diff --git a/.gitattributes b/.gitattributes index 4eae532ef53fa8e97491d57f2284bfacd4b6ca3c..cc87d3d3308d4f5ef56bdca189254abfe752651d 100644 --- a/.gitattributes +++ b/.gitattributes @@ -7081,3 +7081,67 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text 2023/Neural[[:space:]]Texture[[:space:]]Synthesis[[:space:]]With[[:space:]]Guided[[:space:]]Correspondence/0d395694-dab5-4c6a-8e55-9527936bde58_origin.pdf filter=lfs diff=lfs merge=lfs -text 2023/Neural[[:space:]]Transformation[[:space:]]Fields[[:space:]]for[[:space:]]Arbitrary-Styled[[:space:]]Font[[:space:]]Generation/b7d79e74-13fc-423c-9b96-f2beae254d5b_origin.pdf filter=lfs diff=lfs merge=lfs -text 2023/Neural[[:space:]]Vector[[:space:]]Fields_[[:space:]]Implicit[[:space:]]Representation[[:space:]]by[[:space:]]Explicit[[:space:]]Learning/b84bdac4-dddd-4b22-a197-aea99de19635_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Neural[[:space:]]Video[[:space:]]Compression[[:space:]]With[[:space:]]Diverse[[:space:]]Contexts/b731bba7-3b96-4f19-a574-8047f59049bc_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Neural[[:space:]]Volumetric[[:space:]]Memory[[:space:]]for[[:space:]]Visual[[:space:]]Locomotion[[:space:]]Control/dad36583-d1e2-4d95-a7b6-551afd512667_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Neural[[:space:]]Voting[[:space:]]Field[[:space:]]for[[:space:]]Camera-Space[[:space:]]3D[[:space:]]Hand[[:space:]]Pose[[:space:]]Estimation/96d05801-c2f5-4205-ac8b-2fcd00428ae6_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/NeuralDome_[[:space:]]A[[:space:]]Neural[[:space:]]Modeling[[:space:]]Pipeline[[:space:]]on[[:space:]]Multi-View[[:space:]]Human-Object[[:space:]]Interactions/fa0c20cd-9328-405b-b73e-2045fdfa714d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/NeuralEditor_[[:space:]]Editing[[:space:]]Neural[[:space:]]Radiance[[:space:]]Fields[[:space:]]via[[:space:]]Manipulating[[:space:]]Point[[:space:]]Clouds/70b7a137-ced8-4460-839e-a2dd484a031c_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/NeuralField-LDM_[[:space:]]Scene[[:space:]]Generation[[:space:]]With[[:space:]]Hierarchical[[:space:]]Latent[[:space:]]Diffusion[[:space:]]Models/81bfb70e-1652-4684-aa6c-4729296b6ed4_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/NeuralLift-360_[[:space:]]Lifting[[:space:]]an[[:space:]]In-the-Wild[[:space:]]2D[[:space:]]Photo[[:space:]]to[[:space:]]a[[:space:]]3D[[:space:]]Object[[:space:]]With[[:space:]]360deg[[:space:]]Views/90dd78db-4103-415c-8ddc-ba0f83221385_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/NeuralPCI_[[:space:]]Spatio-Temporal[[:space:]]Neural[[:space:]]Field[[:space:]]for[[:space:]]3D[[:space:]]Point[[:space:]]Cloud[[:space:]]Multi-Frame[[:space:]]Non-Linear[[:space:]]Interpolation/865808e9-fa47-4657-baf0-084c6ecae997_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/NeuralUDF_[[:space:]]Learning[[:space:]]Unsigned[[:space:]]Distance[[:space:]]Fields[[:space:]]for[[:space:]]Multi-View[[:space:]]Reconstruction[[:space:]]of[[:space:]]Surfaces[[:space:]]With[[:space:]]Arbitrary[[:space:]]Topologies/cda2b21c-85ea-47d7-952f-4f7b2568e13a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Neuralangelo_[[:space:]]High-Fidelity[[:space:]]Neural[[:space:]]Surface[[:space:]]Reconstruction/55f0b247-41a3-45d9-bec0-4490ab23bed9_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Neuralizer_[[:space:]]General[[:space:]]Neuroimage[[:space:]]Analysis[[:space:]]Without[[:space:]]Re-Training/eee09609-bc86-4647-98dc-269f47c8002f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Neuro-Modulated[[:space:]]Hebbian[[:space:]]Learning[[:space:]]for[[:space:]]Fully[[:space:]]Test-Time[[:space:]]Adaptation/28e8edf6-185c-4dae-9bfc-b17f1c2f7606_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Neuron[[:space:]]Structure[[:space:]]Modeling[[:space:]]for[[:space:]]Generalizable[[:space:]]Remote[[:space:]]Physiological[[:space:]]Measurement/0c500895-2bd8-4cea-80ae-7bc4fd9b8ba5_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/NewsNet_[[:space:]]A[[:space:]]Novel[[:space:]]Dataset[[:space:]]for[[:space:]]Hierarchical[[:space:]]Temporal[[:space:]]Segmentation/d2756edd-2b85-4c25-85cd-99de3bdf0d9a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Next3D_[[:space:]]Generative[[:space:]]Neural[[:space:]]Texture[[:space:]]Rasterization[[:space:]]for[[:space:]]3D-Aware[[:space:]]Head[[:space:]]Avatars/7b615aeb-a836-44b1-beba-de01bf0423b2_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Nighttime[[:space:]]Smartphone[[:space:]]Reflective[[:space:]]Flare[[:space:]]Removal[[:space:]]Using[[:space:]]Optical[[:space:]]Center[[:space:]]Symmetry[[:space:]]Prior/d34edf55-17a4-4aa7-b0c0-de157905c27a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/No[[:space:]]One[[:space:]]Left[[:space:]]Behind_[[:space:]]Improving[[:space:]]the[[:space:]]Worst[[:space:]]Categories[[:space:]]in[[:space:]]Long-Tailed[[:space:]]Learning/aa8d6d83-fa42-452c-97e5-1d8dfd5a2154_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/NoPe-NeRF_[[:space:]]Optimising[[:space:]]Neural[[:space:]]Radiance[[:space:]]Field[[:space:]]With[[:space:]]No[[:space:]]Pose[[:space:]]Prior/2260561b-6c24-4454-9ee0-0e88425e374a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Noisy[[:space:]]Correspondence[[:space:]]Learning[[:space:]]With[[:space:]]Meta[[:space:]]Similarity[[:space:]]Correction/0167b343-6cc5-4960-bab8-e11934c107f0_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/NoisyQuant_[[:space:]]Noisy[[:space:]]Bias-Enhanced[[:space:]]Post-Training[[:space:]]Activation[[:space:]]Quantization[[:space:]]for[[:space:]]Vision[[:space:]]Transformers/68808ca4-cf24-4585-b35b-3e2acef953ea_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/NoisyTwins_[[:space:]]Class-Consistent[[:space:]]and[[:space:]]Diverse[[:space:]]Image[[:space:]]Generation[[:space:]]Through[[:space:]]StyleGANs/c323d83e-f322-4c02-8ecc-1d0c86cf0f70_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Non-Contrastive[[:space:]]Learning[[:space:]]Meets[[:space:]]Language-Image[[:space:]]Pre-Training/fe567c45-5ef2-4da5-8116-0c22fc6e398a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Non-Contrastive[[:space:]]Unsupervised[[:space:]]Learning[[:space:]]of[[:space:]]Physiological[[:space:]]Signals[[:space:]]From[[:space:]]Video/f2bb9236-2bf0-41f5-a980-2c135b55dca1_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Non-Line-of-Sight[[:space:]]Imaging[[:space:]]With[[:space:]]Signal[[:space:]]Superresolution[[:space:]]Network/224cfcfa-2657-4e74-b370-35626fe2670d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Normal-Guided[[:space:]]Garment[[:space:]]UV[[:space:]]Prediction[[:space:]]for[[:space:]]Human[[:space:]]Re-Texturing/6ef67559-1f18-4ef9-b45a-a009b6d372b6_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Normalizing[[:space:]]Flow[[:space:]]Based[[:space:]]Feature[[:space:]]Synthesis[[:space:]]for[[:space:]]Outlier-Aware[[:space:]]Object[[:space:]]Detection/44811963-183d-45a5-8a3b-7dd7378c9155_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Not[[:space:]]All[[:space:]]Image[[:space:]]Regions[[:space:]]Matter_[[:space:]]Masked[[:space:]]Vector[[:space:]]Quantization[[:space:]]for[[:space:]]Autoregressive[[:space:]]Image[[:space:]]Generation/2a88178e-8ce2-4c0f-b3f1-bb99f0452012_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Novel[[:space:]]Class[[:space:]]Discovery[[:space:]]for[[:space:]]3D[[:space:]]Point[[:space:]]Cloud[[:space:]]Semantic[[:space:]]Segmentation/b9953b69-6d05-422e-9f85-28806509ef58_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Novel-View[[:space:]]Acoustic[[:space:]]Synthesis/2dea3110-7f39-4b1f-9a6f-a10e0ea0bd92_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/OCELOT_[[:space:]]Overlapped[[:space:]]Cell[[:space:]]on[[:space:]]Tissue[[:space:]]Dataset[[:space:]]for[[:space:]]Histopathology/052d2272-757d-4e0e-b0c0-b44903da46a8_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/OCTET_[[:space:]]Object-Aware[[:space:]]Counterfactual[[:space:]]Explanations/03950ee8-92a4-474c-99c2-49607c7368f1_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/OPE-SR_[[:space:]]Orthogonal[[:space:]]Position[[:space:]]Encoding[[:space:]]for[[:space:]]Designing[[:space:]]a[[:space:]]Parameter-Free[[:space:]]Upsampling[[:space:]]Module[[:space:]]in[[:space:]]Arbitrary-Scale[[:space:]]Image[[:space:]]Super-Resolution/538d0672-86b3-4cf7-ab0d-b9344b19df37_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/ORCa_[[:space:]]Glossy[[:space:]]Objects[[:space:]]As[[:space:]]Radiance-Field[[:space:]]Cameras/da15cd79-f9cc-4796-854b-9e2af61a06e8_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/OReX_[[:space:]]Object[[:space:]]Reconstruction[[:space:]]From[[:space:]]Planar[[:space:]]Cross-Sections[[:space:]]Using[[:space:]]Neural[[:space:]]Fields/748d73f6-6f8e-4794-a6a8-f014dea611bf_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/OSAN_[[:space:]]A[[:space:]]One-Stage[[:space:]]Alignment[[:space:]]Network[[:space:]]To[[:space:]]Unify[[:space:]]Multimodal[[:space:]]Alignment[[:space:]]and[[:space:]]Unsupervised[[:space:]]Domain[[:space:]]Adaptation/4429e1d5-6fb8-40a5-9804-367d3c0ae901_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/OSRT_[[:space:]]Omnidirectional[[:space:]]Image[[:space:]]Super-Resolution[[:space:]]With[[:space:]]Distortion-Aware[[:space:]]Transformer/b6aea18d-97cd-464e-a4ad-626c19fdc060_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/OT-Filter_[[:space:]]An[[:space:]]Optimal[[:space:]]Transport[[:space:]]Filter[[:space:]]for[[:space:]]Learning[[:space:]]With[[:space:]]Noisy[[:space:]]Labels/69b93138-511e-410f-9d28-563936aa9d2f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/OTAvatar_[[:space:]]One-Shot[[:space:]]Talking[[:space:]]Face[[:space:]]Avatar[[:space:]]With[[:space:]]Controllable[[:space:]]Tri-Plane[[:space:]]Rendering/aea7643c-93a9-4960-bd4a-4385733b4c38_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/OVTrack_[[:space:]]Open-Vocabulary[[:space:]]Multiple[[:space:]]Object[[:space:]]Tracking/092f1f8a-50aa-41bc-8e59-a89dde3c1cc2_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Objaverse_[[:space:]]A[[:space:]]Universe[[:space:]]of[[:space:]]Annotated[[:space:]]3D[[:space:]]Objects/ea3a0563-acc5-4afc-aad6-362a9401875b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Object[[:space:]]Detection[[:space:]]With[[:space:]]Self-Supervised[[:space:]]Scene[[:space:]]Adaptation/b774b6cf-eacd-4660-845f-f32b5ec4dc83_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Object[[:space:]]Discovery[[:space:]]From[[:space:]]Motion-Guided[[:space:]]Tokens/cce05c92-f0c6-4fb5-8422-8cb2ccb76230_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Object[[:space:]]Pop-Up_[[:space:]]Can[[:space:]]We[[:space:]]Infer[[:space:]]3D[[:space:]]Objects[[:space:]]and[[:space:]]Their[[:space:]]Poses[[:space:]]From[[:space:]]Human[[:space:]]Interactions[[:space:]]Alone_/965b59c8-d0c6-413b-852b-108c8ff0bbe7_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Object[[:space:]]Pose[[:space:]]Estimation[[:space:]]With[[:space:]]Statistical[[:space:]]Guarantees_[[:space:]]Conformal[[:space:]]Keypoint[[:space:]]Detection[[:space:]]and[[:space:]]Geometric[[:space:]]Uncertainty[[:space:]]Propagation/a86330a1-e4d7-4fd5-a6ce-7c82335913b8_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Object-Aware[[:space:]]Distillation[[:space:]]Pyramid[[:space:]]for[[:space:]]Open-Vocabulary[[:space:]]Object[[:space:]]Detection/b1741226-bc11-487a-861d-b993de9bcc68_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Object-Goal[[:space:]]Visual[[:space:]]Navigation[[:space:]]via[[:space:]]Effective[[:space:]]Exploration[[:space:]]of[[:space:]]Relations[[:space:]]Among[[:space:]]Historical[[:space:]]Navigation[[:space:]]States/02126feb-2cfd-431e-b8b7-664314fe1580_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/ObjectMatch_[[:space:]]Robust[[:space:]]Registration[[:space:]]Using[[:space:]]Canonical[[:space:]]Object[[:space:]]Correspondences/9d0104e4-40f2-4a9d-bf9d-8065e49cc737_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/ObjectStitch_[[:space:]]Object[[:space:]]Compositing[[:space:]]With[[:space:]]Diffusion[[:space:]]Model/69193070-f3e5-4dd6-8743-b15c03b3e677_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Observation-Centric[[:space:]]SORT_[[:space:]]Rethinking[[:space:]]SORT[[:space:]]for[[:space:]]Robust[[:space:]]Multi-Object[[:space:]]Tracking/b4c4cc9d-52d9-4722-825a-050e1d27cf7f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/OcTr_[[:space:]]Octree-Based[[:space:]]Transformer[[:space:]]for[[:space:]]3D[[:space:]]Object[[:space:]]Detection/b6928c01-d853-4d0d-9874-89b3af0f167a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Occlusion-Free[[:space:]]Scene[[:space:]]Recovery[[:space:]]via[[:space:]]Neural[[:space:]]Radiance[[:space:]]Fields/223bdb66-a260-46a1-86eb-891a595d848a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Octree[[:space:]]Guided[[:space:]]Unoriented[[:space:]]Surface[[:space:]]Reconstruction/0db48127-a127-468b-be90-b27ec5813046_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Omni[[:space:]]Aggregation[[:space:]]Networks[[:space:]]for[[:space:]]Lightweight[[:space:]]Image[[:space:]]Super-Resolution/39175e81-b86e-49a9-967e-e590c3dbdaa0_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Omni3D_[[:space:]]A[[:space:]]Large[[:space:]]Benchmark[[:space:]]and[[:space:]]Model[[:space:]]for[[:space:]]3D[[:space:]]Object[[:space:]]Detection[[:space:]]in[[:space:]]the[[:space:]]Wild/2efbd142-17ac-4a1f-b83f-1ae025c49292_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/OmniAL_[[:space:]]A[[:space:]]Unified[[:space:]]CNN[[:space:]]Framework[[:space:]]for[[:space:]]Unsupervised[[:space:]]Anomaly[[:space:]]Localization/390b7405-6295-46ac-87cb-a78d4a95dd63_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/OmniAvatar_[[:space:]]Geometry-Guided[[:space:]]Controllable[[:space:]]3D[[:space:]]Head[[:space:]]Synthesis/13b82072-fc3f-4f4d-8e9e-44b54de26be7_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/OmniCity_[[:space:]]Omnipotent[[:space:]]City[[:space:]]Understanding[[:space:]]With[[:space:]]Multi-Level[[:space:]]and[[:space:]]Multi-View[[:space:]]Images/1ecc869d-a020-42e0-bc1d-485949718b11_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/OmniMAE_[[:space:]]Single[[:space:]]Model[[:space:]]Masked[[:space:]]Pretraining[[:space:]]on[[:space:]]Images[[:space:]]and[[:space:]]Videos/2fdd9b20-eaac-4038-a247-56813890042d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/OmniObject3D_[[:space:]]Large-Vocabulary[[:space:]]3D[[:space:]]Object[[:space:]]Dataset[[:space:]]for[[:space:]]Realistic[[:space:]]Perception,[[:space:]]Reconstruction[[:space:]]and[[:space:]]Generation/243023a1-e64c-4bdc-a9c5-c30584f803ba_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/OmniVidar_[[:space:]]Omnidirectional[[:space:]]Depth[[:space:]]Estimation[[:space:]]From[[:space:]]Multi-Fisheye[[:space:]]Images/e34f49e7-aba6-408f-9647-b38f210ab7c3_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/Omnimatte3D_[[:space:]]Associating[[:space:]]Objects[[:space:]]and[[:space:]]Their[[:space:]]Effects[[:space:]]in[[:space:]]Unconstrained[[:space:]]Monocular[[:space:]]Video/5700f2ea-1d90-4ebb-a4e2-33ec303e3c7c_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/On[[:space:]]Calibrating[[:space:]]Semantic[[:space:]]Segmentation[[:space:]]Models_[[:space:]]Analyses[[:space:]]and[[:space:]]an[[:space:]]Algorithm/f35086d8-d1cb-49d4-972d-c5bbdb5412ed_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/On[[:space:]]Data[[:space:]]Scaling[[:space:]]in[[:space:]]Masked[[:space:]]Image[[:space:]]Modeling/23a2cca2-7ab3-48ca-b93b-e899d7bc87f3_origin.pdf filter=lfs diff=lfs merge=lfs -text +2023/On[[:space:]]Distillation[[:space:]]of[[:space:]]Guided[[:space:]]Diffusion[[:space:]]Models/2a7f0724-3ab3-4aec-b0e2-13e799a34131_origin.pdf filter=lfs diff=lfs merge=lfs -text diff --git a/2023/Neural Video Compression With Diverse Contexts/b731bba7-3b96-4f19-a574-8047f59049bc_content_list.json b/2023/Neural Video Compression With Diverse Contexts/b731bba7-3b96-4f19-a574-8047f59049bc_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..50aa79863482355744119b47cbbbb308cb009f34 --- /dev/null +++ b/2023/Neural Video Compression With Diverse Contexts/b731bba7-3b96-4f19-a574-8047f59049bc_content_list.json @@ -0,0 +1,1379 @@ +[ + { + "type": "text", + "text": "Neural Video Compression with Diverse Contexts", + "text_level": 1, + "bbox": [ + 233, + 130, + 736, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jiahao Li, Bin Li, Yan Lu \nMicrosoft Research Asia", + "bbox": [ + 380, + 181, + 588, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{li.jiahao, libin, yanlu}@microsoft.com", + "bbox": [ + 308, + 218, + 655, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "For any video codecs, the coding efficiency highly relies on whether the current signal to be encoded can find the relevant contexts from the previous reconstructed signals. Traditional codec has verified more contexts bring substantial coding gain, but in a time-consuming manner. However, for the emerging neural video codec (NVC), its contexts are still limited, leading to low compression ratio. To boost NVC, this paper proposes increasing the context diversity in both temporal and spatial dimensions. First, we guide the model to learn hierarchical quality patterns across frames, which enriches long-term and yet high-quality temporal contexts. Furthermore, to tap the potential of optical flow-based coding framework, we introduce a group-based offset diversity where the cross-group interaction is proposed for better context mining. In addition, this paper also adopts a quadtree-based partition to increase spatial context diversity when encoding the latent representation in parallel. Experiments show that our codec obtains $23.5\\%$ bitrate saving over previous SOTA NVC. Better yet, our codec has surpassed the under-developing next generation traditional codec/ECM in both RGB and YUV420 colorspaces, in terms of PSNR. The codes are at https://github.com/microsoft/DCVC.", + "bbox": [ + 76, + 301, + 473, + 648 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 676, + 209, + 694 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The philosophy of video codec is that, for the current signal to be encoded, the codec will find the relevant contexts (e.g., various predictions as the contexts) from previous reconstructed signals to reduce the spatial-temporal redundancy. The more relevant contexts are, the higher bitrate saving is achieved.", + "bbox": [ + 75, + 703, + 468, + 794 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "If looking back the development of traditional CODECs (from H.261 [17] in 1988 to H.266 [7] in 2020), we find that the coding gain mainly comes from the continuously expanded coding modes, where each mode uses a specific manner to extract and utilize context. For example, the numbers of intra prediction directions [42] in H.264, H.265, H.266 are 9, 35, and 65, respectively. So many modes can", + "bbox": [ + 75, + 794, + 470, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/aade64995bb23ac8e4c264999b6e9cf627d3fc0b5899c0996b22bb4e42959693.jpg", + "image_caption": [ + "Figure 1. Average results on UVG, MCL-JCV, and HEVC datasets. All traditional codecs use their best compression ratio configuration. DCVC-HEM [29] is the previous SOTA NVC and only has released the model for RGB colorspace." + ], + "image_footnote": [], + "bbox": [ + 501, + 267, + 890, + 431 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "extract diverse contexts to reduce redundancy, but also bring huge complexity as rate distortion optimization (RDO) is used to search the best mode. For encoding a 1080p frame, the under-developing ECM (the prototype of next generation traditional codec) needs up to half an hour [49]. Although some DL-based methods [24,51,52] proposed accelerating traditional codecs, the complexity is still very high.", + "bbox": [ + 496, + 520, + 890, + 627 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "By contrast, neural video codec (NVC) changes the extraction and utilization of context from hand-crafted design to automatic-learned manner. Mainstream frameworks of NVC can be classified into residual coding-based [1, 13, 31, 32, 34, 36, 47, 59, 61] and condition coding-based [21, 27-29, 33, 38, 50]. The residual coding explicitly uses the predicted frame as the context, and the context utilization is restricted to use subtraction for redundancy removal. By comparison, conditional coding implicitly learns feature domain contexts. The high dimension contexts can carry richer information to facilitate encoding, decoding, as well as entropy modelling.", + "bbox": [ + 496, + 627, + 892, + 808 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, for most NVCs, the manners of context extraction and utilization are still limited, e.g., only using optical flow to explore temporal correlation. This makes NVC easily suffer from the uncertainty [12, 16, 37] in parameters or fall into local optimum [25]. One solution is adding traditional codec-like coding modes into NVC [25]. But it brings", + "bbox": [ + 496, + 810, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "22616", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "large computational complexity as RDO is used. So the question comes up: how to better learn and use the contexts while yielding low computational cost?", + "bbox": [ + 75, + 90, + 468, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To this end, based on DCVC (deep contextual video compression) [28] framework and its following work DCVC-HEM [29], we propose a new model DCVC-DC which efficiently utilizes the Diverse Contexts to further boost compression ratio. At first, we guide DCVC-DC to learn hierarchical quality pattern across frames. With this guidance during the training, the long-term and yet high-quality contexts which are vital for the reconstruction of the following frames are implicitly learned during the feature propagation. This helps further exploit the long-range temporal correlation in video and effectively alleviate the quality degradation problem existed in most NVCs. In addition, we adopt the offset diversity [8] to strengthen the optical flow-based codec, where multiple offsets can reduce the warping errors for complex or large motions. In particular, inspired by the weighted prediction in traditional codec, the offsets are divided into groups and the cross-group fusion is proposed to improve the temporal context mining.", + "bbox": [ + 75, + 136, + 470, + 407 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Besides from temporal dimension, this paper also proposes increasing the spatial context diversity when encoding the latent representation. Based on recent checkerboard model [19] and dual spatial model [29, 56], we design a quadtree-based partition to improve the distribution estimation. When compared with [19, 29], the types of correlation modelling are more diverse hence the model has a larger chance to find more relevant context.", + "bbox": [ + 75, + 409, + 468, + 527 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "It is noted that all our designs are parallel-efficient. To further reduce the computational cost, we also adopt depthwise separable convolution [10], and assign unequal channel numbers for features with different resolutions. Experiments show that our DCVC-DC achieves much higher efficiency over previous SOTA NVC and pushes the compression ratio to a new height. When compared with DCVCHEM [29], $23.5\\%$ bitrate saving is achieved while MACs (multiply-accumulate operations) are reduced by $19.4\\%$ . Better yet, besides H.266-VTM 17.0, our codec also already outperforms ECM-5.0 (its best compression ratio configuration for low delay coding is used) in both RGB and YUV420 colorspaces, as shown in Fig. 1. To the best of our knowledge, this is the first NVC which can achieve such accomplishment. In summary, our contributions are:", + "bbox": [ + 75, + 529, + 470, + 755 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose efficiently increasing context diversity to boost NVC. Diverse contexts are complementary to each other and have larger chance to provide good reference for reducing redundancy.", + "- From temporal dimension, we guide model to extract high-quality contexts to alleviate the quality degradation problem. In addition, the group-based offset diversity is designed for better temporal context mining." + ], + "bbox": [ + 96, + 768, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- From spatial dimension, we adopt a quadtree-based partition for latent representation. This provides diverse spatial contexts for better entropy coding.", + "- Our DCVC-DC obtains $23.5\\%$ bitrate saving over the previous SOTA NVC. In particular, our DCVC-DC has surpassed the best traditional codec ECM in both RGB and YUV420 colorspaces, which is an important milestone in the development of NVC." + ], + "bbox": [ + 517, + 90, + 890, + 220 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 232, + 640, + 247 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Neural Image Compression", + "text_level": 1, + "bbox": [ + 500, + 257, + 746, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Most neural image CODECs are based on hyperprior [4] where some bits are first used to provide basic contexts for entropy coding. Then, the auto-regressive prior [40] proposes using neighbour contexts to capture spatial correlation. Recently works [18, 26, 44, 45] propose extracting global or long-range contexts to further boost performance. These show more diverse contexts bring substantial coding gain for neural image codec.", + "bbox": [ + 496, + 279, + 890, + 401 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Neural Video Compression", + "text_level": 1, + "bbox": [ + 500, + 409, + 741, + 425 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent years also have witnessed the prosperity of NVC. The pioneering DVC [34] follows traditional codec. It uses optical flow network to generate prediction frame, then its residual with the current frame is coded. Many subsequent works also adopt this residual coding-based framework and refine the modules therein. For example, [31, 43, 47] proposed motion prediction to further reduce redundancy. Optical flow estimation in scale-space [1] was designed to handle complex motion. Yang et al. [61] utilized recurrent autoencoder to improve coding efficiency.", + "bbox": [ + 496, + 431, + 890, + 583 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Residual coding explicitly generates predicted frame in pixel domain as the context and only uses the subtraction to remove redundancy. By comparison, conditional coding has stronger extensibility. The definition, learning, and usage manner of condition can be flexibly designed. In [33, 38], temporally conditional entropy models were designed. [27] uses conditional coding to encode the foreground contents. Li et al. proposed DCVC [28] to learn feature domain contexts to increase context capacity. Then DCVC-TCM [50] adopts feature propagation to boost performance. Recently, DCVC-HEM [29] designs the hybrid entropy model utilizing both spatial and temporal contexts.", + "bbox": [ + 496, + 584, + 890, + 763 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, the coding modes in most NVCs are still limited when compared with traditional codec. For example, traditional codec adopts translational/affine motion models, geometric partition, bi-prediction, and so on modes to extract diverse temporal contexts [7]. By contrast, existing NVCs usually only relies on single optical flow, which is easily influenced by epistemic uncertainty [12, 16, 37] in model parameters. The recent work [25] also shows such NVCs easily fall into local optimum when coding mode is", + "bbox": [ + 496, + 765, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "22617", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/91c53fd11f964f018075f8bad8cd27199ed184db2cd34e9f8673f2eab7d071e4.jpg", + "image_caption": [ + "Figure 2. Framework overview of our DCVC-DC. $x_{t}$ and $\\hat{x}_t$ are the input and reconstructed frames. $C_t$ is the learned temporal context as the condition for coding $x_{t}$ . $F_{t}$ is the propagated but unprocessed feature used for next frame. In the loss term, $r_t$ means the bit cost for coding whole frame. $dist(\\cdot)$ is distortion function. $\\lambda$ and $w_{t}$ are the global and frame-level weights, respectively. The number in the spatial partition for $\\hat{y}_t$ represent the coding order index." + ], + "image_footnote": [], + "bbox": [ + 81, + 99, + 517, + 333 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/cfcc72d2798cbf87cfc50b3ccc1ae55c764493d36deacac373d993aff64d33c0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 519, + 99, + 885, + 333 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "limited. So [25] designed many additional modes like traditional codec, and RDO is used to search the best mode. Such method inevitably brings large computational cost. By comparison, our model has no additional inference cost when exploiting high-quality temporal contexts. Our offset diversity and quadtree partition are also time-efficient designs in providing diverse contexts.", + "bbox": [ + 75, + 414, + 470, + 521 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Proposed Method", + "text_level": 1, + "bbox": [ + 76, + 535, + 250, + 551 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 76, + 559, + 187, + 574 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To achieve higher compression ratio, our codec is built on the more flexible conditional coding rather than the residual coding. The framework of our DCVC-DC is illustrated in Fig. 2. It is noted our DCVC-DC is designed for low-delay coding as it can be applied in more scenarios, e.g., real-time communication. As shown in Fig. 2, for coding each frame $x_{t}$ with frame index $t$ , our coding pipeline contains three core steps: $f_{motion}$ , $f_{T\\text{context}}$ , and $f_{frame}$ . At first, $f_{motion}$ uses optical flow network to estimate the motion vector (MV) $v_{t}$ , then $v_{t}$ is encoded and decoded as $\\hat{v}_{t}$ . Second, based on $\\hat{v}_{t}$ and the propagated feature $F_{t-1}$ from the previous frame, $f_{T\\text{context}}$ extracts the motion-aligned temporal context feature $C_{t}$ . At last, conditioned on $C_{t}$ , $f_{frame}$ encodes $x_{t}$ into quantized latent representation $\\hat{y}_{t}$ . After entropy coding, the output frame $\\hat{x}_{t}$ is reconstructed via the decoder and frame generator. At the same time, $F_{t}$ is also generated and propagated to the next frame. It is noted that, our DCVC-DC is based on DCVC-HEM [29]. When compared with DCVC-HEM, this paper redesigns the modules to exploit Diverse Contexts from both temporal (Section 3.2 and 3.3) and spatial (Section 3.4)", + "bbox": [ + 75, + 583, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "dimensions.", + "bbox": [ + 500, + 415, + 583, + 429 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Hierarchical Quality Structure", + "text_level": 1, + "bbox": [ + 500, + 439, + 772, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Traditional codec widely adopts hierarchical quality structure, where frames are assigned into different layers and then use different QPs (quantization parameters). This design originates from scalable video coding [48] but also improves the performance for general low delay coding from two aspects. One is periodically improving the quality can alleviate the error propagation, as shown in Fig. 3. During the inter prediction, the high-quality reference frames enable the codec to find the more accurate MV during motion estimation. Meanwhile, the motion compensated prediction is also high-quality, leading to smaller prediction error. Another aspect is that, powered by multiple reference frame selection and weighted prediction mechanisms, the prediction combinations from the nearest reference frame and long-range high-quality reference frame are more diverse. The work [30] investigates many settings on frame quality and reference frame selection, and concludes that the hierarchical quality structure with referencing both the nearest frame and farther high-quality frame achieves the best performance.", + "bbox": [ + 496, + 462, + 890, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Inspired by the success in traditional codec, we are thinking whether we can equip NVC with the hierarchical quality structure and let NVC also enjoy the benefits. Considering the recent neural CODECs [11,29] also support variable bitrate in single model, one straightforward solution is following the traditional codec and directly assigning hierarchical QPs during the inference of NVC. However, not like traditional codec uses well-defined rules to perform motion estimation and motion compensation (MEMC), NVC uses neural net", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "22618", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2ca98f5fce310bd2f63647b2335caef438745cc784471bde5d203dcbf9ccf4ed.jpg", + "image_caption": [ + "Figure 3. Hierarchical quality structure in H.266-VTM 17.0 and our NVC. This example is from BasketballPass video from HEVC D dataset. The average bpp (bits per pixel) and PSNR of H.266 are 0.056 and 35.10. Our DCVC-DC is with 0.045 and 36.13." + ], + "image_footnote": [], + "bbox": [ + 81, + 95, + 464, + 253 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "works and MEMC is often in feature domain. For NVC, the advantage of such design is that it is automatic learned and has larger potential to achieve better performance. The disadvantage is that it has wore robustness and generalization ability for out-of-distribution quality pattern. Thus, if we directly feed NVC the hierarchical QPs during the testing like [22], it may not well adapt to the hierarchical quality pattern, and the MEMC may get sub-optimal performance.", + "bbox": [ + 75, + 319, + 467, + 439 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To this end, we propose guiding the NVC to learn the hierarchical quality pattern during the training. Specifically, we add a weight $w_{t}$ for each frame in the rate-distortion loss, as shown in Fig. 2. The setting of $w_{t}$ follows the hierarchical structure. Powered by this hierarchical distortion loss, both high-quality output frame $\\hat{x}_{t}$ and feature $F_{t}$ containing many high-frequency details are periodically generated. They are very helpful for improving the MEMC effectiveness and then alleviate the error propagation problem that many other NVCs suffer from. In addition, via the cascaded training across multiple frames, the feature propagation chain is formed. The high-quality contexts which are vital for the reconstruction of the following frames are automatically learned and kept in long range. Thus, for the encoding of $x_{t}$ , $F_{t-1}$ not only contains the short-term contexts extracted from $x_{t-1}$ , but also provides long-term and continually-updated high-quality contexts from many previous frames. Such diverse $F_{t-1}$ helps further exploit the temporal correlation across many frames and then boost the compression ratio. Fig. 3 also shows the quality pattern of our NVC. We can see that our DCVC-DC achieves better average quality while with smaller bit cost than H.266.", + "bbox": [ + 75, + 441, + 467, + 773 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Group-Based Offset Diversity", + "text_level": 1, + "bbox": [ + 76, + 785, + 339, + 801 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Due to the various motions between frames, directly using the unprocessed $F_{t - 1}$ without motion alignment is hard for codec to capture temporal correspondence. Therefore, we follow existing NVCs and use optical flow network to extract motion aligned temporal context $C_t$ via $f_{T\\text{context}}(F_{t - 1}, \\hat{v}_t)$ , where $\\hat{v}_t$ is the decoded MV. However,", + "bbox": [ + 75, + 810, + 467, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/65fae7d00d04e6b4565a477a4e0ee6f4a56aaae079f25fc49c7d60c09841ee9c.jpg", + "image_caption": [ + "Figure 4. Cross-group fusion module. In this example, the group number $G$ is 4 and the offset number $N$ of each group is 2." + ], + "image_footnote": [], + "bbox": [ + 500, + 95, + 890, + 202 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "in most existing NVCs, $f_{T\\text{context}}$ is only the warping operation with the single MV. Such single motion-based alignment is not robust to complex motions or occlusions. The works [9, 54, 58] show deformable alignment gets better results for video restoration as each location has multiple offsets to capture temporal correspondence. So recently it is also applied into NVC [23]. However, the training of deformable alignment is not stable and the overflow of offsets degrades the performance [8, 58]. In addition, the number of offsets is limited to the size of deformable convolutional kernel. Thus, this paper adopts the more flexible design called offset diversity [8]. Meanwhile, the decoded MV is used as the base offset to stabilize the training as [9].", + "bbox": [ + 496, + 250, + 890, + 445 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Fig. 2, our $f_{T\\text{context}}$ consists of two core sub-modules: offset prediction and cross-group fusion. At first, offset prediction uses the decoded MV $\\hat{v}_t$ to predict the residual offsets $d_t$ , where $\\hat{x}_{t-1}$ and $F_{t-1}$ are also warped and fed as the auxiliary information. The $d_t$ adds the base offset $\\hat{v}_t$ to obtain the final offsets $o_t$ . At the same time, offset prediction also generates the modulation mask $m_t$ which can be regarded as an attention that reflects the confidence of offsets. It is noted that $F_{t-1}$ is divided into $G$ groups along the channel dimension, and each group has separate $N$ offsets. Thus, there are a total of $G \\times N$ offsets learned. The diverse offsets are complementary to each other, and help codec cope with complex motion and occlusion.", + "bbox": [ + 496, + 446, + 890, + 642 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In addition, motivated by the channel shuffle operation [62] which improves the information flow in the CNN backbone, we customize a group-level interaction mechanism to further tap the potential of offset diversity for NVC. In particular, after warping each group with multiple offsets and applying the corresponding masks, we will reorder all groups before the fusion, as shown in Fig. 4. If using $g_{i}^{j}$ to represent the $i$ -th group warped with its $j$ -th offset, the features before reordering are $g_{0}^{0}, \\dots, g_{0}^{N-1}, g_{1}^{0}, \\dots, g_{1}^{N-1}, \\dots, g_{G-1}^{0}, \\dots, g_{G-1}^{N-1}$ , where the offset order is primary and group order is secondary. Then we reorder them as $g_{0}^{0}, \\dots, g_{G-1}^{0}, g_{1}^{1}, \\dots, g_{G-1}^{1}, \\dots, g_{G-1}^{N-1}, \\dots, g_{G-1}^{N-1}$ , where the group order is primary instead. The following fusion operation will fuse every $N$ contiguous groups into one group. Therefore, during this process, the group reordering enables more cross-group interactions without increasing", + "bbox": [ + 496, + 643, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "22619", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/89ce4849cdaa74b4c7c581f0991ef2995c43cbd600ea3b30602aa310d02470c8.jpg", + "image_caption": [ + "Figure 5. Entropy coding with quadtree partition. The number means the coding order index. During the 4 coding steps, the $\\hat{y}_{t-1}$ from previous frame, hyper prior $\\hat{z}_t$ , and temporal context $C_t$ are also used for entropy modelling." + ], + "image_footnote": [], + "bbox": [ + 81, + 93, + 467, + 345 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "complexity. This design also enjoys the similar benefit with the weighted prediction from different reference frames in traditional codec. Via the cross-group fusion, more diverse combinations in extracting temporal contexts from different groups are introduced and further improve the effectiveness of offset diversity.", + "bbox": [ + 75, + 420, + 468, + 512 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Quadtree Partition-Based Entropy Coding", + "text_level": 1, + "bbox": [ + 76, + 525, + 439, + 542 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After obtaining temporal context $C_t$ via our offset diversity module, the input frame $x_{t}$ will be encoded and quantized as $\\hat{y}_t$ , conditioned on $C_t$ , as shown in Fig. 2. We need to estimate the probability mass function (PMF) of $\\hat{y}_t$ for its arithmetic coding. In this process, how to build an accurate entropy model to estimate the PMF of $\\hat{y}_t$ is vital for the compression efficiency.", + "bbox": [ + 75, + 550, + 468, + 655 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Many neural codecs adopt the auto-regressive model [40] as entropy model. However, it seriously slows down the coding speed. By contrast, the checkerboard model [19] proposes coding the even positions of $\\hat{y}_t$ first, and then use them to predict the PMF of the odd positions in parallel. Recently the dual spatial model [29] improves it by utilizing the correlation along channel dimension. However, the neighbours used for entropy modelling in [19, 29] are still limited when compared with auto-regressive model. Thus, inspired from [41, 46], this paper proposes a finer-grained coding manner via the quadtree partition, where diverse spatial contexts are exploited to improve entropy modelling.", + "bbox": [ + 75, + 657, + 468, + 838 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Fig. 5, we first divide $\\hat{y}_t$ into four groups along the channel dimension. Then each group is partitioned into non-overlapped $2\\times 2$ patches in spatial dimension. The whole entropy coding is divided into four steps,", + "bbox": [ + 75, + 839, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and each step codes the different positions associated with the corresponding indexes in Fig. 5. At 0th step, all positions with index 0 of all patches are coded at the same time. It is noted, for the four groups, the positions with index 0 are different from each other. Thus, for every spatial position, there are one fourth channels (i.e., one group) encoded. In the subsequent 1st, 2nd, and 3rd steps, all positions coded in previous steps are used for predicting the PMF of the positions coded in the current step, and different spatial positions are coded for different groups in each step.", + "bbox": [ + 496, + 90, + 890, + 242 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During this process, more diverse neighbours are utilized. If considering the 8 spatial neighbours for a position, the auto-regressive model [40] uses 4 (left, top-left, top, top-right) neighbours for every position if not considering the boundary region. The checkerboard and dual spatial models [19,29] uses 0 and 4 (left, top, right, bottom) neighbours for the 0th and 1st steps, respectively. By contrast, as shown in Fig. 5, our DCVC-DC uses 0, 4, 4, and 8 neighbours for the four steps, respectively. On average, the neighbour number in DCVC-DC is 2 times of that of [19, 29] and is same with that of auto-regressive model. However, our model is much more time-efficient than auto-regressive model as all positions in each step can be coded in parallel. In addition, our model also exploits the cross-channel correlation, which is like [29] but in a refined way. For example, at the 3rd step, for one specific position of a group, the other channels at the same position are already coded from different groups in previous steps, and they can be used as the contexts for the entropy modelling in this step. This helps further squeeze the redundancy. Overall, our quadtree partition-based solution makes the entropy coding benefit from the finer-grained and diverse contexts, which fully mines the correlation from both spatial and channel dimensions.", + "bbox": [ + 496, + 243, + 892, + 590 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Implementation", + "text_level": 1, + "bbox": [ + 500, + 604, + 658, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our DCVC-DC is based on DCVC-HEM [29] but focuses on exploiting Diverse Contexts to further boost performance. In addition, we also make the following improvements to obtain better tradeoff between performance and complexity. The first is that, considering depthwise separable convolution [10] can reduce the computation cost while alleviating over-fitting, we widely use it to replace the normal convolution in the basic block design. The second is that we use the unequal channel number settings for features with different resolutions, where the higher resolution feature is assigned with smaller channel number for acceleration. The third is that we move partial quantization operations to higher resolution in the encoder, which helps achieve more precise bit allocation. The harmonization of encoding and quantization also brings some compression ratio improvements. The section 4.3 verifies the effectiveness of these structure optimizations, and the detailed network structures can be found in supplementary materials.", + "bbox": [ + 496, + 628, + 892, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "22620", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1e433e77b60fdaff6aa4e5ead59590cab7c9fc64f01f2cea49c3dfe50e4fe467.jpg", + "image_caption": [ + "Figure 6. Rate and distortion curves. The comparison is in RGB colorspace measured with PSNR. More results including the corresponding MS-SSIM curves and the comparison in YUV420 colorspace are in supplementary materials." + ], + "image_footnote": [], + "bbox": [ + 80, + 90, + 279, + 229 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/74ec3365b457d92b2328efc384a3722e16d496d36bcfd15bf3abd6ef4be8e531.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 279, + 90, + 480, + 229 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8af78d728ffa426c113f3d37e8e7e7b9aaeef9e0a2929803f983246efbbdda01.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 480, + 90, + 681, + 229 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b97c9dc01dce74bd66cc11175621edb25b1caae0931db3d9be4eeacdfc0b109d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 681, + 90, + 880, + 229 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/7affc2c93a0b44c625d915bb4c5543db790943202d2c38c3d1dec491c37ff7eb.jpg", + "table_caption": [ + "Table 1. BD-Rate (%) comparison in RGB colorspace measured with PSNR. The anchor is VTM-17.0." + ], + "table_footnote": [], + "table_body": "
UVGMCL-JCVHEVC BHEVC CHEVC DHEVC EHEVC RGBAverage
VTM-17.00.00.00.00.00.00.00.00.0
HM-16.2536.441.538.836.033.744.039.438.5
ECM-5.0-10.0-12.2-11.5-13.4-13.5-10.9-11.1-11.8
CANF-VC [21]73.070.864.476.263.1118.079.977.9
DCVC [28]166.1121.6123.2143.298.0266.1113.4147.4
DCVC-TCM [50]44.151.040.266.337.082.724.449.4
DCVC-HEM [29]1.18.65.122.22.420.5-9.97.1
Our DCVC-DC-19.1-11.3-12.0-10.3-26.1-18.0-27.6-17.8
", + "bbox": [ + 122, + 297, + 843, + 465 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In particular, our DCVC-DC also already outperforms ECM in YUV420 colorspace. Not like traditional codec needs many hand-crafted changes in designing coding tools for different colorspaces, DCVC-DC only just needs simple adoptions based on a existing model trained for RGB. Without changing the network structure, we only up-sample the UV to use the unified input interface with RGB. Correspondingly, after obtaining the reconstructed frame, it is down-sampled on UV. In addition, our model for YUV420 just needs a simple finetune training based on the model trained for RGB.", + "bbox": [ + 75, + 470, + 468, + 636 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experimental Results", + "text_level": 1, + "bbox": [ + 76, + 654, + 282, + 671 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 76, + 679, + 282, + 696 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. For training, we follow most existing NVCs and use Video-90k [60]. For testing YUV420 videos, HEVC B~E [6], UVG [39], and MCL-JCV [57] are used. Their raw format is YUV420, so there is no any change before feeding them to NVC. For testing RGB videos, as these testsets have no RGB format, most existing NVCs use BT.601 (default in FFmpeg) to convert them from YUV420 to RGB. Actually, JPEG AI [2,3] adopts BT.709 because using BT.709 obtains higher compression ratio under similar visual quality. Thus, this paper follows JPEG AI and uses BT.709 for all codecs during testing RGB. It is noted that the relative bitrate comparisons between different codecs are similar in BT.601 and BT.709. The supplementary ma", + "bbox": [ + 75, + 703, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "terials show the results using BT.601. In addition, we follow [29, 50] and also test HEVC RGB dataset [15] when testing RGB videos, and there is no format change as HEVC RGB dataset itself is in RGB format.", + "bbox": [ + 496, + 470, + 890, + 531 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Test Conditions. We follow [29, 50] and test 96 frames for each video, and the intra period is set as 32. The low delay encoding setting is used, as the same with most existing works [1, 28, 34]. BD-Rate [5] is used to measure the compression ratio, where negative numbers indicate bitrate saving and positive numbers indicate bitrate increase.", + "bbox": [ + 496, + 532, + 890, + 625 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our benchmarks include HM [20] and VTM [55] which represent the best H.265 and H.266 encoder, respectively. In particular, we also compare with ECM [14] which is the prototype of next generation traditional codec. For the codec setting, we follow [29, 50] and further use 10-bit as the intermediate representation when testing RGB, which leads to better compression ratio for the three traditional codecs. The detailed settings are shown in supplementary materials. As for the NVC benchmarks, we compare with the recent SOTA models including CANF-VC [21], DCVC [28], DCVC-TCM [50], and DCVC-HEM [29].", + "bbox": [ + 496, + 626, + 890, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Model Training. We adopt the multi-stage training strategy as [29, 50]. Our model also supports variable bitrate in single model [29], so different $\\lambda$ values are used in different optimization steps. We follow [29] and use $4\\lambda$ values (85, 170, 380, 840). But different from [29] using constant distortion weight in the loss, this paper proposes using hierarchical weight setting on $w_{t}$ for the distortion term (the", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "22621", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ecc99482c4f1fad9f00376e153cddd72f99b04166031d6506ccf1b38e5e2dd59.jpg", + "table_caption": [ + "Table 2. BD-Rate (%) comparison in RGB colorspace measured with MS-SSIM. The anchor is VTM-17.0." + ], + "table_footnote": [], + "table_body": "
UVGMCL-JCVHEVC BHEVC CHEVC DHEVC EHEVC RGBAverage
VTM-17.00.00.00.00.00.00.00.00.0
HM-16.2531.138.836.635.233.041.136.636.1
ECM-5.0-9.1-11.1-10.2-11.7-11.0-9.9-9.8-10.4
CANF-VC [21]46.526.043.530.917.9173.057.756.5
DCVC [28]64.927.554.439.715.2210.451.366.2
DCVC-TCM [50]1.0-10.8-11.7-15.2-29.016.7-22.2-10.2
DCVC-HEM [29]-25.2-36.3-38.0-38.3-48.1-25.8-43.6-36.5
Our DCVC-DC-32.6-44.8-47.8-49.8-58.2-45.8-54.4-47.6
", + "bbox": [ + 124, + 114, + 843, + 280 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/725e2695db4ae035b7601b87958faf1968e9f61fc03882c560d92d8d6161cb9f.jpg", + "table_caption": [ + "Table 3. BD-Rate (%) comparison in YUV420 colorspace measured with PSNR. The anchor is VTM-17.0." + ], + "table_footnote": [], + "table_body": "
UVGMCL-JCVHEVC BHEVC CHEVC DHEVC EAverage
VTM-17.00.00.00.00.00.00.00.0
HM-16.2536.742.539.233.330.040.737.1
ECM-5.0-10.6-13.7-12.6-14.7-14.9-12.1-13.1
Our DCVC-DC-17.8-12.0-10.8-12.4-28.5-20.4-17.0
", + "bbox": [ + 176, + 316, + 789, + 412 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "whole loss is defined in Fig. 2). Considering our training set Vimeo-90k only has 7 frames for each video, we refer traditional codec setting and set the pattern size as 4. The $w_{t}$ settings for 4 consecutive frames are (0.5, 1.2, 0.5, 0.9).", + "bbox": [ + 75, + 428, + 468, + 491 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Comparisons with Previous SOTA Methods", + "text_level": 1, + "bbox": [ + 76, + 498, + 446, + 513 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "RGB colorspace. Table 1 and 2 show the BR-rate comparison using RGB videos in terms of PSNR and MS-SSIM, respectively. From Table 1, we find our codec achieves significant compression ratio improvement over VTM on every dataset, and there is an average of $17.8\\%$ bitrate saving. By contrast, the other neural CODEs are still worse than VTM. If using DCVC-HEM as anchor, our average bitrate saving is $23.5\\%$ . In addition, our DCVC-DC also outperform ECM from Table 1. If using ECM as anchor, an average of $6.4\\%$ bitrate saving is achieved.", + "bbox": [ + 75, + 522, + 467, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fig. 6 shows the rate-distortion curves. From the curves, we can see our DCVC-DC achieves the SOTA compression ratio in wide bitrate range. When using MS-SSIM as quality metric, our DCVC-DC shows larger improvement. As shown in Table 2, DCVC-DC has an average of $47.6\\%$ bitrate saving over VTM. By contrast, the corresponding number of ECM over VTM is only $10.4\\%$ .", + "bbox": [ + 75, + 672, + 467, + 779 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "It is noted that Table 1 and 2 use the RGB video with BT.709 conversion. If with BT.601 conversion, the relative bitrate saving is similar with that in BT.709. For example, with BT.601 conversion, DCVC-DC over VTM has an average of $18.0\\%$ bitrate saving in terms of PSNR. More results with BT.601 can be seen in supplementary materials.", + "bbox": [ + 75, + 779, + 467, + 869 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "YUV420 colorspace. Actually traditional codecs are mainly optimized in YUV420. Thus, the comparison in", + "bbox": [ + 76, + 869, + 467, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "YUV420 is also very important for evaluating the progress of NVC over traditional codec. The corresponding results are shown in Table 3. The numbers in this table are calculated using the weighted PSNR for the three components of YUV. The weights are $(6,1,1)/8$ , which are consistent with that in standard committee [53]. As most NVCs have no corresponding released models for YUV420, Table 3 only reports the numbers of our NVC. We can see that DCVC-DC has an average of $17.0\\%$ bitrate saving over VTM. If only considering the Y component, an average $15.3\\%$ bitrate saving is achieved over VTM. Better yet, our DCVC-DC also outperforms ECM in YUV420 on average, as shown in Table 3. This is an important milestone in the development of NVC. It is noted our codec uses the same network structure for both RGB and YUV420 colorspaces, where only different refinunings are used during training. This shows the simplicity and strong extensibility of our codec on the optimization for different input colorspaces.", + "bbox": [ + 496, + 428, + 890, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 710, + 651, + 726 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To verify the effectiveness of each component, we conduct comprehensive ablation studies. For simplification, the HEVC datasets in RGB colorspace are used here. The average BD-Rate in terms of PSNR is shown.", + "bbox": [ + 496, + 734, + 890, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Diverse Contexts. Table 4 shows the study on the effectiveness of diverse contexts. First, from the comparison between $M_e$ and $M_d$ , we can see that the BD-rate is reduced from $21.3\\%$ to $14.7\\%$ . This large difference shows the substantial coding gain of our qaudtree partition-based entropy coding, and verifies the advantages of diverse spatial and channel contexts via finer-grained partition.", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "22622", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c22f58e743e86ef2e7225182deaeb9eb40b49a229fd9969b3732e8af440178b7.jpg", + "table_caption": [ + "Table 4. Ablation Study on Diverse Contexts." + ], + "table_footnote": [], + "table_body": "
MaMbMcMdMe
Hierarchical quality structure✓
Offset diversity w/ cross-group✓✓
Offset diversity w/o cross-group [8]✓
Quadtree partition based model✓✓✓✓
Dual spatial model [29]✓
BD-Rate(%)0.08.412.114.721.3
", + "bbox": [ + 99, + 114, + 447, + 287 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "From temporal dimension, we also design hierarchical quality structure and offset diversity with cross-group interaction. In Table 4, based on $M_{d}$ , we first test the original offset diversity [8] without cross-group interaction, i.e., removing the reorder operation in Fig. 4. However, it $(M_{c})$ only brings $2.6\\%$ BD-rate difference. By contrast, powered by our cross-group interaction, the potential of offset diversion is fully tapped, and $M_{b}$ reduces the BD-rate number by $6.3\\%$ over $M_{d}$ . At last, based on $M_{b}$ , we evaluate the hierarchical quality structure, i.e., $M_{a}$ . The $8.4\\%$ gap shows learning high-quality contexts brings large benefits to the mining of temporal correlation across many frames.", + "bbox": [ + 75, + 294, + 468, + 474 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Structure optimization. Although our codec learns utilizing diverse contexts in efficient manner, we still purse better tradeoff between compression ratio and computational cost. Therefore, we further optimize our model in network structure. Table 5 shows the study. Based on the $M_{a}$ (same with that in Table 4), we first implement the depthwise separable convolution into codec. $M_{h}$ shows widely using depthwise separable convolution to replace the normal convolution not only significantly reduces the MACs, but also brings some compression ratio improvements.", + "bbox": [ + 75, + 476, + 468, + 626 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The second acceleration is that we use the unequal channel number settings. Not like many existing NVCs use the same channel number for features with different resolutions, we propose assigning the larger number for low resolution feature to increase the latent representation capacity while using the smaller number for the high resolution feature to accelerate model. The performance of $M_{g}$ verifies the effectiveness of our improvement. In addition, many existing NVCs perform the quantization at the low-resolution latent representation after encoding. To achieve more precise rate adjustment, this paper moves partial quantization operations to the higher resolution in the encoder. $M_{f}$ shows the integration of encoding and quantization brings some BD-rate improvements with negligible MAC change.", + "bbox": [ + 75, + 627, + 468, + 838 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Complexity", + "text_level": 1, + "bbox": [ + 76, + 847, + 202, + 863 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The complexity comparison is shown in Table 6. We find the MACs of our DCVC-DC are reduced by $19.4\\%$ when", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/d6943a336157e62405948371d8cb56279be4a28fbe2d476f134656fd3ec831cb.jpg", + "table_caption": [ + "Table 5. Ablation Study on Structure Optimization." + ], + "table_footnote": [], + "table_body": "
MfMgMhMa
Quant at high resolution✓
Unequal channel setting✓✓
Depthwise separable conv✓✓✓
MACs2642G2642G2939G3456G
BD-Rate (%)0.01.12.43.5
", + "bbox": [ + 506, + 114, + 890, + 227 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e0bec5e8158c76a54b506bfc95acbc507bb24a42f71fa51b5f265fb2ca354a67.jpg", + "table_caption": [ + "Table 6. Complexity comparison." + ], + "table_footnote": [ + "Note: Tested on NVIDIA 2080TI with using $1080\\mathrm{p}$ as input." + ], + "table_body": "
MACsEncoding TimeDecoding Time
DCVC-HEM [29]3279G890ms652ms
Our DCVC-DC2642G1005ms765ms
", + "bbox": [ + 509, + 266, + 885, + 321 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "compared with DCVC-HEM [29]. However, the actual encoding and decoding time is higher. This is because currently the computational density of depthwise convolution is not as high as normal convolution under the same MAC condition. But through the customized optimization [35], it can be further accelerated in the future. From another perspective, considering that our DCVC-DC has $23.5\\%$ bitrate saving over previous SOTA DCVC-HEM [29], such increase degree in running time is a price worth paying. By contrast, ECM brings $13.1\\%$ (Table 3) improvement over its predecessor VTM, but the encoding complexity is more than 4 times [49] of VTM.", + "bbox": [ + 496, + 345, + 890, + 526 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion and Limitation", + "text_level": 1, + "bbox": [ + 500, + 551, + 748, + 566 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we have presented how to utilize diverse contexts to further boost NVC. From temporal dimension, the model is guided to extract long-term and yet high-quality contexts to alleviate error propagation and exploit long range correlation. The offset diversity with cross-group interaction provides complementary motion alignments to handle complex motion. From spatial dimension, the fine-grained quadtree-based partition is proposed to increase spatial context diversity. Powered by our techniques, the compression ratio of NVC has been pushed to new height. Our DCVC-DC has surpassed ECM in both RGB and YUV420 colorspaces, which is an important milestone in the development of NVC.", + "bbox": [ + 496, + 580, + 890, + 775 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "During the training, to learn the hierarchical quality pattern, we still use the fixed distortion weights which are similar with those in traditional codec. This may not be the best choice for NVC. Actually, reinforcement learning is good at solving such kind of time series weight decision problem. In the future, we will investigate utilizing reinforcement learning to help NVC make better weight decision with considering the temporal dependency.", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "22623", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Eirikur Agustsson, David Minnen, Nick Johnston, Johannes Balle, Sung Jin Hwang, and George Toderici. Scale-space flow for end-to-end optimized video compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8503-8512, 2020. 1, 2, 6", + "[2] E. Alshina, J. Ascenso, T. Ebrahimi, F. Pereira, and T. Richter. [AHG 11] Brief information about JPEG AI CfP status. In JVET-AA0047, 2022. 6", + "[3] Anchors · JPEG-AI MMSP Challenge. https://jpegai.github.io/7-anchors/, 2022. Accessed: 2022-11-02. 6", + "[4] Johannes Balle, David Minnen, Saurabh Singh, Sung Jin Hwang, and Nick Johnston. Variational image compression with a scale hyperprior. 6th International Conference on Learning Representations, ICLR, 2018. 2", + "[5] Gisle Bjontegaard. Calculation of average PSNR differences between RD-curves. VCEG-M33, 2001. 6", + "[6] Frank Bossen et al. Common test conditions and software reference configurations. In JCTVC-L1100, 2013. 6", + "[7] Benjamin Bross, Ye-Kui Wang, Yan Ye, Shan Liu, Jianle Chen, Gary J Sullivan, and Jens-Rainer Ohm. Overview of the versatile video coding (VVC) standard and its applications. IEEE Transactions on Circuits and Systems for Video Technology, 31(10):3736-3764, 2021. 1, 2", + "[8] Kelvin CK Chan, Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Understanding deformable alignment in video super-resolution. In Proceedings of the AAAI conference on artificial intelligence, 2021. 2, 4, 8", + "[9] Kelvin CK Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. BasicVSR++: Improving video superresolution with enhanced propagation and alignment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5972-5981, 2022. 4", + "[10] François Chollet. Xception: Deep learning with depthwise separable convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1251-1258, 2017. 2, 5", + "[11] Ze Cui, Jing Wang, Shangyin Gao, Tiansheng Guo, Yihui Feng, and Bo Bai. Asymmetric gained deep image compression with continuous rate adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10532-10541, 2021. 3", + "[12] Armen Der Kiureghian and Ove Ditlevsen. Aleatory or epistemic? does it matter? Structural safety, 31(2):105-112, 2009. 1, 2", + "[13] Abdelaziz Djelouah, Joaquim Campos, Simone Schaub-Meyer, and Christopher Schroers. Neural inter-frame compression for video coding. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2019. 1", + "[14] ECM-5.0. https://vgit.hhi.fraunhofer.de/ecm/ECM, 2022. Accessed: 2022-11-02. 6", + "[15] D Flynn, K Sharman, and C Rosewarne. Common test conditions and software reference configurations for hevc range" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "extensions, document jctvc-n1006. Joint Collaborative Team Video Coding ITU-T SG, 16, 2013. 6", + "[16] Yarin Gal. Uncertainty in Deep Learning. PhD thesis, University of Cambridge, 2016. 1, 2", + "[17] Bernd Girod, Eckehard G Steinbach, and Niko Faerber. Comparison of the H. 263 and H. 261 video compression standards. In Standards and Common Interfaces for Video Information Systems: A Critical Review, 1995. 1", + "[18] Zongyu Guo, Zhizheng Zhang, Runsen Feng, and Zhibo Chen. Causal contextual prediction for learned image compression. IEEE Transactions on Circuits and Systems for Video Technology, 32(4):2329-2341, 2021. 2", + "[19] Dailan He, Yaoyan Zheng, Baocheng Sun, Yan Wang, and Hongwei Qin. Checkerboard context model for efficient learned image compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14771-14780, 2021. 2, 5", + "[20] HM-16.25. https://vgit.hhi.fraunhofer.de/jvet/HM/, 2022. Accessed: 2022-11-02. 6", + "[21] Yung-Han Ho, Chih-Peng Chang, Peng-Yu Chen, Alessandro Gnutti, and Wen-Hsiao Peng. Canf-vc: Conditional augmented normalizing flows for video compression. European Conference on Computer Vision, 2022. 1, 6, 7", + "[22] Zhihao Hu, Guo Lu, Jinyang Guo, Shan Liu, Wei Jiang, and Dong Xu. Coarse-to-fine deep video coding with hyperprior-guided mode prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5921-5930, 2022. 4", + "[23] Zhihao Hu, Guo Lu, and Dong Xu. FVC: A new framework towards deep video compression in feature space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1502-1511, 2021. 4", + "[24] Yan Huang, Li Song, and Ebroul Izquierdo. CNN accelerated intra video coding, where is the upper bound? In 2019 Picture Coding Symposium (PCS), pages 1-5. IEEE, 2019. 1", + "[25] Shuai Huo, Dong Liu, Li Li, Siwei Ma, Feng Wu, and Wen Gao. Towards hybrid-optimization video coding. arXiv preprint arXiv:2207.05565, 2022. 1, 2, 3", + "[26] Jun-Hyuk Kim, Byeongho Heo, and Jong-Seok Lee. Joint global and local hierarchical priors for learned image compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5992-6001, 2022. 2", + "[27] Théo Ladune, Pierick Philippe, Wassim Hamidouche, Lu Zhang, and Olivier Déforges. Conditional coding for flexible learned video compression. In Neural Compression: From Information Theory to Applications - Workshop @ ICLR, 2021. 1, 2", + "[28] Jiahao Li, Bin Li, and Yan Lu. Deep contextual video compression. Advances in Neural Information Processing Systems, 34, 2021. 1, 2, 6, 7", + "[29] Jiahao Li, Bin Li, and Yan Lu. Hybrid spatial-temporal entropy modelling for neural video compression. In Proceedings of the 30th ACM International Conference on Multimedia, pages 1503-1511, 2022. 1, 2, 3, 5, 6, 7, 8", + "[30] Chong Soon Lim, SMT Naing, V Wahadaniah, and X Jing. Reference lists for B pictures under low delay constraints." + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "22624", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "document JCTVC-D093, ITU-T/ISO/IEC Joint Collaborative Team on Video Coding (JCT-VC), Daegu, Korea, 2011. 3", + "[31] Jianping Lin, Dong Liu, Houqiang Li, and Feng Wu. M-LVC: multiple frames prediction for learned video compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 1, 2", + "[32] Haojie Liu, Ming Lu, Zhan Ma, Fan Wang, Zhihuang Xie, Xun Cao, and Yao Wang. Neural video coding using multiscale motion compensation and spatiotemporal context model. IEEE Transactions on Circuits and Systems for Video Technology, 2020. 1", + "[33] Jerry Liu, Shenlong Wang, Wei-Chiu Ma, Meet Shah, Rui Hu, Pranaab Dhawan, and Raquel Urtasun. Conditional entropy coding for efficient video compression. In European Conference on Computer Vision, pages 453-468. Springer, 2020. 1, 2", + "[34] Guo Lu, Wanli Ouyang, Dong Xu, Xiaoyun Zhang, Chunlei Cai, and Zhiyong Gao. DVC: an end-to-end deep video compression framework. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11006-11015, 2019. 1, 2, 6", + "[35] Gangzhao Lu, Weizhe Zhang, and Zheng Wang. Optimizing depthwise separable convolution operations on gpus. IEEE Transactions on Parallel and Distributed Systems, 33(1):70-87, 2021. 8", + "[36] Guo Lu, Xiaoyun Zhang, Wanli Ouyang, Li Chen, Zhiyong Gao, and Dong Xu. An end-to-end learning framework for video compression. IEEE transactions on pattern analysis and machine intelligence, 2020. 1", + "[37] Wufei Ma, Jiahao Li, Bin Li, and Yan Lu. Uncertainty-aware deep video compression with ensembles, 2021. 1, 2", + "[38] Fabian Mentzer, George Toderici, David Minnen, Sung-Jin Hwang, Sergi Caelles, Mario Lucic, and Eirikur Agustsson. Vct: A video compression transformer. arXiv preprint arXiv:2206.07307, 2022. 1, 2", + "[39] Alexandre Mercat, Marko Viitanen, and Jarno Vanne. UVG dataset: 50/120fps 4k sequences for video codec analysis and development. In Proceedings of the 11th ACM Multimedia Systems Conference, pages 297-302, 2020. 6", + "[40] David Minnen, Johannes Ballé, and George D Toderici. Joint autoregressive and hierarchical priors for learned image compression. Advances in neural information processing systems, 31, 2018. 2, 5", + "[41] Ken M Nakanishi, Shin-ichi Maeda, Takeru Miyato, and Daisuke Okanohara. Neural multi-scale image compression. In Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2–6, 2018, Revised Selected Papers, Part VI 14, pages 718–732. Springer, 2019. 5", + "[42] Jonathan Pfaff, Alexey Filippov, Shan Liu, Xin Zhao, Jianle Chen, Santiago De-Luxan-Hernandez, Thomas Wiegand, Vasily Rufitskiy, Adarsh Krishnan Ramasubramonian, and Geert Van der Auwera. Intra prediction and mode coding in VLC. IEEE Transactions on Circuits and Systems for Video Technology, 31(10):3834-3847, 2021. 1", + "[43] Reza Pourreza, Hoang Le, Amir Said, Guillaume Sautiere, and Auke Wiggers. Boosting neural video codec" + ], + "bbox": [ + 78, + 90, + 470, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "by exploiting hierarchical redundancy. arXiv preprint arXiv:2208.04303, 2022. 2", + "[44] Yichen Qian, Ming Lin, Xiuyu Sun, Zhiyu Tan, and Rong Jin. Entroformer: A transformer-based entropy model for learned image compression. arXiv preprint arXiv:2202.05492, 2022. 2", + "[45] Yichen Qian, Zhiyu Tan, Xiuyu Sun, Ming Lin, Dongyang Li, Zhenhong Sun, Hao Li, and Rong Jin. Learning accurate entropy model with global reference for image compression. arXiv preprint arXiv:2010.08321, 2020. 2", + "[46] Scott Reed, Aaron Oord, Nal Kalchbrenner, Sergio Gomez Colmenarejo, Ziyu Wang, Yutian Chen, Dan Belov, and Nando Freitas. Parallel multiscale autoregressive density estimation. In International Conference on Machine Learning, pages 2912-2921. PMLR, 2017. 5", + "[47] Oren Rippel, Alexander G Anderson, Kedar Tatwawadi, Sanjay Nair, Craig Lytle, and Lubomir Bourdev. ELF-VC: Efficient learned flexible-rate video coding. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14479-14488, October 2021. 1, 2", + "[48] Heiko Schwarz, Detlev Marpe, and Thomas Wiegand. Analysis of hierarchical B pictures and MCTF. In 2006 IEEE International Conference on Multimedia and Expo, pages 1929-1932. IEEE, 2006. 3", + "[49] Vadim Seregin, Jie Chen, Fabrice Leannec, and Kai Zhang. JVET AHG report: ECM software development (AHG6). In JVET-AA0006, 2022. 1, 8", + "[50] Xihua Sheng, Jiahao Li, Bin Li, Li Li, Dong Liu, and Yan Lu. Temporal Context Mining for Learned Video Compression. IEEE Transactions on Multimedia, 2022. 1, 2, 6, 7", + "[51] Hui Su, Mingliang Chen, Alexander Bokov, Debargha Mukherjee, Yunqing Wang, and Yue Chen. Machine learning accelerated transform search for av1. In 2019 Picture Coding Symposium (PCS), pages 1-5. IEEE, 2019. 1", + "[52] Hui Su, Chi-Yo Tsai, Yunqing Wang, and Yaowu Xu. Machine learning accelerated partition search for video encoding. In 2019 IEEE International Conference on Image Processing (ICIP), pages 2661-2665. IEEE, 2019. 1", + "[53] Gary J Sullivan and Jens-Rainer Ohm. Meeting report of the fourth meeting of the joint collaborative team on video coding (jct-vc), daegu, kr, 20-28 january 2011. Document JCTVC-D500, Daegu, KR, 2011. 7", + "[54] Yapeng Tian, Yulun Zhang, Yun Fu, and Chenliang Xu. TDAN: Temporally-deformable alignment network for video super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3360-3369, 2020. 4", + "[55] VTM-17.0. https://vcert.hhi.fraunhofer.de/jvet/VVCSoftware_VTM/, 2022. Accessed: 2022-11-02.6", + "[56] Guo-Hua Wang, Jiahao Li, Bin Li, and Yan Lu. EVC: Towards Real-Time Neural Image Compression with Mask Decay. In International Conference on Learning Representations, 2023. 2", + "[57] Haiqiang Wang, Weihao Gan, Sudeng Hu, Joe Yuchieh Lin, Lina Jin, Longguang Song, Ping Wang, Ioannis Katsavounidis, Anne Aaron, and C-C Jay Kuo. MCL-JCV: a JND-based" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "22625", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "H. 264/AVC video quality assessment dataset. In 2016 IEEE International Conference on Image Processing (ICIP), pages 1509-1513. IEEE, 2016. 6", + "[58] Xintao Wang, Kelvin CK Chan, Ke Yu, Chao Dong, and Chen Change Loy. EDVR: Video restoration with enhanced deformable convolutional networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 0–0, 2019. 4", + "[59] Chao-Yuan Wu, Nayan Singhal, and Philipp Krahenbuhl. Video compression through image interpolation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 416-431, 2018. 1", + "[60] Tianfan Xue, Baian Chen, Jiajun Wu, Donglai Wei, and William T Freeman. Video enhancement with task-oriented flow. International Journal of Computer Vision (IJCV), 127(8):1106-1125, 2019. 6", + "[61] Ren Yang, Fabian Mentzer, Luc Van Gool, and Radu Timofte. Learning for video compression with recurrent autoencoder and recurrent probability model. IEEE Journal of Selected Topics in Signal Processing, 15(2):388-401, 2021. 1, 2", + "[62] Xiangyu Zhang, Xinyu Zhou, Mengxiao Lin, and Jian Sun. Shufflenet: An extremely efficient convolutional neural network for mobile devices. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6848-6856, 2018. 4" + ], + "bbox": [ + 78, + 90, + 468, + 455 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "22626", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Neural Video Compression With Diverse Contexts/b731bba7-3b96-4f19-a574-8047f59049bc_model.json b/2023/Neural Video Compression With Diverse Contexts/b731bba7-3b96-4f19-a574-8047f59049bc_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ccca78d739f51958d90147313a0f90fffbe6e9d7 --- /dev/null +++ b/2023/Neural Video Compression With Diverse Contexts/b731bba7-3b96-4f19-a574-8047f59049bc_model.json @@ -0,0 +1,2147 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.131, + 0.737, + 0.154 + ], + "angle": 0, + "content": "Neural Video Compression with Diverse Contexts" + }, + { + "type": "text", + "bbox": [ + 0.382, + 0.182, + 0.589, + 0.216 + ], + "angle": 0, + "content": "Jiahao Li, Bin Li, Yan Lu \nMicrosoft Research Asia" + }, + { + "type": "text", + "bbox": [ + 0.31, + 0.219, + 0.656, + 0.234 + ], + "angle": 0, + "content": "{li.jiahao, libin, yanlu}@microsoft.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.302, + 0.474, + 0.65 + ], + "angle": 0, + "content": "For any video codecs, the coding efficiency highly relies on whether the current signal to be encoded can find the relevant contexts from the previous reconstructed signals. Traditional codec has verified more contexts bring substantial coding gain, but in a time-consuming manner. However, for the emerging neural video codec (NVC), its contexts are still limited, leading to low compression ratio. To boost NVC, this paper proposes increasing the context diversity in both temporal and spatial dimensions. First, we guide the model to learn hierarchical quality patterns across frames, which enriches long-term and yet high-quality temporal contexts. Furthermore, to tap the potential of optical flow-based coding framework, we introduce a group-based offset diversity where the cross-group interaction is proposed for better context mining. In addition, this paper also adopts a quadtree-based partition to increase spatial context diversity when encoding the latent representation in parallel. Experiments show that our codec obtains \\(23.5\\%\\) bitrate saving over previous SOTA NVC. Better yet, our codec has surpassed the under-developing next generation traditional codec/ECM in both RGB and YUV420 colorspaces, in terms of PSNR. The codes are at https://github.com/microsoft/DCVC." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.678, + 0.21, + 0.695 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.47, + 0.795 + ], + "angle": 0, + "content": "The philosophy of video codec is that, for the current signal to be encoded, the codec will find the relevant contexts (e.g., various predictions as the contexts) from previous reconstructed signals to reduce the spatial-temporal redundancy. The more relevant contexts are, the higher bitrate saving is achieved." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.471, + 0.902 + ], + "angle": 0, + "content": "If looking back the development of traditional CODECs (from H.261 [17] in 1988 to H.266 [7] in 2020), we find that the coding gain mainly comes from the continuously expanded coding modes, where each mode uses a specific manner to extract and utilize context. For example, the numbers of intra prediction directions [42] in H.264, H.265, H.266 are 9, 35, and 65, respectively. So many modes can" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.268, + 0.892, + 0.433 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.447, + 0.895, + 0.504 + ], + "angle": 0, + "content": "Figure 1. Average results on UVG, MCL-JCV, and HEVC datasets. All traditional codecs use their best compression ratio configuration. DCVC-HEM [29] is the previous SOTA NVC and only has released the model for RGB colorspace." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.521, + 0.892, + 0.628 + ], + "angle": 0, + "content": "extract diverse contexts to reduce redundancy, but also bring huge complexity as rate distortion optimization (RDO) is used to search the best mode. For encoding a 1080p frame, the under-developing ECM (the prototype of next generation traditional codec) needs up to half an hour [49]. Although some DL-based methods [24,51,52] proposed accelerating traditional codecs, the complexity is still very high." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.628, + 0.893, + 0.809 + ], + "angle": 0, + "content": "By contrast, neural video codec (NVC) changes the extraction and utilization of context from hand-crafted design to automatic-learned manner. Mainstream frameworks of NVC can be classified into residual coding-based [1, 13, 31, 32, 34, 36, 47, 59, 61] and condition coding-based [21, 27-29, 33, 38, 50]. The residual coding explicitly uses the predicted frame as the context, and the context utilization is restricted to use subtraction for redundancy removal. By comparison, conditional coding implicitly learns feature domain contexts. The high dimension contexts can carry richer information to facilitate encoding, decoding, as well as entropy modelling." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.895, + 0.902 + ], + "angle": 0, + "content": "However, for most NVCs, the manners of context extraction and utilization are still limited, e.g., only using optical flow to explore temporal correlation. This makes NVC easily suffer from the uncertainty [12, 16, 37] in parameters or fall into local optimum [25]. One solution is adding traditional codec-like coding modes into NVC [25]. But it brings" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "22616" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.137 + ], + "angle": 0, + "content": "large computational complexity as RDO is used. So the question comes up: how to better learn and use the contexts while yielding low computational cost?" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.137, + 0.471, + 0.408 + ], + "angle": 0, + "content": "To this end, based on DCVC (deep contextual video compression) [28] framework and its following work DCVC-HEM [29], we propose a new model DCVC-DC which efficiently utilizes the Diverse Contexts to further boost compression ratio. At first, we guide DCVC-DC to learn hierarchical quality pattern across frames. With this guidance during the training, the long-term and yet high-quality contexts which are vital for the reconstruction of the following frames are implicitly learned during the feature propagation. This helps further exploit the long-range temporal correlation in video and effectively alleviate the quality degradation problem existed in most NVCs. In addition, we adopt the offset diversity [8] to strengthen the optical flow-based codec, where multiple offsets can reduce the warping errors for complex or large motions. In particular, inspired by the weighted prediction in traditional codec, the offsets are divided into groups and the cross-group fusion is proposed to improve the temporal context mining." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.41, + 0.47, + 0.529 + ], + "angle": 0, + "content": "Besides from temporal dimension, this paper also proposes increasing the spatial context diversity when encoding the latent representation. Based on recent checkerboard model [19] and dual spatial model [29, 56], we design a quadtree-based partition to improve the distribution estimation. When compared with [19, 29], the types of correlation modelling are more diverse hence the model has a larger chance to find more relevant context." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.53, + 0.471, + 0.756 + ], + "angle": 0, + "content": "It is noted that all our designs are parallel-efficient. To further reduce the computational cost, we also adopt depthwise separable convolution [10], and assign unequal channel numbers for features with different resolutions. Experiments show that our DCVC-DC achieves much higher efficiency over previous SOTA NVC and pushes the compression ratio to a new height. When compared with DCVCHEM [29], \\(23.5\\%\\) bitrate saving is achieved while MACs (multiply-accumulate operations) are reduced by \\(19.4\\%\\). Better yet, besides H.266-VTM 17.0, our codec also already outperforms ECM-5.0 (its best compression ratio configuration for low delay coding is used) in both RGB and YUV420 colorspaces, as shown in Fig. 1. To the best of our knowledge, this is the first NVC which can achieve such accomplishment. In summary, our contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.77, + 0.468, + 0.83 + ], + "angle": 0, + "content": "- We propose efficiently increasing context diversity to boost NVC. Diverse contexts are complementary to each other and have larger chance to provide good reference for reducing redundancy." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.84, + 0.468, + 0.901 + ], + "angle": 0, + "content": "- From temporal dimension, we guide model to extract high-quality contexts to alleviate the quality degradation problem. In addition, the group-based offset diversity is designed for better temporal context mining." + }, + { + "type": "list", + "bbox": [ + 0.097, + 0.77, + 0.468, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.092, + 0.891, + 0.137 + ], + "angle": 0, + "content": "- From spatial dimension, we adopt a quadtree-based partition for latent representation. This provides diverse spatial contexts for better entropy coding." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.146, + 0.892, + 0.221 + ], + "angle": 0, + "content": "- Our DCVC-DC obtains \\(23.5\\%\\) bitrate saving over the previous SOTA NVC. In particular, our DCVC-DC has surpassed the best traditional codec ECM in both RGB and YUV420 colorspaces, which is an important milestone in the development of NVC." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.092, + 0.892, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.233, + 0.642, + 0.248 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.258, + 0.747, + 0.273 + ], + "angle": 0, + "content": "2.1. Neural Image Compression" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.28, + 0.892, + 0.402 + ], + "angle": 0, + "content": "Most neural image CODECs are based on hyperprior [4] where some bits are first used to provide basic contexts for entropy coding. Then, the auto-regressive prior [40] proposes using neighbour contexts to capture spatial correlation. Recently works [18, 26, 44, 45] propose extracting global or long-range contexts to further boost performance. These show more diverse contexts bring substantial coding gain for neural image codec." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.41, + 0.743, + 0.426 + ], + "angle": 0, + "content": "2.2. Neural Video Compression" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.433, + 0.892, + 0.584 + ], + "angle": 0, + "content": "Recent years also have witnessed the prosperity of NVC. The pioneering DVC [34] follows traditional codec. It uses optical flow network to generate prediction frame, then its residual with the current frame is coded. Many subsequent works also adopt this residual coding-based framework and refine the modules therein. For example, [31, 43, 47] proposed motion prediction to further reduce redundancy. Optical flow estimation in scale-space [1] was designed to handle complex motion. Yang et al. [61] utilized recurrent autoencoder to improve coding efficiency." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.585, + 0.892, + 0.765 + ], + "angle": 0, + "content": "Residual coding explicitly generates predicted frame in pixel domain as the context and only uses the subtraction to remove redundancy. By comparison, conditional coding has stronger extensibility. The definition, learning, and usage manner of condition can be flexibly designed. In [33, 38], temporally conditional entropy models were designed. [27] uses conditional coding to encode the foreground contents. Li et al. proposed DCVC [28] to learn feature domain contexts to increase context capacity. Then DCVC-TCM [50] adopts feature propagation to boost performance. Recently, DCVC-HEM [29] designs the hybrid entropy model utilizing both spatial and temporal contexts." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.766, + 0.892, + 0.901 + ], + "angle": 0, + "content": "However, the coding modes in most NVCs are still limited when compared with traditional codec. For example, traditional codec adopts translational/affine motion models, geometric partition, bi-prediction, and so on modes to extract diverse temporal contexts [7]. By contrast, existing NVCs usually only relies on single optical flow, which is easily influenced by epistemic uncertainty [12, 16, 37] in model parameters. The recent work [25] also shows such NVCs easily fall into local optimum when coding mode is" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "22617" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.101, + 0.518, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.101, + 0.887, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.343, + 0.895, + 0.4 + ], + "angle": 0, + "content": "Figure 2. Framework overview of our DCVC-DC. \\( x_{t} \\) and \\( \\hat{x}_t \\) are the input and reconstructed frames. \\( C_t \\) is the learned temporal context as the condition for coding \\( x_{t} \\). \\( F_{t} \\) is the propagated but unprocessed feature used for next frame. In the loss term, \\( r_t \\) means the bit cost for coding whole frame. \\( dist(\\cdot) \\) is distortion function. \\( \\lambda \\) and \\( w_{t} \\) are the global and frame-level weights, respectively. The number in the spatial partition for \\( \\hat{y}_t \\) represent the coding order index." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.415, + 0.471, + 0.522 + ], + "angle": 0, + "content": "limited. So [25] designed many additional modes like traditional codec, and RDO is used to search the best mode. Such method inevitably brings large computational cost. By comparison, our model has no additional inference cost when exploiting high-quality temporal contexts. Our offset diversity and quadtree partition are also time-efficient designs in providing diverse contexts." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.536, + 0.251, + 0.552 + ], + "angle": 0, + "content": "3. Proposed Method" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.56, + 0.188, + 0.575 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.584, + 0.471, + 0.903 + ], + "angle": 0, + "content": "To achieve higher compression ratio, our codec is built on the more flexible conditional coding rather than the residual coding. The framework of our DCVC-DC is illustrated in Fig. 2. It is noted our DCVC-DC is designed for low-delay coding as it can be applied in more scenarios, e.g., real-time communication. As shown in Fig. 2, for coding each frame \\( x_{t} \\) with frame index \\( t \\), our coding pipeline contains three core steps: \\( f_{motion} \\), \\( f_{T\\text{context}} \\), and \\( f_{frame} \\). At first, \\( f_{motion} \\) uses optical flow network to estimate the motion vector (MV) \\( v_{t} \\), then \\( v_{t} \\) is encoded and decoded as \\( \\hat{v}_{t} \\). Second, based on \\( \\hat{v}_{t} \\) and the propagated feature \\( F_{t-1} \\) from the previous frame, \\( f_{T\\text{context}} \\) extracts the motion-aligned temporal context feature \\( C_{t} \\). At last, conditioned on \\( C_{t} \\), \\( f_{frame} \\) encodes \\( x_{t} \\) into quantized latent representation \\( \\hat{y}_{t} \\). After entropy coding, the output frame \\( \\hat{x}_{t} \\) is reconstructed via the decoder and frame generator. At the same time, \\( F_{t} \\) is also generated and propagated to the next frame. It is noted that, our DCVC-DC is based on DCVC-HEM [29]. When compared with DCVC-HEM, this paper redesigns the modules to exploit Diverse Contexts from both temporal (Section 3.2 and 3.3) and spatial (Section 3.4)" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.416, + 0.584, + 0.43 + ], + "angle": 0, + "content": "dimensions." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.44, + 0.774, + 0.456 + ], + "angle": 0, + "content": "3.2. Hierarchical Quality Structure" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.463, + 0.892, + 0.765 + ], + "angle": 0, + "content": "Traditional codec widely adopts hierarchical quality structure, where frames are assigned into different layers and then use different QPs (quantization parameters). This design originates from scalable video coding [48] but also improves the performance for general low delay coding from two aspects. One is periodically improving the quality can alleviate the error propagation, as shown in Fig. 3. During the inter prediction, the high-quality reference frames enable the codec to find the more accurate MV during motion estimation. Meanwhile, the motion compensated prediction is also high-quality, leading to smaller prediction error. Another aspect is that, powered by multiple reference frame selection and weighted prediction mechanisms, the prediction combinations from the nearest reference frame and long-range high-quality reference frame are more diverse. The work [30] investigates many settings on frame quality and reference frame selection, and concludes that the hierarchical quality structure with referencing both the nearest frame and farther high-quality frame achieves the best performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Inspired by the success in traditional codec, we are thinking whether we can equip NVC with the hierarchical quality structure and let NVC also enjoy the benefits. Considering the recent neural CODECs [11,29] also support variable bitrate in single model, one straightforward solution is following the traditional codec and directly assigning hierarchical QPs during the inference of NVC. However, not like traditional codec uses well-defined rules to perform motion estimation and motion compensation (MEMC), NVC uses neural net" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22618" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.096, + 0.465, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.257, + 0.47, + 0.313 + ], + "angle": 0, + "content": "Figure 3. Hierarchical quality structure in H.266-VTM 17.0 and our NVC. This example is from BasketballPass video from HEVC D dataset. The average bpp (bits per pixel) and PSNR of H.266 are 0.056 and 35.10. Our DCVC-DC is with 0.045 and 36.13." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.32, + 0.468, + 0.44 + ], + "angle": 0, + "content": "works and MEMC is often in feature domain. For NVC, the advantage of such design is that it is automatic learned and has larger potential to achieve better performance. The disadvantage is that it has wore robustness and generalization ability for out-of-distribution quality pattern. Thus, if we directly feed NVC the hierarchical QPs during the testing like [22], it may not well adapt to the hierarchical quality pattern, and the MEMC may get sub-optimal performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.442, + 0.468, + 0.775 + ], + "angle": 0, + "content": "To this end, we propose guiding the NVC to learn the hierarchical quality pattern during the training. Specifically, we add a weight \\( w_{t} \\) for each frame in the rate-distortion loss, as shown in Fig. 2. The setting of \\( w_{t} \\) follows the hierarchical structure. Powered by this hierarchical distortion loss, both high-quality output frame \\( \\hat{x}_{t} \\) and feature \\( F_{t} \\) containing many high-frequency details are periodically generated. They are very helpful for improving the MEMC effectiveness and then alleviate the error propagation problem that many other NVCs suffer from. In addition, via the cascaded training across multiple frames, the feature propagation chain is formed. The high-quality contexts which are vital for the reconstruction of the following frames are automatically learned and kept in long range. Thus, for the encoding of \\( x_{t} \\), \\( F_{t-1} \\) not only contains the short-term contexts extracted from \\( x_{t-1} \\), but also provides long-term and continually-updated high-quality contexts from many previous frames. Such diverse \\( F_{t-1} \\) helps further exploit the temporal correlation across many frames and then boost the compression ratio. Fig. 3 also shows the quality pattern of our NVC. We can see that our DCVC-DC achieves better average quality while with smaller bit cost than H.266." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.786, + 0.341, + 0.803 + ], + "angle": 0, + "content": "3.3. Group-Based Offset Diversity" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.468, + 0.902 + ], + "angle": 0, + "content": "Due to the various motions between frames, directly using the unprocessed \\(F_{t - 1}\\) without motion alignment is hard for codec to capture temporal correspondence. Therefore, we follow existing NVCs and use optical flow network to extract motion aligned temporal context \\(C_t\\) via \\(f_{T\\text{context}}(F_{t - 1}, \\hat{v}_t)\\), where \\(\\hat{v}_t\\) is the decoded MV. However," + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.096, + 0.892, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.217, + 0.892, + 0.244 + ], + "angle": 0, + "content": "Figure 4. Cross-group fusion module. In this example, the group number \\( G \\) is 4 and the offset number \\( N \\) of each group is 2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.251, + 0.892, + 0.446 + ], + "angle": 0, + "content": "in most existing NVCs, \\( f_{T\\text{context}} \\) is only the warping operation with the single MV. Such single motion-based alignment is not robust to complex motions or occlusions. The works [9, 54, 58] show deformable alignment gets better results for video restoration as each location has multiple offsets to capture temporal correspondence. So recently it is also applied into NVC [23]. However, the training of deformable alignment is not stable and the overflow of offsets degrades the performance [8, 58]. In addition, the number of offsets is limited to the size of deformable convolutional kernel. Thus, this paper adopts the more flexible design called offset diversity [8]. Meanwhile, the decoded MV is used as the base offset to stabilize the training as [9]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.448, + 0.892, + 0.643 + ], + "angle": 0, + "content": "As shown in Fig. 2, our \\( f_{T\\text{context}} \\) consists of two core sub-modules: offset prediction and cross-group fusion. At first, offset prediction uses the decoded MV \\( \\hat{v}_t \\) to predict the residual offsets \\( d_t \\), where \\( \\hat{x}_{t-1} \\) and \\( F_{t-1} \\) are also warped and fed as the auxiliary information. The \\( d_t \\) adds the base offset \\( \\hat{v}_t \\) to obtain the final offsets \\( o_t \\). At the same time, offset prediction also generates the modulation mask \\( m_t \\) which can be regarded as an attention that reflects the confidence of offsets. It is noted that \\( F_{t-1} \\) is divided into \\( G \\) groups along the channel dimension, and each group has separate \\( N \\) offsets. Thus, there are a total of \\( G \\times N \\) offsets learned. The diverse offsets are complementary to each other, and help codec cope with complex motion and occlusion." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.645, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In addition, motivated by the channel shuffle operation [62] which improves the information flow in the CNN backbone, we customize a group-level interaction mechanism to further tap the potential of offset diversity for NVC. In particular, after warping each group with multiple offsets and applying the corresponding masks, we will reorder all groups before the fusion, as shown in Fig. 4. If using \\( g_{i}^{j} \\) to represent the \\( i \\)-th group warped with its \\( j \\)-th offset, the features before reordering are \\( g_{0}^{0}, \\dots, g_{0}^{N-1}, g_{1}^{0}, \\dots, g_{1}^{N-1}, \\dots, g_{G-1}^{0}, \\dots, g_{G-1}^{N-1} \\), where the offset order is primary and group order is secondary. Then we reorder them as \\( g_{0}^{0}, \\dots, g_{G-1}^{0}, g_{1}^{1}, \\dots, g_{G-1}^{1}, \\dots, g_{G-1}^{N-1}, \\dots, g_{G-1}^{N-1} \\), where the group order is primary instead. The following fusion operation will fuse every \\( N \\) contiguous groups into one group. Therefore, during this process, the group reordering enables more cross-group interactions without increasing" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22619" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.094, + 0.468, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.354, + 0.472, + 0.412 + ], + "angle": 0, + "content": "Figure 5. Entropy coding with quadtree partition. The number means the coding order index. During the 4 coding steps, the \\(\\hat{y}_{t-1}\\) from previous frame, hyper prior \\(\\hat{z}_t\\), and temporal context \\(C_t\\) are also used for entropy modelling." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.421, + 0.47, + 0.513 + ], + "angle": 0, + "content": "complexity. This design also enjoys the similar benefit with the weighted prediction from different reference frames in traditional codec. Via the cross-group fusion, more diverse combinations in extracting temporal contexts from different groups are introduced and further improve the effectiveness of offset diversity." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.526, + 0.441, + 0.543 + ], + "angle": 0, + "content": "3.4. Quadtree Partition-Based Entropy Coding" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.551, + 0.47, + 0.656 + ], + "angle": 0, + "content": "After obtaining temporal context \\( C_t \\) via our offset diversity module, the input frame \\( x_{t} \\) will be encoded and quantized as \\( \\hat{y}_t \\), conditioned on \\( C_t \\), as shown in Fig. 2. We need to estimate the probability mass function (PMF) of \\( \\hat{y}_t \\) for its arithmetic coding. In this process, how to build an accurate entropy model to estimate the PMF of \\( \\hat{y}_t \\) is vital for the compression efficiency." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.658, + 0.47, + 0.839 + ], + "angle": 0, + "content": "Many neural codecs adopt the auto-regressive model [40] as entropy model. However, it seriously slows down the coding speed. By contrast, the checkerboard model [19] proposes coding the even positions of \\(\\hat{y}_t\\) first, and then use them to predict the PMF of the odd positions in parallel. Recently the dual spatial model [29] improves it by utilizing the correlation along channel dimension. However, the neighbours used for entropy modelling in [19, 29] are still limited when compared with auto-regressive model. Thus, inspired from [41, 46], this paper proposes a finer-grained coding manner via the quadtree partition, where diverse spatial contexts are exploited to improve entropy modelling." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.902 + ], + "angle": 0, + "content": "As shown in Fig. 5, we first divide \\(\\hat{y}_t\\) into four groups along the channel dimension. Then each group is partitioned into non-overlapped \\(2\\times 2\\) patches in spatial dimension. The whole entropy coding is divided into four steps," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.243 + ], + "angle": 0, + "content": "and each step codes the different positions associated with the corresponding indexes in Fig. 5. At 0th step, all positions with index 0 of all patches are coded at the same time. It is noted, for the four groups, the positions with index 0 are different from each other. Thus, for every spatial position, there are one fourth channels (i.e., one group) encoded. In the subsequent 1st, 2nd, and 3rd steps, all positions coded in previous steps are used for predicting the PMF of the positions coded in the current step, and different spatial positions are coded for different groups in each step." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.244, + 0.893, + 0.591 + ], + "angle": 0, + "content": "During this process, more diverse neighbours are utilized. If considering the 8 spatial neighbours for a position, the auto-regressive model [40] uses 4 (left, top-left, top, top-right) neighbours for every position if not considering the boundary region. The checkerboard and dual spatial models [19,29] uses 0 and 4 (left, top, right, bottom) neighbours for the 0th and 1st steps, respectively. By contrast, as shown in Fig. 5, our DCVC-DC uses 0, 4, 4, and 8 neighbours for the four steps, respectively. On average, the neighbour number in DCVC-DC is 2 times of that of [19, 29] and is same with that of auto-regressive model. However, our model is much more time-efficient than auto-regressive model as all positions in each step can be coded in parallel. In addition, our model also exploits the cross-channel correlation, which is like [29] but in a refined way. For example, at the 3rd step, for one specific position of a group, the other channels at the same position are already coded from different groups in previous steps, and they can be used as the contexts for the entropy modelling in this step. This helps further squeeze the redundancy. Overall, our quadtree partition-based solution makes the entropy coding benefit from the finer-grained and diverse contexts, which fully mines the correlation from both spatial and channel dimensions." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.605, + 0.66, + 0.621 + ], + "angle": 0, + "content": "3.5. Implementation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Our DCVC-DC is based on DCVC-HEM [29] but focuses on exploiting Diverse Contexts to further boost performance. In addition, we also make the following improvements to obtain better tradeoff between performance and complexity. The first is that, considering depthwise separable convolution [10] can reduce the computation cost while alleviating over-fitting, we widely use it to replace the normal convolution in the basic block design. The second is that we use the unequal channel number settings for features with different resolutions, where the higher resolution feature is assigned with smaller channel number for acceleration. The third is that we move partial quantization operations to higher resolution in the encoder, which helps achieve more precise bit allocation. The harmonization of encoding and quantization also brings some compression ratio improvements. The section 4.3 verifies the effectiveness of these structure optimizations, and the detailed network structures can be found in supplementary materials." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22620" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.091, + 0.28, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.28, + 0.091, + 0.481, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.481, + 0.091, + 0.682, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.682, + 0.091, + 0.882, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.239, + 0.892, + 0.268 + ], + "angle": 0, + "content": "Figure 6. Rate and distortion curves. The comparison is in RGB colorspace measured with PSNR. More results including the corresponding MS-SSIM curves and the comparison in YUV420 colorspace are in supplementary materials." + }, + { + "type": "table_caption", + "bbox": [ + 0.179, + 0.274, + 0.789, + 0.289 + ], + "angle": 0, + "content": "Table 1. BD-Rate (%) comparison in RGB colorspace measured with PSNR. The anchor is VTM-17.0." + }, + { + "type": "table", + "bbox": [ + 0.124, + 0.299, + 0.844, + 0.466 + ], + "angle": 0, + "content": "
UVGMCL-JCVHEVC BHEVC CHEVC DHEVC EHEVC RGBAverage
VTM-17.00.00.00.00.00.00.00.00.0
HM-16.2536.441.538.836.033.744.039.438.5
ECM-5.0-10.0-12.2-11.5-13.4-13.5-10.9-11.1-11.8
CANF-VC [21]73.070.864.476.263.1118.079.977.9
DCVC [28]166.1121.6123.2143.298.0266.1113.4147.4
DCVC-TCM [50]44.151.040.266.337.082.724.449.4
DCVC-HEM [29]1.18.65.122.22.420.5-9.97.1
Our DCVC-DC-19.1-11.3-12.0-10.3-26.1-18.0-27.6-17.8
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.472, + 0.47, + 0.637 + ], + "angle": 0, + "content": "In particular, our DCVC-DC also already outperforms ECM in YUV420 colorspace. Not like traditional codec needs many hand-crafted changes in designing coding tools for different colorspaces, DCVC-DC only just needs simple adoptions based on a existing model trained for RGB. Without changing the network structure, we only up-sample the UV to use the unified input interface with RGB. Correspondingly, after obtaining the reconstructed frame, it is down-sampled on UV. In addition, our model for YUV420 just needs a simple finetune training based on the model trained for RGB." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.655, + 0.283, + 0.672 + ], + "angle": 0, + "content": "4. Experimental Results" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.68, + 0.284, + 0.697 + ], + "angle": 0, + "content": "4.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Datasets. For training, we follow most existing NVCs and use Video-90k [60]. For testing YUV420 videos, HEVC B~E [6], UVG [39], and MCL-JCV [57] are used. Their raw format is YUV420, so there is no any change before feeding them to NVC. For testing RGB videos, as these testsets have no RGB format, most existing NVCs use BT.601 (default in FFmpeg) to convert them from YUV420 to RGB. Actually, JPEG AI [2,3] adopts BT.709 because using BT.709 obtains higher compression ratio under similar visual quality. Thus, this paper follows JPEG AI and uses BT.709 for all codecs during testing RGB. It is noted that the relative bitrate comparisons between different codecs are similar in BT.601 and BT.709. The supplementary ma" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.472, + 0.892, + 0.532 + ], + "angle": 0, + "content": "terials show the results using BT.601. In addition, we follow [29, 50] and also test HEVC RGB dataset [15] when testing RGB videos, and there is no format change as HEVC RGB dataset itself is in RGB format." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.534, + 0.892, + 0.625 + ], + "angle": 0, + "content": "Test Conditions. We follow [29, 50] and test 96 frames for each video, and the intra period is set as 32. The low delay encoding setting is used, as the same with most existing works [1, 28, 34]. BD-Rate [5] is used to measure the compression ratio, where negative numbers indicate bitrate saving and positive numbers indicate bitrate increase." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.627, + 0.892, + 0.793 + ], + "angle": 0, + "content": "Our benchmarks include HM [20] and VTM [55] which represent the best H.265 and H.266 encoder, respectively. In particular, we also compare with ECM [14] which is the prototype of next generation traditional codec. For the codec setting, we follow [29, 50] and further use 10-bit as the intermediate representation when testing RGB, which leads to better compression ratio for the three traditional codecs. The detailed settings are shown in supplementary materials. As for the NVC benchmarks, we compare with the recent SOTA models including CANF-VC [21], DCVC [28], DCVC-TCM [50], and DCVC-HEM [29]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Model Training. We adopt the multi-stage training strategy as [29, 50]. Our model also supports variable bitrate in single model [29], so different \\(\\lambda\\) values are used in different optimization steps. We follow [29] and use \\(4\\lambda\\) values (85, 170, 380, 840). But different from [29] using constant distortion weight in the loss, this paper proposes using hierarchical weight setting on \\(w_{t}\\) for the distortion term (the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.958 + ], + "angle": 0, + "content": "22621" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.168, + 0.089, + 0.8, + 0.104 + ], + "angle": 0, + "content": "Table 2. BD-Rate (%) comparison in RGB colorspace measured with MS-SSIM. The anchor is VTM-17.0." + }, + { + "type": "table", + "bbox": [ + 0.125, + 0.115, + 0.844, + 0.281 + ], + "angle": 0, + "content": "
UVGMCL-JCVHEVC BHEVC CHEVC DHEVC EHEVC RGBAverage
VTM-17.00.00.00.00.00.00.00.00.0
HM-16.2531.138.836.635.233.041.136.636.1
ECM-5.0-9.1-11.1-10.2-11.7-11.0-9.9-9.8-10.4
CANF-VC [21]46.526.043.530.917.9173.057.756.5
DCVC [28]64.927.554.439.715.2210.451.366.2
DCVC-TCM [50]1.0-10.8-11.7-15.2-29.016.7-22.2-10.2
DCVC-HEM [29]-25.2-36.3-38.0-38.3-48.1-25.8-43.6-36.5
Our DCVC-DC-32.6-44.8-47.8-49.8-58.2-45.8-54.4-47.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.168, + 0.292, + 0.8, + 0.307 + ], + "angle": 0, + "content": "Table 3. BD-Rate (%) comparison in YUV420 colorspace measured with PSNR. The anchor is VTM-17.0." + }, + { + "type": "table", + "bbox": [ + 0.177, + 0.318, + 0.79, + 0.414 + ], + "angle": 0, + "content": "
UVGMCL-JCVHEVC BHEVC CHEVC DHEVC EAverage
VTM-17.00.00.00.00.00.00.00.0
HM-16.2536.742.539.233.330.040.737.1
ECM-5.0-10.6-13.7-12.6-14.7-14.9-12.1-13.1
Our DCVC-DC-17.8-12.0-10.8-12.4-28.5-20.4-17.0
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.429, + 0.47, + 0.492 + ], + "angle": 0, + "content": "whole loss is defined in Fig. 2). Considering our training set Vimeo-90k only has 7 frames for each video, we refer traditional codec setting and set the pattern size as 4. The \\( w_{t} \\) settings for 4 consecutive frames are (0.5, 1.2, 0.5, 0.9)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.499, + 0.447, + 0.515 + ], + "angle": 0, + "content": "4.2. Comparisons with Previous SOTA Methods" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.523, + 0.468, + 0.674 + ], + "angle": 0, + "content": "RGB colorspace. Table 1 and 2 show the BR-rate comparison using RGB videos in terms of PSNR and MS-SSIM, respectively. From Table 1, we find our codec achieves significant compression ratio improvement over VTM on every dataset, and there is an average of \\(17.8\\%\\) bitrate saving. By contrast, the other neural CODEs are still worse than VTM. If using DCVC-HEM as anchor, our average bitrate saving is \\(23.5\\%\\). In addition, our DCVC-DC also outperform ECM from Table 1. If using ECM as anchor, an average of \\(6.4\\%\\) bitrate saving is achieved." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.674, + 0.468, + 0.78 + ], + "angle": 0, + "content": "Fig. 6 shows the rate-distortion curves. From the curves, we can see our DCVC-DC achieves the SOTA compression ratio in wide bitrate range. When using MS-SSIM as quality metric, our DCVC-DC shows larger improvement. As shown in Table 2, DCVC-DC has an average of \\(47.6\\%\\) bitrate saving over VTM. By contrast, the corresponding number of ECM over VTM is only \\(10.4\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.468, + 0.87 + ], + "angle": 0, + "content": "It is noted that Table 1 and 2 use the RGB video with BT.709 conversion. If with BT.601 conversion, the relative bitrate saving is similar with that in BT.709. For example, with BT.601 conversion, DCVC-DC over VTM has an average of \\(18.0\\%\\) bitrate saving in terms of PSNR. More results with BT.601 can be seen in supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.468, + 0.901 + ], + "angle": 0, + "content": "YUV420 colorspace. Actually traditional codecs are mainly optimized in YUV420. Thus, the comparison in" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.429, + 0.892, + 0.701 + ], + "angle": 0, + "content": "YUV420 is also very important for evaluating the progress of NVC over traditional codec. The corresponding results are shown in Table 3. The numbers in this table are calculated using the weighted PSNR for the three components of YUV. The weights are \\((6,1,1)/8\\), which are consistent with that in standard committee [53]. As most NVCs have no corresponding released models for YUV420, Table 3 only reports the numbers of our NVC. We can see that DCVC-DC has an average of \\(17.0\\%\\) bitrate saving over VTM. If only considering the Y component, an average \\(15.3\\%\\) bitrate saving is achieved over VTM. Better yet, our DCVC-DC also outperforms ECM in YUV420 on average, as shown in Table 3. This is an important milestone in the development of NVC. It is noted our codec uses the same network structure for both RGB and YUV420 colorspaces, where only different refinunings are used during training. This shows the simplicity and strong extensibility of our codec on the optimization for different input colorspaces." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.711, + 0.653, + 0.727 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.794 + ], + "angle": 0, + "content": "To verify the effectiveness of each component, we conduct comprehensive ablation studies. For simplification, the HEVC datasets in RGB colorspace are used here. The average BD-Rate in terms of PSNR is shown." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Diverse Contexts. Table 4 shows the study on the effectiveness of diverse contexts. First, from the comparison between \\( M_e \\) and \\( M_d \\), we can see that the BD-rate is reduced from \\( 21.3\\% \\) to \\( 14.7\\% \\). This large difference shows the substantial coding gain of our qaudtree partition-based entropy coding, and verifies the advantages of diverse spatial and channel contexts via finer-grained partition." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22622" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.138, + 0.09, + 0.409, + 0.103 + ], + "angle": 0, + "content": "Table 4. Ablation Study on Diverse Contexts." + }, + { + "type": "table", + "bbox": [ + 0.1, + 0.115, + 0.449, + 0.289 + ], + "angle": 0, + "content": "
MaMbMcMdMe
Hierarchical quality structure✓
Offset diversity w/ cross-group✓✓
Offset diversity w/o cross-group [8]✓
Quadtree partition based model✓✓✓✓
Dual spatial model [29]✓
BD-Rate(%)0.08.412.114.721.3
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.295, + 0.47, + 0.476 + ], + "angle": 0, + "content": "From temporal dimension, we also design hierarchical quality structure and offset diversity with cross-group interaction. In Table 4, based on \\( M_{d} \\), we first test the original offset diversity [8] without cross-group interaction, i.e., removing the reorder operation in Fig. 4. However, it \\( (M_{c}) \\) only brings \\( 2.6\\% \\) BD-rate difference. By contrast, powered by our cross-group interaction, the potential of offset diversion is fully tapped, and \\( M_{b} \\) reduces the BD-rate number by \\( 6.3\\% \\) over \\( M_{d} \\). At last, based on \\( M_{b} \\), we evaluate the hierarchical quality structure, i.e., \\( M_{a} \\). The \\( 8.4\\% \\) gap shows learning high-quality contexts brings large benefits to the mining of temporal correlation across many frames." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.477, + 0.469, + 0.627 + ], + "angle": 0, + "content": "Structure optimization. Although our codec learns utilizing diverse contexts in efficient manner, we still purse better tradeoff between compression ratio and computational cost. Therefore, we further optimize our model in network structure. Table 5 shows the study. Based on the \\( M_{a} \\) (same with that in Table 4), we first implement the depthwise separable convolution into codec. \\( M_{h} \\) shows widely using depthwise separable convolution to replace the normal convolution not only significantly reduces the MACs, but also brings some compression ratio improvements." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.628, + 0.469, + 0.839 + ], + "angle": 0, + "content": "The second acceleration is that we use the unequal channel number settings. Not like many existing NVCs use the same channel number for features with different resolutions, we propose assigning the larger number for low resolution feature to increase the latent representation capacity while using the smaller number for the high resolution feature to accelerate model. The performance of \\( M_{g} \\) verifies the effectiveness of our improvement. In addition, many existing NVCs perform the quantization at the low-resolution latent representation after encoding. To achieve more precise rate adjustment, this paper moves partial quantization operations to the higher resolution in the encoder. \\( M_{f} \\) shows the integration of encoding and quantization brings some BD-rate improvements with negligible MAC change." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.848, + 0.203, + 0.864 + ], + "angle": 0, + "content": "4.4. Complexity" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "The complexity comparison is shown in Table 6. We find the MACs of our DCVC-DC are reduced by \\(19.4\\%\\) when" + }, + { + "type": "table_caption", + "bbox": [ + 0.543, + 0.09, + 0.848, + 0.104 + ], + "angle": 0, + "content": "Table 5. Ablation Study on Structure Optimization." + }, + { + "type": "table", + "bbox": [ + 0.508, + 0.115, + 0.892, + 0.228 + ], + "angle": 0, + "content": "
MfMgMhMa
Quant at high resolution✓
Unequal channel setting✓✓
Depthwise separable conv✓✓✓
MACs2642G2642G2939G3456G
BD-Rate (%)0.01.12.43.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.596, + 0.241, + 0.796, + 0.256 + ], + "angle": 0, + "content": "Table 6. Complexity comparison." + }, + { + "type": "table", + "bbox": [ + 0.511, + 0.267, + 0.886, + 0.323 + ], + "angle": 0, + "content": "
MACsEncoding TimeDecoding Time
DCVC-HEM [29]3279G890ms652ms
Our DCVC-DC2642G1005ms765ms
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.505, + 0.327, + 0.825, + 0.34 + ], + "angle": 0, + "content": "Note: Tested on NVIDIA 2080TI with using \\(1080\\mathrm{p}\\) as input." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.347, + 0.892, + 0.527 + ], + "angle": 0, + "content": "compared with DCVC-HEM [29]. However, the actual encoding and decoding time is higher. This is because currently the computational density of depthwise convolution is not as high as normal convolution under the same MAC condition. But through the customized optimization [35], it can be further accelerated in the future. From another perspective, considering that our DCVC-DC has \\(23.5\\%\\) bitrate saving over previous SOTA DCVC-HEM [29], such increase degree in running time is a price worth paying. By contrast, ECM brings \\(13.1\\%\\) (Table 3) improvement over its predecessor VTM, but the encoding complexity is more than 4 times [49] of VTM." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.552, + 0.749, + 0.568 + ], + "angle": 0, + "content": "5. Conclusion and Limitation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.581, + 0.892, + 0.776 + ], + "angle": 0, + "content": "In this paper, we have presented how to utilize diverse contexts to further boost NVC. From temporal dimension, the model is guided to extract long-term and yet high-quality contexts to alleviate error propagation and exploit long range correlation. The offset diversity with cross-group interaction provides complementary motion alignments to handle complex motion. From spatial dimension, the fine-grained quadtree-based partition is proposed to increase spatial context diversity. Powered by our techniques, the compression ratio of NVC has been pushed to new height. Our DCVC-DC has surpassed ECM in both RGB and YUV420 colorspaces, which is an important milestone in the development of NVC." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "During the training, to learn the hierarchical quality pattern, we still use the fixed distortion weights which are similar with those in traditional codec. This may not be the best choice for NVC. Actually, reinforcement learning is good at solving such kind of time series weight decision problem. In the future, we will investigate utilizing reinforcement learning to help NVC make better weight decision with considering the temporal dependency." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22623" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.198 + ], + "angle": 0, + "content": "[1] Eirikur Agustsson, David Minnen, Nick Johnston, Johannes Balle, Sung Jin Hwang, and George Toderici. Scale-space flow for end-to-end optimized video compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8503-8512, 2020. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.201, + 0.472, + 0.242 + ], + "angle": 0, + "content": "[2] E. Alshina, J. Ascenso, T. Ebrahimi, F. Pereira, and T. Richter. [AHG 11] Brief information about JPEG AI CfP status. In JVET-AA0047, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.244, + 0.47, + 0.284 + ], + "angle": 0, + "content": "[3] Anchors · JPEG-AI MMSP Challenge. https://jpegai.github.io/7-anchors/, 2022. Accessed: 2022-11-02. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.287, + 0.47, + 0.342 + ], + "angle": 0, + "content": "[4] Johannes Balle, David Minnen, Saurabh Singh, Sung Jin Hwang, and Nick Johnston. Variational image compression with a scale hyperprior. 6th International Conference on Learning Representations, ICLR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.344, + 0.47, + 0.372 + ], + "angle": 0, + "content": "[5] Gisle Bjontegaard. Calculation of average PSNR differences between RD-curves. VCEG-M33, 2001. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.374, + 0.47, + 0.401 + ], + "angle": 0, + "content": "[6] Frank Bossen et al. Common test conditions and software reference configurations. In JCTVC-L1100, 2013. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.403, + 0.47, + 0.471 + ], + "angle": 0, + "content": "[7] Benjamin Bross, Ye-Kui Wang, Yan Ye, Shan Liu, Jianle Chen, Gary J Sullivan, and Jens-Rainer Ohm. Overview of the versatile video coding (VVC) standard and its applications. IEEE Transactions on Circuits and Systems for Video Technology, 31(10):3736-3764, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.474, + 0.47, + 0.529 + ], + "angle": 0, + "content": "[8] Kelvin CK Chan, Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Understanding deformable alignment in video super-resolution. In Proceedings of the AAAI conference on artificial intelligence, 2021. 2, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.531, + 0.47, + 0.6 + ], + "angle": 0, + "content": "[9] Kelvin CK Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. BasicVSR++: Improving video superresolution with enhanced propagation and alignment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5972-5981, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.602, + 0.47, + 0.657 + ], + "angle": 0, + "content": "[10] François Chollet. Xception: Deep learning with depthwise separable convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1251-1258, 2017. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.659, + 0.47, + 0.728 + ], + "angle": 0, + "content": "[11] Ze Cui, Jing Wang, Shangyin Gao, Tiansheng Guo, Yihui Feng, and Bo Bai. Asymmetric gained deep image compression with continuous rate adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10532-10541, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.73, + 0.47, + 0.77 + ], + "angle": 0, + "content": "[12] Armen Der Kiureghian and Ove Ditlevsen. Aleatory or epistemic? does it matter? Structural safety, 31(2):105-112, 2009. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.773, + 0.47, + 0.841 + ], + "angle": 0, + "content": "[13] Abdelaziz Djelouah, Joaquim Campos, Simone Schaub-Meyer, and Christopher Schroers. Neural inter-frame compression for video coding. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.844, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[14] ECM-5.0. https://vgit.hhi.fraunhofer.de/ecm/ECM, 2022. Accessed: 2022-11-02. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[15] D Flynn, K Sharman, and C Rosewarne. Common test conditions and software reference configurations for hevc range" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.116, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "extensions, document jctvc-n1006. Joint Collaborative Team Video Coding ITU-T SG, 16, 2013. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.15 + ], + "angle": 0, + "content": "[16] Yarin Gal. Uncertainty in Deep Learning. PhD thesis, University of Cambridge, 2016. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.151, + 0.892, + 0.206 + ], + "angle": 0, + "content": "[17] Bernd Girod, Eckehard G Steinbach, and Niko Faerber. Comparison of the H. 263 and H. 261 video compression standards. In Standards and Common Interfaces for Video Information Systems: A Critical Review, 1995. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.208, + 0.892, + 0.262 + ], + "angle": 0, + "content": "[18] Zongyu Guo, Zhizheng Zhang, Runsen Feng, and Zhibo Chen. Causal contextual prediction for learned image compression. IEEE Transactions on Circuits and Systems for Video Technology, 32(4):2329-2341, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.264, + 0.892, + 0.333 + ], + "angle": 0, + "content": "[19] Dailan He, Yaoyan Zheng, Baocheng Sun, Yan Wang, and Hongwei Qin. Checkerboard context model for efficient learned image compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14771-14780, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.335, + 0.892, + 0.362 + ], + "angle": 0, + "content": "[20] HM-16.25. https://vgit.hhi.fraunhofer.de/jvet/HM/, 2022. Accessed: 2022-11-02. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.364, + 0.892, + 0.419 + ], + "angle": 0, + "content": "[21] Yung-Han Ho, Chih-Peng Chang, Peng-Yu Chen, Alessandro Gnutti, and Wen-Hsiao Peng. Canf-vc: Conditional augmented normalizing flows for video compression. European Conference on Computer Vision, 2022. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.421, + 0.892, + 0.489 + ], + "angle": 0, + "content": "[22] Zhihao Hu, Guo Lu, Jinyang Guo, Shan Liu, Wei Jiang, and Dong Xu. Coarse-to-fine deep video coding with hyperprior-guided mode prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5921-5930, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.491, + 0.892, + 0.545 + ], + "angle": 0, + "content": "[23] Zhihao Hu, Guo Lu, and Dong Xu. FVC: A new framework towards deep video compression in feature space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1502-1511, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.547, + 0.892, + 0.589 + ], + "angle": 0, + "content": "[24] Yan Huang, Li Song, and Ebroul Izquierdo. CNN accelerated intra video coding, where is the upper bound? In 2019 Picture Coding Symposium (PCS), pages 1-5. IEEE, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.591, + 0.892, + 0.631 + ], + "angle": 0, + "content": "[25] Shuai Huo, Dong Liu, Li Li, Siwei Ma, Feng Wu, and Wen Gao. Towards hybrid-optimization video coding. arXiv preprint arXiv:2207.05565, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.633, + 0.892, + 0.701 + ], + "angle": 0, + "content": "[26] Jun-Hyuk Kim, Byeongho Heo, and Jong-Seok Lee. Joint global and local hierarchical priors for learned image compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5992-6001, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.703, + 0.892, + 0.771 + ], + "angle": 0, + "content": "[27] Théo Ladune, Pierick Philippe, Wassim Hamidouche, Lu Zhang, and Olivier Déforges. Conditional coding for flexible learned video compression. In Neural Compression: From Information Theory to Applications - Workshop @ ICLR, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.774, + 0.892, + 0.814 + ], + "angle": 0, + "content": "[28] Jiahao Li, Bin Li, and Yan Lu. Deep contextual video compression. Advances in Neural Information Processing Systems, 34, 2021. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.817, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[29] Jiahao Li, Bin Li, and Yan Lu. Hybrid spatial-temporal entropy modelling for neural video compression. In Proceedings of the 30th ACM International Conference on Multimedia, pages 1503-1511, 2022. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[30] Chong Soon Lim, SMT Naing, V Wahadaniah, and X Jing. Reference lists for B pictures under low delay constraints." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "22624" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.469, + 0.133 + ], + "angle": 0, + "content": "document JCTVC-D093, ITU-T/ISO/IEC Joint Collaborative Team on Video Coding (JCT-VC), Daegu, Korea, 2011. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.134, + 0.47, + 0.19 + ], + "angle": 0, + "content": "[31] Jianping Lin, Dong Liu, Houqiang Li, and Feng Wu. M-LVC: multiple frames prediction for learned video compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.19, + 0.471, + 0.259 + ], + "angle": 0, + "content": "[32] Haojie Liu, Ming Lu, Zhan Ma, Fan Wang, Zhihuang Xie, Xun Cao, and Yao Wang. Neural video coding using multiscale motion compensation and spatiotemporal context model. IEEE Transactions on Circuits and Systems for Video Technology, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.26, + 0.471, + 0.328 + ], + "angle": 0, + "content": "[33] Jerry Liu, Shenlong Wang, Wei-Chiu Ma, Meet Shah, Rui Hu, Pranaab Dhawan, and Raquel Urtasun. Conditional entropy coding for efficient video compression. In European Conference on Computer Vision, pages 453-468. Springer, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.33, + 0.471, + 0.398 + ], + "angle": 0, + "content": "[34] Guo Lu, Wanli Ouyang, Dong Xu, Xiaoyun Zhang, Chunlei Cai, and Zhiyong Gao. DVC: an end-to-end deep video compression framework. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11006-11015, 2019. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.4, + 0.471, + 0.453 + ], + "angle": 0, + "content": "[35] Gangzhao Lu, Weizhe Zhang, and Zheng Wang. Optimizing depthwise separable convolution operations on gpus. IEEE Transactions on Parallel and Distributed Systems, 33(1):70-87, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.455, + 0.471, + 0.51 + ], + "angle": 0, + "content": "[36] Guo Lu, Xiaoyun Zhang, Wanli Ouyang, Li Chen, Zhiyong Gao, and Dong Xu. An end-to-end learning framework for video compression. IEEE transactions on pattern analysis and machine intelligence, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.511, + 0.471, + 0.538 + ], + "angle": 0, + "content": "[37] Wufei Ma, Jiahao Li, Bin Li, and Yan Lu. Uncertainty-aware deep video compression with ensembles, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.539, + 0.471, + 0.594 + ], + "angle": 0, + "content": "[38] Fabian Mentzer, George Toderici, David Minnen, Sung-Jin Hwang, Sergi Caelles, Mario Lucic, and Eirikur Agustsson. Vct: A video compression transformer. arXiv preprint arXiv:2206.07307, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.595, + 0.471, + 0.65 + ], + "angle": 0, + "content": "[39] Alexandre Mercat, Marko Viitanen, and Jarno Vanne. UVG dataset: 50/120fps 4k sequences for video codec analysis and development. In Proceedings of the 11th ACM Multimedia Systems Conference, pages 297-302, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.651, + 0.471, + 0.705 + ], + "angle": 0, + "content": "[40] David Minnen, Johannes Ballé, and George D Toderici. Joint autoregressive and hierarchical priors for learned image compression. Advances in neural information processing systems, 31, 2018. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.706, + 0.471, + 0.788 + ], + "angle": 0, + "content": "[41] Ken M Nakanishi, Shin-ichi Maeda, Takeru Miyato, and Daisuke Okanohara. Neural multi-scale image compression. In Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2–6, 2018, Revised Selected Papers, Part VI 14, pages 718–732. Springer, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.471, + 0.872 + ], + "angle": 0, + "content": "[42] Jonathan Pfaff, Alexey Filippov, Shan Liu, Xin Zhao, Jianle Chen, Santiago De-Luxan-Hernandez, Thomas Wiegand, Vasily Rufitskiy, Adarsh Krishnan Ramasubramonian, and Geert Van der Auwera. Intra prediction and mode coding in VLC. IEEE Transactions on Circuits and Systems for Video Technology, 31(10):3834-3847, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.471, + 0.902 + ], + "angle": 0, + "content": "[43] Reza Pourreza, Hoang Le, Amir Said, Guillaume Sautiere, and Auke Wiggers. Boosting neural video codec" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "by exploiting hierarchical redundancy. arXiv preprint arXiv:2208.04303, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[44] Yichen Qian, Ming Lin, Xiuyu Sun, Zhiyu Tan, and Rong Jin. Entroformer: A transformer-based entropy model for learned image compression. arXiv preprint arXiv:2202.05492, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.234 + ], + "angle": 0, + "content": "[45] Yichen Qian, Zhiyu Tan, Xiuyu Sun, Ming Lin, Dongyang Li, Zhenhong Sun, Hao Li, and Rong Jin. Learning accurate entropy model with global reference for image compression. arXiv preprint arXiv:2010.08321, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.235, + 0.892, + 0.305 + ], + "angle": 0, + "content": "[46] Scott Reed, Aaron Oord, Nal Kalchbrenner, Sergio Gomez Colmenarejo, Ziyu Wang, Yutian Chen, Dan Belov, and Nando Freitas. Parallel multiscale autoregressive density estimation. In International Conference on Machine Learning, pages 2912-2921. PMLR, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.306, + 0.892, + 0.375 + ], + "angle": 0, + "content": "[47] Oren Rippel, Alexander G Anderson, Kedar Tatwawadi, Sanjay Nair, Craig Lytle, and Lubomir Bourdev. ELF-VC: Efficient learned flexible-rate video coding. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14479-14488, October 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.377, + 0.892, + 0.431 + ], + "angle": 0, + "content": "[48] Heiko Schwarz, Detlev Marpe, and Thomas Wiegand. Analysis of hierarchical B pictures and MCTF. In 2006 IEEE International Conference on Multimedia and Expo, pages 1929-1932. IEEE, 2006. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.433, + 0.892, + 0.473 + ], + "angle": 0, + "content": "[49] Vadim Seregin, Jie Chen, Fabrice Leannec, and Kai Zhang. JVET AHG report: ECM software development (AHG6). In JVET-AA0006, 2022. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.476, + 0.892, + 0.517 + ], + "angle": 0, + "content": "[50] Xihua Sheng, Jiahao Li, Bin Li, Li Li, Dong Liu, and Yan Lu. Temporal Context Mining for Learned Video Compression. IEEE Transactions on Multimedia, 2022. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.519, + 0.892, + 0.574 + ], + "angle": 0, + "content": "[51] Hui Su, Mingliang Chen, Alexander Bokov, Debargha Mukherjee, Yunqing Wang, and Yue Chen. Machine learning accelerated transform search for av1. In 2019 Picture Coding Symposium (PCS), pages 1-5. IEEE, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.576, + 0.892, + 0.631 + ], + "angle": 0, + "content": "[52] Hui Su, Chi-Yo Tsai, Yunqing Wang, and Yaowu Xu. Machine learning accelerated partition search for video encoding. In 2019 IEEE International Conference on Image Processing (ICIP), pages 2661-2665. IEEE, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.632, + 0.892, + 0.687 + ], + "angle": 0, + "content": "[53] Gary J Sullivan and Jens-Rainer Ohm. Meeting report of the fourth meeting of the joint collaborative team on video coding (jct-vc), daegu, kr, 20-28 january 2011. Document JCTVC-D500, Daegu, KR, 2011. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.689, + 0.892, + 0.757 + ], + "angle": 0, + "content": "[54] Yapeng Tian, Yulun Zhang, Yun Fu, and Chenliang Xu. TDAN: Temporally-deformable alignment network for video super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3360-3369, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.759, + 0.892, + 0.8 + ], + "angle": 0, + "content": "[55] VTM-17.0. https://vcert.hhi.fraunhofer.de/jvet/VVCSoftware_VTM/, 2022. Accessed: 2022-11-02.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.802, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[56] Guo-Hua Wang, Jiahao Li, Bin Li, and Yan Lu. EVC: Towards Real-Time Neural Image Compression with Mask Decay. In International Conference on Learning Representations, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[57] Haiqiang Wang, Weihao Gan, Sudeng Hu, Joe Yuchieh Lin, Lina Jin, Longguang Song, Ping Wang, Ioannis Katsavounidis, Anne Aaron, and C-C Jay Kuo. MCL-JCV: a JND-based" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22625" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "H. 264/AVC video quality assessment dataset. In 2016 IEEE International Conference on Image Processing (ICIP), pages 1509-1513. IEEE, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.205 + ], + "angle": 0, + "content": "[58] Xintao Wang, Kelvin CK Chan, Ke Yu, Chao Dong, and Chen Change Loy. EDVR: Video restoration with enhanced deformable convolutional networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 0–0, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.47, + 0.261 + ], + "angle": 0, + "content": "[59] Chao-Yuan Wu, Nayan Singhal, and Philipp Krahenbuhl. Video compression through image interpolation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 416-431, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.263, + 0.47, + 0.317 + ], + "angle": 0, + "content": "[60] Tianfan Xue, Baian Chen, Jiajun Wu, Donglai Wei, and William T Freeman. Video enhancement with task-oriented flow. International Journal of Computer Vision (IJCV), 127(8):1106-1125, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.32, + 0.47, + 0.387 + ], + "angle": 0, + "content": "[61] Ren Yang, Fabian Mentzer, Luc Van Gool, and Radu Timofte. Learning for video compression with recurrent autoencoder and recurrent probability model. IEEE Journal of Selected Topics in Signal Processing, 15(2):388-401, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.39, + 0.469, + 0.457 + ], + "angle": 0, + "content": "[62] Xiangyu Zhang, Xinyu Zhou, Mengxiao Lin, and Jian Sun. Shufflenet: An extremely efficient convolutional neural network for mobile devices. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6848-6856, 2018. 4" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22626" + } + ] +] \ No newline at end of file diff --git a/2023/Neural Video Compression With Diverse Contexts/b731bba7-3b96-4f19-a574-8047f59049bc_origin.pdf b/2023/Neural Video Compression With Diverse Contexts/b731bba7-3b96-4f19-a574-8047f59049bc_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ad5cfbc81571971716d1e0642b55267158457807 --- /dev/null +++ b/2023/Neural Video Compression With Diverse Contexts/b731bba7-3b96-4f19-a574-8047f59049bc_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bad8d1d6d2fd805d57794612446d8398f064a4731a97717ad76413f903a83587 +size 2026673 diff --git a/2023/Neural Video Compression With Diverse Contexts/full.md b/2023/Neural Video Compression With Diverse Contexts/full.md new file mode 100644 index 0000000000000000000000000000000000000000..140fbc3d79ae18e4f80c027e4daa7c5710e8fffa --- /dev/null +++ b/2023/Neural Video Compression With Diverse Contexts/full.md @@ -0,0 +1,277 @@ +# Neural Video Compression with Diverse Contexts + +Jiahao Li, Bin Li, Yan Lu +Microsoft Research Asia + +{li.jiahao, libin, yanlu}@microsoft.com + +# Abstract + +For any video codecs, the coding efficiency highly relies on whether the current signal to be encoded can find the relevant contexts from the previous reconstructed signals. Traditional codec has verified more contexts bring substantial coding gain, but in a time-consuming manner. However, for the emerging neural video codec (NVC), its contexts are still limited, leading to low compression ratio. To boost NVC, this paper proposes increasing the context diversity in both temporal and spatial dimensions. First, we guide the model to learn hierarchical quality patterns across frames, which enriches long-term and yet high-quality temporal contexts. Furthermore, to tap the potential of optical flow-based coding framework, we introduce a group-based offset diversity where the cross-group interaction is proposed for better context mining. In addition, this paper also adopts a quadtree-based partition to increase spatial context diversity when encoding the latent representation in parallel. Experiments show that our codec obtains $23.5\%$ bitrate saving over previous SOTA NVC. Better yet, our codec has surpassed the under-developing next generation traditional codec/ECM in both RGB and YUV420 colorspaces, in terms of PSNR. The codes are at https://github.com/microsoft/DCVC. + +# 1. Introduction + +The philosophy of video codec is that, for the current signal to be encoded, the codec will find the relevant contexts (e.g., various predictions as the contexts) from previous reconstructed signals to reduce the spatial-temporal redundancy. The more relevant contexts are, the higher bitrate saving is achieved. + +If looking back the development of traditional CODECs (from H.261 [17] in 1988 to H.266 [7] in 2020), we find that the coding gain mainly comes from the continuously expanded coding modes, where each mode uses a specific manner to extract and utilize context. For example, the numbers of intra prediction directions [42] in H.264, H.265, H.266 are 9, 35, and 65, respectively. So many modes can + +![](images/aade64995bb23ac8e4c264999b6e9cf627d3fc0b5899c0996b22bb4e42959693.jpg) +Figure 1. Average results on UVG, MCL-JCV, and HEVC datasets. All traditional codecs use their best compression ratio configuration. DCVC-HEM [29] is the previous SOTA NVC and only has released the model for RGB colorspace. + +extract diverse contexts to reduce redundancy, but also bring huge complexity as rate distortion optimization (RDO) is used to search the best mode. For encoding a 1080p frame, the under-developing ECM (the prototype of next generation traditional codec) needs up to half an hour [49]. Although some DL-based methods [24,51,52] proposed accelerating traditional codecs, the complexity is still very high. + +By contrast, neural video codec (NVC) changes the extraction and utilization of context from hand-crafted design to automatic-learned manner. Mainstream frameworks of NVC can be classified into residual coding-based [1, 13, 31, 32, 34, 36, 47, 59, 61] and condition coding-based [21, 27-29, 33, 38, 50]. The residual coding explicitly uses the predicted frame as the context, and the context utilization is restricted to use subtraction for redundancy removal. By comparison, conditional coding implicitly learns feature domain contexts. The high dimension contexts can carry richer information to facilitate encoding, decoding, as well as entropy modelling. + +However, for most NVCs, the manners of context extraction and utilization are still limited, e.g., only using optical flow to explore temporal correlation. This makes NVC easily suffer from the uncertainty [12, 16, 37] in parameters or fall into local optimum [25]. One solution is adding traditional codec-like coding modes into NVC [25]. But it brings + +large computational complexity as RDO is used. So the question comes up: how to better learn and use the contexts while yielding low computational cost? + +To this end, based on DCVC (deep contextual video compression) [28] framework and its following work DCVC-HEM [29], we propose a new model DCVC-DC which efficiently utilizes the Diverse Contexts to further boost compression ratio. At first, we guide DCVC-DC to learn hierarchical quality pattern across frames. With this guidance during the training, the long-term and yet high-quality contexts which are vital for the reconstruction of the following frames are implicitly learned during the feature propagation. This helps further exploit the long-range temporal correlation in video and effectively alleviate the quality degradation problem existed in most NVCs. In addition, we adopt the offset diversity [8] to strengthen the optical flow-based codec, where multiple offsets can reduce the warping errors for complex or large motions. In particular, inspired by the weighted prediction in traditional codec, the offsets are divided into groups and the cross-group fusion is proposed to improve the temporal context mining. + +Besides from temporal dimension, this paper also proposes increasing the spatial context diversity when encoding the latent representation. Based on recent checkerboard model [19] and dual spatial model [29, 56], we design a quadtree-based partition to improve the distribution estimation. When compared with [19, 29], the types of correlation modelling are more diverse hence the model has a larger chance to find more relevant context. + +It is noted that all our designs are parallel-efficient. To further reduce the computational cost, we also adopt depthwise separable convolution [10], and assign unequal channel numbers for features with different resolutions. Experiments show that our DCVC-DC achieves much higher efficiency over previous SOTA NVC and pushes the compression ratio to a new height. When compared with DCVCHEM [29], $23.5\%$ bitrate saving is achieved while MACs (multiply-accumulate operations) are reduced by $19.4\%$ . Better yet, besides H.266-VTM 17.0, our codec also already outperforms ECM-5.0 (its best compression ratio configuration for low delay coding is used) in both RGB and YUV420 colorspaces, as shown in Fig. 1. To the best of our knowledge, this is the first NVC which can achieve such accomplishment. In summary, our contributions are: + +- We propose efficiently increasing context diversity to boost NVC. Diverse contexts are complementary to each other and have larger chance to provide good reference for reducing redundancy. +- From temporal dimension, we guide model to extract high-quality contexts to alleviate the quality degradation problem. In addition, the group-based offset diversity is designed for better temporal context mining. + +- From spatial dimension, we adopt a quadtree-based partition for latent representation. This provides diverse spatial contexts for better entropy coding. +- Our DCVC-DC obtains $23.5\%$ bitrate saving over the previous SOTA NVC. In particular, our DCVC-DC has surpassed the best traditional codec ECM in both RGB and YUV420 colorspaces, which is an important milestone in the development of NVC. + +# 2. Related Work + +# 2.1. Neural Image Compression + +Most neural image CODECs are based on hyperprior [4] where some bits are first used to provide basic contexts for entropy coding. Then, the auto-regressive prior [40] proposes using neighbour contexts to capture spatial correlation. Recently works [18, 26, 44, 45] propose extracting global or long-range contexts to further boost performance. These show more diverse contexts bring substantial coding gain for neural image codec. + +# 2.2. Neural Video Compression + +Recent years also have witnessed the prosperity of NVC. The pioneering DVC [34] follows traditional codec. It uses optical flow network to generate prediction frame, then its residual with the current frame is coded. Many subsequent works also adopt this residual coding-based framework and refine the modules therein. For example, [31, 43, 47] proposed motion prediction to further reduce redundancy. Optical flow estimation in scale-space [1] was designed to handle complex motion. Yang et al. [61] utilized recurrent autoencoder to improve coding efficiency. + +Residual coding explicitly generates predicted frame in pixel domain as the context and only uses the subtraction to remove redundancy. By comparison, conditional coding has stronger extensibility. The definition, learning, and usage manner of condition can be flexibly designed. In [33, 38], temporally conditional entropy models were designed. [27] uses conditional coding to encode the foreground contents. Li et al. proposed DCVC [28] to learn feature domain contexts to increase context capacity. Then DCVC-TCM [50] adopts feature propagation to boost performance. Recently, DCVC-HEM [29] designs the hybrid entropy model utilizing both spatial and temporal contexts. + +However, the coding modes in most NVCs are still limited when compared with traditional codec. For example, traditional codec adopts translational/affine motion models, geometric partition, bi-prediction, and so on modes to extract diverse temporal contexts [7]. By contrast, existing NVCs usually only relies on single optical flow, which is easily influenced by epistemic uncertainty [12, 16, 37] in model parameters. The recent work [25] also shows such NVCs easily fall into local optimum when coding mode is + +![](images/91c53fd11f964f018075f8bad8cd27199ed184db2cd34e9f8673f2eab7d071e4.jpg) +Figure 2. Framework overview of our DCVC-DC. $x_{t}$ and $\hat{x}_t$ are the input and reconstructed frames. $C_t$ is the learned temporal context as the condition for coding $x_{t}$ . $F_{t}$ is the propagated but unprocessed feature used for next frame. In the loss term, $r_t$ means the bit cost for coding whole frame. $dist(\cdot)$ is distortion function. $\lambda$ and $w_{t}$ are the global and frame-level weights, respectively. The number in the spatial partition for $\hat{y}_t$ represent the coding order index. + +![](images/cfcc72d2798cbf87cfc50b3ccc1ae55c764493d36deacac373d993aff64d33c0.jpg) + +limited. So [25] designed many additional modes like traditional codec, and RDO is used to search the best mode. Such method inevitably brings large computational cost. By comparison, our model has no additional inference cost when exploiting high-quality temporal contexts. Our offset diversity and quadtree partition are also time-efficient designs in providing diverse contexts. + +# 3. Proposed Method + +# 3.1. Overview + +To achieve higher compression ratio, our codec is built on the more flexible conditional coding rather than the residual coding. The framework of our DCVC-DC is illustrated in Fig. 2. It is noted our DCVC-DC is designed for low-delay coding as it can be applied in more scenarios, e.g., real-time communication. As shown in Fig. 2, for coding each frame $x_{t}$ with frame index $t$ , our coding pipeline contains three core steps: $f_{motion}$ , $f_{T\text{context}}$ , and $f_{frame}$ . At first, $f_{motion}$ uses optical flow network to estimate the motion vector (MV) $v_{t}$ , then $v_{t}$ is encoded and decoded as $\hat{v}_{t}$ . Second, based on $\hat{v}_{t}$ and the propagated feature $F_{t-1}$ from the previous frame, $f_{T\text{context}}$ extracts the motion-aligned temporal context feature $C_{t}$ . At last, conditioned on $C_{t}$ , $f_{frame}$ encodes $x_{t}$ into quantized latent representation $\hat{y}_{t}$ . After entropy coding, the output frame $\hat{x}_{t}$ is reconstructed via the decoder and frame generator. At the same time, $F_{t}$ is also generated and propagated to the next frame. It is noted that, our DCVC-DC is based on DCVC-HEM [29]. When compared with DCVC-HEM, this paper redesigns the modules to exploit Diverse Contexts from both temporal (Section 3.2 and 3.3) and spatial (Section 3.4) + +dimensions. + +# 3.2. Hierarchical Quality Structure + +Traditional codec widely adopts hierarchical quality structure, where frames are assigned into different layers and then use different QPs (quantization parameters). This design originates from scalable video coding [48] but also improves the performance for general low delay coding from two aspects. One is periodically improving the quality can alleviate the error propagation, as shown in Fig. 3. During the inter prediction, the high-quality reference frames enable the codec to find the more accurate MV during motion estimation. Meanwhile, the motion compensated prediction is also high-quality, leading to smaller prediction error. Another aspect is that, powered by multiple reference frame selection and weighted prediction mechanisms, the prediction combinations from the nearest reference frame and long-range high-quality reference frame are more diverse. The work [30] investigates many settings on frame quality and reference frame selection, and concludes that the hierarchical quality structure with referencing both the nearest frame and farther high-quality frame achieves the best performance. + +Inspired by the success in traditional codec, we are thinking whether we can equip NVC with the hierarchical quality structure and let NVC also enjoy the benefits. Considering the recent neural CODECs [11,29] also support variable bitrate in single model, one straightforward solution is following the traditional codec and directly assigning hierarchical QPs during the inference of NVC. However, not like traditional codec uses well-defined rules to perform motion estimation and motion compensation (MEMC), NVC uses neural net + +![](images/2ca98f5fce310bd2f63647b2335caef438745cc784471bde5d203dcbf9ccf4ed.jpg) +Figure 3. Hierarchical quality structure in H.266-VTM 17.0 and our NVC. This example is from BasketballPass video from HEVC D dataset. The average bpp (bits per pixel) and PSNR of H.266 are 0.056 and 35.10. Our DCVC-DC is with 0.045 and 36.13. + +works and MEMC is often in feature domain. For NVC, the advantage of such design is that it is automatic learned and has larger potential to achieve better performance. The disadvantage is that it has wore robustness and generalization ability for out-of-distribution quality pattern. Thus, if we directly feed NVC the hierarchical QPs during the testing like [22], it may not well adapt to the hierarchical quality pattern, and the MEMC may get sub-optimal performance. + +To this end, we propose guiding the NVC to learn the hierarchical quality pattern during the training. Specifically, we add a weight $w_{t}$ for each frame in the rate-distortion loss, as shown in Fig. 2. The setting of $w_{t}$ follows the hierarchical structure. Powered by this hierarchical distortion loss, both high-quality output frame $\hat{x}_{t}$ and feature $F_{t}$ containing many high-frequency details are periodically generated. They are very helpful for improving the MEMC effectiveness and then alleviate the error propagation problem that many other NVCs suffer from. In addition, via the cascaded training across multiple frames, the feature propagation chain is formed. The high-quality contexts which are vital for the reconstruction of the following frames are automatically learned and kept in long range. Thus, for the encoding of $x_{t}$ , $F_{t-1}$ not only contains the short-term contexts extracted from $x_{t-1}$ , but also provides long-term and continually-updated high-quality contexts from many previous frames. Such diverse $F_{t-1}$ helps further exploit the temporal correlation across many frames and then boost the compression ratio. Fig. 3 also shows the quality pattern of our NVC. We can see that our DCVC-DC achieves better average quality while with smaller bit cost than H.266. + +# 3.3. Group-Based Offset Diversity + +Due to the various motions between frames, directly using the unprocessed $F_{t - 1}$ without motion alignment is hard for codec to capture temporal correspondence. Therefore, we follow existing NVCs and use optical flow network to extract motion aligned temporal context $C_t$ via $f_{T\text{context}}(F_{t - 1}, \hat{v}_t)$ , where $\hat{v}_t$ is the decoded MV. However, + +![](images/65fae7d00d04e6b4565a477a4e0ee6f4a56aaae079f25fc49c7d60c09841ee9c.jpg) +Figure 4. Cross-group fusion module. In this example, the group number $G$ is 4 and the offset number $N$ of each group is 2. + +in most existing NVCs, $f_{T\text{context}}$ is only the warping operation with the single MV. Such single motion-based alignment is not robust to complex motions or occlusions. The works [9, 54, 58] show deformable alignment gets better results for video restoration as each location has multiple offsets to capture temporal correspondence. So recently it is also applied into NVC [23]. However, the training of deformable alignment is not stable and the overflow of offsets degrades the performance [8, 58]. In addition, the number of offsets is limited to the size of deformable convolutional kernel. Thus, this paper adopts the more flexible design called offset diversity [8]. Meanwhile, the decoded MV is used as the base offset to stabilize the training as [9]. + +As shown in Fig. 2, our $f_{T\text{context}}$ consists of two core sub-modules: offset prediction and cross-group fusion. At first, offset prediction uses the decoded MV $\hat{v}_t$ to predict the residual offsets $d_t$ , where $\hat{x}_{t-1}$ and $F_{t-1}$ are also warped and fed as the auxiliary information. The $d_t$ adds the base offset $\hat{v}_t$ to obtain the final offsets $o_t$ . At the same time, offset prediction also generates the modulation mask $m_t$ which can be regarded as an attention that reflects the confidence of offsets. It is noted that $F_{t-1}$ is divided into $G$ groups along the channel dimension, and each group has separate $N$ offsets. Thus, there are a total of $G \times N$ offsets learned. The diverse offsets are complementary to each other, and help codec cope with complex motion and occlusion. + +In addition, motivated by the channel shuffle operation [62] which improves the information flow in the CNN backbone, we customize a group-level interaction mechanism to further tap the potential of offset diversity for NVC. In particular, after warping each group with multiple offsets and applying the corresponding masks, we will reorder all groups before the fusion, as shown in Fig. 4. If using $g_{i}^{j}$ to represent the $i$ -th group warped with its $j$ -th offset, the features before reordering are $g_{0}^{0}, \dots, g_{0}^{N-1}, g_{1}^{0}, \dots, g_{1}^{N-1}, \dots, g_{G-1}^{0}, \dots, g_{G-1}^{N-1}$ , where the offset order is primary and group order is secondary. Then we reorder them as $g_{0}^{0}, \dots, g_{G-1}^{0}, g_{1}^{1}, \dots, g_{G-1}^{1}, \dots, g_{G-1}^{N-1}, \dots, g_{G-1}^{N-1}$ , where the group order is primary instead. The following fusion operation will fuse every $N$ contiguous groups into one group. Therefore, during this process, the group reordering enables more cross-group interactions without increasing + +![](images/89ce4849cdaa74b4c7c581f0991ef2995c43cbd600ea3b30602aa310d02470c8.jpg) +Figure 5. Entropy coding with quadtree partition. The number means the coding order index. During the 4 coding steps, the $\hat{y}_{t-1}$ from previous frame, hyper prior $\hat{z}_t$ , and temporal context $C_t$ are also used for entropy modelling. + +complexity. This design also enjoys the similar benefit with the weighted prediction from different reference frames in traditional codec. Via the cross-group fusion, more diverse combinations in extracting temporal contexts from different groups are introduced and further improve the effectiveness of offset diversity. + +# 3.4. Quadtree Partition-Based Entropy Coding + +After obtaining temporal context $C_t$ via our offset diversity module, the input frame $x_{t}$ will be encoded and quantized as $\hat{y}_t$ , conditioned on $C_t$ , as shown in Fig. 2. We need to estimate the probability mass function (PMF) of $\hat{y}_t$ for its arithmetic coding. In this process, how to build an accurate entropy model to estimate the PMF of $\hat{y}_t$ is vital for the compression efficiency. + +Many neural codecs adopt the auto-regressive model [40] as entropy model. However, it seriously slows down the coding speed. By contrast, the checkerboard model [19] proposes coding the even positions of $\hat{y}_t$ first, and then use them to predict the PMF of the odd positions in parallel. Recently the dual spatial model [29] improves it by utilizing the correlation along channel dimension. However, the neighbours used for entropy modelling in [19, 29] are still limited when compared with auto-regressive model. Thus, inspired from [41, 46], this paper proposes a finer-grained coding manner via the quadtree partition, where diverse spatial contexts are exploited to improve entropy modelling. + +As shown in Fig. 5, we first divide $\hat{y}_t$ into four groups along the channel dimension. Then each group is partitioned into non-overlapped $2\times 2$ patches in spatial dimension. The whole entropy coding is divided into four steps, + +and each step codes the different positions associated with the corresponding indexes in Fig. 5. At 0th step, all positions with index 0 of all patches are coded at the same time. It is noted, for the four groups, the positions with index 0 are different from each other. Thus, for every spatial position, there are one fourth channels (i.e., one group) encoded. In the subsequent 1st, 2nd, and 3rd steps, all positions coded in previous steps are used for predicting the PMF of the positions coded in the current step, and different spatial positions are coded for different groups in each step. + +During this process, more diverse neighbours are utilized. If considering the 8 spatial neighbours for a position, the auto-regressive model [40] uses 4 (left, top-left, top, top-right) neighbours for every position if not considering the boundary region. The checkerboard and dual spatial models [19,29] uses 0 and 4 (left, top, right, bottom) neighbours for the 0th and 1st steps, respectively. By contrast, as shown in Fig. 5, our DCVC-DC uses 0, 4, 4, and 8 neighbours for the four steps, respectively. On average, the neighbour number in DCVC-DC is 2 times of that of [19, 29] and is same with that of auto-regressive model. However, our model is much more time-efficient than auto-regressive model as all positions in each step can be coded in parallel. In addition, our model also exploits the cross-channel correlation, which is like [29] but in a refined way. For example, at the 3rd step, for one specific position of a group, the other channels at the same position are already coded from different groups in previous steps, and they can be used as the contexts for the entropy modelling in this step. This helps further squeeze the redundancy. Overall, our quadtree partition-based solution makes the entropy coding benefit from the finer-grained and diverse contexts, which fully mines the correlation from both spatial and channel dimensions. + +# 3.5. Implementation + +Our DCVC-DC is based on DCVC-HEM [29] but focuses on exploiting Diverse Contexts to further boost performance. In addition, we also make the following improvements to obtain better tradeoff between performance and complexity. The first is that, considering depthwise separable convolution [10] can reduce the computation cost while alleviating over-fitting, we widely use it to replace the normal convolution in the basic block design. The second is that we use the unequal channel number settings for features with different resolutions, where the higher resolution feature is assigned with smaller channel number for acceleration. The third is that we move partial quantization operations to higher resolution in the encoder, which helps achieve more precise bit allocation. The harmonization of encoding and quantization also brings some compression ratio improvements. The section 4.3 verifies the effectiveness of these structure optimizations, and the detailed network structures can be found in supplementary materials. + +![](images/1e433e77b60fdaff6aa4e5ead59590cab7c9fc64f01f2cea49c3dfe50e4fe467.jpg) +Figure 6. Rate and distortion curves. The comparison is in RGB colorspace measured with PSNR. More results including the corresponding MS-SSIM curves and the comparison in YUV420 colorspace are in supplementary materials. + +![](images/74ec3365b457d92b2328efc384a3722e16d496d36bcfd15bf3abd6ef4be8e531.jpg) + +![](images/8af78d728ffa426c113f3d37e8e7e7b9aaeef9e0a2929803f983246efbbdda01.jpg) + +![](images/b97c9dc01dce74bd66cc11175621edb25b1caae0931db3d9be4eeacdfc0b109d.jpg) + +Table 1. BD-Rate (%) comparison in RGB colorspace measured with PSNR. The anchor is VTM-17.0. + +
UVGMCL-JCVHEVC BHEVC CHEVC DHEVC EHEVC RGBAverage
VTM-17.00.00.00.00.00.00.00.00.0
HM-16.2536.441.538.836.033.744.039.438.5
ECM-5.0-10.0-12.2-11.5-13.4-13.5-10.9-11.1-11.8
CANF-VC [21]73.070.864.476.263.1118.079.977.9
DCVC [28]166.1121.6123.2143.298.0266.1113.4147.4
DCVC-TCM [50]44.151.040.266.337.082.724.449.4
DCVC-HEM [29]1.18.65.122.22.420.5-9.97.1
Our DCVC-DC-19.1-11.3-12.0-10.3-26.1-18.0-27.6-17.8
+ +In particular, our DCVC-DC also already outperforms ECM in YUV420 colorspace. Not like traditional codec needs many hand-crafted changes in designing coding tools for different colorspaces, DCVC-DC only just needs simple adoptions based on a existing model trained for RGB. Without changing the network structure, we only up-sample the UV to use the unified input interface with RGB. Correspondingly, after obtaining the reconstructed frame, it is down-sampled on UV. In addition, our model for YUV420 just needs a simple finetune training based on the model trained for RGB. + +# 4. Experimental Results + +# 4.1. Experimental Settings + +Datasets. For training, we follow most existing NVCs and use Video-90k [60]. For testing YUV420 videos, HEVC B~E [6], UVG [39], and MCL-JCV [57] are used. Their raw format is YUV420, so there is no any change before feeding them to NVC. For testing RGB videos, as these testsets have no RGB format, most existing NVCs use BT.601 (default in FFmpeg) to convert them from YUV420 to RGB. Actually, JPEG AI [2,3] adopts BT.709 because using BT.709 obtains higher compression ratio under similar visual quality. Thus, this paper follows JPEG AI and uses BT.709 for all codecs during testing RGB. It is noted that the relative bitrate comparisons between different codecs are similar in BT.601 and BT.709. The supplementary ma + +terials show the results using BT.601. In addition, we follow [29, 50] and also test HEVC RGB dataset [15] when testing RGB videos, and there is no format change as HEVC RGB dataset itself is in RGB format. + +Test Conditions. We follow [29, 50] and test 96 frames for each video, and the intra period is set as 32. The low delay encoding setting is used, as the same with most existing works [1, 28, 34]. BD-Rate [5] is used to measure the compression ratio, where negative numbers indicate bitrate saving and positive numbers indicate bitrate increase. + +Our benchmarks include HM [20] and VTM [55] which represent the best H.265 and H.266 encoder, respectively. In particular, we also compare with ECM [14] which is the prototype of next generation traditional codec. For the codec setting, we follow [29, 50] and further use 10-bit as the intermediate representation when testing RGB, which leads to better compression ratio for the three traditional codecs. The detailed settings are shown in supplementary materials. As for the NVC benchmarks, we compare with the recent SOTA models including CANF-VC [21], DCVC [28], DCVC-TCM [50], and DCVC-HEM [29]. + +Model Training. We adopt the multi-stage training strategy as [29, 50]. Our model also supports variable bitrate in single model [29], so different $\lambda$ values are used in different optimization steps. We follow [29] and use $4\lambda$ values (85, 170, 380, 840). But different from [29] using constant distortion weight in the loss, this paper proposes using hierarchical weight setting on $w_{t}$ for the distortion term (the + +Table 2. BD-Rate (%) comparison in RGB colorspace measured with MS-SSIM. The anchor is VTM-17.0. + +
UVGMCL-JCVHEVC BHEVC CHEVC DHEVC EHEVC RGBAverage
VTM-17.00.00.00.00.00.00.00.00.0
HM-16.2531.138.836.635.233.041.136.636.1
ECM-5.0-9.1-11.1-10.2-11.7-11.0-9.9-9.8-10.4
CANF-VC [21]46.526.043.530.917.9173.057.756.5
DCVC [28]64.927.554.439.715.2210.451.366.2
DCVC-TCM [50]1.0-10.8-11.7-15.2-29.016.7-22.2-10.2
DCVC-HEM [29]-25.2-36.3-38.0-38.3-48.1-25.8-43.6-36.5
Our DCVC-DC-32.6-44.8-47.8-49.8-58.2-45.8-54.4-47.6
+ +Table 3. BD-Rate (%) comparison in YUV420 colorspace measured with PSNR. The anchor is VTM-17.0. + +
UVGMCL-JCVHEVC BHEVC CHEVC DHEVC EAverage
VTM-17.00.00.00.00.00.00.00.0
HM-16.2536.742.539.233.330.040.737.1
ECM-5.0-10.6-13.7-12.6-14.7-14.9-12.1-13.1
Our DCVC-DC-17.8-12.0-10.8-12.4-28.5-20.4-17.0
+ +whole loss is defined in Fig. 2). Considering our training set Vimeo-90k only has 7 frames for each video, we refer traditional codec setting and set the pattern size as 4. The $w_{t}$ settings for 4 consecutive frames are (0.5, 1.2, 0.5, 0.9). + +# 4.2. Comparisons with Previous SOTA Methods + +RGB colorspace. Table 1 and 2 show the BR-rate comparison using RGB videos in terms of PSNR and MS-SSIM, respectively. From Table 1, we find our codec achieves significant compression ratio improvement over VTM on every dataset, and there is an average of $17.8\%$ bitrate saving. By contrast, the other neural CODEs are still worse than VTM. If using DCVC-HEM as anchor, our average bitrate saving is $23.5\%$ . In addition, our DCVC-DC also outperform ECM from Table 1. If using ECM as anchor, an average of $6.4\%$ bitrate saving is achieved. + +Fig. 6 shows the rate-distortion curves. From the curves, we can see our DCVC-DC achieves the SOTA compression ratio in wide bitrate range. When using MS-SSIM as quality metric, our DCVC-DC shows larger improvement. As shown in Table 2, DCVC-DC has an average of $47.6\%$ bitrate saving over VTM. By contrast, the corresponding number of ECM over VTM is only $10.4\%$ . + +It is noted that Table 1 and 2 use the RGB video with BT.709 conversion. If with BT.601 conversion, the relative bitrate saving is similar with that in BT.709. For example, with BT.601 conversion, DCVC-DC over VTM has an average of $18.0\%$ bitrate saving in terms of PSNR. More results with BT.601 can be seen in supplementary materials. + +YUV420 colorspace. Actually traditional codecs are mainly optimized in YUV420. Thus, the comparison in + +YUV420 is also very important for evaluating the progress of NVC over traditional codec. The corresponding results are shown in Table 3. The numbers in this table are calculated using the weighted PSNR for the three components of YUV. The weights are $(6,1,1)/8$ , which are consistent with that in standard committee [53]. As most NVCs have no corresponding released models for YUV420, Table 3 only reports the numbers of our NVC. We can see that DCVC-DC has an average of $17.0\%$ bitrate saving over VTM. If only considering the Y component, an average $15.3\%$ bitrate saving is achieved over VTM. Better yet, our DCVC-DC also outperforms ECM in YUV420 on average, as shown in Table 3. This is an important milestone in the development of NVC. It is noted our codec uses the same network structure for both RGB and YUV420 colorspaces, where only different refinunings are used during training. This shows the simplicity and strong extensibility of our codec on the optimization for different input colorspaces. + +# 4.3. Ablation Study + +To verify the effectiveness of each component, we conduct comprehensive ablation studies. For simplification, the HEVC datasets in RGB colorspace are used here. The average BD-Rate in terms of PSNR is shown. + +Diverse Contexts. Table 4 shows the study on the effectiveness of diverse contexts. First, from the comparison between $M_e$ and $M_d$ , we can see that the BD-rate is reduced from $21.3\%$ to $14.7\%$ . This large difference shows the substantial coding gain of our qaudtree partition-based entropy coding, and verifies the advantages of diverse spatial and channel contexts via finer-grained partition. + +Table 4. Ablation Study on Diverse Contexts. + +
MaMbMcMdMe
Hierarchical quality structure✓
Offset diversity w/ cross-group✓✓
Offset diversity w/o cross-group [8]✓
Quadtree partition based model✓✓✓✓
Dual spatial model [29]✓
BD-Rate(%)0.08.412.114.721.3
+ +From temporal dimension, we also design hierarchical quality structure and offset diversity with cross-group interaction. In Table 4, based on $M_{d}$ , we first test the original offset diversity [8] without cross-group interaction, i.e., removing the reorder operation in Fig. 4. However, it $(M_{c})$ only brings $2.6\%$ BD-rate difference. By contrast, powered by our cross-group interaction, the potential of offset diversion is fully tapped, and $M_{b}$ reduces the BD-rate number by $6.3\%$ over $M_{d}$ . At last, based on $M_{b}$ , we evaluate the hierarchical quality structure, i.e., $M_{a}$ . The $8.4\%$ gap shows learning high-quality contexts brings large benefits to the mining of temporal correlation across many frames. + +Structure optimization. Although our codec learns utilizing diverse contexts in efficient manner, we still purse better tradeoff between compression ratio and computational cost. Therefore, we further optimize our model in network structure. Table 5 shows the study. Based on the $M_{a}$ (same with that in Table 4), we first implement the depthwise separable convolution into codec. $M_{h}$ shows widely using depthwise separable convolution to replace the normal convolution not only significantly reduces the MACs, but also brings some compression ratio improvements. + +The second acceleration is that we use the unequal channel number settings. Not like many existing NVCs use the same channel number for features with different resolutions, we propose assigning the larger number for low resolution feature to increase the latent representation capacity while using the smaller number for the high resolution feature to accelerate model. The performance of $M_{g}$ verifies the effectiveness of our improvement. In addition, many existing NVCs perform the quantization at the low-resolution latent representation after encoding. To achieve more precise rate adjustment, this paper moves partial quantization operations to the higher resolution in the encoder. $M_{f}$ shows the integration of encoding and quantization brings some BD-rate improvements with negligible MAC change. + +# 4.4. Complexity + +The complexity comparison is shown in Table 6. We find the MACs of our DCVC-DC are reduced by $19.4\%$ when + +Table 5. Ablation Study on Structure Optimization. + +
MfMgMhMa
Quant at high resolution✓
Unequal channel setting✓✓
Depthwise separable conv✓✓✓
MACs2642G2642G2939G3456G
BD-Rate (%)0.01.12.43.5
+ +Table 6. Complexity comparison. + +
MACsEncoding TimeDecoding Time
DCVC-HEM [29]3279G890ms652ms
Our DCVC-DC2642G1005ms765ms
+ +Note: Tested on NVIDIA 2080TI with using $1080\mathrm{p}$ as input. + +compared with DCVC-HEM [29]. However, the actual encoding and decoding time is higher. This is because currently the computational density of depthwise convolution is not as high as normal convolution under the same MAC condition. But through the customized optimization [35], it can be further accelerated in the future. From another perspective, considering that our DCVC-DC has $23.5\%$ bitrate saving over previous SOTA DCVC-HEM [29], such increase degree in running time is a price worth paying. By contrast, ECM brings $13.1\%$ (Table 3) improvement over its predecessor VTM, but the encoding complexity is more than 4 times [49] of VTM. + +# 5. Conclusion and Limitation + +In this paper, we have presented how to utilize diverse contexts to further boost NVC. From temporal dimension, the model is guided to extract long-term and yet high-quality contexts to alleviate error propagation and exploit long range correlation. The offset diversity with cross-group interaction provides complementary motion alignments to handle complex motion. From spatial dimension, the fine-grained quadtree-based partition is proposed to increase spatial context diversity. Powered by our techniques, the compression ratio of NVC has been pushed to new height. Our DCVC-DC has surpassed ECM in both RGB and YUV420 colorspaces, which is an important milestone in the development of NVC. + +During the training, to learn the hierarchical quality pattern, we still use the fixed distortion weights which are similar with those in traditional codec. This may not be the best choice for NVC. Actually, reinforcement learning is good at solving such kind of time series weight decision problem. In the future, we will investigate utilizing reinforcement learning to help NVC make better weight decision with considering the temporal dependency. + +# References + +[1] Eirikur Agustsson, David Minnen, Nick Johnston, Johannes Balle, Sung Jin Hwang, and George Toderici. Scale-space flow for end-to-end optimized video compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8503-8512, 2020. 1, 2, 6 +[2] E. Alshina, J. Ascenso, T. Ebrahimi, F. Pereira, and T. Richter. [AHG 11] Brief information about JPEG AI CfP status. In JVET-AA0047, 2022. 6 +[3] Anchors · JPEG-AI MMSP Challenge. https://jpegai.github.io/7-anchors/, 2022. Accessed: 2022-11-02. 6 +[4] Johannes Balle, David Minnen, Saurabh Singh, Sung Jin Hwang, and Nick Johnston. Variational image compression with a scale hyperprior. 6th International Conference on Learning Representations, ICLR, 2018. 2 +[5] Gisle Bjontegaard. Calculation of average PSNR differences between RD-curves. VCEG-M33, 2001. 6 +[6] Frank Bossen et al. Common test conditions and software reference configurations. In JCTVC-L1100, 2013. 6 +[7] Benjamin Bross, Ye-Kui Wang, Yan Ye, Shan Liu, Jianle Chen, Gary J Sullivan, and Jens-Rainer Ohm. Overview of the versatile video coding (VVC) standard and its applications. IEEE Transactions on Circuits and Systems for Video Technology, 31(10):3736-3764, 2021. 1, 2 +[8] Kelvin CK Chan, Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Understanding deformable alignment in video super-resolution. In Proceedings of the AAAI conference on artificial intelligence, 2021. 2, 4, 8 +[9] Kelvin CK Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. BasicVSR++: Improving video superresolution with enhanced propagation and alignment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5972-5981, 2022. 4 +[10] François Chollet. Xception: Deep learning with depthwise separable convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1251-1258, 2017. 2, 5 +[11] Ze Cui, Jing Wang, Shangyin Gao, Tiansheng Guo, Yihui Feng, and Bo Bai. Asymmetric gained deep image compression with continuous rate adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10532-10541, 2021. 3 +[12] Armen Der Kiureghian and Ove Ditlevsen. Aleatory or epistemic? does it matter? Structural safety, 31(2):105-112, 2009. 1, 2 +[13] Abdelaziz Djelouah, Joaquim Campos, Simone Schaub-Meyer, and Christopher Schroers. Neural inter-frame compression for video coding. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2019. 1 +[14] ECM-5.0. https://vgit.hhi.fraunhofer.de/ecm/ECM, 2022. Accessed: 2022-11-02. 6 +[15] D Flynn, K Sharman, and C Rosewarne. Common test conditions and software reference configurations for hevc range + +extensions, document jctvc-n1006. Joint Collaborative Team Video Coding ITU-T SG, 16, 2013. 6 +[16] Yarin Gal. Uncertainty in Deep Learning. PhD thesis, University of Cambridge, 2016. 1, 2 +[17] Bernd Girod, Eckehard G Steinbach, and Niko Faerber. Comparison of the H. 263 and H. 261 video compression standards. In Standards and Common Interfaces for Video Information Systems: A Critical Review, 1995. 1 +[18] Zongyu Guo, Zhizheng Zhang, Runsen Feng, and Zhibo Chen. Causal contextual prediction for learned image compression. IEEE Transactions on Circuits and Systems for Video Technology, 32(4):2329-2341, 2021. 2 +[19] Dailan He, Yaoyan Zheng, Baocheng Sun, Yan Wang, and Hongwei Qin. Checkerboard context model for efficient learned image compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14771-14780, 2021. 2, 5 +[20] HM-16.25. https://vgit.hhi.fraunhofer.de/jvet/HM/, 2022. Accessed: 2022-11-02. 6 +[21] Yung-Han Ho, Chih-Peng Chang, Peng-Yu Chen, Alessandro Gnutti, and Wen-Hsiao Peng. Canf-vc: Conditional augmented normalizing flows for video compression. European Conference on Computer Vision, 2022. 1, 6, 7 +[22] Zhihao Hu, Guo Lu, Jinyang Guo, Shan Liu, Wei Jiang, and Dong Xu. Coarse-to-fine deep video coding with hyperprior-guided mode prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5921-5930, 2022. 4 +[23] Zhihao Hu, Guo Lu, and Dong Xu. FVC: A new framework towards deep video compression in feature space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1502-1511, 2021. 4 +[24] Yan Huang, Li Song, and Ebroul Izquierdo. CNN accelerated intra video coding, where is the upper bound? In 2019 Picture Coding Symposium (PCS), pages 1-5. IEEE, 2019. 1 +[25] Shuai Huo, Dong Liu, Li Li, Siwei Ma, Feng Wu, and Wen Gao. Towards hybrid-optimization video coding. arXiv preprint arXiv:2207.05565, 2022. 1, 2, 3 +[26] Jun-Hyuk Kim, Byeongho Heo, and Jong-Seok Lee. Joint global and local hierarchical priors for learned image compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5992-6001, 2022. 2 +[27] Théo Ladune, Pierick Philippe, Wassim Hamidouche, Lu Zhang, and Olivier Déforges. Conditional coding for flexible learned video compression. In Neural Compression: From Information Theory to Applications - Workshop @ ICLR, 2021. 1, 2 +[28] Jiahao Li, Bin Li, and Yan Lu. Deep contextual video compression. Advances in Neural Information Processing Systems, 34, 2021. 1, 2, 6, 7 +[29] Jiahao Li, Bin Li, and Yan Lu. Hybrid spatial-temporal entropy modelling for neural video compression. In Proceedings of the 30th ACM International Conference on Multimedia, pages 1503-1511, 2022. 1, 2, 3, 5, 6, 7, 8 +[30] Chong Soon Lim, SMT Naing, V Wahadaniah, and X Jing. Reference lists for B pictures under low delay constraints. + +document JCTVC-D093, ITU-T/ISO/IEC Joint Collaborative Team on Video Coding (JCT-VC), Daegu, Korea, 2011. 3 +[31] Jianping Lin, Dong Liu, Houqiang Li, and Feng Wu. M-LVC: multiple frames prediction for learned video compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 1, 2 +[32] Haojie Liu, Ming Lu, Zhan Ma, Fan Wang, Zhihuang Xie, Xun Cao, and Yao Wang. Neural video coding using multiscale motion compensation and spatiotemporal context model. IEEE Transactions on Circuits and Systems for Video Technology, 2020. 1 +[33] Jerry Liu, Shenlong Wang, Wei-Chiu Ma, Meet Shah, Rui Hu, Pranaab Dhawan, and Raquel Urtasun. Conditional entropy coding for efficient video compression. In European Conference on Computer Vision, pages 453-468. Springer, 2020. 1, 2 +[34] Guo Lu, Wanli Ouyang, Dong Xu, Xiaoyun Zhang, Chunlei Cai, and Zhiyong Gao. DVC: an end-to-end deep video compression framework. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11006-11015, 2019. 1, 2, 6 +[35] Gangzhao Lu, Weizhe Zhang, and Zheng Wang. Optimizing depthwise separable convolution operations on gpus. IEEE Transactions on Parallel and Distributed Systems, 33(1):70-87, 2021. 8 +[36] Guo Lu, Xiaoyun Zhang, Wanli Ouyang, Li Chen, Zhiyong Gao, and Dong Xu. An end-to-end learning framework for video compression. IEEE transactions on pattern analysis and machine intelligence, 2020. 1 +[37] Wufei Ma, Jiahao Li, Bin Li, and Yan Lu. Uncertainty-aware deep video compression with ensembles, 2021. 1, 2 +[38] Fabian Mentzer, George Toderici, David Minnen, Sung-Jin Hwang, Sergi Caelles, Mario Lucic, and Eirikur Agustsson. Vct: A video compression transformer. arXiv preprint arXiv:2206.07307, 2022. 1, 2 +[39] Alexandre Mercat, Marko Viitanen, and Jarno Vanne. UVG dataset: 50/120fps 4k sequences for video codec analysis and development. In Proceedings of the 11th ACM Multimedia Systems Conference, pages 297-302, 2020. 6 +[40] David Minnen, Johannes Ballé, and George D Toderici. Joint autoregressive and hierarchical priors for learned image compression. Advances in neural information processing systems, 31, 2018. 2, 5 +[41] Ken M Nakanishi, Shin-ichi Maeda, Takeru Miyato, and Daisuke Okanohara. Neural multi-scale image compression. In Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2–6, 2018, Revised Selected Papers, Part VI 14, pages 718–732. Springer, 2019. 5 +[42] Jonathan Pfaff, Alexey Filippov, Shan Liu, Xin Zhao, Jianle Chen, Santiago De-Luxan-Hernandez, Thomas Wiegand, Vasily Rufitskiy, Adarsh Krishnan Ramasubramonian, and Geert Van der Auwera. Intra prediction and mode coding in VLC. IEEE Transactions on Circuits and Systems for Video Technology, 31(10):3834-3847, 2021. 1 +[43] Reza Pourreza, Hoang Le, Amir Said, Guillaume Sautiere, and Auke Wiggers. Boosting neural video codec + +by exploiting hierarchical redundancy. arXiv preprint arXiv:2208.04303, 2022. 2 +[44] Yichen Qian, Ming Lin, Xiuyu Sun, Zhiyu Tan, and Rong Jin. Entroformer: A transformer-based entropy model for learned image compression. arXiv preprint arXiv:2202.05492, 2022. 2 +[45] Yichen Qian, Zhiyu Tan, Xiuyu Sun, Ming Lin, Dongyang Li, Zhenhong Sun, Hao Li, and Rong Jin. Learning accurate entropy model with global reference for image compression. arXiv preprint arXiv:2010.08321, 2020. 2 +[46] Scott Reed, Aaron Oord, Nal Kalchbrenner, Sergio Gomez Colmenarejo, Ziyu Wang, Yutian Chen, Dan Belov, and Nando Freitas. Parallel multiscale autoregressive density estimation. In International Conference on Machine Learning, pages 2912-2921. PMLR, 2017. 5 +[47] Oren Rippel, Alexander G Anderson, Kedar Tatwawadi, Sanjay Nair, Craig Lytle, and Lubomir Bourdev. ELF-VC: Efficient learned flexible-rate video coding. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14479-14488, October 2021. 1, 2 +[48] Heiko Schwarz, Detlev Marpe, and Thomas Wiegand. Analysis of hierarchical B pictures and MCTF. In 2006 IEEE International Conference on Multimedia and Expo, pages 1929-1932. IEEE, 2006. 3 +[49] Vadim Seregin, Jie Chen, Fabrice Leannec, and Kai Zhang. JVET AHG report: ECM software development (AHG6). In JVET-AA0006, 2022. 1, 8 +[50] Xihua Sheng, Jiahao Li, Bin Li, Li Li, Dong Liu, and Yan Lu. Temporal Context Mining for Learned Video Compression. IEEE Transactions on Multimedia, 2022. 1, 2, 6, 7 +[51] Hui Su, Mingliang Chen, Alexander Bokov, Debargha Mukherjee, Yunqing Wang, and Yue Chen. Machine learning accelerated transform search for av1. In 2019 Picture Coding Symposium (PCS), pages 1-5. IEEE, 2019. 1 +[52] Hui Su, Chi-Yo Tsai, Yunqing Wang, and Yaowu Xu. Machine learning accelerated partition search for video encoding. In 2019 IEEE International Conference on Image Processing (ICIP), pages 2661-2665. IEEE, 2019. 1 +[53] Gary J Sullivan and Jens-Rainer Ohm. Meeting report of the fourth meeting of the joint collaborative team on video coding (jct-vc), daegu, kr, 20-28 january 2011. Document JCTVC-D500, Daegu, KR, 2011. 7 +[54] Yapeng Tian, Yulun Zhang, Yun Fu, and Chenliang Xu. TDAN: Temporally-deformable alignment network for video super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3360-3369, 2020. 4 +[55] VTM-17.0. https://vcert.hhi.fraunhofer.de/jvet/VVCSoftware_VTM/, 2022. Accessed: 2022-11-02.6 +[56] Guo-Hua Wang, Jiahao Li, Bin Li, and Yan Lu. EVC: Towards Real-Time Neural Image Compression with Mask Decay. In International Conference on Learning Representations, 2023. 2 +[57] Haiqiang Wang, Weihao Gan, Sudeng Hu, Joe Yuchieh Lin, Lina Jin, Longguang Song, Ping Wang, Ioannis Katsavounidis, Anne Aaron, and C-C Jay Kuo. MCL-JCV: a JND-based + +H. 264/AVC video quality assessment dataset. In 2016 IEEE International Conference on Image Processing (ICIP), pages 1509-1513. IEEE, 2016. 6 +[58] Xintao Wang, Kelvin CK Chan, Ke Yu, Chao Dong, and Chen Change Loy. EDVR: Video restoration with enhanced deformable convolutional networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 0–0, 2019. 4 +[59] Chao-Yuan Wu, Nayan Singhal, and Philipp Krahenbuhl. Video compression through image interpolation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 416-431, 2018. 1 +[60] Tianfan Xue, Baian Chen, Jiajun Wu, Donglai Wei, and William T Freeman. Video enhancement with task-oriented flow. International Journal of Computer Vision (IJCV), 127(8):1106-1125, 2019. 6 +[61] Ren Yang, Fabian Mentzer, Luc Van Gool, and Radu Timofte. Learning for video compression with recurrent autoencoder and recurrent probability model. IEEE Journal of Selected Topics in Signal Processing, 15(2):388-401, 2021. 1, 2 +[62] Xiangyu Zhang, Xinyu Zhou, Mengxiao Lin, and Jian Sun. Shufflenet: An extremely efficient convolutional neural network for mobile devices. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6848-6856, 2018. 4 \ No newline at end of file diff --git a/2023/Neural Video Compression With Diverse Contexts/images.zip b/2023/Neural Video Compression With Diverse Contexts/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b76683daab5e693e25067ff3b5af2ffb62e594dc --- /dev/null +++ b/2023/Neural Video Compression With Diverse Contexts/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9078a52607ab8addd1301afa95898d2286b77c34fa8aef7c4df1c5cb30a23343 +size 464430 diff --git a/2023/Neural Video Compression With Diverse Contexts/layout.json b/2023/Neural Video Compression With Diverse Contexts/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d69a84f29bf470cae9be45126f903783fc2bc0cb --- /dev/null +++ b/2023/Neural Video Compression With Diverse Contexts/layout.json @@ -0,0 +1,8821 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 143, + 103, + 451, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 103, + 451, + 121 + ], + "spans": [ + { + "bbox": [ + 143, + 103, + 451, + 121 + ], + "type": "text", + "content": "Neural Video Compression with Diverse Contexts" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 233, + 144, + 360, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 144, + 360, + 171 + ], + "spans": [ + { + "bbox": [ + 233, + 144, + 360, + 171 + ], + "type": "text", + "content": "Jiahao Li, Bin Li, Yan Lu \nMicrosoft Research Asia" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 189, + 173, + 401, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 189, + 173, + 401, + 185 + ], + "spans": [ + { + "bbox": [ + 189, + 173, + 401, + 185 + ], + "type": "text", + "content": "{li.jiahao, libin, yanlu}@microsoft.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 239, + 290, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 239, + 290, + 514 + ], + "spans": [ + { + "bbox": [ + 47, + 239, + 290, + 514 + ], + "type": "text", + "content": "For any video codecs, the coding efficiency highly relies on whether the current signal to be encoded can find the relevant contexts from the previous reconstructed signals. Traditional codec has verified more contexts bring substantial coding gain, but in a time-consuming manner. However, for the emerging neural video codec (NVC), its contexts are still limited, leading to low compression ratio. To boost NVC, this paper proposes increasing the context diversity in both temporal and spatial dimensions. First, we guide the model to learn hierarchical quality patterns across frames, which enriches long-term and yet high-quality temporal contexts. Furthermore, to tap the potential of optical flow-based coding framework, we introduce a group-based offset diversity where the cross-group interaction is proposed for better context mining. In addition, this paper also adopts a quadtree-based partition to increase spatial context diversity when encoding the latent representation in parallel. Experiments show that our codec obtains " + }, + { + "bbox": [ + 47, + 239, + 290, + 514 + ], + "type": "inline_equation", + "content": "23.5\\%" + }, + { + "bbox": [ + 47, + 239, + 290, + 514 + ], + "type": "text", + "content": " bitrate saving over previous SOTA NVC. Better yet, our codec has surpassed the under-developing next generation traditional codec/ECM in both RGB and YUV420 colorspaces, in terms of PSNR. The codes are at https://github.com/microsoft/DCVC." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 536, + 128, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 536, + 128, + 550 + ], + "spans": [ + { + "bbox": [ + 47, + 536, + 128, + 550 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 557, + 287, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 287, + 629 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 287, + 629 + ], + "type": "text", + "content": "The philosophy of video codec is that, for the current signal to be encoded, the codec will find the relevant contexts (e.g., various predictions as the contexts) from previous reconstructed signals to reduce the spatial-temporal redundancy. The more relevant contexts are, the higher bitrate saving is achieved." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 629, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 714 + ], + "type": "text", + "content": "If looking back the development of traditional CODECs (from H.261 [17] in 1988 to H.266 [7] in 2020), we find that the coding gain mainly comes from the continuously expanded coding modes, where each mode uses a specific manner to extract and utilize context. For example, the numbers of intra prediction directions [42] in H.264, H.265, H.266 are 9, 35, and 65, respectively. So many modes can" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 307, + 212, + 545, + 342 + ], + "blocks": [ + { + "bbox": [ + 307, + 212, + 545, + 342 + ], + "lines": [ + { + "bbox": [ + 307, + 212, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 307, + 212, + 545, + 342 + ], + "type": "image", + "image_path": "aade64995bb23ac8e4c264999b6e9cf627d3fc0b5899c0996b22bb4e42959693.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 354, + 547, + 399 + ], + "lines": [ + { + "bbox": [ + 304, + 354, + 547, + 399 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 547, + 399 + ], + "type": "text", + "content": "Figure 1. Average results on UVG, MCL-JCV, and HEVC datasets. All traditional codecs use their best compression ratio configuration. DCVC-HEM [29] is the previous SOTA NVC and only has released the model for RGB colorspace." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 412, + 545, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 412, + 545, + 497 + ], + "spans": [ + { + "bbox": [ + 304, + 412, + 545, + 497 + ], + "type": "text", + "content": "extract diverse contexts to reduce redundancy, but also bring huge complexity as rate distortion optimization (RDO) is used to search the best mode. For encoding a 1080p frame, the under-developing ECM (the prototype of next generation traditional codec) needs up to half an hour [49]. Although some DL-based methods [24,51,52] proposed accelerating traditional codecs, the complexity is still very high." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 497, + 546, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 497, + 546, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 497, + 546, + 640 + ], + "type": "text", + "content": "By contrast, neural video codec (NVC) changes the extraction and utilization of context from hand-crafted design to automatic-learned manner. Mainstream frameworks of NVC can be classified into residual coding-based [1, 13, 31, 32, 34, 36, 47, 59, 61] and condition coding-based [21, 27-29, 33, 38, 50]. The residual coding explicitly uses the predicted frame as the context, and the context utilization is restricted to use subtraction for redundancy removal. By comparison, conditional coding implicitly learns feature domain contexts. The high dimension contexts can carry richer information to facilitate encoding, decoding, as well as entropy modelling." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 642, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 547, + 714 + ], + "type": "text", + "content": "However, for most NVCs, the manners of context extraction and utilization are still limited, e.g., only using optical flow to explore temporal correlation. This makes NVC easily suffer from the uncertainty [12, 16, 37] in parameters or fall into local optimum [25]. One solution is adding traditional codec-like coding modes into NVC [25]. But it brings" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "22616" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": "large computational complexity as RDO is used. So the question comes up: how to better learn and use the contexts while yielding low computational cost?" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 108, + 288, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 108, + 288, + 323 + ], + "spans": [ + { + "bbox": [ + 46, + 108, + 288, + 323 + ], + "type": "text", + "content": "To this end, based on DCVC (deep contextual video compression) [28] framework and its following work DCVC-HEM [29], we propose a new model DCVC-DC which efficiently utilizes the Diverse Contexts to further boost compression ratio. At first, we guide DCVC-DC to learn hierarchical quality pattern across frames. With this guidance during the training, the long-term and yet high-quality contexts which are vital for the reconstruction of the following frames are implicitly learned during the feature propagation. This helps further exploit the long-range temporal correlation in video and effectively alleviate the quality degradation problem existed in most NVCs. In addition, we adopt the offset diversity [8] to strengthen the optical flow-based codec, where multiple offsets can reduce the warping errors for complex or large motions. In particular, inspired by the weighted prediction in traditional codec, the offsets are divided into groups and the cross-group fusion is proposed to improve the temporal context mining." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 324, + 287, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 324, + 287, + 418 + ], + "spans": [ + { + "bbox": [ + 46, + 324, + 287, + 418 + ], + "type": "text", + "content": "Besides from temporal dimension, this paper also proposes increasing the spatial context diversity when encoding the latent representation. Based on recent checkerboard model [19] and dual spatial model [29, 56], we design a quadtree-based partition to improve the distribution estimation. When compared with [19, 29], the types of correlation modelling are more diverse hence the model has a larger chance to find more relevant context." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 419, + 288, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 419, + 288, + 598 + ], + "spans": [ + { + "bbox": [ + 46, + 419, + 288, + 598 + ], + "type": "text", + "content": "It is noted that all our designs are parallel-efficient. To further reduce the computational cost, we also adopt depthwise separable convolution [10], and assign unequal channel numbers for features with different resolutions. Experiments show that our DCVC-DC achieves much higher efficiency over previous SOTA NVC and pushes the compression ratio to a new height. When compared with DCVCHEM [29], " + }, + { + "bbox": [ + 46, + 419, + 288, + 598 + ], + "type": "inline_equation", + "content": "23.5\\%" + }, + { + "bbox": [ + 46, + 419, + 288, + 598 + ], + "type": "text", + "content": " bitrate saving is achieved while MACs (multiply-accumulate operations) are reduced by " + }, + { + "bbox": [ + 46, + 419, + 288, + 598 + ], + "type": "inline_equation", + "content": "19.4\\%" + }, + { + "bbox": [ + 46, + 419, + 288, + 598 + ], + "type": "text", + "content": ". Better yet, besides H.266-VTM 17.0, our codec also already outperforms ECM-5.0 (its best compression ratio configuration for low delay coding is used) in both RGB and YUV420 colorspaces, as shown in Fig. 1. To the best of our knowledge, this is the first NVC which can achieve such accomplishment. In summary, our contributions are:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 609, + 286, + 713 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 59, + 609, + 286, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 609, + 286, + 657 + ], + "spans": [ + { + "bbox": [ + 59, + 609, + 286, + 657 + ], + "type": "text", + "content": "- We propose efficiently increasing context diversity to boost NVC. Diverse contexts are complementary to each other and have larger chance to provide good reference for reducing redundancy." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 665, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 665, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 665, + 286, + 713 + ], + "type": "text", + "content": "- From temporal dimension, we guide model to extract high-quality contexts to alleviate the quality degradation problem. In addition, the group-based offset diversity is designed for better temporal context mining." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 317, + 72, + 545, + 175 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 317, + 72, + 545, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 72, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 545, + 108 + ], + "type": "text", + "content": "- From spatial dimension, we adopt a quadtree-based partition for latent representation. This provides diverse spatial contexts for better entropy coding." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 317, + 115, + 545, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 115, + 545, + 175 + ], + "spans": [ + { + "bbox": [ + 317, + 115, + 545, + 175 + ], + "type": "text", + "content": "- Our DCVC-DC obtains " + }, + { + "bbox": [ + 317, + 115, + 545, + 175 + ], + "type": "inline_equation", + "content": "23.5\\%" + }, + { + "bbox": [ + 317, + 115, + 545, + 175 + ], + "type": "text", + "content": " bitrate saving over the previous SOTA NVC. In particular, our DCVC-DC has surpassed the best traditional codec ECM in both RGB and YUV420 colorspaces, which is an important milestone in the development of NVC." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 184, + 392, + 196 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 184, + 392, + 196 + ], + "spans": [ + { + "bbox": [ + 306, + 184, + 392, + 196 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 204, + 457, + 216 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 204, + 457, + 216 + ], + "spans": [ + { + "bbox": [ + 306, + 204, + 457, + 216 + ], + "type": "text", + "content": "2.1. Neural Image Compression" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 221, + 545, + 318 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 221, + 545, + 318 + ], + "spans": [ + { + "bbox": [ + 304, + 221, + 545, + 318 + ], + "type": "text", + "content": "Most neural image CODECs are based on hyperprior [4] where some bits are first used to provide basic contexts for entropy coding. Then, the auto-regressive prior [40] proposes using neighbour contexts to capture spatial correlation. Recently works [18, 26, 44, 45] propose extracting global or long-range contexts to further boost performance. These show more diverse contexts bring substantial coding gain for neural image codec." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 324, + 454, + 337 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 324, + 454, + 337 + ], + "spans": [ + { + "bbox": [ + 306, + 324, + 454, + 337 + ], + "type": "text", + "content": "2.2. Neural Video Compression" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 342, + 545, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 342, + 545, + 462 + ], + "spans": [ + { + "bbox": [ + 304, + 342, + 545, + 462 + ], + "type": "text", + "content": "Recent years also have witnessed the prosperity of NVC. The pioneering DVC [34] follows traditional codec. It uses optical flow network to generate prediction frame, then its residual with the current frame is coded. Many subsequent works also adopt this residual coding-based framework and refine the modules therein. For example, [31, 43, 47] proposed motion prediction to further reduce redundancy. Optical flow estimation in scale-space [1] was designed to handle complex motion. Yang et al. [61] utilized recurrent autoencoder to improve coding efficiency." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 463, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 463, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 463, + 545, + 605 + ], + "type": "text", + "content": "Residual coding explicitly generates predicted frame in pixel domain as the context and only uses the subtraction to remove redundancy. By comparison, conditional coding has stronger extensibility. The definition, learning, and usage manner of condition can be flexibly designed. In [33, 38], temporally conditional entropy models were designed. [27] uses conditional coding to encode the foreground contents. Li et al. proposed DCVC [28] to learn feature domain contexts to increase context capacity. Then DCVC-TCM [50] adopts feature propagation to boost performance. Recently, DCVC-HEM [29] designs the hybrid entropy model utilizing both spatial and temporal contexts." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 606, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 606, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 606, + 545, + 713 + ], + "type": "text", + "content": "However, the coding modes in most NVCs are still limited when compared with traditional codec. For example, traditional codec adopts translational/affine motion models, geometric partition, bi-prediction, and so on modes to extract diverse temporal contexts [7]. By contrast, existing NVCs usually only relies on single optical flow, which is easily influenced by epistemic uncertainty [12, 16, 37] in model parameters. The recent work [25] also shows such NVCs easily fall into local optimum when coding mode is" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "22617" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 79, + 317, + 264 + ], + "blocks": [ + { + "bbox": [ + 50, + 79, + 317, + 264 + ], + "lines": [ + { + "bbox": [ + 50, + 79, + 317, + 264 + ], + "spans": [ + { + "bbox": [ + 50, + 79, + 317, + 264 + ], + "type": "image", + "image_path": "91c53fd11f964f018075f8bad8cd27199ed184db2cd34e9f8673f2eab7d071e4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 271, + 547, + 316 + ], + "lines": [ + { + "bbox": [ + 46, + 271, + 547, + 316 + ], + "spans": [ + { + "bbox": [ + 46, + 271, + 547, + 316 + ], + "type": "text", + "content": "Figure 2. Framework overview of our DCVC-DC. " + }, + { + "bbox": [ + 46, + 271, + 547, + 316 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 46, + 271, + 547, + 316 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 271, + 547, + 316 + ], + "type": "inline_equation", + "content": "\\hat{x}_t" + }, + { + "bbox": [ + 46, + 271, + 547, + 316 + ], + "type": "text", + "content": " are the input and reconstructed frames. " + }, + { + "bbox": [ + 46, + 271, + 547, + 316 + ], + "type": "inline_equation", + "content": "C_t" + }, + { + "bbox": [ + 46, + 271, + 547, + 316 + ], + "type": "text", + "content": " is the learned temporal context as the condition for coding " + }, + { + "bbox": [ + 46, + 271, + 547, + 316 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 46, + 271, + 547, + 316 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 271, + 547, + 316 + ], + "type": "inline_equation", + "content": "F_{t}" + }, + { + "bbox": [ + 46, + 271, + 547, + 316 + ], + "type": "text", + "content": " is the propagated but unprocessed feature used for next frame. In the loss term, " + }, + { + "bbox": [ + 46, + 271, + 547, + 316 + ], + "type": "inline_equation", + "content": "r_t" + }, + { + "bbox": [ + 46, + 271, + 547, + 316 + ], + "type": "text", + "content": " means the bit cost for coding whole frame. " + }, + { + "bbox": [ + 46, + 271, + 547, + 316 + ], + "type": "inline_equation", + "content": "dist(\\cdot)" + }, + { + "bbox": [ + 46, + 271, + 547, + 316 + ], + "type": "text", + "content": " is distortion function. " + }, + { + "bbox": [ + 46, + 271, + 547, + 316 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 271, + 547, + 316 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 271, + 547, + 316 + ], + "type": "inline_equation", + "content": "w_{t}" + }, + { + "bbox": [ + 46, + 271, + 547, + 316 + ], + "type": "text", + "content": " are the global and frame-level weights, respectively. The number in the spatial partition for " + }, + { + "bbox": [ + 46, + 271, + 547, + 316 + ], + "type": "inline_equation", + "content": "\\hat{y}_t" + }, + { + "bbox": [ + 46, + 271, + 547, + 316 + ], + "type": "text", + "content": " represent the coding order index." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 318, + 79, + 542, + 264 + ], + "blocks": [ + { + "bbox": [ + 318, + 79, + 542, + 264 + ], + "lines": [ + { + "bbox": [ + 318, + 79, + 542, + 264 + ], + "spans": [ + { + "bbox": [ + 318, + 79, + 542, + 264 + ], + "type": "image", + "image_path": "cfcc72d2798cbf87cfc50b3ccc1ae55c764493d36deacac373d993aff64d33c0.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 328, + 288, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 328, + 288, + 413 + ], + "spans": [ + { + "bbox": [ + 46, + 328, + 288, + 413 + ], + "type": "text", + "content": "limited. So [25] designed many additional modes like traditional codec, and RDO is used to search the best mode. Such method inevitably brings large computational cost. By comparison, our model has no additional inference cost when exploiting high-quality temporal contexts. Our offset diversity and quadtree partition are also time-efficient designs in providing diverse contexts." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 424, + 153, + 437 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 424, + 153, + 437 + ], + "spans": [ + { + "bbox": [ + 47, + 424, + 153, + 437 + ], + "type": "text", + "content": "3. Proposed Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 443, + 115, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 443, + 115, + 455 + ], + "spans": [ + { + "bbox": [ + 47, + 443, + 115, + 455 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "type": "text", + "content": "To achieve higher compression ratio, our codec is built on the more flexible conditional coding rather than the residual coding. The framework of our DCVC-DC is illustrated in Fig. 2. It is noted our DCVC-DC is designed for low-delay coding as it can be applied in more scenarios, e.g., real-time communication. As shown in Fig. 2, for coding each frame " + }, + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "type": "text", + "content": " with frame index " + }, + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "type": "text", + "content": ", our coding pipeline contains three core steps: " + }, + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "type": "inline_equation", + "content": "f_{motion}" + }, + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "type": "inline_equation", + "content": "f_{T\\text{context}}" + }, + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "type": "inline_equation", + "content": "f_{frame}" + }, + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "type": "text", + "content": ". At first, " + }, + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "type": "inline_equation", + "content": "f_{motion}" + }, + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "type": "text", + "content": " uses optical flow network to estimate the motion vector (MV) " + }, + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "type": "inline_equation", + "content": "v_{t}" + }, + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "type": "text", + "content": ", then " + }, + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "type": "inline_equation", + "content": "v_{t}" + }, + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "type": "text", + "content": " is encoded and decoded as " + }, + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\hat{v}_{t}" + }, + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "type": "text", + "content": ". Second, based on " + }, + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\hat{v}_{t}" + }, + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "type": "text", + "content": " and the propagated feature " + }, + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "type": "inline_equation", + "content": "F_{t-1}" + }, + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "type": "text", + "content": " from the previous frame, " + }, + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "type": "inline_equation", + "content": "f_{T\\text{context}}" + }, + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "type": "text", + "content": " extracts the motion-aligned temporal context feature " + }, + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "type": "inline_equation", + "content": "C_{t}" + }, + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "type": "text", + "content": ". At last, conditioned on " + }, + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "type": "inline_equation", + "content": "C_{t}" + }, + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "type": "inline_equation", + "content": "f_{frame}" + }, + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "type": "text", + "content": " encodes " + }, + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "type": "text", + "content": " into quantized latent representation " + }, + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\hat{y}_{t}" + }, + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "type": "text", + "content": ". After entropy coding, the output frame " + }, + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\hat{x}_{t}" + }, + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "type": "text", + "content": " is reconstructed via the decoder and frame generator. At the same time, " + }, + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "type": "inline_equation", + "content": "F_{t}" + }, + { + "bbox": [ + 46, + 462, + 288, + 715 + ], + "type": "text", + "content": " is also generated and propagated to the next frame. It is noted that, our DCVC-DC is based on DCVC-HEM [29]. When compared with DCVC-HEM, this paper redesigns the modules to exploit Diverse Contexts from both temporal (Section 3.2 and 3.3) and spatial (Section 3.4)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 329, + 357, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 329, + 357, + 340 + ], + "spans": [ + { + "bbox": [ + 306, + 329, + 357, + 340 + ], + "type": "text", + "content": "dimensions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 348, + 473, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 348, + 473, + 361 + ], + "spans": [ + { + "bbox": [ + 306, + 348, + 473, + 361 + ], + "type": "text", + "content": "3.2. Hierarchical Quality Structure" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 366, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 366, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 366, + 545, + 605 + ], + "type": "text", + "content": "Traditional codec widely adopts hierarchical quality structure, where frames are assigned into different layers and then use different QPs (quantization parameters). This design originates from scalable video coding [48] but also improves the performance for general low delay coding from two aspects. One is periodically improving the quality can alleviate the error propagation, as shown in Fig. 3. During the inter prediction, the high-quality reference frames enable the codec to find the more accurate MV during motion estimation. Meanwhile, the motion compensated prediction is also high-quality, leading to smaller prediction error. Another aspect is that, powered by multiple reference frame selection and weighted prediction mechanisms, the prediction combinations from the nearest reference frame and long-range high-quality reference frame are more diverse. The work [30] investigates many settings on frame quality and reference frame selection, and concludes that the hierarchical quality structure with referencing both the nearest frame and farther high-quality frame achieves the best performance." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "Inspired by the success in traditional codec, we are thinking whether we can equip NVC with the hierarchical quality structure and let NVC also enjoy the benefits. Considering the recent neural CODECs [11,29] also support variable bitrate in single model, one straightforward solution is following the traditional codec and directly assigning hierarchical QPs during the inference of NVC. However, not like traditional codec uses well-defined rules to perform motion estimation and motion compensation (MEMC), NVC uses neural net" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22618" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 76, + 284, + 201 + ], + "blocks": [ + { + "bbox": [ + 50, + 76, + 284, + 201 + ], + "lines": [ + { + "bbox": [ + 50, + 76, + 284, + 201 + ], + "spans": [ + { + "bbox": [ + 50, + 76, + 284, + 201 + ], + "type": "image", + "image_path": "2ca98f5fce310bd2f63647b2335caef438745cc784471bde5d203dcbf9ccf4ed.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 203, + 287, + 247 + ], + "lines": [ + { + "bbox": [ + 46, + 203, + 287, + 247 + ], + "spans": [ + { + "bbox": [ + 46, + 203, + 287, + 247 + ], + "type": "text", + "content": "Figure 3. Hierarchical quality structure in H.266-VTM 17.0 and our NVC. This example is from BasketballPass video from HEVC D dataset. The average bpp (bits per pixel) and PSNR of H.266 are 0.056 and 35.10. Our DCVC-DC is with 0.045 and 36.13." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 253, + 286, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 253, + 286, + 348 + ], + "spans": [ + { + "bbox": [ + 46, + 253, + 286, + 348 + ], + "type": "text", + "content": "works and MEMC is often in feature domain. For NVC, the advantage of such design is that it is automatic learned and has larger potential to achieve better performance. The disadvantage is that it has wore robustness and generalization ability for out-of-distribution quality pattern. Thus, if we directly feed NVC the hierarchical QPs during the testing like [22], it may not well adapt to the hierarchical quality pattern, and the MEMC may get sub-optimal performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 350, + 286, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 350, + 286, + 613 + ], + "spans": [ + { + "bbox": [ + 46, + 350, + 286, + 613 + ], + "type": "text", + "content": "To this end, we propose guiding the NVC to learn the hierarchical quality pattern during the training. Specifically, we add a weight " + }, + { + "bbox": [ + 46, + 350, + 286, + 613 + ], + "type": "inline_equation", + "content": "w_{t}" + }, + { + "bbox": [ + 46, + 350, + 286, + 613 + ], + "type": "text", + "content": " for each frame in the rate-distortion loss, as shown in Fig. 2. The setting of " + }, + { + "bbox": [ + 46, + 350, + 286, + 613 + ], + "type": "inline_equation", + "content": "w_{t}" + }, + { + "bbox": [ + 46, + 350, + 286, + 613 + ], + "type": "text", + "content": " follows the hierarchical structure. Powered by this hierarchical distortion loss, both high-quality output frame " + }, + { + "bbox": [ + 46, + 350, + 286, + 613 + ], + "type": "inline_equation", + "content": "\\hat{x}_{t}" + }, + { + "bbox": [ + 46, + 350, + 286, + 613 + ], + "type": "text", + "content": " and feature " + }, + { + "bbox": [ + 46, + 350, + 286, + 613 + ], + "type": "inline_equation", + "content": "F_{t}" + }, + { + "bbox": [ + 46, + 350, + 286, + 613 + ], + "type": "text", + "content": " containing many high-frequency details are periodically generated. They are very helpful for improving the MEMC effectiveness and then alleviate the error propagation problem that many other NVCs suffer from. In addition, via the cascaded training across multiple frames, the feature propagation chain is formed. The high-quality contexts which are vital for the reconstruction of the following frames are automatically learned and kept in long range. Thus, for the encoding of " + }, + { + "bbox": [ + 46, + 350, + 286, + 613 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 46, + 350, + 286, + 613 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 350, + 286, + 613 + ], + "type": "inline_equation", + "content": "F_{t-1}" + }, + { + "bbox": [ + 46, + 350, + 286, + 613 + ], + "type": "text", + "content": " not only contains the short-term contexts extracted from " + }, + { + "bbox": [ + 46, + 350, + 286, + 613 + ], + "type": "inline_equation", + "content": "x_{t-1}" + }, + { + "bbox": [ + 46, + 350, + 286, + 613 + ], + "type": "text", + "content": ", but also provides long-term and continually-updated high-quality contexts from many previous frames. Such diverse " + }, + { + "bbox": [ + 46, + 350, + 286, + 613 + ], + "type": "inline_equation", + "content": "F_{t-1}" + }, + { + "bbox": [ + 46, + 350, + 286, + 613 + ], + "type": "text", + "content": " helps further exploit the temporal correlation across many frames and then boost the compression ratio. Fig. 3 also shows the quality pattern of our NVC. We can see that our DCVC-DC achieves better average quality while with smaller bit cost than H.266." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 622, + 208, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 622, + 208, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 622, + 208, + 635 + ], + "type": "text", + "content": "3.3. Group-Based Offset Diversity" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 642, + 286, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 286, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 286, + 714 + ], + "type": "text", + "content": "Due to the various motions between frames, directly using the unprocessed " + }, + { + "bbox": [ + 46, + 642, + 286, + 714 + ], + "type": "inline_equation", + "content": "F_{t - 1}" + }, + { + "bbox": [ + 46, + 642, + 286, + 714 + ], + "type": "text", + "content": " without motion alignment is hard for codec to capture temporal correspondence. Therefore, we follow existing NVCs and use optical flow network to extract motion aligned temporal context " + }, + { + "bbox": [ + 46, + 642, + 286, + 714 + ], + "type": "inline_equation", + "content": "C_t" + }, + { + "bbox": [ + 46, + 642, + 286, + 714 + ], + "type": "text", + "content": " via " + }, + { + "bbox": [ + 46, + 642, + 286, + 714 + ], + "type": "inline_equation", + "content": "f_{T\\text{context}}(F_{t - 1}, \\hat{v}_t)" + }, + { + "bbox": [ + 46, + 642, + 286, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 642, + 286, + 714 + ], + "type": "inline_equation", + "content": "\\hat{v}_t" + }, + { + "bbox": [ + 46, + 642, + 286, + 714 + ], + "type": "text", + "content": " is the decoded MV. However," + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 306, + 76, + 545, + 160 + ], + "blocks": [ + { + "bbox": [ + 306, + 76, + 545, + 160 + ], + "lines": [ + { + "bbox": [ + 306, + 76, + 545, + 160 + ], + "spans": [ + { + "bbox": [ + 306, + 76, + 545, + 160 + ], + "type": "image", + "image_path": "65fae7d00d04e6b4565a477a4e0ee6f4a56aaae079f25fc49c7d60c09841ee9c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 171, + 545, + 193 + ], + "lines": [ + { + "bbox": [ + 305, + 171, + 545, + 193 + ], + "spans": [ + { + "bbox": [ + 305, + 171, + 545, + 193 + ], + "type": "text", + "content": "Figure 4. Cross-group fusion module. In this example, the group number " + }, + { + "bbox": [ + 305, + 171, + 545, + 193 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 305, + 171, + 545, + 193 + ], + "type": "text", + "content": " is 4 and the offset number " + }, + { + "bbox": [ + 305, + 171, + 545, + 193 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 305, + 171, + 545, + 193 + ], + "type": "text", + "content": " of each group is 2." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 198, + 545, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 198, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 304, + 198, + 545, + 353 + ], + "type": "text", + "content": "in most existing NVCs, " + }, + { + "bbox": [ + 304, + 198, + 545, + 353 + ], + "type": "inline_equation", + "content": "f_{T\\text{context}}" + }, + { + "bbox": [ + 304, + 198, + 545, + 353 + ], + "type": "text", + "content": " is only the warping operation with the single MV. Such single motion-based alignment is not robust to complex motions or occlusions. The works [9, 54, 58] show deformable alignment gets better results for video restoration as each location has multiple offsets to capture temporal correspondence. So recently it is also applied into NVC [23]. However, the training of deformable alignment is not stable and the overflow of offsets degrades the performance [8, 58]. In addition, the number of offsets is limited to the size of deformable convolutional kernel. Thus, this paper adopts the more flexible design called offset diversity [8]. Meanwhile, the decoded MV is used as the base offset to stabilize the training as [9]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 354, + 545, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 354, + 545, + 509 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 545, + 509 + ], + "type": "text", + "content": "As shown in Fig. 2, our " + }, + { + "bbox": [ + 304, + 354, + 545, + 509 + ], + "type": "inline_equation", + "content": "f_{T\\text{context}}" + }, + { + "bbox": [ + 304, + 354, + 545, + 509 + ], + "type": "text", + "content": " consists of two core sub-modules: offset prediction and cross-group fusion. At first, offset prediction uses the decoded MV " + }, + { + "bbox": [ + 304, + 354, + 545, + 509 + ], + "type": "inline_equation", + "content": "\\hat{v}_t" + }, + { + "bbox": [ + 304, + 354, + 545, + 509 + ], + "type": "text", + "content": " to predict the residual offsets " + }, + { + "bbox": [ + 304, + 354, + 545, + 509 + ], + "type": "inline_equation", + "content": "d_t" + }, + { + "bbox": [ + 304, + 354, + 545, + 509 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 354, + 545, + 509 + ], + "type": "inline_equation", + "content": "\\hat{x}_{t-1}" + }, + { + "bbox": [ + 304, + 354, + 545, + 509 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 354, + 545, + 509 + ], + "type": "inline_equation", + "content": "F_{t-1}" + }, + { + "bbox": [ + 304, + 354, + 545, + 509 + ], + "type": "text", + "content": " are also warped and fed as the auxiliary information. The " + }, + { + "bbox": [ + 304, + 354, + 545, + 509 + ], + "type": "inline_equation", + "content": "d_t" + }, + { + "bbox": [ + 304, + 354, + 545, + 509 + ], + "type": "text", + "content": " adds the base offset " + }, + { + "bbox": [ + 304, + 354, + 545, + 509 + ], + "type": "inline_equation", + "content": "\\hat{v}_t" + }, + { + "bbox": [ + 304, + 354, + 545, + 509 + ], + "type": "text", + "content": " to obtain the final offsets " + }, + { + "bbox": [ + 304, + 354, + 545, + 509 + ], + "type": "inline_equation", + "content": "o_t" + }, + { + "bbox": [ + 304, + 354, + 545, + 509 + ], + "type": "text", + "content": ". At the same time, offset prediction also generates the modulation mask " + }, + { + "bbox": [ + 304, + 354, + 545, + 509 + ], + "type": "inline_equation", + "content": "m_t" + }, + { + "bbox": [ + 304, + 354, + 545, + 509 + ], + "type": "text", + "content": " which can be regarded as an attention that reflects the confidence of offsets. It is noted that " + }, + { + "bbox": [ + 304, + 354, + 545, + 509 + ], + "type": "inline_equation", + "content": "F_{t-1}" + }, + { + "bbox": [ + 304, + 354, + 545, + 509 + ], + "type": "text", + "content": " is divided into " + }, + { + "bbox": [ + 304, + 354, + 545, + 509 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 304, + 354, + 545, + 509 + ], + "type": "text", + "content": " groups along the channel dimension, and each group has separate " + }, + { + "bbox": [ + 304, + 354, + 545, + 509 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 354, + 545, + 509 + ], + "type": "text", + "content": " offsets. Thus, there are a total of " + }, + { + "bbox": [ + 304, + 354, + 545, + 509 + ], + "type": "inline_equation", + "content": "G \\times N" + }, + { + "bbox": [ + 304, + 354, + 545, + 509 + ], + "type": "text", + "content": " offsets learned. The diverse offsets are complementary to each other, and help codec cope with complex motion and occlusion." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": "In addition, motivated by the channel shuffle operation [62] which improves the information flow in the CNN backbone, we customize a group-level interaction mechanism to further tap the potential of offset diversity for NVC. In particular, after warping each group with multiple offsets and applying the corresponding masks, we will reorder all groups before the fusion, as shown in Fig. 4. If using " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "g_{i}^{j}" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": " to represent the " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": "-th group warped with its " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": "-th offset, the features before reordering are " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "g_{0}^{0}, \\dots, g_{0}^{N-1}, g_{1}^{0}, \\dots, g_{1}^{N-1}, \\dots, g_{G-1}^{0}, \\dots, g_{G-1}^{N-1}" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": ", where the offset order is primary and group order is secondary. Then we reorder them as " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "g_{0}^{0}, \\dots, g_{G-1}^{0}, g_{1}^{1}, \\dots, g_{G-1}^{1}, \\dots, g_{G-1}^{N-1}, \\dots, g_{G-1}^{N-1}" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": ", where the group order is primary instead. The following fusion operation will fuse every " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": " contiguous groups into one group. Therefore, during this process, the group reordering enables more cross-group interactions without increasing" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22619" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 74, + 286, + 274 + ], + "blocks": [ + { + "bbox": [ + 50, + 74, + 286, + 274 + ], + "lines": [ + { + "bbox": [ + 50, + 74, + 286, + 274 + ], + "spans": [ + { + "bbox": [ + 50, + 74, + 286, + 274 + ], + "type": "image", + "image_path": "89ce4849cdaa74b4c7c581f0991ef2995c43cbd600ea3b30602aa310d02470c8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 280, + 288, + 326 + ], + "lines": [ + { + "bbox": [ + 46, + 280, + 288, + 326 + ], + "spans": [ + { + "bbox": [ + 46, + 280, + 288, + 326 + ], + "type": "text", + "content": "Figure 5. Entropy coding with quadtree partition. The number means the coding order index. During the 4 coding steps, the " + }, + { + "bbox": [ + 46, + 280, + 288, + 326 + ], + "type": "inline_equation", + "content": "\\hat{y}_{t-1}" + }, + { + "bbox": [ + 46, + 280, + 288, + 326 + ], + "type": "text", + "content": " from previous frame, hyper prior " + }, + { + "bbox": [ + 46, + 280, + 288, + 326 + ], + "type": "inline_equation", + "content": "\\hat{z}_t" + }, + { + "bbox": [ + 46, + 280, + 288, + 326 + ], + "type": "text", + "content": ", and temporal context " + }, + { + "bbox": [ + 46, + 280, + 288, + 326 + ], + "type": "inline_equation", + "content": "C_t" + }, + { + "bbox": [ + 46, + 280, + 288, + 326 + ], + "type": "text", + "content": " are also used for entropy modelling." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 333, + 287, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 333, + 287, + 406 + ], + "spans": [ + { + "bbox": [ + 46, + 333, + 287, + 406 + ], + "type": "text", + "content": "complexity. This design also enjoys the similar benefit with the weighted prediction from different reference frames in traditional codec. Via the cross-group fusion, more diverse combinations in extracting temporal contexts from different groups are introduced and further improve the effectiveness of offset diversity." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 416, + 269, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 416, + 269, + 430 + ], + "spans": [ + { + "bbox": [ + 47, + 416, + 269, + 430 + ], + "type": "text", + "content": "3.4. Quadtree Partition-Based Entropy Coding" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 436, + 287, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 436, + 287, + 519 + ], + "spans": [ + { + "bbox": [ + 46, + 436, + 287, + 519 + ], + "type": "text", + "content": "After obtaining temporal context " + }, + { + "bbox": [ + 46, + 436, + 287, + 519 + ], + "type": "inline_equation", + "content": "C_t" + }, + { + "bbox": [ + 46, + 436, + 287, + 519 + ], + "type": "text", + "content": " via our offset diversity module, the input frame " + }, + { + "bbox": [ + 46, + 436, + 287, + 519 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 46, + 436, + 287, + 519 + ], + "type": "text", + "content": " will be encoded and quantized as " + }, + { + "bbox": [ + 46, + 436, + 287, + 519 + ], + "type": "inline_equation", + "content": "\\hat{y}_t" + }, + { + "bbox": [ + 46, + 436, + 287, + 519 + ], + "type": "text", + "content": ", conditioned on " + }, + { + "bbox": [ + 46, + 436, + 287, + 519 + ], + "type": "inline_equation", + "content": "C_t" + }, + { + "bbox": [ + 46, + 436, + 287, + 519 + ], + "type": "text", + "content": ", as shown in Fig. 2. We need to estimate the probability mass function (PMF) of " + }, + { + "bbox": [ + 46, + 436, + 287, + 519 + ], + "type": "inline_equation", + "content": "\\hat{y}_t" + }, + { + "bbox": [ + 46, + 436, + 287, + 519 + ], + "type": "text", + "content": " for its arithmetic coding. In this process, how to build an accurate entropy model to estimate the PMF of " + }, + { + "bbox": [ + 46, + 436, + 287, + 519 + ], + "type": "inline_equation", + "content": "\\hat{y}_t" + }, + { + "bbox": [ + 46, + 436, + 287, + 519 + ], + "type": "text", + "content": " is vital for the compression efficiency." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 521, + 287, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 287, + 664 + ], + "type": "text", + "content": "Many neural codecs adopt the auto-regressive model [40] as entropy model. However, it seriously slows down the coding speed. By contrast, the checkerboard model [19] proposes coding the even positions of " + }, + { + "bbox": [ + 46, + 521, + 287, + 664 + ], + "type": "inline_equation", + "content": "\\hat{y}_t" + }, + { + "bbox": [ + 46, + 521, + 287, + 664 + ], + "type": "text", + "content": " first, and then use them to predict the PMF of the odd positions in parallel. Recently the dual spatial model [29] improves it by utilizing the correlation along channel dimension. However, the neighbours used for entropy modelling in [19, 29] are still limited when compared with auto-regressive model. Thus, inspired from [41, 46], this paper proposes a finer-grained coding manner via the quadtree partition, where diverse spatial contexts are exploited to improve entropy modelling." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": "As shown in Fig. 5, we first divide " + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\hat{y}_t" + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": " into four groups along the channel dimension. Then each group is partitioned into non-overlapped " + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "inline_equation", + "content": "2\\times 2" + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": " patches in spatial dimension. The whole entropy coding is divided into four steps," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "text", + "content": "and each step codes the different positions associated with the corresponding indexes in Fig. 5. At 0th step, all positions with index 0 of all patches are coded at the same time. It is noted, for the four groups, the positions with index 0 are different from each other. Thus, for every spatial position, there are one fourth channels (i.e., one group) encoded. In the subsequent 1st, 2nd, and 3rd steps, all positions coded in previous steps are used for predicting the PMF of the positions coded in the current step, and different spatial positions are coded for different groups in each step." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 193, + 546, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 193, + 546, + 468 + ], + "spans": [ + { + "bbox": [ + 304, + 193, + 546, + 468 + ], + "type": "text", + "content": "During this process, more diverse neighbours are utilized. If considering the 8 spatial neighbours for a position, the auto-regressive model [40] uses 4 (left, top-left, top, top-right) neighbours for every position if not considering the boundary region. The checkerboard and dual spatial models [19,29] uses 0 and 4 (left, top, right, bottom) neighbours for the 0th and 1st steps, respectively. By contrast, as shown in Fig. 5, our DCVC-DC uses 0, 4, 4, and 8 neighbours for the four steps, respectively. On average, the neighbour number in DCVC-DC is 2 times of that of [19, 29] and is same with that of auto-regressive model. However, our model is much more time-efficient than auto-regressive model as all positions in each step can be coded in parallel. In addition, our model also exploits the cross-channel correlation, which is like [29] but in a refined way. For example, at the 3rd step, for one specific position of a group, the other channels at the same position are already coded from different groups in previous steps, and they can be used as the contexts for the entropy modelling in this step. This helps further squeeze the redundancy. Overall, our quadtree partition-based solution makes the entropy coding benefit from the finer-grained and diverse contexts, which fully mines the correlation from both spatial and channel dimensions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 479, + 403, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 479, + 403, + 491 + ], + "spans": [ + { + "bbox": [ + 306, + 479, + 403, + 491 + ], + "type": "text", + "content": "3.5. Implementation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 498, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 546, + 713 + ], + "type": "text", + "content": "Our DCVC-DC is based on DCVC-HEM [29] but focuses on exploiting Diverse Contexts to further boost performance. In addition, we also make the following improvements to obtain better tradeoff between performance and complexity. The first is that, considering depthwise separable convolution [10] can reduce the computation cost while alleviating over-fitting, we widely use it to replace the normal convolution in the basic block design. The second is that we use the unequal channel number settings for features with different resolutions, where the higher resolution feature is assigned with smaller channel number for acceleration. The third is that we move partial quantization operations to higher resolution in the encoder, which helps achieve more precise bit allocation. The harmonization of encoding and quantization also brings some compression ratio improvements. The section 4.3 verifies the effectiveness of these structure optimizations, and the detailed network structures can be found in supplementary materials." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22620" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 72, + 171, + 182 + ], + "blocks": [ + { + "bbox": [ + 49, + 72, + 171, + 182 + ], + "lines": [ + { + "bbox": [ + 49, + 72, + 171, + 182 + ], + "spans": [ + { + "bbox": [ + 49, + 72, + 171, + 182 + ], + "type": "image", + "image_path": "1e433e77b60fdaff6aa4e5ead59590cab7c9fc64f01f2cea49c3dfe50e4fe467.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 189, + 545, + 212 + ], + "lines": [ + { + "bbox": [ + 46, + 189, + 545, + 212 + ], + "spans": [ + { + "bbox": [ + 46, + 189, + 545, + 212 + ], + "type": "text", + "content": "Figure 6. Rate and distortion curves. The comparison is in RGB colorspace measured with PSNR. More results including the corresponding MS-SSIM curves and the comparison in YUV420 colorspace are in supplementary materials." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 171, + 72, + 294, + 182 + ], + "blocks": [ + { + "bbox": [ + 171, + 72, + 294, + 182 + ], + "lines": [ + { + "bbox": [ + 171, + 72, + 294, + 182 + ], + "spans": [ + { + "bbox": [ + 171, + 72, + 294, + 182 + ], + "type": "image", + "image_path": "74ec3365b457d92b2328efc384a3722e16d496d36bcfd15bf3abd6ef4be8e531.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 294, + 72, + 417, + 182 + ], + "blocks": [ + { + "bbox": [ + 294, + 72, + 417, + 182 + ], + "lines": [ + { + "bbox": [ + 294, + 72, + 417, + 182 + ], + "spans": [ + { + "bbox": [ + 294, + 72, + 417, + 182 + ], + "type": "image", + "image_path": "8af78d728ffa426c113f3d37e8e7e7b9aaeef9e0a2929803f983246efbbdda01.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 417, + 72, + 539, + 182 + ], + "blocks": [ + { + "bbox": [ + 417, + 72, + 539, + 182 + ], + "lines": [ + { + "bbox": [ + 417, + 72, + 539, + 182 + ], + "spans": [ + { + "bbox": [ + 417, + 72, + 539, + 182 + ], + "type": "image", + "image_path": "b97c9dc01dce74bd66cc11175621edb25b1caae0931db3d9be4eeacdfc0b109d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 75, + 236, + 516, + 369 + ], + "blocks": [ + { + "bbox": [ + 109, + 217, + 482, + 228 + ], + "lines": [ + { + "bbox": [ + 109, + 217, + 482, + 228 + ], + "spans": [ + { + "bbox": [ + 109, + 217, + 482, + 228 + ], + "type": "text", + "content": "Table 1. BD-Rate (%) comparison in RGB colorspace measured with PSNR. The anchor is VTM-17.0." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 75, + 236, + 516, + 369 + ], + "lines": [ + { + "bbox": [ + 75, + 236, + 516, + 369 + ], + "spans": [ + { + "bbox": [ + 75, + 236, + 516, + 369 + ], + "type": "table", + "html": "
UVGMCL-JCVHEVC BHEVC CHEVC DHEVC EHEVC RGBAverage
VTM-17.00.00.00.00.00.00.00.00.0
HM-16.2536.441.538.836.033.744.039.438.5
ECM-5.0-10.0-12.2-11.5-13.4-13.5-10.9-11.1-11.8
CANF-VC [21]73.070.864.476.263.1118.079.977.9
DCVC [28]166.1121.6123.2143.298.0266.1113.4147.4
DCVC-TCM [50]44.151.040.266.337.082.724.449.4
DCVC-HEM [29]1.18.65.122.22.420.5-9.97.1
Our DCVC-DC-19.1-11.3-12.0-10.3-26.1-18.0-27.6-17.8
", + "image_path": "7affc2c93a0b44c625d915bb4c5543db790943202d2c38c3d1dec491c37ff7eb.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 373, + 287, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 373, + 287, + 504 + ], + "spans": [ + { + "bbox": [ + 46, + 373, + 287, + 504 + ], + "type": "text", + "content": "In particular, our DCVC-DC also already outperforms ECM in YUV420 colorspace. Not like traditional codec needs many hand-crafted changes in designing coding tools for different colorspaces, DCVC-DC only just needs simple adoptions based on a existing model trained for RGB. Without changing the network structure, we only up-sample the UV to use the unified input interface with RGB. Correspondingly, after obtaining the reconstructed frame, it is down-sampled on UV. In addition, our model for YUV420 just needs a simple finetune training based on the model trained for RGB." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 518, + 173, + 532 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 518, + 173, + 532 + ], + "spans": [ + { + "bbox": [ + 47, + 518, + 173, + 532 + ], + "type": "text", + "content": "4. Experimental Results" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 538, + 173, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 538, + 173, + 552 + ], + "spans": [ + { + "bbox": [ + 47, + 538, + 173, + 552 + ], + "type": "text", + "content": "4.1. Experimental Settings" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 288, + 715 + ], + "type": "text", + "content": "Datasets. For training, we follow most existing NVCs and use Video-90k [60]. For testing YUV420 videos, HEVC B~E [6], UVG [39], and MCL-JCV [57] are used. Their raw format is YUV420, so there is no any change before feeding them to NVC. For testing RGB videos, as these testsets have no RGB format, most existing NVCs use BT.601 (default in FFmpeg) to convert them from YUV420 to RGB. Actually, JPEG AI [2,3] adopts BT.709 because using BT.709 obtains higher compression ratio under similar visual quality. Thus, this paper follows JPEG AI and uses BT.709 for all codecs during testing RGB. It is noted that the relative bitrate comparisons between different codecs are similar in BT.601 and BT.709. The supplementary ma" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 373, + 545, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 373, + 545, + 421 + ], + "spans": [ + { + "bbox": [ + 304, + 373, + 545, + 421 + ], + "type": "text", + "content": "terials show the results using BT.601. In addition, we follow [29, 50] and also test HEVC RGB dataset [15] when testing RGB videos, and there is no format change as HEVC RGB dataset itself is in RGB format." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 422, + 545, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 422, + 545, + 495 + ], + "spans": [ + { + "bbox": [ + 304, + 422, + 545, + 495 + ], + "type": "text", + "content": "Test Conditions. We follow [29, 50] and test 96 frames for each video, and the intra period is set as 32. The low delay encoding setting is used, as the same with most existing works [1, 28, 34]. BD-Rate [5] is used to measure the compression ratio, where negative numbers indicate bitrate saving and positive numbers indicate bitrate increase." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 496, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 496, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 496, + 545, + 628 + ], + "type": "text", + "content": "Our benchmarks include HM [20] and VTM [55] which represent the best H.265 and H.266 encoder, respectively. In particular, we also compare with ECM [14] which is the prototype of next generation traditional codec. For the codec setting, we follow [29, 50] and further use 10-bit as the intermediate representation when testing RGB, which leads to better compression ratio for the three traditional codecs. The detailed settings are shown in supplementary materials. As for the NVC benchmarks, we compare with the recent SOTA models including CANF-VC [21], DCVC [28], DCVC-TCM [50], and DCVC-HEM [29]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "Model Training. We adopt the multi-stage training strategy as [29, 50]. Our model also supports variable bitrate in single model [29], so different " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " values are used in different optimization steps. We follow [29] and use " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "4\\lambda" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " values (85, 170, 380, 840). But different from [29] using constant distortion weight in the loss, this paper proposes using hierarchical weight setting on " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "w_{t}" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " for the distortion term (the" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "22621" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 76, + 91, + 516, + 222 + ], + "blocks": [ + { + "bbox": [ + 102, + 70, + 489, + 82 + ], + "lines": [ + { + "bbox": [ + 102, + 70, + 489, + 82 + ], + "spans": [ + { + "bbox": [ + 102, + 70, + 489, + 82 + ], + "type": "text", + "content": "Table 2. BD-Rate (%) comparison in RGB colorspace measured with MS-SSIM. The anchor is VTM-17.0." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 76, + 91, + 516, + 222 + ], + "lines": [ + { + "bbox": [ + 76, + 91, + 516, + 222 + ], + "spans": [ + { + "bbox": [ + 76, + 91, + 516, + 222 + ], + "type": "table", + "html": "
UVGMCL-JCVHEVC BHEVC CHEVC DHEVC EHEVC RGBAverage
VTM-17.00.00.00.00.00.00.00.00.0
HM-16.2531.138.836.635.233.041.136.636.1
ECM-5.0-9.1-11.1-10.2-11.7-11.0-9.9-9.8-10.4
CANF-VC [21]46.526.043.530.917.9173.057.756.5
DCVC [28]64.927.554.439.715.2210.451.366.2
DCVC-TCM [50]1.0-10.8-11.7-15.2-29.016.7-22.2-10.2
DCVC-HEM [29]-25.2-36.3-38.0-38.3-48.1-25.8-43.6-36.5
Our DCVC-DC-32.6-44.8-47.8-49.8-58.2-45.8-54.4-47.6
", + "image_path": "ecc99482c4f1fad9f00376e153cddd72f99b04166031d6506ccf1b38e5e2dd59.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 108, + 251, + 483, + 327 + ], + "blocks": [ + { + "bbox": [ + 102, + 231, + 489, + 243 + ], + "lines": [ + { + "bbox": [ + 102, + 231, + 489, + 243 + ], + "spans": [ + { + "bbox": [ + 102, + 231, + 489, + 243 + ], + "type": "text", + "content": "Table 3. BD-Rate (%) comparison in YUV420 colorspace measured with PSNR. The anchor is VTM-17.0." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 108, + 251, + 483, + 327 + ], + "lines": [ + { + "bbox": [ + 108, + 251, + 483, + 327 + ], + "spans": [ + { + "bbox": [ + 108, + 251, + 483, + 327 + ], + "type": "table", + "html": "
UVGMCL-JCVHEVC BHEVC CHEVC DHEVC EAverage
VTM-17.00.00.00.00.00.00.00.0
HM-16.2536.742.539.233.330.040.737.1
ECM-5.0-10.6-13.7-12.6-14.7-14.9-12.1-13.1
Our DCVC-DC-17.8-12.0-10.8-12.4-28.5-20.4-17.0
", + "image_path": "725e2695db4ae035b7601b87958faf1968e9f61fc03882c560d92d8d6161cb9f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 339, + 287, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 339, + 287, + 389 + ], + "spans": [ + { + "bbox": [ + 46, + 339, + 287, + 389 + ], + "type": "text", + "content": "whole loss is defined in Fig. 2). Considering our training set Vimeo-90k only has 7 frames for each video, we refer traditional codec setting and set the pattern size as 4. The " + }, + { + "bbox": [ + 46, + 339, + 287, + 389 + ], + "type": "inline_equation", + "content": "w_{t}" + }, + { + "bbox": [ + 46, + 339, + 287, + 389 + ], + "type": "text", + "content": " settings for 4 consecutive frames are (0.5, 1.2, 0.5, 0.9)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 395, + 273, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 395, + 273, + 407 + ], + "spans": [ + { + "bbox": [ + 47, + 395, + 273, + 407 + ], + "type": "text", + "content": "4.2. Comparisons with Previous SOTA Methods" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 414, + 286, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 414, + 286, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 414, + 286, + 533 + ], + "type": "text", + "content": "RGB colorspace. Table 1 and 2 show the BR-rate comparison using RGB videos in terms of PSNR and MS-SSIM, respectively. From Table 1, we find our codec achieves significant compression ratio improvement over VTM on every dataset, and there is an average of " + }, + { + "bbox": [ + 46, + 414, + 286, + 533 + ], + "type": "inline_equation", + "content": "17.8\\%" + }, + { + "bbox": [ + 46, + 414, + 286, + 533 + ], + "type": "text", + "content": " bitrate saving. By contrast, the other neural CODEs are still worse than VTM. If using DCVC-HEM as anchor, our average bitrate saving is " + }, + { + "bbox": [ + 46, + 414, + 286, + 533 + ], + "type": "inline_equation", + "content": "23.5\\%" + }, + { + "bbox": [ + 46, + 414, + 286, + 533 + ], + "type": "text", + "content": ". In addition, our DCVC-DC also outperform ECM from Table 1. If using ECM as anchor, an average of " + }, + { + "bbox": [ + 46, + 414, + 286, + 533 + ], + "type": "inline_equation", + "content": "6.4\\%" + }, + { + "bbox": [ + 46, + 414, + 286, + 533 + ], + "type": "text", + "content": " bitrate saving is achieved." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 533, + 286, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 286, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 286, + 617 + ], + "type": "text", + "content": "Fig. 6 shows the rate-distortion curves. From the curves, we can see our DCVC-DC achieves the SOTA compression ratio in wide bitrate range. When using MS-SSIM as quality metric, our DCVC-DC shows larger improvement. As shown in Table 2, DCVC-DC has an average of " + }, + { + "bbox": [ + 46, + 533, + 286, + 617 + ], + "type": "inline_equation", + "content": "47.6\\%" + }, + { + "bbox": [ + 46, + 533, + 286, + 617 + ], + "type": "text", + "content": " bitrate saving over VTM. By contrast, the corresponding number of ECM over VTM is only " + }, + { + "bbox": [ + 46, + 533, + 286, + 617 + ], + "type": "inline_equation", + "content": "10.4\\%" + }, + { + "bbox": [ + 46, + 533, + 286, + 617 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 617, + 286, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 286, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 286, + 689 + ], + "type": "text", + "content": "It is noted that Table 1 and 2 use the RGB video with BT.709 conversion. If with BT.601 conversion, the relative bitrate saving is similar with that in BT.709. For example, with BT.601 conversion, DCVC-DC over VTM has an average of " + }, + { + "bbox": [ + 46, + 617, + 286, + 689 + ], + "type": "inline_equation", + "content": "18.0\\%" + }, + { + "bbox": [ + 46, + 617, + 286, + 689 + ], + "type": "text", + "content": " bitrate saving in terms of PSNR. More results with BT.601 can be seen in supplementary materials." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "type": "text", + "content": "YUV420 colorspace. Actually traditional codecs are mainly optimized in YUV420. Thus, the comparison in" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 339, + 545, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 339, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 339, + 545, + 555 + ], + "type": "text", + "content": "YUV420 is also very important for evaluating the progress of NVC over traditional codec. The corresponding results are shown in Table 3. The numbers in this table are calculated using the weighted PSNR for the three components of YUV. The weights are " + }, + { + "bbox": [ + 304, + 339, + 545, + 555 + ], + "type": "inline_equation", + "content": "(6,1,1)/8" + }, + { + "bbox": [ + 304, + 339, + 545, + 555 + ], + "type": "text", + "content": ", which are consistent with that in standard committee [53]. As most NVCs have no corresponding released models for YUV420, Table 3 only reports the numbers of our NVC. We can see that DCVC-DC has an average of " + }, + { + "bbox": [ + 304, + 339, + 545, + 555 + ], + "type": "inline_equation", + "content": "17.0\\%" + }, + { + "bbox": [ + 304, + 339, + 545, + 555 + ], + "type": "text", + "content": " bitrate saving over VTM. If only considering the Y component, an average " + }, + { + "bbox": [ + 304, + 339, + 545, + 555 + ], + "type": "inline_equation", + "content": "15.3\\%" + }, + { + "bbox": [ + 304, + 339, + 545, + 555 + ], + "type": "text", + "content": " bitrate saving is achieved over VTM. Better yet, our DCVC-DC also outperforms ECM in YUV420 on average, as shown in Table 3. This is an important milestone in the development of NVC. It is noted our codec uses the same network structure for both RGB and YUV420 colorspaces, where only different refinunings are used during training. This shows the simplicity and strong extensibility of our codec on the optimization for different input colorspaces." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 563, + 399, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 563, + 399, + 575 + ], + "spans": [ + { + "bbox": [ + 306, + 563, + 399, + 575 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 582, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 628 + ], + "type": "text", + "content": "To verify the effectiveness of each component, we conduct comprehensive ablation studies. For simplification, the HEVC datasets in RGB colorspace are used here. The average BD-Rate in terms of PSNR is shown." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "Diverse Contexts. Table 4 shows the study on the effectiveness of diverse contexts. First, from the comparison between " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "M_e" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "M_d" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": ", we can see that the BD-rate is reduced from " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "21.3\\%" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "14.7\\%" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": ". This large difference shows the substantial coding gain of our qaudtree partition-based entropy coding, and verifies the advantages of diverse spatial and channel contexts via finer-grained partition." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22622" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 91, + 274, + 228 + ], + "blocks": [ + { + "bbox": [ + 84, + 71, + 250, + 81 + ], + "lines": [ + { + "bbox": [ + 84, + 71, + 250, + 81 + ], + "spans": [ + { + "bbox": [ + 84, + 71, + 250, + 81 + ], + "type": "text", + "content": "Table 4. Ablation Study on Diverse Contexts." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 91, + 274, + 228 + ], + "lines": [ + { + "bbox": [ + 61, + 91, + 274, + 228 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 274, + 228 + ], + "type": "table", + "html": "
MaMbMcMdMe
Hierarchical quality structure✓
Offset diversity w/ cross-group✓✓
Offset diversity w/o cross-group [8]✓
Quadtree partition based model✓✓✓✓
Dual spatial model [29]✓
BD-Rate(%)0.08.412.114.721.3
", + "image_path": "c22f58e743e86ef2e7225182deaeb9eb40b49a229fd9969b3732e8af440178b7.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 233, + 287, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 233, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 46, + 233, + 287, + 376 + ], + "type": "text", + "content": "From temporal dimension, we also design hierarchical quality structure and offset diversity with cross-group interaction. In Table 4, based on " + }, + { + "bbox": [ + 46, + 233, + 287, + 376 + ], + "type": "inline_equation", + "content": "M_{d}" + }, + { + "bbox": [ + 46, + 233, + 287, + 376 + ], + "type": "text", + "content": ", we first test the original offset diversity [8] without cross-group interaction, i.e., removing the reorder operation in Fig. 4. However, it " + }, + { + "bbox": [ + 46, + 233, + 287, + 376 + ], + "type": "inline_equation", + "content": "(M_{c})" + }, + { + "bbox": [ + 46, + 233, + 287, + 376 + ], + "type": "text", + "content": " only brings " + }, + { + "bbox": [ + 46, + 233, + 287, + 376 + ], + "type": "inline_equation", + "content": "2.6\\%" + }, + { + "bbox": [ + 46, + 233, + 287, + 376 + ], + "type": "text", + "content": " BD-rate difference. By contrast, powered by our cross-group interaction, the potential of offset diversion is fully tapped, and " + }, + { + "bbox": [ + 46, + 233, + 287, + 376 + ], + "type": "inline_equation", + "content": "M_{b}" + }, + { + "bbox": [ + 46, + 233, + 287, + 376 + ], + "type": "text", + "content": " reduces the BD-rate number by " + }, + { + "bbox": [ + 46, + 233, + 287, + 376 + ], + "type": "inline_equation", + "content": "6.3\\%" + }, + { + "bbox": [ + 46, + 233, + 287, + 376 + ], + "type": "text", + "content": " over " + }, + { + "bbox": [ + 46, + 233, + 287, + 376 + ], + "type": "inline_equation", + "content": "M_{d}" + }, + { + "bbox": [ + 46, + 233, + 287, + 376 + ], + "type": "text", + "content": ". At last, based on " + }, + { + "bbox": [ + 46, + 233, + 287, + 376 + ], + "type": "inline_equation", + "content": "M_{b}" + }, + { + "bbox": [ + 46, + 233, + 287, + 376 + ], + "type": "text", + "content": ", we evaluate the hierarchical quality structure, i.e., " + }, + { + "bbox": [ + 46, + 233, + 287, + 376 + ], + "type": "inline_equation", + "content": "M_{a}" + }, + { + "bbox": [ + 46, + 233, + 287, + 376 + ], + "type": "text", + "content": ". The " + }, + { + "bbox": [ + 46, + 233, + 287, + 376 + ], + "type": "inline_equation", + "content": "8.4\\%" + }, + { + "bbox": [ + 46, + 233, + 287, + 376 + ], + "type": "text", + "content": " gap shows learning high-quality contexts brings large benefits to the mining of temporal correlation across many frames." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 377, + 287, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 377, + 287, + 496 + ], + "spans": [ + { + "bbox": [ + 46, + 377, + 287, + 496 + ], + "type": "text", + "content": "Structure optimization. Although our codec learns utilizing diverse contexts in efficient manner, we still purse better tradeoff between compression ratio and computational cost. Therefore, we further optimize our model in network structure. Table 5 shows the study. Based on the " + }, + { + "bbox": [ + 46, + 377, + 287, + 496 + ], + "type": "inline_equation", + "content": "M_{a}" + }, + { + "bbox": [ + 46, + 377, + 287, + 496 + ], + "type": "text", + "content": " (same with that in Table 4), we first implement the depthwise separable convolution into codec. " + }, + { + "bbox": [ + 46, + 377, + 287, + 496 + ], + "type": "inline_equation", + "content": "M_{h}" + }, + { + "bbox": [ + 46, + 377, + 287, + 496 + ], + "type": "text", + "content": " shows widely using depthwise separable convolution to replace the normal convolution not only significantly reduces the MACs, but also brings some compression ratio improvements." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 497, + 287, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 497, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 497, + 287, + 664 + ], + "type": "text", + "content": "The second acceleration is that we use the unequal channel number settings. Not like many existing NVCs use the same channel number for features with different resolutions, we propose assigning the larger number for low resolution feature to increase the latent representation capacity while using the smaller number for the high resolution feature to accelerate model. The performance of " + }, + { + "bbox": [ + 46, + 497, + 287, + 664 + ], + "type": "inline_equation", + "content": "M_{g}" + }, + { + "bbox": [ + 46, + 497, + 287, + 664 + ], + "type": "text", + "content": " verifies the effectiveness of our improvement. In addition, many existing NVCs perform the quantization at the low-resolution latent representation after encoding. To achieve more precise rate adjustment, this paper moves partial quantization operations to the higher resolution in the encoder. " + }, + { + "bbox": [ + 46, + 497, + 287, + 664 + ], + "type": "inline_equation", + "content": "M_{f}" + }, + { + "bbox": [ + 46, + 497, + 287, + 664 + ], + "type": "text", + "content": " shows the integration of encoding and quantization brings some BD-rate improvements with negligible MAC change." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 671, + 124, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 124, + 684 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 124, + 684 + ], + "type": "text", + "content": "4.4. Complexity" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "The complexity comparison is shown in Table 6. We find the MACs of our DCVC-DC are reduced by " + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "inline_equation", + "content": "19.4\\%" + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": " when" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 310, + 91, + 545, + 180 + ], + "blocks": [ + { + "bbox": [ + 332, + 71, + 518, + 82 + ], + "lines": [ + { + "bbox": [ + 332, + 71, + 518, + 82 + ], + "spans": [ + { + "bbox": [ + 332, + 71, + 518, + 82 + ], + "type": "text", + "content": "Table 5. Ablation Study on Structure Optimization." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 310, + 91, + 545, + 180 + ], + "lines": [ + { + "bbox": [ + 310, + 91, + 545, + 180 + ], + "spans": [ + { + "bbox": [ + 310, + 91, + 545, + 180 + ], + "type": "table", + "html": "
MfMgMhMa
Quant at high resolution✓
Unequal channel setting✓✓
Depthwise separable conv✓✓✓
MACs2642G2642G2939G3456G
BD-Rate (%)0.01.12.43.5
", + "image_path": "d6943a336157e62405948371d8cb56279be4a28fbe2d476f134656fd3ec831cb.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 312, + 211, + 542, + 255 + ], + "blocks": [ + { + "bbox": [ + 364, + 190, + 487, + 202 + ], + "lines": [ + { + "bbox": [ + 364, + 190, + 487, + 202 + ], + "spans": [ + { + "bbox": [ + 364, + 190, + 487, + 202 + ], + "type": "text", + "content": "Table 6. Complexity comparison." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 312, + 211, + 542, + 255 + ], + "lines": [ + { + "bbox": [ + 312, + 211, + 542, + 255 + ], + "spans": [ + { + "bbox": [ + 312, + 211, + 542, + 255 + ], + "type": "table", + "html": "
MACsEncoding TimeDecoding Time
DCVC-HEM [29]3279G890ms652ms
Our DCVC-DC2642G1005ms765ms
", + "image_path": "e0bec5e8158c76a54b506bfc95acbc507bb24a42f71fa51b5f265fb2ca354a67.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 309, + 258, + 504, + 269 + ], + "lines": [ + { + "bbox": [ + 309, + 258, + 504, + 269 + ], + "spans": [ + { + "bbox": [ + 309, + 258, + 504, + 269 + ], + "type": "text", + "content": "Note: Tested on NVIDIA 2080TI with using " + }, + { + "bbox": [ + 309, + 258, + 504, + 269 + ], + "type": "inline_equation", + "content": "1080\\mathrm{p}" + }, + { + "bbox": [ + 309, + 258, + 504, + 269 + ], + "type": "text", + "content": " as input." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 274, + 545, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 274, + 545, + 417 + ], + "spans": [ + { + "bbox": [ + 304, + 274, + 545, + 417 + ], + "type": "text", + "content": "compared with DCVC-HEM [29]. However, the actual encoding and decoding time is higher. This is because currently the computational density of depthwise convolution is not as high as normal convolution under the same MAC condition. But through the customized optimization [35], it can be further accelerated in the future. From another perspective, considering that our DCVC-DC has " + }, + { + "bbox": [ + 304, + 274, + 545, + 417 + ], + "type": "inline_equation", + "content": "23.5\\%" + }, + { + "bbox": [ + 304, + 274, + 545, + 417 + ], + "type": "text", + "content": " bitrate saving over previous SOTA DCVC-HEM [29], such increase degree in running time is a price worth paying. By contrast, ECM brings " + }, + { + "bbox": [ + 304, + 274, + 545, + 417 + ], + "type": "inline_equation", + "content": "13.1\\%" + }, + { + "bbox": [ + 304, + 274, + 545, + 417 + ], + "type": "text", + "content": " (Table 3) improvement over its predecessor VTM, but the encoding complexity is more than 4 times [49] of VTM." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 437, + 458, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 437, + 458, + 449 + ], + "spans": [ + { + "bbox": [ + 306, + 437, + 458, + 449 + ], + "type": "text", + "content": "5. Conclusion and Limitation" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 460, + 545, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 545, + 614 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 545, + 614 + ], + "type": "text", + "content": "In this paper, we have presented how to utilize diverse contexts to further boost NVC. From temporal dimension, the model is guided to extract long-term and yet high-quality contexts to alleviate error propagation and exploit long range correlation. The offset diversity with cross-group interaction provides complementary motion alignments to handle complex motion. From spatial dimension, the fine-grained quadtree-based partition is proposed to increase spatial context diversity. Powered by our techniques, the compression ratio of NVC has been pushed to new height. Our DCVC-DC has surpassed ECM in both RGB and YUV420 colorspaces, which is an important milestone in the development of NVC." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "During the training, to learn the hierarchical quality pattern, we still use the fixed distortion weights which are similar with those in traditional codec. This may not be the best choice for NVC. Actually, reinforcement learning is good at solving such kind of time series weight decision problem. In the future, we will investigate utilizing reinforcement learning to help NVC make better weight decision with considering the temporal dependency." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22623" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "type": "text", + "content": "[1] Eirikur Agustsson, David Minnen, Nick Johnston, Johannes Balle, Sung Jin Hwang, and George Toderici. Scale-space flow for end-to-end optimized video compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8503-8512, 2020. 1, 2, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 159, + 288, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 159, + 288, + 191 + ], + "spans": [ + { + "bbox": [ + 53, + 159, + 288, + 191 + ], + "type": "text", + "content": "[2] E. Alshina, J. Ascenso, T. Ebrahimi, F. Pereira, and T. Richter. [AHG 11] Brief information about JPEG AI CfP status. In JVET-AA0047, 2022. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 193, + 287, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 193, + 287, + 224 + ], + "spans": [ + { + "bbox": [ + 53, + 193, + 287, + 224 + ], + "type": "text", + "content": "[3] Anchors · JPEG-AI MMSP Challenge. https://jpegai.github.io/7-anchors/, 2022. Accessed: 2022-11-02. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 227, + 287, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 227, + 287, + 270 + ], + "spans": [ + { + "bbox": [ + 53, + 227, + 287, + 270 + ], + "type": "text", + "content": "[4] Johannes Balle, David Minnen, Saurabh Singh, Sung Jin Hwang, and Nick Johnston. Variational image compression with a scale hyperprior. 6th International Conference on Learning Representations, ICLR, 2018. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 272, + 287, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 272, + 287, + 294 + ], + "spans": [ + { + "bbox": [ + 53, + 272, + 287, + 294 + ], + "type": "text", + "content": "[5] Gisle Bjontegaard. Calculation of average PSNR differences between RD-curves. VCEG-M33, 2001. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 296, + 287, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 296, + 287, + 317 + ], + "spans": [ + { + "bbox": [ + 53, + 296, + 287, + 317 + ], + "type": "text", + "content": "[6] Frank Bossen et al. Common test conditions and software reference configurations. In JCTVC-L1100, 2013. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 319, + 287, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 319, + 287, + 373 + ], + "spans": [ + { + "bbox": [ + 53, + 319, + 287, + 373 + ], + "type": "text", + "content": "[7] Benjamin Bross, Ye-Kui Wang, Yan Ye, Shan Liu, Jianle Chen, Gary J Sullivan, and Jens-Rainer Ohm. Overview of the versatile video coding (VVC) standard and its applications. IEEE Transactions on Circuits and Systems for Video Technology, 31(10):3736-3764, 2021. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 375, + 287, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 375, + 287, + 418 + ], + "spans": [ + { + "bbox": [ + 53, + 375, + 287, + 418 + ], + "type": "text", + "content": "[8] Kelvin CK Chan, Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Understanding deformable alignment in video super-resolution. In Proceedings of the AAAI conference on artificial intelligence, 2021. 2, 4, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 420, + 287, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 420, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 53, + 420, + 287, + 475 + ], + "type": "text", + "content": "[9] Kelvin CK Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. BasicVSR++: Improving video superresolution with enhanced propagation and alignment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5972-5981, 2022. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 476, + 287, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 476, + 287, + 520 + ], + "spans": [ + { + "bbox": [ + 48, + 476, + 287, + 520 + ], + "type": "text", + "content": "[10] François Chollet. Xception: Deep learning with depthwise separable convolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1251-1258, 2017. 2, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 521, + 287, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 521, + 287, + 576 + ], + "spans": [ + { + "bbox": [ + 48, + 521, + 287, + 576 + ], + "type": "text", + "content": "[11] Ze Cui, Jing Wang, Shangyin Gao, Tiansheng Guo, Yihui Feng, and Bo Bai. Asymmetric gained deep image compression with continuous rate adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10532-10541, 2021. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 578, + 287, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 287, + 609 + ], + "type": "text", + "content": "[12] Armen Der Kiureghian and Ove Ditlevsen. Aleatory or epistemic? does it matter? Structural safety, 31(2):105-112, 2009. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 612, + 287, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 612, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 48, + 612, + 287, + 666 + ], + "type": "text", + "content": "[13] Abdelaziz Djelouah, Joaquim Campos, Simone Schaub-Meyer, and Christopher Schroers. Neural inter-frame compression for video coding. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2019. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 668, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 668, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 668, + 287, + 689 + ], + "type": "text", + "content": "[14] ECM-5.0. https://vgit.hhi.fraunhofer.de/ecm/ECM, 2022. Accessed: 2022-11-02. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[15] D Flynn, K Sharman, and C Rosewarne. Common test conditions and software reference configurations for hevc range" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "extensions, document jctvc-n1006. Joint Collaborative Team Video Coding ITU-T SG, 16, 2013. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 96, + 545, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 118 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 118 + ], + "type": "text", + "content": "[16] Yarin Gal. Uncertainty in Deep Learning. PhD thesis, University of Cambridge, 2016. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 119, + 545, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 119, + 545, + 163 + ], + "spans": [ + { + "bbox": [ + 307, + 119, + 545, + 163 + ], + "type": "text", + "content": "[17] Bernd Girod, Eckehard G Steinbach, and Niko Faerber. Comparison of the H. 263 and H. 261 video compression standards. In Standards and Common Interfaces for Video Information Systems: A Critical Review, 1995. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 164, + 545, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 164, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 307, + 164, + 545, + 207 + ], + "type": "text", + "content": "[18] Zongyu Guo, Zhizheng Zhang, Runsen Feng, and Zhibo Chen. Causal contextual prediction for learned image compression. IEEE Transactions on Circuits and Systems for Video Technology, 32(4):2329-2341, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 209, + 545, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 209, + 545, + 263 + ], + "spans": [ + { + "bbox": [ + 307, + 209, + 545, + 263 + ], + "type": "text", + "content": "[19] Dailan He, Yaoyan Zheng, Baocheng Sun, Yan Wang, and Hongwei Qin. Checkerboard context model for efficient learned image compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14771-14780, 2021. 2, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 265, + 545, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 265, + 545, + 286 + ], + "spans": [ + { + "bbox": [ + 307, + 265, + 545, + 286 + ], + "type": "text", + "content": "[20] HM-16.25. https://vgit.hhi.fraunhofer.de/jvet/HM/, 2022. Accessed: 2022-11-02. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 288, + 545, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 288, + 545, + 331 + ], + "spans": [ + { + "bbox": [ + 307, + 288, + 545, + 331 + ], + "type": "text", + "content": "[21] Yung-Han Ho, Chih-Peng Chang, Peng-Yu Chen, Alessandro Gnutti, and Wen-Hsiao Peng. Canf-vc: Conditional augmented normalizing flows for video compression. European Conference on Computer Vision, 2022. 1, 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 333, + 545, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 333, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 307, + 333, + 545, + 387 + ], + "type": "text", + "content": "[22] Zhihao Hu, Guo Lu, Jinyang Guo, Shan Liu, Wei Jiang, and Dong Xu. Coarse-to-fine deep video coding with hyperprior-guided mode prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5921-5930, 2022. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 388, + 545, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 545, + 431 + ], + "type": "text", + "content": "[23] Zhihao Hu, Guo Lu, and Dong Xu. FVC: A new framework towards deep video compression in feature space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1502-1511, 2021. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 433, + 545, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 433, + 545, + 466 + ], + "spans": [ + { + "bbox": [ + 307, + 433, + 545, + 466 + ], + "type": "text", + "content": "[24] Yan Huang, Li Song, and Ebroul Izquierdo. CNN accelerated intra video coding, where is the upper bound? In 2019 Picture Coding Symposium (PCS), pages 1-5. IEEE, 2019. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 468, + 545, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 468, + 545, + 499 + ], + "spans": [ + { + "bbox": [ + 307, + 468, + 545, + 499 + ], + "type": "text", + "content": "[25] Shuai Huo, Dong Liu, Li Li, Siwei Ma, Feng Wu, and Wen Gao. Towards hybrid-optimization video coding. arXiv preprint arXiv:2207.05565, 2022. 1, 2, 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 501, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 501, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 307, + 501, + 545, + 555 + ], + "type": "text", + "content": "[26] Jun-Hyuk Kim, Byeongho Heo, and Jong-Seok Lee. Joint global and local hierarchical priors for learned image compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5992-6001, 2022. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 556, + 545, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 556, + 545, + 610 + ], + "spans": [ + { + "bbox": [ + 307, + 556, + 545, + 610 + ], + "type": "text", + "content": "[27] Théo Ladune, Pierick Philippe, Wassim Hamidouche, Lu Zhang, and Olivier Déforges. Conditional coding for flexible learned video compression. In Neural Compression: From Information Theory to Applications - Workshop @ ICLR, 2021. 1, 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 613, + 545, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 545, + 644 + ], + "type": "text", + "content": "[28] Jiahao Li, Bin Li, and Yan Lu. Deep contextual video compression. Advances in Neural Information Processing Systems, 34, 2021. 1, 2, 6, 7" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "type": "text", + "content": "[29] Jiahao Li, Bin Li, and Yan Lu. Hybrid spatial-temporal entropy modelling for neural video compression. In Proceedings of the 30th ACM International Conference on Multimedia, pages 1503-1511, 2022. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "text", + "content": "[30] Chong Soon Lim, SMT Naing, V Wahadaniah, and X Jing. Reference lists for B pictures under low delay constraints." + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "22624" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 105 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 105 + ], + "type": "text", + "content": "document JCTVC-D093, ITU-T/ISO/IEC Joint Collaborative Team on Video Coding (JCT-VC), Daegu, Korea, 2011. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 287, + 150 + ], + "type": "text", + "content": "[31] Jianping Lin, Dong Liu, Houqiang Li, and Feng Wu. M-LVC: multiple frames prediction for learned video compression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 150, + 288, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 150, + 288, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 150, + 288, + 205 + ], + "type": "text", + "content": "[32] Haojie Liu, Ming Lu, Zhan Ma, Fan Wang, Zhihuang Xie, Xun Cao, and Yao Wang. Neural video coding using multiscale motion compensation and spatiotemporal context model. IEEE Transactions on Circuits and Systems for Video Technology, 2020. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 205, + 288, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 205, + 288, + 259 + ], + "spans": [ + { + "bbox": [ + 48, + 205, + 288, + 259 + ], + "type": "text", + "content": "[33] Jerry Liu, Shenlong Wang, Wei-Chiu Ma, Meet Shah, Rui Hu, Pranaab Dhawan, and Raquel Urtasun. Conditional entropy coding for efficient video compression. In European Conference on Computer Vision, pages 453-468. Springer, 2020. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 261, + 288, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 261, + 288, + 315 + ], + "spans": [ + { + "bbox": [ + 48, + 261, + 288, + 315 + ], + "type": "text", + "content": "[34] Guo Lu, Wanli Ouyang, Dong Xu, Xiaoyun Zhang, Chunlei Cai, and Zhiyong Gao. DVC: an end-to-end deep video compression framework. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11006-11015, 2019. 1, 2, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 316, + 288, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 316, + 288, + 358 + ], + "spans": [ + { + "bbox": [ + 48, + 316, + 288, + 358 + ], + "type": "text", + "content": "[35] Gangzhao Lu, Weizhe Zhang, and Zheng Wang. Optimizing depthwise separable convolution operations on gpus. IEEE Transactions on Parallel and Distributed Systems, 33(1):70-87, 2021. 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 360, + 288, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 360, + 288, + 403 + ], + "spans": [ + { + "bbox": [ + 48, + 360, + 288, + 403 + ], + "type": "text", + "content": "[36] Guo Lu, Xiaoyun Zhang, Wanli Ouyang, Li Chen, Zhiyong Gao, and Dong Xu. An end-to-end learning framework for video compression. IEEE transactions on pattern analysis and machine intelligence, 2020. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 404, + 288, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 404, + 288, + 426 + ], + "spans": [ + { + "bbox": [ + 48, + 404, + 288, + 426 + ], + "type": "text", + "content": "[37] Wufei Ma, Jiahao Li, Bin Li, and Yan Lu. Uncertainty-aware deep video compression with ensembles, 2021. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 426, + 288, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 426, + 288, + 470 + ], + "spans": [ + { + "bbox": [ + 48, + 426, + 288, + 470 + ], + "type": "text", + "content": "[38] Fabian Mentzer, George Toderici, David Minnen, Sung-Jin Hwang, Sergi Caelles, Mario Lucic, and Eirikur Agustsson. Vct: A video compression transformer. arXiv preprint arXiv:2206.07307, 2022. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 471, + 288, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 471, + 288, + 514 + ], + "spans": [ + { + "bbox": [ + 48, + 471, + 288, + 514 + ], + "type": "text", + "content": "[39] Alexandre Mercat, Marko Viitanen, and Jarno Vanne. UVG dataset: 50/120fps 4k sequences for video codec analysis and development. In Proceedings of the 11th ACM Multimedia Systems Conference, pages 297-302, 2020. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 515, + 288, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 515, + 288, + 558 + ], + "spans": [ + { + "bbox": [ + 48, + 515, + 288, + 558 + ], + "type": "text", + "content": "[40] David Minnen, Johannes Ballé, and George D Toderici. Joint autoregressive and hierarchical priors for learned image compression. Advances in neural information processing systems, 31, 2018. 2, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 559, + 288, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 288, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 288, + 624 + ], + "type": "text", + "content": "[41] Ken M Nakanishi, Shin-ichi Maeda, Takeru Miyato, and Daisuke Okanohara. Neural multi-scale image compression. In Computer Vision-ACCV 2018: 14th Asian Conference on Computer Vision, Perth, Australia, December 2–6, 2018, Revised Selected Papers, Part VI 14, pages 718–732. Springer, 2019. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 625, + 288, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 288, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 288, + 690 + ], + "type": "text", + "content": "[42] Jonathan Pfaff, Alexey Filippov, Shan Liu, Xin Zhao, Jianle Chen, Santiago De-Luxan-Hernandez, Thomas Wiegand, Vasily Rufitskiy, Adarsh Krishnan Ramasubramonian, and Geert Van der Auwera. Intra prediction and mode coding in VLC. IEEE Transactions on Circuits and Systems for Video Technology, 31(10):3834-3847, 2021. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 714 + ], + "type": "text", + "content": "[43] Reza Pourreza, Hoang Le, Amir Said, Guillaume Sautiere, and Auke Wiggers. Boosting neural video codec" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "by exploiting hierarchical redundancy. arXiv preprint arXiv:2208.04303, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 96, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 139 + ], + "type": "text", + "content": "[44] Yichen Qian, Ming Lin, Xiuyu Sun, Zhiyu Tan, and Rong Jin. Entroformer: A transformer-based entropy model for learned image compression. arXiv preprint arXiv:2202.05492, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 141, + 545, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 185 + ], + "type": "text", + "content": "[45] Yichen Qian, Zhiyu Tan, Xiuyu Sun, Ming Lin, Dongyang Li, Zhenhong Sun, Hao Li, and Rong Jin. Learning accurate entropy model with global reference for image compression. arXiv preprint arXiv:2010.08321, 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 186, + 545, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 241 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 241 + ], + "type": "text", + "content": "[46] Scott Reed, Aaron Oord, Nal Kalchbrenner, Sergio Gomez Colmenarejo, Ziyu Wang, Yutian Chen, Dan Belov, and Nando Freitas. Parallel multiscale autoregressive density estimation. In International Conference on Machine Learning, pages 2912-2921. PMLR, 2017. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 242, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 242, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 307, + 242, + 545, + 297 + ], + "type": "text", + "content": "[47] Oren Rippel, Alexander G Anderson, Kedar Tatwawadi, Sanjay Nair, Craig Lytle, and Lubomir Bourdev. ELF-VC: Efficient learned flexible-rate video coding. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14479-14488, October 2021. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 298, + 545, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 298, + 545, + 341 + ], + "spans": [ + { + "bbox": [ + 307, + 298, + 545, + 341 + ], + "type": "text", + "content": "[48] Heiko Schwarz, Detlev Marpe, and Thomas Wiegand. Analysis of hierarchical B pictures and MCTF. In 2006 IEEE International Conference on Multimedia and Expo, pages 1929-1932. IEEE, 2006. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 342, + 545, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 342, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 307, + 342, + 545, + 374 + ], + "type": "text", + "content": "[49] Vadim Seregin, Jie Chen, Fabrice Leannec, and Kai Zhang. JVET AHG report: ECM software development (AHG6). In JVET-AA0006, 2022. 1, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 376, + 545, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 376, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 307, + 376, + 545, + 409 + ], + "type": "text", + "content": "[50] Xihua Sheng, Jiahao Li, Bin Li, Li Li, Dong Liu, and Yan Lu. Temporal Context Mining for Learned Video Compression. IEEE Transactions on Multimedia, 2022. 1, 2, 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 411, + 545, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 411, + 545, + 454 + ], + "spans": [ + { + "bbox": [ + 307, + 411, + 545, + 454 + ], + "type": "text", + "content": "[51] Hui Su, Mingliang Chen, Alexander Bokov, Debargha Mukherjee, Yunqing Wang, and Yue Chen. Machine learning accelerated transform search for av1. In 2019 Picture Coding Symposium (PCS), pages 1-5. IEEE, 2019. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 456, + 545, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 456, + 545, + 499 + ], + "spans": [ + { + "bbox": [ + 307, + 456, + 545, + 499 + ], + "type": "text", + "content": "[52] Hui Su, Chi-Yo Tsai, Yunqing Wang, and Yaowu Xu. Machine learning accelerated partition search for video encoding. In 2019 IEEE International Conference on Image Processing (ICIP), pages 2661-2665. IEEE, 2019. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 500, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 500, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 307, + 500, + 545, + 544 + ], + "type": "text", + "content": "[53] Gary J Sullivan and Jens-Rainer Ohm. Meeting report of the fourth meeting of the joint collaborative team on video coding (jct-vc), daegu, kr, 20-28 january 2011. Document JCTVC-D500, Daegu, KR, 2011. 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 545, + 545, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 545, + 545, + 599 + ], + "spans": [ + { + "bbox": [ + 307, + 545, + 545, + 599 + ], + "type": "text", + "content": "[54] Yapeng Tian, Yulun Zhang, Yun Fu, and Chenliang Xu. TDAN: Temporally-deformable alignment network for video super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3360-3369, 2020. 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 601, + 545, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 601, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 307, + 601, + 545, + 633 + ], + "type": "text", + "content": "[55] VTM-17.0. https://vcert.hhi.fraunhofer.de/jvet/VVCSoftware_VTM/, 2022. Accessed: 2022-11-02.6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 635, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 545, + 678 + ], + "type": "text", + "content": "[56] Guo-Hua Wang, Jiahao Li, Bin Li, and Yan Lu. EVC: Towards Real-Time Neural Image Compression with Mask Decay. In International Conference on Learning Representations, 2023. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "text", + "content": "[57] Haiqiang Wang, Weihao Gan, Sudeng Hu, Joe Yuchieh Lin, Lina Jin, Longguang Song, Ping Wang, Ioannis Katsavounidis, Anne Aaron, and C-C Jay Kuo. MCL-JCV: a JND-based" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22625" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 361 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 106 + ], + "type": "text", + "content": "H. 264/AVC video quality assessment dataset. In 2016 IEEE International Conference on Image Processing (ICIP), pages 1509-1513. IEEE, 2016. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "type": "text", + "content": "[58] Xintao Wang, Kelvin CK Chan, Ke Yu, Chao Dong, and Chen Change Loy. EDVR: Video restoration with enhanced deformable convolutional networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 0–0, 2019. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 287, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 287, + 206 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 287, + 206 + ], + "type": "text", + "content": "[59] Chao-Yuan Wu, Nayan Singhal, and Philipp Krahenbuhl. Video compression through image interpolation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 416-431, 2018. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 208, + 287, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 208, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 48, + 208, + 287, + 251 + ], + "type": "text", + "content": "[60] Tianfan Xue, Baian Chen, Jiajun Wu, Donglai Wei, and William T Freeman. Video enhancement with task-oriented flow. International Journal of Computer Vision (IJCV), 127(8):1106-1125, 2019. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 253, + 287, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 253, + 287, + 306 + ], + "spans": [ + { + "bbox": [ + 48, + 253, + 287, + 306 + ], + "type": "text", + "content": "[61] Ren Yang, Fabian Mentzer, Luc Van Gool, and Radu Timofte. Learning for video compression with recurrent autoencoder and recurrent probability model. IEEE Journal of Selected Topics in Signal Processing, 15(2):388-401, 2021. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 308, + 287, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 308, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 48, + 308, + 287, + 361 + ], + "type": "text", + "content": "[62] Xiangyu Zhang, Xinyu Zhou, Mengxiao Lin, and Jian Sun. Shufflenet: An extremely efficient convolutional neural network for mobile devices. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6848-6856, 2018. 4" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22626" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Neural Volumetric Memory for Visual Locomotion Control/dad36583-d1e2-4d95-a7b6-551afd512667_content_list.json b/2023/Neural Volumetric Memory for Visual Locomotion Control/dad36583-d1e2-4d95-a7b6-551afd512667_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..cdb2130ae9024556618506c5b0037b90b0d1849b --- /dev/null +++ b/2023/Neural Volumetric Memory for Visual Locomotion Control/dad36583-d1e2-4d95-a7b6-551afd512667_content_list.json @@ -0,0 +1,1476 @@ +[ + { + "type": "text", + "text": "Neural Volumetric Memory for Visual Locomotion Control", + "text_level": 1, + "bbox": [ + 186, + 117, + 782, + 138 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ruihan Yang1 Ge Yang2,3 Xiaolong Wang1", + "bbox": [ + 292, + 167, + 661, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ UC San Diego $^{2}$ Institute of AI and Fundamental Interactions $^{3}$ MIT CSAIL", + "bbox": [ + 179, + 186, + 785, + 204 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/37fbd98d0a807e7d96e00c56a762f543b5e37e2c7ebc8d833ab517e20a0e6662.jpg", + "image_caption": [ + "Figure 1. Real-World Deployment: We deploy our policy in real-world environments with Stones, Stairs, Stages, Obstacles, and Unstructured Terrain. Our robot utilizes the volumetric memory of the surrounding 3D structure to successfully traverse complex environments." + ], + "image_footnote": [], + "bbox": [ + 80, + 215, + 892, + 412 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 477, + 313, + 494 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Legged robots have the potential to expand the reach of autonomy beyond paved roads. In this work, we consider the difficult problem of locomotion on challenging terrains using a single forward-facing depth camera. Due to the partial observability of the problem, the robot has to rely on past observations to infer the terrain currently beneath it. To solve this problem, we follow the paradigm in computer vision that explicitly models the 3D geometry of the scene and propose Neural Volumetric Memory (NVM), a geometric memory architecture that explicitly accounts for the SE(3) equivariance of the 3D world. NVM aggregates feature volumes from multiple camera views by first bringing them back to the ego-centric frame of the robot. We test the learned visual-locomotion policy on a physical robot and show that our approach, which explicitly introduces geometric priors during training, offers superior performance than more naive methods. We also include ablation studies and show that the representations stored in the neural volumetric memory capture sufficient geometric information to reconstruct the scene. Our project page with videos is https://rchalyang.github.io/NVM", + "bbox": [ + 75, + 510, + 472, + 828 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 843, + 209, + 859 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Consider difficult locomotion tasks such as walking up and down a flight of stairs and stepping over gaps (Figure 1).", + "bbox": [ + 75, + 869, + 468, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The control of such behaviors requires tight coupling with perception because vision is needed to provide details of the terrain right beneath the robot and the 3D scene immediately around it. This problem is also partially-observable. Immediately relevant terrain information is often occluded from the robot's current frame of observation, forcing it to rely on past observations for control decisions. For this reason, while blind controllers that are learned in simulation using reinforcement learning have achieved impressive results in agility and robustness [33, 36, 38], there are clear limitations on how much they can do. How to incorporate perception into the pipeline to produce an integrated visuomotor controller thus remains an open problem.", + "bbox": [ + 496, + 478, + 893, + 676 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A recent line of work combines perception with locomotion using ego-centric cameras mounted on the robot. The predominant approach for addressing partial observability is to do frame-stacking, where the robot maintains a visual buffer of recent images. This naive heuristic suffers from two major problems: first, frame-stacking on a moving robot ignores the equivariance structure of the 3D environment, making learning a lot more difficult as policy success now relies on being able to learn to account for spurious changes in camera poses. A second but subtler issue is that biological systems do not have the ability to save detailed visual observations pixel-by-pixel. These concerns motivate the creation of an intermediary, short-term memory mechanism to functionally aggregate streams of obser", + "bbox": [ + 496, + 689, + 895, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1430", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c5c8690fe4b432b9bf5793d57eee7ca061c0c7956d1c08dd5da9ce40760374da.jpg", + "image_caption": [ + "Figure 2. Overview of Simulated Environment & Real World Environment: Our simulated environments are shown on the left and real-world environments are shown on the right. For the real-world environment, the corresponding visual observations for each real-world environment are shown in the bottom row. All policies are trained in the simulation and transferred into the real world without fine-tuning." + ], + "image_footnote": [], + "bbox": [ + 76, + 87, + 893, + 239 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "vation into a single, coherent representation of the world.", + "bbox": [ + 76, + 304, + 455, + 319 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Motivated by these observations, we introduce a novel volumetric memory architecture for legged locomotion control. Our architecture consists of a 2D to 3D feature volume encoder, and a pose estimator that can estimate the relative camera poses given two input images. When combined, the two networks constitute a Neural Volumetric Memory (NVM) that takes as input a sequence of depth images taken by a forward-looking, ego-centric camera and fuses them together into a coherent latent representation for locomotion control. We encourage the memory space to be $SE(3)$ equivariant to changes in the camera pose of the robot by incorporating translation and rotation operations based on estimated relative poses from the pose network. This inverse transformation allows NVM to align feature volumes from the past to the present ego-centric frame, making both integrating over multiple timesteps into a coherent scene representation and learning a policy, less difficult.", + "bbox": [ + 75, + 321, + 472, + 580 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our training pipeline follows a two-step teacher-student process where the primary goal of the first stage is to produce behaviors in the form of a policy. After training completes, this policy can traverse these difficult terrains robustly, but it relies on privileged sensory information such as an elevation map and ground-truth velocity. Elevation maps obtained in the real world are often biased, incomplete, and full of errors [40], whereas ground-truth velocity information is typically only available in instrumented environments. Hence in the visuomotor distillation stage of the pipeline, which still runs in the simulator, we feed the stream of ego-centric views from the forward depth camera into the neural volumetric memory. We feed the content of this memory into a small policy network and train everything end-to-end including the two network components of the NVM using a behavior cloning loss where the state-only policy acts as the teacher. For completeness, we offer an additional self-supervised learning objective (Figure 4) that relies on novel-view consistency for learning. The end product of this visuomotor distillation pipeline is a memory-equipped visuomotor policy that can operate directly on the", + "bbox": [ + 75, + 584, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "UniTree A1 robot hardware (see Figure 1). A single policy is used to handle all environments covered by this paper. We provide comprehensive experiment and ablation studies in both simulation and the real world, and show that our method outperforms all baselines by a large margin. It is thus essential to model the 3D structure of the environment.", + "bbox": [ + 496, + 304, + 893, + 395 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 498, + 411, + 640, + 426 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Blind Locomotion Controllers: Legged locomotion using the proprioception of the robot has been studied for decades [3, 16, 42, 51, 61, 68]. Traditionally, controllers are developed with model-based methods [2, 4, 5, 9, 10, 15, 17, 59]. However, most model-based methods fail to generalize to unseen environments in the real world. In order to attain better generalization and robustness, recent works [22, 24, 25, 32, 33, 36, 48, 60, 65] utilize model-free reinforcement learning to train neural network controllers in simulation and directly transfer the learned policies to the real robot. These methods are able to achieve high levels of agility and adapt over diverse terrains using just the proprioceptive state of the robot. The real-world utility of these robots, however, is also limited, since these blind controllers are unable to perceive the surrounding. In this paper, we propose to integrate perception to facilitate control, with a focus on a better 3D understanding of the environments.", + "bbox": [ + 496, + 436, + 893, + 696 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Locomotion with Exteroceptive Perception: The elevation map [12, 13, 34, 46] around the robot has been used to perform optimization and planning for foot placement [8, 11, 47, 63, 66, 71]. Recent RL controller [40, 53] also uses elevation maps in an end-to-end manner. Using the elevation map to represent the surrounding is straightforward for foot placement optimization, but the elevation map obtained in the real world is either noisy or requires a complicated hardware pipeline. Another option is using depth or RGB images to provide visual observation for the controller. Hierarchical control systems [26, 39, 55, 58, 69] are proposed to utilize depth reading while the low-level", + "bbox": [ + 496, + 719, + 895, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1431", + "bbox": [ + 483, + 944, + 513, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/965c9e7a5b0ea50782226b52dd9d5ba31082b5c82f36c2aff77046d3fb38feaf.jpg", + "image_caption": [ + "Figure 3. Neural Volumetric Memory (NVM): Our NVM module extracts 3D feature volumes from every visual observation in the near past, transform them with corresponding estimated transformation into present robot frame, and fuse the transformed feature volumes into a neural volumetric memory for decision-making." + ], + "image_footnote": [], + "bbox": [ + 78, + 87, + 897, + 262 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "controllers still rely on only the proprioceptive state of the robot. To get the most out of the depth image, Yang et al. [67] and Imai et al. [23] propose to learn visual RL policy to output the raw joint angles and enable the robot to maneuver in the wild avoiding different kinds of obstacles. Agarwal et al. [1] and Loquercioo et al. [37] train a policy using RL to traverse a large variety of terrains. Though these works show encouraging results in specific tasks, they have not fully exploited the 3D structure of the environment. On the other hand, we propose to train a novel neural volumetric memory together with policy learning, which allows our robot to walk through more challenging terrains.", + "bbox": [ + 75, + 334, + 472, + 518 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Volumetric Cognitive Map: How to represent and understand the 3D scene has been studied by the computer vision community for a long time. Hartley et al. [20] and Wiles et al. [64] provide representation built on geometry principles and camera motions. Learning-based approaches are proposed [21, 28, 29, 35, 52, 62, 64, 70] to learn 3D representation from 2D inputs. Neural 3D representations [14, 19, 41, 44, 45, 56, 57] has also shown great result in multiple computer vision tasks [14, 18, 19]. Our work is closely related to Lai et al. [35] where the deep voxel feature representing the entire scene and the camera trajectory are learned from video clips without odometry in a self-supervised manner. Instead of focusing on view synthesis, we aggregate the 3D feature volume to form a short-term memory representation for decision-making.", + "bbox": [ + 75, + 534, + 472, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Teacher-Student Training: To accelerate training and utilize the privileged information in simulation, Miki et al. [40] follow a teacher-student training framework, where the teacher policy is trained with privileged information and then distilled into student policy where the privileged information is not available. Multiple works in different research domains [6, 7, 27, 30, 31, 43, 50] follow this framework as well. We share a similar pipeline as [1, 39] where elevation", + "bbox": [ + 75, + 780, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "map-based policies are trained first and the elevation-map-based policies are distilled visual policies afterward.", + "bbox": [ + 498, + 334, + 890, + 367 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Volumetric Memory for Legged Locomotion", + "text_level": 1, + "bbox": [ + 498, + 380, + 890, + 398 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Legged locomotion using ego-centric camera views is intrinsically a partially-observed problem. To make the control problem tangible, our robot needs to aggregate information from previous frames and correctly infer the occluded terrain underneath it. During locomotion, the camera mounted directly on the robot chassis undergo large and spurious changes in pose, making integrating individual frames to a coherent representation non-trivial.", + "bbox": [ + 496, + 407, + 890, + 527 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To account for these camera pose changes, we propose neural volumetric memory (NVM) — a 3D representation format for scene features. It takes as input a sequence of visual observations and outputs a single 3D feature volume representing the surrounding 3D structure.", + "bbox": [ + 496, + 529, + 893, + 604 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Baking-in $SE(3)$ Equivariance in NVM", + "text_level": 1, + "bbox": [ + 498, + 616, + 841, + 633 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our training procedure encourages the learned memory space to be $SE(3)$ equivariant by applying translation and rotation transformations on the feature volume before aggregating additional information from new observations.", + "bbox": [ + 496, + 641, + 890, + 700 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As the first step (Figure 3), an encoder network receives a sequence of consecutive observations $O_{t}..O_{t - n + 1}$ . Instead of generating a compact feature embedding in the shape of a tuple, the 3D encoder Enc3D extracts feature volumes $V_{t}.V_{t - n + 1}$ . A second network Encpose estimates the relative changes in camera pose between the present frame $O_{t}^{v}$ and frames $O_{t}^{v}.O_{t - n + 1}^{v}$ from the past. This sequence of estimate allows NVM to transform the feature volume $V_{t}.V_{t - n + 1}$ from the past to the present frame using the transformations $T_{t}^{t}.T_{t - n + 1}^{t}$ . We annotate the transformed feature volume as $\\hat{V}_{t}. \\hat{V}_{t - n + 1}$ . These canonicalized volumes can now be aggregated together by element-wise sum to produce a fused feature volume $\\hat{V}_{t}^{m}$ for the surround-", + "bbox": [ + 496, + 702, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1432", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ing 3D structure. This scheme allows NVM to disentangle the symmetry structure of the ambient 3D space from the \"content\" specific to the scene. By doing so, we encourage $SE(3)$ equivariance in the memory space.", + "bbox": [ + 75, + 90, + 468, + 151 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We provide a detailed description of all components in our NVM module in the remaining section.", + "bbox": [ + 75, + 151, + 468, + 181 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3D Encoder: Our 3D encoder $Enc_{3D}$ takes a 2D depth image as input and outputs a 3D feature volume. We parameterize this encoder with a set of 2D convolution layers followed by 3D convolution layers. Specifically, we generate a 2D feature map with the shape $(C, H, W)$ using 2D convolution, then reshape the feature map channel-wise into a volume tensor with the shape $(C / D, D, H, W)$ . We chose a voxel-grid size of $D = 12$ , $H = 6$ , $W = 6$ . The 3D convolution layers are added, to further refine the 3D features.", + "bbox": [ + 75, + 181, + 470, + 316 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nV _ {i} = E n c _ {3 D} (O _ {i} ^ {v}) \\quad i = t.. t - n + 1\n$$\n", + "text_format": "latex", + "bbox": [ + 147, + 328, + 398, + 345 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Pose Encoder: Our pose encoder $Enc_{\\mathrm{pose}}$ takes two depth images and estimates the transformation in the latent space between two frames. Specifically, our pose encoder outputs 6D vectors representing the camera pose (rotation and translation) and we parameterize the pose encoder with a set of 2D convs. When constructing the NVM, the pose encoder is used to estimate the transformation $T_{t}^{t}\\ldots T_{t - n + 1}^{t}$ between previous frames and the present frame by comparing previous visual observation $O_{t}^{v}\\ldots O_{t - n + 1}^{v}$ and present $O_{t}^{v}$ .", + "bbox": [ + 75, + 356, + 468, + 492 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nT _ {i} ^ {j} = \\operatorname {E n c} _ {\\text {p o s e}} \\left(O _ {i} ^ {v}, O _ {j} ^ {v}\\right) \\in S E (3)\n$$\n", + "text_format": "latex", + "bbox": [ + 160, + 503, + 383, + 522 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fusing 3D feature: Given feature volumes from different frames and the estimated transformation from different frames to the present frame. We transform the feature volume $V_{t}..V_{t - n + 1}$ with the corresponding transformation $T_{t}^{t}.T_{t - n + 1}^{t}$ . Particularly, for each position $p = (i,j,k)^T$ in 3D deep voxel, we transform it to $\\hat{p} = (\\hat{i},\\hat{j},\\hat{k})^T$ with:", + "bbox": [ + 75, + 531, + 468, + 625 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {p} = R p + t\n$$\n", + "text_format": "latex", + "bbox": [ + 230, + 636, + 313, + 651 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $R$ is a 3x3 rotation matrix from $T_{t}^{t}\\ldots T_{t - n + 1}^{t}$ and $t$ is the translation vector from $T_{t}^{t}\\ldots T_{t - n + 1}^{t}$ , and we note the transformation function as $f$ .", + "bbox": [ + 75, + 662, + 468, + 708 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {V} _ {i} = f \\left(V _ {i}, T _ {i} ^ {t}\\right) \\quad i = t.. t - n + 1\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 717, + 383, + 734 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since directly performing transformation in latent space might cause misalignment due to the coarse of voxel representation, we use two additional 3D conv layers to refine the transformed feature volume. We then aggregate transformed feature volumes to get Neural Volumetric Memory $\\hat{V}_t^m$ representing the surrounding 3D structure in the present frame by computing the mean over feature volumes:", + "bbox": [ + 75, + 746, + 468, + 852 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {V} _ {t} ^ {m} = \\frac {1}{n} \\sum_ {i = t - n + 1} ^ {t} \\hat {V} _ {i}\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 862, + 344, + 902 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/bd9f9f00c1227a7833e3ae53f3cef80a4864abbc2bbb66bf80e951cd3f47fd1b.jpg", + "image_caption": [ + "Figure 4. Self-Supervised Learning: we train a separate decoder to predict the visual observation in the different frames given one visual observation and the estimated transformation between two frames, similar to Lai et al. [35]." + ], + "image_footnote": [], + "bbox": [ + 501, + 89, + 897, + 210 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Learning NVM via Self-Supervision", + "text_level": 1, + "bbox": [ + 500, + 305, + 810, + 321 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Although the behavior cloning objective is sufficient in producing a good policy, being equivariant to translation and rotation automatically offers a standalone, self-supervised learning objective just for the neural volumetric memory. Shown in Figure 4, one can assume that the surrounding 3D scene remains static between frames. Because the camera is forward-looking, we can canonicalized feature volume from earlier frames, and use it to predict a camera image taken at later steps. This self-consistency loss, computed as the pixel reconstruction error, encourages the memory to be $SE(3)$ equivariant.", + "bbox": [ + 496, + 328, + 890, + 494 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For given visual observation history over a short span of time $O_{t}^{v}..O_{t - n + 1}^{v}$ , we use the most dated visual observation $O_{t - n + 1}^{v}$ in the visual history and predict all frames. We extract a 3D feature volume from $O_{t - n + 1}^{v}: V_{t - n + 1} = Enc_{3D}(O_{t - n + 1}^{v})$ , estimate the relative camera pose $T_{t}^{t - n + 1}..T_{t - n + 1}^{t - n + 1}$ between $O_{t}^{v}.O_{t - n + 1}^{v}$ and the most dated $O_{t - n + 1}^{v}$ . We transform the extracted 3D feature volume $V_{t - n + 1}$ into different frames with the corresponding transformations and then use a decoder $Dec$ to predict the visual observation $\\hat{O}_{t..t - n + 1}^{v}$ rendered in different frames. The Decoder $Dec$ takes a deep 3D feature volume as input and outputs a 2D image. The decoder first reshapes the deep 3D feature volume with shape $(C,D,H,W)$ into shape $(C*D,H,W)$ and applies a set of 2D conv layers to map the reshaped features to images. Given the predicted visual observation $\\hat{O}_{t..t - n + 1}^{v}$ and the original visual observation $O_{t}^{v}.O_{t - n + 1}^{v}$ in different frames, we compute the prediction loss with pixel-wise L1 loss. The 3D feature encoder and the pose encoder are both trained in this predictive self-supervised training task. Specifically, we can formulate this 3D representation learning task as follows:", + "bbox": [ + 496, + 494, + 892, + 814 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} V _ {t - n + 1} = \\operatorname {E n c} _ {3 D} \\left(O _ {t - n + 1} ^ {v}\\right) \\\\ \\hat {O} _ {i} ^ {v} = \\operatorname {D e c} \\left(f \\left(V _ {t - n + 1}, T _ {i} ^ {t - n + 1}\\right)\\right) i = t.. t - n + 1 \\\\ L _ {r e c} = \\frac {1}{n} \\sum_ {i = t} ^ {t - n + 1} \\left| O _ {i} ^ {v} - \\hat {O} _ {i} ^ {v} \\right| \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 820, + 875, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1433", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $L_{rec}$ represents the L1 loss. Since the information contained in different visual observations is naturally different, perfect prediction in our task is not feasible. However, such a predictive learning task could still improve the 3D structural understanding of the surrounding and encourages the SE(3) equivariance over the learned 3D latent space, for this reconstruction loss encourages the disentanglement between the surrounding 3D structure and the camera pose.", + "bbox": [ + 75, + 90, + 472, + 212 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Teacher Policy and Visuomotor Distillation", + "text_level": 1, + "bbox": [ + 76, + 220, + 444, + 237 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Image observations introduce additional overhead and slow training down as much as 40 times1. For this reason, we bootstrap learning the visuomotor policy by first training a state-space policy that depends on the elevation map.", + "bbox": [ + 75, + 244, + 470, + 306 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Learning a state-only policy We use proximal policy optimization (PPO [54]) to train a single teacher policy $\\hat{\\pi}$ in all four environments. We parameterize the teacher policy as a 4-layer ReLU network with [256, 128, 128, 64] latent neurons. To provide terrain information, we include two elevation maps, one larger and coarser than the other. Overall, the observation space for the teacher policy include:", + "bbox": [ + 75, + 324, + 468, + 431 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- proprioceptive state - $\\mathbb{R}^{63}$ A 63-dimension vector consists of the gravity vector in the body frame $\\mathbf{g}_t$ , joint rotations $\\mathbf{q}_t$ , joint velocity $\\dot{\\mathbf{q}}_t$ , and previous actions $\\mathbf{a}_{t-1}$ .", + "- privileged information - $\\mathbb{R}^{14}$ A 14-dimension vector including linear velocity $\\mathbf{v}_t$ and angular velocity $\\omega_t$ of the robot, and environment parameters $\\mathbf{e}_t$ .", + "- a dense elevation map $m_{dense} - \\mathbb{R}^{21 \\times 26}$ with 5cm spacing between the grid points, which is only slightly bigger than the robot.", + "- a sparse elevation map $m_{sparse} - \\mathbb{R}^{10 \\times 19}$ with $10\\mathrm{cm}$ spacing. This covers a much larger area." + ], + "bbox": [ + 76, + 439, + 470, + 623 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We provide a more detailed description of the observation in the appendix.", + "bbox": [ + 75, + 632, + 468, + 664 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Reward function : We use the same reward function for teacher's training in all environments. The reward contains the following terms: 1) Reward for moving forward at a target velocity (in our experiment: $0.4\\mathrm{m / s}$ ). 2) an energy penalty to encourage the robot to use as little torque as possible. 3) Height tracking reward encouraging the robot to walk while maintaining a specific relative height to the ground (in our experiment: $0.265\\mathrm{m}$ ). 4) AMP reward [49] encouraging the agent to produce natural gaits.", + "bbox": [ + 75, + 681, + 468, + 818 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} R = \\alpha_ {\\text {f o r w a r d}} * R _ {\\text {f o r w a r d}} + \\alpha_ {\\text {e n e r g y}} * R _ {\\text {e n e r g y}} \\\\ + \\alpha_ {\\text {h e i g h t}} * R _ {\\text {h e i g h t}} + \\alpha_ {\\text {a m p}} * R _ {\\text {a m p}} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 829, + 408, + 864 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/72a857023805d5200176ef101d101571b8774c8154bb1500b31a118ea30c13ed.jpg", + "table_caption": [ + "Table 1. Comparison between observation spaces. State-only policy (w/ elevation-map) vs. vision policy" + ], + "table_footnote": [], + "table_body": "
Joint statesVelocityElevation mapDepth map
Teacher policy✓✓✓×
Visual policy✓××✓
", + "bbox": [ + 503, + 119, + 883, + 195 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We include the details of each reward term and the corresponding weights in the appendix", + "bbox": [ + 498, + 212, + 890, + 244 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Visuomotor distillation: Given the privileged teacher policy from stage one, we then distill the privileged teacher policy $\\hat{\\pi}$ into the visual policy $\\pi$ without privileged information. Our visual policy utilizes proprioceptive state and visual observation for decision-making. The proprioceptive state $O_{prop}$ is the same as we used to train privileged elevation-based policy in stage one while the visual observation contains a sequence of depth images ( $O_{t}^{v}..O_{t-n+1}^{v}$ , $n=5$ in our experiment) with a shape $64 \\times 64$ from a depth camera mounted on the head of the robot. The observation space comparison between the privileged teacher policy and the visual policy is provided in Table 1, and the specific position of the camera is provided in Figure 2.", + "bbox": [ + 496, + 260, + 890, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Figure 3, our visual policy first processes the sequential visual observation with our proposed NVM module. After obtaining NVM with shape $(C,D,H,W)$ , our visual policy flattens the NVM into a 2D feature map with shape $(C\\times D,H,W)$ and utilizes a set of 2D conv layers and a single linear layer to get a 1D visual feature. Our visual policy encodes the proprioceptive state with a neural network and gets the proprioceptive feature. The proprioceptive feature and visual feature are then concatenated and fed into a final network to get the actual action for the robot.", + "bbox": [ + 496, + 455, + 890, + 607 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The visual policy is trained with behavior cloning by minimizing the L1 distance between actions from teacher policy $\\hat{\\pi}$ and visual policy $\\pi$ ,", + "bbox": [ + 496, + 607, + 890, + 654 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {B C} = | \\hat {\\pi} (p) - \\pi (o) |\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 661, + 769, + 679 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{p}$ is the privileged observation (including elevation maps), and $\\mathbf{o}$ is the observation (including visual observations). Combined with our proposed self-supervised task, the overall training loss for our NVM is given by:", + "bbox": [ + 496, + 686, + 890, + 748 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL = \\lambda_ {B C} L _ {B C} + \\lambda_ {r e c} L _ {r e c}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 756, + 785, + 771 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In our experiment, we use $\\lambda_{BC} = 1$ and $\\lambda_{rec} = 0.01$ .", + "bbox": [ + 498, + 780, + 854, + 795 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 806, + 633, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Simulation Environments", + "text_level": 1, + "bbox": [ + 500, + 830, + 733, + 847 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate NVM on four challenging terrains shown in Figure 2. Our intention is to evaluate the global and local perception ability of agents, as well as the robustness of", + "bbox": [ + 496, + 854, + 893, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "1We are able to achieve 40k fps using just elevation map. Using $64 \\times 64$ images, it slows down to 1k fps.", + "bbox": [ + 75, + 875, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1434", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/c5db4398abde028676dfb3d1fd44355e8a5299a822fdb056ba5388927f651413.jpg", + "table_caption": [ + "Table 2. Evaluation of All Policies: We evaluate all methods in all four environments and show their Traversing Rate and Success Rate. The privileged elevation-map-based policy is noted as Teacher, and its performance is regarded as the upper bound for the rest methods." + ], + "table_footnote": [], + "table_body": "
Traversing Rate (%) ↑Success Rate (%) ↑
ScenariosStagesStairsStonesObstaclesStagesStairsStonesObstacles
Teacher97.9±2.094.5±1.889.6±3.192.8±1.895.6±2.386.6±2.770.6±5.287.8±2.7
NaiveCNN76.4±26.981.5±28.629.4±10.663.3±22.553.9±19.771.3±25.40.5±0.845.2±16.4
NaiveCNN-MS72.1±25.484.5±29.726.4±9.363.7±22.748.9±18.973.1±26.20.0±0.048.1±18.3
NaiveCNN-RNN68.3±24.182.4±29.726.3±9.367.9±23.948.9±18.079.1±28.92.2±1.549.8±17.9
LocoTransformer74.0±26.585.5±29.925.2±8.965.2±22.953.3±19.375.8±26.80.9±1.146.7±16.9
NVM79.9±28.185.5±30.047.4±16.774.1±26.065.0±23.279.8±28.214.4±6.061.6±21.9
", + "bbox": [ + 81, + 127, + 888, + 268 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the learned locomotion skills. Our environments consist of Stages including multiple separate stages with gaps of different sizes between stages; Obstacles including multiple tall cuboids for the agent to avoid; Stairs including stairs with different heights and widths; Stones including numerous small stones of different shapes for the agent to step over. Stages, Stairs, and Stones environments evaluate the agent's ability to reason about the nearby surrounding and perform precise foot placement. Obstacle environment requires the agent to be able to perform long-term planning and avoid obstacles beforehand.", + "bbox": [ + 75, + 292, + 472, + 457 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Baselines", + "text_level": 1, + "bbox": [ + 76, + 465, + 187, + 481 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare our method against baselines that do not contain SE(3) equivariance priors:", + "bbox": [ + 76, + 489, + 468, + 521 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- NaiveCNN which learns visual features with a 2D CNN encoder followed by a linear projection. Visual features and proprioceptive features are concatenated. The action is then computed from the concatenated feature.", + "- NaiveCNN-RNN which shares a similar structure with NaiveCNN, but before computing action from the concatenated feature, NaiveCNN-RNN baseline utilizes a 2-layer GRU to provide memory mechanism for the agent", + "- LocoTransformer which utilizes a shared transformer model to perform cross-modal reasoning between visual tokens from 2D visual feature map and proprioceptive features, instead of directly concatenating visual feature and proprioceptive feature.", + "- Multi-Step NaiveCNN (NaiveCNN-MS) which use a 2D CNN to extract visual feature from each depth image and fuse them together to get the overall visual feature, instead of stacking multiple depth images and processing the stacked visual observation with a 2D CNN." + ], + "bbox": [ + 76, + 529, + 468, + 814 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For NaiveCNN and LocoTransformer, the visual observations in the near past are stacked in the channel dimension and serve as input for 2D encoder. All these baselines and our method share the same 2D CNN encoder structure for a fair comparison.", + "bbox": [ + 75, + 824, + 468, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Simulated Results", + "text_level": 1, + "bbox": [ + 500, + 291, + 676, + 306 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation metrics: We evaluate the final policies of all methods for 640 episodes in each simulated environment and compare the performance. We evaluate different methods by two metrics: traversing rate(\\%) and success rate(\\%). The traversing rate is defined by the distance the robot moves dividing the reachable distance (the distance between the end of the environment and starting position), and the Success Rate is defined by the ratio of the robot reaching the end of the environment.", + "bbox": [ + 496, + 318, + 893, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluated all methods including the privileged teacher policy (noted as Teacher) across all environments and the results are provided in Table 2. Since in stage two of our pipeline, visual policies are trained by imitation learning (our method with self-supervised learning as well), the performance of the privileged teacher policy can be regarded as the upper bound for the performance of visual policies. We can see that our NVM outperforms all the baselines in all the environments, especially the Stones environment which requires the agent to estimate the 3D structure of the surroundings from visual observations and provide accurate foot placement. Surprisingly, it turns out our method traversed a longer distance in the obstacle environment, where the understanding of 3D structure is not required since it's possible to avoid the obstacles with only the latest visual observation. In this case, we conjecture that the 3D understanding of the environment helps the agent to keep closer to the trajectory of the teacher policy would take since the teacher could exploit the 3D structure of the surroundings to find the optimal path.", + "bbox": [ + 496, + 458, + 893, + 760 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We find that though NaiveCNN-MS introduces the mechanism of processing each frame of visual observation separately, there is no improvement compared with NaiveCNN. This comparison indicates that the improvement brought by our NVM is actually from explicitly modeling the 3D transformation instead of introducing more computation. Compared with NaiveCNN, the memory-enabled NaiveCNN-RNN also shows no improvement, which indicates that the memory mechanism provided by RNN is not better than, if", + "bbox": [ + 496, + 763, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1435", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/5295f174b48a6623a861f649d5048f077b6bd31715a722251b72a4ae8fbaf8b9.jpg", + "table_caption": [ + "Table 3. Ablation study. We scale the grid resolution of the neural volumetric memory and the horizon of the past kept in memory. By default, our NVM uses use $D = 12$ , $H = W = 6$ , $n = 5$ ." + ], + "table_footnote": [], + "table_body": "
Traversing Rate (%) ↑Success Rate (%) ↑
ScenariosStagesStairsStonesObstaclesStagesStairsStonesObstacles
NVM79.9±28.185.5±30.047.4±16.774.1±26.065.0±23.279.8±28.214.4±6.061.6±21.9
NVM(D=3)65.4±23.768.0±25.338.6±14.546.5±17.052.8±19.452.2±19.75.9±4.136.1±13.5
NVM(D=6)79.3±27.883.1±29.249.8±17.666.2±23.664.2±23.074.7±26.414.7±6.250.2±19.3
NVM(D=4x4)72.8±26.570.1±25.939.1±14.466.4±24.054.8±21.462.8±23.711.3±6.048.6±18.4
NVM(D=8x8)78.9±27.873.4±26.341.8±15.064.5±22.961.4±22.364.8±23.912.2±4.850.0±18.3
NVM(n=3)70.1±24.876.6±26.937.6±13.364.3±23.043.9±16.867.0±23.90.6±0.848.9±18.1
NVM(n=9)78.4±27.685.9±30.139.8±14.169.8±24.661.6±22.080.0±28.55.8±3.256.1±19.9
", + "bbox": [ + 81, + 128, + 888, + 305 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "not worse, naively stacking the visual observations in the near past.", + "bbox": [ + 75, + 329, + 470, + 359 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Real World Results", + "text_level": 1, + "bbox": [ + 76, + 376, + 264, + 391 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To validate the performance of our method beyond simulation, we constructed multiple real-world experiment scenes as shown on the right of Figure 2. We evaluated our method in three scenes: Stages: we build 4 discrete stages with $15\\mathrm{cm} / 20\\mathrm{cm} / 30\\mathrm{cm}$ between stages. Stairs: We build a stair consisting of 3 ladders with heights of $3\\mathrm{cm} / 6\\mathrm{cm} / 10\\mathrm{cm}$ . Obstacles: We place multiple cuboid obstacles in the narrow hallway for the robot to avoid. We deploy the policies in real-world scenes and record the distance robot moved for each episode. We terminate the episode after a certain amount of time or until the robot reaches the end of the environment or the robot falls. We repeat the experiment for each method and environment five times. Real-world trajectories are also provided in Figure 1. Since NaiveCNN, NaiveCNN-RNN, LocoTransformer show similar performance in simulation and LocoTransformer is known to be more robust in unseen scenarios and objects in the real world, we compared our NVM with LocoTransformer in these three scenarios. The quantitative results are shown in Table 4. We find that our NVM moved significantly further in Stages and Stairs while slightly further in Obstacles. This is because Stages and Stairs environment require the robot to understand the surrounding 3D structure to place the foot while Obstacles only requires the ability to perform high-level planning with visual observation. We observe that in Stages and Stairs environment, agents trained with LocoTransformer can place the front leg across the stage or over the ladder, but not the rear legs when the gap between stages is large or the ladder is tall, which demonstrates that lack of understanding of the surrounding 3D structure. We also found that even though LocoTransformer agent could move a reasonable distance in simulation in Stairs and Stages environment, it fails in", + "bbox": [ + 73, + 401, + 473, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/bd7b19c92e76b35233b6d023b311f186b0ca8dcd53c0f8174b7cb466e763bf2c.jpg", + "table_caption": [ + "Table 4. Evaluating the visual policy in the real world" + ], + "table_footnote": [], + "table_body": "
Distance Moved (m) ↑
ScenariosStagesStairsObstacles
NVM5.4±0.34.0±0.07.3±0.1
LocoTransformer3.4±0.12.1±0.86.7±0.4
", + "bbox": [ + 511, + 352, + 877, + 436 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the real world. The real-world results indicate that the understanding of the surrounding 3D structures is essential for legged robots to traverse complex terrains in the real world and our NVM module provides a high-quality volumetric memory of the surrounding 3D structure.", + "bbox": [ + 496, + 454, + 893, + 531 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Visualization of SSL task", + "text_level": 1, + "bbox": [ + 498, + 542, + 730, + 559 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Aiming to gather insight into our learned 3D features volume, we visualized the synthesized visual observation $\\hat{O}_t^v$ in our self-supervised learning task along with the present visual observation $O_t^v$ (the ground truth in our SSL task), and most dated visual observation $O_{t - n + 1}^v$ (the input in our SSL task) in Figure 5. Even though we apply extensive data augmentation over the raw visual observation (the second image for every case), the synthesized visual observations are clearly indicating the robustness of our model. In the synthesized visual observation for $Stair$ and $Obstacle$ environment, we can see the overall structure of the environment is well-preserved. In Figure 5b, The obstacles in the previous frame are transformed to the position in the current frame without much distortion. In Figure 5c, the boundary of the nearest ladder is clear, and the second nearest ladder is also observable in the present frame. We have to note that the reason the decoder can not recover the details away from the robot is in the input for SSL the large values are clipped. For more complex Stages and Stones environments, we found that even though the terrain details are not perfectly reconstructed, the essential regions for decision-making are reconstructed which means the learned 3D voxel", + "bbox": [ + 496, + 568, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1436", + "bbox": [ + 483, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5cdfdfdfff859ff42c5df33b8d3d43adbb51e7b3e89254457d7a3cdd9984a7a4.jpg", + "image_caption": [ + "(a) Stages" + ], + "image_footnote": [], + "bbox": [ + 78, + 92, + 480, + 183 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e6f2145722811db0516d8f3f33dbafa54e9063633b284bd2bdfbca7eba5131c7.jpg", + "image_caption": [ + "(b) Obstacles" + ], + "image_footnote": [], + "bbox": [ + 490, + 92, + 893, + 183 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fcddc846ab074ba48db9a679d83d06568e909c217d6c7d1e69123936a3388ed5.jpg", + "image_caption": [ + "(c) Stairs" + ], + "image_footnote": [], + "bbox": [ + 78, + 202, + 480, + 292 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1a8f9e1f53f7f204c7ba3a870cc35854f92a0c7b407d9724dd94a17c5becf98c.jpg", + "image_caption": [ + "(d) Stones", + "Figure 5. Visual reconstruction from learned decoder. We visualize the synthesized visual observation in our self-supervised task. For every tuple, the first image shows the robot moving in the environment, the second image is the input visual observation $O_{t - n + 1}^v$ , the third image is the synthesized visual observation using 3D feature volume $V_{t - n + 1}$ extracted from $O_{t - n + 1}^v$ and the estimated relative camera between $O_{t - n + 1}^v$ and robot's ego-centric frame, and the fourth image is the actual visual observation captured by the robot (the ground-truth in our self-supervised task). For the input visual observation, we apply extensive data augmentation to the image to improve the robustness of our model." + ], + "image_footnote": [], + "bbox": [ + 490, + 202, + 893, + 292 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "features capture the most important parts in the visual observations. In Figure 5a, we can see the gaps between stages nearby are well reconstructed and in Figure 5d, we can see the reconstructed stone in front of the robot.", + "bbox": [ + 75, + 415, + 468, + 474 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 484, + 228, + 500 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "we perform multiple sets of ablation studies, and the results are shown in Table 3, with the purpose of studying the influence of different components in our method.", + "bbox": [ + 75, + 508, + 468, + 551 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The Resolution of NVM: We provide our NVM module voxel features with different resolutions by slightly changing the structure of the 2D encoder, which resulting 2D feature maps in different shapes and use different depth parameter $D$ . From Table 3, we find that the larger depth value $D$ generally provides better performance, since larger $D$ converts the 2D feature map into more fine-grained voxel feature which results in fine-grained neural volumetric memory, and it's supposed to be easier for the agent to reason about the foot placement over complex terrain with more fine-grained neural volumetric memory. For $H$ and $W$ , we found that $6x6$ which provides enough detail horizontally without introducing much noise vertically works best.", + "bbox": [ + 75, + 551, + 468, + 750 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The Length of Visual History of NVM: We provide our NVM model with different lengths of visual history by changing the number of stored visual observation frames. As shown in Table 3, the performance of $NVM$ improves when the length of visual history increases to 5 from 3, we think this is because the longer visual history provides longer 3D memory of the surrounding for the agent to safely placing the rear feet. But when the length of visual history increases to 9 from 5, the performance drops, we hypothesize that it is because when the visual history becomes too", + "bbox": [ + 75, + 750, + 468, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "long, it's hard for the pose encoder to estimate the accurate transformation between the oldest frame and the current frame, and the surplus visual information might be redundant for decision-making over terrains.", + "bbox": [ + 496, + 415, + 892, + 474 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 493, + 617, + 508 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose a short-term memory mechanism, the Neural Volumetric Memory (NVM), to enable legged robots to traverse complex terrains with ease. NVM aggregates knowledge from sequential visual observation in the near past by explicitly modeling the translation and rotation between previous frames and the present frame. The explicit modeling of the transformation and self-supervised task also encourage the $SE(3)$ equivariance in the memory space. Our neural volumetric memory module enables the legged robot to traverse complex terrain in simulation and the real world and achieve better sim-to-real generalization results. This shows our neural volumetric memory provides reliable surrounding 3D structure information for the robot to perform decision-making and new possibility of utilizing visual observation for legged-robot.", + "bbox": [ + 496, + 521, + 892, + 748 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This project was supported, in part, by NSF CCF-2112665 (TILOS), NSF CAREER Award IIS-2240014, NSF 1730158 CI-New: Cognitive Hardware and Software Ecosystem Community Infrastructure (CHASE-CI), NSF ACI-1541349 CC*DNI Pacific Research Platform, Amazon Research Award and gifts from Qualcomm. Ge Yang is supported by the National Science Foundation Institute for Artificial Intelligence and Fundamental Interactions (IAIFI, https://iaifi.org/) under the Cooperative Agreement PHY-2019786. We thank Yuxiang Yang for providing help regarding hardware setup.", + "bbox": [ + 496, + 750, + 893, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1437", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Ananye Agarwal, Ashish Kumar, Jitendra Malik, and Deepak Pathak. Legged locomotion in challenging terrains using egocentric vision. In 6th Annual Conference on Robot Learning, 2022. 3", + "[2] Gerardo Bledt and Sangbae Kim. Extracting legged locomotion heuristics with regularized predictive control. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 406-412. IEEE, 2020. 2", + "[3] Gerardo Bledt, Matthew J Powell, Benjamin Katz, Jared Di Carlo, Patrick M Wensing, and Sangbae Kim. Mit cheetah 3: Design and control of a robust, dynamic quadruped robot. In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 2245–2252. IEEE, 2018. 2", + "[4] Jan Carius, René Ranftl, Vladlen Koltun, and Marco Hutter. Trajectory optimization for legged robots with slipping motions. IEEE Robotics and Automation Letters, 4(3):3013-3020, 2019. 2", + "[5] Jared Di Carlo, Patrick M. Wensing, Benjamin Katz, Gerardo Bledt, and Sangbae Kim. Dynamic locomotion in the MIT cheetah 3 through convex model-predictive control. In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems, IROS 2018, Madrid, Spain, October 1-5, 2018, pages 1-9. IEEE, 2018. 2", + "[6] Dian Chen, Brady Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Learning by cheating. In Conference on Robot Learning, pages 66-75. PMLR, 2020. 3", + "[7] Tao Chen, Jie Xu, and Pulkit Agrawal. A system for general in-hand object re-orientation. In Conference on Robot Learning, pages 297–307. PMLR, 2022. 3", + "[8] Annett Chilian and Heiko Hirschmüller. Stereo camera based navigation of mobile robots on rough terrain. In 2009 IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 4571-4576. IEEE, 2009. 2", + "[9] Jared Di Carlo, Patrick M Wensing, Benjamin Katz, Gerardo Bledt, and Sangbae Kim. Dynamic locomotion in the mit cheetah 3 through convex model-predictive control. In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1-9. IEEE, 2018. 2", + "[10] Yanran Ding, Abhishek Panda, and Hae-Won Park. Realtime model predictive control for versatile dynamic motions in quadrupedal robots. In 2019 International Conference on Robotics and Automation (ICRA), pages 8484-8490. IEEE, 2019. 2", + "[11] Peter Fankhauser, Marko Bjelonic, C. Dario Bellicoso, Takahiro Miki, and Marco Hutter. Robust rough-terrain locomotion with a quadrupedal robot. In 2018 IEEE International Conference on Robotics and Automation (ICRA), pages 5761-5768, 2018. 2", + "[12] Péter Fankhauser, Michael Bloesch, Christian Gehring, Marco Hutter, and Roland Siegwart. Robot-centric elevation mapping with uncertainty estimates. In International Conference on Climbing and Walking Robots (CLAWAR), 2014. 2", + "[13] Péter Fankhauser, Michael Bloesch, and Marco Hutter. Probabilistic terrain mapping for mobile robots with unclear" + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tain localization. IEEE Robotics and Automation Letters, 3(4):3019-3026, 2018. 2", + "[14] Hsiao-Yu Fish Tung, Ricson Cheng, and Katerina Fragkiadaki. Learning spatial common sense with geometry-aware recurrent networks. In cvpr. 3", + "[15] Christian Gehring, Stelian Coros, Marco Hutter, Michael Blösch, Mark A. Hoepflinger, and Roland Siegwart. Control of dynamic gaits for a quadrupedal robot. In 2013 IEEE International Conference on Robotics and Automation, Karlsruhe, Germany, May 6-10, 2013, pages 3287-3292. IEEE, 2013. 2", + "[16] Hartmut Geyer, Andre Seyfarth, and Reinhard Blickhan. Positive force feedback in bouncing gaits? Proceedings of the Royal Society of London. Series B: Biological Sciences, 270(1529):2173-2183, 2003. 2", + "[17] Ruben Grandia, Farbod Farshidian, Alexey Dosovitskiy, René Ranftl, and Marco Hutter. Frequency-aware model predictive control. IEEE Robotics and Automation Letters, 4(2):1517-1524, 2019. 2", + "[18] Adam W Harley, Shrinidhi Kowshika Lakshmikanth, Paul Schydlo, and Katerina Fragkiadaki. Tracking emerges by looking around static scenes, with neural 3d mapping. In eccv, 2020. 3", + "[19] Adam W. Harley, Fangyu Li, Shrinidhi K. Lakshmikanth, Xian Zhou, Hsiao-Yu Fish Tung, and Katerina Fragkiadaki. Learning from unlabelled videos using contrastive predictive neural 3d mapping. In ICLR, 2019. 3", + "[20] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. Cambridge university press, 2003. 3", + "[21] Ronghang Hu, Nikhila Ravi, Alexander C. Berg, and Deepak Pathak. Worldsheet: Wrapping the world in a 3d sheet for view synthesis from a single image. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2021. 3", + "[22] Jemin Hwangbo, J. Lee, A. Dosovitskiy, Dario Bellicoso, Vassilios Tsounis, V. Koltun, and M. Hutter. Learning agile and dynamic motor skills for legged robots. Science Robotics, 4, 2019. 2", + "[23] Chieko Imai, Minghao Zhang, Yuchen Zhang, Marcin Kierebinski, Ruihan Yang, Yuzhe Qin, and Xiaolong Wang. Vision-guided quadrupedal locomotion in the wild with multi-modal delay randomization. In 2022 IEEE/RSJ international conference on intelligent robots and systems (IROS). 3", + "[24] Atil Iscen, Ken Caluwaerts, Jie Tan, Tingnan Zhang, Erwin Coumans, Vikas Sindhwani, and Vincent Vanhoucke. Policies modulating trajectory generators. In 2nd Annual Conference on Robot Learning, CoRL 2018, Zürich, Switzerland, 29-31 October 2018, Proceedings, volume 87 of Proceedings of Machine Learning Research, pages 916-926. PMLR, 2018. 2", + "[25] Deepali Jain, Atil Iscen, and Ken Caluwaerts. Hierarchical reinforcement learning for quadruped locomotion. In 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems, IROS 2019, Macau, SAR, China, November 3-8, 2019, pages 7551-7557. IEEE, 2019. 2" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "1438", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Deepali Jain, Atil Iscen, and Ken Caluwaerts. From pixels to legs: Hierarchical learning of quadruped locomotion. arXiv preprint arXiv:2011.11722, 2020. 2", + "[27] Joel Janai, Fatma Güney, Aseem Behl, Andreas Geiger, et al. Computer vision for autonomous vehicles: Problems, datasets and state of the art. Foundations and Trends® in Computer Graphics and Vision, 12(1-3):1-308, 2020. 3", + "[28] Danilo Jimenez Rezende, SM Eslami, Shakir Mohamed, Peter Battaglia, Max Jaderberg, and Nicolas Heess. Unsupervised learning of 3d structure from images. Advances in neural information processing systems, 29:4996-5004, 2016. 3", + "[29] Angjoo Kanazawa, Shubham Tulsiani, Alexei A Efros, and Jitendra Malik. Learning category-specific mesh reconstruction from image collections. In ECCV, 2018. 3", + "[30] Elia Kaufmann, Antonio Loquercio, Rene Ranftl, Alexey Dosovitskiy, Vladlen Koltun, and Davide Scaramuzza. Deep drone racing: Learning agile flight in dynamic environments. In Conference on Robot Learning, pages 133-145. PMLR, 2018. 3", + "[31] Elia Kaufmann, Antonio Loquercio, René Ranftl, Matthias Müller, Vladlen Koltun, and Davide Scaramuzza. Deep drone acrobatics. arXiv preprint arXiv:2006.05768, 2020. 3", + "[32] Nate Kohl and Peter Stone. Policy gradient reinforcement learning for fast quadrupedal locomotion. In Proceedings of the 2004 IEEE International Conference on Robotics and Automation, ICRA 2004, April 26 - May 1, 2004, New Orleans, LA, USA, pages 2619-2624. IEEE, 2004. 2", + "[33] Ashish Kumar, Zipeng Fu, Deepak Pathak, and Jitendra Malik. Rma: Rapid motor adaptation for legged robot. Robotics: Science and Systems, 2021. 1, 2", + "[34] In-So Kweon and Takeo Kanade. High-resolution terrain map from multiple sensor data. IEEE Transactions on Pattern Analysis and Machine Intelligence, 14(2):278-292, 1992. 2", + "[35] Zihang Lai, Sifei Liu, Alexei A Efros, and Xiaolong Wang. Video autoencoder: self-supervised disentanglement of 3d structure and motion. In ICCV, 2021. 3, 4", + "[36] Joonho Lee, Jemin Hwangbo, Lorenz Wellhausen, Vladlen Koltun, and Marco Hutter. Learning quadrupedal locomotion over challenging terrain. Science robotics, 5(47), 2020. 1, 2", + "[37] Antonio Loquercio, Ashish Kumar, and Jitendra Malik. Learning visual locomotion with cross-modal supervision, 2022. 3", + "[38] G. Margolis, Ge Yang, Kartik Paigwar, Tao Chen, and Pulkit Agrawal. Rapid locomotion via reinforcement learning. ArXiv, abs/2205.02824, 2022. 1", + "[39] Gabriel B Margolis, Tao Chen, Kartik Paigwar, Xiang Fu, Donghyun Kim, Sang bae Kim, and Pulkit Agrawal. Learning to jump from pixels. In 5th Annual Conference on Robot Learning, 2021. 2, 3", + "[40] Takahiro Miki, Joonho Lee, Jemin Hwangbo, Lorenz Wellhausen, Vladlen Koltun, and Marco Hutter. Learning robust perceptive locomotion for quadrupedal robots in the wild. Sci Robot, 7(62):eabk2822, Jan. 2022. 2, 3", + "[41] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf:" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 3", + "[42] Hirofumi Miura and Isao Shimoyama. Dynamic walk of a biped. The International Journal of Robotics Research, 3(2):60-74, 1984. 2", + "[43] Tongzhou Mu, Zhan Ling, Fanbo Xiang, Derek Yang, Xu-anlin Li, Stone Tao, Zhiao Huang, Zhiwei Jia, and Hao Su. Maniskill: Generalizable manipulation skill benchmark with large-scale demonstrations. arXiv preprint arXiv:2107.14483, 2021. 3", + "[44] Siva Karthik Mustikovela, Varun Jampani, Shalini De Mello, Sifei Liu, Umar Iqbal, Carsten Rother, and Jan Kautz. Self-supervised viewpoint learning from image collections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3971-3981, 2020. 3", + "[45] Thu Nguyen-Phuoc, Chuan Li, Lucas Theis, Christian Richardt, and Yong-Liang Yang. Hologan: Unsupervised learning of 3d representations from natural images. In Proceedings of the IEEE International Conference on Computer Vision, pages 7588–7597, 2019. 3", + "[46] Yiyuan Pan, Xuecheng Xu, Yue Wang, Xiaqing Ding, and Rong Xiong. GPU accelerated real-time traversability mapping. In 2019 IEEE International Conference on Robotics and Biomimetics (ROBIO), pages 734-740. IEEE, 2019. 2", + "[47] Xue Bin Peng, Glen Berseth, KangKang Yin, and Michiel Van De Panne. Deeploco: Dynamic locomotion skills using hierarchical deep reinforcement learning. ACM Transactions on Graphics (TOG), 36(4):1-13, 2017. 2", + "[48] Xue Bin Peng, Erwin Coumans, Tingnan Zhang, Tsang-Wei Edward Lee, Jie Tan, and Sergey Levine. Learning agile robotic locomotion skills by imitating animals. In Robotics: Science and Systems, 07 2020. 2", + "[49] Xue Bin Peng, Ze Ma, Pieter Abbeel, Sergey Levine, and Angjoo Kanazawa. Amp: Adversarial motion priors for stylized physics-based character control. ACM Trans. Graph., 40(4), July 2021. 5", + "[50] Lerrel Pinto, Marcin Andrychowicz, Peter Welinder, Wojciech Zaremba, and Pieter Abbeel. Asymmetric actor critic for image-based robot learning. arXiv preprint arXiv:1710.06542, 2017. 3", + "[51] Marc H Raibert. Hopping in legged systems—modeling and simulation for the two-dimensional one-legged case. IEEE Transactions on Systems, Man, and Cybernetics, SMC-14(3):451–463, 1984. 2", + "[52] Chris Rockwell, David F. Fouhey, and Justin Johnson. Pixelsynth: Generating a 3d-consistent experience from a single image. In ICCV, 2021. 3", + "[53] Nikita Rudin, David Hoeller, Philipp Reist, and Marco Hutter. Learning to walk in minutes using massively parallel deep reinforcement learning. In Conference on Robot Learning, pages 91-100. PMLR, 2022. 2", + "[54] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 5", + "[55] Mingyo Seo, Ryan Gupta, Yifeng Zhu, Alexy Skoutnev, Luis Sentis, and Yuke Zhu. Learning to walk by steering: Perceptive quadrupedal locomotion in dynamic environments. In arXiv preprint arXiv:2209.09233, 2022. 2" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "1439", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[56] Vincent Sitzmann, Justus Thies, Felix Heide, Matthias Nießner, Gordon Wetzstein, and Michael Zollhofer. Deepvoxels: Learning persistent 3d feature embeddings. In CVPR, 2019. 3", + "[57] Vincent Sitzmann, Michael Zollhöfer, and Gordon Wetzstein. Scene representation networks: Continuous 3d-structure-aware neural scene representations. In NeurIPS, 2019. 3", + "[58] Maks Sorokin, Jie Tan, C. Karen Liu, and Sehoon Ha. Learning to navigate sidewalks in outdoor environments. IEEE Robotics and Automation Letters, 7(2):3906-3913, 2022. 2", + "[59] Yu Sun, Wyatt L. Ubellacker, Wen-Loong Ma, Xiang Zhang, Changhao Wang, Noel V. Csomay-Shanklin, Masayoshi Tomizuka, Koushil Sreenath, and Aaron D. Ames. Online learning of unknown dynamics for model-based controllers in legged locomotion. IEEE Robotics and Automation Letters (RA-L), 2021. 2", + "[60] Jie Tan, Tingnan Zhang, Erwin Coumans, Atil Iscen, Yunfei Bai, Danijar Hafner, Steven Bohez, and Vincent Vanhoucke. Sim-to-real: Learning agile locomotion for quadruped robots. In Hadas Kress-Gazit, Siddhartha S. Srinivasa, Tom Howard, and Nikolay Atanasov, editors, Robotics: Science and Systems XIV, Carnegie Mellon University, Pittsburgh, Pennsylvania, USA, June 26-30, 2018, 2018. 2", + "[61] Nick Torkos and Michiel van de Panne. Footprint-based quadruped motion synthesis. In Proceedings of the Graphics Interface 1998 Conference, June 18-20, 1998, Vancouver, BC, Canada, pages 151-160, June 1998. 2", + "[62] Shubham Tulsiani, Tinghui Zhou, Alexei A Efros, and Jitendra Malik. Multi-view supervision for single-view reconstruction via differentiable ray consistency. In CVPR, 2017. 3", + "[63] Martin Wermelinger, Péter Fankhauser, Remo Diethelm, Philipp Krüsi, Roland Siegwart, and Marco Hutter. Navigation planning for legged robots in challenging terrain. In 2016 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1184-1189. IEEE, 2016. 2", + "[64] Olivia Wiles, Georgia Gkioxari, Richard Szeliski, and Justin Johnson. Synsin: End-to-end view synthesis from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7467-7477, 2020. 3", + "[65] Zhaoming Xie, Xingye Da, Buck Babich, Animesh Garg, and Michiel van de Panne. Glide: Generalizable quadrupedal locomotion in diverse environments with a centroidal model. CoRR, abs/2104.09771, 2021. 2", + "[66] Zhaoming Xie, Hung Yu Ling, Nam Hee Kim, and Michiel van de Panne. Allsteps: curriculum-driven learning of stepping stone skills. In Computer Graphics Forum, volume 39, pages 213-224. Wiley Online Library, 2020. 2", + "[67] Ruihan Yang, Minghao Zhang, Nicklas Hansen, Huazhe Xu, and Xiaolong Wang. Learning vision-guided quadrupedal locomotion end-to-end with cross-modal transformers. In International Conference on Learning Representations, 2022. 3" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[68] KangKang Yin, Kevin Loken, and Michiel Van de Panne. Simbicon: Simple biped locomotion control. ACM Transactions on Graphics (TOG), 26(3):105-es, 2007. 2", + "[69] Wenhao Yu, Deepali Jain, Alejandro Escontrela, Atil Iscen, Peng Xu, Erwin Coumans, Sehoon Ha, Jie Tan, and Tingnan Zhang. Visual-locomotion: Learning to walk on complex terrains with vision. In Aleksandra Faust, David Hsu, and Gerhard Neumann, editors, Proceedings of the 5th Conference on Robot Learning, volume 164 of Proceedings of Machine Learning Research, pages 1291-1302. PMLR, 08-11 Nov 2022. 2", + "[70] Tinghui Zhou, Matthew Brown, Noah Snavely, and David G Lowe. Unsupervised learning of depth and ego-motion from video. In CVPR, 2017. 3", + "[71] Matt Zucker, Nathan Ratliff, Martin Stolle, Joel Chestnutt, J Andrew Bagnell, Christopher G Atkeson, and James Kuffner. Optimization and learning for rough terrain legged locomotion. The International Journal of Robotics Research, 30(2):175-191, 2011. 2" + ], + "bbox": [ + 501, + 92, + 890, + 358 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "1440", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Neural Volumetric Memory for Visual Locomotion Control/dad36583-d1e2-4d95-a7b6-551afd512667_model.json b/2023/Neural Volumetric Memory for Visual Locomotion Control/dad36583-d1e2-4d95-a7b6-551afd512667_model.json new file mode 100644 index 0000000000000000000000000000000000000000..30bcbd4eb5357a7f7bd36cf3cceb12275e7b0873 --- /dev/null +++ b/2023/Neural Volumetric Memory for Visual Locomotion Control/dad36583-d1e2-4d95-a7b6-551afd512667_model.json @@ -0,0 +1,2356 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.187, + 0.118, + 0.783, + 0.14 + ], + "angle": 0, + "content": "Neural Volumetric Memory for Visual Locomotion Control" + }, + { + "type": "text", + "bbox": [ + 0.293, + 0.168, + 0.662, + 0.187 + ], + "angle": 0, + "content": "Ruihan Yang1 Ge Yang2,3 Xiaolong Wang1" + }, + { + "type": "text", + "bbox": [ + 0.18, + 0.187, + 0.786, + 0.205 + ], + "angle": 0, + "content": "\\(^{1}\\)UC San Diego \\(^{2}\\)Institute of AI and Fundamental Interactions \\(^{3}\\)MIT CSAIL" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.216, + 0.893, + 0.413 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.424, + 0.893, + 0.454 + ], + "angle": 0, + "content": "Figure 1. Real-World Deployment: We deploy our policy in real-world environments with Stones, Stairs, Stages, Obstacles, and Unstructured Terrain. Our robot utilizes the volumetric memory of the surrounding 3D structure to successfully traverse complex environments." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.478, + 0.314, + 0.495 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.511, + 0.473, + 0.829 + ], + "angle": 0, + "content": "Legged robots have the potential to expand the reach of autonomy beyond paved roads. In this work, we consider the difficult problem of locomotion on challenging terrains using a single forward-facing depth camera. Due to the partial observability of the problem, the robot has to rely on past observations to infer the terrain currently beneath it. To solve this problem, we follow the paradigm in computer vision that explicitly models the 3D geometry of the scene and propose Neural Volumetric Memory (NVM), a geometric memory architecture that explicitly accounts for the SE(3) equivariance of the 3D world. NVM aggregates feature volumes from multiple camera views by first bringing them back to the ego-centric frame of the robot. We test the learned visual-locomotion policy on a physical robot and show that our approach, which explicitly introduces geometric priors during training, offers superior performance than more naive methods. We also include ablation studies and show that the representations stored in the neural volumetric memory capture sufficient geometric information to reconstruct the scene. Our project page with videos is https://rchalyang.github.io/NVM" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.844, + 0.21, + 0.86 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Consider difficult locomotion tasks such as walking up and down a flight of stairs and stepping over gaps (Figure 1)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.479, + 0.895, + 0.678 + ], + "angle": 0, + "content": "The control of such behaviors requires tight coupling with perception because vision is needed to provide details of the terrain right beneath the robot and the 3D scene immediately around it. This problem is also partially-observable. Immediately relevant terrain information is often occluded from the robot's current frame of observation, forcing it to rely on past observations for control decisions. For this reason, while blind controllers that are learned in simulation using reinforcement learning have achieved impressive results in agility and robustness [33, 36, 38], there are clear limitations on how much they can do. How to incorporate perception into the pipeline to produce an integrated visuomotor controller thus remains an open problem." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.896, + 0.903 + ], + "angle": 0, + "content": "A recent line of work combines perception with locomotion using ego-centric cameras mounted on the robot. The predominant approach for addressing partial observability is to do frame-stacking, where the robot maintains a visual buffer of recent images. This naive heuristic suffers from two major problems: first, frame-stacking on a moving robot ignores the equivariance structure of the 3D environment, making learning a lot more difficult as policy success now relies on being able to learn to account for spurious changes in camera poses. A second but subtler issue is that biological systems do not have the ability to save detailed visual observations pixel-by-pixel. These concerns motivate the creation of an intermediary, short-term memory mechanism to functionally aggregate streams of obser" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1430" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.088, + 0.895, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.25, + 0.895, + 0.294 + ], + "angle": 0, + "content": "Figure 2. Overview of Simulated Environment & Real World Environment: Our simulated environments are shown on the left and real-world environments are shown on the right. For the real-world environment, the corresponding visual observations for each real-world environment are shown in the bottom row. All policies are trained in the simulation and transferred into the real world without fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.305, + 0.457, + 0.32 + ], + "angle": 0, + "content": "vation into a single, coherent representation of the world." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.323, + 0.473, + 0.581 + ], + "angle": 0, + "content": "Motivated by these observations, we introduce a novel volumetric memory architecture for legged locomotion control. Our architecture consists of a 2D to 3D feature volume encoder, and a pose estimator that can estimate the relative camera poses given two input images. When combined, the two networks constitute a Neural Volumetric Memory (NVM) that takes as input a sequence of depth images taken by a forward-looking, ego-centric camera and fuses them together into a coherent latent representation for locomotion control. We encourage the memory space to be \\( SE(3) \\) equivariant to changes in the camera pose of the robot by incorporating translation and rotation operations based on estimated relative poses from the pose network. This inverse transformation allows NVM to align feature volumes from the past to the present ego-centric frame, making both integrating over multiple timesteps into a coherent scene representation and learning a policy, less difficult." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.585, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Our training pipeline follows a two-step teacher-student process where the primary goal of the first stage is to produce behaviors in the form of a policy. After training completes, this policy can traverse these difficult terrains robustly, but it relies on privileged sensory information such as an elevation map and ground-truth velocity. Elevation maps obtained in the real world are often biased, incomplete, and full of errors [40], whereas ground-truth velocity information is typically only available in instrumented environments. Hence in the visuomotor distillation stage of the pipeline, which still runs in the simulator, we feed the stream of ego-centric views from the forward depth camera into the neural volumetric memory. We feed the content of this memory into a small policy network and train everything end-to-end including the two network components of the NVM using a behavior cloning loss where the state-only policy acts as the teacher. For completeness, we offer an additional self-supervised learning objective (Figure 4) that relies on novel-view consistency for learning. The end product of this visuomotor distillation pipeline is a memory-equipped visuomotor policy that can operate directly on the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.305, + 0.895, + 0.396 + ], + "angle": 0, + "content": "UniTree A1 robot hardware (see Figure 1). A single policy is used to handle all environments covered by this paper. We provide comprehensive experiment and ablation studies in both simulation and the real world, and show that our method outperforms all baselines by a large margin. It is thus essential to model the 3D structure of the environment." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.412, + 0.642, + 0.428 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.438, + 0.895, + 0.697 + ], + "angle": 0, + "content": "Blind Locomotion Controllers: Legged locomotion using the proprioception of the robot has been studied for decades [3, 16, 42, 51, 61, 68]. Traditionally, controllers are developed with model-based methods [2, 4, 5, 9, 10, 15, 17, 59]. However, most model-based methods fail to generalize to unseen environments in the real world. In order to attain better generalization and robustness, recent works [22, 24, 25, 32, 33, 36, 48, 60, 65] utilize model-free reinforcement learning to train neural network controllers in simulation and directly transfer the learned policies to the real robot. These methods are able to achieve high levels of agility and adapt over diverse terrains using just the proprioceptive state of the robot. The real-world utility of these robots, however, is also limited, since these blind controllers are unable to perceive the surrounding. In this paper, we propose to integrate perception to facilitate control, with a focus on a better 3D understanding of the environments." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.897, + 0.903 + ], + "angle": 0, + "content": "Locomotion with Exteroceptive Perception: The elevation map [12, 13, 34, 46] around the robot has been used to perform optimization and planning for foot placement [8, 11, 47, 63, 66, 71]. Recent RL controller [40, 53] also uses elevation maps in an end-to-end manner. Using the elevation map to represent the surrounding is straightforward for foot placement optimization, but the elevation map obtained in the real world is either noisy or requires a complicated hardware pipeline. Another option is using depth or RGB images to provide visual observation for the controller. Hierarchical control systems [26, 39, 55, 58, 69] are proposed to utilize depth reading while the low-level" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "1431" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.079, + 0.088, + 0.898, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.276, + 0.895, + 0.32 + ], + "angle": 0, + "content": "Figure 3. Neural Volumetric Memory (NVM): Our NVM module extracts 3D feature volumes from every visual observation in the near past, transform them with corresponding estimated transformation into present robot frame, and fuse the transformed feature volumes into a neural volumetric memory for decision-making." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.335, + 0.473, + 0.519 + ], + "angle": 0, + "content": "controllers still rely on only the proprioceptive state of the robot. To get the most out of the depth image, Yang et al. [67] and Imai et al. [23] propose to learn visual RL policy to output the raw joint angles and enable the robot to maneuver in the wild avoiding different kinds of obstacles. Agarwal et al. [1] and Loquercioo et al. [37] train a policy using RL to traverse a large variety of terrains. Though these works show encouraging results in specific tasks, they have not fully exploited the 3D structure of the environment. On the other hand, we propose to train a novel neural volumetric memory together with policy learning, which allows our robot to walk through more challenging terrains." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.535, + 0.473, + 0.765 + ], + "angle": 0, + "content": "Volumetric Cognitive Map: How to represent and understand the 3D scene has been studied by the computer vision community for a long time. Hartley et al. [20] and Wiles et al. [64] provide representation built on geometry principles and camera motions. Learning-based approaches are proposed [21, 28, 29, 35, 52, 62, 64, 70] to learn 3D representation from 2D inputs. Neural 3D representations [14, 19, 41, 44, 45, 56, 57] has also shown great result in multiple computer vision tasks [14, 18, 19]. Our work is closely related to Lai et al. [35] where the deep voxel feature representing the entire scene and the camera trajectory are learned from video clips without odometry in a self-supervised manner. Instead of focusing on view synthesis, we aggregate the 3D feature volume to form a short-term memory representation for decision-making." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.473, + 0.904 + ], + "angle": 0, + "content": "Teacher-Student Training: To accelerate training and utilize the privileged information in simulation, Miki et al. [40] follow a teacher-student training framework, where the teacher policy is trained with privileged information and then distilled into student policy where the privileged information is not available. Multiple works in different research domains [6, 7, 27, 30, 31, 43, 50] follow this framework as well. We share a similar pipeline as [1, 39] where elevation" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.335, + 0.892, + 0.368 + ], + "angle": 0, + "content": "map-based policies are trained first and the elevation-map-based policies are distilled visual policies afterward." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.381, + 0.892, + 0.4 + ], + "angle": 0, + "content": "3. Volumetric Memory for Legged Locomotion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.408, + 0.892, + 0.528 + ], + "angle": 0, + "content": "Legged locomotion using ego-centric camera views is intrinsically a partially-observed problem. To make the control problem tangible, our robot needs to aggregate information from previous frames and correctly infer the occluded terrain underneath it. During locomotion, the camera mounted directly on the robot chassis undergo large and spurious changes in pose, making integrating individual frames to a coherent representation non-trivial." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.53, + 0.894, + 0.605 + ], + "angle": 0, + "content": "To account for these camera pose changes, we propose neural volumetric memory (NVM) — a 3D representation format for scene features. It takes as input a sequence of visual observations and outputs a single 3D feature volume representing the surrounding 3D structure." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.617, + 0.842, + 0.634 + ], + "angle": 0, + "content": "3.1. Baking-in \\(SE(3)\\) Equivariance in NVM" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.642, + 0.892, + 0.702 + ], + "angle": 0, + "content": "Our training procedure encourages the learned memory space to be \\(SE(3)\\) equivariant by applying translation and rotation transformations on the feature volume before aggregating additional information from new observations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.703, + 0.895, + 0.903 + ], + "angle": 0, + "content": "As the first step (Figure 3), an encoder network receives a sequence of consecutive observations \\( O_{t}..O_{t - n + 1} \\). Instead of generating a compact feature embedding in the shape of a tuple, the 3D encoder Enc3D extracts feature volumes \\( V_{t}.V_{t - n + 1} \\). A second network Encpose estimates the relative changes in camera pose between the present frame \\( O_{t}^{v} \\) and frames \\( O_{t}^{v}.O_{t - n + 1}^{v} \\) from the past. This sequence of estimate allows NVM to transform the feature volume \\( V_{t}.V_{t - n + 1} \\) from the past to the present frame using the transformations \\( T_{t}^{t}.T_{t - n + 1}^{t} \\). We annotate the transformed feature volume as \\( \\hat{V}_{t}. \\hat{V}_{t - n + 1} \\). These canonicalized volumes can now be aggregated together by element-wise sum to produce a fused feature volume \\( \\hat{V}_{t}^{m} \\) for the surround-" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1432" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "ing 3D structure. This scheme allows NVM to disentangle the symmetry structure of the ambient 3D space from the \"content\" specific to the scene. By doing so, we encourage \\(SE(3)\\) equivariance in the memory space." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.152, + 0.469, + 0.182 + ], + "angle": 0, + "content": "We provide a detailed description of all components in our NVM module in the remaining section." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.182, + 0.471, + 0.318 + ], + "angle": 0, + "content": "3D Encoder: Our 3D encoder \\( Enc_{3D} \\) takes a 2D depth image as input and outputs a 3D feature volume. We parameterize this encoder with a set of 2D convolution layers followed by 3D convolution layers. Specifically, we generate a 2D feature map with the shape \\( (C, H, W) \\) using 2D convolution, then reshape the feature map channel-wise into a volume tensor with the shape \\( (C / D, D, H, W) \\). We chose a voxel-grid size of \\( D = 12 \\), \\( H = 6 \\), \\( W = 6 \\). The 3D convolution layers are added, to further refine the 3D features." + }, + { + "type": "equation", + "bbox": [ + 0.148, + 0.329, + 0.399, + 0.346 + ], + "angle": 0, + "content": "\\[\nV _ {i} = E n c _ {3 D} (O _ {i} ^ {v}) \\quad i = t.. t - n + 1\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.357, + 0.47, + 0.493 + ], + "angle": 0, + "content": "Pose Encoder: Our pose encoder \\( Enc_{\\mathrm{pose}} \\) takes two depth images and estimates the transformation in the latent space between two frames. Specifically, our pose encoder outputs 6D vectors representing the camera pose (rotation and translation) and we parameterize the pose encoder with a set of 2D convs. When constructing the NVM, the pose encoder is used to estimate the transformation \\( T_{t}^{t}\\ldots T_{t - n + 1}^{t} \\) between previous frames and the present frame by comparing previous visual observation \\( O_{t}^{v}\\ldots O_{t - n + 1}^{v} \\) and present \\( O_{t}^{v} \\)." + }, + { + "type": "equation", + "bbox": [ + 0.161, + 0.504, + 0.384, + 0.523 + ], + "angle": 0, + "content": "\\[\nT _ {i} ^ {j} = \\operatorname {E n c} _ {\\text {p o s e}} \\left(O _ {i} ^ {v}, O _ {j} ^ {v}\\right) \\in S E (3)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.532, + 0.469, + 0.626 + ], + "angle": 0, + "content": "Fusing 3D feature: Given feature volumes from different frames and the estimated transformation from different frames to the present frame. We transform the feature volume \\( V_{t}..V_{t - n + 1} \\) with the corresponding transformation \\( T_{t}^{t}.T_{t - n + 1}^{t} \\). Particularly, for each position \\( p = (i,j,k)^T \\) in 3D deep voxel, we transform it to \\( \\hat{p} = (\\hat{i},\\hat{j},\\hat{k})^T \\) with:" + }, + { + "type": "equation", + "bbox": [ + 0.232, + 0.637, + 0.315, + 0.652 + ], + "angle": 0, + "content": "\\[\n\\hat {p} = R p + t\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.663, + 0.47, + 0.709 + ], + "angle": 0, + "content": "where \\( R \\) is a 3x3 rotation matrix from \\( T_{t}^{t}\\ldots T_{t - n + 1}^{t} \\) and \\( t \\) is the translation vector from \\( T_{t}^{t}\\ldots T_{t - n + 1}^{t} \\), and we note the transformation function as \\( f \\)." + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.718, + 0.385, + 0.736 + ], + "angle": 0, + "content": "\\[\n\\hat {V} _ {i} = f \\left(V _ {i}, T _ {i} ^ {t}\\right) \\quad i = t.. t - n + 1\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.747, + 0.469, + 0.853 + ], + "angle": 0, + "content": "Since directly performing transformation in latent space might cause misalignment due to the coarse of voxel representation, we use two additional 3D conv layers to refine the transformed feature volume. We then aggregate transformed feature volumes to get Neural Volumetric Memory \\(\\hat{V}_t^m\\) representing the surrounding 3D structure in the present frame by computing the mean over feature volumes:" + }, + { + "type": "equation", + "bbox": [ + 0.202, + 0.863, + 0.345, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\hat {V} _ {t} ^ {m} = \\frac {1}{n} \\sum_ {i = t - n + 1} ^ {t} \\hat {V} _ {i}\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.09, + 0.898, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.227, + 0.894, + 0.282 + ], + "angle": 0, + "content": "Figure 4. Self-Supervised Learning: we train a separate decoder to predict the visual observation in the different frames given one visual observation and the estimated transformation between two frames, similar to Lai et al. [35]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.306, + 0.811, + 0.322 + ], + "angle": 0, + "content": "3.2. Learning NVM via Self-Supervision" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.329, + 0.892, + 0.495 + ], + "angle": 0, + "content": "Although the behavior cloning objective is sufficient in producing a good policy, being equivariant to translation and rotation automatically offers a standalone, self-supervised learning objective just for the neural volumetric memory. Shown in Figure 4, one can assume that the surrounding 3D scene remains static between frames. Because the camera is forward-looking, we can canonicalized feature volume from earlier frames, and use it to predict a camera image taken at later steps. This self-consistency loss, computed as the pixel reconstruction error, encourages the memory to be \\(SE(3)\\) equivariant." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.496, + 0.893, + 0.815 + ], + "angle": 0, + "content": "For given visual observation history over a short span of time \\( O_{t}^{v}..O_{t - n + 1}^{v} \\), we use the most dated visual observation \\( O_{t - n + 1}^{v} \\) in the visual history and predict all frames. We extract a 3D feature volume from \\( O_{t - n + 1}^{v}: V_{t - n + 1} = Enc_{3D}(O_{t - n + 1}^{v}) \\), estimate the relative camera pose \\( T_{t}^{t - n + 1}..T_{t - n + 1}^{t - n + 1} \\) between \\( O_{t}^{v}.O_{t - n + 1}^{v} \\) and the most dated \\( O_{t - n + 1}^{v} \\). We transform the extracted 3D feature volume \\( V_{t - n + 1} \\) into different frames with the corresponding transformations and then use a decoder \\( Dec \\) to predict the visual observation \\( \\hat{O}_{t..t - n + 1}^{v} \\) rendered in different frames. The Decoder \\( Dec \\) takes a deep 3D feature volume as input and outputs a 2D image. The decoder first reshapes the deep 3D feature volume with shape \\( (C,D,H,W) \\) into shape \\( (C*D,H,W) \\) and applies a set of 2D conv layers to map the reshaped features to images. Given the predicted visual observation \\( \\hat{O}_{t..t - n + 1}^{v} \\) and the original visual observation \\( O_{t}^{v}.O_{t - n + 1}^{v} \\) in different frames, we compute the prediction loss with pixel-wise L1 loss. The 3D feature encoder and the pose encoder are both trained in this predictive self-supervised training task. Specifically, we can formulate this 3D representation learning task as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.821, + 0.877, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} V _ {t - n + 1} = \\operatorname {E n c} _ {3 D} \\left(O _ {t - n + 1} ^ {v}\\right) \\\\ \\hat {O} _ {i} ^ {v} = \\operatorname {D e c} \\left(f \\left(V _ {t - n + 1}, T _ {i} ^ {t - n + 1}\\right)\\right) i = t.. t - n + 1 \\\\ L _ {r e c} = \\frac {1}{n} \\sum_ {i = t} ^ {t - n + 1} \\left| O _ {i} ^ {v} - \\hat {O} _ {i} ^ {v} \\right| \\\\ \\end{array}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1433" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.213 + ], + "angle": 0, + "content": "where \\( L_{rec} \\) represents the L1 loss. Since the information contained in different visual observations is naturally different, perfect prediction in our task is not feasible. However, such a predictive learning task could still improve the 3D structural understanding of the surrounding and encourages the SE(3) equivariance over the learned 3D latent space, for this reconstruction loss encourages the disentanglement between the surrounding 3D structure and the camera pose." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.222, + 0.445, + 0.238 + ], + "angle": 0, + "content": "3.3. Teacher Policy and Visuomotor Distillation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.245, + 0.471, + 0.307 + ], + "angle": 0, + "content": "Image observations introduce additional overhead and slow training down as much as 40 times1. For this reason, we bootstrap learning the visuomotor policy by first training a state-space policy that depends on the elevation map." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.325, + 0.47, + 0.432 + ], + "angle": 0, + "content": "Learning a state-only policy We use proximal policy optimization (PPO [54]) to train a single teacher policy \\(\\hat{\\pi}\\) in all four environments. We parameterize the teacher policy as a 4-layer ReLU network with [256, 128, 128, 64] latent neurons. To provide terrain information, we include two elevation maps, one larger and coarser than the other. Overall, the observation space for the teacher policy include:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.44, + 0.47, + 0.488 + ], + "angle": 0, + "content": "- proprioceptive state - \\(\\mathbb{R}^{63}\\) A 63-dimension vector consists of the gravity vector in the body frame \\(\\mathbf{g}_t\\), joint rotations \\(\\mathbf{q}_t\\), joint velocity \\(\\dot{\\mathbf{q}}_t\\), and previous actions \\(\\mathbf{a}_{t-1}\\)." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.491, + 0.472, + 0.537 + ], + "angle": 0, + "content": "- privileged information - \\(\\mathbb{R}^{14}\\) A 14-dimension vector including linear velocity \\(\\mathbf{v}_t\\) and angular velocity \\(\\omega_t\\) of the robot, and environment parameters \\(\\mathbf{e}_t\\)." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.541, + 0.47, + 0.587 + ], + "angle": 0, + "content": "- a dense elevation map \\( m_{dense} - \\mathbb{R}^{21 \\times 26} \\) with 5cm spacing between the grid points, which is only slightly bigger than the robot." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.591, + 0.47, + 0.624 + ], + "angle": 0, + "content": "- a sparse elevation map \\( m_{sparse} - \\mathbb{R}^{10 \\times 19} \\) with \\( 10\\mathrm{cm} \\) spacing. This covers a much larger area." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.44, + 0.472, + 0.624 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.633, + 0.47, + 0.665 + ], + "angle": 0, + "content": "We provide a more detailed description of the observation in the appendix." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.682, + 0.47, + 0.819 + ], + "angle": 0, + "content": "Reward function : We use the same reward function for teacher's training in all environments. The reward contains the following terms: 1) Reward for moving forward at a target velocity (in our experiment: \\(0.4\\mathrm{m / s}\\)). 2) an energy penalty to encourage the robot to use as little torque as possible. 3) Height tracking reward encouraging the robot to walk while maintaining a specific relative height to the ground (in our experiment: \\(0.265\\mathrm{m}\\)). 4) AMP reward [49] encouraging the agent to produce natural gaits." + }, + { + "type": "equation", + "bbox": [ + 0.136, + 0.83, + 0.409, + 0.866 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} R = \\alpha_ {\\text {f o r w a r d}} * R _ {\\text {f o r w a r d}} + \\alpha_ {\\text {e n e r g y}} * R _ {\\text {e n e r g y}} \\\\ + \\alpha_ {\\text {h e i g h t}} * R _ {\\text {h e i g h t}} + \\alpha_ {\\text {a m p}} * R _ {\\text {a m p}} \\\\ \\end{array}\n\\]" + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.875, + 0.471, + 0.902 + ], + "angle": 0, + "content": "1We are able to achieve 40k fps using just elevation map. Using \\(64 \\times 64\\) images, it slows down to 1k fps." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.089, + 0.892, + 0.118 + ], + "angle": 0, + "content": "Table 1. Comparison between observation spaces. State-only policy (w/ elevation-map) vs. vision policy" + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.12, + 0.884, + 0.196 + ], + "angle": 0, + "content": "
Joint statesVelocityElevation mapDepth map
Teacher policy✓✓✓×
Visual policy✓××✓
" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.213, + 0.892, + 0.245 + ], + "angle": 0, + "content": "We include the details of each reward term and the corresponding weights in the appendix" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.261, + 0.892, + 0.457 + ], + "angle": 0, + "content": "Visuomotor distillation: Given the privileged teacher policy from stage one, we then distill the privileged teacher policy \\(\\hat{\\pi}\\) into the visual policy \\(\\pi\\) without privileged information. Our visual policy utilizes proprioceptive state and visual observation for decision-making. The proprioceptive state \\(O_{prop}\\) is the same as we used to train privileged elevation-based policy in stage one while the visual observation contains a sequence of depth images (\\(O_{t}^{v}..O_{t-n+1}^{v}\\), \\(n=5\\) in our experiment) with a shape \\(64 \\times 64\\) from a depth camera mounted on the head of the robot. The observation space comparison between the privileged teacher policy and the visual policy is provided in Table 1, and the specific position of the camera is provided in Figure 2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.457, + 0.892, + 0.608 + ], + "angle": 0, + "content": "As shown in Figure 3, our visual policy first processes the sequential visual observation with our proposed NVM module. After obtaining NVM with shape \\((C,D,H,W)\\), our visual policy flattens the NVM into a 2D feature map with shape \\((C\\times D,H,W)\\) and utilizes a set of 2D conv layers and a single linear layer to get a 1D visual feature. Our visual policy encodes the proprioceptive state with a neural network and gets the proprioceptive feature. The proprioceptive feature and visual feature are then concatenated and fed into a final network to get the actual action for the robot." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.608, + 0.892, + 0.655 + ], + "angle": 0, + "content": "The visual policy is trained with behavior cloning by minimizing the L1 distance between actions from teacher policy \\(\\hat{\\pi}\\) and visual policy \\(\\pi\\)," + }, + { + "type": "equation", + "bbox": [ + 0.621, + 0.662, + 0.771, + 0.68 + ], + "angle": 0, + "content": "\\[\nL _ {B C} = | \\hat {\\pi} (p) - \\pi (o) |\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.687, + 0.892, + 0.749 + ], + "angle": 0, + "content": "where \\(\\mathbf{p}\\) is the privileged observation (including elevation maps), and \\(\\mathbf{o}\\) is the observation (including visual observations). Combined with our proposed self-supervised task, the overall training loss for our NVM is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.606, + 0.757, + 0.787, + 0.772 + ], + "angle": 0, + "content": "\\[\nL = \\lambda_ {B C} L _ {B C} + \\lambda_ {r e c} L _ {r e c}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.781, + 0.855, + 0.796 + ], + "angle": 0, + "content": "In our experiment, we use \\(\\lambda_{BC} = 1\\) and \\(\\lambda_{rec} = 0.01\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.808, + 0.634, + 0.825 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.832, + 0.734, + 0.848 + ], + "angle": 0, + "content": "4.1. Simulation Environments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.894, + 0.902 + ], + "angle": 0, + "content": "We evaluate NVM on four challenging terrains shown in Figure 2. Our intention is to evaluate the global and local perception ability of agents, as well as the robustness of" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1434" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.892, + 0.119 + ], + "angle": 0, + "content": "Table 2. Evaluation of All Policies: We evaluate all methods in all four environments and show their Traversing Rate and Success Rate. The privileged elevation-map-based policy is noted as Teacher, and its performance is regarded as the upper bound for the rest methods." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.128, + 0.89, + 0.269 + ], + "angle": 0, + "content": "
Traversing Rate (%) ↑Success Rate (%) ↑
ScenariosStagesStairsStonesObstaclesStagesStairsStonesObstacles
Teacher97.9±2.094.5±1.889.6±3.192.8±1.895.6±2.386.6±2.770.6±5.287.8±2.7
NaiveCNN76.4±26.981.5±28.629.4±10.663.3±22.553.9±19.771.3±25.40.5±0.845.2±16.4
NaiveCNN-MS72.1±25.484.5±29.726.4±9.363.7±22.748.9±18.973.1±26.20.0±0.048.1±18.3
NaiveCNN-RNN68.3±24.182.4±29.726.3±9.367.9±23.948.9±18.079.1±28.92.2±1.549.8±17.9
LocoTransformer74.0±26.585.5±29.925.2±8.965.2±22.953.3±19.375.8±26.80.9±1.146.7±16.9
NVM79.9±28.185.5±30.047.4±16.774.1±26.065.0±23.279.8±28.214.4±6.061.6±21.9
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.293, + 0.473, + 0.458 + ], + "angle": 0, + "content": "the learned locomotion skills. Our environments consist of Stages including multiple separate stages with gaps of different sizes between stages; Obstacles including multiple tall cuboids for the agent to avoid; Stairs including stairs with different heights and widths; Stones including numerous small stones of different shapes for the agent to step over. Stages, Stairs, and Stones environments evaluate the agent's ability to reason about the nearby surrounding and perform precise foot placement. Obstacle environment requires the agent to be able to perform long-term planning and avoid obstacles beforehand." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.467, + 0.188, + 0.482 + ], + "angle": 0, + "content": "4.2. Baselines" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.49, + 0.47, + 0.522 + ], + "angle": 0, + "content": "We compare our method against baselines that do not contain SE(3) equivariance priors:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.53, + 0.47, + 0.59 + ], + "angle": 0, + "content": "- NaiveCNN which learns visual features with a 2D CNN encoder followed by a linear projection. Visual features and proprioceptive features are concatenated. The action is then computed from the concatenated feature." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.595, + 0.47, + 0.656 + ], + "angle": 0, + "content": "- NaiveCNN-RNN which shares a similar structure with NaiveCNN, but before computing action from the concatenated feature, NaiveCNN-RNN baseline utilizes a 2-layer GRU to provide memory mechanism for the agent" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.66, + 0.47, + 0.735 + ], + "angle": 0, + "content": "- LocoTransformer which utilizes a shared transformer model to perform cross-modal reasoning between visual tokens from 2D visual feature map and proprioceptive features, instead of directly concatenating visual feature and proprioceptive feature." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.74, + 0.47, + 0.815 + ], + "angle": 0, + "content": "- Multi-Step NaiveCNN (NaiveCNN-MS) which use a 2D CNN to extract visual feature from each depth image and fuse them together to get the overall visual feature, instead of stacking multiple depth images and processing the stacked visual observation with a 2D CNN." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.53, + 0.47, + 0.815 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.903 + ], + "angle": 0, + "content": "For NaiveCNN and LocoTransformer, the visual observations in the near past are stacked in the channel dimension and serve as input for 2D encoder. All these baselines and our method share the same 2D CNN encoder structure for a fair comparison." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.292, + 0.677, + 0.308 + ], + "angle": 0, + "content": "4.3. Simulated Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.319, + 0.895, + 0.456 + ], + "angle": 0, + "content": "Evaluation metrics: We evaluate the final policies of all methods for 640 episodes in each simulated environment and compare the performance. We evaluate different methods by two metrics: traversing rate(\\%) and success rate(\\%). The traversing rate is defined by the distance the robot moves dividing the reachable distance (the distance between the end of the environment and starting position), and the Success Rate is defined by the ratio of the robot reaching the end of the environment." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.459, + 0.895, + 0.761 + ], + "angle": 0, + "content": "We evaluated all methods including the privileged teacher policy (noted as Teacher) across all environments and the results are provided in Table 2. Since in stage two of our pipeline, visual policies are trained by imitation learning (our method with self-supervised learning as well), the performance of the privileged teacher policy can be regarded as the upper bound for the performance of visual policies. We can see that our NVM outperforms all the baselines in all the environments, especially the Stones environment which requires the agent to estimate the 3D structure of the surroundings from visual observations and provide accurate foot placement. Surprisingly, it turns out our method traversed a longer distance in the obstacle environment, where the understanding of 3D structure is not required since it's possible to avoid the obstacles with only the latest visual observation. In this case, we conjecture that the 3D understanding of the environment helps the agent to keep closer to the trajectory of the teacher policy would take since the teacher could exploit the 3D structure of the surroundings to find the optimal path." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.895, + 0.903 + ], + "angle": 0, + "content": "We find that though NaiveCNN-MS introduces the mechanism of processing each frame of visual observation separately, there is no improvement compared with NaiveCNN. This comparison indicates that the improvement brought by our NVM is actually from explicitly modeling the 3D transformation instead of introducing more computation. Compared with NaiveCNN, the memory-enabled NaiveCNN-RNN also shows no improvement, which indicates that the memory mechanism provided by RNN is not better than, if" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1435" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.895, + 0.119 + ], + "angle": 0, + "content": "Table 3. Ablation study. We scale the grid resolution of the neural volumetric memory and the horizon of the past kept in memory. By default, our NVM uses use \\( D = 12 \\), \\( H = W = 6 \\), \\( n = 5 \\)." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.13, + 0.89, + 0.306 + ], + "angle": 0, + "content": "
Traversing Rate (%) ↑Success Rate (%) ↑
ScenariosStagesStairsStonesObstaclesStagesStairsStonesObstacles
NVM79.9±28.185.5±30.047.4±16.774.1±26.065.0±23.279.8±28.214.4±6.061.6±21.9
NVM(D=3)65.4±23.768.0±25.338.6±14.546.5±17.052.8±19.452.2±19.75.9±4.136.1±13.5
NVM(D=6)79.3±27.883.1±29.249.8±17.666.2±23.664.2±23.074.7±26.414.7±6.250.2±19.3
NVM(D=4x4)72.8±26.570.1±25.939.1±14.466.4±24.054.8±21.462.8±23.711.3±6.048.6±18.4
NVM(D=8x8)78.9±27.873.4±26.341.8±15.064.5±22.961.4±22.364.8±23.912.2±4.850.0±18.3
NVM(n=3)70.1±24.876.6±26.937.6±13.364.3±23.043.9±16.867.0±23.90.6±0.848.9±18.1
NVM(n=9)78.4±27.685.9±30.139.8±14.169.8±24.661.6±22.080.0±28.55.8±3.256.1±19.9
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.33, + 0.471, + 0.36 + ], + "angle": 0, + "content": "not worse, naively stacking the visual observations in the near past." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.377, + 0.265, + 0.392 + ], + "angle": 0, + "content": "4.4. Real World Results" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.402, + 0.474, + 0.903 + ], + "angle": 0, + "content": "To validate the performance of our method beyond simulation, we constructed multiple real-world experiment scenes as shown on the right of Figure 2. We evaluated our method in three scenes: Stages: we build 4 discrete stages with \\(15\\mathrm{cm} / 20\\mathrm{cm} / 30\\mathrm{cm}\\) between stages. Stairs: We build a stair consisting of 3 ladders with heights of \\(3\\mathrm{cm} / 6\\mathrm{cm} / 10\\mathrm{cm}\\). Obstacles: We place multiple cuboid obstacles in the narrow hallway for the robot to avoid. We deploy the policies in real-world scenes and record the distance robot moved for each episode. We terminate the episode after a certain amount of time or until the robot reaches the end of the environment or the robot falls. We repeat the experiment for each method and environment five times. Real-world trajectories are also provided in Figure 1. Since NaiveCNN, NaiveCNN-RNN, LocoTransformer show similar performance in simulation and LocoTransformer is known to be more robust in unseen scenarios and objects in the real world, we compared our NVM with LocoTransformer in these three scenarios. The quantitative results are shown in Table 4. We find that our NVM moved significantly further in Stages and Stairs while slightly further in Obstacles. This is because Stages and Stairs environment require the robot to understand the surrounding 3D structure to place the foot while Obstacles only requires the ability to perform high-level planning with visual observation. We observe that in Stages and Stairs environment, agents trained with LocoTransformer can place the front leg across the stage or over the ladder, but not the rear legs when the gap between stages is large or the ladder is tall, which demonstrates that lack of understanding of the surrounding 3D structure. We also found that even though LocoTransformer agent could move a reasonable distance in simulation in Stairs and Stages environment, it fails in" + }, + { + "type": "table_caption", + "bbox": [ + 0.536, + 0.328, + 0.859, + 0.343 + ], + "angle": 0, + "content": "Table 4. Evaluating the visual policy in the real world" + }, + { + "type": "table", + "bbox": [ + 0.513, + 0.353, + 0.878, + 0.438 + ], + "angle": 0, + "content": "
Distance Moved (m) ↑
ScenariosStagesStairsObstacles
NVM5.4±0.34.0±0.07.3±0.1
LocoTransformer3.4±0.12.1±0.86.7±0.4
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.455, + 0.895, + 0.532 + ], + "angle": 0, + "content": "the real world. The real-world results indicate that the understanding of the surrounding 3D structures is essential for legged robots to traverse complex terrains in the real world and our NVM module provides a high-quality volumetric memory of the surrounding 3D structure." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.544, + 0.731, + 0.56 + ], + "angle": 0, + "content": "4.5. Visualization of SSL task" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.569, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Aiming to gather insight into our learned 3D features volume, we visualized the synthesized visual observation \\(\\hat{O}_t^v\\) in our self-supervised learning task along with the present visual observation \\(O_t^v\\) (the ground truth in our SSL task), and most dated visual observation \\(O_{t - n + 1}^v\\) (the input in our SSL task) in Figure 5. Even though we apply extensive data augmentation over the raw visual observation (the second image for every case), the synthesized visual observations are clearly indicating the robustness of our model. In the synthesized visual observation for \\(Stair\\) and \\(Obstacle\\) environment, we can see the overall structure of the environment is well-preserved. In Figure 5b, The obstacles in the previous frame are transformed to the position in the current frame without much distortion. In Figure 5c, the boundary of the nearest ladder is clear, and the second nearest ladder is also observable in the present frame. We have to note that the reason the decoder can not recover the details away from the robot is in the input for SSL the large values are clipped. For more complex Stages and Stones environments, we found that even though the terrain details are not perfectly reconstructed, the essential regions for decision-making are reconstructed which means the learned 3D voxel" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "1436" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.093, + 0.482, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.253, + 0.186, + 0.306, + 0.198 + ], + "angle": 0, + "content": "(a) Stages" + }, + { + "type": "image", + "bbox": [ + 0.491, + 0.093, + 0.895, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.657, + 0.186, + 0.726, + 0.197 + ], + "angle": 0, + "content": "(b) Obstacles" + }, + { + "type": "image", + "bbox": [ + 0.079, + 0.203, + 0.482, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.255, + 0.295, + 0.304, + 0.306 + ], + "angle": 0, + "content": "(c) Stairs" + }, + { + "type": "image", + "bbox": [ + 0.491, + 0.203, + 0.895, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.665, + 0.295, + 0.72, + 0.306 + ], + "angle": 0, + "content": "(d) Stones" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.314, + 0.894, + 0.397 + ], + "angle": 0, + "content": "Figure 5. Visual reconstruction from learned decoder. We visualize the synthesized visual observation in our self-supervised task. For every tuple, the first image shows the robot moving in the environment, the second image is the input visual observation \\( O_{t - n + 1}^v \\), the third image is the synthesized visual observation using 3D feature volume \\( V_{t - n + 1} \\) extracted from \\( O_{t - n + 1}^v \\) and the estimated relative camera between \\( O_{t - n + 1}^v \\) and robot's ego-centric frame, and the fourth image is the actual visual observation captured by the robot (the ground-truth in our self-supervised task). For the input visual observation, we apply extensive data augmentation to the image to improve the robustness of our model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.416, + 0.47, + 0.475 + ], + "angle": 0, + "content": "features capture the most important parts in the visual observations. In Figure 5a, we can see the gaps between stages nearby are well reconstructed and in Figure 5d, we can see the reconstructed stone in front of the robot." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.485, + 0.23, + 0.501 + ], + "angle": 0, + "content": "4.6. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.509, + 0.469, + 0.553 + ], + "angle": 0, + "content": "we perform multiple sets of ablation studies, and the results are shown in Table 3, with the purpose of studying the influence of different components in our method." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.553, + 0.469, + 0.75 + ], + "angle": 0, + "content": "The Resolution of NVM: We provide our NVM module voxel features with different resolutions by slightly changing the structure of the 2D encoder, which resulting 2D feature maps in different shapes and use different depth parameter \\( D \\). From Table 3, we find that the larger depth value \\( D \\) generally provides better performance, since larger \\( D \\) converts the 2D feature map into more fine-grained voxel feature which results in fine-grained neural volumetric memory, and it's supposed to be easier for the agent to reason about the foot placement over complex terrain with more fine-grained neural volumetric memory. For \\( H \\) and \\( W \\), we found that \\( 6x6 \\) which provides enough detail horizontally without introducing much noise vertically works best." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.47, + 0.902 + ], + "angle": 0, + "content": "The Length of Visual History of NVM: We provide our NVM model with different lengths of visual history by changing the number of stored visual observation frames. As shown in Table 3, the performance of \\(NVM\\) improves when the length of visual history increases to 5 from 3, we think this is because the longer visual history provides longer 3D memory of the surrounding for the agent to safely placing the rear feet. But when the length of visual history increases to 9 from 5, the performance drops, we hypothesize that it is because when the visual history becomes too" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.416, + 0.893, + 0.476 + ], + "angle": 0, + "content": "long, it's hard for the pose encoder to estimate the accurate transformation between the oldest frame and the current frame, and the surplus visual information might be redundant for decision-making over terrains." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.494, + 0.619, + 0.51 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.522, + 0.893, + 0.749 + ], + "angle": 0, + "content": "We propose a short-term memory mechanism, the Neural Volumetric Memory (NVM), to enable legged robots to traverse complex terrains with ease. NVM aggregates knowledge from sequential visual observation in the near past by explicitly modeling the translation and rotation between previous frames and the present frame. The explicit modeling of the transformation and self-supervised task also encourage the \\(SE(3)\\) equivariance in the memory space. Our neural volumetric memory module enables the legged robot to traverse complex terrain in simulation and the real world and achieve better sim-to-real generalization results. This shows our neural volumetric memory provides reliable surrounding 3D structure information for the robot to perform decision-making and new possibility of utilizing visual observation for legged-robot." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.751, + 0.894, + 0.902 + ], + "angle": 0, + "content": "Acknowledgements. This project was supported, in part, by NSF CCF-2112665 (TILOS), NSF CAREER Award IIS-2240014, NSF 1730158 CI-New: Cognitive Hardware and Software Ecosystem Community Infrastructure (CHASE-CI), NSF ACI-1541349 CC*DNI Pacific Research Platform, Amazon Research Award and gifts from Qualcomm. Ge Yang is supported by the National Science Foundation Institute for Artificial Intelligence and Fundamental Interactions (IAIFI, https://iaifi.org/) under the Cooperative Agreement PHY-2019786. We thank Yuxiang Yang for providing help regarding hardware setup." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1437" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.17 + ], + "angle": 0, + "content": "[1] Ananye Agarwal, Ashish Kumar, Jitendra Malik, and Deepak Pathak. Legged locomotion in challenging terrains using egocentric vision. In 6th Annual Conference on Robot Learning, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.171, + 0.472, + 0.227 + ], + "angle": 0, + "content": "[2] Gerardo Bledt and Sangbae Kim. Extracting legged locomotion heuristics with regularized predictive control. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 406-412. IEEE, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.228, + 0.471, + 0.31 + ], + "angle": 0, + "content": "[3] Gerardo Bledt, Matthew J Powell, Benjamin Katz, Jared Di Carlo, Patrick M Wensing, and Sangbae Kim. Mit cheetah 3: Design and control of a robust, dynamic quadruped robot. In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 2245–2252. IEEE, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.311, + 0.471, + 0.367 + ], + "angle": 0, + "content": "[4] Jan Carius, René Ranftl, Vladlen Koltun, and Marco Hutter. Trajectory optimization for legged robots with slipping motions. IEEE Robotics and Automation Letters, 4(3):3013-3020, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.368, + 0.471, + 0.451 + ], + "angle": 0, + "content": "[5] Jared Di Carlo, Patrick M. Wensing, Benjamin Katz, Gerardo Bledt, and Sangbae Kim. Dynamic locomotion in the MIT cheetah 3 through convex model-predictive control. In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems, IROS 2018, Madrid, Spain, October 1-5, 2018, pages 1-9. IEEE, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.452, + 0.471, + 0.494 + ], + "angle": 0, + "content": "[6] Dian Chen, Brady Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Learning by cheating. In Conference on Robot Learning, pages 66-75. PMLR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.495, + 0.471, + 0.536 + ], + "angle": 0, + "content": "[7] Tao Chen, Jie Xu, and Pulkit Agrawal. A system for general in-hand object re-orientation. In Conference on Robot Learning, pages 297–307. PMLR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.537, + 0.471, + 0.592 + ], + "angle": 0, + "content": "[8] Annett Chilian and Heiko Hirschmüller. Stereo camera based navigation of mobile robots on rough terrain. In 2009 IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 4571-4576. IEEE, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.593, + 0.471, + 0.662 + ], + "angle": 0, + "content": "[9] Jared Di Carlo, Patrick M Wensing, Benjamin Katz, Gerardo Bledt, and Sangbae Kim. Dynamic locomotion in the mit cheetah 3 through convex model-predictive control. In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1-9. IEEE, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.663, + 0.47, + 0.731 + ], + "angle": 0, + "content": "[10] Yanran Ding, Abhishek Panda, and Hae-Won Park. Realtime model predictive control for versatile dynamic motions in quadrupedal robots. In 2019 International Conference on Robotics and Automation (ICRA), pages 8484-8490. IEEE, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.733, + 0.47, + 0.802 + ], + "angle": 0, + "content": "[11] Peter Fankhauser, Marko Bjelonic, C. Dario Bellicoso, Takahiro Miki, and Marco Hutter. Robust rough-terrain locomotion with a quadrupedal robot. In 2018 IEEE International Conference on Robotics and Automation (ICRA), pages 5761-5768, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.47, + 0.871 + ], + "angle": 0, + "content": "[12] Péter Fankhauser, Michael Bloesch, Christian Gehring, Marco Hutter, and Roland Siegwart. Robot-centric elevation mapping with uncertainty estimates. In International Conference on Climbing and Walking Robots (CLAWAR), 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.872, + 0.47, + 0.902 + ], + "angle": 0, + "content": "[13] Péter Fankhauser, Michael Bloesch, and Marco Hutter. Probabilistic terrain mapping for mobile robots with unclear" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "tain localization. IEEE Robotics and Automation Letters, 3(4):3019-3026, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.164 + ], + "angle": 0, + "content": "[14] Hsiao-Yu Fish Tung, Ricson Cheng, and Katerina Fragkiadaki. Learning spatial common sense with geometry-aware recurrent networks. In cvpr. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.165, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[15] Christian Gehring, Stelian Coros, Marco Hutter, Michael Blösch, Mark A. Hoepflinger, and Roland Siegwart. Control of dynamic gaits for a quadrupedal robot. In 2013 IEEE International Conference on Robotics and Automation, Karlsruhe, Germany, May 6-10, 2013, pages 3287-3292. IEEE, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.895, + 0.305 + ], + "angle": 0, + "content": "[16] Hartmut Geyer, Andre Seyfarth, and Reinhard Blickhan. Positive force feedback in bouncing gaits? Proceedings of the Royal Society of London. Series B: Biological Sciences, 270(1529):2173-2183, 2003. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.307, + 0.892, + 0.361 + ], + "angle": 0, + "content": "[17] Ruben Grandia, Farbod Farshidian, Alexey Dosovitskiy, René Ranftl, and Marco Hutter. Frequency-aware model predictive control. IEEE Robotics and Automation Letters, 4(2):1517-1524, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.363, + 0.892, + 0.419 + ], + "angle": 0, + "content": "[18] Adam W Harley, Shrinidhi Kowshika Lakshmikanth, Paul Schydlo, and Katerina Fragkiadaki. Tracking emerges by looking around static scenes, with neural 3d mapping. In eccv, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.421, + 0.892, + 0.476 + ], + "angle": 0, + "content": "[19] Adam W. Harley, Fangyu Li, Shrinidhi K. Lakshmikanth, Xian Zhou, Hsiao-Yu Fish Tung, and Katerina Fragkiadaki. Learning from unlabelled videos using contrastive predictive neural 3d mapping. In ICLR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.478, + 0.892, + 0.518 + ], + "angle": 0, + "content": "[20] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. Cambridge university press, 2003. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.521, + 0.892, + 0.588 + ], + "angle": 0, + "content": "[21] Ronghang Hu, Nikhila Ravi, Alexander C. Berg, and Deepak Pathak. Worldsheet: Wrapping the world in a 3d sheet for view synthesis from a single image. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.591, + 0.892, + 0.646 + ], + "angle": 0, + "content": "[22] Jemin Hwangbo, J. Lee, A. Dosovitskiy, Dario Bellicoso, Vassilios Tsounis, V. Koltun, and M. Hutter. Learning agile and dynamic motor skills for legged robots. Science Robotics, 4, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.649, + 0.892, + 0.732 + ], + "angle": 0, + "content": "[23] Chieko Imai, Minghao Zhang, Yuchen Zhang, Marcin Kierebinski, Ruihan Yang, Yuzhe Qin, and Xiaolong Wang. Vision-guided quadrupedal locomotion in the wild with multi-modal delay randomization. In 2022 IEEE/RSJ international conference on intelligent robots and systems (IROS). 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.734, + 0.892, + 0.829 + ], + "angle": 0, + "content": "[24] Atil Iscen, Ken Caluwaerts, Jie Tan, Tingnan Zhang, Erwin Coumans, Vikas Sindhwani, and Vincent Vanhoucke. Policies modulating trajectory generators. In 2nd Annual Conference on Robot Learning, CoRL 2018, Zürich, Switzerland, 29-31 October 2018, Proceedings, volume 87 of Proceedings of Machine Learning Research, pages 916-926. PMLR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[25] Deepali Jain, Atil Iscen, and Ken Caluwaerts. Hierarchical reinforcement learning for quadruped locomotion. In 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems, IROS 2019, Macau, SAR, China, November 3-8, 2019, pages 7551-7557. IEEE, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.895, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1438" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[26] Deepali Jain, Atil Iscen, and Ken Caluwaerts. From pixels to legs: Hierarchical learning of quadruped locomotion. arXiv preprint arXiv:2011.11722, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.19 + ], + "angle": 0, + "content": "[27] Joel Janai, Fatma Güney, Aseem Behl, Andreas Geiger, et al. Computer vision for autonomous vehicles: Problems, datasets and state of the art. Foundations and Trends® in Computer Graphics and Vision, 12(1-3):1-308, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.469, + 0.247 + ], + "angle": 0, + "content": "[28] Danilo Jimenez Rezende, SM Eslami, Shakir Mohamed, Peter Battaglia, Max Jaderberg, and Nicolas Heess. Unsupervised learning of 3d structure from images. Advances in neural information processing systems, 29:4996-5004, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.249, + 0.469, + 0.29 + ], + "angle": 0, + "content": "[29] Angjoo Kanazawa, Shubham Tulsiani, Alexei A Efros, and Jitendra Malik. Learning category-specific mesh reconstruction from image collections. In ECCV, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.292, + 0.469, + 0.359 + ], + "angle": 0, + "content": "[30] Elia Kaufmann, Antonio Loquercio, Rene Ranftl, Alexey Dosovitskiy, Vladlen Koltun, and Davide Scaramuzza. Deep drone racing: Learning agile flight in dynamic environments. In Conference on Robot Learning, pages 133-145. PMLR, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.362, + 0.469, + 0.416 + ], + "angle": 0, + "content": "[31] Elia Kaufmann, Antonio Loquercio, René Ranftl, Matthias Müller, Vladlen Koltun, and Davide Scaramuzza. Deep drone acrobatics. arXiv preprint arXiv:2006.05768, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.419, + 0.469, + 0.488 + ], + "angle": 0, + "content": "[32] Nate Kohl and Peter Stone. Policy gradient reinforcement learning for fast quadrupedal locomotion. In Proceedings of the 2004 IEEE International Conference on Robotics and Automation, ICRA 2004, April 26 - May 1, 2004, New Orleans, LA, USA, pages 2619-2624. IEEE, 2004. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.49, + 0.469, + 0.531 + ], + "angle": 0, + "content": "[33] Ashish Kumar, Zipeng Fu, Deepak Pathak, and Jitendra Malik. Rma: Rapid motor adaptation for legged robot. Robotics: Science and Systems, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.532, + 0.469, + 0.586 + ], + "angle": 0, + "content": "[34] In-So Kweon and Takeo Kanade. High-resolution terrain map from multiple sensor data. IEEE Transactions on Pattern Analysis and Machine Intelligence, 14(2):278-292, 1992. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.589, + 0.469, + 0.629 + ], + "angle": 0, + "content": "[35] Zihang Lai, Sifei Liu, Alexei A Efros, and Xiaolong Wang. Video autoencoder: self-supervised disentanglement of 3d structure and motion. In ICCV, 2021. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.632, + 0.469, + 0.673 + ], + "angle": 0, + "content": "[36] Joonho Lee, Jemin Hwangbo, Lorenz Wellhausen, Vladlen Koltun, and Marco Hutter. Learning quadrupedal locomotion over challenging terrain. Science robotics, 5(47), 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.675, + 0.469, + 0.714 + ], + "angle": 0, + "content": "[37] Antonio Loquercio, Ashish Kumar, and Jitendra Malik. Learning visual locomotion with cross-modal supervision, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.717, + 0.469, + 0.758 + ], + "angle": 0, + "content": "[38] G. Margolis, Ge Yang, Kartik Paigwar, Tao Chen, and Pulkit Agrawal. Rapid locomotion via reinforcement learning. ArXiv, abs/2205.02824, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.76, + 0.469, + 0.815 + ], + "angle": 0, + "content": "[39] Gabriel B Margolis, Tao Chen, Kartik Paigwar, Xiang Fu, Donghyun Kim, Sang bae Kim, and Pulkit Agrawal. Learning to jump from pixels. In 5th Annual Conference on Robot Learning, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[40] Takahiro Miki, Joonho Lee, Jemin Hwangbo, Lorenz Wellhausen, Vladlen Koltun, and Marco Hutter. Learning robust perceptive locomotion for quadrupedal robots in the wild. Sci Robot, 7(62):eabk2822, Jan. 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[41] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf:" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.121, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[42] Hirofumi Miura and Isao Shimoyama. Dynamic walk of a biped. The International Journal of Robotics Research, 3(2):60-74, 1984. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.163, + 0.892, + 0.231 + ], + "angle": 0, + "content": "[43] Tongzhou Mu, Zhan Ling, Fanbo Xiang, Derek Yang, Xu-anlin Li, Stone Tao, Zhiao Huang, Zhiwei Jia, and Hao Su. Maniskill: Generalizable manipulation skill benchmark with large-scale demonstrations. arXiv preprint arXiv:2107.14483, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.233, + 0.892, + 0.301 + ], + "angle": 0, + "content": "[44] Siva Karthik Mustikovela, Varun Jampani, Shalini De Mello, Sifei Liu, Umar Iqbal, Carsten Rother, and Jan Kautz. Self-supervised viewpoint learning from image collections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3971-3981, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.302, + 0.892, + 0.371 + ], + "angle": 0, + "content": "[45] Thu Nguyen-Phuoc, Chuan Li, Lucas Theis, Christian Richardt, and Yong-Liang Yang. Hologan: Unsupervised learning of 3d representations from natural images. In Proceedings of the IEEE International Conference on Computer Vision, pages 7588–7597, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.372, + 0.892, + 0.427 + ], + "angle": 0, + "content": "[46] Yiyuan Pan, Xuecheng Xu, Yue Wang, Xiaqing Ding, and Rong Xiong. GPU accelerated real-time traversability mapping. In 2019 IEEE International Conference on Robotics and Biomimetics (ROBIO), pages 734-740. IEEE, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.428, + 0.892, + 0.482 + ], + "angle": 0, + "content": "[47] Xue Bin Peng, Glen Berseth, KangKang Yin, and Michiel Van De Panne. Deeploco: Dynamic locomotion skills using hierarchical deep reinforcement learning. ACM Transactions on Graphics (TOG), 36(4):1-13, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.484, + 0.892, + 0.537 + ], + "angle": 0, + "content": "[48] Xue Bin Peng, Erwin Coumans, Tingnan Zhang, Tsang-Wei Edward Lee, Jie Tan, and Sergey Levine. Learning agile robotic locomotion skills by imitating animals. In Robotics: Science and Systems, 07 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.539, + 0.892, + 0.594 + ], + "angle": 0, + "content": "[49] Xue Bin Peng, Ze Ma, Pieter Abbeel, Sergey Levine, and Angjoo Kanazawa. Amp: Adversarial motion priors for stylized physics-based character control. ACM Trans. Graph., 40(4), July 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.595, + 0.892, + 0.649 + ], + "angle": 0, + "content": "[50] Lerrel Pinto, Marcin Andrychowicz, Peter Welinder, Wojciech Zaremba, and Pieter Abbeel. Asymmetric actor critic for image-based robot learning. arXiv preprint arXiv:1710.06542, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.65, + 0.892, + 0.704 + ], + "angle": 0, + "content": "[51] Marc H Raibert. Hopping in legged systems—modeling and simulation for the two-dimensional one-legged case. IEEE Transactions on Systems, Man, and Cybernetics, SMC-14(3):451–463, 1984. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.706, + 0.892, + 0.748 + ], + "angle": 0, + "content": "[52] Chris Rockwell, David F. Fouhey, and Justin Johnson. Pixelsynth: Generating a 3d-consistent experience from a single image. In ICCV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.749, + 0.892, + 0.803 + ], + "angle": 0, + "content": "[53] Nikita Rudin, David Hoeller, Philipp Reist, and Marco Hutter. Learning to walk in minutes using massively parallel deep reinforcement learning. In Conference on Robot Learning, pages 91-100. PMLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.804, + 0.892, + 0.845 + ], + "angle": 0, + "content": "[54] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[55] Mingyo Seo, Ryan Gupta, Yifeng Zhu, Alexy Skoutnev, Luis Sentis, and Yuke Zhu. Learning to walk by steering: Perceptive quadrupedal locomotion in dynamic environments. In arXiv preprint arXiv:2209.09233, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1439" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[56] Vincent Sitzmann, Justus Thies, Felix Heide, Matthias Nießner, Gordon Wetzstein, and Michael Zollhofer. Deepvoxels: Learning persistent 3d feature embeddings. In CVPR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.47, + 0.205 + ], + "angle": 0, + "content": "[57] Vincent Sitzmann, Michael Zollhöfer, and Gordon Wetzstein. Scene representation networks: Continuous 3d-structure-aware neural scene representations. In NeurIPS, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.471, + 0.248 + ], + "angle": 0, + "content": "[58] Maks Sorokin, Jie Tan, C. Karen Liu, and Sehoon Ha. Learning to navigate sidewalks in outdoor environments. IEEE Robotics and Automation Letters, 7(2):3906-3913, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.25, + 0.47, + 0.333 + ], + "angle": 0, + "content": "[59] Yu Sun, Wyatt L. Ubellacker, Wen-Loong Ma, Xiang Zhang, Changhao Wang, Noel V. Csomay-Shanklin, Masayoshi Tomizuka, Koushil Sreenath, and Aaron D. Ames. Online learning of unknown dynamics for model-based controllers in legged locomotion. IEEE Robotics and Automation Letters (RA-L), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.335, + 0.47, + 0.445 + ], + "angle": 0, + "content": "[60] Jie Tan, Tingnan Zhang, Erwin Coumans, Atil Iscen, Yunfei Bai, Danijar Hafner, Steven Bohez, and Vincent Vanhoucke. Sim-to-real: Learning agile locomotion for quadruped robots. In Hadas Kress-Gazit, Siddhartha S. Srinivasa, Tom Howard, and Nikolay Atanasov, editors, Robotics: Science and Systems XIV, Carnegie Mellon University, Pittsburgh, Pennsylvania, USA, June 26-30, 2018, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.447, + 0.469, + 0.503 + ], + "angle": 0, + "content": "[61] Nick Torkos and Michiel van de Panne. Footprint-based quadruped motion synthesis. In Proceedings of the Graphics Interface 1998 Conference, June 18-20, 1998, Vancouver, BC, Canada, pages 151-160, June 1998. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.505, + 0.469, + 0.558 + ], + "angle": 0, + "content": "[62] Shubham Tulsiani, Tinghui Zhou, Alexei A Efros, and Jitendra Malik. Multi-view supervision for single-view reconstruction via differentiable ray consistency. In CVPR, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.562, + 0.469, + 0.643 + ], + "angle": 0, + "content": "[63] Martin Wermelinger, Péter Fankhauser, Remo Diethelm, Philipp Krüsi, Roland Siegwart, and Marco Hutter. Navigation planning for legged robots in challenging terrain. In 2016 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1184-1189. IEEE, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.646, + 0.469, + 0.715 + ], + "angle": 0, + "content": "[64] Olivia Wiles, Georgia Gkioxari, Richard Szeliski, and Justin Johnson. Synsin: End-to-end view synthesis from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7467-7477, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.718, + 0.469, + 0.772 + ], + "angle": 0, + "content": "[65] Zhaoming Xie, Xingye Da, Buck Babich, Animesh Garg, and Michiel van de Panne. Glide: Generalizable quadrupedal locomotion in diverse environments with a centroidal model. CoRR, abs/2104.09771, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.469, + 0.83 + ], + "angle": 0, + "content": "[66] Zhaoming Xie, Hung Yu Ling, Nam Hee Kim, and Michiel van de Panne. Allsteps: curriculum-driven learning of stepping stone skills. In Computer Graphics Forum, volume 39, pages 213-224. Wiley Online Library, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.469, + 0.899 + ], + "angle": 0, + "content": "[67] Ruihan Yang, Minghao Zhang, Nicklas Hansen, Huazhe Xu, and Xiaolong Wang. Learning vision-guided quadrupedal locomotion end-to-end with cross-modal transformers. In International Conference on Learning Representations, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[68] KangKang Yin, Kevin Loken, and Michiel Van de Panne. Simbicon: Simple biped locomotion control. ACM Transactions on Graphics (TOG), 26(3):105-es, 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[69] Wenhao Yu, Deepali Jain, Alejandro Escontrela, Atil Iscen, Peng Xu, Erwin Coumans, Sehoon Ha, Jie Tan, and Tingnan Zhang. Visual-locomotion: Learning to walk on complex terrains with vision. In Aleksandra Faust, David Hsu, and Gerhard Neumann, editors, Proceedings of the 5th Conference on Robot Learning, volume 164 of Proceedings of Machine Learning Research, pages 1291-1302. PMLR, 08-11 Nov 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.248, + 0.892, + 0.288 + ], + "angle": 0, + "content": "[70] Tinghui Zhou, Matthew Brown, Noah Snavely, and David G Lowe. Unsupervised learning of depth and ego-motion from video. In CVPR, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.29, + 0.892, + 0.359 + ], + "angle": 0, + "content": "[71] Matt Zucker, Nathan Ratliff, Martin Stolle, Joel Chestnutt, J Andrew Bagnell, Christopher G Atkeson, and James Kuffner. Optimization and learning for rough terrain legged locomotion. The International Journal of Robotics Research, 30(2):175-191, 2011. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1440" + } + ] +] \ No newline at end of file diff --git a/2023/Neural Volumetric Memory for Visual Locomotion Control/dad36583-d1e2-4d95-a7b6-551afd512667_origin.pdf b/2023/Neural Volumetric Memory for Visual Locomotion Control/dad36583-d1e2-4d95-a7b6-551afd512667_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6dbf64deff2493dffad87bdbc2d2ac3250c68912 --- /dev/null +++ b/2023/Neural Volumetric Memory for Visual Locomotion Control/dad36583-d1e2-4d95-a7b6-551afd512667_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:793a368559ed0460135e8601eed8387fc2d5e058c1434d27855467aaba707cb4 +size 3432342 diff --git a/2023/Neural Volumetric Memory for Visual Locomotion Control/full.md b/2023/Neural Volumetric Memory for Visual Locomotion Control/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7e7e14d9cd8513cbd8f55ddde07556007f1c46c7 --- /dev/null +++ b/2023/Neural Volumetric Memory for Visual Locomotion Control/full.md @@ -0,0 +1,315 @@ +# Neural Volumetric Memory for Visual Locomotion Control + +Ruihan Yang1 Ge Yang2,3 Xiaolong Wang1 + +$^{1}$ UC San Diego $^{2}$ Institute of AI and Fundamental Interactions $^{3}$ MIT CSAIL + +![](images/37fbd98d0a807e7d96e00c56a762f543b5e37e2c7ebc8d833ab517e20a0e6662.jpg) +Figure 1. Real-World Deployment: We deploy our policy in real-world environments with Stones, Stairs, Stages, Obstacles, and Unstructured Terrain. Our robot utilizes the volumetric memory of the surrounding 3D structure to successfully traverse complex environments. + +# Abstract + +Legged robots have the potential to expand the reach of autonomy beyond paved roads. In this work, we consider the difficult problem of locomotion on challenging terrains using a single forward-facing depth camera. Due to the partial observability of the problem, the robot has to rely on past observations to infer the terrain currently beneath it. To solve this problem, we follow the paradigm in computer vision that explicitly models the 3D geometry of the scene and propose Neural Volumetric Memory (NVM), a geometric memory architecture that explicitly accounts for the SE(3) equivariance of the 3D world. NVM aggregates feature volumes from multiple camera views by first bringing them back to the ego-centric frame of the robot. We test the learned visual-locomotion policy on a physical robot and show that our approach, which explicitly introduces geometric priors during training, offers superior performance than more naive methods. We also include ablation studies and show that the representations stored in the neural volumetric memory capture sufficient geometric information to reconstruct the scene. Our project page with videos is https://rchalyang.github.io/NVM + +# 1. Introduction + +Consider difficult locomotion tasks such as walking up and down a flight of stairs and stepping over gaps (Figure 1). + +The control of such behaviors requires tight coupling with perception because vision is needed to provide details of the terrain right beneath the robot and the 3D scene immediately around it. This problem is also partially-observable. Immediately relevant terrain information is often occluded from the robot's current frame of observation, forcing it to rely on past observations for control decisions. For this reason, while blind controllers that are learned in simulation using reinforcement learning have achieved impressive results in agility and robustness [33, 36, 38], there are clear limitations on how much they can do. How to incorporate perception into the pipeline to produce an integrated visuomotor controller thus remains an open problem. + +A recent line of work combines perception with locomotion using ego-centric cameras mounted on the robot. The predominant approach for addressing partial observability is to do frame-stacking, where the robot maintains a visual buffer of recent images. This naive heuristic suffers from two major problems: first, frame-stacking on a moving robot ignores the equivariance structure of the 3D environment, making learning a lot more difficult as policy success now relies on being able to learn to account for spurious changes in camera poses. A second but subtler issue is that biological systems do not have the ability to save detailed visual observations pixel-by-pixel. These concerns motivate the creation of an intermediary, short-term memory mechanism to functionally aggregate streams of obser + +![](images/c5c8690fe4b432b9bf5793d57eee7ca061c0c7956d1c08dd5da9ce40760374da.jpg) +Figure 2. Overview of Simulated Environment & Real World Environment: Our simulated environments are shown on the left and real-world environments are shown on the right. For the real-world environment, the corresponding visual observations for each real-world environment are shown in the bottom row. All policies are trained in the simulation and transferred into the real world without fine-tuning. + +vation into a single, coherent representation of the world. + +Motivated by these observations, we introduce a novel volumetric memory architecture for legged locomotion control. Our architecture consists of a 2D to 3D feature volume encoder, and a pose estimator that can estimate the relative camera poses given two input images. When combined, the two networks constitute a Neural Volumetric Memory (NVM) that takes as input a sequence of depth images taken by a forward-looking, ego-centric camera and fuses them together into a coherent latent representation for locomotion control. We encourage the memory space to be $SE(3)$ equivariant to changes in the camera pose of the robot by incorporating translation and rotation operations based on estimated relative poses from the pose network. This inverse transformation allows NVM to align feature volumes from the past to the present ego-centric frame, making both integrating over multiple timesteps into a coherent scene representation and learning a policy, less difficult. + +Our training pipeline follows a two-step teacher-student process where the primary goal of the first stage is to produce behaviors in the form of a policy. After training completes, this policy can traverse these difficult terrains robustly, but it relies on privileged sensory information such as an elevation map and ground-truth velocity. Elevation maps obtained in the real world are often biased, incomplete, and full of errors [40], whereas ground-truth velocity information is typically only available in instrumented environments. Hence in the visuomotor distillation stage of the pipeline, which still runs in the simulator, we feed the stream of ego-centric views from the forward depth camera into the neural volumetric memory. We feed the content of this memory into a small policy network and train everything end-to-end including the two network components of the NVM using a behavior cloning loss where the state-only policy acts as the teacher. For completeness, we offer an additional self-supervised learning objective (Figure 4) that relies on novel-view consistency for learning. The end product of this visuomotor distillation pipeline is a memory-equipped visuomotor policy that can operate directly on the + +UniTree A1 robot hardware (see Figure 1). A single policy is used to handle all environments covered by this paper. We provide comprehensive experiment and ablation studies in both simulation and the real world, and show that our method outperforms all baselines by a large margin. It is thus essential to model the 3D structure of the environment. + +# 2. Related Work + +Blind Locomotion Controllers: Legged locomotion using the proprioception of the robot has been studied for decades [3, 16, 42, 51, 61, 68]. Traditionally, controllers are developed with model-based methods [2, 4, 5, 9, 10, 15, 17, 59]. However, most model-based methods fail to generalize to unseen environments in the real world. In order to attain better generalization and robustness, recent works [22, 24, 25, 32, 33, 36, 48, 60, 65] utilize model-free reinforcement learning to train neural network controllers in simulation and directly transfer the learned policies to the real robot. These methods are able to achieve high levels of agility and adapt over diverse terrains using just the proprioceptive state of the robot. The real-world utility of these robots, however, is also limited, since these blind controllers are unable to perceive the surrounding. In this paper, we propose to integrate perception to facilitate control, with a focus on a better 3D understanding of the environments. + +Locomotion with Exteroceptive Perception: The elevation map [12, 13, 34, 46] around the robot has been used to perform optimization and planning for foot placement [8, 11, 47, 63, 66, 71]. Recent RL controller [40, 53] also uses elevation maps in an end-to-end manner. Using the elevation map to represent the surrounding is straightforward for foot placement optimization, but the elevation map obtained in the real world is either noisy or requires a complicated hardware pipeline. Another option is using depth or RGB images to provide visual observation for the controller. Hierarchical control systems [26, 39, 55, 58, 69] are proposed to utilize depth reading while the low-level + +![](images/965c9e7a5b0ea50782226b52dd9d5ba31082b5c82f36c2aff77046d3fb38feaf.jpg) +Figure 3. Neural Volumetric Memory (NVM): Our NVM module extracts 3D feature volumes from every visual observation in the near past, transform them with corresponding estimated transformation into present robot frame, and fuse the transformed feature volumes into a neural volumetric memory for decision-making. + +controllers still rely on only the proprioceptive state of the robot. To get the most out of the depth image, Yang et al. [67] and Imai et al. [23] propose to learn visual RL policy to output the raw joint angles and enable the robot to maneuver in the wild avoiding different kinds of obstacles. Agarwal et al. [1] and Loquercioo et al. [37] train a policy using RL to traverse a large variety of terrains. Though these works show encouraging results in specific tasks, they have not fully exploited the 3D structure of the environment. On the other hand, we propose to train a novel neural volumetric memory together with policy learning, which allows our robot to walk through more challenging terrains. + +Volumetric Cognitive Map: How to represent and understand the 3D scene has been studied by the computer vision community for a long time. Hartley et al. [20] and Wiles et al. [64] provide representation built on geometry principles and camera motions. Learning-based approaches are proposed [21, 28, 29, 35, 52, 62, 64, 70] to learn 3D representation from 2D inputs. Neural 3D representations [14, 19, 41, 44, 45, 56, 57] has also shown great result in multiple computer vision tasks [14, 18, 19]. Our work is closely related to Lai et al. [35] where the deep voxel feature representing the entire scene and the camera trajectory are learned from video clips without odometry in a self-supervised manner. Instead of focusing on view synthesis, we aggregate the 3D feature volume to form a short-term memory representation for decision-making. + +Teacher-Student Training: To accelerate training and utilize the privileged information in simulation, Miki et al. [40] follow a teacher-student training framework, where the teacher policy is trained with privileged information and then distilled into student policy where the privileged information is not available. Multiple works in different research domains [6, 7, 27, 30, 31, 43, 50] follow this framework as well. We share a similar pipeline as [1, 39] where elevation + +map-based policies are trained first and the elevation-map-based policies are distilled visual policies afterward. + +# 3. Volumetric Memory for Legged Locomotion + +Legged locomotion using ego-centric camera views is intrinsically a partially-observed problem. To make the control problem tangible, our robot needs to aggregate information from previous frames and correctly infer the occluded terrain underneath it. During locomotion, the camera mounted directly on the robot chassis undergo large and spurious changes in pose, making integrating individual frames to a coherent representation non-trivial. + +To account for these camera pose changes, we propose neural volumetric memory (NVM) — a 3D representation format for scene features. It takes as input a sequence of visual observations and outputs a single 3D feature volume representing the surrounding 3D structure. + +# 3.1. Baking-in $SE(3)$ Equivariance in NVM + +Our training procedure encourages the learned memory space to be $SE(3)$ equivariant by applying translation and rotation transformations on the feature volume before aggregating additional information from new observations. + +As the first step (Figure 3), an encoder network receives a sequence of consecutive observations $O_{t}..O_{t - n + 1}$ . Instead of generating a compact feature embedding in the shape of a tuple, the 3D encoder Enc3D extracts feature volumes $V_{t}.V_{t - n + 1}$ . A second network Encpose estimates the relative changes in camera pose between the present frame $O_{t}^{v}$ and frames $O_{t}^{v}.O_{t - n + 1}^{v}$ from the past. This sequence of estimate allows NVM to transform the feature volume $V_{t}.V_{t - n + 1}$ from the past to the present frame using the transformations $T_{t}^{t}.T_{t - n + 1}^{t}$ . We annotate the transformed feature volume as $\hat{V}_{t}. \hat{V}_{t - n + 1}$ . These canonicalized volumes can now be aggregated together by element-wise sum to produce a fused feature volume $\hat{V}_{t}^{m}$ for the surround- + +ing 3D structure. This scheme allows NVM to disentangle the symmetry structure of the ambient 3D space from the "content" specific to the scene. By doing so, we encourage $SE(3)$ equivariance in the memory space. + +We provide a detailed description of all components in our NVM module in the remaining section. + +3D Encoder: Our 3D encoder $Enc_{3D}$ takes a 2D depth image as input and outputs a 3D feature volume. We parameterize this encoder with a set of 2D convolution layers followed by 3D convolution layers. Specifically, we generate a 2D feature map with the shape $(C, H, W)$ using 2D convolution, then reshape the feature map channel-wise into a volume tensor with the shape $(C / D, D, H, W)$ . We chose a voxel-grid size of $D = 12$ , $H = 6$ , $W = 6$ . The 3D convolution layers are added, to further refine the 3D features. + +$$ +V _ {i} = E n c _ {3 D} (O _ {i} ^ {v}) \quad i = t.. t - n + 1 +$$ + +Pose Encoder: Our pose encoder $Enc_{\mathrm{pose}}$ takes two depth images and estimates the transformation in the latent space between two frames. Specifically, our pose encoder outputs 6D vectors representing the camera pose (rotation and translation) and we parameterize the pose encoder with a set of 2D convs. When constructing the NVM, the pose encoder is used to estimate the transformation $T_{t}^{t}\ldots T_{t - n + 1}^{t}$ between previous frames and the present frame by comparing previous visual observation $O_{t}^{v}\ldots O_{t - n + 1}^{v}$ and present $O_{t}^{v}$ . + +$$ +T _ {i} ^ {j} = \operatorname {E n c} _ {\text {p o s e}} \left(O _ {i} ^ {v}, O _ {j} ^ {v}\right) \in S E (3) +$$ + +Fusing 3D feature: Given feature volumes from different frames and the estimated transformation from different frames to the present frame. We transform the feature volume $V_{t}..V_{t - n + 1}$ with the corresponding transformation $T_{t}^{t}.T_{t - n + 1}^{t}$ . Particularly, for each position $p = (i,j,k)^T$ in 3D deep voxel, we transform it to $\hat{p} = (\hat{i},\hat{j},\hat{k})^T$ with: + +$$ +\hat {p} = R p + t +$$ + +where $R$ is a 3x3 rotation matrix from $T_{t}^{t}\ldots T_{t - n + 1}^{t}$ and $t$ is the translation vector from $T_{t}^{t}\ldots T_{t - n + 1}^{t}$ , and we note the transformation function as $f$ . + +$$ +\hat {V} _ {i} = f \left(V _ {i}, T _ {i} ^ {t}\right) \quad i = t.. t - n + 1 +$$ + +Since directly performing transformation in latent space might cause misalignment due to the coarse of voxel representation, we use two additional 3D conv layers to refine the transformed feature volume. We then aggregate transformed feature volumes to get Neural Volumetric Memory $\hat{V}_t^m$ representing the surrounding 3D structure in the present frame by computing the mean over feature volumes: + +$$ +\hat {V} _ {t} ^ {m} = \frac {1}{n} \sum_ {i = t - n + 1} ^ {t} \hat {V} _ {i} +$$ + +![](images/bd9f9f00c1227a7833e3ae53f3cef80a4864abbc2bbb66bf80e951cd3f47fd1b.jpg) +Figure 4. Self-Supervised Learning: we train a separate decoder to predict the visual observation in the different frames given one visual observation and the estimated transformation between two frames, similar to Lai et al. [35]. + +# 3.2. Learning NVM via Self-Supervision + +Although the behavior cloning objective is sufficient in producing a good policy, being equivariant to translation and rotation automatically offers a standalone, self-supervised learning objective just for the neural volumetric memory. Shown in Figure 4, one can assume that the surrounding 3D scene remains static between frames. Because the camera is forward-looking, we can canonicalized feature volume from earlier frames, and use it to predict a camera image taken at later steps. This self-consistency loss, computed as the pixel reconstruction error, encourages the memory to be $SE(3)$ equivariant. + +For given visual observation history over a short span of time $O_{t}^{v}..O_{t - n + 1}^{v}$ , we use the most dated visual observation $O_{t - n + 1}^{v}$ in the visual history and predict all frames. We extract a 3D feature volume from $O_{t - n + 1}^{v}: V_{t - n + 1} = Enc_{3D}(O_{t - n + 1}^{v})$ , estimate the relative camera pose $T_{t}^{t - n + 1}..T_{t - n + 1}^{t - n + 1}$ between $O_{t}^{v}.O_{t - n + 1}^{v}$ and the most dated $O_{t - n + 1}^{v}$ . We transform the extracted 3D feature volume $V_{t - n + 1}$ into different frames with the corresponding transformations and then use a decoder $Dec$ to predict the visual observation $\hat{O}_{t..t - n + 1}^{v}$ rendered in different frames. The Decoder $Dec$ takes a deep 3D feature volume as input and outputs a 2D image. The decoder first reshapes the deep 3D feature volume with shape $(C,D,H,W)$ into shape $(C*D,H,W)$ and applies a set of 2D conv layers to map the reshaped features to images. Given the predicted visual observation $\hat{O}_{t..t - n + 1}^{v}$ and the original visual observation $O_{t}^{v}.O_{t - n + 1}^{v}$ in different frames, we compute the prediction loss with pixel-wise L1 loss. The 3D feature encoder and the pose encoder are both trained in this predictive self-supervised training task. Specifically, we can formulate this 3D representation learning task as follows: + +$$ +\begin{array}{l} V _ {t - n + 1} = \operatorname {E n c} _ {3 D} \left(O _ {t - n + 1} ^ {v}\right) \\ \hat {O} _ {i} ^ {v} = \operatorname {D e c} \left(f \left(V _ {t - n + 1}, T _ {i} ^ {t - n + 1}\right)\right) i = t.. t - n + 1 \\ L _ {r e c} = \frac {1}{n} \sum_ {i = t} ^ {t - n + 1} \left| O _ {i} ^ {v} - \hat {O} _ {i} ^ {v} \right| \\ \end{array} +$$ + +where $L_{rec}$ represents the L1 loss. Since the information contained in different visual observations is naturally different, perfect prediction in our task is not feasible. However, such a predictive learning task could still improve the 3D structural understanding of the surrounding and encourages the SE(3) equivariance over the learned 3D latent space, for this reconstruction loss encourages the disentanglement between the surrounding 3D structure and the camera pose. + +# 3.3. Teacher Policy and Visuomotor Distillation + +Image observations introduce additional overhead and slow training down as much as 40 times1. For this reason, we bootstrap learning the visuomotor policy by first training a state-space policy that depends on the elevation map. + +Learning a state-only policy We use proximal policy optimization (PPO [54]) to train a single teacher policy $\hat{\pi}$ in all four environments. We parameterize the teacher policy as a 4-layer ReLU network with [256, 128, 128, 64] latent neurons. To provide terrain information, we include two elevation maps, one larger and coarser than the other. Overall, the observation space for the teacher policy include: + +- proprioceptive state - $\mathbb{R}^{63}$ A 63-dimension vector consists of the gravity vector in the body frame $\mathbf{g}_t$ , joint rotations $\mathbf{q}_t$ , joint velocity $\dot{\mathbf{q}}_t$ , and previous actions $\mathbf{a}_{t-1}$ . +- privileged information - $\mathbb{R}^{14}$ A 14-dimension vector including linear velocity $\mathbf{v}_t$ and angular velocity $\omega_t$ of the robot, and environment parameters $\mathbf{e}_t$ . +- a dense elevation map $m_{dense} - \mathbb{R}^{21 \times 26}$ with 5cm spacing between the grid points, which is only slightly bigger than the robot. +- a sparse elevation map $m_{sparse} - \mathbb{R}^{10 \times 19}$ with $10\mathrm{cm}$ spacing. This covers a much larger area. + +We provide a more detailed description of the observation in the appendix. + +Reward function : We use the same reward function for teacher's training in all environments. The reward contains the following terms: 1) Reward for moving forward at a target velocity (in our experiment: $0.4\mathrm{m / s}$ ). 2) an energy penalty to encourage the robot to use as little torque as possible. 3) Height tracking reward encouraging the robot to walk while maintaining a specific relative height to the ground (in our experiment: $0.265\mathrm{m}$ ). 4) AMP reward [49] encouraging the agent to produce natural gaits. + +$$ +\begin{array}{l} R = \alpha_ {\text {f o r w a r d}} * R _ {\text {f o r w a r d}} + \alpha_ {\text {e n e r g y}} * R _ {\text {e n e r g y}} \\ + \alpha_ {\text {h e i g h t}} * R _ {\text {h e i g h t}} + \alpha_ {\text {a m p}} * R _ {\text {a m p}} \\ \end{array} +$$ + +Table 1. Comparison between observation spaces. State-only policy (w/ elevation-map) vs. vision policy + +
Joint statesVelocityElevation mapDepth map
Teacher policy✓✓✓×
Visual policy✓××✓
+ +We include the details of each reward term and the corresponding weights in the appendix + +Visuomotor distillation: Given the privileged teacher policy from stage one, we then distill the privileged teacher policy $\hat{\pi}$ into the visual policy $\pi$ without privileged information. Our visual policy utilizes proprioceptive state and visual observation for decision-making. The proprioceptive state $O_{prop}$ is the same as we used to train privileged elevation-based policy in stage one while the visual observation contains a sequence of depth images ( $O_{t}^{v}..O_{t-n+1}^{v}$ , $n=5$ in our experiment) with a shape $64 \times 64$ from a depth camera mounted on the head of the robot. The observation space comparison between the privileged teacher policy and the visual policy is provided in Table 1, and the specific position of the camera is provided in Figure 2. + +As shown in Figure 3, our visual policy first processes the sequential visual observation with our proposed NVM module. After obtaining NVM with shape $(C,D,H,W)$ , our visual policy flattens the NVM into a 2D feature map with shape $(C\times D,H,W)$ and utilizes a set of 2D conv layers and a single linear layer to get a 1D visual feature. Our visual policy encodes the proprioceptive state with a neural network and gets the proprioceptive feature. The proprioceptive feature and visual feature are then concatenated and fed into a final network to get the actual action for the robot. + +The visual policy is trained with behavior cloning by minimizing the L1 distance between actions from teacher policy $\hat{\pi}$ and visual policy $\pi$ , + +$$ +L _ {B C} = | \hat {\pi} (p) - \pi (o) | +$$ + +where $\mathbf{p}$ is the privileged observation (including elevation maps), and $\mathbf{o}$ is the observation (including visual observations). Combined with our proposed self-supervised task, the overall training loss for our NVM is given by: + +$$ +L = \lambda_ {B C} L _ {B C} + \lambda_ {r e c} L _ {r e c} +$$ + +In our experiment, we use $\lambda_{BC} = 1$ and $\lambda_{rec} = 0.01$ . + +# 4. Experiments + +# 4.1. Simulation Environments + +We evaluate NVM on four challenging terrains shown in Figure 2. Our intention is to evaluate the global and local perception ability of agents, as well as the robustness of + +Table 2. Evaluation of All Policies: We evaluate all methods in all four environments and show their Traversing Rate and Success Rate. The privileged elevation-map-based policy is noted as Teacher, and its performance is regarded as the upper bound for the rest methods. + +
Traversing Rate (%) ↑Success Rate (%) ↑
ScenariosStagesStairsStonesObstaclesStagesStairsStonesObstacles
Teacher97.9±2.094.5±1.889.6±3.192.8±1.895.6±2.386.6±2.770.6±5.287.8±2.7
NaiveCNN76.4±26.981.5±28.629.4±10.663.3±22.553.9±19.771.3±25.40.5±0.845.2±16.4
NaiveCNN-MS72.1±25.484.5±29.726.4±9.363.7±22.748.9±18.973.1±26.20.0±0.048.1±18.3
NaiveCNN-RNN68.3±24.182.4±29.726.3±9.367.9±23.948.9±18.079.1±28.92.2±1.549.8±17.9
LocoTransformer74.0±26.585.5±29.925.2±8.965.2±22.953.3±19.375.8±26.80.9±1.146.7±16.9
NVM79.9±28.185.5±30.047.4±16.774.1±26.065.0±23.279.8±28.214.4±6.061.6±21.9
+ +the learned locomotion skills. Our environments consist of Stages including multiple separate stages with gaps of different sizes between stages; Obstacles including multiple tall cuboids for the agent to avoid; Stairs including stairs with different heights and widths; Stones including numerous small stones of different shapes for the agent to step over. Stages, Stairs, and Stones environments evaluate the agent's ability to reason about the nearby surrounding and perform precise foot placement. Obstacle environment requires the agent to be able to perform long-term planning and avoid obstacles beforehand. + +# 4.2. Baselines + +We compare our method against baselines that do not contain SE(3) equivariance priors: + +- NaiveCNN which learns visual features with a 2D CNN encoder followed by a linear projection. Visual features and proprioceptive features are concatenated. The action is then computed from the concatenated feature. +- NaiveCNN-RNN which shares a similar structure with NaiveCNN, but before computing action from the concatenated feature, NaiveCNN-RNN baseline utilizes a 2-layer GRU to provide memory mechanism for the agent +- LocoTransformer which utilizes a shared transformer model to perform cross-modal reasoning between visual tokens from 2D visual feature map and proprioceptive features, instead of directly concatenating visual feature and proprioceptive feature. +- Multi-Step NaiveCNN (NaiveCNN-MS) which use a 2D CNN to extract visual feature from each depth image and fuse them together to get the overall visual feature, instead of stacking multiple depth images and processing the stacked visual observation with a 2D CNN. + +For NaiveCNN and LocoTransformer, the visual observations in the near past are stacked in the channel dimension and serve as input for 2D encoder. All these baselines and our method share the same 2D CNN encoder structure for a fair comparison. + +# 4.3. Simulated Results + +Evaluation metrics: We evaluate the final policies of all methods for 640 episodes in each simulated environment and compare the performance. We evaluate different methods by two metrics: traversing rate(\%) and success rate(\%). The traversing rate is defined by the distance the robot moves dividing the reachable distance (the distance between the end of the environment and starting position), and the Success Rate is defined by the ratio of the robot reaching the end of the environment. + +We evaluated all methods including the privileged teacher policy (noted as Teacher) across all environments and the results are provided in Table 2. Since in stage two of our pipeline, visual policies are trained by imitation learning (our method with self-supervised learning as well), the performance of the privileged teacher policy can be regarded as the upper bound for the performance of visual policies. We can see that our NVM outperforms all the baselines in all the environments, especially the Stones environment which requires the agent to estimate the 3D structure of the surroundings from visual observations and provide accurate foot placement. Surprisingly, it turns out our method traversed a longer distance in the obstacle environment, where the understanding of 3D structure is not required since it's possible to avoid the obstacles with only the latest visual observation. In this case, we conjecture that the 3D understanding of the environment helps the agent to keep closer to the trajectory of the teacher policy would take since the teacher could exploit the 3D structure of the surroundings to find the optimal path. + +We find that though NaiveCNN-MS introduces the mechanism of processing each frame of visual observation separately, there is no improvement compared with NaiveCNN. This comparison indicates that the improvement brought by our NVM is actually from explicitly modeling the 3D transformation instead of introducing more computation. Compared with NaiveCNN, the memory-enabled NaiveCNN-RNN also shows no improvement, which indicates that the memory mechanism provided by RNN is not better than, if + +Table 3. Ablation study. We scale the grid resolution of the neural volumetric memory and the horizon of the past kept in memory. By default, our NVM uses use $D = 12$ , $H = W = 6$ , $n = 5$ . + +
Traversing Rate (%) ↑Success Rate (%) ↑
ScenariosStagesStairsStonesObstaclesStagesStairsStonesObstacles
NVM79.9±28.185.5±30.047.4±16.774.1±26.065.0±23.279.8±28.214.4±6.061.6±21.9
NVM(D=3)65.4±23.768.0±25.338.6±14.546.5±17.052.8±19.452.2±19.75.9±4.136.1±13.5
NVM(D=6)79.3±27.883.1±29.249.8±17.666.2±23.664.2±23.074.7±26.414.7±6.250.2±19.3
NVM(D=4x4)72.8±26.570.1±25.939.1±14.466.4±24.054.8±21.462.8±23.711.3±6.048.6±18.4
NVM(D=8x8)78.9±27.873.4±26.341.8±15.064.5±22.961.4±22.364.8±23.912.2±4.850.0±18.3
NVM(n=3)70.1±24.876.6±26.937.6±13.364.3±23.043.9±16.867.0±23.90.6±0.848.9±18.1
NVM(n=9)78.4±27.685.9±30.139.8±14.169.8±24.661.6±22.080.0±28.55.8±3.256.1±19.9
+ +not worse, naively stacking the visual observations in the near past. + +# 4.4. Real World Results + +To validate the performance of our method beyond simulation, we constructed multiple real-world experiment scenes as shown on the right of Figure 2. We evaluated our method in three scenes: Stages: we build 4 discrete stages with $15\mathrm{cm} / 20\mathrm{cm} / 30\mathrm{cm}$ between stages. Stairs: We build a stair consisting of 3 ladders with heights of $3\mathrm{cm} / 6\mathrm{cm} / 10\mathrm{cm}$ . Obstacles: We place multiple cuboid obstacles in the narrow hallway for the robot to avoid. We deploy the policies in real-world scenes and record the distance robot moved for each episode. We terminate the episode after a certain amount of time or until the robot reaches the end of the environment or the robot falls. We repeat the experiment for each method and environment five times. Real-world trajectories are also provided in Figure 1. Since NaiveCNN, NaiveCNN-RNN, LocoTransformer show similar performance in simulation and LocoTransformer is known to be more robust in unseen scenarios and objects in the real world, we compared our NVM with LocoTransformer in these three scenarios. The quantitative results are shown in Table 4. We find that our NVM moved significantly further in Stages and Stairs while slightly further in Obstacles. This is because Stages and Stairs environment require the robot to understand the surrounding 3D structure to place the foot while Obstacles only requires the ability to perform high-level planning with visual observation. We observe that in Stages and Stairs environment, agents trained with LocoTransformer can place the front leg across the stage or over the ladder, but not the rear legs when the gap between stages is large or the ladder is tall, which demonstrates that lack of understanding of the surrounding 3D structure. We also found that even though LocoTransformer agent could move a reasonable distance in simulation in Stairs and Stages environment, it fails in + +Table 4. Evaluating the visual policy in the real world + +
Distance Moved (m) ↑
ScenariosStagesStairsObstacles
NVM5.4±0.34.0±0.07.3±0.1
LocoTransformer3.4±0.12.1±0.86.7±0.4
+ +the real world. The real-world results indicate that the understanding of the surrounding 3D structures is essential for legged robots to traverse complex terrains in the real world and our NVM module provides a high-quality volumetric memory of the surrounding 3D structure. + +# 4.5. Visualization of SSL task + +Aiming to gather insight into our learned 3D features volume, we visualized the synthesized visual observation $\hat{O}_t^v$ in our self-supervised learning task along with the present visual observation $O_t^v$ (the ground truth in our SSL task), and most dated visual observation $O_{t - n + 1}^v$ (the input in our SSL task) in Figure 5. Even though we apply extensive data augmentation over the raw visual observation (the second image for every case), the synthesized visual observations are clearly indicating the robustness of our model. In the synthesized visual observation for $Stair$ and $Obstacle$ environment, we can see the overall structure of the environment is well-preserved. In Figure 5b, The obstacles in the previous frame are transformed to the position in the current frame without much distortion. In Figure 5c, the boundary of the nearest ladder is clear, and the second nearest ladder is also observable in the present frame. We have to note that the reason the decoder can not recover the details away from the robot is in the input for SSL the large values are clipped. For more complex Stages and Stones environments, we found that even though the terrain details are not perfectly reconstructed, the essential regions for decision-making are reconstructed which means the learned 3D voxel + +![](images/5cdfdfdfff859ff42c5df33b8d3d43adbb51e7b3e89254457d7a3cdd9984a7a4.jpg) +(a) Stages + +![](images/e6f2145722811db0516d8f3f33dbafa54e9063633b284bd2bdfbca7eba5131c7.jpg) +(b) Obstacles + +![](images/fcddc846ab074ba48db9a679d83d06568e909c217d6c7d1e69123936a3388ed5.jpg) +(c) Stairs + +![](images/1a8f9e1f53f7f204c7ba3a870cc35854f92a0c7b407d9724dd94a17c5becf98c.jpg) +(d) Stones +Figure 5. Visual reconstruction from learned decoder. We visualize the synthesized visual observation in our self-supervised task. For every tuple, the first image shows the robot moving in the environment, the second image is the input visual observation $O_{t - n + 1}^v$ , the third image is the synthesized visual observation using 3D feature volume $V_{t - n + 1}$ extracted from $O_{t - n + 1}^v$ and the estimated relative camera between $O_{t - n + 1}^v$ and robot's ego-centric frame, and the fourth image is the actual visual observation captured by the robot (the ground-truth in our self-supervised task). For the input visual observation, we apply extensive data augmentation to the image to improve the robustness of our model. + +features capture the most important parts in the visual observations. In Figure 5a, we can see the gaps between stages nearby are well reconstructed and in Figure 5d, we can see the reconstructed stone in front of the robot. + +# 4.6. Ablation Study + +we perform multiple sets of ablation studies, and the results are shown in Table 3, with the purpose of studying the influence of different components in our method. + +The Resolution of NVM: We provide our NVM module voxel features with different resolutions by slightly changing the structure of the 2D encoder, which resulting 2D feature maps in different shapes and use different depth parameter $D$ . From Table 3, we find that the larger depth value $D$ generally provides better performance, since larger $D$ converts the 2D feature map into more fine-grained voxel feature which results in fine-grained neural volumetric memory, and it's supposed to be easier for the agent to reason about the foot placement over complex terrain with more fine-grained neural volumetric memory. For $H$ and $W$ , we found that $6x6$ which provides enough detail horizontally without introducing much noise vertically works best. + +The Length of Visual History of NVM: We provide our NVM model with different lengths of visual history by changing the number of stored visual observation frames. As shown in Table 3, the performance of $NVM$ improves when the length of visual history increases to 5 from 3, we think this is because the longer visual history provides longer 3D memory of the surrounding for the agent to safely placing the rear feet. But when the length of visual history increases to 9 from 5, the performance drops, we hypothesize that it is because when the visual history becomes too + +long, it's hard for the pose encoder to estimate the accurate transformation between the oldest frame and the current frame, and the surplus visual information might be redundant for decision-making over terrains. + +# 5. Conclusion + +We propose a short-term memory mechanism, the Neural Volumetric Memory (NVM), to enable legged robots to traverse complex terrains with ease. NVM aggregates knowledge from sequential visual observation in the near past by explicitly modeling the translation and rotation between previous frames and the present frame. The explicit modeling of the transformation and self-supervised task also encourage the $SE(3)$ equivariance in the memory space. Our neural volumetric memory module enables the legged robot to traverse complex terrain in simulation and the real world and achieve better sim-to-real generalization results. This shows our neural volumetric memory provides reliable surrounding 3D structure information for the robot to perform decision-making and new possibility of utilizing visual observation for legged-robot. + +Acknowledgements. This project was supported, in part, by NSF CCF-2112665 (TILOS), NSF CAREER Award IIS-2240014, NSF 1730158 CI-New: Cognitive Hardware and Software Ecosystem Community Infrastructure (CHASE-CI), NSF ACI-1541349 CC*DNI Pacific Research Platform, Amazon Research Award and gifts from Qualcomm. Ge Yang is supported by the National Science Foundation Institute for Artificial Intelligence and Fundamental Interactions (IAIFI, https://iaifi.org/) under the Cooperative Agreement PHY-2019786. We thank Yuxiang Yang for providing help regarding hardware setup. + +# References + +[1] Ananye Agarwal, Ashish Kumar, Jitendra Malik, and Deepak Pathak. Legged locomotion in challenging terrains using egocentric vision. In 6th Annual Conference on Robot Learning, 2022. 3 +[2] Gerardo Bledt and Sangbae Kim. Extracting legged locomotion heuristics with regularized predictive control. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 406-412. IEEE, 2020. 2 +[3] Gerardo Bledt, Matthew J Powell, Benjamin Katz, Jared Di Carlo, Patrick M Wensing, and Sangbae Kim. Mit cheetah 3: Design and control of a robust, dynamic quadruped robot. In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 2245–2252. IEEE, 2018. 2 +[4] Jan Carius, René Ranftl, Vladlen Koltun, and Marco Hutter. Trajectory optimization for legged robots with slipping motions. IEEE Robotics and Automation Letters, 4(3):3013-3020, 2019. 2 +[5] Jared Di Carlo, Patrick M. Wensing, Benjamin Katz, Gerardo Bledt, and Sangbae Kim. Dynamic locomotion in the MIT cheetah 3 through convex model-predictive control. In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems, IROS 2018, Madrid, Spain, October 1-5, 2018, pages 1-9. IEEE, 2018. 2 +[6] Dian Chen, Brady Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Learning by cheating. In Conference on Robot Learning, pages 66-75. PMLR, 2020. 3 +[7] Tao Chen, Jie Xu, and Pulkit Agrawal. A system for general in-hand object re-orientation. In Conference on Robot Learning, pages 297–307. PMLR, 2022. 3 +[8] Annett Chilian and Heiko Hirschmüller. Stereo camera based navigation of mobile robots on rough terrain. In 2009 IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 4571-4576. IEEE, 2009. 2 +[9] Jared Di Carlo, Patrick M Wensing, Benjamin Katz, Gerardo Bledt, and Sangbae Kim. Dynamic locomotion in the mit cheetah 3 through convex model-predictive control. In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1-9. IEEE, 2018. 2 +[10] Yanran Ding, Abhishek Panda, and Hae-Won Park. Realtime model predictive control for versatile dynamic motions in quadrupedal robots. In 2019 International Conference on Robotics and Automation (ICRA), pages 8484-8490. IEEE, 2019. 2 +[11] Peter Fankhauser, Marko Bjelonic, C. Dario Bellicoso, Takahiro Miki, and Marco Hutter. Robust rough-terrain locomotion with a quadrupedal robot. In 2018 IEEE International Conference on Robotics and Automation (ICRA), pages 5761-5768, 2018. 2 +[12] Péter Fankhauser, Michael Bloesch, Christian Gehring, Marco Hutter, and Roland Siegwart. Robot-centric elevation mapping with uncertainty estimates. In International Conference on Climbing and Walking Robots (CLAWAR), 2014. 2 +[13] Péter Fankhauser, Michael Bloesch, and Marco Hutter. Probabilistic terrain mapping for mobile robots with unclear + +tain localization. IEEE Robotics and Automation Letters, 3(4):3019-3026, 2018. 2 +[14] Hsiao-Yu Fish Tung, Ricson Cheng, and Katerina Fragkiadaki. Learning spatial common sense with geometry-aware recurrent networks. In cvpr. 3 +[15] Christian Gehring, Stelian Coros, Marco Hutter, Michael Blösch, Mark A. Hoepflinger, and Roland Siegwart. Control of dynamic gaits for a quadrupedal robot. In 2013 IEEE International Conference on Robotics and Automation, Karlsruhe, Germany, May 6-10, 2013, pages 3287-3292. IEEE, 2013. 2 +[16] Hartmut Geyer, Andre Seyfarth, and Reinhard Blickhan. Positive force feedback in bouncing gaits? Proceedings of the Royal Society of London. Series B: Biological Sciences, 270(1529):2173-2183, 2003. 2 +[17] Ruben Grandia, Farbod Farshidian, Alexey Dosovitskiy, René Ranftl, and Marco Hutter. Frequency-aware model predictive control. IEEE Robotics and Automation Letters, 4(2):1517-1524, 2019. 2 +[18] Adam W Harley, Shrinidhi Kowshika Lakshmikanth, Paul Schydlo, and Katerina Fragkiadaki. Tracking emerges by looking around static scenes, with neural 3d mapping. In eccv, 2020. 3 +[19] Adam W. Harley, Fangyu Li, Shrinidhi K. Lakshmikanth, Xian Zhou, Hsiao-Yu Fish Tung, and Katerina Fragkiadaki. Learning from unlabelled videos using contrastive predictive neural 3d mapping. In ICLR, 2019. 3 +[20] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. Cambridge university press, 2003. 3 +[21] Ronghang Hu, Nikhila Ravi, Alexander C. Berg, and Deepak Pathak. Worldsheet: Wrapping the world in a 3d sheet for view synthesis from a single image. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2021. 3 +[22] Jemin Hwangbo, J. Lee, A. Dosovitskiy, Dario Bellicoso, Vassilios Tsounis, V. Koltun, and M. Hutter. Learning agile and dynamic motor skills for legged robots. Science Robotics, 4, 2019. 2 +[23] Chieko Imai, Minghao Zhang, Yuchen Zhang, Marcin Kierebinski, Ruihan Yang, Yuzhe Qin, and Xiaolong Wang. Vision-guided quadrupedal locomotion in the wild with multi-modal delay randomization. In 2022 IEEE/RSJ international conference on intelligent robots and systems (IROS). 3 +[24] Atil Iscen, Ken Caluwaerts, Jie Tan, Tingnan Zhang, Erwin Coumans, Vikas Sindhwani, and Vincent Vanhoucke. Policies modulating trajectory generators. In 2nd Annual Conference on Robot Learning, CoRL 2018, Zürich, Switzerland, 29-31 October 2018, Proceedings, volume 87 of Proceedings of Machine Learning Research, pages 916-926. PMLR, 2018. 2 +[25] Deepali Jain, Atil Iscen, and Ken Caluwaerts. Hierarchical reinforcement learning for quadruped locomotion. In 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems, IROS 2019, Macau, SAR, China, November 3-8, 2019, pages 7551-7557. IEEE, 2019. 2 + +[26] Deepali Jain, Atil Iscen, and Ken Caluwaerts. From pixels to legs: Hierarchical learning of quadruped locomotion. arXiv preprint arXiv:2011.11722, 2020. 2 +[27] Joel Janai, Fatma Güney, Aseem Behl, Andreas Geiger, et al. Computer vision for autonomous vehicles: Problems, datasets and state of the art. Foundations and Trends® in Computer Graphics and Vision, 12(1-3):1-308, 2020. 3 +[28] Danilo Jimenez Rezende, SM Eslami, Shakir Mohamed, Peter Battaglia, Max Jaderberg, and Nicolas Heess. Unsupervised learning of 3d structure from images. Advances in neural information processing systems, 29:4996-5004, 2016. 3 +[29] Angjoo Kanazawa, Shubham Tulsiani, Alexei A Efros, and Jitendra Malik. Learning category-specific mesh reconstruction from image collections. In ECCV, 2018. 3 +[30] Elia Kaufmann, Antonio Loquercio, Rene Ranftl, Alexey Dosovitskiy, Vladlen Koltun, and Davide Scaramuzza. Deep drone racing: Learning agile flight in dynamic environments. In Conference on Robot Learning, pages 133-145. PMLR, 2018. 3 +[31] Elia Kaufmann, Antonio Loquercio, René Ranftl, Matthias Müller, Vladlen Koltun, and Davide Scaramuzza. Deep drone acrobatics. arXiv preprint arXiv:2006.05768, 2020. 3 +[32] Nate Kohl and Peter Stone. Policy gradient reinforcement learning for fast quadrupedal locomotion. In Proceedings of the 2004 IEEE International Conference on Robotics and Automation, ICRA 2004, April 26 - May 1, 2004, New Orleans, LA, USA, pages 2619-2624. IEEE, 2004. 2 +[33] Ashish Kumar, Zipeng Fu, Deepak Pathak, and Jitendra Malik. Rma: Rapid motor adaptation for legged robot. Robotics: Science and Systems, 2021. 1, 2 +[34] In-So Kweon and Takeo Kanade. High-resolution terrain map from multiple sensor data. IEEE Transactions on Pattern Analysis and Machine Intelligence, 14(2):278-292, 1992. 2 +[35] Zihang Lai, Sifei Liu, Alexei A Efros, and Xiaolong Wang. Video autoencoder: self-supervised disentanglement of 3d structure and motion. In ICCV, 2021. 3, 4 +[36] Joonho Lee, Jemin Hwangbo, Lorenz Wellhausen, Vladlen Koltun, and Marco Hutter. Learning quadrupedal locomotion over challenging terrain. Science robotics, 5(47), 2020. 1, 2 +[37] Antonio Loquercio, Ashish Kumar, and Jitendra Malik. Learning visual locomotion with cross-modal supervision, 2022. 3 +[38] G. Margolis, Ge Yang, Kartik Paigwar, Tao Chen, and Pulkit Agrawal. Rapid locomotion via reinforcement learning. ArXiv, abs/2205.02824, 2022. 1 +[39] Gabriel B Margolis, Tao Chen, Kartik Paigwar, Xiang Fu, Donghyun Kim, Sang bae Kim, and Pulkit Agrawal. Learning to jump from pixels. In 5th Annual Conference on Robot Learning, 2021. 2, 3 +[40] Takahiro Miki, Joonho Lee, Jemin Hwangbo, Lorenz Wellhausen, Vladlen Koltun, and Marco Hutter. Learning robust perceptive locomotion for quadrupedal robots in the wild. Sci Robot, 7(62):eabk2822, Jan. 2022. 2, 3 +[41] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: + +Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 3 +[42] Hirofumi Miura and Isao Shimoyama. Dynamic walk of a biped. The International Journal of Robotics Research, 3(2):60-74, 1984. 2 +[43] Tongzhou Mu, Zhan Ling, Fanbo Xiang, Derek Yang, Xu-anlin Li, Stone Tao, Zhiao Huang, Zhiwei Jia, and Hao Su. Maniskill: Generalizable manipulation skill benchmark with large-scale demonstrations. arXiv preprint arXiv:2107.14483, 2021. 3 +[44] Siva Karthik Mustikovela, Varun Jampani, Shalini De Mello, Sifei Liu, Umar Iqbal, Carsten Rother, and Jan Kautz. Self-supervised viewpoint learning from image collections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3971-3981, 2020. 3 +[45] Thu Nguyen-Phuoc, Chuan Li, Lucas Theis, Christian Richardt, and Yong-Liang Yang. Hologan: Unsupervised learning of 3d representations from natural images. In Proceedings of the IEEE International Conference on Computer Vision, pages 7588–7597, 2019. 3 +[46] Yiyuan Pan, Xuecheng Xu, Yue Wang, Xiaqing Ding, and Rong Xiong. GPU accelerated real-time traversability mapping. In 2019 IEEE International Conference on Robotics and Biomimetics (ROBIO), pages 734-740. IEEE, 2019. 2 +[47] Xue Bin Peng, Glen Berseth, KangKang Yin, and Michiel Van De Panne. Deeploco: Dynamic locomotion skills using hierarchical deep reinforcement learning. ACM Transactions on Graphics (TOG), 36(4):1-13, 2017. 2 +[48] Xue Bin Peng, Erwin Coumans, Tingnan Zhang, Tsang-Wei Edward Lee, Jie Tan, and Sergey Levine. Learning agile robotic locomotion skills by imitating animals. In Robotics: Science and Systems, 07 2020. 2 +[49] Xue Bin Peng, Ze Ma, Pieter Abbeel, Sergey Levine, and Angjoo Kanazawa. Amp: Adversarial motion priors for stylized physics-based character control. ACM Trans. Graph., 40(4), July 2021. 5 +[50] Lerrel Pinto, Marcin Andrychowicz, Peter Welinder, Wojciech Zaremba, and Pieter Abbeel. Asymmetric actor critic for image-based robot learning. arXiv preprint arXiv:1710.06542, 2017. 3 +[51] Marc H Raibert. Hopping in legged systems—modeling and simulation for the two-dimensional one-legged case. IEEE Transactions on Systems, Man, and Cybernetics, SMC-14(3):451–463, 1984. 2 +[52] Chris Rockwell, David F. Fouhey, and Justin Johnson. Pixelsynth: Generating a 3d-consistent experience from a single image. In ICCV, 2021. 3 +[53] Nikita Rudin, David Hoeller, Philipp Reist, and Marco Hutter. Learning to walk in minutes using massively parallel deep reinforcement learning. In Conference on Robot Learning, pages 91-100. PMLR, 2022. 2 +[54] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 5 +[55] Mingyo Seo, Ryan Gupta, Yifeng Zhu, Alexy Skoutnev, Luis Sentis, and Yuke Zhu. Learning to walk by steering: Perceptive quadrupedal locomotion in dynamic environments. In arXiv preprint arXiv:2209.09233, 2022. 2 + +[56] Vincent Sitzmann, Justus Thies, Felix Heide, Matthias Nießner, Gordon Wetzstein, and Michael Zollhofer. Deepvoxels: Learning persistent 3d feature embeddings. In CVPR, 2019. 3 +[57] Vincent Sitzmann, Michael Zollhöfer, and Gordon Wetzstein. Scene representation networks: Continuous 3d-structure-aware neural scene representations. In NeurIPS, 2019. 3 +[58] Maks Sorokin, Jie Tan, C. Karen Liu, and Sehoon Ha. Learning to navigate sidewalks in outdoor environments. IEEE Robotics and Automation Letters, 7(2):3906-3913, 2022. 2 +[59] Yu Sun, Wyatt L. Ubellacker, Wen-Loong Ma, Xiang Zhang, Changhao Wang, Noel V. Csomay-Shanklin, Masayoshi Tomizuka, Koushil Sreenath, and Aaron D. Ames. Online learning of unknown dynamics for model-based controllers in legged locomotion. IEEE Robotics and Automation Letters (RA-L), 2021. 2 +[60] Jie Tan, Tingnan Zhang, Erwin Coumans, Atil Iscen, Yunfei Bai, Danijar Hafner, Steven Bohez, and Vincent Vanhoucke. Sim-to-real: Learning agile locomotion for quadruped robots. In Hadas Kress-Gazit, Siddhartha S. Srinivasa, Tom Howard, and Nikolay Atanasov, editors, Robotics: Science and Systems XIV, Carnegie Mellon University, Pittsburgh, Pennsylvania, USA, June 26-30, 2018, 2018. 2 +[61] Nick Torkos and Michiel van de Panne. Footprint-based quadruped motion synthesis. In Proceedings of the Graphics Interface 1998 Conference, June 18-20, 1998, Vancouver, BC, Canada, pages 151-160, June 1998. 2 +[62] Shubham Tulsiani, Tinghui Zhou, Alexei A Efros, and Jitendra Malik. Multi-view supervision for single-view reconstruction via differentiable ray consistency. In CVPR, 2017. 3 +[63] Martin Wermelinger, Péter Fankhauser, Remo Diethelm, Philipp Krüsi, Roland Siegwart, and Marco Hutter. Navigation planning for legged robots in challenging terrain. In 2016 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1184-1189. IEEE, 2016. 2 +[64] Olivia Wiles, Georgia Gkioxari, Richard Szeliski, and Justin Johnson. Synsin: End-to-end view synthesis from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7467-7477, 2020. 3 +[65] Zhaoming Xie, Xingye Da, Buck Babich, Animesh Garg, and Michiel van de Panne. Glide: Generalizable quadrupedal locomotion in diverse environments with a centroidal model. CoRR, abs/2104.09771, 2021. 2 +[66] Zhaoming Xie, Hung Yu Ling, Nam Hee Kim, and Michiel van de Panne. Allsteps: curriculum-driven learning of stepping stone skills. In Computer Graphics Forum, volume 39, pages 213-224. Wiley Online Library, 2020. 2 +[67] Ruihan Yang, Minghao Zhang, Nicklas Hansen, Huazhe Xu, and Xiaolong Wang. Learning vision-guided quadrupedal locomotion end-to-end with cross-modal transformers. In International Conference on Learning Representations, 2022. 3 + +[68] KangKang Yin, Kevin Loken, and Michiel Van de Panne. Simbicon: Simple biped locomotion control. ACM Transactions on Graphics (TOG), 26(3):105-es, 2007. 2 +[69] Wenhao Yu, Deepali Jain, Alejandro Escontrela, Atil Iscen, Peng Xu, Erwin Coumans, Sehoon Ha, Jie Tan, and Tingnan Zhang. Visual-locomotion: Learning to walk on complex terrains with vision. In Aleksandra Faust, David Hsu, and Gerhard Neumann, editors, Proceedings of the 5th Conference on Robot Learning, volume 164 of Proceedings of Machine Learning Research, pages 1291-1302. PMLR, 08-11 Nov 2022. 2 +[70] Tinghui Zhou, Matthew Brown, Noah Snavely, and David G Lowe. Unsupervised learning of depth and ego-motion from video. In CVPR, 2017. 3 +[71] Matt Zucker, Nathan Ratliff, Martin Stolle, Joel Chestnutt, J Andrew Bagnell, Christopher G Atkeson, and James Kuffner. Optimization and learning for rough terrain legged locomotion. The International Journal of Robotics Research, 30(2):175-191, 2011. 2 \ No newline at end of file diff --git a/2023/Neural Volumetric Memory for Visual Locomotion Control/images.zip b/2023/Neural Volumetric Memory for Visual Locomotion Control/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..db8371209ca4c787f05b683ce976f42647f17018 --- /dev/null +++ b/2023/Neural Volumetric Memory for Visual Locomotion Control/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8ad1483094039d9f770122f90da7df9a6cd7e856a5dd8230e33894245018ef5d +size 596972 diff --git a/2023/Neural Volumetric Memory for Visual Locomotion Control/layout.json b/2023/Neural Volumetric Memory for Visual Locomotion Control/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d699e340cc57c8fa4e91f52be5bfd941ddfaac7f --- /dev/null +++ b/2023/Neural Volumetric Memory for Visual Locomotion Control/layout.json @@ -0,0 +1,9178 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 114, + 93, + 479, + 110 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 93, + 479, + 110 + ], + "spans": [ + { + "bbox": [ + 114, + 93, + 479, + 110 + ], + "type": "text", + "content": "Neural Volumetric Memory for Visual Locomotion Control" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 179, + 133, + 405, + 148 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 133, + 405, + 148 + ], + "spans": [ + { + "bbox": [ + 179, + 133, + 405, + 148 + ], + "type": "text", + "content": "Ruihan Yang1 Ge Yang2,3 Xiaolong Wang1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 110, + 148, + 481, + 162 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 148, + 481, + 162 + ], + "spans": [ + { + "bbox": [ + 110, + 148, + 481, + 162 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 110, + 148, + 481, + 162 + ], + "type": "text", + "content": "UC San Diego " + }, + { + "bbox": [ + 110, + 148, + 481, + 162 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 110, + 148, + 481, + 162 + ], + "type": "text", + "content": "Institute of AI and Fundamental Interactions " + }, + { + "bbox": [ + 110, + 148, + 481, + 162 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 110, + 148, + 481, + 162 + ], + "type": "text", + "content": "MIT CSAIL" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 49, + 171, + 546, + 327 + ], + "blocks": [ + { + "bbox": [ + 49, + 171, + 546, + 327 + ], + "lines": [ + { + "bbox": [ + 49, + 171, + 546, + 327 + ], + "spans": [ + { + "bbox": [ + 49, + 171, + 546, + 327 + ], + "type": "image", + "image_path": "37fbd98d0a807e7d96e00c56a762f543b5e37e2c7ebc8d833ab517e20a0e6662.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 335, + 546, + 359 + ], + "lines": [ + { + "bbox": [ + 46, + 335, + 546, + 359 + ], + "spans": [ + { + "bbox": [ + 46, + 335, + 546, + 359 + ], + "type": "text", + "content": "Figure 1. Real-World Deployment: We deploy our policy in real-world environments with Stones, Stairs, Stages, Obstacles, and Unstructured Terrain. Our robot utilizes the volumetric memory of the surrounding 3D structure to successfully traverse complex environments." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 378, + 192, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 378, + 192, + 392 + ], + "spans": [ + { + "bbox": [ + 143, + 378, + 192, + 392 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 404, + 289, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 404, + 289, + 656 + ], + "spans": [ + { + "bbox": [ + 46, + 404, + 289, + 656 + ], + "type": "text", + "content": "Legged robots have the potential to expand the reach of autonomy beyond paved roads. In this work, we consider the difficult problem of locomotion on challenging terrains using a single forward-facing depth camera. Due to the partial observability of the problem, the robot has to rely on past observations to infer the terrain currently beneath it. To solve this problem, we follow the paradigm in computer vision that explicitly models the 3D geometry of the scene and propose Neural Volumetric Memory (NVM), a geometric memory architecture that explicitly accounts for the SE(3) equivariance of the 3D world. NVM aggregates feature volumes from multiple camera views by first bringing them back to the ego-centric frame of the robot. We test the learned visual-locomotion policy on a physical robot and show that our approach, which explicitly introduces geometric priors during training, offers superior performance than more naive methods. We also include ablation studies and show that the representations stored in the neural volumetric memory capture sufficient geometric information to reconstruct the scene. Our project page with videos is https://rchalyang.github.io/NVM" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 668, + 128, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 668, + 128, + 681 + ], + "spans": [ + { + "bbox": [ + 47, + 668, + 128, + 681 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "content": "Consider difficult locomotion tasks such as walking up and down a flight of stairs and stepping over gaps (Figure 1)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 379, + 547, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 379, + 547, + 536 + ], + "spans": [ + { + "bbox": [ + 304, + 379, + 547, + 536 + ], + "type": "text", + "content": "The control of such behaviors requires tight coupling with perception because vision is needed to provide details of the terrain right beneath the robot and the 3D scene immediately around it. This problem is also partially-observable. Immediately relevant terrain information is often occluded from the robot's current frame of observation, forcing it to rely on past observations for control decisions. For this reason, while blind controllers that are learned in simulation using reinforcement learning have achieved impressive results in agility and robustness [33, 36, 38], there are clear limitations on how much they can do. How to incorporate perception into the pipeline to produce an integrated visuomotor controller thus remains an open problem." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 546, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 548, + 715 + ], + "type": "text", + "content": "A recent line of work combines perception with locomotion using ego-centric cameras mounted on the robot. The predominant approach for addressing partial observability is to do frame-stacking, where the robot maintains a visual buffer of recent images. This naive heuristic suffers from two major problems: first, frame-stacking on a moving robot ignores the equivariance structure of the 3D environment, making learning a lot more difficult as policy success now relies on being able to learn to account for spurious changes in camera poses. A second but subtler issue is that biological systems do not have the ability to save detailed visual observations pixel-by-pixel. These concerns motivate the creation of an intermediary, short-term memory mechanism to functionally aggregate streams of obser" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1430" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 69, + 547, + 190 + ], + "blocks": [ + { + "bbox": [ + 47, + 69, + 547, + 190 + ], + "lines": [ + { + "bbox": [ + 47, + 69, + 547, + 190 + ], + "spans": [ + { + "bbox": [ + 47, + 69, + 547, + 190 + ], + "type": "image", + "image_path": "c5c8690fe4b432b9bf5793d57eee7ca061c0c7956d1c08dd5da9ce40760374da.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 198, + 547, + 232 + ], + "lines": [ + { + "bbox": [ + 46, + 198, + 547, + 232 + ], + "spans": [ + { + "bbox": [ + 46, + 198, + 547, + 232 + ], + "type": "text", + "content": "Figure 2. Overview of Simulated Environment & Real World Environment: Our simulated environments are shown on the left and real-world environments are shown on the right. For the real-world environment, the corresponding visual observations for each real-world environment are shown in the bottom row. All policies are trained in the simulation and transferred into the real world without fine-tuning." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 241, + 279, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 241, + 279, + 253 + ], + "spans": [ + { + "bbox": [ + 47, + 241, + 279, + 253 + ], + "type": "text", + "content": "vation into a single, coherent representation of the world." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 255, + 289, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 255, + 289, + 460 + ], + "spans": [ + { + "bbox": [ + 46, + 255, + 289, + 460 + ], + "type": "text", + "content": "Motivated by these observations, we introduce a novel volumetric memory architecture for legged locomotion control. Our architecture consists of a 2D to 3D feature volume encoder, and a pose estimator that can estimate the relative camera poses given two input images. When combined, the two networks constitute a Neural Volumetric Memory (NVM) that takes as input a sequence of depth images taken by a forward-looking, ego-centric camera and fuses them together into a coherent latent representation for locomotion control. We encourage the memory space to be " + }, + { + "bbox": [ + 46, + 255, + 289, + 460 + ], + "type": "inline_equation", + "content": "SE(3)" + }, + { + "bbox": [ + 46, + 255, + 289, + 460 + ], + "type": "text", + "content": " equivariant to changes in the camera pose of the robot by incorporating translation and rotation operations based on estimated relative poses from the pose network. This inverse transformation allows NVM to align feature volumes from the past to the present ego-centric frame, making both integrating over multiple timesteps into a coherent scene representation and learning a policy, less difficult." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 463, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 463, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 463, + 289, + 715 + ], + "type": "text", + "content": "Our training pipeline follows a two-step teacher-student process where the primary goal of the first stage is to produce behaviors in the form of a policy. After training completes, this policy can traverse these difficult terrains robustly, but it relies on privileged sensory information such as an elevation map and ground-truth velocity. Elevation maps obtained in the real world are often biased, incomplete, and full of errors [40], whereas ground-truth velocity information is typically only available in instrumented environments. Hence in the visuomotor distillation stage of the pipeline, which still runs in the simulator, we feed the stream of ego-centric views from the forward depth camera into the neural volumetric memory. We feed the content of this memory into a small policy network and train everything end-to-end including the two network components of the NVM using a behavior cloning loss where the state-only policy acts as the teacher. For completeness, we offer an additional self-supervised learning objective (Figure 4) that relies on novel-view consistency for learning. The end product of this visuomotor distillation pipeline is a memory-equipped visuomotor policy that can operate directly on the" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 241, + 547, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 241, + 547, + 313 + ], + "spans": [ + { + "bbox": [ + 304, + 241, + 547, + 313 + ], + "type": "text", + "content": "UniTree A1 robot hardware (see Figure 1). A single policy is used to handle all environments covered by this paper. We provide comprehensive experiment and ablation studies in both simulation and the real world, and show that our method outperforms all baselines by a large margin. It is thus essential to model the 3D structure of the environment." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 326, + 392, + 338 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 326, + 392, + 338 + ], + "spans": [ + { + "bbox": [ + 305, + 326, + 392, + 338 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 346, + 547, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 346, + 547, + 552 + ], + "spans": [ + { + "bbox": [ + 304, + 346, + 547, + 552 + ], + "type": "text", + "content": "Blind Locomotion Controllers: Legged locomotion using the proprioception of the robot has been studied for decades [3, 16, 42, 51, 61, 68]. Traditionally, controllers are developed with model-based methods [2, 4, 5, 9, 10, 15, 17, 59]. However, most model-based methods fail to generalize to unseen environments in the real world. In order to attain better generalization and robustness, recent works [22, 24, 25, 32, 33, 36, 48, 60, 65] utilize model-free reinforcement learning to train neural network controllers in simulation and directly transfer the learned policies to the real robot. These methods are able to achieve high levels of agility and adapt over diverse terrains using just the proprioceptive state of the robot. The real-world utility of these robots, however, is also limited, since these blind controllers are unable to perceive the surrounding. In this paper, we propose to integrate perception to facilitate control, with a focus on a better 3D understanding of the environments." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 570, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 548, + 715 + ], + "type": "text", + "content": "Locomotion with Exteroceptive Perception: The elevation map [12, 13, 34, 46] around the robot has been used to perform optimization and planning for foot placement [8, 11, 47, 63, 66, 71]. Recent RL controller [40, 53] also uses elevation maps in an end-to-end manner. Using the elevation map to represent the surrounding is straightforward for foot placement optimization, but the elevation map obtained in the real world is either noisy or requires a complicated hardware pipeline. Another option is using depth or RGB images to provide visual observation for the controller. Hierarchical control systems [26, 39, 55, 58, 69] are proposed to utilize depth reading while the low-level" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "text", + "content": "1431" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 69, + 549, + 208 + ], + "blocks": [ + { + "bbox": [ + 48, + 69, + 549, + 208 + ], + "lines": [ + { + "bbox": [ + 48, + 69, + 549, + 208 + ], + "spans": [ + { + "bbox": [ + 48, + 69, + 549, + 208 + ], + "type": "image", + "image_path": "965c9e7a5b0ea50782226b52dd9d5ba31082b5c82f36c2aff77046d3fb38feaf.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 218, + 547, + 253 + ], + "lines": [ + { + "bbox": [ + 45, + 218, + 547, + 253 + ], + "spans": [ + { + "bbox": [ + 45, + 218, + 547, + 253 + ], + "type": "text", + "content": "Figure 3. Neural Volumetric Memory (NVM): Our NVM module extracts 3D feature volumes from every visual observation in the near past, transform them with corresponding estimated transformation into present robot frame, and fuse the transformed feature volumes into a neural volumetric memory for decision-making." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 265, + 289, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 265, + 289, + 411 + ], + "spans": [ + { + "bbox": [ + 46, + 265, + 289, + 411 + ], + "type": "text", + "content": "controllers still rely on only the proprioceptive state of the robot. To get the most out of the depth image, Yang et al. [67] and Imai et al. [23] propose to learn visual RL policy to output the raw joint angles and enable the robot to maneuver in the wild avoiding different kinds of obstacles. Agarwal et al. [1] and Loquercioo et al. [37] train a policy using RL to traverse a large variety of terrains. Though these works show encouraging results in specific tasks, they have not fully exploited the 3D structure of the environment. On the other hand, we propose to train a novel neural volumetric memory together with policy learning, which allows our robot to walk through more challenging terrains." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 423, + 289, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 423, + 289, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 423, + 289, + 605 + ], + "type": "text", + "content": "Volumetric Cognitive Map: How to represent and understand the 3D scene has been studied by the computer vision community for a long time. Hartley et al. [20] and Wiles et al. [64] provide representation built on geometry principles and camera motions. Learning-based approaches are proposed [21, 28, 29, 35, 52, 62, 64, 70] to learn 3D representation from 2D inputs. Neural 3D representations [14, 19, 41, 44, 45, 56, 57] has also shown great result in multiple computer vision tasks [14, 18, 19]. Our work is closely related to Lai et al. [35] where the deep voxel feature representing the entire scene and the camera trajectory are learned from video clips without odometry in a self-supervised manner. Instead of focusing on view synthesis, we aggregate the 3D feature volume to form a short-term memory representation for decision-making." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 618, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 289, + 715 + ], + "type": "text", + "content": "Teacher-Student Training: To accelerate training and utilize the privileged information in simulation, Miki et al. [40] follow a teacher-student training framework, where the teacher policy is trained with privileged information and then distilled into student policy where the privileged information is not available. Multiple works in different research domains [6, 7, 27, 30, 31, 43, 50] follow this framework as well. We share a similar pipeline as [1, 39] where elevation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 265, + 545, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 265, + 545, + 291 + ], + "spans": [ + { + "bbox": [ + 305, + 265, + 545, + 291 + ], + "type": "text", + "content": "map-based policies are trained first and the elevation-map-based policies are distilled visual policies afterward." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 301, + 545, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 301, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 305, + 301, + 545, + 316 + ], + "type": "text", + "content": "3. Volumetric Memory for Legged Locomotion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 323, + 545, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 323, + 545, + 418 + ], + "spans": [ + { + "bbox": [ + 304, + 323, + 545, + 418 + ], + "type": "text", + "content": "Legged locomotion using ego-centric camera views is intrinsically a partially-observed problem. To make the control problem tangible, our robot needs to aggregate information from previous frames and correctly infer the occluded terrain underneath it. During locomotion, the camera mounted directly on the robot chassis undergo large and spurious changes in pose, making integrating individual frames to a coherent representation non-trivial." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 419, + 547, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 419, + 547, + 479 + ], + "spans": [ + { + "bbox": [ + 304, + 419, + 547, + 479 + ], + "type": "text", + "content": "To account for these camera pose changes, we propose neural volumetric memory (NVM) — a 3D representation format for scene features. It takes as input a sequence of visual observations and outputs a single 3D feature volume representing the surrounding 3D structure." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 488, + 515, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 488, + 515, + 502 + ], + "spans": [ + { + "bbox": [ + 305, + 488, + 515, + 502 + ], + "type": "text", + "content": "3.1. Baking-in " + }, + { + "bbox": [ + 305, + 488, + 515, + 502 + ], + "type": "inline_equation", + "content": "SE(3)" + }, + { + "bbox": [ + 305, + 488, + 515, + 502 + ], + "type": "text", + "content": " Equivariance in NVM" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 508, + 545, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 508, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 508, + 545, + 555 + ], + "type": "text", + "content": "Our training procedure encourages the learned memory space to be " + }, + { + "bbox": [ + 304, + 508, + 545, + 555 + ], + "type": "inline_equation", + "content": "SE(3)" + }, + { + "bbox": [ + 304, + 508, + 545, + 555 + ], + "type": "text", + "content": " equivariant by applying translation and rotation transformations on the feature volume before aggregating additional information from new observations." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 556, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 556, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 556, + 547, + 715 + ], + "type": "text", + "content": "As the first step (Figure 3), an encoder network receives a sequence of consecutive observations " + }, + { + "bbox": [ + 304, + 556, + 547, + 715 + ], + "type": "inline_equation", + "content": "O_{t}..O_{t - n + 1}" + }, + { + "bbox": [ + 304, + 556, + 547, + 715 + ], + "type": "text", + "content": ". Instead of generating a compact feature embedding in the shape of a tuple, the 3D encoder Enc3D extracts feature volumes " + }, + { + "bbox": [ + 304, + 556, + 547, + 715 + ], + "type": "inline_equation", + "content": "V_{t}.V_{t - n + 1}" + }, + { + "bbox": [ + 304, + 556, + 547, + 715 + ], + "type": "text", + "content": ". A second network Encpose estimates the relative changes in camera pose between the present frame " + }, + { + "bbox": [ + 304, + 556, + 547, + 715 + ], + "type": "inline_equation", + "content": "O_{t}^{v}" + }, + { + "bbox": [ + 304, + 556, + 547, + 715 + ], + "type": "text", + "content": " and frames " + }, + { + "bbox": [ + 304, + 556, + 547, + 715 + ], + "type": "inline_equation", + "content": "O_{t}^{v}.O_{t - n + 1}^{v}" + }, + { + "bbox": [ + 304, + 556, + 547, + 715 + ], + "type": "text", + "content": " from the past. This sequence of estimate allows NVM to transform the feature volume " + }, + { + "bbox": [ + 304, + 556, + 547, + 715 + ], + "type": "inline_equation", + "content": "V_{t}.V_{t - n + 1}" + }, + { + "bbox": [ + 304, + 556, + 547, + 715 + ], + "type": "text", + "content": " from the past to the present frame using the transformations " + }, + { + "bbox": [ + 304, + 556, + 547, + 715 + ], + "type": "inline_equation", + "content": "T_{t}^{t}.T_{t - n + 1}^{t}" + }, + { + "bbox": [ + 304, + 556, + 547, + 715 + ], + "type": "text", + "content": ". We annotate the transformed feature volume as " + }, + { + "bbox": [ + 304, + 556, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\hat{V}_{t}. \\hat{V}_{t - n + 1}" + }, + { + "bbox": [ + 304, + 556, + 547, + 715 + ], + "type": "text", + "content": ". These canonicalized volumes can now be aggregated together by element-wise sum to produce a fused feature volume " + }, + { + "bbox": [ + 304, + 556, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\hat{V}_{t}^{m}" + }, + { + "bbox": [ + 304, + 556, + 547, + 715 + ], + "type": "text", + "content": " for the surround-" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1432" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "ing 3D structure. This scheme allows NVM to disentangle the symmetry structure of the ambient 3D space from the \"content\" specific to the scene. By doing so, we encourage " + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "SE(3)" + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": " equivariance in the memory space." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 120, + 287, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 120, + 287, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 120, + 287, + 144 + ], + "type": "text", + "content": "We provide a detailed description of all components in our NVM module in the remaining section." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 144, + 288, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 144, + 288, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 144, + 288, + 251 + ], + "type": "text", + "content": "3D Encoder: Our 3D encoder " + }, + { + "bbox": [ + 46, + 144, + 288, + 251 + ], + "type": "inline_equation", + "content": "Enc_{3D}" + }, + { + "bbox": [ + 46, + 144, + 288, + 251 + ], + "type": "text", + "content": " takes a 2D depth image as input and outputs a 3D feature volume. We parameterize this encoder with a set of 2D convolution layers followed by 3D convolution layers. Specifically, we generate a 2D feature map with the shape " + }, + { + "bbox": [ + 46, + 144, + 288, + 251 + ], + "type": "inline_equation", + "content": "(C, H, W)" + }, + { + "bbox": [ + 46, + 144, + 288, + 251 + ], + "type": "text", + "content": " using 2D convolution, then reshape the feature map channel-wise into a volume tensor with the shape " + }, + { + "bbox": [ + 46, + 144, + 288, + 251 + ], + "type": "inline_equation", + "content": "(C / D, D, H, W)" + }, + { + "bbox": [ + 46, + 144, + 288, + 251 + ], + "type": "text", + "content": ". We chose a voxel-grid size of " + }, + { + "bbox": [ + 46, + 144, + 288, + 251 + ], + "type": "inline_equation", + "content": "D = 12" + }, + { + "bbox": [ + 46, + 144, + 288, + 251 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 144, + 288, + 251 + ], + "type": "inline_equation", + "content": "H = 6" + }, + { + "bbox": [ + 46, + 144, + 288, + 251 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 144, + 288, + 251 + ], + "type": "inline_equation", + "content": "W = 6" + }, + { + "bbox": [ + 46, + 144, + 288, + 251 + ], + "type": "text", + "content": ". The 3D convolution layers are added, to further refine the 3D features." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 90, + 260, + 244, + 274 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 260, + 244, + 274 + ], + "spans": [ + { + "bbox": [ + 90, + 260, + 244, + 274 + ], + "type": "interline_equation", + "content": "V _ {i} = E n c _ {3 D} (O _ {i} ^ {v}) \\quad i = t.. t - n + 1", + "image_path": "db47828aca6b38e9bfec336d2adc4cf2dd036579dd6619fe18dc27cf9bd8054d.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 282, + 287, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 282, + 287, + 390 + ], + "spans": [ + { + "bbox": [ + 46, + 282, + 287, + 390 + ], + "type": "text", + "content": "Pose Encoder: Our pose encoder " + }, + { + "bbox": [ + 46, + 282, + 287, + 390 + ], + "type": "inline_equation", + "content": "Enc_{\\mathrm{pose}}" + }, + { + "bbox": [ + 46, + 282, + 287, + 390 + ], + "type": "text", + "content": " takes two depth images and estimates the transformation in the latent space between two frames. Specifically, our pose encoder outputs 6D vectors representing the camera pose (rotation and translation) and we parameterize the pose encoder with a set of 2D convs. When constructing the NVM, the pose encoder is used to estimate the transformation " + }, + { + "bbox": [ + 46, + 282, + 287, + 390 + ], + "type": "inline_equation", + "content": "T_{t}^{t}\\ldots T_{t - n + 1}^{t}" + }, + { + "bbox": [ + 46, + 282, + 287, + 390 + ], + "type": "text", + "content": " between previous frames and the present frame by comparing previous visual observation " + }, + { + "bbox": [ + 46, + 282, + 287, + 390 + ], + "type": "inline_equation", + "content": "O_{t}^{v}\\ldots O_{t - n + 1}^{v}" + }, + { + "bbox": [ + 46, + 282, + 287, + 390 + ], + "type": "text", + "content": " and present " + }, + { + "bbox": [ + 46, + 282, + 287, + 390 + ], + "type": "inline_equation", + "content": "O_{t}^{v}" + }, + { + "bbox": [ + 46, + 282, + 287, + 390 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 98, + 399, + 235, + 414 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 98, + 399, + 235, + 414 + ], + "spans": [ + { + "bbox": [ + 98, + 399, + 235, + 414 + ], + "type": "interline_equation", + "content": "T _ {i} ^ {j} = \\operatorname {E n c} _ {\\text {p o s e}} \\left(O _ {i} ^ {v}, O _ {j} ^ {v}\\right) \\in S E (3)", + "image_path": "07e32a1a8edfdd574070b15b0739c94b15fe1287f1657745e3812b444b045186.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 421, + 287, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 421, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 421, + 287, + 495 + ], + "type": "text", + "content": "Fusing 3D feature: Given feature volumes from different frames and the estimated transformation from different frames to the present frame. We transform the feature volume " + }, + { + "bbox": [ + 46, + 421, + 287, + 495 + ], + "type": "inline_equation", + "content": "V_{t}..V_{t - n + 1}" + }, + { + "bbox": [ + 46, + 421, + 287, + 495 + ], + "type": "text", + "content": " with the corresponding transformation " + }, + { + "bbox": [ + 46, + 421, + 287, + 495 + ], + "type": "inline_equation", + "content": "T_{t}^{t}.T_{t - n + 1}^{t}" + }, + { + "bbox": [ + 46, + 421, + 287, + 495 + ], + "type": "text", + "content": ". Particularly, for each position " + }, + { + "bbox": [ + 46, + 421, + 287, + 495 + ], + "type": "inline_equation", + "content": "p = (i,j,k)^T" + }, + { + "bbox": [ + 46, + 421, + 287, + 495 + ], + "type": "text", + "content": " in 3D deep voxel, we transform it to " + }, + { + "bbox": [ + 46, + 421, + 287, + 495 + ], + "type": "inline_equation", + "content": "\\hat{p} = (\\hat{i},\\hat{j},\\hat{k})^T" + }, + { + "bbox": [ + 46, + 421, + 287, + 495 + ], + "type": "text", + "content": " with:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 141, + 504, + 192, + 516 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 504, + 192, + 516 + ], + "spans": [ + { + "bbox": [ + 141, + 504, + 192, + 516 + ], + "type": "interline_equation", + "content": "\\hat {p} = R p + t", + "image_path": "d2d3e81a113cb8065bd9b8e30d4dfcda4ac4dcdb2b94367545faafc7570988db.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 525, + 287, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 525, + 287, + 561 + ], + "spans": [ + { + "bbox": [ + 46, + 525, + 287, + 561 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 525, + 287, + 561 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 46, + 525, + 287, + 561 + ], + "type": "text", + "content": " is a 3x3 rotation matrix from " + }, + { + "bbox": [ + 46, + 525, + 287, + 561 + ], + "type": "inline_equation", + "content": "T_{t}^{t}\\ldots T_{t - n + 1}^{t}" + }, + { + "bbox": [ + 46, + 525, + 287, + 561 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 525, + 287, + 561 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 525, + 287, + 561 + ], + "type": "text", + "content": " is the translation vector from " + }, + { + "bbox": [ + 46, + 525, + 287, + 561 + ], + "type": "inline_equation", + "content": "T_{t}^{t}\\ldots T_{t - n + 1}^{t}" + }, + { + "bbox": [ + 46, + 525, + 287, + 561 + ], + "type": "text", + "content": ", and we note the transformation function as " + }, + { + "bbox": [ + 46, + 525, + 287, + 561 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 525, + 287, + 561 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 97, + 568, + 235, + 582 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 568, + 235, + 582 + ], + "spans": [ + { + "bbox": [ + 97, + 568, + 235, + 582 + ], + "type": "interline_equation", + "content": "\\hat {V} _ {i} = f \\left(V _ {i}, T _ {i} ^ {t}\\right) \\quad i = t.. t - n + 1", + "image_path": "634208e5c4c8a2b3df7e0e706f53481731dc01d3d4f10dd66665e754c6ce2faf.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 591, + 287, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 591, + 287, + 675 + ], + "spans": [ + { + "bbox": [ + 46, + 591, + 287, + 675 + ], + "type": "text", + "content": "Since directly performing transformation in latent space might cause misalignment due to the coarse of voxel representation, we use two additional 3D conv layers to refine the transformed feature volume. We then aggregate transformed feature volumes to get Neural Volumetric Memory " + }, + { + "bbox": [ + 46, + 591, + 287, + 675 + ], + "type": "inline_equation", + "content": "\\hat{V}_t^m" + }, + { + "bbox": [ + 46, + 591, + 287, + 675 + ], + "type": "text", + "content": " representing the surrounding 3D structure in the present frame by computing the mean over feature volumes:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 123, + 683, + 211, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 683, + 211, + 715 + ], + "spans": [ + { + "bbox": [ + 123, + 683, + 211, + 715 + ], + "type": "interline_equation", + "content": "\\hat {V} _ {t} ^ {m} = \\frac {1}{n} \\sum_ {i = t - n + 1} ^ {t} \\hat {V} _ {i}", + "image_path": "f7f9be0dda0046c3824946a5dff19b7eb0ef680ba91b3d3d2fb00e109a64709a.jpg" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 307, + 71, + 549, + 167 + ], + "blocks": [ + { + "bbox": [ + 307, + 71, + 549, + 167 + ], + "lines": [ + { + "bbox": [ + 307, + 71, + 549, + 167 + ], + "spans": [ + { + "bbox": [ + 307, + 71, + 549, + 167 + ], + "type": "image", + "image_path": "bd9f9f00c1227a7833e3ae53f3cef80a4864abbc2bbb66bf80e951cd3f47fd1b.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 179, + 547, + 223 + ], + "lines": [ + { + "bbox": [ + 304, + 179, + 547, + 223 + ], + "spans": [ + { + "bbox": [ + 304, + 179, + 547, + 223 + ], + "type": "text", + "content": "Figure 4. Self-Supervised Learning: we train a separate decoder to predict the visual observation in the different frames given one visual observation and the estimated transformation between two frames, similar to Lai et al. [35]." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 242, + 496, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 242, + 496, + 255 + ], + "spans": [ + { + "bbox": [ + 306, + 242, + 496, + 255 + ], + "type": "text", + "content": "3.2. Learning NVM via Self-Supervision" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 260, + 545, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 260, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 304, + 260, + 545, + 392 + ], + "type": "text", + "content": "Although the behavior cloning objective is sufficient in producing a good policy, being equivariant to translation and rotation automatically offers a standalone, self-supervised learning objective just for the neural volumetric memory. Shown in Figure 4, one can assume that the surrounding 3D scene remains static between frames. Because the camera is forward-looking, we can canonicalized feature volume from earlier frames, and use it to predict a camera image taken at later steps. This self-consistency loss, computed as the pixel reconstruction error, encourages the memory to be " + }, + { + "bbox": [ + 304, + 260, + 545, + 392 + ], + "type": "inline_equation", + "content": "SE(3)" + }, + { + "bbox": [ + 304, + 260, + 545, + 392 + ], + "type": "text", + "content": " equivariant." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 392, + 546, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 392, + 546, + 645 + ], + "spans": [ + { + "bbox": [ + 304, + 392, + 546, + 645 + ], + "type": "text", + "content": "For given visual observation history over a short span of time " + }, + { + "bbox": [ + 304, + 392, + 546, + 645 + ], + "type": "inline_equation", + "content": "O_{t}^{v}..O_{t - n + 1}^{v}" + }, + { + "bbox": [ + 304, + 392, + 546, + 645 + ], + "type": "text", + "content": ", we use the most dated visual observation " + }, + { + "bbox": [ + 304, + 392, + 546, + 645 + ], + "type": "inline_equation", + "content": "O_{t - n + 1}^{v}" + }, + { + "bbox": [ + 304, + 392, + 546, + 645 + ], + "type": "text", + "content": " in the visual history and predict all frames. We extract a 3D feature volume from " + }, + { + "bbox": [ + 304, + 392, + 546, + 645 + ], + "type": "inline_equation", + "content": "O_{t - n + 1}^{v}: V_{t - n + 1} = Enc_{3D}(O_{t - n + 1}^{v})" + }, + { + "bbox": [ + 304, + 392, + 546, + 645 + ], + "type": "text", + "content": ", estimate the relative camera pose " + }, + { + "bbox": [ + 304, + 392, + 546, + 645 + ], + "type": "inline_equation", + "content": "T_{t}^{t - n + 1}..T_{t - n + 1}^{t - n + 1}" + }, + { + "bbox": [ + 304, + 392, + 546, + 645 + ], + "type": "text", + "content": " between " + }, + { + "bbox": [ + 304, + 392, + 546, + 645 + ], + "type": "inline_equation", + "content": "O_{t}^{v}.O_{t - n + 1}^{v}" + }, + { + "bbox": [ + 304, + 392, + 546, + 645 + ], + "type": "text", + "content": " and the most dated " + }, + { + "bbox": [ + 304, + 392, + 546, + 645 + ], + "type": "inline_equation", + "content": "O_{t - n + 1}^{v}" + }, + { + "bbox": [ + 304, + 392, + 546, + 645 + ], + "type": "text", + "content": ". We transform the extracted 3D feature volume " + }, + { + "bbox": [ + 304, + 392, + 546, + 645 + ], + "type": "inline_equation", + "content": "V_{t - n + 1}" + }, + { + "bbox": [ + 304, + 392, + 546, + 645 + ], + "type": "text", + "content": " into different frames with the corresponding transformations and then use a decoder " + }, + { + "bbox": [ + 304, + 392, + 546, + 645 + ], + "type": "inline_equation", + "content": "Dec" + }, + { + "bbox": [ + 304, + 392, + 546, + 645 + ], + "type": "text", + "content": " to predict the visual observation " + }, + { + "bbox": [ + 304, + 392, + 546, + 645 + ], + "type": "inline_equation", + "content": "\\hat{O}_{t..t - n + 1}^{v}" + }, + { + "bbox": [ + 304, + 392, + 546, + 645 + ], + "type": "text", + "content": " rendered in different frames. The Decoder " + }, + { + "bbox": [ + 304, + 392, + 546, + 645 + ], + "type": "inline_equation", + "content": "Dec" + }, + { + "bbox": [ + 304, + 392, + 546, + 645 + ], + "type": "text", + "content": " takes a deep 3D feature volume as input and outputs a 2D image. The decoder first reshapes the deep 3D feature volume with shape " + }, + { + "bbox": [ + 304, + 392, + 546, + 645 + ], + "type": "inline_equation", + "content": "(C,D,H,W)" + }, + { + "bbox": [ + 304, + 392, + 546, + 645 + ], + "type": "text", + "content": " into shape " + }, + { + "bbox": [ + 304, + 392, + 546, + 645 + ], + "type": "inline_equation", + "content": "(C*D,H,W)" + }, + { + "bbox": [ + 304, + 392, + 546, + 645 + ], + "type": "text", + "content": " and applies a set of 2D conv layers to map the reshaped features to images. Given the predicted visual observation " + }, + { + "bbox": [ + 304, + 392, + 546, + 645 + ], + "type": "inline_equation", + "content": "\\hat{O}_{t..t - n + 1}^{v}" + }, + { + "bbox": [ + 304, + 392, + 546, + 645 + ], + "type": "text", + "content": " and the original visual observation " + }, + { + "bbox": [ + 304, + 392, + 546, + 645 + ], + "type": "inline_equation", + "content": "O_{t}^{v}.O_{t - n + 1}^{v}" + }, + { + "bbox": [ + 304, + 392, + 546, + 645 + ], + "type": "text", + "content": " in different frames, we compute the prediction loss with pixel-wise L1 loss. The 3D feature encoder and the pose encoder are both trained in this predictive self-supervised training task. Specifically, we can formulate this 3D representation learning task as follows:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 650, + 536, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 650, + 536, + 715 + ], + "spans": [ + { + "bbox": [ + 314, + 650, + 536, + 715 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} V _ {t - n + 1} = \\operatorname {E n c} _ {3 D} \\left(O _ {t - n + 1} ^ {v}\\right) \\\\ \\hat {O} _ {i} ^ {v} = \\operatorname {D e c} \\left(f \\left(V _ {t - n + 1}, T _ {i} ^ {t - n + 1}\\right)\\right) i = t.. t - n + 1 \\\\ L _ {r e c} = \\frac {1}{n} \\sum_ {i = t} ^ {t - n + 1} \\left| O _ {i} ^ {v} - \\hat {O} _ {i} ^ {v} \\right| \\\\ \\end{array}", + "image_path": "50331e84dc4539f823381cb8a6eb9afe8d3c57ab4854d09694a1274c33233c81.jpg" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1433" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "inline_equation", + "content": "L_{rec}" + }, + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "content": " represents the L1 loss. Since the information contained in different visual observations is naturally different, perfect prediction in our task is not feasible. However, such a predictive learning task could still improve the 3D structural understanding of the surrounding and encourages the SE(3) equivariance over the learned 3D latent space, for this reconstruction loss encourages the disentanglement between the surrounding 3D structure and the camera pose." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 175, + 272, + 188 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 175, + 272, + 188 + ], + "spans": [ + { + "bbox": [ + 47, + 175, + 272, + 188 + ], + "type": "text", + "content": "3.3. Teacher Policy and Visuomotor Distillation" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 194, + 288, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 194, + 288, + 243 + ], + "spans": [ + { + "bbox": [ + 46, + 194, + 288, + 243 + ], + "type": "text", + "content": "Image observations introduce additional overhead and slow training down as much as 40 times1. For this reason, we bootstrap learning the visuomotor policy by first training a state-space policy that depends on the elevation map." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 257, + 287, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 257, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 46, + 257, + 287, + 342 + ], + "type": "text", + "content": "Learning a state-only policy We use proximal policy optimization (PPO [54]) to train a single teacher policy " + }, + { + "bbox": [ + 46, + 257, + 287, + 342 + ], + "type": "inline_equation", + "content": "\\hat{\\pi}" + }, + { + "bbox": [ + 46, + 257, + 287, + 342 + ], + "type": "text", + "content": " in all four environments. We parameterize the teacher policy as a 4-layer ReLU network with [256, 128, 128, 64] latent neurons. To provide terrain information, we include two elevation maps, one larger and coarser than the other. Overall, the observation space for the teacher policy include:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 348, + 288, + 494 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 47, + 348, + 287, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 348, + 287, + 386 + ], + "spans": [ + { + "bbox": [ + 47, + 348, + 287, + 386 + ], + "type": "text", + "content": "- proprioceptive state - " + }, + { + "bbox": [ + 47, + 348, + 287, + 386 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{63}" + }, + { + "bbox": [ + 47, + 348, + 287, + 386 + ], + "type": "text", + "content": " A 63-dimension vector consists of the gravity vector in the body frame " + }, + { + "bbox": [ + 47, + 348, + 287, + 386 + ], + "type": "inline_equation", + "content": "\\mathbf{g}_t" + }, + { + "bbox": [ + 47, + 348, + 287, + 386 + ], + "type": "text", + "content": ", joint rotations " + }, + { + "bbox": [ + 47, + 348, + 287, + 386 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_t" + }, + { + "bbox": [ + 47, + 348, + 287, + 386 + ], + "type": "text", + "content": ", joint velocity " + }, + { + "bbox": [ + 47, + 348, + 287, + 386 + ], + "type": "inline_equation", + "content": "\\dot{\\mathbf{q}}_t" + }, + { + "bbox": [ + 47, + 348, + 287, + 386 + ], + "type": "text", + "content": ", and previous actions " + }, + { + "bbox": [ + 47, + 348, + 287, + 386 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_{t-1}" + }, + { + "bbox": [ + 47, + 348, + 287, + 386 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 388, + 288, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 388, + 288, + 425 + ], + "spans": [ + { + "bbox": [ + 47, + 388, + 288, + 425 + ], + "type": "text", + "content": "- privileged information - " + }, + { + "bbox": [ + 47, + 388, + 288, + 425 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{14}" + }, + { + "bbox": [ + 47, + 388, + 288, + 425 + ], + "type": "text", + "content": " A 14-dimension vector including linear velocity " + }, + { + "bbox": [ + 47, + 388, + 288, + 425 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_t" + }, + { + "bbox": [ + 47, + 388, + 288, + 425 + ], + "type": "text", + "content": " and angular velocity " + }, + { + "bbox": [ + 47, + 388, + 288, + 425 + ], + "type": "inline_equation", + "content": "\\omega_t" + }, + { + "bbox": [ + 47, + 388, + 288, + 425 + ], + "type": "text", + "content": " of the robot, and environment parameters " + }, + { + "bbox": [ + 47, + 388, + 288, + 425 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_t" + }, + { + "bbox": [ + 47, + 388, + 288, + 425 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 428, + 287, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 428, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 47, + 428, + 287, + 464 + ], + "type": "text", + "content": "- a dense elevation map " + }, + { + "bbox": [ + 47, + 428, + 287, + 464 + ], + "type": "inline_equation", + "content": "m_{dense} - \\mathbb{R}^{21 \\times 26}" + }, + { + "bbox": [ + 47, + 428, + 287, + 464 + ], + "type": "text", + "content": " with 5cm spacing between the grid points, which is only slightly bigger than the robot." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 468, + 287, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 468, + 287, + 494 + ], + "spans": [ + { + "bbox": [ + 47, + 468, + 287, + 494 + ], + "type": "text", + "content": "- a sparse elevation map " + }, + { + "bbox": [ + 47, + 468, + 287, + 494 + ], + "type": "inline_equation", + "content": "m_{sparse} - \\mathbb{R}^{10 \\times 19}" + }, + { + "bbox": [ + 47, + 468, + 287, + 494 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 47, + 468, + 287, + 494 + ], + "type": "inline_equation", + "content": "10\\mathrm{cm}" + }, + { + "bbox": [ + 47, + 468, + 287, + 494 + ], + "type": "text", + "content": " spacing. This covers a much larger area." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 46, + 501, + 287, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 501, + 287, + 526 + ], + "spans": [ + { + "bbox": [ + 46, + 501, + 287, + 526 + ], + "type": "text", + "content": "We provide a more detailed description of the observation in the appendix." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 540, + 287, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 540, + 287, + 648 + ], + "spans": [ + { + "bbox": [ + 46, + 540, + 287, + 648 + ], + "type": "text", + "content": "Reward function : We use the same reward function for teacher's training in all environments. The reward contains the following terms: 1) Reward for moving forward at a target velocity (in our experiment: " + }, + { + "bbox": [ + 46, + 540, + 287, + 648 + ], + "type": "inline_equation", + "content": "0.4\\mathrm{m / s}" + }, + { + "bbox": [ + 46, + 540, + 287, + 648 + ], + "type": "text", + "content": "). 2) an energy penalty to encourage the robot to use as little torque as possible. 3) Height tracking reward encouraging the robot to walk while maintaining a specific relative height to the ground (in our experiment: " + }, + { + "bbox": [ + 46, + 540, + 287, + 648 + ], + "type": "inline_equation", + "content": "0.265\\mathrm{m}" + }, + { + "bbox": [ + 46, + 540, + 287, + 648 + ], + "type": "text", + "content": "). 4) AMP reward [49] encouraging the agent to produce natural gaits." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 83, + 657, + 250, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 657, + 250, + 685 + ], + "spans": [ + { + "bbox": [ + 83, + 657, + 250, + 685 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} R = \\alpha_ {\\text {f o r w a r d}} * R _ {\\text {f o r w a r d}} + \\alpha_ {\\text {e n e r g y}} * R _ {\\text {e n e r g y}} \\\\ + \\alpha_ {\\text {h e i g h t}} * R _ {\\text {h e i g h t}} + \\alpha_ {\\text {a m p}} * R _ {\\text {a m p}} \\\\ \\end{array}", + "image_path": "b741ac7f18197ff4017433ddf08223b1f271fc1e0b146f10946e538af9e1a44b.jpg" + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 308, + 95, + 541, + 155 + ], + "blocks": [ + { + "bbox": [ + 305, + 70, + 545, + 93 + ], + "lines": [ + { + "bbox": [ + 305, + 70, + 545, + 93 + ], + "spans": [ + { + "bbox": [ + 305, + 70, + 545, + 93 + ], + "type": "text", + "content": "Table 1. Comparison between observation spaces. State-only policy (w/ elevation-map) vs. vision policy" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 95, + 541, + 155 + ], + "lines": [ + { + "bbox": [ + 308, + 95, + 541, + 155 + ], + "spans": [ + { + "bbox": [ + 308, + 95, + 541, + 155 + ], + "type": "table", + "html": "
Joint statesVelocityElevation mapDepth map
Teacher policy✓✓✓×
Visual policy✓××✓
", + "image_path": "72a857023805d5200176ef101d101571b8774c8154bb1500b31a118ea30c13ed.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 168, + 545, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 168, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 305, + 168, + 545, + 194 + ], + "type": "text", + "content": "We include the details of each reward term and the corresponding weights in the appendix" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 206, + 545, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 206, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 304, + 206, + 545, + 361 + ], + "type": "text", + "content": "Visuomotor distillation: Given the privileged teacher policy from stage one, we then distill the privileged teacher policy " + }, + { + "bbox": [ + 304, + 206, + 545, + 361 + ], + "type": "inline_equation", + "content": "\\hat{\\pi}" + }, + { + "bbox": [ + 304, + 206, + 545, + 361 + ], + "type": "text", + "content": " into the visual policy " + }, + { + "bbox": [ + 304, + 206, + 545, + 361 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 304, + 206, + 545, + 361 + ], + "type": "text", + "content": " without privileged information. Our visual policy utilizes proprioceptive state and visual observation for decision-making. The proprioceptive state " + }, + { + "bbox": [ + 304, + 206, + 545, + 361 + ], + "type": "inline_equation", + "content": "O_{prop}" + }, + { + "bbox": [ + 304, + 206, + 545, + 361 + ], + "type": "text", + "content": " is the same as we used to train privileged elevation-based policy in stage one while the visual observation contains a sequence of depth images (" + }, + { + "bbox": [ + 304, + 206, + 545, + 361 + ], + "type": "inline_equation", + "content": "O_{t}^{v}..O_{t-n+1}^{v}" + }, + { + "bbox": [ + 304, + 206, + 545, + 361 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 206, + 545, + 361 + ], + "type": "inline_equation", + "content": "n=5" + }, + { + "bbox": [ + 304, + 206, + 545, + 361 + ], + "type": "text", + "content": " in our experiment) with a shape " + }, + { + "bbox": [ + 304, + 206, + 545, + 361 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 304, + 206, + 545, + 361 + ], + "type": "text", + "content": " from a depth camera mounted on the head of the robot. The observation space comparison between the privileged teacher policy and the visual policy is provided in Table 1, and the specific position of the camera is provided in Figure 2." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 361, + 545, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 361, + 545, + 481 + ], + "spans": [ + { + "bbox": [ + 304, + 361, + 545, + 481 + ], + "type": "text", + "content": "As shown in Figure 3, our visual policy first processes the sequential visual observation with our proposed NVM module. After obtaining NVM with shape " + }, + { + "bbox": [ + 304, + 361, + 545, + 481 + ], + "type": "inline_equation", + "content": "(C,D,H,W)" + }, + { + "bbox": [ + 304, + 361, + 545, + 481 + ], + "type": "text", + "content": ", our visual policy flattens the NVM into a 2D feature map with shape " + }, + { + "bbox": [ + 304, + 361, + 545, + 481 + ], + "type": "inline_equation", + "content": "(C\\times D,H,W)" + }, + { + "bbox": [ + 304, + 361, + 545, + 481 + ], + "type": "text", + "content": " and utilizes a set of 2D conv layers and a single linear layer to get a 1D visual feature. Our visual policy encodes the proprioceptive state with a neural network and gets the proprioceptive feature. The proprioceptive feature and visual feature are then concatenated and fed into a final network to get the actual action for the robot." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 481, + 545, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 481, + 545, + 518 + ], + "spans": [ + { + "bbox": [ + 304, + 481, + 545, + 518 + ], + "type": "text", + "content": "The visual policy is trained with behavior cloning by minimizing the L1 distance between actions from teacher policy " + }, + { + "bbox": [ + 304, + 481, + 545, + 518 + ], + "type": "inline_equation", + "content": "\\hat{\\pi}" + }, + { + "bbox": [ + 304, + 481, + 545, + 518 + ], + "type": "text", + "content": " and visual policy " + }, + { + "bbox": [ + 304, + 481, + 545, + 518 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 304, + 481, + 545, + 518 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 380, + 524, + 471, + 538 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 524, + 471, + 538 + ], + "spans": [ + { + "bbox": [ + 380, + 524, + 471, + 538 + ], + "type": "interline_equation", + "content": "L _ {B C} = | \\hat {\\pi} (p) - \\pi (o) |", + "image_path": "ffe4dec5ee70962d20d3150b058bcf2808ff614502968fecfcef579044445840.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 544, + 545, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 544, + 545, + 593 + ], + "spans": [ + { + "bbox": [ + 304, + 544, + 545, + 593 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 544, + 545, + 593 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 304, + 544, + 545, + 593 + ], + "type": "text", + "content": " is the privileged observation (including elevation maps), and " + }, + { + "bbox": [ + 304, + 544, + 545, + 593 + ], + "type": "inline_equation", + "content": "\\mathbf{o}" + }, + { + "bbox": [ + 304, + 544, + 545, + 593 + ], + "type": "text", + "content": " is the observation (including visual observations). Combined with our proposed self-supervised task, the overall training loss for our NVM is given by:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 370, + 599, + 481, + 611 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 599, + 481, + 611 + ], + "spans": [ + { + "bbox": [ + 370, + 599, + 481, + 611 + ], + "type": "interline_equation", + "content": "L = \\lambda_ {B C} L _ {B C} + \\lambda_ {r e c} L _ {r e c}", + "image_path": "29d870ec647c5ec3672f778ef3cd9bb4d26dc4ba49ab1d3acab4cf01540295b1.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 305, + 618, + 523, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 618, + 523, + 630 + ], + "spans": [ + { + "bbox": [ + 305, + 618, + 523, + 630 + ], + "type": "text", + "content": "In our experiment, we use " + }, + { + "bbox": [ + 305, + 618, + 523, + 630 + ], + "type": "inline_equation", + "content": "\\lambda_{BC} = 1" + }, + { + "bbox": [ + 305, + 618, + 523, + 630 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 618, + 523, + 630 + ], + "type": "inline_equation", + "content": "\\lambda_{rec} = 0.01" + }, + { + "bbox": [ + 305, + 618, + 523, + 630 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 639, + 388, + 653 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 639, + 388, + 653 + ], + "spans": [ + { + "bbox": [ + 306, + 639, + 388, + 653 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 658, + 449, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 658, + 449, + 671 + ], + "spans": [ + { + "bbox": [ + 306, + 658, + 449, + 671 + ], + "type": "text", + "content": "4.1. Simulation Environments" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 677, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 714 + ], + "type": "text", + "content": "We evaluate NVM on four challenging terrains shown in Figure 2. Our intention is to evaluate the global and local perception ability of agents, as well as the robustness of" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 46, + 693, + 288, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 693, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 693, + 288, + 714 + ], + "type": "text", + "content": "1We are able to achieve 40k fps using just elevation map. Using " + }, + { + "bbox": [ + 46, + 693, + 288, + 714 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 46, + 693, + 288, + 714 + ], + "type": "text", + "content": " images, it slows down to 1k fps." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1434" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 101, + 544, + 213 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 545, + 94 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 545, + 94 + ], + "type": "text", + "content": "Table 2. Evaluation of All Policies: We evaluate all methods in all four environments and show their Traversing Rate and Success Rate. The privileged elevation-map-based policy is noted as Teacher, and its performance is regarded as the upper bound for the rest methods." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 101, + 544, + 213 + ], + "lines": [ + { + "bbox": [ + 50, + 101, + 544, + 213 + ], + "spans": [ + { + "bbox": [ + 50, + 101, + 544, + 213 + ], + "type": "table", + "html": "
Traversing Rate (%) ↑Success Rate (%) ↑
ScenariosStagesStairsStonesObstaclesStagesStairsStonesObstacles
Teacher97.9±2.094.5±1.889.6±3.192.8±1.895.6±2.386.6±2.770.6±5.287.8±2.7
NaiveCNN76.4±26.981.5±28.629.4±10.663.3±22.553.9±19.771.3±25.40.5±0.845.2±16.4
NaiveCNN-MS72.1±25.484.5±29.726.4±9.363.7±22.748.9±18.973.1±26.20.0±0.048.1±18.3
NaiveCNN-RNN68.3±24.182.4±29.726.3±9.367.9±23.948.9±18.079.1±28.92.2±1.549.8±17.9
LocoTransformer74.0±26.585.5±29.925.2±8.965.2±22.953.3±19.375.8±26.80.9±1.146.7±16.9
NVM79.9±28.185.5±30.047.4±16.774.1±26.065.0±23.279.8±28.214.4±6.061.6±21.9
", + "image_path": "c5db4398abde028676dfb3d1fd44355e8a5299a822fdb056ba5388927f651413.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 232, + 289, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 232, + 289, + 362 + ], + "spans": [ + { + "bbox": [ + 46, + 232, + 289, + 362 + ], + "type": "text", + "content": "the learned locomotion skills. Our environments consist of Stages including multiple separate stages with gaps of different sizes between stages; Obstacles including multiple tall cuboids for the agent to avoid; Stairs including stairs with different heights and widths; Stones including numerous small stones of different shapes for the agent to step over. Stages, Stairs, and Stones environments evaluate the agent's ability to reason about the nearby surrounding and perform precise foot placement. Obstacle environment requires the agent to be able to perform long-term planning and avoid obstacles beforehand." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 369, + 115, + 381 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 369, + 115, + 381 + ], + "spans": [ + { + "bbox": [ + 47, + 369, + 115, + 381 + ], + "type": "text", + "content": "4.2. Baselines" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 388, + 287, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 388, + 287, + 413 + ], + "spans": [ + { + "bbox": [ + 47, + 388, + 287, + 413 + ], + "type": "text", + "content": "We compare our method against baselines that do not contain SE(3) equivariance priors:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 419, + 287, + 645 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 47, + 419, + 287, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 419, + 287, + 467 + ], + "spans": [ + { + "bbox": [ + 47, + 419, + 287, + 467 + ], + "type": "text", + "content": "- NaiveCNN which learns visual features with a 2D CNN encoder followed by a linear projection. Visual features and proprioceptive features are concatenated. The action is then computed from the concatenated feature." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 471, + 287, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 471, + 287, + 519 + ], + "spans": [ + { + "bbox": [ + 47, + 471, + 287, + 519 + ], + "type": "text", + "content": "- NaiveCNN-RNN which shares a similar structure with NaiveCNN, but before computing action from the concatenated feature, NaiveCNN-RNN baseline utilizes a 2-layer GRU to provide memory mechanism for the agent" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 522, + 287, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 522, + 287, + 582 + ], + "spans": [ + { + "bbox": [ + 47, + 522, + 287, + 582 + ], + "type": "text", + "content": "- LocoTransformer which utilizes a shared transformer model to perform cross-modal reasoning between visual tokens from 2D visual feature map and proprioceptive features, instead of directly concatenating visual feature and proprioceptive feature." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 586, + 287, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 586, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 47, + 586, + 287, + 645 + ], + "type": "text", + "content": "- Multi-Step NaiveCNN (NaiveCNN-MS) which use a 2D CNN to extract visual feature from each depth image and fuse them together to get the overall visual feature, instead of stacking multiple depth images and processing the stacked visual observation with a 2D CNN." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": "For NaiveCNN and LocoTransformer, the visual observations in the near past are stacked in the channel dimension and serve as input for 2D encoder. All these baselines and our method share the same 2D CNN encoder structure for a fair comparison." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 231, + 414, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 231, + 414, + 243 + ], + "spans": [ + { + "bbox": [ + 306, + 231, + 414, + 243 + ], + "type": "text", + "content": "4.3. Simulated Results" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 252, + 547, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 252, + 547, + 361 + ], + "spans": [ + { + "bbox": [ + 304, + 252, + 547, + 361 + ], + "type": "text", + "content": "Evaluation metrics: We evaluate the final policies of all methods for 640 episodes in each simulated environment and compare the performance. We evaluate different methods by two metrics: traversing rate(\\%) and success rate(\\%). The traversing rate is defined by the distance the robot moves dividing the reachable distance (the distance between the end of the environment and starting position), and the Success Rate is defined by the ratio of the robot reaching the end of the environment." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 363, + 547, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 363, + 547, + 602 + ], + "spans": [ + { + "bbox": [ + 304, + 363, + 547, + 602 + ], + "type": "text", + "content": "We evaluated all methods including the privileged teacher policy (noted as Teacher) across all environments and the results are provided in Table 2. Since in stage two of our pipeline, visual policies are trained by imitation learning (our method with self-supervised learning as well), the performance of the privileged teacher policy can be regarded as the upper bound for the performance of visual policies. We can see that our NVM outperforms all the baselines in all the environments, especially the Stones environment which requires the agent to estimate the 3D structure of the surroundings from visual observations and provide accurate foot placement. Surprisingly, it turns out our method traversed a longer distance in the obstacle environment, where the understanding of 3D structure is not required since it's possible to avoid the obstacles with only the latest visual observation. In this case, we conjecture that the 3D understanding of the environment helps the agent to keep closer to the trajectory of the teacher policy would take since the teacher could exploit the 3D structure of the surroundings to find the optimal path." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": "We find that though NaiveCNN-MS introduces the mechanism of processing each frame of visual observation separately, there is no improvement compared with NaiveCNN. This comparison indicates that the improvement brought by our NVM is actually from explicitly modeling the 3D transformation instead of introducing more computation. Compared with NaiveCNN, the memory-enabled NaiveCNN-RNN also shows no improvement, which indicates that the memory mechanism provided by RNN is not better than, if" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1435" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 102, + 544, + 242 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 547, + 94 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 547, + 94 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 547, + 94 + ], + "type": "text", + "content": "Table 3. Ablation study. We scale the grid resolution of the neural volumetric memory and the horizon of the past kept in memory. By default, our NVM uses use " + }, + { + "bbox": [ + 47, + 70, + 547, + 94 + ], + "type": "inline_equation", + "content": "D = 12" + }, + { + "bbox": [ + 47, + 70, + 547, + 94 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 70, + 547, + 94 + ], + "type": "inline_equation", + "content": "H = W = 6" + }, + { + "bbox": [ + 47, + 70, + 547, + 94 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 70, + 547, + 94 + ], + "type": "inline_equation", + "content": "n = 5" + }, + { + "bbox": [ + 47, + 70, + 547, + 94 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 102, + 544, + 242 + ], + "lines": [ + { + "bbox": [ + 50, + 102, + 544, + 242 + ], + "spans": [ + { + "bbox": [ + 50, + 102, + 544, + 242 + ], + "type": "table", + "html": "
Traversing Rate (%) ↑Success Rate (%) ↑
ScenariosStagesStairsStonesObstaclesStagesStairsStonesObstacles
NVM79.9±28.185.5±30.047.4±16.774.1±26.065.0±23.279.8±28.214.4±6.061.6±21.9
NVM(D=3)65.4±23.768.0±25.338.6±14.546.5±17.052.8±19.452.2±19.75.9±4.136.1±13.5
NVM(D=6)79.3±27.883.1±29.249.8±17.666.2±23.664.2±23.074.7±26.414.7±6.250.2±19.3
NVM(D=4x4)72.8±26.570.1±25.939.1±14.466.4±24.054.8±21.462.8±23.711.3±6.048.6±18.4
NVM(D=8x8)78.9±27.873.4±26.341.8±15.064.5±22.961.4±22.364.8±23.912.2±4.850.0±18.3
NVM(n=3)70.1±24.876.6±26.937.6±13.364.3±23.043.9±16.867.0±23.90.6±0.848.9±18.1
NVM(n=9)78.4±27.685.9±30.139.8±14.169.8±24.661.6±22.080.0±28.55.8±3.256.1±19.9
", + "image_path": "5295f174b48a6623a861f649d5048f077b6bd31715a722251b72a4ae8fbaf8b9.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 261, + 288, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 261, + 288, + 285 + ], + "spans": [ + { + "bbox": [ + 46, + 261, + 288, + 285 + ], + "type": "text", + "content": "not worse, naively stacking the visual observations in the near past." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 298, + 162, + 310 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 298, + 162, + 310 + ], + "spans": [ + { + "bbox": [ + 47, + 298, + 162, + 310 + ], + "type": "text", + "content": "4.4. Real World Results" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 45, + 318, + 290, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 318, + 290, + 715 + ], + "spans": [ + { + "bbox": [ + 45, + 318, + 290, + 715 + ], + "type": "text", + "content": "To validate the performance of our method beyond simulation, we constructed multiple real-world experiment scenes as shown on the right of Figure 2. We evaluated our method in three scenes: Stages: we build 4 discrete stages with " + }, + { + "bbox": [ + 45, + 318, + 290, + 715 + ], + "type": "inline_equation", + "content": "15\\mathrm{cm} / 20\\mathrm{cm} / 30\\mathrm{cm}" + }, + { + "bbox": [ + 45, + 318, + 290, + 715 + ], + "type": "text", + "content": " between stages. Stairs: We build a stair consisting of 3 ladders with heights of " + }, + { + "bbox": [ + 45, + 318, + 290, + 715 + ], + "type": "inline_equation", + "content": "3\\mathrm{cm} / 6\\mathrm{cm} / 10\\mathrm{cm}" + }, + { + "bbox": [ + 45, + 318, + 290, + 715 + ], + "type": "text", + "content": ". Obstacles: We place multiple cuboid obstacles in the narrow hallway for the robot to avoid. We deploy the policies in real-world scenes and record the distance robot moved for each episode. We terminate the episode after a certain amount of time or until the robot reaches the end of the environment or the robot falls. We repeat the experiment for each method and environment five times. Real-world trajectories are also provided in Figure 1. Since NaiveCNN, NaiveCNN-RNN, LocoTransformer show similar performance in simulation and LocoTransformer is known to be more robust in unseen scenarios and objects in the real world, we compared our NVM with LocoTransformer in these three scenarios. The quantitative results are shown in Table 4. We find that our NVM moved significantly further in Stages and Stairs while slightly further in Obstacles. This is because Stages and Stairs environment require the robot to understand the surrounding 3D structure to place the foot while Obstacles only requires the ability to perform high-level planning with visual observation. We observe that in Stages and Stairs environment, agents trained with LocoTransformer can place the front leg across the stage or over the ladder, but not the rear legs when the gap between stages is large or the ladder is tall, which demonstrates that lack of understanding of the surrounding 3D structure. We also found that even though LocoTransformer agent could move a reasonable distance in simulation in Stairs and Stages environment, it fails in" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 313, + 279, + 537, + 346 + ], + "blocks": [ + { + "bbox": [ + 328, + 259, + 525, + 271 + ], + "lines": [ + { + "bbox": [ + 328, + 259, + 525, + 271 + ], + "spans": [ + { + "bbox": [ + 328, + 259, + 525, + 271 + ], + "type": "text", + "content": "Table 4. Evaluating the visual policy in the real world" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 313, + 279, + 537, + 346 + ], + "lines": [ + { + "bbox": [ + 313, + 279, + 537, + 346 + ], + "spans": [ + { + "bbox": [ + 313, + 279, + 537, + 346 + ], + "type": "table", + "html": "
Distance Moved (m) ↑
ScenariosStagesStairsObstacles
NVM5.4±0.34.0±0.07.3±0.1
LocoTransformer3.4±0.12.1±0.86.7±0.4
", + "image_path": "bd7b19c92e76b35233b6d023b311f186b0ca8dcd53c0f8174b7cb466e763bf2c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 360, + 547, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 360, + 547, + 421 + ], + "spans": [ + { + "bbox": [ + 304, + 360, + 547, + 421 + ], + "type": "text", + "content": "the real world. The real-world results indicate that the understanding of the surrounding 3D structures is essential for legged robots to traverse complex terrains in the real world and our NVM module provides a high-quality volumetric memory of the surrounding 3D structure." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 430, + 447, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 430, + 447, + 443 + ], + "spans": [ + { + "bbox": [ + 305, + 430, + 447, + 443 + ], + "type": "text", + "content": "4.5. Visualization of SSL task" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 450, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 450, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 547, + 715 + ], + "type": "text", + "content": "Aiming to gather insight into our learned 3D features volume, we visualized the synthesized visual observation " + }, + { + "bbox": [ + 304, + 450, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\hat{O}_t^v" + }, + { + "bbox": [ + 304, + 450, + 547, + 715 + ], + "type": "text", + "content": " in our self-supervised learning task along with the present visual observation " + }, + { + "bbox": [ + 304, + 450, + 547, + 715 + ], + "type": "inline_equation", + "content": "O_t^v" + }, + { + "bbox": [ + 304, + 450, + 547, + 715 + ], + "type": "text", + "content": " (the ground truth in our SSL task), and most dated visual observation " + }, + { + "bbox": [ + 304, + 450, + 547, + 715 + ], + "type": "inline_equation", + "content": "O_{t - n + 1}^v" + }, + { + "bbox": [ + 304, + 450, + 547, + 715 + ], + "type": "text", + "content": " (the input in our SSL task) in Figure 5. Even though we apply extensive data augmentation over the raw visual observation (the second image for every case), the synthesized visual observations are clearly indicating the robustness of our model. In the synthesized visual observation for " + }, + { + "bbox": [ + 304, + 450, + 547, + 715 + ], + "type": "inline_equation", + "content": "Stair" + }, + { + "bbox": [ + 304, + 450, + 547, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 450, + 547, + 715 + ], + "type": "inline_equation", + "content": "Obstacle" + }, + { + "bbox": [ + 304, + 450, + 547, + 715 + ], + "type": "text", + "content": " environment, we can see the overall structure of the environment is well-preserved. In Figure 5b, The obstacles in the previous frame are transformed to the position in the current frame without much distortion. In Figure 5c, the boundary of the nearest ladder is clear, and the second nearest ladder is also observable in the present frame. We have to note that the reason the decoder can not recover the details away from the robot is in the input for SSL the large values are clipped. For more complex Stages and Stones environments, we found that even though the terrain details are not perfectly reconstructed, the essential regions for decision-making are reconstructed which means the learned 3D voxel" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 317, + 757 + ], + "type": "text", + "content": "1436" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 73, + 294, + 145 + ], + "blocks": [ + { + "bbox": [ + 48, + 73, + 294, + 145 + ], + "lines": [ + { + "bbox": [ + 48, + 73, + 294, + 145 + ], + "spans": [ + { + "bbox": [ + 48, + 73, + 294, + 145 + ], + "type": "image", + "image_path": "5cdfdfdfff859ff42c5df33b8d3d43adbb51e7b3e89254457d7a3cdd9984a7a4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 154, + 147, + 187, + 156 + ], + "lines": [ + { + "bbox": [ + 154, + 147, + 187, + 156 + ], + "spans": [ + { + "bbox": [ + 154, + 147, + 187, + 156 + ], + "type": "text", + "content": "(a) Stages" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 300, + 73, + 547, + 145 + ], + "blocks": [ + { + "bbox": [ + 300, + 73, + 547, + 145 + ], + "lines": [ + { + "bbox": [ + 300, + 73, + 547, + 145 + ], + "spans": [ + { + "bbox": [ + 300, + 73, + 547, + 145 + ], + "type": "image", + "image_path": "e6f2145722811db0516d8f3f33dbafa54e9063633b284bd2bdfbca7eba5131c7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 402, + 147, + 444, + 156 + ], + "lines": [ + { + "bbox": [ + 402, + 147, + 444, + 156 + ], + "spans": [ + { + "bbox": [ + 402, + 147, + 444, + 156 + ], + "type": "text", + "content": "(b) Obstacles" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 48, + 160, + 294, + 232 + ], + "blocks": [ + { + "bbox": [ + 48, + 160, + 294, + 232 + ], + "lines": [ + { + "bbox": [ + 48, + 160, + 294, + 232 + ], + "spans": [ + { + "bbox": [ + 48, + 160, + 294, + 232 + ], + "type": "image", + "image_path": "fcddc846ab074ba48db9a679d83d06568e909c217d6c7d1e69123936a3388ed5.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 156, + 233, + 186, + 242 + ], + "lines": [ + { + "bbox": [ + 156, + 233, + 186, + 242 + ], + "spans": [ + { + "bbox": [ + 156, + 233, + 186, + 242 + ], + "type": "text", + "content": "(c) Stairs" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 300, + 160, + 547, + 232 + ], + "blocks": [ + { + "bbox": [ + 300, + 160, + 547, + 232 + ], + "lines": [ + { + "bbox": [ + 300, + 160, + 547, + 232 + ], + "spans": [ + { + "bbox": [ + 300, + 160, + 547, + 232 + ], + "type": "image", + "image_path": "1a8f9e1f53f7f204c7ba3a870cc35854f92a0c7b407d9724dd94a17c5becf98c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 406, + 233, + 440, + 242 + ], + "lines": [ + { + "bbox": [ + 406, + 233, + 440, + 242 + ], + "spans": [ + { + "bbox": [ + 406, + 233, + 440, + 242 + ], + "type": "text", + "content": "(d) Stones" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 248, + 547, + 314 + ], + "lines": [ + { + "bbox": [ + 46, + 248, + 547, + 314 + ], + "spans": [ + { + "bbox": [ + 46, + 248, + 547, + 314 + ], + "type": "text", + "content": "Figure 5. Visual reconstruction from learned decoder. We visualize the synthesized visual observation in our self-supervised task. For every tuple, the first image shows the robot moving in the environment, the second image is the input visual observation " + }, + { + "bbox": [ + 46, + 248, + 547, + 314 + ], + "type": "inline_equation", + "content": "O_{t - n + 1}^v" + }, + { + "bbox": [ + 46, + 248, + 547, + 314 + ], + "type": "text", + "content": ", the third image is the synthesized visual observation using 3D feature volume " + }, + { + "bbox": [ + 46, + 248, + 547, + 314 + ], + "type": "inline_equation", + "content": "V_{t - n + 1}" + }, + { + "bbox": [ + 46, + 248, + 547, + 314 + ], + "type": "text", + "content": " extracted from " + }, + { + "bbox": [ + 46, + 248, + 547, + 314 + ], + "type": "inline_equation", + "content": "O_{t - n + 1}^v" + }, + { + "bbox": [ + 46, + 248, + 547, + 314 + ], + "type": "text", + "content": " and the estimated relative camera between " + }, + { + "bbox": [ + 46, + 248, + 547, + 314 + ], + "type": "inline_equation", + "content": "O_{t - n + 1}^v" + }, + { + "bbox": [ + 46, + 248, + 547, + 314 + ], + "type": "text", + "content": " and robot's ego-centric frame, and the fourth image is the actual visual observation captured by the robot (the ground-truth in our self-supervised task). For the input visual observation, we apply extensive data augmentation to the image to improve the robustness of our model." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 329, + 287, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 329, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 46, + 329, + 287, + 376 + ], + "type": "text", + "content": "features capture the most important parts in the visual observations. In Figure 5a, we can see the gaps between stages nearby are well reconstructed and in Figure 5d, we can see the reconstructed stone in front of the robot." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 384, + 140, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 384, + 140, + 396 + ], + "spans": [ + { + "bbox": [ + 47, + 384, + 140, + 396 + ], + "type": "text", + "content": "4.6. Ablation Study" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 403, + 287, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 403, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 46, + 403, + 287, + 437 + ], + "type": "text", + "content": "we perform multiple sets of ablation studies, and the results are shown in Table 3, with the purpose of studying the influence of different components in our method." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 437, + 287, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 287, + 594 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 287, + 594 + ], + "type": "text", + "content": "The Resolution of NVM: We provide our NVM module voxel features with different resolutions by slightly changing the structure of the 2D encoder, which resulting 2D feature maps in different shapes and use different depth parameter " + }, + { + "bbox": [ + 46, + 437, + 287, + 594 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 437, + 287, + 594 + ], + "type": "text", + "content": ". From Table 3, we find that the larger depth value " + }, + { + "bbox": [ + 46, + 437, + 287, + 594 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 437, + 287, + 594 + ], + "type": "text", + "content": " generally provides better performance, since larger " + }, + { + "bbox": [ + 46, + 437, + 287, + 594 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 437, + 287, + 594 + ], + "type": "text", + "content": " converts the 2D feature map into more fine-grained voxel feature which results in fine-grained neural volumetric memory, and it's supposed to be easier for the agent to reason about the foot placement over complex terrain with more fine-grained neural volumetric memory. For " + }, + { + "bbox": [ + 46, + 437, + 287, + 594 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 46, + 437, + 287, + 594 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 437, + 287, + 594 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 46, + 437, + 287, + 594 + ], + "type": "text", + "content": ", we found that " + }, + { + "bbox": [ + 46, + 437, + 287, + 594 + ], + "type": "inline_equation", + "content": "6x6" + }, + { + "bbox": [ + 46, + 437, + 287, + 594 + ], + "type": "text", + "content": " which provides enough detail horizontally without introducing much noise vertically works best." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": "The Length of Visual History of NVM: We provide our NVM model with different lengths of visual history by changing the number of stored visual observation frames. As shown in Table 3, the performance of " + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "inline_equation", + "content": "NVM" + }, + { + "bbox": [ + 46, + 594, + 287, + 714 + ], + "type": "text", + "content": " improves when the length of visual history increases to 5 from 3, we think this is because the longer visual history provides longer 3D memory of the surrounding for the agent to safely placing the rear feet. But when the length of visual history increases to 9 from 5, the performance drops, we hypothesize that it is because when the visual history becomes too" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 329, + 546, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 329, + 546, + 376 + ], + "spans": [ + { + "bbox": [ + 304, + 329, + 546, + 376 + ], + "type": "text", + "content": "long, it's hard for the pose encoder to estimate the accurate transformation between the oldest frame and the current frame, and the surplus visual information might be redundant for decision-making over terrains." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 391, + 378, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 391, + 378, + 403 + ], + "spans": [ + { + "bbox": [ + 306, + 391, + 378, + 403 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 413, + 546, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 413, + 546, + 593 + ], + "spans": [ + { + "bbox": [ + 304, + 413, + 546, + 593 + ], + "type": "text", + "content": "We propose a short-term memory mechanism, the Neural Volumetric Memory (NVM), to enable legged robots to traverse complex terrains with ease. NVM aggregates knowledge from sequential visual observation in the near past by explicitly modeling the translation and rotation between previous frames and the present frame. The explicit modeling of the transformation and self-supervised task also encourage the " + }, + { + "bbox": [ + 304, + 413, + 546, + 593 + ], + "type": "inline_equation", + "content": "SE(3)" + }, + { + "bbox": [ + 304, + 413, + 546, + 593 + ], + "type": "text", + "content": " equivariance in the memory space. Our neural volumetric memory module enables the legged robot to traverse complex terrain in simulation and the real world and achieve better sim-to-real generalization results. This shows our neural volumetric memory provides reliable surrounding 3D structure information for the robot to perform decision-making and new possibility of utilizing visual observation for legged-robot." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "content": "Acknowledgements. This project was supported, in part, by NSF CCF-2112665 (TILOS), NSF CAREER Award IIS-2240014, NSF 1730158 CI-New: Cognitive Hardware and Software Ecosystem Community Infrastructure (CHASE-CI), NSF ACI-1541349 CC*DNI Pacific Research Platform, Amazon Research Award and gifts from Qualcomm. Ge Yang is supported by the National Science Foundation Institute for Artificial Intelligence and Fundamental Interactions (IAIFI, https://iaifi.org/) under the Cooperative Agreement PHY-2019786. We thank Yuxiang Yang for providing help regarding hardware setup." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1437" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "text", + "content": "[1] Ananye Agarwal, Ashish Kumar, Jitendra Malik, and Deepak Pathak. Legged locomotion in challenging terrains using egocentric vision. In 6th Annual Conference on Robot Learning, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 135, + 288, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 135, + 288, + 179 + ], + "spans": [ + { + "bbox": [ + 53, + 135, + 288, + 179 + ], + "type": "text", + "content": "[2] Gerardo Bledt and Sangbae Kim. Extracting legged locomotion heuristics with regularized predictive control. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 406-412. IEEE, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 180, + 288, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 180, + 288, + 245 + ], + "spans": [ + { + "bbox": [ + 53, + 180, + 288, + 245 + ], + "type": "text", + "content": "[3] Gerardo Bledt, Matthew J Powell, Benjamin Katz, Jared Di Carlo, Patrick M Wensing, and Sangbae Kim. Mit cheetah 3: Design and control of a robust, dynamic quadruped robot. In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 2245–2252. IEEE, 2018. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 246, + 288, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 246, + 288, + 290 + ], + "spans": [ + { + "bbox": [ + 53, + 246, + 288, + 290 + ], + "type": "text", + "content": "[4] Jan Carius, René Ranftl, Vladlen Koltun, and Marco Hutter. Trajectory optimization for legged robots with slipping motions. IEEE Robotics and Automation Letters, 4(3):3013-3020, 2019. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 291, + 288, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 291, + 288, + 357 + ], + "spans": [ + { + "bbox": [ + 53, + 291, + 288, + 357 + ], + "type": "text", + "content": "[5] Jared Di Carlo, Patrick M. Wensing, Benjamin Katz, Gerardo Bledt, and Sangbae Kim. Dynamic locomotion in the MIT cheetah 3 through convex model-predictive control. In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems, IROS 2018, Madrid, Spain, October 1-5, 2018, pages 1-9. IEEE, 2018. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 357, + 288, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 357, + 288, + 391 + ], + "spans": [ + { + "bbox": [ + 53, + 357, + 288, + 391 + ], + "type": "text", + "content": "[6] Dian Chen, Brady Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Learning by cheating. In Conference on Robot Learning, pages 66-75. PMLR, 2020. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 392, + 288, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 392, + 288, + 424 + ], + "spans": [ + { + "bbox": [ + 53, + 392, + 288, + 424 + ], + "type": "text", + "content": "[7] Tao Chen, Jie Xu, and Pulkit Agrawal. A system for general in-hand object re-orientation. In Conference on Robot Learning, pages 297–307. PMLR, 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 425, + 288, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 425, + 288, + 468 + ], + "spans": [ + { + "bbox": [ + 53, + 425, + 288, + 468 + ], + "type": "text", + "content": "[8] Annett Chilian and Heiko Hirschmüller. Stereo camera based navigation of mobile robots on rough terrain. In 2009 IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 4571-4576. IEEE, 2009. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 469, + 288, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 469, + 288, + 524 + ], + "spans": [ + { + "bbox": [ + 53, + 469, + 288, + 524 + ], + "type": "text", + "content": "[9] Jared Di Carlo, Patrick M Wensing, Benjamin Katz, Gerardo Bledt, and Sangbae Kim. Dynamic locomotion in the mit cheetah 3 through convex model-predictive control. In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1-9. IEEE, 2018. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 525, + 287, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 287, + 578 + ], + "type": "text", + "content": "[10] Yanran Ding, Abhishek Panda, and Hae-Won Park. Realtime model predictive control for versatile dynamic motions in quadrupedal robots. In 2019 International Conference on Robotics and Automation (ICRA), pages 8484-8490. IEEE, 2019. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 580, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 580, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 580, + 287, + 635 + ], + "type": "text", + "content": "[11] Peter Fankhauser, Marko Bjelonic, C. Dario Bellicoso, Takahiro Miki, and Marco Hutter. Robust rough-terrain locomotion with a quadrupedal robot. In 2018 IEEE International Conference on Robotics and Automation (ICRA), pages 5761-5768, 2018. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "text", + "content": "[12] Péter Fankhauser, Michael Bloesch, Christian Gehring, Marco Hutter, and Roland Siegwart. Robot-centric elevation mapping with uncertainty estimates. In International Conference on Climbing and Walking Robots (CLAWAR), 2014. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 690, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 690, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 690, + 287, + 714 + ], + "type": "text", + "content": "[13] Péter Fankhauser, Michael Bloesch, and Marco Hutter. Probabilistic terrain mapping for mobile robots with unclear" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "tain localization. IEEE Robotics and Automation Letters, 3(4):3019-3026, 2018. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 96, + 545, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 129 + ], + "type": "text", + "content": "[14] Hsiao-Yu Fish Tung, Ricson Cheng, and Katerina Fragkiadaki. Learning spatial common sense with geometry-aware recurrent networks. In cvpr. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 130, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 130, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 130, + 545, + 195 + ], + "type": "text", + "content": "[15] Christian Gehring, Stelian Coros, Marco Hutter, Michael Blösch, Mark A. Hoepflinger, and Roland Siegwart. Control of dynamic gaits for a quadrupedal robot. In 2013 IEEE International Conference on Robotics and Automation, Karlsruhe, Germany, May 6-10, 2013, pages 3287-3292. IEEE, 2013. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 197, + 547, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 547, + 241 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 547, + 241 + ], + "type": "text", + "content": "[16] Hartmut Geyer, Andre Seyfarth, and Reinhard Blickhan. Positive force feedback in bouncing gaits? Proceedings of the Royal Society of London. Series B: Biological Sciences, 270(1529):2173-2183, 2003. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 243, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 243, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 243, + 545, + 285 + ], + "type": "text", + "content": "[17] Ruben Grandia, Farbod Farshidian, Alexey Dosovitskiy, René Ranftl, and Marco Hutter. Frequency-aware model predictive control. IEEE Robotics and Automation Letters, 4(2):1517-1524, 2019. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 287, + 545, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 287, + 545, + 331 + ], + "spans": [ + { + "bbox": [ + 308, + 287, + 545, + 331 + ], + "type": "text", + "content": "[18] Adam W Harley, Shrinidhi Kowshika Lakshmikanth, Paul Schydlo, and Katerina Fragkiadaki. Tracking emerges by looking around static scenes, with neural 3d mapping. In eccv, 2020. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 333, + 545, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 333, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 308, + 333, + 545, + 376 + ], + "type": "text", + "content": "[19] Adam W. Harley, Fangyu Li, Shrinidhi K. Lakshmikanth, Xian Zhou, Hsiao-Yu Fish Tung, and Katerina Fragkiadaki. Learning from unlabelled videos using contrastive predictive neural 3d mapping. In ICLR, 2019. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 378, + 545, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 378, + 545, + 410 + ], + "spans": [ + { + "bbox": [ + 308, + 378, + 545, + 410 + ], + "type": "text", + "content": "[20] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. Cambridge university press, 2003. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 412, + 545, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 412, + 545, + 465 + ], + "spans": [ + { + "bbox": [ + 308, + 412, + 545, + 465 + ], + "type": "text", + "content": "[21] Ronghang Hu, Nikhila Ravi, Alexander C. Berg, and Deepak Pathak. Worldsheet: Wrapping the world in a 3d sheet for view synthesis from a single image. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2021. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 468, + 545, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 468, + 545, + 511 + ], + "spans": [ + { + "bbox": [ + 308, + 468, + 545, + 511 + ], + "type": "text", + "content": "[22] Jemin Hwangbo, J. Lee, A. Dosovitskiy, Dario Bellicoso, Vassilios Tsounis, V. Koltun, and M. Hutter. Learning agile and dynamic motor skills for legged robots. Science Robotics, 4, 2019. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 514, + 545, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 514, + 545, + 579 + ], + "spans": [ + { + "bbox": [ + 308, + 514, + 545, + 579 + ], + "type": "text", + "content": "[23] Chieko Imai, Minghao Zhang, Yuchen Zhang, Marcin Kierebinski, Ruihan Yang, Yuzhe Qin, and Xiaolong Wang. Vision-guided quadrupedal locomotion in the wild with multi-modal delay randomization. In 2022 IEEE/RSJ international conference on intelligent robots and systems (IROS). 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 581, + 545, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 581, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 308, + 581, + 545, + 656 + ], + "type": "text", + "content": "[24] Atil Iscen, Ken Caluwaerts, Jie Tan, Tingnan Zhang, Erwin Coumans, Vikas Sindhwani, and Vincent Vanhoucke. Policies modulating trajectory generators. In 2nd Annual Conference on Robot Learning, CoRL 2018, Zürich, Switzerland, 29-31 October 2018, Proceedings, volume 87 of Proceedings of Machine Learning Research, pages 916-926. PMLR, 2018. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "type": "text", + "content": "[25] Deepali Jain, Atil Iscen, and Ken Caluwaerts. Hierarchical reinforcement learning for quadruped locomotion. In 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems, IROS 2019, Macau, SAR, China, November 3-8, 2019, pages 7551-7557. IEEE, 2019. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1438" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[26] Deepali Jain, Atil Iscen, and Ken Caluwaerts. From pixels to legs: Hierarchical learning of quadruped locomotion. arXiv preprint arXiv:2011.11722, 2020. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 150 + ], + "type": "text", + "content": "[27] Joel Janai, Fatma Güney, Aseem Behl, Andreas Geiger, et al. Computer vision for autonomous vehicles: Problems, datasets and state of the art. Foundations and Trends® in Computer Graphics and Vision, 12(1-3):1-308, 2020. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 195 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 195 + ], + "type": "text", + "content": "[28] Danilo Jimenez Rezende, SM Eslami, Shakir Mohamed, Peter Battaglia, Max Jaderberg, and Nicolas Heess. Unsupervised learning of 3d structure from images. Advances in neural information processing systems, 29:4996-5004, 2016. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 197, + 287, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 197, + 287, + 229 + ], + "spans": [ + { + "bbox": [ + 48, + 197, + 287, + 229 + ], + "type": "text", + "content": "[29] Angjoo Kanazawa, Shubham Tulsiani, Alexei A Efros, and Jitendra Malik. Learning category-specific mesh reconstruction from image collections. In ECCV, 2018. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 231, + 287, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 231, + 287, + 284 + ], + "spans": [ + { + "bbox": [ + 48, + 231, + 287, + 284 + ], + "type": "text", + "content": "[30] Elia Kaufmann, Antonio Loquercio, Rene Ranftl, Alexey Dosovitskiy, Vladlen Koltun, and Davide Scaramuzza. Deep drone racing: Learning agile flight in dynamic environments. In Conference on Robot Learning, pages 133-145. PMLR, 2018. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 286, + 287, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 286, + 287, + 329 + ], + "spans": [ + { + "bbox": [ + 48, + 286, + 287, + 329 + ], + "type": "text", + "content": "[31] Elia Kaufmann, Antonio Loquercio, René Ranftl, Matthias Müller, Vladlen Koltun, and Davide Scaramuzza. Deep drone acrobatics. arXiv preprint arXiv:2006.05768, 2020. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 331, + 287, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 331, + 287, + 386 + ], + "spans": [ + { + "bbox": [ + 48, + 331, + 287, + 386 + ], + "type": "text", + "content": "[32] Nate Kohl and Peter Stone. Policy gradient reinforcement learning for fast quadrupedal locomotion. In Proceedings of the 2004 IEEE International Conference on Robotics and Automation, ICRA 2004, April 26 - May 1, 2004, New Orleans, LA, USA, pages 2619-2624. IEEE, 2004. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 388, + 287, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 287, + 420 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 287, + 420 + ], + "type": "text", + "content": "[33] Ashish Kumar, Zipeng Fu, Deepak Pathak, and Jitendra Malik. Rma: Rapid motor adaptation for legged robot. Robotics: Science and Systems, 2021. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 421, + 287, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 421, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 48, + 421, + 287, + 464 + ], + "type": "text", + "content": "[34] In-So Kweon and Takeo Kanade. High-resolution terrain map from multiple sensor data. IEEE Transactions on Pattern Analysis and Machine Intelligence, 14(2):278-292, 1992. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 466, + 287, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 466, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 48, + 466, + 287, + 498 + ], + "type": "text", + "content": "[35] Zihang Lai, Sifei Liu, Alexei A Efros, and Xiaolong Wang. Video autoencoder: self-supervised disentanglement of 3d structure and motion. In ICCV, 2021. 3, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 500, + 287, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 500, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 500, + 287, + 533 + ], + "type": "text", + "content": "[36] Joonho Lee, Jemin Hwangbo, Lorenz Wellhausen, Vladlen Koltun, and Marco Hutter. Learning quadrupedal locomotion over challenging terrain. Science robotics, 5(47), 2020. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 534, + 287, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 534, + 287, + 565 + ], + "spans": [ + { + "bbox": [ + 48, + 534, + 287, + 565 + ], + "type": "text", + "content": "[37] Antonio Loquercio, Ashish Kumar, and Jitendra Malik. Learning visual locomotion with cross-modal supervision, 2022. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 567, + 287, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 287, + 600 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 287, + 600 + ], + "type": "text", + "content": "[38] G. Margolis, Ge Yang, Kartik Paigwar, Tao Chen, and Pulkit Agrawal. Rapid locomotion via reinforcement learning. ArXiv, abs/2205.02824, 2022. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 601, + 287, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 287, + 645 + ], + "type": "text", + "content": "[39] Gabriel B Margolis, Tao Chen, Kartik Paigwar, Xiang Fu, Donghyun Kim, Sang bae Kim, and Pulkit Agrawal. Learning to jump from pixels. In 5th Annual Conference on Robot Learning, 2021. 2, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "text", + "content": "[40] Takahiro Miki, Joonho Lee, Jemin Hwangbo, Lorenz Wellhausen, Vladlen Koltun, and Marco Hutter. Learning robust perceptive locomotion for quadrupedal robots in the wild. Sci Robot, 7(62):eabk2822, Jan. 2022. 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[41] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf:" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 95, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 95, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 308, + 95, + 545, + 128 + ], + "type": "text", + "content": "[42] Hirofumi Miura and Isao Shimoyama. Dynamic walk of a biped. The International Journal of Robotics Research, 3(2):60-74, 1984. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 129, + 545, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 129, + 545, + 182 + ], + "spans": [ + { + "bbox": [ + 308, + 129, + 545, + 182 + ], + "type": "text", + "content": "[43] Tongzhou Mu, Zhan Ling, Fanbo Xiang, Derek Yang, Xu-anlin Li, Stone Tao, Zhiao Huang, Zhiwei Jia, and Hao Su. Maniskill: Generalizable manipulation skill benchmark with large-scale demonstrations. arXiv preprint arXiv:2107.14483, 2021. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 184, + 545, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 184, + 545, + 238 + ], + "spans": [ + { + "bbox": [ + 308, + 184, + 545, + 238 + ], + "type": "text", + "content": "[44] Siva Karthik Mustikovela, Varun Jampani, Shalini De Mello, Sifei Liu, Umar Iqbal, Carsten Rother, and Jan Kautz. Self-supervised viewpoint learning from image collections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3971-3981, 2020. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 239, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 239, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 308, + 239, + 545, + 293 + ], + "type": "text", + "content": "[45] Thu Nguyen-Phuoc, Chuan Li, Lucas Theis, Christian Richardt, and Yong-Liang Yang. Hologan: Unsupervised learning of 3d representations from natural images. In Proceedings of the IEEE International Conference on Computer Vision, pages 7588–7597, 2019. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 294, + 545, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 294, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 308, + 294, + 545, + 338 + ], + "type": "text", + "content": "[46] Yiyuan Pan, Xuecheng Xu, Yue Wang, Xiaqing Ding, and Rong Xiong. GPU accelerated real-time traversability mapping. In 2019 IEEE International Conference on Robotics and Biomimetics (ROBIO), pages 734-740. IEEE, 2019. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 338, + 545, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 338, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 308, + 338, + 545, + 381 + ], + "type": "text", + "content": "[47] Xue Bin Peng, Glen Berseth, KangKang Yin, and Michiel Van De Panne. Deeploco: Dynamic locomotion skills using hierarchical deep reinforcement learning. ACM Transactions on Graphics (TOG), 36(4):1-13, 2017. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 383, + 545, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 383, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 308, + 383, + 545, + 425 + ], + "type": "text", + "content": "[48] Xue Bin Peng, Erwin Coumans, Tingnan Zhang, Tsang-Wei Edward Lee, Jie Tan, and Sergey Levine. Learning agile robotic locomotion skills by imitating animals. In Robotics: Science and Systems, 07 2020. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 426, + 545, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 426, + 545, + 470 + ], + "spans": [ + { + "bbox": [ + 308, + 426, + 545, + 470 + ], + "type": "text", + "content": "[49] Xue Bin Peng, Ze Ma, Pieter Abbeel, Sergey Levine, and Angjoo Kanazawa. Amp: Adversarial motion priors for stylized physics-based character control. ACM Trans. Graph., 40(4), July 2021. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 471, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 471, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 308, + 471, + 545, + 514 + ], + "type": "text", + "content": "[50] Lerrel Pinto, Marcin Andrychowicz, Peter Welinder, Wojciech Zaremba, and Pieter Abbeel. Asymmetric actor critic for image-based robot learning. arXiv preprint arXiv:1710.06542, 2017. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 514, + 545, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 514, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 308, + 514, + 545, + 557 + ], + "type": "text", + "content": "[51] Marc H Raibert. Hopping in legged systems—modeling and simulation for the two-dimensional one-legged case. IEEE Transactions on Systems, Man, and Cybernetics, SMC-14(3):451–463, 1984. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 559, + 545, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 559, + 545, + 592 + ], + "spans": [ + { + "bbox": [ + 308, + 559, + 545, + 592 + ], + "type": "text", + "content": "[52] Chris Rockwell, David F. Fouhey, and Justin Johnson. Pixelsynth: Generating a 3d-consistent experience from a single image. In ICCV, 2021. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 593, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 593, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 308, + 593, + 545, + 635 + ], + "type": "text", + "content": "[53] Nikita Rudin, David Hoeller, Philipp Reist, and Marco Hutter. Learning to walk in minutes using massively parallel deep reinforcement learning. In Conference on Robot Learning, pages 91-100. PMLR, 2022. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 636, + 545, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 636, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 308, + 636, + 545, + 669 + ], + "type": "text", + "content": "[54] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 5" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "text", + "content": "[55] Mingyo Seo, Ryan Gupta, Yifeng Zhu, Alexy Skoutnev, Luis Sentis, and Yuke Zhu. Learning to walk by steering: Perceptive quadrupedal locomotion in dynamic environments. In arXiv preprint arXiv:2209.09233, 2022. 2" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1439" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[56] Vincent Sitzmann, Justus Thies, Felix Heide, Matthias Nießner, Gordon Wetzstein, and Michael Zollhofer. Deepvoxels: Learning persistent 3d feature embeddings. In CVPR, 2019. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 162 + ], + "type": "text", + "content": "[57] Vincent Sitzmann, Michael Zollhöfer, and Gordon Wetzstein. Scene representation networks: Continuous 3d-structure-aware neural scene representations. In NeurIPS, 2019. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 288, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 288, + 196 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 288, + 196 + ], + "type": "text", + "content": "[58] Maks Sorokin, Jie Tan, C. Karen Liu, and Sehoon Ha. Learning to navigate sidewalks in outdoor environments. IEEE Robotics and Automation Letters, 7(2):3906-3913, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 198, + 287, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 198, + 287, + 263 + ], + "spans": [ + { + "bbox": [ + 48, + 198, + 287, + 263 + ], + "type": "text", + "content": "[59] Yu Sun, Wyatt L. Ubellacker, Wen-Loong Ma, Xiang Zhang, Changhao Wang, Noel V. Csomay-Shanklin, Masayoshi Tomizuka, Koushil Sreenath, and Aaron D. Ames. Online learning of unknown dynamics for model-based controllers in legged locomotion. IEEE Robotics and Automation Letters (RA-L), 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 265, + 287, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 265, + 287, + 352 + ], + "spans": [ + { + "bbox": [ + 48, + 265, + 287, + 352 + ], + "type": "text", + "content": "[60] Jie Tan, Tingnan Zhang, Erwin Coumans, Atil Iscen, Yunfei Bai, Danijar Hafner, Steven Bohez, and Vincent Vanhoucke. Sim-to-real: Learning agile locomotion for quadruped robots. In Hadas Kress-Gazit, Siddhartha S. Srinivasa, Tom Howard, and Nikolay Atanasov, editors, Robotics: Science and Systems XIV, Carnegie Mellon University, Pittsburgh, Pennsylvania, USA, June 26-30, 2018, 2018. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 354, + 287, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 354, + 287, + 398 + ], + "spans": [ + { + "bbox": [ + 48, + 354, + 287, + 398 + ], + "type": "text", + "content": "[61] Nick Torkos and Michiel van de Panne. Footprint-based quadruped motion synthesis. In Proceedings of the Graphics Interface 1998 Conference, June 18-20, 1998, Vancouver, BC, Canada, pages 151-160, June 1998. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 399, + 287, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 287, + 441 + ], + "type": "text", + "content": "[62] Shubham Tulsiani, Tinghui Zhou, Alexei A Efros, and Jitendra Malik. Multi-view supervision for single-view reconstruction via differentiable ray consistency. In CVPR, 2017. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 445, + 287, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 445, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 48, + 445, + 287, + 509 + ], + "type": "text", + "content": "[63] Martin Wermelinger, Péter Fankhauser, Remo Diethelm, Philipp Krüsi, Roland Siegwart, and Marco Hutter. Navigation planning for legged robots in challenging terrain. In 2016 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1184-1189. IEEE, 2016. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 511, + 287, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 511, + 287, + 566 + ], + "spans": [ + { + "bbox": [ + 48, + 511, + 287, + 566 + ], + "type": "text", + "content": "[64] Olivia Wiles, Georgia Gkioxari, Richard Szeliski, and Justin Johnson. Synsin: End-to-end view synthesis from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7467-7477, 2020. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 568, + 287, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 287, + 611 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 287, + 611 + ], + "type": "text", + "content": "[65] Zhaoming Xie, Xingye Da, Buck Babich, Animesh Garg, and Michiel van de Panne. Glide: Generalizable quadrupedal locomotion in diverse environments with a centroidal model. CoRR, abs/2104.09771, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 613, + 287, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 657 + ], + "type": "text", + "content": "[66] Zhaoming Xie, Hung Yu Ling, Nam Hee Kim, and Michiel van de Panne. Allsteps: curriculum-driven learning of stepping stone skills. In Computer Graphics Forum, volume 39, pages 213-224. Wiley Online Library, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[67] Ruihan Yang, Minghao Zhang, Nicklas Hansen, Huazhe Xu, and Xiaolong Wang. Learning vision-guided quadrupedal locomotion end-to-end with cross-modal transformers. In International Conference on Learning Representations, 2022. 3" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 284 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[68] KangKang Yin, Kevin Loken, and Michiel Van de Panne. Simbicon: Simple biped locomotion control. ACM Transactions on Graphics (TOG), 26(3):105-es, 2007. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 107, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 194 + ], + "type": "text", + "content": "[69] Wenhao Yu, Deepali Jain, Alejandro Escontrela, Atil Iscen, Peng Xu, Erwin Coumans, Sehoon Ha, Jie Tan, and Tingnan Zhang. Visual-locomotion: Learning to walk on complex terrains with vision. In Aleksandra Faust, David Hsu, and Gerhard Neumann, editors, Proceedings of the 5th Conference on Robot Learning, volume 164 of Proceedings of Machine Learning Research, pages 1291-1302. PMLR, 08-11 Nov 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 196, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 196, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 196, + 545, + 228 + ], + "type": "text", + "content": "[70] Tinghui Zhou, Matthew Brown, Noah Snavely, and David G Lowe. Unsupervised learning of depth and ego-motion from video. In CVPR, 2017. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 229, + 545, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 229, + 545, + 284 + ], + "spans": [ + { + "bbox": [ + 307, + 229, + 545, + 284 + ], + "type": "text", + "content": "[71] Matt Zucker, Nathan Ratliff, Martin Stolle, Joel Chestnutt, J Andrew Bagnell, Christopher G Atkeson, and James Kuffner. Optimization and learning for rough terrain legged locomotion. The International Journal of Robotics Research, 30(2):175-191, 2011. 2" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1440" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Neural Voting Field for Camera-Space 3D Hand Pose Estimation/96d05801-c2f5-4205-ac8b-2fcd00428ae6_content_list.json b/2023/Neural Voting Field for Camera-Space 3D Hand Pose Estimation/96d05801-c2f5-4205-ac8b-2fcd00428ae6_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a2593cb10cd428694827207a797ef1ecda853d3b --- /dev/null +++ b/2023/Neural Voting Field for Camera-Space 3D Hand Pose Estimation/96d05801-c2f5-4205-ac8b-2fcd00428ae6_content_list.json @@ -0,0 +1,1605 @@ +[ + { + "type": "text", + "text": "Neural Voting Field for Camera-Space 3D Hand Pose Estimation", + "text_level": 1, + "bbox": [ + 156, + 130, + 813, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Lin Huang*", + "bbox": [ + 156, + 180, + 259, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$g^{1*}$ Chung-Ching Lin", + "bbox": [ + 241, + 180, + 483, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Lijuan Wang2 Junsong Yuan1 \n1University at Buffalo", + "bbox": [ + 241, + 199, + 545, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kevin Lin² Lin Lia", + "bbox": [ + 558, + 180, + 781, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zicheng Liu", + "bbox": [ + 620, + 199, + 727, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2Microsoft", + "bbox": [ + 565, + 217, + 653, + 232 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://linhuang17.github.io/NVF", + "bbox": [ + 338, + 236, + 625, + 250 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/201f7b5bc242e0e13edced3ca156e9b7fc97c83a6a18e05619536c61672bd131.jpg", + "image_caption": [ + "Figure 1. We present Neural Voting Field (NVF), a 3D implicit function for camera-space 3D hand pose estimation from an RGB image. Given an RGB input (a), for a 3D point in camera space, NVF can predict its signed distance to hand surface and a set of 4D offset vectors (1D voting weight and 3D directional vector to each joint). 3D pose is then obtained via dense 3D point-wise voting. Rendered predictions: (b) Hand mesh generated by Marching Cubes from the predicted signed distances (c-h) Predicted 1D voting weight from each predicted near-surface 3D point to a joint. (Brighter line means larger weight, darker or no line means smaller weight) (i) Estimated 3D hand pose." + ], + "image_footnote": [], + "bbox": [ + 76, + 270, + 893, + 366 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 450, + 313, + 465 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We present a unified framework for camera-space 3D hand pose estimation from a single RGB image based on 3D implicit representation. As opposed to recent works, most of which first adopt holistic or pixel-level dense regression to obtain relative 3D hand pose and then follow with complex second-stage operations for 3D global root or scale recovery, we propose a novel unified 3D dense regression scheme to estimate camera-space 3D hand pose via dense 3D pointwise voting in camera frustum. Through direct dense modeling in 3D domain inspired by Pixel-aligned Implicit Functions for 3D detailed reconstruction, our proposed Neural Voting Field (NVF) fully models 3D dense local evidence and hand global geometry, helping to alleviate common 2D-to-3D ambiguities. Specifically, for a 3D query point in camera frustum and its pixel-aligned image feature, NVF, represented by a Multi-Layer Perceptron, regresses: (i) its signed distance to the hand surface; (ii) a set of 4D offset vectors (1D voting weight and 3D directional vector to each hand joint). Following a vote-casting scheme, 4D offset vectors from near-surface points are selected to calculate the 3D hand joint coordinates by a weighted average. Experiments demonstrate that NVF outperforms existing state-of-the-art algorithms on FreiHAND dataset for camera-space 3D hand pose estimation. We also adapt NVF to the classic task of root-relative 3D hand pose estimation, for which NVF also obtains state-of-the-art results on HO3D dataset.", + "bbox": [ + 75, + 483, + 472, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 501, + 450, + 630, + 465 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Monocular 3D hand pose estimation, which aims to recover 3D locations of hand joints from an RGB image, has attracted enormous attention and made remarkable progress in recent years. As a long-standing task in computer vision, it remains challenging due to its highly articulated structure, large variations in orientations, severe (self-)occlusion, and inherent 2D-to-3D scale and depth ambiguity.", + "bbox": [ + 496, + 476, + 890, + 582 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Owing to the aforementioned difficulties, most existing works [4-7, 15, 19, 26, 32, 35, 36, 45, 52, 53, 59, 60, 63] focused on one aspect of this general problem, which is to estimate root-relative 3D hand pose (i.e., 3D joint coordinates relative to a pre-defined root joint, such as hand wrist). While accurate 2D-to-3D root-relative pose estimation is essential for numerous applications in Virtual/Augmented Reality, there are various interactive tasks in which having root-relative hand joint coordinates alone is insufficient. For instance, being able to recover camera-space 3D hand joint coordinates in an AR view enables the user to directly use hands to manipulate virtual objects moving in 3D space.", + "bbox": [ + 496, + 583, + 892, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To recover robust camera-space 3D hand pose, there are two key design elements: (1) the ability to exploit dense local evidence. Specifically, as demonstrated in previous works [14, 16, 25, 29, 34, 42, 43, 47, 49, 54-56], dense regression-based methods are more effective than holistic regression-based counterparts for handling highly articulated 3D pose structure, attributed to its ability to maintain the input data spatial structure and fully exploit local evidence; (2) the ability to reason 3D hand global geome", + "bbox": [ + 496, + 763, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Work done during Lin Huang's internship with Microsoft.", + "bbox": [ + 99, + 886, + 415, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "8969", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/04dace699167e4242252c429efc5336c637acb660e55e5a4d9bbe7af4e631939.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodFirst StageSecond Stage
Iqbal et al. [29]2D-DenseScale Estimation
ObMan [22]HolisticRoot Estimation
I2L-MeshNet [42]1D-DenseRoot Depth Estimation
CMR [8]2D-Dense+SpiralConvRegistration
Hasson et al. [21]HolisticModel Fitting
NVF (Ours)Unified 3D-DenseWeighted Average
", + "bbox": [ + 78, + 88, + 475, + 178 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Table 1. Comparison of representative absolute 3D hand pose estimation schemes. Please refer to Sec. 2.1 for more details.", + "bbox": [ + 76, + 184, + 468, + 212 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "try. As shown in previous literature [13, 29, 33], given 2D evidence and camera intrinsic parameters, reasonable understanding towards target object 3D structure/geometry is crucial to alleviate 2D-to-3D depth ambiguity, which is the key to accurately locate 3D hand pose in camera space.", + "bbox": [ + 75, + 220, + 468, + 296 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To fully integrate both elements into our algorithm design in a unified manner, we connect with Pixel-aligned Implicit Function (PIFu) [24, 27, 49, 50, 62]. Through direct dense modeling in 3D domain with pixel-aligned local features, PIFu-based methods reconstruct highly detailed 3D human geometry from an RGB image in a unified way, showing its ability to model high-frequency local details such as clothing wrinkles while generating complete global geometry including largely occluded region such as the back of a person. Inspired by these results, we propose a novel unified 3D dense regression scheme based on a 3D implicit function for robust camera-space 3D hand pose estimation. Specifically, for each of the 3D query points densely sampled in camera frustum and its pixel-aligned image feature, unlike PIFu predicting occupancy value for each point, our proposed Neural Voting Field (NVF) regresses: (i) the signed distance between the point and the hand surface; (ii) a set of 4D offset vectors (1D voting weight and 3D directional vector from the point to each joint). Following a vote-casting scheme, 4D offset vectors from near-surface points (i.e., points for which the predicted signed distance is below a threshold) are selected to calculate the 3D hand joint coordinates by a weighted average.", + "bbox": [ + 75, + 297, + 468, + 643 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Most existing works for camera-space 3D hand pose estimation, as shown in Tab. 1, follow a two-stage estimation scheme. They first adopt holistic or pixel-level dense regression to obtain 2D and relative 3D hand poses and then follow with complex second-stage processing such as fitting, registration, using a separate network for 3D global root location or scale estimation. NVF instead provides a unified solution via direct dense modeling in 3D camera space followed by a simple weighted average operation, which enables reasoning about 3D dense local evidence and hand global geometry. As shown in Fig. 1, NVF makes solid 3D point-wise prediction and overall distribution of signed distance and voting weight even in highly occluded regions, leading to accurate camera-space pose estimation. In Sec. 4, we show that NVF noticeably outperforms two baselines based on holistic regression and 2D dense regression. Besides, NVF exhibits state-of-the-art performance", + "bbox": [ + 75, + 645, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "for the task of camera-space 3D hand pose estimation on FreiHAND dataset. We also adapt NVF to the classic task of root-relative 3D hand pose estimation, for which NVF also achieves state-of-the-art results on HO3D dataset.", + "bbox": [ + 496, + 90, + 890, + 150 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Since estimating absolute 3D pose from an RGB image is an ill-posed problem due to scale and depth ambiguity [13,29], in Sec. 4.4, we also provide ablation analysis on hand scale based on results from NVF and the baselines.", + "bbox": [ + 496, + 151, + 890, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This work makes the following contributions:", + "bbox": [ + 517, + 212, + 821, + 226 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We propose Neural Voting Field (NVF), as the first 3D implicit representation-based unified solution to estimate camera-space 3D hand pose.", + "2. NVF follows a novel unified 3D dense regression scheme to estimate camera-space 3D hand pose via dense 3D point-wise voting in camera frustum.", + "3. NVF outperforms baseline methods based on holistic and 2D dense regression and achieves state-of-the-art results on absolute and relative hand pose estimation." + ], + "bbox": [ + 511, + 229, + 890, + 372 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 383, + 640, + 398 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Camera-Space 3D Hand Pose Estimation", + "text_level": 1, + "bbox": [ + 500, + 407, + 849, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As shown in Tab. 1, most existing works for monocular RGB input-based camera-space 3D hand pose estimation generally follow a two-stage estimation scheme. Iqbal et al. [29] recovers 3D hand pose in camera space up to a scaling factor by solving a quadratic equation given the camera intrinsic parameters and the estimated scale-normalized and root-relative 2.5D pose via 2D Dense regression. Extra global scale estimation is required for camera-space absolute 3D hand pose recovery. Another line of work [8,21,22] first obtains 2D pose or root-relative 3D hand pose estimations via 2D dense or holistic regression. Then, Hasson et al. [21] fits a 3D hand model to the initial estimations and recovers 3D hand pose in camera space from the input RGB video. Both ObMan [22] and CMR [8] work on a single RGB input and directly recover 3D global root from the initial 2D pose and root-relative 3D pose/mesh estimations. I2L-MeshNet [42] proposes a 1D dense regression scheme to predict per-lixel likelihood on 1D heat map for recovering root-relative 2.5D pose, and then uses a separate network (i.e., RootNet [41]) for the absolute root depth estimation.", + "bbox": [ + 496, + 430, + 890, + 733 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Root-relative 3D Hand Pose Estimation", + "text_level": 1, + "bbox": [ + 500, + 741, + 836, + 756 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For single RGB input-based root-relative 3D hand pose estimation, some early methods [52, 59, 63] adopt classic frameworks such as Convolutional Neural Network and Multi-Layer Perceptron in combination with holistic or dense regression scheme for scale-invariant and rootrelative 3D hand joint positions recovery. After the introduction of the 3D parametric MANO [48] model, tremendous literature [3, 4, 9, 22, 45, 53, 60] attempts to integrate MANO into the learning pipeline and directly regress", + "bbox": [ + 496, + 763, + 890, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "8970", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/5c5b97a25cfe87446c39d98302f4648cde69e9cbef49246741415c21b4f95b2b.jpg", + "image_caption": [ + "Figure 2. Overview of Neural Voting Field (NVF): We first extract a feature map $F$ from the input RGB image $I$ by an image encoder. Then, for a 3D query point $\\pmb{p}$ sampled in camera space, we define an implicit function $f_{\\mathrm{NVF}}$ realized by a MLP, which maps from pixel-aligned image feature $S(F, \\pi(\\pmb{p}))$ and depth $Z(\\pmb{p})$ of $\\pmb{p}$ to its signed distance $s$ and a set of 4D offset vectors $V$ . Each 4D offset vector $v$ contains a 1D voting weight and a 3D directional vector to each joint. The 3D hand pose is then obtained by dense 3D point-wise voting." + ], + "image_footnote": [], + "bbox": [ + 99, + 85, + 465, + 284 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ef3ea652f3cfe22d6df35f499c4a15b8e0e6c716de1d7b6ff2d444fb2d78e63a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 455, + 98, + 854, + 276 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "MANO parameters from the image. Meanwhile, MANO model-free approaches [6] continue to develop and more works also start to investigate the dependencies among hand joints as hands are inherently structured. Thus, geometric deep learning-based models (e.g., Graph Convolution, Spiral Convolution) [5, 7, 8, 15, 32, 53] and Transformer-based models [19, 35, 36, 45] have also been proposed.", + "bbox": [ + 75, + 356, + 468, + 464 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. 3D Implicit Representation", + "text_level": 1, + "bbox": [ + 76, + 469, + 321, + 484 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recent methods [1, 2, 10, 17, 37, 40, 46, 51] have shown that, as a continuous and differentiable function realized by Multi-Layer Perceptron, memory efficient deep implicit function can represent detailed 3D shapes with arbitrary topology and have no resolution limitations. It can also be extended to RGB image-based 3D reconstruction [40, 44, 46, 57]. Grasping Field [30] proposes to reconstruct hands grasping object via predicting signed distances for hand and object given the input point cloud or RGB image. Unlike previous implicit functions relying on a global feature vector, given an input RGB image with cleanly segmented foreground, Pixel-aligned Implicit Function [49] (PIFu) utilizes pixel-aligned image feature and depth value of each 3D point for occupancy prediction. PIFu-based methods have shown the ability to generate high-fidelity 3D human geometry with more clothes details [27, 31, 50, 62] and robust 6D pose estimation. For instance, Neural Correspondence Field [24] predicts signed distance and 3D object coordinate for each 3D point to establish dense 3D-3D correspondences for robust 6D object pose estimation.", + "bbox": [ + 75, + 492, + 470, + 795 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. The Proposed Method", + "text_level": 1, + "bbox": [ + 76, + 806, + 287, + 823 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 76, + 830, + 217, + 845 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Pixel-Aligned Implicit Function (PIFu) [49]. PIFu, by learning an implicit function in 3D domain with pixel-aligned features, generates detailed 3D human geometry in", + "bbox": [ + 75, + 854, + 468, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "cluding largely occluded regions from a segmented RGB image. Specifically, given a segmented image $I:\\mathbb{R}^2\\mapsto \\mathbb{R}^3$ PIFu first extracts a feature map $F:\\mathbb{R}^2\\mapsto \\mathbb{R}^C$ with $C$ channels by a fully convolutional image encoder. Then, for a 3D point $\\pmb {p}\\in \\mathbb{R}^3$ in the camera space and its 2D projection $\\pi (\\pmb {p}):\\mathbb{R}^3\\mapsto \\mathbb{R}^2$ on $I$ , the implicit function, represented by a Multi-Layer Perceptron (MLP) $f_{\\mathrm{PIF_u}}:\\mathbb{R}^C\\times \\mathbb{R}\\mapsto$ $\\mathbb{R}$ as $f_{\\mathrm{PIF_u}}(S(F,\\pi (\\pmb {p})),Z(\\pmb {p})) = y$ maps the pixel-aligned feature $S(F,\\pi (\\pmb {p}))$ and the depth $Z(\\pmb {p})$ of $\\pmb{p}$ to the occupancy $y\\in [0,1]$ (1 means $\\pmb{p}$ is inside the human surface and 0 means otherwise). $S$ is the sampling function (e.g., bilinear interpolation) to sample the value of $F$ at pixel $\\pi (\\pmb {p})$", + "bbox": [ + 496, + 356, + 890, + 537 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Signed Distance Function (SDF) [12, 46]. Given a 3D shape $M \\in \\mathcal{M}$ , a signed distance function is a continuous function $\\phi: \\mathbb{R}^3 \\times \\mathcal{M} \\mapsto \\mathbb{R}$ as $\\phi(\\pmb{p}, M) = s$ , mapping a 3D point $\\pmb{p}$ to the signed distance $s$ between $\\pmb{p}$ and the shape surface. The sign denotes if $\\pmb{p}$ is outside (positive) or inside (negative) of a watertight surface, with the iso-surface $\\phi(\\pmb{p}, M) = 0$ implicitly representing the surface.", + "bbox": [ + 496, + 546, + 890, + 652 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Dense Offset-based Pose Re-Parameterization", + "text_level": 1, + "bbox": [ + 500, + 662, + 885, + 679 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As opposed to holistic regression directly mapping the input data to sparse and absolute joint positions and 2D dense regression which mainly exploits the 2D input local evidence, our method aims to directly model in 3D domain and thus fully exploit the 3D dense local evidence and hand global structure. To achieve this goal, given a common sparse and explicit pose representation (i.e., a set of 3D coordinates of $T$ hand joints) as $J = \\{j_{t}\\}_{t=1}^{T} \\in \\mathcal{J}$ , $j_{t} \\in \\mathbb{R}^{3}$ , we parameterize it as a set of dense 3D point-wise offset-based representation. As shown in previous works [16,54], it is non-ideal to directly learn the 3D offset vector from each 3D point to a joint due to the large variance of offsets and far away joints are also beyond the scope of receptive field. Thus, given a camera-space 3D hand pose $J$ and", + "bbox": [ + 496, + 689, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "8971", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "shape $M$ , we define the pose parameterization as a function $\\psi$ , mapping from a 3D point $\\pmb{p}$ in the camera space to a set of 4D offset vectors $V = \\{\\pmb{v}_t\\}_{t=1}^T$ , $\\pmb{v}_t \\in \\mathbb{R}^4$ :", + "bbox": [ + 76, + 90, + 468, + 137 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\psi : \\mathbb {R} ^ {3} \\times \\mathcal {J} \\times \\mathcal {M} \\mapsto \\mathbb {R} ^ {T \\times 4} \\text {a s} \\psi (\\boldsymbol {p}, J, M) = V. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 146, + 468, + 162 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Each offset vector $\\pmb{v}_t = (w_t, \\pmb{d}_t)$ consists of a 1D voting weight $w_t \\in \\mathbb{R}$ as point-wise closeness to hand joint $j_t$ :", + "bbox": [ + 76, + 174, + 468, + 205 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nw _ {t} = \\left\\{ \\begin{array}{c c} 1 - \\frac {\\| \\boldsymbol {j} _ {t} - \\boldsymbol {p} \\| _ {2}}{r} & | s | < \\delta \\text {a n d} \\| \\boldsymbol {j} _ {t} - \\boldsymbol {p} \\| _ {2} \\leq r \\text {a n d} \\boldsymbol {p} \\in B _ {t} ^ {K}, \\\\ 0 & \\text {o t h e r w i s e}; \\end{array} \\right. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 214, + 496, + 262 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and a 3D unit directional vector $\\pmb{d}_t \\in \\mathbb{R}^3$ as point-wise direction to hand joint $j_t$ :", + "bbox": [ + 76, + 263, + 468, + 294 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {d} _ {t} = \\left\\{ \\begin{array}{c c} \\frac {\\boldsymbol {j} _ {t} - \\boldsymbol {p}}{\\| \\boldsymbol {j} _ {t} - \\boldsymbol {p} \\| _ {2}} & | s | < \\delta \\text {a n d} \\| \\boldsymbol {j} _ {t} - \\boldsymbol {p} \\| _ {2} \\leq r \\text {a n d} \\boldsymbol {p} \\in B _ {t} ^ {K}, \\\\ 0 & \\text {o t h e r w i s e}; \\end{array} \\right. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 303, + 468, + 354 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $s = \\phi(\\pmb{p}, M)$ and $|s| < \\delta$ means points that are within clamping distance $\\delta$ from the hand surface. $r$ denotes the radius of a 3D ball centered at each joint and $B_t^K$ denotes the set of $K$ -Nearest Neighbors (KNN) to joint $\\pmb{j}_t$ among all the near-surface points (i.e., points with $|s| < \\delta$ ). Thus, among all the 3D points that are in the hand surface vicinity, we assign the non-zero 4D offset vector $\\pmb{v}_t$ to the $K$ nearest points inside the 3D ball with radius $r$ of joint $\\pmb{j}_t$ .", + "bbox": [ + 76, + 356, + 468, + 478 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Neural Voting Field", + "text_level": 1, + "bbox": [ + 76, + 484, + 264, + 501 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To fully exploit the 3D local evidence and hand global geometry for camera-space 3D hand pose estimation, based on the aforementioned signed distance function and pose parameterization, we unify both modalities using a pixel-aligned implicit function. Following the proposed pipeline in Fig. 2, we first extract a $C$ -channel feature map $F$ from the input RGB image $I$ by a hourglass network $g(I; \\boldsymbol{\\eta})$ as used in [24, 49]. Then, for a 3D query point $\\pmb{p}$ sampled in camera space, we define a continuous implicit function $f_{\\mathrm{NVF}}$ realized by a MLP, which maps from pixel-aligned image feature $S(F, \\pi(\\pmb{p}))$ and depth $Z(\\pmb{p})$ of point $\\pmb{p}$ to its signed distance $s$ and a set of 4D offset vectors $V$ :", + "bbox": [ + 75, + 508, + 468, + 690 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} f _ {\\mathrm {N V F}}: \\mathbb {R} ^ {C} \\times \\mathbb {R} \\mapsto \\mathbb {R} \\times \\mathbb {R} ^ {T \\times 4} \\quad \\text {a s} \\\\ f _ {\\mathrm {s}}: (C (E _ {\\mathrm {s}} (\\cdot)), Z (\\cdot), \\boldsymbol {a}) \\quad (\\cdot , V) \\end{array} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 700, + 468, + 723 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {\\mathrm {N V F}} \\big (S (F, \\pi (\\boldsymbol {p})), Z (\\boldsymbol {p}); \\boldsymbol {\\theta} \\big) = (s, V),\n$$\n", + "text_format": "latex", + "bbox": [ + 147, + 717, + 398, + 733 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\eta$ and $\\theta$ are the parameters of the hourglass $g$ and MLP $f_{\\mathrm{NVF}}$ , respectively. We use a pinhole camera model for 3D-to-2D projection. For each 3D point, besides the set of 4D offset vectors, NVF also predicts its signed distance to the hand surface, helping to learn the global hand geometry and also to select near-surface 3D points by thresholding the signed distances. The near-surface 3D points are then considered as valid 3D voters to be used in the voting stage.", + "bbox": [ + 75, + 743, + 468, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Optimization. During training, we sample $N$ 3D points $P = \\{\\pmb{p}_n\\}_{n=1}^N$ in camera frustum for each image and apply", + "bbox": [ + 76, + 869, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "direct supervision to both predicted signed distance $s_n$ as $L_s$ and the predicted set of 4D offset vectors $V_n$ as $L_V$ . Given the ground-truth signed distance $\\hat{s}_n = \\phi(\\pmb{p}_n, \\hat{M})$ and set of offset vectors $\\hat{V}_n = \\psi(\\pmb{p}_n, \\hat{J}, \\hat{M})$ , where $\\hat{J}$ and $\\hat{M}$ are the ground-truth pose and mesh, $L_s$ and $L_V$ are defined as:", + "bbox": [ + 498, + 90, + 890, + 166 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {s} = \\frac {1}{N} \\sum_ {n = 1} ^ {N} | \\operatorname {c l a m p} (\\hat {s} _ {n}, \\delta) - \\operatorname {c l a m p} (s _ {n}, \\delta) |, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 544, + 167, + 890, + 208 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {V} = \\frac {1}{N} \\sum_ {n = 1} ^ {N} \\mathbb {1} \\left(\\left| \\hat {s} _ {n} \\right| < \\delta\\right) H \\left(\\hat {V} _ {n}, V _ {n}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 210, + 890, + 252 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathrm{clamp}(s,\\delta)\\coloneqq \\min (\\delta ,\\max (-\\delta ,s))$ uses $\\delta$ to control the distance from the hand surface over which we want to main a metric SDF, following [46]. $H$ is the Huber loss [28] and $\\mathbb{1}(\\cdot)$ is an indicator function that selects near-surface points (i.e., points that are within clamping distance $\\delta$ from the hand surface). Then, We jointly train the hourglass network $g$ and MLP $f_{\\mathrm{NVF}}$ end-to-end by solving the following optimization problem with $\\lambda$ as the balancing weight:", + "bbox": [ + 496, + 257, + 890, + 377 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\eta} ^ {\\star}, \\boldsymbol {\\theta} ^ {\\star} = \\underset {\\boldsymbol {\\eta}, \\boldsymbol {\\theta}} {\\arg \\min } L _ {s} + \\lambda L _ {V}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 378, + 890, + 404 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Sampling of 3D Query Points. During training, unlike voxel-based methods which require discretization of ground-truth 3D meshes, we directly sample 3D points on the fly from the ground-truth mesh following [24, 49]. To make NVF fully exploit meaningful local evidence and understand the hand global geometry, it is crucial to focus on the hand by sampling the query points more densely around the hand surface during training. Thus, given the 3D camera-space hand model, we first sample 12500 points nearby the surface, 1000 points inside the bounding sphere of the model, and 1000 points inside the camera frustum. Among all these points, we then sample 2500 points inside the model and 2500 points outside to form a set of 5000 points in total for training. During inference, the points are directly sampled at centers of voxels that fill up the same camera frustum without any ground-truth information used.", + "bbox": [ + 496, + 410, + 890, + 652 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Dense 3D Point-to-Joint Voting.", + "text_level": 1, + "bbox": [ + 500, + 659, + 777, + 674 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "During inference, assuming $N$ 3D points $P = \\{\\pmb{p}_n\\}_{n=1}^N$ sampled in camera frustum for each image, for each point $\\pmb{p}_n$ that is considered as a valid 3D voter (i.e., the predicted signed distance $s_n$ is below the threshold $\\delta$ ), based on its predicted set of 4D offset vectors $V_n = \\{\\pmb{v}_t^n\\}_{t=1}^T$ , we first convert each 4D offset vector $\\pmb{v}_t^n = (w_t^n, \\pmb{d}_t^n)$ into the actual 3D offset $\\pmb{o}_t^n$ from $\\pmb{p}_n$ to joint $\\pmb{j}_t$ based on Eq. 2 and Eq. 3:", + "bbox": [ + 496, + 681, + 890, + 787 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {o} _ {t} ^ {n} = \\mathbb {1} \\left(\\left| s _ {n} \\right| < \\delta\\right) \\left[ r \\left(1 - w _ {t} ^ {n}\\right) \\boldsymbol {d} _ {t} ^ {n} \\right]. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 792, + 890, + 809 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then, each camera-space joint location $\\mathbf{j}_t$ is calculated simply by a weighted average over each voter's 3D offset prediction:", + "bbox": [ + 496, + 814, + 890, + 858 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {j} _ {t} = \\sum_ {n = 1} ^ {N} \\frac {\\mathbb {1} \\left(\\left| s _ {n} \\right| < \\delta\\right) w _ {t} ^ {n} \\left(\\boldsymbol {o} _ {t} ^ {n} + \\boldsymbol {p} _ {n}\\right)}{\\sum_ {n = 1} ^ {N} \\mathbb {1} \\left(\\left| s _ {n} \\right| < \\delta\\right) w _ {t} ^ {n}}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 862, + 890, + 904 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "8972", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, we obtain the final camera-space 3D hand pose as $J = \\{j_{t}\\}_{t=1}^{T}$ . Furthermore, instead of using all the near-surface points as valid voters, we can also only focus on using near-surface points with high predicted voting weight, as a higher voting weight means the point is more confident and closer to a hand joint. Therefore, we use voting fraction as a parameter to control the fraction of near-surface points with the highest voting weight for voting. Related ablation studies will be given in Sec. 4.7.", + "bbox": [ + 76, + 90, + 472, + 227 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 241, + 210, + 258 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Baseline Methods", + "text_level": 1, + "bbox": [ + 76, + 266, + 250, + 281 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To demonstrate the effectiveness of our proposed 3D dense regression scheme against classic holistic regression and 2D dense regression, we design two directly comparable baselines that share the same architecture of the hourglass network and the MLP as NVF described in Sec. 3. Specifically, similar to NVF, both baselines first extract $C$ -channel feature map from the input RGB image $I$ . Then, unlike our $f_{\\mathrm{NVF}}$ which takes a sampled 3D query point's pixel-aligned feature $S(F,\\pi (\\pmb {p}))$ and its depth $Z(\\pmb {p})$ to predict its signed distance $s$ and the set of 4D offset vectors $V$ , the two baselines take the following steps.", + "bbox": [ + 75, + 290, + 472, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baseline-Holistic. We directly apply a global average pooling $\\alpha$ to its feature map $F^{\\prime}$ extracted by the hourglass network $g(I;\\pmb {\\eta}^{\\prime})$ and generate a $C$ -dimension feature vector $\\alpha (F^{\\prime})$ . The feature vector is then passed through the MLP $f_{\\mathrm{HOL}}:\\mathbb{R}^{C}\\mapsto \\mathbb{R}^{T\\times 3}$ as $f_{\\mathrm{HOL}}(\\alpha (F^{\\prime});\\pmb {\\theta}^{\\prime}) = J^{\\prime}$ $f_{\\mathrm{HOL}}$ directly outputs $T$ 3D hand joint coordinates $J^{\\prime} = \\{\\pmb{j}_{t}^{\\prime}\\}_{t = 1}^{T}$ in camera space. Given the ground-truth pose $\\hat{J}$ , it is trained by solving the following optimization problem with a loss function $L_{J^{\\prime}} = H(\\hat{J},J^{\\prime})$ ..", + "bbox": [ + 76, + 462, + 472, + 599 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\eta} ^ {\\prime \\star}, \\boldsymbol {\\theta} ^ {\\prime \\star} = \\underset {\\boldsymbol {\\eta} ^ {\\prime}, \\boldsymbol {\\theta} ^ {\\prime}} {\\arg \\min } L _ {J ^ {\\prime}}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 604, + 468, + 633 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baseline-2D-Dense. We adopt a 2D dense regression scheme to estimate the camera-space pose via dense 2D pixel-wise voting, which is different from NVF which is based on dense 3D point-wise voting. Given the feature map $F''$ extracted by the hourglass network $g(I; \\boldsymbol{\\eta}^{\\prime \\prime})$ , we only pass the pixel-aligned feature $F''(\\boldsymbol{u})$ at each 2D pixel location $\\boldsymbol{u}$ through the MLP $f_{\\mathrm{DEN}}: \\mathbb{R}^C \\mapsto \\mathbb{R} \\times \\mathbb{R}^{T \\times 4}$ as $f_{\\mathrm{DEN}}(F''(\\boldsymbol{u}); \\boldsymbol{\\theta}'') = (e, V'')$ . $f_{\\mathrm{DEN}}$ outputs the probability $e \\in [0,1]$ that the hand is present at $\\boldsymbol{u}$ and also a set of 4D vectors $V'' = \\{\\boldsymbol{v}_t''\\}_{t=1}^T$ . Each 4D vector $\\boldsymbol{v}_t'' = (w_t'', j_t'')$ contains a 1D voting weight $w_t''$ and 3D joint coordinate $j_t''$ . For the 1D voting weight, it is defined similarly to the one in NVF. We first use the camera ray for each pixel to find its first 3D intersection point with the 3D hand model in the camera space. Then, we directly follow Eq. 2 to calculate the pixel's voting weight as the closeness to each joint. It is trained by solving the following optimization", + "bbox": [ + 75, + 638, + 472, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "tion problem with $\\lambda''$ as the balancing weight:", + "bbox": [ + 498, + 89, + 807, + 107 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\boldsymbol {\\eta} ^ {\\prime \\prime \\star}, \\boldsymbol {\\theta} ^ {\\prime \\prime \\star} = \\underset {\\boldsymbol {\\eta} ^ {\\prime \\prime}, \\boldsymbol {\\theta} ^ {\\prime \\prime}} {\\arg \\min } L _ {e} + \\lambda^ {\\prime \\prime} L _ {V ^ {\\prime \\prime}} (11) \\\\ = \\underset {\\boldsymbol {\\eta} ^ {\\prime \\prime}, \\boldsymbol {\\theta} ^ {\\prime \\prime}} {\\arg \\min } \\frac {1}{L} \\left[ \\sum_ {l = 1} ^ {L} E \\left(\\hat {e} _ {l}, e _ {l}\\right) + \\lambda^ {\\prime \\prime} \\sum_ {l = 1} ^ {L} \\hat {e} _ {l} H \\left(\\hat {V} _ {l} ^ {\\prime \\prime}, V _ {l} ^ {\\prime \\prime}\\right) \\right], (12) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 504, + 112, + 890, + 205 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $L$ is the number of pixels, $E$ is the binary cross-entropy loss, $\\hat{e}_l$ is given by the ground-truth hand mask, and $\\hat{V}_l^{\\prime \\prime}$ is the ground-truth set of 4D vectors. During inference, for pixel $\\pmb{u}_l$ with predicted probability $e_l > 0.5$ (i.e., predicted as the foreground pixel), its predicted voting weight and 3D joint location are used to calculate the camera-space pose $J^{\\prime \\prime} = \\{\\dot{j}_t^{\\prime \\prime}\\}_{t = 1}^{T}$ by a weighted average, similar to Eq. 9.", + "bbox": [ + 498, + 214, + 893, + 324 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Dataset and Evaluation Metrics", + "text_level": 1, + "bbox": [ + 500, + 329, + 779, + 344 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "FreiHAND Dataset [64]. As a real 3D hand dataset, it provides 130,240 training and 3,960 evaluation samples. The result of the evaluation set is evaluated via the official server for online evaluation. Since FreiHAND provides hand scale during evaluation time, we provide an ablation on hand scale for analysis of 2D-to-3D scale and depth ambiguity.", + "bbox": [ + 496, + 353, + 890, + 444 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Metrics. For FreiHAND, we evaluate for the task of camera-space 3D hand pose estimation. As in [8,64], we report the unaligned version of Mean Joint Error (mm) which is measured in camera space as CS-MJE and the Area-Under-the-Curve of PCK (Percentage of Correct Keypoints) vs. error thresholds as CS-AUC to evaluate the absolute 3D hand pose. We also define and report a 3D Translational Error (mm) (TE) [23] and Depth Error (mm) (DE) to evaluate the predicted and ground-truth pose's centroid and the depth-axis of the centroid, respectively. The results of TE and DE will be used in analysis for baseline studies, especially for the analysis of scale and depth ambiguity.", + "bbox": [ + 496, + 445, + 892, + 626 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "HO3D Dataset [18]. As a real 3D hand-object dataset, it contains 66,034 training samples and 11,524 evaluation samples. We use the official server for online evaluation as the annotations of the evaluation set are not available.", + "bbox": [ + 496, + 627, + 890, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Metrics. For HO3D, we evaluate for the task of rootrelative 3D hand pose estimation. As in [18, 19], we report the Mean Joint Error (mm) after scale-translation alignment of the root as MJE and corresponding AUC to evaluate the pose. We also follow [45] to add ground-truth root to the predicted root-relative pose and report error (mm) as RS-MJE as it is equivalent to measuring in root-relative space.", + "bbox": [ + 496, + 686, + 890, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Complement (Comp) Dataset [7]. As a synthetic 3D hand dataset, it provides 1,520 poses and 71 viewpoints for research use. We only use it for pre-training when needed.", + "bbox": [ + 496, + 794, + 890, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Implementation Details", + "text_level": 1, + "bbox": [ + 500, + 847, + 718, + 863 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For NVF and two baselines, the network architecture is adopted from PIFu [49]. Specifically, we use hourglass net", + "bbox": [ + 500, + 869, + 890, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "8973", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/7461c6fc6319113b746718d4d7078b060c9ecf4f7552bbeaffb02f8a2837dad8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodExtra DataHand CropHand ScaleCS-MJE↓
ObMan [22]-✓✗85.2
MANO CNN [64]-✓✗71.3
I2L-MeshNet [42]-✓✗60.3
CMR-SG-RN18 [8]-✓✗49.7
CMR-SG-RN50 [8]-✓✗48.8
Baseline-Holisitc-✗✗54.5
Baseline-2D-Dense-✗✗53.2
CS-NVF (Ours)-✗✗47.2
Baseline-Holisitc-✗✓50.4
Baseline-2D-Dense-✗✓49.0
CS-NVF (Ours)-✗✓42.4
CS-NVF (Ours)Comp*✗✗44.6
CS-NVF (Ours)Comp*✗✓39.3
", + "bbox": [ + 78, + 88, + 491, + 321 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "work with the output stride of 4 and output channel of 256 as the image encoder. All the MLPs $f_{\\mathrm{NVF}}$ , $f_{\\mathrm{HOL}}$ , and $f_{\\mathrm{DEN}}$ have four hidden fully-connected layers with the number of neurons as (1024, 512, 256, 128). Unless stated otherwise, during training for NVF, the clamping distance $\\delta$ is set to 5 mm. The KNN is set to 1024. The ball radius $r$ is set to 80 mm. During inference, the sampling step size of 3D points is 16 mm in all three axes. The voting fraction is set to 50%. Unlike PIFu, no image segmentation is applied. Camera intrinsic parameters are assumed known. Base on the above setting, for camera-space 3D hand pose estimation, we directly sample 3D points in camera space, as described in Sec. 3.3. We call NVF working in camera space as CS-NVF. Unlike other methods in Tab. 2, CS-NVF and both baselines take the original image as input without hand detection or cropping. For the task of root-relative 3D hand pose estimation, we follow similar sampling strategy but sample in a hand root-relative 3D cube instead of the camera frustum. We then call NVF working in root-relative space as RS-NVF. RS-NVF takes cropped hand-centered image. Please refer to the supplementary for more details.", + "bbox": [ + 75, + 378, + 472, + 698 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Baseline Studies", + "text_level": 1, + "bbox": [ + 76, + 700, + 238, + 715 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, based on the results from CS-NVF and both baselines (as in Tab. 2, Tab. 3, Fig. 3, and Fig. 4), we investigate the impact of 2D-to-3D scale and depth ambiguity for the task of absolute 3D hand pose estimation and the effectiveness of using 3D dense regression over both 2D dense regression and holistic regression.", + "bbox": [ + 75, + 718, + 468, + 809 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "2D-to-3D Scale and Depth Ambiguity. Estimating absolute 3D pose from an RGB image is an ill-posed problem due to 2D-to-3D scale and depth ambiguity [13, 29]. In view of this matter, FreiHAND [64] provides hand scale during evaluation for participants to use. As most existing works do not use the provided hand scale and thus, no", + "bbox": [ + 75, + 810, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/02a2bfc56059a68fe8eb0364a80bc7bae21013a2f686f3ad1c99cb0cab405264.jpg", + "table_caption": [ + "Table 2. Comparison for absolute 3D hand pose on FreiHAND.*: pre-training on Comp. Note that our CS-NVF and two baselines take the original image without hand detection or cropping." + ], + "table_footnote": [], + "table_body": "
MethodExtra DataHand ScaleTE↓DE↓
Baseline-Holisitc-X50.649.1
Baseline-2D-Dense-X49.247.9
CS-NVF (Ours)-X43.642.4
Baseline-Holisitc-✓46.945.5
Baseline-2D-Dense-✓45.343.9
CS-NVF (Ours)-✓38.937.8
Baseline-HolisitcComp*X48.747.1
Baseline-2D-DenseComp*X47.946.4
CS-NVF (Ours)Comp*X41.540.4
Baseline-HolisitcComp*✓41.740.1
Baseline-2D-DenseComp*✓40.538.8
CS-NVF (Ours)Comp*✓36.535.5
", + "bbox": [ + 522, + 103, + 893, + 321 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Comparison of 3D Translational and Depth Error for absolute 3D hand pose on FreiHAND. \\*: pre-training on Comp.", + "bbox": [ + 511, + 325, + 906, + 354 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a79c1edbd3877a69c57f267f417210b4558e36cde782abb28345014c76ec10c8.jpg", + "image_caption": [ + "Figure 3. 3D PCK for absolute 3D hand pose on FreiHAND." + ], + "image_footnote": [], + "bbox": [ + 501, + 377, + 890, + 601 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "relevant ablation on hand scale exists, we provide ablation results on hand scale in Tab. 2 and Tab. 3 for CS-NVF and two baselines. First, based on results in Tab. 2 and Tab. 3, we observe that, for all three pipelines, the main cause of the error between the ground-truth and predicted hand joints is the global translation and especially the depth component as TE is close to CS-MJE and DE is close to TE. For instance, CS-NVF with hand scale and Comp pre-training achieves $39.3\\mathrm{mm}$ CS-MJE, while TE and DE are $36.5\\mathrm{mm}$ and $35.5\\mathrm{mm}$ , respectively. This validates the fact that depth ambiguity is indeed one of the main challenges for estimating accurate absolute 3D hand pose from a single RGB image. Moreover, for CS-NVF and the two baselines, using hand scale leads to significant improvements in terms of all three metrics. It is also worth noting that, for all three pipelines, results only using hand scale outperform results using only extra data. For instance, CS-NVF with hand scale achieves $42.4\\mathrm{mm}$ CS-MJE, while CS-NVF with Comp pre-training", + "bbox": [ + 496, + 628, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "8974", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ee33d041aa02eea1f78ed4791a18e2ba076dcd420e9cf3c2e06bb11b389f1eb2.jpg", + "image_caption": [ + "Figure 4. Qualitative results for absolute and relative 3D pose: (b) Predicted hand mask (Baseline) or mesh generated by Marching Cubes from the predicted signed distances (NVF). (c-h) Predicted 1D voting weight from each predicted foreground pixel (Baseline) or predicted near-surface 3D point (NVF) to a joint. (Brighter line means larger weight, darker or no line means smaller weight) (i) Estimated 3D pose. NVF makes solid point-wise prediction and overall distribution of signed distances and voting weights in highly occluded region." + ], + "image_footnote": [], + "bbox": [ + 81, + 85, + 893, + 321 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b5ff383930dd7e3582a71c01555fde7d5d3322d2f149490364b34fc52e5af2f0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodMJE↓AUC↑RS-MJE↓
Hasson et al. [20]36.90.369-
Pose2Mesh [11]33.30.48033.2
ObMan [22]31.80.46155.2
Liu et al. [38]31.70.46330.0
Hampali et al. [18]30.40.494-
METRO [35]28.90.504-
I2L-MeshNet [42]26.00.52926.8
Keypoint Trans. [19]25.70.553-
ArtiBoost [58]25.30.532-
Zheng et al. [61]25.10.541-
HandOccNet [45]24.00.55724.9
RS-NVF (Ours)21.80.61023.2
", + "bbox": [ + 89, + 386, + 455, + 589 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Comparison for relative 3D hand pose on HO3D.", + "bbox": [ + 89, + 593, + 452, + 607 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "has $44.6 \\mathrm{~mm}$ CJ-MJE. These verify that a reasonable understanding towards the hand scale would help for robust depth learning and alleviate depth ambiguity, which is also aligned with existing literature [13, 29, 33].", + "bbox": [ + 75, + 619, + 468, + 681 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Dense Regression vs. Holistic Regression. In Tab. 2 and Fig. 3, we can observe that both CS-NVF and Baseline-2D-Dense consistently outperform Baseline-Holistic on CS-MJE and CS-AUC under different training settings. For instance, without using hand scale, Baseline-2D-Dense is 1.3 mm better than Baseline-Holistic in CS-MJE, and CS-NVF can be 7.3 mm better. Since three pipelines share the same network architecture, this supports our point that dense regression is more effective than holistic regression for handling highly articulated pose data.", + "bbox": [ + 75, + 685, + 468, + 835 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3D Dense Regression vs. 2D Dense Regression. Quantitatively, in Tab. 2 and Fig. 3, it can be observed that CS-NVF consistently outperforms Baseline-2D-Dense by a large margin on CS-MJE and CS-AUC under different", + "bbox": [ + 75, + 839, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "training settings. For instance, without using hand scale, CS-NVF can be $6\\mathrm{mm}$ better on CS-MJE. Qualitatively, as indicated by the white circles in column (c-h) of Fig. 4, both methods are able to find valid voters (Baseline-2D-Dense: pixels with predicted probability $e > 0.5$ . CS-NVF: 3D points with predicted signed distance $|s| < \\delta$ ). However, CS-NVF generally provides not only more number of valid voters, but also more reasonable element-wise prediction and overall distribution of the voting weight than Baseline-2D-Dense. Specifically, for the number of voters, since CS-NVF models all the 3D points around the hand surface and Baseline-2D-Dense can only model foreground pixels, CS-NVF generally can find and use more valid voters. As for element-wise prediction and overall distribution of the voting weight, CS-NVF is generally able to predict large voting weight for points close to hand joint and small value for points far away, resulting in well-shaped symmetric distribution around the corresponding joint, even in occluded regions. Note that brighter line indicates larger voting weight, darker or no line indicates smaller voting weight. Moreover, as shown in column (b) for the mesh rendering of CS-NVF (obtained using Marching Cubes [39] from predicted signed distances), even in highly occluded region, CS-NVF provides solid signed distance distribution showing its ability to reason the global hand geometry. Overall, these demonstrate that CS-NVF, through direct dense modeling in 3D domain, can better model 3D dense local evidence, the relation between voters and 3D hand pose, and also the global hand geometry, compared with 2D dense regression-based counterpart, which leads to robust 3D hand pose estimation.", + "bbox": [ + 496, + 388, + 893, + 843 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Main Results: Absolute 3D Hand Pose", + "text_level": 1, + "bbox": [ + 498, + 849, + 831, + 864 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Tab. 2 and Fig. 3, we also compare CS-NVF with current state-of-the-art methods on FreiHAND for absolute 3D", + "bbox": [ + 498, + 869, + 892, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "8975", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6877abd3f39ecb9f95eecd3c285713a89eb1ded1e62d23e418d0ec0f5345b74d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Proposed MethodEvaluation MetricVoting Fraction (%)Clamping Distance (mm)KNN
25507510015102025651210242048
CS-NVF (Ours)CS-MJE↓39.939.339.540.239.339.340.040.140.339.539.339.3
CS-AUC↑0.360.370.360.360.370.370.360.360.350.360.370.37
RS-NVF (Ours)MJE↓22.221.821.922.021.921.821.821.923.622.621.821.8
AUC↑0.600.610.610.600.600.610.610.610.590.600.610.61
", + "bbox": [ + 78, + 87, + 895, + 183 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a4268fd0031444e54b8868a370150b169dd7bd01ee2c39554a953f41c7b9f364.jpg", + "table_caption": [ + "Table 5. Ablation study on parameter choices for absolute and relative 3D hand pose: Voting fraction, clamping distance, and KNN" + ], + "table_footnote": [], + "table_body": "
Step SizeCS-NVF (Ours)
#PointsCS-MJE↓CS-AUC↑FPS↑
41882k39.20.371
8230k39.20.378
1628k39.30.3734
248k39.80.3650
323k42.40.3456
", + "bbox": [ + 86, + 209, + 455, + 319 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Ablation study on sampling step size for absolute 3D hand pose on FreiHAND. #points: number of sampled 3D points.", + "bbox": [ + 75, + 324, + 468, + 353 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "hand pose estimation. In Tab. 2, without the use of hand scale and extra data as a fair comparison, CS-NVF surpasses previous state-of-the-art methods, achieving $1.6\\mathrm{mm}$ better than CMR [8]. Note that, unlike other methods, CS-NVF takes the original image as input without hand detection or cropping. Additional hand scale information used without extra data improves the CS-MJE of CS-NVF further to $42.4\\mathrm{mm}$ . Additional Comp pre-training improves the CS-MJE of CS-NVF to $44.6\\mathrm{mm}$ without hand scale and $39.3\\mathrm{mm}$ with hand scale. Referring to Fig. 3, CS-NVF also achieves state-of-the-art performance on CS-AUC.", + "bbox": [ + 75, + 362, + 468, + 527 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. Main Results: Relative 3D Hand Pose", + "text_level": 1, + "bbox": [ + 76, + 535, + 403, + 550 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Tab. 4, we compare RS-NVF with current state-of-the-art methods on HO3D for relative 3D hand pose estimation. RS-NVF significantly outperforms all these methods in all three metrics. Especially, RS-NVF outmatches HandOccNet [45] by $2.2\\mathrm{mm}$ on MJE, $1.7\\mathrm{mm}$ on RS-MJE, and also has 0.610 AUC against 0.557 AUC from HandOccNet.", + "bbox": [ + 75, + 555, + 468, + 647 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.7. Ablation Studies", + "text_level": 1, + "bbox": [ + 76, + 652, + 241, + 667 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Sampling Step Size. During inference, 3D query points are sampled at centers of voxels that fill up the 3D space. Since the choices of different sampling step sizes can influence both accuracy and efficiency, in Tab. 6, we show the ablation on sampling step size for CS-NVF. With $16\\mathrm{mm}$ as the step size, CS-NVF achieves 34 FPS and $39.3\\mathrm{mm}$ on CS-MJE given 28k points. Reducing the step size to $4\\mathrm{mm}$ yields 1,882k points and 1 FPS with slightly improved CS-MJE of $39.2\\mathrm{mm}$ . Increasing to $32\\mathrm{mm}$ step size yields 3k points and improves the FPS to 56 with still competitive $42.4\\mathrm{mm}$ on CS-MJE. For RS-NVF not shown in the table, 4 and 8 mm step size both yield MJE of $21.7\\mathrm{mm}$ and AUC of 0.62. We can observe that, while step size affects FPS, NVF is relatively insensitive regarding the performance. In all other reported results, we use $16\\mathrm{mm}$ as the default step size.", + "bbox": [ + 75, + 674, + 470, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Voting Fraction. Since the predicted point-wise voting weight reflects the confidence and closeness of a given point towards a certain joint, besides using all the near-surface points as valid voters, we can use a specified fraction of near-surface points with the highest voting weight for voting during inference. In Tab. 5, we observe that using all the near-surface points (100%) yields competitive results while 50% fraction leads to the overall better performance for both CS-NVF and RS-NVF. The results are aligned with our algorithm design since points far from each joint are not expected to make meaningful contributions. In all other reported results, we use 50% as the default voting fraction.", + "bbox": [ + 496, + 213, + 893, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Clamping Distance. During training, clamping distance $\\delta$ is used to concentrate NVF's capacity on learning high-quality point-wise prediction (i.e., signed distance and set of 4D offset vectors) near the hand surface. Ablation in Tab. 5 shows that $5\\mathrm{mm}$ generates the overall better performance for both CS-NVF and RS-NVF. In all other reported results, we use $5\\mathrm{mm}$ as the default clamping distance.", + "bbox": [ + 496, + 393, + 893, + 500 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "$K$ -Nearest Neighbors (KNN). During training, among all the sampled points that are in the hand surface vicinity, nonzero 4D offset vector will only be assigned to the $K$ nearest points inside a 3D ball of each hand joint. As shown in previous works [16, 54], estimating offsets for large KNN is unnecessary and far away joints are beyond the scope of receptive field. Small KNN might also cause poor performance due to the sparse learning scheme. Ablation in Tab. 5 shows that 1024 KNN generates the overall better performance for both CS-NVF and RS-NVF. In all other reported results, we use 1024 as the default KNN.", + "bbox": [ + 496, + 500, + 893, + 667 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusions", + "text_level": 1, + "bbox": [ + 500, + 679, + 627, + 694 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose Neural Voting Field (NVF), as the first 3D implicit representation-based unified solution to estimate camera-space 3D hand pose given a single RGB image. NVF follows a novel unified 3D dense regression scheme to estimate 3D hand pose in 3D camera space via dense 3D point-wise voting in camera frustum. Studies with both holistic regression baseline and 2D dense regression baseline verify NVF's ability to fully exploit 3D dense local evidence and model global hand structure/geometry, which are essential for precisely locating 3D hand joints in 3D camera space. NVF also demonstrates state-of-the-art results on both tasks of camera-space 3D hand pose estimation and root-relative 3D hand pose estimation.", + "bbox": [ + 496, + 703, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8976", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Matan Atzmon and Yaron Lipman. SAL: Sign agnostic learning of shapes from raw data. In CVPR, 2020. 3", + "[2] Matan Atzmon and Yaron Lipman. SALD: Sign agnostic learning with derivatives. In ICLR, 2021. 3", + "[3] Seungryul Baek, Kwang In Kim, and Tae-Kyun Kim. Pushing the envelope for rgb-based dense 3d hand pose estimation via neural rendering. In CVPR, 2019. 2", + "[4] Adnane Boukhayma, Rodrigo de Bem, and Philip HS Torr. 3d hand shape and pose from images in the wild. In CVPR, 2019. 1, 2", + "[5] Yujun Cai, Liuhao Ge, Jun Liu, Jianfei Cai, Tat-Jen Cham, Junsong Yuan, and Nadia Magnenat Thalmann. Exploiting spatial-temporal relationships for 3d pose estimation via graph convolutional networks. In ICCV, 2019. 1, 3", + "[6] Ping Chen, Yujin Chen, Dong Yang, Fangyin Wu, Qin Li, Qingpei Xia, and Yong Tan. I2uv-handnet: Image-to-uv prediction network for accurate and high-fidelity 3d hand mesh modeling. In ICCV, 2021. 1, 3", + "[7] Xingyu Chen, Yufeng Liu, Yajiao Dong, Xiong Zhang, Chongyang Ma, Yanmin Xiong, Yuan Zhang, and Xiaoyan Guo. Mobrecon: Mobile-friendly hand mesh reconstruction from monocular image. In CVPR, 2022. 1, 3, 5", + "[8] Xingyu Chen, Yufeng Liu, Chongyang Ma, Jianlong Chang, Huayan Wang, Tian Chen, Xiaoyan Guo, Pengfei Wan, and Wen Zheng. Camera-space hand mesh recovery via semantic aggregation and adaptive 2d-1d registration. In CVPR, 2021. 2, 3, 5, 6, 8", + "[9] Yujin Chen, Zhigang Tu, Di Kang, Linchao Bao, Ying Zhang, Xuefei Zhe, Ruizhi Chen, and Junsong Yuan. Model-based 3d hand reconstruction via self-supervised learning. In CVPR, 2021. 2", + "[10] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In CVPR, 2019. 3", + "[11] Hongsuk Choi, Gyeongsik Moon, and Kyoung Mu Lee. Pose2mesh: Graph convolutional network for 3d human pose and mesh recovery from a 2d human pose. In ECCV, 2020. 7", + "[12] Brian Curless and Marc Levoy. A volumetric method for building complex models from range images. SIGGRAPH, 1996. 3", + "[13] Tom van Dijk and Guido de Croon. How do neural networks see depth in single images? In ICCV, 2019. 2, 6, 7", + "[14] Linpu Fang, Xingyan Liu, Li Liu, Hang Xu, and Wenxiong Kang. Jgr-p2o: Joint graph reasoning based pixel-to-offset prediction network for 3d hand pose estimation from a single depth image. In ECCV, 2020. 1", + "[15] Liuhao Ge, Zhou Ren, Yuncheng Li, Zehao Xue, Yingying Wang, Jianfei Cai, and Junsong Yuan. 3d hand shape and pose estimation from a single rgb image. In CVPR, 2019. 1, 3", + "[16] Liuhao Ge, Zhou Ren, and Junsong Yuan. Point-to-point regression pointnet for 3d hand pose estimation. In ECCV, 2018. 1, 3, 8", + "[17] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. In ICML, 2020. 3" + ], + "bbox": [ + 78, + 114, + 467, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[18] Shreyas Hampali, Mahdi Rad, Markus Oberweger, and Vincent Lepetit. Honnotate: A method for 3d annotation of hand and object poses. In CVPR, 2020. 5, 7", + "[19] Shreyas Hampali, Sayan Deb Sarkar, Mahdi Rad, and Vincent Lepetit. Keypoint transformer: Solving joint identification in challenging hands and object interactions for accurate 3d pose estimation. In CVPR, 2022. 1, 3, 5, 7", + "[20] Yana Hasson, Bugra Tekin, Federica Bogo, Ivan Laptev, Marc Pollefeys, and Cordelia Schmid. Leveraging photometric consistency over time for sparsely supervised hand-object reconstruction. In CVPR, 2020. 7", + "[21] Yana Hasson, Gül Varol, Cordelia Schmid, and Ivan Laptev. Towards unconstrained joint hand-object reconstruction from rgb videos. In 3DV, 2021. 2", + "[22] Yana Hasson, Gül Varol, Dimitris Tzionas, Igor Kalevatykh, Michael J. Black, Ivan Laptev, and Cordelia Schmid. Learning joint reconstruction of hands and manipulated objects. In CVPR, 2019. 2, 6, 7", + "[23] Tomás Hodań, Jiří Matas, and Štepan Obdržálek. On evaluation of 6d object pose estimation. In ECCV, 2016. 5", + "[24] Lin Huang, Tomas Hodan, Lingni Ma, Linguang Zhang, Luan Tran, Christopher Twigg, Po-Chen Wu, Junsong Yuan, Cem Keskin, and Robert Wang. Neural correspondence field for object pose estimation. In ECCV, 2022. 2, 3, 4", + "[25] Lin Huang, Jianchao Tan, Ji Liu, and Junsong Yuan. Handtransformer: non-autoregressive structured modeling for 3d hand pose estimation. In ECCV, 2020. 1", + "[26] Lin Huang, Boshen Zhang, Zhilin Guo, Yang Xiao, Zhiguo Cao, and Junsong Yuan. Survey on depth and rgb image-based 3d hand shape and pose estimation. VRIH, 2021. 1", + "[27] Zeng Huang, Yuanlu Xu, Christoph Lassner, Hao Li, and Tony Tung. Arch: Animatable reconstruction of clothed humans. In CVPR, 2020. 2, 3", + "[28] Peter J Huber. Robust estimation of a location parameter. Breakthroughs in statistics, 1992. 4", + "[29] Umar Iqbal, Pavlo Molchanov, Thomas Breuel Juergen Gall, and Jan Kautz. Hand pose estimation via latent 2.5d heatmap regression. In ECCV, 2018. 1, 2, 6, 7", + "[30] Korrawe Karunratanakul, Jinlong Yang, Yan Zhang, Michael J. Black, Krikamol Muandet, and Siyu Tang. Grasping field: Learning implicit representations for human grasps. In 3DV, 2020. 3", + "[31] Nilesh Kulkarni, Justin Johnson, and David F Fouhey. What's behind the couch? directed ray distance functions (drdf) for 3d scene reconstruction. In arXiv e-prints, 2021. 3", + "[32] Dominik Kulon, Riza Alp Guler, Iasonas Kokkinos, Michael M Bronstein, and Stefanos Zafeiriou. Weakly-supervised mesh-convolutional hand reconstruction in the wild. In CVPR, 2020. 1, 3", + "[33] Vincent Lepetit, Francesc Moreno-Noguer, and Pascal Fua. Epnp: An accurate o (n) solution to the pnp problem. IJCV, 2009. 2, 7", + "[34] Shile Li and Dongheui Lee. Point-to-pose voting based hand pose estimation using residual permutation equivariant layer. In CVPR, 2019. 1", + "[35] Kevin Lin, Lijuan Wang, and Zicheng Liu. End-to-end human pose and mesh reconstruction with transformers. In CVPR, 2021. 1, 3, 7" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "8977", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[36] Kevin Lin, Lijuan Wang, and Zicheng Liu. Mesh graphormer. In ICCV, 2021. 1, 3", + "[37] Feng Liu, Luan Tran, and Xiaoming Liu. Fully understanding generic objects: Modeling, segmentation, and reconstruction. In CVPR, 2021. 3", + "[38] Shaowei Liu, Hanwen Jiang, Jiarui Xu, Sifei Liu, and Xiaolong Wang. Semi-supervised 3d hand-object poses estimation with interactions in time. In CVPR, 2021. 7", + "[39] William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3d surface construction algorithm. SIGGRAPH, 1987. 7", + "[40] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In CVPR, 2019. 3", + "[41] Gyeongsik Moon, Ju Yong Chang, and Kyoung Mu Lee. Camera distance-aware top-down approach for 3d multiperson pose estimation from a single rgb image. In ICCV, 2019. 2", + "[42] Gyeongsik Moon and Kyoung Mu Lee. I2l-meshnet: Image-to-lixel prediction network for accurate 3d human pose and mesh estimation from a single rgb image. In ECCV, 2020. 1, 2, 6, 7", + "[43] Alejandro Newell, Kaiyu Yang, and Jia Deng. Stacked hourglass networks for human pose estimation. In ECCV, 2016. 1", + "[44] Michael Oechsle, Lars Mescheder, Michael Niemeyer, Thilo Strauss, and Andreas Geiger. Texture fields: Learning texture representations in function space. In ICCV, 2019. 3", + "[45] JoonKyu Park, Yeonguk Oh, Gyeongsik Moon, Hongsuk Choi, and Kyoung Mu Lee. Handocctnet: Occlusion-robust 3d hand mesh estimation network. In CVPR, 2022. 1, 2, 3, 5, 7, 8", + "[46] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In CVPR, 2019. 3, 4", + "[47] Georgios Pavlakos, Xiaowei Zhou, Konstantinos G Derpanis, and Kostas Daniilidis. Coarse-to-fine volumetric prediction for single-image 3d human pose. In CVPR, 2017. 1", + "[48] Javier Romero, Dimitrios Tzionas, and Michael J Black. Embodied hands: Modeling and capturing hands and bodies together. TOG, 2022. 2", + "[49] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In ICCV, 2019. 1, 2, 3, 4, 5", + "[50] Shunsuke Saito, Tomas Simon, Jason Saragih, and Hanbyul Joo. Pifuhd: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization. In CVPR, 2020. 2, 3", + "[51] Vincent Sitzmann, Eric Chan, Richard Tucker, Noah Snavely, and Gordon Wetzstein. Metasdf: Meta-learning signed distance functions. In NeurIPS, 2020. 3", + "[52] Adrian Spurr, Jie Song, Seonwook Park, and Otmar Hilliges. Cross-modal deep variational hand pose estimation. In CVPR, 2018. 1, 2" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[53] Tze Ho Elden Tse, Kwang In Kim, Ales Leonardis, and Hyung Jin Chang. Collaborative learning for hand and object reconstruction with attention-guided graph convolution. In CVPR, 2022. 1, 2, 3", + "[54] Chengde Wan, Thomas Probst, Luc Van Gool, and Angela Yao. Dense 3d regression for hand pose estimation. In CVPR, 2018. 1, 3, 8", + "[55] Shih-En Wei, Varun Ramakrishna, Takeo Kanade, and Yaser Sheikh. Convolutional pose machines. In CVPR, 2016. 1", + "[56] Fu Xiong, Boshen Zhang, Yang Xiao, Zhiguo Cao, Taidong Yu, Joey Tianyi Zhou, and Junsong Yuan. A2j: Anchor-to-joint regression network for 3d articulated pose estimation from a single depth image. In ICCV, 2019. 1", + "[57] Qiangeng Xu, Weiyue Wang, Duygu Ceylan, Radomir Mech, and Ulrich Neumann. Disn: Deep implicit surface network for high-quality single-view 3d reconstruction. In NeurIPS, 2019. 3", + "[58] Lixin Yang, Kailin Li, Xinyu Zhan, Jun Lv, Wenqiang Xu, Jiefeng Li, and Cewu Lu. Artibost: Boosting articulated 3d hand-object pose estimation via online exploration and synthesis. In CVPR, 2022. 7", + "[59] Linlin Yang, Shile Li, Dongheui Lee, and Angela Yao. Aligning latent spaces for 3d hand pose estimation. In ICCV, 2019. 1, 2", + "[60] Xiong Zhang, Qiang Li, Hong Mo, Wenbo Zhang, and Wen Zheng. End-to-end hand mesh recovery from a monocular rgb image. In ICCV, 2019. 1, 2", + "[61] Xiaozheng Zheng, Pengfei Ren, Haifeng Sun, Jingyu Wang, Qi Qi, and Jianxin Liao. Joint-aware regression: Rethinking regression-based method for 3d hand pose estimation. In BMVC, 2021. 7", + "[62] Zerong Zheng, Tao Yu, Yebin Liu, and Qionghai Dai. Pamir: Parametric model-conditioned implicit representation for image-based human reconstruction. TPAMI, 2021. 2, 3", + "[63] Christian Zimmermann and Thomas Brox. Learning to estimate 3d hand pose from single rgb images. In ICCV, 2017. 1, 2", + "[64] Christian Zimmermann, Duygu Ceylan, Jimei Yang, Bryan Russell, Max Argus, and Thomas Brox. Freihand: A dataset for markerless capture of hand pose and shape from single rgb images. In ICCV, 2019. 5, 6" + ], + "bbox": [ + 501, + 92, + 890, + 686 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "8978", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Neural Voting Field for Camera-Space 3D Hand Pose Estimation/96d05801-c2f5-4205-ac8b-2fcd00428ae6_model.json b/2023/Neural Voting Field for Camera-Space 3D Hand Pose Estimation/96d05801-c2f5-4205-ac8b-2fcd00428ae6_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8d9d2cd5a9c4110e16a4fb2bbc41f0de152d488d --- /dev/null +++ b/2023/Neural Voting Field for Camera-Space 3D Hand Pose Estimation/96d05801-c2f5-4205-ac8b-2fcd00428ae6_model.json @@ -0,0 +1,2277 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.158, + 0.131, + 0.815, + 0.154 + ], + "angle": 0, + "content": "Neural Voting Field for Camera-Space 3D Hand Pose Estimation" + }, + { + "type": "text", + "bbox": [ + 0.158, + 0.181, + 0.261, + 0.2 + ], + "angle": 0, + "content": "Lin Huang*" + }, + { + "type": "text", + "bbox": [ + 0.242, + 0.181, + 0.484, + 0.2 + ], + "angle": 0, + "content": "\\(g^{1*}\\) Chung-Ching Lin" + }, + { + "type": "text", + "bbox": [ + 0.242, + 0.2, + 0.546, + 0.234 + ], + "angle": 0, + "content": "Lijuan Wang2 Junsong Yuan1 \n1University at Buffalo" + }, + { + "type": "text", + "bbox": [ + 0.559, + 0.181, + 0.782, + 0.2 + ], + "angle": 0, + "content": "Kevin Lin² Lin Lia" + }, + { + "type": "text", + "bbox": [ + 0.621, + 0.2, + 0.728, + 0.217 + ], + "angle": 0, + "content": "Zicheng Liu" + }, + { + "type": "text", + "bbox": [ + 0.566, + 0.218, + 0.654, + 0.233 + ], + "angle": 0, + "content": "2Microsoft" + }, + { + "type": "text", + "bbox": [ + 0.339, + 0.237, + 0.626, + 0.251 + ], + "angle": 0, + "content": "https://linhuang17.github.io/NVF" + }, + { + "type": "image", + "bbox": [ + 0.078, + 0.271, + 0.895, + 0.367 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.369, + 0.895, + 0.44 + ], + "angle": 0, + "content": "Figure 1. We present Neural Voting Field (NVF), a 3D implicit function for camera-space 3D hand pose estimation from an RGB image. Given an RGB input (a), for a 3D point in camera space, NVF can predict its signed distance to hand surface and a set of 4D offset vectors (1D voting weight and 3D directional vector to each joint). 3D pose is then obtained via dense 3D point-wise voting. Rendered predictions: (b) Hand mesh generated by Marching Cubes from the predicted signed distances (c-h) Predicted 1D voting weight from each predicted near-surface 3D point to a joint. (Brighter line means larger weight, darker or no line means smaller weight) (i) Estimated 3D hand pose." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.451, + 0.314, + 0.467 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.484, + 0.473, + 0.876 + ], + "angle": 0, + "content": "We present a unified framework for camera-space 3D hand pose estimation from a single RGB image based on 3D implicit representation. As opposed to recent works, most of which first adopt holistic or pixel-level dense regression to obtain relative 3D hand pose and then follow with complex second-stage operations for 3D global root or scale recovery, we propose a novel unified 3D dense regression scheme to estimate camera-space 3D hand pose via dense 3D pointwise voting in camera frustum. Through direct dense modeling in 3D domain inspired by Pixel-aligned Implicit Functions for 3D detailed reconstruction, our proposed Neural Voting Field (NVF) fully models 3D dense local evidence and hand global geometry, helping to alleviate common 2D-to-3D ambiguities. Specifically, for a 3D query point in camera frustum and its pixel-aligned image feature, NVF, represented by a Multi-Layer Perceptron, regresses: (i) its signed distance to the hand surface; (ii) a set of 4D offset vectors (1D voting weight and 3D directional vector to each hand joint). Following a vote-casting scheme, 4D offset vectors from near-surface points are selected to calculate the 3D hand joint coordinates by a weighted average. Experiments demonstrate that NVF outperforms existing state-of-the-art algorithms on FreiHAND dataset for camera-space 3D hand pose estimation. We also adapt NVF to the classic task of root-relative 3D hand pose estimation, for which NVF also obtains state-of-the-art results on HO3D dataset." + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.451, + 0.632, + 0.467 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.477, + 0.892, + 0.583 + ], + "angle": 0, + "content": "Monocular 3D hand pose estimation, which aims to recover 3D locations of hand joints from an RGB image, has attracted enormous attention and made remarkable progress in recent years. As a long-standing task in computer vision, it remains challenging due to its highly articulated structure, large variations in orientations, severe (self-)occlusion, and inherent 2D-to-3D scale and depth ambiguity." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.893, + 0.765 + ], + "angle": 0, + "content": "Owing to the aforementioned difficulties, most existing works [4-7, 15, 19, 26, 32, 35, 36, 45, 52, 53, 59, 60, 63] focused on one aspect of this general problem, which is to estimate root-relative 3D hand pose (i.e., 3D joint coordinates relative to a pre-defined root joint, such as hand wrist). While accurate 2D-to-3D root-relative pose estimation is essential for numerous applications in Virtual/Augmented Reality, there are various interactive tasks in which having root-relative hand joint coordinates alone is insufficient. For instance, being able to recover camera-space 3D hand joint coordinates in an AR view enables the user to directly use hands to manipulate virtual objects moving in 3D space." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.895, + 0.903 + ], + "angle": 0, + "content": "To recover robust camera-space 3D hand pose, there are two key design elements: (1) the ability to exploit dense local evidence. Specifically, as demonstrated in previous works [14, 16, 25, 29, 34, 42, 43, 47, 49, 54-56], dense regression-based methods are more effective than holistic regression-based counterparts for handling highly articulated 3D pose structure, attributed to its ability to maintain the input data spatial structure and fully exploit local evidence; (2) the ability to reason 3D hand global geome" + }, + { + "type": "page_footnote", + "bbox": [ + 0.101, + 0.887, + 0.416, + 0.9 + ], + "angle": 0, + "content": "*Work done during Lin Huang's internship with Microsoft." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8969" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.079, + 0.089, + 0.477, + 0.179 + ], + "angle": 0, + "content": "
MethodFirst StageSecond Stage
Iqbal et al. [29]2D-DenseScale Estimation
ObMan [22]HolisticRoot Estimation
I2L-MeshNet [42]1D-DenseRoot Depth Estimation
CMR [8]2D-Dense+SpiralConvRegistration
Hasson et al. [21]HolisticModel Fitting
NVF (Ours)Unified 3D-DenseWeighted Average
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.185, + 0.47, + 0.213 + ], + "angle": 0, + "content": "Table 1. Comparison of representative absolute 3D hand pose estimation schemes. Please refer to Sec. 2.1 for more details." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.222, + 0.469, + 0.297 + ], + "angle": 0, + "content": "try. As shown in previous literature [13, 29, 33], given 2D evidence and camera intrinsic parameters, reasonable understanding towards target object 3D structure/geometry is crucial to alleviate 2D-to-3D depth ambiguity, which is the key to accurately locate 3D hand pose in camera space." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.298, + 0.47, + 0.644 + ], + "angle": 0, + "content": "To fully integrate both elements into our algorithm design in a unified manner, we connect with Pixel-aligned Implicit Function (PIFu) [24, 27, 49, 50, 62]. Through direct dense modeling in 3D domain with pixel-aligned local features, PIFu-based methods reconstruct highly detailed 3D human geometry from an RGB image in a unified way, showing its ability to model high-frequency local details such as clothing wrinkles while generating complete global geometry including largely occluded region such as the back of a person. Inspired by these results, we propose a novel unified 3D dense regression scheme based on a 3D implicit function for robust camera-space 3D hand pose estimation. Specifically, for each of the 3D query points densely sampled in camera frustum and its pixel-aligned image feature, unlike PIFu predicting occupancy value for each point, our proposed Neural Voting Field (NVF) regresses: (i) the signed distance between the point and the hand surface; (ii) a set of 4D offset vectors (1D voting weight and 3D directional vector from the point to each joint). Following a vote-casting scheme, 4D offset vectors from near-surface points (i.e., points for which the predicted signed distance is below a threshold) are selected to calculate the 3D hand joint coordinates by a weighted average." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.646, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Most existing works for camera-space 3D hand pose estimation, as shown in Tab. 1, follow a two-stage estimation scheme. They first adopt holistic or pixel-level dense regression to obtain 2D and relative 3D hand poses and then follow with complex second-stage processing such as fitting, registration, using a separate network for 3D global root location or scale estimation. NVF instead provides a unified solution via direct dense modeling in 3D camera space followed by a simple weighted average operation, which enables reasoning about 3D dense local evidence and hand global geometry. As shown in Fig. 1, NVF makes solid 3D point-wise prediction and overall distribution of signed distance and voting weight even in highly occluded regions, leading to accurate camera-space pose estimation. In Sec. 4, we show that NVF noticeably outperforms two baselines based on holistic regression and 2D dense regression. Besides, NVF exhibits state-of-the-art performance" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.151 + ], + "angle": 0, + "content": "for the task of camera-space 3D hand pose estimation on FreiHAND dataset. We also adapt NVF to the classic task of root-relative 3D hand pose estimation, for which NVF also achieves state-of-the-art results on HO3D dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.152, + 0.892, + 0.212 + ], + "angle": 0, + "content": "Since estimating absolute 3D pose from an RGB image is an ill-posed problem due to scale and depth ambiguity [13,29], in Sec. 4.4, we also provide ablation analysis on hand scale based on results from NVF and the baselines." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.213, + 0.822, + 0.227 + ], + "angle": 0, + "content": "This work makes the following contributions:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.23, + 0.892, + 0.275 + ], + "angle": 0, + "content": "1. We propose Neural Voting Field (NVF), as the first 3D implicit representation-based unified solution to estimate camera-space 3D hand pose." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.279, + 0.892, + 0.324 + ], + "angle": 0, + "content": "2. NVF follows a novel unified 3D dense regression scheme to estimate camera-space 3D hand pose via dense 3D point-wise voting in camera frustum." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.328, + 0.892, + 0.373 + ], + "angle": 0, + "content": "3. NVF outperforms baseline methods based on holistic and 2D dense regression and achieves state-of-the-art results on absolute and relative hand pose estimation." + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.23, + 0.892, + 0.373 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.384, + 0.642, + 0.399 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.408, + 0.851, + 0.424 + ], + "angle": 0, + "content": "2.1. Camera-Space 3D Hand Pose Estimation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.431, + 0.892, + 0.734 + ], + "angle": 0, + "content": "As shown in Tab. 1, most existing works for monocular RGB input-based camera-space 3D hand pose estimation generally follow a two-stage estimation scheme. Iqbal et al. [29] recovers 3D hand pose in camera space up to a scaling factor by solving a quadratic equation given the camera intrinsic parameters and the estimated scale-normalized and root-relative 2.5D pose via 2D Dense regression. Extra global scale estimation is required for camera-space absolute 3D hand pose recovery. Another line of work [8,21,22] first obtains 2D pose or root-relative 3D hand pose estimations via 2D dense or holistic regression. Then, Hasson et al. [21] fits a 3D hand model to the initial estimations and recovers 3D hand pose in camera space from the input RGB video. Both ObMan [22] and CMR [8] work on a single RGB input and directly recover 3D global root from the initial 2D pose and root-relative 3D pose/mesh estimations. I2L-MeshNet [42] proposes a 1D dense regression scheme to predict per-lixel likelihood on 1D heat map for recovering root-relative 2.5D pose, and then uses a separate network (i.e., RootNet [41]) for the absolute root depth estimation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.742, + 0.838, + 0.757 + ], + "angle": 0, + "content": "2.2. Root-relative 3D Hand Pose Estimation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.903 + ], + "angle": 0, + "content": "For single RGB input-based root-relative 3D hand pose estimation, some early methods [52, 59, 63] adopt classic frameworks such as Convolutional Neural Network and Multi-Layer Perceptron in combination with holistic or dense regression scheme for scale-invariant and rootrelative 3D hand joint positions recovery. After the introduction of the 3D parametric MANO [48] model, tremendous literature [3, 4, 9, 22, 45, 53, 60] attempts to integrate MANO into the learning pipeline and directly regress" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8970" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.087, + 0.466, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.456, + 0.099, + 0.856, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.291, + 0.892, + 0.349 + ], + "angle": 0, + "content": "Figure 2. Overview of Neural Voting Field (NVF): We first extract a feature map \\( F \\) from the input RGB image \\( I \\) by an image encoder. Then, for a 3D query point \\( \\pmb{p} \\) sampled in camera space, we define an implicit function \\( f_{\\mathrm{NVF}} \\) realized by a MLP, which maps from pixel-aligned image feature \\( S(F, \\pi(\\pmb{p})) \\) and depth \\( Z(\\pmb{p}) \\) of \\( \\pmb{p} \\) to its signed distance \\( s \\) and a set of 4D offset vectors \\( V \\). Each 4D offset vector \\( v \\) contains a 1D voting weight and a 3D directional vector to each joint. The 3D hand pose is then obtained by dense 3D point-wise voting." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.357, + 0.47, + 0.465 + ], + "angle": 0, + "content": "MANO parameters from the image. Meanwhile, MANO model-free approaches [6] continue to develop and more works also start to investigate the dependencies among hand joints as hands are inherently structured. Thus, geometric deep learning-based models (e.g., Graph Convolution, Spiral Convolution) [5, 7, 8, 15, 32, 53] and Transformer-based models [19, 35, 36, 45] have also been proposed." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.47, + 0.322, + 0.486 + ], + "angle": 0, + "content": "2.3. 3D Implicit Representation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.493, + 0.471, + 0.796 + ], + "angle": 0, + "content": "Recent methods [1, 2, 10, 17, 37, 40, 46, 51] have shown that, as a continuous and differentiable function realized by Multi-Layer Perceptron, memory efficient deep implicit function can represent detailed 3D shapes with arbitrary topology and have no resolution limitations. It can also be extended to RGB image-based 3D reconstruction [40, 44, 46, 57]. Grasping Field [30] proposes to reconstruct hands grasping object via predicting signed distances for hand and object given the input point cloud or RGB image. Unlike previous implicit functions relying on a global feature vector, given an input RGB image with cleanly segmented foreground, Pixel-aligned Implicit Function [49] (PIFu) utilizes pixel-aligned image feature and depth value of each 3D point for occupancy prediction. PIFu-based methods have shown the ability to generate high-fidelity 3D human geometry with more clothes details [27, 31, 50, 62] and robust 6D pose estimation. For instance, Neural Correspondence Field [24] predicts signed distance and 3D object coordinate for each 3D point to establish dense 3D-3D correspondences for robust 6D object pose estimation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.807, + 0.289, + 0.824 + ], + "angle": 0, + "content": "3. The Proposed Method" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.831, + 0.218, + 0.846 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.855, + 0.469, + 0.903 + ], + "angle": 0, + "content": "Pixel-Aligned Implicit Function (PIFu) [49]. PIFu, by learning an implicit function in 3D domain with pixel-aligned features, generates detailed 3D human geometry in" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.357, + 0.892, + 0.539 + ], + "angle": 0, + "content": "cluding largely occluded regions from a segmented RGB image. Specifically, given a segmented image \\(I:\\mathbb{R}^2\\mapsto \\mathbb{R}^3\\) PIFu first extracts a feature map \\(F:\\mathbb{R}^2\\mapsto \\mathbb{R}^C\\) with \\(C\\) channels by a fully convolutional image encoder. Then, for a 3D point \\(\\pmb {p}\\in \\mathbb{R}^3\\) in the camera space and its 2D projection \\(\\pi (\\pmb {p}):\\mathbb{R}^3\\mapsto \\mathbb{R}^2\\) on \\(I\\) , the implicit function, represented by a Multi-Layer Perceptron (MLP) \\(f_{\\mathrm{PIF_u}}:\\mathbb{R}^C\\times \\mathbb{R}\\mapsto\\) \\(\\mathbb{R}\\) as \\(f_{\\mathrm{PIF_u}}(S(F,\\pi (\\pmb {p})),Z(\\pmb {p})) = y\\) maps the pixel-aligned feature \\(S(F,\\pi (\\pmb {p}))\\) and the depth \\(Z(\\pmb {p})\\) of \\(\\pmb{p}\\) to the occupancy \\(y\\in [0,1]\\) (1 means \\(\\pmb{p}\\) is inside the human surface and 0 means otherwise). \\(S\\) is the sampling function (e.g., bilinear interpolation) to sample the value of \\(F\\) at pixel \\(\\pi (\\pmb {p})\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.547, + 0.892, + 0.654 + ], + "angle": 0, + "content": "Signed Distance Function (SDF) [12, 46]. Given a 3D shape \\(M \\in \\mathcal{M}\\), a signed distance function is a continuous function \\(\\phi: \\mathbb{R}^3 \\times \\mathcal{M} \\mapsto \\mathbb{R}\\) as \\(\\phi(\\pmb{p}, M) = s\\), mapping a 3D point \\(\\pmb{p}\\) to the signed distance \\(s\\) between \\(\\pmb{p}\\) and the shape surface. The sign denotes if \\(\\pmb{p}\\) is outside (positive) or inside (negative) of a watertight surface, with the iso-surface \\(\\phi(\\pmb{p}, M) = 0\\) implicitly representing the surface." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.664, + 0.887, + 0.68 + ], + "angle": 0, + "content": "3.2. Dense Offset-based Pose Re-Parameterization" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.892, + 0.901 + ], + "angle": 0, + "content": "As opposed to holistic regression directly mapping the input data to sparse and absolute joint positions and 2D dense regression which mainly exploits the 2D input local evidence, our method aims to directly model in 3D domain and thus fully exploit the 3D dense local evidence and hand global structure. To achieve this goal, given a common sparse and explicit pose representation (i.e., a set of 3D coordinates of \\( T \\) hand joints) as \\( J = \\{j_{t}\\}_{t=1}^{T} \\in \\mathcal{J} \\), \\( j_{t} \\in \\mathbb{R}^{3} \\), we parameterize it as a set of dense 3D point-wise offset-based representation. As shown in previous works [16,54], it is non-ideal to directly learn the 3D offset vector from each 3D point to a joint due to the large variance of offsets and far away joints are also beyond the scope of receptive field. Thus, given a camera-space 3D hand pose \\( J \\) and" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "8971" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.138 + ], + "angle": 0, + "content": "shape \\(M\\), we define the pose parameterization as a function \\(\\psi\\), mapping from a 3D point \\(\\pmb{p}\\) in the camera space to a set of 4D offset vectors \\(V = \\{\\pmb{v}_t\\}_{t=1}^T\\), \\(\\pmb{v}_t \\in \\mathbb{R}^4\\):" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.147, + 0.469, + 0.164 + ], + "angle": 0, + "content": "\\[\n\\psi : \\mathbb {R} ^ {3} \\times \\mathcal {J} \\times \\mathcal {M} \\mapsto \\mathbb {R} ^ {T \\times 4} \\text {a s} \\psi (\\boldsymbol {p}, J, M) = V. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.175, + 0.469, + 0.207 + ], + "angle": 0, + "content": "Each offset vector \\(\\pmb{v}_t = (w_t, \\pmb{d}_t)\\) consists of a 1D voting weight \\(w_t \\in \\mathbb{R}\\) as point-wise closeness to hand joint \\(j_t\\):" + }, + { + "type": "equation", + "bbox": [ + 0.078, + 0.215, + 0.498, + 0.263 + ], + "angle": 0, + "content": "\\[\nw _ {t} = \\left\\{ \\begin{array}{c c} 1 - \\frac {\\| \\boldsymbol {j} _ {t} - \\boldsymbol {p} \\| _ {2}}{r} & | s | < \\delta \\text {a n d} \\| \\boldsymbol {j} _ {t} - \\boldsymbol {p} \\| _ {2} \\leq r \\text {a n d} \\boldsymbol {p} \\in B _ {t} ^ {K}, \\\\ 0 & \\text {o t h e r w i s e}; \\end{array} \\right. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.264, + 0.469, + 0.295 + ], + "angle": 0, + "content": "and a 3D unit directional vector \\( \\pmb{d}_t \\in \\mathbb{R}^3 \\) as point-wise direction to hand joint \\( j_t \\):" + }, + { + "type": "equation", + "bbox": [ + 0.077, + 0.304, + 0.469, + 0.356 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {d} _ {t} = \\left\\{ \\begin{array}{c c} \\frac {\\boldsymbol {j} _ {t} - \\boldsymbol {p}}{\\| \\boldsymbol {j} _ {t} - \\boldsymbol {p} \\| _ {2}} & | s | < \\delta \\text {a n d} \\| \\boldsymbol {j} _ {t} - \\boldsymbol {p} \\| _ {2} \\leq r \\text {a n d} \\boldsymbol {p} \\in B _ {t} ^ {K}, \\\\ 0 & \\text {o t h e r w i s e}; \\end{array} \\right. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.357, + 0.469, + 0.479 + ], + "angle": 0, + "content": "where \\( s = \\phi(\\pmb{p}, M) \\) and \\( |s| < \\delta \\) means points that are within clamping distance \\( \\delta \\) from the hand surface. \\( r \\) denotes the radius of a 3D ball centered at each joint and \\( B_t^K \\) denotes the set of \\( K \\)-Nearest Neighbors (KNN) to joint \\( \\pmb{j}_t \\) among all the near-surface points (i.e., points with \\( |s| < \\delta \\)). Thus, among all the 3D points that are in the hand surface vicinity, we assign the non-zero 4D offset vector \\( \\pmb{v}_t \\) to the \\( K \\) nearest points inside the 3D ball with radius \\( r \\) of joint \\( \\pmb{j}_t \\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.486, + 0.265, + 0.502 + ], + "angle": 0, + "content": "3.3. Neural Voting Field" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.51, + 0.469, + 0.691 + ], + "angle": 0, + "content": "To fully exploit the 3D local evidence and hand global geometry for camera-space 3D hand pose estimation, based on the aforementioned signed distance function and pose parameterization, we unify both modalities using a pixel-aligned implicit function. Following the proposed pipeline in Fig. 2, we first extract a \\(C\\)-channel feature map \\(F\\) from the input RGB image \\(I\\) by a hourglass network \\(g(I; \\boldsymbol{\\eta})\\) as used in [24, 49]. Then, for a 3D query point \\(\\pmb{p}\\) sampled in camera space, we define a continuous implicit function \\(f_{\\mathrm{NVF}}\\) realized by a MLP, which maps from pixel-aligned image feature \\(S(F, \\pi(\\pmb{p}))\\) and depth \\(Z(\\pmb{p})\\) of point \\(\\pmb{p}\\) to its signed distance \\(s\\) and a set of 4D offset vectors \\(V\\):" + }, + { + "type": "equation", + "bbox": [ + 0.146, + 0.701, + 0.469, + 0.724 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} f _ {\\mathrm {N V F}}: \\mathbb {R} ^ {C} \\times \\mathbb {R} \\mapsto \\mathbb {R} \\times \\mathbb {R} ^ {T \\times 4} \\quad \\text {a s} \\\\ f _ {\\mathrm {s}}: (C (E _ {\\mathrm {s}} (\\cdot)), Z (\\cdot), \\boldsymbol {a}) \\quad (\\cdot , V) \\end{array} \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.148, + 0.718, + 0.4, + 0.734 + ], + "angle": 0, + "content": "\\[\nf _ {\\mathrm {N V F}} \\big (S (F, \\pi (\\boldsymbol {p})), Z (\\boldsymbol {p}); \\boldsymbol {\\theta} \\big) = (s, V),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.744, + 0.469, + 0.865 + ], + "angle": 0, + "content": "where \\(\\eta\\) and \\(\\theta\\) are the parameters of the hourglass \\(g\\) and MLP \\(f_{\\mathrm{NVF}}\\), respectively. We use a pinhole camera model for 3D-to-2D projection. For each 3D point, besides the set of 4D offset vectors, NVF also predicts its signed distance to the hand surface, helping to learn the global hand geometry and also to select near-surface 3D points by thresholding the signed distances. The near-surface 3D points are then considered as valid 3D voters to be used in the voting stage." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.903 + ], + "angle": 0, + "content": "Optimization. During training, we sample \\(N\\) 3D points \\(P = \\{\\pmb{p}_n\\}_{n=1}^N\\) in camera frustum for each image and apply" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.167 + ], + "angle": 0, + "content": "direct supervision to both predicted signed distance \\( s_n \\) as \\( L_s \\) and the predicted set of 4D offset vectors \\( V_n \\) as \\( L_V \\). Given the ground-truth signed distance \\( \\hat{s}_n = \\phi(\\pmb{p}_n, \\hat{M}) \\) and set of offset vectors \\( \\hat{V}_n = \\psi(\\pmb{p}_n, \\hat{J}, \\hat{M}) \\), where \\( \\hat{J} \\) and \\( \\hat{M} \\) are the ground-truth pose and mesh, \\( L_s \\) and \\( L_V \\) are defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.545, + 0.168, + 0.892, + 0.209 + ], + "angle": 0, + "content": "\\[\nL _ {s} = \\frac {1}{N} \\sum_ {n = 1} ^ {N} | \\operatorname {c l a m p} (\\hat {s} _ {n}, \\delta) - \\operatorname {c l a m p} (s _ {n}, \\delta) |, \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.542, + 0.212, + 0.892, + 0.253 + ], + "angle": 0, + "content": "\\[\nL _ {V} = \\frac {1}{N} \\sum_ {n = 1} ^ {N} \\mathbb {1} \\left(\\left| \\hat {s} _ {n} \\right| < \\delta\\right) H \\left(\\hat {V} _ {n}, V _ {n}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.258, + 0.892, + 0.378 + ], + "angle": 0, + "content": "where \\(\\mathrm{clamp}(s,\\delta)\\coloneqq \\min (\\delta ,\\max (-\\delta ,s))\\) uses \\(\\delta\\) to control the distance from the hand surface over which we want to main a metric SDF, following [46]. \\(H\\) is the Huber loss [28] and \\(\\mathbb{1}(\\cdot)\\) is an indicator function that selects near-surface points (i.e., points that are within clamping distance \\(\\delta\\) from the hand surface). Then, We jointly train the hourglass network \\(g\\) and MLP \\(f_{\\mathrm{NVF}}\\) end-to-end by solving the following optimization problem with \\(\\lambda\\) as the balancing weight:" + }, + { + "type": "equation", + "bbox": [ + 0.595, + 0.38, + 0.891, + 0.405 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\eta} ^ {\\star}, \\boldsymbol {\\theta} ^ {\\star} = \\underset {\\boldsymbol {\\eta}, \\boldsymbol {\\theta}} {\\arg \\min } L _ {s} + \\lambda L _ {V}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.411, + 0.892, + 0.653 + ], + "angle": 0, + "content": "Sampling of 3D Query Points. During training, unlike voxel-based methods which require discretization of ground-truth 3D meshes, we directly sample 3D points on the fly from the ground-truth mesh following [24, 49]. To make NVF fully exploit meaningful local evidence and understand the hand global geometry, it is crucial to focus on the hand by sampling the query points more densely around the hand surface during training. Thus, given the 3D camera-space hand model, we first sample 12500 points nearby the surface, 1000 points inside the bounding sphere of the model, and 1000 points inside the camera frustum. Among all these points, we then sample 2500 points inside the model and 2500 points outside to form a set of 5000 points in total for training. During inference, the points are directly sampled at centers of voxels that fill up the same camera frustum without any ground-truth information used." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.66, + 0.779, + 0.675 + ], + "angle": 0, + "content": "3.4. Dense 3D Point-to-Joint Voting." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.682, + 0.892, + 0.789 + ], + "angle": 0, + "content": "During inference, assuming \\(N\\) 3D points \\(P = \\{\\pmb{p}_n\\}_{n=1}^N\\) sampled in camera frustum for each image, for each point \\(\\pmb{p}_n\\) that is considered as a valid 3D voter (i.e., the predicted signed distance \\(s_n\\) is below the threshold \\(\\delta\\)), based on its predicted set of 4D offset vectors \\(V_n = \\{\\pmb{v}_t^n\\}_{t=1}^T\\), we first convert each 4D offset vector \\(\\pmb{v}_t^n = (w_t^n, \\pmb{d}_t^n)\\) into the actual 3D offset \\(\\pmb{o}_t^n\\) from \\(\\pmb{p}_n\\) to joint \\(\\pmb{j}_t\\) based on Eq. 2 and Eq. 3:" + }, + { + "type": "equation", + "bbox": [ + 0.58, + 0.793, + 0.891, + 0.81 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {o} _ {t} ^ {n} = \\mathbb {1} \\left(\\left| s _ {n} \\right| < \\delta\\right) \\left[ r \\left(1 - w _ {t} ^ {n}\\right) \\boldsymbol {d} _ {t} ^ {n} \\right]. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.815, + 0.891, + 0.859 + ], + "angle": 0, + "content": "Then, each camera-space joint location \\( \\mathbf{j}_t \\) is calculated simply by a weighted average over each voter's 3D offset prediction:" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.863, + 0.891, + 0.905 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {j} _ {t} = \\sum_ {n = 1} ^ {N} \\frac {\\mathbb {1} \\left(\\left| s _ {n} \\right| < \\delta\\right) w _ {t} ^ {n} \\left(\\boldsymbol {o} _ {t} ^ {n} + \\boldsymbol {p} _ {n}\\right)}{\\sum_ {n = 1} ^ {N} \\mathbb {1} \\left(\\left| s _ {n} \\right| < \\delta\\right) w _ {t} ^ {n}}. \\tag {9}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8972" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.228 + ], + "angle": 0, + "content": "Finally, we obtain the final camera-space 3D hand pose as \\( J = \\{j_{t}\\}_{t=1}^{T} \\). Furthermore, instead of using all the near-surface points as valid voters, we can also only focus on using near-surface points with high predicted voting weight, as a higher voting weight means the point is more confident and closer to a hand joint. Therefore, we use voting fraction as a parameter to control the fraction of near-surface points with the highest voting weight for voting. Related ablation studies will be given in Sec. 4.7." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.242, + 0.212, + 0.259 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.267, + 0.25, + 0.282 + ], + "angle": 0, + "content": "4.1. Baseline Methods" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.291, + 0.473, + 0.457 + ], + "angle": 0, + "content": "To demonstrate the effectiveness of our proposed 3D dense regression scheme against classic holistic regression and 2D dense regression, we design two directly comparable baselines that share the same architecture of the hourglass network and the MLP as NVF described in Sec. 3. Specifically, similar to NVF, both baselines first extract \\(C\\)-channel feature map from the input RGB image \\(I\\). Then, unlike our \\(f_{\\mathrm{NVF}}\\) which takes a sampled 3D query point's pixel-aligned feature \\(S(F,\\pi (\\pmb {p}))\\) and its depth \\(Z(\\pmb {p})\\) to predict its signed distance \\(s\\) and the set of 4D offset vectors \\(V\\), the two baselines take the following steps." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.463, + 0.473, + 0.601 + ], + "angle": 0, + "content": "Baseline-Holistic. We directly apply a global average pooling \\(\\alpha\\) to its feature map \\(F^{\\prime}\\) extracted by the hourglass network \\(g(I;\\pmb {\\eta}^{\\prime})\\) and generate a \\(C\\) -dimension feature vector \\(\\alpha (F^{\\prime})\\) . The feature vector is then passed through the MLP \\(f_{\\mathrm{HOL}}:\\mathbb{R}^{C}\\mapsto \\mathbb{R}^{T\\times 3}\\) as \\(f_{\\mathrm{HOL}}(\\alpha (F^{\\prime});\\pmb {\\theta}^{\\prime}) = J^{\\prime}\\) \\(f_{\\mathrm{HOL}}\\) directly outputs \\(T\\) 3D hand joint coordinates \\(J^{\\prime} = \\{\\pmb{j}_{t}^{\\prime}\\}_{t = 1}^{T}\\) in camera space. Given the ground-truth pose \\(\\hat{J}\\) , it is trained by solving the following optimization problem with a loss function \\(L_{J^{\\prime}} = H(\\hat{J},J^{\\prime})\\) .." + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.606, + 0.47, + 0.634 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\eta} ^ {\\prime \\star}, \\boldsymbol {\\theta} ^ {\\prime \\star} = \\underset {\\boldsymbol {\\eta} ^ {\\prime}, \\boldsymbol {\\theta} ^ {\\prime}} {\\arg \\min } L _ {J ^ {\\prime}}. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.64, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Baseline-2D-Dense. We adopt a 2D dense regression scheme to estimate the camera-space pose via dense 2D pixel-wise voting, which is different from NVF which is based on dense 3D point-wise voting. Given the feature map \\( F'' \\) extracted by the hourglass network \\( g(I; \\boldsymbol{\\eta}^{\\prime \\prime}) \\), we only pass the pixel-aligned feature \\( F''(\\boldsymbol{u}) \\) at each 2D pixel location \\( \\boldsymbol{u} \\) through the MLP \\( f_{\\mathrm{DEN}}: \\mathbb{R}^C \\mapsto \\mathbb{R} \\times \\mathbb{R}^{T \\times 4} \\) as \\( f_{\\mathrm{DEN}}(F''(\\boldsymbol{u}); \\boldsymbol{\\theta}'') = (e, V'') \\). \\( f_{\\mathrm{DEN}} \\) outputs the probability \\( e \\in [0,1] \\) that the hand is present at \\( \\boldsymbol{u} \\) and also a set of 4D vectors \\( V'' = \\{\\boldsymbol{v}_t''\\}_{t=1}^T \\). Each 4D vector \\( \\boldsymbol{v}_t'' = (w_t'', j_t'') \\) contains a 1D voting weight \\( w_t'' \\) and 3D joint coordinate \\( j_t'' \\). For the 1D voting weight, it is defined similarly to the one in NVF. We first use the camera ray for each pixel to find its first 3D intersection point with the 3D hand model in the camera space. Then, we directly follow Eq. 2 to calculate the pixel's voting weight as the closeness to each joint. It is trained by solving the following optimization" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.09, + 0.808, + 0.108 + ], + "angle": 0, + "content": "tion problem with \\(\\lambda''\\) as the balancing weight:" + }, + { + "type": "equation", + "bbox": [ + 0.506, + 0.113, + 0.892, + 0.207 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\boldsymbol {\\eta} ^ {\\prime \\prime \\star}, \\boldsymbol {\\theta} ^ {\\prime \\prime \\star} = \\underset {\\boldsymbol {\\eta} ^ {\\prime \\prime}, \\boldsymbol {\\theta} ^ {\\prime \\prime}} {\\arg \\min } L _ {e} + \\lambda^ {\\prime \\prime} L _ {V ^ {\\prime \\prime}} (11) \\\\ = \\underset {\\boldsymbol {\\eta} ^ {\\prime \\prime}, \\boldsymbol {\\theta} ^ {\\prime \\prime}} {\\arg \\min } \\frac {1}{L} \\left[ \\sum_ {l = 1} ^ {L} E \\left(\\hat {e} _ {l}, e _ {l}\\right) + \\lambda^ {\\prime \\prime} \\sum_ {l = 1} ^ {L} \\hat {e} _ {l} H \\left(\\hat {V} _ {l} ^ {\\prime \\prime}, V _ {l} ^ {\\prime \\prime}\\right) \\right], (12) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.215, + 0.895, + 0.325 + ], + "angle": 0, + "content": "where \\(L\\) is the number of pixels, \\(E\\) is the binary cross-entropy loss, \\(\\hat{e}_l\\) is given by the ground-truth hand mask, and \\(\\hat{V}_l^{\\prime \\prime}\\) is the ground-truth set of 4D vectors. During inference, for pixel \\(\\pmb{u}_l\\) with predicted probability \\(e_l > 0.5\\) (i.e., predicted as the foreground pixel), its predicted voting weight and 3D joint location are used to calculate the camera-space pose \\(J^{\\prime \\prime} = \\{\\dot{j}_t^{\\prime \\prime}\\}_{t = 1}^{T}\\) by a weighted average, similar to Eq. 9." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.33, + 0.78, + 0.345 + ], + "angle": 0, + "content": "4.2. Dataset and Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.354, + 0.892, + 0.445 + ], + "angle": 0, + "content": "FreiHAND Dataset [64]. As a real 3D hand dataset, it provides 130,240 training and 3,960 evaluation samples. The result of the evaluation set is evaluated via the official server for online evaluation. Since FreiHAND provides hand scale during evaluation time, we provide an ablation on hand scale for analysis of 2D-to-3D scale and depth ambiguity." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.446, + 0.893, + 0.627 + ], + "angle": 0, + "content": "Metrics. For FreiHAND, we evaluate for the task of camera-space 3D hand pose estimation. As in [8,64], we report the unaligned version of Mean Joint Error (mm) which is measured in camera space as CS-MJE and the Area-Under-the-Curve of PCK (Percentage of Correct Keypoints) vs. error thresholds as CS-AUC to evaluate the absolute 3D hand pose. We also define and report a 3D Translational Error (mm) (TE) [23] and Depth Error (mm) (DE) to evaluate the predicted and ground-truth pose's centroid and the depth-axis of the centroid, respectively. The results of TE and DE will be used in analysis for baseline studies, especially for the analysis of scale and depth ambiguity." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.628, + 0.892, + 0.687 + ], + "angle": 0, + "content": "HO3D Dataset [18]. As a real 3D hand-object dataset, it contains 66,034 training samples and 11,524 evaluation samples. We use the official server for online evaluation as the annotations of the evaluation set are not available." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.688, + 0.892, + 0.794 + ], + "angle": 0, + "content": "Metrics. For HO3D, we evaluate for the task of rootrelative 3D hand pose estimation. As in [18, 19], we report the Mean Joint Error (mm) after scale-translation alignment of the root as MJE and corresponding AUC to evaluate the pose. We also follow [45] to add ground-truth root to the predicted root-relative pose and report error (mm) as RS-MJE as it is equivalent to measuring in root-relative space." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.84 + ], + "angle": 0, + "content": "Complement (Comp) Dataset [7]. As a synthetic 3D hand dataset, it provides 1,520 poses and 71 viewpoints for research use. We only use it for pre-training when needed." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.848, + 0.719, + 0.864 + ], + "angle": 0, + "content": "4.3. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "For NVF and two baselines, the network architecture is adopted from PIFu [49]. Specifically, we use hourglass net" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8973" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.08, + 0.089, + 0.493, + 0.322 + ], + "angle": 0, + "content": "
MethodExtra DataHand CropHand ScaleCS-MJE↓
ObMan [22]-✓✗85.2
MANO CNN [64]-✓✗71.3
I2L-MeshNet [42]-✓✗60.3
CMR-SG-RN18 [8]-✓✗49.7
CMR-SG-RN50 [8]-✓✗48.8
Baseline-Holisitc-✗✗54.5
Baseline-2D-Dense-✗✗53.2
CS-NVF (Ours)-✗✗47.2
Baseline-Holisitc-✗✓50.4
Baseline-2D-Dense-✗✓49.0
CS-NVF (Ours)-✗✓42.4
CS-NVF (Ours)Comp*✗✗44.6
CS-NVF (Ours)Comp*✗✓39.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.327, + 0.47, + 0.37 + ], + "angle": 0, + "content": "Table 2. Comparison for absolute 3D hand pose on FreiHAND.*: pre-training on Comp. Note that our CS-NVF and two baselines take the original image without hand detection or cropping." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.38, + 0.473, + 0.699 + ], + "angle": 0, + "content": "work with the output stride of 4 and output channel of 256 as the image encoder. All the MLPs \\( f_{\\mathrm{NVF}} \\), \\( f_{\\mathrm{HOL}} \\), and \\( f_{\\mathrm{DEN}} \\) have four hidden fully-connected layers with the number of neurons as (1024, 512, 256, 128). Unless stated otherwise, during training for NVF, the clamping distance \\( \\delta \\) is set to 5 mm. The KNN is set to 1024. The ball radius \\( r \\) is set to 80 mm. During inference, the sampling step size of 3D points is 16 mm in all three axes. The voting fraction is set to 50%. Unlike PIFu, no image segmentation is applied. Camera intrinsic parameters are assumed known. Base on the above setting, for camera-space 3D hand pose estimation, we directly sample 3D points in camera space, as described in Sec. 3.3. We call NVF working in camera space as CS-NVF. Unlike other methods in Tab. 2, CS-NVF and both baselines take the original image as input without hand detection or cropping. For the task of root-relative 3D hand pose estimation, we follow similar sampling strategy but sample in a hand root-relative 3D cube instead of the camera frustum. We then call NVF working in root-relative space as RS-NVF. RS-NVF takes cropped hand-centered image. Please refer to the supplementary for more details." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.701, + 0.24, + 0.716 + ], + "angle": 0, + "content": "4.4. Baseline Studies" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.719, + 0.47, + 0.81 + ], + "angle": 0, + "content": "In this section, based on the results from CS-NVF and both baselines (as in Tab. 2, Tab. 3, Fig. 3, and Fig. 4), we investigate the impact of 2D-to-3D scale and depth ambiguity for the task of absolute 3D hand pose estimation and the effectiveness of using 3D dense regression over both 2D dense regression and holistic regression." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.471, + 0.903 + ], + "angle": 0, + "content": "2D-to-3D Scale and Depth Ambiguity. Estimating absolute 3D pose from an RGB image is an ill-posed problem due to 2D-to-3D scale and depth ambiguity [13, 29]. In view of this matter, FreiHAND [64] provides hand scale during evaluation for participants to use. As most existing works do not use the provided hand scale and thus, no" + }, + { + "type": "table", + "bbox": [ + 0.523, + 0.104, + 0.895, + 0.322 + ], + "angle": 0, + "content": "
MethodExtra DataHand ScaleTE↓DE↓
Baseline-Holisitc-X50.649.1
Baseline-2D-Dense-X49.247.9
CS-NVF (Ours)-X43.642.4
Baseline-Holisitc-✓46.945.5
Baseline-2D-Dense-✓45.343.9
CS-NVF (Ours)-✓38.937.8
Baseline-HolisitcComp*X48.747.1
Baseline-2D-DenseComp*X47.946.4
CS-NVF (Ours)Comp*X41.540.4
Baseline-HolisitcComp*✓41.740.1
Baseline-2D-DenseComp*✓40.538.8
CS-NVF (Ours)Comp*✓36.535.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.326, + 0.907, + 0.356 + ], + "angle": 0, + "content": "Table 3. Comparison of 3D Translational and Depth Error for absolute 3D hand pose on FreiHAND. \\*: pre-training on Comp." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.378, + 0.892, + 0.602 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.507, + 0.607, + 0.885, + 0.622 + ], + "angle": 0, + "content": "Figure 3. 3D PCK for absolute 3D hand pose on FreiHAND." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.895, + 0.903 + ], + "angle": 0, + "content": "relevant ablation on hand scale exists, we provide ablation results on hand scale in Tab. 2 and Tab. 3 for CS-NVF and two baselines. First, based on results in Tab. 2 and Tab. 3, we observe that, for all three pipelines, the main cause of the error between the ground-truth and predicted hand joints is the global translation and especially the depth component as TE is close to CS-MJE and DE is close to TE. For instance, CS-NVF with hand scale and Comp pre-training achieves \\(39.3\\mathrm{mm}\\) CS-MJE, while TE and DE are \\(36.5\\mathrm{mm}\\) and \\(35.5\\mathrm{mm}\\), respectively. This validates the fact that depth ambiguity is indeed one of the main challenges for estimating accurate absolute 3D hand pose from a single RGB image. Moreover, for CS-NVF and the two baselines, using hand scale leads to significant improvements in terms of all three metrics. It is also worth noting that, for all three pipelines, results only using hand scale outperform results using only extra data. For instance, CS-NVF with hand scale achieves \\(42.4\\mathrm{mm}\\) CS-MJE, while CS-NVF with Comp pre-training" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8974" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.087, + 0.895, + 0.323 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.325, + 0.895, + 0.382 + ], + "angle": 0, + "content": "Figure 4. Qualitative results for absolute and relative 3D pose: (b) Predicted hand mask (Baseline) or mesh generated by Marching Cubes from the predicted signed distances (NVF). (c-h) Predicted 1D voting weight from each predicted foreground pixel (Baseline) or predicted near-surface 3D point (NVF) to a joint. (Brighter line means larger weight, darker or no line means smaller weight) (i) Estimated 3D pose. NVF makes solid point-wise prediction and overall distribution of signed distances and voting weights in highly occluded region." + }, + { + "type": "table", + "bbox": [ + 0.091, + 0.387, + 0.456, + 0.59 + ], + "angle": 0, + "content": "
MethodMJE↓AUC↑RS-MJE↓
Hasson et al. [20]36.90.369-
Pose2Mesh [11]33.30.48033.2
ObMan [22]31.80.46155.2
Liu et al. [38]31.70.46330.0
Hampali et al. [18]30.40.494-
METRO [35]28.90.504-
I2L-MeshNet [42]26.00.52926.8
Keypoint Trans. [19]25.70.553-
ArtiBoost [58]25.30.532-
Zheng et al. [61]25.10.541-
HandOccNet [45]24.00.55724.9
RS-NVF (Ours)21.80.61023.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.594, + 0.454, + 0.608 + ], + "angle": 0, + "content": "Table 4. Comparison for relative 3D hand pose on HO3D." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.621, + 0.47, + 0.682 + ], + "angle": 0, + "content": "has \\(44.6 \\mathrm{~mm}\\) CJ-MJE. These verify that a reasonable understanding towards the hand scale would help for robust depth learning and alleviate depth ambiguity, which is also aligned with existing literature [13, 29, 33]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.686, + 0.47, + 0.837 + ], + "angle": 0, + "content": "Dense Regression vs. Holistic Regression. In Tab. 2 and Fig. 3, we can observe that both CS-NVF and Baseline-2D-Dense consistently outperform Baseline-Holistic on CS-MJE and CS-AUC under different training settings. For instance, without using hand scale, Baseline-2D-Dense is 1.3 mm better than Baseline-Holistic in CS-MJE, and CS-NVF can be 7.3 mm better. Since three pipelines share the same network architecture, this supports our point that dense regression is more effective than holistic regression for handling highly articulated pose data." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.902 + ], + "angle": 0, + "content": "3D Dense Regression vs. 2D Dense Regression. Quantitatively, in Tab. 2 and Fig. 3, it can be observed that CS-NVF consistently outperforms Baseline-2D-Dense by a large margin on CS-MJE and CS-AUC under different" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.39, + 0.895, + 0.844 + ], + "angle": 0, + "content": "training settings. For instance, without using hand scale, CS-NVF can be \\(6\\mathrm{mm}\\) better on CS-MJE. Qualitatively, as indicated by the white circles in column (c-h) of Fig. 4, both methods are able to find valid voters (Baseline-2D-Dense: pixels with predicted probability \\(e > 0.5\\). CS-NVF: 3D points with predicted signed distance \\(|s| < \\delta\\)). However, CS-NVF generally provides not only more number of valid voters, but also more reasonable element-wise prediction and overall distribution of the voting weight than Baseline-2D-Dense. Specifically, for the number of voters, since CS-NVF models all the 3D points around the hand surface and Baseline-2D-Dense can only model foreground pixels, CS-NVF generally can find and use more valid voters. As for element-wise prediction and overall distribution of the voting weight, CS-NVF is generally able to predict large voting weight for points close to hand joint and small value for points far away, resulting in well-shaped symmetric distribution around the corresponding joint, even in occluded regions. Note that brighter line indicates larger voting weight, darker or no line indicates smaller voting weight. Moreover, as shown in column (b) for the mesh rendering of CS-NVF (obtained using Marching Cubes [39] from predicted signed distances), even in highly occluded region, CS-NVF provides solid signed distance distribution showing its ability to reason the global hand geometry. Overall, these demonstrate that CS-NVF, through direct dense modeling in 3D domain, can better model 3D dense local evidence, the relation between voters and 3D hand pose, and also the global hand geometry, compared with 2D dense regression-based counterpart, which leads to robust 3D hand pose estimation." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.85, + 0.833, + 0.865 + ], + "angle": 0, + "content": "4.5. Main Results: Absolute 3D Hand Pose" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.893, + 0.9 + ], + "angle": 0, + "content": "In Tab. 2 and Fig. 3, we also compare CS-NVF with current state-of-the-art methods on FreiHAND for absolute 3D" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8975" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.079, + 0.088, + 0.897, + 0.184 + ], + "angle": 0, + "content": "
Proposed MethodEvaluation MetricVoting Fraction (%)Clamping Distance (mm)KNN
25507510015102025651210242048
CS-NVF (Ours)CS-MJE↓39.939.339.540.239.339.340.040.140.339.539.339.3
CS-AUC↑0.360.370.360.360.370.370.360.360.350.360.370.37
RS-NVF (Ours)MJE↓22.221.821.922.021.921.821.821.923.622.621.821.8
AUC↑0.600.610.610.600.600.610.610.610.590.600.610.61
" + }, + { + "type": "table_caption", + "bbox": [ + 0.08, + 0.189, + 0.89, + 0.204 + ], + "angle": 0, + "content": "Table 5. Ablation study on parameter choices for absolute and relative 3D hand pose: Voting fraction, clamping distance, and KNN" + }, + { + "type": "table", + "bbox": [ + 0.088, + 0.21, + 0.457, + 0.32 + ], + "angle": 0, + "content": "
Step SizeCS-NVF (Ours)
#PointsCS-MJE↓CS-AUC↑FPS↑
41882k39.20.371
8230k39.20.378
1628k39.30.3734
248k39.80.3650
323k42.40.3456
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.325, + 0.47, + 0.354 + ], + "angle": 0, + "content": "Table 6. Ablation study on sampling step size for absolute 3D hand pose on FreiHAND. #points: number of sampled 3D points." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.363, + 0.47, + 0.529 + ], + "angle": 0, + "content": "hand pose estimation. In Tab. 2, without the use of hand scale and extra data as a fair comparison, CS-NVF surpasses previous state-of-the-art methods, achieving \\(1.6\\mathrm{mm}\\) better than CMR [8]. Note that, unlike other methods, CS-NVF takes the original image as input without hand detection or cropping. Additional hand scale information used without extra data improves the CS-MJE of CS-NVF further to \\(42.4\\mathrm{mm}\\). Additional Comp pre-training improves the CS-MJE of CS-NVF to \\(44.6\\mathrm{mm}\\) without hand scale and \\(39.3\\mathrm{mm}\\) with hand scale. Referring to Fig. 3, CS-NVF also achieves state-of-the-art performance on CS-AUC." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.536, + 0.405, + 0.551 + ], + "angle": 0, + "content": "4.6. Main Results: Relative 3D Hand Pose" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.556, + 0.47, + 0.648 + ], + "angle": 0, + "content": "In Tab. 4, we compare RS-NVF with current state-of-the-art methods on HO3D for relative 3D hand pose estimation. RS-NVF significantly outperforms all these methods in all three metrics. Especially, RS-NVF outmatches HandOccNet [45] by \\(2.2\\mathrm{mm}\\) on MJE, \\(1.7\\mathrm{mm}\\) on RS-MJE, and also has 0.610 AUC against 0.557 AUC from HandOccNet." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.654, + 0.242, + 0.669 + ], + "angle": 0, + "content": "4.7. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Sampling Step Size. During inference, 3D query points are sampled at centers of voxels that fill up the 3D space. Since the choices of different sampling step sizes can influence both accuracy and efficiency, in Tab. 6, we show the ablation on sampling step size for CS-NVF. With \\(16\\mathrm{mm}\\) as the step size, CS-NVF achieves 34 FPS and \\(39.3\\mathrm{mm}\\) on CS-MJE given 28k points. Reducing the step size to \\(4\\mathrm{mm}\\) yields 1,882k points and 1 FPS with slightly improved CS-MJE of \\(39.2\\mathrm{mm}\\). Increasing to \\(32\\mathrm{mm}\\) step size yields 3k points and improves the FPS to 56 with still competitive \\(42.4\\mathrm{mm}\\) on CS-MJE. For RS-NVF not shown in the table, 4 and 8 mm step size both yield MJE of \\(21.7\\mathrm{mm}\\) and AUC of 0.62. We can observe that, while step size affects FPS, NVF is relatively insensitive regarding the performance. In all other reported results, we use \\(16\\mathrm{mm}\\) as the default step size." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.214, + 0.895, + 0.395 + ], + "angle": 0, + "content": "Voting Fraction. Since the predicted point-wise voting weight reflects the confidence and closeness of a given point towards a certain joint, besides using all the near-surface points as valid voters, we can use a specified fraction of near-surface points with the highest voting weight for voting during inference. In Tab. 5, we observe that using all the near-surface points (100%) yields competitive results while 50% fraction leads to the overall better performance for both CS-NVF and RS-NVF. The results are aligned with our algorithm design since points far from each joint are not expected to make meaningful contributions. In all other reported results, we use 50% as the default voting fraction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.395, + 0.895, + 0.501 + ], + "angle": 0, + "content": "Clamping Distance. During training, clamping distance \\(\\delta\\) is used to concentrate NVF's capacity on learning high-quality point-wise prediction (i.e., signed distance and set of 4D offset vectors) near the hand surface. Ablation in Tab. 5 shows that \\(5\\mathrm{mm}\\) generates the overall better performance for both CS-NVF and RS-NVF. In all other reported results, we use \\(5\\mathrm{mm}\\) as the default clamping distance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.501, + 0.895, + 0.668 + ], + "angle": 0, + "content": "\\(K\\)-Nearest Neighbors (KNN). During training, among all the sampled points that are in the hand surface vicinity, nonzero 4D offset vector will only be assigned to the \\(K\\) nearest points inside a 3D ball of each hand joint. As shown in previous works [16, 54], estimating offsets for large KNN is unnecessary and far away joints are beyond the scope of receptive field. Small KNN might also cause poor performance due to the sparse learning scheme. Ablation in Tab. 5 shows that 1024 KNN generates the overall better performance for both CS-NVF and RS-NVF. In all other reported results, we use 1024 as the default KNN." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.68, + 0.628, + 0.695 + ], + "angle": 0, + "content": "5. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.895, + 0.901 + ], + "angle": 0, + "content": "We propose Neural Voting Field (NVF), as the first 3D implicit representation-based unified solution to estimate camera-space 3D hand pose given a single RGB image. NVF follows a novel unified 3D dense regression scheme to estimate 3D hand pose in 3D camera space via dense 3D point-wise voting in camera frustum. Studies with both holistic regression baseline and 2D dense regression baseline verify NVF's ability to fully exploit 3D dense local evidence and model global hand structure/geometry, which are essential for precisely locating 3D hand joints in 3D camera space. NVF also demonstrates state-of-the-art results on both tasks of camera-space 3D hand pose estimation and root-relative 3D hand pose estimation." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8976" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.468, + 0.143 + ], + "angle": 0, + "content": "[1] Matan Atzmon and Yaron Lipman. SAL: Sign agnostic learning of shapes from raw data. In CVPR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.145, + 0.468, + 0.171 + ], + "angle": 0, + "content": "[2] Matan Atzmon and Yaron Lipman. SALD: Sign agnostic learning with derivatives. In ICLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.468, + 0.213 + ], + "angle": 0, + "content": "[3] Seungryul Baek, Kwang In Kim, and Tae-Kyun Kim. Pushing the envelope for rgb-based dense 3d hand pose estimation via neural rendering. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.215, + 0.468, + 0.255 + ], + "angle": 0, + "content": "[4] Adnane Boukhayma, Rodrigo de Bem, and Philip HS Torr. 3d hand shape and pose from images in the wild. In CVPR, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.257, + 0.468, + 0.312 + ], + "angle": 0, + "content": "[5] Yujun Cai, Liuhao Ge, Jun Liu, Jianfei Cai, Tat-Jen Cham, Junsong Yuan, and Nadia Magnenat Thalmann. Exploiting spatial-temporal relationships for 3d pose estimation via graph convolutional networks. In ICCV, 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.314, + 0.468, + 0.368 + ], + "angle": 0, + "content": "[6] Ping Chen, Yujin Chen, Dong Yang, Fangyin Wu, Qin Li, Qingpei Xia, and Yong Tan. I2uv-handnet: Image-to-uv prediction network for accurate and high-fidelity 3d hand mesh modeling. In ICCV, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.37, + 0.468, + 0.424 + ], + "angle": 0, + "content": "[7] Xingyu Chen, Yufeng Liu, Yajiao Dong, Xiong Zhang, Chongyang Ma, Yanmin Xiong, Yuan Zhang, and Xiaoyan Guo. Mobrecon: Mobile-friendly hand mesh reconstruction from monocular image. In CVPR, 2022. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.425, + 0.468, + 0.493 + ], + "angle": 0, + "content": "[8] Xingyu Chen, Yufeng Liu, Chongyang Ma, Jianlong Chang, Huayan Wang, Tian Chen, Xiaoyan Guo, Pengfei Wan, and Wen Zheng. Camera-space hand mesh recovery via semantic aggregation and adaptive 2d-1d registration. In CVPR, 2021. 2, 3, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.495, + 0.468, + 0.548 + ], + "angle": 0, + "content": "[9] Yujin Chen, Zhigang Tu, Di Kang, Linchao Bao, Ying Zhang, Xuefei Zhe, Ruizhi Chen, and Junsong Yuan. Model-based 3d hand reconstruction via self-supervised learning. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.551, + 0.468, + 0.578 + ], + "angle": 0, + "content": "[10] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In CVPR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.58, + 0.468, + 0.632 + ], + "angle": 0, + "content": "[11] Hongsuk Choi, Gyeongsik Moon, and Kyoung Mu Lee. Pose2mesh: Graph convolutional network for 3d human pose and mesh recovery from a 2d human pose. In ECCV, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.635, + 0.468, + 0.675 + ], + "angle": 0, + "content": "[12] Brian Curless and Marc Levoy. A volumetric method for building complex models from range images. SIGGRAPH, 1996. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.678, + 0.468, + 0.704 + ], + "angle": 0, + "content": "[13] Tom van Dijk and Guido de Croon. How do neural networks see depth in single images? In ICCV, 2019. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.706, + 0.468, + 0.76 + ], + "angle": 0, + "content": "[14] Linpu Fang, Xingyan Liu, Li Liu, Hang Xu, and Wenxiong Kang. Jgr-p2o: Joint graph reasoning based pixel-to-offset prediction network for 3d hand pose estimation from a single depth image. In ECCV, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.762, + 0.468, + 0.814 + ], + "angle": 0, + "content": "[15] Liuhao Ge, Zhou Ren, Yuncheng Li, Zehao Xue, Yingying Wang, Jianfei Cai, and Junsong Yuan. 3d hand shape and pose estimation from a single rgb image. In CVPR, 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.468, + 0.857 + ], + "angle": 0, + "content": "[16] Liuhao Ge, Zhou Ren, and Junsong Yuan. Point-to-point regression pointnet for 3d hand pose estimation. In ECCV, 2018. 1, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.468, + 0.9 + ], + "angle": 0, + "content": "[17] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. In ICML, 2020. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.468, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[18] Shreyas Hampali, Mahdi Rad, Markus Oberweger, and Vincent Lepetit. Honnotate: A method for 3d annotation of hand and object poses. In CVPR, 2020. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.189 + ], + "angle": 0, + "content": "[19] Shreyas Hampali, Sayan Deb Sarkar, Mahdi Rad, and Vincent Lepetit. Keypoint transformer: Solving joint identification in challenging hands and object interactions for accurate 3d pose estimation. In CVPR, 2022. 1, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.191, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[20] Yana Hasson, Bugra Tekin, Federica Bogo, Ivan Laptev, Marc Pollefeys, and Cordelia Schmid. Leveraging photometric consistency over time for sparsely supervised hand-object reconstruction. In CVPR, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.247, + 0.892, + 0.287 + ], + "angle": 0, + "content": "[21] Yana Hasson, Gül Varol, Cordelia Schmid, and Ivan Laptev. Towards unconstrained joint hand-object reconstruction from rgb videos. In 3DV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.289, + 0.892, + 0.342 + ], + "angle": 0, + "content": "[22] Yana Hasson, Gül Varol, Dimitris Tzionas, Igor Kalevatykh, Michael J. Black, Ivan Laptev, and Cordelia Schmid. Learning joint reconstruction of hands and manipulated objects. In CVPR, 2019. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.344, + 0.892, + 0.371 + ], + "angle": 0, + "content": "[23] Tomás Hodań, Jiří Matas, and Štepan Obdržálek. On evaluation of 6d object pose estimation. In ECCV, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.373, + 0.892, + 0.426 + ], + "angle": 0, + "content": "[24] Lin Huang, Tomas Hodan, Lingni Ma, Linguang Zhang, Luan Tran, Christopher Twigg, Po-Chen Wu, Junsong Yuan, Cem Keskin, and Robert Wang. Neural correspondence field for object pose estimation. In ECCV, 2022. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.428, + 0.892, + 0.468 + ], + "angle": 0, + "content": "[25] Lin Huang, Jianchao Tan, Ji Liu, and Junsong Yuan. Handtransformer: non-autoregressive structured modeling for 3d hand pose estimation. In ECCV, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.47, + 0.892, + 0.51 + ], + "angle": 0, + "content": "[26] Lin Huang, Boshen Zhang, Zhilin Guo, Yang Xiao, Zhiguo Cao, and Junsong Yuan. Survey on depth and rgb image-based 3d hand shape and pose estimation. VRIH, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.512, + 0.892, + 0.551 + ], + "angle": 0, + "content": "[27] Zeng Huang, Yuanlu Xu, Christoph Lassner, Hao Li, and Tony Tung. Arch: Animatable reconstruction of clothed humans. In CVPR, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.553, + 0.892, + 0.58 + ], + "angle": 0, + "content": "[28] Peter J Huber. Robust estimation of a location parameter. Breakthroughs in statistics, 1992. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.582, + 0.892, + 0.622 + ], + "angle": 0, + "content": "[29] Umar Iqbal, Pavlo Molchanov, Thomas Breuel Juergen Gall, and Jan Kautz. Hand pose estimation via latent 2.5d heatmap regression. In ECCV, 2018. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.624, + 0.892, + 0.678 + ], + "angle": 0, + "content": "[30] Korrawe Karunratanakul, Jinlong Yang, Yan Zhang, Michael J. Black, Krikamol Muandet, and Siyu Tang. Grasping field: Learning implicit representations for human grasps. In 3DV, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.68, + 0.892, + 0.719 + ], + "angle": 0, + "content": "[31] Nilesh Kulkarni, Justin Johnson, and David F Fouhey. What's behind the couch? directed ray distance functions (drdf) for 3d scene reconstruction. In arXiv e-prints, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.721, + 0.892, + 0.774 + ], + "angle": 0, + "content": "[32] Dominik Kulon, Riza Alp Guler, Iasonas Kokkinos, Michael M Bronstein, and Stefanos Zafeiriou. Weakly-supervised mesh-convolutional hand reconstruction in the wild. In CVPR, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.776, + 0.892, + 0.816 + ], + "angle": 0, + "content": "[33] Vincent Lepetit, Francesc Moreno-Noguer, and Pascal Fua. Epnp: An accurate o (n) solution to the pnp problem. IJCV, 2009. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.818, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[34] Shile Li and Dongheui Lee. Point-to-pose voting based hand pose estimation using residual permutation equivariant layer. In CVPR, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.86, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[35] Kevin Lin, Lijuan Wang, and Zicheng Liu. End-to-end human pose and mesh reconstruction with transformers. In CVPR, 2021. 1, 3, 7" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8977" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.12 + ], + "angle": 0, + "content": "[36] Kevin Lin, Lijuan Wang, and Zicheng Liu. Mesh graphormer. In ICCV, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.122, + 0.469, + 0.163 + ], + "angle": 0, + "content": "[37] Feng Liu, Luan Tran, and Xiaoming Liu. Fully understanding generic objects: Modeling, segmentation, and reconstruction. In CVPR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.166, + 0.469, + 0.207 + ], + "angle": 0, + "content": "[38] Shaowei Liu, Hanwen Jiang, Jiarui Xu, Sifei Liu, and Xiaolong Wang. Semi-supervised 3d hand-object poses estimation with interactions in time. In CVPR, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.21, + 0.469, + 0.25 + ], + "angle": 0, + "content": "[39] William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3d surface construction algorithm. SIGGRAPH, 1987. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.253, + 0.469, + 0.307 + ], + "angle": 0, + "content": "[40] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In CVPR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.311, + 0.469, + 0.364 + ], + "angle": 0, + "content": "[41] Gyeongsik Moon, Ju Yong Chang, and Kyoung Mu Lee. Camera distance-aware top-down approach for 3d multiperson pose estimation from a single rgb image. In ICCV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.368, + 0.469, + 0.422 + ], + "angle": 0, + "content": "[42] Gyeongsik Moon and Kyoung Mu Lee. I2l-meshnet: Image-to-lixel prediction network for accurate 3d human pose and mesh estimation from a single rgb image. In ECCV, 2020. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.426, + 0.469, + 0.465 + ], + "angle": 0, + "content": "[43] Alejandro Newell, Kaiyu Yang, and Jia Deng. Stacked hourglass networks for human pose estimation. In ECCV, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.469, + 0.469, + 0.511 + ], + "angle": 0, + "content": "[44] Michael Oechsle, Lars Mescheder, Michael Niemeyer, Thilo Strauss, and Andreas Geiger. Texture fields: Learning texture representations in function space. In ICCV, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.513, + 0.469, + 0.567 + ], + "angle": 0, + "content": "[45] JoonKyu Park, Yeonguk Oh, Gyeongsik Moon, Hongsuk Choi, and Kyoung Mu Lee. Handocctnet: Occlusion-robust 3d hand mesh estimation network. In CVPR, 2022. 1, 2, 3, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.571, + 0.469, + 0.624 + ], + "angle": 0, + "content": "[46] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In CVPR, 2019. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.628, + 0.469, + 0.669 + ], + "angle": 0, + "content": "[47] Georgios Pavlakos, Xiaowei Zhou, Konstantinos G Derpanis, and Kostas Daniilidis. Coarse-to-fine volumetric prediction for single-image 3d human pose. In CVPR, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.672, + 0.469, + 0.712 + ], + "angle": 0, + "content": "[48] Javier Romero, Dimitrios Tzionas, and Michael J Black. Embodied hands: Modeling and capturing hands and bodies together. TOG, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.716, + 0.469, + 0.769 + ], + "angle": 0, + "content": "[49] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In ICCV, 2019. 1, 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.773, + 0.469, + 0.813 + ], + "angle": 0, + "content": "[50] Shunsuke Saito, Tomas Simon, Jason Saragih, and Hanbyul Joo. Pifuhd: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization. In CVPR, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.816, + 0.469, + 0.857 + ], + "angle": 0, + "content": "[51] Vincent Sitzmann, Eric Chan, Richard Tucker, Noah Snavely, and Gordon Wetzstein. Metasdf: Meta-learning signed distance functions. In NeurIPS, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.86, + 0.469, + 0.899 + ], + "angle": 0, + "content": "[52] Adrian Spurr, Jie Song, Seonwook Park, and Otmar Hilliges. Cross-modal deep variational hand pose estimation. In CVPR, 2018. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[53] Tze Ho Elden Tse, Kwang In Kim, Ales Leonardis, and Hyung Jin Chang. Collaborative learning for hand and object reconstruction with attention-guided graph convolution. In CVPR, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[54] Chengde Wan, Thomas Probst, Luc Van Gool, and Angela Yao. Dense 3d regression for hand pose estimation. In CVPR, 2018. 1, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.193, + 0.892, + 0.22 + ], + "angle": 0, + "content": "[55] Shih-En Wei, Varun Ramakrishna, Takeo Kanade, and Yaser Sheikh. Convolutional pose machines. In CVPR, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.222, + 0.892, + 0.276 + ], + "angle": 0, + "content": "[56] Fu Xiong, Boshen Zhang, Yang Xiao, Zhiguo Cao, Taidong Yu, Joey Tianyi Zhou, and Junsong Yuan. A2j: Anchor-to-joint regression network for 3d articulated pose estimation from a single depth image. In ICCV, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.278, + 0.892, + 0.331 + ], + "angle": 0, + "content": "[57] Qiangeng Xu, Weiyue Wang, Duygu Ceylan, Radomir Mech, and Ulrich Neumann. Disn: Deep implicit surface network for high-quality single-view 3d reconstruction. In NeurIPS, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.334, + 0.892, + 0.389 + ], + "angle": 0, + "content": "[58] Lixin Yang, Kailin Li, Xinyu Zhan, Jun Lv, Wenqiang Xu, Jiefeng Li, and Cewu Lu. Artibost: Boosting articulated 3d hand-object pose estimation via online exploration and synthesis. In CVPR, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.391, + 0.892, + 0.431 + ], + "angle": 0, + "content": "[59] Linlin Yang, Shile Li, Dongheui Lee, and Angela Yao. Aligning latent spaces for 3d hand pose estimation. In ICCV, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.434, + 0.892, + 0.475 + ], + "angle": 0, + "content": "[60] Xiong Zhang, Qiang Li, Hong Mo, Wenbo Zhang, and Wen Zheng. End-to-end hand mesh recovery from a monocular rgb image. In ICCV, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.477, + 0.892, + 0.531 + ], + "angle": 0, + "content": "[61] Xiaozheng Zheng, Pengfei Ren, Haifeng Sun, Jingyu Wang, Qi Qi, and Jianxin Liao. Joint-aware regression: Rethinking regression-based method for 3d hand pose estimation. In BMVC, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.533, + 0.892, + 0.587 + ], + "angle": 0, + "content": "[62] Zerong Zheng, Tao Yu, Yebin Liu, and Qionghai Dai. Pamir: Parametric model-conditioned implicit representation for image-based human reconstruction. TPAMI, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.59, + 0.892, + 0.63 + ], + "angle": 0, + "content": "[63] Christian Zimmermann and Thomas Brox. Learning to estimate 3d hand pose from single rgb images. In ICCV, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.633, + 0.892, + 0.688 + ], + "angle": 0, + "content": "[64] Christian Zimmermann, Duygu Ceylan, Jimei Yang, Bryan Russell, Max Argus, and Thomas Brox. Freihand: A dataset for markerless capture of hand pose and shape from single rgb images. In ICCV, 2019. 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.688 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8978" + } + ] +] \ No newline at end of file diff --git a/2023/Neural Voting Field for Camera-Space 3D Hand Pose Estimation/96d05801-c2f5-4205-ac8b-2fcd00428ae6_origin.pdf b/2023/Neural Voting Field for Camera-Space 3D Hand Pose Estimation/96d05801-c2f5-4205-ac8b-2fcd00428ae6_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ee3d637aa1b00d809139c5d1529241b1aef7485f --- /dev/null +++ b/2023/Neural Voting Field for Camera-Space 3D Hand Pose Estimation/96d05801-c2f5-4205-ac8b-2fcd00428ae6_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d483290eef404d62cbf13b13b69f0ecb0b98108277830384ad01908568ae37c4 +size 1949674 diff --git a/2023/Neural Voting Field for Camera-Space 3D Hand Pose Estimation/full.md b/2023/Neural Voting Field for Camera-Space 3D Hand Pose Estimation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..cf2b32b4fcf3009837d41641aea05d89e5ad6d0c --- /dev/null +++ b/2023/Neural Voting Field for Camera-Space 3D Hand Pose Estimation/full.md @@ -0,0 +1,332 @@ +# Neural Voting Field for Camera-Space 3D Hand Pose Estimation + +Lin Huang* + +$g^{1*}$ Chung-Ching Lin + +Lijuan Wang2 Junsong Yuan1 +1University at Buffalo + +Kevin Lin² Lin Lia + +Zicheng Liu + +2Microsoft + +https://linhuang17.github.io/NVF + +![](images/201f7b5bc242e0e13edced3ca156e9b7fc97c83a6a18e05619536c61672bd131.jpg) +Figure 1. We present Neural Voting Field (NVF), a 3D implicit function for camera-space 3D hand pose estimation from an RGB image. Given an RGB input (a), for a 3D point in camera space, NVF can predict its signed distance to hand surface and a set of 4D offset vectors (1D voting weight and 3D directional vector to each joint). 3D pose is then obtained via dense 3D point-wise voting. Rendered predictions: (b) Hand mesh generated by Marching Cubes from the predicted signed distances (c-h) Predicted 1D voting weight from each predicted near-surface 3D point to a joint. (Brighter line means larger weight, darker or no line means smaller weight) (i) Estimated 3D hand pose. + +# Abstract + +We present a unified framework for camera-space 3D hand pose estimation from a single RGB image based on 3D implicit representation. As opposed to recent works, most of which first adopt holistic or pixel-level dense regression to obtain relative 3D hand pose and then follow with complex second-stage operations for 3D global root or scale recovery, we propose a novel unified 3D dense regression scheme to estimate camera-space 3D hand pose via dense 3D pointwise voting in camera frustum. Through direct dense modeling in 3D domain inspired by Pixel-aligned Implicit Functions for 3D detailed reconstruction, our proposed Neural Voting Field (NVF) fully models 3D dense local evidence and hand global geometry, helping to alleviate common 2D-to-3D ambiguities. Specifically, for a 3D query point in camera frustum and its pixel-aligned image feature, NVF, represented by a Multi-Layer Perceptron, regresses: (i) its signed distance to the hand surface; (ii) a set of 4D offset vectors (1D voting weight and 3D directional vector to each hand joint). Following a vote-casting scheme, 4D offset vectors from near-surface points are selected to calculate the 3D hand joint coordinates by a weighted average. Experiments demonstrate that NVF outperforms existing state-of-the-art algorithms on FreiHAND dataset for camera-space 3D hand pose estimation. We also adapt NVF to the classic task of root-relative 3D hand pose estimation, for which NVF also obtains state-of-the-art results on HO3D dataset. + +# 1. Introduction + +Monocular 3D hand pose estimation, which aims to recover 3D locations of hand joints from an RGB image, has attracted enormous attention and made remarkable progress in recent years. As a long-standing task in computer vision, it remains challenging due to its highly articulated structure, large variations in orientations, severe (self-)occlusion, and inherent 2D-to-3D scale and depth ambiguity. + +Owing to the aforementioned difficulties, most existing works [4-7, 15, 19, 26, 32, 35, 36, 45, 52, 53, 59, 60, 63] focused on one aspect of this general problem, which is to estimate root-relative 3D hand pose (i.e., 3D joint coordinates relative to a pre-defined root joint, such as hand wrist). While accurate 2D-to-3D root-relative pose estimation is essential for numerous applications in Virtual/Augmented Reality, there are various interactive tasks in which having root-relative hand joint coordinates alone is insufficient. For instance, being able to recover camera-space 3D hand joint coordinates in an AR view enables the user to directly use hands to manipulate virtual objects moving in 3D space. + +To recover robust camera-space 3D hand pose, there are two key design elements: (1) the ability to exploit dense local evidence. Specifically, as demonstrated in previous works [14, 16, 25, 29, 34, 42, 43, 47, 49, 54-56], dense regression-based methods are more effective than holistic regression-based counterparts for handling highly articulated 3D pose structure, attributed to its ability to maintain the input data spatial structure and fully exploit local evidence; (2) the ability to reason 3D hand global geome + +
MethodFirst StageSecond Stage
Iqbal et al. [29]2D-DenseScale Estimation
ObMan [22]HolisticRoot Estimation
I2L-MeshNet [42]1D-DenseRoot Depth Estimation
CMR [8]2D-Dense+SpiralConvRegistration
Hasson et al. [21]HolisticModel Fitting
NVF (Ours)Unified 3D-DenseWeighted Average
+ +Table 1. Comparison of representative absolute 3D hand pose estimation schemes. Please refer to Sec. 2.1 for more details. + +try. As shown in previous literature [13, 29, 33], given 2D evidence and camera intrinsic parameters, reasonable understanding towards target object 3D structure/geometry is crucial to alleviate 2D-to-3D depth ambiguity, which is the key to accurately locate 3D hand pose in camera space. + +To fully integrate both elements into our algorithm design in a unified manner, we connect with Pixel-aligned Implicit Function (PIFu) [24, 27, 49, 50, 62]. Through direct dense modeling in 3D domain with pixel-aligned local features, PIFu-based methods reconstruct highly detailed 3D human geometry from an RGB image in a unified way, showing its ability to model high-frequency local details such as clothing wrinkles while generating complete global geometry including largely occluded region such as the back of a person. Inspired by these results, we propose a novel unified 3D dense regression scheme based on a 3D implicit function for robust camera-space 3D hand pose estimation. Specifically, for each of the 3D query points densely sampled in camera frustum and its pixel-aligned image feature, unlike PIFu predicting occupancy value for each point, our proposed Neural Voting Field (NVF) regresses: (i) the signed distance between the point and the hand surface; (ii) a set of 4D offset vectors (1D voting weight and 3D directional vector from the point to each joint). Following a vote-casting scheme, 4D offset vectors from near-surface points (i.e., points for which the predicted signed distance is below a threshold) are selected to calculate the 3D hand joint coordinates by a weighted average. + +Most existing works for camera-space 3D hand pose estimation, as shown in Tab. 1, follow a two-stage estimation scheme. They first adopt holistic or pixel-level dense regression to obtain 2D and relative 3D hand poses and then follow with complex second-stage processing such as fitting, registration, using a separate network for 3D global root location or scale estimation. NVF instead provides a unified solution via direct dense modeling in 3D camera space followed by a simple weighted average operation, which enables reasoning about 3D dense local evidence and hand global geometry. As shown in Fig. 1, NVF makes solid 3D point-wise prediction and overall distribution of signed distance and voting weight even in highly occluded regions, leading to accurate camera-space pose estimation. In Sec. 4, we show that NVF noticeably outperforms two baselines based on holistic regression and 2D dense regression. Besides, NVF exhibits state-of-the-art performance + +for the task of camera-space 3D hand pose estimation on FreiHAND dataset. We also adapt NVF to the classic task of root-relative 3D hand pose estimation, for which NVF also achieves state-of-the-art results on HO3D dataset. + +Since estimating absolute 3D pose from an RGB image is an ill-posed problem due to scale and depth ambiguity [13,29], in Sec. 4.4, we also provide ablation analysis on hand scale based on results from NVF and the baselines. + +This work makes the following contributions: + +1. We propose Neural Voting Field (NVF), as the first 3D implicit representation-based unified solution to estimate camera-space 3D hand pose. +2. NVF follows a novel unified 3D dense regression scheme to estimate camera-space 3D hand pose via dense 3D point-wise voting in camera frustum. +3. NVF outperforms baseline methods based on holistic and 2D dense regression and achieves state-of-the-art results on absolute and relative hand pose estimation. + +# 2. Related Work + +# 2.1. Camera-Space 3D Hand Pose Estimation + +As shown in Tab. 1, most existing works for monocular RGB input-based camera-space 3D hand pose estimation generally follow a two-stage estimation scheme. Iqbal et al. [29] recovers 3D hand pose in camera space up to a scaling factor by solving a quadratic equation given the camera intrinsic parameters and the estimated scale-normalized and root-relative 2.5D pose via 2D Dense regression. Extra global scale estimation is required for camera-space absolute 3D hand pose recovery. Another line of work [8,21,22] first obtains 2D pose or root-relative 3D hand pose estimations via 2D dense or holistic regression. Then, Hasson et al. [21] fits a 3D hand model to the initial estimations and recovers 3D hand pose in camera space from the input RGB video. Both ObMan [22] and CMR [8] work on a single RGB input and directly recover 3D global root from the initial 2D pose and root-relative 3D pose/mesh estimations. I2L-MeshNet [42] proposes a 1D dense regression scheme to predict per-lixel likelihood on 1D heat map for recovering root-relative 2.5D pose, and then uses a separate network (i.e., RootNet [41]) for the absolute root depth estimation. + +# 2.2. Root-relative 3D Hand Pose Estimation + +For single RGB input-based root-relative 3D hand pose estimation, some early methods [52, 59, 63] adopt classic frameworks such as Convolutional Neural Network and Multi-Layer Perceptron in combination with holistic or dense regression scheme for scale-invariant and rootrelative 3D hand joint positions recovery. After the introduction of the 3D parametric MANO [48] model, tremendous literature [3, 4, 9, 22, 45, 53, 60] attempts to integrate MANO into the learning pipeline and directly regress + +![](images/5c5b97a25cfe87446c39d98302f4648cde69e9cbef49246741415c21b4f95b2b.jpg) +Figure 2. Overview of Neural Voting Field (NVF): We first extract a feature map $F$ from the input RGB image $I$ by an image encoder. Then, for a 3D query point $\pmb{p}$ sampled in camera space, we define an implicit function $f_{\mathrm{NVF}}$ realized by a MLP, which maps from pixel-aligned image feature $S(F, \pi(\pmb{p}))$ and depth $Z(\pmb{p})$ of $\pmb{p}$ to its signed distance $s$ and a set of 4D offset vectors $V$ . Each 4D offset vector $v$ contains a 1D voting weight and a 3D directional vector to each joint. The 3D hand pose is then obtained by dense 3D point-wise voting. + +![](images/ef3ea652f3cfe22d6df35f499c4a15b8e0e6c716de1d7b6ff2d444fb2d78e63a.jpg) + +MANO parameters from the image. Meanwhile, MANO model-free approaches [6] continue to develop and more works also start to investigate the dependencies among hand joints as hands are inherently structured. Thus, geometric deep learning-based models (e.g., Graph Convolution, Spiral Convolution) [5, 7, 8, 15, 32, 53] and Transformer-based models [19, 35, 36, 45] have also been proposed. + +# 2.3. 3D Implicit Representation + +Recent methods [1, 2, 10, 17, 37, 40, 46, 51] have shown that, as a continuous and differentiable function realized by Multi-Layer Perceptron, memory efficient deep implicit function can represent detailed 3D shapes with arbitrary topology and have no resolution limitations. It can also be extended to RGB image-based 3D reconstruction [40, 44, 46, 57]. Grasping Field [30] proposes to reconstruct hands grasping object via predicting signed distances for hand and object given the input point cloud or RGB image. Unlike previous implicit functions relying on a global feature vector, given an input RGB image with cleanly segmented foreground, Pixel-aligned Implicit Function [49] (PIFu) utilizes pixel-aligned image feature and depth value of each 3D point for occupancy prediction. PIFu-based methods have shown the ability to generate high-fidelity 3D human geometry with more clothes details [27, 31, 50, 62] and robust 6D pose estimation. For instance, Neural Correspondence Field [24] predicts signed distance and 3D object coordinate for each 3D point to establish dense 3D-3D correspondences for robust 6D object pose estimation. + +# 3. The Proposed Method + +# 3.1. Preliminaries + +Pixel-Aligned Implicit Function (PIFu) [49]. PIFu, by learning an implicit function in 3D domain with pixel-aligned features, generates detailed 3D human geometry in + +cluding largely occluded regions from a segmented RGB image. Specifically, given a segmented image $I:\mathbb{R}^2\mapsto \mathbb{R}^3$ PIFu first extracts a feature map $F:\mathbb{R}^2\mapsto \mathbb{R}^C$ with $C$ channels by a fully convolutional image encoder. Then, for a 3D point $\pmb {p}\in \mathbb{R}^3$ in the camera space and its 2D projection $\pi (\pmb {p}):\mathbb{R}^3\mapsto \mathbb{R}^2$ on $I$ , the implicit function, represented by a Multi-Layer Perceptron (MLP) $f_{\mathrm{PIF_u}}:\mathbb{R}^C\times \mathbb{R}\mapsto$ $\mathbb{R}$ as $f_{\mathrm{PIF_u}}(S(F,\pi (\pmb {p})),Z(\pmb {p})) = y$ maps the pixel-aligned feature $S(F,\pi (\pmb {p}))$ and the depth $Z(\pmb {p})$ of $\pmb{p}$ to the occupancy $y\in [0,1]$ (1 means $\pmb{p}$ is inside the human surface and 0 means otherwise). $S$ is the sampling function (e.g., bilinear interpolation) to sample the value of $F$ at pixel $\pi (\pmb {p})$ + +Signed Distance Function (SDF) [12, 46]. Given a 3D shape $M \in \mathcal{M}$ , a signed distance function is a continuous function $\phi: \mathbb{R}^3 \times \mathcal{M} \mapsto \mathbb{R}$ as $\phi(\pmb{p}, M) = s$ , mapping a 3D point $\pmb{p}$ to the signed distance $s$ between $\pmb{p}$ and the shape surface. The sign denotes if $\pmb{p}$ is outside (positive) or inside (negative) of a watertight surface, with the iso-surface $\phi(\pmb{p}, M) = 0$ implicitly representing the surface. + +# 3.2. Dense Offset-based Pose Re-Parameterization + +As opposed to holistic regression directly mapping the input data to sparse and absolute joint positions and 2D dense regression which mainly exploits the 2D input local evidence, our method aims to directly model in 3D domain and thus fully exploit the 3D dense local evidence and hand global structure. To achieve this goal, given a common sparse and explicit pose representation (i.e., a set of 3D coordinates of $T$ hand joints) as $J = \{j_{t}\}_{t=1}^{T} \in \mathcal{J}$ , $j_{t} \in \mathbb{R}^{3}$ , we parameterize it as a set of dense 3D point-wise offset-based representation. As shown in previous works [16,54], it is non-ideal to directly learn the 3D offset vector from each 3D point to a joint due to the large variance of offsets and far away joints are also beyond the scope of receptive field. Thus, given a camera-space 3D hand pose $J$ and + +shape $M$ , we define the pose parameterization as a function $\psi$ , mapping from a 3D point $\pmb{p}$ in the camera space to a set of 4D offset vectors $V = \{\pmb{v}_t\}_{t=1}^T$ , $\pmb{v}_t \in \mathbb{R}^4$ : + +$$ +\psi : \mathbb {R} ^ {3} \times \mathcal {J} \times \mathcal {M} \mapsto \mathbb {R} ^ {T \times 4} \text {a s} \psi (\boldsymbol {p}, J, M) = V. \tag {1} +$$ + +Each offset vector $\pmb{v}_t = (w_t, \pmb{d}_t)$ consists of a 1D voting weight $w_t \in \mathbb{R}$ as point-wise closeness to hand joint $j_t$ : + +$$ +w _ {t} = \left\{ \begin{array}{c c} 1 - \frac {\| \boldsymbol {j} _ {t} - \boldsymbol {p} \| _ {2}}{r} & | s | < \delta \text {a n d} \| \boldsymbol {j} _ {t} - \boldsymbol {p} \| _ {2} \leq r \text {a n d} \boldsymbol {p} \in B _ {t} ^ {K}, \\ 0 & \text {o t h e r w i s e}; \end{array} \right. \tag {2} +$$ + +and a 3D unit directional vector $\pmb{d}_t \in \mathbb{R}^3$ as point-wise direction to hand joint $j_t$ : + +$$ +\boldsymbol {d} _ {t} = \left\{ \begin{array}{c c} \frac {\boldsymbol {j} _ {t} - \boldsymbol {p}}{\| \boldsymbol {j} _ {t} - \boldsymbol {p} \| _ {2}} & | s | < \delta \text {a n d} \| \boldsymbol {j} _ {t} - \boldsymbol {p} \| _ {2} \leq r \text {a n d} \boldsymbol {p} \in B _ {t} ^ {K}, \\ 0 & \text {o t h e r w i s e}; \end{array} \right. \tag {3} +$$ + +where $s = \phi(\pmb{p}, M)$ and $|s| < \delta$ means points that are within clamping distance $\delta$ from the hand surface. $r$ denotes the radius of a 3D ball centered at each joint and $B_t^K$ denotes the set of $K$ -Nearest Neighbors (KNN) to joint $\pmb{j}_t$ among all the near-surface points (i.e., points with $|s| < \delta$ ). Thus, among all the 3D points that are in the hand surface vicinity, we assign the non-zero 4D offset vector $\pmb{v}_t$ to the $K$ nearest points inside the 3D ball with radius $r$ of joint $\pmb{j}_t$ . + +# 3.3. Neural Voting Field + +To fully exploit the 3D local evidence and hand global geometry for camera-space 3D hand pose estimation, based on the aforementioned signed distance function and pose parameterization, we unify both modalities using a pixel-aligned implicit function. Following the proposed pipeline in Fig. 2, we first extract a $C$ -channel feature map $F$ from the input RGB image $I$ by a hourglass network $g(I; \boldsymbol{\eta})$ as used in [24, 49]. Then, for a 3D query point $\pmb{p}$ sampled in camera space, we define a continuous implicit function $f_{\mathrm{NVF}}$ realized by a MLP, which maps from pixel-aligned image feature $S(F, \pi(\pmb{p}))$ and depth $Z(\pmb{p})$ of point $\pmb{p}$ to its signed distance $s$ and a set of 4D offset vectors $V$ : + +$$ +\begin{array}{l} f _ {\mathrm {N V F}}: \mathbb {R} ^ {C} \times \mathbb {R} \mapsto \mathbb {R} \times \mathbb {R} ^ {T \times 4} \quad \text {a s} \\ f _ {\mathrm {s}}: (C (E _ {\mathrm {s}} (\cdot)), Z (\cdot), \boldsymbol {a}) \quad (\cdot , V) \end{array} \tag {4} +$$ + +$$ +f _ {\mathrm {N V F}} \big (S (F, \pi (\boldsymbol {p})), Z (\boldsymbol {p}); \boldsymbol {\theta} \big) = (s, V), +$$ + +where $\eta$ and $\theta$ are the parameters of the hourglass $g$ and MLP $f_{\mathrm{NVF}}$ , respectively. We use a pinhole camera model for 3D-to-2D projection. For each 3D point, besides the set of 4D offset vectors, NVF also predicts its signed distance to the hand surface, helping to learn the global hand geometry and also to select near-surface 3D points by thresholding the signed distances. The near-surface 3D points are then considered as valid 3D voters to be used in the voting stage. + +Optimization. During training, we sample $N$ 3D points $P = \{\pmb{p}_n\}_{n=1}^N$ in camera frustum for each image and apply + +direct supervision to both predicted signed distance $s_n$ as $L_s$ and the predicted set of 4D offset vectors $V_n$ as $L_V$ . Given the ground-truth signed distance $\hat{s}_n = \phi(\pmb{p}_n, \hat{M})$ and set of offset vectors $\hat{V}_n = \psi(\pmb{p}_n, \hat{J}, \hat{M})$ , where $\hat{J}$ and $\hat{M}$ are the ground-truth pose and mesh, $L_s$ and $L_V$ are defined as: + +$$ +L _ {s} = \frac {1}{N} \sum_ {n = 1} ^ {N} | \operatorname {c l a m p} (\hat {s} _ {n}, \delta) - \operatorname {c l a m p} (s _ {n}, \delta) |, \tag {5} +$$ + +$$ +L _ {V} = \frac {1}{N} \sum_ {n = 1} ^ {N} \mathbb {1} \left(\left| \hat {s} _ {n} \right| < \delta\right) H \left(\hat {V} _ {n}, V _ {n}\right), \tag {6} +$$ + +where $\mathrm{clamp}(s,\delta)\coloneqq \min (\delta ,\max (-\delta ,s))$ uses $\delta$ to control the distance from the hand surface over which we want to main a metric SDF, following [46]. $H$ is the Huber loss [28] and $\mathbb{1}(\cdot)$ is an indicator function that selects near-surface points (i.e., points that are within clamping distance $\delta$ from the hand surface). Then, We jointly train the hourglass network $g$ and MLP $f_{\mathrm{NVF}}$ end-to-end by solving the following optimization problem with $\lambda$ as the balancing weight: + +$$ +\boldsymbol {\eta} ^ {\star}, \boldsymbol {\theta} ^ {\star} = \underset {\boldsymbol {\eta}, \boldsymbol {\theta}} {\arg \min } L _ {s} + \lambda L _ {V}. \tag {7} +$$ + +Sampling of 3D Query Points. During training, unlike voxel-based methods which require discretization of ground-truth 3D meshes, we directly sample 3D points on the fly from the ground-truth mesh following [24, 49]. To make NVF fully exploit meaningful local evidence and understand the hand global geometry, it is crucial to focus on the hand by sampling the query points more densely around the hand surface during training. Thus, given the 3D camera-space hand model, we first sample 12500 points nearby the surface, 1000 points inside the bounding sphere of the model, and 1000 points inside the camera frustum. Among all these points, we then sample 2500 points inside the model and 2500 points outside to form a set of 5000 points in total for training. During inference, the points are directly sampled at centers of voxels that fill up the same camera frustum without any ground-truth information used. + +# 3.4. Dense 3D Point-to-Joint Voting. + +During inference, assuming $N$ 3D points $P = \{\pmb{p}_n\}_{n=1}^N$ sampled in camera frustum for each image, for each point $\pmb{p}_n$ that is considered as a valid 3D voter (i.e., the predicted signed distance $s_n$ is below the threshold $\delta$ ), based on its predicted set of 4D offset vectors $V_n = \{\pmb{v}_t^n\}_{t=1}^T$ , we first convert each 4D offset vector $\pmb{v}_t^n = (w_t^n, \pmb{d}_t^n)$ into the actual 3D offset $\pmb{o}_t^n$ from $\pmb{p}_n$ to joint $\pmb{j}_t$ based on Eq. 2 and Eq. 3: + +$$ +\boldsymbol {o} _ {t} ^ {n} = \mathbb {1} \left(\left| s _ {n} \right| < \delta\right) \left[ r \left(1 - w _ {t} ^ {n}\right) \boldsymbol {d} _ {t} ^ {n} \right]. \tag {8} +$$ + +Then, each camera-space joint location $\mathbf{j}_t$ is calculated simply by a weighted average over each voter's 3D offset prediction: + +$$ +\boldsymbol {j} _ {t} = \sum_ {n = 1} ^ {N} \frac {\mathbb {1} \left(\left| s _ {n} \right| < \delta\right) w _ {t} ^ {n} \left(\boldsymbol {o} _ {t} ^ {n} + \boldsymbol {p} _ {n}\right)}{\sum_ {n = 1} ^ {N} \mathbb {1} \left(\left| s _ {n} \right| < \delta\right) w _ {t} ^ {n}}. \tag {9} +$$ + +Finally, we obtain the final camera-space 3D hand pose as $J = \{j_{t}\}_{t=1}^{T}$ . Furthermore, instead of using all the near-surface points as valid voters, we can also only focus on using near-surface points with high predicted voting weight, as a higher voting weight means the point is more confident and closer to a hand joint. Therefore, we use voting fraction as a parameter to control the fraction of near-surface points with the highest voting weight for voting. Related ablation studies will be given in Sec. 4.7. + +# 4. Experiments + +# 4.1. Baseline Methods + +To demonstrate the effectiveness of our proposed 3D dense regression scheme against classic holistic regression and 2D dense regression, we design two directly comparable baselines that share the same architecture of the hourglass network and the MLP as NVF described in Sec. 3. Specifically, similar to NVF, both baselines first extract $C$ -channel feature map from the input RGB image $I$ . Then, unlike our $f_{\mathrm{NVF}}$ which takes a sampled 3D query point's pixel-aligned feature $S(F,\pi (\pmb {p}))$ and its depth $Z(\pmb {p})$ to predict its signed distance $s$ and the set of 4D offset vectors $V$ , the two baselines take the following steps. + +Baseline-Holistic. We directly apply a global average pooling $\alpha$ to its feature map $F^{\prime}$ extracted by the hourglass network $g(I;\pmb {\eta}^{\prime})$ and generate a $C$ -dimension feature vector $\alpha (F^{\prime})$ . The feature vector is then passed through the MLP $f_{\mathrm{HOL}}:\mathbb{R}^{C}\mapsto \mathbb{R}^{T\times 3}$ as $f_{\mathrm{HOL}}(\alpha (F^{\prime});\pmb {\theta}^{\prime}) = J^{\prime}$ $f_{\mathrm{HOL}}$ directly outputs $T$ 3D hand joint coordinates $J^{\prime} = \{\pmb{j}_{t}^{\prime}\}_{t = 1}^{T}$ in camera space. Given the ground-truth pose $\hat{J}$ , it is trained by solving the following optimization problem with a loss function $L_{J^{\prime}} = H(\hat{J},J^{\prime})$ .. + +$$ +\boldsymbol {\eta} ^ {\prime \star}, \boldsymbol {\theta} ^ {\prime \star} = \underset {\boldsymbol {\eta} ^ {\prime}, \boldsymbol {\theta} ^ {\prime}} {\arg \min } L _ {J ^ {\prime}}. \tag {10} +$$ + +Baseline-2D-Dense. We adopt a 2D dense regression scheme to estimate the camera-space pose via dense 2D pixel-wise voting, which is different from NVF which is based on dense 3D point-wise voting. Given the feature map $F''$ extracted by the hourglass network $g(I; \boldsymbol{\eta}^{\prime \prime})$ , we only pass the pixel-aligned feature $F''(\boldsymbol{u})$ at each 2D pixel location $\boldsymbol{u}$ through the MLP $f_{\mathrm{DEN}}: \mathbb{R}^C \mapsto \mathbb{R} \times \mathbb{R}^{T \times 4}$ as $f_{\mathrm{DEN}}(F''(\boldsymbol{u}); \boldsymbol{\theta}'') = (e, V'')$ . $f_{\mathrm{DEN}}$ outputs the probability $e \in [0,1]$ that the hand is present at $\boldsymbol{u}$ and also a set of 4D vectors $V'' = \{\boldsymbol{v}_t''\}_{t=1}^T$ . Each 4D vector $\boldsymbol{v}_t'' = (w_t'', j_t'')$ contains a 1D voting weight $w_t''$ and 3D joint coordinate $j_t''$ . For the 1D voting weight, it is defined similarly to the one in NVF. We first use the camera ray for each pixel to find its first 3D intersection point with the 3D hand model in the camera space. Then, we directly follow Eq. 2 to calculate the pixel's voting weight as the closeness to each joint. It is trained by solving the following optimization + +tion problem with $\lambda''$ as the balancing weight: + +$$ +\begin{array}{l} \boldsymbol {\eta} ^ {\prime \prime \star}, \boldsymbol {\theta} ^ {\prime \prime \star} = \underset {\boldsymbol {\eta} ^ {\prime \prime}, \boldsymbol {\theta} ^ {\prime \prime}} {\arg \min } L _ {e} + \lambda^ {\prime \prime} L _ {V ^ {\prime \prime}} (11) \\ = \underset {\boldsymbol {\eta} ^ {\prime \prime}, \boldsymbol {\theta} ^ {\prime \prime}} {\arg \min } \frac {1}{L} \left[ \sum_ {l = 1} ^ {L} E \left(\hat {e} _ {l}, e _ {l}\right) + \lambda^ {\prime \prime} \sum_ {l = 1} ^ {L} \hat {e} _ {l} H \left(\hat {V} _ {l} ^ {\prime \prime}, V _ {l} ^ {\prime \prime}\right) \right], (12) \\ \end{array} +$$ + +where $L$ is the number of pixels, $E$ is the binary cross-entropy loss, $\hat{e}_l$ is given by the ground-truth hand mask, and $\hat{V}_l^{\prime \prime}$ is the ground-truth set of 4D vectors. During inference, for pixel $\pmb{u}_l$ with predicted probability $e_l > 0.5$ (i.e., predicted as the foreground pixel), its predicted voting weight and 3D joint location are used to calculate the camera-space pose $J^{\prime \prime} = \{\dot{j}_t^{\prime \prime}\}_{t = 1}^{T}$ by a weighted average, similar to Eq. 9. + +# 4.2. Dataset and Evaluation Metrics + +FreiHAND Dataset [64]. As a real 3D hand dataset, it provides 130,240 training and 3,960 evaluation samples. The result of the evaluation set is evaluated via the official server for online evaluation. Since FreiHAND provides hand scale during evaluation time, we provide an ablation on hand scale for analysis of 2D-to-3D scale and depth ambiguity. + +Metrics. For FreiHAND, we evaluate for the task of camera-space 3D hand pose estimation. As in [8,64], we report the unaligned version of Mean Joint Error (mm) which is measured in camera space as CS-MJE and the Area-Under-the-Curve of PCK (Percentage of Correct Keypoints) vs. error thresholds as CS-AUC to evaluate the absolute 3D hand pose. We also define and report a 3D Translational Error (mm) (TE) [23] and Depth Error (mm) (DE) to evaluate the predicted and ground-truth pose's centroid and the depth-axis of the centroid, respectively. The results of TE and DE will be used in analysis for baseline studies, especially for the analysis of scale and depth ambiguity. + +HO3D Dataset [18]. As a real 3D hand-object dataset, it contains 66,034 training samples and 11,524 evaluation samples. We use the official server for online evaluation as the annotations of the evaluation set are not available. + +Metrics. For HO3D, we evaluate for the task of rootrelative 3D hand pose estimation. As in [18, 19], we report the Mean Joint Error (mm) after scale-translation alignment of the root as MJE and corresponding AUC to evaluate the pose. We also follow [45] to add ground-truth root to the predicted root-relative pose and report error (mm) as RS-MJE as it is equivalent to measuring in root-relative space. + +Complement (Comp) Dataset [7]. As a synthetic 3D hand dataset, it provides 1,520 poses and 71 viewpoints for research use. We only use it for pre-training when needed. + +# 4.3. Implementation Details + +For NVF and two baselines, the network architecture is adopted from PIFu [49]. Specifically, we use hourglass net + +
MethodExtra DataHand CropHand ScaleCS-MJE↓
ObMan [22]-✓✗85.2
MANO CNN [64]-✓✗71.3
I2L-MeshNet [42]-✓✗60.3
CMR-SG-RN18 [8]-✓✗49.7
CMR-SG-RN50 [8]-✓✗48.8
Baseline-Holisitc-✗✗54.5
Baseline-2D-Dense-✗✗53.2
CS-NVF (Ours)-✗✗47.2
Baseline-Holisitc-✗✓50.4
Baseline-2D-Dense-✗✓49.0
CS-NVF (Ours)-✗✓42.4
CS-NVF (Ours)Comp*✗✗44.6
CS-NVF (Ours)Comp*✗✓39.3
+ +work with the output stride of 4 and output channel of 256 as the image encoder. All the MLPs $f_{\mathrm{NVF}}$ , $f_{\mathrm{HOL}}$ , and $f_{\mathrm{DEN}}$ have four hidden fully-connected layers with the number of neurons as (1024, 512, 256, 128). Unless stated otherwise, during training for NVF, the clamping distance $\delta$ is set to 5 mm. The KNN is set to 1024. The ball radius $r$ is set to 80 mm. During inference, the sampling step size of 3D points is 16 mm in all three axes. The voting fraction is set to 50%. Unlike PIFu, no image segmentation is applied. Camera intrinsic parameters are assumed known. Base on the above setting, for camera-space 3D hand pose estimation, we directly sample 3D points in camera space, as described in Sec. 3.3. We call NVF working in camera space as CS-NVF. Unlike other methods in Tab. 2, CS-NVF and both baselines take the original image as input without hand detection or cropping. For the task of root-relative 3D hand pose estimation, we follow similar sampling strategy but sample in a hand root-relative 3D cube instead of the camera frustum. We then call NVF working in root-relative space as RS-NVF. RS-NVF takes cropped hand-centered image. Please refer to the supplementary for more details. + +# 4.4. Baseline Studies + +In this section, based on the results from CS-NVF and both baselines (as in Tab. 2, Tab. 3, Fig. 3, and Fig. 4), we investigate the impact of 2D-to-3D scale and depth ambiguity for the task of absolute 3D hand pose estimation and the effectiveness of using 3D dense regression over both 2D dense regression and holistic regression. + +2D-to-3D Scale and Depth Ambiguity. Estimating absolute 3D pose from an RGB image is an ill-posed problem due to 2D-to-3D scale and depth ambiguity [13, 29]. In view of this matter, FreiHAND [64] provides hand scale during evaluation for participants to use. As most existing works do not use the provided hand scale and thus, no + +Table 2. Comparison for absolute 3D hand pose on FreiHAND.*: pre-training on Comp. Note that our CS-NVF and two baselines take the original image without hand detection or cropping. + +
MethodExtra DataHand ScaleTE↓DE↓
Baseline-Holisitc-X50.649.1
Baseline-2D-Dense-X49.247.9
CS-NVF (Ours)-X43.642.4
Baseline-Holisitc-✓46.945.5
Baseline-2D-Dense-✓45.343.9
CS-NVF (Ours)-✓38.937.8
Baseline-HolisitcComp*X48.747.1
Baseline-2D-DenseComp*X47.946.4
CS-NVF (Ours)Comp*X41.540.4
Baseline-HolisitcComp*✓41.740.1
Baseline-2D-DenseComp*✓40.538.8
CS-NVF (Ours)Comp*✓36.535.5
+ +Table 3. Comparison of 3D Translational and Depth Error for absolute 3D hand pose on FreiHAND. \*: pre-training on Comp. + +![](images/a79c1edbd3877a69c57f267f417210b4558e36cde782abb28345014c76ec10c8.jpg) +Figure 3. 3D PCK for absolute 3D hand pose on FreiHAND. + +relevant ablation on hand scale exists, we provide ablation results on hand scale in Tab. 2 and Tab. 3 for CS-NVF and two baselines. First, based on results in Tab. 2 and Tab. 3, we observe that, for all three pipelines, the main cause of the error between the ground-truth and predicted hand joints is the global translation and especially the depth component as TE is close to CS-MJE and DE is close to TE. For instance, CS-NVF with hand scale and Comp pre-training achieves $39.3\mathrm{mm}$ CS-MJE, while TE and DE are $36.5\mathrm{mm}$ and $35.5\mathrm{mm}$ , respectively. This validates the fact that depth ambiguity is indeed one of the main challenges for estimating accurate absolute 3D hand pose from a single RGB image. Moreover, for CS-NVF and the two baselines, using hand scale leads to significant improvements in terms of all three metrics. It is also worth noting that, for all three pipelines, results only using hand scale outperform results using only extra data. For instance, CS-NVF with hand scale achieves $42.4\mathrm{mm}$ CS-MJE, while CS-NVF with Comp pre-training + +![](images/ee33d041aa02eea1f78ed4791a18e2ba076dcd420e9cf3c2e06bb11b389f1eb2.jpg) +Figure 4. Qualitative results for absolute and relative 3D pose: (b) Predicted hand mask (Baseline) or mesh generated by Marching Cubes from the predicted signed distances (NVF). (c-h) Predicted 1D voting weight from each predicted foreground pixel (Baseline) or predicted near-surface 3D point (NVF) to a joint. (Brighter line means larger weight, darker or no line means smaller weight) (i) Estimated 3D pose. NVF makes solid point-wise prediction and overall distribution of signed distances and voting weights in highly occluded region. + +
MethodMJE↓AUC↑RS-MJE↓
Hasson et al. [20]36.90.369-
Pose2Mesh [11]33.30.48033.2
ObMan [22]31.80.46155.2
Liu et al. [38]31.70.46330.0
Hampali et al. [18]30.40.494-
METRO [35]28.90.504-
I2L-MeshNet [42]26.00.52926.8
Keypoint Trans. [19]25.70.553-
ArtiBoost [58]25.30.532-
Zheng et al. [61]25.10.541-
HandOccNet [45]24.00.55724.9
RS-NVF (Ours)21.80.61023.2
+ +Table 4. Comparison for relative 3D hand pose on HO3D. + +has $44.6 \mathrm{~mm}$ CJ-MJE. These verify that a reasonable understanding towards the hand scale would help for robust depth learning and alleviate depth ambiguity, which is also aligned with existing literature [13, 29, 33]. + +Dense Regression vs. Holistic Regression. In Tab. 2 and Fig. 3, we can observe that both CS-NVF and Baseline-2D-Dense consistently outperform Baseline-Holistic on CS-MJE and CS-AUC under different training settings. For instance, without using hand scale, Baseline-2D-Dense is 1.3 mm better than Baseline-Holistic in CS-MJE, and CS-NVF can be 7.3 mm better. Since three pipelines share the same network architecture, this supports our point that dense regression is more effective than holistic regression for handling highly articulated pose data. + +3D Dense Regression vs. 2D Dense Regression. Quantitatively, in Tab. 2 and Fig. 3, it can be observed that CS-NVF consistently outperforms Baseline-2D-Dense by a large margin on CS-MJE and CS-AUC under different + +training settings. For instance, without using hand scale, CS-NVF can be $6\mathrm{mm}$ better on CS-MJE. Qualitatively, as indicated by the white circles in column (c-h) of Fig. 4, both methods are able to find valid voters (Baseline-2D-Dense: pixels with predicted probability $e > 0.5$ . CS-NVF: 3D points with predicted signed distance $|s| < \delta$ ). However, CS-NVF generally provides not only more number of valid voters, but also more reasonable element-wise prediction and overall distribution of the voting weight than Baseline-2D-Dense. Specifically, for the number of voters, since CS-NVF models all the 3D points around the hand surface and Baseline-2D-Dense can only model foreground pixels, CS-NVF generally can find and use more valid voters. As for element-wise prediction and overall distribution of the voting weight, CS-NVF is generally able to predict large voting weight for points close to hand joint and small value for points far away, resulting in well-shaped symmetric distribution around the corresponding joint, even in occluded regions. Note that brighter line indicates larger voting weight, darker or no line indicates smaller voting weight. Moreover, as shown in column (b) for the mesh rendering of CS-NVF (obtained using Marching Cubes [39] from predicted signed distances), even in highly occluded region, CS-NVF provides solid signed distance distribution showing its ability to reason the global hand geometry. Overall, these demonstrate that CS-NVF, through direct dense modeling in 3D domain, can better model 3D dense local evidence, the relation between voters and 3D hand pose, and also the global hand geometry, compared with 2D dense regression-based counterpart, which leads to robust 3D hand pose estimation. + +# 4.5. Main Results: Absolute 3D Hand Pose + +In Tab. 2 and Fig. 3, we also compare CS-NVF with current state-of-the-art methods on FreiHAND for absolute 3D + +
Proposed MethodEvaluation MetricVoting Fraction (%)Clamping Distance (mm)KNN
25507510015102025651210242048
CS-NVF (Ours)CS-MJE↓39.939.339.540.239.339.340.040.140.339.539.339.3
CS-AUC↑0.360.370.360.360.370.370.360.360.350.360.370.37
RS-NVF (Ours)MJE↓22.221.821.922.021.921.821.821.923.622.621.821.8
AUC↑0.600.610.610.600.600.610.610.610.590.600.610.61
+ +Table 5. Ablation study on parameter choices for absolute and relative 3D hand pose: Voting fraction, clamping distance, and KNN + +
Step SizeCS-NVF (Ours)
#PointsCS-MJE↓CS-AUC↑FPS↑
41882k39.20.371
8230k39.20.378
1628k39.30.3734
248k39.80.3650
323k42.40.3456
+ +Table 6. Ablation study on sampling step size for absolute 3D hand pose on FreiHAND. #points: number of sampled 3D points. + +hand pose estimation. In Tab. 2, without the use of hand scale and extra data as a fair comparison, CS-NVF surpasses previous state-of-the-art methods, achieving $1.6\mathrm{mm}$ better than CMR [8]. Note that, unlike other methods, CS-NVF takes the original image as input without hand detection or cropping. Additional hand scale information used without extra data improves the CS-MJE of CS-NVF further to $42.4\mathrm{mm}$ . Additional Comp pre-training improves the CS-MJE of CS-NVF to $44.6\mathrm{mm}$ without hand scale and $39.3\mathrm{mm}$ with hand scale. Referring to Fig. 3, CS-NVF also achieves state-of-the-art performance on CS-AUC. + +# 4.6. Main Results: Relative 3D Hand Pose + +In Tab. 4, we compare RS-NVF with current state-of-the-art methods on HO3D for relative 3D hand pose estimation. RS-NVF significantly outperforms all these methods in all three metrics. Especially, RS-NVF outmatches HandOccNet [45] by $2.2\mathrm{mm}$ on MJE, $1.7\mathrm{mm}$ on RS-MJE, and also has 0.610 AUC against 0.557 AUC from HandOccNet. + +# 4.7. Ablation Studies + +Sampling Step Size. During inference, 3D query points are sampled at centers of voxels that fill up the 3D space. Since the choices of different sampling step sizes can influence both accuracy and efficiency, in Tab. 6, we show the ablation on sampling step size for CS-NVF. With $16\mathrm{mm}$ as the step size, CS-NVF achieves 34 FPS and $39.3\mathrm{mm}$ on CS-MJE given 28k points. Reducing the step size to $4\mathrm{mm}$ yields 1,882k points and 1 FPS with slightly improved CS-MJE of $39.2\mathrm{mm}$ . Increasing to $32\mathrm{mm}$ step size yields 3k points and improves the FPS to 56 with still competitive $42.4\mathrm{mm}$ on CS-MJE. For RS-NVF not shown in the table, 4 and 8 mm step size both yield MJE of $21.7\mathrm{mm}$ and AUC of 0.62. We can observe that, while step size affects FPS, NVF is relatively insensitive regarding the performance. In all other reported results, we use $16\mathrm{mm}$ as the default step size. + +Voting Fraction. Since the predicted point-wise voting weight reflects the confidence and closeness of a given point towards a certain joint, besides using all the near-surface points as valid voters, we can use a specified fraction of near-surface points with the highest voting weight for voting during inference. In Tab. 5, we observe that using all the near-surface points (100%) yields competitive results while 50% fraction leads to the overall better performance for both CS-NVF and RS-NVF. The results are aligned with our algorithm design since points far from each joint are not expected to make meaningful contributions. In all other reported results, we use 50% as the default voting fraction. + +Clamping Distance. During training, clamping distance $\delta$ is used to concentrate NVF's capacity on learning high-quality point-wise prediction (i.e., signed distance and set of 4D offset vectors) near the hand surface. Ablation in Tab. 5 shows that $5\mathrm{mm}$ generates the overall better performance for both CS-NVF and RS-NVF. In all other reported results, we use $5\mathrm{mm}$ as the default clamping distance. + +$K$ -Nearest Neighbors (KNN). During training, among all the sampled points that are in the hand surface vicinity, nonzero 4D offset vector will only be assigned to the $K$ nearest points inside a 3D ball of each hand joint. As shown in previous works [16, 54], estimating offsets for large KNN is unnecessary and far away joints are beyond the scope of receptive field. Small KNN might also cause poor performance due to the sparse learning scheme. Ablation in Tab. 5 shows that 1024 KNN generates the overall better performance for both CS-NVF and RS-NVF. In all other reported results, we use 1024 as the default KNN. + +# 5. Conclusions + +We propose Neural Voting Field (NVF), as the first 3D implicit representation-based unified solution to estimate camera-space 3D hand pose given a single RGB image. NVF follows a novel unified 3D dense regression scheme to estimate 3D hand pose in 3D camera space via dense 3D point-wise voting in camera frustum. Studies with both holistic regression baseline and 2D dense regression baseline verify NVF's ability to fully exploit 3D dense local evidence and model global hand structure/geometry, which are essential for precisely locating 3D hand joints in 3D camera space. NVF also demonstrates state-of-the-art results on both tasks of camera-space 3D hand pose estimation and root-relative 3D hand pose estimation. + +# References + +[1] Matan Atzmon and Yaron Lipman. SAL: Sign agnostic learning of shapes from raw data. In CVPR, 2020. 3 +[2] Matan Atzmon and Yaron Lipman. SALD: Sign agnostic learning with derivatives. In ICLR, 2021. 3 +[3] Seungryul Baek, Kwang In Kim, and Tae-Kyun Kim. Pushing the envelope for rgb-based dense 3d hand pose estimation via neural rendering. In CVPR, 2019. 2 +[4] Adnane Boukhayma, Rodrigo de Bem, and Philip HS Torr. 3d hand shape and pose from images in the wild. In CVPR, 2019. 1, 2 +[5] Yujun Cai, Liuhao Ge, Jun Liu, Jianfei Cai, Tat-Jen Cham, Junsong Yuan, and Nadia Magnenat Thalmann. Exploiting spatial-temporal relationships for 3d pose estimation via graph convolutional networks. In ICCV, 2019. 1, 3 +[6] Ping Chen, Yujin Chen, Dong Yang, Fangyin Wu, Qin Li, Qingpei Xia, and Yong Tan. I2uv-handnet: Image-to-uv prediction network for accurate and high-fidelity 3d hand mesh modeling. In ICCV, 2021. 1, 3 +[7] Xingyu Chen, Yufeng Liu, Yajiao Dong, Xiong Zhang, Chongyang Ma, Yanmin Xiong, Yuan Zhang, and Xiaoyan Guo. Mobrecon: Mobile-friendly hand mesh reconstruction from monocular image. In CVPR, 2022. 1, 3, 5 +[8] Xingyu Chen, Yufeng Liu, Chongyang Ma, Jianlong Chang, Huayan Wang, Tian Chen, Xiaoyan Guo, Pengfei Wan, and Wen Zheng. Camera-space hand mesh recovery via semantic aggregation and adaptive 2d-1d registration. In CVPR, 2021. 2, 3, 5, 6, 8 +[9] Yujin Chen, Zhigang Tu, Di Kang, Linchao Bao, Ying Zhang, Xuefei Zhe, Ruizhi Chen, and Junsong Yuan. Model-based 3d hand reconstruction via self-supervised learning. In CVPR, 2021. 2 +[10] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In CVPR, 2019. 3 +[11] Hongsuk Choi, Gyeongsik Moon, and Kyoung Mu Lee. Pose2mesh: Graph convolutional network for 3d human pose and mesh recovery from a 2d human pose. In ECCV, 2020. 7 +[12] Brian Curless and Marc Levoy. A volumetric method for building complex models from range images. SIGGRAPH, 1996. 3 +[13] Tom van Dijk and Guido de Croon. How do neural networks see depth in single images? In ICCV, 2019. 2, 6, 7 +[14] Linpu Fang, Xingyan Liu, Li Liu, Hang Xu, and Wenxiong Kang. Jgr-p2o: Joint graph reasoning based pixel-to-offset prediction network for 3d hand pose estimation from a single depth image. In ECCV, 2020. 1 +[15] Liuhao Ge, Zhou Ren, Yuncheng Li, Zehao Xue, Yingying Wang, Jianfei Cai, and Junsong Yuan. 3d hand shape and pose estimation from a single rgb image. In CVPR, 2019. 1, 3 +[16] Liuhao Ge, Zhou Ren, and Junsong Yuan. Point-to-point regression pointnet for 3d hand pose estimation. In ECCV, 2018. 1, 3, 8 +[17] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. In ICML, 2020. 3 + +[18] Shreyas Hampali, Mahdi Rad, Markus Oberweger, and Vincent Lepetit. Honnotate: A method for 3d annotation of hand and object poses. In CVPR, 2020. 5, 7 +[19] Shreyas Hampali, Sayan Deb Sarkar, Mahdi Rad, and Vincent Lepetit. Keypoint transformer: Solving joint identification in challenging hands and object interactions for accurate 3d pose estimation. In CVPR, 2022. 1, 3, 5, 7 +[20] Yana Hasson, Bugra Tekin, Federica Bogo, Ivan Laptev, Marc Pollefeys, and Cordelia Schmid. Leveraging photometric consistency over time for sparsely supervised hand-object reconstruction. In CVPR, 2020. 7 +[21] Yana Hasson, Gül Varol, Cordelia Schmid, and Ivan Laptev. Towards unconstrained joint hand-object reconstruction from rgb videos. In 3DV, 2021. 2 +[22] Yana Hasson, Gül Varol, Dimitris Tzionas, Igor Kalevatykh, Michael J. Black, Ivan Laptev, and Cordelia Schmid. Learning joint reconstruction of hands and manipulated objects. In CVPR, 2019. 2, 6, 7 +[23] Tomás Hodań, Jiří Matas, and Štepan Obdržálek. On evaluation of 6d object pose estimation. In ECCV, 2016. 5 +[24] Lin Huang, Tomas Hodan, Lingni Ma, Linguang Zhang, Luan Tran, Christopher Twigg, Po-Chen Wu, Junsong Yuan, Cem Keskin, and Robert Wang. Neural correspondence field for object pose estimation. In ECCV, 2022. 2, 3, 4 +[25] Lin Huang, Jianchao Tan, Ji Liu, and Junsong Yuan. Handtransformer: non-autoregressive structured modeling for 3d hand pose estimation. In ECCV, 2020. 1 +[26] Lin Huang, Boshen Zhang, Zhilin Guo, Yang Xiao, Zhiguo Cao, and Junsong Yuan. Survey on depth and rgb image-based 3d hand shape and pose estimation. VRIH, 2021. 1 +[27] Zeng Huang, Yuanlu Xu, Christoph Lassner, Hao Li, and Tony Tung. Arch: Animatable reconstruction of clothed humans. In CVPR, 2020. 2, 3 +[28] Peter J Huber. Robust estimation of a location parameter. Breakthroughs in statistics, 1992. 4 +[29] Umar Iqbal, Pavlo Molchanov, Thomas Breuel Juergen Gall, and Jan Kautz. Hand pose estimation via latent 2.5d heatmap regression. In ECCV, 2018. 1, 2, 6, 7 +[30] Korrawe Karunratanakul, Jinlong Yang, Yan Zhang, Michael J. Black, Krikamol Muandet, and Siyu Tang. Grasping field: Learning implicit representations for human grasps. In 3DV, 2020. 3 +[31] Nilesh Kulkarni, Justin Johnson, and David F Fouhey. What's behind the couch? directed ray distance functions (drdf) for 3d scene reconstruction. In arXiv e-prints, 2021. 3 +[32] Dominik Kulon, Riza Alp Guler, Iasonas Kokkinos, Michael M Bronstein, and Stefanos Zafeiriou. Weakly-supervised mesh-convolutional hand reconstruction in the wild. In CVPR, 2020. 1, 3 +[33] Vincent Lepetit, Francesc Moreno-Noguer, and Pascal Fua. Epnp: An accurate o (n) solution to the pnp problem. IJCV, 2009. 2, 7 +[34] Shile Li and Dongheui Lee. Point-to-pose voting based hand pose estimation using residual permutation equivariant layer. In CVPR, 2019. 1 +[35] Kevin Lin, Lijuan Wang, and Zicheng Liu. End-to-end human pose and mesh reconstruction with transformers. In CVPR, 2021. 1, 3, 7 + +[36] Kevin Lin, Lijuan Wang, and Zicheng Liu. Mesh graphormer. In ICCV, 2021. 1, 3 +[37] Feng Liu, Luan Tran, and Xiaoming Liu. Fully understanding generic objects: Modeling, segmentation, and reconstruction. In CVPR, 2021. 3 +[38] Shaowei Liu, Hanwen Jiang, Jiarui Xu, Sifei Liu, and Xiaolong Wang. Semi-supervised 3d hand-object poses estimation with interactions in time. In CVPR, 2021. 7 +[39] William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3d surface construction algorithm. SIGGRAPH, 1987. 7 +[40] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In CVPR, 2019. 3 +[41] Gyeongsik Moon, Ju Yong Chang, and Kyoung Mu Lee. Camera distance-aware top-down approach for 3d multiperson pose estimation from a single rgb image. In ICCV, 2019. 2 +[42] Gyeongsik Moon and Kyoung Mu Lee. I2l-meshnet: Image-to-lixel prediction network for accurate 3d human pose and mesh estimation from a single rgb image. In ECCV, 2020. 1, 2, 6, 7 +[43] Alejandro Newell, Kaiyu Yang, and Jia Deng. Stacked hourglass networks for human pose estimation. In ECCV, 2016. 1 +[44] Michael Oechsle, Lars Mescheder, Michael Niemeyer, Thilo Strauss, and Andreas Geiger. Texture fields: Learning texture representations in function space. In ICCV, 2019. 3 +[45] JoonKyu Park, Yeonguk Oh, Gyeongsik Moon, Hongsuk Choi, and Kyoung Mu Lee. Handocctnet: Occlusion-robust 3d hand mesh estimation network. In CVPR, 2022. 1, 2, 3, 5, 7, 8 +[46] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In CVPR, 2019. 3, 4 +[47] Georgios Pavlakos, Xiaowei Zhou, Konstantinos G Derpanis, and Kostas Daniilidis. Coarse-to-fine volumetric prediction for single-image 3d human pose. In CVPR, 2017. 1 +[48] Javier Romero, Dimitrios Tzionas, and Michael J Black. Embodied hands: Modeling and capturing hands and bodies together. TOG, 2022. 2 +[49] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In ICCV, 2019. 1, 2, 3, 4, 5 +[50] Shunsuke Saito, Tomas Simon, Jason Saragih, and Hanbyul Joo. Pifuhd: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization. In CVPR, 2020. 2, 3 +[51] Vincent Sitzmann, Eric Chan, Richard Tucker, Noah Snavely, and Gordon Wetzstein. Metasdf: Meta-learning signed distance functions. In NeurIPS, 2020. 3 +[52] Adrian Spurr, Jie Song, Seonwook Park, and Otmar Hilliges. Cross-modal deep variational hand pose estimation. In CVPR, 2018. 1, 2 + +[53] Tze Ho Elden Tse, Kwang In Kim, Ales Leonardis, and Hyung Jin Chang. Collaborative learning for hand and object reconstruction with attention-guided graph convolution. In CVPR, 2022. 1, 2, 3 +[54] Chengde Wan, Thomas Probst, Luc Van Gool, and Angela Yao. Dense 3d regression for hand pose estimation. In CVPR, 2018. 1, 3, 8 +[55] Shih-En Wei, Varun Ramakrishna, Takeo Kanade, and Yaser Sheikh. Convolutional pose machines. In CVPR, 2016. 1 +[56] Fu Xiong, Boshen Zhang, Yang Xiao, Zhiguo Cao, Taidong Yu, Joey Tianyi Zhou, and Junsong Yuan. A2j: Anchor-to-joint regression network for 3d articulated pose estimation from a single depth image. In ICCV, 2019. 1 +[57] Qiangeng Xu, Weiyue Wang, Duygu Ceylan, Radomir Mech, and Ulrich Neumann. Disn: Deep implicit surface network for high-quality single-view 3d reconstruction. In NeurIPS, 2019. 3 +[58] Lixin Yang, Kailin Li, Xinyu Zhan, Jun Lv, Wenqiang Xu, Jiefeng Li, and Cewu Lu. Artibost: Boosting articulated 3d hand-object pose estimation via online exploration and synthesis. In CVPR, 2022. 7 +[59] Linlin Yang, Shile Li, Dongheui Lee, and Angela Yao. Aligning latent spaces for 3d hand pose estimation. In ICCV, 2019. 1, 2 +[60] Xiong Zhang, Qiang Li, Hong Mo, Wenbo Zhang, and Wen Zheng. End-to-end hand mesh recovery from a monocular rgb image. In ICCV, 2019. 1, 2 +[61] Xiaozheng Zheng, Pengfei Ren, Haifeng Sun, Jingyu Wang, Qi Qi, and Jianxin Liao. Joint-aware regression: Rethinking regression-based method for 3d hand pose estimation. In BMVC, 2021. 7 +[62] Zerong Zheng, Tao Yu, Yebin Liu, and Qionghai Dai. Pamir: Parametric model-conditioned implicit representation for image-based human reconstruction. TPAMI, 2021. 2, 3 +[63] Christian Zimmermann and Thomas Brox. Learning to estimate 3d hand pose from single rgb images. In ICCV, 2017. 1, 2 +[64] Christian Zimmermann, Duygu Ceylan, Jimei Yang, Bryan Russell, Max Argus, and Thomas Brox. Freihand: A dataset for markerless capture of hand pose and shape from single rgb images. In ICCV, 2019. 5, 6 \ No newline at end of file diff --git a/2023/Neural Voting Field for Camera-Space 3D Hand Pose Estimation/images.zip b/2023/Neural Voting Field for Camera-Space 3D Hand Pose Estimation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..5331a0ca080a710bf4a69dbf43cb842cde07f915 --- /dev/null +++ b/2023/Neural Voting Field for Camera-Space 3D Hand Pose Estimation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:88a98fa6fe7fa2d974f0f9e7db71afc26d1de672701c738128b57465b1a20327 +size 565009 diff --git a/2023/Neural Voting Field for Camera-Space 3D Hand Pose Estimation/layout.json b/2023/Neural Voting Field for Camera-Space 3D Hand Pose Estimation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..0c3354fadb771cb770043e98fad728f669a1e2c3 --- /dev/null +++ b/2023/Neural Voting Field for Camera-Space 3D Hand Pose Estimation/layout.json @@ -0,0 +1,10506 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 96, + 103, + 498, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 103, + 498, + 121 + ], + "spans": [ + { + "bbox": [ + 96, + 103, + 498, + 121 + ], + "type": "text", + "content": "Neural Voting Field for Camera-Space 3D Hand Pose Estimation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 96, + 143, + 159, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 143, + 159, + 158 + ], + "spans": [ + { + "bbox": [ + 96, + 143, + 159, + 158 + ], + "type": "text", + "content": "Lin Huang*" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 148, + 143, + 296, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 143, + 296, + 158 + ], + "spans": [ + { + "bbox": [ + 148, + 143, + 296, + 158 + ], + "type": "inline_equation", + "content": "g^{1*}" + }, + { + "bbox": [ + 148, + 143, + 296, + 158 + ], + "type": "text", + "content": " Chung-Ching Lin" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 148, + 158, + 334, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 158, + 334, + 185 + ], + "spans": [ + { + "bbox": [ + 148, + 158, + 334, + 185 + ], + "type": "text", + "content": "Lijuan Wang2 Junsong Yuan1 \n1University at Buffalo" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 342, + 143, + 478, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 143, + 478, + 158 + ], + "spans": [ + { + "bbox": [ + 342, + 143, + 478, + 158 + ], + "type": "text", + "content": "Kevin Lin² Lin Lia" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 380, + 158, + 445, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 158, + 445, + 171 + ], + "spans": [ + { + "bbox": [ + 380, + 158, + 445, + 171 + ], + "type": "text", + "content": "Zicheng Liu" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 346, + 172, + 400, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 346, + 172, + 400, + 184 + ], + "spans": [ + { + "bbox": [ + 346, + 172, + 400, + 184 + ], + "type": "text", + "content": "2Microsoft" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 207, + 187, + 383, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 207, + 187, + 383, + 198 + ], + "spans": [ + { + "bbox": [ + 207, + 187, + 383, + 198 + ], + "type": "text", + "content": "https://linhuang17.github.io/NVF" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 47, + 214, + 547, + 290 + ], + "blocks": [ + { + "bbox": [ + 47, + 214, + 547, + 290 + ], + "lines": [ + { + "bbox": [ + 47, + 214, + 547, + 290 + ], + "spans": [ + { + "bbox": [ + 47, + 214, + 547, + 290 + ], + "type": "image", + "image_path": "201f7b5bc242e0e13edced3ca156e9b7fc97c83a6a18e05619536c61672bd131.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 292, + 547, + 348 + ], + "lines": [ + { + "bbox": [ + 46, + 292, + 547, + 348 + ], + "spans": [ + { + "bbox": [ + 46, + 292, + 547, + 348 + ], + "type": "text", + "content": "Figure 1. We present Neural Voting Field (NVF), a 3D implicit function for camera-space 3D hand pose estimation from an RGB image. Given an RGB input (a), for a 3D point in camera space, NVF can predict its signed distance to hand surface and a set of 4D offset vectors (1D voting weight and 3D directional vector to each joint). 3D pose is then obtained via dense 3D point-wise voting. Rendered predictions: (b) Hand mesh generated by Marching Cubes from the predicted signed distances (c-h) Predicted 1D voting weight from each predicted near-surface 3D point to a joint. (Brighter line means larger weight, darker or no line means smaller weight) (i) Estimated 3D hand pose." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 143, + 357, + 192, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 357, + 192, + 369 + ], + "spans": [ + { + "bbox": [ + 143, + 357, + 192, + 369 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 383, + 289, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 383, + 289, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 383, + 289, + 693 + ], + "type": "text", + "content": "We present a unified framework for camera-space 3D hand pose estimation from a single RGB image based on 3D implicit representation. As opposed to recent works, most of which first adopt holistic or pixel-level dense regression to obtain relative 3D hand pose and then follow with complex second-stage operations for 3D global root or scale recovery, we propose a novel unified 3D dense regression scheme to estimate camera-space 3D hand pose via dense 3D pointwise voting in camera frustum. Through direct dense modeling in 3D domain inspired by Pixel-aligned Implicit Functions for 3D detailed reconstruction, our proposed Neural Voting Field (NVF) fully models 3D dense local evidence and hand global geometry, helping to alleviate common 2D-to-3D ambiguities. Specifically, for a 3D query point in camera frustum and its pixel-aligned image feature, NVF, represented by a Multi-Layer Perceptron, regresses: (i) its signed distance to the hand surface; (ii) a set of 4D offset vectors (1D voting weight and 3D directional vector to each hand joint). Following a vote-casting scheme, 4D offset vectors from near-surface points are selected to calculate the 3D hand joint coordinates by a weighted average. Experiments demonstrate that NVF outperforms existing state-of-the-art algorithms on FreiHAND dataset for camera-space 3D hand pose estimation. We also adapt NVF to the classic task of root-relative 3D hand pose estimation, for which NVF also obtains state-of-the-art results on HO3D dataset." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 357, + 386, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 357, + 386, + 369 + ], + "spans": [ + { + "bbox": [ + 307, + 357, + 386, + 369 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 377, + 545, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 377, + 545, + 461 + ], + "spans": [ + { + "bbox": [ + 304, + 377, + 545, + 461 + ], + "type": "text", + "content": "Monocular 3D hand pose estimation, which aims to recover 3D locations of hand joints from an RGB image, has attracted enormous attention and made remarkable progress in recent years. As a long-standing task in computer vision, it remains challenging due to its highly articulated structure, large variations in orientations, severe (self-)occlusion, and inherent 2D-to-3D scale and depth ambiguity." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 462, + 546, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 546, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 546, + 605 + ], + "type": "text", + "content": "Owing to the aforementioned difficulties, most existing works [4-7, 15, 19, 26, 32, 35, 36, 45, 52, 53, 59, 60, 63] focused on one aspect of this general problem, which is to estimate root-relative 3D hand pose (i.e., 3D joint coordinates relative to a pre-defined root joint, such as hand wrist). While accurate 2D-to-3D root-relative pose estimation is essential for numerous applications in Virtual/Augmented Reality, there are various interactive tasks in which having root-relative hand joint coordinates alone is insufficient. For instance, being able to recover camera-space 3D hand joint coordinates in an AR view enables the user to directly use hands to manipulate virtual objects moving in 3D space." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": "To recover robust camera-space 3D hand pose, there are two key design elements: (1) the ability to exploit dense local evidence. Specifically, as demonstrated in previous works [14, 16, 25, 29, 34, 42, 43, 47, 49, 54-56], dense regression-based methods are more effective than holistic regression-based counterparts for handling highly articulated 3D pose structure, attributed to its ability to maintain the input data spatial structure and fully exploit local evidence; (2) the ability to reason 3D hand global geome" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 702, + 254, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 702, + 254, + 712 + ], + "spans": [ + { + "bbox": [ + 61, + 702, + 254, + 712 + ], + "type": "text", + "content": "*Work done during Lin Huang's internship with Microsoft." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8969" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 70, + 291, + 141 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 291, + 141 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 291, + 141 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 291, + 141 + ], + "type": "table", + "html": "
MethodFirst StageSecond Stage
Iqbal et al. [29]2D-DenseScale Estimation
ObMan [22]HolisticRoot Estimation
I2L-MeshNet [42]1D-DenseRoot Depth Estimation
CMR [8]2D-Dense+SpiralConvRegistration
Hasson et al. [21]HolisticModel Fitting
NVF (Ours)Unified 3D-DenseWeighted Average
", + "image_path": "04dace699167e4242252c429efc5336c637acb660e55e5a4d9bbe7af4e631939.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 146, + 287, + 168 + ], + "lines": [ + { + "bbox": [ + 47, + 146, + 287, + 168 + ], + "spans": [ + { + "bbox": [ + 47, + 146, + 287, + 168 + ], + "type": "text", + "content": "Table 1. Comparison of representative absolute 3D hand pose estimation schemes. Please refer to Sec. 2.1 for more details." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 175, + 287, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 175, + 287, + 235 + ], + "spans": [ + { + "bbox": [ + 46, + 175, + 287, + 235 + ], + "type": "text", + "content": "try. As shown in previous literature [13, 29, 33], given 2D evidence and camera intrinsic parameters, reasonable understanding towards target object 3D structure/geometry is crucial to alleviate 2D-to-3D depth ambiguity, which is the key to accurately locate 3D hand pose in camera space." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 236, + 287, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 236, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 46, + 236, + 287, + 510 + ], + "type": "text", + "content": "To fully integrate both elements into our algorithm design in a unified manner, we connect with Pixel-aligned Implicit Function (PIFu) [24, 27, 49, 50, 62]. Through direct dense modeling in 3D domain with pixel-aligned local features, PIFu-based methods reconstruct highly detailed 3D human geometry from an RGB image in a unified way, showing its ability to model high-frequency local details such as clothing wrinkles while generating complete global geometry including largely occluded region such as the back of a person. Inspired by these results, we propose a novel unified 3D dense regression scheme based on a 3D implicit function for robust camera-space 3D hand pose estimation. Specifically, for each of the 3D query points densely sampled in camera frustum and its pixel-aligned image feature, unlike PIFu predicting occupancy value for each point, our proposed Neural Voting Field (NVF) regresses: (i) the signed distance between the point and the hand surface; (ii) a set of 4D offset vectors (1D voting weight and 3D directional vector from the point to each joint). Following a vote-casting scheme, 4D offset vectors from near-surface points (i.e., points for which the predicted signed distance is below a threshold) are selected to calculate the 3D hand joint coordinates by a weighted average." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 511, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 511, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 511, + 288, + 714 + ], + "type": "text", + "content": "Most existing works for camera-space 3D hand pose estimation, as shown in Tab. 1, follow a two-stage estimation scheme. They first adopt holistic or pixel-level dense regression to obtain 2D and relative 3D hand poses and then follow with complex second-stage processing such as fitting, registration, using a separate network for 3D global root location or scale estimation. NVF instead provides a unified solution via direct dense modeling in 3D camera space followed by a simple weighted average operation, which enables reasoning about 3D dense local evidence and hand global geometry. As shown in Fig. 1, NVF makes solid 3D point-wise prediction and overall distribution of signed distance and voting weight even in highly occluded regions, leading to accurate camera-space pose estimation. In Sec. 4, we show that NVF noticeably outperforms two baselines based on holistic regression and 2D dense regression. Besides, NVF exhibits state-of-the-art performance" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 119 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 119 + ], + "type": "text", + "content": "for the task of camera-space 3D hand pose estimation on FreiHAND dataset. We also adapt NVF to the classic task of root-relative 3D hand pose estimation, for which NVF also achieves state-of-the-art results on HO3D dataset." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 120, + 545, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 120, + 545, + 167 + ], + "spans": [ + { + "bbox": [ + 304, + 120, + 545, + 167 + ], + "type": "text", + "content": "Since estimating absolute 3D pose from an RGB image is an ill-posed problem due to scale and depth ambiguity [13,29], in Sec. 4.4, we also provide ablation analysis on hand scale based on results from NVF and the baselines." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 168, + 503, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 168, + 503, + 179 + ], + "spans": [ + { + "bbox": [ + 317, + 168, + 503, + 179 + ], + "type": "text", + "content": "This work makes the following contributions:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 182, + 545, + 295 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 313, + 182, + 545, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 182, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 313, + 182, + 545, + 217 + ], + "type": "text", + "content": "1. We propose Neural Voting Field (NVF), as the first 3D implicit representation-based unified solution to estimate camera-space 3D hand pose." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 220, + 545, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 220, + 545, + 256 + ], + "spans": [ + { + "bbox": [ + 313, + 220, + 545, + 256 + ], + "type": "text", + "content": "2. NVF follows a novel unified 3D dense regression scheme to estimate camera-space 3D hand pose via dense 3D point-wise voting in camera frustum." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 259, + 545, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 259, + 545, + 295 + ], + "spans": [ + { + "bbox": [ + 313, + 259, + 545, + 295 + ], + "type": "text", + "content": "3. NVF outperforms baseline methods based on holistic and 2D dense regression and achieves state-of-the-art results on absolute and relative hand pose estimation." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 304, + 392, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 304, + 392, + 316 + ], + "spans": [ + { + "bbox": [ + 306, + 304, + 392, + 316 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 323, + 520, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 323, + 520, + 335 + ], + "spans": [ + { + "bbox": [ + 306, + 323, + 520, + 335 + ], + "type": "text", + "content": "2.1. Camera-Space 3D Hand Pose Estimation" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 341, + 545, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 341, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 341, + 545, + 581 + ], + "type": "text", + "content": "As shown in Tab. 1, most existing works for monocular RGB input-based camera-space 3D hand pose estimation generally follow a two-stage estimation scheme. Iqbal et al. [29] recovers 3D hand pose in camera space up to a scaling factor by solving a quadratic equation given the camera intrinsic parameters and the estimated scale-normalized and root-relative 2.5D pose via 2D Dense regression. Extra global scale estimation is required for camera-space absolute 3D hand pose recovery. Another line of work [8,21,22] first obtains 2D pose or root-relative 3D hand pose estimations via 2D dense or holistic regression. Then, Hasson et al. [21] fits a 3D hand model to the initial estimations and recovers 3D hand pose in camera space from the input RGB video. Both ObMan [22] and CMR [8] work on a single RGB input and directly recover 3D global root from the initial 2D pose and root-relative 3D pose/mesh estimations. I2L-MeshNet [42] proposes a 1D dense regression scheme to predict per-lixel likelihood on 1D heat map for recovering root-relative 2.5D pose, and then uses a separate network (i.e., RootNet [41]) for the absolute root depth estimation." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 587, + 512, + 599 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 587, + 512, + 599 + ], + "spans": [ + { + "bbox": [ + 306, + 587, + 512, + 599 + ], + "type": "text", + "content": "2.2. Root-relative 3D Hand Pose Estimation" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 715 + ], + "type": "text", + "content": "For single RGB input-based root-relative 3D hand pose estimation, some early methods [52, 59, 63] adopt classic frameworks such as Convolutional Neural Network and Multi-Layer Perceptron in combination with holistic or dense regression scheme for scale-invariant and rootrelative 3D hand joint positions recovery. After the introduction of the 3D parametric MANO [48] model, tremendous literature [3, 4, 9, 22, 45, 53, 60] attempts to integrate MANO into the learning pipeline and directly regress" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8970" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 68, + 285, + 225 + ], + "blocks": [ + { + "bbox": [ + 61, + 68, + 285, + 225 + ], + "lines": [ + { + "bbox": [ + 61, + 68, + 285, + 225 + ], + "spans": [ + { + "bbox": [ + 61, + 68, + 285, + 225 + ], + "type": "image", + "image_path": "5c5b97a25cfe87446c39d98302f4648cde69e9cbef49246741415c21b4f95b2b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 230, + 545, + 276 + ], + "lines": [ + { + "bbox": [ + 45, + 230, + 545, + 276 + ], + "spans": [ + { + "bbox": [ + 45, + 230, + 545, + 276 + ], + "type": "text", + "content": "Figure 2. Overview of Neural Voting Field (NVF): We first extract a feature map " + }, + { + "bbox": [ + 45, + 230, + 545, + 276 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 45, + 230, + 545, + 276 + ], + "type": "text", + "content": " from the input RGB image " + }, + { + "bbox": [ + 45, + 230, + 545, + 276 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 45, + 230, + 545, + 276 + ], + "type": "text", + "content": " by an image encoder. Then, for a 3D query point " + }, + { + "bbox": [ + 45, + 230, + 545, + 276 + ], + "type": "inline_equation", + "content": "\\pmb{p}" + }, + { + "bbox": [ + 45, + 230, + 545, + 276 + ], + "type": "text", + "content": " sampled in camera space, we define an implicit function " + }, + { + "bbox": [ + 45, + 230, + 545, + 276 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{NVF}}" + }, + { + "bbox": [ + 45, + 230, + 545, + 276 + ], + "type": "text", + "content": " realized by a MLP, which maps from pixel-aligned image feature " + }, + { + "bbox": [ + 45, + 230, + 545, + 276 + ], + "type": "inline_equation", + "content": "S(F, \\pi(\\pmb{p}))" + }, + { + "bbox": [ + 45, + 230, + 545, + 276 + ], + "type": "text", + "content": " and depth " + }, + { + "bbox": [ + 45, + 230, + 545, + 276 + ], + "type": "inline_equation", + "content": "Z(\\pmb{p})" + }, + { + "bbox": [ + 45, + 230, + 545, + 276 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 45, + 230, + 545, + 276 + ], + "type": "inline_equation", + "content": "\\pmb{p}" + }, + { + "bbox": [ + 45, + 230, + 545, + 276 + ], + "type": "text", + "content": " to its signed distance " + }, + { + "bbox": [ + 45, + 230, + 545, + 276 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 45, + 230, + 545, + 276 + ], + "type": "text", + "content": " and a set of 4D offset vectors " + }, + { + "bbox": [ + 45, + 230, + 545, + 276 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 45, + 230, + 545, + 276 + ], + "type": "text", + "content": ". Each 4D offset vector " + }, + { + "bbox": [ + 45, + 230, + 545, + 276 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 45, + 230, + 545, + 276 + ], + "type": "text", + "content": " contains a 1D voting weight and a 3D directional vector to each joint. The 3D hand pose is then obtained by dense 3D point-wise voting." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 279, + 78, + 523, + 219 + ], + "blocks": [ + { + "bbox": [ + 279, + 78, + 523, + 219 + ], + "lines": [ + { + "bbox": [ + 279, + 78, + 523, + 219 + ], + "spans": [ + { + "bbox": [ + 279, + 78, + 523, + 219 + ], + "type": "image", + "image_path": "ef3ea652f3cfe22d6df35f499c4a15b8e0e6c716de1d7b6ff2d444fb2d78e63a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 282, + 287, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 282, + 287, + 368 + ], + "spans": [ + { + "bbox": [ + 46, + 282, + 287, + 368 + ], + "type": "text", + "content": "MANO parameters from the image. Meanwhile, MANO model-free approaches [6] continue to develop and more works also start to investigate the dependencies among hand joints as hands are inherently structured. Thus, geometric deep learning-based models (e.g., Graph Convolution, Spiral Convolution) [5, 7, 8, 15, 32, 53] and Transformer-based models [19, 35, 36, 45] have also been proposed." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 372, + 197, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 372, + 197, + 384 + ], + "spans": [ + { + "bbox": [ + 47, + 372, + 197, + 384 + ], + "type": "text", + "content": "2.3. 3D Implicit Representation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 390, + 288, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 390, + 288, + 630 + ], + "spans": [ + { + "bbox": [ + 46, + 390, + 288, + 630 + ], + "type": "text", + "content": "Recent methods [1, 2, 10, 17, 37, 40, 46, 51] have shown that, as a continuous and differentiable function realized by Multi-Layer Perceptron, memory efficient deep implicit function can represent detailed 3D shapes with arbitrary topology and have no resolution limitations. It can also be extended to RGB image-based 3D reconstruction [40, 44, 46, 57]. Grasping Field [30] proposes to reconstruct hands grasping object via predicting signed distances for hand and object given the input point cloud or RGB image. Unlike previous implicit functions relying on a global feature vector, given an input RGB image with cleanly segmented foreground, Pixel-aligned Implicit Function [49] (PIFu) utilizes pixel-aligned image feature and depth value of each 3D point for occupancy prediction. PIFu-based methods have shown the ability to generate high-fidelity 3D human geometry with more clothes details [27, 31, 50, 62] and robust 6D pose estimation. For instance, Neural Correspondence Field [24] predicts signed distance and 3D object coordinate for each 3D point to establish dense 3D-3D correspondences for robust 6D object pose estimation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 639, + 176, + 652 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 639, + 176, + 652 + ], + "spans": [ + { + "bbox": [ + 47, + 639, + 176, + 652 + ], + "type": "text", + "content": "3. The Proposed Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 658, + 133, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 133, + 670 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 133, + 670 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "content": "Pixel-Aligned Implicit Function (PIFu) [49]. PIFu, by learning an implicit function in 3D domain with pixel-aligned features, generates detailed 3D human geometry in" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 282, + 545, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 282, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 304, + 282, + 545, + 426 + ], + "type": "text", + "content": "cluding largely occluded regions from a segmented RGB image. Specifically, given a segmented image " + }, + { + "bbox": [ + 304, + 282, + 545, + 426 + ], + "type": "inline_equation", + "content": "I:\\mathbb{R}^2\\mapsto \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 282, + 545, + 426 + ], + "type": "text", + "content": " PIFu first extracts a feature map " + }, + { + "bbox": [ + 304, + 282, + 545, + 426 + ], + "type": "inline_equation", + "content": "F:\\mathbb{R}^2\\mapsto \\mathbb{R}^C" + }, + { + "bbox": [ + 304, + 282, + 545, + 426 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 304, + 282, + 545, + 426 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 282, + 545, + 426 + ], + "type": "text", + "content": " channels by a fully convolutional image encoder. Then, for a 3D point " + }, + { + "bbox": [ + 304, + 282, + 545, + 426 + ], + "type": "inline_equation", + "content": "\\pmb {p}\\in \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 282, + 545, + 426 + ], + "type": "text", + "content": " in the camera space and its 2D projection " + }, + { + "bbox": [ + 304, + 282, + 545, + 426 + ], + "type": "inline_equation", + "content": "\\pi (\\pmb {p}):\\mathbb{R}^3\\mapsto \\mathbb{R}^2" + }, + { + "bbox": [ + 304, + 282, + 545, + 426 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 304, + 282, + 545, + 426 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 282, + 545, + 426 + ], + "type": "text", + "content": " , the implicit function, represented by a Multi-Layer Perceptron (MLP) " + }, + { + "bbox": [ + 304, + 282, + 545, + 426 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{PIF_u}}:\\mathbb{R}^C\\times \\mathbb{R}\\mapsto" + }, + { + "bbox": [ + 304, + 282, + 545, + 426 + ], + "type": "inline_equation", + "content": "\\mathbb{R}" + }, + { + "bbox": [ + 304, + 282, + 545, + 426 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 304, + 282, + 545, + 426 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{PIF_u}}(S(F,\\pi (\\pmb {p})),Z(\\pmb {p})) = y" + }, + { + "bbox": [ + 304, + 282, + 545, + 426 + ], + "type": "text", + "content": " maps the pixel-aligned feature " + }, + { + "bbox": [ + 304, + 282, + 545, + 426 + ], + "type": "inline_equation", + "content": "S(F,\\pi (\\pmb {p}))" + }, + { + "bbox": [ + 304, + 282, + 545, + 426 + ], + "type": "text", + "content": " and the depth " + }, + { + "bbox": [ + 304, + 282, + 545, + 426 + ], + "type": "inline_equation", + "content": "Z(\\pmb {p})" + }, + { + "bbox": [ + 304, + 282, + 545, + 426 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 304, + 282, + 545, + 426 + ], + "type": "inline_equation", + "content": "\\pmb{p}" + }, + { + "bbox": [ + 304, + 282, + 545, + 426 + ], + "type": "text", + "content": " to the occupancy " + }, + { + "bbox": [ + 304, + 282, + 545, + 426 + ], + "type": "inline_equation", + "content": "y\\in [0,1]" + }, + { + "bbox": [ + 304, + 282, + 545, + 426 + ], + "type": "text", + "content": " (1 means " + }, + { + "bbox": [ + 304, + 282, + 545, + 426 + ], + "type": "inline_equation", + "content": "\\pmb{p}" + }, + { + "bbox": [ + 304, + 282, + 545, + 426 + ], + "type": "text", + "content": " is inside the human surface and 0 means otherwise). " + }, + { + "bbox": [ + 304, + 282, + 545, + 426 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 282, + 545, + 426 + ], + "type": "text", + "content": " is the sampling function (e.g., bilinear interpolation) to sample the value of " + }, + { + "bbox": [ + 304, + 282, + 545, + 426 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 304, + 282, + 545, + 426 + ], + "type": "text", + "content": " at pixel " + }, + { + "bbox": [ + 304, + 282, + 545, + 426 + ], + "type": "inline_equation", + "content": "\\pi (\\pmb {p})" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 433, + 545, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 433, + 545, + 517 + ], + "spans": [ + { + "bbox": [ + 304, + 433, + 545, + 517 + ], + "type": "text", + "content": "Signed Distance Function (SDF) [12, 46]. Given a 3D shape " + }, + { + "bbox": [ + 304, + 433, + 545, + 517 + ], + "type": "inline_equation", + "content": "M \\in \\mathcal{M}" + }, + { + "bbox": [ + 304, + 433, + 545, + 517 + ], + "type": "text", + "content": ", a signed distance function is a continuous function " + }, + { + "bbox": [ + 304, + 433, + 545, + 517 + ], + "type": "inline_equation", + "content": "\\phi: \\mathbb{R}^3 \\times \\mathcal{M} \\mapsto \\mathbb{R}" + }, + { + "bbox": [ + 304, + 433, + 545, + 517 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 304, + 433, + 545, + 517 + ], + "type": "inline_equation", + "content": "\\phi(\\pmb{p}, M) = s" + }, + { + "bbox": [ + 304, + 433, + 545, + 517 + ], + "type": "text", + "content": ", mapping a 3D point " + }, + { + "bbox": [ + 304, + 433, + 545, + 517 + ], + "type": "inline_equation", + "content": "\\pmb{p}" + }, + { + "bbox": [ + 304, + 433, + 545, + 517 + ], + "type": "text", + "content": " to the signed distance " + }, + { + "bbox": [ + 304, + 433, + 545, + 517 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 433, + 545, + 517 + ], + "type": "text", + "content": " between " + }, + { + "bbox": [ + 304, + 433, + 545, + 517 + ], + "type": "inline_equation", + "content": "\\pmb{p}" + }, + { + "bbox": [ + 304, + 433, + 545, + 517 + ], + "type": "text", + "content": " and the shape surface. The sign denotes if " + }, + { + "bbox": [ + 304, + 433, + 545, + 517 + ], + "type": "inline_equation", + "content": "\\pmb{p}" + }, + { + "bbox": [ + 304, + 433, + 545, + 517 + ], + "type": "text", + "content": " is outside (positive) or inside (negative) of a watertight surface, with the iso-surface " + }, + { + "bbox": [ + 304, + 433, + 545, + 517 + ], + "type": "inline_equation", + "content": "\\phi(\\pmb{p}, M) = 0" + }, + { + "bbox": [ + 304, + 433, + 545, + 517 + ], + "type": "text", + "content": " implicitly representing the surface." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 525, + 542, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 525, + 542, + 538 + ], + "spans": [ + { + "bbox": [ + 306, + 525, + 542, + 538 + ], + "type": "text", + "content": "3.2. Dense Offset-based Pose Re-Parameterization" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "text", + "content": "As opposed to holistic regression directly mapping the input data to sparse and absolute joint positions and 2D dense regression which mainly exploits the 2D input local evidence, our method aims to directly model in 3D domain and thus fully exploit the 3D dense local evidence and hand global structure. To achieve this goal, given a common sparse and explicit pose representation (i.e., a set of 3D coordinates of " + }, + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "text", + "content": " hand joints) as " + }, + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "inline_equation", + "content": "J = \\{j_{t}\\}_{t=1}^{T} \\in \\mathcal{J}" + }, + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "inline_equation", + "content": "j_{t} \\in \\mathbb{R}^{3}" + }, + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "text", + "content": ", we parameterize it as a set of dense 3D point-wise offset-based representation. As shown in previous works [16,54], it is non-ideal to directly learn the 3D offset vector from each 3D point to a joint due to the large variance of offsets and far away joints are also beyond the scope of receptive field. Thus, given a camera-space 3D hand pose " + }, + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "text", + "content": " and" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "8971" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "text", + "content": "shape " + }, + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "text", + "content": ", we define the pose parameterization as a function " + }, + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "text", + "content": ", mapping from a 3D point " + }, + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "inline_equation", + "content": "\\pmb{p}" + }, + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "text", + "content": " in the camera space to a set of 4D offset vectors " + }, + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "inline_equation", + "content": "V = \\{\\pmb{v}_t\\}_{t=1}^T" + }, + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "inline_equation", + "content": "\\pmb{v}_t \\in \\mathbb{R}^4" + }, + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 116, + 287, + 129 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 116, + 287, + 129 + ], + "spans": [ + { + "bbox": [ + 61, + 116, + 287, + 129 + ], + "type": "interline_equation", + "content": "\\psi : \\mathbb {R} ^ {3} \\times \\mathcal {J} \\times \\mathcal {M} \\mapsto \\mathbb {R} ^ {T \\times 4} \\text {a s} \\psi (\\boldsymbol {p}, J, M) = V. \\tag {1}", + "image_path": "a13796218207fb07d04c485ce47307fef855d60a92b2afd203a95192e2109709.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 138, + 287, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 138, + 287, + 163 + ], + "spans": [ + { + "bbox": [ + 47, + 138, + 287, + 163 + ], + "type": "text", + "content": "Each offset vector " + }, + { + "bbox": [ + 47, + 138, + 287, + 163 + ], + "type": "inline_equation", + "content": "\\pmb{v}_t = (w_t, \\pmb{d}_t)" + }, + { + "bbox": [ + 47, + 138, + 287, + 163 + ], + "type": "text", + "content": " consists of a 1D voting weight " + }, + { + "bbox": [ + 47, + 138, + 287, + 163 + ], + "type": "inline_equation", + "content": "w_t \\in \\mathbb{R}" + }, + { + "bbox": [ + 47, + 138, + 287, + 163 + ], + "type": "text", + "content": " as point-wise closeness to hand joint " + }, + { + "bbox": [ + 47, + 138, + 287, + 163 + ], + "type": "inline_equation", + "content": "j_t" + }, + { + "bbox": [ + 47, + 138, + 287, + 163 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 170, + 304, + 208 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 170, + 304, + 208 + ], + "spans": [ + { + "bbox": [ + 47, + 170, + 304, + 208 + ], + "type": "interline_equation", + "content": "w _ {t} = \\left\\{ \\begin{array}{c c} 1 - \\frac {\\| \\boldsymbol {j} _ {t} - \\boldsymbol {p} \\| _ {2}}{r} & | s | < \\delta \\text {a n d} \\| \\boldsymbol {j} _ {t} - \\boldsymbol {p} \\| _ {2} \\leq r \\text {a n d} \\boldsymbol {p} \\in B _ {t} ^ {K}, \\\\ 0 & \\text {o t h e r w i s e}; \\end{array} \\right. \\tag {2}", + "image_path": "43fb717b28beec9e9c22cb87c13e3111c7c67acc2d04cf0e53eb3c89570f155f.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 209, + 287, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 209, + 287, + 233 + ], + "spans": [ + { + "bbox": [ + 47, + 209, + 287, + 233 + ], + "type": "text", + "content": "and a 3D unit directional vector " + }, + { + "bbox": [ + 47, + 209, + 287, + 233 + ], + "type": "inline_equation", + "content": "\\pmb{d}_t \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 47, + 209, + 287, + 233 + ], + "type": "text", + "content": " as point-wise direction to hand joint " + }, + { + "bbox": [ + 47, + 209, + 287, + 233 + ], + "type": "inline_equation", + "content": "j_t" + }, + { + "bbox": [ + 47, + 209, + 287, + 233 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 240, + 287, + 281 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 240, + 287, + 281 + ], + "spans": [ + { + "bbox": [ + 47, + 240, + 287, + 281 + ], + "type": "interline_equation", + "content": "\\boldsymbol {d} _ {t} = \\left\\{ \\begin{array}{c c} \\frac {\\boldsymbol {j} _ {t} - \\boldsymbol {p}}{\\| \\boldsymbol {j} _ {t} - \\boldsymbol {p} \\| _ {2}} & | s | < \\delta \\text {a n d} \\| \\boldsymbol {j} _ {t} - \\boldsymbol {p} \\| _ {2} \\leq r \\text {a n d} \\boldsymbol {p} \\in B _ {t} ^ {K}, \\\\ 0 & \\text {o t h e r w i s e}; \\end{array} \\right. \\tag {3}", + "image_path": "792e85a4f10c4ff5ec2a1c7e99697896ff5f3ad92bbef0c2b4215e50f6e80432.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 282, + 287, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 282, + 287, + 379 + ], + "spans": [ + { + "bbox": [ + 47, + 282, + 287, + 379 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 282, + 287, + 379 + ], + "type": "inline_equation", + "content": "s = \\phi(\\pmb{p}, M)" + }, + { + "bbox": [ + 47, + 282, + 287, + 379 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 282, + 287, + 379 + ], + "type": "inline_equation", + "content": "|s| < \\delta" + }, + { + "bbox": [ + 47, + 282, + 287, + 379 + ], + "type": "text", + "content": " means points that are within clamping distance " + }, + { + "bbox": [ + 47, + 282, + 287, + 379 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 47, + 282, + 287, + 379 + ], + "type": "text", + "content": " from the hand surface. " + }, + { + "bbox": [ + 47, + 282, + 287, + 379 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 47, + 282, + 287, + 379 + ], + "type": "text", + "content": " denotes the radius of a 3D ball centered at each joint and " + }, + { + "bbox": [ + 47, + 282, + 287, + 379 + ], + "type": "inline_equation", + "content": "B_t^K" + }, + { + "bbox": [ + 47, + 282, + 287, + 379 + ], + "type": "text", + "content": " denotes the set of " + }, + { + "bbox": [ + 47, + 282, + 287, + 379 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 47, + 282, + 287, + 379 + ], + "type": "text", + "content": "-Nearest Neighbors (KNN) to joint " + }, + { + "bbox": [ + 47, + 282, + 287, + 379 + ], + "type": "inline_equation", + "content": "\\pmb{j}_t" + }, + { + "bbox": [ + 47, + 282, + 287, + 379 + ], + "type": "text", + "content": " among all the near-surface points (i.e., points with " + }, + { + "bbox": [ + 47, + 282, + 287, + 379 + ], + "type": "inline_equation", + "content": "|s| < \\delta" + }, + { + "bbox": [ + 47, + 282, + 287, + 379 + ], + "type": "text", + "content": "). Thus, among all the 3D points that are in the hand surface vicinity, we assign the non-zero 4D offset vector " + }, + { + "bbox": [ + 47, + 282, + 287, + 379 + ], + "type": "inline_equation", + "content": "\\pmb{v}_t" + }, + { + "bbox": [ + 47, + 282, + 287, + 379 + ], + "type": "text", + "content": " to the " + }, + { + "bbox": [ + 47, + 282, + 287, + 379 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 47, + 282, + 287, + 379 + ], + "type": "text", + "content": " nearest points inside the 3D ball with radius " + }, + { + "bbox": [ + 47, + 282, + 287, + 379 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 47, + 282, + 287, + 379 + ], + "type": "text", + "content": " of joint " + }, + { + "bbox": [ + 47, + 282, + 287, + 379 + ], + "type": "inline_equation", + "content": "\\pmb{j}_t" + }, + { + "bbox": [ + 47, + 282, + 287, + 379 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 384, + 162, + 397 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 384, + 162, + 397 + ], + "spans": [ + { + "bbox": [ + 47, + 384, + 162, + 397 + ], + "type": "text", + "content": "3.3. Neural Voting Field" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 403, + 287, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 403, + 287, + 547 + ], + "spans": [ + { + "bbox": [ + 46, + 403, + 287, + 547 + ], + "type": "text", + "content": "To fully exploit the 3D local evidence and hand global geometry for camera-space 3D hand pose estimation, based on the aforementioned signed distance function and pose parameterization, we unify both modalities using a pixel-aligned implicit function. Following the proposed pipeline in Fig. 2, we first extract a " + }, + { + "bbox": [ + 46, + 403, + 287, + 547 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 403, + 287, + 547 + ], + "type": "text", + "content": "-channel feature map " + }, + { + "bbox": [ + 46, + 403, + 287, + 547 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 46, + 403, + 287, + 547 + ], + "type": "text", + "content": " from the input RGB image " + }, + { + "bbox": [ + 46, + 403, + 287, + 547 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 403, + 287, + 547 + ], + "type": "text", + "content": " by a hourglass network " + }, + { + "bbox": [ + 46, + 403, + 287, + 547 + ], + "type": "inline_equation", + "content": "g(I; \\boldsymbol{\\eta})" + }, + { + "bbox": [ + 46, + 403, + 287, + 547 + ], + "type": "text", + "content": " as used in [24, 49]. Then, for a 3D query point " + }, + { + "bbox": [ + 46, + 403, + 287, + 547 + ], + "type": "inline_equation", + "content": "\\pmb{p}" + }, + { + "bbox": [ + 46, + 403, + 287, + 547 + ], + "type": "text", + "content": " sampled in camera space, we define a continuous implicit function " + }, + { + "bbox": [ + 46, + 403, + 287, + 547 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{NVF}}" + }, + { + "bbox": [ + 46, + 403, + 287, + 547 + ], + "type": "text", + "content": " realized by a MLP, which maps from pixel-aligned image feature " + }, + { + "bbox": [ + 46, + 403, + 287, + 547 + ], + "type": "inline_equation", + "content": "S(F, \\pi(\\pmb{p}))" + }, + { + "bbox": [ + 46, + 403, + 287, + 547 + ], + "type": "text", + "content": " and depth " + }, + { + "bbox": [ + 46, + 403, + 287, + 547 + ], + "type": "inline_equation", + "content": "Z(\\pmb{p})" + }, + { + "bbox": [ + 46, + 403, + 287, + 547 + ], + "type": "text", + "content": " of point " + }, + { + "bbox": [ + 46, + 403, + 287, + 547 + ], + "type": "inline_equation", + "content": "\\pmb{p}" + }, + { + "bbox": [ + 46, + 403, + 287, + 547 + ], + "type": "text", + "content": " to its signed distance " + }, + { + "bbox": [ + 46, + 403, + 287, + 547 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 403, + 287, + 547 + ], + "type": "text", + "content": " and a set of 4D offset vectors " + }, + { + "bbox": [ + 46, + 403, + 287, + 547 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 46, + 403, + 287, + 547 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 89, + 555, + 287, + 573 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 555, + 287, + 573 + ], + "spans": [ + { + "bbox": [ + 89, + 555, + 287, + 573 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} f _ {\\mathrm {N V F}}: \\mathbb {R} ^ {C} \\times \\mathbb {R} \\mapsto \\mathbb {R} \\times \\mathbb {R} ^ {T \\times 4} \\quad \\text {a s} \\\\ f _ {\\mathrm {s}}: (C (E _ {\\mathrm {s}} (\\cdot)), Z (\\cdot), \\boldsymbol {a}) \\quad (\\cdot , V) \\end{array} \\tag {4}", + "image_path": "bab17bd93eaff8d4d40a78d9fde3aefc7836780c72fd7ee201aef9027b1ae3d3.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 90, + 568, + 244, + 581 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 568, + 244, + 581 + ], + "spans": [ + { + "bbox": [ + 90, + 568, + 244, + 581 + ], + "type": "interline_equation", + "content": "f _ {\\mathrm {N V F}} \\big (S (F, \\pi (\\boldsymbol {p})), Z (\\boldsymbol {p}); \\boldsymbol {\\theta} \\big) = (s, V),", + "image_path": "9f5622e13b3f0f5edf0cdc05f5f0a88512694b7da14ca5ccc37baf8eb33bff9d.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 589, + 287, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 589, + 287, + 685 + ], + "spans": [ + { + "bbox": [ + 46, + 589, + 287, + 685 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 589, + 287, + 685 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 46, + 589, + 287, + 685 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 589, + 287, + 685 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 589, + 287, + 685 + ], + "type": "text", + "content": " are the parameters of the hourglass " + }, + { + "bbox": [ + 46, + 589, + 287, + 685 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 46, + 589, + 287, + 685 + ], + "type": "text", + "content": " and MLP " + }, + { + "bbox": [ + 46, + 589, + 287, + 685 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{NVF}}" + }, + { + "bbox": [ + 46, + 589, + 287, + 685 + ], + "type": "text", + "content": ", respectively. We use a pinhole camera model for 3D-to-2D projection. For each 3D point, besides the set of 4D offset vectors, NVF also predicts its signed distance to the hand surface, helping to learn the global hand geometry and also to select near-surface 3D points by thresholding the signed distances. The near-surface 3D points are then considered as valid 3D voters to be used in the voting stage." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "content": "Optimization. During training, we sample " + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "content": " 3D points " + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "inline_equation", + "content": "P = \\{\\pmb{p}_n\\}_{n=1}^N" + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "content": " in camera frustum for each image and apply" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 72, + 545, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 132 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 132 + ], + "type": "text", + "content": "direct supervision to both predicted signed distance " + }, + { + "bbox": [ + 305, + 72, + 545, + 132 + ], + "type": "inline_equation", + "content": "s_n" + }, + { + "bbox": [ + 305, + 72, + 545, + 132 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 305, + 72, + 545, + 132 + ], + "type": "inline_equation", + "content": "L_s" + }, + { + "bbox": [ + 305, + 72, + 545, + 132 + ], + "type": "text", + "content": " and the predicted set of 4D offset vectors " + }, + { + "bbox": [ + 305, + 72, + 545, + 132 + ], + "type": "inline_equation", + "content": "V_n" + }, + { + "bbox": [ + 305, + 72, + 545, + 132 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 305, + 72, + 545, + 132 + ], + "type": "inline_equation", + "content": "L_V" + }, + { + "bbox": [ + 305, + 72, + 545, + 132 + ], + "type": "text", + "content": ". Given the ground-truth signed distance " + }, + { + "bbox": [ + 305, + 72, + 545, + 132 + ], + "type": "inline_equation", + "content": "\\hat{s}_n = \\phi(\\pmb{p}_n, \\hat{M})" + }, + { + "bbox": [ + 305, + 72, + 545, + 132 + ], + "type": "text", + "content": " and set of offset vectors " + }, + { + "bbox": [ + 305, + 72, + 545, + 132 + ], + "type": "inline_equation", + "content": "\\hat{V}_n = \\psi(\\pmb{p}_n, \\hat{J}, \\hat{M})" + }, + { + "bbox": [ + 305, + 72, + 545, + 132 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 305, + 72, + 545, + 132 + ], + "type": "inline_equation", + "content": "\\hat{J}" + }, + { + "bbox": [ + 305, + 72, + 545, + 132 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 72, + 545, + 132 + ], + "type": "inline_equation", + "content": "\\hat{M}" + }, + { + "bbox": [ + 305, + 72, + 545, + 132 + ], + "type": "text", + "content": " are the ground-truth pose and mesh, " + }, + { + "bbox": [ + 305, + 72, + 545, + 132 + ], + "type": "inline_equation", + "content": "L_s" + }, + { + "bbox": [ + 305, + 72, + 545, + 132 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 72, + 545, + 132 + ], + "type": "inline_equation", + "content": "L_V" + }, + { + "bbox": [ + 305, + 72, + 545, + 132 + ], + "type": "text", + "content": " are defined as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 333, + 133, + 545, + 165 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 133, + 545, + 165 + ], + "spans": [ + { + "bbox": [ + 333, + 133, + 545, + 165 + ], + "type": "interline_equation", + "content": "L _ {s} = \\frac {1}{N} \\sum_ {n = 1} ^ {N} | \\operatorname {c l a m p} (\\hat {s} _ {n}, \\delta) - \\operatorname {c l a m p} (s _ {n}, \\delta) |, \\tag {5}", + "image_path": "6274a1505964ecf55cbba7985d2f711b5aec730daf44298b57337b181359a537.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 331, + 167, + 545, + 200 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 167, + 545, + 200 + ], + "spans": [ + { + "bbox": [ + 331, + 167, + 545, + 200 + ], + "type": "interline_equation", + "content": "L _ {V} = \\frac {1}{N} \\sum_ {n = 1} ^ {N} \\mathbb {1} \\left(\\left| \\hat {s} _ {n} \\right| < \\delta\\right) H \\left(\\hat {V} _ {n}, V _ {n}\\right), \\tag {6}", + "image_path": "eacf2b9a016da3c60def06ae4bf2a77b97e68bea7f20a69df079a5efe1eee469.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 204, + 545, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 204, + 545, + 299 + ], + "spans": [ + { + "bbox": [ + 304, + 204, + 545, + 299 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 204, + 545, + 299 + ], + "type": "inline_equation", + "content": "\\mathrm{clamp}(s,\\delta)\\coloneqq \\min (\\delta ,\\max (-\\delta ,s))" + }, + { + "bbox": [ + 304, + 204, + 545, + 299 + ], + "type": "text", + "content": " uses " + }, + { + "bbox": [ + 304, + 204, + 545, + 299 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 304, + 204, + 545, + 299 + ], + "type": "text", + "content": " to control the distance from the hand surface over which we want to main a metric SDF, following [46]. " + }, + { + "bbox": [ + 304, + 204, + 545, + 299 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 304, + 204, + 545, + 299 + ], + "type": "text", + "content": " is the Huber loss [28] and " + }, + { + "bbox": [ + 304, + 204, + 545, + 299 + ], + "type": "inline_equation", + "content": "\\mathbb{1}(\\cdot)" + }, + { + "bbox": [ + 304, + 204, + 545, + 299 + ], + "type": "text", + "content": " is an indicator function that selects near-surface points (i.e., points that are within clamping distance " + }, + { + "bbox": [ + 304, + 204, + 545, + 299 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 304, + 204, + 545, + 299 + ], + "type": "text", + "content": " from the hand surface). Then, We jointly train the hourglass network " + }, + { + "bbox": [ + 304, + 204, + 545, + 299 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 304, + 204, + 545, + 299 + ], + "type": "text", + "content": " and MLP " + }, + { + "bbox": [ + 304, + 204, + 545, + 299 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{NVF}}" + }, + { + "bbox": [ + 304, + 204, + 545, + 299 + ], + "type": "text", + "content": " end-to-end by solving the following optimization problem with " + }, + { + "bbox": [ + 304, + 204, + 545, + 299 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 204, + 545, + 299 + ], + "type": "text", + "content": " as the balancing weight:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 364, + 300, + 545, + 320 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 300, + 545, + 320 + ], + "spans": [ + { + "bbox": [ + 364, + 300, + 545, + 320 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\eta} ^ {\\star}, \\boldsymbol {\\theta} ^ {\\star} = \\underset {\\boldsymbol {\\eta}, \\boldsymbol {\\theta}} {\\arg \\min } L _ {s} + \\lambda L _ {V}. \\tag {7}", + "image_path": "2100e01162ea7315844dfab7c10982bd8f51f47f563fca63bc131581c569066e.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 325, + 545, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 325, + 545, + 517 + ], + "spans": [ + { + "bbox": [ + 304, + 325, + 545, + 517 + ], + "type": "text", + "content": "Sampling of 3D Query Points. During training, unlike voxel-based methods which require discretization of ground-truth 3D meshes, we directly sample 3D points on the fly from the ground-truth mesh following [24, 49]. To make NVF fully exploit meaningful local evidence and understand the hand global geometry, it is crucial to focus on the hand by sampling the query points more densely around the hand surface during training. Thus, given the 3D camera-space hand model, we first sample 12500 points nearby the surface, 1000 points inside the bounding sphere of the model, and 1000 points inside the camera frustum. Among all these points, we then sample 2500 points inside the model and 2500 points outside to form a set of 5000 points in total for training. During inference, the points are directly sampled at centers of voxels that fill up the same camera frustum without any ground-truth information used." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 522, + 476, + 534 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 522, + 476, + 534 + ], + "spans": [ + { + "bbox": [ + 306, + 522, + 476, + 534 + ], + "type": "text", + "content": "3.4. Dense 3D Point-to-Joint Voting." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 540, + 545, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 540, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 304, + 540, + 545, + 624 + ], + "type": "text", + "content": "During inference, assuming " + }, + { + "bbox": [ + 304, + 540, + 545, + 624 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 540, + 545, + 624 + ], + "type": "text", + "content": " 3D points " + }, + { + "bbox": [ + 304, + 540, + 545, + 624 + ], + "type": "inline_equation", + "content": "P = \\{\\pmb{p}_n\\}_{n=1}^N" + }, + { + "bbox": [ + 304, + 540, + 545, + 624 + ], + "type": "text", + "content": " sampled in camera frustum for each image, for each point " + }, + { + "bbox": [ + 304, + 540, + 545, + 624 + ], + "type": "inline_equation", + "content": "\\pmb{p}_n" + }, + { + "bbox": [ + 304, + 540, + 545, + 624 + ], + "type": "text", + "content": " that is considered as a valid 3D voter (i.e., the predicted signed distance " + }, + { + "bbox": [ + 304, + 540, + 545, + 624 + ], + "type": "inline_equation", + "content": "s_n" + }, + { + "bbox": [ + 304, + 540, + 545, + 624 + ], + "type": "text", + "content": " is below the threshold " + }, + { + "bbox": [ + 304, + 540, + 545, + 624 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 304, + 540, + 545, + 624 + ], + "type": "text", + "content": "), based on its predicted set of 4D offset vectors " + }, + { + "bbox": [ + 304, + 540, + 545, + 624 + ], + "type": "inline_equation", + "content": "V_n = \\{\\pmb{v}_t^n\\}_{t=1}^T" + }, + { + "bbox": [ + 304, + 540, + 545, + 624 + ], + "type": "text", + "content": ", we first convert each 4D offset vector " + }, + { + "bbox": [ + 304, + 540, + 545, + 624 + ], + "type": "inline_equation", + "content": "\\pmb{v}_t^n = (w_t^n, \\pmb{d}_t^n)" + }, + { + "bbox": [ + 304, + 540, + 545, + 624 + ], + "type": "text", + "content": " into the actual 3D offset " + }, + { + "bbox": [ + 304, + 540, + 545, + 624 + ], + "type": "inline_equation", + "content": "\\pmb{o}_t^n" + }, + { + "bbox": [ + 304, + 540, + 545, + 624 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 304, + 540, + 545, + 624 + ], + "type": "inline_equation", + "content": "\\pmb{p}_n" + }, + { + "bbox": [ + 304, + 540, + 545, + 624 + ], + "type": "text", + "content": " to joint " + }, + { + "bbox": [ + 304, + 540, + 545, + 624 + ], + "type": "inline_equation", + "content": "\\pmb{j}_t" + }, + { + "bbox": [ + 304, + 540, + 545, + 624 + ], + "type": "text", + "content": " based on Eq. 2 and Eq. 3:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 354, + 628, + 545, + 641 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 628, + 545, + 641 + ], + "spans": [ + { + "bbox": [ + 354, + 628, + 545, + 641 + ], + "type": "interline_equation", + "content": "\\boldsymbol {o} _ {t} ^ {n} = \\mathbb {1} \\left(\\left| s _ {n} \\right| < \\delta\\right) \\left[ r \\left(1 - w _ {t} ^ {n}\\right) \\boldsymbol {d} _ {t} ^ {n} \\right]. \\tag {8}", + "image_path": "afe8363430c3f22b9af19a3fbb27352ee53de43d02f6e30c1aad963c9bac1225.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 645, + 545, + 680 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 645, + 545, + 680 + ], + "spans": [ + { + "bbox": [ + 304, + 645, + 545, + 680 + ], + "type": "text", + "content": "Then, each camera-space joint location " + }, + { + "bbox": [ + 304, + 645, + 545, + 680 + ], + "type": "inline_equation", + "content": "\\mathbf{j}_t" + }, + { + "bbox": [ + 304, + 645, + 545, + 680 + ], + "type": "text", + "content": " is calculated simply by a weighted average over each voter's 3D offset prediction:" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 348, + 683, + 545, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 683, + 545, + 716 + ], + "spans": [ + { + "bbox": [ + 348, + 683, + 545, + 716 + ], + "type": "interline_equation", + "content": "\\boldsymbol {j} _ {t} = \\sum_ {n = 1} ^ {N} \\frac {\\mathbb {1} \\left(\\left| s _ {n} \\right| < \\delta\\right) w _ {t} ^ {n} \\left(\\boldsymbol {o} _ {t} ^ {n} + \\boldsymbol {p} _ {n}\\right)}{\\sum_ {n = 1} ^ {N} \\mathbb {1} \\left(\\left| s _ {n} \\right| < \\delta\\right) w _ {t} ^ {n}}. \\tag {9}", + "image_path": "74ac22c146ed9e9f887e0e9031dea91ac158c62607fc4ff27c7854dcec77c47f.jpg" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8972" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 180 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 180 + ], + "type": "text", + "content": "Finally, we obtain the final camera-space 3D hand pose as " + }, + { + "bbox": [ + 47, + 72, + 289, + 180 + ], + "type": "inline_equation", + "content": "J = \\{j_{t}\\}_{t=1}^{T}" + }, + { + "bbox": [ + 47, + 72, + 289, + 180 + ], + "type": "text", + "content": ". Furthermore, instead of using all the near-surface points as valid voters, we can also only focus on using near-surface points with high predicted voting weight, as a higher voting weight means the point is more confident and closer to a hand joint. Therefore, we use voting fraction as a parameter to control the fraction of near-surface points with the highest voting weight for voting. Related ablation studies will be given in Sec. 4.7." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 191, + 129, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 191, + 129, + 205 + ], + "spans": [ + { + "bbox": [ + 47, + 191, + 129, + 205 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 211, + 153, + 223 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 211, + 153, + 223 + ], + "spans": [ + { + "bbox": [ + 47, + 211, + 153, + 223 + ], + "type": "text", + "content": "4.1. Baseline Methods" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 230, + 289, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 230, + 289, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 230, + 289, + 361 + ], + "type": "text", + "content": "To demonstrate the effectiveness of our proposed 3D dense regression scheme against classic holistic regression and 2D dense regression, we design two directly comparable baselines that share the same architecture of the hourglass network and the MLP as NVF described in Sec. 3. Specifically, similar to NVF, both baselines first extract " + }, + { + "bbox": [ + 46, + 230, + 289, + 361 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 230, + 289, + 361 + ], + "type": "text", + "content": "-channel feature map from the input RGB image " + }, + { + "bbox": [ + 46, + 230, + 289, + 361 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 230, + 289, + 361 + ], + "type": "text", + "content": ". Then, unlike our " + }, + { + "bbox": [ + 46, + 230, + 289, + 361 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{NVF}}" + }, + { + "bbox": [ + 46, + 230, + 289, + 361 + ], + "type": "text", + "content": " which takes a sampled 3D query point's pixel-aligned feature " + }, + { + "bbox": [ + 46, + 230, + 289, + 361 + ], + "type": "inline_equation", + "content": "S(F,\\pi (\\pmb {p}))" + }, + { + "bbox": [ + 46, + 230, + 289, + 361 + ], + "type": "text", + "content": " and its depth " + }, + { + "bbox": [ + 46, + 230, + 289, + 361 + ], + "type": "inline_equation", + "content": "Z(\\pmb {p})" + }, + { + "bbox": [ + 46, + 230, + 289, + 361 + ], + "type": "text", + "content": " to predict its signed distance " + }, + { + "bbox": [ + 46, + 230, + 289, + 361 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 230, + 289, + 361 + ], + "type": "text", + "content": " and the set of 4D offset vectors " + }, + { + "bbox": [ + 46, + 230, + 289, + 361 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 46, + 230, + 289, + 361 + ], + "type": "text", + "content": ", the two baselines take the following steps." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 366, + 289, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 366, + 289, + 475 + ], + "spans": [ + { + "bbox": [ + 47, + 366, + 289, + 475 + ], + "type": "text", + "content": "Baseline-Holistic. We directly apply a global average pooling " + }, + { + "bbox": [ + 47, + 366, + 289, + 475 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 47, + 366, + 289, + 475 + ], + "type": "text", + "content": " to its feature map " + }, + { + "bbox": [ + 47, + 366, + 289, + 475 + ], + "type": "inline_equation", + "content": "F^{\\prime}" + }, + { + "bbox": [ + 47, + 366, + 289, + 475 + ], + "type": "text", + "content": " extracted by the hourglass network " + }, + { + "bbox": [ + 47, + 366, + 289, + 475 + ], + "type": "inline_equation", + "content": "g(I;\\pmb {\\eta}^{\\prime})" + }, + { + "bbox": [ + 47, + 366, + 289, + 475 + ], + "type": "text", + "content": " and generate a " + }, + { + "bbox": [ + 47, + 366, + 289, + 475 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 47, + 366, + 289, + 475 + ], + "type": "text", + "content": " -dimension feature vector " + }, + { + "bbox": [ + 47, + 366, + 289, + 475 + ], + "type": "inline_equation", + "content": "\\alpha (F^{\\prime})" + }, + { + "bbox": [ + 47, + 366, + 289, + 475 + ], + "type": "text", + "content": " . The feature vector is then passed through the MLP " + }, + { + "bbox": [ + 47, + 366, + 289, + 475 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{HOL}}:\\mathbb{R}^{C}\\mapsto \\mathbb{R}^{T\\times 3}" + }, + { + "bbox": [ + 47, + 366, + 289, + 475 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 47, + 366, + 289, + 475 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{HOL}}(\\alpha (F^{\\prime});\\pmb {\\theta}^{\\prime}) = J^{\\prime}" + }, + { + "bbox": [ + 47, + 366, + 289, + 475 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{HOL}}" + }, + { + "bbox": [ + 47, + 366, + 289, + 475 + ], + "type": "text", + "content": " directly outputs " + }, + { + "bbox": [ + 47, + 366, + 289, + 475 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 47, + 366, + 289, + 475 + ], + "type": "text", + "content": " 3D hand joint coordinates " + }, + { + "bbox": [ + 47, + 366, + 289, + 475 + ], + "type": "inline_equation", + "content": "J^{\\prime} = \\{\\pmb{j}_{t}^{\\prime}\\}_{t = 1}^{T}" + }, + { + "bbox": [ + 47, + 366, + 289, + 475 + ], + "type": "text", + "content": " in camera space. Given the ground-truth pose " + }, + { + "bbox": [ + 47, + 366, + 289, + 475 + ], + "type": "inline_equation", + "content": "\\hat{J}" + }, + { + "bbox": [ + 47, + 366, + 289, + 475 + ], + "type": "text", + "content": " , it is trained by solving the following optimization problem with a loss function " + }, + { + "bbox": [ + 47, + 366, + 289, + 475 + ], + "type": "inline_equation", + "content": "L_{J^{\\prime}} = H(\\hat{J},J^{\\prime})" + }, + { + "bbox": [ + 47, + 366, + 289, + 475 + ], + "type": "text", + "content": " .." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 116, + 479, + 287, + 502 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 479, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 116, + 479, + 287, + 502 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\eta} ^ {\\prime \\star}, \\boldsymbol {\\theta} ^ {\\prime \\star} = \\underset {\\boldsymbol {\\eta} ^ {\\prime}, \\boldsymbol {\\theta} ^ {\\prime}} {\\arg \\min } L _ {J ^ {\\prime}}. \\tag {10}", + "image_path": "c47ecae746dcbce4ea4074c8a460c2005c789757b121ceb73e274cfdccb484d6.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 506, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 506, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 506, + 289, + 715 + ], + "type": "text", + "content": "Baseline-2D-Dense. We adopt a 2D dense regression scheme to estimate the camera-space pose via dense 2D pixel-wise voting, which is different from NVF which is based on dense 3D point-wise voting. Given the feature map " + }, + { + "bbox": [ + 46, + 506, + 289, + 715 + ], + "type": "inline_equation", + "content": "F''" + }, + { + "bbox": [ + 46, + 506, + 289, + 715 + ], + "type": "text", + "content": " extracted by the hourglass network " + }, + { + "bbox": [ + 46, + 506, + 289, + 715 + ], + "type": "inline_equation", + "content": "g(I; \\boldsymbol{\\eta}^{\\prime \\prime})" + }, + { + "bbox": [ + 46, + 506, + 289, + 715 + ], + "type": "text", + "content": ", we only pass the pixel-aligned feature " + }, + { + "bbox": [ + 46, + 506, + 289, + 715 + ], + "type": "inline_equation", + "content": "F''(\\boldsymbol{u})" + }, + { + "bbox": [ + 46, + 506, + 289, + 715 + ], + "type": "text", + "content": " at each 2D pixel location " + }, + { + "bbox": [ + 46, + 506, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\boldsymbol{u}" + }, + { + "bbox": [ + 46, + 506, + 289, + 715 + ], + "type": "text", + "content": " through the MLP " + }, + { + "bbox": [ + 46, + 506, + 289, + 715 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{DEN}}: \\mathbb{R}^C \\mapsto \\mathbb{R} \\times \\mathbb{R}^{T \\times 4}" + }, + { + "bbox": [ + 46, + 506, + 289, + 715 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 46, + 506, + 289, + 715 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{DEN}}(F''(\\boldsymbol{u}); \\boldsymbol{\\theta}'') = (e, V'')" + }, + { + "bbox": [ + 46, + 506, + 289, + 715 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 506, + 289, + 715 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{DEN}}" + }, + { + "bbox": [ + 46, + 506, + 289, + 715 + ], + "type": "text", + "content": " outputs the probability " + }, + { + "bbox": [ + 46, + 506, + 289, + 715 + ], + "type": "inline_equation", + "content": "e \\in [0,1]" + }, + { + "bbox": [ + 46, + 506, + 289, + 715 + ], + "type": "text", + "content": " that the hand is present at " + }, + { + "bbox": [ + 46, + 506, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\boldsymbol{u}" + }, + { + "bbox": [ + 46, + 506, + 289, + 715 + ], + "type": "text", + "content": " and also a set of 4D vectors " + }, + { + "bbox": [ + 46, + 506, + 289, + 715 + ], + "type": "inline_equation", + "content": "V'' = \\{\\boldsymbol{v}_t''\\}_{t=1}^T" + }, + { + "bbox": [ + 46, + 506, + 289, + 715 + ], + "type": "text", + "content": ". Each 4D vector " + }, + { + "bbox": [ + 46, + 506, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\boldsymbol{v}_t'' = (w_t'', j_t'')" + }, + { + "bbox": [ + 46, + 506, + 289, + 715 + ], + "type": "text", + "content": " contains a 1D voting weight " + }, + { + "bbox": [ + 46, + 506, + 289, + 715 + ], + "type": "inline_equation", + "content": "w_t''" + }, + { + "bbox": [ + 46, + 506, + 289, + 715 + ], + "type": "text", + "content": " and 3D joint coordinate " + }, + { + "bbox": [ + 46, + 506, + 289, + 715 + ], + "type": "inline_equation", + "content": "j_t''" + }, + { + "bbox": [ + 46, + 506, + 289, + 715 + ], + "type": "text", + "content": ". For the 1D voting weight, it is defined similarly to the one in NVF. We first use the camera ray for each pixel to find its first 3D intersection point with the 3D hand model in the camera space. Then, we directly follow Eq. 2 to calculate the pixel's voting weight as the closeness to each joint. It is trained by solving the following optimization" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 71, + 494, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 71, + 494, + 85 + ], + "spans": [ + { + "bbox": [ + 305, + 71, + 494, + 85 + ], + "type": "text", + "content": "tion problem with " + }, + { + "bbox": [ + 305, + 71, + 494, + 85 + ], + "type": "inline_equation", + "content": "\\lambda''" + }, + { + "bbox": [ + 305, + 71, + 494, + 85 + ], + "type": "text", + "content": " as the balancing weight:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 309, + 89, + 545, + 163 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 89, + 545, + 163 + ], + "spans": [ + { + "bbox": [ + 309, + 89, + 545, + 163 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\boldsymbol {\\eta} ^ {\\prime \\prime \\star}, \\boldsymbol {\\theta} ^ {\\prime \\prime \\star} = \\underset {\\boldsymbol {\\eta} ^ {\\prime \\prime}, \\boldsymbol {\\theta} ^ {\\prime \\prime}} {\\arg \\min } L _ {e} + \\lambda^ {\\prime \\prime} L _ {V ^ {\\prime \\prime}} (11) \\\\ = \\underset {\\boldsymbol {\\eta} ^ {\\prime \\prime}, \\boldsymbol {\\theta} ^ {\\prime \\prime}} {\\arg \\min } \\frac {1}{L} \\left[ \\sum_ {l = 1} ^ {L} E \\left(\\hat {e} _ {l}, e _ {l}\\right) + \\lambda^ {\\prime \\prime} \\sum_ {l = 1} ^ {L} \\hat {e} _ {l} H \\left(\\hat {V} _ {l} ^ {\\prime \\prime}, V _ {l} ^ {\\prime \\prime}\\right) \\right], (12) \\\\ \\end{array}", + "image_path": "40f69ef62733fa784bc5a8155a7cbe2eded0f49b08768ac878e75036327c0ff2.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 170, + 547, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 170, + 547, + 257 + ], + "spans": [ + { + "bbox": [ + 305, + 170, + 547, + 257 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 170, + 547, + 257 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 305, + 170, + 547, + 257 + ], + "type": "text", + "content": " is the number of pixels, " + }, + { + "bbox": [ + 305, + 170, + 547, + 257 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 305, + 170, + 547, + 257 + ], + "type": "text", + "content": " is the binary cross-entropy loss, " + }, + { + "bbox": [ + 305, + 170, + 547, + 257 + ], + "type": "inline_equation", + "content": "\\hat{e}_l" + }, + { + "bbox": [ + 305, + 170, + 547, + 257 + ], + "type": "text", + "content": " is given by the ground-truth hand mask, and " + }, + { + "bbox": [ + 305, + 170, + 547, + 257 + ], + "type": "inline_equation", + "content": "\\hat{V}_l^{\\prime \\prime}" + }, + { + "bbox": [ + 305, + 170, + 547, + 257 + ], + "type": "text", + "content": " is the ground-truth set of 4D vectors. During inference, for pixel " + }, + { + "bbox": [ + 305, + 170, + 547, + 257 + ], + "type": "inline_equation", + "content": "\\pmb{u}_l" + }, + { + "bbox": [ + 305, + 170, + 547, + 257 + ], + "type": "text", + "content": " with predicted probability " + }, + { + "bbox": [ + 305, + 170, + 547, + 257 + ], + "type": "inline_equation", + "content": "e_l > 0.5" + }, + { + "bbox": [ + 305, + 170, + 547, + 257 + ], + "type": "text", + "content": " (i.e., predicted as the foreground pixel), its predicted voting weight and 3D joint location are used to calculate the camera-space pose " + }, + { + "bbox": [ + 305, + 170, + 547, + 257 + ], + "type": "inline_equation", + "content": "J^{\\prime \\prime} = \\{\\dot{j}_t^{\\prime \\prime}\\}_{t = 1}^{T}" + }, + { + "bbox": [ + 305, + 170, + 547, + 257 + ], + "type": "text", + "content": " by a weighted average, similar to Eq. 9." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 261, + 477, + 273 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 261, + 477, + 273 + ], + "spans": [ + { + "bbox": [ + 306, + 261, + 477, + 273 + ], + "type": "text", + "content": "4.2. Dataset and Evaluation Metrics" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 280, + 545, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 280, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 304, + 280, + 545, + 352 + ], + "type": "text", + "content": "FreiHAND Dataset [64]. As a real 3D hand dataset, it provides 130,240 training and 3,960 evaluation samples. The result of the evaluation set is evaluated via the official server for online evaluation. Since FreiHAND provides hand scale during evaluation time, we provide an ablation on hand scale for analysis of 2D-to-3D scale and depth ambiguity." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 353, + 546, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 353, + 546, + 496 + ], + "spans": [ + { + "bbox": [ + 304, + 353, + 546, + 496 + ], + "type": "text", + "content": "Metrics. For FreiHAND, we evaluate for the task of camera-space 3D hand pose estimation. As in [8,64], we report the unaligned version of Mean Joint Error (mm) which is measured in camera space as CS-MJE and the Area-Under-the-Curve of PCK (Percentage of Correct Keypoints) vs. error thresholds as CS-AUC to evaluate the absolute 3D hand pose. We also define and report a 3D Translational Error (mm) (TE) [23] and Depth Error (mm) (DE) to evaluate the predicted and ground-truth pose's centroid and the depth-axis of the centroid, respectively. The results of TE and DE will be used in analysis for baseline studies, especially for the analysis of scale and depth ambiguity." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 497, + 545, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 497, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 497, + 545, + 544 + ], + "type": "text", + "content": "HO3D Dataset [18]. As a real 3D hand-object dataset, it contains 66,034 training samples and 11,524 evaluation samples. We use the official server for online evaluation as the annotations of the evaluation set are not available." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 544, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 544, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 544, + 545, + 628 + ], + "type": "text", + "content": "Metrics. For HO3D, we evaluate for the task of rootrelative 3D hand pose estimation. As in [18, 19], we report the Mean Joint Error (mm) after scale-translation alignment of the root as MJE and corresponding AUC to evaluate the pose. We also follow [45] to add ground-truth root to the predicted root-relative pose and report error (mm) as RS-MJE as it is equivalent to measuring in root-relative space." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 629, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 665 + ], + "type": "text", + "content": "Complement (Comp) Dataset [7]. As a synthetic 3D hand dataset, it provides 1,520 poses and 71 viewpoints for research use. We only use it for pre-training when needed." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 671, + 440, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 671, + 440, + 684 + ], + "spans": [ + { + "bbox": [ + 306, + 671, + 440, + 684 + ], + "type": "text", + "content": "4.3. Implementation Details" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "content": "For NVF and two baselines, the network architecture is adopted from PIFu [49]. Specifically, we use hourglass net" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8973" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 70, + 301, + 255 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 301, + 255 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 301, + 255 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 301, + 255 + ], + "type": "table", + "html": "
MethodExtra DataHand CropHand ScaleCS-MJE↓
ObMan [22]-✓✗85.2
MANO CNN [64]-✓✗71.3
I2L-MeshNet [42]-✓✗60.3
CMR-SG-RN18 [8]-✓✗49.7
CMR-SG-RN50 [8]-✓✗48.8
Baseline-Holisitc-✗✗54.5
Baseline-2D-Dense-✗✗53.2
CS-NVF (Ours)-✗✗47.2
Baseline-Holisitc-✗✓50.4
Baseline-2D-Dense-✗✓49.0
CS-NVF (Ours)-✗✓42.4
CS-NVF (Ours)Comp*✗✗44.6
CS-NVF (Ours)Comp*✗✓39.3
", + "image_path": "7461c6fc6319113b746718d4d7078b060c9ecf4f7552bbeaffb02f8a2837dad8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 300, + 289, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 300, + 289, + 553 + ], + "spans": [ + { + "bbox": [ + 46, + 300, + 289, + 553 + ], + "type": "text", + "content": "work with the output stride of 4 and output channel of 256 as the image encoder. All the MLPs " + }, + { + "bbox": [ + 46, + 300, + 289, + 553 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{NVF}}" + }, + { + "bbox": [ + 46, + 300, + 289, + 553 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 300, + 289, + 553 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{HOL}}" + }, + { + "bbox": [ + 46, + 300, + 289, + 553 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 300, + 289, + 553 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{DEN}}" + }, + { + "bbox": [ + 46, + 300, + 289, + 553 + ], + "type": "text", + "content": " have four hidden fully-connected layers with the number of neurons as (1024, 512, 256, 128). Unless stated otherwise, during training for NVF, the clamping distance " + }, + { + "bbox": [ + 46, + 300, + 289, + 553 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 46, + 300, + 289, + 553 + ], + "type": "text", + "content": " is set to 5 mm. The KNN is set to 1024. The ball radius " + }, + { + "bbox": [ + 46, + 300, + 289, + 553 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 46, + 300, + 289, + 553 + ], + "type": "text", + "content": " is set to 80 mm. During inference, the sampling step size of 3D points is 16 mm in all three axes. The voting fraction is set to 50%. Unlike PIFu, no image segmentation is applied. Camera intrinsic parameters are assumed known. Base on the above setting, for camera-space 3D hand pose estimation, we directly sample 3D points in camera space, as described in Sec. 3.3. We call NVF working in camera space as CS-NVF. Unlike other methods in Tab. 2, CS-NVF and both baselines take the original image as input without hand detection or cropping. For the task of root-relative 3D hand pose estimation, we follow similar sampling strategy but sample in a hand root-relative 3D cube instead of the camera frustum. We then call NVF working in root-relative space as RS-NVF. RS-NVF takes cropped hand-centered image. Please refer to the supplementary for more details." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 555, + 146, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 555, + 146, + 567 + ], + "spans": [ + { + "bbox": [ + 47, + 555, + 146, + 567 + ], + "type": "text", + "content": "4.4. Baseline Studies" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 569, + 287, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 569, + 287, + 641 + ], + "spans": [ + { + "bbox": [ + 46, + 569, + 287, + 641 + ], + "type": "text", + "content": "In this section, based on the results from CS-NVF and both baselines (as in Tab. 2, Tab. 3, Fig. 3, and Fig. 4), we investigate the impact of 2D-to-3D scale and depth ambiguity for the task of absolute 3D hand pose estimation and the effectiveness of using 3D dense regression over both 2D dense regression and holistic regression." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "type": "text", + "content": "2D-to-3D Scale and Depth Ambiguity. Estimating absolute 3D pose from an RGB image is an ill-posed problem due to 2D-to-3D scale and depth ambiguity [13, 29]. In view of this matter, FreiHAND [64] provides hand scale during evaluation for participants to use. As most existing works do not use the provided hand scale and thus, no" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 320, + 82, + 547, + 255 + ], + "blocks": [ + { + "bbox": [ + 47, + 258, + 287, + 293 + ], + "lines": [ + { + "bbox": [ + 47, + 258, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 47, + 258, + 287, + 293 + ], + "type": "text", + "content": "Table 2. Comparison for absolute 3D hand pose on FreiHAND.*: pre-training on Comp. Note that our CS-NVF and two baselines take the original image without hand detection or cropping." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 320, + 82, + 547, + 255 + ], + "lines": [ + { + "bbox": [ + 320, + 82, + 547, + 255 + ], + "spans": [ + { + "bbox": [ + 320, + 82, + 547, + 255 + ], + "type": "table", + "html": "
MethodExtra DataHand ScaleTE↓DE↓
Baseline-Holisitc-X50.649.1
Baseline-2D-Dense-X49.247.9
CS-NVF (Ours)-X43.642.4
Baseline-Holisitc-✓46.945.5
Baseline-2D-Dense-✓45.343.9
CS-NVF (Ours)-✓38.937.8
Baseline-HolisitcComp*X48.747.1
Baseline-2D-DenseComp*X47.946.4
CS-NVF (Ours)Comp*X41.540.4
Baseline-HolisitcComp*✓41.740.1
Baseline-2D-DenseComp*✓40.538.8
CS-NVF (Ours)Comp*✓36.535.5
", + "image_path": "02a2bfc56059a68fe8eb0364a80bc7bae21013a2f686f3ad1c99cb0cab405264.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 258, + 555, + 281 + ], + "lines": [ + { + "bbox": [ + 313, + 258, + 555, + 281 + ], + "spans": [ + { + "bbox": [ + 313, + 258, + 555, + 281 + ], + "type": "text", + "content": "Table 3. Comparison of 3D Translational and Depth Error for absolute 3D hand pose on FreiHAND. \\*: pre-training on Comp." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 307, + 299, + 545, + 476 + ], + "blocks": [ + { + "bbox": [ + 307, + 299, + 545, + 476 + ], + "lines": [ + { + "bbox": [ + 307, + 299, + 545, + 476 + ], + "spans": [ + { + "bbox": [ + 307, + 299, + 545, + 476 + ], + "type": "image", + "image_path": "a79c1edbd3877a69c57f267f417210b4558e36cde782abb28345014c76ec10c8.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 310, + 480, + 541, + 492 + ], + "lines": [ + { + "bbox": [ + 310, + 480, + 541, + 492 + ], + "spans": [ + { + "bbox": [ + 310, + 480, + 541, + 492 + ], + "type": "text", + "content": "Figure 3. 3D PCK for absolute 3D hand pose on FreiHAND." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 498, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 547, + 715 + ], + "type": "text", + "content": "relevant ablation on hand scale exists, we provide ablation results on hand scale in Tab. 2 and Tab. 3 for CS-NVF and two baselines. First, based on results in Tab. 2 and Tab. 3, we observe that, for all three pipelines, the main cause of the error between the ground-truth and predicted hand joints is the global translation and especially the depth component as TE is close to CS-MJE and DE is close to TE. For instance, CS-NVF with hand scale and Comp pre-training achieves " + }, + { + "bbox": [ + 304, + 498, + 547, + 715 + ], + "type": "inline_equation", + "content": "39.3\\mathrm{mm}" + }, + { + "bbox": [ + 304, + 498, + 547, + 715 + ], + "type": "text", + "content": " CS-MJE, while TE and DE are " + }, + { + "bbox": [ + 304, + 498, + 547, + 715 + ], + "type": "inline_equation", + "content": "36.5\\mathrm{mm}" + }, + { + "bbox": [ + 304, + 498, + 547, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 498, + 547, + 715 + ], + "type": "inline_equation", + "content": "35.5\\mathrm{mm}" + }, + { + "bbox": [ + 304, + 498, + 547, + 715 + ], + "type": "text", + "content": ", respectively. This validates the fact that depth ambiguity is indeed one of the main challenges for estimating accurate absolute 3D hand pose from a single RGB image. Moreover, for CS-NVF and the two baselines, using hand scale leads to significant improvements in terms of all three metrics. It is also worth noting that, for all three pipelines, results only using hand scale outperform results using only extra data. For instance, CS-NVF with hand scale achieves " + }, + { + "bbox": [ + 304, + 498, + 547, + 715 + ], + "type": "inline_equation", + "content": "42.4\\mathrm{mm}" + }, + { + "bbox": [ + 304, + 498, + 547, + 715 + ], + "type": "text", + "content": " CS-MJE, while CS-NVF with Comp pre-training" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8974" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 68, + 547, + 255 + ], + "blocks": [ + { + "bbox": [ + 50, + 68, + 547, + 255 + ], + "lines": [ + { + "bbox": [ + 50, + 68, + 547, + 255 + ], + "spans": [ + { + "bbox": [ + 50, + 68, + 547, + 255 + ], + "type": "image", + "image_path": "ee33d041aa02eea1f78ed4791a18e2ba076dcd420e9cf3c2e06bb11b389f1eb2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 257, + 547, + 302 + ], + "lines": [ + { + "bbox": [ + 46, + 257, + 547, + 302 + ], + "spans": [ + { + "bbox": [ + 46, + 257, + 547, + 302 + ], + "type": "text", + "content": "Figure 4. Qualitative results for absolute and relative 3D pose: (b) Predicted hand mask (Baseline) or mesh generated by Marching Cubes from the predicted signed distances (NVF). (c-h) Predicted 1D voting weight from each predicted foreground pixel (Baseline) or predicted near-surface 3D point (NVF) to a joint. (Brighter line means larger weight, darker or no line means smaller weight) (i) Estimated 3D pose. NVF makes solid point-wise prediction and overall distribution of signed distances and voting weights in highly occluded region." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 55, + 306, + 279, + 467 + ], + "blocks": [ + { + "bbox": [ + 55, + 306, + 279, + 467 + ], + "lines": [ + { + "bbox": [ + 55, + 306, + 279, + 467 + ], + "spans": [ + { + "bbox": [ + 55, + 306, + 279, + 467 + ], + "type": "table", + "html": "
MethodMJE↓AUC↑RS-MJE↓
Hasson et al. [20]36.90.369-
Pose2Mesh [11]33.30.48033.2
ObMan [22]31.80.46155.2
Liu et al. [38]31.70.46330.0
Hampali et al. [18]30.40.494-
METRO [35]28.90.504-
I2L-MeshNet [42]26.00.52926.8
Keypoint Trans. [19]25.70.553-
ArtiBoost [58]25.30.532-
Zheng et al. [61]25.10.541-
HandOccNet [45]24.00.55724.9
RS-NVF (Ours)21.80.61023.2
", + "image_path": "b5ff383930dd7e3582a71c01555fde7d5d3322d2f149490364b34fc52e5af2f0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 470, + 277, + 481 + ], + "lines": [ + { + "bbox": [ + 55, + 470, + 277, + 481 + ], + "spans": [ + { + "bbox": [ + 55, + 470, + 277, + 481 + ], + "type": "text", + "content": "Table 4. Comparison for relative 3D hand pose on HO3D." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 491, + 287, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 491, + 287, + 540 + ], + "spans": [ + { + "bbox": [ + 46, + 491, + 287, + 540 + ], + "type": "text", + "content": "has " + }, + { + "bbox": [ + 46, + 491, + 287, + 540 + ], + "type": "inline_equation", + "content": "44.6 \\mathrm{~mm}" + }, + { + "bbox": [ + 46, + 491, + 287, + 540 + ], + "type": "text", + "content": " CJ-MJE. These verify that a reasonable understanding towards the hand scale would help for robust depth learning and alleviate depth ambiguity, which is also aligned with existing literature [13, 29, 33]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 543, + 287, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 543, + 287, + 662 + ], + "spans": [ + { + "bbox": [ + 46, + 543, + 287, + 662 + ], + "type": "text", + "content": "Dense Regression vs. Holistic Regression. In Tab. 2 and Fig. 3, we can observe that both CS-NVF and Baseline-2D-Dense consistently outperform Baseline-Holistic on CS-MJE and CS-AUC under different training settings. For instance, without using hand scale, Baseline-2D-Dense is 1.3 mm better than Baseline-Holistic in CS-MJE, and CS-NVF can be 7.3 mm better. Since three pipelines share the same network architecture, this supports our point that dense regression is more effective than holistic regression for handling highly articulated pose data." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": "3D Dense Regression vs. 2D Dense Regression. Quantitatively, in Tab. 2 and Fig. 3, it can be observed that CS-NVF consistently outperforms Baseline-2D-Dense by a large margin on CS-MJE and CS-AUC under different" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 308, + 547, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 308, + 547, + 668 + ], + "spans": [ + { + "bbox": [ + 304, + 308, + 547, + 668 + ], + "type": "text", + "content": "training settings. For instance, without using hand scale, CS-NVF can be " + }, + { + "bbox": [ + 304, + 308, + 547, + 668 + ], + "type": "inline_equation", + "content": "6\\mathrm{mm}" + }, + { + "bbox": [ + 304, + 308, + 547, + 668 + ], + "type": "text", + "content": " better on CS-MJE. Qualitatively, as indicated by the white circles in column (c-h) of Fig. 4, both methods are able to find valid voters (Baseline-2D-Dense: pixels with predicted probability " + }, + { + "bbox": [ + 304, + 308, + 547, + 668 + ], + "type": "inline_equation", + "content": "e > 0.5" + }, + { + "bbox": [ + 304, + 308, + 547, + 668 + ], + "type": "text", + "content": ". CS-NVF: 3D points with predicted signed distance " + }, + { + "bbox": [ + 304, + 308, + 547, + 668 + ], + "type": "inline_equation", + "content": "|s| < \\delta" + }, + { + "bbox": [ + 304, + 308, + 547, + 668 + ], + "type": "text", + "content": "). However, CS-NVF generally provides not only more number of valid voters, but also more reasonable element-wise prediction and overall distribution of the voting weight than Baseline-2D-Dense. Specifically, for the number of voters, since CS-NVF models all the 3D points around the hand surface and Baseline-2D-Dense can only model foreground pixels, CS-NVF generally can find and use more valid voters. As for element-wise prediction and overall distribution of the voting weight, CS-NVF is generally able to predict large voting weight for points close to hand joint and small value for points far away, resulting in well-shaped symmetric distribution around the corresponding joint, even in occluded regions. Note that brighter line indicates larger voting weight, darker or no line indicates smaller voting weight. Moreover, as shown in column (b) for the mesh rendering of CS-NVF (obtained using Marching Cubes [39] from predicted signed distances), even in highly occluded region, CS-NVF provides solid signed distance distribution showing its ability to reason the global hand geometry. Overall, these demonstrate that CS-NVF, through direct dense modeling in 3D domain, can better model 3D dense local evidence, the relation between voters and 3D hand pose, and also the global hand geometry, compared with 2D dense regression-based counterpart, which leads to robust 3D hand pose estimation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 673, + 509, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 673, + 509, + 685 + ], + "spans": [ + { + "bbox": [ + 305, + 673, + 509, + 685 + ], + "type": "text", + "content": "4.5. Main Results: Absolute 3D Hand Pose" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 689, + 546, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 546, + 712 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 546, + 712 + ], + "type": "text", + "content": "In Tab. 2 and Fig. 3, we also compare CS-NVF with current state-of-the-art methods on FreiHAND for absolute 3D" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8975" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 69, + 548, + 145 + ], + "blocks": [ + { + "bbox": [ + 48, + 69, + 548, + 145 + ], + "lines": [ + { + "bbox": [ + 48, + 69, + 548, + 145 + ], + "spans": [ + { + "bbox": [ + 48, + 69, + 548, + 145 + ], + "type": "table", + "html": "
Proposed MethodEvaluation MetricVoting Fraction (%)Clamping Distance (mm)KNN
25507510015102025651210242048
CS-NVF (Ours)CS-MJE↓39.939.339.540.239.339.340.040.140.339.539.339.3
CS-AUC↑0.360.370.360.360.370.370.360.360.350.360.370.37
RS-NVF (Ours)MJE↓22.221.821.922.021.921.821.821.923.622.621.821.8
AUC↑0.600.610.610.600.600.610.610.610.590.600.610.61
", + "image_path": "6877abd3f39ecb9f95eecd3c285713a89eb1ded1e62d23e418d0ec0f5345b74d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 53, + 166, + 279, + 253 + ], + "blocks": [ + { + "bbox": [ + 48, + 149, + 544, + 161 + ], + "lines": [ + { + "bbox": [ + 48, + 149, + 544, + 161 + ], + "spans": [ + { + "bbox": [ + 48, + 149, + 544, + 161 + ], + "type": "text", + "content": "Table 5. Ablation study on parameter choices for absolute and relative 3D hand pose: Voting fraction, clamping distance, and KNN" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 53, + 166, + 279, + 253 + ], + "lines": [ + { + "bbox": [ + 53, + 166, + 279, + 253 + ], + "spans": [ + { + "bbox": [ + 53, + 166, + 279, + 253 + ], + "type": "table", + "html": "
Step SizeCS-NVF (Ours)
#PointsCS-MJE↓CS-AUC↑FPS↑
41882k39.20.371
8230k39.20.378
1628k39.30.3734
248k39.80.3650
323k42.40.3456
", + "image_path": "a4268fd0031444e54b8868a370150b169dd7bd01ee2c39554a953f41c7b9f364.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 257, + 287, + 280 + ], + "lines": [ + { + "bbox": [ + 46, + 257, + 287, + 280 + ], + "spans": [ + { + "bbox": [ + 46, + 257, + 287, + 280 + ], + "type": "text", + "content": "Table 6. Ablation study on sampling step size for absolute 3D hand pose on FreiHAND. #points: number of sampled 3D points." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 287, + 287, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 287, + 287, + 418 + ], + "spans": [ + { + "bbox": [ + 46, + 287, + 287, + 418 + ], + "type": "text", + "content": "hand pose estimation. In Tab. 2, without the use of hand scale and extra data as a fair comparison, CS-NVF surpasses previous state-of-the-art methods, achieving " + }, + { + "bbox": [ + 46, + 287, + 287, + 418 + ], + "type": "inline_equation", + "content": "1.6\\mathrm{mm}" + }, + { + "bbox": [ + 46, + 287, + 287, + 418 + ], + "type": "text", + "content": " better than CMR [8]. Note that, unlike other methods, CS-NVF takes the original image as input without hand detection or cropping. Additional hand scale information used without extra data improves the CS-MJE of CS-NVF further to " + }, + { + "bbox": [ + 46, + 287, + 287, + 418 + ], + "type": "inline_equation", + "content": "42.4\\mathrm{mm}" + }, + { + "bbox": [ + 46, + 287, + 287, + 418 + ], + "type": "text", + "content": ". Additional Comp pre-training improves the CS-MJE of CS-NVF to " + }, + { + "bbox": [ + 46, + 287, + 287, + 418 + ], + "type": "inline_equation", + "content": "44.6\\mathrm{mm}" + }, + { + "bbox": [ + 46, + 287, + 287, + 418 + ], + "type": "text", + "content": " without hand scale and " + }, + { + "bbox": [ + 46, + 287, + 287, + 418 + ], + "type": "inline_equation", + "content": "39.3\\mathrm{mm}" + }, + { + "bbox": [ + 46, + 287, + 287, + 418 + ], + "type": "text", + "content": " with hand scale. Referring to Fig. 3, CS-NVF also achieves state-of-the-art performance on CS-AUC." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 424, + 247, + 436 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 424, + 247, + 436 + ], + "spans": [ + { + "bbox": [ + 47, + 424, + 247, + 436 + ], + "type": "text", + "content": "4.6. Main Results: Relative 3D Hand Pose" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 440, + 287, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 440, + 287, + 513 + ], + "spans": [ + { + "bbox": [ + 46, + 440, + 287, + 513 + ], + "type": "text", + "content": "In Tab. 4, we compare RS-NVF with current state-of-the-art methods on HO3D for relative 3D hand pose estimation. RS-NVF significantly outperforms all these methods in all three metrics. Especially, RS-NVF outmatches HandOccNet [45] by " + }, + { + "bbox": [ + 46, + 440, + 287, + 513 + ], + "type": "inline_equation", + "content": "2.2\\mathrm{mm}" + }, + { + "bbox": [ + 46, + 440, + 287, + 513 + ], + "type": "text", + "content": " on MJE, " + }, + { + "bbox": [ + 46, + 440, + 287, + 513 + ], + "type": "inline_equation", + "content": "1.7\\mathrm{mm}" + }, + { + "bbox": [ + 46, + 440, + 287, + 513 + ], + "type": "text", + "content": " on RS-MJE, and also has 0.610 AUC against 0.557 AUC from HandOccNet." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 517, + 148, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 517, + 148, + 529 + ], + "spans": [ + { + "bbox": [ + 47, + 517, + 148, + 529 + ], + "type": "text", + "content": "4.7. Ablation Studies" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 534, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 288, + 713 + ], + "type": "text", + "content": "Sampling Step Size. During inference, 3D query points are sampled at centers of voxels that fill up the 3D space. Since the choices of different sampling step sizes can influence both accuracy and efficiency, in Tab. 6, we show the ablation on sampling step size for CS-NVF. With " + }, + { + "bbox": [ + 46, + 534, + 288, + 713 + ], + "type": "inline_equation", + "content": "16\\mathrm{mm}" + }, + { + "bbox": [ + 46, + 534, + 288, + 713 + ], + "type": "text", + "content": " as the step size, CS-NVF achieves 34 FPS and " + }, + { + "bbox": [ + 46, + 534, + 288, + 713 + ], + "type": "inline_equation", + "content": "39.3\\mathrm{mm}" + }, + { + "bbox": [ + 46, + 534, + 288, + 713 + ], + "type": "text", + "content": " on CS-MJE given 28k points. Reducing the step size to " + }, + { + "bbox": [ + 46, + 534, + 288, + 713 + ], + "type": "inline_equation", + "content": "4\\mathrm{mm}" + }, + { + "bbox": [ + 46, + 534, + 288, + 713 + ], + "type": "text", + "content": " yields 1,882k points and 1 FPS with slightly improved CS-MJE of " + }, + { + "bbox": [ + 46, + 534, + 288, + 713 + ], + "type": "inline_equation", + "content": "39.2\\mathrm{mm}" + }, + { + "bbox": [ + 46, + 534, + 288, + 713 + ], + "type": "text", + "content": ". Increasing to " + }, + { + "bbox": [ + 46, + 534, + 288, + 713 + ], + "type": "inline_equation", + "content": "32\\mathrm{mm}" + }, + { + "bbox": [ + 46, + 534, + 288, + 713 + ], + "type": "text", + "content": " step size yields 3k points and improves the FPS to 56 with still competitive " + }, + { + "bbox": [ + 46, + 534, + 288, + 713 + ], + "type": "inline_equation", + "content": "42.4\\mathrm{mm}" + }, + { + "bbox": [ + 46, + 534, + 288, + 713 + ], + "type": "text", + "content": " on CS-MJE. For RS-NVF not shown in the table, 4 and 8 mm step size both yield MJE of " + }, + { + "bbox": [ + 46, + 534, + 288, + 713 + ], + "type": "inline_equation", + "content": "21.7\\mathrm{mm}" + }, + { + "bbox": [ + 46, + 534, + 288, + 713 + ], + "type": "text", + "content": " and AUC of 0.62. We can observe that, while step size affects FPS, NVF is relatively insensitive regarding the performance. In all other reported results, we use " + }, + { + "bbox": [ + 46, + 534, + 288, + 713 + ], + "type": "inline_equation", + "content": "16\\mathrm{mm}" + }, + { + "bbox": [ + 46, + 534, + 288, + 713 + ], + "type": "text", + "content": " as the default step size." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 169, + 547, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 169, + 547, + 312 + ], + "spans": [ + { + "bbox": [ + 304, + 169, + 547, + 312 + ], + "type": "text", + "content": "Voting Fraction. Since the predicted point-wise voting weight reflects the confidence and closeness of a given point towards a certain joint, besides using all the near-surface points as valid voters, we can use a specified fraction of near-surface points with the highest voting weight for voting during inference. In Tab. 5, we observe that using all the near-surface points (100%) yields competitive results while 50% fraction leads to the overall better performance for both CS-NVF and RS-NVF. The results are aligned with our algorithm design since points far from each joint are not expected to make meaningful contributions. In all other reported results, we use 50% as the default voting fraction." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 312, + 547, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 312, + 547, + 396 + ], + "spans": [ + { + "bbox": [ + 304, + 312, + 547, + 396 + ], + "type": "text", + "content": "Clamping Distance. During training, clamping distance " + }, + { + "bbox": [ + 304, + 312, + 547, + 396 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 304, + 312, + 547, + 396 + ], + "type": "text", + "content": " is used to concentrate NVF's capacity on learning high-quality point-wise prediction (i.e., signed distance and set of 4D offset vectors) near the hand surface. Ablation in Tab. 5 shows that " + }, + { + "bbox": [ + 304, + 312, + 547, + 396 + ], + "type": "inline_equation", + "content": "5\\mathrm{mm}" + }, + { + "bbox": [ + 304, + 312, + 547, + 396 + ], + "type": "text", + "content": " generates the overall better performance for both CS-NVF and RS-NVF. In all other reported results, we use " + }, + { + "bbox": [ + 304, + 312, + 547, + 396 + ], + "type": "inline_equation", + "content": "5\\mathrm{mm}" + }, + { + "bbox": [ + 304, + 312, + 547, + 396 + ], + "type": "text", + "content": " as the default clamping distance." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 396, + 547, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 396, + 547, + 529 + ], + "spans": [ + { + "bbox": [ + 304, + 396, + 547, + 529 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 396, + 547, + 529 + ], + "type": "text", + "content": "-Nearest Neighbors (KNN). During training, among all the sampled points that are in the hand surface vicinity, nonzero 4D offset vector will only be assigned to the " + }, + { + "bbox": [ + 304, + 396, + 547, + 529 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 396, + 547, + 529 + ], + "type": "text", + "content": " nearest points inside a 3D ball of each hand joint. As shown in previous works [16, 54], estimating offsets for large KNN is unnecessary and far away joints are beyond the scope of receptive field. Small KNN might also cause poor performance due to the sparse learning scheme. Ablation in Tab. 5 shows that 1024 KNN generates the overall better performance for both CS-NVF and RS-NVF. In all other reported results, we use 1024 as the default KNN." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 538, + 384, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 538, + 384, + 550 + ], + "spans": [ + { + "bbox": [ + 306, + 538, + 384, + 550 + ], + "type": "text", + "content": "5. Conclusions" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "text", + "content": "We propose Neural Voting Field (NVF), as the first 3D implicit representation-based unified solution to estimate camera-space 3D hand pose given a single RGB image. NVF follows a novel unified 3D dense regression scheme to estimate 3D hand pose in 3D camera space via dense 3D point-wise voting in camera frustum. Studies with both holistic regression baseline and 2D dense regression baseline verify NVF's ability to fully exploit 3D dense local evidence and model global hand structure/geometry, which are essential for precisely locating 3D hand joints in 3D camera space. NVF also demonstrates state-of-the-art results on both tasks of camera-space 3D hand pose estimation and root-relative 3D hand pose estimation." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8976" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 286, + 712 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 53, + 91, + 286, + 113 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 286, + 113 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 286, + 113 + ], + "type": "text", + "content": "[1] Matan Atzmon and Yaron Lipman. SAL: Sign agnostic learning of shapes from raw data. In CVPR, 2020. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 114, + 286, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 114, + 286, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 114, + 286, + 135 + ], + "type": "text", + "content": "[2] Matan Atzmon and Yaron Lipman. SALD: Sign agnostic learning with derivatives. In ICLR, 2021. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 137, + 286, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 286, + 168 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 286, + 168 + ], + "type": "text", + "content": "[3] Seungryul Baek, Kwang In Kim, and Tae-Kyun Kim. Pushing the envelope for rgb-based dense 3d hand pose estimation via neural rendering. In CVPR, 2019. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 170, + 286, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 170, + 286, + 201 + ], + "spans": [ + { + "bbox": [ + 53, + 170, + 286, + 201 + ], + "type": "text", + "content": "[4] Adnane Boukhayma, Rodrigo de Bem, and Philip HS Torr. 3d hand shape and pose from images in the wild. In CVPR, 2019. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 203, + 286, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 203, + 286, + 247 + ], + "spans": [ + { + "bbox": [ + 53, + 203, + 286, + 247 + ], + "type": "text", + "content": "[5] Yujun Cai, Liuhao Ge, Jun Liu, Jianfei Cai, Tat-Jen Cham, Junsong Yuan, and Nadia Magnenat Thalmann. Exploiting spatial-temporal relationships for 3d pose estimation via graph convolutional networks. In ICCV, 2019. 1, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 248, + 286, + 291 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 248, + 286, + 291 + ], + "spans": [ + { + "bbox": [ + 53, + 248, + 286, + 291 + ], + "type": "text", + "content": "[6] Ping Chen, Yujin Chen, Dong Yang, Fangyin Wu, Qin Li, Qingpei Xia, and Yong Tan. I2uv-handnet: Image-to-uv prediction network for accurate and high-fidelity 3d hand mesh modeling. In ICCV, 2021. 1, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 293, + 286, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 293, + 286, + 335 + ], + "spans": [ + { + "bbox": [ + 53, + 293, + 286, + 335 + ], + "type": "text", + "content": "[7] Xingyu Chen, Yufeng Liu, Yajiao Dong, Xiong Zhang, Chongyang Ma, Yanmin Xiong, Yuan Zhang, and Xiaoyan Guo. Mobrecon: Mobile-friendly hand mesh reconstruction from monocular image. In CVPR, 2022. 1, 3, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 336, + 286, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 336, + 286, + 390 + ], + "spans": [ + { + "bbox": [ + 53, + 336, + 286, + 390 + ], + "type": "text", + "content": "[8] Xingyu Chen, Yufeng Liu, Chongyang Ma, Jianlong Chang, Huayan Wang, Tian Chen, Xiaoyan Guo, Pengfei Wan, and Wen Zheng. Camera-space hand mesh recovery via semantic aggregation and adaptive 2d-1d registration. In CVPR, 2021. 2, 3, 5, 6, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 392, + 286, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 392, + 286, + 434 + ], + "spans": [ + { + "bbox": [ + 53, + 392, + 286, + 434 + ], + "type": "text", + "content": "[9] Yujin Chen, Zhigang Tu, Di Kang, Linchao Bao, Ying Zhang, Xuefei Zhe, Ruizhi Chen, and Junsong Yuan. Model-based 3d hand reconstruction via self-supervised learning. In CVPR, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 436, + 286, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 436, + 286, + 457 + ], + "spans": [ + { + "bbox": [ + 48, + 436, + 286, + 457 + ], + "type": "text", + "content": "[10] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In CVPR, 2019. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 459, + 286, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 459, + 286, + 500 + ], + "spans": [ + { + "bbox": [ + 48, + 459, + 286, + 500 + ], + "type": "text", + "content": "[11] Hongsuk Choi, Gyeongsik Moon, and Kyoung Mu Lee. Pose2mesh: Graph convolutional network for 3d human pose and mesh recovery from a 2d human pose. In ECCV, 2020. 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 502, + 286, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 502, + 286, + 534 + ], + "spans": [ + { + "bbox": [ + 48, + 502, + 286, + 534 + ], + "type": "text", + "content": "[12] Brian Curless and Marc Levoy. A volumetric method for building complex models from range images. SIGGRAPH, 1996. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 536, + 286, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 286, + 557 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 286, + 557 + ], + "type": "text", + "content": "[13] Tom van Dijk and Guido de Croon. How do neural networks see depth in single images? In ICCV, 2019. 2, 6, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 559, + 286, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 286, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 286, + 601 + ], + "type": "text", + "content": "[14] Linpu Fang, Xingyan Liu, Li Liu, Hang Xu, and Wenxiong Kang. Jgr-p2o: Joint graph reasoning based pixel-to-offset prediction network for 3d hand pose estimation from a single depth image. In ECCV, 2020. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 603, + 286, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 603, + 286, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 603, + 286, + 644 + ], + "type": "text", + "content": "[15] Liuhao Ge, Zhou Ren, Yuncheng Li, Zehao Xue, Yingying Wang, Jianfei Cai, and Junsong Yuan. 3d hand shape and pose estimation from a single rgb image. In CVPR, 2019. 1, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 647, + 286, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 286, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 286, + 678 + ], + "type": "text", + "content": "[16] Liuhao Ge, Zhou Ren, and Junsong Yuan. Point-to-point regression pointnet for 3d hand pose estimation. In ECCV, 2018. 1, 3, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 681, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 286, + 712 + ], + "type": "text", + "content": "[17] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. In ICML, 2020. 3" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 37, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[18] Shreyas Hampali, Mahdi Rad, Markus Oberweger, and Vincent Lepetit. Honnotate: A method for 3d annotation of hand and object poses. In CVPR, 2020. 5, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 107, + 545, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 149 + ], + "type": "text", + "content": "[19] Shreyas Hampali, Sayan Deb Sarkar, Mahdi Rad, and Vincent Lepetit. Keypoint transformer: Solving joint identification in challenging hands and object interactions for accurate 3d pose estimation. In CVPR, 2022. 1, 3, 5, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 151, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 151, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 307, + 151, + 545, + 194 + ], + "type": "text", + "content": "[20] Yana Hasson, Bugra Tekin, Federica Bogo, Ivan Laptev, Marc Pollefeys, and Cordelia Schmid. Leveraging photometric consistency over time for sparsely supervised hand-object reconstruction. In CVPR, 2020. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 195, + 545, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 195, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 307, + 195, + 545, + 227 + ], + "type": "text", + "content": "[21] Yana Hasson, Gül Varol, Cordelia Schmid, and Ivan Laptev. Towards unconstrained joint hand-object reconstruction from rgb videos. In 3DV, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 228, + 545, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 228, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 307, + 228, + 545, + 270 + ], + "type": "text", + "content": "[22] Yana Hasson, Gül Varol, Dimitris Tzionas, Igor Kalevatykh, Michael J. Black, Ivan Laptev, and Cordelia Schmid. Learning joint reconstruction of hands and manipulated objects. In CVPR, 2019. 2, 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 272, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 272, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 307, + 272, + 545, + 293 + ], + "type": "text", + "content": "[23] Tomás Hodań, Jiří Matas, and Štepan Obdržálek. On evaluation of 6d object pose estimation. In ECCV, 2016. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 295, + 545, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 295, + 545, + 337 + ], + "spans": [ + { + "bbox": [ + 307, + 295, + 545, + 337 + ], + "type": "text", + "content": "[24] Lin Huang, Tomas Hodan, Lingni Ma, Linguang Zhang, Luan Tran, Christopher Twigg, Po-Chen Wu, Junsong Yuan, Cem Keskin, and Robert Wang. Neural correspondence field for object pose estimation. In ECCV, 2022. 2, 3, 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 338, + 545, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 338, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 307, + 338, + 545, + 370 + ], + "type": "text", + "content": "[25] Lin Huang, Jianchao Tan, Ji Liu, and Junsong Yuan. Handtransformer: non-autoregressive structured modeling for 3d hand pose estimation. In ECCV, 2020. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 372, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 372, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 307, + 372, + 545, + 403 + ], + "type": "text", + "content": "[26] Lin Huang, Boshen Zhang, Zhilin Guo, Yang Xiao, Zhiguo Cao, and Junsong Yuan. Survey on depth and rgb image-based 3d hand shape and pose estimation. VRIH, 2021. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 405, + 545, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 405, + 545, + 436 + ], + "spans": [ + { + "bbox": [ + 307, + 405, + 545, + 436 + ], + "type": "text", + "content": "[27] Zeng Huang, Yuanlu Xu, Christoph Lassner, Hao Li, and Tony Tung. Arch: Animatable reconstruction of clothed humans. In CVPR, 2020. 2, 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 437, + 545, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 437, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 307, + 437, + 545, + 459 + ], + "type": "text", + "content": "[28] Peter J Huber. Robust estimation of a location parameter. Breakthroughs in statistics, 1992. 4" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 460, + 545, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 460, + 545, + 492 + ], + "spans": [ + { + "bbox": [ + 307, + 460, + 545, + 492 + ], + "type": "text", + "content": "[29] Umar Iqbal, Pavlo Molchanov, Thomas Breuel Juergen Gall, and Jan Kautz. Hand pose estimation via latent 2.5d heatmap regression. In ECCV, 2018. 1, 2, 6, 7" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 494, + 545, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 494, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 307, + 494, + 545, + 536 + ], + "type": "text", + "content": "[30] Korrawe Karunratanakul, Jinlong Yang, Yan Zhang, Michael J. Black, Krikamol Muandet, and Siyu Tang. Grasping field: Learning implicit representations for human grasps. In 3DV, 2020. 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 538, + 545, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 538, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 307, + 538, + 545, + 569 + ], + "type": "text", + "content": "[31] Nilesh Kulkarni, Justin Johnson, and David F Fouhey. What's behind the couch? directed ray distance functions (drdf) for 3d scene reconstruction. In arXiv e-prints, 2021. 3" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 571, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 571, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 307, + 571, + 545, + 613 + ], + "type": "text", + "content": "[32] Dominik Kulon, Riza Alp Guler, Iasonas Kokkinos, Michael M Bronstein, and Stefanos Zafeiriou. Weakly-supervised mesh-convolutional hand reconstruction in the wild. In CVPR, 2020. 1, 3" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 307, + 614, + 545, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 614, + 545, + 646 + ], + "spans": [ + { + "bbox": [ + 307, + 614, + 545, + 646 + ], + "type": "text", + "content": "[33] Vincent Lepetit, Francesc Moreno-Noguer, and Pascal Fua. Epnp: An accurate o (n) solution to the pnp problem. IJCV, 2009. 2, 7" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 307, + 647, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 678 + ], + "type": "text", + "content": "[34] Shile Li and Dongheui Lee. Point-to-pose voting based hand pose estimation using residual permutation equivariant layer. In CVPR, 2019. 1" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 307, + 681, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 681, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 681, + 545, + 712 + ], + "type": "text", + "content": "[35] Kevin Lin, Lijuan Wang, and Zicheng Liu. End-to-end human pose and mesh reconstruction with transformers. In CVPR, 2021. 1, 3, 7" + } + ] + } + ], + "index": 36 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8977" + } + ] + } + ], + "index": 38 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 95 + ], + "type": "text", + "content": "[36] Kevin Lin, Lijuan Wang, and Zicheng Liu. Mesh graphormer. In ICCV, 2021. 1, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 96, + 287, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 96, + 287, + 129 + ], + "spans": [ + { + "bbox": [ + 49, + 96, + 287, + 129 + ], + "type": "text", + "content": "[37] Feng Liu, Luan Tran, and Xiaoming Liu. Fully understanding generic objects: Modeling, segmentation, and reconstruction. In CVPR, 2021. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 131, + 287, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 131, + 287, + 163 + ], + "spans": [ + { + "bbox": [ + 49, + 131, + 287, + 163 + ], + "type": "text", + "content": "[38] Shaowei Liu, Hanwen Jiang, Jiarui Xu, Sifei Liu, and Xiaolong Wang. Semi-supervised 3d hand-object poses estimation with interactions in time. In CVPR, 2021. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 166, + 287, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 166, + 287, + 198 + ], + "spans": [ + { + "bbox": [ + 49, + 166, + 287, + 198 + ], + "type": "text", + "content": "[39] William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3d surface construction algorithm. SIGGRAPH, 1987. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 200, + 287, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 200, + 287, + 243 + ], + "spans": [ + { + "bbox": [ + 49, + 200, + 287, + 243 + ], + "type": "text", + "content": "[40] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In CVPR, 2019. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 246, + 287, + 288 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 246, + 287, + 288 + ], + "spans": [ + { + "bbox": [ + 49, + 246, + 287, + 288 + ], + "type": "text", + "content": "[41] Gyeongsik Moon, Ju Yong Chang, and Kyoung Mu Lee. Camera distance-aware top-down approach for 3d multiperson pose estimation from a single rgb image. In ICCV, 2019. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 291, + 287, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 291, + 287, + 334 + ], + "spans": [ + { + "bbox": [ + 49, + 291, + 287, + 334 + ], + "type": "text", + "content": "[42] Gyeongsik Moon and Kyoung Mu Lee. I2l-meshnet: Image-to-lixel prediction network for accurate 3d human pose and mesh estimation from a single rgb image. In ECCV, 2020. 1, 2, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 337, + 287, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 337, + 287, + 368 + ], + "spans": [ + { + "bbox": [ + 49, + 337, + 287, + 368 + ], + "type": "text", + "content": "[43] Alejandro Newell, Kaiyu Yang, and Jia Deng. Stacked hourglass networks for human pose estimation. In ECCV, 2016. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 371, + 287, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 371, + 287, + 404 + ], + "spans": [ + { + "bbox": [ + 49, + 371, + 287, + 404 + ], + "type": "text", + "content": "[44] Michael Oechsle, Lars Mescheder, Michael Niemeyer, Thilo Strauss, and Andreas Geiger. Texture fields: Learning texture representations in function space. In ICCV, 2019. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 406, + 287, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 406, + 287, + 449 + ], + "spans": [ + { + "bbox": [ + 49, + 406, + 287, + 449 + ], + "type": "text", + "content": "[45] JoonKyu Park, Yeonguk Oh, Gyeongsik Moon, Hongsuk Choi, and Kyoung Mu Lee. Handocctnet: Occlusion-robust 3d hand mesh estimation network. In CVPR, 2022. 1, 2, 3, 5, 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 452, + 287, + 494 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 452, + 287, + 494 + ], + "spans": [ + { + "bbox": [ + 49, + 452, + 287, + 494 + ], + "type": "text", + "content": "[46] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In CVPR, 2019. 3, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 497, + 287, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 497, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 49, + 497, + 287, + 529 + ], + "type": "text", + "content": "[47] Georgios Pavlakos, Xiaowei Zhou, Konstantinos G Derpanis, and Kostas Daniilidis. Coarse-to-fine volumetric prediction for single-image 3d human pose. In CVPR, 2017. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 532, + 287, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 532, + 287, + 563 + ], + "spans": [ + { + "bbox": [ + 49, + 532, + 287, + 563 + ], + "type": "text", + "content": "[48] Javier Romero, Dimitrios Tzionas, and Michael J Black. Embodied hands: Modeling and capturing hands and bodies together. TOG, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 567, + 287, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 567, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 49, + 567, + 287, + 609 + ], + "type": "text", + "content": "[49] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In ICCV, 2019. 1, 2, 3, 4, 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 49, + 612, + 287, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 612, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 49, + 612, + 287, + 643 + ], + "type": "text", + "content": "[50] Shunsuke Saito, Tomas Simon, Jason Saragih, and Hanbyul Joo. Pifuhd: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization. In CVPR, 2020. 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 49, + 646, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 646, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 49, + 646, + 287, + 678 + ], + "type": "text", + "content": "[51] Vincent Sitzmann, Eric Chan, Richard Tucker, Noah Snavely, and Gordon Wetzstein. Metasdf: Meta-learning signed distance functions. In NeurIPS, 2020. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 49, + 681, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 681, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 681, + 287, + 712 + ], + "type": "text", + "content": "[52] Adrian Spurr, Jie Song, Seonwook Park, and Otmar Hilliges. Cross-modal deep variational hand pose estimation. In CVPR, 2018. 1, 2" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 544 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[53] Tze Ho Elden Tse, Kwang In Kim, Ales Leonardis, and Hyung Jin Chang. Collaborative learning for hand and object reconstruction with attention-guided graph convolution. In CVPR, 2022. 1, 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "text", + "content": "[54] Chengde Wan, Thomas Probst, Luc Van Gool, and Angela Yao. Dense 3d regression for hand pose estimation. In CVPR, 2018. 1, 3, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 152, + 545, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 174 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 174 + ], + "type": "text", + "content": "[55] Shih-En Wei, Varun Ramakrishna, Takeo Kanade, and Yaser Sheikh. Convolutional pose machines. In CVPR, 2016. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 175, + 545, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 218 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 218 + ], + "type": "text", + "content": "[56] Fu Xiong, Boshen Zhang, Yang Xiao, Zhiguo Cao, Taidong Yu, Joey Tianyi Zhou, and Junsong Yuan. A2j: Anchor-to-joint regression network for 3d articulated pose estimation from a single depth image. In ICCV, 2019. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 220, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 545, + 262 + ], + "type": "text", + "content": "[57] Qiangeng Xu, Weiyue Wang, Duygu Ceylan, Radomir Mech, and Ulrich Neumann. Disn: Deep implicit surface network for high-quality single-view 3d reconstruction. In NeurIPS, 2019. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 264, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 264, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 264, + 545, + 308 + ], + "type": "text", + "content": "[58] Lixin Yang, Kailin Li, Xinyu Zhan, Jun Lv, Wenqiang Xu, Jiefeng Li, and Cewu Lu. Artibost: Boosting articulated 3d hand-object pose estimation via online exploration and synthesis. In CVPR, 2022. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 309, + 545, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 309, + 545, + 341 + ], + "spans": [ + { + "bbox": [ + 307, + 309, + 545, + 341 + ], + "type": "text", + "content": "[59] Linlin Yang, Shile Li, Dongheui Lee, and Angela Yao. Aligning latent spaces for 3d hand pose estimation. In ICCV, 2019. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 343, + 545, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 343, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 307, + 343, + 545, + 376 + ], + "type": "text", + "content": "[60] Xiong Zhang, Qiang Li, Hong Mo, Wenbo Zhang, and Wen Zheng. End-to-end hand mesh recovery from a monocular rgb image. In ICCV, 2019. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 377, + 545, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 377, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 307, + 377, + 545, + 420 + ], + "type": "text", + "content": "[61] Xiaozheng Zheng, Pengfei Ren, Haifeng Sun, Jingyu Wang, Qi Qi, and Jianxin Liao. Joint-aware regression: Rethinking regression-based method for 3d hand pose estimation. In BMVC, 2021. 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 422, + 545, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 422, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 307, + 422, + 545, + 464 + ], + "type": "text", + "content": "[62] Zerong Zheng, Tao Yu, Yebin Liu, and Qionghai Dai. Pamir: Parametric model-conditioned implicit representation for image-based human reconstruction. TPAMI, 2021. 2, 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 467, + 545, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 467, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 307, + 467, + 545, + 498 + ], + "type": "text", + "content": "[63] Christian Zimmermann and Thomas Brox. Learning to estimate 3d hand pose from single rgb images. In ICCV, 2017. 1, 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 501, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 501, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 307, + 501, + 545, + 544 + ], + "type": "text", + "content": "[64] Christian Zimmermann, Duygu Ceylan, Jimei Yang, Bryan Russell, Max Argus, and Thomas Brox. Freihand: A dataset for markerless capture of hand pose and shape from single rgb images. In ICCV, 2019. 5, 6" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "8978" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/NeuralDome_ A Neural Modeling Pipeline on Multi-View Human-Object Interactions/fa0c20cd-9328-405b-b73e-2045fdfa714d_content_list.json b/2023/NeuralDome_ A Neural Modeling Pipeline on Multi-View Human-Object Interactions/fa0c20cd-9328-405b-b73e-2045fdfa714d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..2f8cb7020ddca8c32e536365e294d2eaa23f02b8 --- /dev/null +++ b/2023/NeuralDome_ A Neural Modeling Pipeline on Multi-View Human-Object Interactions/fa0c20cd-9328-405b-b73e-2045fdfa714d_content_list.json @@ -0,0 +1,1766 @@ +[ + { + "type": "text", + "text": "NeuralDome: A Neural Modeling Pipeline on Multi-View Human-Object Interactions", + "text_level": 1, + "bbox": [ + 117, + 128, + 852, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Juze Zhang $^{1,2,3,4,*}$ , Haimin Luo $^{1,4,5,*}$ , Hongdi Yang $^{1,4}$ , Xinru Xu $^{1,4}$ , Qianyang Wu $^{1,4}$ , Ye Shi $^{1,4}$ , Jingyi Yu $^{1,4}$ , Lan Xu $^{1,4,\\dagger}$ , Jingya Wang $^{1,4,\\dagger}$", + "bbox": [ + 112, + 200, + 879, + 238 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "$^{1}$ ShanghaiTech University $^{2}$ Shanghai Advanced Research Institute, Chinese Academy of Sciences $^{3}$ University of Chinese Academy of Sciences", + "$^{4}$ Shanghai Engineering Research Center of Intelligent Vision and Imaging $^{5}$ LumiAni Technology $\\{\\mathrm{zhangjz,luohm,yanghd,xuxr2022,wuqy2022,shiye,yujingyi,xulan1,wangjingya}\\} @shanghaiitech.edu.cn$" + ], + "bbox": [ + 86, + 238, + 903, + 310 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a6ddac9037905908066efb62551cd861d33d9aae1ede8cb2fffc35ad11317161.jpg", + "image_caption": [ + "Figure 1. Our NeuralDome pipeline for processing multi-view video sequences on human object interactions. NeuralDome supports tracking, modeling, and rendering of individual human subjects and objects. To validate NeuralDome, we collect a large dataset HODome over a total of 71 million video frames across 76 viewpoints and process the datasets using NeuralDome for a variety of inference and neural modeling and rendering tasks." + ], + "image_footnote": [], + "bbox": [ + 112, + 383, + 857, + 570 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 656, + 313, + 672 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Humans constantly interact with objects in daily life tasks. Capturing such processes and subsequently conducting visual inferences from a fixed viewpoint suffers from occlusions, shape and texture ambiguities, motions, etc. To mitigate the problem, it is essential to build a training dataset that captures free-viewpoint interactions. We construct a dense multi-view dome to acquire a complex human object interaction dataset, named HODome, that consists of $\\sim 71M$ frames on 10 subjects interacting with 23 objects. To process the HODome dataset, we develop NeuralDome, a layer-wise neural processing pipeline tai", + "bbox": [ + 75, + 690, + 470, + 857 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "lored for multi-view video inputs to conduct accurate tracking, geometry reconstruction and free-view rendering, for both human subjects and objects. Extensive experiments on the HODome dataset demonstrate the effectiveness of NeuralDome on a variety of inference, modeling, and rendering tasks. Both the dataset and the NeuralDome tools will be disseminated to the community for further development, which can be found at https://juzezhang.github.io/NeuralDome", + "bbox": [ + 500, + 657, + 893, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 828, + 630, + 844 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A key task of computer vision is to understand how humans interact with the surrounding world, by faithfully capturing and subsequently reproducing the process via mod", + "bbox": [ + 500, + 856, + 892, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* These authors contributed equally.", + "bbox": [ + 76, + 875, + 272, + 888 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "†Corresponding author.", + "bbox": [ + 76, + 888, + 205, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "8834", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "eling and rendering. Successful solutions benefit broad applications ranging from sports training to vocational education, digital entertainment to tele-medicine.", + "bbox": [ + 75, + 90, + 468, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Early solutions [11, 12, 51] that reconstruct dynamic meshes with per-frame texture maps are time-consuming and vulnerable to occlusions or lack of textures. Recent advances in neural rendering [37, 62, 71] bring huge potential for human-centric modeling. Most notably, the variants of Neural Radiance Field (NeRF) [37] achieve compelling novel view synthesis, which can enable real-time rendering performance [38, 59, 67] even for dynamic scenes [45, 63, 77], and can be extended to the generative setting without per-scene training [23, 69, 80]. However, less attention is paid to the rich and diverse interactions between humans and objects, mainly due to the severe lack of dense-view human-object datasets. Actually, existing datasets of human-object interactions are mostly based on optical markers [61] or sparse RGB/RGBD sensors [6, 26], without sufficient appearance supervision for neural rendering tasks. As a result, the literature on neural human-object rendering [21,57] is surprisingly sparse, let alone further exploring the real-time or generative directions. Besides, existing neural techniques [52, 77] suffer from tedious training procedures due to the human-object occlusion, and hence infeasible for building a large-scale dataset. In a nutshell, despite the recent tremendous thriving of neural rendering, the lack of both a rich dataset and an efficient reconstruction scheme constitute barriers in human-object modeling.", + "bbox": [ + 75, + 138, + 472, + 516 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we present NeuralDome, a neural pipeline that takes multi-view dome capture as inputs and conducts accurate 3D modeling and photo-realistic rendering of complex human-object interaction. As shown in Fig. 1, NeuralDome exploits layer-wise neural modeling to produce rich and multi-modality outputs including the geometry of dynamic human, object shapes and tracked poses, as well as a free-view rendering of the sequence.", + "bbox": [ + 75, + 518, + 468, + 638 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Specifically, we first capture a novel human-object dome (HODome) dataset that consists of 274 human-object interacting sequences, covering 23 diverse 3D objects and 10 human subjects (5 males and 5 females) in various apparels. We record multi-view video sequences of natural interactions between the human subjects and the objects where each sequence is about 60s in length using a dome with 76 RGB cameras, resulting in 71 million video frames. We also provide an accurate pre-scanned 3D template for each object and utilize sparse optical markers to track individual objects throughout the sequences.", + "bbox": [ + 75, + 642, + 468, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To process the HODome dataset, we adopt an extended Neural Radiance Field (NeRF) pipeline. The brute-force adoption of off-the-shelf neural techniques such as Instant-NSR [79] and Instant-NGP [38], although effective, do not separate objects from human subjects and therefore lack sufficient fidelity to model their interactions. We instead", + "bbox": [ + 75, + 810, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "introduce a layer-wise neural processing pipeline. Specifically, we first perform a joint optimization based on the dense inputs for accurately tracking human motions using the parametric SMPL-X model [43] as well as localizing objects using template meshes. We then propose an efficient layer-wise neural rendering scheme where the humans and objects are formulated as a pose-embedded dynamic NeRF and a static NeRF with tracked 6-DoF rigid poses, respectively. Such a layer-wise representation effectively exploits temporal information and robustly tackles the occluded regions under interactions. We further introduce an object-aware ray sampling strategy to mitigate artifacts during layer-wise training, as well as template-aware geometry regularizers to enforce contact-aware deformations. Through weak segmentation supervision, we obtain the decoupled and occlusion-free appearances for both the humans and the objects at a high fidelity amenable for training the input multi-view inputs for a variety of tasks from monocular motion capture to free-view rendering from sparse multi-view inputs.", + "bbox": [ + 496, + 90, + 890, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To summarize, our main contributions include:", + "bbox": [ + 519, + 393, + 830, + 407 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce NeuralDome, a neural pipeline, to accurately track humans and objects, conduct layer-wise geometry reconstruction, and enable novel-view synthesis, from multi-view HOI video inputs.", + "- We collect a comprehensive dataset that we call HODome that will be disseminated to the community, with both raw data and the output modalities including separated geometry and rendering of individual objects and human subjects, their tracking results, free-view rendering results, etc.", + "- We demonstrate using the dataset to train networks for a variety of visual inference tasks with complex human object interactions." + ], + "bbox": [ + 517, + 419, + 890, + 625 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 500, + 640, + 648, + 656 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Neural Human Rendering", + "text_level": 1, + "bbox": [ + 500, + 665, + 736, + 681 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Various 3D data representations have been explored for neural human rendering, such as point-clouds [4, 41, 72], textured meshes [14, 28, 29, 54], and volumes [31, 32]. Implicit occupancy function-based methods [17, 19, 48, 49] can recover detailed 3D human geometry from sparse 2D images without faithful appearance synthesis. The recent NeRF [37] technique brings huge potential for 3D photorealistic view synthesis [9, 10, 30, 34, 36, 38, 39, 64, 69, 75] and geometry modeling [7, 33, 68]. Further explorations extend it to dynamic scenes [7, 35, 42, 46, 67, 77], especially for humans. Existing works equip NeRF with pose-embeddings [23, 27, 40, 45, 80], learnable skinning weights [25, 44, 70] and even generalization across individuals [23, 66, 80]. However, these works only focus on", + "bbox": [ + 496, + 688, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "8835", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/478840b07deb591a49c4581a19f0f57c9e0287d2fafd35f34f6dd703335dd2a0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Datasets#Cam-view# Frame(M)ResolutionFpsMarkerObj. Num.Human Annot.Novel viewsGeometryHuman TexturedNeural RepresentationObject Appearance
NTU [26]3~ 341920 × 108030×NA✓NA××××
PiGr [50]10.1960 × 5405×NA✓NA×××✓
GRAB [61]×NANANA✓51××××××
PROX [16]10.11920 × 108030✓NA✓××××✓
BEHAVE [6]40.152048 × 153630×20×××××✓
InterCap [18]60.071920 × 108030×10✓××××✓
Our76713840 × 216060✓23✓✓✓✓✓✓
", + "bbox": [ + 99, + 89, + 870, + 191 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 1. Dataset Comparisons. We compare our proposed HODome dataset with the existing publicly available human-object dataset. HODome has the largest scale of human-object interactions in terms of the number of frames (#Frame), camera view (#Cam-view), and modality. \"Obj. Num.\" represents the object number. \"Human Annot.\" represents annotation from professional annotators.", + "bbox": [ + 75, + 202, + 893, + 247 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e0899a0cd871ac212bc82fc31dd32ab46ad0a0d3b0462149c3af250f16a9e4f8.jpg", + "image_caption": [ + "Figure 2. HODome Modality. HODome features multiple modalities of data format and annotations, including a) Color image, b) Human-object mask (Homask), c) MoCap (SMPL-X parameters and located object template), d) Geometry, e) digital assets in neural representations, f) Novel view and g) Keypoints (25 for bodies and 42 for hands with human annotation)." + ], + "image_footnote": [], + "bbox": [ + 78, + 267, + 467, + 446 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "a single person. The most relevant works [53, 77] aim to model multi-person interactions, but they cannot handle more complex interactions due to the lack of large-scale human-object interactions dataset.", + "bbox": [ + 75, + 566, + 468, + 628 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Human-object Modeling", + "text_level": 1, + "bbox": [ + 76, + 635, + 302, + 652 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Only a few works [6, 16, 18, 61, 73, 78] consider to jointly model whole-body interactions. Early solutions [11, 12, 51] that reconstruct meshes with per-frame texture maps, which are vulnerable to occlusions and lack of textures. Recently, several works [6, 16, 18, 61] explore the relationship via several interaction constraints, such as contact map [6, 18, 61], spatial arrangement [78] and physically plausible constraint [74]. However, these methods only produce a relative arrangement. The most relevant works [21, 56, 58] aim to render human-object interactions with volumetric fusion [56], neural texturing blending [58] and volumetric rendering [21]. However, without sufficient appearance supervision, their quality is limited. Comparably, our layerwise neural representations can be converted to high-quality geometry and support separate free-view rendering for human or object.", + "bbox": [ + 75, + 657, + 468, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Human-centric Dataset", + "text_level": 1, + "bbox": [ + 500, + 270, + 718, + 286 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A variety of human-centric datasets have been developed for human-only capturing or rendering tasks. Early datasets combine multi-camera RGB video capture with synchronized ground-truth 3D skeletons [20, 22, 55], parametric model [24, 65], scanned mesh [15, 76], without human-object interaction. Recent datasets capture human-object interactions using optical markers [61] or sparse RGB sensors [6, 26], without sufficient appearance supervision. As a result, the literature on neural human-object rendering [21, 57] is surprisingly rare. High-end works [11, 13] use dense cameras for reconstruction and rendering of humans and objects through mesh reconstruction and motion tracking, but without texture and neural representations. To fill this gap, we propose NeuralDome for capturing and rendering human-object interactions, facilitating various generative human-object tasks.", + "bbox": [ + 496, + 294, + 892, + 537 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. HODome Dataset", + "text_level": 1, + "bbox": [ + 500, + 551, + 671, + 568 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We present the HODome dataset for capturing and rendering photo-realistic human-object interactions. It consists of 274 human-object interacting sequences, covering 23 diverse 3D objects and 10 subjects (5 males and 5 females) under various apparels. For each sequence, we record the naturally interacting scene for about 60s using 76 RGB cameras for dense-view at $3840 \\times 2160$ resolution and 60 frame-per-seconds (fps), resulting in roughly 71 million frames. HODome consists of rich labels covering different aspects of HOI capturing and rendering labels (see Fig.2). See Tab. 1 for comparison with other datasets.", + "bbox": [ + 496, + 577, + 890, + 743 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Data Capturing System", + "text_level": 1, + "bbox": [ + 500, + 755, + 718, + 772 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To construct the HODome dataset, we use 76 Z-CAM cinema cameras with sufficient appearance supervision for neural rendering tasks and 16 Optitrack MoCap cameras [1] for accurate human-object tracking tasks. The Z-CAM system and Optitrack system are synchronized to record the RGB and MoCap data together. We use a publicly available tool [2] to estimate the intrinsic camera parameters and extrinsic camera parameters.", + "bbox": [ + 496, + 779, + 892, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "8836", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/50ae51d5779015aa428705f59a1fa4f33feb545fe7b54a145b6f8f2f4479273f.jpg", + "image_caption": [ + "Figure 3. Overview of NeuralDome. Given the 76-view RGB stream as input, we first jointly track human skeletal motion and object rigid motion. Then with the tracked motion priors, we decouple the human-object interaction scenes via a layer-wise neural rendering scheme, to generate human/object renderings separately and corresponding acceptable segmentation maps. Blending with the high-fidelity captured views, we obtain the layer-wise neural representations in our HODome dataset." + ], + "image_footnote": [], + "bbox": [ + 99, + 88, + 877, + 253 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Dataset Modality", + "text_level": 1, + "bbox": [ + 76, + 345, + 246, + 363 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Human MoCap and Object 6D Pose. For appearance realism, we do not place markers on human actors. Thus we detect the human joints from 76 RGB images by running the whole-body Openpose [8] to perform markerless MoCap. And we treat each object as rigid and solve the rigid object pose estimation from markers using the Iterative Closest Point(ICP) algorithm [5, 81]. See Fig. 3 for the pipeline of our method.", + "bbox": [ + 75, + 371, + 468, + 491 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Human-Object Neural Representations. Under the dense-view setting, one can apply off-the-shell neural techniques, i.e., Instant-NSR [79] and Instant-NGP [38] to efficiently obtain per-frame geometry and novel-view synthesis of the whole human-object sequence, respectively. However, neither schemes separate object from human subject. We therefore propose a layer-wise neural scheme to separately recover neural representations of human and object, enabling both high-fidelity geometry reconstruction and free-view appearance rendering (Sec. 4.2).", + "bbox": [ + 75, + 494, + 468, + 645 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Data Annotation. Our dataset includes 23 objects with varying scales and interaction types. Each object was pre-scanned using off-the-shelf multi-view software packages [2]. Further, following the previous methods [6, 61], we annotated its pseudo contact label that computes from a threshold distance. To provide more accurate results for the quantitative benchmark, we annotate a separate quantitative subset as our test set with human-annotated segmentation and hand joints by the professional annotator. Our datasets will be available for research purposes.", + "bbox": [ + 75, + 647, + 470, + 800 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Neural Modeling on HODome", + "text_level": 1, + "bbox": [ + 76, + 814, + 354, + 830 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We introduce a neural pipeline to produce the rich digital assets from the dense-view input of each human-object interaction (HOI) sequence in HODome, including accurate tracking, high-quality geometry reconstruction and novel-", + "bbox": [ + 75, + 839, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "view rendering. As illustrated in Fig. 3, given 76-view videos, we first perform a joint optimization scheme to accurately capture both the human skeletal motions and object rigid motion (Sec. 4.1). Then, based on the tracked human-object motions and shape priors, we decouple the humans and objects in HOI scenarios via a layer-wise neural human-object rendering scheme and a corresponding HOI-aware optimization strategy (Sec. 4.2). Such digital assets of humans, objects and the entire HOI scenes can naturally enable further geometry and appearance analysis of HOI scenarios.", + "bbox": [ + 496, + 345, + 890, + 512 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Human-object Tracking", + "text_level": 1, + "bbox": [ + 498, + 523, + 723, + 540 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here we introduce our human-object tracking scheme with the aid of object markers.", + "bbox": [ + 498, + 547, + 890, + 578 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Tracking initialization. We initialize the human tracking process by fitting the SMPL-X model to the keypoints of each view. Note that we utilize the off-the-shell toolbox Easymocap [3] to get the initial per-frame parameters, i.e., pose $\\theta_{t}$ , shape $\\beta_{t}$ , facial expression $\\psi_{t}$ , and translation $\\gamma_{t}$ on each frame $t$ . As for the object, following GRAB [61], we regard each object as rigid and only estimate the rotation $R_{t} \\in \\mathbb{S0}(3)$ and rigid translation $T_{t} \\in \\mathbb{R}^{3}$ with respect to its pre-scanned template. Specifically, we conduct the rigid-ICP technique [5, 81] to compute the rigid transformation between the per-frame markers, which serves as the initial object pose.", + "bbox": [ + 496, + 580, + 892, + 762 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Joint optimization for human-object tracking. With human and object pose initialization, we further perform a joint optimization scheme to ensure correct human-object contact. We impose constraints to ensure plausible interactions as the following form:", + "bbox": [ + 496, + 763, + 893, + 839 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} E \\left(\\beta_ {t}, \\theta_ {t}, \\psi_ {t}, \\gamma_ {t}, R _ {t}, T _ {t}\\right) = E _ {\\text {s m p l}} + \\lambda_ {\\text {c o n t a c t}} E _ {\\text {c o n t a c t}} + \\\\ \\lambda_ {\\text {h o m a s k}} E _ {\\text {h o m a s k}} + \\lambda_ {\\text {m a k e r}} E _ {\\text {m a k e r}}, \\tag {1} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 516, + 851, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "8837", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8fc6c6ae5126ab084ccaa7a81b4d86b3a38a6acf3098c59b56958722e75a476a.jpg", + "image_caption": [ + "Figure 4. Layer-wise Neural Human-Object Rendering Details. We sample a ray uniformly in the bounding box of the performer and near the template for the object. We condition both the human and object layer with appearance codes $(l_o^c, l_h^c)$ . The dynamic human layer contains an additional pose embedding $(E_p)$ conditioned deformation field. The sample colors and densities are merged, sorted and accumulated into pixel colors." + ], + "image_footnote": [], + "bbox": [ + 76, + 88, + 472, + 303 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $E_{\\mathrm{spl}}$ is the multi-view data fitting term describing the $\\ell_2$ distance between estimated joints and detected joints, following SMPLify-X [43] and PROX [16]. Besides, we impose contact term $E_{\\mathrm{contact}}$ , human-object silhouette loss term $E_{\\mathrm{homask}}$ , and object marker align term $E_{\\mathrm{maker}}$ to ensure the plausibility of human-object interaction tracking. Due to page limitation, we have to defer more details of these terms in the Appendix.", + "bbox": [ + 75, + 440, + 468, + 561 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Layer-wise Neural Human-Object Rendering", + "text_level": 1, + "bbox": [ + 76, + 571, + 460, + 589 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here we introduce our layer-wise image generation pipeline via neural rendering, without tedious manual efforts for geometry separation.", + "bbox": [ + 75, + 595, + 468, + 641 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Layered neural human-object modeling. For ease of instances separation, we represent the human-object interaction (HOI) scenes as a continuous layered neural radiance field, following ST-NeRF [77], as shown in Fig. 4. To leverage the shape prior imposed by the captured parametric SMPL-X model, we adopt a pose embedded dynamic NeRF similar to HumanNeRF [80] as the human layer. We use the skeletal pose to bridge the live frames to canonical space and an additional deformation MLP to learn subtle non-rigid deformation. Rather than using the pixel-aligned image feature, we leverage latent codes for time-varying human appearance capture to get rid of input images. Opposite to the human layer, we model objects as rigid static radiance fields in canonical space. The live frames are transformed to canonical space via object poses to maintain a globally consistent density field. Similarly, we adopt appearance latent codes for the time-varying shadow during interactions.", + "bbox": [ + 75, + 643, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dynamic human-object volume rendering. We utilize the layered volume rendering technique described in ST-NeRF [77] to render our neural human-object scenes. For a camera ray intersecting with the $i_{th}$ entity at any times-tamp, we compute the ray segment as the depth of intersection points $d_f^j$ and $d_f^i$ . We evenly partition each segment into $N$ bins and sample one point uniformly from each bin:", + "bbox": [ + 496, + 90, + 890, + 196 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\np _ {j} ^ {i} \\sim \\mathcal {U} \\left[ d _ {n} ^ {i} + \\frac {j - 1}{N} \\left(d _ {f} ^ {i} - d _ {n} ^ {i}\\right), d _ {n} ^ {i} + \\frac {j}{N} \\left(d _ {f} ^ {i} - d _ {n} ^ {i}\\right) \\right], j \\in [ 1, 2, \\dots , N ], \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 496, + 215, + 893, + 258 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{U}$ means uniform distribution. We impose the shape prior encoded in the tracked object template, on the ray sampling scheme for efficient training. We compute the ray-object intersection and uniformly sample a few points in the narrow segment where the first intersection point lies. For human, we simply sample in the ray-bounding box intersection segment. The samples from different segments are then merged and sorted by depth into $M$ samples in total. We then compute a pixel's color by accumulating the radiance at sampled points:", + "bbox": [ + 496, + 258, + 890, + 410 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nC = \\sum_ {i = 1} ^ {M} T \\left(p _ {i}\\right) \\left[ \\left(1 - e ^ {- \\sigma_ {p _ {i}} \\delta_ {p _ {i}}}\\right) c _ {p _ {i}} \\right], \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 420, + 890, + 460 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\delta_{p_i}$ is the distance between adjacent points, $c_{p_i}$ and $\\sigma_{p_i}$ are color and density, and $T(p_i) = \\prod_j^{i - 1}e^{-\\sigma_{p_j}\\delta_{p_j}}$", + "bbox": [ + 496, + 467, + 890, + 501 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "HOI-aware training scheme. Here we introduce an effective optimization scheme to train our neural layer-wise HOI model. We first leverage a photometric loss:", + "bbox": [ + 496, + 501, + 890, + 546 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c} = \\sum_ {\\mathbf {r} \\in \\mathcal {R}} \\| C _ {\\mathbf {r}} - \\hat {C} _ {\\mathbf {r}} \\| _ {2} ^ {2}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 556, + 890, + 590 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{r}$ is a ray in the training ray set $\\mathcal{R}$ , $C_{\\mathbf{r}}$ and $\\hat{C}_{\\mathbf{r}}$ are the corresponding rendered and observed colors.", + "bbox": [ + 496, + 602, + 890, + 633 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Note that in HOI scenarios, the human layer inevitably intersects with other object layers. This fact introduces ambiguities and cannot be addressed by a simple photometric loss. Thus we design three additional regularizers based on the tracked object template to alleviate this issue. To enforce the object layers to be solid surface radiance fields that do not contribute to human appearance, we adopt an occupancy and a sparsity regularizer inside and outside the object template, i.e.,", + "bbox": [ + 496, + 633, + 890, + 770 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {o} = \\sum_ {p \\in \\mathcal {O}} \\left(\\Omega_ {-} (p, \\mathcal {M}) \\| e ^ {(- \\sigma_ {p})} \\| _ {2} ^ {2} + \\Omega_ {+} (p, \\mathcal {M}) \\| \\sigma_ {p} \\| _ {2} ^ {2}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 780, + 890, + 814 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{M}$ is the template mesh, $\\mathcal{O}$ is a set of points randomly sampled in the object's bounding box, $\\Omega_{+}$ and $\\Omega_{-}$ indicate whether a point locates outside or inside the mesh.", + "bbox": [ + 496, + 824, + 890, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Recall that we have encoded human-object contact prior by jointly tracking human-object at the skeletal-pose level", + "bbox": [ + 496, + 869, + 890, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "8838", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d0913aeafdcba991c57124898a981e3dace7c18fb8fd34efe736909eabb7b8e0.jpg", + "image_caption": [ + "Figure 5. The geometry, MoCap and neural modeling results from our dataset. HODome includes various interaction sequences, such as \"holding a vase\", \"opening the cupboard\", \"sitting on the sofa\" and \"moving the table\"." + ], + "image_footnote": [], + "bbox": [ + 114, + 87, + 861, + 314 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "(Eq. 1). However, due to the misalignment of SMPL shape and actual human geometry, the human body and clothes may still overlap the object space, which goes against real-world physics. Therefore, we implicitly constrain the deformation net to predict contact-aware non-rigid deformation:", + "bbox": [ + 75, + 364, + 472, + 441 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {h} = \\sum_ {p \\in \\mathcal {H}} \\Omega_ {-} (p, \\mathcal {M}) \\| \\sigma_ {p} \\| _ {2} ^ {2}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 178, + 449, + 468, + 481 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathcal{H}$ is the set of random samples in the human's bounding box.", + "bbox": [ + 75, + 489, + 468, + 518 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We further explore a weakly-supervising scheme that utilizes object texture cues to help decouple the human and object entities. The key observation is that, with accurate object tracking, the object radiance field fused across frames can be rendered with higher confidence. That is, we train the entire scene first and render the objects only to capture views, and the pixels similar to the captured ones are labeled as object pixels. In this way, we obtain coarse object segmentation maps $S$ that serve as pseudo supervision: we adopt label-wise integration [53] to render the layer labels $s$ of rays:", + "bbox": [ + 75, + 520, + 473, + 685 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {s} = \\sum_ {i = 1} ^ {M} T \\left(p _ {i}\\right) \\left[ \\left(1 - e ^ {- \\sigma_ {p _ {i}} \\delta_ {p _ {i}}}\\right) \\mathbf {l} _ {p _ {i}} \\right], \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 685, + 468, + 724 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathbf{l}_{p_i}$ is one-hot label indicates which layer the point $p_i$ belongs to. We apply a semantic loss for object rays $S_o$ :", + "bbox": [ + 76, + 729, + 467, + 761 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s} = \\sum_ {\\mathbf {r} \\in \\mathcal {S} _ {o}} \\| \\mathbf {s} _ {\\mathbf {r}} - \\hat {\\mathbf {s}} _ {\\mathbf {r}} \\| _ {2} ^ {2}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 768, + 468, + 800 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\hat{\\mathbf{s}}_{\\mathbf{r}}$ is the pseudo semantic label of ray $\\mathbf{r}$ . Note that additional loss can be applied if accurate labels are provided.", + "bbox": [ + 75, + 808, + 468, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Layer-wise neural representation in HODome. To further use the data, we can render reliable segmentation maps and occlusion-free appearances for humans and objects respectively. However, the renderings tend to be blurred since", + "bbox": [ + 75, + 840, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the information of all the frames is fused. Hence we introduce an enhancement scheme to construct separate high-quality per-frame neural representation for human/object: we blend the visible regions of the high-fidelity observed input views to rendered human/object images using the segmentation maps. The Instant-NGP [38] and Instant-NSR [79] techniques are then applied to reconstruct separate human/object in tens of seconds. Finally, we obtain layer-wise representation which enables real-time photorealistic rendering and benefits the thorough analysis of HOI scenarios.", + "bbox": [ + 496, + 364, + 893, + 531 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementations. To build our neural representations, we train our layered models using Adam optimizer with a learning rate that starts from $5e^{-4}$ and decays exponentially. Using the 76-view 4K resolution videos, the training process takes 8-10 hours on a single NVIDIA 3090 GPU for acceptable appearance and segmentation rendering. Thanks to Instant-NGP [38] and Instant-NSR [79], the layer-wise neural modeling process is conducted in seconds per frame.", + "bbox": [ + 496, + 532, + 893, + 655 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 500, + 670, + 633, + 688 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we compare our neural pipeline with existing state-of-the-arts. We first demonstrate the effectiveness and the generalization ability of our neural modeling approach (5.1). Next, we implement NeuralDome on three specific tasks (5.2).", + "bbox": [ + 496, + 696, + 893, + 773 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Analysis of Neural Rendering", + "text_level": 1, + "bbox": [ + 500, + 784, + 764, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparisons. Here we compare our layer-wise neural human-object rendering approach, denoted as 'Ours-layer', with recent state-of-the-art methods, i.e., ST-NeRF [77], NeuralBody(NB) [45], in human-object interaction scenario. For fair comparisons, we select 20 cameras surrounding the performer for training and others for evaluation. We", + "bbox": [ + 496, + 809, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "8839", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/15fa79560dc22a3b92c6a65e9be6944aeacef22aec6a805f821c06047b8caaf6.jpg", + "image_caption": [ + "Figure 6. Qualitative comparisons of novel view synthesis results. We show ground truth and synthesized images of novel view for NeuralBody [45], and ST-NeRF [77] and our layered human-object representation. Our approach achieves the best performance. We further illustrate our high-fidelity neural representations in HODome in the last column." + ], + "image_footnote": [], + "bbox": [ + 81, + 85, + 472, + 324 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/48f0734af74b4dac3b0d5e285ab95b41e536ce2d64f3d2b5c784217dbafb95d1.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsNB [45]ST-NeRF [77]Ours
ScenesPSNR↑SSIM↑PSNR↑SSIM↑PSNR↑SSIM↑
Bigsofa19.330.89624.020.88632.280.958
Sofa26.730.96528.490.95835.620.987
Table21.940.93322.440.90027.880.949
Average22.670.93124.990.91531.930.964
", + "bbox": [ + 81, + 421, + 465, + 511 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "remove the background layer and train ST-NeRF without the semantic label as ours. Fig. 6 shows several appearance synthesis results. NeuralBody has no ability to model objects which leads to artifacts in the density field. ST-NeRF fails as the bounding boxes of the performer and object almost completely overlap. Our method achieves both human and object modeling and rendering and thus further enables generating high-fidelity neural representation denoted as 'Ours-HODome'. Tab. 2 further illustrates our method significantly outperforms the baselines.", + "bbox": [ + 75, + 566, + 468, + 717 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation. Here we further evaluate how our scheme contributes to the generated layer-wise neural representations. Fig. 7 shows the single human layer where the performer is \"sitting\" in the air. Note that we have no ground truth of the specific human/object layer, thus we conduct qualitative evaluations only. Let w/o pseudo segmentation and w/o blending denote our human assets generated without pseudo semantic loss and blending-based enhancement scheme. It demonstrates that the pseudo semantic loss effectively helps decouple the human and object layer and our blending scheme further boosts the appearance fidelity of the blurred temporal-fused model to a photo-realistic level.", + "bbox": [ + 75, + 719, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/446b1aead1a9ec1ebb02f849db33f917efe52cfa62691cd186a8d5f5b706234e.jpg", + "image_caption": [ + "Figure 7. Qualitative evaluation of each strategy's contribution to layer-wise neural representation generation. The weakly-supervising scheme effectively helps decouple human and object and our blending scheme generates high-quality digital assets." + ], + "image_footnote": [], + "bbox": [ + 501, + 87, + 883, + 332 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2f627f24f5f6fafeadd38b05f7b5fc451963b69356868c10218ec703c4a18981.jpg", + "table_caption": [ + "Table 2. Quantitative comparisons of novel view synthesized appearance on different human-object interaction sequences." + ], + "table_footnote": [], + "table_body": "
MethodSeparated evaluationJoint evaluation
MPJPE↓PA-MPJPE↓Chamferh↓Chamfero↓V2V↓p.V2V↓
Fit to input16.116.9486.6618.1461.5728.92
PHOSA [78]14.886.9477.6216.6454.5926.09
CHORE [73]10.216.1486.587.6944.9314.93
", + "bbox": [ + 503, + 401, + 892, + 460 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Human object capture benchmark. \"Fit to input\" represents the vanilla method that fits the object template to image and capture human with Frankmocap [47].", + "bbox": [ + 498, + 472, + 890, + 513 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Task and benchmark", + "text_level": 1, + "bbox": [ + 500, + 527, + 697, + 542 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Human-object Capture Benchmark. HODome provides multi-view sequences with synchronized ground truth of capturing. To demonstrate the capability, we provide a benchmark for human-object capture, shown in Tab. 3. For detailed metric explanation please refer to [6, 73, 74]. PHOSA [78] used the contact map from hand-crafted annotations, suffering from coarse contact information. While CHORE [73] outperforms it with the distance fields predicted from neural networks. It shows the importance of a prepared human-object dataset that supports data-driven human-object capture model. Check out the full capturing benchmark of the dataset at https://arxiv.org/abs/2212.07626.", + "bbox": [ + 496, + 551, + 892, + 747 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Human-object Geometry Reconstruction Benchmark. Benefiting from the dense-view setting, various human-only geometry reconstruction methods can step towards a human-object geometry reconstruction setting. We can benchmark the state-of-the-art algorithms in sparse view geometry reconstruction tasks. Hence, we evaluate PIFu [48] by training on a subset of HODome. As shown in Fig. 8, we compare the quality result between original PIFu and PIFu-trained on HODome and in-the-wild images. Compared to the original PIFu, our training set can improve the quality of", + "bbox": [ + 496, + 750, + 892, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "8840", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/74525579354cdace85ffdff903c9f173f7f0e827cea1481936cbb59b002f45ef.jpg", + "image_caption": [ + "NA" + ], + "image_footnote": [], + "bbox": [ + 80, + 89, + 142, + 170 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b63b870fe6f5b57c7d5848cec3cfa58168197534f846f10c099e04164648f771.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 147, + 90, + 205, + 170 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/da7ac4250173b8fbe463747dcf3901144873cfd38441ff40a1b78fbba28d1a23.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 210, + 90, + 281, + 170 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5f8cd77a43bfb8f83a5905cb23fe91454cd31d9a4f871c3fc89678f59a9ec847.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 284, + 90, + 362, + 170 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/891609b21faf132d8da293426b7344ea2c7c6629ec176f5f1ec63aab2a0247ba.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 367, + 90, + 444, + 170 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/624a47ec8b979c3074ab298f8e13364e51ac1952c655cf81a1da138f235135c1.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 80, + 175, + 143, + 253 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/16d70660e3a1c3a6b6fb9ddebfd579b9323b00d0b64ddd3d6b664e05ee9f8169.jpg", + "image_caption": [ + "PIFu-Origin" + ], + "image_footnote": [], + "bbox": [ + 236, + 170, + 299, + 255 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e724f1aece2df2ce899a9b73ea27ddf65efccc4f724ab6b09ac8444f4681b927.jpg", + "image_caption": [ + "Monocular PIFu-Trained", + "Figure 8. Qualitative evaluation on human-object reconstruction. We compare the original PIFu [48] with trained PIFu on both images from our dataset and in-the-wild images. The original PIFu fails to predict the object depth, and PIFu trained on our dataset can approximately predict the shape of the object." + ], + "image_footnote": [], + "bbox": [ + 313, + 166, + 370, + 255 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/de5e2daad2b958c849cb53506715a8e29d0e1cdb5296626ed52be20e94845e54.jpg", + "image_caption": [ + "Six View PIFu-Trained" + ], + "image_footnote": [], + "bbox": [ + 390, + 170, + 447, + 255 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/4ebb5f0a76ef37b8e8de9bfc114317a86ada32544d6585086e6df3db56777611.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodP2S ×10-4↓Chamfer ×10-4↓
Origin PIFu [48]38.72640.947
Monocular PIFu-trained14.65314.483
6-View PIFu-trained3.3764.901
", + "bbox": [ + 106, + 356, + 439, + 410 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/403a7a18ef53b879b8218d2668e84f17b8f170dcaffb82a5ec78f4a2bcafab05.jpg", + "table_caption": [ + "Table 4. Geometry reconstruction benchmark. Origin PIFu [48] uses the pre-trained model to infer. Monocular PIFu is trained on a single view. 6-View PIFu is trained on 6-view inputs." + ], + "table_footnote": [], + "table_body": "
MethodPSNR↑SSIM↑
IBRnet [69]21.430.892
Neuray [30]23.340.909
NeuralHumanFVV [60]21.690.914
NeuralHOIFVV23.100.912
", + "bbox": [ + 151, + 470, + 393, + 536 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Neural rendering benchmark in sparse-view setting.", + "bbox": [ + 84, + 546, + 459, + 561 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the reconstructed shape of the objects. For further quantitative analysis, we evaluate the performance using P2S and CD in Tab. 4.", + "bbox": [ + 76, + 575, + 468, + 619 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Sparse-View Human-object Rendering Benchmark. Our HODome supports various neural rendering tasks, even for human-object interactions. We provide a benchmark on sparse-view rendering tasks and evaluate on IBRNet [69], NeuRay [30] and NeuralHumanFVV [60]. Besides, we also provide a baseline, named NeuralHOIFVV, which uses the projection of trained PIFu's [48] results as depth input, then applies the Neural Blending method presented in [79] to obtain novel-view synthesis. More details about NeuralHOIFVV are referred to the Appendix. We analyze the above methods on HODome and in-the-wild images. The qualitative and quantitative results are shown in Fig. 9 and Tab. 5. Our proposed NeuralHOIFVV can obtain generative novel-view synthesis on HODome and in-the-wild inputs.", + "bbox": [ + 76, + 623, + 468, + 834 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Limitations", + "text_level": 1, + "bbox": [ + 76, + 845, + 202, + 859 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Although NeuralDome can provide accurate capturing results and layer-wise neural representations, we also want", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6eaa6ccfcf504c35f5ade452b9f6517ce9131c8f2ccb5cd221ba76e2a1101d8d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 95, + 570, + 180 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1daaf82378f503cfe4ba5ffccbdab163354242a30edae5969bc5c2eac14393dc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 584, + 92, + 627, + 180 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4a0643b8bb8ad2571537467f4d4f233b4f36de28b6667780198a66abadaa30d6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 627, + 93, + 687, + 178 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/57fd79734fae9b65d5873a0b5b4e059735e8b181dc25b53fce4e97e88d8056c9.jpg", + "image_caption": [ + "GT IBRNet" + ], + "image_footnote": [], + "bbox": [ + 622, + 178, + 687, + 258 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7134b8c3401540046b7b5fd75b7c2d5c59c8f868e8e115e0a33c06a91acf6766.jpg", + "image_caption": [ + "NeuRay" + ], + "image_footnote": [], + "bbox": [ + 691, + 92, + 750, + 260 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1943accd934a35f81527c8be64e66948d3fef9fede39a25320683f904b982774.jpg", + "image_caption": [ + "Neural HumanFVV", + "Figure 9. Qualitative evaluation on neural human-object rendering. We show the comparison of IBRNet [69], NeuRay [30], NeuralHumanFVV [60] and our proposed NeuralHOIFVV. We render one novel view between two input camera views." + ], + "image_footnote": [], + "bbox": [ + 756, + 92, + 815, + 262 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e21a01a06ded1545afa492c11daab668908d812f13ae91bcacc0da7bb875cec6.jpg", + "image_caption": [ + "Neural HOIFVV" + ], + "image_footnote": [], + "bbox": [ + 821, + 93, + 877, + 261 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "to highlight some potential limitations of this pipeline. First, NeuralDome only considers single-person interaction with objects and the single-person task is already quite challenging for current research. It is non-trivial to extend the current algorithm to multi-person settings, especially in a crowd scene. Secondly, the reconstruction of 3D holistic scene is not covered in our pipeline. We leave the joint modeling of humans, objects and scene in the future work. Moreover, our dataset was collected under fixed illumination conditions with few backgrounds variance, limiting its generalization ability for other environments.", + "bbox": [ + 496, + 369, + 890, + 536 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 564, + 617, + 579 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have presented HODome, the first dataset to jointly capture and render human-object interaction using dense RGB cameras and the Optitrack system. To process the HODome dataset, we have further developed NeuralDome, a neural pipeline that can accurately track humans and objects, conduct layer-wise geometry reconstruction, and enable novel-view synthesis for both the subjects and objects. We believe the dataset could boost the development of capturing and rendering human-object interactions and is very valuable to the community for a wide range of human-object capturing and rendering tasks.", + "bbox": [ + 496, + 594, + 890, + 760 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement This work was supported by the Shanghai Sailing Program (21YF1429400, 22YF1428800), Shanghai Local College Capacity Building Program (23010503100, 22010502800), NSFC programs (61976138, 61977047), the National Key Research and Development Program (2018YFB2100500), STCSM (2015F0203-000-06), SHMEC (2019-01-07-00-01-E00003) and Shanghai Frontiers Science Center of Humancentered Artificial Intelligence (ShangHAI).", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8841", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Naturalpoint, inc. motion capture systems. https://optitrack.com/. 6.3", + "[2] Reality capture. https://www.capturingreality.com/realitycapture. 3, 4", + "[3] Easymocap - make human motion capture easier. Github, 2021.4", + "[4] Kara-Ali Aliev, Artem Sevastopolsky, Maria Kolos, Dmitry Ulyanov, and Victor Lempitsky. Neural point-based graphics. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16, pages 696-712. Springer, 2020. 2", + "[5] Paul J Besl and Neil D McKay. Method for registration of 3-d shapes. In Sensor fusion IV: control paradigms and data structures, volume 1611, pages 586-606. Spie, 1992. 4", + "[6] Bharat Lal Bhatnagar, Xianghui Xie, Ilya A Petrov, Cristian Sminchisescu, Christian Theobalt, and Gerard Pons-Moll. Behave: Dataset and method for tracking human object interactions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15935-15946, 2022. 2, 3, 4, 7", + "[7] Hongrui Cai, Wanquan Feng, Xuetao Feng, Yan Wang, and Juyong Zhang. Neural surface reconstruction of dynamic scenes with monocular rgb-d camera. In Thirty-sixth Conference on Neural Information Processing Systems (NeurIPS), 2022. 2", + "[8] Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Realtime multi-person 2d pose estimation using part affinity fields. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7291-7299, 2017. 4", + "[9] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. arXiv preprint arXiv:2203.09517, 2022. 2", + "[10] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14124-14133, 2021. 2", + "[11] Alvaro Collet, Ming Chuang, Pat Sweeney, Don Gillett, Dennis Evseev, David Calabrese, Hugues Hoppe, Adam Kirk, and Steve Sullivan. High-quality streamable free-viewpoint video. ACM Transactions on Graphics (TOG), 34(4):69, 2015. 2, 3", + "[12] Mingsong Dou, Philip Davidson, Sean Ryan Fanello, Sameh Khamis, Adarsh Kowdle, Christoph Rhemann, Vladimir Tankovich, and Shahram Izadi. Motion2fusion: Real-time volumetric performance capture. ACM Trans. Graph., 36(6):246:1-246:16, Nov. 2017. 2, 3", + "[13] Kaiwen Guo, Peter Lincoln, Philip Davidson, Jay Busch, Xueming Yu, Matt Whalen, Geoff Harvey, Sergio Orts-Escolano, Rohit Pandey, Jason Dourgarian, et al. The relightables: Volumetric performance capture of humans with realistic relighting. ACM Transactions on Graphics (ToG), 38(6):1–19, 2019. 3", + "[14] Marc Habermann, Lingjie Liu, Weipeng Xu, Michael Zollhoefer, Gerard Pons-Moll, and Christian Theobalt. Real-time" + ], + "bbox": [ + 78, + 116, + 467, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "deep dynamic characters. ACM Transactions on Graphics (TOG), 40(4):1-16, 2021. 2", + "[15] Nils Hasler, Carsten Stoll, Martin Sunkel, Bodo Rosenhahn, and H-P Seidel. A statistical model of human pose and body shape. In Computer graphics forum, volume 28, pages 337-346. Wiley Online Library, 2009. 3", + "[16] Mohamed Hassan, Vasileios Choutas, Dimitrios Tzionas, and Michael J Black. Resolving 3d human pose ambiguities with 3d scene constraints. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2282-2292, 2019. 3, 5", + "[17] Tong He, Yuanlu Xu, Shunsuke Saito, Stefano Soatto, and Tony Tung. Arch++: Animation-ready clothed human reconstruction revisited. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11046-11056, 2021. 2", + "[18] Yinghao Huang, Omid Taheri, Michael J Black, and Dimitrios Tzionas. Intercap: Joint markerless 3d tracking of humans and objects in interaction. In DAGM German Conference on Pattern Recognition, pages 281-299. Springer, 2022. 3", + "[19] Zeng Huang, Yuanlu Xu, Christoph Lassner, Hao Li, and Tony Tung. Arch: Animatable reconstruction of clothed humans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3093-3102, 2020. 2", + "[20] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3.6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE Transactions on Pattern Analysis and Machine Intelligence, 36(7):1325-1339, jul 2014. 3", + "[21] Yuheng Jiang, Suyi Jiang, Guoxing Sun, Zhuo Su, Kaiwen Guo, Minye Wu, Jingyi Yu, and Lan Xu. Neuralhofusion: Neural volumetric rendering under human-object interactions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6155-6165, 2022. 2, 3", + "[22] Hanbyul Joo, Tomas Simon, Xulong Li, Hao Liu, Lei Tan, Lin Gui, Sean Banerjee, Timothy Scott Godisart, Bart Nabbe, Iain Matthews, Takeo Kanade, Shohei Nobuhara, and Yaser Sheikh. Panoptic studio: A massively multiview system for social interaction capture. 2017. 3", + "[23] Youngjoong Kwon, Dahun Kim, Duygu Ceylan, and Henry Fuchs. Neural human performer: Learning generalizable radiance fields for human performance rendering. Advances in Neural Information Processing Systems, 34:24741-24752, 2021. 2", + "[24] Christoph Lassner, Javier Romero, Martin Kiefel, Federica Bogo, Michael J. Black, and Peter V. Gehler. Unite the people: Closing the loop between 3d and 2d human representations. In IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), July 2017. 3", + "[25] Ruilong Li, Julian Tanke, Minh Vo, Michael Zollhofer, Jurgen Gall, Angjoo Kanazawa, and Christoph Lassner. Tava: Template-free animatable volumetric actors. arXiv preprint arXiv:2206.08929, 2022. 2", + "[26] Jun Liu, Amir Shahroudy, Mauricio Perez, Gang Wang, Ling-Yu Duan, and Alex C Kot. Nturgb+ d 120: A large-" + ], + "bbox": [ + 503, + 93, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "8842", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "scale benchmark for 3d human activity understanding. IEEE transactions on pattern analysis and machine intelligence, 42(10):2684-2701, 2019. 2, 3", + "[27] Lingjie Liu, Marc Habermann, Viktor Rudnev, Kripasindhu Sarkar, Jiatao Gu, and Christian Theobalt. Neural actor: Neural free-view synthesis of human actors with pose control. ACM Trans. Graph. (ACM SIGGRAPH Asia), 2021. 2", + "[28] Lingjie Liu, Weipeng Xu, Marc Habermann, Michael Zollhöfer, Florian Bernard, Hyeongwoo Kim, Wenping Wang, and Christian Theobalt. Neural human video rendering by learning dynamic textures and rendering-to-video translation. arXiv preprint arXiv:2001.04947, 2020. 2", + "[29] Lingjie Liu, Weipeng Xu, Michael Zollhoefer, Hyeongwoo Kim, Florian Bernard, Marc Habermann, Wenping Wang, and Christian Theobalt. Neural rendering and reenactment of human actor videos. ACM Transactions on Graphics (TOG), 38(5):1-14, 2019. 2", + "[30] Yuan Liu, Sida Peng, Lingjie Liu, Qianqian Wang, Peng Wang, Christian Theobalt, Xiaowei Zhou, and Wenping Wang. Neural rays for occlusion-aware image-based rendering. In CVPR, 2022. 2, 8", + "[31] Stephen Lombardi, Tomas Simon, Jason Saragih, Gabriel Schwartz, Andreas Lehrmann, and Yaser Sheikh. Neural volumes: Learning dynamic renderable volumes from images. arXiv preprint arXiv:1906.07751, 2019. 2", + "[32] Stephen Lombardi, Tomas Simon, Gabriel Schwartz, Michael Zollhoefer, Yaser Sheikh, and Jason Saragih. Mixture of volumetric primitives for efficient neural rendering. arXiv preprint arXiv:2103.01954, 2021. 2", + "[33] Xiaoxiao Long, Cheng Lin, Peng Wang, Taku Komura, and Wenping Wang. Sparseneus: Fast generalizable neural surface reconstruction from sparse views. ECCV, 2022. 2", + "[34] H. Luo, A. Chen, Q. Zhang, B. Pang, M. Wu, L. Xu, and J. Yu. Convolutional neural opacity radiance fields. In 2021 IEEE International Conference on Computational Photography (ICCP), pages 1-12, Los Alamitos, CA, USA, may 2021. IEEE Computer Society. 2", + "[35] Haimin Luo, Teng Xu, Yuheng Jiang, Chenglin Zhou, Qiwei Qiu, Yingliang Zhang, Wei Yang, Lan Xu, and Jingyi Yu. Artemis: Articulated neural pets with appearance and motion synthesis. ACM Trans. Graph., 41(4), jul 2022. 2", + "[36] Quan Meng, Anpei Chen, Haimin Luo, Minye Wu, Hao Su, Lan Xu, Xuming He, and Jingyi Yu. Gnerf: Gan-based neural radiance field without posed camera. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6351-6361, 2021. 2", + "[37] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European conference on computer vision, pages 405-421. Springer, 2020. 2", + "[38] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, July 2022. 2, 4, 6", + "[39] Michael Niemeyer, Jonathan T. Barron, Ben Mildenhall, Mehdi S. M. Sajjadi, Andreas Geiger, and Noha Radwan." + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 2", + "[40] Atsuhiro Noguchi, Xiao Sun, Stephen Lin, and Tatsuya Harada. Neural articulated radiance field. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5762-5772, 2021. 2", + "[41] Anqi Pang, Xin Chen, Haimin Luo, Minye Wu, Jingyi Yu, and Lan Xu. Few-shot neural human performance rendering from sparse rgbd videos, 2021. 2", + "[42] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M. Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. ACM Trans. Graph., 40(6), dec 2021. 2", + "[43] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3d hands, face, and body from a single image. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10975-10985, 2019. 2, 5", + "[44] Sida Peng, Junting Dong, Qianqian Wang, Shangzhan Zhang, Qing Shuai, Xiaowei Zhou, and Hujun Bao. Animatable neural radiance fields for modeling dynamic human bodies. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14314-14323, 2021. 2", + "[45] Sida Peng, Yuanqing Zhang, Yinghao Xu, Qianqian Wang, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In CVPR, 2021. 2, 6, 7", + "[46] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, jun 2021. 2", + "[47] Yu Rong, Takaaki Shiratori, and Hanbyul Joo. Frankmocap: A monocular 3d whole-body pose estimation system via regression and integration. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1749-1759, 2021. 7", + "[48] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2304-2314, 2019. 2, 7, 8", + "[49] Shunsuke Saito, Tomas Simon, Jason Saragih, and Hanbyul Joo. Pifuhd: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 84-93, 2020. 2", + "[50] Manolis Savva, Angel X Chang, Pat Hanrahan, Matthew Fisher, and Matthias Nießner. Pigraphs: learning interaction snapshots from observations. ACM Transactions on Graphics (TOG), 35(4):1-12, 2016. 3", + "[51] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-Motion Revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2, 3" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "8843", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[52] Qing Shuai, Chen Geng, Qi Fang, Sida Peng, Wenhao Shen, Xiaowei Zhou, and Hujun Bao. Novel view synthesis of human interactions from sparse multi-view videos. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-10, 2022. 2", + "[53] Qing Shuai, Chen Geng, Qi Fang, Sida Peng, Wenhao Shen, Xiaowei Zhou, and Hujun Bao. Novel view synthesis of human interactions from sparse multi-view videos. In ACM SIGGRAPH 2022 Conference Proceedings, SIGGRAPH '22, New York, NY, USA, 2022. Association for Computing Machinery. 3, 6", + "[54] Aliaksandra Shysheya, Egor Zakharov, Kara-Ali Aliev, Renat Bashirov, Egor Burkov, Karim Iskakov, Aleksei Ivakhnenko, Yury Malkov, Igor Pasechnik, Dmitry Ulyanov, et al. Textured neural avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2387-2397, 2019. 2", + "[55] Leonid Sigal, Alexandru O Balan, and Michael J Black. Humaneva: Synchronized video and motion capture dataset and baseline algorithm for evaluation of articulated human motion. International journal of computer vision, 87(1):4-27, 2010. 3", + "[56] Zhuo Su, Lan Xu, Zerong Zheng, Tao Yu, Yebin Liu, and Lu Fang. Robustfusion: Human volumetric capture with data-driven visual cues using a rgbd camera. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, Computer Vision – ECCV 2020, pages 246–264, Cham, 2020. Springer International Publishing. 3", + "[57] Guoxing Sun, Xin Chen, Yizhang Chen, Anqi Pang, Pei Lin, Yuheng Jiang, Lan Xu, Jingya Wang, and Jingyi Yu. Neural free-viewpoint performance rendering under complex human-object interactions. In Proceedings of the 29th ACM International Conference on Multimedia, 2021. 2, 3", + "[58] Guoxing Sun, Xin Chen, Yizhang Chen, Anqi Pang, Pei Lin, Yuheng Jiang, Lan Xu, Jingyi Yu, and Jingya Wang. Neural free-viewpoint performance rendering under complex human-object interactions. In Proceedings of the 29th ACM International Conference on Multimedia, pages 4651-4660, 2021. 3", + "[59] Xin Suo, Yuheng Jiang, Pei Lin, Yingliang Zhang, Minye Wu, Kaiwen Guo, and Lan Xu. Neuralhumanfvv: Realtime neural volumetric human performance rendering using rgb cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6226-6237, 2021. 2", + "[60] Xin Suo, Yuheng Jiang, Pei Lin, Yingliang Zhang, Minye Wu, Kaiwen Guo, and Lan Xu. Neuralhumanfvv: Real-time neural volumetric human performance rendering using rgb cameras. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6222-6233, 2021. 8", + "[61] Omid Taheri, Nima Ghorbani, Michael J Black, and Dimitrios Tzionas. Grab: A dataset of whole-body human grasping of objects. In European conference on computer vision, pages 581-600. Springer, 2020. 2, 3, 4", + "[62] Ayush Tewari, Justus Thies, Ben Mildenhall, Pratul Srinivasan, Edgar Tretschk, Yifan Wang, Christoph Lassner, Vincent Sitzmann, Ricardo Martin-Brualla, Stephen Lom-" + ], + "bbox": [ + 78, + 90, + 470, + 901 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "bardi, Tomas Simon, Christian Theobalt, Matthias Niessner, Jonathan T. Barron, Gordon Wetzstein, Michael Zollhoefer, and Vladislav Golyanik. Advances in neural rendering, 2021. 2", + "[63] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhofer, Christoph Lassner, and Christian Theobalt. Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12959-12970, 2021. 2", + "[64] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T Barron, and Pratul P Srinivasan. Ref-nerf: Structured view-dependent appearance for neural radiance fields. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5481-5490. IEEE, 2022. 2", + "[65] Timo Von Marcard, Roberto Henschel, Michael J Black, Bodo Rosenhahn, and Gerard Pons-Moll. Recovering accurate 3d human pose in the wild using imus and a moving camera. In Proceedings of the European Conference on Computer Vision (ECCV), pages 601-617, 2018. 3", + "[66] Liao Wang, Ziyu Wang, Pei Lin, Yuheng Jiang, Xin Suo, Minye Wu, Lan Xu, and Jingyi Yu. ibutter: Neural interactive bullet time generator for human free-viewpoint rendering. In Proceedings of the 29th ACM International Conference on Multimedia, pages 4641-4650, 2021. 2", + "[67] Liao Wang, Jiakai Zhang, Xinhang Liu, Fuqiang Zhao, Yanshun Zhang, Yingliang Zhang, Minye Wu, Jingyi Yu, and Lan Xu. Fourier plenoctrees for dynamic radiance field rendering in real-time. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13524-13534, 2022. 2", + "[68] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. NeurIPS, 2021. 2", + "[69] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul Srinivasan, Howard Zhou, Jonathan T. Barron, Ricardo MartinBrualla, Noah Snively, and Thomas Funkhouser. Ibrnet: Learning multi-view image-based rendering. In CVPR, 2021. 2,8", + "[70] Shaofei Wang, Katja Schwarz, Andreas Geiger, and Siyu Tang. Arah: animatable volume rendering of articulated human sdfs. arXiv preprint arXiv:2210.10036, 2022. 2", + "[71] Minye Wu, Yuehao Wang, Qiang Hu, and Jingyi Yu. Multi-view neural human rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 2", + "[72] Minye Wu, Yuehao Wang, Qiang Hu, and Jingyi Yu. Multi-view neural human rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1682-1691, 2020. 2", + "[73] Xianghui Xie, Bharat Lal Bhatnagar, and Gerard Pons-Moll. Chore: Contact, human and object reconstruction from a single rgb image. arXiv preprint arXiv:2204.02445, 2022. 3, 7", + "[74] Hongwei Yi, Chun-Hao P Huang, Dimitrios Tzionas, Muhammed Kocabas, Mohamed Hassan, Siyu Tang, Justus Thies, and Michael J Black. Human-aware object placement" + ], + "bbox": [ + 501, + 92, + 893, + 901 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "8844", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "for visual environment reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3959-3970, 2022. 3, 7", + "[75] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. Plenoctrees for real-time rendering of neural radiance fields. In arXiv, 2021. 2", + "[76] Chao Zhang, Sergi Pujades, Michael J. Black, and Gerard Pons-Moll. Detailed, accurate, human shape estimation from clothed 3d scan sequences. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), July 2017. 3", + "[77] Jiakai Zhang, Xinhang Liu, Xinyi Ye, Fuqiang Zhao, Yanshun Zhang, Minye Wu, Yingliang Zhang, Lan Xu, and Jingyi Yu. Editable free-viewpoint video using a layered neural representation. ACM Trans. Graph., 40(4), July 2021. 2, 3, 5, 6, 7", + "[78] Jason Y Zhang, Sam Pepose, Hanbyul Joo, Deva Ramanan, Jitendra Malik, and Angjoo Kanazawa. Perceiving 3d human-object spatial arrangements from a single image in the wild. In European conference on computer vision, pages 34-51. Springer, 2020. 3, 7", + "[79] Fuqiang Zhao, Yuheng Jiang, Kaixin Yao, Jiakai Zhang, Liao Wang, Haizhao Dai, Yuhui Zhong, Yingliang Zhang, Minye Wu, Lan Xu, et al. Human performance modeling and rendering via neural animated mesh. arXiv preprint arXiv:2209.08468, 2022. 2, 4, 6, 8", + "[80] Fuqiang Zhao, Wei Yang, Jiakai Zhang, Pei Lin, Yingliang Zhang, Jingyi Yu, and Lan Xu. Humannerf: Efficiently generated human radiance field from sparse inputs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7743-7753, 2022. 2, 5", + "[81] Qian-Yi Zhou, Jaesik Park, and Vladlen Koltun. Open3d: A modern library for 3d data processing. arXiv preprint arXiv:1801.09847, 2018. 4" + ], + "bbox": [ + 78, + 92, + 468, + 556 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "8845", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2023/NeuralDome_ A Neural Modeling Pipeline on Multi-View Human-Object Interactions/fa0c20cd-9328-405b-b73e-2045fdfa714d_model.json b/2023/NeuralDome_ A Neural Modeling Pipeline on Multi-View Human-Object Interactions/fa0c20cd-9328-405b-b73e-2045fdfa714d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..23b31aa02ca68a6711c76dbb16c28c2369e5d407 --- /dev/null +++ b/2023/NeuralDome_ A Neural Modeling Pipeline on Multi-View Human-Object Interactions/fa0c20cd-9328-405b-b73e-2045fdfa714d_model.json @@ -0,0 +1,2787 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.118, + 0.13, + 0.853, + 0.175 + ], + "angle": 0, + "content": "NeuralDome: A Neural Modeling Pipeline on Multi-View Human-Object Interactions" + }, + { + "type": "text", + "bbox": [ + 0.113, + 0.202, + 0.88, + 0.239 + ], + "angle": 0, + "content": "Juze Zhang \\(^{1,2,3,4,*}\\), Haimin Luo \\(^{1,4,5,*}\\), Hongdi Yang \\(^{1,4}\\), Xinru Xu \\(^{1,4}\\), Qianyang Wu \\(^{1,4}\\), Ye Shi \\(^{1,4}\\), Jingyi Yu \\(^{1,4}\\), Lan Xu \\(^{1,4,\\dagger}\\), Jingya Wang \\(^{1,4,\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.108, + 0.239, + 0.888, + 0.275 + ], + "angle": 0, + "content": "\\(^{1}\\) ShanghaiTech University \\(^{2}\\) Shanghai Advanced Research Institute, Chinese Academy of Sciences \\(^{3}\\) University of Chinese Academy of Sciences" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.275, + 0.904, + 0.311 + ], + "angle": 0, + "content": "\\(^{4}\\) Shanghai Engineering Research Center of Intelligent Vision and Imaging \\(^{5}\\) LumiAni Technology \\(\\{\\mathrm{zhangjz,luohm,yanghd,xuxr2022,wuqy2022,shiye,yujingyi,xulan1,wangjingya}\\} @shanghaiitech.edu.cn\\)" + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.239, + 0.904, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.114, + 0.385, + 0.858, + 0.571 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.581, + 0.893, + 0.637 + ], + "angle": 0, + "content": "Figure 1. Our NeuralDome pipeline for processing multi-view video sequences on human object interactions. NeuralDome supports tracking, modeling, and rendering of individual human subjects and objects. To validate NeuralDome, we collect a large dataset HODome over a total of 71 million video frames across 76 viewpoints and process the datasets using NeuralDome for a variety of inference and neural modeling and rendering tasks." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.657, + 0.314, + 0.673 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.691, + 0.471, + 0.858 + ], + "angle": 0, + "content": "Humans constantly interact with objects in daily life tasks. Capturing such processes and subsequently conducting visual inferences from a fixed viewpoint suffers from occlusions, shape and texture ambiguities, motions, etc. To mitigate the problem, it is essential to build a training dataset that captures free-viewpoint interactions. We construct a dense multi-view dome to acquire a complex human object interaction dataset, named HODome, that consists of \\(\\sim 71M\\) frames on 10 subjects interacting with 23 objects. To process the HODome dataset, we develop NeuralDome, a layer-wise neural processing pipeline tai" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.659, + 0.894, + 0.794 + ], + "angle": 0, + "content": "lored for multi-view video inputs to conduct accurate tracking, geometry reconstruction and free-view rendering, for both human subjects and objects. Extensive experiments on the HODome dataset demonstrate the effectiveness of NeuralDome on a variety of inference, modeling, and rendering tasks. Both the dataset and the NeuralDome tools will be disseminated to the community for further development, which can be found at https://juzezhang.github.io/NeuralDome" + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.829, + 0.631, + 0.845 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.857, + 0.893, + 0.903 + ], + "angle": 0, + "content": "A key task of computer vision is to understand how humans interact with the surrounding world, by faithfully capturing and subsequently reproducing the process via mod" + }, + { + "type": "page_footnote", + "bbox": [ + 0.078, + 0.876, + 0.273, + 0.89 + ], + "angle": 0, + "content": "* These authors contributed equally." + }, + { + "type": "page_footnote", + "bbox": [ + 0.078, + 0.89, + 0.207, + 0.902 + ], + "angle": 0, + "content": "†Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.876, + 0.273, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8834" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.137 + ], + "angle": 0, + "content": "eling and rendering. Successful solutions benefit broad applications ranging from sports training to vocational education, digital entertainment to tele-medicine." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.139, + 0.473, + 0.517 + ], + "angle": 0, + "content": "Early solutions [11, 12, 51] that reconstruct dynamic meshes with per-frame texture maps are time-consuming and vulnerable to occlusions or lack of textures. Recent advances in neural rendering [37, 62, 71] bring huge potential for human-centric modeling. Most notably, the variants of Neural Radiance Field (NeRF) [37] achieve compelling novel view synthesis, which can enable real-time rendering performance [38, 59, 67] even for dynamic scenes [45, 63, 77], and can be extended to the generative setting without per-scene training [23, 69, 80]. However, less attention is paid to the rich and diverse interactions between humans and objects, mainly due to the severe lack of dense-view human-object datasets. Actually, existing datasets of human-object interactions are mostly based on optical markers [61] or sparse RGB/RGBD sensors [6, 26], without sufficient appearance supervision for neural rendering tasks. As a result, the literature on neural human-object rendering [21,57] is surprisingly sparse, let alone further exploring the real-time or generative directions. Besides, existing neural techniques [52, 77] suffer from tedious training procedures due to the human-object occlusion, and hence infeasible for building a large-scale dataset. In a nutshell, despite the recent tremendous thriving of neural rendering, the lack of both a rich dataset and an efficient reconstruction scheme constitute barriers in human-object modeling." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.52, + 0.47, + 0.64 + ], + "angle": 0, + "content": "In this paper, we present NeuralDome, a neural pipeline that takes multi-view dome capture as inputs and conducts accurate 3D modeling and photo-realistic rendering of complex human-object interaction. As shown in Fig. 1, NeuralDome exploits layer-wise neural modeling to produce rich and multi-modality outputs including the geometry of dynamic human, object shapes and tracked poses, as well as a free-view rendering of the sequence." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.643, + 0.47, + 0.809 + ], + "angle": 0, + "content": "Specifically, we first capture a novel human-object dome (HODome) dataset that consists of 274 human-object interacting sequences, covering 23 diverse 3D objects and 10 human subjects (5 males and 5 females) in various apparels. We record multi-view video sequences of natural interactions between the human subjects and the objects where each sequence is about 60s in length using a dome with 76 RGB cameras, resulting in 71 million video frames. We also provide an accurate pre-scanned 3D template for each object and utilize sparse optical markers to track individual objects throughout the sequences." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.47, + 0.901 + ], + "angle": 0, + "content": "To process the HODome dataset, we adopt an extended Neural Radiance Field (NeRF) pipeline. The brute-force adoption of off-the-shelf neural techniques such as Instant-NSR [79] and Instant-NGP [38], although effective, do not separate objects from human subjects and therefore lack sufficient fidelity to model their interactions. We instead" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.394 + ], + "angle": 0, + "content": "introduce a layer-wise neural processing pipeline. Specifically, we first perform a joint optimization based on the dense inputs for accurately tracking human motions using the parametric SMPL-X model [43] as well as localizing objects using template meshes. We then propose an efficient layer-wise neural rendering scheme where the humans and objects are formulated as a pose-embedded dynamic NeRF and a static NeRF with tracked 6-DoF rigid poses, respectively. Such a layer-wise representation effectively exploits temporal information and robustly tackles the occluded regions under interactions. We further introduce an object-aware ray sampling strategy to mitigate artifacts during layer-wise training, as well as template-aware geometry regularizers to enforce contact-aware deformations. Through weak segmentation supervision, we obtain the decoupled and occlusion-free appearances for both the humans and the objects at a high fidelity amenable for training the input multi-view inputs for a variety of tasks from monocular motion capture to free-view rendering from sparse multi-view inputs." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.394, + 0.831, + 0.408 + ], + "angle": 0, + "content": "To summarize, our main contributions include:" + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.42, + 0.891, + 0.48 + ], + "angle": 0, + "content": "- We introduce NeuralDome, a neural pipeline, to accurately track humans and objects, conduct layer-wise geometry reconstruction, and enable novel-view synthesis, from multi-view HOI video inputs." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.485, + 0.892, + 0.576 + ], + "angle": 0, + "content": "- We collect a comprehensive dataset that we call HODome that will be disseminated to the community, with both raw data and the output modalities including separated geometry and rendering of individual objects and human subjects, their tracking results, free-view rendering results, etc." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.582, + 0.892, + 0.626 + ], + "angle": 0, + "content": "- We demonstrate using the dataset to train networks for a variety of visual inference tasks with complex human object interactions." + }, + { + "type": "list", + "bbox": [ + 0.519, + 0.42, + 0.892, + 0.626 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.641, + 0.65, + 0.657 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.666, + 0.738, + 0.683 + ], + "angle": 0, + "content": "2.1. Neural Human Rendering" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Various 3D data representations have been explored for neural human rendering, such as point-clouds [4, 41, 72], textured meshes [14, 28, 29, 54], and volumes [31, 32]. Implicit occupancy function-based methods [17, 19, 48, 49] can recover detailed 3D human geometry from sparse 2D images without faithful appearance synthesis. The recent NeRF [37] technique brings huge potential for 3D photorealistic view synthesis [9, 10, 30, 34, 36, 38, 39, 64, 69, 75] and geometry modeling [7, 33, 68]. Further explorations extend it to dynamic scenes [7, 35, 42, 46, 67, 77], especially for humans. Existing works equip NeRF with pose-embeddings [23, 27, 40, 45, 80], learnable skinning weights [25, 44, 70] and even generalization across individuals [23, 66, 80]. However, these works only focus on" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8835" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.1, + 0.09, + 0.872, + 0.193 + ], + "angle": 0, + "content": "
Datasets#Cam-view# Frame(M)ResolutionFpsMarkerObj. Num.Human Annot.Novel viewsGeometryHuman TexturedNeural RepresentationObject Appearance
NTU [26]3~ 341920 × 108030×NA✓NA××××
PiGr [50]10.1960 × 5405×NA✓NA×××✓
GRAB [61]×NANANA✓51××××××
PROX [16]10.11920 × 108030✓NA✓××××✓
BEHAVE [6]40.152048 × 153630×20×××××✓
InterCap [18]60.071920 × 108030×10✓××××✓
Our76713840 × 216060✓23✓✓✓✓✓✓
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.203, + 0.894, + 0.248 + ], + "angle": 0, + "content": "Table 1. Dataset Comparisons. We compare our proposed HODome dataset with the existing publicly available human-object dataset. HODome has the largest scale of human-object interactions in terms of the number of frames (#Frame), camera view (#Cam-view), and modality. \"Obj. Num.\" represents the object number. \"Human Annot.\" represents annotation from professional annotators." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.268, + 0.468, + 0.447 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.458, + 0.472, + 0.542 + ], + "angle": 0, + "content": "Figure 2. HODome Modality. HODome features multiple modalities of data format and annotations, including a) Color image, b) Human-object mask (Homask), c) MoCap (SMPL-X parameters and located object template), d) Geometry, e) digital assets in neural representations, f) Novel view and g) Keypoints (25 for bodies and 42 for hands with human annotation)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.567, + 0.47, + 0.629 + ], + "angle": 0, + "content": "a single person. The most relevant works [53, 77] aim to model multi-person interactions, but they cannot handle more complex interactions due to the lack of large-scale human-object interactions dataset." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.636, + 0.303, + 0.654 + ], + "angle": 0, + "content": "2.2. Human-object Modeling" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Only a few works [6, 16, 18, 61, 73, 78] consider to jointly model whole-body interactions. Early solutions [11, 12, 51] that reconstruct meshes with per-frame texture maps, which are vulnerable to occlusions and lack of textures. Recently, several works [6, 16, 18, 61] explore the relationship via several interaction constraints, such as contact map [6, 18, 61], spatial arrangement [78] and physically plausible constraint [74]. However, these methods only produce a relative arrangement. The most relevant works [21, 56, 58] aim to render human-object interactions with volumetric fusion [56], neural texturing blending [58] and volumetric rendering [21]. However, without sufficient appearance supervision, their quality is limited. Comparably, our layerwise neural representations can be converted to high-quality geometry and support separate free-view rendering for human or object." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.271, + 0.719, + 0.287 + ], + "angle": 0, + "content": "2.3. Human-centric Dataset" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.295, + 0.893, + 0.538 + ], + "angle": 0, + "content": "A variety of human-centric datasets have been developed for human-only capturing or rendering tasks. Early datasets combine multi-camera RGB video capture with synchronized ground-truth 3D skeletons [20, 22, 55], parametric model [24, 65], scanned mesh [15, 76], without human-object interaction. Recent datasets capture human-object interactions using optical markers [61] or sparse RGB sensors [6, 26], without sufficient appearance supervision. As a result, the literature on neural human-object rendering [21, 57] is surprisingly rare. High-end works [11, 13] use dense cameras for reconstruction and rendering of humans and objects through mesh reconstruction and motion tracking, but without texture and neural representations. To fill this gap, we propose NeuralDome for capturing and rendering human-object interactions, facilitating various generative human-object tasks." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.552, + 0.673, + 0.569 + ], + "angle": 0, + "content": "3. HODome Dataset" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.578, + 0.892, + 0.744 + ], + "angle": 0, + "content": "We present the HODome dataset for capturing and rendering photo-realistic human-object interactions. It consists of 274 human-object interacting sequences, covering 23 diverse 3D objects and 10 subjects (5 males and 5 females) under various apparels. For each sequence, we record the naturally interacting scene for about 60s using 76 RGB cameras for dense-view at \\(3840 \\times 2160\\) resolution and 60 frame-per-seconds (fps), resulting in roughly 71 million frames. HODome consists of rich labels covering different aspects of HOI capturing and rendering labels (see Fig.2). See Tab. 1 for comparison with other datasets." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.756, + 0.719, + 0.773 + ], + "angle": 0, + "content": "3.1. Data Capturing System" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.893, + 0.903 + ], + "angle": 0, + "content": "To construct the HODome dataset, we use 76 Z-CAM cinema cameras with sufficient appearance supervision for neural rendering tasks and 16 Optitrack MoCap cameras [1] for accurate human-object tracking tasks. The Z-CAM system and Optitrack system are synchronized to record the RGB and MoCap data together. We use a publicly available tool [2] to estimate the intrinsic camera parameters and extrinsic camera parameters." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8836" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.089, + 0.878, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.265, + 0.895, + 0.322 + ], + "angle": 0, + "content": "Figure 3. Overview of NeuralDome. Given the 76-view RGB stream as input, we first jointly track human skeletal motion and object rigid motion. Then with the tracked motion priors, we decouple the human-object interaction scenes via a layer-wise neural rendering scheme, to generate human/object renderings separately and corresponding acceptable segmentation maps. Blending with the high-fidelity captured views, we obtain the layer-wise neural representations in our HODome dataset." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.347, + 0.248, + 0.364 + ], + "angle": 0, + "content": "3.2. Dataset Modality" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.372, + 0.47, + 0.492 + ], + "angle": 0, + "content": "Human MoCap and Object 6D Pose. For appearance realism, we do not place markers on human actors. Thus we detect the human joints from 76 RGB images by running the whole-body Openpose [8] to perform markerless MoCap. And we treat each object as rigid and solve the rigid object pose estimation from markers using the Iterative Closest Point(ICP) algorithm [5, 81]. See Fig. 3 for the pipeline of our method." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.495, + 0.47, + 0.646 + ], + "angle": 0, + "content": "Human-Object Neural Representations. Under the dense-view setting, one can apply off-the-shell neural techniques, i.e., Instant-NSR [79] and Instant-NGP [38] to efficiently obtain per-frame geometry and novel-view synthesis of the whole human-object sequence, respectively. However, neither schemes separate object from human subject. We therefore propose a layer-wise neural scheme to separately recover neural representations of human and object, enabling both high-fidelity geometry reconstruction and free-view appearance rendering (Sec. 4.2)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.648, + 0.471, + 0.801 + ], + "angle": 0, + "content": "Data Annotation. Our dataset includes 23 objects with varying scales and interaction types. Each object was pre-scanned using off-the-shelf multi-view software packages [2]. Further, following the previous methods [6, 61], we annotated its pseudo contact label that computes from a threshold distance. To provide more accurate results for the quantitative benchmark, we annotate a separate quantitative subset as our test set with human-annotated segmentation and hand joints by the professional annotator. Our datasets will be available for research purposes." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.815, + 0.355, + 0.832 + ], + "angle": 0, + "content": "4. Neural Modeling on HODome" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.903 + ], + "angle": 0, + "content": "We introduce a neural pipeline to produce the rich digital assets from the dense-view input of each human-object interaction (HOI) sequence in HODome, including accurate tracking, high-quality geometry reconstruction and novel-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.347, + 0.892, + 0.513 + ], + "angle": 0, + "content": "view rendering. As illustrated in Fig. 3, given 76-view videos, we first perform a joint optimization scheme to accurately capture both the human skeletal motions and object rigid motion (Sec. 4.1). Then, based on the tracked human-object motions and shape priors, we decouple the humans and objects in HOI scenarios via a layer-wise neural human-object rendering scheme and a corresponding HOI-aware optimization strategy (Sec. 4.2). Such digital assets of humans, objects and the entire HOI scenes can naturally enable further geometry and appearance analysis of HOI scenarios." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.525, + 0.724, + 0.541 + ], + "angle": 0, + "content": "4.1. Human-object Tracking" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.549, + 0.892, + 0.579 + ], + "angle": 0, + "content": "Here we introduce our human-object tracking scheme with the aid of object markers." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.581, + 0.893, + 0.763 + ], + "angle": 0, + "content": "Tracking initialization. We initialize the human tracking process by fitting the SMPL-X model to the keypoints of each view. Note that we utilize the off-the-shell toolbox Easymocap [3] to get the initial per-frame parameters, i.e., pose \\(\\theta_{t}\\), shape \\(\\beta_{t}\\), facial expression \\(\\psi_{t}\\), and translation \\(\\gamma_{t}\\) on each frame \\(t\\). As for the object, following GRAB [61], we regard each object as rigid and only estimate the rotation \\(R_{t} \\in \\mathbb{S0}(3)\\) and rigid translation \\(T_{t} \\in \\mathbb{R}^{3}\\) with respect to its pre-scanned template. Specifically, we conduct the rigid-ICP technique [5, 81] to compute the rigid transformation between the per-frame markers, which serves as the initial object pose." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.764, + 0.894, + 0.84 + ], + "angle": 0, + "content": "Joint optimization for human-object tracking. With human and object pose initialization, we further perform a joint optimization scheme to ensure correct human-object contact. We impose constraints to ensure plausible interactions as the following form:" + }, + { + "type": "equation", + "bbox": [ + 0.517, + 0.852, + 0.892, + 0.901 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} E \\left(\\beta_ {t}, \\theta_ {t}, \\psi_ {t}, \\gamma_ {t}, R _ {t}, T _ {t}\\right) = E _ {\\text {s m p l}} + \\lambda_ {\\text {c o n t a c t}} E _ {\\text {c o n t a c t}} + \\\\ \\lambda_ {\\text {h o m a s k}} E _ {\\text {h o m a s k}} + \\lambda_ {\\text {m a k e r}} E _ {\\text {m a k e r}}, \\tag {1} \\\\ \\end{array}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8837" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.089, + 0.473, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.316, + 0.473, + 0.414 + ], + "angle": 0, + "content": "Figure 4. Layer-wise Neural Human-Object Rendering Details. We sample a ray uniformly in the bounding box of the performer and near the template for the object. We condition both the human and object layer with appearance codes \\((l_o^c, l_h^c)\\). The dynamic human layer contains an additional pose embedding \\((E_p)\\) conditioned deformation field. The sample colors and densities are merged, sorted and accumulated into pixel colors." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.441, + 0.47, + 0.563 + ], + "angle": 0, + "content": "where \\( E_{\\mathrm{spl}} \\) is the multi-view data fitting term describing the \\( \\ell_2 \\) distance between estimated joints and detected joints, following SMPLify-X [43] and PROX [16]. Besides, we impose contact term \\( E_{\\mathrm{contact}} \\), human-object silhouette loss term \\( E_{\\mathrm{homask}} \\), and object marker align term \\( E_{\\mathrm{maker}} \\) to ensure the plausibility of human-object interaction tracking. Due to page limitation, we have to defer more details of these terms in the Appendix." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.573, + 0.462, + 0.59 + ], + "angle": 0, + "content": "4.2. Layer-wise Neural Human-Object Rendering" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.597, + 0.47, + 0.642 + ], + "angle": 0, + "content": "Here we introduce our layer-wise image generation pipeline via neural rendering, without tedious manual efforts for geometry separation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.645, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Layered neural human-object modeling. For ease of instances separation, we represent the human-object interaction (HOI) scenes as a continuous layered neural radiance field, following ST-NeRF [77], as shown in Fig. 4. To leverage the shape prior imposed by the captured parametric SMPL-X model, we adopt a pose embedded dynamic NeRF similar to HumanNeRF [80] as the human layer. We use the skeletal pose to bridge the live frames to canonical space and an additional deformation MLP to learn subtle non-rigid deformation. Rather than using the pixel-aligned image feature, we leverage latent codes for time-varying human appearance capture to get rid of input images. Opposite to the human layer, we model objects as rigid static radiance fields in canonical space. The live frames are transformed to canonical space via object poses to maintain a globally consistent density field. Similarly, we adopt appearance latent codes for the time-varying shadow during interactions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.198 + ], + "angle": 0, + "content": "Dynamic human-object volume rendering. We utilize the layered volume rendering technique described in ST-NeRF [77] to render our neural human-object scenes. For a camera ray intersecting with the \\(i_{th}\\) entity at any times-tamp, we compute the ray segment as the depth of intersection points \\(d_f^j\\) and \\(d_f^i\\). We evenly partition each segment into \\(N\\) bins and sample one point uniformly from each bin:" + }, + { + "type": "equation", + "bbox": [ + 0.498, + 0.217, + 0.895, + 0.259 + ], + "angle": 0, + "content": "\\[\np _ {j} ^ {i} \\sim \\mathcal {U} \\left[ d _ {n} ^ {i} + \\frac {j - 1}{N} \\left(d _ {f} ^ {i} - d _ {n} ^ {i}\\right), d _ {n} ^ {i} + \\frac {j}{N} \\left(d _ {f} ^ {i} - d _ {n} ^ {i}\\right) \\right], j \\in [ 1, 2, \\dots , N ], \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.259, + 0.892, + 0.411 + ], + "angle": 0, + "content": "where \\(\\mathcal{U}\\) means uniform distribution. We impose the shape prior encoded in the tracked object template, on the ray sampling scheme for efficient training. We compute the ray-object intersection and uniformly sample a few points in the narrow segment where the first intersection point lies. For human, we simply sample in the ray-bounding box intersection segment. The samples from different segments are then merged and sorted by depth into \\(M\\) samples in total. We then compute a pixel's color by accumulating the radiance at sampled points:" + }, + { + "type": "equation", + "bbox": [ + 0.578, + 0.421, + 0.892, + 0.462 + ], + "angle": 0, + "content": "\\[\nC = \\sum_ {i = 1} ^ {M} T \\left(p _ {i}\\right) \\left[ \\left(1 - e ^ {- \\sigma_ {p _ {i}} \\delta_ {p _ {i}}}\\right) c _ {p _ {i}} \\right], \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.468, + 0.892, + 0.502 + ], + "angle": 0, + "content": "where \\(\\delta_{p_i}\\) is the distance between adjacent points, \\(c_{p_i}\\) and \\(\\sigma_{p_i}\\) are color and density, and \\(T(p_i) = \\prod_j^{i - 1}e^{-\\sigma_{p_j}\\delta_{p_j}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.502, + 0.892, + 0.547 + ], + "angle": 0, + "content": "HOI-aware training scheme. Here we introduce an effective optimization scheme to train our neural layer-wise HOI model. We first leverage a photometric loss:" + }, + { + "type": "equation", + "bbox": [ + 0.616, + 0.558, + 0.892, + 0.591 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c} = \\sum_ {\\mathbf {r} \\in \\mathcal {R}} \\| C _ {\\mathbf {r}} - \\hat {C} _ {\\mathbf {r}} \\| _ {2} ^ {2}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.603, + 0.891, + 0.634 + ], + "angle": 0, + "content": "where \\(\\mathbf{r}\\) is a ray in the training ray set \\(\\mathcal{R}\\), \\(C_{\\mathbf{r}}\\) and \\(\\hat{C}_{\\mathbf{r}}\\) are the corresponding rendered and observed colors." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.635, + 0.892, + 0.771 + ], + "angle": 0, + "content": "Note that in HOI scenarios, the human layer inevitably intersects with other object layers. This fact introduces ambiguities and cannot be addressed by a simple photometric loss. Thus we design three additional regularizers based on the tracked object template to alleviate this issue. To enforce the object layers to be solid surface radiance fields that do not contribute to human appearance, we adopt an occupancy and a sparsity regularizer inside and outside the object template, i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.512, + 0.781, + 0.892, + 0.815 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {o} = \\sum_ {p \\in \\mathcal {O}} \\left(\\Omega_ {-} (p, \\mathcal {M}) \\| e ^ {(- \\sigma_ {p})} \\| _ {2} ^ {2} + \\Omega_ {+} (p, \\mathcal {M}) \\| \\sigma_ {p} \\| _ {2} ^ {2}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.891, + 0.87 + ], + "angle": 0, + "content": "where \\(\\mathcal{M}\\) is the template mesh, \\(\\mathcal{O}\\) is a set of points randomly sampled in the object's bounding box, \\(\\Omega_{+}\\) and \\(\\Omega_{-}\\) indicate whether a point locates outside or inside the mesh." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Recall that we have encoded human-object contact prior by jointly tracking human-object at the skeletal-pose level" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8838" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.116, + 0.088, + 0.862, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.324, + 0.895, + 0.355 + ], + "angle": 0, + "content": "Figure 5. The geometry, MoCap and neural modeling results from our dataset. HODome includes various interaction sequences, such as \"holding a vase\", \"opening the cupboard\", \"sitting on the sofa\" and \"moving the table\"." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.366, + 0.473, + 0.442 + ], + "angle": 0, + "content": "(Eq. 1). However, due to the misalignment of SMPL shape and actual human geometry, the human body and clothes may still overlap the object space, which goes against real-world physics. Therefore, we implicitly constrain the deformation net to predict contact-aware non-rigid deformation:" + }, + { + "type": "equation", + "bbox": [ + 0.179, + 0.45, + 0.47, + 0.482 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {h} = \\sum_ {p \\in \\mathcal {H}} \\Omega_ {-} (p, \\mathcal {M}) \\| \\sigma_ {p} \\| _ {2} ^ {2}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.491, + 0.47, + 0.52 + ], + "angle": 0, + "content": "where \\(\\mathcal{H}\\) is the set of random samples in the human's bounding box." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.521, + 0.474, + 0.686 + ], + "angle": 0, + "content": "We further explore a weakly-supervising scheme that utilizes object texture cues to help decouple the human and object entities. The key observation is that, with accurate object tracking, the object radiance field fused across frames can be rendered with higher confidence. That is, we train the entire scene first and render the objects only to capture views, and the pixels similar to the captured ones are labeled as object pixels. In this way, we obtain coarse object segmentation maps \\( S \\) that serve as pseudo supervision: we adopt label-wise integration [53] to render the layer labels \\( s \\) of rays:" + }, + { + "type": "equation", + "bbox": [ + 0.158, + 0.686, + 0.47, + 0.726 + ], + "angle": 0, + "content": "\\[\n\\mathbf {s} = \\sum_ {i = 1} ^ {M} T \\left(p _ {i}\\right) \\left[ \\left(1 - e ^ {- \\sigma_ {p _ {i}} \\delta_ {p _ {i}}}\\right) \\mathbf {l} _ {p _ {i}} \\right], \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.73, + 0.468, + 0.762 + ], + "angle": 0, + "content": "where \\(\\mathbf{l}_{p_i}\\) is one-hot label indicates which layer the point \\(p_i\\) belongs to. We apply a semantic loss for object rays \\(S_o\\):" + }, + { + "type": "equation", + "bbox": [ + 0.196, + 0.769, + 0.47, + 0.801 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s} = \\sum_ {\\mathbf {r} \\in \\mathcal {S} _ {o}} \\| \\mathbf {s} _ {\\mathbf {r}} - \\hat {\\mathbf {s}} _ {\\mathbf {r}} \\| _ {2} ^ {2}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.809, + 0.47, + 0.84 + ], + "angle": 0, + "content": "where \\(\\hat{\\mathbf{s}}_{\\mathbf{r}}\\) is the pseudo semantic label of ray \\(\\mathbf{r}\\). Note that additional loss can be applied if accurate labels are provided." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.841, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Layer-wise neural representation in HODome. To further use the data, we can render reliable segmentation maps and occlusion-free appearances for humans and objects respectively. However, the renderings tend to be blurred since" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.366, + 0.895, + 0.532 + ], + "angle": 0, + "content": "the information of all the frames is fused. Hence we introduce an enhancement scheme to construct separate high-quality per-frame neural representation for human/object: we blend the visible regions of the high-fidelity observed input views to rendered human/object images using the segmentation maps. The Instant-NGP [38] and Instant-NSR [79] techniques are then applied to reconstruct separate human/object in tens of seconds. Finally, we obtain layer-wise representation which enables real-time photorealistic rendering and benefits the thorough analysis of HOI scenarios." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.534, + 0.895, + 0.656 + ], + "angle": 0, + "content": "Implementations. To build our neural representations, we train our layered models using Adam optimizer with a learning rate that starts from \\(5e^{-4}\\) and decays exponentially. Using the 76-view 4K resolution videos, the training process takes 8-10 hours on a single NVIDIA 3090 GPU for acceptable appearance and segmentation rendering. Thanks to Instant-NGP [38] and Instant-NSR [79], the layer-wise neural modeling process is conducted in seconds per frame." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.671, + 0.634, + 0.689 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.697, + 0.894, + 0.774 + ], + "angle": 0, + "content": "In this section, we compare our neural pipeline with existing state-of-the-arts. We first demonstrate the effectiveness and the generalization ability of our neural modeling approach (5.1). Next, we implement NeuralDome on three specific tasks (5.2)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.785, + 0.766, + 0.802 + ], + "angle": 0, + "content": "5.1. Analysis of Neural Rendering" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Comparisons. Here we compare our layer-wise neural human-object rendering approach, denoted as 'Ours-layer', with recent state-of-the-art methods, i.e., ST-NeRF [77], NeuralBody(NB) [45], in human-object interaction scenario. For fair comparisons, we select 20 cameras surrounding the performer for training and others for evaluation. We" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8839" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.087, + 0.473, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.335, + 0.47, + 0.417 + ], + "angle": 0, + "content": "Figure 6. Qualitative comparisons of novel view synthesis results. We show ground truth and synthesized images of novel view for NeuralBody [45], and ST-NeRF [77] and our layered human-object representation. Our approach achieves the best performance. We further illustrate our high-fidelity neural representations in HODome in the last column." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.422, + 0.466, + 0.512 + ], + "angle": 0, + "content": "
MethodsNB [45]ST-NeRF [77]Ours
ScenesPSNR↑SSIM↑PSNR↑SSIM↑PSNR↑SSIM↑
Bigsofa19.330.89624.020.88632.280.958
Sofa26.730.96528.490.95835.620.987
Table21.940.93322.440.90027.880.949
Average22.670.93124.990.91531.930.964
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.524, + 0.47, + 0.552 + ], + "angle": 0, + "content": "Table 2. Quantitative comparisons of novel view synthesized appearance on different human-object interaction sequences." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.567, + 0.47, + 0.718 + ], + "angle": 0, + "content": "remove the background layer and train ST-NeRF without the semantic label as ours. Fig. 6 shows several appearance synthesis results. NeuralBody has no ability to model objects which leads to artifacts in the density field. ST-NeRF fails as the bounding boxes of the performer and object almost completely overlap. Our method achieves both human and object modeling and rendering and thus further enables generating high-fidelity neural representation denoted as 'Ours-HODome'. Tab. 2 further illustrates our method significantly outperforms the baselines." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Evaluation. Here we further evaluate how our scheme contributes to the generated layer-wise neural representations. Fig. 7 shows the single human layer where the performer is \"sitting\" in the air. Note that we have no ground truth of the specific human/object layer, thus we conduct qualitative evaluations only. Let w/o pseudo segmentation and w/o blending denote our human assets generated without pseudo semantic loss and blending-based enhancement scheme. It demonstrates that the pseudo semantic loss effectively helps decouple the human and object layer and our blending scheme further boosts the appearance fidelity of the blurred temporal-fused model to a photo-realistic level." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.088, + 0.885, + 0.333 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.343, + 0.893, + 0.399 + ], + "angle": 0, + "content": "Figure 7. Qualitative evaluation of each strategy's contribution to layer-wise neural representation generation. The weakly-supervising scheme effectively helps decouple human and object and our blending scheme generates high-quality digital assets." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.402, + 0.893, + 0.462 + ], + "angle": 0, + "content": "
MethodSeparated evaluationJoint evaluation
MPJPE↓PA-MPJPE↓Chamferh↓Chamfero↓V2V↓p.V2V↓
Fit to input16.116.9486.6618.1461.5728.92
PHOSA [78]14.886.9477.6216.6454.5926.09
CHORE [73]10.216.1486.587.6944.9314.93
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.473, + 0.892, + 0.515 + ], + "angle": 0, + "content": "Table 3. Human object capture benchmark. \"Fit to input\" represents the vanilla method that fits the object template to image and capture human with Frankmocap [47]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.529, + 0.698, + 0.544 + ], + "angle": 0, + "content": "5.2. Task and benchmark" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.552, + 0.893, + 0.748 + ], + "angle": 0, + "content": "Human-object Capture Benchmark. HODome provides multi-view sequences with synchronized ground truth of capturing. To demonstrate the capability, we provide a benchmark for human-object capture, shown in Tab. 3. For detailed metric explanation please refer to [6, 73, 74]. PHOSA [78] used the contact map from hand-crafted annotations, suffering from coarse contact information. While CHORE [73] outperforms it with the distance fields predicted from neural networks. It shows the importance of a prepared human-object dataset that supports data-driven human-object capture model. Check out the full capturing benchmark of the dataset at https://arxiv.org/abs/2212.07626." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Human-object Geometry Reconstruction Benchmark. Benefiting from the dense-view setting, various human-only geometry reconstruction methods can step towards a human-object geometry reconstruction setting. We can benchmark the state-of-the-art algorithms in sparse view geometry reconstruction tasks. Hence, we evaluate PIFu [48] by training on a subset of HODome. As shown in Fig. 8, we compare the quality result between original PIFu and PIFu-trained on HODome and in-the-wild images. Compared to the original PIFu, our training set can improve the quality of" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8840" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.09, + 0.143, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.148, + 0.091, + 0.207, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.212, + 0.091, + 0.282, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.285, + 0.091, + 0.364, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.091, + 0.445, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.176, + 0.144, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.102, + 0.26, + 0.129, + 0.269 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image_caption", + "bbox": [ + 0.175, + 0.204, + 0.196, + 0.214 + ], + "angle": 0, + "content": "NA" + }, + { + "type": "image", + "bbox": [ + 0.238, + 0.171, + 0.3, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.224, + 0.26, + 0.281, + 0.269 + ], + "angle": 0, + "content": "PIFu-Origin" + }, + { + "type": "image", + "bbox": [ + 0.315, + 0.167, + 0.372, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.306, + 0.257, + 0.366, + 0.274 + ], + "angle": 0, + "content": "Monocular PIFu-Trained" + }, + { + "type": "image", + "bbox": [ + 0.391, + 0.171, + 0.448, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.38, + 0.257, + 0.441, + 0.273 + ], + "angle": 0, + "content": "Six View PIFu-Trained" + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.284, + 0.47, + 0.354 + ], + "angle": 0, + "content": "Figure 8. Qualitative evaluation on human-object reconstruction. We compare the original PIFu [48] with trained PIFu on both images from our dataset and in-the-wild images. The original PIFu fails to predict the object depth, and PIFu trained on our dataset can approximately predict the shape of the object." + }, + { + "type": "table", + "bbox": [ + 0.107, + 0.357, + 0.441, + 0.411 + ], + "angle": 0, + "content": "
MethodP2S ×10-4↓Chamfer ×10-4↓
Origin PIFu [48]38.72640.947
Monocular PIFu-trained14.65314.483
6-View PIFu-trained3.3764.901
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.422, + 0.469, + 0.464 + ], + "angle": 0, + "content": "Table 4. Geometry reconstruction benchmark. Origin PIFu [48] uses the pre-trained model to infer. Monocular PIFu is trained on a single view. 6-View PIFu is trained on 6-view inputs." + }, + { + "type": "table", + "bbox": [ + 0.153, + 0.471, + 0.395, + 0.537 + ], + "angle": 0, + "content": "
MethodPSNR↑SSIM↑
IBRnet [69]21.430.892
Neuray [30]23.340.909
NeuralHumanFVV [60]21.690.914
NeuralHOIFVV23.100.912
" + }, + { + "type": "table_caption", + "bbox": [ + 0.086, + 0.547, + 0.46, + 0.562 + ], + "angle": 0, + "content": "Table 5. Neural rendering benchmark in sparse-view setting." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.577, + 0.469, + 0.621 + ], + "angle": 0, + "content": "the reconstructed shape of the objects. For further quantitative analysis, we evaluate the performance using P2S and CD in Tab. 4." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.624, + 0.469, + 0.835 + ], + "angle": 0, + "content": "Sparse-View Human-object Rendering Benchmark. Our HODome supports various neural rendering tasks, even for human-object interactions. We provide a benchmark on sparse-view rendering tasks and evaluate on IBRNet [69], NeuRay [30] and NeuralHumanFVV [60]. Besides, we also provide a baseline, named NeuralHOIFVV, which uses the projection of trained PIFu's [48] results as depth input, then applies the Neural Blending method presented in [79] to obtain novel-view synthesis. More details about NeuralHOIFVV are referred to the Appendix. We analyze the above methods on HODome and in-the-wild images. The qualitative and quantitative results are shown in Fig. 9 and Tab. 5. Our proposed NeuralHOIFVV can obtain generative novel-view synthesis on HODome and in-the-wild inputs." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.847, + 0.204, + 0.861 + ], + "angle": 0, + "content": "5.3. Limitations" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Although NeuralDome can provide accurate capturing results and layer-wise neural representations, we also want" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.096, + 0.571, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.586, + 0.093, + 0.628, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.629, + 0.094, + 0.688, + 0.179 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.623, + 0.179, + 0.688, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.592, + 0.265, + 0.678, + 0.275 + ], + "angle": 0, + "content": "GT IBRNet" + }, + { + "type": "image", + "bbox": [ + 0.692, + 0.093, + 0.751, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.704, + 0.265, + 0.742, + 0.275 + ], + "angle": 0, + "content": "NeuRay" + }, + { + "type": "image", + "bbox": [ + 0.758, + 0.093, + 0.816, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.761, + 0.265, + 0.819, + 0.281 + ], + "angle": 0, + "content": "Neural HumanFVV" + }, + { + "type": "image", + "bbox": [ + 0.823, + 0.094, + 0.879, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.837, + 0.265, + 0.875, + 0.281 + ], + "angle": 0, + "content": "Neural HOIFVV" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.292, + 0.892, + 0.348 + ], + "angle": 0, + "content": "Figure 9. Qualitative evaluation on neural human-object rendering. We show the comparison of IBRNet [69], NeuRay [30], NeuralHumanFVV [60] and our proposed NeuralHOIFVV. We render one novel view between two input camera views." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.371, + 0.892, + 0.537 + ], + "angle": 0, + "content": "to highlight some potential limitations of this pipeline. First, NeuralDome only considers single-person interaction with objects and the single-person task is already quite challenging for current research. It is non-trivial to extend the current algorithm to multi-person settings, especially in a crowd scene. Secondly, the reconstruction of 3D holistic scene is not covered in our pipeline. We leave the joint modeling of humans, objects and scene in the future work. Moreover, our dataset was collected under fixed illumination conditions with few backgrounds variance, limiting its generalization ability for other environments." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.565, + 0.618, + 0.58 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.595, + 0.892, + 0.761 + ], + "angle": 0, + "content": "We have presented HODome, the first dataset to jointly capture and render human-object interaction using dense RGB cameras and the Optitrack system. To process the HODome dataset, we have further developed NeuralDome, a neural pipeline that can accurately track humans and objects, conduct layer-wise geometry reconstruction, and enable novel-view synthesis for both the subjects and objects. We believe the dataset could boost the development of capturing and rendering human-object interactions and is very valuable to the community for a wide range of human-object capturing and rendering tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgement This work was supported by the Shanghai Sailing Program (21YF1429400, 22YF1428800), Shanghai Local College Capacity Building Program (23010503100, 22010502800), NSFC programs (61976138, 61977047), the National Key Research and Development Program (2018YFB2100500), STCSM (2015F0203-000-06), SHMEC (2019-01-07-00-01-E00003) and Shanghai Frontiers Science Center of Humancentered Artificial Intelligence (ShangHAI)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "8841" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.117, + 0.468, + 0.143 + ], + "angle": 0, + "content": "[1] Naturalpoint, inc. motion capture systems. https://optitrack.com/. 6.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.146, + 0.468, + 0.172 + ], + "angle": 0, + "content": "[2] Reality capture. https://www.capturingreality.com/realitycapture. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.176, + 0.468, + 0.201 + ], + "angle": 0, + "content": "[3] Easymocap - make human motion capture easier. Github, 2021.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.205, + 0.468, + 0.273 + ], + "angle": 0, + "content": "[4] Kara-Ali Aliev, Artem Sevastopolsky, Maria Kolos, Dmitry Ulyanov, and Victor Lempitsky. Neural point-based graphics. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16, pages 696-712. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.276, + 0.468, + 0.317 + ], + "angle": 0, + "content": "[5] Paul J Besl and Neil D McKay. Method for registration of 3-d shapes. In Sensor fusion IV: control paradigms and data structures, volume 1611, pages 586-606. Spie, 1992. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.32, + 0.468, + 0.401 + ], + "angle": 0, + "content": "[6] Bharat Lal Bhatnagar, Xianghui Xie, Ilya A Petrov, Cristian Sminchisescu, Christian Theobalt, and Gerard Pons-Moll. Behave: Dataset and method for tracking human object interactions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15935-15946, 2022. 2, 3, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.404, + 0.468, + 0.471 + ], + "angle": 0, + "content": "[7] Hongrui Cai, Wanquan Feng, Xuetao Feng, Yan Wang, and Juyong Zhang. Neural surface reconstruction of dynamic scenes with monocular rgb-d camera. In Thirty-sixth Conference on Neural Information Processing Systems (NeurIPS), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.475, + 0.468, + 0.531 + ], + "angle": 0, + "content": "[8] Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Realtime multi-person 2d pose estimation using part affinity fields. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7291-7299, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.533, + 0.468, + 0.573 + ], + "angle": 0, + "content": "[9] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. arXiv preprint arXiv:2203.09517, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.576, + 0.468, + 0.644 + ], + "angle": 0, + "content": "[10] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14124-14133, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.647, + 0.468, + 0.714 + ], + "angle": 0, + "content": "[11] Alvaro Collet, Ming Chuang, Pat Sweeney, Don Gillett, Dennis Evseev, David Calabrese, Hugues Hoppe, Adam Kirk, and Steve Sullivan. High-quality streamable free-viewpoint video. ACM Transactions on Graphics (TOG), 34(4):69, 2015. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.718, + 0.468, + 0.785 + ], + "angle": 0, + "content": "[12] Mingsong Dou, Philip Davidson, Sean Ryan Fanello, Sameh Khamis, Adarsh Kowdle, Christoph Rhemann, Vladimir Tankovich, and Shahram Izadi. Motion2fusion: Real-time volumetric performance capture. ACM Trans. Graph., 36(6):246:1-246:16, Nov. 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.468, + 0.87 + ], + "angle": 0, + "content": "[13] Kaiwen Guo, Peter Lincoln, Philip Davidson, Jay Busch, Xueming Yu, Matt Whalen, Geoff Harvey, Sergio Orts-Escolano, Rohit Pandey, Jason Dourgarian, et al. The relightables: Volumetric performance capture of humans with realistic relighting. ACM Transactions on Graphics (ToG), 38(6):1–19, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.874, + 0.468, + 0.9 + ], + "angle": 0, + "content": "[14] Marc Habermann, Lingjie Liu, Weipeng Xu, Michael Zollhoefer, Gerard Pons-Moll, and Christian Theobalt. Real-time" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.117, + 0.468, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.094, + 0.892, + 0.119 + ], + "angle": 0, + "content": "deep dynamic characters. ACM Transactions on Graphics (TOG), 40(4):1-16, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[15] Nils Hasler, Carsten Stoll, Martin Sunkel, Bodo Rosenhahn, and H-P Seidel. A statistical model of human pose and body shape. In Computer graphics forum, volume 28, pages 337-346. Wiley Online Library, 2009. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.178, + 0.892, + 0.244 + ], + "angle": 0, + "content": "[16] Mohamed Hassan, Vasileios Choutas, Dimitrios Tzionas, and Michael J Black. Resolving 3d human pose ambiguities with 3d scene constraints. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2282-2292, 2019. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.247, + 0.892, + 0.313 + ], + "angle": 0, + "content": "[17] Tong He, Yuanlu Xu, Shunsuke Saito, Stefano Soatto, and Tony Tung. Arch++: Animation-ready clothed human reconstruction revisited. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11046-11056, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.317, + 0.892, + 0.383 + ], + "angle": 0, + "content": "[18] Yinghao Huang, Omid Taheri, Michael J Black, and Dimitrios Tzionas. Intercap: Joint markerless 3d tracking of humans and objects in interaction. In DAGM German Conference on Pattern Recognition, pages 281-299. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.386, + 0.892, + 0.453 + ], + "angle": 0, + "content": "[19] Zeng Huang, Yuanlu Xu, Christoph Lassner, Hao Li, and Tony Tung. Arch: Animatable reconstruction of clothed humans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3093-3102, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.456, + 0.892, + 0.524 + ], + "angle": 0, + "content": "[20] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3.6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE Transactions on Pattern Analysis and Machine Intelligence, 36(7):1325-1339, jul 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.526, + 0.892, + 0.607 + ], + "angle": 0, + "content": "[21] Yuheng Jiang, Suyi Jiang, Guoxing Sun, Zhuo Su, Kaiwen Guo, Minye Wu, Jingyi Yu, and Lan Xu. Neuralhofusion: Neural volumetric rendering under human-object interactions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6155-6165, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.609, + 0.892, + 0.677 + ], + "angle": 0, + "content": "[22] Hanbyul Joo, Tomas Simon, Xulong Li, Hao Liu, Lei Tan, Lin Gui, Sean Banerjee, Timothy Scott Godisart, Bart Nabbe, Iain Matthews, Takeo Kanade, Shohei Nobuhara, and Yaser Sheikh. Panoptic studio: A massively multiview system for social interaction capture. 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.679, + 0.892, + 0.746 + ], + "angle": 0, + "content": "[23] Youngjoong Kwon, Dahun Kim, Duygu Ceylan, and Henry Fuchs. Neural human performer: Learning generalizable radiance fields for human performance rendering. Advances in Neural Information Processing Systems, 34:24741-24752, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.749, + 0.892, + 0.816 + ], + "angle": 0, + "content": "[24] Christoph Lassner, Javier Romero, Martin Kiefel, Federica Bogo, Michael J. Black, and Peter V. Gehler. Unite the people: Closing the loop between 3d and 2d human representations. In IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), July 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.818, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[25] Ruilong Li, Julian Tanke, Minh Vo, Michael Zollhofer, Jurgen Gall, Angjoo Kanazawa, and Christoph Lassner. Tava: Template-free animatable volumetric actors. arXiv preprint arXiv:2206.08929, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.874, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[26] Jun Liu, Amir Shahroudy, Mauricio Perez, Gang Wang, Ling-Yu Duan, and Alex C Kot. Nturgb+ d 120: A large-" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.094, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8842" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "scale benchmark for 3d human activity understanding. IEEE transactions on pattern analysis and machine intelligence, 42(10):2684-2701, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.191 + ], + "angle": 0, + "content": "[27] Lingjie Liu, Marc Habermann, Viktor Rudnev, Kripasindhu Sarkar, Jiatao Gu, and Christian Theobalt. Neural actor: Neural free-view synthesis of human actors with pose control. ACM Trans. Graph. (ACM SIGGRAPH Asia), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.471, + 0.262 + ], + "angle": 0, + "content": "[28] Lingjie Liu, Weipeng Xu, Marc Habermann, Michael Zollhöfer, Florian Bernard, Hyeongwoo Kim, Wenping Wang, and Christian Theobalt. Neural human video rendering by learning dynamic textures and rendering-to-video translation. arXiv preprint arXiv:2001.04947, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.264, + 0.471, + 0.332 + ], + "angle": 0, + "content": "[29] Lingjie Liu, Weipeng Xu, Michael Zollhoefer, Hyeongwoo Kim, Florian Bernard, Marc Habermann, Wenping Wang, and Christian Theobalt. Neural rendering and reenactment of human actor videos. ACM Transactions on Graphics (TOG), 38(5):1-14, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.334, + 0.471, + 0.389 + ], + "angle": 0, + "content": "[30] Yuan Liu, Sida Peng, Lingjie Liu, Qianqian Wang, Peng Wang, Christian Theobalt, Xiaowei Zhou, and Wenping Wang. Neural rays for occlusion-aware image-based rendering. In CVPR, 2022. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.391, + 0.471, + 0.446 + ], + "angle": 0, + "content": "[31] Stephen Lombardi, Tomas Simon, Jason Saragih, Gabriel Schwartz, Andreas Lehrmann, and Yaser Sheikh. Neural volumes: Learning dynamic renderable volumes from images. arXiv preprint arXiv:1906.07751, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.448, + 0.47, + 0.503 + ], + "angle": 0, + "content": "[32] Stephen Lombardi, Tomas Simon, Gabriel Schwartz, Michael Zollhoefer, Yaser Sheikh, and Jason Saragih. Mixture of volumetric primitives for efficient neural rendering. arXiv preprint arXiv:2103.01954, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.505, + 0.471, + 0.546 + ], + "angle": 0, + "content": "[33] Xiaoxiao Long, Cheng Lin, Peng Wang, Taku Komura, and Wenping Wang. Sparseneus: Fast generalizable neural surface reconstruction from sparse views. ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.548, + 0.47, + 0.617 + ], + "angle": 0, + "content": "[34] H. Luo, A. Chen, Q. Zhang, B. Pang, M. Wu, L. Xu, and J. Yu. Convolutional neural opacity radiance fields. In 2021 IEEE International Conference on Computational Photography (ICCP), pages 1-12, Los Alamitos, CA, USA, may 2021. IEEE Computer Society. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.619, + 0.47, + 0.674 + ], + "angle": 0, + "content": "[35] Haimin Luo, Teng Xu, Yuheng Jiang, Chenglin Zhou, Qiwei Qiu, Yingliang Zhang, Wei Yang, Lan Xu, and Jingyi Yu. Artemis: Articulated neural pets with appearance and motion synthesis. ACM Trans. Graph., 41(4), jul 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.676, + 0.47, + 0.744 + ], + "angle": 0, + "content": "[36] Quan Meng, Anpei Chen, Haimin Luo, Minye Wu, Hao Su, Lan Xu, Xuming He, and Jingyi Yu. Gnerf: Gan-based neural radiance field without posed camera. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6351-6361, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.746, + 0.47, + 0.815 + ], + "angle": 0, + "content": "[37] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European conference on computer vision, pages 405-421. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.47, + 0.872 + ], + "angle": 0, + "content": "[38] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, July 2022. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.874, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[39] Michael Niemeyer, Jonathan T. Barron, Ben Mildenhall, Mehdi S. M. Sajjadi, Andreas Geiger, and Noha Radwan." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.893, + 0.134 + ], + "angle": 0, + "content": "Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.893, + 0.19 + ], + "angle": 0, + "content": "[40] Atsuhiro Noguchi, Xiao Sun, Stephen Lin, and Tatsuya Harada. Neural articulated radiance field. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5762-5772, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.893, + 0.232 + ], + "angle": 0, + "content": "[41] Anqi Pang, Xin Chen, Haimin Luo, Minye Wu, Jingyi Yu, and Lan Xu. Few-shot neural human performance rendering from sparse rgbd videos, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.234, + 0.893, + 0.301 + ], + "angle": 0, + "content": "[42] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M. Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. ACM Trans. Graph., 40(6), dec 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.303, + 0.894, + 0.385 + ], + "angle": 0, + "content": "[43] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3d hands, face, and body from a single image. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10975-10985, 2019. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.387, + 0.893, + 0.455 + ], + "angle": 0, + "content": "[44] Sida Peng, Junting Dong, Qianqian Wang, Shangzhan Zhang, Qing Shuai, Xiaowei Zhou, and Hujun Bao. Animatable neural radiance fields for modeling dynamic human bodies. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14314-14323, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.457, + 0.893, + 0.523 + ], + "angle": 0, + "content": "[45] Sida Peng, Yuanqing Zhang, Yinghao Xu, Qianqian Wang, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In CVPR, 2021. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.526, + 0.893, + 0.58 + ], + "angle": 0, + "content": "[46] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, jun 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.582, + 0.893, + 0.649 + ], + "angle": 0, + "content": "[47] Yu Rong, Takaaki Shiratori, and Hanbyul Joo. Frankmocap: A monocular 3d whole-body pose estimation system via regression and integration. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1749-1759, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.651, + 0.893, + 0.733 + ], + "angle": 0, + "content": "[48] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2304-2314, 2019. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.735, + 0.894, + 0.803 + ], + "angle": 0, + "content": "[49] Shunsuke Saito, Tomas Simon, Jason Saragih, and Hanbyul Joo. Pifuhd: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 84-93, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.805, + 0.893, + 0.859 + ], + "angle": 0, + "content": "[50] Manolis Savva, Angel X Chang, Pat Hanrahan, Matthew Fisher, and Matthias Nießner. Pigraphs: learning interaction snapshots from observations. ACM Transactions on Graphics (TOG), 35(4):1-12, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.861, + 0.893, + 0.901 + ], + "angle": 0, + "content": "[51] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-Motion Revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8843" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[52] Qing Shuai, Chen Geng, Qi Fang, Sida Peng, Wenhao Shen, Xiaowei Zhou, and Hujun Bao. Novel view synthesis of human interactions from sparse multi-view videos. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-10, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.164, + 0.472, + 0.248 + ], + "angle": 0, + "content": "[53] Qing Shuai, Chen Geng, Qi Fang, Sida Peng, Wenhao Shen, Xiaowei Zhou, and Hujun Bao. Novel view synthesis of human interactions from sparse multi-view videos. In ACM SIGGRAPH 2022 Conference Proceedings, SIGGRAPH '22, New York, NY, USA, 2022. Association for Computing Machinery. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.249, + 0.472, + 0.332 + ], + "angle": 0, + "content": "[54] Aliaksandra Shysheya, Egor Zakharov, Kara-Ali Aliev, Renat Bashirov, Egor Burkov, Karim Iskakov, Aleksei Ivakhnenko, Yury Malkov, Igor Pasechnik, Dmitry Ulyanov, et al. Textured neural avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2387-2397, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.334, + 0.472, + 0.403 + ], + "angle": 0, + "content": "[55] Leonid Sigal, Alexandru O Balan, and Michael J Black. Humaneva: Synchronized video and motion capture dataset and baseline algorithm for evaluation of articulated human motion. International journal of computer vision, 87(1):4-27, 2010. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.404, + 0.472, + 0.489 + ], + "angle": 0, + "content": "[56] Zhuo Su, Lan Xu, Zerong Zheng, Tao Yu, Yebin Liu, and Lu Fang. Robustfusion: Human volumetric capture with data-driven visual cues using a rgbd camera. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, Computer Vision – ECCV 2020, pages 246–264, Cham, 2020. Springer International Publishing. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.49, + 0.472, + 0.559 + ], + "angle": 0, + "content": "[57] Guoxing Sun, Xin Chen, Yizhang Chen, Anqi Pang, Pei Lin, Yuheng Jiang, Lan Xu, Jingya Wang, and Jingyi Yu. Neural free-viewpoint performance rendering under complex human-object interactions. In Proceedings of the 29th ACM International Conference on Multimedia, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.561, + 0.472, + 0.643 + ], + "angle": 0, + "content": "[58] Guoxing Sun, Xin Chen, Yizhang Chen, Anqi Pang, Pei Lin, Yuheng Jiang, Lan Xu, Jingyi Yu, and Jingya Wang. Neural free-viewpoint performance rendering under complex human-object interactions. In Proceedings of the 29th ACM International Conference on Multimedia, pages 4651-4660, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.645, + 0.472, + 0.729 + ], + "angle": 0, + "content": "[59] Xin Suo, Yuheng Jiang, Pei Lin, Yingliang Zhang, Minye Wu, Kaiwen Guo, and Lan Xu. Neuralhumanfvv: Realtime neural volumetric human performance rendering using rgb cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6226-6237, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.731, + 0.472, + 0.801 + ], + "angle": 0, + "content": "[60] Xin Suo, Yuheng Jiang, Pei Lin, Yingliang Zhang, Minye Wu, Kaiwen Guo, and Lan Xu. Neuralhumanfvv: Real-time neural volumetric human performance rendering using rgb cameras. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6222-6233, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.472, + 0.858 + ], + "angle": 0, + "content": "[61] Omid Taheri, Nima Ghorbani, Michael J Black, and Dimitrios Tzionas. Grab: A dataset of whole-body human grasping of objects. In European conference on computer vision, pages 581-600. Springer, 2020. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.472, + 0.902 + ], + "angle": 0, + "content": "[62] Ayush Tewari, Justus Thies, Ben Mildenhall, Pratul Srinivasan, Edgar Tretschk, Yifan Wang, Christoph Lassner, Vincent Sitzmann, Ricardo Martin-Brualla, Stephen Lom-" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.148 + ], + "angle": 0, + "content": "bardi, Tomas Simon, Christian Theobalt, Matthias Niessner, Jonathan T. Barron, Gordon Wetzstein, Michael Zollhoefer, and Vladislav Golyanik. Advances in neural rendering, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.894, + 0.233 + ], + "angle": 0, + "content": "[63] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhofer, Christoph Lassner, and Christian Theobalt. Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12959-12970, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.234, + 0.894, + 0.304 + ], + "angle": 0, + "content": "[64] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T Barron, and Pratul P Srinivasan. Ref-nerf: Structured view-dependent appearance for neural radiance fields. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5481-5490. IEEE, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.305, + 0.894, + 0.375 + ], + "angle": 0, + "content": "[65] Timo Von Marcard, Roberto Henschel, Michael J Black, Bodo Rosenhahn, and Gerard Pons-Moll. Recovering accurate 3d human pose in the wild using imus and a moving camera. In Proceedings of the European Conference on Computer Vision (ECCV), pages 601-617, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.376, + 0.894, + 0.445 + ], + "angle": 0, + "content": "[66] Liao Wang, Ziyu Wang, Pei Lin, Yuheng Jiang, Xin Suo, Minye Wu, Lan Xu, and Jingyi Yu. ibutter: Neural interactive bullet time generator for human free-viewpoint rendering. In Proceedings of the 29th ACM International Conference on Multimedia, pages 4641-4650, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.447, + 0.894, + 0.53 + ], + "angle": 0, + "content": "[67] Liao Wang, Jiakai Zhang, Xinhang Liu, Fuqiang Zhao, Yanshun Zhang, Yingliang Zhang, Minye Wu, Jingyi Yu, and Lan Xu. Fourier plenoctrees for dynamic radiance field rendering in real-time. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13524-13534, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.531, + 0.894, + 0.586 + ], + "angle": 0, + "content": "[68] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.588, + 0.894, + 0.657 + ], + "angle": 0, + "content": "[69] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul Srinivasan, Howard Zhou, Jonathan T. Barron, Ricardo MartinBrualla, Noah Snively, and Thomas Funkhouser. Ibrnet: Learning multi-view image-based rendering. In CVPR, 2021. 2,8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.659, + 0.894, + 0.701 + ], + "angle": 0, + "content": "[70] Shaofei Wang, Katja Schwarz, Andreas Geiger, and Siyu Tang. Arah: animatable volume rendering of articulated human sdfs. arXiv preprint arXiv:2210.10036, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.702, + 0.894, + 0.758 + ], + "angle": 0, + "content": "[71] Minye Wu, Yuehao Wang, Qiang Hu, and Jingyi Yu. Multi-view neural human rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.759, + 0.894, + 0.815 + ], + "angle": 0, + "content": "[72] Minye Wu, Yuehao Wang, Qiang Hu, and Jingyi Yu. Multi-view neural human rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1682-1691, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.816, + 0.894, + 0.858 + ], + "angle": 0, + "content": "[73] Xianghui Xie, Bharat Lal Bhatnagar, and Gerard Pons-Moll. Chore: Contact, human and object reconstruction from a single rgb image. arXiv preprint arXiv:2204.02445, 2022. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.894, + 0.902 + ], + "angle": 0, + "content": "[74] Hongwei Yi, Chun-Hao P Huang, Dimitrios Tzionas, Muhammed Kocabas, Mohamed Hassan, Siyu Tang, Justus Thies, and Michael J Black. Human-aware object placement" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "8844" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.093, + 0.47, + 0.134 + ], + "angle": 0, + "content": "for visual environment reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3959-3970, 2022. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.176 + ], + "angle": 0, + "content": "[75] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. Plenoctrees for real-time rendering of neural radiance fields. In arXiv, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.469, + 0.233 + ], + "angle": 0, + "content": "[76] Chao Zhang, Sergi Pujades, Michael J. Black, and Gerard Pons-Moll. Detailed, accurate, human shape estimation from clothed 3d scan sequences. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), July 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.235, + 0.469, + 0.303 + ], + "angle": 0, + "content": "[77] Jiakai Zhang, Xinhang Liu, Xinyi Ye, Fuqiang Zhao, Yanshun Zhang, Minye Wu, Yingliang Zhang, Lan Xu, and Jingyi Yu. Editable free-viewpoint video using a layered neural representation. ACM Trans. Graph., 40(4), July 2021. 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.305, + 0.469, + 0.374 + ], + "angle": 0, + "content": "[78] Jason Y Zhang, Sam Pepose, Hanbyul Joo, Deva Ramanan, Jitendra Malik, and Angjoo Kanazawa. Perceiving 3d human-object spatial arrangements from a single image in the wild. In European conference on computer vision, pages 34-51. Springer, 2020. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.376, + 0.469, + 0.444 + ], + "angle": 0, + "content": "[79] Fuqiang Zhao, Yuheng Jiang, Kaixin Yao, Jiakai Zhang, Liao Wang, Haizhao Dai, Yuhui Zhong, Yingliang Zhang, Minye Wu, Lan Xu, et al. Human performance modeling and rendering via neural animated mesh. arXiv preprint arXiv:2209.08468, 2022. 2, 4, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.446, + 0.469, + 0.515 + ], + "angle": 0, + "content": "[80] Fuqiang Zhao, Wei Yang, Jiakai Zhang, Pei Lin, Yingliang Zhang, Jingyi Yu, and Lan Xu. Humannerf: Efficiently generated human radiance field from sparse inputs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7743-7753, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.517, + 0.469, + 0.557 + ], + "angle": 0, + "content": "[81] Qian-Yi Zhou, Jaesik Park, and Vladlen Koltun. Open3d: A modern library for 3d data processing. arXiv preprint arXiv:1801.09847, 2018. 4" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.47, + 0.557 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.515, + 0.956 + ], + "angle": 0, + "content": "8845" + } + ] +] \ No newline at end of file diff --git a/2023/NeuralDome_ A Neural Modeling Pipeline on Multi-View Human-Object Interactions/fa0c20cd-9328-405b-b73e-2045fdfa714d_origin.pdf b/2023/NeuralDome_ A Neural Modeling Pipeline on Multi-View Human-Object Interactions/fa0c20cd-9328-405b-b73e-2045fdfa714d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ccd2e100e196ceda7807926d3656e09e6a3cc1e4 --- /dev/null +++ b/2023/NeuralDome_ A Neural Modeling Pipeline on Multi-View Human-Object Interactions/fa0c20cd-9328-405b-b73e-2045fdfa714d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fff205d2f1adad80996e98756608b180e9d0f44ef7a472afeaf80e3f20d9590f +size 6114228 diff --git a/2023/NeuralDome_ A Neural Modeling Pipeline on Multi-View Human-Object Interactions/full.md b/2023/NeuralDome_ A Neural Modeling Pipeline on Multi-View Human-Object Interactions/full.md new file mode 100644 index 0000000000000000000000000000000000000000..795e8eec892852ffa6c01fff11f353a9bda7b1da --- /dev/null +++ b/2023/NeuralDome_ A Neural Modeling Pipeline on Multi-View Human-Object Interactions/full.md @@ -0,0 +1,365 @@ +# NeuralDome: A Neural Modeling Pipeline on Multi-View Human-Object Interactions + +Juze Zhang $^{1,2,3,4,*}$ , Haimin Luo $^{1,4,5,*}$ , Hongdi Yang $^{1,4}$ , Xinru Xu $^{1,4}$ , Qianyang Wu $^{1,4}$ , Ye Shi $^{1,4}$ , Jingyi Yu $^{1,4}$ , Lan Xu $^{1,4,\dagger}$ , Jingya Wang $^{1,4,\dagger}$ + +$^{1}$ ShanghaiTech University $^{2}$ Shanghai Advanced Research Institute, Chinese Academy of Sciences $^{3}$ University of Chinese Academy of Sciences +$^{4}$ Shanghai Engineering Research Center of Intelligent Vision and Imaging $^{5}$ LumiAni Technology $\{\mathrm{zhangjz,luohm,yanghd,xuxr2022,wuqy2022,shiye,yujingyi,xulan1,wangjingya}\} @shanghaiitech.edu.cn$ + +![](images/a6ddac9037905908066efb62551cd861d33d9aae1ede8cb2fffc35ad11317161.jpg) +Figure 1. Our NeuralDome pipeline for processing multi-view video sequences on human object interactions. NeuralDome supports tracking, modeling, and rendering of individual human subjects and objects. To validate NeuralDome, we collect a large dataset HODome over a total of 71 million video frames across 76 viewpoints and process the datasets using NeuralDome for a variety of inference and neural modeling and rendering tasks. + +# Abstract + +Humans constantly interact with objects in daily life tasks. Capturing such processes and subsequently conducting visual inferences from a fixed viewpoint suffers from occlusions, shape and texture ambiguities, motions, etc. To mitigate the problem, it is essential to build a training dataset that captures free-viewpoint interactions. We construct a dense multi-view dome to acquire a complex human object interaction dataset, named HODome, that consists of $\sim 71M$ frames on 10 subjects interacting with 23 objects. To process the HODome dataset, we develop NeuralDome, a layer-wise neural processing pipeline tai + +lored for multi-view video inputs to conduct accurate tracking, geometry reconstruction and free-view rendering, for both human subjects and objects. Extensive experiments on the HODome dataset demonstrate the effectiveness of NeuralDome on a variety of inference, modeling, and rendering tasks. Both the dataset and the NeuralDome tools will be disseminated to the community for further development, which can be found at https://juzezhang.github.io/NeuralDome + +# 1. Introduction + +A key task of computer vision is to understand how humans interact with the surrounding world, by faithfully capturing and subsequently reproducing the process via mod + +eling and rendering. Successful solutions benefit broad applications ranging from sports training to vocational education, digital entertainment to tele-medicine. + +Early solutions [11, 12, 51] that reconstruct dynamic meshes with per-frame texture maps are time-consuming and vulnerable to occlusions or lack of textures. Recent advances in neural rendering [37, 62, 71] bring huge potential for human-centric modeling. Most notably, the variants of Neural Radiance Field (NeRF) [37] achieve compelling novel view synthesis, which can enable real-time rendering performance [38, 59, 67] even for dynamic scenes [45, 63, 77], and can be extended to the generative setting without per-scene training [23, 69, 80]. However, less attention is paid to the rich and diverse interactions between humans and objects, mainly due to the severe lack of dense-view human-object datasets. Actually, existing datasets of human-object interactions are mostly based on optical markers [61] or sparse RGB/RGBD sensors [6, 26], without sufficient appearance supervision for neural rendering tasks. As a result, the literature on neural human-object rendering [21,57] is surprisingly sparse, let alone further exploring the real-time or generative directions. Besides, existing neural techniques [52, 77] suffer from tedious training procedures due to the human-object occlusion, and hence infeasible for building a large-scale dataset. In a nutshell, despite the recent tremendous thriving of neural rendering, the lack of both a rich dataset and an efficient reconstruction scheme constitute barriers in human-object modeling. + +In this paper, we present NeuralDome, a neural pipeline that takes multi-view dome capture as inputs and conducts accurate 3D modeling and photo-realistic rendering of complex human-object interaction. As shown in Fig. 1, NeuralDome exploits layer-wise neural modeling to produce rich and multi-modality outputs including the geometry of dynamic human, object shapes and tracked poses, as well as a free-view rendering of the sequence. + +Specifically, we first capture a novel human-object dome (HODome) dataset that consists of 274 human-object interacting sequences, covering 23 diverse 3D objects and 10 human subjects (5 males and 5 females) in various apparels. We record multi-view video sequences of natural interactions between the human subjects and the objects where each sequence is about 60s in length using a dome with 76 RGB cameras, resulting in 71 million video frames. We also provide an accurate pre-scanned 3D template for each object and utilize sparse optical markers to track individual objects throughout the sequences. + +To process the HODome dataset, we adopt an extended Neural Radiance Field (NeRF) pipeline. The brute-force adoption of off-the-shelf neural techniques such as Instant-NSR [79] and Instant-NGP [38], although effective, do not separate objects from human subjects and therefore lack sufficient fidelity to model their interactions. We instead + +introduce a layer-wise neural processing pipeline. Specifically, we first perform a joint optimization based on the dense inputs for accurately tracking human motions using the parametric SMPL-X model [43] as well as localizing objects using template meshes. We then propose an efficient layer-wise neural rendering scheme where the humans and objects are formulated as a pose-embedded dynamic NeRF and a static NeRF with tracked 6-DoF rigid poses, respectively. Such a layer-wise representation effectively exploits temporal information and robustly tackles the occluded regions under interactions. We further introduce an object-aware ray sampling strategy to mitigate artifacts during layer-wise training, as well as template-aware geometry regularizers to enforce contact-aware deformations. Through weak segmentation supervision, we obtain the decoupled and occlusion-free appearances for both the humans and the objects at a high fidelity amenable for training the input multi-view inputs for a variety of tasks from monocular motion capture to free-view rendering from sparse multi-view inputs. + +To summarize, our main contributions include: + +- We introduce NeuralDome, a neural pipeline, to accurately track humans and objects, conduct layer-wise geometry reconstruction, and enable novel-view synthesis, from multi-view HOI video inputs. +- We collect a comprehensive dataset that we call HODome that will be disseminated to the community, with both raw data and the output modalities including separated geometry and rendering of individual objects and human subjects, their tracking results, free-view rendering results, etc. +- We demonstrate using the dataset to train networks for a variety of visual inference tasks with complex human object interactions. + +# 2. Related Works + +# 2.1. Neural Human Rendering + +Various 3D data representations have been explored for neural human rendering, such as point-clouds [4, 41, 72], textured meshes [14, 28, 29, 54], and volumes [31, 32]. Implicit occupancy function-based methods [17, 19, 48, 49] can recover detailed 3D human geometry from sparse 2D images without faithful appearance synthesis. The recent NeRF [37] technique brings huge potential for 3D photorealistic view synthesis [9, 10, 30, 34, 36, 38, 39, 64, 69, 75] and geometry modeling [7, 33, 68]. Further explorations extend it to dynamic scenes [7, 35, 42, 46, 67, 77], especially for humans. Existing works equip NeRF with pose-embeddings [23, 27, 40, 45, 80], learnable skinning weights [25, 44, 70] and even generalization across individuals [23, 66, 80]. However, these works only focus on + +
Datasets#Cam-view# Frame(M)ResolutionFpsMarkerObj. Num.Human Annot.Novel viewsGeometryHuman TexturedNeural RepresentationObject Appearance
NTU [26]3~ 341920 × 108030×NA✓NA××××
PiGr [50]10.1960 × 5405×NA✓NA×××✓
GRAB [61]×NANANA✓51××××××
PROX [16]10.11920 × 108030✓NA✓××××✓
BEHAVE [6]40.152048 × 153630×20×××××✓
InterCap [18]60.071920 × 108030×10✓××××✓
Our76713840 × 216060✓23✓✓✓✓✓✓
+ +Table 1. Dataset Comparisons. We compare our proposed HODome dataset with the existing publicly available human-object dataset. HODome has the largest scale of human-object interactions in terms of the number of frames (#Frame), camera view (#Cam-view), and modality. "Obj. Num." represents the object number. "Human Annot." represents annotation from professional annotators. + +![](images/e0899a0cd871ac212bc82fc31dd32ab46ad0a0d3b0462149c3af250f16a9e4f8.jpg) +Figure 2. HODome Modality. HODome features multiple modalities of data format and annotations, including a) Color image, b) Human-object mask (Homask), c) MoCap (SMPL-X parameters and located object template), d) Geometry, e) digital assets in neural representations, f) Novel view and g) Keypoints (25 for bodies and 42 for hands with human annotation). + +a single person. The most relevant works [53, 77] aim to model multi-person interactions, but they cannot handle more complex interactions due to the lack of large-scale human-object interactions dataset. + +# 2.2. Human-object Modeling + +Only a few works [6, 16, 18, 61, 73, 78] consider to jointly model whole-body interactions. Early solutions [11, 12, 51] that reconstruct meshes with per-frame texture maps, which are vulnerable to occlusions and lack of textures. Recently, several works [6, 16, 18, 61] explore the relationship via several interaction constraints, such as contact map [6, 18, 61], spatial arrangement [78] and physically plausible constraint [74]. However, these methods only produce a relative arrangement. The most relevant works [21, 56, 58] aim to render human-object interactions with volumetric fusion [56], neural texturing blending [58] and volumetric rendering [21]. However, without sufficient appearance supervision, their quality is limited. Comparably, our layerwise neural representations can be converted to high-quality geometry and support separate free-view rendering for human or object. + +# 2.3. Human-centric Dataset + +A variety of human-centric datasets have been developed for human-only capturing or rendering tasks. Early datasets combine multi-camera RGB video capture with synchronized ground-truth 3D skeletons [20, 22, 55], parametric model [24, 65], scanned mesh [15, 76], without human-object interaction. Recent datasets capture human-object interactions using optical markers [61] or sparse RGB sensors [6, 26], without sufficient appearance supervision. As a result, the literature on neural human-object rendering [21, 57] is surprisingly rare. High-end works [11, 13] use dense cameras for reconstruction and rendering of humans and objects through mesh reconstruction and motion tracking, but without texture and neural representations. To fill this gap, we propose NeuralDome for capturing and rendering human-object interactions, facilitating various generative human-object tasks. + +# 3. HODome Dataset + +We present the HODome dataset for capturing and rendering photo-realistic human-object interactions. It consists of 274 human-object interacting sequences, covering 23 diverse 3D objects and 10 subjects (5 males and 5 females) under various apparels. For each sequence, we record the naturally interacting scene for about 60s using 76 RGB cameras for dense-view at $3840 \times 2160$ resolution and 60 frame-per-seconds (fps), resulting in roughly 71 million frames. HODome consists of rich labels covering different aspects of HOI capturing and rendering labels (see Fig.2). See Tab. 1 for comparison with other datasets. + +# 3.1. Data Capturing System + +To construct the HODome dataset, we use 76 Z-CAM cinema cameras with sufficient appearance supervision for neural rendering tasks and 16 Optitrack MoCap cameras [1] for accurate human-object tracking tasks. The Z-CAM system and Optitrack system are synchronized to record the RGB and MoCap data together. We use a publicly available tool [2] to estimate the intrinsic camera parameters and extrinsic camera parameters. + +![](images/50ae51d5779015aa428705f59a1fa4f33feb545fe7b54a145b6f8f2f4479273f.jpg) +Figure 3. Overview of NeuralDome. Given the 76-view RGB stream as input, we first jointly track human skeletal motion and object rigid motion. Then with the tracked motion priors, we decouple the human-object interaction scenes via a layer-wise neural rendering scheme, to generate human/object renderings separately and corresponding acceptable segmentation maps. Blending with the high-fidelity captured views, we obtain the layer-wise neural representations in our HODome dataset. + +# 3.2. Dataset Modality + +Human MoCap and Object 6D Pose. For appearance realism, we do not place markers on human actors. Thus we detect the human joints from 76 RGB images by running the whole-body Openpose [8] to perform markerless MoCap. And we treat each object as rigid and solve the rigid object pose estimation from markers using the Iterative Closest Point(ICP) algorithm [5, 81]. See Fig. 3 for the pipeline of our method. + +Human-Object Neural Representations. Under the dense-view setting, one can apply off-the-shell neural techniques, i.e., Instant-NSR [79] and Instant-NGP [38] to efficiently obtain per-frame geometry and novel-view synthesis of the whole human-object sequence, respectively. However, neither schemes separate object from human subject. We therefore propose a layer-wise neural scheme to separately recover neural representations of human and object, enabling both high-fidelity geometry reconstruction and free-view appearance rendering (Sec. 4.2). + +Data Annotation. Our dataset includes 23 objects with varying scales and interaction types. Each object was pre-scanned using off-the-shelf multi-view software packages [2]. Further, following the previous methods [6, 61], we annotated its pseudo contact label that computes from a threshold distance. To provide more accurate results for the quantitative benchmark, we annotate a separate quantitative subset as our test set with human-annotated segmentation and hand joints by the professional annotator. Our datasets will be available for research purposes. + +# 4. Neural Modeling on HODome + +We introduce a neural pipeline to produce the rich digital assets from the dense-view input of each human-object interaction (HOI) sequence in HODome, including accurate tracking, high-quality geometry reconstruction and novel- + +view rendering. As illustrated in Fig. 3, given 76-view videos, we first perform a joint optimization scheme to accurately capture both the human skeletal motions and object rigid motion (Sec. 4.1). Then, based on the tracked human-object motions and shape priors, we decouple the humans and objects in HOI scenarios via a layer-wise neural human-object rendering scheme and a corresponding HOI-aware optimization strategy (Sec. 4.2). Such digital assets of humans, objects and the entire HOI scenes can naturally enable further geometry and appearance analysis of HOI scenarios. + +# 4.1. Human-object Tracking + +Here we introduce our human-object tracking scheme with the aid of object markers. + +Tracking initialization. We initialize the human tracking process by fitting the SMPL-X model to the keypoints of each view. Note that we utilize the off-the-shell toolbox Easymocap [3] to get the initial per-frame parameters, i.e., pose $\theta_{t}$ , shape $\beta_{t}$ , facial expression $\psi_{t}$ , and translation $\gamma_{t}$ on each frame $t$ . As for the object, following GRAB [61], we regard each object as rigid and only estimate the rotation $R_{t} \in \mathbb{S0}(3)$ and rigid translation $T_{t} \in \mathbb{R}^{3}$ with respect to its pre-scanned template. Specifically, we conduct the rigid-ICP technique [5, 81] to compute the rigid transformation between the per-frame markers, which serves as the initial object pose. + +Joint optimization for human-object tracking. With human and object pose initialization, we further perform a joint optimization scheme to ensure correct human-object contact. We impose constraints to ensure plausible interactions as the following form: + +$$ +\begin{array}{l} E \left(\beta_ {t}, \theta_ {t}, \psi_ {t}, \gamma_ {t}, R _ {t}, T _ {t}\right) = E _ {\text {s m p l}} + \lambda_ {\text {c o n t a c t}} E _ {\text {c o n t a c t}} + \\ \lambda_ {\text {h o m a s k}} E _ {\text {h o m a s k}} + \lambda_ {\text {m a k e r}} E _ {\text {m a k e r}}, \tag {1} \\ \end{array} +$$ + +![](images/8fc6c6ae5126ab084ccaa7a81b4d86b3a38a6acf3098c59b56958722e75a476a.jpg) +Figure 4. Layer-wise Neural Human-Object Rendering Details. We sample a ray uniformly in the bounding box of the performer and near the template for the object. We condition both the human and object layer with appearance codes $(l_o^c, l_h^c)$ . The dynamic human layer contains an additional pose embedding $(E_p)$ conditioned deformation field. The sample colors and densities are merged, sorted and accumulated into pixel colors. + +where $E_{\mathrm{spl}}$ is the multi-view data fitting term describing the $\ell_2$ distance between estimated joints and detected joints, following SMPLify-X [43] and PROX [16]. Besides, we impose contact term $E_{\mathrm{contact}}$ , human-object silhouette loss term $E_{\mathrm{homask}}$ , and object marker align term $E_{\mathrm{maker}}$ to ensure the plausibility of human-object interaction tracking. Due to page limitation, we have to defer more details of these terms in the Appendix. + +# 4.2. Layer-wise Neural Human-Object Rendering + +Here we introduce our layer-wise image generation pipeline via neural rendering, without tedious manual efforts for geometry separation. + +Layered neural human-object modeling. For ease of instances separation, we represent the human-object interaction (HOI) scenes as a continuous layered neural radiance field, following ST-NeRF [77], as shown in Fig. 4. To leverage the shape prior imposed by the captured parametric SMPL-X model, we adopt a pose embedded dynamic NeRF similar to HumanNeRF [80] as the human layer. We use the skeletal pose to bridge the live frames to canonical space and an additional deformation MLP to learn subtle non-rigid deformation. Rather than using the pixel-aligned image feature, we leverage latent codes for time-varying human appearance capture to get rid of input images. Opposite to the human layer, we model objects as rigid static radiance fields in canonical space. The live frames are transformed to canonical space via object poses to maintain a globally consistent density field. Similarly, we adopt appearance latent codes for the time-varying shadow during interactions. + +Dynamic human-object volume rendering. We utilize the layered volume rendering technique described in ST-NeRF [77] to render our neural human-object scenes. For a camera ray intersecting with the $i_{th}$ entity at any times-tamp, we compute the ray segment as the depth of intersection points $d_f^j$ and $d_f^i$ . We evenly partition each segment into $N$ bins and sample one point uniformly from each bin: + +$$ +p _ {j} ^ {i} \sim \mathcal {U} \left[ d _ {n} ^ {i} + \frac {j - 1}{N} \left(d _ {f} ^ {i} - d _ {n} ^ {i}\right), d _ {n} ^ {i} + \frac {j}{N} \left(d _ {f} ^ {i} - d _ {n} ^ {i}\right) \right], j \in [ 1, 2, \dots , N ], \tag {2} +$$ + +where $\mathcal{U}$ means uniform distribution. We impose the shape prior encoded in the tracked object template, on the ray sampling scheme for efficient training. We compute the ray-object intersection and uniformly sample a few points in the narrow segment where the first intersection point lies. For human, we simply sample in the ray-bounding box intersection segment. The samples from different segments are then merged and sorted by depth into $M$ samples in total. We then compute a pixel's color by accumulating the radiance at sampled points: + +$$ +C = \sum_ {i = 1} ^ {M} T \left(p _ {i}\right) \left[ \left(1 - e ^ {- \sigma_ {p _ {i}} \delta_ {p _ {i}}}\right) c _ {p _ {i}} \right], \tag {3} +$$ + +where $\delta_{p_i}$ is the distance between adjacent points, $c_{p_i}$ and $\sigma_{p_i}$ are color and density, and $T(p_i) = \prod_j^{i - 1}e^{-\sigma_{p_j}\delta_{p_j}}$ + +HOI-aware training scheme. Here we introduce an effective optimization scheme to train our neural layer-wise HOI model. We first leverage a photometric loss: + +$$ +\mathcal {L} _ {c} = \sum_ {\mathbf {r} \in \mathcal {R}} \| C _ {\mathbf {r}} - \hat {C} _ {\mathbf {r}} \| _ {2} ^ {2}, \tag {4} +$$ + +where $\mathbf{r}$ is a ray in the training ray set $\mathcal{R}$ , $C_{\mathbf{r}}$ and $\hat{C}_{\mathbf{r}}$ are the corresponding rendered and observed colors. + +Note that in HOI scenarios, the human layer inevitably intersects with other object layers. This fact introduces ambiguities and cannot be addressed by a simple photometric loss. Thus we design three additional regularizers based on the tracked object template to alleviate this issue. To enforce the object layers to be solid surface radiance fields that do not contribute to human appearance, we adopt an occupancy and a sparsity regularizer inside and outside the object template, i.e., + +$$ +\mathcal {L} _ {o} = \sum_ {p \in \mathcal {O}} \left(\Omega_ {-} (p, \mathcal {M}) \| e ^ {(- \sigma_ {p})} \| _ {2} ^ {2} + \Omega_ {+} (p, \mathcal {M}) \| \sigma_ {p} \| _ {2} ^ {2}\right), \tag {5} +$$ + +where $\mathcal{M}$ is the template mesh, $\mathcal{O}$ is a set of points randomly sampled in the object's bounding box, $\Omega_{+}$ and $\Omega_{-}$ indicate whether a point locates outside or inside the mesh. + +Recall that we have encoded human-object contact prior by jointly tracking human-object at the skeletal-pose level + +![](images/d0913aeafdcba991c57124898a981e3dace7c18fb8fd34efe736909eabb7b8e0.jpg) +Figure 5. The geometry, MoCap and neural modeling results from our dataset. HODome includes various interaction sequences, such as "holding a vase", "opening the cupboard", "sitting on the sofa" and "moving the table". + +(Eq. 1). However, due to the misalignment of SMPL shape and actual human geometry, the human body and clothes may still overlap the object space, which goes against real-world physics. Therefore, we implicitly constrain the deformation net to predict contact-aware non-rigid deformation: + +$$ +\mathcal {L} _ {h} = \sum_ {p \in \mathcal {H}} \Omega_ {-} (p, \mathcal {M}) \| \sigma_ {p} \| _ {2} ^ {2}, \tag {6} +$$ + +where $\mathcal{H}$ is the set of random samples in the human's bounding box. + +We further explore a weakly-supervising scheme that utilizes object texture cues to help decouple the human and object entities. The key observation is that, with accurate object tracking, the object radiance field fused across frames can be rendered with higher confidence. That is, we train the entire scene first and render the objects only to capture views, and the pixels similar to the captured ones are labeled as object pixels. In this way, we obtain coarse object segmentation maps $S$ that serve as pseudo supervision: we adopt label-wise integration [53] to render the layer labels $s$ of rays: + +$$ +\mathbf {s} = \sum_ {i = 1} ^ {M} T \left(p _ {i}\right) \left[ \left(1 - e ^ {- \sigma_ {p _ {i}} \delta_ {p _ {i}}}\right) \mathbf {l} _ {p _ {i}} \right], \tag {7} +$$ + +where $\mathbf{l}_{p_i}$ is one-hot label indicates which layer the point $p_i$ belongs to. We apply a semantic loss for object rays $S_o$ : + +$$ +\mathcal {L} _ {s} = \sum_ {\mathbf {r} \in \mathcal {S} _ {o}} \| \mathbf {s} _ {\mathbf {r}} - \hat {\mathbf {s}} _ {\mathbf {r}} \| _ {2} ^ {2}, \tag {8} +$$ + +where $\hat{\mathbf{s}}_{\mathbf{r}}$ is the pseudo semantic label of ray $\mathbf{r}$ . Note that additional loss can be applied if accurate labels are provided. + +Layer-wise neural representation in HODome. To further use the data, we can render reliable segmentation maps and occlusion-free appearances for humans and objects respectively. However, the renderings tend to be blurred since + +the information of all the frames is fused. Hence we introduce an enhancement scheme to construct separate high-quality per-frame neural representation for human/object: we blend the visible regions of the high-fidelity observed input views to rendered human/object images using the segmentation maps. The Instant-NGP [38] and Instant-NSR [79] techniques are then applied to reconstruct separate human/object in tens of seconds. Finally, we obtain layer-wise representation which enables real-time photorealistic rendering and benefits the thorough analysis of HOI scenarios. + +Implementations. To build our neural representations, we train our layered models using Adam optimizer with a learning rate that starts from $5e^{-4}$ and decays exponentially. Using the 76-view 4K resolution videos, the training process takes 8-10 hours on a single NVIDIA 3090 GPU for acceptable appearance and segmentation rendering. Thanks to Instant-NGP [38] and Instant-NSR [79], the layer-wise neural modeling process is conducted in seconds per frame. + +# 5. Experiments + +In this section, we compare our neural pipeline with existing state-of-the-arts. We first demonstrate the effectiveness and the generalization ability of our neural modeling approach (5.1). Next, we implement NeuralDome on three specific tasks (5.2). + +# 5.1. Analysis of Neural Rendering + +Comparisons. Here we compare our layer-wise neural human-object rendering approach, denoted as 'Ours-layer', with recent state-of-the-art methods, i.e., ST-NeRF [77], NeuralBody(NB) [45], in human-object interaction scenario. For fair comparisons, we select 20 cameras surrounding the performer for training and others for evaluation. We + +![](images/15fa79560dc22a3b92c6a65e9be6944aeacef22aec6a805f821c06047b8caaf6.jpg) +Figure 6. Qualitative comparisons of novel view synthesis results. We show ground truth and synthesized images of novel view for NeuralBody [45], and ST-NeRF [77] and our layered human-object representation. Our approach achieves the best performance. We further illustrate our high-fidelity neural representations in HODome in the last column. + +
MethodsNB [45]ST-NeRF [77]Ours
ScenesPSNR↑SSIM↑PSNR↑SSIM↑PSNR↑SSIM↑
Bigsofa19.330.89624.020.88632.280.958
Sofa26.730.96528.490.95835.620.987
Table21.940.93322.440.90027.880.949
Average22.670.93124.990.91531.930.964
+ +remove the background layer and train ST-NeRF without the semantic label as ours. Fig. 6 shows several appearance synthesis results. NeuralBody has no ability to model objects which leads to artifacts in the density field. ST-NeRF fails as the bounding boxes of the performer and object almost completely overlap. Our method achieves both human and object modeling and rendering and thus further enables generating high-fidelity neural representation denoted as 'Ours-HODome'. Tab. 2 further illustrates our method significantly outperforms the baselines. + +Evaluation. Here we further evaluate how our scheme contributes to the generated layer-wise neural representations. Fig. 7 shows the single human layer where the performer is "sitting" in the air. Note that we have no ground truth of the specific human/object layer, thus we conduct qualitative evaluations only. Let w/o pseudo segmentation and w/o blending denote our human assets generated without pseudo semantic loss and blending-based enhancement scheme. It demonstrates that the pseudo semantic loss effectively helps decouple the human and object layer and our blending scheme further boosts the appearance fidelity of the blurred temporal-fused model to a photo-realistic level. + +![](images/446b1aead1a9ec1ebb02f849db33f917efe52cfa62691cd186a8d5f5b706234e.jpg) +Figure 7. Qualitative evaluation of each strategy's contribution to layer-wise neural representation generation. The weakly-supervising scheme effectively helps decouple human and object and our blending scheme generates high-quality digital assets. + +Table 2. Quantitative comparisons of novel view synthesized appearance on different human-object interaction sequences. + +
MethodSeparated evaluationJoint evaluation
MPJPE↓PA-MPJPE↓Chamferh↓Chamfero↓V2V↓p.V2V↓
Fit to input16.116.9486.6618.1461.5728.92
PHOSA [78]14.886.9477.6216.6454.5926.09
CHORE [73]10.216.1486.587.6944.9314.93
+ +Table 3. Human object capture benchmark. "Fit to input" represents the vanilla method that fits the object template to image and capture human with Frankmocap [47]. + +# 5.2. Task and benchmark + +Human-object Capture Benchmark. HODome provides multi-view sequences with synchronized ground truth of capturing. To demonstrate the capability, we provide a benchmark for human-object capture, shown in Tab. 3. For detailed metric explanation please refer to [6, 73, 74]. PHOSA [78] used the contact map from hand-crafted annotations, suffering from coarse contact information. While CHORE [73] outperforms it with the distance fields predicted from neural networks. It shows the importance of a prepared human-object dataset that supports data-driven human-object capture model. Check out the full capturing benchmark of the dataset at https://arxiv.org/abs/2212.07626. + +Human-object Geometry Reconstruction Benchmark. Benefiting from the dense-view setting, various human-only geometry reconstruction methods can step towards a human-object geometry reconstruction setting. We can benchmark the state-of-the-art algorithms in sparse view geometry reconstruction tasks. Hence, we evaluate PIFu [48] by training on a subset of HODome. As shown in Fig. 8, we compare the quality result between original PIFu and PIFu-trained on HODome and in-the-wild images. Compared to the original PIFu, our training set can improve the quality of + +![](images/74525579354cdace85ffdff903c9f173f7f0e827cea1481936cbb59b002f45ef.jpg) +NA + +![](images/b63b870fe6f5b57c7d5848cec3cfa58168197534f846f10c099e04164648f771.jpg) + +![](images/da7ac4250173b8fbe463747dcf3901144873cfd38441ff40a1b78fbba28d1a23.jpg) + +![](images/5f8cd77a43bfb8f83a5905cb23fe91454cd31d9a4f871c3fc89678f59a9ec847.jpg) + +![](images/891609b21faf132d8da293426b7344ea2c7c6629ec176f5f1ec63aab2a0247ba.jpg) + +![](images/624a47ec8b979c3074ab298f8e13364e51ac1952c655cf81a1da138f235135c1.jpg) +Input + +![](images/16d70660e3a1c3a6b6fb9ddebfd579b9323b00d0b64ddd3d6b664e05ee9f8169.jpg) +PIFu-Origin + +![](images/e724f1aece2df2ce899a9b73ea27ddf65efccc4f724ab6b09ac8444f4681b927.jpg) +Monocular PIFu-Trained +Figure 8. Qualitative evaluation on human-object reconstruction. We compare the original PIFu [48] with trained PIFu on both images from our dataset and in-the-wild images. The original PIFu fails to predict the object depth, and PIFu trained on our dataset can approximately predict the shape of the object. + +![](images/de5e2daad2b958c849cb53506715a8e29d0e1cdb5296626ed52be20e94845e54.jpg) +Six View PIFu-Trained + +
MethodP2S ×10-4↓Chamfer ×10-4↓
Origin PIFu [48]38.72640.947
Monocular PIFu-trained14.65314.483
6-View PIFu-trained3.3764.901
+ +Table 4. Geometry reconstruction benchmark. Origin PIFu [48] uses the pre-trained model to infer. Monocular PIFu is trained on a single view. 6-View PIFu is trained on 6-view inputs. + +
MethodPSNR↑SSIM↑
IBRnet [69]21.430.892
Neuray [30]23.340.909
NeuralHumanFVV [60]21.690.914
NeuralHOIFVV23.100.912
+ +Table 5. Neural rendering benchmark in sparse-view setting. + +the reconstructed shape of the objects. For further quantitative analysis, we evaluate the performance using P2S and CD in Tab. 4. + +Sparse-View Human-object Rendering Benchmark. Our HODome supports various neural rendering tasks, even for human-object interactions. We provide a benchmark on sparse-view rendering tasks and evaluate on IBRNet [69], NeuRay [30] and NeuralHumanFVV [60]. Besides, we also provide a baseline, named NeuralHOIFVV, which uses the projection of trained PIFu's [48] results as depth input, then applies the Neural Blending method presented in [79] to obtain novel-view synthesis. More details about NeuralHOIFVV are referred to the Appendix. We analyze the above methods on HODome and in-the-wild images. The qualitative and quantitative results are shown in Fig. 9 and Tab. 5. Our proposed NeuralHOIFVV can obtain generative novel-view synthesis on HODome and in-the-wild inputs. + +# 5.3. Limitations + +Although NeuralDome can provide accurate capturing results and layer-wise neural representations, we also want + +![](images/6eaa6ccfcf504c35f5ade452b9f6517ce9131c8f2ccb5cd221ba76e2a1101d8d.jpg) + +![](images/1daaf82378f503cfe4ba5ffccbdab163354242a30edae5969bc5c2eac14393dc.jpg) + +![](images/4a0643b8bb8ad2571537467f4d4f233b4f36de28b6667780198a66abadaa30d6.jpg) + +![](images/57fd79734fae9b65d5873a0b5b4e059735e8b181dc25b53fce4e97e88d8056c9.jpg) +GT IBRNet + +![](images/7134b8c3401540046b7b5fd75b7c2d5c59c8f868e8e115e0a33c06a91acf6766.jpg) +NeuRay + +![](images/1943accd934a35f81527c8be64e66948d3fef9fede39a25320683f904b982774.jpg) +Neural HumanFVV +Figure 9. Qualitative evaluation on neural human-object rendering. We show the comparison of IBRNet [69], NeuRay [30], NeuralHumanFVV [60] and our proposed NeuralHOIFVV. We render one novel view between two input camera views. + +![](images/e21a01a06ded1545afa492c11daab668908d812f13ae91bcacc0da7bb875cec6.jpg) +Neural HOIFVV + +to highlight some potential limitations of this pipeline. First, NeuralDome only considers single-person interaction with objects and the single-person task is already quite challenging for current research. It is non-trivial to extend the current algorithm to multi-person settings, especially in a crowd scene. Secondly, the reconstruction of 3D holistic scene is not covered in our pipeline. We leave the joint modeling of humans, objects and scene in the future work. Moreover, our dataset was collected under fixed illumination conditions with few backgrounds variance, limiting its generalization ability for other environments. + +# 6. Conclusion + +We have presented HODome, the first dataset to jointly capture and render human-object interaction using dense RGB cameras and the Optitrack system. To process the HODome dataset, we have further developed NeuralDome, a neural pipeline that can accurately track humans and objects, conduct layer-wise geometry reconstruction, and enable novel-view synthesis for both the subjects and objects. We believe the dataset could boost the development of capturing and rendering human-object interactions and is very valuable to the community for a wide range of human-object capturing and rendering tasks. + +Acknowledgement This work was supported by the Shanghai Sailing Program (21YF1429400, 22YF1428800), Shanghai Local College Capacity Building Program (23010503100, 22010502800), NSFC programs (61976138, 61977047), the National Key Research and Development Program (2018YFB2100500), STCSM (2015F0203-000-06), SHMEC (2019-01-07-00-01-E00003) and Shanghai Frontiers Science Center of Humancentered Artificial Intelligence (ShangHAI). + +# References + +[1] Naturalpoint, inc. motion capture systems. https://optitrack.com/. 6.3 +[2] Reality capture. https://www.capturingreality.com/realitycapture. 3, 4 +[3] Easymocap - make human motion capture easier. Github, 2021.4 +[4] Kara-Ali Aliev, Artem Sevastopolsky, Maria Kolos, Dmitry Ulyanov, and Victor Lempitsky. Neural point-based graphics. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16, pages 696-712. Springer, 2020. 2 +[5] Paul J Besl and Neil D McKay. Method for registration of 3-d shapes. In Sensor fusion IV: control paradigms and data structures, volume 1611, pages 586-606. Spie, 1992. 4 +[6] Bharat Lal Bhatnagar, Xianghui Xie, Ilya A Petrov, Cristian Sminchisescu, Christian Theobalt, and Gerard Pons-Moll. Behave: Dataset and method for tracking human object interactions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15935-15946, 2022. 2, 3, 4, 7 +[7] Hongrui Cai, Wanquan Feng, Xuetao Feng, Yan Wang, and Juyong Zhang. Neural surface reconstruction of dynamic scenes with monocular rgb-d camera. In Thirty-sixth Conference on Neural Information Processing Systems (NeurIPS), 2022. 2 +[8] Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Realtime multi-person 2d pose estimation using part affinity fields. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7291-7299, 2017. 4 +[9] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. arXiv preprint arXiv:2203.09517, 2022. 2 +[10] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14124-14133, 2021. 2 +[11] Alvaro Collet, Ming Chuang, Pat Sweeney, Don Gillett, Dennis Evseev, David Calabrese, Hugues Hoppe, Adam Kirk, and Steve Sullivan. High-quality streamable free-viewpoint video. ACM Transactions on Graphics (TOG), 34(4):69, 2015. 2, 3 +[12] Mingsong Dou, Philip Davidson, Sean Ryan Fanello, Sameh Khamis, Adarsh Kowdle, Christoph Rhemann, Vladimir Tankovich, and Shahram Izadi. Motion2fusion: Real-time volumetric performance capture. ACM Trans. Graph., 36(6):246:1-246:16, Nov. 2017. 2, 3 +[13] Kaiwen Guo, Peter Lincoln, Philip Davidson, Jay Busch, Xueming Yu, Matt Whalen, Geoff Harvey, Sergio Orts-Escolano, Rohit Pandey, Jason Dourgarian, et al. The relightables: Volumetric performance capture of humans with realistic relighting. ACM Transactions on Graphics (ToG), 38(6):1–19, 2019. 3 +[14] Marc Habermann, Lingjie Liu, Weipeng Xu, Michael Zollhoefer, Gerard Pons-Moll, and Christian Theobalt. Real-time + +deep dynamic characters. ACM Transactions on Graphics (TOG), 40(4):1-16, 2021. 2 +[15] Nils Hasler, Carsten Stoll, Martin Sunkel, Bodo Rosenhahn, and H-P Seidel. A statistical model of human pose and body shape. In Computer graphics forum, volume 28, pages 337-346. Wiley Online Library, 2009. 3 +[16] Mohamed Hassan, Vasileios Choutas, Dimitrios Tzionas, and Michael J Black. Resolving 3d human pose ambiguities with 3d scene constraints. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2282-2292, 2019. 3, 5 +[17] Tong He, Yuanlu Xu, Shunsuke Saito, Stefano Soatto, and Tony Tung. Arch++: Animation-ready clothed human reconstruction revisited. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11046-11056, 2021. 2 +[18] Yinghao Huang, Omid Taheri, Michael J Black, and Dimitrios Tzionas. Intercap: Joint markerless 3d tracking of humans and objects in interaction. In DAGM German Conference on Pattern Recognition, pages 281-299. Springer, 2022. 3 +[19] Zeng Huang, Yuanlu Xu, Christoph Lassner, Hao Li, and Tony Tung. Arch: Animatable reconstruction of clothed humans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3093-3102, 2020. 2 +[20] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3.6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE Transactions on Pattern Analysis and Machine Intelligence, 36(7):1325-1339, jul 2014. 3 +[21] Yuheng Jiang, Suyi Jiang, Guoxing Sun, Zhuo Su, Kaiwen Guo, Minye Wu, Jingyi Yu, and Lan Xu. Neuralhofusion: Neural volumetric rendering under human-object interactions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6155-6165, 2022. 2, 3 +[22] Hanbyul Joo, Tomas Simon, Xulong Li, Hao Liu, Lei Tan, Lin Gui, Sean Banerjee, Timothy Scott Godisart, Bart Nabbe, Iain Matthews, Takeo Kanade, Shohei Nobuhara, and Yaser Sheikh. Panoptic studio: A massively multiview system for social interaction capture. 2017. 3 +[23] Youngjoong Kwon, Dahun Kim, Duygu Ceylan, and Henry Fuchs. Neural human performer: Learning generalizable radiance fields for human performance rendering. Advances in Neural Information Processing Systems, 34:24741-24752, 2021. 2 +[24] Christoph Lassner, Javier Romero, Martin Kiefel, Federica Bogo, Michael J. Black, and Peter V. Gehler. Unite the people: Closing the loop between 3d and 2d human representations. In IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), July 2017. 3 +[25] Ruilong Li, Julian Tanke, Minh Vo, Michael Zollhofer, Jurgen Gall, Angjoo Kanazawa, and Christoph Lassner. Tava: Template-free animatable volumetric actors. arXiv preprint arXiv:2206.08929, 2022. 2 +[26] Jun Liu, Amir Shahroudy, Mauricio Perez, Gang Wang, Ling-Yu Duan, and Alex C Kot. Nturgb+ d 120: A large- + +scale benchmark for 3d human activity understanding. IEEE transactions on pattern analysis and machine intelligence, 42(10):2684-2701, 2019. 2, 3 +[27] Lingjie Liu, Marc Habermann, Viktor Rudnev, Kripasindhu Sarkar, Jiatao Gu, and Christian Theobalt. Neural actor: Neural free-view synthesis of human actors with pose control. ACM Trans. Graph. (ACM SIGGRAPH Asia), 2021. 2 +[28] Lingjie Liu, Weipeng Xu, Marc Habermann, Michael Zollhöfer, Florian Bernard, Hyeongwoo Kim, Wenping Wang, and Christian Theobalt. Neural human video rendering by learning dynamic textures and rendering-to-video translation. arXiv preprint arXiv:2001.04947, 2020. 2 +[29] Lingjie Liu, Weipeng Xu, Michael Zollhoefer, Hyeongwoo Kim, Florian Bernard, Marc Habermann, Wenping Wang, and Christian Theobalt. Neural rendering and reenactment of human actor videos. ACM Transactions on Graphics (TOG), 38(5):1-14, 2019. 2 +[30] Yuan Liu, Sida Peng, Lingjie Liu, Qianqian Wang, Peng Wang, Christian Theobalt, Xiaowei Zhou, and Wenping Wang. Neural rays for occlusion-aware image-based rendering. In CVPR, 2022. 2, 8 +[31] Stephen Lombardi, Tomas Simon, Jason Saragih, Gabriel Schwartz, Andreas Lehrmann, and Yaser Sheikh. Neural volumes: Learning dynamic renderable volumes from images. arXiv preprint arXiv:1906.07751, 2019. 2 +[32] Stephen Lombardi, Tomas Simon, Gabriel Schwartz, Michael Zollhoefer, Yaser Sheikh, and Jason Saragih. Mixture of volumetric primitives for efficient neural rendering. arXiv preprint arXiv:2103.01954, 2021. 2 +[33] Xiaoxiao Long, Cheng Lin, Peng Wang, Taku Komura, and Wenping Wang. Sparseneus: Fast generalizable neural surface reconstruction from sparse views. ECCV, 2022. 2 +[34] H. Luo, A. Chen, Q. Zhang, B. Pang, M. Wu, L. Xu, and J. Yu. Convolutional neural opacity radiance fields. In 2021 IEEE International Conference on Computational Photography (ICCP), pages 1-12, Los Alamitos, CA, USA, may 2021. IEEE Computer Society. 2 +[35] Haimin Luo, Teng Xu, Yuheng Jiang, Chenglin Zhou, Qiwei Qiu, Yingliang Zhang, Wei Yang, Lan Xu, and Jingyi Yu. Artemis: Articulated neural pets with appearance and motion synthesis. ACM Trans. Graph., 41(4), jul 2022. 2 +[36] Quan Meng, Anpei Chen, Haimin Luo, Minye Wu, Hao Su, Lan Xu, Xuming He, and Jingyi Yu. Gnerf: Gan-based neural radiance field without posed camera. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6351-6361, 2021. 2 +[37] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European conference on computer vision, pages 405-421. Springer, 2020. 2 +[38] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, July 2022. 2, 4, 6 +[39] Michael Niemeyer, Jonathan T. Barron, Ben Mildenhall, Mehdi S. M. Sajjadi, Andreas Geiger, and Noha Radwan. + +Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 2 +[40] Atsuhiro Noguchi, Xiao Sun, Stephen Lin, and Tatsuya Harada. Neural articulated radiance field. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5762-5772, 2021. 2 +[41] Anqi Pang, Xin Chen, Haimin Luo, Minye Wu, Jingyi Yu, and Lan Xu. Few-shot neural human performance rendering from sparse rgbd videos, 2021. 2 +[42] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M. Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. ACM Trans. Graph., 40(6), dec 2021. 2 +[43] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3d hands, face, and body from a single image. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10975-10985, 2019. 2, 5 +[44] Sida Peng, Junting Dong, Qianqian Wang, Shangzhan Zhang, Qing Shuai, Xiaowei Zhou, and Hujun Bao. Animatable neural radiance fields for modeling dynamic human bodies. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14314-14323, 2021. 2 +[45] Sida Peng, Yuanqing Zhang, Yinghao Xu, Qianqian Wang, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In CVPR, 2021. 2, 6, 7 +[46] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, jun 2021. 2 +[47] Yu Rong, Takaaki Shiratori, and Hanbyul Joo. Frankmocap: A monocular 3d whole-body pose estimation system via regression and integration. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1749-1759, 2021. 7 +[48] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2304-2314, 2019. 2, 7, 8 +[49] Shunsuke Saito, Tomas Simon, Jason Saragih, and Hanbyul Joo. Pifuhd: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 84-93, 2020. 2 +[50] Manolis Savva, Angel X Chang, Pat Hanrahan, Matthew Fisher, and Matthias Nießner. Pigraphs: learning interaction snapshots from observations. ACM Transactions on Graphics (TOG), 35(4):1-12, 2016. 3 +[51] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-Motion Revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2, 3 + +[52] Qing Shuai, Chen Geng, Qi Fang, Sida Peng, Wenhao Shen, Xiaowei Zhou, and Hujun Bao. Novel view synthesis of human interactions from sparse multi-view videos. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-10, 2022. 2 +[53] Qing Shuai, Chen Geng, Qi Fang, Sida Peng, Wenhao Shen, Xiaowei Zhou, and Hujun Bao. Novel view synthesis of human interactions from sparse multi-view videos. In ACM SIGGRAPH 2022 Conference Proceedings, SIGGRAPH '22, New York, NY, USA, 2022. Association for Computing Machinery. 3, 6 +[54] Aliaksandra Shysheya, Egor Zakharov, Kara-Ali Aliev, Renat Bashirov, Egor Burkov, Karim Iskakov, Aleksei Ivakhnenko, Yury Malkov, Igor Pasechnik, Dmitry Ulyanov, et al. Textured neural avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2387-2397, 2019. 2 +[55] Leonid Sigal, Alexandru O Balan, and Michael J Black. Humaneva: Synchronized video and motion capture dataset and baseline algorithm for evaluation of articulated human motion. International journal of computer vision, 87(1):4-27, 2010. 3 +[56] Zhuo Su, Lan Xu, Zerong Zheng, Tao Yu, Yebin Liu, and Lu Fang. Robustfusion: Human volumetric capture with data-driven visual cues using a rgbd camera. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, Computer Vision – ECCV 2020, pages 246–264, Cham, 2020. Springer International Publishing. 3 +[57] Guoxing Sun, Xin Chen, Yizhang Chen, Anqi Pang, Pei Lin, Yuheng Jiang, Lan Xu, Jingya Wang, and Jingyi Yu. Neural free-viewpoint performance rendering under complex human-object interactions. In Proceedings of the 29th ACM International Conference on Multimedia, 2021. 2, 3 +[58] Guoxing Sun, Xin Chen, Yizhang Chen, Anqi Pang, Pei Lin, Yuheng Jiang, Lan Xu, Jingyi Yu, and Jingya Wang. Neural free-viewpoint performance rendering under complex human-object interactions. In Proceedings of the 29th ACM International Conference on Multimedia, pages 4651-4660, 2021. 3 +[59] Xin Suo, Yuheng Jiang, Pei Lin, Yingliang Zhang, Minye Wu, Kaiwen Guo, and Lan Xu. Neuralhumanfvv: Realtime neural volumetric human performance rendering using rgb cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6226-6237, 2021. 2 +[60] Xin Suo, Yuheng Jiang, Pei Lin, Yingliang Zhang, Minye Wu, Kaiwen Guo, and Lan Xu. Neuralhumanfvv: Real-time neural volumetric human performance rendering using rgb cameras. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6222-6233, 2021. 8 +[61] Omid Taheri, Nima Ghorbani, Michael J Black, and Dimitrios Tzionas. Grab: A dataset of whole-body human grasping of objects. In European conference on computer vision, pages 581-600. Springer, 2020. 2, 3, 4 +[62] Ayush Tewari, Justus Thies, Ben Mildenhall, Pratul Srinivasan, Edgar Tretschk, Yifan Wang, Christoph Lassner, Vincent Sitzmann, Ricardo Martin-Brualla, Stephen Lom- + +bardi, Tomas Simon, Christian Theobalt, Matthias Niessner, Jonathan T. Barron, Gordon Wetzstein, Michael Zollhoefer, and Vladislav Golyanik. Advances in neural rendering, 2021. 2 +[63] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhofer, Christoph Lassner, and Christian Theobalt. Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12959-12970, 2021. 2 +[64] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T Barron, and Pratul P Srinivasan. Ref-nerf: Structured view-dependent appearance for neural radiance fields. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5481-5490. IEEE, 2022. 2 +[65] Timo Von Marcard, Roberto Henschel, Michael J Black, Bodo Rosenhahn, and Gerard Pons-Moll. Recovering accurate 3d human pose in the wild using imus and a moving camera. In Proceedings of the European Conference on Computer Vision (ECCV), pages 601-617, 2018. 3 +[66] Liao Wang, Ziyu Wang, Pei Lin, Yuheng Jiang, Xin Suo, Minye Wu, Lan Xu, and Jingyi Yu. ibutter: Neural interactive bullet time generator for human free-viewpoint rendering. In Proceedings of the 29th ACM International Conference on Multimedia, pages 4641-4650, 2021. 2 +[67] Liao Wang, Jiakai Zhang, Xinhang Liu, Fuqiang Zhao, Yanshun Zhang, Yingliang Zhang, Minye Wu, Jingyi Yu, and Lan Xu. Fourier plenoctrees for dynamic radiance field rendering in real-time. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13524-13534, 2022. 2 +[68] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. NeurIPS, 2021. 2 +[69] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul Srinivasan, Howard Zhou, Jonathan T. Barron, Ricardo MartinBrualla, Noah Snively, and Thomas Funkhouser. Ibrnet: Learning multi-view image-based rendering. In CVPR, 2021. 2,8 +[70] Shaofei Wang, Katja Schwarz, Andreas Geiger, and Siyu Tang. Arah: animatable volume rendering of articulated human sdfs. arXiv preprint arXiv:2210.10036, 2022. 2 +[71] Minye Wu, Yuehao Wang, Qiang Hu, and Jingyi Yu. Multi-view neural human rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 2 +[72] Minye Wu, Yuehao Wang, Qiang Hu, and Jingyi Yu. Multi-view neural human rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1682-1691, 2020. 2 +[73] Xianghui Xie, Bharat Lal Bhatnagar, and Gerard Pons-Moll. Chore: Contact, human and object reconstruction from a single rgb image. arXiv preprint arXiv:2204.02445, 2022. 3, 7 +[74] Hongwei Yi, Chun-Hao P Huang, Dimitrios Tzionas, Muhammed Kocabas, Mohamed Hassan, Siyu Tang, Justus Thies, and Michael J Black. Human-aware object placement + +for visual environment reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3959-3970, 2022. 3, 7 +[75] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. Plenoctrees for real-time rendering of neural radiance fields. In arXiv, 2021. 2 +[76] Chao Zhang, Sergi Pujades, Michael J. Black, and Gerard Pons-Moll. Detailed, accurate, human shape estimation from clothed 3d scan sequences. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), July 2017. 3 +[77] Jiakai Zhang, Xinhang Liu, Xinyi Ye, Fuqiang Zhao, Yanshun Zhang, Minye Wu, Yingliang Zhang, Lan Xu, and Jingyi Yu. Editable free-viewpoint video using a layered neural representation. ACM Trans. Graph., 40(4), July 2021. 2, 3, 5, 6, 7 +[78] Jason Y Zhang, Sam Pepose, Hanbyul Joo, Deva Ramanan, Jitendra Malik, and Angjoo Kanazawa. Perceiving 3d human-object spatial arrangements from a single image in the wild. In European conference on computer vision, pages 34-51. Springer, 2020. 3, 7 +[79] Fuqiang Zhao, Yuheng Jiang, Kaixin Yao, Jiakai Zhang, Liao Wang, Haizhao Dai, Yuhui Zhong, Yingliang Zhang, Minye Wu, Lan Xu, et al. Human performance modeling and rendering via neural animated mesh. arXiv preprint arXiv:2209.08468, 2022. 2, 4, 6, 8 +[80] Fuqiang Zhao, Wei Yang, Jiakai Zhang, Pei Lin, Yingliang Zhang, Jingyi Yu, and Lan Xu. Humannerf: Efficiently generated human radiance field from sparse inputs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7743-7753, 2022. 2, 5 +[81] Qian-Yi Zhou, Jaesik Park, and Vladlen Koltun. Open3d: A modern library for 3d data processing. arXiv preprint arXiv:1801.09847, 2018. 4 \ No newline at end of file diff --git a/2023/NeuralDome_ A Neural Modeling Pipeline on Multi-View Human-Object Interactions/images.zip b/2023/NeuralDome_ A Neural Modeling Pipeline on Multi-View Human-Object Interactions/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..abbbf8205648c718fa206543d28ad35957df442f --- /dev/null +++ b/2023/NeuralDome_ A Neural Modeling Pipeline on Multi-View Human-Object Interactions/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1e7f87d676e39e31330607f088816dd3454929c4ab1589b7bee4485d2d60a72a +size 639705 diff --git a/2023/NeuralDome_ A Neural Modeling Pipeline on Multi-View Human-Object Interactions/layout.json b/2023/NeuralDome_ A Neural Modeling Pipeline on Multi-View Human-Object Interactions/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e6577fb0c15ff28bc2f044b24ebd94cc6c8b7ef4 --- /dev/null +++ b/2023/NeuralDome_ A Neural Modeling Pipeline on Multi-View Human-Object Interactions/layout.json @@ -0,0 +1,9676 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 72, + 102, + 522, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 102, + 522, + 138 + ], + "spans": [ + { + "bbox": [ + 72, + 102, + 522, + 138 + ], + "type": "text", + "content": "NeuralDome: A Neural Modeling Pipeline on Multi-View Human-Object Interactions" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 159, + 538, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 159, + 538, + 189 + ], + "spans": [ + { + "bbox": [ + 69, + 159, + 538, + 189 + ], + "type": "text", + "content": "Juze Zhang " + }, + { + "bbox": [ + 69, + 159, + 538, + 189 + ], + "type": "inline_equation", + "content": "^{1,2,3,4,*}" + }, + { + "bbox": [ + 69, + 159, + 538, + 189 + ], + "type": "text", + "content": ", Haimin Luo " + }, + { + "bbox": [ + 69, + 159, + 538, + 189 + ], + "type": "inline_equation", + "content": "^{1,4,5,*}" + }, + { + "bbox": [ + 69, + 159, + 538, + 189 + ], + "type": "text", + "content": ", Hongdi Yang " + }, + { + "bbox": [ + 69, + 159, + 538, + 189 + ], + "type": "inline_equation", + "content": "^{1,4}" + }, + { + "bbox": [ + 69, + 159, + 538, + 189 + ], + "type": "text", + "content": ", Xinru Xu " + }, + { + "bbox": [ + 69, + 159, + 538, + 189 + ], + "type": "inline_equation", + "content": "^{1,4}" + }, + { + "bbox": [ + 69, + 159, + 538, + 189 + ], + "type": "text", + "content": ", Qianyang Wu " + }, + { + "bbox": [ + 69, + 159, + 538, + 189 + ], + "type": "inline_equation", + "content": "^{1,4}" + }, + { + "bbox": [ + 69, + 159, + 538, + 189 + ], + "type": "text", + "content": ", Ye Shi " + }, + { + "bbox": [ + 69, + 159, + 538, + 189 + ], + "type": "inline_equation", + "content": "^{1,4}" + }, + { + "bbox": [ + 69, + 159, + 538, + 189 + ], + "type": "text", + "content": ", Jingyi Yu " + }, + { + "bbox": [ + 69, + 159, + 538, + 189 + ], + "type": "inline_equation", + "content": "^{1,4}" + }, + { + "bbox": [ + 69, + 159, + 538, + 189 + ], + "type": "text", + "content": ", Lan Xu " + }, + { + "bbox": [ + 69, + 159, + 538, + 189 + ], + "type": "inline_equation", + "content": "^{1,4,\\dagger}" + }, + { + "bbox": [ + 69, + 159, + 538, + 189 + ], + "type": "text", + "content": ", Jingya Wang " + }, + { + "bbox": [ + 69, + 159, + 538, + 189 + ], + "type": "inline_equation", + "content": "^{1,4,\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 189, + 553, + 246 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 66, + 189, + 543, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 189, + 543, + 217 + ], + "spans": [ + { + "bbox": [ + 66, + 189, + 543, + 217 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 66, + 189, + 543, + 217 + ], + "type": "text", + "content": " ShanghaiTech University " + }, + { + "bbox": [ + 66, + 189, + 543, + 217 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 66, + 189, + 543, + 217 + ], + "type": "text", + "content": " Shanghai Advanced Research Institute, Chinese Academy of Sciences " + }, + { + "bbox": [ + 66, + 189, + 543, + 217 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 66, + 189, + 543, + 217 + ], + "type": "text", + "content": " University of Chinese Academy of Sciences" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 217, + 553, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 217, + 553, + 246 + ], + "spans": [ + { + "bbox": [ + 53, + 217, + 553, + 246 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 53, + 217, + 553, + 246 + ], + "type": "text", + "content": " Shanghai Engineering Research Center of Intelligent Vision and Imaging " + }, + { + "bbox": [ + 53, + 217, + 553, + 246 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 53, + 217, + 553, + 246 + ], + "type": "text", + "content": " LumiAni Technology " + }, + { + "bbox": [ + 53, + 217, + 553, + 246 + ], + "type": "inline_equation", + "content": "\\{\\mathrm{zhangjz,luohm,yanghd,xuxr2022,wuqy2022,shiye,yujingyi,xulan1,wangjingya}\\} @shanghaiitech.edu.cn" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "type": "image", + "bbox": [ + 69, + 304, + 525, + 452 + ], + "blocks": [ + { + "bbox": [ + 69, + 304, + 525, + 452 + ], + "lines": [ + { + "bbox": [ + 69, + 304, + 525, + 452 + ], + "spans": [ + { + "bbox": [ + 69, + 304, + 525, + 452 + ], + "type": "image", + "image_path": "a6ddac9037905908066efb62551cd861d33d9aae1ede8cb2fffc35ad11317161.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 460, + 546, + 504 + ], + "lines": [ + { + "bbox": [ + 45, + 460, + 546, + 504 + ], + "spans": [ + { + "bbox": [ + 45, + 460, + 546, + 504 + ], + "type": "text", + "content": "Figure 1. Our NeuralDome pipeline for processing multi-view video sequences on human object interactions. NeuralDome supports tracking, modeling, and rendering of individual human subjects and objects. To validate NeuralDome, we collect a large dataset HODome over a total of 71 million video frames across 76 viewpoints and process the datasets using NeuralDome for a variety of inference and neural modeling and rendering tasks." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 143, + 520, + 192, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 520, + 192, + 533 + ], + "spans": [ + { + "bbox": [ + 143, + 520, + 192, + 533 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 547, + 288, + 679 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 547, + 288, + 679 + ], + "spans": [ + { + "bbox": [ + 46, + 547, + 288, + 679 + ], + "type": "text", + "content": "Humans constantly interact with objects in daily life tasks. Capturing such processes and subsequently conducting visual inferences from a fixed viewpoint suffers from occlusions, shape and texture ambiguities, motions, etc. To mitigate the problem, it is essential to build a training dataset that captures free-viewpoint interactions. We construct a dense multi-view dome to acquire a complex human object interaction dataset, named HODome, that consists of " + }, + { + "bbox": [ + 46, + 547, + 288, + 679 + ], + "type": "inline_equation", + "content": "\\sim 71M" + }, + { + "bbox": [ + 46, + 547, + 288, + 679 + ], + "type": "text", + "content": " frames on 10 subjects interacting with 23 objects. To process the HODome dataset, we develop NeuralDome, a layer-wise neural processing pipeline tai" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 521, + 547, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 521, + 547, + 628 + ], + "spans": [ + { + "bbox": [ + 306, + 521, + 547, + 628 + ], + "type": "text", + "content": "lored for multi-view video inputs to conduct accurate tracking, geometry reconstruction and free-view rendering, for both human subjects and objects. Extensive experiments on the HODome dataset demonstrate the effectiveness of NeuralDome on a variety of inference, modeling, and rendering tasks. Both the dataset and the NeuralDome tools will be disseminated to the community for further development, which can be found at https://juzezhang.github.io/NeuralDome" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 656, + 386, + 669 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 656, + 386, + 669 + ], + "spans": [ + { + "bbox": [ + 306, + 656, + 386, + 669 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 678, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 678, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 306, + 678, + 546, + 715 + ], + "type": "text", + "content": "A key task of computer vision is to understand how humans interact with the surrounding world, by faithfully capturing and subsequently reproducing the process via mod" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 693, + 167, + 704 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 693, + 167, + 704 + ], + "spans": [ + { + "bbox": [ + 47, + 693, + 167, + 704 + ], + "type": "text", + "content": "* These authors contributed equally." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 704, + 126, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 704, + 126, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 704, + 126, + 714 + ], + "type": "text", + "content": "†Corresponding author." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8834" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": "eling and rendering. Successful solutions benefit broad applications ranging from sports training to vocational education, digital entertainment to tele-medicine." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 110, + 289, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 110, + 289, + 409 + ], + "spans": [ + { + "bbox": [ + 46, + 110, + 289, + 409 + ], + "type": "text", + "content": "Early solutions [11, 12, 51] that reconstruct dynamic meshes with per-frame texture maps are time-consuming and vulnerable to occlusions or lack of textures. Recent advances in neural rendering [37, 62, 71] bring huge potential for human-centric modeling. Most notably, the variants of Neural Radiance Field (NeRF) [37] achieve compelling novel view synthesis, which can enable real-time rendering performance [38, 59, 67] even for dynamic scenes [45, 63, 77], and can be extended to the generative setting without per-scene training [23, 69, 80]. However, less attention is paid to the rich and diverse interactions between humans and objects, mainly due to the severe lack of dense-view human-object datasets. Actually, existing datasets of human-object interactions are mostly based on optical markers [61] or sparse RGB/RGBD sensors [6, 26], without sufficient appearance supervision for neural rendering tasks. As a result, the literature on neural human-object rendering [21,57] is surprisingly sparse, let alone further exploring the real-time or generative directions. Besides, existing neural techniques [52, 77] suffer from tedious training procedures due to the human-object occlusion, and hence infeasible for building a large-scale dataset. In a nutshell, despite the recent tremendous thriving of neural rendering, the lack of both a rich dataset and an efficient reconstruction scheme constitute barriers in human-object modeling." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 411, + 287, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 411, + 287, + 506 + ], + "spans": [ + { + "bbox": [ + 46, + 411, + 287, + 506 + ], + "type": "text", + "content": "In this paper, we present NeuralDome, a neural pipeline that takes multi-view dome capture as inputs and conducts accurate 3D modeling and photo-realistic rendering of complex human-object interaction. As shown in Fig. 1, NeuralDome exploits layer-wise neural modeling to produce rich and multi-modality outputs including the geometry of dynamic human, object shapes and tracked poses, as well as a free-view rendering of the sequence." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 509, + 287, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 509, + 287, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 509, + 287, + 640 + ], + "type": "text", + "content": "Specifically, we first capture a novel human-object dome (HODome) dataset that consists of 274 human-object interacting sequences, covering 23 diverse 3D objects and 10 human subjects (5 males and 5 females) in various apparels. We record multi-view video sequences of natural interactions between the human subjects and the objects where each sequence is about 60s in length using a dome with 76 RGB cameras, resulting in 71 million video frames. We also provide an accurate pre-scanned 3D template for each object and utilize sparse optical markers to track individual objects throughout the sequences." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "text", + "content": "To process the HODome dataset, we adopt an extended Neural Radiance Field (NeRF) pipeline. The brute-force adoption of off-the-shelf neural techniques such as Instant-NSR [79] and Instant-NGP [38], although effective, do not separate objects from human subjects and therefore lack sufficient fidelity to model their interactions. We instead" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 312 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 312 + ], + "type": "text", + "content": "introduce a layer-wise neural processing pipeline. Specifically, we first perform a joint optimization based on the dense inputs for accurately tracking human motions using the parametric SMPL-X model [43] as well as localizing objects using template meshes. We then propose an efficient layer-wise neural rendering scheme where the humans and objects are formulated as a pose-embedded dynamic NeRF and a static NeRF with tracked 6-DoF rigid poses, respectively. Such a layer-wise representation effectively exploits temporal information and robustly tackles the occluded regions under interactions. We further introduce an object-aware ray sampling strategy to mitigate artifacts during layer-wise training, as well as template-aware geometry regularizers to enforce contact-aware deformations. Through weak segmentation supervision, we obtain the decoupled and occlusion-free appearances for both the humans and the objects at a high fidelity amenable for training the input multi-view inputs for a variety of tasks from monocular motion capture to free-view rendering from sparse multi-view inputs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 318, + 312, + 508, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 312, + 508, + 323 + ], + "spans": [ + { + "bbox": [ + 318, + 312, + 508, + 323 + ], + "type": "text", + "content": "To summarize, our main contributions include:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 332, + 545, + 495 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 317, + 332, + 545, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 332, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 317, + 332, + 545, + 380 + ], + "type": "text", + "content": "- We introduce NeuralDome, a neural pipeline, to accurately track humans and objects, conduct layer-wise geometry reconstruction, and enable novel-view synthesis, from multi-view HOI video inputs." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 317, + 384, + 545, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 384, + 545, + 456 + ], + "spans": [ + { + "bbox": [ + 317, + 384, + 545, + 456 + ], + "type": "text", + "content": "- We collect a comprehensive dataset that we call HODome that will be disseminated to the community, with both raw data and the output modalities including separated geometry and rendering of individual objects and human subjects, their tracking results, free-view rendering results, etc." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 317, + 460, + 545, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 460, + 545, + 495 + ], + "spans": [ + { + "bbox": [ + 317, + 460, + 545, + 495 + ], + "type": "text", + "content": "- We demonstrate using the dataset to train networks for a variety of visual inference tasks with complex human object interactions." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 507, + 397, + 520 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 507, + 397, + 520 + ], + "spans": [ + { + "bbox": [ + 306, + 507, + 397, + 520 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 527, + 451, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 527, + 451, + 540 + ], + "spans": [ + { + "bbox": [ + 306, + 527, + 451, + 540 + ], + "type": "text", + "content": "2.1. Neural Human Rendering" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": "Various 3D data representations have been explored for neural human rendering, such as point-clouds [4, 41, 72], textured meshes [14, 28, 29, 54], and volumes [31, 32]. Implicit occupancy function-based methods [17, 19, 48, 49] can recover detailed 3D human geometry from sparse 2D images without faithful appearance synthesis. The recent NeRF [37] technique brings huge potential for 3D photorealistic view synthesis [9, 10, 30, 34, 36, 38, 39, 64, 69, 75] and geometry modeling [7, 33, 68]. Further explorations extend it to dynamic scenes [7, 35, 42, 46, 67, 77], especially for humans. Existing works equip NeRF with pose-embeddings [23, 27, 40, 45, 80], learnable skinning weights [25, 44, 70] and even generalization across individuals [23, 66, 80]. However, these works only focus on" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8835" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 71, + 533, + 152 + ], + "blocks": [ + { + "bbox": [ + 61, + 71, + 533, + 152 + ], + "lines": [ + { + "bbox": [ + 61, + 71, + 533, + 152 + ], + "spans": [ + { + "bbox": [ + 61, + 71, + 533, + 152 + ], + "type": "table", + "html": "
Datasets#Cam-view# Frame(M)ResolutionFpsMarkerObj. Num.Human Annot.Novel viewsGeometryHuman TexturedNeural RepresentationObject Appearance
NTU [26]3~ 341920 × 108030×NA✓NA××××
PiGr [50]10.1960 × 5405×NA✓NA×××✓
GRAB [61]×NANANA✓51××××××
PROX [16]10.11920 × 108030✓NA✓××××✓
BEHAVE [6]40.152048 × 153630×20×××××✓
InterCap [18]60.071920 × 108030×10✓××××✓
Our76713840 × 216060✓23✓✓✓✓✓✓
", + "image_path": "478840b07deb591a49c4581a19f0f57c9e0287d2fafd35f34f6dd703335dd2a0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 160, + 547, + 196 + ], + "lines": [ + { + "bbox": [ + 46, + 160, + 547, + 196 + ], + "spans": [ + { + "bbox": [ + 46, + 160, + 547, + 196 + ], + "type": "text", + "content": "Table 1. Dataset Comparisons. We compare our proposed HODome dataset with the existing publicly available human-object dataset. HODome has the largest scale of human-object interactions in terms of the number of frames (#Frame), camera view (#Cam-view), and modality. \"Obj. Num.\" represents the object number. \"Human Annot.\" represents annotation from professional annotators." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 48, + 212, + 286, + 354 + ], + "blocks": [ + { + "bbox": [ + 48, + 212, + 286, + 354 + ], + "lines": [ + { + "bbox": [ + 48, + 212, + 286, + 354 + ], + "spans": [ + { + "bbox": [ + 48, + 212, + 286, + 354 + ], + "type": "image", + "image_path": "e0899a0cd871ac212bc82fc31dd32ab46ad0a0d3b0462149c3af250f16a9e4f8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 362, + 288, + 429 + ], + "lines": [ + { + "bbox": [ + 46, + 362, + 288, + 429 + ], + "spans": [ + { + "bbox": [ + 46, + 362, + 288, + 429 + ], + "type": "text", + "content": "Figure 2. HODome Modality. HODome features multiple modalities of data format and annotations, including a) Color image, b) Human-object mask (Homask), c) MoCap (SMPL-X parameters and located object template), d) Geometry, e) digital assets in neural representations, f) Novel view and g) Keypoints (25 for bodies and 42 for hands with human annotation)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 449, + 287, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 287, + 498 + ], + "type": "text", + "content": "a single person. The most relevant works [53, 77] aim to model multi-person interactions, but they cannot handle more complex interactions due to the lack of large-scale human-object interactions dataset." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 503, + 185, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 503, + 185, + 517 + ], + "spans": [ + { + "bbox": [ + 47, + 503, + 185, + 517 + ], + "type": "text", + "content": "2.2. Human-object Modeling" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 521, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 287, + 715 + ], + "type": "text", + "content": "Only a few works [6, 16, 18, 61, 73, 78] consider to jointly model whole-body interactions. Early solutions [11, 12, 51] that reconstruct meshes with per-frame texture maps, which are vulnerable to occlusions and lack of textures. Recently, several works [6, 16, 18, 61] explore the relationship via several interaction constraints, such as contact map [6, 18, 61], spatial arrangement [78] and physically plausible constraint [74]. However, these methods only produce a relative arrangement. The most relevant works [21, 56, 58] aim to render human-object interactions with volumetric fusion [56], neural texturing blending [58] and volumetric rendering [21]. However, without sufficient appearance supervision, their quality is limited. Comparably, our layerwise neural representations can be converted to high-quality geometry and support separate free-view rendering for human or object." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 214, + 440, + 227 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 214, + 440, + 227 + ], + "spans": [ + { + "bbox": [ + 306, + 214, + 440, + 227 + ], + "type": "text", + "content": "2.3. Human-centric Dataset" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 233, + 546, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 233, + 546, + 426 + ], + "spans": [ + { + "bbox": [ + 304, + 233, + 546, + 426 + ], + "type": "text", + "content": "A variety of human-centric datasets have been developed for human-only capturing or rendering tasks. Early datasets combine multi-camera RGB video capture with synchronized ground-truth 3D skeletons [20, 22, 55], parametric model [24, 65], scanned mesh [15, 76], without human-object interaction. Recent datasets capture human-object interactions using optical markers [61] or sparse RGB sensors [6, 26], without sufficient appearance supervision. As a result, the literature on neural human-object rendering [21, 57] is surprisingly rare. High-end works [11, 13] use dense cameras for reconstruction and rendering of humans and objects through mesh reconstruction and motion tracking, but without texture and neural representations. To fill this gap, we propose NeuralDome for capturing and rendering human-object interactions, facilitating various generative human-object tasks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 437, + 411, + 450 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 437, + 411, + 450 + ], + "spans": [ + { + "bbox": [ + 306, + 437, + 411, + 450 + ], + "type": "text", + "content": "3. HODome Dataset" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 457, + 545, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 457, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 304, + 457, + 545, + 589 + ], + "type": "text", + "content": "We present the HODome dataset for capturing and rendering photo-realistic human-object interactions. It consists of 274 human-object interacting sequences, covering 23 diverse 3D objects and 10 subjects (5 males and 5 females) under various apparels. For each sequence, we record the naturally interacting scene for about 60s using 76 RGB cameras for dense-view at " + }, + { + "bbox": [ + 304, + 457, + 545, + 589 + ], + "type": "inline_equation", + "content": "3840 \\times 2160" + }, + { + "bbox": [ + 304, + 457, + 545, + 589 + ], + "type": "text", + "content": " resolution and 60 frame-per-seconds (fps), resulting in roughly 71 million frames. HODome consists of rich labels covering different aspects of HOI capturing and rendering labels (see Fig.2). See Tab. 1 for comparison with other datasets." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 598, + 440, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 598, + 440, + 612 + ], + "spans": [ + { + "bbox": [ + 306, + 598, + 440, + 612 + ], + "type": "text", + "content": "3.1. Data Capturing System" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 617, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 546, + 715 + ], + "type": "text", + "content": "To construct the HODome dataset, we use 76 Z-CAM cinema cameras with sufficient appearance supervision for neural rendering tasks and 16 Optitrack MoCap cameras [1] for accurate human-object tracking tasks. The Z-CAM system and Optitrack system are synchronized to record the RGB and MoCap data together. We use a publicly available tool [2] to estimate the intrinsic camera parameters and extrinsic camera parameters." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8836" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 70, + 537, + 201 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 537, + 201 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 537, + 201 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 537, + 201 + ], + "type": "image", + "image_path": "50ae51d5779015aa428705f59a1fa4f33feb545fe7b54a145b6f8f2f4479273f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 209, + 547, + 255 + ], + "lines": [ + { + "bbox": [ + 46, + 209, + 547, + 255 + ], + "spans": [ + { + "bbox": [ + 46, + 209, + 547, + 255 + ], + "type": "text", + "content": "Figure 3. Overview of NeuralDome. Given the 76-view RGB stream as input, we first jointly track human skeletal motion and object rigid motion. Then with the tracked motion priors, we decouple the human-object interaction scenes via a layer-wise neural rendering scheme, to generate human/object renderings separately and corresponding acceptable segmentation maps. Blending with the high-fidelity captured views, we obtain the layer-wise neural representations in our HODome dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 274, + 151, + 288 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 274, + 151, + 288 + ], + "spans": [ + { + "bbox": [ + 47, + 274, + 151, + 288 + ], + "type": "text", + "content": "3.2. Dataset Modality" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 294, + 287, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 294, + 287, + 389 + ], + "spans": [ + { + "bbox": [ + 46, + 294, + 287, + 389 + ], + "type": "text", + "content": "Human MoCap and Object 6D Pose. For appearance realism, we do not place markers on human actors. Thus we detect the human joints from 76 RGB images by running the whole-body Openpose [8] to perform markerless MoCap. And we treat each object as rigid and solve the rigid object pose estimation from markers using the Iterative Closest Point(ICP) algorithm [5, 81]. See Fig. 3 for the pipeline of our method." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 392, + 287, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 392, + 287, + 511 + ], + "spans": [ + { + "bbox": [ + 46, + 392, + 287, + 511 + ], + "type": "text", + "content": "Human-Object Neural Representations. Under the dense-view setting, one can apply off-the-shell neural techniques, i.e., Instant-NSR [79] and Instant-NGP [38] to efficiently obtain per-frame geometry and novel-view synthesis of the whole human-object sequence, respectively. However, neither schemes separate object from human subject. We therefore propose a layer-wise neural scheme to separately recover neural representations of human and object, enabling both high-fidelity geometry reconstruction and free-view appearance rendering (Sec. 4.2)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 513, + 288, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 513, + 288, + 634 + ], + "spans": [ + { + "bbox": [ + 46, + 513, + 288, + 634 + ], + "type": "text", + "content": "Data Annotation. Our dataset includes 23 objects with varying scales and interaction types. Each object was pre-scanned using off-the-shelf multi-view software packages [2]. Further, following the previous methods [6, 61], we annotated its pseudo contact label that computes from a threshold distance. To provide more accurate results for the quantitative benchmark, we annotate a separate quantitative subset as our test set with human-annotated segmentation and hand joints by the professional annotator. Our datasets will be available for research purposes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 645, + 217, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 645, + 217, + 658 + ], + "spans": [ + { + "bbox": [ + 47, + 645, + 217, + 658 + ], + "type": "text", + "content": "4. Neural Modeling on HODome" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": "We introduce a neural pipeline to produce the rich digital assets from the dense-view input of each human-object interaction (HOI) sequence in HODome, including accurate tracking, high-quality geometry reconstruction and novel-" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 274, + 545, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 274, + 545, + 406 + ], + "spans": [ + { + "bbox": [ + 304, + 274, + 545, + 406 + ], + "type": "text", + "content": "view rendering. As illustrated in Fig. 3, given 76-view videos, we first perform a joint optimization scheme to accurately capture both the human skeletal motions and object rigid motion (Sec. 4.1). Then, based on the tracked human-object motions and shape priors, we decouple the humans and objects in HOI scenarios via a layer-wise neural human-object rendering scheme and a corresponding HOI-aware optimization strategy (Sec. 4.2). Such digital assets of humans, objects and the entire HOI scenes can naturally enable further geometry and appearance analysis of HOI scenarios." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 415, + 443, + 428 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 415, + 443, + 428 + ], + "spans": [ + { + "bbox": [ + 305, + 415, + 443, + 428 + ], + "type": "text", + "content": "4.1. Human-object Tracking" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 434, + 545, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 434, + 545, + 458 + ], + "spans": [ + { + "bbox": [ + 305, + 434, + 545, + 458 + ], + "type": "text", + "content": "Here we introduce our human-object tracking scheme with the aid of object markers." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 460, + 546, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 546, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 546, + 604 + ], + "type": "text", + "content": "Tracking initialization. We initialize the human tracking process by fitting the SMPL-X model to the keypoints of each view. Note that we utilize the off-the-shell toolbox Easymocap [3] to get the initial per-frame parameters, i.e., pose " + }, + { + "bbox": [ + 304, + 460, + 546, + 604 + ], + "type": "inline_equation", + "content": "\\theta_{t}" + }, + { + "bbox": [ + 304, + 460, + 546, + 604 + ], + "type": "text", + "content": ", shape " + }, + { + "bbox": [ + 304, + 460, + 546, + 604 + ], + "type": "inline_equation", + "content": "\\beta_{t}" + }, + { + "bbox": [ + 304, + 460, + 546, + 604 + ], + "type": "text", + "content": ", facial expression " + }, + { + "bbox": [ + 304, + 460, + 546, + 604 + ], + "type": "inline_equation", + "content": "\\psi_{t}" + }, + { + "bbox": [ + 304, + 460, + 546, + 604 + ], + "type": "text", + "content": ", and translation " + }, + { + "bbox": [ + 304, + 460, + 546, + 604 + ], + "type": "inline_equation", + "content": "\\gamma_{t}" + }, + { + "bbox": [ + 304, + 460, + 546, + 604 + ], + "type": "text", + "content": " on each frame " + }, + { + "bbox": [ + 304, + 460, + 546, + 604 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 460, + 546, + 604 + ], + "type": "text", + "content": ". As for the object, following GRAB [61], we regard each object as rigid and only estimate the rotation " + }, + { + "bbox": [ + 304, + 460, + 546, + 604 + ], + "type": "inline_equation", + "content": "R_{t} \\in \\mathbb{S0}(3)" + }, + { + "bbox": [ + 304, + 460, + 546, + 604 + ], + "type": "text", + "content": " and rigid translation " + }, + { + "bbox": [ + 304, + 460, + 546, + 604 + ], + "type": "inline_equation", + "content": "T_{t} \\in \\mathbb{R}^{3}" + }, + { + "bbox": [ + 304, + 460, + 546, + 604 + ], + "type": "text", + "content": " with respect to its pre-scanned template. Specifically, we conduct the rigid-ICP technique [5, 81] to compute the rigid transformation between the per-frame markers, which serves as the initial object pose." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 605, + 547, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 665 + ], + "type": "text", + "content": "Joint optimization for human-object tracking. With human and object pose initialization, we further perform a joint optimization scheme to ensure correct human-object contact. We impose constraints to ensure plausible interactions as the following form:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 316, + 674, + 545, + 713 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 674, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 674, + 545, + 713 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} E \\left(\\beta_ {t}, \\theta_ {t}, \\psi_ {t}, \\gamma_ {t}, R _ {t}, T _ {t}\\right) = E _ {\\text {s m p l}} + \\lambda_ {\\text {c o n t a c t}} E _ {\\text {c o n t a c t}} + \\\\ \\lambda_ {\\text {h o m a s k}} E _ {\\text {h o m a s k}} + \\lambda_ {\\text {m a k e r}} E _ {\\text {m a k e r}}, \\tag {1} \\\\ \\end{array}", + "image_path": "e6bf97627adf4751dbb8f130339643d7443f729da8681fbe4a15d2b67c6b4a56.jpg" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8837" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 70, + 289, + 240 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 289, + 240 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 289, + 240 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 289, + 240 + ], + "type": "image", + "image_path": "8fc6c6ae5126ab084ccaa7a81b4d86b3a38a6acf3098c59b56958722e75a476a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 250, + 289, + 327 + ], + "lines": [ + { + "bbox": [ + 46, + 250, + 289, + 327 + ], + "spans": [ + { + "bbox": [ + 46, + 250, + 289, + 327 + ], + "type": "text", + "content": "Figure 4. Layer-wise Neural Human-Object Rendering Details. We sample a ray uniformly in the bounding box of the performer and near the template for the object. We condition both the human and object layer with appearance codes " + }, + { + "bbox": [ + 46, + 250, + 289, + 327 + ], + "type": "inline_equation", + "content": "(l_o^c, l_h^c)" + }, + { + "bbox": [ + 46, + 250, + 289, + 327 + ], + "type": "text", + "content": ". The dynamic human layer contains an additional pose embedding " + }, + { + "bbox": [ + 46, + 250, + 289, + 327 + ], + "type": "inline_equation", + "content": "(E_p)" + }, + { + "bbox": [ + 46, + 250, + 289, + 327 + ], + "type": "text", + "content": " conditioned deformation field. The sample colors and densities are merged, sorted and accumulated into pixel colors." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 349, + 287, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 349, + 287, + 445 + ], + "spans": [ + { + "bbox": [ + 46, + 349, + 287, + 445 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 349, + 287, + 445 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{spl}}" + }, + { + "bbox": [ + 46, + 349, + 287, + 445 + ], + "type": "text", + "content": " is the multi-view data fitting term describing the " + }, + { + "bbox": [ + 46, + 349, + 287, + 445 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 46, + 349, + 287, + 445 + ], + "type": "text", + "content": " distance between estimated joints and detected joints, following SMPLify-X [43] and PROX [16]. Besides, we impose contact term " + }, + { + "bbox": [ + 46, + 349, + 287, + 445 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{contact}}" + }, + { + "bbox": [ + 46, + 349, + 287, + 445 + ], + "type": "text", + "content": ", human-object silhouette loss term " + }, + { + "bbox": [ + 46, + 349, + 287, + 445 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{homask}}" + }, + { + "bbox": [ + 46, + 349, + 287, + 445 + ], + "type": "text", + "content": ", and object marker align term " + }, + { + "bbox": [ + 46, + 349, + 287, + 445 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{maker}}" + }, + { + "bbox": [ + 46, + 349, + 287, + 445 + ], + "type": "text", + "content": " to ensure the plausibility of human-object interaction tracking. Due to page limitation, we have to defer more details of these terms in the Appendix." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 453, + 282, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 453, + 282, + 467 + ], + "spans": [ + { + "bbox": [ + 47, + 453, + 282, + 467 + ], + "type": "text", + "content": "4.2. Layer-wise Neural Human-Object Rendering" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 472, + 287, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 472, + 287, + 508 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 287, + 508 + ], + "type": "text", + "content": "Here we introduce our layer-wise image generation pipeline via neural rendering, without tedious manual efforts for geometry separation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 510, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 288, + 714 + ], + "type": "text", + "content": "Layered neural human-object modeling. For ease of instances separation, we represent the human-object interaction (HOI) scenes as a continuous layered neural radiance field, following ST-NeRF [77], as shown in Fig. 4. To leverage the shape prior imposed by the captured parametric SMPL-X model, we adopt a pose embedded dynamic NeRF similar to HumanNeRF [80] as the human layer. We use the skeletal pose to bridge the live frames to canonical space and an additional deformation MLP to learn subtle non-rigid deformation. Rather than using the pixel-aligned image feature, we leverage latent codes for time-varying human appearance capture to get rid of input images. Opposite to the human layer, we model objects as rigid static radiance fields in canonical space. The live frames are transformed to canonical space via object poses to maintain a globally consistent density field. Similarly, we adopt appearance latent codes for the time-varying shadow during interactions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": "Dynamic human-object volume rendering. We utilize the layered volume rendering technique described in ST-NeRF [77] to render our neural human-object scenes. For a camera ray intersecting with the " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "i_{th}" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " entity at any times-tamp, we compute the ray segment as the depth of intersection points " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "d_f^j" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "d_f^i" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": ". We evenly partition each segment into " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " bins and sample one point uniformly from each bin:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 171, + 547, + 205 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 171, + 547, + 205 + ], + "spans": [ + { + "bbox": [ + 304, + 171, + 547, + 205 + ], + "type": "interline_equation", + "content": "p _ {j} ^ {i} \\sim \\mathcal {U} \\left[ d _ {n} ^ {i} + \\frac {j - 1}{N} \\left(d _ {f} ^ {i} - d _ {n} ^ {i}\\right), d _ {n} ^ {i} + \\frac {j}{N} \\left(d _ {f} ^ {i} - d _ {n} ^ {i}\\right) \\right], j \\in [ 1, 2, \\dots , N ], \\tag {2}", + "image_path": "4d34b02812a35918a1d4d94fd0c8bb7602a5e5820e0af3655db4aac247ae5769.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 205, + 545, + 325 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 205, + 545, + 325 + ], + "spans": [ + { + "bbox": [ + 304, + 205, + 545, + 325 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 205, + 545, + 325 + ], + "type": "inline_equation", + "content": "\\mathcal{U}" + }, + { + "bbox": [ + 304, + 205, + 545, + 325 + ], + "type": "text", + "content": " means uniform distribution. We impose the shape prior encoded in the tracked object template, on the ray sampling scheme for efficient training. We compute the ray-object intersection and uniformly sample a few points in the narrow segment where the first intersection point lies. For human, we simply sample in the ray-bounding box intersection segment. The samples from different segments are then merged and sorted by depth into " + }, + { + "bbox": [ + 304, + 205, + 545, + 325 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 205, + 545, + 325 + ], + "type": "text", + "content": " samples in total. We then compute a pixel's color by accumulating the radiance at sampled points:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 353, + 333, + 545, + 365 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 333, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 353, + 333, + 545, + 365 + ], + "type": "interline_equation", + "content": "C = \\sum_ {i = 1} ^ {M} T \\left(p _ {i}\\right) \\left[ \\left(1 - e ^ {- \\sigma_ {p _ {i}} \\delta_ {p _ {i}}}\\right) c _ {p _ {i}} \\right], \\tag {3}", + "image_path": "4291c42ade8a81740fa2cecaef81c50c5252c45d06ee365134a5cf287b066376.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 370, + 545, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 370, + 545, + 397 + ], + "spans": [ + { + "bbox": [ + 304, + 370, + 545, + 397 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 370, + 545, + 397 + ], + "type": "inline_equation", + "content": "\\delta_{p_i}" + }, + { + "bbox": [ + 304, + 370, + 545, + 397 + ], + "type": "text", + "content": " is the distance between adjacent points, " + }, + { + "bbox": [ + 304, + 370, + 545, + 397 + ], + "type": "inline_equation", + "content": "c_{p_i}" + }, + { + "bbox": [ + 304, + 370, + 545, + 397 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 370, + 545, + 397 + ], + "type": "inline_equation", + "content": "\\sigma_{p_i}" + }, + { + "bbox": [ + 304, + 370, + 545, + 397 + ], + "type": "text", + "content": " are color and density, and " + }, + { + "bbox": [ + 304, + 370, + 545, + 397 + ], + "type": "inline_equation", + "content": "T(p_i) = \\prod_j^{i - 1}e^{-\\sigma_{p_j}\\delta_{p_j}}" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 397, + 545, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 397, + 545, + 433 + ], + "spans": [ + { + "bbox": [ + 304, + 397, + 545, + 433 + ], + "type": "text", + "content": "HOI-aware training scheme. Here we introduce an effective optimization scheme to train our neural layer-wise HOI model. We first leverage a photometric loss:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 376, + 441, + 545, + 468 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 441, + 545, + 468 + ], + "spans": [ + { + "bbox": [ + 376, + 441, + 545, + 468 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c} = \\sum_ {\\mathbf {r} \\in \\mathcal {R}} \\| C _ {\\mathbf {r}} - \\hat {C} _ {\\mathbf {r}} \\| _ {2} ^ {2}, \\tag {4}", + "image_path": "b0e44997d785f22a367cc989c320087071cb5643d6f02c7eea9b2d3aad368929.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 477, + 545, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 477, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 304, + 477, + 545, + 502 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 477, + 545, + 502 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 304, + 477, + 545, + 502 + ], + "type": "text", + "content": " is a ray in the training ray set " + }, + { + "bbox": [ + 304, + 477, + 545, + 502 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 304, + 477, + 545, + 502 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 477, + 545, + 502 + ], + "type": "inline_equation", + "content": "C_{\\mathbf{r}}" + }, + { + "bbox": [ + 304, + 477, + 545, + 502 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 477, + 545, + 502 + ], + "type": "inline_equation", + "content": "\\hat{C}_{\\mathbf{r}}" + }, + { + "bbox": [ + 304, + 477, + 545, + 502 + ], + "type": "text", + "content": " are the corresponding rendered and observed colors." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 502, + 545, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 502, + 545, + 610 + ], + "spans": [ + { + "bbox": [ + 304, + 502, + 545, + 610 + ], + "type": "text", + "content": "Note that in HOI scenarios, the human layer inevitably intersects with other object layers. This fact introduces ambiguities and cannot be addressed by a simple photometric loss. Thus we design three additional regularizers based on the tracked object template to alleviate this issue. To enforce the object layers to be solid surface radiance fields that do not contribute to human appearance, we adopt an occupancy and a sparsity regularizer inside and outside the object template, i.e.," + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 618, + 545, + 645 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 545, + 645 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {o} = \\sum_ {p \\in \\mathcal {O}} \\left(\\Omega_ {-} (p, \\mathcal {M}) \\| e ^ {(- \\sigma_ {p})} \\| _ {2} ^ {2} + \\Omega_ {+} (p, \\mathcal {M}) \\| \\sigma_ {p} \\| _ {2} ^ {2}\\right), \\tag {5}", + "image_path": "fc10f86d08fbf17b7f538dff0d7979ca586a3a4d8bd984b0bd90b0d5e7bda009.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 653, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 689 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 653, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 304, + 653, + 545, + 689 + ], + "type": "text", + "content": " is the template mesh, " + }, + { + "bbox": [ + 304, + 653, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\mathcal{O}" + }, + { + "bbox": [ + 304, + 653, + 545, + 689 + ], + "type": "text", + "content": " is a set of points randomly sampled in the object's bounding box, " + }, + { + "bbox": [ + 304, + 653, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\Omega_{+}" + }, + { + "bbox": [ + 304, + 653, + 545, + 689 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 653, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\Omega_{-}" + }, + { + "bbox": [ + 304, + 653, + 545, + 689 + ], + "type": "text", + "content": " indicate whether a point locates outside or inside the mesh." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "content": "Recall that we have encoded human-object contact prior by jointly tracking human-object at the skeletal-pose level" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8838" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 69, + 527, + 249 + ], + "blocks": [ + { + "bbox": [ + 70, + 69, + 527, + 249 + ], + "lines": [ + { + "bbox": [ + 70, + 69, + 527, + 249 + ], + "spans": [ + { + "bbox": [ + 70, + 69, + 527, + 249 + ], + "type": "image", + "image_path": "d0913aeafdcba991c57124898a981e3dace7c18fb8fd34efe736909eabb7b8e0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 256, + 547, + 281 + ], + "lines": [ + { + "bbox": [ + 46, + 256, + 547, + 281 + ], + "spans": [ + { + "bbox": [ + 46, + 256, + 547, + 281 + ], + "type": "text", + "content": "Figure 5. The geometry, MoCap and neural modeling results from our dataset. HODome includes various interaction sequences, such as \"holding a vase\", \"opening the cupboard\", \"sitting on the sofa\" and \"moving the table\"." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 289, + 289, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 289, + 289, + 350 + ], + "spans": [ + { + "bbox": [ + 46, + 289, + 289, + 350 + ], + "type": "text", + "content": "(Eq. 1). However, due to the misalignment of SMPL shape and actual human geometry, the human body and clothes may still overlap the object space, which goes against real-world physics. Therefore, we implicitly constrain the deformation net to predict contact-aware non-rigid deformation:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 109, + 356, + 287, + 381 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 109, + 356, + 287, + 381 + ], + "spans": [ + { + "bbox": [ + 109, + 356, + 287, + 381 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {h} = \\sum_ {p \\in \\mathcal {H}} \\Omega_ {-} (p, \\mathcal {M}) \\| \\sigma_ {p} \\| _ {2} ^ {2}, \\tag {6}", + "image_path": "45aaefc8d07b31c3700f3aa9074c132b191b997f07f29b9ad2c706cb109558e3.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 388, + 287, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 388, + 287, + 411 + ], + "spans": [ + { + "bbox": [ + 46, + 388, + 287, + 411 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 388, + 287, + 411 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 46, + 388, + 287, + 411 + ], + "type": "text", + "content": " is the set of random samples in the human's bounding box." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 412, + 290, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 412, + 290, + 543 + ], + "spans": [ + { + "bbox": [ + 46, + 412, + 290, + 543 + ], + "type": "text", + "content": "We further explore a weakly-supervising scheme that utilizes object texture cues to help decouple the human and object entities. The key observation is that, with accurate object tracking, the object radiance field fused across frames can be rendered with higher confidence. That is, we train the entire scene first and render the objects only to capture views, and the pixels similar to the captured ones are labeled as object pixels. In this way, we obtain coarse object segmentation maps " + }, + { + "bbox": [ + 46, + 412, + 290, + 543 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 412, + 290, + 543 + ], + "type": "text", + "content": " that serve as pseudo supervision: we adopt label-wise integration [53] to render the layer labels " + }, + { + "bbox": [ + 46, + 412, + 290, + 543 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 412, + 290, + 543 + ], + "type": "text", + "content": " of rays:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 96, + 543, + 287, + 574 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 543, + 287, + 574 + ], + "spans": [ + { + "bbox": [ + 96, + 543, + 287, + 574 + ], + "type": "interline_equation", + "content": "\\mathbf {s} = \\sum_ {i = 1} ^ {M} T \\left(p _ {i}\\right) \\left[ \\left(1 - e ^ {- \\sigma_ {p _ {i}} \\delta_ {p _ {i}}}\\right) \\mathbf {l} _ {p _ {i}} \\right], \\tag {7}", + "image_path": "e512c2e5fc820171516816134b3d03806b4db59c7f91d0c03583ea8dc3b3c167.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 578, + 286, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 578, + 286, + 603 + ], + "spans": [ + { + "bbox": [ + 47, + 578, + 286, + 603 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 578, + 286, + 603 + ], + "type": "inline_equation", + "content": "\\mathbf{l}_{p_i}" + }, + { + "bbox": [ + 47, + 578, + 286, + 603 + ], + "type": "text", + "content": " is one-hot label indicates which layer the point " + }, + { + "bbox": [ + 47, + 578, + 286, + 603 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 47, + 578, + 286, + 603 + ], + "type": "text", + "content": " belongs to. We apply a semantic loss for object rays " + }, + { + "bbox": [ + 47, + 578, + 286, + 603 + ], + "type": "inline_equation", + "content": "S_o" + }, + { + "bbox": [ + 47, + 578, + 286, + 603 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 119, + 609, + 287, + 634 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 609, + 287, + 634 + ], + "spans": [ + { + "bbox": [ + 119, + 609, + 287, + 634 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s} = \\sum_ {\\mathbf {r} \\in \\mathcal {S} _ {o}} \\| \\mathbf {s} _ {\\mathbf {r}} - \\hat {\\mathbf {s}} _ {\\mathbf {r}} \\| _ {2} ^ {2}, \\tag {8}", + "image_path": "294dcfa4142812fd8393c715767d57fe833005f67e2f5ba2787d73dec8174f2b.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 640, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 640, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 640, + 287, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 640, + 287, + 665 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{s}}_{\\mathbf{r}}" + }, + { + "bbox": [ + 46, + 640, + 287, + 665 + ], + "type": "text", + "content": " is the pseudo semantic label of ray " + }, + { + "bbox": [ + 46, + 640, + 287, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 46, + 640, + 287, + 665 + ], + "type": "text", + "content": ". Note that additional loss can be applied if accurate labels are provided." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 666, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 666, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 666, + 287, + 714 + ], + "type": "text", + "content": "Layer-wise neural representation in HODome. To further use the data, we can render reliable segmentation maps and occlusion-free appearances for humans and objects respectively. However, the renderings tend to be blurred since" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 289, + 547, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 289, + 547, + 421 + ], + "spans": [ + { + "bbox": [ + 304, + 289, + 547, + 421 + ], + "type": "text", + "content": "the information of all the frames is fused. Hence we introduce an enhancement scheme to construct separate high-quality per-frame neural representation for human/object: we blend the visible regions of the high-fidelity observed input views to rendered human/object images using the segmentation maps. The Instant-NGP [38] and Instant-NSR [79] techniques are then applied to reconstruct separate human/object in tens of seconds. Finally, we obtain layer-wise representation which enables real-time photorealistic rendering and benefits the thorough analysis of HOI scenarios." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 422, + 547, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 422, + 547, + 519 + ], + "spans": [ + { + "bbox": [ + 304, + 422, + 547, + 519 + ], + "type": "text", + "content": "Implementations. To build our neural representations, we train our layered models using Adam optimizer with a learning rate that starts from " + }, + { + "bbox": [ + 304, + 422, + 547, + 519 + ], + "type": "inline_equation", + "content": "5e^{-4}" + }, + { + "bbox": [ + 304, + 422, + 547, + 519 + ], + "type": "text", + "content": " and decays exponentially. Using the 76-view 4K resolution videos, the training process takes 8-10 hours on a single NVIDIA 3090 GPU for acceptable appearance and segmentation rendering. Thanks to Instant-NGP [38] and Instant-NSR [79], the layer-wise neural modeling process is conducted in seconds per frame." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 531, + 388, + 545 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 531, + 388, + 545 + ], + "spans": [ + { + "bbox": [ + 306, + 531, + 388, + 545 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 552, + 547, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 552, + 547, + 613 + ], + "spans": [ + { + "bbox": [ + 304, + 552, + 547, + 613 + ], + "type": "text", + "content": "In this section, we compare our neural pipeline with existing state-of-the-arts. We first demonstrate the effectiveness and the generalization ability of our neural modeling approach (5.1). Next, we implement NeuralDome on three specific tasks (5.2)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 621, + 468, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 621, + 468, + 635 + ], + "spans": [ + { + "bbox": [ + 306, + 621, + 468, + 635 + ], + "type": "text", + "content": "5.1. Analysis of Neural Rendering" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "content": "Comparisons. Here we compare our layer-wise neural human-object rendering approach, denoted as 'Ours-layer', with recent state-of-the-art methods, i.e., ST-NeRF [77], NeuralBody(NB) [45], in human-object interaction scenario. For fair comparisons, we select 20 cameras surrounding the performer for training and others for evaluation. We" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8839" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 68, + 289, + 257 + ], + "blocks": [ + { + "bbox": [ + 50, + 68, + 289, + 257 + ], + "lines": [ + { + "bbox": [ + 50, + 68, + 289, + 257 + ], + "spans": [ + { + "bbox": [ + 50, + 68, + 289, + 257 + ], + "type": "image", + "image_path": "15fa79560dc22a3b92c6a65e9be6944aeacef22aec6a805f821c06047b8caaf6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 265, + 287, + 330 + ], + "lines": [ + { + "bbox": [ + 46, + 265, + 287, + 330 + ], + "spans": [ + { + "bbox": [ + 46, + 265, + 287, + 330 + ], + "type": "text", + "content": "Figure 6. Qualitative comparisons of novel view synthesis results. We show ground truth and synthesized images of novel view for NeuralBody [45], and ST-NeRF [77] and our layered human-object representation. Our approach achieves the best performance. We further illustrate our high-fidelity neural representations in HODome in the last column." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 334, + 285, + 405 + ], + "blocks": [ + { + "bbox": [ + 50, + 334, + 285, + 405 + ], + "lines": [ + { + "bbox": [ + 50, + 334, + 285, + 405 + ], + "spans": [ + { + "bbox": [ + 50, + 334, + 285, + 405 + ], + "type": "table", + "html": "
MethodsNB [45]ST-NeRF [77]Ours
ScenesPSNR↑SSIM↑PSNR↑SSIM↑PSNR↑SSIM↑
Bigsofa19.330.89624.020.88632.280.958
Sofa26.730.96528.490.95835.620.987
Table21.940.93322.440.90027.880.949
Average22.670.93124.990.91531.930.964
", + "image_path": "48f0734af74b4dac3b0d5e285ab95b41e536ce2d64f3d2b5c784217dbafb95d1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 449, + 287, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 287, + 568 + ], + "type": "text", + "content": "remove the background layer and train ST-NeRF without the semantic label as ours. Fig. 6 shows several appearance synthesis results. NeuralBody has no ability to model objects which leads to artifacts in the density field. ST-NeRF fails as the bounding boxes of the performer and object almost completely overlap. Our method achieves both human and object modeling and rendering and thus further enables generating high-fidelity neural representation denoted as 'Ours-HODome'. Tab. 2 further illustrates our method significantly outperforms the baselines." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 288, + 714 + ], + "type": "text", + "content": "Evaluation. Here we further evaluate how our scheme contributes to the generated layer-wise neural representations. Fig. 7 shows the single human layer where the performer is \"sitting\" in the air. Note that we have no ground truth of the specific human/object layer, thus we conduct qualitative evaluations only. Let w/o pseudo segmentation and w/o blending denote our human assets generated without pseudo semantic loss and blending-based enhancement scheme. It demonstrates that the pseudo semantic loss effectively helps decouple the human and object layer and our blending scheme further boosts the appearance fidelity of the blurred temporal-fused model to a photo-realistic level." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 307, + 69, + 541, + 263 + ], + "blocks": [ + { + "bbox": [ + 307, + 69, + 541, + 263 + ], + "lines": [ + { + "bbox": [ + 307, + 69, + 541, + 263 + ], + "spans": [ + { + "bbox": [ + 307, + 69, + 541, + 263 + ], + "type": "image", + "image_path": "446b1aead1a9ec1ebb02f849db33f917efe52cfa62691cd186a8d5f5b706234e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 271, + 546, + 316 + ], + "lines": [ + { + "bbox": [ + 305, + 271, + 546, + 316 + ], + "spans": [ + { + "bbox": [ + 305, + 271, + 546, + 316 + ], + "type": "text", + "content": "Figure 7. Qualitative evaluation of each strategy's contribution to layer-wise neural representation generation. The weakly-supervising scheme effectively helps decouple human and object and our blending scheme generates high-quality digital assets." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 308, + 318, + 546, + 365 + ], + "blocks": [ + { + "bbox": [ + 46, + 415, + 287, + 437 + ], + "lines": [ + { + "bbox": [ + 46, + 415, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 46, + 415, + 287, + 437 + ], + "type": "text", + "content": "Table 2. Quantitative comparisons of novel view synthesized appearance on different human-object interaction sequences." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 318, + 546, + 365 + ], + "lines": [ + { + "bbox": [ + 308, + 318, + 546, + 365 + ], + "spans": [ + { + "bbox": [ + 308, + 318, + 546, + 365 + ], + "type": "table", + "html": "
MethodSeparated evaluationJoint evaluation
MPJPE↓PA-MPJPE↓Chamferh↓Chamfero↓V2V↓p.V2V↓
Fit to input16.116.9486.6618.1461.5728.92
PHOSA [78]14.886.9477.6216.6454.5926.09
CHORE [73]10.216.1486.587.6944.9314.93
", + "image_path": "2f627f24f5f6fafeadd38b05f7b5fc451963b69356868c10218ec703c4a18981.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 374, + 545, + 407 + ], + "lines": [ + { + "bbox": [ + 305, + 374, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 305, + 374, + 545, + 407 + ], + "type": "text", + "content": "Table 3. Human object capture benchmark. \"Fit to input\" represents the vanilla method that fits the object template to image and capture human with Frankmocap [47]." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 418, + 427, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 418, + 427, + 430 + ], + "spans": [ + { + "bbox": [ + 306, + 418, + 427, + 430 + ], + "type": "text", + "content": "5.2. Task and benchmark" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 437, + 546, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 546, + 592 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 546, + 592 + ], + "type": "text", + "content": "Human-object Capture Benchmark. HODome provides multi-view sequences with synchronized ground truth of capturing. To demonstrate the capability, we provide a benchmark for human-object capture, shown in Tab. 3. For detailed metric explanation please refer to [6, 73, 74]. PHOSA [78] used the contact map from hand-crafted annotations, suffering from coarse contact information. While CHORE [73] outperforms it with the distance fields predicted from neural networks. It shows the importance of a prepared human-object dataset that supports data-driven human-object capture model. Check out the full capturing benchmark of the dataset at https://arxiv.org/abs/2212.07626." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "type": "text", + "content": "Human-object Geometry Reconstruction Benchmark. Benefiting from the dense-view setting, various human-only geometry reconstruction methods can step towards a human-object geometry reconstruction setting. We can benchmark the state-of-the-art algorithms in sparse view geometry reconstruction tasks. Hence, we evaluate PIFu [48] by training on a subset of HODome. As shown in Fig. 8, we compare the quality result between original PIFu and PIFu-trained on HODome and in-the-wild images. Compared to the original PIFu, our training set can improve the quality of" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8840" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 71, + 87, + 135 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 87, + 135 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 87, + 135 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 87, + 135 + ], + "type": "image", + "image_path": "74525579354cdace85ffdff903c9f173f7f0e827cea1481936cbb59b002f45ef.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 107, + 161, + 119, + 169 + ], + "lines": [ + { + "bbox": [ + 107, + 161, + 119, + 169 + ], + "spans": [ + { + "bbox": [ + 107, + 161, + 119, + 169 + ], + "type": "text", + "content": "NA" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 90, + 72, + 126, + 135 + ], + "blocks": [ + { + "bbox": [ + 90, + 72, + 126, + 135 + ], + "lines": [ + { + "bbox": [ + 90, + 72, + 126, + 135 + ], + "spans": [ + { + "bbox": [ + 90, + 72, + 126, + 135 + ], + "type": "image", + "image_path": "b63b870fe6f5b57c7d5848cec3cfa58168197534f846f10c099e04164648f771.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 129, + 72, + 172, + 135 + ], + "blocks": [ + { + "bbox": [ + 129, + 72, + 172, + 135 + ], + "lines": [ + { + "bbox": [ + 129, + 72, + 172, + 135 + ], + "spans": [ + { + "bbox": [ + 129, + 72, + 172, + 135 + ], + "type": "image", + "image_path": "da7ac4250173b8fbe463747dcf3901144873cfd38441ff40a1b78fbba28d1a23.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 174, + 72, + 222, + 135 + ], + "blocks": [ + { + "bbox": [ + 174, + 72, + 222, + 135 + ], + "lines": [ + { + "bbox": [ + 174, + 72, + 222, + 135 + ], + "spans": [ + { + "bbox": [ + 174, + 72, + 222, + 135 + ], + "type": "image", + "image_path": "5f8cd77a43bfb8f83a5905cb23fe91454cd31d9a4f871c3fc89678f59a9ec847.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 225, + 72, + 272, + 135 + ], + "blocks": [ + { + "bbox": [ + 225, + 72, + 272, + 135 + ], + "lines": [ + { + "bbox": [ + 225, + 72, + 272, + 135 + ], + "spans": [ + { + "bbox": [ + 225, + 72, + 272, + 135 + ], + "type": "image", + "image_path": "891609b21faf132d8da293426b7344ea2c7c6629ec176f5f1ec63aab2a0247ba.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 49, + 139, + 88, + 201 + ], + "blocks": [ + { + "bbox": [ + 49, + 139, + 88, + 201 + ], + "lines": [ + { + "bbox": [ + 49, + 139, + 88, + 201 + ], + "spans": [ + { + "bbox": [ + 49, + 139, + 88, + 201 + ], + "type": "image", + "image_path": "624a47ec8b979c3074ab298f8e13364e51ac1952c655cf81a1da138f235135c1.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 62, + 205, + 78, + 213 + ], + "lines": [ + { + "bbox": [ + 62, + 205, + 78, + 213 + ], + "spans": [ + { + "bbox": [ + 62, + 205, + 78, + 213 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 145, + 135, + 183, + 202 + ], + "blocks": [ + { + "bbox": [ + 145, + 135, + 183, + 202 + ], + "lines": [ + { + "bbox": [ + 145, + 135, + 183, + 202 + ], + "spans": [ + { + "bbox": [ + 145, + 135, + 183, + 202 + ], + "type": "image", + "image_path": "16d70660e3a1c3a6b6fb9ddebfd579b9323b00d0b64ddd3d6b664e05ee9f8169.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 137, + 205, + 171, + 213 + ], + "lines": [ + { + "bbox": [ + 137, + 205, + 171, + 213 + ], + "spans": [ + { + "bbox": [ + 137, + 205, + 171, + 213 + ], + "type": "text", + "content": "PIFu-Origin" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 192, + 132, + 227, + 202 + ], + "blocks": [ + { + "bbox": [ + 192, + 132, + 227, + 202 + ], + "lines": [ + { + "bbox": [ + 192, + 132, + 227, + 202 + ], + "spans": [ + { + "bbox": [ + 192, + 132, + 227, + 202 + ], + "type": "image", + "image_path": "e724f1aece2df2ce899a9b73ea27ddf65efccc4f724ab6b09ac8444f4681b927.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 187, + 203, + 223, + 217 + ], + "lines": [ + { + "bbox": [ + 187, + 203, + 223, + 217 + ], + "spans": [ + { + "bbox": [ + 187, + 203, + 223, + 217 + ], + "type": "text", + "content": "Monocular PIFu-Trained" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 47, + 224, + 287, + 280 + ], + "lines": [ + { + "bbox": [ + 47, + 224, + 287, + 280 + ], + "spans": [ + { + "bbox": [ + 47, + 224, + 287, + 280 + ], + "type": "text", + "content": "Figure 8. Qualitative evaluation on human-object reconstruction. We compare the original PIFu [48] with trained PIFu on both images from our dataset and in-the-wild images. The original PIFu fails to predict the object depth, and PIFu trained on our dataset can approximately predict the shape of the object." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 239, + 135, + 274, + 202 + ], + "blocks": [ + { + "bbox": [ + 239, + 135, + 274, + 202 + ], + "lines": [ + { + "bbox": [ + 239, + 135, + 274, + 202 + ], + "spans": [ + { + "bbox": [ + 239, + 135, + 274, + 202 + ], + "type": "image", + "image_path": "de5e2daad2b958c849cb53506715a8e29d0e1cdb5296626ed52be20e94845e54.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 232, + 203, + 269, + 216 + ], + "lines": [ + { + "bbox": [ + 232, + 203, + 269, + 216 + ], + "spans": [ + { + "bbox": [ + 232, + 203, + 269, + 216 + ], + "type": "text", + "content": "Six View PIFu-Trained" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 65, + 282, + 269, + 325 + ], + "blocks": [ + { + "bbox": [ + 65, + 282, + 269, + 325 + ], + "lines": [ + { + "bbox": [ + 65, + 282, + 269, + 325 + ], + "spans": [ + { + "bbox": [ + 65, + 282, + 269, + 325 + ], + "type": "table", + "html": "
MethodP2S ×10-4↓Chamfer ×10-4↓
Origin PIFu [48]38.72640.947
Monocular PIFu-trained14.65314.483
6-View PIFu-trained3.3764.901
", + "image_path": "4ebb5f0a76ef37b8e8de9bfc114317a86ada32544d6585086e6df3db56777611.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "type": "table", + "bbox": [ + 93, + 373, + 241, + 425 + ], + "blocks": [ + { + "bbox": [ + 47, + 334, + 287, + 367 + ], + "lines": [ + { + "bbox": [ + 47, + 334, + 287, + 367 + ], + "spans": [ + { + "bbox": [ + 47, + 334, + 287, + 367 + ], + "type": "text", + "content": "Table 4. Geometry reconstruction benchmark. Origin PIFu [48] uses the pre-trained model to infer. Monocular PIFu is trained on a single view. 6-View PIFu is trained on 6-view inputs." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 93, + 373, + 241, + 425 + ], + "lines": [ + { + "bbox": [ + 93, + 373, + 241, + 425 + ], + "spans": [ + { + "bbox": [ + 93, + 373, + 241, + 425 + ], + "type": "table", + "html": "
MethodPSNR↑SSIM↑
IBRnet [69]21.430.892
Neuray [30]23.340.909
NeuralHumanFVV [60]21.690.914
NeuralHOIFVV23.100.912
", + "image_path": "403a7a18ef53b879b8218d2668e84f17b8f170dcaffb82a5ec78f4a2bcafab05.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 52, + 433, + 281, + 445 + ], + "lines": [ + { + "bbox": [ + 52, + 433, + 281, + 445 + ], + "spans": [ + { + "bbox": [ + 52, + 433, + 281, + 445 + ], + "type": "text", + "content": "Table 5. Neural rendering benchmark in sparse-view setting." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 456, + 287, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 456, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 47, + 456, + 287, + 491 + ], + "type": "text", + "content": "the reconstructed shape of the objects. For further quantitative analysis, we evaluate the performance using P2S and CD in Tab. 4." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 47, + 494, + 287, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 494, + 287, + 661 + ], + "spans": [ + { + "bbox": [ + 47, + 494, + 287, + 661 + ], + "type": "text", + "content": "Sparse-View Human-object Rendering Benchmark. Our HODome supports various neural rendering tasks, even for human-object interactions. We provide a benchmark on sparse-view rendering tasks and evaluate on IBRNet [69], NeuRay [30] and NeuralHumanFVV [60]. Besides, we also provide a baseline, named NeuralHOIFVV, which uses the projection of trained PIFu's [48] results as depth input, then applies the Neural Blending method presented in [79] to obtain novel-view synthesis. More details about NeuralHOIFVV are referred to the Appendix. We analyze the above methods on HODome and in-the-wild images. The qualitative and quantitative results are shown in Fig. 9 and Tab. 5. Our proposed NeuralHOIFVV can obtain generative novel-view synthesis on HODome and in-the-wild inputs." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 47, + 670, + 124, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 124, + 681 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 124, + 681 + ], + "type": "text", + "content": "5.3. Limitations" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "Although NeuralDome can provide accurate capturing results and layer-wise neural representations, we also want" + } + ] + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 310, + 76, + 349, + 143 + ], + "blocks": [ + { + "bbox": [ + 310, + 76, + 349, + 143 + ], + "lines": [ + { + "bbox": [ + 310, + 76, + 349, + 143 + ], + "spans": [ + { + "bbox": [ + 310, + 76, + 349, + 143 + ], + "type": "image", + "image_path": "6eaa6ccfcf504c35f5ade452b9f6517ce9131c8f2ccb5cd221ba76e2a1101d8d.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 358, + 73, + 384, + 143 + ], + "blocks": [ + { + "bbox": [ + 358, + 73, + 384, + 143 + ], + "lines": [ + { + "bbox": [ + 358, + 73, + 384, + 143 + ], + "spans": [ + { + "bbox": [ + 358, + 73, + 384, + 143 + ], + "type": "image", + "image_path": "1daaf82378f503cfe4ba5ffccbdab163354242a30edae5969bc5c2eac14393dc.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 384, + 74, + 421, + 141 + ], + "blocks": [ + { + "bbox": [ + 384, + 74, + 421, + 141 + ], + "lines": [ + { + "bbox": [ + 384, + 74, + 421, + 141 + ], + "spans": [ + { + "bbox": [ + 384, + 74, + 421, + 141 + ], + "type": "image", + "image_path": "4a0643b8bb8ad2571537467f4d4f233b4f36de28b6667780198a66abadaa30d6.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 381, + 141, + 421, + 205 + ], + "blocks": [ + { + "bbox": [ + 381, + 141, + 421, + 205 + ], + "lines": [ + { + "bbox": [ + 381, + 141, + 421, + 205 + ], + "spans": [ + { + "bbox": [ + 381, + 141, + 421, + 205 + ], + "type": "image", + "image_path": "57fd79734fae9b65d5873a0b5b4e059735e8b181dc25b53fce4e97e88d8056c9.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 362, + 209, + 414, + 217 + ], + "lines": [ + { + "bbox": [ + 362, + 209, + 414, + 217 + ], + "spans": [ + { + "bbox": [ + 362, + 209, + 414, + 217 + ], + "type": "text", + "content": "GT IBRNet" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 423, + 73, + 459, + 206 + ], + "blocks": [ + { + "bbox": [ + 423, + 73, + 459, + 206 + ], + "lines": [ + { + "bbox": [ + 423, + 73, + 459, + 206 + ], + "spans": [ + { + "bbox": [ + 423, + 73, + 459, + 206 + ], + "type": "image", + "image_path": "7134b8c3401540046b7b5fd75b7c2d5c59c8f868e8e115e0a33c06a91acf6766.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 430, + 209, + 454, + 217 + ], + "lines": [ + { + "bbox": [ + 430, + 209, + 454, + 217 + ], + "spans": [ + { + "bbox": [ + 430, + 209, + 454, + 217 + ], + "type": "text", + "content": "NeuRay" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 463, + 73, + 499, + 208 + ], + "blocks": [ + { + "bbox": [ + 463, + 73, + 499, + 208 + ], + "lines": [ + { + "bbox": [ + 463, + 73, + 499, + 208 + ], + "spans": [ + { + "bbox": [ + 463, + 73, + 499, + 208 + ], + "type": "image", + "image_path": "1943accd934a35f81527c8be64e66948d3fef9fede39a25320683f904b982774.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 465, + 209, + 501, + 222 + ], + "lines": [ + { + "bbox": [ + 465, + 209, + 501, + 222 + ], + "spans": [ + { + "bbox": [ + 465, + 209, + 501, + 222 + ], + "type": "text", + "content": "Neural HumanFVV" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 231, + 545, + 275 + ], + "lines": [ + { + "bbox": [ + 305, + 231, + 545, + 275 + ], + "spans": [ + { + "bbox": [ + 305, + 231, + 545, + 275 + ], + "type": "text", + "content": "Figure 9. Qualitative evaluation on neural human-object rendering. We show the comparison of IBRNet [69], NeuRay [30], NeuralHumanFVV [60] and our proposed NeuralHOIFVV. We render one novel view between two input camera views." + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 503, + 74, + 537, + 207 + ], + "blocks": [ + { + "bbox": [ + 503, + 74, + 537, + 207 + ], + "lines": [ + { + "bbox": [ + 503, + 74, + 537, + 207 + ], + "spans": [ + { + "bbox": [ + 503, + 74, + 537, + 207 + ], + "type": "image", + "image_path": "e21a01a06ded1545afa492c11daab668908d812f13ae91bcacc0da7bb875cec6.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 512, + 209, + 535, + 222 + ], + "lines": [ + { + "bbox": [ + 512, + 209, + 535, + 222 + ], + "spans": [ + { + "bbox": [ + 512, + 209, + 535, + 222 + ], + "type": "text", + "content": "Neural HOIFVV" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "bbox": [ + 304, + 293, + 545, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 293, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 304, + 293, + 545, + 425 + ], + "type": "text", + "content": "to highlight some potential limitations of this pipeline. First, NeuralDome only considers single-person interaction with objects and the single-person task is already quite challenging for current research. It is non-trivial to extend the current algorithm to multi-person settings, especially in a crowd scene. Secondly, the reconstruction of 3D holistic scene is not covered in our pipeline. We leave the joint modeling of humans, objects and scene in the future work. Moreover, our dataset was collected under fixed illumination conditions with few backgrounds variance, limiting its generalization ability for other environments." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 306, + 447, + 378, + 459 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 447, + 378, + 459 + ], + "spans": [ + { + "bbox": [ + 306, + 447, + 378, + 459 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 304, + 471, + 545, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 471, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 304, + 471, + 545, + 602 + ], + "type": "text", + "content": "We have presented HODome, the first dataset to jointly capture and render human-object interaction using dense RGB cameras and the Optitrack system. To process the HODome dataset, we have further developed NeuralDome, a neural pipeline that can accurately track humans and objects, conduct layer-wise geometry reconstruction, and enable novel-view synthesis for both the subjects and objects. We believe the dataset could boost the development of capturing and rendering human-object interactions and is very valuable to the community for a wide range of human-object capturing and rendering tasks." + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgement This work was supported by the Shanghai Sailing Program (21YF1429400, 22YF1428800), Shanghai Local College Capacity Building Program (23010503100, 22010502800), NSFC programs (61976138, 61977047), the National Key Research and Development Program (2018YFB2100500), STCSM (2015F0203-000-06), SHMEC (2019-01-07-00-01-E00003) and Shanghai Frontiers Science Center of Humancentered Artificial Intelligence (ShangHAI)." + } + ] + } + ], + "index": 38 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "8841" + } + ] + } + ], + "index": 39 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 92, + 286, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 92, + 286, + 113 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 92, + 286, + 113 + ], + "spans": [ + { + "bbox": [ + 53, + 92, + 286, + 113 + ], + "type": "text", + "content": "[1] Naturalpoint, inc. motion capture systems. https://optitrack.com/. 6.3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 115, + 286, + 136 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 115, + 286, + 136 + ], + "spans": [ + { + "bbox": [ + 53, + 115, + 286, + 136 + ], + "type": "text", + "content": "[2] Reality capture. https://www.capturingreality.com/realitycapture. 3, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 139, + 286, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 139, + 286, + 159 + ], + "spans": [ + { + "bbox": [ + 53, + 139, + 286, + 159 + ], + "type": "text", + "content": "[3] Easymocap - make human motion capture easier. Github, 2021.4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 162, + 286, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 162, + 286, + 216 + ], + "spans": [ + { + "bbox": [ + 53, + 162, + 286, + 216 + ], + "type": "text", + "content": "[4] Kara-Ali Aliev, Artem Sevastopolsky, Maria Kolos, Dmitry Ulyanov, and Victor Lempitsky. Neural point-based graphics. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXII 16, pages 696-712. Springer, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 218, + 286, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 218, + 286, + 251 + ], + "spans": [ + { + "bbox": [ + 53, + 218, + 286, + 251 + ], + "type": "text", + "content": "[5] Paul J Besl and Neil D McKay. Method for registration of 3-d shapes. In Sensor fusion IV: control paradigms and data structures, volume 1611, pages 586-606. Spie, 1992. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 253, + 286, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 253, + 286, + 317 + ], + "spans": [ + { + "bbox": [ + 53, + 253, + 286, + 317 + ], + "type": "text", + "content": "[6] Bharat Lal Bhatnagar, Xianghui Xie, Ilya A Petrov, Cristian Sminchisescu, Christian Theobalt, and Gerard Pons-Moll. Behave: Dataset and method for tracking human object interactions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15935-15946, 2022. 2, 3, 4, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 319, + 286, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 319, + 286, + 373 + ], + "spans": [ + { + "bbox": [ + 53, + 319, + 286, + 373 + ], + "type": "text", + "content": "[7] Hongrui Cai, Wanquan Feng, Xuetao Feng, Yan Wang, and Juyong Zhang. Neural surface reconstruction of dynamic scenes with monocular rgb-d camera. In Thirty-sixth Conference on Neural Information Processing Systems (NeurIPS), 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 376, + 286, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 376, + 286, + 420 + ], + "spans": [ + { + "bbox": [ + 53, + 376, + 286, + 420 + ], + "type": "text", + "content": "[8] Zhe Cao, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Realtime multi-person 2d pose estimation using part affinity fields. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7291-7299, 2017. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 422, + 286, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 422, + 286, + 453 + ], + "spans": [ + { + "bbox": [ + 53, + 422, + 286, + 453 + ], + "type": "text", + "content": "[9] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. arXiv preprint arXiv:2203.09517, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 456, + 286, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 286, + 510 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 286, + 510 + ], + "type": "text", + "content": "[10] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14124-14133, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 512, + 286, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 512, + 286, + 565 + ], + "spans": [ + { + "bbox": [ + 48, + 512, + 286, + 565 + ], + "type": "text", + "content": "[11] Alvaro Collet, Ming Chuang, Pat Sweeney, Don Gillett, Dennis Evseev, David Calabrese, Hugues Hoppe, Adam Kirk, and Steve Sullivan. High-quality streamable free-viewpoint video. ACM Transactions on Graphics (TOG), 34(4):69, 2015. 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 568, + 286, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 286, + 621 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 286, + 621 + ], + "type": "text", + "content": "[12] Mingsong Dou, Philip Davidson, Sean Ryan Fanello, Sameh Khamis, Adarsh Kowdle, Christoph Rhemann, Vladimir Tankovich, and Shahram Izadi. Motion2fusion: Real-time volumetric performance capture. ACM Trans. Graph., 36(6):246:1-246:16, Nov. 2017. 2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 624, + 286, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 286, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 286, + 689 + ], + "type": "text", + "content": "[13] Kaiwen Guo, Peter Lincoln, Philip Davidson, Jay Busch, Xueming Yu, Matt Whalen, Geoff Harvey, Sergio Orts-Escolano, Rohit Pandey, Jason Dourgarian, et al. The relightables: Volumetric performance capture of humans with realistic relighting. ACM Transactions on Graphics (ToG), 38(6):1–19, 2019. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 692, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 692, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 692, + 286, + 712 + ], + "type": "text", + "content": "[14] Marc Habermann, Lingjie Liu, Weipeng Xu, Michael Zollhoefer, Gerard Pons-Moll, and Christian Theobalt. Real-time" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 74, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 326, + 74, + 545, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 74, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 326, + 74, + 545, + 94 + ], + "type": "text", + "content": "deep dynamic characters. ACM Transactions on Graphics (TOG), 40(4):1-16, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 96, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 139 + ], + "type": "text", + "content": "[15] Nils Hasler, Carsten Stoll, Martin Sunkel, Bodo Rosenhahn, and H-P Seidel. A statistical model of human pose and body shape. In Computer graphics forum, volume 28, pages 337-346. Wiley Online Library, 2009. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 140, + 545, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 140, + 545, + 193 + ], + "spans": [ + { + "bbox": [ + 308, + 140, + 545, + 193 + ], + "type": "text", + "content": "[16] Mohamed Hassan, Vasileios Choutas, Dimitrios Tzionas, and Michael J Black. Resolving 3d human pose ambiguities with 3d scene constraints. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2282-2292, 2019. 3, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 195, + 545, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 195, + 545, + 247 + ], + "spans": [ + { + "bbox": [ + 308, + 195, + 545, + 247 + ], + "type": "text", + "content": "[17] Tong He, Yuanlu Xu, Shunsuke Saito, Stefano Soatto, and Tony Tung. Arch++: Animation-ready clothed human reconstruction revisited. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11046-11056, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 251, + 545, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 251, + 545, + 303 + ], + "spans": [ + { + "bbox": [ + 308, + 251, + 545, + 303 + ], + "type": "text", + "content": "[18] Yinghao Huang, Omid Taheri, Michael J Black, and Dimitrios Tzionas. Intercap: Joint markerless 3d tracking of humans and objects in interaction. In DAGM German Conference on Pattern Recognition, pages 281-299. Springer, 2022. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 305, + 545, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 305, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 308, + 305, + 545, + 358 + ], + "type": "text", + "content": "[19] Zeng Huang, Yuanlu Xu, Christoph Lassner, Hao Li, and Tony Tung. Arch: Animatable reconstruction of clothed humans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3093-3102, 2020. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 361, + 545, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 361, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 308, + 361, + 545, + 415 + ], + "type": "text", + "content": "[20] Catalin Ionescu, Dragos Papava, Vlad Olaru, and Cristian Sminchisescu. Human3.6m: Large scale datasets and predictive methods for 3d human sensing in natural environments. IEEE Transactions on Pattern Analysis and Machine Intelligence, 36(7):1325-1339, jul 2014. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 416, + 545, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 416, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 308, + 416, + 545, + 480 + ], + "type": "text", + "content": "[21] Yuheng Jiang, Suyi Jiang, Guoxing Sun, Zhuo Su, Kaiwen Guo, Minye Wu, Jingyi Yu, and Lan Xu. Neuralhofusion: Neural volumetric rendering under human-object interactions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6155-6165, 2022. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 482, + 545, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 482, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 308, + 482, + 545, + 536 + ], + "type": "text", + "content": "[22] Hanbyul Joo, Tomas Simon, Xulong Li, Hao Liu, Lei Tan, Lin Gui, Sean Banerjee, Timothy Scott Godisart, Bart Nabbe, Iain Matthews, Takeo Kanade, Shohei Nobuhara, and Yaser Sheikh. Panoptic studio: A massively multiview system for social interaction capture. 2017. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 537, + 545, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 537, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 308, + 537, + 545, + 590 + ], + "type": "text", + "content": "[23] Youngjoong Kwon, Dahun Kim, Duygu Ceylan, and Henry Fuchs. Neural human performer: Learning generalizable radiance fields for human performance rendering. Advances in Neural Information Processing Systems, 34:24741-24752, 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 593, + 545, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 593, + 545, + 646 + ], + "spans": [ + { + "bbox": [ + 308, + 593, + 545, + 646 + ], + "type": "text", + "content": "[24] Christoph Lassner, Javier Romero, Martin Kiefel, Federica Bogo, Michael J. Black, and Peter V. Gehler. Unite the people: Closing the loop between 3d and 2d human representations. In IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), July 2017. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "type": "text", + "content": "[25] Ruilong Li, Julian Tanke, Minh Vo, Michael Zollhofer, Jurgen Gall, Angjoo Kanazawa, and Christoph Lassner. Tava: Template-free animatable volumetric actors. arXiv preprint arXiv:2206.08929, 2022. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 692, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 692, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 692, + 545, + 712 + ], + "type": "text", + "content": "[26] Jun Liu, Amir Shahroudy, Mauricio Perez, Gang Wang, Ling-Yu Duan, and Alex C Kot. Nturgb+ d 120: A large-" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8842" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "text", + "content": "scale benchmark for 3d human activity understanding. IEEE transactions on pattern analysis and machine intelligence, 42(10):2684-2701, 2019. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 151 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 151 + ], + "type": "text", + "content": "[27] Lingjie Liu, Marc Habermann, Viktor Rudnev, Kripasindhu Sarkar, Jiatao Gu, and Christian Theobalt. Neural actor: Neural free-view synthesis of human actors with pose control. ACM Trans. Graph. (ACM SIGGRAPH Asia), 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 288, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 288, + 207 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 288, + 207 + ], + "type": "text", + "content": "[28] Lingjie Liu, Weipeng Xu, Marc Habermann, Michael Zollhöfer, Florian Bernard, Hyeongwoo Kim, Wenping Wang, and Christian Theobalt. Neural human video rendering by learning dynamic textures and rendering-to-video translation. arXiv preprint arXiv:2001.04947, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 209, + 288, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 288, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 288, + 262 + ], + "type": "text", + "content": "[29] Lingjie Liu, Weipeng Xu, Michael Zollhoefer, Hyeongwoo Kim, Florian Bernard, Marc Habermann, Wenping Wang, and Christian Theobalt. Neural rendering and reenactment of human actor videos. ACM Transactions on Graphics (TOG), 38(5):1-14, 2019. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 264, + 288, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 264, + 288, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 264, + 288, + 308 + ], + "type": "text", + "content": "[30] Yuan Liu, Sida Peng, Lingjie Liu, Qianqian Wang, Peng Wang, Christian Theobalt, Xiaowei Zhou, and Wenping Wang. Neural rays for occlusion-aware image-based rendering. In CVPR, 2022. 2, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 309, + 288, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 309, + 288, + 353 + ], + "spans": [ + { + "bbox": [ + 48, + 309, + 288, + 353 + ], + "type": "text", + "content": "[31] Stephen Lombardi, Tomas Simon, Jason Saragih, Gabriel Schwartz, Andreas Lehrmann, and Yaser Sheikh. Neural volumes: Learning dynamic renderable volumes from images. arXiv preprint arXiv:1906.07751, 2019. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 354, + 287, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 354, + 287, + 398 + ], + "spans": [ + { + "bbox": [ + 48, + 354, + 287, + 398 + ], + "type": "text", + "content": "[32] Stephen Lombardi, Tomas Simon, Gabriel Schwartz, Michael Zollhoefer, Yaser Sheikh, and Jason Saragih. Mixture of volumetric primitives for efficient neural rendering. arXiv preprint arXiv:2103.01954, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 399, + 288, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 288, + 432 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 288, + 432 + ], + "type": "text", + "content": "[33] Xiaoxiao Long, Cheng Lin, Peng Wang, Taku Komura, and Wenping Wang. Sparseneus: Fast generalizable neural surface reconstruction from sparse views. ECCV, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 434, + 287, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 434, + 287, + 488 + ], + "spans": [ + { + "bbox": [ + 48, + 434, + 287, + 488 + ], + "type": "text", + "content": "[34] H. Luo, A. Chen, Q. Zhang, B. Pang, M. Wu, L. Xu, and J. Yu. Convolutional neural opacity radiance fields. In 2021 IEEE International Conference on Computational Photography (ICCP), pages 1-12, Los Alamitos, CA, USA, may 2021. IEEE Computer Society. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 490, + 287, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 490, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 490, + 287, + 533 + ], + "type": "text", + "content": "[35] Haimin Luo, Teng Xu, Yuheng Jiang, Chenglin Zhou, Qiwei Qiu, Yingliang Zhang, Wei Yang, Lan Xu, and Jingyi Yu. Artemis: Articulated neural pets with appearance and motion synthesis. ACM Trans. Graph., 41(4), jul 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 535, + 287, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 535, + 287, + 589 + ], + "spans": [ + { + "bbox": [ + 48, + 535, + 287, + 589 + ], + "type": "text", + "content": "[36] Quan Meng, Anpei Chen, Haimin Luo, Minye Wu, Hao Su, Lan Xu, Xuming He, and Jingyi Yu. Gnerf: Gan-based neural radiance field without posed camera. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6351-6361, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 590, + 287, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 645 + ], + "type": "text", + "content": "[37] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European conference on computer vision, pages 405-421. Springer, 2020. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 647, + 287, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 690 + ], + "type": "text", + "content": "[38] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, July 2022. 2, 4, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "type": "text", + "content": "[39] Michael Niemeyer, Jonathan T. Barron, Ben Mildenhall, Mehdi S. M. Sajjadi, Andreas Geiger, and Noha Radwan." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 326, + 73, + 546, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 546, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 546, + 106 + ], + "type": "text", + "content": "Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 107, + 546, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 546, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 546, + 150 + ], + "type": "text", + "content": "[40] Atsuhiro Noguchi, Xiao Sun, Stephen Lin, and Tatsuya Harada. Neural articulated radiance field. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5762-5772, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 152, + 546, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 546, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 546, + 183 + ], + "type": "text", + "content": "[41] Anqi Pang, Xin Chen, Haimin Luo, Minye Wu, Jingyi Yu, and Lan Xu. Few-shot neural human performance rendering from sparse rgbd videos, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 185, + 546, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 185, + 546, + 238 + ], + "spans": [ + { + "bbox": [ + 307, + 185, + 546, + 238 + ], + "type": "text", + "content": "[42] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M. Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. ACM Trans. Graph., 40(6), dec 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 239, + 547, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 239, + 547, + 304 + ], + "spans": [ + { + "bbox": [ + 307, + 239, + 547, + 304 + ], + "type": "text", + "content": "[43] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3d hands, face, and body from a single image. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10975-10985, 2019. 2, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 306, + 546, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 306, + 546, + 360 + ], + "spans": [ + { + "bbox": [ + 307, + 306, + 546, + 360 + ], + "type": "text", + "content": "[44] Sida Peng, Junting Dong, Qianqian Wang, Shangzhan Zhang, Qing Shuai, Xiaowei Zhou, and Hujun Bao. Animatable neural radiance fields for modeling dynamic human bodies. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14314-14323, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 361, + 546, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 361, + 546, + 414 + ], + "spans": [ + { + "bbox": [ + 307, + 361, + 546, + 414 + ], + "type": "text", + "content": "[45] Sida Peng, Yuanqing Zhang, Yinghao Xu, Qianqian Wang, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In CVPR, 2021. 2, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 416, + 546, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 416, + 546, + 459 + ], + "spans": [ + { + "bbox": [ + 307, + 416, + 546, + 459 + ], + "type": "text", + "content": "[46] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, jun 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 460, + 546, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 460, + 546, + 514 + ], + "spans": [ + { + "bbox": [ + 307, + 460, + 546, + 514 + ], + "type": "text", + "content": "[47] Yu Rong, Takaaki Shiratori, and Hanbyul Joo. Frankmocap: A monocular 3d whole-body pose estimation system via regression and integration. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1749-1759, 2021. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 515, + 546, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 515, + 546, + 580 + ], + "spans": [ + { + "bbox": [ + 307, + 515, + 546, + 580 + ], + "type": "text", + "content": "[48] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2304-2314, 2019. 2, 7, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 582, + 547, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 582, + 547, + 635 + ], + "spans": [ + { + "bbox": [ + 307, + 582, + 547, + 635 + ], + "type": "text", + "content": "[49] Shunsuke Saito, Tomas Simon, Jason Saragih, and Hanbyul Joo. Pifuhd: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 84-93, 2020. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 637, + 546, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 637, + 546, + 680 + ], + "spans": [ + { + "bbox": [ + 307, + 637, + 546, + 680 + ], + "type": "text", + "content": "[50] Manolis Savva, Angel X Chang, Pat Hanrahan, Matthew Fisher, and Matthias Nießner. Pigraphs: learning interaction snapshots from observations. ACM Transactions on Graphics (TOG), 35(4):1-12, 2016. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 681, + 546, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 681, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 681, + 546, + 713 + ], + "type": "text", + "content": "[51] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-Motion Revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2, 3" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8843" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[52] Qing Shuai, Chen Geng, Qi Fang, Sida Peng, Wenhao Shen, Xiaowei Zhou, and Hujun Bao. Novel view synthesis of human interactions from sparse multi-view videos. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-10, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 288, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 288, + 196 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 288, + 196 + ], + "type": "text", + "content": "[53] Qing Shuai, Chen Geng, Qi Fang, Sida Peng, Wenhao Shen, Xiaowei Zhou, and Hujun Bao. Novel view synthesis of human interactions from sparse multi-view videos. In ACM SIGGRAPH 2022 Conference Proceedings, SIGGRAPH '22, New York, NY, USA, 2022. Association for Computing Machinery. 3, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 197, + 288, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 197, + 288, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 197, + 288, + 262 + ], + "type": "text", + "content": "[54] Aliaksandra Shysheya, Egor Zakharov, Kara-Ali Aliev, Renat Bashirov, Egor Burkov, Karim Iskakov, Aleksei Ivakhnenko, Yury Malkov, Igor Pasechnik, Dmitry Ulyanov, et al. Textured neural avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2387-2397, 2019. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 264, + 288, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 264, + 288, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 264, + 288, + 319 + ], + "type": "text", + "content": "[55] Leonid Sigal, Alexandru O Balan, and Michael J Black. Humaneva: Synchronized video and motion capture dataset and baseline algorithm for evaluation of articulated human motion. International journal of computer vision, 87(1):4-27, 2010. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 319, + 288, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 319, + 288, + 387 + ], + "spans": [ + { + "bbox": [ + 48, + 319, + 288, + 387 + ], + "type": "text", + "content": "[56] Zhuo Su, Lan Xu, Zerong Zheng, Tao Yu, Yebin Liu, and Lu Fang. Robustfusion: Human volumetric capture with data-driven visual cues using a rgbd camera. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, Computer Vision – ECCV 2020, pages 246–264, Cham, 2020. Springer International Publishing. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 388, + 288, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 288, + 442 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 288, + 442 + ], + "type": "text", + "content": "[57] Guoxing Sun, Xin Chen, Yizhang Chen, Anqi Pang, Pei Lin, Yuheng Jiang, Lan Xu, Jingya Wang, and Jingyi Yu. Neural free-viewpoint performance rendering under complex human-object interactions. In Proceedings of the 29th ACM International Conference on Multimedia, 2021. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 444, + 288, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 444, + 288, + 509 + ], + "spans": [ + { + "bbox": [ + 48, + 444, + 288, + 509 + ], + "type": "text", + "content": "[58] Guoxing Sun, Xin Chen, Yizhang Chen, Anqi Pang, Pei Lin, Yuheng Jiang, Lan Xu, Jingyi Yu, and Jingya Wang. Neural free-viewpoint performance rendering under complex human-object interactions. In Proceedings of the 29th ACM International Conference on Multimedia, pages 4651-4660, 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 510, + 288, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 510, + 288, + 577 + ], + "spans": [ + { + "bbox": [ + 48, + 510, + 288, + 577 + ], + "type": "text", + "content": "[59] Xin Suo, Yuheng Jiang, Pei Lin, Yingliang Zhang, Minye Wu, Kaiwen Guo, and Lan Xu. Neuralhumanfvv: Realtime neural volumetric human performance rendering using rgb cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6226-6237, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 578, + 288, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 288, + 634 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 288, + 634 + ], + "type": "text", + "content": "[60] Xin Suo, Yuheng Jiang, Pei Lin, Yingliang Zhang, Minye Wu, Kaiwen Guo, and Lan Xu. Neuralhumanfvv: Real-time neural volumetric human performance rendering using rgb cameras. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6222-6233, 2021. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 635, + 288, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 288, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 288, + 679 + ], + "type": "text", + "content": "[61] Omid Taheri, Nima Ghorbani, Michael J Black, and Dimitrios Tzionas. Grab: A dataset of whole-body human grasping of objects. In European conference on computer vision, pages 581-600. Springer, 2020. 2, 3, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 680, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 288, + 714 + ], + "type": "text", + "content": "[62] Ayush Tewari, Justus Thies, Ben Mildenhall, Pratul Srinivasan, Edgar Tretschk, Yifan Wang, Christoph Lassner, Vincent Sitzmann, Ricardo Martin-Brualla, Stephen Lom-" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 117 + ], + "type": "text", + "content": "bardi, Tomas Simon, Christian Theobalt, Matthias Niessner, Jonathan T. Barron, Gordon Wetzstein, Michael Zollhoefer, and Vladislav Golyanik. Advances in neural rendering, 2021. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 307, + 118, + 547, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 547, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 547, + 184 + ], + "type": "text", + "content": "[63] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhofer, Christoph Lassner, and Christian Theobalt. Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12959-12970, 2021. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 185, + 547, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 185, + 547, + 240 + ], + "spans": [ + { + "bbox": [ + 307, + 185, + 547, + 240 + ], + "type": "text", + "content": "[64] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T Barron, and Pratul P Srinivasan. Ref-nerf: Structured view-dependent appearance for neural radiance fields. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5481-5490. IEEE, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 241, + 547, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 241, + 547, + 297 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 547, + 297 + ], + "type": "text", + "content": "[65] Timo Von Marcard, Roberto Henschel, Michael J Black, Bodo Rosenhahn, and Gerard Pons-Moll. Recovering accurate 3d human pose in the wild using imus and a moving camera. In Proceedings of the European Conference on Computer Vision (ECCV), pages 601-617, 2018. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 297, + 547, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 297, + 547, + 352 + ], + "spans": [ + { + "bbox": [ + 307, + 297, + 547, + 352 + ], + "type": "text", + "content": "[66] Liao Wang, Ziyu Wang, Pei Lin, Yuheng Jiang, Xin Suo, Minye Wu, Lan Xu, and Jingyi Yu. ibutter: Neural interactive bullet time generator for human free-viewpoint rendering. In Proceedings of the 29th ACM International Conference on Multimedia, pages 4641-4650, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 354, + 547, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 354, + 547, + 419 + ], + "spans": [ + { + "bbox": [ + 307, + 354, + 547, + 419 + ], + "type": "text", + "content": "[67] Liao Wang, Jiakai Zhang, Xinhang Liu, Fuqiang Zhao, Yanshun Zhang, Yingliang Zhang, Minye Wu, Jingyi Yu, and Lan Xu. Fourier plenoctrees for dynamic radiance field rendering in real-time. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13524-13534, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 420, + 547, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 420, + 547, + 464 + ], + "spans": [ + { + "bbox": [ + 307, + 420, + 547, + 464 + ], + "type": "text", + "content": "[68] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. NeurIPS, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 465, + 547, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 465, + 547, + 520 + ], + "spans": [ + { + "bbox": [ + 307, + 465, + 547, + 520 + ], + "type": "text", + "content": "[69] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul Srinivasan, Howard Zhou, Jonathan T. Barron, Ricardo MartinBrualla, Noah Snively, and Thomas Funkhouser. Ibrnet: Learning multi-view image-based rendering. In CVPR, 2021. 2,8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 521, + 547, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 521, + 547, + 555 + ], + "spans": [ + { + "bbox": [ + 307, + 521, + 547, + 555 + ], + "type": "text", + "content": "[70] Shaofei Wang, Katja Schwarz, Andreas Geiger, and Siyu Tang. Arah: animatable volume rendering of articulated human sdfs. arXiv preprint arXiv:2210.10036, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 555, + 547, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 555, + 547, + 600 + ], + "spans": [ + { + "bbox": [ + 307, + 555, + 547, + 600 + ], + "type": "text", + "content": "[71] Minye Wu, Yuehao Wang, Qiang Hu, and Jingyi Yu. Multi-view neural human rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 601, + 547, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 601, + 547, + 645 + ], + "spans": [ + { + "bbox": [ + 307, + 601, + 547, + 645 + ], + "type": "text", + "content": "[72] Minye Wu, Yuehao Wang, Qiang Hu, and Jingyi Yu. Multi-view neural human rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1682-1691, 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 646, + 547, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 646, + 547, + 679 + ], + "spans": [ + { + "bbox": [ + 307, + 646, + 547, + 679 + ], + "type": "text", + "content": "[73] Xianghui Xie, Bharat Lal Bhatnagar, and Gerard Pons-Moll. Chore: Contact, human and object reconstruction from a single rgb image. arXiv preprint arXiv:2204.02445, 2022. 3, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 680, + 547, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 547, + 714 + ], + "type": "text", + "content": "[74] Hongwei Yi, Chun-Hao P Huang, Dimitrios Tzionas, Muhammed Kocabas, Mohamed Hassan, Siyu Tang, Justus Thies, and Michael J Black. Human-aware object placement" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "8844" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 441 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 66, + 73, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 73, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 73, + 287, + 106 + ], + "type": "text", + "content": "for visual environment reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3959-3970, 2022. 3, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 139 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 139 + ], + "type": "text", + "content": "[75] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. Plenoctrees for real-time rendering of neural radiance fields. In arXiv, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 287, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 287, + 184 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 287, + 184 + ], + "type": "text", + "content": "[76] Chao Zhang, Sergi Pujades, Michael J. Black, and Gerard Pons-Moll. Detailed, accurate, human shape estimation from clothed 3d scan sequences. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), July 2017. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 186, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 287, + 239 + ], + "type": "text", + "content": "[77] Jiakai Zhang, Xinhang Liu, Xinyi Ye, Fuqiang Zhao, Yanshun Zhang, Minye Wu, Yingliang Zhang, Lan Xu, and Jingyi Yu. Editable free-viewpoint video using a layered neural representation. ACM Trans. Graph., 40(4), July 2021. 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 241, + 287, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 241, + 287, + 296 + ], + "spans": [ + { + "bbox": [ + 48, + 241, + 287, + 296 + ], + "type": "text", + "content": "[78] Jason Y Zhang, Sam Pepose, Hanbyul Joo, Deva Ramanan, Jitendra Malik, and Angjoo Kanazawa. Perceiving 3d human-object spatial arrangements from a single image in the wild. In European conference on computer vision, pages 34-51. Springer, 2020. 3, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 297, + 287, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 297, + 287, + 351 + ], + "spans": [ + { + "bbox": [ + 48, + 297, + 287, + 351 + ], + "type": "text", + "content": "[79] Fuqiang Zhao, Yuheng Jiang, Kaixin Yao, Jiakai Zhang, Liao Wang, Haizhao Dai, Yuhui Zhong, Yingliang Zhang, Minye Wu, Lan Xu, et al. Human performance modeling and rendering via neural animated mesh. arXiv preprint arXiv:2209.08468, 2022. 2, 4, 6, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 353, + 287, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 353, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 48, + 353, + 287, + 407 + ], + "type": "text", + "content": "[80] Fuqiang Zhao, Wei Yang, Jiakai Zhang, Pei Lin, Yingliang Zhang, Jingyi Yu, and Lan Xu. Humannerf: Efficiently generated human radiance field from sparse inputs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7743-7753, 2022. 2, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 409, + 287, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 409, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 48, + 409, + 287, + 441 + ], + "type": "text", + "content": "[81] Qian-Yi Zhou, Jaesik Park, and Vladlen Koltun. Open3d: A modern library for 3d data processing. arXiv preprint arXiv:1801.09847, 2018. 4" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8845" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/NeuralEditor_ Editing Neural Radiance Fields via Manipulating Point Clouds/70b7a137-ced8-4460-839e-a2dd484a031c_content_list.json b/2023/NeuralEditor_ Editing Neural Radiance Fields via Manipulating Point Clouds/70b7a137-ced8-4460-839e-a2dd484a031c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..402649b821a54e1244458adf0fd511042843854d --- /dev/null +++ b/2023/NeuralEditor_ Editing Neural Radiance Fields via Manipulating Point Clouds/70b7a137-ced8-4460-839e-a2dd484a031c_content_list.json @@ -0,0 +1,1372 @@ +[ + { + "type": "text", + "text": "NeuralEditor: Editing Neural Radiance Fields via Manipulating Point Clouds", + "text_level": 1, + "bbox": [ + 93, + 130, + 875, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jun-Kun Chen $^{1\\dagger}$", + "bbox": [ + 251, + 179, + 385, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jipeng Lyu $^{2\\dagger}$", + "bbox": [ + 426, + 180, + 529, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yu-Xiong Wang", + "bbox": [ + 571, + 180, + 709, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1University of Illinois at Urbana-Champaign", + "bbox": [ + 109, + 199, + 462, + 218 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Peking University $\\dagger$ Equal Contribution", + "bbox": [ + 501, + 200, + 854, + 218 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{junkun3，yxw}@illinois.edu lvjipeng@pku.edu.cn", + "bbox": [ + 194, + 220, + 769, + 237 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4589dde9f2e2500c54fe59004e1e22d8c9108ed27ca90fd9432a6d04afdc8926.jpg", + "image_caption": [ + "Figure 1. Our NeuralEditor offers native support for general and flexible shape editing of neural radiance fields via manipulating point clouds. By generating a precise point cloud of the scene with a novel point cloud-guided NeRF model, our NeuralEditor produces high-fidelity rendering results in both shape deformation and more challenging scene morphing tasks." + ], + "image_footnote": [], + "bbox": [ + 80, + 263, + 300, + 420 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1b9a1be3d3a7295491938e1aeb98a2f36f0f998e2d7cf65aba738962a02a830c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 307, + 263, + 890, + 425 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 488, + 313, + 503 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This paper proposes NeuralEditor that enables neural radiance fields (NeRFs) natively editable for general shape editing tasks. Despite their impressive results on novel-view synthesis, it remains a fundamental challenge for NeRFs to edit the shape of the scene. Our key insight is to exploit the explicit point cloud representation as the underlying structure to construct NeRFs, inspired by the intuitive interpretation of NeRF rendering as a process that projects or \"plots\" the associated 3D point cloud to a 2D image plane. To this end, NeuralEditor introduces a novel rendering scheme based on deterministic integration within K-D tree-guided density-adaptive voxels, which produces both high-quality rendering results and precise point clouds through optimization. NeuralEditor then performs shape editing via mapping associated points between point clouds. Extensive evaluation shows that NeuralEditor achieves state-of-the-art performance in both shape deformation and scene morphing tasks. Notably, NeuralEditor supports both zero-shot inference and further fine-tuning over the edited scene. Our code, benchmark, and demo video are available at immortalco.github.io/NeuralEditor.", + "bbox": [ + 75, + 508, + 473, + 828 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 845, + 207, + 859 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Perhaps the most memorable shot of the film Transformers, Optimus Prime is seamlessly transformed between a", + "bbox": [ + 75, + 869, + 468, + 900 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "humanoid and a Peterbilt truck – such free-form editing of 3D objects and scenes is a fundamental task in 3D computer vision and computer graphics, directly impacting applications such as visual simulation, movie, and game industries. In these applications, often we are required to manipulate a scene or objects in the scene by editing or modifying its shape, color, light condition, etc., and generate visually-faithful rendering results on the edited scene efficiently. Among the various editing operations, shape editing has received continued attention but remains challenging, where the scene is deformed in a human-guided way, while all of its visual attributes (e.g., shape, color, brightness, and light condition) are supposed to be natural and consistent with the ambient environment.", + "bbox": [ + 496, + 489, + 892, + 699 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "State-of-the-art rendering models are based on implicit neural representations, as exemplified by neural radiance field (NeRF) [27] and its variants [3, 33, 37, 39, 48]. Despite their impressive novel-view synthesis results, most of the NeRF models substantially lack the ability for users to adjust, edit, or modify the shape of scene objects. On the other hand, shape editing operations can be natively applied to explicit 3D representations such as point clouds and meshes.", + "bbox": [ + 496, + 702, + 890, + 823 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Inspired by this, we propose NeuralEditor – a general and flexible approach to editing neural radiance fields via manipulating point clouds (Fig. 1). Our key insight is to benefit from the best of both worlds: the superiority in rendering performance from implicit neural representation", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "12439", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "combined with the ease of editing from explicit point cloud representation. NeuralEditor enables us to perform a wide spectrum of shape editing operations in a consistent way.", + "bbox": [ + 78, + 90, + 467, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Such introduction of point clouds into NeRF for general shape editing is rooted in our interpretation of NeRF rendering as a process that projects or \"plots\" the associated 3D point cloud to a 2D image plane. Conceptually, with a dense enough point cloud where each point has an opacity and its color is defined as a function of viewing direction, directly plotting the point cloud would achieve similar visual effects (i.e., transparency and view-dependent colors) that are rendered by NeRF. This intrinsic integration between NeRF and point clouds underscores the advantage of our NeuralEditor over existing mesh-based NeRF editing methods such as NeRF-Editing [51], Deforming-NeRF [44], and CageNeRF [30], where the process of constructing and optimizing the mesh is separated from the NeRF modeling, making them time-consuming. More importantly, with the point cloud constructed for a scene, the shape editing can be natively defined as and easily solved by just moving each point into the new, edited position and re-plotting the point cloud. Therefore, our approach supports more general scene editing operations which are difficult to achieve via mesh-guided space deformation.", + "bbox": [ + 80, + 138, + 467, + 454 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The key component in our NeuralEditor lies in a point cloud-guided NeRF model that natively supports general shape editing operations. While the recent method Point-NeRF [43] has demonstrated improved novel-view synthesis capability based on point clouds, it is not supportive to shape editing. Our idea then is to exploit the underlying point cloud in ways of not only optimizing its structure and features (e.g., adaptive voxels) for rendering, but also extracting additional useful attributes (e.g., normal vectors) to guide the editing process. To this end, we introduce K-D trees [4] to construct density-adaptive voxels for efficient and stable rendering, together with a novel deterministic integration strategy. Moreover, we model the color with the Phong reflection [31] to decompose the specular color and better represent the scene geometry.", + "bbox": [ + 80, + 458, + 467, + 684 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "With a much more precise point cloud attributed to these improvements, our NeuralEditor achieves high-fidelity rendering results on deformed scenes compared with prior work as shown in Fig. 1, even in a zero-shot inference manner without additional training. Through fast fine-tuning, the visual quality of the deformed scene is further enhanced, almost perfectly consistent with the surrounding light condition. In addition, under the guidance of a point cloud diffusion model [24], NeuralEditor can be naturally extended for smooth scene morphing across multiple scenes, which is difficult for existing NeRF editing work.", + "bbox": [ + 80, + 686, + 467, + 852 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are four-fold. (1) We introduce NeuralEditor, a flexible and versatile approach that makes neural radiance fields editable through manipulating point", + "bbox": [ + 78, + 854, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "clouds. (2) We propose a point cloud-guided NeRF model based on K-D trees and deterministic integration, which produces precise point clouds and supports general scene editing. (3) Due to the lack of publicly available benchmarks for shape editing, we construct and release a reproducible benchmark that promotes future research on shape editing. (4) We investigate a wide range of shape editing tasks, covering both shape deformation (as studied in existing NeRF editing work) and challenging scene morphing (a novel task addressed here). NeuralEditor achieves state-of-the-art performance on all shape editing tasks in a unified framework, without extra information or supervision.", + "bbox": [ + 503, + 90, + 890, + 271 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 503, + 286, + 638, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Neural Scene Representation. Traditional methods model scenes with explicit [2, 11, 15, 20, 34, 38] or implicit [6, 19, 26, 28, 40] 3D geometric or shape representations. Initiated by NeRF [27], leveraging implicit neural networks to represent scenes and perform novel-view synthesis has become a fast-developing field in 3D vision [8,9]. While most of the follow-up work focuses on improving aspects such as the rendering realism [3, 10, 21, 37], efficiency [33, 41, 48], and cross-scene generalization [5, 39, 43, 49], the scene editing capability is substantially missing in the NeRF family which we address in this paper. In addition, we exploit K-D tree-guided point clouds as the underlying structure, different from other NeRF variants based on octrees [21, 33, 48] or plain voxels [41].", + "bbox": [ + 503, + 311, + 890, + 521 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Point-Based NeRFs. Recently, using point clouds to build a NeRF model has shown better encoding of scene shape and improved rendering performance, as represented by PointNeRF [43]. PointNeRF proposes a point initialization network to produce the initial point cloud together with the point features, which is further optimized by a pruning and growing strategy. While both PointNeRF and our NeuralEditor employ point clouds as the underlying structure, NeuralEditor better exploits useful information within the point clouds: PointNeRF only directly uses the locations of points; by contrast, NeuralEditor considers the point cloud more as a geometrical shape and extracts relevant information like normal vectors, which plays an important role in rendering and shape editing. Importantly, our approach is designed to support scene editing, in contrast to PointNeRF.", + "bbox": [ + 503, + 523, + 890, + 748 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Scene Editing via NeRFs. Different types of scene editing have been studied under NeRFs. EditNeRF [22], ObjectNeRF [45], and DistillNeRF [18] perform simple shape and color editing for objects specified with human-input scribble, pixel, segment, language, etc. NeuPhysics [32] edits a dynamic scene via physics parameters. CCNeRF [35] proposes an explicit NeRF representation with tensor rank decomposition to support scene composition. INSP-Net [42] considers filter editing like denoising. Such work cannot address 3D shape editing and only sup", + "bbox": [ + 503, + 750, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "12440", + "bbox": [ + 480, + 946, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ports simple editing operations, like object selection, similarity transformation, or limited shape deformation.", + "bbox": [ + 75, + 90, + 468, + 119 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D Shape Editing. Traditional representation methods support keypoint-based shape editing [13, 14, 16, 25, 36, 47, 52] with meshes [29, 50], which cannot be directly applied to implicit representations used by NeRF. Existing NeRF editing work primarily studies a particular shape editing task, mesh deformation, and addresses it in a common paradigm [30, 44, 51]: A mesh of the scene is first constructed by either exporting it from a trained NeRF with the Marching Cubes algorithm [23], or optimizing close-to-surface cages along with training. After the user deforms the mesh, the deformed scene is rendered by deforming the space and bending the viewing rays in the original scene with the trained NeRF. Doing so requires extra efforts to convert implicit scene representation to explicit mesh, which might not be precise enough, and only supports continuous shape editing that can be converted to space deformation. On the contrary, our NeuralEditor directly maintains and utilizes alternative explicit scene representation – the point cloud which is intrinsically integrated with NeRF, making NeuralEditor require no extra efforts and support more general shape editing tasks like scene morphing. NeuralEditor supports both zero-shot inference and further fine-tuning over the edited scene, while prior work cannot.", + "bbox": [ + 75, + 121, + 472, + 470 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. NeuralEditor: Point Cloud-Guided NeRF", + "text_level": 1, + "bbox": [ + 76, + 481, + 449, + 498 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We propose a novel point cloud-guided NeRF model, NeuralEditor - it not only achieves realistic rendering results in the novel-view synthesis task, but also produces a point cloud that precisely describes the shape of the scene, thus facilitating general shape editing tasks. As illustrated in Fig. 2, we leverage the K-D trees [4] to construct density-adaptive voxels (which also naturally enable us to skip empty spaces), and introduce deterministic spline integration for rendering. We use the Phong reflection to model the color along with the normal vectors obtained from the underlying point cloud. With our enhanced point cloud optimization, NeuralEditor obtains much more precise underlying point clouds, compared with noisy and imprecise outputs of state-of-the-art PointNeRF [43] (as shown in Sec. 5).", + "bbox": [ + 75, + 506, + 472, + 719 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. K-D Tree-Guided Voxels", + "text_level": 1, + "bbox": [ + 76, + 726, + 303, + 742 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To render with points, we construct multi-scale density-adaptive voxels based on K-D trees [4], namely, $K - D$ voxels. K-D trees are a data structure constructed on $K$ -dimensional points, where $K = 3$ for 3D points. As a special decision tree, K-D tree's each node divides the point set into two equal-sized parts with axis-parallel criterion.", + "bbox": [ + 75, + 750, + 468, + 840 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For each K-D tree's node, we compute its bounding box by taking the minimum and maximum $x, y, z$ coordinates in its subtree and with proper padding margins. As we divide the points in a top-down manner in one of the $x, y, z$ direc", + "bbox": [ + 75, + 840, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/bd5934b42f8753048540e844e4d330c236474cd2f2bb976fc87fea27eb17524a.jpg", + "image_caption": [ + "Figure 2. Our K-D Voxels. Column 1: original point clouds constructed from two scenes, colored with normal vector directions. Column 2: upper-level voxels which coarsely represent the shape. Column 3: lower-level voxels which tightly cover the shape. Right: Visualization of K-D voxels on a 2D point cloud. Each color represents boxes of nodes on each level of the K-D tree. Lower-level boxes containing fewer points cover the shape more tightly, and vice versa for higher-level boxes." + ], + "image_footnote": [], + "bbox": [ + 514, + 89, + 883, + 186 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tions, in each layer of the K-D tree, different nodes' bounding boxes are mutually exclusive. Therefore, the bounding boxes can be regarded as voxels. As boxes in the upper layers contain more points (larger voxels), while those in the lower layers contain fewer points (smaller voxels), we naively obtain a multi-scale voxel construction from one K-D tree. As shown in Fig. 2, voxels from the large to small scales represent the shape of the scene from coarse to fine.", + "bbox": [ + 496, + 305, + 893, + 428 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Rendering Over K-D Voxels", + "text_level": 1, + "bbox": [ + 500, + 436, + 751, + 454 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We now introduce a rendering scheme that exploits K-D voxels to perform all the sub-procedures associated with rendering in a unified way. This scheme enables us to render more naturally, efficiently, and even deterministically, meanwhile it also simplifies some widely-adopted design choices in conventional NeRF rendering.", + "bbox": [ + 496, + 460, + 893, + 551 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Skipping Empty Spaces. In NeRF rendering, we are supposed to focus only on the surface of scene objects. As shown in Fig. 2, all our K-D voxels are produced to stick to the surface of objects, which the point cloud is constructed to describe. Such a property allows us to avoid explicitly \"skipping\" empty spaces, which often requires extra consideration in most NeRF models - only considering the space inside a voxel automatically focuses on the surface; as the depth of the voxel's node goes deeper, it becomes closer to the surface. Moreover, during the construction of the K-D tree, the points at each node are divided within its sub-nodes, and the node's voxel fully covers all its sub-nodes. This further provides us with a native top-down recursive procedure to locate the voxels intersected with the querying ray: We start from the root node, and recurse on the sub-nodes until we (1) reach a pre-set node depth (or equivalently, a pre-set voxel size) and then query within the associated voxel, or (2) stop recursion on non-intersected nodes.", + "bbox": [ + 496, + 551, + 895, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Density-Adaptive Rendering. An important design in NeRFs is the coarse-to-fine strategy for density-adaptive rendering, so that more points are sampled in high-volume density areas. Our K-D voxels natively support such a design without additional bells and whistles. This is because", + "bbox": [ + 496, + 825, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "12441", + "bbox": [ + 480, + 944, + 517, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/8286100af116b003cc588dc64b936796675e5dfe7fe4d68a7a8d758cbc05da6a.jpg", + "image_caption": [ + "Figure 3. Our NeuralEditor architecture. We propose deterministic spline integration for KNN-based point features over each K-D tree-guided density-adaptive voxel, and model the color via Phong reflection with normal vectors estimated from the point cloud's shape." + ], + "image_footnote": [], + "bbox": [ + 200, + 78, + 772, + 236 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "voxels in the same K-D tree layer contain the same number of points. As the point density can be regarded as an approximation of volume density, all such voxels have the same density. Therefore, we directly use K-D voxels to guide the density-adaptive rendering. Specifically, we conduct the rendering process at the voxels of some bottom layers in the K-D tree. For each querying ray, we use the aforementioned recursive procedure to locate the minimal intersected voxels that are deep enough. Here \"minimal\" means that the ray intersects with the node's voxel, but does not intersect with any sub-node's voxel. These intersected voxels divide the querying ray into several segments (Fig. 3). The segments covered by a voxel are close to the surface and used for rendering, while those not covered are in empty spaces.", + "bbox": [ + 75, + 273, + 472, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Deterministic Spline Integration. DIVER [41] shows that deterministic integration outperforms stochastic integration in NeRF rendering. So we perform a deterministic integration to obtain the segment's feature within each voxel. Since we do not necessarily have points at the voxel's vertices, the trilinear interpolation used in DIVER is not feasible here. Instead, we use spline integration. For the $i$ -th intersected voxel in the ray passing order, we uniformly select points in this segment, and integrate their features to obtain the average feature $f_{i}$ of the segment of the $i$ -th voxel:", + "bbox": [ + 75, + 484, + 468, + 636 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {i} = \\frac {1}{r _ {i} - l _ {i}} \\int_ {l _ {i}} ^ {r _ {i}} \\mathbf {f e a t u r e} (o + t \\cdot d) \\mathrm {d} t, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 642, + 468, + 676 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $[l_i, r_i]$ is the intersection interval, and $o$ and $d$ are the source and direction of the querying ray, respectively. This average feature $f_i$ can be interpreted as the feature of a representative point $p_i$ located somewhere in the segment.", + "bbox": [ + 76, + 676, + 468, + 738 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "KNN-Based Feature Aggregation. For each uniformly selected point $q$ in the segment during spline integration, we obtain its feature via weighted interpolation from the features of its $K$ nearest neighbors (KNN) in the point cloud:", + "bbox": [ + 76, + 739, + 468, + 799 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {f e a t u r e} (q) = \\sum_ {p _ {j} \\in \\operatorname {K N N} (q; K)} k _ {j} e _ {j}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 806, + 468, + 844 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{k_{j}\\right\\} = \\operatorname *{SoftMax}_{p_{j}\\in \\mathrm{KNN}(q;K)}\\left(\\log \\gamma_{j} - \\log \\| q - p_{j}\\|_{2}^{2}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 844, + 431, + 868 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where for each point $p_j$ in the point cloud, we parameterize its confidence $\\gamma_j$ and point feature $e_j$ as in PointNeRF.", + "bbox": [ + 76, + 869, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5bf97e2419a70f2570269ef2a1b209b53783419124e6c97d6266575431f3f7fa.jpg", + "image_caption": [ + "Figure 4. In the two scenes of NeRF Synthetic [27], NeuralEditor optimizes the rough initial point cloud to a precise point cloud. The points are colored with their normal vectors." + ], + "image_footnote": [], + "bbox": [ + 522, + 273, + 869, + 452 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Phong Reflection-Based Color Modeling with Point Cloud Normal Vectors. To obtain the volume density $\\sigma_{i}$ and color $c_{i}$ of the representative point $p_i$ in the $i$ -th voxel, we use the Phong reflection model [31]. As we have the underlying point cloud, we use Open3D [53] to estimate the normal vector for each point, and integrate these vectors over the interval to get an average normal vector $n_i$ . Such information better characterizes the shape of point clouds (scenes), which plays an important role in Phong-based rendering and also implicitly facilitates optimizing the point clouds (Sec. 3.3). Consistent with RefNeRF [37], we use multiple multilayer perceptrons (MLPs) to model other attributes, including volume density, tint, roughness, and diffuse and specular color, and use the Phong formula to calculate the final $c_{i}$ with these attributes. Finally, we aggregate $c_{i}$ of all segments on the ray to obtain the final color:", + "bbox": [ + 496, + 503, + 892, + 744 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nc _ {\\text {p i x e l}} = \\sum_ {i \\geq 1} \\tau_ {i} \\cdot \\left(1 - \\exp \\left(- \\left(r _ {i} - l _ {i}\\right) \\sigma_ {i}\\right) \\cdot c _ {i}, \\right.\n$$\n", + "text_format": "latex", + "bbox": [ + 545, + 753, + 841, + 786 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tau_ {i} = \\exp \\left(- \\sum_ {i ^ {\\prime} = 1} ^ {i - 1} \\left(r _ {i ^ {\\prime}} - l _ {i ^ {\\prime}}\\right) \\sigma_ {i ^ {\\prime}}\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 786, + 890, + 832 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Point Cloud Optimization", + "text_level": 1, + "bbox": [ + 500, + 845, + 735, + 863 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Point Cloud Initialization. To start training, we need a coarse initial point cloud. Consistent with PointNeRF, we", + "bbox": [ + 500, + 869, + 892, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "12442", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "use a point generation network, which consists of a multiview stereo (MVS) model [12] based on a 3D convolutional neural network (CNN), to generate the points' coordinates and confidence values, and another 2D CNN [46] to generate their initial features. This network was pre-trained on the DTU training dataset [1], and can generalize to other datasets and scenes. As shown in Fig. 4, the initial point cloud generated by such a network is coarse and noisy.", + "bbox": [ + 75, + 90, + 472, + 210 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Explicit Optimization via Pruning and Growing. We perform a similar pruning and growing procedure as in PointNeRF, to prune outliers with low confidence $\\gamma_{j}$ and fill holes in the point clouds. We make several important modifications over PointNeRF, and integrate this procedure with our deterministic integration (details in the supplementary).", + "bbox": [ + 75, + 212, + 468, + 301 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implicit Optimization with Normal Vectors. In addition to the explicit optimization, the point cloud is also optimized implicitly during training through the adjustment of point confidence $\\gamma_{j}$ . When computing the average normal vector for rendering, we aggregate normal vectors of nearby points weighted with their distance and confidence, where the confidence of noisy or inaccurate points with potentially abnormal normal vectors is adjusted accordingly. Moreover, we apply the normal vector regularization losses from RefNeRF [37] to supervise the points' confidence w.r.t. their normal vectors. These strategies collectively provide implicit but more tailored ways to optimize the point clouds. With both explicit and implicit optimization, NeuralEditor obtains very precise point clouds (Fig. 4).", + "bbox": [ + 75, + 301, + 473, + 515 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Shape Editing with NeuralEditor", + "text_level": 1, + "bbox": [ + 76, + 526, + 379, + 542 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Formulation of General Shape Editing Tasks. We define the shape editing tasks based on indexed point clouds. To this end, we first re-define an indexed point cloud $P$ as a mapping from a point index $j$ to the corresponding point $p_j$ ,", + "bbox": [ + 75, + 551, + 468, + 628 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nP: j \\rightarrow p _ {j}, \\text {w h e r e} p _ {j} \\in \\mathbb {R} ^ {3}, j = 1, \\dots , | P |. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 124, + 632, + 468, + 648 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A shape editing task is defined as another indexed point cloud $Q(P)$", + "bbox": [ + 75, + 657, + 468, + 688 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nQ (P): j \\rightarrow q _ {j}, \\text {w h e r e} q _ {j} \\in \\mathbb {R} ^ {3} \\cup \\{\\varnothing \\}, j = 1, \\dots , | P |, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 693, + 468, + 710 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "describing a shape editing task whether the $j$ -th point moves from $p_j$ to $q_j$ or is deleted in the deformation if $q_j = \\emptyset$ . With this definition, $Q(P)$ can be an arbitrary point cloud with points properly matched to points in $P$ by same indices, regardless of connectivity or continuity.", + "bbox": [ + 75, + 719, + 468, + 795 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our formulation represents a broad range of shape editing tasks. The mesh editing tasks in NeRF-Editing [51], Deforming-NeRF [44], and CageNeRF [30] can be more simply and clearly defined here. For example, in NeRF-Editing, a mesh is exported from a general NeRF model, deformed manually, and converted to an \"offset\" or a continuous space deformation. We can depict such a task without", + "bbox": [ + 75, + 795, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/cfa7eafab6f5def118dad4763b9eb0193fc659617d8d194f6fa1ca2f094ba3f8.jpg", + "image_caption": [ + "Figure 5. Infinitesimal surface transformation (IST). (a) As the view-dependent colors are modeled as absolute viewing directions, they (solid arrows at the right) are different from the correct colors (dashed arrows at the right) after deformation. We solve this by (b) constructing a local coordinate system near the $j$ -th point and (c) modeling IST from the edited scene to the original scene with the coordinate systems, so as to (d) redirect the viewing direction to the original scene when rendering the edited scene. \"offsets,\" by recording only the final location for each point without extra information. Notably, our formulation even models those whose deformation is not continuous in the space, e.g., cutting a scene into two parts, and thus cannot be covered and solved by NeRF-Editing, Deforming-NeRF, or CageNeRF, as shown in the supplementary." + ], + "image_footnote": [], + "bbox": [ + 501, + 89, + 679, + 195 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/fca9b06889fcc3e42dc57debb9f4d2dab78274a0368ef6398e046a5d8115d411.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 681, + 89, + 893, + 195 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Editing Shape by Moving Points. We design our shape editing scheme with NeuralEditor. This is achieved by interpreting NeRF rendering as \"plotting\" the sampled points over the viewing ray. If we render the scene by naively plotting the point cloud $P$ , the shape editing task can be addressed by replacing each point's coordinate from $p_j$ to $q_j$ . For NeuralEditor, we similarly replace the underlying point cloud from $P$ to $Q(P)$ , while maintaining the confidence values and features. This method is general and can also be applied to any point-based NeRF model like PointNeRF.", + "bbox": [ + 496, + 402, + 890, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Correcting View-Dependence with Infinitesimal Surface Transformation (IST). The editing method above can already obtain reasonable results. However, as illustrated in Fig. 5, the modeled view-dependent colors record the absolute viewing direction, making them incorrect after deformations that change their orientation.", + "bbox": [ + 496, + 553, + 890, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To solve this issue, we model the infinitesimal surface transformation (IST) for each point to redirect the viewing ray in the correct direction. We construct a local coordinate system for each point to represent the orientation of the infinitesimal surface, using its normal vector and two point indices that are neighbors of the $j$ -th point in both $P$ and $Q(P)$ . By comparing these two coordinate systems, we can obtain an affine transformation $\\mathrm{IST}_j$ for the $j$ -th point to redirect the querying view direction (Fig. 5). This procedure is different from modeling space deformation [30, 44, 51], as we only need to model a simple affine transformation at each point, while those methods model a complicated, continuous, and non-linear deformation in the whole space.", + "bbox": [ + 496, + 643, + 892, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our proposed method requires a precise point cloud with normal vector-based color modeling. It is thus incompatible with PointNeRF, as PointNeRF is unable to obtain a desired point cloud to estimate the surface normal vectors.", + "bbox": [ + 496, + 839, + 893, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "12443", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Fine-Tuning on Deformed Scene. Using the shape editing scheme introduced above, we can apply shape deformation on the scene modeled by our NeuralEditor without any modification to the model architecture or rendering scheme, which means that the resulting model is still a valid, fully functional NeuralEditor. Therefore, we can further fine-tune NeuralEditor on the deformed scene if the ground truth is available. We can even fine-tune the infinitesimal surface transformation with other parameters, to rapidly adjust toward better ambient consistency. This makes NeuralEditor desirable in practice, since in most applications, the final goal is not a zero-shot inference, but to fit the deformed scene well with reduced cost. By supporting fine-tuning, our NeuralEditor aligns well with and achieves this goal.", + "bbox": [ + 75, + 90, + 472, + 301 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As another point-based NeRF model, PointNeRF supports fine-tuning but cannot leverage infinitesimal surface transformation fine-tuning to further optimize the performance. On the other hand, mesh-based NeRF editing models [30, 44, 51] do not support fine-tuning well: With deforming the space instead of the scene, these models' rendering scheme has highly changed. In their rendering process, a ray may go through a long, irregular way to reach the scene's surface. As the modeled space deformation might not be precise, it could be hard to tune the irregular space well, and even hurt other parts of the trained NeRF model. Such issues occur especially for some spaces with non-uniform density, since most of their model components (e.g., positional encodings, voxels) are not designed to deal with non-uniform spaces. Among all these methods, only our NeuralEditor has complete support for fine-tuning.", + "bbox": [ + 75, + 301, + 473, + 544 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiment", + "text_level": 1, + "bbox": [ + 76, + 558, + 202, + 574 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Point Cloud Generation. The underlying point cloud is fundamental to all editing tasks. Fig. 6 first provides a qualitative comparison of point clouds generated by our NeuralEditor and PointNeRF [43] on NeRF Synthetic [27]. Ours are much more precise with sharper details, e.g., the mayonnaise on the Hotdog's sausage, the uneven texture on the Chair's cushion, the edge of the Mic's stand, and the Lego brick's studs. By contrast, PointNeRF's point clouds are blurred and noisy, lose most of the details, and even contain obvious shape deflects on the Hotdog's plate and Chair's backrest. This shows that while the point cloud generation task is challenging, NeuralEditor generates a superprecise point cloud which is crucial for shape editing tasks.", + "bbox": [ + 75, + 583, + 468, + 779 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Experimental Settings. We mainly conduct experiments based on scenes from the NeRF Synthetic (NS) dataset. NS is a widely-used NeRF benchmark constructed from Blender [7] scenes. Due to the lack of publicly available benchmarks for shape editing, we use Blender to construct a reproducible benchmark, including the ground truth of edited scenes for evaluation and fine-tuning. Our shape editing tasks cover all eight scenes in NS, while prior", + "bbox": [ + 75, + 780, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/02a70b85b1e99d27d431dac987eb4762b95b169bb6a63eb8ecc38593f79e6b2c.jpg", + "image_caption": [ + "Figure 6. NeuralEditor generates much more precise point clouds than PointNeRF [43] in the four scenes of NeRF Synthetic [27]. The points are colored with their normal vectors." + ], + "image_footnote": [], + "bbox": [ + 517, + 88, + 880, + 362 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/bc6bc7668bbcdc03a475b7cb90437a3cb9c2b590519e5aae2ac1ef1ff30e88e1.jpg", + "image_caption": [ + "Figure 7. With too coarse cages, DeformingNeRF [44] is unable to perform the deformation faithfully, leading to poor results." + ], + "image_footnote": [], + "bbox": [ + 501, + 412, + 890, + 474 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "work [30,44,51] only picks a few scenes. The provided images for NS scenes are with opacity, and there is no requirement for the background color. We evaluate and visualize the results on a black background, for better contrast and clearer detail visualization. In the supplementary, we show the results on a white background with same conclusions.", + "bbox": [ + 498, + 507, + 890, + 597 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Shape Editing Tasks. We evaluate our model on two types of shape editing tasks, as shown in Fig. 1:", + "bbox": [ + 498, + 598, + 890, + 627 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(I) Shape (Mesh) Deformation Task. We consider the shape deformation task as in [30, 44, 51]: deform the shape of a scene in a human-guided way. To construct our deformation tasks from NS and obtain the ground truth, we apply the shape deformation simultaneously to the scene and our point cloud within the provided Blender file. We perform both zero-shot inference and fine-tuning, and compare our rendering results with the ground truth. As demonstrated in Figs. 7 and 8, our deformation tasks are much more precise and aggressive, compared with those in previous work [30, 44, 51]. In the supplementary, we also design a non-continuous deformation task and deformation tasks on the real-world dataset Tanks and Temples [17] (with zero-shot inference only, as there is no ground truth available).", + "(II) Scene Morphing Task. We address a more challenging shape editing task that has not been investigated in prior NeRF editing work: the scene morphing task. Given two scenes $A$ and $B$ , we should construct a path to gradually" + ], + "bbox": [ + 496, + 628, + 893, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "12444", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/8dad0916d0c0ad591e03aa6c149a8319e832ceec95835714224d012874f9bcc7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelZero-Shot Inference, PSNR ↑Fine-Tune for 1 Epoch, PSNR ↑
ChairHotdogLegoDrumsFicusMaterialsMicShipChairHotdogLegoDrumsFicusMaterialsMicShip
DeformingNeRF [44]18.84-13.10-------------
PointNeRF [43]22.2125.9524.5621.0024.2421.2126.7721.1930.1136.0831.4527.1631.4827.5534.3428.90
Naive Plotting24.9127.0125.6421.2926.2221.6527.6322.2932.0136.3831.7228.0933.2130.3135.1530.01
NeuralEditor w/o IST24.9227.0225.6521.2926.2421.6427.6422.2832.2436.6932.7928.3033.3430.4035.2830.08
NeuralEditor (Ours)25.8527.4927.4621.8427.1923.1827.7524.1632.5337.2232.9528.3533.5330.8235.4630.44
", + "bbox": [ + 78, + 89, + 890, + 169 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. NeuralEditor significantly and consistently outperforms PointNeRF and Naive Plotting on all deformed scenes of NeRF Synthetic in peak signal-to-noise ratio (PSNR), in both zero-shot inference and fine-tuning settings. Our infinitesimal surface transformation (IST) effectively improves the results by correcting the view-dependent colors. With the precise point cloud generated by NeuralEditor, even Naive Plotting consistently outperforms PointNeRF. Comparison results under other metrics are in the supplementary.", + "bbox": [ + 75, + 170, + 892, + 224 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f74a14115e3021e730674e5b0bd209b5f0ffc0552d64ad1f60b5c780cebbc1f2.jpg", + "image_caption": [ + "Figure 8. NeuralEditor produces superior rendering results to PointNeRF, with significantly fewer artifacts in zero-shot inference. Finetuning further improves the consistency of rendering with the ambient environment. We use a black background for better visualization." + ], + "image_footnote": [], + "bbox": [ + 80, + 226, + 890, + 428 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "change one scene to the other, and the intermediate scenes should have reasonable appearances. We are required to render all intermediate scenes. For this task, we use the point cloud diffusion model [24] to generate intermediate point clouds with latent space interpolation between $A$ and $B$ as in [24], and we introduce a K-D tree-based [4] matching algorithm to match the adjacent points to fix the indices of the intermediate scenes. To render an intermediate scene, we apply the shape transformation to the NeuralEditor models trained for scenes $A$ and $B$ , and then interpolate the rendering features to obtain the features of the intermediate scene for rendering.", + "bbox": [ + 75, + 462, + 468, + 642 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "NeuralEditor Variants. (1) Full NeuralEditor: Our complete model with all components. (2) NeuralEditor without infinitesimal surface transformation (IST): We remove the maintenance and optimization of infinitesimal surface transformation in scene editing. This key variant of NeuralEditor enables us to evaluate the importance of IST as well as other components of our model. The full ablation study of NeuralEditor is in the supplementary.", + "bbox": [ + 75, + 643, + 468, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Baselines. We compare NeuralEditor against different types of baselines as follows. (1) Naive Plotting: We use the point cloud generated by NeuralEditor, computing each point's opacity and view-dependent colors with their point features. We render the scene by directly plotting/projecting the point cloud to the camera plane. (2) PointNeRF [43]: For the shape deformation task, we apply the same deformation to the point clouds generated by PointNeRF. For the scene morphing task, we apply the matching algorithm to", + "bbox": [ + 75, + 763, + 468, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the point clouds generated by PointNeRF and the same intermediate point clouds generated by the point cloud diffusion model [24] for fairness. (3) DeformingNeRF [44]: DeformingNeRF is not compatible with the scene morphing task. For the shape deformation task, we perform the same deformation on vertices of given cages. Note that DeformingNeRF only released trained models for Lego and Chair, so we can only evaluate it on these two scenes. While other models [30,51] support NeRF-based shape deformation via cages or exported meshes, we were unable to use them as baselines - they did not provide executable code nor their deformed scenes for us to evaluate on their tasks.", + "bbox": [ + 496, + 462, + 892, + 642 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Shape (Mesh) Deformation Results. The qualitative comparison is shown in Fig. 8. Both PointNeRF and Naive Plotting have many artifacts, like blur, wrong color, black or white shadows, noise, etc., whereas our powerful NeuralEditor produces clean and realistic rendering results. After fine-tuning, NeuralEditor shows a significant improvement with better rendering results than PointNeRF, indicating that NeuralEditor is able to achieve higher consistency with the surrounding ambient environment. Notably, in the Materials scene (the 6th scene from left), only our NeuralEditor generates reasonable reflection, while both baselines show blurry and visually messy results. Also in the Drums scene, the bottom face of the gong is not visible in any of the training views in the original scene, so all models render poor results in zero-shot inference. However, after only fast fine-tuning, NeuralEditor is able to precisely model the previously unknown surface and generate a sub-", + "bbox": [ + 496, + 643, + 892, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "12445", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e5fdd963b7c5b0e7575fa808d7bac345a35a0ce9533eb71732a090c11ce98b2f.jpg", + "image_caption": [ + "Figure 9. Our NeuralEditor produces smooth morphing results between Chair, Hotdog, Lego, and Mic in the NeRF Synthetic dataset, while PointNeRF produces results with blurry textures, black shadows, and gloomy, non-smooth colors. The rendering results in the looped morphing process are arranged in the shape of the numerical digit \"3,\" indicated by the dividing lines and arrows." + ], + "image_footnote": [], + "bbox": [ + 153, + 88, + 821, + 435 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "stantially better result than PointNeRF, highlighting NeuralEditor's strength in fast-fitting. All these results demonstrate that NeuralEditor can handle various visual effects and make them consistent in the deformed scene. We provide the figure with higher resolution in the supplementary.", + "bbox": [ + 75, + 481, + 468, + 558 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The quantitative comparison is summarized in Table 1. We observe that: (1) Our NeuralEditor consistently outperforms all the baselines and variants for both zero-shot inference and fine-tuning settings. (2) With the precise point cloud generated by NeuralEditor, the Naive Plotting baseline even consistently outperforms PointNeRF. (3) Our 'w/o IST' variant has a comparable performance to Naive Plotting with the same point cloud and features in the zero-shot inference setting, but after fine-tuning its performance is significantly higher than Naive Plotting, validating the capability of NeuralEditor in NeRF modeling.", + "bbox": [ + 75, + 561, + 468, + 728 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Notably, DeformingNeRF [44] performs poorly in our benchmark with significantly lower metric values. As shown in Fig. 7, the cages provided by DeformingNeRF are too coarse, and cannot even cover the whole scene. Therefore, DeformingNeRF cannot faithfully perform the precise deformation in our benchmark, leading to poor rendering results. On the contrary, both PointNeRF and our NeuralEditor at least faithfully perform the deformation, showing that point cloud is necessary for precise shape editing.", + "bbox": [ + 75, + 729, + 468, + 867 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Scene Morphing Results. The morphing results between 4 NeRF Synthetic scenes are shown in Fig. 9. The", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "morphing process starts from Chair, morphs to Hotdog, Lego, Mic, and at last turns back to Chair. NeuralEditor produces smooth rendering results on the point cloud diffusion-guided [24] intermediate scenes, mixing the textures of the two scenes in a reasonable way. In comparison, the rendering results produced by PointNeRF are unsatisfactory, with blurry textures, black shadows, and gloomy, non-smooth colors. These results show that our NeuralEditor can render challenging intermediate morphing scenes and achieve decent results with only the input of moved points.", + "bbox": [ + 496, + 482, + 890, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 643, + 617, + 659 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper proposes NeuralEditor, a point cloud-guided NeRF model that supports general shape editing tasks by manipulating the underlying point clouds. Empirical evaluation shows NeuralEditor to produce rendering results of much higher quality than baselines in a zero-shot inference manner, further significantly improving after fast finetuning. NeuralEditor even supports smooth scene morphing between multiple scenes, which is difficult for prior work. We hope that our work can inspire more research on point cloud-guided NeRFs and 3D shape and scene editing tasks. Acknowledgement. This work was supported in part by NSF Grant 2106825, NIFA Award 2020-67021-32799, the Jump ARCHES endowment, the NCSA Fellows program, the IBM-Illinois Discovery Accelerator Institute, the Illinois-Inspor Partnership, and the Amazon Research Award. This work used NVIDIA GPUs at NCSA Delta through allocation CIS220014 from the ACCESS program. We thank the authors of NeRF [27] for their help in processing Blender files of the NS dataset.", + "bbox": [ + 496, + 666, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "12446", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Henrik Aanaes, Rasmus Ramsbøl Jensen, George Vogiatzis, Engin Tola, and Anders Bjorholm Dahl. Large-scale data for multiple-view stereopsis. International Journal of Computer Vision, 120(2):153-168, 2016. 5", + "[2] Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas J. Guibas. Learning representations and generative models for 3D point clouds. In ICML, 2018. 2", + "[3] Jonathan T. Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P. Srinivasan. Mip-NeRF: A multiscale representation for antiailiasing neural radiance fields. In ICCV, 2021. 1, 2", + "[4] Jon Louis Bentley. Multidimensional binary search trees used for associative searching. Commun. ACM, 18(9):509-517, 1975. 2, 3, 7", + "[5] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. MVSNeRF: Fast generalizable radiance field reconstruction from multi-view stereo. In ICCV, 2021. 2", + "[6] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In CVPR, 2019. 2", + "[7] Blender Online Community. *Blender - a 3D modelling and rendering package*. Blender Foundation, Stichting Blender Foundation, Amsterdam, 2018. 6", + "[8] Frank Dellaert and Lin Yen-Chen. Neural volume rendering: NeRF and beyond. arXiv:2101.05204, 2021. 2", + "[9] Kyle Gao, Yina Gao, Hongjie He, Denning Lu, Linlin Xu, and Jonathan Li. NeRF: Neural radiance field in 3D vision, a comprehensive review. arXiv:2101.05204, 2021. 2", + "[10] Yuan-Chen Guo, Di Kang, Linchao Bao, Yu He, and Song-Hai Zhang. NeRFReN: Neural radiance fields with reflections. In CVPR, 2022. 2", + "[11] Peter Hedman, Tobias Ritschel, George Drettakis, and Gabriel Brostow. Scalable inside-out image-based rendering. ACM Trans. Graph., 35(6):231:1-231:11, 2016. 2", + "[12] Po-Han Huang, Kevin Matzen, Johannes Kopf, Narendra Ahuja, and Jia-Bin Huang. DeepMVS: Learning multi-view stereopsis. In CVPR, 2018. 5", + "[13] Alec Jacobson, Ilya Baran, Ladislav Kavan, Jovan Popovic, and Olga Sorkine. Fast automatic skinning transformations. ACM Trans. Graph., 31(4), 2012. 3", + "[14] Tomas Jakab, Richard Tucker, Ameesh Makadia, Jiajun Wu, Noah Snavely, and Angjoo Kanazawa. Keypointdeformer: Unsupervised 3D keypoint discovery for shape control. In CVPR, 2021. 3", + "[15] Mengqi Ji, Juergen Gall, Haitian Zheng, Yebin Liu, and Lu Fang. Surfacenet: An end-to-end 3D neural network for multiview stereopsis. In ICCV, 2017. 2", + "[16] Tao Ju, Scott Schaefer, and Joe Warren. Mean value coordinates for closed triangular meshes. ACM Trans. Graph., 24(3):561-566, 2005. 3", + "[17] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and Temples: Benchmarking large-scale scene reconstruction. ACM Trans. Graph., 36(4), 2017. 6", + "[18] Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitzmann. Decomposing NeRF for editing via feature field distillation. In NeurlPS, 2022. 2" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[19] Marc Levoy and Pat Hanrahan. Light field rendering. In SIGGRAPH, 1996. 2", + "[20] Fayao Liu, Chunhua Shen, Guosheng Lin, and Ian Reid. Learning depth from single monocular images using deep convolutional neural fields. IEEE Transactions on Pattern Analysis and Machine Intelligence, 38(10):2024-2039, 2015. 2", + "[21] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. In NeurIPS, 2020. 2", + "[22] Steven Liu, Xiuming Zhang, Zhoutong Zhang, Richard Zhang, Junyan Zhu, and Bryan C. Russell. Editing conditional radiance fields. In ICCV, 2021. 2", + "[23] William E. Lorensen and Harvey E. Cline. Marching cubes: A high resolution 3D surface construction algorithm. In SIGGRAPH, 1987. 3", + "[24] Shitong Luo and Wei Hu. Diffusion probabilistic models for 3D point cloud generation. In CVPR, 2021. 2, 7, 8", + "[25] Bruce Merry, Patrick Marais, and James Gain. Animation space: A truly linear framework for character animation. ACM Trans. Graph., 25(4):1400-1423, 2006. 3", + "[26] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3D reconstruction in function space. In CVPR, 2019. 2", + "[27] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2, 4, 6, 8", + "[28] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Differentiable volumetric rendering: Learning implicit 3D representations without 3D supervision. In CVPR, 2020. 2", + "[29] Jesús Nieto and Toni Susin. Cage based deformations: A survey. Lecture Notes in Computational Vision and Biomechanics, 7:75-99, 2013. 3", + "[30] Yicong Peng, Yichao Yan, Shengqi Liu, Yuhao Cheng, Shanyan Guan, Bowen Pan, Guangtao Zhai, and Xiaokang Yang. CageNeRF: Cage-based neural radiance field for generalized 3D deformation and animation. In NeurIPS, 2022. 2, 3, 5, 6, 7", + "[31] Bui Tuong Phong. Illumination for computer generated pictures. Commun. ACM, 18(6):311-317, 1975. 2, 4", + "[32] Yi-Ling Qiao, Alexander Gao, and Ming C. Lin. NeuPhysics: Editable neural geometry and physics from monocular videos. In NeurIPS, 2022. 2", + "[33] Sara Fridovich-Keil and Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In CVPR, 2022. 1, 2", + "[34] Noah Snavely, Steven M. Seitz, and Richard Szeliski. Photo tourism: Exploring photo collections in 3D. ACM Trans. Graph., 25(3):835-846, 2006. 2", + "[35] Jiaxiang Tang, Xiaokang Chen, Jingbo Wang, and Gang Zeng. Compressible-composable NeRF via rank-residual decomposition. In NeurIPS, 2022. 2" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "12447", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[36] Jean-Marc Thiery, Julien Tierny, and Tamy Boubekeur. Jacobians and hessian of mean value coordinates for closed triangular meshes. The Visual Computer, 30(9):981-995, 2014. 3", + "[37] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T. Barron, and Pratul P. Srinivasan. Ref-NeRF: Structured view-dependent appearance for neural radiance fields. In CVPR, 2022. 1, 2, 4, 5", + "[38] Nanyang Wang, Yinda Zhang, Zhuwen Li, Yanwei Fu, Wei Liu, and Yu-Gang Jiang. Pixel2Mesh: Generating 3D mesh models from single RGB images. In ECCV, 2018. 2", + "[39] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul Srinivasan, Howard Zhou, Jonathan T. Barron, Ricardo MartinBrualla, Noah Snavely, and Thomas Funkhouser. IBRNet: Learning multi-view image-based rendering. In CVPR, 2021. 1, 2", + "[40] Suttisak Wizadwongsa, Pakkapon Phongthawee, Jiraphon Yenphraphai, and Supasorn Suwajanakorn. NeX: Real-time view synthesis with neural basis expansion. In CVPR, 2021. 2", + "[41] Liwen Wu, Jae Yong Lee, Anand Bhattachad, Yu-Xiong Wang, and David Forsyth. DIVeR: Real-time and accurate neural radiance fields with deterministic integration for volume rendering. In CVPR, 2022. 2, 4", + "[42] Dejia Xu, Peihao Wang, Yifan Jiang, Zhiwen Fan, and Zhangyang Wang. Signal processing for implicit neural representations. In NeurIPS, 2022. 2", + "[43] Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-NeRF: Point-based neural radiance fields. In CVPR, 2021. 2, 3, 6, 7", + "[44] Tianhan Xu and Tatsuya Harada. Deforming radiance fields with cages. In ECCV, 2022. 2, 3, 5, 6, 7, 8", + "[45] Bangbang Yang, Yinda Zhang, Yinghao Xu, Yijin Li, Han Zhou, Hujun Bao, Guofeng Zhang, and Zhaopeng Cui. Learning object-compositional neural radiance field for editing scene rendering. In ICCV, 2021. 2", + "[46] Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. MVSNet: Depth inference for unstructured multiview stereo. In ECCV, 2018. 5", + "[47] Wang Yifan, Noam Aigerman, Vladimir G Kim, Siddhartha Chaudhuri, and Olga Sorkine-Hornung. Neural cages for detail-preserving 3D deformations. In CVPR, 2020. 3", + "[48] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. PlenOctrees for real-time rendering of neural radiance fields. In ICCV, 2021. 1, 2", + "[49] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelNeRF: Neural radiance fields from one or few images. In CVPR, 2021. 2", + "[50] Yu-Jie Yuan, Yu-Kun Lai, Tong Wu, Lin Gao, and Ligang Liu. A revisit of shape editing techniques: From the geometric to the neural viewpoint. Journal of Computer Science and Technology, 36(3):520-554, 2021. 3", + "[51] Yu-Jie Yuan, Yang tian Sun, Yu-Kun Lai, Yuewen Ma, Rongfei Jia, and Lin Gao. NeRF-Editing: Geometry editing of neural radiance fields. In CVPR, 2022. 2, 3, 5, 6, 7" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[52] Yuzhe Zhang, Jianmin Zheng, and Yiyu Cai. Proxy-driven free-form deformation by topology-adjustable control lattice. Computers & Graphics, 89:167-177, 2020. 3", + "[53] Qian-Yi Zhou, Jaesik Park, and Vladlen Koltun. Open3D: A modern library for 3D data processing. arXiv:1801.09847, 2018. 4" + ], + "bbox": [ + 501, + 90, + 892, + 175 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "12448", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/NeuralEditor_ Editing Neural Radiance Fields via Manipulating Point Clouds/70b7a137-ced8-4460-839e-a2dd484a031c_model.json b/2023/NeuralEditor_ Editing Neural Radiance Fields via Manipulating Point Clouds/70b7a137-ced8-4460-839e-a2dd484a031c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d4fe1f6365fdfaa2c446e0cf55a67d82c8cc9508 --- /dev/null +++ b/2023/NeuralEditor_ Editing Neural Radiance Fields via Manipulating Point Clouds/70b7a137-ced8-4460-839e-a2dd484a031c_model.json @@ -0,0 +1,1969 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.094, + 0.131, + 0.877, + 0.153 + ], + "angle": 0, + "content": "NeuralEditor: Editing Neural Radiance Fields via Manipulating Point Clouds" + }, + { + "type": "text", + "bbox": [ + 0.253, + 0.18, + 0.387, + 0.199 + ], + "angle": 0, + "content": "Jun-Kun Chen\\(^{1\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.427, + 0.181, + 0.531, + 0.199 + ], + "angle": 0, + "content": "Jipeng Lyu\\(^{2\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.573, + 0.181, + 0.71, + 0.199 + ], + "angle": 0, + "content": "Yu-Xiong Wang" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.2, + 0.463, + 0.219 + ], + "angle": 0, + "content": "1University of Illinois at Urbana-Champaign" + }, + { + "type": "text", + "bbox": [ + 0.503, + 0.201, + 0.855, + 0.219 + ], + "angle": 0, + "content": "\\(^{2}\\)Peking University \\(\\dagger\\) Equal Contribution" + }, + { + "type": "text", + "bbox": [ + 0.195, + 0.221, + 0.771, + 0.238 + ], + "angle": 0, + "content": "{junkun3，yxw}@illinois.edu lvjipeng@pku.edu.cn" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.264, + 0.302, + 0.421 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.308, + 0.264, + 0.892, + 0.426 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.436, + 0.893, + 0.478 + ], + "angle": 0, + "content": "Figure 1. Our NeuralEditor offers native support for general and flexible shape editing of neural radiance fields via manipulating point clouds. By generating a precise point cloud of the scene with a novel point cloud-guided NeRF model, our NeuralEditor produces high-fidelity rendering results in both shape deformation and more challenging scene morphing tasks." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.489, + 0.314, + 0.504 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.51, + 0.474, + 0.829 + ], + "angle": 0, + "content": "This paper proposes NeuralEditor that enables neural radiance fields (NeRFs) natively editable for general shape editing tasks. Despite their impressive results on novel-view synthesis, it remains a fundamental challenge for NeRFs to edit the shape of the scene. Our key insight is to exploit the explicit point cloud representation as the underlying structure to construct NeRFs, inspired by the intuitive interpretation of NeRF rendering as a process that projects or \"plots\" the associated 3D point cloud to a 2D image plane. To this end, NeuralEditor introduces a novel rendering scheme based on deterministic integration within K-D tree-guided density-adaptive voxels, which produces both high-quality rendering results and precise point clouds through optimization. NeuralEditor then performs shape editing via mapping associated points between point clouds. Extensive evaluation shows that NeuralEditor achieves state-of-the-art performance in both shape deformation and scene morphing tasks. Notably, NeuralEditor supports both zero-shot inference and further fine-tuning over the edited scene. Our code, benchmark, and demo video are available at immortalco.github.io/NeuralEditor." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.846, + 0.208, + 0.86 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Perhaps the most memorable shot of the film Transformers, Optimus Prime is seamlessly transformed between a" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.49, + 0.893, + 0.7 + ], + "angle": 0, + "content": "humanoid and a Peterbilt truck – such free-form editing of 3D objects and scenes is a fundamental task in 3D computer vision and computer graphics, directly impacting applications such as visual simulation, movie, and game industries. In these applications, often we are required to manipulate a scene or objects in the scene by editing or modifying its shape, color, light condition, etc., and generate visually-faithful rendering results on the edited scene efficiently. Among the various editing operations, shape editing has received continued attention but remains challenging, where the scene is deformed in a human-guided way, while all of its visual attributes (e.g., shape, color, brightness, and light condition) are supposed to be natural and consistent with the ambient environment." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.703, + 0.892, + 0.824 + ], + "angle": 0, + "content": "State-of-the-art rendering models are based on implicit neural representations, as exemplified by neural radiance field (NeRF) [27] and its variants [3, 33, 37, 39, 48]. Despite their impressive novel-view synthesis results, most of the NeRF models substantially lack the ability for users to adjust, edit, or modify the shape of scene objects. On the other hand, shape editing operations can be natively applied to explicit 3D representations such as point clouds and meshes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Inspired by this, we propose NeuralEditor – a general and flexible approach to editing neural radiance fields via manipulating point clouds (Fig. 1). Our key insight is to benefit from the best of both worlds: the superiority in rendering performance from implicit neural representation" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12439" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.137 + ], + "angle": 0, + "content": "combined with the ease of editing from explicit point cloud representation. NeuralEditor enables us to perform a wide spectrum of shape editing operations in a consistent way." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.139, + 0.468, + 0.455 + ], + "angle": 0, + "content": "Such introduction of point clouds into NeRF for general shape editing is rooted in our interpretation of NeRF rendering as a process that projects or \"plots\" the associated 3D point cloud to a 2D image plane. Conceptually, with a dense enough point cloud where each point has an opacity and its color is defined as a function of viewing direction, directly plotting the point cloud would achieve similar visual effects (i.e., transparency and view-dependent colors) that are rendered by NeRF. This intrinsic integration between NeRF and point clouds underscores the advantage of our NeuralEditor over existing mesh-based NeRF editing methods such as NeRF-Editing [51], Deforming-NeRF [44], and CageNeRF [30], where the process of constructing and optimizing the mesh is separated from the NeRF modeling, making them time-consuming. More importantly, with the point cloud constructed for a scene, the shape editing can be natively defined as and easily solved by just moving each point into the new, edited position and re-plotting the point cloud. Therefore, our approach supports more general scene editing operations which are difficult to achieve via mesh-guided space deformation." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.459, + 0.468, + 0.685 + ], + "angle": 0, + "content": "The key component in our NeuralEditor lies in a point cloud-guided NeRF model that natively supports general shape editing operations. While the recent method Point-NeRF [43] has demonstrated improved novel-view synthesis capability based on point clouds, it is not supportive to shape editing. Our idea then is to exploit the underlying point cloud in ways of not only optimizing its structure and features (e.g., adaptive voxels) for rendering, but also extracting additional useful attributes (e.g., normal vectors) to guide the editing process. To this end, we introduce K-D trees [4] to construct density-adaptive voxels for efficient and stable rendering, together with a novel deterministic integration strategy. Moreover, we model the color with the Phong reflection [31] to decompose the specular color and better represent the scene geometry." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.687, + 0.468, + 0.853 + ], + "angle": 0, + "content": "With a much more precise point cloud attributed to these improvements, our NeuralEditor achieves high-fidelity rendering results on deformed scenes compared with prior work as shown in Fig. 1, even in a zero-shot inference manner without additional training. Through fast fine-tuning, the visual quality of the deformed scene is further enhanced, almost perfectly consistent with the surrounding light condition. In addition, under the guidance of a point cloud diffusion model [24], NeuralEditor can be naturally extended for smooth scene morphing across multiple scenes, which is difficult for existing NeRF editing work." + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.856, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Our contributions are four-fold. (1) We introduce NeuralEditor, a flexible and versatile approach that makes neural radiance fields editable through manipulating point" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.092, + 0.891, + 0.272 + ], + "angle": 0, + "content": "clouds. (2) We propose a point cloud-guided NeRF model based on K-D trees and deterministic integration, which produces precise point clouds and supports general scene editing. (3) Due to the lack of publicly available benchmarks for shape editing, we construct and release a reproducible benchmark that promotes future research on shape editing. (4) We investigate a wide range of shape editing tasks, covering both shape deformation (as studied in existing NeRF editing work) and challenging scene morphing (a novel task addressed here). NeuralEditor achieves state-of-the-art performance on all shape editing tasks in a unified framework, without extra information or supervision." + }, + { + "type": "title", + "bbox": [ + 0.504, + 0.287, + 0.64, + 0.302 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.313, + 0.891, + 0.522 + ], + "angle": 0, + "content": "Neural Scene Representation. Traditional methods model scenes with explicit [2, 11, 15, 20, 34, 38] or implicit [6, 19, 26, 28, 40] 3D geometric or shape representations. Initiated by NeRF [27], leveraging implicit neural networks to represent scenes and perform novel-view synthesis has become a fast-developing field in 3D vision [8,9]. While most of the follow-up work focuses on improving aspects such as the rendering realism [3, 10, 21, 37], efficiency [33, 41, 48], and cross-scene generalization [5, 39, 43, 49], the scene editing capability is substantially missing in the NeRF family which we address in this paper. In addition, we exploit K-D tree-guided point clouds as the underlying structure, different from other NeRF variants based on octrees [21, 33, 48] or plain voxels [41]." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.524, + 0.891, + 0.749 + ], + "angle": 0, + "content": "Point-Based NeRFs. Recently, using point clouds to build a NeRF model has shown better encoding of scene shape and improved rendering performance, as represented by PointNeRF [43]. PointNeRF proposes a point initialization network to produce the initial point cloud together with the point features, which is further optimized by a pruning and growing strategy. While both PointNeRF and our NeuralEditor employ point clouds as the underlying structure, NeuralEditor better exploits useful information within the point clouds: PointNeRF only directly uses the locations of points; by contrast, NeuralEditor considers the point cloud more as a geometrical shape and extracts relevant information like normal vectors, which plays an important role in rendering and shape editing. Importantly, our approach is designed to support scene editing, in contrast to PointNeRF." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.75, + 0.891, + 0.901 + ], + "angle": 0, + "content": "Scene Editing via NeRFs. Different types of scene editing have been studied under NeRFs. EditNeRF [22], ObjectNeRF [45], and DistillNeRF [18] perform simple shape and color editing for objects specified with human-input scribble, pixel, segment, language, etc. NeuPhysics [32] edits a dynamic scene via physics parameters. CCNeRF [35] proposes an explicit NeRF representation with tensor rank decomposition to support scene composition. INSP-Net [42] considers filter editing like denoising. Such work cannot address 3D shape editing and only sup" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.947, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12440" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.121 + ], + "angle": 0, + "content": "ports simple editing operations, like object selection, similarity transformation, or limited shape deformation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.122, + 0.473, + 0.471 + ], + "angle": 0, + "content": "3D Shape Editing. Traditional representation methods support keypoint-based shape editing [13, 14, 16, 25, 36, 47, 52] with meshes [29, 50], which cannot be directly applied to implicit representations used by NeRF. Existing NeRF editing work primarily studies a particular shape editing task, mesh deformation, and addresses it in a common paradigm [30, 44, 51]: A mesh of the scene is first constructed by either exporting it from a trained NeRF with the Marching Cubes algorithm [23], or optimizing close-to-surface cages along with training. After the user deforms the mesh, the deformed scene is rendered by deforming the space and bending the viewing rays in the original scene with the trained NeRF. Doing so requires extra efforts to convert implicit scene representation to explicit mesh, which might not be precise enough, and only supports continuous shape editing that can be converted to space deformation. On the contrary, our NeuralEditor directly maintains and utilizes alternative explicit scene representation – the point cloud which is intrinsically integrated with NeRF, making NeuralEditor require no extra efforts and support more general shape editing tasks like scene morphing. NeuralEditor supports both zero-shot inference and further fine-tuning over the edited scene, while prior work cannot." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.482, + 0.45, + 0.499 + ], + "angle": 0, + "content": "3. NeuralEditor: Point Cloud-Guided NeRF" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.507, + 0.473, + 0.72 + ], + "angle": 0, + "content": "We propose a novel point cloud-guided NeRF model, NeuralEditor - it not only achieves realistic rendering results in the novel-view synthesis task, but also produces a point cloud that precisely describes the shape of the scene, thus facilitating general shape editing tasks. As illustrated in Fig. 2, we leverage the K-D trees [4] to construct density-adaptive voxels (which also naturally enable us to skip empty spaces), and introduce deterministic spline integration for rendering. We use the Phong reflection to model the color along with the normal vectors obtained from the underlying point cloud. With our enhanced point cloud optimization, NeuralEditor obtains much more precise underlying point clouds, compared with noisy and imprecise outputs of state-of-the-art PointNeRF [43] (as shown in Sec. 5)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.727, + 0.305, + 0.743 + ], + "angle": 0, + "content": "3.1. K-D Tree-Guided Voxels" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.47, + 0.841 + ], + "angle": 0, + "content": "To render with points, we construct multi-scale density-adaptive voxels based on K-D trees [4], namely, \\(K - D\\) voxels. K-D trees are a data structure constructed on \\(K\\)-dimensional points, where \\(K = 3\\) for 3D points. As a special decision tree, K-D tree's each node divides the point set into two equal-sized parts with axis-parallel criterion." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.841, + 0.471, + 0.903 + ], + "angle": 0, + "content": "For each K-D tree's node, we compute its bounding box by taking the minimum and maximum \\( x, y, z \\) coordinates in its subtree and with proper padding margins. As we divide the points in a top-down manner in one of the \\( x, y, z \\) direc" + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.09, + 0.885, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.188, + 0.895, + 0.3 + ], + "angle": 0, + "content": "Figure 2. Our K-D Voxels. Column 1: original point clouds constructed from two scenes, colored with normal vector directions. Column 2: upper-level voxels which coarsely represent the shape. Column 3: lower-level voxels which tightly cover the shape. Right: Visualization of K-D voxels on a 2D point cloud. Each color represents boxes of nodes on each level of the K-D tree. Lower-level boxes containing fewer points cover the shape more tightly, and vice versa for higher-level boxes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.306, + 0.895, + 0.429 + ], + "angle": 0, + "content": "tions, in each layer of the K-D tree, different nodes' bounding boxes are mutually exclusive. Therefore, the bounding boxes can be regarded as voxels. As boxes in the upper layers contain more points (larger voxels), while those in the lower layers contain fewer points (smaller voxels), we naively obtain a multi-scale voxel construction from one K-D tree. As shown in Fig. 2, voxels from the large to small scales represent the shape of the scene from coarse to fine." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.438, + 0.753, + 0.455 + ], + "angle": 0, + "content": "3.2. Rendering Over K-D Voxels" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.462, + 0.895, + 0.553 + ], + "angle": 0, + "content": "We now introduce a rendering scheme that exploits K-D voxels to perform all the sub-procedures associated with rendering in a unified way. This scheme enables us to render more naturally, efficiently, and even deterministically, meanwhile it also simplifies some widely-adopted design choices in conventional NeRF rendering." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.553, + 0.896, + 0.825 + ], + "angle": 0, + "content": "Skipping Empty Spaces. In NeRF rendering, we are supposed to focus only on the surface of scene objects. As shown in Fig. 2, all our K-D voxels are produced to stick to the surface of objects, which the point cloud is constructed to describe. Such a property allows us to avoid explicitly \"skipping\" empty spaces, which often requires extra consideration in most NeRF models - only considering the space inside a voxel automatically focuses on the surface; as the depth of the voxel's node goes deeper, it becomes closer to the surface. Moreover, during the construction of the K-D tree, the points at each node are divided within its sub-nodes, and the node's voxel fully covers all its sub-nodes. This further provides us with a native top-down recursive procedure to locate the voxels intersected with the querying ray: We start from the root node, and recurse on the sub-nodes until we (1) reach a pre-set node depth (or equivalently, a pre-set voxel size) and then query within the associated voxel, or (2) stop recursion on non-intersected nodes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Density-Adaptive Rendering. An important design in NeRFs is the coarse-to-fine strategy for density-adaptive rendering, so that more points are sampled in high-volume density areas. Our K-D voxels natively support such a design without additional bells and whistles. This is because" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "12441" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.202, + 0.079, + 0.773, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.241, + 0.894, + 0.27 + ], + "angle": 0, + "content": "Figure 3. Our NeuralEditor architecture. We propose deterministic spline integration for KNN-based point features over each K-D tree-guided density-adaptive voxel, and model the color via Phong reflection with normal vectors estimated from the point cloud's shape." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.274, + 0.473, + 0.485 + ], + "angle": 0, + "content": "voxels in the same K-D tree layer contain the same number of points. As the point density can be regarded as an approximation of volume density, all such voxels have the same density. Therefore, we directly use K-D voxels to guide the density-adaptive rendering. Specifically, we conduct the rendering process at the voxels of some bottom layers in the K-D tree. For each querying ray, we use the aforementioned recursive procedure to locate the minimal intersected voxels that are deep enough. Here \"minimal\" means that the ray intersects with the node's voxel, but does not intersect with any sub-node's voxel. These intersected voxels divide the querying ray into several segments (Fig. 3). The segments covered by a voxel are close to the surface and used for rendering, while those not covered are in empty spaces." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.486, + 0.47, + 0.637 + ], + "angle": 0, + "content": "Deterministic Spline Integration. DIVER [41] shows that deterministic integration outperforms stochastic integration in NeRF rendering. So we perform a deterministic integration to obtain the segment's feature within each voxel. Since we do not necessarily have points at the voxel's vertices, the trilinear interpolation used in DIVER is not feasible here. Instead, we use spline integration. For the \\(i\\)-th intersected voxel in the ray passing order, we uniformly select points in this segment, and integrate their features to obtain the average feature \\(f_{i}\\) of the segment of the \\(i\\)-th voxel:" + }, + { + "type": "equation", + "bbox": [ + 0.144, + 0.643, + 0.47, + 0.677 + ], + "angle": 0, + "content": "\\[\nf _ {i} = \\frac {1}{r _ {i} - l _ {i}} \\int_ {l _ {i}} ^ {r _ {i}} \\mathbf {f e a t u r e} (o + t \\cdot d) \\mathrm {d} t, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.678, + 0.47, + 0.739 + ], + "angle": 0, + "content": "where \\([l_i, r_i]\\) is the intersection interval, and \\(o\\) and \\(d\\) are the source and direction of the querying ray, respectively. This average feature \\(f_i\\) can be interpreted as the feature of a representative point \\(p_i\\) located somewhere in the segment." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.74, + 0.469, + 0.8 + ], + "angle": 0, + "content": "KNN-Based Feature Aggregation. For each uniformly selected point \\( q \\) in the segment during spline integration, we obtain its feature via weighted interpolation from the features of its \\( K \\) nearest neighbors (KNN) in the point cloud:" + }, + { + "type": "equation", + "bbox": [ + 0.114, + 0.807, + 0.47, + 0.845 + ], + "angle": 0, + "content": "\\[\n\\operatorname {f e a t u r e} (q) = \\sum_ {p _ {j} \\in \\operatorname {K N N} (q; K)} k _ {j} e _ {j}, \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.116, + 0.845, + 0.432, + 0.869 + ], + "angle": 0, + "content": "\\[\n\\left\\{k_{j}\\right\\} = \\operatorname *{SoftMax}_{p_{j}\\in \\mathrm{KNN}(q;K)}\\left(\\log \\gamma_{j} - \\log \\| q - p_{j}\\|_{2}^{2}\\right),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.903 + ], + "angle": 0, + "content": "where for each point \\( p_j \\) in the point cloud, we parameterize its confidence \\( \\gamma_j \\) and point feature \\( e_j \\) as in PointNeRF." + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.274, + 0.87, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.453, + 0.893, + 0.495 + ], + "angle": 0, + "content": "Figure 4. In the two scenes of NeRF Synthetic [27], NeuralEditor optimizes the rough initial point cloud to a precise point cloud. The points are colored with their normal vectors." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.504, + 0.893, + 0.746 + ], + "angle": 0, + "content": "Phong Reflection-Based Color Modeling with Point Cloud Normal Vectors. To obtain the volume density \\(\\sigma_{i}\\) and color \\(c_{i}\\) of the representative point \\(p_i\\) in the \\(i\\)-th voxel, we use the Phong reflection model [31]. As we have the underlying point cloud, we use Open3D [53] to estimate the normal vector for each point, and integrate these vectors over the interval to get an average normal vector \\(n_i\\). Such information better characterizes the shape of point clouds (scenes), which plays an important role in Phong-based rendering and also implicitly facilitates optimizing the point clouds (Sec. 3.3). Consistent with RefNeRF [37], we use multiple multilayer perceptrons (MLPs) to model other attributes, including volume density, tint, roughness, and diffuse and specular color, and use the Phong formula to calculate the final \\(c_{i}\\) with these attributes. Finally, we aggregate \\(c_{i}\\) of all segments on the ray to obtain the final color:" + }, + { + "type": "equation", + "bbox": [ + 0.546, + 0.755, + 0.843, + 0.787 + ], + "angle": 0, + "content": "\\[\nc _ {\\text {p i x e l}} = \\sum_ {i \\geq 1} \\tau_ {i} \\cdot \\left(1 - \\exp \\left(- \\left(r _ {i} - l _ {i}\\right) \\sigma_ {i}\\right) \\cdot c _ {i}, \\right.\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.549, + 0.787, + 0.891, + 0.833 + ], + "angle": 0, + "content": "\\[\n\\tau_ {i} = \\exp \\left(- \\sum_ {i ^ {\\prime} = 1} ^ {i - 1} \\left(r _ {i ^ {\\prime}} - l _ {i ^ {\\prime}}\\right) \\sigma_ {i ^ {\\prime}}\\right). \\tag {3}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.847, + 0.736, + 0.864 + ], + "angle": 0, + "content": "3.3. Point Cloud Optimization" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Point Cloud Initialization. To start training, we need a coarse initial point cloud. Consistent with PointNeRF, we" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12442" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.212 + ], + "angle": 0, + "content": "use a point generation network, which consists of a multiview stereo (MVS) model [12] based on a 3D convolutional neural network (CNN), to generate the points' coordinates and confidence values, and another 2D CNN [46] to generate their initial features. This network was pre-trained on the DTU training dataset [1], and can generalize to other datasets and scenes. As shown in Fig. 4, the initial point cloud generated by such a network is coarse and noisy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.213, + 0.47, + 0.303 + ], + "angle": 0, + "content": "Explicit Optimization via Pruning and Growing. We perform a similar pruning and growing procedure as in PointNeRF, to prune outliers with low confidence \\(\\gamma_{j}\\) and fill holes in the point clouds. We make several important modifications over PointNeRF, and integrate this procedure with our deterministic integration (details in the supplementary)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.303, + 0.474, + 0.516 + ], + "angle": 0, + "content": "Implicit Optimization with Normal Vectors. In addition to the explicit optimization, the point cloud is also optimized implicitly during training through the adjustment of point confidence \\(\\gamma_{j}\\). When computing the average normal vector for rendering, we aggregate normal vectors of nearby points weighted with their distance and confidence, where the confidence of noisy or inaccurate points with potentially abnormal normal vectors is adjusted accordingly. Moreover, we apply the normal vector regularization losses from RefNeRF [37] to supervise the points' confidence w.r.t. their normal vectors. These strategies collectively provide implicit but more tailored ways to optimize the point clouds. With both explicit and implicit optimization, NeuralEditor obtains very precise point clouds (Fig. 4)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.527, + 0.38, + 0.543 + ], + "angle": 0, + "content": "4. Shape Editing with NeuralEditor" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.552, + 0.47, + 0.629 + ], + "angle": 0, + "content": "Formulation of General Shape Editing Tasks. We define the shape editing tasks based on indexed point clouds. To this end, we first re-define an indexed point cloud \\( P \\) as a mapping from a point index \\( j \\) to the corresponding point \\( p_j \\)," + }, + { + "type": "equation", + "bbox": [ + 0.125, + 0.633, + 0.469, + 0.65 + ], + "angle": 0, + "content": "\\[\nP: j \\rightarrow p _ {j}, \\text {w h e r e} p _ {j} \\in \\mathbb {R} ^ {3}, j = 1, \\dots , | P |. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.47, + 0.689 + ], + "angle": 0, + "content": "A shape editing task is defined as another indexed point cloud \\(Q(P)\\)" + }, + { + "type": "equation", + "bbox": [ + 0.085, + 0.694, + 0.47, + 0.712 + ], + "angle": 0, + "content": "\\[\nQ (P): j \\rightarrow q _ {j}, \\text {w h e r e} q _ {j} \\in \\mathbb {R} ^ {3} \\cup \\{\\varnothing \\}, j = 1, \\dots , | P |, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.47, + 0.796 + ], + "angle": 0, + "content": "describing a shape editing task whether the \\(j\\)-th point moves from \\(p_j\\) to \\(q_j\\) or is deleted in the deformation if \\(q_j = \\emptyset\\). With this definition, \\(Q(P)\\) can be an arbitrary point cloud with points properly matched to points in \\(P\\) by same indices, regardless of connectivity or continuity." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Our formulation represents a broad range of shape editing tasks. The mesh editing tasks in NeRF-Editing [51], Deforming-NeRF [44], and CageNeRF [30] can be more simply and clearly defined here. For example, in NeRF-Editing, a mesh is exported from a general NeRF model, deformed manually, and converted to an \"offset\" or a continuous space deformation. We can depict such a task without" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.09, + 0.681, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.683, + 0.09, + 0.894, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.197, + 0.895, + 0.403 + ], + "angle": 0, + "content": "Figure 5. Infinitesimal surface transformation (IST). (a) As the view-dependent colors are modeled as absolute viewing directions, they (solid arrows at the right) are different from the correct colors (dashed arrows at the right) after deformation. We solve this by (b) constructing a local coordinate system near the \\( j \\)-th point and (c) modeling IST from the edited scene to the original scene with the coordinate systems, so as to (d) redirect the viewing direction to the original scene when rendering the edited scene. \"offsets,\" by recording only the final location for each point without extra information. Notably, our formulation even models those whose deformation is not continuous in the space, e.g., cutting a scene into two parts, and thus cannot be covered and solved by NeRF-Editing, Deforming-NeRF, or CageNeRF, as shown in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.403, + 0.892, + 0.553 + ], + "angle": 0, + "content": "Editing Shape by Moving Points. We design our shape editing scheme with NeuralEditor. This is achieved by interpreting NeRF rendering as \"plotting\" the sampled points over the viewing ray. If we render the scene by naively plotting the point cloud \\( P \\), the shape editing task can be addressed by replacing each point's coordinate from \\( p_j \\) to \\( q_j \\). For NeuralEditor, we similarly replace the underlying point cloud from \\( P \\) to \\( Q(P) \\), while maintaining the confidence values and features. This method is general and can also be applied to any point-based NeRF model like PointNeRF." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.554, + 0.892, + 0.644 + ], + "angle": 0, + "content": "Correcting View-Dependence with Infinitesimal Surface Transformation (IST). The editing method above can already obtain reasonable results. However, as illustrated in Fig. 5, the modeled view-dependent colors record the absolute viewing direction, making them incorrect after deformations that change their orientation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.893, + 0.84 + ], + "angle": 0, + "content": "To solve this issue, we model the infinitesimal surface transformation (IST) for each point to redirect the viewing ray in the correct direction. We construct a local coordinate system for each point to represent the orientation of the infinitesimal surface, using its normal vector and two point indices that are neighbors of the \\(j\\)-th point in both \\(P\\) and \\(Q(P)\\). By comparing these two coordinate systems, we can obtain an affine transformation \\(\\mathrm{IST}_j\\) for the \\(j\\)-th point to redirect the querying view direction (Fig. 5). This procedure is different from modeling space deformation [30, 44, 51], as we only need to model a simple affine transformation at each point, while those methods model a complicated, continuous, and non-linear deformation in the whole space." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Our proposed method requires a precise point cloud with normal vector-based color modeling. It is thus incompatible with PointNeRF, as PointNeRF is unable to obtain a desired point cloud to estimate the surface normal vectors." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12443" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.302 + ], + "angle": 0, + "content": "Fine-Tuning on Deformed Scene. Using the shape editing scheme introduced above, we can apply shape deformation on the scene modeled by our NeuralEditor without any modification to the model architecture or rendering scheme, which means that the resulting model is still a valid, fully functional NeuralEditor. Therefore, we can further fine-tune NeuralEditor on the deformed scene if the ground truth is available. We can even fine-tune the infinitesimal surface transformation with other parameters, to rapidly adjust toward better ambient consistency. This makes NeuralEditor desirable in practice, since in most applications, the final goal is not a zero-shot inference, but to fit the deformed scene well with reduced cost. By supporting fine-tuning, our NeuralEditor aligns well with and achieves this goal." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.303, + 0.474, + 0.545 + ], + "angle": 0, + "content": "As another point-based NeRF model, PointNeRF supports fine-tuning but cannot leverage infinitesimal surface transformation fine-tuning to further optimize the performance. On the other hand, mesh-based NeRF editing models [30, 44, 51] do not support fine-tuning well: With deforming the space instead of the scene, these models' rendering scheme has highly changed. In their rendering process, a ray may go through a long, irregular way to reach the scene's surface. As the modeled space deformation might not be precise, it could be hard to tune the irregular space well, and even hurt other parts of the trained NeRF model. Such issues occur especially for some spaces with non-uniform density, since most of their model components (e.g., positional encodings, voxels) are not designed to deal with non-uniform spaces. Among all these methods, only our NeuralEditor has complete support for fine-tuning." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.559, + 0.203, + 0.575 + ], + "angle": 0, + "content": "5. Experiment" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.584, + 0.47, + 0.78 + ], + "angle": 0, + "content": "Point Cloud Generation. The underlying point cloud is fundamental to all editing tasks. Fig. 6 first provides a qualitative comparison of point clouds generated by our NeuralEditor and PointNeRF [43] on NeRF Synthetic [27]. Ours are much more precise with sharper details, e.g., the mayonnaise on the Hotdog's sausage, the uneven texture on the Chair's cushion, the edge of the Mic's stand, and the Lego brick's studs. By contrast, PointNeRF's point clouds are blurred and noisy, lose most of the details, and even contain obvious shape deflects on the Hotdog's plate and Chair's backrest. This shows that while the point cloud generation task is challenging, NeuralEditor generates a superprecise point cloud which is crucial for shape editing tasks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Experimental Settings. We mainly conduct experiments based on scenes from the NeRF Synthetic (NS) dataset. NS is a widely-used NeRF benchmark constructed from Blender [7] scenes. Due to the lack of publicly available benchmarks for shape editing, we use Blender to construct a reproducible benchmark, including the ground truth of edited scenes for evaluation and fine-tuning. Our shape editing tasks cover all eight scenes in NS, while prior" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.089, + 0.882, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.366, + 0.892, + 0.407 + ], + "angle": 0, + "content": "Figure 6. NeuralEditor generates much more precise point clouds than PointNeRF [43] in the four scenes of NeRF Synthetic [27]. The points are colored with their normal vectors." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.413, + 0.892, + 0.475 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.476, + 0.892, + 0.504 + ], + "angle": 0, + "content": "Figure 7. With too coarse cages, DeformingNeRF [44] is unable to perform the deformation faithfully, leading to poor results." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.508, + 0.892, + 0.598 + ], + "angle": 0, + "content": "work [30,44,51] only picks a few scenes. The provided images for NS scenes are with opacity, and there is no requirement for the background color. We evaluate and visualize the results on a black background, for better contrast and clearer detail visualization. In the supplementary, we show the results on a white background with same conclusions." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.599, + 0.892, + 0.628 + ], + "angle": 0, + "content": "Shape Editing Tasks. We evaluate our model on two types of shape editing tasks, as shown in Fig. 1:" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.63, + 0.893, + 0.84 + ], + "angle": 0, + "content": "(I) Shape (Mesh) Deformation Task. We consider the shape deformation task as in [30, 44, 51]: deform the shape of a scene in a human-guided way. To construct our deformation tasks from NS and obtain the ground truth, we apply the shape deformation simultaneously to the scene and our point cloud within the provided Blender file. We perform both zero-shot inference and fine-tuning, and compare our rendering results with the ground truth. As demonstrated in Figs. 7 and 8, our deformation tasks are much more precise and aggressive, compared with those in previous work [30, 44, 51]. In the supplementary, we also design a non-continuous deformation task and deformation tasks on the real-world dataset Tanks and Temples [17] (with zero-shot inference only, as there is no ground truth available)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.841, + 0.894, + 0.902 + ], + "angle": 0, + "content": "(II) Scene Morphing Task. We address a more challenging shape editing task that has not been investigated in prior NeRF editing work: the scene morphing task. Given two scenes \\( A \\) and \\( B \\), we should construct a path to gradually" + }, + { + "type": "list", + "bbox": [ + 0.498, + 0.63, + 0.894, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12444" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.08, + 0.09, + 0.891, + 0.17 + ], + "angle": 0, + "content": "
ModelZero-Shot Inference, PSNR ↑Fine-Tune for 1 Epoch, PSNR ↑
ChairHotdogLegoDrumsFicusMaterialsMicShipChairHotdogLegoDrumsFicusMaterialsMicShip
DeformingNeRF [44]18.84-13.10-------------
PointNeRF [43]22.2125.9524.5621.0024.2421.2126.7721.1930.1136.0831.4527.1631.4827.5534.3428.90
Naive Plotting24.9127.0125.6421.2926.2221.6527.6322.2932.0136.3831.7228.0933.2130.3135.1530.01
NeuralEditor w/o IST24.9227.0225.6521.2926.2421.6427.6422.2832.2436.6932.7928.3033.3430.4035.2830.08
NeuralEditor (Ours)25.8527.4927.4621.8427.1923.1827.7524.1632.5337.2232.9528.3533.5330.8235.4630.44
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.171, + 0.893, + 0.226 + ], + "angle": 0, + "content": "Table 1. NeuralEditor significantly and consistently outperforms PointNeRF and Naive Plotting on all deformed scenes of NeRF Synthetic in peak signal-to-noise ratio (PSNR), in both zero-shot inference and fine-tuning settings. Our infinitesimal surface transformation (IST) effectively improves the results by correcting the view-dependent colors. With the precise point cloud generated by NeuralEditor, even Naive Plotting consistently outperforms PointNeRF. Comparison results under other metrics are in the supplementary." + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.227, + 0.891, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.432, + 0.893, + 0.459 + ], + "angle": 0, + "content": "Figure 8. NeuralEditor produces superior rendering results to PointNeRF, with significantly fewer artifacts in zero-shot inference. Finetuning further improves the consistency of rendering with the ambient environment. We use a black background for better visualization." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.463, + 0.47, + 0.643 + ], + "angle": 0, + "content": "change one scene to the other, and the intermediate scenes should have reasonable appearances. We are required to render all intermediate scenes. For this task, we use the point cloud diffusion model [24] to generate intermediate point clouds with latent space interpolation between \\( A \\) and \\( B \\) as in [24], and we introduce a K-D tree-based [4] matching algorithm to match the adjacent points to fix the indices of the intermediate scenes. To render an intermediate scene, we apply the shape transformation to the NeuralEditor models trained for scenes \\( A \\) and \\( B \\), and then interpolate the rendering features to obtain the features of the intermediate scene for rendering." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.47, + 0.765 + ], + "angle": 0, + "content": "NeuralEditor Variants. (1) Full NeuralEditor: Our complete model with all components. (2) NeuralEditor without infinitesimal surface transformation (IST): We remove the maintenance and optimization of infinitesimal surface transformation in scene editing. This key variant of NeuralEditor enables us to evaluate the importance of IST as well as other components of our model. The full ablation study of NeuralEditor is in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Baselines. We compare NeuralEditor against different types of baselines as follows. (1) Naive Plotting: We use the point cloud generated by NeuralEditor, computing each point's opacity and view-dependent colors with their point features. We render the scene by directly plotting/projecting the point cloud to the camera plane. (2) PointNeRF [43]: For the shape deformation task, we apply the same deformation to the point clouds generated by PointNeRF. For the scene morphing task, we apply the matching algorithm to" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.463, + 0.893, + 0.643 + ], + "angle": 0, + "content": "the point clouds generated by PointNeRF and the same intermediate point clouds generated by the point cloud diffusion model [24] for fairness. (3) DeformingNeRF [44]: DeformingNeRF is not compatible with the scene morphing task. For the shape deformation task, we perform the same deformation on vertices of given cages. Note that DeformingNeRF only released trained models for Lego and Chair, so we can only evaluate it on these two scenes. While other models [30,51] support NeRF-based shape deformation via cages or exported meshes, we were unable to use them as baselines - they did not provide executable code nor their deformed scenes for us to evaluate on their tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Shape (Mesh) Deformation Results. The qualitative comparison is shown in Fig. 8. Both PointNeRF and Naive Plotting have many artifacts, like blur, wrong color, black or white shadows, noise, etc., whereas our powerful NeuralEditor produces clean and realistic rendering results. After fine-tuning, NeuralEditor shows a significant improvement with better rendering results than PointNeRF, indicating that NeuralEditor is able to achieve higher consistency with the surrounding ambient environment. Notably, in the Materials scene (the 6th scene from left), only our NeuralEditor generates reasonable reflection, while both baselines show blurry and visually messy results. Also in the Drums scene, the bottom face of the gong is not visible in any of the training views in the original scene, so all models render poor results in zero-shot inference. However, after only fast fine-tuning, NeuralEditor is able to precisely model the previously unknown surface and generate a sub-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12445" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.155, + 0.089, + 0.823, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.437, + 0.895, + 0.478 + ], + "angle": 0, + "content": "Figure 9. Our NeuralEditor produces smooth morphing results between Chair, Hotdog, Lego, and Mic in the NeRF Synthetic dataset, while PointNeRF produces results with blurry textures, black shadows, and gloomy, non-smooth colors. The rendering results in the looped morphing process are arranged in the shape of the numerical digit \"3,\" indicated by the dividing lines and arrows." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.482, + 0.47, + 0.559 + ], + "angle": 0, + "content": "stantially better result than PointNeRF, highlighting NeuralEditor's strength in fast-fitting. All these results demonstrate that NeuralEditor can handle various visual effects and make them consistent in the deformed scene. We provide the figure with higher resolution in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.562, + 0.47, + 0.729 + ], + "angle": 0, + "content": "The quantitative comparison is summarized in Table 1. We observe that: (1) Our NeuralEditor consistently outperforms all the baselines and variants for both zero-shot inference and fine-tuning settings. (2) With the precise point cloud generated by NeuralEditor, the Naive Plotting baseline even consistently outperforms PointNeRF. (3) Our 'w/o IST' variant has a comparable performance to Naive Plotting with the same point cloud and features in the zero-shot inference setting, but after fine-tuning its performance is significantly higher than Naive Plotting, validating the capability of NeuralEditor in NeRF modeling." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.731, + 0.47, + 0.868 + ], + "angle": 0, + "content": "Notably, DeformingNeRF [44] performs poorly in our benchmark with significantly lower metric values. As shown in Fig. 7, the cages provided by DeformingNeRF are too coarse, and cannot even cover the whole scene. Therefore, DeformingNeRF cannot faithfully perform the precise deformation in our benchmark, leading to poor rendering results. On the contrary, both PointNeRF and our NeuralEditor at least faithfully perform the deformation, showing that point cloud is necessary for precise shape editing." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Scene Morphing Results. The morphing results between 4 NeRF Synthetic scenes are shown in Fig. 9. The" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.483, + 0.892, + 0.634 + ], + "angle": 0, + "content": "morphing process starts from Chair, morphs to Hotdog, Lego, Mic, and at last turns back to Chair. NeuralEditor produces smooth rendering results on the point cloud diffusion-guided [24] intermediate scenes, mixing the textures of the two scenes in a reasonable way. In comparison, the rendering results produced by PointNeRF are unsatisfactory, with blurry textures, black shadows, and gloomy, non-smooth colors. These results show that our NeuralEditor can render challenging intermediate morphing scenes and achieve decent results with only the input of moved points." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.645, + 0.619, + 0.66 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.667, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This paper proposes NeuralEditor, a point cloud-guided NeRF model that supports general shape editing tasks by manipulating the underlying point clouds. Empirical evaluation shows NeuralEditor to produce rendering results of much higher quality than baselines in a zero-shot inference manner, further significantly improving after fast finetuning. NeuralEditor even supports smooth scene morphing between multiple scenes, which is difficult for prior work. We hope that our work can inspire more research on point cloud-guided NeRFs and 3D shape and scene editing tasks. Acknowledgement. This work was supported in part by NSF Grant 2106825, NIFA Award 2020-67021-32799, the Jump ARCHES endowment, the NCSA Fellows program, the IBM-Illinois Discovery Accelerator Institute, the Illinois-Inspor Partnership, and the Amazon Research Award. This work used NVIDIA GPUs at NCSA Delta through allocation CIS220014 from the ACCESS program. We thank the authors of NeRF [27] for their help in processing Blender files of the NS dataset." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12446" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.17 + ], + "angle": 0, + "content": "[1] Henrik Aanaes, Rasmus Ramsbøl Jensen, George Vogiatzis, Engin Tola, and Anders Bjorholm Dahl. Large-scale data for multiple-view stereopsis. International Journal of Computer Vision, 120(2):153-168, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.171, + 0.47, + 0.213 + ], + "angle": 0, + "content": "[2] Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas J. Guibas. Learning representations and generative models for 3D point clouds. In ICML, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.213, + 0.471, + 0.269 + ], + "angle": 0, + "content": "[3] Jonathan T. Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P. Srinivasan. Mip-NeRF: A multiscale representation for antiailiasing neural radiance fields. In ICCV, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.269, + 0.47, + 0.311 + ], + "angle": 0, + "content": "[4] Jon Louis Bentley. Multidimensional binary search trees used for associative searching. Commun. ACM, 18(9):509-517, 1975. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.311, + 0.47, + 0.367 + ], + "angle": 0, + "content": "[5] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. MVSNeRF: Fast generalizable radiance field reconstruction from multi-view stereo. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.367, + 0.47, + 0.396 + ], + "angle": 0, + "content": "[6] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.396, + 0.471, + 0.437 + ], + "angle": 0, + "content": "[7] Blender Online Community. *Blender - a 3D modelling and rendering package*. Blender Foundation, Stichting Blender Foundation, Amsterdam, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.438, + 0.47, + 0.465 + ], + "angle": 0, + "content": "[8] Frank Dellaert and Lin Yen-Chen. Neural volume rendering: NeRF and beyond. arXiv:2101.05204, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.466, + 0.47, + 0.508 + ], + "angle": 0, + "content": "[9] Kyle Gao, Yina Gao, Hongjie He, Denning Lu, Linlin Xu, and Jonathan Li. NeRF: Neural radiance field in 3D vision, a comprehensive review. arXiv:2101.05204, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.509, + 0.47, + 0.55 + ], + "angle": 0, + "content": "[10] Yuan-Chen Guo, Di Kang, Linchao Bao, Yu He, and Song-Hai Zhang. NeRFReN: Neural radiance fields with reflections. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.55, + 0.47, + 0.592 + ], + "angle": 0, + "content": "[11] Peter Hedman, Tobias Ritschel, George Drettakis, and Gabriel Brostow. Scalable inside-out image-based rendering. ACM Trans. Graph., 35(6):231:1-231:11, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.592, + 0.47, + 0.634 + ], + "angle": 0, + "content": "[12] Po-Han Huang, Kevin Matzen, Johannes Kopf, Narendra Ahuja, and Jia-Bin Huang. DeepMVS: Learning multi-view stereopsis. In CVPR, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.634, + 0.47, + 0.676 + ], + "angle": 0, + "content": "[13] Alec Jacobson, Ilya Baran, Ladislav Kavan, Jovan Popovic, and Olga Sorkine. Fast automatic skinning transformations. ACM Trans. Graph., 31(4), 2012. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.677, + 0.47, + 0.732 + ], + "angle": 0, + "content": "[14] Tomas Jakab, Richard Tucker, Ameesh Makadia, Jiajun Wu, Noah Snavely, and Angjoo Kanazawa. Keypointdeformer: Unsupervised 3D keypoint discovery for shape control. In CVPR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.732, + 0.47, + 0.774 + ], + "angle": 0, + "content": "[15] Mengqi Ji, Juergen Gall, Haitian Zheng, Yebin Liu, and Lu Fang. Surfacenet: An end-to-end 3D neural network for multiview stereopsis. In ICCV, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.774, + 0.47, + 0.816 + ], + "angle": 0, + "content": "[16] Tao Ju, Scott Schaefer, and Joe Warren. Mean value coordinates for closed triangular meshes. ACM Trans. Graph., 24(3):561-566, 2005. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.816, + 0.47, + 0.859 + ], + "angle": 0, + "content": "[17] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and Temples: Benchmarking large-scale scene reconstruction. ACM Trans. Graph., 36(4), 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[18] Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitzmann. Decomposing NeRF for editing via feature field distillation. In NeurlPS, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "[19] Marc Levoy and Pat Hanrahan. Light field rendering. In SIGGRAPH, 1996. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.121, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[20] Fayao Liu, Chunhua Shen, Guosheng Lin, and Ian Reid. Learning depth from single monocular images using deep convolutional neural fields. IEEE Transactions on Pattern Analysis and Machine Intelligence, 38(10):2024-2039, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.191, + 0.892, + 0.234 + ], + "angle": 0, + "content": "[21] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. In NeurIPS, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.235, + 0.892, + 0.279 + ], + "angle": 0, + "content": "[22] Steven Liu, Xiuming Zhang, Zhoutong Zhang, Richard Zhang, Junyan Zhu, and Bryan C. Russell. Editing conditional radiance fields. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.28, + 0.892, + 0.322 + ], + "angle": 0, + "content": "[23] William E. Lorensen and Harvey E. Cline. Marching cubes: A high resolution 3D surface construction algorithm. In SIGGRAPH, 1987. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.323, + 0.892, + 0.352 + ], + "angle": 0, + "content": "[24] Shitong Luo and Wei Hu. Diffusion probabilistic models for 3D point cloud generation. In CVPR, 2021. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.353, + 0.892, + 0.396 + ], + "angle": 0, + "content": "[25] Bruce Merry, Patrick Marais, and James Gain. Animation space: A truly linear framework for character animation. ACM Trans. Graph., 25(4):1400-1423, 2006. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.396, + 0.892, + 0.452 + ], + "angle": 0, + "content": "[26] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3D reconstruction in function space. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.453, + 0.892, + 0.511 + ], + "angle": 0, + "content": "[27] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2, 4, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.511, + 0.892, + 0.568 + ], + "angle": 0, + "content": "[28] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Differentiable volumetric rendering: Learning implicit 3D representations without 3D supervision. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.569, + 0.892, + 0.611 + ], + "angle": 0, + "content": "[29] Jesús Nieto and Toni Susin. Cage based deformations: A survey. Lecture Notes in Computational Vision and Biomechanics, 7:75-99, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.612, + 0.892, + 0.683 + ], + "angle": 0, + "content": "[30] Yicong Peng, Yichao Yan, Shengqi Liu, Yuhao Cheng, Shanyan Guan, Bowen Pan, Guangtao Zhai, and Xiaokang Yang. CageNeRF: Cage-based neural radiance field for generalized 3D deformation and animation. In NeurIPS, 2022. 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.684, + 0.892, + 0.712 + ], + "angle": 0, + "content": "[31] Bui Tuong Phong. Illumination for computer generated pictures. Commun. ACM, 18(6):311-317, 1975. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.712, + 0.892, + 0.756 + ], + "angle": 0, + "content": "[32] Yi-Ling Qiao, Alexander Gao, and Ming C. Lin. NeuPhysics: Editable neural geometry and physics from monocular videos. In NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.756, + 0.892, + 0.813 + ], + "angle": 0, + "content": "[33] Sara Fridovich-Keil and Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In CVPR, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.814, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[34] Noah Snavely, Steven M. Seitz, and Richard Szeliski. Photo tourism: Exploring photo collections in 3D. ACM Trans. Graph., 25(3):835-846, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.857, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[35] Jiaxiang Tang, Xiaokang Chen, Jingbo Wang, and Gang Zeng. Compressible-composable NeRF via rank-residual decomposition. In NeurIPS, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12447" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.146 + ], + "angle": 0, + "content": "[36] Jean-Marc Thiery, Julien Tierny, and Tamy Boubekeur. Jacobians and hessian of mean value coordinates for closed triangular meshes. The Visual Computer, 30(9):981-995, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.15, + 0.47, + 0.207 + ], + "angle": 0, + "content": "[37] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T. Barron, and Pratul P. Srinivasan. Ref-NeRF: Structured view-dependent appearance for neural radiance fields. In CVPR, 2022. 1, 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.208, + 0.47, + 0.251 + ], + "angle": 0, + "content": "[38] Nanyang Wang, Yinda Zhang, Zhuwen Li, Yanwei Fu, Wei Liu, and Yu-Gang Jiang. Pixel2Mesh: Generating 3D mesh models from single RGB images. In ECCV, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.252, + 0.47, + 0.32 + ], + "angle": 0, + "content": "[39] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul Srinivasan, Howard Zhou, Jonathan T. Barron, Ricardo MartinBrualla, Noah Snavely, and Thomas Funkhouser. IBRNet: Learning multi-view image-based rendering. In CVPR, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.323, + 0.47, + 0.377 + ], + "angle": 0, + "content": "[40] Suttisak Wizadwongsa, Pakkapon Phongthawee, Jiraphon Yenphraphai, and Supasorn Suwajanakorn. NeX: Real-time view synthesis with neural basis expansion. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.381, + 0.47, + 0.436 + ], + "angle": 0, + "content": "[41] Liwen Wu, Jae Yong Lee, Anand Bhattachad, Yu-Xiong Wang, and David Forsyth. DIVeR: Real-time and accurate neural radiance fields with deterministic integration for volume rendering. In CVPR, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.438, + 0.47, + 0.479 + ], + "angle": 0, + "content": "[42] Dejia Xu, Peihao Wang, Yifan Jiang, Zhiwen Fan, and Zhangyang Wang. Signal processing for implicit neural representations. In NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.482, + 0.47, + 0.524 + ], + "angle": 0, + "content": "[43] Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-NeRF: Point-based neural radiance fields. In CVPR, 2021. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.526, + 0.47, + 0.553 + ], + "angle": 0, + "content": "[44] Tianhan Xu and Tatsuya Harada. Deforming radiance fields with cages. In ECCV, 2022. 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.556, + 0.47, + 0.611 + ], + "angle": 0, + "content": "[45] Bangbang Yang, Yinda Zhang, Yinghao Xu, Yijin Li, Han Zhou, Hujun Bao, Guofeng Zhang, and Zhaopeng Cui. Learning object-compositional neural radiance field for editing scene rendering. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.613, + 0.47, + 0.654 + ], + "angle": 0, + "content": "[46] Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. MVSNet: Depth inference for unstructured multiview stereo. In ECCV, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.657, + 0.47, + 0.699 + ], + "angle": 0, + "content": "[47] Wang Yifan, Noam Aigerman, Vladimir G Kim, Siddhartha Chaudhuri, and Olga Sorkine-Hornung. Neural cages for detail-preserving 3D deformations. In CVPR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.701, + 0.47, + 0.742 + ], + "angle": 0, + "content": "[48] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. PlenOctrees for real-time rendering of neural radiance fields. In ICCV, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.745, + 0.469, + 0.785 + ], + "angle": 0, + "content": "[49] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelNeRF: Neural radiance fields from one or few images. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.788, + 0.47, + 0.843 + ], + "angle": 0, + "content": "[50] Yu-Jie Yuan, Yu-Kun Lai, Tong Wu, Lin Gao, and Ligang Liu. A revisit of shape editing techniques: From the geometric to the neural viewpoint. Journal of Computer Science and Technology, 36(3):520-554, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.47, + 0.899 + ], + "angle": 0, + "content": "[51] Yu-Jie Yuan, Yang tian Sun, Yu-Kun Lai, Yuewen Ma, Rongfei Jia, and Lin Gao. NeRF-Editing: Geometry editing of neural radiance fields. In CVPR, 2022. 2, 3, 5, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.135 + ], + "angle": 0, + "content": "[52] Yuzhe Zhang, Jianmin Zheng, and Yiyu Cai. Proxy-driven free-form deformation by topology-adjustable control lattice. Computers & Graphics, 89:167-177, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.893, + 0.176 + ], + "angle": 0, + "content": "[53] Qian-Yi Zhou, Jaesik Park, and Vladlen Koltun. Open3D: A modern library for 3D data processing. arXiv:1801.09847, 2018. 4" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.176 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12448" + } + ] +] \ No newline at end of file diff --git a/2023/NeuralEditor_ Editing Neural Radiance Fields via Manipulating Point Clouds/70b7a137-ced8-4460-839e-a2dd484a031c_origin.pdf b/2023/NeuralEditor_ Editing Neural Radiance Fields via Manipulating Point Clouds/70b7a137-ced8-4460-839e-a2dd484a031c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..821039bdd82ecfe44f4ac7bc8b863fefeb034530 --- /dev/null +++ b/2023/NeuralEditor_ Editing Neural Radiance Fields via Manipulating Point Clouds/70b7a137-ced8-4460-839e-a2dd484a031c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:aad0c454ac8b04f5ff494e3ba850f3e310f7dc84b1ddb8f8847a298a54bb091c +size 8217762 diff --git a/2023/NeuralEditor_ Editing Neural Radiance Fields via Manipulating Point Clouds/full.md b/2023/NeuralEditor_ Editing Neural Radiance Fields via Manipulating Point Clouds/full.md new file mode 100644 index 0000000000000000000000000000000000000000..59e28563f1ee50872b156a8578003e978271413d --- /dev/null +++ b/2023/NeuralEditor_ Editing Neural Radiance Fields via Manipulating Point Clouds/full.md @@ -0,0 +1,274 @@ +# NeuralEditor: Editing Neural Radiance Fields via Manipulating Point Clouds + +Jun-Kun Chen $^{1\dagger}$ + +Jipeng Lyu $^{2\dagger}$ + +Yu-Xiong Wang + +1University of Illinois at Urbana-Champaign + +$^{2}$ Peking University $\dagger$ Equal Contribution + +{junkun3，yxw}@illinois.edu lvjipeng@pku.edu.cn + +![](images/4589dde9f2e2500c54fe59004e1e22d8c9108ed27ca90fd9432a6d04afdc8926.jpg) +Figure 1. Our NeuralEditor offers native support for general and flexible shape editing of neural radiance fields via manipulating point clouds. By generating a precise point cloud of the scene with a novel point cloud-guided NeRF model, our NeuralEditor produces high-fidelity rendering results in both shape deformation and more challenging scene morphing tasks. + +![](images/1b9a1be3d3a7295491938e1aeb98a2f36f0f998e2d7cf65aba738962a02a830c.jpg) + +# Abstract + +This paper proposes NeuralEditor that enables neural radiance fields (NeRFs) natively editable for general shape editing tasks. Despite their impressive results on novel-view synthesis, it remains a fundamental challenge for NeRFs to edit the shape of the scene. Our key insight is to exploit the explicit point cloud representation as the underlying structure to construct NeRFs, inspired by the intuitive interpretation of NeRF rendering as a process that projects or "plots" the associated 3D point cloud to a 2D image plane. To this end, NeuralEditor introduces a novel rendering scheme based on deterministic integration within K-D tree-guided density-adaptive voxels, which produces both high-quality rendering results and precise point clouds through optimization. NeuralEditor then performs shape editing via mapping associated points between point clouds. Extensive evaluation shows that NeuralEditor achieves state-of-the-art performance in both shape deformation and scene morphing tasks. Notably, NeuralEditor supports both zero-shot inference and further fine-tuning over the edited scene. Our code, benchmark, and demo video are available at immortalco.github.io/NeuralEditor. + +# 1. Introduction + +Perhaps the most memorable shot of the film Transformers, Optimus Prime is seamlessly transformed between a + +humanoid and a Peterbilt truck – such free-form editing of 3D objects and scenes is a fundamental task in 3D computer vision and computer graphics, directly impacting applications such as visual simulation, movie, and game industries. In these applications, often we are required to manipulate a scene or objects in the scene by editing or modifying its shape, color, light condition, etc., and generate visually-faithful rendering results on the edited scene efficiently. Among the various editing operations, shape editing has received continued attention but remains challenging, where the scene is deformed in a human-guided way, while all of its visual attributes (e.g., shape, color, brightness, and light condition) are supposed to be natural and consistent with the ambient environment. + +State-of-the-art rendering models are based on implicit neural representations, as exemplified by neural radiance field (NeRF) [27] and its variants [3, 33, 37, 39, 48]. Despite their impressive novel-view synthesis results, most of the NeRF models substantially lack the ability for users to adjust, edit, or modify the shape of scene objects. On the other hand, shape editing operations can be natively applied to explicit 3D representations such as point clouds and meshes. + +Inspired by this, we propose NeuralEditor – a general and flexible approach to editing neural radiance fields via manipulating point clouds (Fig. 1). Our key insight is to benefit from the best of both worlds: the superiority in rendering performance from implicit neural representation + +combined with the ease of editing from explicit point cloud representation. NeuralEditor enables us to perform a wide spectrum of shape editing operations in a consistent way. + +Such introduction of point clouds into NeRF for general shape editing is rooted in our interpretation of NeRF rendering as a process that projects or "plots" the associated 3D point cloud to a 2D image plane. Conceptually, with a dense enough point cloud where each point has an opacity and its color is defined as a function of viewing direction, directly plotting the point cloud would achieve similar visual effects (i.e., transparency and view-dependent colors) that are rendered by NeRF. This intrinsic integration between NeRF and point clouds underscores the advantage of our NeuralEditor over existing mesh-based NeRF editing methods such as NeRF-Editing [51], Deforming-NeRF [44], and CageNeRF [30], where the process of constructing and optimizing the mesh is separated from the NeRF modeling, making them time-consuming. More importantly, with the point cloud constructed for a scene, the shape editing can be natively defined as and easily solved by just moving each point into the new, edited position and re-plotting the point cloud. Therefore, our approach supports more general scene editing operations which are difficult to achieve via mesh-guided space deformation. + +The key component in our NeuralEditor lies in a point cloud-guided NeRF model that natively supports general shape editing operations. While the recent method Point-NeRF [43] has demonstrated improved novel-view synthesis capability based on point clouds, it is not supportive to shape editing. Our idea then is to exploit the underlying point cloud in ways of not only optimizing its structure and features (e.g., adaptive voxels) for rendering, but also extracting additional useful attributes (e.g., normal vectors) to guide the editing process. To this end, we introduce K-D trees [4] to construct density-adaptive voxels for efficient and stable rendering, together with a novel deterministic integration strategy. Moreover, we model the color with the Phong reflection [31] to decompose the specular color and better represent the scene geometry. + +With a much more precise point cloud attributed to these improvements, our NeuralEditor achieves high-fidelity rendering results on deformed scenes compared with prior work as shown in Fig. 1, even in a zero-shot inference manner without additional training. Through fast fine-tuning, the visual quality of the deformed scene is further enhanced, almost perfectly consistent with the surrounding light condition. In addition, under the guidance of a point cloud diffusion model [24], NeuralEditor can be naturally extended for smooth scene morphing across multiple scenes, which is difficult for existing NeRF editing work. + +Our contributions are four-fold. (1) We introduce NeuralEditor, a flexible and versatile approach that makes neural radiance fields editable through manipulating point + +clouds. (2) We propose a point cloud-guided NeRF model based on K-D trees and deterministic integration, which produces precise point clouds and supports general scene editing. (3) Due to the lack of publicly available benchmarks for shape editing, we construct and release a reproducible benchmark that promotes future research on shape editing. (4) We investigate a wide range of shape editing tasks, covering both shape deformation (as studied in existing NeRF editing work) and challenging scene morphing (a novel task addressed here). NeuralEditor achieves state-of-the-art performance on all shape editing tasks in a unified framework, without extra information or supervision. + +# 2. Related Work + +Neural Scene Representation. Traditional methods model scenes with explicit [2, 11, 15, 20, 34, 38] or implicit [6, 19, 26, 28, 40] 3D geometric or shape representations. Initiated by NeRF [27], leveraging implicit neural networks to represent scenes and perform novel-view synthesis has become a fast-developing field in 3D vision [8,9]. While most of the follow-up work focuses on improving aspects such as the rendering realism [3, 10, 21, 37], efficiency [33, 41, 48], and cross-scene generalization [5, 39, 43, 49], the scene editing capability is substantially missing in the NeRF family which we address in this paper. In addition, we exploit K-D tree-guided point clouds as the underlying structure, different from other NeRF variants based on octrees [21, 33, 48] or plain voxels [41]. + +Point-Based NeRFs. Recently, using point clouds to build a NeRF model has shown better encoding of scene shape and improved rendering performance, as represented by PointNeRF [43]. PointNeRF proposes a point initialization network to produce the initial point cloud together with the point features, which is further optimized by a pruning and growing strategy. While both PointNeRF and our NeuralEditor employ point clouds as the underlying structure, NeuralEditor better exploits useful information within the point clouds: PointNeRF only directly uses the locations of points; by contrast, NeuralEditor considers the point cloud more as a geometrical shape and extracts relevant information like normal vectors, which plays an important role in rendering and shape editing. Importantly, our approach is designed to support scene editing, in contrast to PointNeRF. + +Scene Editing via NeRFs. Different types of scene editing have been studied under NeRFs. EditNeRF [22], ObjectNeRF [45], and DistillNeRF [18] perform simple shape and color editing for objects specified with human-input scribble, pixel, segment, language, etc. NeuPhysics [32] edits a dynamic scene via physics parameters. CCNeRF [35] proposes an explicit NeRF representation with tensor rank decomposition to support scene composition. INSP-Net [42] considers filter editing like denoising. Such work cannot address 3D shape editing and only sup + +ports simple editing operations, like object selection, similarity transformation, or limited shape deformation. + +3D Shape Editing. Traditional representation methods support keypoint-based shape editing [13, 14, 16, 25, 36, 47, 52] with meshes [29, 50], which cannot be directly applied to implicit representations used by NeRF. Existing NeRF editing work primarily studies a particular shape editing task, mesh deformation, and addresses it in a common paradigm [30, 44, 51]: A mesh of the scene is first constructed by either exporting it from a trained NeRF with the Marching Cubes algorithm [23], or optimizing close-to-surface cages along with training. After the user deforms the mesh, the deformed scene is rendered by deforming the space and bending the viewing rays in the original scene with the trained NeRF. Doing so requires extra efforts to convert implicit scene representation to explicit mesh, which might not be precise enough, and only supports continuous shape editing that can be converted to space deformation. On the contrary, our NeuralEditor directly maintains and utilizes alternative explicit scene representation – the point cloud which is intrinsically integrated with NeRF, making NeuralEditor require no extra efforts and support more general shape editing tasks like scene morphing. NeuralEditor supports both zero-shot inference and further fine-tuning over the edited scene, while prior work cannot. + +# 3. NeuralEditor: Point Cloud-Guided NeRF + +We propose a novel point cloud-guided NeRF model, NeuralEditor - it not only achieves realistic rendering results in the novel-view synthesis task, but also produces a point cloud that precisely describes the shape of the scene, thus facilitating general shape editing tasks. As illustrated in Fig. 2, we leverage the K-D trees [4] to construct density-adaptive voxels (which also naturally enable us to skip empty spaces), and introduce deterministic spline integration for rendering. We use the Phong reflection to model the color along with the normal vectors obtained from the underlying point cloud. With our enhanced point cloud optimization, NeuralEditor obtains much more precise underlying point clouds, compared with noisy and imprecise outputs of state-of-the-art PointNeRF [43] (as shown in Sec. 5). + +# 3.1. K-D Tree-Guided Voxels + +To render with points, we construct multi-scale density-adaptive voxels based on K-D trees [4], namely, $K - D$ voxels. K-D trees are a data structure constructed on $K$ -dimensional points, where $K = 3$ for 3D points. As a special decision tree, K-D tree's each node divides the point set into two equal-sized parts with axis-parallel criterion. + +For each K-D tree's node, we compute its bounding box by taking the minimum and maximum $x, y, z$ coordinates in its subtree and with proper padding margins. As we divide the points in a top-down manner in one of the $x, y, z$ direc + +![](images/bd5934b42f8753048540e844e4d330c236474cd2f2bb976fc87fea27eb17524a.jpg) +Figure 2. Our K-D Voxels. Column 1: original point clouds constructed from two scenes, colored with normal vector directions. Column 2: upper-level voxels which coarsely represent the shape. Column 3: lower-level voxels which tightly cover the shape. Right: Visualization of K-D voxels on a 2D point cloud. Each color represents boxes of nodes on each level of the K-D tree. Lower-level boxes containing fewer points cover the shape more tightly, and vice versa for higher-level boxes. + +tions, in each layer of the K-D tree, different nodes' bounding boxes are mutually exclusive. Therefore, the bounding boxes can be regarded as voxels. As boxes in the upper layers contain more points (larger voxels), while those in the lower layers contain fewer points (smaller voxels), we naively obtain a multi-scale voxel construction from one K-D tree. As shown in Fig. 2, voxels from the large to small scales represent the shape of the scene from coarse to fine. + +# 3.2. Rendering Over K-D Voxels + +We now introduce a rendering scheme that exploits K-D voxels to perform all the sub-procedures associated with rendering in a unified way. This scheme enables us to render more naturally, efficiently, and even deterministically, meanwhile it also simplifies some widely-adopted design choices in conventional NeRF rendering. + +Skipping Empty Spaces. In NeRF rendering, we are supposed to focus only on the surface of scene objects. As shown in Fig. 2, all our K-D voxels are produced to stick to the surface of objects, which the point cloud is constructed to describe. Such a property allows us to avoid explicitly "skipping" empty spaces, which often requires extra consideration in most NeRF models - only considering the space inside a voxel automatically focuses on the surface; as the depth of the voxel's node goes deeper, it becomes closer to the surface. Moreover, during the construction of the K-D tree, the points at each node are divided within its sub-nodes, and the node's voxel fully covers all its sub-nodes. This further provides us with a native top-down recursive procedure to locate the voxels intersected with the querying ray: We start from the root node, and recurse on the sub-nodes until we (1) reach a pre-set node depth (or equivalently, a pre-set voxel size) and then query within the associated voxel, or (2) stop recursion on non-intersected nodes. + +Density-Adaptive Rendering. An important design in NeRFs is the coarse-to-fine strategy for density-adaptive rendering, so that more points are sampled in high-volume density areas. Our K-D voxels natively support such a design without additional bells and whistles. This is because + +![](images/8286100af116b003cc588dc64b936796675e5dfe7fe4d68a7a8d758cbc05da6a.jpg) +Figure 3. Our NeuralEditor architecture. We propose deterministic spline integration for KNN-based point features over each K-D tree-guided density-adaptive voxel, and model the color via Phong reflection with normal vectors estimated from the point cloud's shape. + +voxels in the same K-D tree layer contain the same number of points. As the point density can be regarded as an approximation of volume density, all such voxels have the same density. Therefore, we directly use K-D voxels to guide the density-adaptive rendering. Specifically, we conduct the rendering process at the voxels of some bottom layers in the K-D tree. For each querying ray, we use the aforementioned recursive procedure to locate the minimal intersected voxels that are deep enough. Here "minimal" means that the ray intersects with the node's voxel, but does not intersect with any sub-node's voxel. These intersected voxels divide the querying ray into several segments (Fig. 3). The segments covered by a voxel are close to the surface and used for rendering, while those not covered are in empty spaces. + +Deterministic Spline Integration. DIVER [41] shows that deterministic integration outperforms stochastic integration in NeRF rendering. So we perform a deterministic integration to obtain the segment's feature within each voxel. Since we do not necessarily have points at the voxel's vertices, the trilinear interpolation used in DIVER is not feasible here. Instead, we use spline integration. For the $i$ -th intersected voxel in the ray passing order, we uniformly select points in this segment, and integrate their features to obtain the average feature $f_{i}$ of the segment of the $i$ -th voxel: + +$$ +f _ {i} = \frac {1}{r _ {i} - l _ {i}} \int_ {l _ {i}} ^ {r _ {i}} \mathbf {f e a t u r e} (o + t \cdot d) \mathrm {d} t, \tag {1} +$$ + +where $[l_i, r_i]$ is the intersection interval, and $o$ and $d$ are the source and direction of the querying ray, respectively. This average feature $f_i$ can be interpreted as the feature of a representative point $p_i$ located somewhere in the segment. + +KNN-Based Feature Aggregation. For each uniformly selected point $q$ in the segment during spline integration, we obtain its feature via weighted interpolation from the features of its $K$ nearest neighbors (KNN) in the point cloud: + +$$ +\operatorname {f e a t u r e} (q) = \sum_ {p _ {j} \in \operatorname {K N N} (q; K)} k _ {j} e _ {j}, \tag {2} +$$ + +$$ +\left\{k_{j}\right\} = \operatorname *{SoftMax}_{p_{j}\in \mathrm{KNN}(q;K)}\left(\log \gamma_{j} - \log \| q - p_{j}\|_{2}^{2}\right), +$$ + +where for each point $p_j$ in the point cloud, we parameterize its confidence $\gamma_j$ and point feature $e_j$ as in PointNeRF. + +![](images/5bf97e2419a70f2570269ef2a1b209b53783419124e6c97d6266575431f3f7fa.jpg) +Figure 4. In the two scenes of NeRF Synthetic [27], NeuralEditor optimizes the rough initial point cloud to a precise point cloud. The points are colored with their normal vectors. + +Phong Reflection-Based Color Modeling with Point Cloud Normal Vectors. To obtain the volume density $\sigma_{i}$ and color $c_{i}$ of the representative point $p_i$ in the $i$ -th voxel, we use the Phong reflection model [31]. As we have the underlying point cloud, we use Open3D [53] to estimate the normal vector for each point, and integrate these vectors over the interval to get an average normal vector $n_i$ . Such information better characterizes the shape of point clouds (scenes), which plays an important role in Phong-based rendering and also implicitly facilitates optimizing the point clouds (Sec. 3.3). Consistent with RefNeRF [37], we use multiple multilayer perceptrons (MLPs) to model other attributes, including volume density, tint, roughness, and diffuse and specular color, and use the Phong formula to calculate the final $c_{i}$ with these attributes. Finally, we aggregate $c_{i}$ of all segments on the ray to obtain the final color: + +$$ +c _ {\text {p i x e l}} = \sum_ {i \geq 1} \tau_ {i} \cdot \left(1 - \exp \left(- \left(r _ {i} - l _ {i}\right) \sigma_ {i}\right) \cdot c _ {i}, \right. +$$ + +$$ +\tau_ {i} = \exp \left(- \sum_ {i ^ {\prime} = 1} ^ {i - 1} \left(r _ {i ^ {\prime}} - l _ {i ^ {\prime}}\right) \sigma_ {i ^ {\prime}}\right). \tag {3} +$$ + +# 3.3. Point Cloud Optimization + +Point Cloud Initialization. To start training, we need a coarse initial point cloud. Consistent with PointNeRF, we + +use a point generation network, which consists of a multiview stereo (MVS) model [12] based on a 3D convolutional neural network (CNN), to generate the points' coordinates and confidence values, and another 2D CNN [46] to generate their initial features. This network was pre-trained on the DTU training dataset [1], and can generalize to other datasets and scenes. As shown in Fig. 4, the initial point cloud generated by such a network is coarse and noisy. + +Explicit Optimization via Pruning and Growing. We perform a similar pruning and growing procedure as in PointNeRF, to prune outliers with low confidence $\gamma_{j}$ and fill holes in the point clouds. We make several important modifications over PointNeRF, and integrate this procedure with our deterministic integration (details in the supplementary). + +Implicit Optimization with Normal Vectors. In addition to the explicit optimization, the point cloud is also optimized implicitly during training through the adjustment of point confidence $\gamma_{j}$ . When computing the average normal vector for rendering, we aggregate normal vectors of nearby points weighted with their distance and confidence, where the confidence of noisy or inaccurate points with potentially abnormal normal vectors is adjusted accordingly. Moreover, we apply the normal vector regularization losses from RefNeRF [37] to supervise the points' confidence w.r.t. their normal vectors. These strategies collectively provide implicit but more tailored ways to optimize the point clouds. With both explicit and implicit optimization, NeuralEditor obtains very precise point clouds (Fig. 4). + +# 4. Shape Editing with NeuralEditor + +Formulation of General Shape Editing Tasks. We define the shape editing tasks based on indexed point clouds. To this end, we first re-define an indexed point cloud $P$ as a mapping from a point index $j$ to the corresponding point $p_j$ , + +$$ +P: j \rightarrow p _ {j}, \text {w h e r e} p _ {j} \in \mathbb {R} ^ {3}, j = 1, \dots , | P |. \tag {4} +$$ + +A shape editing task is defined as another indexed point cloud $Q(P)$ + +$$ +Q (P): j \rightarrow q _ {j}, \text {w h e r e} q _ {j} \in \mathbb {R} ^ {3} \cup \{\varnothing \}, j = 1, \dots , | P |, \tag {5} +$$ + +describing a shape editing task whether the $j$ -th point moves from $p_j$ to $q_j$ or is deleted in the deformation if $q_j = \emptyset$ . With this definition, $Q(P)$ can be an arbitrary point cloud with points properly matched to points in $P$ by same indices, regardless of connectivity or continuity. + +Our formulation represents a broad range of shape editing tasks. The mesh editing tasks in NeRF-Editing [51], Deforming-NeRF [44], and CageNeRF [30] can be more simply and clearly defined here. For example, in NeRF-Editing, a mesh is exported from a general NeRF model, deformed manually, and converted to an "offset" or a continuous space deformation. We can depict such a task without + +![](images/cfa7eafab6f5def118dad4763b9eb0193fc659617d8d194f6fa1ca2f094ba3f8.jpg) +Figure 5. Infinitesimal surface transformation (IST). (a) As the view-dependent colors are modeled as absolute viewing directions, they (solid arrows at the right) are different from the correct colors (dashed arrows at the right) after deformation. We solve this by (b) constructing a local coordinate system near the $j$ -th point and (c) modeling IST from the edited scene to the original scene with the coordinate systems, so as to (d) redirect the viewing direction to the original scene when rendering the edited scene. "offsets," by recording only the final location for each point without extra information. Notably, our formulation even models those whose deformation is not continuous in the space, e.g., cutting a scene into two parts, and thus cannot be covered and solved by NeRF-Editing, Deforming-NeRF, or CageNeRF, as shown in the supplementary. + +![](images/fca9b06889fcc3e42dc57debb9f4d2dab78274a0368ef6398e046a5d8115d411.jpg) + +Editing Shape by Moving Points. We design our shape editing scheme with NeuralEditor. This is achieved by interpreting NeRF rendering as "plotting" the sampled points over the viewing ray. If we render the scene by naively plotting the point cloud $P$ , the shape editing task can be addressed by replacing each point's coordinate from $p_j$ to $q_j$ . For NeuralEditor, we similarly replace the underlying point cloud from $P$ to $Q(P)$ , while maintaining the confidence values and features. This method is general and can also be applied to any point-based NeRF model like PointNeRF. + +Correcting View-Dependence with Infinitesimal Surface Transformation (IST). The editing method above can already obtain reasonable results. However, as illustrated in Fig. 5, the modeled view-dependent colors record the absolute viewing direction, making them incorrect after deformations that change their orientation. + +To solve this issue, we model the infinitesimal surface transformation (IST) for each point to redirect the viewing ray in the correct direction. We construct a local coordinate system for each point to represent the orientation of the infinitesimal surface, using its normal vector and two point indices that are neighbors of the $j$ -th point in both $P$ and $Q(P)$ . By comparing these two coordinate systems, we can obtain an affine transformation $\mathrm{IST}_j$ for the $j$ -th point to redirect the querying view direction (Fig. 5). This procedure is different from modeling space deformation [30, 44, 51], as we only need to model a simple affine transformation at each point, while those methods model a complicated, continuous, and non-linear deformation in the whole space. + +Our proposed method requires a precise point cloud with normal vector-based color modeling. It is thus incompatible with PointNeRF, as PointNeRF is unable to obtain a desired point cloud to estimate the surface normal vectors. + +Fine-Tuning on Deformed Scene. Using the shape editing scheme introduced above, we can apply shape deformation on the scene modeled by our NeuralEditor without any modification to the model architecture or rendering scheme, which means that the resulting model is still a valid, fully functional NeuralEditor. Therefore, we can further fine-tune NeuralEditor on the deformed scene if the ground truth is available. We can even fine-tune the infinitesimal surface transformation with other parameters, to rapidly adjust toward better ambient consistency. This makes NeuralEditor desirable in practice, since in most applications, the final goal is not a zero-shot inference, but to fit the deformed scene well with reduced cost. By supporting fine-tuning, our NeuralEditor aligns well with and achieves this goal. + +As another point-based NeRF model, PointNeRF supports fine-tuning but cannot leverage infinitesimal surface transformation fine-tuning to further optimize the performance. On the other hand, mesh-based NeRF editing models [30, 44, 51] do not support fine-tuning well: With deforming the space instead of the scene, these models' rendering scheme has highly changed. In their rendering process, a ray may go through a long, irregular way to reach the scene's surface. As the modeled space deformation might not be precise, it could be hard to tune the irregular space well, and even hurt other parts of the trained NeRF model. Such issues occur especially for some spaces with non-uniform density, since most of their model components (e.g., positional encodings, voxels) are not designed to deal with non-uniform spaces. Among all these methods, only our NeuralEditor has complete support for fine-tuning. + +# 5. Experiment + +Point Cloud Generation. The underlying point cloud is fundamental to all editing tasks. Fig. 6 first provides a qualitative comparison of point clouds generated by our NeuralEditor and PointNeRF [43] on NeRF Synthetic [27]. Ours are much more precise with sharper details, e.g., the mayonnaise on the Hotdog's sausage, the uneven texture on the Chair's cushion, the edge of the Mic's stand, and the Lego brick's studs. By contrast, PointNeRF's point clouds are blurred and noisy, lose most of the details, and even contain obvious shape deflects on the Hotdog's plate and Chair's backrest. This shows that while the point cloud generation task is challenging, NeuralEditor generates a superprecise point cloud which is crucial for shape editing tasks. + +Experimental Settings. We mainly conduct experiments based on scenes from the NeRF Synthetic (NS) dataset. NS is a widely-used NeRF benchmark constructed from Blender [7] scenes. Due to the lack of publicly available benchmarks for shape editing, we use Blender to construct a reproducible benchmark, including the ground truth of edited scenes for evaluation and fine-tuning. Our shape editing tasks cover all eight scenes in NS, while prior + +![](images/02a70b85b1e99d27d431dac987eb4762b95b169bb6a63eb8ecc38593f79e6b2c.jpg) +Figure 6. NeuralEditor generates much more precise point clouds than PointNeRF [43] in the four scenes of NeRF Synthetic [27]. The points are colored with their normal vectors. + +![](images/bc6bc7668bbcdc03a475b7cb90437a3cb9c2b590519e5aae2ac1ef1ff30e88e1.jpg) +Figure 7. With too coarse cages, DeformingNeRF [44] is unable to perform the deformation faithfully, leading to poor results. + +work [30,44,51] only picks a few scenes. The provided images for NS scenes are with opacity, and there is no requirement for the background color. We evaluate and visualize the results on a black background, for better contrast and clearer detail visualization. In the supplementary, we show the results on a white background with same conclusions. + +Shape Editing Tasks. We evaluate our model on two types of shape editing tasks, as shown in Fig. 1: + +(I) Shape (Mesh) Deformation Task. We consider the shape deformation task as in [30, 44, 51]: deform the shape of a scene in a human-guided way. To construct our deformation tasks from NS and obtain the ground truth, we apply the shape deformation simultaneously to the scene and our point cloud within the provided Blender file. We perform both zero-shot inference and fine-tuning, and compare our rendering results with the ground truth. As demonstrated in Figs. 7 and 8, our deformation tasks are much more precise and aggressive, compared with those in previous work [30, 44, 51]. In the supplementary, we also design a non-continuous deformation task and deformation tasks on the real-world dataset Tanks and Temples [17] (with zero-shot inference only, as there is no ground truth available). +(II) Scene Morphing Task. We address a more challenging shape editing task that has not been investigated in prior NeRF editing work: the scene morphing task. Given two scenes $A$ and $B$ , we should construct a path to gradually + +
ModelZero-Shot Inference, PSNR ↑Fine-Tune for 1 Epoch, PSNR ↑
ChairHotdogLegoDrumsFicusMaterialsMicShipChairHotdogLegoDrumsFicusMaterialsMicShip
DeformingNeRF [44]18.84-13.10-------------
PointNeRF [43]22.2125.9524.5621.0024.2421.2126.7721.1930.1136.0831.4527.1631.4827.5534.3428.90
Naive Plotting24.9127.0125.6421.2926.2221.6527.6322.2932.0136.3831.7228.0933.2130.3135.1530.01
NeuralEditor w/o IST24.9227.0225.6521.2926.2421.6427.6422.2832.2436.6932.7928.3033.3430.4035.2830.08
NeuralEditor (Ours)25.8527.4927.4621.8427.1923.1827.7524.1632.5337.2232.9528.3533.5330.8235.4630.44
+ +Table 1. NeuralEditor significantly and consistently outperforms PointNeRF and Naive Plotting on all deformed scenes of NeRF Synthetic in peak signal-to-noise ratio (PSNR), in both zero-shot inference and fine-tuning settings. Our infinitesimal surface transformation (IST) effectively improves the results by correcting the view-dependent colors. With the precise point cloud generated by NeuralEditor, even Naive Plotting consistently outperforms PointNeRF. Comparison results under other metrics are in the supplementary. + +![](images/f74a14115e3021e730674e5b0bd209b5f0ffc0552d64ad1f60b5c780cebbc1f2.jpg) +Figure 8. NeuralEditor produces superior rendering results to PointNeRF, with significantly fewer artifacts in zero-shot inference. Finetuning further improves the consistency of rendering with the ambient environment. We use a black background for better visualization. + +change one scene to the other, and the intermediate scenes should have reasonable appearances. We are required to render all intermediate scenes. For this task, we use the point cloud diffusion model [24] to generate intermediate point clouds with latent space interpolation between $A$ and $B$ as in [24], and we introduce a K-D tree-based [4] matching algorithm to match the adjacent points to fix the indices of the intermediate scenes. To render an intermediate scene, we apply the shape transformation to the NeuralEditor models trained for scenes $A$ and $B$ , and then interpolate the rendering features to obtain the features of the intermediate scene for rendering. + +NeuralEditor Variants. (1) Full NeuralEditor: Our complete model with all components. (2) NeuralEditor without infinitesimal surface transformation (IST): We remove the maintenance and optimization of infinitesimal surface transformation in scene editing. This key variant of NeuralEditor enables us to evaluate the importance of IST as well as other components of our model. The full ablation study of NeuralEditor is in the supplementary. + +Baselines. We compare NeuralEditor against different types of baselines as follows. (1) Naive Plotting: We use the point cloud generated by NeuralEditor, computing each point's opacity and view-dependent colors with their point features. We render the scene by directly plotting/projecting the point cloud to the camera plane. (2) PointNeRF [43]: For the shape deformation task, we apply the same deformation to the point clouds generated by PointNeRF. For the scene morphing task, we apply the matching algorithm to + +the point clouds generated by PointNeRF and the same intermediate point clouds generated by the point cloud diffusion model [24] for fairness. (3) DeformingNeRF [44]: DeformingNeRF is not compatible with the scene morphing task. For the shape deformation task, we perform the same deformation on vertices of given cages. Note that DeformingNeRF only released trained models for Lego and Chair, so we can only evaluate it on these two scenes. While other models [30,51] support NeRF-based shape deformation via cages or exported meshes, we were unable to use them as baselines - they did not provide executable code nor their deformed scenes for us to evaluate on their tasks. + +Shape (Mesh) Deformation Results. The qualitative comparison is shown in Fig. 8. Both PointNeRF and Naive Plotting have many artifacts, like blur, wrong color, black or white shadows, noise, etc., whereas our powerful NeuralEditor produces clean and realistic rendering results. After fine-tuning, NeuralEditor shows a significant improvement with better rendering results than PointNeRF, indicating that NeuralEditor is able to achieve higher consistency with the surrounding ambient environment. Notably, in the Materials scene (the 6th scene from left), only our NeuralEditor generates reasonable reflection, while both baselines show blurry and visually messy results. Also in the Drums scene, the bottom face of the gong is not visible in any of the training views in the original scene, so all models render poor results in zero-shot inference. However, after only fast fine-tuning, NeuralEditor is able to precisely model the previously unknown surface and generate a sub- + +![](images/e5fdd963b7c5b0e7575fa808d7bac345a35a0ce9533eb71732a090c11ce98b2f.jpg) +Figure 9. Our NeuralEditor produces smooth morphing results between Chair, Hotdog, Lego, and Mic in the NeRF Synthetic dataset, while PointNeRF produces results with blurry textures, black shadows, and gloomy, non-smooth colors. The rendering results in the looped morphing process are arranged in the shape of the numerical digit "3," indicated by the dividing lines and arrows. + +stantially better result than PointNeRF, highlighting NeuralEditor's strength in fast-fitting. All these results demonstrate that NeuralEditor can handle various visual effects and make them consistent in the deformed scene. We provide the figure with higher resolution in the supplementary. + +The quantitative comparison is summarized in Table 1. We observe that: (1) Our NeuralEditor consistently outperforms all the baselines and variants for both zero-shot inference and fine-tuning settings. (2) With the precise point cloud generated by NeuralEditor, the Naive Plotting baseline even consistently outperforms PointNeRF. (3) Our 'w/o IST' variant has a comparable performance to Naive Plotting with the same point cloud and features in the zero-shot inference setting, but after fine-tuning its performance is significantly higher than Naive Plotting, validating the capability of NeuralEditor in NeRF modeling. + +Notably, DeformingNeRF [44] performs poorly in our benchmark with significantly lower metric values. As shown in Fig. 7, the cages provided by DeformingNeRF are too coarse, and cannot even cover the whole scene. Therefore, DeformingNeRF cannot faithfully perform the precise deformation in our benchmark, leading to poor rendering results. On the contrary, both PointNeRF and our NeuralEditor at least faithfully perform the deformation, showing that point cloud is necessary for precise shape editing. + +Scene Morphing Results. The morphing results between 4 NeRF Synthetic scenes are shown in Fig. 9. The + +morphing process starts from Chair, morphs to Hotdog, Lego, Mic, and at last turns back to Chair. NeuralEditor produces smooth rendering results on the point cloud diffusion-guided [24] intermediate scenes, mixing the textures of the two scenes in a reasonable way. In comparison, the rendering results produced by PointNeRF are unsatisfactory, with blurry textures, black shadows, and gloomy, non-smooth colors. These results show that our NeuralEditor can render challenging intermediate morphing scenes and achieve decent results with only the input of moved points. + +# 6. Conclusion + +This paper proposes NeuralEditor, a point cloud-guided NeRF model that supports general shape editing tasks by manipulating the underlying point clouds. Empirical evaluation shows NeuralEditor to produce rendering results of much higher quality than baselines in a zero-shot inference manner, further significantly improving after fast finetuning. NeuralEditor even supports smooth scene morphing between multiple scenes, which is difficult for prior work. We hope that our work can inspire more research on point cloud-guided NeRFs and 3D shape and scene editing tasks. Acknowledgement. This work was supported in part by NSF Grant 2106825, NIFA Award 2020-67021-32799, the Jump ARCHES endowment, the NCSA Fellows program, the IBM-Illinois Discovery Accelerator Institute, the Illinois-Inspor Partnership, and the Amazon Research Award. This work used NVIDIA GPUs at NCSA Delta through allocation CIS220014 from the ACCESS program. We thank the authors of NeRF [27] for their help in processing Blender files of the NS dataset. + +# References + +[1] Henrik Aanaes, Rasmus Ramsbøl Jensen, George Vogiatzis, Engin Tola, and Anders Bjorholm Dahl. Large-scale data for multiple-view stereopsis. International Journal of Computer Vision, 120(2):153-168, 2016. 5 +[2] Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas J. Guibas. Learning representations and generative models for 3D point clouds. In ICML, 2018. 2 +[3] Jonathan T. Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P. Srinivasan. Mip-NeRF: A multiscale representation for antiailiasing neural radiance fields. In ICCV, 2021. 1, 2 +[4] Jon Louis Bentley. Multidimensional binary search trees used for associative searching. Commun. ACM, 18(9):509-517, 1975. 2, 3, 7 +[5] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. MVSNeRF: Fast generalizable radiance field reconstruction from multi-view stereo. In ICCV, 2021. 2 +[6] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In CVPR, 2019. 2 +[7] Blender Online Community. *Blender - a 3D modelling and rendering package*. Blender Foundation, Stichting Blender Foundation, Amsterdam, 2018. 6 +[8] Frank Dellaert and Lin Yen-Chen. Neural volume rendering: NeRF and beyond. arXiv:2101.05204, 2021. 2 +[9] Kyle Gao, Yina Gao, Hongjie He, Denning Lu, Linlin Xu, and Jonathan Li. NeRF: Neural radiance field in 3D vision, a comprehensive review. arXiv:2101.05204, 2021. 2 +[10] Yuan-Chen Guo, Di Kang, Linchao Bao, Yu He, and Song-Hai Zhang. NeRFReN: Neural radiance fields with reflections. In CVPR, 2022. 2 +[11] Peter Hedman, Tobias Ritschel, George Drettakis, and Gabriel Brostow. Scalable inside-out image-based rendering. ACM Trans. Graph., 35(6):231:1-231:11, 2016. 2 +[12] Po-Han Huang, Kevin Matzen, Johannes Kopf, Narendra Ahuja, and Jia-Bin Huang. DeepMVS: Learning multi-view stereopsis. In CVPR, 2018. 5 +[13] Alec Jacobson, Ilya Baran, Ladislav Kavan, Jovan Popovic, and Olga Sorkine. Fast automatic skinning transformations. ACM Trans. Graph., 31(4), 2012. 3 +[14] Tomas Jakab, Richard Tucker, Ameesh Makadia, Jiajun Wu, Noah Snavely, and Angjoo Kanazawa. Keypointdeformer: Unsupervised 3D keypoint discovery for shape control. In CVPR, 2021. 3 +[15] Mengqi Ji, Juergen Gall, Haitian Zheng, Yebin Liu, and Lu Fang. Surfacenet: An end-to-end 3D neural network for multiview stereopsis. In ICCV, 2017. 2 +[16] Tao Ju, Scott Schaefer, and Joe Warren. Mean value coordinates for closed triangular meshes. ACM Trans. Graph., 24(3):561-566, 2005. 3 +[17] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and Temples: Benchmarking large-scale scene reconstruction. ACM Trans. Graph., 36(4), 2017. 6 +[18] Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitzmann. Decomposing NeRF for editing via feature field distillation. In NeurlPS, 2022. 2 + +[19] Marc Levoy and Pat Hanrahan. Light field rendering. In SIGGRAPH, 1996. 2 +[20] Fayao Liu, Chunhua Shen, Guosheng Lin, and Ian Reid. Learning depth from single monocular images using deep convolutional neural fields. IEEE Transactions on Pattern Analysis and Machine Intelligence, 38(10):2024-2039, 2015. 2 +[21] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. In NeurIPS, 2020. 2 +[22] Steven Liu, Xiuming Zhang, Zhoutong Zhang, Richard Zhang, Junyan Zhu, and Bryan C. Russell. Editing conditional radiance fields. In ICCV, 2021. 2 +[23] William E. Lorensen and Harvey E. Cline. Marching cubes: A high resolution 3D surface construction algorithm. In SIGGRAPH, 1987. 3 +[24] Shitong Luo and Wei Hu. Diffusion probabilistic models for 3D point cloud generation. In CVPR, 2021. 2, 7, 8 +[25] Bruce Merry, Patrick Marais, and James Gain. Animation space: A truly linear framework for character animation. ACM Trans. Graph., 25(4):1400-1423, 2006. 3 +[26] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3D reconstruction in function space. In CVPR, 2019. 2 +[27] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2, 4, 6, 8 +[28] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Differentiable volumetric rendering: Learning implicit 3D representations without 3D supervision. In CVPR, 2020. 2 +[29] Jesús Nieto and Toni Susin. Cage based deformations: A survey. Lecture Notes in Computational Vision and Biomechanics, 7:75-99, 2013. 3 +[30] Yicong Peng, Yichao Yan, Shengqi Liu, Yuhao Cheng, Shanyan Guan, Bowen Pan, Guangtao Zhai, and Xiaokang Yang. CageNeRF: Cage-based neural radiance field for generalized 3D deformation and animation. In NeurIPS, 2022. 2, 3, 5, 6, 7 +[31] Bui Tuong Phong. Illumination for computer generated pictures. Commun. ACM, 18(6):311-317, 1975. 2, 4 +[32] Yi-Ling Qiao, Alexander Gao, and Ming C. Lin. NeuPhysics: Editable neural geometry and physics from monocular videos. In NeurIPS, 2022. 2 +[33] Sara Fridovich-Keil and Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In CVPR, 2022. 1, 2 +[34] Noah Snavely, Steven M. Seitz, and Richard Szeliski. Photo tourism: Exploring photo collections in 3D. ACM Trans. Graph., 25(3):835-846, 2006. 2 +[35] Jiaxiang Tang, Xiaokang Chen, Jingbo Wang, and Gang Zeng. Compressible-composable NeRF via rank-residual decomposition. In NeurIPS, 2022. 2 + +[36] Jean-Marc Thiery, Julien Tierny, and Tamy Boubekeur. Jacobians and hessian of mean value coordinates for closed triangular meshes. The Visual Computer, 30(9):981-995, 2014. 3 +[37] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T. Barron, and Pratul P. Srinivasan. Ref-NeRF: Structured view-dependent appearance for neural radiance fields. In CVPR, 2022. 1, 2, 4, 5 +[38] Nanyang Wang, Yinda Zhang, Zhuwen Li, Yanwei Fu, Wei Liu, and Yu-Gang Jiang. Pixel2Mesh: Generating 3D mesh models from single RGB images. In ECCV, 2018. 2 +[39] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul Srinivasan, Howard Zhou, Jonathan T. Barron, Ricardo MartinBrualla, Noah Snavely, and Thomas Funkhouser. IBRNet: Learning multi-view image-based rendering. In CVPR, 2021. 1, 2 +[40] Suttisak Wizadwongsa, Pakkapon Phongthawee, Jiraphon Yenphraphai, and Supasorn Suwajanakorn. NeX: Real-time view synthesis with neural basis expansion. In CVPR, 2021. 2 +[41] Liwen Wu, Jae Yong Lee, Anand Bhattachad, Yu-Xiong Wang, and David Forsyth. DIVeR: Real-time and accurate neural radiance fields with deterministic integration for volume rendering. In CVPR, 2022. 2, 4 +[42] Dejia Xu, Peihao Wang, Yifan Jiang, Zhiwen Fan, and Zhangyang Wang. Signal processing for implicit neural representations. In NeurIPS, 2022. 2 +[43] Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-NeRF: Point-based neural radiance fields. In CVPR, 2021. 2, 3, 6, 7 +[44] Tianhan Xu and Tatsuya Harada. Deforming radiance fields with cages. In ECCV, 2022. 2, 3, 5, 6, 7, 8 +[45] Bangbang Yang, Yinda Zhang, Yinghao Xu, Yijin Li, Han Zhou, Hujun Bao, Guofeng Zhang, and Zhaopeng Cui. Learning object-compositional neural radiance field for editing scene rendering. In ICCV, 2021. 2 +[46] Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. MVSNet: Depth inference for unstructured multiview stereo. In ECCV, 2018. 5 +[47] Wang Yifan, Noam Aigerman, Vladimir G Kim, Siddhartha Chaudhuri, and Olga Sorkine-Hornung. Neural cages for detail-preserving 3D deformations. In CVPR, 2020. 3 +[48] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. PlenOctrees for real-time rendering of neural radiance fields. In ICCV, 2021. 1, 2 +[49] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelNeRF: Neural radiance fields from one or few images. In CVPR, 2021. 2 +[50] Yu-Jie Yuan, Yu-Kun Lai, Tong Wu, Lin Gao, and Ligang Liu. A revisit of shape editing techniques: From the geometric to the neural viewpoint. Journal of Computer Science and Technology, 36(3):520-554, 2021. 3 +[51] Yu-Jie Yuan, Yang tian Sun, Yu-Kun Lai, Yuewen Ma, Rongfei Jia, and Lin Gao. NeRF-Editing: Geometry editing of neural radiance fields. In CVPR, 2022. 2, 3, 5, 6, 7 + +[52] Yuzhe Zhang, Jianmin Zheng, and Yiyu Cai. Proxy-driven free-form deformation by topology-adjustable control lattice. Computers & Graphics, 89:167-177, 2020. 3 +[53] Qian-Yi Zhou, Jaesik Park, and Vladlen Koltun. Open3D: A modern library for 3D data processing. arXiv:1801.09847, 2018. 4 \ No newline at end of file diff --git a/2023/NeuralEditor_ Editing Neural Radiance Fields via Manipulating Point Clouds/images.zip b/2023/NeuralEditor_ Editing Neural Radiance Fields via Manipulating Point Clouds/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1839fc66b003c81bbaf350c2600f8643d69f9dbd --- /dev/null +++ b/2023/NeuralEditor_ Editing Neural Radiance Fields via Manipulating Point Clouds/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0bab3c1cdfcb7b7f1895b068490ac959f61bb6e5419b4222b723c35dde9716d8 +size 583922 diff --git a/2023/NeuralEditor_ Editing Neural Radiance Fields via Manipulating Point Clouds/layout.json b/2023/NeuralEditor_ Editing Neural Radiance Fields via Manipulating Point Clouds/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..51b171c2607b902f9cb8e961fea15d21c46c2017 --- /dev/null +++ b/2023/NeuralEditor_ Editing Neural Radiance Fields via Manipulating Point Clouds/layout.json @@ -0,0 +1,7134 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 57, + 103, + 536, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 103, + 536, + 121 + ], + "spans": [ + { + "bbox": [ + 57, + 103, + 536, + 121 + ], + "type": "text", + "content": "NeuralEditor: Editing Neural Radiance Fields via Manipulating Point Clouds" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 154, + 142, + 236, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 154, + 142, + 236, + 157 + ], + "spans": [ + { + "bbox": [ + 154, + 142, + 236, + 157 + ], + "type": "text", + "content": "Jun-Kun Chen" + }, + { + "bbox": [ + 154, + 142, + 236, + 157 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 261, + 143, + 324, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 261, + 143, + 324, + 157 + ], + "spans": [ + { + "bbox": [ + 261, + 143, + 324, + 157 + ], + "type": "text", + "content": "Jipeng Lyu" + }, + { + "bbox": [ + 261, + 143, + 324, + 157 + ], + "type": "inline_equation", + "content": "^{2\\dagger}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 350, + 143, + 434, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 143, + 434, + 157 + ], + "spans": [ + { + "bbox": [ + 350, + 143, + 434, + 157 + ], + "type": "text", + "content": "Yu-Xiong Wang" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 158, + 283, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 158, + 283, + 173 + ], + "spans": [ + { + "bbox": [ + 67, + 158, + 283, + 173 + ], + "type": "text", + "content": "1University of Illinois at Urbana-Champaign" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 307, + 159, + 523, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 159, + 523, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 159, + 523, + 173 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 307, + 159, + 523, + 173 + ], + "type": "text", + "content": "Peking University " + }, + { + "bbox": [ + 307, + 159, + 523, + 173 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 307, + 159, + 523, + 173 + ], + "type": "text", + "content": " Equal Contribution" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 119, + 175, + 471, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 175, + 471, + 188 + ], + "spans": [ + { + "bbox": [ + 119, + 175, + 471, + 188 + ], + "type": "text", + "content": "{junkun3，yxw}@illinois.edu lvjipeng@pku.edu.cn" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 49, + 209, + 184, + 333 + ], + "blocks": [ + { + "bbox": [ + 49, + 209, + 184, + 333 + ], + "lines": [ + { + "bbox": [ + 49, + 209, + 184, + 333 + ], + "spans": [ + { + "bbox": [ + 49, + 209, + 184, + 333 + ], + "type": "image", + "image_path": "4589dde9f2e2500c54fe59004e1e22d8c9108ed27ca90fd9432a6d04afdc8926.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 345, + 546, + 378 + ], + "lines": [ + { + "bbox": [ + 46, + 345, + 546, + 378 + ], + "spans": [ + { + "bbox": [ + 46, + 345, + 546, + 378 + ], + "type": "text", + "content": "Figure 1. Our NeuralEditor offers native support for general and flexible shape editing of neural radiance fields via manipulating point clouds. By generating a precise point cloud of the scene with a novel point cloud-guided NeRF model, our NeuralEditor produces high-fidelity rendering results in both shape deformation and more challenging scene morphing tasks." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 188, + 209, + 545, + 337 + ], + "blocks": [ + { + "bbox": [ + 188, + 209, + 545, + 337 + ], + "lines": [ + { + "bbox": [ + 188, + 209, + 545, + 337 + ], + "spans": [ + { + "bbox": [ + 188, + 209, + 545, + 337 + ], + "type": "image", + "image_path": "1b9a1be3d3a7295491938e1aeb98a2f36f0f998e2d7cf65aba738962a02a830c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 143, + 387, + 192, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 387, + 192, + 399 + ], + "spans": [ + { + "bbox": [ + 143, + 387, + 192, + 399 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 403, + 290, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 403, + 290, + 656 + ], + "spans": [ + { + "bbox": [ + 46, + 403, + 290, + 656 + ], + "type": "text", + "content": "This paper proposes NeuralEditor that enables neural radiance fields (NeRFs) natively editable for general shape editing tasks. Despite their impressive results on novel-view synthesis, it remains a fundamental challenge for NeRFs to edit the shape of the scene. Our key insight is to exploit the explicit point cloud representation as the underlying structure to construct NeRFs, inspired by the intuitive interpretation of NeRF rendering as a process that projects or \"plots\" the associated 3D point cloud to a 2D image plane. To this end, NeuralEditor introduces a novel rendering scheme based on deterministic integration within K-D tree-guided density-adaptive voxels, which produces both high-quality rendering results and precise point clouds through optimization. NeuralEditor then performs shape editing via mapping associated points between point clouds. Extensive evaluation shows that NeuralEditor achieves state-of-the-art performance in both shape deformation and scene morphing tasks. Notably, NeuralEditor supports both zero-shot inference and further fine-tuning over the edited scene. Our code, benchmark, and demo video are available at immortalco.github.io/NeuralEditor." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 670, + 127, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 127, + 681 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 127, + 681 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": "Perhaps the most memorable shot of the film Transformers, Optimus Prime is seamlessly transformed between a" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 388, + 546, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 388, + 546, + 554 + ], + "spans": [ + { + "bbox": [ + 304, + 388, + 546, + 554 + ], + "type": "text", + "content": "humanoid and a Peterbilt truck – such free-form editing of 3D objects and scenes is a fundamental task in 3D computer vision and computer graphics, directly impacting applications such as visual simulation, movie, and game industries. In these applications, often we are required to manipulate a scene or objects in the scene by editing or modifying its shape, color, light condition, etc., and generate visually-faithful rendering results on the edited scene efficiently. Among the various editing operations, shape editing has received continued attention but remains challenging, where the scene is deformed in a human-guided way, while all of its visual attributes (e.g., shape, color, brightness, and light condition) are supposed to be natural and consistent with the ambient environment." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 556, + 545, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 556, + 545, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 556, + 545, + 652 + ], + "type": "text", + "content": "State-of-the-art rendering models are based on implicit neural representations, as exemplified by neural radiance field (NeRF) [27] and its variants [3, 33, 37, 39, 48]. Despite their impressive novel-view synthesis results, most of the NeRF models substantially lack the ability for users to adjust, edit, or modify the shape of scene objects. On the other hand, shape editing operations can be natively applied to explicit 3D representations such as point clouds and meshes." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "Inspired by this, we propose NeuralEditor – a general and flexible approach to editing neural radiance fields via manipulating point clouds (Fig. 1). Our key insight is to benefit from the best of both worlds: the superiority in rendering performance from implicit neural representation" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12439" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 286, + 108 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 286, + 108 + ], + "type": "text", + "content": "combined with the ease of editing from explicit point cloud representation. NeuralEditor enables us to perform a wide spectrum of shape editing operations in a consistent way." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 110, + 286, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 110, + 286, + 360 + ], + "spans": [ + { + "bbox": [ + 49, + 110, + 286, + 360 + ], + "type": "text", + "content": "Such introduction of point clouds into NeRF for general shape editing is rooted in our interpretation of NeRF rendering as a process that projects or \"plots\" the associated 3D point cloud to a 2D image plane. Conceptually, with a dense enough point cloud where each point has an opacity and its color is defined as a function of viewing direction, directly plotting the point cloud would achieve similar visual effects (i.e., transparency and view-dependent colors) that are rendered by NeRF. This intrinsic integration between NeRF and point clouds underscores the advantage of our NeuralEditor over existing mesh-based NeRF editing methods such as NeRF-Editing [51], Deforming-NeRF [44], and CageNeRF [30], where the process of constructing and optimizing the mesh is separated from the NeRF modeling, making them time-consuming. More importantly, with the point cloud constructed for a scene, the shape editing can be natively defined as and easily solved by just moving each point into the new, edited position and re-plotting the point cloud. Therefore, our approach supports more general scene editing operations which are difficult to achieve via mesh-guided space deformation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 363, + 286, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 363, + 286, + 542 + ], + "spans": [ + { + "bbox": [ + 49, + 363, + 286, + 542 + ], + "type": "text", + "content": "The key component in our NeuralEditor lies in a point cloud-guided NeRF model that natively supports general shape editing operations. While the recent method Point-NeRF [43] has demonstrated improved novel-view synthesis capability based on point clouds, it is not supportive to shape editing. Our idea then is to exploit the underlying point cloud in ways of not only optimizing its structure and features (e.g., adaptive voxels) for rendering, but also extracting additional useful attributes (e.g., normal vectors) to guide the editing process. To this end, we introduce K-D trees [4] to construct density-adaptive voxels for efficient and stable rendering, together with a novel deterministic integration strategy. Moreover, we model the color with the Phong reflection [31] to decompose the specular color and better represent the scene geometry." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 544, + 286, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 544, + 286, + 675 + ], + "spans": [ + { + "bbox": [ + 49, + 544, + 286, + 675 + ], + "type": "text", + "content": "With a much more precise point cloud attributed to these improvements, our NeuralEditor achieves high-fidelity rendering results on deformed scenes compared with prior work as shown in Fig. 1, even in a zero-shot inference manner without additional training. Through fast fine-tuning, the visual quality of the deformed scene is further enhanced, almost perfectly consistent with the surrounding light condition. In addition, under the guidance of a point cloud diffusion model [24], NeuralEditor can be naturally extended for smooth scene morphing across multiple scenes, which is difficult for existing NeRF editing work." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 677, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 677, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 677, + 286, + 713 + ], + "type": "text", + "content": "Our contributions are four-fold. (1) We introduce NeuralEditor, a flexible and versatile approach that makes neural radiance fields editable through manipulating point" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 308, + 72, + 545, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 72, + 545, + 215 + ], + "spans": [ + { + "bbox": [ + 308, + 72, + 545, + 215 + ], + "type": "text", + "content": "clouds. (2) We propose a point cloud-guided NeRF model based on K-D trees and deterministic integration, which produces precise point clouds and supports general scene editing. (3) Due to the lack of publicly available benchmarks for shape editing, we construct and release a reproducible benchmark that promotes future research on shape editing. (4) We investigate a wide range of shape editing tasks, covering both shape deformation (as studied in existing NeRF editing work) and challenging scene morphing (a novel task addressed here). NeuralEditor achieves state-of-the-art performance on all shape editing tasks in a unified framework, without extra information or supervision." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 308, + 227, + 391, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 227, + 391, + 239 + ], + "spans": [ + { + "bbox": [ + 308, + 227, + 391, + 239 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 308, + 247, + 545, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 247, + 545, + 413 + ], + "spans": [ + { + "bbox": [ + 308, + 247, + 545, + 413 + ], + "type": "text", + "content": "Neural Scene Representation. Traditional methods model scenes with explicit [2, 11, 15, 20, 34, 38] or implicit [6, 19, 26, 28, 40] 3D geometric or shape representations. Initiated by NeRF [27], leveraging implicit neural networks to represent scenes and perform novel-view synthesis has become a fast-developing field in 3D vision [8,9]. While most of the follow-up work focuses on improving aspects such as the rendering realism [3, 10, 21, 37], efficiency [33, 41, 48], and cross-scene generalization [5, 39, 43, 49], the scene editing capability is substantially missing in the NeRF family which we address in this paper. In addition, we exploit K-D tree-guided point clouds as the underlying structure, different from other NeRF variants based on octrees [21, 33, 48] or plain voxels [41]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 308, + 415, + 545, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 415, + 545, + 593 + ], + "spans": [ + { + "bbox": [ + 308, + 415, + 545, + 593 + ], + "type": "text", + "content": "Point-Based NeRFs. Recently, using point clouds to build a NeRF model has shown better encoding of scene shape and improved rendering performance, as represented by PointNeRF [43]. PointNeRF proposes a point initialization network to produce the initial point cloud together with the point features, which is further optimized by a pruning and growing strategy. While both PointNeRF and our NeuralEditor employ point clouds as the underlying structure, NeuralEditor better exploits useful information within the point clouds: PointNeRF only directly uses the locations of points; by contrast, NeuralEditor considers the point cloud more as a geometrical shape and extracts relevant information like normal vectors, which plays an important role in rendering and shape editing. Importantly, our approach is designed to support scene editing, in contrast to PointNeRF." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 308, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 594, + 545, + 713 + ], + "type": "text", + "content": "Scene Editing via NeRFs. Different types of scene editing have been studied under NeRFs. EditNeRF [22], ObjectNeRF [45], and DistillNeRF [18] perform simple shape and color editing for objects specified with human-input scribble, pixel, segment, language, etc. NeuPhysics [32] edits a dynamic scene via physics parameters. CCNeRF [35] proposes an explicit NeRF representation with tensor rank decomposition to support scene composition. INSP-Net [42] considers filter editing like denoising. Such work cannot address 3D shape editing and only sup" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 750, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 750, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 750, + 317, + 757 + ], + "type": "text", + "content": "12440" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "content": "ports simple editing operations, like object selection, similarity transformation, or limited shape deformation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 96, + 289, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 96, + 289, + 373 + ], + "spans": [ + { + "bbox": [ + 46, + 96, + 289, + 373 + ], + "type": "text", + "content": "3D Shape Editing. Traditional representation methods support keypoint-based shape editing [13, 14, 16, 25, 36, 47, 52] with meshes [29, 50], which cannot be directly applied to implicit representations used by NeRF. Existing NeRF editing work primarily studies a particular shape editing task, mesh deformation, and addresses it in a common paradigm [30, 44, 51]: A mesh of the scene is first constructed by either exporting it from a trained NeRF with the Marching Cubes algorithm [23], or optimizing close-to-surface cages along with training. After the user deforms the mesh, the deformed scene is rendered by deforming the space and bending the viewing rays in the original scene with the trained NeRF. Doing so requires extra efforts to convert implicit scene representation to explicit mesh, which might not be precise enough, and only supports continuous shape editing that can be converted to space deformation. On the contrary, our NeuralEditor directly maintains and utilizes alternative explicit scene representation – the point cloud which is intrinsically integrated with NeRF, making NeuralEditor require no extra efforts and support more general shape editing tasks like scene morphing. NeuralEditor supports both zero-shot inference and further fine-tuning over the edited scene, while prior work cannot." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 381, + 275, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 381, + 275, + 395 + ], + "spans": [ + { + "bbox": [ + 47, + 381, + 275, + 395 + ], + "type": "text", + "content": "3. NeuralEditor: Point Cloud-Guided NeRF" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 401, + 289, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 401, + 289, + 570 + ], + "spans": [ + { + "bbox": [ + 46, + 401, + 289, + 570 + ], + "type": "text", + "content": "We propose a novel point cloud-guided NeRF model, NeuralEditor - it not only achieves realistic rendering results in the novel-view synthesis task, but also produces a point cloud that precisely describes the shape of the scene, thus facilitating general shape editing tasks. As illustrated in Fig. 2, we leverage the K-D trees [4] to construct density-adaptive voxels (which also naturally enable us to skip empty spaces), and introduce deterministic spline integration for rendering. We use the Phong reflection to model the color along with the normal vectors obtained from the underlying point cloud. With our enhanced point cloud optimization, NeuralEditor obtains much more precise underlying point clouds, compared with noisy and imprecise outputs of state-of-the-art PointNeRF [43] (as shown in Sec. 5)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 575, + 186, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 575, + 186, + 588 + ], + "spans": [ + { + "bbox": [ + 47, + 575, + 186, + 588 + ], + "type": "text", + "content": "3.1. K-D Tree-Guided Voxels" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 594, + 287, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 666 + ], + "type": "text", + "content": "To render with points, we construct multi-scale density-adaptive voxels based on K-D trees [4], namely, " + }, + { + "bbox": [ + 46, + 594, + 287, + 666 + ], + "type": "inline_equation", + "content": "K - D" + }, + { + "bbox": [ + 46, + 594, + 287, + 666 + ], + "type": "text", + "content": " voxels. K-D trees are a data structure constructed on " + }, + { + "bbox": [ + 46, + 594, + 287, + 666 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 594, + 287, + 666 + ], + "type": "text", + "content": "-dimensional points, where " + }, + { + "bbox": [ + 46, + 594, + 287, + 666 + ], + "type": "inline_equation", + "content": "K = 3" + }, + { + "bbox": [ + 46, + 594, + 287, + 666 + ], + "type": "text", + "content": " for 3D points. As a special decision tree, K-D tree's each node divides the point set into two equal-sized parts with axis-parallel criterion." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 666, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 666, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 666, + 288, + 715 + ], + "type": "text", + "content": "For each K-D tree's node, we compute its bounding box by taking the minimum and maximum " + }, + { + "bbox": [ + 46, + 666, + 288, + 715 + ], + "type": "inline_equation", + "content": "x, y, z" + }, + { + "bbox": [ + 46, + 666, + 288, + 715 + ], + "type": "text", + "content": " coordinates in its subtree and with proper padding margins. As we divide the points in a top-down manner in one of the " + }, + { + "bbox": [ + 46, + 666, + 288, + 715 + ], + "type": "inline_equation", + "content": "x, y, z" + }, + { + "bbox": [ + 46, + 666, + 288, + 715 + ], + "type": "text", + "content": " direc" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 315, + 71, + 541, + 148 + ], + "blocks": [ + { + "bbox": [ + 315, + 71, + 541, + 148 + ], + "lines": [ + { + "bbox": [ + 315, + 71, + 541, + 148 + ], + "spans": [ + { + "bbox": [ + 315, + 71, + 541, + 148 + ], + "type": "image", + "image_path": "bd5934b42f8753048540e844e4d330c236474cd2f2bb976fc87fea27eb17524a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 148, + 547, + 237 + ], + "lines": [ + { + "bbox": [ + 304, + 148, + 547, + 237 + ], + "spans": [ + { + "bbox": [ + 304, + 148, + 547, + 237 + ], + "type": "text", + "content": "Figure 2. Our K-D Voxels. Column 1: original point clouds constructed from two scenes, colored with normal vector directions. Column 2: upper-level voxels which coarsely represent the shape. Column 3: lower-level voxels which tightly cover the shape. Right: Visualization of K-D voxels on a 2D point cloud. Each color represents boxes of nodes on each level of the K-D tree. Lower-level boxes containing fewer points cover the shape more tightly, and vice versa for higher-level boxes." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 242, + 547, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 242, + 547, + 339 + ], + "spans": [ + { + "bbox": [ + 304, + 242, + 547, + 339 + ], + "type": "text", + "content": "tions, in each layer of the K-D tree, different nodes' bounding boxes are mutually exclusive. Therefore, the bounding boxes can be regarded as voxels. As boxes in the upper layers contain more points (larger voxels), while those in the lower layers contain fewer points (smaller voxels), we naively obtain a multi-scale voxel construction from one K-D tree. As shown in Fig. 2, voxels from the large to small scales represent the shape of the scene from coarse to fine." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 346, + 460, + 360 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 346, + 460, + 360 + ], + "spans": [ + { + "bbox": [ + 306, + 346, + 460, + 360 + ], + "type": "text", + "content": "3.2. Rendering Over K-D Voxels" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 365, + 547, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 365, + 547, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 547, + 437 + ], + "type": "text", + "content": "We now introduce a rendering scheme that exploits K-D voxels to perform all the sub-procedures associated with rendering in a unified way. This scheme enables us to render more naturally, efficiently, and even deterministically, meanwhile it also simplifies some widely-adopted design choices in conventional NeRF rendering." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 437, + 548, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 548, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 548, + 653 + ], + "type": "text", + "content": "Skipping Empty Spaces. In NeRF rendering, we are supposed to focus only on the surface of scene objects. As shown in Fig. 2, all our K-D voxels are produced to stick to the surface of objects, which the point cloud is constructed to describe. Such a property allows us to avoid explicitly \"skipping\" empty spaces, which often requires extra consideration in most NeRF models - only considering the space inside a voxel automatically focuses on the surface; as the depth of the voxel's node goes deeper, it becomes closer to the surface. Moreover, during the construction of the K-D tree, the points at each node are divided within its sub-nodes, and the node's voxel fully covers all its sub-nodes. This further provides us with a native top-down recursive procedure to locate the voxels intersected with the querying ray: We start from the root node, and recurse on the sub-nodes until we (1) reach a pre-set node depth (or equivalently, a pre-set voxel size) and then query within the associated voxel, or (2) stop recursion on non-intersected nodes." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "content": "Density-Adaptive Rendering. An important design in NeRFs is the coarse-to-fine strategy for density-adaptive rendering, so that more points are sampled in high-volume density areas. Our K-D voxels natively support such a design without additional bells and whistles. This is because" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 758 + ], + "type": "text", + "content": "12441" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 123, + 62, + 473, + 187 + ], + "blocks": [ + { + "bbox": [ + 123, + 62, + 473, + 187 + ], + "lines": [ + { + "bbox": [ + 123, + 62, + 473, + 187 + ], + "spans": [ + { + "bbox": [ + 123, + 62, + 473, + 187 + ], + "type": "image", + "image_path": "8286100af116b003cc588dc64b936796675e5dfe7fe4d68a7a8d758cbc05da6a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 190, + 547, + 213 + ], + "lines": [ + { + "bbox": [ + 46, + 190, + 547, + 213 + ], + "spans": [ + { + "bbox": [ + 46, + 190, + 547, + 213 + ], + "type": "text", + "content": "Figure 3. Our NeuralEditor architecture. We propose deterministic spline integration for KNN-based point features over each K-D tree-guided density-adaptive voxel, and model the color via Phong reflection with normal vectors estimated from the point cloud's shape." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 217, + 289, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 217, + 289, + 384 + ], + "spans": [ + { + "bbox": [ + 46, + 217, + 289, + 384 + ], + "type": "text", + "content": "voxels in the same K-D tree layer contain the same number of points. As the point density can be regarded as an approximation of volume density, all such voxels have the same density. Therefore, we directly use K-D voxels to guide the density-adaptive rendering. Specifically, we conduct the rendering process at the voxels of some bottom layers in the K-D tree. For each querying ray, we use the aforementioned recursive procedure to locate the minimal intersected voxels that are deep enough. Here \"minimal\" means that the ray intersects with the node's voxel, but does not intersect with any sub-node's voxel. These intersected voxels divide the querying ray into several segments (Fig. 3). The segments covered by a voxel are close to the surface and used for rendering, while those not covered are in empty spaces." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 384, + 287, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 384, + 287, + 504 + ], + "spans": [ + { + "bbox": [ + 46, + 384, + 287, + 504 + ], + "type": "text", + "content": "Deterministic Spline Integration. DIVER [41] shows that deterministic integration outperforms stochastic integration in NeRF rendering. So we perform a deterministic integration to obtain the segment's feature within each voxel. Since we do not necessarily have points at the voxel's vertices, the trilinear interpolation used in DIVER is not feasible here. Instead, we use spline integration. For the " + }, + { + "bbox": [ + 46, + 384, + 287, + 504 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 384, + 287, + 504 + ], + "type": "text", + "content": "-th intersected voxel in the ray passing order, we uniformly select points in this segment, and integrate their features to obtain the average feature " + }, + { + "bbox": [ + 46, + 384, + 287, + 504 + ], + "type": "inline_equation", + "content": "f_{i}" + }, + { + "bbox": [ + 46, + 384, + 287, + 504 + ], + "type": "text", + "content": " of the segment of the " + }, + { + "bbox": [ + 46, + 384, + 287, + 504 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 384, + 287, + 504 + ], + "type": "text", + "content": "-th voxel:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 88, + 509, + 287, + 536 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 509, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 88, + 509, + 287, + 536 + ], + "type": "interline_equation", + "content": "f _ {i} = \\frac {1}{r _ {i} - l _ {i}} \\int_ {l _ {i}} ^ {r _ {i}} \\mathbf {f e a t u r e} (o + t \\cdot d) \\mathrm {d} t, \\tag {1}", + "image_path": "800a741b216eef4ffdf791801e2153daa529533def619930d7b67650415f551d.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 536, + 287, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 536, + 287, + 585 + ], + "spans": [ + { + "bbox": [ + 47, + 536, + 287, + 585 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 536, + 287, + 585 + ], + "type": "inline_equation", + "content": "[l_i, r_i]" + }, + { + "bbox": [ + 47, + 536, + 287, + 585 + ], + "type": "text", + "content": " is the intersection interval, and " + }, + { + "bbox": [ + 47, + 536, + 287, + 585 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 47, + 536, + 287, + 585 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 536, + 287, + 585 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 47, + 536, + 287, + 585 + ], + "type": "text", + "content": " are the source and direction of the querying ray, respectively. This average feature " + }, + { + "bbox": [ + 47, + 536, + 287, + 585 + ], + "type": "inline_equation", + "content": "f_i" + }, + { + "bbox": [ + 47, + 536, + 287, + 585 + ], + "type": "text", + "content": " can be interpreted as the feature of a representative point " + }, + { + "bbox": [ + 47, + 536, + 287, + 585 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 47, + 536, + 287, + 585 + ], + "type": "text", + "content": " located somewhere in the segment." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 586, + 287, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 586, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 47, + 586, + 287, + 633 + ], + "type": "text", + "content": "KNN-Based Feature Aggregation. For each uniformly selected point " + }, + { + "bbox": [ + 47, + 586, + 287, + 633 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 47, + 586, + 287, + 633 + ], + "type": "text", + "content": " in the segment during spline integration, we obtain its feature via weighted interpolation from the features of its " + }, + { + "bbox": [ + 47, + 586, + 287, + 633 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 47, + 586, + 287, + 633 + ], + "type": "text", + "content": " nearest neighbors (KNN) in the point cloud:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 69, + 639, + 287, + 669 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 639, + 287, + 669 + ], + "spans": [ + { + "bbox": [ + 69, + 639, + 287, + 669 + ], + "type": "interline_equation", + "content": "\\operatorname {f e a t u r e} (q) = \\sum_ {p _ {j} \\in \\operatorname {K N N} (q; K)} k _ {j} e _ {j}, \\tag {2}", + "image_path": "0655900d8651050c45ca7b715a792536ae8377dac20506e8592681ce94bf9932.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 70, + 669, + 264, + 688 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 669, + 264, + 688 + ], + "spans": [ + { + "bbox": [ + 70, + 669, + 264, + 688 + ], + "type": "interline_equation", + "content": "\\left\\{k_{j}\\right\\} = \\operatorname *{SoftMax}_{p_{j}\\in \\mathrm{KNN}(q;K)}\\left(\\log \\gamma_{j} - \\log \\| q - p_{j}\\|_{2}^{2}\\right),", + "image_path": "12262c852871886728036447b694a6ced14462c281896ee49fe676348df66331.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "content": "where for each point " + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "inline_equation", + "content": "p_j" + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "content": " in the point cloud, we parameterize its confidence " + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\gamma_j" + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "content": " and point feature " + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "inline_equation", + "content": "e_j" + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "content": " as in PointNeRF." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 320, + 217, + 532, + 358 + ], + "blocks": [ + { + "bbox": [ + 320, + 217, + 532, + 358 + ], + "lines": [ + { + "bbox": [ + 320, + 217, + 532, + 358 + ], + "spans": [ + { + "bbox": [ + 320, + 217, + 532, + 358 + ], + "type": "image", + "image_path": "5bf97e2419a70f2570269ef2a1b209b53783419124e6c97d6266575431f3f7fa.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 358, + 546, + 392 + ], + "lines": [ + { + "bbox": [ + 305, + 358, + 546, + 392 + ], + "spans": [ + { + "bbox": [ + 305, + 358, + 546, + 392 + ], + "type": "text", + "content": "Figure 4. In the two scenes of NeRF Synthetic [27], NeuralEditor optimizes the rough initial point cloud to a precise point cloud. The points are colored with their normal vectors." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 399, + 546, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 399, + 546, + 590 + ], + "spans": [ + { + "bbox": [ + 304, + 399, + 546, + 590 + ], + "type": "text", + "content": "Phong Reflection-Based Color Modeling with Point Cloud Normal Vectors. To obtain the volume density " + }, + { + "bbox": [ + 304, + 399, + 546, + 590 + ], + "type": "inline_equation", + "content": "\\sigma_{i}" + }, + { + "bbox": [ + 304, + 399, + 546, + 590 + ], + "type": "text", + "content": " and color " + }, + { + "bbox": [ + 304, + 399, + 546, + 590 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 304, + 399, + 546, + 590 + ], + "type": "text", + "content": " of the representative point " + }, + { + "bbox": [ + 304, + 399, + 546, + 590 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 304, + 399, + 546, + 590 + ], + "type": "text", + "content": " in the " + }, + { + "bbox": [ + 304, + 399, + 546, + 590 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 399, + 546, + 590 + ], + "type": "text", + "content": "-th voxel, we use the Phong reflection model [31]. As we have the underlying point cloud, we use Open3D [53] to estimate the normal vector for each point, and integrate these vectors over the interval to get an average normal vector " + }, + { + "bbox": [ + 304, + 399, + 546, + 590 + ], + "type": "inline_equation", + "content": "n_i" + }, + { + "bbox": [ + 304, + 399, + 546, + 590 + ], + "type": "text", + "content": ". Such information better characterizes the shape of point clouds (scenes), which plays an important role in Phong-based rendering and also implicitly facilitates optimizing the point clouds (Sec. 3.3). Consistent with RefNeRF [37], we use multiple multilayer perceptrons (MLPs) to model other attributes, including volume density, tint, roughness, and diffuse and specular color, and use the Phong formula to calculate the final " + }, + { + "bbox": [ + 304, + 399, + 546, + 590 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 304, + 399, + 546, + 590 + ], + "type": "text", + "content": " with these attributes. Finally, we aggregate " + }, + { + "bbox": [ + 304, + 399, + 546, + 590 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 304, + 399, + 546, + 590 + ], + "type": "text", + "content": " of all segments on the ray to obtain the final color:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 334, + 597, + 515, + 623 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 597, + 515, + 623 + ], + "spans": [ + { + "bbox": [ + 334, + 597, + 515, + 623 + ], + "type": "interline_equation", + "content": "c _ {\\text {p i x e l}} = \\sum_ {i \\geq 1} \\tau_ {i} \\cdot \\left(1 - \\exp \\left(- \\left(r _ {i} - l _ {i}\\right) \\sigma_ {i}\\right) \\cdot c _ {i}, \\right.", + "image_path": "ca417be18fa62cfcc0fc053b5aa13c06326b09136cb5fa4fcf23dadf1941edb4.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 335, + 623, + 545, + 659 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 623, + 545, + 659 + ], + "spans": [ + { + "bbox": [ + 335, + 623, + 545, + 659 + ], + "type": "interline_equation", + "content": "\\tau_ {i} = \\exp \\left(- \\sum_ {i ^ {\\prime} = 1} ^ {i - 1} \\left(r _ {i ^ {\\prime}} - l _ {i ^ {\\prime}}\\right) \\sigma_ {i ^ {\\prime}}\\right). \\tag {3}", + "image_path": "537e7545d566181e8334bcdc2b5e69dd76e718b08f990725128db4c7d4dadf70.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 670, + 450, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 670, + 450, + 684 + ], + "spans": [ + { + "bbox": [ + 306, + 670, + 450, + 684 + ], + "type": "text", + "content": "3.3. Point Cloud Optimization" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 689, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 546, + 713 + ], + "type": "text", + "content": "Point Cloud Initialization. To start training, we need a coarse initial point cloud. Consistent with PointNeRF, we" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12442" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "text", + "content": "use a point generation network, which consists of a multiview stereo (MVS) model [12] based on a 3D convolutional neural network (CNN), to generate the points' coordinates and confidence values, and another 2D CNN [46] to generate their initial features. This network was pre-trained on the DTU training dataset [1], and can generalize to other datasets and scenes. As shown in Fig. 4, the initial point cloud generated by such a network is coarse and noisy." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 168, + 287, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 168, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 46, + 168, + 287, + 239 + ], + "type": "text", + "content": "Explicit Optimization via Pruning and Growing. We perform a similar pruning and growing procedure as in PointNeRF, to prune outliers with low confidence " + }, + { + "bbox": [ + 46, + 168, + 287, + 239 + ], + "type": "inline_equation", + "content": "\\gamma_{j}" + }, + { + "bbox": [ + 46, + 168, + 287, + 239 + ], + "type": "text", + "content": " and fill holes in the point clouds. We make several important modifications over PointNeRF, and integrate this procedure with our deterministic integration (details in the supplementary)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 239, + 290, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 239, + 290, + 408 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 290, + 408 + ], + "type": "text", + "content": "Implicit Optimization with Normal Vectors. In addition to the explicit optimization, the point cloud is also optimized implicitly during training through the adjustment of point confidence " + }, + { + "bbox": [ + 46, + 239, + 290, + 408 + ], + "type": "inline_equation", + "content": "\\gamma_{j}" + }, + { + "bbox": [ + 46, + 239, + 290, + 408 + ], + "type": "text", + "content": ". When computing the average normal vector for rendering, we aggregate normal vectors of nearby points weighted with their distance and confidence, where the confidence of noisy or inaccurate points with potentially abnormal normal vectors is adjusted accordingly. Moreover, we apply the normal vector regularization losses from RefNeRF [37] to supervise the points' confidence w.r.t. their normal vectors. These strategies collectively provide implicit but more tailored ways to optimize the point clouds. With both explicit and implicit optimization, NeuralEditor obtains very precise point clouds (Fig. 4)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 417, + 232, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 417, + 232, + 430 + ], + "spans": [ + { + "bbox": [ + 47, + 417, + 232, + 430 + ], + "type": "text", + "content": "4. Shape Editing with NeuralEditor" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 437, + 287, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 287, + 498 + ], + "type": "text", + "content": "Formulation of General Shape Editing Tasks. We define the shape editing tasks based on indexed point clouds. To this end, we first re-define an indexed point cloud " + }, + { + "bbox": [ + 46, + 437, + 287, + 498 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 437, + 287, + 498 + ], + "type": "text", + "content": " as a mapping from a point index " + }, + { + "bbox": [ + 46, + 437, + 287, + 498 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 437, + 287, + 498 + ], + "type": "text", + "content": " to the corresponding point " + }, + { + "bbox": [ + 46, + 437, + 287, + 498 + ], + "type": "inline_equation", + "content": "p_j" + }, + { + "bbox": [ + 46, + 437, + 287, + 498 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 76, + 501, + 287, + 514 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 501, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 76, + 501, + 287, + 514 + ], + "type": "interline_equation", + "content": "P: j \\rightarrow p _ {j}, \\text {w h e r e} p _ {j} \\in \\mathbb {R} ^ {3}, j = 1, \\dots , | P |. \\tag {4}", + "image_path": "d38c4204b1a19357b4712e223dbf8a5285dba70e45ae66d88686e8a131820c2c.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 521, + 287, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 287, + 545 + ], + "type": "text", + "content": "A shape editing task is defined as another indexed point cloud " + }, + { + "bbox": [ + 46, + 521, + 287, + 545 + ], + "type": "inline_equation", + "content": "Q(P)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 52, + 549, + 287, + 563 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 549, + 287, + 563 + ], + "spans": [ + { + "bbox": [ + 52, + 549, + 287, + 563 + ], + "type": "interline_equation", + "content": "Q (P): j \\rightarrow q _ {j}, \\text {w h e r e} q _ {j} \\in \\mathbb {R} ^ {3} \\cup \\{\\varnothing \\}, j = 1, \\dots , | P |, \\tag {5}", + "image_path": "9957d69de0060844ae98c6a310cd814c32b7a66579579a0bec172a587961f94d.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 570, + 287, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 630 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 630 + ], + "type": "text", + "content": "describing a shape editing task whether the " + }, + { + "bbox": [ + 46, + 570, + 287, + 630 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 570, + 287, + 630 + ], + "type": "text", + "content": "-th point moves from " + }, + { + "bbox": [ + 46, + 570, + 287, + 630 + ], + "type": "inline_equation", + "content": "p_j" + }, + { + "bbox": [ + 46, + 570, + 287, + 630 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 570, + 287, + 630 + ], + "type": "inline_equation", + "content": "q_j" + }, + { + "bbox": [ + 46, + 570, + 287, + 630 + ], + "type": "text", + "content": " or is deleted in the deformation if " + }, + { + "bbox": [ + 46, + 570, + 287, + 630 + ], + "type": "inline_equation", + "content": "q_j = \\emptyset" + }, + { + "bbox": [ + 46, + 570, + 287, + 630 + ], + "type": "text", + "content": ". With this definition, " + }, + { + "bbox": [ + 46, + 570, + 287, + 630 + ], + "type": "inline_equation", + "content": "Q(P)" + }, + { + "bbox": [ + 46, + 570, + 287, + 630 + ], + "type": "text", + "content": " can be an arbitrary point cloud with points properly matched to points in " + }, + { + "bbox": [ + 46, + 570, + 287, + 630 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 570, + 287, + 630 + ], + "type": "text", + "content": " by same indices, regardless of connectivity or continuity." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "text", + "content": "Our formulation represents a broad range of shape editing tasks. The mesh editing tasks in NeRF-Editing [51], Deforming-NeRF [44], and CageNeRF [30] can be more simply and clearly defined here. For example, in NeRF-Editing, a mesh is exported from a general NeRF model, deformed manually, and converted to an \"offset\" or a continuous space deformation. We can depict such a task without" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 307, + 71, + 416, + 155 + ], + "blocks": [ + { + "bbox": [ + 307, + 71, + 416, + 155 + ], + "lines": [ + { + "bbox": [ + 307, + 71, + 416, + 155 + ], + "spans": [ + { + "bbox": [ + 307, + 71, + 416, + 155 + ], + "type": "image", + "image_path": "cfa7eafab6f5def118dad4763b9eb0193fc659617d8d194f6fa1ca2f094ba3f8.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 156, + 547, + 319 + ], + "lines": [ + { + "bbox": [ + 304, + 156, + 547, + 319 + ], + "spans": [ + { + "bbox": [ + 304, + 156, + 547, + 319 + ], + "type": "text", + "content": "Figure 5. Infinitesimal surface transformation (IST). (a) As the view-dependent colors are modeled as absolute viewing directions, they (solid arrows at the right) are different from the correct colors (dashed arrows at the right) after deformation. We solve this by (b) constructing a local coordinate system near the " + }, + { + "bbox": [ + 304, + 156, + 547, + 319 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 156, + 547, + 319 + ], + "type": "text", + "content": "-th point and (c) modeling IST from the edited scene to the original scene with the coordinate systems, so as to (d) redirect the viewing direction to the original scene when rendering the edited scene. \"offsets,\" by recording only the final location for each point without extra information. Notably, our formulation even models those whose deformation is not continuous in the space, e.g., cutting a scene into two parts, and thus cannot be covered and solved by NeRF-Editing, Deforming-NeRF, or CageNeRF, as shown in the supplementary." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 417, + 71, + 547, + 155 + ], + "blocks": [ + { + "bbox": [ + 417, + 71, + 547, + 155 + ], + "lines": [ + { + "bbox": [ + 417, + 71, + 547, + 155 + ], + "spans": [ + { + "bbox": [ + 417, + 71, + 547, + 155 + ], + "type": "image", + "image_path": "fca9b06889fcc3e42dc57debb9f4d2dab78274a0368ef6398e046a5d8115d411.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 319, + 545, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 319, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 319, + 545, + 437 + ], + "type": "text", + "content": "Editing Shape by Moving Points. We design our shape editing scheme with NeuralEditor. This is achieved by interpreting NeRF rendering as \"plotting\" the sampled points over the viewing ray. If we render the scene by naively plotting the point cloud " + }, + { + "bbox": [ + 304, + 319, + 545, + 437 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 319, + 545, + 437 + ], + "type": "text", + "content": ", the shape editing task can be addressed by replacing each point's coordinate from " + }, + { + "bbox": [ + 304, + 319, + 545, + 437 + ], + "type": "inline_equation", + "content": "p_j" + }, + { + "bbox": [ + 304, + 319, + 545, + 437 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 319, + 545, + 437 + ], + "type": "inline_equation", + "content": "q_j" + }, + { + "bbox": [ + 304, + 319, + 545, + 437 + ], + "type": "text", + "content": ". For NeuralEditor, we similarly replace the underlying point cloud from " + }, + { + "bbox": [ + 304, + 319, + 545, + 437 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 319, + 545, + 437 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 319, + 545, + 437 + ], + "type": "inline_equation", + "content": "Q(P)" + }, + { + "bbox": [ + 304, + 319, + 545, + 437 + ], + "type": "text", + "content": ", while maintaining the confidence values and features. This method is general and can also be applied to any point-based NeRF model like PointNeRF." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 438, + 545, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 438, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 438, + 545, + 510 + ], + "type": "text", + "content": "Correcting View-Dependence with Infinitesimal Surface Transformation (IST). The editing method above can already obtain reasonable results. However, as illustrated in Fig. 5, the modeled view-dependent colors record the absolute viewing direction, making them incorrect after deformations that change their orientation." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 510, + 546, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 546, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 546, + 665 + ], + "type": "text", + "content": "To solve this issue, we model the infinitesimal surface transformation (IST) for each point to redirect the viewing ray in the correct direction. We construct a local coordinate system for each point to represent the orientation of the infinitesimal surface, using its normal vector and two point indices that are neighbors of the " + }, + { + "bbox": [ + 304, + 510, + 546, + 665 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 510, + 546, + 665 + ], + "type": "text", + "content": "-th point in both " + }, + { + "bbox": [ + 304, + 510, + 546, + 665 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 510, + 546, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 510, + 546, + 665 + ], + "type": "inline_equation", + "content": "Q(P)" + }, + { + "bbox": [ + 304, + 510, + 546, + 665 + ], + "type": "text", + "content": ". By comparing these two coordinate systems, we can obtain an affine transformation " + }, + { + "bbox": [ + 304, + 510, + 546, + 665 + ], + "type": "inline_equation", + "content": "\\mathrm{IST}_j" + }, + { + "bbox": [ + 304, + 510, + 546, + 665 + ], + "type": "text", + "content": " for the " + }, + { + "bbox": [ + 304, + 510, + 546, + 665 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 510, + 546, + 665 + ], + "type": "text", + "content": "-th point to redirect the querying view direction (Fig. 5). This procedure is different from modeling space deformation [30, 44, 51], as we only need to model a simple affine transformation at each point, while those methods model a complicated, continuous, and non-linear deformation in the whole space." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": "Our proposed method requires a precise point cloud with normal vector-based color modeling. It is thus incompatible with PointNeRF, as PointNeRF is unable to obtain a desired point cloud to estimate the surface normal vectors." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "12443" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 239 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 239 + ], + "type": "text", + "content": "Fine-Tuning on Deformed Scene. Using the shape editing scheme introduced above, we can apply shape deformation on the scene modeled by our NeuralEditor without any modification to the model architecture or rendering scheme, which means that the resulting model is still a valid, fully functional NeuralEditor. Therefore, we can further fine-tune NeuralEditor on the deformed scene if the ground truth is available. We can even fine-tune the infinitesimal surface transformation with other parameters, to rapidly adjust toward better ambient consistency. This makes NeuralEditor desirable in practice, since in most applications, the final goal is not a zero-shot inference, but to fit the deformed scene well with reduced cost. By supporting fine-tuning, our NeuralEditor aligns well with and achieves this goal." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 239, + 290, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 239, + 290, + 431 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 290, + 431 + ], + "type": "text", + "content": "As another point-based NeRF model, PointNeRF supports fine-tuning but cannot leverage infinitesimal surface transformation fine-tuning to further optimize the performance. On the other hand, mesh-based NeRF editing models [30, 44, 51] do not support fine-tuning well: With deforming the space instead of the scene, these models' rendering scheme has highly changed. In their rendering process, a ray may go through a long, irregular way to reach the scene's surface. As the modeled space deformation might not be precise, it could be hard to tune the irregular space well, and even hurt other parts of the trained NeRF model. Such issues occur especially for some spaces with non-uniform density, since most of their model components (e.g., positional encodings, voxels) are not designed to deal with non-uniform spaces. Among all these methods, only our NeuralEditor has complete support for fine-tuning." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 442, + 124, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 442, + 124, + 455 + ], + "spans": [ + { + "bbox": [ + 47, + 442, + 124, + 455 + ], + "type": "text", + "content": "5. Experiment" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 462, + 287, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 462, + 287, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 462, + 287, + 617 + ], + "type": "text", + "content": "Point Cloud Generation. The underlying point cloud is fundamental to all editing tasks. Fig. 6 first provides a qualitative comparison of point clouds generated by our NeuralEditor and PointNeRF [43] on NeRF Synthetic [27]. Ours are much more precise with sharper details, e.g., the mayonnaise on the Hotdog's sausage, the uneven texture on the Chair's cushion, the edge of the Mic's stand, and the Lego brick's studs. By contrast, PointNeRF's point clouds are blurred and noisy, lose most of the details, and even contain obvious shape deflects on the Hotdog's plate and Chair's backrest. This shows that while the point cloud generation task is challenging, NeuralEditor generates a superprecise point cloud which is crucial for shape editing tasks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "content": "Experimental Settings. We mainly conduct experiments based on scenes from the NeRF Synthetic (NS) dataset. NS is a widely-used NeRF benchmark constructed from Blender [7] scenes. Due to the lack of publicly available benchmarks for shape editing, we use Blender to construct a reproducible benchmark, including the ground truth of edited scenes for evaluation and fine-tuning. Our shape editing tasks cover all eight scenes in NS, while prior" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 317, + 70, + 539, + 287 + ], + "blocks": [ + { + "bbox": [ + 317, + 70, + 539, + 287 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 539, + 287 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 539, + 287 + ], + "type": "image", + "image_path": "02a70b85b1e99d27d431dac987eb4762b95b169bb6a63eb8ecc38593f79e6b2c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 289, + 545, + 322 + ], + "lines": [ + { + "bbox": [ + 305, + 289, + 545, + 322 + ], + "spans": [ + { + "bbox": [ + 305, + 289, + 545, + 322 + ], + "type": "text", + "content": "Figure 6. NeuralEditor generates much more precise point clouds than PointNeRF [43] in the four scenes of NeRF Synthetic [27]. The points are colored with their normal vectors." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 307, + 327, + 545, + 376 + ], + "blocks": [ + { + "bbox": [ + 307, + 327, + 545, + 376 + ], + "lines": [ + { + "bbox": [ + 307, + 327, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 307, + 327, + 545, + 376 + ], + "type": "image", + "image_path": "bc6bc7668bbcdc03a475b7cb90437a3cb9c2b590519e5aae2ac1ef1ff30e88e1.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 376, + 545, + 399 + ], + "lines": [ + { + "bbox": [ + 306, + 376, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 306, + 376, + 545, + 399 + ], + "type": "text", + "content": "Figure 7. With too coarse cages, DeformingNeRF [44] is unable to perform the deformation faithfully, leading to poor results." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 402, + 545, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 402, + 545, + 473 + ], + "spans": [ + { + "bbox": [ + 305, + 402, + 545, + 473 + ], + "type": "text", + "content": "work [30,44,51] only picks a few scenes. The provided images for NS scenes are with opacity, and there is no requirement for the background color. We evaluate and visualize the results on a black background, for better contrast and clearer detail visualization. In the supplementary, we show the results on a white background with same conclusions." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 474, + 545, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 474, + 545, + 497 + ], + "spans": [ + { + "bbox": [ + 305, + 474, + 545, + 497 + ], + "type": "text", + "content": "Shape Editing Tasks. We evaluate our model on two types of shape editing tasks, as shown in Fig. 1:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 498, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 304, + 498, + 546, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 546, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 546, + 665 + ], + "type": "text", + "content": "(I) Shape (Mesh) Deformation Task. We consider the shape deformation task as in [30, 44, 51]: deform the shape of a scene in a human-guided way. To construct our deformation tasks from NS and obtain the ground truth, we apply the shape deformation simultaneously to the scene and our point cloud within the provided Blender file. We perform both zero-shot inference and fine-tuning, and compare our rendering results with the ground truth. As demonstrated in Figs. 7 and 8, our deformation tasks are much more precise and aggressive, compared with those in previous work [30, 44, 51]. In the supplementary, we also design a non-continuous deformation task and deformation tasks on the real-world dataset Tanks and Temples [17] (with zero-shot inference only, as there is no ground truth available)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 666, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 666, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 666, + 547, + 714 + ], + "type": "text", + "content": "(II) Scene Morphing Task. We address a more challenging shape editing task that has not been investigated in prior NeRF editing work: the scene morphing task. Given two scenes " + }, + { + "bbox": [ + 305, + 666, + 547, + 714 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 305, + 666, + 547, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 666, + 547, + 714 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 305, + 666, + 547, + 714 + ], + "type": "text", + "content": ", we should construct a path to gradually" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12444" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 71, + 545, + 134 + ], + "blocks": [ + { + "bbox": [ + 48, + 71, + 545, + 134 + ], + "lines": [ + { + "bbox": [ + 48, + 71, + 545, + 134 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 545, + 134 + ], + "type": "table", + "html": "
ModelZero-Shot Inference, PSNR ↑Fine-Tune for 1 Epoch, PSNR ↑
ChairHotdogLegoDrumsFicusMaterialsMicShipChairHotdogLegoDrumsFicusMaterialsMicShip
DeformingNeRF [44]18.84-13.10-------------
PointNeRF [43]22.2125.9524.5621.0024.2421.2126.7721.1930.1136.0831.4527.1631.4827.5534.3428.90
Naive Plotting24.9127.0125.6421.2926.2221.6527.6322.2932.0136.3831.7228.0933.2130.3135.1530.01
NeuralEditor w/o IST24.9227.0225.6521.2926.2421.6427.6422.2832.2436.6932.7928.3033.3430.4035.2830.08
NeuralEditor (Ours)25.8527.4927.4621.8427.1923.1827.7524.1632.5337.2232.9528.3533.5330.8235.4630.44
", + "image_path": "8dad0916d0c0ad591e03aa6c149a8319e832ceec95835714224d012874f9bcc7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 135, + 546, + 178 + ], + "lines": [ + { + "bbox": [ + 46, + 135, + 546, + 178 + ], + "spans": [ + { + "bbox": [ + 46, + 135, + 546, + 178 + ], + "type": "text", + "content": "Table 1. NeuralEditor significantly and consistently outperforms PointNeRF and Naive Plotting on all deformed scenes of NeRF Synthetic in peak signal-to-noise ratio (PSNR), in both zero-shot inference and fine-tuning settings. Our infinitesimal surface transformation (IST) effectively improves the results by correcting the view-dependent colors. With the precise point cloud generated by NeuralEditor, even Naive Plotting consistently outperforms PointNeRF. Comparison results under other metrics are in the supplementary." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 49, + 179, + 545, + 339 + ], + "blocks": [ + { + "bbox": [ + 49, + 179, + 545, + 339 + ], + "lines": [ + { + "bbox": [ + 49, + 179, + 545, + 339 + ], + "spans": [ + { + "bbox": [ + 49, + 179, + 545, + 339 + ], + "type": "image", + "image_path": "f74a14115e3021e730674e5b0bd209b5f0ffc0552d64ad1f60b5c780cebbc1f2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 342, + 546, + 363 + ], + "lines": [ + { + "bbox": [ + 46, + 342, + 546, + 363 + ], + "spans": [ + { + "bbox": [ + 46, + 342, + 546, + 363 + ], + "type": "text", + "content": "Figure 8. NeuralEditor produces superior rendering results to PointNeRF, with significantly fewer artifacts in zero-shot inference. Finetuning further improves the consistency of rendering with the ambient environment. We use a black background for better visualization." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 366, + 287, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 366, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 46, + 366, + 287, + 509 + ], + "type": "text", + "content": "change one scene to the other, and the intermediate scenes should have reasonable appearances. We are required to render all intermediate scenes. For this task, we use the point cloud diffusion model [24] to generate intermediate point clouds with latent space interpolation between " + }, + { + "bbox": [ + 46, + 366, + 287, + 509 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 46, + 366, + 287, + 509 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 366, + 287, + 509 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 46, + 366, + 287, + 509 + ], + "type": "text", + "content": " as in [24], and we introduce a K-D tree-based [4] matching algorithm to match the adjacent points to fix the indices of the intermediate scenes. To render an intermediate scene, we apply the shape transformation to the NeuralEditor models trained for scenes " + }, + { + "bbox": [ + 46, + 366, + 287, + 509 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 46, + 366, + 287, + 509 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 366, + 287, + 509 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 46, + 366, + 287, + 509 + ], + "type": "text", + "content": ", and then interpolate the rendering features to obtain the features of the intermediate scene for rendering." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 510, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 287, + 605 + ], + "type": "text", + "content": "NeuralEditor Variants. (1) Full NeuralEditor: Our complete model with all components. (2) NeuralEditor without infinitesimal surface transformation (IST): We remove the maintenance and optimization of infinitesimal surface transformation in scene editing. This key variant of NeuralEditor enables us to evaluate the importance of IST as well as other components of our model. The full ablation study of NeuralEditor is in the supplementary." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "content": "Baselines. We compare NeuralEditor against different types of baselines as follows. (1) Naive Plotting: We use the point cloud generated by NeuralEditor, computing each point's opacity and view-dependent colors with their point features. We render the scene by directly plotting/projecting the point cloud to the camera plane. (2) PointNeRF [43]: For the shape deformation task, we apply the same deformation to the point clouds generated by PointNeRF. For the scene morphing task, we apply the matching algorithm to" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 366, + 546, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 366, + 546, + 509 + ], + "spans": [ + { + "bbox": [ + 304, + 366, + 546, + 509 + ], + "type": "text", + "content": "the point clouds generated by PointNeRF and the same intermediate point clouds generated by the point cloud diffusion model [24] for fairness. (3) DeformingNeRF [44]: DeformingNeRF is not compatible with the scene morphing task. For the shape deformation task, we perform the same deformation on vertices of given cages. Note that DeformingNeRF only released trained models for Lego and Chair, so we can only evaluate it on these two scenes. While other models [30,51] support NeRF-based shape deformation via cages or exported meshes, we were unable to use them as baselines - they did not provide executable code nor their deformed scenes for us to evaluate on their tasks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 510, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 546, + 713 + ], + "type": "text", + "content": "Shape (Mesh) Deformation Results. The qualitative comparison is shown in Fig. 8. Both PointNeRF and Naive Plotting have many artifacts, like blur, wrong color, black or white shadows, noise, etc., whereas our powerful NeuralEditor produces clean and realistic rendering results. After fine-tuning, NeuralEditor shows a significant improvement with better rendering results than PointNeRF, indicating that NeuralEditor is able to achieve higher consistency with the surrounding ambient environment. Notably, in the Materials scene (the 6th scene from left), only our NeuralEditor generates reasonable reflection, while both baselines show blurry and visually messy results. Also in the Drums scene, the bottom face of the gong is not visible in any of the training views in the original scene, so all models render poor results in zero-shot inference. However, after only fast fine-tuning, NeuralEditor is able to precisely model the previously unknown surface and generate a sub-" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12445" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 94, + 70, + 503, + 345 + ], + "blocks": [ + { + "bbox": [ + 94, + 70, + 503, + 345 + ], + "lines": [ + { + "bbox": [ + 94, + 70, + 503, + 345 + ], + "spans": [ + { + "bbox": [ + 94, + 70, + 503, + 345 + ], + "type": "image", + "image_path": "e5fdd963b7c5b0e7575fa808d7bac345a35a0ce9533eb71732a090c11ce98b2f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 346, + 547, + 378 + ], + "lines": [ + { + "bbox": [ + 46, + 346, + 547, + 378 + ], + "spans": [ + { + "bbox": [ + 46, + 346, + 547, + 378 + ], + "type": "text", + "content": "Figure 9. Our NeuralEditor produces smooth morphing results between Chair, Hotdog, Lego, and Mic in the NeRF Synthetic dataset, while PointNeRF produces results with blurry textures, black shadows, and gloomy, non-smooth colors. The rendering results in the looped morphing process are arranged in the shape of the numerical digit \"3,\" indicated by the dividing lines and arrows." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 381, + 287, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 381, + 287, + 442 + ], + "spans": [ + { + "bbox": [ + 46, + 381, + 287, + 442 + ], + "type": "text", + "content": "stantially better result than PointNeRF, highlighting NeuralEditor's strength in fast-fitting. All these results demonstrate that NeuralEditor can handle various visual effects and make them consistent in the deformed scene. We provide the figure with higher resolution in the supplementary." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 445, + 287, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 445, + 287, + 577 + ], + "spans": [ + { + "bbox": [ + 46, + 445, + 287, + 577 + ], + "type": "text", + "content": "The quantitative comparison is summarized in Table 1. We observe that: (1) Our NeuralEditor consistently outperforms all the baselines and variants for both zero-shot inference and fine-tuning settings. (2) With the precise point cloud generated by NeuralEditor, the Naive Plotting baseline even consistently outperforms PointNeRF. (3) Our 'w/o IST' variant has a comparable performance to Naive Plotting with the same point cloud and features in the zero-shot inference setting, but after fine-tuning its performance is significantly higher than Naive Plotting, validating the capability of NeuralEditor in NeRF modeling." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 578, + 287, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 578, + 287, + 687 + ], + "spans": [ + { + "bbox": [ + 46, + 578, + 287, + 687 + ], + "type": "text", + "content": "Notably, DeformingNeRF [44] performs poorly in our benchmark with significantly lower metric values. As shown in Fig. 7, the cages provided by DeformingNeRF are too coarse, and cannot even cover the whole scene. Therefore, DeformingNeRF cannot faithfully perform the precise deformation in our benchmark, leading to poor rendering results. On the contrary, both PointNeRF and our NeuralEditor at least faithfully perform the deformation, showing that point cloud is necessary for precise shape editing." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "Scene Morphing Results. The morphing results between 4 NeRF Synthetic scenes are shown in Fig. 9. The" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 382, + 545, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 382, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 304, + 382, + 545, + 502 + ], + "type": "text", + "content": "morphing process starts from Chair, morphs to Hotdog, Lego, Mic, and at last turns back to Chair. NeuralEditor produces smooth rendering results on the point cloud diffusion-guided [24] intermediate scenes, mixing the textures of the two scenes in a reasonable way. In comparison, the rendering results produced by PointNeRF are unsatisfactory, with blurry textures, black shadows, and gloomy, non-smooth colors. These results show that our NeuralEditor can render challenging intermediate morphing scenes and achieve decent results with only the input of moved points." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 510, + 378, + 522 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 510, + 378, + 522 + ], + "spans": [ + { + "bbox": [ + 306, + 510, + 378, + 522 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 528, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 528, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 528, + 545, + 713 + ], + "type": "text", + "content": "This paper proposes NeuralEditor, a point cloud-guided NeRF model that supports general shape editing tasks by manipulating the underlying point clouds. Empirical evaluation shows NeuralEditor to produce rendering results of much higher quality than baselines in a zero-shot inference manner, further significantly improving after fast finetuning. NeuralEditor even supports smooth scene morphing between multiple scenes, which is difficult for prior work. We hope that our work can inspire more research on point cloud-guided NeRFs and 3D shape and scene editing tasks. Acknowledgement. This work was supported in part by NSF Grant 2106825, NIFA Award 2020-67021-32799, the Jump ARCHES endowment, the NCSA Fellows program, the IBM-Illinois Discovery Accelerator Institute, the Illinois-Inspor Partnership, and the Amazon Research Award. This work used NVIDIA GPUs at NCSA Delta through allocation CIS220014 from the ACCESS program. We thank the authors of NeRF [27] for their help in processing Blender files of the NS dataset." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "12446" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "text", + "content": "[1] Henrik Aanaes, Rasmus Ramsbøl Jensen, George Vogiatzis, Engin Tola, and Anders Bjorholm Dahl. Large-scale data for multiple-view stereopsis. International Journal of Computer Vision, 120(2):153-168, 2016. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 135, + 287, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 135, + 287, + 168 + ], + "spans": [ + { + "bbox": [ + 53, + 135, + 287, + 168 + ], + "type": "text", + "content": "[2] Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas J. Guibas. Learning representations and generative models for 3D point clouds. In ICML, 2018. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 168, + 288, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 168, + 288, + 213 + ], + "spans": [ + { + "bbox": [ + 53, + 168, + 288, + 213 + ], + "type": "text", + "content": "[3] Jonathan T. Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P. Srinivasan. Mip-NeRF: A multiscale representation for antiailiasing neural radiance fields. In ICCV, 2021. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 213, + 287, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 213, + 287, + 246 + ], + "spans": [ + { + "bbox": [ + 53, + 213, + 287, + 246 + ], + "type": "text", + "content": "[4] Jon Louis Bentley. Multidimensional binary search trees used for associative searching. Commun. ACM, 18(9):509-517, 1975. 2, 3, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 246, + 287, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 246, + 287, + 290 + ], + "spans": [ + { + "bbox": [ + 53, + 246, + 287, + 290 + ], + "type": "text", + "content": "[5] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. MVSNeRF: Fast generalizable radiance field reconstruction from multi-view stereo. In ICCV, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 290, + 287, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 290, + 287, + 313 + ], + "spans": [ + { + "bbox": [ + 53, + 290, + 287, + 313 + ], + "type": "text", + "content": "[6] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In CVPR, 2019. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 313, + 288, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 313, + 288, + 346 + ], + "spans": [ + { + "bbox": [ + 53, + 313, + 288, + 346 + ], + "type": "text", + "content": "[7] Blender Online Community. *Blender - a 3D modelling and rendering package*. Blender Foundation, Stichting Blender Foundation, Amsterdam, 2018. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 346, + 287, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 346, + 287, + 368 + ], + "spans": [ + { + "bbox": [ + 53, + 346, + 287, + 368 + ], + "type": "text", + "content": "[8] Frank Dellaert and Lin Yen-Chen. Neural volume rendering: NeRF and beyond. arXiv:2101.05204, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 369, + 287, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 369, + 287, + 402 + ], + "spans": [ + { + "bbox": [ + 53, + 369, + 287, + 402 + ], + "type": "text", + "content": "[9] Kyle Gao, Yina Gao, Hongjie He, Denning Lu, Linlin Xu, and Jonathan Li. NeRF: Neural radiance field in 3D vision, a comprehensive review. arXiv:2101.05204, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 403, + 287, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 403, + 287, + 435 + ], + "spans": [ + { + "bbox": [ + 48, + 403, + 287, + 435 + ], + "type": "text", + "content": "[10] Yuan-Chen Guo, Di Kang, Linchao Bao, Yu He, and Song-Hai Zhang. NeRFReN: Neural radiance fields with reflections. In CVPR, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 435, + 287, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 435, + 287, + 468 + ], + "spans": [ + { + "bbox": [ + 48, + 435, + 287, + 468 + ], + "type": "text", + "content": "[11] Peter Hedman, Tobias Ritschel, George Drettakis, and Gabriel Brostow. Scalable inside-out image-based rendering. ACM Trans. Graph., 35(6):231:1-231:11, 2016. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 468, + 287, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 468, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 48, + 468, + 287, + 502 + ], + "type": "text", + "content": "[12] Po-Han Huang, Kevin Matzen, Johannes Kopf, Narendra Ahuja, and Jia-Bin Huang. DeepMVS: Learning multi-view stereopsis. In CVPR, 2018. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 502, + 287, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 502, + 287, + 535 + ], + "spans": [ + { + "bbox": [ + 48, + 502, + 287, + 535 + ], + "type": "text", + "content": "[13] Alec Jacobson, Ilya Baran, Ladislav Kavan, Jovan Popovic, and Olga Sorkine. Fast automatic skinning transformations. ACM Trans. Graph., 31(4), 2012. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 536, + 287, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 287, + 579 + ], + "type": "text", + "content": "[14] Tomas Jakab, Richard Tucker, Ameesh Makadia, Jiajun Wu, Noah Snavely, and Angjoo Kanazawa. Keypointdeformer: Unsupervised 3D keypoint discovery for shape control. In CVPR, 2021. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 579, + 287, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 579, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 579, + 287, + 613 + ], + "type": "text", + "content": "[15] Mengqi Ji, Juergen Gall, Haitian Zheng, Yebin Liu, and Lu Fang. Surfacenet: An end-to-end 3D neural network for multiview stereopsis. In ICCV, 2017. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 613, + 287, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 646 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 646 + ], + "type": "text", + "content": "[16] Tao Ju, Scott Schaefer, and Joe Warren. Mean value coordinates for closed triangular meshes. ACM Trans. Graph., 24(3):561-566, 2005. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 646, + 287, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 646, + 287, + 680 + ], + "spans": [ + { + "bbox": [ + 48, + 646, + 287, + 680 + ], + "type": "text", + "content": "[17] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and Temples: Benchmarking large-scale scene reconstruction. ACM Trans. Graph., 36(4), 2017. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[18] Sosuke Kobayashi, Eiichi Matsumoto, and Vincent Sitzmann. Decomposing NeRF for editing via feature field distillation. In NeurlPS, 2022. 2" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 37, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "type": "text", + "content": "[19] Marc Levoy and Pat Hanrahan. Light field rendering. In SIGGRAPH, 1996. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 95, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 95, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 95, + 545, + 150 + ], + "type": "text", + "content": "[20] Fayao Liu, Chunhua Shen, Guosheng Lin, and Ian Reid. Learning depth from single monocular images using deep convolutional neural fields. IEEE Transactions on Pattern Analysis and Machine Intelligence, 38(10):2024-2039, 2015. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 151, + 545, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 151, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 307, + 151, + 545, + 185 + ], + "type": "text", + "content": "[21] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. In NeurIPS, 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 186, + 545, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 220 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 220 + ], + "type": "text", + "content": "[22] Steven Liu, Xiuming Zhang, Zhoutong Zhang, Richard Zhang, Junyan Zhu, and Bryan C. Russell. Editing conditional radiance fields. In ICCV, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 221, + 545, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 221, + 545, + 255 + ], + "spans": [ + { + "bbox": [ + 307, + 221, + 545, + 255 + ], + "type": "text", + "content": "[23] William E. Lorensen and Harvey E. Cline. Marching cubes: A high resolution 3D surface construction algorithm. In SIGGRAPH, 1987. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 255, + 545, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 255, + 545, + 278 + ], + "spans": [ + { + "bbox": [ + 307, + 255, + 545, + 278 + ], + "type": "text", + "content": "[24] Shitong Luo and Wei Hu. Diffusion probabilistic models for 3D point cloud generation. In CVPR, 2021. 2, 7, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 279, + 545, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 279, + 545, + 313 + ], + "spans": [ + { + "bbox": [ + 307, + 279, + 545, + 313 + ], + "type": "text", + "content": "[25] Bruce Merry, Patrick Marais, and James Gain. Animation space: A truly linear framework for character animation. ACM Trans. Graph., 25(4):1400-1423, 2006. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 313, + 545, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 313, + 545, + 357 + ], + "spans": [ + { + "bbox": [ + 307, + 313, + 545, + 357 + ], + "type": "text", + "content": "[26] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3D reconstruction in function space. In CVPR, 2019. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 358, + 545, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 358, + 545, + 404 + ], + "spans": [ + { + "bbox": [ + 307, + 358, + 545, + 404 + ], + "type": "text", + "content": "[27] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2, 4, 6, 8" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 404, + 545, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 404, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 307, + 404, + 545, + 449 + ], + "type": "text", + "content": "[28] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Differentiable volumetric rendering: Learning implicit 3D representations without 3D supervision. In CVPR, 2020. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 450, + 545, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 450, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 307, + 450, + 545, + 483 + ], + "type": "text", + "content": "[29] Jesús Nieto and Toni Susin. Cage based deformations: A survey. Lecture Notes in Computational Vision and Biomechanics, 7:75-99, 2013. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 484, + 545, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 484, + 545, + 540 + ], + "spans": [ + { + "bbox": [ + 307, + 484, + 545, + 540 + ], + "type": "text", + "content": "[30] Yicong Peng, Yichao Yan, Shengqi Liu, Yuhao Cheng, Shanyan Guan, Bowen Pan, Guangtao Zhai, and Xiaokang Yang. CageNeRF: Cage-based neural radiance field for generalized 3D deformation and animation. In NeurIPS, 2022. 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 541, + 545, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 541, + 545, + 563 + ], + "spans": [ + { + "bbox": [ + 307, + 541, + 545, + 563 + ], + "type": "text", + "content": "[31] Bui Tuong Phong. Illumination for computer generated pictures. Commun. ACM, 18(6):311-317, 1975. 2, 4" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 563, + 545, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 563, + 545, + 598 + ], + "spans": [ + { + "bbox": [ + 307, + 563, + 545, + 598 + ], + "type": "text", + "content": "[32] Yi-Ling Qiao, Alexander Gao, and Ming C. Lin. NeuPhysics: Editable neural geometry and physics from monocular videos. In NeurIPS, 2022. 2" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 307, + 598, + 545, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 598, + 545, + 643 + ], + "spans": [ + { + "bbox": [ + 307, + 598, + 545, + 643 + ], + "type": "text", + "content": "[33] Sara Fridovich-Keil and Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In CVPR, 2022. 1, 2" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 307, + 644, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 644, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 644, + 545, + 678 + ], + "type": "text", + "content": "[34] Noah Snavely, Steven M. Seitz, and Richard Szeliski. Photo tourism: Exploring photo collections in 3D. ACM Trans. Graph., 25(3):835-846, 2006. 2" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 307, + 678, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 678, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 678, + 545, + 713 + ], + "type": "text", + "content": "[35] Jiaxiang Tang, Xiaokang Chen, Jingbo Wang, and Gang Zeng. Compressible-composable NeRF via rank-residual decomposition. In NeurIPS, 2022. 2" + } + ] + } + ], + "index": 36 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "12447" + } + ] + } + ], + "index": 38 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "type": "text", + "content": "[36] Jean-Marc Thiery, Julien Tierny, and Tamy Boubekeur. Jacobians and hessian of mean value coordinates for closed triangular meshes. The Visual Computer, 30(9):981-995, 2014. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 163 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 163 + ], + "type": "text", + "content": "[37] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T. Barron, and Pratul P. Srinivasan. Ref-NeRF: Structured view-dependent appearance for neural radiance fields. In CVPR, 2022. 1, 2, 4, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 164, + 287, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 164, + 287, + 198 + ], + "spans": [ + { + "bbox": [ + 48, + 164, + 287, + 198 + ], + "type": "text", + "content": "[38] Nanyang Wang, Yinda Zhang, Zhuwen Li, Yanwei Fu, Wei Liu, and Yu-Gang Jiang. Pixel2Mesh: Generating 3D mesh models from single RGB images. In ECCV, 2018. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 199, + 287, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 199, + 287, + 253 + ], + "spans": [ + { + "bbox": [ + 48, + 199, + 287, + 253 + ], + "type": "text", + "content": "[39] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul Srinivasan, Howard Zhou, Jonathan T. Barron, Ricardo MartinBrualla, Noah Snavely, and Thomas Funkhouser. IBRNet: Learning multi-view image-based rendering. In CVPR, 2021. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 255, + 287, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 255, + 287, + 298 + ], + "spans": [ + { + "bbox": [ + 48, + 255, + 287, + 298 + ], + "type": "text", + "content": "[40] Suttisak Wizadwongsa, Pakkapon Phongthawee, Jiraphon Yenphraphai, and Supasorn Suwajanakorn. NeX: Real-time view synthesis with neural basis expansion. In CVPR, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 301, + 287, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 301, + 287, + 345 + ], + "spans": [ + { + "bbox": [ + 48, + 301, + 287, + 345 + ], + "type": "text", + "content": "[41] Liwen Wu, Jae Yong Lee, Anand Bhattachad, Yu-Xiong Wang, and David Forsyth. DIVeR: Real-time and accurate neural radiance fields with deterministic integration for volume rendering. In CVPR, 2022. 2, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 346, + 287, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 346, + 287, + 379 + ], + "spans": [ + { + "bbox": [ + 48, + 346, + 287, + 379 + ], + "type": "text", + "content": "[42] Dejia Xu, Peihao Wang, Yifan Jiang, Zhiwen Fan, and Zhangyang Wang. Signal processing for implicit neural representations. In NeurIPS, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 381, + 287, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 381, + 287, + 415 + ], + "spans": [ + { + "bbox": [ + 48, + 381, + 287, + 415 + ], + "type": "text", + "content": "[43] Qiangeng Xu, Zexiang Xu, Julien Philip, Sai Bi, Zhixin Shu, Kalyan Sunkavalli, and Ulrich Neumann. Point-NeRF: Point-based neural radiance fields. In CVPR, 2021. 2, 3, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 416, + 287, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 416, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 48, + 416, + 287, + 437 + ], + "type": "text", + "content": "[44] Tianhan Xu and Tatsuya Harada. Deforming radiance fields with cages. In ECCV, 2022. 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 440, + 287, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 440, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 48, + 440, + 287, + 483 + ], + "type": "text", + "content": "[45] Bangbang Yang, Yinda Zhang, Yinghao Xu, Yijin Li, Han Zhou, Hujun Bao, Guofeng Zhang, and Zhaopeng Cui. Learning object-compositional neural radiance field for editing scene rendering. In ICCV, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 485, + 287, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 485, + 287, + 517 + ], + "spans": [ + { + "bbox": [ + 48, + 485, + 287, + 517 + ], + "type": "text", + "content": "[46] Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. MVSNet: Depth inference for unstructured multiview stereo. In ECCV, 2018. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 520, + 287, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 520, + 287, + 553 + ], + "spans": [ + { + "bbox": [ + 48, + 520, + 287, + 553 + ], + "type": "text", + "content": "[47] Wang Yifan, Noam Aigerman, Vladimir G Kim, Siddhartha Chaudhuri, and Olga Sorkine-Hornung. Neural cages for detail-preserving 3D deformations. In CVPR, 2020. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 555, + 287, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 555, + 287, + 587 + ], + "spans": [ + { + "bbox": [ + 48, + 555, + 287, + 587 + ], + "type": "text", + "content": "[48] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. PlenOctrees for real-time rendering of neural radiance fields. In ICCV, 2021. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 590, + 287, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 621 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 621 + ], + "type": "text", + "content": "[49] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelNeRF: Neural radiance fields from one or few images. In CVPR, 2021. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "type": "text", + "content": "[50] Yu-Jie Yuan, Yu-Kun Lai, Tong Wu, Lin Gao, and Ligang Liu. A revisit of shape editing techniques: From the geometric to the neural viewpoint. Journal of Computer Science and Technology, 36(3):520-554, 2021. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[51] Yu-Jie Yuan, Yang tian Sun, Yu-Kun Lai, Yuewen Ma, Rongfei Jia, and Lin Gao. NeRF-Editing: Geometry editing of neural radiance fields. In CVPR, 2022. 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 546, + 139 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 307, + 72, + 546, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 546, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 546, + 106 + ], + "type": "text", + "content": "[52] Yuzhe Zhang, Jianmin Zheng, and Yiyu Cai. Proxy-driven free-form deformation by topology-adjustable control lattice. Computers & Graphics, 89:167-177, 2020. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 107, + 546, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 546, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 546, + 139 + ], + "type": "text", + "content": "[53] Qian-Yi Zhou, Jaesik Park, and Vladlen Koltun. Open3D: A modern library for 3D data processing. arXiv:1801.09847, 2018. 4" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "12448" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/NeuralField-LDM_ Scene Generation With Hierarchical Latent Diffusion Models/81bfb70e-1652-4684-aa6c-4729296b6ed4_content_list.json b/2023/NeuralField-LDM_ Scene Generation With Hierarchical Latent Diffusion Models/81bfb70e-1652-4684-aa6c-4729296b6ed4_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..049dc9381bc8ff65b9a564fcc402488dd1269f10 --- /dev/null +++ b/2023/NeuralField-LDM_ Scene Generation With Hierarchical Latent Diffusion Models/81bfb70e-1652-4684-aa6c-4729296b6ed4_content_list.json @@ -0,0 +1,1685 @@ +[ + { + "type": "text", + "text": "NeuralField-LDM: Scene Generation with Hierarchical Latent Diffusion Models", + "text_level": 1, + "bbox": [ + 80, + 108, + 888, + 128 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Seung Wook Kim $^{1,2,3*}$ Bradley Brown $^{1,5*†}$ Kangxue Yin $^{1}$ Karsten Kreis $^{1}$ Katja Schwarz $^{6†}$", + "bbox": [ + 86, + 157, + 883, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Daiqing Li", + "bbox": [ + 189, + 176, + 285, + 195 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Robin Rombach $^{7\\dagger}$", + "bbox": [ + 326, + 176, + 468, + 195 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Antonio Torralba", + "bbox": [ + 511, + 176, + 656, + 195 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sanja Fidler1,2,3", + "bbox": [ + 697, + 176, + 821, + 195 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ NVIDIA $^{2}$ University of Toronto $^{3}$ Vector Institute $^{4}$ CSAIL, MIT $^{5}$ University of Waterloo", + "bbox": [ + 197, + 196, + 779, + 210 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "6University of Tübingen, Tübingen AI Center 7LMU Munich", + "bbox": [ + 300, + 213, + 676, + 229 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fa25ef0555bea1ba507b5b1f853b4ad8e1acf043ea8867875ede42fa25075a26.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 234, + 890, + 345 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a7c0c2a5c50443499d118db2c92fd0798cee8f89ece0862a8e9153c7b05d74a0.jpg", + "image_caption": [ + "Figure 1. We introduce NeuralField-LDM, a generative model for complex open-world 3D scenes. This figure contains a panorama constructed from NeuralField-LDM's generated scene. We visualize different parts of the scene by placing cameras on them." + ], + "image_footnote": [], + "bbox": [ + 80, + 345, + 228, + 421 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/22cf9cbbf8ae11a68b0fd11712a7c9faabafac4044bd2b69985eb6d8ec6d2165.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 243, + 345, + 395, + 421 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/dbb85e17ddc79672b02f433511e29dfa4d317a403bd2192d0ce4b6c948398a3e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 410, + 345, + 560, + 421 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f816f399f8022c2169295cabfa1b364ea4ffb738f2b78f8bcf11db12fb5001df.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 575, + 345, + 725, + 421 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c4102ba302da853b621c88a79df028cebaabdc3234d53438413f2ac55e858c6b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 740, + 344, + 890, + 421 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 463, + 313, + 478 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Automatically generating high-quality real world 3D scenes is of enormous interest for applications such as virtual reality and robotics simulation. Towards this goal, we introduce NeuralField-LDM, a generative model capable of synthesizing complex 3D environments. We leverage Latent Diffusion Models that have been successfully utilized for efficient high-quality 2D content creation. We first train a scene auto-encoder to express a set of image and pose pairs as a neural field, represented as density and feature voxel grids that can be projected to produce novel views of the scene. To further compress this representation, we train a latent-autoencoder that maps the voxel grids to a set of latent representations. A hierarchical diffusion model is then fit to the latents to complete the scene generation pipeline. We achieve a substantial improvement over existing state-of-the-art scene generation models. Additionally, we show how NeuralField-LDM can be used for a variety of 3D content creation applications, including conditional scene generation, scene inpainting and scene style manipulation.", + "bbox": [ + 75, + 494, + 473, + 782 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 810, + 209, + 825 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "There has been increasing interest in modelling 3D real-world scenes for use in virtual reality, game design, digi-", + "bbox": [ + 76, + 835, + 468, + 866 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tal twin creation and more. However, designing 3D worlds by hand is a challenging and time-consuming process, requiring 3D modeling expertise and artistic talent. Recently, we have seen success in automating 3D content creation via 3D generative models that output individual object assets [15,46,73]. Although a great step forward, automating the generation of real-world scenes remains an important open problem and would unlock many applications ranging from scalably generating a diverse array of environments for training AI agents (e.g. autonomous vehicles) to the design of realistic open-world video games. In this work, we take a step towards this goal with NeuralField-LDM (NF-LDM), a generative model capable of synthesizing complex real-world 3D scenes. NF-LDM is trained on a collection of posed camera images and depth measurements which are easier to obtain than explicit ground-truth 3D data, offering a scalable way to synthesize 3D scenes.", + "bbox": [ + 500, + 464, + 892, + 720 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent approaches [2, 6, 8] tackle the same problem of generating 3D scenes, albeit on less complex data. In [6, 8], a latent distribution is mapped to a set of scenes using adversarial training, and in GAUDI [2], a denoising diffusion model is fit to a set of scene latents learned using an auto-decoder. These models all have an inherent weakness of attempting to capture the entire scene into a single vector that conditions a neural radiance field. In practice, we find that this limits the ability to fit complex scene distributions.", + "bbox": [ + 496, + 727, + 893, + 864 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, diffusion models have emerged as a very powerful class of generative models, capable of generating high-", + "bbox": [ + 500, + 869, + 892, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution.", + "bbox": [ + 99, + 875, + 212, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "†Work done during an internship at NVIDIA.", + "bbox": [ + 102, + 887, + 341, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "8496", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "quality images, point clouds and videos [18, 24, 39, 48, 53, 73, 78]. Yet, due to the nature of our task, where image data must be mapped to a shared 3D scene without an explicit ground truth 3D representation, straightforward approaches fitting a diffusion model directly to data are infeasible.", + "bbox": [ + 75, + 90, + 467, + 165 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In NeuralField-LDM, we learn to model scenes using a three-stage pipeline. First, we learn an auto-encoder that encodes scenes into a neural field, represented as density and feature voxel grids. Inspired by the success of latent diffusion models for images [53], we learn to model the distribution of our scene voxels in latent space to focus the generative capacity on core parts of the scene and not the extraneous details captured by our voxel auto-encoders. Specifically, a latent-autoencoder decomposes the scene voxels into a 3D coarse, 2D fine and 1D global latent. Hierarchical diffusion models are then trained on the tri-latent representation to generate novel 3D scenes. We show how NF-LDM enables applications such as scene editing, birdeye view conditional generation and style adaptation. Finally, we demonstrate how score distillation [46] can be used to optimize the quality of generated neural fields, allowing us to leverage the representations learned from state-of-the-art image diffusion models that have been exposed to orders of magnitude more data.", + "bbox": [ + 75, + 167, + 467, + 450 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are: 1) We introduce NF-LDM, a hierarchical diffusion model capable of generating complex open-world 3D scenes and achieving state of the art scene generation results on four challenging datasets. 2) We extend NF-LDM to semantic birds-eye view conditional scene generation, style modification and 3D scene editing.", + "bbox": [ + 75, + 454, + 467, + 544 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 551, + 217, + 566 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2D Generative Models In past years, generative adversarial networks (GANs) [3, 17, 27, 41, 58] and likelihood-based approaches [33, 49, 51, 67] enabled high-resolution photorealistic image synthesis. Due to their quality, GANs are used in a multitude of downstream applications ranging from steerable content creation [30, 34, 36, 37, 61, 76] to data driven simulation [26, 31, 32, 34]. Recently, autoregressive models and score-based models, e.g. diffusion models, demonstrate better distribution coverage while preserving high sample quality [10, 11, 13, 20, 22, 43, 48, 53, 54, 68]. Since evaluation and optimization of these approaches in pixel space is computationally expensive, [53, 68] apply them to latent space, achieving state-of-the-art image synthesis at megapixel resolution. As our approach operates on 3D scenes, computational efficiency is crucial. Hence, we build upon [53] and train our model in latent space.", + "bbox": [ + 75, + 577, + 467, + 816 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Novel View Synthesis In their seminal work [42], Mildenhall et al. introduce Neural Radiance Fields (NeRF) as a powerful 3D representation. PixelNeRF [72] and IBR-Net [70] propose to condition NeRF on aggregated features from multiple views to enable novel view synthesis from", + "bbox": [ + 75, + 824, + 467, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "a sparse set of views. Another line of works scale NeRF to large-scale indoor and outdoor scenes [40, 50, 74, 75]. Recently, Nerfusion [75] predicts local radiance fields and fuses them into a scene representation using a recurrent neural network. Similarly, we construct a latent scene representation by aggregating features across multiple views. Different from the aforementioned methods, our approach is a generative model capable of synthesizing novel scenes.", + "bbox": [ + 496, + 90, + 890, + 212 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D Diffusion Models A few recent works propose to apply denoising diffusion models (DDM) [20, 22, 63] on point clouds for 3D shape generation [39, 73, 78]. While PVD [78] trains on point clouds directly, DPM [39] and LION [73] use a shape latent variable. Similar to LION, we design a hierarchical model by training separate conditional DDMs. However, our approach generates both texture and geometry of a scene without needing 3D ground truth as supervision.", + "bbox": [ + 496, + 218, + 890, + 339 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D-Aware Generative Models 3D-aware generative models synthesize images while providing explicit control over the camera pose and potentially other scene properties, like object shape and appearance. SGAM [62] generates a 3D scene by autoregressively generating sensor data and building a 3D map. Several previous approaches generate NeRFs of single objects with conditional coordinate-based MLPs [7, 44, 59]. GSN [8] conditions a coordinate-based MLP on a \"floor plan\", i.e., a 2D feature map, to model more complex indoor scenes. EG3D [6] and VoxGRAF [60] use convolutional backbones to generate 3D representations. All of these approaches rely on adversarial training. Instead, we train a DDM on voxels in latent space. The work closest to ours is GAUDI [2], which first trains an auto-decoder and subsequently trains a DDM on the learned latent codes. Instead of using a global latent code, we encode scenes onto voxel grids and train a hierarchical DDM to optimally combine global and local features.", + "bbox": [ + 496, + 345, + 890, + 617 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. NeuralField-LDM", + "text_level": 1, + "bbox": [ + 500, + 632, + 676, + 647 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our objective is to train a generative model to synthesize 3D scenes that can be rendered to any viewpoint. We assume access to a dataset $\\{(i,\\kappa ,\\rho)\\}_{1..N}$ which consists of $N$ RGB images $i$ and their camera poses $\\kappa$ , along with a depth measurement $\\rho$ that can be either sparse (e.g. Lidar points) or dense. The generative model must learn to model both the texture and geometry distributions of the dataset in 3D by learning solely from the sensor observations, which is a highly non-trivial problem.", + "bbox": [ + 496, + 657, + 890, + 794 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Past work typically tackles this problem with a generative adversarial network (GAN) framework [6, 8, 59, 60]. They produce an intermediate 3D representation and render images for a given viewpoint with volume rendering [25, 42]. Discriminator losses then ensure that the 3D representation produces a valid image from any viewpoint. However, GANs come with notorious training instability", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "8497", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b1a3e5073629a6e122828874c18ea1f3ac41f1f030a4b040e2bbff35986fe11f.jpg", + "image_caption": [ + "Figure 2. Overview of NeuralField-LDM. We first encode RGB images with camera poses into a neural field represented by density and feature voxel grids. We compress the neural field into smaller latent spaces and fit a hierarchical latent diffusion model on the latent space. Sampled latents can then be decoded into a neural field that can be rendered into a given viewpoint." + ], + "image_footnote": [], + "bbox": [ + 122, + 84, + 843, + 224 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and mode dropping behaviors [1, 16, 35]. Denoising Diffusion models [20] (DDMs) have recently emerged as an alternative to GANs that avoid the aforementioned disadvantages [53, 56, 57]. However, DDMs model the data likelihood explicitly and are trained to reconstruct the training data. Thus, they have been used in limited scenarios [73, 77] since ground-truth 3D data is not readily available at scale.", + "bbox": [ + 75, + 282, + 468, + 388 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To tackle the challenging problem of generating an entire scene with texture and geometry, we take inspiration from latent diffusion models (LDM) [53], which first construct an intermediate latent distribution of the training data then fit a diffusion model on the latent distribution. In Sec. 3.1, we introduce a scene auto-encoder that encodes the set of RGB images into a neural field representation consisting of density and feature voxel grids. To accurately capture a scene, the voxel grids' spatial dimension needs to be much larger than what current state-of-the-art LDMs can model. In Sec. 3.2, we show how we can further compress and decompose the explicit voxel grids into compressed latent representations to facilitate learning the data distribution. Finally, Sec. 3.3 introduces a latent diffusion model that models the latent distributions in a hierarchical manner. Fig. 2 shows an overview of our method, which we name NeuralField-LDM (NF-LDM). We provide training and additional architecture details in the supplementary.", + "bbox": [ + 75, + 388, + 470, + 661 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Scene Auto-Encoder", + "text_level": 1, + "bbox": [ + 76, + 666, + 269, + 680 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The goal of the scene auto-encoder is to obtain a 3D representation of the scene from input images by learning to reconstruct them. Fig. 3 depicts the auto-encoding process. The scene encoder is a 2D CNN and processes each RGB image $i_{1..N}$ separately, producing a $\\mathbb{R}^{H\\times W\\times (D + C)}$ dimensional 2D tensor for each image, where $H$ and $W$ are smaller than $i$ 's size. We follow a similar procedure to Lift-Splat-Shoot (LSS) [45] to lift each 2D image feature map and combine them in the common voxel-based 3D neural field. We build a discrete frustum of size $H\\times W\\times D$ with the camera poses $\\kappa$ for each image. This frustum contains image features and density values for each pixel, along a pre-defined discrete set of $D$ depths. Unlike LSS, we take the first $D$ channels of the 2D CNN's output and use them as", + "bbox": [ + 75, + 688, + 470, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "density values. That is, the $d$ th channel of the CNN's output at pixel $(h,w)$ becomes the density value of the frustum entry at $(h,w,d)$ . Motivated by the volume rendering equation [42], we get the occupancy weight $O$ of each element $(h,w,d)$ in the frustum using the density values $\\sigma \\geq 0$ :", + "bbox": [ + 496, + 282, + 890, + 361 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nO (h, w, d) = \\exp \\left(- \\sum_ {j = 0} ^ {d - 1} \\sigma_ {(h, w, j)} \\delta_ {j}\\right) \\left(1 - \\exp \\left(- \\sigma_ {(h, w, d)} \\delta_ {d}\\right)\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 498, + 369, + 890, + 425 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $h, w$ denotes the pixel coordinate of the frustum and $\\delta_{j}$ is the distance between each depth in the frustum. Using the occupancy weights, we put the last $C$ channels of the CNN's output into the frustum $F$ :", + "bbox": [ + 496, + 426, + 893, + 486 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nF (h, w, d) = \\left[ O (h, w, d) \\phi (h, w), \\sigma (h, w, d) \\right] \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 544, + 496, + 890, + 513 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\phi (h,w)$ denotes the $C$ -channeled feature vector at pixel $(h,w)$ which is scaled by $O(h,w,d)$ for $F$ at depth $d$", + "bbox": [ + 496, + 523, + 890, + 553 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "After constructing the frustum for each view, we transform the frustums to world coordinates and fuse them into a shared 3D neural field, represented as density and feature voxel grids. Let $V_{\\text{Density}}$ and $V_{\\text{Feat}}$ denote the density and feature grid, respectively. This formulation of representing a scene with density and feature grids has been explored before [65] for optimization-based scene reconstruction and we utilize it as an intermediate representation for our scene auto-encoder. $V_{\\text{Density}, \\text{Feat}}$ have the same spatial size, and each voxel in $V$ represents a region in the world coordinate system. For each voxel indexed by $(x, y, z)$ , we pool all densities and features of the corresponding frustum entries. In this paper, we simply take the mean of the pooled features. More sophisticated pooling functions (e.g. attention) can be used, which we leave as future work.", + "bbox": [ + 496, + 553, + 890, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Finally, we perform volume rendering using the camera poses $\\kappa$ to project $V$ onto a 2D feature map. We trilinearly interpolate the values on each voxel to get the feature and density for each sampling point along the camera rays. 2D features are then fed into a CNN decoder that produces the output image $\\hat{i}$ . We denote rendering of voxels to output images as $\\hat{i} = r(V,\\kappa)$ . From the volume rendering process, we also get the expected depth $\\hat{\\rho}$ along each ray [50]. The scene", + "bbox": [ + 496, + 780, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "8498", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/01e062238090e1950f0f6a8a49e12e05623432fbde673f153e72ed7cd1dde40f.jpg", + "image_caption": [ + "Figure 3. Scene Auto-Encoder: Each input image is processed with a 2D CNN then lifted up to 3D and merged into the shared voxel grids. Density prediction is not shown here for brevity." + ], + "image_footnote": [], + "bbox": [ + 86, + 87, + 457, + 253 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "auto-encoding pipeline is trained with an image reconstruction loss $||i - \\hat{i}||$ and a depth supervision loss $||\\rho - \\hat{\\rho}||$ . In the case of sparse depth measurements, we only supervise the pixels with recorded depth. We can further improve the quality of the auto-encoder with adversarial loss as in VQGAN [13] or by doing a few optimization steps at inference time, which we discuss in the supplementary.", + "bbox": [ + 75, + 311, + 470, + 417 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Latent Voxel Auto-Encoder", + "text_level": 1, + "bbox": [ + 76, + 426, + 323, + 441 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "It is possible to fit a generative model on voxel grids obtained from Sec. 3.1. However, to capture real-world scenes, the dimensionality of the representation needs to be much larger than what SOTA diffusion models can be trained on. For example, Imagen [56] trains DDMs on $256 \\times 256$ RGB images, and we use voxels of size $128 \\times 128 \\times 32$ with 32 channels. We thus introduce a latent auto-encoder (LAE) that compresses voxels into a 128-dimensional global latent as well as coarse (3D) and fine (2D) quantized latents with channel dimensions of four and spatial dimensions $32 \\times 32 \\times 16$ and $128 \\times 128$ respectively.", + "bbox": [ + 75, + 446, + 468, + 613 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We concatenate $V_{\\text{Density}}$ and $V_{\\text{Feat}}$ along the channel dimension and use separate CNN encoders to encode the voxel grid $V$ into a hierarchy of three latents: 1D global latent $g$ , 3D coarse latent $c$ , and 2D fine latent $f$ , as shown in Fig. 2. The intuition for this design is that $g$ is responsible for representing the global properties of the scene, such as the time of the day, $c$ represents coarse 3D scene structure, and $f$ is a 2D tensor with the same horizontal size $X \\times Y$ as $V$ , which gives further details for each location $(x, y)$ in bird's eye view perspective. We empirically found that 2D CNNs perform similarly to 3D CNNs while being more efficient, thus we use 2D CNNs throughout. To use 2D CNNs for the 3D input $V$ , we concatenate $V$ 's vertical axis along the channel dimension and feed it to the encoders. We also add latent regularizations to avoid high variance latent spaces [53]. For the 1D vector $g$ , we use a small KL-penalty via the reparameterization trick [33], and for $c$ and $f$ , we impose a vector-quantization [13,69] layer to regularize them.", + "bbox": [ + 75, + 613, + 470, + 885 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The CNN decoder is similarly a 2D CNN, and takes $c$", + "bbox": [ + 96, + 885, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ed6be463d777247a4f659c4585dc74e0fbf29f520d27bbf5657b0030a8db5bd5.jpg", + "image_caption": [ + "Figure 4. Hierarchical LDM. Top: LDM $\\psi_{g}$ for KL-regulairzed global latent $g$ . Middle: LDM $\\psi_{c}$ for vector-quantized coarse latent $c$ . Bottom: LDM $\\psi_{f}$ for vector-quantized fine latent $f$ . All LDMs optionally take an additional conditioning variable as input, such as a Bird's Eye View segmentation map as depicted here." + ], + "image_footnote": [], + "bbox": [ + 517, + 87, + 870, + 353 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "concatenated along vertical axis, as the initial input. The decoder uses conditional group normalization layers [71] with $g$ as the conditioning variable. Lastly, we concatenate $f$ to an intermediate tensor in the decoder. The latent decoder outputs $\\hat{V}$ which is the reconstructed voxel. LAE is trained with the voxel reconstruction loss $||V - \\hat{V}||$ along with the image reconstruction loss $||i - \\hat{i}||$ where $\\hat{i} = r(\\hat{V},\\kappa)$ . Note that the image reconstruction loss only helps with the learning of LAE, and the scene auto-encoder is kept fixed.", + "bbox": [ + 496, + 440, + 892, + 577 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Hierarchical Latent Diffusion Models", + "text_level": 1, + "bbox": [ + 500, + 585, + 823, + 602 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given the latent variables $g, c, f$ that represent a voxel-based scene representation $V$ , we define our generative model as $p(V, g, c, f) = p(V | g, c, f) p(f | g, c) p(c | g) p(g)$ with Denoising Diffusion Models (DDMs) [21]. In general, DDMs with discrete time steps have a fixed Markovian forward process $q(x_{t} | x_{t-1})$ where $q(x_{0})$ denotes the data distribution and $q(x_{T})$ is defined to be close to the standard normal distribution, where we use the subscript to denote the time step. DDMs then learn to revert the forward process $p_{\\theta}(x_{t-1} | x_{t})$ with learnable parameters $\\theta$ . It can be shown that learning the reverse process is equivalent to learning to denoise $x_{t}$ to $x_{0}$ for all timesteps $t$ [21, 24] by reducing the following loss:", + "bbox": [ + 496, + 609, + 892, + 805 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\underset {t, \\epsilon , x _ {0}} {\\mathbb {E}} \\left[ w \\left(\\lambda_ {t}\\right) \\| x _ {0} - \\hat {x} _ {\\theta} \\left(x _ {t}, t\\right) \\| _ {2} ^ {2} \\right] \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 805, + 890, + 834 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $t$ is sampled from a uniform distribution for timesteps, $\\epsilon$ is sampled from the standard normal to noise the data $x_0$ , $w(\\lambda_t)$ is a timestep dependent weighting constant, and $\\hat{x}_{\\theta}$ denotes the learned denoising model.", + "bbox": [ + 496, + 839, + 892, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "8499", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We train our hierarchical LDM with the following losses:", + "bbox": [ + 96, + 90, + 467, + 104 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\underset {t, \\epsilon , g _ {0}} {\\mathbb {E}} \\left[ w \\left(\\lambda_ {t}\\right) \\| g _ {0} - \\psi_ {g} \\left(g _ {t}, t\\right) \\| _ {2} ^ {2} \\right] \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 108, + 467, + 137 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\underset {t, \\epsilon , g _ {0}, c _ {0}} {\\mathbb {E}} \\left[ w \\left(\\lambda_ {t}\\right) \\| c _ {0} - \\psi_ {c} \\left(c _ {t}, g _ {0}, t\\right) \\| _ {2} ^ {2} \\right] \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 140, + 467, + 167 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\underset {t, \\epsilon , g _ {0}, c _ {0}, f _ {0}} {\\mathbb {E}} \\left[ w \\left(\\lambda_ {t}\\right) \\| f _ {0} - \\psi_ {f} \\left(f _ {t}, g _ {0}, c _ {0}, t\\right) \\| _ {2} ^ {2} \\right] \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 171, + 467, + 200 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\psi$ denotes the learnable denoising networks for $g, c, f$ . Fig. 4 visualizes the diffusion models. $\\psi_{g}$ is implemented with linear layers with skip connections and $\\psi_{c}$ and $\\psi_{f}$ adopt the U-net architecture [55]. $g$ is fed into $\\psi_{c}$ and $\\psi_{f}$ with conditional group normalization layers. $c$ is interpolated and concatenated to the input to $\\psi_{f}$ . The camera poses contain the trajectory the camera is travelling, and this information can be useful for modelling a 3D scene as it tells the model where to focus on generating. Therefore, we concatenate the camera trajectory information to $g$ and also learn to sample it. For brevity, we still call the concatenated vector $g$ . For conditional generation, each $\\psi$ takes the conditioning variable as input with cross-attention layers [53].", + "bbox": [ + 75, + 205, + 467, + 400 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Each $\\psi$ can be trained in parallel and, once trained, can be sampled one after another following the hierarchy. In practice, we use the v-prediction parameterization [57] that has been shown to have better convergence and training stability [56, 57]. Once $g, c, f$ are sampled, we can use the latent decoder from Sec. 3.2 to construct the voxel $V$ which represents the neural field for the sampled scene.. Following the volume rendering and decoding step in Sec. 3.1, the sampled scene can be visualized from desired viewpoints.", + "bbox": [ + 75, + 402, + 467, + 537 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Post-Optimizing Generated Neural Fields", + "text_level": 1, + "bbox": [ + 76, + 541, + 431, + 556 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Samples generated from our model on real-world data contain reasonable texture and geometry (Fig. 10), but can be further optimized by leveraging recent advances in 2D image diffusion models trained on orders of magnitude more data. Specifically, we iteratively update initially generated voxels, $V$ , by rendering viewpoints from the scene and applying Score Distillation Sampling (SDS) [46] loss on each image independently:", + "bbox": [ + 75, + 564, + 467, + 685 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {V} L _ {S D S} = \\underset {\\epsilon , t, \\kappa} {\\mathbb {E}} \\left[ w \\left(\\lambda_ {t}\\right) \\left(\\epsilon - \\hat {\\epsilon_ {\\theta}} (r (V, \\kappa), t)\\right) \\frac {\\partial r (V , \\kappa)}{\\partial V} \\right] \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 690, + 467, + 720 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\kappa$ is sampled uniformly in a $6m^2$ region around the origin of the scene with random rotation about the vertical axis, $w(\\lambda_t)$ is the weighting schedule used to train $\\hat{\\epsilon_{\\theta}}$ and $t\\sim U[0.02T,0.2T]$ where $T$ is the amount of noise steps used to train $\\hat{\\epsilon_{\\theta}}$ . Note that for latent diffusion models, the noise prediction step is applied after encoding $r(V,\\kappa)$ to the LDM's latent space and the partial gradient term is updated appropriately. For $\\hat{\\epsilon_{\\theta}}$ , we use an off-the-shelf latent diffusion model [53], finetuned to condition on CLIP image embeddings $[47]^1$ . We found that CLIP contains a representativeness matrix", + "bbox": [ + 75, + 728, + 467, + 878 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2ea9770e5dc69ebb6ad3226e32dc2ba9354bf165faa9180aaa077abab306d5ae.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 90, + 694, + 140 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2ae539322fb9ecdeec9f45be497ef0576bd2c9760c6a10da7cdb04fab91cebe8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 696, + 90, + 887, + 140 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d07685b02488898282a4648785c86b7c290ce42a96a0acba94c7bfaeb60c1f2d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 142, + 630, + 193 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/41d03d8227e08ae330ef93128c0d7a9d288883beab02da68729370592d46a572.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 630, + 142, + 887, + 193 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a398490987c93ea73f453a89f7ccccb079ac1c51e836915a2e3ba17d9dce9694.jpg", + "image_caption": [ + "Figure 5. Datasets: Top-left: VizDoom [29]. Top-right: Replica [64]. Middle: Carla [12]. Bottom: AVD. For Carla and AVD, we visualize a subset of available cameras." + ], + "image_footnote": [], + "bbox": [ + 501, + 195, + 630, + 247 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b9446f3503473f385534e348384da5b56e2395971ea38b7be5cb23b178879374.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 630, + 195, + 887, + 247 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "tation of the quality of images that the LDM is able to interpret: denoising an image while conditioning on CLIP image embeddings of our model's samples produced images with similar geometry distortions and texture errors. We leverage this property by optimizing $L_{SDS}$ with negative guidance. Letting $y, y'$ be CLIP embeddings of clean image conditioning (e.g. dataset images) and artifact conditioning (e.g. samples) respectively, we perform classifier-free guidance [23] with conditioning vector $y$ , but replace the unconditional embedding with $y'$ . As shown in the supplementary, this is equivalent (up to scale) to sampling from $\\frac{p(x|y)^{\\alpha}}{p(x|y')}$ at each denoising step where $\\alpha$ controls the trade-off between sampling towards dataset images and away from images with artifacts. We stress that this post-optimization is only successful due to the strong scene prior contained in our voxel samples, as shown by our comparison to running optimization on randomly initialized voxels in the supplementary.", + "bbox": [ + 496, + 314, + 890, + 575 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 588, + 632, + 604 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate NeuralField-LDM on the following four datasets (Fig. 5). Each dataset contains RGB images and a depth measurement with their corresponding camera poses.", + "bbox": [ + 496, + 613, + 890, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "VizDoom [28] consists of front-view sensor observations obtained by navigating inside a synthetic game environment. We use the dataset provided by [9], which contains 34 trajectories with a length of 600.", + "bbox": [ + 496, + 659, + 890, + 718 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Replica [64] is a dataset of high-quality reconstructions of 18 indoor scenes, containing 101 front-view trajectories with a length of 100. We use the dataset provided by [9].", + "bbox": [ + 496, + 719, + 890, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Carla [12] is an open-source simulation platform for self-driving research. We mimic the camera settings for a self-driving car, by placing six cameras (front-left, front, front-right, back-left, back, back-right), covering 360 degrees with some overlaps, and move the car in a randomly sampled direction and distance for 10 timesteps. We sample 43K datapoints, each containing 60 images.", + "bbox": [ + 496, + 765, + 890, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "AVD is an in-house dataset of human driving recordings in roads and parking lots. It has ten cameras with varying", + "bbox": [ + 498, + 869, + 890, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "$^{1}$ https://github.com/justinpinkney/stable-diffusion", + "bbox": [ + 94, + 886, + 359, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "8500", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b1c935c6e1e32d6f692e5f9c1bd56ae3253af7d2de0ffd29b6205c266d678831.jpg", + "image_caption": [ + "Figure 6. Generated Scenes: The top three rows are samples from Carla, and the bottom three rows are samples from AVD." + ], + "image_footnote": [], + "bbox": [ + 80, + 88, + 467, + 445 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/77e6bdd80d513de7c69a7233eab5c29539258efc4d04ccc3faa0ec87d6944e9c.jpg", + "image_caption": [ + "Figure 7. We run marching-cubes [38] on the density voxels to visualize the geometry of the samples generated by NF-LDM. The structure of the scene is reflected well in the mesh." + ], + "image_footnote": [], + "bbox": [ + 114, + 481, + 426, + 743 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "lens types along with Lidar for depth measurement. It has 73K sequences, each with 8 frames extracted at 10 fps.", + "bbox": [ + 76, + 791, + 467, + 821 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Baseline Comparisons", + "text_level": 1, + "bbox": [ + 76, + 832, + 284, + 848 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Unconditional Generation We evaluate the unconditional generation performance of NF-LDM by comparing it with baseline models. All results are without the post", + "bbox": [ + 76, + 854, + 467, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/f19b24e71ef2e1eb056f18cba4e2556f876cb148df02f0e4819dd602a5ef1a4e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
CriterionMethodDepthVizDoomReplica
FID (↓)GRAF [59]X47.5065.37
π-GAN [4]X143.55166.55
GSN [8]✓37.2141.75
GAUDI [2]✓33.7018.75
NF-LDM✓19.54*14.59
", + "bbox": [ + 524, + 89, + 864, + 181 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/dd519d0099e98d4851b80bcae57e754b5e391f5482462a985bf75d9cc1ab13d9.jpg", + "table_caption": [ + "Table 1. FID [19] scores on VizDoom and Replica. NF-LDM outperforms all baseline models. Baseline numbers are from [2]." + ], + "table_footnote": [], + "table_body": "
CriterionMethodDepthCarlaAVD
FID (↓)EG3D [5]×76.89194.34
GSN [8]✓75.45166.07
NF-LDM✓35.6954.26
", + "bbox": [ + 531, + 220, + 857, + 292 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. FID [19] scores on Carla and AVD datasets. Baseline models have trouble learning the distribution of complex outdoor datasets, in particular AVD, while NF-LDM models them well.", + "bbox": [ + 500, + 295, + 890, + 335 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/bc961cb0f79b63395ccc524c7e7dd1fb29c581b3531c929804c3e55a4306d39b.jpg", + "image_caption": [ + "Figure 8. Reconstructing held-out scenes not seen during training." + ], + "image_footnote": [], + "bbox": [ + 517, + 340, + 867, + 656 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "optimization step (Sec. 3.4), unless specified. Tab. 1 shows the results on VizDoom and Replica. GRAF [59] and $\\pi$ -GAN [4], which do not utilize ground truth depth in training, have shown successes in modelling single objects, but they exhibit worse performance than others that leverages depth information for modelling scenes. GAUDI [2] is an auto-decoder-based diffusion model. Their auto-decoder optimizes a small per-scene latent to reconstruct its matching scene. GAUDI comes with the advantage that learning the generative model is simple as it only needs to model the small dimensional latent distribution that acts as the key to their corresponding scenes. On the contrary, NF-LDM is trained on the latents that are a decomposition of the explicit 3D neural field. Therefore, GAUDI puts more modelling", + "bbox": [ + 496, + 688, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "8501", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/1d0a529dc27450600a0621064058a0cc84837ec3f839377a72b62cc6342bd5d4.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
CriterionMethodDepthCarlaAVD
FVD (↓)EG3D [5]X134.941232.38
GSN [8]✓184.301659.81
NF-LDM✓91.80242.50
", + "bbox": [ + 109, + 88, + 434, + 157 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "capacity into the auto-decoder part, and NF-LDM puts it more into the generative model part. We attribute our improvement over GAUDI to our expressive hierarchical LDM that can model the details of the scenes better. In VizDoom, only one scene exists, and each sequence contains several hundred steps covering a large area in the scene, which our voxels were not large enough to encode. Therefore, we chunked each VizDoom trajectory to be 50 steps long.", + "bbox": [ + 75, + 212, + 468, + 333 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tab. 2 shows results on complex outdoor datasets: Carla and AVD. We compare with EG3D [5] and GSN [8]. Both are GAN-based 3D generative models, but GSN utilizes ground truth depth measurements. Note that we did not include GAUDI [2] as the code was not available. NF-LDM achieves the best performance, and both baseline models have difficulty modelling the real outdoor dataset (AVD). Fig. 6 compares the samples from different models.", + "bbox": [ + 75, + 335, + 468, + 455 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Since the datasets are composed of frame sequences, we can treat them as videos and further evaluate with Fréchet Video Distance (FVD) [66] to compare the distributions of the dataset and sampled sequences. This can quantify samples' 3D structure by how natural the rendered sequence from a moving camera is. For EG3D and GSN, we randomly sample a trajectory from the datasets and for NF-LDM, we sample a trajectory from the global latent diffusion model. Tab. 3 shows that NF-LDM achieves the best results. We empirically observed GSN sometimes produced slightly inconsistent rendering, which could attribute to its lower FVD score than EG3D's. We also visualize the geometry of NF-LDM's samples by running marching-cubes [38] on the density voxels. Fig. 7 shows that our samples produce a coarse but realistic geometry.", + "bbox": [ + 75, + 458, + 468, + 685 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablations We evaluate the hierarchical structure of NF-LDM. Tab. 4 shows an ablation study on Carla. The model with the full hierarchy achieves the best performance. The global latent makes it easier for the other LDMs to sample as conditioning on the global properties of the scene (e.g. time of day) narrows down the distribution they need to model. The 2D fine latent helps retain the residual information missing in the 3D coarse latent, thus improving the latent auto-encoder and, consequently, the LDMs.", + "bbox": [ + 75, + 686, + 468, + 823 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Scene Reconstruction Unlike previous approaches, NF-LDM has an explicit scene auto-encoder that can be used for scene reconstruction. GAUDI [2] is auto-decoder based, so it is not trivial to infer a latent for a new scene. GSN [8] can invert a new scene using a GAN inversion method [52,", + "bbox": [ + 75, + 824, + 468, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f3c51d533f81ba0ba02a18e1db53a31f2109d36dce5e90c3ae8dd46b9c47178b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 85, + 890, + 189 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/372c91bb060305a99c8da1b2ebfd42c8d16a6925b4c6cc3fa0e4c07c0768bf4e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 195, + 890, + 277 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e460ab0240c0a8197bc6f8243058061d111636897779ccd35fde311b785be1c6.jpg", + "image_caption": [ + "Figure 9. BEV-Conditioned Synthesis: NF-LDM allows controllable generation by editing the BEV segmentation map. From the initial sample, we add trees (green) and then edit the location of the car (blue). Note the ego car is at the center and thus not rendered." + ], + "image_footnote": [], + "bbox": [ + 500, + 282, + 890, + 363 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5a6e738645d16aefa1b7a951c7fc110ab868a74ee9375cae6052cc7f7c293017.jpg", + "table_caption": [ + "Table 3. FVD [66] scores on Carla and AVD Datasets. As for FID, baseline models have trouble learning to model complex datasets." + ], + "table_footnote": [], + "table_body": "
Coarse lat. c+ Fine lat. f+ Global lat. g
FID (↓)46.4343.5235.69
", + "bbox": [ + 531, + 443, + 857, + 481 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. FID [19] on ablating the choice of hierarchy on the Carla dataset. The first column is for training both LAE and LDM only with the coarse latent. The last column is our full model.", + "bbox": [ + 498, + 482, + 890, + 523 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "79], but as Fig. 8 shows, it fails to get the details of the scene correct. Our scene auto-encoder generalizes well and is scalable as the number of scenes grow.", + "bbox": [ + 498, + 545, + 890, + 592 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Applications and Limitations", + "text_level": 1, + "bbox": [ + 500, + 603, + 759, + 619 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Conditional Synthesis NF-LDM can utilize additional conditioning signals for controllable generation. In this paper, we consider Bird's Eye View (BEV) segmentation maps, but our model can be extended to other conditioning variables. We use cross attention layers [53], which have been shown to be effective for conditional synthesis. Fig. 9 shows that NF-LDM follows the given BEV map faithfully and how the map can be edited for controllable synthesis.", + "bbox": [ + 496, + 627, + 890, + 748 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Scene Editing Image diffusion models can be used for image inpainting without explicit training on the task [53, 63]. We leverage this property to edit scenes in 3D by resampling a region in the 3D coarse latent $c$ . Specifically, at each denoising step, we noise the region to be kept and concatenate with the region being sampled, and pass it through the diffusion model. We use reconstruction guidance [24] to better harmonize the sampled and kept regions. After we get a new $c$ , the fine latent is also re-sampled conditioned on $c$ . Fig. 11 shows results on scene editing with NF-LDM.", + "bbox": [ + 496, + 750, + 892, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "8502", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f6f3e6e358c5a07224b7f0a5b202839d588539fd0048e6c5e6c9587662c45bf8.jpg", + "image_caption": [ + "Figure 10. Panoramas from NF-LDM's samples: From the initial sample at the top, we apply post-optimization with Score Distillation Sampling [46] (Sec. 3.4). (a) demonstrates improved sample quality. (b) showcases style modification by conditioning on evening scenes." + ], + "image_footnote": [], + "bbox": [ + 138, + 88, + 831, + 349 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/97bdf0f0cb0156bf24f2cc3015518c5a49502b30d7eb93a34bd500039ac93a1d.jpg", + "image_caption": [ + "Figure 11. Scene Editing: We use the 3D coarse latent $c$ for scene editing. From the initial sample indicated by light green, we resample a part of the latent, indicated by dark green." + ], + "image_footnote": [], + "bbox": [ + 81, + 388, + 468, + 726 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Post-Optimization Fig. 10 shows how post-optimization (Sec. 3.4) can improve the quality of NF-LDM's initial sample while retaining the 3D structure. In addition to improving quality, we can also modify scene properties, such as time of day and weather, by conditioning the LDM on images with the desired properties. SDS-based style modification is effective for changes where a set of", + "bbox": [ + 75, + 794, + 472, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "clean image data is available with the desired property and is reasonably close to our dataset's domain (e.g. street images for AVD). In the supplementary, we also provide results experimenting with directional CLIP loss [14] to quickly finetune our scene decoder for a given text prompt.", + "bbox": [ + 496, + 393, + 890, + 469 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations NF-LDM's hierarchical structure and three stage pipeline allows us to achieve high-quality generations and reconstructions, but it comes with a degradation in training time and sampling speed. In this work, the neural field representation is based on dense voxel grids, and it becomes expensive to volume render and learn the diffusion models as they get larger. Therefore, exploring alternative sparse representations is a promising future direction. Lastly, our method requires multi-view images which limits data availability and therefore risks universal problems in generative modelling of overfitting. For example, we found that output samples in AVD had limited diversity because the dataset itself was recorded in a limited number of scenes.", + "bbox": [ + 496, + 469, + 892, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 693, + 619, + 709 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduced NeuralField-LDM (NF-LDM), a generative model for complex 3D environments. NF-LDM first constructs an expressive latent distribution by encoding input images into a 3D neural field representation which is further compressed into more abstract latent spaces. Then, our proposed hierarchical LDM is fit onto the latent spaces, achieving state-of-the-art performance on 3D scene generation. NF-LDM enables a diverse set of applications, including controllable scene generation and scene editing. Future directions include exploring more efficient sparse voxel representations, training on larger-scale real-world data and learning to continuously expand generated scenes.", + "bbox": [ + 496, + 719, + 893, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8503", + "bbox": [ + 482, + 944, + 514, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Martin Arjovsky and Léon Bottou. Towards principled methods for training generative adversarial networks. In ICLR, 2017. 3", + "[2] Miguel Ángel Bautista, Pengsheng Guo, Samira Abnar, Walter Talbott, Alexander Toshev, Zhuoyuan Chen, Laurent Dinh, Shuangfei Zhai, Hanlin Goh, Daniel Ulbricht, Afshin Dehghan, and Josh M. Susskind. GAUDI: A neural architect for immersive 3d scene generation. 2022. 1, 2, 6, 7", + "[3] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale GAN training for high fidelity natural image synthesis. In ICLR, 2019. 2", + "[4] Eric Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. arXiv, 2012.00926, 2020. 6", + "[5] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In arXiv, 2021. 6, 7", + "[6] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In CVPR, 2022. 1, 2", + "[7] Eric R. Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. Pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In CVPR, 2021. 2", + "[8] Terrance DeVries, Miguel Angel Bautista, Nitish Srivastava, Graham W. Taylor, and Joshua M. Susskind. Unconstrained scene generation with locally conditioned radiance fields. In ICCV, 2021. 1, 2, 6, 7", + "[9] Terrance DeVries, Miguel Angel Bautista, Nitish Srivastava, Graham W. Taylor, and Joshua M. Susskind. Unconstrained scene generation with locally conditioned radiance fields. arXiv, 2021. 5", + "[10] Prafulla Dhariwal and Alexander Quinn Nichol. Diffusion models beat gans on image synthesis. In Marc'Aurelio Ranzato, Alina Beygelzimer, Yann N. Dauphin, Percy Liang, and Jennifer Wortman Vaughan, editors, NeurIPS, 2021. 2", + "[11] Tim Dockhorn, Arash Vahdat, and Karsten Kreis. Score-based generative modeling with critically-damped Langevin diffusion. In ICLR, 2022. 2", + "[12] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. CARLA: An open urban driving simulator. 2017. 5", + "[13] Patrick Esser, Robin Rombach, and Björn Ommer. Taming transformers for high-resolution image synthesis. In CVPR, 2021. 2, 4", + "[14] Rinon Gal, Or Patashnik, Haggai Maron, Gal Chechik, and Daniel Cohen-Or. Stylegan-nada: Clip-guided domain adaptation of image generators, 2021. 8", + "[15] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja" + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. In Advances In Neural Information Processing Systems, 2022. 1", + "[16] Ian Goodfellow. Nips 2016 tutorial: Generative adversarial networks. arXiv preprint arXiv:1701.00160, 2016. 3", + "[17] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron C. Courville, and Yoshua Bengio. Generative adversarial nets. In NeurIPS, 2014. 2", + "[18] William Harvey, Saeid Naderiparizi, Vaden Masrani, Christian Weilbach, and Frank Wood. Flexible diffusion modeling of long videos, 2022. 2", + "[19] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. 6, 7", + "[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Hugo Larochelle, Marc' Aurelio Ranzato, Raia Hadsell, Maria-Florina Balcan, and Hsuan-Tien Lin, editors, NeurIPS, 2020. 2, 3", + "[21] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 6840-6851. Curran Associates, Inc., 2020. 4", + "[22] Jonathan Ho, Chitwan Sahara, William Chan, David J. Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. J. Mach. Learn. Res., 2022. 2", + "[23] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 5", + "[24] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. arXiv:2204.03458, 2022. 2, 4, 7", + "[25] James T Kajiya and Brian P Von Herzen. Ray tracing volume densities. ACM SIGGRAPH computer graphics, 18(3):165-174, 1984. 2", + "[26] Amlan Kar, Aayush Prakash, Ming-Yu Liu, Eric Cameracci, Justin Yuan, Matt Rusiniak, David Acuna, Antonio Torralba, and Sanja Fidler. Meta-sim: Learning to generate synthetic datasets. In ICCV, 2019. 2", + "[27] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. In CVPR, 2020. 2", + "[28] Michal Kempka, Marek Wydmuch, Grzegorz Runc, Jakub Toczek, and Wojciech Jaskowski. Vizdoom: A doom-based ai research platform for visual reinforcement learning. In 2016 IEEE conference on computational intelligence and games (CIG), pages 1-8. IEEE, 2016. 5", + "[29] Margret Keuper, Siyu Tang, Bjoern Andres, Thomas Brox, and Bernt Schiele. Motion segmentation & multiple object tracking by correlation co-clustering. IEEE TPAMI, 2018. 5", + "[30] Seung Wook Kim, Karsten Kreis, Daiqing Li, Antonio Torralba, and Sanja Fidler. Polymorphic-gan: Generating aligned samples across multiple domains with learned morph maps. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Jun. 2022. 2" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "8504", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[31] Seung Wook Kim, Jonah Philion, Antonio Torralba, and Sanja Fidler. Drivegan: Towards a controllable high-quality neural simulation. In CVPR, 2021. 2", + "[32] Seung Wook Kim, Yuhao Zhou, Jonah Philion, Antonio Torralba, and Sanja Fidler. Learning to simulate dynamic environments with gamean. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1231-1240, 2020. 2", + "[33] Diederik P. Kingma and Max Welling. Auto-encoding variational bayes. *ICLR*, 2014. 2, 4", + "[34] Daiqing Li, Huan Ling, Seung Wook Kim, Karsten Kreis, Adela Barriuso, Sanja Fidler, and Antonio Torralba. Big-datasetgan: Synthesizing imagenet with pixel-wise annotations. arXiv, 2201.04684, 2022. 2", + "[35] Ke Li and Jitendra Malik. On the implicit assumptions of gans. arXiv preprint arXiv:1811.12402, 2018. 3", + "[36] Yiyi Liao, Katja Schwarz, Lars Mescheder, and Andreas Geiger. Towards unsupervised learning of generative models for 3d controllable image synthesis. In CVPR, 2020. 2", + "[37] Huan Ling, Karsten Kreis, Daiqing Li, Seung Wook Kim, Antonio Torralba, and Sanja Fidler. Editgan: High-precision semantic image editing. In NeurIPS, 2021. 2", + "[38] William E. Lorensen and Harvey E. Cline. Marching cubes: A high resolution 3d surface construction algorithm. In ACM Trans. on Graphics, 1987. 6, 7", + "[39] Shitong Luo and Wei Hu. Diffusion probabilistic models for 3d point cloud generation. In CVPR, 2021. 2", + "[40] Ricardo Martin-Brualla, Noha Radwan, Mehdi S. M. Sajjadi, Jonathan T. Barron, Alexey Dosovitskiy, and Daniel Duckworth. NeRF in the Wild: Neural Radiance Fields for Unconstrained Photo Collections. In CVPR, 2021. 2", + "[41] Lars Mescheder, Andreas Geiger, and Sebastian Nowozin. Which training methods for gans do actually converge? In Proc. of the International Conf. on Machine learning (ICML), 2018. 2", + "[42] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 2, 3", + "[43] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. GLIDE: towards photorealistic image generation and editing with text-guided diffusion models. In Kamalika Chaudhuri, Stefanie Jegelka, Le Song, Csaba Szepesvári, Gang Niu, and Sivan Sabato, editors, Proc. of the International Conf. on Machine learning (ICML), 2022. 2", + "[44] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In CVPR, 2021. 2", + "[45] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In ECCV, 2020. 3", + "[46] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv, 2022. 1, 2, 5, 8" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[47] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. arXiv, 2103.00020, 2021. 5", + "[48] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with CLIP latents. arXiv, abs/2204.06125, 2022. 2", + "[49] Ali Razavi, Aaron van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with VQ-VAE-2. In NeurIPS, 2019. 2", + "[50] Konstantinos Rematas, Andrew Liu, Pratul P. Srinivasan, Jonathan T. Barron, Andrea Tagliasacchi, Thomas A. Funkhouser, and Vittorio Ferrari. Urban radiance fields. In CVPR, 2022. 2, 3", + "[51] Danilo Jimenez Rezende, Shakir Mohamed, and Daan Wierstra. Stochastic backpropagation and approximate inference in deep generative models. In Proc. of the International Conf. on Machine learning (ICML), 2014. 2", + "[52] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a stylegan encoder for image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2287–2296, 2021. 7", + "[53] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 3, 4, 5, 7", + "[54] Robin Rombach, Patrick Esser, and Björn Ommer. Geometry-free view synthesis: Transformers and no 3d priors. In ICCV, 2021. 2", + "[55] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234-241. Springer, 2015. 5", + "[56] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 3, 4, 5", + "[57] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. arXiv preprint arXiv:2202.00512, 2022. 3, 5", + "[58] Axel Sauer, Katja Schwarz, and Andreas Geiger. StyleganxI: Scaling stylegan to large diverse datasets. ACM Trans. on Graphics, 2022. 2", + "[59] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. Graf: Generative radiance fields for 3d-aware image synthesis. 2020. 2, 6", + "[60] Katja Schwarz, Axel Sauer, Michael Niemeyer, Yiyi Liao, and Andreas Geiger. Voxgraf: Fast 3d-aware image synthesis with sparse voxel grids. In NeurIPS, 2022. 2", + "[61] Yujun Shen, Jinjin Gu, Xiaou Tang, and Bolei Zhou. Interpreting the latent space of gans for semantic face editing. In CVPR, 2020. 2" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "8505", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[62] Yuan Shen, Wei-Chiu Ma, and Shenlong Wang. SGAM: Building a virtual 3d world through simultaneous generation and mapping. In Thirty-Sixth Conference on Neural Information Processing Systems, 2022. 2", + "[63] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 2, 7", + "[64] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J. Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, Anton Clarkson, Mingfei Yan, Brian Budge, Yajie Yan, Xiaqing Pan, June Yon, Yuyang Zou, Kimberly Leon, Nigel Carter, Jesus Briales, Tyler Gillingham, Elias Mueggler, Luis Pesqueira, Manolis Savva, Dhruv Batra, Hauke M. Strasdat, Renzo De Nardi, Michael Goesele, Steven Lovegrove, and Richard Newcombe. The Replica dataset: A digital replica of indoor spaces. arXiv, 1906.05797, 2019. 5", + "[65] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. CVPR, 2022. 3", + "[66] Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 7", + "[67] Arash Vahdat and Jan Kautz. NVAE: A deep hierarchical variational autoencoder. In Hugo Larochelle, Marc'Aurelio Ranzato, Raia Hadsell, Maria-Florina Balcan, and Hsuan-Tien Lin, editors, NeurIPS, 2020. 2", + "[68] Arash Vahdat, Karsten Kreis, and Jan Kautz. Score-based generative modeling in latent space. In NeurIPS, 2021. 2", + "[69] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 4", + "[70] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P. Srinivasan, Howard Zhou, Jonathan T. Barron, Ricardo Martin-Brualla, Noah Snavely, and Thomas A. Funkhouser. Ibrnet: Learning multi-view image-based rendering. In CVPR, 2021. 2", + "[71] Yuxin Wu and Kaiming He. Group normalization. In Proceedings of the European conference on computer vision (ECCV), pages 3-19, 2018. 4", + "[72] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelNeRF: Neural radiance fields from one or few images. In CVPR, 2021. 2", + "[73] Xiaohui Zeng, Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, and Karsten Kreis. LION: latent point diffusion models for 3d shape generation. In NeurIPS, 2022. 1, 2, 3", + "[74] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv, 2010.07492, 2020. 2", + "[75] Xiaoshuai Zhang, Sai Bi, Kalyan Sunkavalli, Hao Su, and Zexiang Xu. Nerfusion: Fusing radiance fields for large-scale scene reconstruction. In CVPR, 2022. 2", + "[76] Yuxuan Zhang, Wenzheng Chen, Huan Ling, Jun Gao, Yi-nan Zhang, Antonio Torralba, and Sanja Fidler. Image gans" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "meet differentiable rendering for inverse graphics and interpretable 3d neural rendering. In ICLR, 2021. 2", + "[77] Yan Zheng, Lemeng Wu, Xingchao Liu, Zhen Chen, Qiang Liu, and Qixing Huang. Neural volumetric mesh generator. arXiv preprint arXiv:2210.03158, 2022. 3", + "[78] Linqi Zhou, Yilun Du, and Jiajun Wu. 3d shape generation and completion through point-voxel diffusion. In ICCV, 2021. 2", + "[79] Jiapeng Zhu, Yujun Shen, Deli Zhao, and Bolei Zhou. Indomain gan inversion for real image editing. In European conference on computer vision, pages 592-608. Springer, 2020. 7" + ], + "bbox": [ + 503, + 92, + 890, + 260 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "8506", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/NeuralField-LDM_ Scene Generation With Hierarchical Latent Diffusion Models/81bfb70e-1652-4684-aa6c-4729296b6ed4_model.json b/2023/NeuralField-LDM_ Scene Generation With Hierarchical Latent Diffusion Models/81bfb70e-1652-4684-aa6c-4729296b6ed4_model.json new file mode 100644 index 0000000000000000000000000000000000000000..98800bf3e73be9d9231dc21f12099dcb30c80d95 --- /dev/null +++ b/2023/NeuralField-LDM_ Scene Generation With Hierarchical Latent Diffusion Models/81bfb70e-1652-4684-aa6c-4729296b6ed4_model.json @@ -0,0 +1,2554 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.081, + 0.109, + 0.89, + 0.13 + ], + "angle": 0, + "content": "NeuralField-LDM: Scene Generation with Hierarchical Latent Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.159, + 0.885, + 0.178 + ], + "angle": 0, + "content": "Seung Wook Kim\\(^{1,2,3*}\\) Bradley Brown\\(^{1,5*†}\\) Kangxue Yin\\(^{1}\\) Karsten Kreis\\(^{1}\\) Katja Schwarz\\(^{6†}\\)" + }, + { + "type": "text", + "bbox": [ + 0.19, + 0.178, + 0.286, + 0.196 + ], + "angle": 0, + "content": "Daiqing Li" + }, + { + "type": "text", + "bbox": [ + 0.327, + 0.178, + 0.47, + 0.196 + ], + "angle": 0, + "content": "Robin Rombach\\(^{7\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.178, + 0.657, + 0.196 + ], + "angle": 0, + "content": "Antonio Torralba" + }, + { + "type": "text", + "bbox": [ + 0.698, + 0.178, + 0.823, + 0.196 + ], + "angle": 0, + "content": "Sanja Fidler1,2,3" + }, + { + "type": "text", + "bbox": [ + 0.199, + 0.197, + 0.78, + 0.212 + ], + "angle": 0, + "content": "\\(^{1}\\)NVIDIA \\(^{2}\\)University of Toronto \\(^{3}\\)Vector Institute \\(^{4}\\)CSAIL, MIT \\(^{5}\\)University of Waterloo" + }, + { + "type": "text", + "bbox": [ + 0.301, + 0.214, + 0.678, + 0.23 + ], + "angle": 0, + "content": "6University of Tübingen, Tübingen AI Center 7LMU Munich" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.236, + 0.891, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.346, + 0.23, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.346, + 0.396, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.411, + 0.346, + 0.561, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.576, + 0.346, + 0.726, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.741, + 0.345, + 0.891, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.423, + 0.893, + 0.451 + ], + "angle": 0, + "content": "Figure 1. We introduce NeuralField-LDM, a generative model for complex open-world 3D scenes. This figure contains a panorama constructed from NeuralField-LDM's generated scene. We visualize different parts of the scene by placing cameras on them." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.464, + 0.314, + 0.479 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.495, + 0.474, + 0.783 + ], + "angle": 0, + "content": "Automatically generating high-quality real world 3D scenes is of enormous interest for applications such as virtual reality and robotics simulation. Towards this goal, we introduce NeuralField-LDM, a generative model capable of synthesizing complex 3D environments. We leverage Latent Diffusion Models that have been successfully utilized for efficient high-quality 2D content creation. We first train a scene auto-encoder to express a set of image and pose pairs as a neural field, represented as density and feature voxel grids that can be projected to produce novel views of the scene. To further compress this representation, we train a latent-autoencoder that maps the voxel grids to a set of latent representations. A hierarchical diffusion model is then fit to the latents to complete the scene generation pipeline. We achieve a substantial improvement over existing state-of-the-art scene generation models. Additionally, we show how NeuralField-LDM can be used for a variety of 3D content creation applications, including conditional scene generation, scene inpainting and scene style manipulation." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.811, + 0.21, + 0.827 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.836, + 0.47, + 0.867 + ], + "angle": 0, + "content": "There has been increasing interest in modelling 3D real-world scenes for use in virtual reality, game design, digi-" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.465, + 0.893, + 0.722 + ], + "angle": 0, + "content": "tal twin creation and more. However, designing 3D worlds by hand is a challenging and time-consuming process, requiring 3D modeling expertise and artistic talent. Recently, we have seen success in automating 3D content creation via 3D generative models that output individual object assets [15,46,73]. Although a great step forward, automating the generation of real-world scenes remains an important open problem and would unlock many applications ranging from scalably generating a diverse array of environments for training AI agents (e.g. autonomous vehicles) to the design of realistic open-world video games. In this work, we take a step towards this goal with NeuralField-LDM (NF-LDM), a generative model capable of synthesizing complex real-world 3D scenes. NF-LDM is trained on a collection of posed camera images and depth measurements which are easier to obtain than explicit ground-truth 3D data, offering a scalable way to synthesize 3D scenes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.728, + 0.894, + 0.865 + ], + "angle": 0, + "content": "Recent approaches [2, 6, 8] tackle the same problem of generating 3D scenes, albeit on less complex data. In [6, 8], a latent distribution is mapped to a set of scenes using adversarial training, and in GAUDI [2], a denoising diffusion model is fit to a set of scene latents learned using an auto-decoder. These models all have an inherent weakness of attempting to capture the entire scene into a single vector that conditions a neural radiance field. In practice, we find that this limits the ability to fit complex scene distributions." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Recently, diffusion models have emerged as a very powerful class of generative models, capable of generating high-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.1, + 0.875, + 0.213, + 0.888 + ], + "angle": 0, + "content": "* Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.103, + 0.888, + 0.342, + 0.901 + ], + "angle": 0, + "content": "†Work done during an internship at NVIDIA." + }, + { + "type": "list", + "bbox": [ + 0.1, + 0.871, + 0.342, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8496" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.166 + ], + "angle": 0, + "content": "quality images, point clouds and videos [18, 24, 39, 48, 53, 73, 78]. Yet, due to the nature of our task, where image data must be mapped to a shared 3D scene without an explicit ground truth 3D representation, straightforward approaches fitting a diffusion model directly to data are infeasible." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.168, + 0.468, + 0.452 + ], + "angle": 0, + "content": "In NeuralField-LDM, we learn to model scenes using a three-stage pipeline. First, we learn an auto-encoder that encodes scenes into a neural field, represented as density and feature voxel grids. Inspired by the success of latent diffusion models for images [53], we learn to model the distribution of our scene voxels in latent space to focus the generative capacity on core parts of the scene and not the extraneous details captured by our voxel auto-encoders. Specifically, a latent-autoencoder decomposes the scene voxels into a 3D coarse, 2D fine and 1D global latent. Hierarchical diffusion models are then trained on the tri-latent representation to generate novel 3D scenes. We show how NF-LDM enables applications such as scene editing, birdeye view conditional generation and style adaptation. Finally, we demonstrate how score distillation [46] can be used to optimize the quality of generated neural fields, allowing us to leverage the representations learned from state-of-the-art image diffusion models that have been exposed to orders of magnitude more data." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.455, + 0.468, + 0.545 + ], + "angle": 0, + "content": "Our contributions are: 1) We introduce NF-LDM, a hierarchical diffusion model capable of generating complex open-world 3D scenes and achieving state of the art scene generation results on four challenging datasets. 2) We extend NF-LDM to semantic birds-eye view conditional scene generation, style modification and 3D scene editing." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.553, + 0.218, + 0.568 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.578, + 0.468, + 0.818 + ], + "angle": 0, + "content": "2D Generative Models In past years, generative adversarial networks (GANs) [3, 17, 27, 41, 58] and likelihood-based approaches [33, 49, 51, 67] enabled high-resolution photorealistic image synthesis. Due to their quality, GANs are used in a multitude of downstream applications ranging from steerable content creation [30, 34, 36, 37, 61, 76] to data driven simulation [26, 31, 32, 34]. Recently, autoregressive models and score-based models, e.g. diffusion models, demonstrate better distribution coverage while preserving high sample quality [10, 11, 13, 20, 22, 43, 48, 53, 54, 68]. Since evaluation and optimization of these approaches in pixel space is computationally expensive, [53, 68] apply them to latent space, achieving state-of-the-art image synthesis at megapixel resolution. As our approach operates on 3D scenes, computational efficiency is crucial. Hence, we build upon [53] and train our model in latent space." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.468, + 0.9 + ], + "angle": 0, + "content": "Novel View Synthesis In their seminal work [42], Mildenhall et al. introduce Neural Radiance Fields (NeRF) as a powerful 3D representation. PixelNeRF [72] and IBR-Net [70] propose to condition NeRF on aggregated features from multiple views to enable novel view synthesis from" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.213 + ], + "angle": 0, + "content": "a sparse set of views. Another line of works scale NeRF to large-scale indoor and outdoor scenes [40, 50, 74, 75]. Recently, Nerfusion [75] predicts local radiance fields and fuses them into a scene representation using a recurrent neural network. Similarly, we construct a latent scene representation by aggregating features across multiple views. Different from the aforementioned methods, our approach is a generative model capable of synthesizing novel scenes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.219, + 0.892, + 0.34 + ], + "angle": 0, + "content": "3D Diffusion Models A few recent works propose to apply denoising diffusion models (DDM) [20, 22, 63] on point clouds for 3D shape generation [39, 73, 78]. While PVD [78] trains on point clouds directly, DPM [39] and LION [73] use a shape latent variable. Similar to LION, we design a hierarchical model by training separate conditional DDMs. However, our approach generates both texture and geometry of a scene without needing 3D ground truth as supervision." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.346, + 0.892, + 0.618 + ], + "angle": 0, + "content": "3D-Aware Generative Models 3D-aware generative models synthesize images while providing explicit control over the camera pose and potentially other scene properties, like object shape and appearance. SGAM [62] generates a 3D scene by autoregressively generating sensor data and building a 3D map. Several previous approaches generate NeRFs of single objects with conditional coordinate-based MLPs [7, 44, 59]. GSN [8] conditions a coordinate-based MLP on a \"floor plan\", i.e., a 2D feature map, to model more complex indoor scenes. EG3D [6] and VoxGRAF [60] use convolutional backbones to generate 3D representations. All of these approaches rely on adversarial training. Instead, we train a DDM on voxels in latent space. The work closest to ours is GAUDI [2], which first trains an auto-decoder and subsequently trains a DDM on the learned latent codes. Instead of using a global latent code, we encode scenes onto voxel grids and train a hierarchical DDM to optimally combine global and local features." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.633, + 0.677, + 0.648 + ], + "angle": 0, + "content": "3. NeuralField-LDM" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.892, + 0.795 + ], + "angle": 0, + "content": "Our objective is to train a generative model to synthesize 3D scenes that can be rendered to any viewpoint. We assume access to a dataset \\(\\{(i,\\kappa ,\\rho)\\}_{1..N}\\) which consists of \\(N\\) RGB images \\(i\\) and their camera poses \\(\\kappa\\), along with a depth measurement \\(\\rho\\) that can be either sparse (e.g. Lidar points) or dense. The generative model must learn to model both the texture and geometry distributions of the dataset in 3D by learning solely from the sensor observations, which is a highly non-trivial problem." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Past work typically tackles this problem with a generative adversarial network (GAN) framework [6, 8, 59, 60]. They produce an intermediate 3D representation and render images for a given viewpoint with volume rendering [25, 42]. Discriminator losses then ensure that the 3D representation produces a valid image from any viewpoint. However, GANs come with notorious training instability" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8497" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.124, + 0.085, + 0.844, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.225, + 0.895, + 0.268 + ], + "angle": 0, + "content": "Figure 2. Overview of NeuralField-LDM. We first encode RGB images with camera poses into a neural field represented by density and feature voxel grids. We compress the neural field into smaller latent spaces and fit a hierarchical latent diffusion model on the latent space. Sampled latents can then be decoded into a neural field that can be rendered into a given viewpoint." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.284, + 0.47, + 0.39 + ], + "angle": 0, + "content": "and mode dropping behaviors [1, 16, 35]. Denoising Diffusion models [20] (DDMs) have recently emerged as an alternative to GANs that avoid the aforementioned disadvantages [53, 56, 57]. However, DDMs model the data likelihood explicitly and are trained to reconstruct the training data. Thus, they have been used in limited scenarios [73, 77] since ground-truth 3D data is not readily available at scale." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.39, + 0.471, + 0.662 + ], + "angle": 0, + "content": "To tackle the challenging problem of generating an entire scene with texture and geometry, we take inspiration from latent diffusion models (LDM) [53], which first construct an intermediate latent distribution of the training data then fit a diffusion model on the latent distribution. In Sec. 3.1, we introduce a scene auto-encoder that encodes the set of RGB images into a neural field representation consisting of density and feature voxel grids. To accurately capture a scene, the voxel grids' spatial dimension needs to be much larger than what current state-of-the-art LDMs can model. In Sec. 3.2, we show how we can further compress and decompose the explicit voxel grids into compressed latent representations to facilitate learning the data distribution. Finally, Sec. 3.3 introduces a latent diffusion model that models the latent distributions in a hierarchical manner. Fig. 2 shows an overview of our method, which we name NeuralField-LDM (NF-LDM). We provide training and additional architecture details in the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.667, + 0.271, + 0.681 + ], + "angle": 0, + "content": "3.1. Scene Auto-Encoder" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.471, + 0.901 + ], + "angle": 0, + "content": "The goal of the scene auto-encoder is to obtain a 3D representation of the scene from input images by learning to reconstruct them. Fig. 3 depicts the auto-encoding process. The scene encoder is a 2D CNN and processes each RGB image \\( i_{1..N} \\) separately, producing a \\( \\mathbb{R}^{H\\times W\\times (D + C)} \\) dimensional 2D tensor for each image, where \\( H \\) and \\( W \\) are smaller than \\( i \\)'s size. We follow a similar procedure to Lift-Splat-Shoot (LSS) [45] to lift each 2D image feature map and combine them in the common voxel-based 3D neural field. We build a discrete frustum of size \\( H\\times W\\times D \\) with the camera poses \\( \\kappa \\) for each image. This frustum contains image features and density values for each pixel, along a pre-defined discrete set of \\( D \\) depths. Unlike LSS, we take the first \\( D \\) channels of the 2D CNN's output and use them as" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.284, + 0.892, + 0.362 + ], + "angle": 0, + "content": "density values. That is, the \\(d\\)th channel of the CNN's output at pixel \\((h,w)\\) becomes the density value of the frustum entry at \\((h,w,d)\\). Motivated by the volume rendering equation [42], we get the occupancy weight \\(O\\) of each element \\((h,w,d)\\) in the frustum using the density values \\(\\sigma \\geq 0\\):" + }, + { + "type": "equation", + "bbox": [ + 0.499, + 0.37, + 0.892, + 0.426 + ], + "angle": 0, + "content": "\\[\nO (h, w, d) = \\exp \\left(- \\sum_ {j = 0} ^ {d - 1} \\sigma_ {(h, w, j)} \\delta_ {j}\\right) \\left(1 - \\exp \\left(- \\sigma_ {(h, w, d)} \\delta_ {d}\\right)\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.427, + 0.894, + 0.487 + ], + "angle": 0, + "content": "where \\( h, w \\) denotes the pixel coordinate of the frustum and \\( \\delta_{j} \\) is the distance between each depth in the frustum. Using the occupancy weights, we put the last \\( C \\) channels of the CNN's output into the frustum \\( F \\):" + }, + { + "type": "equation", + "bbox": [ + 0.545, + 0.497, + 0.892, + 0.515 + ], + "angle": 0, + "content": "\\[\nF (h, w, d) = \\left[ O (h, w, d) \\phi (h, w), \\sigma (h, w, d) \\right] \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.524, + 0.892, + 0.554 + ], + "angle": 0, + "content": "where \\(\\phi (h,w)\\) denotes the \\(C\\) -channeled feature vector at pixel \\((h,w)\\) which is scaled by \\(O(h,w,d)\\) for \\(F\\) at depth \\(d\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.554, + 0.892, + 0.78 + ], + "angle": 0, + "content": "After constructing the frustum for each view, we transform the frustums to world coordinates and fuse them into a shared 3D neural field, represented as density and feature voxel grids. Let \\( V_{\\text{Density}} \\) and \\( V_{\\text{Feat}} \\) denote the density and feature grid, respectively. This formulation of representing a scene with density and feature grids has been explored before [65] for optimization-based scene reconstruction and we utilize it as an intermediate representation for our scene auto-encoder. \\( V_{\\text{Density}, \\text{Feat}} \\) have the same spatial size, and each voxel in \\( V \\) represents a region in the world coordinate system. For each voxel indexed by \\( (x, y, z) \\), we pool all densities and features of the corresponding frustum entries. In this paper, we simply take the mean of the pooled features. More sophisticated pooling functions (e.g. attention) can be used, which we leave as future work." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Finally, we perform volume rendering using the camera poses \\(\\kappa\\) to project \\(V\\) onto a 2D feature map. We trilinearly interpolate the values on each voxel to get the feature and density for each sampling point along the camera rays. 2D features are then fed into a CNN decoder that produces the output image \\(\\hat{i}\\). We denote rendering of voxels to output images as \\(\\hat{i} = r(V,\\kappa)\\). From the volume rendering process, we also get the expected depth \\(\\hat{\\rho}\\) along each ray [50]. The scene" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8498" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.087, + 0.088, + 0.458, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.255, + 0.47, + 0.297 + ], + "angle": 0, + "content": "Figure 3. Scene Auto-Encoder: Each input image is processed with a 2D CNN then lifted up to 3D and merged into the shared voxel grids. Density prediction is not shown here for brevity." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.313, + 0.471, + 0.419 + ], + "angle": 0, + "content": "auto-encoding pipeline is trained with an image reconstruction loss \\( ||i - \\hat{i}|| \\) and a depth supervision loss \\( ||\\rho - \\hat{\\rho}|| \\). In the case of sparse depth measurements, we only supervise the pixels with recorded depth. We can further improve the quality of the auto-encoder with adversarial loss as in VQGAN [13] or by doing a few optimization steps at inference time, which we discuss in the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.428, + 0.325, + 0.443 + ], + "angle": 0, + "content": "3.2. Latent Voxel Auto-Encoder" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.448, + 0.47, + 0.614 + ], + "angle": 0, + "content": "It is possible to fit a generative model on voxel grids obtained from Sec. 3.1. However, to capture real-world scenes, the dimensionality of the representation needs to be much larger than what SOTA diffusion models can be trained on. For example, Imagen [56] trains DDMs on \\(256 \\times 256\\) RGB images, and we use voxels of size \\(128 \\times 128 \\times 32\\) with 32 channels. We thus introduce a latent auto-encoder (LAE) that compresses voxels into a 128-dimensional global latent as well as coarse (3D) and fine (2D) quantized latents with channel dimensions of four and spatial dimensions \\(32 \\times 32 \\times 16\\) and \\(128 \\times 128\\) respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.471, + 0.886 + ], + "angle": 0, + "content": "We concatenate \\( V_{\\text{Density}} \\) and \\( V_{\\text{Feat}} \\) along the channel dimension and use separate CNN encoders to encode the voxel grid \\( V \\) into a hierarchy of three latents: 1D global latent \\( g \\), 3D coarse latent \\( c \\), and 2D fine latent \\( f \\), as shown in Fig. 2. The intuition for this design is that \\( g \\) is responsible for representing the global properties of the scene, such as the time of the day, \\( c \\) represents coarse 3D scene structure, and \\( f \\) is a 2D tensor with the same horizontal size \\( X \\times Y \\) as \\( V \\), which gives further details for each location \\( (x, y) \\) in bird's eye view perspective. We empirically found that 2D CNNs perform similarly to 3D CNNs while being more efficient, thus we use 2D CNNs throughout. To use 2D CNNs for the 3D input \\( V \\), we concatenate \\( V \\)'s vertical axis along the channel dimension and feed it to the encoders. We also add latent regularizations to avoid high variance latent spaces [53]. For the 1D vector \\( g \\), we use a small KL-penalty via the reparameterization trick [33], and for \\( c \\) and \\( f \\), we impose a vector-quantization [13,69] layer to regularize them." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.469, + 0.901 + ], + "angle": 0, + "content": "The CNN decoder is similarly a 2D CNN, and takes \\( c \\)" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.088, + 0.872, + 0.354 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.357, + 0.893, + 0.428 + ], + "angle": 0, + "content": "Figure 4. Hierarchical LDM. Top: LDM \\(\\psi_{g}\\) for KL-regulairzed global latent \\(g\\). Middle: LDM \\(\\psi_{c}\\) for vector-quantized coarse latent \\(c\\). Bottom: LDM \\(\\psi_{f}\\) for vector-quantized fine latent \\(f\\). All LDMs optionally take an additional conditioning variable as input, such as a Bird's Eye View segmentation map as depicted here." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.441, + 0.893, + 0.578 + ], + "angle": 0, + "content": "concatenated along vertical axis, as the initial input. The decoder uses conditional group normalization layers [71] with \\( g \\) as the conditioning variable. Lastly, we concatenate \\( f \\) to an intermediate tensor in the decoder. The latent decoder outputs \\( \\hat{V} \\) which is the reconstructed voxel. LAE is trained with the voxel reconstruction loss \\( ||V - \\hat{V}|| \\) along with the image reconstruction loss \\( ||i - \\hat{i}|| \\) where \\( \\hat{i} = r(\\hat{V},\\kappa) \\). Note that the image reconstruction loss only helps with the learning of LAE, and the scene auto-encoder is kept fixed." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.587, + 0.825, + 0.603 + ], + "angle": 0, + "content": "3.3. Hierarchical Latent Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.61, + 0.893, + 0.806 + ], + "angle": 0, + "content": "Given the latent variables \\( g, c, f \\) that represent a voxel-based scene representation \\( V \\), we define our generative model as \\( p(V, g, c, f) = p(V | g, c, f) p(f | g, c) p(c | g) p(g) \\) with Denoising Diffusion Models (DDMs) [21]. In general, DDMs with discrete time steps have a fixed Markovian forward process \\( q(x_{t} | x_{t-1}) \\) where \\( q(x_{0}) \\) denotes the data distribution and \\( q(x_{T}) \\) is defined to be close to the standard normal distribution, where we use the subscript to denote the time step. DDMs then learn to revert the forward process \\( p_{\\theta}(x_{t-1} | x_{t}) \\) with learnable parameters \\( \\theta \\). It can be shown that learning the reverse process is equivalent to learning to denoise \\( x_{t} \\) to \\( x_{0} \\) for all timesteps \\( t \\) [21, 24] by reducing the following loss:" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.806, + 0.892, + 0.835 + ], + "angle": 0, + "content": "\\[\n\\underset {t, \\epsilon , x _ {0}} {\\mathbb {E}} \\left[ w \\left(\\lambda_ {t}\\right) \\| x _ {0} - \\hat {x} _ {\\theta} \\left(x _ {t}, t\\right) \\| _ {2} ^ {2} \\right] \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.893, + 0.901 + ], + "angle": 0, + "content": "where \\( t \\) is sampled from a uniform distribution for timesteps, \\( \\epsilon \\) is sampled from the standard normal to noise the data \\( x_0 \\), \\( w(\\lambda_t) \\) is a timestep dependent weighting constant, and \\( \\hat{x}_{\\theta} \\) denotes the learned denoising model." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8499" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.098, + 0.092, + 0.468, + 0.106 + ], + "angle": 0, + "content": "We train our hierarchical LDM with the following losses:" + }, + { + "type": "equation", + "bbox": [ + 0.131, + 0.109, + 0.468, + 0.138 + ], + "angle": 0, + "content": "\\[\n\\underset {t, \\epsilon , g _ {0}} {\\mathbb {E}} \\left[ w \\left(\\lambda_ {t}\\right) \\| g _ {0} - \\psi_ {g} \\left(g _ {t}, t\\right) \\| _ {2} ^ {2} \\right] \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.132, + 0.141, + 0.468, + 0.169 + ], + "angle": 0, + "content": "\\[\n\\underset {t, \\epsilon , g _ {0}, c _ {0}} {\\mathbb {E}} \\left[ w \\left(\\lambda_ {t}\\right) \\| c _ {0} - \\psi_ {c} \\left(c _ {t}, g _ {0}, t\\right) \\| _ {2} ^ {2} \\right] \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.132, + 0.172, + 0.468, + 0.201 + ], + "angle": 0, + "content": "\\[\n\\underset {t, \\epsilon , g _ {0}, c _ {0}, f _ {0}} {\\mathbb {E}} \\left[ w \\left(\\lambda_ {t}\\right) \\| f _ {0} - \\psi_ {f} \\left(f _ {t}, g _ {0}, c _ {0}, t\\right) \\| _ {2} ^ {2} \\right] \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.206, + 0.468, + 0.401 + ], + "angle": 0, + "content": "where \\(\\psi\\) denotes the learnable denoising networks for \\(g, c, f\\). Fig. 4 visualizes the diffusion models. \\(\\psi_{g}\\) is implemented with linear layers with skip connections and \\(\\psi_{c}\\) and \\(\\psi_{f}\\) adopt the U-net architecture [55]. \\(g\\) is fed into \\(\\psi_{c}\\) and \\(\\psi_{f}\\) with conditional group normalization layers. \\(c\\) is interpolated and concatenated to the input to \\(\\psi_{f}\\). The camera poses contain the trajectory the camera is travelling, and this information can be useful for modelling a 3D scene as it tells the model where to focus on generating. Therefore, we concatenate the camera trajectory information to \\(g\\) and also learn to sample it. For brevity, we still call the concatenated vector \\(g\\). For conditional generation, each \\(\\psi\\) takes the conditioning variable as input with cross-attention layers [53]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.403, + 0.468, + 0.538 + ], + "angle": 0, + "content": "Each \\(\\psi\\) can be trained in parallel and, once trained, can be sampled one after another following the hierarchy. In practice, we use the v-prediction parameterization [57] that has been shown to have better convergence and training stability [56, 57]. Once \\(g, c, f\\) are sampled, we can use the latent decoder from Sec. 3.2 to construct the voxel \\(V\\) which represents the neural field for the sampled scene.. Following the volume rendering and decoding step in Sec. 3.1, the sampled scene can be visualized from desired viewpoints." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.542, + 0.432, + 0.557 + ], + "angle": 0, + "content": "3.4. Post-Optimizing Generated Neural Fields" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.565, + 0.468, + 0.686 + ], + "angle": 0, + "content": "Samples generated from our model on real-world data contain reasonable texture and geometry (Fig. 10), but can be further optimized by leveraging recent advances in 2D image diffusion models trained on orders of magnitude more data. Specifically, we iteratively update initially generated voxels, \\( V \\), by rendering viewpoints from the scene and applying Score Distillation Sampling (SDS) [46] loss on each image independently:" + }, + { + "type": "equation", + "bbox": [ + 0.085, + 0.691, + 0.468, + 0.722 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {V} L _ {S D S} = \\underset {\\epsilon , t, \\kappa} {\\mathbb {E}} \\left[ w \\left(\\lambda_ {t}\\right) \\left(\\epsilon - \\hat {\\epsilon_ {\\theta}} (r (V, \\kappa), t)\\right) \\frac {\\partial r (V , \\kappa)}{\\partial V} \\right] \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.729, + 0.468, + 0.88 + ], + "angle": 0, + "content": "where \\(\\kappa\\) is sampled uniformly in a \\(6m^2\\) region around the origin of the scene with random rotation about the vertical axis, \\(w(\\lambda_t)\\) is the weighting schedule used to train \\(\\hat{\\epsilon_{\\theta}}\\) and \\(t\\sim U[0.02T,0.2T]\\) where \\(T\\) is the amount of noise steps used to train \\(\\hat{\\epsilon_{\\theta}}\\). Note that for latent diffusion models, the noise prediction step is applied after encoding \\(r(V,\\kappa)\\) to the LDM's latent space and the partial gradient term is updated appropriately. For \\(\\hat{\\epsilon_{\\theta}}\\), we use an off-the-shelf latent diffusion model [53], finetuned to condition on CLIP image embeddings \\([47]^1\\). We found that CLIP contains a representativeness matrix" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.091, + 0.695, + 0.141 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.091, + 0.888, + 0.141 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.143, + 0.631, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.631, + 0.143, + 0.888, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.196, + 0.631, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.631, + 0.196, + 0.888, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.254, + 0.892, + 0.295 + ], + "angle": 0, + "content": "Figure 5. Datasets: Top-left: VizDoom [29]. Top-right: Replica [64]. Middle: Carla [12]. Bottom: AVD. For Carla and AVD, we visualize a subset of available cameras." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.315, + 0.892, + 0.576 + ], + "angle": 0, + "content": "tation of the quality of images that the LDM is able to interpret: denoising an image while conditioning on CLIP image embeddings of our model's samples produced images with similar geometry distortions and texture errors. We leverage this property by optimizing \\(L_{SDS}\\) with negative guidance. Letting \\(y, y'\\) be CLIP embeddings of clean image conditioning (e.g. dataset images) and artifact conditioning (e.g. samples) respectively, we perform classifier-free guidance [23] with conditioning vector \\(y\\), but replace the unconditional embedding with \\(y'\\). As shown in the supplementary, this is equivalent (up to scale) to sampling from \\(\\frac{p(x|y)^{\\alpha}}{p(x|y')}\\) at each denoising step where \\(\\alpha\\) controls the trade-off between sampling towards dataset images and away from images with artifacts. We stress that this post-optimization is only successful due to the strong scene prior contained in our voxel samples, as shown by our comparison to running optimization on randomly initialized voxels in the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.589, + 0.633, + 0.606 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.892, + 0.659 + ], + "angle": 0, + "content": "We evaluate NeuralField-LDM on the following four datasets (Fig. 5). Each dataset contains RGB images and a depth measurement with their corresponding camera poses." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.892, + 0.719 + ], + "angle": 0, + "content": "VizDoom [28] consists of front-view sensor observations obtained by navigating inside a synthetic game environment. We use the dataset provided by [9], which contains 34 trajectories with a length of 600." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.765 + ], + "angle": 0, + "content": "Replica [64] is a dataset of high-quality reconstructions of 18 indoor scenes, containing 101 front-view trajectories with a length of 100. We use the dataset provided by [9]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.766, + 0.892, + 0.87 + ], + "angle": 0, + "content": "Carla [12] is an open-source simulation platform for self-driving research. We mimic the camera settings for a self-driving car, by placing six cameras (front-left, front, front-right, back-left, back, back-right), covering 360 degrees with some overlaps, and move the car in a randomly sampled direction and distance for 10 timesteps. We sample 43K datapoints, each containing 60 images." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "AVD is an in-house dataset of human driving recordings in roads and parking lots. It has ten cameras with varying" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.36, + 0.9 + ], + "angle": 0, + "content": "\\(^{1}\\)https://github.com/justinpinkney/stable-diffusion" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8500" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.468, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.078, + 0.451, + 0.468, + 0.478 + ], + "angle": 0, + "content": "Figure 6. Generated Scenes: The top three rows are samples from Carla, and the bottom three rows are samples from AVD." + }, + { + "type": "image", + "bbox": [ + 0.116, + 0.482, + 0.428, + 0.744 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.746, + 0.468, + 0.786 + ], + "angle": 0, + "content": "Figure 7. We run marching-cubes [38] on the density voxels to visualize the geometry of the samples generated by NF-LDM. The structure of the scene is reflected well in the mesh." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.792, + 0.468, + 0.822 + ], + "angle": 0, + "content": "lens types along with Lidar for depth measurement. It has 73K sequences, each with 8 frames extracted at 10 fps." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.833, + 0.285, + 0.849 + ], + "angle": 0, + "content": "4.1. Baseline Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Unconditional Generation We evaluate the unconditional generation performance of NF-LDM by comparing it with baseline models. All results are without the post" + }, + { + "type": "table", + "bbox": [ + 0.525, + 0.09, + 0.866, + 0.182 + ], + "angle": 0, + "content": "
CriterionMethodDepthVizDoomReplica
FID (↓)GRAF [59]X47.5065.37
π-GAN [4]X143.55166.55
GSN [8]✓37.2141.75
GAUDI [2]✓33.7018.75
NF-LDM✓19.54*14.59
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.184, + 0.892, + 0.213 + ], + "angle": 0, + "content": "Table 1. FID [19] scores on VizDoom and Replica. NF-LDM outperforms all baseline models. Baseline numbers are from [2]." + }, + { + "type": "table", + "bbox": [ + 0.532, + 0.221, + 0.858, + 0.293 + ], + "angle": 0, + "content": "
CriterionMethodDepthCarlaAVD
FID (↓)EG3D [5]×76.89194.34
GSN [8]✓75.45166.07
NF-LDM✓35.6954.26
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.296, + 0.892, + 0.337 + ], + "angle": 0, + "content": "Table 2. FID [19] scores on Carla and AVD datasets. Baseline models have trouble learning the distribution of complex outdoor datasets, in particular AVD, while NF-LDM models them well." + }, + { + "type": "image", + "bbox": [ + 0.519, + 0.342, + 0.869, + 0.657 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.661, + 0.89, + 0.676 + ], + "angle": 0, + "content": "Figure 8. Reconstructing held-out scenes not seen during training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.901 + ], + "angle": 0, + "content": "optimization step (Sec. 3.4), unless specified. Tab. 1 shows the results on VizDoom and Replica. GRAF [59] and \\(\\pi\\)-GAN [4], which do not utilize ground truth depth in training, have shown successes in modelling single objects, but they exhibit worse performance than others that leverages depth information for modelling scenes. GAUDI [2] is an auto-decoder-based diffusion model. Their auto-decoder optimizes a small per-scene latent to reconstruct its matching scene. GAUDI comes with the advantage that learning the generative model is simple as it only needs to model the small dimensional latent distribution that acts as the key to their corresponding scenes. On the contrary, NF-LDM is trained on the latents that are a decomposition of the explicit 3D neural field. Therefore, GAUDI puts more modelling" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "8501" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.11, + 0.089, + 0.435, + 0.158 + ], + "angle": 0, + "content": "
CriterionMethodDepthCarlaAVD
FVD (↓)EG3D [5]X134.941232.38
GSN [8]✓184.301659.81
NF-LDM✓91.80242.50
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.162, + 0.47, + 0.19 + ], + "angle": 0, + "content": "Table 3. FVD [66] scores on Carla and AVD Datasets. As for FID, baseline models have trouble learning to model complex datasets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.213, + 0.47, + 0.334 + ], + "angle": 0, + "content": "capacity into the auto-decoder part, and NF-LDM puts it more into the generative model part. We attribute our improvement over GAUDI to our expressive hierarchical LDM that can model the details of the scenes better. In VizDoom, only one scene exists, and each sequence contains several hundred steps covering a large area in the scene, which our voxels were not large enough to encode. Therefore, we chunked each VizDoom trajectory to be 50 steps long." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.336, + 0.469, + 0.457 + ], + "angle": 0, + "content": "Tab. 2 shows results on complex outdoor datasets: Carla and AVD. We compare with EG3D [5] and GSN [8]. Both are GAN-based 3D generative models, but GSN utilizes ground truth depth measurements. Note that we did not include GAUDI [2] as the code was not available. NF-LDM achieves the best performance, and both baseline models have difficulty modelling the real outdoor dataset (AVD). Fig. 6 compares the samples from different models." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.459, + 0.47, + 0.686 + ], + "angle": 0, + "content": "Since the datasets are composed of frame sequences, we can treat them as videos and further evaluate with Fréchet Video Distance (FVD) [66] to compare the distributions of the dataset and sampled sequences. This can quantify samples' 3D structure by how natural the rendered sequence from a moving camera is. For EG3D and GSN, we randomly sample a trajectory from the datasets and for NF-LDM, we sample a trajectory from the global latent diffusion model. Tab. 3 shows that NF-LDM achieves the best results. We empirically observed GSN sometimes produced slightly inconsistent rendering, which could attribute to its lower FVD score than EG3D's. We also visualize the geometry of NF-LDM's samples by running marching-cubes [38] on the density voxels. Fig. 7 shows that our samples produce a coarse but realistic geometry." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.688, + 0.469, + 0.824 + ], + "angle": 0, + "content": "Ablations We evaluate the hierarchical structure of NF-LDM. Tab. 4 shows an ablation study on Carla. The model with the full hierarchy achieves the best performance. The global latent makes it easier for the other LDMs to sample as conditioning on the global properties of the scene (e.g. time of day) narrows down the distribution they need to model. The 2D fine latent helps retain the residual information missing in the 3D coarse latent, thus improving the latent auto-encoder and, consequently, the LDMs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Scene Reconstruction Unlike previous approaches, NF-LDM has an explicit scene auto-encoder that can be used for scene reconstruction. GAUDI [2] is auto-decoder based, so it is not trivial to infer a latent for a new scene. GSN [8] can invert a new scene using a GAN inversion method [52," + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.087, + 0.892, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.196, + 0.892, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.283, + 0.892, + 0.364 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.369, + 0.893, + 0.437 + ], + "angle": 0, + "content": "Figure 9. BEV-Conditioned Synthesis: NF-LDM allows controllable generation by editing the BEV segmentation map. From the initial sample, we add trees (green) and then edit the location of the car (blue). Note the ego car is at the center and thus not rendered." + }, + { + "type": "table", + "bbox": [ + 0.532, + 0.444, + 0.858, + 0.482 + ], + "angle": 0, + "content": "
Coarse lat. c+ Fine lat. f+ Global lat. g
FID (↓)46.4343.5235.69
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.483, + 0.892, + 0.525 + ], + "angle": 0, + "content": "Table 4. FID [19] on ablating the choice of hierarchy on the Carla dataset. The first column is for training both LAE and LDM only with the coarse latent. The last column is our full model." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.546, + 0.892, + 0.593 + ], + "angle": 0, + "content": "79], but as Fig. 8 shows, it fails to get the details of the scene correct. Our scene auto-encoder generalizes well and is scalable as the number of scenes grow." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.604, + 0.76, + 0.62 + ], + "angle": 0, + "content": "4.2. Applications and Limitations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.628, + 0.892, + 0.749 + ], + "angle": 0, + "content": "Conditional Synthesis NF-LDM can utilize additional conditioning signals for controllable generation. In this paper, we consider Bird's Eye View (BEV) segmentation maps, but our model can be extended to other conditioning variables. We use cross attention layers [53], which have been shown to be effective for conditional synthesis. Fig. 9 shows that NF-LDM follows the given BEV map faithfully and how the map can be edited for controllable synthesis." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Scene Editing Image diffusion models can be used for image inpainting without explicit training on the task [53, 63]. We leverage this property to edit scenes in 3D by resampling a region in the 3D coarse latent \\( c \\). Specifically, at each denoising step, we noise the region to be kept and concatenate with the region being sampled, and pass it through the diffusion model. We use reconstruction guidance [24] to better harmonize the sampled and kept regions. After we get a new \\( c \\), the fine latent is also re-sampled conditioned on \\( c \\). Fig. 11 shows results on scene editing with NF-LDM." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8502" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.139, + 0.089, + 0.833, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.35, + 0.895, + 0.38 + ], + "angle": 0, + "content": "Figure 10. Panoramas from NF-LDM's samples: From the initial sample at the top, we apply post-optimization with Score Distillation Sampling [46] (Sec. 3.4). (a) demonstrates improved sample quality. (b) showcases style modification by conditioning on evening scenes." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.39, + 0.47, + 0.727 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.731, + 0.471, + 0.775 + ], + "angle": 0, + "content": "Figure 11. Scene Editing: We use the 3D coarse latent \\( c \\) for scene editing. From the initial sample indicated by light green, we resample a part of the latent, indicated by dark green." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Post-Optimization Fig. 10 shows how post-optimization (Sec. 3.4) can improve the quality of NF-LDM's initial sample while retaining the 3D structure. In addition to improving quality, we can also modify scene properties, such as time of day and weather, by conditioning the LDM on images with the desired properties. SDS-based style modification is effective for changes where a set of" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.394, + 0.892, + 0.47 + ], + "angle": 0, + "content": "clean image data is available with the desired property and is reasonably close to our dataset's domain (e.g. street images for AVD). In the supplementary, we also provide results experimenting with directional CLIP loss [14] to quickly finetune our scene decoder for a given text prompt." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.47, + 0.893, + 0.68 + ], + "angle": 0, + "content": "Limitations NF-LDM's hierarchical structure and three stage pipeline allows us to achieve high-quality generations and reconstructions, but it comes with a degradation in training time and sampling speed. In this work, the neural field representation is based on dense voxel grids, and it becomes expensive to volume render and learn the diffusion models as they get larger. Therefore, exploring alternative sparse representations is a promising future direction. Lastly, our method requires multi-view images which limits data availability and therefore risks universal problems in generative modelling of overfitting. For example, we found that output samples in AVD had limited diversity because the dataset itself was recorded in a limited number of scenes." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.694, + 0.62, + 0.71 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.903 + ], + "angle": 0, + "content": "We introduced NeuralField-LDM (NF-LDM), a generative model for complex 3D environments. NF-LDM first constructs an expressive latent distribution by encoding input images into a 3D neural field representation which is further compressed into more abstract latent spaces. Then, our proposed hierarchical LDM is fit onto the latent spaces, achieving state-of-the-art performance on 3D scene generation. NF-LDM enables a diverse set of applications, including controllable scene generation and scene editing. Future directions include exploring more efficient sparse voxel representations, training on larger-scale real-world data and learning to continuously expand generated scenes." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.958 + ], + "angle": 0, + "content": "8503" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.156 + ], + "angle": 0, + "content": "[1] Martin Arjovsky and Léon Bottou. Towards principled methods for training generative adversarial networks. In ICLR, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.159, + 0.472, + 0.228 + ], + "angle": 0, + "content": "[2] Miguel Ángel Bautista, Pengsheng Guo, Samira Abnar, Walter Talbott, Alexander Toshev, Zhuoyuan Chen, Laurent Dinh, Shuangfei Zhai, Hanlin Goh, Daniel Ulbricht, Afshin Dehghan, and Josh M. Susskind. GAUDI: A neural architect for immersive 3d scene generation. 2022. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.231, + 0.47, + 0.271 + ], + "angle": 0, + "content": "[3] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale GAN training for high fidelity natural image synthesis. In ICLR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.273, + 0.468, + 0.327 + ], + "angle": 0, + "content": "[4] Eric Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. arXiv, 2012.00926, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.33, + 0.468, + 0.4 + ], + "angle": 0, + "content": "[5] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In arXiv, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.402, + 0.468, + 0.47 + ], + "angle": 0, + "content": "[6] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In CVPR, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.472, + 0.468, + 0.526 + ], + "angle": 0, + "content": "[7] Eric R. Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. Pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.529, + 0.468, + 0.584 + ], + "angle": 0, + "content": "[8] Terrance DeVries, Miguel Angel Bautista, Nitish Srivastava, Graham W. Taylor, and Joshua M. Susskind. Unconstrained scene generation with locally conditioned radiance fields. In ICCV, 2021. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.586, + 0.468, + 0.64 + ], + "angle": 0, + "content": "[9] Terrance DeVries, Miguel Angel Bautista, Nitish Srivastava, Graham W. Taylor, and Joshua M. Susskind. Unconstrained scene generation with locally conditioned radiance fields. arXiv, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.643, + 0.468, + 0.698 + ], + "angle": 0, + "content": "[10] Prafulla Dhariwal and Alexander Quinn Nichol. Diffusion models beat gans on image synthesis. In Marc'Aurelio Ranzato, Alina Beygelzimer, Yann N. Dauphin, Percy Liang, and Jennifer Wortman Vaughan, editors, NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.7, + 0.468, + 0.741 + ], + "angle": 0, + "content": "[11] Tim Dockhorn, Arash Vahdat, and Karsten Kreis. Score-based generative modeling with critically-damped Langevin diffusion. In ICLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.743, + 0.468, + 0.784 + ], + "angle": 0, + "content": "[12] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. CARLA: An open urban driving simulator. 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.786, + 0.468, + 0.827 + ], + "angle": 0, + "content": "[13] Patrick Esser, Robin Rombach, and Björn Ommer. Taming transformers for high-resolution image synthesis. In CVPR, 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.83, + 0.468, + 0.871 + ], + "angle": 0, + "content": "[14] Rinon Gal, Or Patashnik, Haggai Maron, Gal Chechik, and Daniel Cohen-Or. Stylegan-nada: Clip-guided domain adaptation of image generators, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.468, + 0.902 + ], + "angle": 0, + "content": "[15] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. In Advances In Neural Information Processing Systems, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.136, + 0.892, + 0.163 + ], + "angle": 0, + "content": "[16] Ian Goodfellow. Nips 2016 tutorial: Generative adversarial networks. arXiv preprint arXiv:1701.00160, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.165, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[17] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron C. Courville, and Yoshua Bengio. Generative adversarial nets. In NeurIPS, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.221, + 0.892, + 0.262 + ], + "angle": 0, + "content": "[18] William Harvey, Saeid Naderiparizi, Vaden Masrani, Christian Weilbach, and Frank Wood. Flexible diffusion modeling of long videos, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.264, + 0.892, + 0.319 + ], + "angle": 0, + "content": "[19] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.321, + 0.892, + 0.375 + ], + "angle": 0, + "content": "[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Hugo Larochelle, Marc' Aurelio Ranzato, Raia Hadsell, Maria-Florina Balcan, and Hsuan-Tien Lin, editors, NeurIPS, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.377, + 0.892, + 0.446 + ], + "angle": 0, + "content": "[21] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 6840-6851. Curran Associates, Inc., 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.448, + 0.892, + 0.502 + ], + "angle": 0, + "content": "[22] Jonathan Ho, Chitwan Sahara, William Chan, David J. Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. J. Mach. Learn. Res., 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.504, + 0.892, + 0.533 + ], + "angle": 0, + "content": "[23] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.534, + 0.892, + 0.574 + ], + "angle": 0, + "content": "[24] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. arXiv:2204.03458, 2022. 2, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.576, + 0.892, + 0.617 + ], + "angle": 0, + "content": "[25] James T Kajiya and Brian P Von Herzen. Ray tracing volume densities. ACM SIGGRAPH computer graphics, 18(3):165-174, 1984. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.619, + 0.892, + 0.674 + ], + "angle": 0, + "content": "[26] Amlan Kar, Aayush Prakash, Ming-Yu Liu, Eric Cameracci, Justin Yuan, Matt Rusiniak, David Acuna, Antonio Torralba, and Sanja Fidler. Meta-sim: Learning to generate synthetic datasets. In ICCV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.676, + 0.892, + 0.717 + ], + "angle": 0, + "content": "[27] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.719, + 0.892, + 0.788 + ], + "angle": 0, + "content": "[28] Michal Kempka, Marek Wydmuch, Grzegorz Runc, Jakub Toczek, and Wojciech Jaskowski. Vizdoom: A doom-based ai research platform for visual reinforcement learning. In 2016 IEEE conference on computational intelligence and games (CIG), pages 1-8. IEEE, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.789, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[29] Margret Keuper, Siyu Tang, Bjoern Andres, Thomas Brox, and Bernt Schiele. Motion segmentation & multiple object tracking by correlation co-clustering. IEEE TPAMI, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[30] Seung Wook Kim, Karsten Kreis, Daiqing Li, Antonio Torralba, and Sanja Fidler. Polymorphic-gan: Generating aligned samples across multiple domains with learned morph maps. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Jun. 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8504" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[31] Seung Wook Kim, Jonah Philion, Antonio Torralba, and Sanja Fidler. Drivegan: Towards a controllable high-quality neural simulation. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.472, + 0.207 + ], + "angle": 0, + "content": "[32] Seung Wook Kim, Yuhao Zhou, Jonah Philion, Antonio Torralba, and Sanja Fidler. Learning to simulate dynamic environments with gamean. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1231-1240, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.208, + 0.471, + 0.235 + ], + "angle": 0, + "content": "[33] Diederik P. Kingma and Max Welling. Auto-encoding variational bayes. *ICLR*, 2014. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.238, + 0.471, + 0.293 + ], + "angle": 0, + "content": "[34] Daiqing Li, Huan Ling, Seung Wook Kim, Karsten Kreis, Adela Barriuso, Sanja Fidler, and Antonio Torralba. Big-datasetgan: Synthesizing imagenet with pixel-wise annotations. arXiv, 2201.04684, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.295, + 0.471, + 0.323 + ], + "angle": 0, + "content": "[35] Ke Li and Jitendra Malik. On the implicit assumptions of gans. arXiv preprint arXiv:1811.12402, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.325, + 0.471, + 0.367 + ], + "angle": 0, + "content": "[36] Yiyi Liao, Katja Schwarz, Lars Mescheder, and Andreas Geiger. Towards unsupervised learning of generative models for 3d controllable image synthesis. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.369, + 0.471, + 0.41 + ], + "angle": 0, + "content": "[37] Huan Ling, Karsten Kreis, Daiqing Li, Seung Wook Kim, Antonio Torralba, and Sanja Fidler. Editgan: High-precision semantic image editing. In NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.412, + 0.471, + 0.453 + ], + "angle": 0, + "content": "[38] William E. Lorensen and Harvey E. Cline. Marching cubes: A high resolution 3d surface construction algorithm. In ACM Trans. on Graphics, 1987. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.456, + 0.471, + 0.484 + ], + "angle": 0, + "content": "[39] Shitong Luo and Wei Hu. Diffusion probabilistic models for 3d point cloud generation. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.486, + 0.471, + 0.541 + ], + "angle": 0, + "content": "[40] Ricardo Martin-Brualla, Noha Radwan, Mehdi S. M. Sajjadi, Jonathan T. Barron, Alexey Dosovitskiy, and Daniel Duckworth. NeRF in the Wild: Neural Radiance Fields for Unconstrained Photo Collections. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.543, + 0.471, + 0.598 + ], + "angle": 0, + "content": "[41] Lars Mescheder, Andreas Geiger, and Sebastian Nowozin. Which training methods for gans do actually converge? In Proc. of the International Conf. on Machine learning (ICML), 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.601, + 0.471, + 0.656 + ], + "angle": 0, + "content": "[42] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.658, + 0.471, + 0.769 + ], + "angle": 0, + "content": "[43] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. GLIDE: towards photorealistic image generation and editing with text-guided diffusion models. In Kamalika Chaudhuri, Stefanie Jegelka, Le Song, Csaba Szepesvári, Gang Niu, and Sivan Sabato, editors, Proc. of the International Conf. on Machine learning (ICML), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.772, + 0.471, + 0.813 + ], + "angle": 0, + "content": "[44] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.816, + 0.471, + 0.856 + ], + "angle": 0, + "content": "[45] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In ECCV, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.471, + 0.9 + ], + "angle": 0, + "content": "[46] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv, 2022. 1, 2, 5, 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.161 + ], + "angle": 0, + "content": "[47] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. arXiv, 2103.00020, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[48] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with CLIP latents. arXiv, abs/2204.06125, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.206, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[49] Ali Razavi, Aaron van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with VQ-VAE-2. In NeurIPS, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.303 + ], + "angle": 0, + "content": "[50] Konstantinos Rematas, Andrew Liu, Pratul P. Srinivasan, Jonathan T. Barron, Andrea Tagliasacchi, Thomas A. Funkhouser, and Vittorio Ferrari. Urban radiance fields. In CVPR, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.305, + 0.892, + 0.361 + ], + "angle": 0, + "content": "[51] Danilo Jimenez Rezende, Shakir Mohamed, and Daan Wierstra. Stochastic backpropagation and approximate inference in deep generative models. In Proc. of the International Conf. on Machine learning (ICML), 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.362, + 0.892, + 0.432 + ], + "angle": 0, + "content": "[52] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a stylegan encoder for image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2287–2296, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.434, + 0.892, + 0.488 + ], + "angle": 0, + "content": "[53] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 3, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.49, + 0.892, + 0.531 + ], + "angle": 0, + "content": "[54] Robin Rombach, Patrick Esser, and Björn Ommer. Geometry-free view synthesis: Transformers and no 3d priors. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.533, + 0.892, + 0.602 + ], + "angle": 0, + "content": "[55] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234-241. Springer, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.604, + 0.892, + 0.687 + ], + "angle": 0, + "content": "[56] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.688, + 0.892, + 0.729 + ], + "angle": 0, + "content": "[57] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. arXiv preprint arXiv:2202.00512, 2022. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.731, + 0.892, + 0.772 + ], + "angle": 0, + "content": "[58] Axel Sauer, Katja Schwarz, and Andreas Geiger. StyleganxI: Scaling stylegan to large diverse datasets. ACM Trans. on Graphics, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.774, + 0.892, + 0.814 + ], + "angle": 0, + "content": "[59] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. Graf: Generative radiance fields for 3d-aware image synthesis. 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.817, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[60] Katja Schwarz, Axel Sauer, Michael Niemeyer, Yiyi Liao, and Andreas Geiger. Voxgraf: Fast 3d-aware image synthesis with sparse voxel grids. In NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.86, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[61] Yujun Shen, Jinjin Gu, Xiaou Tang, and Bolei Zhou. Interpreting the latent space of gans for semantic face editing. In CVPR, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.957 + ], + "angle": 0, + "content": "8505" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[62] Yuan Shen, Wei-Chiu Ma, and Shenlong Wang. SGAM: Building a virtual 3d world through simultaneous generation and mapping. In Thirty-Sixth Conference on Neural Information Processing Systems, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.47, + 0.205 + ], + "angle": 0, + "content": "[63] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.207, + 0.471, + 0.344 + ], + "angle": 0, + "content": "[64] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J. Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, Anton Clarkson, Mingfei Yan, Brian Budge, Yajie Yan, Xiaqing Pan, June Yon, Yuyang Zou, Kimberly Leon, Nigel Carter, Jesus Briales, Tyler Gillingham, Elias Mueggler, Luis Pesqueira, Manolis Savva, Dhruv Batra, Hauke M. Strasdat, Renzo De Nardi, Michael Goesele, Steven Lovegrove, and Richard Newcombe. The Replica dataset: A digital replica of indoor spaces. arXiv, 1906.05797, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.346, + 0.469, + 0.387 + ], + "angle": 0, + "content": "[65] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.389, + 0.469, + 0.444 + ], + "angle": 0, + "content": "[66] Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.446, + 0.469, + 0.5 + ], + "angle": 0, + "content": "[67] Arash Vahdat and Jan Kautz. NVAE: A deep hierarchical variational autoencoder. In Hugo Larochelle, Marc'Aurelio Ranzato, Raia Hadsell, Maria-Florina Balcan, and Hsuan-Tien Lin, editors, NeurIPS, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.502, + 0.469, + 0.53 + ], + "angle": 0, + "content": "[68] Arash Vahdat, Karsten Kreis, and Jan Kautz. Score-based generative modeling in latent space. In NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.532, + 0.469, + 0.572 + ], + "angle": 0, + "content": "[69] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.575, + 0.469, + 0.642 + ], + "angle": 0, + "content": "[70] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P. Srinivasan, Howard Zhou, Jonathan T. Barron, Ricardo Martin-Brualla, Noah Snavely, and Thomas A. Funkhouser. Ibrnet: Learning multi-view image-based rendering. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.645, + 0.469, + 0.686 + ], + "angle": 0, + "content": "[71] Yuxin Wu and Kaiming He. Group normalization. In Proceedings of the European conference on computer vision (ECCV), pages 3-19, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.688, + 0.469, + 0.728 + ], + "angle": 0, + "content": "[72] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelNeRF: Neural radiance fields from one or few images. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.731, + 0.469, + 0.784 + ], + "angle": 0, + "content": "[73] Xiaohui Zeng, Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, and Karsten Kreis. LION: latent point diffusion models for 3d shape generation. In NeurIPS, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.788, + 0.469, + 0.828 + ], + "angle": 0, + "content": "[74] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv, 2010.07492, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.831, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[75] Xiaoshuai Zhang, Sai Bi, Kalyan Sunkavalli, Hao Su, and Zexiang Xu. Nerfusion: Fusing radiance fields for large-scale scene reconstruction. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[76] Yuxuan Zhang, Wenzheng Chen, Huan Ling, Jun Gao, Yi-nan Zhang, Antonio Torralba, and Sanja Fidler. Image gans" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "meet differentiable rendering for inverse graphics and interpretable 3d neural rendering. In ICLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.163 + ], + "angle": 0, + "content": "[77] Yan Zheng, Lemeng Wu, Xingchao Liu, Zhen Chen, Qiang Liu, and Qixing Huang. Neural volumetric mesh generator. arXiv preprint arXiv:2210.03158, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.165, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[78] Linqi Zhou, Yilun Du, and Jiajun Wu. 3d shape generation and completion through point-voxel diffusion. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.207, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[79] Jiapeng Zhu, Yujun Shen, Deli Zhao, and Bolei Zhou. Indomain gan inversion for real image editing. In European conference on computer vision, pages 592-608. Springer, 2020. 7" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8506" + } + ] +] \ No newline at end of file diff --git a/2023/NeuralField-LDM_ Scene Generation With Hierarchical Latent Diffusion Models/81bfb70e-1652-4684-aa6c-4729296b6ed4_origin.pdf b/2023/NeuralField-LDM_ Scene Generation With Hierarchical Latent Diffusion Models/81bfb70e-1652-4684-aa6c-4729296b6ed4_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..87c7ad238e2170c3da1e46991f1698b325d3c851 --- /dev/null +++ b/2023/NeuralField-LDM_ Scene Generation With Hierarchical Latent Diffusion Models/81bfb70e-1652-4684-aa6c-4729296b6ed4_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:794e1fefe2b40be6cd1161afbb9508b5ac756c8ab01dd4a41d63cd0831fe0ae2 +size 3723672 diff --git a/2023/NeuralField-LDM_ Scene Generation With Hierarchical Latent Diffusion Models/full.md b/2023/NeuralField-LDM_ Scene Generation With Hierarchical Latent Diffusion Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..31b511e56826a55bdd61b0b9ef3dbed9aec22e01 --- /dev/null +++ b/2023/NeuralField-LDM_ Scene Generation With Hierarchical Latent Diffusion Models/full.md @@ -0,0 +1,341 @@ +# NeuralField-LDM: Scene Generation with Hierarchical Latent Diffusion Models + +Seung Wook Kim $^{1,2,3*}$ Bradley Brown $^{1,5*†}$ Kangxue Yin $^{1}$ Karsten Kreis $^{1}$ Katja Schwarz $^{6†}$ + +Daiqing Li + +Robin Rombach $^{7\dagger}$ + +Antonio Torralba + +Sanja Fidler1,2,3 + +$^{1}$ NVIDIA $^{2}$ University of Toronto $^{3}$ Vector Institute $^{4}$ CSAIL, MIT $^{5}$ University of Waterloo + +6University of Tübingen, Tübingen AI Center 7LMU Munich + +![](images/fa25ef0555bea1ba507b5b1f853b4ad8e1acf043ea8867875ede42fa25075a26.jpg) + +![](images/a7c0c2a5c50443499d118db2c92fd0798cee8f89ece0862a8e9153c7b05d74a0.jpg) +Figure 1. We introduce NeuralField-LDM, a generative model for complex open-world 3D scenes. This figure contains a panorama constructed from NeuralField-LDM's generated scene. We visualize different parts of the scene by placing cameras on them. + +![](images/22cf9cbbf8ae11a68b0fd11712a7c9faabafac4044bd2b69985eb6d8ec6d2165.jpg) + +![](images/dbb85e17ddc79672b02f433511e29dfa4d317a403bd2192d0ce4b6c948398a3e.jpg) + +![](images/f816f399f8022c2169295cabfa1b364ea4ffb738f2b78f8bcf11db12fb5001df.jpg) + +![](images/c4102ba302da853b621c88a79df028cebaabdc3234d53438413f2ac55e858c6b.jpg) + +# Abstract + +Automatically generating high-quality real world 3D scenes is of enormous interest for applications such as virtual reality and robotics simulation. Towards this goal, we introduce NeuralField-LDM, a generative model capable of synthesizing complex 3D environments. We leverage Latent Diffusion Models that have been successfully utilized for efficient high-quality 2D content creation. We first train a scene auto-encoder to express a set of image and pose pairs as a neural field, represented as density and feature voxel grids that can be projected to produce novel views of the scene. To further compress this representation, we train a latent-autoencoder that maps the voxel grids to a set of latent representations. A hierarchical diffusion model is then fit to the latents to complete the scene generation pipeline. We achieve a substantial improvement over existing state-of-the-art scene generation models. Additionally, we show how NeuralField-LDM can be used for a variety of 3D content creation applications, including conditional scene generation, scene inpainting and scene style manipulation. + +# 1. Introduction + +There has been increasing interest in modelling 3D real-world scenes for use in virtual reality, game design, digi- + +tal twin creation and more. However, designing 3D worlds by hand is a challenging and time-consuming process, requiring 3D modeling expertise and artistic talent. Recently, we have seen success in automating 3D content creation via 3D generative models that output individual object assets [15,46,73]. Although a great step forward, automating the generation of real-world scenes remains an important open problem and would unlock many applications ranging from scalably generating a diverse array of environments for training AI agents (e.g. autonomous vehicles) to the design of realistic open-world video games. In this work, we take a step towards this goal with NeuralField-LDM (NF-LDM), a generative model capable of synthesizing complex real-world 3D scenes. NF-LDM is trained on a collection of posed camera images and depth measurements which are easier to obtain than explicit ground-truth 3D data, offering a scalable way to synthesize 3D scenes. + +Recent approaches [2, 6, 8] tackle the same problem of generating 3D scenes, albeit on less complex data. In [6, 8], a latent distribution is mapped to a set of scenes using adversarial training, and in GAUDI [2], a denoising diffusion model is fit to a set of scene latents learned using an auto-decoder. These models all have an inherent weakness of attempting to capture the entire scene into a single vector that conditions a neural radiance field. In practice, we find that this limits the ability to fit complex scene distributions. + +Recently, diffusion models have emerged as a very powerful class of generative models, capable of generating high- + +quality images, point clouds and videos [18, 24, 39, 48, 53, 73, 78]. Yet, due to the nature of our task, where image data must be mapped to a shared 3D scene without an explicit ground truth 3D representation, straightforward approaches fitting a diffusion model directly to data are infeasible. + +In NeuralField-LDM, we learn to model scenes using a three-stage pipeline. First, we learn an auto-encoder that encodes scenes into a neural field, represented as density and feature voxel grids. Inspired by the success of latent diffusion models for images [53], we learn to model the distribution of our scene voxels in latent space to focus the generative capacity on core parts of the scene and not the extraneous details captured by our voxel auto-encoders. Specifically, a latent-autoencoder decomposes the scene voxels into a 3D coarse, 2D fine and 1D global latent. Hierarchical diffusion models are then trained on the tri-latent representation to generate novel 3D scenes. We show how NF-LDM enables applications such as scene editing, birdeye view conditional generation and style adaptation. Finally, we demonstrate how score distillation [46] can be used to optimize the quality of generated neural fields, allowing us to leverage the representations learned from state-of-the-art image diffusion models that have been exposed to orders of magnitude more data. + +Our contributions are: 1) We introduce NF-LDM, a hierarchical diffusion model capable of generating complex open-world 3D scenes and achieving state of the art scene generation results on four challenging datasets. 2) We extend NF-LDM to semantic birds-eye view conditional scene generation, style modification and 3D scene editing. + +# 2. Related Work + +2D Generative Models In past years, generative adversarial networks (GANs) [3, 17, 27, 41, 58] and likelihood-based approaches [33, 49, 51, 67] enabled high-resolution photorealistic image synthesis. Due to their quality, GANs are used in a multitude of downstream applications ranging from steerable content creation [30, 34, 36, 37, 61, 76] to data driven simulation [26, 31, 32, 34]. Recently, autoregressive models and score-based models, e.g. diffusion models, demonstrate better distribution coverage while preserving high sample quality [10, 11, 13, 20, 22, 43, 48, 53, 54, 68]. Since evaluation and optimization of these approaches in pixel space is computationally expensive, [53, 68] apply them to latent space, achieving state-of-the-art image synthesis at megapixel resolution. As our approach operates on 3D scenes, computational efficiency is crucial. Hence, we build upon [53] and train our model in latent space. + +Novel View Synthesis In their seminal work [42], Mildenhall et al. introduce Neural Radiance Fields (NeRF) as a powerful 3D representation. PixelNeRF [72] and IBR-Net [70] propose to condition NeRF on aggregated features from multiple views to enable novel view synthesis from + +a sparse set of views. Another line of works scale NeRF to large-scale indoor and outdoor scenes [40, 50, 74, 75]. Recently, Nerfusion [75] predicts local radiance fields and fuses them into a scene representation using a recurrent neural network. Similarly, we construct a latent scene representation by aggregating features across multiple views. Different from the aforementioned methods, our approach is a generative model capable of synthesizing novel scenes. + +3D Diffusion Models A few recent works propose to apply denoising diffusion models (DDM) [20, 22, 63] on point clouds for 3D shape generation [39, 73, 78]. While PVD [78] trains on point clouds directly, DPM [39] and LION [73] use a shape latent variable. Similar to LION, we design a hierarchical model by training separate conditional DDMs. However, our approach generates both texture and geometry of a scene without needing 3D ground truth as supervision. + +3D-Aware Generative Models 3D-aware generative models synthesize images while providing explicit control over the camera pose and potentially other scene properties, like object shape and appearance. SGAM [62] generates a 3D scene by autoregressively generating sensor data and building a 3D map. Several previous approaches generate NeRFs of single objects with conditional coordinate-based MLPs [7, 44, 59]. GSN [8] conditions a coordinate-based MLP on a "floor plan", i.e., a 2D feature map, to model more complex indoor scenes. EG3D [6] and VoxGRAF [60] use convolutional backbones to generate 3D representations. All of these approaches rely on adversarial training. Instead, we train a DDM on voxels in latent space. The work closest to ours is GAUDI [2], which first trains an auto-decoder and subsequently trains a DDM on the learned latent codes. Instead of using a global latent code, we encode scenes onto voxel grids and train a hierarchical DDM to optimally combine global and local features. + +# 3. NeuralField-LDM + +Our objective is to train a generative model to synthesize 3D scenes that can be rendered to any viewpoint. We assume access to a dataset $\{(i,\kappa ,\rho)\}_{1..N}$ which consists of $N$ RGB images $i$ and their camera poses $\kappa$ , along with a depth measurement $\rho$ that can be either sparse (e.g. Lidar points) or dense. The generative model must learn to model both the texture and geometry distributions of the dataset in 3D by learning solely from the sensor observations, which is a highly non-trivial problem. + +Past work typically tackles this problem with a generative adversarial network (GAN) framework [6, 8, 59, 60]. They produce an intermediate 3D representation and render images for a given viewpoint with volume rendering [25, 42]. Discriminator losses then ensure that the 3D representation produces a valid image from any viewpoint. However, GANs come with notorious training instability + +![](images/b1a3e5073629a6e122828874c18ea1f3ac41f1f030a4b040e2bbff35986fe11f.jpg) +Figure 2. Overview of NeuralField-LDM. We first encode RGB images with camera poses into a neural field represented by density and feature voxel grids. We compress the neural field into smaller latent spaces and fit a hierarchical latent diffusion model on the latent space. Sampled latents can then be decoded into a neural field that can be rendered into a given viewpoint. + +and mode dropping behaviors [1, 16, 35]. Denoising Diffusion models [20] (DDMs) have recently emerged as an alternative to GANs that avoid the aforementioned disadvantages [53, 56, 57]. However, DDMs model the data likelihood explicitly and are trained to reconstruct the training data. Thus, they have been used in limited scenarios [73, 77] since ground-truth 3D data is not readily available at scale. + +To tackle the challenging problem of generating an entire scene with texture and geometry, we take inspiration from latent diffusion models (LDM) [53], which first construct an intermediate latent distribution of the training data then fit a diffusion model on the latent distribution. In Sec. 3.1, we introduce a scene auto-encoder that encodes the set of RGB images into a neural field representation consisting of density and feature voxel grids. To accurately capture a scene, the voxel grids' spatial dimension needs to be much larger than what current state-of-the-art LDMs can model. In Sec. 3.2, we show how we can further compress and decompose the explicit voxel grids into compressed latent representations to facilitate learning the data distribution. Finally, Sec. 3.3 introduces a latent diffusion model that models the latent distributions in a hierarchical manner. Fig. 2 shows an overview of our method, which we name NeuralField-LDM (NF-LDM). We provide training and additional architecture details in the supplementary. + +# 3.1. Scene Auto-Encoder + +The goal of the scene auto-encoder is to obtain a 3D representation of the scene from input images by learning to reconstruct them. Fig. 3 depicts the auto-encoding process. The scene encoder is a 2D CNN and processes each RGB image $i_{1..N}$ separately, producing a $\mathbb{R}^{H\times W\times (D + C)}$ dimensional 2D tensor for each image, where $H$ and $W$ are smaller than $i$ 's size. We follow a similar procedure to Lift-Splat-Shoot (LSS) [45] to lift each 2D image feature map and combine them in the common voxel-based 3D neural field. We build a discrete frustum of size $H\times W\times D$ with the camera poses $\kappa$ for each image. This frustum contains image features and density values for each pixel, along a pre-defined discrete set of $D$ depths. Unlike LSS, we take the first $D$ channels of the 2D CNN's output and use them as + +density values. That is, the $d$ th channel of the CNN's output at pixel $(h,w)$ becomes the density value of the frustum entry at $(h,w,d)$ . Motivated by the volume rendering equation [42], we get the occupancy weight $O$ of each element $(h,w,d)$ in the frustum using the density values $\sigma \geq 0$ : + +$$ +O (h, w, d) = \exp \left(- \sum_ {j = 0} ^ {d - 1} \sigma_ {(h, w, j)} \delta_ {j}\right) \left(1 - \exp \left(- \sigma_ {(h, w, d)} \delta_ {d}\right)\right) \tag {1} +$$ + +where $h, w$ denotes the pixel coordinate of the frustum and $\delta_{j}$ is the distance between each depth in the frustum. Using the occupancy weights, we put the last $C$ channels of the CNN's output into the frustum $F$ : + +$$ +F (h, w, d) = \left[ O (h, w, d) \phi (h, w), \sigma (h, w, d) \right] \tag {2} +$$ + +where $\phi (h,w)$ denotes the $C$ -channeled feature vector at pixel $(h,w)$ which is scaled by $O(h,w,d)$ for $F$ at depth $d$ + +After constructing the frustum for each view, we transform the frustums to world coordinates and fuse them into a shared 3D neural field, represented as density and feature voxel grids. Let $V_{\text{Density}}$ and $V_{\text{Feat}}$ denote the density and feature grid, respectively. This formulation of representing a scene with density and feature grids has been explored before [65] for optimization-based scene reconstruction and we utilize it as an intermediate representation for our scene auto-encoder. $V_{\text{Density}, \text{Feat}}$ have the same spatial size, and each voxel in $V$ represents a region in the world coordinate system. For each voxel indexed by $(x, y, z)$ , we pool all densities and features of the corresponding frustum entries. In this paper, we simply take the mean of the pooled features. More sophisticated pooling functions (e.g. attention) can be used, which we leave as future work. + +Finally, we perform volume rendering using the camera poses $\kappa$ to project $V$ onto a 2D feature map. We trilinearly interpolate the values on each voxel to get the feature and density for each sampling point along the camera rays. 2D features are then fed into a CNN decoder that produces the output image $\hat{i}$ . We denote rendering of voxels to output images as $\hat{i} = r(V,\kappa)$ . From the volume rendering process, we also get the expected depth $\hat{\rho}$ along each ray [50]. The scene + +![](images/01e062238090e1950f0f6a8a49e12e05623432fbde673f153e72ed7cd1dde40f.jpg) +Figure 3. Scene Auto-Encoder: Each input image is processed with a 2D CNN then lifted up to 3D and merged into the shared voxel grids. Density prediction is not shown here for brevity. + +auto-encoding pipeline is trained with an image reconstruction loss $||i - \hat{i}||$ and a depth supervision loss $||\rho - \hat{\rho}||$ . In the case of sparse depth measurements, we only supervise the pixels with recorded depth. We can further improve the quality of the auto-encoder with adversarial loss as in VQGAN [13] or by doing a few optimization steps at inference time, which we discuss in the supplementary. + +# 3.2. Latent Voxel Auto-Encoder + +It is possible to fit a generative model on voxel grids obtained from Sec. 3.1. However, to capture real-world scenes, the dimensionality of the representation needs to be much larger than what SOTA diffusion models can be trained on. For example, Imagen [56] trains DDMs on $256 \times 256$ RGB images, and we use voxels of size $128 \times 128 \times 32$ with 32 channels. We thus introduce a latent auto-encoder (LAE) that compresses voxels into a 128-dimensional global latent as well as coarse (3D) and fine (2D) quantized latents with channel dimensions of four and spatial dimensions $32 \times 32 \times 16$ and $128 \times 128$ respectively. + +We concatenate $V_{\text{Density}}$ and $V_{\text{Feat}}$ along the channel dimension and use separate CNN encoders to encode the voxel grid $V$ into a hierarchy of three latents: 1D global latent $g$ , 3D coarse latent $c$ , and 2D fine latent $f$ , as shown in Fig. 2. The intuition for this design is that $g$ is responsible for representing the global properties of the scene, such as the time of the day, $c$ represents coarse 3D scene structure, and $f$ is a 2D tensor with the same horizontal size $X \times Y$ as $V$ , which gives further details for each location $(x, y)$ in bird's eye view perspective. We empirically found that 2D CNNs perform similarly to 3D CNNs while being more efficient, thus we use 2D CNNs throughout. To use 2D CNNs for the 3D input $V$ , we concatenate $V$ 's vertical axis along the channel dimension and feed it to the encoders. We also add latent regularizations to avoid high variance latent spaces [53]. For the 1D vector $g$ , we use a small KL-penalty via the reparameterization trick [33], and for $c$ and $f$ , we impose a vector-quantization [13,69] layer to regularize them. + +The CNN decoder is similarly a 2D CNN, and takes $c$ + +![](images/ed6be463d777247a4f659c4585dc74e0fbf29f520d27bbf5657b0030a8db5bd5.jpg) +Figure 4. Hierarchical LDM. Top: LDM $\psi_{g}$ for KL-regulairzed global latent $g$ . Middle: LDM $\psi_{c}$ for vector-quantized coarse latent $c$ . Bottom: LDM $\psi_{f}$ for vector-quantized fine latent $f$ . All LDMs optionally take an additional conditioning variable as input, such as a Bird's Eye View segmentation map as depicted here. + +concatenated along vertical axis, as the initial input. The decoder uses conditional group normalization layers [71] with $g$ as the conditioning variable. Lastly, we concatenate $f$ to an intermediate tensor in the decoder. The latent decoder outputs $\hat{V}$ which is the reconstructed voxel. LAE is trained with the voxel reconstruction loss $||V - \hat{V}||$ along with the image reconstruction loss $||i - \hat{i}||$ where $\hat{i} = r(\hat{V},\kappa)$ . Note that the image reconstruction loss only helps with the learning of LAE, and the scene auto-encoder is kept fixed. + +# 3.3. Hierarchical Latent Diffusion Models + +Given the latent variables $g, c, f$ that represent a voxel-based scene representation $V$ , we define our generative model as $p(V, g, c, f) = p(V | g, c, f) p(f | g, c) p(c | g) p(g)$ with Denoising Diffusion Models (DDMs) [21]. In general, DDMs with discrete time steps have a fixed Markovian forward process $q(x_{t} | x_{t-1})$ where $q(x_{0})$ denotes the data distribution and $q(x_{T})$ is defined to be close to the standard normal distribution, where we use the subscript to denote the time step. DDMs then learn to revert the forward process $p_{\theta}(x_{t-1} | x_{t})$ with learnable parameters $\theta$ . It can be shown that learning the reverse process is equivalent to learning to denoise $x_{t}$ to $x_{0}$ for all timesteps $t$ [21, 24] by reducing the following loss: + +$$ +\underset {t, \epsilon , x _ {0}} {\mathbb {E}} \left[ w \left(\lambda_ {t}\right) \| x _ {0} - \hat {x} _ {\theta} \left(x _ {t}, t\right) \| _ {2} ^ {2} \right] \tag {3} +$$ + +where $t$ is sampled from a uniform distribution for timesteps, $\epsilon$ is sampled from the standard normal to noise the data $x_0$ , $w(\lambda_t)$ is a timestep dependent weighting constant, and $\hat{x}_{\theta}$ denotes the learned denoising model. + +We train our hierarchical LDM with the following losses: + +$$ +\underset {t, \epsilon , g _ {0}} {\mathbb {E}} \left[ w \left(\lambda_ {t}\right) \| g _ {0} - \psi_ {g} \left(g _ {t}, t\right) \| _ {2} ^ {2} \right] \tag {4} +$$ + +$$ +\underset {t, \epsilon , g _ {0}, c _ {0}} {\mathbb {E}} \left[ w \left(\lambda_ {t}\right) \| c _ {0} - \psi_ {c} \left(c _ {t}, g _ {0}, t\right) \| _ {2} ^ {2} \right] \tag {5} +$$ + +$$ +\underset {t, \epsilon , g _ {0}, c _ {0}, f _ {0}} {\mathbb {E}} \left[ w \left(\lambda_ {t}\right) \| f _ {0} - \psi_ {f} \left(f _ {t}, g _ {0}, c _ {0}, t\right) \| _ {2} ^ {2} \right] \tag {6} +$$ + +where $\psi$ denotes the learnable denoising networks for $g, c, f$ . Fig. 4 visualizes the diffusion models. $\psi_{g}$ is implemented with linear layers with skip connections and $\psi_{c}$ and $\psi_{f}$ adopt the U-net architecture [55]. $g$ is fed into $\psi_{c}$ and $\psi_{f}$ with conditional group normalization layers. $c$ is interpolated and concatenated to the input to $\psi_{f}$ . The camera poses contain the trajectory the camera is travelling, and this information can be useful for modelling a 3D scene as it tells the model where to focus on generating. Therefore, we concatenate the camera trajectory information to $g$ and also learn to sample it. For brevity, we still call the concatenated vector $g$ . For conditional generation, each $\psi$ takes the conditioning variable as input with cross-attention layers [53]. + +Each $\psi$ can be trained in parallel and, once trained, can be sampled one after another following the hierarchy. In practice, we use the v-prediction parameterization [57] that has been shown to have better convergence and training stability [56, 57]. Once $g, c, f$ are sampled, we can use the latent decoder from Sec. 3.2 to construct the voxel $V$ which represents the neural field for the sampled scene.. Following the volume rendering and decoding step in Sec. 3.1, the sampled scene can be visualized from desired viewpoints. + +# 3.4. Post-Optimizing Generated Neural Fields + +Samples generated from our model on real-world data contain reasonable texture and geometry (Fig. 10), but can be further optimized by leveraging recent advances in 2D image diffusion models trained on orders of magnitude more data. Specifically, we iteratively update initially generated voxels, $V$ , by rendering viewpoints from the scene and applying Score Distillation Sampling (SDS) [46] loss on each image independently: + +$$ +\nabla_ {V} L _ {S D S} = \underset {\epsilon , t, \kappa} {\mathbb {E}} \left[ w \left(\lambda_ {t}\right) \left(\epsilon - \hat {\epsilon_ {\theta}} (r (V, \kappa), t)\right) \frac {\partial r (V , \kappa)}{\partial V} \right] \tag {7} +$$ + +where $\kappa$ is sampled uniformly in a $6m^2$ region around the origin of the scene with random rotation about the vertical axis, $w(\lambda_t)$ is the weighting schedule used to train $\hat{\epsilon_{\theta}}$ and $t\sim U[0.02T,0.2T]$ where $T$ is the amount of noise steps used to train $\hat{\epsilon_{\theta}}$ . Note that for latent diffusion models, the noise prediction step is applied after encoding $r(V,\kappa)$ to the LDM's latent space and the partial gradient term is updated appropriately. For $\hat{\epsilon_{\theta}}$ , we use an off-the-shelf latent diffusion model [53], finetuned to condition on CLIP image embeddings $[47]^1$ . We found that CLIP contains a representativeness matrix + +![](images/2ea9770e5dc69ebb6ad3226e32dc2ba9354bf165faa9180aaa077abab306d5ae.jpg) + +![](images/2ae539322fb9ecdeec9f45be497ef0576bd2c9760c6a10da7cdb04fab91cebe8.jpg) + +![](images/d07685b02488898282a4648785c86b7c290ce42a96a0acba94c7bfaeb60c1f2d.jpg) + +![](images/41d03d8227e08ae330ef93128c0d7a9d288883beab02da68729370592d46a572.jpg) + +![](images/a398490987c93ea73f453a89f7ccccb079ac1c51e836915a2e3ba17d9dce9694.jpg) +Figure 5. Datasets: Top-left: VizDoom [29]. Top-right: Replica [64]. Middle: Carla [12]. Bottom: AVD. For Carla and AVD, we visualize a subset of available cameras. + +![](images/b9446f3503473f385534e348384da5b56e2395971ea38b7be5cb23b178879374.jpg) + +tation of the quality of images that the LDM is able to interpret: denoising an image while conditioning on CLIP image embeddings of our model's samples produced images with similar geometry distortions and texture errors. We leverage this property by optimizing $L_{SDS}$ with negative guidance. Letting $y, y'$ be CLIP embeddings of clean image conditioning (e.g. dataset images) and artifact conditioning (e.g. samples) respectively, we perform classifier-free guidance [23] with conditioning vector $y$ , but replace the unconditional embedding with $y'$ . As shown in the supplementary, this is equivalent (up to scale) to sampling from $\frac{p(x|y)^{\alpha}}{p(x|y')}$ at each denoising step where $\alpha$ controls the trade-off between sampling towards dataset images and away from images with artifacts. We stress that this post-optimization is only successful due to the strong scene prior contained in our voxel samples, as shown by our comparison to running optimization on randomly initialized voxels in the supplementary. + +# 4. Experiments + +We evaluate NeuralField-LDM on the following four datasets (Fig. 5). Each dataset contains RGB images and a depth measurement with their corresponding camera poses. + +VizDoom [28] consists of front-view sensor observations obtained by navigating inside a synthetic game environment. We use the dataset provided by [9], which contains 34 trajectories with a length of 600. + +Replica [64] is a dataset of high-quality reconstructions of 18 indoor scenes, containing 101 front-view trajectories with a length of 100. We use the dataset provided by [9]. + +Carla [12] is an open-source simulation platform for self-driving research. We mimic the camera settings for a self-driving car, by placing six cameras (front-left, front, front-right, back-left, back, back-right), covering 360 degrees with some overlaps, and move the car in a randomly sampled direction and distance for 10 timesteps. We sample 43K datapoints, each containing 60 images. + +AVD is an in-house dataset of human driving recordings in roads and parking lots. It has ten cameras with varying + +![](images/b1c935c6e1e32d6f692e5f9c1bd56ae3253af7d2de0ffd29b6205c266d678831.jpg) +Figure 6. Generated Scenes: The top three rows are samples from Carla, and the bottom three rows are samples from AVD. + +![](images/77e6bdd80d513de7c69a7233eab5c29539258efc4d04ccc3faa0ec87d6944e9c.jpg) +Figure 7. We run marching-cubes [38] on the density voxels to visualize the geometry of the samples generated by NF-LDM. The structure of the scene is reflected well in the mesh. + +lens types along with Lidar for depth measurement. It has 73K sequences, each with 8 frames extracted at 10 fps. + +# 4.1. Baseline Comparisons + +Unconditional Generation We evaluate the unconditional generation performance of NF-LDM by comparing it with baseline models. All results are without the post + +
CriterionMethodDepthVizDoomReplica
FID (↓)GRAF [59]X47.5065.37
π-GAN [4]X143.55166.55
GSN [8]✓37.2141.75
GAUDI [2]✓33.7018.75
NF-LDM✓19.54*14.59
+ +Table 1. FID [19] scores on VizDoom and Replica. NF-LDM outperforms all baseline models. Baseline numbers are from [2]. + +
CriterionMethodDepthCarlaAVD
FID (↓)EG3D [5]×76.89194.34
GSN [8]✓75.45166.07
NF-LDM✓35.6954.26
+ +Table 2. FID [19] scores on Carla and AVD datasets. Baseline models have trouble learning the distribution of complex outdoor datasets, in particular AVD, while NF-LDM models them well. + +![](images/bc961cb0f79b63395ccc524c7e7dd1fb29c581b3531c929804c3e55a4306d39b.jpg) +Figure 8. Reconstructing held-out scenes not seen during training. + +optimization step (Sec. 3.4), unless specified. Tab. 1 shows the results on VizDoom and Replica. GRAF [59] and $\pi$ -GAN [4], which do not utilize ground truth depth in training, have shown successes in modelling single objects, but they exhibit worse performance than others that leverages depth information for modelling scenes. GAUDI [2] is an auto-decoder-based diffusion model. Their auto-decoder optimizes a small per-scene latent to reconstruct its matching scene. GAUDI comes with the advantage that learning the generative model is simple as it only needs to model the small dimensional latent distribution that acts as the key to their corresponding scenes. On the contrary, NF-LDM is trained on the latents that are a decomposition of the explicit 3D neural field. Therefore, GAUDI puts more modelling + +
CriterionMethodDepthCarlaAVD
FVD (↓)EG3D [5]X134.941232.38
GSN [8]✓184.301659.81
NF-LDM✓91.80242.50
+ +capacity into the auto-decoder part, and NF-LDM puts it more into the generative model part. We attribute our improvement over GAUDI to our expressive hierarchical LDM that can model the details of the scenes better. In VizDoom, only one scene exists, and each sequence contains several hundred steps covering a large area in the scene, which our voxels were not large enough to encode. Therefore, we chunked each VizDoom trajectory to be 50 steps long. + +Tab. 2 shows results on complex outdoor datasets: Carla and AVD. We compare with EG3D [5] and GSN [8]. Both are GAN-based 3D generative models, but GSN utilizes ground truth depth measurements. Note that we did not include GAUDI [2] as the code was not available. NF-LDM achieves the best performance, and both baseline models have difficulty modelling the real outdoor dataset (AVD). Fig. 6 compares the samples from different models. + +Since the datasets are composed of frame sequences, we can treat them as videos and further evaluate with Fréchet Video Distance (FVD) [66] to compare the distributions of the dataset and sampled sequences. This can quantify samples' 3D structure by how natural the rendered sequence from a moving camera is. For EG3D and GSN, we randomly sample a trajectory from the datasets and for NF-LDM, we sample a trajectory from the global latent diffusion model. Tab. 3 shows that NF-LDM achieves the best results. We empirically observed GSN sometimes produced slightly inconsistent rendering, which could attribute to its lower FVD score than EG3D's. We also visualize the geometry of NF-LDM's samples by running marching-cubes [38] on the density voxels. Fig. 7 shows that our samples produce a coarse but realistic geometry. + +Ablations We evaluate the hierarchical structure of NF-LDM. Tab. 4 shows an ablation study on Carla. The model with the full hierarchy achieves the best performance. The global latent makes it easier for the other LDMs to sample as conditioning on the global properties of the scene (e.g. time of day) narrows down the distribution they need to model. The 2D fine latent helps retain the residual information missing in the 3D coarse latent, thus improving the latent auto-encoder and, consequently, the LDMs. + +Scene Reconstruction Unlike previous approaches, NF-LDM has an explicit scene auto-encoder that can be used for scene reconstruction. GAUDI [2] is auto-decoder based, so it is not trivial to infer a latent for a new scene. GSN [8] can invert a new scene using a GAN inversion method [52, + +![](images/f3c51d533f81ba0ba02a18e1db53a31f2109d36dce5e90c3ae8dd46b9c47178b.jpg) + +![](images/372c91bb060305a99c8da1b2ebfd42c8d16a6925b4c6cc3fa0e4c07c0768bf4e.jpg) + +![](images/e460ab0240c0a8197bc6f8243058061d111636897779ccd35fde311b785be1c6.jpg) +Figure 9. BEV-Conditioned Synthesis: NF-LDM allows controllable generation by editing the BEV segmentation map. From the initial sample, we add trees (green) and then edit the location of the car (blue). Note the ego car is at the center and thus not rendered. + +Table 3. FVD [66] scores on Carla and AVD Datasets. As for FID, baseline models have trouble learning to model complex datasets. + +
Coarse lat. c+ Fine lat. f+ Global lat. g
FID (↓)46.4343.5235.69
+ +Table 4. FID [19] on ablating the choice of hierarchy on the Carla dataset. The first column is for training both LAE and LDM only with the coarse latent. The last column is our full model. + +79], but as Fig. 8 shows, it fails to get the details of the scene correct. Our scene auto-encoder generalizes well and is scalable as the number of scenes grow. + +# 4.2. Applications and Limitations + +Conditional Synthesis NF-LDM can utilize additional conditioning signals for controllable generation. In this paper, we consider Bird's Eye View (BEV) segmentation maps, but our model can be extended to other conditioning variables. We use cross attention layers [53], which have been shown to be effective for conditional synthesis. Fig. 9 shows that NF-LDM follows the given BEV map faithfully and how the map can be edited for controllable synthesis. + +Scene Editing Image diffusion models can be used for image inpainting without explicit training on the task [53, 63]. We leverage this property to edit scenes in 3D by resampling a region in the 3D coarse latent $c$ . Specifically, at each denoising step, we noise the region to be kept and concatenate with the region being sampled, and pass it through the diffusion model. We use reconstruction guidance [24] to better harmonize the sampled and kept regions. After we get a new $c$ , the fine latent is also re-sampled conditioned on $c$ . Fig. 11 shows results on scene editing with NF-LDM. + +![](images/f6f3e6e358c5a07224b7f0a5b202839d588539fd0048e6c5e6c9587662c45bf8.jpg) +Figure 10. Panoramas from NF-LDM's samples: From the initial sample at the top, we apply post-optimization with Score Distillation Sampling [46] (Sec. 3.4). (a) demonstrates improved sample quality. (b) showcases style modification by conditioning on evening scenes. + +![](images/97bdf0f0cb0156bf24f2cc3015518c5a49502b30d7eb93a34bd500039ac93a1d.jpg) +Figure 11. Scene Editing: We use the 3D coarse latent $c$ for scene editing. From the initial sample indicated by light green, we resample a part of the latent, indicated by dark green. + +Post-Optimization Fig. 10 shows how post-optimization (Sec. 3.4) can improve the quality of NF-LDM's initial sample while retaining the 3D structure. In addition to improving quality, we can also modify scene properties, such as time of day and weather, by conditioning the LDM on images with the desired properties. SDS-based style modification is effective for changes where a set of + +clean image data is available with the desired property and is reasonably close to our dataset's domain (e.g. street images for AVD). In the supplementary, we also provide results experimenting with directional CLIP loss [14] to quickly finetune our scene decoder for a given text prompt. + +Limitations NF-LDM's hierarchical structure and three stage pipeline allows us to achieve high-quality generations and reconstructions, but it comes with a degradation in training time and sampling speed. In this work, the neural field representation is based on dense voxel grids, and it becomes expensive to volume render and learn the diffusion models as they get larger. Therefore, exploring alternative sparse representations is a promising future direction. Lastly, our method requires multi-view images which limits data availability and therefore risks universal problems in generative modelling of overfitting. For example, we found that output samples in AVD had limited diversity because the dataset itself was recorded in a limited number of scenes. + +# 5. Conclusion + +We introduced NeuralField-LDM (NF-LDM), a generative model for complex 3D environments. NF-LDM first constructs an expressive latent distribution by encoding input images into a 3D neural field representation which is further compressed into more abstract latent spaces. Then, our proposed hierarchical LDM is fit onto the latent spaces, achieving state-of-the-art performance on 3D scene generation. NF-LDM enables a diverse set of applications, including controllable scene generation and scene editing. Future directions include exploring more efficient sparse voxel representations, training on larger-scale real-world data and learning to continuously expand generated scenes. + +# References + +[1] Martin Arjovsky and Léon Bottou. Towards principled methods for training generative adversarial networks. In ICLR, 2017. 3 +[2] Miguel Ángel Bautista, Pengsheng Guo, Samira Abnar, Walter Talbott, Alexander Toshev, Zhuoyuan Chen, Laurent Dinh, Shuangfei Zhai, Hanlin Goh, Daniel Ulbricht, Afshin Dehghan, and Josh M. Susskind. GAUDI: A neural architect for immersive 3d scene generation. 2022. 1, 2, 6, 7 +[3] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale GAN training for high fidelity natural image synthesis. In ICLR, 2019. 2 +[4] Eric Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. arXiv, 2012.00926, 2020. 6 +[5] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In arXiv, 2021. 6, 7 +[6] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In CVPR, 2022. 1, 2 +[7] Eric R. Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. Pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In CVPR, 2021. 2 +[8] Terrance DeVries, Miguel Angel Bautista, Nitish Srivastava, Graham W. Taylor, and Joshua M. Susskind. Unconstrained scene generation with locally conditioned radiance fields. In ICCV, 2021. 1, 2, 6, 7 +[9] Terrance DeVries, Miguel Angel Bautista, Nitish Srivastava, Graham W. Taylor, and Joshua M. Susskind. Unconstrained scene generation with locally conditioned radiance fields. arXiv, 2021. 5 +[10] Prafulla Dhariwal and Alexander Quinn Nichol. Diffusion models beat gans on image synthesis. In Marc'Aurelio Ranzato, Alina Beygelzimer, Yann N. Dauphin, Percy Liang, and Jennifer Wortman Vaughan, editors, NeurIPS, 2021. 2 +[11] Tim Dockhorn, Arash Vahdat, and Karsten Kreis. Score-based generative modeling with critically-damped Langevin diffusion. In ICLR, 2022. 2 +[12] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. CARLA: An open urban driving simulator. 2017. 5 +[13] Patrick Esser, Robin Rombach, and Björn Ommer. Taming transformers for high-resolution image synthesis. In CVPR, 2021. 2, 4 +[14] Rinon Gal, Or Patashnik, Haggai Maron, Gal Chechik, and Daniel Cohen-Or. Stylegan-nada: Clip-guided domain adaptation of image generators, 2021. 8 +[15] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja + +Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. In Advances In Neural Information Processing Systems, 2022. 1 +[16] Ian Goodfellow. Nips 2016 tutorial: Generative adversarial networks. arXiv preprint arXiv:1701.00160, 2016. 3 +[17] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron C. Courville, and Yoshua Bengio. Generative adversarial nets. In NeurIPS, 2014. 2 +[18] William Harvey, Saeid Naderiparizi, Vaden Masrani, Christian Weilbach, and Frank Wood. Flexible diffusion modeling of long videos, 2022. 2 +[19] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. 6, 7 +[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Hugo Larochelle, Marc' Aurelio Ranzato, Raia Hadsell, Maria-Florina Balcan, and Hsuan-Tien Lin, editors, NeurIPS, 2020. 2, 3 +[21] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 6840-6851. Curran Associates, Inc., 2020. 4 +[22] Jonathan Ho, Chitwan Sahara, William Chan, David J. Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. J. Mach. Learn. Res., 2022. 2 +[23] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 5 +[24] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. arXiv:2204.03458, 2022. 2, 4, 7 +[25] James T Kajiya and Brian P Von Herzen. Ray tracing volume densities. ACM SIGGRAPH computer graphics, 18(3):165-174, 1984. 2 +[26] Amlan Kar, Aayush Prakash, Ming-Yu Liu, Eric Cameracci, Justin Yuan, Matt Rusiniak, David Acuna, Antonio Torralba, and Sanja Fidler. Meta-sim: Learning to generate synthetic datasets. In ICCV, 2019. 2 +[27] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. In CVPR, 2020. 2 +[28] Michal Kempka, Marek Wydmuch, Grzegorz Runc, Jakub Toczek, and Wojciech Jaskowski. Vizdoom: A doom-based ai research platform for visual reinforcement learning. In 2016 IEEE conference on computational intelligence and games (CIG), pages 1-8. IEEE, 2016. 5 +[29] Margret Keuper, Siyu Tang, Bjoern Andres, Thomas Brox, and Bernt Schiele. Motion segmentation & multiple object tracking by correlation co-clustering. IEEE TPAMI, 2018. 5 +[30] Seung Wook Kim, Karsten Kreis, Daiqing Li, Antonio Torralba, and Sanja Fidler. Polymorphic-gan: Generating aligned samples across multiple domains with learned morph maps. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Jun. 2022. 2 + +[31] Seung Wook Kim, Jonah Philion, Antonio Torralba, and Sanja Fidler. Drivegan: Towards a controllable high-quality neural simulation. In CVPR, 2021. 2 +[32] Seung Wook Kim, Yuhao Zhou, Jonah Philion, Antonio Torralba, and Sanja Fidler. Learning to simulate dynamic environments with gamean. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1231-1240, 2020. 2 +[33] Diederik P. Kingma and Max Welling. Auto-encoding variational bayes. *ICLR*, 2014. 2, 4 +[34] Daiqing Li, Huan Ling, Seung Wook Kim, Karsten Kreis, Adela Barriuso, Sanja Fidler, and Antonio Torralba. Big-datasetgan: Synthesizing imagenet with pixel-wise annotations. arXiv, 2201.04684, 2022. 2 +[35] Ke Li and Jitendra Malik. On the implicit assumptions of gans. arXiv preprint arXiv:1811.12402, 2018. 3 +[36] Yiyi Liao, Katja Schwarz, Lars Mescheder, and Andreas Geiger. Towards unsupervised learning of generative models for 3d controllable image synthesis. In CVPR, 2020. 2 +[37] Huan Ling, Karsten Kreis, Daiqing Li, Seung Wook Kim, Antonio Torralba, and Sanja Fidler. Editgan: High-precision semantic image editing. In NeurIPS, 2021. 2 +[38] William E. Lorensen and Harvey E. Cline. Marching cubes: A high resolution 3d surface construction algorithm. In ACM Trans. on Graphics, 1987. 6, 7 +[39] Shitong Luo and Wei Hu. Diffusion probabilistic models for 3d point cloud generation. In CVPR, 2021. 2 +[40] Ricardo Martin-Brualla, Noha Radwan, Mehdi S. M. Sajjadi, Jonathan T. Barron, Alexey Dosovitskiy, and Daniel Duckworth. NeRF in the Wild: Neural Radiance Fields for Unconstrained Photo Collections. In CVPR, 2021. 2 +[41] Lars Mescheder, Andreas Geiger, and Sebastian Nowozin. Which training methods for gans do actually converge? In Proc. of the International Conf. on Machine learning (ICML), 2018. 2 +[42] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 2, 3 +[43] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. GLIDE: towards photorealistic image generation and editing with text-guided diffusion models. In Kamalika Chaudhuri, Stefanie Jegelka, Le Song, Csaba Szepesvári, Gang Niu, and Sivan Sabato, editors, Proc. of the International Conf. on Machine learning (ICML), 2022. 2 +[44] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In CVPR, 2021. 2 +[45] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In ECCV, 2020. 3 +[46] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv, 2022. 1, 2, 5, 8 + +[47] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. arXiv, 2103.00020, 2021. 5 +[48] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with CLIP latents. arXiv, abs/2204.06125, 2022. 2 +[49] Ali Razavi, Aaron van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with VQ-VAE-2. In NeurIPS, 2019. 2 +[50] Konstantinos Rematas, Andrew Liu, Pratul P. Srinivasan, Jonathan T. Barron, Andrea Tagliasacchi, Thomas A. Funkhouser, and Vittorio Ferrari. Urban radiance fields. In CVPR, 2022. 2, 3 +[51] Danilo Jimenez Rezende, Shakir Mohamed, and Daan Wierstra. Stochastic backpropagation and approximate inference in deep generative models. In Proc. of the International Conf. on Machine learning (ICML), 2014. 2 +[52] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a stylegan encoder for image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2287–2296, 2021. 7 +[53] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 3, 4, 5, 7 +[54] Robin Rombach, Patrick Esser, and Björn Ommer. Geometry-free view synthesis: Transformers and no 3d priors. In ICCV, 2021. 2 +[55] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234-241. Springer, 2015. 5 +[56] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 3, 4, 5 +[57] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. arXiv preprint arXiv:2202.00512, 2022. 3, 5 +[58] Axel Sauer, Katja Schwarz, and Andreas Geiger. StyleganxI: Scaling stylegan to large diverse datasets. ACM Trans. on Graphics, 2022. 2 +[59] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. Graf: Generative radiance fields for 3d-aware image synthesis. 2020. 2, 6 +[60] Katja Schwarz, Axel Sauer, Michael Niemeyer, Yiyi Liao, and Andreas Geiger. Voxgraf: Fast 3d-aware image synthesis with sparse voxel grids. In NeurIPS, 2022. 2 +[61] Yujun Shen, Jinjin Gu, Xiaou Tang, and Bolei Zhou. Interpreting the latent space of gans for semantic face editing. In CVPR, 2020. 2 + +[62] Yuan Shen, Wei-Chiu Ma, and Shenlong Wang. SGAM: Building a virtual 3d world through simultaneous generation and mapping. In Thirty-Sixth Conference on Neural Information Processing Systems, 2022. 2 +[63] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 2, 7 +[64] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J. Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, Anton Clarkson, Mingfei Yan, Brian Budge, Yajie Yan, Xiaqing Pan, June Yon, Yuyang Zou, Kimberly Leon, Nigel Carter, Jesus Briales, Tyler Gillingham, Elias Mueggler, Luis Pesqueira, Manolis Savva, Dhruv Batra, Hauke M. Strasdat, Renzo De Nardi, Michael Goesele, Steven Lovegrove, and Richard Newcombe. The Replica dataset: A digital replica of indoor spaces. arXiv, 1906.05797, 2019. 5 +[65] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. CVPR, 2022. 3 +[66] Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 7 +[67] Arash Vahdat and Jan Kautz. NVAE: A deep hierarchical variational autoencoder. In Hugo Larochelle, Marc'Aurelio Ranzato, Raia Hadsell, Maria-Florina Balcan, and Hsuan-Tien Lin, editors, NeurIPS, 2020. 2 +[68] Arash Vahdat, Karsten Kreis, and Jan Kautz. Score-based generative modeling in latent space. In NeurIPS, 2021. 2 +[69] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 4 +[70] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P. Srinivasan, Howard Zhou, Jonathan T. Barron, Ricardo Martin-Brualla, Noah Snavely, and Thomas A. Funkhouser. Ibrnet: Learning multi-view image-based rendering. In CVPR, 2021. 2 +[71] Yuxin Wu and Kaiming He. Group normalization. In Proceedings of the European conference on computer vision (ECCV), pages 3-19, 2018. 4 +[72] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelNeRF: Neural radiance fields from one or few images. In CVPR, 2021. 2 +[73] Xiaohui Zeng, Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, and Karsten Kreis. LION: latent point diffusion models for 3d shape generation. In NeurIPS, 2022. 1, 2, 3 +[74] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv, 2010.07492, 2020. 2 +[75] Xiaoshuai Zhang, Sai Bi, Kalyan Sunkavalli, Hao Su, and Zexiang Xu. Nerfusion: Fusing radiance fields for large-scale scene reconstruction. In CVPR, 2022. 2 +[76] Yuxuan Zhang, Wenzheng Chen, Huan Ling, Jun Gao, Yi-nan Zhang, Antonio Torralba, and Sanja Fidler. Image gans + +meet differentiable rendering for inverse graphics and interpretable 3d neural rendering. In ICLR, 2021. 2 +[77] Yan Zheng, Lemeng Wu, Xingchao Liu, Zhen Chen, Qiang Liu, and Qixing Huang. Neural volumetric mesh generator. arXiv preprint arXiv:2210.03158, 2022. 3 +[78] Linqi Zhou, Yilun Du, and Jiajun Wu. 3d shape generation and completion through point-voxel diffusion. In ICCV, 2021. 2 +[79] Jiapeng Zhu, Yujun Shen, Deli Zhao, and Bolei Zhou. Indomain gan inversion for real image editing. In European conference on computer vision, pages 592-608. Springer, 2020. 7 \ No newline at end of file diff --git a/2023/NeuralField-LDM_ Scene Generation With Hierarchical Latent Diffusion Models/images.zip b/2023/NeuralField-LDM_ Scene Generation With Hierarchical Latent Diffusion Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a795357c3f02e6c954bdfa68d0622afe7f0dba5d --- /dev/null +++ b/2023/NeuralField-LDM_ Scene Generation With Hierarchical Latent Diffusion Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c7c0f1b9dda45eb2a59c8c84ac66060451b9369a791af1b653a23910f472825e +size 913874 diff --git a/2023/NeuralField-LDM_ Scene Generation With Hierarchical Latent Diffusion Models/layout.json b/2023/NeuralField-LDM_ Scene Generation With Hierarchical Latent Diffusion Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..64533897e0992799209cb2f8ad32782753e92adb --- /dev/null +++ b/2023/NeuralField-LDM_ Scene Generation With Hierarchical Latent Diffusion Models/layout.json @@ -0,0 +1,10685 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 49, + 86, + 544, + 102 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 86, + 544, + 102 + ], + "spans": [ + { + "bbox": [ + 49, + 86, + 544, + 102 + ], + "type": "text", + "content": "NeuralField-LDM: Scene Generation with Hierarchical Latent Diffusion Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 125, + 541, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 541, + 140 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 541, + 140 + ], + "type": "text", + "content": "Seung Wook Kim" + }, + { + "bbox": [ + 53, + 125, + 541, + 140 + ], + "type": "inline_equation", + "content": "^{1,2,3*}" + }, + { + "bbox": [ + 53, + 125, + 541, + 140 + ], + "type": "text", + "content": " Bradley Brown" + }, + { + "bbox": [ + 53, + 125, + 541, + 140 + ], + "type": "inline_equation", + "content": "^{1,5*†}" + }, + { + "bbox": [ + 53, + 125, + 541, + 140 + ], + "type": "text", + "content": " Kangxue Yin" + }, + { + "bbox": [ + 53, + 125, + 541, + 140 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 125, + 541, + 140 + ], + "type": "text", + "content": " Karsten Kreis" + }, + { + "bbox": [ + 53, + 125, + 541, + 140 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 125, + 541, + 140 + ], + "type": "text", + "content": " Katja Schwarz" + }, + { + "bbox": [ + 53, + 125, + 541, + 140 + ], + "type": "inline_equation", + "content": "^{6†}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 116, + 140, + 175, + 155 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 140, + 175, + 155 + ], + "spans": [ + { + "bbox": [ + 116, + 140, + 175, + 155 + ], + "type": "text", + "content": "Daiqing Li" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 200, + 140, + 287, + 155 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 140, + 287, + 155 + ], + "spans": [ + { + "bbox": [ + 200, + 140, + 287, + 155 + ], + "type": "text", + "content": "Robin Rombach" + }, + { + "bbox": [ + 200, + 140, + 287, + 155 + ], + "type": "inline_equation", + "content": "^{7\\dagger}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 140, + 402, + 155 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 140, + 402, + 155 + ], + "spans": [ + { + "bbox": [ + 313, + 140, + 402, + 155 + ], + "type": "text", + "content": "Antonio Torralba" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 427, + 140, + 503, + 155 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 427, + 140, + 503, + 155 + ], + "spans": [ + { + "bbox": [ + 427, + 140, + 503, + 155 + ], + "type": "text", + "content": "Sanja Fidler1,2,3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 121, + 156, + 477, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 156, + 477, + 167 + ], + "spans": [ + { + "bbox": [ + 121, + 156, + 477, + 167 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 121, + 156, + 477, + 167 + ], + "type": "text", + "content": "NVIDIA " + }, + { + "bbox": [ + 121, + 156, + 477, + 167 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 121, + 156, + 477, + 167 + ], + "type": "text", + "content": "University of Toronto " + }, + { + "bbox": [ + 121, + 156, + 477, + 167 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 121, + 156, + 477, + 167 + ], + "type": "text", + "content": "Vector Institute " + }, + { + "bbox": [ + 121, + 156, + 477, + 167 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 121, + 156, + 477, + 167 + ], + "type": "text", + "content": "CSAIL, MIT " + }, + { + "bbox": [ + 121, + 156, + 477, + 167 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 121, + 156, + 477, + 167 + ], + "type": "text", + "content": "University of Waterloo" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 184, + 169, + 414, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 169, + 414, + 182 + ], + "spans": [ + { + "bbox": [ + 184, + 169, + 414, + 182 + ], + "type": "text", + "content": "6University of Tübingen, Tübingen AI Center 7LMU Munich" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 49, + 186, + 545, + 274 + ], + "blocks": [ + { + "bbox": [ + 49, + 186, + 545, + 274 + ], + "lines": [ + { + "bbox": [ + 49, + 186, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 49, + 186, + 545, + 274 + ], + "type": "image", + "image_path": "fa25ef0555bea1ba507b5b1f853b4ad8e1acf043ea8867875ede42fa25075a26.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 49, + 274, + 140, + 334 + ], + "blocks": [ + { + "bbox": [ + 49, + 274, + 140, + 334 + ], + "lines": [ + { + "bbox": [ + 49, + 274, + 140, + 334 + ], + "spans": [ + { + "bbox": [ + 49, + 274, + 140, + 334 + ], + "type": "image", + "image_path": "a7c0c2a5c50443499d118db2c92fd0798cee8f89ece0862a8e9153c7b05d74a0.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 335, + 546, + 357 + ], + "lines": [ + { + "bbox": [ + 47, + 335, + 546, + 357 + ], + "spans": [ + { + "bbox": [ + 47, + 335, + 546, + 357 + ], + "type": "text", + "content": "Figure 1. We introduce NeuralField-LDM, a generative model for complex open-world 3D scenes. This figure contains a panorama constructed from NeuralField-LDM's generated scene. We visualize different parts of the scene by placing cameras on them." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 149, + 274, + 242, + 334 + ], + "blocks": [ + { + "bbox": [ + 149, + 274, + 242, + 334 + ], + "lines": [ + { + "bbox": [ + 149, + 274, + 242, + 334 + ], + "spans": [ + { + "bbox": [ + 149, + 274, + 242, + 334 + ], + "type": "image", + "image_path": "22cf9cbbf8ae11a68b0fd11712a7c9faabafac4044bd2b69985eb6d8ec6d2165.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 251, + 274, + 343, + 334 + ], + "blocks": [ + { + "bbox": [ + 251, + 274, + 343, + 334 + ], + "lines": [ + { + "bbox": [ + 251, + 274, + 343, + 334 + ], + "spans": [ + { + "bbox": [ + 251, + 274, + 343, + 334 + ], + "type": "image", + "image_path": "dbb85e17ddc79672b02f433511e29dfa4d317a403bd2192d0ce4b6c948398a3e.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 352, + 274, + 444, + 334 + ], + "blocks": [ + { + "bbox": [ + 352, + 274, + 444, + 334 + ], + "lines": [ + { + "bbox": [ + 352, + 274, + 444, + 334 + ], + "spans": [ + { + "bbox": [ + 352, + 274, + 444, + 334 + ], + "type": "image", + "image_path": "f816f399f8022c2169295cabfa1b364ea4ffb738f2b78f8bcf11db12fb5001df.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 453, + 273, + 545, + 334 + ], + "blocks": [ + { + "bbox": [ + 453, + 273, + 545, + 334 + ], + "lines": [ + { + "bbox": [ + 453, + 273, + 545, + 334 + ], + "spans": [ + { + "bbox": [ + 453, + 273, + 545, + 334 + ], + "type": "image", + "image_path": "c4102ba302da853b621c88a79df028cebaabdc3234d53438413f2ac55e858c6b.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 143, + 367, + 192, + 379 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 367, + 192, + 379 + ], + "spans": [ + { + "bbox": [ + 143, + 367, + 192, + 379 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 46, + 392, + 290, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 392, + 290, + 620 + ], + "spans": [ + { + "bbox": [ + 46, + 392, + 290, + 620 + ], + "type": "text", + "content": "Automatically generating high-quality real world 3D scenes is of enormous interest for applications such as virtual reality and robotics simulation. Towards this goal, we introduce NeuralField-LDM, a generative model capable of synthesizing complex 3D environments. We leverage Latent Diffusion Models that have been successfully utilized for efficient high-quality 2D content creation. We first train a scene auto-encoder to express a set of image and pose pairs as a neural field, represented as density and feature voxel grids that can be projected to produce novel views of the scene. To further compress this representation, we train a latent-autoencoder that maps the voxel grids to a set of latent representations. A hierarchical diffusion model is then fit to the latents to complete the scene generation pipeline. We achieve a substantial improvement over existing state-of-the-art scene generation models. Additionally, we show how NeuralField-LDM can be used for a variety of 3D content creation applications, including conditional scene generation, scene inpainting and scene style manipulation." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 47, + 642, + 128, + 654 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 642, + 128, + 654 + ], + "spans": [ + { + "bbox": [ + 47, + 642, + 128, + 654 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 47, + 662, + 287, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 662, + 287, + 686 + ], + "spans": [ + { + "bbox": [ + 47, + 662, + 287, + 686 + ], + "type": "text", + "content": "There has been increasing interest in modelling 3D real-world scenes for use in virtual reality, game design, digi-" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 368, + 546, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 368, + 546, + 571 + ], + "spans": [ + { + "bbox": [ + 306, + 368, + 546, + 571 + ], + "type": "text", + "content": "tal twin creation and more. However, designing 3D worlds by hand is a challenging and time-consuming process, requiring 3D modeling expertise and artistic talent. Recently, we have seen success in automating 3D content creation via 3D generative models that output individual object assets [15,46,73]. Although a great step forward, automating the generation of real-world scenes remains an important open problem and would unlock many applications ranging from scalably generating a diverse array of environments for training AI agents (e.g. autonomous vehicles) to the design of realistic open-world video games. In this work, we take a step towards this goal with NeuralField-LDM (NF-LDM), a generative model capable of synthesizing complex real-world 3D scenes. NF-LDM is trained on a collection of posed camera images and depth measurements which are easier to obtain than explicit ground-truth 3D data, offering a scalable way to synthesize 3D scenes." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 576, + 547, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 576, + 547, + 685 + ], + "spans": [ + { + "bbox": [ + 304, + 576, + 547, + 685 + ], + "type": "text", + "content": "Recent approaches [2, 6, 8] tackle the same problem of generating 3D scenes, albeit on less complex data. In [6, 8], a latent distribution is mapped to a set of scenes using adversarial training, and in GAUDI [2], a denoising diffusion model is fit to a set of scene latents learned using an auto-decoder. These models all have an inherent weakness of attempting to capture the entire scene into a single vector that conditions a neural radiance field. In practice, we find that this limits the ability to fit complex scene distributions." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 689, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 546, + 714 + ], + "type": "text", + "content": "Recently, diffusion models have emerged as a very powerful class of generative models, capable of generating high-" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 693, + 130, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 693, + 130, + 703 + ], + "spans": [ + { + "bbox": [ + 61, + 693, + 130, + 703 + ], + "type": "text", + "content": "* Equal contribution." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 63, + 703, + 209, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 703, + 209, + 713 + ], + "spans": [ + { + "bbox": [ + 63, + 703, + 209, + 713 + ], + "type": "text", + "content": "†Work done during an internship at NVIDIA." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8496" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 131 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 131 + ], + "type": "text", + "content": "quality images, point clouds and videos [18, 24, 39, 48, 53, 73, 78]. Yet, due to the nature of our task, where image data must be mapped to a shared 3D scene without an explicit ground truth 3D representation, straightforward approaches fitting a diffusion model directly to data are infeasible." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 133, + 286, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 133, + 286, + 357 + ], + "spans": [ + { + "bbox": [ + 46, + 133, + 286, + 357 + ], + "type": "text", + "content": "In NeuralField-LDM, we learn to model scenes using a three-stage pipeline. First, we learn an auto-encoder that encodes scenes into a neural field, represented as density and feature voxel grids. Inspired by the success of latent diffusion models for images [53], we learn to model the distribution of our scene voxels in latent space to focus the generative capacity on core parts of the scene and not the extraneous details captured by our voxel auto-encoders. Specifically, a latent-autoencoder decomposes the scene voxels into a 3D coarse, 2D fine and 1D global latent. Hierarchical diffusion models are then trained on the tri-latent representation to generate novel 3D scenes. We show how NF-LDM enables applications such as scene editing, birdeye view conditional generation and style adaptation. Finally, we demonstrate how score distillation [46] can be used to optimize the quality of generated neural fields, allowing us to leverage the representations learned from state-of-the-art image diffusion models that have been exposed to orders of magnitude more data." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 360, + 286, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 360, + 286, + 431 + ], + "spans": [ + { + "bbox": [ + 46, + 360, + 286, + 431 + ], + "type": "text", + "content": "Our contributions are: 1) We introduce NF-LDM, a hierarchical diffusion model capable of generating complex open-world 3D scenes and achieving state of the art scene generation results on four challenging datasets. 2) We extend NF-LDM to semantic birds-eye view conditional scene generation, style modification and 3D scene editing." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 437, + 133, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 437, + 133, + 449 + ], + "spans": [ + { + "bbox": [ + 47, + 437, + 133, + 449 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 457, + 286, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 457, + 286, + 647 + ], + "spans": [ + { + "bbox": [ + 46, + 457, + 286, + 647 + ], + "type": "text", + "content": "2D Generative Models In past years, generative adversarial networks (GANs) [3, 17, 27, 41, 58] and likelihood-based approaches [33, 49, 51, 67] enabled high-resolution photorealistic image synthesis. Due to their quality, GANs are used in a multitude of downstream applications ranging from steerable content creation [30, 34, 36, 37, 61, 76] to data driven simulation [26, 31, 32, 34]. Recently, autoregressive models and score-based models, e.g. diffusion models, demonstrate better distribution coverage while preserving high sample quality [10, 11, 13, 20, 22, 43, 48, 53, 54, 68]. Since evaluation and optimization of these approaches in pixel space is computationally expensive, [53, 68] apply them to latent space, achieving state-of-the-art image synthesis at megapixel resolution. As our approach operates on 3D scenes, computational efficiency is crucial. Hence, we build upon [53] and train our model in latent space." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 653, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 286, + 712 + ], + "type": "text", + "content": "Novel View Synthesis In their seminal work [42], Mildenhall et al. introduce Neural Radiance Fields (NeRF) as a powerful 3D representation. PixelNeRF [72] and IBR-Net [70] propose to condition NeRF on aggregated features from multiple views to enable novel view synthesis from" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": "a sparse set of views. Another line of works scale NeRF to large-scale indoor and outdoor scenes [40, 50, 74, 75]. Recently, Nerfusion [75] predicts local radiance fields and fuses them into a scene representation using a recurrent neural network. Similarly, we construct a latent scene representation by aggregating features across multiple views. Different from the aforementioned methods, our approach is a generative model capable of synthesizing novel scenes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 173, + 545, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 173, + 545, + 269 + ], + "spans": [ + { + "bbox": [ + 304, + 173, + 545, + 269 + ], + "type": "text", + "content": "3D Diffusion Models A few recent works propose to apply denoising diffusion models (DDM) [20, 22, 63] on point clouds for 3D shape generation [39, 73, 78]. While PVD [78] trains on point clouds directly, DPM [39] and LION [73] use a shape latent variable. Similar to LION, we design a hierarchical model by training separate conditional DDMs. However, our approach generates both texture and geometry of a scene without needing 3D ground truth as supervision." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 274, + 545, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 274, + 545, + 489 + ], + "spans": [ + { + "bbox": [ + 304, + 274, + 545, + 489 + ], + "type": "text", + "content": "3D-Aware Generative Models 3D-aware generative models synthesize images while providing explicit control over the camera pose and potentially other scene properties, like object shape and appearance. SGAM [62] generates a 3D scene by autoregressively generating sensor data and building a 3D map. Several previous approaches generate NeRFs of single objects with conditional coordinate-based MLPs [7, 44, 59]. GSN [8] conditions a coordinate-based MLP on a \"floor plan\", i.e., a 2D feature map, to model more complex indoor scenes. EG3D [6] and VoxGRAF [60] use convolutional backbones to generate 3D representations. All of these approaches rely on adversarial training. Instead, we train a DDM on voxels in latent space. The work closest to ours is GAUDI [2], which first trains an auto-decoder and subsequently trains a DDM on the learned latent codes. Instead of using a global latent code, we encode scenes onto voxel grids and train a hierarchical DDM to optimally combine global and local features." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 501, + 414, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 501, + 414, + 513 + ], + "spans": [ + { + "bbox": [ + 306, + 501, + 414, + 513 + ], + "type": "text", + "content": "3. NeuralField-LDM" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 521, + 545, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 629 + ], + "type": "text", + "content": "Our objective is to train a generative model to synthesize 3D scenes that can be rendered to any viewpoint. We assume access to a dataset " + }, + { + "bbox": [ + 304, + 521, + 545, + 629 + ], + "type": "inline_equation", + "content": "\\{(i,\\kappa ,\\rho)\\}_{1..N}" + }, + { + "bbox": [ + 304, + 521, + 545, + 629 + ], + "type": "text", + "content": " which consists of " + }, + { + "bbox": [ + 304, + 521, + 545, + 629 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 521, + 545, + 629 + ], + "type": "text", + "content": " RGB images " + }, + { + "bbox": [ + 304, + 521, + 545, + 629 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 521, + 545, + 629 + ], + "type": "text", + "content": " and their camera poses " + }, + { + "bbox": [ + 304, + 521, + 545, + 629 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 304, + 521, + 545, + 629 + ], + "type": "text", + "content": ", along with a depth measurement " + }, + { + "bbox": [ + 304, + 521, + 545, + 629 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 304, + 521, + 545, + 629 + ], + "type": "text", + "content": " that can be either sparse (e.g. Lidar points) or dense. The generative model must learn to model both the texture and geometry distributions of the dataset in 3D by learning solely from the sensor observations, which is a highly non-trivial problem." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "Past work typically tackles this problem with a generative adversarial network (GAN) framework [6, 8, 59, 60]. They produce an intermediate 3D representation and render images for a given viewpoint with volume rendering [25, 42]. Discriminator losses then ensure that the 3D representation produces a valid image from any viewpoint. However, GANs come with notorious training instability" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8497" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 75, + 67, + 516, + 178 + ], + "blocks": [ + { + "bbox": [ + 75, + 67, + 516, + 178 + ], + "lines": [ + { + "bbox": [ + 75, + 67, + 516, + 178 + ], + "spans": [ + { + "bbox": [ + 75, + 67, + 516, + 178 + ], + "type": "image", + "image_path": "b1a3e5073629a6e122828874c18ea1f3ac41f1f030a4b040e2bbff35986fe11f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 178, + 547, + 212 + ], + "lines": [ + { + "bbox": [ + 46, + 178, + 547, + 212 + ], + "spans": [ + { + "bbox": [ + 46, + 178, + 547, + 212 + ], + "type": "text", + "content": "Figure 2. Overview of NeuralField-LDM. We first encode RGB images with camera poses into a neural field represented by density and feature voxel grids. We compress the neural field into smaller latent spaces and fit a hierarchical latent diffusion model on the latent space. Sampled latents can then be decoded into a neural field that can be rendered into a given viewpoint." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 224, + 287, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 224, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 46, + 224, + 287, + 308 + ], + "type": "text", + "content": "and mode dropping behaviors [1, 16, 35]. Denoising Diffusion models [20] (DDMs) have recently emerged as an alternative to GANs that avoid the aforementioned disadvantages [53, 56, 57]. However, DDMs model the data likelihood explicitly and are trained to reconstruct the training data. Thus, they have been used in limited scenarios [73, 77] since ground-truth 3D data is not readily available at scale." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 308, + 288, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 308, + 288, + 524 + ], + "spans": [ + { + "bbox": [ + 46, + 308, + 288, + 524 + ], + "type": "text", + "content": "To tackle the challenging problem of generating an entire scene with texture and geometry, we take inspiration from latent diffusion models (LDM) [53], which first construct an intermediate latent distribution of the training data then fit a diffusion model on the latent distribution. In Sec. 3.1, we introduce a scene auto-encoder that encodes the set of RGB images into a neural field representation consisting of density and feature voxel grids. To accurately capture a scene, the voxel grids' spatial dimension needs to be much larger than what current state-of-the-art LDMs can model. In Sec. 3.2, we show how we can further compress and decompose the explicit voxel grids into compressed latent representations to facilitate learning the data distribution. Finally, Sec. 3.3 introduces a latent diffusion model that models the latent distributions in a hierarchical manner. Fig. 2 shows an overview of our method, which we name NeuralField-LDM (NF-LDM). We provide training and additional architecture details in the supplementary." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 528, + 165, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 528, + 165, + 539 + ], + "spans": [ + { + "bbox": [ + 47, + 528, + 165, + 539 + ], + "type": "text", + "content": "3.1. Scene Auto-Encoder" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "content": "The goal of the scene auto-encoder is to obtain a 3D representation of the scene from input images by learning to reconstruct them. Fig. 3 depicts the auto-encoding process. The scene encoder is a 2D CNN and processes each RGB image " + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "inline_equation", + "content": "i_{1..N}" + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "content": " separately, producing a " + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{H\\times W\\times (D + C)}" + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "content": " dimensional 2D tensor for each image, where " + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "content": " are smaller than " + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "content": "'s size. We follow a similar procedure to Lift-Splat-Shoot (LSS) [45] to lift each 2D image feature map and combine them in the common voxel-based 3D neural field. We build a discrete frustum of size " + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "inline_equation", + "content": "H\\times W\\times D" + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "content": " with the camera poses " + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "content": " for each image. This frustum contains image features and density values for each pixel, along a pre-defined discrete set of " + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "content": " depths. Unlike LSS, we take the first " + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "content": " channels of the 2D CNN's output and use them as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 224, + 545, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 224, + 545, + 286 + ], + "spans": [ + { + "bbox": [ + 304, + 224, + 545, + 286 + ], + "type": "text", + "content": "density values. That is, the " + }, + { + "bbox": [ + 304, + 224, + 545, + 286 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 224, + 545, + 286 + ], + "type": "text", + "content": "th channel of the CNN's output at pixel " + }, + { + "bbox": [ + 304, + 224, + 545, + 286 + ], + "type": "inline_equation", + "content": "(h,w)" + }, + { + "bbox": [ + 304, + 224, + 545, + 286 + ], + "type": "text", + "content": " becomes the density value of the frustum entry at " + }, + { + "bbox": [ + 304, + 224, + 545, + 286 + ], + "type": "inline_equation", + "content": "(h,w,d)" + }, + { + "bbox": [ + 304, + 224, + 545, + 286 + ], + "type": "text", + "content": ". Motivated by the volume rendering equation [42], we get the occupancy weight " + }, + { + "bbox": [ + 304, + 224, + 545, + 286 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 304, + 224, + 545, + 286 + ], + "type": "text", + "content": " of each element " + }, + { + "bbox": [ + 304, + 224, + 545, + 286 + ], + "type": "inline_equation", + "content": "(h,w,d)" + }, + { + "bbox": [ + 304, + 224, + 545, + 286 + ], + "type": "text", + "content": " in the frustum using the density values " + }, + { + "bbox": [ + 304, + 224, + 545, + 286 + ], + "type": "inline_equation", + "content": "\\sigma \\geq 0" + }, + { + "bbox": [ + 304, + 224, + 545, + 286 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 293, + 545, + 337 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 293, + 545, + 337 + ], + "spans": [ + { + "bbox": [ + 305, + 293, + 545, + 337 + ], + "type": "interline_equation", + "content": "O (h, w, d) = \\exp \\left(- \\sum_ {j = 0} ^ {d - 1} \\sigma_ {(h, w, j)} \\delta_ {j}\\right) \\left(1 - \\exp \\left(- \\sigma_ {(h, w, d)} \\delta_ {d}\\right)\\right) \\tag {1}", + "image_path": "d35d3d184a4ccdfa6e6aa7d631520485ad88f309f4922fe0263f9680fda8b2b5.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 338, + 547, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 338, + 547, + 385 + ], + "spans": [ + { + "bbox": [ + 304, + 338, + 547, + 385 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 338, + 547, + 385 + ], + "type": "inline_equation", + "content": "h, w" + }, + { + "bbox": [ + 304, + 338, + 547, + 385 + ], + "type": "text", + "content": " denotes the pixel coordinate of the frustum and " + }, + { + "bbox": [ + 304, + 338, + 547, + 385 + ], + "type": "inline_equation", + "content": "\\delta_{j}" + }, + { + "bbox": [ + 304, + 338, + 547, + 385 + ], + "type": "text", + "content": " is the distance between each depth in the frustum. Using the occupancy weights, we put the last " + }, + { + "bbox": [ + 304, + 338, + 547, + 385 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 338, + 547, + 385 + ], + "type": "text", + "content": " channels of the CNN's output into the frustum " + }, + { + "bbox": [ + 304, + 338, + 547, + 385 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 304, + 338, + 547, + 385 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 333, + 393, + 545, + 407 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 393, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 333, + 393, + 545, + 407 + ], + "type": "interline_equation", + "content": "F (h, w, d) = \\left[ O (h, w, d) \\phi (h, w), \\sigma (h, w, d) \\right] \\tag {2}", + "image_path": "587ea6dffc86c6f0d3d4e4a28bc914b29cd8df33708da8beff8499743dc42efd.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 415, + 545, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 415, + 545, + 438 + ], + "spans": [ + { + "bbox": [ + 304, + 415, + 545, + 438 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 415, + 545, + 438 + ], + "type": "inline_equation", + "content": "\\phi (h,w)" + }, + { + "bbox": [ + 304, + 415, + 545, + 438 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 304, + 415, + 545, + 438 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 415, + 545, + 438 + ], + "type": "text", + "content": " -channeled feature vector at pixel " + }, + { + "bbox": [ + 304, + 415, + 545, + 438 + ], + "type": "inline_equation", + "content": "(h,w)" + }, + { + "bbox": [ + 304, + 415, + 545, + 438 + ], + "type": "text", + "content": " which is scaled by " + }, + { + "bbox": [ + 304, + 415, + 545, + 438 + ], + "type": "inline_equation", + "content": "O(h,w,d)" + }, + { + "bbox": [ + 304, + 415, + 545, + 438 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 304, + 415, + 545, + 438 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 304, + 415, + 545, + 438 + ], + "type": "text", + "content": " at depth " + }, + { + "bbox": [ + 304, + 415, + 545, + 438 + ], + "type": "inline_equation", + "content": "d" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 438, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 438, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 438, + 545, + 617 + ], + "type": "text", + "content": "After constructing the frustum for each view, we transform the frustums to world coordinates and fuse them into a shared 3D neural field, represented as density and feature voxel grids. Let " + }, + { + "bbox": [ + 304, + 438, + 545, + 617 + ], + "type": "inline_equation", + "content": "V_{\\text{Density}}" + }, + { + "bbox": [ + 304, + 438, + 545, + 617 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 438, + 545, + 617 + ], + "type": "inline_equation", + "content": "V_{\\text{Feat}}" + }, + { + "bbox": [ + 304, + 438, + 545, + 617 + ], + "type": "text", + "content": " denote the density and feature grid, respectively. This formulation of representing a scene with density and feature grids has been explored before [65] for optimization-based scene reconstruction and we utilize it as an intermediate representation for our scene auto-encoder. " + }, + { + "bbox": [ + 304, + 438, + 545, + 617 + ], + "type": "inline_equation", + "content": "V_{\\text{Density}, \\text{Feat}}" + }, + { + "bbox": [ + 304, + 438, + 545, + 617 + ], + "type": "text", + "content": " have the same spatial size, and each voxel in " + }, + { + "bbox": [ + 304, + 438, + 545, + 617 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 304, + 438, + 545, + 617 + ], + "type": "text", + "content": " represents a region in the world coordinate system. For each voxel indexed by " + }, + { + "bbox": [ + 304, + 438, + 545, + 617 + ], + "type": "inline_equation", + "content": "(x, y, z)" + }, + { + "bbox": [ + 304, + 438, + 545, + 617 + ], + "type": "text", + "content": ", we pool all densities and features of the corresponding frustum entries. In this paper, we simply take the mean of the pooled features. More sophisticated pooling functions (e.g. attention) can be used, which we leave as future work." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "content": "Finally, we perform volume rendering using the camera poses " + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "content": " to project " + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "content": " onto a 2D feature map. We trilinearly interpolate the values on each voxel to get the feature and density for each sampling point along the camera rays. 2D features are then fed into a CNN decoder that produces the output image " + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\hat{i}" + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "content": ". We denote rendering of voxels to output images as " + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\hat{i} = r(V,\\kappa)" + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "content": ". From the volume rendering process, we also get the expected depth " + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\hat{\\rho}" + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "content": " along each ray [50]. The scene" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8498" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 69, + 280, + 201 + ], + "blocks": [ + { + "bbox": [ + 53, + 69, + 280, + 201 + ], + "lines": [ + { + "bbox": [ + 53, + 69, + 280, + 201 + ], + "spans": [ + { + "bbox": [ + 53, + 69, + 280, + 201 + ], + "type": "image", + "image_path": "01e062238090e1950f0f6a8a49e12e05623432fbde673f153e72ed7cd1dde40f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 201, + 287, + 235 + ], + "lines": [ + { + "bbox": [ + 47, + 201, + 287, + 235 + ], + "spans": [ + { + "bbox": [ + 47, + 201, + 287, + 235 + ], + "type": "text", + "content": "Figure 3. Scene Auto-Encoder: Each input image is processed with a 2D CNN then lifted up to 3D and merged into the shared voxel grids. Density prediction is not shown here for brevity." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 247, + 288, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 247, + 288, + 331 + ], + "spans": [ + { + "bbox": [ + 46, + 247, + 288, + 331 + ], + "type": "text", + "content": "auto-encoding pipeline is trained with an image reconstruction loss " + }, + { + "bbox": [ + 46, + 247, + 288, + 331 + ], + "type": "inline_equation", + "content": "||i - \\hat{i}||" + }, + { + "bbox": [ + 46, + 247, + 288, + 331 + ], + "type": "text", + "content": " and a depth supervision loss " + }, + { + "bbox": [ + 46, + 247, + 288, + 331 + ], + "type": "inline_equation", + "content": "||\\rho - \\hat{\\rho}||" + }, + { + "bbox": [ + 46, + 247, + 288, + 331 + ], + "type": "text", + "content": ". In the case of sparse depth measurements, we only supervise the pixels with recorded depth. We can further improve the quality of the auto-encoder with adversarial loss as in VQGAN [13] or by doing a few optimization steps at inference time, which we discuss in the supplementary." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 338, + 198, + 350 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 338, + 198, + 350 + ], + "spans": [ + { + "bbox": [ + 47, + 338, + 198, + 350 + ], + "type": "text", + "content": "3.2. Latent Voxel Auto-Encoder" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 354, + 287, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 354, + 287, + 486 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 287, + 486 + ], + "type": "text", + "content": "It is possible to fit a generative model on voxel grids obtained from Sec. 3.1. However, to capture real-world scenes, the dimensionality of the representation needs to be much larger than what SOTA diffusion models can be trained on. For example, Imagen [56] trains DDMs on " + }, + { + "bbox": [ + 46, + 354, + 287, + 486 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 46, + 354, + 287, + 486 + ], + "type": "text", + "content": " RGB images, and we use voxels of size " + }, + { + "bbox": [ + 46, + 354, + 287, + 486 + ], + "type": "inline_equation", + "content": "128 \\times 128 \\times 32" + }, + { + "bbox": [ + 46, + 354, + 287, + 486 + ], + "type": "text", + "content": " with 32 channels. We thus introduce a latent auto-encoder (LAE) that compresses voxels into a 128-dimensional global latent as well as coarse (3D) and fine (2D) quantized latents with channel dimensions of four and spatial dimensions " + }, + { + "bbox": [ + 46, + 354, + 287, + 486 + ], + "type": "inline_equation", + "content": "32 \\times 32 \\times 16" + }, + { + "bbox": [ + 46, + 354, + 287, + 486 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 354, + 287, + 486 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 46, + 354, + 287, + 486 + ], + "type": "text", + "content": " respectively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 486, + 288, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 288, + 701 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 288, + 701 + ], + "type": "text", + "content": "We concatenate " + }, + { + "bbox": [ + 46, + 486, + 288, + 701 + ], + "type": "inline_equation", + "content": "V_{\\text{Density}}" + }, + { + "bbox": [ + 46, + 486, + 288, + 701 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 486, + 288, + 701 + ], + "type": "inline_equation", + "content": "V_{\\text{Feat}}" + }, + { + "bbox": [ + 46, + 486, + 288, + 701 + ], + "type": "text", + "content": " along the channel dimension and use separate CNN encoders to encode the voxel grid " + }, + { + "bbox": [ + 46, + 486, + 288, + 701 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 46, + 486, + 288, + 701 + ], + "type": "text", + "content": " into a hierarchy of three latents: 1D global latent " + }, + { + "bbox": [ + 46, + 486, + 288, + 701 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 46, + 486, + 288, + 701 + ], + "type": "text", + "content": ", 3D coarse latent " + }, + { + "bbox": [ + 46, + 486, + 288, + 701 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 486, + 288, + 701 + ], + "type": "text", + "content": ", and 2D fine latent " + }, + { + "bbox": [ + 46, + 486, + 288, + 701 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 486, + 288, + 701 + ], + "type": "text", + "content": ", as shown in Fig. 2. The intuition for this design is that " + }, + { + "bbox": [ + 46, + 486, + 288, + 701 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 46, + 486, + 288, + 701 + ], + "type": "text", + "content": " is responsible for representing the global properties of the scene, such as the time of the day, " + }, + { + "bbox": [ + 46, + 486, + 288, + 701 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 486, + 288, + 701 + ], + "type": "text", + "content": " represents coarse 3D scene structure, and " + }, + { + "bbox": [ + 46, + 486, + 288, + 701 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 486, + 288, + 701 + ], + "type": "text", + "content": " is a 2D tensor with the same horizontal size " + }, + { + "bbox": [ + 46, + 486, + 288, + 701 + ], + "type": "inline_equation", + "content": "X \\times Y" + }, + { + "bbox": [ + 46, + 486, + 288, + 701 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 46, + 486, + 288, + 701 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 46, + 486, + 288, + 701 + ], + "type": "text", + "content": ", which gives further details for each location " + }, + { + "bbox": [ + 46, + 486, + 288, + 701 + ], + "type": "inline_equation", + "content": "(x, y)" + }, + { + "bbox": [ + 46, + 486, + 288, + 701 + ], + "type": "text", + "content": " in bird's eye view perspective. We empirically found that 2D CNNs perform similarly to 3D CNNs while being more efficient, thus we use 2D CNNs throughout. To use 2D CNNs for the 3D input " + }, + { + "bbox": [ + 46, + 486, + 288, + 701 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 46, + 486, + 288, + 701 + ], + "type": "text", + "content": ", we concatenate " + }, + { + "bbox": [ + 46, + 486, + 288, + 701 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 46, + 486, + 288, + 701 + ], + "type": "text", + "content": "'s vertical axis along the channel dimension and feed it to the encoders. We also add latent regularizations to avoid high variance latent spaces [53]. For the 1D vector " + }, + { + "bbox": [ + 46, + 486, + 288, + 701 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 46, + 486, + 288, + 701 + ], + "type": "text", + "content": ", we use a small KL-penalty via the reparameterization trick [33], and for " + }, + { + "bbox": [ + 46, + 486, + 288, + 701 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 486, + 288, + 701 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 486, + 288, + 701 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 486, + 288, + 701 + ], + "type": "text", + "content": ", we impose a vector-quantization [13,69] layer to regularize them." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "text", + "content": "The CNN decoder is similarly a 2D CNN, and takes " + }, + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "inline_equation", + "content": "c" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 317, + 69, + 533, + 280 + ], + "blocks": [ + { + "bbox": [ + 317, + 69, + 533, + 280 + ], + "lines": [ + { + "bbox": [ + 317, + 69, + 533, + 280 + ], + "spans": [ + { + "bbox": [ + 317, + 69, + 533, + 280 + ], + "type": "image", + "image_path": "ed6be463d777247a4f659c4585dc74e0fbf29f520d27bbf5657b0030a8db5bd5.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 282, + 546, + 338 + ], + "lines": [ + { + "bbox": [ + 305, + 282, + 546, + 338 + ], + "spans": [ + { + "bbox": [ + 305, + 282, + 546, + 338 + ], + "type": "text", + "content": "Figure 4. Hierarchical LDM. Top: LDM " + }, + { + "bbox": [ + 305, + 282, + 546, + 338 + ], + "type": "inline_equation", + "content": "\\psi_{g}" + }, + { + "bbox": [ + 305, + 282, + 546, + 338 + ], + "type": "text", + "content": " for KL-regulairzed global latent " + }, + { + "bbox": [ + 305, + 282, + 546, + 338 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 305, + 282, + 546, + 338 + ], + "type": "text", + "content": ". Middle: LDM " + }, + { + "bbox": [ + 305, + 282, + 546, + 338 + ], + "type": "inline_equation", + "content": "\\psi_{c}" + }, + { + "bbox": [ + 305, + 282, + 546, + 338 + ], + "type": "text", + "content": " for vector-quantized coarse latent " + }, + { + "bbox": [ + 305, + 282, + 546, + 338 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 305, + 282, + 546, + 338 + ], + "type": "text", + "content": ". Bottom: LDM " + }, + { + "bbox": [ + 305, + 282, + 546, + 338 + ], + "type": "inline_equation", + "content": "\\psi_{f}" + }, + { + "bbox": [ + 305, + 282, + 546, + 338 + ], + "type": "text", + "content": " for vector-quantized fine latent " + }, + { + "bbox": [ + 305, + 282, + 546, + 338 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 305, + 282, + 546, + 338 + ], + "type": "text", + "content": ". All LDMs optionally take an additional conditioning variable as input, such as a Bird's Eye View segmentation map as depicted here." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 349, + 546, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 349, + 546, + 457 + ], + "spans": [ + { + "bbox": [ + 304, + 349, + 546, + 457 + ], + "type": "text", + "content": "concatenated along vertical axis, as the initial input. The decoder uses conditional group normalization layers [71] with " + }, + { + "bbox": [ + 304, + 349, + 546, + 457 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 304, + 349, + 546, + 457 + ], + "type": "text", + "content": " as the conditioning variable. Lastly, we concatenate " + }, + { + "bbox": [ + 304, + 349, + 546, + 457 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 349, + 546, + 457 + ], + "type": "text", + "content": " to an intermediate tensor in the decoder. The latent decoder outputs " + }, + { + "bbox": [ + 304, + 349, + 546, + 457 + ], + "type": "inline_equation", + "content": "\\hat{V}" + }, + { + "bbox": [ + 304, + 349, + 546, + 457 + ], + "type": "text", + "content": " which is the reconstructed voxel. LAE is trained with the voxel reconstruction loss " + }, + { + "bbox": [ + 304, + 349, + 546, + 457 + ], + "type": "inline_equation", + "content": "||V - \\hat{V}||" + }, + { + "bbox": [ + 304, + 349, + 546, + 457 + ], + "type": "text", + "content": " along with the image reconstruction loss " + }, + { + "bbox": [ + 304, + 349, + 546, + 457 + ], + "type": "inline_equation", + "content": "||i - \\hat{i}||" + }, + { + "bbox": [ + 304, + 349, + 546, + 457 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 349, + 546, + 457 + ], + "type": "inline_equation", + "content": "\\hat{i} = r(\\hat{V},\\kappa)" + }, + { + "bbox": [ + 304, + 349, + 546, + 457 + ], + "type": "text", + "content": ". Note that the image reconstruction loss only helps with the learning of LAE, and the scene auto-encoder is kept fixed." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 464, + 504, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 464, + 504, + 477 + ], + "spans": [ + { + "bbox": [ + 306, + 464, + 504, + 477 + ], + "type": "text", + "content": "3.3. Hierarchical Latent Diffusion Models" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 483, + 546, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 483, + 546, + 638 + ], + "spans": [ + { + "bbox": [ + 304, + 483, + 546, + 638 + ], + "type": "text", + "content": "Given the latent variables " + }, + { + "bbox": [ + 304, + 483, + 546, + 638 + ], + "type": "inline_equation", + "content": "g, c, f" + }, + { + "bbox": [ + 304, + 483, + 546, + 638 + ], + "type": "text", + "content": " that represent a voxel-based scene representation " + }, + { + "bbox": [ + 304, + 483, + 546, + 638 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 304, + 483, + 546, + 638 + ], + "type": "text", + "content": ", we define our generative model as " + }, + { + "bbox": [ + 304, + 483, + 546, + 638 + ], + "type": "inline_equation", + "content": "p(V, g, c, f) = p(V | g, c, f) p(f | g, c) p(c | g) p(g)" + }, + { + "bbox": [ + 304, + 483, + 546, + 638 + ], + "type": "text", + "content": " with Denoising Diffusion Models (DDMs) [21]. In general, DDMs with discrete time steps have a fixed Markovian forward process " + }, + { + "bbox": [ + 304, + 483, + 546, + 638 + ], + "type": "inline_equation", + "content": "q(x_{t} | x_{t-1})" + }, + { + "bbox": [ + 304, + 483, + 546, + 638 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 483, + 546, + 638 + ], + "type": "inline_equation", + "content": "q(x_{0})" + }, + { + "bbox": [ + 304, + 483, + 546, + 638 + ], + "type": "text", + "content": " denotes the data distribution and " + }, + { + "bbox": [ + 304, + 483, + 546, + 638 + ], + "type": "inline_equation", + "content": "q(x_{T})" + }, + { + "bbox": [ + 304, + 483, + 546, + 638 + ], + "type": "text", + "content": " is defined to be close to the standard normal distribution, where we use the subscript to denote the time step. DDMs then learn to revert the forward process " + }, + { + "bbox": [ + 304, + 483, + 546, + 638 + ], + "type": "inline_equation", + "content": "p_{\\theta}(x_{t-1} | x_{t})" + }, + { + "bbox": [ + 304, + 483, + 546, + 638 + ], + "type": "text", + "content": " with learnable parameters " + }, + { + "bbox": [ + 304, + 483, + 546, + 638 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 483, + 546, + 638 + ], + "type": "text", + "content": ". It can be shown that learning the reverse process is equivalent to learning to denoise " + }, + { + "bbox": [ + 304, + 483, + 546, + 638 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 304, + 483, + 546, + 638 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 483, + 546, + 638 + ], + "type": "inline_equation", + "content": "x_{0}" + }, + { + "bbox": [ + 304, + 483, + 546, + 638 + ], + "type": "text", + "content": " for all timesteps " + }, + { + "bbox": [ + 304, + 483, + 546, + 638 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 483, + 546, + 638 + ], + "type": "text", + "content": " [21, 24] by reducing the following loss:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 361, + 638, + 545, + 661 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 638, + 545, + 661 + ], + "spans": [ + { + "bbox": [ + 361, + 638, + 545, + 661 + ], + "type": "interline_equation", + "content": "\\underset {t, \\epsilon , x _ {0}} {\\mathbb {E}} \\left[ w \\left(\\lambda_ {t}\\right) \\| x _ {0} - \\hat {x} _ {\\theta} \\left(x _ {t}, t\\right) \\| _ {2} ^ {2} \\right] \\tag {3}", + "image_path": "e37dda5de6b0b34d58c4fe69d772880367fd30a0d1d7f7e851d977f66ad33e37.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 665, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 546, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 665, + 546, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 665, + 546, + 713 + ], + "type": "text", + "content": " is sampled from a uniform distribution for timesteps, " + }, + { + "bbox": [ + 304, + 665, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 304, + 665, + 546, + 713 + ], + "type": "text", + "content": " is sampled from the standard normal to noise the data " + }, + { + "bbox": [ + 304, + 665, + 546, + 713 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 304, + 665, + 546, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 665, + 546, + 713 + ], + "type": "inline_equation", + "content": "w(\\lambda_t)" + }, + { + "bbox": [ + 304, + 665, + 546, + 713 + ], + "type": "text", + "content": " is a timestep dependent weighting constant, and " + }, + { + "bbox": [ + 304, + 665, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\hat{x}_{\\theta}" + }, + { + "bbox": [ + 304, + 665, + 546, + 713 + ], + "type": "text", + "content": " denotes the learned denoising model." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8499" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 59, + 72, + 286, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 72, + 286, + 83 + ], + "spans": [ + { + "bbox": [ + 59, + 72, + 286, + 83 + ], + "type": "text", + "content": "We train our hierarchical LDM with the following losses:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 80, + 86, + 286, + 109 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 86, + 286, + 109 + ], + "spans": [ + { + "bbox": [ + 80, + 86, + 286, + 109 + ], + "type": "interline_equation", + "content": "\\underset {t, \\epsilon , g _ {0}} {\\mathbb {E}} \\left[ w \\left(\\lambda_ {t}\\right) \\| g _ {0} - \\psi_ {g} \\left(g _ {t}, t\\right) \\| _ {2} ^ {2} \\right] \\tag {4}", + "image_path": "1d1d581b524425a857cb03f27d3160610b3ce9f1a130ba744dabde36cbccfde8.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 80, + 111, + 286, + 133 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 111, + 286, + 133 + ], + "spans": [ + { + "bbox": [ + 80, + 111, + 286, + 133 + ], + "type": "interline_equation", + "content": "\\underset {t, \\epsilon , g _ {0}, c _ {0}} {\\mathbb {E}} \\left[ w \\left(\\lambda_ {t}\\right) \\| c _ {0} - \\psi_ {c} \\left(c _ {t}, g _ {0}, t\\right) \\| _ {2} ^ {2} \\right] \\tag {5}", + "image_path": "479902462c472ea1f3efced09641f92c1cacc0946ff85c4600bf0ed97dc49065.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 80, + 136, + 286, + 159 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 136, + 286, + 159 + ], + "spans": [ + { + "bbox": [ + 80, + 136, + 286, + 159 + ], + "type": "interline_equation", + "content": "\\underset {t, \\epsilon , g _ {0}, c _ {0}, f _ {0}} {\\mathbb {E}} \\left[ w \\left(\\lambda_ {t}\\right) \\| f _ {0} - \\psi_ {f} \\left(f _ {t}, g _ {0}, c _ {0}, t\\right) \\| _ {2} ^ {2} \\right] \\tag {6}", + "image_path": "37dfbdb8e3162351bd44191731057b7f0e8f34f9fe866dbe94bb3641bdb324a5.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 163, + 286, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 163, + 286, + 317 + ], + "spans": [ + { + "bbox": [ + 46, + 163, + 286, + 317 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 163, + 286, + 317 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 46, + 163, + 286, + 317 + ], + "type": "text", + "content": " denotes the learnable denoising networks for " + }, + { + "bbox": [ + 46, + 163, + 286, + 317 + ], + "type": "inline_equation", + "content": "g, c, f" + }, + { + "bbox": [ + 46, + 163, + 286, + 317 + ], + "type": "text", + "content": ". Fig. 4 visualizes the diffusion models. " + }, + { + "bbox": [ + 46, + 163, + 286, + 317 + ], + "type": "inline_equation", + "content": "\\psi_{g}" + }, + { + "bbox": [ + 46, + 163, + 286, + 317 + ], + "type": "text", + "content": " is implemented with linear layers with skip connections and " + }, + { + "bbox": [ + 46, + 163, + 286, + 317 + ], + "type": "inline_equation", + "content": "\\psi_{c}" + }, + { + "bbox": [ + 46, + 163, + 286, + 317 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 163, + 286, + 317 + ], + "type": "inline_equation", + "content": "\\psi_{f}" + }, + { + "bbox": [ + 46, + 163, + 286, + 317 + ], + "type": "text", + "content": " adopt the U-net architecture [55]. " + }, + { + "bbox": [ + 46, + 163, + 286, + 317 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 46, + 163, + 286, + 317 + ], + "type": "text", + "content": " is fed into " + }, + { + "bbox": [ + 46, + 163, + 286, + 317 + ], + "type": "inline_equation", + "content": "\\psi_{c}" + }, + { + "bbox": [ + 46, + 163, + 286, + 317 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 163, + 286, + 317 + ], + "type": "inline_equation", + "content": "\\psi_{f}" + }, + { + "bbox": [ + 46, + 163, + 286, + 317 + ], + "type": "text", + "content": " with conditional group normalization layers. " + }, + { + "bbox": [ + 46, + 163, + 286, + 317 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 163, + 286, + 317 + ], + "type": "text", + "content": " is interpolated and concatenated to the input to " + }, + { + "bbox": [ + 46, + 163, + 286, + 317 + ], + "type": "inline_equation", + "content": "\\psi_{f}" + }, + { + "bbox": [ + 46, + 163, + 286, + 317 + ], + "type": "text", + "content": ". The camera poses contain the trajectory the camera is travelling, and this information can be useful for modelling a 3D scene as it tells the model where to focus on generating. Therefore, we concatenate the camera trajectory information to " + }, + { + "bbox": [ + 46, + 163, + 286, + 317 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 46, + 163, + 286, + 317 + ], + "type": "text", + "content": " and also learn to sample it. For brevity, we still call the concatenated vector " + }, + { + "bbox": [ + 46, + 163, + 286, + 317 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 46, + 163, + 286, + 317 + ], + "type": "text", + "content": ". For conditional generation, each " + }, + { + "bbox": [ + 46, + 163, + 286, + 317 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 46, + 163, + 286, + 317 + ], + "type": "text", + "content": " takes the conditioning variable as input with cross-attention layers [53]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 319, + 286, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 319, + 286, + 426 + ], + "spans": [ + { + "bbox": [ + 46, + 319, + 286, + 426 + ], + "type": "text", + "content": "Each " + }, + { + "bbox": [ + 46, + 319, + 286, + 426 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 46, + 319, + 286, + 426 + ], + "type": "text", + "content": " can be trained in parallel and, once trained, can be sampled one after another following the hierarchy. In practice, we use the v-prediction parameterization [57] that has been shown to have better convergence and training stability [56, 57]. Once " + }, + { + "bbox": [ + 46, + 319, + 286, + 426 + ], + "type": "inline_equation", + "content": "g, c, f" + }, + { + "bbox": [ + 46, + 319, + 286, + 426 + ], + "type": "text", + "content": " are sampled, we can use the latent decoder from Sec. 3.2 to construct the voxel " + }, + { + "bbox": [ + 46, + 319, + 286, + 426 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 46, + 319, + 286, + 426 + ], + "type": "text", + "content": " which represents the neural field for the sampled scene.. Following the volume rendering and decoding step in Sec. 3.1, the sampled scene can be visualized from desired viewpoints." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 429, + 264, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 429, + 264, + 441 + ], + "spans": [ + { + "bbox": [ + 47, + 429, + 264, + 441 + ], + "type": "text", + "content": "3.4. Post-Optimizing Generated Neural Fields" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 447, + 286, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 447, + 286, + 543 + ], + "spans": [ + { + "bbox": [ + 46, + 447, + 286, + 543 + ], + "type": "text", + "content": "Samples generated from our model on real-world data contain reasonable texture and geometry (Fig. 10), but can be further optimized by leveraging recent advances in 2D image diffusion models trained on orders of magnitude more data. Specifically, we iteratively update initially generated voxels, " + }, + { + "bbox": [ + 46, + 447, + 286, + 543 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 46, + 447, + 286, + 543 + ], + "type": "text", + "content": ", by rendering viewpoints from the scene and applying Score Distillation Sampling (SDS) [46] loss on each image independently:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 52, + 547, + 286, + 571 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 547, + 286, + 571 + ], + "spans": [ + { + "bbox": [ + 52, + 547, + 286, + 571 + ], + "type": "interline_equation", + "content": "\\nabla_ {V} L _ {S D S} = \\underset {\\epsilon , t, \\kappa} {\\mathbb {E}} \\left[ w \\left(\\lambda_ {t}\\right) \\left(\\epsilon - \\hat {\\epsilon_ {\\theta}} (r (V, \\kappa), t)\\right) \\frac {\\partial r (V , \\kappa)}{\\partial V} \\right] \\tag {7}", + "image_path": "686d427cb1aaf26d735984be7d9c2b765b15d40cdd68c2e92e0aa97acb62a920.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 577, + 286, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 577, + 286, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 577, + 286, + 696 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 577, + 286, + 696 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 46, + 577, + 286, + 696 + ], + "type": "text", + "content": " is sampled uniformly in a " + }, + { + "bbox": [ + 46, + 577, + 286, + 696 + ], + "type": "inline_equation", + "content": "6m^2" + }, + { + "bbox": [ + 46, + 577, + 286, + 696 + ], + "type": "text", + "content": " region around the origin of the scene with random rotation about the vertical axis, " + }, + { + "bbox": [ + 46, + 577, + 286, + 696 + ], + "type": "inline_equation", + "content": "w(\\lambda_t)" + }, + { + "bbox": [ + 46, + 577, + 286, + 696 + ], + "type": "text", + "content": " is the weighting schedule used to train " + }, + { + "bbox": [ + 46, + 577, + 286, + 696 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon_{\\theta}}" + }, + { + "bbox": [ + 46, + 577, + 286, + 696 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 577, + 286, + 696 + ], + "type": "inline_equation", + "content": "t\\sim U[0.02T,0.2T]" + }, + { + "bbox": [ + 46, + 577, + 286, + 696 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 46, + 577, + 286, + 696 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 577, + 286, + 696 + ], + "type": "text", + "content": " is the amount of noise steps used to train " + }, + { + "bbox": [ + 46, + 577, + 286, + 696 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon_{\\theta}}" + }, + { + "bbox": [ + 46, + 577, + 286, + 696 + ], + "type": "text", + "content": ". Note that for latent diffusion models, the noise prediction step is applied after encoding " + }, + { + "bbox": [ + 46, + 577, + 286, + 696 + ], + "type": "inline_equation", + "content": "r(V,\\kappa)" + }, + { + "bbox": [ + 46, + 577, + 286, + 696 + ], + "type": "text", + "content": " to the LDM's latent space and the partial gradient term is updated appropriately. For " + }, + { + "bbox": [ + 46, + 577, + 286, + 696 + ], + "type": "inline_equation", + "content": "\\hat{\\epsilon_{\\theta}}" + }, + { + "bbox": [ + 46, + 577, + 286, + 696 + ], + "type": "text", + "content": ", we use an off-the-shelf latent diffusion model [53], finetuned to condition on CLIP image embeddings " + }, + { + "bbox": [ + 46, + 577, + 286, + 696 + ], + "type": "inline_equation", + "content": "[47]^1" + }, + { + "bbox": [ + 46, + 577, + 286, + 696 + ], + "type": "text", + "content": ". We found that CLIP contains a representativeness matrix" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 307, + 72, + 425, + 111 + ], + "blocks": [ + { + "bbox": [ + 307, + 72, + 425, + 111 + ], + "lines": [ + { + "bbox": [ + 307, + 72, + 425, + 111 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 425, + 111 + ], + "type": "image", + "image_path": "2ea9770e5dc69ebb6ad3226e32dc2ba9354bf165faa9180aaa077abab306d5ae.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 426, + 72, + 543, + 111 + ], + "blocks": [ + { + "bbox": [ + 426, + 72, + 543, + 111 + ], + "lines": [ + { + "bbox": [ + 426, + 72, + 543, + 111 + ], + "spans": [ + { + "bbox": [ + 426, + 72, + 543, + 111 + ], + "type": "image", + "image_path": "2ae539322fb9ecdeec9f45be497ef0576bd2c9760c6a10da7cdb04fab91cebe8.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 307, + 113, + 386, + 153 + ], + "blocks": [ + { + "bbox": [ + 307, + 113, + 386, + 153 + ], + "lines": [ + { + "bbox": [ + 307, + 113, + 386, + 153 + ], + "spans": [ + { + "bbox": [ + 307, + 113, + 386, + 153 + ], + "type": "image", + "image_path": "d07685b02488898282a4648785c86b7c290ce42a96a0acba94c7bfaeb60c1f2d.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 386, + 113, + 543, + 153 + ], + "blocks": [ + { + "bbox": [ + 386, + 113, + 543, + 153 + ], + "lines": [ + { + "bbox": [ + 386, + 113, + 543, + 153 + ], + "spans": [ + { + "bbox": [ + 386, + 113, + 543, + 153 + ], + "type": "image", + "image_path": "41d03d8227e08ae330ef93128c0d7a9d288883beab02da68729370592d46a572.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 307, + 155, + 386, + 196 + ], + "blocks": [ + { + "bbox": [ + 307, + 155, + 386, + 196 + ], + "lines": [ + { + "bbox": [ + 307, + 155, + 386, + 196 + ], + "spans": [ + { + "bbox": [ + 307, + 155, + 386, + 196 + ], + "type": "image", + "image_path": "a398490987c93ea73f453a89f7ccccb079ac1c51e836915a2e3ba17d9dce9694.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 201, + 545, + 233 + ], + "lines": [ + { + "bbox": [ + 306, + 201, + 545, + 233 + ], + "spans": [ + { + "bbox": [ + 306, + 201, + 545, + 233 + ], + "type": "text", + "content": "Figure 5. Datasets: Top-left: VizDoom [29]. Top-right: Replica [64]. Middle: Carla [12]. Bottom: AVD. For Carla and AVD, we visualize a subset of available cameras." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 386, + 155, + 543, + 196 + ], + "blocks": [ + { + "bbox": [ + 386, + 155, + 543, + 196 + ], + "lines": [ + { + "bbox": [ + 386, + 155, + 543, + 196 + ], + "spans": [ + { + "bbox": [ + 386, + 155, + 543, + 196 + ], + "type": "image", + "image_path": "b9446f3503473f385534e348384da5b56e2395971ea38b7be5cb23b178879374.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 249, + 545, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 249, + 545, + 456 + ], + "spans": [ + { + "bbox": [ + 304, + 249, + 545, + 456 + ], + "type": "text", + "content": "tation of the quality of images that the LDM is able to interpret: denoising an image while conditioning on CLIP image embeddings of our model's samples produced images with similar geometry distortions and texture errors. We leverage this property by optimizing " + }, + { + "bbox": [ + 304, + 249, + 545, + 456 + ], + "type": "inline_equation", + "content": "L_{SDS}" + }, + { + "bbox": [ + 304, + 249, + 545, + 456 + ], + "type": "text", + "content": " with negative guidance. Letting " + }, + { + "bbox": [ + 304, + 249, + 545, + 456 + ], + "type": "inline_equation", + "content": "y, y'" + }, + { + "bbox": [ + 304, + 249, + 545, + 456 + ], + "type": "text", + "content": " be CLIP embeddings of clean image conditioning (e.g. dataset images) and artifact conditioning (e.g. samples) respectively, we perform classifier-free guidance [23] with conditioning vector " + }, + { + "bbox": [ + 304, + 249, + 545, + 456 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 249, + 545, + 456 + ], + "type": "text", + "content": ", but replace the unconditional embedding with " + }, + { + "bbox": [ + 304, + 249, + 545, + 456 + ], + "type": "inline_equation", + "content": "y'" + }, + { + "bbox": [ + 304, + 249, + 545, + 456 + ], + "type": "text", + "content": ". As shown in the supplementary, this is equivalent (up to scale) to sampling from " + }, + { + "bbox": [ + 304, + 249, + 545, + 456 + ], + "type": "inline_equation", + "content": "\\frac{p(x|y)^{\\alpha}}{p(x|y')}" + }, + { + "bbox": [ + 304, + 249, + 545, + 456 + ], + "type": "text", + "content": " at each denoising step where " + }, + { + "bbox": [ + 304, + 249, + 545, + 456 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 249, + 545, + 456 + ], + "type": "text", + "content": " controls the trade-off between sampling towards dataset images and away from images with artifacts. We stress that this post-optimization is only successful due to the strong scene prior contained in our voxel samples, as shown by our comparison to running optimization on randomly initialized voxels in the supplementary." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 466, + 387, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 466, + 387, + 479 + ], + "spans": [ + { + "bbox": [ + 306, + 466, + 387, + 479 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 486, + 545, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 545, + 521 + ], + "type": "text", + "content": "We evaluate NeuralField-LDM on the following four datasets (Fig. 5). Each dataset contains RGB images and a depth measurement with their corresponding camera poses." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 522, + 545, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 545, + 569 + ], + "type": "text", + "content": "VizDoom [28] consists of front-view sensor observations obtained by navigating inside a synthetic game environment. We use the dataset provided by [9], which contains 34 trajectories with a length of 600." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 570, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 605 + ], + "type": "text", + "content": "Replica [64] is a dataset of high-quality reconstructions of 18 indoor scenes, containing 101 front-view trajectories with a length of 100. We use the dataset provided by [9]." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 606, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 606, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 606, + 545, + 689 + ], + "type": "text", + "content": "Carla [12] is an open-source simulation platform for self-driving research. We mimic the camera settings for a self-driving car, by placing six cameras (front-left, front, front-right, back-left, back, back-right), covering 360 degrees with some overlaps, and move the car in a randomly sampled direction and distance for 10 timesteps. We sample 43K datapoints, each containing 60 images." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "content": "AVD is an in-house dataset of human driving recordings in roads and parking lots. It has ten cameras with varying" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 58, + 702, + 220, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 220, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 220, + 712 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 58, + 702, + 220, + 712 + ], + "type": "text", + "content": "https://github.com/justinpinkney/stable-diffusion" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8500" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 286, + 353 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 286, + 353 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 286, + 353 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 286, + 353 + ], + "type": "image", + "image_path": "b1c935c6e1e32d6f692e5f9c1bd56ae3253af7d2de0ffd29b6205c266d678831.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 357, + 286, + 378 + ], + "lines": [ + { + "bbox": [ + 47, + 357, + 286, + 378 + ], + "spans": [ + { + "bbox": [ + 47, + 357, + 286, + 378 + ], + "type": "text", + "content": "Figure 6. Generated Scenes: The top three rows are samples from Carla, and the bottom three rows are samples from AVD." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 70, + 381, + 261, + 589 + ], + "blocks": [ + { + "bbox": [ + 70, + 381, + 261, + 589 + ], + "lines": [ + { + "bbox": [ + 70, + 381, + 261, + 589 + ], + "spans": [ + { + "bbox": [ + 70, + 381, + 261, + 589 + ], + "type": "image", + "image_path": "77e6bdd80d513de7c69a7233eab5c29539258efc4d04ccc3faa0ec87d6944e9c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 590, + 286, + 622 + ], + "lines": [ + { + "bbox": [ + 47, + 590, + 286, + 622 + ], + "spans": [ + { + "bbox": [ + 47, + 590, + 286, + 622 + ], + "type": "text", + "content": "Figure 7. We run marching-cubes [38] on the density voxels to visualize the geometry of the samples generated by NF-LDM. The structure of the scene is reflected well in the mesh." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 627, + 286, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 627, + 286, + 651 + ], + "spans": [ + { + "bbox": [ + 47, + 627, + 286, + 651 + ], + "type": "text", + "content": "lens types along with Lidar for depth measurement. It has 73K sequences, each with 8 frames extracted at 10 fps." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 659, + 174, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 659, + 174, + 672 + ], + "spans": [ + { + "bbox": [ + 47, + 659, + 174, + 672 + ], + "type": "text", + "content": "4.1. Baseline Comparisons" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 677, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 286, + 713 + ], + "type": "text", + "content": "Unconditional Generation We evaluate the unconditional generation performance of NF-LDM by comparing it with baseline models. All results are without the post" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 321, + 71, + 529, + 144 + ], + "blocks": [ + { + "bbox": [ + 321, + 71, + 529, + 144 + ], + "lines": [ + { + "bbox": [ + 321, + 71, + 529, + 144 + ], + "spans": [ + { + "bbox": [ + 321, + 71, + 529, + 144 + ], + "type": "table", + "html": "
CriterionMethodDepthVizDoomReplica
FID (↓)GRAF [59]X47.5065.37
π-GAN [4]X143.55166.55
GSN [8]✓37.2141.75
GAUDI [2]✓33.7018.75
NF-LDM✓19.54*14.59
", + "image_path": "f19b24e71ef2e1eb056f18cba4e2556f876cb148df02f0e4819dd602a5ef1a4e.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 325, + 175, + 525, + 232 + ], + "blocks": [ + { + "bbox": [ + 306, + 145, + 545, + 168 + ], + "lines": [ + { + "bbox": [ + 306, + 145, + 545, + 168 + ], + "spans": [ + { + "bbox": [ + 306, + 145, + 545, + 168 + ], + "type": "text", + "content": "Table 1. FID [19] scores on VizDoom and Replica. NF-LDM outperforms all baseline models. Baseline numbers are from [2]." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 325, + 175, + 525, + 232 + ], + "lines": [ + { + "bbox": [ + 325, + 175, + 525, + 232 + ], + "spans": [ + { + "bbox": [ + 325, + 175, + 525, + 232 + ], + "type": "table", + "html": "
CriterionMethodDepthCarlaAVD
FID (↓)EG3D [5]×76.89194.34
GSN [8]✓75.45166.07
NF-LDM✓35.6954.26
", + "image_path": "dd519d0099e98d4851b80bcae57e754b5e391f5482462a985bf75d9cc1ab13d9.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 234, + 545, + 266 + ], + "lines": [ + { + "bbox": [ + 306, + 234, + 545, + 266 + ], + "spans": [ + { + "bbox": [ + 306, + 234, + 545, + 266 + ], + "type": "text", + "content": "Table 2. FID [19] scores on Carla and AVD datasets. Baseline models have trouble learning the distribution of complex outdoor datasets, in particular AVD, while NF-LDM models them well." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 317, + 270, + 531, + 520 + ], + "blocks": [ + { + "bbox": [ + 317, + 270, + 531, + 520 + ], + "lines": [ + { + "bbox": [ + 317, + 270, + 531, + 520 + ], + "spans": [ + { + "bbox": [ + 317, + 270, + 531, + 520 + ], + "type": "image", + "image_path": "bc961cb0f79b63395ccc524c7e7dd1fb29c581b3531c929804c3e55a4306d39b.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 523, + 544, + 535 + ], + "lines": [ + { + "bbox": [ + 306, + 523, + 544, + 535 + ], + "spans": [ + { + "bbox": [ + 306, + 523, + 544, + 535 + ], + "type": "text", + "content": "Figure 8. Reconstructing held-out scenes not seen during training." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": "optimization step (Sec. 3.4), unless specified. Tab. 1 shows the results on VizDoom and Replica. GRAF [59] and " + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": "-GAN [4], which do not utilize ground truth depth in training, have shown successes in modelling single objects, but they exhibit worse performance than others that leverages depth information for modelling scenes. GAUDI [2] is an auto-decoder-based diffusion model. Their auto-decoder optimizes a small per-scene latent to reconstruct its matching scene. GAUDI comes with the advantage that learning the generative model is simple as it only needs to model the small dimensional latent distribution that acts as the key to their corresponding scenes. On the contrary, NF-LDM is trained on the latents that are a decomposition of the explicit 3D neural field. Therefore, GAUDI puts more modelling" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "8501" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 67, + 70, + 266, + 125 + ], + "blocks": [ + { + "bbox": [ + 67, + 70, + 266, + 125 + ], + "lines": [ + { + "bbox": [ + 67, + 70, + 266, + 125 + ], + "spans": [ + { + "bbox": [ + 67, + 70, + 266, + 125 + ], + "type": "table", + "html": "
CriterionMethodDepthCarlaAVD
FVD (↓)EG3D [5]X134.941232.38
GSN [8]✓184.301659.81
NF-LDM✓91.80242.50
", + "image_path": "1d0a529dc27450600a0621064058a0cc84837ec3f839377a72b62cc6342bd5d4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 168, + 287, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 168, + 287, + 264 + ], + "spans": [ + { + "bbox": [ + 46, + 168, + 287, + 264 + ], + "type": "text", + "content": "capacity into the auto-decoder part, and NF-LDM puts it more into the generative model part. We attribute our improvement over GAUDI to our expressive hierarchical LDM that can model the details of the scenes better. In VizDoom, only one scene exists, and each sequence contains several hundred steps covering a large area in the scene, which our voxels were not large enough to encode. Therefore, we chunked each VizDoom trajectory to be 50 steps long." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 266, + 287, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 287, + 361 + ], + "type": "text", + "content": "Tab. 2 shows results on complex outdoor datasets: Carla and AVD. We compare with EG3D [5] and GSN [8]. Both are GAN-based 3D generative models, but GSN utilizes ground truth depth measurements. Note that we did not include GAUDI [2] as the code was not available. NF-LDM achieves the best performance, and both baseline models have difficulty modelling the real outdoor dataset (AVD). Fig. 6 compares the samples from different models." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 363, + 287, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 363, + 287, + 543 + ], + "spans": [ + { + "bbox": [ + 46, + 363, + 287, + 543 + ], + "type": "text", + "content": "Since the datasets are composed of frame sequences, we can treat them as videos and further evaluate with Fréchet Video Distance (FVD) [66] to compare the distributions of the dataset and sampled sequences. This can quantify samples' 3D structure by how natural the rendered sequence from a moving camera is. For EG3D and GSN, we randomly sample a trajectory from the datasets and for NF-LDM, we sample a trajectory from the global latent diffusion model. Tab. 3 shows that NF-LDM achieves the best results. We empirically observed GSN sometimes produced slightly inconsistent rendering, which could attribute to its lower FVD score than EG3D's. We also visualize the geometry of NF-LDM's samples by running marching-cubes [38] on the density voxels. Fig. 7 shows that our samples produce a coarse but realistic geometry." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 544, + 287, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 544, + 287, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 544, + 287, + 652 + ], + "type": "text", + "content": "Ablations We evaluate the hierarchical structure of NF-LDM. Tab. 4 shows an ablation study on Carla. The model with the full hierarchy achieves the best performance. The global latent makes it easier for the other LDMs to sample as conditioning on the global properties of the scene (e.g. time of day) narrows down the distribution they need to model. The 2D fine latent helps retain the residual information missing in the 3D coarse latent, thus improving the latent auto-encoder and, consequently, the LDMs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "text", + "content": "Scene Reconstruction Unlike previous approaches, NF-LDM has an explicit scene auto-encoder that can be used for scene reconstruction. GAUDI [2] is auto-decoder based, so it is not trivial to infer a latent for a new scene. GSN [8] can invert a new scene using a GAN inversion method [52," + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 306, + 68, + 545, + 150 + ], + "blocks": [ + { + "bbox": [ + 306, + 68, + 545, + 150 + ], + "lines": [ + { + "bbox": [ + 306, + 68, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 306, + 68, + 545, + 150 + ], + "type": "image", + "image_path": "f3c51d533f81ba0ba02a18e1db53a31f2109d36dce5e90c3ae8dd46b9c47178b.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 306, + 155, + 545, + 220 + ], + "blocks": [ + { + "bbox": [ + 306, + 155, + 545, + 220 + ], + "lines": [ + { + "bbox": [ + 306, + 155, + 545, + 220 + ], + "spans": [ + { + "bbox": [ + 306, + 155, + 545, + 220 + ], + "type": "image", + "image_path": "372c91bb060305a99c8da1b2ebfd42c8d16a6925b4c6cc3fa0e4c07c0768bf4e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 306, + 224, + 545, + 288 + ], + "blocks": [ + { + "bbox": [ + 306, + 224, + 545, + 288 + ], + "lines": [ + { + "bbox": [ + 306, + 224, + 545, + 288 + ], + "spans": [ + { + "bbox": [ + 306, + 224, + 545, + 288 + ], + "type": "image", + "image_path": "e460ab0240c0a8197bc6f8243058061d111636897779ccd35fde311b785be1c6.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 292, + 546, + 346 + ], + "lines": [ + { + "bbox": [ + 305, + 292, + 546, + 346 + ], + "spans": [ + { + "bbox": [ + 305, + 292, + 546, + 346 + ], + "type": "text", + "content": "Figure 9. BEV-Conditioned Synthesis: NF-LDM allows controllable generation by editing the BEV segmentation map. From the initial sample, we add trees (green) and then edit the location of the car (blue). Note the ego car is at the center and thus not rendered." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 325, + 351, + 525, + 381 + ], + "blocks": [ + { + "bbox": [ + 47, + 128, + 287, + 150 + ], + "lines": [ + { + "bbox": [ + 47, + 128, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 47, + 128, + 287, + 150 + ], + "type": "text", + "content": "Table 3. FVD [66] scores on Carla and AVD Datasets. As for FID, baseline models have trouble learning to model complex datasets." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 325, + 351, + 525, + 381 + ], + "lines": [ + { + "bbox": [ + 325, + 351, + 525, + 381 + ], + "spans": [ + { + "bbox": [ + 325, + 351, + 525, + 381 + ], + "type": "table", + "html": "
Coarse lat. c+ Fine lat. f+ Global lat. g
FID (↓)46.4343.5235.69
", + "image_path": "5a6e738645d16aefa1b7a951c7fc110ab868a74ee9375cae6052cc7f7c293017.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 382, + 545, + 415 + ], + "lines": [ + { + "bbox": [ + 305, + 382, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 305, + 382, + 545, + 415 + ], + "type": "text", + "content": "Table 4. FID [19] on ablating the choice of hierarchy on the Carla dataset. The first column is for training both LAE and LDM only with the coarse latent. The last column is our full model." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 432, + 545, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 432, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 305, + 432, + 545, + 469 + ], + "type": "text", + "content": "79], but as Fig. 8 shows, it fails to get the details of the scene correct. Our scene auto-encoder generalizes well and is scalable as the number of scenes grow." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 478, + 465, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 478, + 465, + 491 + ], + "spans": [ + { + "bbox": [ + 306, + 478, + 465, + 491 + ], + "type": "text", + "content": "4.2. Applications and Limitations" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 497, + 545, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 497, + 545, + 593 + ], + "spans": [ + { + "bbox": [ + 304, + 497, + 545, + 593 + ], + "type": "text", + "content": "Conditional Synthesis NF-LDM can utilize additional conditioning signals for controllable generation. In this paper, we consider Bird's Eye View (BEV) segmentation maps, but our model can be extended to other conditioning variables. We use cross attention layers [53], which have been shown to be effective for conditional synthesis. Fig. 9 shows that NF-LDM follows the given BEV map faithfully and how the map can be edited for controllable synthesis." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "content": "Scene Editing Image diffusion models can be used for image inpainting without explicit training on the task [53, 63]. We leverage this property to edit scenes in 3D by resampling a region in the 3D coarse latent " + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "content": ". Specifically, at each denoising step, we noise the region to be kept and concatenate with the region being sampled, and pass it through the diffusion model. We use reconstruction guidance [24] to better harmonize the sampled and kept regions. After we get a new " + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "content": ", the fine latent is also re-sampled conditioned on " + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 594, + 546, + 713 + ], + "type": "text", + "content": ". Fig. 11 shows results on scene editing with NF-LDM." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8502" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 85, + 70, + 509, + 277 + ], + "blocks": [ + { + "bbox": [ + 85, + 70, + 509, + 277 + ], + "lines": [ + { + "bbox": [ + 85, + 70, + 509, + 277 + ], + "spans": [ + { + "bbox": [ + 85, + 70, + 509, + 277 + ], + "type": "image", + "image_path": "f6f3e6e358c5a07224b7f0a5b202839d588539fd0048e6c5e6c9587662c45bf8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 277, + 547, + 300 + ], + "lines": [ + { + "bbox": [ + 46, + 277, + 547, + 300 + ], + "spans": [ + { + "bbox": [ + 46, + 277, + 547, + 300 + ], + "type": "text", + "content": "Figure 10. Panoramas from NF-LDM's samples: From the initial sample at the top, we apply post-optimization with Score Distillation Sampling [46] (Sec. 3.4). (a) demonstrates improved sample quality. (b) showcases style modification by conditioning on evening scenes." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 50, + 308, + 287, + 575 + ], + "blocks": [ + { + "bbox": [ + 50, + 308, + 287, + 575 + ], + "lines": [ + { + "bbox": [ + 50, + 308, + 287, + 575 + ], + "spans": [ + { + "bbox": [ + 50, + 308, + 287, + 575 + ], + "type": "image", + "image_path": "97bdf0f0cb0156bf24f2cc3015518c5a49502b30d7eb93a34bd500039ac93a1d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 578, + 288, + 613 + ], + "lines": [ + { + "bbox": [ + 46, + 578, + 288, + 613 + ], + "spans": [ + { + "bbox": [ + 46, + 578, + 288, + 613 + ], + "type": "text", + "content": "Figure 11. Scene Editing: We use the 3D coarse latent " + }, + { + "bbox": [ + 46, + 578, + 288, + 613 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 578, + 288, + 613 + ], + "type": "text", + "content": " for scene editing. From the initial sample indicated by light green, we resample a part of the latent, indicated by dark green." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 629, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 289, + 715 + ], + "type": "text", + "content": "Post-Optimization Fig. 10 shows how post-optimization (Sec. 3.4) can improve the quality of NF-LDM's initial sample while retaining the 3D structure. In addition to improving quality, we can also modify scene properties, such as time of day and weather, by conditioning the LDM on images with the desired properties. SDS-based style modification is effective for changes where a set of" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 312, + 545, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 312, + 545, + 372 + ], + "spans": [ + { + "bbox": [ + 304, + 312, + 545, + 372 + ], + "type": "text", + "content": "clean image data is available with the desired property and is reasonably close to our dataset's domain (e.g. street images for AVD). In the supplementary, we also provide results experimenting with directional CLIP loss [14] to quickly finetune our scene decoder for a given text prompt." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 372, + 546, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 372, + 546, + 538 + ], + "spans": [ + { + "bbox": [ + 304, + 372, + 546, + 538 + ], + "type": "text", + "content": "Limitations NF-LDM's hierarchical structure and three stage pipeline allows us to achieve high-quality generations and reconstructions, but it comes with a degradation in training time and sampling speed. In this work, the neural field representation is based on dense voxel grids, and it becomes expensive to volume render and learn the diffusion models as they get larger. Therefore, exploring alternative sparse representations is a promising future direction. Lastly, our method requires multi-view images which limits data availability and therefore risks universal problems in generative modelling of overfitting. For example, we found that output samples in AVD had limited diversity because the dataset itself was recorded in a limited number of scenes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 549, + 379, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 549, + 379, + 562 + ], + "spans": [ + { + "bbox": [ + 306, + 549, + 379, + 562 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 715 + ], + "type": "text", + "content": "We introduced NeuralField-LDM (NF-LDM), a generative model for complex 3D environments. NF-LDM first constructs an expressive latent distribution by encoding input images into a 3D neural field representation which is further compressed into more abstract latent spaces. Then, our proposed hierarchical LDM is fit onto the latent spaces, achieving state-of-the-art performance on 3D scene generation. NF-LDM enables a diverse set of applications, including controllable scene generation and scene editing. Future directions include exploring more efficient sparse voxel representations, training on larger-scale real-world data and learning to continuously expand generated scenes." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "text", + "content": "8503" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "text", + "content": "[1] Martin Arjovsky and Léon Bottou. Towards principled methods for training generative adversarial networks. In ICLR, 2017. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 125, + 288, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 288, + 180 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 288, + 180 + ], + "type": "text", + "content": "[2] Miguel Ángel Bautista, Pengsheng Guo, Samira Abnar, Walter Talbott, Alexander Toshev, Zhuoyuan Chen, Laurent Dinh, Shuangfei Zhai, Hanlin Goh, Daniel Ulbricht, Afshin Dehghan, and Josh M. Susskind. GAUDI: A neural architect for immersive 3d scene generation. 2022. 1, 2, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 182, + 287, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 182, + 287, + 214 + ], + "spans": [ + { + "bbox": [ + 53, + 182, + 287, + 214 + ], + "type": "text", + "content": "[3] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale GAN training for high fidelity natural image synthesis. In ICLR, 2019. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 216, + 286, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 216, + 286, + 258 + ], + "spans": [ + { + "bbox": [ + 53, + 216, + 286, + 258 + ], + "type": "text", + "content": "[4] Eric Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. arXiv, 2012.00926, 2020. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 261, + 286, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 261, + 286, + 316 + ], + "spans": [ + { + "bbox": [ + 53, + 261, + 286, + 316 + ], + "type": "text", + "content": "[5] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In arXiv, 2021. 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 318, + 286, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 318, + 286, + 372 + ], + "spans": [ + { + "bbox": [ + 53, + 318, + 286, + 372 + ], + "type": "text", + "content": "[6] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In CVPR, 2022. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 373, + 286, + 416 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 373, + 286, + 416 + ], + "spans": [ + { + "bbox": [ + 53, + 373, + 286, + 416 + ], + "type": "text", + "content": "[7] Eric R. Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. Pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In CVPR, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 418, + 286, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 418, + 286, + 462 + ], + "spans": [ + { + "bbox": [ + 53, + 418, + 286, + 462 + ], + "type": "text", + "content": "[8] Terrance DeVries, Miguel Angel Bautista, Nitish Srivastava, Graham W. Taylor, and Joshua M. Susskind. Unconstrained scene generation with locally conditioned radiance fields. In ICCV, 2021. 1, 2, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 464, + 286, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 464, + 286, + 506 + ], + "spans": [ + { + "bbox": [ + 53, + 464, + 286, + 506 + ], + "type": "text", + "content": "[9] Terrance DeVries, Miguel Angel Bautista, Nitish Srivastava, Graham W. Taylor, and Joshua M. Susskind. Unconstrained scene generation with locally conditioned radiance fields. arXiv, 2021. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 509, + 286, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 509, + 286, + 552 + ], + "spans": [ + { + "bbox": [ + 48, + 509, + 286, + 552 + ], + "type": "text", + "content": "[10] Prafulla Dhariwal and Alexander Quinn Nichol. Diffusion models beat gans on image synthesis. In Marc'Aurelio Ranzato, Alina Beygelzimer, Yann N. Dauphin, Percy Liang, and Jennifer Wortman Vaughan, editors, NeurIPS, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 554, + 286, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 554, + 286, + 586 + ], + "spans": [ + { + "bbox": [ + 48, + 554, + 286, + 586 + ], + "type": "text", + "content": "[11] Tim Dockhorn, Arash Vahdat, and Karsten Kreis. Score-based generative modeling with critically-damped Langevin diffusion. In ICLR, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 588, + 286, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 588, + 286, + 620 + ], + "spans": [ + { + "bbox": [ + 48, + 588, + 286, + 620 + ], + "type": "text", + "content": "[12] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. CARLA: An open urban driving simulator. 2017. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 622, + 286, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 622, + 286, + 654 + ], + "spans": [ + { + "bbox": [ + 48, + 622, + 286, + 654 + ], + "type": "text", + "content": "[13] Patrick Esser, Robin Rombach, and Björn Ommer. Taming transformers for high-resolution image synthesis. In CVPR, 2021. 2, 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 657, + 286, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 657, + 286, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 657, + 286, + 689 + ], + "type": "text", + "content": "[14] Rinon Gal, Or Patashnik, Haggai Maron, Gal Chechik, and Daniel Cohen-Or. Stylegan-nada: Clip-guided domain adaptation of image generators, 2021. 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 691, + 286, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 286, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 286, + 714 + ], + "type": "text", + "content": "[15] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "text", + "content": "Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. In Advances In Neural Information Processing Systems, 2022. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 107, + 545, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 107, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 308, + 107, + 545, + 129 + ], + "type": "text", + "content": "[16] Ian Goodfellow. Nips 2016 tutorial: Generative adversarial networks. arXiv preprint arXiv:1701.00160, 2016. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 130, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 130, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 308, + 130, + 545, + 173 + ], + "type": "text", + "content": "[17] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron C. Courville, and Yoshua Bengio. Generative adversarial nets. In NeurIPS, 2014. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 175, + 545, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 175, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 308, + 175, + 545, + 207 + ], + "type": "text", + "content": "[18] William Harvey, Saeid Naderiparizi, Vaden Masrani, Christian Weilbach, and Frank Wood. Flexible diffusion modeling of long videos, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 209, + 545, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 209, + 545, + 252 + ], + "spans": [ + { + "bbox": [ + 308, + 209, + 545, + 252 + ], + "type": "text", + "content": "[19] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 254, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 254, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 308, + 254, + 545, + 297 + ], + "type": "text", + "content": "[20] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Hugo Larochelle, Marc' Aurelio Ranzato, Raia Hadsell, Maria-Florina Balcan, and Hsuan-Tien Lin, editors, NeurIPS, 2020. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 298, + 545, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 298, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 308, + 298, + 545, + 353 + ], + "type": "text", + "content": "[21] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 6840-6851. Curran Associates, Inc., 2020. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 354, + 545, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 354, + 545, + 397 + ], + "spans": [ + { + "bbox": [ + 308, + 354, + 545, + 397 + ], + "type": "text", + "content": "[22] Jonathan Ho, Chitwan Sahara, William Chan, David J. Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. J. Mach. Learn. Res., 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 399, + 545, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 399, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 308, + 399, + 545, + 422 + ], + "type": "text", + "content": "[23] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 422, + 545, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 422, + 545, + 454 + ], + "spans": [ + { + "bbox": [ + 308, + 422, + 545, + 454 + ], + "type": "text", + "content": "[24] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. arXiv:2204.03458, 2022. 2, 4, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 456, + 545, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 456, + 545, + 488 + ], + "spans": [ + { + "bbox": [ + 308, + 456, + 545, + 488 + ], + "type": "text", + "content": "[25] James T Kajiya and Brian P Von Herzen. Ray tracing volume densities. ACM SIGGRAPH computer graphics, 18(3):165-174, 1984. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 490, + 545, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 490, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 308, + 490, + 545, + 533 + ], + "type": "text", + "content": "[26] Amlan Kar, Aayush Prakash, Ming-Yu Liu, Eric Cameracci, Justin Yuan, Matt Rusiniak, David Acuna, Antonio Torralba, and Sanja Fidler. Meta-sim: Learning to generate synthetic datasets. In ICCV, 2019. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 535, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 535, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 308, + 535, + 545, + 567 + ], + "type": "text", + "content": "[27] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. In CVPR, 2020. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 569, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 569, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 569, + 545, + 624 + ], + "type": "text", + "content": "[28] Michal Kempka, Marek Wydmuch, Grzegorz Runc, Jakub Toczek, and Wojciech Jaskowski. Vizdoom: A doom-based ai research platform for visual reinforcement learning. In 2016 IEEE conference on computational intelligence and games (CIG), pages 1-8. IEEE, 2016. 5" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 624, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 624, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 308, + 624, + 545, + 657 + ], + "type": "text", + "content": "[29] Margret Keuper, Siyu Tang, Bjoern Andres, Thomas Brox, and Bernt Schiele. Motion segmentation & multiple object tracking by correlation co-clustering. IEEE TPAMI, 2018. 5" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "type": "text", + "content": "[30] Seung Wook Kim, Karsten Kreis, Daiqing Li, Antonio Torralba, and Sanja Fidler. Polymorphic-gan: Generating aligned samples across multiple domains with learned morph maps. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Jun. 2022. 2" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8504" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[31] Seung Wook Kim, Jonah Philion, Antonio Torralba, and Sanja Fidler. Drivegan: Towards a controllable high-quality neural simulation. In CVPR, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 288, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 288, + 163 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 288, + 163 + ], + "type": "text", + "content": "[32] Seung Wook Kim, Yuhao Zhou, Jonah Philion, Antonio Torralba, and Sanja Fidler. Learning to simulate dynamic environments with gamean. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1231-1240, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 164, + 288, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 164, + 288, + 186 + ], + "spans": [ + { + "bbox": [ + 48, + 164, + 288, + 186 + ], + "type": "text", + "content": "[33] Diederik P. Kingma and Max Welling. Auto-encoding variational bayes. *ICLR*, 2014. 2, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 188, + 288, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 188, + 288, + 232 + ], + "spans": [ + { + "bbox": [ + 48, + 188, + 288, + 232 + ], + "type": "text", + "content": "[34] Daiqing Li, Huan Ling, Seung Wook Kim, Karsten Kreis, Adela Barriuso, Sanja Fidler, and Antonio Torralba. Big-datasetgan: Synthesizing imagenet with pixel-wise annotations. arXiv, 2201.04684, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 233, + 288, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 233, + 288, + 255 + ], + "spans": [ + { + "bbox": [ + 48, + 233, + 288, + 255 + ], + "type": "text", + "content": "[35] Ke Li and Jitendra Malik. On the implicit assumptions of gans. arXiv preprint arXiv:1811.12402, 2018. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 257, + 288, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 257, + 288, + 290 + ], + "spans": [ + { + "bbox": [ + 48, + 257, + 288, + 290 + ], + "type": "text", + "content": "[36] Yiyi Liao, Katja Schwarz, Lars Mescheder, and Andreas Geiger. Towards unsupervised learning of generative models for 3d controllable image synthesis. In CVPR, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 292, + 288, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 292, + 288, + 324 + ], + "spans": [ + { + "bbox": [ + 48, + 292, + 288, + 324 + ], + "type": "text", + "content": "[37] Huan Ling, Karsten Kreis, Daiqing Li, Seung Wook Kim, Antonio Torralba, and Sanja Fidler. Editgan: High-precision semantic image editing. In NeurIPS, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 326, + 288, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 326, + 288, + 358 + ], + "spans": [ + { + "bbox": [ + 48, + 326, + 288, + 358 + ], + "type": "text", + "content": "[38] William E. Lorensen and Harvey E. Cline. Marching cubes: A high resolution 3d surface construction algorithm. In ACM Trans. on Graphics, 1987. 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 361, + 288, + 383 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 361, + 288, + 383 + ], + "spans": [ + { + "bbox": [ + 48, + 361, + 288, + 383 + ], + "type": "text", + "content": "[39] Shitong Luo and Wei Hu. Diffusion probabilistic models for 3d point cloud generation. In CVPR, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 384, + 288, + 428 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 384, + 288, + 428 + ], + "spans": [ + { + "bbox": [ + 48, + 384, + 288, + 428 + ], + "type": "text", + "content": "[40] Ricardo Martin-Brualla, Noha Radwan, Mehdi S. M. Sajjadi, Jonathan T. Barron, Alexey Dosovitskiy, and Daniel Duckworth. NeRF in the Wild: Neural Radiance Fields for Unconstrained Photo Collections. In CVPR, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 430, + 288, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 430, + 288, + 473 + ], + "spans": [ + { + "bbox": [ + 48, + 430, + 288, + 473 + ], + "type": "text", + "content": "[41] Lars Mescheder, Andreas Geiger, and Sebastian Nowozin. Which training methods for gans do actually converge? In Proc. of the International Conf. on Machine learning (ICML), 2018. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 475, + 288, + 519 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 475, + 288, + 519 + ], + "spans": [ + { + "bbox": [ + 48, + 475, + 288, + 519 + ], + "type": "text", + "content": "[42] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 521, + 288, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 521, + 288, + 609 + ], + "spans": [ + { + "bbox": [ + 48, + 521, + 288, + 609 + ], + "type": "text", + "content": "[43] Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. GLIDE: towards photorealistic image generation and editing with text-guided diffusion models. In Kamalika Chaudhuri, Stefanie Jegelka, Le Song, Csaba Szepesvári, Gang Niu, and Sivan Sabato, editors, Proc. of the International Conf. on Machine learning (ICML), 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 611, + 288, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 611, + 288, + 643 + ], + "spans": [ + { + "bbox": [ + 48, + 611, + 288, + 643 + ], + "type": "text", + "content": "[44] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In CVPR, 2021. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 646, + 288, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 646, + 288, + 677 + ], + "spans": [ + { + "bbox": [ + 48, + 646, + 288, + 677 + ], + "type": "text", + "content": "[45] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In ECCV, 2020. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 680, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 288, + 712 + ], + "type": "text", + "content": "[46] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv, 2022. 1, 2, 5, 8" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "text", + "content": "[47] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. arXiv, 2103.00020, 2021. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 129, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 161 + ], + "type": "text", + "content": "[48] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with CLIP latents. arXiv, abs/2204.06125, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 163, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 195 + ], + "type": "text", + "content": "[49] Ali Razavi, Aaron van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with VQ-VAE-2. In NeurIPS, 2019. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 197, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 239 + ], + "type": "text", + "content": "[50] Konstantinos Rematas, Andrew Liu, Pratul P. Srinivasan, Jonathan T. Barron, Andrea Tagliasacchi, Thomas A. Funkhouser, and Vittorio Ferrari. Urban radiance fields. In CVPR, 2022. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 241, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 241, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 545, + 285 + ], + "type": "text", + "content": "[51] Danilo Jimenez Rezende, Shakir Mohamed, and Daan Wierstra. Stochastic backpropagation and approximate inference in deep generative models. In Proc. of the International Conf. on Machine learning (ICML), 2014. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 286, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 286, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 307, + 286, + 545, + 342 + ], + "type": "text", + "content": "[52] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: a stylegan encoder for image-to-image translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2287–2296, 2021. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 343, + 545, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 343, + 545, + 386 + ], + "spans": [ + { + "bbox": [ + 307, + 343, + 545, + 386 + ], + "type": "text", + "content": "[53] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 3, 4, 5, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 388, + 545, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 545, + 420 + ], + "type": "text", + "content": "[54] Robin Rombach, Patrick Esser, and Björn Ommer. Geometry-free view synthesis: Transformers and no 3d priors. In ICCV, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 422, + 545, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 422, + 545, + 476 + ], + "spans": [ + { + "bbox": [ + 307, + 422, + 545, + 476 + ], + "type": "text", + "content": "[55] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234-241. Springer, 2015. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 478, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 478, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 307, + 478, + 545, + 544 + ], + "type": "text", + "content": "[56] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 3, 4, 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 544, + 545, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 544, + 545, + 577 + ], + "spans": [ + { + "bbox": [ + 307, + 544, + 545, + 577 + ], + "type": "text", + "content": "[57] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. arXiv preprint arXiv:2202.00512, 2022. 3, 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 578, + 545, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 578, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 307, + 578, + 545, + 611 + ], + "type": "text", + "content": "[58] Axel Sauer, Katja Schwarz, and Andreas Geiger. StyleganxI: Scaling stylegan to large diverse datasets. ACM Trans. on Graphics, 2022. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 613, + 545, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 545, + 644 + ], + "type": "text", + "content": "[59] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. Graf: Generative radiance fields for 3d-aware image synthesis. 2020. 2, 6" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 647, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 679 + ], + "type": "text", + "content": "[60] Katja Schwarz, Axel Sauer, Michael Niemeyer, Yiyi Liao, and Andreas Geiger. Voxgraf: Fast 3d-aware image synthesis with sparse voxel grids. In NeurIPS, 2022. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 681, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 681, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 681, + 545, + 712 + ], + "type": "text", + "content": "[61] Yujun Shen, Jinjin Gu, Xiaou Tang, and Bolei Zhou. Interpreting the latent space of gans for semantic face editing. In CVPR, 2020. 2" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "8505" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[62] Yuan Shen, Wei-Chiu Ma, and Shenlong Wang. SGAM: Building a virtual 3d world through simultaneous generation and mapping. In Thirty-Sixth Conference on Neural Information Processing Systems, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 287, + 162 + ], + "type": "text", + "content": "[63] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 2, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 163, + 288, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 163, + 288, + 272 + ], + "spans": [ + { + "bbox": [ + 49, + 163, + 288, + 272 + ], + "type": "text", + "content": "[64] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J. Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, Anton Clarkson, Mingfei Yan, Brian Budge, Yajie Yan, Xiaqing Pan, June Yon, Yuyang Zou, Kimberly Leon, Nigel Carter, Jesus Briales, Tyler Gillingham, Elias Mueggler, Luis Pesqueira, Manolis Savva, Dhruv Batra, Hauke M. Strasdat, Renzo De Nardi, Michael Goesele, Steven Lovegrove, and Richard Newcombe. The Replica dataset: A digital replica of indoor spaces. arXiv, 1906.05797, 2019. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 274, + 287, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 274, + 287, + 306 + ], + "spans": [ + { + "bbox": [ + 49, + 274, + 287, + 306 + ], + "type": "text", + "content": "[65] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. CVPR, 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 308, + 287, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 308, + 287, + 351 + ], + "spans": [ + { + "bbox": [ + 49, + 308, + 287, + 351 + ], + "type": "text", + "content": "[66] Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 353, + 287, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 353, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 49, + 353, + 287, + 396 + ], + "type": "text", + "content": "[67] Arash Vahdat and Jan Kautz. NVAE: A deep hierarchical variational autoencoder. In Hugo Larochelle, Marc'Aurelio Ranzato, Raia Hadsell, Maria-Florina Balcan, and Hsuan-Tien Lin, editors, NeurIPS, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 397, + 287, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 397, + 287, + 419 + ], + "spans": [ + { + "bbox": [ + 49, + 397, + 287, + 419 + ], + "type": "text", + "content": "[68] Arash Vahdat, Karsten Kreis, and Jan Kautz. Score-based generative modeling in latent space. In NeurIPS, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 421, + 287, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 421, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 49, + 421, + 287, + 453 + ], + "type": "text", + "content": "[69] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 455, + 287, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 455, + 287, + 508 + ], + "spans": [ + { + "bbox": [ + 49, + 455, + 287, + 508 + ], + "type": "text", + "content": "[70] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P. Srinivasan, Howard Zhou, Jonathan T. Barron, Ricardo Martin-Brualla, Noah Snavely, and Thomas A. Funkhouser. Ibrnet: Learning multi-view image-based rendering. In CVPR, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 510, + 287, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 510, + 287, + 543 + ], + "spans": [ + { + "bbox": [ + 49, + 510, + 287, + 543 + ], + "type": "text", + "content": "[71] Yuxin Wu and Kaiming He. Group normalization. In Proceedings of the European conference on computer vision (ECCV), pages 3-19, 2018. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 544, + 287, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 544, + 287, + 576 + ], + "spans": [ + { + "bbox": [ + 49, + 544, + 287, + 576 + ], + "type": "text", + "content": "[72] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelNeRF: Neural radiance fields from one or few images. In CVPR, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 578, + 287, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 578, + 287, + 620 + ], + "spans": [ + { + "bbox": [ + 49, + 578, + 287, + 620 + ], + "type": "text", + "content": "[73] Xiaohui Zeng, Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, and Karsten Kreis. LION: latent point diffusion models for 3d shape generation. In NeurIPS, 2022. 1, 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 624, + 287, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 624, + 287, + 655 + ], + "spans": [ + { + "bbox": [ + 49, + 624, + 287, + 655 + ], + "type": "text", + "content": "[74] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv, 2010.07492, 2020. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 658, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 658, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 49, + 658, + 287, + 689 + ], + "type": "text", + "content": "[75] Xiaoshuai Zhang, Sai Bi, Kalyan Sunkavalli, Hao Su, and Zexiang Xu. Nerfusion: Fusing radiance fields for large-scale scene reconstruction. In CVPR, 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 49, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 49, + 691, + 287, + 713 + ], + "type": "text", + "content": "[76] Yuxuan Zhang, Wenzheng Chen, Huan Ling, Jun Gao, Yi-nan Zhang, Antonio Torralba, and Sanja Fidler. Image gans" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 206 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "meet differentiable rendering for inverse graphics and interpretable 3d neural rendering. In ICLR, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 96, + 545, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 129 + ], + "type": "text", + "content": "[77] Yan Zheng, Lemeng Wu, Xingchao Liu, Zhen Chen, Qiang Liu, and Qixing Huang. Neural volumetric mesh generator. arXiv preprint arXiv:2210.03158, 2022. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 130, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 130, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 308, + 130, + 545, + 161 + ], + "type": "text", + "content": "[78] Linqi Zhou, Yilun Du, and Jiajun Wu. 3d shape generation and completion through point-voxel diffusion. In ICCV, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 163, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 163, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 308, + 163, + 545, + 206 + ], + "type": "text", + "content": "[79] Jiapeng Zhu, Yujun Shen, Deli Zhao, and Bolei Zhou. Indomain gan inversion for real image editing. In European conference on computer vision, pages 592-608. Springer, 2020. 7" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "8506" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/NeuralLift-360_ Lifting an In-the-Wild 2D Photo to a 3D Object With 360deg Views/90dd78db-4103-415c-8ddc-ba0f83221385_content_list.json b/2023/NeuralLift-360_ Lifting an In-the-Wild 2D Photo to a 3D Object With 360deg Views/90dd78db-4103-415c-8ddc-ba0f83221385_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..03a76d084abc2c73ec95c6d3a5b4fce7fff5c8fb --- /dev/null +++ b/2023/NeuralLift-360_ Lifting an In-the-Wild 2D Photo to a 3D Object With 360deg Views/90dd78db-4103-415c-8ddc-ba0f83221385_content_list.json @@ -0,0 +1,1575 @@ +[ + { + "type": "text", + "text": "NeuralLift-360: Lifting An In-the-wild 2D Photo to A 3D Object with $360^{\\circ}$ Views", + "text_level": 1, + "bbox": [ + 78, + 130, + 893, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dejia Xu, Yifan Jiang, Peihao Wang, Zhiwen Fan, Yi Wang, Zhangyang Wang \nVITA Group, University of Texas at Austin", + "bbox": [ + 174, + 180, + 794, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{dejia,atlaswang}@utexas.edu", + "bbox": [ + 361, + 218, + 609, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Virtual reality and augmented reality (XR) bring increasing demand for 3D content generation. However, creating high-quality 3D content requires tedious work from a human expert. In this work, we study the challenging task of lifting a single image to a 3D object and, for the first time, demonstrate the ability to generate a plausible 3D object with $360^{\\circ}$ views that corresponds well with the given reference image. By conditioning on the reference image, our model can fulfill the everlasting curiosity for synthesizing novel views of objects from images. Our technique sheds light on a promising direction of easing the workflows for 3D artists and XR designers. We propose a novel framework, dubbed NeuralLift-360, that utilizes a depth-aware neural radiance representation (NeRF) and learns to craft the scene guided by denoising diffusion models. By introducing a ranking loss, our NeuralLift-360 can be guided with rough depth estimation in the wild. We also adopt a CLIP-guided sampling strategy for the diffusion prior to provide coherent guidance. Extensive experiments demonstrate that our NeuralLift-360 significantly outperforms existing state-of-the-art baselines. Project page: https://vita-group.github.io/NeuralLift-360/", + "bbox": [ + 76, + 300, + 473, + 632 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 662, + 209, + 679 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Creating 3D content has been a long-standing problem in computer vision. This problem enables various applications in game studios, home decoration, virtual reality, and augmented reality. Over the past few decades, the manual task has dominated real scenarios, which requires tedious professional expert modeling. Modern artists rely on special software tools (e.g., Blender, Maya3D, 3DS Max, etc.) and time-demanding manual adjustments to realize imaginations and transform them into virtual objects. Meanwhile, automatic 3D content creation pipelines serve as effective tools to facilitate human efforts. These pipelines typically capture hundreds of images and leverage multi-view stereo [71] to quickly model fine-grained virtual landscapes.", + "bbox": [ + 76, + 688, + 470, + 885 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "More recently, researchers have started aiming at a more", + "bbox": [ + 96, + 886, + 468, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ambitious goal, to create a 3D object from a single image [9, 10, 12, 21, 27, 51, 66, 68]. This enables broad applications since it greatly reduces the prerequisite to a minimal request. Existing approaches can be mainly categorized into two major directions. One line of work utilizes learning priors from large-scale datasets with multi-view images [10, 21, 51, 66, 68]. These approaches usually learn a conditional neural network to predict 3D information based on input images. However, due to their poor generalization ability, drastic performance degradation is observed when the testing image is out-of-distribution.", + "bbox": [ + 496, + 268, + 893, + 436 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Another direction constructs the pipeline on top of the depth estimation techniques [65]. Under the guidance of monocular depth estimation networks [43, 44], the 2D image is firstly back-projected to a 3D data format (e.g., point cloud or multi-plane image) and then re-projected into novel views. After that, advanced image inpainting techniques are then adopted to fill missing holes [52] produced during the projection. However, most of them can be highly affected by the quality of the estimated depth maps. Though LeRes [73] attempted to rectify the predicted depth by refining the projected 3D point cloud, their results do not generalize well to an arbitrary image in the wild. Overall, the aforementioned approaches are either adopted in limited scenarios (e.g. face or toy examples [4, 75]), or only produce limited viewing directions [8, 70] when being applied to scenes in the wild. Different from all these approaches, we focus on a more challenging task and for the first time, show promising results by lifting a single in-the-wild image into a 3D object with $360^{\\circ}$ novel views.", + "bbox": [ + 496, + 441, + 895, + 729 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Attracted by the dramatic progress of neural volumetric rendering on 3D reconstruction tasks, we consider building our framework based on Neural Radiance Fields (NeRFs) [35]. The original NeRF takes hundreds of training views and their camera poses as inputs to learn an implicit representation. Subsequent models dedicate tremendous efforts [8, 19, 70, 75] to apply NeRF to sparse training views. The most similar work [70] to our method proposes to optimize a NeRF using only a single image and its corresponding depth map. However, it renders limited views from a small range of angles, and the prerequisite of a high-quality", + "bbox": [ + 496, + 734, + 895, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "4479", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "depth map largely constrains its practical usage.", + "bbox": [ + 76, + 90, + 393, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the above issues, we propose a novel framework, coined as NeuralLift-360, which aims to ease the creation of 3D assets by building a bridge to convert diverse in-the-wild 2D photos to sophisticated 3D contents in $360^{\\circ}$ views and enable its automation. The major challenge in our work is that the content on the back side is hidden and hard to hallucinate. To tackle these hurdles, we consider the diffusion priors together with the monocular depth estimator as the cues for hallucination. Modern diffusion models [41,48,50] are trained on a massive dataset (e.g., 5B text-to-image pairs [24]). During inference time, they can generate impressive photorealistic photos based on simple text inputs. By adopting these learning priors with CLIP [40] guidance, NeuralLift-360 can generate plausible 3D consistent instances that correspond well to the given reference image while only requiring minimal additional input, the correct user prompts. Moreover, rather than simply taking the depth map from the pre-trained depth estimator as geometry supervision, we propose to use the relative ranking information from the rough depth during the training process. This simple strategy is observed to robustly mitigate geometry errors for depth estimations in the wild.", + "bbox": [ + 75, + 107, + 468, + 439 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions can be summarized as follows,", + "bbox": [ + 96, + 440, + 419, + 454 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Given a single image in the wild, we demonstrate promising results of them being lifted to 3D. We use NeRF as an effective scene representation and integrate prior knowledge from the diffusion model.", + "- We propose a CLIP-guided sampling strategy that effectively marries the prior knowledge from the diffusion model with the reference image.", + "- When the reference image is hard to describe exactly, we finetune the diffusion model on the single image while maintaining its ability to generate diverse contents to guide the NeRF training.", + "- We introduce scale-invariant depth supervision that uses the ranking information. This design alleviates the need for accurate multi-view consistent depth estimation and broadens the application of our algorithms." + ], + "bbox": [ + 96, + 472, + 468, + 742 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 76, + 758, + 225, + 776 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1.3D from Single Image", + "text_level": 1, + "bbox": [ + 76, + 785, + 281, + 801 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Inferring 3D worlds from a single image has been a popular direction over the past years. Previous works try to solve this problem by firstly adopting a monocular depth estimator (e.g., MiDaS [44]) to predict the 3D geometry and then using multi-plane images (MPI) [20, 52] or point cloud [36] to render artistic effects. Meanwhile, the missing", + "bbox": [ + 75, + 810, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "holes from novel views are inpainted by a pre-trained modern deep network [76]. While these approaches are powerful and efficient, the depth estimator could be unstable, and the inpainting techniques may produce artifacts that make the rendered images look fake. While our method also applies a monocular depth estimator as the geometry cues, we propose a depth ranking loss that can mitigate the instability of depth estimation in the wild. Later approaches [28, 75] train their model on a large-scale 3D assets dataset. During inference time, they directly predict a 3D object using a single image. However, their performance faces drastic degradation when being applied to an in-the-wild image, due to the domain gap. The most similar work to ours is SinNeRF [70], which trains a neural radiance field only on a single RGB image and its corresponding depth map. However, their approach requires a high-quality depth map for geometry guidance, and it can only render views from a small range of angles. Thanks to the proposed depth ranking loss, our method can reconstruct impressive geometry information only using guidance from a pre-trained monocular depth estimator.", + "bbox": [ + 496, + 90, + 890, + 407 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Neural Radiance Fields", + "text_level": 1, + "bbox": [ + 500, + 416, + 715, + 431 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Neural Radiance Fields (NeRFs) [34] have shown encouraging performance for view synthesis. After its origin, tremendous efforts have been devoted to improving it [2,3,5,6,11,14,16,18,31,37,45,57,59,61,63,67,75]. Unlike most NeRF-like models, which at least require several different views to reconstruct scenes, we focus on reconstructing 3D content with single views. Analogously, recent works [18, 39] try to generate 3D content from text inputs directly. Although they also rely on the text prompt as guidance, their rendered objects are only conditioned on text sentences without image constraints, which can not reach our goal. Different from them, we aim at lifting a single image (either real or synthetic) to photorealistic 3D objects.", + "bbox": [ + 496, + 440, + 890, + 636 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Preliminaries", + "text_level": 1, + "bbox": [ + 500, + 648, + 637, + 664 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Neural Radiance Field NeRFs [34] render images by sampling 5D coordinates (location $(x,y,z)$ and viewing direction $(\\theta ,\\phi)$ ) along camera rays. After mapping them to color $(r,g,b)$ and volume density $\\sigma$ , additional volumetric rendering techniques are adopted to fuse the information along the ray. Given $\\mathbf{o}$ as the camera origin and $\\pmb{d}$ as the ray direction, 3D point locations along the ray can be expressed as $\\pmb {r}(t) = \\pmb {o} + td$ The predicted color accumulated is defined as follows:", + "bbox": [ + 496, + 674, + 890, + 809 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {C} (\\boldsymbol {r}) = \\int_ {t _ {n}} ^ {t _ {f}} T (t) \\sigma (\\boldsymbol {r} (t)) \\boldsymbol {c} (\\boldsymbol {r} (t), \\boldsymbol {d}) d t, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 816, + 890, + 853 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $T(t) = \\exp \\left(-\\int_{t_n}^t\\sigma (\\boldsymbol {r}(s))ds\\right)$ , $\\sigma (\\cdot)$ and $c(\\cdot ,\\cdot)$ are densities and color predictions. To numerically evaluate", + "bbox": [ + 496, + 863, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "4480", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/725e14362f498810b5d86d0cb388287e55dd33039ddbcee24e2307a522486150.jpg", + "image_caption": [ + "Figure 1. Overview of the proposed pipeline. NeuralLift-360 learns to recover a 3D object from a single reference image $y$ . We derive a prior distillation loss $\\mathcal{L}_{\\mathrm{diff}}$ for CLIP-guided diffusion prior and utilize rough depth estimation with the help of a ranking loss $\\mathcal{L}_{\\mathrm{ranking}}$ ." + ], + "image_footnote": [], + "bbox": [ + 163, + 87, + 816, + 330 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Eq. 1, the continuous integral is estimated using quadrature [34]. Our model is built upon Instant-NGP [37], which reduces computation cost by inferring color and density from a multiresolution hash table implemented by CUDA.", + "bbox": [ + 75, + 378, + 470, + 441 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Denoising Diffusion Model Denoising diffusion models generate images by gradually denoising from a gaussian noise $p(\\mathbf{x}_T) = \\mathcal{N}(\\mathbf{0},\\mathbf{I})$ and transforming into a certain data distribution. The forward diffusion process $q(\\pmb {x}_t|\\pmb {x}_{t - 1})$ adds Gaussian noise to the image $\\pmb {x}_t$ . The marginal distribution can be written as: $q(\\pmb {x}_t\\mid \\pmb {x_0}) = \\mathcal{N}\\left(\\alpha_t\\pmb {x_0},\\sigma_t^2\\pmb {I}\\right)$ where $\\alpha_{t}$ and $\\sigma_{t}$ are designed to converge to $\\mathcal{N}(\\mathbf{0},\\mathbf{I})$ when $t$ is at the end of the forward process [23, 56]. The reverse diffusion process $p(\\pmb{x}_{t - 1}|\\pmb {x}_t)$ learns to denoise. Given infinitesimal timesteps, the reverse diffusion process can be approximated with Gaussian [56] related with an optimal MSE denoiser [53]. The diffusion models are designed as noise estimators $\\epsilon_{\\theta}(\\pmb {x}_t,t)$ taking noisy images as input and estimating the noise. They are trained via optimizing the weighted evidence lower bound (ELBO) [17, 23]:", + "bbox": [ + 75, + 460, + 468, + 690 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {E L B O}} (\\theta) = \\mathbb {E} \\left[ w (t) \\| \\boldsymbol {\\epsilon} _ {\\theta} (\\alpha_ {t} \\pmb {x} _ {0} + \\sigma_ {t} \\pmb {\\epsilon}; t) - \\boldsymbol {\\epsilon} \\| _ {2} ^ {2} \\right], \\quad (2)\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 700, + 468, + 727 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})$ , $w(t)$ is a weighting function. In practice, setting $w(t) = 1$ delivers good performance [17]. Sampling from a diffusion model can be either stochastic [17] or deterministic [54]. After sampling $\\boldsymbol{x}_T \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})$ , we can gradually reduce the noise level and reach a clean image with high quality at the end of the iterative process.", + "bbox": [ + 75, + 737, + 468, + 829 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 76, + 844, + 168, + 859 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Overview. In this section, we propose our method, dubbed NeuralLift-360, which reconstructs a 3D object", + "bbox": [ + 75, + 869, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "from a single image. An overview of NeuralLift-360 is illustrated in Fig. 1. NeuralLift-360 combines the best of the two worlds: NeRF and diffusion model, where the former is leveraged to universally model a 3D scene and the latter is utilized to hallucinate the unseen views. To inject knowledge of a diffusion model into a radiance field, we derive a training framework in Sec. 4.1, whose supervision consists of two parts: (1) a prior distillation loss from a generative model (Sec. 4.2), (2) a straight loss on the given image (Sec. 4.4). Moreover, we introduce a domain adaption technique to adapt the diffusion prior to in-the-wild images (Sec. 4.3).", + "bbox": [ + 496, + 378, + 890, + 545 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Probabilistic-Driven 3D Lifting", + "text_level": 1, + "bbox": [ + 498, + 553, + 777, + 569 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We begin by formulating our setting along the way, introducing our notations. Given an image $\\pmb{y} \\in \\mathbb{R}^N$ and its text description $\\pmb{z} \\in \\mathbb{R}^D$ , NeuralLift-360 intends to reconstruct a 3D scene $\\pmb{V} \\in \\mathbb{R}^M$ , where $\\pmb{V}$ can be either a radiance volume or the network parameters of a NeRF. We regard $\\pmb{y}$ , $\\pmb{z}$ and $\\pmb{V}$ as random variables. To reconstruct $\\pmb{V}$ from $\\pmb{y}$ and $\\pmb{z}$ , we maximize a log-posterior and apply Bayesian rule:", + "bbox": [ + 496, + 577, + 890, + 683 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\log p (\\boldsymbol {V} | \\boldsymbol {y}, \\boldsymbol {z}) = \\log p (\\boldsymbol {y} | \\boldsymbol {V}, \\boldsymbol {z}) + \\log p (\\boldsymbol {V} | \\boldsymbol {z}) + c o n s t. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 508, + 691, + 890, + 724 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "One can see the prior term $p(\\mathbf{V}|z)$ provides complementary information to lift an image into 3D space. However, specifying this prior can be intractable, and directly optimizing objective Eq. 3 is also sampling inefficient. Instead, we introduce camera pose $\\Phi \\in \\mathbb{S}\\mathbb{O}(3)\\times \\mathbb{R}^3$ as a latent variable, then the likelihood term can be rewritten as:", + "bbox": [ + 496, + 734, + 890, + 825 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} p (\\boldsymbol {y} | \\boldsymbol {V}, \\boldsymbol {z}) = \\int p (\\boldsymbol {y} | \\boldsymbol {V}, \\boldsymbol {\\Phi}, \\boldsymbol {z}) p (\\boldsymbol {\\Phi} | \\boldsymbol {V}, \\boldsymbol {z}) d \\boldsymbol {\\Phi} (4) \\\\ = \\int p (\\boldsymbol {y} | \\boldsymbol {V}, \\boldsymbol {\\Phi}, \\boldsymbol {z}) \\frac {p (\\boldsymbol {V} | \\boldsymbol {\\Phi} , \\boldsymbol {z}) p (\\boldsymbol {\\Phi})}{p (\\boldsymbol {V} | \\boldsymbol {z})} d \\boldsymbol {\\Phi}, (5) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 832, + 890, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "4481", + "bbox": [ + 480, + 944, + 513, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where we assume $\\Phi$ is independent of $z$ . Merging Eq. 3 and Eq. 5, we can get rid of the implicit prior term and obtain a sampling efficient ELBO of the original objective as below:", + "bbox": [ + 76, + 90, + 468, + 137 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\log \\mathbb {E} _ {\\boldsymbol {\\Phi} \\sim p (\\boldsymbol {\\Phi})} p (\\boldsymbol {y} | \\boldsymbol {V}, \\boldsymbol {\\Phi}, \\boldsymbol {z}) p (\\boldsymbol {V} | \\boldsymbol {\\Phi}, \\boldsymbol {z}) (6) \\\\ \\geq \\mathbb {E} _ {\\boldsymbol {\\Phi} \\sim p (\\boldsymbol {\\Phi})} \\left[ \\log p (\\boldsymbol {y} | \\boldsymbol {V}, \\boldsymbol {\\Phi}, \\boldsymbol {z}) + \\log p (\\boldsymbol {V} | \\boldsymbol {\\Phi}, \\boldsymbol {z}) \\right]. (7) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 146, + 468, + 181 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Furthermore, we define $p(\\boldsymbol{y}|\\boldsymbol{V}, \\boldsymbol{\\Phi}, \\boldsymbol{z}) = p(\\boldsymbol{y}|h(\\boldsymbol{V}, \\boldsymbol{\\Phi}), \\boldsymbol{z})$ and $p(\\boldsymbol{V}|\\boldsymbol{\\Phi}, \\boldsymbol{z}) = p(h(\\boldsymbol{V}, \\boldsymbol{\\Phi})|\\boldsymbol{z})$ , where $h(\\boldsymbol{V}, \\boldsymbol{\\Phi})$ renders the 3D scene with respect to the camera pose $\\boldsymbol{\\Phi}$ . Then we derive our final training objective for NeuralLift-360:", + "bbox": [ + 76, + 190, + 468, + 251 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = - \\mathbb {E} _ {\\boldsymbol {\\Phi}} \\left[ \\underbrace {\\log p (\\boldsymbol {y} \\mid h (\\boldsymbol {V} , \\boldsymbol {\\Phi}) , \\boldsymbol {z})} _ {\\text {r e f e r e n c e d l o s s}} + \\underbrace {\\log p (h (\\boldsymbol {V} , \\boldsymbol {\\Phi}) \\mid \\boldsymbol {z})} _ {\\text {n o n - r e f e r e n c e d l o s s}} \\right]. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 83, + 258, + 468, + 324 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "At each training step, we sample a camera pose with respect to a distribution (more on Sec. 5.1), and then render the view associated with this camera parameter (via volume rendering Eq. 1). The loss computed on the rendered image consists of two parts: image-referenced guidance and reference-free regularization. The referenced loss will enforce $V$ to have a close appearance with given $y$ , while the non-referenced term hallucinates the image information with prior knowledge. In the next section, we introduce a diffusion model to implement our loss function Eq. 8.", + "bbox": [ + 76, + 334, + 468, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. CLIP-guided Diffusion Priors", + "text_level": 1, + "bbox": [ + 76, + 493, + 341, + 508 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Diffusion models [17, 54-56], which synthesize images by gradually denoising from a gaussian noise, have undeniably been successful in the realm of image generation, editing, and processing tasks. When trained on large-scale datasets for text-to-image tasks, these models can accurately represent the distribution of images based on textual descriptions [41, 42, 48]. Consequently, we are inspired to utilize a diffusion model's generative prior to complement the 3D information. To accomplish this, we re-parameterize our Eq. 8 with a diffusion model. Nevertheless, the derived loss function contains both discriminative and generative terms. Taking inspiration from classifier guidance techniques [26, 30, 56], we show that we can leverage the training loss of a diffusion model to surrogate the generative part and utilize CLIP space similarity to model the discriminative part.", + "bbox": [ + 75, + 516, + 468, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Following the derivation in [17] and [32], we lower bound Eq. 8 by:", + "bbox": [ + 76, + 758, + 468, + 789 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\log p _ {\\theta} (\\boldsymbol {y} | \\boldsymbol {x}, \\boldsymbol {z}) + \\log p _ {\\theta} (\\boldsymbol {x} | \\boldsymbol {z}) \\\\ \\geq - \\mathbb {E} _ {t, \\epsilon \\sim \\mathcal {N} (\\mathbf {0}, \\mathbf {I})} \\left[ \\mathcal {D} _ {\\mathrm {K L}} \\left(q \\left(\\boldsymbol {x} _ {t - 1} \\mid \\boldsymbol {x} _ {t}, \\boldsymbol {x} _ {0}\\right) \\| p _ {\\theta} \\left(\\boldsymbol {x} _ {t - 1} \\mid \\boldsymbol {x} _ {t}, \\boldsymbol {y}, \\boldsymbol {z}\\right) \\right] \\right. \\\\ = - \\mathbb {E} _ {t, \\boldsymbol {\\epsilon} \\sim \\mathcal {N} (\\mathbf {0}, \\boldsymbol {I})} \\left[ w (t) \\| \\nabla \\log p _ {\\theta} (\\boldsymbol {x} _ {t} | \\boldsymbol {y}, \\boldsymbol {z}) - \\boldsymbol {\\epsilon} \\| _ {2} ^ {2} \\right], \\tag {9} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 78, + 797, + 470, + 862 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\pmb{x}_t = \\alpha_t\\pmb{x}_0 + \\sigma_t\\pmb{\\epsilon}$ and $w(t)$ is a constant dependent on noise scale $\\alpha_{t}$ and $\\sigma_t$ . The expectation of $t\\sim \\mathcal{U}\\{1,\\dots ,T\\}$", + "bbox": [ + 76, + 869, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "is usually taken over a uniform distribution. We provide detailed derivation in supplementary. At first glance, we may directly plug in the pre-trained denoiser $\\epsilon_{\\theta}(\\pmb{x}_t;\\pmb{z})$ of a diffusion model to surrogate $\\nabla p_{\\theta}(\\pmb{x}_t|\\pmb{y},\\pmb{z})$ similar to [39]. However, note that $p_{\\theta}(\\pmb{x}_t|\\pmb{y},\\pmb{z})$ is also conditioned on the reference image $\\pmb{y}$ , which requires further efforts to inject the reference image information into the diffusion process.", + "bbox": [ + 496, + 90, + 890, + 196 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To this end, NeuralLift-360 leverages a Mixed Reference Guidance that guides the diffusion process with a mixture of references. Similar to classifier guidance techniques [26, 30, 32, 56], we rewrite $p_{\\theta}(\\boldsymbol{x}_t|\\boldsymbol{y}, \\boldsymbol{z}) = p_{\\theta}(\\boldsymbol{y}|\\boldsymbol{x}_t, \\boldsymbol{z})p_{\\theta}(\\boldsymbol{x}_t|\\boldsymbol{z}) / p_{\\theta}(\\boldsymbol{y}|\\boldsymbol{z})$ , and we adopt classifier-free guidance to model $p_{\\theta}(\\boldsymbol{x}_t|\\boldsymbol{z})$ , then we can obtain a modified score function [55] as below:", + "bbox": [ + 496, + 198, + 890, + 303 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\boldsymbol {\\epsilon} _ {\\theta} (\\boldsymbol {x} _ {t}; \\boldsymbol {y}, \\boldsymbol {z}) = (1 + w) \\boldsymbol {\\epsilon} _ {\\theta} (\\boldsymbol {x} _ {t}; \\boldsymbol {z}) - w \\boldsymbol {\\epsilon} _ {\\theta} (\\boldsymbol {x} _ {t}) \\\\ + \\sqrt {1 - \\alpha_ {t}} \\nabla \\log p _ {\\theta} (\\boldsymbol {y} | \\boldsymbol {x} _ {t}, \\boldsymbol {z}), \\tag {10} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 545, + 318, + 890, + 354 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $w$ is a hyper-parameter that controls the strength of the classifier guidance. We further measure the discrimination with a distance metric on the feature space $p_{\\theta}(\\boldsymbol{y}|\\boldsymbol{x}_t, \\boldsymbol{z}) \\propto \\exp\\left[-\\phi\\left(\\frac{\\boldsymbol{x}_t - \\sigma_t \\boldsymbol{\\epsilon}_{\\theta}(\\boldsymbol{x}_t; \\boldsymbol{z})}{\\alpha_t}, \\boldsymbol{y}\\right)\\right]$ . Specifically, we choose the inner product on the CLIP embedding space [40] as the distance metric, i.e., $\\log p_{\\theta}(\\boldsymbol{y}|\\boldsymbol{x}_t, \\boldsymbol{z}) \\propto -\\left\\langle F\\left(\\frac{\\boldsymbol{x}_t - \\sigma_t \\boldsymbol{\\epsilon}_{\\theta}(\\boldsymbol{x}_t)}{\\alpha_t}\\right), F(\\boldsymbol{y})\\right\\rangle$ , which spurs on semantic similarity between denoised rendered image and reference image $\\boldsymbol{y}$ , where $F$ is the CLIP image encoder.", + "bbox": [ + 496, + 369, + 890, + 518 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In each training iteration, we alternate two camera pose sampling strategies: (1) we fix a camera pose $\\Phi_0$ (i.e., a Delta distribution) as the reference view, and when rendering on that view, we only penalize the RGB differences by modeling $p(\\boldsymbol{y} | h(\\boldsymbol{V}, \\boldsymbol{\\Phi}), \\boldsymbol{z}) = \\mathcal{N}(\\boldsymbol{y} | h(\\boldsymbol{V}, \\boldsymbol{\\Phi}), \\sigma^2 \\boldsymbol{I})$ , (2) we stochastically sample camera poses, and adopt the diffusion ELBO (Eq. 9) to estimate the evidence of rendered image together with prior $p_{\\theta}(\\boldsymbol{x} | \\boldsymbol{z})$ . Combining all together, we can summarize our training loss as below:", + "bbox": [ + 496, + 518, + 890, + 655 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} = \\frac {1}{\\sigma^ {2}} \\underbrace {\\| h (\\boldsymbol {V} , \\boldsymbol {\\Phi} _ {0}) - \\boldsymbol {y} \\| _ {2} ^ {2}} _ {\\mathcal {L} _ {\\text {p h o t o m e t r i c}}} \\tag {11} \\\\ + \\underbrace {\\mathbb {E} _ {\\boldsymbol {\\Phi} , t , \\boldsymbol {\\epsilon} \\sim \\mathcal {N} (\\boldsymbol {0} , \\boldsymbol {I})} \\left[ w (t) \\| \\boldsymbol {\\epsilon} _ {\\theta} (\\alpha_ {t} h (\\boldsymbol {V} , \\boldsymbol {\\Phi}) + \\sigma_ {t} \\boldsymbol {\\epsilon} | \\boldsymbol {y} , \\boldsymbol {z}) - \\boldsymbol {\\epsilon} \\| _ {2} ^ {2} \\right]} _ {\\mathcal {L} _ {\\mathrm {d i f f}}}. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 667, + 901, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Compared with DreamFusion [39], NeuralLift-360's training objective grounds the 3D scene with the given image which enables the generated scene to have consistent appearance with the specified object.", + "bbox": [ + 496, + 772, + 890, + 833 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.3. Domain Adaption to In-the-wild Images", + "text_level": 1, + "bbox": [ + 500, + 845, + 841, + 862 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A primary challenge, however, lies in assigning an appropriate text prompt for our reference image. Given that", + "bbox": [ + 496, + 869, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4482", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c160a1d560897fd34353942460214ed10cf85d921bccb454939bd5ead57c525a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 88, + 155, + 143 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/42721362303136b9f0d4c3afa5614b7e9326a62cf2614d01f64872c6fca20353.jpg", + "image_caption": [ + "(a) Pre-trained" + ], + "image_footnote": [], + "bbox": [ + 158, + 88, + 233, + 143 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/dcc6c2ab0768b4ae14aaa71f1b11d37ea2c94bfd676956c62b305d0e270f13ab.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 236, + 88, + 310, + 143 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0ca5b2e3e3d4841da341e3c15d2b921165e698d4ba14c4f6bad700ffc8bbadcd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 313, + 88, + 390, + 143 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/428a4ee7e2027c92cbd8c6df576749ac85becc2b47b0a24a471f977851348bac.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 392, + 88, + 468, + 143 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d45179ef0aa73b6c8bff223309ce4e4b56a3741cff17d4c47c23e06dd2f8196d.jpg", + "image_caption": [ + "Figure 2. After fine-tuning, our model can generate diverse images similar to the reference image (b) while not identical." + ], + "image_footnote": [], + "bbox": [ + 80, + 159, + 155, + 218 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/faaf30362df5684fbcfaf8eb6b4c80acf115e1a7e0dda21815099486d27d41ef.jpg", + "image_caption": [ + "(c) Fine-tuned" + ], + "image_footnote": [], + "bbox": [ + 156, + 159, + 232, + 218 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/599cc7a413195a7fa7d519255148d3e27d323095ebf7fe517fd24034e7771f2f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 233, + 159, + 310, + 218 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/95620d1a82f2c61d9a7e888d725bbe9f56a2ddd7f0fd50a571b7c947544255ad.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 313, + 159, + 387, + 218 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/91e36f2d1e4951c26124c9a0ae489047d3ad56f323414eb38a6b9e61c84f90c0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 390, + 159, + 464, + 218 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "large diffusion models are predominantly trained for text-to-image tasks, they typically require a text embedding as conditional input. For images in the wild, it is generally hard to exactly describe an image since images contain more fine-grained information than texts.", + "bbox": [ + 75, + 279, + 467, + 354 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Adapting to in-the-wild images necessitates finetuning the Stable Diffusion model on the reference image Textual inversion [15] and Dreambooth [49] enable finetuning the Stable Diffusion model on a limited number of images for a subject-driven generation. However, in our situation, we only have access to a single image, which significantly hampers the effectiveness of these methods. Imagic [22] and UniTune [60] support finetuning on a single image, but their primary objective is to perform real image editing, which can be quite successful even if the diffusion model overfits the given image. Nevertheless, in our context, we need the diffusion model to generate images resembling the provided reference but not identical to it. If the supervision for each view is the same, NeRF will otherwise produce an isotropic surface.", + "bbox": [ + 75, + 354, + 467, + 580 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We draw inspiration from Pivotal Tuning [47] and Imagic [22] when fine-tuning Stable Diffusion [48] on our own. To start with, we optimize the CLIP text embedding $z$ so that it fits our reference image $y$ better. During fine-tuning, we use Eq. 2 to supervise the noise estimate. As we expect that our denoised result will resemble the reference image, we also apply a CLIP image loss to the estimated denoised image. $\\hat{x}_0$ ,", + "bbox": [ + 75, + 580, + 467, + 703 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {f i n e t u n e}} = \\mathbb {E} _ {t, \\epsilon} \\left[ \\| \\boldsymbol {\\epsilon} - \\boldsymbol {\\epsilon} _ {\\theta} (\\mathbf {x} _ {t}; \\boldsymbol {z}) \\| _ {2} ^ {2} \\right] \\\\ - \\mathbb {E} _ {t, \\epsilon} \\left\\langle F \\left(\\frac {\\boldsymbol {x} _ {t} - \\sigma_ {t} \\boldsymbol {\\epsilon} _ {\\theta} (\\boldsymbol {x} _ {t} ; \\boldsymbol {z})}{\\alpha_ {t}}\\right), F (\\boldsymbol {y}) \\right\\rangle , \\tag {12} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 710, + 467, + 771 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $F(\\cdot)$ is the pre-trained CLIP image encoder.", + "bbox": [ + 76, + 779, + 411, + 794 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then, we fix the optimized text embedding $z^{*}$ and finetune the diffusion model $\\epsilon_{\\theta}(\\cdot)$ using the same loss function as above. To preserve the generalization ability of the diffusion model, we introduce several augmentations during the finetuning. Specifically, in each iteration, we perform random image-level augmentations such as flip, transpose, shift, scale, and rotate. As shown in Fig. 2, we are able", + "bbox": [ + 75, + 795, + 467, + 898 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "to overcome the overfitting issue and preserve the diffusion model's ability to generate diverse similar but not identical contents. Similar to Imagic [22] and UniTune [47], we observe that after finetuning, the optimized text embedding will have an overfitting issue and it is critical to perform interpolation between the original text embedding and the optimized text embedding to preserve the diffusion model's ability to generate diverse contents. We obtain $z' = \\eta z^* + (1 - \\eta)z$ , where $\\eta$ is a weighting hyper-parameter.", + "bbox": [ + 496, + 90, + 890, + 227 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During sampling, we aim to amplify the conditional likelihood for our optimized embedding $z^{*}$ , while decreasing both the unconditional likelihood and the conditional likelihood for the original embedding $z$ . Hence for 50% of all iterations, we use the original Eq. 10, while for the others, we replace the unconditional sampling $\\epsilon_{\\theta}(\\boldsymbol{x}_t)$ in Eq. 10 with conditional sampling using the original text embedding $z$ :", + "bbox": [ + 496, + 227, + 890, + 332 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\widetilde {\\boldsymbol {\\epsilon}} _ {\\theta} \\left(\\boldsymbol {x} _ {t}; \\boldsymbol {y}, \\boldsymbol {z} ^ {\\prime}, \\boldsymbol {z}\\right) = (1 + w) \\boldsymbol {\\epsilon} _ {\\theta} \\left(\\boldsymbol {x} _ {t}; \\boldsymbol {z} ^ {\\prime}\\right) \\tag {13} \\\\ - w \\epsilon_ {\\theta} (\\boldsymbol {x} _ {t}; \\boldsymbol {z}) + \\sqrt {1 - \\alpha_ {t}} \\nabla \\log p _ {\\theta} (\\boldsymbol {y} | \\boldsymbol {x} _ {t}, \\boldsymbol {z} ^ {\\prime}), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 344, + 890, + 381 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Supervision from Rough Depth", + "text_level": 1, + "bbox": [ + 500, + 393, + 774, + 409 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "While the aforementioned pipeline laid the foundation for the success of our method, we observe that obtaining a high-quality 3D object without geometry regularization is inherently challenging, if relying solely on learning priors. As shown in SinNeRF [70] and DS-NeRF [8], accurate depth information can largely help improve the rendering quality. However, their impressive outcomes are contingent upon the 3D consistency of the accessible high-quality depth, a requirement that is not easily fulfilled in real-life scenarios.", + "bbox": [ + 496, + 417, + 890, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In spite of the significant progress in monocular depth estimation for in-the-wild scenarios, current state-of-the-art depth estimators are insufficient for recovering 3D shapes. Due to unknown camera baselines and stereoscopic post-processing in data, these methods [7,25,44,64,69,72] utilize loss functions invariant to shift and scale or adopt ranking loss that utilizes relative information. Therefore, their results cannot be used to reconstruct plausible 3D scene shapes directly due to unknown depth shifts [74]. Though there are also attempts [74] that improve off-the-shelf depth estimators, they generally can not generalize well to unseen objects in the wild.", + "bbox": [ + 496, + 568, + 890, + 748 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given that the acquired depth is imperfect, the absolute depth scale lacks the reliability necessary for recovering a high-quality 3D object. In contrast to previous depth-supervised NeRFs [8, 70], which rely solely on precise values from the depth map, our approach represents the first attempt to maximize the utilization of depth information by adopting the relative ranking information. Drawing inspiration from [7, 38], we adopt a pairwise depth ranking loss to supervise the rendered depth from our NeRF backbone. In this way, we enforce the depth we rendered to be consistent", + "bbox": [ + 496, + 750, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "4483", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "with the depth from the reference view in terms of internal ranking instead of actual value. The loss is formulated as follows,", + "bbox": [ + 75, + 90, + 468, + 135 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r a n k i n g}} = \\left\\{ \\begin{array}{l l} \\max \\left(z _ {j _ {k}} - z _ {i _ {k}}, 0\\right), & r _ {k} \\in \\{> \\} \\\\ \\max \\left(z _ {i _ {k}} - z _ {j _ {k}}, 0\\right), & r _ {k} \\in \\{< \\} \\end{array} \\right. \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 141, + 468, + 180 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $r_k$ refers to the pairwise ranking relationship between pixel $z_{i_k}$ and $z_{j_k}$ in the monocular depth estimation. This loss function is used alongside the photo-metric loss and diffusion loss mentioned in Eq. 11 and serves as additional supervision on geometry.", + "bbox": [ + 75, + 185, + 468, + 261 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 268, + 209, + 284 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we will first describe a unique training recipe that we propose to train our framework, then the implementation details.", + "bbox": [ + 75, + 290, + 468, + 334 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Training Recipe", + "text_level": 1, + "bbox": [ + 76, + 338, + 236, + 354 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Diverse Camera Pose and Intrinsic Sampling During training, we observe that a diverse camera pose and intrinsicics are essential for the model to learn a coherent 3D object. If the camera pose is not sampled from a diverse distribution, the 3D object will collapse towards the major viewing directions and thus generate unpleasant results. Specifically, we alter between two pose sampling strategies; (1) we have a fixed camera pose $\\Phi_0$ (i.e., a Delta distribution) for the reference view, (2) we uniformly sample the camera rotation on the surface of a unit sphere and then uniformly sample the radius in [0.4, 1.0] to locate the camera. We start with the Delta distribution, and as training progresses, we gradually shift to the other sophisticated camera pose distribution. The camera intrinsics provide additional augmentation on the viewing direction. By uniformly sampling the field of view in [50, 70], we zoom in and out to get different resolution projections for the scene. We also add Gaussian noise to the camera pose sampling phase. Such perturbations help us render more diverse images during training.", + "bbox": [ + 75, + 357, + 468, + 645 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Foreground Aware Diffusion Prior ERNIE-ViLG [13] utilizes a weighting mechanism during the training of diffusion models to let the model generate. Using an additional object detection model, their weighted loss forces the model to focus on salient regions. Similarly, we hope to strengthen the guidance on the foreground object region. We propose to weight the score matching distance with the predicted density. Specifically, we penalize the regions with higher density so that our NeRF gains more guidance on the foreground regions. Hence, for each rendered image, we first obtain the bounding box for the non-zero density regions. Then, we extend the $w(t)$ in $\\mathcal{L}_{\\mathrm{diff}}$ (Eq. 11) into a spatial function:", + "bbox": [ + 75, + 662, + 468, + 858 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nW _ {i, j} = \\left\\{ \\begin{array}{l l} 2, & \\text {i f} (i, j) \\text {i s i n s i d e t h e b o u n d i n g b o x} \\\\ 1, & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 864, + 468, + 904 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Geometry Regularization We also include a series of geometry regularizations to avoid undesired artifacts in our 3D object. To avoid foggy and semi-transparent artifacts at the back of the scene, we utilize the orient loss from Ref-NeRF [61] to prevent the surface normals from facing backward from the camera. We also incorporate depth smoothness loss [62], sparsity loss, and distortion loss [3, 58] to avoid the NeRF filling empty space with floaters. Full details for these regularizations are provided in the supplementary material.", + "bbox": [ + 496, + 90, + 890, + 242 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Timestep Annealing At the start of training, the quality of the rendered images is poor, and we need the diffusion prior to provide more thorough guidance. As a result, we need to add more noise to perturb the rendered image. As the training progresses, the quality of rendered images gradually improves, and we choose to reduce the timestep range accordingly. By annealing the timestep where we add noise, we are lowering the noise level of the perturbation, and the noise estimate becomes more accurate.", + "bbox": [ + 496, + 250, + 890, + 383 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Evaluation Workflow", + "text_level": 1, + "bbox": [ + 500, + 393, + 697, + 407 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct extensive experiments using both text-to-image synthetic images and real-world images. Synthetic images in our experiments are generated using PNDM Sampler [29] with 50 sampling steps. Real images are either from CO3Dv2 dataset [46] or randomly crawled from the Google search engine. For each image, we use a pre-trained background removal framework [1] to segment the foreground objects. Then we obtain their depth using LeRes [73] with the \"Boost Your Own depth\" [33] strategy. [33] inferences the monocular depth estimation network on multiple patches and then fuses the results for better details. We compare with previous sparse view NeRFs, including Depth-supervised NeRF [8] and SinNeRF [70]. We also compare with DietNeRF [19] that uses CLIP as guidance.", + "bbox": [ + 496, + 416, + 890, + 627 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Comparisons", + "text_level": 1, + "bbox": [ + 500, + 636, + 638, + 651 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We present quantitative results in eight scenes. We report the average CLIP distance between rendered image and the reference image as a metric for how well the object is represented. During testing, we use a CLIP model different from training time and evaluate on held out viewing directions with a larger distance-to-origin than training. We render 100 images on each scene for each method. As shown in Tab. 1, our method achieves the best CLIP distance among existing approaches.", + "bbox": [ + 496, + 659, + 890, + 794 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We also provide visual comparisons in Fig. 3. Since DSNeRF and SinNeRF rely on the exact value of the available depth information, their performance using monocular depth estimation is far from satisfactory. It can also be seen that the DietNeRF suffers from unconstrained geometry. Our method, on the other hand, is able to synthesize visually pleasing novel views with 3D consistency.", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "4484", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2fa5abf08f79858788196ec86283d10e346a75da03b8b32250595477e658dccc.jpg", + "image_caption": [ + "Figure 3. Visual comparisons for novel view synthesis results. We provide video comparisons in supplementary." + ], + "image_footnote": [], + "bbox": [ + 99, + 87, + 869, + 873 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "4485", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ff3c4fdd375b5f1de9a5db6dfb3757260018fe029f63d6de1866ace6f89a070c.jpg", + "table_caption": [ + "Table 1. Quantitative evaluation by comparing the proposed method with previous state-of-the-art approaches." + ], + "table_footnote": [], + "table_body": "
CLIP Distance ↓AppleBenchBaseballHydrantRemoteSuitcaseStatueSandwichAverage
DSNeRF [8]0.53090.54470.55060.62260.64410.55330.59920.61100.5821
DietNeRF [19]0.42450.55790.45450.51900.57180.48900.59470.51790.5162
SinNeRF [70]0.51920.57690.62300.63430.64960.57050.56050.60150.5918
NeuralLift-360 (Ours)0.39140.47850.39740.51460.49160.41860.44110.46550.4498
", + "bbox": [ + 122, + 114, + 848, + 185 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5b63a43d9c64fc35ae3add3e12b667955b411b2ac73dc82001d688914b37a7da.jpg", + "image_caption": [ + "Figure 4. Ablation study on the baseline variants. Given a single reference image, we run different variants to render its $360^{\\circ}$ views." + ], + "image_footnote": [], + "bbox": [ + 102, + 190, + 867, + 334 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/eaf753534202c57de45f178e63bbb73cb703165fb5d44bd6fe894366ac993bf9.jpg", + "image_caption": [ + "Figure 5. Ablation study on CLIP guidance for the diffusion prior." + ], + "image_footnote": [], + "bbox": [ + 99, + 367, + 444, + 691 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 723, + 230, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Variants of supervision. In this section, we perform detailed ablation studies on different components of our framework. \"w/o $\\mathcal{L}_{\\mathrm{ranking}}$ \" replaces the ranking loss with $L_{1}$ loss on depth. \"w/o $\\mathcal{L}_{\\mathrm{diff}}$ \" replaces the CLIP guided diffusion prior with CLIP. \"w/o fine-tuning\" refers to using the pre-trained diffusion model. As shown in Fig. 4, our full model delivers the best performance, while other variants have different issues. \"w/o $\\mathcal{L}_{\\mathrm{ranking}}$ \" relies on the exact value from depth information and suffers from unreliable depth. \"w/o $\\mathcal{L}_{\\mathrm{diff}}$ \" uses CLIP as a prior, and the optimized result is", + "bbox": [ + 75, + 750, + 470, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "blurry and collapsed. \"w/o fine-tuning\" uses a pre-trained diffusion model, since the text prompt is not very accurate, the resulting object has lettuce which doesn't exist in the reference image.", + "bbox": [ + 498, + 368, + 890, + 429 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of CLIP guided diffusion prior We validate the effectiveness of CLIP guidance in our diffusion prior design in Fig. 5. It can be seen that without the CLIP guidance, although the model produces a 3D object, it doesn't look like the reference image and has a strange shape. In comparison, our full model generates a better shape and looks more similar to the reference image.", + "bbox": [ + 498, + 460, + 892, + 566 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusions and Future Work", + "text_level": 1, + "bbox": [ + 500, + 594, + 777, + 609 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present NeuralLift-360, a novel framework to lift an in-the-wild 2D photo into a 3D object with 360-degree views. NeuralLift-360 learns probabilistic-driven 3D lifting with CLIP-guided diffusion priors and mitigates the depth errors by a scale-invariant depth ranking loss. Comprehensive experiments are conducted on real and synthetic images, where NeuralLift-360 outperforms the current state-of-the-art methods. While our current implementation is based on Stable Diffusion [48] version 1.4, the framework should also work well with other diffusion models such as Imagen [50] or DALL-E2 [41], once they are open-sourced.", + "bbox": [ + 496, + 625, + 890, + 790 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. Despite the encouraging visual results rendered by NeuralLift-360, the target resolution $(128 \\times 128)$ is still behind large generative models [41, 50]. Moreover, challenging cases (e.g., multiple objects with occlusion) are not included in our assumptions, which need to be further explored. Going forward, we will investigate how to expand our method to more general scenarios.", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "4486", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Image background removal. https://github.com/OPHoperHPO/image-background-remove-tool. 6", + "[2] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 2", + "[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5470–5479, 2022. 2, 6", + "[4] Angel X. Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. ShapeNet: An Information-Rich 3D Model Repository. Technical Report arXiv:1512.03012 [cs.GR], Stanford University — Princeton University — Toyota Technological Institute at Chicago, 2015. 1", + "[5] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14124-14133, 2021. 2", + "[6] Tianlong Chen, Peihao Wang, Zhiwen Fan, and Zhangyang Wang. Aug-nerf: Training stronger neural radiance fields with triple-level physically-grounded augmentations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15191-15202, 2022. 2", + "[7] Weifeng Chen, Zhao Fu, Dawei Yang, and Jia Deng. Single-image depth perception in the wild. Advances in neural information processing systems, 29, 2016. 5", + "[8] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised nerf: Fewer views and faster training for free. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12882-12891, 2022. 1, 5, 6, 7, 8", + "[9] Jean-Denis Durou, Maurizio Falcone, and Manuela Sagona. Numerical methods for shape-from-shading: A new survey with benchmarks. Computer Vision and Image Understanding, 109(1):22-43, 2008. 1", + "[10] Haoqiang Fan, Hao Su, and Leonidas J Guibas. A point set generation network for 3d object reconstruction from a single image. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 605-613, 2017. 1", + "[11] Zhiwen Fan, Yifan Jiang, Peihao Wang, Xinyu Gong, Dejia Xu, and Zhangyang Wang. Unified implicit neural stylization. arXiv preprint arXiv:2204.01943, 2022. 2", + "[12] Paolo Favaro and Stefano Soatto. A geometric approach to shape from defocus. IEEE Transactions on Pattern Analysis and Machine Intelligence, 27(3):406-417, 2005. 1", + "[13] Zhida Feng, Zhenyu Zhang, Xintong Yu, Yewei Fang, Lanxin Li, Xuyi Chen, Yuxiang Lu, Jiaxiang Liu, Weichong" + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Yin, Shikun Feng, et al. Ernie-vilg 2.0: Improving text-to-image diffusion model with knowledge-enhanced mixture-of-denoising-experts. arXiv preprint arXiv:2210.15257, 2022.6", + "[14] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5501–5510, 2022. 2", + "[15] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 5", + "[16] Yuan-Chen Guo, Di Kang, Linchao Bao, Yu He, and Song-Hai Zhang. Nerfren: Neural radiance fields with reflections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18409-18418, 2022. 2", + "[17] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 3, 4", + "[18] Ajay Jain, Ben Mildenhall, Jonathan T Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 867-876, 2022. 2", + "[19] Ajay Jain, Matthew Tancik, and Pieter Abbeel. Putting nerf on a diet: Semantically consistent few-shot view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5885-5894, 2021. 1, 6, 7, 8", + "[20] Varun Jampani, Huiwen Chang, Kyle Sargent, Abhishek Kar, Richard Tucker, Michael Krainin, Dominik Kaeser, William T Freeman, David Salesin, Brian Curless, et al. Slide: Single image 3d photography with soft layering and depth-aware inpainting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12518-12527, 2021. 2", + "[21] Adrian Johnston, Ravi Garg, Gustavo Carneiro, Ian Reid, and Anton van den Hengel. Scaling cnns for high resolution volumetric reconstruction from a single image. In Proceedings of the IEEE International Conference on Computer Vision Workshops, pages 939-948, 2017. 1", + "[22] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagc: Text-based real image editing with diffusion models. arXiv preprint arXiv:2210.09276, 2022. 5", + "[23] Diederik Kingma, Tim Salimans, Ben Poole, and Jonathan Ho. Variational diffusion models. Advances in neural information processing systems, 34:21696-21707, 2021. 3", + "[24] Hyung-Kwon Ko, Gwanmo Park, Hyeon Jeon, Jaemin Jo, Juho Kim, and Jinwook Seo. Large-scale text-to-image generation models for visual artists' creative works. arXiv preprint arXiv:2210.08477, 2022. 2", + "[25] Manuel Lang, Alexander Hornung, Oliver Wang, Steven Poulakos, Aljoscha Smolic, and Markus Gross. Nonlinear disparity mapping for stereoscopic 3d. ACM Transactions on Graphics (TOG), 29(4):1-10, 2010. 5" + ], + "bbox": [ + 503, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "4487", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Yann LeCun, Sumit Chopra, Raia Hadsell, M Ranzato, and Fujie Huang. A tutorial on energy-based learning. Predicting structured data, 1(0), 2006. 4", + "[27] Zhengqi Li and Noah Snavely. Megadepth: Learning single-view depth prediction from internet photos. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2041-2050, 2018. 1", + "[28] Kai-En Lin, Lin Yen-Chen, Wei-Sheng Lai, Tsung-Yi Lin, Yi-Chang Shih, and Ravi Ramamoorthi. Vision transformer for nef-based view synthesis from a single input image. arXiv preprint arXiv:2207.05736, 2022. 2", + "[29] Luping Liu, Yi Ren, Zhijie Lin, and Zhou Zhao. Pseudo numerical methods for diffusion models on manifolds. arXiv preprint arXiv:2202.09778, 2022. 6", + "[30] Xihui Liu, Dong Huk Park, Samaneh Azadi, Gong Zhang, Arman Chopikyan, Yuxiao Hu, Humphrey Shi, Anna Rohrbach, and Trevor Darrell. More control for free! image synthesis with semantic diffusion guidance. 2021. 4", + "[31] Yuan Liu, Sida Peng, Lingjie Liu, Qianqian Wang, Peng Wang, Christian Theobalt, Xiaowei Zhou, and Wenping Wang. Neural rays for occlusion-aware image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7824-7833, 2022. 2", + "[32] Calvin Luo. Understanding diffusion models: A unified perspective. arXiv preprint arXiv:2208.11970, 2022. 4", + "[33] S. Mahdi H. Miangoleh, Sebastian Dille, Long Mai, Sylvain Paris, and Yaqiz Aksoy. Boosting monocular depth estimation models to high-resolution via content-adaptive multi-resolution merging. 2021. 6", + "[34] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European conference on computer vision, pages 405-421. Springer, 2020. 2, 3", + "[35] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1", + "[36] Fangzhou Mu, Jian Wang, Yicheng Wu, and Yin Li. 3d photo stylization: Learning to generate stylized novel views from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16273-16282, 2022. 2", + "[37] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. arXiv preprint arXiv:2201.05989, 2022. 2, 3", + "[38] Georgios Pavlakos, Xiaowei Zhou, and Kostas Daniilidis. Ordinal depth supervision for 3d human pose estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7307-7316, 2018. 5", + "[39] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 2, 4", + "[40] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry," + ], + "bbox": [ + 78, + 90, + 468, + 902 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2, 4", + "[41] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2, 4, 8", + "[42] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 4", + "[43] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12179-12188, 2021. 1", + "[44] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE transactions on pattern analysis and machine intelligence, 2020. 1, 2, 5", + "[45] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding up neural radiance fields with thousands of tiny mlps. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14335-14345, 2021. 2", + "[46] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In International Conference on Computer Vision, 2021. 6", + "[47] Daniel Roich, Ron Mokady, Amit H Bermano, and Daniel Cohen-Or. Pivotal tuning for latent-based editing of real images. ACM Transactions on Graphics (TOG), 42(1):1-13, 2022. 5", + "[48] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2, 4, 5, 8", + "[49] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. arXiv preprint arXiv:2208.12242, 2022. 5", + "[50] Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 2, 8", + "[51] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2304–2314, 2019. 1", + "[52] Meng-Li Shih, Shih-Yang Su, Johannes Kopf, and Jia-Bin Huang. 3d photography using context-aware layered depth" + ], + "bbox": [ + 501, + 90, + 890, + 902 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "4488", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + " inpainting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8028-8038, 2020. 1, 2", + "[53] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256-2265. PMLR, 2015. 3", + "[54] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 3, 4", + "[55] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in Neural Information Processing Systems, 32, 2019. 4", + "[56] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 3, 4", + "[57] Mohammed Suhail, Carlos Esteves, Leonid Signal, and Ameesh Makadia. Light field neural rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8269-8279, 2022. 2", + "[58] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Improved direct voxel grid optimization for radiance fields reconstruction. arxiv.cs.GR 2206.05085, 2022. 6", + "[59] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5459-5469, 2022. 2", + "[60] Dani Valevski, Matan Kalman, Yossi Matias, and Yaniv Leviathan. Unitune: Text-driven image editing by fine tuning an image generation model on a single image. arXiv preprint arXiv:2210.09477, 2022. 5", + "[61] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T Barron, and Pratul P Srinivasan. Ref-nerf: Structured view-dependent appearance for neural radiance fields. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5481-5490. IEEE, 2022. 2, 6", + "[62] Chaoyang Wang, José Miguel Buenapasada, Rui Zhu, and Simon Lucey. Learning depth from monocular videos using direct methods. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2022-2030, 2018. 6", + "[63] Can Wang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Text-and-image driven manipulation of neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3835–3844, 2022. 2", + "[64] Chaoyang Wang, Simon Lucey, Federico Peruzzi, and Oliver Wang. Web stereo video supervision for depth prediction from dynamic scenes. In 2019 International Conference on 3D Vision (3DV), pages 348-357. IEEE, 2019. 5", + "[65] Lizhen Wang, Xiaochen Zhao, Tao Yu, Songtao Wang, and Yebin Liu. Normalgan: Learning detailed 3d human from a single rgb-d image. In European Conference on Computer Vision, pages 430-446. Springer, 2020. 1" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[66] Nanyang Wang, Yinda Zhang, Zhuwen Li, Yanwei Fu, Wei Liu, and Yu-Gang Jiang. Pixel2mesh: Generating 3d mesh models from single rgb images. In Proceedings of the European conference on computer vision (ECCV), pages 52-67, 2018. 1", + "[67] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P Srinivasan, Howard Zhou, Jonathan T Barron, Ricardo Martin-Brualla, Noah Snavely, and Thomas Funkhouser. Ibrnet: Learning multi-view image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4690-4699, 2021. 2", + "[68] Jiajun Wu, Yifan Wang, Tianfan Xue, Xingyuan Sun, Bill Freeman, and Josh Tenenbaum. Marrnet: 3d shape reconstruction via 2.5 d sketches. Advances in neural information processing systems, 30, 2017. 1", + "[69] Ke Xian, Chunhua Shen, Zhiguo Cao, Hao Lu, Yang Xiao, Ruibo Li, and Zhenbo Luo. Monocular relative depth perception with web stereo data supervision. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 311-320, 2018. 5", + "[70] Dejia Xu, Yifan Jiang, Peihao Wang, Zhiwen Fan, Humphrey Shi, and Zhangyang Wang. Sinnerf: Training neural radiance fields on complex scenes from a single image. arXiv preprint arXiv:2204.00928, 2022. 1, 2, 5, 6, 7, 8", + "[71] Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. Mvsnet: Depth inference for unstructured multi-view stereo. In Proceedings of the European conference on computer vision (ECCV), pages 767-783, 2018. 1", + "[72] Wei Yin, Xinlong Wang, Chunhua Shen, Yifan Liu, Zhi Tian, Songcen Xu, Changming Sun, and Dou Renyin. Diversedepth: Affine-invariant depth prediction using diverse data. arXiv preprint arXiv:2002.00569, 2020. 5", + "[73] Wei Yin, Jianming Zhang, Oliver Wang, Simon Niklaus, Long Mai, Simon Chen, and Chunhua Shen. Learning to recover 3d scene shape from a single image. In Proc. IEEE Conf. Comp. Vis. Patt. Recogn. (CVPR), 2021. 1, 6", + "[74] Wei Yin, Jianming Zhang, Oliver Wang, Simon Niklaus, Long Mai, Simon Chen, and Chunhua Shen. Learning to recover 3d scene shape from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 204-213, 2021. 5", + "[75] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4578-4587, 2021. 1, 2", + "[76] Jiahui Yu, Zhe Lin, Jimei Yang, Xiaohui Shen, Xin Lu, and Thomas S Huang. Free-form image inpainting with gated convolution. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4471-4480, 2019. 2" + ], + "bbox": [ + 501, + 92, + 890, + 796 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "4489", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/NeuralLift-360_ Lifting an In-the-Wild 2D Photo to a 3D Object With 360deg Views/90dd78db-4103-415c-8ddc-ba0f83221385_model.json b/2023/NeuralLift-360_ Lifting an In-the-Wild 2D Photo to a 3D Object With 360deg Views/90dd78db-4103-415c-8ddc-ba0f83221385_model.json new file mode 100644 index 0000000000000000000000000000000000000000..37d3f5c7df3f48a37cc1956cb4d957af0bf7e794 --- /dev/null +++ b/2023/NeuralLift-360_ Lifting an In-the-Wild 2D Photo to a 3D Object With 360deg Views/90dd78db-4103-415c-8ddc-ba0f83221385_model.json @@ -0,0 +1,2422 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.079, + 0.131, + 0.895, + 0.153 + ], + "angle": 0, + "content": "NeuralLift-360: Lifting An In-the-wild 2D Photo to A 3D Object with \\(360^{\\circ}\\) Views" + }, + { + "type": "text", + "bbox": [ + 0.176, + 0.181, + 0.795, + 0.217 + ], + "angle": 0, + "content": "Dejia Xu, Yifan Jiang, Peihao Wang, Zhiwen Fan, Yi Wang, Zhangyang Wang \nVITA Group, University of Texas at Austin" + }, + { + "type": "text", + "bbox": [ + 0.362, + 0.219, + 0.611, + 0.234 + ], + "angle": 0, + "content": "{dejia,atlaswang}@utexas.edu" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.301, + 0.474, + 0.633 + ], + "angle": 0, + "content": "Virtual reality and augmented reality (XR) bring increasing demand for 3D content generation. However, creating high-quality 3D content requires tedious work from a human expert. In this work, we study the challenging task of lifting a single image to a 3D object and, for the first time, demonstrate the ability to generate a plausible 3D object with \\(360^{\\circ}\\) views that corresponds well with the given reference image. By conditioning on the reference image, our model can fulfill the everlasting curiosity for synthesizing novel views of objects from images. Our technique sheds light on a promising direction of easing the workflows for 3D artists and XR designers. We propose a novel framework, dubbed NeuralLift-360, that utilizes a depth-aware neural radiance representation (NeRF) and learns to craft the scene guided by denoising diffusion models. By introducing a ranking loss, our NeuralLift-360 can be guided with rough depth estimation in the wild. We also adopt a CLIP-guided sampling strategy for the diffusion prior to provide coherent guidance. Extensive experiments demonstrate that our NeuralLift-360 significantly outperforms existing state-of-the-art baselines. Project page: https://vita-group.github.io/NeuralLift-360/" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.664, + 0.21, + 0.68 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.689, + 0.471, + 0.886 + ], + "angle": 0, + "content": "Creating 3D content has been a long-standing problem in computer vision. This problem enables various applications in game studios, home decoration, virtual reality, and augmented reality. Over the past few decades, the manual task has dominated real scenarios, which requires tedious professional expert modeling. Modern artists rely on special software tools (e.g., Blender, Maya3D, 3DS Max, etc.) and time-demanding manual adjustments to realize imaginations and transform them into virtual objects. Meanwhile, automatic 3D content creation pipelines serve as effective tools to facilitate human efforts. These pipelines typically capture hundreds of images and leverage multi-view stereo [71] to quickly model fine-grained virtual landscapes." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.887, + 0.47, + 0.902 + ], + "angle": 0, + "content": "More recently, researchers have started aiming at a more" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.27, + 0.895, + 0.438 + ], + "angle": 0, + "content": "ambitious goal, to create a 3D object from a single image [9, 10, 12, 21, 27, 51, 66, 68]. This enables broad applications since it greatly reduces the prerequisite to a minimal request. Existing approaches can be mainly categorized into two major directions. One line of work utilizes learning priors from large-scale datasets with multi-view images [10, 21, 51, 66, 68]. These approaches usually learn a conditional neural network to predict 3D information based on input images. However, due to their poor generalization ability, drastic performance degradation is observed when the testing image is out-of-distribution." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.442, + 0.896, + 0.73 + ], + "angle": 0, + "content": "Another direction constructs the pipeline on top of the depth estimation techniques [65]. Under the guidance of monocular depth estimation networks [43, 44], the 2D image is firstly back-projected to a 3D data format (e.g., point cloud or multi-plane image) and then re-projected into novel views. After that, advanced image inpainting techniques are then adopted to fill missing holes [52] produced during the projection. However, most of them can be highly affected by the quality of the estimated depth maps. Though LeRes [73] attempted to rectify the predicted depth by refining the projected 3D point cloud, their results do not generalize well to an arbitrary image in the wild. Overall, the aforementioned approaches are either adopted in limited scenarios (e.g. face or toy examples [4, 75]), or only produce limited viewing directions [8, 70] when being applied to scenes in the wild. Different from all these approaches, we focus on a more challenging task and for the first time, show promising results by lifting a single in-the-wild image into a 3D object with \\(360^{\\circ}\\) novel views." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.897, + 0.903 + ], + "angle": 0, + "content": "Attracted by the dramatic progress of neural volumetric rendering on 3D reconstruction tasks, we consider building our framework based on Neural Radiance Fields (NeRFs) [35]. The original NeRF takes hundreds of training views and their camera poses as inputs to learn an implicit representation. Subsequent models dedicate tremendous efforts [8, 19, 70, 75] to apply NeRF to sparse training views. The most similar work [70] to our method proposes to optimize a NeRF using only a single image and its corresponding depth map. However, it renders limited views from a small range of angles, and the prerequisite of a high-quality" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4479" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.395, + 0.107 + ], + "angle": 0, + "content": "depth map largely constrains its practical usage." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.108, + 0.47, + 0.44 + ], + "angle": 0, + "content": "To address the above issues, we propose a novel framework, coined as NeuralLift-360, which aims to ease the creation of 3D assets by building a bridge to convert diverse in-the-wild 2D photos to sophisticated 3D contents in \\(360^{\\circ}\\) views and enable its automation. The major challenge in our work is that the content on the back side is hidden and hard to hallucinate. To tackle these hurdles, we consider the diffusion priors together with the monocular depth estimator as the cues for hallucination. Modern diffusion models [41,48,50] are trained on a massive dataset (e.g., 5B text-to-image pairs [24]). During inference time, they can generate impressive photorealistic photos based on simple text inputs. By adopting these learning priors with CLIP [40] guidance, NeuralLift-360 can generate plausible 3D consistent instances that correspond well to the given reference image while only requiring minimal additional input, the correct user prompts. Moreover, rather than simply taking the depth map from the pre-trained depth estimator as geometry supervision, we propose to use the relative ranking information from the rough depth during the training process. This simple strategy is observed to robustly mitigate geometry errors for depth estimations in the wild." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.441, + 0.421, + 0.455 + ], + "angle": 0, + "content": "Our contributions can be summarized as follows," + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.473, + 0.469, + 0.534 + ], + "angle": 0, + "content": "- Given a single image in the wild, we demonstrate promising results of them being lifted to 3D. We use NeRF as an effective scene representation and integrate prior knowledge from the diffusion model." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.548, + 0.469, + 0.593 + ], + "angle": 0, + "content": "- We propose a CLIP-guided sampling strategy that effectively marries the prior knowledge from the diffusion model with the reference image." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.608, + 0.469, + 0.668 + ], + "angle": 0, + "content": "- When the reference image is hard to describe exactly, we finetune the diffusion model on the single image while maintaining its ability to generate diverse contents to guide the NeRF training." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.682, + 0.469, + 0.743 + ], + "angle": 0, + "content": "- We introduce scale-invariant depth supervision that uses the ranking information. This design alleviates the need for accurate multi-view consistent depth estimation and broadens the application of our algorithms." + }, + { + "type": "list", + "bbox": [ + 0.097, + 0.473, + 0.469, + 0.743 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.76, + 0.227, + 0.777 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.786, + 0.282, + 0.803 + ], + "angle": 0, + "content": "2.1.3D from Single Image" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Inferring 3D worlds from a single image has been a popular direction over the past years. Previous works try to solve this problem by firstly adopting a monocular depth estimator (e.g., MiDaS [44]) to predict the 3D geometry and then using multi-plane images (MPI) [20, 52] or point cloud [36] to render artistic effects. Meanwhile, the missing" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.409 + ], + "angle": 0, + "content": "holes from novel views are inpainted by a pre-trained modern deep network [76]. While these approaches are powerful and efficient, the depth estimator could be unstable, and the inpainting techniques may produce artifacts that make the rendered images look fake. While our method also applies a monocular depth estimator as the geometry cues, we propose a depth ranking loss that can mitigate the instability of depth estimation in the wild. Later approaches [28, 75] train their model on a large-scale 3D assets dataset. During inference time, they directly predict a 3D object using a single image. However, their performance faces drastic degradation when being applied to an in-the-wild image, due to the domain gap. The most similar work to ours is SinNeRF [70], which trains a neural radiance field only on a single RGB image and its corresponding depth map. However, their approach requires a high-quality depth map for geometry guidance, and it can only render views from a small range of angles. Thanks to the proposed depth ranking loss, our method can reconstruct impressive geometry information only using guidance from a pre-trained monocular depth estimator." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.417, + 0.716, + 0.432 + ], + "angle": 0, + "content": "2.2. Neural Radiance Fields" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.441, + 0.892, + 0.637 + ], + "angle": 0, + "content": "Neural Radiance Fields (NeRFs) [34] have shown encouraging performance for view synthesis. After its origin, tremendous efforts have been devoted to improving it [2,3,5,6,11,14,16,18,31,37,45,57,59,61,63,67,75]. Unlike most NeRF-like models, which at least require several different views to reconstruct scenes, we focus on reconstructing 3D content with single views. Analogously, recent works [18, 39] try to generate 3D content from text inputs directly. Although they also rely on the text prompt as guidance, their rendered objects are only conditioned on text sentences without image constraints, which can not reach our goal. Different from them, we aim at lifting a single image (either real or synthetic) to photorealistic 3D objects." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.65, + 0.638, + 0.665 + ], + "angle": 0, + "content": "3. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.892, + 0.81 + ], + "angle": 0, + "content": "Neural Radiance Field NeRFs [34] render images by sampling 5D coordinates (location \\((x,y,z)\\) and viewing direction \\((\\theta ,\\phi)\\)) along camera rays. After mapping them to color \\((r,g,b)\\) and volume density \\(\\sigma\\), additional volumetric rendering techniques are adopted to fuse the information along the ray. Given \\(\\mathbf{o}\\) as the camera origin and \\(\\pmb{d}\\) as the ray direction, 3D point locations along the ray can be expressed as \\(\\pmb {r}(t) = \\pmb {o} + td\\) The predicted color accumulated is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.818, + 0.892, + 0.854 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {C} (\\boldsymbol {r}) = \\int_ {t _ {n}} ^ {t _ {f}} T (t) \\sigma (\\boldsymbol {r} (t)) \\boldsymbol {c} (\\boldsymbol {r} (t), \\boldsymbol {d}) d t, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.864, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where \\( T(t) = \\exp \\left(-\\int_{t_n}^t\\sigma (\\boldsymbol {r}(s))ds\\right) \\), \\( \\sigma (\\cdot) \\) and \\( c(\\cdot ,\\cdot) \\) are densities and color predictions. To numerically evaluate" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4480" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.165, + 0.088, + 0.817, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.339, + 0.895, + 0.369 + ], + "angle": 0, + "content": "Figure 1. Overview of the proposed pipeline. NeuralLift-360 learns to recover a 3D object from a single reference image \\( y \\). We derive a prior distillation loss \\( \\mathcal{L}_{\\mathrm{diff}} \\) for CLIP-guided diffusion prior and utilize rough depth estimation with the help of a ranking loss \\( \\mathcal{L}_{\\mathrm{ranking}} \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.379, + 0.471, + 0.442 + ], + "angle": 0, + "content": "Eq. 1, the continuous integral is estimated using quadrature [34]. Our model is built upon Instant-NGP [37], which reduces computation cost by inferring color and density from a multiresolution hash table implemented by CUDA." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.462, + 0.47, + 0.691 + ], + "angle": 0, + "content": "Denoising Diffusion Model Denoising diffusion models generate images by gradually denoising from a gaussian noise \\( p(\\mathbf{x}_T) = \\mathcal{N}(\\mathbf{0},\\mathbf{I}) \\) and transforming into a certain data distribution. The forward diffusion process \\( q(\\pmb {x}_t|\\pmb {x}_{t - 1}) \\) adds Gaussian noise to the image \\( \\pmb {x}_t \\). The marginal distribution can be written as: \\( q(\\pmb {x}_t\\mid \\pmb {x_0}) = \\mathcal{N}\\left(\\alpha_t\\pmb {x_0},\\sigma_t^2\\pmb {I}\\right) \\) where \\( \\alpha_{t} \\) and \\( \\sigma_{t} \\) are designed to converge to \\( \\mathcal{N}(\\mathbf{0},\\mathbf{I}) \\) when \\( t \\) is at the end of the forward process [23, 56]. The reverse diffusion process \\( p(\\pmb{x}_{t - 1}|\\pmb {x}_t) \\) learns to denoise. Given infinitesimal timesteps, the reverse diffusion process can be approximated with Gaussian [56] related with an optimal MSE denoiser [53]. The diffusion models are designed as noise estimators \\( \\epsilon_{\\theta}(\\pmb {x}_t,t) \\) taking noisy images as input and estimating the noise. They are trained via optimizing the weighted evidence lower bound (ELBO) [17, 23]:" + }, + { + "type": "equation", + "bbox": [ + 0.098, + 0.701, + 0.47, + 0.728 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {E L B O}} (\\theta) = \\mathbb {E} \\left[ w (t) \\| \\boldsymbol {\\epsilon} _ {\\theta} (\\alpha_ {t} \\pmb {x} _ {0} + \\sigma_ {t} \\pmb {\\epsilon}; t) - \\boldsymbol {\\epsilon} \\| _ {2} ^ {2} \\right], \\quad (2)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.738, + 0.469, + 0.83 + ], + "angle": 0, + "content": "where \\(\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})\\), \\(w(t)\\) is a weighting function. In practice, setting \\(w(t) = 1\\) delivers good performance [17]. Sampling from a diffusion model can be either stochastic [17] or deterministic [54]. After sampling \\(\\boldsymbol{x}_T \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})\\), we can gradually reduce the noise level and reach a clean image with high quality at the end of the iterative process." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.845, + 0.169, + 0.86 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Overview. In this section, we propose our method, dubbed NeuralLift-360, which reconstructs a 3D object" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.379, + 0.892, + 0.546 + ], + "angle": 0, + "content": "from a single image. An overview of NeuralLift-360 is illustrated in Fig. 1. NeuralLift-360 combines the best of the two worlds: NeRF and diffusion model, where the former is leveraged to universally model a 3D scene and the latter is utilized to hallucinate the unseen views. To inject knowledge of a diffusion model into a radiance field, we derive a training framework in Sec. 4.1, whose supervision consists of two parts: (1) a prior distillation loss from a generative model (Sec. 4.2), (2) a straight loss on the given image (Sec. 4.4). Moreover, we introduce a domain adaption technique to adapt the diffusion prior to in-the-wild images (Sec. 4.3)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.554, + 0.779, + 0.57 + ], + "angle": 0, + "content": "4.1. Probabilistic-Driven 3D Lifting" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.578, + 0.892, + 0.684 + ], + "angle": 0, + "content": "We begin by formulating our setting along the way, introducing our notations. Given an image \\(\\pmb{y} \\in \\mathbb{R}^N\\) and its text description \\(\\pmb{z} \\in \\mathbb{R}^D\\), NeuralLift-360 intends to reconstruct a 3D scene \\(\\pmb{V} \\in \\mathbb{R}^M\\), where \\(\\pmb{V}\\) can be either a radiance volume or the network parameters of a NeRF. We regard \\(\\pmb{y}\\), \\(\\pmb{z}\\) and \\(\\pmb{V}\\) as random variables. To reconstruct \\(\\pmb{V}\\) from \\(\\pmb{y}\\) and \\(\\pmb{z}\\), we maximize a log-posterior and apply Bayesian rule:" + }, + { + "type": "equation", + "bbox": [ + 0.509, + 0.693, + 0.892, + 0.725 + ], + "angle": 0, + "content": "\\[\n\\log p (\\boldsymbol {V} | \\boldsymbol {y}, \\boldsymbol {z}) = \\log p (\\boldsymbol {y} | \\boldsymbol {V}, \\boldsymbol {z}) + \\log p (\\boldsymbol {V} | \\boldsymbol {z}) + c o n s t. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.826 + ], + "angle": 0, + "content": "One can see the prior term \\( p(\\mathbf{V}|z) \\) provides complementary information to lift an image into 3D space. However, specifying this prior can be intractable, and directly optimizing objective Eq. 3 is also sampling inefficient. Instead, we introduce camera pose \\( \\Phi \\in \\mathbb{S}\\mathbb{O}(3)\\times \\mathbb{R}^3 \\) as a latent variable, then the likelihood term can be rewritten as:" + }, + { + "type": "equation", + "bbox": [ + 0.516, + 0.833, + 0.892, + 0.899 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} p (\\boldsymbol {y} | \\boldsymbol {V}, \\boldsymbol {z}) = \\int p (\\boldsymbol {y} | \\boldsymbol {V}, \\boldsymbol {\\Phi}, \\boldsymbol {z}) p (\\boldsymbol {\\Phi} | \\boldsymbol {V}, \\boldsymbol {z}) d \\boldsymbol {\\Phi} (4) \\\\ = \\int p (\\boldsymbol {y} | \\boldsymbol {V}, \\boldsymbol {\\Phi}, \\boldsymbol {z}) \\frac {p (\\boldsymbol {V} | \\boldsymbol {\\Phi} , \\boldsymbol {z}) p (\\boldsymbol {\\Phi})}{p (\\boldsymbol {V} | \\boldsymbol {z})} d \\boldsymbol {\\Phi}, (5) \\\\ \\end{array}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "4481" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.138 + ], + "angle": 0, + "content": "where we assume \\(\\Phi\\) is independent of \\(z\\). Merging Eq. 3 and Eq. 5, we can get rid of the implicit prior term and obtain a sampling efficient ELBO of the original objective as below:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.147, + 0.469, + 0.183 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\log \\mathbb {E} _ {\\boldsymbol {\\Phi} \\sim p (\\boldsymbol {\\Phi})} p (\\boldsymbol {y} | \\boldsymbol {V}, \\boldsymbol {\\Phi}, \\boldsymbol {z}) p (\\boldsymbol {V} | \\boldsymbol {\\Phi}, \\boldsymbol {z}) (6) \\\\ \\geq \\mathbb {E} _ {\\boldsymbol {\\Phi} \\sim p (\\boldsymbol {\\Phi})} \\left[ \\log p (\\boldsymbol {y} | \\boldsymbol {V}, \\boldsymbol {\\Phi}, \\boldsymbol {z}) + \\log p (\\boldsymbol {V} | \\boldsymbol {\\Phi}, \\boldsymbol {z}) \\right]. (7) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.191, + 0.469, + 0.252 + ], + "angle": 0, + "content": "Furthermore, we define \\( p(\\boldsymbol{y}|\\boldsymbol{V}, \\boldsymbol{\\Phi}, \\boldsymbol{z}) = p(\\boldsymbol{y}|h(\\boldsymbol{V}, \\boldsymbol{\\Phi}), \\boldsymbol{z}) \\) and \\( p(\\boldsymbol{V}|\\boldsymbol{\\Phi}, \\boldsymbol{z}) = p(h(\\boldsymbol{V}, \\boldsymbol{\\Phi})|\\boldsymbol{z}) \\), where \\( h(\\boldsymbol{V}, \\boldsymbol{\\Phi}) \\) renders the 3D scene with respect to the camera pose \\( \\boldsymbol{\\Phi} \\). Then we derive our final training objective for NeuralLift-360:" + }, + { + "type": "equation", + "bbox": [ + 0.084, + 0.259, + 0.469, + 0.325 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = - \\mathbb {E} _ {\\boldsymbol {\\Phi}} \\left[ \\underbrace {\\log p (\\boldsymbol {y} \\mid h (\\boldsymbol {V} , \\boldsymbol {\\Phi}) , \\boldsymbol {z})} _ {\\text {r e f e r e n c e d l o s s}} + \\underbrace {\\log p (h (\\boldsymbol {V} , \\boldsymbol {\\Phi}) \\mid \\boldsymbol {z})} _ {\\text {n o n - r e f e r e n c e d l o s s}} \\right]. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.335, + 0.469, + 0.486 + ], + "angle": 0, + "content": "At each training step, we sample a camera pose with respect to a distribution (more on Sec. 5.1), and then render the view associated with this camera parameter (via volume rendering Eq. 1). The loss computed on the rendered image consists of two parts: image-referenced guidance and reference-free regularization. The referenced loss will enforce \\( V \\) to have a close appearance with given \\( y \\), while the non-referenced term hallucinates the image information with prior knowledge. In the next section, we introduce a diffusion model to implement our loss function Eq. 8." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.494, + 0.342, + 0.51 + ], + "angle": 0, + "content": "4.2. CLIP-guided Diffusion Priors" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.517, + 0.469, + 0.759 + ], + "angle": 0, + "content": "Diffusion models [17, 54-56], which synthesize images by gradually denoising from a gaussian noise, have undeniably been successful in the realm of image generation, editing, and processing tasks. When trained on large-scale datasets for text-to-image tasks, these models can accurately represent the distribution of images based on textual descriptions [41, 42, 48]. Consequently, we are inspired to utilize a diffusion model's generative prior to complement the 3D information. To accomplish this, we re-parameterize our Eq. 8 with a diffusion model. Nevertheless, the derived loss function contains both discriminative and generative terms. Taking inspiration from classifier guidance techniques [26, 30, 56], we show that we can leverage the training loss of a diffusion model to surrogate the generative part and utilize CLIP space similarity to model the discriminative part." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.759, + 0.469, + 0.79 + ], + "angle": 0, + "content": "Following the derivation in [17] and [32], we lower bound Eq. 8 by:" + }, + { + "type": "equation", + "bbox": [ + 0.08, + 0.799, + 0.472, + 0.863 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\log p _ {\\theta} (\\boldsymbol {y} | \\boldsymbol {x}, \\boldsymbol {z}) + \\log p _ {\\theta} (\\boldsymbol {x} | \\boldsymbol {z}) \\\\ \\geq - \\mathbb {E} _ {t, \\epsilon \\sim \\mathcal {N} (\\mathbf {0}, \\mathbf {I})} \\left[ \\mathcal {D} _ {\\mathrm {K L}} \\left(q \\left(\\boldsymbol {x} _ {t - 1} \\mid \\boldsymbol {x} _ {t}, \\boldsymbol {x} _ {0}\\right) \\| p _ {\\theta} \\left(\\boldsymbol {x} _ {t - 1} \\mid \\boldsymbol {x} _ {t}, \\boldsymbol {y}, \\boldsymbol {z}\\right) \\right] \\right. \\\\ = - \\mathbb {E} _ {t, \\boldsymbol {\\epsilon} \\sim \\mathcal {N} (\\mathbf {0}, \\boldsymbol {I})} \\left[ w (t) \\| \\nabla \\log p _ {\\theta} (\\boldsymbol {x} _ {t} | \\boldsymbol {y}, \\boldsymbol {z}) - \\boldsymbol {\\epsilon} \\| _ {2} ^ {2} \\right], \\tag {9} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.902 + ], + "angle": 0, + "content": "where \\(\\pmb{x}_t = \\alpha_t\\pmb{x}_0 + \\sigma_t\\pmb{\\epsilon}\\) and \\(w(t)\\) is a constant dependent on noise scale \\(\\alpha_{t}\\) and \\(\\sigma_t\\) . The expectation of \\(t\\sim \\mathcal{U}\\{1,\\dots ,T\\}\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.198 + ], + "angle": 0, + "content": "is usually taken over a uniform distribution. We provide detailed derivation in supplementary. At first glance, we may directly plug in the pre-trained denoiser \\(\\epsilon_{\\theta}(\\pmb{x}_t;\\pmb{z})\\) of a diffusion model to surrogate \\(\\nabla p_{\\theta}(\\pmb{x}_t|\\pmb{y},\\pmb{z})\\) similar to [39]. However, note that \\(p_{\\theta}(\\pmb{x}_t|\\pmb{y},\\pmb{z})\\) is also conditioned on the reference image \\(\\pmb{y}\\), which requires further efforts to inject the reference image information into the diffusion process." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.199, + 0.892, + 0.304 + ], + "angle": 0, + "content": "To this end, NeuralLift-360 leverages a Mixed Reference Guidance that guides the diffusion process with a mixture of references. Similar to classifier guidance techniques [26, 30, 32, 56], we rewrite \\( p_{\\theta}(\\boldsymbol{x}_t|\\boldsymbol{y}, \\boldsymbol{z}) = p_{\\theta}(\\boldsymbol{y}|\\boldsymbol{x}_t, \\boldsymbol{z})p_{\\theta}(\\boldsymbol{x}_t|\\boldsymbol{z}) / p_{\\theta}(\\boldsymbol{y}|\\boldsymbol{z}) \\), and we adopt classifier-free guidance to model \\( p_{\\theta}(\\boldsymbol{x}_t|\\boldsymbol{z}) \\), then we can obtain a modified score function [55] as below:" + }, + { + "type": "equation", + "bbox": [ + 0.547, + 0.319, + 0.891, + 0.356 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\boldsymbol {\\epsilon} _ {\\theta} (\\boldsymbol {x} _ {t}; \\boldsymbol {y}, \\boldsymbol {z}) = (1 + w) \\boldsymbol {\\epsilon} _ {\\theta} (\\boldsymbol {x} _ {t}; \\boldsymbol {z}) - w \\boldsymbol {\\epsilon} _ {\\theta} (\\boldsymbol {x} _ {t}) \\\\ + \\sqrt {1 - \\alpha_ {t}} \\nabla \\log p _ {\\theta} (\\boldsymbol {y} | \\boldsymbol {x} _ {t}, \\boldsymbol {z}), \\tag {10} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.37, + 0.892, + 0.519 + ], + "angle": 0, + "content": "where \\(w\\) is a hyper-parameter that controls the strength of the classifier guidance. We further measure the discrimination with a distance metric on the feature space \\(p_{\\theta}(\\boldsymbol{y}|\\boldsymbol{x}_t, \\boldsymbol{z}) \\propto \\exp\\left[-\\phi\\left(\\frac{\\boldsymbol{x}_t - \\sigma_t \\boldsymbol{\\epsilon}_{\\theta}(\\boldsymbol{x}_t; \\boldsymbol{z})}{\\alpha_t}, \\boldsymbol{y}\\right)\\right]\\). Specifically, we choose the inner product on the CLIP embedding space [40] as the distance metric, i.e., \\(\\log p_{\\theta}(\\boldsymbol{y}|\\boldsymbol{x}_t, \\boldsymbol{z}) \\propto -\\left\\langle F\\left(\\frac{\\boldsymbol{x}_t - \\sigma_t \\boldsymbol{\\epsilon}_{\\theta}(\\boldsymbol{x}_t)}{\\alpha_t}\\right), F(\\boldsymbol{y})\\right\\rangle\\), which spurs on semantic similarity between denoised rendered image and reference image \\(\\boldsymbol{y}\\), where \\(F\\) is the CLIP image encoder." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.52, + 0.892, + 0.656 + ], + "angle": 0, + "content": "In each training iteration, we alternate two camera pose sampling strategies: (1) we fix a camera pose \\(\\Phi_0\\) (i.e., a Delta distribution) as the reference view, and when rendering on that view, we only penalize the RGB differences by modeling \\(p(\\boldsymbol{y} | h(\\boldsymbol{V}, \\boldsymbol{\\Phi}), \\boldsymbol{z}) = \\mathcal{N}(\\boldsymbol{y} | h(\\boldsymbol{V}, \\boldsymbol{\\Phi}), \\sigma^2 \\boldsymbol{I})\\), (2) we stochastically sample camera poses, and adopt the diffusion ELBO (Eq. 9) to estimate the evidence of rendered image together with prior \\(p_{\\theta}(\\boldsymbol{x} | \\boldsymbol{z})\\). Combining all together, we can summarize our training loss as below:" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.668, + 0.902, + 0.76 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} = \\frac {1}{\\sigma^ {2}} \\underbrace {\\| h (\\boldsymbol {V} , \\boldsymbol {\\Phi} _ {0}) - \\boldsymbol {y} \\| _ {2} ^ {2}} _ {\\mathcal {L} _ {\\text {p h o t o m e t r i c}}} \\tag {11} \\\\ + \\underbrace {\\mathbb {E} _ {\\boldsymbol {\\Phi} , t , \\boldsymbol {\\epsilon} \\sim \\mathcal {N} (\\boldsymbol {0} , \\boldsymbol {I})} \\left[ w (t) \\| \\boldsymbol {\\epsilon} _ {\\theta} (\\alpha_ {t} h (\\boldsymbol {V} , \\boldsymbol {\\Phi}) + \\sigma_ {t} \\boldsymbol {\\epsilon} | \\boldsymbol {y} , \\boldsymbol {z}) - \\boldsymbol {\\epsilon} \\| _ {2} ^ {2} \\right]} _ {\\mathcal {L} _ {\\mathrm {d i f f}}}. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.773, + 0.892, + 0.834 + ], + "angle": 0, + "content": "Compared with DreamFusion [39], NeuralLift-360's training objective grounds the 3D scene with the given image which enables the generated scene to have consistent appearance with the specified object." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.846, + 0.843, + 0.863 + ], + "angle": 0, + "content": "4.3. Domain Adaption to In-the-wild Images" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "A primary challenge, however, lies in assigning an appropriate text prompt for our reference image. Given that" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4482" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.156, + 0.145 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.159, + 0.089, + 0.234, + 0.145 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.237, + 0.089, + 0.312, + 0.145 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.315, + 0.089, + 0.391, + 0.145 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.393, + 0.089, + 0.469, + 0.145 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.197, + 0.147, + 0.274, + 0.159 + ], + "angle": 0, + "content": "(a) Pre-trained" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.16, + 0.156, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.158, + 0.16, + 0.233, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.16, + 0.311, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.314, + 0.16, + 0.388, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.391, + 0.16, + 0.465, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.235, + 0.221, + 0.311, + 0.233 + ], + "angle": 0, + "content": "(c) Fine-tuned" + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.239, + 0.468, + 0.266 + ], + "angle": 0, + "content": "Figure 2. After fine-tuning, our model can generate diverse images similar to the reference image (b) while not identical." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.28, + 0.468, + 0.355 + ], + "angle": 0, + "content": "large diffusion models are predominantly trained for text-to-image tasks, they typically require a text embedding as conditional input. For images in the wild, it is generally hard to exactly describe an image since images contain more fine-grained information than texts." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.356, + 0.468, + 0.581 + ], + "angle": 0, + "content": "Adapting to in-the-wild images necessitates finetuning the Stable Diffusion model on the reference image Textual inversion [15] and Dreambooth [49] enable finetuning the Stable Diffusion model on a limited number of images for a subject-driven generation. However, in our situation, we only have access to a single image, which significantly hampers the effectiveness of these methods. Imagic [22] and UniTune [60] support finetuning on a single image, but their primary objective is to perform real image editing, which can be quite successful even if the diffusion model overfits the given image. Nevertheless, in our context, we need the diffusion model to generate images resembling the provided reference but not identical to it. If the supervision for each view is the same, NeRF will otherwise produce an isotropic surface." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.582, + 0.468, + 0.704 + ], + "angle": 0, + "content": "We draw inspiration from Pivotal Tuning [47] and Imagic [22] when fine-tuning Stable Diffusion [48] on our own. To start with, we optimize the CLIP text embedding \\( z \\) so that it fits our reference image \\( y \\) better. During fine-tuning, we use Eq. 2 to supervise the noise estimate. As we expect that our denoised result will resemble the reference image, we also apply a CLIP image loss to the estimated denoised image. \\( \\hat{x}_0 \\)," + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.712, + 0.468, + 0.772 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {f i n e t u n e}} = \\mathbb {E} _ {t, \\epsilon} \\left[ \\| \\boldsymbol {\\epsilon} - \\boldsymbol {\\epsilon} _ {\\theta} (\\mathbf {x} _ {t}; \\boldsymbol {z}) \\| _ {2} ^ {2} \\right] \\\\ - \\mathbb {E} _ {t, \\epsilon} \\left\\langle F \\left(\\frac {\\boldsymbol {x} _ {t} - \\sigma_ {t} \\boldsymbol {\\epsilon} _ {\\theta} (\\boldsymbol {x} _ {t} ; \\boldsymbol {z})}{\\alpha_ {t}}\\right), F (\\boldsymbol {y}) \\right\\rangle , \\tag {12} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.78, + 0.412, + 0.795 + ], + "angle": 0, + "content": "where \\(F(\\cdot)\\) is the pre-trained CLIP image encoder." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.468, + 0.9 + ], + "angle": 0, + "content": "Then, we fix the optimized text embedding \\( z^{*} \\) and finetune the diffusion model \\( \\epsilon_{\\theta}(\\cdot) \\) using the same loss function as above. To preserve the generalization ability of the diffusion model, we introduce several augmentations during the finetuning. Specifically, in each iteration, we perform random image-level augmentations such as flip, transpose, shift, scale, and rotate. As shown in Fig. 2, we are able" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.228 + ], + "angle": 0, + "content": "to overcome the overfitting issue and preserve the diffusion model's ability to generate diverse similar but not identical contents. Similar to Imagic [22] and UniTune [47], we observe that after finetuning, the optimized text embedding will have an overfitting issue and it is critical to perform interpolation between the original text embedding and the optimized text embedding to preserve the diffusion model's ability to generate diverse contents. We obtain \\( z' = \\eta z^* + (1 - \\eta)z \\), where \\( \\eta \\) is a weighting hyper-parameter." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.228, + 0.892, + 0.333 + ], + "angle": 0, + "content": "During sampling, we aim to amplify the conditional likelihood for our optimized embedding \\( z^{*} \\), while decreasing both the unconditional likelihood and the conditional likelihood for the original embedding \\( z \\). Hence for 50% of all iterations, we use the original Eq. 10, while for the others, we replace the unconditional sampling \\( \\epsilon_{\\theta}(\\boldsymbol{x}_t) \\) in Eq. 10 with conditional sampling using the original text embedding \\( z \\):" + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.345, + 0.891, + 0.382 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\widetilde {\\boldsymbol {\\epsilon}} _ {\\theta} \\left(\\boldsymbol {x} _ {t}; \\boldsymbol {y}, \\boldsymbol {z} ^ {\\prime}, \\boldsymbol {z}\\right) = (1 + w) \\boldsymbol {\\epsilon} _ {\\theta} \\left(\\boldsymbol {x} _ {t}; \\boldsymbol {z} ^ {\\prime}\\right) \\tag {13} \\\\ - w \\epsilon_ {\\theta} (\\boldsymbol {x} _ {t}; \\boldsymbol {z}) + \\sqrt {1 - \\alpha_ {t}} \\nabla \\log p _ {\\theta} (\\boldsymbol {y} | \\boldsymbol {x} _ {t}, \\boldsymbol {z} ^ {\\prime}), \\\\ \\end{array}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.394, + 0.776, + 0.41 + ], + "angle": 0, + "content": "4.4. Supervision from Rough Depth" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.418, + 0.892, + 0.567 + ], + "angle": 0, + "content": "While the aforementioned pipeline laid the foundation for the success of our method, we observe that obtaining a high-quality 3D object without geometry regularization is inherently challenging, if relying solely on learning priors. As shown in SinNeRF [70] and DS-NeRF [8], accurate depth information can largely help improve the rendering quality. However, their impressive outcomes are contingent upon the 3D consistency of the accessible high-quality depth, a requirement that is not easily fulfilled in real-life scenarios." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.569, + 0.892, + 0.749 + ], + "angle": 0, + "content": "In spite of the significant progress in monocular depth estimation for in-the-wild scenarios, current state-of-the-art depth estimators are insufficient for recovering 3D shapes. Due to unknown camera baselines and stereoscopic post-processing in data, these methods [7,25,44,64,69,72] utilize loss functions invariant to shift and scale or adopt ranking loss that utilizes relative information. Therefore, their results cannot be used to reconstruct plausible 3D scene shapes directly due to unknown depth shifts [74]. Though there are also attempts [74] that improve off-the-shelf depth estimators, they generally can not generalize well to unseen objects in the wild." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Given that the acquired depth is imperfect, the absolute depth scale lacks the reliability necessary for recovering a high-quality 3D object. In contrast to previous depth-supervised NeRFs [8, 70], which rely solely on precise values from the depth map, our approach represents the first attempt to maximize the utilization of depth information by adopting the relative ranking information. Drawing inspiration from [7, 38], we adopt a pairwise depth ranking loss to supervise the rendered depth from our NeRF backbone. In this way, we enforce the depth we rendered to be consistent" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.515, + 0.957 + ], + "angle": 0, + "content": "4483" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.136 + ], + "angle": 0, + "content": "with the depth from the reference view in terms of internal ranking instead of actual value. The loss is formulated as follows," + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.142, + 0.47, + 0.181 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r a n k i n g}} = \\left\\{ \\begin{array}{l l} \\max \\left(z _ {j _ {k}} - z _ {i _ {k}}, 0\\right), & r _ {k} \\in \\{> \\} \\\\ \\max \\left(z _ {i _ {k}} - z _ {j _ {k}}, 0\\right), & r _ {k} \\in \\{< \\} \\end{array} \\right. \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.186, + 0.469, + 0.262 + ], + "angle": 0, + "content": "where \\( r_k \\) refers to the pairwise ranking relationship between pixel \\( z_{i_k} \\) and \\( z_{j_k} \\) in the monocular depth estimation. This loss function is used alongside the photo-metric loss and diffusion loss mentioned in Eq. 11 and serves as additional supervision on geometry." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.269, + 0.21, + 0.285 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.291, + 0.469, + 0.335 + ], + "angle": 0, + "content": "In this section, we will first describe a unique training recipe that we propose to train our framework, then the implementation details." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.339, + 0.238, + 0.355 + ], + "angle": 0, + "content": "5.1. Training Recipe" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.358, + 0.47, + 0.646 + ], + "angle": 0, + "content": "Diverse Camera Pose and Intrinsic Sampling During training, we observe that a diverse camera pose and intrinsicics are essential for the model to learn a coherent 3D object. If the camera pose is not sampled from a diverse distribution, the 3D object will collapse towards the major viewing directions and thus generate unpleasant results. Specifically, we alter between two pose sampling strategies; (1) we have a fixed camera pose \\(\\Phi_0\\) (i.e., a Delta distribution) for the reference view, (2) we uniformly sample the camera rotation on the surface of a unit sphere and then uniformly sample the radius in [0.4, 1.0] to locate the camera. We start with the Delta distribution, and as training progresses, we gradually shift to the other sophisticated camera pose distribution. The camera intrinsics provide additional augmentation on the viewing direction. By uniformly sampling the field of view in [50, 70], we zoom in and out to get different resolution projections for the scene. We also add Gaussian noise to the camera pose sampling phase. Such perturbations help us render more diverse images during training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.663, + 0.469, + 0.859 + ], + "angle": 0, + "content": "Foreground Aware Diffusion Prior ERNIE-ViLG [13] utilizes a weighting mechanism during the training of diffusion models to let the model generate. Using an additional object detection model, their weighted loss forces the model to focus on salient regions. Similarly, we hope to strengthen the guidance on the foreground object region. We propose to weight the score matching distance with the predicted density. Specifically, we penalize the regions with higher density so that our NeRF gains more guidance on the foreground regions. Hence, for each rendered image, we first obtain the bounding box for the non-zero density regions. Then, we extend the \\( w(t) \\) in \\( \\mathcal{L}_{\\mathrm{diff}} \\) (Eq. 11) into a spatial function:" + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.865, + 0.469, + 0.905 + ], + "angle": 0, + "content": "\\[\nW _ {i, j} = \\left\\{ \\begin{array}{l l} 2, & \\text {i f} (i, j) \\text {i s i n s i d e t h e b o u n d i n g b o x} \\\\ 1, & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.243 + ], + "angle": 0, + "content": "Geometry Regularization We also include a series of geometry regularizations to avoid undesired artifacts in our 3D object. To avoid foggy and semi-transparent artifacts at the back of the scene, we utilize the orient loss from Ref-NeRF [61] to prevent the surface normals from facing backward from the camera. We also incorporate depth smoothness loss [62], sparsity loss, and distortion loss [3, 58] to avoid the NeRF filling empty space with floaters. Full details for these regularizations are provided in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.25, + 0.892, + 0.385 + ], + "angle": 0, + "content": "Timestep Annealing At the start of training, the quality of the rendered images is poor, and we need the diffusion prior to provide more thorough guidance. As a result, we need to add more noise to perturb the rendered image. As the training progresses, the quality of rendered images gradually improves, and we choose to reduce the timestep range accordingly. By annealing the timestep where we add noise, we are lowering the noise level of the perturbation, and the noise estimate becomes more accurate." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.394, + 0.699, + 0.408 + ], + "angle": 0, + "content": "5.2. Evaluation Workflow" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.417, + 0.892, + 0.628 + ], + "angle": 0, + "content": "We conduct extensive experiments using both text-to-image synthetic images and real-world images. Synthetic images in our experiments are generated using PNDM Sampler [29] with 50 sampling steps. Real images are either from CO3Dv2 dataset [46] or randomly crawled from the Google search engine. For each image, we use a pre-trained background removal framework [1] to segment the foreground objects. Then we obtain their depth using LeRes [73] with the \"Boost Your Own depth\" [33] strategy. [33] inferences the monocular depth estimation network on multiple patches and then fuses the results for better details. We compare with previous sparse view NeRFs, including Depth-supervised NeRF [8] and SinNeRF [70]. We also compare with DietNeRF [19] that uses CLIP as guidance." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.637, + 0.64, + 0.652 + ], + "angle": 0, + "content": "5.3. Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.892, + 0.795 + ], + "angle": 0, + "content": "We present quantitative results in eight scenes. We report the average CLIP distance between rendered image and the reference image as a metric for how well the object is represented. During testing, we use a CLIP model different from training time and evaluate on held out viewing directions with a larger distance-to-origin than training. We render 100 images on each scene for each method. As shown in Tab. 1, our method achieves the best CLIP distance among existing approaches." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We also provide visual comparisons in Fig. 3. Since DSNeRF and SinNeRF rely on the exact value of the available depth information, their performance using monocular depth estimation is far from satisfactory. It can also be seen that the DietNeRF suffers from unconstrained geometry. Our method, on the other hand, is able to synthesize visually pleasing novel views with 3D consistency." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4484" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.088, + 0.87, + 0.874 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.15, + 0.883, + 0.819, + 0.899 + ], + "angle": 0, + "content": "Figure 3. Visual comparisons for novel view synthesis results. We provide video comparisons in supplementary." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.515, + 0.957 + ], + "angle": 0, + "content": "4485" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.158, + 0.09, + 0.812, + 0.104 + ], + "angle": 0, + "content": "Table 1. Quantitative evaluation by comparing the proposed method with previous state-of-the-art approaches." + }, + { + "type": "table", + "bbox": [ + 0.123, + 0.115, + 0.849, + 0.186 + ], + "angle": 0, + "content": "
CLIP Distance ↓AppleBenchBaseballHydrantRemoteSuitcaseStatueSandwichAverage
DSNeRF [8]0.53090.54470.55060.62260.64410.55330.59920.61100.5821
DietNeRF [19]0.42450.55790.45450.51900.57180.48900.59470.51790.5162
SinNeRF [70]0.51920.57690.62300.63430.64960.57050.56050.60150.5918
NeuralLift-360 (Ours)0.39140.47850.39740.51460.49160.41860.44110.46550.4498
" + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.191, + 0.868, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.093, + 0.343, + 0.877, + 0.357 + ], + "angle": 0, + "content": "Figure 4. Ablation study on the baseline variants. Given a single reference image, we run different variants to render its \\(360^{\\circ}\\) views." + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.368, + 0.446, + 0.692 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.703, + 0.47, + 0.717 + ], + "angle": 0, + "content": "Figure 5. Ablation study on CLIP guidance for the diffusion prior." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.724, + 0.231, + 0.74 + ], + "angle": 0, + "content": "5.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Variants of supervision. In this section, we perform detailed ablation studies on different components of our framework. \"w/o \\(\\mathcal{L}_{\\mathrm{ranking}}\\) \" replaces the ranking loss with \\(L_{1}\\) loss on depth. \"w/o \\(\\mathcal{L}_{\\mathrm{diff}}\\) \" replaces the CLIP guided diffusion prior with CLIP. \"w/o fine-tuning\" refers to using the pre-trained diffusion model. As shown in Fig. 4, our full model delivers the best performance, while other variants have different issues. \"w/o \\(\\mathcal{L}_{\\mathrm{ranking}}\\) \" relies on the exact value from depth information and suffers from unreliable depth. \"w/o \\(\\mathcal{L}_{\\mathrm{diff}}\\) \" uses CLIP as a prior, and the optimized result is" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.369, + 0.892, + 0.43 + ], + "angle": 0, + "content": "blurry and collapsed. \"w/o fine-tuning\" uses a pre-trained diffusion model, since the text prompt is not very accurate, the resulting object has lettuce which doesn't exist in the reference image." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.462, + 0.893, + 0.568 + ], + "angle": 0, + "content": "Effectiveness of CLIP guided diffusion prior We validate the effectiveness of CLIP guidance in our diffusion prior design in Fig. 5. It can be seen that without the CLIP guidance, although the model produces a 3D object, it doesn't look like the reference image and has a strange shape. In comparison, our full model generates a better shape and looks more similar to the reference image." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.595, + 0.779, + 0.611 + ], + "angle": 0, + "content": "6. Conclusions and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.625, + 0.892, + 0.791 + ], + "angle": 0, + "content": "We present NeuralLift-360, a novel framework to lift an in-the-wild 2D photo into a 3D object with 360-degree views. NeuralLift-360 learns probabilistic-driven 3D lifting with CLIP-guided diffusion priors and mitigates the depth errors by a scale-invariant depth ranking loss. Comprehensive experiments are conducted on real and synthetic images, where NeuralLift-360 outperforms the current state-of-the-art methods. While our current implementation is based on Stable Diffusion [48] version 1.4, the framework should also work well with other diffusion models such as Imagen [50] or DALL-E2 [41], once they are open-sourced." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Limitations. Despite the encouraging visual results rendered by NeuralLift-360, the target resolution \\((128 \\times 128)\\) is still behind large generative models [41, 50]. Moreover, challenging cases (e.g., multiple objects with occlusion) are not included in our assumptions, which need to be further explored. Going forward, we will investigate how to expand our method to more general scenarios." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4486" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.156 + ], + "angle": 0, + "content": "[1] Image background removal. https://github.com/OPHoperHPO/image-background-remove-tool. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.16, + 0.472, + 0.241 + ], + "angle": 0, + "content": "[2] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.245, + 0.471, + 0.313 + ], + "angle": 0, + "content": "[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5470–5479, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.316, + 0.47, + 0.413 + ], + "angle": 0, + "content": "[4] Angel X. Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. ShapeNet: An Information-Rich 3D Model Repository. Technical Report arXiv:1512.03012 [cs.GR], Stanford University — Princeton University — Toyota Technological Institute at Chicago, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.415, + 0.47, + 0.484 + ], + "angle": 0, + "content": "[5] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14124-14133, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.486, + 0.469, + 0.555 + ], + "angle": 0, + "content": "[6] Tianlong Chen, Peihao Wang, Zhiwen Fan, and Zhangyang Wang. Aug-nerf: Training stronger neural radiance fields with triple-level physically-grounded augmentations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15191-15202, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.557, + 0.469, + 0.599 + ], + "angle": 0, + "content": "[7] Weifeng Chen, Zhao Fu, Dawei Yang, and Jia Deng. Single-image depth perception in the wild. Advances in neural information processing systems, 29, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.601, + 0.47, + 0.669 + ], + "angle": 0, + "content": "[8] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised nerf: Fewer views and faster training for free. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12882-12891, 2022. 1, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.672, + 0.469, + 0.727 + ], + "angle": 0, + "content": "[9] Jean-Denis Durou, Maurizio Falcone, and Manuela Sagona. Numerical methods for shape-from-shading: A new survey with benchmarks. Computer Vision and Image Understanding, 109(1):22-43, 2008. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.729, + 0.469, + 0.784 + ], + "angle": 0, + "content": "[10] Haoqiang Fan, Hao Su, and Leonidas J Guibas. A point set generation network for 3d object reconstruction from a single image. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 605-613, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.786, + 0.469, + 0.828 + ], + "angle": 0, + "content": "[11] Zhiwen Fan, Yifan Jiang, Peihao Wang, Xinyu Gong, Dejia Xu, and Zhangyang Wang. Unified implicit neural stylization. arXiv preprint arXiv:2204.01943, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.83, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[12] Paolo Favaro and Stefano Soatto. A geometric approach to shape from defocus. IEEE Transactions on Pattern Analysis and Machine Intelligence, 27(3):406-417, 2005. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.902 + ], + "angle": 0, + "content": "[13] Zhida Feng, Zhenyu Zhang, Xintong Yu, Yewei Fang, Lanxin Li, Xuyi Chen, Yuxiang Lu, Jiaxiang Liu, Weichong" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "Yin, Shikun Feng, et al. Ernie-vilg 2.0: Improving text-to-image diffusion model with knowledge-enhanced mixture-of-denoising-experts. arXiv preprint arXiv:2210.15257, 2022.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.149, + 0.894, + 0.218 + ], + "angle": 0, + "content": "[14] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5501–5510, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.219, + 0.892, + 0.286 + ], + "angle": 0, + "content": "[15] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.288, + 0.892, + 0.355 + ], + "angle": 0, + "content": "[16] Yuan-Chen Guo, Di Kang, Linchao Bao, Yu He, and Song-Hai Zhang. Nerfren: Neural radiance fields with reflections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18409-18418, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.357, + 0.892, + 0.399 + ], + "angle": 0, + "content": "[17] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.4, + 0.892, + 0.468 + ], + "angle": 0, + "content": "[18] Ajay Jain, Ben Mildenhall, Jonathan T Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 867-876, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.47, + 0.892, + 0.525 + ], + "angle": 0, + "content": "[19] Ajay Jain, Matthew Tancik, and Pieter Abbeel. Putting nerf on a diet: Semantically consistent few-shot view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5885-5894, 2021. 1, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.526, + 0.892, + 0.621 + ], + "angle": 0, + "content": "[20] Varun Jampani, Huiwen Chang, Kyle Sargent, Abhishek Kar, Richard Tucker, Michael Krainin, Dominik Kaeser, William T Freeman, David Salesin, Brian Curless, et al. Slide: Single image 3d photography with soft layering and depth-aware inpainting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12518-12527, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.623, + 0.892, + 0.691 + ], + "angle": 0, + "content": "[21] Adrian Johnston, Ravi Garg, Gustavo Carneiro, Ian Reid, and Anton van den Hengel. Scaling cnns for high resolution volumetric reconstruction from a single image. In Proceedings of the IEEE International Conference on Computer Vision Workshops, pages 939-948, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.692, + 0.892, + 0.748 + ], + "angle": 0, + "content": "[22] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagc: Text-based real image editing with diffusion models. arXiv preprint arXiv:2210.09276, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.748, + 0.892, + 0.789 + ], + "angle": 0, + "content": "[23] Diederik Kingma, Tim Salimans, Ben Poole, and Jonathan Ho. Variational diffusion models. Advances in neural information processing systems, 34:21696-21707, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.79, + 0.892, + 0.845 + ], + "angle": 0, + "content": "[24] Hyung-Kwon Ko, Gwanmo Park, Hyeon Jeon, Jaemin Jo, Juho Kim, and Jinwook Seo. Large-scale text-to-image generation models for visual artists' creative works. arXiv preprint arXiv:2210.08477, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[25] Manuel Lang, Alexander Hornung, Oliver Wang, Steven Poulakos, Aljoscha Smolic, and Markus Gross. Nonlinear disparity mapping for stereoscopic 3d. ACM Transactions on Graphics (TOG), 29(4):1-10, 2010. 5" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4487" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[26] Yann LeCun, Sumit Chopra, Raia Hadsell, M Ranzato, and Fujie Huang. A tutorial on energy-based learning. Predicting structured data, 1(0), 2006. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.134, + 0.47, + 0.19 + ], + "angle": 0, + "content": "[27] Zhengqi Li and Noah Snavely. Megadepth: Learning single-view depth prediction from internet photos. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2041-2050, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.19, + 0.47, + 0.246 + ], + "angle": 0, + "content": "[28] Kai-En Lin, Lin Yen-Chen, Wei-Sheng Lai, Tsung-Yi Lin, Yi-Chang Shih, and Ravi Ramamoorthi. Vision transformer for nef-based view synthesis from a single input image. arXiv preprint arXiv:2207.05736, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.246, + 0.47, + 0.287 + ], + "angle": 0, + "content": "[29] Luping Liu, Yi Ren, Zhijie Lin, and Zhou Zhao. Pseudo numerical methods for diffusion models on manifolds. arXiv preprint arXiv:2202.09778, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.287, + 0.47, + 0.344 + ], + "angle": 0, + "content": "[30] Xihui Liu, Dong Huk Park, Samaneh Azadi, Gong Zhang, Arman Chopikyan, Yuxiao Hu, Humphrey Shi, Anna Rohrbach, and Trevor Darrell. More control for free! image synthesis with semantic diffusion guidance. 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.344, + 0.47, + 0.426 + ], + "angle": 0, + "content": "[31] Yuan Liu, Sida Peng, Lingjie Liu, Qianqian Wang, Peng Wang, Christian Theobalt, Xiaowei Zhou, and Wenping Wang. Neural rays for occlusion-aware image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7824-7833, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.426, + 0.47, + 0.455 + ], + "angle": 0, + "content": "[32] Calvin Luo. Understanding diffusion models: A unified perspective. arXiv preprint arXiv:2208.11970, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.455, + 0.47, + 0.512 + ], + "angle": 0, + "content": "[33] S. Mahdi H. Miangoleh, Sebastian Dille, Long Mai, Sylvain Paris, and Yaqiz Aksoy. Boosting monocular depth estimation models to high-resolution via content-adaptive multi-resolution merging. 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.511, + 0.47, + 0.58 + ], + "angle": 0, + "content": "[34] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European conference on computer vision, pages 405-421. Springer, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.58, + 0.47, + 0.649 + ], + "angle": 0, + "content": "[35] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.65, + 0.47, + 0.718 + ], + "angle": 0, + "content": "[36] Fangzhou Mu, Jian Wang, Yicheng Wu, and Yin Li. 3d photo stylization: Learning to generate stylized novel views from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16273-16282, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.719, + 0.47, + 0.775 + ], + "angle": 0, + "content": "[37] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. arXiv preprint arXiv:2201.05989, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.47, + 0.831 + ], + "angle": 0, + "content": "[38] Georgios Pavlakos, Xiaowei Zhou, and Kostas Daniilidis. Ordinal depth supervision for 3d human pose estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7307-7316, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.47, + 0.873 + ], + "angle": 0, + "content": "[39] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.47, + 0.903 + ], + "angle": 0, + "content": "[40] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry," + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.092, + 0.892, + 0.149 + ], + "angle": 0, + "content": "Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[41] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.204, + 0.892, + 0.272 + ], + "angle": 0, + "content": "[42] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.272, + 0.892, + 0.33 + ], + "angle": 0, + "content": "[43] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12179-12188, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.33, + 0.892, + 0.399 + ], + "angle": 0, + "content": "[44] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE transactions on pattern analysis and machine intelligence, 2020. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.399, + 0.892, + 0.468 + ], + "angle": 0, + "content": "[45] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding up neural radiance fields with thousands of tiny mlps. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14335-14345, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.469, + 0.892, + 0.539 + ], + "angle": 0, + "content": "[46] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In International Conference on Computer Vision, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.539, + 0.892, + 0.593 + ], + "angle": 0, + "content": "[47] Daniel Roich, Ron Mokady, Amit H Bermano, and Daniel Cohen-Or. Pivotal tuning for latent-based editing of real images. ACM Transactions on Graphics (TOG), 42(1):1-13, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.593, + 0.892, + 0.663 + ], + "angle": 0, + "content": "[48] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2, 4, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.663, + 0.892, + 0.72 + ], + "angle": 0, + "content": "[49] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. arXiv preprint arXiv:2208.12242, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.72, + 0.892, + 0.803 + ], + "angle": 0, + "content": "[50] Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.803, + 0.892, + 0.873 + ], + "angle": 0, + "content": "[51] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2304–2314, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.903 + ], + "angle": 0, + "content": "[52] Meng-Li Shih, Shih-Yang Su, Johannes Kopf, and Jia-Bin Huang. 3d photography using context-aware layered depth" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4488" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.133 + ], + "angle": 0, + "content": " inpainting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8028-8038, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.134, + 0.47, + 0.202 + ], + "angle": 0, + "content": "[53] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256-2265. PMLR, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.204, + 0.471, + 0.245 + ], + "angle": 0, + "content": "[54] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.246, + 0.471, + 0.287 + ], + "angle": 0, + "content": "[55] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in Neural Information Processing Systems, 32, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.288, + 0.471, + 0.343 + ], + "angle": 0, + "content": "[56] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.344, + 0.471, + 0.399 + ], + "angle": 0, + "content": "[57] Mohammed Suhail, Carlos Esteves, Leonid Signal, and Ameesh Makadia. Light field neural rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8269-8279, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.4, + 0.471, + 0.44 + ], + "angle": 0, + "content": "[58] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Improved direct voxel grid optimization for radiance fields reconstruction. arxiv.cs.GR 2206.05085, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.442, + 0.471, + 0.509 + ], + "angle": 0, + "content": "[59] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5459-5469, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.511, + 0.471, + 0.566 + ], + "angle": 0, + "content": "[60] Dani Valevski, Matan Kalman, Yossi Matias, and Yaniv Leviathan. Unitune: Text-driven image editing by fine tuning an image generation model on a single image. arXiv preprint arXiv:2210.09477, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.567, + 0.471, + 0.648 + ], + "angle": 0, + "content": "[61] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T Barron, and Pratul P Srinivasan. Ref-nerf: Structured view-dependent appearance for neural radiance fields. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5481-5490. IEEE, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.65, + 0.471, + 0.718 + ], + "angle": 0, + "content": "[62] Chaoyang Wang, José Miguel Buenapasada, Rui Zhu, and Simon Lucey. Learning depth from monocular videos using direct methods. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2022-2030, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.72, + 0.471, + 0.789 + ], + "angle": 0, + "content": "[63] Can Wang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Text-and-image driven manipulation of neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3835–3844, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.471, + 0.844 + ], + "angle": 0, + "content": "[64] Chaoyang Wang, Simon Lucey, Federico Peruzzi, and Oliver Wang. Web stereo video supervision for depth prediction from dynamic scenes. In 2019 International Conference on 3D Vision (3DV), pages 348-357. IEEE, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[65] Lizhen Wang, Xiaochen Zhao, Tao Yu, Songtao Wang, and Yebin Liu. Normalgan: Learning detailed 3d human from a single rgb-d image. In European Conference on Computer Vision, pages 430-446. Springer, 2020. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.16 + ], + "angle": 0, + "content": "[66] Nanyang Wang, Yinda Zhang, Zhuwen Li, Yanwei Fu, Wei Liu, and Yu-Gang Jiang. Pixel2mesh: Generating 3d mesh models from single rgb images. In Proceedings of the European conference on computer vision (ECCV), pages 52-67, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[67] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P Srinivasan, Howard Zhou, Jonathan T Barron, Ricardo Martin-Brualla, Noah Snavely, and Thomas Funkhouser. Ibrnet: Learning multi-view image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4690-4699, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.247, + 0.892, + 0.303 + ], + "angle": 0, + "content": "[68] Jiajun Wu, Yifan Wang, Tianfan Xue, Xingyuan Sun, Bill Freeman, and Josh Tenenbaum. Marrnet: 3d shape reconstruction via 2.5 d sketches. Advances in neural information processing systems, 30, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.304, + 0.892, + 0.373 + ], + "angle": 0, + "content": "[69] Ke Xian, Chunhua Shen, Zhiguo Cao, Hao Lu, Yang Xiao, Ruibo Li, and Zhenbo Luo. Monocular relative depth perception with web stereo data supervision. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 311-320, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.374, + 0.892, + 0.429 + ], + "angle": 0, + "content": "[70] Dejia Xu, Yifan Jiang, Peihao Wang, Zhiwen Fan, Humphrey Shi, and Zhangyang Wang. Sinnerf: Training neural radiance fields on complex scenes from a single image. arXiv preprint arXiv:2204.00928, 2022. 1, 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.431, + 0.892, + 0.487 + ], + "angle": 0, + "content": "[71] Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. Mvsnet: Depth inference for unstructured multi-view stereo. In Proceedings of the European conference on computer vision (ECCV), pages 767-783, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.488, + 0.892, + 0.543 + ], + "angle": 0, + "content": "[72] Wei Yin, Xinlong Wang, Chunhua Shen, Yifan Liu, Zhi Tian, Songcen Xu, Changming Sun, and Dou Renyin. Diversedepth: Affine-invariant depth prediction using diverse data. arXiv preprint arXiv:2002.00569, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.544, + 0.892, + 0.599 + ], + "angle": 0, + "content": "[73] Wei Yin, Jianming Zhang, Oliver Wang, Simon Niklaus, Long Mai, Simon Chen, and Chunhua Shen. Learning to recover 3d scene shape from a single image. In Proc. IEEE Conf. Comp. Vis. Patt. Recogn. (CVPR), 2021. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.601, + 0.892, + 0.669 + ], + "angle": 0, + "content": "[74] Wei Yin, Jianming Zhang, Oliver Wang, Simon Niklaus, Long Mai, Simon Chen, and Chunhua Shen. Learning to recover 3d scene shape from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 204-213, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.671, + 0.892, + 0.739 + ], + "angle": 0, + "content": "[75] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4578-4587, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.741, + 0.892, + 0.797 + ], + "angle": 0, + "content": "[76] Jiahui Yu, Zhe Lin, Jimei Yang, Xiaohui Shen, Xin Lu, and Thomas S Huang. Free-form image inpainting with gated convolution. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4471-4480, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.797 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4489" + } + ] +] \ No newline at end of file diff --git a/2023/NeuralLift-360_ Lifting an In-the-Wild 2D Photo to a 3D Object With 360deg Views/90dd78db-4103-415c-8ddc-ba0f83221385_origin.pdf b/2023/NeuralLift-360_ Lifting an In-the-Wild 2D Photo to a 3D Object With 360deg Views/90dd78db-4103-415c-8ddc-ba0f83221385_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c1cfebb5bcd851702cbbd9bb9001b5a90ec29e95 --- /dev/null +++ b/2023/NeuralLift-360_ Lifting an In-the-Wild 2D Photo to a 3D Object With 360deg Views/90dd78db-4103-415c-8ddc-ba0f83221385_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f61b645ed6bb674da1fe357f58d0d533f9965fe80aef9d2089f2d385f2d370e5 +size 494925 diff --git a/2023/NeuralLift-360_ Lifting an In-the-Wild 2D Photo to a 3D Object With 360deg Views/full.md b/2023/NeuralLift-360_ Lifting an In-the-Wild 2D Photo to a 3D Object With 360deg Views/full.md new file mode 100644 index 0000000000000000000000000000000000000000..661943753db4db5d467fc4219fe3f6c7661d4081 --- /dev/null +++ b/2023/NeuralLift-360_ Lifting an In-the-Wild 2D Photo to a 3D Object With 360deg Views/full.md @@ -0,0 +1,339 @@ +# NeuralLift-360: Lifting An In-the-wild 2D Photo to A 3D Object with $360^{\circ}$ Views + +Dejia Xu, Yifan Jiang, Peihao Wang, Zhiwen Fan, Yi Wang, Zhangyang Wang +VITA Group, University of Texas at Austin + +{dejia,atlaswang}@utexas.edu + +# Abstract + +Virtual reality and augmented reality (XR) bring increasing demand for 3D content generation. However, creating high-quality 3D content requires tedious work from a human expert. In this work, we study the challenging task of lifting a single image to a 3D object and, for the first time, demonstrate the ability to generate a plausible 3D object with $360^{\circ}$ views that corresponds well with the given reference image. By conditioning on the reference image, our model can fulfill the everlasting curiosity for synthesizing novel views of objects from images. Our technique sheds light on a promising direction of easing the workflows for 3D artists and XR designers. We propose a novel framework, dubbed NeuralLift-360, that utilizes a depth-aware neural radiance representation (NeRF) and learns to craft the scene guided by denoising diffusion models. By introducing a ranking loss, our NeuralLift-360 can be guided with rough depth estimation in the wild. We also adopt a CLIP-guided sampling strategy for the diffusion prior to provide coherent guidance. Extensive experiments demonstrate that our NeuralLift-360 significantly outperforms existing state-of-the-art baselines. Project page: https://vita-group.github.io/NeuralLift-360/ + +# 1. Introduction + +Creating 3D content has been a long-standing problem in computer vision. This problem enables various applications in game studios, home decoration, virtual reality, and augmented reality. Over the past few decades, the manual task has dominated real scenarios, which requires tedious professional expert modeling. Modern artists rely on special software tools (e.g., Blender, Maya3D, 3DS Max, etc.) and time-demanding manual adjustments to realize imaginations and transform them into virtual objects. Meanwhile, automatic 3D content creation pipelines serve as effective tools to facilitate human efforts. These pipelines typically capture hundreds of images and leverage multi-view stereo [71] to quickly model fine-grained virtual landscapes. + +More recently, researchers have started aiming at a more + +ambitious goal, to create a 3D object from a single image [9, 10, 12, 21, 27, 51, 66, 68]. This enables broad applications since it greatly reduces the prerequisite to a minimal request. Existing approaches can be mainly categorized into two major directions. One line of work utilizes learning priors from large-scale datasets with multi-view images [10, 21, 51, 66, 68]. These approaches usually learn a conditional neural network to predict 3D information based on input images. However, due to their poor generalization ability, drastic performance degradation is observed when the testing image is out-of-distribution. + +Another direction constructs the pipeline on top of the depth estimation techniques [65]. Under the guidance of monocular depth estimation networks [43, 44], the 2D image is firstly back-projected to a 3D data format (e.g., point cloud or multi-plane image) and then re-projected into novel views. After that, advanced image inpainting techniques are then adopted to fill missing holes [52] produced during the projection. However, most of them can be highly affected by the quality of the estimated depth maps. Though LeRes [73] attempted to rectify the predicted depth by refining the projected 3D point cloud, their results do not generalize well to an arbitrary image in the wild. Overall, the aforementioned approaches are either adopted in limited scenarios (e.g. face or toy examples [4, 75]), or only produce limited viewing directions [8, 70] when being applied to scenes in the wild. Different from all these approaches, we focus on a more challenging task and for the first time, show promising results by lifting a single in-the-wild image into a 3D object with $360^{\circ}$ novel views. + +Attracted by the dramatic progress of neural volumetric rendering on 3D reconstruction tasks, we consider building our framework based on Neural Radiance Fields (NeRFs) [35]. The original NeRF takes hundreds of training views and their camera poses as inputs to learn an implicit representation. Subsequent models dedicate tremendous efforts [8, 19, 70, 75] to apply NeRF to sparse training views. The most similar work [70] to our method proposes to optimize a NeRF using only a single image and its corresponding depth map. However, it renders limited views from a small range of angles, and the prerequisite of a high-quality + +depth map largely constrains its practical usage. + +To address the above issues, we propose a novel framework, coined as NeuralLift-360, which aims to ease the creation of 3D assets by building a bridge to convert diverse in-the-wild 2D photos to sophisticated 3D contents in $360^{\circ}$ views and enable its automation. The major challenge in our work is that the content on the back side is hidden and hard to hallucinate. To tackle these hurdles, we consider the diffusion priors together with the monocular depth estimator as the cues for hallucination. Modern diffusion models [41,48,50] are trained on a massive dataset (e.g., 5B text-to-image pairs [24]). During inference time, they can generate impressive photorealistic photos based on simple text inputs. By adopting these learning priors with CLIP [40] guidance, NeuralLift-360 can generate plausible 3D consistent instances that correspond well to the given reference image while only requiring minimal additional input, the correct user prompts. Moreover, rather than simply taking the depth map from the pre-trained depth estimator as geometry supervision, we propose to use the relative ranking information from the rough depth during the training process. This simple strategy is observed to robustly mitigate geometry errors for depth estimations in the wild. + +Our contributions can be summarized as follows, + +- Given a single image in the wild, we demonstrate promising results of them being lifted to 3D. We use NeRF as an effective scene representation and integrate prior knowledge from the diffusion model. +- We propose a CLIP-guided sampling strategy that effectively marries the prior knowledge from the diffusion model with the reference image. +- When the reference image is hard to describe exactly, we finetune the diffusion model on the single image while maintaining its ability to generate diverse contents to guide the NeRF training. +- We introduce scale-invariant depth supervision that uses the ranking information. This design alleviates the need for accurate multi-view consistent depth estimation and broadens the application of our algorithms. + +# 2. Related Works + +# 2.1.3D from Single Image + +Inferring 3D worlds from a single image has been a popular direction over the past years. Previous works try to solve this problem by firstly adopting a monocular depth estimator (e.g., MiDaS [44]) to predict the 3D geometry and then using multi-plane images (MPI) [20, 52] or point cloud [36] to render artistic effects. Meanwhile, the missing + +holes from novel views are inpainted by a pre-trained modern deep network [76]. While these approaches are powerful and efficient, the depth estimator could be unstable, and the inpainting techniques may produce artifacts that make the rendered images look fake. While our method also applies a monocular depth estimator as the geometry cues, we propose a depth ranking loss that can mitigate the instability of depth estimation in the wild. Later approaches [28, 75] train their model on a large-scale 3D assets dataset. During inference time, they directly predict a 3D object using a single image. However, their performance faces drastic degradation when being applied to an in-the-wild image, due to the domain gap. The most similar work to ours is SinNeRF [70], which trains a neural radiance field only on a single RGB image and its corresponding depth map. However, their approach requires a high-quality depth map for geometry guidance, and it can only render views from a small range of angles. Thanks to the proposed depth ranking loss, our method can reconstruct impressive geometry information only using guidance from a pre-trained monocular depth estimator. + +# 2.2. Neural Radiance Fields + +Neural Radiance Fields (NeRFs) [34] have shown encouraging performance for view synthesis. After its origin, tremendous efforts have been devoted to improving it [2,3,5,6,11,14,16,18,31,37,45,57,59,61,63,67,75]. Unlike most NeRF-like models, which at least require several different views to reconstruct scenes, we focus on reconstructing 3D content with single views. Analogously, recent works [18, 39] try to generate 3D content from text inputs directly. Although they also rely on the text prompt as guidance, their rendered objects are only conditioned on text sentences without image constraints, which can not reach our goal. Different from them, we aim at lifting a single image (either real or synthetic) to photorealistic 3D objects. + +# 3. Preliminaries + +Neural Radiance Field NeRFs [34] render images by sampling 5D coordinates (location $(x,y,z)$ and viewing direction $(\theta ,\phi)$ ) along camera rays. After mapping them to color $(r,g,b)$ and volume density $\sigma$ , additional volumetric rendering techniques are adopted to fuse the information along the ray. Given $\mathbf{o}$ as the camera origin and $\pmb{d}$ as the ray direction, 3D point locations along the ray can be expressed as $\pmb {r}(t) = \pmb {o} + td$ The predicted color accumulated is defined as follows: + +$$ +\boldsymbol {C} (\boldsymbol {r}) = \int_ {t _ {n}} ^ {t _ {f}} T (t) \sigma (\boldsymbol {r} (t)) \boldsymbol {c} (\boldsymbol {r} (t), \boldsymbol {d}) d t, \tag {1} +$$ + +where $T(t) = \exp \left(-\int_{t_n}^t\sigma (\boldsymbol {r}(s))ds\right)$ , $\sigma (\cdot)$ and $c(\cdot ,\cdot)$ are densities and color predictions. To numerically evaluate + +![](images/725e14362f498810b5d86d0cb388287e55dd33039ddbcee24e2307a522486150.jpg) +Figure 1. Overview of the proposed pipeline. NeuralLift-360 learns to recover a 3D object from a single reference image $y$ . We derive a prior distillation loss $\mathcal{L}_{\mathrm{diff}}$ for CLIP-guided diffusion prior and utilize rough depth estimation with the help of a ranking loss $\mathcal{L}_{\mathrm{ranking}}$ . + +Eq. 1, the continuous integral is estimated using quadrature [34]. Our model is built upon Instant-NGP [37], which reduces computation cost by inferring color and density from a multiresolution hash table implemented by CUDA. + +Denoising Diffusion Model Denoising diffusion models generate images by gradually denoising from a gaussian noise $p(\mathbf{x}_T) = \mathcal{N}(\mathbf{0},\mathbf{I})$ and transforming into a certain data distribution. The forward diffusion process $q(\pmb {x}_t|\pmb {x}_{t - 1})$ adds Gaussian noise to the image $\pmb {x}_t$ . The marginal distribution can be written as: $q(\pmb {x}_t\mid \pmb {x_0}) = \mathcal{N}\left(\alpha_t\pmb {x_0},\sigma_t^2\pmb {I}\right)$ where $\alpha_{t}$ and $\sigma_{t}$ are designed to converge to $\mathcal{N}(\mathbf{0},\mathbf{I})$ when $t$ is at the end of the forward process [23, 56]. The reverse diffusion process $p(\pmb{x}_{t - 1}|\pmb {x}_t)$ learns to denoise. Given infinitesimal timesteps, the reverse diffusion process can be approximated with Gaussian [56] related with an optimal MSE denoiser [53]. The diffusion models are designed as noise estimators $\epsilon_{\theta}(\pmb {x}_t,t)$ taking noisy images as input and estimating the noise. They are trained via optimizing the weighted evidence lower bound (ELBO) [17, 23]: + +$$ +\mathcal {L} _ {\mathrm {E L B O}} (\theta) = \mathbb {E} \left[ w (t) \| \boldsymbol {\epsilon} _ {\theta} (\alpha_ {t} \pmb {x} _ {0} + \sigma_ {t} \pmb {\epsilon}; t) - \boldsymbol {\epsilon} \| _ {2} ^ {2} \right], \quad (2) +$$ + +where $\epsilon \sim \mathcal{N}(\mathbf{0},\mathbf{I})$ , $w(t)$ is a weighting function. In practice, setting $w(t) = 1$ delivers good performance [17]. Sampling from a diffusion model can be either stochastic [17] or deterministic [54]. After sampling $\boldsymbol{x}_T \sim \mathcal{N}(\mathbf{0},\mathbf{I})$ , we can gradually reduce the noise level and reach a clean image with high quality at the end of the iterative process. + +# 4. Method + +Overview. In this section, we propose our method, dubbed NeuralLift-360, which reconstructs a 3D object + +from a single image. An overview of NeuralLift-360 is illustrated in Fig. 1. NeuralLift-360 combines the best of the two worlds: NeRF and diffusion model, where the former is leveraged to universally model a 3D scene and the latter is utilized to hallucinate the unseen views. To inject knowledge of a diffusion model into a radiance field, we derive a training framework in Sec. 4.1, whose supervision consists of two parts: (1) a prior distillation loss from a generative model (Sec. 4.2), (2) a straight loss on the given image (Sec. 4.4). Moreover, we introduce a domain adaption technique to adapt the diffusion prior to in-the-wild images (Sec. 4.3). + +# 4.1. Probabilistic-Driven 3D Lifting + +We begin by formulating our setting along the way, introducing our notations. Given an image $\pmb{y} \in \mathbb{R}^N$ and its text description $\pmb{z} \in \mathbb{R}^D$ , NeuralLift-360 intends to reconstruct a 3D scene $\pmb{V} \in \mathbb{R}^M$ , where $\pmb{V}$ can be either a radiance volume or the network parameters of a NeRF. We regard $\pmb{y}$ , $\pmb{z}$ and $\pmb{V}$ as random variables. To reconstruct $\pmb{V}$ from $\pmb{y}$ and $\pmb{z}$ , we maximize a log-posterior and apply Bayesian rule: + +$$ +\log p (\boldsymbol {V} | \boldsymbol {y}, \boldsymbol {z}) = \log p (\boldsymbol {y} | \boldsymbol {V}, \boldsymbol {z}) + \log p (\boldsymbol {V} | \boldsymbol {z}) + c o n s t. \tag {3} +$$ + +One can see the prior term $p(\mathbf{V}|z)$ provides complementary information to lift an image into 3D space. However, specifying this prior can be intractable, and directly optimizing objective Eq. 3 is also sampling inefficient. Instead, we introduce camera pose $\Phi \in \mathbb{S}\mathbb{O}(3)\times \mathbb{R}^3$ as a latent variable, then the likelihood term can be rewritten as: + +$$ +\begin{array}{l} p (\boldsymbol {y} | \boldsymbol {V}, \boldsymbol {z}) = \int p (\boldsymbol {y} | \boldsymbol {V}, \boldsymbol {\Phi}, \boldsymbol {z}) p (\boldsymbol {\Phi} | \boldsymbol {V}, \boldsymbol {z}) d \boldsymbol {\Phi} (4) \\ = \int p (\boldsymbol {y} | \boldsymbol {V}, \boldsymbol {\Phi}, \boldsymbol {z}) \frac {p (\boldsymbol {V} | \boldsymbol {\Phi} , \boldsymbol {z}) p (\boldsymbol {\Phi})}{p (\boldsymbol {V} | \boldsymbol {z})} d \boldsymbol {\Phi}, (5) \\ \end{array} +$$ + +where we assume $\Phi$ is independent of $z$ . Merging Eq. 3 and Eq. 5, we can get rid of the implicit prior term and obtain a sampling efficient ELBO of the original objective as below: + +$$ +\begin{array}{l} \log \mathbb {E} _ {\boldsymbol {\Phi} \sim p (\boldsymbol {\Phi})} p (\boldsymbol {y} | \boldsymbol {V}, \boldsymbol {\Phi}, \boldsymbol {z}) p (\boldsymbol {V} | \boldsymbol {\Phi}, \boldsymbol {z}) (6) \\ \geq \mathbb {E} _ {\boldsymbol {\Phi} \sim p (\boldsymbol {\Phi})} \left[ \log p (\boldsymbol {y} | \boldsymbol {V}, \boldsymbol {\Phi}, \boldsymbol {z}) + \log p (\boldsymbol {V} | \boldsymbol {\Phi}, \boldsymbol {z}) \right]. (7) \\ \end{array} +$$ + +Furthermore, we define $p(\boldsymbol{y}|\boldsymbol{V}, \boldsymbol{\Phi}, \boldsymbol{z}) = p(\boldsymbol{y}|h(\boldsymbol{V}, \boldsymbol{\Phi}), \boldsymbol{z})$ and $p(\boldsymbol{V}|\boldsymbol{\Phi}, \boldsymbol{z}) = p(h(\boldsymbol{V}, \boldsymbol{\Phi})|\boldsymbol{z})$ , where $h(\boldsymbol{V}, \boldsymbol{\Phi})$ renders the 3D scene with respect to the camera pose $\boldsymbol{\Phi}$ . Then we derive our final training objective for NeuralLift-360: + +$$ +\mathcal {L} = - \mathbb {E} _ {\boldsymbol {\Phi}} \left[ \underbrace {\log p (\boldsymbol {y} \mid h (\boldsymbol {V} , \boldsymbol {\Phi}) , \boldsymbol {z})} _ {\text {r e f e r e n c e d l o s s}} + \underbrace {\log p (h (\boldsymbol {V} , \boldsymbol {\Phi}) \mid \boldsymbol {z})} _ {\text {n o n - r e f e r e n c e d l o s s}} \right]. \tag {8} +$$ + +At each training step, we sample a camera pose with respect to a distribution (more on Sec. 5.1), and then render the view associated with this camera parameter (via volume rendering Eq. 1). The loss computed on the rendered image consists of two parts: image-referenced guidance and reference-free regularization. The referenced loss will enforce $V$ to have a close appearance with given $y$ , while the non-referenced term hallucinates the image information with prior knowledge. In the next section, we introduce a diffusion model to implement our loss function Eq. 8. + +# 4.2. CLIP-guided Diffusion Priors + +Diffusion models [17, 54-56], which synthesize images by gradually denoising from a gaussian noise, have undeniably been successful in the realm of image generation, editing, and processing tasks. When trained on large-scale datasets for text-to-image tasks, these models can accurately represent the distribution of images based on textual descriptions [41, 42, 48]. Consequently, we are inspired to utilize a diffusion model's generative prior to complement the 3D information. To accomplish this, we re-parameterize our Eq. 8 with a diffusion model. Nevertheless, the derived loss function contains both discriminative and generative terms. Taking inspiration from classifier guidance techniques [26, 30, 56], we show that we can leverage the training loss of a diffusion model to surrogate the generative part and utilize CLIP space similarity to model the discriminative part. + +Following the derivation in [17] and [32], we lower bound Eq. 8 by: + +$$ +\begin{array}{l} \log p _ {\theta} (\boldsymbol {y} | \boldsymbol {x}, \boldsymbol {z}) + \log p _ {\theta} (\boldsymbol {x} | \boldsymbol {z}) \\ \geq - \mathbb {E} _ {t, \epsilon \sim \mathcal {N} (\mathbf {0}, \mathbf {I})} \left[ \mathcal {D} _ {\mathrm {K L}} \left(q \left(\boldsymbol {x} _ {t - 1} \mid \boldsymbol {x} _ {t}, \boldsymbol {x} _ {0}\right) \| p _ {\theta} \left(\boldsymbol {x} _ {t - 1} \mid \boldsymbol {x} _ {t}, \boldsymbol {y}, \boldsymbol {z}\right) \right] \right. \\ = - \mathbb {E} _ {t, \boldsymbol {\epsilon} \sim \mathcal {N} (\mathbf {0}, \boldsymbol {I})} \left[ w (t) \| \nabla \log p _ {\theta} (\boldsymbol {x} _ {t} | \boldsymbol {y}, \boldsymbol {z}) - \boldsymbol {\epsilon} \| _ {2} ^ {2} \right], \tag {9} \\ \end{array} +$$ + +where $\pmb{x}_t = \alpha_t\pmb{x}_0 + \sigma_t\pmb{\epsilon}$ and $w(t)$ is a constant dependent on noise scale $\alpha_{t}$ and $\sigma_t$ . The expectation of $t\sim \mathcal{U}\{1,\dots ,T\}$ + +is usually taken over a uniform distribution. We provide detailed derivation in supplementary. At first glance, we may directly plug in the pre-trained denoiser $\epsilon_{\theta}(\pmb{x}_t;\pmb{z})$ of a diffusion model to surrogate $\nabla p_{\theta}(\pmb{x}_t|\pmb{y},\pmb{z})$ similar to [39]. However, note that $p_{\theta}(\pmb{x}_t|\pmb{y},\pmb{z})$ is also conditioned on the reference image $\pmb{y}$ , which requires further efforts to inject the reference image information into the diffusion process. + +To this end, NeuralLift-360 leverages a Mixed Reference Guidance that guides the diffusion process with a mixture of references. Similar to classifier guidance techniques [26, 30, 32, 56], we rewrite $p_{\theta}(\boldsymbol{x}_t|\boldsymbol{y}, \boldsymbol{z}) = p_{\theta}(\boldsymbol{y}|\boldsymbol{x}_t, \boldsymbol{z})p_{\theta}(\boldsymbol{x}_t|\boldsymbol{z}) / p_{\theta}(\boldsymbol{y}|\boldsymbol{z})$ , and we adopt classifier-free guidance to model $p_{\theta}(\boldsymbol{x}_t|\boldsymbol{z})$ , then we can obtain a modified score function [55] as below: + +$$ +\begin{array}{l} \boldsymbol {\epsilon} _ {\theta} (\boldsymbol {x} _ {t}; \boldsymbol {y}, \boldsymbol {z}) = (1 + w) \boldsymbol {\epsilon} _ {\theta} (\boldsymbol {x} _ {t}; \boldsymbol {z}) - w \boldsymbol {\epsilon} _ {\theta} (\boldsymbol {x} _ {t}) \\ + \sqrt {1 - \alpha_ {t}} \nabla \log p _ {\theta} (\boldsymbol {y} | \boldsymbol {x} _ {t}, \boldsymbol {z}), \tag {10} \\ \end{array} +$$ + +where $w$ is a hyper-parameter that controls the strength of the classifier guidance. We further measure the discrimination with a distance metric on the feature space $p_{\theta}(\boldsymbol{y}|\boldsymbol{x}_t, \boldsymbol{z}) \propto \exp\left[-\phi\left(\frac{\boldsymbol{x}_t - \sigma_t \boldsymbol{\epsilon}_{\theta}(\boldsymbol{x}_t; \boldsymbol{z})}{\alpha_t}, \boldsymbol{y}\right)\right]$ . Specifically, we choose the inner product on the CLIP embedding space [40] as the distance metric, i.e., $\log p_{\theta}(\boldsymbol{y}|\boldsymbol{x}_t, \boldsymbol{z}) \propto -\left\langle F\left(\frac{\boldsymbol{x}_t - \sigma_t \boldsymbol{\epsilon}_{\theta}(\boldsymbol{x}_t)}{\alpha_t}\right), F(\boldsymbol{y})\right\rangle$ , which spurs on semantic similarity between denoised rendered image and reference image $\boldsymbol{y}$ , where $F$ is the CLIP image encoder. + +In each training iteration, we alternate two camera pose sampling strategies: (1) we fix a camera pose $\Phi_0$ (i.e., a Delta distribution) as the reference view, and when rendering on that view, we only penalize the RGB differences by modeling $p(\boldsymbol{y} | h(\boldsymbol{V}, \boldsymbol{\Phi}), \boldsymbol{z}) = \mathcal{N}(\boldsymbol{y} | h(\boldsymbol{V}, \boldsymbol{\Phi}), \sigma^2 \boldsymbol{I})$ , (2) we stochastically sample camera poses, and adopt the diffusion ELBO (Eq. 9) to estimate the evidence of rendered image together with prior $p_{\theta}(\boldsymbol{x} | \boldsymbol{z})$ . Combining all together, we can summarize our training loss as below: + +$$ +\begin{array}{l} \mathcal {L} = \frac {1}{\sigma^ {2}} \underbrace {\| h (\boldsymbol {V} , \boldsymbol {\Phi} _ {0}) - \boldsymbol {y} \| _ {2} ^ {2}} _ {\mathcal {L} _ {\text {p h o t o m e t r i c}}} \tag {11} \\ + \underbrace {\mathbb {E} _ {\boldsymbol {\Phi} , t , \boldsymbol {\epsilon} \sim \mathcal {N} (\boldsymbol {0} , \boldsymbol {I})} \left[ w (t) \| \boldsymbol {\epsilon} _ {\theta} (\alpha_ {t} h (\boldsymbol {V} , \boldsymbol {\Phi}) + \sigma_ {t} \boldsymbol {\epsilon} | \boldsymbol {y} , \boldsymbol {z}) - \boldsymbol {\epsilon} \| _ {2} ^ {2} \right]} _ {\mathcal {L} _ {\mathrm {d i f f}}}. \\ \end{array} +$$ + +Compared with DreamFusion [39], NeuralLift-360's training objective grounds the 3D scene with the given image which enables the generated scene to have consistent appearance with the specified object. + +# 4.3. Domain Adaption to In-the-wild Images + +A primary challenge, however, lies in assigning an appropriate text prompt for our reference image. Given that + +![](images/c160a1d560897fd34353942460214ed10cf85d921bccb454939bd5ead57c525a.jpg) + +![](images/42721362303136b9f0d4c3afa5614b7e9326a62cf2614d01f64872c6fca20353.jpg) +(a) Pre-trained + +![](images/dcc6c2ab0768b4ae14aaa71f1b11d37ea2c94bfd676956c62b305d0e270f13ab.jpg) + +![](images/0ca5b2e3e3d4841da341e3c15d2b921165e698d4ba14c4f6bad700ffc8bbadcd.jpg) + +![](images/428a4ee7e2027c92cbd8c6df576749ac85becc2b47b0a24a471f977851348bac.jpg) + +![](images/d45179ef0aa73b6c8bff223309ce4e4b56a3741cff17d4c47c23e06dd2f8196d.jpg) +Figure 2. After fine-tuning, our model can generate diverse images similar to the reference image (b) while not identical. + +![](images/faaf30362df5684fbcfaf8eb6b4c80acf115e1a7e0dda21815099486d27d41ef.jpg) +(c) Fine-tuned + +![](images/599cc7a413195a7fa7d519255148d3e27d323095ebf7fe517fd24034e7771f2f.jpg) + +![](images/95620d1a82f2c61d9a7e888d725bbe9f56a2ddd7f0fd50a571b7c947544255ad.jpg) + +![](images/91e36f2d1e4951c26124c9a0ae489047d3ad56f323414eb38a6b9e61c84f90c0.jpg) + +large diffusion models are predominantly trained for text-to-image tasks, they typically require a text embedding as conditional input. For images in the wild, it is generally hard to exactly describe an image since images contain more fine-grained information than texts. + +Adapting to in-the-wild images necessitates finetuning the Stable Diffusion model on the reference image Textual inversion [15] and Dreambooth [49] enable finetuning the Stable Diffusion model on a limited number of images for a subject-driven generation. However, in our situation, we only have access to a single image, which significantly hampers the effectiveness of these methods. Imagic [22] and UniTune [60] support finetuning on a single image, but their primary objective is to perform real image editing, which can be quite successful even if the diffusion model overfits the given image. Nevertheless, in our context, we need the diffusion model to generate images resembling the provided reference but not identical to it. If the supervision for each view is the same, NeRF will otherwise produce an isotropic surface. + +We draw inspiration from Pivotal Tuning [47] and Imagic [22] when fine-tuning Stable Diffusion [48] on our own. To start with, we optimize the CLIP text embedding $z$ so that it fits our reference image $y$ better. During fine-tuning, we use Eq. 2 to supervise the noise estimate. As we expect that our denoised result will resemble the reference image, we also apply a CLIP image loss to the estimated denoised image. $\hat{x}_0$ , + +$$ +\begin{array}{l} \mathcal {L} _ {\text {f i n e t u n e}} = \mathbb {E} _ {t, \epsilon} \left[ \| \boldsymbol {\epsilon} - \boldsymbol {\epsilon} _ {\theta} (\mathbf {x} _ {t}; \boldsymbol {z}) \| _ {2} ^ {2} \right] \\ - \mathbb {E} _ {t, \epsilon} \left\langle F \left(\frac {\boldsymbol {x} _ {t} - \sigma_ {t} \boldsymbol {\epsilon} _ {\theta} (\boldsymbol {x} _ {t} ; \boldsymbol {z})}{\alpha_ {t}}\right), F (\boldsymbol {y}) \right\rangle , \tag {12} \\ \end{array} +$$ + +where $F(\cdot)$ is the pre-trained CLIP image encoder. + +Then, we fix the optimized text embedding $z^{*}$ and finetune the diffusion model $\epsilon_{\theta}(\cdot)$ using the same loss function as above. To preserve the generalization ability of the diffusion model, we introduce several augmentations during the finetuning. Specifically, in each iteration, we perform random image-level augmentations such as flip, transpose, shift, scale, and rotate. As shown in Fig. 2, we are able + +to overcome the overfitting issue and preserve the diffusion model's ability to generate diverse similar but not identical contents. Similar to Imagic [22] and UniTune [47], we observe that after finetuning, the optimized text embedding will have an overfitting issue and it is critical to perform interpolation between the original text embedding and the optimized text embedding to preserve the diffusion model's ability to generate diverse contents. We obtain $z' = \eta z^* + (1 - \eta)z$ , where $\eta$ is a weighting hyper-parameter. + +During sampling, we aim to amplify the conditional likelihood for our optimized embedding $z^{*}$ , while decreasing both the unconditional likelihood and the conditional likelihood for the original embedding $z$ . Hence for 50% of all iterations, we use the original Eq. 10, while for the others, we replace the unconditional sampling $\epsilon_{\theta}(\boldsymbol{x}_t)$ in Eq. 10 with conditional sampling using the original text embedding $z$ : + +$$ +\begin{array}{l} \widetilde {\boldsymbol {\epsilon}} _ {\theta} \left(\boldsymbol {x} _ {t}; \boldsymbol {y}, \boldsymbol {z} ^ {\prime}, \boldsymbol {z}\right) = (1 + w) \boldsymbol {\epsilon} _ {\theta} \left(\boldsymbol {x} _ {t}; \boldsymbol {z} ^ {\prime}\right) \tag {13} \\ - w \epsilon_ {\theta} (\boldsymbol {x} _ {t}; \boldsymbol {z}) + \sqrt {1 - \alpha_ {t}} \nabla \log p _ {\theta} (\boldsymbol {y} | \boldsymbol {x} _ {t}, \boldsymbol {z} ^ {\prime}), \\ \end{array} +$$ + +# 4.4. Supervision from Rough Depth + +While the aforementioned pipeline laid the foundation for the success of our method, we observe that obtaining a high-quality 3D object without geometry regularization is inherently challenging, if relying solely on learning priors. As shown in SinNeRF [70] and DS-NeRF [8], accurate depth information can largely help improve the rendering quality. However, their impressive outcomes are contingent upon the 3D consistency of the accessible high-quality depth, a requirement that is not easily fulfilled in real-life scenarios. + +In spite of the significant progress in monocular depth estimation for in-the-wild scenarios, current state-of-the-art depth estimators are insufficient for recovering 3D shapes. Due to unknown camera baselines and stereoscopic post-processing in data, these methods [7,25,44,64,69,72] utilize loss functions invariant to shift and scale or adopt ranking loss that utilizes relative information. Therefore, their results cannot be used to reconstruct plausible 3D scene shapes directly due to unknown depth shifts [74]. Though there are also attempts [74] that improve off-the-shelf depth estimators, they generally can not generalize well to unseen objects in the wild. + +Given that the acquired depth is imperfect, the absolute depth scale lacks the reliability necessary for recovering a high-quality 3D object. In contrast to previous depth-supervised NeRFs [8, 70], which rely solely on precise values from the depth map, our approach represents the first attempt to maximize the utilization of depth information by adopting the relative ranking information. Drawing inspiration from [7, 38], we adopt a pairwise depth ranking loss to supervise the rendered depth from our NeRF backbone. In this way, we enforce the depth we rendered to be consistent + +with the depth from the reference view in terms of internal ranking instead of actual value. The loss is formulated as follows, + +$$ +\mathcal {L} _ {\text {r a n k i n g}} = \left\{ \begin{array}{l l} \max \left(z _ {j _ {k}} - z _ {i _ {k}}, 0\right), & r _ {k} \in \{> \} \\ \max \left(z _ {i _ {k}} - z _ {j _ {k}}, 0\right), & r _ {k} \in \{< \} \end{array} \right. \tag {14} +$$ + +where $r_k$ refers to the pairwise ranking relationship between pixel $z_{i_k}$ and $z_{j_k}$ in the monocular depth estimation. This loss function is used alongside the photo-metric loss and diffusion loss mentioned in Eq. 11 and serves as additional supervision on geometry. + +# 5. Experiments + +In this section, we will first describe a unique training recipe that we propose to train our framework, then the implementation details. + +# 5.1. Training Recipe + +Diverse Camera Pose and Intrinsic Sampling During training, we observe that a diverse camera pose and intrinsicics are essential for the model to learn a coherent 3D object. If the camera pose is not sampled from a diverse distribution, the 3D object will collapse towards the major viewing directions and thus generate unpleasant results. Specifically, we alter between two pose sampling strategies; (1) we have a fixed camera pose $\Phi_0$ (i.e., a Delta distribution) for the reference view, (2) we uniformly sample the camera rotation on the surface of a unit sphere and then uniformly sample the radius in [0.4, 1.0] to locate the camera. We start with the Delta distribution, and as training progresses, we gradually shift to the other sophisticated camera pose distribution. The camera intrinsics provide additional augmentation on the viewing direction. By uniformly sampling the field of view in [50, 70], we zoom in and out to get different resolution projections for the scene. We also add Gaussian noise to the camera pose sampling phase. Such perturbations help us render more diverse images during training. + +Foreground Aware Diffusion Prior ERNIE-ViLG [13] utilizes a weighting mechanism during the training of diffusion models to let the model generate. Using an additional object detection model, their weighted loss forces the model to focus on salient regions. Similarly, we hope to strengthen the guidance on the foreground object region. We propose to weight the score matching distance with the predicted density. Specifically, we penalize the regions with higher density so that our NeRF gains more guidance on the foreground regions. Hence, for each rendered image, we first obtain the bounding box for the non-zero density regions. Then, we extend the $w(t)$ in $\mathcal{L}_{\mathrm{diff}}$ (Eq. 11) into a spatial function: + +$$ +W _ {i, j} = \left\{ \begin{array}{l l} 2, & \text {i f} (i, j) \text {i s i n s i d e t h e b o u n d i n g b o x} \\ 1, & \text {o t h e r w i s e .} \end{array} \right. \tag {15} +$$ + +Geometry Regularization We also include a series of geometry regularizations to avoid undesired artifacts in our 3D object. To avoid foggy and semi-transparent artifacts at the back of the scene, we utilize the orient loss from Ref-NeRF [61] to prevent the surface normals from facing backward from the camera. We also incorporate depth smoothness loss [62], sparsity loss, and distortion loss [3, 58] to avoid the NeRF filling empty space with floaters. Full details for these regularizations are provided in the supplementary material. + +Timestep Annealing At the start of training, the quality of the rendered images is poor, and we need the diffusion prior to provide more thorough guidance. As a result, we need to add more noise to perturb the rendered image. As the training progresses, the quality of rendered images gradually improves, and we choose to reduce the timestep range accordingly. By annealing the timestep where we add noise, we are lowering the noise level of the perturbation, and the noise estimate becomes more accurate. + +# 5.2. Evaluation Workflow + +We conduct extensive experiments using both text-to-image synthetic images and real-world images. Synthetic images in our experiments are generated using PNDM Sampler [29] with 50 sampling steps. Real images are either from CO3Dv2 dataset [46] or randomly crawled from the Google search engine. For each image, we use a pre-trained background removal framework [1] to segment the foreground objects. Then we obtain their depth using LeRes [73] with the "Boost Your Own depth" [33] strategy. [33] inferences the monocular depth estimation network on multiple patches and then fuses the results for better details. We compare with previous sparse view NeRFs, including Depth-supervised NeRF [8] and SinNeRF [70]. We also compare with DietNeRF [19] that uses CLIP as guidance. + +# 5.3. Comparisons + +We present quantitative results in eight scenes. We report the average CLIP distance between rendered image and the reference image as a metric for how well the object is represented. During testing, we use a CLIP model different from training time and evaluate on held out viewing directions with a larger distance-to-origin than training. We render 100 images on each scene for each method. As shown in Tab. 1, our method achieves the best CLIP distance among existing approaches. + +We also provide visual comparisons in Fig. 3. Since DSNeRF and SinNeRF rely on the exact value of the available depth information, their performance using monocular depth estimation is far from satisfactory. It can also be seen that the DietNeRF suffers from unconstrained geometry. Our method, on the other hand, is able to synthesize visually pleasing novel views with 3D consistency. + +![](images/2fa5abf08f79858788196ec86283d10e346a75da03b8b32250595477e658dccc.jpg) +Figure 3. Visual comparisons for novel view synthesis results. We provide video comparisons in supplementary. + +Table 1. Quantitative evaluation by comparing the proposed method with previous state-of-the-art approaches. + +
CLIP Distance ↓AppleBenchBaseballHydrantRemoteSuitcaseStatueSandwichAverage
DSNeRF [8]0.53090.54470.55060.62260.64410.55330.59920.61100.5821
DietNeRF [19]0.42450.55790.45450.51900.57180.48900.59470.51790.5162
SinNeRF [70]0.51920.57690.62300.63430.64960.57050.56050.60150.5918
NeuralLift-360 (Ours)0.39140.47850.39740.51460.49160.41860.44110.46550.4498
+ +![](images/5b63a43d9c64fc35ae3add3e12b667955b411b2ac73dc82001d688914b37a7da.jpg) +Figure 4. Ablation study on the baseline variants. Given a single reference image, we run different variants to render its $360^{\circ}$ views. + +![](images/eaf753534202c57de45f178e63bbb73cb703165fb5d44bd6fe894366ac993bf9.jpg) +Figure 5. Ablation study on CLIP guidance for the diffusion prior. + +# 5.4. Ablation Study + +Variants of supervision. In this section, we perform detailed ablation studies on different components of our framework. "w/o $\mathcal{L}_{\mathrm{ranking}}$ " replaces the ranking loss with $L_{1}$ loss on depth. "w/o $\mathcal{L}_{\mathrm{diff}}$ " replaces the CLIP guided diffusion prior with CLIP. "w/o fine-tuning" refers to using the pre-trained diffusion model. As shown in Fig. 4, our full model delivers the best performance, while other variants have different issues. "w/o $\mathcal{L}_{\mathrm{ranking}}$ " relies on the exact value from depth information and suffers from unreliable depth. "w/o $\mathcal{L}_{\mathrm{diff}}$ " uses CLIP as a prior, and the optimized result is + +blurry and collapsed. "w/o fine-tuning" uses a pre-trained diffusion model, since the text prompt is not very accurate, the resulting object has lettuce which doesn't exist in the reference image. + +Effectiveness of CLIP guided diffusion prior We validate the effectiveness of CLIP guidance in our diffusion prior design in Fig. 5. It can be seen that without the CLIP guidance, although the model produces a 3D object, it doesn't look like the reference image and has a strange shape. In comparison, our full model generates a better shape and looks more similar to the reference image. + +# 6. Conclusions and Future Work + +We present NeuralLift-360, a novel framework to lift an in-the-wild 2D photo into a 3D object with 360-degree views. NeuralLift-360 learns probabilistic-driven 3D lifting with CLIP-guided diffusion priors and mitigates the depth errors by a scale-invariant depth ranking loss. Comprehensive experiments are conducted on real and synthetic images, where NeuralLift-360 outperforms the current state-of-the-art methods. While our current implementation is based on Stable Diffusion [48] version 1.4, the framework should also work well with other diffusion models such as Imagen [50] or DALL-E2 [41], once they are open-sourced. + +Limitations. Despite the encouraging visual results rendered by NeuralLift-360, the target resolution $(128 \times 128)$ is still behind large generative models [41, 50]. Moreover, challenging cases (e.g., multiple objects with occlusion) are not included in our assumptions, which need to be further explored. Going forward, we will investigate how to expand our method to more general scenarios. + +# References + +[1] Image background removal. https://github.com/OPHoperHPO/image-background-remove-tool. 6 +[2] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 2 +[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5470–5479, 2022. 2, 6 +[4] Angel X. Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. ShapeNet: An Information-Rich 3D Model Repository. Technical Report arXiv:1512.03012 [cs.GR], Stanford University — Princeton University — Toyota Technological Institute at Chicago, 2015. 1 +[5] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14124-14133, 2021. 2 +[6] Tianlong Chen, Peihao Wang, Zhiwen Fan, and Zhangyang Wang. Aug-nerf: Training stronger neural radiance fields with triple-level physically-grounded augmentations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15191-15202, 2022. 2 +[7] Weifeng Chen, Zhao Fu, Dawei Yang, and Jia Deng. Single-image depth perception in the wild. Advances in neural information processing systems, 29, 2016. 5 +[8] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised nerf: Fewer views and faster training for free. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12882-12891, 2022. 1, 5, 6, 7, 8 +[9] Jean-Denis Durou, Maurizio Falcone, and Manuela Sagona. Numerical methods for shape-from-shading: A new survey with benchmarks. Computer Vision and Image Understanding, 109(1):22-43, 2008. 1 +[10] Haoqiang Fan, Hao Su, and Leonidas J Guibas. A point set generation network for 3d object reconstruction from a single image. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 605-613, 2017. 1 +[11] Zhiwen Fan, Yifan Jiang, Peihao Wang, Xinyu Gong, Dejia Xu, and Zhangyang Wang. Unified implicit neural stylization. arXiv preprint arXiv:2204.01943, 2022. 2 +[12] Paolo Favaro and Stefano Soatto. A geometric approach to shape from defocus. IEEE Transactions on Pattern Analysis and Machine Intelligence, 27(3):406-417, 2005. 1 +[13] Zhida Feng, Zhenyu Zhang, Xintong Yu, Yewei Fang, Lanxin Li, Xuyi Chen, Yuxiang Lu, Jiaxiang Liu, Weichong + +Yin, Shikun Feng, et al. Ernie-vilg 2.0: Improving text-to-image diffusion model with knowledge-enhanced mixture-of-denoising-experts. arXiv preprint arXiv:2210.15257, 2022.6 +[14] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5501–5510, 2022. 2 +[15] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 5 +[16] Yuan-Chen Guo, Di Kang, Linchao Bao, Yu He, and Song-Hai Zhang. Nerfren: Neural radiance fields with reflections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18409-18418, 2022. 2 +[17] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 3, 4 +[18] Ajay Jain, Ben Mildenhall, Jonathan T Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 867-876, 2022. 2 +[19] Ajay Jain, Matthew Tancik, and Pieter Abbeel. Putting nerf on a diet: Semantically consistent few-shot view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5885-5894, 2021. 1, 6, 7, 8 +[20] Varun Jampani, Huiwen Chang, Kyle Sargent, Abhishek Kar, Richard Tucker, Michael Krainin, Dominik Kaeser, William T Freeman, David Salesin, Brian Curless, et al. Slide: Single image 3d photography with soft layering and depth-aware inpainting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12518-12527, 2021. 2 +[21] Adrian Johnston, Ravi Garg, Gustavo Carneiro, Ian Reid, and Anton van den Hengel. Scaling cnns for high resolution volumetric reconstruction from a single image. In Proceedings of the IEEE International Conference on Computer Vision Workshops, pages 939-948, 2017. 1 +[22] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagc: Text-based real image editing with diffusion models. arXiv preprint arXiv:2210.09276, 2022. 5 +[23] Diederik Kingma, Tim Salimans, Ben Poole, and Jonathan Ho. Variational diffusion models. Advances in neural information processing systems, 34:21696-21707, 2021. 3 +[24] Hyung-Kwon Ko, Gwanmo Park, Hyeon Jeon, Jaemin Jo, Juho Kim, and Jinwook Seo. Large-scale text-to-image generation models for visual artists' creative works. arXiv preprint arXiv:2210.08477, 2022. 2 +[25] Manuel Lang, Alexander Hornung, Oliver Wang, Steven Poulakos, Aljoscha Smolic, and Markus Gross. Nonlinear disparity mapping for stereoscopic 3d. ACM Transactions on Graphics (TOG), 29(4):1-10, 2010. 5 + +[26] Yann LeCun, Sumit Chopra, Raia Hadsell, M Ranzato, and Fujie Huang. A tutorial on energy-based learning. Predicting structured data, 1(0), 2006. 4 +[27] Zhengqi Li and Noah Snavely. Megadepth: Learning single-view depth prediction from internet photos. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2041-2050, 2018. 1 +[28] Kai-En Lin, Lin Yen-Chen, Wei-Sheng Lai, Tsung-Yi Lin, Yi-Chang Shih, and Ravi Ramamoorthi. Vision transformer for nef-based view synthesis from a single input image. arXiv preprint arXiv:2207.05736, 2022. 2 +[29] Luping Liu, Yi Ren, Zhijie Lin, and Zhou Zhao. Pseudo numerical methods for diffusion models on manifolds. arXiv preprint arXiv:2202.09778, 2022. 6 +[30] Xihui Liu, Dong Huk Park, Samaneh Azadi, Gong Zhang, Arman Chopikyan, Yuxiao Hu, Humphrey Shi, Anna Rohrbach, and Trevor Darrell. More control for free! image synthesis with semantic diffusion guidance. 2021. 4 +[31] Yuan Liu, Sida Peng, Lingjie Liu, Qianqian Wang, Peng Wang, Christian Theobalt, Xiaowei Zhou, and Wenping Wang. Neural rays for occlusion-aware image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7824-7833, 2022. 2 +[32] Calvin Luo. Understanding diffusion models: A unified perspective. arXiv preprint arXiv:2208.11970, 2022. 4 +[33] S. Mahdi H. Miangoleh, Sebastian Dille, Long Mai, Sylvain Paris, and Yaqiz Aksoy. Boosting monocular depth estimation models to high-resolution via content-adaptive multi-resolution merging. 2021. 6 +[34] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European conference on computer vision, pages 405-421. Springer, 2020. 2, 3 +[35] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1 +[36] Fangzhou Mu, Jian Wang, Yicheng Wu, and Yin Li. 3d photo stylization: Learning to generate stylized novel views from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16273-16282, 2022. 2 +[37] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. arXiv preprint arXiv:2201.05989, 2022. 2, 3 +[38] Georgios Pavlakos, Xiaowei Zhou, and Kostas Daniilidis. Ordinal depth supervision for 3d human pose estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7307-7316, 2018. 5 +[39] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 2, 4 +[40] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, + +Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2, 4 +[41] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2, 4, 8 +[42] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 4 +[43] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12179-12188, 2021. 1 +[44] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE transactions on pattern analysis and machine intelligence, 2020. 1, 2, 5 +[45] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding up neural radiance fields with thousands of tiny mlps. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14335-14345, 2021. 2 +[46] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In International Conference on Computer Vision, 2021. 6 +[47] Daniel Roich, Ron Mokady, Amit H Bermano, and Daniel Cohen-Or. Pivotal tuning for latent-based editing of real images. ACM Transactions on Graphics (TOG), 42(1):1-13, 2022. 5 +[48] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2, 4, 5, 8 +[49] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. arXiv preprint arXiv:2208.12242, 2022. 5 +[50] Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 2, 8 +[51] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2304–2314, 2019. 1 +[52] Meng-Li Shih, Shih-Yang Su, Johannes Kopf, and Jia-Bin Huang. 3d photography using context-aware layered depth + +inpainting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8028-8038, 2020. 1, 2 +[53] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256-2265. PMLR, 2015. 3 +[54] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 3, 4 +[55] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in Neural Information Processing Systems, 32, 2019. 4 +[56] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 3, 4 +[57] Mohammed Suhail, Carlos Esteves, Leonid Signal, and Ameesh Makadia. Light field neural rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8269-8279, 2022. 2 +[58] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Improved direct voxel grid optimization for radiance fields reconstruction. arxiv.cs.GR 2206.05085, 2022. 6 +[59] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5459-5469, 2022. 2 +[60] Dani Valevski, Matan Kalman, Yossi Matias, and Yaniv Leviathan. Unitune: Text-driven image editing by fine tuning an image generation model on a single image. arXiv preprint arXiv:2210.09477, 2022. 5 +[61] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T Barron, and Pratul P Srinivasan. Ref-nerf: Structured view-dependent appearance for neural radiance fields. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5481-5490. IEEE, 2022. 2, 6 +[62] Chaoyang Wang, José Miguel Buenapasada, Rui Zhu, and Simon Lucey. Learning depth from monocular videos using direct methods. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2022-2030, 2018. 6 +[63] Can Wang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Text-and-image driven manipulation of neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3835–3844, 2022. 2 +[64] Chaoyang Wang, Simon Lucey, Federico Peruzzi, and Oliver Wang. Web stereo video supervision for depth prediction from dynamic scenes. In 2019 International Conference on 3D Vision (3DV), pages 348-357. IEEE, 2019. 5 +[65] Lizhen Wang, Xiaochen Zhao, Tao Yu, Songtao Wang, and Yebin Liu. Normalgan: Learning detailed 3d human from a single rgb-d image. In European Conference on Computer Vision, pages 430-446. Springer, 2020. 1 + +[66] Nanyang Wang, Yinda Zhang, Zhuwen Li, Yanwei Fu, Wei Liu, and Yu-Gang Jiang. Pixel2mesh: Generating 3d mesh models from single rgb images. In Proceedings of the European conference on computer vision (ECCV), pages 52-67, 2018. 1 +[67] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P Srinivasan, Howard Zhou, Jonathan T Barron, Ricardo Martin-Brualla, Noah Snavely, and Thomas Funkhouser. Ibrnet: Learning multi-view image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4690-4699, 2021. 2 +[68] Jiajun Wu, Yifan Wang, Tianfan Xue, Xingyuan Sun, Bill Freeman, and Josh Tenenbaum. Marrnet: 3d shape reconstruction via 2.5 d sketches. Advances in neural information processing systems, 30, 2017. 1 +[69] Ke Xian, Chunhua Shen, Zhiguo Cao, Hao Lu, Yang Xiao, Ruibo Li, and Zhenbo Luo. Monocular relative depth perception with web stereo data supervision. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 311-320, 2018. 5 +[70] Dejia Xu, Yifan Jiang, Peihao Wang, Zhiwen Fan, Humphrey Shi, and Zhangyang Wang. Sinnerf: Training neural radiance fields on complex scenes from a single image. arXiv preprint arXiv:2204.00928, 2022. 1, 2, 5, 6, 7, 8 +[71] Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. Mvsnet: Depth inference for unstructured multi-view stereo. In Proceedings of the European conference on computer vision (ECCV), pages 767-783, 2018. 1 +[72] Wei Yin, Xinlong Wang, Chunhua Shen, Yifan Liu, Zhi Tian, Songcen Xu, Changming Sun, and Dou Renyin. Diversedepth: Affine-invariant depth prediction using diverse data. arXiv preprint arXiv:2002.00569, 2020. 5 +[73] Wei Yin, Jianming Zhang, Oliver Wang, Simon Niklaus, Long Mai, Simon Chen, and Chunhua Shen. Learning to recover 3d scene shape from a single image. In Proc. IEEE Conf. Comp. Vis. Patt. Recogn. (CVPR), 2021. 1, 6 +[74] Wei Yin, Jianming Zhang, Oliver Wang, Simon Niklaus, Long Mai, Simon Chen, and Chunhua Shen. Learning to recover 3d scene shape from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 204-213, 2021. 5 +[75] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4578-4587, 2021. 1, 2 +[76] Jiahui Yu, Zhe Lin, Jimei Yang, Xiaohui Shen, Xin Lu, and Thomas S Huang. Free-form image inpainting with gated convolution. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4471-4480, 2019. 2 \ No newline at end of file diff --git a/2023/NeuralLift-360_ Lifting an In-the-Wild 2D Photo to a 3D Object With 360deg Views/images.zip b/2023/NeuralLift-360_ Lifting an In-the-Wild 2D Photo to a 3D Object With 360deg Views/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7245e0808a4ec4dcca868de937a3a4850f820889 --- /dev/null +++ b/2023/NeuralLift-360_ Lifting an In-the-Wild 2D Photo to a 3D Object With 360deg Views/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f97453e6e1d1689bc7e6fc7ba96f8e7cbe73195f5f989439feb101ebdba9d1aa +size 593132 diff --git a/2023/NeuralLift-360_ Lifting an In-the-Wild 2D Photo to a 3D Object With 360deg Views/layout.json b/2023/NeuralLift-360_ Lifting an In-the-Wild 2D Photo to a 3D Object With 360deg Views/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..8977418c606b10c87b162191212d0f57c04bfe1e --- /dev/null +++ b/2023/NeuralLift-360_ Lifting an In-the-Wild 2D Photo to a 3D Object With 360deg Views/layout.json @@ -0,0 +1,9272 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 48, + 103, + 547, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 103, + 547, + 121 + ], + "spans": [ + { + "bbox": [ + 48, + 103, + 547, + 121 + ], + "type": "text", + "content": "NeuralLift-360: Lifting An In-the-wild 2D Photo to A 3D Object with " + }, + { + "bbox": [ + 48, + 103, + 547, + 121 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 48, + 103, + 547, + 121 + ], + "type": "text", + "content": " Views" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 107, + 143, + 486, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 143, + 486, + 171 + ], + "spans": [ + { + "bbox": [ + 107, + 143, + 486, + 171 + ], + "type": "text", + "content": "Dejia Xu, Yifan Jiang, Peihao Wang, Zhiwen Fan, Yi Wang, Zhangyang Wang \nVITA Group, University of Texas at Austin" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 221, + 173, + 373, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 221, + 173, + 373, + 185 + ], + "spans": [ + { + "bbox": [ + 221, + 173, + 373, + 185 + ], + "type": "text", + "content": "{dejia,atlaswang}@utexas.edu" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 238, + 290, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 238, + 290, + 501 + ], + "spans": [ + { + "bbox": [ + 47, + 238, + 290, + 501 + ], + "type": "text", + "content": "Virtual reality and augmented reality (XR) bring increasing demand for 3D content generation. However, creating high-quality 3D content requires tedious work from a human expert. In this work, we study the challenging task of lifting a single image to a 3D object and, for the first time, demonstrate the ability to generate a plausible 3D object with " + }, + { + "bbox": [ + 47, + 238, + 290, + 501 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 47, + 238, + 290, + 501 + ], + "type": "text", + "content": " views that corresponds well with the given reference image. By conditioning on the reference image, our model can fulfill the everlasting curiosity for synthesizing novel views of objects from images. Our technique sheds light on a promising direction of easing the workflows for 3D artists and XR designers. We propose a novel framework, dubbed NeuralLift-360, that utilizes a depth-aware neural radiance representation (NeRF) and learns to craft the scene guided by denoising diffusion models. By introducing a ranking loss, our NeuralLift-360 can be guided with rough depth estimation in the wild. We also adopt a CLIP-guided sampling strategy for the diffusion prior to provide coherent guidance. Extensive experiments demonstrate that our NeuralLift-360 significantly outperforms existing state-of-the-art baselines. Project page: https://vita-group.github.io/NeuralLift-360/" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 525, + 128, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 525, + 128, + 538 + ], + "spans": [ + { + "bbox": [ + 47, + 525, + 128, + 538 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 545, + 288, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 545, + 288, + 701 + ], + "spans": [ + { + "bbox": [ + 47, + 545, + 288, + 701 + ], + "type": "text", + "content": "Creating 3D content has been a long-standing problem in computer vision. This problem enables various applications in game studios, home decoration, virtual reality, and augmented reality. Over the past few decades, the manual task has dominated real scenarios, which requires tedious professional expert modeling. Modern artists rely on special software tools (e.g., Blender, Maya3D, 3DS Max, etc.) and time-demanding manual adjustments to realize imaginations and transform them into virtual objects. Meanwhile, automatic 3D content creation pipelines serve as effective tools to facilitate human efforts. These pipelines typically capture hundreds of images and leverage multi-view stereo [71] to quickly model fine-grained virtual landscapes." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 59, + 702, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 702, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 59, + 702, + 287, + 714 + ], + "type": "text", + "content": "More recently, researchers have started aiming at a more" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 213, + 547, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 213, + 547, + 346 + ], + "spans": [ + { + "bbox": [ + 304, + 213, + 547, + 346 + ], + "type": "text", + "content": "ambitious goal, to create a 3D object from a single image [9, 10, 12, 21, 27, 51, 66, 68]. This enables broad applications since it greatly reduces the prerequisite to a minimal request. Existing approaches can be mainly categorized into two major directions. One line of work utilizes learning priors from large-scale datasets with multi-view images [10, 21, 51, 66, 68]. These approaches usually learn a conditional neural network to predict 3D information based on input images. However, due to their poor generalization ability, drastic performance degradation is observed when the testing image is out-of-distribution." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 350, + 548, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 350, + 548, + 578 + ], + "spans": [ + { + "bbox": [ + 304, + 350, + 548, + 578 + ], + "type": "text", + "content": "Another direction constructs the pipeline on top of the depth estimation techniques [65]. Under the guidance of monocular depth estimation networks [43, 44], the 2D image is firstly back-projected to a 3D data format (e.g., point cloud or multi-plane image) and then re-projected into novel views. After that, advanced image inpainting techniques are then adopted to fill missing holes [52] produced during the projection. However, most of them can be highly affected by the quality of the estimated depth maps. Though LeRes [73] attempted to rectify the predicted depth by refining the projected 3D point cloud, their results do not generalize well to an arbitrary image in the wild. Overall, the aforementioned approaches are either adopted in limited scenarios (e.g. face or toy examples [4, 75]), or only produce limited viewing directions [8, 70] when being applied to scenes in the wild. Different from all these approaches, we focus on a more challenging task and for the first time, show promising results by lifting a single in-the-wild image into a 3D object with " + }, + { + "bbox": [ + 304, + 350, + 548, + 578 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 304, + 350, + 548, + 578 + ], + "type": "text", + "content": " novel views." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 582, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 548, + 715 + ], + "type": "text", + "content": "Attracted by the dramatic progress of neural volumetric rendering on 3D reconstruction tasks, we consider building our framework based on Neural Radiance Fields (NeRFs) [35]. The original NeRF takes hundreds of training views and their camera poses as inputs to learn an implicit representation. Subsequent models dedicate tremendous efforts [8, 19, 70, 75] to apply NeRF to sparse training views. The most similar work [70] to our method proposes to optimize a NeRF using only a single image and its corresponding depth map. However, it renders limited views from a small range of angles, and the prerequisite of a high-quality" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4479" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 241, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 241, + 84 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 241, + 84 + ], + "type": "text", + "content": "depth map largely constrains its practical usage." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 85, + 287, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 85, + 287, + 348 + ], + "spans": [ + { + "bbox": [ + 46, + 85, + 287, + 348 + ], + "type": "text", + "content": "To address the above issues, we propose a novel framework, coined as NeuralLift-360, which aims to ease the creation of 3D assets by building a bridge to convert diverse in-the-wild 2D photos to sophisticated 3D contents in " + }, + { + "bbox": [ + 46, + 85, + 287, + 348 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 46, + 85, + 287, + 348 + ], + "type": "text", + "content": " views and enable its automation. The major challenge in our work is that the content on the back side is hidden and hard to hallucinate. To tackle these hurdles, we consider the diffusion priors together with the monocular depth estimator as the cues for hallucination. Modern diffusion models [41,48,50] are trained on a massive dataset (e.g., 5B text-to-image pairs [24]). During inference time, they can generate impressive photorealistic photos based on simple text inputs. By adopting these learning priors with CLIP [40] guidance, NeuralLift-360 can generate plausible 3D consistent instances that correspond well to the given reference image while only requiring minimal additional input, the correct user prompts. Moreover, rather than simply taking the depth map from the pre-trained depth estimator as geometry supervision, we propose to use the relative ranking information from the rough depth during the training process. This simple strategy is observed to robustly mitigate geometry errors for depth estimations in the wild." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 59, + 349, + 257, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 349, + 257, + 360 + ], + "spans": [ + { + "bbox": [ + 59, + 349, + 257, + 360 + ], + "type": "text", + "content": "Our contributions can be summarized as follows," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 374, + 287, + 588 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 59, + 374, + 287, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 374, + 287, + 422 + ], + "spans": [ + { + "bbox": [ + 59, + 374, + 287, + 422 + ], + "type": "text", + "content": "- Given a single image in the wild, we demonstrate promising results of them being lifted to 3D. We use NeRF as an effective scene representation and integrate prior knowledge from the diffusion model." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 434, + 287, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 434, + 287, + 469 + ], + "spans": [ + { + "bbox": [ + 59, + 434, + 287, + 469 + ], + "type": "text", + "content": "- We propose a CLIP-guided sampling strategy that effectively marries the prior knowledge from the diffusion model with the reference image." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 481, + 287, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 481, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 59, + 481, + 287, + 529 + ], + "type": "text", + "content": "- When the reference image is hard to describe exactly, we finetune the diffusion model on the single image while maintaining its ability to generate diverse contents to guide the NeRF training." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 59, + 540, + 287, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 540, + 287, + 588 + ], + "spans": [ + { + "bbox": [ + 59, + 540, + 287, + 588 + ], + "type": "text", + "content": "- We introduce scale-invariant depth supervision that uses the ranking information. This design alleviates the need for accurate multi-view consistent depth estimation and broadens the application of our algorithms." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 601, + 138, + 615 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 601, + 138, + 615 + ], + "spans": [ + { + "bbox": [ + 47, + 601, + 138, + 615 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 622, + 172, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 622, + 172, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 622, + 172, + 635 + ], + "type": "text", + "content": "2.1.3D from Single Image" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "text", + "content": "Inferring 3D worlds from a single image has been a popular direction over the past years. Previous works try to solve this problem by firstly adopting a monocular depth estimator (e.g., MiDaS [44]) to predict the 3D geometry and then using multi-plane images (MPI) [20, 52] or point cloud [36] to render artistic effects. Meanwhile, the missing" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 72, + 545, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 323 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 323 + ], + "type": "text", + "content": "holes from novel views are inpainted by a pre-trained modern deep network [76]. While these approaches are powerful and efficient, the depth estimator could be unstable, and the inpainting techniques may produce artifacts that make the rendered images look fake. While our method also applies a monocular depth estimator as the geometry cues, we propose a depth ranking loss that can mitigate the instability of depth estimation in the wild. Later approaches [28, 75] train their model on a large-scale 3D assets dataset. During inference time, they directly predict a 3D object using a single image. However, their performance faces drastic degradation when being applied to an in-the-wild image, due to the domain gap. The most similar work to ours is SinNeRF [70], which trains a neural radiance field only on a single RGB image and its corresponding depth map. However, their approach requires a high-quality depth map for geometry guidance, and it can only render views from a small range of angles. Thanks to the proposed depth ranking loss, our method can reconstruct impressive geometry information only using guidance from a pre-trained monocular depth estimator." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 330, + 438, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 330, + 438, + 342 + ], + "spans": [ + { + "bbox": [ + 306, + 330, + 438, + 342 + ], + "type": "text", + "content": "2.2. Neural Radiance Fields" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 349, + 545, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 349, + 545, + 504 + ], + "spans": [ + { + "bbox": [ + 304, + 349, + 545, + 504 + ], + "type": "text", + "content": "Neural Radiance Fields (NeRFs) [34] have shown encouraging performance for view synthesis. After its origin, tremendous efforts have been devoted to improving it [2,3,5,6,11,14,16,18,31,37,45,57,59,61,63,67,75]. Unlike most NeRF-like models, which at least require several different views to reconstruct scenes, we focus on reconstructing 3D content with single views. Analogously, recent works [18, 39] try to generate 3D content from text inputs directly. Although they also rely on the text prompt as guidance, their rendered objects are only conditioned on text sentences without image constraints, which can not reach our goal. Different from them, we aim at lifting a single image (either real or synthetic) to photorealistic 3D objects." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 514, + 390, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 514, + 390, + 526 + ], + "spans": [ + { + "bbox": [ + 306, + 514, + 390, + 526 + ], + "type": "text", + "content": "3. Preliminaries" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 534, + 545, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 545, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 545, + 641 + ], + "type": "text", + "content": "Neural Radiance Field NeRFs [34] render images by sampling 5D coordinates (location " + }, + { + "bbox": [ + 304, + 534, + 545, + 641 + ], + "type": "inline_equation", + "content": "(x,y,z)" + }, + { + "bbox": [ + 304, + 534, + 545, + 641 + ], + "type": "text", + "content": " and viewing direction " + }, + { + "bbox": [ + 304, + 534, + 545, + 641 + ], + "type": "inline_equation", + "content": "(\\theta ,\\phi)" + }, + { + "bbox": [ + 304, + 534, + 545, + 641 + ], + "type": "text", + "content": ") along camera rays. After mapping them to color " + }, + { + "bbox": [ + 304, + 534, + 545, + 641 + ], + "type": "inline_equation", + "content": "(r,g,b)" + }, + { + "bbox": [ + 304, + 534, + 545, + 641 + ], + "type": "text", + "content": " and volume density " + }, + { + "bbox": [ + 304, + 534, + 545, + 641 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 304, + 534, + 545, + 641 + ], + "type": "text", + "content": ", additional volumetric rendering techniques are adopted to fuse the information along the ray. Given " + }, + { + "bbox": [ + 304, + 534, + 545, + 641 + ], + "type": "inline_equation", + "content": "\\mathbf{o}" + }, + { + "bbox": [ + 304, + 534, + 545, + 641 + ], + "type": "text", + "content": " as the camera origin and " + }, + { + "bbox": [ + 304, + 534, + 545, + 641 + ], + "type": "inline_equation", + "content": "\\pmb{d}" + }, + { + "bbox": [ + 304, + 534, + 545, + 641 + ], + "type": "text", + "content": " as the ray direction, 3D point locations along the ray can be expressed as " + }, + { + "bbox": [ + 304, + 534, + 545, + 641 + ], + "type": "inline_equation", + "content": "\\pmb {r}(t) = \\pmb {o} + td" + }, + { + "bbox": [ + 304, + 534, + 545, + 641 + ], + "type": "text", + "content": " The predicted color accumulated is defined as follows:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 345, + 647, + 545, + 676 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 647, + 545, + 676 + ], + "spans": [ + { + "bbox": [ + 345, + 647, + 545, + 676 + ], + "type": "interline_equation", + "content": "\\boldsymbol {C} (\\boldsymbol {r}) = \\int_ {t _ {n}} ^ {t _ {f}} T (t) \\sigma (\\boldsymbol {r} (t)) \\boldsymbol {c} (\\boldsymbol {r} (t), \\boldsymbol {d}) d t, \\tag {1}", + "image_path": "c13c9067f60ecb13d1b84c5f6e3b2d449db35c51092c468d8ef91fbc7c166cbd.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 684, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 684, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 684, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 684, + 545, + 713 + ], + "type": "inline_equation", + "content": "T(t) = \\exp \\left(-\\int_{t_n}^t\\sigma (\\boldsymbol {r}(s))ds\\right)" + }, + { + "bbox": [ + 304, + 684, + 545, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 684, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\sigma (\\cdot)" + }, + { + "bbox": [ + 304, + 684, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 684, + 545, + 713 + ], + "type": "inline_equation", + "content": "c(\\cdot ,\\cdot)" + }, + { + "bbox": [ + 304, + 684, + 545, + 713 + ], + "type": "text", + "content": " are densities and color predictions. To numerically evaluate" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4480" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 100, + 69, + 500, + 262 + ], + "blocks": [ + { + "bbox": [ + 100, + 69, + 500, + 262 + ], + "lines": [ + { + "bbox": [ + 100, + 69, + 500, + 262 + ], + "spans": [ + { + "bbox": [ + 100, + 69, + 500, + 262 + ], + "type": "image", + "image_path": "725e14362f498810b5d86d0cb388287e55dd33039ddbcee24e2307a522486150.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 268, + 547, + 292 + ], + "lines": [ + { + "bbox": [ + 46, + 268, + 547, + 292 + ], + "spans": [ + { + "bbox": [ + 46, + 268, + 547, + 292 + ], + "type": "text", + "content": "Figure 1. Overview of the proposed pipeline. NeuralLift-360 learns to recover a 3D object from a single reference image " + }, + { + "bbox": [ + 46, + 268, + 547, + 292 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 268, + 547, + 292 + ], + "type": "text", + "content": ". We derive a prior distillation loss " + }, + { + "bbox": [ + 46, + 268, + 547, + 292 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{diff}}" + }, + { + "bbox": [ + 46, + 268, + 547, + 292 + ], + "type": "text", + "content": " for CLIP-guided diffusion prior and utilize rough depth estimation with the help of a ranking loss " + }, + { + "bbox": [ + 46, + 268, + 547, + 292 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ranking}}" + }, + { + "bbox": [ + 46, + 268, + 547, + 292 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 300, + 288, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 300, + 288, + 350 + ], + "spans": [ + { + "bbox": [ + 46, + 300, + 288, + 350 + ], + "type": "text", + "content": "Eq. 1, the continuous integral is estimated using quadrature [34]. Our model is built upon Instant-NGP [37], which reduces computation cost by inferring color and density from a multiresolution hash table implemented by CUDA." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 365, + 287, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 365, + 287, + 547 + ], + "spans": [ + { + "bbox": [ + 46, + 365, + 287, + 547 + ], + "type": "text", + "content": "Denoising Diffusion Model Denoising diffusion models generate images by gradually denoising from a gaussian noise " + }, + { + "bbox": [ + 46, + 365, + 287, + 547 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}_T) = \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 46, + 365, + 287, + 547 + ], + "type": "text", + "content": " and transforming into a certain data distribution. The forward diffusion process " + }, + { + "bbox": [ + 46, + 365, + 287, + 547 + ], + "type": "inline_equation", + "content": "q(\\pmb {x}_t|\\pmb {x}_{t - 1})" + }, + { + "bbox": [ + 46, + 365, + 287, + 547 + ], + "type": "text", + "content": " adds Gaussian noise to the image " + }, + { + "bbox": [ + 46, + 365, + 287, + 547 + ], + "type": "inline_equation", + "content": "\\pmb {x}_t" + }, + { + "bbox": [ + 46, + 365, + 287, + 547 + ], + "type": "text", + "content": ". The marginal distribution can be written as: " + }, + { + "bbox": [ + 46, + 365, + 287, + 547 + ], + "type": "inline_equation", + "content": "q(\\pmb {x}_t\\mid \\pmb {x_0}) = \\mathcal{N}\\left(\\alpha_t\\pmb {x_0},\\sigma_t^2\\pmb {I}\\right)" + }, + { + "bbox": [ + 46, + 365, + 287, + 547 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 46, + 365, + 287, + 547 + ], + "type": "inline_equation", + "content": "\\alpha_{t}" + }, + { + "bbox": [ + 46, + 365, + 287, + 547 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 365, + 287, + 547 + ], + "type": "inline_equation", + "content": "\\sigma_{t}" + }, + { + "bbox": [ + 46, + 365, + 287, + 547 + ], + "type": "text", + "content": " are designed to converge to " + }, + { + "bbox": [ + 46, + 365, + 287, + 547 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 46, + 365, + 287, + 547 + ], + "type": "text", + "content": " when " + }, + { + "bbox": [ + 46, + 365, + 287, + 547 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 365, + 287, + 547 + ], + "type": "text", + "content": " is at the end of the forward process [23, 56]. The reverse diffusion process " + }, + { + "bbox": [ + 46, + 365, + 287, + 547 + ], + "type": "inline_equation", + "content": "p(\\pmb{x}_{t - 1}|\\pmb {x}_t)" + }, + { + "bbox": [ + 46, + 365, + 287, + 547 + ], + "type": "text", + "content": " learns to denoise. Given infinitesimal timesteps, the reverse diffusion process can be approximated with Gaussian [56] related with an optimal MSE denoiser [53]. The diffusion models are designed as noise estimators " + }, + { + "bbox": [ + 46, + 365, + 287, + 547 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\pmb {x}_t,t)" + }, + { + "bbox": [ + 46, + 365, + 287, + 547 + ], + "type": "text", + "content": " taking noisy images as input and estimating the noise. They are trained via optimizing the weighted evidence lower bound (ELBO) [17, 23]:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 555, + 287, + 576 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 555, + 287, + 576 + ], + "spans": [ + { + "bbox": [ + 59, + 555, + 287, + 576 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {E L B O}} (\\theta) = \\mathbb {E} \\left[ w (t) \\| \\boldsymbol {\\epsilon} _ {\\theta} (\\alpha_ {t} \\pmb {x} _ {0} + \\sigma_ {t} \\pmb {\\epsilon}; t) - \\boldsymbol {\\epsilon} \\| _ {2} ^ {2} \\right], \\quad (2)", + "image_path": "18c8be283788b908571e145c049a9fb92bf2ac51cfd950d452777f89bc1e03ec.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 584, + 287, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 584, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 46, + 584, + 287, + 657 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 584, + 287, + 657 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 46, + 584, + 287, + 657 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 584, + 287, + 657 + ], + "type": "inline_equation", + "content": "w(t)" + }, + { + "bbox": [ + 46, + 584, + 287, + 657 + ], + "type": "text", + "content": " is a weighting function. In practice, setting " + }, + { + "bbox": [ + 46, + 584, + 287, + 657 + ], + "type": "inline_equation", + "content": "w(t) = 1" + }, + { + "bbox": [ + 46, + 584, + 287, + 657 + ], + "type": "text", + "content": " delivers good performance [17]. Sampling from a diffusion model can be either stochastic [17] or deterministic [54]. After sampling " + }, + { + "bbox": [ + 46, + 584, + 287, + 657 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}_T \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 46, + 584, + 287, + 657 + ], + "type": "text", + "content": ", we can gradually reduce the noise level and reach a clean image with high quality at the end of the iterative process." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 669, + 103, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 669, + 103, + 681 + ], + "spans": [ + { + "bbox": [ + 47, + 669, + 103, + 681 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "text", + "content": "Overview. In this section, we propose our method, dubbed NeuralLift-360, which reconstructs a 3D object" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 300, + 545, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 300, + 545, + 432 + ], + "spans": [ + { + "bbox": [ + 304, + 300, + 545, + 432 + ], + "type": "text", + "content": "from a single image. An overview of NeuralLift-360 is illustrated in Fig. 1. NeuralLift-360 combines the best of the two worlds: NeRF and diffusion model, where the former is leveraged to universally model a 3D scene and the latter is utilized to hallucinate the unseen views. To inject knowledge of a diffusion model into a radiance field, we derive a training framework in Sec. 4.1, whose supervision consists of two parts: (1) a prior distillation loss from a generative model (Sec. 4.2), (2) a straight loss on the given image (Sec. 4.4). Moreover, we introduce a domain adaption technique to adapt the diffusion prior to in-the-wild images (Sec. 4.3)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 438, + 476, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 438, + 476, + 451 + ], + "spans": [ + { + "bbox": [ + 305, + 438, + 476, + 451 + ], + "type": "text", + "content": "4.1. Probabilistic-Driven 3D Lifting" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 457, + 545, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 457, + 545, + 541 + ], + "spans": [ + { + "bbox": [ + 304, + 457, + 545, + 541 + ], + "type": "text", + "content": "We begin by formulating our setting along the way, introducing our notations. Given an image " + }, + { + "bbox": [ + 304, + 457, + 545, + 541 + ], + "type": "inline_equation", + "content": "\\pmb{y} \\in \\mathbb{R}^N" + }, + { + "bbox": [ + 304, + 457, + 545, + 541 + ], + "type": "text", + "content": " and its text description " + }, + { + "bbox": [ + 304, + 457, + 545, + 541 + ], + "type": "inline_equation", + "content": "\\pmb{z} \\in \\mathbb{R}^D" + }, + { + "bbox": [ + 304, + 457, + 545, + 541 + ], + "type": "text", + "content": ", NeuralLift-360 intends to reconstruct a 3D scene " + }, + { + "bbox": [ + 304, + 457, + 545, + 541 + ], + "type": "inline_equation", + "content": "\\pmb{V} \\in \\mathbb{R}^M" + }, + { + "bbox": [ + 304, + 457, + 545, + 541 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 457, + 545, + 541 + ], + "type": "inline_equation", + "content": "\\pmb{V}" + }, + { + "bbox": [ + 304, + 457, + 545, + 541 + ], + "type": "text", + "content": " can be either a radiance volume or the network parameters of a NeRF. We regard " + }, + { + "bbox": [ + 304, + 457, + 545, + 541 + ], + "type": "inline_equation", + "content": "\\pmb{y}" + }, + { + "bbox": [ + 304, + 457, + 545, + 541 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 457, + 545, + 541 + ], + "type": "inline_equation", + "content": "\\pmb{z}" + }, + { + "bbox": [ + 304, + 457, + 545, + 541 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 457, + 545, + 541 + ], + "type": "inline_equation", + "content": "\\pmb{V}" + }, + { + "bbox": [ + 304, + 457, + 545, + 541 + ], + "type": "text", + "content": " as random variables. To reconstruct " + }, + { + "bbox": [ + 304, + 457, + 545, + 541 + ], + "type": "inline_equation", + "content": "\\pmb{V}" + }, + { + "bbox": [ + 304, + 457, + 545, + 541 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 304, + 457, + 545, + 541 + ], + "type": "inline_equation", + "content": "\\pmb{y}" + }, + { + "bbox": [ + 304, + 457, + 545, + 541 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 457, + 545, + 541 + ], + "type": "inline_equation", + "content": "\\pmb{z}" + }, + { + "bbox": [ + 304, + 457, + 545, + 541 + ], + "type": "text", + "content": ", we maximize a log-posterior and apply Bayesian rule:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 311, + 548, + 545, + 574 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 548, + 545, + 574 + ], + "spans": [ + { + "bbox": [ + 311, + 548, + 545, + 574 + ], + "type": "interline_equation", + "content": "\\log p (\\boldsymbol {V} | \\boldsymbol {y}, \\boldsymbol {z}) = \\log p (\\boldsymbol {y} | \\boldsymbol {V}, \\boldsymbol {z}) + \\log p (\\boldsymbol {V} | \\boldsymbol {z}) + c o n s t. \\tag {3}", + "image_path": "3734b27ab9d6b91c318d0f05165da85b7791084eba2d76a944ae34f78125ef61.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 582, + 545, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 654 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 654 + ], + "type": "text", + "content": "One can see the prior term " + }, + { + "bbox": [ + 304, + 582, + 545, + 654 + ], + "type": "inline_equation", + "content": "p(\\mathbf{V}|z)" + }, + { + "bbox": [ + 304, + 582, + 545, + 654 + ], + "type": "text", + "content": " provides complementary information to lift an image into 3D space. However, specifying this prior can be intractable, and directly optimizing objective Eq. 3 is also sampling inefficient. Instead, we introduce camera pose " + }, + { + "bbox": [ + 304, + 582, + 545, + 654 + ], + "type": "inline_equation", + "content": "\\Phi \\in \\mathbb{S}\\mathbb{O}(3)\\times \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 582, + 545, + 654 + ], + "type": "text", + "content": " as a latent variable, then the likelihood term can be rewritten as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 315, + 659, + 545, + 712 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 659, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 315, + 659, + 545, + 712 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} p (\\boldsymbol {y} | \\boldsymbol {V}, \\boldsymbol {z}) = \\int p (\\boldsymbol {y} | \\boldsymbol {V}, \\boldsymbol {\\Phi}, \\boldsymbol {z}) p (\\boldsymbol {\\Phi} | \\boldsymbol {V}, \\boldsymbol {z}) d \\boldsymbol {\\Phi} (4) \\\\ = \\int p (\\boldsymbol {y} | \\boldsymbol {V}, \\boldsymbol {\\Phi}, \\boldsymbol {z}) \\frac {p (\\boldsymbol {V} | \\boldsymbol {\\Phi} , \\boldsymbol {z}) p (\\boldsymbol {\\Phi})}{p (\\boldsymbol {V} | \\boldsymbol {z})} d \\boldsymbol {\\Phi}, (5) \\\\ \\end{array}", + "image_path": "d3bd3e2db564bec013c09f07b304499f3675f2b9612990aed636ca3484a1f31b.jpg" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "type": "text", + "content": "4481" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "text", + "content": "where we assume " + }, + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "text", + "content": " is independent of " + }, + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "text", + "content": ". Merging Eq. 3 and Eq. 5, we can get rid of the implicit prior term and obtain a sampling efficient ELBO of the original objective as below:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 116, + 287, + 144 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 116, + 287, + 144 + ], + "spans": [ + { + "bbox": [ + 61, + 116, + 287, + 144 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\log \\mathbb {E} _ {\\boldsymbol {\\Phi} \\sim p (\\boldsymbol {\\Phi})} p (\\boldsymbol {y} | \\boldsymbol {V}, \\boldsymbol {\\Phi}, \\boldsymbol {z}) p (\\boldsymbol {V} | \\boldsymbol {\\Phi}, \\boldsymbol {z}) (6) \\\\ \\geq \\mathbb {E} _ {\\boldsymbol {\\Phi} \\sim p (\\boldsymbol {\\Phi})} \\left[ \\log p (\\boldsymbol {y} | \\boldsymbol {V}, \\boldsymbol {\\Phi}, \\boldsymbol {z}) + \\log p (\\boldsymbol {V} | \\boldsymbol {\\Phi}, \\boldsymbol {z}) \\right]. (7) \\\\ \\end{array}", + "image_path": "f3854c3765a344494aeaec433b07910a2bad1c39dd64ae1090915f24982fea87.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 151, + 287, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 151, + 287, + 199 + ], + "spans": [ + { + "bbox": [ + 47, + 151, + 287, + 199 + ], + "type": "text", + "content": "Furthermore, we define " + }, + { + "bbox": [ + 47, + 151, + 287, + 199 + ], + "type": "inline_equation", + "content": "p(\\boldsymbol{y}|\\boldsymbol{V}, \\boldsymbol{\\Phi}, \\boldsymbol{z}) = p(\\boldsymbol{y}|h(\\boldsymbol{V}, \\boldsymbol{\\Phi}), \\boldsymbol{z})" + }, + { + "bbox": [ + 47, + 151, + 287, + 199 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 151, + 287, + 199 + ], + "type": "inline_equation", + "content": "p(\\boldsymbol{V}|\\boldsymbol{\\Phi}, \\boldsymbol{z}) = p(h(\\boldsymbol{V}, \\boldsymbol{\\Phi})|\\boldsymbol{z})" + }, + { + "bbox": [ + 47, + 151, + 287, + 199 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 151, + 287, + 199 + ], + "type": "inline_equation", + "content": "h(\\boldsymbol{V}, \\boldsymbol{\\Phi})" + }, + { + "bbox": [ + 47, + 151, + 287, + 199 + ], + "type": "text", + "content": " renders the 3D scene with respect to the camera pose " + }, + { + "bbox": [ + 47, + 151, + 287, + 199 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\Phi}" + }, + { + "bbox": [ + 47, + 151, + 287, + 199 + ], + "type": "text", + "content": ". Then we derive our final training objective for NeuralLift-360:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 51, + 205, + 287, + 257 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 205, + 287, + 257 + ], + "spans": [ + { + "bbox": [ + 51, + 205, + 287, + 257 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = - \\mathbb {E} _ {\\boldsymbol {\\Phi}} \\left[ \\underbrace {\\log p (\\boldsymbol {y} \\mid h (\\boldsymbol {V} , \\boldsymbol {\\Phi}) , \\boldsymbol {z})} _ {\\text {r e f e r e n c e d l o s s}} + \\underbrace {\\log p (h (\\boldsymbol {V} , \\boldsymbol {\\Phi}) \\mid \\boldsymbol {z})} _ {\\text {n o n - r e f e r e n c e d l o s s}} \\right]. \\tag {8}", + "image_path": "56427521445ca91b281b7c8af8a44f3c9d2639fed7f7818ef2ee47fd9756604e.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 265, + 287, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 265, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 47, + 265, + 287, + 384 + ], + "type": "text", + "content": "At each training step, we sample a camera pose with respect to a distribution (more on Sec. 5.1), and then render the view associated with this camera parameter (via volume rendering Eq. 1). The loss computed on the rendered image consists of two parts: image-referenced guidance and reference-free regularization. The referenced loss will enforce " + }, + { + "bbox": [ + 47, + 265, + 287, + 384 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 47, + 265, + 287, + 384 + ], + "type": "text", + "content": " to have a close appearance with given " + }, + { + "bbox": [ + 47, + 265, + 287, + 384 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 47, + 265, + 287, + 384 + ], + "type": "text", + "content": ", while the non-referenced term hallucinates the image information with prior knowledge. In the next section, we introduce a diffusion model to implement our loss function Eq. 8." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 391, + 209, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 391, + 209, + 403 + ], + "spans": [ + { + "bbox": [ + 47, + 391, + 209, + 403 + ], + "type": "text", + "content": "4.2. CLIP-guided Diffusion Priors" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 409, + 287, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 409, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 46, + 409, + 287, + 601 + ], + "type": "text", + "content": "Diffusion models [17, 54-56], which synthesize images by gradually denoising from a gaussian noise, have undeniably been successful in the realm of image generation, editing, and processing tasks. When trained on large-scale datasets for text-to-image tasks, these models can accurately represent the distribution of images based on textual descriptions [41, 42, 48]. Consequently, we are inspired to utilize a diffusion model's generative prior to complement the 3D information. To accomplish this, we re-parameterize our Eq. 8 with a diffusion model. Nevertheless, the derived loss function contains both discriminative and generative terms. Taking inspiration from classifier guidance techniques [26, 30, 56], we show that we can leverage the training loss of a diffusion model to surrogate the generative part and utilize CLIP space similarity to model the discriminative part." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 601, + 287, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 601, + 287, + 625 + ], + "spans": [ + { + "bbox": [ + 47, + 601, + 287, + 625 + ], + "type": "text", + "content": "Following the derivation in [17] and [32], we lower bound Eq. 8 by:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 632, + 288, + 683 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 632, + 288, + 683 + ], + "spans": [ + { + "bbox": [ + 48, + 632, + 288, + 683 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\log p _ {\\theta} (\\boldsymbol {y} | \\boldsymbol {x}, \\boldsymbol {z}) + \\log p _ {\\theta} (\\boldsymbol {x} | \\boldsymbol {z}) \\\\ \\geq - \\mathbb {E} _ {t, \\epsilon \\sim \\mathcal {N} (\\mathbf {0}, \\mathbf {I})} \\left[ \\mathcal {D} _ {\\mathrm {K L}} \\left(q \\left(\\boldsymbol {x} _ {t - 1} \\mid \\boldsymbol {x} _ {t}, \\boldsymbol {x} _ {0}\\right) \\| p _ {\\theta} \\left(\\boldsymbol {x} _ {t - 1} \\mid \\boldsymbol {x} _ {t}, \\boldsymbol {y}, \\boldsymbol {z}\\right) \\right] \\right. \\\\ = - \\mathbb {E} _ {t, \\boldsymbol {\\epsilon} \\sim \\mathcal {N} (\\mathbf {0}, \\boldsymbol {I})} \\left[ w (t) \\| \\nabla \\log p _ {\\theta} (\\boldsymbol {x} _ {t} | \\boldsymbol {y}, \\boldsymbol {z}) - \\boldsymbol {\\epsilon} \\| _ {2} ^ {2} \\right], \\tag {9} \\\\ \\end{array}", + "image_path": "0d58c1d6cc4cb0541193576eee12c8b8d97878a78678ffd0124b5ec15eba654b.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{x}_t = \\alpha_t\\pmb{x}_0 + \\sigma_t\\pmb{\\epsilon}" + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "w(t)" + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": " is a constant dependent on noise scale " + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\alpha_{t}" + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\sigma_t" + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": " . The expectation of " + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "t\\sim \\mathcal{U}\\{1,\\dots ,T\\}" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": "is usually taken over a uniform distribution. We provide detailed derivation in supplementary. At first glance, we may directly plug in the pre-trained denoiser " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\pmb{x}_t;\\pmb{z})" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " of a diffusion model to surrogate " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "\\nabla p_{\\theta}(\\pmb{x}_t|\\pmb{y},\\pmb{z})" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " similar to [39]. However, note that " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\pmb{x}_t|\\pmb{y},\\pmb{z})" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " is also conditioned on the reference image " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "\\pmb{y}" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": ", which requires further efforts to inject the reference image information into the diffusion process." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 157, + 545, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 157, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 304, + 157, + 545, + 240 + ], + "type": "text", + "content": "To this end, NeuralLift-360 leverages a Mixed Reference Guidance that guides the diffusion process with a mixture of references. Similar to classifier guidance techniques [26, 30, 32, 56], we rewrite " + }, + { + "bbox": [ + 304, + 157, + 545, + 240 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\boldsymbol{x}_t|\\boldsymbol{y}, \\boldsymbol{z}) = p_{\\theta}(\\boldsymbol{y}|\\boldsymbol{x}_t, \\boldsymbol{z})p_{\\theta}(\\boldsymbol{x}_t|\\boldsymbol{z}) / p_{\\theta}(\\boldsymbol{y}|\\boldsymbol{z})" + }, + { + "bbox": [ + 304, + 157, + 545, + 240 + ], + "type": "text", + "content": ", and we adopt classifier-free guidance to model " + }, + { + "bbox": [ + 304, + 157, + 545, + 240 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\boldsymbol{x}_t|\\boldsymbol{z})" + }, + { + "bbox": [ + 304, + 157, + 545, + 240 + ], + "type": "text", + "content": ", then we can obtain a modified score function [55] as below:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 334, + 252, + 545, + 281 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 252, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 334, + 252, + 545, + 281 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\boldsymbol {\\epsilon} _ {\\theta} (\\boldsymbol {x} _ {t}; \\boldsymbol {y}, \\boldsymbol {z}) = (1 + w) \\boldsymbol {\\epsilon} _ {\\theta} (\\boldsymbol {x} _ {t}; \\boldsymbol {z}) - w \\boldsymbol {\\epsilon} _ {\\theta} (\\boldsymbol {x} _ {t}) \\\\ + \\sqrt {1 - \\alpha_ {t}} \\nabla \\log p _ {\\theta} (\\boldsymbol {y} | \\boldsymbol {x} _ {t}, \\boldsymbol {z}), \\tag {10} \\\\ \\end{array}", + "image_path": "faf961cb35f724554f7eb0aba26f50646620fb8e180c44d0c5bcc6e7555c66d6.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 293, + 545, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 293, + 545, + 411 + ], + "spans": [ + { + "bbox": [ + 304, + 293, + 545, + 411 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 293, + 545, + 411 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 304, + 293, + 545, + 411 + ], + "type": "text", + "content": " is a hyper-parameter that controls the strength of the classifier guidance. We further measure the discrimination with a distance metric on the feature space " + }, + { + "bbox": [ + 304, + 293, + 545, + 411 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\boldsymbol{y}|\\boldsymbol{x}_t, \\boldsymbol{z}) \\propto \\exp\\left[-\\phi\\left(\\frac{\\boldsymbol{x}_t - \\sigma_t \\boldsymbol{\\epsilon}_{\\theta}(\\boldsymbol{x}_t; \\boldsymbol{z})}{\\alpha_t}, \\boldsymbol{y}\\right)\\right]" + }, + { + "bbox": [ + 304, + 293, + 545, + 411 + ], + "type": "text", + "content": ". Specifically, we choose the inner product on the CLIP embedding space [40] as the distance metric, i.e., " + }, + { + "bbox": [ + 304, + 293, + 545, + 411 + ], + "type": "inline_equation", + "content": "\\log p_{\\theta}(\\boldsymbol{y}|\\boldsymbol{x}_t, \\boldsymbol{z}) \\propto -\\left\\langle F\\left(\\frac{\\boldsymbol{x}_t - \\sigma_t \\boldsymbol{\\epsilon}_{\\theta}(\\boldsymbol{x}_t)}{\\alpha_t}\\right), F(\\boldsymbol{y})\\right\\rangle" + }, + { + "bbox": [ + 304, + 293, + 545, + 411 + ], + "type": "text", + "content": ", which spurs on semantic similarity between denoised rendered image and reference image " + }, + { + "bbox": [ + 304, + 293, + 545, + 411 + ], + "type": "inline_equation", + "content": "\\boldsymbol{y}" + }, + { + "bbox": [ + 304, + 293, + 545, + 411 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 293, + 545, + 411 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 304, + 293, + 545, + 411 + ], + "type": "text", + "content": " is the CLIP image encoder." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 411, + 545, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 411, + 545, + 519 + ], + "spans": [ + { + "bbox": [ + 304, + 411, + 545, + 519 + ], + "type": "text", + "content": "In each training iteration, we alternate two camera pose sampling strategies: (1) we fix a camera pose " + }, + { + "bbox": [ + 304, + 411, + 545, + 519 + ], + "type": "inline_equation", + "content": "\\Phi_0" + }, + { + "bbox": [ + 304, + 411, + 545, + 519 + ], + "type": "text", + "content": " (i.e., a Delta distribution) as the reference view, and when rendering on that view, we only penalize the RGB differences by modeling " + }, + { + "bbox": [ + 304, + 411, + 545, + 519 + ], + "type": "inline_equation", + "content": "p(\\boldsymbol{y} | h(\\boldsymbol{V}, \\boldsymbol{\\Phi}), \\boldsymbol{z}) = \\mathcal{N}(\\boldsymbol{y} | h(\\boldsymbol{V}, \\boldsymbol{\\Phi}), \\sigma^2 \\boldsymbol{I})" + }, + { + "bbox": [ + 304, + 411, + 545, + 519 + ], + "type": "text", + "content": ", (2) we stochastically sample camera poses, and adopt the diffusion ELBO (Eq. 9) to estimate the evidence of rendered image together with prior " + }, + { + "bbox": [ + 304, + 411, + 545, + 519 + ], + "type": "inline_equation", + "content": "p_{\\theta}(\\boldsymbol{x} | \\boldsymbol{z})" + }, + { + "bbox": [ + 304, + 411, + 545, + 519 + ], + "type": "text", + "content": ". Combining all together, we can summarize our training loss as below:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 529, + 552, + 601 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 529, + 552, + 601 + ], + "spans": [ + { + "bbox": [ + 306, + 529, + 552, + 601 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} = \\frac {1}{\\sigma^ {2}} \\underbrace {\\| h (\\boldsymbol {V} , \\boldsymbol {\\Phi} _ {0}) - \\boldsymbol {y} \\| _ {2} ^ {2}} _ {\\mathcal {L} _ {\\text {p h o t o m e t r i c}}} \\tag {11} \\\\ + \\underbrace {\\mathbb {E} _ {\\boldsymbol {\\Phi} , t , \\boldsymbol {\\epsilon} \\sim \\mathcal {N} (\\boldsymbol {0} , \\boldsymbol {I})} \\left[ w (t) \\| \\boldsymbol {\\epsilon} _ {\\theta} (\\alpha_ {t} h (\\boldsymbol {V} , \\boldsymbol {\\Phi}) + \\sigma_ {t} \\boldsymbol {\\epsilon} | \\boldsymbol {y} , \\boldsymbol {z}) - \\boldsymbol {\\epsilon} \\| _ {2} ^ {2} \\right]} _ {\\mathcal {L} _ {\\mathrm {d i f f}}}. \\\\ \\end{array}", + "image_path": "9ba3f8534827c2615c6995a5d5846a5fe3c8e0b6fe77e10b959578688cd6e6b4.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 612, + 545, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 612, + 545, + 660 + ], + "spans": [ + { + "bbox": [ + 304, + 612, + 545, + 660 + ], + "type": "text", + "content": "Compared with DreamFusion [39], NeuralLift-360's training objective grounds the 3D scene with the given image which enables the generated scene to have consistent appearance with the specified object." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 670, + 515, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 670, + 515, + 683 + ], + "spans": [ + { + "bbox": [ + 306, + 670, + 515, + 683 + ], + "type": "text", + "content": "4.3. Domain Adaption to In-the-wild Images" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "content": "A primary challenge, however, lies in assigning an appropriate text prompt for our reference image. Given that" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4482" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 95, + 114 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 95, + 114 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 95, + 114 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 95, + 114 + ], + "type": "image", + "image_path": "c160a1d560897fd34353942460214ed10cf85d921bccb454939bd5ead57c525a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 97, + 70, + 143, + 114 + ], + "blocks": [ + { + "bbox": [ + 97, + 70, + 143, + 114 + ], + "lines": [ + { + "bbox": [ + 97, + 70, + 143, + 114 + ], + "spans": [ + { + "bbox": [ + 97, + 70, + 143, + 114 + ], + "type": "image", + "image_path": "42721362303136b9f0d4c3afa5614b7e9326a62cf2614d01f64872c6fca20353.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 120, + 116, + 167, + 125 + ], + "lines": [ + { + "bbox": [ + 120, + 116, + 167, + 125 + ], + "spans": [ + { + "bbox": [ + 120, + 116, + 167, + 125 + ], + "type": "text", + "content": "(a) Pre-trained" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 145, + 70, + 190, + 114 + ], + "blocks": [ + { + "bbox": [ + 145, + 70, + 190, + 114 + ], + "lines": [ + { + "bbox": [ + 145, + 70, + 190, + 114 + ], + "spans": [ + { + "bbox": [ + 145, + 70, + 190, + 114 + ], + "type": "image", + "image_path": "dcc6c2ab0768b4ae14aaa71f1b11d37ea2c94bfd676956c62b305d0e270f13ab.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 192, + 70, + 239, + 114 + ], + "blocks": [ + { + "bbox": [ + 192, + 70, + 239, + 114 + ], + "lines": [ + { + "bbox": [ + 192, + 70, + 239, + 114 + ], + "spans": [ + { + "bbox": [ + 192, + 70, + 239, + 114 + ], + "type": "image", + "image_path": "0ca5b2e3e3d4841da341e3c15d2b921165e698d4ba14c4f6bad700ffc8bbadcd.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 240, + 70, + 287, + 114 + ], + "blocks": [ + { + "bbox": [ + 240, + 70, + 287, + 114 + ], + "lines": [ + { + "bbox": [ + 240, + 70, + 287, + 114 + ], + "spans": [ + { + "bbox": [ + 240, + 70, + 287, + 114 + ], + "type": "image", + "image_path": "428a4ee7e2027c92cbd8c6df576749ac85becc2b47b0a24a471f977851348bac.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 49, + 126, + 95, + 173 + ], + "blocks": [ + { + "bbox": [ + 49, + 126, + 95, + 173 + ], + "lines": [ + { + "bbox": [ + 49, + 126, + 95, + 173 + ], + "spans": [ + { + "bbox": [ + 49, + 126, + 95, + 173 + ], + "type": "image", + "image_path": "d45179ef0aa73b6c8bff223309ce4e4b56a3741cff17d4c47c23e06dd2f8196d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 189, + 286, + 210 + ], + "lines": [ + { + "bbox": [ + 47, + 189, + 286, + 210 + ], + "spans": [ + { + "bbox": [ + 47, + 189, + 286, + 210 + ], + "type": "text", + "content": "Figure 2. After fine-tuning, our model can generate diverse images similar to the reference image (b) while not identical." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 96, + 126, + 142, + 173 + ], + "blocks": [ + { + "bbox": [ + 96, + 126, + 142, + 173 + ], + "lines": [ + { + "bbox": [ + 96, + 126, + 142, + 173 + ], + "spans": [ + { + "bbox": [ + 96, + 126, + 142, + 173 + ], + "type": "image", + "image_path": "faaf30362df5684fbcfaf8eb6b4c80acf115e1a7e0dda21815099486d27d41ef.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 143, + 175, + 190, + 184 + ], + "lines": [ + { + "bbox": [ + 143, + 175, + 190, + 184 + ], + "spans": [ + { + "bbox": [ + 143, + 175, + 190, + 184 + ], + "type": "text", + "content": "(c) Fine-tuned" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 143, + 126, + 190, + 173 + ], + "blocks": [ + { + "bbox": [ + 143, + 126, + 190, + 173 + ], + "lines": [ + { + "bbox": [ + 143, + 126, + 190, + 173 + ], + "spans": [ + { + "bbox": [ + 143, + 126, + 190, + 173 + ], + "type": "image", + "image_path": "599cc7a413195a7fa7d519255148d3e27d323095ebf7fe517fd24034e7771f2f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 192, + 126, + 237, + 173 + ], + "blocks": [ + { + "bbox": [ + 192, + 126, + 237, + 173 + ], + "lines": [ + { + "bbox": [ + 192, + 126, + 237, + 173 + ], + "spans": [ + { + "bbox": [ + 192, + 126, + 237, + 173 + ], + "type": "image", + "image_path": "95620d1a82f2c61d9a7e888d725bbe9f56a2ddd7f0fd50a571b7c947544255ad.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 239, + 126, + 284, + 173 + ], + "blocks": [ + { + "bbox": [ + 239, + 126, + 284, + 173 + ], + "lines": [ + { + "bbox": [ + 239, + 126, + 284, + 173 + ], + "spans": [ + { + "bbox": [ + 239, + 126, + 284, + 173 + ], + "type": "image", + "image_path": "91e36f2d1e4951c26124c9a0ae489047d3ad56f323414eb38a6b9e61c84f90c0.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 221, + 286, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 221, + 286, + 281 + ], + "spans": [ + { + "bbox": [ + 46, + 221, + 286, + 281 + ], + "type": "text", + "content": "large diffusion models are predominantly trained for text-to-image tasks, they typically require a text embedding as conditional input. For images in the wild, it is generally hard to exactly describe an image since images contain more fine-grained information than texts." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 281, + 286, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 281, + 286, + 460 + ], + "spans": [ + { + "bbox": [ + 46, + 281, + 286, + 460 + ], + "type": "text", + "content": "Adapting to in-the-wild images necessitates finetuning the Stable Diffusion model on the reference image Textual inversion [15] and Dreambooth [49] enable finetuning the Stable Diffusion model on a limited number of images for a subject-driven generation. However, in our situation, we only have access to a single image, which significantly hampers the effectiveness of these methods. Imagic [22] and UniTune [60] support finetuning on a single image, but their primary objective is to perform real image editing, which can be quite successful even if the diffusion model overfits the given image. Nevertheless, in our context, we need the diffusion model to generate images resembling the provided reference but not identical to it. If the supervision for each view is the same, NeRF will otherwise produce an isotropic surface." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 460, + 286, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 460, + 286, + 557 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 286, + 557 + ], + "type": "text", + "content": "We draw inspiration from Pivotal Tuning [47] and Imagic [22] when fine-tuning Stable Diffusion [48] on our own. To start with, we optimize the CLIP text embedding " + }, + { + "bbox": [ + 46, + 460, + 286, + 557 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 46, + 460, + 286, + 557 + ], + "type": "text", + "content": " so that it fits our reference image " + }, + { + "bbox": [ + 46, + 460, + 286, + 557 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 460, + 286, + 557 + ], + "type": "text", + "content": " better. During fine-tuning, we use Eq. 2 to supervise the noise estimate. As we expect that our denoised result will resemble the reference image, we also apply a CLIP image loss to the estimated denoised image. " + }, + { + "bbox": [ + 46, + 460, + 286, + 557 + ], + "type": "inline_equation", + "content": "\\hat{x}_0" + }, + { + "bbox": [ + 46, + 460, + 286, + 557 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 58, + 563, + 286, + 611 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 563, + 286, + 611 + ], + "spans": [ + { + "bbox": [ + 58, + 563, + 286, + 611 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\text {f i n e t u n e}} = \\mathbb {E} _ {t, \\epsilon} \\left[ \\| \\boldsymbol {\\epsilon} - \\boldsymbol {\\epsilon} _ {\\theta} (\\mathbf {x} _ {t}; \\boldsymbol {z}) \\| _ {2} ^ {2} \\right] \\\\ - \\mathbb {E} _ {t, \\epsilon} \\left\\langle F \\left(\\frac {\\boldsymbol {x} _ {t} - \\sigma_ {t} \\boldsymbol {\\epsilon} _ {\\theta} (\\boldsymbol {x} _ {t} ; \\boldsymbol {z})}{\\alpha_ {t}}\\right), F (\\boldsymbol {y}) \\right\\rangle , \\tag {12} \\\\ \\end{array}", + "image_path": "bd8f47132b52c9c928ec2dccd3134abb069a354e2fb99dbfae2cf306752c8d3c.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 47, + 617, + 252, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 617, + 252, + 629 + ], + "spans": [ + { + "bbox": [ + 47, + 617, + 252, + 629 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 617, + 252, + 629 + ], + "type": "inline_equation", + "content": "F(\\cdot)" + }, + { + "bbox": [ + 47, + 617, + 252, + 629 + ], + "type": "text", + "content": " is the pre-trained CLIP image encoder." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 46, + 630, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 286, + 712 + ], + "type": "text", + "content": "Then, we fix the optimized text embedding " + }, + { + "bbox": [ + 46, + 630, + 286, + 712 + ], + "type": "inline_equation", + "content": "z^{*}" + }, + { + "bbox": [ + 46, + 630, + 286, + 712 + ], + "type": "text", + "content": " and finetune the diffusion model " + }, + { + "bbox": [ + 46, + 630, + 286, + 712 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\cdot)" + }, + { + "bbox": [ + 46, + 630, + 286, + 712 + ], + "type": "text", + "content": " using the same loss function as above. To preserve the generalization ability of the diffusion model, we introduce several augmentations during the finetuning. Specifically, in each iteration, we perform random image-level augmentations such as flip, transpose, shift, scale, and rotate. As shown in Fig. 2, we are able" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "content": "to overcome the overfitting issue and preserve the diffusion model's ability to generate diverse similar but not identical contents. Similar to Imagic [22] and UniTune [47], we observe that after finetuning, the optimized text embedding will have an overfitting issue and it is critical to perform interpolation between the original text embedding and the optimized text embedding to preserve the diffusion model's ability to generate diverse contents. We obtain " + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "inline_equation", + "content": "z' = \\eta z^* + (1 - \\eta)z" + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "content": " is a weighting hyper-parameter." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 180, + 545, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 180, + 545, + 263 + ], + "spans": [ + { + "bbox": [ + 304, + 180, + 545, + 263 + ], + "type": "text", + "content": "During sampling, we aim to amplify the conditional likelihood for our optimized embedding " + }, + { + "bbox": [ + 304, + 180, + 545, + 263 + ], + "type": "inline_equation", + "content": "z^{*}" + }, + { + "bbox": [ + 304, + 180, + 545, + 263 + ], + "type": "text", + "content": ", while decreasing both the unconditional likelihood and the conditional likelihood for the original embedding " + }, + { + "bbox": [ + 304, + 180, + 545, + 263 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 304, + 180, + 545, + 263 + ], + "type": "text", + "content": ". Hence for 50% of all iterations, we use the original Eq. 10, while for the others, we replace the unconditional sampling " + }, + { + "bbox": [ + 304, + 180, + 545, + 263 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\boldsymbol{x}_t)" + }, + { + "bbox": [ + 304, + 180, + 545, + 263 + ], + "type": "text", + "content": " in Eq. 10 with conditional sampling using the original text embedding " + }, + { + "bbox": [ + 304, + 180, + 545, + 263 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 304, + 180, + 545, + 263 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 322, + 273, + 545, + 302 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 273, + 545, + 302 + ], + "spans": [ + { + "bbox": [ + 322, + 273, + 545, + 302 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\widetilde {\\boldsymbol {\\epsilon}} _ {\\theta} \\left(\\boldsymbol {x} _ {t}; \\boldsymbol {y}, \\boldsymbol {z} ^ {\\prime}, \\boldsymbol {z}\\right) = (1 + w) \\boldsymbol {\\epsilon} _ {\\theta} \\left(\\boldsymbol {x} _ {t}; \\boldsymbol {z} ^ {\\prime}\\right) \\tag {13} \\\\ - w \\epsilon_ {\\theta} (\\boldsymbol {x} _ {t}; \\boldsymbol {z}) + \\sqrt {1 - \\alpha_ {t}} \\nabla \\log p _ {\\theta} (\\boldsymbol {y} | \\boldsymbol {x} _ {t}, \\boldsymbol {z} ^ {\\prime}), \\\\ \\end{array}", + "image_path": "7fe0d4e924c071788fea4e22f3883643d359d929372d2b19b1940623181dc85a.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 312, + 474, + 324 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 312, + 474, + 324 + ], + "spans": [ + { + "bbox": [ + 306, + 312, + 474, + 324 + ], + "type": "text", + "content": "4.4. Supervision from Rough Depth" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 331, + 545, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 331, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 331, + 545, + 449 + ], + "type": "text", + "content": "While the aforementioned pipeline laid the foundation for the success of our method, we observe that obtaining a high-quality 3D object without geometry regularization is inherently challenging, if relying solely on learning priors. As shown in SinNeRF [70] and DS-NeRF [8], accurate depth information can largely help improve the rendering quality. However, their impressive outcomes are contingent upon the 3D consistency of the accessible high-quality depth, a requirement that is not easily fulfilled in real-life scenarios." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 450, + 545, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 450, + 545, + 593 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 545, + 593 + ], + "type": "text", + "content": "In spite of the significant progress in monocular depth estimation for in-the-wild scenarios, current state-of-the-art depth estimators are insufficient for recovering 3D shapes. Due to unknown camera baselines and stereoscopic post-processing in data, these methods [7,25,44,64,69,72] utilize loss functions invariant to shift and scale or adopt ranking loss that utilizes relative information. Therefore, their results cannot be used to reconstruct plausible 3D scene shapes directly due to unknown depth shifts [74]. Though there are also attempts [74] that improve off-the-shelf depth estimators, they generally can not generalize well to unseen objects in the wild." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "Given that the acquired depth is imperfect, the absolute depth scale lacks the reliability necessary for recovering a high-quality 3D object. In contrast to previous depth-supervised NeRFs [8, 70], which rely solely on precise values from the depth map, our approach represents the first attempt to maximize the utilization of depth information by adopting the relative ranking information. Drawing inspiration from [7, 38], we adopt a pairwise depth ranking loss to supervise the rendered depth from our NeRF backbone. In this way, we enforce the depth we rendered to be consistent" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4483" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "type": "text", + "content": "with the depth from the reference view in terms of internal ranking instead of actual value. The loss is formulated as follows," + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 112, + 287, + 143 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 112, + 287, + 143 + ], + "spans": [ + { + "bbox": [ + 67, + 112, + 287, + 143 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r a n k i n g}} = \\left\\{ \\begin{array}{l l} \\max \\left(z _ {j _ {k}} - z _ {i _ {k}}, 0\\right), & r _ {k} \\in \\{> \\} \\\\ \\max \\left(z _ {i _ {k}} - z _ {j _ {k}}, 0\\right), & r _ {k} \\in \\{< \\} \\end{array} \\right. \\tag {14}", + "image_path": "9136c2926836df6af9adce4793133c3bb9925ef8e957817f581207cb08b3c36f.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 147, + 287, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 147, + 287, + 207 + ], + "spans": [ + { + "bbox": [ + 46, + 147, + 287, + 207 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 147, + 287, + 207 + ], + "type": "inline_equation", + "content": "r_k" + }, + { + "bbox": [ + 46, + 147, + 287, + 207 + ], + "type": "text", + "content": " refers to the pairwise ranking relationship between pixel " + }, + { + "bbox": [ + 46, + 147, + 287, + 207 + ], + "type": "inline_equation", + "content": "z_{i_k}" + }, + { + "bbox": [ + 46, + 147, + 287, + 207 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 147, + 287, + 207 + ], + "type": "inline_equation", + "content": "z_{j_k}" + }, + { + "bbox": [ + 46, + 147, + 287, + 207 + ], + "type": "text", + "content": " in the monocular depth estimation. This loss function is used alongside the photo-metric loss and diffusion loss mentioned in Eq. 11 and serves as additional supervision on geometry." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 213, + 128, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 213, + 128, + 225 + ], + "spans": [ + { + "bbox": [ + 47, + 213, + 128, + 225 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 230, + 287, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 230, + 287, + 265 + ], + "spans": [ + { + "bbox": [ + 46, + 230, + 287, + 265 + ], + "type": "text", + "content": "In this section, we will first describe a unique training recipe that we propose to train our framework, then the implementation details." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 268, + 145, + 281 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 268, + 145, + 281 + ], + "spans": [ + { + "bbox": [ + 47, + 268, + 145, + 281 + ], + "type": "text", + "content": "5.1. Training Recipe" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 283, + 287, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 283, + 287, + 511 + ], + "spans": [ + { + "bbox": [ + 46, + 283, + 287, + 511 + ], + "type": "text", + "content": "Diverse Camera Pose and Intrinsic Sampling During training, we observe that a diverse camera pose and intrinsicics are essential for the model to learn a coherent 3D object. If the camera pose is not sampled from a diverse distribution, the 3D object will collapse towards the major viewing directions and thus generate unpleasant results. Specifically, we alter between two pose sampling strategies; (1) we have a fixed camera pose " + }, + { + "bbox": [ + 46, + 283, + 287, + 511 + ], + "type": "inline_equation", + "content": "\\Phi_0" + }, + { + "bbox": [ + 46, + 283, + 287, + 511 + ], + "type": "text", + "content": " (i.e., a Delta distribution) for the reference view, (2) we uniformly sample the camera rotation on the surface of a unit sphere and then uniformly sample the radius in [0.4, 1.0] to locate the camera. We start with the Delta distribution, and as training progresses, we gradually shift to the other sophisticated camera pose distribution. The camera intrinsics provide additional augmentation on the viewing direction. By uniformly sampling the field of view in [50, 70], we zoom in and out to get different resolution projections for the scene. We also add Gaussian noise to the camera pose sampling phase. Such perturbations help us render more diverse images during training." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 525, + 287, + 680 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 525, + 287, + 680 + ], + "spans": [ + { + "bbox": [ + 46, + 525, + 287, + 680 + ], + "type": "text", + "content": "Foreground Aware Diffusion Prior ERNIE-ViLG [13] utilizes a weighting mechanism during the training of diffusion models to let the model generate. Using an additional object detection model, their weighted loss forces the model to focus on salient regions. Similarly, we hope to strengthen the guidance on the foreground object region. We propose to weight the score matching distance with the predicted density. Specifically, we penalize the regions with higher density so that our NeRF gains more guidance on the foreground regions. Hence, for each rendered image, we first obtain the bounding box for the non-zero density regions. Then, we extend the " + }, + { + "bbox": [ + 46, + 525, + 287, + 680 + ], + "type": "inline_equation", + "content": "w(t)" + }, + { + "bbox": [ + 46, + 525, + 287, + 680 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 46, + 525, + 287, + 680 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{diff}}" + }, + { + "bbox": [ + 46, + 525, + 287, + 680 + ], + "type": "text", + "content": " (Eq. 11) into a spatial function:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 59, + 685, + 287, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 685, + 287, + 716 + ], + "spans": [ + { + "bbox": [ + 59, + 685, + 287, + 716 + ], + "type": "interline_equation", + "content": "W _ {i, j} = \\left\\{ \\begin{array}{l l} 2, & \\text {i f} (i, j) \\text {i s i n s i d e t h e b o u n d i n g b o x} \\\\ 1, & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {15}", + "image_path": "01d231248fd779026e2c8ca7e9da2ff3433c3186e34ba96ce3ffe38489c0b608.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "text", + "content": "Geometry Regularization We also include a series of geometry regularizations to avoid undesired artifacts in our 3D object. To avoid foggy and semi-transparent artifacts at the back of the scene, we utilize the orient loss from Ref-NeRF [61] to prevent the surface normals from facing backward from the camera. We also incorporate depth smoothness loss [62], sparsity loss, and distortion loss [3, 58] to avoid the NeRF filling empty space with floaters. Full details for these regularizations are provided in the supplementary material." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 198, + 545, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 198, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 304, + 198, + 545, + 304 + ], + "type": "text", + "content": "Timestep Annealing At the start of training, the quality of the rendered images is poor, and we need the diffusion prior to provide more thorough guidance. As a result, we need to add more noise to perturb the rendered image. As the training progresses, the quality of rendered images gradually improves, and we choose to reduce the timestep range accordingly. By annealing the timestep where we add noise, we are lowering the noise level of the perturbation, and the noise estimate becomes more accurate." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 312, + 427, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 312, + 427, + 323 + ], + "spans": [ + { + "bbox": [ + 306, + 312, + 427, + 323 + ], + "type": "text", + "content": "5.2. Evaluation Workflow" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 330, + 545, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 330, + 545, + 497 + ], + "spans": [ + { + "bbox": [ + 304, + 330, + 545, + 497 + ], + "type": "text", + "content": "We conduct extensive experiments using both text-to-image synthetic images and real-world images. Synthetic images in our experiments are generated using PNDM Sampler [29] with 50 sampling steps. Real images are either from CO3Dv2 dataset [46] or randomly crawled from the Google search engine. For each image, we use a pre-trained background removal framework [1] to segment the foreground objects. Then we obtain their depth using LeRes [73] with the \"Boost Your Own depth\" [33] strategy. [33] inferences the monocular depth estimation network on multiple patches and then fuses the results for better details. We compare with previous sparse view NeRFs, including Depth-supervised NeRF [8] and SinNeRF [70]. We also compare with DietNeRF [19] that uses CLIP as guidance." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 504, + 391, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 504, + 391, + 516 + ], + "spans": [ + { + "bbox": [ + 306, + 504, + 391, + 516 + ], + "type": "text", + "content": "5.3. Comparisons" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 522, + 545, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 545, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 545, + 629 + ], + "type": "text", + "content": "We present quantitative results in eight scenes. We report the average CLIP distance between rendered image and the reference image as a metric for how well the object is represented. During testing, we use a CLIP model different from training time and evaluate on held out viewing directions with a larger distance-to-origin than training. We render 100 images on each scene for each method. As shown in Tab. 1, our method achieves the best CLIP distance among existing approaches." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "We also provide visual comparisons in Fig. 3. Since DSNeRF and SinNeRF rely on the exact value of the available depth information, their performance using monocular depth estimation is far from satisfactory. It can also be seen that the DietNeRF suffers from unconstrained geometry. Our method, on the other hand, is able to synthesize visually pleasing novel views with 3D consistency." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4484" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 69, + 532, + 692 + ], + "blocks": [ + { + "bbox": [ + 61, + 69, + 532, + 692 + ], + "lines": [ + { + "bbox": [ + 61, + 69, + 532, + 692 + ], + "spans": [ + { + "bbox": [ + 61, + 69, + 532, + 692 + ], + "type": "image", + "image_path": "2fa5abf08f79858788196ec86283d10e346a75da03b8b32250595477e658dccc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 91, + 699, + 501, + 712 + ], + "lines": [ + { + "bbox": [ + 91, + 699, + 501, + 712 + ], + "spans": [ + { + "bbox": [ + 91, + 699, + 501, + 712 + ], + "type": "text", + "content": "Figure 3. Visual comparisons for novel view synthesis results. We provide video comparisons in supplementary." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4485" + } + ] + } + ], + "index": 2 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 75, + 91, + 519, + 147 + ], + "blocks": [ + { + "bbox": [ + 96, + 71, + 496, + 82 + ], + "lines": [ + { + "bbox": [ + 96, + 71, + 496, + 82 + ], + "spans": [ + { + "bbox": [ + 96, + 71, + 496, + 82 + ], + "type": "text", + "content": "Table 1. Quantitative evaluation by comparing the proposed method with previous state-of-the-art approaches." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 75, + 91, + 519, + 147 + ], + "lines": [ + { + "bbox": [ + 75, + 91, + 519, + 147 + ], + "spans": [ + { + "bbox": [ + 75, + 91, + 519, + 147 + ], + "type": "table", + "html": "
CLIP Distance ↓AppleBenchBaseballHydrantRemoteSuitcaseStatueSandwichAverage
DSNeRF [8]0.53090.54470.55060.62260.64410.55330.59920.61100.5821
DietNeRF [19]0.42450.55790.45450.51900.57180.48900.59470.51790.5162
SinNeRF [70]0.51920.57690.62300.63430.64960.57050.56050.60150.5918
NeuralLift-360 (Ours)0.39140.47850.39740.51460.49160.41860.44110.46550.4498
", + "image_path": "ff3c4fdd375b5f1de9a5db6dfb3757260018fe029f63d6de1866ace6f89a070c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 63, + 151, + 531, + 265 + ], + "blocks": [ + { + "bbox": [ + 63, + 151, + 531, + 265 + ], + "lines": [ + { + "bbox": [ + 63, + 151, + 531, + 265 + ], + "spans": [ + { + "bbox": [ + 63, + 151, + 531, + 265 + ], + "type": "image", + "image_path": "5b63a43d9c64fc35ae3add3e12b667955b411b2ac73dc82001d688914b37a7da.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 56, + 271, + 536, + 282 + ], + "lines": [ + { + "bbox": [ + 56, + 271, + 536, + 282 + ], + "spans": [ + { + "bbox": [ + 56, + 271, + 536, + 282 + ], + "type": "text", + "content": "Figure 4. Ablation study on the baseline variants. Given a single reference image, we run different variants to render its " + }, + { + "bbox": [ + 56, + 271, + 536, + 282 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 56, + 271, + 536, + 282 + ], + "type": "text", + "content": " views." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 61, + 291, + 272, + 548 + ], + "blocks": [ + { + "bbox": [ + 61, + 291, + 272, + 548 + ], + "lines": [ + { + "bbox": [ + 61, + 291, + 272, + 548 + ], + "spans": [ + { + "bbox": [ + 61, + 291, + 272, + 548 + ], + "type": "image", + "image_path": "eaf753534202c57de45f178e63bbb73cb703165fb5d44bd6fe894366ac993bf9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 556, + 287, + 567 + ], + "lines": [ + { + "bbox": [ + 47, + 556, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 47, + 556, + 287, + 567 + ], + "type": "text", + "content": "Figure 5. Ablation study on CLIP guidance for the diffusion prior." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 573, + 141, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 573, + 141, + 586 + ], + "spans": [ + { + "bbox": [ + 47, + 573, + 141, + 586 + ], + "type": "text", + "content": "5.4. Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "type": "text", + "content": "Variants of supervision. In this section, we perform detailed ablation studies on different components of our framework. \"w/o " + }, + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ranking}}" + }, + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "type": "text", + "content": " \" replaces the ranking loss with " + }, + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "type": "text", + "content": " loss on depth. \"w/o " + }, + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{diff}}" + }, + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "type": "text", + "content": " \" replaces the CLIP guided diffusion prior with CLIP. \"w/o fine-tuning\" refers to using the pre-trained diffusion model. As shown in Fig. 4, our full model delivers the best performance, while other variants have different issues. \"w/o " + }, + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ranking}}" + }, + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "type": "text", + "content": " \" relies on the exact value from depth information and suffers from unreliable depth. \"w/o " + }, + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{diff}}" + }, + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "type": "text", + "content": " \" uses CLIP as a prior, and the optimized result is" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 292, + 545, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 292, + 545, + 340 + ], + "spans": [ + { + "bbox": [ + 305, + 292, + 545, + 340 + ], + "type": "text", + "content": "blurry and collapsed. \"w/o fine-tuning\" uses a pre-trained diffusion model, since the text prompt is not very accurate, the resulting object has lettuce which doesn't exist in the reference image." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 365, + 546, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 365, + 546, + 449 + ], + "spans": [ + { + "bbox": [ + 305, + 365, + 546, + 449 + ], + "type": "text", + "content": "Effectiveness of CLIP guided diffusion prior We validate the effectiveness of CLIP guidance in our diffusion prior design in Fig. 5. It can be seen that without the CLIP guidance, although the model produces a 3D object, it doesn't look like the reference image and has a strange shape. In comparison, our full model generates a better shape and looks more similar to the reference image." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 471, + 476, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 471, + 476, + 483 + ], + "spans": [ + { + "bbox": [ + 306, + 471, + 476, + 483 + ], + "type": "text", + "content": "6. Conclusions and Future Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 495, + 545, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 495, + 545, + 626 + ], + "spans": [ + { + "bbox": [ + 304, + 495, + 545, + 626 + ], + "type": "text", + "content": "We present NeuralLift-360, a novel framework to lift an in-the-wild 2D photo into a 3D object with 360-degree views. NeuralLift-360 learns probabilistic-driven 3D lifting with CLIP-guided diffusion priors and mitigates the depth errors by a scale-invariant depth ranking loss. Comprehensive experiments are conducted on real and synthetic images, where NeuralLift-360 outperforms the current state-of-the-art methods. While our current implementation is based on Stable Diffusion [48] version 1.4, the framework should also work well with other diffusion models such as Imagen [50] or DALL-E2 [41], once they are open-sourced." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "Limitations. Despite the encouraging visual results rendered by NeuralLift-360, the target resolution " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "(128 \\times 128)" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " is still behind large generative models [41, 50]. Moreover, challenging cases (e.g., multiple objects with occlusion) are not included in our assumptions, which need to be further explored. Going forward, we will investigate how to expand our method to more general scenarios." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4486" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "text", + "content": "[1] Image background removal. https://github.com/OPHoperHPO/image-background-remove-tool. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 126, + 288, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 126, + 288, + 190 + ], + "spans": [ + { + "bbox": [ + 53, + 126, + 288, + 190 + ], + "type": "text", + "content": "[2] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 194, + 288, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 194, + 288, + 247 + ], + "spans": [ + { + "bbox": [ + 53, + 194, + 288, + 247 + ], + "type": "text", + "content": "[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5470–5479, 2022. 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 250, + 287, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 250, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 53, + 250, + 287, + 327 + ], + "type": "text", + "content": "[4] Angel X. Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. ShapeNet: An Information-Rich 3D Model Repository. Technical Report arXiv:1512.03012 [cs.GR], Stanford University — Princeton University — Toyota Technological Institute at Chicago, 2015. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 328, + 287, + 383 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 328, + 287, + 383 + ], + "spans": [ + { + "bbox": [ + 53, + 328, + 287, + 383 + ], + "type": "text", + "content": "[5] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14124-14133, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 384, + 287, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 384, + 287, + 439 + ], + "spans": [ + { + "bbox": [ + 53, + 384, + 287, + 439 + ], + "type": "text", + "content": "[6] Tianlong Chen, Peihao Wang, Zhiwen Fan, and Zhangyang Wang. Aug-nerf: Training stronger neural radiance fields with triple-level physically-grounded augmentations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15191-15202, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 441, + 287, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 441, + 287, + 474 + ], + "spans": [ + { + "bbox": [ + 53, + 441, + 287, + 474 + ], + "type": "text", + "content": "[7] Weifeng Chen, Zhao Fu, Dawei Yang, and Jia Deng. Single-image depth perception in the wild. Advances in neural information processing systems, 29, 2016. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 475, + 287, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 475, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 53, + 475, + 287, + 529 + ], + "type": "text", + "content": "[8] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised nerf: Fewer views and faster training for free. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12882-12891, 2022. 1, 5, 6, 7, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 532, + 287, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 532, + 287, + 575 + ], + "spans": [ + { + "bbox": [ + 53, + 532, + 287, + 575 + ], + "type": "text", + "content": "[9] Jean-Denis Durou, Maurizio Falcone, and Manuela Sagona. Numerical methods for shape-from-shading: A new survey with benchmarks. Computer Vision and Image Understanding, 109(1):22-43, 2008. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 577, + 287, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 577, + 287, + 620 + ], + "spans": [ + { + "bbox": [ + 48, + 577, + 287, + 620 + ], + "type": "text", + "content": "[10] Haoqiang Fan, Hao Su, and Leonidas J Guibas. A point set generation network for 3d object reconstruction from a single image. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 605-613, 2017. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 622, + 287, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 622, + 287, + 655 + ], + "spans": [ + { + "bbox": [ + 48, + 622, + 287, + 655 + ], + "type": "text", + "content": "[11] Zhiwen Fan, Yifan Jiang, Peihao Wang, Xinyu Gong, Dejia Xu, and Zhangyang Wang. Unified implicit neural stylization. arXiv preprint arXiv:2204.01943, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 657, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 657, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 657, + 287, + 689 + ], + "type": "text", + "content": "[12] Paolo Favaro and Stefano Soatto. A geometric approach to shape from defocus. IEEE Transactions on Pattern Analysis and Machine Intelligence, 27(3):406-417, 2005. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "text", + "content": "[13] Zhida Feng, Zhenyu Zhang, Xintong Yu, Yewei Fang, Lanxin Li, Xuyi Chen, Yuxiang Lu, Jiaxiang Liu, Weichong" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 116 + ], + "type": "text", + "content": "Yin, Shikun Feng, et al. Ernie-vilg 2.0: Improving text-to-image diffusion model with knowledge-enhanced mixture-of-denoising-experts. arXiv preprint arXiv:2210.15257, 2022.6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 118, + 547, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 118, + 547, + 172 + ], + "spans": [ + { + "bbox": [ + 308, + 118, + 547, + 172 + ], + "type": "text", + "content": "[14] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5501–5510, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 173, + 545, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 173, + 545, + 226 + ], + "spans": [ + { + "bbox": [ + 308, + 173, + 545, + 226 + ], + "type": "text", + "content": "[15] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 228, + 545, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 228, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 308, + 228, + 545, + 281 + ], + "type": "text", + "content": "[16] Yuan-Chen Guo, Di Kang, Linchao Bao, Yu He, and Song-Hai Zhang. Nerfren: Neural radiance fields with reflections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18409-18418, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 282, + 545, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 282, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 308, + 282, + 545, + 316 + ], + "type": "text", + "content": "[17] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 3, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 316, + 545, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 316, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 308, + 316, + 545, + 370 + ], + "type": "text", + "content": "[18] Ajay Jain, Ben Mildenhall, Jonathan T Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 867-876, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 372, + 545, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 372, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 308, + 372, + 545, + 415 + ], + "type": "text", + "content": "[19] Ajay Jain, Matthew Tancik, and Pieter Abbeel. Putting nerf on a diet: Semantically consistent few-shot view synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5885-5894, 2021. 1, 6, 7, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 416, + 545, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 416, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 308, + 416, + 545, + 491 + ], + "type": "text", + "content": "[20] Varun Jampani, Huiwen Chang, Kyle Sargent, Abhishek Kar, Richard Tucker, Michael Krainin, Dominik Kaeser, William T Freeman, David Salesin, Brian Curless, et al. Slide: Single image 3d photography with soft layering and depth-aware inpainting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12518-12527, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 493, + 545, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 493, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 308, + 493, + 545, + 547 + ], + "type": "text", + "content": "[21] Adrian Johnston, Ravi Garg, Gustavo Carneiro, Ian Reid, and Anton van den Hengel. Scaling cnns for high resolution volumetric reconstruction from a single image. In Proceedings of the IEEE International Conference on Computer Vision Workshops, pages 939-948, 2017. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 548, + 545, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 548, + 545, + 592 + ], + "spans": [ + { + "bbox": [ + 308, + 548, + 545, + 592 + ], + "type": "text", + "content": "[22] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagc: Text-based real image editing with diffusion models. arXiv preprint arXiv:2210.09276, 2022. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 592, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 592, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 592, + 545, + 624 + ], + "type": "text", + "content": "[23] Diederik Kingma, Tim Salimans, Ben Poole, and Jonathan Ho. Variational diffusion models. Advances in neural information processing systems, 34:21696-21707, 2021. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 625, + 545, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 625, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 308, + 625, + 545, + 669 + ], + "type": "text", + "content": "[24] Hyung-Kwon Ko, Gwanmo Park, Hyeon Jeon, Jaemin Jo, Juho Kim, and Jinwook Seo. Large-scale text-to-image generation models for visual artists' creative works. arXiv preprint arXiv:2210.08477, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "text", + "content": "[25] Manuel Lang, Alexander Hornung, Oliver Wang, Steven Poulakos, Aljoscha Smolic, and Markus Gross. Nonlinear disparity mapping for stereoscopic 3d. ACM Transactions on Graphics (TOG), 29(4):1-10, 2010. 5" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4487" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 715 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[26] Yann LeCun, Sumit Chopra, Raia Hadsell, M Ranzato, and Fujie Huang. A tutorial on energy-based learning. Predicting structured data, 1(0), 2006. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 287, + 150 + ], + "type": "text", + "content": "[27] Zhengqi Li and Noah Snavely. Megadepth: Learning single-view depth prediction from internet photos. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2041-2050, 2018. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 150, + 287, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 150, + 287, + 194 + ], + "spans": [ + { + "bbox": [ + 48, + 150, + 287, + 194 + ], + "type": "text", + "content": "[28] Kai-En Lin, Lin Yen-Chen, Wei-Sheng Lai, Tsung-Yi Lin, Yi-Chang Shih, and Ravi Ramamoorthi. Vision transformer for nef-based view synthesis from a single input image. arXiv preprint arXiv:2207.05736, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 194, + 287, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 194, + 287, + 227 + ], + "spans": [ + { + "bbox": [ + 48, + 194, + 287, + 227 + ], + "type": "text", + "content": "[29] Luping Liu, Yi Ren, Zhijie Lin, and Zhou Zhao. Pseudo numerical methods for diffusion models on manifolds. arXiv preprint arXiv:2202.09778, 2022. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 227, + 287, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 227, + 287, + 272 + ], + "spans": [ + { + "bbox": [ + 48, + 227, + 287, + 272 + ], + "type": "text", + "content": "[30] Xihui Liu, Dong Huk Park, Samaneh Azadi, Gong Zhang, Arman Chopikyan, Yuxiao Hu, Humphrey Shi, Anna Rohrbach, and Trevor Darrell. More control for free! image synthesis with semantic diffusion guidance. 2021. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 272, + 287, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 272, + 287, + 337 + ], + "spans": [ + { + "bbox": [ + 48, + 272, + 287, + 337 + ], + "type": "text", + "content": "[31] Yuan Liu, Sida Peng, Lingjie Liu, Qianqian Wang, Peng Wang, Christian Theobalt, Xiaowei Zhou, and Wenping Wang. Neural rays for occlusion-aware image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7824-7833, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 337, + 287, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 337, + 287, + 360 + ], + "spans": [ + { + "bbox": [ + 48, + 337, + 287, + 360 + ], + "type": "text", + "content": "[32] Calvin Luo. Understanding diffusion models: A unified perspective. arXiv preprint arXiv:2208.11970, 2022. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 360, + 287, + 405 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 360, + 287, + 405 + ], + "spans": [ + { + "bbox": [ + 48, + 360, + 287, + 405 + ], + "type": "text", + "content": "[33] S. Mahdi H. Miangoleh, Sebastian Dille, Long Mai, Sylvain Paris, and Yaqiz Aksoy. Boosting monocular depth estimation models to high-resolution via content-adaptive multi-resolution merging. 2021. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 404, + 287, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 404, + 287, + 459 + ], + "spans": [ + { + "bbox": [ + 48, + 404, + 287, + 459 + ], + "type": "text", + "content": "[34] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European conference on computer vision, pages 405-421. Springer, 2020. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 459, + 287, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 459, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 48, + 459, + 287, + 514 + ], + "type": "text", + "content": "[35] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 514, + 287, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 287, + 568 + ], + "type": "text", + "content": "[36] Fangzhou Mu, Jian Wang, Yicheng Wu, and Yin Li. 3d photo stylization: Learning to generate stylized novel views from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16273-16282, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 569, + 287, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 569, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 569, + 287, + 613 + ], + "type": "text", + "content": "[37] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. arXiv preprint arXiv:2201.05989, 2022. 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 614, + 287, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 287, + 658 + ], + "type": "text", + "content": "[38] Georgios Pavlakos, Xiaowei Zhou, and Kostas Daniilidis. Ordinal depth supervision for 3d human pose estimation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7307-7316, 2018. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 658, + 287, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 691 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 691 + ], + "type": "text", + "content": "[39] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 2, 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 691, + 287, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 715 + ], + "type": "text", + "content": "[40] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry," + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 545, + 715 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 327, + 72, + 545, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 72, + 545, + 118 + ], + "spans": [ + { + "bbox": [ + 327, + 72, + 545, + 118 + ], + "type": "text", + "content": "Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2, 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "type": "text", + "content": "[41] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2, 4, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 161, + 545, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 161, + 545, + 215 + ], + "spans": [ + { + "bbox": [ + 307, + 161, + 545, + 215 + ], + "type": "text", + "content": "[42] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 215, + 545, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 215, + 545, + 261 + ], + "spans": [ + { + "bbox": [ + 307, + 215, + 545, + 261 + ], + "type": "text", + "content": "[43] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12179-12188, 2021. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 261, + 545, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 261, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 307, + 261, + 545, + 316 + ], + "type": "text", + "content": "[44] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE transactions on pattern analysis and machine intelligence, 2020. 1, 2, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 316, + 545, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 316, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 307, + 316, + 545, + 370 + ], + "type": "text", + "content": "[45] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding up neural radiance fields with thousands of tiny mlps. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14335-14345, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 371, + 545, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 371, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 307, + 371, + 545, + 426 + ], + "type": "text", + "content": "[46] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In International Conference on Computer Vision, 2021. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 426, + 545, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 426, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 307, + 426, + 545, + 469 + ], + "type": "text", + "content": "[47] Daniel Roich, Ron Mokady, Amit H Bermano, and Daniel Cohen-Or. Pivotal tuning for latent-based editing of real images. ACM Transactions on Graphics (TOG), 42(1):1-13, 2022. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 469, + 545, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 469, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 307, + 469, + 545, + 525 + ], + "type": "text", + "content": "[48] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2, 4, 5, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 525, + 545, + 570 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 525, + 545, + 570 + ], + "spans": [ + { + "bbox": [ + 307, + 525, + 545, + 570 + ], + "type": "text", + "content": "[49] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. arXiv preprint arXiv:2208.12242, 2022. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 570, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 570, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 307, + 570, + 545, + 635 + ], + "type": "text", + "content": "[50] Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 2, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 635, + 545, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 545, + 691 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 545, + 691 + ], + "type": "text", + "content": "[51] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2304–2314, 2019. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 691, + 545, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 715 + ], + "type": "text", + "content": "[52] Meng-Li Shih, Shih-Yang Su, Johannes Kopf, and Jia-Bin Huang. 3d photography using context-aware layered depth" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4488" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 105 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 105 + ], + "type": "text", + "content": " inpainting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8028-8038, 2020. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 287, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 287, + 159 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 287, + 159 + ], + "type": "text", + "content": "[53] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256-2265. PMLR, 2015. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 161, + 288, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 161, + 288, + 194 + ], + "spans": [ + { + "bbox": [ + 48, + 161, + 288, + 194 + ], + "type": "text", + "content": "[54] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 3, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 194, + 288, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 194, + 288, + 227 + ], + "spans": [ + { + "bbox": [ + 48, + 194, + 288, + 227 + ], + "type": "text", + "content": "[55] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in Neural Information Processing Systems, 32, 2019. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 228, + 288, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 228, + 288, + 271 + ], + "spans": [ + { + "bbox": [ + 48, + 228, + 288, + 271 + ], + "type": "text", + "content": "[56] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 3, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 272, + 288, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 272, + 288, + 316 + ], + "spans": [ + { + "bbox": [ + 48, + 272, + 288, + 316 + ], + "type": "text", + "content": "[57] Mohammed Suhail, Carlos Esteves, Leonid Signal, and Ameesh Makadia. Light field neural rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8269-8279, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 316, + 288, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 316, + 288, + 348 + ], + "spans": [ + { + "bbox": [ + 48, + 316, + 288, + 348 + ], + "type": "text", + "content": "[58] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Improved direct voxel grid optimization for radiance fields reconstruction. arxiv.cs.GR 2206.05085, 2022. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 350, + 288, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 350, + 288, + 403 + ], + "spans": [ + { + "bbox": [ + 48, + 350, + 288, + 403 + ], + "type": "text", + "content": "[59] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5459-5469, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 404, + 288, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 404, + 288, + 448 + ], + "spans": [ + { + "bbox": [ + 48, + 404, + 288, + 448 + ], + "type": "text", + "content": "[60] Dani Valevski, Matan Kalman, Yossi Matias, and Yaniv Leviathan. Unitune: Text-driven image editing by fine tuning an image generation model on a single image. arXiv preprint arXiv:2210.09477, 2022. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 449, + 288, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 449, + 288, + 513 + ], + "spans": [ + { + "bbox": [ + 48, + 449, + 288, + 513 + ], + "type": "text", + "content": "[61] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T Barron, and Pratul P Srinivasan. Ref-nerf: Structured view-dependent appearance for neural radiance fields. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5481-5490. IEEE, 2022. 2, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 514, + 288, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 288, + 568 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 288, + 568 + ], + "type": "text", + "content": "[62] Chaoyang Wang, José Miguel Buenapasada, Rui Zhu, and Simon Lucey. Learning depth from monocular videos using direct methods. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2022-2030, 2018. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 570, + 288, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 570, + 288, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 570, + 288, + 624 + ], + "type": "text", + "content": "[63] Can Wang, Menglei Chai, Mingming He, Dongdong Chen, and Jing Liao. Clip-nerf: Text-and-image driven manipulation of neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3835–3844, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 625, + 288, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 288, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 288, + 668 + ], + "type": "text", + "content": "[64] Chaoyang Wang, Simon Lucey, Federico Peruzzi, and Oliver Wang. Web stereo video supervision for depth prediction from dynamic scenes. In 2019 International Conference on 3D Vision (3DV), pages 348-357. IEEE, 2019. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 669, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 288, + 713 + ], + "type": "text", + "content": "[65] Lizhen Wang, Xiaochen Zhao, Tao Yu, Songtao Wang, and Yebin Liu. Normalgan: Learning detailed 3d human from a single rgb-d image. In European Conference on Computer Vision, pages 430-446. Springer, 2020. 1" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 631 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 126 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 126 + ], + "type": "text", + "content": "[66] Nanyang Wang, Yinda Zhang, Zhuwen Li, Yanwei Fu, Wei Liu, and Yu-Gang Jiang. Pixel2mesh: Generating 3d mesh models from single rgb images. In Proceedings of the European conference on computer vision (ECCV), pages 52-67, 2018. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 129, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 194 + ], + "type": "text", + "content": "[67] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P Srinivasan, Howard Zhou, Jonathan T Barron, Ricardo Martin-Brualla, Noah Snavely, and Thomas Funkhouser. Ibrnet: Learning multi-view image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4690-4699, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 195, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 195, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 195, + 545, + 239 + ], + "type": "text", + "content": "[68] Jiajun Wu, Yifan Wang, Tianfan Xue, Xingyuan Sun, Bill Freeman, and Josh Tenenbaum. Marrnet: 3d shape reconstruction via 2.5 d sketches. Advances in neural information processing systems, 30, 2017. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 240, + 545, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 240, + 545, + 295 + ], + "spans": [ + { + "bbox": [ + 307, + 240, + 545, + 295 + ], + "type": "text", + "content": "[69] Ke Xian, Chunhua Shen, Zhiguo Cao, Hao Lu, Yang Xiao, Ruibo Li, and Zhenbo Luo. Monocular relative depth perception with web stereo data supervision. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 311-320, 2018. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 296, + 545, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 296, + 545, + 339 + ], + "spans": [ + { + "bbox": [ + 307, + 296, + 545, + 339 + ], + "type": "text", + "content": "[70] Dejia Xu, Yifan Jiang, Peihao Wang, Zhiwen Fan, Humphrey Shi, and Zhangyang Wang. Sinnerf: Training neural radiance fields on complex scenes from a single image. arXiv preprint arXiv:2204.00928, 2022. 1, 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 341, + 545, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 341, + 545, + 385 + ], + "spans": [ + { + "bbox": [ + 307, + 341, + 545, + 385 + ], + "type": "text", + "content": "[71] Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. Mvsnet: Depth inference for unstructured multi-view stereo. In Proceedings of the European conference on computer vision (ECCV), pages 767-783, 2018. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 386, + 545, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 386, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 307, + 386, + 545, + 430 + ], + "type": "text", + "content": "[72] Wei Yin, Xinlong Wang, Chunhua Shen, Yifan Liu, Zhi Tian, Songcen Xu, Changming Sun, and Dou Renyin. Diversedepth: Affine-invariant depth prediction using diverse data. arXiv preprint arXiv:2002.00569, 2020. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 430, + 545, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 430, + 545, + 474 + ], + "spans": [ + { + "bbox": [ + 307, + 430, + 545, + 474 + ], + "type": "text", + "content": "[73] Wei Yin, Jianming Zhang, Oliver Wang, Simon Niklaus, Long Mai, Simon Chen, and Chunhua Shen. Learning to recover 3d scene shape from a single image. In Proc. IEEE Conf. Comp. Vis. Patt. Recogn. (CVPR), 2021. 1, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 475, + 545, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 475, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 307, + 475, + 545, + 529 + ], + "type": "text", + "content": "[74] Wei Yin, Jianming Zhang, Oliver Wang, Simon Niklaus, Long Mai, Simon Chen, and Chunhua Shen. Learning to recover 3d scene shape from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 204-213, 2021. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 531, + 545, + 585 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 531, + 545, + 585 + ], + "spans": [ + { + "bbox": [ + 307, + 531, + 545, + 585 + ], + "type": "text", + "content": "[75] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4578-4587, 2021. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 586, + 545, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 586, + 545, + 631 + ], + "spans": [ + { + "bbox": [ + 307, + 586, + 545, + 631 + ], + "type": "text", + "content": "[76] Jiahui Yu, Zhe Lin, Jimei Yang, Xiaohui Shen, Xin Lu, and Thomas S Huang. Free-form image inpainting with gated convolution. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4471-4480, 2019. 2" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4489" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/NeuralPCI_ Spatio-Temporal Neural Field for 3D Point Cloud Multi-Frame Non-Linear Interpolation/865808e9-fa47-4657-baf0-084c6ecae997_content_list.json b/2023/NeuralPCI_ Spatio-Temporal Neural Field for 3D Point Cloud Multi-Frame Non-Linear Interpolation/865808e9-fa47-4657-baf0-084c6ecae997_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b9c001fbcfc965aa8020bc0726c7923dc9192f21 --- /dev/null +++ b/2023/NeuralPCI_ Spatio-Temporal Neural Field for 3D Point Cloud Multi-Frame Non-Linear Interpolation/865808e9-fa47-4657-baf0-084c6ecae997_content_list.json @@ -0,0 +1,1484 @@ +[ + { + "type": "text", + "text": "NeuralPCI: Spatio-temporal Neural Field for 3D Point Cloud Multi-frame Non-linear Interpolation", + "text_level": 1, + "bbox": [ + 109, + 128, + 859, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zehan Zheng*, Danni Wu*, Ruisi Lu, Fan Lu, Guang Chen†, Changjun Jiang Tongji University", + "bbox": [ + 166, + 203, + 799, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{zhengzehan, woodannie, lrs910, lufan, guangchen, cjjiang}@tongji.edu.cn", + "bbox": [ + 166, + 241, + 800, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 287, + 313, + 304 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, there has been a significant increase in focus on the interpolation task of computer vision. Despite the tremendous advancement of video interpolation, point cloud interpolation remains insufficiently explored. Meanwhile, the existence of numerous nonlinear large motions in real-world scenarios makes the point cloud interpolation task more challenging. In light of these issues, we present NeuralPCI: an end-to-end 4D spatio-temporal Neural field for 3D Point Cloud Interpolation, which implicitly integrates multi-frame information to handle nonlinear large motions for both indoor and outdoor scenarios. Furthermore, we construct a new multi-frame point cloud interpolation dataset called NL-Drive for large nonlinear motions in autonomous driving scenes to better demonstrate the superiority of our method. Ultimately, NeuralPCI achieves state-of-the-art performance on both DHB (Dynamic Human Bodies) and NL-Drive datasets. Beyond the interpolation task, our method can be naturally extended to point cloud extrapolation, morphing, and auto-labeling, which indicates its substantial potential in other domains. Codes are available at https://github.com/ispc-lab/NeuralPCI.", + "bbox": [ + 73, + 316, + 473, + 633 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 657, + 209, + 672 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In the field of computer vision, sequential point clouds are frequently utilized in many applications, such as VR/AR techniques [11, 38, 49] and autonomous driving [4, 32, 46]. The relatively low frequency of LiDAR compared to other sensors, i.e., $10 - 20\\mathrm{Hz}$ , impedes exploration for high temporal resolution point clouds [47]. Therefore, interpolation tasks for point cloud sequences, which have not been substantially investigated, are receiving increasing attention.", + "bbox": [ + 75, + 681, + 468, + 803 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With the similar goal of obtaining a smooth sequence with high temporal resolution, we can draw inspiration from the video frame interpolation (VFI) task. Several VFI methods [6,8,20,35,45,48] concentrate on nonlinear movements in the real world. They take multiple frames as input and", + "bbox": [ + 75, + 804, + 470, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/efc8019b7730120d5f4ea61ebc6cad08679a1c0f3bfacd5a2c7fb07d0cd9dd66.jpg", + "image_caption": [ + "Figure 1. Common cases of nonlinear motions in autonomous driving scenarios. Spatially uniform linear interpolation (3) using the middle two frames of the point cloud differs significantly from the actual situation (●), so it is necessary to take multiple point clouds into consideration for nonlinear interpolation." + ], + "image_footnote": [], + "bbox": [ + 509, + 287, + 879, + 450 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "generate explicit multi-frame fusion results based on flow estimation [6, 8, 20, 45, 48] or transformer [35]. Nonetheless, due to the unique structure of point clouds [31], it is non-trivial to extend VFI methods to the 3D domain.", + "bbox": [ + 496, + 537, + 892, + 597 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Some early works [18, 19] rely on stereo images to generate pseudo-LiDAR point cloud interpolation. For pure point cloud input, previous methods [22,47] take two consecutive frames as input and output the point cloud at a given intermediate moment. However, with limited two input frames, these approaches can only produce linear interpolation results [22], or perform nonlinear compensation by fusing input frames in the feature dimension linearly [47], which is inherently a data-driven approach to learning the dataset-specified distribution of nonlinear motions rather than an actual nonlinear interpolation. Only when the frame rate of the input point cloud sequence is high enough or the object motion is small enough, can the two adjacent point clouds satisfy the linear motion assumption. Nonetheless, there are numerous nonlinear motions in real-world cases. For instance, as illustrated in Fig. 1, the result of linear interpolation between two adjacent point cloud frames has a large deviation from the actual situation. A point cloud sequence rather than just two point cloud frames allows us to view further into the past and future. Neighboring multiple point", + "bbox": [ + 496, + 599, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution. † Corresponding author.", + "bbox": [ + 99, + 886, + 346, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "909", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "clouds contain additional spatial-temporal cues, namely different perspectives, complementary geometry, and implicit high-order motion information. Therefore, it is time to rethink the point cloud interpolation task with an expanded design space, which is an open challenge yet.", + "bbox": [ + 75, + 90, + 468, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Methods that explicitly fuse multiple point cloud frames generally just approximate the motion model over time, which actually simplifies real-world complex motion. The neural field provides a more elegant way to parameterize the continuous point cloud sequence implicitly. Inspired by NeRF [25] whose view synthesis of images is essentially an interpolation, we propose NeuralPCI, a neural field to exploit the spatial and temporal information of multi-frame point clouds. We build a 4D neural spatio-temporal field, which takes sequential 3D point clouds and the independent interpolation time as input, and predicts the in-between or future point cloud at the given time. Moreover, NeuralPCI is optimized on runtime in a self-supervised manner, without relying on costly ground truths, which makes it free from the out-of-the-distribution generalization problem. Our method can be flexibly applied to segmentation auto-labeling and morphing. Besides, we newly construct a challenging multi-frame point cloud interpolation dataset called NL-Drive from public autonomous driving datasets. Finally, we achieve state-of-the-art performance on indoor DHB dataset and outdoor NL-Drive dataset.", + "bbox": [ + 75, + 167, + 472, + 484 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main contributions are summarized as follows:", + "bbox": [ + 96, + 484, + 434, + 500 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel multi-frame point cloud interpolation algorithm to deal with the nonlinear complex motion in real-world indoor and outdoor scenarios.", + "- We introduce a 4D spatio-temporal neural field to integrate motion information implicitly over space and time to generate the in-between point cloud frames at the arbitrary given time.", + "- A flexible unified framework to conduct both the interpolation and extrapolation, facilitating several applications as well." + ], + "bbox": [ + 96, + 508, + 467, + 659 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 678, + 218, + 694 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Video Frame Interpolation. Most VFI approaches are based on optical flow estimation, focusing on the motion cues between two consecutive input frames. These methods warp source frames with the aid of the optical flow to generate the intermediate frame [1, 13, 26, 33]. In order to deal with complex motions, some works exploit nonlinear information by expanding the design domain to multiple consecutive frames [6, 8, 20, 35, 45]. QVI [45] approximates the flow-based velocity and acceleration of the quadratic motion model explicitly. Following QVI, EQVI [20] improves the training strategy. Recently, Dutta et al. [8] utilizes space-time convolution to adaptively switch motion models with discontinuous motions. These works inspire us", + "bbox": [ + 75, + 704, + 472, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "to design a point cloud interpolation network with multiple input frames. However, it is still challenging to extend these methods to the unordered and unstructured point cloud.", + "bbox": [ + 496, + 90, + 890, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Point Cloud Interpolation. Existing point cloud interpolation methods try to find point-to-point correspondences between two point cloud reference frames. An intuitive way is to utilize scene flow, the extension of optical flow in the 3D domain. PointINet [22] warps two input frames with bi-directional flows to the intermediate frame, then samples the two warped results adaptively and fuses them with attentive weights. This approach is performed under the linear motion assumption and relies heavily on the accuracy of the scene flow backbone. IDEA-Net [47] proposes an alternative method to solve the correlation between two input frames by learning a one-to-one alignment matrix and refining linear interpolation results with a trajectory compensation module. However, the one-to-one correspondence assumption limits its application for large-scale outdoor point cloud datasets. Moreover, the higher-order motion information in the temporal domain is overlooked [8] when taking two frames as input. Therefore, there remain challenges to capturing and modeling the complex nonlinear motion in the real world. To address this issue, we propose a novel neural field that takes advantage of the more comprehensive spatio-temporal information of multiple point cloud frames.", + "bbox": [ + 496, + 136, + 892, + 468 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Neural Implicit Representation. Distinct from the common paradigm of learning-based methods, neural fields are fitted to a single degraded sample rather than a large dataset of samples. The neural field can be seen as a parameterization of diverse input types, such as 2D images [5, 14], 3D shapes [24, 27], etc. Since NeRF [25] presents a novel neural radiance field, which encodes a scene with spatial location and view direction as input and outputs volume rendering, several dynamic scene synthesis studies [7, 10, 16, 17, 28, 30, 36, 43] have been proposed to exploit the representation ability of neural radiance fields for dynamic scenes. Based on the linear motion assumption, Li et al. [17] proposes a time-variant continuous neural representation for space-time view synthesis. NSFP [16] presents a neural prior to regularize scene flow implicitly. Following NSFP, Wang et al. [37] introduce a neural trajectory prior to representing the trajectories as a vector field. Coordinate-based networks show great potential for encoding a continuous input domain over arbitrary dimensions at an arbitrary resolution. Our work utilizes and exploits the ability of the coordinate-based network to represent continuous spatio-temporal motions of dynamic nonlinear scenes.", + "bbox": [ + 496, + 469, + 892, + 801 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 500, + 815, + 635, + 832 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this section, we propose a novel end-to-end 4D spatio-temporal neural field for 3D point cloud interpolation named NeuralPCI. Firstly, we state the problem of multi-frame point cloud interpolation. Sec. 3.1 then pro", + "bbox": [ + 496, + 840, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "910", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ea647606e55ef10498368221cf09f4372a4ef7e02363d21c648cac5fbdec5214.jpg", + "image_caption": [ + "Figure 2. Overview of our proposed NeuralPCI. The 4D neural field is constructed by encoding the spatio-temporal coordinate of the multi-frame input point clouds via a coordinate-based multi-layer perceptron network. In the inference stage after self-supervised optimization, NeuralPCI receives a reference point cloud and an arbitrary interpolation frame moment as input to generate the point cloud of the associated spatio-temporal location." + ], + "image_footnote": [], + "bbox": [ + 119, + 88, + 851, + 353 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "vides an overview of NeuralPCI's architecture and design philosophy. Following this, we explain the details of how to construct the 4D spatio-temporal neural field and integrate neighboring multi-frame nonlinear information of point clouds in Secs. 3.2 and 3.3. Finally, in Sec. 3.4, we further elaborate on the self-supervised optimization manner of the whole neural field.", + "bbox": [ + 75, + 429, + 470, + 532 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Problem Formulation. Let $P \\in \\mathbb{R}^{N \\times 3}$ be one frame of a dynamic point cloud sequence, with $N$ sampled points at the time $t \\in \\mathbb{R}$ . Given a low temporal resolution sequence of $M$ frames of point clouds $S = \\{P_0, P_1, \\dots, P_{M-1}\\}$ with their corresponding timestamps $T = \\{t_0, t_1, \\dots, t_{M-1}\\}$ , the goal of NeuralPCI is to predict the intermediate point cloud frame $P_i$ at an arbitrary given time $t_i$ for interpolation. Interpolating $n \\in \\mathbb{Z}^+$ frames of point clouds at equal intervals between every two consecutive frames yields a sequence of point clouds with $n + 1$ times higher temporal resolution.", + "bbox": [ + 75, + 534, + 468, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For the multi-frame point cloud interpolation task, we define the input as every four consecutive point cloud frames in the sequence $S$ and assume that these frames are equally time-spaced for convenience. The output is the temporally uniform interpolation result of $n$ point cloud frames between the middle two frames. Then, sliding the window of multi-frame inputs traverses the entire sequence.", + "bbox": [ + 75, + 686, + 468, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. NeuralPCI Overview", + "text_level": 1, + "bbox": [ + 76, + 801, + 274, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Mildenhall et al. [25] build the neural radiance field by inputting a series of 2D images with different viewpoints, and then generate the unknown image under a new viewpoint using neural volume rendering. Although the huge gap between images and point clouds makes it non-trivial to", + "bbox": [ + 75, + 824, + 468, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "apply the neural field technique to the point cloud interpolation task, we still want to find a way to address the nonlinear interpolation problem of multi-frame point clouds under a similar design philosophy. That is, encoding a sequence of 3D point clouds at different moments to construct a 4D neural spatio-temporal field and then feeding an arbitrary interpolation frame moment into the network to generate the point cloud of the associated spatio-temporal location.", + "bbox": [ + 496, + 430, + 890, + 550 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Consequently, we propose NeuralPCI whose overall structure is depicted in Fig. 2. It can be divided into three main sections for elaboration. In the optimization stage, the 4D neural field is constructed by encoding the spatiotemporal coordinate of the multi-frame input point cloud via a coordinate-based Multi-Layer Perceptron (MLP) network. For each point cloud frame of the input, the interpolation time is set to the corresponding timestamps of four input frames in order to provide the network with the ability to generate the point cloud of the respective spatio-temporal position, and then optimize the neural field through self-supervised losses. In the inference stage, we run forward the neural field with the spatio-temporal coordinate of a reference point cloud and the moment of the interpolation frame as input to obtain the corresponding in-between point cloud.", + "bbox": [ + 496, + 551, + 892, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. 4D Spatio-temporal Neural Field", + "text_level": 1, + "bbox": [ + 500, + 786, + 790, + 803 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Following the statement in [44], a field is a quantity defined for all spatial and/or temporal coordinates, and a neural field is a field that is parameterized fully or in part by a neural network. Here, we use a coordinate-based MLP network to represent the scenes, which takes as input the 3D spatial coordinate $\\mathbf{x} \\in \\mathbb{R}^3$ and the 1D temporal coordi", + "bbox": [ + 496, + 810, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "911", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "nate $t\\in \\mathbb{R}$ , and produces as output the motions $\\Delta \\mathbf{x}\\in \\mathbb{R}^3$ The MLP is parameterized by $\\Theta$ and it can be viewed as a mapping from the coordinate field to the motion field.", + "bbox": [ + 75, + 90, + 468, + 135 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Furthermore, we leverage an independent input, the interpolation frame time $t_{intp} \\in \\mathbb{R}$ , to provide MLP with time cues and regulate the scene motion output at the interpolation moment relative to the original input spatio-temporal coordinates. Then, we conduct the point-wise addition of the interpolation frame motion and the input point cloud to generate the final output of the neural field, which is the 3D point cloud coordinate at the interpolation frame moment $\\mathbf{x}_{intp} \\in \\mathbb{R}^3$ . Formally, the neural field $\\Phi$ is defined as:", + "bbox": [ + 75, + 136, + 470, + 273 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Phi (\\mathbf {z}; \\Theta): \\mathbf {x} _ {i n t p} = \\mathbf {x} + F _ {\\Theta} (\\mathbf {z}) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 289, + 468, + 306 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{z}$ represents the spatial-temporal coordinate input of point clouds, and $F_{\\Theta}$ is defined as:", + "bbox": [ + 76, + 320, + 468, + 351 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF _ {\\Theta}: \\mathbb {R} ^ {5} \\rightarrow \\mathbb {R} ^ {3}, \\quad \\Delta \\mathbf {x} = F _ {\\Theta} (\\mathbf {z}) = F _ {\\Theta} (\\mathbf {x}, t, t _ {i n t p}) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 364, + 468, + 382 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In such a manner, we implicitly construct a 4D, i.e., $(x,y,z,t)$ in physical, spatio-temporal neural field to represent the entire scenarios of sequential point clouds. Eventually, we are able to utilize the continuity of the dedicated neural field to smoothly interpolate the point cloud at an arbitrary in-between moment.", + "bbox": [ + 75, + 397, + 468, + 487 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Multi-frame Nonlinear Integration", + "text_level": 1, + "bbox": [ + 76, + 498, + 380, + 513 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In order to integrate multi-frame information, an intuitive way is utilizing the existing pair-frame point cloud interpolation algorithm for every two frames among the multiframe point cloud input and fuse the predicted results. We adopt it as a baseline model (see Supplementary Material), and it turns out that directly fusing multiple intermediate predictions by random sampling leads to worse results.", + "bbox": [ + 75, + 522, + 468, + 627 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Besides, an alternative way is to explicitly model the nonlinear kinematic equations taking advantage of multiframe point clouds. We follow the nonlinear video interpolation algorithm [45] and extend it to the 3D domain to formulate the high order equation of point clouds. Nonetheless, the explicit modeling approach is also ineffective in real-world scenarios with complex motions (see Supplementary Material). Consequently, we integrate multi-frame point clouds more effectively using spatio-temporal neural fields.", + "bbox": [ + 75, + 628, + 468, + 777 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our proposed NeuralPCI does not impose a restriction on the frame number of the input point clouds, and thus can be naturally expanded to multi-frame point clouds. When the number of multi-frame inputs is 4, for example, the input set containing a point cloud sequence and corresponding timestamps are $S = \\{P_0, P_1, P_2, P_3\\}$ and $T = \\{t_0, t_1, t_2, t_3\\}$ , respectively. As shown in Fig. 2, the neural field receives the point cloud $P_0$ and time $t_0$", + "bbox": [ + 75, + 779, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "as input, and the time step for interpolation $t_{intp}$ is adjusted to one of $\\{t_0, t_1, t_2, t_3\\}$ to yield the predicted point clouds $\\{\\hat{P}_0^{t_0}, \\hat{P}_0^{t_1}, \\hat{P}_0^{t_2}, \\hat{P}_0^{t_3}\\}$ . The loss function (described in Sec. 3.4) is then calculated between the prediction $\\hat{P}_i^{t_j}$ and each of the four input frames of point clouds $P_j$ . The entire point cloud sequence is traversed through the same operation, with the spatio-temporal neural field end-to-end optimized meanwhile.", + "bbox": [ + 496, + 90, + 890, + 210 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Each input frame serves as a constraint to supervise optimization, allowing us to incorporate the information from multiple frames of point clouds more elegantly. Owing to the continuity, smoothness, and excellent fitting ability of MLP, the derivative of the final spatio-temporal neural field function with respect to the interpolated frame time $t_{intp}$ is an implicit higher-order continuous function, which can better handle complex motion scenes and can yield smooth interpolation outputs.", + "bbox": [ + 496, + 212, + 890, + 348 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Self-supervised Optimization", + "text_level": 1, + "bbox": [ + 498, + 354, + 759, + 371 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "NeuralPCI optimizes the weights of the neural field in a self-supervised manner. As illustrated in Fig. 2, the interpolation time is adjusted as each timestamp of inputs to generate predictions for all input point cloud frames, and the neural field back-propagates the gradients to update the samplespecified parameters by minimizing the following distribution loss between the input and predicted point clouds.", + "bbox": [ + 496, + 378, + 890, + 483 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "CD Loss. Chamfer Distance (CD) [9] measures the distribution difference between two point clouds. We adopt CD as the main term in the loss function, which can be expressed as the following equation:", + "bbox": [ + 496, + 483, + 890, + 540 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {C D} = \\frac {1}{N} \\sum_ {\\hat {p} _ {i} \\in \\hat {P}} \\min _ {p _ {i} \\in P} \\| \\hat {p} _ {i} - p _ {i} \\| _ {2} ^ {2} + \\frac {1}{N} \\sum_ {p _ {i} \\in P} \\min _ {\\hat {p} _ {i} \\in \\hat {P}} \\| p _ {i} - \\hat {p} _ {i} \\| _ {2} ^ {2} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 508, + 544, + 890, + 594 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $P$ and $\\hat{P}$ are the input and predicted point cloud frames. $p_i$ and $\\hat{p}_i$ represent the points in the respective point clouds. $\\| .\\| _2$ denotes the $L_{2}$ norm of the spatial coordinate.", + "bbox": [ + 496, + 597, + 890, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "EMD Loss. Earth Mover's Distance (EMD) [34] calculates the corresponding points by solving the optimal transmission matrix of the two point clouds. We minimize the EMD loss to encourage the two point clouds to have the same density distribution, calculated as:", + "bbox": [ + 496, + 643, + 890, + 718 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {E M D} = \\min _ {\\phi : \\hat {P} \\rightarrow P} \\frac {1}{N} \\sum_ {\\hat {p} \\in \\hat {P}} \\| \\hat {p} - \\phi (\\hat {p}) \\| _ {2} ^ {2} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 724, + 890, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\phi : \\hat{P} \\to P$ denotes a bijection from $\\hat{P}$ to $P$ . Due to its high computational complexity, we only use it in the loss function for sparse point clouds.", + "bbox": [ + 496, + 761, + 890, + 806 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Smoothness Loss. To better regulate the estimated motion, we introduce the smoothness loss, which facilitates the interpolation of point clouds for local rigid motions and is utilized in autonomous driving scenarios.", + "bbox": [ + 496, + 806, + 890, + 863 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {S} = \\sum_ {p _ {i} \\in P} \\frac {1}{| N (p _ {i}) |} \\sum_ {p _ {j} \\in N (p _ {i})} \\| \\Delta \\mathbf {x} _ {j} - \\Delta \\mathbf {x} _ {i} \\| _ {2} ^ {2} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 867, + 890, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "912", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $N(p_i)$ stands for the nearest neighbors of the $i^{th}$ point $p_i$ and $|\\cdot|$ denotes the number of points. And $\\Delta \\mathbf{x}_i$ denotes the scene flow estimation from point cloud $P$ to $\\hat{P}$ (defined in Eq. (2)) of the $i^{th}$ point $p_i$ .", + "bbox": [ + 75, + 90, + 468, + 151 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Total Loss Function. We utilize the above three terms of loss and balance them with separately given weights (defined as $\\alpha, \\beta, \\gamma$ in Eq. (6)) to obtain the total loss. We define multiple input frames and their corresponding timestamps as the input set. Then, we traverse the input set by assigning each point cloud of input frames as the reference and generate predictions of all input frames at their corresponding time steps. The overall loss is computed by summing the losses of every pair of point clouds as Eq. (7).", + "bbox": [ + 75, + 152, + 470, + 287 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Psi = \\alpha \\mathcal {L} _ {C D} + \\beta \\mathcal {L} _ {E M D} + \\gamma \\mathcal {L} _ {S} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 300, + 468, + 316 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\sum_ {P _ {i} \\in S} \\sum_ {t _ {j} \\in T} \\Psi \\left(P _ {t _ {j}}, \\hat {P} _ {i} ^ {t _ {j}}\\right) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 329, + 468, + 364 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $S$ and $T$ are the multiple input frames and their corresponding timestamps. $P_{t_j}$ denotes the input frame at time $t_j$ and $\\hat{P}_i^{t_j}$ denotes the neural field prediction when $P_i$ is the reference frame and the interpolation time is $t_j$ .", + "bbox": [ + 76, + 378, + 468, + 443 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 457, + 209, + 474 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 76, + 483, + 266, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We evaluate NeuralPCI in both indoor and outdoor datasets. Indoor Dynamic Human Bodies dataset (DHB) [47] contains point cloud sequences for the non-rigid deformable 3D human motion with sampled 1024 points. We construct a challenging multi-frame interpolation dataset named Nonliner-Drive (NL-Drive) dataset. Based on the principle of hard-sample selection and the diversity of scenarios, NL-Drive dataset contains point cloud sequences with large nonlinear movements from three public large-scale autonomous driving datasets: KITTI [12], Argoverse [3] and Nuscenes [2]. More details of NL-Drive dataset are provided in Supplementary Material.", + "bbox": [ + 75, + 507, + 468, + 688 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines. To demonstrate the performance of NeuralPCI, we compare our method with previous SOTA approaches, namely IDEA-Net [47] and PointINet [22] for the interpolation task and MoNet [21] for the extrapolation task. We reproduced the results on DHB Dataset of IDEA-Net and PointINet using official implementation. Moreover, we train PointINet and MoNet on NL-Drive dataset according to the official training setting (Since the training code of IDEA-Net has not been released yet, we do not report its results here). In addition, we utilize state-of-the-art scene flow estimation methods, i.e., neural-based NSFP [16] and recurrent-based PV-RAFT [39], with linear interpolation to produce corresponding results on both DHB and NL-Drive datasets for comprehensive comparison. Remarkably, we", + "bbox": [ + 75, + 688, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "evaluate all the optimization-based methods directly on the test set without using the training set.", + "bbox": [ + 498, + 90, + 890, + 121 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Metrics. We adopt CD and EMD as quantitative evaluation metrics following [21, 22, 47], which are described in Eqs. (3) and (4), respectively.", + "bbox": [ + 498, + 121, + 890, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 500, + 178, + 718, + 193 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We implement NeuralPCI with PyTorch [29]. We define our NeuralPCI as a coordinate-based 8-layer MLP architecture with 512 units per layer and adopt LeakyReLU as the activation function. The network weights are randomly initialized with the Adam [15] optimizer (the $lr$ is 0.001). For each sample, the maximum optimization step is limited to 1000 iterations. The spatio-temporal coordinates of the point clouds are position-encoded by a sinusoidal function and fed into the MLP, while the interpolation time is inserted into the penultimate hidden layer to control the final output. All experiments were conducted on a single NVIDIA GeForce RTX 3090 GPU. For more implementation details, please refer to Supplementary Material.", + "bbox": [ + 496, + 200, + 890, + 398 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Evaluation of Point Cloud Interpolation", + "text_level": 1, + "bbox": [ + 500, + 407, + 841, + 425 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Results on DHB dataset. The quantitative comparison on DHB dataset is displayed in Table 1, where our NeuralPCI outperforms other baseline approaches by a large margin. In particular, our overall CD is about half that of other baselines, and our overall EMD is $40\\%$ lower than the suboptimal PV-RAFT [39]. As can be seen, IDEA-Net [47] does not perform well in all scenarios, but in contrast, our method achieves the best results in every single scene, especially in the Squat scenario. This demonstrates the flexibility and adaptability of the neural field in various indoor scenarios. Furthermore, Fig. 3 clearly exhibits the outcomes of qualitative experiments. The interpolation results of PointINet [22] and IDEA-Net contain several outlier noise points and lack lots of detailed information in local areas, such as hair, hands, and skirt hems. Instead, our method benefits from the higher-order implicit function of the neural field, which better handles these complex motions.", + "bbox": [ + 496, + 431, + 890, + 688 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Results on NL-Drive dataset. Table 2 shows results on NL-Drive dataset, where NeuralPCI achieves the best results on most frames and overall metrics. For instance, our method reaches comparable EMD error in the intermediate frames and significantly reduces the CD error, eventually outperforming SOTA by $24.5\\%$ and $4\\%$ in the overall results of CD and EMD metrics, respectively. In the qualitative experiments, we show in detail the nonlinear motion in the outdoor autonomous driving scenario as well as the interpolation frame comparison results in Fig. 5. This indicates that NeuralPCI is scalable to large dense point clouds. As noted in the local zoomed-in view, the vehicle edges are clearer and sharper in the results of our method, while they are blurrier and noisier in the results of PointINet [22].", + "bbox": [ + 496, + 688, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "913", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/647ebd2579cebf4efc9c83f05b5719011b048d75091c585bcfea5c8bb4f2a6bb.jpg", + "table_caption": [ + "Table 1. Quantitative comparison $\\left( {\\times {10}^{-3}}\\right)$ with other open-sourced methods on DHB-Dataset [47]. Baseline methods include IDEA-Net [47], PointINet [22] and the results of linear interpolation of scene flow estimated by NSFP [17] and PV-RAFT [39]." + ], + "table_footnote": [], + "table_body": "
MethodsLongdressLootRed&BlackSoldierSquatSwingOverall
CDEMDCDEMDCDEMDCDEMDCDEMDCDEMDCD↓EMD↓
IDEA-Net0.896.010.868.620.9410.341.6330.070.626.681.246.931.0212.03
PointINet0.9810.870.8512.100.8710.680.9712.390.9013.991.4514.810.9612.25
NSFP1.047.450.817.130.978.140.685.251.147.973.0911.391.227.81
PV-RAFT1.036.880.825.990.947.030.915.310.572.811.4210.540.926.14
NeuralPCI0.704.360.614.760.674.790.594.630.030.020.532.220.543.68
", + "bbox": [ + 81, + 127, + 893, + 244 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/9abd6701c14ed17a1850a9eac91ae0fe8fa873fc25567dd27f62e6d6915cc173.jpg", + "table_caption": [ + "Table 2. Quantitative comparison with other open-sourced methods on NL-Drive Dataset. Type indicates whether the interpolation results are based on forward, backward, bidirectional flow, or neural field. Frame-1, Frame-2 and Frame-3 refer to the three interpolation frames located at equal intervals between the two intermediate input frames." + ], + "table_footnote": [], + "table_body": "
MethodsTypeFrame-1Frame-2Frame-3Average
CDEMDCDEMDCDEMDCD ↓EMD ↓
NSFPforward flow0.9495.181.75132.302.55168.911.75132.13
backward flow2.53168.751.74132.190.9595.231.74132.05
PV-RAFTforward flow1.36104.571.92146.871.63169.821.64140.42
backward flow1.58173.181.85145.481.30102.711.58140.46
PointINetbi-directional flow0.9397.481.24110.221.0195.651.06101.12
NeuralPCIneural field0.7289.030.94113.450.7488.610.8097.03
", + "bbox": [ + 83, + 309, + 890, + 441 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1a6defa161db8749d1b52d41238b60791b41cd929e14d8bfa05b042bdce12318.jpg", + "image_caption": [ + "Figure 3. Qualitative results on DHB dataset. Each column represents one interpolation frame result in the point cloud sequence, where our method is more consistent with the ground truth and preserves local details better than previous SOTA methods." + ], + "image_footnote": [], + "bbox": [ + 78, + 460, + 472, + 705 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Additional Evaluation. To evaluate the capability of NeuralPCI under large motions, we conduct additional experiments to observe the robustness of each method with increasing time interval between input frames. As illustrated in Fig. 4, the results on both datasets show that the performance of our method still lies at the optimum under large motion interpolation.", + "bbox": [ + 75, + 794, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5b073cfc5595db6a6f2563170776391586281f2ca8ba7e7af716771f13fbf3b9.jpg", + "image_caption": [ + "Figure 4. Quantitative experiments at different intervals. The CD error increases as the input frame interval grows. Our method is always optimal at different time intervals and has good robustness under long-distance motion." + ], + "image_footnote": [], + "bbox": [ + 503, + 462, + 890, + 584 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Evaluation of Point Cloud Extrapolation", + "text_level": 1, + "bbox": [ + 500, + 665, + 846, + 681 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As a parameterization of continuous point clouds over space and time, NeuralPCI is generalizable to predict the near future frame by adjusting the timestamp, making it more flexible than existing methods. We conduct extrapolation experiments on NL-Drive dataset with the same input settings as the interpolation task, while the outputs are four consecutive point cloud frames in the future. According to recent works about point cloud prediction [21, 23, 40, 41], we adopt the state-of-the-art method MoNet [21] and also the linear extrapolation results based on scene flow from PV-RAFT [39] and NSFP [16] as baseline methods.", + "bbox": [ + 496, + 688, + 890, + 853 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "A quantitative comparison with baseline methods is shown in Table 3. Flow-based methods suffer a sharp growth in error when the time for the future frame increases.", + "bbox": [ + 498, + 854, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "914", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1e44e8c631839509492b3e33965483a7b6a03e71df422c6d1220ce160d701efe.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 127, + 89, + 480, + 181 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/df9422f70d89d7b302c2415196feaa8b075935001bb9c0d01d4266311a49f7a8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 482, + 90, + 839, + 181 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1ae22fe54cf1a26ab0c81d70412c92e7f5a7b54e3aad1d5594b7cc90ff10c380.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 127, + 183, + 480, + 253 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f67088808d9b7e3c34e8c68087f1e65df80a89a80c01d9738fd729e2a1eee587.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 482, + 183, + 839, + 253 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1c91a1c210f22fe128dd1870378fa2d2cf2c1b2b85d512b35e667829d6ff7343.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 127, + 253, + 480, + 325 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7005eb14ca927cf96ee26d9b01dc67f8579b547fc00590c6ce17feef77d70ea3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 482, + 253, + 839, + 325 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/839da1ef7b7fd22a8ef10fa388a7eba5fdebacfa0ffdb8e734632ff4126e71bb.jpg", + "image_caption": [ + "Figure 5. Qualitative results on NL-Drive dataset. We transform four frames of the input point cloud to the same coordinate system, and the overall motion of the point cloud sequence is depicted in the first row. The following three rows compare the interpolation results, demonstrating that our approach is more accurate and robust, whereas interpolation results of PointINet [22] are noisier." + ], + "image_footnote": [], + "bbox": [ + 127, + 327, + 480, + 398 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cacc45ebce82dfec4a6aaea9d3873d73263916e4222f61450ef312ee15090644.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 482, + 327, + 839, + 398 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/548b788f3b7bd30acf34b98f56328c842b49ec38e57aa0cab8bd9f111a976074.jpg", + "table_caption": [ + "Table 3. Quantitative comparison with other baseline methods for point cloud extrapolation on NL-Drive dataset. Here we adopt MoNet [21] and linear extrapolation results based on scene flow from PV-RAFT [39] and NSFP [16]. Frames 1-4 refer to the consecutive extrapolation frames after the last input frame." + ], + "table_footnote": [], + "table_body": "
MethodsFrame 1Frame 2Frame 3Frame 4
CDEMDCDEMDCDEMDCD↓EMD↓
NSFP4.70209.965.45233.336.24254.726.61268.44
PV-RAFT2.05206.933.90248.946.55293.2710.02377.25
MoNet0.6681.900.96108.411.28135.961.37159.20
NeuralPCI0.7884.261.20108.431.61135.421.87156.64
", + "bbox": [ + 81, + 542, + 472, + 650 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our method surpasses MoNet in terms of EMD error in the last two frames and achieves suboptimal results in the other remaining frames. MoNet's RNN module predicts each frame recurrently, while our neural field lacks constraints on the future direction. Despite that, NeuralPCI still achieves comparable results, indicating its flexibility to accomplish both inter-/extra-polation in a unified framework.", + "bbox": [ + 75, + 664, + 468, + 772 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 784, + 230, + 800 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Contributions of key modules in NeuralPCI, namely the multi-frame integration, NN-intp, EMD loss, smoothness loss, and network enhancement are shown in Table 4.", + "bbox": [ + 75, + 808, + 468, + 853 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Multi-frame and NN-intp. We begin with a simple neural field with pair-frame input and optimized using only CD loss. With the multi-frame integration (ID 2&7), NeuralPCI", + "bbox": [ + 75, + 854, + 470, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/0fab5fa2c0987a8f7f704e00c17db8451b04584da7dfb021063ef5c294775152.jpg", + "table_caption": [ + "Table 4. Quantitative results of ablation studies on DHB dataset [47] $(\\times 10^{-3})$ and NL-Drive dataset. Methods A~F denote CD loss, multi-frame integration, NN-intp, EMD loss, smoothness loss and network enhancement, respectively." + ], + "table_footnote": [], + "table_body": "
DatasetsIDMethodsMetrics
ABCDEFCD ↓EMD ↓
DHB1✓0.655.24
2✓✓0.595.08
3✓✓✓0.574.28
4✓✓✓✓0.563.84
5✓✓✓✓✓0.543.68
NL-Drive6✓0.86114.31
7✓✓0.84112.03
8✓✓✓0.81104.71
9✓✓✓✓0.8299.25
10✓✓✓✓✓0.8097.03
", + "bbox": [ + 504, + 529, + 888, + 705 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "gains $9.2\\%$ and $3.1\\%$ reductions in CD and EMD error on DHB dataset, and $2.3\\%$ and $2.0\\%$ on NL-Drive dataset. By adopting the nearest neighbor in the time domain as the reference frame (ID 3&8), the long-term error growth is effectively reduced, and another EMD improvement of $15.7\\%$ and $6.5\\%$ is achieved on the respective datasets.", + "bbox": [ + 496, + 717, + 890, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "EMD and smoothness loss. To prevent the network from overfitting on a single CD metric, we introduce additional loss terms to regulate the output. The extra EMD loss and smoothness loss help NeuralPCI achieve a significant reduction in EMD metric error, namely $10.3\\%$ (ID 4) on DHB dataset and $5.2\\%$ (ID 9) on NL-Drive dataset.", + "bbox": [ + 496, + 809, + 892, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "915", + "bbox": [ + 485, + 944, + 509, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dd2dbfa7b79527dbbda6b6b27f735acdfd5e1a483613edacab5357c2e9429bed.jpg", + "image_caption": [ + "Figure 6. Visual results of NeuralPCI based auto-labeling. We use NeuralPCI to take the labeled keyframe point clouds as input, output the interpolation results, and automatically assign labels to the intermediate frames. The second row shows the results of the auto-labeling, which intuitively achieves high labeling accuracy." + ], + "image_footnote": [], + "bbox": [ + 114, + 90, + 851, + 311 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0f853ba7e4a362fde33e0037e810a6448298bb7f7c49f653db71e2e9d6fc7422.jpg", + "image_caption": [ + "Figure 7. The process of point clouds morphing between the airplane and the chair samples in ModelNet40 [42] using NeuralPCI." + ], + "image_footnote": [], + "bbox": [ + 81, + 373, + 465, + 470 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Network enhancement. The impact of network structure is future investigated. Previously, NeuralPCI utilized an 8-layer 256-hidden-unit MLP with a ReLU activation function and received the direct spatio-temporal coordinate as input. Here, we introduce a sinusoidal function-based position encoding, switch to the LeakyReLU activation function, and increase the width of the MLP. In the end, the network enhancement brings another improvement of $3.6\\%$ and $4.2\\%$ (ID 5) on DHB dataset and $2.4\\%$ and $2.2\\%$ (ID 10) on NL-Drive dataset.", + "bbox": [ + 75, + 521, + 468, + 672 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. Applications", + "text_level": 1, + "bbox": [ + 76, + 681, + 210, + 698 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Auto-labeling. The annotated point clouds with perpoint segmentation GT in Nuscenes [2] are at $2\\mathrm{Hz}$ , only one tenth of the dataset. Even so, the labeling workload was laborious and enormous with a total of 1,400,000,000 points. Here, we explore the potential of NeuralPCI to generate keyframe-based interpolation results and assign point-wise labels to unannotated intermediate frames, which shows remarkable capability as depicted in Fig. 6. We use four keyframes as input to predict intermediate point clouds, so the predicted outputs could inherit the corresponding keyframe labels in order. Then the kNN algorithm is applied between the unlabeled intermediate frame and the labeled predicted frame to annotate each ground-truth point.", + "bbox": [ + 75, + 704, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Point cloud morphing. In addition to the conventional meaning of nonlinear motion under indoor and outdoor scenes, the transformation relationship across point clouds with totally different topological shapes is also needed. This transformation, i.e., point cloud morphing, is of interest for computer graphics simulation and data enhancement. We deploy NeuralPCI to output interpolation point clouds between two different classes of object point clouds under the ModelNet40 [42] dataset to establish the process of point cloud morphing. As illustrated in Fig. 7, our method enables a more natural and smooth transformation between different shapes compared to random sampling.", + "bbox": [ + 496, + 376, + 893, + 559 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 498, + 570, + 619, + 585 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we redefine the input domain of the point cloud interpolation task as multiple consecutive frames instead of the two consecutive frames used in previous works, which increases the receptive field of the time domain. To achieve that, we presented NeuralPCI, a 4D spatio-temporal neural field for 3D point cloud interpolation that is able to implicitly integrate multi-frame information to handle nonlinear large motions. Our approach achieves state-of-the-art results in both indoor and outdoor datasets. Since neuralfield-based methods are optimized at runtime, the application of our method is limited in terms of real-time efficiency. Based on NeuralPCI, further development can be considered in the future by improving its real-time performance and generalization over unknown samples.", + "bbox": [ + 496, + 595, + 892, + 806 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments This work is supported by Shanghai Municipal Science and Technology Major Project (No.2018SHZDZX01), ZJ Lab, and Shanghai Center for Brain Science and Brain-Inspired Technology, and the Shanghai Rising Star Program (No.21QC1400900) and Tongji-Westwell Autonomous Vehicle Joint Lab Project.", + "bbox": [ + 496, + 810, + 893, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "916", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Wenbo Bao, Wei-Sheng Lai, Chao Ma, Xiaoyun Zhang, Zhiyong Gao, and Ming-Hsuan Yang. Depth-aware video frame interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3703-3712, 2019. 2", + "[2] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 5, 8", + "[3] Ming-Fang Chang, John Lambert, Patsorn Sangkloy, Jagjeet Singh, Slawomir Bak, Andrew Hartnett, De Wang, Peter Carr, Simon Lucey, Deva Ramanan, et al. Argoverse: 3d tracking and forecasting with rich maps. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8748-8757, 2019. 5", + "[4] Xieyuanli Chen, Shijie Li, Benedikt Mersch, Louis Wiesmann, Jürgen Gall, Jens Behley, and Cyril Stachniss. Moving object segmentation in 3d lidar data: A learning-based approach exploiting sequential data. IEEE Robotics and Automation Letters, 6(4):6529-6536, 2021. 1", + "[5] Zhiqin Chen. IM-NET: Learning implicit fields for generative shape modeling. PhD thesis, Applied Sciences: School of Computing Science, 2019. 2", + "[6] Zhixiang Chi, Rasoul Mohammadi Nasiri, Zheng Liu, Juwei Lu, Jin Tang, and Konstantinos N Plataniotis. All at once: Temporally adaptive multi-frame interpolation with advanced motion modeling. In European Conference on Computer Vision, pages 107-123. Springer, 2020. 1, 2", + "[7] Yilun Du, Yinan Zhang, Hong-Xing Yu, Joshua B Tenenbaum, and Jiajun Wu. Neural radiance flow for 4d view synthesis and video processing. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 14304-14314. IEEE Computer Society, 2021. 2", + "[8] Saikat Dutta, Arulkumar Subramaniam, and Anurag Mittal. Non-linear motion estimation for video frame interpolation using space-time convolutions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1726-1731, 2022. 1, 2", + "[9] Haoqiang Fan, Hao Su, and Leonidas J Guibas. A point set generation network for 3d object reconstruction from a single image. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 605-613, 2017. 4", + "[10] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5712-5721, 2021. 2", + "[11] Daniel Garrido, Rui Rodrigues, A Augusto Sousa, Joao Jacob, and Daniel Castro Silva. Point cloud interaction and manipulation in virtual reality. In 2021 5th International Conference on Artificial Intelligence and Virtual Reality (AIVR), pages 15–20, 2021. 1", + "[12] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "suite. In 2012 IEEE conference on computer vision and pattern recognition, pages 3354-3361. IEEE, 2012. 5", + "[13] Huaizu Jiang, Deqing Sun, Varun Jampani, Ming-Hsuan Yang, Erik Learned-Miller, and Jan Kautz. Super slomo: High quality estimation of multiple intermediate frames for video interpolation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9000-9008, 2018. 2", + "[14] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. Advances in Neural Information Processing Systems, 34:852-863, 2021. 2", + "[15] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 5", + "[16] Xueqian Li, Jhony Kaesemodel Pontes, and Simon Lucey. Neural scene flow prior. Advances in Neural Information Processing Systems, 34:7838-7851, 2021. 2, 5, 6, 7", + "[17] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6498-6508, 2021. 2, 6", + "[18] Haojie Liu, Kang Liao, Chunyu Lin, Yao Zhao, and Yulan Guo. Pseudo-lidar point cloud interpolation based on 3d motion representation and spatial supervision. IEEE Transactions on Intelligent Transportation Systems, 23(7):6379-6389, 2021. 1", + "[19] Haojie Liu, Kang Liao, Chunyu Lin, Yao Zhao, and Meiqin Liu. Plin: A network for pseudo-lidar point cloud interpolation. Sensors, 20(6):1573, 2020. 1", + "[20] Yihao Liu, Liangbin Xie, Li Siyao, Wenxiu Sun, Yu Qiao, and Chao Dong. Enhanced quadratic video interpolation. In European Conference on Computer Vision, pages 41-56. Springer, 2020. 1, 2", + "[21] Fan Lu, Guang Chen, Zhijun Li, Lijun Zhang, Yinlong Liu, Sanqing Qu, and Alois Knoll. Monet: Motion-based point cloud prediction network. IEEE Transactions on Intelligent Transportation Systems, 2021. 5, 6, 7", + "[22] Fan Lu, Guang Chen, Sanqing Qu, Zhijun Li, Yinlong Liu, and Alois Knoll. Pointinet: Point cloud frame interpolation network. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 2251-2259, 2021. 1, 2, 5, 6, 7", + "[23] Benedikt Mersch, Xieyuanli Chen, Jens Behley, and Cyril Stachniss. Self-supervised point cloud prediction using 3d spatio-temporal convolutional networks. In Conference on Robot Learning, pages 1444-1454. PMLR, 2022. 6", + "[24] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4460-4470, 2019. 2", + "[25] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2, 3" + ], + "bbox": [ + 503, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "917", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Junheum Park, Keunsoo Ko, Chul Lee, and Chang-Su Kim. Bmbc: Bilateral motion estimation with bilateral cost volume for video interpolation. In European Conference on Computer Vision, pages 109-125. Springer, 2020. 2", + "[27] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 165-174, 2019. 2", + "[28] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865-5874, 2021. 2", + "[29] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 5", + "[30] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10318–10327, 2021. 2", + "[31] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 1", + "[32] Charles R Qi, Yin Zhou, Mahyar Najibi, Pei Sun, Khoa Vo, Boyang Deng, and Dragomir Anguelov. Offboard 3d object detection from point cloud sequences. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6134-6144, 2021. 1", + "[33] Fitssum A Reda, Deqing Sun, Aysegul Dundar, Mohammad Shoeybi, Guilin Liu, Kevin J Shih, Andrew Tao, Jan Kautz, and Bryan Catanzaro. Unsupervised video interpolation using cycle consistency. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 892-900, 2019. 2", + "[34] Yossi Rubner, Carlo Tomasi, and Leonidas J Guibas. The earth mover's distance as a metric for image retrieval. International journal of computer vision, 40(2):99-121, 2000. 4", + "[35] Zhihao Shi, Xiangyu Xu, Xiaohong Liu, Jun Chen, and Ming-Hsuan Yang. Video frame interpolation transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17482-17491, 2022. 1, 2", + "[36] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Christoph Lassner, and Christian Theobalt. Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12959-12970, 2021. 2", + "[37] Chaoyang Wang, Xueqian Li, Jhony Kaesemodel Pontes, and Simon Lucey. Neural prior for trajectory estimation." + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6532-6542, 2022. 2", + "[38] Haiyan Wang, Liang Yang, Xuejian Rong, Jinglun Feng, and Yingli Tian. Self-supervised 4d spatio-temporal feature learning via order prediction of sequential point cloud clips. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3762-3771, 2021. 1", + "[39] Yi Wei, Ziyi Wang, Yongming Rao, Jiwen Lu, and Jie Zhou. Pv-raft: point-voxel correlation fields for scene flow estimation of point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6954-6963, 2021. 5, 6, 7", + "[40] Xinshuo Weng, Junyu Nan, Kuan-Hui Lee, Rowan McAllister, Adrien Gaidon, Nicholas Rhinehart, and Kris M Kitani. S2net: Stochastic sequential pointcloud forecasting. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXVII, pages 549-564. Springer, 2022. 6", + "[41] Xinshuo Weng, Jianren Wang, Sergey Levine, Kris Kitani, and Nicholas Rhinehart. Inverting the pose forecasting pipeline with spf2: Sequential pointcloud forecasting for sequential pose forecasting. In Conference on robot learning, pages 11-20. PMLR, 2021. 6", + "[42] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1912-1920, 2015. 8", + "[43] Wenqi Xian, Jia-Bin Huang, Johannes Kopf, and Changil Kim. Space-time neural irradiance fields for free-viewpoint video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9421-9431, 2021. 2", + "[44] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond. In Computer Graphics Forum, volume 41, pages 641-676. Wiley Online Library, 2022. 3", + "[45] Xiangyu Xu, Li Siyao, Wenxiu Sun, Qian Yin, and Ming-Hsuan Yang. Quadratic video interpolation. Advances in Neural Information Processing Systems, 32, 2019. 1, 2, 4", + "[46] Bin Yang, Min Bai, Ming Liang, Wenyuan Zeng, and Raquel Urtasun. Auto4d: Learning to label 4d objects from sequential point clouds. arXiv preprint arXiv:2101.06586, 2021. 1", + "[47] Yiming Zeng, Yue Qian, Qijian Zhang, Junhui Hou, Yixuan Yuan, and Ying He. Idea-net: Dynamic 3d point cloud interpolation via deep embedding alignment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6338-6347, 2022. 1, 2, 5, 6, 7", + "[48] Haoxian Zhang, Ronggang Wang, and Yang Zhao. Multiframe pyramid refinement network for video frame interpolation. IEEE Access, 7:130610-130621, 2019. 1", + "[49] Zihao Zhang, Lei Hu, Xiaoming Deng, and Shihong Xia. Sequential 3d human pose estimation using adaptive point cloud sampling strategy. In *IJCAI*, pages 1330–1337, 2021. 1" + ], + "bbox": [ + 503, + 92, + 890, + 891 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "918", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/NeuralPCI_ Spatio-Temporal Neural Field for 3D Point Cloud Multi-Frame Non-Linear Interpolation/865808e9-fa47-4657-baf0-084c6ecae997_model.json b/2023/NeuralPCI_ Spatio-Temporal Neural Field for 3D Point Cloud Multi-Frame Non-Linear Interpolation/865808e9-fa47-4657-baf0-084c6ecae997_model.json new file mode 100644 index 0000000000000000000000000000000000000000..fab30ceccb232c5c30c9b6dd523a9ed3e2244090 --- /dev/null +++ b/2023/NeuralPCI_ Spatio-Temporal Neural Field for 3D Point Cloud Multi-Frame Non-Linear Interpolation/865808e9-fa47-4657-baf0-084c6ecae997_model.json @@ -0,0 +1,2068 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.111, + 0.13, + 0.861, + 0.177 + ], + "angle": 0, + "content": "NeuralPCI: Spatio-temporal Neural Field for 3D Point Cloud Multi-frame Non-linear Interpolation" + }, + { + "type": "text", + "bbox": [ + 0.168, + 0.204, + 0.8, + 0.24 + ], + "angle": 0, + "content": "Zehan Zheng*, Danni Wu*, Ruisi Lu, Fan Lu, Guang Chen†, Changjun Jiang Tongji University" + }, + { + "type": "text", + "bbox": [ + 0.167, + 0.242, + 0.802, + 0.258 + ], + "angle": 0, + "content": "{zhengzehan, woodannie, lrs910, lufan, guangchen, cjjiang}@tongji.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.289, + 0.314, + 0.305 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.074, + 0.318, + 0.474, + 0.635 + ], + "angle": 0, + "content": "In recent years, there has been a significant increase in focus on the interpolation task of computer vision. Despite the tremendous advancement of video interpolation, point cloud interpolation remains insufficiently explored. Meanwhile, the existence of numerous nonlinear large motions in real-world scenarios makes the point cloud interpolation task more challenging. In light of these issues, we present NeuralPCI: an end-to-end 4D spatio-temporal Neural field for 3D Point Cloud Interpolation, which implicitly integrates multi-frame information to handle nonlinear large motions for both indoor and outdoor scenarios. Furthermore, we construct a new multi-frame point cloud interpolation dataset called NL-Drive for large nonlinear motions in autonomous driving scenes to better demonstrate the superiority of our method. Ultimately, NeuralPCI achieves state-of-the-art performance on both DHB (Dynamic Human Bodies) and NL-Drive datasets. Beyond the interpolation task, our method can be naturally extended to point cloud extrapolation, morphing, and auto-labeling, which indicates its substantial potential in other domains. Codes are available at https://github.com/ispc-lab/NeuralPCI." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.658, + 0.21, + 0.673 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.683, + 0.47, + 0.804 + ], + "angle": 0, + "content": "In the field of computer vision, sequential point clouds are frequently utilized in many applications, such as VR/AR techniques [11, 38, 49] and autonomous driving [4, 32, 46]. The relatively low frequency of LiDAR compared to other sensors, i.e., \\(10 - 20\\mathrm{Hz}\\), impedes exploration for high temporal resolution point clouds [47]. Therefore, interpolation tasks for point cloud sequences, which have not been substantially investigated, are receiving increasing attention." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.805, + 0.471, + 0.88 + ], + "angle": 0, + "content": "With the similar goal of obtaining a smooth sequence with high temporal resolution, we can draw inspiration from the video frame interpolation (VFI) task. Several VFI methods [6,8,20,35,45,48] concentrate on nonlinear movements in the real world. They take multiple frames as input and" + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.289, + 0.88, + 0.452 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.455, + 0.895, + 0.525 + ], + "angle": 0, + "content": "Figure 1. Common cases of nonlinear motions in autonomous driving scenarios. Spatially uniform linear interpolation (3) using the middle two frames of the point cloud differs significantly from the actual situation (●), so it is necessary to take multiple point clouds into consideration for nonlinear interpolation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.538, + 0.893, + 0.598 + ], + "angle": 0, + "content": "generate explicit multi-frame fusion results based on flow estimation [6, 8, 20, 45, 48] or transformer [35]. Nonetheless, due to the unique structure of point clouds [31], it is non-trivial to extend VFI methods to the 3D domain." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.6, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Some early works [18, 19] rely on stereo images to generate pseudo-LiDAR point cloud interpolation. For pure point cloud input, previous methods [22,47] take two consecutive frames as input and output the point cloud at a given intermediate moment. However, with limited two input frames, these approaches can only produce linear interpolation results [22], or perform nonlinear compensation by fusing input frames in the feature dimension linearly [47], which is inherently a data-driven approach to learning the dataset-specified distribution of nonlinear motions rather than an actual nonlinear interpolation. Only when the frame rate of the input point cloud sequence is high enough or the object motion is small enough, can the two adjacent point clouds satisfy the linear motion assumption. Nonetheless, there are numerous nonlinear motions in real-world cases. For instance, as illustrated in Fig. 1, the result of linear interpolation between two adjacent point cloud frames has a large deviation from the actual situation. A point cloud sequence rather than just two point cloud frames allows us to view further into the past and future. Neighboring multiple point" + }, + { + "type": "page_footnote", + "bbox": [ + 0.101, + 0.887, + 0.348, + 0.901 + ], + "angle": 0, + "content": "* Equal contribution. † Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "909" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.168 + ], + "angle": 0, + "content": "clouds contain additional spatial-temporal cues, namely different perspectives, complementary geometry, and implicit high-order motion information. Therefore, it is time to rethink the point cloud interpolation task with an expanded design space, which is an open challenge yet." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.168, + 0.473, + 0.485 + ], + "angle": 0, + "content": "Methods that explicitly fuse multiple point cloud frames generally just approximate the motion model over time, which actually simplifies real-world complex motion. The neural field provides a more elegant way to parameterize the continuous point cloud sequence implicitly. Inspired by NeRF [25] whose view synthesis of images is essentially an interpolation, we propose NeuralPCI, a neural field to exploit the spatial and temporal information of multi-frame point clouds. We build a 4D neural spatio-temporal field, which takes sequential 3D point clouds and the independent interpolation time as input, and predicts the in-between or future point cloud at the given time. Moreover, NeuralPCI is optimized on runtime in a self-supervised manner, without relying on costly ground truths, which makes it free from the out-of-the-distribution generalization problem. Our method can be flexibly applied to segmentation auto-labeling and morphing. Besides, we newly construct a challenging multi-frame point cloud interpolation dataset called NL-Drive from public autonomous driving datasets. Finally, we achieve state-of-the-art performance on indoor DHB dataset and outdoor NL-Drive dataset." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.486, + 0.436, + 0.501 + ], + "angle": 0, + "content": "Our main contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.51, + 0.468, + 0.553 + ], + "angle": 0, + "content": "- We propose a novel multi-frame point cloud interpolation algorithm to deal with the nonlinear complex motion in real-world indoor and outdoor scenarios." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.556, + 0.468, + 0.616 + ], + "angle": 0, + "content": "- We introduce a 4D spatio-temporal neural field to integrate motion information implicitly over space and time to generate the in-between point cloud frames at the arbitrary given time." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.617, + 0.468, + 0.66 + ], + "angle": 0, + "content": "- A flexible unified framework to conduct both the interpolation and extrapolation, facilitating several applications as well." + }, + { + "type": "list", + "bbox": [ + 0.097, + 0.51, + 0.468, + 0.66 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.679, + 0.22, + 0.695 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.473, + 0.901 + ], + "angle": 0, + "content": "Video Frame Interpolation. Most VFI approaches are based on optical flow estimation, focusing on the motion cues between two consecutive input frames. These methods warp source frames with the aid of the optical flow to generate the intermediate frame [1, 13, 26, 33]. In order to deal with complex motions, some works exploit nonlinear information by expanding the design domain to multiple consecutive frames [6, 8, 20, 35, 45]. QVI [45] approximates the flow-based velocity and acceleration of the quadratic motion model explicitly. Following QVI, EQVI [20] improves the training strategy. Recently, Dutta et al. [8] utilizes space-time convolution to adaptively switch motion models with discontinuous motions. These works inspire us" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.137 + ], + "angle": 0, + "content": "to design a point cloud interpolation network with multiple input frames. However, it is still challenging to extend these methods to the unordered and unstructured point cloud." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.137, + 0.893, + 0.469 + ], + "angle": 0, + "content": "Point Cloud Interpolation. Existing point cloud interpolation methods try to find point-to-point correspondences between two point cloud reference frames. An intuitive way is to utilize scene flow, the extension of optical flow in the 3D domain. PointINet [22] warps two input frames with bi-directional flows to the intermediate frame, then samples the two warped results adaptively and fuses them with attentive weights. This approach is performed under the linear motion assumption and relies heavily on the accuracy of the scene flow backbone. IDEA-Net [47] proposes an alternative method to solve the correlation between two input frames by learning a one-to-one alignment matrix and refining linear interpolation results with a trajectory compensation module. However, the one-to-one correspondence assumption limits its application for large-scale outdoor point cloud datasets. Moreover, the higher-order motion information in the temporal domain is overlooked [8] when taking two frames as input. Therefore, there remain challenges to capturing and modeling the complex nonlinear motion in the real world. To address this issue, we propose a novel neural field that takes advantage of the more comprehensive spatio-temporal information of multiple point cloud frames." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.47, + 0.893, + 0.802 + ], + "angle": 0, + "content": "Neural Implicit Representation. Distinct from the common paradigm of learning-based methods, neural fields are fitted to a single degraded sample rather than a large dataset of samples. The neural field can be seen as a parameterization of diverse input types, such as 2D images [5, 14], 3D shapes [24, 27], etc. Since NeRF [25] presents a novel neural radiance field, which encodes a scene with spatial location and view direction as input and outputs volume rendering, several dynamic scene synthesis studies [7, 10, 16, 17, 28, 30, 36, 43] have been proposed to exploit the representation ability of neural radiance fields for dynamic scenes. Based on the linear motion assumption, Li et al. [17] proposes a time-variant continuous neural representation for space-time view synthesis. NSFP [16] presents a neural prior to regularize scene flow implicitly. Following NSFP, Wang et al. [37] introduce a neural trajectory prior to representing the trajectories as a vector field. Coordinate-based networks show great potential for encoding a continuous input domain over arbitrary dimensions at an arbitrary resolution. Our work utilizes and exploits the ability of the coordinate-based network to represent continuous spatio-temporal motions of dynamic nonlinear scenes." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.816, + 0.636, + 0.833 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.841, + 0.894, + 0.903 + ], + "angle": 0, + "content": "In this section, we propose a novel end-to-end 4D spatio-temporal neural field for 3D point cloud interpolation named NeuralPCI. Firstly, we state the problem of multi-frame point cloud interpolation. Sec. 3.1 then pro" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "910" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.12, + 0.089, + 0.852, + 0.354 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.362, + 0.895, + 0.418 + ], + "angle": 0, + "content": "Figure 2. Overview of our proposed NeuralPCI. The 4D neural field is constructed by encoding the spatio-temporal coordinate of the multi-frame input point clouds via a coordinate-based multi-layer perceptron network. In the inference stage after self-supervised optimization, NeuralPCI receives a reference point cloud and an arbitrary interpolation frame moment as input to generate the point cloud of the associated spatio-temporal location." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.43, + 0.471, + 0.534 + ], + "angle": 0, + "content": "vides an overview of NeuralPCI's architecture and design philosophy. Following this, we explain the details of how to construct the 4D spatio-temporal neural field and integrate neighboring multi-frame nonlinear information of point clouds in Secs. 3.2 and 3.3. Finally, in Sec. 3.4, we further elaborate on the self-supervised optimization manner of the whole neural field." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.535, + 0.47, + 0.687 + ], + "angle": 0, + "content": "Problem Formulation. Let \\( P \\in \\mathbb{R}^{N \\times 3} \\) be one frame of a dynamic point cloud sequence, with \\( N \\) sampled points at the time \\( t \\in \\mathbb{R} \\). Given a low temporal resolution sequence of \\( M \\) frames of point clouds \\( S = \\{P_0, P_1, \\dots, P_{M-1}\\} \\) with their corresponding timestamps \\( T = \\{t_0, t_1, \\dots, t_{M-1}\\} \\), the goal of NeuralPCI is to predict the intermediate point cloud frame \\( P_i \\) at an arbitrary given time \\( t_i \\) for interpolation. Interpolating \\( n \\in \\mathbb{Z}^+ \\) frames of point clouds at equal intervals between every two consecutive frames yields a sequence of point clouds with \\( n + 1 \\) times higher temporal resolution." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.687, + 0.47, + 0.793 + ], + "angle": 0, + "content": "For the multi-frame point cloud interpolation task, we define the input as every four consecutive point cloud frames in the sequence \\( S \\) and assume that these frames are equally time-spaced for convenience. The output is the temporally uniform interpolation result of \\( n \\) point cloud frames between the middle two frames. Then, sliding the window of multi-frame inputs traverses the entire sequence." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.802, + 0.276, + 0.817 + ], + "angle": 0, + "content": "3.1. NeuralPCI Overview" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Mildenhall et al. [25] build the neural radiance field by inputting a series of 2D images with different viewpoints, and then generate the unknown image under a new viewpoint using neural volume rendering. Although the huge gap between images and point clouds makes it non-trivial to" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.431, + 0.892, + 0.551 + ], + "angle": 0, + "content": "apply the neural field technique to the point cloud interpolation task, we still want to find a way to address the nonlinear interpolation problem of multi-frame point clouds under a similar design philosophy. That is, encoding a sequence of 3D point clouds at different moments to construct a 4D neural spatio-temporal field and then feeding an arbitrary interpolation frame moment into the network to generate the point cloud of the associated spatio-temporal location." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.552, + 0.893, + 0.779 + ], + "angle": 0, + "content": "Consequently, we propose NeuralPCI whose overall structure is depicted in Fig. 2. It can be divided into three main sections for elaboration. In the optimization stage, the 4D neural field is constructed by encoding the spatiotemporal coordinate of the multi-frame input point cloud via a coordinate-based Multi-Layer Perceptron (MLP) network. For each point cloud frame of the input, the interpolation time is set to the corresponding timestamps of four input frames in order to provide the network with the ability to generate the point cloud of the respective spatio-temporal position, and then optimize the neural field through self-supervised losses. In the inference stage, we run forward the neural field with the spatio-temporal coordinate of a reference point cloud and the moment of the interpolation frame as input to obtain the corresponding in-between point cloud." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.787, + 0.791, + 0.804 + ], + "angle": 0, + "content": "3.2. 4D Spatio-temporal Neural Field" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Following the statement in [44], a field is a quantity defined for all spatial and/or temporal coordinates, and a neural field is a field that is parameterized fully or in part by a neural network. Here, we use a coordinate-based MLP network to represent the scenes, which takes as input the 3D spatial coordinate \\(\\mathbf{x} \\in \\mathbb{R}^3\\) and the 1D temporal coordi" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "911" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.47, + 0.136 + ], + "angle": 0, + "content": "nate \\(t\\in \\mathbb{R}\\) , and produces as output the motions \\(\\Delta \\mathbf{x}\\in \\mathbb{R}^3\\) The MLP is parameterized by \\(\\Theta\\) and it can be viewed as a mapping from the coordinate field to the motion field." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.137, + 0.471, + 0.274 + ], + "angle": 0, + "content": "Furthermore, we leverage an independent input, the interpolation frame time \\( t_{intp} \\in \\mathbb{R} \\), to provide MLP with time cues and regulate the scene motion output at the interpolation moment relative to the original input spatio-temporal coordinates. Then, we conduct the point-wise addition of the interpolation frame motion and the input point cloud to generate the final output of the neural field, which is the 3D point cloud coordinate at the interpolation frame moment \\( \\mathbf{x}_{intp} \\in \\mathbb{R}^3 \\). Formally, the neural field \\( \\Phi \\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.29, + 0.469, + 0.307 + ], + "angle": 0, + "content": "\\[\n\\Phi (\\mathbf {z}; \\Theta): \\mathbf {x} _ {i n t p} = \\mathbf {x} + F _ {\\Theta} (\\mathbf {z}) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.321, + 0.47, + 0.352 + ], + "angle": 0, + "content": "where \\(\\mathbf{z}\\) represents the spatial-temporal coordinate input of point clouds, and \\(F_{\\Theta}\\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.096, + 0.365, + 0.469, + 0.383 + ], + "angle": 0, + "content": "\\[\nF _ {\\Theta}: \\mathbb {R} ^ {5} \\rightarrow \\mathbb {R} ^ {3}, \\quad \\Delta \\mathbf {x} = F _ {\\Theta} (\\mathbf {z}) = F _ {\\Theta} (\\mathbf {x}, t, t _ {i n t p}) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.398, + 0.469, + 0.488 + ], + "angle": 0, + "content": "In such a manner, we implicitly construct a 4D, i.e., \\((x,y,z,t)\\) in physical, spatio-temporal neural field to represent the entire scenarios of sequential point clouds. Eventually, we are able to utilize the continuity of the dedicated neural field to smoothly interpolate the point cloud at an arbitrary in-between moment." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.499, + 0.382, + 0.515 + ], + "angle": 0, + "content": "3.3. Multi-frame Nonlinear Integration" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.523, + 0.469, + 0.628 + ], + "angle": 0, + "content": "In order to integrate multi-frame information, an intuitive way is utilizing the existing pair-frame point cloud interpolation algorithm for every two frames among the multiframe point cloud input and fuse the predicted results. We adopt it as a baseline model (see Supplementary Material), and it turns out that directly fusing multiple intermediate predictions by random sampling leads to worse results." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.469, + 0.779 + ], + "angle": 0, + "content": "Besides, an alternative way is to explicitly model the nonlinear kinematic equations taking advantage of multiframe point clouds. We follow the nonlinear video interpolation algorithm [45] and extend it to the 3D domain to formulate the high order equation of point clouds. Nonetheless, the explicit modeling approach is also ineffective in real-world scenarios with complex motions (see Supplementary Material). Consequently, we integrate multi-frame point clouds more effectively using spatio-temporal neural fields." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Our proposed NeuralPCI does not impose a restriction on the frame number of the input point clouds, and thus can be naturally expanded to multi-frame point clouds. When the number of multi-frame inputs is 4, for example, the input set containing a point cloud sequence and corresponding timestamps are \\( S = \\{P_0, P_1, P_2, P_3\\} \\) and \\( T = \\{t_0, t_1, t_2, t_3\\} \\), respectively. As shown in Fig. 2, the neural field receives the point cloud \\( P_0 \\) and time \\( t_0 \\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.212 + ], + "angle": 0, + "content": "as input, and the time step for interpolation \\( t_{intp} \\) is adjusted to one of \\( \\{t_0, t_1, t_2, t_3\\} \\) to yield the predicted point clouds \\( \\{\\hat{P}_0^{t_0}, \\hat{P}_0^{t_1}, \\hat{P}_0^{t_2}, \\hat{P}_0^{t_3}\\} \\). The loss function (described in Sec. 3.4) is then calculated between the prediction \\( \\hat{P}_i^{t_j} \\) and each of the four input frames of point clouds \\( P_j \\). The entire point cloud sequence is traversed through the same operation, with the spatio-temporal neural field end-to-end optimized meanwhile." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.213, + 0.892, + 0.349 + ], + "angle": 0, + "content": "Each input frame serves as a constraint to supervise optimization, allowing us to incorporate the information from multiple frames of point clouds more elegantly. Owing to the continuity, smoothness, and excellent fitting ability of MLP, the derivative of the final spatio-temporal neural field function with respect to the interpolated frame time \\( t_{intp} \\) is an implicit higher-order continuous function, which can better handle complex motion scenes and can yield smooth interpolation outputs." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.356, + 0.76, + 0.372 + ], + "angle": 0, + "content": "3.4. Self-supervised Optimization" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.379, + 0.892, + 0.484 + ], + "angle": 0, + "content": "NeuralPCI optimizes the weights of the neural field in a self-supervised manner. As illustrated in Fig. 2, the interpolation time is adjusted as each timestamp of inputs to generate predictions for all input point cloud frames, and the neural field back-propagates the gradients to update the samplespecified parameters by minimizing the following distribution loss between the input and predicted point clouds." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.484, + 0.892, + 0.541 + ], + "angle": 0, + "content": "CD Loss. Chamfer Distance (CD) [9] measures the distribution difference between two point clouds. We adopt CD as the main term in the loss function, which can be expressed as the following equation:" + }, + { + "type": "equation", + "bbox": [ + 0.509, + 0.545, + 0.892, + 0.595 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {C D} = \\frac {1}{N} \\sum_ {\\hat {p} _ {i} \\in \\hat {P}} \\min _ {p _ {i} \\in P} \\| \\hat {p} _ {i} - p _ {i} \\| _ {2} ^ {2} + \\frac {1}{N} \\sum_ {p _ {i} \\in P} \\min _ {\\hat {p} _ {i} \\in \\hat {P}} \\| p _ {i} - \\hat {p} _ {i} \\| _ {2} ^ {2} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.598, + 0.891, + 0.644 + ], + "angle": 0, + "content": "where \\(P\\) and \\(\\hat{P}\\) are the input and predicted point cloud frames. \\(p_i\\) and \\(\\hat{p}_i\\) represent the points in the respective point clouds. \\(\\| .\\| _2\\) denotes the \\(L_{2}\\) norm of the spatial coordinate." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.892, + 0.719 + ], + "angle": 0, + "content": "EMD Loss. Earth Mover's Distance (EMD) [34] calculates the corresponding points by solving the optimal transmission matrix of the two point clouds. We minimize the EMD loss to encourage the two point clouds to have the same density distribution, calculated as:" + }, + { + "type": "equation", + "bbox": [ + 0.567, + 0.725, + 0.891, + 0.765 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {E M D} = \\min _ {\\phi : \\hat {P} \\rightarrow P} \\frac {1}{N} \\sum_ {\\hat {p} \\in \\hat {P}} \\| \\hat {p} - \\phi (\\hat {p}) \\| _ {2} ^ {2} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.762, + 0.891, + 0.808 + ], + "angle": 0, + "content": "where \\(\\phi : \\hat{P} \\to P\\) denotes a bijection from \\(\\hat{P}\\) to \\(P\\). Due to its high computational complexity, we only use it in the loss function for sparse point clouds." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.808, + 0.892, + 0.864 + ], + "angle": 0, + "content": "Smoothness Loss. To better regulate the estimated motion, we introduce the smoothness loss, which facilitates the interpolation of point clouds for local rigid motions and is utilized in autonomous driving scenarios." + }, + { + "type": "equation", + "bbox": [ + 0.555, + 0.868, + 0.891, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {S} = \\sum_ {p _ {i} \\in P} \\frac {1}{| N (p _ {i}) |} \\sum_ {p _ {j} \\in N (p _ {i})} \\| \\Delta \\mathbf {x} _ {j} - \\Delta \\mathbf {x} _ {i} \\| _ {2} ^ {2} \\tag {5}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "912" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.47, + 0.152 + ], + "angle": 0, + "content": "where \\(N(p_i)\\) stands for the nearest neighbors of the \\(i^{th}\\) point \\(p_i\\) and \\(|\\cdot|\\) denotes the number of points. And \\(\\Delta \\mathbf{x}_i\\) denotes the scene flow estimation from point cloud \\(P\\) to \\(\\hat{P}\\) (defined in Eq. (2)) of the \\(i^{th}\\) point \\(p_i\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.153, + 0.471, + 0.289 + ], + "angle": 0, + "content": "Total Loss Function. We utilize the above three terms of loss and balance them with separately given weights (defined as \\(\\alpha, \\beta, \\gamma\\) in Eq. (6)) to obtain the total loss. We define multiple input frames and their corresponding timestamps as the input set. Then, we traverse the input set by assigning each point cloud of input frames as the reference and generate predictions of all input frames at their corresponding time steps. The overall loss is computed by summing the losses of every pair of point clouds as Eq. (7)." + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.301, + 0.47, + 0.318 + ], + "angle": 0, + "content": "\\[\n\\Psi = \\alpha \\mathcal {L} _ {C D} + \\beta \\mathcal {L} _ {E M D} + \\gamma \\mathcal {L} _ {S} \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.176, + 0.33, + 0.469, + 0.365 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\sum_ {P _ {i} \\in S} \\sum_ {t _ {j} \\in T} \\Psi \\left(P _ {t _ {j}}, \\hat {P} _ {i} ^ {t _ {j}}\\right) \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.379, + 0.47, + 0.444 + ], + "angle": 0, + "content": "where \\( S \\) and \\( T \\) are the multiple input frames and their corresponding timestamps. \\( P_{t_j} \\) denotes the input frame at time \\( t_j \\) and \\( \\hat{P}_i^{t_j} \\) denotes the neural field prediction when \\( P_i \\) is the reference frame and the interpolation time is \\( t_j \\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.458, + 0.21, + 0.475 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.484, + 0.267, + 0.501 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.508, + 0.47, + 0.689 + ], + "angle": 0, + "content": "Datasets. We evaluate NeuralPCI in both indoor and outdoor datasets. Indoor Dynamic Human Bodies dataset (DHB) [47] contains point cloud sequences for the non-rigid deformable 3D human motion with sampled 1024 points. We construct a challenging multi-frame interpolation dataset named Nonliner-Drive (NL-Drive) dataset. Based on the principle of hard-sample selection and the diversity of scenarios, NL-Drive dataset contains point cloud sequences with large nonlinear movements from three public large-scale autonomous driving datasets: KITTI [12], Argoverse [3] and Nuscenes [2]. More details of NL-Drive dataset are provided in Supplementary Material." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Baselines. To demonstrate the performance of NeuralPCI, we compare our method with previous SOTA approaches, namely IDEA-Net [47] and PointINet [22] for the interpolation task and MoNet [21] for the extrapolation task. We reproduced the results on DHB Dataset of IDEA-Net and PointINet using official implementation. Moreover, we train PointINet and MoNet on NL-Drive dataset according to the official training setting (Since the training code of IDEA-Net has not been released yet, we do not report its results here). In addition, we utilize state-of-the-art scene flow estimation methods, i.e., neural-based NSFP [16] and recurrent-based PV-RAFT [39], with linear interpolation to produce corresponding results on both DHB and NL-Drive datasets for comprehensive comparison. Remarkably, we" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "evaluate all the optimization-based methods directly on the test set without using the training set." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.122, + 0.892, + 0.168 + ], + "angle": 0, + "content": "Metrics. We adopt CD and EMD as quantitative evaluation metrics following [21, 22, 47], which are described in Eqs. (3) and (4), respectively." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.179, + 0.719, + 0.194 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.201, + 0.892, + 0.399 + ], + "angle": 0, + "content": "We implement NeuralPCI with PyTorch [29]. We define our NeuralPCI as a coordinate-based 8-layer MLP architecture with 512 units per layer and adopt LeakyReLU as the activation function. The network weights are randomly initialized with the Adam [15] optimizer (the \\(lr\\) is 0.001). For each sample, the maximum optimization step is limited to 1000 iterations. The spatio-temporal coordinates of the point clouds are position-encoded by a sinusoidal function and fed into the MLP, while the interpolation time is inserted into the penultimate hidden layer to control the final output. All experiments were conducted on a single NVIDIA GeForce RTX 3090 GPU. For more implementation details, please refer to Supplementary Material." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.409, + 0.843, + 0.426 + ], + "angle": 0, + "content": "4.3. Evaluation of Point Cloud Interpolation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.433, + 0.892, + 0.689 + ], + "angle": 0, + "content": "Results on DHB dataset. The quantitative comparison on DHB dataset is displayed in Table 1, where our NeuralPCI outperforms other baseline approaches by a large margin. In particular, our overall CD is about half that of other baselines, and our overall EMD is \\(40\\%\\) lower than the suboptimal PV-RAFT [39]. As can be seen, IDEA-Net [47] does not perform well in all scenarios, but in contrast, our method achieves the best results in every single scene, especially in the Squat scenario. This demonstrates the flexibility and adaptability of the neural field in various indoor scenarios. Furthermore, Fig. 3 clearly exhibits the outcomes of qualitative experiments. The interpolation results of PointINet [22] and IDEA-Net contain several outlier noise points and lack lots of detailed information in local areas, such as hair, hands, and skirt hems. Instead, our method benefits from the higher-order implicit function of the neural field, which better handles these complex motions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Results on NL-Drive dataset. Table 2 shows results on NL-Drive dataset, where NeuralPCI achieves the best results on most frames and overall metrics. For instance, our method reaches comparable EMD error in the intermediate frames and significantly reduces the CD error, eventually outperforming SOTA by \\(24.5\\%\\) and \\(4\\%\\) in the overall results of CD and EMD metrics, respectively. In the qualitative experiments, we show in detail the nonlinear motion in the outdoor autonomous driving scenario as well as the interpolation frame comparison results in Fig. 5. This indicates that NeuralPCI is scalable to large dense point clouds. As noted in the local zoomed-in view, the vehicle edges are clearer and sharper in the results of our method, while they are blurrier and noisier in the results of PointINet [22]." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "913" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.892, + 0.118 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison \\( \\left( {\\times {10}^{-3}}\\right) \\) with other open-sourced methods on DHB-Dataset [47]. Baseline methods include IDEA-Net [47], PointINet [22] and the results of linear interpolation of scene flow estimated by NSFP [17] and PV-RAFT [39]." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.128, + 0.894, + 0.246 + ], + "angle": 0, + "content": "
MethodsLongdressLootRed&BlackSoldierSquatSwingOverall
CDEMDCDEMDCDEMDCDEMDCDEMDCDEMDCD↓EMD↓
IDEA-Net0.896.010.868.620.9410.341.6330.070.626.681.246.931.0212.03
PointINet0.9810.870.8512.100.8710.680.9712.390.9013.991.4514.810.9612.25
NSFP1.047.450.817.130.978.140.685.251.147.973.0911.391.227.81
PV-RAFT1.036.880.825.990.947.030.915.310.572.811.4210.540.926.14
NeuralPCI0.704.360.614.760.674.790.594.630.030.020.532.220.543.68
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.257, + 0.893, + 0.3 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison with other open-sourced methods on NL-Drive Dataset. Type indicates whether the interpolation results are based on forward, backward, bidirectional flow, or neural field. Frame-1, Frame-2 and Frame-3 refer to the three interpolation frames located at equal intervals between the two intermediate input frames." + }, + { + "type": "table", + "bbox": [ + 0.084, + 0.31, + 0.892, + 0.442 + ], + "angle": 0, + "content": "
MethodsTypeFrame-1Frame-2Frame-3Average
CDEMDCDEMDCDEMDCD ↓EMD ↓
NSFPforward flow0.9495.181.75132.302.55168.911.75132.13
backward flow2.53168.751.74132.190.9595.231.74132.05
PV-RAFTforward flow1.36104.571.92146.871.63169.821.64140.42
backward flow1.58173.181.85145.481.30102.711.58140.46
PointINetbi-directional flow0.9397.481.24110.221.0195.651.06101.12
NeuralPCIneural field0.7289.030.94113.450.7488.610.8097.03
" + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.462, + 0.473, + 0.706 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.715, + 0.47, + 0.772 + ], + "angle": 0, + "content": "Figure 3. Qualitative results on DHB dataset. Each column represents one interpolation frame result in the point cloud sequence, where our method is more consistent with the ground truth and preserves local details better than previous SOTA methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Additional Evaluation. To evaluate the capability of NeuralPCI under large motions, we conduct additional experiments to observe the robustness of each method with increasing time interval between input frames. As illustrated in Fig. 4, the results on both datasets show that the performance of our method still lies at the optimum under large motion interpolation." + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.463, + 0.892, + 0.585 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.597, + 0.892, + 0.653 + ], + "angle": 0, + "content": "Figure 4. Quantitative experiments at different intervals. The CD error increases as the input frame interval grows. Our method is always optimal at different time intervals and has good robustness under long-distance motion." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.666, + 0.848, + 0.683 + ], + "angle": 0, + "content": "4.4. Evaluation of Point Cloud Extrapolation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.854 + ], + "angle": 0, + "content": "As a parameterization of continuous point clouds over space and time, NeuralPCI is generalizable to predict the near future frame by adjusting the timestamp, making it more flexible than existing methods. We conduct extrapolation experiments on NL-Drive dataset with the same input settings as the interpolation task, while the outputs are four consecutive point cloud frames in the future. According to recent works about point cloud prediction [21, 23, 40, 41], we adopt the state-of-the-art method MoNet [21] and also the linear extrapolation results based on scene flow from PV-RAFT [39] and NSFP [16] as baseline methods." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "A quantitative comparison with baseline methods is shown in Table 3. Flow-based methods suffer a sharp growth in error when the time for the future frame increases." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "914" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.129, + 0.09, + 0.482, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.483, + 0.091, + 0.841, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.129, + 0.184, + 0.482, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.483, + 0.184, + 0.841, + 0.255 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.129, + 0.255, + 0.482, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.483, + 0.255, + 0.841, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.129, + 0.328, + 0.482, + 0.399 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.483, + 0.328, + 0.841, + 0.399 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.412, + 0.895, + 0.456 + ], + "angle": 0, + "content": "Figure 5. Qualitative results on NL-Drive dataset. We transform four frames of the input point cloud to the same coordinate system, and the overall motion of the point cloud sequence is depicted in the first row. The following three rows compare the interpolation results, demonstrating that our approach is more accurate and robust, whereas interpolation results of PointINet [22] are noisier." + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.464, + 0.473, + 0.534 + ], + "angle": 0, + "content": "Table 3. Quantitative comparison with other baseline methods for point cloud extrapolation on NL-Drive dataset. Here we adopt MoNet [21] and linear extrapolation results based on scene flow from PV-RAFT [39] and NSFP [16]. Frames 1-4 refer to the consecutive extrapolation frames after the last input frame." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.544, + 0.473, + 0.651 + ], + "angle": 0, + "content": "
MethodsFrame 1Frame 2Frame 3Frame 4
CDEMDCDEMDCDEMDCD↓EMD↓
NSFP4.70209.965.45233.336.24254.726.61268.44
PV-RAFT2.05206.933.90248.946.55293.2710.02377.25
MoNet0.6681.900.96108.411.28135.961.37159.20
NeuralPCI0.7884.261.20108.431.61135.421.87156.64
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.665, + 0.47, + 0.773 + ], + "angle": 0, + "content": "Our method surpasses MoNet in terms of EMD error in the last two frames and achieves suboptimal results in the other remaining frames. MoNet's RNN module predicts each frame recurrently, while our neural field lacks constraints on the future direction. Despite that, NeuralPCI still achieves comparable results, indicating its flexibility to accomplish both inter-/extra-polation in a unified framework." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.785, + 0.231, + 0.801 + ], + "angle": 0, + "content": "4.5. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.809, + 0.47, + 0.854 + ], + "angle": 0, + "content": "Contributions of key modules in NeuralPCI, namely the multi-frame integration, NN-intp, EMD loss, smoothness loss, and network enhancement are shown in Table 4." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Multi-frame and NN-intp. We begin with a simple neural field with pair-frame input and optimized using only CD loss. With the multi-frame integration (ID 2&7), NeuralPCI" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.464, + 0.895, + 0.521 + ], + "angle": 0, + "content": "Table 4. Quantitative results of ablation studies on DHB dataset [47] \\((\\times 10^{-3})\\) and NL-Drive dataset. Methods A~F denote CD loss, multi-frame integration, NN-intp, EMD loss, smoothness loss and network enhancement, respectively." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.53, + 0.89, + 0.706 + ], + "angle": 0, + "content": "
DatasetsIDMethodsMetrics
ABCDEFCD ↓EMD ↓
DHB1✓0.655.24
2✓✓0.595.08
3✓✓✓0.574.28
4✓✓✓✓0.563.84
5✓✓✓✓✓0.543.68
NL-Drive6✓0.86114.31
7✓✓0.84112.03
8✓✓✓0.81104.71
9✓✓✓✓0.8299.25
10✓✓✓✓✓0.8097.03
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.718, + 0.892, + 0.809 + ], + "angle": 0, + "content": "gains \\(9.2\\%\\) and \\(3.1\\%\\) reductions in CD and EMD error on DHB dataset, and \\(2.3\\%\\) and \\(2.0\\%\\) on NL-Drive dataset. By adopting the nearest neighbor in the time domain as the reference frame (ID 3&8), the long-term error growth is effectively reduced, and another EMD improvement of \\(15.7\\%\\) and \\(6.5\\%\\) is achieved on the respective datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.893, + 0.9 + ], + "angle": 0, + "content": "EMD and smoothness loss. To prevent the network from overfitting on a single CD metric, we introduce additional loss terms to regulate the output. The extra EMD loss and smoothness loss help NeuralPCI achieve a significant reduction in EMD metric error, namely \\(10.3\\%\\) (ID 4) on DHB dataset and \\(5.2\\%\\) (ID 9) on NL-Drive dataset." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.945, + 0.511, + 0.957 + ], + "angle": 0, + "content": "915" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.116, + 0.091, + 0.852, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.323, + 0.895, + 0.366 + ], + "angle": 0, + "content": "Figure 6. Visual results of NeuralPCI based auto-labeling. We use NeuralPCI to take the labeled keyframe point clouds as input, output the interpolation results, and automatically assign labels to the intermediate frames. The second row shows the results of the auto-labeling, which intuitively achieves high labeling accuracy." + }, + { + "type": "image", + "bbox": [ + 0.083, + 0.374, + 0.466, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.483, + 0.47, + 0.511 + ], + "angle": 0, + "content": "Figure 7. The process of point clouds morphing between the airplane and the chair samples in ModelNet40 [42] using NeuralPCI." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.522, + 0.47, + 0.673 + ], + "angle": 0, + "content": "Network enhancement. The impact of network structure is future investigated. Previously, NeuralPCI utilized an 8-layer 256-hidden-unit MLP with a ReLU activation function and received the direct spatio-temporal coordinate as input. Here, we introduce a sinusoidal function-based position encoding, switch to the LeakyReLU activation function, and increase the width of the MLP. In the end, the network enhancement brings another improvement of \\(3.6\\%\\) and \\(4.2\\%\\) (ID 5) on DHB dataset and \\(2.4\\%\\) and \\(2.2\\%\\) (ID 10) on NL-Drive dataset." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.682, + 0.212, + 0.699 + ], + "angle": 0, + "content": "4.6. Applications" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Auto-labeling. The annotated point clouds with perpoint segmentation GT in Nuscenes [2] are at \\(2\\mathrm{Hz}\\), only one tenth of the dataset. Even so, the labeling workload was laborious and enormous with a total of 1,400,000,000 points. Here, we explore the potential of NeuralPCI to generate keyframe-based interpolation results and assign point-wise labels to unannotated intermediate frames, which shows remarkable capability as depicted in Fig. 6. We use four keyframes as input to predict intermediate point clouds, so the predicted outputs could inherit the corresponding keyframe labels in order. Then the kNN algorithm is applied between the unlabeled intermediate frame and the labeled predicted frame to annotate each ground-truth point." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.377, + 0.895, + 0.56 + ], + "angle": 0, + "content": "Point cloud morphing. In addition to the conventional meaning of nonlinear motion under indoor and outdoor scenes, the transformation relationship across point clouds with totally different topological shapes is also needed. This transformation, i.e., point cloud morphing, is of interest for computer graphics simulation and data enhancement. We deploy NeuralPCI to output interpolation point clouds between two different classes of object point clouds under the ModelNet40 [42] dataset to establish the process of point cloud morphing. As illustrated in Fig. 7, our method enables a more natural and smooth transformation between different shapes compared to random sampling." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.571, + 0.62, + 0.587 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.596, + 0.893, + 0.807 + ], + "angle": 0, + "content": "In this paper, we redefine the input domain of the point cloud interpolation task as multiple consecutive frames instead of the two consecutive frames used in previous works, which increases the receptive field of the time domain. To achieve that, we presented NeuralPCI, a 4D spatio-temporal neural field for 3D point cloud interpolation that is able to implicitly integrate multi-frame information to handle nonlinear large motions. Our approach achieves state-of-the-art results in both indoor and outdoor datasets. Since neuralfield-based methods are optimized at runtime, the application of our method is limited in terms of real-time efficiency. Based on NeuralPCI, further development can be considered in the future by improving its real-time performance and generalization over unknown samples." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Acknowledgments This work is supported by Shanghai Municipal Science and Technology Major Project (No.2018SHZDZX01), ZJ Lab, and Shanghai Center for Brain Science and Brain-Inspired Technology, and the Shanghai Rising Star Program (No.21QC1400900) and Tongji-Westwell Autonomous Vehicle Joint Lab Project." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "916" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.185 + ], + "angle": 0, + "content": "[1] Wenbo Bao, Wei-Sheng Lai, Chao Ma, Xiaoyun Zhang, Zhiyong Gao, and Ming-Hsuan Yang. Depth-aware video frame interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3703-3712, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.187, + 0.472, + 0.272 + ], + "angle": 0, + "content": "[2] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.273, + 0.472, + 0.356 + ], + "angle": 0, + "content": "[3] Ming-Fang Chang, John Lambert, Patsorn Sangkloy, Jagjeet Singh, Slawomir Bak, Andrew Hartnett, De Wang, Peter Carr, Simon Lucey, Deva Ramanan, et al. Argoverse: 3d tracking and forecasting with rich maps. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8748-8757, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.358, + 0.47, + 0.427 + ], + "angle": 0, + "content": "[4] Xieyuanli Chen, Shijie Li, Benedikt Mersch, Louis Wiesmann, Jürgen Gall, Jens Behley, and Cyril Stachniss. Moving object segmentation in 3d lidar data: A learning-based approach exploiting sequential data. IEEE Robotics and Automation Letters, 6(4):6529-6536, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.429, + 0.47, + 0.47 + ], + "angle": 0, + "content": "[5] Zhiqin Chen. IM-NET: Learning implicit fields for generative shape modeling. PhD thesis, Applied Sciences: School of Computing Science, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.472, + 0.47, + 0.543 + ], + "angle": 0, + "content": "[6] Zhixiang Chi, Rasoul Mohammadi Nasiri, Zheng Liu, Juwei Lu, Jin Tang, and Konstantinos N Plataniotis. All at once: Temporally adaptive multi-frame interpolation with advanced motion modeling. In European Conference on Computer Vision, pages 107-123. Springer, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.545, + 0.47, + 0.613 + ], + "angle": 0, + "content": "[7] Yilun Du, Yinan Zhang, Hong-Xing Yu, Joshua B Tenenbaum, and Jiajun Wu. Neural radiance flow for 4d view synthesis and video processing. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 14304-14314. IEEE Computer Society, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.616, + 0.472, + 0.685 + ], + "angle": 0, + "content": "[8] Saikat Dutta, Arulkumar Subramaniam, and Anurag Mittal. Non-linear motion estimation for video frame interpolation using space-time convolutions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1726-1731, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.687, + 0.47, + 0.743 + ], + "angle": 0, + "content": "[9] Haoqiang Fan, Hao Su, and Leonidas J Guibas. A point set generation network for 3d object reconstruction from a single image. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 605-613, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.745, + 0.469, + 0.8 + ], + "angle": 0, + "content": "[10] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5712-5721, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[11] Daniel Garrido, Rui Rodrigues, A Augusto Sousa, Joao Jacob, and Daniel Castro Silva. Point cloud interaction and manipulation in virtual reality. In 2021 5th International Conference on Artificial Intelligence and Virtual Reality (AIVR), pages 15–20, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[12] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "suite. In 2012 IEEE conference on computer vision and pattern recognition, pages 3354-3361. IEEE, 2012. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.121, + 0.892, + 0.203 + ], + "angle": 0, + "content": "[13] Huaizu Jiang, Deqing Sun, Varun Jampani, Ming-Hsuan Yang, Erik Learned-Miller, and Jan Kautz. Super slomo: High quality estimation of multiple intermediate frames for video interpolation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9000-9008, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.205, + 0.892, + 0.259 + ], + "angle": 0, + "content": "[14] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. Advances in Neural Information Processing Systems, 34:852-863, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.261, + 0.892, + 0.3 + ], + "angle": 0, + "content": "[15] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.302, + 0.892, + 0.343 + ], + "angle": 0, + "content": "[16] Xueqian Li, Jhony Kaesemodel Pontes, and Simon Lucey. Neural scene flow prior. Advances in Neural Information Processing Systems, 34:7838-7851, 2021. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.344, + 0.892, + 0.412 + ], + "angle": 0, + "content": "[17] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6498-6508, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.414, + 0.892, + 0.481 + ], + "angle": 0, + "content": "[18] Haojie Liu, Kang Liao, Chunyu Lin, Yao Zhao, and Yulan Guo. Pseudo-lidar point cloud interpolation based on 3d motion representation and spatial supervision. IEEE Transactions on Intelligent Transportation Systems, 23(7):6379-6389, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.484, + 0.892, + 0.524 + ], + "angle": 0, + "content": "[19] Haojie Liu, Kang Liao, Chunyu Lin, Yao Zhao, and Meiqin Liu. Plin: A network for pseudo-lidar point cloud interpolation. Sensors, 20(6):1573, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.526, + 0.892, + 0.58 + ], + "angle": 0, + "content": "[20] Yihao Liu, Liangbin Xie, Li Siyao, Wenxiu Sun, Yu Qiao, and Chao Dong. Enhanced quadratic video interpolation. In European Conference on Computer Vision, pages 41-56. Springer, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.582, + 0.892, + 0.636 + ], + "angle": 0, + "content": "[21] Fan Lu, Guang Chen, Zhijun Li, Lijun Zhang, Yinlong Liu, Sanqing Qu, and Alois Knoll. Monet: Motion-based point cloud prediction network. IEEE Transactions on Intelligent Transportation Systems, 2021. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.637, + 0.892, + 0.704 + ], + "angle": 0, + "content": "[22] Fan Lu, Guang Chen, Sanqing Qu, Zhijun Li, Yinlong Liu, and Alois Knoll. Pointinet: Point cloud frame interpolation network. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 2251-2259, 2021. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.707, + 0.892, + 0.762 + ], + "angle": 0, + "content": "[23] Benedikt Mersch, Xieyuanli Chen, Jens Behley, and Cyril Stachniss. Self-supervised point cloud prediction using 3d spatio-temporal convolutional networks. In Conference on Robot Learning, pages 1444-1454. PMLR, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.763, + 0.892, + 0.831 + ], + "angle": 0, + "content": "[24] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4460-4470, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[25] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.956 + ], + "angle": 0, + "content": "917" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[26] Junheum Park, Keunsoo Ko, Chul Lee, and Chang-Su Kim. Bmbc: Bilateral motion estimation with bilateral cost volume for video interpolation. In European Conference on Computer Vision, pages 109-125. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.47, + 0.218 + ], + "angle": 0, + "content": "[27] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 165-174, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.219, + 0.47, + 0.287 + ], + "angle": 0, + "content": "[28] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865-5874, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.288, + 0.47, + 0.369 + ], + "angle": 0, + "content": "[29] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.371, + 0.47, + 0.44 + ], + "angle": 0, + "content": "[30] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10318–10327, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.441, + 0.47, + 0.509 + ], + "angle": 0, + "content": "[31] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.511, + 0.471, + 0.58 + ], + "angle": 0, + "content": "[32] Charles R Qi, Yin Zhou, Mahyar Najibi, Pei Sun, Khoa Vo, Boyang Deng, and Dragomir Anguelov. Offboard 3d object detection from point cloud sequences. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6134-6144, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.581, + 0.47, + 0.662 + ], + "angle": 0, + "content": "[33] Fitssum A Reda, Deqing Sun, Aysegul Dundar, Mohammad Shoeybi, Guilin Liu, Kevin J Shih, Andrew Tao, Jan Kautz, and Bryan Catanzaro. Unsupervised video interpolation using cycle consistency. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 892-900, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.664, + 0.47, + 0.718 + ], + "angle": 0, + "content": "[34] Yossi Rubner, Carlo Tomasi, and Leonidas J Guibas. The earth mover's distance as a metric for image retrieval. International journal of computer vision, 40(2):99-121, 2000. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.72, + 0.47, + 0.787 + ], + "angle": 0, + "content": "[35] Zhihao Shi, Xiangyu Xu, Xiaohong Liu, Jun Chen, and Ming-Hsuan Yang. Video frame interpolation transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17482-17491, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.47, + 0.872 + ], + "angle": 0, + "content": "[36] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Christoph Lassner, and Christian Theobalt. Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12959-12970, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[37] Chaoyang Wang, Xueqian Li, Jhony Kaesemodel Pontes, and Simon Lucey. Neural prior for trajectory estimation." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6532-6542, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[38] Haiyan Wang, Liang Yang, Xuejian Rong, Jinglun Feng, and Yingli Tian. Self-supervised 4d spatio-temporal feature learning via order prediction of sequential point cloud clips. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3762-3771, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.192, + 0.892, + 0.26 + ], + "angle": 0, + "content": "[39] Yi Wei, Ziyi Wang, Yongming Rao, Jiwen Lu, and Jie Zhou. Pv-raft: point-voxel correlation fields for scene flow estimation of point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6954-6963, 2021. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.262, + 0.892, + 0.345 + ], + "angle": 0, + "content": "[40] Xinshuo Weng, Junyu Nan, Kuan-Hui Lee, Rowan McAllister, Adrien Gaidon, Nicholas Rhinehart, and Kris M Kitani. S2net: Stochastic sequential pointcloud forecasting. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXVII, pages 549-564. Springer, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.347, + 0.892, + 0.415 + ], + "angle": 0, + "content": "[41] Xinshuo Weng, Jianren Wang, Sergey Levine, Kris Kitani, and Nicholas Rhinehart. Inverting the pose forecasting pipeline with spf2: Sequential pointcloud forecasting for sequential pose forecasting. In Conference on robot learning, pages 11-20. PMLR, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.417, + 0.892, + 0.485 + ], + "angle": 0, + "content": "[42] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1912-1920, 2015. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.487, + 0.892, + 0.555 + ], + "angle": 0, + "content": "[43] Wenqi Xian, Jia-Bin Huang, Johannes Kopf, and Changil Kim. Space-time neural irradiance fields for free-viewpoint video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9421-9431, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.557, + 0.892, + 0.626 + ], + "angle": 0, + "content": "[44] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond. In Computer Graphics Forum, volume 41, pages 641-676. Wiley Online Library, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.628, + 0.892, + 0.668 + ], + "angle": 0, + "content": "[45] Xiangyu Xu, Li Siyao, Wenxiu Sun, Qian Yin, and Ming-Hsuan Yang. Quadratic video interpolation. Advances in Neural Information Processing Systems, 32, 2019. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.669, + 0.892, + 0.724 + ], + "angle": 0, + "content": "[46] Bin Yang, Min Bai, Ming Liang, Wenyuan Zeng, and Raquel Urtasun. Auto4d: Learning to label 4d objects from sequential point clouds. arXiv preprint arXiv:2101.06586, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.726, + 0.892, + 0.794 + ], + "angle": 0, + "content": "[47] Yiming Zeng, Yue Qian, Qijian Zhang, Junhui Hou, Yixuan Yuan, and Ying He. Idea-net: Dynamic 3d point cloud interpolation via deep embedding alignment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6338-6347, 2022. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.797, + 0.892, + 0.837 + ], + "angle": 0, + "content": "[48] Haoxian Zhang, Ronggang Wang, and Yang Zhao. Multiframe pyramid refinement network for video frame interpolation. IEEE Access, 7:130610-130621, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.839, + 0.892, + 0.892 + ], + "angle": 0, + "content": "[49] Zihao Zhang, Lei Hu, Xiaoming Deng, and Shihong Xia. Sequential 3d human pose estimation using adaptive point cloud sampling strategy. In *IJCAI*, pages 1330–1337, 2021. 1" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.892 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "918" + } + ] +] \ No newline at end of file diff --git a/2023/NeuralPCI_ Spatio-Temporal Neural Field for 3D Point Cloud Multi-Frame Non-Linear Interpolation/865808e9-fa47-4657-baf0-084c6ecae997_origin.pdf b/2023/NeuralPCI_ Spatio-Temporal Neural Field for 3D Point Cloud Multi-Frame Non-Linear Interpolation/865808e9-fa47-4657-baf0-084c6ecae997_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4170b39f6d59ef0ef1d04b13a36a7ae79880155e --- /dev/null +++ b/2023/NeuralPCI_ Spatio-Temporal Neural Field for 3D Point Cloud Multi-Frame Non-Linear Interpolation/865808e9-fa47-4657-baf0-084c6ecae997_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f645a3d8e138af74c27a875d1f21826c1c1f777a3e214078ceeff2c28c22fe32 +size 3941205 diff --git a/2023/NeuralPCI_ Spatio-Temporal Neural Field for 3D Point Cloud Multi-Frame Non-Linear Interpolation/full.md b/2023/NeuralPCI_ Spatio-Temporal Neural Field for 3D Point Cloud Multi-Frame Non-Linear Interpolation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4382ff21bf5f08d96cf36ee3f295ed7585bd94dd --- /dev/null +++ b/2023/NeuralPCI_ Spatio-Temporal Neural Field for 3D Point Cloud Multi-Frame Non-Linear Interpolation/full.md @@ -0,0 +1,293 @@ +# NeuralPCI: Spatio-temporal Neural Field for 3D Point Cloud Multi-frame Non-linear Interpolation + +Zehan Zheng*, Danni Wu*, Ruisi Lu, Fan Lu, Guang Chen†, Changjun Jiang Tongji University + +{zhengzehan, woodannie, lrs910, lufan, guangchen, cjjiang}@tongji.edu.cn + +# Abstract + +In recent years, there has been a significant increase in focus on the interpolation task of computer vision. Despite the tremendous advancement of video interpolation, point cloud interpolation remains insufficiently explored. Meanwhile, the existence of numerous nonlinear large motions in real-world scenarios makes the point cloud interpolation task more challenging. In light of these issues, we present NeuralPCI: an end-to-end 4D spatio-temporal Neural field for 3D Point Cloud Interpolation, which implicitly integrates multi-frame information to handle nonlinear large motions for both indoor and outdoor scenarios. Furthermore, we construct a new multi-frame point cloud interpolation dataset called NL-Drive for large nonlinear motions in autonomous driving scenes to better demonstrate the superiority of our method. Ultimately, NeuralPCI achieves state-of-the-art performance on both DHB (Dynamic Human Bodies) and NL-Drive datasets. Beyond the interpolation task, our method can be naturally extended to point cloud extrapolation, morphing, and auto-labeling, which indicates its substantial potential in other domains. Codes are available at https://github.com/ispc-lab/NeuralPCI. + +# 1. Introduction + +In the field of computer vision, sequential point clouds are frequently utilized in many applications, such as VR/AR techniques [11, 38, 49] and autonomous driving [4, 32, 46]. The relatively low frequency of LiDAR compared to other sensors, i.e., $10 - 20\mathrm{Hz}$ , impedes exploration for high temporal resolution point clouds [47]. Therefore, interpolation tasks for point cloud sequences, which have not been substantially investigated, are receiving increasing attention. + +With the similar goal of obtaining a smooth sequence with high temporal resolution, we can draw inspiration from the video frame interpolation (VFI) task. Several VFI methods [6,8,20,35,45,48] concentrate on nonlinear movements in the real world. They take multiple frames as input and + +![](images/efc8019b7730120d5f4ea61ebc6cad08679a1c0f3bfacd5a2c7fb07d0cd9dd66.jpg) +Figure 1. Common cases of nonlinear motions in autonomous driving scenarios. Spatially uniform linear interpolation (3) using the middle two frames of the point cloud differs significantly from the actual situation (●), so it is necessary to take multiple point clouds into consideration for nonlinear interpolation. + +generate explicit multi-frame fusion results based on flow estimation [6, 8, 20, 45, 48] or transformer [35]. Nonetheless, due to the unique structure of point clouds [31], it is non-trivial to extend VFI methods to the 3D domain. + +Some early works [18, 19] rely on stereo images to generate pseudo-LiDAR point cloud interpolation. For pure point cloud input, previous methods [22,47] take two consecutive frames as input and output the point cloud at a given intermediate moment. However, with limited two input frames, these approaches can only produce linear interpolation results [22], or perform nonlinear compensation by fusing input frames in the feature dimension linearly [47], which is inherently a data-driven approach to learning the dataset-specified distribution of nonlinear motions rather than an actual nonlinear interpolation. Only when the frame rate of the input point cloud sequence is high enough or the object motion is small enough, can the two adjacent point clouds satisfy the linear motion assumption. Nonetheless, there are numerous nonlinear motions in real-world cases. For instance, as illustrated in Fig. 1, the result of linear interpolation between two adjacent point cloud frames has a large deviation from the actual situation. A point cloud sequence rather than just two point cloud frames allows us to view further into the past and future. Neighboring multiple point + +clouds contain additional spatial-temporal cues, namely different perspectives, complementary geometry, and implicit high-order motion information. Therefore, it is time to rethink the point cloud interpolation task with an expanded design space, which is an open challenge yet. + +Methods that explicitly fuse multiple point cloud frames generally just approximate the motion model over time, which actually simplifies real-world complex motion. The neural field provides a more elegant way to parameterize the continuous point cloud sequence implicitly. Inspired by NeRF [25] whose view synthesis of images is essentially an interpolation, we propose NeuralPCI, a neural field to exploit the spatial and temporal information of multi-frame point clouds. We build a 4D neural spatio-temporal field, which takes sequential 3D point clouds and the independent interpolation time as input, and predicts the in-between or future point cloud at the given time. Moreover, NeuralPCI is optimized on runtime in a self-supervised manner, without relying on costly ground truths, which makes it free from the out-of-the-distribution generalization problem. Our method can be flexibly applied to segmentation auto-labeling and morphing. Besides, we newly construct a challenging multi-frame point cloud interpolation dataset called NL-Drive from public autonomous driving datasets. Finally, we achieve state-of-the-art performance on indoor DHB dataset and outdoor NL-Drive dataset. + +Our main contributions are summarized as follows: + +- We propose a novel multi-frame point cloud interpolation algorithm to deal with the nonlinear complex motion in real-world indoor and outdoor scenarios. +- We introduce a 4D spatio-temporal neural field to integrate motion information implicitly over space and time to generate the in-between point cloud frames at the arbitrary given time. +- A flexible unified framework to conduct both the interpolation and extrapolation, facilitating several applications as well. + +# 2. Related Work + +Video Frame Interpolation. Most VFI approaches are based on optical flow estimation, focusing on the motion cues between two consecutive input frames. These methods warp source frames with the aid of the optical flow to generate the intermediate frame [1, 13, 26, 33]. In order to deal with complex motions, some works exploit nonlinear information by expanding the design domain to multiple consecutive frames [6, 8, 20, 35, 45]. QVI [45] approximates the flow-based velocity and acceleration of the quadratic motion model explicitly. Following QVI, EQVI [20] improves the training strategy. Recently, Dutta et al. [8] utilizes space-time convolution to adaptively switch motion models with discontinuous motions. These works inspire us + +to design a point cloud interpolation network with multiple input frames. However, it is still challenging to extend these methods to the unordered and unstructured point cloud. + +Point Cloud Interpolation. Existing point cloud interpolation methods try to find point-to-point correspondences between two point cloud reference frames. An intuitive way is to utilize scene flow, the extension of optical flow in the 3D domain. PointINet [22] warps two input frames with bi-directional flows to the intermediate frame, then samples the two warped results adaptively and fuses them with attentive weights. This approach is performed under the linear motion assumption and relies heavily on the accuracy of the scene flow backbone. IDEA-Net [47] proposes an alternative method to solve the correlation between two input frames by learning a one-to-one alignment matrix and refining linear interpolation results with a trajectory compensation module. However, the one-to-one correspondence assumption limits its application for large-scale outdoor point cloud datasets. Moreover, the higher-order motion information in the temporal domain is overlooked [8] when taking two frames as input. Therefore, there remain challenges to capturing and modeling the complex nonlinear motion in the real world. To address this issue, we propose a novel neural field that takes advantage of the more comprehensive spatio-temporal information of multiple point cloud frames. + +Neural Implicit Representation. Distinct from the common paradigm of learning-based methods, neural fields are fitted to a single degraded sample rather than a large dataset of samples. The neural field can be seen as a parameterization of diverse input types, such as 2D images [5, 14], 3D shapes [24, 27], etc. Since NeRF [25] presents a novel neural radiance field, which encodes a scene with spatial location and view direction as input and outputs volume rendering, several dynamic scene synthesis studies [7, 10, 16, 17, 28, 30, 36, 43] have been proposed to exploit the representation ability of neural radiance fields for dynamic scenes. Based on the linear motion assumption, Li et al. [17] proposes a time-variant continuous neural representation for space-time view synthesis. NSFP [16] presents a neural prior to regularize scene flow implicitly. Following NSFP, Wang et al. [37] introduce a neural trajectory prior to representing the trajectories as a vector field. Coordinate-based networks show great potential for encoding a continuous input domain over arbitrary dimensions at an arbitrary resolution. Our work utilizes and exploits the ability of the coordinate-based network to represent continuous spatio-temporal motions of dynamic nonlinear scenes. + +# 3. Methodology + +In this section, we propose a novel end-to-end 4D spatio-temporal neural field for 3D point cloud interpolation named NeuralPCI. Firstly, we state the problem of multi-frame point cloud interpolation. Sec. 3.1 then pro + +![](images/ea647606e55ef10498368221cf09f4372a4ef7e02363d21c648cac5fbdec5214.jpg) +Figure 2. Overview of our proposed NeuralPCI. The 4D neural field is constructed by encoding the spatio-temporal coordinate of the multi-frame input point clouds via a coordinate-based multi-layer perceptron network. In the inference stage after self-supervised optimization, NeuralPCI receives a reference point cloud and an arbitrary interpolation frame moment as input to generate the point cloud of the associated spatio-temporal location. + +vides an overview of NeuralPCI's architecture and design philosophy. Following this, we explain the details of how to construct the 4D spatio-temporal neural field and integrate neighboring multi-frame nonlinear information of point clouds in Secs. 3.2 and 3.3. Finally, in Sec. 3.4, we further elaborate on the self-supervised optimization manner of the whole neural field. + +Problem Formulation. Let $P \in \mathbb{R}^{N \times 3}$ be one frame of a dynamic point cloud sequence, with $N$ sampled points at the time $t \in \mathbb{R}$ . Given a low temporal resolution sequence of $M$ frames of point clouds $S = \{P_0, P_1, \dots, P_{M-1}\}$ with their corresponding timestamps $T = \{t_0, t_1, \dots, t_{M-1}\}$ , the goal of NeuralPCI is to predict the intermediate point cloud frame $P_i$ at an arbitrary given time $t_i$ for interpolation. Interpolating $n \in \mathbb{Z}^+$ frames of point clouds at equal intervals between every two consecutive frames yields a sequence of point clouds with $n + 1$ times higher temporal resolution. + +For the multi-frame point cloud interpolation task, we define the input as every four consecutive point cloud frames in the sequence $S$ and assume that these frames are equally time-spaced for convenience. The output is the temporally uniform interpolation result of $n$ point cloud frames between the middle two frames. Then, sliding the window of multi-frame inputs traverses the entire sequence. + +# 3.1. NeuralPCI Overview + +Mildenhall et al. [25] build the neural radiance field by inputting a series of 2D images with different viewpoints, and then generate the unknown image under a new viewpoint using neural volume rendering. Although the huge gap between images and point clouds makes it non-trivial to + +apply the neural field technique to the point cloud interpolation task, we still want to find a way to address the nonlinear interpolation problem of multi-frame point clouds under a similar design philosophy. That is, encoding a sequence of 3D point clouds at different moments to construct a 4D neural spatio-temporal field and then feeding an arbitrary interpolation frame moment into the network to generate the point cloud of the associated spatio-temporal location. + +Consequently, we propose NeuralPCI whose overall structure is depicted in Fig. 2. It can be divided into three main sections for elaboration. In the optimization stage, the 4D neural field is constructed by encoding the spatiotemporal coordinate of the multi-frame input point cloud via a coordinate-based Multi-Layer Perceptron (MLP) network. For each point cloud frame of the input, the interpolation time is set to the corresponding timestamps of four input frames in order to provide the network with the ability to generate the point cloud of the respective spatio-temporal position, and then optimize the neural field through self-supervised losses. In the inference stage, we run forward the neural field with the spatio-temporal coordinate of a reference point cloud and the moment of the interpolation frame as input to obtain the corresponding in-between point cloud. + +# 3.2. 4D Spatio-temporal Neural Field + +Following the statement in [44], a field is a quantity defined for all spatial and/or temporal coordinates, and a neural field is a field that is parameterized fully or in part by a neural network. Here, we use a coordinate-based MLP network to represent the scenes, which takes as input the 3D spatial coordinate $\mathbf{x} \in \mathbb{R}^3$ and the 1D temporal coordi + +nate $t\in \mathbb{R}$ , and produces as output the motions $\Delta \mathbf{x}\in \mathbb{R}^3$ The MLP is parameterized by $\Theta$ and it can be viewed as a mapping from the coordinate field to the motion field. + +Furthermore, we leverage an independent input, the interpolation frame time $t_{intp} \in \mathbb{R}$ , to provide MLP with time cues and regulate the scene motion output at the interpolation moment relative to the original input spatio-temporal coordinates. Then, we conduct the point-wise addition of the interpolation frame motion and the input point cloud to generate the final output of the neural field, which is the 3D point cloud coordinate at the interpolation frame moment $\mathbf{x}_{intp} \in \mathbb{R}^3$ . Formally, the neural field $\Phi$ is defined as: + +$$ +\Phi (\mathbf {z}; \Theta): \mathbf {x} _ {i n t p} = \mathbf {x} + F _ {\Theta} (\mathbf {z}) \tag {1} +$$ + +where $\mathbf{z}$ represents the spatial-temporal coordinate input of point clouds, and $F_{\Theta}$ is defined as: + +$$ +F _ {\Theta}: \mathbb {R} ^ {5} \rightarrow \mathbb {R} ^ {3}, \quad \Delta \mathbf {x} = F _ {\Theta} (\mathbf {z}) = F _ {\Theta} (\mathbf {x}, t, t _ {i n t p}) \tag {2} +$$ + +In such a manner, we implicitly construct a 4D, i.e., $(x,y,z,t)$ in physical, spatio-temporal neural field to represent the entire scenarios of sequential point clouds. Eventually, we are able to utilize the continuity of the dedicated neural field to smoothly interpolate the point cloud at an arbitrary in-between moment. + +# 3.3. Multi-frame Nonlinear Integration + +In order to integrate multi-frame information, an intuitive way is utilizing the existing pair-frame point cloud interpolation algorithm for every two frames among the multiframe point cloud input and fuse the predicted results. We adopt it as a baseline model (see Supplementary Material), and it turns out that directly fusing multiple intermediate predictions by random sampling leads to worse results. + +Besides, an alternative way is to explicitly model the nonlinear kinematic equations taking advantage of multiframe point clouds. We follow the nonlinear video interpolation algorithm [45] and extend it to the 3D domain to formulate the high order equation of point clouds. Nonetheless, the explicit modeling approach is also ineffective in real-world scenarios with complex motions (see Supplementary Material). Consequently, we integrate multi-frame point clouds more effectively using spatio-temporal neural fields. + +Our proposed NeuralPCI does not impose a restriction on the frame number of the input point clouds, and thus can be naturally expanded to multi-frame point clouds. When the number of multi-frame inputs is 4, for example, the input set containing a point cloud sequence and corresponding timestamps are $S = \{P_0, P_1, P_2, P_3\}$ and $T = \{t_0, t_1, t_2, t_3\}$ , respectively. As shown in Fig. 2, the neural field receives the point cloud $P_0$ and time $t_0$ + +as input, and the time step for interpolation $t_{intp}$ is adjusted to one of $\{t_0, t_1, t_2, t_3\}$ to yield the predicted point clouds $\{\hat{P}_0^{t_0}, \hat{P}_0^{t_1}, \hat{P}_0^{t_2}, \hat{P}_0^{t_3}\}$ . The loss function (described in Sec. 3.4) is then calculated between the prediction $\hat{P}_i^{t_j}$ and each of the four input frames of point clouds $P_j$ . The entire point cloud sequence is traversed through the same operation, with the spatio-temporal neural field end-to-end optimized meanwhile. + +Each input frame serves as a constraint to supervise optimization, allowing us to incorporate the information from multiple frames of point clouds more elegantly. Owing to the continuity, smoothness, and excellent fitting ability of MLP, the derivative of the final spatio-temporal neural field function with respect to the interpolated frame time $t_{intp}$ is an implicit higher-order continuous function, which can better handle complex motion scenes and can yield smooth interpolation outputs. + +# 3.4. Self-supervised Optimization + +NeuralPCI optimizes the weights of the neural field in a self-supervised manner. As illustrated in Fig. 2, the interpolation time is adjusted as each timestamp of inputs to generate predictions for all input point cloud frames, and the neural field back-propagates the gradients to update the samplespecified parameters by minimizing the following distribution loss between the input and predicted point clouds. + +CD Loss. Chamfer Distance (CD) [9] measures the distribution difference between two point clouds. We adopt CD as the main term in the loss function, which can be expressed as the following equation: + +$$ +\mathcal {L} _ {C D} = \frac {1}{N} \sum_ {\hat {p} _ {i} \in \hat {P}} \min _ {p _ {i} \in P} \| \hat {p} _ {i} - p _ {i} \| _ {2} ^ {2} + \frac {1}{N} \sum_ {p _ {i} \in P} \min _ {\hat {p} _ {i} \in \hat {P}} \| p _ {i} - \hat {p} _ {i} \| _ {2} ^ {2} \tag {3} +$$ + +where $P$ and $\hat{P}$ are the input and predicted point cloud frames. $p_i$ and $\hat{p}_i$ represent the points in the respective point clouds. $\| .\| _2$ denotes the $L_{2}$ norm of the spatial coordinate. + +EMD Loss. Earth Mover's Distance (EMD) [34] calculates the corresponding points by solving the optimal transmission matrix of the two point clouds. We minimize the EMD loss to encourage the two point clouds to have the same density distribution, calculated as: + +$$ +\mathcal {L} _ {E M D} = \min _ {\phi : \hat {P} \rightarrow P} \frac {1}{N} \sum_ {\hat {p} \in \hat {P}} \| \hat {p} - \phi (\hat {p}) \| _ {2} ^ {2} \tag {4} +$$ + +where $\phi : \hat{P} \to P$ denotes a bijection from $\hat{P}$ to $P$ . Due to its high computational complexity, we only use it in the loss function for sparse point clouds. + +Smoothness Loss. To better regulate the estimated motion, we introduce the smoothness loss, which facilitates the interpolation of point clouds for local rigid motions and is utilized in autonomous driving scenarios. + +$$ +\mathcal {L} _ {S} = \sum_ {p _ {i} \in P} \frac {1}{| N (p _ {i}) |} \sum_ {p _ {j} \in N (p _ {i})} \| \Delta \mathbf {x} _ {j} - \Delta \mathbf {x} _ {i} \| _ {2} ^ {2} \tag {5} +$$ + +where $N(p_i)$ stands for the nearest neighbors of the $i^{th}$ point $p_i$ and $|\cdot|$ denotes the number of points. And $\Delta \mathbf{x}_i$ denotes the scene flow estimation from point cloud $P$ to $\hat{P}$ (defined in Eq. (2)) of the $i^{th}$ point $p_i$ . + +Total Loss Function. We utilize the above three terms of loss and balance them with separately given weights (defined as $\alpha, \beta, \gamma$ in Eq. (6)) to obtain the total loss. We define multiple input frames and their corresponding timestamps as the input set. Then, we traverse the input set by assigning each point cloud of input frames as the reference and generate predictions of all input frames at their corresponding time steps. The overall loss is computed by summing the losses of every pair of point clouds as Eq. (7). + +$$ +\Psi = \alpha \mathcal {L} _ {C D} + \beta \mathcal {L} _ {E M D} + \gamma \mathcal {L} _ {S} \tag {6} +$$ + +$$ +\mathcal {L} = \sum_ {P _ {i} \in S} \sum_ {t _ {j} \in T} \Psi \left(P _ {t _ {j}}, \hat {P} _ {i} ^ {t _ {j}}\right) \tag {7} +$$ + +where $S$ and $T$ are the multiple input frames and their corresponding timestamps. $P_{t_j}$ denotes the input frame at time $t_j$ and $\hat{P}_i^{t_j}$ denotes the neural field prediction when $P_i$ is the reference frame and the interpolation time is $t_j$ . + +# 4. Experiments + +# 4.1. Experimental Setup + +Datasets. We evaluate NeuralPCI in both indoor and outdoor datasets. Indoor Dynamic Human Bodies dataset (DHB) [47] contains point cloud sequences for the non-rigid deformable 3D human motion with sampled 1024 points. We construct a challenging multi-frame interpolation dataset named Nonliner-Drive (NL-Drive) dataset. Based on the principle of hard-sample selection and the diversity of scenarios, NL-Drive dataset contains point cloud sequences with large nonlinear movements from three public large-scale autonomous driving datasets: KITTI [12], Argoverse [3] and Nuscenes [2]. More details of NL-Drive dataset are provided in Supplementary Material. + +Baselines. To demonstrate the performance of NeuralPCI, we compare our method with previous SOTA approaches, namely IDEA-Net [47] and PointINet [22] for the interpolation task and MoNet [21] for the extrapolation task. We reproduced the results on DHB Dataset of IDEA-Net and PointINet using official implementation. Moreover, we train PointINet and MoNet on NL-Drive dataset according to the official training setting (Since the training code of IDEA-Net has not been released yet, we do not report its results here). In addition, we utilize state-of-the-art scene flow estimation methods, i.e., neural-based NSFP [16] and recurrent-based PV-RAFT [39], with linear interpolation to produce corresponding results on both DHB and NL-Drive datasets for comprehensive comparison. Remarkably, we + +evaluate all the optimization-based methods directly on the test set without using the training set. + +Metrics. We adopt CD and EMD as quantitative evaluation metrics following [21, 22, 47], which are described in Eqs. (3) and (4), respectively. + +# 4.2. Implementation Details + +We implement NeuralPCI with PyTorch [29]. We define our NeuralPCI as a coordinate-based 8-layer MLP architecture with 512 units per layer and adopt LeakyReLU as the activation function. The network weights are randomly initialized with the Adam [15] optimizer (the $lr$ is 0.001). For each sample, the maximum optimization step is limited to 1000 iterations. The spatio-temporal coordinates of the point clouds are position-encoded by a sinusoidal function and fed into the MLP, while the interpolation time is inserted into the penultimate hidden layer to control the final output. All experiments were conducted on a single NVIDIA GeForce RTX 3090 GPU. For more implementation details, please refer to Supplementary Material. + +# 4.3. Evaluation of Point Cloud Interpolation + +Results on DHB dataset. The quantitative comparison on DHB dataset is displayed in Table 1, where our NeuralPCI outperforms other baseline approaches by a large margin. In particular, our overall CD is about half that of other baselines, and our overall EMD is $40\%$ lower than the suboptimal PV-RAFT [39]. As can be seen, IDEA-Net [47] does not perform well in all scenarios, but in contrast, our method achieves the best results in every single scene, especially in the Squat scenario. This demonstrates the flexibility and adaptability of the neural field in various indoor scenarios. Furthermore, Fig. 3 clearly exhibits the outcomes of qualitative experiments. The interpolation results of PointINet [22] and IDEA-Net contain several outlier noise points and lack lots of detailed information in local areas, such as hair, hands, and skirt hems. Instead, our method benefits from the higher-order implicit function of the neural field, which better handles these complex motions. + +Results on NL-Drive dataset. Table 2 shows results on NL-Drive dataset, where NeuralPCI achieves the best results on most frames and overall metrics. For instance, our method reaches comparable EMD error in the intermediate frames and significantly reduces the CD error, eventually outperforming SOTA by $24.5\%$ and $4\%$ in the overall results of CD and EMD metrics, respectively. In the qualitative experiments, we show in detail the nonlinear motion in the outdoor autonomous driving scenario as well as the interpolation frame comparison results in Fig. 5. This indicates that NeuralPCI is scalable to large dense point clouds. As noted in the local zoomed-in view, the vehicle edges are clearer and sharper in the results of our method, while they are blurrier and noisier in the results of PointINet [22]. + +Table 1. Quantitative comparison $\left( {\times {10}^{-3}}\right)$ with other open-sourced methods on DHB-Dataset [47]. Baseline methods include IDEA-Net [47], PointINet [22] and the results of linear interpolation of scene flow estimated by NSFP [17] and PV-RAFT [39]. + +
MethodsLongdressLootRed&BlackSoldierSquatSwingOverall
CDEMDCDEMDCDEMDCDEMDCDEMDCDEMDCD↓EMD↓
IDEA-Net0.896.010.868.620.9410.341.6330.070.626.681.246.931.0212.03
PointINet0.9810.870.8512.100.8710.680.9712.390.9013.991.4514.810.9612.25
NSFP1.047.450.817.130.978.140.685.251.147.973.0911.391.227.81
PV-RAFT1.036.880.825.990.947.030.915.310.572.811.4210.540.926.14
NeuralPCI0.704.360.614.760.674.790.594.630.030.020.532.220.543.68
+ +Table 2. Quantitative comparison with other open-sourced methods on NL-Drive Dataset. Type indicates whether the interpolation results are based on forward, backward, bidirectional flow, or neural field. Frame-1, Frame-2 and Frame-3 refer to the three interpolation frames located at equal intervals between the two intermediate input frames. + +
MethodsTypeFrame-1Frame-2Frame-3Average
CDEMDCDEMDCDEMDCD ↓EMD ↓
NSFPforward flow0.9495.181.75132.302.55168.911.75132.13
backward flow2.53168.751.74132.190.9595.231.74132.05
PV-RAFTforward flow1.36104.571.92146.871.63169.821.64140.42
backward flow1.58173.181.85145.481.30102.711.58140.46
PointINetbi-directional flow0.9397.481.24110.221.0195.651.06101.12
NeuralPCIneural field0.7289.030.94113.450.7488.610.8097.03
+ +![](images/1a6defa161db8749d1b52d41238b60791b41cd929e14d8bfa05b042bdce12318.jpg) +Figure 3. Qualitative results on DHB dataset. Each column represents one interpolation frame result in the point cloud sequence, where our method is more consistent with the ground truth and preserves local details better than previous SOTA methods. + +Additional Evaluation. To evaluate the capability of NeuralPCI under large motions, we conduct additional experiments to observe the robustness of each method with increasing time interval between input frames. As illustrated in Fig. 4, the results on both datasets show that the performance of our method still lies at the optimum under large motion interpolation. + +![](images/5b073cfc5595db6a6f2563170776391586281f2ca8ba7e7af716771f13fbf3b9.jpg) +Figure 4. Quantitative experiments at different intervals. The CD error increases as the input frame interval grows. Our method is always optimal at different time intervals and has good robustness under long-distance motion. + +# 4.4. Evaluation of Point Cloud Extrapolation + +As a parameterization of continuous point clouds over space and time, NeuralPCI is generalizable to predict the near future frame by adjusting the timestamp, making it more flexible than existing methods. We conduct extrapolation experiments on NL-Drive dataset with the same input settings as the interpolation task, while the outputs are four consecutive point cloud frames in the future. According to recent works about point cloud prediction [21, 23, 40, 41], we adopt the state-of-the-art method MoNet [21] and also the linear extrapolation results based on scene flow from PV-RAFT [39] and NSFP [16] as baseline methods. + +A quantitative comparison with baseline methods is shown in Table 3. Flow-based methods suffer a sharp growth in error when the time for the future frame increases. + +![](images/1e44e8c631839509492b3e33965483a7b6a03e71df422c6d1220ce160d701efe.jpg) + +![](images/df9422f70d89d7b302c2415196feaa8b075935001bb9c0d01d4266311a49f7a8.jpg) + +![](images/1ae22fe54cf1a26ab0c81d70412c92e7f5a7b54e3aad1d5594b7cc90ff10c380.jpg) + +![](images/f67088808d9b7e3c34e8c68087f1e65df80a89a80c01d9738fd729e2a1eee587.jpg) + +![](images/1c91a1c210f22fe128dd1870378fa2d2cf2c1b2b85d512b35e667829d6ff7343.jpg) + +![](images/7005eb14ca927cf96ee26d9b01dc67f8579b547fc00590c6ce17feef77d70ea3.jpg) + +![](images/839da1ef7b7fd22a8ef10fa388a7eba5fdebacfa0ffdb8e734632ff4126e71bb.jpg) +Figure 5. Qualitative results on NL-Drive dataset. We transform four frames of the input point cloud to the same coordinate system, and the overall motion of the point cloud sequence is depicted in the first row. The following three rows compare the interpolation results, demonstrating that our approach is more accurate and robust, whereas interpolation results of PointINet [22] are noisier. + +![](images/cacc45ebce82dfec4a6aaea9d3873d73263916e4222f61450ef312ee15090644.jpg) + +Table 3. Quantitative comparison with other baseline methods for point cloud extrapolation on NL-Drive dataset. Here we adopt MoNet [21] and linear extrapolation results based on scene flow from PV-RAFT [39] and NSFP [16]. Frames 1-4 refer to the consecutive extrapolation frames after the last input frame. + +
MethodsFrame 1Frame 2Frame 3Frame 4
CDEMDCDEMDCDEMDCD↓EMD↓
NSFP4.70209.965.45233.336.24254.726.61268.44
PV-RAFT2.05206.933.90248.946.55293.2710.02377.25
MoNet0.6681.900.96108.411.28135.961.37159.20
NeuralPCI0.7884.261.20108.431.61135.421.87156.64
+ +Our method surpasses MoNet in terms of EMD error in the last two frames and achieves suboptimal results in the other remaining frames. MoNet's RNN module predicts each frame recurrently, while our neural field lacks constraints on the future direction. Despite that, NeuralPCI still achieves comparable results, indicating its flexibility to accomplish both inter-/extra-polation in a unified framework. + +# 4.5. Ablation Study + +Contributions of key modules in NeuralPCI, namely the multi-frame integration, NN-intp, EMD loss, smoothness loss, and network enhancement are shown in Table 4. + +Multi-frame and NN-intp. We begin with a simple neural field with pair-frame input and optimized using only CD loss. With the multi-frame integration (ID 2&7), NeuralPCI + +Table 4. Quantitative results of ablation studies on DHB dataset [47] $(\times 10^{-3})$ and NL-Drive dataset. Methods A~F denote CD loss, multi-frame integration, NN-intp, EMD loss, smoothness loss and network enhancement, respectively. + +
DatasetsIDMethodsMetrics
ABCDEFCD ↓EMD ↓
DHB1✓0.655.24
2✓✓0.595.08
3✓✓✓0.574.28
4✓✓✓✓0.563.84
5✓✓✓✓✓0.543.68
NL-Drive6✓0.86114.31
7✓✓0.84112.03
8✓✓✓0.81104.71
9✓✓✓✓0.8299.25
10✓✓✓✓✓0.8097.03
+ +gains $9.2\%$ and $3.1\%$ reductions in CD and EMD error on DHB dataset, and $2.3\%$ and $2.0\%$ on NL-Drive dataset. By adopting the nearest neighbor in the time domain as the reference frame (ID 3&8), the long-term error growth is effectively reduced, and another EMD improvement of $15.7\%$ and $6.5\%$ is achieved on the respective datasets. + +EMD and smoothness loss. To prevent the network from overfitting on a single CD metric, we introduce additional loss terms to regulate the output. The extra EMD loss and smoothness loss help NeuralPCI achieve a significant reduction in EMD metric error, namely $10.3\%$ (ID 4) on DHB dataset and $5.2\%$ (ID 9) on NL-Drive dataset. + +![](images/dd2dbfa7b79527dbbda6b6b27f735acdfd5e1a483613edacab5357c2e9429bed.jpg) +Figure 6. Visual results of NeuralPCI based auto-labeling. We use NeuralPCI to take the labeled keyframe point clouds as input, output the interpolation results, and automatically assign labels to the intermediate frames. The second row shows the results of the auto-labeling, which intuitively achieves high labeling accuracy. + +![](images/0f853ba7e4a362fde33e0037e810a6448298bb7f7c49f653db71e2e9d6fc7422.jpg) +Figure 7. The process of point clouds morphing between the airplane and the chair samples in ModelNet40 [42] using NeuralPCI. + +Network enhancement. The impact of network structure is future investigated. Previously, NeuralPCI utilized an 8-layer 256-hidden-unit MLP with a ReLU activation function and received the direct spatio-temporal coordinate as input. Here, we introduce a sinusoidal function-based position encoding, switch to the LeakyReLU activation function, and increase the width of the MLP. In the end, the network enhancement brings another improvement of $3.6\%$ and $4.2\%$ (ID 5) on DHB dataset and $2.4\%$ and $2.2\%$ (ID 10) on NL-Drive dataset. + +# 4.6. Applications + +Auto-labeling. The annotated point clouds with perpoint segmentation GT in Nuscenes [2] are at $2\mathrm{Hz}$ , only one tenth of the dataset. Even so, the labeling workload was laborious and enormous with a total of 1,400,000,000 points. Here, we explore the potential of NeuralPCI to generate keyframe-based interpolation results and assign point-wise labels to unannotated intermediate frames, which shows remarkable capability as depicted in Fig. 6. We use four keyframes as input to predict intermediate point clouds, so the predicted outputs could inherit the corresponding keyframe labels in order. Then the kNN algorithm is applied between the unlabeled intermediate frame and the labeled predicted frame to annotate each ground-truth point. + +Point cloud morphing. In addition to the conventional meaning of nonlinear motion under indoor and outdoor scenes, the transformation relationship across point clouds with totally different topological shapes is also needed. This transformation, i.e., point cloud morphing, is of interest for computer graphics simulation and data enhancement. We deploy NeuralPCI to output interpolation point clouds between two different classes of object point clouds under the ModelNet40 [42] dataset to establish the process of point cloud morphing. As illustrated in Fig. 7, our method enables a more natural and smooth transformation between different shapes compared to random sampling. + +# 5. Conclusion + +In this paper, we redefine the input domain of the point cloud interpolation task as multiple consecutive frames instead of the two consecutive frames used in previous works, which increases the receptive field of the time domain. To achieve that, we presented NeuralPCI, a 4D spatio-temporal neural field for 3D point cloud interpolation that is able to implicitly integrate multi-frame information to handle nonlinear large motions. Our approach achieves state-of-the-art results in both indoor and outdoor datasets. Since neuralfield-based methods are optimized at runtime, the application of our method is limited in terms of real-time efficiency. Based on NeuralPCI, further development can be considered in the future by improving its real-time performance and generalization over unknown samples. + +Acknowledgments This work is supported by Shanghai Municipal Science and Technology Major Project (No.2018SHZDZX01), ZJ Lab, and Shanghai Center for Brain Science and Brain-Inspired Technology, and the Shanghai Rising Star Program (No.21QC1400900) and Tongji-Westwell Autonomous Vehicle Joint Lab Project. + +# References + +[1] Wenbo Bao, Wei-Sheng Lai, Chao Ma, Xiaoyun Zhang, Zhiyong Gao, and Ming-Hsuan Yang. Depth-aware video frame interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3703-3712, 2019. 2 +[2] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 5, 8 +[3] Ming-Fang Chang, John Lambert, Patsorn Sangkloy, Jagjeet Singh, Slawomir Bak, Andrew Hartnett, De Wang, Peter Carr, Simon Lucey, Deva Ramanan, et al. Argoverse: 3d tracking and forecasting with rich maps. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8748-8757, 2019. 5 +[4] Xieyuanli Chen, Shijie Li, Benedikt Mersch, Louis Wiesmann, Jürgen Gall, Jens Behley, and Cyril Stachniss. Moving object segmentation in 3d lidar data: A learning-based approach exploiting sequential data. IEEE Robotics and Automation Letters, 6(4):6529-6536, 2021. 1 +[5] Zhiqin Chen. IM-NET: Learning implicit fields for generative shape modeling. PhD thesis, Applied Sciences: School of Computing Science, 2019. 2 +[6] Zhixiang Chi, Rasoul Mohammadi Nasiri, Zheng Liu, Juwei Lu, Jin Tang, and Konstantinos N Plataniotis. All at once: Temporally adaptive multi-frame interpolation with advanced motion modeling. In European Conference on Computer Vision, pages 107-123. Springer, 2020. 1, 2 +[7] Yilun Du, Yinan Zhang, Hong-Xing Yu, Joshua B Tenenbaum, and Jiajun Wu. Neural radiance flow for 4d view synthesis and video processing. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 14304-14314. IEEE Computer Society, 2021. 2 +[8] Saikat Dutta, Arulkumar Subramaniam, and Anurag Mittal. Non-linear motion estimation for video frame interpolation using space-time convolutions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1726-1731, 2022. 1, 2 +[9] Haoqiang Fan, Hao Su, and Leonidas J Guibas. A point set generation network for 3d object reconstruction from a single image. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 605-613, 2017. 4 +[10] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5712-5721, 2021. 2 +[11] Daniel Garrido, Rui Rodrigues, A Augusto Sousa, Joao Jacob, and Daniel Castro Silva. Point cloud interaction and manipulation in virtual reality. In 2021 5th International Conference on Artificial Intelligence and Virtual Reality (AIVR), pages 15–20, 2021. 1 +[12] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark + +suite. In 2012 IEEE conference on computer vision and pattern recognition, pages 3354-3361. IEEE, 2012. 5 +[13] Huaizu Jiang, Deqing Sun, Varun Jampani, Ming-Hsuan Yang, Erik Learned-Miller, and Jan Kautz. Super slomo: High quality estimation of multiple intermediate frames for video interpolation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9000-9008, 2018. 2 +[14] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. Advances in Neural Information Processing Systems, 34:852-863, 2021. 2 +[15] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 5 +[16] Xueqian Li, Jhony Kaesemodel Pontes, and Simon Lucey. Neural scene flow prior. Advances in Neural Information Processing Systems, 34:7838-7851, 2021. 2, 5, 6, 7 +[17] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6498-6508, 2021. 2, 6 +[18] Haojie Liu, Kang Liao, Chunyu Lin, Yao Zhao, and Yulan Guo. Pseudo-lidar point cloud interpolation based on 3d motion representation and spatial supervision. IEEE Transactions on Intelligent Transportation Systems, 23(7):6379-6389, 2021. 1 +[19] Haojie Liu, Kang Liao, Chunyu Lin, Yao Zhao, and Meiqin Liu. Plin: A network for pseudo-lidar point cloud interpolation. Sensors, 20(6):1573, 2020. 1 +[20] Yihao Liu, Liangbin Xie, Li Siyao, Wenxiu Sun, Yu Qiao, and Chao Dong. Enhanced quadratic video interpolation. In European Conference on Computer Vision, pages 41-56. Springer, 2020. 1, 2 +[21] Fan Lu, Guang Chen, Zhijun Li, Lijun Zhang, Yinlong Liu, Sanqing Qu, and Alois Knoll. Monet: Motion-based point cloud prediction network. IEEE Transactions on Intelligent Transportation Systems, 2021. 5, 6, 7 +[22] Fan Lu, Guang Chen, Sanqing Qu, Zhijun Li, Yinlong Liu, and Alois Knoll. Pointinet: Point cloud frame interpolation network. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 2251-2259, 2021. 1, 2, 5, 6, 7 +[23] Benedikt Mersch, Xieyuanli Chen, Jens Behley, and Cyril Stachniss. Self-supervised point cloud prediction using 3d spatio-temporal convolutional networks. In Conference on Robot Learning, pages 1444-1454. PMLR, 2022. 6 +[24] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4460-4470, 2019. 2 +[25] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2, 3 + +[26] Junheum Park, Keunsoo Ko, Chul Lee, and Chang-Su Kim. Bmbc: Bilateral motion estimation with bilateral cost volume for video interpolation. In European Conference on Computer Vision, pages 109-125. Springer, 2020. 2 +[27] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 165-174, 2019. 2 +[28] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865-5874, 2021. 2 +[29] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 5 +[30] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10318–10327, 2021. 2 +[31] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 1 +[32] Charles R Qi, Yin Zhou, Mahyar Najibi, Pei Sun, Khoa Vo, Boyang Deng, and Dragomir Anguelov. Offboard 3d object detection from point cloud sequences. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6134-6144, 2021. 1 +[33] Fitssum A Reda, Deqing Sun, Aysegul Dundar, Mohammad Shoeybi, Guilin Liu, Kevin J Shih, Andrew Tao, Jan Kautz, and Bryan Catanzaro. Unsupervised video interpolation using cycle consistency. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 892-900, 2019. 2 +[34] Yossi Rubner, Carlo Tomasi, and Leonidas J Guibas. The earth mover's distance as a metric for image retrieval. International journal of computer vision, 40(2):99-121, 2000. 4 +[35] Zhihao Shi, Xiangyu Xu, Xiaohong Liu, Jun Chen, and Ming-Hsuan Yang. Video frame interpolation transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17482-17491, 2022. 1, 2 +[36] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Christoph Lassner, and Christian Theobalt. Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12959-12970, 2021. 2 +[37] Chaoyang Wang, Xueqian Li, Jhony Kaesemodel Pontes, and Simon Lucey. Neural prior for trajectory estimation. + +In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6532-6542, 2022. 2 +[38] Haiyan Wang, Liang Yang, Xuejian Rong, Jinglun Feng, and Yingli Tian. Self-supervised 4d spatio-temporal feature learning via order prediction of sequential point cloud clips. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3762-3771, 2021. 1 +[39] Yi Wei, Ziyi Wang, Yongming Rao, Jiwen Lu, and Jie Zhou. Pv-raft: point-voxel correlation fields for scene flow estimation of point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6954-6963, 2021. 5, 6, 7 +[40] Xinshuo Weng, Junyu Nan, Kuan-Hui Lee, Rowan McAllister, Adrien Gaidon, Nicholas Rhinehart, and Kris M Kitani. S2net: Stochastic sequential pointcloud forecasting. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXVII, pages 549-564. Springer, 2022. 6 +[41] Xinshuo Weng, Jianren Wang, Sergey Levine, Kris Kitani, and Nicholas Rhinehart. Inverting the pose forecasting pipeline with spf2: Sequential pointcloud forecasting for sequential pose forecasting. In Conference on robot learning, pages 11-20. PMLR, 2021. 6 +[42] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1912-1920, 2015. 8 +[43] Wenqi Xian, Jia-Bin Huang, Johannes Kopf, and Changil Kim. Space-time neural irradiance fields for free-viewpoint video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9421-9431, 2021. 2 +[44] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond. In Computer Graphics Forum, volume 41, pages 641-676. Wiley Online Library, 2022. 3 +[45] Xiangyu Xu, Li Siyao, Wenxiu Sun, Qian Yin, and Ming-Hsuan Yang. Quadratic video interpolation. Advances in Neural Information Processing Systems, 32, 2019. 1, 2, 4 +[46] Bin Yang, Min Bai, Ming Liang, Wenyuan Zeng, and Raquel Urtasun. Auto4d: Learning to label 4d objects from sequential point clouds. arXiv preprint arXiv:2101.06586, 2021. 1 +[47] Yiming Zeng, Yue Qian, Qijian Zhang, Junhui Hou, Yixuan Yuan, and Ying He. Idea-net: Dynamic 3d point cloud interpolation via deep embedding alignment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6338-6347, 2022. 1, 2, 5, 6, 7 +[48] Haoxian Zhang, Ronggang Wang, and Yang Zhao. Multiframe pyramid refinement network for video frame interpolation. IEEE Access, 7:130610-130621, 2019. 1 +[49] Zihao Zhang, Lei Hu, Xiaoming Deng, and Shihong Xia. Sequential 3d human pose estimation using adaptive point cloud sampling strategy. In *IJCAI*, pages 1330–1337, 2021. 1 \ No newline at end of file diff --git a/2023/NeuralPCI_ Spatio-Temporal Neural Field for 3D Point Cloud Multi-Frame Non-Linear Interpolation/images.zip b/2023/NeuralPCI_ Spatio-Temporal Neural Field for 3D Point Cloud Multi-Frame Non-Linear Interpolation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..89b3b249e8af46b38754c049ca67d1083db35c13 --- /dev/null +++ b/2023/NeuralPCI_ Spatio-Temporal Neural Field for 3D Point Cloud Multi-Frame Non-Linear Interpolation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cf619310e0e9a55c9d96783616e25fdbd40cb3785d7a11ca5979c45937d5f802 +size 744170 diff --git a/2023/NeuralPCI_ Spatio-Temporal Neural Field for 3D Point Cloud Multi-Frame Non-Linear Interpolation/layout.json b/2023/NeuralPCI_ Spatio-Temporal Neural Field for 3D Point Cloud Multi-Frame Non-Linear Interpolation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3587aac5459485c6ebd4bd814b271a8a08a8b305 --- /dev/null +++ b/2023/NeuralPCI_ Spatio-Temporal Neural Field for 3D Point Cloud Multi-Frame Non-Linear Interpolation/layout.json @@ -0,0 +1,7936 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 67, + 102, + 526, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 102, + 526, + 140 + ], + "spans": [ + { + "bbox": [ + 67, + 102, + 526, + 140 + ], + "type": "text", + "content": "NeuralPCI: Spatio-temporal Neural Field for 3D Point Cloud Multi-frame Non-linear Interpolation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 102, + 161, + 489, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 161, + 489, + 190 + ], + "spans": [ + { + "bbox": [ + 102, + 161, + 489, + 190 + ], + "type": "text", + "content": "Zehan Zheng*, Danni Wu*, Ruisi Lu, Fan Lu, Guang Chen†, Changjun Jiang Tongji University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 102, + 191, + 490, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 191, + 490, + 204 + ], + "spans": [ + { + "bbox": [ + 102, + 191, + 490, + 204 + ], + "type": "text", + "content": "{zhengzehan, woodannie, lrs910, lufan, guangchen, cjjiang}@tongji.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 228, + 192, + 241 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 228, + 192, + 241 + ], + "spans": [ + { + "bbox": [ + 143, + 228, + 192, + 241 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 251, + 290, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 251, + 290, + 502 + ], + "spans": [ + { + "bbox": [ + 45, + 251, + 290, + 502 + ], + "type": "text", + "content": "In recent years, there has been a significant increase in focus on the interpolation task of computer vision. Despite the tremendous advancement of video interpolation, point cloud interpolation remains insufficiently explored. Meanwhile, the existence of numerous nonlinear large motions in real-world scenarios makes the point cloud interpolation task more challenging. In light of these issues, we present NeuralPCI: an end-to-end 4D spatio-temporal Neural field for 3D Point Cloud Interpolation, which implicitly integrates multi-frame information to handle nonlinear large motions for both indoor and outdoor scenarios. Furthermore, we construct a new multi-frame point cloud interpolation dataset called NL-Drive for large nonlinear motions in autonomous driving scenes to better demonstrate the superiority of our method. Ultimately, NeuralPCI achieves state-of-the-art performance on both DHB (Dynamic Human Bodies) and NL-Drive datasets. Beyond the interpolation task, our method can be naturally extended to point cloud extrapolation, morphing, and auto-labeling, which indicates its substantial potential in other domains. Codes are available at https://github.com/ispc-lab/NeuralPCI." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 521, + 128, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 521, + 128, + 533 + ], + "spans": [ + { + "bbox": [ + 47, + 521, + 128, + 533 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 540, + 287, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 540, + 287, + 636 + ], + "spans": [ + { + "bbox": [ + 46, + 540, + 287, + 636 + ], + "type": "text", + "content": "In the field of computer vision, sequential point clouds are frequently utilized in many applications, such as VR/AR techniques [11, 38, 49] and autonomous driving [4, 32, 46]. The relatively low frequency of LiDAR compared to other sensors, i.e., " + }, + { + "bbox": [ + 46, + 540, + 287, + 636 + ], + "type": "inline_equation", + "content": "10 - 20\\mathrm{Hz}" + }, + { + "bbox": [ + 46, + 540, + 287, + 636 + ], + "type": "text", + "content": ", impedes exploration for high temporal resolution point clouds [47]. Therefore, interpolation tasks for point cloud sequences, which have not been substantially investigated, are receiving increasing attention." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 637, + 288, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 637, + 288, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 637, + 288, + 696 + ], + "type": "text", + "content": "With the similar goal of obtaining a smooth sequence with high temporal resolution, we can draw inspiration from the video frame interpolation (VFI) task. Several VFI methods [6,8,20,35,45,48] concentrate on nonlinear movements in the real world. They take multiple frames as input and" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 312, + 228, + 538, + 357 + ], + "blocks": [ + { + "bbox": [ + 312, + 228, + 538, + 357 + ], + "lines": [ + { + "bbox": [ + 312, + 228, + 538, + 357 + ], + "spans": [ + { + "bbox": [ + 312, + 228, + 538, + 357 + ], + "type": "image", + "image_path": "efc8019b7730120d5f4ea61ebc6cad08679a1c0f3bfacd5a2c7fb07d0cd9dd66.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 360, + 547, + 415 + ], + "lines": [ + { + "bbox": [ + 304, + 360, + 547, + 415 + ], + "spans": [ + { + "bbox": [ + 304, + 360, + 547, + 415 + ], + "type": "text", + "content": "Figure 1. Common cases of nonlinear motions in autonomous driving scenarios. Spatially uniform linear interpolation (3) using the middle two frames of the point cloud differs significantly from the actual situation (●), so it is necessary to take multiple point clouds into consideration for nonlinear interpolation." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 426, + 546, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 546, + 473 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 546, + 473 + ], + "type": "text", + "content": "generate explicit multi-frame fusion results based on flow estimation [6, 8, 20, 45, 48] or transformer [35]. Nonetheless, due to the unique structure of point clouds [31], it is non-trivial to extend VFI methods to the 3D domain." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 475, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 475, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 475, + 547, + 714 + ], + "type": "text", + "content": "Some early works [18, 19] rely on stereo images to generate pseudo-LiDAR point cloud interpolation. For pure point cloud input, previous methods [22,47] take two consecutive frames as input and output the point cloud at a given intermediate moment. However, with limited two input frames, these approaches can only produce linear interpolation results [22], or perform nonlinear compensation by fusing input frames in the feature dimension linearly [47], which is inherently a data-driven approach to learning the dataset-specified distribution of nonlinear motions rather than an actual nonlinear interpolation. Only when the frame rate of the input point cloud sequence is high enough or the object motion is small enough, can the two adjacent point clouds satisfy the linear motion assumption. Nonetheless, there are numerous nonlinear motions in real-world cases. For instance, as illustrated in Fig. 1, the result of linear interpolation between two adjacent point cloud frames has a large deviation from the actual situation. A point cloud sequence rather than just two point cloud frames allows us to view further into the past and future. Neighboring multiple point" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 702, + 212, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 702, + 212, + 713 + ], + "spans": [ + { + "bbox": [ + 61, + 702, + 212, + 713 + ], + "type": "text", + "content": "* Equal contribution. † Corresponding author." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "909" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "content": "clouds contain additional spatial-temporal cues, namely different perspectives, complementary geometry, and implicit high-order motion information. Therefore, it is time to rethink the point cloud interpolation task with an expanded design space, which is an open challenge yet." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 133, + 289, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 133, + 289, + 384 + ], + "spans": [ + { + "bbox": [ + 46, + 133, + 289, + 384 + ], + "type": "text", + "content": "Methods that explicitly fuse multiple point cloud frames generally just approximate the motion model over time, which actually simplifies real-world complex motion. The neural field provides a more elegant way to parameterize the continuous point cloud sequence implicitly. Inspired by NeRF [25] whose view synthesis of images is essentially an interpolation, we propose NeuralPCI, a neural field to exploit the spatial and temporal information of multi-frame point clouds. We build a 4D neural spatio-temporal field, which takes sequential 3D point clouds and the independent interpolation time as input, and predicts the in-between or future point cloud at the given time. Moreover, NeuralPCI is optimized on runtime in a self-supervised manner, without relying on costly ground truths, which makes it free from the out-of-the-distribution generalization problem. Our method can be flexibly applied to segmentation auto-labeling and morphing. Besides, we newly construct a challenging multi-frame point cloud interpolation dataset called NL-Drive from public autonomous driving datasets. Finally, we achieve state-of-the-art performance on indoor DHB dataset and outdoor NL-Drive dataset." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 59, + 384, + 266, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 384, + 266, + 396 + ], + "spans": [ + { + "bbox": [ + 59, + 384, + 266, + 396 + ], + "type": "text", + "content": "Our main contributions are summarized as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 403, + 286, + 522 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 59, + 403, + 286, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 403, + 286, + 437 + ], + "spans": [ + { + "bbox": [ + 59, + 403, + 286, + 437 + ], + "type": "text", + "content": "- We propose a novel multi-frame point cloud interpolation algorithm to deal with the nonlinear complex motion in real-world indoor and outdoor scenarios." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 440, + 286, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 440, + 286, + 487 + ], + "spans": [ + { + "bbox": [ + 59, + 440, + 286, + 487 + ], + "type": "text", + "content": "- We introduce a 4D spatio-temporal neural field to integrate motion information implicitly over space and time to generate the in-between point cloud frames at the arbitrary given time." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 488, + 286, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 488, + 286, + 522 + ], + "spans": [ + { + "bbox": [ + 59, + 488, + 286, + 522 + ], + "type": "text", + "content": "- A flexible unified framework to conduct both the interpolation and extrapolation, facilitating several applications as well." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 537, + 134, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 537, + 134, + 550 + ], + "spans": [ + { + "bbox": [ + 47, + 537, + 134, + 550 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 558, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 289, + 713 + ], + "type": "text", + "content": "Video Frame Interpolation. Most VFI approaches are based on optical flow estimation, focusing on the motion cues between two consecutive input frames. These methods warp source frames with the aid of the optical flow to generate the intermediate frame [1, 13, 26, 33]. In order to deal with complex motions, some works exploit nonlinear information by expanding the design domain to multiple consecutive frames [6, 8, 20, 35, 45]. QVI [45] approximates the flow-based velocity and acceleration of the quadratic motion model explicitly. Following QVI, EQVI [20] improves the training strategy. Recently, Dutta et al. [8] utilizes space-time convolution to adaptively switch motion models with discontinuous motions. These works inspire us" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "content": "to design a point cloud interpolation network with multiple input frames. However, it is still challenging to extend these methods to the unordered and unstructured point cloud." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 108, + 546, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 108, + 546, + 371 + ], + "spans": [ + { + "bbox": [ + 304, + 108, + 546, + 371 + ], + "type": "text", + "content": "Point Cloud Interpolation. Existing point cloud interpolation methods try to find point-to-point correspondences between two point cloud reference frames. An intuitive way is to utilize scene flow, the extension of optical flow in the 3D domain. PointINet [22] warps two input frames with bi-directional flows to the intermediate frame, then samples the two warped results adaptively and fuses them with attentive weights. This approach is performed under the linear motion assumption and relies heavily on the accuracy of the scene flow backbone. IDEA-Net [47] proposes an alternative method to solve the correlation between two input frames by learning a one-to-one alignment matrix and refining linear interpolation results with a trajectory compensation module. However, the one-to-one correspondence assumption limits its application for large-scale outdoor point cloud datasets. Moreover, the higher-order motion information in the temporal domain is overlooked [8] when taking two frames as input. Therefore, there remain challenges to capturing and modeling the complex nonlinear motion in the real world. To address this issue, we propose a novel neural field that takes advantage of the more comprehensive spatio-temporal information of multiple point cloud frames." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 372, + 546, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 372, + 546, + 635 + ], + "spans": [ + { + "bbox": [ + 304, + 372, + 546, + 635 + ], + "type": "text", + "content": "Neural Implicit Representation. Distinct from the common paradigm of learning-based methods, neural fields are fitted to a single degraded sample rather than a large dataset of samples. The neural field can be seen as a parameterization of diverse input types, such as 2D images [5, 14], 3D shapes [24, 27], etc. Since NeRF [25] presents a novel neural radiance field, which encodes a scene with spatial location and view direction as input and outputs volume rendering, several dynamic scene synthesis studies [7, 10, 16, 17, 28, 30, 36, 43] have been proposed to exploit the representation ability of neural radiance fields for dynamic scenes. Based on the linear motion assumption, Li et al. [17] proposes a time-variant continuous neural representation for space-time view synthesis. NSFP [16] presents a neural prior to regularize scene flow implicitly. Following NSFP, Wang et al. [37] introduce a neural trajectory prior to representing the trajectories as a vector field. Coordinate-based networks show great potential for encoding a continuous input domain over arbitrary dimensions at an arbitrary resolution. Our work utilizes and exploits the ability of the coordinate-based network to represent continuous spatio-temporal motions of dynamic nonlinear scenes." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 646, + 389, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 646, + 389, + 659 + ], + "spans": [ + { + "bbox": [ + 306, + 646, + 389, + 659 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 666, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 666, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 547, + 715 + ], + "type": "text", + "content": "In this section, we propose a novel end-to-end 4D spatio-temporal neural field for 3D point cloud interpolation named NeuralPCI. Firstly, we state the problem of multi-frame point cloud interpolation. Sec. 3.1 then pro" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "910" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 73, + 70, + 521, + 280 + ], + "blocks": [ + { + "bbox": [ + 73, + 70, + 521, + 280 + ], + "lines": [ + { + "bbox": [ + 73, + 70, + 521, + 280 + ], + "spans": [ + { + "bbox": [ + 73, + 70, + 521, + 280 + ], + "type": "image", + "image_path": "ea647606e55ef10498368221cf09f4372a4ef7e02363d21c648cac5fbdec5214.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 286, + 547, + 331 + ], + "lines": [ + { + "bbox": [ + 45, + 286, + 547, + 331 + ], + "spans": [ + { + "bbox": [ + 45, + 286, + 547, + 331 + ], + "type": "text", + "content": "Figure 2. Overview of our proposed NeuralPCI. The 4D neural field is constructed by encoding the spatio-temporal coordinate of the multi-frame input point clouds via a coordinate-based multi-layer perceptron network. In the inference stage after self-supervised optimization, NeuralPCI receives a reference point cloud and an arbitrary interpolation frame moment as input to generate the point cloud of the associated spatio-temporal location." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 340, + 288, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 340, + 288, + 422 + ], + "spans": [ + { + "bbox": [ + 46, + 340, + 288, + 422 + ], + "type": "text", + "content": "vides an overview of NeuralPCI's architecture and design philosophy. Following this, we explain the details of how to construct the 4D spatio-temporal neural field and integrate neighboring multi-frame nonlinear information of point clouds in Secs. 3.2 and 3.3. Finally, in Sec. 3.4, we further elaborate on the self-supervised optimization manner of the whole neural field." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 423, + 287, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 423, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 423, + 287, + 544 + ], + "type": "text", + "content": "Problem Formulation. Let " + }, + { + "bbox": [ + 46, + 423, + 287, + 544 + ], + "type": "inline_equation", + "content": "P \\in \\mathbb{R}^{N \\times 3}" + }, + { + "bbox": [ + 46, + 423, + 287, + 544 + ], + "type": "text", + "content": " be one frame of a dynamic point cloud sequence, with " + }, + { + "bbox": [ + 46, + 423, + 287, + 544 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 423, + 287, + 544 + ], + "type": "text", + "content": " sampled points at the time " + }, + { + "bbox": [ + 46, + 423, + 287, + 544 + ], + "type": "inline_equation", + "content": "t \\in \\mathbb{R}" + }, + { + "bbox": [ + 46, + 423, + 287, + 544 + ], + "type": "text", + "content": ". Given a low temporal resolution sequence of " + }, + { + "bbox": [ + 46, + 423, + 287, + 544 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 423, + 287, + 544 + ], + "type": "text", + "content": " frames of point clouds " + }, + { + "bbox": [ + 46, + 423, + 287, + 544 + ], + "type": "inline_equation", + "content": "S = \\{P_0, P_1, \\dots, P_{M-1}\\}" + }, + { + "bbox": [ + 46, + 423, + 287, + 544 + ], + "type": "text", + "content": " with their corresponding timestamps " + }, + { + "bbox": [ + 46, + 423, + 287, + 544 + ], + "type": "inline_equation", + "content": "T = \\{t_0, t_1, \\dots, t_{M-1}\\}" + }, + { + "bbox": [ + 46, + 423, + 287, + 544 + ], + "type": "text", + "content": ", the goal of NeuralPCI is to predict the intermediate point cloud frame " + }, + { + "bbox": [ + 46, + 423, + 287, + 544 + ], + "type": "inline_equation", + "content": "P_i" + }, + { + "bbox": [ + 46, + 423, + 287, + 544 + ], + "type": "text", + "content": " at an arbitrary given time " + }, + { + "bbox": [ + 46, + 423, + 287, + 544 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 46, + 423, + 287, + 544 + ], + "type": "text", + "content": " for interpolation. Interpolating " + }, + { + "bbox": [ + 46, + 423, + 287, + 544 + ], + "type": "inline_equation", + "content": "n \\in \\mathbb{Z}^+" + }, + { + "bbox": [ + 46, + 423, + 287, + 544 + ], + "type": "text", + "content": " frames of point clouds at equal intervals between every two consecutive frames yields a sequence of point clouds with " + }, + { + "bbox": [ + 46, + 423, + 287, + 544 + ], + "type": "inline_equation", + "content": "n + 1" + }, + { + "bbox": [ + 46, + 423, + 287, + 544 + ], + "type": "text", + "content": " times higher temporal resolution." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 544, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 544, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 544, + 287, + 628 + ], + "type": "text", + "content": "For the multi-frame point cloud interpolation task, we define the input as every four consecutive point cloud frames in the sequence " + }, + { + "bbox": [ + 46, + 544, + 287, + 628 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 544, + 287, + 628 + ], + "type": "text", + "content": " and assume that these frames are equally time-spaced for convenience. The output is the temporally uniform interpolation result of " + }, + { + "bbox": [ + 46, + 544, + 287, + 628 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 544, + 287, + 628 + ], + "type": "text", + "content": " point cloud frames between the middle two frames. Then, sliding the window of multi-frame inputs traverses the entire sequence." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 635, + 168, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 168, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 168, + 647 + ], + "type": "text", + "content": "3.1. NeuralPCI Overview" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": "Mildenhall et al. [25] build the neural radiance field by inputting a series of 2D images with different viewpoints, and then generate the unknown image under a new viewpoint using neural volume rendering. Although the huge gap between images and point clouds makes it non-trivial to" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 341, + 545, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 341, + 545, + 436 + ], + "spans": [ + { + "bbox": [ + 304, + 341, + 545, + 436 + ], + "type": "text", + "content": "apply the neural field technique to the point cloud interpolation task, we still want to find a way to address the nonlinear interpolation problem of multi-frame point clouds under a similar design philosophy. That is, encoding a sequence of 3D point clouds at different moments to construct a 4D neural spatio-temporal field and then feeding an arbitrary interpolation frame moment into the network to generate the point cloud of the associated spatio-temporal location." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 437, + 546, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 546, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 546, + 616 + ], + "type": "text", + "content": "Consequently, we propose NeuralPCI whose overall structure is depicted in Fig. 2. It can be divided into three main sections for elaboration. In the optimization stage, the 4D neural field is constructed by encoding the spatiotemporal coordinate of the multi-frame input point cloud via a coordinate-based Multi-Layer Perceptron (MLP) network. For each point cloud frame of the input, the interpolation time is set to the corresponding timestamps of four input frames in order to provide the network with the ability to generate the point cloud of the respective spatio-temporal position, and then optimize the neural field through self-supervised losses. In the inference stage, we run forward the neural field with the spatio-temporal coordinate of a reference point cloud and the moment of the interpolation frame as input to obtain the corresponding in-between point cloud." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 623, + 484, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 623, + 484, + 636 + ], + "spans": [ + { + "bbox": [ + 306, + 623, + 484, + 636 + ], + "type": "text", + "content": "3.2. 4D Spatio-temporal Neural Field" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": "Following the statement in [44], a field is a quantity defined for all spatial and/or temporal coordinates, and a neural field is a field that is parameterized fully or in part by a neural network. Here, we use a coordinate-based MLP network to represent the scenes, which takes as input the 3D spatial coordinate " + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": " and the 1D temporal coordi" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "911" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "type": "text", + "content": "nate " + }, + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "type": "inline_equation", + "content": "t\\in \\mathbb{R}" + }, + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "type": "text", + "content": " , and produces as output the motions " + }, + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{x}\\in \\mathbb{R}^3" + }, + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "type": "text", + "content": " The MLP is parameterized by " + }, + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "type": "text", + "content": " and it can be viewed as a mapping from the coordinate field to the motion field." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 108, + 288, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 108, + 288, + 217 + ], + "spans": [ + { + "bbox": [ + 46, + 108, + 288, + 217 + ], + "type": "text", + "content": "Furthermore, we leverage an independent input, the interpolation frame time " + }, + { + "bbox": [ + 46, + 108, + 288, + 217 + ], + "type": "inline_equation", + "content": "t_{intp} \\in \\mathbb{R}" + }, + { + "bbox": [ + 46, + 108, + 288, + 217 + ], + "type": "text", + "content": ", to provide MLP with time cues and regulate the scene motion output at the interpolation moment relative to the original input spatio-temporal coordinates. Then, we conduct the point-wise addition of the interpolation frame motion and the input point cloud to generate the final output of the neural field, which is the 3D point cloud coordinate at the interpolation frame moment " + }, + { + "bbox": [ + 46, + 108, + 288, + 217 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{intp} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 46, + 108, + 288, + 217 + ], + "type": "text", + "content": ". Formally, the neural field " + }, + { + "bbox": [ + 46, + 108, + 288, + 217 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 46, + 108, + 288, + 217 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 102, + 229, + 287, + 243 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 229, + 287, + 243 + ], + "spans": [ + { + "bbox": [ + 102, + 229, + 287, + 243 + ], + "type": "interline_equation", + "content": "\\Phi (\\mathbf {z}; \\Theta): \\mathbf {x} _ {i n t p} = \\mathbf {x} + F _ {\\Theta} (\\mathbf {z}) \\tag {1}", + "image_path": "75acee6284b974eaefa2622e684971e4660c48e459ba397659387392996e2fb3.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 254, + 287, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 254, + 287, + 278 + ], + "spans": [ + { + "bbox": [ + 47, + 254, + 287, + 278 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 254, + 287, + 278 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 47, + 254, + 287, + 278 + ], + "type": "text", + "content": " represents the spatial-temporal coordinate input of point clouds, and " + }, + { + "bbox": [ + 47, + 254, + 287, + 278 + ], + "type": "inline_equation", + "content": "F_{\\Theta}" + }, + { + "bbox": [ + 47, + 254, + 287, + 278 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 289, + 287, + 303 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 289, + 287, + 303 + ], + "spans": [ + { + "bbox": [ + 58, + 289, + 287, + 303 + ], + "type": "interline_equation", + "content": "F _ {\\Theta}: \\mathbb {R} ^ {5} \\rightarrow \\mathbb {R} ^ {3}, \\quad \\Delta \\mathbf {x} = F _ {\\Theta} (\\mathbf {z}) = F _ {\\Theta} (\\mathbf {x}, t, t _ {i n t p}) \\tag {2}", + "image_path": "cef3281b8dd27f53a4440daa3fcdda43ed1e5b5607cb85ddd2792b7bb7b481f3.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 315, + 287, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 315, + 287, + 386 + ], + "spans": [ + { + "bbox": [ + 46, + 315, + 287, + 386 + ], + "type": "text", + "content": "In such a manner, we implicitly construct a 4D, i.e., " + }, + { + "bbox": [ + 46, + 315, + 287, + 386 + ], + "type": "inline_equation", + "content": "(x,y,z,t)" + }, + { + "bbox": [ + 46, + 315, + 287, + 386 + ], + "type": "text", + "content": " in physical, spatio-temporal neural field to represent the entire scenarios of sequential point clouds. Eventually, we are able to utilize the continuity of the dedicated neural field to smoothly interpolate the point cloud at an arbitrary in-between moment." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 395, + 233, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 395, + 233, + 407 + ], + "spans": [ + { + "bbox": [ + 47, + 395, + 233, + 407 + ], + "type": "text", + "content": "3.3. Multi-frame Nonlinear Integration" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 414, + 287, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 414, + 287, + 497 + ], + "spans": [ + { + "bbox": [ + 46, + 414, + 287, + 497 + ], + "type": "text", + "content": "In order to integrate multi-frame information, an intuitive way is utilizing the existing pair-frame point cloud interpolation algorithm for every two frames among the multiframe point cloud input and fuse the predicted results. We adopt it as a baseline model (see Supplementary Material), and it turns out that directly fusing multiple intermediate predictions by random sampling leads to worse results." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 498, + 287, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 616 + ], + "type": "text", + "content": "Besides, an alternative way is to explicitly model the nonlinear kinematic equations taking advantage of multiframe point clouds. We follow the nonlinear video interpolation algorithm [45] and extend it to the 3D domain to formulate the high order equation of point clouds. Nonetheless, the explicit modeling approach is also ineffective in real-world scenarios with complex motions (see Supplementary Material). Consequently, we integrate multi-frame point clouds more effectively using spatio-temporal neural fields." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": "Our proposed NeuralPCI does not impose a restriction on the frame number of the input point clouds, and thus can be naturally expanded to multi-frame point clouds. When the number of multi-frame inputs is 4, for example, the input set containing a point cloud sequence and corresponding timestamps are " + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "inline_equation", + "content": "S = \\{P_0, P_1, P_2, P_3\\}" + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "inline_equation", + "content": "T = \\{t_0, t_1, t_2, t_3\\}" + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": ", respectively. As shown in Fig. 2, the neural field receives the point cloud " + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "inline_equation", + "content": "P_0" + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "text", + "content": " and time " + }, + { + "bbox": [ + 46, + 617, + 287, + 713 + ], + "type": "inline_equation", + "content": "t_0" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "text", + "content": "as input, and the time step for interpolation " + }, + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "inline_equation", + "content": "t_{intp}" + }, + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "text", + "content": " is adjusted to one of " + }, + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "inline_equation", + "content": "\\{t_0, t_1, t_2, t_3\\}" + }, + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "text", + "content": " to yield the predicted point clouds " + }, + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "inline_equation", + "content": "\\{\\hat{P}_0^{t_0}, \\hat{P}_0^{t_1}, \\hat{P}_0^{t_2}, \\hat{P}_0^{t_3}\\}" + }, + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "text", + "content": ". The loss function (described in Sec. 3.4) is then calculated between the prediction " + }, + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "inline_equation", + "content": "\\hat{P}_i^{t_j}" + }, + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "text", + "content": " and each of the four input frames of point clouds " + }, + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "inline_equation", + "content": "P_j" + }, + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "text", + "content": ". The entire point cloud sequence is traversed through the same operation, with the spatio-temporal neural field end-to-end optimized meanwhile." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 168, + 545, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 168, + 545, + 276 + ], + "spans": [ + { + "bbox": [ + 304, + 168, + 545, + 276 + ], + "type": "text", + "content": "Each input frame serves as a constraint to supervise optimization, allowing us to incorporate the information from multiple frames of point clouds more elegantly. Owing to the continuity, smoothness, and excellent fitting ability of MLP, the derivative of the final spatio-temporal neural field function with respect to the interpolated frame time " + }, + { + "bbox": [ + 304, + 168, + 545, + 276 + ], + "type": "inline_equation", + "content": "t_{intp}" + }, + { + "bbox": [ + 304, + 168, + 545, + 276 + ], + "type": "text", + "content": " is an implicit higher-order continuous function, which can better handle complex motion scenes and can yield smooth interpolation outputs." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 281, + 465, + 294 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 281, + 465, + 294 + ], + "spans": [ + { + "bbox": [ + 305, + 281, + 465, + 294 + ], + "type": "text", + "content": "3.4. Self-supervised Optimization" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 300, + 545, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 300, + 545, + 383 + ], + "spans": [ + { + "bbox": [ + 304, + 300, + 545, + 383 + ], + "type": "text", + "content": "NeuralPCI optimizes the weights of the neural field in a self-supervised manner. As illustrated in Fig. 2, the interpolation time is adjusted as each timestamp of inputs to generate predictions for all input point cloud frames, and the neural field back-propagates the gradients to update the samplespecified parameters by minimizing the following distribution loss between the input and predicted point clouds." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 383, + 545, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 383, + 545, + 428 + ], + "spans": [ + { + "bbox": [ + 304, + 383, + 545, + 428 + ], + "type": "text", + "content": "CD Loss. Chamfer Distance (CD) [9] measures the distribution difference between two point clouds. We adopt CD as the main term in the loss function, which can be expressed as the following equation:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 311, + 431, + 545, + 471 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 431, + 545, + 471 + ], + "spans": [ + { + "bbox": [ + 311, + 431, + 545, + 471 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {C D} = \\frac {1}{N} \\sum_ {\\hat {p} _ {i} \\in \\hat {P}} \\min _ {p _ {i} \\in P} \\| \\hat {p} _ {i} - p _ {i} \\| _ {2} ^ {2} + \\frac {1}{N} \\sum_ {p _ {i} \\in P} \\min _ {\\hat {p} _ {i} \\in \\hat {P}} \\| p _ {i} - \\hat {p} _ {i} \\| _ {2} ^ {2} \\tag {3}", + "image_path": "cc7dd98e2851607f4b3bbfa52348e9d8376e09cf56b33a58e4584fcf8a87ca71.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 473, + 545, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 473, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 473, + 545, + 510 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 473, + 545, + 510 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 473, + 545, + 510 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 473, + 545, + 510 + ], + "type": "inline_equation", + "content": "\\hat{P}" + }, + { + "bbox": [ + 304, + 473, + 545, + 510 + ], + "type": "text", + "content": " are the input and predicted point cloud frames. " + }, + { + "bbox": [ + 304, + 473, + 545, + 510 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 304, + 473, + 545, + 510 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 473, + 545, + 510 + ], + "type": "inline_equation", + "content": "\\hat{p}_i" + }, + { + "bbox": [ + 304, + 473, + 545, + 510 + ], + "type": "text", + "content": " represent the points in the respective point clouds. " + }, + { + "bbox": [ + 304, + 473, + 545, + 510 + ], + "type": "inline_equation", + "content": "\\| .\\| _2" + }, + { + "bbox": [ + 304, + 473, + 545, + 510 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 304, + 473, + 545, + 510 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 304, + 473, + 545, + 510 + ], + "type": "text", + "content": " norm of the spatial coordinate." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 510, + 545, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 569 + ], + "type": "text", + "content": "EMD Loss. Earth Mover's Distance (EMD) [34] calculates the corresponding points by solving the optimal transmission matrix of the two point clouds. We minimize the EMD loss to encourage the two point clouds to have the same density distribution, calculated as:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 347, + 574, + 545, + 605 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 574, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 347, + 574, + 545, + 605 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {E M D} = \\min _ {\\phi : \\hat {P} \\rightarrow P} \\frac {1}{N} \\sum_ {\\hat {p} \\in \\hat {P}} \\| \\hat {p} - \\phi (\\hat {p}) \\| _ {2} ^ {2} \\tag {4}", + "image_path": "bb29a16a2b12861d2a35edf86a8ec8f208673fcc1d2a8f2eb26ba12053656af0.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 603, + 545, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 603, + 545, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 603, + 545, + 639 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 603, + 545, + 639 + ], + "type": "inline_equation", + "content": "\\phi : \\hat{P} \\to P" + }, + { + "bbox": [ + 304, + 603, + 545, + 639 + ], + "type": "text", + "content": " denotes a bijection from " + }, + { + "bbox": [ + 304, + 603, + 545, + 639 + ], + "type": "inline_equation", + "content": "\\hat{P}" + }, + { + "bbox": [ + 304, + 603, + 545, + 639 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 603, + 545, + 639 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 603, + 545, + 639 + ], + "type": "text", + "content": ". Due to its high computational complexity, we only use it in the loss function for sparse point clouds." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 639, + 545, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 639, + 545, + 684 + ], + "spans": [ + { + "bbox": [ + 304, + 639, + 545, + 684 + ], + "type": "text", + "content": "Smoothness Loss. To better regulate the estimated motion, we introduce the smoothness loss, which facilitates the interpolation of point clouds for local rigid motions and is utilized in autonomous driving scenarios." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 339, + 687, + 545, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 687, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 339, + 687, + 545, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {S} = \\sum_ {p _ {i} \\in P} \\frac {1}{| N (p _ {i}) |} \\sum_ {p _ {j} \\in N (p _ {i})} \\| \\Delta \\mathbf {x} _ {j} - \\Delta \\mathbf {x} _ {i} \\| _ {2} ^ {2} \\tag {5}", + "image_path": "efb1897734ccff84a691d08cc9c5667c41fa3c7a461d3d802513763d4ab4e7de.jpg" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "912" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "N(p_i)" + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": " stands for the nearest neighbors of the " + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": " point " + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "|\\cdot|" + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": " denotes the number of points. And " + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{x}_i" + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": " denotes the scene flow estimation from point cloud " + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "\\hat{P}" + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": " (defined in Eq. (2)) of the " + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": " point " + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 121, + 288, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 121, + 288, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 121, + 288, + 228 + ], + "type": "text", + "content": "Total Loss Function. We utilize the above three terms of loss and balance them with separately given weights (defined as " + }, + { + "bbox": [ + 46, + 121, + 288, + 228 + ], + "type": "inline_equation", + "content": "\\alpha, \\beta, \\gamma" + }, + { + "bbox": [ + 46, + 121, + 288, + 228 + ], + "type": "text", + "content": " in Eq. (6)) to obtain the total loss. We define multiple input frames and their corresponding timestamps as the input set. Then, we traverse the input set by assigning each point cloud of input frames as the reference and generate predictions of all input frames at their corresponding time steps. The overall loss is computed by summing the losses of every pair of point clouds as Eq. (7)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 102, + 238, + 287, + 251 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 238, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 102, + 238, + 287, + 251 + ], + "type": "interline_equation", + "content": "\\Psi = \\alpha \\mathcal {L} _ {C D} + \\beta \\mathcal {L} _ {E M D} + \\gamma \\mathcal {L} _ {S} \\tag {6}", + "image_path": "46e79815bbc1afc4cab10f3163d732fbe768f3854390bd7abdc5a5f35f98049f.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 107, + 261, + 287, + 289 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 261, + 287, + 289 + ], + "spans": [ + { + "bbox": [ + 107, + 261, + 287, + 289 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\sum_ {P _ {i} \\in S} \\sum_ {t _ {j} \\in T} \\Psi \\left(P _ {t _ {j}}, \\hat {P} _ {i} ^ {t _ {j}}\\right) \\tag {7}", + "image_path": "62f51fd4076b3b98f7230269533718a74f57030846863ff77352a535b9165e4b.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 300, + 287, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 300, + 287, + 351 + ], + "spans": [ + { + "bbox": [ + 47, + 300, + 287, + 351 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 300, + 287, + 351 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 47, + 300, + 287, + 351 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 300, + 287, + 351 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 47, + 300, + 287, + 351 + ], + "type": "text", + "content": " are the multiple input frames and their corresponding timestamps. " + }, + { + "bbox": [ + 47, + 300, + 287, + 351 + ], + "type": "inline_equation", + "content": "P_{t_j}" + }, + { + "bbox": [ + 47, + 300, + 287, + 351 + ], + "type": "text", + "content": " denotes the input frame at time " + }, + { + "bbox": [ + 47, + 300, + 287, + 351 + ], + "type": "inline_equation", + "content": "t_j" + }, + { + "bbox": [ + 47, + 300, + 287, + 351 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 300, + 287, + 351 + ], + "type": "inline_equation", + "content": "\\hat{P}_i^{t_j}" + }, + { + "bbox": [ + 47, + 300, + 287, + 351 + ], + "type": "text", + "content": " denotes the neural field prediction when " + }, + { + "bbox": [ + 47, + 300, + 287, + 351 + ], + "type": "inline_equation", + "content": "P_i" + }, + { + "bbox": [ + 47, + 300, + 287, + 351 + ], + "type": "text", + "content": " is the reference frame and the interpolation time is " + }, + { + "bbox": [ + 47, + 300, + 287, + 351 + ], + "type": "inline_equation", + "content": "t_j" + }, + { + "bbox": [ + 47, + 300, + 287, + 351 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 362, + 128, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 362, + 128, + 376 + ], + "spans": [ + { + "bbox": [ + 47, + 362, + 128, + 376 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 383, + 163, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 383, + 163, + 396 + ], + "spans": [ + { + "bbox": [ + 47, + 383, + 163, + 396 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 402, + 287, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 402, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 46, + 402, + 287, + 545 + ], + "type": "text", + "content": "Datasets. We evaluate NeuralPCI in both indoor and outdoor datasets. Indoor Dynamic Human Bodies dataset (DHB) [47] contains point cloud sequences for the non-rigid deformable 3D human motion with sampled 1024 points. We construct a challenging multi-frame interpolation dataset named Nonliner-Drive (NL-Drive) dataset. Based on the principle of hard-sample selection and the diversity of scenarios, NL-Drive dataset contains point cloud sequences with large nonlinear movements from three public large-scale autonomous driving datasets: KITTI [12], Argoverse [3] and Nuscenes [2]. More details of NL-Drive dataset are provided in Supplementary Material." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 545, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 288, + 714 + ], + "type": "text", + "content": "Baselines. To demonstrate the performance of NeuralPCI, we compare our method with previous SOTA approaches, namely IDEA-Net [47] and PointINet [22] for the interpolation task and MoNet [21] for the extrapolation task. We reproduced the results on DHB Dataset of IDEA-Net and PointINet using official implementation. Moreover, we train PointINet and MoNet on NL-Drive dataset according to the official training setting (Since the training code of IDEA-Net has not been released yet, we do not report its results here). In addition, we utilize state-of-the-art scene flow estimation methods, i.e., neural-based NSFP [16] and recurrent-based PV-RAFT [39], with linear interpolation to produce corresponding results on both DHB and NL-Drive datasets for comprehensive comparison. Remarkably, we" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": "evaluate all the optimization-based methods directly on the test set without using the training set." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 96, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 96, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 305, + 96, + 545, + 133 + ], + "type": "text", + "content": "Metrics. We adopt CD and EMD as quantitative evaluation metrics following [21, 22, 47], which are described in Eqs. (3) and (4), respectively." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 141, + 440, + 153 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 141, + 440, + 153 + ], + "spans": [ + { + "bbox": [ + 306, + 141, + 440, + 153 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 159, + 545, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 159, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 304, + 159, + 545, + 316 + ], + "type": "text", + "content": "We implement NeuralPCI with PyTorch [29]. We define our NeuralPCI as a coordinate-based 8-layer MLP architecture with 512 units per layer and adopt LeakyReLU as the activation function. The network weights are randomly initialized with the Adam [15] optimizer (the " + }, + { + "bbox": [ + 304, + 159, + 545, + 316 + ], + "type": "inline_equation", + "content": "lr" + }, + { + "bbox": [ + 304, + 159, + 545, + 316 + ], + "type": "text", + "content": " is 0.001). For each sample, the maximum optimization step is limited to 1000 iterations. The spatio-temporal coordinates of the point clouds are position-encoded by a sinusoidal function and fed into the MLP, while the interpolation time is inserted into the penultimate hidden layer to control the final output. All experiments were conducted on a single NVIDIA GeForce RTX 3090 GPU. For more implementation details, please refer to Supplementary Material." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 323, + 515, + 337 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 323, + 515, + 337 + ], + "spans": [ + { + "bbox": [ + 306, + 323, + 515, + 337 + ], + "type": "text", + "content": "4.3. Evaluation of Point Cloud Interpolation" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 342, + 545, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 342, + 545, + 545 + ], + "spans": [ + { + "bbox": [ + 304, + 342, + 545, + 545 + ], + "type": "text", + "content": "Results on DHB dataset. The quantitative comparison on DHB dataset is displayed in Table 1, where our NeuralPCI outperforms other baseline approaches by a large margin. In particular, our overall CD is about half that of other baselines, and our overall EMD is " + }, + { + "bbox": [ + 304, + 342, + 545, + 545 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 304, + 342, + 545, + 545 + ], + "type": "text", + "content": " lower than the suboptimal PV-RAFT [39]. As can be seen, IDEA-Net [47] does not perform well in all scenarios, but in contrast, our method achieves the best results in every single scene, especially in the Squat scenario. This demonstrates the flexibility and adaptability of the neural field in various indoor scenarios. Furthermore, Fig. 3 clearly exhibits the outcomes of qualitative experiments. The interpolation results of PointINet [22] and IDEA-Net contain several outlier noise points and lack lots of detailed information in local areas, such as hair, hands, and skirt hems. Instead, our method benefits from the higher-order implicit function of the neural field, which better handles these complex motions." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": "Results on NL-Drive dataset. Table 2 shows results on NL-Drive dataset, where NeuralPCI achieves the best results on most frames and overall metrics. For instance, our method reaches comparable EMD error in the intermediate frames and significantly reduces the CD error, eventually outperforming SOTA by " + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "inline_equation", + "content": "24.5\\%" + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "inline_equation", + "content": "4\\%" + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": " in the overall results of CD and EMD metrics, respectively. In the qualitative experiments, we show in detail the nonlinear motion in the outdoor autonomous driving scenario as well as the interpolation frame comparison results in Fig. 5. This indicates that NeuralPCI is scalable to large dense point clouds. As noted in the local zoomed-in view, the vehicle edges are clearer and sharper in the results of our method, while they are blurrier and noisier in the results of PointINet [22]." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "913" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 101, + 547, + 194 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 545, + 93 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 545, + 93 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 545, + 93 + ], + "type": "text", + "content": "Table 1. Quantitative comparison " + }, + { + "bbox": [ + 47, + 70, + 545, + 93 + ], + "type": "inline_equation", + "content": "\\left( {\\times {10}^{-3}}\\right)" + }, + { + "bbox": [ + 47, + 70, + 545, + 93 + ], + "type": "text", + "content": " with other open-sourced methods on DHB-Dataset [47]. Baseline methods include IDEA-Net [47], PointINet [22] and the results of linear interpolation of scene flow estimated by NSFP [17] and PV-RAFT [39]." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 101, + 547, + 194 + ], + "lines": [ + { + "bbox": [ + 50, + 101, + 547, + 194 + ], + "spans": [ + { + "bbox": [ + 50, + 101, + 547, + 194 + ], + "type": "table", + "html": "
MethodsLongdressLootRed&BlackSoldierSquatSwingOverall
CDEMDCDEMDCDEMDCDEMDCDEMDCDEMDCD↓EMD↓
IDEA-Net0.896.010.868.620.9410.341.6330.070.626.681.246.931.0212.03
PointINet0.9810.870.8512.100.8710.680.9712.390.9013.991.4514.810.9612.25
NSFP1.047.450.817.130.978.140.685.251.147.973.0911.391.227.81
PV-RAFT1.036.880.825.990.947.030.915.310.572.811.4210.540.926.14
NeuralPCI0.704.360.614.760.674.790.594.630.030.020.532.220.543.68
", + "image_path": "647ebd2579cebf4efc9c83f05b5719011b048d75091c585bcfea5c8bb4f2a6bb.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 51, + 245, + 545, + 350 + ], + "blocks": [ + { + "bbox": [ + 46, + 203, + 546, + 237 + ], + "lines": [ + { + "bbox": [ + 46, + 203, + 546, + 237 + ], + "spans": [ + { + "bbox": [ + 46, + 203, + 546, + 237 + ], + "type": "text", + "content": "Table 2. Quantitative comparison with other open-sourced methods on NL-Drive Dataset. Type indicates whether the interpolation results are based on forward, backward, bidirectional flow, or neural field. Frame-1, Frame-2 and Frame-3 refer to the three interpolation frames located at equal intervals between the two intermediate input frames." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 51, + 245, + 545, + 350 + ], + "lines": [ + { + "bbox": [ + 51, + 245, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 51, + 245, + 545, + 350 + ], + "type": "table", + "html": "
MethodsTypeFrame-1Frame-2Frame-3Average
CDEMDCDEMDCDEMDCD ↓EMD ↓
NSFPforward flow0.9495.181.75132.302.55168.911.75132.13
backward flow2.53168.751.74132.190.9595.231.74132.05
PV-RAFTforward flow1.36104.571.92146.871.63169.821.64140.42
backward flow1.58173.181.85145.481.30102.711.58140.46
PointINetbi-directional flow0.9397.481.24110.221.0195.651.06101.12
NeuralPCIneural field0.7289.030.94113.450.7488.610.8097.03
", + "image_path": "9abd6701c14ed17a1850a9eac91ae0fe8fa873fc25567dd27f62e6d6915cc173.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 48, + 365, + 289, + 559 + ], + "blocks": [ + { + "bbox": [ + 48, + 365, + 289, + 559 + ], + "lines": [ + { + "bbox": [ + 48, + 365, + 289, + 559 + ], + "spans": [ + { + "bbox": [ + 48, + 365, + 289, + 559 + ], + "type": "image", + "image_path": "1a6defa161db8749d1b52d41238b60791b41cd929e14d8bfa05b042bdce12318.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 566, + 287, + 611 + ], + "lines": [ + { + "bbox": [ + 46, + 566, + 287, + 611 + ], + "spans": [ + { + "bbox": [ + 46, + 566, + 287, + 611 + ], + "type": "text", + "content": "Figure 3. Qualitative results on DHB dataset. Each column represents one interpolation frame result in the point cloud sequence, where our method is more consistent with the ground truth and preserves local details better than previous SOTA methods." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": "Additional Evaluation. To evaluate the capability of NeuralPCI under large motions, we conduct additional experiments to observe the robustness of each method with increasing time interval between input frames. As illustrated in Fig. 4, the results on both datasets show that the performance of our method still lies at the optimum under large motion interpolation." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 308, + 366, + 545, + 463 + ], + "blocks": [ + { + "bbox": [ + 308, + 366, + 545, + 463 + ], + "lines": [ + { + "bbox": [ + 308, + 366, + 545, + 463 + ], + "spans": [ + { + "bbox": [ + 308, + 366, + 545, + 463 + ], + "type": "image", + "image_path": "5b073cfc5595db6a6f2563170776391586281f2ca8ba7e7af716771f13fbf3b9.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 472, + 545, + 517 + ], + "lines": [ + { + "bbox": [ + 305, + 472, + 545, + 517 + ], + "spans": [ + { + "bbox": [ + 305, + 472, + 545, + 517 + ], + "type": "text", + "content": "Figure 4. Quantitative experiments at different intervals. The CD error increases as the input frame interval grows. Our method is always optimal at different time intervals and has good robustness under long-distance motion." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 527, + 518, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 527, + 518, + 540 + ], + "spans": [ + { + "bbox": [ + 306, + 527, + 518, + 540 + ], + "type": "text", + "content": "4.4. Evaluation of Point Cloud Extrapolation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 545, + 545, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 676 + ], + "type": "text", + "content": "As a parameterization of continuous point clouds over space and time, NeuralPCI is generalizable to predict the near future frame by adjusting the timestamp, making it more flexible than existing methods. We conduct extrapolation experiments on NL-Drive dataset with the same input settings as the interpolation task, while the outputs are four consecutive point cloud frames in the future. According to recent works about point cloud prediction [21, 23, 40, 41], we adopt the state-of-the-art method MoNet [21] and also the linear extrapolation results based on scene flow from PV-RAFT [39] and NSFP [16] as baseline methods." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": "A quantitative comparison with baseline methods is shown in Table 3. Flow-based methods suffer a sharp growth in error when the time for the future frame increases." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "914" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 78, + 71, + 294, + 144 + ], + "blocks": [ + { + "bbox": [ + 78, + 71, + 294, + 144 + ], + "lines": [ + { + "bbox": [ + 78, + 71, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 78, + 71, + 294, + 144 + ], + "type": "image", + "image_path": "1e44e8c631839509492b3e33965483a7b6a03e71df422c6d1220ce160d701efe.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 295, + 72, + 514, + 144 + ], + "blocks": [ + { + "bbox": [ + 295, + 72, + 514, + 144 + ], + "lines": [ + { + "bbox": [ + 295, + 72, + 514, + 144 + ], + "spans": [ + { + "bbox": [ + 295, + 72, + 514, + 144 + ], + "type": "image", + "image_path": "df9422f70d89d7b302c2415196feaa8b075935001bb9c0d01d4266311a49f7a8.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 78, + 145, + 294, + 201 + ], + "blocks": [ + { + "bbox": [ + 78, + 145, + 294, + 201 + ], + "lines": [ + { + "bbox": [ + 78, + 145, + 294, + 201 + ], + "spans": [ + { + "bbox": [ + 78, + 145, + 294, + 201 + ], + "type": "image", + "image_path": "1ae22fe54cf1a26ab0c81d70412c92e7f5a7b54e3aad1d5594b7cc90ff10c380.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 295, + 145, + 514, + 201 + ], + "blocks": [ + { + "bbox": [ + 295, + 145, + 514, + 201 + ], + "lines": [ + { + "bbox": [ + 295, + 145, + 514, + 201 + ], + "spans": [ + { + "bbox": [ + 295, + 145, + 514, + 201 + ], + "type": "image", + "image_path": "f67088808d9b7e3c34e8c68087f1e65df80a89a80c01d9738fd729e2a1eee587.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 78, + 201, + 294, + 258 + ], + "blocks": [ + { + "bbox": [ + 78, + 201, + 294, + 258 + ], + "lines": [ + { + "bbox": [ + 78, + 201, + 294, + 258 + ], + "spans": [ + { + "bbox": [ + 78, + 201, + 294, + 258 + ], + "type": "image", + "image_path": "1c91a1c210f22fe128dd1870378fa2d2cf2c1b2b85d512b35e667829d6ff7343.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 295, + 201, + 514, + 258 + ], + "blocks": [ + { + "bbox": [ + 295, + 201, + 514, + 258 + ], + "lines": [ + { + "bbox": [ + 295, + 201, + 514, + 258 + ], + "spans": [ + { + "bbox": [ + 295, + 201, + 514, + 258 + ], + "type": "image", + "image_path": "7005eb14ca927cf96ee26d9b01dc67f8579b547fc00590c6ce17feef77d70ea3.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 78, + 259, + 294, + 316 + ], + "blocks": [ + { + "bbox": [ + 78, + 259, + 294, + 316 + ], + "lines": [ + { + "bbox": [ + 78, + 259, + 294, + 316 + ], + "spans": [ + { + "bbox": [ + 78, + 259, + 294, + 316 + ], + "type": "image", + "image_path": "839da1ef7b7fd22a8ef10fa388a7eba5fdebacfa0ffdb8e734632ff4126e71bb.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 326, + 547, + 361 + ], + "lines": [ + { + "bbox": [ + 46, + 326, + 547, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 326, + 547, + 361 + ], + "type": "text", + "content": "Figure 5. Qualitative results on NL-Drive dataset. We transform four frames of the input point cloud to the same coordinate system, and the overall motion of the point cloud sequence is depicted in the first row. The following three rows compare the interpolation results, demonstrating that our approach is more accurate and robust, whereas interpolation results of PointINet [22] are noisier." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 295, + 259, + 514, + 316 + ], + "blocks": [ + { + "bbox": [ + 295, + 259, + 514, + 316 + ], + "lines": [ + { + "bbox": [ + 295, + 259, + 514, + 316 + ], + "spans": [ + { + "bbox": [ + 295, + 259, + 514, + 316 + ], + "type": "image", + "image_path": "cacc45ebce82dfec4a6aaea9d3873d73263916e4222f61450ef312ee15090644.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 50, + 430, + 289, + 515 + ], + "blocks": [ + { + "bbox": [ + 47, + 367, + 289, + 422 + ], + "lines": [ + { + "bbox": [ + 47, + 367, + 289, + 422 + ], + "spans": [ + { + "bbox": [ + 47, + 367, + 289, + 422 + ], + "type": "text", + "content": "Table 3. Quantitative comparison with other baseline methods for point cloud extrapolation on NL-Drive dataset. Here we adopt MoNet [21] and linear extrapolation results based on scene flow from PV-RAFT [39] and NSFP [16]. Frames 1-4 refer to the consecutive extrapolation frames after the last input frame." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 430, + 289, + 515 + ], + "lines": [ + { + "bbox": [ + 50, + 430, + 289, + 515 + ], + "spans": [ + { + "bbox": [ + 50, + 430, + 289, + 515 + ], + "type": "table", + "html": "
MethodsFrame 1Frame 2Frame 3Frame 4
CDEMDCDEMDCDEMDCD↓EMD↓
NSFP4.70209.965.45233.336.24254.726.61268.44
PV-RAFT2.05206.933.90248.946.55293.2710.02377.25
MoNet0.6681.900.96108.411.28135.961.37159.20
NeuralPCI0.7884.261.20108.431.61135.421.87156.64
", + "image_path": "548b788f3b7bd30acf34b98f56328c842b49ec38e57aa0cab8bd9f111a976074.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 526, + 287, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 526, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 46, + 526, + 287, + 612 + ], + "type": "text", + "content": "Our method surpasses MoNet in terms of EMD error in the last two frames and achieves suboptimal results in the other remaining frames. MoNet's RNN module predicts each frame recurrently, while our neural field lacks constraints on the future direction. Despite that, NeuralPCI still achieves comparable results, indicating its flexibility to accomplish both inter-/extra-polation in a unified framework." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 621, + 141, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 621, + 141, + 634 + ], + "spans": [ + { + "bbox": [ + 47, + 621, + 141, + 634 + ], + "type": "text", + "content": "4.5. Ablation Study" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 640, + 287, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 640, + 287, + 676 + ], + "spans": [ + { + "bbox": [ + 46, + 640, + 287, + 676 + ], + "type": "text", + "content": "Contributions of key modules in NeuralPCI, namely the multi-frame integration, NN-intp, EMD loss, smoothness loss, and network enhancement are shown in Table 4." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "content": "Multi-frame and NN-intp. We begin with a simple neural field with pair-frame input and optimized using only CD loss. With the multi-frame integration (ID 2&7), NeuralPCI" + } + ] + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 309, + 419, + 544, + 559 + ], + "blocks": [ + { + "bbox": [ + 305, + 367, + 547, + 412 + ], + "lines": [ + { + "bbox": [ + 305, + 367, + 547, + 412 + ], + "spans": [ + { + "bbox": [ + 305, + 367, + 547, + 412 + ], + "type": "text", + "content": "Table 4. Quantitative results of ablation studies on DHB dataset [47] " + }, + { + "bbox": [ + 305, + 367, + 547, + 412 + ], + "type": "inline_equation", + "content": "(\\times 10^{-3})" + }, + { + "bbox": [ + 305, + 367, + 547, + 412 + ], + "type": "text", + "content": " and NL-Drive dataset. Methods A~F denote CD loss, multi-frame integration, NN-intp, EMD loss, smoothness loss and network enhancement, respectively." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 419, + 544, + 559 + ], + "lines": [ + { + "bbox": [ + 309, + 419, + 544, + 559 + ], + "spans": [ + { + "bbox": [ + 309, + 419, + 544, + 559 + ], + "type": "table", + "html": "
DatasetsIDMethodsMetrics
ABCDEFCD ↓EMD ↓
DHB1✓0.655.24
2✓✓0.595.08
3✓✓✓0.574.28
4✓✓✓✓0.563.84
5✓✓✓✓✓0.543.68
NL-Drive6✓0.86114.31
7✓✓0.84112.03
8✓✓✓0.81104.71
9✓✓✓✓0.8299.25
10✓✓✓✓✓0.8097.03
", + "image_path": "0fab5fa2c0987a8f7f704e00c17db8451b04584da7dfb021063ef5c294775152.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 568, + 545, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 568, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 568, + 545, + 640 + ], + "type": "text", + "content": "gains " + }, + { + "bbox": [ + 304, + 568, + 545, + 640 + ], + "type": "inline_equation", + "content": "9.2\\%" + }, + { + "bbox": [ + 304, + 568, + 545, + 640 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 568, + 545, + 640 + ], + "type": "inline_equation", + "content": "3.1\\%" + }, + { + "bbox": [ + 304, + 568, + 545, + 640 + ], + "type": "text", + "content": " reductions in CD and EMD error on DHB dataset, and " + }, + { + "bbox": [ + 304, + 568, + 545, + 640 + ], + "type": "inline_equation", + "content": "2.3\\%" + }, + { + "bbox": [ + 304, + 568, + 545, + 640 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 568, + 545, + 640 + ], + "type": "inline_equation", + "content": "2.0\\%" + }, + { + "bbox": [ + 304, + 568, + 545, + 640 + ], + "type": "text", + "content": " on NL-Drive dataset. By adopting the nearest neighbor in the time domain as the reference frame (ID 3&8), the long-term error growth is effectively reduced, and another EMD improvement of " + }, + { + "bbox": [ + 304, + 568, + 545, + 640 + ], + "type": "inline_equation", + "content": "15.7\\%" + }, + { + "bbox": [ + 304, + 568, + 545, + 640 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 568, + 545, + 640 + ], + "type": "inline_equation", + "content": "6.5\\%" + }, + { + "bbox": [ + 304, + 568, + 545, + 640 + ], + "type": "text", + "content": " is achieved on the respective datasets." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 641, + 546, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 546, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 546, + 712 + ], + "type": "text", + "content": "EMD and smoothness loss. To prevent the network from overfitting on a single CD metric, we introduce additional loss terms to regulate the output. The extra EMD loss and smoothness loss help NeuralPCI achieve a significant reduction in EMD metric error, namely " + }, + { + "bbox": [ + 304, + 641, + 546, + 712 + ], + "type": "inline_equation", + "content": "10.3\\%" + }, + { + "bbox": [ + 304, + 641, + 546, + 712 + ], + "type": "text", + "content": " (ID 4) on DHB dataset and " + }, + { + "bbox": [ + 304, + 641, + 546, + 712 + ], + "type": "inline_equation", + "content": "5.2\\%" + }, + { + "bbox": [ + 304, + 641, + 546, + 712 + ], + "type": "text", + "content": " (ID 9) on NL-Drive dataset." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "type": "text", + "content": "915" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 72, + 521, + 247 + ], + "blocks": [ + { + "bbox": [ + 70, + 72, + 521, + 247 + ], + "lines": [ + { + "bbox": [ + 70, + 72, + 521, + 247 + ], + "spans": [ + { + "bbox": [ + 70, + 72, + 521, + 247 + ], + "type": "image", + "image_path": "dd2dbfa7b79527dbbda6b6b27f735acdfd5e1a483613edacab5357c2e9429bed.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 255, + 547, + 289 + ], + "lines": [ + { + "bbox": [ + 46, + 255, + 547, + 289 + ], + "spans": [ + { + "bbox": [ + 46, + 255, + 547, + 289 + ], + "type": "text", + "content": "Figure 6. Visual results of NeuralPCI based auto-labeling. We use NeuralPCI to take the labeled keyframe point clouds as input, output the interpolation results, and automatically assign labels to the intermediate frames. The second row shows the results of the auto-labeling, which intuitively achieves high labeling accuracy." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 50, + 296, + 285, + 373 + ], + "blocks": [ + { + "bbox": [ + 50, + 296, + 285, + 373 + ], + "lines": [ + { + "bbox": [ + 50, + 296, + 285, + 373 + ], + "spans": [ + { + "bbox": [ + 50, + 296, + 285, + 373 + ], + "type": "image", + "image_path": "0f853ba7e4a362fde33e0037e810a6448298bb7f7c49f653db71e2e9d6fc7422.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 382, + 287, + 404 + ], + "lines": [ + { + "bbox": [ + 46, + 382, + 287, + 404 + ], + "spans": [ + { + "bbox": [ + 46, + 382, + 287, + 404 + ], + "type": "text", + "content": "Figure 7. The process of point clouds morphing between the airplane and the chair samples in ModelNet40 [42] using NeuralPCI." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 413, + 287, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 413, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 413, + 287, + 533 + ], + "type": "text", + "content": "Network enhancement. The impact of network structure is future investigated. Previously, NeuralPCI utilized an 8-layer 256-hidden-unit MLP with a ReLU activation function and received the direct spatio-temporal coordinate as input. Here, we introduce a sinusoidal function-based position encoding, switch to the LeakyReLU activation function, and increase the width of the MLP. In the end, the network enhancement brings another improvement of " + }, + { + "bbox": [ + 46, + 413, + 287, + 533 + ], + "type": "inline_equation", + "content": "3.6\\%" + }, + { + "bbox": [ + 46, + 413, + 287, + 533 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 413, + 287, + 533 + ], + "type": "inline_equation", + "content": "4.2\\%" + }, + { + "bbox": [ + 46, + 413, + 287, + 533 + ], + "type": "text", + "content": " (ID 5) on DHB dataset and " + }, + { + "bbox": [ + 46, + 413, + 287, + 533 + ], + "type": "inline_equation", + "content": "2.4\\%" + }, + { + "bbox": [ + 46, + 413, + 287, + 533 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 413, + 287, + 533 + ], + "type": "inline_equation", + "content": "2.2\\%" + }, + { + "bbox": [ + 46, + 413, + 287, + 533 + ], + "type": "text", + "content": " (ID 10) on NL-Drive dataset." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 540, + 129, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 540, + 129, + 553 + ], + "spans": [ + { + "bbox": [ + 47, + 540, + 129, + 553 + ], + "type": "text", + "content": "4.6. Applications" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "content": "Auto-labeling. The annotated point clouds with perpoint segmentation GT in Nuscenes [2] are at " + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "inline_equation", + "content": "2\\mathrm{Hz}" + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "content": ", only one tenth of the dataset. Even so, the labeling workload was laborious and enormous with a total of 1,400,000,000 points. Here, we explore the potential of NeuralPCI to generate keyframe-based interpolation results and assign point-wise labels to unannotated intermediate frames, which shows remarkable capability as depicted in Fig. 6. We use four keyframes as input to predict intermediate point clouds, so the predicted outputs could inherit the corresponding keyframe labels in order. Then the kNN algorithm is applied between the unlabeled intermediate frame and the labeled predicted frame to annotate each ground-truth point." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 298, + 547, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 298, + 547, + 443 + ], + "spans": [ + { + "bbox": [ + 304, + 298, + 547, + 443 + ], + "type": "text", + "content": "Point cloud morphing. In addition to the conventional meaning of nonlinear motion under indoor and outdoor scenes, the transformation relationship across point clouds with totally different topological shapes is also needed. This transformation, i.e., point cloud morphing, is of interest for computer graphics simulation and data enhancement. We deploy NeuralPCI to output interpolation point clouds between two different classes of object point clouds under the ModelNet40 [42] dataset to establish the process of point cloud morphing. As illustrated in Fig. 7, our method enables a more natural and smooth transformation between different shapes compared to random sampling." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 452, + 379, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 452, + 379, + 464 + ], + "spans": [ + { + "bbox": [ + 305, + 452, + 379, + 464 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 472, + 546, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 546, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 546, + 639 + ], + "type": "text", + "content": "In this paper, we redefine the input domain of the point cloud interpolation task as multiple consecutive frames instead of the two consecutive frames used in previous works, which increases the receptive field of the time domain. To achieve that, we presented NeuralPCI, a 4D spatio-temporal neural field for 3D point cloud interpolation that is able to implicitly integrate multi-frame information to handle nonlinear large motions. Our approach achieves state-of-the-art results in both indoor and outdoor datasets. Since neuralfield-based methods are optimized at runtime, the application of our method is limited in terms of real-time efficiency. Based on NeuralPCI, further development can be considered in the future by improving its real-time performance and generalization over unknown samples." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "text", + "content": "Acknowledgments This work is supported by Shanghai Municipal Science and Technology Major Project (No.2018SHZDZX01), ZJ Lab, and Shanghai Center for Brain Science and Brain-Inspired Technology, and the Shanghai Rising Star Program (No.21QC1400900) and Tongji-Westwell Autonomous Vehicle Joint Lab Project." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "916" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 146 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 146 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 146 + ], + "type": "text", + "content": "[1] Wenbo Bao, Wei-Sheng Lai, Chao Ma, Xiaoyun Zhang, Zhiyong Gao, and Ming-Hsuan Yang. Depth-aware video frame interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3703-3712, 2019. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 148, + 288, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 288, + 215 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 288, + 215 + ], + "type": "text", + "content": "[2] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 5, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 216, + 288, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 216, + 288, + 281 + ], + "spans": [ + { + "bbox": [ + 53, + 216, + 288, + 281 + ], + "type": "text", + "content": "[3] Ming-Fang Chang, John Lambert, Patsorn Sangkloy, Jagjeet Singh, Slawomir Bak, Andrew Hartnett, De Wang, Peter Carr, Simon Lucey, Deva Ramanan, et al. Argoverse: 3d tracking and forecasting with rich maps. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8748-8757, 2019. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 283, + 287, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 283, + 287, + 338 + ], + "spans": [ + { + "bbox": [ + 53, + 283, + 287, + 338 + ], + "type": "text", + "content": "[4] Xieyuanli Chen, Shijie Li, Benedikt Mersch, Louis Wiesmann, Jürgen Gall, Jens Behley, and Cyril Stachniss. Moving object segmentation in 3d lidar data: A learning-based approach exploiting sequential data. IEEE Robotics and Automation Letters, 6(4):6529-6536, 2021. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 339, + 287, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 339, + 287, + 372 + ], + "spans": [ + { + "bbox": [ + 53, + 339, + 287, + 372 + ], + "type": "text", + "content": "[5] Zhiqin Chen. IM-NET: Learning implicit fields for generative shape modeling. PhD thesis, Applied Sciences: School of Computing Science, 2019. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 373, + 287, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 373, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 53, + 373, + 287, + 430 + ], + "type": "text", + "content": "[6] Zhixiang Chi, Rasoul Mohammadi Nasiri, Zheng Liu, Juwei Lu, Jin Tang, and Konstantinos N Plataniotis. All at once: Temporally adaptive multi-frame interpolation with advanced motion modeling. In European Conference on Computer Vision, pages 107-123. Springer, 2020. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 431, + 287, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 431, + 287, + 485 + ], + "spans": [ + { + "bbox": [ + 53, + 431, + 287, + 485 + ], + "type": "text", + "content": "[7] Yilun Du, Yinan Zhang, Hong-Xing Yu, Joshua B Tenenbaum, and Jiajun Wu. Neural radiance flow for 4d view synthesis and video processing. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 14304-14314. IEEE Computer Society, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 487, + 288, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 487, + 288, + 542 + ], + "spans": [ + { + "bbox": [ + 53, + 487, + 288, + 542 + ], + "type": "text", + "content": "[8] Saikat Dutta, Arulkumar Subramaniam, and Anurag Mittal. Non-linear motion estimation for video frame interpolation using space-time convolutions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1726-1731, 2022. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 544, + 287, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 544, + 287, + 588 + ], + "spans": [ + { + "bbox": [ + 53, + 544, + 287, + 588 + ], + "type": "text", + "content": "[9] Haoqiang Fan, Hao Su, and Leonidas J Guibas. A point set generation network for 3d object reconstruction from a single image. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 605-613, 2017. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 590, + 287, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 633 + ], + "type": "text", + "content": "[10] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5712-5721, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "text", + "content": "[11] Daniel Garrido, Rui Rodrigues, A Augusto Sousa, Joao Jacob, and Daniel Castro Silva. Point cloud interaction and manipulation in virtual reality. In 2021 5th International Conference on Artificial Intelligence and Virtual Reality (AIVR), pages 15–20, 2021. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[12] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "suite. In 2012 IEEE conference on computer vision and pattern recognition, pages 3354-3361. IEEE, 2012. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 95, + 545, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 95, + 545, + 160 + ], + "spans": [ + { + "bbox": [ + 308, + 95, + 545, + 160 + ], + "type": "text", + "content": "[13] Huaizu Jiang, Deqing Sun, Varun Jampani, Ming-Hsuan Yang, Erik Learned-Miller, and Jan Kautz. Super slomo: High quality estimation of multiple intermediate frames for video interpolation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9000-9008, 2018. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 162, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 162, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 308, + 162, + 545, + 205 + ], + "type": "text", + "content": "[14] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. Advances in Neural Information Processing Systems, 34:852-863, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 206, + 545, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 206, + 545, + 237 + ], + "spans": [ + { + "bbox": [ + 308, + 206, + 545, + 237 + ], + "type": "text", + "content": "[15] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 239, + 545, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 239, + 545, + 271 + ], + "spans": [ + { + "bbox": [ + 308, + 239, + 545, + 271 + ], + "type": "text", + "content": "[16] Xueqian Li, Jhony Kaesemodel Pontes, and Simon Lucey. Neural scene flow prior. Advances in Neural Information Processing Systems, 34:7838-7851, 2021. 2, 5, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 272, + 545, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 272, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 308, + 272, + 545, + 326 + ], + "type": "text", + "content": "[17] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6498-6508, 2021. 2, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 327, + 545, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 327, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 308, + 327, + 545, + 380 + ], + "type": "text", + "content": "[18] Haojie Liu, Kang Liao, Chunyu Lin, Yao Zhao, and Yulan Guo. Pseudo-lidar point cloud interpolation based on 3d motion representation and spatial supervision. IEEE Transactions on Intelligent Transportation Systems, 23(7):6379-6389, 2021. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 383, + 545, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 383, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 308, + 383, + 545, + 415 + ], + "type": "text", + "content": "[19] Haojie Liu, Kang Liao, Chunyu Lin, Yao Zhao, and Meiqin Liu. Plin: A network for pseudo-lidar point cloud interpolation. Sensors, 20(6):1573, 2020. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 416, + 545, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 416, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 308, + 416, + 545, + 459 + ], + "type": "text", + "content": "[20] Yihao Liu, Liangbin Xie, Li Siyao, Wenxiu Sun, Yu Qiao, and Chao Dong. Enhanced quadratic video interpolation. In European Conference on Computer Vision, pages 41-56. Springer, 2020. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 460, + 545, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 460, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 308, + 460, + 545, + 503 + ], + "type": "text", + "content": "[21] Fan Lu, Guang Chen, Zhijun Li, Lijun Zhang, Yinlong Liu, Sanqing Qu, and Alois Knoll. Monet: Motion-based point cloud prediction network. IEEE Transactions on Intelligent Transportation Systems, 2021. 5, 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 504, + 545, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 504, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 308, + 504, + 545, + 557 + ], + "type": "text", + "content": "[22] Fan Lu, Guang Chen, Sanqing Qu, Zhijun Li, Yinlong Liu, and Alois Knoll. Pointinet: Point cloud frame interpolation network. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 2251-2259, 2021. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 559, + 545, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 559, + 545, + 603 + ], + "spans": [ + { + "bbox": [ + 308, + 559, + 545, + 603 + ], + "type": "text", + "content": "[23] Benedikt Mersch, Xieyuanli Chen, Jens Behley, and Cyril Stachniss. Self-supervised point cloud prediction using 3d spatio-temporal convolutional networks. In Conference on Robot Learning, pages 1444-1454. PMLR, 2022. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 604, + 545, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 604, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 308, + 604, + 545, + 658 + ], + "type": "text", + "content": "[24] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4460-4470, 2019. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "text", + "content": "[25] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2, 3" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "917" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[26] Junheum Park, Keunsoo Ko, Chul Lee, and Chang-Su Kim. Bmbc: Bilateral motion estimation with bilateral cost volume for video interpolation. In European Conference on Computer Vision, pages 109-125. Springer, 2020. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 172 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 172 + ], + "type": "text", + "content": "[27] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 165-174, 2019. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 173, + 287, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 173, + 287, + 227 + ], + "spans": [ + { + "bbox": [ + 48, + 173, + 287, + 227 + ], + "type": "text", + "content": "[28] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865-5874, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 228, + 287, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 228, + 287, + 292 + ], + "spans": [ + { + "bbox": [ + 48, + 228, + 287, + 292 + ], + "type": "text", + "content": "[29] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 293, + 287, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 293, + 287, + 348 + ], + "spans": [ + { + "bbox": [ + 48, + 293, + 287, + 348 + ], + "type": "text", + "content": "[30] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10318–10327, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 349, + 287, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 349, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 48, + 349, + 287, + 403 + ], + "type": "text", + "content": "[31] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 404, + 288, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 404, + 288, + 459 + ], + "spans": [ + { + "bbox": [ + 48, + 404, + 288, + 459 + ], + "type": "text", + "content": "[32] Charles R Qi, Yin Zhou, Mahyar Najibi, Pei Sun, Khoa Vo, Boyang Deng, and Dragomir Anguelov. Offboard 3d object detection from point cloud sequences. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6134-6144, 2021. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 460, + 287, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 460, + 287, + 524 + ], + "spans": [ + { + "bbox": [ + 48, + 460, + 287, + 524 + ], + "type": "text", + "content": "[33] Fitssum A Reda, Deqing Sun, Aysegul Dundar, Mohammad Shoeybi, Guilin Liu, Kevin J Shih, Andrew Tao, Jan Kautz, and Bryan Catanzaro. Unsupervised video interpolation using cycle consistency. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 892-900, 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 525, + 287, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 287, + 568 + ], + "type": "text", + "content": "[34] Yossi Rubner, Carlo Tomasi, and Leonidas J Guibas. The earth mover's distance as a metric for image retrieval. International journal of computer vision, 40(2):99-121, 2000. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 570, + 287, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 570, + 287, + 623 + ], + "spans": [ + { + "bbox": [ + 48, + 570, + 287, + 623 + ], + "type": "text", + "content": "[35] Zhihao Shi, Xiangyu Xu, Xiaohong Liu, Jun Chen, and Ming-Hsuan Yang. Video frame interpolation transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17482-17491, 2022. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 624, + 287, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 690 + ], + "type": "text", + "content": "[36] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Christoph Lassner, and Christian Theobalt. Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12959-12970, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[37] Chaoyang Wang, Xueqian Li, Jhony Kaesemodel Pontes, and Simon Lucey. Neural prior for trajectory estimation." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 706 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6532-6542, 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 308, + 96, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 150 + ], + "type": "text", + "content": "[38] Haiyan Wang, Liang Yang, Xuejian Rong, Jinglun Feng, and Yingli Tian. Self-supervised 4d spatio-temporal feature learning via order prediction of sequential point cloud clips. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3762-3771, 2021. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 152, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 152, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 308, + 152, + 545, + 205 + ], + "type": "text", + "content": "[39] Yi Wei, Ziyi Wang, Yongming Rao, Jiwen Lu, and Jie Zhou. Pv-raft: point-voxel correlation fields for scene flow estimation of point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6954-6963, 2021. 5, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 207, + 545, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 207, + 545, + 273 + ], + "spans": [ + { + "bbox": [ + 308, + 207, + 545, + 273 + ], + "type": "text", + "content": "[40] Xinshuo Weng, Junyu Nan, Kuan-Hui Lee, Rowan McAllister, Adrien Gaidon, Nicholas Rhinehart, and Kris M Kitani. S2net: Stochastic sequential pointcloud forecasting. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXVII, pages 549-564. Springer, 2022. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 274, + 545, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 274, + 545, + 328 + ], + "spans": [ + { + "bbox": [ + 308, + 274, + 545, + 328 + ], + "type": "text", + "content": "[41] Xinshuo Weng, Jianren Wang, Sergey Levine, Kris Kitani, and Nicholas Rhinehart. Inverting the pose forecasting pipeline with spf2: Sequential pointcloud forecasting for sequential pose forecasting. In Conference on robot learning, pages 11-20. PMLR, 2021. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 330, + 545, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 330, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 308, + 330, + 545, + 384 + ], + "type": "text", + "content": "[42] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1912-1920, 2015. 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 385, + 545, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 385, + 545, + 439 + ], + "spans": [ + { + "bbox": [ + 308, + 385, + 545, + 439 + ], + "type": "text", + "content": "[43] Wenqi Xian, Jia-Bin Huang, Johannes Kopf, and Changil Kim. Space-time neural irradiance fields for free-viewpoint video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9421-9431, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 441, + 545, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 441, + 545, + 495 + ], + "spans": [ + { + "bbox": [ + 308, + 441, + 545, + 495 + ], + "type": "text", + "content": "[44] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond. In Computer Graphics Forum, volume 41, pages 641-676. Wiley Online Library, 2022. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 497, + 545, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 497, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 308, + 497, + 545, + 529 + ], + "type": "text", + "content": "[45] Xiangyu Xu, Li Siyao, Wenxiu Sun, Qian Yin, and Ming-Hsuan Yang. Quadratic video interpolation. Advances in Neural Information Processing Systems, 32, 2019. 1, 2, 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 529, + 545, + 573 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 529, + 545, + 573 + ], + "spans": [ + { + "bbox": [ + 308, + 529, + 545, + 573 + ], + "type": "text", + "content": "[46] Bin Yang, Min Bai, Ming Liang, Wenyuan Zeng, and Raquel Urtasun. Auto4d: Learning to label 4d objects from sequential point clouds. arXiv preprint arXiv:2101.06586, 2021. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 574, + 545, + 628 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 574, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 308, + 574, + 545, + 628 + ], + "type": "text", + "content": "[47] Yiming Zeng, Yue Qian, Qijian Zhang, Junhui Hou, Yixuan Yuan, and Ying He. Idea-net: Dynamic 3d point cloud interpolation via deep embedding alignment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6338-6347, 2022. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 631, + 545, + 662 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 631, + 545, + 662 + ], + "spans": [ + { + "bbox": [ + 308, + 631, + 545, + 662 + ], + "type": "text", + "content": "[48] Haoxian Zhang, Ronggang Wang, and Yang Zhao. Multiframe pyramid refinement network for video frame interpolation. IEEE Access, 7:130610-130621, 2019. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 664, + 545, + 706 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 664, + 545, + 706 + ], + "spans": [ + { + "bbox": [ + 308, + 664, + 545, + 706 + ], + "type": "text", + "content": "[49] Zihao Zhang, Lei Hu, Xiaoming Deng, and Shihong Xia. Sequential 3d human pose estimation using adaptive point cloud sampling strategy. In *IJCAI*, pages 1330–1337, 2021. 1" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "918" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/NeuralUDF_ Learning Unsigned Distance Fields for Multi-View Reconstruction of Surfaces With Arbitrary Topologies/cda2b21c-85ea-47d7-952f-4f7b2568e13a_content_list.json b/2023/NeuralUDF_ Learning Unsigned Distance Fields for Multi-View Reconstruction of Surfaces With Arbitrary Topologies/cda2b21c-85ea-47d7-952f-4f7b2568e13a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..956c424fefa82c1d32b616b5d0f88a213a0e7e81 --- /dev/null +++ b/2023/NeuralUDF_ Learning Unsigned Distance Fields for Multi-View Reconstruction of Surfaces With Arbitrary Topologies/cda2b21c-85ea-47d7-952f-4f7b2568e13a_content_list.json @@ -0,0 +1,2089 @@ +[ + { + "type": "text", + "text": "NeuralUDF: Learning Unsigned Distance Fields for Multi-view Reconstruction of Surfaces with Arbitrary Topologies", + "text_level": 1, + "bbox": [ + 137, + 130, + 828, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xiaoxiao Long $^{1,2\\dagger}$ Cheng Lin $^{2*}$ Lingjie Liu $^{3}$ Yuan Liu $^{1}$ Peng Wang $^{1}$ Christian Theobalt $^{3}$ Taku Komura $^{1}$ Wenping Wang $^{4*}$ $^{1}$ The University of Hong Kong $^{2}$ Tencent Games \n $^{3}$ Max Planck Institute for Informatics $^{4}$ Texas A&M University", + "bbox": [ + 174, + 200, + 792, + 281 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 313, + 313, + 329 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We present a novel method, called NeuralUDF, for reconstructing surfaces with arbitrary topologies from 2D images via volume rendering. Recent advances in neural rendering based reconstruction have achieved compelling results. However, these methods are limited to objects with closed surfaces since they adopt Signed Distance Function (SDF) as surface representation which requires the target shape to be divided into inside and outside. In this paper, we propose to represent surfaces as the Unsigned Distance Function (UDF) and develop a new volume rendering scheme to learn the neural UDF representation. Specifically, a new density function that correlates the property of UDF with the volume rendering scheme is introduced for robust optimization of the UDF fields. Experiments on the DTU and DeepFashion3D datasets show that our method not only enables high-quality reconstruction of non-closed shapes with complex typologies, but also achieves comparable performance to the SDF based methods on the reconstruction of closed surfaces. Visit our project page at https://www.xxlong.site/NeuralUDF/.", + "bbox": [ + 75, + 344, + 473, + 648 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 657, + 209, + 674 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Reconstructing high-quality surfaces from multi-view images is a long-standing problem in computer vision and computer graphics. Neural implicit fields have become an emerging trend in recent advances due to its superior capability of representing surfaces of complex geometry. NeRF [35] and its variants [2, 26, 27, 36, 40, 52] have recently achieved compelling results in novel view synthesis. For each point in the 3D space, NeRF-based methods learn two neural implicit functions based on volume rendering: a volume density function and a view-dependent color function. Despite their successes in novel view synthesis, NeRF-based methods still struggle to faithfully reconstruct the ac", + "bbox": [ + 75, + 683, + 468, + 864 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/708d98f8ae5d6eca6b29dd87cb20d25df7f26c6f32520a4926a6f6a1c7b52a11.jpg", + "image_caption": [ + "Figure 1. We show three groups of multi-view reconstruction results generated by our proposed NeuralUDF and NeuS [46] respectively. Our method is able to faithfully reconstruct the high-quality geometries for both the closed and open surfaces, while NeuS can only model shapes as closed surfaces, thus leading to inconsistent typologies and erroneous geometries." + ], + "image_footnote": [], + "bbox": [ + 511, + 313, + 877, + 571 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "curate scene geometry from multi-view inputs, because of the difficulty in extracting high-quality surfaces from the representation of volume density.", + "bbox": [ + 496, + 671, + 893, + 715 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "VolSDF [49] and NeuS [46] incorporate the Signed Distance Function (SDF) into volume rendering to facilitate high-quality surface reconstruction in a NeRF framework. However, as a continuous function with clearly defined inside/outside, SDF is limited to modeling only closed watertight surfaces. Although there have been efforts to modify the SDF representation by learning an additional truncation function (e.g., 3PSDF [5], TSDF [44]), their surface representations are still built upon the definition of SDF, thus they are not suitable for representing complex topologies.", + "bbox": [ + 496, + 717, + 890, + 869 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We therefore propose to employ the Unsigned Distance Function (UDF) to represent surfaces in volume rendering.", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding authors.", + "bbox": [ + 99, + 875, + 233, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "This work was conducted during an internship at Tencent Games.", + "bbox": [ + 99, + 887, + 452, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "20834", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With a surface represented by its zero level set without signs, UDF is a unified representation with higher-degree of freedom for both closed and open surfaces, thus making it possible to reconstruct shapes with arbitrary topologies.", + "bbox": [ + 75, + 90, + 468, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "There are two major challenges in learning a neural UDF field by volume rendering. First, UDF is not occlusion-aware, while the formulation of volume rendering with distance fields requires to estimate surface occlusion for points. Considering a camera ray intersecting with a surface, SDF assumes that the surface is closed and distinguishes the inside/outside with signs, so that a negative SDF value on the ray clearly indicates an occluded point inside the surface. In contrast, UDF does not impose the closed surface assumption and always gives non-negative values along the ray. Hence, the UDF value of a point alone cannot be used to infer occlusion.", + "bbox": [ + 75, + 152, + 470, + 330 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The second challenge is that the UDF is not differentiable at its zero-level sets. The non-differentiability at the zero-level sets imposes barriers in the learning of UDF field. The gradients are ill-defined near the iso-surface, leading to difficulty in optimization. Consequently, the distance field surrounding the iso-surface is not accurate and the exact zero-level set of a learned UDF cannot be identified in a stable manner.", + "bbox": [ + 75, + 332, + 468, + 450 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we present a novel method for learning neural UDF fields by volume rendering. We introduce a density function that correlates the property of the UDF representation with the volume rendering process, which effectively tackles the aforementioned challenges induced by the unsigned representation and enables robust learning of surfaces with arbitrary topologies. Experiments on DTU [15] and DeepFashion3D [53] datasets show that our method not only enables the high-quality reconstruction of non-closed shapes with complex topologies, but also achieves comparable performance to the SDF based methods on the reconstruction of closed surfaces.", + "bbox": [ + 75, + 454, + 470, + 633 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions can be summarized as:", + "bbox": [ + 96, + 635, + 369, + 648 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We incorporate the UDF into volume rendering, which extends the representation of the underlying geometry of neural radiance fields.", + "- We introduce an effective density function that correlates the property of the UDF with the volume rendering process, thus enabling robust optimization of the distance fields.", + "- Our method achieves SOTA results for reconstructing high-quality surfaces with various topologies (closed or open) using the UDF in volume rendering." + ], + "bbox": [ + 94, + 648, + 467, + 801 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 814, + 218, + 830 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Classic multi-view 3D reconstruction. 3D reconstruction from multiple images is a fundamental problem in computer vision. Traditional reconstruction methods usually adopt discrete representations, such as voxel grids [16,", + "bbox": [ + 75, + 839, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "17, 19, 23, 43, 44], 3D point clouds [10, 25], and depth maps [4, 11, 13, 29, 31, 32, 42, 45, 47, 48]. The depth map representation is flexible and appropriate for parallel computation, so it is the most widely used representation. The depth map based methods first compute depth maps by analyzing the correlation between multiple input images, and then fuse the depth maps into a point cloud, which is then further processed into a mesh via meshing algorithms, such as Poisson surface reconstruction [20], ball-pivoting [3] and Delaunay triangulation [24]. However, the traditional methods are sensitive to reflections, image noises, and varying illuminations, thus usually yielding incomplete and noisy reconstructions.", + "bbox": [ + 496, + 90, + 890, + 286 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Neural rendering based surface reconstruction. The recent advances on neural implicit representations have achieved promising results in multi-view 3D reconstruction [8,18,22,28,37,38,46,49-51]. These methods typically represent the surfaces to be reconstructed as Signed Distance Function (SDF), and utilize surface rendering or volume rendering to render the pixel colors of the input images for optimizing the SDF fields. The surface rendering based methods [37, 50] require extra masks for optimization and cannot handle objects with complex structures. The volume rendering based works [8, 38, 46, 49, 51] get rid of the extra masks and achieve more accurate reconstructions. However, these methods adopt SDF as representation so they are limited to closed surfaces and their performance degrade a lot for non-closed surfaces.", + "bbox": [ + 496, + 287, + 890, + 512 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Surface representations. There are various representations to model 3D geometric surfaces. NeRF [35] uses a density field to represent the underlying geometry of a target scene. Although NeRF achieves huge successes in novel view synthesis, it's difficult to extract high-quality geometry from such a density representation. Signed Distance Function (SDF) is widely used for geometry modeling and surface reconstruction [1,6,8,12,33,34,38,39,46,49,51], but SDF is only applicable to objects with closed surfaces. To deal with open surfaces, there are some efforts to learn an additional truncation function (e.g. 3PSDF [5], TSDF [44]). However, their representations are just built on the modification of SDF, and thus are still subject to flexibility of SDF and struggle to model complex surfaces. NDF [7] uses a neural UDF field to reconstruct surfaces, but they explicitly rely on 3D point clouds for supervising the optimization.", + "bbox": [ + 496, + 513, + 890, + 755 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 768, + 589, + 784 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our goal is to reconstruct a surface $S$ from a set of posed input images $\\{\\mathcal{I}_k\\}$ of a 3D model. We represent the surface as the zero level set of an Unsigned Distance Function (UDF), and propose a novel volume rendering scheme to learn the UDF. We will introduce our method in three parts: 1) We discuss the relationship between the unsigned distance function and the density function of the volume ren", + "bbox": [ + 496, + 794, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "20835", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0c34ec97745c9acfd781109e16531c66efe7b61efd60c57536fdb2d5a6d23a70.jpg", + "image_caption": [ + "(a) Biased density function", + "Figure 2. Illustration of (a) biased density function and (b) unbiased density function for UDF. Suppose that the ray has one single intersection point $t^*$ with the surface. (a) If the density function is modeled as a bell-shaped probability distribution, the weight function attains a local maximum before $t^*$ . (b) If the density function is modeled as a monotonically increasing function, the weight function can attain its local maximum at $t^*$ ." + ], + "image_footnote": [], + "bbox": [ + 86, + 88, + 269, + 219 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/bf32a3758bdcbf8280746fc26f7131225aca1456213c3dee09f76453afc9bf86.jpg", + "image_caption": [ + "(b) Unbiased density function" + ], + "image_footnote": [], + "bbox": [ + 274, + 103, + 459, + 219 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "dering equation; 2) To derive a density function induced by the UDF, we propose a differentiable indicator function to transform the UDF into its associated density field; 3) In addition to the loss terms commonly used in the prior works, we introduce an iso-surface regularizer to improve the stability of the UDF field near the zero level set. This UDF-based volume rendering scheme enables robust optimization of NeuralUDF for reconstructing both closed surfaces and open surfaces.", + "bbox": [ + 75, + 342, + 472, + 479 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Volume Rendering", + "text_level": 1, + "bbox": [ + 76, + 487, + 256, + 503 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Scene representation. Our method NeuralUDF represents the scene to be reconstructed by two neural implicit functions: 1) an unsigned distance function $f_{u}(\\mathbf{x})$ that maps a 3D point $\\mathbf{x} \\in \\mathbb{R}^3$ to its unsigned distance to the closest object surface; 2) a view-dependent color function $c(\\mathbf{x},\\mathbf{v})$ that encodes the color associated with a point $\\mathbf{x} \\in \\mathbb{R}^3$ and a viewing direction $\\mathbf{v} \\in \\mathbb{R}^3$ .", + "bbox": [ + 75, + 511, + 468, + 617 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Volume rendering. We consider a ray $\\mathbf{r}$ emanating from the camera center $\\mathbf{o} \\in \\mathbb{R}^3$ in direction $\\mathbf{v} \\in \\mathbb{R}^3$ ( $\\| \\mathbf{v} \\| = 1$ ), which is parameterized by $\\{\\mathbf{r}(t) = \\mathbf{o} + t\\mathbf{v} \\mid t \\geq 0\\}$ . Volume rendering is used to calculate the color $C$ of the pixel corresponding to the ray $\\mathbf{r}$ . The accumulated color of the camera ray $\\mathbf{r}$ is given by $C(\\mathbf{r}) = \\int_0^{+\\infty} T(t) \\sigma(t) \\mathbf{c}(\\mathbf{r}(t), \\mathbf{v}) \\, \\mathrm{d}t$ , where $\\sigma(t)$ is the volume density, $T(t) = \\exp \\left(-\\int_0^t \\sigma(u) \\, \\mathrm{d}u\\right)$ denotes the accumulated transmittance along the ray, and the product of $T(t) \\sigma(t)$ is the weight function of a point, denoted as $w(t)$ , i.e., $w(t) = T(t) \\sigma(t)$ .", + "bbox": [ + 75, + 617, + 470, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Modeling Density with UDF", + "text_level": 1, + "bbox": [ + 76, + 786, + 328, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the NeRF framework, the density $\\sigma(t)$ is predicted by an MLP, where the points near the surfaces having larger density values and the points far from the surfaces having smaller densities. In order to apply volume rendering to learning distance field $f$ , we need to define a mapping $\\Omega$ to transform the distance function $f(\\mathbf{r}(t))$ into a density field", + "bbox": [ + 75, + 809, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$\\sigma (t)$ , that is $\\sigma (t) = \\Omega (f(\\mathbf{r}(t)))$", + "bbox": [ + 498, + 90, + 712, + 106 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unbiased volume density function. Inspired by NeRF, we may consider defining the density function $\\sigma(t) = \\Omega(f(\\mathbf{r}(t)))$ as a bell-shaped distribution model centered at zero, so that the surface points intersecting with a ray (UDF values are zero) have the maximum density values. However, as illustrated in Fig 2, if the mapping function is straightforwardly defined in this way, a noticeable bias will be induced in the weight function $w(t)$ since its local maximum is not at the intersection point $t^*$ where the ray intersects with the surfaces. This can result in conspicuous inaccuracy in the reconstructed surfaces, which is demonstrated in [46].", + "bbox": [ + 496, + 106, + 890, + 286 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To define an appropriate mapping, we need to first review the relationship between the density function $\\sigma(t)$ and a distance field $f(t)$ . Suppose that there is only one single intersection of the ray and a surface. It is shown in NeuS [46] that, to make the weight function $w(t)$ unbiased, the density function should be in the form of a monotonically increasing function, which attains the maximum at an infinite point rather than at the intersection point.", + "bbox": [ + 496, + 287, + 892, + 407 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Accordingly, NeuS [46] chooses to use a Sigmoid function $\\Phi_{\\kappa}(x) = (1 + e^{-\\kappa x})^{-1}$ to achieve the goal. The accumulated transmittance $T(t)$ is then defined as $T(t) = \\Phi_{\\kappa}(f_s(t))$ , where $f_{s}(t)$ is the signed distance function (SDF). The density function $\\sigma_{s}(t)$ induced by SDF is formulated as", + "bbox": [ + 498, + 407, + 890, + 500 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\sigma_ {s} (t) = \\Omega_ {s} \\left(f _ {s} (\\mathbf {r} (t))\\right) = \\kappa | \\cos (\\theta) | \\left(1 - \\Phi_ {\\kappa} \\left(f _ {s} (\\mathbf {r} (t))\\right)\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 512, + 890, + 542 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\kappa$ is the parameter of Sigmoid function $\\Phi_{\\kappa}(x)$ , and $\\theta$ is the angle between the view direction $\\mathbf{v}$ and the surface normal $\\mathbf{n}$ .", + "bbox": [ + 496, + 542, + 890, + 585 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unbiased volume density for UDF. However, the unbiased density formulation induced by SDF [46] cannot be directly applied to UDF. The formulation of UDF near the surface can be given by", + "bbox": [ + 496, + 588, + 890, + 650 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nf _ {u} (\\mathbf {r} (t)) = \\left\\{ \\begin{array}{l l} | \\cos (\\theta) \\cdot (t - t ^ {*}) |, & t \\neq t ^ {*} \\\\ 0, & t = t ^ {*} \\end{array} \\right. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 660, + 890, + 694 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\theta$ is the angle between the view direction $\\mathbf{v}$ and the surface normal $\\mathbf{n}$ . Obviously, the UDF value is not monotonous along a ray when it crosses a surface, where the mapping $t\\rightarrow f_u(\\mathbf{r}(t))$ is non-injective. This leads to difficulties in modeling the monotonicity between $t$ and the density function $\\sigma (t) = \\Omega (f_u(\\mathbf{r}(t)))$ , given that an unbiased density function is expected to be monotonically increasing.", + "bbox": [ + 496, + 703, + 890, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To leverage the unbiased property derived from the monotonic density function, we need to model the injective mapping from $t$ to the density function $\\sigma_{u}(t)$ with the formulation of UDF $f_{u}(\\mathbf{r}(t))$ . Since the monotonicity of UDF along a ray only changes at the intersection point $t^{*}$ , we can", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "20836", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "use a visibility indicator function $\\Psi(t)$ to reflect the existence of the intersection and thus capture the monotonicity change:", + "bbox": [ + 75, + 90, + 468, + 135 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Psi (t) = \\left\\{ \\begin{array}{l l} 1, & t < t ^ {*} \\\\ 0, & t > t ^ {*}. \\end{array} \\right. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 191, + 147, + 468, + 179 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The indicator function $\\Psi(t)$ indicates the visibility of the points sampled along the ray $\\mathbf{r}$ , where 1 means that the point is visible in the free space while 0 means that the point is occluded. Note this just gives a simple hint for the indicator function, while the actual formulation will be derived later.", + "bbox": [ + 75, + 185, + 468, + 260 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With $\\Psi (t)$ , given an unbiased density function $\\sigma (t) = \\Omega_s(f(\\mathbf{r}(t)))$ that is monotonically increasing, the density function for UDF can be defined as", + "bbox": [ + 76, + 261, + 468, + 306 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\sigma_ {u} (t) = \\Psi (t) \\Omega_ {s} \\left(f _ {u} (\\mathbf {r} (t))\\right) + (1 - \\Psi (t)) \\Omega_ {s} \\left(- f _ {u} (\\mathbf {r} (t))\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 83, + 311, + 468, + 340 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\Omega_s$ can be any monotonic density function that is unbiasedly designed, while in this paper we directly use the formulation of Eq.1. Since unsigned distance function $f_u$ is not differentiable at $t^*$ , the domain scope of $\\sigma_u(t)$ is defined as $\\{t \\in \\mathbb{R}^1, t \\neq t^*\\}$ .", + "bbox": [ + 75, + 357, + 468, + 431 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Zero level set. The visibility indicator function $\\Psi$ plays a vital role in the formulation of volume density for UDF, because it essentially encodes the distribution of the geometric surface. The key to the visibility is to find the intersection point $t^*$ of surfaces and the ray, whose UDF value is zero, i.e., $f_{u}(\\mathbf{r}(t^{*})) = 0$ .", + "bbox": [ + 75, + 431, + 468, + 523 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "It is nontrivial to find the zero level set for a UDF field. In practice, it's impossible to locate a point with exact zero UDF value, and a trade-off to balance accuracy and efficiency is to find points whose UDF values are smaller than a pre-defined threshold $\\epsilon$ . As a result, the approximated surfaces will be relaxed from exact zero level set to be a $\\epsilon$ -bounded surface band $S = \\{\\mathbf{x} \\in \\mathbb{R}^3 \\mid f_u(\\mathbf{x}) < \\epsilon\\}$ .", + "bbox": [ + 75, + 523, + 468, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, such a definition of the indicator function involving $\\epsilon$ bounding brings about the following difficulties in learning the UDF field. This indicator function design with $\\epsilon$ bounding operation incurs a set of barriers that impede learning the UDF field: 1) choosing an appropriate $\\epsilon$ is difficult; 2) a fixed $\\epsilon$ cannot be adjusted to adapt the UDF field in different optimization stages; 3) the step-like indicator function is not differentiable. These limitations make the optimization of UDF unstable and easy to collapse.", + "bbox": [ + 75, + 628, + 468, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Visibility indicator function. To overcome the limitations mentioned above, we leverage probabilistic models to make the indicator function $\\Psi(t)$ differentiable. Therefore, we propose to approximate the surfaces by a soft probabilistic distribution instead of using a hard $\\epsilon$ bound. We measure the probability of a point $\\mathbf{r}(t)$ being on a surface, named as surface existence probability $h(\\mathbf{r}(t))$ abbreviated as $h(t)$ . This can be achieved by transforming the UDF values into probabilities via a bell-shaped distribution centered at 0. Here", + "bbox": [ + 75, + 763, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "we use a logistic distribution $\\phi_{\\beta}(x)$ to define the probability $h(t_i)$ :", + "bbox": [ + 498, + 90, + 888, + 121 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nh \\left(t _ {i}\\right) = 1 - \\exp \\left(- \\alpha \\cdot \\phi_ {\\beta} \\left(f _ {u} \\left(\\mathbf {r} \\left(t _ {i}\\right)\\right)\\right) \\cdot \\delta_ {i}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 133, + 890, + 148 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\phi_ {\\beta} (x) = \\beta e ^ {- \\beta x} / \\left(1 + e ^ {- \\beta x}\\right) ^ {2} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 171, + 890, + 191 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $t_i$ is the $i$ -th sample in the ray $\\mathbf{r}$ , $h(t_i)$ denotes the surface existence probability of the sample $t_i$ , $\\delta_i = t_{i+1} - t_i$ is the distance between adjacent samples, and $\\phi_\\beta(t)$ is a logistic distribution function. Note that the standard deviation of $\\phi_\\beta(x)$ is given by $1 / \\beta$ , where $\\beta$ is a trainable parameter, and $1 / \\beta$ approaches zero as the network training converges; $\\alpha$ is a constant scalar to scale the distribution.", + "bbox": [ + 498, + 198, + 890, + 301 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After obtaining the surface existence probabilities of the sampled points on the ray $\\mathbf{r}$ , we can measure the probability of whether a point is occluded by any surfaces or not. Since local UDF values do not indicate occlusion, we accumulate the probabilities of the points $\\mathbf{r}(t_j)$ ahead of a query point $\\mathbf{r}(t_i)$ along the ray:", + "bbox": [ + 498, + 303, + 890, + 393 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\Psi} \\left(t _ {i}\\right) = \\prod_ {j = 1} ^ {i - 1} \\left(1 - h \\left(t _ {j}\\right)\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 609, + 405, + 890, + 446 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $h(t_i) \\in [0,1]$ measures the probability of a point being on the surfaces. From this equation, we can see that $\\hat{\\Psi}(t_i)$ will drop dramatically to zero once there is a point that has high surface existence probability on the ray. Compared to using a hard bound, the probability accumulation strategy makes the indicator function $\\hat{\\Psi}(t)$ differentiable and learnable, which adapts to learning various surface geometries. The function $\\hat{\\Psi}(t)$ will become sharper when $1/\\beta$ approaches 0 as the network training converges, resulting in more accurate approximation of the zero level set.", + "bbox": [ + 498, + 453, + 890, + 604 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Gradient-aware indicator. Although the formulation of $\\hat{\\Psi}(t)$ in Eq. 7 is effective for identifying surfaces in learning the UDF field, it can induce bias in the indicator. As shown in Figure 3, the surface existence probability $h(t_i)$ increases when the ray approaches a surface, and the accumulated probability $\\hat{\\Psi}(t_i)$ can attain zero before the point $t^*$ where the ray intersects with the surface.", + "bbox": [ + 498, + 604, + 890, + 709 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To alleviate this problem, we modify the indicator function $\\hat{\\Psi}(t)$ with a masking function $m(t)$", + "bbox": [ + 498, + 710, + 888, + 741 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\Psi} (t _ {i}) = \\prod_ {j = 1} ^ {i - 1} \\left(1 - h \\left(t _ {j}\\right) \\cdot m \\left(t _ {j}\\right)\\right) \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 752, + 890, + 794 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nm \\left(t _ {j}\\right) = \\left\\{ \\begin{array}{l l} 0, & \\cos \\left(\\theta_ {j + 1}\\right) < 0 \\\\ 1, & \\cos \\left(\\theta_ {j + 1}\\right) \\geq 0 \\end{array} \\right. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 797, + 890, + 830 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\theta_{j + 1}$ denotes the angle between the ray direction and the gradient of $f_{u}(\\mathbf{r}(t_{j + 1}))$ . This formulation exploits the geometric property of UDF surface, that is, the gradient direction of UDF flips across its zero level set. Therefore,", + "bbox": [ + 498, + 839, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "footer", + "text": "20837", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b81c1fe9f1cc7ab184aab0d5c24b584cb184cd43a08b6f821b46167b3d02ee70.jpg", + "image_caption": [ + "(a) Naive indicator func." + ], + "image_footnote": [], + "bbox": [ + 81, + 89, + 269, + 265 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e23405a73e68d984869ff1184cb8ce2eebe0f738205dac767b37c112adc60de5.jpg", + "image_caption": [ + "(b) Grad-aware indicator func.", + "Figure 3. Illustration of (a) naive visibility indicator function and (b) gradient-aware visibility indicator function. The naive indicator function will attain zero before the ray intersects with surface point $t^*$ . Incorporating the UDF gradients information, the gradient-aware indicator function attains zero exactly at the intersection point $t^*$ ." + ], + "image_footnote": [], + "bbox": [ + 281, + 90, + 464, + 266 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the points with negative $\\cos (\\theta_{j + 1})$ are located ahead of the intersection point $t^{*}$ ; we mask out these points and only accumulate the surface existence probability once the ray reaches the surface. The final formulation jointly leverages the insights from both UDF value and gradient, leading to more reliable surface identification.", + "bbox": [ + 75, + 378, + 467, + 467 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Optimizing UDF Fields", + "text_level": 1, + "bbox": [ + 76, + 481, + 290, + 496 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Iso-surface regularizer. With the volume rendering scheme mentioned above, NeuralUDF learns a UDF field from the images by minimizing the differences between rendered images and ground truth images. However, unlike SDF representation that is defined for closed surface, the intrinsic property of the UDF field makes it more difficult to optimize. First, the UDF has a higher degree of freedom accommodating arbitrary topologies instead of only closed surfaces. Second, the UDF is not differentiable at the zero level set; the values and gradients of the UDF field close to the zero level set are therefore hard to learn. This usually results in a wide strip around the iso-surface where the UDF values are distributed irregularly. Therefore, we introduce a simple yet effective geometric regularizer term during training to improve the stability of the UDF field:", + "bbox": [ + 75, + 506, + 467, + 732 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r e g}} = \\frac {1}{M} \\sum_ {k, i} \\exp (- \\tau \\cdot f _ {u} (\\mathbf {r} _ {k} (t _ {k, i}))), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 744, + 467, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\tau$ is a constant scalar to scale the UDF values, and $M$ is the number of sampled rays per optimization step. This formulation effectively precludes the UDF values of the non-surface points from being zero, which thus encourages the simplicity and clearness of the zero level set in the field. The experiments show that this iso-surface regularizer excels at improving the surface quality (see Sec. 4.3).", + "bbox": [ + 75, + 794, + 467, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Loss functions. We train NeuralUDF by enforcing the consistency of the rendered colors and the ground truth colors of the input images without using 3D ground truth shapes. Specifically, we optimize the neural networks and two trainable probabilistic parameters $\\kappa$ of Eq. 1 and $\\beta$ of Eq. 6 by randomly sampling a batch of pixels.", + "bbox": [ + 496, + 90, + 890, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The overall loss function is defined as a weighted sum of several loss terms:", + "bbox": [ + 498, + 181, + 890, + 210 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {c o l o r}} + \\lambda_ {0} \\mathcal {L} _ {\\text {p a t c h}} + \\lambda_ {1} \\mathcal {L} _ {\\text {e i k}} + \\lambda_ {2} \\mathcal {L} _ {\\text {r e g}} (+ \\gamma \\mathcal {L} _ {\\text {m a s k}}). \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 223, + 890, + 239 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{\\mathrm{color}}$ is the L1 color loss to minimize the differences between the rendered pixel colors and the ground truth pixel colors of the input images. Following the prior works [46, 49, 50], the Eikonal term $\\mathcal{L}_{eik}$ is used to regularize the neural distance field to have unit norm of gradients, and the optional mask loss term $\\mathcal{L}_{\\mathrm{mask}}$ can be included if extra masks are provided. $\\mathcal{L}_{\\mathrm{patch}}$ is a patch-based color loss term introduced in [30], which renders the colors of a local patch of an input image using patch blending strategy and then enforces the consistency of the rendered colors and ground truth colors at the patch level. (See the supplementary materials for the details.)", + "bbox": [ + 496, + 250, + 890, + 431 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 498, + 444, + 630, + 460 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental settings", + "text_level": 1, + "bbox": [ + 500, + 469, + 702, + 484 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We conduct evaluations on the DTU [15] dataset and Deepfashion3D [53] dataset. DTU is a multi-view dataset containing objects with closed surfaces, and each scene consists of 49 or 64 images with $1600 \\times 1200$ resolution. We use the same 15 scenes following the prior works for comparison. To validate the effectiveness of our method on objects with open surfaces, we adopt Deepfashion3D dataset for evaluation. Deepfashion3D is a large-scale real 3D garment dataset containing 563 diverse cloth items, and provides the ground truth point clouds of the captured garments. To facilitate the multi-view 3D reconstruction task, we randomly choose 12 garments from different categories of the dataset, and then scale the point clouds to be in the unit sphere. We render each point cloud into 72 images with $1024 \\times 1024$ resolution. The rendered images are used for reconstruction. We further collect a set of synthetic plant objects for further qualitative comparisons. Each plant object is rendered into 324 images with $256 \\times 256$ resolution. Implementation details. We use MLPs to model the unsigned distance field and color radiance field. We use the Adam optimizer with a global learning rate 5e-4 for the network training, specially we set the learning rate of the UDF MLP as 1e-4. We sample 512 rays per batch and train our model for 300k iterations on a single NVIDIA RTX 2080Ti GPU. Our network architecture and initialization scheme are similar to those of prior works [46, 49, 50]. We leverage MeshUDF [14], an extended marching cube algorithm", + "bbox": [ + 496, + 493, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "20838", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e242aabf8e2646899a0f44bba7960b2ea0cb6001c9b04d79823652da3abae17f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 88, + 183, + 167 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/299a72f5e0baeb12419eea0fa76d9c0f057678222ca4b051b4c0f2db4f9baf17.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 106, + 172, + 179, + 262 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2709eed36efc1360d8a22ba02e98a7fb12ffff7487d824f540ee63499dc6655f.jpg", + "image_caption": [ + "Reference Images", + "Figure 4. Qualitative comparisons with SOTAs on DeepFashion3D [53] dataset. The GT meshes are reconstructed from the GT point clouds provided by the DeepFashion3D dataset via Ball Pivoting meshing algorithm [3]. Thanks to the UDF representation, NeuralUDF accurately recovers the garments with open surfaces. However, the SDF-based methods can only model the garments as closed surfaces, thus leading to inconsistent topologies and erroneous geometries. (See supplementary materials for more results.)" + ], + "image_footnote": [], + "bbox": [ + 117, + 266, + 169, + 364 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c0547579cc7ed34a37e939b8df7be412cf4a7aecad236a606931a999e88d16d2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 209, + 89, + 294, + 167 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4968797747c75c8c49414bcebaa9e56c008868bc64b3870bddf15011a92481ed.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 218, + 174, + 289, + 261 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6970deb48840a414a41653849168bf5d08d4588efb5e36deea701c48f1479d5f.jpg", + "image_caption": [ + "GT meshes" + ], + "image_footnote": [], + "bbox": [ + 205, + 266, + 307, + 363 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9ffd3458043d718de8565dcb9313e15749704e03da1b0395f8d1056760122f4e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 344, + 90, + 428, + 167 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6a9e7f918528cf3ebad80496ba3397f1636c824462e44d83d092f846e95ea2cd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 354, + 174, + 424, + 262 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d8af9485e5e787590b24da6d063c35bed6efece4fa50df85e9a840eec44822ad.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 341, + 266, + 444, + 362 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2f2e904f26b537e612f76d8fcbcf36c0f66b84136c9d31a666cb3ce7414eed37.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 478, + 90, + 563, + 167 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6030a40666a7272ba9e263344e497fa2f0c6f08cf205754cdd069bc86b4cdbef.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 491, + 174, + 560, + 262 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/19b7bdea0f27bda7885bd733afb14088d387e816dc831591e7655ea8512d90e1.jpg", + "image_caption": [ + "NeuS [46]" + ], + "image_footnote": [], + "bbox": [ + 478, + 266, + 578, + 363 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4084c6e7e2383ca5ee4bb9309fce142e7cd06c825cf986163f344b40502a6db7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 614, + 90, + 697, + 167 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d25fafb8dc21c7f4c88c572bfe127a652a16c998aad1a180f798443107dfde85.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 625, + 174, + 696, + 262 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d8abcb8765f197281e7f3e3474629167d8036b90b07fd97cd16976eead352335.jpg", + "image_caption": [ + "VolSDF [49]" + ], + "image_footnote": [], + "bbox": [ + 614, + 266, + 715, + 364 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/cc1c7c4339160dd12ddfc76d1cca44dc4f89f5b0191c7c9695783a697add3561.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 750, + 90, + 833, + 167 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6dde80a8a022de71a1175eab430a119b525219ba27d715a0b85057c5c89230de.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 761, + 174, + 831, + 262 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2ace1ce08f1096aa78f01f094bbb21813de616247a1b11e51358e5ab74024a34.jpg", + "image_caption": [ + "NeRF [35]" + ], + "image_footnote": [], + "bbox": [ + 751, + 266, + 854, + 364 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/f635dd88d93497d9e15b89df3c0951591b57798a1a0a402c3ed4a8a8e001c2ed.jpg", + "table_caption": [], + "table_footnote": [ + "$\\dagger$ Optimization using extra object masks.", + "LS-C denotes Long-Sleeve Coat, SS-C denotes Short-Sleeve Coat, NS-C denotes None-Sleeve Coat, LS-D denotes Long-Sleeve Dress, SS-D denotes Short-Sleeve Dress, NS-D denotes None-Sleeve Dress." + ], + "table_body": "
ScanLS-C0SS-D0NS-D0NS-C0Skirt0LS-D0Trouser0NS-D1Trouser1LS-C1Skirt1SS-C0Mean
Colmap [41]2.952.913.583.063.233.263.093.112.953.162.972.953.10
IDR†[50]10.858.205.009.636.766.336.323.323.756.082.447.026.31
Nerf [35]6.387.845.035.454.697.864.824.636.047.178.246.896.25
VolSdf [49]5.924.794.548.577.025.966.974.367.248.737.748.846.72
Neus [46]3.184.824.784.993.735.715.892.215.893.602.445.134.36
NeuralWarp [9]4.7111.245.408.564.105.835.792.002.076.509.616.946.06
Ours1.922.052.361.581.334.112.471.501.632.472.162.151.97
", + "bbox": [ + 89, + 448, + 877, + 551 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Quantitative evaluation results on Deepfashion3D [53] dataset.", + "bbox": [ + 271, + 593, + 694, + 607 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "designed for UDF, to extract explicit mesh from the learned UDF field.", + "bbox": [ + 75, + 625, + 468, + 654 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "A hierarchical sampling strategy is used to sample points along a ray in a coarse-to-fine manner for volume rendering. We first uniformly sample points on the ray and then iteratively conduct importance sampling [35] to sample more points on top of coarse probability estimation. The probability of coarse sampling is computed based on the density function $\\sigma_u(t)$ with fixed and predefined parameters $\\kappa$ and $\\beta$ , while the probability of fine sampling is computed with the learned $\\kappa$ and $\\beta$ .", + "bbox": [ + 75, + 657, + 468, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. We compare NeuralUDF with the following state-of-the-art baselines: 1) a widely used classic reconstruction method COLAMP [41], which we use to reconstruct a mesh from the point cloud generated by COLMAP via Screened Poisson Surface Reconstruction [21]; 2) neural implicit methods utilizing surface rendering, IDR [50] and UNISURF [38]; 3) neural implicit methods utilizing", + "bbox": [ + 75, + 795, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "volume rendering, NeuS [46], VolSDF [49] and NeuralWarp [9]; 4) NeRF [35]. Although NeRF is not designed for surface reconstruction, we include NeRF for comparisons since its density representation enables successful novel view synthesis on objects with arbitrary topologies. Following NeuS [46], we use a density threshold of 25 on DTU (50 on Deepfashion3D) to extract surfaces from NeRF for comparisons. Note that, IDR takes extra masks as input for optimization while ours and the other methods do not.", + "bbox": [ + 496, + 625, + 890, + 761 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparisons", + "text_level": 1, + "bbox": [ + 500, + 771, + 638, + 787 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation on non-closed objects. We conduct evaluations on DeepFashion3D dataset whose objects are non-closed, and measure the Chamfer Distances $(10^{-3})$ between the predicted meshes with ground truth point clouds. As shown in Table 1, our method outperforms the other methods by a large margin. This is because our method can accurately reconstruct the surfaces with arbitrary topologies", + "bbox": [ + 496, + 794, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "20839", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c679f9a0b7dca0b6723e5c55d9bfcfe9fbedac314565d9264bab1ac7fca914f8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 114, + 95, + 240, + 167 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/492d5cb01e8f4ec4a3f0c8b2c8ff094b166ad5d1be6c560966f1809d68806502.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 114, + 171, + 238, + 243 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7df5b1a4d6d4c1aa652f7e3e254cc716ca478e12aaf4cc50c16a3a65578f7ab6.jpg", + "image_caption": [ + "Reference Images", + "Figure 5. Qualitative comparisons with SOTAs on DTU [15] dataset.(See the supplementary materials for more results.)" + ], + "image_footnote": [], + "bbox": [ + 114, + 244, + 236, + 316 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5cbc6191404c5e142cd74502f356e119366045fe351b214c091bc9ff3925ddf0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 256, + 94, + 364, + 175 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/08bc81eaaa0f082acbe792256ee2e7a288d4a613732cb4d2798c1dbba8fe59d3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 272, + 178, + 344, + 242 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/128346cefeb343db08d4663ce8e62cb5b21752df92ff3002967b074f86510af9.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 269, + 244, + 344, + 316 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e25b2b7f2395985ae9c770d86947199343550c60abae68507a27b321daf1f41a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 379, + 94, + 488, + 176 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/408f7f76a4f2343a997affcb7b81f3332173aacb21acbaf4f058e749da798540.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 393, + 178, + 465, + 242 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e839afe1720a683ec56a01d090738901224461d9dab5ec062738dcef4b160032.jpg", + "image_caption": [ + "NeuS [46]" + ], + "image_footnote": [], + "bbox": [ + 392, + 244, + 468, + 315 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3171d1bf9b8b9481cd1f8c64c05031f98880259aa0d09908127361f755bd00b2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 92, + 609, + 172 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f0b7375b52d41b0b43ab10fc53a8816413904bda949bfe8e54fd435818d30ce0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 176, + 588, + 242 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7c6148ef0de344de72a366ec523540a6a4bf1aeb846fb4c6d36b2d9091233f54.jpg", + "image_caption": [ + "VolSDF [49]" + ], + "image_footnote": [], + "bbox": [ + 514, + 244, + 591, + 315 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/001ec6c4bba3259fac69d85d429761d68df4211302fb6f6501a5ef9646abab13.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 620, + 87, + 728, + 172 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/66b7334e3214922b4bfd6d9ca47560bf9aa901b5e41da365c6a931182c065aed.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 638, + 176, + 710, + 241 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7b84905db6a1d7532728af65eaee0da712ef56c6ff04c02dace75ecbe3142dfc.jpg", + "image_caption": [ + "NeRF [35]" + ], + "image_footnote": [], + "bbox": [ + 637, + 244, + 712, + 315 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/99b19a304992c3da9f2223019f8b55c9b4aac4317eb2a12e25434d9d65972f4c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 746, + 90, + 852, + 172 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8c4c75f75a1dab9cbddb85333073be17e792a370c33ea1b36fe37dc94c038b10.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 761, + 176, + 831, + 241 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a92aaa181450ea1df784b097f1cad301ab19009f457f00147db230cc9b46c854.jpg", + "image_caption": [ + "IDR [50]" + ], + "image_footnote": [], + "bbox": [ + 758, + 244, + 834, + 315 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/61e65738bb3ccc451ced233ab31e944e06ba330f9be012345a17f2ab6bcd4fd5.jpg", + "table_caption": [], + "table_footnote": [ + "$\\dagger$ Optimization using extra object masks." + ], + "table_body": "
Scan243740556365698397105106110114118122Mean
Colmap [41]0.812.050.731.221.791.581.023.051.402.051.001.320.490.781.171.36
IDR†[50]1.631.870.630.481.040.790.771.331.160.760.670.900.420.510.530.90
Nerf [35]1.901.601.850.582.281.271.471.672.051.070.882.531.061.150.961.49
UniSurf [38]1.321.361.720.441.350.790.801.491.370.890.591.470.460.590.621.02
VolSdf [49]1.141.260.810.491.250.700.721.291.180.700.661.080.420.610.550.86
Neus [46]1.371.210.730.401.200.700.721.011.160.820.661.690.390.490.510.87
NeuralWarp [9]0.490.710.380.380.790.810.821.201.060.680.660.740.410.630.510.68
Ours0.691.180.670.440.900.660.671.320.940.950.570.860.370.560.550.75
", + "bbox": [ + 99, + 359, + 867, + 481 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Quantitative evaluation results on DTU [15] dataset.", + "bbox": [ + 302, + 496, + 663, + 508 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "while the SDF based methods can only model the garments as closed surfaces and thus exhibit large errors. In Figure 4, the SDF based methods including NeuS, VolSDF and IDR enforce strong assumption on the closeness of the target shape and attempt to model all the shapes as closed surfaces, thus inevitably leading to large errors and inconsistent topologies for the garments. For example, for the first row, VolSDF attempts to use a double-layered surface to represent the thin cloth; optimizing this tiny SDF band is difficult and the results contain many noticeable holes. For the second row, NeuS gives a closed surface for the garment model while the cuffs are even incorrectly connected. Note although IDR takes extra masks for optimization, the reconstructed results still exhibit large quantitative errors.", + "bbox": [ + 75, + 521, + 468, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation on closed objects. We further perform evaluation on DTU dataset whose objects are closed and watertight, and measure the Chamfer Distances between the predicted meshes and ground truth point clouds. Although our method adopts UDF as representation and does not impose closed surface assumption, our method still achieves comparable performance with the SDF based methods that are particularly designed for closed shapes.", + "bbox": [ + 75, + 733, + 468, + 853 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Table 2, in terms of the averaged Chamfer Distance, NeuralUDF outperforms most of the SOTA methods except the most recent work NeuralWarp [8]. Neural-", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Warp is a fine-tuning method that further optimizes the surfaces obtained from VolSDF [49] by minimizing the differences between the warped source images and the reference image. However, NeuralWarp heavily relies on the reconstructed geometry provided by other methods to generate reasonable warped source images, so it cannot be trained from scratch. The qualitative comparisons with SOTAs are provided in Figure 5, where our reconstruction results contain fewer artifacts than the SDF based methods, e.g., the roof of the House and the Apple.", + "bbox": [ + 496, + 521, + 890, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablations and Analysis", + "text_level": 1, + "bbox": [ + 500, + 681, + 712, + 696 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation studies. We conduct ablation studies (Figure 6 and Figure 7) to investigate the individual contribution of the important designs of our method. The first key design is the gradient-aware visibility indicator function for effectively indicating the surfaces. For convenience, we call the visibility indicator function that is not gradient-aware as naive indicator function. As shown in Figure 6, the reconstruction result using the naive indicator function contains larger errors than the result using the gradient-aware indicator function, where the Chamfer Distance (CD) of the two reconstructions are 1.31 and 0.75 respectively (the lower is better). We can also see that the reconstruction using the naive function lacks geometric details, where the teeth and", + "bbox": [ + 496, + 703, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "20840", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7f7e8d1d4201e5d61e877a0779d333081ce57048c1f0c6a0b4f0eeeacefb41a2.jpg", + "image_caption": [ + "(a) Naive indicator function" + ], + "image_footnote": [], + "bbox": [ + 83, + 89, + 205, + 210 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/31c1d23292fc68f8f429730d3cc5d004e5b21e062ffce9f65db8a4248c5617a5.jpg", + "image_caption": [ + "(b) Gradient-aware indicator function" + ], + "image_footnote": [], + "bbox": [ + 215, + 90, + 339, + 213 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/85e0c2f0093ed06b50b4809ccd74bdf97aae13ac2edf1c8988bdcdec04a00355.jpg", + "image_caption": [ + "(c) Patch blending" + ], + "image_footnote": [], + "bbox": [ + 344, + 92, + 470, + 212 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "eye holes of the Skull are not reconstructed accurately. This is because the naive indicator function has inherent bias and the measured visibility is not consistent with the true surface intersection, thus leading to inaccurate reconstruction.", + "bbox": [ + 75, + 291, + 468, + 353 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We leverage an iso-surface regularizer to make the UDF field more stable and accurate around zero level sets. The reconstructions with or without the regularizer are shown in Figure 7. Although the Chamfer Distances of the two results are close, the mesh without the regularizer is much more noisy and has noticeable artifacts while the mesh with the regularizer is much cleaner and smoother. To give an in-depth analysis, we plot the UDF values and the cosine values of the angle between the ray direction and UDF gradients along a selected ray (the marked red point of (b,c)). We can see that, without the regularizer, the signs of the $\\cos (\\theta)$ values change multiple times around the zero level set, which indicates the gradients are not stable. In contrast, with the regularizer, the gradients are more stable, where the gradient direction only flips at the surface intersection.", + "bbox": [ + 75, + 357, + 468, + 583 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Also, inspired by [30], the patch blending loss term is included in the optimization. With this term, the reconstruction can be further improved from 0.75 to 0.66 in terms of Chamfer Distance.", + "bbox": [ + 75, + 588, + 470, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "More types of shapes. We additionally show results on several challenging objects with complex open surfaces. As shown in Figure 8, our method successfully reconstructs the challenging plant models with non-manifold geometries, while the reconstruction results of NeuS [46] are incomplete and contain numerous holes. This is because SDF is essentially not a suitable representation for many topologies like the thin leaves. Therefore, optimizing a SDF field for these kinds of models is deserved to be difficult and the reconstructed surfaces are naturally erroneous.", + "bbox": [ + 75, + 654, + 468, + 805 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. Compared to the SDF based methods, NeuralUDF will degrade on textureless objects that lack enough distinguishable features, and produces smoother results that may lack some geometric details. This is because NeuralUDF does not impose closed surface assumption and is not differentiable at zero-level sets (see details in the Supp.).", + "bbox": [ + 75, + 809, + 468, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/de32a3cd349c56e222bf269a1a0ac82360820e3ec4702daae15477e052b1f349.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 511, + 97, + 622, + 209 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/635b208294a045cd28b474ab370f80513ebacb52d8d35306bcfc2672c6d17608.jpg", + "image_caption": [ + "CD: 2.47" + ], + "image_footnote": [], + "bbox": [ + 637, + 98, + 738, + 208 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/625fab0d2d5831f3392c9df5166a1c38592304c7f996fe6ab1573325c0cd3934.jpg", + "image_caption": [ + "CD: 2.49 (c) w/o regularizer" + ], + "image_footnote": [], + "bbox": [ + 764, + 98, + 864, + 208 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/60a51f16ca4e5e28aa130b839df915df0823cc67fffdd6e727b9384b1538d78f.jpg", + "image_caption": [ + "Figure 6. Ablation studies on gradient-aware visibility indicator function and patch blending loss term.", + "(a) Reference Image", + "(d) The curves of UDF values and $\\cos (\\theta)$ values", + "Figure 7. Ablation study on the iso-surface regularizer." + ], + "image_footnote": [], + "bbox": [ + 504, + 233, + 689, + 321 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8adaa7428266edcb54815f623de0c08df91c918d2ba3ba47f82612d465b15f5b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 689, + 233, + 887, + 321 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7c0345feff3e7963df5c10345d4c095375a37f2e189e1d0a7444fac532dddcc7.jpg", + "image_caption": [ + "Figure 8. Qualitative comparisons with NeuS on two plants." + ], + "image_footnote": [], + "bbox": [ + 511, + 364, + 879, + 565 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusions", + "text_level": 1, + "bbox": [ + 500, + 590, + 627, + 606 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduce NeuralUDF, a NeRF-based method that uses neural UDF fields for multiview reconstruction of surfaces with arbitrary topologies. Our key idea is to introduce a novel volume density function that effectively correlates the UDF representation with the volume rendering scheme to learn neural UDF fields. The experiments show that, unlike the SDF based works that can only model closed shapes, our method can reconstruct more types of shapes with either open or closed surfaces, thus extending the underlying geometric representation of neural volume rendering to various topologies.", + "bbox": [ + 496, + 617, + 890, + 784 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 500, + 799, + 668, + 815 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Xiaoxiao Long is supported by Hong Kong PhD Fellowship Scheme. Taku Komura is partly supported by the Hong Kong ITF Grant ITS/319/21FP. Lingjie Liu and Christian Theobalt have been supported by the ERC Consolidator Grant 4DReply (770784).", + "bbox": [ + 496, + 824, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "20841", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Matan Atzmon and Yaron Lipman. Sal: Sign agnostic learning of shapes from raw data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2565-2574, 2020. 2", + "[2] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 1", + "[3] Fausto Bernardini, Joshua Mittleman, Holly Rushmeier, Cláudio Silva, and Gabriel Taubin. The ball-pivoting algorithm for surface reconstruction. IEEE transactions on visualization and computer graphics, 5(4):349-359, 1999. 2, 6", + "[4] Neill DF Campbell, George Vogiatzis, Carlos Hernández, and Roberto Cipolla. Using multiple hypotheses to improve depth-maps for multi-view stereo. In European Conference on Computer Vision, pages 766-779. Springer, 2008. 2", + "[5] Weikai Chen, Cheng Lin, Weiyang Li, and Bo Yang. 3psdf: Three-pole signed distance function for learning surfaces with arbitrary topologies. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18522-18531, 2022. 1, 2", + "[6] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5939-5948, 2019. 2", + "[7] Julian Chibane, Gerard Pons-Moll, et al. Neural unsigned distance fields for implicit function learning. Advances in Neural Information Processing Systems, 33:21638-21652, 2020. 2", + "[8] François Darmon, Bénédicte Bascle, Jean-Clément Devaux, Pascal Monasse, and Mathieu Aubry. Improving neural implicit surfaces geometry with patch warping. arXiv preprint arXiv:2112.09648, 2021. 2, 7", + "[9] François Darmon, Bénédicte Bascle, Jean-Clément Devaux, Pascal Monasse, and Mathieu Aubry. Improving neural implicit surfaces geometry with patch warping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6260–6269, 2022. 6, 7", + "[10] Yasutaka Furukawa and Jean Ponce. Accurate, dense, and robust multiview stereopsis. IEEE transactions on pattern analysis and machine intelligence, 32(8):1362-1376, 2009. 2", + "[11] Silvano Galliani, Katrin Lasinger, and Konrad Schindler. Massively parallel multiview stereopsis by surface normal diffusion. In Proceedings of the IEEE International Conference on Computer Vision, pages 873-881, 2015. 2", + "[12] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. arXiv preprint arXiv:2002.10099, 2020. 2", + "[13] Xiaodong Gu, Zhiwen Fan, Siyu Zhu, Zuozhuo Dai, Feitong Tan, and Ping Tan. Cascade cost volume for high-resolution multi-view stereo and stereo matching. In Proceedings of" + ], + "bbox": [ + 76, + 114, + 472, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2495-2504, 2020. 2", + "[14] Benoit Guillard, Federico Stella, and Pascal Fua. Meshud: Fast and differentiable meshing of unsigned distance field networks. arXiv preprint arXiv:2111.14549, 2021. 5", + "[15] Rasmus Jensen, Anders Dahl, George Vogiatzis, Engin Tola, and Henrik Aanæs. Large scale multi-view stereopsis evaluation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 406-413, 2014. 2, 5, 7", + "[16] Mengqi Ji, Juergen Gall, Haitian Zheng, Yebin Liu, and Lu Fang. Surfacenet: An end-to-end 3d neural network for multiview stereopsis. In Proceedings of the IEEE International Conference on Computer Vision, pages 2307-2315, 2017. 2", + "[17] Mengqi Ji, Jinzhi Zhang, Qionghai Dai, and Lu Fang. SurfaceNet+: An end-to-end 3d neural network for very sparse multi-view stereopsis. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(11):4078-4093, 2020. 2", + "[18] Yue Jiang, Dantong Ji, Zhizhong Han, and Matthias Zwicker. Sdfdiff: Differentiable rendering of signed distance fields for 3d shape optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1251-1261, 2020. 2", + "[19] Abhishek Kar, Christian Hane, and Jitendra Malik. Learning a multi-view stereo machine. Advances in neural information processing systems, 30, 2017. 2", + "[20] Michael Kazhdan, Matthew Bolitho, and Hugues Hoppe. Poisson surface reconstruction. In Proceedings of the fourth Eurographics symposium on Geometry processing, volume 7, 2006. 2", + "[21] Michael Kazhdan and Hugues Hoppe. Screened poisson surface reconstruction. ACM Transactions on Graphics (ToG), 32(3):1-13, 2013. 6", + "[22] Petr Kellnhofer, Lars C Jebe, Andrew Jones, Ryan Spicer, Kari Pulli, and Gordon Wetzstein. Neural lumigraph rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4287-4297, 2021. 2", + "[23] Kiriakos N Kutulakos and Steven M Seitz. A theory of shape by space carving. International journal of computer vision, 38(3):199-218, 2000. 2", + "[24] Der-Tsai Lee and Bruce J Schachter. Two algorithms for constructing a delaunay triangulation. International Journal of Computer & Information Sciences, 9(3):219-242, 1980. 2", + "[25] Maxime Lhuillier and Long Quan. A quasi-dense approach to surface reconstruction from uncalibrated images. IEEE transactions on pattern analysis and machine intelligence, 27(3):418-433, 2005. 2", + "[26] Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. Barf: Bundle-adjusting neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5741-5751, 2021. 1", + "[27] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. Advances in Neural Information Processing Systems, 33:15651-15663, 2020. 1", + "[28] Shaohui Liu, Yinda Zhang, Songyou Peng, Boxin Shi, Marc Pollefeys, and Zhaopeng Cui. Dist: Rendering deep implicit" + ], + "bbox": [ + 501, + 92, + 890, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "20842", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "signed distance function with differentiable sphere tracing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2019-2028, 2020. 2", + "[29] Xiaoxiao Long, Cheng Lin, Lingjie Liu, Wei Li, Christian Theobalt, Ruigang Yang, and Wenping Wang. Adaptive surface normal constraint for depth estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12849-12858, 2021. 2", + "[30] Xiaoxiao Long, Cheng Lin, Peng Wang, Taku Komura, and Wenping Wang. Sparseneus: Fast generalizable neural surface reconstruction from sparse views. arXiv preprint arXiv:2206.05737, 2022. 5, 8", + "[31] Xiaoxiao Long, Lingjie Liu, Wei Li, Christian Theobalt, and Wenping Wang. Multi-view depth estimation using epipolar spatio-temporal networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8258-8267, 2021. 2", + "[32] Xiaoxiao Long, Lingjie Liu, Christian Theobalt, and Wenping Wang. Occlusion-aware depth estimation with adaptive normal constraints. In European Conference on Computer Vision, pages 640-657. Springer, 2020. 2", + "[33] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4460-4470, 2019. 2", + "[34] Mateusz Michalkiewicz, Jhony K Pontes, Dominic Jack, Mahsa Baktashmotlagh, and Anders Eriksson. Implicit surface representations as layers in neural networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4743-4752, 2019. 2", + "[35] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European conference on computer vision, pages 405-421. Springer, 2020. 1, 2, 6, 7", + "[36] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11453-11464, 2021. 1", + "[37] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3504-3515, 2020. 2", + "[38] Michael Oechsle, Songyou Peng, and Andreas Geiger. Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5589-5599, 2021. 2, 6, 7", + "[39] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 165-174, 2019. 2", + "[40] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865-5874, 2021. 1", + "[41] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 6, 7", + "[42] Johannes L Schonberger, Enliang Zheng, Jan-Michael Frahm, and Marc Pollefeys. Pixelwise view selection for unstructured multi-view stereo. In European Conference on Computer Vision, pages 501-518. Springer, 2016. 2", + "[43] Steven M Seitz and Charles R Dyer. Photorealistic scene reconstruction by voxel coloring. International Journal of Computer Vision, 35(2):151-173, 1999. 2", + "[44] Jiaming Sun, Yiming Xie, Linghao Chen, Xiaowei Zhou, and Hujun Bao. Neuralrecon: Real-time coherent 3d reconstruction from monocular video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15598-15607, 2021. 1, 2", + "[45] Engin Tola, Christoph Strecha, and Pascal Fua. Efficient large-scale multi-view stereo for ultra high-resolution image sets. Machine Vision and Applications, 23(5):903-920, 2012. 2", + "[46] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. Advances in Neural Information Processing Systems, 34, 2021. 1, 2, 3, 5, 6, 7, 8", + "[47] Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. Mvsnet: Depth inference for unstructured multi-view stereo. In Proceedings of the European Conference on Computer Vision (ECCV), pages 767-783, 2018. 2", + "[48] Yao Yao, Zixin Luo, Shiwei Li, Tianwei Shen, Tian Fang, and Long Quan. Recurrent mvsnet for high-resolution multi-view stereo depth inference. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5525-5534, 2019. 2", + "[49] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems, 34, 2021. 1, 2, 5, 6, 7", + "[50] Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Basri Ronen, and Yaron Lipman. Multiview neural surface reconstruction by disentangling geometry and appearance. Advances in Neural Information Processing Systems, 33:2492-2502, 2020. 2, 5, 6, 7", + "[51] Jingyang Zhang, Yao Yao, and Long Quan. Learning signed distance field for multi-view surface reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6525-6534, 2021. 2", + "[52] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492, 2020. 1", + "[53] Heming Zhu, Yu Cao, Hang Jin, Weikai Chen, Dong Du, Zhangye Wang, Shuguang Cui, and Xiaoguang Han. Deep fashion3d: A dataset and benchmark for 3d garment reconstruction from single images. In European Conference on Computer Vision, pages 512-530. Springer, 2020. 2, 5, 6" + ], + "bbox": [ + 501, + 92, + 893, + 897 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "20843", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/NeuralUDF_ Learning Unsigned Distance Fields for Multi-View Reconstruction of Surfaces With Arbitrary Topologies/cda2b21c-85ea-47d7-952f-4f7b2568e13a_model.json b/2023/NeuralUDF_ Learning Unsigned Distance Fields for Multi-View Reconstruction of Surfaces With Arbitrary Topologies/cda2b21c-85ea-47d7-952f-4f7b2568e13a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..551ebd2d64f603d173125f9a724a5af14c983dec --- /dev/null +++ b/2023/NeuralUDF_ Learning Unsigned Distance Fields for Multi-View Reconstruction of Surfaces With Arbitrary Topologies/cda2b21c-85ea-47d7-952f-4f7b2568e13a_model.json @@ -0,0 +1,2893 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.138, + 0.131, + 0.83, + 0.177 + ], + "angle": 0, + "content": "NeuralUDF: Learning Unsigned Distance Fields for Multi-view Reconstruction of Surfaces with Arbitrary Topologies" + }, + { + "type": "text", + "bbox": [ + 0.176, + 0.202, + 0.793, + 0.282 + ], + "angle": 0, + "content": "Xiaoxiao Long\\(^{1,2\\dagger}\\) Cheng Lin\\(^{2*}\\) Lingjie Liu\\(^{3}\\) Yuan Liu\\(^{1}\\) Peng Wang\\(^{1}\\) Christian Theobalt\\(^{3}\\) Taku Komura\\(^{1}\\) Wenping Wang\\(^{4*}\\) \n\\(^{1}\\)The University of Hong Kong \\(^{2}\\)Tencent Games \n\\(^{3}\\)Max Planck Institute for Informatics \\(^{4}\\)Texas A&M University" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.314, + 0.314, + 0.33 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.345, + 0.474, + 0.649 + ], + "angle": 0, + "content": "We present a novel method, called NeuralUDF, for reconstructing surfaces with arbitrary topologies from 2D images via volume rendering. Recent advances in neural rendering based reconstruction have achieved compelling results. However, these methods are limited to objects with closed surfaces since they adopt Signed Distance Function (SDF) as surface representation which requires the target shape to be divided into inside and outside. In this paper, we propose to represent surfaces as the Unsigned Distance Function (UDF) and develop a new volume rendering scheme to learn the neural UDF representation. Specifically, a new density function that correlates the property of UDF with the volume rendering scheme is introduced for robust optimization of the UDF fields. Experiments on the DTU and DeepFashion3D datasets show that our method not only enables high-quality reconstruction of non-closed shapes with complex typologies, but also achieves comparable performance to the SDF based methods on the reconstruction of closed surfaces. Visit our project page at https://www.xxlong.site/NeuralUDF/." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.659, + 0.21, + 0.675 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.684, + 0.47, + 0.866 + ], + "angle": 0, + "content": "Reconstructing high-quality surfaces from multi-view images is a long-standing problem in computer vision and computer graphics. Neural implicit fields have become an emerging trend in recent advances due to its superior capability of representing surfaces of complex geometry. NeRF [35] and its variants [2, 26, 27, 36, 40, 52] have recently achieved compelling results in novel view synthesis. For each point in the 3D space, NeRF-based methods learn two neural implicit functions based on volume rendering: a volume density function and a view-dependent color function. Despite their successes in novel view synthesis, NeRF-based methods still struggle to faithfully reconstruct the ac" + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.314, + 0.878, + 0.572 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.574, + 0.892, + 0.658 + ], + "angle": 0, + "content": "Figure 1. We show three groups of multi-view reconstruction results generated by our proposed NeuralUDF and NeuS [46] respectively. Our method is able to faithfully reconstruct the high-quality geometries for both the closed and open surfaces, while NeuS can only model shapes as closed surfaces, thus leading to inconsistent typologies and erroneous geometries." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.672, + 0.894, + 0.717 + ], + "angle": 0, + "content": "curate scene geometry from multi-view inputs, because of the difficulty in extracting high-quality surfaces from the representation of volume density." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.718, + 0.892, + 0.87 + ], + "angle": 0, + "content": "VolSDF [49] and NeuS [46] incorporate the Signed Distance Function (SDF) into volume rendering to facilitate high-quality surface reconstruction in a NeRF framework. However, as a continuous function with clearly defined inside/outside, SDF is limited to modeling only closed watertight surfaces. Although there have been efforts to modify the SDF representation by learning an additional truncation function (e.g., 3PSDF [5], TSDF [44]), their surface representations are still built upon the definition of SDF, thus they are not suitable for representing complex topologies." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We therefore propose to employ the Unsigned Distance Function (UDF) to represent surfaces in volume rendering." + }, + { + "type": "page_footnote", + "bbox": [ + 0.1, + 0.875, + 0.234, + 0.888 + ], + "angle": 0, + "content": "*Corresponding authors." + }, + { + "type": "page_footnote", + "bbox": [ + 0.1, + 0.888, + 0.453, + 0.9 + ], + "angle": 0, + "content": "This work was conducted during an internship at Tencent Games." + }, + { + "type": "list", + "bbox": [ + 0.1, + 0.871, + 0.453, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "20834" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "With a surface represented by its zero level set without signs, UDF is a unified representation with higher-degree of freedom for both closed and open surfaces, thus making it possible to reconstruct shapes with arbitrary topologies." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.153, + 0.471, + 0.332 + ], + "angle": 0, + "content": "There are two major challenges in learning a neural UDF field by volume rendering. First, UDF is not occlusion-aware, while the formulation of volume rendering with distance fields requires to estimate surface occlusion for points. Considering a camera ray intersecting with a surface, SDF assumes that the surface is closed and distinguishes the inside/outside with signs, so that a negative SDF value on the ray clearly indicates an occluded point inside the surface. In contrast, UDF does not impose the closed surface assumption and always gives non-negative values along the ray. Hence, the UDF value of a point alone cannot be used to infer occlusion." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.333, + 0.47, + 0.452 + ], + "angle": 0, + "content": "The second challenge is that the UDF is not differentiable at its zero-level sets. The non-differentiability at the zero-level sets imposes barriers in the learning of UDF field. The gradients are ill-defined near the iso-surface, leading to difficulty in optimization. Consequently, the distance field surrounding the iso-surface is not accurate and the exact zero-level set of a learned UDF cannot be identified in a stable manner." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.455, + 0.471, + 0.634 + ], + "angle": 0, + "content": "In this paper, we present a novel method for learning neural UDF fields by volume rendering. We introduce a density function that correlates the property of the UDF representation with the volume rendering process, which effectively tackles the aforementioned challenges induced by the unsigned representation and enables robust learning of surfaces with arbitrary topologies. Experiments on DTU [15] and DeepFashion3D [53] datasets show that our method not only enables the high-quality reconstruction of non-closed shapes with complex topologies, but also achieves comparable performance to the SDF based methods on the reconstruction of closed surfaces." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.636, + 0.37, + 0.649 + ], + "angle": 0, + "content": "Our contributions can be summarized as:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.65, + 0.468, + 0.693 + ], + "angle": 0, + "content": "- We incorporate the UDF into volume rendering, which extends the representation of the underlying geometry of neural radiance fields." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.697, + 0.468, + 0.755 + ], + "angle": 0, + "content": "- We introduce an effective density function that correlates the property of the UDF with the volume rendering process, thus enabling robust optimization of the distance fields." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.758, + 0.468, + 0.803 + ], + "angle": 0, + "content": "- Our method achieves SOTA results for reconstructing high-quality surfaces with various topologies (closed or open) using the UDF in volume rendering." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.65, + 0.468, + 0.803 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.815, + 0.219, + 0.831 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Classic multi-view 3D reconstruction. 3D reconstruction from multiple images is a fundamental problem in computer vision. Traditional reconstruction methods usually adopt discrete representations, such as voxel grids [16," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.287 + ], + "angle": 0, + "content": "17, 19, 23, 43, 44], 3D point clouds [10, 25], and depth maps [4, 11, 13, 29, 31, 32, 42, 45, 47, 48]. The depth map representation is flexible and appropriate for parallel computation, so it is the most widely used representation. The depth map based methods first compute depth maps by analyzing the correlation between multiple input images, and then fuse the depth maps into a point cloud, which is then further processed into a mesh via meshing algorithms, such as Poisson surface reconstruction [20], ball-pivoting [3] and Delaunay triangulation [24]. However, the traditional methods are sensitive to reflections, image noises, and varying illuminations, thus usually yielding incomplete and noisy reconstructions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.288, + 0.892, + 0.513 + ], + "angle": 0, + "content": "Neural rendering based surface reconstruction. The recent advances on neural implicit representations have achieved promising results in multi-view 3D reconstruction [8,18,22,28,37,38,46,49-51]. These methods typically represent the surfaces to be reconstructed as Signed Distance Function (SDF), and utilize surface rendering or volume rendering to render the pixel colors of the input images for optimizing the SDF fields. The surface rendering based methods [37, 50] require extra masks for optimization and cannot handle objects with complex structures. The volume rendering based works [8, 38, 46, 49, 51] get rid of the extra masks and achieve more accurate reconstructions. However, these methods adopt SDF as representation so they are limited to closed surfaces and their performance degrade a lot for non-closed surfaces." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.515, + 0.892, + 0.756 + ], + "angle": 0, + "content": "Surface representations. There are various representations to model 3D geometric surfaces. NeRF [35] uses a density field to represent the underlying geometry of a target scene. Although NeRF achieves huge successes in novel view synthesis, it's difficult to extract high-quality geometry from such a density representation. Signed Distance Function (SDF) is widely used for geometry modeling and surface reconstruction [1,6,8,12,33,34,38,39,46,49,51], but SDF is only applicable to objects with closed surfaces. To deal with open surfaces, there are some efforts to learn an additional truncation function (e.g. 3PSDF [5], TSDF [44]). However, their representations are just built on the modification of SDF, and thus are still subject to flexibility of SDF and struggle to model complex surfaces. NDF [7] uses a neural UDF field to reconstruct surfaces, but they explicitly rely on 3D point clouds for supervising the optimization." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.77, + 0.591, + 0.785 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Our goal is to reconstruct a surface \\(S\\) from a set of posed input images \\(\\{\\mathcal{I}_k\\}\\) of a 3D model. We represent the surface as the zero level set of an Unsigned Distance Function (UDF), and propose a novel volume rendering scheme to learn the UDF. We will introduce our method in three parts: 1) We discuss the relationship between the unsigned distance function and the density function of the volume ren" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20835" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.089, + 0.27, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.22, + 0.25, + 0.232 + ], + "angle": 0, + "content": "(a) Biased density function" + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.104, + 0.46, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.277, + 0.22, + 0.454, + 0.232 + ], + "angle": 0, + "content": "(b) Unbiased density function" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.234, + 0.47, + 0.332 + ], + "angle": 0, + "content": "Figure 2. Illustration of (a) biased density function and (b) unbiased density function for UDF. Suppose that the ray has one single intersection point \\( t^* \\) with the surface. (a) If the density function is modeled as a bell-shaped probability distribution, the weight function attains a local maximum before \\( t^* \\). (b) If the density function is modeled as a monotonically increasing function, the weight function can attain its local maximum at \\( t^* \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.343, + 0.473, + 0.48 + ], + "angle": 0, + "content": "dering equation; 2) To derive a density function induced by the UDF, we propose a differentiable indicator function to transform the UDF into its associated density field; 3) In addition to the loss terms commonly used in the prior works, we introduce an iso-surface regularizer to improve the stability of the UDF field near the zero level set. This UDF-based volume rendering scheme enables robust optimization of NeuralUDF for reconstructing both closed surfaces and open surfaces." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.488, + 0.258, + 0.505 + ], + "angle": 0, + "content": "3.1. Volume Rendering" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.512, + 0.47, + 0.618 + ], + "angle": 0, + "content": "Scene representation. Our method NeuralUDF represents the scene to be reconstructed by two neural implicit functions: 1) an unsigned distance function \\( f_{u}(\\mathbf{x}) \\) that maps a 3D point \\( \\mathbf{x} \\in \\mathbb{R}^3 \\) to its unsigned distance to the closest object surface; 2) a view-dependent color function \\( c(\\mathbf{x},\\mathbf{v}) \\) that encodes the color associated with a point \\( \\mathbf{x} \\in \\mathbb{R}^3 \\) and a viewing direction \\( \\mathbf{v} \\in \\mathbb{R}^3 \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.618, + 0.471, + 0.78 + ], + "angle": 0, + "content": "Volume rendering. We consider a ray \\(\\mathbf{r}\\) emanating from the camera center \\(\\mathbf{o} \\in \\mathbb{R}^3\\) in direction \\(\\mathbf{v} \\in \\mathbb{R}^3\\) (\\(\\| \\mathbf{v} \\| = 1\\)), which is parameterized by \\(\\{\\mathbf{r}(t) = \\mathbf{o} + t\\mathbf{v} \\mid t \\geq 0\\}\\). Volume rendering is used to calculate the color \\(C\\) of the pixel corresponding to the ray \\(\\mathbf{r}\\). The accumulated color of the camera ray \\(\\mathbf{r}\\) is given by \\(C(\\mathbf{r}) = \\int_0^{+\\infty} T(t) \\sigma(t) \\mathbf{c}(\\mathbf{r}(t), \\mathbf{v}) \\, \\mathrm{d}t\\), where \\(\\sigma(t)\\) is the volume density, \\(T(t) = \\exp \\left(-\\int_0^t \\sigma(u) \\, \\mathrm{d}u\\right)\\) denotes the accumulated transmittance along the ray, and the product of \\(T(t) \\sigma(t)\\) is the weight function of a point, denoted as \\(w(t)\\), i.e., \\(w(t) = T(t) \\sigma(t)\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.787, + 0.329, + 0.803 + ], + "angle": 0, + "content": "3.2. Modeling Density with UDF" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.471, + 0.903 + ], + "angle": 0, + "content": "In the NeRF framework, the density \\(\\sigma(t)\\) is predicted by an MLP, where the points near the surfaces having larger density values and the points far from the surfaces having smaller densities. In order to apply volume rendering to learning distance field \\(f\\), we need to define a mapping \\(\\Omega\\) to transform the distance function \\(f(\\mathbf{r}(t))\\) into a density field" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.713, + 0.107 + ], + "angle": 0, + "content": "\\(\\sigma (t)\\) , that is \\(\\sigma (t) = \\Omega (f(\\mathbf{r}(t)))\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.107, + 0.892, + 0.287 + ], + "angle": 0, + "content": "Unbiased volume density function. Inspired by NeRF, we may consider defining the density function \\(\\sigma(t) = \\Omega(f(\\mathbf{r}(t)))\\) as a bell-shaped distribution model centered at zero, so that the surface points intersecting with a ray (UDF values are zero) have the maximum density values. However, as illustrated in Fig 2, if the mapping function is straightforwardly defined in this way, a noticeable bias will be induced in the weight function \\(w(t)\\) since its local maximum is not at the intersection point \\(t^*\\) where the ray intersects with the surfaces. This can result in conspicuous inaccuracy in the reconstructed surfaces, which is demonstrated in [46]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.288, + 0.893, + 0.409 + ], + "angle": 0, + "content": "To define an appropriate mapping, we need to first review the relationship between the density function \\(\\sigma(t)\\) and a distance field \\(f(t)\\). Suppose that there is only one single intersection of the ray and a surface. It is shown in NeuS [46] that, to make the weight function \\(w(t)\\) unbiased, the density function should be in the form of a monotonically increasing function, which attains the maximum at an infinite point rather than at the intersection point." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.409, + 0.892, + 0.5 + ], + "angle": 0, + "content": "Accordingly, NeuS [46] chooses to use a Sigmoid function \\(\\Phi_{\\kappa}(x) = (1 + e^{-\\kappa x})^{-1}\\) to achieve the goal. The accumulated transmittance \\(T(t)\\) is then defined as \\(T(t) = \\Phi_{\\kappa}(f_s(t))\\), where \\(f_{s}(t)\\) is the signed distance function (SDF). The density function \\(\\sigma_{s}(t)\\) induced by SDF is formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.51, + 0.513, + 0.891, + 0.544 + ], + "angle": 0, + "content": "\\[\n\\sigma_ {s} (t) = \\Omega_ {s} \\left(f _ {s} (\\mathbf {r} (t))\\right) = \\kappa | \\cos (\\theta) | \\left(1 - \\Phi_ {\\kappa} \\left(f _ {s} (\\mathbf {r} (t))\\right)\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.544, + 0.892, + 0.587 + ], + "angle": 0, + "content": "where \\(\\kappa\\) is the parameter of Sigmoid function \\(\\Phi_{\\kappa}(x)\\), and \\(\\theta\\) is the angle between the view direction \\(\\mathbf{v}\\) and the surface normal \\(\\mathbf{n}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.589, + 0.892, + 0.651 + ], + "angle": 0, + "content": "Unbiased volume density for UDF. However, the unbiased density formulation induced by SDF [46] cannot be directly applied to UDF. The formulation of UDF near the surface can be given by" + }, + { + "type": "equation", + "bbox": [ + 0.549, + 0.661, + 0.892, + 0.695 + ], + "angle": 0, + "content": "\\[\nf _ {u} (\\mathbf {r} (t)) = \\left\\{ \\begin{array}{l l} | \\cos (\\theta) \\cdot (t - t ^ {*}) |, & t \\neq t ^ {*} \\\\ 0, & t = t ^ {*} \\end{array} \\right. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.892, + 0.825 + ], + "angle": 0, + "content": "where \\(\\theta\\) is the angle between the view direction \\(\\mathbf{v}\\) and the surface normal \\(\\mathbf{n}\\). Obviously, the UDF value is not monotonous along a ray when it crosses a surface, where the mapping \\(t\\rightarrow f_u(\\mathbf{r}(t))\\) is non-injective. This leads to difficulties in modeling the monotonicity between \\(t\\) and the density function \\(\\sigma (t) = \\Omega (f_u(\\mathbf{r}(t)))\\), given that an unbiased density function is expected to be monotonically increasing." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "To leverage the unbiased property derived from the monotonic density function, we need to model the injective mapping from \\( t \\) to the density function \\( \\sigma_{u}(t) \\) with the formulation of UDF \\( f_{u}(\\mathbf{r}(t)) \\). Since the monotonicity of UDF along a ray only changes at the intersection point \\( t^{*} \\), we can" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "20836" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.136 + ], + "angle": 0, + "content": "use a visibility indicator function \\(\\Psi(t)\\) to reflect the existence of the intersection and thus capture the monotonicity change:" + }, + { + "type": "equation", + "bbox": [ + 0.192, + 0.148, + 0.469, + 0.18 + ], + "angle": 0, + "content": "\\[\n\\Psi (t) = \\left\\{ \\begin{array}{l l} 1, & t < t ^ {*} \\\\ 0, & t > t ^ {*}. \\end{array} \\right. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.186, + 0.469, + 0.261 + ], + "angle": 0, + "content": "The indicator function \\(\\Psi(t)\\) indicates the visibility of the points sampled along the ray \\(\\mathbf{r}\\), where 1 means that the point is visible in the free space while 0 means that the point is occluded. Note this just gives a simple hint for the indicator function, while the actual formulation will be derived later." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.262, + 0.469, + 0.307 + ], + "angle": 0, + "content": "With \\(\\Psi (t)\\), given an unbiased density function \\(\\sigma (t) = \\Omega_s(f(\\mathbf{r}(t)))\\) that is monotonically increasing, the density function for UDF can be defined as" + }, + { + "type": "equation", + "bbox": [ + 0.084, + 0.312, + 0.469, + 0.342 + ], + "angle": 0, + "content": "\\[\n\\sigma_ {u} (t) = \\Psi (t) \\Omega_ {s} \\left(f _ {u} (\\mathbf {r} (t))\\right) + (1 - \\Psi (t)) \\Omega_ {s} \\left(- f _ {u} (\\mathbf {r} (t))\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.358, + 0.469, + 0.433 + ], + "angle": 0, + "content": "where \\(\\Omega_s\\) can be any monotonic density function that is unbiasedly designed, while in this paper we directly use the formulation of Eq.1. Since unsigned distance function \\(f_u\\) is not differentiable at \\(t^*\\), the domain scope of \\(\\sigma_u(t)\\) is defined as \\(\\{t \\in \\mathbb{R}^1, t \\neq t^*\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.433, + 0.469, + 0.524 + ], + "angle": 0, + "content": "Zero level set. The visibility indicator function \\(\\Psi\\) plays a vital role in the formulation of volume density for UDF, because it essentially encodes the distribution of the geometric surface. The key to the visibility is to find the intersection point \\(t^*\\) of surfaces and the ray, whose UDF value is zero, i.e., \\(f_{u}(\\mathbf{r}(t^{*})) = 0\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.524, + 0.469, + 0.63 + ], + "angle": 0, + "content": "It is nontrivial to find the zero level set for a UDF field. In practice, it's impossible to locate a point with exact zero UDF value, and a trade-off to balance accuracy and efficiency is to find points whose UDF values are smaller than a pre-defined threshold \\(\\epsilon\\). As a result, the approximated surfaces will be relaxed from exact zero level set to be a \\(\\epsilon\\)-bounded surface band \\(S = \\{\\mathbf{x} \\in \\mathbb{R}^3 \\mid f_u(\\mathbf{x}) < \\epsilon\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.63, + 0.469, + 0.765 + ], + "angle": 0, + "content": "However, such a definition of the indicator function involving \\(\\epsilon\\) bounding brings about the following difficulties in learning the UDF field. This indicator function design with \\(\\epsilon\\) bounding operation incurs a set of barriers that impede learning the UDF field: 1) choosing an appropriate \\(\\epsilon\\) is difficult; 2) a fixed \\(\\epsilon\\) cannot be adjusted to adapt the UDF field in different optimization stages; 3) the step-like indicator function is not differentiable. These limitations make the optimization of UDF unstable and easy to collapse." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Visibility indicator function. To overcome the limitations mentioned above, we leverage probabilistic models to make the indicator function \\(\\Psi(t)\\) differentiable. Therefore, we propose to approximate the surfaces by a soft probabilistic distribution instead of using a hard \\(\\epsilon\\) bound. We measure the probability of a point \\(\\mathbf{r}(t)\\) being on a surface, named as surface existence probability \\(h(\\mathbf{r}(t))\\) abbreviated as \\(h(t)\\). This can be achieved by transforming the UDF values into probabilities via a bell-shaped distribution centered at 0. Here" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.89, + 0.122 + ], + "angle": 0, + "content": "we use a logistic distribution \\(\\phi_{\\beta}(x)\\) to define the probability \\(h(t_i)\\):" + }, + { + "type": "equation", + "bbox": [ + 0.551, + 0.134, + 0.892, + 0.15 + ], + "angle": 0, + "content": "\\[\nh \\left(t _ {i}\\right) = 1 - \\exp \\left(- \\alpha \\cdot \\phi_ {\\beta} \\left(f _ {u} \\left(\\mathbf {r} \\left(t _ {i}\\right)\\right)\\right) \\cdot \\delta_ {i}\\right), \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.593, + 0.172, + 0.892, + 0.193 + ], + "angle": 0, + "content": "\\[\n\\phi_ {\\beta} (x) = \\beta e ^ {- \\beta x} / \\left(1 + e ^ {- \\beta x}\\right) ^ {2} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.199, + 0.892, + 0.303 + ], + "angle": 0, + "content": "where \\( t_i \\) is the \\( i \\)-th sample in the ray \\( \\mathbf{r} \\), \\( h(t_i) \\) denotes the surface existence probability of the sample \\( t_i \\), \\( \\delta_i = t_{i+1} - t_i \\) is the distance between adjacent samples, and \\( \\phi_\\beta(t) \\) is a logistic distribution function. Note that the standard deviation of \\( \\phi_\\beta(x) \\) is given by \\( 1 / \\beta \\), where \\( \\beta \\) is a trainable parameter, and \\( 1 / \\beta \\) approaches zero as the network training converges; \\( \\alpha \\) is a constant scalar to scale the distribution." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.304, + 0.892, + 0.395 + ], + "angle": 0, + "content": "After obtaining the surface existence probabilities of the sampled points on the ray \\(\\mathbf{r}\\), we can measure the probability of whether a point is occluded by any surfaces or not. Since local UDF values do not indicate occlusion, we accumulate the probabilities of the points \\(\\mathbf{r}(t_j)\\) ahead of a query point \\(\\mathbf{r}(t_i)\\) along the ray:" + }, + { + "type": "equation", + "bbox": [ + 0.61, + 0.406, + 0.892, + 0.448 + ], + "angle": 0, + "content": "\\[\n\\hat {\\Psi} \\left(t _ {i}\\right) = \\prod_ {j = 1} ^ {i - 1} \\left(1 - h \\left(t _ {j}\\right)\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.454, + 0.892, + 0.605 + ], + "angle": 0, + "content": "where \\( h(t_i) \\in [0,1] \\) measures the probability of a point being on the surfaces. From this equation, we can see that \\( \\hat{\\Psi}(t_i) \\) will drop dramatically to zero once there is a point that has high surface existence probability on the ray. Compared to using a hard bound, the probability accumulation strategy makes the indicator function \\( \\hat{\\Psi}(t) \\) differentiable and learnable, which adapts to learning various surface geometries. The function \\( \\hat{\\Psi}(t) \\) will become sharper when \\( 1/\\beta \\) approaches 0 as the network training converges, resulting in more accurate approximation of the zero level set." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.606, + 0.892, + 0.71 + ], + "angle": 0, + "content": "Gradient-aware indicator. Although the formulation of \\(\\hat{\\Psi}(t)\\) in Eq. 7 is effective for identifying surfaces in learning the UDF field, it can induce bias in the indicator. As shown in Figure 3, the surface existence probability \\(h(t_i)\\) increases when the ray approaches a surface, and the accumulated probability \\(\\hat{\\Psi}(t_i)\\) can attain zero before the point \\(t^*\\) where the ray intersects with the surface." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.711, + 0.89, + 0.742 + ], + "angle": 0, + "content": "To alleviate this problem, we modify the indicator function \\(\\hat{\\Psi}(t)\\) with a masking function \\(m(t)\\)" + }, + { + "type": "equation", + "bbox": [ + 0.591, + 0.753, + 0.892, + 0.795 + ], + "angle": 0, + "content": "\\[\n\\hat {\\Psi} (t _ {i}) = \\prod_ {j = 1} ^ {i - 1} \\left(1 - h \\left(t _ {j}\\right) \\cdot m \\left(t _ {j}\\right)\\right) \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.586, + 0.798, + 0.892, + 0.832 + ], + "angle": 0, + "content": "\\[\nm \\left(t _ {j}\\right) = \\left\\{ \\begin{array}{l l} 0, & \\cos \\left(\\theta_ {j + 1}\\right) < 0 \\\\ 1, & \\cos \\left(\\theta_ {j + 1}\\right) \\geq 0 \\end{array} \\right. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where \\(\\theta_{j + 1}\\) denotes the angle between the ray direction and the gradient of \\(f_{u}(\\mathbf{r}(t_{j + 1}))\\). This formulation exploits the geometric property of UDF surface, that is, the gradient direction of UDF flips across its zero level set. Therefore," + }, + { + "type": "footer", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20837" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.09, + 0.27, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.103, + 0.267, + 0.25, + 0.279 + ], + "angle": 0, + "content": "(a) Naive indicator func." + }, + { + "type": "image", + "bbox": [ + 0.282, + 0.091, + 0.465, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.283, + 0.268, + 0.465, + 0.28 + ], + "angle": 0, + "content": "(b) Grad-aware indicator func." + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.281, + 0.468, + 0.364 + ], + "angle": 0, + "content": "Figure 3. Illustration of (a) naive visibility indicator function and (b) gradient-aware visibility indicator function. The naive indicator function will attain zero before the ray intersects with surface point \\( t^* \\). Incorporating the UDF gradients information, the gradient-aware indicator function attains zero exactly at the intersection point \\( t^* \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.379, + 0.468, + 0.468 + ], + "angle": 0, + "content": "the points with negative \\(\\cos (\\theta_{j + 1})\\) are located ahead of the intersection point \\(t^{*}\\); we mask out these points and only accumulate the surface existence probability once the ray reaches the surface. The final formulation jointly leverages the insights from both UDF value and gradient, leading to more reliable surface identification." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.482, + 0.292, + 0.497 + ], + "angle": 0, + "content": "3.3. Optimizing UDF Fields" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.507, + 0.468, + 0.733 + ], + "angle": 0, + "content": "Iso-surface regularizer. With the volume rendering scheme mentioned above, NeuralUDF learns a UDF field from the images by minimizing the differences between rendered images and ground truth images. However, unlike SDF representation that is defined for closed surface, the intrinsic property of the UDF field makes it more difficult to optimize. First, the UDF has a higher degree of freedom accommodating arbitrary topologies instead of only closed surfaces. Second, the UDF is not differentiable at the zero level set; the values and gradients of the UDF field close to the zero level set are therefore hard to learn. This usually results in a wide strip around the iso-surface where the UDF values are distributed irregularly. Therefore, we introduce a simple yet effective geometric regularizer term during training to improve the stability of the UDF field:" + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.745, + 0.468, + 0.78 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r e g}} = \\frac {1}{M} \\sum_ {k, i} \\exp (- \\tau \\cdot f _ {u} (\\mathbf {r} _ {k} (t _ {k, i}))), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.468, + 0.901 + ], + "angle": 0, + "content": "where \\(\\tau\\) is a constant scalar to scale the UDF values, and \\(M\\) is the number of sampled rays per optimization step. This formulation effectively precludes the UDF values of the non-surface points from being zero, which thus encourages the simplicity and clearness of the zero level set in the field. The experiments show that this iso-surface regularizer excels at improving the surface quality (see Sec. 4.3)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.182 + ], + "angle": 0, + "content": "Loss functions. We train NeuralUDF by enforcing the consistency of the rendered colors and the ground truth colors of the input images without using 3D ground truth shapes. Specifically, we optimize the neural networks and two trainable probabilistic parameters \\(\\kappa\\) of Eq. 1 and \\(\\beta\\) of Eq. 6 by randomly sampling a batch of pixels." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.183, + 0.892, + 0.211 + ], + "angle": 0, + "content": "The overall loss function is defined as a weighted sum of several loss terms:" + }, + { + "type": "equation", + "bbox": [ + 0.508, + 0.224, + 0.892, + 0.24 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {c o l o r}} + \\lambda_ {0} \\mathcal {L} _ {\\text {p a t c h}} + \\lambda_ {1} \\mathcal {L} _ {\\text {e i k}} + \\lambda_ {2} \\mathcal {L} _ {\\text {r e g}} (+ \\gamma \\mathcal {L} _ {\\text {m a s k}}). \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.251, + 0.892, + 0.432 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{\\mathrm{color}}\\) is the L1 color loss to minimize the differences between the rendered pixel colors and the ground truth pixel colors of the input images. Following the prior works [46, 49, 50], the Eikonal term \\(\\mathcal{L}_{eik}\\) is used to regularize the neural distance field to have unit norm of gradients, and the optional mask loss term \\(\\mathcal{L}_{\\mathrm{mask}}\\) can be included if extra masks are provided. \\(\\mathcal{L}_{\\mathrm{patch}}\\) is a patch-based color loss term introduced in [30], which renders the colors of a local patch of an input image using patch blending strategy and then enforces the consistency of the rendered colors and ground truth colors at the patch level. (See the supplementary materials for the details.)" + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.445, + 0.632, + 0.462 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.47, + 0.703, + 0.485 + ], + "angle": 0, + "content": "4.1. Experimental settings" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.494, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Datasets. We conduct evaluations on the DTU [15] dataset and Deepfashion3D [53] dataset. DTU is a multi-view dataset containing objects with closed surfaces, and each scene consists of 49 or 64 images with \\(1600 \\times 1200\\) resolution. We use the same 15 scenes following the prior works for comparison. To validate the effectiveness of our method on objects with open surfaces, we adopt Deepfashion3D dataset for evaluation. Deepfashion3D is a large-scale real 3D garment dataset containing 563 diverse cloth items, and provides the ground truth point clouds of the captured garments. To facilitate the multi-view 3D reconstruction task, we randomly choose 12 garments from different categories of the dataset, and then scale the point clouds to be in the unit sphere. We render each point cloud into 72 images with \\(1024 \\times 1024\\) resolution. The rendered images are used for reconstruction. We further collect a set of synthetic plant objects for further qualitative comparisons. Each plant object is rendered into 324 images with \\(256 \\times 256\\) resolution. Implementation details. We use MLPs to model the unsigned distance field and color radiance field. We use the Adam optimizer with a global learning rate 5e-4 for the network training, specially we set the learning rate of the UDF MLP as 1e-4. We sample 512 rays per batch and train our model for 300k iterations on a single NVIDIA RTX 2080Ti GPU. Our network architecture and initialization scheme are similar to those of prior works [46, 49, 50]. We leverage MeshUDF [14], an extended marching cube algorithm" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20838" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.089, + 0.184, + 0.169 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.174, + 0.18, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.119, + 0.267, + 0.171, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.366, + 0.196, + 0.378 + ], + "angle": 0, + "content": "Reference Images" + }, + { + "type": "image", + "bbox": [ + 0.21, + 0.09, + 0.295, + 0.169 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.175, + 0.29, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.207, + 0.267, + 0.308, + 0.364 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.218, + 0.365, + 0.288, + 0.377 + ], + "angle": 0, + "content": "GT meshes" + }, + { + "type": "image", + "bbox": [ + 0.345, + 0.091, + 0.429, + 0.169 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.355, + 0.175, + 0.425, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.342, + 0.267, + 0.445, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.374, + 0.365, + 0.404, + 0.376 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.48, + 0.091, + 0.564, + 0.169 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.492, + 0.175, + 0.561, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.48, + 0.267, + 0.58, + 0.364 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.495, + 0.365, + 0.558, + 0.377 + ], + "angle": 0, + "content": "NeuS [46]" + }, + { + "type": "image", + "bbox": [ + 0.616, + 0.091, + 0.699, + 0.169 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.627, + 0.175, + 0.697, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.616, + 0.267, + 0.716, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.626, + 0.365, + 0.703, + 0.377 + ], + "angle": 0, + "content": "VolSDF [49]" + }, + { + "type": "image", + "bbox": [ + 0.75, + 0.091, + 0.834, + 0.169 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.763, + 0.175, + 0.833, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.752, + 0.267, + 0.855, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.772, + 0.365, + 0.837, + 0.377 + ], + "angle": 0, + "content": "NeRF [35]" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.38, + 0.894, + 0.437 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparisons with SOTAs on DeepFashion3D [53] dataset. The GT meshes are reconstructed from the GT point clouds provided by the DeepFashion3D dataset via Ball Pivoting meshing algorithm [3]. Thanks to the UDF representation, NeuralUDF accurately recovers the garments with open surfaces. However, the SDF-based methods can only model the garments as closed surfaces, thus leading to inconsistent topologies and erroneous geometries. (See supplementary materials for more results.)" + }, + { + "type": "table", + "bbox": [ + 0.091, + 0.449, + 0.878, + 0.553 + ], + "angle": 0, + "content": "
ScanLS-C0SS-D0NS-D0NS-C0Skirt0LS-D0Trouser0NS-D1Trouser1LS-C1Skirt1SS-C0Mean
Colmap [41]2.952.913.583.063.233.263.093.112.953.162.972.953.10
IDR†[50]10.858.205.009.636.766.336.323.323.756.082.447.026.31
Nerf [35]6.387.845.035.454.697.864.824.636.047.178.246.896.25
VolSdf [49]5.924.794.548.577.025.966.974.367.248.737.748.846.72
Neus [46]3.184.824.784.993.735.715.892.215.893.602.445.134.36
NeuralWarp [9]4.7111.245.408.564.105.835.792.002.076.509.616.946.06
Ours1.922.052.361.581.334.112.471.501.632.472.162.151.97
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.102, + 0.554, + 0.32, + 0.566 + ], + "angle": 0, + "content": "\\(\\dagger\\) Optimization using extra object masks." + }, + { + "type": "table_footnote", + "bbox": [ + 0.11, + 0.566, + 0.876, + 0.59 + ], + "angle": 0, + "content": "LS-C denotes Long-Sleeve Coat, SS-C denotes Short-Sleeve Coat, NS-C denotes None-Sleeve Coat, LS-D denotes Long-Sleeve Dress, SS-D denotes Short-Sleeve Dress, NS-D denotes None-Sleeve Dress." + }, + { + "type": "list", + "bbox": [ + 0.102, + 0.554, + 0.876, + 0.59 + ], + "angle": 0, + "content": null + }, + { + "type": "table_caption", + "bbox": [ + 0.272, + 0.594, + 0.696, + 0.608 + ], + "angle": 0, + "content": "Table 1. Quantitative evaluation results on Deepfashion3D [53] dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.625, + 0.47, + 0.655 + ], + "angle": 0, + "content": "designed for UDF, to extract explicit mesh from the learned UDF field." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.658, + 0.469, + 0.793 + ], + "angle": 0, + "content": "A hierarchical sampling strategy is used to sample points along a ray in a coarse-to-fine manner for volume rendering. We first uniformly sample points on the ray and then iteratively conduct importance sampling [35] to sample more points on top of coarse probability estimation. The probability of coarse sampling is computed based on the density function \\(\\sigma_u(t)\\) with fixed and predefined parameters \\(\\kappa\\) and \\(\\beta\\), while the probability of fine sampling is computed with the learned \\(\\kappa\\) and \\(\\beta\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Baselines. We compare NeuralUDF with the following state-of-the-art baselines: 1) a widely used classic reconstruction method COLAMP [41], which we use to reconstruct a mesh from the point cloud generated by COLMAP via Screened Poisson Surface Reconstruction [21]; 2) neural implicit methods utilizing surface rendering, IDR [50] and UNISURF [38]; 3) neural implicit methods utilizing" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.625, + 0.892, + 0.762 + ], + "angle": 0, + "content": "volume rendering, NeuS [46], VolSDF [49] and NeuralWarp [9]; 4) NeRF [35]. Although NeRF is not designed for surface reconstruction, we include NeRF for comparisons since its density representation enables successful novel view synthesis on objects with arbitrary topologies. Following NeuS [46], we use a density threshold of 25 on DTU (50 on Deepfashion3D) to extract surfaces from NeRF for comparisons. Note that, IDR takes extra masks as input for optimization while ours and the other methods do not." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.772, + 0.64, + 0.788 + ], + "angle": 0, + "content": "4.2. Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Evaluation on non-closed objects. We conduct evaluations on DeepFashion3D dataset whose objects are non-closed, and measure the Chamfer Distances \\((10^{-3})\\) between the predicted meshes with ground truth point clouds. As shown in Table 1, our method outperforms the other methods by a large margin. This is because our method can accurately reconstruct the surfaces with arbitrary topologies" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20839" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.115, + 0.096, + 0.241, + 0.168 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.116, + 0.172, + 0.239, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.116, + 0.245, + 0.238, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.125, + 0.318, + 0.233, + 0.33 + ], + "angle": 0, + "content": "Reference Images" + }, + { + "type": "image", + "bbox": [ + 0.257, + 0.095, + 0.366, + 0.176 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.273, + 0.179, + 0.345, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.271, + 0.246, + 0.346, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.296, + 0.318, + 0.326, + 0.328 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.38, + 0.095, + 0.49, + 0.177 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.395, + 0.179, + 0.467, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.393, + 0.246, + 0.469, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.402, + 0.317, + 0.464, + 0.33 + ], + "angle": 0, + "content": "NeuS [46]" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.093, + 0.61, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.178, + 0.589, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.246, + 0.592, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.519, + 0.317, + 0.596, + 0.33 + ], + "angle": 0, + "content": "VolSDF [49]" + }, + { + "type": "image", + "bbox": [ + 0.621, + 0.088, + 0.73, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.178, + 0.711, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.638, + 0.246, + 0.713, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.645, + 0.317, + 0.711, + 0.33 + ], + "angle": 0, + "content": "NeRF [35]" + }, + { + "type": "image", + "bbox": [ + 0.747, + 0.091, + 0.854, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.762, + 0.178, + 0.833, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.759, + 0.246, + 0.836, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.771, + 0.317, + 0.827, + 0.33 + ], + "angle": 0, + "content": "IDR [50]" + }, + { + "type": "image_caption", + "bbox": [ + 0.13, + 0.334, + 0.838, + 0.348 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparisons with SOTAs on DTU [15] dataset.(See the supplementary materials for more results.)" + }, + { + "type": "table", + "bbox": [ + 0.1, + 0.361, + 0.868, + 0.482 + ], + "angle": 0, + "content": "
Scan243740556365698397105106110114118122Mean
Colmap [41]0.812.050.731.221.791.581.023.051.402.051.001.320.490.781.171.36
IDR†[50]1.631.870.630.481.040.790.771.331.160.760.670.900.420.510.530.90
Nerf [35]1.901.601.850.582.281.271.471.672.051.070.882.531.061.150.961.49
UniSurf [38]1.321.361.720.441.350.790.801.491.370.890.591.470.460.590.621.02
VolSdf [49]1.141.260.810.491.250.700.721.291.180.700.661.080.420.610.550.86
Neus [46]1.371.210.730.401.200.700.721.011.160.820.661.690.390.490.510.87
NeuralWarp [9]0.490.710.380.380.790.810.821.201.060.680.660.740.410.630.510.68
Ours0.691.180.670.440.900.660.671.320.940.950.570.860.370.560.550.75
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.11, + 0.484, + 0.34, + 0.497 + ], + "angle": 0, + "content": "\\(\\dagger\\) Optimization using extra object masks." + }, + { + "type": "table_caption", + "bbox": [ + 0.303, + 0.497, + 0.665, + 0.51 + ], + "angle": 0, + "content": "Table 2. Quantitative evaluation results on DTU [15] dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.522, + 0.47, + 0.734 + ], + "angle": 0, + "content": "while the SDF based methods can only model the garments as closed surfaces and thus exhibit large errors. In Figure 4, the SDF based methods including NeuS, VolSDF and IDR enforce strong assumption on the closeness of the target shape and attempt to model all the shapes as closed surfaces, thus inevitably leading to large errors and inconsistent topologies for the garments. For example, for the first row, VolSDF attempts to use a double-layered surface to represent the thin cloth; optimizing this tiny SDF band is difficult and the results contain many noticeable holes. For the second row, NeuS gives a closed surface for the garment model while the cuffs are even incorrectly connected. Note although IDR takes extra masks for optimization, the reconstructed results still exhibit large quantitative errors." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.734, + 0.469, + 0.854 + ], + "angle": 0, + "content": "Evaluation on closed objects. We further perform evaluation on DTU dataset whose objects are closed and watertight, and measure the Chamfer Distances between the predicted meshes and ground truth point clouds. Although our method adopts UDF as representation and does not impose closed surface assumption, our method still achieves comparable performance with the SDF based methods that are particularly designed for closed shapes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "As shown in Table 2, in terms of the averaged Chamfer Distance, NeuralUDF outperforms most of the SOTA methods except the most recent work NeuralWarp [8]. Neural-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.522, + 0.892, + 0.673 + ], + "angle": 0, + "content": "Warp is a fine-tuning method that further optimizes the surfaces obtained from VolSDF [49] by minimizing the differences between the warped source images and the reference image. However, NeuralWarp heavily relies on the reconstructed geometry provided by other methods to generate reasonable warped source images, so it cannot be trained from scratch. The qualitative comparisons with SOTAs are provided in Figure 5, where our reconstruction results contain fewer artifacts than the SDF based methods, e.g., the roof of the House and the Apple." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.682, + 0.714, + 0.698 + ], + "angle": 0, + "content": "4.3. Ablations and Analysis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Ablation studies. We conduct ablation studies (Figure 6 and Figure 7) to investigate the individual contribution of the important designs of our method. The first key design is the gradient-aware visibility indicator function for effectively indicating the surfaces. For convenience, we call the visibility indicator function that is not gradient-aware as naive indicator function. As shown in Figure 6, the reconstruction result using the naive indicator function contains larger errors than the result using the gradient-aware indicator function, where the Chamfer Distance (CD) of the two reconstructions are 1.31 and 0.75 respectively (the lower is better). We can also see that the reconstruction using the naive function lacks geometric details, where the teeth and" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20840" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.09, + 0.207, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.08, + 0.215, + 0.182, + 0.236 + ], + "angle": 0, + "content": "(a) Naive indicator function" + }, + { + "type": "image", + "bbox": [ + 0.217, + 0.092, + 0.34, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.215, + 0.328, + 0.236 + ], + "angle": 0, + "content": "(b) Gradient-aware indicator function" + }, + { + "type": "image", + "bbox": [ + 0.346, + 0.093, + 0.472, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.349, + 0.215, + 0.447, + 0.227 + ], + "angle": 0, + "content": "(c) Patch blending" + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.242, + 0.47, + 0.271 + ], + "angle": 0, + "content": "Figure 6. Ablation studies on gradient-aware visibility indicator function and patch blending loss term." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.292, + 0.47, + 0.354 + ], + "angle": 0, + "content": "eye holes of the Skull are not reconstructed accurately. This is because the naive indicator function has inherent bias and the measured visibility is not consistent with the true surface intersection, thus leading to inaccurate reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.358, + 0.47, + 0.584 + ], + "angle": 0, + "content": "We leverage an iso-surface regularizer to make the UDF field more stable and accurate around zero level sets. The reconstructions with or without the regularizer are shown in Figure 7. Although the Chamfer Distances of the two results are close, the mesh without the regularizer is much more noisy and has noticeable artifacts while the mesh with the regularizer is much cleaner and smoother. To give an in-depth analysis, we plot the UDF values and the cosine values of the angle between the ray direction and UDF gradients along a selected ray (the marked red point of (b,c)). We can see that, without the regularizer, the signs of the \\(\\cos (\\theta)\\) values change multiple times around the zero level set, which indicates the gradients are not stable. In contrast, with the regularizer, the gradients are more stable, where the gradient direction only flips at the surface intersection." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.589, + 0.471, + 0.649 + ], + "angle": 0, + "content": "Also, inspired by [30], the patch blending loss term is included in the optimization. With this term, the reconstruction can be further improved from 0.75 to 0.66 in terms of Chamfer Distance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.655, + 0.47, + 0.806 + ], + "angle": 0, + "content": "More types of shapes. We additionally show results on several challenging objects with complex open surfaces. As shown in Figure 8, our method successfully reconstructs the challenging plant models with non-manifold geometries, while the reconstruction results of NeuS [46] are incomplete and contain numerous holes. This is because SDF is essentially not a suitable representation for many topologies like the thin leaves. Therefore, optimizing a SDF field for these kinds of models is deserved to be difficult and the reconstructed surfaces are naturally erroneous." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Limitations. Compared to the SDF based methods, NeuralUDF will degrade on textureless objects that lack enough distinguishable features, and produces smoother results that may lack some geometric details. This is because NeuralUDF does not impose closed surface assumption and is not differentiable at zero-level sets (see details in the Supp.)." + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.098, + 0.623, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.508, + 0.219, + 0.631, + 0.232 + ], + "angle": 0, + "content": "(a) Reference Image" + }, + { + "type": "image", + "bbox": [ + 0.638, + 0.099, + 0.739, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.656, + 0.209, + 0.72, + 0.22 + ], + "angle": 0, + "content": "CD: 2.47" + }, + { + "type": "image", + "bbox": [ + 0.765, + 0.099, + 0.865, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.762, + 0.209, + 0.875, + 0.23 + ], + "angle": 0, + "content": "CD: 2.49 (c) w/o regularizer" + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.234, + 0.69, + 0.323 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.551, + 0.323, + 0.838, + 0.335 + ], + "angle": 0, + "content": "(d) The curves of UDF values and \\(\\cos (\\theta)\\) values" + }, + { + "type": "image", + "bbox": [ + 0.69, + 0.234, + 0.888, + 0.323 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.532, + 0.336, + 0.86, + 0.35 + ], + "angle": 0, + "content": "Figure 7. Ablation study on the iso-surface regularizer." + }, + { + "type": "image", + "bbox": [ + 0.512, + 0.365, + 0.88, + 0.566 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.517, + 0.569, + 0.874, + 0.584 + ], + "angle": 0, + "content": "Figure 8. Qualitative comparisons with NeuS on two plants." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.592, + 0.628, + 0.607 + ], + "angle": 0, + "content": "5. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.618, + 0.892, + 0.785 + ], + "angle": 0, + "content": "We introduce NeuralUDF, a NeRF-based method that uses neural UDF fields for multiview reconstruction of surfaces with arbitrary topologies. Our key idea is to introduce a novel volume density function that effectively correlates the UDF representation with the volume rendering scheme to learn neural UDF fields. The experiments show that, unlike the SDF based works that can only model closed shapes, our method can reconstruct more types of shapes with either open or closed surfaces, thus extending the underlying geometric representation of neural volume rendering to various topologies." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.8, + 0.669, + 0.816 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Xiaoxiao Long is supported by Hong Kong PhD Fellowship Scheme. Taku Komura is partly supported by the Hong Kong ITF Grant ITS/319/21FP. Lingjie Liu and Christian Theobalt have been supported by the ERC Consolidator Grant 4DReply (770784)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20841" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.473, + 0.171 + ], + "angle": 0, + "content": "[1] Matan Atzmon and Yaron Lipman. Sal: Sign agnostic learning of shapes from raw data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2565-2574, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.472, + 0.256 + ], + "angle": 0, + "content": "[2] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.258, + 0.472, + 0.327 + ], + "angle": 0, + "content": "[3] Fausto Bernardini, Joshua Mittleman, Holly Rushmeier, Cláudio Silva, and Gabriel Taubin. The ball-pivoting algorithm for surface reconstruction. IEEE transactions on visualization and computer graphics, 5(4):349-359, 1999. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.329, + 0.472, + 0.386 + ], + "angle": 0, + "content": "[4] Neill DF Campbell, George Vogiatzis, Carlos Hernández, and Roberto Cipolla. Using multiple hypotheses to improve depth-maps for multi-view stereo. In European Conference on Computer Vision, pages 766-779. Springer, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.387, + 0.472, + 0.457 + ], + "angle": 0, + "content": "[5] Weikai Chen, Cheng Lin, Weiyang Li, and Bo Yang. 3psdf: Three-pole signed distance function for learning surfaces with arbitrary topologies. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18522-18531, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.458, + 0.472, + 0.514 + ], + "angle": 0, + "content": "[6] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5939-5948, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.515, + 0.472, + 0.57 + ], + "angle": 0, + "content": "[7] Julian Chibane, Gerard Pons-Moll, et al. Neural unsigned distance fields for implicit function learning. Advances in Neural Information Processing Systems, 33:21638-21652, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.572, + 0.472, + 0.628 + ], + "angle": 0, + "content": "[8] François Darmon, Bénédicte Bascle, Jean-Clément Devaux, Pascal Monasse, and Mathieu Aubry. Improving neural implicit surfaces geometry with patch warping. arXiv preprint arXiv:2112.09648, 2021. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.63, + 0.472, + 0.699 + ], + "angle": 0, + "content": "[9] François Darmon, Bénédicte Bascle, Jean-Clément Devaux, Pascal Monasse, and Mathieu Aubry. Improving neural implicit surfaces geometry with patch warping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6260–6269, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.701, + 0.472, + 0.755 + ], + "angle": 0, + "content": "[10] Yasutaka Furukawa and Jean Ponce. Accurate, dense, and robust multiview stereopsis. IEEE transactions on pattern analysis and machine intelligence, 32(8):1362-1376, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.758, + 0.472, + 0.814 + ], + "angle": 0, + "content": "[11] Silvano Galliani, Katrin Lasinger, and Konrad Schindler. Massively parallel multiview stereopsis by surface normal diffusion. In Proceedings of the IEEE International Conference on Computer Vision, pages 873-881, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.816, + 0.472, + 0.857 + ], + "angle": 0, + "content": "[12] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. arXiv preprint arXiv:2002.10099, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.859, + 0.472, + 0.902 + ], + "angle": 0, + "content": "[13] Xiaodong Gu, Zhiwen Fan, Siyu Zhu, Zuozhuo Dai, Feitong Tan, and Ping Tan. Cascade cost volume for high-resolution multi-view stereo and stereo matching. In Proceedings of" + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.116, + 0.473, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2495-2504, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.163 + ], + "angle": 0, + "content": "[14] Benoit Guillard, Federico Stella, and Pascal Fua. Meshud: Fast and differentiable meshing of unsigned distance field networks. arXiv preprint arXiv:2111.14549, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.165, + 0.892, + 0.22 + ], + "angle": 0, + "content": "[15] Rasmus Jensen, Anders Dahl, George Vogiatzis, Engin Tola, and Henrik Aanæs. Large scale multi-view stereopsis evaluation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 406-413, 2014. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.222, + 0.892, + 0.277 + ], + "angle": 0, + "content": "[16] Mengqi Ji, Juergen Gall, Haitian Zheng, Yebin Liu, and Lu Fang. Surfacenet: An end-to-end 3d neural network for multiview stereopsis. In Proceedings of the IEEE International Conference on Computer Vision, pages 2307-2315, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.279, + 0.892, + 0.333 + ], + "angle": 0, + "content": "[17] Mengqi Ji, Jinzhi Zhang, Qionghai Dai, and Lu Fang. SurfaceNet+: An end-to-end 3d neural network for very sparse multi-view stereopsis. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(11):4078-4093, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.335, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[18] Yue Jiang, Dantong Ji, Zhizhong Han, and Matthias Zwicker. Sdfdiff: Differentiable rendering of signed distance fields for 3d shape optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1251-1261, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.405, + 0.892, + 0.447 + ], + "angle": 0, + "content": "[19] Abhishek Kar, Christian Hane, and Jitendra Malik. Learning a multi-view stereo machine. Advances in neural information processing systems, 30, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.448, + 0.892, + 0.502 + ], + "angle": 0, + "content": "[20] Michael Kazhdan, Matthew Bolitho, and Hugues Hoppe. Poisson surface reconstruction. In Proceedings of the fourth Eurographics symposium on Geometry processing, volume 7, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.504, + 0.892, + 0.545 + ], + "angle": 0, + "content": "[21] Michael Kazhdan and Hugues Hoppe. Screened poisson surface reconstruction. ACM Transactions on Graphics (ToG), 32(3):1-13, 2013. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.547, + 0.892, + 0.615 + ], + "angle": 0, + "content": "[22] Petr Kellnhofer, Lars C Jebe, Andrew Jones, Ryan Spicer, Kari Pulli, and Gordon Wetzstein. Neural lumigraph rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4287-4297, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.617, + 0.892, + 0.658 + ], + "angle": 0, + "content": "[23] Kiriakos N Kutulakos and Steven M Seitz. A theory of shape by space carving. International journal of computer vision, 38(3):199-218, 2000. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.66, + 0.892, + 0.702 + ], + "angle": 0, + "content": "[24] Der-Tsai Lee and Bruce J Schachter. Two algorithms for constructing a delaunay triangulation. International Journal of Computer & Information Sciences, 9(3):219-242, 1980. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.704, + 0.892, + 0.758 + ], + "angle": 0, + "content": "[25] Maxime Lhuillier and Long Quan. A quasi-dense approach to surface reconstruction from uncalibrated images. IEEE transactions on pattern analysis and machine intelligence, 27(3):418-433, 2005. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.76, + 0.892, + 0.815 + ], + "angle": 0, + "content": "[26] Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. Barf: Bundle-adjusting neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5741-5751, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.817, + 0.892, + 0.87 + ], + "angle": 0, + "content": "[27] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. Advances in Neural Information Processing Systems, 33:15651-15663, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[28] Shaohui Liu, Yinda Zhang, Songyou Peng, Boxin Shi, Marc Pollefeys, and Zhaopeng Cui. Dist: Rendering deep implicit" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20842" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "signed distance function with differentiable sphere tracing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2019-2028, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.135, + 0.47, + 0.204 + ], + "angle": 0, + "content": "[29] Xiaoxiao Long, Cheng Lin, Lingjie Liu, Wei Li, Christian Theobalt, Ruigang Yang, and Wenping Wang. Adaptive surface normal constraint for depth estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12849-12858, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.205, + 0.471, + 0.26 + ], + "angle": 0, + "content": "[30] Xiaoxiao Long, Cheng Lin, Peng Wang, Taku Komura, and Wenping Wang. Sparseneus: Fast generalizable neural surface reconstruction from sparse views. arXiv preprint arXiv:2206.05737, 2022. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.261, + 0.471, + 0.329 + ], + "angle": 0, + "content": "[31] Xiaoxiao Long, Lingjie Liu, Wei Li, Christian Theobalt, and Wenping Wang. Multi-view depth estimation using epipolar spatio-temporal networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8258-8267, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.33, + 0.471, + 0.385 + ], + "angle": 0, + "content": "[32] Xiaoxiao Long, Lingjie Liu, Christian Theobalt, and Wenping Wang. Occlusion-aware depth estimation with adaptive normal constraints. In European Conference on Computer Vision, pages 640-657. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.386, + 0.471, + 0.455 + ], + "angle": 0, + "content": "[33] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4460-4470, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.456, + 0.471, + 0.524 + ], + "angle": 0, + "content": "[34] Mateusz Michalkiewicz, Jhony K Pontes, Dominic Jack, Mahsa Baktashmotlagh, and Anders Eriksson. Implicit surface representations as layers in neural networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4743-4752, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.525, + 0.471, + 0.594 + ], + "angle": 0, + "content": "[35] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European conference on computer vision, pages 405-421. Springer, 2020. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.595, + 0.471, + 0.663 + ], + "angle": 0, + "content": "[36] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11453-11464, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.664, + 0.471, + 0.734 + ], + "angle": 0, + "content": "[37] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3504-3515, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.735, + 0.471, + 0.803 + ], + "angle": 0, + "content": "[38] Michael Oechsle, Songyou Peng, and Andreas Geiger. Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5589-5599, 2021. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.471, + 0.873 + ], + "angle": 0, + "content": "[39] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 165-174, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.471, + 0.9 + ], + "angle": 0, + "content": "[40] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865-5874, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.189 + ], + "angle": 0, + "content": "[41] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.19, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[42] Johannes L Schonberger, Enliang Zheng, Jan-Michael Frahm, and Marc Pollefeys. Pixelwise view selection for unstructured multi-view stereo. In European Conference on Computer Vision, pages 501-518. Springer, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.246, + 0.894, + 0.286 + ], + "angle": 0, + "content": "[43] Steven M Seitz and Charles R Dyer. Photorealistic scene reconstruction by voxel coloring. International Journal of Computer Vision, 35(2):151-173, 1999. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.287, + 0.892, + 0.356 + ], + "angle": 0, + "content": "[44] Jiaming Sun, Yiming Xie, Linghao Chen, Xiaowei Zhou, and Hujun Bao. Neuralrecon: Real-time coherent 3d reconstruction from monocular video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15598-15607, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.357, + 0.892, + 0.41 + ], + "angle": 0, + "content": "[45] Engin Tola, Christoph Strecha, and Pascal Fua. Efficient large-scale multi-view stereo for ultra high-resolution image sets. Machine Vision and Applications, 23(5):903-920, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.411, + 0.892, + 0.48 + ], + "angle": 0, + "content": "[46] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. Advances in Neural Information Processing Systems, 34, 2021. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.481, + 0.892, + 0.536 + ], + "angle": 0, + "content": "[47] Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. Mvsnet: Depth inference for unstructured multi-view stereo. In Proceedings of the European Conference on Computer Vision (ECCV), pages 767-783, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.537, + 0.892, + 0.605 + ], + "angle": 0, + "content": "[48] Yao Yao, Zixin Luo, Shiwei Li, Tianwei Shen, Tian Fang, and Long Quan. Recurrent mvsnet for high-resolution multi-view stereo depth inference. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5525-5534, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.606, + 0.892, + 0.659 + ], + "angle": 0, + "content": "[49] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems, 34, 2021. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.661, + 0.892, + 0.73 + ], + "angle": 0, + "content": "[50] Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Basri Ronen, and Yaron Lipman. Multiview neural surface reconstruction by disentangling geometry and appearance. Advances in Neural Information Processing Systems, 33:2492-2502, 2020. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.731, + 0.892, + 0.786 + ], + "angle": 0, + "content": "[51] Jingyang Zhang, Yao Yao, and Long Quan. Learning signed distance field for multi-view surface reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6525-6534, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.786, + 0.892, + 0.828 + ], + "angle": 0, + "content": "[52] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.828, + 0.892, + 0.898 + ], + "angle": 0, + "content": "[53] Heming Zhu, Yu Cao, Hang Jin, Weikai Chen, Dong Du, Zhangye Wang, Shuguang Cui, and Xiaoguang Han. Deep fashion3d: A dataset and benchmark for 3d garment reconstruction from single images. In European Conference on Computer Vision, pages 512-530. Springer, 2020. 2, 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.898 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20843" + } + ] +] \ No newline at end of file diff --git a/2023/NeuralUDF_ Learning Unsigned Distance Fields for Multi-View Reconstruction of Surfaces With Arbitrary Topologies/cda2b21c-85ea-47d7-952f-4f7b2568e13a_origin.pdf b/2023/NeuralUDF_ Learning Unsigned Distance Fields for Multi-View Reconstruction of Surfaces With Arbitrary Topologies/cda2b21c-85ea-47d7-952f-4f7b2568e13a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..bb0f62142c3ea775993f8d8d8241f05149a15f03 --- /dev/null +++ b/2023/NeuralUDF_ Learning Unsigned Distance Fields for Multi-View Reconstruction of Surfaces With Arbitrary Topologies/cda2b21c-85ea-47d7-952f-4f7b2568e13a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4180a88f2075f3f2da1dbf41f7750f6ab4db7f776472e677c431ac532d860f0a +size 1519216 diff --git a/2023/NeuralUDF_ Learning Unsigned Distance Fields for Multi-View Reconstruction of Surfaces With Arbitrary Topologies/full.md b/2023/NeuralUDF_ Learning Unsigned Distance Fields for Multi-View Reconstruction of Surfaces With Arbitrary Topologies/full.md new file mode 100644 index 0000000000000000000000000000000000000000..bb2cad831f62b5ecd5a6288375e55e06e2091502 --- /dev/null +++ b/2023/NeuralUDF_ Learning Unsigned Distance Fields for Multi-View Reconstruction of Surfaces With Arbitrary Topologies/full.md @@ -0,0 +1,416 @@ +# NeuralUDF: Learning Unsigned Distance Fields for Multi-view Reconstruction of Surfaces with Arbitrary Topologies + +Xiaoxiao Long $^{1,2\dagger}$ Cheng Lin $^{2*}$ Lingjie Liu $^{3}$ Yuan Liu $^{1}$ Peng Wang $^{1}$ Christian Theobalt $^{3}$ Taku Komura $^{1}$ Wenping Wang $^{4*}$ $^{1}$ The University of Hong Kong $^{2}$ Tencent Games + $^{3}$ Max Planck Institute for Informatics $^{4}$ Texas A&M University + +# Abstract + +We present a novel method, called NeuralUDF, for reconstructing surfaces with arbitrary topologies from 2D images via volume rendering. Recent advances in neural rendering based reconstruction have achieved compelling results. However, these methods are limited to objects with closed surfaces since they adopt Signed Distance Function (SDF) as surface representation which requires the target shape to be divided into inside and outside. In this paper, we propose to represent surfaces as the Unsigned Distance Function (UDF) and develop a new volume rendering scheme to learn the neural UDF representation. Specifically, a new density function that correlates the property of UDF with the volume rendering scheme is introduced for robust optimization of the UDF fields. Experiments on the DTU and DeepFashion3D datasets show that our method not only enables high-quality reconstruction of non-closed shapes with complex typologies, but also achieves comparable performance to the SDF based methods on the reconstruction of closed surfaces. Visit our project page at https://www.xxlong.site/NeuralUDF/. + +# 1. Introduction + +Reconstructing high-quality surfaces from multi-view images is a long-standing problem in computer vision and computer graphics. Neural implicit fields have become an emerging trend in recent advances due to its superior capability of representing surfaces of complex geometry. NeRF [35] and its variants [2, 26, 27, 36, 40, 52] have recently achieved compelling results in novel view synthesis. For each point in the 3D space, NeRF-based methods learn two neural implicit functions based on volume rendering: a volume density function and a view-dependent color function. Despite their successes in novel view synthesis, NeRF-based methods still struggle to faithfully reconstruct the ac + +![](images/708d98f8ae5d6eca6b29dd87cb20d25df7f26c6f32520a4926a6f6a1c7b52a11.jpg) +Figure 1. We show three groups of multi-view reconstruction results generated by our proposed NeuralUDF and NeuS [46] respectively. Our method is able to faithfully reconstruct the high-quality geometries for both the closed and open surfaces, while NeuS can only model shapes as closed surfaces, thus leading to inconsistent typologies and erroneous geometries. + +curate scene geometry from multi-view inputs, because of the difficulty in extracting high-quality surfaces from the representation of volume density. + +VolSDF [49] and NeuS [46] incorporate the Signed Distance Function (SDF) into volume rendering to facilitate high-quality surface reconstruction in a NeRF framework. However, as a continuous function with clearly defined inside/outside, SDF is limited to modeling only closed watertight surfaces. Although there have been efforts to modify the SDF representation by learning an additional truncation function (e.g., 3PSDF [5], TSDF [44]), their surface representations are still built upon the definition of SDF, thus they are not suitable for representing complex topologies. + +We therefore propose to employ the Unsigned Distance Function (UDF) to represent surfaces in volume rendering. + +With a surface represented by its zero level set without signs, UDF is a unified representation with higher-degree of freedom for both closed and open surfaces, thus making it possible to reconstruct shapes with arbitrary topologies. + +There are two major challenges in learning a neural UDF field by volume rendering. First, UDF is not occlusion-aware, while the formulation of volume rendering with distance fields requires to estimate surface occlusion for points. Considering a camera ray intersecting with a surface, SDF assumes that the surface is closed and distinguishes the inside/outside with signs, so that a negative SDF value on the ray clearly indicates an occluded point inside the surface. In contrast, UDF does not impose the closed surface assumption and always gives non-negative values along the ray. Hence, the UDF value of a point alone cannot be used to infer occlusion. + +The second challenge is that the UDF is not differentiable at its zero-level sets. The non-differentiability at the zero-level sets imposes barriers in the learning of UDF field. The gradients are ill-defined near the iso-surface, leading to difficulty in optimization. Consequently, the distance field surrounding the iso-surface is not accurate and the exact zero-level set of a learned UDF cannot be identified in a stable manner. + +In this paper, we present a novel method for learning neural UDF fields by volume rendering. We introduce a density function that correlates the property of the UDF representation with the volume rendering process, which effectively tackles the aforementioned challenges induced by the unsigned representation and enables robust learning of surfaces with arbitrary topologies. Experiments on DTU [15] and DeepFashion3D [53] datasets show that our method not only enables the high-quality reconstruction of non-closed shapes with complex topologies, but also achieves comparable performance to the SDF based methods on the reconstruction of closed surfaces. + +Our contributions can be summarized as: + +- We incorporate the UDF into volume rendering, which extends the representation of the underlying geometry of neural radiance fields. +- We introduce an effective density function that correlates the property of the UDF with the volume rendering process, thus enabling robust optimization of the distance fields. +- Our method achieves SOTA results for reconstructing high-quality surfaces with various topologies (closed or open) using the UDF in volume rendering. + +# 2. Related Work + +Classic multi-view 3D reconstruction. 3D reconstruction from multiple images is a fundamental problem in computer vision. Traditional reconstruction methods usually adopt discrete representations, such as voxel grids [16, + +17, 19, 23, 43, 44], 3D point clouds [10, 25], and depth maps [4, 11, 13, 29, 31, 32, 42, 45, 47, 48]. The depth map representation is flexible and appropriate for parallel computation, so it is the most widely used representation. The depth map based methods first compute depth maps by analyzing the correlation between multiple input images, and then fuse the depth maps into a point cloud, which is then further processed into a mesh via meshing algorithms, such as Poisson surface reconstruction [20], ball-pivoting [3] and Delaunay triangulation [24]. However, the traditional methods are sensitive to reflections, image noises, and varying illuminations, thus usually yielding incomplete and noisy reconstructions. + +Neural rendering based surface reconstruction. The recent advances on neural implicit representations have achieved promising results in multi-view 3D reconstruction [8,18,22,28,37,38,46,49-51]. These methods typically represent the surfaces to be reconstructed as Signed Distance Function (SDF), and utilize surface rendering or volume rendering to render the pixel colors of the input images for optimizing the SDF fields. The surface rendering based methods [37, 50] require extra masks for optimization and cannot handle objects with complex structures. The volume rendering based works [8, 38, 46, 49, 51] get rid of the extra masks and achieve more accurate reconstructions. However, these methods adopt SDF as representation so they are limited to closed surfaces and their performance degrade a lot for non-closed surfaces. + +Surface representations. There are various representations to model 3D geometric surfaces. NeRF [35] uses a density field to represent the underlying geometry of a target scene. Although NeRF achieves huge successes in novel view synthesis, it's difficult to extract high-quality geometry from such a density representation. Signed Distance Function (SDF) is widely used for geometry modeling and surface reconstruction [1,6,8,12,33,34,38,39,46,49,51], but SDF is only applicable to objects with closed surfaces. To deal with open surfaces, there are some efforts to learn an additional truncation function (e.g. 3PSDF [5], TSDF [44]). However, their representations are just built on the modification of SDF, and thus are still subject to flexibility of SDF and struggle to model complex surfaces. NDF [7] uses a neural UDF field to reconstruct surfaces, but they explicitly rely on 3D point clouds for supervising the optimization. + +# 3. Method + +Our goal is to reconstruct a surface $S$ from a set of posed input images $\{\mathcal{I}_k\}$ of a 3D model. We represent the surface as the zero level set of an Unsigned Distance Function (UDF), and propose a novel volume rendering scheme to learn the UDF. We will introduce our method in three parts: 1) We discuss the relationship between the unsigned distance function and the density function of the volume ren + +![](images/0c34ec97745c9acfd781109e16531c66efe7b61efd60c57536fdb2d5a6d23a70.jpg) +(a) Biased density function +Figure 2. Illustration of (a) biased density function and (b) unbiased density function for UDF. Suppose that the ray has one single intersection point $t^*$ with the surface. (a) If the density function is modeled as a bell-shaped probability distribution, the weight function attains a local maximum before $t^*$ . (b) If the density function is modeled as a monotonically increasing function, the weight function can attain its local maximum at $t^*$ . + +![](images/bf32a3758bdcbf8280746fc26f7131225aca1456213c3dee09f76453afc9bf86.jpg) +(b) Unbiased density function + +dering equation; 2) To derive a density function induced by the UDF, we propose a differentiable indicator function to transform the UDF into its associated density field; 3) In addition to the loss terms commonly used in the prior works, we introduce an iso-surface regularizer to improve the stability of the UDF field near the zero level set. This UDF-based volume rendering scheme enables robust optimization of NeuralUDF for reconstructing both closed surfaces and open surfaces. + +# 3.1. Volume Rendering + +Scene representation. Our method NeuralUDF represents the scene to be reconstructed by two neural implicit functions: 1) an unsigned distance function $f_{u}(\mathbf{x})$ that maps a 3D point $\mathbf{x} \in \mathbb{R}^3$ to its unsigned distance to the closest object surface; 2) a view-dependent color function $c(\mathbf{x},\mathbf{v})$ that encodes the color associated with a point $\mathbf{x} \in \mathbb{R}^3$ and a viewing direction $\mathbf{v} \in \mathbb{R}^3$ . + +Volume rendering. We consider a ray $\mathbf{r}$ emanating from the camera center $\mathbf{o} \in \mathbb{R}^3$ in direction $\mathbf{v} \in \mathbb{R}^3$ ( $\| \mathbf{v} \| = 1$ ), which is parameterized by $\{\mathbf{r}(t) = \mathbf{o} + t\mathbf{v} \mid t \geq 0\}$ . Volume rendering is used to calculate the color $C$ of the pixel corresponding to the ray $\mathbf{r}$ . The accumulated color of the camera ray $\mathbf{r}$ is given by $C(\mathbf{r}) = \int_0^{+\infty} T(t) \sigma(t) \mathbf{c}(\mathbf{r}(t), \mathbf{v}) \, \mathrm{d}t$ , where $\sigma(t)$ is the volume density, $T(t) = \exp \left(-\int_0^t \sigma(u) \, \mathrm{d}u\right)$ denotes the accumulated transmittance along the ray, and the product of $T(t) \sigma(t)$ is the weight function of a point, denoted as $w(t)$ , i.e., $w(t) = T(t) \sigma(t)$ . + +# 3.2. Modeling Density with UDF + +In the NeRF framework, the density $\sigma(t)$ is predicted by an MLP, where the points near the surfaces having larger density values and the points far from the surfaces having smaller densities. In order to apply volume rendering to learning distance field $f$ , we need to define a mapping $\Omega$ to transform the distance function $f(\mathbf{r}(t))$ into a density field + +$\sigma (t)$ , that is $\sigma (t) = \Omega (f(\mathbf{r}(t)))$ + +Unbiased volume density function. Inspired by NeRF, we may consider defining the density function $\sigma(t) = \Omega(f(\mathbf{r}(t)))$ as a bell-shaped distribution model centered at zero, so that the surface points intersecting with a ray (UDF values are zero) have the maximum density values. However, as illustrated in Fig 2, if the mapping function is straightforwardly defined in this way, a noticeable bias will be induced in the weight function $w(t)$ since its local maximum is not at the intersection point $t^*$ where the ray intersects with the surfaces. This can result in conspicuous inaccuracy in the reconstructed surfaces, which is demonstrated in [46]. + +To define an appropriate mapping, we need to first review the relationship between the density function $\sigma(t)$ and a distance field $f(t)$ . Suppose that there is only one single intersection of the ray and a surface. It is shown in NeuS [46] that, to make the weight function $w(t)$ unbiased, the density function should be in the form of a monotonically increasing function, which attains the maximum at an infinite point rather than at the intersection point. + +Accordingly, NeuS [46] chooses to use a Sigmoid function $\Phi_{\kappa}(x) = (1 + e^{-\kappa x})^{-1}$ to achieve the goal. The accumulated transmittance $T(t)$ is then defined as $T(t) = \Phi_{\kappa}(f_s(t))$ , where $f_{s}(t)$ is the signed distance function (SDF). The density function $\sigma_{s}(t)$ induced by SDF is formulated as + +$$ +\sigma_ {s} (t) = \Omega_ {s} \left(f _ {s} (\mathbf {r} (t))\right) = \kappa | \cos (\theta) | \left(1 - \Phi_ {\kappa} \left(f _ {s} (\mathbf {r} (t))\right)\right), \tag {1} +$$ + +where $\kappa$ is the parameter of Sigmoid function $\Phi_{\kappa}(x)$ , and $\theta$ is the angle between the view direction $\mathbf{v}$ and the surface normal $\mathbf{n}$ . + +Unbiased volume density for UDF. However, the unbiased density formulation induced by SDF [46] cannot be directly applied to UDF. The formulation of UDF near the surface can be given by + +$$ +f _ {u} (\mathbf {r} (t)) = \left\{ \begin{array}{l l} | \cos (\theta) \cdot (t - t ^ {*}) |, & t \neq t ^ {*} \\ 0, & t = t ^ {*} \end{array} \right. \tag {2} +$$ + +where $\theta$ is the angle between the view direction $\mathbf{v}$ and the surface normal $\mathbf{n}$ . Obviously, the UDF value is not monotonous along a ray when it crosses a surface, where the mapping $t\rightarrow f_u(\mathbf{r}(t))$ is non-injective. This leads to difficulties in modeling the monotonicity between $t$ and the density function $\sigma (t) = \Omega (f_u(\mathbf{r}(t)))$ , given that an unbiased density function is expected to be monotonically increasing. + +To leverage the unbiased property derived from the monotonic density function, we need to model the injective mapping from $t$ to the density function $\sigma_{u}(t)$ with the formulation of UDF $f_{u}(\mathbf{r}(t))$ . Since the monotonicity of UDF along a ray only changes at the intersection point $t^{*}$ , we can + +use a visibility indicator function $\Psi(t)$ to reflect the existence of the intersection and thus capture the monotonicity change: + +$$ +\Psi (t) = \left\{ \begin{array}{l l} 1, & t < t ^ {*} \\ 0, & t > t ^ {*}. \end{array} \right. \tag {3} +$$ + +The indicator function $\Psi(t)$ indicates the visibility of the points sampled along the ray $\mathbf{r}$ , where 1 means that the point is visible in the free space while 0 means that the point is occluded. Note this just gives a simple hint for the indicator function, while the actual formulation will be derived later. + +With $\Psi (t)$ , given an unbiased density function $\sigma (t) = \Omega_s(f(\mathbf{r}(t)))$ that is monotonically increasing, the density function for UDF can be defined as + +$$ +\sigma_ {u} (t) = \Psi (t) \Omega_ {s} \left(f _ {u} (\mathbf {r} (t))\right) + (1 - \Psi (t)) \Omega_ {s} \left(- f _ {u} (\mathbf {r} (t))\right), \tag {4} +$$ + +where $\Omega_s$ can be any monotonic density function that is unbiasedly designed, while in this paper we directly use the formulation of Eq.1. Since unsigned distance function $f_u$ is not differentiable at $t^*$ , the domain scope of $\sigma_u(t)$ is defined as $\{t \in \mathbb{R}^1, t \neq t^*\}$ . + +Zero level set. The visibility indicator function $\Psi$ plays a vital role in the formulation of volume density for UDF, because it essentially encodes the distribution of the geometric surface. The key to the visibility is to find the intersection point $t^*$ of surfaces and the ray, whose UDF value is zero, i.e., $f_{u}(\mathbf{r}(t^{*})) = 0$ . + +It is nontrivial to find the zero level set for a UDF field. In practice, it's impossible to locate a point with exact zero UDF value, and a trade-off to balance accuracy and efficiency is to find points whose UDF values are smaller than a pre-defined threshold $\epsilon$ . As a result, the approximated surfaces will be relaxed from exact zero level set to be a $\epsilon$ -bounded surface band $S = \{\mathbf{x} \in \mathbb{R}^3 \mid f_u(\mathbf{x}) < \epsilon\}$ . + +However, such a definition of the indicator function involving $\epsilon$ bounding brings about the following difficulties in learning the UDF field. This indicator function design with $\epsilon$ bounding operation incurs a set of barriers that impede learning the UDF field: 1) choosing an appropriate $\epsilon$ is difficult; 2) a fixed $\epsilon$ cannot be adjusted to adapt the UDF field in different optimization stages; 3) the step-like indicator function is not differentiable. These limitations make the optimization of UDF unstable and easy to collapse. + +Visibility indicator function. To overcome the limitations mentioned above, we leverage probabilistic models to make the indicator function $\Psi(t)$ differentiable. Therefore, we propose to approximate the surfaces by a soft probabilistic distribution instead of using a hard $\epsilon$ bound. We measure the probability of a point $\mathbf{r}(t)$ being on a surface, named as surface existence probability $h(\mathbf{r}(t))$ abbreviated as $h(t)$ . This can be achieved by transforming the UDF values into probabilities via a bell-shaped distribution centered at 0. Here + +we use a logistic distribution $\phi_{\beta}(x)$ to define the probability $h(t_i)$ : + +$$ +h \left(t _ {i}\right) = 1 - \exp \left(- \alpha \cdot \phi_ {\beta} \left(f _ {u} \left(\mathbf {r} \left(t _ {i}\right)\right)\right) \cdot \delta_ {i}\right), \tag {5} +$$ + +$$ +\phi_ {\beta} (x) = \beta e ^ {- \beta x} / \left(1 + e ^ {- \beta x}\right) ^ {2} \tag {6} +$$ + +where $t_i$ is the $i$ -th sample in the ray $\mathbf{r}$ , $h(t_i)$ denotes the surface existence probability of the sample $t_i$ , $\delta_i = t_{i+1} - t_i$ is the distance between adjacent samples, and $\phi_\beta(t)$ is a logistic distribution function. Note that the standard deviation of $\phi_\beta(x)$ is given by $1 / \beta$ , where $\beta$ is a trainable parameter, and $1 / \beta$ approaches zero as the network training converges; $\alpha$ is a constant scalar to scale the distribution. + +After obtaining the surface existence probabilities of the sampled points on the ray $\mathbf{r}$ , we can measure the probability of whether a point is occluded by any surfaces or not. Since local UDF values do not indicate occlusion, we accumulate the probabilities of the points $\mathbf{r}(t_j)$ ahead of a query point $\mathbf{r}(t_i)$ along the ray: + +$$ +\hat {\Psi} \left(t _ {i}\right) = \prod_ {j = 1} ^ {i - 1} \left(1 - h \left(t _ {j}\right)\right), \tag {7} +$$ + +where $h(t_i) \in [0,1]$ measures the probability of a point being on the surfaces. From this equation, we can see that $\hat{\Psi}(t_i)$ will drop dramatically to zero once there is a point that has high surface existence probability on the ray. Compared to using a hard bound, the probability accumulation strategy makes the indicator function $\hat{\Psi}(t)$ differentiable and learnable, which adapts to learning various surface geometries. The function $\hat{\Psi}(t)$ will become sharper when $1/\beta$ approaches 0 as the network training converges, resulting in more accurate approximation of the zero level set. + +Gradient-aware indicator. Although the formulation of $\hat{\Psi}(t)$ in Eq. 7 is effective for identifying surfaces in learning the UDF field, it can induce bias in the indicator. As shown in Figure 3, the surface existence probability $h(t_i)$ increases when the ray approaches a surface, and the accumulated probability $\hat{\Psi}(t_i)$ can attain zero before the point $t^*$ where the ray intersects with the surface. + +To alleviate this problem, we modify the indicator function $\hat{\Psi}(t)$ with a masking function $m(t)$ + +$$ +\hat {\Psi} (t _ {i}) = \prod_ {j = 1} ^ {i - 1} \left(1 - h \left(t _ {j}\right) \cdot m \left(t _ {j}\right)\right) \tag {8} +$$ + +$$ +m \left(t _ {j}\right) = \left\{ \begin{array}{l l} 0, & \cos \left(\theta_ {j + 1}\right) < 0 \\ 1, & \cos \left(\theta_ {j + 1}\right) \geq 0 \end{array} \right. \tag {9} +$$ + +where $\theta_{j + 1}$ denotes the angle between the ray direction and the gradient of $f_{u}(\mathbf{r}(t_{j + 1}))$ . This formulation exploits the geometric property of UDF surface, that is, the gradient direction of UDF flips across its zero level set. Therefore, + +![](images/b81c1fe9f1cc7ab184aab0d5c24b584cb184cd43a08b6f821b46167b3d02ee70.jpg) +(a) Naive indicator func. + +![](images/e23405a73e68d984869ff1184cb8ce2eebe0f738205dac767b37c112adc60de5.jpg) +(b) Grad-aware indicator func. +Figure 3. Illustration of (a) naive visibility indicator function and (b) gradient-aware visibility indicator function. The naive indicator function will attain zero before the ray intersects with surface point $t^*$ . Incorporating the UDF gradients information, the gradient-aware indicator function attains zero exactly at the intersection point $t^*$ . + +the points with negative $\cos (\theta_{j + 1})$ are located ahead of the intersection point $t^{*}$ ; we mask out these points and only accumulate the surface existence probability once the ray reaches the surface. The final formulation jointly leverages the insights from both UDF value and gradient, leading to more reliable surface identification. + +# 3.3. Optimizing UDF Fields + +Iso-surface regularizer. With the volume rendering scheme mentioned above, NeuralUDF learns a UDF field from the images by minimizing the differences between rendered images and ground truth images. However, unlike SDF representation that is defined for closed surface, the intrinsic property of the UDF field makes it more difficult to optimize. First, the UDF has a higher degree of freedom accommodating arbitrary topologies instead of only closed surfaces. Second, the UDF is not differentiable at the zero level set; the values and gradients of the UDF field close to the zero level set are therefore hard to learn. This usually results in a wide strip around the iso-surface where the UDF values are distributed irregularly. Therefore, we introduce a simple yet effective geometric regularizer term during training to improve the stability of the UDF field: + +$$ +\mathcal {L} _ {\text {r e g}} = \frac {1}{M} \sum_ {k, i} \exp (- \tau \cdot f _ {u} (\mathbf {r} _ {k} (t _ {k, i}))), \tag {10} +$$ + +where $\tau$ is a constant scalar to scale the UDF values, and $M$ is the number of sampled rays per optimization step. This formulation effectively precludes the UDF values of the non-surface points from being zero, which thus encourages the simplicity and clearness of the zero level set in the field. The experiments show that this iso-surface regularizer excels at improving the surface quality (see Sec. 4.3). + +Loss functions. We train NeuralUDF by enforcing the consistency of the rendered colors and the ground truth colors of the input images without using 3D ground truth shapes. Specifically, we optimize the neural networks and two trainable probabilistic parameters $\kappa$ of Eq. 1 and $\beta$ of Eq. 6 by randomly sampling a batch of pixels. + +The overall loss function is defined as a weighted sum of several loss terms: + +$$ +\mathcal {L} = \mathcal {L} _ {\text {c o l o r}} + \lambda_ {0} \mathcal {L} _ {\text {p a t c h}} + \lambda_ {1} \mathcal {L} _ {\text {e i k}} + \lambda_ {2} \mathcal {L} _ {\text {r e g}} (+ \gamma \mathcal {L} _ {\text {m a s k}}). \tag {11} +$$ + +where $\mathcal{L}_{\mathrm{color}}$ is the L1 color loss to minimize the differences between the rendered pixel colors and the ground truth pixel colors of the input images. Following the prior works [46, 49, 50], the Eikonal term $\mathcal{L}_{eik}$ is used to regularize the neural distance field to have unit norm of gradients, and the optional mask loss term $\mathcal{L}_{\mathrm{mask}}$ can be included if extra masks are provided. $\mathcal{L}_{\mathrm{patch}}$ is a patch-based color loss term introduced in [30], which renders the colors of a local patch of an input image using patch blending strategy and then enforces the consistency of the rendered colors and ground truth colors at the patch level. (See the supplementary materials for the details.) + +# 4. Experiments + +# 4.1. Experimental settings + +Datasets. We conduct evaluations on the DTU [15] dataset and Deepfashion3D [53] dataset. DTU is a multi-view dataset containing objects with closed surfaces, and each scene consists of 49 or 64 images with $1600 \times 1200$ resolution. We use the same 15 scenes following the prior works for comparison. To validate the effectiveness of our method on objects with open surfaces, we adopt Deepfashion3D dataset for evaluation. Deepfashion3D is a large-scale real 3D garment dataset containing 563 diverse cloth items, and provides the ground truth point clouds of the captured garments. To facilitate the multi-view 3D reconstruction task, we randomly choose 12 garments from different categories of the dataset, and then scale the point clouds to be in the unit sphere. We render each point cloud into 72 images with $1024 \times 1024$ resolution. The rendered images are used for reconstruction. We further collect a set of synthetic plant objects for further qualitative comparisons. Each plant object is rendered into 324 images with $256 \times 256$ resolution. Implementation details. We use MLPs to model the unsigned distance field and color radiance field. We use the Adam optimizer with a global learning rate 5e-4 for the network training, specially we set the learning rate of the UDF MLP as 1e-4. We sample 512 rays per batch and train our model for 300k iterations on a single NVIDIA RTX 2080Ti GPU. Our network architecture and initialization scheme are similar to those of prior works [46, 49, 50]. We leverage MeshUDF [14], an extended marching cube algorithm + +![](images/e242aabf8e2646899a0f44bba7960b2ea0cb6001c9b04d79823652da3abae17f.jpg) + +![](images/299a72f5e0baeb12419eea0fa76d9c0f057678222ca4b051b4c0f2db4f9baf17.jpg) + +![](images/2709eed36efc1360d8a22ba02e98a7fb12ffff7487d824f540ee63499dc6655f.jpg) +Reference Images +Figure 4. Qualitative comparisons with SOTAs on DeepFashion3D [53] dataset. The GT meshes are reconstructed from the GT point clouds provided by the DeepFashion3D dataset via Ball Pivoting meshing algorithm [3]. Thanks to the UDF representation, NeuralUDF accurately recovers the garments with open surfaces. However, the SDF-based methods can only model the garments as closed surfaces, thus leading to inconsistent topologies and erroneous geometries. (See supplementary materials for more results.) + +![](images/c0547579cc7ed34a37e939b8df7be412cf4a7aecad236a606931a999e88d16d2.jpg) + +![](images/4968797747c75c8c49414bcebaa9e56c008868bc64b3870bddf15011a92481ed.jpg) + +![](images/6970deb48840a414a41653849168bf5d08d4588efb5e36deea701c48f1479d5f.jpg) +GT meshes + +![](images/9ffd3458043d718de8565dcb9313e15749704e03da1b0395f8d1056760122f4e.jpg) + +![](images/6a9e7f918528cf3ebad80496ba3397f1636c824462e44d83d092f846e95ea2cd.jpg) + +![](images/d8af9485e5e787590b24da6d063c35bed6efece4fa50df85e9a840eec44822ad.jpg) +Ours + +![](images/2f2e904f26b537e612f76d8fcbcf36c0f66b84136c9d31a666cb3ce7414eed37.jpg) + +![](images/6030a40666a7272ba9e263344e497fa2f0c6f08cf205754cdd069bc86b4cdbef.jpg) + +![](images/19b7bdea0f27bda7885bd733afb14088d387e816dc831591e7655ea8512d90e1.jpg) +NeuS [46] + +![](images/4084c6e7e2383ca5ee4bb9309fce142e7cd06c825cf986163f344b40502a6db7.jpg) + +![](images/d25fafb8dc21c7f4c88c572bfe127a652a16c998aad1a180f798443107dfde85.jpg) + +![](images/d8abcb8765f197281e7f3e3474629167d8036b90b07fd97cd16976eead352335.jpg) +VolSDF [49] + +![](images/cc1c7c4339160dd12ddfc76d1cca44dc4f89f5b0191c7c9695783a697add3561.jpg) + +![](images/6dde80a8a022de71a1175eab430a119b525219ba27d715a0b85057c5c89230de.jpg) + +![](images/2ace1ce08f1096aa78f01f094bbb21813de616247a1b11e51358e5ab74024a34.jpg) +NeRF [35] + +
ScanLS-C0SS-D0NS-D0NS-C0Skirt0LS-D0Trouser0NS-D1Trouser1LS-C1Skirt1SS-C0Mean
Colmap [41]2.952.913.583.063.233.263.093.112.953.162.972.953.10
IDR†[50]10.858.205.009.636.766.336.323.323.756.082.447.026.31
Nerf [35]6.387.845.035.454.697.864.824.636.047.178.246.896.25
VolSdf [49]5.924.794.548.577.025.966.974.367.248.737.748.846.72
Neus [46]3.184.824.784.993.735.715.892.215.893.602.445.134.36
NeuralWarp [9]4.7111.245.408.564.105.835.792.002.076.509.616.946.06
Ours1.922.052.361.581.334.112.471.501.632.472.162.151.97
+ +$\dagger$ Optimization using extra object masks. +LS-C denotes Long-Sleeve Coat, SS-C denotes Short-Sleeve Coat, NS-C denotes None-Sleeve Coat, LS-D denotes Long-Sleeve Dress, SS-D denotes Short-Sleeve Dress, NS-D denotes None-Sleeve Dress. + +Table 1. Quantitative evaluation results on Deepfashion3D [53] dataset. + +designed for UDF, to extract explicit mesh from the learned UDF field. + +A hierarchical sampling strategy is used to sample points along a ray in a coarse-to-fine manner for volume rendering. We first uniformly sample points on the ray and then iteratively conduct importance sampling [35] to sample more points on top of coarse probability estimation. The probability of coarse sampling is computed based on the density function $\sigma_u(t)$ with fixed and predefined parameters $\kappa$ and $\beta$ , while the probability of fine sampling is computed with the learned $\kappa$ and $\beta$ . + +Baselines. We compare NeuralUDF with the following state-of-the-art baselines: 1) a widely used classic reconstruction method COLAMP [41], which we use to reconstruct a mesh from the point cloud generated by COLMAP via Screened Poisson Surface Reconstruction [21]; 2) neural implicit methods utilizing surface rendering, IDR [50] and UNISURF [38]; 3) neural implicit methods utilizing + +volume rendering, NeuS [46], VolSDF [49] and NeuralWarp [9]; 4) NeRF [35]. Although NeRF is not designed for surface reconstruction, we include NeRF for comparisons since its density representation enables successful novel view synthesis on objects with arbitrary topologies. Following NeuS [46], we use a density threshold of 25 on DTU (50 on Deepfashion3D) to extract surfaces from NeRF for comparisons. Note that, IDR takes extra masks as input for optimization while ours and the other methods do not. + +# 4.2. Comparisons + +Evaluation on non-closed objects. We conduct evaluations on DeepFashion3D dataset whose objects are non-closed, and measure the Chamfer Distances $(10^{-3})$ between the predicted meshes with ground truth point clouds. As shown in Table 1, our method outperforms the other methods by a large margin. This is because our method can accurately reconstruct the surfaces with arbitrary topologies + +![](images/c679f9a0b7dca0b6723e5c55d9bfcfe9fbedac314565d9264bab1ac7fca914f8.jpg) + +![](images/492d5cb01e8f4ec4a3f0c8b2c8ff094b166ad5d1be6c560966f1809d68806502.jpg) + +![](images/7df5b1a4d6d4c1aa652f7e3e254cc716ca478e12aaf4cc50c16a3a65578f7ab6.jpg) +Reference Images +Figure 5. Qualitative comparisons with SOTAs on DTU [15] dataset.(See the supplementary materials for more results.) + +![](images/5cbc6191404c5e142cd74502f356e119366045fe351b214c091bc9ff3925ddf0.jpg) + +![](images/08bc81eaaa0f082acbe792256ee2e7a288d4a613732cb4d2798c1dbba8fe59d3.jpg) + +![](images/128346cefeb343db08d4663ce8e62cb5b21752df92ff3002967b074f86510af9.jpg) +Ours + +![](images/e25b2b7f2395985ae9c770d86947199343550c60abae68507a27b321daf1f41a.jpg) + +![](images/408f7f76a4f2343a997affcb7b81f3332173aacb21acbaf4f058e749da798540.jpg) + +![](images/e839afe1720a683ec56a01d090738901224461d9dab5ec062738dcef4b160032.jpg) +NeuS [46] + +![](images/3171d1bf9b8b9481cd1f8c64c05031f98880259aa0d09908127361f755bd00b2.jpg) + +![](images/f0b7375b52d41b0b43ab10fc53a8816413904bda949bfe8e54fd435818d30ce0.jpg) + +![](images/7c6148ef0de344de72a366ec523540a6a4bf1aeb846fb4c6d36b2d9091233f54.jpg) +VolSDF [49] + +![](images/001ec6c4bba3259fac69d85d429761d68df4211302fb6f6501a5ef9646abab13.jpg) + +![](images/66b7334e3214922b4bfd6d9ca47560bf9aa901b5e41da365c6a931182c065aed.jpg) + +![](images/7b84905db6a1d7532728af65eaee0da712ef56c6ff04c02dace75ecbe3142dfc.jpg) +NeRF [35] + +![](images/99b19a304992c3da9f2223019f8b55c9b4aac4317eb2a12e25434d9d65972f4c.jpg) + +![](images/8c4c75f75a1dab9cbddb85333073be17e792a370c33ea1b36fe37dc94c038b10.jpg) + +![](images/a92aaa181450ea1df784b097f1cad301ab19009f457f00147db230cc9b46c854.jpg) +IDR [50] + +
Scan243740556365698397105106110114118122Mean
Colmap [41]0.812.050.731.221.791.581.023.051.402.051.001.320.490.781.171.36
IDR†[50]1.631.870.630.481.040.790.771.331.160.760.670.900.420.510.530.90
Nerf [35]1.901.601.850.582.281.271.471.672.051.070.882.531.061.150.961.49
UniSurf [38]1.321.361.720.441.350.790.801.491.370.890.591.470.460.590.621.02
VolSdf [49]1.141.260.810.491.250.700.721.291.180.700.661.080.420.610.550.86
Neus [46]1.371.210.730.401.200.700.721.011.160.820.661.690.390.490.510.87
NeuralWarp [9]0.490.710.380.380.790.810.821.201.060.680.660.740.410.630.510.68
Ours0.691.180.670.440.900.660.671.320.940.950.570.860.370.560.550.75
+ +$\dagger$ Optimization using extra object masks. + +Table 2. Quantitative evaluation results on DTU [15] dataset. + +while the SDF based methods can only model the garments as closed surfaces and thus exhibit large errors. In Figure 4, the SDF based methods including NeuS, VolSDF and IDR enforce strong assumption on the closeness of the target shape and attempt to model all the shapes as closed surfaces, thus inevitably leading to large errors and inconsistent topologies for the garments. For example, for the first row, VolSDF attempts to use a double-layered surface to represent the thin cloth; optimizing this tiny SDF band is difficult and the results contain many noticeable holes. For the second row, NeuS gives a closed surface for the garment model while the cuffs are even incorrectly connected. Note although IDR takes extra masks for optimization, the reconstructed results still exhibit large quantitative errors. + +Evaluation on closed objects. We further perform evaluation on DTU dataset whose objects are closed and watertight, and measure the Chamfer Distances between the predicted meshes and ground truth point clouds. Although our method adopts UDF as representation and does not impose closed surface assumption, our method still achieves comparable performance with the SDF based methods that are particularly designed for closed shapes. + +As shown in Table 2, in terms of the averaged Chamfer Distance, NeuralUDF outperforms most of the SOTA methods except the most recent work NeuralWarp [8]. Neural- + +Warp is a fine-tuning method that further optimizes the surfaces obtained from VolSDF [49] by minimizing the differences between the warped source images and the reference image. However, NeuralWarp heavily relies on the reconstructed geometry provided by other methods to generate reasonable warped source images, so it cannot be trained from scratch. The qualitative comparisons with SOTAs are provided in Figure 5, where our reconstruction results contain fewer artifacts than the SDF based methods, e.g., the roof of the House and the Apple. + +# 4.3. Ablations and Analysis + +Ablation studies. We conduct ablation studies (Figure 6 and Figure 7) to investigate the individual contribution of the important designs of our method. The first key design is the gradient-aware visibility indicator function for effectively indicating the surfaces. For convenience, we call the visibility indicator function that is not gradient-aware as naive indicator function. As shown in Figure 6, the reconstruction result using the naive indicator function contains larger errors than the result using the gradient-aware indicator function, where the Chamfer Distance (CD) of the two reconstructions are 1.31 and 0.75 respectively (the lower is better). We can also see that the reconstruction using the naive function lacks geometric details, where the teeth and + +![](images/7f7e8d1d4201e5d61e877a0779d333081ce57048c1f0c6a0b4f0eeeacefb41a2.jpg) +(a) Naive indicator function + +![](images/31c1d23292fc68f8f429730d3cc5d004e5b21e062ffce9f65db8a4248c5617a5.jpg) +(b) Gradient-aware indicator function + +![](images/85e0c2f0093ed06b50b4809ccd74bdf97aae13ac2edf1c8988bdcdec04a00355.jpg) +(c) Patch blending + +eye holes of the Skull are not reconstructed accurately. This is because the naive indicator function has inherent bias and the measured visibility is not consistent with the true surface intersection, thus leading to inaccurate reconstruction. + +We leverage an iso-surface regularizer to make the UDF field more stable and accurate around zero level sets. The reconstructions with or without the regularizer are shown in Figure 7. Although the Chamfer Distances of the two results are close, the mesh without the regularizer is much more noisy and has noticeable artifacts while the mesh with the regularizer is much cleaner and smoother. To give an in-depth analysis, we plot the UDF values and the cosine values of the angle between the ray direction and UDF gradients along a selected ray (the marked red point of (b,c)). We can see that, without the regularizer, the signs of the $\cos (\theta)$ values change multiple times around the zero level set, which indicates the gradients are not stable. In contrast, with the regularizer, the gradients are more stable, where the gradient direction only flips at the surface intersection. + +Also, inspired by [30], the patch blending loss term is included in the optimization. With this term, the reconstruction can be further improved from 0.75 to 0.66 in terms of Chamfer Distance. + +More types of shapes. We additionally show results on several challenging objects with complex open surfaces. As shown in Figure 8, our method successfully reconstructs the challenging plant models with non-manifold geometries, while the reconstruction results of NeuS [46] are incomplete and contain numerous holes. This is because SDF is essentially not a suitable representation for many topologies like the thin leaves. Therefore, optimizing a SDF field for these kinds of models is deserved to be difficult and the reconstructed surfaces are naturally erroneous. + +Limitations. Compared to the SDF based methods, NeuralUDF will degrade on textureless objects that lack enough distinguishable features, and produces smoother results that may lack some geometric details. This is because NeuralUDF does not impose closed surface assumption and is not differentiable at zero-level sets (see details in the Supp.). + +![](images/de32a3cd349c56e222bf269a1a0ac82360820e3ec4702daae15477e052b1f349.jpg) + +![](images/635b208294a045cd28b474ab370f80513ebacb52d8d35306bcfc2672c6d17608.jpg) +CD: 2.47 + +![](images/625fab0d2d5831f3392c9df5166a1c38592304c7f996fe6ab1573325c0cd3934.jpg) +CD: 2.49 (c) w/o regularizer + +![](images/60a51f16ca4e5e28aa130b839df915df0823cc67fffdd6e727b9384b1538d78f.jpg) +Figure 6. Ablation studies on gradient-aware visibility indicator function and patch blending loss term. +(a) Reference Image +(d) The curves of UDF values and $\cos (\theta)$ values +Figure 7. Ablation study on the iso-surface regularizer. + +![](images/8adaa7428266edcb54815f623de0c08df91c918d2ba3ba47f82612d465b15f5b.jpg) + +![](images/7c0345feff3e7963df5c10345d4c095375a37f2e189e1d0a7444fac532dddcc7.jpg) +Figure 8. Qualitative comparisons with NeuS on two plants. + +# 5. Conclusions + +We introduce NeuralUDF, a NeRF-based method that uses neural UDF fields for multiview reconstruction of surfaces with arbitrary topologies. Our key idea is to introduce a novel volume density function that effectively correlates the UDF representation with the volume rendering scheme to learn neural UDF fields. The experiments show that, unlike the SDF based works that can only model closed shapes, our method can reconstruct more types of shapes with either open or closed surfaces, thus extending the underlying geometric representation of neural volume rendering to various topologies. + +# Acknowledgements + +Xiaoxiao Long is supported by Hong Kong PhD Fellowship Scheme. Taku Komura is partly supported by the Hong Kong ITF Grant ITS/319/21FP. Lingjie Liu and Christian Theobalt have been supported by the ERC Consolidator Grant 4DReply (770784). + +# References + +[1] Matan Atzmon and Yaron Lipman. Sal: Sign agnostic learning of shapes from raw data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2565-2574, 2020. 2 +[2] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 1 +[3] Fausto Bernardini, Joshua Mittleman, Holly Rushmeier, Cláudio Silva, and Gabriel Taubin. The ball-pivoting algorithm for surface reconstruction. IEEE transactions on visualization and computer graphics, 5(4):349-359, 1999. 2, 6 +[4] Neill DF Campbell, George Vogiatzis, Carlos Hernández, and Roberto Cipolla. Using multiple hypotheses to improve depth-maps for multi-view stereo. In European Conference on Computer Vision, pages 766-779. Springer, 2008. 2 +[5] Weikai Chen, Cheng Lin, Weiyang Li, and Bo Yang. 3psdf: Three-pole signed distance function for learning surfaces with arbitrary topologies. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18522-18531, 2022. 1, 2 +[6] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5939-5948, 2019. 2 +[7] Julian Chibane, Gerard Pons-Moll, et al. Neural unsigned distance fields for implicit function learning. Advances in Neural Information Processing Systems, 33:21638-21652, 2020. 2 +[8] François Darmon, Bénédicte Bascle, Jean-Clément Devaux, Pascal Monasse, and Mathieu Aubry. Improving neural implicit surfaces geometry with patch warping. arXiv preprint arXiv:2112.09648, 2021. 2, 7 +[9] François Darmon, Bénédicte Bascle, Jean-Clément Devaux, Pascal Monasse, and Mathieu Aubry. Improving neural implicit surfaces geometry with patch warping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6260–6269, 2022. 6, 7 +[10] Yasutaka Furukawa and Jean Ponce. Accurate, dense, and robust multiview stereopsis. IEEE transactions on pattern analysis and machine intelligence, 32(8):1362-1376, 2009. 2 +[11] Silvano Galliani, Katrin Lasinger, and Konrad Schindler. Massively parallel multiview stereopsis by surface normal diffusion. In Proceedings of the IEEE International Conference on Computer Vision, pages 873-881, 2015. 2 +[12] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. arXiv preprint arXiv:2002.10099, 2020. 2 +[13] Xiaodong Gu, Zhiwen Fan, Siyu Zhu, Zuozhuo Dai, Feitong Tan, and Ping Tan. Cascade cost volume for high-resolution multi-view stereo and stereo matching. In Proceedings of + +the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2495-2504, 2020. 2 +[14] Benoit Guillard, Federico Stella, and Pascal Fua. Meshud: Fast and differentiable meshing of unsigned distance field networks. arXiv preprint arXiv:2111.14549, 2021. 5 +[15] Rasmus Jensen, Anders Dahl, George Vogiatzis, Engin Tola, and Henrik Aanæs. Large scale multi-view stereopsis evaluation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 406-413, 2014. 2, 5, 7 +[16] Mengqi Ji, Juergen Gall, Haitian Zheng, Yebin Liu, and Lu Fang. Surfacenet: An end-to-end 3d neural network for multiview stereopsis. In Proceedings of the IEEE International Conference on Computer Vision, pages 2307-2315, 2017. 2 +[17] Mengqi Ji, Jinzhi Zhang, Qionghai Dai, and Lu Fang. SurfaceNet+: An end-to-end 3d neural network for very sparse multi-view stereopsis. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(11):4078-4093, 2020. 2 +[18] Yue Jiang, Dantong Ji, Zhizhong Han, and Matthias Zwicker. Sdfdiff: Differentiable rendering of signed distance fields for 3d shape optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1251-1261, 2020. 2 +[19] Abhishek Kar, Christian Hane, and Jitendra Malik. Learning a multi-view stereo machine. Advances in neural information processing systems, 30, 2017. 2 +[20] Michael Kazhdan, Matthew Bolitho, and Hugues Hoppe. Poisson surface reconstruction. In Proceedings of the fourth Eurographics symposium on Geometry processing, volume 7, 2006. 2 +[21] Michael Kazhdan and Hugues Hoppe. Screened poisson surface reconstruction. ACM Transactions on Graphics (ToG), 32(3):1-13, 2013. 6 +[22] Petr Kellnhofer, Lars C Jebe, Andrew Jones, Ryan Spicer, Kari Pulli, and Gordon Wetzstein. Neural lumigraph rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4287-4297, 2021. 2 +[23] Kiriakos N Kutulakos and Steven M Seitz. A theory of shape by space carving. International journal of computer vision, 38(3):199-218, 2000. 2 +[24] Der-Tsai Lee and Bruce J Schachter. Two algorithms for constructing a delaunay triangulation. International Journal of Computer & Information Sciences, 9(3):219-242, 1980. 2 +[25] Maxime Lhuillier and Long Quan. A quasi-dense approach to surface reconstruction from uncalibrated images. IEEE transactions on pattern analysis and machine intelligence, 27(3):418-433, 2005. 2 +[26] Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. Barf: Bundle-adjusting neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5741-5751, 2021. 1 +[27] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. Advances in Neural Information Processing Systems, 33:15651-15663, 2020. 1 +[28] Shaohui Liu, Yinda Zhang, Songyou Peng, Boxin Shi, Marc Pollefeys, and Zhaopeng Cui. Dist: Rendering deep implicit + +signed distance function with differentiable sphere tracing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2019-2028, 2020. 2 +[29] Xiaoxiao Long, Cheng Lin, Lingjie Liu, Wei Li, Christian Theobalt, Ruigang Yang, and Wenping Wang. Adaptive surface normal constraint for depth estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12849-12858, 2021. 2 +[30] Xiaoxiao Long, Cheng Lin, Peng Wang, Taku Komura, and Wenping Wang. Sparseneus: Fast generalizable neural surface reconstruction from sparse views. arXiv preprint arXiv:2206.05737, 2022. 5, 8 +[31] Xiaoxiao Long, Lingjie Liu, Wei Li, Christian Theobalt, and Wenping Wang. Multi-view depth estimation using epipolar spatio-temporal networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8258-8267, 2021. 2 +[32] Xiaoxiao Long, Lingjie Liu, Christian Theobalt, and Wenping Wang. Occlusion-aware depth estimation with adaptive normal constraints. In European Conference on Computer Vision, pages 640-657. Springer, 2020. 2 +[33] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4460-4470, 2019. 2 +[34] Mateusz Michalkiewicz, Jhony K Pontes, Dominic Jack, Mahsa Baktashmotlagh, and Anders Eriksson. Implicit surface representations as layers in neural networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4743-4752, 2019. 2 +[35] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European conference on computer vision, pages 405-421. Springer, 2020. 1, 2, 6, 7 +[36] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11453-11464, 2021. 1 +[37] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3504-3515, 2020. 2 +[38] Michael Oechsle, Songyou Peng, and Andreas Geiger. Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5589-5599, 2021. 2, 6, 7 +[39] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 165-174, 2019. 2 +[40] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo + +Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865-5874, 2021. 1 +[41] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 6, 7 +[42] Johannes L Schonberger, Enliang Zheng, Jan-Michael Frahm, and Marc Pollefeys. Pixelwise view selection for unstructured multi-view stereo. In European Conference on Computer Vision, pages 501-518. Springer, 2016. 2 +[43] Steven M Seitz and Charles R Dyer. Photorealistic scene reconstruction by voxel coloring. International Journal of Computer Vision, 35(2):151-173, 1999. 2 +[44] Jiaming Sun, Yiming Xie, Linghao Chen, Xiaowei Zhou, and Hujun Bao. Neuralrecon: Real-time coherent 3d reconstruction from monocular video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15598-15607, 2021. 1, 2 +[45] Engin Tola, Christoph Strecha, and Pascal Fua. Efficient large-scale multi-view stereo for ultra high-resolution image sets. Machine Vision and Applications, 23(5):903-920, 2012. 2 +[46] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. Advances in Neural Information Processing Systems, 34, 2021. 1, 2, 3, 5, 6, 7, 8 +[47] Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. Mvsnet: Depth inference for unstructured multi-view stereo. In Proceedings of the European Conference on Computer Vision (ECCV), pages 767-783, 2018. 2 +[48] Yao Yao, Zixin Luo, Shiwei Li, Tianwei Shen, Tian Fang, and Long Quan. Recurrent mvsnet for high-resolution multi-view stereo depth inference. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5525-5534, 2019. 2 +[49] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems, 34, 2021. 1, 2, 5, 6, 7 +[50] Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Basri Ronen, and Yaron Lipman. Multiview neural surface reconstruction by disentangling geometry and appearance. Advances in Neural Information Processing Systems, 33:2492-2502, 2020. 2, 5, 6, 7 +[51] Jingyang Zhang, Yao Yao, and Long Quan. Learning signed distance field for multi-view surface reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6525-6534, 2021. 2 +[52] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492, 2020. 1 +[53] Heming Zhu, Yu Cao, Hang Jin, Weikai Chen, Dong Du, Zhangye Wang, Shuguang Cui, and Xiaoguang Han. Deep fashion3d: A dataset and benchmark for 3d garment reconstruction from single images. In European Conference on Computer Vision, pages 512-530. Springer, 2020. 2, 5, 6 \ No newline at end of file diff --git a/2023/NeuralUDF_ Learning Unsigned Distance Fields for Multi-View Reconstruction of Surfaces With Arbitrary Topologies/images.zip b/2023/NeuralUDF_ Learning Unsigned Distance Fields for Multi-View Reconstruction of Surfaces With Arbitrary Topologies/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..5449515e6127f2edabafb76fedad9ecac7b5738c --- /dev/null +++ b/2023/NeuralUDF_ Learning Unsigned Distance Fields for Multi-View Reconstruction of Surfaces With Arbitrary Topologies/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5a5be6af32a2e5197f2ed8ca6782fff9183da529045761b7b3056218b6aca042 +size 540344 diff --git a/2023/NeuralUDF_ Learning Unsigned Distance Fields for Multi-View Reconstruction of Surfaces With Arbitrary Topologies/layout.json b/2023/NeuralUDF_ Learning Unsigned Distance Fields for Multi-View Reconstruction of Surfaces With Arbitrary Topologies/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..69220d88c50c08b3ed9a3030eef19259883d70a1 --- /dev/null +++ b/2023/NeuralUDF_ Learning Unsigned Distance Fields for Multi-View Reconstruction of Surfaces With Arbitrary Topologies/layout.json @@ -0,0 +1,12105 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 84, + 103, + 507, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 103, + 507, + 140 + ], + "spans": [ + { + "bbox": [ + 84, + 103, + 507, + 140 + ], + "type": "text", + "content": "NeuralUDF: Learning Unsigned Distance Fields for Multi-view Reconstruction of Surfaces with Arbitrary Topologies" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 107, + 159, + 485, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 159, + 485, + 223 + ], + "spans": [ + { + "bbox": [ + 107, + 159, + 485, + 223 + ], + "type": "text", + "content": "Xiaoxiao Long" + }, + { + "bbox": [ + 107, + 159, + 485, + 223 + ], + "type": "inline_equation", + "content": "^{1,2\\dagger}" + }, + { + "bbox": [ + 107, + 159, + 485, + 223 + ], + "type": "text", + "content": " Cheng Lin" + }, + { + "bbox": [ + 107, + 159, + 485, + 223 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 107, + 159, + 485, + 223 + ], + "type": "text", + "content": " Lingjie Liu" + }, + { + "bbox": [ + 107, + 159, + 485, + 223 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 107, + 159, + 485, + 223 + ], + "type": "text", + "content": " Yuan Liu" + }, + { + "bbox": [ + 107, + 159, + 485, + 223 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 107, + 159, + 485, + 223 + ], + "type": "text", + "content": " Peng Wang" + }, + { + "bbox": [ + 107, + 159, + 485, + 223 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 107, + 159, + 485, + 223 + ], + "type": "text", + "content": " Christian Theobalt" + }, + { + "bbox": [ + 107, + 159, + 485, + 223 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 107, + 159, + 485, + 223 + ], + "type": "text", + "content": " Taku Komura" + }, + { + "bbox": [ + 107, + 159, + 485, + 223 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 107, + 159, + 485, + 223 + ], + "type": "text", + "content": " Wenping Wang" + }, + { + "bbox": [ + 107, + 159, + 485, + 223 + ], + "type": "inline_equation", + "content": "^{4*}" + }, + { + "bbox": [ + 107, + 159, + 485, + 223 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 107, + 159, + 485, + 223 + ], + "type": "text", + "content": "The University of Hong Kong " + }, + { + "bbox": [ + 107, + 159, + 485, + 223 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 107, + 159, + 485, + 223 + ], + "type": "text", + "content": "Tencent Games \n" + }, + { + "bbox": [ + 107, + 159, + 485, + 223 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 107, + 159, + 485, + 223 + ], + "type": "text", + "content": "Max Planck Institute for Informatics " + }, + { + "bbox": [ + 107, + 159, + 485, + 223 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 107, + 159, + 485, + 223 + ], + "type": "text", + "content": "Texas A&M University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 248, + 192, + 261 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 248, + 192, + 261 + ], + "spans": [ + { + "bbox": [ + 143, + 248, + 192, + 261 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 273, + 290, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 273, + 290, + 514 + ], + "spans": [ + { + "bbox": [ + 46, + 273, + 290, + 514 + ], + "type": "text", + "content": "We present a novel method, called NeuralUDF, for reconstructing surfaces with arbitrary topologies from 2D images via volume rendering. Recent advances in neural rendering based reconstruction have achieved compelling results. However, these methods are limited to objects with closed surfaces since they adopt Signed Distance Function (SDF) as surface representation which requires the target shape to be divided into inside and outside. In this paper, we propose to represent surfaces as the Unsigned Distance Function (UDF) and develop a new volume rendering scheme to learn the neural UDF representation. Specifically, a new density function that correlates the property of UDF with the volume rendering scheme is introduced for robust optimization of the UDF fields. Experiments on the DTU and DeepFashion3D datasets show that our method not only enables high-quality reconstruction of non-closed shapes with complex typologies, but also achieves comparable performance to the SDF based methods on the reconstruction of closed surfaces. Visit our project page at https://www.xxlong.site/NeuralUDF/." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 521, + 128, + 534 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 521, + 128, + 534 + ], + "spans": [ + { + "bbox": [ + 47, + 521, + 128, + 534 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 541, + 287, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 541, + 287, + 685 + ], + "spans": [ + { + "bbox": [ + 46, + 541, + 287, + 685 + ], + "type": "text", + "content": "Reconstructing high-quality surfaces from multi-view images is a long-standing problem in computer vision and computer graphics. Neural implicit fields have become an emerging trend in recent advances due to its superior capability of representing surfaces of complex geometry. NeRF [35] and its variants [2, 26, 27, 36, 40, 52] have recently achieved compelling results in novel view synthesis. For each point in the 3D space, NeRF-based methods learn two neural implicit functions based on volume rendering: a volume density function and a view-dependent color function. Despite their successes in novel view synthesis, NeRF-based methods still struggle to faithfully reconstruct the ac" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 313, + 248, + 537, + 453 + ], + "blocks": [ + { + "bbox": [ + 313, + 248, + 537, + 453 + ], + "lines": [ + { + "bbox": [ + 313, + 248, + 537, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 248, + 537, + 453 + ], + "type": "image", + "image_path": "708d98f8ae5d6eca6b29dd87cb20d25df7f26c6f32520a4926a6f6a1c7b52a11.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 454, + 545, + 521 + ], + "lines": [ + { + "bbox": [ + 304, + 454, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 454, + 545, + 521 + ], + "type": "text", + "content": "Figure 1. We show three groups of multi-view reconstruction results generated by our proposed NeuralUDF and NeuS [46] respectively. Our method is able to faithfully reconstruct the high-quality geometries for both the closed and open surfaces, while NeuS can only model shapes as closed surfaces, thus leading to inconsistent typologies and erroneous geometries." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 532, + 547, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 532, + 547, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 532, + 547, + 567 + ], + "type": "text", + "content": "curate scene geometry from multi-view inputs, because of the difficulty in extracting high-quality surfaces from the representation of volume density." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 568, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 568, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 568, + 545, + 689 + ], + "type": "text", + "content": "VolSDF [49] and NeuS [46] incorporate the Signed Distance Function (SDF) into volume rendering to facilitate high-quality surface reconstruction in a NeRF framework. However, as a continuous function with clearly defined inside/outside, SDF is limited to modeling only closed watertight surfaces. Although there have been efforts to modify the SDF representation by learning an additional truncation function (e.g., 3PSDF [5], TSDF [44]), their surface representations are still built upon the definition of SDF, thus they are not suitable for representing complex topologies." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "We therefore propose to employ the Unsigned Distance Function (UDF) to represent surfaces in volume rendering." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 693, + 143, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 693, + 143, + 703 + ], + "spans": [ + { + "bbox": [ + 61, + 693, + 143, + 703 + ], + "type": "text", + "content": "*Corresponding authors." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 61, + 703, + 277, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 703, + 277, + 712 + ], + "spans": [ + { + "bbox": [ + 61, + 703, + 277, + 712 + ], + "type": "text", + "content": "This work was conducted during an internship at Tencent Games." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20834" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "With a surface represented by its zero level set without signs, UDF is a unified representation with higher-degree of freedom for both closed and open surfaces, thus making it possible to reconstruct shapes with arbitrary topologies." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 121, + 288, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 121, + 288, + 262 + ], + "spans": [ + { + "bbox": [ + 46, + 121, + 288, + 262 + ], + "type": "text", + "content": "There are two major challenges in learning a neural UDF field by volume rendering. First, UDF is not occlusion-aware, while the formulation of volume rendering with distance fields requires to estimate surface occlusion for points. Considering a camera ray intersecting with a surface, SDF assumes that the surface is closed and distinguishes the inside/outside with signs, so that a negative SDF value on the ray clearly indicates an occluded point inside the surface. In contrast, UDF does not impose the closed surface assumption and always gives non-negative values along the ray. Hence, the UDF value of a point alone cannot be used to infer occlusion." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 263, + 287, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 263, + 287, + 357 + ], + "spans": [ + { + "bbox": [ + 46, + 263, + 287, + 357 + ], + "type": "text", + "content": "The second challenge is that the UDF is not differentiable at its zero-level sets. The non-differentiability at the zero-level sets imposes barriers in the learning of UDF field. The gradients are ill-defined near the iso-surface, leading to difficulty in optimization. Consequently, the distance field surrounding the iso-surface is not accurate and the exact zero-level set of a learned UDF cannot be identified in a stable manner." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 360, + 288, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 360, + 288, + 502 + ], + "spans": [ + { + "bbox": [ + 46, + 360, + 288, + 502 + ], + "type": "text", + "content": "In this paper, we present a novel method for learning neural UDF fields by volume rendering. We introduce a density function that correlates the property of the UDF representation with the volume rendering process, which effectively tackles the aforementioned challenges induced by the unsigned representation and enables robust learning of surfaces with arbitrary topologies. Experiments on DTU [15] and DeepFashion3D [53] datasets show that our method not only enables the high-quality reconstruction of non-closed shapes with complex topologies, but also achieves comparable performance to the SDF based methods on the reconstruction of closed surfaces." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 503, + 226, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 503, + 226, + 514 + ], + "spans": [ + { + "bbox": [ + 59, + 503, + 226, + 514 + ], + "type": "text", + "content": "Our contributions can be summarized as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 514, + 286, + 635 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 58, + 514, + 286, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 514, + 286, + 548 + ], + "spans": [ + { + "bbox": [ + 58, + 514, + 286, + 548 + ], + "type": "text", + "content": "- We incorporate the UDF into volume rendering, which extends the representation of the underlying geometry of neural radiance fields." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 58, + 552, + 286, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 552, + 286, + 597 + ], + "spans": [ + { + "bbox": [ + 58, + 552, + 286, + 597 + ], + "type": "text", + "content": "- We introduce an effective density function that correlates the property of the UDF with the volume rendering process, thus enabling robust optimization of the distance fields." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 58, + 600, + 286, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 600, + 286, + 635 + ], + "spans": [ + { + "bbox": [ + 58, + 600, + 286, + 635 + ], + "type": "text", + "content": "- Our method achieves SOTA results for reconstructing high-quality surfaces with various topologies (closed or open) using the UDF in volume rendering." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 645, + 134, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 645, + 134, + 658 + ], + "spans": [ + { + "bbox": [ + 47, + 645, + 134, + 658 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 286, + 713 + ], + "type": "text", + "content": "Classic multi-view 3D reconstruction. 3D reconstruction from multiple images is a fundamental problem in computer vision. Traditional reconstruction methods usually adopt discrete representations, such as voxel grids [16," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 72, + 545, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 227 + ], + "type": "text", + "content": "17, 19, 23, 43, 44], 3D point clouds [10, 25], and depth maps [4, 11, 13, 29, 31, 32, 42, 45, 47, 48]. The depth map representation is flexible and appropriate for parallel computation, so it is the most widely used representation. The depth map based methods first compute depth maps by analyzing the correlation between multiple input images, and then fuse the depth maps into a point cloud, which is then further processed into a mesh via meshing algorithms, such as Poisson surface reconstruction [20], ball-pivoting [3] and Delaunay triangulation [24]. However, the traditional methods are sensitive to reflections, image noises, and varying illuminations, thus usually yielding incomplete and noisy reconstructions." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 228, + 545, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 228, + 545, + 406 + ], + "spans": [ + { + "bbox": [ + 304, + 228, + 545, + 406 + ], + "type": "text", + "content": "Neural rendering based surface reconstruction. The recent advances on neural implicit representations have achieved promising results in multi-view 3D reconstruction [8,18,22,28,37,38,46,49-51]. These methods typically represent the surfaces to be reconstructed as Signed Distance Function (SDF), and utilize surface rendering or volume rendering to render the pixel colors of the input images for optimizing the SDF fields. The surface rendering based methods [37, 50] require extra masks for optimization and cannot handle objects with complex structures. The volume rendering based works [8, 38, 46, 49, 51] get rid of the extra masks and achieve more accurate reconstructions. However, these methods adopt SDF as representation so they are limited to closed surfaces and their performance degrade a lot for non-closed surfaces." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 407, + 545, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 407, + 545, + 598 + ], + "spans": [ + { + "bbox": [ + 304, + 407, + 545, + 598 + ], + "type": "text", + "content": "Surface representations. There are various representations to model 3D geometric surfaces. NeRF [35] uses a density field to represent the underlying geometry of a target scene. Although NeRF achieves huge successes in novel view synthesis, it's difficult to extract high-quality geometry from such a density representation. Signed Distance Function (SDF) is widely used for geometry modeling and surface reconstruction [1,6,8,12,33,34,38,39,46,49,51], but SDF is only applicable to objects with closed surfaces. To deal with open surfaces, there are some efforts to learn an additional truncation function (e.g. 3PSDF [5], TSDF [44]). However, their representations are just built on the modification of SDF, and thus are still subject to flexibility of SDF and struggle to model complex surfaces. NDF [7] uses a neural UDF field to reconstruct surfaces, but they explicitly rely on 3D point clouds for supervising the optimization." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 609, + 361, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 609, + 361, + 621 + ], + "spans": [ + { + "bbox": [ + 306, + 609, + 361, + 621 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": "Our goal is to reconstruct a surface " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": " from a set of posed input images " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{I}_k\\}" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": " of a 3D model. We represent the surface as the zero level set of an Unsigned Distance Function (UDF), and propose a novel volume rendering scheme to learn the UDF. We will introduce our method in three parts: 1) We discuss the relationship between the unsigned distance function and the density function of the volume ren" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20835" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 70, + 165, + 174 + ], + "blocks": [ + { + "bbox": [ + 53, + 70, + 165, + 174 + ], + "lines": [ + { + "bbox": [ + 53, + 70, + 165, + 174 + ], + "spans": [ + { + "bbox": [ + 53, + 70, + 165, + 174 + ], + "type": "image", + "image_path": "0c34ec97745c9acfd781109e16531c66efe7b61efd60c57536fdb2d5a6d23a70.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 174, + 153, + 183 + ], + "lines": [ + { + "bbox": [ + 54, + 174, + 153, + 183 + ], + "spans": [ + { + "bbox": [ + 54, + 174, + 153, + 183 + ], + "type": "text", + "content": "(a) Biased density function" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 185, + 287, + 262 + ], + "lines": [ + { + "bbox": [ + 46, + 185, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 46, + 185, + 287, + 262 + ], + "type": "text", + "content": "Figure 2. Illustration of (a) biased density function and (b) unbiased density function for UDF. Suppose that the ray has one single intersection point " + }, + { + "bbox": [ + 46, + 185, + 287, + 262 + ], + "type": "inline_equation", + "content": "t^*" + }, + { + "bbox": [ + 46, + 185, + 287, + 262 + ], + "type": "text", + "content": " with the surface. (a) If the density function is modeled as a bell-shaped probability distribution, the weight function attains a local maximum before " + }, + { + "bbox": [ + 46, + 185, + 287, + 262 + ], + "type": "inline_equation", + "content": "t^*" + }, + { + "bbox": [ + 46, + 185, + 287, + 262 + ], + "type": "text", + "content": ". (b) If the density function is modeled as a monotonically increasing function, the weight function can attain its local maximum at " + }, + { + "bbox": [ + 46, + 185, + 287, + 262 + ], + "type": "inline_equation", + "content": "t^*" + }, + { + "bbox": [ + 46, + 185, + 287, + 262 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 168, + 82, + 281, + 174 + ], + "blocks": [ + { + "bbox": [ + 168, + 82, + 281, + 174 + ], + "lines": [ + { + "bbox": [ + 168, + 82, + 281, + 174 + ], + "spans": [ + { + "bbox": [ + 168, + 82, + 281, + 174 + ], + "type": "image", + "image_path": "bf32a3758bdcbf8280746fc26f7131225aca1456213c3dee09f76453afc9bf86.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 169, + 174, + 277, + 183 + ], + "lines": [ + { + "bbox": [ + 169, + 174, + 277, + 183 + ], + "spans": [ + { + "bbox": [ + 169, + 174, + 277, + 183 + ], + "type": "text", + "content": "(b) Unbiased density function" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 271, + 289, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 271, + 289, + 380 + ], + "spans": [ + { + "bbox": [ + 46, + 271, + 289, + 380 + ], + "type": "text", + "content": "dering equation; 2) To derive a density function induced by the UDF, we propose a differentiable indicator function to transform the UDF into its associated density field; 3) In addition to the loss terms commonly used in the prior works, we introduce an iso-surface regularizer to improve the stability of the UDF field near the zero level set. This UDF-based volume rendering scheme enables robust optimization of NeuralUDF for reconstructing both closed surfaces and open surfaces." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 386, + 157, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 386, + 157, + 399 + ], + "spans": [ + { + "bbox": [ + 47, + 386, + 157, + 399 + ], + "type": "text", + "content": "3.1. Volume Rendering" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 405, + 287, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 405, + 287, + 489 + ], + "spans": [ + { + "bbox": [ + 46, + 405, + 287, + 489 + ], + "type": "text", + "content": "Scene representation. Our method NeuralUDF represents the scene to be reconstructed by two neural implicit functions: 1) an unsigned distance function " + }, + { + "bbox": [ + 46, + 405, + 287, + 489 + ], + "type": "inline_equation", + "content": "f_{u}(\\mathbf{x})" + }, + { + "bbox": [ + 46, + 405, + 287, + 489 + ], + "type": "text", + "content": " that maps a 3D point " + }, + { + "bbox": [ + 46, + 405, + 287, + 489 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 46, + 405, + 287, + 489 + ], + "type": "text", + "content": " to its unsigned distance to the closest object surface; 2) a view-dependent color function " + }, + { + "bbox": [ + 46, + 405, + 287, + 489 + ], + "type": "inline_equation", + "content": "c(\\mathbf{x},\\mathbf{v})" + }, + { + "bbox": [ + 46, + 405, + 287, + 489 + ], + "type": "text", + "content": " that encodes the color associated with a point " + }, + { + "bbox": [ + 46, + 405, + 287, + 489 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 46, + 405, + 287, + 489 + ], + "type": "text", + "content": " and a viewing direction " + }, + { + "bbox": [ + 46, + 405, + 287, + 489 + ], + "type": "inline_equation", + "content": "\\mathbf{v} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 46, + 405, + 287, + 489 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 489, + 288, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 489, + 288, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 489, + 288, + 617 + ], + "type": "text", + "content": "Volume rendering. We consider a ray " + }, + { + "bbox": [ + 46, + 489, + 288, + 617 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 46, + 489, + 288, + 617 + ], + "type": "text", + "content": " emanating from the camera center " + }, + { + "bbox": [ + 46, + 489, + 288, + 617 + ], + "type": "inline_equation", + "content": "\\mathbf{o} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 46, + 489, + 288, + 617 + ], + "type": "text", + "content": " in direction " + }, + { + "bbox": [ + 46, + 489, + 288, + 617 + ], + "type": "inline_equation", + "content": "\\mathbf{v} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 46, + 489, + 288, + 617 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 46, + 489, + 288, + 617 + ], + "type": "inline_equation", + "content": "\\| \\mathbf{v} \\| = 1" + }, + { + "bbox": [ + 46, + 489, + 288, + 617 + ], + "type": "text", + "content": "), which is parameterized by " + }, + { + "bbox": [ + 46, + 489, + 288, + 617 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{r}(t) = \\mathbf{o} + t\\mathbf{v} \\mid t \\geq 0\\}" + }, + { + "bbox": [ + 46, + 489, + 288, + 617 + ], + "type": "text", + "content": ". Volume rendering is used to calculate the color " + }, + { + "bbox": [ + 46, + 489, + 288, + 617 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 489, + 288, + 617 + ], + "type": "text", + "content": " of the pixel corresponding to the ray " + }, + { + "bbox": [ + 46, + 489, + 288, + 617 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 46, + 489, + 288, + 617 + ], + "type": "text", + "content": ". The accumulated color of the camera ray " + }, + { + "bbox": [ + 46, + 489, + 288, + 617 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 46, + 489, + 288, + 617 + ], + "type": "text", + "content": " is given by " + }, + { + "bbox": [ + 46, + 489, + 288, + 617 + ], + "type": "inline_equation", + "content": "C(\\mathbf{r}) = \\int_0^{+\\infty} T(t) \\sigma(t) \\mathbf{c}(\\mathbf{r}(t), \\mathbf{v}) \\, \\mathrm{d}t" + }, + { + "bbox": [ + 46, + 489, + 288, + 617 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 489, + 288, + 617 + ], + "type": "inline_equation", + "content": "\\sigma(t)" + }, + { + "bbox": [ + 46, + 489, + 288, + 617 + ], + "type": "text", + "content": " is the volume density, " + }, + { + "bbox": [ + 46, + 489, + 288, + 617 + ], + "type": "inline_equation", + "content": "T(t) = \\exp \\left(-\\int_0^t \\sigma(u) \\, \\mathrm{d}u\\right)" + }, + { + "bbox": [ + 46, + 489, + 288, + 617 + ], + "type": "text", + "content": " denotes the accumulated transmittance along the ray, and the product of " + }, + { + "bbox": [ + 46, + 489, + 288, + 617 + ], + "type": "inline_equation", + "content": "T(t) \\sigma(t)" + }, + { + "bbox": [ + 46, + 489, + 288, + 617 + ], + "type": "text", + "content": " is the weight function of a point, denoted as " + }, + { + "bbox": [ + 46, + 489, + 288, + 617 + ], + "type": "inline_equation", + "content": "w(t)" + }, + { + "bbox": [ + 46, + 489, + 288, + 617 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 46, + 489, + 288, + 617 + ], + "type": "inline_equation", + "content": "w(t) = T(t) \\sigma(t)" + }, + { + "bbox": [ + 46, + 489, + 288, + 617 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 623, + 201, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 623, + 201, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 623, + 201, + 635 + ], + "type": "text", + "content": "3.2. Modeling Density with UDF" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "content": "In the NeRF framework, the density " + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\sigma(t)" + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "content": " is predicted by an MLP, where the points near the surfaces having larger density values and the points far from the surfaces having smaller densities. In order to apply volume rendering to learning distance field " + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "content": ", we need to define a mapping " + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "content": " to transform the distance function " + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "inline_equation", + "content": "f(\\mathbf{r}(t))" + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "content": " into a density field" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 72, + 436, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 436, + 84 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 436, + 84 + ], + "type": "inline_equation", + "content": "\\sigma (t)" + }, + { + "bbox": [ + 305, + 72, + 436, + 84 + ], + "type": "text", + "content": " , that is " + }, + { + "bbox": [ + 305, + 72, + 436, + 84 + ], + "type": "inline_equation", + "content": "\\sigma (t) = \\Omega (f(\\mathbf{r}(t)))" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 84, + 545, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 84, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 304, + 84, + 545, + 227 + ], + "type": "text", + "content": "Unbiased volume density function. Inspired by NeRF, we may consider defining the density function " + }, + { + "bbox": [ + 304, + 84, + 545, + 227 + ], + "type": "inline_equation", + "content": "\\sigma(t) = \\Omega(f(\\mathbf{r}(t)))" + }, + { + "bbox": [ + 304, + 84, + 545, + 227 + ], + "type": "text", + "content": " as a bell-shaped distribution model centered at zero, so that the surface points intersecting with a ray (UDF values are zero) have the maximum density values. However, as illustrated in Fig 2, if the mapping function is straightforwardly defined in this way, a noticeable bias will be induced in the weight function " + }, + { + "bbox": [ + 304, + 84, + 545, + 227 + ], + "type": "inline_equation", + "content": "w(t)" + }, + { + "bbox": [ + 304, + 84, + 545, + 227 + ], + "type": "text", + "content": " since its local maximum is not at the intersection point " + }, + { + "bbox": [ + 304, + 84, + 545, + 227 + ], + "type": "inline_equation", + "content": "t^*" + }, + { + "bbox": [ + 304, + 84, + 545, + 227 + ], + "type": "text", + "content": " where the ray intersects with the surfaces. This can result in conspicuous inaccuracy in the reconstructed surfaces, which is demonstrated in [46]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 228, + 546, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 228, + 546, + 323 + ], + "spans": [ + { + "bbox": [ + 304, + 228, + 546, + 323 + ], + "type": "text", + "content": "To define an appropriate mapping, we need to first review the relationship between the density function " + }, + { + "bbox": [ + 304, + 228, + 546, + 323 + ], + "type": "inline_equation", + "content": "\\sigma(t)" + }, + { + "bbox": [ + 304, + 228, + 546, + 323 + ], + "type": "text", + "content": " and a distance field " + }, + { + "bbox": [ + 304, + 228, + 546, + 323 + ], + "type": "inline_equation", + "content": "f(t)" + }, + { + "bbox": [ + 304, + 228, + 546, + 323 + ], + "type": "text", + "content": ". Suppose that there is only one single intersection of the ray and a surface. It is shown in NeuS [46] that, to make the weight function " + }, + { + "bbox": [ + 304, + 228, + 546, + 323 + ], + "type": "inline_equation", + "content": "w(t)" + }, + { + "bbox": [ + 304, + 228, + 546, + 323 + ], + "type": "text", + "content": " unbiased, the density function should be in the form of a monotonically increasing function, which attains the maximum at an infinite point rather than at the intersection point." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 323, + 545, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 323, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 305, + 323, + 545, + 396 + ], + "type": "text", + "content": "Accordingly, NeuS [46] chooses to use a Sigmoid function " + }, + { + "bbox": [ + 305, + 323, + 545, + 396 + ], + "type": "inline_equation", + "content": "\\Phi_{\\kappa}(x) = (1 + e^{-\\kappa x})^{-1}" + }, + { + "bbox": [ + 305, + 323, + 545, + 396 + ], + "type": "text", + "content": " to achieve the goal. The accumulated transmittance " + }, + { + "bbox": [ + 305, + 323, + 545, + 396 + ], + "type": "inline_equation", + "content": "T(t)" + }, + { + "bbox": [ + 305, + 323, + 545, + 396 + ], + "type": "text", + "content": " is then defined as " + }, + { + "bbox": [ + 305, + 323, + 545, + 396 + ], + "type": "inline_equation", + "content": "T(t) = \\Phi_{\\kappa}(f_s(t))" + }, + { + "bbox": [ + 305, + 323, + 545, + 396 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 305, + 323, + 545, + 396 + ], + "type": "inline_equation", + "content": "f_{s}(t)" + }, + { + "bbox": [ + 305, + 323, + 545, + 396 + ], + "type": "text", + "content": " is the signed distance function (SDF). The density function " + }, + { + "bbox": [ + 305, + 323, + 545, + 396 + ], + "type": "inline_equation", + "content": "\\sigma_{s}(t)" + }, + { + "bbox": [ + 305, + 323, + 545, + 396 + ], + "type": "text", + "content": " induced by SDF is formulated as" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 406, + 545, + 430 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 406, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 312, + 406, + 545, + 430 + ], + "type": "interline_equation", + "content": "\\sigma_ {s} (t) = \\Omega_ {s} \\left(f _ {s} (\\mathbf {r} (t))\\right) = \\kappa | \\cos (\\theta) | \\left(1 - \\Phi_ {\\kappa} \\left(f _ {s} (\\mathbf {r} (t))\\right)\\right), \\tag {1}", + "image_path": "e825ee437f6c9310d970af737c029c91d398818b469943f9c2912a38c80cc1fa.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 430, + 545, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 430, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 304, + 430, + 545, + 464 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 430, + 545, + 464 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 304, + 430, + 545, + 464 + ], + "type": "text", + "content": " is the parameter of Sigmoid function " + }, + { + "bbox": [ + 304, + 430, + 545, + 464 + ], + "type": "inline_equation", + "content": "\\Phi_{\\kappa}(x)" + }, + { + "bbox": [ + 304, + 430, + 545, + 464 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 430, + 545, + 464 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 430, + 545, + 464 + ], + "type": "text", + "content": " is the angle between the view direction " + }, + { + "bbox": [ + 304, + 430, + 545, + 464 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 304, + 430, + 545, + 464 + ], + "type": "text", + "content": " and the surface normal " + }, + { + "bbox": [ + 304, + 430, + 545, + 464 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 304, + 430, + 545, + 464 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 466, + 545, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 466, + 545, + 515 + ], + "spans": [ + { + "bbox": [ + 304, + 466, + 545, + 515 + ], + "type": "text", + "content": "Unbiased volume density for UDF. However, the unbiased density formulation induced by SDF [46] cannot be directly applied to UDF. The formulation of UDF near the surface can be given by" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 335, + 523, + 545, + 550 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 523, + 545, + 550 + ], + "spans": [ + { + "bbox": [ + 335, + 523, + 545, + 550 + ], + "type": "interline_equation", + "content": "f _ {u} (\\mathbf {r} (t)) = \\left\\{ \\begin{array}{l l} | \\cos (\\theta) \\cdot (t - t ^ {*}) |, & t \\neq t ^ {*} \\\\ 0, & t = t ^ {*} \\end{array} \\right. \\tag {2}", + "image_path": "a109fd86d127bdb666999e58aaa6ac4b76cc940b16e3e9230b6bab8a4895d718.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 557, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 545, + 653 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 557, + 545, + 653 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 557, + 545, + 653 + ], + "type": "text", + "content": " is the angle between the view direction " + }, + { + "bbox": [ + 304, + 557, + 545, + 653 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 304, + 557, + 545, + 653 + ], + "type": "text", + "content": " and the surface normal " + }, + { + "bbox": [ + 304, + 557, + 545, + 653 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 304, + 557, + 545, + 653 + ], + "type": "text", + "content": ". Obviously, the UDF value is not monotonous along a ray when it crosses a surface, where the mapping " + }, + { + "bbox": [ + 304, + 557, + 545, + 653 + ], + "type": "inline_equation", + "content": "t\\rightarrow f_u(\\mathbf{r}(t))" + }, + { + "bbox": [ + 304, + 557, + 545, + 653 + ], + "type": "text", + "content": " is non-injective. This leads to difficulties in modeling the monotonicity between " + }, + { + "bbox": [ + 304, + 557, + 545, + 653 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 557, + 545, + 653 + ], + "type": "text", + "content": " and the density function " + }, + { + "bbox": [ + 304, + 557, + 545, + 653 + ], + "type": "inline_equation", + "content": "\\sigma (t) = \\Omega (f_u(\\mathbf{r}(t)))" + }, + { + "bbox": [ + 304, + 557, + 545, + 653 + ], + "type": "text", + "content": ", given that an unbiased density function is expected to be monotonically increasing." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "To leverage the unbiased property derived from the monotonic density function, we need to model the injective mapping from " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": " to the density function " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\sigma_{u}(t)" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": " with the formulation of UDF " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "f_{u}(\\mathbf{r}(t))" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": ". Since the monotonicity of UDF along a ray only changes at the intersection point " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "t^{*}" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": ", we can" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20836" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "type": "text", + "content": "use a visibility indicator function " + }, + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "type": "inline_equation", + "content": "\\Psi(t)" + }, + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "type": "text", + "content": " to reflect the existence of the intersection and thus capture the monotonicity change:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 117, + 117, + 287, + 142 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 117, + 287, + 142 + ], + "spans": [ + { + "bbox": [ + 117, + 117, + 287, + 142 + ], + "type": "interline_equation", + "content": "\\Psi (t) = \\left\\{ \\begin{array}{l l} 1, & t < t ^ {*} \\\\ 0, & t > t ^ {*}. \\end{array} \\right. \\tag {3}", + "image_path": "0fa978868e2c93cba29fd42f781885048e3fb6dd8ce3c1174730406296618867.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 147, + 287, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 147, + 287, + 206 + ], + "spans": [ + { + "bbox": [ + 46, + 147, + 287, + 206 + ], + "type": "text", + "content": "The indicator function " + }, + { + "bbox": [ + 46, + 147, + 287, + 206 + ], + "type": "inline_equation", + "content": "\\Psi(t)" + }, + { + "bbox": [ + 46, + 147, + 287, + 206 + ], + "type": "text", + "content": " indicates the visibility of the points sampled along the ray " + }, + { + "bbox": [ + 46, + 147, + 287, + 206 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 46, + 147, + 287, + 206 + ], + "type": "text", + "content": ", where 1 means that the point is visible in the free space while 0 means that the point is occluded. Note this just gives a simple hint for the indicator function, while the actual formulation will be derived later." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 207, + 287, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 207, + 287, + 243 + ], + "spans": [ + { + "bbox": [ + 47, + 207, + 287, + 243 + ], + "type": "text", + "content": "With " + }, + { + "bbox": [ + 47, + 207, + 287, + 243 + ], + "type": "inline_equation", + "content": "\\Psi (t)" + }, + { + "bbox": [ + 47, + 207, + 287, + 243 + ], + "type": "text", + "content": ", given an unbiased density function " + }, + { + "bbox": [ + 47, + 207, + 287, + 243 + ], + "type": "inline_equation", + "content": "\\sigma (t) = \\Omega_s(f(\\mathbf{r}(t)))" + }, + { + "bbox": [ + 47, + 207, + 287, + 243 + ], + "type": "text", + "content": " that is monotonically increasing, the density function for UDF can be defined as" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 51, + 247, + 287, + 270 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 247, + 287, + 270 + ], + "spans": [ + { + "bbox": [ + 51, + 247, + 287, + 270 + ], + "type": "interline_equation", + "content": "\\sigma_ {u} (t) = \\Psi (t) \\Omega_ {s} \\left(f _ {u} (\\mathbf {r} (t))\\right) + (1 - \\Psi (t)) \\Omega_ {s} \\left(- f _ {u} (\\mathbf {r} (t))\\right), \\tag {4}", + "image_path": "a05b4d18e9b5f3eddd353cfbefb8279ff9d49096a41debf11aab051d042a7d48.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 283, + 287, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 283, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 46, + 283, + 287, + 342 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 283, + 287, + 342 + ], + "type": "inline_equation", + "content": "\\Omega_s" + }, + { + "bbox": [ + 46, + 283, + 287, + 342 + ], + "type": "text", + "content": " can be any monotonic density function that is unbiasedly designed, while in this paper we directly use the formulation of Eq.1. Since unsigned distance function " + }, + { + "bbox": [ + 46, + 283, + 287, + 342 + ], + "type": "inline_equation", + "content": "f_u" + }, + { + "bbox": [ + 46, + 283, + 287, + 342 + ], + "type": "text", + "content": " is not differentiable at " + }, + { + "bbox": [ + 46, + 283, + 287, + 342 + ], + "type": "inline_equation", + "content": "t^*" + }, + { + "bbox": [ + 46, + 283, + 287, + 342 + ], + "type": "text", + "content": ", the domain scope of " + }, + { + "bbox": [ + 46, + 283, + 287, + 342 + ], + "type": "inline_equation", + "content": "\\sigma_u(t)" + }, + { + "bbox": [ + 46, + 283, + 287, + 342 + ], + "type": "text", + "content": " is defined as " + }, + { + "bbox": [ + 46, + 283, + 287, + 342 + ], + "type": "inline_equation", + "content": "\\{t \\in \\mathbb{R}^1, t \\neq t^*\\}" + }, + { + "bbox": [ + 46, + 283, + 287, + 342 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 342, + 287, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 342, + 287, + 415 + ], + "spans": [ + { + "bbox": [ + 46, + 342, + 287, + 415 + ], + "type": "text", + "content": "Zero level set. The visibility indicator function " + }, + { + "bbox": [ + 46, + 342, + 287, + 415 + ], + "type": "inline_equation", + "content": "\\Psi" + }, + { + "bbox": [ + 46, + 342, + 287, + 415 + ], + "type": "text", + "content": " plays a vital role in the formulation of volume density for UDF, because it essentially encodes the distribution of the geometric surface. The key to the visibility is to find the intersection point " + }, + { + "bbox": [ + 46, + 342, + 287, + 415 + ], + "type": "inline_equation", + "content": "t^*" + }, + { + "bbox": [ + 46, + 342, + 287, + 415 + ], + "type": "text", + "content": " of surfaces and the ray, whose UDF value is zero, i.e., " + }, + { + "bbox": [ + 46, + 342, + 287, + 415 + ], + "type": "inline_equation", + "content": "f_{u}(\\mathbf{r}(t^{*})) = 0" + }, + { + "bbox": [ + 46, + 342, + 287, + 415 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 415, + 287, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 415, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 46, + 415, + 287, + 498 + ], + "type": "text", + "content": "It is nontrivial to find the zero level set for a UDF field. In practice, it's impossible to locate a point with exact zero UDF value, and a trade-off to balance accuracy and efficiency is to find points whose UDF values are smaller than a pre-defined threshold " + }, + { + "bbox": [ + 46, + 415, + 287, + 498 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 46, + 415, + 287, + 498 + ], + "type": "text", + "content": ". As a result, the approximated surfaces will be relaxed from exact zero level set to be a " + }, + { + "bbox": [ + 46, + 415, + 287, + 498 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 46, + 415, + 287, + 498 + ], + "type": "text", + "content": "-bounded surface band " + }, + { + "bbox": [ + 46, + 415, + 287, + 498 + ], + "type": "inline_equation", + "content": "S = \\{\\mathbf{x} \\in \\mathbb{R}^3 \\mid f_u(\\mathbf{x}) < \\epsilon\\}" + }, + { + "bbox": [ + 46, + 415, + 287, + 498 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "type": "text", + "content": "However, such a definition of the indicator function involving " + }, + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "type": "text", + "content": " bounding brings about the following difficulties in learning the UDF field. This indicator function design with " + }, + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "type": "text", + "content": " bounding operation incurs a set of barriers that impede learning the UDF field: 1) choosing an appropriate " + }, + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "type": "text", + "content": " is difficult; 2) a fixed " + }, + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 46, + 498, + 287, + 605 + ], + "type": "text", + "content": " cannot be adjusted to adapt the UDF field in different optimization stages; 3) the step-like indicator function is not differentiable. These limitations make the optimization of UDF unstable and easy to collapse." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": "Visibility indicator function. To overcome the limitations mentioned above, we leverage probabilistic models to make the indicator function " + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\Psi(t)" + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": " differentiable. Therefore, we propose to approximate the surfaces by a soft probabilistic distribution instead of using a hard " + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": " bound. We measure the probability of a point " + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{r}(t)" + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": " being on a surface, named as surface existence probability " + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "h(\\mathbf{r}(t))" + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": " abbreviated as " + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "h(t)" + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": ". This can be achieved by transforming the UDF values into probabilities via a bell-shaped distribution centered at 0. Here" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 72, + 544, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 544, + 96 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 544, + 96 + ], + "type": "text", + "content": "we use a logistic distribution " + }, + { + "bbox": [ + 305, + 72, + 544, + 96 + ], + "type": "inline_equation", + "content": "\\phi_{\\beta}(x)" + }, + { + "bbox": [ + 305, + 72, + 544, + 96 + ], + "type": "text", + "content": " to define the probability " + }, + { + "bbox": [ + 305, + 72, + 544, + 96 + ], + "type": "inline_equation", + "content": "h(t_i)" + }, + { + "bbox": [ + 305, + 72, + 544, + 96 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 337, + 106, + 545, + 118 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 106, + 545, + 118 + ], + "spans": [ + { + "bbox": [ + 337, + 106, + 545, + 118 + ], + "type": "interline_equation", + "content": "h \\left(t _ {i}\\right) = 1 - \\exp \\left(- \\alpha \\cdot \\phi_ {\\beta} \\left(f _ {u} \\left(\\mathbf {r} \\left(t _ {i}\\right)\\right)\\right) \\cdot \\delta_ {i}\\right), \\tag {5}", + "image_path": "2a6d0f221b119828c465eda036f852fd2f99e8e715e73c95a0d23fcb6bfc7467.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 362, + 136, + 545, + 152 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 136, + 545, + 152 + ], + "spans": [ + { + "bbox": [ + 362, + 136, + 545, + 152 + ], + "type": "interline_equation", + "content": "\\phi_ {\\beta} (x) = \\beta e ^ {- \\beta x} / \\left(1 + e ^ {- \\beta x}\\right) ^ {2} \\tag {6}", + "image_path": "8d058cf51ae87c0365b4803b12d5e073c410a0f8f6afbb05a8f8eba7d92e6d45.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 157, + 545, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 157, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 305, + 157, + 545, + 239 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 157, + 545, + 239 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 305, + 157, + 545, + 239 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 305, + 157, + 545, + 239 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 305, + 157, + 545, + 239 + ], + "type": "text", + "content": "-th sample in the ray " + }, + { + "bbox": [ + 305, + 157, + 545, + 239 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 305, + 157, + 545, + 239 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 157, + 545, + 239 + ], + "type": "inline_equation", + "content": "h(t_i)" + }, + { + "bbox": [ + 305, + 157, + 545, + 239 + ], + "type": "text", + "content": " denotes the surface existence probability of the sample " + }, + { + "bbox": [ + 305, + 157, + 545, + 239 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 305, + 157, + 545, + 239 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 157, + 545, + 239 + ], + "type": "inline_equation", + "content": "\\delta_i = t_{i+1} - t_i" + }, + { + "bbox": [ + 305, + 157, + 545, + 239 + ], + "type": "text", + "content": " is the distance between adjacent samples, and " + }, + { + "bbox": [ + 305, + 157, + 545, + 239 + ], + "type": "inline_equation", + "content": "\\phi_\\beta(t)" + }, + { + "bbox": [ + 305, + 157, + 545, + 239 + ], + "type": "text", + "content": " is a logistic distribution function. Note that the standard deviation of " + }, + { + "bbox": [ + 305, + 157, + 545, + 239 + ], + "type": "inline_equation", + "content": "\\phi_\\beta(x)" + }, + { + "bbox": [ + 305, + 157, + 545, + 239 + ], + "type": "text", + "content": " is given by " + }, + { + "bbox": [ + 305, + 157, + 545, + 239 + ], + "type": "inline_equation", + "content": "1 / \\beta" + }, + { + "bbox": [ + 305, + 157, + 545, + 239 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 305, + 157, + 545, + 239 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 305, + 157, + 545, + 239 + ], + "type": "text", + "content": " is a trainable parameter, and " + }, + { + "bbox": [ + 305, + 157, + 545, + 239 + ], + "type": "inline_equation", + "content": "1 / \\beta" + }, + { + "bbox": [ + 305, + 157, + 545, + 239 + ], + "type": "text", + "content": " approaches zero as the network training converges; " + }, + { + "bbox": [ + 305, + 157, + 545, + 239 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 305, + 157, + 545, + 239 + ], + "type": "text", + "content": " is a constant scalar to scale the distribution." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 240, + 545, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 240, + 545, + 312 + ], + "spans": [ + { + "bbox": [ + 305, + 240, + 545, + 312 + ], + "type": "text", + "content": "After obtaining the surface existence probabilities of the sampled points on the ray " + }, + { + "bbox": [ + 305, + 240, + 545, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 305, + 240, + 545, + 312 + ], + "type": "text", + "content": ", we can measure the probability of whether a point is occluded by any surfaces or not. Since local UDF values do not indicate occlusion, we accumulate the probabilities of the points " + }, + { + "bbox": [ + 305, + 240, + 545, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{r}(t_j)" + }, + { + "bbox": [ + 305, + 240, + 545, + 312 + ], + "type": "text", + "content": " ahead of a query point " + }, + { + "bbox": [ + 305, + 240, + 545, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{r}(t_i)" + }, + { + "bbox": [ + 305, + 240, + 545, + 312 + ], + "type": "text", + "content": " along the ray:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 373, + 321, + 545, + 354 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 321, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 373, + 321, + 545, + 354 + ], + "type": "interline_equation", + "content": "\\hat {\\Psi} \\left(t _ {i}\\right) = \\prod_ {j = 1} ^ {i - 1} \\left(1 - h \\left(t _ {j}\\right)\\right), \\tag {7}", + "image_path": "581c93a15cdb5237d98308aae4261a9c3507a1ae83fba56b4bf40d8fd8d82268.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 359, + 545, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 359, + 545, + 479 + ], + "spans": [ + { + "bbox": [ + 305, + 359, + 545, + 479 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 359, + 545, + 479 + ], + "type": "inline_equation", + "content": "h(t_i) \\in [0,1]" + }, + { + "bbox": [ + 305, + 359, + 545, + 479 + ], + "type": "text", + "content": " measures the probability of a point being on the surfaces. From this equation, we can see that " + }, + { + "bbox": [ + 305, + 359, + 545, + 479 + ], + "type": "inline_equation", + "content": "\\hat{\\Psi}(t_i)" + }, + { + "bbox": [ + 305, + 359, + 545, + 479 + ], + "type": "text", + "content": " will drop dramatically to zero once there is a point that has high surface existence probability on the ray. Compared to using a hard bound, the probability accumulation strategy makes the indicator function " + }, + { + "bbox": [ + 305, + 359, + 545, + 479 + ], + "type": "inline_equation", + "content": "\\hat{\\Psi}(t)" + }, + { + "bbox": [ + 305, + 359, + 545, + 479 + ], + "type": "text", + "content": " differentiable and learnable, which adapts to learning various surface geometries. The function " + }, + { + "bbox": [ + 305, + 359, + 545, + 479 + ], + "type": "inline_equation", + "content": "\\hat{\\Psi}(t)" + }, + { + "bbox": [ + 305, + 359, + 545, + 479 + ], + "type": "text", + "content": " will become sharper when " + }, + { + "bbox": [ + 305, + 359, + 545, + 479 + ], + "type": "inline_equation", + "content": "1/\\beta" + }, + { + "bbox": [ + 305, + 359, + 545, + 479 + ], + "type": "text", + "content": " approaches 0 as the network training converges, resulting in more accurate approximation of the zero level set." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 479, + 545, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 479, + 545, + 562 + ], + "spans": [ + { + "bbox": [ + 305, + 479, + 545, + 562 + ], + "type": "text", + "content": "Gradient-aware indicator. Although the formulation of " + }, + { + "bbox": [ + 305, + 479, + 545, + 562 + ], + "type": "inline_equation", + "content": "\\hat{\\Psi}(t)" + }, + { + "bbox": [ + 305, + 479, + 545, + 562 + ], + "type": "text", + "content": " in Eq. 7 is effective for identifying surfaces in learning the UDF field, it can induce bias in the indicator. As shown in Figure 3, the surface existence probability " + }, + { + "bbox": [ + 305, + 479, + 545, + 562 + ], + "type": "inline_equation", + "content": "h(t_i)" + }, + { + "bbox": [ + 305, + 479, + 545, + 562 + ], + "type": "text", + "content": " increases when the ray approaches a surface, and the accumulated probability " + }, + { + "bbox": [ + 305, + 479, + 545, + 562 + ], + "type": "inline_equation", + "content": "\\hat{\\Psi}(t_i)" + }, + { + "bbox": [ + 305, + 479, + 545, + 562 + ], + "type": "text", + "content": " can attain zero before the point " + }, + { + "bbox": [ + 305, + 479, + 545, + 562 + ], + "type": "inline_equation", + "content": "t^*" + }, + { + "bbox": [ + 305, + 479, + 545, + 562 + ], + "type": "text", + "content": " where the ray intersects with the surface." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 563, + 544, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 563, + 544, + 587 + ], + "spans": [ + { + "bbox": [ + 305, + 563, + 544, + 587 + ], + "type": "text", + "content": "To alleviate this problem, we modify the indicator function " + }, + { + "bbox": [ + 305, + 563, + 544, + 587 + ], + "type": "inline_equation", + "content": "\\hat{\\Psi}(t)" + }, + { + "bbox": [ + 305, + 563, + 544, + 587 + ], + "type": "text", + "content": " with a masking function " + }, + { + "bbox": [ + 305, + 563, + 544, + 587 + ], + "type": "inline_equation", + "content": "m(t)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 361, + 596, + 545, + 629 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 596, + 545, + 629 + ], + "spans": [ + { + "bbox": [ + 361, + 596, + 545, + 629 + ], + "type": "interline_equation", + "content": "\\hat {\\Psi} (t _ {i}) = \\prod_ {j = 1} ^ {i - 1} \\left(1 - h \\left(t _ {j}\\right) \\cdot m \\left(t _ {j}\\right)\\right) \\tag {8}", + "image_path": "e6d183fc5aedfbdc04ce00a4dc003951bb26e3c34cdd810974305c438733c1cd.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 358, + 632, + 545, + 658 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 632, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 358, + 632, + 545, + 658 + ], + "type": "interline_equation", + "content": "m \\left(t _ {j}\\right) = \\left\\{ \\begin{array}{l l} 0, & \\cos \\left(\\theta_ {j + 1}\\right) < 0 \\\\ 1, & \\cos \\left(\\theta_ {j + 1}\\right) \\geq 0 \\end{array} \\right. \\tag {9}", + "image_path": "a6b9ec425e858431208af5fa70fa1c3a80e761654c8d7634e162974876ad796b.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\theta_{j + 1}" + }, + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "type": "text", + "content": " denotes the angle between the ray direction and the gradient of " + }, + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "f_{u}(\\mathbf{r}(t_{j + 1}))" + }, + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "type": "text", + "content": ". This formulation exploits the geometric property of UDF surface, that is, the gradient direction of UDF flips across its zero level set. Therefore," + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20837" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 165, + 210 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 165, + 210 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 165, + 210 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 165, + 210 + ], + "type": "image", + "image_path": "b81c1fe9f1cc7ab184aab0d5c24b584cb184cd43a08b6f821b46167b3d02ee70.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 63, + 211, + 153, + 220 + ], + "lines": [ + { + "bbox": [ + 63, + 211, + 153, + 220 + ], + "spans": [ + { + "bbox": [ + 63, + 211, + 153, + 220 + ], + "type": "text", + "content": "(a) Naive indicator func." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 172, + 72, + 284, + 211 + ], + "blocks": [ + { + "bbox": [ + 172, + 72, + 284, + 211 + ], + "lines": [ + { + "bbox": [ + 172, + 72, + 284, + 211 + ], + "spans": [ + { + "bbox": [ + 172, + 72, + 284, + 211 + ], + "type": "image", + "image_path": "e23405a73e68d984869ff1184cb8ce2eebe0f738205dac767b37c112adc60de5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 173, + 212, + 284, + 221 + ], + "lines": [ + { + "bbox": [ + 173, + 212, + 284, + 221 + ], + "spans": [ + { + "bbox": [ + 173, + 212, + 284, + 221 + ], + "type": "text", + "content": "(b) Grad-aware indicator func." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 222, + 286, + 288 + ], + "lines": [ + { + "bbox": [ + 46, + 222, + 286, + 288 + ], + "spans": [ + { + "bbox": [ + 46, + 222, + 286, + 288 + ], + "type": "text", + "content": "Figure 3. Illustration of (a) naive visibility indicator function and (b) gradient-aware visibility indicator function. The naive indicator function will attain zero before the ray intersects with surface point " + }, + { + "bbox": [ + 46, + 222, + 286, + 288 + ], + "type": "inline_equation", + "content": "t^*" + }, + { + "bbox": [ + 46, + 222, + 286, + 288 + ], + "type": "text", + "content": ". Incorporating the UDF gradients information, the gradient-aware indicator function attains zero exactly at the intersection point " + }, + { + "bbox": [ + 46, + 222, + 286, + 288 + ], + "type": "inline_equation", + "content": "t^*" + }, + { + "bbox": [ + 46, + 222, + 286, + 288 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 300, + 286, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 300, + 286, + 370 + ], + "spans": [ + { + "bbox": [ + 46, + 300, + 286, + 370 + ], + "type": "text", + "content": "the points with negative " + }, + { + "bbox": [ + 46, + 300, + 286, + 370 + ], + "type": "inline_equation", + "content": "\\cos (\\theta_{j + 1})" + }, + { + "bbox": [ + 46, + 300, + 286, + 370 + ], + "type": "text", + "content": " are located ahead of the intersection point " + }, + { + "bbox": [ + 46, + 300, + 286, + 370 + ], + "type": "inline_equation", + "content": "t^{*}" + }, + { + "bbox": [ + 46, + 300, + 286, + 370 + ], + "type": "text", + "content": "; we mask out these points and only accumulate the surface existence probability once the ray reaches the surface. The final formulation jointly leverages the insights from both UDF value and gradient, leading to more reliable surface identification." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 381, + 178, + 393 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 381, + 178, + 393 + ], + "spans": [ + { + "bbox": [ + 47, + 381, + 178, + 393 + ], + "type": "text", + "content": "3.3. Optimizing UDF Fields" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 401, + 286, + 580 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 401, + 286, + 580 + ], + "spans": [ + { + "bbox": [ + 46, + 401, + 286, + 580 + ], + "type": "text", + "content": "Iso-surface regularizer. With the volume rendering scheme mentioned above, NeuralUDF learns a UDF field from the images by minimizing the differences between rendered images and ground truth images. However, unlike SDF representation that is defined for closed surface, the intrinsic property of the UDF field makes it more difficult to optimize. First, the UDF has a higher degree of freedom accommodating arbitrary topologies instead of only closed surfaces. Second, the UDF is not differentiable at the zero level set; the values and gradients of the UDF field close to the zero level set are therefore hard to learn. This usually results in a wide strip around the iso-surface where the UDF values are distributed irregularly. Therefore, we introduce a simple yet effective geometric regularizer term during training to improve the stability of the UDF field:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 87, + 590, + 286, + 617 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 590, + 286, + 617 + ], + "spans": [ + { + "bbox": [ + 87, + 590, + 286, + 617 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r e g}} = \\frac {1}{M} \\sum_ {k, i} \\exp (- \\tau \\cdot f _ {u} (\\mathbf {r} _ {k} (t _ {k, i}))), \\tag {10}", + "image_path": "abd5a7fbdfabe53d26e8295db9ba77fd07eca647902235da75b3767ffdd6f2c5.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "type": "text", + "content": " is a constant scalar to scale the UDF values, and " + }, + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 629, + 286, + 713 + ], + "type": "text", + "content": " is the number of sampled rays per optimization step. This formulation effectively precludes the UDF values of the non-surface points from being zero, which thus encourages the simplicity and clearness of the zero level set in the field. The experiments show that this iso-surface regularizer excels at improving the surface quality (see Sec. 4.3)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": "Loss functions. We train NeuralUDF by enforcing the consistency of the rendered colors and the ground truth colors of the input images without using 3D ground truth shapes. Specifically, we optimize the neural networks and two trainable probabilistic parameters " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": " of Eq. 1 and " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": " of Eq. 6 by randomly sampling a batch of pixels." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 144, + 545, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 144, + 545, + 167 + ], + "spans": [ + { + "bbox": [ + 305, + 144, + 545, + 167 + ], + "type": "text", + "content": "The overall loss function is defined as a weighted sum of several loss terms:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 310, + 177, + 545, + 190 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 177, + 545, + 190 + ], + "spans": [ + { + "bbox": [ + 310, + 177, + 545, + 190 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {c o l o r}} + \\lambda_ {0} \\mathcal {L} _ {\\text {p a t c h}} + \\lambda_ {1} \\mathcal {L} _ {\\text {e i k}} + \\lambda_ {2} \\mathcal {L} _ {\\text {r e g}} (+ \\gamma \\mathcal {L} _ {\\text {m a s k}}). \\tag {11}", + "image_path": "904fde290068536b1ee71e31baf87336d426898ecb81c3abe8ef988657ab799f.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 198, + 545, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 198, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 304, + 198, + 545, + 342 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 198, + 545, + 342 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{color}}" + }, + { + "bbox": [ + 304, + 198, + 545, + 342 + ], + "type": "text", + "content": " is the L1 color loss to minimize the differences between the rendered pixel colors and the ground truth pixel colors of the input images. Following the prior works [46, 49, 50], the Eikonal term " + }, + { + "bbox": [ + 304, + 198, + 545, + 342 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{eik}" + }, + { + "bbox": [ + 304, + 198, + 545, + 342 + ], + "type": "text", + "content": " is used to regularize the neural distance field to have unit norm of gradients, and the optional mask loss term " + }, + { + "bbox": [ + 304, + 198, + 545, + 342 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{mask}}" + }, + { + "bbox": [ + 304, + 198, + 545, + 342 + ], + "type": "text", + "content": " can be included if extra masks are provided. " + }, + { + "bbox": [ + 304, + 198, + 545, + 342 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{patch}}" + }, + { + "bbox": [ + 304, + 198, + 545, + 342 + ], + "type": "text", + "content": " is a patch-based color loss term introduced in [30], which renders the colors of a local patch of an input image using patch blending strategy and then enforces the consistency of the rendered colors and ground truth colors at the patch level. (See the supplementary materials for the details.)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 352, + 386, + 365 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 352, + 386, + 365 + ], + "spans": [ + { + "bbox": [ + 305, + 352, + 386, + 365 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 372, + 430, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 372, + 430, + 384 + ], + "spans": [ + { + "bbox": [ + 306, + 372, + 430, + 384 + ], + "type": "text", + "content": "4.1. Experimental settings" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 391, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 391, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 391, + 545, + 713 + ], + "type": "text", + "content": "Datasets. We conduct evaluations on the DTU [15] dataset and Deepfashion3D [53] dataset. DTU is a multi-view dataset containing objects with closed surfaces, and each scene consists of 49 or 64 images with " + }, + { + "bbox": [ + 304, + 391, + 545, + 713 + ], + "type": "inline_equation", + "content": "1600 \\times 1200" + }, + { + "bbox": [ + 304, + 391, + 545, + 713 + ], + "type": "text", + "content": " resolution. We use the same 15 scenes following the prior works for comparison. To validate the effectiveness of our method on objects with open surfaces, we adopt Deepfashion3D dataset for evaluation. Deepfashion3D is a large-scale real 3D garment dataset containing 563 diverse cloth items, and provides the ground truth point clouds of the captured garments. To facilitate the multi-view 3D reconstruction task, we randomly choose 12 garments from different categories of the dataset, and then scale the point clouds to be in the unit sphere. We render each point cloud into 72 images with " + }, + { + "bbox": [ + 304, + 391, + 545, + 713 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 304, + 391, + 545, + 713 + ], + "type": "text", + "content": " resolution. The rendered images are used for reconstruction. We further collect a set of synthetic plant objects for further qualitative comparisons. Each plant object is rendered into 324 images with " + }, + { + "bbox": [ + 304, + 391, + 545, + 713 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 304, + 391, + 545, + 713 + ], + "type": "text", + "content": " resolution. Implementation details. We use MLPs to model the unsigned distance field and color radiance field. We use the Adam optimizer with a global learning rate 5e-4 for the network training, specially we set the learning rate of the UDF MLP as 1e-4. We sample 512 rays per batch and train our model for 300k iterations on a single NVIDIA RTX 2080Ti GPU. Our network architecture and initialization scheme are similar to those of prior works [46, 49, 50]. We leverage MeshUDF [14], an extended marching cube algorithm" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20838" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 112, + 133 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 112, + 133 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 112, + 133 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 112, + 133 + ], + "type": "image", + "image_path": "e242aabf8e2646899a0f44bba7960b2ea0cb6001c9b04d79823652da3abae17f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 65, + 137, + 110, + 208 + ], + "blocks": [ + { + "bbox": [ + 65, + 137, + 110, + 208 + ], + "lines": [ + { + "bbox": [ + 65, + 137, + 110, + 208 + ], + "spans": [ + { + "bbox": [ + 65, + 137, + 110, + 208 + ], + "type": "image", + "image_path": "299a72f5e0baeb12419eea0fa76d9c0f057678222ca4b051b4c0f2db4f9baf17.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 72, + 211, + 104, + 289 + ], + "blocks": [ + { + "bbox": [ + 72, + 211, + 104, + 289 + ], + "lines": [ + { + "bbox": [ + 72, + 211, + 104, + 289 + ], + "spans": [ + { + "bbox": [ + 72, + 211, + 104, + 289 + ], + "type": "image", + "image_path": "2709eed36efc1360d8a22ba02e98a7fb12ffff7487d824f540ee63499dc6655f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 289, + 119, + 299 + ], + "lines": [ + { + "bbox": [ + 54, + 289, + 119, + 299 + ], + "spans": [ + { + "bbox": [ + 54, + 289, + 119, + 299 + ], + "type": "text", + "content": "Reference Images" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 300, + 547, + 346 + ], + "lines": [ + { + "bbox": [ + 46, + 300, + 547, + 346 + ], + "spans": [ + { + "bbox": [ + 46, + 300, + 547, + 346 + ], + "type": "text", + "content": "Figure 4. Qualitative comparisons with SOTAs on DeepFashion3D [53] dataset. The GT meshes are reconstructed from the GT point clouds provided by the DeepFashion3D dataset via Ball Pivoting meshing algorithm [3]. Thanks to the UDF representation, NeuralUDF accurately recovers the garments with open surfaces. However, the SDF-based methods can only model the garments as closed surfaces, thus leading to inconsistent topologies and erroneous geometries. (See supplementary materials for more results.)" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 128, + 71, + 180, + 133 + ], + "blocks": [ + { + "bbox": [ + 128, + 71, + 180, + 133 + ], + "lines": [ + { + "bbox": [ + 128, + 71, + 180, + 133 + ], + "spans": [ + { + "bbox": [ + 128, + 71, + 180, + 133 + ], + "type": "image", + "image_path": "c0547579cc7ed34a37e939b8df7be412cf4a7aecad236a606931a999e88d16d2.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 134, + 138, + 177, + 207 + ], + "blocks": [ + { + "bbox": [ + 134, + 138, + 177, + 207 + ], + "lines": [ + { + "bbox": [ + 134, + 138, + 177, + 207 + ], + "spans": [ + { + "bbox": [ + 134, + 138, + 177, + 207 + ], + "type": "image", + "image_path": "4968797747c75c8c49414bcebaa9e56c008868bc64b3870bddf15011a92481ed.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 126, + 211, + 188, + 288 + ], + "blocks": [ + { + "bbox": [ + 126, + 211, + 188, + 288 + ], + "lines": [ + { + "bbox": [ + 126, + 211, + 188, + 288 + ], + "spans": [ + { + "bbox": [ + 126, + 211, + 188, + 288 + ], + "type": "image", + "image_path": "6970deb48840a414a41653849168bf5d08d4588efb5e36deea701c48f1479d5f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 133, + 289, + 176, + 298 + ], + "lines": [ + { + "bbox": [ + 133, + 289, + 176, + 298 + ], + "spans": [ + { + "bbox": [ + 133, + 289, + 176, + 298 + ], + "type": "text", + "content": "GT meshes" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 211, + 72, + 262, + 133 + ], + "blocks": [ + { + "bbox": [ + 211, + 72, + 262, + 133 + ], + "lines": [ + { + "bbox": [ + 211, + 72, + 262, + 133 + ], + "spans": [ + { + "bbox": [ + 211, + 72, + 262, + 133 + ], + "type": "image", + "image_path": "9ffd3458043d718de8565dcb9313e15749704e03da1b0395f8d1056760122f4e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 217, + 138, + 260, + 208 + ], + "blocks": [ + { + "bbox": [ + 217, + 138, + 260, + 208 + ], + "lines": [ + { + "bbox": [ + 217, + 138, + 260, + 208 + ], + "spans": [ + { + "bbox": [ + 217, + 138, + 260, + 208 + ], + "type": "image", + "image_path": "6a9e7f918528cf3ebad80496ba3397f1636c824462e44d83d092f846e95ea2cd.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 209, + 211, + 272, + 287 + ], + "blocks": [ + { + "bbox": [ + 209, + 211, + 272, + 287 + ], + "lines": [ + { + "bbox": [ + 209, + 211, + 272, + 287 + ], + "spans": [ + { + "bbox": [ + 209, + 211, + 272, + 287 + ], + "type": "image", + "image_path": "d8af9485e5e787590b24da6d063c35bed6efece4fa50df85e9a840eec44822ad.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 228, + 289, + 247, + 297 + ], + "lines": [ + { + "bbox": [ + 228, + 289, + 247, + 297 + ], + "spans": [ + { + "bbox": [ + 228, + 289, + 247, + 297 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 293, + 72, + 345, + 133 + ], + "blocks": [ + { + "bbox": [ + 293, + 72, + 345, + 133 + ], + "lines": [ + { + "bbox": [ + 293, + 72, + 345, + 133 + ], + "spans": [ + { + "bbox": [ + 293, + 72, + 345, + 133 + ], + "type": "image", + "image_path": "2f2e904f26b537e612f76d8fcbcf36c0f66b84136c9d31a666cb3ce7414eed37.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 301, + 138, + 343, + 208 + ], + "blocks": [ + { + "bbox": [ + 301, + 138, + 343, + 208 + ], + "lines": [ + { + "bbox": [ + 301, + 138, + 343, + 208 + ], + "spans": [ + { + "bbox": [ + 301, + 138, + 343, + 208 + ], + "type": "image", + "image_path": "6030a40666a7272ba9e263344e497fa2f0c6f08cf205754cdd069bc86b4cdbef.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 293, + 211, + 354, + 288 + ], + "blocks": [ + { + "bbox": [ + 293, + 211, + 354, + 288 + ], + "lines": [ + { + "bbox": [ + 293, + 211, + 354, + 288 + ], + "spans": [ + { + "bbox": [ + 293, + 211, + 354, + 288 + ], + "type": "image", + "image_path": "19b7bdea0f27bda7885bd733afb14088d387e816dc831591e7655ea8512d90e1.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 302, + 289, + 341, + 298 + ], + "lines": [ + { + "bbox": [ + 302, + 289, + 341, + 298 + ], + "spans": [ + { + "bbox": [ + 302, + 289, + 341, + 298 + ], + "type": "text", + "content": "NeuS [46]" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 376, + 72, + 427, + 133 + ], + "blocks": [ + { + "bbox": [ + 376, + 72, + 427, + 133 + ], + "lines": [ + { + "bbox": [ + 376, + 72, + 427, + 133 + ], + "spans": [ + { + "bbox": [ + 376, + 72, + 427, + 133 + ], + "type": "image", + "image_path": "4084c6e7e2383ca5ee4bb9309fce142e7cd06c825cf986163f344b40502a6db7.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 383, + 138, + 426, + 208 + ], + "blocks": [ + { + "bbox": [ + 383, + 138, + 426, + 208 + ], + "lines": [ + { + "bbox": [ + 383, + 138, + 426, + 208 + ], + "spans": [ + { + "bbox": [ + 383, + 138, + 426, + 208 + ], + "type": "image", + "image_path": "d25fafb8dc21c7f4c88c572bfe127a652a16c998aad1a180f798443107dfde85.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 376, + 211, + 438, + 289 + ], + "blocks": [ + { + "bbox": [ + 376, + 211, + 438, + 289 + ], + "lines": [ + { + "bbox": [ + 376, + 211, + 438, + 289 + ], + "spans": [ + { + "bbox": [ + 376, + 211, + 438, + 289 + ], + "type": "image", + "image_path": "d8abcb8765f197281e7f3e3474629167d8036b90b07fd97cd16976eead352335.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 383, + 289, + 430, + 298 + ], + "lines": [ + { + "bbox": [ + 383, + 289, + 430, + 298 + ], + "spans": [ + { + "bbox": [ + 383, + 289, + 430, + 298 + ], + "type": "text", + "content": "VolSDF [49]" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 459, + 72, + 510, + 133 + ], + "blocks": [ + { + "bbox": [ + 459, + 72, + 510, + 133 + ], + "lines": [ + { + "bbox": [ + 459, + 72, + 510, + 133 + ], + "spans": [ + { + "bbox": [ + 459, + 72, + 510, + 133 + ], + "type": "image", + "image_path": "cc1c7c4339160dd12ddfc76d1cca44dc4f89f5b0191c7c9695783a697add3561.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 466, + 138, + 509, + 208 + ], + "blocks": [ + { + "bbox": [ + 466, + 138, + 509, + 208 + ], + "lines": [ + { + "bbox": [ + 466, + 138, + 509, + 208 + ], + "spans": [ + { + "bbox": [ + 466, + 138, + 509, + 208 + ], + "type": "image", + "image_path": "6dde80a8a022de71a1175eab430a119b525219ba27d715a0b85057c5c89230de.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 460, + 211, + 523, + 289 + ], + "blocks": [ + { + "bbox": [ + 460, + 211, + 523, + 289 + ], + "lines": [ + { + "bbox": [ + 460, + 211, + 523, + 289 + ], + "spans": [ + { + "bbox": [ + 460, + 211, + 523, + 289 + ], + "type": "image", + "image_path": "2ace1ce08f1096aa78f01f094bbb21813de616247a1b11e51358e5ab74024a34.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 472, + 289, + 512, + 298 + ], + "lines": [ + { + "bbox": [ + 472, + 289, + 512, + 298 + ], + "spans": [ + { + "bbox": [ + 472, + 289, + 512, + 298 + ], + "type": "text", + "content": "NeRF [35]" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "table", + "bbox": [ + 55, + 355, + 537, + 437 + ], + "blocks": [ + { + "bbox": [ + 55, + 355, + 537, + 437 + ], + "lines": [ + { + "bbox": [ + 55, + 355, + 537, + 437 + ], + "spans": [ + { + "bbox": [ + 55, + 355, + 537, + 437 + ], + "type": "table", + "html": "
ScanLS-C0SS-D0NS-D0NS-C0Skirt0LS-D0Trouser0NS-D1Trouser1LS-C1Skirt1SS-C0Mean
Colmap [41]2.952.913.583.063.233.263.093.112.953.162.972.953.10
IDR†[50]10.858.205.009.636.766.336.323.323.756.082.447.026.31
Nerf [35]6.387.845.035.454.697.864.824.636.047.178.246.896.25
VolSdf [49]5.924.794.548.577.025.966.974.367.248.737.748.846.72
Neus [46]3.184.824.784.993.735.715.892.215.893.602.445.134.36
NeuralWarp [9]4.7111.245.408.564.105.835.792.002.076.509.616.946.06
Ours1.922.052.361.581.334.112.471.501.632.472.162.151.97
", + "image_path": "f635dd88d93497d9e15b89df3c0951591b57798a1a0a402c3ed4a8a8e001c2ed.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 62, + 438, + 195, + 448 + ], + "lines": [ + { + "bbox": [ + 62, + 438, + 195, + 448 + ], + "spans": [ + { + "bbox": [ + 62, + 438, + 195, + 448 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 62, + 438, + 195, + 448 + ], + "type": "text", + "content": " Optimization using extra object masks." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "table_footnote" + }, + { + "bbox": [ + 67, + 448, + 536, + 467 + ], + "lines": [ + { + "bbox": [ + 67, + 448, + 536, + 467 + ], + "spans": [ + { + "bbox": [ + 67, + 448, + 536, + 467 + ], + "type": "text", + "content": "LS-C denotes Long-Sleeve Coat, SS-C denotes Short-Sleeve Coat, NS-C denotes None-Sleeve Coat, LS-D denotes Long-Sleeve Dress, SS-D denotes Short-Sleeve Dress, NS-D denotes None-Sleeve Dress." + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 25 + }, + { + "bbox": [ + 166, + 470, + 425, + 481 + ], + "lines": [ + { + "bbox": [ + 166, + 470, + 425, + 481 + ], + "spans": [ + { + "bbox": [ + 166, + 470, + 425, + 481 + ], + "type": "text", + "content": "Table 1. Quantitative evaluation results on Deepfashion3D [53] dataset." + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 495, + 287, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 495, + 287, + 518 + ], + "spans": [ + { + "bbox": [ + 46, + 495, + 287, + 518 + ], + "type": "text", + "content": "designed for UDF, to extract explicit mesh from the learned UDF field." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 46, + 521, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 287, + 628 + ], + "type": "text", + "content": "A hierarchical sampling strategy is used to sample points along a ray in a coarse-to-fine manner for volume rendering. We first uniformly sample points on the ray and then iteratively conduct importance sampling [35] to sample more points on top of coarse probability estimation. The probability of coarse sampling is computed based on the density function " + }, + { + "bbox": [ + 46, + 521, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\sigma_u(t)" + }, + { + "bbox": [ + 46, + 521, + 287, + 628 + ], + "type": "text", + "content": " with fixed and predefined parameters " + }, + { + "bbox": [ + 46, + 521, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 46, + 521, + 287, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 521, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 521, + 287, + 628 + ], + "type": "text", + "content": ", while the probability of fine sampling is computed with the learned " + }, + { + "bbox": [ + 46, + 521, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 46, + 521, + 287, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 521, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 521, + 287, + 628 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 287, + 713 + ], + "type": "text", + "content": "Baselines. We compare NeuralUDF with the following state-of-the-art baselines: 1) a widely used classic reconstruction method COLAMP [41], which we use to reconstruct a mesh from the point cloud generated by COLMAP via Screened Poisson Surface Reconstruction [21]; 2) neural implicit methods utilizing surface rendering, IDR [50] and UNISURF [38]; 3) neural implicit methods utilizing" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 304, + 495, + 545, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 495, + 545, + 603 + ], + "spans": [ + { + "bbox": [ + 304, + 495, + 545, + 603 + ], + "type": "text", + "content": "volume rendering, NeuS [46], VolSDF [49] and NeuralWarp [9]; 4) NeRF [35]. Although NeRF is not designed for surface reconstruction, we include NeRF for comparisons since its density representation enables successful novel view synthesis on objects with arbitrary topologies. Following NeuS [46], we use a density threshold of 25 on DTU (50 on Deepfashion3D) to extract surfaces from NeRF for comparisons. Note that, IDR takes extra masks as input for optimization while ours and the other methods do not." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 306, + 611, + 391, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 611, + 391, + 624 + ], + "spans": [ + { + "bbox": [ + 306, + 611, + 391, + 624 + ], + "type": "text", + "content": "4.2. Comparisons" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": "Evaluation on non-closed objects. We conduct evaluations on DeepFashion3D dataset whose objects are non-closed, and measure the Chamfer Distances " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "(10^{-3})" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": " between the predicted meshes with ground truth point clouds. As shown in Table 1, our method outperforms the other methods by a large margin. This is because our method can accurately reconstruct the surfaces with arbitrary topologies" + } + ] + } + ], + "index": 35 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20839" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 76, + 147, + 133 + ], + "blocks": [ + { + "bbox": [ + 70, + 76, + 147, + 133 + ], + "lines": [ + { + "bbox": [ + 70, + 76, + 147, + 133 + ], + "spans": [ + { + "bbox": [ + 70, + 76, + 147, + 133 + ], + "type": "image", + "image_path": "c679f9a0b7dca0b6723e5c55d9bfcfe9fbedac314565d9264bab1ac7fca914f8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 70, + 136, + 146, + 193 + ], + "blocks": [ + { + "bbox": [ + 70, + 136, + 146, + 193 + ], + "lines": [ + { + "bbox": [ + 70, + 136, + 146, + 193 + ], + "spans": [ + { + "bbox": [ + 70, + 136, + 146, + 193 + ], + "type": "image", + "image_path": "492d5cb01e8f4ec4a3f0c8b2c8ff094b166ad5d1be6c560966f1809d68806502.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 70, + 194, + 145, + 251 + ], + "blocks": [ + { + "bbox": [ + 70, + 194, + 145, + 251 + ], + "lines": [ + { + "bbox": [ + 70, + 194, + 145, + 251 + ], + "spans": [ + { + "bbox": [ + 70, + 194, + 145, + 251 + ], + "type": "image", + "image_path": "7df5b1a4d6d4c1aa652f7e3e254cc716ca478e12aaf4cc50c16a3a65578f7ab6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 76, + 251, + 142, + 261 + ], + "lines": [ + { + "bbox": [ + 76, + 251, + 142, + 261 + ], + "spans": [ + { + "bbox": [ + 76, + 251, + 142, + 261 + ], + "type": "text", + "content": "Reference Images" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 79, + 264, + 512, + 275 + ], + "lines": [ + { + "bbox": [ + 79, + 264, + 512, + 275 + ], + "spans": [ + { + "bbox": [ + 79, + 264, + 512, + 275 + ], + "type": "text", + "content": "Figure 5. Qualitative comparisons with SOTAs on DTU [15] dataset.(See the supplementary materials for more results.)" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 157, + 75, + 223, + 139 + ], + "blocks": [ + { + "bbox": [ + 157, + 75, + 223, + 139 + ], + "lines": [ + { + "bbox": [ + 157, + 75, + 223, + 139 + ], + "spans": [ + { + "bbox": [ + 157, + 75, + 223, + 139 + ], + "type": "image", + "image_path": "5cbc6191404c5e142cd74502f356e119366045fe351b214c091bc9ff3925ddf0.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 167, + 141, + 211, + 192 + ], + "blocks": [ + { + "bbox": [ + 167, + 141, + 211, + 192 + ], + "lines": [ + { + "bbox": [ + 167, + 141, + 211, + 192 + ], + "spans": [ + { + "bbox": [ + 167, + 141, + 211, + 192 + ], + "type": "image", + "image_path": "08bc81eaaa0f082acbe792256ee2e7a288d4a613732cb4d2798c1dbba8fe59d3.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 165, + 194, + 211, + 251 + ], + "blocks": [ + { + "bbox": [ + 165, + 194, + 211, + 251 + ], + "lines": [ + { + "bbox": [ + 165, + 194, + 211, + 251 + ], + "spans": [ + { + "bbox": [ + 165, + 194, + 211, + 251 + ], + "type": "image", + "image_path": "128346cefeb343db08d4663ce8e62cb5b21752df92ff3002967b074f86510af9.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 181, + 251, + 199, + 259 + ], + "lines": [ + { + "bbox": [ + 181, + 251, + 199, + 259 + ], + "spans": [ + { + "bbox": [ + 181, + 251, + 199, + 259 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 232, + 75, + 299, + 140 + ], + "blocks": [ + { + "bbox": [ + 232, + 75, + 299, + 140 + ], + "lines": [ + { + "bbox": [ + 232, + 75, + 299, + 140 + ], + "spans": [ + { + "bbox": [ + 232, + 75, + 299, + 140 + ], + "type": "image", + "image_path": "e25b2b7f2395985ae9c770d86947199343550c60abae68507a27b321daf1f41a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 241, + 141, + 285, + 192 + ], + "blocks": [ + { + "bbox": [ + 241, + 141, + 285, + 192 + ], + "lines": [ + { + "bbox": [ + 241, + 141, + 285, + 192 + ], + "spans": [ + { + "bbox": [ + 241, + 141, + 285, + 192 + ], + "type": "image", + "image_path": "408f7f76a4f2343a997affcb7b81f3332173aacb21acbaf4f058e749da798540.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 240, + 194, + 287, + 250 + ], + "blocks": [ + { + "bbox": [ + 240, + 194, + 287, + 250 + ], + "lines": [ + { + "bbox": [ + 240, + 194, + 287, + 250 + ], + "spans": [ + { + "bbox": [ + 240, + 194, + 287, + 250 + ], + "type": "image", + "image_path": "e839afe1720a683ec56a01d090738901224461d9dab5ec062738dcef4b160032.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 246, + 251, + 283, + 261 + ], + "lines": [ + { + "bbox": [ + 246, + 251, + 283, + 261 + ], + "spans": [ + { + "bbox": [ + 246, + 251, + 283, + 261 + ], + "type": "text", + "content": "NeuS [46]" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 307, + 73, + 373, + 137 + ], + "blocks": [ + { + "bbox": [ + 307, + 73, + 373, + 137 + ], + "lines": [ + { + "bbox": [ + 307, + 73, + 373, + 137 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 373, + 137 + ], + "type": "image", + "image_path": "3171d1bf9b8b9481cd1f8c64c05031f98880259aa0d09908127361f755bd00b2.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 317, + 140, + 360, + 192 + ], + "blocks": [ + { + "bbox": [ + 317, + 140, + 360, + 192 + ], + "lines": [ + { + "bbox": [ + 317, + 140, + 360, + 192 + ], + "spans": [ + { + "bbox": [ + 317, + 140, + 360, + 192 + ], + "type": "image", + "image_path": "f0b7375b52d41b0b43ab10fc53a8816413904bda949bfe8e54fd435818d30ce0.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 315, + 194, + 362, + 250 + ], + "blocks": [ + { + "bbox": [ + 315, + 194, + 362, + 250 + ], + "lines": [ + { + "bbox": [ + 315, + 194, + 362, + 250 + ], + "spans": [ + { + "bbox": [ + 315, + 194, + 362, + 250 + ], + "type": "image", + "image_path": "7c6148ef0de344de72a366ec523540a6a4bf1aeb846fb4c6d36b2d9091233f54.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 317, + 251, + 364, + 261 + ], + "lines": [ + { + "bbox": [ + 317, + 251, + 364, + 261 + ], + "spans": [ + { + "bbox": [ + 317, + 251, + 364, + 261 + ], + "type": "text", + "content": "VolSDF [49]" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 380, + 69, + 446, + 137 + ], + "blocks": [ + { + "bbox": [ + 380, + 69, + 446, + 137 + ], + "lines": [ + { + "bbox": [ + 380, + 69, + 446, + 137 + ], + "spans": [ + { + "bbox": [ + 380, + 69, + 446, + 137 + ], + "type": "image", + "image_path": "001ec6c4bba3259fac69d85d429761d68df4211302fb6f6501a5ef9646abab13.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 391, + 140, + 435, + 191 + ], + "blocks": [ + { + "bbox": [ + 391, + 140, + 435, + 191 + ], + "lines": [ + { + "bbox": [ + 391, + 140, + 435, + 191 + ], + "spans": [ + { + "bbox": [ + 391, + 140, + 435, + 191 + ], + "type": "image", + "image_path": "66b7334e3214922b4bfd6d9ca47560bf9aa901b5e41da365c6a931182c065aed.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 390, + 194, + 436, + 250 + ], + "blocks": [ + { + "bbox": [ + 390, + 194, + 436, + 250 + ], + "lines": [ + { + "bbox": [ + 390, + 194, + 436, + 250 + ], + "spans": [ + { + "bbox": [ + 390, + 194, + 436, + 250 + ], + "type": "image", + "image_path": "7b84905db6a1d7532728af65eaee0da712ef56c6ff04c02dace75ecbe3142dfc.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 394, + 251, + 435, + 261 + ], + "lines": [ + { + "bbox": [ + 394, + 251, + 435, + 261 + ], + "spans": [ + { + "bbox": [ + 394, + 251, + 435, + 261 + ], + "type": "text", + "content": "NeRF [35]" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 457, + 72, + 522, + 137 + ], + "blocks": [ + { + "bbox": [ + 457, + 72, + 522, + 137 + ], + "lines": [ + { + "bbox": [ + 457, + 72, + 522, + 137 + ], + "spans": [ + { + "bbox": [ + 457, + 72, + 522, + 137 + ], + "type": "image", + "image_path": "99b19a304992c3da9f2223019f8b55c9b4aac4317eb2a12e25434d9d65972f4c.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 466, + 140, + 509, + 191 + ], + "blocks": [ + { + "bbox": [ + 466, + 140, + 509, + 191 + ], + "lines": [ + { + "bbox": [ + 466, + 140, + 509, + 191 + ], + "spans": [ + { + "bbox": [ + 466, + 140, + 509, + 191 + ], + "type": "image", + "image_path": "8c4c75f75a1dab9cbddb85333073be17e792a370c33ea1b36fe37dc94c038b10.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 464, + 194, + 511, + 250 + ], + "blocks": [ + { + "bbox": [ + 464, + 194, + 511, + 250 + ], + "lines": [ + { + "bbox": [ + 464, + 194, + 511, + 250 + ], + "spans": [ + { + "bbox": [ + 464, + 194, + 511, + 250 + ], + "type": "image", + "image_path": "a92aaa181450ea1df784b097f1cad301ab19009f457f00147db230cc9b46c854.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 471, + 251, + 506, + 261 + ], + "lines": [ + { + "bbox": [ + 471, + 251, + 506, + 261 + ], + "spans": [ + { + "bbox": [ + 471, + 251, + 506, + 261 + ], + "type": "text", + "content": "IDR [50]" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "table", + "bbox": [ + 61, + 285, + 531, + 381 + ], + "blocks": [ + { + "bbox": [ + 61, + 285, + 531, + 381 + ], + "lines": [ + { + "bbox": [ + 61, + 285, + 531, + 381 + ], + "spans": [ + { + "bbox": [ + 61, + 285, + 531, + 381 + ], + "type": "table", + "html": "
Scan243740556365698397105106110114118122Mean
Colmap [41]0.812.050.731.221.791.581.023.051.402.051.001.320.490.781.171.36
IDR†[50]1.631.870.630.481.040.790.771.331.160.760.670.900.420.510.530.90
Nerf [35]1.901.601.850.582.281.271.471.672.051.070.882.531.061.150.961.49
UniSurf [38]1.321.361.720.441.350.790.801.491.370.890.591.470.460.590.621.02
VolSdf [49]1.141.260.810.491.250.700.721.291.180.700.661.080.420.610.550.86
Neus [46]1.371.210.730.401.200.700.721.011.160.820.661.690.390.490.510.87
NeuralWarp [9]0.490.710.380.380.790.810.821.201.060.680.660.740.410.630.510.68
Ours0.691.180.670.440.900.660.671.320.940.950.570.860.370.560.550.75
", + "image_path": "61e65738bb3ccc451ced233ab31e944e06ba330f9be012345a17f2ab6bcd4fd5.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 67, + 383, + 208, + 393 + ], + "lines": [ + { + "bbox": [ + 67, + 383, + 208, + 393 + ], + "spans": [ + { + "bbox": [ + 67, + 383, + 208, + 393 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 67, + 383, + 208, + 393 + ], + "type": "text", + "content": " Optimization using extra object masks." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 25 + }, + { + "bbox": [ + 185, + 393, + 406, + 403 + ], + "lines": [ + { + "bbox": [ + 185, + 393, + 406, + 403 + ], + "spans": [ + { + "bbox": [ + 185, + 393, + 406, + 403 + ], + "type": "text", + "content": "Table 2. Quantitative evaluation results on DTU [15] dataset." + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 413, + 287, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 413, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 413, + 287, + 581 + ], + "type": "text", + "content": "while the SDF based methods can only model the garments as closed surfaces and thus exhibit large errors. In Figure 4, the SDF based methods including NeuS, VolSDF and IDR enforce strong assumption on the closeness of the target shape and attempt to model all the shapes as closed surfaces, thus inevitably leading to large errors and inconsistent topologies for the garments. For example, for the first row, VolSDF attempts to use a double-layered surface to represent the thin cloth; optimizing this tiny SDF band is difficult and the results contain many noticeable holes. For the second row, NeuS gives a closed surface for the garment model while the cuffs are even incorrectly connected. Note although IDR takes extra masks for optimization, the reconstructed results still exhibit large quantitative errors." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 46, + 581, + 287, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 581, + 287, + 676 + ], + "spans": [ + { + "bbox": [ + 46, + 581, + 287, + 676 + ], + "type": "text", + "content": "Evaluation on closed objects. We further perform evaluation on DTU dataset whose objects are closed and watertight, and measure the Chamfer Distances between the predicted meshes and ground truth point clouds. Although our method adopts UDF as representation and does not impose closed surface assumption, our method still achieves comparable performance with the SDF based methods that are particularly designed for closed shapes." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "As shown in Table 2, in terms of the averaged Chamfer Distance, NeuralUDF outperforms most of the SOTA methods except the most recent work NeuralWarp [8]. Neural-" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 304, + 413, + 545, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 413, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 413, + 545, + 533 + ], + "type": "text", + "content": "Warp is a fine-tuning method that further optimizes the surfaces obtained from VolSDF [49] by minimizing the differences between the warped source images and the reference image. However, NeuralWarp heavily relies on the reconstructed geometry provided by other methods to generate reasonable warped source images, so it cannot be trained from scratch. The qualitative comparisons with SOTAs are provided in Figure 5, where our reconstruction results contain fewer artifacts than the SDF based methods, e.g., the roof of the House and the Apple." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 306, + 540, + 436, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 540, + 436, + 552 + ], + "spans": [ + { + "bbox": [ + 306, + 540, + 436, + 552 + ], + "type": "text", + "content": "4.3. Ablations and Analysis" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 304, + 557, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 545, + 713 + ], + "type": "text", + "content": "Ablation studies. We conduct ablation studies (Figure 6 and Figure 7) to investigate the individual contribution of the important designs of our method. The first key design is the gradient-aware visibility indicator function for effectively indicating the surfaces. For convenience, we call the visibility indicator function that is not gradient-aware as naive indicator function. As shown in Figure 6, the reconstruction result using the naive indicator function contains larger errors than the result using the gradient-aware indicator function, where the Chamfer Distance (CD) of the two reconstructions are 1.31 and 0.75 respectively (the lower is better). We can also see that the reconstruction using the naive function lacks geometric details, where the teeth and" + } + ] + } + ], + "index": 33 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20840" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 71, + 126, + 167 + ], + "blocks": [ + { + "bbox": [ + 51, + 71, + 126, + 167 + ], + "lines": [ + { + "bbox": [ + 51, + 71, + 126, + 167 + ], + "spans": [ + { + "bbox": [ + 51, + 71, + 126, + 167 + ], + "type": "image", + "image_path": "7f7e8d1d4201e5d61e877a0779d333081ce57048c1f0c6a0b4f0eeeacefb41a2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 48, + 170, + 111, + 186 + ], + "lines": [ + { + "bbox": [ + 48, + 170, + 111, + 186 + ], + "spans": [ + { + "bbox": [ + 48, + 170, + 111, + 186 + ], + "type": "text", + "content": "(a) Naive indicator function" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 132, + 72, + 208, + 169 + ], + "blocks": [ + { + "bbox": [ + 132, + 72, + 208, + 169 + ], + "lines": [ + { + "bbox": [ + 132, + 72, + 208, + 169 + ], + "spans": [ + { + "bbox": [ + 132, + 72, + 208, + 169 + ], + "type": "image", + "image_path": "31c1d23292fc68f8f429730d3cc5d004e5b21e062ffce9f65db8a4248c5617a5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 130, + 170, + 200, + 186 + ], + "lines": [ + { + "bbox": [ + 130, + 170, + 200, + 186 + ], + "spans": [ + { + "bbox": [ + 130, + 170, + 200, + 186 + ], + "type": "text", + "content": "(b) Gradient-aware indicator function" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 211, + 73, + 288, + 168 + ], + "blocks": [ + { + "bbox": [ + 211, + 73, + 288, + 168 + ], + "lines": [ + { + "bbox": [ + 211, + 73, + 288, + 168 + ], + "spans": [ + { + "bbox": [ + 211, + 73, + 288, + 168 + ], + "type": "image", + "image_path": "85e0c2f0093ed06b50b4809ccd74bdf97aae13ac2edf1c8988bdcdec04a00355.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 213, + 170, + 273, + 179 + ], + "lines": [ + { + "bbox": [ + 213, + 170, + 273, + 179 + ], + "spans": [ + { + "bbox": [ + 213, + 170, + 273, + 179 + ], + "type": "text", + "content": "(c) Patch blending" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 231, + 287, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 231, + 287, + 280 + ], + "spans": [ + { + "bbox": [ + 46, + 231, + 287, + 280 + ], + "type": "text", + "content": "eye holes of the Skull are not reconstructed accurately. This is because the naive indicator function has inherent bias and the measured visibility is not consistent with the true surface intersection, thus leading to inaccurate reconstruction." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 283, + 287, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 283, + 287, + 462 + ], + "spans": [ + { + "bbox": [ + 46, + 283, + 287, + 462 + ], + "type": "text", + "content": "We leverage an iso-surface regularizer to make the UDF field more stable and accurate around zero level sets. The reconstructions with or without the regularizer are shown in Figure 7. Although the Chamfer Distances of the two results are close, the mesh without the regularizer is much more noisy and has noticeable artifacts while the mesh with the regularizer is much cleaner and smoother. To give an in-depth analysis, we plot the UDF values and the cosine values of the angle between the ray direction and UDF gradients along a selected ray (the marked red point of (b,c)). We can see that, without the regularizer, the signs of the " + }, + { + "bbox": [ + 46, + 283, + 287, + 462 + ], + "type": "inline_equation", + "content": "\\cos (\\theta)" + }, + { + "bbox": [ + 46, + 283, + 287, + 462 + ], + "type": "text", + "content": " values change multiple times around the zero level set, which indicates the gradients are not stable. In contrast, with the regularizer, the gradients are more stable, where the gradient direction only flips at the surface intersection." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 466, + 288, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 466, + 288, + 514 + ], + "spans": [ + { + "bbox": [ + 46, + 466, + 288, + 514 + ], + "type": "text", + "content": "Also, inspired by [30], the patch blending loss term is included in the optimization. With this term, the reconstruction can be further improved from 0.75 to 0.66 in terms of Chamfer Distance." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 518, + 287, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 518, + 287, + 638 + ], + "spans": [ + { + "bbox": [ + 46, + 518, + 287, + 638 + ], + "type": "text", + "content": "More types of shapes. We additionally show results on several challenging objects with complex open surfaces. As shown in Figure 8, our method successfully reconstructs the challenging plant models with non-manifold geometries, while the reconstruction results of NeuS [46] are incomplete and contain numerous holes. This is because SDF is essentially not a suitable representation for many topologies like the thin leaves. Therefore, optimizing a SDF field for these kinds of models is deserved to be difficult and the reconstructed surfaces are naturally erroneous." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "text", + "content": "Limitations. Compared to the SDF based methods, NeuralUDF will degrade on textureless objects that lack enough distinguishable features, and produces smoother results that may lack some geometric details. This is because NeuralUDF does not impose closed surface assumption and is not differentiable at zero-level sets (see details in the Supp.)." + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 313, + 77, + 381, + 166 + ], + "blocks": [ + { + "bbox": [ + 313, + 77, + 381, + 166 + ], + "lines": [ + { + "bbox": [ + 313, + 77, + 381, + 166 + ], + "spans": [ + { + "bbox": [ + 313, + 77, + 381, + 166 + ], + "type": "image", + "image_path": "de32a3cd349c56e222bf269a1a0ac82360820e3ec4702daae15477e052b1f349.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 390, + 78, + 452, + 165 + ], + "blocks": [ + { + "bbox": [ + 390, + 78, + 452, + 165 + ], + "lines": [ + { + "bbox": [ + 390, + 78, + 452, + 165 + ], + "spans": [ + { + "bbox": [ + 390, + 78, + 452, + 165 + ], + "type": "image", + "image_path": "635b208294a045cd28b474ab370f80513ebacb52d8d35306bcfc2672c6d17608.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 401, + 165, + 440, + 174 + ], + "lines": [ + { + "bbox": [ + 401, + 165, + 440, + 174 + ], + "spans": [ + { + "bbox": [ + 401, + 165, + 440, + 174 + ], + "type": "text", + "content": "CD: 2.47" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 468, + 78, + 529, + 165 + ], + "blocks": [ + { + "bbox": [ + 468, + 78, + 529, + 165 + ], + "lines": [ + { + "bbox": [ + 468, + 78, + 529, + 165 + ], + "spans": [ + { + "bbox": [ + 468, + 78, + 529, + 165 + ], + "type": "image", + "image_path": "625fab0d2d5831f3392c9df5166a1c38592304c7f996fe6ab1573325c0cd3934.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 466, + 165, + 535, + 182 + ], + "lines": [ + { + "bbox": [ + 466, + 165, + 535, + 182 + ], + "spans": [ + { + "bbox": [ + 466, + 165, + 535, + 182 + ], + "type": "text", + "content": "CD: 2.49 (c) w/o regularizer" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 309, + 185, + 422, + 255 + ], + "blocks": [ + { + "bbox": [ + 47, + 191, + 287, + 214 + ], + "lines": [ + { + "bbox": [ + 47, + 191, + 287, + 214 + ], + "spans": [ + { + "bbox": [ + 47, + 191, + 287, + 214 + ], + "type": "text", + "content": "Figure 6. Ablation studies on gradient-aware visibility indicator function and patch blending loss term." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 310, + 173, + 386, + 183 + ], + "lines": [ + { + "bbox": [ + 310, + 173, + 386, + 183 + ], + "spans": [ + { + "bbox": [ + 310, + 173, + 386, + 183 + ], + "type": "text", + "content": "(a) Reference Image" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 309, + 185, + 422, + 255 + ], + "lines": [ + { + "bbox": [ + 309, + 185, + 422, + 255 + ], + "spans": [ + { + "bbox": [ + 309, + 185, + 422, + 255 + ], + "type": "image", + "image_path": "60a51f16ca4e5e28aa130b839df915df0823cc67fffdd6e727b9384b1538d78f.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 337, + 255, + 512, + 265 + ], + "lines": [ + { + "bbox": [ + 337, + 255, + 512, + 265 + ], + "spans": [ + { + "bbox": [ + 337, + 255, + 512, + 265 + ], + "type": "text", + "content": "(d) The curves of UDF values and " + }, + { + "bbox": [ + 337, + 255, + 512, + 265 + ], + "type": "inline_equation", + "content": "\\cos (\\theta)" + }, + { + "bbox": [ + 337, + 255, + 512, + 265 + ], + "type": "text", + "content": " values" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 325, + 266, + 526, + 277 + ], + "lines": [ + { + "bbox": [ + 325, + 266, + 526, + 277 + ], + "spans": [ + { + "bbox": [ + 325, + 266, + 526, + 277 + ], + "type": "text", + "content": "Figure 7. Ablation study on the iso-surface regularizer." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 422, + 185, + 543, + 255 + ], + "blocks": [ + { + "bbox": [ + 422, + 185, + 543, + 255 + ], + "lines": [ + { + "bbox": [ + 422, + 185, + 543, + 255 + ], + "spans": [ + { + "bbox": [ + 422, + 185, + 543, + 255 + ], + "type": "image", + "image_path": "8adaa7428266edcb54815f623de0c08df91c918d2ba3ba47f82612d465b15f5b.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 313, + 289, + 538, + 448 + ], + "blocks": [ + { + "bbox": [ + 313, + 289, + 538, + 448 + ], + "lines": [ + { + "bbox": [ + 313, + 289, + 538, + 448 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 538, + 448 + ], + "type": "image", + "image_path": "7c0345feff3e7963df5c10345d4c095375a37f2e189e1d0a7444fac532dddcc7.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 316, + 450, + 534, + 462 + ], + "lines": [ + { + "bbox": [ + 316, + 450, + 534, + 462 + ], + "spans": [ + { + "bbox": [ + 316, + 450, + 534, + 462 + ], + "type": "text", + "content": "Figure 8. Qualitative comparisons with NeuS on two plants." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 468, + 384, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 468, + 384, + 480 + ], + "spans": [ + { + "bbox": [ + 306, + 468, + 384, + 480 + ], + "type": "text", + "content": "5. Conclusions" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 489, + 545, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 489, + 545, + 621 + ], + "spans": [ + { + "bbox": [ + 304, + 489, + 545, + 621 + ], + "type": "text", + "content": "We introduce NeuralUDF, a NeRF-based method that uses neural UDF fields for multiview reconstruction of surfaces with arbitrary topologies. Our key idea is to introduce a novel volume density function that effectively correlates the UDF representation with the volume rendering scheme to learn neural UDF fields. The experiments show that, unlike the SDF based works that can only model closed shapes, our method can reconstruct more types of shapes with either open or closed surfaces, thus extending the underlying geometric representation of neural volume rendering to various topologies." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 306, + 633, + 409, + 646 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 633, + 409, + 646 + ], + "spans": [ + { + "bbox": [ + 306, + 633, + 409, + 646 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": "Xiaoxiao Long is supported by Hong Kong PhD Fellowship Scheme. Taku Komura is partly supported by the Hong Kong ITF Grant ITS/319/21FP. Lingjie Liu and Christian Theobalt have been supported by the ERC Consolidator Grant 4DReply (770784)." + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20841" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 91, + 289, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 289, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 289, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 289, + 135 + ], + "type": "text", + "content": "[1] Matan Atzmon and Yaron Lipman. Sal: Sign agnostic learning of shapes from raw data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2565-2574, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 288, + 202 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 288, + 202 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 288, + 202 + ], + "type": "text", + "content": "[2] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 204, + 288, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 204, + 288, + 258 + ], + "spans": [ + { + "bbox": [ + 53, + 204, + 288, + 258 + ], + "type": "text", + "content": "[3] Fausto Bernardini, Joshua Mittleman, Holly Rushmeier, Cláudio Silva, and Gabriel Taubin. The ball-pivoting algorithm for surface reconstruction. IEEE transactions on visualization and computer graphics, 5(4):349-359, 1999. 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 260, + 288, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 260, + 288, + 305 + ], + "spans": [ + { + "bbox": [ + 53, + 260, + 288, + 305 + ], + "type": "text", + "content": "[4] Neill DF Campbell, George Vogiatzis, Carlos Hernández, and Roberto Cipolla. Using multiple hypotheses to improve depth-maps for multi-view stereo. In European Conference on Computer Vision, pages 766-779. Springer, 2008. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 306, + 288, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 306, + 288, + 361 + ], + "spans": [ + { + "bbox": [ + 53, + 306, + 288, + 361 + ], + "type": "text", + "content": "[5] Weikai Chen, Cheng Lin, Weiyang Li, and Bo Yang. 3psdf: Three-pole signed distance function for learning surfaces with arbitrary topologies. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18522-18531, 2022. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 362, + 288, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 362, + 288, + 407 + ], + "spans": [ + { + "bbox": [ + 53, + 362, + 288, + 407 + ], + "type": "text", + "content": "[6] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5939-5948, 2019. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 407, + 288, + 451 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 407, + 288, + 451 + ], + "spans": [ + { + "bbox": [ + 53, + 407, + 288, + 451 + ], + "type": "text", + "content": "[7] Julian Chibane, Gerard Pons-Moll, et al. Neural unsigned distance fields for implicit function learning. Advances in Neural Information Processing Systems, 33:21638-21652, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 453, + 288, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 453, + 288, + 497 + ], + "spans": [ + { + "bbox": [ + 53, + 453, + 288, + 497 + ], + "type": "text", + "content": "[8] François Darmon, Bénédicte Bascle, Jean-Clément Devaux, Pascal Monasse, and Mathieu Aubry. Improving neural implicit surfaces geometry with patch warping. arXiv preprint arXiv:2112.09648, 2021. 2, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 498, + 288, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 498, + 288, + 553 + ], + "spans": [ + { + "bbox": [ + 53, + 498, + 288, + 553 + ], + "type": "text", + "content": "[9] François Darmon, Bénédicte Bascle, Jean-Clément Devaux, Pascal Monasse, and Mathieu Aubry. Improving neural implicit surfaces geometry with patch warping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6260–6269, 2022. 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 555, + 288, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 555, + 288, + 597 + ], + "spans": [ + { + "bbox": [ + 47, + 555, + 288, + 597 + ], + "type": "text", + "content": "[10] Yasutaka Furukawa and Jean Ponce. Accurate, dense, and robust multiview stereopsis. IEEE transactions on pattern analysis and machine intelligence, 32(8):1362-1376, 2009. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 600, + 288, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 600, + 288, + 644 + ], + "spans": [ + { + "bbox": [ + 47, + 600, + 288, + 644 + ], + "type": "text", + "content": "[11] Silvano Galliani, Katrin Lasinger, and Konrad Schindler. Massively parallel multiview stereopsis by surface normal diffusion. In Proceedings of the IEEE International Conference on Computer Vision, pages 873-881, 2015. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 646, + 288, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 646, + 288, + 678 + ], + "spans": [ + { + "bbox": [ + 47, + 646, + 288, + 678 + ], + "type": "text", + "content": "[12] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. arXiv preprint arXiv:2002.10099, 2020. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 680, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 680, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 680, + 288, + 714 + ], + "type": "text", + "content": "[13] Xiaodong Gu, Zhiwen Fan, Siyu Zhu, Zuozhuo Dai, Feitong Tan, and Ping Tan. Cascade cost volume for high-resolution multi-view stereo and stereo matching. In Proceedings of" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 714 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2495-2504, 2020. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 96, + 545, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 129 + ], + "type": "text", + "content": "[14] Benoit Guillard, Federico Stella, and Pascal Fua. Meshud: Fast and differentiable meshing of unsigned distance field networks. arXiv preprint arXiv:2111.14549, 2021. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 130, + 545, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 130, + 545, + 174 + ], + "spans": [ + { + "bbox": [ + 307, + 130, + 545, + 174 + ], + "type": "text", + "content": "[15] Rasmus Jensen, Anders Dahl, George Vogiatzis, Engin Tola, and Henrik Aanæs. Large scale multi-view stereopsis evaluation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 406-413, 2014. 2, 5, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 175, + 545, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 219 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 219 + ], + "type": "text", + "content": "[16] Mengqi Ji, Juergen Gall, Haitian Zheng, Yebin Liu, and Lu Fang. Surfacenet: An end-to-end 3d neural network for multiview stereopsis. In Proceedings of the IEEE International Conference on Computer Vision, pages 2307-2315, 2017. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 220, + 545, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 545, + 263 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 545, + 263 + ], + "type": "text", + "content": "[17] Mengqi Ji, Jinzhi Zhang, Qionghai Dai, and Lu Fang. SurfaceNet+: An end-to-end 3d neural network for very sparse multi-view stereopsis. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(11):4078-4093, 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 265, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 265, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 265, + 545, + 319 + ], + "type": "text", + "content": "[18] Yue Jiang, Dantong Ji, Zhizhong Han, and Matthias Zwicker. Sdfdiff: Differentiable rendering of signed distance fields for 3d shape optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1251-1261, 2020. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 320, + 545, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 320, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 307, + 320, + 545, + 354 + ], + "type": "text", + "content": "[19] Abhishek Kar, Christian Hane, and Jitendra Malik. Learning a multi-view stereo machine. Advances in neural information processing systems, 30, 2017. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 354, + 545, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 354, + 545, + 397 + ], + "spans": [ + { + "bbox": [ + 307, + 354, + 545, + 397 + ], + "type": "text", + "content": "[20] Michael Kazhdan, Matthew Bolitho, and Hugues Hoppe. Poisson surface reconstruction. In Proceedings of the fourth Eurographics symposium on Geometry processing, volume 7, 2006. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 399, + 545, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 545, + 431 + ], + "type": "text", + "content": "[21] Michael Kazhdan and Hugues Hoppe. Screened poisson surface reconstruction. ACM Transactions on Graphics (ToG), 32(3):1-13, 2013. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 433, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 433, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 307, + 433, + 545, + 487 + ], + "type": "text", + "content": "[22] Petr Kellnhofer, Lars C Jebe, Andrew Jones, Ryan Spicer, Kari Pulli, and Gordon Wetzstein. Neural lumigraph rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4287-4297, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 488, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 488, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 307, + 488, + 545, + 521 + ], + "type": "text", + "content": "[23] Kiriakos N Kutulakos and Steven M Seitz. A theory of shape by space carving. International journal of computer vision, 38(3):199-218, 2000. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 522, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 522, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 307, + 522, + 545, + 555 + ], + "type": "text", + "content": "[24] Der-Tsai Lee and Bruce J Schachter. Two algorithms for constructing a delaunay triangulation. International Journal of Computer & Information Sciences, 9(3):219-242, 1980. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 557, + 545, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 557, + 545, + 600 + ], + "spans": [ + { + "bbox": [ + 307, + 557, + 545, + 600 + ], + "type": "text", + "content": "[25] Maxime Lhuillier and Long Quan. A quasi-dense approach to surface reconstruction from uncalibrated images. IEEE transactions on pattern analysis and machine intelligence, 27(3):418-433, 2005. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 601, + 545, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 601, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 307, + 601, + 545, + 645 + ], + "type": "text", + "content": "[26] Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. Barf: Bundle-adjusting neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5741-5751, 2021. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "type": "text", + "content": "[27] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. Advances in Neural Information Processing Systems, 33:15651-15663, 2020. 1" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 691, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 714 + ], + "type": "text", + "content": "[28] Shaohui Liu, Yinda Zhang, Songyou Peng, Boxin Shi, Marc Pollefeys, and Zhaopeng Cui. Dist: Rendering deep implicit" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20842" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "text", + "content": "signed distance function with differentiable sphere tracing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2019-2028, 2020. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 287, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 287, + 161 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 287, + 161 + ], + "type": "text", + "content": "[29] Xiaoxiao Long, Cheng Lin, Lingjie Liu, Wei Li, Christian Theobalt, Ruigang Yang, and Wenping Wang. Adaptive surface normal constraint for depth estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12849-12858, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 162, + 288, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 162, + 288, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 162, + 288, + 205 + ], + "type": "text", + "content": "[30] Xiaoxiao Long, Cheng Lin, Peng Wang, Taku Komura, and Wenping Wang. Sparseneus: Fast generalizable neural surface reconstruction from sparse views. arXiv preprint arXiv:2206.05737, 2022. 5, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 206, + 288, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 206, + 288, + 260 + ], + "spans": [ + { + "bbox": [ + 48, + 206, + 288, + 260 + ], + "type": "text", + "content": "[31] Xiaoxiao Long, Lingjie Liu, Wei Li, Christian Theobalt, and Wenping Wang. Multi-view depth estimation using epipolar spatio-temporal networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8258-8267, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 261, + 288, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 261, + 288, + 304 + ], + "spans": [ + { + "bbox": [ + 48, + 261, + 288, + 304 + ], + "type": "text", + "content": "[32] Xiaoxiao Long, Lingjie Liu, Christian Theobalt, and Wenping Wang. Occlusion-aware depth estimation with adaptive normal constraints. In European Conference on Computer Vision, pages 640-657. Springer, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 305, + 288, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 305, + 288, + 360 + ], + "spans": [ + { + "bbox": [ + 48, + 305, + 288, + 360 + ], + "type": "text", + "content": "[33] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4460-4470, 2019. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 361, + 288, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 361, + 288, + 415 + ], + "spans": [ + { + "bbox": [ + 48, + 361, + 288, + 415 + ], + "type": "text", + "content": "[34] Mateusz Michalkiewicz, Jhony K Pontes, Dominic Jack, Mahsa Baktashmotlagh, and Anders Eriksson. Implicit surface representations as layers in neural networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4743-4752, 2019. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 415, + 288, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 288, + 470 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 288, + 470 + ], + "type": "text", + "content": "[35] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European conference on computer vision, pages 405-421. Springer, 2020. 1, 2, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 471, + 288, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 471, + 288, + 525 + ], + "spans": [ + { + "bbox": [ + 48, + 471, + 288, + 525 + ], + "type": "text", + "content": "[36] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11453-11464, 2021. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 525, + 288, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 288, + 581 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 288, + 581 + ], + "type": "text", + "content": "[37] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3504-3515, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 582, + 288, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 582, + 288, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 582, + 288, + 635 + ], + "type": "text", + "content": "[38] Michael Oechsle, Songyou Peng, and Andreas Geiger. Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5589-5599, 2021. 2, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 636, + 288, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 288, + 691 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 288, + 691 + ], + "type": "text", + "content": "[39] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 165-174, 2019. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 691, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 712 + ], + "type": "text", + "content": "[40] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 711 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "text", + "content": "Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865-5874, 2021. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 107, + 545, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 149 + ], + "type": "text", + "content": "[41] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 150, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 150, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 307, + 150, + 545, + 194 + ], + "type": "text", + "content": "[42] Johannes L Schonberger, Enliang Zheng, Jan-Michael Frahm, and Marc Pollefeys. Pixelwise view selection for unstructured multi-view stereo. In European Conference on Computer Vision, pages 501-518. Springer, 2016. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 194, + 547, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 194, + 547, + 226 + ], + "spans": [ + { + "bbox": [ + 307, + 194, + 547, + 226 + ], + "type": "text", + "content": "[43] Steven M Seitz and Charles R Dyer. Photorealistic scene reconstruction by voxel coloring. International Journal of Computer Vision, 35(2):151-173, 1999. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 227, + 545, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 227, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 307, + 227, + 545, + 281 + ], + "type": "text", + "content": "[44] Jiaming Sun, Yiming Xie, Linghao Chen, Xiaowei Zhou, and Hujun Bao. Neuralrecon: Real-time coherent 3d reconstruction from monocular video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15598-15607, 2021. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 282, + 545, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 282, + 545, + 324 + ], + "spans": [ + { + "bbox": [ + 307, + 282, + 545, + 324 + ], + "type": "text", + "content": "[45] Engin Tola, Christoph Strecha, and Pascal Fua. Efficient large-scale multi-view stereo for ultra high-resolution image sets. Machine Vision and Applications, 23(5):903-920, 2012. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 325, + 545, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 325, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 307, + 325, + 545, + 380 + ], + "type": "text", + "content": "[46] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. Advances in Neural Information Processing Systems, 34, 2021. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 380, + 545, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 380, + 545, + 424 + ], + "spans": [ + { + "bbox": [ + 307, + 380, + 545, + 424 + ], + "type": "text", + "content": "[47] Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. Mvsnet: Depth inference for unstructured multi-view stereo. In Proceedings of the European Conference on Computer Vision (ECCV), pages 767-783, 2018. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 425, + 545, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 425, + 545, + 479 + ], + "spans": [ + { + "bbox": [ + 307, + 425, + 545, + 479 + ], + "type": "text", + "content": "[48] Yao Yao, Zixin Luo, Shiwei Li, Tianwei Shen, Tian Fang, and Long Quan. Recurrent mvsnet for high-resolution multi-view stereo depth inference. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5525-5534, 2019. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 479, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 479, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 307, + 479, + 545, + 521 + ], + "type": "text", + "content": "[49] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems, 34, 2021. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 523, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 523, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 307, + 523, + 545, + 578 + ], + "type": "text", + "content": "[50] Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Basri Ronen, and Yaron Lipman. Multiview neural surface reconstruction by disentangling geometry and appearance. Advances in Neural Information Processing Systems, 33:2492-2502, 2020. 2, 5, 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 578, + 545, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 578, + 545, + 622 + ], + "spans": [ + { + "bbox": [ + 307, + 578, + 545, + 622 + ], + "type": "text", + "content": "[51] Jingyang Zhang, Yao Yao, and Long Quan. Learning signed distance field for multi-view surface reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6525-6534, 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 622, + 545, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 622, + 545, + 655 + ], + "spans": [ + { + "bbox": [ + 307, + 622, + 545, + 655 + ], + "type": "text", + "content": "[52] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492, 2020. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 655, + 545, + 711 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 655, + 545, + 711 + ], + "spans": [ + { + "bbox": [ + 307, + 655, + 545, + 711 + ], + "type": "text", + "content": "[53] Heming Zhu, Yu Cao, Hang Jin, Weikai Chen, Dong Du, Zhangye Wang, Shuguang Cui, and Xiaoguang Han. Deep fashion3d: A dataset and benchmark for 3d garment reconstruction from single images. In European Conference on Computer Vision, pages 512-530. Springer, 2020. 2, 5, 6" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20843" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Neuralangelo_ High-Fidelity Neural Surface Reconstruction/55f0b247-41a3-45d9-bec0-4490ab23bed9_content_list.json b/2023/Neuralangelo_ High-Fidelity Neural Surface Reconstruction/55f0b247-41a3-45d9-bec0-4490ab23bed9_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8327854d1e9b599b45fe37dfd770a38a17bf813c --- /dev/null +++ b/2023/Neuralangelo_ High-Fidelity Neural Surface Reconstruction/55f0b247-41a3-45d9-bec0-4490ab23bed9_content_list.json @@ -0,0 +1,1752 @@ +[ + { + "type": "text", + "text": "Neuralangelo: High-Fidelity Neural Surface Reconstruction", + "text_level": 1, + "bbox": [ + 181, + 130, + 787, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhaoshuo Li $^{1,2}$ Thomas Müller $^{1}$ Alex Evans $^{1}$ Russell H. Taylor $^{2}$ Mathias Unberath $^{2}$ Ming-Yu Liu $^{1}$ Chen-Hsuan Lin $^{1}$ $^{1}$ NVIDIA Research $^{2}$ Johns Hopkins University https://research.nvidia.com/labs/dir/neuralangelo", + "bbox": [ + 114, + 178, + 852, + 260 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/58f396abf90bacdda4f6894b60e1a077eefff62560d389ed557b1b5c4dcc227b.jpg", + "image_caption": [ + "Figure 1. We present Neuralangelo, a framework for high-fidelity 3D surface reconstruction from RGB images using neural volume rendering, even without auxiliary data such as segmentation or depth. Shown in the figure is an extracted 3D mesh of a courthouse." + ], + "image_footnote": [], + "bbox": [ + 78, + 280, + 895, + 578 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 637, + 313, + 652 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Neural surface reconstruction has been shown to be powerful for recovering dense 3D surfaces via image-based neural rendering. However, current methods struggle to recover detailed structures of real-world scenes. To address the issue, we present Neuralangelo, which combines the representation power of multi-resolution 3D hash grids with neural surface rendering. Two key ingredients enable our approach: (1) numerical gradients for computing higher-order derivatives as a smoothing operation and (2) coarse-to-fine optimization on the hash grids controlling different levels of details. Even without auxiliary inputs such as depth, Neuralangelo can effectively recover dense 3D surface structures from multi-view images with fidelity significantly surpassing previous methods, enabling detailed large-scale scene reconstruction from RGB video captures.", + "bbox": [ + 73, + 666, + 473, + 893 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 501, + 637, + 630, + 652 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D surface reconstruction aims to recover dense geometric scene structures from multiple images observed at different viewpoints [9]. The recovered surfaces provide structural information useful for many downstream applications, such as 3D asset generation for augmented/virtual/mixed reality or environment mapping for autonomous navigation of robotics. Photogrammetric surface reconstruction using a monocular RGB camera is of particular interest, as it equips users with the capability of casually creating digital twins of the real world using ubiquitous mobile devices.", + "bbox": [ + 496, + 667, + 895, + 819 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Classically, multi-view stereo algorithms [6, 16, 29, 34] had been the method of choice for sparse 3D reconstruction. An inherent drawback of these algorithms, however, is their inability to handle ambiguous observations, e.g. regions with large areas of homogeneous colors, repetitive texture", + "bbox": [ + 496, + 824, + 895, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "8456", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "patterns, or strong color variations. This would result in inaccurate reconstructions with noisy or missing surfaces. Recently, neural surface reconstruction methods [36,41,42] have shown great potential in addressing these limitations. This new class of methods uses coordinate-based multi-layer perceptrons (MLPs) to represent the scene as an implicit function, such as occupancy fields [25] or signed distance functions (SDF) [36,41,42]. Leveraging the inherent continuity of MLPs and neural volume rendering [22], these techniques allow the optimized surfaces to meaningfully interpolate between spatial locations, resulting in smooth and complete surface representations.", + "bbox": [ + 75, + 90, + 472, + 271 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite the superiority of neural surface reconstruction methods over classical approaches, the recovered fidelity of current methods does not scale well with the capacity of MLPs. Recently, Müller et al. [23] proposed a new scalable representation, referred to as Instant NGP (Neural Graphics Primitives). Instant NGP introduces a hybrid 3D grid structure with a multi-resolution hash encoding and a lightweight MLP that is more expressive with a memory footprint log-linear to the resolution. The proposed hybrid representation greatly increases the representation power of neural fields and has achieved great success at representing very fine-grained details for a wide variety of tasks, such as object shape representation and novel view synthesis problems.", + "bbox": [ + 75, + 273, + 472, + 468 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose Neuralangelo for high-fidelity surface reconstruction (Fig. 1). Neuralangelo adopts Instant NGP as a neural SDF representation of the underlying 3D scene, optimized from multi-view image observations via neural surface rendering [36]. We present two findings central to fully unlocking the potentials of multi-resolution hash encodings. First, using numerical gradients to compute higher-order derivatives, such as surface normals for the eikonal regularization [8, 12, 20, 42], is critical to stabilizing the optimization. Second, a progressive optimization schedule plays an important role in recovering the structures at different levels of details. We combine these two key ingredients and, via extensive experiments on standard benchmarks and real-world scenes, demonstrate significant improvements over image-based neural surface reconstruction methods in both reconstruction accuracy and view synthesis quality.", + "bbox": [ + 75, + 469, + 472, + 710 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, we present the following contributions:", + "bbox": [ + 96, + 710, + 442, + 724 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We present the Neuralangelo framework to naturally incorporate the representation power of multi-resolution hash encoding [23] into neural SDF representations.", + "- We present two simple techniques to improve the quality of hash-encoded surface reconstruction: higher-order derivatives with numerical gradients and coarse-to-fine optimization with a progressive level of details.", + "- We empirically demonstrate the effectiveness of Neuralangelo on various datasets, showing significant improvements over previous methods." + ], + "bbox": [ + 91, + 739, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 500, + 89, + 637, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multi-view surface reconstruction. Early image-based photogrammetry techniques use a volumetric occupancy grid to represent the scene [4, 16, 17, 29, 32]. Each voxel is visited and marked occupied if strict color constancy between the corresponding projected image pixels is satisfied. The photometric consistency assumption typically fails due to autoexposure or non-Lambertian materials, which are ubiquitous in the real world. Relaxing such color constancy constraints across views is important for realistic 3D reconstruction.", + "bbox": [ + 496, + 119, + 893, + 255 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Follow-up methods typically start with 3D point clouds from multi-view stereo techniques [6,7,28,34] and then perform dense surface reconstruction [13,14]. Reliance on the quality of the generated point clouds often leads to missing or noisy surfaces. Recent learning-based approaches augment the point cloud generation process with learned image features and cost volume construction [2,10,40]. However, these approaches are inherently limited by the resolution of the cost volume and fail to recover geometric details.", + "bbox": [ + 496, + 256, + 893, + 392 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Neural Radiance Fields (NeRF). NeRF [22] achieves remarkable photorealistic view synthesis with view-dependent effects. NeRF encodes 3D scenes with an MLP mapping 3D spatial locations to color and volume density. These predictions are composited into pixel colors using neural volume rendering. A problem of NeRF and its variants [1,30,43,46], however, is the question of how an isosurface of the volume density could be defined to represent the underlying 3D geometry. Current practice often relies on heuristic thresholding on the density values; due to insufficient constraints on the level sets, however, such surfaces are often noisy and may not model the scene structures accurately [36,41]. Therefore, more direct modeling of surfaces is preferred for photogrammetric surface reconstruction problems.", + "bbox": [ + 496, + 397, + 893, + 609 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Neural surface reconstruction. For scene representations with better-defined 3D surfaces, implicit functions such as occupancy grids [24, 25] or SDFs [42] are preferred over simple volume density fields. To integrate with neural volume rendering [22], different techniques [36, 41] have been proposed to reparametrize the underlying representations back to volume density. These designs of neural implicit functions enable more accurate surface prediction with view synthesis capabilities of unsacrificed quality [42].", + "bbox": [ + 496, + 614, + 893, + 750 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Follow-up works extend the above approaches to real-time at the cost of surface fidelity [18, 37], while others [3,5,44] use auxiliary information to enhance the reconstruction results. Notably, NeuralWarp [3] uses patch warping given co-visibility information from structure-from-motion (SfM) to guide surface optimization, but the patchwise planar assumption fails to capture highly-varying surfaces [3]. Other methods [5,45] utilize sparse point clouds from SfM to supervise the SDF, but their performances are upper-bounded by the quality of the point clouds, as with", + "bbox": [ + 496, + 750, + 893, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "8457", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "classical approaches [45]. The use of monocular depth and segmentation as auxiliary data has also been explored with unconstrained image collections [31] or using scene representations with hash encodings [44]. In contrast, our work Neuralangelo builds upon hash encodings [23] to recover surfaces but without the need for auxiliary inputs used in prior work [3, 5, 31, 44, 45]. Concurrent work [38] also proposes coarse-to-fine optimization for improved surface details, where a displacement network corrects the shape predicted by a coarse network. In contrast, we use hierarchical hash grids and control the level of details based on our analysis of higher-order derivatives.", + "bbox": [ + 75, + 90, + 472, + 272 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Approach", + "text_level": 1, + "bbox": [ + 76, + 285, + 184, + 301 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Neuralangelo reconstructs dense structures of the scene from multi-view images. Neuralangelo samples 3D locations along camera view directions and uses a multi-resolution hash encoding to encode the positions. The encoded features are input to an SDF MLP and a color MLP to composite images using SDF-based volume rendering.", + "bbox": [ + 75, + 310, + 468, + 401 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 76, + 410, + 217, + 424 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Neural volume rendering. NeRF [22] represents a 3D scene as volume density and color fields. Given a posed camera and a ray direction, the volume rendering scheme integrates the color radiance of sampled points along the ray. The $i$ -th sampled 3D position $\\mathbf{x}_i$ is at a distance $t_i$ from the camera center. The volume density $\\sigma_i$ and color $\\mathbf{c}_i$ of each sampled point are predicted using a coordinate MLP. The rendered color of a given pixel is approximated as the Riemann sum:", + "bbox": [ + 75, + 433, + 468, + 554 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {c}} (\\mathbf {o}, \\mathbf {d}) = \\sum_ {i = 1} ^ {N} w _ {i} \\mathbf {c} _ {i}, \\quad \\text {w h e r e} w _ {i} = T _ {i} \\alpha_ {i}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 564, + 468, + 606 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here, $\\alpha_{i} = 1 - \\exp (-\\sigma_{i}\\delta_{i})$ is the opacity of the $i$ -th ray segment, $\\delta_{i} = t_{i + 1} - t_{i}$ is the distance between adjacent samples, and $T_{i} = \\Pi_{j = 1}^{i - 1}(1 - \\alpha_{j})$ is the accumulated transmittance, indicating the fraction of light that reaches the camera. To supervise the network, a color loss is used between input images $\\mathbf{c}$ and rendered images $\\hat{\\mathbf{c}}$ :", + "bbox": [ + 75, + 614, + 472, + 705 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {R G B}} = \\left\\| \\hat {\\mathbf {c}} - \\mathbf {c} \\right\\| _ {1}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 207, + 717, + 468, + 734 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, surfaces are not clearly defined using such density formulation. Extracting surfaces from density-based representation often leads to noisy and unrealistic results [36, 41].", + "bbox": [ + 75, + 744, + 470, + 789 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Volume rendering of SDF. One of the most common surface representations is SDF. The surface $S$ of an SDF can be implicitly represented by its zero-level set, i.e., $S = \\{\\mathbf{x} \\in \\mathbb{R}^3 | f(\\mathbf{x}) = 0\\}$ , where $f(\\mathbf{x})$ is the SDF value. In the context of neural SDFs, Wang et al. [36] proposed to convert volume density predictions in NeRF to SDF representations with a logistic function to allow optimization with neural volume", + "bbox": [ + 75, + 794, + 470, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "rendering. Given a 3D point $\\mathbf{x}_i$ and SDF value $f(\\mathbf{x}_i)$ , the corresponding opacity value $\\alpha_i$ used in Eq. 1 is computed as", + "bbox": [ + 498, + 90, + 890, + 121 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha_ {i} = \\max \\left(\\frac {\\Phi_ {s} (f (\\mathbf {x} _ {i})) - \\Phi_ {s} (f (\\mathbf {x} _ {i + 1}))}{\\Phi_ {s} (f (\\mathbf {x} _ {i}))}, 0\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 544, + 131, + 890, + 165 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\Phi_s$ is the sigmoid function. In this work, we use the same SDF-based volume rendering formulation [36].", + "bbox": [ + 498, + 167, + 890, + 198 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multi-resolution hash encoding. Recently, multi-resolution hash encoding proposed by Müller et al. [23] has shown great scalability for neural scene representations, generating fine-grained details for tasks such as novel view synthesis. In Neuralangelo, we adopt the representation power of hash encoding to recover high-fidelity surfaces.", + "bbox": [ + 498, + 203, + 893, + 292 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The hash encoding uses multi-resolution grids, with each grid cell corner mapped to a hash entry. Each hash entry stores the encoding feature. Let $\\{V_1,\\dots,V_L\\}$ be the set of different spatial grid resolutions. Given an input position $\\mathbf{x}_i$ , we map it to the corresponding position at each grid resolution $V_{l}$ as $\\mathbf{x}_{i,l} = \\mathbf{x}_i\\cdot V_l$ . The feature vector $\\gamma_l(\\mathbf{x}_{i,l})\\in \\mathbb{R}^c$ given resolution $V_{l}$ is obtained via trilinear interpolation of hash entries at the grid cell corners. The encoding features across all spatial resolutions are concatenated together, forming a $\\gamma (\\mathbf{x}_i)\\in \\mathbb{R}^{cL}$ feature vector:", + "bbox": [ + 498, + 294, + 893, + 445 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\gamma \\left(\\mathbf {x} _ {i}\\right) = \\left(\\gamma_ {1} \\left(\\mathbf {x} _ {i, 1}\\right), \\dots , \\gamma_ {L} \\left(\\mathbf {x} _ {i, L}\\right)\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 452, + 890, + 470 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The encoded features are then passed to a shallow MLP.", + "bbox": [ + 498, + 477, + 870, + 491 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "One alternative to hash encoding is sparse voxel structures [30, 33, 39, 43], where each grid corner is uniquely defined without collision. However, volumetric feature grids require hierarchical spatial decomposition (e.g. octrees) to make the parameter count tractable; otherwise, the memory would grow cubically with spatial resolution. Given such hierarchy, finer voxel resolutions by design cannot recover surfaces that are misrepresented by the coarser resolutions [33]. Hash encoding instead assumes no spatial hierarchy and resolves collision automatically based on gradient averaging [23].", + "bbox": [ + 496, + 492, + 890, + 643 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Numerical Gradient Computation", + "text_level": 1, + "bbox": [ + 500, + 650, + 797, + 666 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We show in this section that the analytical gradient w.r.t. position of hash encoding suffers from localities. Therefore, optimization updates only propagate to local hash grids, lacking non-local smoothness. We propose a simple fix to such a locality problem by using numerical gradients. An overview is shown in Fig. 2.", + "bbox": [ + 496, + 674, + 893, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A special property of SDF is its differentiability with a gradient of the unit norm. The gradient of SDF satisfies the eikonal equation $\\| \\nabla f(\\mathbf{x})\\| _2 = 1$ (almost everywhere). To enforce the optimized neural representation to be a valid SDF, the eikonal loss [8] is typically imposed on the SDF predictions:", + "bbox": [ + 496, + 763, + 893, + 854 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {e i k}} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left(\\| \\nabla f (\\mathbf {x} _ {i}) \\| _ {2} - 1\\right) ^ {2}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 862, + 890, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "8458", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/75ebad35cb2bbf8018124cffdf59a378eb3f80743bbce21148a13ad52a882865.jpg", + "image_caption": [ + "Figure 2. Using numerical gradients for higher-order derivatives distributes the back-propagation updates beyond the local hash grid cell, thus becoming a smoothed version of analytical gradients." + ], + "image_footnote": [], + "bbox": [ + 88, + 89, + 467, + 210 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $N$ is the total number of sampled points. To allow for end-to-end optimization, a double backward operation on the SDF prediction $f(\\mathbf{x})$ is required.", + "bbox": [ + 75, + 289, + 470, + 335 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The de facto method for computing surface normals of SDFs $\\nabla f(\\mathbf{x})$ is to use analytical gradients [36,41,42]. Analytical gradients of hash encoding w.r.t. position, however, are not continuous across space under trilinear interpolation. To find the sampling location in a voxel grid, each 3D point $\\mathbf{x}_i$ would first be scaled by the grid resolution $V_{l}$ , written as $\\mathbf{x}_{i,l} = \\mathbf{x}_i\\cdot V_l$ . Let the coefficient for (tri-)linear interpolation be $\\beta = \\mathbf{x}_{i,l} - \\left\\lfloor \\mathbf{x}_{i,l}\\right\\rfloor$ . The resulting feature vectors are", + "bbox": [ + 75, + 334, + 472, + 455 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\gamma_ {l} \\left(\\mathbf {x} _ {i, l}\\right) = \\gamma_ {l} \\left(\\left\\lfloor \\mathbf {x} _ {i, l} \\right\\rfloor\\right) \\cdot (1 - \\beta) + \\gamma_ {l} \\left(\\left\\lceil \\mathbf {x} _ {i, l} \\right\\rceil\\right) \\cdot \\beta , \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 465, + 470, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the rounded position $\\lfloor \\mathbf{x}_{i,l}\\rfloor ,\\lceil \\mathbf{x}_{i,l}\\rceil$ correspond to the local grid cell corners. We note that rounding operations $\\lfloor \\cdot \\rfloor$ and $\\lceil \\cdot \\rceil$ are non-differentiable. As a result, the derivative of hash encoding w.r.t. the position can be obtained as", + "bbox": [ + 75, + 494, + 468, + 555 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\frac {\\partial \\gamma_ {l} (\\mathbf {x} _ {i , l})}{\\partial \\mathbf {x} _ {i}} = \\gamma_ {l} (\\lfloor \\mathbf {x} _ {i, l} \\rfloor) \\cdot (- \\frac {\\partial \\beta}{\\partial \\mathbf {x} _ {i , l}}) + \\gamma_ {l} (\\lceil \\mathbf {x} _ {i, l} \\rceil) \\cdot \\frac {\\partial \\beta}{\\partial \\mathbf {x} _ {i , l}} \\\\ = \\gamma_ {l} (\\left\\lfloor \\mathbf {x} _ {i, l} \\right\\rfloor) \\cdot (- V _ {l}) + \\gamma_ {l} (\\left\\lceil \\mathbf {x} _ {i, l} \\right\\rceil) \\cdot V _ {l}. \\tag {7} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 564, + 468, + 619 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The derivative of hash encoding is local, i.e., when $\\mathbf{x}_i$ moves across grid cell borders, the corresponding hash entries will be different. Therefore, the eikonal loss defined in Eq. 5 only back-propagates to the locally sampled hash entries, i.e. $\\gamma_l(\\lfloor \\mathbf{x}_{i,l}\\rfloor)$ and $\\gamma_l(\\lceil \\mathbf{x}_{i,l}\\rceil)$ . When continuous surfaces (e.g. a flat wall) span multiple grid cells, these grid cells should produce coherent surface normals without sudden transitions. To ensure consistency in surface representation, joint optimization of these grid cells is desirable. However, the analytical gradient is limited to local grid cells, unless all corresponding grid cells happen to be sampled and optimized simultaneously. Such sampling is not always guaranteed.", + "bbox": [ + 75, + 628, + 468, + 809 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To overcome the locality of the analytical gradient of hash encoding, we propose to compute the surface normals using numerical gradients. If the step size of the numerical gradient is smaller than the grid size of hash encoding, the numerical gradient would be equivalent to the analytical gradient; otherwise, hash entries of multiple grid cells", + "bbox": [ + 75, + 810, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "would participate in the surface normal computation. Backpropagating through the surface normals thus allows hash entries of multiple grids to receive optimization updates simultaneously. Intuitively, numerical gradients with carefully chosen step sizes can be interpreted as a smoothing operation on the analytical gradient expression. An alternative of normal supervision is a teacher-student curriculum [35, 47], where the predicted noisy normals are driven towards MLP outputs to exploit the smoothness of MLPs. However, analytical gradients from such teacher-student losses still only back-propagate to local grid cells for hash encoding. In contrast, numerical gradients solve the locality issue without the need of additional networks.", + "bbox": [ + 496, + 90, + 893, + 286 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To compute the surface normals using the numerical gradient, additional SDF samples are needed. Given a sampled point $\\mathbf{x}_i = (x_i,y_i,z_i)$ , we additionally sample two points along each axis of the canonical coordinate around $x_{i}$ within a vicinity of a step size of $\\epsilon$ . For example, the $x$ -component of the surface normal can be found as", + "bbox": [ + 496, + 287, + 893, + 377 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\nabla_ {x} f (\\mathbf {x} _ {i}) = \\frac {f \\left(\\gamma \\left(\\mathbf {x} _ {i} + \\boldsymbol {\\epsilon} _ {x}\\right)\\right) - f \\left(\\gamma \\left(\\mathbf {x} _ {i} - \\boldsymbol {\\epsilon} _ {x}\\right)\\right)}{2 \\epsilon}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 382, + 892, + 414 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\epsilon_{x} = [\\epsilon, 0, 0]$ . In total, six additional SDF samples are required for numerical surface normal computation.", + "bbox": [ + 496, + 420, + 890, + 452 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Progressive Levels of Details", + "text_level": 1, + "bbox": [ + 498, + 459, + 754, + 474 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Coarse-to-fine optimization can better shape the loss landscape to avoid falling into false local minima. Such a strategy has found many applications in computer vision, such as image-based registration [19,21,26]. Neuralangelo also adopts a coarse-to-fine optimization scheme to reconstruct the surfaces with progressive levels of details. Using numerical gradients for the higher-order derivatives naturally enables Neuralangelo to perform coarse-to-fine optimization from two perspectives.", + "bbox": [ + 496, + 482, + 893, + 619 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Step size $\\epsilon$ . As previously discussed, numerical gradients can be interpreted as a smoothing operation where the step size $\\epsilon$ controls the resolution and the amount of recovered details. Imposing $\\mathcal{L}_{\\mathrm{eik}}$ with a larger $\\epsilon$ for numerical surface normal computation ensures the surface normal is consistent at a larger scale, thus producing consistent and continuous surfaces. On the other hand, imposing $\\mathcal{L}_{\\mathrm{eik}}$ with a smaller $\\epsilon$ affects a smaller region and avoids smoothing details. In practice, we initialize the step size $\\epsilon$ to the coarsest hash grid size and exponentially decrease it matching different hash grid sizes throughout the optimization process.", + "bbox": [ + 496, + 623, + 893, + 790 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Hash grid resolution $V$ . If all hash grids are activated from the start of the optimization, to capture geometric details, fine hash grids must first \"unlearn\" from the coarse optimization with large step size $\\epsilon$ and \"relearn\" with a smaller $\\epsilon$ . If such a process is unsuccessful due to converged optimization, geometric details would be lost. Therefore, we only enable an initial set of coarse hash grids and progressively activate", + "bbox": [ + 496, + 794, + 893, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "8459", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/aa07006df64d5950aea79119972ef9c9246d1e8b40b82cd7a60d66768e6013d0.jpg", + "image_caption": [ + "Figure 3. Qualitative comparison on the DTU benchmark [11]. Neuralangelo produces more accurate and higher-fidelity surfaces." + ], + "image_footnote": [], + "bbox": [ + 76, + 88, + 893, + 354 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "finer hash grids throughout optimization when $\\epsilon$ decreases to their spatial size. The relearning process can thus be avoided to better capture the details. In practice, we also apply weight decay over all parameters to avoid single-resolution features dominating the final results.", + "bbox": [ + 75, + 405, + 470, + 481 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Optimization", + "text_level": 1, + "bbox": [ + 76, + 488, + 215, + 503 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To further encourage the smoothness of the reconstructed surfaces, we impose a prior by regularizing the mean curvature of SDF. The mean curvature is computed from discrete Laplacian similar to the surface normal computation, otherwise, the second-order analytical gradients of hash encoding are zero everywhere when using trilinear interpolation. The curvature loss $\\mathcal{L}_{\\mathrm{curv}}$ is defined as:", + "bbox": [ + 75, + 512, + 470, + 617 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {c u r v}} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left| \\nabla^ {2} f (\\mathbf {x} _ {i}) \\right|. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 625, + 468, + 666 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We note that the samples used for the surface normal computation in Eq. 8 are sufficient for curvature computation.", + "bbox": [ + 76, + 672, + 470, + 703 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The total loss is defined as the weighted sum of losses:", + "bbox": [ + 96, + 704, + 457, + 718 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\mathrm {R G B}} + w _ {\\mathrm {e i k}} \\mathcal {L} _ {\\mathrm {e i k}} + w _ {\\mathrm {c u r v}} \\mathcal {L} _ {\\mathrm {c u r v}}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 727, + 468, + 744 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "All network parameters, including MLPs and hash encoding, are trained jointly end-to-end.", + "bbox": [ + 76, + 752, + 470, + 782 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 794, + 209, + 811 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. Following prior work, we conduct experiments on 15 object-centric scenes of the DTU dataset [11]. Each scene has 49 or 64 images captured by a robot-held monocular RGB camera. The ground truth is obtained from a structured-light scanner. We further conduct experiments on 6 scenes", + "bbox": [ + 76, + 824, + 470, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "of the Tanks and Temples dataset [15], including large-scale indoor/outdoor scenes. Each scene contains 263 to 1107 images captured using a hand-held monocular RGB camera. The ground truth is obtained using a LiDAR sensor.", + "bbox": [ + 496, + 405, + 893, + 465 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation details. Our hash encoding resolution spans $2^{5}$ to $2^{11}$ with 16 levels. Each hash entry has a channel size of 8. The maximum number of hash entries of each resolution is $2^{22}$ . We activate 4 and 8 hash resolutions at the beginning of optimization for DTU dataset and Tanks and Temples respectively, due to differences in scene scales. We enable a new hash resolution every 5000 iterations when the step size $\\epsilon$ equals its grid cell size. For all experiments, we do not utilize auxiliary data such as segmentation or depth during the optimization process.", + "bbox": [ + 496, + 470, + 892, + 621 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation criteria. We report Chamfer distance and F1 score for surface evaluation [11, 15]. We use peak signal-to-noise ratio (PSNR) to report image synthesis qualities.", + "bbox": [ + 496, + 626, + 893, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. DTU Benchmark", + "text_level": 1, + "bbox": [ + 500, + 681, + 668, + 695 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We show qualitative results in Fig. 3 and quantitative results in Table 1. On average, Neuralangelo achieves the lowest Chamfer distance and the highest PSNR, even without using auxiliary inputs. The result suggests that Neuralangelo is more generally applicable than prior work when recovering surfaces and synthesizing images, despite not performing best in every individual scene.", + "bbox": [ + 496, + 704, + 893, + 809 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We further ablate Neuralangelo against the following conditions: 1) AG: analytical gradients, 2) $\\mathrm{AG + P}$ : analytical gradients and progressive activating hash resolutions, 3) NG: numerical gradients with varying $\\epsilon$ . Fig. 4 shows the results qualitatively. AG produces noisy surfaces, even with hash resolutions progressively activated (AG+P). NG improves", + "bbox": [ + 496, + 810, + 893, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "8460", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/71f810f95a3ffcb08e0c9efc4c5b5fa1872c8b2edfd37c17ab9e23582ed2f384.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 80, + 89, + 238, + 224 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c07bb2e00958992cfbd75cd3c96dec67113af298b84f477b1cb0a9b5137f1255.jpg", + "image_caption": [ + "AG" + ], + "image_footnote": [], + "bbox": [ + 243, + 90, + 401, + 224 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2d79fb8d5d9826cbb9b4feadd0986438bd89feb2f0532af362ec668c4a67bcf7.jpg", + "image_caption": [ + "AG+P", + "Figure 4. Qualitative comparison of different coarse-to-fine optimization scheme. When using the analytical gradient (AG and $\\mathrm{AG + P}$ ), coarse surfaces often contain artifacts. While using numerical gradients (NG) leads to a better coarse shape, details are also smoothed. Our solution $(\\mathrm{NG} + \\mathrm{P})$ produces both smooth surfaces and fine details." + ], + "image_footnote": [], + "bbox": [ + 403, + 90, + 562, + 224 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/62c4f776a7767f6fb5334eeee9c5c34bf35928499d773f69ca6d7b15240cc68a.jpg", + "image_caption": [ + "NG" + ], + "image_footnote": [], + "bbox": [ + 563, + 90, + 723, + 224 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7426cfc657beb8d7e82ba0307ff20f336da469f1e77942152e53152ee1319d51.jpg", + "image_caption": [ + "NG+P (Ours)" + ], + "image_footnote": [], + "bbox": [ + 725, + 90, + 885, + 224 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/4e3cd3538bd55c1024bf40961aefba9a884356b3b56d0e0ce296fd4697a2cc8a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
243740556365698397105106110114118122Mean
NeRF [22]1.901.601.850.582.281.271.471.672.051.070.882.531.061.150.961.49
VolSDF [41]1.141.260.810.491.250.700.721.291.180.700.661.080.420.610.550.86
NeuS [36]1.001.370.930.431.100.650.571.481.090.830.521.200.350.490.540.84
HF-NeuS [38]0.761.320.700.391.060.630.631.151.120.800.521.220.330.490.500.77
RegSDF [45] †0.601.410.640.431.340.620.600.900.921.020.600.590.300.410.390.72
NeuralWarp [3] †0.490.710.380.380.790.810.821.201.060.680.660.740.410.630.510.68
AG0.671.040.840.391.431.231.111.241.540.850.501.010.370.510.440.88
AG+P0.590.950.460.341.190.700.791.191.370.690.490.930.330.440.440.73
NG0.480.810.430.350.890.710.611.261.060.740.470.790.330.450.430.65
NG+P (Ours)0.370.720.350.350.870.540.531.290.970.730.470.740.320.410.430.61
RegSDF [45] †24.7823.0623.4722.2128.5725.5321.8128.8926.8127.9124.7125.1326.8421.6728.2525.31
NeuS [36]26.6223.6426.4325.5930.6132.8329.2433.7126.8531.9732.1828.9228.4135.0034.8129.79
VolSDF [41]26.2825.6126.5526.7631.5731.5029.3833.2328.0332.1333.1631.4930.3334.9034.7530.38
NeRF [22]26.2425.7426.7927.5731.9631.5029.5832.7828.3532.0833.4931.5431.0035.5935.5130.65
AG29.9724.9823.1130.2730.6031.2729.2734.2227.4733.0933.8529.9829.4135.6935.1130.55
AG+P30.1224.6329.5930.2931.6032.0429.8534.1927.8233.2333.9529.1529.4435.9935.6731.17
NG30.3425.1430.2030.7931.7231.8629.8134.3628.0133.4534.3830.3929.8836.0235.7431.47
NG+P (Ours)30.6427.7832.7034.1835.1535.8931.4736.8230.1335.9236.6132.6031.2038.4138.0533.84
", + "bbox": [ + 76, + 306, + 895, + 566 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Quantitative results on DTU dataset [11]. Neuralangelo achieves the best reconstruction accuracy and image synthesis quality. Best result. Second best result. $\\dagger$ Requires 3D points from SfM. Best viewed in color.", + "bbox": [ + 75, + 577, + 893, + 606 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the smoothness of the surface, sacrificing details. Our setup $(\\mathrm{NG} + \\mathrm{P})$ produces both smooth surfaces and fine details.", + "bbox": [ + 75, + 632, + 468, + 662 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Tanks and Temples", + "text_level": 1, + "bbox": [ + 76, + 676, + 261, + 694 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As no public result is available for Tanks and Temples, we train NeuS [36] and NeuralWarp [3] following our setup. We also report classical multi-view stereo results using COLMAP [27]. As COLMAP and NeuralWarp do not support view synthesis, we only report PSNR from NeuS. Results are summarized in Fig. 5 and Table 2.", + "bbox": [ + 75, + 702, + 470, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Neuralangelo achieves the highest average PSNR and performs best in terms of F1 score. Comparing against NeuS [36], we can recover high-fidelity surfaces with intricate details. We find that the dense surfaces generated from COLMAP are sensitive to outliers in the sparse point cloud. We also find that NeuralWarp often predicts surfaces for the sky and backgrounds potentially due to their color rendering", + "bbox": [ + 75, + 795, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "scheme following VolSDF [41]. The additional surfaces predicted for backgrounds are counted as outliers and worsen F1 scores significantly. We instead follow NeuS [36] and use an additional network [46] to model the background.", + "bbox": [ + 496, + 632, + 893, + 691 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Similar to the DTU results, using the analytical gradient produces noisy surfaces and thus leads to a low F1 score. We further note that the reconstruction of Courthouse shown in Figs. 1 and 5 are the same building of different sides, demonstrating the capability of Neuralangelo for large-scale granular reconstruction.", + "bbox": [ + 496, + 694, + 893, + 785 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Level of Details", + "text_level": 1, + "bbox": [ + 500, + 799, + 653, + 814 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As Neuralangelo progressively optimizes the hash features of increasing resolution, we inspect the progressive level of details similar to NGLOD [33]. We show a qualitative visualization in Fig. 6. While some surfaces are entirely missed by coarse levels, for example, the tree, table, and", + "bbox": [ + 496, + 825, + 893, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "8461", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/67c4bce5337f42302fa3cd96147fca72e45ebcc21b7b0a5158358f407eadd2bc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 93, + 359, + 233 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e38973828cda37f01f6bc7ee658681a5d63e56e5039c168136126a2992c6c58f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 367, + 93, + 620, + 233 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a10bc5bbf58fb5d5eba7c874bbeace98b88df4b599161990a6156dee0f29ba12.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 630, + 93, + 883, + 233 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2da33cdffd118fcfb89d09e0be0e270fc28d477ba949b1338aa165010089f9f7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 241, + 357, + 364 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a18f9fc0ba2dbcc2450c94cd56489696111aeb083b95c806dfd8af9f0511eb07.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 367, + 239, + 620, + 363 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/69a68951d6161edd18924e59ec2a63ee06b8dc50ae51c752ae55e019c49d2381.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 630, + 239, + 883, + 363 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6ecb24f0630b1554cbead62a326a6d539c633717e5bda433db8f5f1d05179f28.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 375, + 357, + 494 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1615ec4c42d2595472e0ed0a2981383050c0d11bab0bc47f8caf6c5dcea8df2b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 367, + 369, + 620, + 494 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f8f2d5e4bde5788e86c434f140b05612665682cb61f2e7bae950d8ce46a89230.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 630, + 369, + 883, + 494 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ea8b9fda6fa086f86c622b4ed0462512b09b44dbe8e85d9e763c817a4dd691d1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 501, + 357, + 625 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0b4f59f39ea594ca8f0ee72166f7e166d053ecb8e17fbbb3c8984210fcf960cf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 367, + 501, + 620, + 625 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/eed056bf9dd56dfb4fec284ee73dfc5f353fa5053cdcd6056f3a5673b3aa903d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 630, + 501, + 883, + 625 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1f6d66d561729430cea1aa243bfca2a77ad690457872229c5576df9e8c24456c.jpg", + "image_caption": [ + "Figure 5. Qualitative comparison on Tanks and Temples dataset [15]. Neuralangelo captures the scene details better compared to other baseline approaches, while baseline approaches have missing or noisy surfaces." + ], + "image_footnote": [], + "bbox": [ + 83, + 637, + 357, + 757 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/502eef194b30ad0652e8292c05c570b693fedeff00956eb8b0c053235db66b76.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 367, + 633, + 620, + 756 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/eac1f6398a4dffb4cea3d41125fedf42e597e6186b6a5c204b0cd5da1f9cafb5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 630, + 633, + 883, + 756 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "bike rack, these structures are recovered by finer resolutions successfully. The ability to recover missing surfaces demonstrates the advantages of our spatial hierarchy-free design.", + "bbox": [ + 75, + 824, + 470, + 869 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Moreover, we note that flat surfaces are predicted at sufficiently high resolutions (around Level 8 in this example).", + "bbox": [ + 76, + 869, + 472, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Thus, only relying on the continuity of local cells of coarse resolutions is not sufficient to reconstruct large continuous surfaces. The result motivates the use of the numerical gradients for the higher-order derivatives, such that backpropagation is beyond local grid cells.", + "bbox": [ + 496, + 824, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "8462", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8320594a831f27d587fc460e0fc72f4ee0071a01f83b88170f9e247dce384a7a.jpg", + "image_caption": [ + "Level 4, $V_{4} = 74$" + ], + "image_footnote": [], + "bbox": [ + 81, + 103, + 281, + 172 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f47dac0ea1fa0f539c5cd69c83ebf67eea011e9f2803f2efd54a3b0afec40248.jpg", + "image_caption": [ + "Level $8,V_{8} = 223$", + "Figure 6. Results at different hash resolutions. While some structures, such as the tree, table, and bike rack, are missed at coarse resolutions (Level 4). Finer resolutions can progressively recover these missing surfaces. Flat continuous surfaces also require sufficiently fine resolutions to predict (Level 8). The result motivates the non-local updates when using numerical gradients for higher-order derivatives." + ], + "image_footnote": [], + "bbox": [ + 282, + 102, + 483, + 172 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5cbfcbb7ecfd4bd072bd1ef91b0286740808c7fce0d3f2f23a7e9a93865624b5.jpg", + "image_caption": [ + "Level 12, $V_{12} = 676$" + ], + "image_footnote": [], + "bbox": [ + 486, + 102, + 686, + 172 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/236977ccbd1808e95fa4b1500a10c0c4ef0ede1af924d5764cff7ed12cceed40.jpg", + "image_caption": [ + "Level 16, $V_{16} = 2048$" + ], + "image_footnote": [], + "bbox": [ + 687, + 102, + 888, + 172 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/bacea7a5e0d366188a43b65067457396e1766d7d9793a6c1ef7c721424c13599.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
F1 Score ↑PSNR ↑
NeuralWarp [3]COLMAP [27]NeuS [36]AGAG+PNGNG+P (Ours)NeuS [36]AGAG+PNGNG+P (Ours)
Barn0.220.550.290.220.310.630.7026.3626.9126.6926.1428.57
Caterpillar0.180.010.290.230.240.300.3625.2126.0425.1226.1627.81
Courthouse0.080.110.170.080.090.240.2823.5525.4325.6325.0627.23
Ignatius0.020.220.830.720.730.850.8923.2722.6922.7323.7823.67
Meetingroom0.080.190.240.040.050.270.3225.3828.1328.0527.4430.70
Truck0.350.190.450.330.370.440.4823.7123.8923.9522.9925.43
Mean0.150.210.380.270.300.450.5024.5825.5125.3625.2627.24
", + "bbox": [ + 76, + 258, + 890, + 398 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 2. Quantitative results on Tanks and Temples dataset [15]. Neuralangelo achieves the best surface reconstruction quality and performs best on average in terms of image synthesis. Best result. Second best result. Best viewed in color.", + "bbox": [ + 75, + 409, + 892, + 436 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1fc4cc79b98d1130719e1ed5f4239b95ccf98c402839a1538b9ee80163530954.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 81, + 465, + 205, + 561 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a138717460f089da41ba1890ec75547940740bdec19e65976076f28dc647f56b.jpg", + "image_caption": [ + "w/o $\\mathcal{L}_{\\mathrm{curv}}$" + ], + "image_footnote": [], + "bbox": [ + 210, + 465, + 336, + 561 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1b25b09ec3140664a0b0fadb43241a8209c40759ade3fb2ef412caa0a23ac69c.jpg", + "image_caption": [ + "w/ $\\mathcal{L}_{\\mathrm{curv}}$" + ], + "image_footnote": [], + "bbox": [ + 341, + 465, + 465, + 561 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0892a03f597996cb264649a124d1a9d95065c6189c580a0d1d2ace4e3682268a.jpg", + "image_caption": [ + "(b)", + "Figure 7. Ablation results. (a) Surface smoothness improves with curvature regularization $\\mathcal{L}_{\\mathrm{curv}}$ . (c) Concave shapes are better formed with topology warmup." + ], + "image_footnote": [], + "bbox": [ + 81, + 574, + 207, + 672 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c3c3eb14135ab9e63ca9576391a2e65d113669f7b93eb2e036d0694d0f2f6bb9.jpg", + "image_caption": [ + "w/o topology warmup" + ], + "image_footnote": [], + "bbox": [ + 210, + 574, + 336, + 672 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fac9df24cadf56f3931a617d79f972db2673500a5cf88fc3ee1dacbd78e3efbd.jpg", + "image_caption": [ + "w/ topology warmup" + ], + "image_footnote": [], + "bbox": [ + 341, + 574, + 465, + 672 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablations", + "text_level": 1, + "bbox": [ + 76, + 776, + 187, + 790 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "**Curvature regularization.** We ablate the necessity of curvature regularization in Neuralangelo and compare the results in Fig. 7(a). Intuitively, $\\mathcal{L}_{\\mathrm{curv}}$ acts as a smoothness prior by minimizing surface curvatures. Without $\\mathcal{L}_{\\mathrm{curv}}$ , we find that the surfaces tend to have undesirable sharp transitions. By using $\\mathcal{L}_{\\mathrm{curv}}$ , the surface noises are removed.", + "bbox": [ + 75, + 809, + 470, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Topology warmup. We follow prior work and initialize the SDF approximately as a sphere [42]. With an initial spherical shape, using $\\mathcal{L}_{\\mathrm{curv}}$ also makes concave shapes difficult to form because $\\mathcal{L}_{\\mathrm{curv}}$ preserves topology by preventing singularities in curvature. Thus, instead of applying $\\mathcal{L}_{\\mathrm{curv}}$ from the beginning of the optimization process, we use a short warmup period that linearly increases the curvature loss strength. We find this strategy particularly helpful for concave regions, as shown in Fig. 7(b).", + "bbox": [ + 496, + 464, + 893, + 601 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 628, + 617, + 643 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduce Neuralangelo, an approach for photogrammetric neural surface reconstruction. The findings of Neuralangelo are simple yet effective: using numerical gradients for higher-order derivatives and a coarse-to-fine optimization strategy. Neuralangelo unlocks the representation power of multi-resolution hash encoding for neural surface reconstruction modeled as SDF. We show that Neuralangelo effectively recovers dense scene structures of both object-centric captures and large-scale indoor/outdoor scenes with extremely high fidelity, enabling detailed large-scale scene reconstruction from RGB videos. Our method currently samples pixels from images randomly without tracking their statistics and errors. Therefore, we use long training iterations to reduce the stochastics and ensure sufficient sampling of details. It is our future work to explore a more efficient sampling strategy to accelerate the training process.", + "bbox": [ + 496, + 659, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8463", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 2", + "[2] Rui Chen, Songfang Han, Jing Xu, and Hao Su. Point-based multi-view stereo network. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1538-1547, 2019. 2", + "[3] François Darmon, Bénédicte Bascle, Jean-Clément Devaux, Pascal Monasse, and Mathieu Aubry. Improving neural implicit surfaces geometry with patch warping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6260–6269, 2022. 2, 3, 6, 8", + "[4] Jeremy S De Bonet and Paul Viola. Poxels: Probabilistic voxelized volume reconstruction. In Proceedings of International Conference on Computer Vision (ICCV), volume 2, 1999. 2", + "[5] Qiancheng Fu, Qingshan Xu, Yew-Soon Ong, and Wenbing Tao. Geo-neus: Geometry-consistent neural implicit surfaces learning for multi-view reconstruction. arXiv preprint arXiv:2205.15848, 2022. 2, 3", + "[6] Yasutaka Furukawa and Jean Ponce. Accurate, dense, and robust multiview stereopsis. IEEE transactions on pattern analysis and machine intelligence, 32(8):1362-1376, 2009. 1, 2", + "[7] Silvano Galliani, Katrin Lasinger, and Konrad Schindler. Massively parallel multiview stereopsis by surface normal diffusion. In Proceedings of the IEEE International Conference on Computer Vision, pages 873-881, 2015. 2", + "[8] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. arXiv preprint arXiv:2002.10099, 2020. 2, 3", + "[9] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. Cambridge university press, 2003. 1", + "[10] Po-Han Huang, Kevin Matzen, Johannes Kopf, Narendra Ahuja, and Jia-Bin Huang. Deepmvs: Learning multi-view stereopsis. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2821-2830, 2018. 2", + "[11] Rasmus Jensen, Anders Dahl, George Vogiatzis, Engil Tola, and Henrik Aanæs. Large scale multi-view stereopsis evaluation. In 2014 IEEE Conference on Computer Vision and Pattern Recognition, pages 406-413. IEEE, 2014. 5, 6", + "[12] Yue Jiang, Dantong Ji, Zhizhong Han, and Matthias Zwicker. Sdfdiff: Differentiable rendering of signed distance fields for 3d shape optimization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1251-1261, 2020. 2", + "[13] Michael Kazhdan, Matthew Bolitho, and Hugues Hoppe. Poisson surface reconstruction. In Proceedings of the fourth Eurographics symposium on Geometry processing, volume 7, 2006. 2", + "[14] Michael Kazhdan and Hugues Hoppe. Screened poisson surface reconstruction. ACM Transactions on Graphics (ToG), 32(3):1-13, 2013. 2" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG), 36(4):1-13, 2017. 5, 7, 8", + "[16] Kiriakos N Kutulakos and Steven M Seitz. A theory of shape by space carving. International journal of computer vision, 38(3):199-218, 2000. 1, 2", + "[17] Aldo Laurentini. The visual hull concept for silhouette-based image understanding. IEEE Transactions on pattern analysis and machine intelligence, 16(2):150-162, 1994. 2", + "[18] Hai Li, Xingrui Yang, Hongjia Zhai, Yuqian Liu, Hujun Bao, and Guofeng Zhang. Vox-surf: Voxel-based implicit surface representation. IEEE Transactions on Visualization and Computer Graphics, 2022. 2", + "[19] Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. Barf: Bundle-adjusting neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5741-5751, 2021. 4", + "[20] Chen-Hsuan Lin, Chaoyang Wang, and Simon Lucey. Sdf-srn: Learning signed distance 3d object reconstruction from static images. Advances in Neural Information Processing Systems, 33:11453-11464, 2020. 2", + "[21] Bruce D Lucas, Takeo Kanade, et al. An iterative image registration technique with an application to stereo vision, volume 81. Vancouver, 1981. 4", + "[22] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European conference on computer vision, pages 405-421. Springer, 2020. 2, 3, 6", + "[23] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. arXiv preprint arXiv:2201.05989, 2022. 2, 3", + "[24] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2020. 2", + "[25] Michael Oechsle, Songyou Peng, and Andreas Geiger. Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5589-5599, 2021. 2", + "[26] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865-5874, 2021. 4", + "[27] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 6, 8", + "[28] Johannes L Schonberger, Enliang Zheng, Jan-Michael Frahm, and Marc Pollefeys. Pixelwise view selection for unstructured multi-view stereo. In European conference on computer vision, pages 501-518. Springer, 2016. 2" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "8464", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Steven M Seitz and Charles R Dyer. Photorealistic scene reconstruction by voxel coloring. International Journal of Computer Vision, 35(2):151-173, 1999. 1, 2", + "[30] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5459-5469, 2022. 2, 3", + "[31] Jiaming Sun, Xi Chen, Qianqian Wang, Zhengqi Li, Hadar Averbuch-Elor, Xiaowei Zhou, and Noah Snively. Neural 3d reconstruction in the wild. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-9, 2022. 3", + "[32] Richard Szeliski. Rapid octree construction from image sequences. CVGIP: Image understanding, 58(1):23-32, 1993. 2", + "[33] Towaki Takikawa, Joey Litalien, Kangxue Yin, Karsten Kreis, Charles Loop, Derek Nowrouzezahrai, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Neural geometric level of detail: Real-time rendering with implicit 3d shapes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11358-11367, 2021. 3, 6", + "[34] Engin Tola, Christoph Strecha, and Pascal Fua. Efficient large-scale multi-view stereo for ultra high-resolution image sets. Machine Vision and Applications, 23(5):903-920, 2012. 1, 2", + "[35] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T Barron, and Pratul P Srinivasan. Ref-nerf: Structured view-dependent appearance for neural radiance fields. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5481-5490. IEEE, 2022. 4", + "[36] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021. 2, 3, 4, 6, 8", + "[37] Yiming Wang, Qin Han, Marc Habermann, Kostas Daniilidis, Christian Theobalt, and Lingjie Liu. Neus2: Fast learning of neural implicit surfaces for multi-view reconstruction. arXiv preprint arXiv:2212.05231, 2022. 2", + "[38] Yiqun Wang, Ivan Skorokhodov, and Peter Wonka. Hf-neus: Improved surface reconstruction using high-frequency details. In Advances in Neural Information Processing Systems. 3, 6", + "[39] Tong Wu, Jiaqi Wang, Xingang Pan, Xudong Xu, Christian Theobalt, Ziwei Liu, and Dahua Lin. Voxurf: Voxel-based efficient and accurate neural surface reconstruction. arXiv preprint arXiv:2208.12697, 2022. 3", + "[40] Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. Mvsnet: Depth inference for unstructured multi-view stereo. In Proceedings of the European conference on computer vision (ECCV), pages 767-783, 2018. 2", + "[41] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems, 34:4805-4815, 2021. 2, 3, 4, 6", + "[42] Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Basri Ronen, and Yaron Lipman. Multiview neural surface reconstruction by disentangling geometry and appearance. Advances in Neural Information Processing Systems, 33:2492-2502, 2020. 2, 4, 8" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[43] Alex Yu, Sara Fridovich-Keil, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. arXiv preprint arXiv:2112.05131, 2021. 2, 3", + "[44] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. arXiv preprint arXiv:2206.00665, 2022. 2, 3", + "[45] Jingyang Zhang, Yao Yao, Shiwei Li, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. Critical regularizations for neural surface reconstruction in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6270-6279, 2022. 2, 3, 6", + "[46] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492, 2020. 2, 6", + "[47] Xiuming Zhang, Pratul P Srinivasan, Boyang Deng, Paul Debevec, William T Freeman, and Jonathan T Barron. Nerfactor: Neural factorization of shape and reflectance under an unknown illumination. ACM Transactions on Graphics (TOG), 40(6):1-18, 2021. 4" + ], + "bbox": [ + 501, + 90, + 893, + 387 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "8465", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Neuralangelo_ High-Fidelity Neural Surface Reconstruction/55f0b247-41a3-45d9-bec0-4490ab23bed9_model.json b/2023/Neuralangelo_ High-Fidelity Neural Surface Reconstruction/55f0b247-41a3-45d9-bec0-4490ab23bed9_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f973fd19a26a108acbab48859f4fa164a2dcd01e --- /dev/null +++ b/2023/Neuralangelo_ High-Fidelity Neural Surface Reconstruction/55f0b247-41a3-45d9-bec0-4490ab23bed9_model.json @@ -0,0 +1,2365 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.182, + 0.131, + 0.788, + 0.154 + ], + "angle": 0, + "content": "Neuralangelo: High-Fidelity Neural Surface Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.179, + 0.854, + 0.261 + ], + "angle": 0, + "content": "Zhaoshuo Li\\(^{1,2}\\) Thomas Müller\\(^{1}\\) Alex Evans\\(^{1}\\) Russell H. Taylor\\(^{2}\\) Mathias Unberath\\(^{2}\\) Ming-Yu Liu\\(^{1}\\) Chen-Hsuan Lin\\(^{1}\\) \n\\(^{1}\\)NVIDIA Research \\(^{2}\\)Johns Hopkins University https://research.nvidia.com/labs/dir/neuralangelo" + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.281, + 0.896, + 0.579 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.587, + 0.893, + 0.618 + ], + "angle": 0, + "content": "Figure 1. We present Neuralangelo, a framework for high-fidelity 3D surface reconstruction from RGB images using neural volume rendering, even without auxiliary data such as segmentation or depth. Shown in the figure is an extracted 3D mesh of a courthouse." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.638, + 0.314, + 0.654 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.667, + 0.474, + 0.894 + ], + "angle": 0, + "content": "Neural surface reconstruction has been shown to be powerful for recovering dense 3D surfaces via image-based neural rendering. However, current methods struggle to recover detailed structures of real-world scenes. To address the issue, we present Neuralangelo, which combines the representation power of multi-resolution 3D hash grids with neural surface rendering. Two key ingredients enable our approach: (1) numerical gradients for computing higher-order derivatives as a smoothing operation and (2) coarse-to-fine optimization on the hash grids controlling different levels of details. Even without auxiliary inputs such as depth, Neuralangelo can effectively recover dense 3D surface structures from multi-view images with fidelity significantly surpassing previous methods, enabling detailed large-scale scene reconstruction from RGB video captures." + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.638, + 0.631, + 0.654 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.669, + 0.896, + 0.82 + ], + "angle": 0, + "content": "3D surface reconstruction aims to recover dense geometric scene structures from multiple images observed at different viewpoints [9]. The recovered surfaces provide structural information useful for many downstream applications, such as 3D asset generation for augmented/virtual/mixed reality or environment mapping for autonomous navigation of robotics. Photogrammetric surface reconstruction using a monocular RGB camera is of particular interest, as it equips users with the capability of casually creating digital twins of the real world using ubiquitous mobile devices." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.897, + 0.903 + ], + "angle": 0, + "content": "Classically, multi-view stereo algorithms [6, 16, 29, 34] had been the method of choice for sparse 3D reconstruction. An inherent drawback of these algorithms, however, is their inability to handle ambiguous observations, e.g. regions with large areas of homogeneous colors, repetitive texture" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8456" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.272 + ], + "angle": 0, + "content": "patterns, or strong color variations. This would result in inaccurate reconstructions with noisy or missing surfaces. Recently, neural surface reconstruction methods [36,41,42] have shown great potential in addressing these limitations. This new class of methods uses coordinate-based multi-layer perceptrons (MLPs) to represent the scene as an implicit function, such as occupancy fields [25] or signed distance functions (SDF) [36,41,42]. Leveraging the inherent continuity of MLPs and neural volume rendering [22], these techniques allow the optimized surfaces to meaningfully interpolate between spatial locations, resulting in smooth and complete surface representations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.274, + 0.473, + 0.469 + ], + "angle": 0, + "content": "Despite the superiority of neural surface reconstruction methods over classical approaches, the recovered fidelity of current methods does not scale well with the capacity of MLPs. Recently, Müller et al. [23] proposed a new scalable representation, referred to as Instant NGP (Neural Graphics Primitives). Instant NGP introduces a hybrid 3D grid structure with a multi-resolution hash encoding and a lightweight MLP that is more expressive with a memory footprint log-linear to the resolution. The proposed hybrid representation greatly increases the representation power of neural fields and has achieved great success at representing very fine-grained details for a wide variety of tasks, such as object shape representation and novel view synthesis problems." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.47, + 0.473, + 0.711 + ], + "angle": 0, + "content": "In this paper, we propose Neuralangelo for high-fidelity surface reconstruction (Fig. 1). Neuralangelo adopts Instant NGP as a neural SDF representation of the underlying 3D scene, optimized from multi-view image observations via neural surface rendering [36]. We present two findings central to fully unlocking the potentials of multi-resolution hash encodings. First, using numerical gradients to compute higher-order derivatives, such as surface normals for the eikonal regularization [8, 12, 20, 42], is critical to stabilizing the optimization. Second, a progressive optimization schedule plays an important role in recovering the structures at different levels of details. We combine these two key ingredients and, via extensive experiments on standard benchmarks and real-world scenes, demonstrate significant improvements over image-based neural surface reconstruction methods in both reconstruction accuracy and view synthesis quality." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.712, + 0.443, + 0.726 + ], + "angle": 0, + "content": "In summary, we present the following contributions:" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.74, + 0.468, + 0.784 + ], + "angle": 0, + "content": "- We present the Neuralangelo framework to naturally incorporate the representation power of multi-resolution hash encoding [23] into neural SDF representations." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.79, + 0.468, + 0.85 + ], + "angle": 0, + "content": "- We present two simple techniques to improve the quality of hash-encoded surface reconstruction: higher-order derivatives with numerical gradients and coarse-to-fine optimization with a progressive level of details." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.856, + 0.47, + 0.901 + ], + "angle": 0, + "content": "- We empirically demonstrate the effectiveness of Neuralangelo on various datasets, showing significant improvements over previous methods." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.74, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.09, + 0.638, + 0.106 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.121, + 0.895, + 0.256 + ], + "angle": 0, + "content": "Multi-view surface reconstruction. Early image-based photogrammetry techniques use a volumetric occupancy grid to represent the scene [4, 16, 17, 29, 32]. Each voxel is visited and marked occupied if strict color constancy between the corresponding projected image pixels is satisfied. The photometric consistency assumption typically fails due to autoexposure or non-Lambertian materials, which are ubiquitous in the real world. Relaxing such color constancy constraints across views is important for realistic 3D reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.257, + 0.895, + 0.393 + ], + "angle": 0, + "content": "Follow-up methods typically start with 3D point clouds from multi-view stereo techniques [6,7,28,34] and then perform dense surface reconstruction [13,14]. Reliance on the quality of the generated point clouds often leads to missing or noisy surfaces. Recent learning-based approaches augment the point cloud generation process with learned image features and cost volume construction [2,10,40]. However, these approaches are inherently limited by the resolution of the cost volume and fail to recover geometric details." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.398, + 0.895, + 0.61 + ], + "angle": 0, + "content": "Neural Radiance Fields (NeRF). NeRF [22] achieves remarkable photorealistic view synthesis with view-dependent effects. NeRF encodes 3D scenes with an MLP mapping 3D spatial locations to color and volume density. These predictions are composited into pixel colors using neural volume rendering. A problem of NeRF and its variants [1,30,43,46], however, is the question of how an isosurface of the volume density could be defined to represent the underlying 3D geometry. Current practice often relies on heuristic thresholding on the density values; due to insufficient constraints on the level sets, however, such surfaces are often noisy and may not model the scene structures accurately [36,41]. Therefore, more direct modeling of surfaces is preferred for photogrammetric surface reconstruction problems." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.615, + 0.894, + 0.75 + ], + "angle": 0, + "content": "Neural surface reconstruction. For scene representations with better-defined 3D surfaces, implicit functions such as occupancy grids [24, 25] or SDFs [42] are preferred over simple volume density fields. To integrate with neural volume rendering [22], different techniques [36, 41] have been proposed to reparametrize the underlying representations back to volume density. These designs of neural implicit functions enable more accurate surface prediction with view synthesis capabilities of unsacrificed quality [42]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Follow-up works extend the above approaches to real-time at the cost of surface fidelity [18, 37], while others [3,5,44] use auxiliary information to enhance the reconstruction results. Notably, NeuralWarp [3] uses patch warping given co-visibility information from structure-from-motion (SfM) to guide surface optimization, but the patchwise planar assumption fails to capture highly-varying surfaces [3]. Other methods [5,45] utilize sparse point clouds from SfM to supervise the SDF, but their performances are upper-bounded by the quality of the point clouds, as with" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8457" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.273 + ], + "angle": 0, + "content": "classical approaches [45]. The use of monocular depth and segmentation as auxiliary data has also been explored with unconstrained image collections [31] or using scene representations with hash encodings [44]. In contrast, our work Neuralangelo builds upon hash encodings [23] to recover surfaces but without the need for auxiliary inputs used in prior work [3, 5, 31, 44, 45]. Concurrent work [38] also proposes coarse-to-fine optimization for improved surface details, where a displacement network corrects the shape predicted by a coarse network. In contrast, we use hierarchical hash grids and control the level of details based on our analysis of higher-order derivatives." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.286, + 0.186, + 0.303 + ], + "angle": 0, + "content": "3. Approach" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.311, + 0.47, + 0.402 + ], + "angle": 0, + "content": "Neuralangelo reconstructs dense structures of the scene from multi-view images. Neuralangelo samples 3D locations along camera view directions and uses a multi-resolution hash encoding to encode the positions. The encoded features are input to an SDF MLP and a color MLP to composite images using SDF-based volume rendering." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.411, + 0.218, + 0.425 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.434, + 0.47, + 0.555 + ], + "angle": 0, + "content": "Neural volume rendering. NeRF [22] represents a 3D scene as volume density and color fields. Given a posed camera and a ray direction, the volume rendering scheme integrates the color radiance of sampled points along the ray. The \\(i\\)-th sampled 3D position \\(\\mathbf{x}_i\\) is at a distance \\(t_i\\) from the camera center. The volume density \\(\\sigma_i\\) and color \\(\\mathbf{c}_i\\) of each sampled point are predicted using a coordinate MLP. The rendered color of a given pixel is approximated as the Riemann sum:" + }, + { + "type": "equation", + "bbox": [ + 0.136, + 0.565, + 0.47, + 0.607 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {c}} (\\mathbf {o}, \\mathbf {d}) = \\sum_ {i = 1} ^ {N} w _ {i} \\mathbf {c} _ {i}, \\quad \\text {w h e r e} w _ {i} = T _ {i} \\alpha_ {i}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.616, + 0.473, + 0.707 + ], + "angle": 0, + "content": "Here, \\(\\alpha_{i} = 1 - \\exp (-\\sigma_{i}\\delta_{i})\\) is the opacity of the \\(i\\)-th ray segment, \\(\\delta_{i} = t_{i + 1} - t_{i}\\) is the distance between adjacent samples, and \\(T_{i} = \\Pi_{j = 1}^{i - 1}(1 - \\alpha_{j})\\) is the accumulated transmittance, indicating the fraction of light that reaches the camera. To supervise the network, a color loss is used between input images \\(\\mathbf{c}\\) and rendered images \\(\\hat{\\mathbf{c}}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.208, + 0.718, + 0.47, + 0.735 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {R G B}} = \\left\\| \\hat {\\mathbf {c}} - \\mathbf {c} \\right\\| _ {1}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.745, + 0.472, + 0.79 + ], + "angle": 0, + "content": "However, surfaces are not clearly defined using such density formulation. Extracting surfaces from density-based representation often leads to noisy and unrealistic results [36, 41]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.472, + 0.901 + ], + "angle": 0, + "content": "Volume rendering of SDF. One of the most common surface representations is SDF. The surface \\(S\\) of an SDF can be implicitly represented by its zero-level set, i.e., \\(S = \\{\\mathbf{x} \\in \\mathbb{R}^3 | f(\\mathbf{x}) = 0\\}\\), where \\(f(\\mathbf{x})\\) is the SDF value. In the context of neural SDFs, Wang et al. [36] proposed to convert volume density predictions in NeRF to SDF representations with a logistic function to allow optimization with neural volume" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "rendering. Given a 3D point \\(\\mathbf{x}_i\\) and SDF value \\(f(\\mathbf{x}_i)\\), the corresponding opacity value \\(\\alpha_i\\) used in Eq. 1 is computed as" + }, + { + "type": "equation", + "bbox": [ + 0.545, + 0.132, + 0.892, + 0.166 + ], + "angle": 0, + "content": "\\[\n\\alpha_ {i} = \\max \\left(\\frac {\\Phi_ {s} (f (\\mathbf {x} _ {i})) - \\Phi_ {s} (f (\\mathbf {x} _ {i + 1}))}{\\Phi_ {s} (f (\\mathbf {x} _ {i}))}, 0\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.168, + 0.892, + 0.199 + ], + "angle": 0, + "content": "where \\(\\Phi_s\\) is the sigmoid function. In this work, we use the same SDF-based volume rendering formulation [36]." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.204, + 0.894, + 0.294 + ], + "angle": 0, + "content": "Multi-resolution hash encoding. Recently, multi-resolution hash encoding proposed by Müller et al. [23] has shown great scalability for neural scene representations, generating fine-grained details for tasks such as novel view synthesis. In Neuralangelo, we adopt the representation power of hash encoding to recover high-fidelity surfaces." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.295, + 0.894, + 0.446 + ], + "angle": 0, + "content": "The hash encoding uses multi-resolution grids, with each grid cell corner mapped to a hash entry. Each hash entry stores the encoding feature. Let \\(\\{V_1,\\dots,V_L\\}\\) be the set of different spatial grid resolutions. Given an input position \\(\\mathbf{x}_i\\), we map it to the corresponding position at each grid resolution \\(V_{l}\\) as \\(\\mathbf{x}_{i,l} = \\mathbf{x}_i\\cdot V_l\\). The feature vector \\(\\gamma_l(\\mathbf{x}_{i,l})\\in \\mathbb{R}^c\\) given resolution \\(V_{l}\\) is obtained via trilinear interpolation of hash entries at the grid cell corners. The encoding features across all spatial resolutions are concatenated together, forming a \\(\\gamma (\\mathbf{x}_i)\\in \\mathbb{R}^{cL}\\) feature vector:" + }, + { + "type": "equation", + "bbox": [ + 0.583, + 0.453, + 0.892, + 0.471 + ], + "angle": 0, + "content": "\\[\n\\gamma \\left(\\mathbf {x} _ {i}\\right) = \\left(\\gamma_ {1} \\left(\\mathbf {x} _ {i, 1}\\right), \\dots , \\gamma_ {L} \\left(\\mathbf {x} _ {i, L}\\right)\\right). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.478, + 0.871, + 0.492 + ], + "angle": 0, + "content": "The encoded features are then passed to a shallow MLP." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.493, + 0.892, + 0.644 + ], + "angle": 0, + "content": "One alternative to hash encoding is sparse voxel structures [30, 33, 39, 43], where each grid corner is uniquely defined without collision. However, volumetric feature grids require hierarchical spatial decomposition (e.g. octrees) to make the parameter count tractable; otherwise, the memory would grow cubically with spatial resolution. Given such hierarchy, finer voxel resolutions by design cannot recover surfaces that are misrepresented by the coarser resolutions [33]. Hash encoding instead assumes no spatial hierarchy and resolves collision automatically based on gradient averaging [23]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.651, + 0.798, + 0.667 + ], + "angle": 0, + "content": "3.2. Numerical Gradient Computation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.894, + 0.764 + ], + "angle": 0, + "content": "We show in this section that the analytical gradient w.r.t. position of hash encoding suffers from localities. Therefore, optimization updates only propagate to local hash grids, lacking non-local smoothness. We propose a simple fix to such a locality problem by using numerical gradients. An overview is shown in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.894, + 0.855 + ], + "angle": 0, + "content": "A special property of SDF is its differentiability with a gradient of the unit norm. The gradient of SDF satisfies the eikonal equation \\(\\| \\nabla f(\\mathbf{x})\\| _2 = 1\\) (almost everywhere). To enforce the optimized neural representation to be a valid SDF, the eikonal loss [8] is typically imposed on the SDF predictions:" + }, + { + "type": "equation", + "bbox": [ + 0.583, + 0.863, + 0.892, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {e i k}} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left(\\| \\nabla f (\\mathbf {x} _ {i}) \\| _ {2} - 1\\right) ^ {2}, \\tag {5}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8458" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.089, + 0.09, + 0.468, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.221, + 0.471, + 0.264 + ], + "angle": 0, + "content": "Figure 2. Using numerical gradients for higher-order derivatives distributes the back-propagation updates beyond the local hash grid cell, thus becoming a smoothed version of analytical gradients." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.29, + 0.471, + 0.336 + ], + "angle": 0, + "content": "where \\(N\\) is the total number of sampled points. To allow for end-to-end optimization, a double backward operation on the SDF prediction \\(f(\\mathbf{x})\\) is required." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.335, + 0.473, + 0.457 + ], + "angle": 0, + "content": "The de facto method for computing surface normals of SDFs \\(\\nabla f(\\mathbf{x})\\) is to use analytical gradients [36,41,42]. Analytical gradients of hash encoding w.r.t. position, however, are not continuous across space under trilinear interpolation. To find the sampling location in a voxel grid, each 3D point \\(\\mathbf{x}_i\\) would first be scaled by the grid resolution \\(V_{l}\\), written as \\(\\mathbf{x}_{i,l} = \\mathbf{x}_i\\cdot V_l\\). Let the coefficient for (tri-)linear interpolation be \\(\\beta = \\mathbf{x}_{i,l} - \\left\\lfloor \\mathbf{x}_{i,l}\\right\\rfloor\\). The resulting feature vectors are" + }, + { + "type": "equation", + "bbox": [ + 0.116, + 0.467, + 0.471, + 0.485 + ], + "angle": 0, + "content": "\\[\n\\gamma_ {l} \\left(\\mathbf {x} _ {i, l}\\right) = \\gamma_ {l} \\left(\\left\\lfloor \\mathbf {x} _ {i, l} \\right\\rfloor\\right) \\cdot (1 - \\beta) + \\gamma_ {l} \\left(\\left\\lceil \\mathbf {x} _ {i, l} \\right\\rceil\\right) \\cdot \\beta , \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.495, + 0.47, + 0.556 + ], + "angle": 0, + "content": "where the rounded position \\(\\lfloor \\mathbf{x}_{i,l}\\rfloor ,\\lceil \\mathbf{x}_{i,l}\\rceil\\) correspond to the local grid cell corners. We note that rounding operations \\(\\lfloor \\cdot \\rfloor\\) and \\(\\lceil \\cdot \\rceil\\) are non-differentiable. As a result, the derivative of hash encoding w.r.t. the position can be obtained as" + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.565, + 0.47, + 0.62 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\frac {\\partial \\gamma_ {l} (\\mathbf {x} _ {i , l})}{\\partial \\mathbf {x} _ {i}} = \\gamma_ {l} (\\lfloor \\mathbf {x} _ {i, l} \\rfloor) \\cdot (- \\frac {\\partial \\beta}{\\partial \\mathbf {x} _ {i , l}}) + \\gamma_ {l} (\\lceil \\mathbf {x} _ {i, l} \\rceil) \\cdot \\frac {\\partial \\beta}{\\partial \\mathbf {x} _ {i , l}} \\\\ = \\gamma_ {l} (\\left\\lfloor \\mathbf {x} _ {i, l} \\right\\rfloor) \\cdot (- V _ {l}) + \\gamma_ {l} (\\left\\lceil \\mathbf {x} _ {i, l} \\right\\rceil) \\cdot V _ {l}. \\tag {7} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.47, + 0.81 + ], + "angle": 0, + "content": "The derivative of hash encoding is local, i.e., when \\(\\mathbf{x}_i\\) moves across grid cell borders, the corresponding hash entries will be different. Therefore, the eikonal loss defined in Eq. 5 only back-propagates to the locally sampled hash entries, i.e. \\(\\gamma_l(\\lfloor \\mathbf{x}_{i,l}\\rfloor)\\) and \\(\\gamma_l(\\lceil \\mathbf{x}_{i,l}\\rceil)\\). When continuous surfaces (e.g. a flat wall) span multiple grid cells, these grid cells should produce coherent surface normals without sudden transitions. To ensure consistency in surface representation, joint optimization of these grid cells is desirable. However, the analytical gradient is limited to local grid cells, unless all corresponding grid cells happen to be sampled and optimized simultaneously. Such sampling is not always guaranteed." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.471, + 0.902 + ], + "angle": 0, + "content": "To overcome the locality of the analytical gradient of hash encoding, we propose to compute the surface normals using numerical gradients. If the step size of the numerical gradient is smaller than the grid size of hash encoding, the numerical gradient would be equivalent to the analytical gradient; otherwise, hash entries of multiple grid cells" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.287 + ], + "angle": 0, + "content": "would participate in the surface normal computation. Backpropagating through the surface normals thus allows hash entries of multiple grids to receive optimization updates simultaneously. Intuitively, numerical gradients with carefully chosen step sizes can be interpreted as a smoothing operation on the analytical gradient expression. An alternative of normal supervision is a teacher-student curriculum [35, 47], where the predicted noisy normals are driven towards MLP outputs to exploit the smoothness of MLPs. However, analytical gradients from such teacher-student losses still only back-propagate to local grid cells for hash encoding. In contrast, numerical gradients solve the locality issue without the need of additional networks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.288, + 0.895, + 0.378 + ], + "angle": 0, + "content": "To compute the surface normals using the numerical gradient, additional SDF samples are needed. Given a sampled point \\(\\mathbf{x}_i = (x_i,y_i,z_i)\\), we additionally sample two points along each axis of the canonical coordinate around \\(x_{i}\\) within a vicinity of a step size of \\(\\epsilon\\). For example, the \\(x\\)-component of the surface normal can be found as" + }, + { + "type": "equation", + "bbox": [ + 0.537, + 0.383, + 0.893, + 0.415 + ], + "angle": 0, + "content": "\\[\n\\nabla_ {x} f (\\mathbf {x} _ {i}) = \\frac {f \\left(\\gamma \\left(\\mathbf {x} _ {i} + \\boldsymbol {\\epsilon} _ {x}\\right)\\right) - f \\left(\\gamma \\left(\\mathbf {x} _ {i} - \\boldsymbol {\\epsilon} _ {x}\\right)\\right)}{2 \\epsilon}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.421, + 0.892, + 0.453 + ], + "angle": 0, + "content": "where \\(\\epsilon_{x} = [\\epsilon, 0, 0]\\). In total, six additional SDF samples are required for numerical surface normal computation." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.46, + 0.756, + 0.476 + ], + "angle": 0, + "content": "3.3. Progressive Levels of Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.483, + 0.895, + 0.62 + ], + "angle": 0, + "content": "Coarse-to-fine optimization can better shape the loss landscape to avoid falling into false local minima. Such a strategy has found many applications in computer vision, such as image-based registration [19,21,26]. Neuralangelo also adopts a coarse-to-fine optimization scheme to reconstruct the surfaces with progressive levels of details. Using numerical gradients for the higher-order derivatives naturally enables Neuralangelo to perform coarse-to-fine optimization from two perspectives." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.624, + 0.895, + 0.791 + ], + "angle": 0, + "content": "Step size \\(\\epsilon\\). As previously discussed, numerical gradients can be interpreted as a smoothing operation where the step size \\(\\epsilon\\) controls the resolution and the amount of recovered details. Imposing \\(\\mathcal{L}_{\\mathrm{eik}}\\) with a larger \\(\\epsilon\\) for numerical surface normal computation ensures the surface normal is consistent at a larger scale, thus producing consistent and continuous surfaces. On the other hand, imposing \\(\\mathcal{L}_{\\mathrm{eik}}\\) with a smaller \\(\\epsilon\\) affects a smaller region and avoids smoothing details. In practice, we initialize the step size \\(\\epsilon\\) to the coarsest hash grid size and exponentially decrease it matching different hash grid sizes throughout the optimization process." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Hash grid resolution \\( V \\). If all hash grids are activated from the start of the optimization, to capture geometric details, fine hash grids must first \"unlearn\" from the coarse optimization with large step size \\( \\epsilon \\) and \"relearn\" with a smaller \\( \\epsilon \\). If such a process is unsuccessful due to converged optimization, geometric details would be lost. Therefore, we only enable an initial set of coarse hash grids and progressively activate" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8459" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.089, + 0.895, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.364, + 0.877, + 0.38 + ], + "angle": 0, + "content": "Figure 3. Qualitative comparison on the DTU benchmark [11]. Neuralangelo produces more accurate and higher-fidelity surfaces." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.406, + 0.471, + 0.482 + ], + "angle": 0, + "content": "finer hash grids throughout optimization when \\(\\epsilon\\) decreases to their spatial size. The relearning process can thus be avoided to better capture the details. In practice, we also apply weight decay over all parameters to avoid single-resolution features dominating the final results." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.489, + 0.216, + 0.505 + ], + "angle": 0, + "content": "3.4. Optimization" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.513, + 0.471, + 0.618 + ], + "angle": 0, + "content": "To further encourage the smoothness of the reconstructed surfaces, we impose a prior by regularizing the mean curvature of SDF. The mean curvature is computed from discrete Laplacian similar to the surface normal computation, otherwise, the second-order analytical gradients of hash encoding are zero everywhere when using trilinear interpolation. The curvature loss \\(\\mathcal{L}_{\\mathrm{curv}}\\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.625, + 0.47, + 0.667 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {c u r v}} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left| \\nabla^ {2} f (\\mathbf {x} _ {i}) \\right|. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.674, + 0.471, + 0.704 + ], + "angle": 0, + "content": "We note that the samples used for the surface normal computation in Eq. 8 are sufficient for curvature computation." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.705, + 0.458, + 0.719 + ], + "angle": 0, + "content": "The total loss is defined as the weighted sum of losses:" + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.728, + 0.47, + 0.745 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\mathrm {R G B}} + w _ {\\mathrm {e i k}} \\mathcal {L} _ {\\mathrm {e i k}} + w _ {\\mathrm {c u r v}} \\mathcal {L} _ {\\mathrm {c u r v}}. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.753, + 0.471, + 0.784 + ], + "angle": 0, + "content": "All network parameters, including MLPs and hash encoding, are trained jointly end-to-end." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.795, + 0.21, + 0.813 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.825, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Datasets. Following prior work, we conduct experiments on 15 object-centric scenes of the DTU dataset [11]. Each scene has 49 or 64 images captured by a robot-held monocular RGB camera. The ground truth is obtained from a structured-light scanner. We further conduct experiments on 6 scenes" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.406, + 0.895, + 0.467 + ], + "angle": 0, + "content": "of the Tanks and Temples dataset [15], including large-scale indoor/outdoor scenes. Each scene contains 263 to 1107 images captured using a hand-held monocular RGB camera. The ground truth is obtained using a LiDAR sensor." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.471, + 0.893, + 0.622 + ], + "angle": 0, + "content": "Implementation details. Our hash encoding resolution spans \\(2^{5}\\) to \\(2^{11}\\) with 16 levels. Each hash entry has a channel size of 8. The maximum number of hash entries of each resolution is \\(2^{22}\\). We activate 4 and 8 hash resolutions at the beginning of optimization for DTU dataset and Tanks and Temples respectively, due to differences in scene scales. We enable a new hash resolution every 5000 iterations when the step size \\(\\epsilon\\) equals its grid cell size. For all experiments, we do not utilize auxiliary data such as segmentation or depth during the optimization process." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.627, + 0.894, + 0.673 + ], + "angle": 0, + "content": "Evaluation criteria. We report Chamfer distance and F1 score for surface evaluation [11, 15]. We use peak signal-to-noise ratio (PSNR) to report image synthesis qualities." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.682, + 0.669, + 0.696 + ], + "angle": 0, + "content": "4.1. DTU Benchmark" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.894, + 0.81 + ], + "angle": 0, + "content": "We show qualitative results in Fig. 3 and quantitative results in Table 1. On average, Neuralangelo achieves the lowest Chamfer distance and the highest PSNR, even without using auxiliary inputs. The result suggests that Neuralangelo is more generally applicable than prior work when recovering surfaces and synthesizing images, despite not performing best in every individual scene." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.894, + 0.901 + ], + "angle": 0, + "content": "We further ablate Neuralangelo against the following conditions: 1) AG: analytical gradients, 2) \\(\\mathrm{AG + P}\\): analytical gradients and progressive activating hash resolutions, 3) NG: numerical gradients with varying \\(\\epsilon\\). Fig. 4 shows the results qualitatively. AG produces noisy surfaces, even with hash resolutions progressively activated (AG+P). NG improves" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8460" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.09, + 0.239, + 0.225 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.145, + 0.228, + 0.176, + 0.24 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.244, + 0.091, + 0.402, + 0.225 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.312, + 0.228, + 0.333, + 0.238 + ], + "angle": 0, + "content": "AG" + }, + { + "type": "image", + "bbox": [ + 0.404, + 0.091, + 0.563, + 0.225 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.466, + 0.228, + 0.503, + 0.238 + ], + "angle": 0, + "content": "AG+P" + }, + { + "type": "image", + "bbox": [ + 0.565, + 0.091, + 0.724, + 0.225 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.636, + 0.228, + 0.657, + 0.238 + ], + "angle": 0, + "content": "NG" + }, + { + "type": "image", + "bbox": [ + 0.727, + 0.091, + 0.887, + 0.225 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.771, + 0.228, + 0.846, + 0.239 + ], + "angle": 0, + "content": "NG+P (Ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.25, + 0.895, + 0.293 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparison of different coarse-to-fine optimization scheme. When using the analytical gradient (AG and \\(\\mathrm{AG + P}\\)), coarse surfaces often contain artifacts. While using numerical gradients (NG) leads to a better coarse shape, details are also smoothed. Our solution \\((\\mathrm{NG} + \\mathrm{P})\\) produces both smooth surfaces and fine details." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.307, + 0.896, + 0.568 + ], + "angle": 0, + "content": "
243740556365698397105106110114118122Mean
NeRF [22]1.901.601.850.582.281.271.471.672.051.070.882.531.061.150.961.49
VolSDF [41]1.141.260.810.491.250.700.721.291.180.700.661.080.420.610.550.86
NeuS [36]1.001.370.930.431.100.650.571.481.090.830.521.200.350.490.540.84
HF-NeuS [38]0.761.320.700.391.060.630.631.151.120.800.521.220.330.490.500.77
RegSDF [45] †0.601.410.640.431.340.620.600.900.921.020.600.590.300.410.390.72
NeuralWarp [3] †0.490.710.380.380.790.810.821.201.060.680.660.740.410.630.510.68
AG0.671.040.840.391.431.231.111.241.540.850.501.010.370.510.440.88
AG+P0.590.950.460.341.190.700.791.191.370.690.490.930.330.440.440.73
NG0.480.810.430.350.890.710.611.261.060.740.470.790.330.450.430.65
NG+P (Ours)0.370.720.350.350.870.540.531.290.970.730.470.740.320.410.430.61
RegSDF [45] †24.7823.0623.4722.2128.5725.5321.8128.8926.8127.9124.7125.1326.8421.6728.2525.31
NeuS [36]26.6223.6426.4325.5930.6132.8329.2433.7126.8531.9732.1828.9228.4135.0034.8129.79
VolSDF [41]26.2825.6126.5526.7631.5731.5029.3833.2328.0332.1333.1631.4930.3334.9034.7530.38
NeRF [22]26.2425.7426.7927.5731.9631.5029.5832.7828.3532.0833.4931.5431.0035.5935.5130.65
AG29.9724.9823.1130.2730.6031.2729.2734.2227.4733.0933.8529.9829.4135.6935.1130.55
AG+P30.1224.6329.5930.2931.6032.0429.8534.1927.8233.2333.9529.1529.4435.9935.6731.17
NG30.3425.1430.2030.7931.7231.8629.8134.3628.0133.4534.3830.3929.8836.0235.7431.47
NG+P (Ours)30.6427.7832.7034.1835.1535.8931.4736.8230.1335.9236.6132.6031.2038.4138.0533.84
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.578, + 0.895, + 0.607 + ], + "angle": 0, + "content": "Table 1. Quantitative results on DTU dataset [11]. Neuralangelo achieves the best reconstruction accuracy and image synthesis quality. Best result. Second best result. \\(\\dagger\\) Requires 3D points from SfM. Best viewed in color." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.633, + 0.47, + 0.664 + ], + "angle": 0, + "content": "the smoothness of the surface, sacrificing details. Our setup \\((\\mathrm{NG} + \\mathrm{P})\\) produces both smooth surfaces and fine details." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.678, + 0.262, + 0.695 + ], + "angle": 0, + "content": "4.2. Tanks and Temples" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.703, + 0.471, + 0.794 + ], + "angle": 0, + "content": "As no public result is available for Tanks and Temples, we train NeuS [36] and NeuralWarp [3] following our setup. We also report classical multi-view stereo results using COLMAP [27]. As COLMAP and NeuralWarp do not support view synthesis, we only report PSNR from NeuS. Results are summarized in Fig. 5 and Table 2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Neuralangelo achieves the highest average PSNR and performs best in terms of F1 score. Comparing against NeuS [36], we can recover high-fidelity surfaces with intricate details. We find that the dense surfaces generated from COLMAP are sensitive to outliers in the sparse point cloud. We also find that NeuralWarp often predicts surfaces for the sky and backgrounds potentially due to their color rendering" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.633, + 0.894, + 0.693 + ], + "angle": 0, + "content": "scheme following VolSDF [41]. The additional surfaces predicted for backgrounds are counted as outliers and worsen F1 scores significantly. We instead follow NeuS [36] and use an additional network [46] to model the background." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.695, + 0.895, + 0.786 + ], + "angle": 0, + "content": "Similar to the DTU results, using the analytical gradient produces noisy surfaces and thus leads to a low F1 score. We further note that the reconstruction of Courthouse shown in Figs. 1 and 5 are the same building of different sides, demonstrating the capability of Neuralangelo for large-scale granular reconstruction." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.8, + 0.655, + 0.815 + ], + "angle": 0, + "content": "4.3. Level of Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.895, + 0.902 + ], + "angle": 0, + "content": "As Neuralangelo progressively optimizes the hash features of increasing resolution, we inspect the progressive level of details similar to NGLOD [33]. We show a qualitative visualization in Fig. 6. While some surfaces are entirely missed by coarse levels, for example, the tree, table, and" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "8461" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.083, + 0.094, + 0.36, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.369, + 0.094, + 0.622, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.631, + 0.094, + 0.885, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.242, + 0.359, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.369, + 0.24, + 0.622, + 0.364 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.631, + 0.24, + 0.885, + 0.364 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.375, + 0.359, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.369, + 0.371, + 0.622, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.631, + 0.371, + 0.885, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.502, + 0.359, + 0.626 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.369, + 0.502, + 0.622, + 0.626 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.631, + 0.502, + 0.885, + 0.626 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.638, + 0.359, + 0.758 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.369, + 0.635, + 0.622, + 0.757 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.631, + 0.634, + 0.885, + 0.757 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.775, + 0.895, + 0.804 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparison on Tanks and Temples dataset [15]. Neuralangelo captures the scene details better compared to other baseline approaches, while baseline approaches have missing or noisy surfaces." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.471, + 0.871 + ], + "angle": 0, + "content": "bike rack, these structures are recovered by finer resolutions successfully. The ability to recover missing surfaces demonstrates the advantages of our spatial hierarchy-free design." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Moreover, we note that flat surfaces are predicted at sufficiently high resolutions (around Level 8 in this example)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Thus, only relying on the continuity of local cells of coarse resolutions is not sufficient to reconstruct large continuous surfaces. The result motivates the use of the numerical gradients for the higher-order derivatives, such that backpropagation is beyond local grid cells." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8462" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.104, + 0.282, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.139, + 0.176, + 0.224, + 0.187 + ], + "angle": 0, + "content": "Level 4, \\( V_{4} = 74 \\)" + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.103, + 0.484, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.339, + 0.176, + 0.429, + 0.188 + ], + "angle": 0, + "content": "Level \\(8,V_{8} = 223\\)" + }, + { + "type": "image", + "bbox": [ + 0.487, + 0.103, + 0.687, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.536, + 0.176, + 0.638, + 0.188 + ], + "angle": 0, + "content": "Level 12, \\( V_{12} = 676 \\)" + }, + { + "type": "image", + "bbox": [ + 0.689, + 0.103, + 0.89, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.732, + 0.176, + 0.846, + 0.188 + ], + "angle": 0, + "content": "Level 16, \\(V_{16} = 2048\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.202, + 0.894, + 0.245 + ], + "angle": 0, + "content": "Figure 6. Results at different hash resolutions. While some structures, such as the tree, table, and bike rack, are missed at coarse resolutions (Level 4). Finer resolutions can progressively recover these missing surfaces. Flat continuous surfaces also require sufficiently fine resolutions to predict (Level 8). The result motivates the non-local updates when using numerical gradients for higher-order derivatives." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.259, + 0.892, + 0.399 + ], + "angle": 0, + "content": "
F1 Score ↑PSNR ↑
NeuralWarp [3]COLMAP [27]NeuS [36]AGAG+PNGNG+P (Ours)NeuS [36]AGAG+PNGNG+P (Ours)
Barn0.220.550.290.220.310.630.7026.3626.9126.6926.1428.57
Caterpillar0.180.010.290.230.240.300.3625.2126.0425.1226.1627.81
Courthouse0.080.110.170.080.090.240.2823.5525.4325.6325.0627.23
Ignatius0.020.220.830.720.730.850.8923.2722.6922.7323.7823.67
Meetingroom0.080.190.240.040.050.270.3225.3828.1328.0527.4430.70
Truck0.350.190.450.330.370.440.4823.7123.8923.9522.9925.43
Mean0.150.210.380.270.300.450.5024.5825.5125.3625.2627.24
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.41, + 0.893, + 0.438 + ], + "angle": 0, + "content": "Table 2. Quantitative results on Tanks and Temples dataset [15]. Neuralangelo achieves the best surface reconstruction quality and performs best on average in terms of image synthesis. Best result. Second best result. Best viewed in color." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.466, + 0.206, + 0.563 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.097, + 0.564, + 0.108, + 0.573 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.212, + 0.466, + 0.338, + 0.562 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.252, + 0.564, + 0.297, + 0.574 + ], + "angle": 0, + "content": "w/o \\(\\mathcal{L}_{\\mathrm{curv}}\\)" + }, + { + "type": "image", + "bbox": [ + 0.343, + 0.466, + 0.467, + 0.562 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.388, + 0.564, + 0.428, + 0.574 + ], + "angle": 0, + "content": "w/ \\(\\mathcal{L}_{\\mathrm{curv}}\\)" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.575, + 0.208, + 0.673 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.096, + 0.675, + 0.108, + 0.684 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.212, + 0.575, + 0.338, + 0.673 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.225, + 0.675, + 0.325, + 0.685 + ], + "angle": 0, + "content": "w/o topology warmup" + }, + { + "type": "image", + "bbox": [ + 0.342, + 0.575, + 0.467, + 0.673 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.361, + 0.675, + 0.456, + 0.685 + ], + "angle": 0, + "content": "w/ topology warmup" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.7, + 0.47, + 0.742 + ], + "angle": 0, + "content": "Figure 7. Ablation results. (a) Surface smoothness improves with curvature regularization \\(\\mathcal{L}_{\\mathrm{curv}}\\). (c) Concave shapes are better formed with topology warmup." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.777, + 0.188, + 0.791 + ], + "angle": 0, + "content": "4.4. Ablations" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.471, + 0.901 + ], + "angle": 0, + "content": "**Curvature regularization.** We ablate the necessity of curvature regularization in Neuralangelo and compare the results in Fig. 7(a). Intuitively, \\(\\mathcal{L}_{\\mathrm{curv}}\\) acts as a smoothness prior by minimizing surface curvatures. Without \\(\\mathcal{L}_{\\mathrm{curv}}\\), we find that the surfaces tend to have undesirable sharp transitions. By using \\(\\mathcal{L}_{\\mathrm{curv}}\\), the surface noises are removed." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.465, + 0.895, + 0.602 + ], + "angle": 0, + "content": "Topology warmup. We follow prior work and initialize the SDF approximately as a sphere [42]. With an initial spherical shape, using \\(\\mathcal{L}_{\\mathrm{curv}}\\) also makes concave shapes difficult to form because \\(\\mathcal{L}_{\\mathrm{curv}}\\) preserves topology by preventing singularities in curvature. Thus, instead of applying \\(\\mathcal{L}_{\\mathrm{curv}}\\) from the beginning of the optimization process, we use a short warmup period that linearly increases the curvature loss strength. We find this strategy particularly helpful for concave regions, as shown in Fig. 7(b)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.629, + 0.619, + 0.644 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.895, + 0.901 + ], + "angle": 0, + "content": "We introduce Neuralangelo, an approach for photogrammetric neural surface reconstruction. The findings of Neuralangelo are simple yet effective: using numerical gradients for higher-order derivatives and a coarse-to-fine optimization strategy. Neuralangelo unlocks the representation power of multi-resolution hash encoding for neural surface reconstruction modeled as SDF. We show that Neuralangelo effectively recovers dense scene structures of both object-centric captures and large-scale indoor/outdoor scenes with extremely high fidelity, enabling detailed large-scale scene reconstruction from RGB videos. Our method currently samples pixels from images randomly without tracking their statistics and errors. Therefore, we use long training iterations to reduce the stochastics and ensure sufficient sampling of details. It is our future work to explore a more efficient sampling strategy to accelerate the training process." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8463" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.185, + 0.472, + 0.24 + ], + "angle": 0, + "content": "[2] Rui Chen, Songfang Han, Jing Xu, and Hao Su. Point-based multi-view stereo network. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1538-1547, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.242, + 0.471, + 0.311 + ], + "angle": 0, + "content": "[3] François Darmon, Bénédicte Bascle, Jean-Clément Devaux, Pascal Monasse, and Mathieu Aubry. Improving neural implicit surfaces geometry with patch warping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6260–6269, 2022. 2, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.312, + 0.471, + 0.353 + ], + "angle": 0, + "content": "[4] Jeremy S De Bonet and Paul Viola. Poxels: Probabilistic voxelized volume reconstruction. In Proceedings of International Conference on Computer Vision (ICCV), volume 2, 1999. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.354, + 0.471, + 0.408 + ], + "angle": 0, + "content": "[5] Qiancheng Fu, Qingshan Xu, Yew-Soon Ong, and Wenbing Tao. Geo-neus: Geometry-consistent neural implicit surfaces learning for multi-view reconstruction. arXiv preprint arXiv:2205.15848, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.41, + 0.471, + 0.464 + ], + "angle": 0, + "content": "[6] Yasutaka Furukawa and Jean Ponce. Accurate, dense, and robust multiview stereopsis. IEEE transactions on pattern analysis and machine intelligence, 32(8):1362-1376, 2009. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.466, + 0.471, + 0.522 + ], + "angle": 0, + "content": "[7] Silvano Galliani, Katrin Lasinger, and Konrad Schindler. Massively parallel multiview stereopsis by surface normal diffusion. In Proceedings of the IEEE International Conference on Computer Vision, pages 873-881, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.523, + 0.471, + 0.564 + ], + "angle": 0, + "content": "[8] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. arXiv preprint arXiv:2002.10099, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.565, + 0.471, + 0.605 + ], + "angle": 0, + "content": "[9] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. Cambridge university press, 2003. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.607, + 0.471, + 0.674 + ], + "angle": 0, + "content": "[10] Po-Han Huang, Kevin Matzen, Johannes Kopf, Narendra Ahuja, and Jia-Bin Huang. Deepmvs: Learning multi-view stereopsis. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2821-2830, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.677, + 0.471, + 0.733 + ], + "angle": 0, + "content": "[11] Rasmus Jensen, Anders Dahl, George Vogiatzis, Engil Tola, and Henrik Aanæs. Large scale multi-view stereopsis evaluation. In 2014 IEEE Conference on Computer Vision and Pattern Recognition, pages 406-413. IEEE, 2014. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.734, + 0.471, + 0.801 + ], + "angle": 0, + "content": "[12] Yue Jiang, Dantong Ji, Zhizhong Han, and Matthias Zwicker. Sdfdiff: Differentiable rendering of signed distance fields for 3d shape optimization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1251-1261, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.471, + 0.857 + ], + "angle": 0, + "content": "[13] Michael Kazhdan, Matthew Bolitho, and Hugues Hoppe. Poisson surface reconstruction. In Proceedings of the fourth Eurographics symposium on Geometry processing, volume 7, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.471, + 0.9 + ], + "angle": 0, + "content": "[14] Michael Kazhdan and Hugues Hoppe. Screened poisson surface reconstruction. ACM Transactions on Graphics (ToG), 32(3):1-13, 2013. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.147 + ], + "angle": 0, + "content": "[15] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG), 36(4):1-13, 2017. 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.151, + 0.894, + 0.192 + ], + "angle": 0, + "content": "[16] Kiriakos N Kutulakos and Steven M Seitz. A theory of shape by space carving. International journal of computer vision, 38(3):199-218, 2000. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.194, + 0.893, + 0.236 + ], + "angle": 0, + "content": "[17] Aldo Laurentini. The visual hull concept for silhouette-based image understanding. IEEE Transactions on pattern analysis and machine intelligence, 16(2):150-162, 1994. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.239, + 0.894, + 0.293 + ], + "angle": 0, + "content": "[18] Hai Li, Xingrui Yang, Hongjia Zhai, Yuqian Liu, Hujun Bao, and Guofeng Zhang. Vox-surf: Voxel-based implicit surface representation. IEEE Transactions on Visualization and Computer Graphics, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.296, + 0.893, + 0.352 + ], + "angle": 0, + "content": "[19] Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. Barf: Bundle-adjusting neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5741-5751, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.354, + 0.894, + 0.409 + ], + "angle": 0, + "content": "[20] Chen-Hsuan Lin, Chaoyang Wang, and Simon Lucey. Sdf-srn: Learning signed distance 3d object reconstruction from static images. Advances in Neural Information Processing Systems, 33:11453-11464, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.412, + 0.893, + 0.453 + ], + "angle": 0, + "content": "[21] Bruce D Lucas, Takeo Kanade, et al. An iterative image registration technique with an application to stereo vision, volume 81. Vancouver, 1981. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.456, + 0.894, + 0.526 + ], + "angle": 0, + "content": "[22] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European conference on computer vision, pages 405-421. Springer, 2020. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.528, + 0.894, + 0.583 + ], + "angle": 0, + "content": "[23] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. arXiv preprint arXiv:2201.05989, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.586, + 0.893, + 0.655 + ], + "angle": 0, + "content": "[24] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.657, + 0.893, + 0.728 + ], + "angle": 0, + "content": "[25] Michael Oechsle, Songyou Peng, and Andreas Geiger. Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5589-5599, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.73, + 0.893, + 0.799 + ], + "angle": 0, + "content": "[26] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865-5874, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.801, + 0.893, + 0.843 + ], + "angle": 0, + "content": "[27] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.845, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[28] Johannes L Schonberger, Enliang Zheng, Jan-Michael Frahm, and Marc Pollefeys. Pixelwise view selection for unstructured multi-view stereo. In European conference on computer vision, pages 501-518. Springer, 2016. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8464" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.134 + ], + "angle": 0, + "content": "[29] Steven M Seitz and Charles R Dyer. Photorealistic scene reconstruction by voxel coloring. International Journal of Computer Vision, 35(2):151-173, 1999. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.134, + 0.472, + 0.203 + ], + "angle": 0, + "content": "[30] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5459-5469, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.204, + 0.472, + 0.26 + ], + "angle": 0, + "content": "[31] Jiaming Sun, Xi Chen, Qianqian Wang, Zhengqi Li, Hadar Averbuch-Elor, Xiaowei Zhou, and Noah Snively. Neural 3d reconstruction in the wild. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-9, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.261, + 0.472, + 0.3 + ], + "angle": 0, + "content": "[32] Richard Szeliski. Rapid octree construction from image sequences. CVGIP: Image understanding, 58(1):23-32, 1993. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.302, + 0.472, + 0.385 + ], + "angle": 0, + "content": "[33] Towaki Takikawa, Joey Litalien, Kangxue Yin, Karsten Kreis, Charles Loop, Derek Nowrouzezahrai, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Neural geometric level of detail: Real-time rendering with implicit 3d shapes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11358-11367, 2021. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.386, + 0.472, + 0.44 + ], + "angle": 0, + "content": "[34] Engin Tola, Christoph Strecha, and Pascal Fua. Efficient large-scale multi-view stereo for ultra high-resolution image sets. Machine Vision and Applications, 23(5):903-920, 2012. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.442, + 0.472, + 0.51 + ], + "angle": 0, + "content": "[35] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T Barron, and Pratul P Srinivasan. Ref-nerf: Structured view-dependent appearance for neural radiance fields. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5481-5490. IEEE, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.511, + 0.472, + 0.566 + ], + "angle": 0, + "content": "[36] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021. 2, 3, 4, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.567, + 0.472, + 0.622 + ], + "angle": 0, + "content": "[37] Yiming Wang, Qin Han, Marc Habermann, Kostas Daniilidis, Christian Theobalt, and Lingjie Liu. Neus2: Fast learning of neural implicit surfaces for multi-view reconstruction. arXiv preprint arXiv:2212.05231, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.623, + 0.472, + 0.664 + ], + "angle": 0, + "content": "[38] Yiqun Wang, Ivan Skorokhodov, and Peter Wonka. Hf-neus: Improved surface reconstruction using high-frequency details. In Advances in Neural Information Processing Systems. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.665, + 0.472, + 0.719 + ], + "angle": 0, + "content": "[39] Tong Wu, Jiaqi Wang, Xingang Pan, Xudong Xu, Christian Theobalt, Ziwei Liu, and Dahua Lin. Voxurf: Voxel-based efficient and accurate neural surface reconstruction. arXiv preprint arXiv:2208.12697, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.72, + 0.472, + 0.775 + ], + "angle": 0, + "content": "[40] Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. Mvsnet: Depth inference for unstructured multi-view stereo. In Proceedings of the European conference on computer vision (ECCV), pages 767-783, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.472, + 0.83 + ], + "angle": 0, + "content": "[41] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems, 34:4805-4815, 2021. 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.472, + 0.9 + ], + "angle": 0, + "content": "[42] Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Basri Ronen, and Yaron Lipman. Multiview neural surface reconstruction by disentangling geometry and appearance. Advances in Neural Information Processing Systems, 33:2492-2502, 2020. 2, 4, 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.147 + ], + "angle": 0, + "content": "[43] Alex Yu, Sara Fridovich-Keil, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. arXiv preprint arXiv:2112.05131, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.894, + 0.205 + ], + "angle": 0, + "content": "[44] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. arXiv preprint arXiv:2206.00665, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.206, + 0.894, + 0.275 + ], + "angle": 0, + "content": "[45] Jingyang Zhang, Yao Yao, Shiwei Li, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. Critical regularizations for neural surface reconstruction in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6270-6279, 2022. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.277, + 0.893, + 0.318 + ], + "angle": 0, + "content": "[46] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.319, + 0.894, + 0.388 + ], + "angle": 0, + "content": "[47] Xiuming Zhang, Pratul P Srinivasan, Boyang Deng, Paul Debevec, William T Freeman, and Jonathan T Barron. Nerfactor: Neural factorization of shape and reflectance under an unknown illumination. ACM Transactions on Graphics (TOG), 40(6):1-18, 2021. 4" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.894, + 0.388 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8465" + } + ] +] \ No newline at end of file diff --git a/2023/Neuralangelo_ High-Fidelity Neural Surface Reconstruction/55f0b247-41a3-45d9-bec0-4490ab23bed9_origin.pdf b/2023/Neuralangelo_ High-Fidelity Neural Surface Reconstruction/55f0b247-41a3-45d9-bec0-4490ab23bed9_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..321e264d5af2f63a4899629281f42be394ba10c9 --- /dev/null +++ b/2023/Neuralangelo_ High-Fidelity Neural Surface Reconstruction/55f0b247-41a3-45d9-bec0-4490ab23bed9_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b0cc4b0beb41851cbb94beea73dbacb6ab22b09dd311d39fabdf6e91d6b16764 +size 3189350 diff --git a/2023/Neuralangelo_ High-Fidelity Neural Surface Reconstruction/full.md b/2023/Neuralangelo_ High-Fidelity Neural Surface Reconstruction/full.md new file mode 100644 index 0000000000000000000000000000000000000000..050a55e14a9e4f333d485aef31ee7f7e0cf48c41 --- /dev/null +++ b/2023/Neuralangelo_ High-Fidelity Neural Surface Reconstruction/full.md @@ -0,0 +1,344 @@ +# Neuralangelo: High-Fidelity Neural Surface Reconstruction + +Zhaoshuo Li $^{1,2}$ Thomas Müller $^{1}$ Alex Evans $^{1}$ Russell H. Taylor $^{2}$ Mathias Unberath $^{2}$ Ming-Yu Liu $^{1}$ Chen-Hsuan Lin $^{1}$ $^{1}$ NVIDIA Research $^{2}$ Johns Hopkins University https://research.nvidia.com/labs/dir/neuralangelo + +![](images/58f396abf90bacdda4f6894b60e1a077eefff62560d389ed557b1b5c4dcc227b.jpg) +Figure 1. We present Neuralangelo, a framework for high-fidelity 3D surface reconstruction from RGB images using neural volume rendering, even without auxiliary data such as segmentation or depth. Shown in the figure is an extracted 3D mesh of a courthouse. + +# Abstract + +Neural surface reconstruction has been shown to be powerful for recovering dense 3D surfaces via image-based neural rendering. However, current methods struggle to recover detailed structures of real-world scenes. To address the issue, we present Neuralangelo, which combines the representation power of multi-resolution 3D hash grids with neural surface rendering. Two key ingredients enable our approach: (1) numerical gradients for computing higher-order derivatives as a smoothing operation and (2) coarse-to-fine optimization on the hash grids controlling different levels of details. Even without auxiliary inputs such as depth, Neuralangelo can effectively recover dense 3D surface structures from multi-view images with fidelity significantly surpassing previous methods, enabling detailed large-scale scene reconstruction from RGB video captures. + +# 1. Introduction + +3D surface reconstruction aims to recover dense geometric scene structures from multiple images observed at different viewpoints [9]. The recovered surfaces provide structural information useful for many downstream applications, such as 3D asset generation for augmented/virtual/mixed reality or environment mapping for autonomous navigation of robotics. Photogrammetric surface reconstruction using a monocular RGB camera is of particular interest, as it equips users with the capability of casually creating digital twins of the real world using ubiquitous mobile devices. + +Classically, multi-view stereo algorithms [6, 16, 29, 34] had been the method of choice for sparse 3D reconstruction. An inherent drawback of these algorithms, however, is their inability to handle ambiguous observations, e.g. regions with large areas of homogeneous colors, repetitive texture + +patterns, or strong color variations. This would result in inaccurate reconstructions with noisy or missing surfaces. Recently, neural surface reconstruction methods [36,41,42] have shown great potential in addressing these limitations. This new class of methods uses coordinate-based multi-layer perceptrons (MLPs) to represent the scene as an implicit function, such as occupancy fields [25] or signed distance functions (SDF) [36,41,42]. Leveraging the inherent continuity of MLPs and neural volume rendering [22], these techniques allow the optimized surfaces to meaningfully interpolate between spatial locations, resulting in smooth and complete surface representations. + +Despite the superiority of neural surface reconstruction methods over classical approaches, the recovered fidelity of current methods does not scale well with the capacity of MLPs. Recently, Müller et al. [23] proposed a new scalable representation, referred to as Instant NGP (Neural Graphics Primitives). Instant NGP introduces a hybrid 3D grid structure with a multi-resolution hash encoding and a lightweight MLP that is more expressive with a memory footprint log-linear to the resolution. The proposed hybrid representation greatly increases the representation power of neural fields and has achieved great success at representing very fine-grained details for a wide variety of tasks, such as object shape representation and novel view synthesis problems. + +In this paper, we propose Neuralangelo for high-fidelity surface reconstruction (Fig. 1). Neuralangelo adopts Instant NGP as a neural SDF representation of the underlying 3D scene, optimized from multi-view image observations via neural surface rendering [36]. We present two findings central to fully unlocking the potentials of multi-resolution hash encodings. First, using numerical gradients to compute higher-order derivatives, such as surface normals for the eikonal regularization [8, 12, 20, 42], is critical to stabilizing the optimization. Second, a progressive optimization schedule plays an important role in recovering the structures at different levels of details. We combine these two key ingredients and, via extensive experiments on standard benchmarks and real-world scenes, demonstrate significant improvements over image-based neural surface reconstruction methods in both reconstruction accuracy and view synthesis quality. + +In summary, we present the following contributions: + +- We present the Neuralangelo framework to naturally incorporate the representation power of multi-resolution hash encoding [23] into neural SDF representations. +- We present two simple techniques to improve the quality of hash-encoded surface reconstruction: higher-order derivatives with numerical gradients and coarse-to-fine optimization with a progressive level of details. +- We empirically demonstrate the effectiveness of Neuralangelo on various datasets, showing significant improvements over previous methods. + +# 2. Related work + +Multi-view surface reconstruction. Early image-based photogrammetry techniques use a volumetric occupancy grid to represent the scene [4, 16, 17, 29, 32]. Each voxel is visited and marked occupied if strict color constancy between the corresponding projected image pixels is satisfied. The photometric consistency assumption typically fails due to autoexposure or non-Lambertian materials, which are ubiquitous in the real world. Relaxing such color constancy constraints across views is important for realistic 3D reconstruction. + +Follow-up methods typically start with 3D point clouds from multi-view stereo techniques [6,7,28,34] and then perform dense surface reconstruction [13,14]. Reliance on the quality of the generated point clouds often leads to missing or noisy surfaces. Recent learning-based approaches augment the point cloud generation process with learned image features and cost volume construction [2,10,40]. However, these approaches are inherently limited by the resolution of the cost volume and fail to recover geometric details. + +Neural Radiance Fields (NeRF). NeRF [22] achieves remarkable photorealistic view synthesis with view-dependent effects. NeRF encodes 3D scenes with an MLP mapping 3D spatial locations to color and volume density. These predictions are composited into pixel colors using neural volume rendering. A problem of NeRF and its variants [1,30,43,46], however, is the question of how an isosurface of the volume density could be defined to represent the underlying 3D geometry. Current practice often relies on heuristic thresholding on the density values; due to insufficient constraints on the level sets, however, such surfaces are often noisy and may not model the scene structures accurately [36,41]. Therefore, more direct modeling of surfaces is preferred for photogrammetric surface reconstruction problems. + +Neural surface reconstruction. For scene representations with better-defined 3D surfaces, implicit functions such as occupancy grids [24, 25] or SDFs [42] are preferred over simple volume density fields. To integrate with neural volume rendering [22], different techniques [36, 41] have been proposed to reparametrize the underlying representations back to volume density. These designs of neural implicit functions enable more accurate surface prediction with view synthesis capabilities of unsacrificed quality [42]. + +Follow-up works extend the above approaches to real-time at the cost of surface fidelity [18, 37], while others [3,5,44] use auxiliary information to enhance the reconstruction results. Notably, NeuralWarp [3] uses patch warping given co-visibility information from structure-from-motion (SfM) to guide surface optimization, but the patchwise planar assumption fails to capture highly-varying surfaces [3]. Other methods [5,45] utilize sparse point clouds from SfM to supervise the SDF, but their performances are upper-bounded by the quality of the point clouds, as with + +classical approaches [45]. The use of monocular depth and segmentation as auxiliary data has also been explored with unconstrained image collections [31] or using scene representations with hash encodings [44]. In contrast, our work Neuralangelo builds upon hash encodings [23] to recover surfaces but without the need for auxiliary inputs used in prior work [3, 5, 31, 44, 45]. Concurrent work [38] also proposes coarse-to-fine optimization for improved surface details, where a displacement network corrects the shape predicted by a coarse network. In contrast, we use hierarchical hash grids and control the level of details based on our analysis of higher-order derivatives. + +# 3. Approach + +Neuralangelo reconstructs dense structures of the scene from multi-view images. Neuralangelo samples 3D locations along camera view directions and uses a multi-resolution hash encoding to encode the positions. The encoded features are input to an SDF MLP and a color MLP to composite images using SDF-based volume rendering. + +# 3.1. Preliminaries + +Neural volume rendering. NeRF [22] represents a 3D scene as volume density and color fields. Given a posed camera and a ray direction, the volume rendering scheme integrates the color radiance of sampled points along the ray. The $i$ -th sampled 3D position $\mathbf{x}_i$ is at a distance $t_i$ from the camera center. The volume density $\sigma_i$ and color $\mathbf{c}_i$ of each sampled point are predicted using a coordinate MLP. The rendered color of a given pixel is approximated as the Riemann sum: + +$$ +\hat {\mathbf {c}} (\mathbf {o}, \mathbf {d}) = \sum_ {i = 1} ^ {N} w _ {i} \mathbf {c} _ {i}, \quad \text {w h e r e} w _ {i} = T _ {i} \alpha_ {i}. \tag {1} +$$ + +Here, $\alpha_{i} = 1 - \exp (-\sigma_{i}\delta_{i})$ is the opacity of the $i$ -th ray segment, $\delta_{i} = t_{i + 1} - t_{i}$ is the distance between adjacent samples, and $T_{i} = \Pi_{j = 1}^{i - 1}(1 - \alpha_{j})$ is the accumulated transmittance, indicating the fraction of light that reaches the camera. To supervise the network, a color loss is used between input images $\mathbf{c}$ and rendered images $\hat{\mathbf{c}}$ : + +$$ +\mathcal {L} _ {\mathrm {R G B}} = \left\| \hat {\mathbf {c}} - \mathbf {c} \right\| _ {1}. \tag {2} +$$ + +However, surfaces are not clearly defined using such density formulation. Extracting surfaces from density-based representation often leads to noisy and unrealistic results [36, 41]. + +Volume rendering of SDF. One of the most common surface representations is SDF. The surface $S$ of an SDF can be implicitly represented by its zero-level set, i.e., $S = \{\mathbf{x} \in \mathbb{R}^3 | f(\mathbf{x}) = 0\}$ , where $f(\mathbf{x})$ is the SDF value. In the context of neural SDFs, Wang et al. [36] proposed to convert volume density predictions in NeRF to SDF representations with a logistic function to allow optimization with neural volume + +rendering. Given a 3D point $\mathbf{x}_i$ and SDF value $f(\mathbf{x}_i)$ , the corresponding opacity value $\alpha_i$ used in Eq. 1 is computed as + +$$ +\alpha_ {i} = \max \left(\frac {\Phi_ {s} (f (\mathbf {x} _ {i})) - \Phi_ {s} (f (\mathbf {x} _ {i + 1}))}{\Phi_ {s} (f (\mathbf {x} _ {i}))}, 0\right), \tag {3} +$$ + +where $\Phi_s$ is the sigmoid function. In this work, we use the same SDF-based volume rendering formulation [36]. + +Multi-resolution hash encoding. Recently, multi-resolution hash encoding proposed by Müller et al. [23] has shown great scalability for neural scene representations, generating fine-grained details for tasks such as novel view synthesis. In Neuralangelo, we adopt the representation power of hash encoding to recover high-fidelity surfaces. + +The hash encoding uses multi-resolution grids, with each grid cell corner mapped to a hash entry. Each hash entry stores the encoding feature. Let $\{V_1,\dots,V_L\}$ be the set of different spatial grid resolutions. Given an input position $\mathbf{x}_i$ , we map it to the corresponding position at each grid resolution $V_{l}$ as $\mathbf{x}_{i,l} = \mathbf{x}_i\cdot V_l$ . The feature vector $\gamma_l(\mathbf{x}_{i,l})\in \mathbb{R}^c$ given resolution $V_{l}$ is obtained via trilinear interpolation of hash entries at the grid cell corners. The encoding features across all spatial resolutions are concatenated together, forming a $\gamma (\mathbf{x}_i)\in \mathbb{R}^{cL}$ feature vector: + +$$ +\gamma \left(\mathbf {x} _ {i}\right) = \left(\gamma_ {1} \left(\mathbf {x} _ {i, 1}\right), \dots , \gamma_ {L} \left(\mathbf {x} _ {i, L}\right)\right). \tag {4} +$$ + +The encoded features are then passed to a shallow MLP. + +One alternative to hash encoding is sparse voxel structures [30, 33, 39, 43], where each grid corner is uniquely defined without collision. However, volumetric feature grids require hierarchical spatial decomposition (e.g. octrees) to make the parameter count tractable; otherwise, the memory would grow cubically with spatial resolution. Given such hierarchy, finer voxel resolutions by design cannot recover surfaces that are misrepresented by the coarser resolutions [33]. Hash encoding instead assumes no spatial hierarchy and resolves collision automatically based on gradient averaging [23]. + +# 3.2. Numerical Gradient Computation + +We show in this section that the analytical gradient w.r.t. position of hash encoding suffers from localities. Therefore, optimization updates only propagate to local hash grids, lacking non-local smoothness. We propose a simple fix to such a locality problem by using numerical gradients. An overview is shown in Fig. 2. + +A special property of SDF is its differentiability with a gradient of the unit norm. The gradient of SDF satisfies the eikonal equation $\| \nabla f(\mathbf{x})\| _2 = 1$ (almost everywhere). To enforce the optimized neural representation to be a valid SDF, the eikonal loss [8] is typically imposed on the SDF predictions: + +$$ +\mathcal {L} _ {\mathrm {e i k}} = \frac {1}{N} \sum_ {i = 1} ^ {N} \left(\| \nabla f (\mathbf {x} _ {i}) \| _ {2} - 1\right) ^ {2}, \tag {5} +$$ + +![](images/75ebad35cb2bbf8018124cffdf59a378eb3f80743bbce21148a13ad52a882865.jpg) +Figure 2. Using numerical gradients for higher-order derivatives distributes the back-propagation updates beyond the local hash grid cell, thus becoming a smoothed version of analytical gradients. + +where $N$ is the total number of sampled points. To allow for end-to-end optimization, a double backward operation on the SDF prediction $f(\mathbf{x})$ is required. + +The de facto method for computing surface normals of SDFs $\nabla f(\mathbf{x})$ is to use analytical gradients [36,41,42]. Analytical gradients of hash encoding w.r.t. position, however, are not continuous across space under trilinear interpolation. To find the sampling location in a voxel grid, each 3D point $\mathbf{x}_i$ would first be scaled by the grid resolution $V_{l}$ , written as $\mathbf{x}_{i,l} = \mathbf{x}_i\cdot V_l$ . Let the coefficient for (tri-)linear interpolation be $\beta = \mathbf{x}_{i,l} - \left\lfloor \mathbf{x}_{i,l}\right\rfloor$ . The resulting feature vectors are + +$$ +\gamma_ {l} \left(\mathbf {x} _ {i, l}\right) = \gamma_ {l} \left(\left\lfloor \mathbf {x} _ {i, l} \right\rfloor\right) \cdot (1 - \beta) + \gamma_ {l} \left(\left\lceil \mathbf {x} _ {i, l} \right\rceil\right) \cdot \beta , \tag {6} +$$ + +where the rounded position $\lfloor \mathbf{x}_{i,l}\rfloor ,\lceil \mathbf{x}_{i,l}\rceil$ correspond to the local grid cell corners. We note that rounding operations $\lfloor \cdot \rfloor$ and $\lceil \cdot \rceil$ are non-differentiable. As a result, the derivative of hash encoding w.r.t. the position can be obtained as + +$$ +\begin{array}{l} \frac {\partial \gamma_ {l} (\mathbf {x} _ {i , l})}{\partial \mathbf {x} _ {i}} = \gamma_ {l} (\lfloor \mathbf {x} _ {i, l} \rfloor) \cdot (- \frac {\partial \beta}{\partial \mathbf {x} _ {i , l}}) + \gamma_ {l} (\lceil \mathbf {x} _ {i, l} \rceil) \cdot \frac {\partial \beta}{\partial \mathbf {x} _ {i , l}} \\ = \gamma_ {l} (\left\lfloor \mathbf {x} _ {i, l} \right\rfloor) \cdot (- V _ {l}) + \gamma_ {l} (\left\lceil \mathbf {x} _ {i, l} \right\rceil) \cdot V _ {l}. \tag {7} \\ \end{array} +$$ + +The derivative of hash encoding is local, i.e., when $\mathbf{x}_i$ moves across grid cell borders, the corresponding hash entries will be different. Therefore, the eikonal loss defined in Eq. 5 only back-propagates to the locally sampled hash entries, i.e. $\gamma_l(\lfloor \mathbf{x}_{i,l}\rfloor)$ and $\gamma_l(\lceil \mathbf{x}_{i,l}\rceil)$ . When continuous surfaces (e.g. a flat wall) span multiple grid cells, these grid cells should produce coherent surface normals without sudden transitions. To ensure consistency in surface representation, joint optimization of these grid cells is desirable. However, the analytical gradient is limited to local grid cells, unless all corresponding grid cells happen to be sampled and optimized simultaneously. Such sampling is not always guaranteed. + +To overcome the locality of the analytical gradient of hash encoding, we propose to compute the surface normals using numerical gradients. If the step size of the numerical gradient is smaller than the grid size of hash encoding, the numerical gradient would be equivalent to the analytical gradient; otherwise, hash entries of multiple grid cells + +would participate in the surface normal computation. Backpropagating through the surface normals thus allows hash entries of multiple grids to receive optimization updates simultaneously. Intuitively, numerical gradients with carefully chosen step sizes can be interpreted as a smoothing operation on the analytical gradient expression. An alternative of normal supervision is a teacher-student curriculum [35, 47], where the predicted noisy normals are driven towards MLP outputs to exploit the smoothness of MLPs. However, analytical gradients from such teacher-student losses still only back-propagate to local grid cells for hash encoding. In contrast, numerical gradients solve the locality issue without the need of additional networks. + +To compute the surface normals using the numerical gradient, additional SDF samples are needed. Given a sampled point $\mathbf{x}_i = (x_i,y_i,z_i)$ , we additionally sample two points along each axis of the canonical coordinate around $x_{i}$ within a vicinity of a step size of $\epsilon$ . For example, the $x$ -component of the surface normal can be found as + +$$ +\nabla_ {x} f (\mathbf {x} _ {i}) = \frac {f \left(\gamma \left(\mathbf {x} _ {i} + \boldsymbol {\epsilon} _ {x}\right)\right) - f \left(\gamma \left(\mathbf {x} _ {i} - \boldsymbol {\epsilon} _ {x}\right)\right)}{2 \epsilon}, \tag {8} +$$ + +where $\epsilon_{x} = [\epsilon, 0, 0]$ . In total, six additional SDF samples are required for numerical surface normal computation. + +# 3.3. Progressive Levels of Details + +Coarse-to-fine optimization can better shape the loss landscape to avoid falling into false local minima. Such a strategy has found many applications in computer vision, such as image-based registration [19,21,26]. Neuralangelo also adopts a coarse-to-fine optimization scheme to reconstruct the surfaces with progressive levels of details. Using numerical gradients for the higher-order derivatives naturally enables Neuralangelo to perform coarse-to-fine optimization from two perspectives. + +Step size $\epsilon$ . As previously discussed, numerical gradients can be interpreted as a smoothing operation where the step size $\epsilon$ controls the resolution and the amount of recovered details. Imposing $\mathcal{L}_{\mathrm{eik}}$ with a larger $\epsilon$ for numerical surface normal computation ensures the surface normal is consistent at a larger scale, thus producing consistent and continuous surfaces. On the other hand, imposing $\mathcal{L}_{\mathrm{eik}}$ with a smaller $\epsilon$ affects a smaller region and avoids smoothing details. In practice, we initialize the step size $\epsilon$ to the coarsest hash grid size and exponentially decrease it matching different hash grid sizes throughout the optimization process. + +Hash grid resolution $V$ . If all hash grids are activated from the start of the optimization, to capture geometric details, fine hash grids must first "unlearn" from the coarse optimization with large step size $\epsilon$ and "relearn" with a smaller $\epsilon$ . If such a process is unsuccessful due to converged optimization, geometric details would be lost. Therefore, we only enable an initial set of coarse hash grids and progressively activate + +![](images/aa07006df64d5950aea79119972ef9c9246d1e8b40b82cd7a60d66768e6013d0.jpg) +Figure 3. Qualitative comparison on the DTU benchmark [11]. Neuralangelo produces more accurate and higher-fidelity surfaces. + +finer hash grids throughout optimization when $\epsilon$ decreases to their spatial size. The relearning process can thus be avoided to better capture the details. In practice, we also apply weight decay over all parameters to avoid single-resolution features dominating the final results. + +# 3.4. Optimization + +To further encourage the smoothness of the reconstructed surfaces, we impose a prior by regularizing the mean curvature of SDF. The mean curvature is computed from discrete Laplacian similar to the surface normal computation, otherwise, the second-order analytical gradients of hash encoding are zero everywhere when using trilinear interpolation. The curvature loss $\mathcal{L}_{\mathrm{curv}}$ is defined as: + +$$ +\mathcal {L} _ {\mathrm {c u r v}} = \frac {1}{N} \sum_ {i = 1} ^ {N} \left| \nabla^ {2} f (\mathbf {x} _ {i}) \right|. \tag {9} +$$ + +We note that the samples used for the surface normal computation in Eq. 8 are sufficient for curvature computation. + +The total loss is defined as the weighted sum of losses: + +$$ +\mathcal {L} = \mathcal {L} _ {\mathrm {R G B}} + w _ {\mathrm {e i k}} \mathcal {L} _ {\mathrm {e i k}} + w _ {\mathrm {c u r v}} \mathcal {L} _ {\mathrm {c u r v}}. \tag {10} +$$ + +All network parameters, including MLPs and hash encoding, are trained jointly end-to-end. + +# 4. Experiments + +Datasets. Following prior work, we conduct experiments on 15 object-centric scenes of the DTU dataset [11]. Each scene has 49 or 64 images captured by a robot-held monocular RGB camera. The ground truth is obtained from a structured-light scanner. We further conduct experiments on 6 scenes + +of the Tanks and Temples dataset [15], including large-scale indoor/outdoor scenes. Each scene contains 263 to 1107 images captured using a hand-held monocular RGB camera. The ground truth is obtained using a LiDAR sensor. + +Implementation details. Our hash encoding resolution spans $2^{5}$ to $2^{11}$ with 16 levels. Each hash entry has a channel size of 8. The maximum number of hash entries of each resolution is $2^{22}$ . We activate 4 and 8 hash resolutions at the beginning of optimization for DTU dataset and Tanks and Temples respectively, due to differences in scene scales. We enable a new hash resolution every 5000 iterations when the step size $\epsilon$ equals its grid cell size. For all experiments, we do not utilize auxiliary data such as segmentation or depth during the optimization process. + +Evaluation criteria. We report Chamfer distance and F1 score for surface evaluation [11, 15]. We use peak signal-to-noise ratio (PSNR) to report image synthesis qualities. + +# 4.1. DTU Benchmark + +We show qualitative results in Fig. 3 and quantitative results in Table 1. On average, Neuralangelo achieves the lowest Chamfer distance and the highest PSNR, even without using auxiliary inputs. The result suggests that Neuralangelo is more generally applicable than prior work when recovering surfaces and synthesizing images, despite not performing best in every individual scene. + +We further ablate Neuralangelo against the following conditions: 1) AG: analytical gradients, 2) $\mathrm{AG + P}$ : analytical gradients and progressive activating hash resolutions, 3) NG: numerical gradients with varying $\epsilon$ . Fig. 4 shows the results qualitatively. AG produces noisy surfaces, even with hash resolutions progressively activated (AG+P). NG improves + +![](images/71f810f95a3ffcb08e0c9efc4c5b5fa1872c8b2edfd37c17ab9e23582ed2f384.jpg) +Input + +![](images/c07bb2e00958992cfbd75cd3c96dec67113af298b84f477b1cb0a9b5137f1255.jpg) +AG + +![](images/2d79fb8d5d9826cbb9b4feadd0986438bd89feb2f0532af362ec668c4a67bcf7.jpg) +AG+P +Figure 4. Qualitative comparison of different coarse-to-fine optimization scheme. When using the analytical gradient (AG and $\mathrm{AG + P}$ ), coarse surfaces often contain artifacts. While using numerical gradients (NG) leads to a better coarse shape, details are also smoothed. Our solution $(\mathrm{NG} + \mathrm{P})$ produces both smooth surfaces and fine details. + +![](images/62c4f776a7767f6fb5334eeee9c5c34bf35928499d773f69ca6d7b15240cc68a.jpg) +NG + +![](images/7426cfc657beb8d7e82ba0307ff20f336da469f1e77942152e53152ee1319d51.jpg) +NG+P (Ours) + +
243740556365698397105106110114118122Mean
NeRF [22]1.901.601.850.582.281.271.471.672.051.070.882.531.061.150.961.49
VolSDF [41]1.141.260.810.491.250.700.721.291.180.700.661.080.420.610.550.86
NeuS [36]1.001.370.930.431.100.650.571.481.090.830.521.200.350.490.540.84
HF-NeuS [38]0.761.320.700.391.060.630.631.151.120.800.521.220.330.490.500.77
RegSDF [45] †0.601.410.640.431.340.620.600.900.921.020.600.590.300.410.390.72
NeuralWarp [3] †0.490.710.380.380.790.810.821.201.060.680.660.740.410.630.510.68
AG0.671.040.840.391.431.231.111.241.540.850.501.010.370.510.440.88
AG+P0.590.950.460.341.190.700.791.191.370.690.490.930.330.440.440.73
NG0.480.810.430.350.890.710.611.261.060.740.470.790.330.450.430.65
NG+P (Ours)0.370.720.350.350.870.540.531.290.970.730.470.740.320.410.430.61
RegSDF [45] †24.7823.0623.4722.2128.5725.5321.8128.8926.8127.9124.7125.1326.8421.6728.2525.31
NeuS [36]26.6223.6426.4325.5930.6132.8329.2433.7126.8531.9732.1828.9228.4135.0034.8129.79
VolSDF [41]26.2825.6126.5526.7631.5731.5029.3833.2328.0332.1333.1631.4930.3334.9034.7530.38
NeRF [22]26.2425.7426.7927.5731.9631.5029.5832.7828.3532.0833.4931.5431.0035.5935.5130.65
AG29.9724.9823.1130.2730.6031.2729.2734.2227.4733.0933.8529.9829.4135.6935.1130.55
AG+P30.1224.6329.5930.2931.6032.0429.8534.1927.8233.2333.9529.1529.4435.9935.6731.17
NG30.3425.1430.2030.7931.7231.8629.8134.3628.0133.4534.3830.3929.8836.0235.7431.47
NG+P (Ours)30.6427.7832.7034.1835.1535.8931.4736.8230.1335.9236.6132.6031.2038.4138.0533.84
+ +Table 1. Quantitative results on DTU dataset [11]. Neuralangelo achieves the best reconstruction accuracy and image synthesis quality. Best result. Second best result. $\dagger$ Requires 3D points from SfM. Best viewed in color. + +the smoothness of the surface, sacrificing details. Our setup $(\mathrm{NG} + \mathrm{P})$ produces both smooth surfaces and fine details. + +# 4.2. Tanks and Temples + +As no public result is available for Tanks and Temples, we train NeuS [36] and NeuralWarp [3] following our setup. We also report classical multi-view stereo results using COLMAP [27]. As COLMAP and NeuralWarp do not support view synthesis, we only report PSNR from NeuS. Results are summarized in Fig. 5 and Table 2. + +Neuralangelo achieves the highest average PSNR and performs best in terms of F1 score. Comparing against NeuS [36], we can recover high-fidelity surfaces with intricate details. We find that the dense surfaces generated from COLMAP are sensitive to outliers in the sparse point cloud. We also find that NeuralWarp often predicts surfaces for the sky and backgrounds potentially due to their color rendering + +scheme following VolSDF [41]. The additional surfaces predicted for backgrounds are counted as outliers and worsen F1 scores significantly. We instead follow NeuS [36] and use an additional network [46] to model the background. + +Similar to the DTU results, using the analytical gradient produces noisy surfaces and thus leads to a low F1 score. We further note that the reconstruction of Courthouse shown in Figs. 1 and 5 are the same building of different sides, demonstrating the capability of Neuralangelo for large-scale granular reconstruction. + +# 4.3. Level of Details + +As Neuralangelo progressively optimizes the hash features of increasing resolution, we inspect the progressive level of details similar to NGLOD [33]. We show a qualitative visualization in Fig. 6. While some surfaces are entirely missed by coarse levels, for example, the tree, table, and + +![](images/67c4bce5337f42302fa3cd96147fca72e45ebcc21b7b0a5158358f407eadd2bc.jpg) + +![](images/e38973828cda37f01f6bc7ee658681a5d63e56e5039c168136126a2992c6c58f.jpg) + +![](images/a10bc5bbf58fb5d5eba7c874bbeace98b88df4b599161990a6156dee0f29ba12.jpg) + +![](images/2da33cdffd118fcfb89d09e0be0e270fc28d477ba949b1338aa165010089f9f7.jpg) + +![](images/a18f9fc0ba2dbcc2450c94cd56489696111aeb083b95c806dfd8af9f0511eb07.jpg) + +![](images/69a68951d6161edd18924e59ec2a63ee06b8dc50ae51c752ae55e019c49d2381.jpg) + +![](images/6ecb24f0630b1554cbead62a326a6d539c633717e5bda433db8f5f1d05179f28.jpg) + +![](images/1615ec4c42d2595472e0ed0a2981383050c0d11bab0bc47f8caf6c5dcea8df2b.jpg) + +![](images/f8f2d5e4bde5788e86c434f140b05612665682cb61f2e7bae950d8ce46a89230.jpg) + +![](images/ea8b9fda6fa086f86c622b4ed0462512b09b44dbe8e85d9e763c817a4dd691d1.jpg) + +![](images/0b4f59f39ea594ca8f0ee72166f7e166d053ecb8e17fbbb3c8984210fcf960cf.jpg) + +![](images/eed056bf9dd56dfb4fec284ee73dfc5f353fa5053cdcd6056f3a5673b3aa903d.jpg) + +![](images/1f6d66d561729430cea1aa243bfca2a77ad690457872229c5576df9e8c24456c.jpg) +Figure 5. Qualitative comparison on Tanks and Temples dataset [15]. Neuralangelo captures the scene details better compared to other baseline approaches, while baseline approaches have missing or noisy surfaces. + +![](images/502eef194b30ad0652e8292c05c570b693fedeff00956eb8b0c053235db66b76.jpg) + +![](images/eac1f6398a4dffb4cea3d41125fedf42e597e6186b6a5c204b0cd5da1f9cafb5.jpg) + +bike rack, these structures are recovered by finer resolutions successfully. The ability to recover missing surfaces demonstrates the advantages of our spatial hierarchy-free design. + +Moreover, we note that flat surfaces are predicted at sufficiently high resolutions (around Level 8 in this example). + +Thus, only relying on the continuity of local cells of coarse resolutions is not sufficient to reconstruct large continuous surfaces. The result motivates the use of the numerical gradients for the higher-order derivatives, such that backpropagation is beyond local grid cells. + +![](images/8320594a831f27d587fc460e0fc72f4ee0071a01f83b88170f9e247dce384a7a.jpg) +Level 4, $V_{4} = 74$ + +![](images/f47dac0ea1fa0f539c5cd69c83ebf67eea011e9f2803f2efd54a3b0afec40248.jpg) +Level $8,V_{8} = 223$ +Figure 6. Results at different hash resolutions. While some structures, such as the tree, table, and bike rack, are missed at coarse resolutions (Level 4). Finer resolutions can progressively recover these missing surfaces. Flat continuous surfaces also require sufficiently fine resolutions to predict (Level 8). The result motivates the non-local updates when using numerical gradients for higher-order derivatives. + +![](images/5cbfcbb7ecfd4bd072bd1ef91b0286740808c7fce0d3f2f23a7e9a93865624b5.jpg) +Level 12, $V_{12} = 676$ + +![](images/236977ccbd1808e95fa4b1500a10c0c4ef0ede1af924d5764cff7ed12cceed40.jpg) +Level 16, $V_{16} = 2048$ + +
F1 Score ↑PSNR ↑
NeuralWarp [3]COLMAP [27]NeuS [36]AGAG+PNGNG+P (Ours)NeuS [36]AGAG+PNGNG+P (Ours)
Barn0.220.550.290.220.310.630.7026.3626.9126.6926.1428.57
Caterpillar0.180.010.290.230.240.300.3625.2126.0425.1226.1627.81
Courthouse0.080.110.170.080.090.240.2823.5525.4325.6325.0627.23
Ignatius0.020.220.830.720.730.850.8923.2722.6922.7323.7823.67
Meetingroom0.080.190.240.040.050.270.3225.3828.1328.0527.4430.70
Truck0.350.190.450.330.370.440.4823.7123.8923.9522.9925.43
Mean0.150.210.380.270.300.450.5024.5825.5125.3625.2627.24
+ +Table 2. Quantitative results on Tanks and Temples dataset [15]. Neuralangelo achieves the best surface reconstruction quality and performs best on average in terms of image synthesis. Best result. Second best result. Best viewed in color. + +![](images/1fc4cc79b98d1130719e1ed5f4239b95ccf98c402839a1538b9ee80163530954.jpg) +(a) + +![](images/a138717460f089da41ba1890ec75547940740bdec19e65976076f28dc647f56b.jpg) +w/o $\mathcal{L}_{\mathrm{curv}}$ + +![](images/1b25b09ec3140664a0b0fadb43241a8209c40759ade3fb2ef412caa0a23ac69c.jpg) +w/ $\mathcal{L}_{\mathrm{curv}}$ + +![](images/0892a03f597996cb264649a124d1a9d95065c6189c580a0d1d2ace4e3682268a.jpg) +(b) +Figure 7. Ablation results. (a) Surface smoothness improves with curvature regularization $\mathcal{L}_{\mathrm{curv}}$ . (c) Concave shapes are better formed with topology warmup. + +![](images/c3c3eb14135ab9e63ca9576391a2e65d113669f7b93eb2e036d0694d0f2f6bb9.jpg) +w/o topology warmup + +![](images/fac9df24cadf56f3931a617d79f972db2673500a5cf88fc3ee1dacbd78e3efbd.jpg) +w/ topology warmup + +# 4.4. Ablations + +**Curvature regularization.** We ablate the necessity of curvature regularization in Neuralangelo and compare the results in Fig. 7(a). Intuitively, $\mathcal{L}_{\mathrm{curv}}$ acts as a smoothness prior by minimizing surface curvatures. Without $\mathcal{L}_{\mathrm{curv}}$ , we find that the surfaces tend to have undesirable sharp transitions. By using $\mathcal{L}_{\mathrm{curv}}$ , the surface noises are removed. + +Topology warmup. We follow prior work and initialize the SDF approximately as a sphere [42]. With an initial spherical shape, using $\mathcal{L}_{\mathrm{curv}}$ also makes concave shapes difficult to form because $\mathcal{L}_{\mathrm{curv}}$ preserves topology by preventing singularities in curvature. Thus, instead of applying $\mathcal{L}_{\mathrm{curv}}$ from the beginning of the optimization process, we use a short warmup period that linearly increases the curvature loss strength. We find this strategy particularly helpful for concave regions, as shown in Fig. 7(b). + +# 5. Conclusion + +We introduce Neuralangelo, an approach for photogrammetric neural surface reconstruction. The findings of Neuralangelo are simple yet effective: using numerical gradients for higher-order derivatives and a coarse-to-fine optimization strategy. Neuralangelo unlocks the representation power of multi-resolution hash encoding for neural surface reconstruction modeled as SDF. We show that Neuralangelo effectively recovers dense scene structures of both object-centric captures and large-scale indoor/outdoor scenes with extremely high fidelity, enabling detailed large-scale scene reconstruction from RGB videos. Our method currently samples pixels from images randomly without tracking their statistics and errors. Therefore, we use long training iterations to reduce the stochastics and ensure sufficient sampling of details. It is our future work to explore a more efficient sampling strategy to accelerate the training process. + +# References + +[1] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 2 +[2] Rui Chen, Songfang Han, Jing Xu, and Hao Su. Point-based multi-view stereo network. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1538-1547, 2019. 2 +[3] François Darmon, Bénédicte Bascle, Jean-Clément Devaux, Pascal Monasse, and Mathieu Aubry. Improving neural implicit surfaces geometry with patch warping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6260–6269, 2022. 2, 3, 6, 8 +[4] Jeremy S De Bonet and Paul Viola. Poxels: Probabilistic voxelized volume reconstruction. In Proceedings of International Conference on Computer Vision (ICCV), volume 2, 1999. 2 +[5] Qiancheng Fu, Qingshan Xu, Yew-Soon Ong, and Wenbing Tao. Geo-neus: Geometry-consistent neural implicit surfaces learning for multi-view reconstruction. arXiv preprint arXiv:2205.15848, 2022. 2, 3 +[6] Yasutaka Furukawa and Jean Ponce. Accurate, dense, and robust multiview stereopsis. IEEE transactions on pattern analysis and machine intelligence, 32(8):1362-1376, 2009. 1, 2 +[7] Silvano Galliani, Katrin Lasinger, and Konrad Schindler. Massively parallel multiview stereopsis by surface normal diffusion. In Proceedings of the IEEE International Conference on Computer Vision, pages 873-881, 2015. 2 +[8] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. arXiv preprint arXiv:2002.10099, 2020. 2, 3 +[9] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. Cambridge university press, 2003. 1 +[10] Po-Han Huang, Kevin Matzen, Johannes Kopf, Narendra Ahuja, and Jia-Bin Huang. Deepmvs: Learning multi-view stereopsis. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2821-2830, 2018. 2 +[11] Rasmus Jensen, Anders Dahl, George Vogiatzis, Engil Tola, and Henrik Aanæs. Large scale multi-view stereopsis evaluation. In 2014 IEEE Conference on Computer Vision and Pattern Recognition, pages 406-413. IEEE, 2014. 5, 6 +[12] Yue Jiang, Dantong Ji, Zhizhong Han, and Matthias Zwicker. Sdfdiff: Differentiable rendering of signed distance fields for 3d shape optimization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1251-1261, 2020. 2 +[13] Michael Kazhdan, Matthew Bolitho, and Hugues Hoppe. Poisson surface reconstruction. In Proceedings of the fourth Eurographics symposium on Geometry processing, volume 7, 2006. 2 +[14] Michael Kazhdan and Hugues Hoppe. Screened poisson surface reconstruction. ACM Transactions on Graphics (ToG), 32(3):1-13, 2013. 2 + +[15] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG), 36(4):1-13, 2017. 5, 7, 8 +[16] Kiriakos N Kutulakos and Steven M Seitz. A theory of shape by space carving. International journal of computer vision, 38(3):199-218, 2000. 1, 2 +[17] Aldo Laurentini. The visual hull concept for silhouette-based image understanding. IEEE Transactions on pattern analysis and machine intelligence, 16(2):150-162, 1994. 2 +[18] Hai Li, Xingrui Yang, Hongjia Zhai, Yuqian Liu, Hujun Bao, and Guofeng Zhang. Vox-surf: Voxel-based implicit surface representation. IEEE Transactions on Visualization and Computer Graphics, 2022. 2 +[19] Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. Barf: Bundle-adjusting neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5741-5751, 2021. 4 +[20] Chen-Hsuan Lin, Chaoyang Wang, and Simon Lucey. Sdf-srn: Learning signed distance 3d object reconstruction from static images. Advances in Neural Information Processing Systems, 33:11453-11464, 2020. 2 +[21] Bruce D Lucas, Takeo Kanade, et al. An iterative image registration technique with an application to stereo vision, volume 81. Vancouver, 1981. 4 +[22] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European conference on computer vision, pages 405-421. Springer, 2020. 2, 3, 6 +[23] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. arXiv preprint arXiv:2201.05989, 2022. 2, 3 +[24] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2020. 2 +[25] Michael Oechsle, Songyou Peng, and Andreas Geiger. Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5589-5599, 2021. 2 +[26] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865-5874, 2021. 4 +[27] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 6, 8 +[28] Johannes L Schonberger, Enliang Zheng, Jan-Michael Frahm, and Marc Pollefeys. Pixelwise view selection for unstructured multi-view stereo. In European conference on computer vision, pages 501-518. Springer, 2016. 2 + +[29] Steven M Seitz and Charles R Dyer. Photorealistic scene reconstruction by voxel coloring. International Journal of Computer Vision, 35(2):151-173, 1999. 1, 2 +[30] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5459-5469, 2022. 2, 3 +[31] Jiaming Sun, Xi Chen, Qianqian Wang, Zhengqi Li, Hadar Averbuch-Elor, Xiaowei Zhou, and Noah Snively. Neural 3d reconstruction in the wild. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-9, 2022. 3 +[32] Richard Szeliski. Rapid octree construction from image sequences. CVGIP: Image understanding, 58(1):23-32, 1993. 2 +[33] Towaki Takikawa, Joey Litalien, Kangxue Yin, Karsten Kreis, Charles Loop, Derek Nowrouzezahrai, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Neural geometric level of detail: Real-time rendering with implicit 3d shapes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11358-11367, 2021. 3, 6 +[34] Engin Tola, Christoph Strecha, and Pascal Fua. Efficient large-scale multi-view stereo for ultra high-resolution image sets. Machine Vision and Applications, 23(5):903-920, 2012. 1, 2 +[35] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T Barron, and Pratul P Srinivasan. Ref-nerf: Structured view-dependent appearance for neural radiance fields. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5481-5490. IEEE, 2022. 4 +[36] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021. 2, 3, 4, 6, 8 +[37] Yiming Wang, Qin Han, Marc Habermann, Kostas Daniilidis, Christian Theobalt, and Lingjie Liu. Neus2: Fast learning of neural implicit surfaces for multi-view reconstruction. arXiv preprint arXiv:2212.05231, 2022. 2 +[38] Yiqun Wang, Ivan Skorokhodov, and Peter Wonka. Hf-neus: Improved surface reconstruction using high-frequency details. In Advances in Neural Information Processing Systems. 3, 6 +[39] Tong Wu, Jiaqi Wang, Xingang Pan, Xudong Xu, Christian Theobalt, Ziwei Liu, and Dahua Lin. Voxurf: Voxel-based efficient and accurate neural surface reconstruction. arXiv preprint arXiv:2208.12697, 2022. 3 +[40] Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. Mvsnet: Depth inference for unstructured multi-view stereo. In Proceedings of the European conference on computer vision (ECCV), pages 767-783, 2018. 2 +[41] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems, 34:4805-4815, 2021. 2, 3, 4, 6 +[42] Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Basri Ronen, and Yaron Lipman. Multiview neural surface reconstruction by disentangling geometry and appearance. Advances in Neural Information Processing Systems, 33:2492-2502, 2020. 2, 4, 8 + +[43] Alex Yu, Sara Fridovich-Keil, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. arXiv preprint arXiv:2112.05131, 2021. 2, 3 +[44] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. arXiv preprint arXiv:2206.00665, 2022. 2, 3 +[45] Jingyang Zhang, Yao Yao, Shiwei Li, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. Critical regularizations for neural surface reconstruction in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6270-6279, 2022. 2, 3, 6 +[46] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492, 2020. 2, 6 +[47] Xiuming Zhang, Pratul P Srinivasan, Boyang Deng, Paul Debevec, William T Freeman, and Jonathan T Barron. Nerfactor: Neural factorization of shape and reflectance under an unknown illumination. ACM Transactions on Graphics (TOG), 40(6):1-18, 2021. 4 \ No newline at end of file diff --git a/2023/Neuralangelo_ High-Fidelity Neural Surface Reconstruction/images.zip b/2023/Neuralangelo_ High-Fidelity Neural Surface Reconstruction/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..fea653c5b08dd6b2206a965c10c2f4570c5818de --- /dev/null +++ b/2023/Neuralangelo_ High-Fidelity Neural Surface Reconstruction/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3954461b9afc43bab309508efbb3ab4cc58f1eabbd18c205034d3b6bf76b2c89 +size 965153 diff --git a/2023/Neuralangelo_ High-Fidelity Neural Surface Reconstruction/layout.json b/2023/Neuralangelo_ High-Fidelity Neural Surface Reconstruction/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..52f37fde073598a4e596061dfabf312fc3135a53 --- /dev/null +++ b/2023/Neuralangelo_ High-Fidelity Neural Surface Reconstruction/layout.json @@ -0,0 +1,9172 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 111, + 103, + 482, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 103, + 482, + 121 + ], + "spans": [ + { + "bbox": [ + 111, + 103, + 482, + 121 + ], + "type": "text", + "content": "Neuralangelo: High-Fidelity Neural Surface Reconstruction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 70, + 141, + 522, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 141, + 522, + 206 + ], + "spans": [ + { + "bbox": [ + 70, + 141, + 522, + 206 + ], + "type": "text", + "content": "Zhaoshuo Li" + }, + { + "bbox": [ + 70, + 141, + 522, + 206 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 70, + 141, + 522, + 206 + ], + "type": "text", + "content": " Thomas Müller" + }, + { + "bbox": [ + 70, + 141, + 522, + 206 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 70, + 141, + 522, + 206 + ], + "type": "text", + "content": " Alex Evans" + }, + { + "bbox": [ + 70, + 141, + 522, + 206 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 70, + 141, + 522, + 206 + ], + "type": "text", + "content": " Russell H. Taylor" + }, + { + "bbox": [ + 70, + 141, + 522, + 206 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 70, + 141, + 522, + 206 + ], + "type": "text", + "content": " Mathias Unberath" + }, + { + "bbox": [ + 70, + 141, + 522, + 206 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 70, + 141, + 522, + 206 + ], + "type": "text", + "content": " Ming-Yu Liu" + }, + { + "bbox": [ + 70, + 141, + 522, + 206 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 70, + 141, + 522, + 206 + ], + "type": "text", + "content": " Chen-Hsuan Lin" + }, + { + "bbox": [ + 70, + 141, + 522, + 206 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 70, + 141, + 522, + 206 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 70, + 141, + 522, + 206 + ], + "type": "text", + "content": "NVIDIA Research " + }, + { + "bbox": [ + 70, + 141, + 522, + 206 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 70, + 141, + 522, + 206 + ], + "type": "text", + "content": "Johns Hopkins University https://research.nvidia.com/labs/dir/neuralangelo" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 48, + 222, + 548, + 458 + ], + "blocks": [ + { + "bbox": [ + 48, + 222, + 548, + 458 + ], + "lines": [ + { + "bbox": [ + 48, + 222, + 548, + 458 + ], + "spans": [ + { + "bbox": [ + 48, + 222, + 548, + 458 + ], + "type": "image", + "image_path": "58f396abf90bacdda4f6894b60e1a077eefff62560d389ed557b1b5c4dcc227b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 464, + 546, + 489 + ], + "lines": [ + { + "bbox": [ + 46, + 464, + 546, + 489 + ], + "spans": [ + { + "bbox": [ + 46, + 464, + 546, + 489 + ], + "type": "text", + "content": "Figure 1. We present Neuralangelo, a framework for high-fidelity 3D surface reconstruction from RGB images using neural volume rendering, even without auxiliary data such as segmentation or depth. Shown in the figure is an extracted 3D mesh of a courthouse." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 505, + 192, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 505, + 192, + 517 + ], + "spans": [ + { + "bbox": [ + 143, + 505, + 192, + 517 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 45, + 528, + 290, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 528, + 290, + 708 + ], + "spans": [ + { + "bbox": [ + 45, + 528, + 290, + 708 + ], + "type": "text", + "content": "Neural surface reconstruction has been shown to be powerful for recovering dense 3D surfaces via image-based neural rendering. However, current methods struggle to recover detailed structures of real-world scenes. To address the issue, we present Neuralangelo, which combines the representation power of multi-resolution 3D hash grids with neural surface rendering. Two key ingredients enable our approach: (1) numerical gradients for computing higher-order derivatives as a smoothing operation and (2) coarse-to-fine optimization on the hash grids controlling different levels of details. Even without auxiliary inputs such as depth, Neuralangelo can effectively recover dense 3D surface structures from multi-view images with fidelity significantly surpassing previous methods, enabling detailed large-scale scene reconstruction from RGB video captures." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 307, + 505, + 386, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 505, + 386, + 517 + ], + "spans": [ + { + "bbox": [ + 307, + 505, + 386, + 517 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 529, + 548, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 529, + 548, + 649 + ], + "spans": [ + { + "bbox": [ + 304, + 529, + 548, + 649 + ], + "type": "text", + "content": "3D surface reconstruction aims to recover dense geometric scene structures from multiple images observed at different viewpoints [9]. The recovered surfaces provide structural information useful for many downstream applications, such as 3D asset generation for augmented/virtual/mixed reality or environment mapping for autonomous navigation of robotics. Photogrammetric surface reconstruction using a monocular RGB camera is of particular interest, as it equips users with the capability of casually creating digital twins of the real world using ubiquitous mobile devices." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 653, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 548, + 715 + ], + "type": "text", + "content": "Classically, multi-view stereo algorithms [6, 16, 29, 34] had been the method of choice for sparse 3D reconstruction. An inherent drawback of these algorithms, however, is their inability to handle ambiguous observations, e.g. regions with large areas of homogeneous colors, repetitive texture" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8456" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "type": "text", + "content": "patterns, or strong color variations. This would result in inaccurate reconstructions with noisy or missing surfaces. Recently, neural surface reconstruction methods [36,41,42] have shown great potential in addressing these limitations. This new class of methods uses coordinate-based multi-layer perceptrons (MLPs) to represent the scene as an implicit function, such as occupancy fields [25] or signed distance functions (SDF) [36,41,42]. Leveraging the inherent continuity of MLPs and neural volume rendering [22], these techniques allow the optimized surfaces to meaningfully interpolate between spatial locations, resulting in smooth and complete surface representations." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 217, + 289, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 217, + 289, + 371 + ], + "spans": [ + { + "bbox": [ + 46, + 217, + 289, + 371 + ], + "type": "text", + "content": "Despite the superiority of neural surface reconstruction methods over classical approaches, the recovered fidelity of current methods does not scale well with the capacity of MLPs. Recently, Müller et al. [23] proposed a new scalable representation, referred to as Instant NGP (Neural Graphics Primitives). Instant NGP introduces a hybrid 3D grid structure with a multi-resolution hash encoding and a lightweight MLP that is more expressive with a memory footprint log-linear to the resolution. The proposed hybrid representation greatly increases the representation power of neural fields and has achieved great success at representing very fine-grained details for a wide variety of tasks, such as object shape representation and novel view synthesis problems." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 372, + 289, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 372, + 289, + 563 + ], + "spans": [ + { + "bbox": [ + 46, + 372, + 289, + 563 + ], + "type": "text", + "content": "In this paper, we propose Neuralangelo for high-fidelity surface reconstruction (Fig. 1). Neuralangelo adopts Instant NGP as a neural SDF representation of the underlying 3D scene, optimized from multi-view image observations via neural surface rendering [36]. We present two findings central to fully unlocking the potentials of multi-resolution hash encodings. First, using numerical gradients to compute higher-order derivatives, such as surface normals for the eikonal regularization [8, 12, 20, 42], is critical to stabilizing the optimization. Second, a progressive optimization schedule plays an important role in recovering the structures at different levels of details. We combine these two key ingredients and, via extensive experiments on standard benchmarks and real-world scenes, demonstrate significant improvements over image-based neural surface reconstruction methods in both reconstruction accuracy and view synthesis quality." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 563, + 271, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 563, + 271, + 574 + ], + "spans": [ + { + "bbox": [ + 59, + 563, + 271, + 574 + ], + "type": "text", + "content": "In summary, we present the following contributions:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 586, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 56, + 586, + 286, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 586, + 286, + 620 + ], + "spans": [ + { + "bbox": [ + 56, + 586, + 286, + 620 + ], + "type": "text", + "content": "- We present the Neuralangelo framework to naturally incorporate the representation power of multi-resolution hash encoding [23] into neural SDF representations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 625, + 286, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 286, + 673 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 286, + 673 + ], + "type": "text", + "content": "- We present two simple techniques to improve the quality of hash-encoded surface reconstruction: higher-order derivatives with numerical gradients and coarse-to-fine optimization with a progressive level of details." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 677, + 287, + 713 + ], + "type": "text", + "content": "- We empirically demonstrate the effectiveness of Neuralangelo on various datasets, showing significant improvements over previous methods." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 71, + 390, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 390, + 83 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 390, + 83 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 95, + 547, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 95, + 547, + 202 + ], + "spans": [ + { + "bbox": [ + 304, + 95, + 547, + 202 + ], + "type": "text", + "content": "Multi-view surface reconstruction. Early image-based photogrammetry techniques use a volumetric occupancy grid to represent the scene [4, 16, 17, 29, 32]. Each voxel is visited and marked occupied if strict color constancy between the corresponding projected image pixels is satisfied. The photometric consistency assumption typically fails due to autoexposure or non-Lambertian materials, which are ubiquitous in the real world. Relaxing such color constancy constraints across views is important for realistic 3D reconstruction." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 203, + 547, + 311 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 203, + 547, + 311 + ], + "spans": [ + { + "bbox": [ + 304, + 203, + 547, + 311 + ], + "type": "text", + "content": "Follow-up methods typically start with 3D point clouds from multi-view stereo techniques [6,7,28,34] and then perform dense surface reconstruction [13,14]. Reliance on the quality of the generated point clouds often leads to missing or noisy surfaces. Recent learning-based approaches augment the point cloud generation process with learned image features and cost volume construction [2,10,40]. However, these approaches are inherently limited by the resolution of the cost volume and fail to recover geometric details." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 315, + 547, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 315, + 547, + 483 + ], + "spans": [ + { + "bbox": [ + 304, + 315, + 547, + 483 + ], + "type": "text", + "content": "Neural Radiance Fields (NeRF). NeRF [22] achieves remarkable photorealistic view synthesis with view-dependent effects. NeRF encodes 3D scenes with an MLP mapping 3D spatial locations to color and volume density. These predictions are composited into pixel colors using neural volume rendering. A problem of NeRF and its variants [1,30,43,46], however, is the question of how an isosurface of the volume density could be defined to represent the underlying 3D geometry. Current practice often relies on heuristic thresholding on the density values; due to insufficient constraints on the level sets, however, such surfaces are often noisy and may not model the scene structures accurately [36,41]. Therefore, more direct modeling of surfaces is preferred for photogrammetric surface reconstruction problems." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 487, + 547, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 487, + 547, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 487, + 547, + 594 + ], + "type": "text", + "content": "Neural surface reconstruction. For scene representations with better-defined 3D surfaces, implicit functions such as occupancy grids [24, 25] or SDFs [42] are preferred over simple volume density fields. To integrate with neural volume rendering [22], different techniques [36, 41] have been proposed to reparametrize the underlying representations back to volume density. These designs of neural implicit functions enable more accurate surface prediction with view synthesis capabilities of unsacrificed quality [42]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "type": "text", + "content": "Follow-up works extend the above approaches to real-time at the cost of surface fidelity [18, 37], while others [3,5,44] use auxiliary information to enhance the reconstruction results. Notably, NeuralWarp [3] uses patch warping given co-visibility information from structure-from-motion (SfM) to guide surface optimization, but the patchwise planar assumption fails to capture highly-varying surfaces [3]. Other methods [5,45] utilize sparse point clouds from SfM to supervise the SDF, but their performances are upper-bounded by the quality of the point clouds, as with" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8457" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": "classical approaches [45]. The use of monocular depth and segmentation as auxiliary data has also been explored with unconstrained image collections [31] or using scene representations with hash encodings [44]. In contrast, our work Neuralangelo builds upon hash encodings [23] to recover surfaces but without the need for auxiliary inputs used in prior work [3, 5, 31, 44, 45]. Concurrent work [38] also proposes coarse-to-fine optimization for improved surface details, where a displacement network corrects the shape predicted by a coarse network. In contrast, we use hierarchical hash grids and control the level of details based on our analysis of higher-order derivatives." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 226, + 113, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 226, + 113, + 239 + ], + "spans": [ + { + "bbox": [ + 47, + 226, + 113, + 239 + ], + "type": "text", + "content": "3. Approach" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 246, + 287, + 318 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 246, + 287, + 318 + ], + "spans": [ + { + "bbox": [ + 46, + 246, + 287, + 318 + ], + "type": "text", + "content": "Neuralangelo reconstructs dense structures of the scene from multi-view images. Neuralangelo samples 3D locations along camera view directions and uses a multi-resolution hash encoding to encode the positions. The encoded features are input to an SDF MLP and a color MLP to composite images using SDF-based volume rendering." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 325, + 133, + 336 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 325, + 133, + 336 + ], + "spans": [ + { + "bbox": [ + 47, + 325, + 133, + 336 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 343, + 287, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 343, + 287, + 439 + ], + "spans": [ + { + "bbox": [ + 46, + 343, + 287, + 439 + ], + "type": "text", + "content": "Neural volume rendering. NeRF [22] represents a 3D scene as volume density and color fields. Given a posed camera and a ray direction, the volume rendering scheme integrates the color radiance of sampled points along the ray. The " + }, + { + "bbox": [ + 46, + 343, + 287, + 439 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 343, + 287, + 439 + ], + "type": "text", + "content": "-th sampled 3D position " + }, + { + "bbox": [ + 46, + 343, + 287, + 439 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 46, + 343, + 287, + 439 + ], + "type": "text", + "content": " is at a distance " + }, + { + "bbox": [ + 46, + 343, + 287, + 439 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 46, + 343, + 287, + 439 + ], + "type": "text", + "content": " from the camera center. The volume density " + }, + { + "bbox": [ + 46, + 343, + 287, + 439 + ], + "type": "inline_equation", + "content": "\\sigma_i" + }, + { + "bbox": [ + 46, + 343, + 287, + 439 + ], + "type": "text", + "content": " and color " + }, + { + "bbox": [ + 46, + 343, + 287, + 439 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_i" + }, + { + "bbox": [ + 46, + 343, + 287, + 439 + ], + "type": "text", + "content": " of each sampled point are predicted using a coordinate MLP. The rendered color of a given pixel is approximated as the Riemann sum:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 83, + 447, + 287, + 480 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 447, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 83, + 447, + 287, + 480 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {c}} (\\mathbf {o}, \\mathbf {d}) = \\sum_ {i = 1} ^ {N} w _ {i} \\mathbf {c} _ {i}, \\quad \\text {w h e r e} w _ {i} = T _ {i} \\alpha_ {i}. \\tag {1}", + "image_path": "8989341fe4e4ef1cb00117b144641c69c6f51f0be00e1f5e3a65c4d3f3c99077.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 487, + 289, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 487, + 289, + 559 + ], + "spans": [ + { + "bbox": [ + 46, + 487, + 289, + 559 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 46, + 487, + 289, + 559 + ], + "type": "inline_equation", + "content": "\\alpha_{i} = 1 - \\exp (-\\sigma_{i}\\delta_{i})" + }, + { + "bbox": [ + 46, + 487, + 289, + 559 + ], + "type": "text", + "content": " is the opacity of the " + }, + { + "bbox": [ + 46, + 487, + 289, + 559 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 487, + 289, + 559 + ], + "type": "text", + "content": "-th ray segment, " + }, + { + "bbox": [ + 46, + 487, + 289, + 559 + ], + "type": "inline_equation", + "content": "\\delta_{i} = t_{i + 1} - t_{i}" + }, + { + "bbox": [ + 46, + 487, + 289, + 559 + ], + "type": "text", + "content": " is the distance between adjacent samples, and " + }, + { + "bbox": [ + 46, + 487, + 289, + 559 + ], + "type": "inline_equation", + "content": "T_{i} = \\Pi_{j = 1}^{i - 1}(1 - \\alpha_{j})" + }, + { + "bbox": [ + 46, + 487, + 289, + 559 + ], + "type": "text", + "content": " is the accumulated transmittance, indicating the fraction of light that reaches the camera. To supervise the network, a color loss is used between input images " + }, + { + "bbox": [ + 46, + 487, + 289, + 559 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 46, + 487, + 289, + 559 + ], + "type": "text", + "content": " and rendered images " + }, + { + "bbox": [ + 46, + 487, + 289, + 559 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{c}}" + }, + { + "bbox": [ + 46, + 487, + 289, + 559 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 127, + 568, + 287, + 582 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 568, + 287, + 582 + ], + "spans": [ + { + "bbox": [ + 127, + 568, + 287, + 582 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {R G B}} = \\left\\| \\hat {\\mathbf {c}} - \\mathbf {c} \\right\\| _ {1}. \\tag {2}", + "image_path": "0d806c1cd29e74dbf6cd40d3cca53d24150f716ad462ada1455bff0c21940e2c.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 590, + 288, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 590, + 288, + 625 + ], + "spans": [ + { + "bbox": [ + 46, + 590, + 288, + 625 + ], + "type": "text", + "content": "However, surfaces are not clearly defined using such density formulation. Extracting surfaces from density-based representation often leads to noisy and unrealistic results [36, 41]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "type": "text", + "content": "Volume rendering of SDF. One of the most common surface representations is SDF. The surface " + }, + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "type": "text", + "content": " of an SDF can be implicitly represented by its zero-level set, i.e., " + }, + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "type": "inline_equation", + "content": "S = \\{\\mathbf{x} \\in \\mathbb{R}^3 | f(\\mathbf{x}) = 0\\}" + }, + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "type": "inline_equation", + "content": "f(\\mathbf{x})" + }, + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "type": "text", + "content": " is the SDF value. In the context of neural SDFs, Wang et al. [36] proposed to convert volume density predictions in NeRF to SDF representations with a logistic function to allow optimization with neural volume" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": "rendering. Given a 3D point " + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": " and SDF value " + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "inline_equation", + "content": "f(\\mathbf{x}_i)" + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": ", the corresponding opacity value " + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "inline_equation", + "content": "\\alpha_i" + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": " used in Eq. 1 is computed as" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 333, + 104, + 545, + 131 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 104, + 545, + 131 + ], + "spans": [ + { + "bbox": [ + 333, + 104, + 545, + 131 + ], + "type": "interline_equation", + "content": "\\alpha_ {i} = \\max \\left(\\frac {\\Phi_ {s} (f (\\mathbf {x} _ {i})) - \\Phi_ {s} (f (\\mathbf {x} _ {i + 1}))}{\\Phi_ {s} (f (\\mathbf {x} _ {i}))}, 0\\right), \\tag {3}", + "image_path": "b4e9934fb552a9dc6d2fdae04437fb96912b708afd5c6d86939ee010ae54ac20.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 133, + 545, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 133, + 545, + 157 + ], + "spans": [ + { + "bbox": [ + 305, + 133, + 545, + 157 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 133, + 545, + 157 + ], + "type": "inline_equation", + "content": "\\Phi_s" + }, + { + "bbox": [ + 305, + 133, + 545, + 157 + ], + "type": "text", + "content": " is the sigmoid function. In this work, we use the same SDF-based volume rendering formulation [36]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 161, + 547, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 161, + 547, + 232 + ], + "spans": [ + { + "bbox": [ + 305, + 161, + 547, + 232 + ], + "type": "text", + "content": "Multi-resolution hash encoding. Recently, multi-resolution hash encoding proposed by Müller et al. [23] has shown great scalability for neural scene representations, generating fine-grained details for tasks such as novel view synthesis. In Neuralangelo, we adopt the representation power of hash encoding to recover high-fidelity surfaces." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 233, + 547, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 233, + 547, + 353 + ], + "spans": [ + { + "bbox": [ + 305, + 233, + 547, + 353 + ], + "type": "text", + "content": "The hash encoding uses multi-resolution grids, with each grid cell corner mapped to a hash entry. Each hash entry stores the encoding feature. Let " + }, + { + "bbox": [ + 305, + 233, + 547, + 353 + ], + "type": "inline_equation", + "content": "\\{V_1,\\dots,V_L\\}" + }, + { + "bbox": [ + 305, + 233, + 547, + 353 + ], + "type": "text", + "content": " be the set of different spatial grid resolutions. Given an input position " + }, + { + "bbox": [ + 305, + 233, + 547, + 353 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 305, + 233, + 547, + 353 + ], + "type": "text", + "content": ", we map it to the corresponding position at each grid resolution " + }, + { + "bbox": [ + 305, + 233, + 547, + 353 + ], + "type": "inline_equation", + "content": "V_{l}" + }, + { + "bbox": [ + 305, + 233, + 547, + 353 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 305, + 233, + 547, + 353 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{i,l} = \\mathbf{x}_i\\cdot V_l" + }, + { + "bbox": [ + 305, + 233, + 547, + 353 + ], + "type": "text", + "content": ". The feature vector " + }, + { + "bbox": [ + 305, + 233, + 547, + 353 + ], + "type": "inline_equation", + "content": "\\gamma_l(\\mathbf{x}_{i,l})\\in \\mathbb{R}^c" + }, + { + "bbox": [ + 305, + 233, + 547, + 353 + ], + "type": "text", + "content": " given resolution " + }, + { + "bbox": [ + 305, + 233, + 547, + 353 + ], + "type": "inline_equation", + "content": "V_{l}" + }, + { + "bbox": [ + 305, + 233, + 547, + 353 + ], + "type": "text", + "content": " is obtained via trilinear interpolation of hash entries at the grid cell corners. The encoding features across all spatial resolutions are concatenated together, forming a " + }, + { + "bbox": [ + 305, + 233, + 547, + 353 + ], + "type": "inline_equation", + "content": "\\gamma (\\mathbf{x}_i)\\in \\mathbb{R}^{cL}" + }, + { + "bbox": [ + 305, + 233, + 547, + 353 + ], + "type": "text", + "content": " feature vector:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 356, + 358, + 545, + 373 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 358, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 356, + 358, + 545, + 373 + ], + "type": "interline_equation", + "content": "\\gamma \\left(\\mathbf {x} _ {i}\\right) = \\left(\\gamma_ {1} \\left(\\mathbf {x} _ {i, 1}\\right), \\dots , \\gamma_ {L} \\left(\\mathbf {x} _ {i, L}\\right)\\right). \\tag {4}", + "image_path": "0bc39fcb6cd0a7e70353e3457c5917ab3aec2e7fd715fe4867d5263f4faf0220.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 378, + 533, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 378, + 533, + 389 + ], + "spans": [ + { + "bbox": [ + 305, + 378, + 533, + 389 + ], + "type": "text", + "content": "The encoded features are then passed to a shallow MLP." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 390, + 545, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 390, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 390, + 545, + 510 + ], + "type": "text", + "content": "One alternative to hash encoding is sparse voxel structures [30, 33, 39, 43], where each grid corner is uniquely defined without collision. However, volumetric feature grids require hierarchical spatial decomposition (e.g. octrees) to make the parameter count tractable; otherwise, the memory would grow cubically with spatial resolution. Given such hierarchy, finer voxel resolutions by design cannot recover surfaces that are misrepresented by the coarser resolutions [33]. Hash encoding instead assumes no spatial hierarchy and resolves collision automatically based on gradient averaging [23]." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 515, + 488, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 515, + 488, + 528 + ], + "spans": [ + { + "bbox": [ + 306, + 515, + 488, + 528 + ], + "type": "text", + "content": "3.2. Numerical Gradient Computation" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 534, + 547, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 547, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 547, + 605 + ], + "type": "text", + "content": "We show in this section that the analytical gradient w.r.t. position of hash encoding suffers from localities. Therefore, optimization updates only propagate to local hash grids, lacking non-local smoothness. We propose a simple fix to such a locality problem by using numerical gradients. An overview is shown in Fig. 2." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 605, + 547, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 677 + ], + "type": "text", + "content": "A special property of SDF is its differentiability with a gradient of the unit norm. The gradient of SDF satisfies the eikonal equation " + }, + { + "bbox": [ + 304, + 605, + 547, + 677 + ], + "type": "inline_equation", + "content": "\\| \\nabla f(\\mathbf{x})\\| _2 = 1" + }, + { + "bbox": [ + 304, + 605, + 547, + 677 + ], + "type": "text", + "content": " (almost everywhere). To enforce the optimized neural representation to be a valid SDF, the eikonal loss [8] is typically imposed on the SDF predictions:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 356, + 683, + 545, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 683, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 356, + 683, + 545, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {e i k}} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left(\\| \\nabla f (\\mathbf {x} _ {i}) \\| _ {2} - 1\\right) ^ {2}, \\tag {5}", + "image_path": "10b3bdedd77bb4550d8ac1ba2e4d404d22fc18aeae6d296ca3c18c2723b426da.jpg" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8458" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 54, + 71, + 286, + 167 + ], + "blocks": [ + { + "bbox": [ + 54, + 71, + 286, + 167 + ], + "lines": [ + { + "bbox": [ + 54, + 71, + 286, + 167 + ], + "spans": [ + { + "bbox": [ + 54, + 71, + 286, + 167 + ], + "type": "image", + "image_path": "75ebad35cb2bbf8018124cffdf59a378eb3f80743bbce21148a13ad52a882865.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 175, + 288, + 209 + ], + "lines": [ + { + "bbox": [ + 46, + 175, + 288, + 209 + ], + "spans": [ + { + "bbox": [ + 46, + 175, + 288, + 209 + ], + "type": "text", + "content": "Figure 2. Using numerical gradients for higher-order derivatives distributes the back-propagation updates beyond the local hash grid cell, thus becoming a smoothed version of analytical gradients." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 229, + 288, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 229, + 288, + 266 + ], + "spans": [ + { + "bbox": [ + 46, + 229, + 288, + 266 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 229, + 288, + 266 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 229, + 288, + 266 + ], + "type": "text", + "content": " is the total number of sampled points. To allow for end-to-end optimization, a double backward operation on the SDF prediction " + }, + { + "bbox": [ + 46, + 229, + 288, + 266 + ], + "type": "inline_equation", + "content": "f(\\mathbf{x})" + }, + { + "bbox": [ + 46, + 229, + 288, + 266 + ], + "type": "text", + "content": " is required." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 265, + 289, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 265, + 289, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 265, + 289, + 361 + ], + "type": "text", + "content": "The de facto method for computing surface normals of SDFs " + }, + { + "bbox": [ + 46, + 265, + 289, + 361 + ], + "type": "inline_equation", + "content": "\\nabla f(\\mathbf{x})" + }, + { + "bbox": [ + 46, + 265, + 289, + 361 + ], + "type": "text", + "content": " is to use analytical gradients [36,41,42]. Analytical gradients of hash encoding w.r.t. position, however, are not continuous across space under trilinear interpolation. To find the sampling location in a voxel grid, each 3D point " + }, + { + "bbox": [ + 46, + 265, + 289, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 46, + 265, + 289, + 361 + ], + "type": "text", + "content": " would first be scaled by the grid resolution " + }, + { + "bbox": [ + 46, + 265, + 289, + 361 + ], + "type": "inline_equation", + "content": "V_{l}" + }, + { + "bbox": [ + 46, + 265, + 289, + 361 + ], + "type": "text", + "content": ", written as " + }, + { + "bbox": [ + 46, + 265, + 289, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{i,l} = \\mathbf{x}_i\\cdot V_l" + }, + { + "bbox": [ + 46, + 265, + 289, + 361 + ], + "type": "text", + "content": ". Let the coefficient for (tri-)linear interpolation be " + }, + { + "bbox": [ + 46, + 265, + 289, + 361 + ], + "type": "inline_equation", + "content": "\\beta = \\mathbf{x}_{i,l} - \\left\\lfloor \\mathbf{x}_{i,l}\\right\\rfloor" + }, + { + "bbox": [ + 46, + 265, + 289, + 361 + ], + "type": "text", + "content": ". The resulting feature vectors are" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 70, + 369, + 288, + 384 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 369, + 288, + 384 + ], + "spans": [ + { + "bbox": [ + 70, + 369, + 288, + 384 + ], + "type": "interline_equation", + "content": "\\gamma_ {l} \\left(\\mathbf {x} _ {i, l}\\right) = \\gamma_ {l} \\left(\\left\\lfloor \\mathbf {x} _ {i, l} \\right\\rfloor\\right) \\cdot (1 - \\beta) + \\gamma_ {l} \\left(\\left\\lceil \\mathbf {x} _ {i, l} \\right\\rceil\\right) \\cdot \\beta , \\tag {6}", + "image_path": "49c7e53b212f3bf0d2960f4a364859c954397b8ab6c13e47e7c667007cef1537.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 392, + 287, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 392, + 287, + 440 + ], + "spans": [ + { + "bbox": [ + 46, + 392, + 287, + 440 + ], + "type": "text", + "content": "where the rounded position " + }, + { + "bbox": [ + 46, + 392, + 287, + 440 + ], + "type": "inline_equation", + "content": "\\lfloor \\mathbf{x}_{i,l}\\rfloor ,\\lceil \\mathbf{x}_{i,l}\\rceil" + }, + { + "bbox": [ + 46, + 392, + 287, + 440 + ], + "type": "text", + "content": " correspond to the local grid cell corners. We note that rounding operations " + }, + { + "bbox": [ + 46, + 392, + 287, + 440 + ], + "type": "inline_equation", + "content": "\\lfloor \\cdot \\rfloor" + }, + { + "bbox": [ + 46, + 392, + 287, + 440 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 392, + 287, + 440 + ], + "type": "inline_equation", + "content": "\\lceil \\cdot \\rceil" + }, + { + "bbox": [ + 46, + 392, + 287, + 440 + ], + "type": "text", + "content": " are non-differentiable. As a result, the derivative of hash encoding w.r.t. the position can be obtained as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 447, + 287, + 491 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 447, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 56, + 447, + 287, + 491 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\frac {\\partial \\gamma_ {l} (\\mathbf {x} _ {i , l})}{\\partial \\mathbf {x} _ {i}} = \\gamma_ {l} (\\lfloor \\mathbf {x} _ {i, l} \\rfloor) \\cdot (- \\frac {\\partial \\beta}{\\partial \\mathbf {x} _ {i , l}}) + \\gamma_ {l} (\\lceil \\mathbf {x} _ {i, l} \\rceil) \\cdot \\frac {\\partial \\beta}{\\partial \\mathbf {x} _ {i , l}} \\\\ = \\gamma_ {l} (\\left\\lfloor \\mathbf {x} _ {i, l} \\right\\rfloor) \\cdot (- V _ {l}) + \\gamma_ {l} (\\left\\lceil \\mathbf {x} _ {i, l} \\right\\rceil) \\cdot V _ {l}. \\tag {7} \\\\ \\end{array}", + "image_path": "29d52968ccc528ed0b35861ba8f9ef67ee5488728bf34a586abe960f7c319ad7.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 498, + 287, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 641 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 641 + ], + "type": "text", + "content": "The derivative of hash encoding is local, i.e., when " + }, + { + "bbox": [ + 46, + 498, + 287, + 641 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 46, + 498, + 287, + 641 + ], + "type": "text", + "content": " moves across grid cell borders, the corresponding hash entries will be different. Therefore, the eikonal loss defined in Eq. 5 only back-propagates to the locally sampled hash entries, i.e. " + }, + { + "bbox": [ + 46, + 498, + 287, + 641 + ], + "type": "inline_equation", + "content": "\\gamma_l(\\lfloor \\mathbf{x}_{i,l}\\rfloor)" + }, + { + "bbox": [ + 46, + 498, + 287, + 641 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 498, + 287, + 641 + ], + "type": "inline_equation", + "content": "\\gamma_l(\\lceil \\mathbf{x}_{i,l}\\rceil)" + }, + { + "bbox": [ + 46, + 498, + 287, + 641 + ], + "type": "text", + "content": ". When continuous surfaces (e.g. a flat wall) span multiple grid cells, these grid cells should produce coherent surface normals without sudden transitions. To ensure consistency in surface representation, joint optimization of these grid cells is desirable. However, the analytical gradient is limited to local grid cells, unless all corresponding grid cells happen to be sampled and optimized simultaneously. Such sampling is not always guaranteed." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "content": "To overcome the locality of the analytical gradient of hash encoding, we propose to compute the surface normals using numerical gradients. If the step size of the numerical gradient is smaller than the grid size of hash encoding, the numerical gradient would be equivalent to the analytical gradient; otherwise, hash entries of multiple grid cells" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 547, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 227 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 227 + ], + "type": "text", + "content": "would participate in the surface normal computation. Backpropagating through the surface normals thus allows hash entries of multiple grids to receive optimization updates simultaneously. Intuitively, numerical gradients with carefully chosen step sizes can be interpreted as a smoothing operation on the analytical gradient expression. An alternative of normal supervision is a teacher-student curriculum [35, 47], where the predicted noisy normals are driven towards MLP outputs to exploit the smoothness of MLPs. However, analytical gradients from such teacher-student losses still only back-propagate to local grid cells for hash encoding. In contrast, numerical gradients solve the locality issue without the need of additional networks." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 228, + 547, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 228, + 547, + 299 + ], + "spans": [ + { + "bbox": [ + 304, + 228, + 547, + 299 + ], + "type": "text", + "content": "To compute the surface normals using the numerical gradient, additional SDF samples are needed. Given a sampled point " + }, + { + "bbox": [ + 304, + 228, + 547, + 299 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i = (x_i,y_i,z_i)" + }, + { + "bbox": [ + 304, + 228, + 547, + 299 + ], + "type": "text", + "content": ", we additionally sample two points along each axis of the canonical coordinate around " + }, + { + "bbox": [ + 304, + 228, + 547, + 299 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 304, + 228, + 547, + 299 + ], + "type": "text", + "content": " within a vicinity of a step size of " + }, + { + "bbox": [ + 304, + 228, + 547, + 299 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 304, + 228, + 547, + 299 + ], + "type": "text", + "content": ". For example, the " + }, + { + "bbox": [ + 304, + 228, + 547, + 299 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 228, + 547, + 299 + ], + "type": "text", + "content": "-component of the surface normal can be found as" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 328, + 303, + 546, + 328 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 303, + 546, + 328 + ], + "spans": [ + { + "bbox": [ + 328, + 303, + 546, + 328 + ], + "type": "interline_equation", + "content": "\\nabla_ {x} f (\\mathbf {x} _ {i}) = \\frac {f \\left(\\gamma \\left(\\mathbf {x} _ {i} + \\boldsymbol {\\epsilon} _ {x}\\right)\\right) - f \\left(\\gamma \\left(\\mathbf {x} _ {i} - \\boldsymbol {\\epsilon} _ {x}\\right)\\right)}{2 \\epsilon}, \\tag {8}", + "image_path": "4801853145bbc82c14edb1957e349b23eae6c52bfad0e821f296512f8e9437be.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 333, + 545, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 333, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 304, + 333, + 545, + 358 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 333, + 545, + 358 + ], + "type": "inline_equation", + "content": "\\epsilon_{x} = [\\epsilon, 0, 0]" + }, + { + "bbox": [ + 304, + 333, + 545, + 358 + ], + "type": "text", + "content": ". In total, six additional SDF samples are required for numerical surface normal computation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 364, + 462, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 364, + 462, + 376 + ], + "spans": [ + { + "bbox": [ + 305, + 364, + 462, + 376 + ], + "type": "text", + "content": "3.3. Progressive Levels of Details" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 382, + 547, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 382, + 547, + 491 + ], + "spans": [ + { + "bbox": [ + 304, + 382, + 547, + 491 + ], + "type": "text", + "content": "Coarse-to-fine optimization can better shape the loss landscape to avoid falling into false local minima. Such a strategy has found many applications in computer vision, such as image-based registration [19,21,26]. Neuralangelo also adopts a coarse-to-fine optimization scheme to reconstruct the surfaces with progressive levels of details. Using numerical gradients for the higher-order derivatives naturally enables Neuralangelo to perform coarse-to-fine optimization from two perspectives." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 494, + 547, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 494, + 547, + 626 + ], + "spans": [ + { + "bbox": [ + 304, + 494, + 547, + 626 + ], + "type": "text", + "content": "Step size " + }, + { + "bbox": [ + 304, + 494, + 547, + 626 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 304, + 494, + 547, + 626 + ], + "type": "text", + "content": ". As previously discussed, numerical gradients can be interpreted as a smoothing operation where the step size " + }, + { + "bbox": [ + 304, + 494, + 547, + 626 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 304, + 494, + 547, + 626 + ], + "type": "text", + "content": " controls the resolution and the amount of recovered details. Imposing " + }, + { + "bbox": [ + 304, + 494, + 547, + 626 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{eik}}" + }, + { + "bbox": [ + 304, + 494, + 547, + 626 + ], + "type": "text", + "content": " with a larger " + }, + { + "bbox": [ + 304, + 494, + 547, + 626 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 304, + 494, + 547, + 626 + ], + "type": "text", + "content": " for numerical surface normal computation ensures the surface normal is consistent at a larger scale, thus producing consistent and continuous surfaces. On the other hand, imposing " + }, + { + "bbox": [ + 304, + 494, + 547, + 626 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{eik}}" + }, + { + "bbox": [ + 304, + 494, + 547, + 626 + ], + "type": "text", + "content": " with a smaller " + }, + { + "bbox": [ + 304, + 494, + 547, + 626 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 304, + 494, + 547, + 626 + ], + "type": "text", + "content": " affects a smaller region and avoids smoothing details. In practice, we initialize the step size " + }, + { + "bbox": [ + 304, + 494, + 547, + 626 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 304, + 494, + 547, + 626 + ], + "type": "text", + "content": " to the coarsest hash grid size and exponentially decrease it matching different hash grid sizes throughout the optimization process." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 629, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 547, + 714 + ], + "type": "text", + "content": "Hash grid resolution " + }, + { + "bbox": [ + 304, + 629, + 547, + 714 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 304, + 629, + 547, + 714 + ], + "type": "text", + "content": ". If all hash grids are activated from the start of the optimization, to capture geometric details, fine hash grids must first \"unlearn\" from the coarse optimization with large step size " + }, + { + "bbox": [ + 304, + 629, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 304, + 629, + 547, + 714 + ], + "type": "text", + "content": " and \"relearn\" with a smaller " + }, + { + "bbox": [ + 304, + 629, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 304, + 629, + 547, + 714 + ], + "type": "text", + "content": ". If such a process is unsuccessful due to converged optimization, geometric details would be lost. Therefore, we only enable an initial set of coarse hash grids and progressively activate" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8459" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 70, + 547, + 281 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 547, + 281 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 547, + 281 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 547, + 281 + ], + "type": "image", + "image_path": "aa07006df64d5950aea79119972ef9c9246d1e8b40b82cd7a60d66768e6013d0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 288, + 536, + 300 + ], + "lines": [ + { + "bbox": [ + 54, + 288, + 536, + 300 + ], + "spans": [ + { + "bbox": [ + 54, + 288, + 536, + 300 + ], + "type": "text", + "content": "Figure 3. Qualitative comparison on the DTU benchmark [11]. Neuralangelo produces more accurate and higher-fidelity surfaces." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 321, + 288, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 321, + 288, + 381 + ], + "spans": [ + { + "bbox": [ + 46, + 321, + 288, + 381 + ], + "type": "text", + "content": "finer hash grids throughout optimization when " + }, + { + "bbox": [ + 46, + 321, + 288, + 381 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 46, + 321, + 288, + 381 + ], + "type": "text", + "content": " decreases to their spatial size. The relearning process can thus be avoided to better capture the details. In practice, we also apply weight decay over all parameters to avoid single-resolution features dominating the final results." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 387, + 132, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 387, + 132, + 399 + ], + "spans": [ + { + "bbox": [ + 47, + 387, + 132, + 399 + ], + "type": "text", + "content": "3.4. Optimization" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 406, + 288, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 406, + 288, + 489 + ], + "spans": [ + { + "bbox": [ + 46, + 406, + 288, + 489 + ], + "type": "text", + "content": "To further encourage the smoothness of the reconstructed surfaces, we impose a prior by regularizing the mean curvature of SDF. The mean curvature is computed from discrete Laplacian similar to the surface normal computation, otherwise, the second-order analytical gradients of hash encoding are zero everywhere when using trilinear interpolation. The curvature loss " + }, + { + "bbox": [ + 46, + 406, + 288, + 489 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{curv}}" + }, + { + "bbox": [ + 46, + 406, + 288, + 489 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 110, + 495, + 287, + 528 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 495, + 287, + 528 + ], + "spans": [ + { + "bbox": [ + 110, + 495, + 287, + 528 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {c u r v}} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left| \\nabla^ {2} f (\\mathbf {x} _ {i}) \\right|. \\tag {9}", + "image_path": "d91a6d1ced1ada6e860928636fd04ed5a01fd5a6769a7ac42cbadc16efd1e4d3.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 533, + 288, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 533, + 288, + 557 + ], + "spans": [ + { + "bbox": [ + 47, + 533, + 288, + 557 + ], + "type": "text", + "content": "We note that the samples used for the surface normal computation in Eq. 8 are sufficient for curvature computation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 59, + 558, + 280, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 558, + 280, + 569 + ], + "spans": [ + { + "bbox": [ + 59, + 558, + 280, + 569 + ], + "type": "text", + "content": "The total loss is defined as the weighted sum of losses:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 94, + 576, + 287, + 590 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 576, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 94, + 576, + 287, + 590 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\mathrm {R G B}} + w _ {\\mathrm {e i k}} \\mathcal {L} _ {\\mathrm {e i k}} + w _ {\\mathrm {c u r v}} \\mathcal {L} _ {\\mathrm {c u r v}}. \\tag {10}", + "image_path": "2c6be38e89a8fbc9e05e228dd6ed5a8d30789552297a9b14afbabb9058e5f7c3.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 596, + 288, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 596, + 288, + 620 + ], + "spans": [ + { + "bbox": [ + 47, + 596, + 288, + 620 + ], + "type": "text", + "content": "All network parameters, including MLPs and hash encoding, are trained jointly end-to-end." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 629, + 128, + 643 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 629, + 128, + 643 + ], + "spans": [ + { + "bbox": [ + 47, + 629, + 128, + 643 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 653, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 653, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 653, + 288, + 713 + ], + "type": "text", + "content": "Datasets. Following prior work, we conduct experiments on 15 object-centric scenes of the DTU dataset [11]. Each scene has 49 or 64 images captured by a robot-held monocular RGB camera. The ground truth is obtained from a structured-light scanner. We further conduct experiments on 6 scenes" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 321, + 547, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 321, + 547, + 369 + ], + "spans": [ + { + "bbox": [ + 304, + 321, + 547, + 369 + ], + "type": "text", + "content": "of the Tanks and Temples dataset [15], including large-scale indoor/outdoor scenes. Each scene contains 263 to 1107 images captured using a hand-held monocular RGB camera. The ground truth is obtained using a LiDAR sensor." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 373, + 546, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 373, + 546, + 492 + ], + "spans": [ + { + "bbox": [ + 304, + 373, + 546, + 492 + ], + "type": "text", + "content": "Implementation details. Our hash encoding resolution spans " + }, + { + "bbox": [ + 304, + 373, + 546, + 492 + ], + "type": "inline_equation", + "content": "2^{5}" + }, + { + "bbox": [ + 304, + 373, + 546, + 492 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 373, + 546, + 492 + ], + "type": "inline_equation", + "content": "2^{11}" + }, + { + "bbox": [ + 304, + 373, + 546, + 492 + ], + "type": "text", + "content": " with 16 levels. Each hash entry has a channel size of 8. The maximum number of hash entries of each resolution is " + }, + { + "bbox": [ + 304, + 373, + 546, + 492 + ], + "type": "inline_equation", + "content": "2^{22}" + }, + { + "bbox": [ + 304, + 373, + 546, + 492 + ], + "type": "text", + "content": ". We activate 4 and 8 hash resolutions at the beginning of optimization for DTU dataset and Tanks and Temples respectively, due to differences in scene scales. We enable a new hash resolution every 5000 iterations when the step size " + }, + { + "bbox": [ + 304, + 373, + 546, + 492 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 304, + 373, + 546, + 492 + ], + "type": "text", + "content": " equals its grid cell size. For all experiments, we do not utilize auxiliary data such as segmentation or depth during the optimization process." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 496, + 547, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 496, + 547, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 496, + 547, + 533 + ], + "type": "text", + "content": "Evaluation criteria. We report Chamfer distance and F1 score for surface evaluation [11, 15]. We use peak signal-to-noise ratio (PSNR) to report image synthesis qualities." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 540, + 409, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 540, + 409, + 551 + ], + "spans": [ + { + "bbox": [ + 306, + 540, + 409, + 551 + ], + "type": "text", + "content": "4.1. DTU Benchmark" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 558, + 547, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 547, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 547, + 641 + ], + "type": "text", + "content": "We show qualitative results in Fig. 3 and quantitative results in Table 1. On average, Neuralangelo achieves the lowest Chamfer distance and the highest PSNR, even without using auxiliary inputs. The result suggests that Neuralangelo is more generally applicable than prior work when recovering surfaces and synthesizing images, despite not performing best in every individual scene." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "type": "text", + "content": "We further ablate Neuralangelo against the following conditions: 1) AG: analytical gradients, 2) " + }, + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{AG + P}" + }, + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "type": "text", + "content": ": analytical gradients and progressive activating hash resolutions, 3) NG: numerical gradients with varying " + }, + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "type": "text", + "content": ". Fig. 4 shows the results qualitatively. AG produces noisy surfaces, even with hash resolutions progressively activated (AG+P). NG improves" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8460" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 71, + 146, + 178 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 146, + 178 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 146, + 178 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 146, + 178 + ], + "type": "image", + "image_path": "71f810f95a3ffcb08e0c9efc4c5b5fa1872c8b2edfd37c17ab9e23582ed2f384.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 88, + 180, + 107, + 190 + ], + "lines": [ + { + "bbox": [ + 88, + 180, + 107, + 190 + ], + "spans": [ + { + "bbox": [ + 88, + 180, + 107, + 190 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 149, + 72, + 246, + 178 + ], + "blocks": [ + { + "bbox": [ + 149, + 72, + 246, + 178 + ], + "lines": [ + { + "bbox": [ + 149, + 72, + 246, + 178 + ], + "spans": [ + { + "bbox": [ + 149, + 72, + 246, + 178 + ], + "type": "image", + "image_path": "c07bb2e00958992cfbd75cd3c96dec67113af298b84f477b1cb0a9b5137f1255.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 190, + 180, + 203, + 188 + ], + "lines": [ + { + "bbox": [ + 190, + 180, + 203, + 188 + ], + "spans": [ + { + "bbox": [ + 190, + 180, + 203, + 188 + ], + "type": "text", + "content": "AG" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 247, + 72, + 344, + 178 + ], + "blocks": [ + { + "bbox": [ + 247, + 72, + 344, + 178 + ], + "lines": [ + { + "bbox": [ + 247, + 72, + 344, + 178 + ], + "spans": [ + { + "bbox": [ + 247, + 72, + 344, + 178 + ], + "type": "image", + "image_path": "2d79fb8d5d9826cbb9b4feadd0986438bd89feb2f0532af362ec668c4a67bcf7.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 285, + 180, + 307, + 188 + ], + "lines": [ + { + "bbox": [ + 285, + 180, + 307, + 188 + ], + "spans": [ + { + "bbox": [ + 285, + 180, + 307, + 188 + ], + "type": "text", + "content": "AG+P" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 198, + 547, + 232 + ], + "lines": [ + { + "bbox": [ + 46, + 198, + 547, + 232 + ], + "spans": [ + { + "bbox": [ + 46, + 198, + 547, + 232 + ], + "type": "text", + "content": "Figure 4. Qualitative comparison of different coarse-to-fine optimization scheme. When using the analytical gradient (AG and " + }, + { + "bbox": [ + 46, + 198, + 547, + 232 + ], + "type": "inline_equation", + "content": "\\mathrm{AG + P}" + }, + { + "bbox": [ + 46, + 198, + 547, + 232 + ], + "type": "text", + "content": "), coarse surfaces often contain artifacts. While using numerical gradients (NG) leads to a better coarse shape, details are also smoothed. Our solution " + }, + { + "bbox": [ + 46, + 198, + 547, + 232 + ], + "type": "inline_equation", + "content": "(\\mathrm{NG} + \\mathrm{P})" + }, + { + "bbox": [ + 46, + 198, + 547, + 232 + ], + "type": "text", + "content": " produces both smooth surfaces and fine details." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 345, + 72, + 443, + 178 + ], + "blocks": [ + { + "bbox": [ + 345, + 72, + 443, + 178 + ], + "lines": [ + { + "bbox": [ + 345, + 72, + 443, + 178 + ], + "spans": [ + { + "bbox": [ + 345, + 72, + 443, + 178 + ], + "type": "image", + "image_path": "62c4f776a7767f6fb5334eeee9c5c34bf35928499d773f69ca6d7b15240cc68a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 389, + 180, + 402, + 188 + ], + "lines": [ + { + "bbox": [ + 389, + 180, + 402, + 188 + ], + "spans": [ + { + "bbox": [ + 389, + 180, + 402, + 188 + ], + "type": "text", + "content": "NG" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 444, + 72, + 542, + 178 + ], + "blocks": [ + { + "bbox": [ + 444, + 72, + 542, + 178 + ], + "lines": [ + { + "bbox": [ + 444, + 72, + 542, + 178 + ], + "spans": [ + { + "bbox": [ + 444, + 72, + 542, + 178 + ], + "type": "image", + "image_path": "7426cfc657beb8d7e82ba0307ff20f336da469f1e77942152e53152ee1319d51.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 471, + 180, + 517, + 189 + ], + "lines": [ + { + "bbox": [ + 471, + 180, + 517, + 189 + ], + "spans": [ + { + "bbox": [ + 471, + 180, + 517, + 189 + ], + "type": "text", + "content": "NG+P (Ours)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 47, + 243, + 548, + 449 + ], + "blocks": [ + { + "bbox": [ + 47, + 243, + 548, + 449 + ], + "lines": [ + { + "bbox": [ + 47, + 243, + 548, + 449 + ], + "spans": [ + { + "bbox": [ + 47, + 243, + 548, + 449 + ], + "type": "table", + "html": "
243740556365698397105106110114118122Mean
NeRF [22]1.901.601.850.582.281.271.471.672.051.070.882.531.061.150.961.49
VolSDF [41]1.141.260.810.491.250.700.721.291.180.700.661.080.420.610.550.86
NeuS [36]1.001.370.930.431.100.650.571.481.090.830.521.200.350.490.540.84
HF-NeuS [38]0.761.320.700.391.060.630.631.151.120.800.521.220.330.490.500.77
RegSDF [45] †0.601.410.640.431.340.620.600.900.921.020.600.590.300.410.390.72
NeuralWarp [3] †0.490.710.380.380.790.810.821.201.060.680.660.740.410.630.510.68
AG0.671.040.840.391.431.231.111.241.540.850.501.010.370.510.440.88
AG+P0.590.950.460.341.190.700.791.191.370.690.490.930.330.440.440.73
NG0.480.810.430.350.890.710.611.261.060.740.470.790.330.450.430.65
NG+P (Ours)0.370.720.350.350.870.540.531.290.970.730.470.740.320.410.430.61
RegSDF [45] †24.7823.0623.4722.2128.5725.5321.8128.8926.8127.9124.7125.1326.8421.6728.2525.31
NeuS [36]26.6223.6426.4325.5930.6132.8329.2433.7126.8531.9732.1828.9228.4135.0034.8129.79
VolSDF [41]26.2825.6126.5526.7631.5731.5029.3833.2328.0332.1333.1631.4930.3334.9034.7530.38
NeRF [22]26.2425.7426.7927.5731.9631.5029.5832.7828.3532.0833.4931.5431.0035.5935.5130.65
AG29.9724.9823.1130.2730.6031.2729.2734.2227.4733.0933.8529.9829.4135.6935.1130.55
AG+P30.1224.6329.5930.2931.6032.0429.8534.1927.8233.2333.9529.1529.4435.9935.6731.17
NG30.3425.1430.2030.7931.7231.8629.8134.3628.0133.4534.3830.3929.8836.0235.7431.47
NG+P (Ours)30.6427.7832.7034.1835.1535.8931.4736.8230.1335.9236.6132.6031.2038.4138.0533.84
", + "image_path": "4e3cd3538bd55c1024bf40961aefba9a884356b3b56d0e0ce296fd4697a2cc8a.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 457, + 547, + 480 + ], + "lines": [ + { + "bbox": [ + 46, + 457, + 547, + 480 + ], + "spans": [ + { + "bbox": [ + 46, + 457, + 547, + 480 + ], + "type": "text", + "content": "Table 1. Quantitative results on DTU dataset [11]. Neuralangelo achieves the best reconstruction accuracy and image synthesis quality. Best result. Second best result. " + }, + { + "bbox": [ + 46, + 457, + 547, + 480 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 46, + 457, + 547, + 480 + ], + "type": "text", + "content": " Requires 3D points from SfM. Best viewed in color." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 501, + 287, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 501, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 46, + 501, + 287, + 525 + ], + "type": "text", + "content": "the smoothness of the surface, sacrificing details. Our setup " + }, + { + "bbox": [ + 46, + 501, + 287, + 525 + ], + "type": "inline_equation", + "content": "(\\mathrm{NG} + \\mathrm{P})" + }, + { + "bbox": [ + 46, + 501, + 287, + 525 + ], + "type": "text", + "content": " produces both smooth surfaces and fine details." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 536, + 160, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 536, + 160, + 550 + ], + "spans": [ + { + "bbox": [ + 47, + 536, + 160, + 550 + ], + "type": "text", + "content": "4.2. Tanks and Temples" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 556, + 288, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 556, + 288, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 556, + 288, + 628 + ], + "type": "text", + "content": "As no public result is available for Tanks and Temples, we train NeuS [36] and NeuralWarp [3] following our setup. We also report classical multi-view stereo results using COLMAP [27]. As COLMAP and NeuralWarp do not support view synthesis, we only report PSNR from NeuS. Results are summarized in Fig. 5 and Table 2." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "content": "Neuralangelo achieves the highest average PSNR and performs best in terms of F1 score. Comparing against NeuS [36], we can recover high-fidelity surfaces with intricate details. We find that the dense surfaces generated from COLMAP are sensitive to outliers in the sparse point cloud. We also find that NeuralWarp often predicts surfaces for the sky and backgrounds potentially due to their color rendering" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 501, + 547, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 501, + 547, + 548 + ], + "spans": [ + { + "bbox": [ + 304, + 501, + 547, + 548 + ], + "type": "text", + "content": "scheme following VolSDF [41]. The additional surfaces predicted for backgrounds are counted as outliers and worsen F1 scores significantly. We instead follow NeuS [36] and use an additional network [46] to model the background." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 550, + 547, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 550, + 547, + 622 + ], + "spans": [ + { + "bbox": [ + 304, + 550, + 547, + 622 + ], + "type": "text", + "content": "Similar to the DTU results, using the analytical gradient produces noisy surfaces and thus leads to a low F1 score. We further note that the reconstruction of Courthouse shown in Figs. 1 and 5 are the same building of different sides, demonstrating the capability of Neuralangelo for large-scale granular reconstruction." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 633, + 400, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 633, + 400, + 645 + ], + "spans": [ + { + "bbox": [ + 306, + 633, + 400, + 645 + ], + "type": "text", + "content": "4.3. Level of Details" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 654, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 714 + ], + "type": "text", + "content": "As Neuralangelo progressively optimizes the hash features of increasing resolution, we inspect the progressive level of details similar to NGLOD [33]. We show a qualitative visualization in Fig. 6. While some surfaces are entirely missed by coarse levels, for example, the tree, table, and" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "8461" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 74, + 220, + 185 + ], + "blocks": [ + { + "bbox": [ + 50, + 74, + 220, + 185 + ], + "lines": [ + { + "bbox": [ + 50, + 74, + 220, + 185 + ], + "spans": [ + { + "bbox": [ + 50, + 74, + 220, + 185 + ], + "type": "image", + "image_path": "67c4bce5337f42302fa3cd96147fca72e45ebcc21b7b0a5158358f407eadd2bc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 225, + 74, + 380, + 185 + ], + "blocks": [ + { + "bbox": [ + 225, + 74, + 380, + 185 + ], + "lines": [ + { + "bbox": [ + 225, + 74, + 380, + 185 + ], + "spans": [ + { + "bbox": [ + 225, + 74, + 380, + 185 + ], + "type": "image", + "image_path": "e38973828cda37f01f6bc7ee658681a5d63e56e5039c168136126a2992c6c58f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 386, + 74, + 541, + 185 + ], + "blocks": [ + { + "bbox": [ + 386, + 74, + 541, + 185 + ], + "lines": [ + { + "bbox": [ + 386, + 74, + 541, + 185 + ], + "spans": [ + { + "bbox": [ + 386, + 74, + 541, + 185 + ], + "type": "image", + "image_path": "a10bc5bbf58fb5d5eba7c874bbeace98b88df4b599161990a6156dee0f29ba12.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 49, + 191, + 219, + 289 + ], + "blocks": [ + { + "bbox": [ + 49, + 191, + 219, + 289 + ], + "lines": [ + { + "bbox": [ + 49, + 191, + 219, + 289 + ], + "spans": [ + { + "bbox": [ + 49, + 191, + 219, + 289 + ], + "type": "image", + "image_path": "2da33cdffd118fcfb89d09e0be0e270fc28d477ba949b1338aa165010089f9f7.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 225, + 190, + 380, + 288 + ], + "blocks": [ + { + "bbox": [ + 225, + 190, + 380, + 288 + ], + "lines": [ + { + "bbox": [ + 225, + 190, + 380, + 288 + ], + "spans": [ + { + "bbox": [ + 225, + 190, + 380, + 288 + ], + "type": "image", + "image_path": "a18f9fc0ba2dbcc2450c94cd56489696111aeb083b95c806dfd8af9f0511eb07.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 386, + 190, + 541, + 288 + ], + "blocks": [ + { + "bbox": [ + 386, + 190, + 541, + 288 + ], + "lines": [ + { + "bbox": [ + 386, + 190, + 541, + 288 + ], + "spans": [ + { + "bbox": [ + 386, + 190, + 541, + 288 + ], + "type": "image", + "image_path": "69a68951d6161edd18924e59ec2a63ee06b8dc50ae51c752ae55e019c49d2381.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 50, + 297, + 219, + 392 + ], + "blocks": [ + { + "bbox": [ + 50, + 297, + 219, + 392 + ], + "lines": [ + { + "bbox": [ + 50, + 297, + 219, + 392 + ], + "spans": [ + { + "bbox": [ + 50, + 297, + 219, + 392 + ], + "type": "image", + "image_path": "6ecb24f0630b1554cbead62a326a6d539c633717e5bda433db8f5f1d05179f28.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 225, + 293, + 380, + 392 + ], + "blocks": [ + { + "bbox": [ + 225, + 293, + 380, + 392 + ], + "lines": [ + { + "bbox": [ + 225, + 293, + 380, + 392 + ], + "spans": [ + { + "bbox": [ + 225, + 293, + 380, + 392 + ], + "type": "image", + "image_path": "1615ec4c42d2595472e0ed0a2981383050c0d11bab0bc47f8caf6c5dcea8df2b.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 386, + 293, + 541, + 392 + ], + "blocks": [ + { + "bbox": [ + 386, + 293, + 541, + 392 + ], + "lines": [ + { + "bbox": [ + 386, + 293, + 541, + 392 + ], + "spans": [ + { + "bbox": [ + 386, + 293, + 541, + 392 + ], + "type": "image", + "image_path": "f8f2d5e4bde5788e86c434f140b05612665682cb61f2e7bae950d8ce46a89230.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 50, + 397, + 219, + 495 + ], + "blocks": [ + { + "bbox": [ + 50, + 397, + 219, + 495 + ], + "lines": [ + { + "bbox": [ + 50, + 397, + 219, + 495 + ], + "spans": [ + { + "bbox": [ + 50, + 397, + 219, + 495 + ], + "type": "image", + "image_path": "ea8b9fda6fa086f86c622b4ed0462512b09b44dbe8e85d9e763c817a4dd691d1.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 225, + 397, + 380, + 495 + ], + "blocks": [ + { + "bbox": [ + 225, + 397, + 380, + 495 + ], + "lines": [ + { + "bbox": [ + 225, + 397, + 380, + 495 + ], + "spans": [ + { + "bbox": [ + 225, + 397, + 380, + 495 + ], + "type": "image", + "image_path": "0b4f59f39ea594ca8f0ee72166f7e166d053ecb8e17fbbb3c8984210fcf960cf.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 386, + 397, + 541, + 495 + ], + "blocks": [ + { + "bbox": [ + 386, + 397, + 541, + 495 + ], + "lines": [ + { + "bbox": [ + 386, + 397, + 541, + 495 + ], + "spans": [ + { + "bbox": [ + 386, + 397, + 541, + 495 + ], + "type": "image", + "image_path": "eed056bf9dd56dfb4fec284ee73dfc5f353fa5053cdcd6056f3a5673b3aa903d.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 51, + 505, + 219, + 600 + ], + "blocks": [ + { + "bbox": [ + 51, + 505, + 219, + 600 + ], + "lines": [ + { + "bbox": [ + 51, + 505, + 219, + 600 + ], + "spans": [ + { + "bbox": [ + 51, + 505, + 219, + 600 + ], + "type": "image", + "image_path": "1f6d66d561729430cea1aa243bfca2a77ad690457872229c5576df9e8c24456c.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 613, + 547, + 636 + ], + "lines": [ + { + "bbox": [ + 46, + 613, + 547, + 636 + ], + "spans": [ + { + "bbox": [ + 46, + 613, + 547, + 636 + ], + "type": "text", + "content": "Figure 5. Qualitative comparison on Tanks and Temples dataset [15]. Neuralangelo captures the scene details better compared to other baseline approaches, while baseline approaches have missing or noisy surfaces." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 225, + 502, + 380, + 599 + ], + "blocks": [ + { + "bbox": [ + 225, + 502, + 380, + 599 + ], + "lines": [ + { + "bbox": [ + 225, + 502, + 380, + 599 + ], + "spans": [ + { + "bbox": [ + 225, + 502, + 380, + 599 + ], + "type": "image", + "image_path": "502eef194b30ad0652e8292c05c570b693fedeff00956eb8b0c053235db66b76.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 386, + 502, + 541, + 599 + ], + "blocks": [ + { + "bbox": [ + 386, + 502, + 541, + 599 + ], + "lines": [ + { + "bbox": [ + 386, + 502, + 541, + 599 + ], + "spans": [ + { + "bbox": [ + 386, + 502, + 541, + 599 + ], + "type": "image", + "image_path": "eac1f6398a4dffb4cea3d41125fedf42e597e6186b6a5c204b0cd5da1f9cafb5.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 653, + 288, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 689 + ], + "type": "text", + "content": "bike rack, these structures are recovered by finer resolutions successfully. The ability to recover missing surfaces demonstrates the advantages of our spatial hierarchy-free design." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 47, + 689, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 289, + 715 + ], + "type": "text", + "content": "Moreover, we note that flat surfaces are predicted at sufficiently high resolutions (around Level 8 in this example)." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": "Thus, only relying on the continuity of local cells of coarse resolutions is not sufficient to reconstruct large continuous surfaces. The result motivates the use of the numerical gradients for the higher-order derivatives, such that backpropagation is beyond local grid cells." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8462" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 82, + 172, + 137 + ], + "blocks": [ + { + "bbox": [ + 50, + 82, + 172, + 137 + ], + "lines": [ + { + "bbox": [ + 50, + 82, + 172, + 137 + ], + "spans": [ + { + "bbox": [ + 50, + 82, + 172, + 137 + ], + "type": "image", + "image_path": "8320594a831f27d587fc460e0fc72f4ee0071a01f83b88170f9e247dce384a7a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 85, + 139, + 137, + 148 + ], + "lines": [ + { + "bbox": [ + 85, + 139, + 137, + 148 + ], + "spans": [ + { + "bbox": [ + 85, + 139, + 137, + 148 + ], + "type": "text", + "content": "Level 4, " + }, + { + "bbox": [ + 85, + 139, + 137, + 148 + ], + "type": "inline_equation", + "content": "V_{4} = 74" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 173, + 81, + 296, + 137 + ], + "blocks": [ + { + "bbox": [ + 173, + 81, + 296, + 137 + ], + "lines": [ + { + "bbox": [ + 173, + 81, + 296, + 137 + ], + "spans": [ + { + "bbox": [ + 173, + 81, + 296, + 137 + ], + "type": "image", + "image_path": "f47dac0ea1fa0f539c5cd69c83ebf67eea011e9f2803f2efd54a3b0afec40248.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 207, + 139, + 262, + 148 + ], + "lines": [ + { + "bbox": [ + 207, + 139, + 262, + 148 + ], + "spans": [ + { + "bbox": [ + 207, + 139, + 262, + 148 + ], + "type": "text", + "content": "Level " + }, + { + "bbox": [ + 207, + 139, + 262, + 148 + ], + "type": "inline_equation", + "content": "8,V_{8} = 223" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 159, + 547, + 194 + ], + "lines": [ + { + "bbox": [ + 46, + 159, + 547, + 194 + ], + "spans": [ + { + "bbox": [ + 46, + 159, + 547, + 194 + ], + "type": "text", + "content": "Figure 6. Results at different hash resolutions. While some structures, such as the tree, table, and bike rack, are missed at coarse resolutions (Level 4). Finer resolutions can progressively recover these missing surfaces. Flat continuous surfaces also require sufficiently fine resolutions to predict (Level 8). The result motivates the non-local updates when using numerical gradients for higher-order derivatives." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 298, + 81, + 420, + 137 + ], + "blocks": [ + { + "bbox": [ + 298, + 81, + 420, + 137 + ], + "lines": [ + { + "bbox": [ + 298, + 81, + 420, + 137 + ], + "spans": [ + { + "bbox": [ + 298, + 81, + 420, + 137 + ], + "type": "image", + "image_path": "5cbfcbb7ecfd4bd072bd1ef91b0286740808c7fce0d3f2f23a7e9a93865624b5.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 328, + 139, + 390, + 148 + ], + "lines": [ + { + "bbox": [ + 328, + 139, + 390, + 148 + ], + "spans": [ + { + "bbox": [ + 328, + 139, + 390, + 148 + ], + "type": "text", + "content": "Level 12, " + }, + { + "bbox": [ + 328, + 139, + 390, + 148 + ], + "type": "inline_equation", + "content": "V_{12} = 676" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 421, + 81, + 544, + 137 + ], + "blocks": [ + { + "bbox": [ + 421, + 81, + 544, + 137 + ], + "lines": [ + { + "bbox": [ + 421, + 81, + 544, + 137 + ], + "spans": [ + { + "bbox": [ + 421, + 81, + 544, + 137 + ], + "type": "image", + "image_path": "236977ccbd1808e95fa4b1500a10c0c4ef0ede1af924d5764cff7ed12cceed40.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 447, + 139, + 517, + 148 + ], + "lines": [ + { + "bbox": [ + 447, + 139, + 517, + 148 + ], + "spans": [ + { + "bbox": [ + 447, + 139, + 517, + 148 + ], + "type": "text", + "content": "Level 16, " + }, + { + "bbox": [ + 447, + 139, + 517, + 148 + ], + "type": "inline_equation", + "content": "V_{16} = 2048" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 47, + 205, + 545, + 316 + ], + "blocks": [ + { + "bbox": [ + 47, + 205, + 545, + 316 + ], + "lines": [ + { + "bbox": [ + 47, + 205, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 47, + 205, + 545, + 316 + ], + "type": "table", + "html": "
F1 Score ↑PSNR ↑
NeuralWarp [3]COLMAP [27]NeuS [36]AGAG+PNGNG+P (Ours)NeuS [36]AGAG+PNGNG+P (Ours)
Barn0.220.550.290.220.310.630.7026.3626.9126.6926.1428.57
Caterpillar0.180.010.290.230.240.300.3625.2126.0425.1226.1627.81
Courthouse0.080.110.170.080.090.240.2823.5525.4325.6325.0627.23
Ignatius0.020.220.830.720.730.850.8923.2722.6922.7323.7823.67
Meetingroom0.080.190.240.040.050.270.3225.3828.1328.0527.4430.70
Truck0.350.190.450.330.370.440.4823.7123.8923.9522.9925.43
Mean0.150.210.380.270.300.450.5024.5825.5125.3625.2627.24
", + "image_path": "bacea7a5e0d366188a43b65067457396e1766d7d9793a6c1ef7c721424c13599.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 324, + 546, + 346 + ], + "lines": [ + { + "bbox": [ + 46, + 324, + 546, + 346 + ], + "spans": [ + { + "bbox": [ + 46, + 324, + 546, + 346 + ], + "type": "text", + "content": "Table 2. Quantitative results on Tanks and Temples dataset [15]. Neuralangelo achieves the best surface reconstruction quality and performs best on average in terms of image synthesis. Best result. Second best result. Best viewed in color." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 50, + 369, + 126, + 445 + ], + "blocks": [ + { + "bbox": [ + 50, + 369, + 126, + 445 + ], + "lines": [ + { + "bbox": [ + 50, + 369, + 126, + 445 + ], + "spans": [ + { + "bbox": [ + 50, + 369, + 126, + 445 + ], + "type": "image", + "image_path": "1fc4cc79b98d1130719e1ed5f4239b95ccf98c402839a1538b9ee80163530954.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 59, + 446, + 66, + 453 + ], + "lines": [ + { + "bbox": [ + 59, + 446, + 66, + 453 + ], + "spans": [ + { + "bbox": [ + 59, + 446, + 66, + 453 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 129, + 369, + 206, + 445 + ], + "blocks": [ + { + "bbox": [ + 129, + 369, + 206, + 445 + ], + "lines": [ + { + "bbox": [ + 129, + 369, + 206, + 445 + ], + "spans": [ + { + "bbox": [ + 129, + 369, + 206, + 445 + ], + "type": "image", + "image_path": "a138717460f089da41ba1890ec75547940740bdec19e65976076f28dc647f56b.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 154, + 446, + 181, + 454 + ], + "lines": [ + { + "bbox": [ + 154, + 446, + 181, + 454 + ], + "spans": [ + { + "bbox": [ + 154, + 446, + 181, + 454 + ], + "type": "text", + "content": "w/o " + }, + { + "bbox": [ + 154, + 446, + 181, + 454 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{curv}}" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 209, + 369, + 285, + 445 + ], + "blocks": [ + { + "bbox": [ + 209, + 369, + 285, + 445 + ], + "lines": [ + { + "bbox": [ + 209, + 369, + 285, + 445 + ], + "spans": [ + { + "bbox": [ + 209, + 369, + 285, + 445 + ], + "type": "image", + "image_path": "1b25b09ec3140664a0b0fadb43241a8209c40759ade3fb2ef412caa0a23ac69c.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 237, + 446, + 261, + 454 + ], + "lines": [ + { + "bbox": [ + 237, + 446, + 261, + 454 + ], + "spans": [ + { + "bbox": [ + 237, + 446, + 261, + 454 + ], + "type": "text", + "content": "w/ " + }, + { + "bbox": [ + 237, + 446, + 261, + 454 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{curv}}" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 50, + 455, + 127, + 533 + ], + "blocks": [ + { + "bbox": [ + 50, + 455, + 127, + 533 + ], + "lines": [ + { + "bbox": [ + 50, + 455, + 127, + 533 + ], + "spans": [ + { + "bbox": [ + 50, + 455, + 127, + 533 + ], + "type": "image", + "image_path": "0892a03f597996cb264649a124d1a9d95065c6189c580a0d1d2ace4e3682268a.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 58, + 534, + 66, + 541 + ], + "lines": [ + { + "bbox": [ + 58, + 534, + 66, + 541 + ], + "spans": [ + { + "bbox": [ + 58, + 534, + 66, + 541 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 554, + 287, + 587 + ], + "lines": [ + { + "bbox": [ + 46, + 554, + 287, + 587 + ], + "spans": [ + { + "bbox": [ + 46, + 554, + 287, + 587 + ], + "type": "text", + "content": "Figure 7. Ablation results. (a) Surface smoothness improves with curvature regularization " + }, + { + "bbox": [ + 46, + 554, + 287, + 587 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{curv}}" + }, + { + "bbox": [ + 46, + 554, + 287, + 587 + ], + "type": "text", + "content": ". (c) Concave shapes are better formed with topology warmup." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 129, + 455, + 206, + 533 + ], + "blocks": [ + { + "bbox": [ + 129, + 455, + 206, + 533 + ], + "lines": [ + { + "bbox": [ + 129, + 455, + 206, + 533 + ], + "spans": [ + { + "bbox": [ + 129, + 455, + 206, + 533 + ], + "type": "image", + "image_path": "c3c3eb14135ab9e63ca9576391a2e65d113669f7b93eb2e036d0694d0f2f6bb9.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 137, + 534, + 198, + 542 + ], + "lines": [ + { + "bbox": [ + 137, + 534, + 198, + 542 + ], + "spans": [ + { + "bbox": [ + 137, + 534, + 198, + 542 + ], + "type": "text", + "content": "w/o topology warmup" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 209, + 455, + 285, + 533 + ], + "blocks": [ + { + "bbox": [ + 209, + 455, + 285, + 533 + ], + "lines": [ + { + "bbox": [ + 209, + 455, + 285, + 533 + ], + "spans": [ + { + "bbox": [ + 209, + 455, + 285, + 533 + ], + "type": "image", + "image_path": "fac9df24cadf56f3931a617d79f972db2673500a5cf88fc3ee1dacbd78e3efbd.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 220, + 534, + 279, + 542 + ], + "lines": [ + { + "bbox": [ + 220, + 534, + 279, + 542 + ], + "spans": [ + { + "bbox": [ + 220, + 534, + 279, + 542 + ], + "type": "text", + "content": "w/ topology warmup" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "bbox": [ + 47, + 615, + 115, + 626 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 615, + 115, + 626 + ], + "spans": [ + { + "bbox": [ + 47, + 615, + 115, + 626 + ], + "type": "text", + "content": "4.4. Ablations" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 46, + 641, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 713 + ], + "type": "text", + "content": "**Curvature regularization.** We ablate the necessity of curvature regularization in Neuralangelo and compare the results in Fig. 7(a). Intuitively, " + }, + { + "bbox": [ + 46, + 641, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{curv}}" + }, + { + "bbox": [ + 46, + 641, + 288, + 713 + ], + "type": "text", + "content": " acts as a smoothness prior by minimizing surface curvatures. Without " + }, + { + "bbox": [ + 46, + 641, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{curv}}" + }, + { + "bbox": [ + 46, + 641, + 288, + 713 + ], + "type": "text", + "content": ", we find that the surfaces tend to have undesirable sharp transitions. By using " + }, + { + "bbox": [ + 46, + 641, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{curv}}" + }, + { + "bbox": [ + 46, + 641, + 288, + 713 + ], + "type": "text", + "content": ", the surface noises are removed." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 368, + 547, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 368, + 547, + 476 + ], + "spans": [ + { + "bbox": [ + 304, + 368, + 547, + 476 + ], + "type": "text", + "content": "Topology warmup. We follow prior work and initialize the SDF approximately as a sphere [42]. With an initial spherical shape, using " + }, + { + "bbox": [ + 304, + 368, + 547, + 476 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{curv}}" + }, + { + "bbox": [ + 304, + 368, + 547, + 476 + ], + "type": "text", + "content": " also makes concave shapes difficult to form because " + }, + { + "bbox": [ + 304, + 368, + 547, + 476 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{curv}}" + }, + { + "bbox": [ + 304, + 368, + 547, + 476 + ], + "type": "text", + "content": " preserves topology by preventing singularities in curvature. Thus, instead of applying " + }, + { + "bbox": [ + 304, + 368, + 547, + 476 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{curv}}" + }, + { + "bbox": [ + 304, + 368, + 547, + 476 + ], + "type": "text", + "content": " from the beginning of the optimization process, we use a short warmup period that linearly increases the curvature loss strength. We find this strategy particularly helpful for concave regions, as shown in Fig. 7(b)." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 306, + 498, + 378, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 498, + 378, + 510 + ], + "spans": [ + { + "bbox": [ + 306, + 498, + 378, + 510 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 522, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 547, + 713 + ], + "type": "text", + "content": "We introduce Neuralangelo, an approach for photogrammetric neural surface reconstruction. The findings of Neuralangelo are simple yet effective: using numerical gradients for higher-order derivatives and a coarse-to-fine optimization strategy. Neuralangelo unlocks the representation power of multi-resolution hash encoding for neural surface reconstruction modeled as SDF. We show that Neuralangelo effectively recovers dense scene structures of both object-centric captures and large-scale indoor/outdoor scenes with extremely high fidelity, enabling detailed large-scale scene reconstruction from RGB videos. Our method currently samples pixels from images randomly without tracking their statistics and errors. Therefore, we use long training iterations to reduce the stochastics and ensure sufficient sampling of details. It is our future work to explore a more efficient sampling strategy to accelerate the training process." + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8463" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5855–5864, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 146, + 288, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 146, + 288, + 190 + ], + "spans": [ + { + "bbox": [ + 53, + 146, + 288, + 190 + ], + "type": "text", + "content": "[2] Rui Chen, Songfang Han, Jing Xu, and Hao Su. Point-based multi-view stereo network. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1538-1547, 2019. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 191, + 288, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 191, + 288, + 246 + ], + "spans": [ + { + "bbox": [ + 53, + 191, + 288, + 246 + ], + "type": "text", + "content": "[3] François Darmon, Bénédicte Bascle, Jean-Clément Devaux, Pascal Monasse, and Mathieu Aubry. Improving neural implicit surfaces geometry with patch warping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6260–6269, 2022. 2, 3, 6, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 247, + 288, + 279 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 247, + 288, + 279 + ], + "spans": [ + { + "bbox": [ + 53, + 247, + 288, + 279 + ], + "type": "text", + "content": "[4] Jeremy S De Bonet and Paul Viola. Poxels: Probabilistic voxelized volume reconstruction. In Proceedings of International Conference on Computer Vision (ICCV), volume 2, 1999. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 280, + 288, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 280, + 288, + 323 + ], + "spans": [ + { + "bbox": [ + 53, + 280, + 288, + 323 + ], + "type": "text", + "content": "[5] Qiancheng Fu, Qingshan Xu, Yew-Soon Ong, and Wenbing Tao. Geo-neus: Geometry-consistent neural implicit surfaces learning for multi-view reconstruction. arXiv preprint arXiv:2205.15848, 2022. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 324, + 288, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 324, + 288, + 367 + ], + "spans": [ + { + "bbox": [ + 53, + 324, + 288, + 367 + ], + "type": "text", + "content": "[6] Yasutaka Furukawa and Jean Ponce. Accurate, dense, and robust multiview stereopsis. IEEE transactions on pattern analysis and machine intelligence, 32(8):1362-1376, 2009. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 369, + 288, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 369, + 288, + 413 + ], + "spans": [ + { + "bbox": [ + 53, + 369, + 288, + 413 + ], + "type": "text", + "content": "[7] Silvano Galliani, Katrin Lasinger, and Konrad Schindler. Massively parallel multiview stereopsis by surface normal diffusion. In Proceedings of the IEEE International Conference on Computer Vision, pages 873-881, 2015. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 414, + 288, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 414, + 288, + 446 + ], + "spans": [ + { + "bbox": [ + 53, + 414, + 288, + 446 + ], + "type": "text", + "content": "[8] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. arXiv preprint arXiv:2002.10099, 2020. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 447, + 288, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 447, + 288, + 479 + ], + "spans": [ + { + "bbox": [ + 53, + 447, + 288, + 479 + ], + "type": "text", + "content": "[9] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. Cambridge university press, 2003. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 480, + 288, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 480, + 288, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 480, + 288, + 533 + ], + "type": "text", + "content": "[10] Po-Han Huang, Kevin Matzen, Johannes Kopf, Narendra Ahuja, and Jia-Bin Huang. Deepmvs: Learning multi-view stereopsis. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2821-2830, 2018. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 536, + 288, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 288, + 580 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 288, + 580 + ], + "type": "text", + "content": "[11] Rasmus Jensen, Anders Dahl, George Vogiatzis, Engil Tola, and Henrik Aanæs. Large scale multi-view stereopsis evaluation. In 2014 IEEE Conference on Computer Vision and Pattern Recognition, pages 406-413. IEEE, 2014. 5, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 581, + 288, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 581, + 288, + 634 + ], + "spans": [ + { + "bbox": [ + 48, + 581, + 288, + 634 + ], + "type": "text", + "content": "[12] Yue Jiang, Dantong Ji, Zhizhong Han, and Matthias Zwicker. Sdfdiff: Differentiable rendering of signed distance fields for 3d shape optimization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1251-1261, 2020. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 635, + 288, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 288, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 288, + 678 + ], + "type": "text", + "content": "[13] Michael Kazhdan, Matthew Bolitho, and Hugues Hoppe. Poisson surface reconstruction. In Proceedings of the fourth Eurographics symposium on Geometry processing, volume 7, 2006. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 680, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 288, + 712 + ], + "type": "text", + "content": "[14] Michael Kazhdan and Hugues Hoppe. Screened poisson surface reconstruction. ACM Transactions on Graphics (ToG), 32(3):1-13, 2013. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 307, + 73, + 546, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 546, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 546, + 116 + ], + "type": "text", + "content": "[15] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics (ToG), 36(4):1-13, 2017. 5, 7, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 119, + 547, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 119, + 547, + 152 + ], + "spans": [ + { + "bbox": [ + 307, + 119, + 547, + 152 + ], + "type": "text", + "content": "[16] Kiriakos N Kutulakos and Steven M Seitz. A theory of shape by space carving. International journal of computer vision, 38(3):199-218, 2000. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 153, + 546, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 153, + 546, + 186 + ], + "spans": [ + { + "bbox": [ + 307, + 153, + 546, + 186 + ], + "type": "text", + "content": "[17] Aldo Laurentini. The visual hull concept for silhouette-based image understanding. IEEE Transactions on pattern analysis and machine intelligence, 16(2):150-162, 1994. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 189, + 547, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 189, + 547, + 232 + ], + "spans": [ + { + "bbox": [ + 308, + 189, + 547, + 232 + ], + "type": "text", + "content": "[18] Hai Li, Xingrui Yang, Hongjia Zhai, Yuqian Liu, Hujun Bao, and Guofeng Zhang. Vox-surf: Voxel-based implicit surface representation. IEEE Transactions on Visualization and Computer Graphics, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 234, + 546, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 234, + 546, + 278 + ], + "spans": [ + { + "bbox": [ + 308, + 234, + 546, + 278 + ], + "type": "text", + "content": "[19] Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. Barf: Bundle-adjusting neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5741-5751, 2021. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 280, + 547, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 280, + 547, + 323 + ], + "spans": [ + { + "bbox": [ + 308, + 280, + 547, + 323 + ], + "type": "text", + "content": "[20] Chen-Hsuan Lin, Chaoyang Wang, and Simon Lucey. Sdf-srn: Learning signed distance 3d object reconstruction from static images. Advances in Neural Information Processing Systems, 33:11453-11464, 2020. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 326, + 546, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 326, + 546, + 358 + ], + "spans": [ + { + "bbox": [ + 308, + 326, + 546, + 358 + ], + "type": "text", + "content": "[21] Bruce D Lucas, Takeo Kanade, et al. An iterative image registration technique with an application to stereo vision, volume 81. Vancouver, 1981. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 361, + 547, + 416 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 361, + 547, + 416 + ], + "spans": [ + { + "bbox": [ + 308, + 361, + 547, + 416 + ], + "type": "text", + "content": "[22] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European conference on computer vision, pages 405-421. Springer, 2020. 2, 3, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 418, + 547, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 418, + 547, + 461 + ], + "spans": [ + { + "bbox": [ + 308, + 418, + 547, + 461 + ], + "type": "text", + "content": "[23] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. arXiv preprint arXiv:2201.05989, 2022. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 464, + 546, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 464, + 546, + 518 + ], + "spans": [ + { + "bbox": [ + 308, + 464, + 546, + 518 + ], + "type": "text", + "content": "[24] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 520, + 546, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 520, + 546, + 576 + ], + "spans": [ + { + "bbox": [ + 308, + 520, + 546, + 576 + ], + "type": "text", + "content": "[25] Michael Oechsle, Songyou Peng, and Andreas Geiger. Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5589-5599, 2021. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 578, + 546, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 578, + 546, + 632 + ], + "spans": [ + { + "bbox": [ + 308, + 578, + 546, + 632 + ], + "type": "text", + "content": "[26] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865-5874, 2021. 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 634, + 546, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 634, + 546, + 667 + ], + "spans": [ + { + "bbox": [ + 308, + 634, + 546, + 667 + ], + "type": "text", + "content": "[27] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 6, 8" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 669, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 669, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 669, + 547, + 713 + ], + "type": "text", + "content": "[28] Johannes L Schonberger, Enliang Zheng, Jan-Michael Frahm, and Marc Pollefeys. Pixelwise view selection for unstructured multi-view stereo. In European conference on computer vision, pages 501-518. Springer, 2016. 2" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8464" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 106 + ], + "type": "text", + "content": "[29] Steven M Seitz and Charles R Dyer. Photorealistic scene reconstruction by voxel coloring. International Journal of Computer Vision, 35(2):151-173, 1999. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 288, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 288, + 160 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 288, + 160 + ], + "type": "text", + "content": "[30] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5459-5469, 2022. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 161, + 288, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 161, + 288, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 161, + 288, + 205 + ], + "type": "text", + "content": "[31] Jiaming Sun, Xi Chen, Qianqian Wang, Zhengqi Li, Hadar Averbuch-Elor, Xiaowei Zhou, and Noah Snively. Neural 3d reconstruction in the wild. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-9, 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 206, + 288, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 206, + 288, + 237 + ], + "spans": [ + { + "bbox": [ + 48, + 206, + 288, + 237 + ], + "type": "text", + "content": "[32] Richard Szeliski. Rapid octree construction from image sequences. CVGIP: Image understanding, 58(1):23-32, 1993. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 239, + 288, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 239, + 288, + 304 + ], + "spans": [ + { + "bbox": [ + 48, + 239, + 288, + 304 + ], + "type": "text", + "content": "[33] Towaki Takikawa, Joey Litalien, Kangxue Yin, Karsten Kreis, Charles Loop, Derek Nowrouzezahrai, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Neural geometric level of detail: Real-time rendering with implicit 3d shapes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11358-11367, 2021. 3, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 305, + 288, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 305, + 288, + 348 + ], + "spans": [ + { + "bbox": [ + 48, + 305, + 288, + 348 + ], + "type": "text", + "content": "[34] Engin Tola, Christoph Strecha, and Pascal Fua. Efficient large-scale multi-view stereo for ultra high-resolution image sets. Machine Vision and Applications, 23(5):903-920, 2012. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 350, + 288, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 350, + 288, + 403 + ], + "spans": [ + { + "bbox": [ + 48, + 350, + 288, + 403 + ], + "type": "text", + "content": "[35] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T Barron, and Pratul P Srinivasan. Ref-nerf: Structured view-dependent appearance for neural radiance fields. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5481-5490. IEEE, 2022. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 404, + 288, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 404, + 288, + 448 + ], + "spans": [ + { + "bbox": [ + 48, + 404, + 288, + 448 + ], + "type": "text", + "content": "[36] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021. 2, 3, 4, 6, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 449, + 288, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 449, + 288, + 492 + ], + "spans": [ + { + "bbox": [ + 48, + 449, + 288, + 492 + ], + "type": "text", + "content": "[37] Yiming Wang, Qin Han, Marc Habermann, Kostas Daniilidis, Christian Theobalt, and Lingjie Liu. Neus2: Fast learning of neural implicit surfaces for multi-view reconstruction. arXiv preprint arXiv:2212.05231, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 493, + 288, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 493, + 288, + 525 + ], + "spans": [ + { + "bbox": [ + 48, + 493, + 288, + 525 + ], + "type": "text", + "content": "[38] Yiqun Wang, Ivan Skorokhodov, and Peter Wonka. Hf-neus: Improved surface reconstruction using high-frequency details. In Advances in Neural Information Processing Systems. 3, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 526, + 288, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 526, + 288, + 569 + ], + "spans": [ + { + "bbox": [ + 48, + 526, + 288, + 569 + ], + "type": "text", + "content": "[39] Tong Wu, Jiaqi Wang, Xingang Pan, Xudong Xu, Christian Theobalt, Ziwei Liu, and Dahua Lin. Voxurf: Voxel-based efficient and accurate neural surface reconstruction. arXiv preprint arXiv:2208.12697, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 570, + 288, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 570, + 288, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 570, + 288, + 613 + ], + "type": "text", + "content": "[40] Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. Mvsnet: Depth inference for unstructured multi-view stereo. In Proceedings of the European conference on computer vision (ECCV), pages 767-783, 2018. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 614, + 288, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 288, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 288, + 657 + ], + "type": "text", + "content": "[41] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. Advances in Neural Information Processing Systems, 34:4805-4815, 2021. 2, 3, 4, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "type": "text", + "content": "[42] Lior Yariv, Yoni Kasten, Dror Moran, Meirav Galun, Matan Atzmon, Basri Ronen, and Yaron Lipman. Multiview neural surface reconstruction by disentangling geometry and appearance. Advances in Neural Information Processing Systems, 33:2492-2502, 2020. 2, 4, 8" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 547, + 307 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 307, + 72, + 546, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 546, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 546, + 116 + ], + "type": "text", + "content": "[43] Alex Yu, Sara Fridovich-Keil, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. arXiv preprint arXiv:2112.05131, 2021. 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 547, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 547, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 547, + 162 + ], + "type": "text", + "content": "[44] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. arXiv preprint arXiv:2206.00665, 2022. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 163, + 547, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 547, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 547, + 217 + ], + "type": "text", + "content": "[45] Jingyang Zhang, Yao Yao, Shiwei Li, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. Critical regularizations for neural surface reconstruction in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6270-6279, 2022. 2, 3, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 219, + 546, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 219, + 546, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 219, + 546, + 251 + ], + "type": "text", + "content": "[46] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492, 2020. 2, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 252, + 547, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 252, + 547, + 307 + ], + "spans": [ + { + "bbox": [ + 308, + 252, + 547, + 307 + ], + "type": "text", + "content": "[47] Xiuming Zhang, Pratul P Srinivasan, Boyang Deng, Paul Debevec, William T Freeman, and Jonathan T Barron. Nerfactor: Neural factorization of shape and reflectance under an unknown illumination. ACM Transactions on Graphics (TOG), 40(6):1-18, 2021. 4" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8465" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Neuralizer_ General Neuroimage Analysis Without Re-Training/eee09609-bc86-4647-98dc-269f47c8002f_content_list.json b/2023/Neuralizer_ General Neuroimage Analysis Without Re-Training/eee09609-bc86-4647-98dc-269f47c8002f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b7100edaf8ddb1700ec41011f770f49e25344876 --- /dev/null +++ b/2023/Neuralizer_ General Neuroimage Analysis Without Re-Training/eee09609-bc86-4647-98dc-269f47c8002f_content_list.json @@ -0,0 +1,1940 @@ +[ + { + "type": "text", + "text": "Neuralizer: General Neuroimage Analysis without Re-Training", + "text_level": 1, + "bbox": [ + 166, + 130, + 803, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Steffen Czolbe \nUniversity of Copenhagen and MGH \nper.sc@di.ku.dk", + "bbox": [ + 151, + 181, + 444, + 232 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Adrian V. Dalca \nMIT, and MGH, Harvard Medical School \nadalca@mit.edu", + "bbox": [ + 488, + 181, + 816, + 232 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 312, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Neuroimage processing tasks like segmentation, reconstruction, and registration are central to the study of neuroscience. Robust deep learning strategies and architectures used to solve these tasks are often similar. Yet, when presented with a new task or a dataset with different visual characteristics, practitioners most often need to train a new model, or fine-tune an existing one. This is a time-consuming process that poses a substantial barrier for the thousands of neuroscientists and clinical researchers who often lack the resources or machine-learning expertise to train deep learning models. In practice, this leads to a lack of adoption of deep learning, and neuroscience tools being dominated by classical frameworks.", + "bbox": [ + 75, + 301, + 470, + 496 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We introduce Neuralizer, a single model that generalizes to previously unseen neuroimaging tasks and modalities without the need for re-training or fine-tuning. Tasks do not have to be known a priori, and generalization happens in a single forward pass during inference. The model can solve processing tasks across multiple image modalities, acquisition methods, and datasets, and generalize to tasks and modalities it has not been trained on. Our experiments on coronal slices show that when few annotated subjects are available, our multi-task network outperforms task-specific baselines without training on the task.", + "bbox": [ + 75, + 497, + 468, + 662 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 691, + 207, + 708 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Computational methods for the processing and analysis of neuroimages have enabled a deep understanding of the human brain. The field has also led to advanced patient care by facilitating non-invasive methods of diagnosis and treatment. Recent deep learning research promises to substantially increase the accuracy and speed of neuroimaging analysis methods.", + "bbox": [ + 75, + 719, + 468, + 824 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A drawback of most current deep-learning-based approaches is that each model is limited to solving the task it has been trained on, on the type of data it has been trained on. Generalization to new tasks and domains, such as different acquisition protocols or new segmentation, is", + "bbox": [ + 75, + 825, + 468, + 900 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0d6f1df7db5ce816a8997a97cf48f586d95cde4ec34df20cc3098306dcfeb348.jpg", + "image_caption": [ + "Figure 1. Neuralizer can solve a broad range of image processing tasks, including new ones not seen during training, with a single model by conditioning the prediction on a context set of examples. After training on a diverse set of tasks, the model can generalize to new tasks in a single forward pass without re-training or finetuning. The model is highly flexible, requiring no prior definition of the set of tasks, and can be conditioned with context sets of any length." + ], + "image_footnote": [], + "bbox": [ + 501, + 267, + 890, + 474 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "a main barrier to adoption [62]. Performing neuroimaging tasks like segmentation, registration, reconstruction, or motion correction requires different models for each processing step, despite operating on the same input data and methods exhibiting strong similarities in network architecture [13,45,85]. Yet, designing and training models to solve these tasks on each dataset is prohibitively expensive. To train a deep learning model, a dataset needs to be compiled and often manually annotated, and the network, training, and data loading logic needs to be implemented. All these steps generally require machine learning and neuroimaging expertise. In addition, computational resources like specialized graphics processing hardware and software infrastructure needs to be available. These requirements are particularly problematic in clinical research settings due to a high cost of annotation and a lack of machine learning expertise and hardware.", + "bbox": [ + 496, + 643, + 890, + 898 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "6217", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/56d19b5b78f39d78a44fd8928324dd971826bf61499ca0993eaf583905d221ce.jpg", + "image_caption": [ + "Binary Segmentation" + ], + "image_footnote": [], + "bbox": [ + 78, + 114, + 176, + 265 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/aa387042cbb38ad0a6273668c17cbe0fca1f2f4bec1ba4e7c815bd8d8f33c964.jpg", + "image_caption": [ + "Modality Transfer" + ], + "image_footnote": [], + "bbox": [ + 181, + 114, + 277, + 263 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/731ce385d19df72b67f1a1a12be3bb813740d5dd54b50cdb32502882386a7c11.jpg", + "image_caption": [ + "Super Resolution" + ], + "image_footnote": [], + "bbox": [ + 282, + 114, + 380, + 263 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0b946413a57cb4cd20535159782b092b9f31e4c6dfc29d695bf49be468742941.jpg", + "image_caption": [ + "Figure 2. Example neuroimaging tasks and modalities included in our dataset (top: input images, bottom: output images)." + ], + "image_footnote": [], + "bbox": [ + 385, + 114, + 482, + 263 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/5f1ad77f626c725a8e3c83f5188002286511cc7e13419b3bd57fe2407e99be61.jpg", + "image_caption": [ + "Denoising & Bias Correction", + "Inpainting" + ], + "image_footnote": [], + "bbox": [ + 794, + 114, + 890, + 188 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ce281d27c4b8fb1d5a5b82c3e4a2b1806493c3e5efd855a9e36a4600e8be7f9c.jpg", + "image_caption": [ + "Skull Stripping" + ], + "image_footnote": [], + "bbox": [ + 488, + 114, + 583, + 263 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/71ec469392539fb001ae6c6292a29ab52749c354535f3b4befd191f1ded5954b.jpg", + "image_caption": [ + "Motion Correction" + ], + "image_footnote": [], + "bbox": [ + 588, + 114, + 684, + 263 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2c1406656e88e80f756c46e6d0610e1127065af82e500c46459793f007c831ca.jpg", + "image_caption": [ + "Undersampled Reconstruction" + ], + "image_footnote": [], + "bbox": [ + 691, + 114, + 787, + 263 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/05a870d5c09eb569cdf29856d0a5c2fcbf320ea4b4df440c07baa87c25716a7e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 794, + 190, + 890, + 263 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Contribution", + "text_level": 1, + "bbox": [ + 76, + 316, + 181, + 330 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We introduce Neuralizer, a general-purpose neuroimaging model that can solve a broad range of neuroimaging tasks on diverse image modalities (Fig. 2), without the need for task-specific training or fine-tuning. Neuralizer can solve new tasks, unseen during training, using a set of examples of the new task at inference (Fig. 1)", + "bbox": [ + 75, + 340, + 468, + 430 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Neuralizer involves a convolutional architecture (Fig. 3), that takes as input a context set of examples that define the processing task, and thus does not require prior specification of the tasks. The method enables single-pass generalization during inference and can process any number of reference images in a single pass to inform the prediction.", + "bbox": [ + 75, + 431, + 468, + 521 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As a first method tackling task generalization in neuroimaging, we focus on analyzing the capabilities of such system and presenting general insights, and limit our experiments to 2D. We evaluate our model by comparing the single-pass generalization performance to task-specific baselines conditioned on an equivalent amount of data. We find that Neuralizer outperforms the baselines on tasks where $\\leq 32$ labeled examples are available, despite never training on the task. When generalizing to new segmentation protocols, Neuralizer matches the performance of baselines trained directly on the dataset.", + "bbox": [ + 75, + 523, + 468, + 688 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 704, + 218, + 719 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We give a short introduction to neuroimaging tasks, terminology, and methods. We then provide an overview of fundamental methods for adapting a model to multiple domains, including multi-task learning, few-shot learning, fine-tuning, and data synthesis.", + "bbox": [ + 75, + 729, + 468, + 805 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Neuroimage analysis", + "text_level": 1, + "bbox": [ + 76, + 816, + 272, + 832 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Neuroimage analysis employs computational techniques to study the structure and function of the human brain. Common imaging techniques are structural magnetic resonance imaging (MRI), functional MRI, diffusion tensor", + "bbox": [ + 75, + 840, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "imaging (DTI), computed tomography (CT), and Positron emission tomography (PET). Each imaging method can create diverse images with different characteristics and contrasts, which are further diversified depending on the properties of the acquisition site [60, 110], device, protocol, imaging sequence [56], and use of contrast agents [9,36].", + "bbox": [ + 496, + 316, + 890, + 407 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To analyze these images, a variety of processing tasks are most often combined in a processing pipeline. Common processing tasks include anatomical segmentation [12, 18, 22, 23, 31, 82, 98, 103], skull stripping [49, 49, 53, 87, 94, 109], defacing [2, 39], registration [5-7, 10, 20, 28, 45, 47, 54], modality transfer [79, 80, 97], in-painting [41, 68, 69, 78, 107], super-resolution [21, 58, 72, 73, 104], reconstruction, and de-noising [58, 70, 92], bias field removal [35, 59], surface fitting [48] and parcellation [91, 99].", + "bbox": [ + 496, + 409, + 890, + 545 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multiple toolboxes provide a suite of interoperable software components, most implementing classical optimization strategies. Widely used toolboxes include Freesurfer [31], FSL [51, 95, 106], SPM [34], CIVET [3], BrainSuite [90], HCP pipeline [100], and BrainIAK [57]. Deep-learning-based methods are starting to be included because of their improved accuracy and shorter runtime [13, 49]. While these methods provide solutions for common neuroimaging applications, most are limited to a single task and few modalities. Developers need to manually update the pipelines to include new processing tasks and to support a wider variety of image modalities. This process requires extensive technical expertise and computational resources, often not available to the clinical neuroscientists focusing on scientific questions.", + "bbox": [ + 496, + 546, + 890, + 772 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Multi-task learning", + "text_level": 1, + "bbox": [ + 500, + 786, + 686, + 801 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multi-Task Learning (MTL) frameworks solve multiple tasks simultaneously by exploiting similarities between related tasks [17]. MTL can improve performance and reduce computational cost and development time compared to designing task-specific solutions [26, 88]. In neuroimaging, MTL networks were recently proposed for the simultaneous", + "bbox": [ + 496, + 810, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "6218", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "segmentation and classification of brain tumors by training a single network with separate prediction heads associated with the different tasks [25, 38]. This strategy is challenging to scale as the number of tasks increases, requires prior determination of the set of tasks, and importantly does not enable generalization of the model to new tasks. With Neuralizer, we build on these methods to achieve scalable MTL, without the need for multiple network heads, and importantly with the ability to generalize to new tasks and modalities.", + "bbox": [ + 75, + 90, + 472, + 241 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Fine-tuning", + "text_level": 1, + "bbox": [ + 76, + 252, + 205, + 268 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To tackle problems in the limited data scenarios frequent in medical imaging, neural networks can be pre-trained on a related task with high data availability and then fine-tuned for specific tasks. For example, a common approach involves taking a Res-Net [43] trained on ImageNet [24] and fine-tuning part of the network for a new task [50, 55, 101]. For medical imaging, networks pre-trained on large sets of medical images are available [19], and fine-tuning them to new tasks results in shortened training time and higher accuracy [4, 71]. However, fine-tuning also requires machine learning expertise and computational resources, most often not available in clinical research. Additionally, in scenarios with small datasets, fine-tuning models trained on large vision datasets can be harmful [83].", + "bbox": [ + 75, + 275, + 472, + 487 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.4. Few-shot learning", + "text_level": 1, + "bbox": [ + 76, + 496, + 251, + 512 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Few-shot models generate predictions from just a few labeled examples [66, 84, 86, 105], or in the case of zero-shot methods [14], none at all. Many of these methods require training or fine-tuning. In computer vision, several methods pass a query image, along with a set of support images and labels as input to the model [66, 89, 96, 102]. Natural image segmentation methods [67, 112] use single image-label pairs [61, 111] as support or aggregate information from a larger support set [63]. Recent few-shot learning methods in the medical image segmentation setting [11, 29, 30] operate on a specific anatomical region in a single image modality [42, 114]. Similar Prior-Data Fitted Networks (PFNs) are fitted to multiple datasets at once to learn the training and prediction algorithm [77]. During training, this strategy draws a dataset, a set of data points and their labels from it, masks one of the labels and predicts it. The resulting model aims to generalize to new datasets. PFNs have only been applied to low-dimensional and tabular data [46]. Our solution builds on ideas from these methods, but aims to solve a much larger range of diverse image-to-image tasks on neuroimages of many modalities.", + "bbox": [ + 75, + 520, + 472, + 837 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.5. Data augmentation and synthesis", + "text_level": 1, + "bbox": [ + 76, + 847, + 367, + 863 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Data augmentation increases the diversity of training data by augmenting or modifying existing data [85, 113]. It", + "bbox": [ + 76, + 869, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "improves model robustness to input variability that may not be available in the original training data. In neuroimaging, arbitrary image modalities can be simulated by synthesis of images without requiring any real data [13, 16, 45, 49, 93]. In meta-learning, data augmentation can further be used to generate entirely new tasks [15, 65, 108]. We use data augmentations and further expand existing methods by developing rich neuroimaging task augmentations for generalization to unseen neuroimaging tasks.", + "bbox": [ + 496, + 90, + 893, + 227 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Neuralizer", + "text_level": 1, + "bbox": [ + 500, + 241, + 614, + 257 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We introduce Neuralizer, a multi-task model for neuroimage analysis tasks. In this section, we first define the training framework and adaptations necessary to operate on a diverse range of tasks and input types. We then introduce the model architecture, training, and inference strategies.", + "bbox": [ + 496, + 266, + 893, + 343 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Generalizable multi-task model", + "text_level": 1, + "bbox": [ + 500, + 352, + 772, + 367 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let $T$ represent a set of tasks, with a subset of tasks $T_{\\mathrm{seen}}$ seen during training. Each task consists of input-output image pairs $(x_{t}, y_{t})$ from potentially multiple underlying datasets with input and output spaces $x_{t} \\in \\mathcal{X}, y_{t} \\in \\mathcal{Y}$ .", + "bbox": [ + 496, + 375, + 890, + 436 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To enable generalization to unseen tasks, we condition the model on a context set $C_t = \\{(x_{t,i},y_{t,i})\\}_{i=1}^N$ of input-output image pairs passed to the model alongside the prediction task. The context set defines the desired task, and can vary in size $|C_t| = N$ and is re-sampled from the underlying task-datasets for each input. Fig. 1 gives an example for a modality transfer task.", + "bbox": [ + 496, + 436, + 890, + 541 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We employ a neural network $g_{\\theta}(x_t, C_t) = y_t$ with weights $\\theta$ that applies the task defined by context set $C_t$ to the input neuroimage $x_t$ . We optimize the network using supervised training with the loss", + "bbox": [ + 496, + 542, + 890, + 602 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} \\left(T _ {\\text {s e e n}}; \\theta\\right) = \\mathbb {E} _ {t \\in T _ {\\text {s e e n}}} \\left[ \\mathbb {E} _ {\\left(x _ {t}, y _ {t}, C _ {t}\\right)} \\left[ \\mathcal {L} _ {t} \\left(y _ {t}, g _ {\\theta} \\left(x _ {t}, C _ {t}\\right)\\right) \\right] \\right], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 614, + 890, + 645 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_t$ is a task-specific loss function.", + "bbox": [ + 500, + 646, + 769, + 660 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Design for diverse tasks", + "text_level": 1, + "bbox": [ + 500, + 671, + 718, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To process different tasks with a single model, we carefully select the loss function, neuroimage encodings, and generation of the training set for each task type.", + "bbox": [ + 496, + 694, + 890, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Loss function. Neuralizer solves both segmentation tasks (e.g. anatomical segmentation and skull-stripping via a brain mask), more general and image-to-image tasks (e.g. denoising). We use the Soft Dice Loss [76] for segmentation-like tasks, and the pixel-wise Mean Squared Error $MSE(y_{t}, g(x_{t}, C_{t})) = \\frac{1}{2\\sigma^{2}} \\sum_{p} [y_{tp} - g(x_{t}, C_{t})_{p}]^{2}$ with balancing hyperparameter $\\sigma^{2}$ for other tasks. As the network optimizes multiple tasks during training, the balance of the loss terms can dramatically affect the optimization and resulting performance.", + "bbox": [ + 496, + 752, + 893, + 905 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "6219", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f98a6eaa912876ee02c28a05c983027b5f7e96ca4889be678ceb5cc3377b2e0a.jpg", + "image_caption": [ + "Figure 3. Neuralizer consists of 7 Pairwise-Conv-Avg blocks (right), arranged in a U-Net-like [76, 85] configuration (left). Each Pairwise-Conv-Avg block enables interaction between the input image and the image pairs present in the context set. The block consists of a residual unit, pairwise convolution of each context member with the target, and an averaging of results across the context set to update the representation. The architecture is invariant to context size $N$ ." + ], + "image_footnote": [], + "bbox": [ + 76, + 87, + 416, + 287 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/cca2e8bdd235ecf870b8147531fe736c1aeb1d4f8d5074958ed0bc830beffdd5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 421, + 87, + 893, + 286 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Input and output encoding. For Neuralizer to work on both segmentation and image-to-image tasks, we facilitate simultaneous input of multiple image modalities and masks. We design the input space $\\mathcal{X}$ to accept floating point value images with three channels, and zero-pad any channels unnecessary for a specific task. The output space $\\mathcal{Y}$ follows the same design but uses only one channel.", + "bbox": [ + 75, + 378, + 468, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Training dataset creation. At each training iteration, we first sample a task $t$ from $T_{\\mathrm{seen}}$ , selecting the task-specific dataset (Tab. 1). From this dataset, we sample the input image, ground truth output, and image pairs for the context set. To increase the range of images that can be used to condition the trained model, the image modalities and acquisition protocols of entries of the context set can differ from the input image for some tasks. Supplemental section G contains a detailed description of the training data generator.", + "bbox": [ + 75, + 501, + 468, + 637 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Model architecture", + "text_level": 1, + "bbox": [ + 76, + 648, + 261, + 664 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fig. 3 shows the Neuralizer network architecture, adapted with the concurrently developed [15] - a method that focuses on solving broad segmentation tasks. As the architecture is independent of the task, we omit the task subscript in this section.", + "bbox": [ + 75, + 672, + 468, + 747 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The input image $x$ and the image pairs of the context set $C_i = (x_i, y_i), i = 1, \\dots, N$ are first passed through an embedding layer consisting of a single $1 \\times 1$ convolution with learnable kernels $e_x, e_C$ , to obtain the representations $r_x = x * e_x$ , $r_{C_i} = \\operatorname{cat}(x_i, y_i) * e_C$ where $*$ is the convolution operator. This combines each context image pair to a joint representation $r_{C_i}$ and maps all representations to a uniform channel width $c$ , which is constant throughout the model. Next, we process the representations using multiple Pairwise-Conv-Avg Blocks (explained below), ar", + "bbox": [ + 75, + 750, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/710a2ed9cdd57888092773dd27337213128c15bbfcc6915aac2628ad7a718919.jpg", + "table_caption": [ + "Table 1. Tasks, Modalities, Datasets, and Segmentation classes used in this paper, and involved in training Neuralizer." + ], + "table_footnote": [], + "table_body": "
TasksModalities
Binary SegmentationT1-w.
Modality TransferT2-w.
Super ResolutionMRA
Skull StrippingPD
Motion CorrectionFLAIR
Undersampled ReconstructionADC
Denoising & Bias correctionDWI
InpaintingDTI (17 dir.)
DatasetsSegmentation Classes
OASIS 3 [47, 74]Freesufer protocol, 31
BRATS [8, 9, 75]classes [13, 31]
IXI [1]
ATLAS R2.0 [64]Manually-annotated Hammers
Hammers Atlas [40]Atlas, 96 classes [27, 37, 40]
WMH Challenge [56]
ISLES2022 [81]Brainmasks [31, 49]
", + "bbox": [ + 500, + 417, + 890, + 728 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ranged as a U-Net-like configuration [76,85] to exploit multiple scales. The output $r_x^{\\mathrm{out}}$ of the final Pairwise-Conv-Avg Block is processed by a residual unit [43] and a final $1 \\times 1$ conv layer to map to one output channel. All residual units consist of two $3 \\times 3$ conv layers, a shortcut connection, and GELU activation functions [44].", + "bbox": [ + 496, + 761, + 890, + 851 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Compared to standard CNNs, Neuralizer uses a mechanism to enable knowledge transfer from the context set to the input image. We design the Pairwise-Conv-Avg Block", + "bbox": [ + 498, + 854, + 892, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "6220", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "(Fig. 3, right) to model this interaction. The block maps from representations of the target input $r_x^{\\mathrm{in}}$ and context pairs $r_{C_i}^{\\mathrm{in}}$ to output representations $r_x^{\\mathrm{out}}$ , $r_{C_i}^{\\mathrm{out}}$ of the same size. First, we process each input separately with a residual unit to obtain $r_x^{\\mathrm{int}} = \\mathrm{ResUnit}_x(r_x^{\\mathrm{in}})$ and $r_{C_i}^{\\mathrm{int}} = \\mathrm{ResUnit}_C(r_{C_i}^{\\mathrm{in}})$ . The residual units, which involve two convolutions, operate on the context representations and have shared weights. Second, we pairwise concatenate the context representations with the target representation on the channel dimension: $p_i = \\mathrm{cat}(r_x^{\\mathrm{int}}, r_{C_i}^{\\mathrm{int}})$ . We combine the pairwise representations and reduce the channel size back to $c$ using a $1 \\times 1$ convolution with learnable kernel $k_x$ , and update the target representation by averaging across context members $r_x^{\\mathrm{out}} = r_x^{\\mathrm{int}} + \\frac{1}{N}\\sum_{i=1}^{N}p_i * k_x$ . The context representations are updated with a separate kernel $r_{C_i}^{\\mathrm{out}} = r_{C_i}^{\\mathrm{int}} + p_i * k_C$ . We then re-size the outputs of a Pairwise-Conv-Avg Block before feeding them as input the next block. We experimented with attention-based and weighted average approaches but found that they did not lead to an increased generalization to unseen tasks.", + "bbox": [ + 76, + 90, + 472, + 393 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Task augmentations", + "text_level": 1, + "bbox": [ + 76, + 402, + 266, + 417 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To further diversify the training dataset, we employ task augmentations [15], a group of transformations applied at random to the input, output, and context images. The objective is to increase the diversity of tasks to discourage the model from merely memorizing the tasks in the training data. A list of all task augmentations is summarized in Tab. 2, with more detailed descriptions and visual examples in Supplement C.", + "bbox": [ + 75, + 425, + 468, + 546 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Inference", + "text_level": 1, + "bbox": [ + 76, + 556, + 187, + 569 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During inference, we supply an input image $x_{i}$ and a context set $C_i$ from the desired task. Given these inputs, a simple feed-forward pass through the model provides the prediction $\\hat{y} = g(x,C)$ . To further increase accuracy at test-time, we use context-set bootstrapping [15]. We also increase the context set by sampling with replacement from the context set, and add small affine augmentations.", + "bbox": [ + 75, + 579, + 468, + 685 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 698, + 209, + 714 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We first compare Neuralizer with task-specific networks, which require substantial expertise and compute. We then analyze the effect of the size of the context set, and the", + "bbox": [ + 76, + 723, + 468, + 768 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/4f4e7f6a0c2995295d52c931361db2593a7a3ed70b08e7459af5482c1a44326e.jpg", + "table_caption": [ + "Table 2. Task Augmentations" + ], + "table_footnote": [], + "table_body": "
Task Augmentations
IntensityMappingSyntheticModality
SobelFilterMaskInvert
MaskContourMaskDilation
PermuteChannelsDuplicateChannels
", + "bbox": [ + 78, + 816, + 468, + 909 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "multi-task generalization to unseen segmentation protocols and image modalities. For this first method of large-scale multi-task generalization in neuroimaging, we conduct the experiments on 2D image slices.", + "bbox": [ + 496, + 90, + 890, + 151 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Data", + "text_level": 1, + "bbox": [ + 500, + 160, + 573, + 174 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To create a diverse dataset encompassing a multitude of different modalities, acquisition protocols, devices, and tasks, we pool neuroimages from the public datasets OASIS3 [47, 74], BRATS [8, 9, 75], Atlas R2.0 [64], Hammers Atlas [40], IXI [1], ISLES2022 [81], and the White Matter Hyperintensities Challenge [56]. We segment all subjects with Synthseg [13, 31]. Based on the segmentation, we affinely align the images to the MNI 152 template space [32, 33], and resample to $1\\mathrm{mm}$ isometric resolution at a size of $192 \\times 224 \\times 192\\mathrm{mm}$ . We perform manual quality control of the segmentation and registration by ensuring no segmented areas fall outside of the cropped volume and discard subjects failing this check (4 subjects). We extract a coronal slice of $192 \\times 192\\mathrm{mm}$ , bisecting the frontal Brain stem, Hippocampus, Thalamus, and Lateral ventricles. We rescale image intensities to the $[0, 1]$ interval using dataset-specific percentiles. For full head images, we create a brain mask with Synthstrip [31, 49]. The final dataset contains 2,282 subjects with 15,911 images and segmentation masks across 8 modalities. Subjects of the seven original datasets are split into $80\\%$ for training and validation, $20\\%$ test, with a minimum of 15 test subjects per dataset.", + "bbox": [ + 496, + 183, + 893, + 515 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Models", + "text_level": 1, + "bbox": [ + 500, + 523, + 591, + 537 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Neuralizer-seen. This Neuralizer model includes all tasks available during training. We use this model to evaluate the performance on unseen scans from tasks and modalities that have been included in the training. The model uses the 4-stage architecture shown in Fig. 3 with 64 channels per layer. During training, the context size $|C_i|$ is sampled from $\\mathcal{U}_{\\{1,32\\}}$ at each iteration.", + "bbox": [ + 496, + 547, + 890, + 655 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Neuralizer-unseen. To evaluate Neuralizer performance on tasks and modalities it has not been trained on, we train a family of Neuralizer models where a single task or modality is excluded from the training set. The model architecture of Neuralizer-unseen is identical to Neuralizer-seen.", + "bbox": [ + 496, + 671, + 893, + 746 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baseline-seen. As no established baseline for multi-task and multi-modality models in neuroimaging can tackle the number of tasks we aim for, we compare Neuralizer to an ensemble of task-specific U-Nets [76, 85]. However, training one model for each task and modality requires overwhelming computational resources. To reduce the computational requirement, we follow previous modality-agnostic models [13, 49] and train each model on multiple input modalities. This lowers the number of models to be", + "bbox": [ + 496, + 763, + 893, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "6221", + "bbox": [ + 482, + 945, + 513, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c0fd6b8c4ab958a4866fa6e48d308a57f8a7fbfd10c48d98718dc57123b19619.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 87, + 269, + 194 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2fb8191a3e762802546feea73218571c4caa76075a1acd257d07e3c463b4abb4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 287, + 88, + 475, + 193 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/06758ead4dccdb38c713dc105b69754416a18ba9690bc8090a141449b4d41348.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 88, + 684, + 193 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b3174a3fc67059c080ee525a6a4e5a2fd8a44c5325a7bc8c8afa8eae1bf28f61.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 705, + 88, + 892, + 193 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d9c8c77962f286661c55d3d0979620d756a1d7b37b834ea90c78fa357d8d0643.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 78, + 203, + 269, + 311 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/20705a6f1271ea2172df0597c863044496fa17068ac696b8daddf9a8b5d09a51.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 282, + 203, + 475, + 311 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c732eb5110f46f99c6fb1d1f34776829913682328cbb8c87f2661694e0ef1c00.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 491, + 203, + 684, + 311 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ae87f88a2dd257a524a2cb1e347b10913b005c6aea99e1da76a96975419356fc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 700, + 203, + 893, + 311 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ad495076e3e82ba2851a7bb69018eb31854c3e2d26c32a0eb69559d4bc4668cd.jpg", + "image_caption": [ + "Figure 4. Performance of multi-task Neuralizer and the task-specific baselines on each task, averaged across all modalities in the test set. The tasks being evaluated were included in the training of Neuralizer-seen (orange), held out in Neuralizer-unseen (blue), and specifically trained on by each task-specific baseline (gray). The x-axis is the size of the train/context set, and the y-axis is the Dice/PSNR score. Some points on the x-axis are omitted for better visibility. 'All' refers to all available train data for the task, ranging from 249 to 2,282 subjects depending on the task. The bars denote standard deviation across modalities. We extract results for T1 scans in Supplement D." + ], + "image_footnote": [], + "bbox": [ + 76, + 321, + 269, + 428 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/de6954e83fcf785cfb55148edc6831cb10ae20987179d155976a63a0f0f80a1f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 292, + 321, + 475, + 428 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/789c960037e4757f84c39eb1536dca37df9d20e00206bf3bebb3a821046ae5bb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 491, + 321, + 815, + 428 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "trained to one per task, segmentation class, and modalitytransfer output modality. To compare Baseline-seen with Neuralizer-unseen given an equal amount of data, we train baselines with training set sizes of $\\{1,2,4,8,16,32,\\mathrm{all}\\}$ and employ standard data augmentation.", + "bbox": [ + 75, + 534, + 468, + 609 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We use a 4-stage U-Net architecture with one residual block per layer. The channel width is tuned experimentally for each training dataset size. We select 256 channels when all data is available for training, and 64 channels otherwise. Using larger U-Nets resulted in overfitting and lower performance. Supplement H summarizes model parameter counts and inference costs.", + "bbox": [ + 75, + 609, + 468, + 715 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Training", + "text_level": 1, + "bbox": [ + 76, + 724, + 181, + 742 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We use supervised training, task-specific loss functions, and weigh the MSE loss by selecting $\\sigma^2 = 0.05$ , resulting in the loss terms being of similar magnitude. All models are trained with a batch size of 8, a learning rate of $10^{-4}$ , and the ADAM optimizer [52]. To speed up training, we undersample tasks that the model learns quickly, with sampling weights given in Supplement I.", + "bbox": [ + 75, + 750, + 468, + 854 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In addition to the task augmentations, we use data augmentations via random affine movements, random elastic deformations, and random flips along the sagittal plane.", + "bbox": [ + 75, + 854, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/67d6f6e7811437e39ebb21b71eff8e192b68c9aee78c225c8b8526877df1ae3c.jpg", + "image_caption": [ + "Figure 5. Results averaged across tasks, expressed as relative performance compared to the baseline trained on all data. The tasks being evaluated were included in the training of Neuralizer-seen (orange), held out in Neuralizer-unseen (blue), and specifically trained on by each task-specific baseline (gray). The x-axis is the size of the train/context set, and the y-axis is the relative score. Some points on the x-axis are omitted for better visibility. 'All' refers to all available data for the task, ranging from 249 to 2,282 subjects depending on the task. Bars: standard deviation across tasks/modalities." + ], + "image_footnote": [], + "bbox": [ + 498, + 530, + 890, + 720 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6222", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For Baseline-seen, we reuse the augmentations but remove those that introduce uncertainty in the desired output. The training time for Neuralizer is 7 days on a single A100 GPU. The training time of the baseline models is capped at 5 days. All models use early stopping, ending the training after 25 epochs exhibiting no decrease in validation loss. The model with the lowest validation loss is used for further evaluation on the test set.", + "bbox": [ + 75, + 90, + 472, + 210 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Evaluation", + "text_level": 1, + "bbox": [ + 76, + 231, + 197, + 247 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We evaluate the Dice coefficient for the segmentation and skull stripping tasks, and the Peak Signal-to-Noise Ratio (PSNR) for the image-to-image tasks on the test set. As the low-data regime is of particular interest, we measure performance as a function of context set size for the Neuralizer models and use training set size as an analog for the U-Net models. We evaluate context sets of up to 32 subjects. Larger context sets are possible but come at a linear cost in memory.", + "bbox": [ + 75, + 258, + 472, + 395 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Experiment 1: Baseline Comparison", + "text_level": 1, + "bbox": [ + 76, + 412, + 393, + 430 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To assess if the proposed multi-task approach is competitive with task-specific models, we evaluate the performance and runtime of Neuralizer-seen, Neuralizer-unseen, and Baseline-seen on the test-set of each task. For Neuralizer-unseen, we withhold image modalities using a leave-one-out strategy during training and evaluate on the unseen modalities at test time.", + "bbox": [ + 75, + 440, + 468, + 546 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results. We display the results by task, averaged across modalities in Fig. 4. We also provide an evaluation of using just the T1 modality in Supplement D, since many task-specific networks in neuroimage analysis literature focus on T1 images. We further aggregate performance across all tasks in Fig. 5, and provide tabular results in Supplement E. Both Neuralizer models outperform most task-specific baselines trained on up to 32 samples. When training the baselines on all available data, the baselines outperform Neuralizer-seen by 2 percentage points in relative performance, and Neuralizer-unseen by 3 percentage points. The loss in performance when generalizing to an unseen modality (between Neuralizer-seen and Neuralizer-unseen) is less than 2 percentage points for all context set sizes.", + "bbox": [ + 75, + 575, + 468, + 787 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Training the baseline model to convergence on 32 samples took on average $28.2 \\pm 16.6$ hours per task, using one A100 GPU. Since Neuralizer only requires inference for a new task, it is orders of magnitude faster, requiring less than 0.1 seconds on a GPU and less than 3 seconds on a CPU.", + "bbox": [ + 75, + 791, + 468, + 867 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We provide qualitative samples of the predictions from Neuralizer-seen model in Supplement A, Figures 6-8.", + "bbox": [ + 76, + 869, + 468, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.6. Experiment 2: Context set size analysis", + "text_level": 1, + "bbox": [ + 498, + 90, + 836, + 107 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We assess the few-shot setting that is prevalent in neuroimage analysis, where few annotated images are often available for a new task. We evaluate performance as a function of the number of labeled samples. For Neuralizer, we evaluate the model with context-set sizes of $\\{1,2,4,8,16,32\\}$ unique subjects from the test set. For the baseline, we trained models with reduced training set sizes of the same amount of subjects. To reduce the effect of random training subject selection, we train three separate baselines with $n = 1$ , two baselines with $n = 2$ , and average results of models with the same $n$ .", + "bbox": [ + 496, + 113, + 890, + 280 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results. Tab. 4 and Figs. 4, 5 illustrate the results. For all models, prediction accuracy increases with the availability of labeled data, with diminishing returns. For both Neuralizer models, a context set size of one achieves more than $90\\%$ of the performance attainable with all data. For most tasks, the baseline performs overall worse than both Neuralizer models when $\\leq 32$ labeled samples are available but achieves the best overall performance on larger datasets.", + "bbox": [ + 496, + 290, + 890, + 412 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.7. Experiment 3: Generalization to a new segmentation protocol", + "text_level": 1, + "bbox": [ + 498, + 420, + 890, + 452 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The Hammers Atlas dataset [27,37,40] provides an alternative anatomical segmentation protocol to the widely-used Freesurfer segmentation available for most subjects in the dataset. The shape, size, and amount of annotated regions in the protocols differ drastically. A different image acquisition site also leads to differences in visual characteristics. We use the Hammers Atlas dataset to evaluate Neuralizer-unseen by entirely withholding the dataset and its annotations from training. We evaluate the Dice coefficient of the 14 major anatomical segmentation classes present in the center and right half of the coronal slice.", + "bbox": [ + 496, + 458, + 890, + 625 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results. Tab. 3 illustrates the results. Neuralizer-unseen performs similarly to Neuralizer-seen and the baseline, while not requiring lengthy re-training or fine-tuning on the Hammers Atlas dataset, and not having seen the segmentation protocol. All three models achieve a mean Dice coefficient of 0.84. The largest performance difference is in the third ventricle class, where both Neuralizer models outperform the baseline by at least 0.04 Dice. The Freesurfer segmentation protocol included in the training set of the Neuralizer models also contains a third ventricle class.", + "bbox": [ + 496, + 635, + 890, + 787 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Discussion", + "text_level": 1, + "bbox": [ + 500, + 800, + 612, + 815 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our experiments using modality and segmentation class hold-outs show that Neuralizer can generalize well to unseen neuroimaging tasks. Across all context set sizes, the generalization loss between seen and unseen modalities and segmentation classes is less than 2 percentage points across", + "bbox": [ + 496, + 825, + 890, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "6223", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2dc92878a2517af1737d2685b4a700d0b1814d99759c0e49175f1a5eb5548391.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelTask SeenSegmentation Class (Hammers Atlas)Mean (std)
HipPAGSTGMIGFuGStmInsPCGThaCC3VPrGPoGALG
Baseline-seen✓.88.86.93.92.79.87.82.87.90.80.68.83.77.82.84 (.07)
Neuralizer-seen✓.88.86.92.92.76.88.83.85.90.82.73.86.77.81.84 (.07)
Neuralizer-unseenX.88.87.93.91.78.87.82.85.90.81.72.85.78.81.84 (.06)
", + "bbox": [ + 78, + 88, + 893, + 186 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Segmentation of the Hammers Atlas dataset. For Neuralizer-unseen, this dataset and segmentation protocol is withheld from training. Evaluation of major labels located in the center and right of the coronal slice. See Supplement F for class abbreviations.", + "bbox": [ + 75, + 196, + 892, + 226 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Experiments 1 and 2. On the smaller held-out HammersAtlas segmentation dataset, we find that Neuralizer can generalize to unseen tasks with similar performance. These results show promise that a single Neuralizer model can perform multiple neuroimaging tasks, including generalization to new inference tasks not seen during training.", + "bbox": [ + 75, + 252, + 467, + 343 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In settings with 32 or fewer labeled example images, Neuralizer-unseen outperforms task-specific baselines despite never having seen the task or modality at train time, and taking nearly no effort or compute compared to the baselines which require substantial expertise, manual labor, and compute resources. The performance difference is largest when only one labeled subject is available, but still present at 32 subjects (Fig. 5). Neuralizer provides a performance advantage on smaller datasets likely by exploiting neuroimaging similarities across the many other neuroimaging tasks and datasets available in training.", + "bbox": [ + 75, + 345, + 467, + 511 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "When training the baselines on all available data, they can outperform Neuralizer-seen and Neuralizer-unseen by at most 3 percentage points. The inflection point of identical performance between Neuralizer and the baselines is not covered by the range of context set sizes chosen for training and evaluation due to prohibitive computational costs and is an interesting direction of study. When large annotated datasets are available, the baselines performed best on most tasks. However, training task-specific models comes at a significant cost. As a first step in the proposed problem formulation, Neuralizer offers an alternative with near equal performance, while only requiring seconds to infer any task from the context set.", + "bbox": [ + 75, + 512, + 467, + 708 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations", + "text_level": 1, + "bbox": [ + 76, + 722, + 171, + 738 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We made simplifying assumptions in this first paper demonstrating the potential of multi-task generalization in neuroimaging. The experiments are conducted on 2D data slices. In large part, we did this since running the hundreds of baselines in 3D would be infeasible on our compute cluster. Entire volumetric data also impose a challenging memory requirement on Neuralizer models. To tackle 3D data in the future, we plan to process multiple slices at a time.", + "bbox": [ + 75, + 747, + 467, + 868 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We affinely aligned the neuroimages of the context set to the target image. Early in Neuralizer development, we tried", + "bbox": [ + 76, + 869, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "training on non-aligned inputs but found that it deteriorated performance. The need for affine alignment provides an obstacle to adoption. While existing affine-alignment tools are fast and can be employed, we also believe that this requirement can be removed with further development.", + "bbox": [ + 496, + 252, + 890, + 327 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We originally experimented with tumor and lesion segmentation tasks but found this to be a more challenging scenario. Lesions are spatially heterogeneous, making learning from the context set much harder. We excluded tumor and lesion segmentation masks from the experiments, but plan to study this setting in the future.", + "bbox": [ + 496, + 327, + 890, + 417 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "While we demonstrate the proposed ideas on a broad range of tasks and modalities, neuroimage analysis can involve more domains, tasks, and populations, like image registration, surface-based tasks, CT image domains, and pediatric data. We plan to extend Neuralizer to tackle these in the future.", + "bbox": [ + 496, + 417, + 890, + 507 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 520, + 617, + 536 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Neuralizer performs accurate rapid single-pass, multitask generalization, and even outperforms task-specific baselines in limited data scenarios. Even when a large amount of annotated data is available, Neuralizer often matches baseline performance despite not training on the data. Neuralizer provides clinical researchers and scientists with a single model to solve a wide range of neuroimaging tasks on images of many modalities, and can be easily adapted to new tasks without the prohibitive requirement of retraining or fine-tuning a task-specific model. We believe this will facilitate neuroscience analyses not currently possible.", + "bbox": [ + 496, + 546, + 890, + 726 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 500, + 739, + 658, + 757 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was performed during Steffen Czolbes's visit to the Athinoula A. Martins Center for Biomedical Imaging, MGH. The collaboration was supported by the EliteForsk Travel Grant of the Danish Ministry of Higher Education and Science. The work was further funded by the Novo Nordisk Foundation (grants no. NNF20OC0062606 and NNF17OC0028360), the Lundbeck Foundation (grant no. R218-2016-883), and NIH grants 1R01AG064027 and 1R01AG070988.", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "6224", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] IXI Dataset. 4, 5", + "[2] David Abramian and Anders Eklund. Refacing: Reconstructing anonymized facial features using GANS. International Symposium on Biomedical Imaging, 2019-April:1104-1108,apr 2019.2", + "[3] Yasser Ad-Dab'bagh, O Lyttelton, JS Muehlboeck, C Lepage, D Einarson, K Mok, O Ivanov, RD Vincent, J Lerch, and E Fombonne. The CIVET image-processing environment: a fully automated comprehensive pipeline for anatomical neuroimaging research. In Proceedings of the 12th annual meeting of the organization for human brain mapping, 2006. 2", + "[4] Laith Alzubaidi, Muthana Al-Amidie, Ahmed Al-Asadi, Amjad J. Humaidi, Omran Al-Shamma, Mohammed A. Fadhel, Jinglan Zhang, J. Santamaría, and Ye Duan. Novel Transfer Learning Approach for Medical Imaging with Limited Labeled Data. Cancers 2021, Vol. 13, Page 1590, 13(7):1590, mar 2021. 3", + "[5] Vincent Arsigny, Olivier Commowick, Xavier Pennec, and Nicholas Ayache. A log-euclidean framework for statistics on diffeomorphisms. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 924-931. Springer Verlag, 2006. 2", + "[6] John Ashburner. A fast diffeomorphic image registration algorithm. NeuroImage, 38(1):95-113, oct 2007. 2", + "[7] B. B. Avants, C. L. Epstein, M. Grossman, and J. C. Gee. Symmetric diffeomorphic image registration with cross-correlation: Evaluating automated labeling of elderly and neurodegenerative brain. Medical Image Analysis, 12(1):26-41, feb 2008. 2", + "[8] Spyridon Bakas, Hamed Akbari, Aristeidis Sotiras, Michel Bilello, Martin Rozycki, Justin S. Kirby, John B. Freymann, Keyvan Farahani, and Christos Davatzikos. Advancing The Cancer Genome Atlas glioma MRI collections with expert segmentation labels and radiomic features. Scientific Data, 4, sep 2017. 4, 5", + "[9] Spyridon Bakas, Bjoern Menze, and Others. Identifying the Best Machine Learning Algorithms for Brain Tumor Segmentation, Progression Assessment, and Overall Survival Prediction in the BRATS Challenge. arXiv, 124, nov 2018. 2, 4, 5", + "[10] Guha Balakrishnan, Amy Zhao, Mert R. Sabuncu, John Guttag, and Adrian V. Dalca. VoxelMorph: A Learning Framework for Deformable Medical Image Registration. IEEE Transactions on Medical Imaging, 38(8):1788-1800, feb 2019. 2", + "[11] Cheng Bian, Chenglang Yuan, Kai Ma, Shuang Yu, Dong Wei, and Yefeng Zheng. Domain Adaptation Meets Zero-Shot Learning: An Annotation-Efficient Approach to Multi-Modality Medical Image Segmentation. IEEE Transactions on Medical Imaging, 41(5):1043–1056, may 2022. 3", + "[12] Benjamin Billot, Douglas Greve, Koen Van Leemput, Bruce Fischl, Juan Eugenio Iglesias, and Adrian V Dalca. A learning strategy for contrast-agnostic mri segmentation. arXiv preprint arXiv:2003.01995, 2020. 2" + ], + "bbox": [ + 86, + 114, + 468, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Benjamin Billot, Douglas N Greve, Oula Puonti, Axel Thielscher, Koen Van Leemput, Bruce Fischl, Adrian V Dalca, and Juan Eugenio Iglesias. SynthSeg: Domain Randomisation for Segmentation of Brain Scans of any Contrast and Resolution. jul 2021. 1, 2, 3, 4, 5", + "[14] Maxime Bucher, Tuan-Hung VU, Matthieu Cord, and Patrick Pérez. Zero-Shot Semantic Segmentation. Advances in Neural Information Processing Systems, 32, 2019. 3", + "[15] Victor Ion Butoi, Jose Javier Gonzalez Ortiz, Tianyu Ma, Mert Sabuncu, John Guttag, and Adrian V. Dalca. Universeg: Universal medical image segmentation. arXiv preprint, 2023. 3, 4, 5", + "[16] Víctor M. Campello, Carlos Martín-Isla, Cristian Izquierdo, Steffen E. Petersen, Miguel A.González Ballester, and Karim Lekadir. Combining Multi-Sequence and Synthetic Images for Improved Segmentation of Late Gadolinium Enhancement Cardiac MRI. Lecture Notes in Computer Science, 12009 LNCS:290-299, 2020. 3", + "[17] Rich Caruana, Lorien Pratt, and Sebastian Thrun. Multitask Learning. Machine Learning 1997 28:1, 28(1):41-75, 1997. 2", + "[18] Hao Chen, Qi Dou, Lequan Yu, Jing Qin, and Pheng Ann Heng. VoxResNet: Deep voxelwise residual networks for brain segmentation from 3D MR images. NeuroImage, 170:446-455, apr 2018. 2", + "[19] Sihong Chen, Kai Ma, and Yefeng Zheng. Med3D: Transfer Learning for 3D Medical Image Analysis. apr 2019. 3", + "[20] Steffen Czolbe, Oswin Krause, and Aasa Feragen. Semantic similarity metrics for learned image registration. Proceedings of Machine Learning Research, 2021. 2", + "[21] Adrian V Dalca, Katherine L Bouman, William T Freeman, Natalia S Rost, Mert R Sabuncu, and Polina Golland. Medical image imputation from image collections. IEEE transactions on medical imaging, 38(2):504-514, 2018. 2", + "[22] Adrian V Dalca, John Guttag, and Mert R Sabuncu. Anatomical priors in convolutional networks for unsupervised biomedical segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 9290-9299, 2018. 2", + "[23] Adrian V Dalca, Evan Yu, Polina Golland, Bruce Fischl, Mert R Sabuncu, and Juan Eugenio Iglesias. Unsupervised deep learning for bayesian brain mri segmentation. In Medical Image Computing and Computer Assisted Intervention—MICCAI 2019: 22nd International Conference, Shenzhen, China, October 13–17, 2019, Proceedings, Part III 22, pages 356–365. Springer, 2019. 2", + "[24] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In Conference on Computer Vision and Pattern Recognition, pages 248-255, 2009. 3", + "[25] Francisco Javier Díaz-Pernas, Mario Martínez-Zarzuela, David González-Ortega, and Míriam Antón-Rodríguez. A Deep Learning Approach for Brain Tumor Classification and Segmentation Using a Multiscale Convolutional Neural Network. Healthcare, 9(2):153, feb 2021. 3", + "[26] Theodoras Evgeniou and Massimiliano Pontil. Regularized multi-task learning. Proceedings of the Tenth ACM" + ], + "bbox": [ + 509, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "6225", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "SIGKDD International Conference on Knowledge Discovery and Data Mining, pages 109-117, 2004. 2", + "[27] Isabelle Faillenot, Rolf A. Heckemann, Maud Frot, and Alexander Hammers. Macroanatomy and 3D probabilistic atlas of the human insula. NeuroImage, 150:88-98, apr 2017. 4, 7, 24", + "[28] Mirza Faisal Beg, Michael I Miller, Alain Trouvétrouv, and Laurent Younes. Computing Large Deformation Metric Mappings via Geodesic Flows of Diffeomorphisms. International Journal of Computer Vision, 61(2):139-157, 2005. 2", + "[29] Ruiwei Feng, Xiangshang Zheng, Tianxiang Gao, Jintai Chen, Wenzhe Wang, Danny Z. Chen, and Jian Wu. Interactive Few-Shot Learning: Limited Supervision, Better Medical Image Segmentation. IEEE Transactions on Medical Imaging, 40(10):2575-2588, oct 2021. 3", + "[30] Abdur R Feyjie, Reza Azad, Marco Pedersoli, Claude Kauffman, Ismail Ben Ayed, and Jose Dolz. Semi-supervised few-shot learning for medical image segmentation. mar 2020. 3", + "[31] B Fischl. FreeSurfer. NeuroImage, 62 (2), 774-781, 2012. 2, 4, 5", + "[32] VS Fonov, AC Evans, RC McKinstry, CR Almli, and DL Collins. Unbiased nonlinear average age-appropriate brain templates from birth to adulthood. NeuroImage, 47:S102, jul 2009. 5", + "[33] Vladimir Fonov, Alan C. Evans, Kelly Botteron, C. Robert Almli, Robert C. McKinstry, and D. Louis Collins. Unbiased average age-appropriate atlases for pediatric studies. NeuroImage, 54(1):313-327, jan 2011. 5", + "[34] Richard SJ Frackowiak. Human Brain Function. 2004. 2", + "[35] Tal Goldfryd, Shiri Gordon, and Tammy Riklin Raviv. Deep semi-supervised bias field correction of mr images. Proceedings - International Symposium on Biomedical Imaging, 2021-April:1836-1840, apr 2021. 2", + "[36] Enhao Gong, John M. Pauly, Max Wintermark, and Greg Zaharchuk. Deep learning enables reduced gadolinium dose for contrast-enhanced brain MRI. Journal of Magnetic Resonance Imaging, 48(2):330-340, aug 2018. 2", + "[37] Ioannis S. Gousias, Daniel Rueckert, Rolf A. Heckemann, Leigh E. Dyet, James P. Boardman, A. David Edwards, and Alexander Hammers. Automatic segmentation of brain MRIs of 2-year-olds into 83 regions of interest. NeuroImage, 40(2):672-684, apr 2008. 4, 7, 24", + "[38] Sachin Gupta, Narinder Singh Punn, Sanjay Kumar Sonbhadra, and Sonali Agarwal. MAG-Net: Multi-task Attention Guided Network for Brain Tumor Segmentation and Classification. Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), 13147 LNCS:3-15, 2021. 3", + "[39] J Hale, Nakeisha Schimke, and John Hale. Quickshear defacing for neuroimages Attack Graph Generation on HPC Clusters View project Quickshear Defacing for Neuroimages. Frontiers in psychiatry, 21, 2011. 2", + "[40] Alexander Hammers, Richard Allom, Matthias J. Koepp, Samantha L. Free, Ralph Myers, Louis Lemieux, Tejal N." + ], + "bbox": [ + 86, + 92, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Mitchell, David J. Brooks, and John S. Duncan. Three-dimensional maximum probability atlas of the human brain, with particular reference to the temporal lobe. Human brain mapping, 19(4):224-247, aug 2003. 4, 5, 7, 24", + "[41] Xu Han, Roland Kwitt, Stephen Aylward, Spyridon Bakas, Bjoern Menze, Alexander Asturias, Paul Vespa, John Van Horn, and Marc Niethammer. Brain extraction from normal and pathological images: A joint PCA/Image-Reconstruction approach. NeuroImage, 176:431-445, aug 2018. 2", + "[42] Stine Hansen, Srishti Gautam, Robert Jenssen, and Michael Kampffmeyer. Anomaly Detection-Inspired Few-Shot Medical Image Segmentation Through Self-Supervision With Supervoxels. Medical Image Analysis, 78, mar 2022. 3", + "[43] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 3, 4", + "[44] Dan Hendrycks and Kevin Gimpel. Gaussian Error Linear Units (GELUs). jun 2016. 4", + "[45] Malte Hoffmann, Benjamin Billot, Douglas N. Greve, Juan Eugenio Iglesias, Bruce Fischl, and Adrian V. Dalca. SynthMorph: learning contrast-invariant registration without acquired images. IEEE Transactions on Medical Imaging, apr 2020. 1, 2, 3, 19", + "[46] Noah Hollmann, Samuel Müller, Katharina Eggensperger, and Frank Hutter. TabPFN: A Transformer That Solves Small Tabular Classification Problems in a Second. In NeurIPS 2022 First Table Representation Workshop, jul 2022. 3", + "[47] Andrew Hoopes, Malte Hoffmann, Douglas N Greve, Bruce Fischl, John Guttag, and Adrian V Dalca. Learning the effect of registration hyperparameters with hypermorph. arXiv preprint arXiv:2203.16680, 2022. 2, 4, 5", + "[48] Andrew Hoopes, Juan Eugenio Iglesias, Bruce Fischl, Douglas Greve, and Adrian V Dalca. TopoFit: Rapid Reconstruction of Topologically-Correct Cortical Surfaces. Proceedings of Machine Learning Research-Under Review, pages 1-13, 2022. 2", + "[49] Andrew Hoopes, Jocelyn S. Mora, Adrian V. Dalca, Bruce Fischl, and Malte Hoffmann. SynthStrip: Skull-Stripping for Any Brain Image. mar 2022. 2, 3, 4, 5", + "[50] Minyoung Huh, Pulkit Agrawal, and Alexei A. Efros. What makes ImageNet good for transfer learning? aug 2016. 3", + "[51] Mark Jenkinson, Christian F. Beckmann, Timothy E.J. Behrens, Mark W. Woolrich, and Stephen M. Smith. FSL. NeuroImage, 62(2):782-790, aug 2012. 2", + "[52] Diederik P. Kingma and Jimmy Lei Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations, dec 2015. 6", + "[53] Jens Kleesiek, Gregor Urban, Alexander Hubert, Daniel Schwarz, Klaus Maier-Hein, Martin Bendszus, and Armin Biller. Deep MRI brain extraction: A 3D convolutional neural network for skull stripping. NeuroImage, 129:460–469, apr 2016. 2", + "[54] Arno Klein, Jesper Andersson, Babak A. Ardekani, John Ashburner, Brian Avants, Ming Chang Chiang, Gary E." + ], + "bbox": [ + 509, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "6226", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Christensen, D. Louis Collins, James Gee, Pierre Hellier, Joo Hyun Song, Mark Jenkinson, Claude Lepage, Daniel Rueckert, Paul Thompson, Tom Vercauteren, Roger P. Woods, J. John Mann, and Ramin V. Parsey. Evaluation of 14 nonlinear deformation algorithms applied to human brain MRI registration. NeuroImage, 46(3):786-802, jul 2009. 2", + "[55] Simon Kornblith, Jonathon Shlens, and Quoc V Le Google Brain. Do Better ImageNet Models Transfer Better? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2661-2671, 2019. 3", + "[56] Hugo J. Kuijf, Adrià Casamitjana, D. Louis Collins, Mahsa Dadar, Achilleas Georgiou, Mohsen Ghafoorian, Dakai Jin, April Khademi, Jesse Knight, Hongwei Li, Xavier Llado, J. Matthijs Biesbroek, Miguel Luna, Qaiser Mahmood, Richard Mckinley, Alireza Mehrtash, Sebastien Ourselin, Bo Yong Park, Hyunjin Park, Sang Hyun Park, Simon Pezold, Elodie Puybureau, Jeroen De Bresser, Leticia Rittner, Carole H. Sudre, Sergi Valverde, Veronica Vilaplana, Roland Wiest, Yongchao Xu, Ziyue Xu, Guodong Zeng, Jianguo Zhang, Guoyan Zheng, Rutger Heinen, Christopher Chen, Wiesje Van Der Flier, Frederik Barkhof, Max A. Viergever, Geert Jan Biessels, Simon Andermatt, Mariana Bento, Matt Berseth, Mikhail Belyaev, and M. Jorge Cardoso. Standardized Assessment of Automatic Segmentation of White Matter Hyperintensities and Results of the WMH Segmentation Challenge. IEEE Transactions on Medical Imaging, 38(11):2556-2568, nov 2019. 2, 4, 5", + "[57] Manoj Kumar, Michael J. Anderson, James W. Antony, Christopher Baldassano, Paula P. Brooks, Ming Bo Cai, Po-Hsuan Cameron Chen, Cameron T. Ellis, Gregory Henselman-Petrusek, David Huberdeau, J. Benjamin Hutchinson, Y. Peeta Li, Qihong Lu, Jeremy R. Manning, Anne C. Mennen, Samuel A. Nastase, Hugo Richard, Anna C. Schapiro, Nicolas W. Schuck, Michael Shvartsman, Narayanan Sundaram, Daniel Suo, Javier S. Turek, David Turner, Vy A. Vo, Grant Wallace, Yida Wang, Jamal A. Williams, Hejia Zhang, Xia Zhu, Mihai Capota $^{\\text{一}}$ , Jonathan D. Cohen, Uri Hasson, Kai Li, Peter J. Ramadge, Nicholas B. Turk-Browne, Theodore L. Willke, and Kenneth A. Norman. BrainIAK: The Brain Imaging Analysis Kit. Aperture Neuro, 2021(4), jan 2021. 2", + "[58] Sonia Laguna, Riana Schleicher, Benjamin Billot, Pamela Schaefer, Brenna Mckaig, Joshua N Goldstein, Kevin N Sheth, Matthew S Rosen, W Taylor Kimberly, and Juan Eugenio Iglesias. Super-resolution of portable low-field MRI in real scenarios: integration with denoising and domain adaptation. Medical Imaging with Deep Learning (MIDL), 2022. 2", + "[59] Erik G Learned-Miller and Parvez Ahammad. Joint MRI Bias Removal Using Entropy Minimization Across Images. Advances in Neural Information Processing Systems, 17, 2004. 2", + "[60] Jaein Lee, Eunsong Kang, Eunjin Jeon, and Heung Il Suk. Meta-modulation Network for Domain Generalization in Multi-site fMRI Classification. Lecture Notes in Computer Science, 12905 LNCS:500-509, 2021. 2" + ], + "bbox": [ + 86, + 90, + 472, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[61] Gen Li, Varun Jampani, Laura Sevilla-Lara, Deqing Sun, Jonghyun Kim, and Joongkyu Kim. Adaptive Prototype Learning and Allocation for Few-Shot Segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8334-8343, 2021. 3", + "[62] Xiaoxiao Li, Yufeng Gu, Nicha Dvornek, Lawrence H. Staib, Pamela Ventola, and James S. Duncan. Multi-site fMRI analysis using privacy-preserving federated learning and domain adaptation: ABIDE results. Medical Image Analysis, 65:101765, oct 2020. 1", + "[63] Xiang Li, Tianhan Wei, Yau Pun Chen, Yu-Wing Tai, and Chi-Keung Tang. FSS-1000: A 1000-Class Dataset for Few-Shot Segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2869–2878, 2020. 3", + "[64] Sook-Lei Liew, Bethany P. Lo, Miranda R. Donnelly, Artemis Zavaliangos-Petropulu, Jessica N. Jeong, Giuseppe Barisano, Alexandre Hutton, Julia P. Simon, Julia M. Juliano, Anisha Suri, Zhizhuo Wang, Aisha Abdullah, Jun Kim, Tyler Ard, Nerisa Banaj, Michael R. Borich, Lara A. Boyd, Amy Brodtmann, Cathrin M. Buetefisch, Lei Cao, Jessica M. Cassidy, Valentina Ciullo, Adriana B. Conforto, Steven C. Cramer, Rosalia Dacosta-Aguayo, Ezequiel de la Rosa, Martin Domin, Adrienne N. Dula, Wuwei Feng, Alexandre R. Franco, Fatemeh Geranmayeh, Alexandre Gramfort, Chris M. Gregory, Colleen A. Hanlon, Brenton G. Hordacre, Steven A. Kautz, Mohamed Salah Khlif, Hosung Kim, Jan S. Kirschke, Jingchun Liu, Martin Lotze, Bradley J. MacIntosh, Maria Mataró, Feroze B. Mohamed, Jan E. Nordvik, Gilsoon Park, Amy Pienta, Fabrizio Piras, Shane M. Redman, Kate P. Revill, Mauricio Reyes, Andrew D. Robertson, Na Jin Seo, Surjo R. Soekadar, Gianfranco Spalletta, Alison Sweet, Maria Telenczuk, Gregory Thielman, Lars T. Westlye, Carolee J. Winstein, George F. Wittenberg, Kristin A. Wong, and Chunshui Yu. A large, curated, open-source stroke neuroimaging dataset to improve lesion segmentation algorithms. Scientific Data, 9(1):1-12, jun 2022. 4, 5", + "[65] Jialin Liu, Fei Chao, and Chih-Min Lin. Task Augmentation by Rotating for Meta-Learning. feb 2020. 3", + "[66] Ming-Yu Liu, Xun Huang, Arun Mallya, Tero Karras, Timo Aila, Jaakko Lehtinen, and Jan Kautz. Few-Shot Unsupervised Image-to-Image Translation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10551–10560, 2019. 3", + "[67] Weide Liu, Chi Zhang, Guosheng Lin, and Fayao Liu. CR-Net: Cross-Reference Networks for Few-Shot Segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4165–4173, 2020. 3", + "[68] Xiaoxiao Liu, Marc Niethammer, Roland Kwitt, Nikhil Singh, Matt McCormick, and Stephen Aylward. Low-Rank Atlas Image Analyses in the Presence of Pathologies. IEEE Transactions on Medical Imaging, 34(12):2583-2591, dec 2015. 2", + "[69] Xiaofeng Liu, Fangxu Xing, Chao Yang, C. C.Jay Kuo, Georges El Fakhri, and Jonghye Woo. Symmetric-" + ], + "bbox": [ + 509, + 90, + 893, + 902 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "6227", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Constrained Irregular Structure Inpainting for Brain MRI Registration with Tumor Pathology. Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), 12658 LNCS:80-91, 2021. 2", + "[70] Michael Lustig, David L. Donoho, Juan M. Santos, and John M. Pauly. Compressed sensing MRI: A look at how CS can improve on current imaging techniques. IEEE Signal Processing Magazine, 25(2):72-82, 2008. 2", + "[71] Kushagra Mahajan, Monika Sharma, and Lovekesh Vig. Meta-DermDiagnosis: Few-Shot Skin Disease Identification Using Meta-Learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 730-731, 2020. 3", + "[72] José V. Manjón, Pierrick Coup, Antonio Buades, Vladimir Fonov, D. Louis Collins, and Montserrat Robles. Non-local MRI upsampling. Medical Image Analysis, 14(6):784-792, dec 2010. 2", + "[73] José V. Manjón, Pierrick Coupe, Antonio Buades, D. Louis Collins, and Montserrat Robles. MRI superresolution using self-similarity and image priors. International Journal of Biomedical Imaging, 2010, 2010. 2", + "[74] Daniel S Marcus, Tracy H Wang, Jamie Parker, John G Csernansky, John C Morris, and Randy L Buckner. Open Access Series of Imaging Studies (OASIS): Cross-Sectional MRI Data in Young, Middle Aged, Nondemented, and Demented Older Adults. Journal of Cognitive Neuroscience, 19:1498-1507, 2007. 4, 5", + "[75] Bjoern H. Menze, Andras Jakab, Stefan Bauer, Jayashree Kalpathy-Cramer, Keyvan Farahani, Justin Kirby, Yuliya Burren, Nicole Porz, Johannes Slotboom, Roland Wiest, Lerente Lanczi, Elizabeth Gerstner, Marc Andre Weber, Tal Arbel, Brian B. Avants, Nicholas Ayache, Patricia Buendia, D. Louis Collins, Nicolas Cordier, Jason J. Corso, Antonio Criminisi, Tilak Das, Hervé Delingette, Cagatay Demiralp, Christopher R. Durst, Michel Dojat, Senan Doyle, Joana Festa, Florence Forbes, Ezequiel Geremia, Ben Glocker, Polina Golland, Xiaotao Guo, Andac Hamamci, Khan M. Iftekharuddin, Raj Jena, Nigel M. John, Ender Konukoglu, Danial Lashkari, José Antônio Mariz, Raphael Meier, Sérgio Pereira, Doina Precup, Stephen J. Price, Tammy Riklin Raviv, Syed M.S. Reza, Michael Ryan, Duygu Sarikaya, Lawrence Schwartz, Hoo Chang Shin, Jamie Shotton, Carlos A. Silva, Nuno Sousa, Nagesh K. Subbanna, Gabor Szekely, Thomas J. Taylor, Owen M. Thomas, Nicholas J. Tustison, Gozde Unal, Flor Vasseur, Max Wintermark, Dong Hye Ye, Liang Zhao, Binsheng Zhao, Darko Zikic, Marcel Prastawa, Mauricio Reyes, and Koen Van Leemput. The Multimodal Brain Tumor Image Segmentation Benchmark (BRATS). IEEE Transactions on Medical Imaging, 34(10):1993-2024, oct 2015. 4, 5", + "[76] Fausto Milletari, Nassir Navab, and Seyed Ahmad Ahmadi. V-Net: Fully convolutional neural networks for volumetric medical image segmentation. Proceedings - 2016 4th International Conference on 3D Vision, 3DV 2016, pages 565-571, dec 2016. 3, 4, 5", + "[77] Samuel Müller, Noah Hollmann, Sebastian Pineda Arango, Josif Grabocka, and Frank Hutter. Transformers Can Do" + ], + "bbox": [ + 86, + 90, + 470, + 898 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Bayesian Inference. In International Conference on Learning Representations, mar 2022. 3", + "[78] Bao Nguyen, Adam Feldman, Sarath Bethapudi, Andrew Jennings, and Chris G. Willcocks. Unsupervised region-based anomaly detection in brain mri with adversarial image inpainting. Proceedings - International Symposium on Biomedical Imaging, 2021-April:1127-1131, apr 2021. 2", + "[79] Dong Nie, Roger Trullo, Jun Lian, Caroline Petitjean, Su Ruan, Qian Wang, and Dinggang Shen. Medical image synthesis with context-aware generative adversarial networks. Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), 10435 LNCS:417-425, 2017. 2", + "[80] Alexander F.I. Osman and Nissren M. Tamam. Deep learning-based convolutional neural network for intramodality brain MRI synthesis. Journal of Applied Clinical Medical Physics, 23(4):e13530, apr 2022. 2", + "[81] Moritz Roman Hernandez Petzsche, Ezequiel de la Rosa, Uta Hanning, Roland Wiest, Waldo Enrique Valenzuela Pinilla, Mauricio Reyes, Maria Ines Meyer, Sook-Lei Liew, Florian Kofler, Ivan Ezhov, David Robben, Alexander Hutton, Tassilo Friedrich, Teresa Zarth, Johannes Burkle, The Anh Baran, Bjoern Menze, Gabriel Broocks, Lukas Meyer, Claus Zimmer, Tobias Boeckh-Behrens, Maria Berndt, Benno Ikenberg, Benedikt Wiestler, and Jan S. Kirschke. ISLES 2022: A multi-center magnetic resonance imaging stroke lesion segmentation dataset. jun 2022. 4, 5", + "[82] Kilian M Pohl, Sylvain Bouix, Motoaki Nakamura, Torsten Rohlfing, Robert W McCarley, Ron Kikinis, W Eric L Grimson, Martha E Shenton, and William M Wells. A hierarchical algorithm for mr brain image parcellation. IEEE transactions on medical imaging, 26(9):1201-1212, 2007. 2", + "[83] Maithra Raghu, Chiyuan Zhang, Google Brain, Jon Kleinberg, and Samy Bengio. Transfusion: Understanding Transfer Learning for Medical Imaging. Advances in Neural Information Processing Systems, 32, 2019. 3", + "[84] Sachin Ravi and Hugo Larochelle. Optimization as a Model for Few-Shot Learning. In International Conference on Learning Representations (ICLR), jul 2017. 3", + "[85] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, volume 9351, pages 234-241. Springer Verlag, 2015. 1, 3, 4, 5", + "[86] Edgar Schonfeld, Sayna Ebrahimi, Samarth Sinha, Trevor Darrell, and Zeynep Akata. Generalized Zero- and Few-Shot Learning via Aligned Variational Autoencoders. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8247–8255, 2019. 3", + "[87] F. Segonne, A. M. Dale, E. Busa, M. Glessner, D. Salat, H. K. Hahn, and B. Fischl. A hybrid approach to the skull stripping problem in MRI. NeuroImage, 22(3):1060-1075, jul 2004. 2", + "[88] Ozan Sener and Vladlen Koltun. Multi-Task Learning as Multi-Objective Optimization. Advances in Neural Information Processing Systems, 31, 2018. 2" + ], + "bbox": [ + 509, + 92, + 890, + 900 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "6228", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[89] Jun Seo, Young-Hyun Park, Sung Whan Yoon, and Jaekyun Moon. Task-Adaptive Feature Transformer with Semantic Enrichment for Few-Shot Segmentation. feb 2022. 3", + "[90] David W. Shattuck and Richard M. Leahy. BrainSuite: An automated cortical surface identification tool. Medical Image Analysis, 6(2):129-142, jun 2002. 2", + "[91] X. Shen, F. Tokoglu, X. Papademetris, and R. T. Constable. Groupwise whole-brain parcellation from resting-state fMRI data for network node identification. NeuroImage, 82:403-415, nov 2013. 2", + "[92] Nalini M Singh, Juan Eugenio Iglesias, Elfar Adalsteinson, Adrian V Dalca, and Polina Golland. Joint Frequency and Image Space Learning for MRI Reconstruction and Analysis. Journal of Machine Learning for Biomedical Imaging, 2022:1-28, 2022. 2, 25", + "[93] Youssef Skandarani, Nathan Painchaud, Pierre-Marc Jodoin, and Alain Lalande. On the effectiveness of GAN generated cardiac MRIs for segmentation. may 2020. 3", + "[94] Stephen M. Smith. Fast robust automated brain extraction. Human Brain Mapping, 17(3):143-155, nov 2002. 2", + "[95] Stephen M. Smith, Mark Jenkinson, Mark W. Woolrich, Christian F. Beckmann, Timothy E.J. Behrens, Heidi Johansen-Berg, Peter R. Bannister, Marilena De Luca, Ivana Drobnjak, David E. Flitney, Rami K. Niazy, James Saunders, John Vickers, Yongyue Zhang, Nicola De Stefano, J. Michael Brady, and Paul M. Matthews. Advances in functional and structural MR image analysis and implementation as FSL. NeuroImage, 23(1), 2004. 2", + "[96] Jake Snell, Kevin Swersky, and Twitter Richard Zemel. Prototypical Networks for Few-shot Learning. Advances in Neural Information Processing Systems, 30, 2017. 3", + "[97] Haoliang Sun, Ronak Mehta, Hao H. Zhou, Zhichun Huang, Sterling C. Johnson, Vivek Prabhakaran, and Vikas Singh. DUAL-GLOW: Conditional Flow-Based Generative Model for Modality Transfer. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10611-10620, 2019. 2", + "[98] Nathalie Tzourio-Mazoyer, Brigitte Landeau, Dimitri Papathanassiou, Fabrice Crivello, Octave Etard, Nicolas Delcroix, Bernard Mazoyer, and Marc Joliot. Automated anatomical labeling of activations in spm using a macroscopic anatomical parcellation of the mni mri single-subject brain. Neuroimage, 15(1):273-289, 2002. 2", + "[99] N. Tzourio-Mazoyer, B. Landeau, D. Papathanassiou, F. Crivello, O. Etard, N. Delcroix, B. Mazoyer, and M. Joliot. Automated Anatomical Labeling of Activations in SPM Using a Macroscopic Anatomical Parcellation of the MNI MRI Single-Subject Brain. NeuroImage, 15(1):273–289, jan 2002. 2", + "[100] David C. Van Essen, Stephen M. Smith, Deanna M. Barch, Timothy E.J. Behrens, Essa Yacoub, and Kamil Ugurbil. The WU-Minn Human Connectome Project: An overview. NeuroImage, 80:62-79, oct 2013. 2", + "[101] Oriol Vinyals, Google Deepmind, Charles Blundell, Timothy Lillicrap, Koray Kavukcuoglu, and Daan Wierstra. Matching Networks for One Shot Learning. Advances in Neural Information Processing Systems, 29, 2016. 3" + ], + "bbox": [ + 78, + 92, + 467, + 898 + ], + "page_idx": 12 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[102] Oriol Vinyals, Google Deepmind, Charles Blundell, Timothy Lillicrap, Koray Kavukcuoglu, and Daan Wierstra. Matching Networks for One Shot Learning. Advances in Neural Information Processing Systems, 29, 2016. 3", + "[103] Christian Wachinger, Martin Reuter, and Tassilo Klein. Deepnat: Deep convolutional neural network for segmenting neuroanatomy. NeuroImage, 170:434-445, 2018. 2", + "[104] Qi Wang, Julius Steiglechner, Tobias Lindig, Benjamin Bender, Klaus Scheffler, and Gabriele Lohmann. SuperResolution for Ultra High-Field MR Images. In Medical Imaging with Deep Learning (MIDL), 2022. 2", + "[105] Yaqing Wang, Quanming Yao, James T. Kwok, and Lionel M. Ni. Generalizing from a Few Examples. ACM Computing Surveys, 53(3), jun 2020. 3", + "[106] Mark W. Woolrich, Saad Jbabdi, Brian Patenaude, Michael Chappell, Salima Makni, Timothy Behrens, Christian Beckmann, Mark Jenkinson, and Stephen M. Smith. Bayesian analysis of neuroimaging data in FSL. NeuroImage, 45(1), 2009. 2", + "[107] Xiao Yang, Xu Han, Eunbyung Park, Stephen Aylward, Roland Kwitt, and Marc Niethammer. Registration of Pathological Images. Simulation and synthesis in medical imaging (Workshop), 9968:97, 2016. 2", + "[108] Huaxiu Yao, Long-Kai Huang, Linjun Zhang, Ying Wei, Li Tian, James Zou, Junzhou Huang, and Zhenhui Li. Improving Generalization in Meta-learning via Task Augmentation. In Proceedings of the 38th International Conference on Machine Learning, PMLR, pages 11887-11897. PMLR, jul 2021. 3", + "[109] Chandan Ganesh Bangalore Yogananda, Benjamin C. Wagner, Gowtham K. Murugesan, Ananth Madhuranthakam, and Joseph A. Maldjian. A deep learning pipeline for automatic skull stripping and brain segmentation. Proceedings - International Symposium on Biomedical Imaging, pages 727-731, apr 2019. 2", + "[110] Lin Yuan, Xue Wei, Hui Shen, Ling Li Zeng, and Dewen Hu. Multi-center brain imaging classification using a novel 3d cnn approach. IEEE Access, 6:49925-49934, sep 2018. 2", + "[111] Chi Zhang, Guosheng Lin, Fayao Liu, Jiushuang Guo, Qingyao Wu, and Rui Yao. Pyramid Graph Networks With Connection Attentions for Region-Based One-Shot Semantic Segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9587-9595, 2019. 3", + "[112] Chi Zhang, Guosheng Lin, Fayao Liu, Rui Yao, and Chunhua Shen. CANet: Class-Agnostic Segmentation Networks With Iterative Refinement and Attentive Few-Shot Learning. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 5217-5226, 2019. 3", + "[113] Amy Zhao, Guha Balakrishnan, Frédo Durand, John V Guttag, and Adrian V Dalca. Data Augmentation Using Learned Transformations for One-Shot Medical Image Segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8543-8553, 2019. 3", + "[114] Guoyan Zheng, Chengwen Chu, Daniel L. Belavy, Bulat Ibragimov, Robert Korez, Tomaz Vrtovec, Hugo Hutt," + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "6229", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 12 + }, + { + "type": "text", + "text": "Richard Everson, Judith Meakin, Isabel López Andrade, Ben Glocker, Hao Chen, Qi Dou, Pheng Ann Heng, Chunliang Wang, Daniel Forsberg, Aleš Neubert, Jurgen Fripp, Martin Urschler, Darko Stern, Maria Wimmer, Alexey A. Novikov, Hui Cheng, Gabriele Armbrecht, Dieter Felsenberg, and Shuo Li. Evaluation and comparison of 3D intervertebral disc localization and segmentation methods for 3D T2 MR data: A grand challenge. Medical Image Analysis, 35:327-344, jan 2017. 3", + "bbox": [ + 117, + 90, + 472, + 218 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "6230", + "bbox": [ + 482, + 945, + 516, + 955 + ], + "page_idx": 13 + } +] \ No newline at end of file diff --git a/2023/Neuralizer_ General Neuroimage Analysis Without Re-Training/eee09609-bc86-4647-98dc-269f47c8002f_model.json b/2023/Neuralizer_ General Neuroimage Analysis Without Re-Training/eee09609-bc86-4647-98dc-269f47c8002f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f3e597d7e41a2a8081f1e57e81dcdfb6229f849b --- /dev/null +++ b/2023/Neuralizer_ General Neuroimage Analysis Without Re-Training/eee09609-bc86-4647-98dc-269f47c8002f_model.json @@ -0,0 +1,3187 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.167, + 0.131, + 0.804, + 0.154 + ], + "angle": 0, + "content": "Neuralizer: General Neuroimage Analysis without Re-Training" + }, + { + "type": "text", + "bbox": [ + 0.152, + 0.182, + 0.445, + 0.233 + ], + "angle": 0, + "content": "Steffen Czolbe \nUniversity of Copenhagen and MGH \nper.sc@di.ku.dk" + }, + { + "type": "text", + "bbox": [ + 0.49, + 0.182, + 0.818, + 0.233 + ], + "angle": 0, + "content": "Adrian V. Dalca \nMIT, and MGH, Harvard Medical School \nadalca@mit.edu" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.313, + 0.284 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.302, + 0.471, + 0.497 + ], + "angle": 0, + "content": "Neuroimage processing tasks like segmentation, reconstruction, and registration are central to the study of neuroscience. Robust deep learning strategies and architectures used to solve these tasks are often similar. Yet, when presented with a new task or a dataset with different visual characteristics, practitioners most often need to train a new model, or fine-tune an existing one. This is a time-consuming process that poses a substantial barrier for the thousands of neuroscientists and clinical researchers who often lack the resources or machine-learning expertise to train deep learning models. In practice, this leads to a lack of adoption of deep learning, and neuroscience tools being dominated by classical frameworks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.498, + 0.47, + 0.664 + ], + "angle": 0, + "content": "We introduce Neuralizer, a single model that generalizes to previously unseen neuroimaging tasks and modalities without the need for re-training or fine-tuning. Tasks do not have to be known a priori, and generalization happens in a single forward pass during inference. The model can solve processing tasks across multiple image modalities, acquisition methods, and datasets, and generalize to tasks and modalities it has not been trained on. Our experiments on coronal slices show that when few annotated subjects are available, our multi-task network outperforms task-specific baselines without training on the task." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.693, + 0.208, + 0.709 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.469, + 0.825 + ], + "angle": 0, + "content": "Computational methods for the processing and analysis of neuroimages have enabled a deep understanding of the human brain. The field has also led to advanced patient care by facilitating non-invasive methods of diagnosis and treatment. Recent deep learning research promises to substantially increase the accuracy and speed of neuroimaging analysis methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.469, + 0.901 + ], + "angle": 0, + "content": "A drawback of most current deep-learning-based approaches is that each model is limited to solving the task it has been trained on, on the type of data it has been trained on. Generalization to new tasks and domains, such as different acquisition protocols or new segmentation, is" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.268, + 0.892, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.485, + 0.892, + 0.596 + ], + "angle": 0, + "content": "Figure 1. Neuralizer can solve a broad range of image processing tasks, including new ones not seen during training, with a single model by conditioning the prediction on a context set of examples. After training on a diverse set of tasks, the model can generalize to new tasks in a single forward pass without re-training or finetuning. The model is highly flexible, requiring no prior definition of the set of tasks, and can be conditioned with context sets of any length." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.892, + 0.9 + ], + "angle": 0, + "content": "a main barrier to adoption [62]. Performing neuroimaging tasks like segmentation, registration, reconstruction, or motion correction requires different models for each processing step, despite operating on the same input data and methods exhibiting strong similarities in network architecture [13,45,85]. Yet, designing and training models to solve these tasks on each dataset is prohibitively expensive. To train a deep learning model, a dataset needs to be compiled and often manually annotated, and the network, training, and data loading logic needs to be implemented. All these steps generally require machine learning and neuroimaging expertise. In addition, computational resources like specialized graphics processing hardware and software infrastructure needs to be available. These requirements are particularly problematic in clinical research settings due to a high cost of annotation and a lack of machine learning expertise and hardware." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6217" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.08, + 0.088, + 0.174, + 0.113 + ], + "angle": 0, + "content": "Binary Segmentation" + }, + { + "type": "image_caption", + "bbox": [ + 0.182, + 0.089, + 0.278, + 0.113 + ], + "angle": 0, + "content": "Modality Transfer" + }, + { + "type": "image_caption", + "bbox": [ + 0.284, + 0.089, + 0.382, + 0.113 + ], + "angle": 0, + "content": "Super Resolution" + }, + { + "type": "image_caption", + "bbox": [ + 0.386, + 0.089, + 0.483, + 0.113 + ], + "angle": 0, + "content": "Skull Stripping" + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.115, + 0.177, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.182, + 0.115, + 0.278, + 0.265 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.115, + 0.382, + 0.265 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.386, + 0.115, + 0.483, + 0.265 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.488, + 0.089, + 0.585, + 0.113 + ], + "angle": 0, + "content": "Motion Correction" + }, + { + "type": "image_caption", + "bbox": [ + 0.591, + 0.089, + 0.686, + 0.113 + ], + "angle": 0, + "content": "Undersampled Reconstruction" + }, + { + "type": "image_caption", + "bbox": [ + 0.696, + 0.089, + 0.788, + 0.113 + ], + "angle": 0, + "content": "Denoising & Bias Correction" + }, + { + "type": "image_caption", + "bbox": [ + 0.811, + 0.091, + 0.876, + 0.107 + ], + "angle": 0, + "content": "Inpainting" + }, + { + "type": "image", + "bbox": [ + 0.795, + 0.115, + 0.892, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.489, + 0.115, + 0.584, + 0.265 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.589, + 0.115, + 0.686, + 0.265 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.692, + 0.115, + 0.788, + 0.265 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.795, + 0.191, + 0.891, + 0.265 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.124, + 0.277, + 0.843, + 0.291 + ], + "angle": 0, + "content": "Figure 2. Example neuroimaging tasks and modalities included in our dataset (top: input images, bottom: output images)." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.317, + 0.182, + 0.332 + ], + "angle": 0, + "content": "Contribution" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.341, + 0.469, + 0.431 + ], + "angle": 0, + "content": "We introduce Neuralizer, a general-purpose neuroimaging model that can solve a broad range of neuroimaging tasks on diverse image modalities (Fig. 2), without the need for task-specific training or fine-tuning. Neuralizer can solve new tasks, unseen during training, using a set of examples of the new task at inference (Fig. 1)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.432, + 0.469, + 0.522 + ], + "angle": 0, + "content": "Neuralizer involves a convolutional architecture (Fig. 3), that takes as input a context set of examples that define the processing task, and thus does not require prior specification of the tasks. The method enables single-pass generalization during inference and can process any number of reference images in a single pass to inform the prediction." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.524, + 0.469, + 0.689 + ], + "angle": 0, + "content": "As a first method tackling task generalization in neuroimaging, we focus on analyzing the capabilities of such system and presenting general insights, and limit our experiments to 2D. We evaluate our model by comparing the single-pass generalization performance to task-specific baselines conditioned on an equivalent amount of data. We find that Neuralizer outperforms the baselines on tasks where \\(\\leq 32\\) labeled examples are available, despite never training on the task. When generalizing to new segmentation protocols, Neuralizer matches the performance of baselines trained directly on the dataset." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.705, + 0.219, + 0.72 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.731, + 0.469, + 0.806 + ], + "angle": 0, + "content": "We give a short introduction to neuroimaging tasks, terminology, and methods. We then provide an overview of fundamental methods for adapting a model to multiple domains, including multi-task learning, few-shot learning, fine-tuning, and data synthesis." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.817, + 0.273, + 0.833 + ], + "angle": 0, + "content": "2.1. Neuroimage analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.841, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Neuroimage analysis employs computational techniques to study the structure and function of the human brain. Common imaging techniques are structural magnetic resonance imaging (MRI), functional MRI, diffusion tensor" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.318, + 0.892, + 0.409 + ], + "angle": 0, + "content": "imaging (DTI), computed tomography (CT), and Positron emission tomography (PET). Each imaging method can create diverse images with different characteristics and contrasts, which are further diversified depending on the properties of the acquisition site [60, 110], device, protocol, imaging sequence [56], and use of contrast agents [9,36]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.41, + 0.892, + 0.546 + ], + "angle": 0, + "content": "To analyze these images, a variety of processing tasks are most often combined in a processing pipeline. Common processing tasks include anatomical segmentation [12, 18, 22, 23, 31, 82, 98, 103], skull stripping [49, 49, 53, 87, 94, 109], defacing [2, 39], registration [5-7, 10, 20, 28, 45, 47, 54], modality transfer [79, 80, 97], in-painting [41, 68, 69, 78, 107], super-resolution [21, 58, 72, 73, 104], reconstruction, and de-noising [58, 70, 92], bias field removal [35, 59], surface fitting [48] and parcellation [91, 99]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.547, + 0.892, + 0.773 + ], + "angle": 0, + "content": "Multiple toolboxes provide a suite of interoperable software components, most implementing classical optimization strategies. Widely used toolboxes include Freesurfer [31], FSL [51, 95, 106], SPM [34], CIVET [3], BrainSuite [90], HCP pipeline [100], and BrainIAK [57]. Deep-learning-based methods are starting to be included because of their improved accuracy and shorter runtime [13, 49]. While these methods provide solutions for common neuroimaging applications, most are limited to a single task and few modalities. Developers need to manually update the pipelines to include new processing tasks and to support a wider variety of image modalities. This process requires extensive technical expertise and computational resources, often not available to the clinical neuroscientists focusing on scientific questions." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.787, + 0.687, + 0.803 + ], + "angle": 0, + "content": "2.2. Multi-task learning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Multi-Task Learning (MTL) frameworks solve multiple tasks simultaneously by exploiting similarities between related tasks [17]. MTL can improve performance and reduce computational cost and development time compared to designing task-specific solutions [26, 88]. In neuroimaging, MTL networks were recently proposed for the simultaneous" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6218" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.242 + ], + "angle": 0, + "content": "segmentation and classification of brain tumors by training a single network with separate prediction heads associated with the different tasks [25, 38]. This strategy is challenging to scale as the number of tasks increases, requires prior determination of the set of tasks, and importantly does not enable generalization of the model to new tasks. With Neuralizer, we build on these methods to achieve scalable MTL, without the need for multiple network heads, and importantly with the ability to generalize to new tasks and modalities." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.253, + 0.206, + 0.269 + ], + "angle": 0, + "content": "2.3. Fine-tuning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.276, + 0.473, + 0.488 + ], + "angle": 0, + "content": "To tackle problems in the limited data scenarios frequent in medical imaging, neural networks can be pre-trained on a related task with high data availability and then fine-tuned for specific tasks. For example, a common approach involves taking a Res-Net [43] trained on ImageNet [24] and fine-tuning part of the network for a new task [50, 55, 101]. For medical imaging, networks pre-trained on large sets of medical images are available [19], and fine-tuning them to new tasks results in shortened training time and higher accuracy [4, 71]. However, fine-tuning also requires machine learning expertise and computational resources, most often not available in clinical research. Additionally, in scenarios with small datasets, fine-tuning models trained on large vision datasets can be harmful [83]." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.497, + 0.253, + 0.513 + ], + "angle": 0, + "content": "2.4. Few-shot learning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.521, + 0.473, + 0.838 + ], + "angle": 0, + "content": "Few-shot models generate predictions from just a few labeled examples [66, 84, 86, 105], or in the case of zero-shot methods [14], none at all. Many of these methods require training or fine-tuning. In computer vision, several methods pass a query image, along with a set of support images and labels as input to the model [66, 89, 96, 102]. Natural image segmentation methods [67, 112] use single image-label pairs [61, 111] as support or aggregate information from a larger support set [63]. Recent few-shot learning methods in the medical image segmentation setting [11, 29, 30] operate on a specific anatomical region in a single image modality [42, 114]. Similar Prior-Data Fitted Networks (PFNs) are fitted to multiple datasets at once to learn the training and prediction algorithm [77]. During training, this strategy draws a dataset, a set of data points and their labels from it, masks one of the labels and predicts it. The resulting model aims to generalize to new datasets. PFNs have only been applied to low-dimensional and tabular data [46]. Our solution builds on ideas from these methods, but aims to solve a much larger range of diverse image-to-image tasks on neuroimages of many modalities." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.848, + 0.369, + 0.864 + ], + "angle": 0, + "content": "2.5. Data augmentation and synthesis" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.472, + 0.902 + ], + "angle": 0, + "content": "Data augmentation increases the diversity of training data by augmenting or modifying existing data [85, 113]. It" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.228 + ], + "angle": 0, + "content": "improves model robustness to input variability that may not be available in the original training data. In neuroimaging, arbitrary image modalities can be simulated by synthesis of images without requiring any real data [13, 16, 45, 49, 93]. In meta-learning, data augmentation can further be used to generate entirely new tasks [15, 65, 108]. We use data augmentations and further expand existing methods by developing rich neuroimaging task augmentations for generalization to unseen neuroimaging tasks." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.242, + 0.615, + 0.258 + ], + "angle": 0, + "content": "3. Neuralizer" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.267, + 0.894, + 0.344 + ], + "angle": 0, + "content": "We introduce Neuralizer, a multi-task model for neuroimage analysis tasks. In this section, we first define the training framework and adaptations necessary to operate on a diverse range of tasks and input types. We then introduce the model architecture, training, and inference strategies." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.353, + 0.774, + 0.368 + ], + "angle": 0, + "content": "3.1. Generalizable multi-task model" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.376, + 0.892, + 0.437 + ], + "angle": 0, + "content": "Let \\( T \\) represent a set of tasks, with a subset of tasks \\( T_{\\mathrm{seen}} \\) seen during training. Each task consists of input-output image pairs \\( (x_{t}, y_{t}) \\) from potentially multiple underlying datasets with input and output spaces \\( x_{t} \\in \\mathcal{X}, y_{t} \\in \\mathcal{Y} \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.437, + 0.892, + 0.542 + ], + "angle": 0, + "content": "To enable generalization to unseen tasks, we condition the model on a context set \\( C_t = \\{(x_{t,i},y_{t,i})\\}_{i=1}^N \\) of input-output image pairs passed to the model alongside the prediction task. The context set defines the desired task, and can vary in size \\( |C_t| = N \\) and is re-sampled from the underlying task-datasets for each input. Fig. 1 gives an example for a modality transfer task." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.543, + 0.892, + 0.603 + ], + "angle": 0, + "content": "We employ a neural network \\( g_{\\theta}(x_t, C_t) = y_t \\) with weights \\( \\theta \\) that applies the task defined by context set \\( C_t \\) to the input neuroimage \\( x_t \\). We optimize the network using supervised training with the loss" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.615, + 0.891, + 0.646 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} \\left(T _ {\\text {s e e n}}; \\theta\\right) = \\mathbb {E} _ {t \\in T _ {\\text {s e e n}}} \\left[ \\mathbb {E} _ {\\left(x _ {t}, y _ {t}, C _ {t}\\right)} \\left[ \\mathcal {L} _ {t} \\left(y _ {t}, g _ {\\theta} \\left(x _ {t}, C _ {t}\\right)\\right) \\right] \\right], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.647, + 0.771, + 0.661 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_t\\) is a task-specific loss function." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.672, + 0.719, + 0.687 + ], + "angle": 0, + "content": "3.2. Design for diverse tasks" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.695, + 0.892, + 0.741 + ], + "angle": 0, + "content": "To process different tasks with a single model, we carefully select the loss function, neuroimage encodings, and generation of the training set for each task type." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.753, + 0.895, + 0.906 + ], + "angle": 0, + "content": "Loss function. Neuralizer solves both segmentation tasks (e.g. anatomical segmentation and skull-stripping via a brain mask), more general and image-to-image tasks (e.g. denoising). We use the Soft Dice Loss [76] for segmentation-like tasks, and the pixel-wise Mean Squared Error \\( MSE(y_{t}, g(x_{t}, C_{t})) = \\frac{1}{2\\sigma^{2}} \\sum_{p} [y_{tp} - g(x_{t}, C_{t})_{p}]^{2} \\) with balancing hyperparameter \\( \\sigma^{2} \\) for other tasks. As the network optimizes multiple tasks during training, the balance of the loss terms can dramatically affect the optimization and resulting performance." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6219" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.088, + 0.418, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.422, + 0.088, + 0.895, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.297, + 0.893, + 0.355 + ], + "angle": 0, + "content": "Figure 3. Neuralizer consists of 7 Pairwise-Conv-Avg blocks (right), arranged in a U-Net-like [76, 85] configuration (left). Each Pairwise-Conv-Avg block enables interaction between the input image and the image pairs present in the context set. The block consists of a residual unit, pairwise convolution of each context member with the target, and an averaging of results across the context set to update the representation. The architecture is invariant to context size \\( N \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.38, + 0.47, + 0.485 + ], + "angle": 0, + "content": "Input and output encoding. For Neuralizer to work on both segmentation and image-to-image tasks, we facilitate simultaneous input of multiple image modalities and masks. We design the input space \\(\\mathcal{X}\\) to accept floating point value images with three channels, and zero-pad any channels unnecessary for a specific task. The output space \\(\\mathcal{Y}\\) follows the same design but uses only one channel." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.502, + 0.469, + 0.638 + ], + "angle": 0, + "content": "Training dataset creation. At each training iteration, we first sample a task \\( t \\) from \\( T_{\\mathrm{seen}} \\), selecting the task-specific dataset (Tab. 1). From this dataset, we sample the input image, ground truth output, and image pairs for the context set. To increase the range of images that can be used to condition the trained model, the image modalities and acquisition protocols of entries of the context set can differ from the input image for some tasks. Supplemental section G contains a detailed description of the training data generator." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.65, + 0.262, + 0.665 + ], + "angle": 0, + "content": "3.3. Model architecture" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.673, + 0.469, + 0.748 + ], + "angle": 0, + "content": "Fig. 3 shows the Neuralizer network architecture, adapted with the concurrently developed [15] - a method that focuses on solving broad segmentation tasks. As the architecture is independent of the task, we omit the task subscript in this section." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.47, + 0.901 + ], + "angle": 0, + "content": "The input image \\( x \\) and the image pairs of the context set \\( C_i = (x_i, y_i), i = 1, \\dots, N \\) are first passed through an embedding layer consisting of a single \\( 1 \\times 1 \\) convolution with learnable kernels \\( e_x, e_C \\), to obtain the representations \\( r_x = x * e_x \\), \\( r_{C_i} = \\operatorname{cat}(x_i, y_i) * e_C \\) where \\( * \\) is the convolution operator. This combines each context image pair to a joint representation \\( r_{C_i} \\) and maps all representations to a uniform channel width \\( c \\), which is constant throughout the model. Next, we process the representations using multiple Pairwise-Conv-Avg Blocks (explained below), ar" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.378, + 0.892, + 0.407 + ], + "angle": 0, + "content": "Table 1. Tasks, Modalities, Datasets, and Segmentation classes used in this paper, and involved in training Neuralizer." + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.418, + 0.892, + 0.729 + ], + "angle": 0, + "content": "
TasksModalities
Binary SegmentationT1-w.
Modality TransferT2-w.
Super ResolutionMRA
Skull StrippingPD
Motion CorrectionFLAIR
Undersampled ReconstructionADC
Denoising & Bias correctionDWI
InpaintingDTI (17 dir.)
DatasetsSegmentation Classes
OASIS 3 [47, 74]Freesufer protocol, 31
BRATS [8, 9, 75]classes [13, 31]
IXI [1]
ATLAS R2.0 [64]Manually-annotated Hammers
Hammers Atlas [40]Atlas, 96 classes [27, 37, 40]
WMH Challenge [56]
ISLES2022 [81]Brainmasks [31, 49]
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.762, + 0.892, + 0.852 + ], + "angle": 0, + "content": "ranged as a U-Net-like configuration [76,85] to exploit multiple scales. The output \\( r_x^{\\mathrm{out}} \\) of the final Pairwise-Conv-Avg Block is processed by a residual unit [43] and a final \\( 1 \\times 1 \\) conv layer to map to one output channel. All residual units consist of two \\( 3 \\times 3 \\) conv layers, a shortcut connection, and GELU activation functions [44]." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Compared to standard CNNs, Neuralizer uses a mechanism to enable knowledge transfer from the context set to the input image. We design the Pairwise-Conv-Avg Block" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6220" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.091, + 0.473, + 0.394 + ], + "angle": 0, + "content": "(Fig. 3, right) to model this interaction. The block maps from representations of the target input \\( r_x^{\\mathrm{in}} \\) and context pairs \\( r_{C_i}^{\\mathrm{in}} \\) to output representations \\( r_x^{\\mathrm{out}} \\), \\( r_{C_i}^{\\mathrm{out}} \\) of the same size. First, we process each input separately with a residual unit to obtain \\( r_x^{\\mathrm{int}} = \\mathrm{ResUnit}_x(r_x^{\\mathrm{in}}) \\) and \\( r_{C_i}^{\\mathrm{int}} = \\mathrm{ResUnit}_C(r_{C_i}^{\\mathrm{in}}) \\). The residual units, which involve two convolutions, operate on the context representations and have shared weights. Second, we pairwise concatenate the context representations with the target representation on the channel dimension: \\( p_i = \\mathrm{cat}(r_x^{\\mathrm{int}}, r_{C_i}^{\\mathrm{int}}) \\). We combine the pairwise representations and reduce the channel size back to \\( c \\) using a \\( 1 \\times 1 \\) convolution with learnable kernel \\( k_x \\), and update the target representation by averaging across context members \\( r_x^{\\mathrm{out}} = r_x^{\\mathrm{int}} + \\frac{1}{N}\\sum_{i=1}^{N}p_i * k_x \\). The context representations are updated with a separate kernel \\( r_{C_i}^{\\mathrm{out}} = r_{C_i}^{\\mathrm{int}} + p_i * k_C \\). We then re-size the outputs of a Pairwise-Conv-Avg Block before feeding them as input the next block. We experimented with attention-based and weighted average approaches but found that they did not lead to an increased generalization to unseen tasks." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.403, + 0.267, + 0.419 + ], + "angle": 0, + "content": "3.4. Task augmentations" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.426, + 0.47, + 0.547 + ], + "angle": 0, + "content": "To further diversify the training dataset, we employ task augmentations [15], a group of transformations applied at random to the input, output, and context images. The objective is to increase the diversity of tasks to discourage the model from merely memorizing the tasks in the training data. A list of all task augmentations is summarized in Tab. 2, with more detailed descriptions and visual examples in Supplement C." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.557, + 0.188, + 0.57 + ], + "angle": 0, + "content": "3.5. Inference" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.58, + 0.47, + 0.686 + ], + "angle": 0, + "content": "During inference, we supply an input image \\( x_{i} \\) and a context set \\( C_i \\) from the desired task. Given these inputs, a simple feed-forward pass through the model provides the prediction \\( \\hat{y} = g(x,C) \\). To further increase accuracy at test-time, we use context-set bootstrapping [15]. We also increase the context set by sampling with replacement from the context set, and add small affine augmentations." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.699, + 0.21, + 0.715 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.724, + 0.47, + 0.769 + ], + "angle": 0, + "content": "We first compare Neuralizer with task-specific networks, which require substantial expertise and compute. We then analyze the effect of the size of the context set, and the" + }, + { + "type": "table_caption", + "bbox": [ + 0.185, + 0.792, + 0.362, + 0.806 + ], + "angle": 0, + "content": "Table 2. Task Augmentations" + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.817, + 0.47, + 0.91 + ], + "angle": 0, + "content": "
Task Augmentations
IntensityMappingSyntheticModality
SobelFilterMaskInvert
MaskContourMaskDilation
PermuteChannelsDuplicateChannels
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "multi-task generalization to unseen segmentation protocols and image modalities. For this first method of large-scale multi-task generalization in neuroimaging, we conduct the experiments on 2D image slices." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.161, + 0.575, + 0.175 + ], + "angle": 0, + "content": "4.1. Data" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.184, + 0.895, + 0.516 + ], + "angle": 0, + "content": "To create a diverse dataset encompassing a multitude of different modalities, acquisition protocols, devices, and tasks, we pool neuroimages from the public datasets OASIS3 [47, 74], BRATS [8, 9, 75], Atlas R2.0 [64], Hammers Atlas [40], IXI [1], ISLES2022 [81], and the White Matter Hyperintensities Challenge [56]. We segment all subjects with Synthseg [13, 31]. Based on the segmentation, we affinely align the images to the MNI 152 template space [32, 33], and resample to \\(1\\mathrm{mm}\\) isometric resolution at a size of \\(192 \\times 224 \\times 192\\mathrm{mm}\\). We perform manual quality control of the segmentation and registration by ensuring no segmented areas fall outside of the cropped volume and discard subjects failing this check (4 subjects). We extract a coronal slice of \\(192 \\times 192\\mathrm{mm}\\), bisecting the frontal Brain stem, Hippocampus, Thalamus, and Lateral ventricles. We rescale image intensities to the \\([0, 1]\\) interval using dataset-specific percentiles. For full head images, we create a brain mask with Synthstrip [31, 49]. The final dataset contains 2,282 subjects with 15,911 images and segmentation masks across 8 modalities. Subjects of the seven original datasets are split into \\(80\\%\\) for training and validation, \\(20\\%\\) test, with a minimum of 15 test subjects per dataset." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.525, + 0.593, + 0.539 + ], + "angle": 0, + "content": "4.2. Models" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.548, + 0.892, + 0.656 + ], + "angle": 0, + "content": "Neuralizer-seen. This Neuralizer model includes all tasks available during training. We use this model to evaluate the performance on unseen scans from tasks and modalities that have been included in the training. The model uses the 4-stage architecture shown in Fig. 3 with 64 channels per layer. During training, the context size \\( |C_i| \\) is sampled from \\( \\mathcal{U}_{\\{1,32\\}} \\) at each iteration." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.672, + 0.894, + 0.747 + ], + "angle": 0, + "content": "Neuralizer-unseen. To evaluate Neuralizer performance on tasks and modalities it has not been trained on, we train a family of Neuralizer models where a single task or modality is excluded from the training set. The model architecture of Neuralizer-unseen is identical to Neuralizer-seen." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Baseline-seen. As no established baseline for multi-task and multi-modality models in neuroimaging can tackle the number of tasks we aim for, we compare Neuralizer to an ensemble of task-specific U-Nets [76, 85]. However, training one model for each task and modality requires overwhelming computational resources. To reduce the computational requirement, we follow previous modality-agnostic models [13, 49] and train each model on multiple input modalities. This lowers the number of models to be" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.514, + 0.957 + ], + "angle": 0, + "content": "6221" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.083, + 0.088, + 0.271, + 0.195 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.289, + 0.089, + 0.477, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.089, + 0.685, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.706, + 0.089, + 0.893, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.204, + 0.27, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.204, + 0.477, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.493, + 0.204, + 0.685, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.701, + 0.204, + 0.894, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.077, + 0.322, + 0.27, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.322, + 0.477, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.493, + 0.322, + 0.816, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.438, + 0.894, + 0.507 + ], + "angle": 0, + "content": "Figure 4. Performance of multi-task Neuralizer and the task-specific baselines on each task, averaged across all modalities in the test set. The tasks being evaluated were included in the training of Neuralizer-seen (orange), held out in Neuralizer-unseen (blue), and specifically trained on by each task-specific baseline (gray). The x-axis is the size of the train/context set, and the y-axis is the Dice/PSNR score. Some points on the x-axis are omitted for better visibility. 'All' refers to all available train data for the task, ranging from 249 to 2,282 subjects depending on the task. The bars denote standard deviation across modalities. We extract results for T1 scans in Supplement D." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.535, + 0.47, + 0.61 + ], + "angle": 0, + "content": "trained to one per task, segmentation class, and modalitytransfer output modality. To compare Baseline-seen with Neuralizer-unseen given an equal amount of data, we train baselines with training set sizes of \\(\\{1,2,4,8,16,32,\\mathrm{all}\\}\\) and employ standard data augmentation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.61, + 0.47, + 0.716 + ], + "angle": 0, + "content": "We use a 4-stage U-Net architecture with one residual block per layer. The channel width is tuned experimentally for each training dataset size. We select 256 channels when all data is available for training, and 64 channels otherwise. Using larger U-Nets resulted in overfitting and lower performance. Supplement H summarizes model parameter counts and inference costs." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.726, + 0.182, + 0.743 + ], + "angle": 0, + "content": "4.3. Training" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.47, + 0.856 + ], + "angle": 0, + "content": "We use supervised training, task-specific loss functions, and weigh the MSE loss by selecting \\(\\sigma^2 = 0.05\\), resulting in the loss terms being of similar magnitude. All models are trained with a batch size of 8, a learning rate of \\(10^{-4}\\), and the ADAM optimizer [52]. To speed up training, we undersample tasks that the model learns quickly, with sampling weights given in Supplement I." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.902 + ], + "angle": 0, + "content": "In addition to the task augmentations, we use data augmentations via random affine movements, random elastic deformations, and random flips along the sagittal plane." + }, + { + "type": "image", + "bbox": [ + 0.499, + 0.531, + 0.892, + 0.722 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.731, + 0.895, + 0.87 + ], + "angle": 0, + "content": "Figure 5. Results averaged across tasks, expressed as relative performance compared to the baseline trained on all data. The tasks being evaluated were included in the training of Neuralizer-seen (orange), held out in Neuralizer-unseen (blue), and specifically trained on by each task-specific baseline (gray). The x-axis is the size of the train/context set, and the y-axis is the relative score. Some points on the x-axis are omitted for better visibility. 'All' refers to all available data for the task, ranging from 249 to 2,282 subjects depending on the task. Bars: standard deviation across tasks/modalities." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6222" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.212 + ], + "angle": 0, + "content": "For Baseline-seen, we reuse the augmentations but remove those that introduce uncertainty in the desired output. The training time for Neuralizer is 7 days on a single A100 GPU. The training time of the baseline models is capped at 5 days. All models use early stopping, ending the training after 25 epochs exhibiting no decrease in validation loss. The model with the lowest validation loss is used for further evaluation on the test set." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.232, + 0.199, + 0.248 + ], + "angle": 0, + "content": "4.4. Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.259, + 0.473, + 0.396 + ], + "angle": 0, + "content": "We evaluate the Dice coefficient for the segmentation and skull stripping tasks, and the Peak Signal-to-Noise Ratio (PSNR) for the image-to-image tasks on the test set. As the low-data regime is of particular interest, we measure performance as a function of context set size for the Neuralizer models and use training set size as an analog for the U-Net models. We evaluate context sets of up to 32 subjects. Larger context sets are possible but come at a linear cost in memory." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.414, + 0.395, + 0.431 + ], + "angle": 0, + "content": "4.5. Experiment 1: Baseline Comparison" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.441, + 0.47, + 0.547 + ], + "angle": 0, + "content": "To assess if the proposed multi-task approach is competitive with task-specific models, we evaluate the performance and runtime of Neuralizer-seen, Neuralizer-unseen, and Baseline-seen on the test-set of each task. For Neuralizer-unseen, we withhold image modalities using a leave-one-out strategy during training and evaluate on the unseen modalities at test time." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.577, + 0.47, + 0.789 + ], + "angle": 0, + "content": "Results. We display the results by task, averaged across modalities in Fig. 4. We also provide an evaluation of using just the T1 modality in Supplement D, since many task-specific networks in neuroimage analysis literature focus on T1 images. We further aggregate performance across all tasks in Fig. 5, and provide tabular results in Supplement E. Both Neuralizer models outperform most task-specific baselines trained on up to 32 samples. When training the baselines on all available data, the baselines outperform Neuralizer-seen by 2 percentage points in relative performance, and Neuralizer-unseen by 3 percentage points. The loss in performance when generalizing to an unseen modality (between Neuralizer-seen and Neuralizer-unseen) is less than 2 percentage points for all context set sizes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.792, + 0.47, + 0.868 + ], + "angle": 0, + "content": "Training the baseline model to convergence on 32 samples took on average \\(28.2 \\pm 16.6\\) hours per task, using one A100 GPU. Since Neuralizer only requires inference for a new task, it is orders of magnitude faster, requiring less than 0.1 seconds on a GPU and less than 3 seconds on a CPU." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.902 + ], + "angle": 0, + "content": "We provide qualitative samples of the predictions from Neuralizer-seen model in Supplement A, Figures 6-8." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.091, + 0.838, + 0.108 + ], + "angle": 0, + "content": "4.6. Experiment 2: Context set size analysis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.114, + 0.892, + 0.281 + ], + "angle": 0, + "content": "We assess the few-shot setting that is prevalent in neuroimage analysis, where few annotated images are often available for a new task. We evaluate performance as a function of the number of labeled samples. For Neuralizer, we evaluate the model with context-set sizes of \\(\\{1,2,4,8,16,32\\}\\) unique subjects from the test set. For the baseline, we trained models with reduced training set sizes of the same amount of subjects. To reduce the effect of random training subject selection, we train three separate baselines with \\(n = 1\\), two baselines with \\(n = 2\\), and average results of models with the same \\(n\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.291, + 0.892, + 0.413 + ], + "angle": 0, + "content": "Results. Tab. 4 and Figs. 4, 5 illustrate the results. For all models, prediction accuracy increases with the availability of labeled data, with diminishing returns. For both Neuralizer models, a context set size of one achieves more than \\(90\\%\\) of the performance attainable with all data. For most tasks, the baseline performs overall worse than both Neuralizer models when \\(\\leq 32\\) labeled samples are available but achieves the best overall performance on larger datasets." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.421, + 0.892, + 0.453 + ], + "angle": 0, + "content": "4.7. Experiment 3: Generalization to a new segmentation protocol" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.459, + 0.892, + 0.626 + ], + "angle": 0, + "content": "The Hammers Atlas dataset [27,37,40] provides an alternative anatomical segmentation protocol to the widely-used Freesurfer segmentation available for most subjects in the dataset. The shape, size, and amount of annotated regions in the protocols differ drastically. A different image acquisition site also leads to differences in visual characteristics. We use the Hammers Atlas dataset to evaluate Neuralizer-unseen by entirely withholding the dataset and its annotations from training. We evaluate the Dice coefficient of the 14 major anatomical segmentation classes present in the center and right half of the coronal slice." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.636, + 0.892, + 0.788 + ], + "angle": 0, + "content": "Results. Tab. 3 illustrates the results. Neuralizer-unseen performs similarly to Neuralizer-seen and the baseline, while not requiring lengthy re-training or fine-tuning on the Hammers Atlas dataset, and not having seen the segmentation protocol. All three models achieve a mean Dice coefficient of 0.84. The largest performance difference is in the third ventricle class, where both Neuralizer models outperform the baseline by at least 0.04 Dice. The Freesurfer segmentation protocol included in the training set of the Neuralizer models also contains a third ventricle class." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.801, + 0.613, + 0.816 + ], + "angle": 0, + "content": "5. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Our experiments using modality and segmentation class hold-outs show that Neuralizer can generalize well to unseen neuroimaging tasks. Across all context set sizes, the generalization loss between seen and unseen modalities and segmentation classes is less than 2 percentage points across" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6223" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.08, + 0.089, + 0.894, + 0.187 + ], + "angle": 0, + "content": "
ModelTask SeenSegmentation Class (Hammers Atlas)Mean (std)
HipPAGSTGMIGFuGStmInsPCGThaCC3VPrGPoGALG
Baseline-seen✓.88.86.93.92.79.87.82.87.90.80.68.83.77.82.84 (.07)
Neuralizer-seen✓.88.86.92.92.76.88.83.85.90.82.73.86.77.81.84 (.07)
Neuralizer-unseenX.88.87.93.91.78.87.82.85.90.81.72.85.78.81.84 (.06)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.197, + 0.893, + 0.227 + ], + "angle": 0, + "content": "Table 3. Segmentation of the Hammers Atlas dataset. For Neuralizer-unseen, this dataset and segmentation protocol is withheld from training. Evaluation of major labels located in the center and right of the coronal slice. See Supplement F for class abbreviations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.253, + 0.468, + 0.344 + ], + "angle": 0, + "content": "Experiments 1 and 2. On the smaller held-out HammersAtlas segmentation dataset, we find that Neuralizer can generalize to unseen tasks with similar performance. These results show promise that a single Neuralizer model can perform multiple neuroimaging tasks, including generalization to new inference tasks not seen during training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.346, + 0.468, + 0.512 + ], + "angle": 0, + "content": "In settings with 32 or fewer labeled example images, Neuralizer-unseen outperforms task-specific baselines despite never having seen the task or modality at train time, and taking nearly no effort or compute compared to the baselines which require substantial expertise, manual labor, and compute resources. The performance difference is largest when only one labeled subject is available, but still present at 32 subjects (Fig. 5). Neuralizer provides a performance advantage on smaller datasets likely by exploiting neuroimaging similarities across the many other neuroimaging tasks and datasets available in training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.513, + 0.468, + 0.709 + ], + "angle": 0, + "content": "When training the baselines on all available data, they can outperform Neuralizer-seen and Neuralizer-unseen by at most 3 percentage points. The inflection point of identical performance between Neuralizer and the baselines is not covered by the range of context set sizes chosen for training and evaluation due to prohibitive computational costs and is an interesting direction of study. When large annotated datasets are available, the baselines performed best on most tasks. However, training task-specific models comes at a significant cost. As a first step in the proposed problem formulation, Neuralizer offers an alternative with near equal performance, while only requiring seconds to infer any task from the context set." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.723, + 0.173, + 0.739 + ], + "angle": 0, + "content": "Limitations" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.748, + 0.468, + 0.869 + ], + "angle": 0, + "content": "We made simplifying assumptions in this first paper demonstrating the potential of multi-task generalization in neuroimaging. The experiments are conducted on 2D data slices. In large part, we did this since running the hundreds of baselines in 3D would be infeasible on our compute cluster. Entire volumetric data also impose a challenging memory requirement on Neuralizer models. To tackle 3D data in the future, we plan to process multiple slices at a time." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.468, + 0.901 + ], + "angle": 0, + "content": "We affinely aligned the neuroimages of the context set to the target image. Early in Neuralizer development, we tried" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.253, + 0.892, + 0.328 + ], + "angle": 0, + "content": "training on non-aligned inputs but found that it deteriorated performance. The need for affine alignment provides an obstacle to adoption. While existing affine-alignment tools are fast and can be employed, we also believe that this requirement can be removed with further development." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.328, + 0.892, + 0.418 + ], + "angle": 0, + "content": "We originally experimented with tumor and lesion segmentation tasks but found this to be a more challenging scenario. Lesions are spatially heterogeneous, making learning from the context set much harder. We excluded tumor and lesion segmentation masks from the experiments, but plan to study this setting in the future." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.419, + 0.892, + 0.508 + ], + "angle": 0, + "content": "While we demonstrate the proposed ideas on a broad range of tasks and modalities, neuroimage analysis can involve more domains, tasks, and populations, like image registration, surface-based tasks, CT image domains, and pediatric data. We plan to extend Neuralizer to tackle these in the future." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.521, + 0.619, + 0.537 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.547, + 0.892, + 0.727 + ], + "angle": 0, + "content": "Neuralizer performs accurate rapid single-pass, multitask generalization, and even outperforms task-specific baselines in limited data scenarios. Even when a large amount of annotated data is available, Neuralizer often matches baseline performance despite not training on the data. Neuralizer provides clinical researchers and scientists with a single model to solve a wide range of neuroimaging tasks on images of many modalities, and can be easily adapted to new tasks without the prohibitive requirement of retraining or fine-tuning a task-specific model. We believe this will facilitate neuroscience analyses not currently possible." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.74, + 0.66, + 0.758 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This work was performed during Steffen Czolbes's visit to the Athinoula A. Martins Center for Biomedical Imaging, MGH. The collaboration was supported by the EliteForsk Travel Grant of the Danish Ministry of Higher Education and Science. The work was further funded by the Novo Nordisk Foundation (grants no. NNF20OC0062606 and NNF17OC0028360), the Lundbeck Foundation (grant no. R218-2016-883), and NIH grants 1R01AG064027 and 1R01AG070988." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6224" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.116, + 0.221, + 0.128 + ], + "angle": 0, + "content": "[1] IXI Dataset. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.131, + 0.469, + 0.185 + ], + "angle": 0, + "content": "[2] David Abramian and Anders Eklund. Refacing: Reconstructing anonymized facial features using GANS. International Symposium on Biomedical Imaging, 2019-April:1104-1108,apr 2019.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.095, + 0.187, + 0.469, + 0.282 + ], + "angle": 0, + "content": "[3] Yasser Ad-Dab'bagh, O Lyttelton, JS Muehlboeck, C Lepage, D Einarson, K Mok, O Ivanov, RD Vincent, J Lerch, and E Fombonne. The CIVET image-processing environment: a fully automated comprehensive pipeline for anatomical neuroimaging research. In Proceedings of the 12th annual meeting of the organization for human brain mapping, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.285, + 0.469, + 0.366 + ], + "angle": 0, + "content": "[4] Laith Alzubaidi, Muthana Al-Amidie, Ahmed Al-Asadi, Amjad J. Humaidi, Omran Al-Shamma, Mohammed A. Fadhel, Jinglan Zhang, J. Santamaría, and Ye Duan. Novel Transfer Learning Approach for Medical Imaging with Limited Labeled Data. Cancers 2021, Vol. 13, Page 1590, 13(7):1590, mar 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.369, + 0.469, + 0.437 + ], + "angle": 0, + "content": "[5] Vincent Arsigny, Olivier Commowick, Xavier Pennec, and Nicholas Ayache. A log-euclidean framework for statistics on diffeomorphisms. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 924-931. Springer Verlag, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.439, + 0.468, + 0.465 + ], + "angle": 0, + "content": "[6] John Ashburner. A fast diffeomorphic image registration algorithm. NeuroImage, 38(1):95-113, oct 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.468, + 0.469, + 0.535 + ], + "angle": 0, + "content": "[7] B. B. Avants, C. L. Epstein, M. Grossman, and J. C. Gee. Symmetric diffeomorphic image registration with cross-correlation: Evaluating automated labeling of elderly and neurodegenerative brain. Medical Image Analysis, 12(1):26-41, feb 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.538, + 0.469, + 0.62 + ], + "angle": 0, + "content": "[8] Spyridon Bakas, Hamed Akbari, Aristeidis Sotiras, Michel Bilello, Martin Rozycki, Justin S. Kirby, John B. Freymann, Keyvan Farahani, and Christos Davatzikos. Advancing The Cancer Genome Atlas glioma MRI collections with expert segmentation labels and radiomic features. Scientific Data, 4, sep 2017. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.622, + 0.469, + 0.689 + ], + "angle": 0, + "content": "[9] Spyridon Bakas, Bjoern Menze, and Others. Identifying the Best Machine Learning Algorithms for Brain Tumor Segmentation, Progression Assessment, and Overall Survival Prediction in the BRATS Challenge. arXiv, 124, nov 2018. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.692, + 0.469, + 0.759 + ], + "angle": 0, + "content": "[10] Guha Balakrishnan, Amy Zhao, Mert R. Sabuncu, John Guttag, and Adrian V. Dalca. VoxelMorph: A Learning Framework for Deformable Medical Image Registration. IEEE Transactions on Medical Imaging, 38(8):1788-1800, feb 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.762, + 0.469, + 0.842 + ], + "angle": 0, + "content": "[11] Cheng Bian, Chenglang Yuan, Kai Ma, Shuang Yu, Dong Wei, and Yefeng Zheng. Domain Adaptation Meets Zero-Shot Learning: An Annotation-Efficient Approach to Multi-Modality Medical Image Segmentation. IEEE Transactions on Medical Imaging, 41(5):1043–1056, may 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.846, + 0.468, + 0.9 + ], + "angle": 0, + "content": "[12] Benjamin Billot, Douglas Greve, Koen Van Leemput, Bruce Fischl, Juan Eugenio Iglesias, and Adrian V Dalca. A learning strategy for contrast-agnostic mri segmentation. arXiv preprint arXiv:2003.01995, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.116, + 0.469, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.093, + 0.892, + 0.161 + ], + "angle": 0, + "content": "[13] Benjamin Billot, Douglas N Greve, Oula Puonti, Axel Thielscher, Koen Van Leemput, Bruce Fischl, Adrian V Dalca, and Juan Eugenio Iglesias. SynthSeg: Domain Randomisation for Segmentation of Brain Scans of any Contrast and Resolution. jul 2021. 1, 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.164, + 0.892, + 0.216 + ], + "angle": 0, + "content": "[14] Maxime Bucher, Tuan-Hung VU, Matthieu Cord, and Patrick Pérez. Zero-Shot Semantic Segmentation. Advances in Neural Information Processing Systems, 32, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.219, + 0.892, + 0.273 + ], + "angle": 0, + "content": "[15] Victor Ion Butoi, Jose Javier Gonzalez Ortiz, Tianyu Ma, Mert Sabuncu, John Guttag, and Adrian V. Dalca. Universeg: Universal medical image segmentation. arXiv preprint, 2023. 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.275, + 0.892, + 0.355 + ], + "angle": 0, + "content": "[16] Víctor M. Campello, Carlos Martín-Isla, Cristian Izquierdo, Steffen E. Petersen, Miguel A.González Ballester, and Karim Lekadir. Combining Multi-Sequence and Synthetic Images for Improved Segmentation of Late Gadolinium Enhancement Cardiac MRI. Lecture Notes in Computer Science, 12009 LNCS:290-299, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.358, + 0.892, + 0.397 + ], + "angle": 0, + "content": "[17] Rich Caruana, Lorien Pratt, and Sebastian Thrun. Multitask Learning. Machine Learning 1997 28:1, 28(1):41-75, 1997. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.4, + 0.892, + 0.453 + ], + "angle": 0, + "content": "[18] Hao Chen, Qi Dou, Lequan Yu, Jing Qin, and Pheng Ann Heng. VoxResNet: Deep voxelwise residual networks for brain segmentation from 3D MR images. NeuroImage, 170:446-455, apr 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.456, + 0.892, + 0.482 + ], + "angle": 0, + "content": "[19] Sihong Chen, Kai Ma, and Yefeng Zheng. Med3D: Transfer Learning for 3D Medical Image Analysis. apr 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.484, + 0.892, + 0.524 + ], + "angle": 0, + "content": "[20] Steffen Czolbe, Oswin Krause, and Aasa Feragen. Semantic similarity metrics for learned image registration. Proceedings of Machine Learning Research, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.526, + 0.892, + 0.58 + ], + "angle": 0, + "content": "[21] Adrian V Dalca, Katherine L Bouman, William T Freeman, Natalia S Rost, Mert R Sabuncu, and Polina Golland. Medical image imputation from image collections. IEEE transactions on medical imaging, 38(2):504-514, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.582, + 0.892, + 0.649 + ], + "angle": 0, + "content": "[22] Adrian V Dalca, John Guttag, and Mert R Sabuncu. Anatomical priors in convolutional networks for unsupervised biomedical segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 9290-9299, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.651, + 0.892, + 0.747 + ], + "angle": 0, + "content": "[23] Adrian V Dalca, Evan Yu, Polina Golland, Bruce Fischl, Mert R Sabuncu, and Juan Eugenio Iglesias. Unsupervised deep learning for bayesian brain mri segmentation. In Medical Image Computing and Computer Assisted Intervention—MICCAI 2019: 22nd International Conference, Shenzhen, China, October 13–17, 2019, Proceedings, Part III 22, pages 356–365. Springer, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.749, + 0.892, + 0.802 + ], + "angle": 0, + "content": "[24] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In Conference on Computer Vision and Pattern Recognition, pages 248-255, 2009. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.804, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[25] Francisco Javier Díaz-Pernas, Mario Martínez-Zarzuela, David González-Ortega, and Míriam Antón-Rodríguez. A Deep Learning Approach for Brain Tumor Classification and Segmentation Using a Multiscale Convolutional Neural Network. Healthcare, 9(2):153, feb 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.874, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[26] Theodoras Evgeniou and Massimiliano Pontil. Regularized multi-task learning. Proceedings of the Tenth ACM" + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.956 + ], + "angle": 0, + "content": "6225" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.119, + 0.093, + 0.468, + 0.12 + ], + "angle": 0, + "content": "SIGKDD International Conference on Knowledge Discovery and Data Mining, pages 109-117, 2004. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.122, + 0.47, + 0.176 + ], + "angle": 0, + "content": "[27] Isabelle Faillenot, Rolf A. Heckemann, Maud Frot, and Alexander Hammers. Macroanatomy and 3D probabilistic atlas of the human insula. NeuroImage, 150:88-98, apr 2017. 4, 7, 24" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.179, + 0.468, + 0.246 + ], + "angle": 0, + "content": "[28] Mirza Faisal Beg, Michael I Miller, Alain Trouvétrouv, and Laurent Younes. Computing Large Deformation Metric Mappings via Geodesic Flows of Diffeomorphisms. International Journal of Computer Vision, 61(2):139-157, 2005. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.25, + 0.468, + 0.317 + ], + "angle": 0, + "content": "[29] Ruiwei Feng, Xiangshang Zheng, Tianxiang Gao, Jintai Chen, Wenzhe Wang, Danny Z. Chen, and Jian Wu. Interactive Few-Shot Learning: Limited Supervision, Better Medical Image Segmentation. IEEE Transactions on Medical Imaging, 40(10):2575-2588, oct 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.32, + 0.468, + 0.374 + ], + "angle": 0, + "content": "[30] Abdur R Feyjie, Reza Azad, Marco Pedersoli, Claude Kauffman, Ismail Ben Ayed, and Jose Dolz. Semi-supervised few-shot learning for medical image segmentation. mar 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.377, + 0.468, + 0.403 + ], + "angle": 0, + "content": "[31] B Fischl. FreeSurfer. NeuroImage, 62 (2), 774-781, 2012. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.406, + 0.468, + 0.46 + ], + "angle": 0, + "content": "[32] VS Fonov, AC Evans, RC McKinstry, CR Almli, and DL Collins. Unbiased nonlinear average age-appropriate brain templates from birth to adulthood. NeuroImage, 47:S102, jul 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.463, + 0.468, + 0.517 + ], + "angle": 0, + "content": "[33] Vladimir Fonov, Alan C. Evans, Kelly Botteron, C. Robert Almli, Robert C. McKinstry, and D. Louis Collins. Unbiased average age-appropriate atlases for pediatric studies. NeuroImage, 54(1):313-327, jan 2011. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.519, + 0.46, + 0.532 + ], + "angle": 0, + "content": "[34] Richard SJ Frackowiak. Human Brain Function. 2004. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.535, + 0.468, + 0.589 + ], + "angle": 0, + "content": "[35] Tal Goldfryd, Shiri Gordon, and Tammy Riklin Raviv. Deep semi-supervised bias field correction of mr images. Proceedings - International Symposium on Biomedical Imaging, 2021-April:1836-1840, apr 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.592, + 0.468, + 0.646 + ], + "angle": 0, + "content": "[36] Enhao Gong, John M. Pauly, Max Wintermark, and Greg Zaharchuk. Deep learning enables reduced gadolinium dose for contrast-enhanced brain MRI. Journal of Magnetic Resonance Imaging, 48(2):330-340, aug 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.649, + 0.468, + 0.716 + ], + "angle": 0, + "content": "[37] Ioannis S. Gousias, Daniel Rueckert, Rolf A. Heckemann, Leigh E. Dyet, James P. Boardman, A. David Edwards, and Alexander Hammers. Automatic segmentation of brain MRIs of 2-year-olds into 83 regions of interest. NeuroImage, 40(2):672-684, apr 2008. 4, 7, 24" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.719, + 0.468, + 0.813 + ], + "angle": 0, + "content": "[38] Sachin Gupta, Narinder Singh Punn, Sanjay Kumar Sonbhadra, and Sonali Agarwal. MAG-Net: Multi-task Attention Guided Network for Brain Tumor Segmentation and Classification. Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), 13147 LNCS:3-15, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.817, + 0.468, + 0.871 + ], + "angle": 0, + "content": "[39] J Hale, Nakeisha Schimke, and John Hale. Quickshear defacing for neuroimages Attack Graph Generation on HPC Clusters View project Quickshear Defacing for Neuroimages. Frontiers in psychiatry, 21, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.874, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[40] Alexander Hammers, Richard Allom, Matthias J. Koepp, Samantha L. Free, Ralph Myers, Louis Lemieux, Tejal N." + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.093, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.541, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "Mitchell, David J. Brooks, and John S. Duncan. Three-dimensional maximum probability atlas of the human brain, with particular reference to the temporal lobe. Human brain mapping, 19(4):224-247, aug 2003. 4, 5, 7, 24" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.149, + 0.892, + 0.23 + ], + "angle": 0, + "content": "[41] Xu Han, Roland Kwitt, Stephen Aylward, Spyridon Bakas, Bjoern Menze, Alexander Asturias, Paul Vespa, John Van Horn, and Marc Niethammer. Brain extraction from normal and pathological images: A joint PCA/Image-Reconstruction approach. NeuroImage, 176:431-445, aug 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.233, + 0.892, + 0.299 + ], + "angle": 0, + "content": "[42] Stine Hansen, Srishti Gautam, Robert Jenssen, and Michael Kampffmeyer. Anomaly Detection-Inspired Few-Shot Medical Image Segmentation Through Self-Supervision With Supervoxels. Medical Image Analysis, 78, mar 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.302, + 0.892, + 0.356 + ], + "angle": 0, + "content": "[43] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.359, + 0.892, + 0.384 + ], + "angle": 0, + "content": "[44] Dan Hendrycks and Kevin Gimpel. Gaussian Error Linear Units (GELUs). jun 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.386, + 0.892, + 0.453 + ], + "angle": 0, + "content": "[45] Malte Hoffmann, Benjamin Billot, Douglas N. Greve, Juan Eugenio Iglesias, Bruce Fischl, and Adrian V. Dalca. SynthMorph: learning contrast-invariant registration without acquired images. IEEE Transactions on Medical Imaging, apr 2020. 1, 2, 3, 19" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.456, + 0.892, + 0.522 + ], + "angle": 0, + "content": "[46] Noah Hollmann, Samuel Müller, Katharina Eggensperger, and Frank Hutter. TabPFN: A Transformer That Solves Small Tabular Classification Problems in a Second. In NeurIPS 2022 First Table Representation Workshop, jul 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.525, + 0.892, + 0.579 + ], + "angle": 0, + "content": "[47] Andrew Hoopes, Malte Hoffmann, Douglas N Greve, Bruce Fischl, John Guttag, and Adrian V Dalca. Learning the effect of registration hyperparameters with hypermorph. arXiv preprint arXiv:2203.16680, 2022. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.581, + 0.892, + 0.649 + ], + "angle": 0, + "content": "[48] Andrew Hoopes, Juan Eugenio Iglesias, Bruce Fischl, Douglas Greve, and Adrian V Dalca. TopoFit: Rapid Reconstruction of Topologically-Correct Cortical Surfaces. Proceedings of Machine Learning Research-Under Review, pages 1-13, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.651, + 0.892, + 0.691 + ], + "angle": 0, + "content": "[49] Andrew Hoopes, Jocelyn S. Mora, Adrian V. Dalca, Bruce Fischl, and Malte Hoffmann. SynthStrip: Skull-Stripping for Any Brain Image. mar 2022. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.693, + 0.892, + 0.719 + ], + "angle": 0, + "content": "[50] Minyoung Huh, Pulkit Agrawal, and Alexei A. Efros. What makes ImageNet good for transfer learning? aug 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.721, + 0.892, + 0.761 + ], + "angle": 0, + "content": "[51] Mark Jenkinson, Christian F. Beckmann, Timothy E.J. Behrens, Mark W. Woolrich, and Stephen M. Smith. FSL. NeuroImage, 62(2):782-790, aug 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.763, + 0.892, + 0.802 + ], + "angle": 0, + "content": "[52] Diederik P. Kingma and Jimmy Lei Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations, dec 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.804, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[53] Jens Kleesiek, Gregor Urban, Alexander Hubert, Daniel Schwarz, Klaus Maier-Hein, Martin Bendszus, and Armin Biller. Deep MRI brain extraction: A 3D convolutional neural network for skull stripping. NeuroImage, 129:460–469, apr 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.874, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[54] Arno Klein, Jesper Andersson, Babak A. Ardekani, John Ashburner, Brian Avants, Ming Chang Chiang, Gary E." + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6226" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.118, + 0.092, + 0.47, + 0.188 + ], + "angle": 0, + "content": "Christensen, D. Louis Collins, James Gee, Pierre Hellier, Joo Hyun Song, Mark Jenkinson, Claude Lepage, Daniel Rueckert, Paul Thompson, Tom Vercauteren, Roger P. Woods, J. John Mann, and Ramin V. Parsey. Evaluation of 14 nonlinear deformation algorithms applied to human brain MRI registration. NeuroImage, 46(3):786-802, jul 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.19, + 0.472, + 0.259 + ], + "angle": 0, + "content": "[55] Simon Kornblith, Jonathon Shlens, and Quoc V Le Google Brain. Do Better ImageNet Models Transfer Better? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2661-2671, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.261, + 0.473, + 0.496 + ], + "angle": 0, + "content": "[56] Hugo J. Kuijf, Adrià Casamitjana, D. Louis Collins, Mahsa Dadar, Achilleas Georgiou, Mohsen Ghafoorian, Dakai Jin, April Khademi, Jesse Knight, Hongwei Li, Xavier Llado, J. Matthijs Biesbroek, Miguel Luna, Qaiser Mahmood, Richard Mckinley, Alireza Mehrtash, Sebastien Ourselin, Bo Yong Park, Hyunjin Park, Sang Hyun Park, Simon Pezold, Elodie Puybureau, Jeroen De Bresser, Leticia Rittner, Carole H. Sudre, Sergi Valverde, Veronica Vilaplana, Roland Wiest, Yongchao Xu, Ziyue Xu, Guodong Zeng, Jianguo Zhang, Guoyan Zheng, Rutger Heinen, Christopher Chen, Wiesje Van Der Flier, Frederik Barkhof, Max A. Viergever, Geert Jan Biessels, Simon Andermatt, Mariana Bento, Matt Berseth, Mikhail Belyaev, and M. Jorge Cardoso. Standardized Assessment of Automatic Segmentation of White Matter Hyperintensities and Results of the WMH Segmentation Challenge. IEEE Transactions on Medical Imaging, 38(11):2556-2568, nov 2019. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.497, + 0.473, + 0.691 + ], + "angle": 0, + "content": "[57] Manoj Kumar, Michael J. Anderson, James W. Antony, Christopher Baldassano, Paula P. Brooks, Ming Bo Cai, Po-Hsuan Cameron Chen, Cameron T. Ellis, Gregory Henselman-Petrusek, David Huberdeau, J. Benjamin Hutchinson, Y. Peeta Li, Qihong Lu, Jeremy R. Manning, Anne C. Mennen, Samuel A. Nastase, Hugo Richard, Anna C. Schapiro, Nicolas W. Schuck, Michael Shvartsman, Narayanan Sundaram, Daniel Suo, Javier S. Turek, David Turner, Vy A. Vo, Grant Wallace, Yida Wang, Jamal A. Williams, Hejia Zhang, Xia Zhu, Mihai Capota\\(^{\\text{一}}\\), Jonathan D. Cohen, Uri Hasson, Kai Li, Peter J. Ramadge, Nicholas B. Turk-Browne, Theodore L. Willke, and Kenneth A. Norman. BrainIAK: The Brain Imaging Analysis Kit. Aperture Neuro, 2021(4), jan 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.692, + 0.472, + 0.788 + ], + "angle": 0, + "content": "[58] Sonia Laguna, Riana Schleicher, Benjamin Billot, Pamela Schaefer, Brenna Mckaig, Joshua N Goldstein, Kevin N Sheth, Matthew S Rosen, W Taylor Kimberly, and Juan Eugenio Iglesias. Super-resolution of portable low-field MRI in real scenarios: integration with denoising and domain adaptation. Medical Imaging with Deep Learning (MIDL), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.789, + 0.472, + 0.844 + ], + "angle": 0, + "content": "[59] Erik G Learned-Miller and Parvez Ahammad. Joint MRI Bias Removal Using Entropy Minimization Across Images. Advances in Neural Information Processing Systems, 17, 2004. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.845, + 0.472, + 0.901 + ], + "angle": 0, + "content": "[60] Jaein Lee, Eunsong Kang, Eunjin Jeon, and Heung Il Suk. Meta-modulation Network for Domain Generalization in Multi-site fMRI Classification. Lecture Notes in Computer Science, 12905 LNCS:500-509, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.092, + 0.473, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.092, + 0.894, + 0.175 + ], + "angle": 0, + "content": "[61] Gen Li, Varun Jampani, Laura Sevilla-Lara, Deqing Sun, Jonghyun Kim, and Joongkyu Kim. Adaptive Prototype Learning and Allocation for Few-Shot Segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8334-8343, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.177, + 0.895, + 0.245 + ], + "angle": 0, + "content": "[62] Xiaoxiao Li, Yufeng Gu, Nicha Dvornek, Lawrence H. Staib, Pamela Ventola, and James S. Duncan. Multi-site fMRI analysis using privacy-preserving federated learning and domain adaptation: ABIDE results. Medical Image Analysis, 65:101765, oct 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.247, + 0.895, + 0.316 + ], + "angle": 0, + "content": "[63] Xiang Li, Tianhan Wei, Yau Pun Chen, Yu-Wing Tai, and Chi-Keung Tang. FSS-1000: A 1000-Class Dataset for Few-Shot Segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2869–2878, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.317, + 0.895, + 0.635 + ], + "angle": 0, + "content": "[64] Sook-Lei Liew, Bethany P. Lo, Miranda R. Donnelly, Artemis Zavaliangos-Petropulu, Jessica N. Jeong, Giuseppe Barisano, Alexandre Hutton, Julia P. Simon, Julia M. Juliano, Anisha Suri, Zhizhuo Wang, Aisha Abdullah, Jun Kim, Tyler Ard, Nerisa Banaj, Michael R. Borich, Lara A. Boyd, Amy Brodtmann, Cathrin M. Buetefisch, Lei Cao, Jessica M. Cassidy, Valentina Ciullo, Adriana B. Conforto, Steven C. Cramer, Rosalia Dacosta-Aguayo, Ezequiel de la Rosa, Martin Domin, Adrienne N. Dula, Wuwei Feng, Alexandre R. Franco, Fatemeh Geranmayeh, Alexandre Gramfort, Chris M. Gregory, Colleen A. Hanlon, Brenton G. Hordacre, Steven A. Kautz, Mohamed Salah Khlif, Hosung Kim, Jan S. Kirschke, Jingchun Liu, Martin Lotze, Bradley J. MacIntosh, Maria Mataró, Feroze B. Mohamed, Jan E. Nordvik, Gilsoon Park, Amy Pienta, Fabrizio Piras, Shane M. Redman, Kate P. Revill, Mauricio Reyes, Andrew D. Robertson, Na Jin Seo, Surjo R. Soekadar, Gianfranco Spalletta, Alison Sweet, Maria Telenczuk, Gregory Thielman, Lars T. Westlye, Carolee J. Winstein, George F. Wittenberg, Kristin A. Wong, and Chunshui Yu. A large, curated, open-source stroke neuroimaging dataset to improve lesion segmentation algorithms. Scientific Data, 9(1):1-12, jun 2022. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.635, + 0.892, + 0.663 + ], + "angle": 0, + "content": "[65] Jialin Liu, Fei Chao, and Chih-Min Lin. Task Augmentation by Rotating for Meta-Learning. feb 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.664, + 0.894, + 0.733 + ], + "angle": 0, + "content": "[66] Ming-Yu Liu, Xun Huang, Arun Mallya, Tero Karras, Timo Aila, Jaakko Lehtinen, and Jan Kautz. Few-Shot Unsupervised Image-to-Image Translation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10551–10560, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.734, + 0.894, + 0.802 + ], + "angle": 0, + "content": "[67] Weide Liu, Chi Zhang, Guosheng Lin, and Fayao Liu. CR-Net: Cross-Reference Networks for Few-Shot Segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4165–4173, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.803, + 0.894, + 0.871 + ], + "angle": 0, + "content": "[68] Xiaoxiao Liu, Marc Niethammer, Roland Kwitt, Nikhil Singh, Matt McCormick, and Stephen Aylward. Low-Rank Atlas Image Analyses in the Presence of Pathologies. IEEE Transactions on Medical Imaging, 34(12):2583-2591, dec 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.873, + 0.894, + 0.903 + ], + "angle": 0, + "content": "[69] Xiaofeng Liu, Fangxu Xing, Chao Yang, C. C.Jay Kuo, Georges El Fakhri, and Jonghye Woo. Symmetric-" + }, + { + "type": "list", + "bbox": [ + 0.51, + 0.092, + 0.895, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6227" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.118, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "Constrained Irregular Structure Inpainting for Brain MRI Registration with Tumor Pathology. Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), 12658 LNCS:80-91, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.163, + 0.47, + 0.218 + ], + "angle": 0, + "content": "[70] Michael Lustig, David L. Donoho, Juan M. Santos, and John M. Pauly. Compressed sensing MRI: A look at how CS can improve on current imaging techniques. IEEE Signal Processing Magazine, 25(2):72-82, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.219, + 0.47, + 0.288 + ], + "angle": 0, + "content": "[71] Kushagra Mahajan, Monika Sharma, and Lovekesh Vig. Meta-DermDiagnosis: Few-Shot Skin Disease Identification Using Meta-Learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 730-731, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.289, + 0.47, + 0.343 + ], + "angle": 0, + "content": "[72] José V. Manjón, Pierrick Coup, Antonio Buades, Vladimir Fonov, D. Louis Collins, and Montserrat Robles. Non-local MRI upsampling. Medical Image Analysis, 14(6):784-792, dec 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.344, + 0.472, + 0.4 + ], + "angle": 0, + "content": "[73] José V. Manjón, Pierrick Coupe, Antonio Buades, D. Louis Collins, and Montserrat Robles. MRI superresolution using self-similarity and image priors. International Journal of Biomedical Imaging, 2010, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.401, + 0.47, + 0.483 + ], + "angle": 0, + "content": "[74] Daniel S Marcus, Tracy H Wang, Jamie Parker, John G Csernansky, John C Morris, and Randy L Buckner. Open Access Series of Imaging Studies (OASIS): Cross-Sectional MRI Data in Young, Middle Aged, Nondemented, and Demented Older Adults. Journal of Cognitive Neuroscience, 19:1498-1507, 2007. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.485, + 0.47, + 0.803 + ], + "angle": 0, + "content": "[75] Bjoern H. Menze, Andras Jakab, Stefan Bauer, Jayashree Kalpathy-Cramer, Keyvan Farahani, Justin Kirby, Yuliya Burren, Nicole Porz, Johannes Slotboom, Roland Wiest, Lerente Lanczi, Elizabeth Gerstner, Marc Andre Weber, Tal Arbel, Brian B. Avants, Nicholas Ayache, Patricia Buendia, D. Louis Collins, Nicolas Cordier, Jason J. Corso, Antonio Criminisi, Tilak Das, Hervé Delingette, Cagatay Demiralp, Christopher R. Durst, Michel Dojat, Senan Doyle, Joana Festa, Florence Forbes, Ezequiel Geremia, Ben Glocker, Polina Golland, Xiaotao Guo, Andac Hamamci, Khan M. Iftekharuddin, Raj Jena, Nigel M. John, Ender Konukoglu, Danial Lashkari, José Antônio Mariz, Raphael Meier, Sérgio Pereira, Doina Precup, Stephen J. Price, Tammy Riklin Raviv, Syed M.S. Reza, Michael Ryan, Duygu Sarikaya, Lawrence Schwartz, Hoo Chang Shin, Jamie Shotton, Carlos A. Silva, Nuno Sousa, Nagesh K. Subbanna, Gabor Szekely, Thomas J. Taylor, Owen M. Thomas, Nicholas J. Tustison, Gozde Unal, Flor Vasseur, Max Wintermark, Dong Hye Ye, Liang Zhao, Binsheng Zhao, Darko Zikic, Marcel Prastawa, Mauricio Reyes, and Koen Van Leemput. The Multimodal Brain Tumor Image Segmentation Benchmark (BRATS). IEEE Transactions on Medical Imaging, 34(10):1993-2024, oct 2015. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.804, + 0.47, + 0.872 + ], + "angle": 0, + "content": "[76] Fausto Milletari, Nassir Navab, and Seyed Ahmad Ahmadi. V-Net: Fully convolutional neural networks for volumetric medical image segmentation. Proceedings - 2016 4th International Conference on 3D Vision, 3DV 2016, pages 565-571, dec 2016. 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.873, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[77] Samuel Müller, Noah Hollmann, Sebastian Pineda Arango, Josif Grabocka, and Frank Hutter. Transformers Can Do" + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.092, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.541, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "Bayesian Inference. In International Conference on Learning Representations, mar 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.122, + 0.892, + 0.189 + ], + "angle": 0, + "content": "[78] Bao Nguyen, Adam Feldman, Sarath Bethapudi, Andrew Jennings, and Chris G. Willcocks. Unsupervised region-based anomaly detection in brain mri with adversarial image inpainting. Proceedings - International Symposium on Biomedical Imaging, 2021-April:1127-1131, apr 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.19, + 0.892, + 0.273 + ], + "angle": 0, + "content": "[79] Dong Nie, Roger Trullo, Jun Lian, Caroline Petitjean, Su Ruan, Qian Wang, and Dinggang Shen. Medical image synthesis with context-aware generative adversarial networks. Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), 10435 LNCS:417-425, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.274, + 0.892, + 0.328 + ], + "angle": 0, + "content": "[80] Alexander F.I. Osman and Nissren M. Tamam. Deep learning-based convolutional neural network for intramodality brain MRI synthesis. Journal of Applied Clinical Medical Physics, 23(4):e13530, apr 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.33, + 0.892, + 0.469 + ], + "angle": 0, + "content": "[81] Moritz Roman Hernandez Petzsche, Ezequiel de la Rosa, Uta Hanning, Roland Wiest, Waldo Enrique Valenzuela Pinilla, Mauricio Reyes, Maria Ines Meyer, Sook-Lei Liew, Florian Kofler, Ivan Ezhov, David Robben, Alexander Hutton, Tassilo Friedrich, Teresa Zarth, Johannes Burkle, The Anh Baran, Bjoern Menze, Gabriel Broocks, Lukas Meyer, Claus Zimmer, Tobias Boeckh-Behrens, Maria Berndt, Benno Ikenberg, Benedikt Wiestler, and Jan S. Kirschke. ISLES 2022: A multi-center magnetic resonance imaging stroke lesion segmentation dataset. jun 2022. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.469, + 0.892, + 0.55 + ], + "angle": 0, + "content": "[82] Kilian M Pohl, Sylvain Bouix, Motoaki Nakamura, Torsten Rohlfing, Robert W McCarley, Ron Kikinis, W Eric L Grimson, Martha E Shenton, and William M Wells. A hierarchical algorithm for mr brain image parcellation. IEEE transactions on medical imaging, 26(9):1201-1212, 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.552, + 0.892, + 0.607 + ], + "angle": 0, + "content": "[83] Maithra Raghu, Chiyuan Zhang, Google Brain, Jon Kleinberg, and Samy Bengio. Transfusion: Understanding Transfer Learning for Medical Imaging. Advances in Neural Information Processing Systems, 32, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.608, + 0.892, + 0.65 + ], + "angle": 0, + "content": "[84] Sachin Ravi and Hugo Larochelle. Optimization as a Model for Few-Shot Learning. In International Conference on Learning Representations (ICLR), jul 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.65, + 0.892, + 0.719 + ], + "angle": 0, + "content": "[85] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, volume 9351, pages 234-241. Springer Verlag, 2015. 1, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.72, + 0.892, + 0.801 + ], + "angle": 0, + "content": "[86] Edgar Schonfeld, Sayna Ebrahimi, Samarth Sinha, Trevor Darrell, and Zeynep Akata. Generalized Zero- and Few-Shot Learning via Aligned Variational Autoencoders. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8247–8255, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.802, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[87] F. Segonne, A. M. Dale, E. Busa, M. Glessner, D. Salat, H. K. Hahn, and B. Fischl. A hybrid approach to the skull stripping problem in MRI. NeuroImage, 22(3):1060-1075, jul 2004. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.86, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[88] Ozan Sener and Vladlen Koltun. Multi-Task Learning as Multi-Objective Optimization. Advances in Neural Information Processing Systems, 31, 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6228" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.093, + 0.468, + 0.134 + ], + "angle": 0, + "content": "[89] Jun Seo, Young-Hyun Park, Sung Whan Yoon, and Jaekyun Moon. Task-Adaptive Feature Transformer with Semantic Enrichment for Few-Shot Segmentation. feb 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.136, + 0.468, + 0.177 + ], + "angle": 0, + "content": "[90] David W. Shattuck and Richard M. Leahy. BrainSuite: An automated cortical surface identification tool. Medical Image Analysis, 6(2):129-142, jun 2002. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.179, + 0.468, + 0.233 + ], + "angle": 0, + "content": "[91] X. Shen, F. Tokoglu, X. Papademetris, and R. T. Constable. Groupwise whole-brain parcellation from resting-state fMRI data for network node identification. NeuroImage, 82:403-415, nov 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.236, + 0.468, + 0.305 + ], + "angle": 0, + "content": "[92] Nalini M Singh, Juan Eugenio Iglesias, Elfar Adalsteinson, Adrian V Dalca, and Polina Golland. Joint Frequency and Image Space Learning for MRI Reconstruction and Analysis. Journal of Machine Learning for Biomedical Imaging, 2022:1-28, 2022. 2, 25" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.307, + 0.468, + 0.348 + ], + "angle": 0, + "content": "[93] Youssef Skandarani, Nathan Painchaud, Pierre-Marc Jodoin, and Alain Lalande. On the effectiveness of GAN generated cardiac MRIs for segmentation. may 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.35, + 0.468, + 0.377 + ], + "angle": 0, + "content": "[94] Stephen M. Smith. Fast robust automated brain extraction. Human Brain Mapping, 17(3):143-155, nov 2002. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.379, + 0.468, + 0.489 + ], + "angle": 0, + "content": "[95] Stephen M. Smith, Mark Jenkinson, Mark W. Woolrich, Christian F. Beckmann, Timothy E.J. Behrens, Heidi Johansen-Berg, Peter R. Bannister, Marilena De Luca, Ivana Drobnjak, David E. Flitney, Rami K. Niazy, James Saunders, John Vickers, Yongyue Zhang, Nicola De Stefano, J. Michael Brady, and Paul M. Matthews. Advances in functional and structural MR image analysis and implementation as FSL. NeuroImage, 23(1), 2004. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.491, + 0.468, + 0.533 + ], + "angle": 0, + "content": "[96] Jake Snell, Kevin Swersky, and Twitter Richard Zemel. Prototypical Networks for Few-shot Learning. Advances in Neural Information Processing Systems, 30, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.535, + 0.468, + 0.617 + ], + "angle": 0, + "content": "[97] Haoliang Sun, Ronak Mehta, Hao H. Zhou, Zhichun Huang, Sterling C. Johnson, Vivek Prabhakaran, and Vikas Singh. DUAL-GLOW: Conditional Flow-Based Generative Model for Modality Transfer. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10611-10620, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.619, + 0.468, + 0.701 + ], + "angle": 0, + "content": "[98] Nathalie Tzourio-Mazoyer, Brigitte Landeau, Dimitri Papathanassiou, Fabrice Crivello, Octave Etard, Nicolas Delcroix, Bernard Mazoyer, and Marc Joliot. Automated anatomical labeling of activations in spm using a macroscopic anatomical parcellation of the mni mri single-subject brain. Neuroimage, 15(1):273-289, 2002. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.704, + 0.468, + 0.787 + ], + "angle": 0, + "content": "[99] N. Tzourio-Mazoyer, B. Landeau, D. Papathanassiou, F. Crivello, O. Etard, N. Delcroix, B. Mazoyer, and M. Joliot. Automated Anatomical Labeling of Activations in SPM Using a Macroscopic Anatomical Parcellation of the MNI MRI Single-Subject Brain. NeuroImage, 15(1):273–289, jan 2002. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.468, + 0.843 + ], + "angle": 0, + "content": "[100] David C. Van Essen, Stephen M. Smith, Deanna M. Barch, Timothy E.J. Behrens, Essa Yacoub, and Kamil Ugurbil. The WU-Minn Human Connectome Project: An overview. NeuroImage, 80:62-79, oct 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.468, + 0.9 + ], + "angle": 0, + "content": "[101] Oriol Vinyals, Google Deepmind, Charles Blundell, Timothy Lillicrap, Koray Kavukcuoglu, and Daan Wierstra. Matching Networks for One Shot Learning. Advances in Neural Information Processing Systems, 29, 2016. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.468, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[102] Oriol Vinyals, Google Deepmind, Charles Blundell, Timothy Lillicrap, Koray Kavukcuoglu, and Daan Wierstra. Matching Networks for One Shot Learning. Advances in Neural Information Processing Systems, 29, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.189 + ], + "angle": 0, + "content": "[103] Christian Wachinger, Martin Reuter, and Tassilo Klein. Deepnat: Deep convolutional neural network for segmenting neuroanatomy. NeuroImage, 170:434-445, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.191, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[104] Qi Wang, Julius Steiglechner, Tobias Lindig, Benjamin Bender, Klaus Scheffler, and Gabriele Lohmann. SuperResolution for Ultra High-Field MR Images. In Medical Imaging with Deep Learning (MIDL), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.247, + 0.892, + 0.287 + ], + "angle": 0, + "content": "[105] Yaqing Wang, Quanming Yao, James T. Kwok, and Lionel M. Ni. Generalizing from a Few Examples. ACM Computing Surveys, 53(3), jun 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.289, + 0.892, + 0.357 + ], + "angle": 0, + "content": "[106] Mark W. Woolrich, Saad Jbabdi, Brian Patenaude, Michael Chappell, Salima Makni, Timothy Behrens, Christian Beckmann, Mark Jenkinson, and Stephen M. Smith. Bayesian analysis of neuroimaging data in FSL. NeuroImage, 45(1), 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.359, + 0.892, + 0.413 + ], + "angle": 0, + "content": "[107] Xiao Yang, Xu Han, Eunbyung Park, Stephen Aylward, Roland Kwitt, and Marc Niethammer. Registration of Pathological Images. Simulation and synthesis in medical imaging (Workshop), 9968:97, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.415, + 0.892, + 0.496 + ], + "angle": 0, + "content": "[108] Huaxiu Yao, Long-Kai Huang, Linjun Zhang, Ying Wei, Li Tian, James Zou, Junzhou Huang, and Zhenhui Li. Improving Generalization in Meta-learning via Task Augmentation. In Proceedings of the 38th International Conference on Machine Learning, PMLR, pages 11887-11897. PMLR, jul 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.498, + 0.892, + 0.579 + ], + "angle": 0, + "content": "[109] Chandan Ganesh Bangalore Yogananda, Benjamin C. Wagner, Gowtham K. Murugesan, Ananth Madhuranthakam, and Joseph A. Maldjian. A deep learning pipeline for automatic skull stripping and brain segmentation. Proceedings - International Symposium on Biomedical Imaging, pages 727-731, apr 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.581, + 0.892, + 0.634 + ], + "angle": 0, + "content": "[110] Lin Yuan, Xue Wei, Hui Shen, Ling Li Zeng, and Dewen Hu. Multi-center brain imaging classification using a novel 3d cnn approach. IEEE Access, 6:49925-49934, sep 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.637, + 0.892, + 0.718 + ], + "angle": 0, + "content": "[111] Chi Zhang, Guosheng Lin, Fayao Liu, Jiushuang Guo, Qingyao Wu, and Rui Yao. Pyramid Graph Networks With Connection Attentions for Region-Based One-Shot Semantic Segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9587-9595, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.72, + 0.892, + 0.788 + ], + "angle": 0, + "content": "[112] Chi Zhang, Guosheng Lin, Fayao Liu, Rui Yao, and Chunhua Shen. CANet: Class-Agnostic Segmentation Networks With Iterative Refinement and Attentive Few-Shot Learning. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 5217-5226, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.79, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[113] Amy Zhao, Guha Balakrishnan, Frédo Durand, John V Guttag, and Adrian V Dalca. Data Augmentation Using Learned Transformations for One-Shot Medical Image Segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8543-8553, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[114] Guoyan Zheng, Chengwen Chu, Daniel L. Belavy, Bulat Ibragimov, Robert Korez, Tomaz Vrtovec, Hugo Hutt," + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.956 + ], + "angle": 0, + "content": "6229" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.118, + 0.092, + 0.473, + 0.219 + ], + "angle": 0, + "content": "Richard Everson, Judith Meakin, Isabel López Andrade, Ben Glocker, Hao Chen, Qi Dou, Pheng Ann Heng, Chunliang Wang, Daniel Forsberg, Aleš Neubert, Jurgen Fripp, Martin Urschler, Darko Stern, Maria Wimmer, Alexey A. Novikov, Hui Cheng, Gabriele Armbrecht, Dieter Felsenberg, and Shuo Li. Evaluation and comparison of 3D intervertebral disc localization and segmentation methods for 3D T2 MR data: A grand challenge. Medical Image Analysis, 35:327-344, jan 2017. 3" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.517, + 0.957 + ], + "angle": 0, + "content": "6230" + } + ] +] \ No newline at end of file diff --git a/2023/Neuralizer_ General Neuroimage Analysis Without Re-Training/eee09609-bc86-4647-98dc-269f47c8002f_origin.pdf b/2023/Neuralizer_ General Neuroimage Analysis Without Re-Training/eee09609-bc86-4647-98dc-269f47c8002f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..1d16eecb569806d53e79e3ddf80edae39b3efa9d --- /dev/null +++ b/2023/Neuralizer_ General Neuroimage Analysis Without Re-Training/eee09609-bc86-4647-98dc-269f47c8002f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:178e093432f43b75669c5b6454554dfabbac086c2cb202483692a9fe598db5c7 +size 950872 diff --git a/2023/Neuralizer_ General Neuroimage Analysis Without Re-Training/full.md b/2023/Neuralizer_ General Neuroimage Analysis Without Re-Training/full.md new file mode 100644 index 0000000000000000000000000000000000000000..dd6dea061d458ad96fe200eb56baaa3e23334307 --- /dev/null +++ b/2023/Neuralizer_ General Neuroimage Analysis Without Re-Training/full.md @@ -0,0 +1,405 @@ +# Neuralizer: General Neuroimage Analysis without Re-Training + +Steffen Czolbe +University of Copenhagen and MGH +per.sc@di.ku.dk + +Adrian V. Dalca +MIT, and MGH, Harvard Medical School +adalca@mit.edu + +# Abstract + +Neuroimage processing tasks like segmentation, reconstruction, and registration are central to the study of neuroscience. Robust deep learning strategies and architectures used to solve these tasks are often similar. Yet, when presented with a new task or a dataset with different visual characteristics, practitioners most often need to train a new model, or fine-tune an existing one. This is a time-consuming process that poses a substantial barrier for the thousands of neuroscientists and clinical researchers who often lack the resources or machine-learning expertise to train deep learning models. In practice, this leads to a lack of adoption of deep learning, and neuroscience tools being dominated by classical frameworks. + +We introduce Neuralizer, a single model that generalizes to previously unseen neuroimaging tasks and modalities without the need for re-training or fine-tuning. Tasks do not have to be known a priori, and generalization happens in a single forward pass during inference. The model can solve processing tasks across multiple image modalities, acquisition methods, and datasets, and generalize to tasks and modalities it has not been trained on. Our experiments on coronal slices show that when few annotated subjects are available, our multi-task network outperforms task-specific baselines without training on the task. + +# 1. Introduction + +Computational methods for the processing and analysis of neuroimages have enabled a deep understanding of the human brain. The field has also led to advanced patient care by facilitating non-invasive methods of diagnosis and treatment. Recent deep learning research promises to substantially increase the accuracy and speed of neuroimaging analysis methods. + +A drawback of most current deep-learning-based approaches is that each model is limited to solving the task it has been trained on, on the type of data it has been trained on. Generalization to new tasks and domains, such as different acquisition protocols or new segmentation, is + +![](images/0d6f1df7db5ce816a8997a97cf48f586d95cde4ec34df20cc3098306dcfeb348.jpg) +Figure 1. Neuralizer can solve a broad range of image processing tasks, including new ones not seen during training, with a single model by conditioning the prediction on a context set of examples. After training on a diverse set of tasks, the model can generalize to new tasks in a single forward pass without re-training or finetuning. The model is highly flexible, requiring no prior definition of the set of tasks, and can be conditioned with context sets of any length. + +a main barrier to adoption [62]. Performing neuroimaging tasks like segmentation, registration, reconstruction, or motion correction requires different models for each processing step, despite operating on the same input data and methods exhibiting strong similarities in network architecture [13,45,85]. Yet, designing and training models to solve these tasks on each dataset is prohibitively expensive. To train a deep learning model, a dataset needs to be compiled and often manually annotated, and the network, training, and data loading logic needs to be implemented. All these steps generally require machine learning and neuroimaging expertise. In addition, computational resources like specialized graphics processing hardware and software infrastructure needs to be available. These requirements are particularly problematic in clinical research settings due to a high cost of annotation and a lack of machine learning expertise and hardware. + +![](images/56d19b5b78f39d78a44fd8928324dd971826bf61499ca0993eaf583905d221ce.jpg) +Binary Segmentation + +![](images/aa387042cbb38ad0a6273668c17cbe0fca1f2f4bec1ba4e7c815bd8d8f33c964.jpg) +Modality Transfer + +![](images/731ce385d19df72b67f1a1a12be3bb813740d5dd54b50cdb32502882386a7c11.jpg) +Super Resolution + +![](images/0b946413a57cb4cd20535159782b092b9f31e4c6dfc29d695bf49be468742941.jpg) +Figure 2. Example neuroimaging tasks and modalities included in our dataset (top: input images, bottom: output images). + +![](images/5f1ad77f626c725a8e3c83f5188002286511cc7e13419b3bd57fe2407e99be61.jpg) +Denoising & Bias Correction +Inpainting + +![](images/ce281d27c4b8fb1d5a5b82c3e4a2b1806493c3e5efd855a9e36a4600e8be7f9c.jpg) +Skull Stripping + +![](images/71ec469392539fb001ae6c6292a29ab52749c354535f3b4befd191f1ded5954b.jpg) +Motion Correction + +![](images/2c1406656e88e80f756c46e6d0610e1127065af82e500c46459793f007c831ca.jpg) +Undersampled Reconstruction + +![](images/05a870d5c09eb569cdf29856d0a5c2fcbf320ea4b4df440c07baa87c25716a7e.jpg) + +# Contribution + +We introduce Neuralizer, a general-purpose neuroimaging model that can solve a broad range of neuroimaging tasks on diverse image modalities (Fig. 2), without the need for task-specific training or fine-tuning. Neuralizer can solve new tasks, unseen during training, using a set of examples of the new task at inference (Fig. 1) + +Neuralizer involves a convolutional architecture (Fig. 3), that takes as input a context set of examples that define the processing task, and thus does not require prior specification of the tasks. The method enables single-pass generalization during inference and can process any number of reference images in a single pass to inform the prediction. + +As a first method tackling task generalization in neuroimaging, we focus on analyzing the capabilities of such system and presenting general insights, and limit our experiments to 2D. We evaluate our model by comparing the single-pass generalization performance to task-specific baselines conditioned on an equivalent amount of data. We find that Neuralizer outperforms the baselines on tasks where $\leq 32$ labeled examples are available, despite never training on the task. When generalizing to new segmentation protocols, Neuralizer matches the performance of baselines trained directly on the dataset. + +# 2. Related Work + +We give a short introduction to neuroimaging tasks, terminology, and methods. We then provide an overview of fundamental methods for adapting a model to multiple domains, including multi-task learning, few-shot learning, fine-tuning, and data synthesis. + +# 2.1. Neuroimage analysis + +Neuroimage analysis employs computational techniques to study the structure and function of the human brain. Common imaging techniques are structural magnetic resonance imaging (MRI), functional MRI, diffusion tensor + +imaging (DTI), computed tomography (CT), and Positron emission tomography (PET). Each imaging method can create diverse images with different characteristics and contrasts, which are further diversified depending on the properties of the acquisition site [60, 110], device, protocol, imaging sequence [56], and use of contrast agents [9,36]. + +To analyze these images, a variety of processing tasks are most often combined in a processing pipeline. Common processing tasks include anatomical segmentation [12, 18, 22, 23, 31, 82, 98, 103], skull stripping [49, 49, 53, 87, 94, 109], defacing [2, 39], registration [5-7, 10, 20, 28, 45, 47, 54], modality transfer [79, 80, 97], in-painting [41, 68, 69, 78, 107], super-resolution [21, 58, 72, 73, 104], reconstruction, and de-noising [58, 70, 92], bias field removal [35, 59], surface fitting [48] and parcellation [91, 99]. + +Multiple toolboxes provide a suite of interoperable software components, most implementing classical optimization strategies. Widely used toolboxes include Freesurfer [31], FSL [51, 95, 106], SPM [34], CIVET [3], BrainSuite [90], HCP pipeline [100], and BrainIAK [57]. Deep-learning-based methods are starting to be included because of their improved accuracy and shorter runtime [13, 49]. While these methods provide solutions for common neuroimaging applications, most are limited to a single task and few modalities. Developers need to manually update the pipelines to include new processing tasks and to support a wider variety of image modalities. This process requires extensive technical expertise and computational resources, often not available to the clinical neuroscientists focusing on scientific questions. + +# 2.2. Multi-task learning + +Multi-Task Learning (MTL) frameworks solve multiple tasks simultaneously by exploiting similarities between related tasks [17]. MTL can improve performance and reduce computational cost and development time compared to designing task-specific solutions [26, 88]. In neuroimaging, MTL networks were recently proposed for the simultaneous + +segmentation and classification of brain tumors by training a single network with separate prediction heads associated with the different tasks [25, 38]. This strategy is challenging to scale as the number of tasks increases, requires prior determination of the set of tasks, and importantly does not enable generalization of the model to new tasks. With Neuralizer, we build on these methods to achieve scalable MTL, without the need for multiple network heads, and importantly with the ability to generalize to new tasks and modalities. + +# 2.3. Fine-tuning + +To tackle problems in the limited data scenarios frequent in medical imaging, neural networks can be pre-trained on a related task with high data availability and then fine-tuned for specific tasks. For example, a common approach involves taking a Res-Net [43] trained on ImageNet [24] and fine-tuning part of the network for a new task [50, 55, 101]. For medical imaging, networks pre-trained on large sets of medical images are available [19], and fine-tuning them to new tasks results in shortened training time and higher accuracy [4, 71]. However, fine-tuning also requires machine learning expertise and computational resources, most often not available in clinical research. Additionally, in scenarios with small datasets, fine-tuning models trained on large vision datasets can be harmful [83]. + +# 2.4. Few-shot learning + +Few-shot models generate predictions from just a few labeled examples [66, 84, 86, 105], or in the case of zero-shot methods [14], none at all. Many of these methods require training or fine-tuning. In computer vision, several methods pass a query image, along with a set of support images and labels as input to the model [66, 89, 96, 102]. Natural image segmentation methods [67, 112] use single image-label pairs [61, 111] as support or aggregate information from a larger support set [63]. Recent few-shot learning methods in the medical image segmentation setting [11, 29, 30] operate on a specific anatomical region in a single image modality [42, 114]. Similar Prior-Data Fitted Networks (PFNs) are fitted to multiple datasets at once to learn the training and prediction algorithm [77]. During training, this strategy draws a dataset, a set of data points and their labels from it, masks one of the labels and predicts it. The resulting model aims to generalize to new datasets. PFNs have only been applied to low-dimensional and tabular data [46]. Our solution builds on ideas from these methods, but aims to solve a much larger range of diverse image-to-image tasks on neuroimages of many modalities. + +# 2.5. Data augmentation and synthesis + +Data augmentation increases the diversity of training data by augmenting or modifying existing data [85, 113]. It + +improves model robustness to input variability that may not be available in the original training data. In neuroimaging, arbitrary image modalities can be simulated by synthesis of images without requiring any real data [13, 16, 45, 49, 93]. In meta-learning, data augmentation can further be used to generate entirely new tasks [15, 65, 108]. We use data augmentations and further expand existing methods by developing rich neuroimaging task augmentations for generalization to unseen neuroimaging tasks. + +# 3. Neuralizer + +We introduce Neuralizer, a multi-task model for neuroimage analysis tasks. In this section, we first define the training framework and adaptations necessary to operate on a diverse range of tasks and input types. We then introduce the model architecture, training, and inference strategies. + +# 3.1. Generalizable multi-task model + +Let $T$ represent a set of tasks, with a subset of tasks $T_{\mathrm{seen}}$ seen during training. Each task consists of input-output image pairs $(x_{t}, y_{t})$ from potentially multiple underlying datasets with input and output spaces $x_{t} \in \mathcal{X}, y_{t} \in \mathcal{Y}$ . + +To enable generalization to unseen tasks, we condition the model on a context set $C_t = \{(x_{t,i},y_{t,i})\}_{i=1}^N$ of input-output image pairs passed to the model alongside the prediction task. The context set defines the desired task, and can vary in size $|C_t| = N$ and is re-sampled from the underlying task-datasets for each input. Fig. 1 gives an example for a modality transfer task. + +We employ a neural network $g_{\theta}(x_t, C_t) = y_t$ with weights $\theta$ that applies the task defined by context set $C_t$ to the input neuroimage $x_t$ . We optimize the network using supervised training with the loss + +$$ +\mathcal {L} \left(T _ {\text {s e e n}}; \theta\right) = \mathbb {E} _ {t \in T _ {\text {s e e n}}} \left[ \mathbb {E} _ {\left(x _ {t}, y _ {t}, C _ {t}\right)} \left[ \mathcal {L} _ {t} \left(y _ {t}, g _ {\theta} \left(x _ {t}, C _ {t}\right)\right) \right] \right], \tag {1} +$$ + +where $\mathcal{L}_t$ is a task-specific loss function. + +# 3.2. Design for diverse tasks + +To process different tasks with a single model, we carefully select the loss function, neuroimage encodings, and generation of the training set for each task type. + +Loss function. Neuralizer solves both segmentation tasks (e.g. anatomical segmentation and skull-stripping via a brain mask), more general and image-to-image tasks (e.g. denoising). We use the Soft Dice Loss [76] for segmentation-like tasks, and the pixel-wise Mean Squared Error $MSE(y_{t}, g(x_{t}, C_{t})) = \frac{1}{2\sigma^{2}} \sum_{p} [y_{tp} - g(x_{t}, C_{t})_{p}]^{2}$ with balancing hyperparameter $\sigma^{2}$ for other tasks. As the network optimizes multiple tasks during training, the balance of the loss terms can dramatically affect the optimization and resulting performance. + +![](images/f98a6eaa912876ee02c28a05c983027b5f7e96ca4889be678ceb5cc3377b2e0a.jpg) +Figure 3. Neuralizer consists of 7 Pairwise-Conv-Avg blocks (right), arranged in a U-Net-like [76, 85] configuration (left). Each Pairwise-Conv-Avg block enables interaction between the input image and the image pairs present in the context set. The block consists of a residual unit, pairwise convolution of each context member with the target, and an averaging of results across the context set to update the representation. The architecture is invariant to context size $N$ . + +![](images/cca2e8bdd235ecf870b8147531fe736c1aeb1d4f8d5074958ed0bc830beffdd5.jpg) + +Input and output encoding. For Neuralizer to work on both segmentation and image-to-image tasks, we facilitate simultaneous input of multiple image modalities and masks. We design the input space $\mathcal{X}$ to accept floating point value images with three channels, and zero-pad any channels unnecessary for a specific task. The output space $\mathcal{Y}$ follows the same design but uses only one channel. + +Training dataset creation. At each training iteration, we first sample a task $t$ from $T_{\mathrm{seen}}$ , selecting the task-specific dataset (Tab. 1). From this dataset, we sample the input image, ground truth output, and image pairs for the context set. To increase the range of images that can be used to condition the trained model, the image modalities and acquisition protocols of entries of the context set can differ from the input image for some tasks. Supplemental section G contains a detailed description of the training data generator. + +# 3.3. Model architecture + +Fig. 3 shows the Neuralizer network architecture, adapted with the concurrently developed [15] - a method that focuses on solving broad segmentation tasks. As the architecture is independent of the task, we omit the task subscript in this section. + +The input image $x$ and the image pairs of the context set $C_i = (x_i, y_i), i = 1, \dots, N$ are first passed through an embedding layer consisting of a single $1 \times 1$ convolution with learnable kernels $e_x, e_C$ , to obtain the representations $r_x = x * e_x$ , $r_{C_i} = \operatorname{cat}(x_i, y_i) * e_C$ where $*$ is the convolution operator. This combines each context image pair to a joint representation $r_{C_i}$ and maps all representations to a uniform channel width $c$ , which is constant throughout the model. Next, we process the representations using multiple Pairwise-Conv-Avg Blocks (explained below), ar + +Table 1. Tasks, Modalities, Datasets, and Segmentation classes used in this paper, and involved in training Neuralizer. + +
TasksModalities
Binary SegmentationT1-w.
Modality TransferT2-w.
Super ResolutionMRA
Skull StrippingPD
Motion CorrectionFLAIR
Undersampled ReconstructionADC
Denoising & Bias correctionDWI
InpaintingDTI (17 dir.)
DatasetsSegmentation Classes
OASIS 3 [47, 74]Freesufer protocol, 31
BRATS [8, 9, 75]classes [13, 31]
IXI [1]
ATLAS R2.0 [64]Manually-annotated Hammers
Hammers Atlas [40]Atlas, 96 classes [27, 37, 40]
WMH Challenge [56]
ISLES2022 [81]Brainmasks [31, 49]
+ +ranged as a U-Net-like configuration [76,85] to exploit multiple scales. The output $r_x^{\mathrm{out}}$ of the final Pairwise-Conv-Avg Block is processed by a residual unit [43] and a final $1 \times 1$ conv layer to map to one output channel. All residual units consist of two $3 \times 3$ conv layers, a shortcut connection, and GELU activation functions [44]. + +Compared to standard CNNs, Neuralizer uses a mechanism to enable knowledge transfer from the context set to the input image. We design the Pairwise-Conv-Avg Block + +(Fig. 3, right) to model this interaction. The block maps from representations of the target input $r_x^{\mathrm{in}}$ and context pairs $r_{C_i}^{\mathrm{in}}$ to output representations $r_x^{\mathrm{out}}$ , $r_{C_i}^{\mathrm{out}}$ of the same size. First, we process each input separately with a residual unit to obtain $r_x^{\mathrm{int}} = \mathrm{ResUnit}_x(r_x^{\mathrm{in}})$ and $r_{C_i}^{\mathrm{int}} = \mathrm{ResUnit}_C(r_{C_i}^{\mathrm{in}})$ . The residual units, which involve two convolutions, operate on the context representations and have shared weights. Second, we pairwise concatenate the context representations with the target representation on the channel dimension: $p_i = \mathrm{cat}(r_x^{\mathrm{int}}, r_{C_i}^{\mathrm{int}})$ . We combine the pairwise representations and reduce the channel size back to $c$ using a $1 \times 1$ convolution with learnable kernel $k_x$ , and update the target representation by averaging across context members $r_x^{\mathrm{out}} = r_x^{\mathrm{int}} + \frac{1}{N}\sum_{i=1}^{N}p_i * k_x$ . The context representations are updated with a separate kernel $r_{C_i}^{\mathrm{out}} = r_{C_i}^{\mathrm{int}} + p_i * k_C$ . We then re-size the outputs of a Pairwise-Conv-Avg Block before feeding them as input the next block. We experimented with attention-based and weighted average approaches but found that they did not lead to an increased generalization to unseen tasks. + +# 3.4. Task augmentations + +To further diversify the training dataset, we employ task augmentations [15], a group of transformations applied at random to the input, output, and context images. The objective is to increase the diversity of tasks to discourage the model from merely memorizing the tasks in the training data. A list of all task augmentations is summarized in Tab. 2, with more detailed descriptions and visual examples in Supplement C. + +# 3.5. Inference + +During inference, we supply an input image $x_{i}$ and a context set $C_i$ from the desired task. Given these inputs, a simple feed-forward pass through the model provides the prediction $\hat{y} = g(x,C)$ . To further increase accuracy at test-time, we use context-set bootstrapping [15]. We also increase the context set by sampling with replacement from the context set, and add small affine augmentations. + +# 4. Experiments + +We first compare Neuralizer with task-specific networks, which require substantial expertise and compute. We then analyze the effect of the size of the context set, and the + +Table 2. Task Augmentations + +
Task Augmentations
IntensityMappingSyntheticModality
SobelFilterMaskInvert
MaskContourMaskDilation
PermuteChannelsDuplicateChannels
+ +multi-task generalization to unseen segmentation protocols and image modalities. For this first method of large-scale multi-task generalization in neuroimaging, we conduct the experiments on 2D image slices. + +# 4.1. Data + +To create a diverse dataset encompassing a multitude of different modalities, acquisition protocols, devices, and tasks, we pool neuroimages from the public datasets OASIS3 [47, 74], BRATS [8, 9, 75], Atlas R2.0 [64], Hammers Atlas [40], IXI [1], ISLES2022 [81], and the White Matter Hyperintensities Challenge [56]. We segment all subjects with Synthseg [13, 31]. Based on the segmentation, we affinely align the images to the MNI 152 template space [32, 33], and resample to $1\mathrm{mm}$ isometric resolution at a size of $192 \times 224 \times 192\mathrm{mm}$ . We perform manual quality control of the segmentation and registration by ensuring no segmented areas fall outside of the cropped volume and discard subjects failing this check (4 subjects). We extract a coronal slice of $192 \times 192\mathrm{mm}$ , bisecting the frontal Brain stem, Hippocampus, Thalamus, and Lateral ventricles. We rescale image intensities to the $[0, 1]$ interval using dataset-specific percentiles. For full head images, we create a brain mask with Synthstrip [31, 49]. The final dataset contains 2,282 subjects with 15,911 images and segmentation masks across 8 modalities. Subjects of the seven original datasets are split into $80\%$ for training and validation, $20\%$ test, with a minimum of 15 test subjects per dataset. + +# 4.2. Models + +Neuralizer-seen. This Neuralizer model includes all tasks available during training. We use this model to evaluate the performance on unseen scans from tasks and modalities that have been included in the training. The model uses the 4-stage architecture shown in Fig. 3 with 64 channels per layer. During training, the context size $|C_i|$ is sampled from $\mathcal{U}_{\{1,32\}}$ at each iteration. + +Neuralizer-unseen. To evaluate Neuralizer performance on tasks and modalities it has not been trained on, we train a family of Neuralizer models where a single task or modality is excluded from the training set. The model architecture of Neuralizer-unseen is identical to Neuralizer-seen. + +Baseline-seen. As no established baseline for multi-task and multi-modality models in neuroimaging can tackle the number of tasks we aim for, we compare Neuralizer to an ensemble of task-specific U-Nets [76, 85]. However, training one model for each task and modality requires overwhelming computational resources. To reduce the computational requirement, we follow previous modality-agnostic models [13, 49] and train each model on multiple input modalities. This lowers the number of models to be + +![](images/c0fd6b8c4ab958a4866fa6e48d308a57f8a7fbfd10c48d98718dc57123b19619.jpg) + +![](images/2fb8191a3e762802546feea73218571c4caa76075a1acd257d07e3c463b4abb4.jpg) + +![](images/06758ead4dccdb38c713dc105b69754416a18ba9690bc8090a141449b4d41348.jpg) + +![](images/b3174a3fc67059c080ee525a6a4e5a2fd8a44c5325a7bc8c8afa8eae1bf28f61.jpg) + +![](images/d9c8c77962f286661c55d3d0979620d756a1d7b37b834ea90c78fa357d8d0643.jpg) + +![](images/20705a6f1271ea2172df0597c863044496fa17068ac696b8daddf9a8b5d09a51.jpg) + +![](images/c732eb5110f46f99c6fb1d1f34776829913682328cbb8c87f2661694e0ef1c00.jpg) + +![](images/ae87f88a2dd257a524a2cb1e347b10913b005c6aea99e1da76a96975419356fc.jpg) + +![](images/ad495076e3e82ba2851a7bb69018eb31854c3e2d26c32a0eb69559d4bc4668cd.jpg) +Figure 4. Performance of multi-task Neuralizer and the task-specific baselines on each task, averaged across all modalities in the test set. The tasks being evaluated were included in the training of Neuralizer-seen (orange), held out in Neuralizer-unseen (blue), and specifically trained on by each task-specific baseline (gray). The x-axis is the size of the train/context set, and the y-axis is the Dice/PSNR score. Some points on the x-axis are omitted for better visibility. 'All' refers to all available train data for the task, ranging from 249 to 2,282 subjects depending on the task. The bars denote standard deviation across modalities. We extract results for T1 scans in Supplement D. + +![](images/de6954e83fcf785cfb55148edc6831cb10ae20987179d155976a63a0f0f80a1f.jpg) + +![](images/789c960037e4757f84c39eb1536dca37df9d20e00206bf3bebb3a821046ae5bb.jpg) + +trained to one per task, segmentation class, and modalitytransfer output modality. To compare Baseline-seen with Neuralizer-unseen given an equal amount of data, we train baselines with training set sizes of $\{1,2,4,8,16,32,\mathrm{all}\}$ and employ standard data augmentation. + +We use a 4-stage U-Net architecture with one residual block per layer. The channel width is tuned experimentally for each training dataset size. We select 256 channels when all data is available for training, and 64 channels otherwise. Using larger U-Nets resulted in overfitting and lower performance. Supplement H summarizes model parameter counts and inference costs. + +# 4.3. Training + +We use supervised training, task-specific loss functions, and weigh the MSE loss by selecting $\sigma^2 = 0.05$ , resulting in the loss terms being of similar magnitude. All models are trained with a batch size of 8, a learning rate of $10^{-4}$ , and the ADAM optimizer [52]. To speed up training, we undersample tasks that the model learns quickly, with sampling weights given in Supplement I. + +In addition to the task augmentations, we use data augmentations via random affine movements, random elastic deformations, and random flips along the sagittal plane. + +![](images/67d6f6e7811437e39ebb21b71eff8e192b68c9aee78c225c8b8526877df1ae3c.jpg) +Figure 5. Results averaged across tasks, expressed as relative performance compared to the baseline trained on all data. The tasks being evaluated were included in the training of Neuralizer-seen (orange), held out in Neuralizer-unseen (blue), and specifically trained on by each task-specific baseline (gray). The x-axis is the size of the train/context set, and the y-axis is the relative score. Some points on the x-axis are omitted for better visibility. 'All' refers to all available data for the task, ranging from 249 to 2,282 subjects depending on the task. Bars: standard deviation across tasks/modalities. + +For Baseline-seen, we reuse the augmentations but remove those that introduce uncertainty in the desired output. The training time for Neuralizer is 7 days on a single A100 GPU. The training time of the baseline models is capped at 5 days. All models use early stopping, ending the training after 25 epochs exhibiting no decrease in validation loss. The model with the lowest validation loss is used for further evaluation on the test set. + +# 4.4. Evaluation + +We evaluate the Dice coefficient for the segmentation and skull stripping tasks, and the Peak Signal-to-Noise Ratio (PSNR) for the image-to-image tasks on the test set. As the low-data regime is of particular interest, we measure performance as a function of context set size for the Neuralizer models and use training set size as an analog for the U-Net models. We evaluate context sets of up to 32 subjects. Larger context sets are possible but come at a linear cost in memory. + +# 4.5. Experiment 1: Baseline Comparison + +To assess if the proposed multi-task approach is competitive with task-specific models, we evaluate the performance and runtime of Neuralizer-seen, Neuralizer-unseen, and Baseline-seen on the test-set of each task. For Neuralizer-unseen, we withhold image modalities using a leave-one-out strategy during training and evaluate on the unseen modalities at test time. + +Results. We display the results by task, averaged across modalities in Fig. 4. We also provide an evaluation of using just the T1 modality in Supplement D, since many task-specific networks in neuroimage analysis literature focus on T1 images. We further aggregate performance across all tasks in Fig. 5, and provide tabular results in Supplement E. Both Neuralizer models outperform most task-specific baselines trained on up to 32 samples. When training the baselines on all available data, the baselines outperform Neuralizer-seen by 2 percentage points in relative performance, and Neuralizer-unseen by 3 percentage points. The loss in performance when generalizing to an unseen modality (between Neuralizer-seen and Neuralizer-unseen) is less than 2 percentage points for all context set sizes. + +Training the baseline model to convergence on 32 samples took on average $28.2 \pm 16.6$ hours per task, using one A100 GPU. Since Neuralizer only requires inference for a new task, it is orders of magnitude faster, requiring less than 0.1 seconds on a GPU and less than 3 seconds on a CPU. + +We provide qualitative samples of the predictions from Neuralizer-seen model in Supplement A, Figures 6-8. + +# 4.6. Experiment 2: Context set size analysis + +We assess the few-shot setting that is prevalent in neuroimage analysis, where few annotated images are often available for a new task. We evaluate performance as a function of the number of labeled samples. For Neuralizer, we evaluate the model with context-set sizes of $\{1,2,4,8,16,32\}$ unique subjects from the test set. For the baseline, we trained models with reduced training set sizes of the same amount of subjects. To reduce the effect of random training subject selection, we train three separate baselines with $n = 1$ , two baselines with $n = 2$ , and average results of models with the same $n$ . + +Results. Tab. 4 and Figs. 4, 5 illustrate the results. For all models, prediction accuracy increases with the availability of labeled data, with diminishing returns. For both Neuralizer models, a context set size of one achieves more than $90\%$ of the performance attainable with all data. For most tasks, the baseline performs overall worse than both Neuralizer models when $\leq 32$ labeled samples are available but achieves the best overall performance on larger datasets. + +# 4.7. Experiment 3: Generalization to a new segmentation protocol + +The Hammers Atlas dataset [27,37,40] provides an alternative anatomical segmentation protocol to the widely-used Freesurfer segmentation available for most subjects in the dataset. The shape, size, and amount of annotated regions in the protocols differ drastically. A different image acquisition site also leads to differences in visual characteristics. We use the Hammers Atlas dataset to evaluate Neuralizer-unseen by entirely withholding the dataset and its annotations from training. We evaluate the Dice coefficient of the 14 major anatomical segmentation classes present in the center and right half of the coronal slice. + +Results. Tab. 3 illustrates the results. Neuralizer-unseen performs similarly to Neuralizer-seen and the baseline, while not requiring lengthy re-training or fine-tuning on the Hammers Atlas dataset, and not having seen the segmentation protocol. All three models achieve a mean Dice coefficient of 0.84. The largest performance difference is in the third ventricle class, where both Neuralizer models outperform the baseline by at least 0.04 Dice. The Freesurfer segmentation protocol included in the training set of the Neuralizer models also contains a third ventricle class. + +# 5. Discussion + +Our experiments using modality and segmentation class hold-outs show that Neuralizer can generalize well to unseen neuroimaging tasks. Across all context set sizes, the generalization loss between seen and unseen modalities and segmentation classes is less than 2 percentage points across + +
ModelTask SeenSegmentation Class (Hammers Atlas)Mean (std)
HipPAGSTGMIGFuGStmInsPCGThaCC3VPrGPoGALG
Baseline-seen✓.88.86.93.92.79.87.82.87.90.80.68.83.77.82.84 (.07)
Neuralizer-seen✓.88.86.92.92.76.88.83.85.90.82.73.86.77.81.84 (.07)
Neuralizer-unseenX.88.87.93.91.78.87.82.85.90.81.72.85.78.81.84 (.06)
+ +Table 3. Segmentation of the Hammers Atlas dataset. For Neuralizer-unseen, this dataset and segmentation protocol is withheld from training. Evaluation of major labels located in the center and right of the coronal slice. See Supplement F for class abbreviations. + +Experiments 1 and 2. On the smaller held-out HammersAtlas segmentation dataset, we find that Neuralizer can generalize to unseen tasks with similar performance. These results show promise that a single Neuralizer model can perform multiple neuroimaging tasks, including generalization to new inference tasks not seen during training. + +In settings with 32 or fewer labeled example images, Neuralizer-unseen outperforms task-specific baselines despite never having seen the task or modality at train time, and taking nearly no effort or compute compared to the baselines which require substantial expertise, manual labor, and compute resources. The performance difference is largest when only one labeled subject is available, but still present at 32 subjects (Fig. 5). Neuralizer provides a performance advantage on smaller datasets likely by exploiting neuroimaging similarities across the many other neuroimaging tasks and datasets available in training. + +When training the baselines on all available data, they can outperform Neuralizer-seen and Neuralizer-unseen by at most 3 percentage points. The inflection point of identical performance between Neuralizer and the baselines is not covered by the range of context set sizes chosen for training and evaluation due to prohibitive computational costs and is an interesting direction of study. When large annotated datasets are available, the baselines performed best on most tasks. However, training task-specific models comes at a significant cost. As a first step in the proposed problem formulation, Neuralizer offers an alternative with near equal performance, while only requiring seconds to infer any task from the context set. + +# Limitations + +We made simplifying assumptions in this first paper demonstrating the potential of multi-task generalization in neuroimaging. The experiments are conducted on 2D data slices. In large part, we did this since running the hundreds of baselines in 3D would be infeasible on our compute cluster. Entire volumetric data also impose a challenging memory requirement on Neuralizer models. To tackle 3D data in the future, we plan to process multiple slices at a time. + +We affinely aligned the neuroimages of the context set to the target image. Early in Neuralizer development, we tried + +training on non-aligned inputs but found that it deteriorated performance. The need for affine alignment provides an obstacle to adoption. While existing affine-alignment tools are fast and can be employed, we also believe that this requirement can be removed with further development. + +We originally experimented with tumor and lesion segmentation tasks but found this to be a more challenging scenario. Lesions are spatially heterogeneous, making learning from the context set much harder. We excluded tumor and lesion segmentation masks from the experiments, but plan to study this setting in the future. + +While we demonstrate the proposed ideas on a broad range of tasks and modalities, neuroimage analysis can involve more domains, tasks, and populations, like image registration, surface-based tasks, CT image domains, and pediatric data. We plan to extend Neuralizer to tackle these in the future. + +# 6. Conclusion + +Neuralizer performs accurate rapid single-pass, multitask generalization, and even outperforms task-specific baselines in limited data scenarios. Even when a large amount of annotated data is available, Neuralizer often matches baseline performance despite not training on the data. Neuralizer provides clinical researchers and scientists with a single model to solve a wide range of neuroimaging tasks on images of many modalities, and can be easily adapted to new tasks without the prohibitive requirement of retraining or fine-tuning a task-specific model. We believe this will facilitate neuroscience analyses not currently possible. + +# Acknowledgments + +This work was performed during Steffen Czolbes's visit to the Athinoula A. Martins Center for Biomedical Imaging, MGH. The collaboration was supported by the EliteForsk Travel Grant of the Danish Ministry of Higher Education and Science. The work was further funded by the Novo Nordisk Foundation (grants no. NNF20OC0062606 and NNF17OC0028360), the Lundbeck Foundation (grant no. R218-2016-883), and NIH grants 1R01AG064027 and 1R01AG070988. + +# References + +[1] IXI Dataset. 4, 5 +[2] David Abramian and Anders Eklund. Refacing: Reconstructing anonymized facial features using GANS. International Symposium on Biomedical Imaging, 2019-April:1104-1108,apr 2019.2 +[3] Yasser Ad-Dab'bagh, O Lyttelton, JS Muehlboeck, C Lepage, D Einarson, K Mok, O Ivanov, RD Vincent, J Lerch, and E Fombonne. The CIVET image-processing environment: a fully automated comprehensive pipeline for anatomical neuroimaging research. In Proceedings of the 12th annual meeting of the organization for human brain mapping, 2006. 2 +[4] Laith Alzubaidi, Muthana Al-Amidie, Ahmed Al-Asadi, Amjad J. Humaidi, Omran Al-Shamma, Mohammed A. Fadhel, Jinglan Zhang, J. Santamaría, and Ye Duan. Novel Transfer Learning Approach for Medical Imaging with Limited Labeled Data. Cancers 2021, Vol. 13, Page 1590, 13(7):1590, mar 2021. 3 +[5] Vincent Arsigny, Olivier Commowick, Xavier Pennec, and Nicholas Ayache. A log-euclidean framework for statistics on diffeomorphisms. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 924-931. Springer Verlag, 2006. 2 +[6] John Ashburner. A fast diffeomorphic image registration algorithm. NeuroImage, 38(1):95-113, oct 2007. 2 +[7] B. B. Avants, C. L. Epstein, M. Grossman, and J. C. Gee. Symmetric diffeomorphic image registration with cross-correlation: Evaluating automated labeling of elderly and neurodegenerative brain. Medical Image Analysis, 12(1):26-41, feb 2008. 2 +[8] Spyridon Bakas, Hamed Akbari, Aristeidis Sotiras, Michel Bilello, Martin Rozycki, Justin S. Kirby, John B. Freymann, Keyvan Farahani, and Christos Davatzikos. Advancing The Cancer Genome Atlas glioma MRI collections with expert segmentation labels and radiomic features. Scientific Data, 4, sep 2017. 4, 5 +[9] Spyridon Bakas, Bjoern Menze, and Others. Identifying the Best Machine Learning Algorithms for Brain Tumor Segmentation, Progression Assessment, and Overall Survival Prediction in the BRATS Challenge. arXiv, 124, nov 2018. 2, 4, 5 +[10] Guha Balakrishnan, Amy Zhao, Mert R. Sabuncu, John Guttag, and Adrian V. Dalca. VoxelMorph: A Learning Framework for Deformable Medical Image Registration. IEEE Transactions on Medical Imaging, 38(8):1788-1800, feb 2019. 2 +[11] Cheng Bian, Chenglang Yuan, Kai Ma, Shuang Yu, Dong Wei, and Yefeng Zheng. Domain Adaptation Meets Zero-Shot Learning: An Annotation-Efficient Approach to Multi-Modality Medical Image Segmentation. IEEE Transactions on Medical Imaging, 41(5):1043–1056, may 2022. 3 +[12] Benjamin Billot, Douglas Greve, Koen Van Leemput, Bruce Fischl, Juan Eugenio Iglesias, and Adrian V Dalca. A learning strategy for contrast-agnostic mri segmentation. arXiv preprint arXiv:2003.01995, 2020. 2 + +[13] Benjamin Billot, Douglas N Greve, Oula Puonti, Axel Thielscher, Koen Van Leemput, Bruce Fischl, Adrian V Dalca, and Juan Eugenio Iglesias. SynthSeg: Domain Randomisation for Segmentation of Brain Scans of any Contrast and Resolution. jul 2021. 1, 2, 3, 4, 5 +[14] Maxime Bucher, Tuan-Hung VU, Matthieu Cord, and Patrick Pérez. Zero-Shot Semantic Segmentation. Advances in Neural Information Processing Systems, 32, 2019. 3 +[15] Victor Ion Butoi, Jose Javier Gonzalez Ortiz, Tianyu Ma, Mert Sabuncu, John Guttag, and Adrian V. Dalca. Universeg: Universal medical image segmentation. arXiv preprint, 2023. 3, 4, 5 +[16] Víctor M. Campello, Carlos Martín-Isla, Cristian Izquierdo, Steffen E. Petersen, Miguel A.González Ballester, and Karim Lekadir. Combining Multi-Sequence and Synthetic Images for Improved Segmentation of Late Gadolinium Enhancement Cardiac MRI. Lecture Notes in Computer Science, 12009 LNCS:290-299, 2020. 3 +[17] Rich Caruana, Lorien Pratt, and Sebastian Thrun. Multitask Learning. Machine Learning 1997 28:1, 28(1):41-75, 1997. 2 +[18] Hao Chen, Qi Dou, Lequan Yu, Jing Qin, and Pheng Ann Heng. VoxResNet: Deep voxelwise residual networks for brain segmentation from 3D MR images. NeuroImage, 170:446-455, apr 2018. 2 +[19] Sihong Chen, Kai Ma, and Yefeng Zheng. Med3D: Transfer Learning for 3D Medical Image Analysis. apr 2019. 3 +[20] Steffen Czolbe, Oswin Krause, and Aasa Feragen. Semantic similarity metrics for learned image registration. Proceedings of Machine Learning Research, 2021. 2 +[21] Adrian V Dalca, Katherine L Bouman, William T Freeman, Natalia S Rost, Mert R Sabuncu, and Polina Golland. Medical image imputation from image collections. IEEE transactions on medical imaging, 38(2):504-514, 2018. 2 +[22] Adrian V Dalca, John Guttag, and Mert R Sabuncu. Anatomical priors in convolutional networks for unsupervised biomedical segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 9290-9299, 2018. 2 +[23] Adrian V Dalca, Evan Yu, Polina Golland, Bruce Fischl, Mert R Sabuncu, and Juan Eugenio Iglesias. Unsupervised deep learning for bayesian brain mri segmentation. In Medical Image Computing and Computer Assisted Intervention—MICCAI 2019: 22nd International Conference, Shenzhen, China, October 13–17, 2019, Proceedings, Part III 22, pages 356–365. Springer, 2019. 2 +[24] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In Conference on Computer Vision and Pattern Recognition, pages 248-255, 2009. 3 +[25] Francisco Javier Díaz-Pernas, Mario Martínez-Zarzuela, David González-Ortega, and Míriam Antón-Rodríguez. A Deep Learning Approach for Brain Tumor Classification and Segmentation Using a Multiscale Convolutional Neural Network. Healthcare, 9(2):153, feb 2021. 3 +[26] Theodoras Evgeniou and Massimiliano Pontil. Regularized multi-task learning. Proceedings of the Tenth ACM + +SIGKDD International Conference on Knowledge Discovery and Data Mining, pages 109-117, 2004. 2 +[27] Isabelle Faillenot, Rolf A. Heckemann, Maud Frot, and Alexander Hammers. Macroanatomy and 3D probabilistic atlas of the human insula. NeuroImage, 150:88-98, apr 2017. 4, 7, 24 +[28] Mirza Faisal Beg, Michael I Miller, Alain Trouvétrouv, and Laurent Younes. Computing Large Deformation Metric Mappings via Geodesic Flows of Diffeomorphisms. International Journal of Computer Vision, 61(2):139-157, 2005. 2 +[29] Ruiwei Feng, Xiangshang Zheng, Tianxiang Gao, Jintai Chen, Wenzhe Wang, Danny Z. Chen, and Jian Wu. Interactive Few-Shot Learning: Limited Supervision, Better Medical Image Segmentation. IEEE Transactions on Medical Imaging, 40(10):2575-2588, oct 2021. 3 +[30] Abdur R Feyjie, Reza Azad, Marco Pedersoli, Claude Kauffman, Ismail Ben Ayed, and Jose Dolz. Semi-supervised few-shot learning for medical image segmentation. mar 2020. 3 +[31] B Fischl. FreeSurfer. NeuroImage, 62 (2), 774-781, 2012. 2, 4, 5 +[32] VS Fonov, AC Evans, RC McKinstry, CR Almli, and DL Collins. Unbiased nonlinear average age-appropriate brain templates from birth to adulthood. NeuroImage, 47:S102, jul 2009. 5 +[33] Vladimir Fonov, Alan C. Evans, Kelly Botteron, C. Robert Almli, Robert C. McKinstry, and D. Louis Collins. Unbiased average age-appropriate atlases for pediatric studies. NeuroImage, 54(1):313-327, jan 2011. 5 +[34] Richard SJ Frackowiak. Human Brain Function. 2004. 2 +[35] Tal Goldfryd, Shiri Gordon, and Tammy Riklin Raviv. Deep semi-supervised bias field correction of mr images. Proceedings - International Symposium on Biomedical Imaging, 2021-April:1836-1840, apr 2021. 2 +[36] Enhao Gong, John M. Pauly, Max Wintermark, and Greg Zaharchuk. Deep learning enables reduced gadolinium dose for contrast-enhanced brain MRI. Journal of Magnetic Resonance Imaging, 48(2):330-340, aug 2018. 2 +[37] Ioannis S. Gousias, Daniel Rueckert, Rolf A. Heckemann, Leigh E. Dyet, James P. Boardman, A. David Edwards, and Alexander Hammers. Automatic segmentation of brain MRIs of 2-year-olds into 83 regions of interest. NeuroImage, 40(2):672-684, apr 2008. 4, 7, 24 +[38] Sachin Gupta, Narinder Singh Punn, Sanjay Kumar Sonbhadra, and Sonali Agarwal. MAG-Net: Multi-task Attention Guided Network for Brain Tumor Segmentation and Classification. Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), 13147 LNCS:3-15, 2021. 3 +[39] J Hale, Nakeisha Schimke, and John Hale. Quickshear defacing for neuroimages Attack Graph Generation on HPC Clusters View project Quickshear Defacing for Neuroimages. Frontiers in psychiatry, 21, 2011. 2 +[40] Alexander Hammers, Richard Allom, Matthias J. Koepp, Samantha L. Free, Ralph Myers, Louis Lemieux, Tejal N. + +Mitchell, David J. Brooks, and John S. Duncan. Three-dimensional maximum probability atlas of the human brain, with particular reference to the temporal lobe. Human brain mapping, 19(4):224-247, aug 2003. 4, 5, 7, 24 +[41] Xu Han, Roland Kwitt, Stephen Aylward, Spyridon Bakas, Bjoern Menze, Alexander Asturias, Paul Vespa, John Van Horn, and Marc Niethammer. Brain extraction from normal and pathological images: A joint PCA/Image-Reconstruction approach. NeuroImage, 176:431-445, aug 2018. 2 +[42] Stine Hansen, Srishti Gautam, Robert Jenssen, and Michael Kampffmeyer. Anomaly Detection-Inspired Few-Shot Medical Image Segmentation Through Self-Supervision With Supervoxels. Medical Image Analysis, 78, mar 2022. 3 +[43] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 3, 4 +[44] Dan Hendrycks and Kevin Gimpel. Gaussian Error Linear Units (GELUs). jun 2016. 4 +[45] Malte Hoffmann, Benjamin Billot, Douglas N. Greve, Juan Eugenio Iglesias, Bruce Fischl, and Adrian V. Dalca. SynthMorph: learning contrast-invariant registration without acquired images. IEEE Transactions on Medical Imaging, apr 2020. 1, 2, 3, 19 +[46] Noah Hollmann, Samuel Müller, Katharina Eggensperger, and Frank Hutter. TabPFN: A Transformer That Solves Small Tabular Classification Problems in a Second. In NeurIPS 2022 First Table Representation Workshop, jul 2022. 3 +[47] Andrew Hoopes, Malte Hoffmann, Douglas N Greve, Bruce Fischl, John Guttag, and Adrian V Dalca. Learning the effect of registration hyperparameters with hypermorph. arXiv preprint arXiv:2203.16680, 2022. 2, 4, 5 +[48] Andrew Hoopes, Juan Eugenio Iglesias, Bruce Fischl, Douglas Greve, and Adrian V Dalca. TopoFit: Rapid Reconstruction of Topologically-Correct Cortical Surfaces. Proceedings of Machine Learning Research-Under Review, pages 1-13, 2022. 2 +[49] Andrew Hoopes, Jocelyn S. Mora, Adrian V. Dalca, Bruce Fischl, and Malte Hoffmann. SynthStrip: Skull-Stripping for Any Brain Image. mar 2022. 2, 3, 4, 5 +[50] Minyoung Huh, Pulkit Agrawal, and Alexei A. Efros. What makes ImageNet good for transfer learning? aug 2016. 3 +[51] Mark Jenkinson, Christian F. Beckmann, Timothy E.J. Behrens, Mark W. Woolrich, and Stephen M. Smith. FSL. NeuroImage, 62(2):782-790, aug 2012. 2 +[52] Diederik P. Kingma and Jimmy Lei Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations, dec 2015. 6 +[53] Jens Kleesiek, Gregor Urban, Alexander Hubert, Daniel Schwarz, Klaus Maier-Hein, Martin Bendszus, and Armin Biller. Deep MRI brain extraction: A 3D convolutional neural network for skull stripping. NeuroImage, 129:460–469, apr 2016. 2 +[54] Arno Klein, Jesper Andersson, Babak A. Ardekani, John Ashburner, Brian Avants, Ming Chang Chiang, Gary E. + +Christensen, D. Louis Collins, James Gee, Pierre Hellier, Joo Hyun Song, Mark Jenkinson, Claude Lepage, Daniel Rueckert, Paul Thompson, Tom Vercauteren, Roger P. Woods, J. John Mann, and Ramin V. Parsey. Evaluation of 14 nonlinear deformation algorithms applied to human brain MRI registration. NeuroImage, 46(3):786-802, jul 2009. 2 +[55] Simon Kornblith, Jonathon Shlens, and Quoc V Le Google Brain. Do Better ImageNet Models Transfer Better? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2661-2671, 2019. 3 +[56] Hugo J. Kuijf, Adrià Casamitjana, D. Louis Collins, Mahsa Dadar, Achilleas Georgiou, Mohsen Ghafoorian, Dakai Jin, April Khademi, Jesse Knight, Hongwei Li, Xavier Llado, J. Matthijs Biesbroek, Miguel Luna, Qaiser Mahmood, Richard Mckinley, Alireza Mehrtash, Sebastien Ourselin, Bo Yong Park, Hyunjin Park, Sang Hyun Park, Simon Pezold, Elodie Puybureau, Jeroen De Bresser, Leticia Rittner, Carole H. Sudre, Sergi Valverde, Veronica Vilaplana, Roland Wiest, Yongchao Xu, Ziyue Xu, Guodong Zeng, Jianguo Zhang, Guoyan Zheng, Rutger Heinen, Christopher Chen, Wiesje Van Der Flier, Frederik Barkhof, Max A. Viergever, Geert Jan Biessels, Simon Andermatt, Mariana Bento, Matt Berseth, Mikhail Belyaev, and M. Jorge Cardoso. Standardized Assessment of Automatic Segmentation of White Matter Hyperintensities and Results of the WMH Segmentation Challenge. IEEE Transactions on Medical Imaging, 38(11):2556-2568, nov 2019. 2, 4, 5 +[57] Manoj Kumar, Michael J. Anderson, James W. Antony, Christopher Baldassano, Paula P. Brooks, Ming Bo Cai, Po-Hsuan Cameron Chen, Cameron T. Ellis, Gregory Henselman-Petrusek, David Huberdeau, J. Benjamin Hutchinson, Y. Peeta Li, Qihong Lu, Jeremy R. Manning, Anne C. Mennen, Samuel A. Nastase, Hugo Richard, Anna C. Schapiro, Nicolas W. Schuck, Michael Shvartsman, Narayanan Sundaram, Daniel Suo, Javier S. Turek, David Turner, Vy A. Vo, Grant Wallace, Yida Wang, Jamal A. Williams, Hejia Zhang, Xia Zhu, Mihai Capota $^{\text{一}}$ , Jonathan D. Cohen, Uri Hasson, Kai Li, Peter J. Ramadge, Nicholas B. Turk-Browne, Theodore L. Willke, and Kenneth A. Norman. BrainIAK: The Brain Imaging Analysis Kit. Aperture Neuro, 2021(4), jan 2021. 2 +[58] Sonia Laguna, Riana Schleicher, Benjamin Billot, Pamela Schaefer, Brenna Mckaig, Joshua N Goldstein, Kevin N Sheth, Matthew S Rosen, W Taylor Kimberly, and Juan Eugenio Iglesias. Super-resolution of portable low-field MRI in real scenarios: integration with denoising and domain adaptation. Medical Imaging with Deep Learning (MIDL), 2022. 2 +[59] Erik G Learned-Miller and Parvez Ahammad. Joint MRI Bias Removal Using Entropy Minimization Across Images. Advances in Neural Information Processing Systems, 17, 2004. 2 +[60] Jaein Lee, Eunsong Kang, Eunjin Jeon, and Heung Il Suk. Meta-modulation Network for Domain Generalization in Multi-site fMRI Classification. Lecture Notes in Computer Science, 12905 LNCS:500-509, 2021. 2 + +[61] Gen Li, Varun Jampani, Laura Sevilla-Lara, Deqing Sun, Jonghyun Kim, and Joongkyu Kim. Adaptive Prototype Learning and Allocation for Few-Shot Segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8334-8343, 2021. 3 +[62] Xiaoxiao Li, Yufeng Gu, Nicha Dvornek, Lawrence H. Staib, Pamela Ventola, and James S. Duncan. Multi-site fMRI analysis using privacy-preserving federated learning and domain adaptation: ABIDE results. Medical Image Analysis, 65:101765, oct 2020. 1 +[63] Xiang Li, Tianhan Wei, Yau Pun Chen, Yu-Wing Tai, and Chi-Keung Tang. FSS-1000: A 1000-Class Dataset for Few-Shot Segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2869–2878, 2020. 3 +[64] Sook-Lei Liew, Bethany P. Lo, Miranda R. Donnelly, Artemis Zavaliangos-Petropulu, Jessica N. Jeong, Giuseppe Barisano, Alexandre Hutton, Julia P. Simon, Julia M. Juliano, Anisha Suri, Zhizhuo Wang, Aisha Abdullah, Jun Kim, Tyler Ard, Nerisa Banaj, Michael R. Borich, Lara A. Boyd, Amy Brodtmann, Cathrin M. Buetefisch, Lei Cao, Jessica M. Cassidy, Valentina Ciullo, Adriana B. Conforto, Steven C. Cramer, Rosalia Dacosta-Aguayo, Ezequiel de la Rosa, Martin Domin, Adrienne N. Dula, Wuwei Feng, Alexandre R. Franco, Fatemeh Geranmayeh, Alexandre Gramfort, Chris M. Gregory, Colleen A. Hanlon, Brenton G. Hordacre, Steven A. Kautz, Mohamed Salah Khlif, Hosung Kim, Jan S. Kirschke, Jingchun Liu, Martin Lotze, Bradley J. MacIntosh, Maria Mataró, Feroze B. Mohamed, Jan E. Nordvik, Gilsoon Park, Amy Pienta, Fabrizio Piras, Shane M. Redman, Kate P. Revill, Mauricio Reyes, Andrew D. Robertson, Na Jin Seo, Surjo R. Soekadar, Gianfranco Spalletta, Alison Sweet, Maria Telenczuk, Gregory Thielman, Lars T. Westlye, Carolee J. Winstein, George F. Wittenberg, Kristin A. Wong, and Chunshui Yu. A large, curated, open-source stroke neuroimaging dataset to improve lesion segmentation algorithms. Scientific Data, 9(1):1-12, jun 2022. 4, 5 +[65] Jialin Liu, Fei Chao, and Chih-Min Lin. Task Augmentation by Rotating for Meta-Learning. feb 2020. 3 +[66] Ming-Yu Liu, Xun Huang, Arun Mallya, Tero Karras, Timo Aila, Jaakko Lehtinen, and Jan Kautz. Few-Shot Unsupervised Image-to-Image Translation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10551–10560, 2019. 3 +[67] Weide Liu, Chi Zhang, Guosheng Lin, and Fayao Liu. CR-Net: Cross-Reference Networks for Few-Shot Segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4165–4173, 2020. 3 +[68] Xiaoxiao Liu, Marc Niethammer, Roland Kwitt, Nikhil Singh, Matt McCormick, and Stephen Aylward. Low-Rank Atlas Image Analyses in the Presence of Pathologies. IEEE Transactions on Medical Imaging, 34(12):2583-2591, dec 2015. 2 +[69] Xiaofeng Liu, Fangxu Xing, Chao Yang, C. C.Jay Kuo, Georges El Fakhri, and Jonghye Woo. Symmetric- + +Constrained Irregular Structure Inpainting for Brain MRI Registration with Tumor Pathology. Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), 12658 LNCS:80-91, 2021. 2 +[70] Michael Lustig, David L. Donoho, Juan M. Santos, and John M. Pauly. Compressed sensing MRI: A look at how CS can improve on current imaging techniques. IEEE Signal Processing Magazine, 25(2):72-82, 2008. 2 +[71] Kushagra Mahajan, Monika Sharma, and Lovekesh Vig. Meta-DermDiagnosis: Few-Shot Skin Disease Identification Using Meta-Learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 730-731, 2020. 3 +[72] José V. Manjón, Pierrick Coup, Antonio Buades, Vladimir Fonov, D. Louis Collins, and Montserrat Robles. Non-local MRI upsampling. Medical Image Analysis, 14(6):784-792, dec 2010. 2 +[73] José V. Manjón, Pierrick Coupe, Antonio Buades, D. Louis Collins, and Montserrat Robles. MRI superresolution using self-similarity and image priors. International Journal of Biomedical Imaging, 2010, 2010. 2 +[74] Daniel S Marcus, Tracy H Wang, Jamie Parker, John G Csernansky, John C Morris, and Randy L Buckner. Open Access Series of Imaging Studies (OASIS): Cross-Sectional MRI Data in Young, Middle Aged, Nondemented, and Demented Older Adults. Journal of Cognitive Neuroscience, 19:1498-1507, 2007. 4, 5 +[75] Bjoern H. Menze, Andras Jakab, Stefan Bauer, Jayashree Kalpathy-Cramer, Keyvan Farahani, Justin Kirby, Yuliya Burren, Nicole Porz, Johannes Slotboom, Roland Wiest, Lerente Lanczi, Elizabeth Gerstner, Marc Andre Weber, Tal Arbel, Brian B. Avants, Nicholas Ayache, Patricia Buendia, D. Louis Collins, Nicolas Cordier, Jason J. Corso, Antonio Criminisi, Tilak Das, Hervé Delingette, Cagatay Demiralp, Christopher R. Durst, Michel Dojat, Senan Doyle, Joana Festa, Florence Forbes, Ezequiel Geremia, Ben Glocker, Polina Golland, Xiaotao Guo, Andac Hamamci, Khan M. Iftekharuddin, Raj Jena, Nigel M. John, Ender Konukoglu, Danial Lashkari, José Antônio Mariz, Raphael Meier, Sérgio Pereira, Doina Precup, Stephen J. Price, Tammy Riklin Raviv, Syed M.S. Reza, Michael Ryan, Duygu Sarikaya, Lawrence Schwartz, Hoo Chang Shin, Jamie Shotton, Carlos A. Silva, Nuno Sousa, Nagesh K. Subbanna, Gabor Szekely, Thomas J. Taylor, Owen M. Thomas, Nicholas J. Tustison, Gozde Unal, Flor Vasseur, Max Wintermark, Dong Hye Ye, Liang Zhao, Binsheng Zhao, Darko Zikic, Marcel Prastawa, Mauricio Reyes, and Koen Van Leemput. The Multimodal Brain Tumor Image Segmentation Benchmark (BRATS). IEEE Transactions on Medical Imaging, 34(10):1993-2024, oct 2015. 4, 5 +[76] Fausto Milletari, Nassir Navab, and Seyed Ahmad Ahmadi. V-Net: Fully convolutional neural networks for volumetric medical image segmentation. Proceedings - 2016 4th International Conference on 3D Vision, 3DV 2016, pages 565-571, dec 2016. 3, 4, 5 +[77] Samuel Müller, Noah Hollmann, Sebastian Pineda Arango, Josif Grabocka, and Frank Hutter. Transformers Can Do + +Bayesian Inference. In International Conference on Learning Representations, mar 2022. 3 +[78] Bao Nguyen, Adam Feldman, Sarath Bethapudi, Andrew Jennings, and Chris G. Willcocks. Unsupervised region-based anomaly detection in brain mri with adversarial image inpainting. Proceedings - International Symposium on Biomedical Imaging, 2021-April:1127-1131, apr 2021. 2 +[79] Dong Nie, Roger Trullo, Jun Lian, Caroline Petitjean, Su Ruan, Qian Wang, and Dinggang Shen. Medical image synthesis with context-aware generative adversarial networks. Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), 10435 LNCS:417-425, 2017. 2 +[80] Alexander F.I. Osman and Nissren M. Tamam. Deep learning-based convolutional neural network for intramodality brain MRI synthesis. Journal of Applied Clinical Medical Physics, 23(4):e13530, apr 2022. 2 +[81] Moritz Roman Hernandez Petzsche, Ezequiel de la Rosa, Uta Hanning, Roland Wiest, Waldo Enrique Valenzuela Pinilla, Mauricio Reyes, Maria Ines Meyer, Sook-Lei Liew, Florian Kofler, Ivan Ezhov, David Robben, Alexander Hutton, Tassilo Friedrich, Teresa Zarth, Johannes Burkle, The Anh Baran, Bjoern Menze, Gabriel Broocks, Lukas Meyer, Claus Zimmer, Tobias Boeckh-Behrens, Maria Berndt, Benno Ikenberg, Benedikt Wiestler, and Jan S. Kirschke. ISLES 2022: A multi-center magnetic resonance imaging stroke lesion segmentation dataset. jun 2022. 4, 5 +[82] Kilian M Pohl, Sylvain Bouix, Motoaki Nakamura, Torsten Rohlfing, Robert W McCarley, Ron Kikinis, W Eric L Grimson, Martha E Shenton, and William M Wells. A hierarchical algorithm for mr brain image parcellation. IEEE transactions on medical imaging, 26(9):1201-1212, 2007. 2 +[83] Maithra Raghu, Chiyuan Zhang, Google Brain, Jon Kleinberg, and Samy Bengio. Transfusion: Understanding Transfer Learning for Medical Imaging. Advances in Neural Information Processing Systems, 32, 2019. 3 +[84] Sachin Ravi and Hugo Larochelle. Optimization as a Model for Few-Shot Learning. In International Conference on Learning Representations (ICLR), jul 2017. 3 +[85] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, volume 9351, pages 234-241. Springer Verlag, 2015. 1, 3, 4, 5 +[86] Edgar Schonfeld, Sayna Ebrahimi, Samarth Sinha, Trevor Darrell, and Zeynep Akata. Generalized Zero- and Few-Shot Learning via Aligned Variational Autoencoders. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8247–8255, 2019. 3 +[87] F. Segonne, A. M. Dale, E. Busa, M. Glessner, D. Salat, H. K. Hahn, and B. Fischl. A hybrid approach to the skull stripping problem in MRI. NeuroImage, 22(3):1060-1075, jul 2004. 2 +[88] Ozan Sener and Vladlen Koltun. Multi-Task Learning as Multi-Objective Optimization. Advances in Neural Information Processing Systems, 31, 2018. 2 + +[89] Jun Seo, Young-Hyun Park, Sung Whan Yoon, and Jaekyun Moon. Task-Adaptive Feature Transformer with Semantic Enrichment for Few-Shot Segmentation. feb 2022. 3 +[90] David W. Shattuck and Richard M. Leahy. BrainSuite: An automated cortical surface identification tool. Medical Image Analysis, 6(2):129-142, jun 2002. 2 +[91] X. Shen, F. Tokoglu, X. Papademetris, and R. T. Constable. Groupwise whole-brain parcellation from resting-state fMRI data for network node identification. NeuroImage, 82:403-415, nov 2013. 2 +[92] Nalini M Singh, Juan Eugenio Iglesias, Elfar Adalsteinson, Adrian V Dalca, and Polina Golland. Joint Frequency and Image Space Learning for MRI Reconstruction and Analysis. Journal of Machine Learning for Biomedical Imaging, 2022:1-28, 2022. 2, 25 +[93] Youssef Skandarani, Nathan Painchaud, Pierre-Marc Jodoin, and Alain Lalande. On the effectiveness of GAN generated cardiac MRIs for segmentation. may 2020. 3 +[94] Stephen M. Smith. Fast robust automated brain extraction. Human Brain Mapping, 17(3):143-155, nov 2002. 2 +[95] Stephen M. Smith, Mark Jenkinson, Mark W. Woolrich, Christian F. Beckmann, Timothy E.J. Behrens, Heidi Johansen-Berg, Peter R. Bannister, Marilena De Luca, Ivana Drobnjak, David E. Flitney, Rami K. Niazy, James Saunders, John Vickers, Yongyue Zhang, Nicola De Stefano, J. Michael Brady, and Paul M. Matthews. Advances in functional and structural MR image analysis and implementation as FSL. NeuroImage, 23(1), 2004. 2 +[96] Jake Snell, Kevin Swersky, and Twitter Richard Zemel. Prototypical Networks for Few-shot Learning. Advances in Neural Information Processing Systems, 30, 2017. 3 +[97] Haoliang Sun, Ronak Mehta, Hao H. Zhou, Zhichun Huang, Sterling C. Johnson, Vivek Prabhakaran, and Vikas Singh. DUAL-GLOW: Conditional Flow-Based Generative Model for Modality Transfer. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10611-10620, 2019. 2 +[98] Nathalie Tzourio-Mazoyer, Brigitte Landeau, Dimitri Papathanassiou, Fabrice Crivello, Octave Etard, Nicolas Delcroix, Bernard Mazoyer, and Marc Joliot. Automated anatomical labeling of activations in spm using a macroscopic anatomical parcellation of the mni mri single-subject brain. Neuroimage, 15(1):273-289, 2002. 2 +[99] N. Tzourio-Mazoyer, B. Landeau, D. Papathanassiou, F. Crivello, O. Etard, N. Delcroix, B. Mazoyer, and M. Joliot. Automated Anatomical Labeling of Activations in SPM Using a Macroscopic Anatomical Parcellation of the MNI MRI Single-Subject Brain. NeuroImage, 15(1):273–289, jan 2002. 2 +[100] David C. Van Essen, Stephen M. Smith, Deanna M. Barch, Timothy E.J. Behrens, Essa Yacoub, and Kamil Ugurbil. The WU-Minn Human Connectome Project: An overview. NeuroImage, 80:62-79, oct 2013. 2 +[101] Oriol Vinyals, Google Deepmind, Charles Blundell, Timothy Lillicrap, Koray Kavukcuoglu, and Daan Wierstra. Matching Networks for One Shot Learning. Advances in Neural Information Processing Systems, 29, 2016. 3 + +[102] Oriol Vinyals, Google Deepmind, Charles Blundell, Timothy Lillicrap, Koray Kavukcuoglu, and Daan Wierstra. Matching Networks for One Shot Learning. Advances in Neural Information Processing Systems, 29, 2016. 3 +[103] Christian Wachinger, Martin Reuter, and Tassilo Klein. Deepnat: Deep convolutional neural network for segmenting neuroanatomy. NeuroImage, 170:434-445, 2018. 2 +[104] Qi Wang, Julius Steiglechner, Tobias Lindig, Benjamin Bender, Klaus Scheffler, and Gabriele Lohmann. SuperResolution for Ultra High-Field MR Images. In Medical Imaging with Deep Learning (MIDL), 2022. 2 +[105] Yaqing Wang, Quanming Yao, James T. Kwok, and Lionel M. Ni. Generalizing from a Few Examples. ACM Computing Surveys, 53(3), jun 2020. 3 +[106] Mark W. Woolrich, Saad Jbabdi, Brian Patenaude, Michael Chappell, Salima Makni, Timothy Behrens, Christian Beckmann, Mark Jenkinson, and Stephen M. Smith. Bayesian analysis of neuroimaging data in FSL. NeuroImage, 45(1), 2009. 2 +[107] Xiao Yang, Xu Han, Eunbyung Park, Stephen Aylward, Roland Kwitt, and Marc Niethammer. Registration of Pathological Images. Simulation and synthesis in medical imaging (Workshop), 9968:97, 2016. 2 +[108] Huaxiu Yao, Long-Kai Huang, Linjun Zhang, Ying Wei, Li Tian, James Zou, Junzhou Huang, and Zhenhui Li. Improving Generalization in Meta-learning via Task Augmentation. In Proceedings of the 38th International Conference on Machine Learning, PMLR, pages 11887-11897. PMLR, jul 2021. 3 +[109] Chandan Ganesh Bangalore Yogananda, Benjamin C. Wagner, Gowtham K. Murugesan, Ananth Madhuranthakam, and Joseph A. Maldjian. A deep learning pipeline for automatic skull stripping and brain segmentation. Proceedings - International Symposium on Biomedical Imaging, pages 727-731, apr 2019. 2 +[110] Lin Yuan, Xue Wei, Hui Shen, Ling Li Zeng, and Dewen Hu. Multi-center brain imaging classification using a novel 3d cnn approach. IEEE Access, 6:49925-49934, sep 2018. 2 +[111] Chi Zhang, Guosheng Lin, Fayao Liu, Jiushuang Guo, Qingyao Wu, and Rui Yao. Pyramid Graph Networks With Connection Attentions for Region-Based One-Shot Semantic Segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9587-9595, 2019. 3 +[112] Chi Zhang, Guosheng Lin, Fayao Liu, Rui Yao, and Chunhua Shen. CANet: Class-Agnostic Segmentation Networks With Iterative Refinement and Attentive Few-Shot Learning. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 5217-5226, 2019. 3 +[113] Amy Zhao, Guha Balakrishnan, Frédo Durand, John V Guttag, and Adrian V Dalca. Data Augmentation Using Learned Transformations for One-Shot Medical Image Segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8543-8553, 2019. 3 +[114] Guoyan Zheng, Chengwen Chu, Daniel L. Belavy, Bulat Ibragimov, Robert Korez, Tomaz Vrtovec, Hugo Hutt, + +Richard Everson, Judith Meakin, Isabel López Andrade, Ben Glocker, Hao Chen, Qi Dou, Pheng Ann Heng, Chunliang Wang, Daniel Forsberg, Aleš Neubert, Jurgen Fripp, Martin Urschler, Darko Stern, Maria Wimmer, Alexey A. Novikov, Hui Cheng, Gabriele Armbrecht, Dieter Felsenberg, and Shuo Li. Evaluation and comparison of 3D intervertebral disc localization and segmentation methods for 3D T2 MR data: A grand challenge. Medical Image Analysis, 35:327-344, jan 2017. 3 \ No newline at end of file diff --git a/2023/Neuralizer_ General Neuroimage Analysis Without Re-Training/images.zip b/2023/Neuralizer_ General Neuroimage Analysis Without Re-Training/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1b43b07f2bbc6d0253777adf82a8a16555425c28 --- /dev/null +++ b/2023/Neuralizer_ General Neuroimage Analysis Without Re-Training/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5f8cc5a5c1da5e521db36f883a5f5abaa4aa14bdb534512a6d1f3a2993e63bbb +size 478978 diff --git a/2023/Neuralizer_ General Neuroimage Analysis Without Re-Training/layout.json b/2023/Neuralizer_ General Neuroimage Analysis Without Re-Training/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b6ba43dfb12e81561e8120ce8c63e27313779b81 --- /dev/null +++ b/2023/Neuralizer_ General Neuroimage Analysis Without Re-Training/layout.json @@ -0,0 +1,11069 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 102, + 103, + 492, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 103, + 492, + 121 + ], + "spans": [ + { + "bbox": [ + 102, + 103, + 492, + 121 + ], + "type": "text", + "content": "Neuralizer: General Neuroimage Analysis without Re-Training" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 93, + 144, + 272, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 144, + 272, + 184 + ], + "spans": [ + { + "bbox": [ + 93, + 144, + 272, + 184 + ], + "type": "text", + "content": "Steffen Czolbe \nUniversity of Copenhagen and MGH \nper.sc@di.ku.dk" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 299, + 144, + 500, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 144, + 500, + 184 + ], + "spans": [ + { + "bbox": [ + 299, + 144, + 500, + 184 + ], + "type": "text", + "content": "Adrian V. Dalca \nMIT, and MGH, Harvard Medical School \nadalca@mit.edu" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 213, + 191, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 191, + 224 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 191, + 224 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 239, + 288, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 239, + 288, + 393 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 288, + 393 + ], + "type": "text", + "content": "Neuroimage processing tasks like segmentation, reconstruction, and registration are central to the study of neuroscience. Robust deep learning strategies and architectures used to solve these tasks are often similar. Yet, when presented with a new task or a dataset with different visual characteristics, practitioners most often need to train a new model, or fine-tune an existing one. This is a time-consuming process that poses a substantial barrier for the thousands of neuroscientists and clinical researchers who often lack the resources or machine-learning expertise to train deep learning models. In practice, this leads to a lack of adoption of deep learning, and neuroscience tools being dominated by classical frameworks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 394, + 287, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 394, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 46, + 394, + 287, + 525 + ], + "type": "text", + "content": "We introduce Neuralizer, a single model that generalizes to previously unseen neuroimaging tasks and modalities without the need for re-training or fine-tuning. Tasks do not have to be known a priori, and generalization happens in a single forward pass during inference. The model can solve processing tasks across multiple image modalities, acquisition methods, and datasets, and generalize to tasks and modalities it has not been trained on. Our experiments on coronal slices show that when few annotated subjects are available, our multi-task network outperforms task-specific baselines without training on the task." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 548, + 127, + 561 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 548, + 127, + 561 + ], + "spans": [ + { + "bbox": [ + 47, + 548, + 127, + 561 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 570, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 653 + ], + "type": "text", + "content": "Computational methods for the processing and analysis of neuroimages have enabled a deep understanding of the human brain. The field has also led to advanced patient care by facilitating non-invasive methods of diagnosis and treatment. Recent deep learning research promises to substantially increase the accuracy and speed of neuroimaging analysis methods." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "text", + "content": "A drawback of most current deep-learning-based approaches is that each model is limited to solving the task it has been trained on, on the type of data it has been trained on. Generalization to new tasks and domains, such as different acquisition protocols or new segmentation, is" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 307, + 212, + 545, + 376 + ], + "blocks": [ + { + "bbox": [ + 307, + 212, + 545, + 376 + ], + "lines": [ + { + "bbox": [ + 307, + 212, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 307, + 212, + 545, + 376 + ], + "type": "image", + "image_path": "0d6f1df7db5ce816a8997a97cf48f586d95cde4ec34df20cc3098306dcfeb348.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 384, + 545, + 472 + ], + "lines": [ + { + "bbox": [ + 304, + 384, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 304, + 384, + 545, + 472 + ], + "type": "text", + "content": "Figure 1. Neuralizer can solve a broad range of image processing tasks, including new ones not seen during training, with a single model by conditioning the prediction on a context set of examples. After training on a diverse set of tasks, the model can generalize to new tasks in a single forward pass without re-training or finetuning. The model is highly flexible, requiring no prior definition of the set of tasks, and can be conditioned with context sets of any length." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 510, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 712 + ], + "type": "text", + "content": "a main barrier to adoption [62]. Performing neuroimaging tasks like segmentation, registration, reconstruction, or motion correction requires different models for each processing step, despite operating on the same input data and methods exhibiting strong similarities in network architecture [13,45,85]. Yet, designing and training models to solve these tasks on each dataset is prohibitively expensive. To train a deep learning model, a dataset needs to be compiled and often manually annotated, and the network, training, and data loading logic needs to be implemented. All these steps generally require machine learning and neuroimaging expertise. In addition, computational resources like specialized graphics processing hardware and software infrastructure needs to be available. These requirements are particularly problematic in clinical research settings due to a high cost of annotation and a lack of machine learning expertise and hardware." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6217" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 91, + 108, + 210 + ], + "blocks": [ + { + "bbox": [ + 48, + 69, + 106, + 89 + ], + "lines": [ + { + "bbox": [ + 48, + 69, + 106, + 89 + ], + "spans": [ + { + "bbox": [ + 48, + 69, + 106, + 89 + ], + "type": "text", + "content": "Binary Segmentation" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 48, + 91, + 108, + 210 + ], + "lines": [ + { + "bbox": [ + 48, + 91, + 108, + 210 + ], + "spans": [ + { + "bbox": [ + 48, + 91, + 108, + 210 + ], + "type": "image", + "image_path": "56d19b5b78f39d78a44fd8928324dd971826bf61499ca0993eaf583905d221ce.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 111, + 91, + 170, + 209 + ], + "blocks": [ + { + "bbox": [ + 111, + 70, + 170, + 89 + ], + "lines": [ + { + "bbox": [ + 111, + 70, + 170, + 89 + ], + "spans": [ + { + "bbox": [ + 111, + 70, + 170, + 89 + ], + "type": "text", + "content": "Modality Transfer" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 111, + 91, + 170, + 209 + ], + "lines": [ + { + "bbox": [ + 111, + 91, + 170, + 209 + ], + "spans": [ + { + "bbox": [ + 111, + 91, + 170, + 209 + ], + "type": "image", + "image_path": "aa387042cbb38ad0a6273668c17cbe0fca1f2f4bec1ba4e7c815bd8d8f33c964.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 173, + 91, + 233, + 209 + ], + "blocks": [ + { + "bbox": [ + 173, + 70, + 233, + 89 + ], + "lines": [ + { + "bbox": [ + 173, + 70, + 233, + 89 + ], + "spans": [ + { + "bbox": [ + 173, + 70, + 233, + 89 + ], + "type": "text", + "content": "Super Resolution" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 173, + 91, + 233, + 209 + ], + "lines": [ + { + "bbox": [ + 173, + 91, + 233, + 209 + ], + "spans": [ + { + "bbox": [ + 173, + 91, + 233, + 209 + ], + "type": "image", + "image_path": "731ce385d19df72b67f1a1a12be3bb813740d5dd54b50cdb32502882386a7c11.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 236, + 91, + 295, + 209 + ], + "blocks": [ + { + "bbox": [ + 236, + 91, + 295, + 209 + ], + "lines": [ + { + "bbox": [ + 236, + 91, + 295, + 209 + ], + "spans": [ + { + "bbox": [ + 236, + 91, + 295, + 209 + ], + "type": "image", + "image_path": "0b946413a57cb4cd20535159782b092b9f31e4c6dfc29d695bf49be468742941.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 75, + 219, + 515, + 230 + ], + "lines": [ + { + "bbox": [ + 75, + 219, + 515, + 230 + ], + "spans": [ + { + "bbox": [ + 75, + 219, + 515, + 230 + ], + "type": "text", + "content": "Figure 2. Example neuroimaging tasks and modalities included in our dataset (top: input images, bottom: output images)." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 486, + 91, + 545, + 149 + ], + "blocks": [ + { + "bbox": [ + 425, + 70, + 482, + 89 + ], + "lines": [ + { + "bbox": [ + 425, + 70, + 482, + 89 + ], + "spans": [ + { + "bbox": [ + 425, + 70, + 482, + 89 + ], + "type": "text", + "content": "Denoising & Bias Correction" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 496, + 72, + 536, + 84 + ], + "lines": [ + { + "bbox": [ + 496, + 72, + 536, + 84 + ], + "spans": [ + { + "bbox": [ + 496, + 72, + 536, + 84 + ], + "type": "text", + "content": "Inpainting" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 486, + 91, + 545, + 149 + ], + "lines": [ + { + "bbox": [ + 486, + 91, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 486, + 91, + 545, + 149 + ], + "type": "image", + "image_path": "5f1ad77f626c725a8e3c83f5188002286511cc7e13419b3bd57fe2407e99be61.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 299, + 91, + 357, + 209 + ], + "blocks": [ + { + "bbox": [ + 236, + 70, + 295, + 89 + ], + "lines": [ + { + "bbox": [ + 236, + 70, + 295, + 89 + ], + "spans": [ + { + "bbox": [ + 236, + 70, + 295, + 89 + ], + "type": "text", + "content": "Skull Stripping" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 299, + 91, + 357, + 209 + ], + "lines": [ + { + "bbox": [ + 299, + 91, + 357, + 209 + ], + "spans": [ + { + "bbox": [ + 299, + 91, + 357, + 209 + ], + "type": "image", + "image_path": "ce281d27c4b8fb1d5a5b82c3e4a2b1806493c3e5efd855a9e36a4600e8be7f9c.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 360, + 91, + 419, + 209 + ], + "blocks": [ + { + "bbox": [ + 298, + 70, + 358, + 89 + ], + "lines": [ + { + "bbox": [ + 298, + 70, + 358, + 89 + ], + "spans": [ + { + "bbox": [ + 298, + 70, + 358, + 89 + ], + "type": "text", + "content": "Motion Correction" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 360, + 91, + 419, + 209 + ], + "lines": [ + { + "bbox": [ + 360, + 91, + 419, + 209 + ], + "spans": [ + { + "bbox": [ + 360, + 91, + 419, + 209 + ], + "type": "image", + "image_path": "71ec469392539fb001ae6c6292a29ab52749c354535f3b4befd191f1ded5954b.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 423, + 91, + 482, + 209 + ], + "blocks": [ + { + "bbox": [ + 361, + 70, + 419, + 89 + ], + "lines": [ + { + "bbox": [ + 361, + 70, + 419, + 89 + ], + "spans": [ + { + "bbox": [ + 361, + 70, + 419, + 89 + ], + "type": "text", + "content": "Undersampled Reconstruction" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 423, + 91, + 482, + 209 + ], + "lines": [ + { + "bbox": [ + 423, + 91, + 482, + 209 + ], + "spans": [ + { + "bbox": [ + 423, + 91, + 482, + 209 + ], + "type": "image", + "image_path": "2c1406656e88e80f756c46e6d0610e1127065af82e500c46459793f007c831ca.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 486, + 151, + 545, + 209 + ], + "blocks": [ + { + "bbox": [ + 486, + 151, + 545, + 209 + ], + "lines": [ + { + "bbox": [ + 486, + 151, + 545, + 209 + ], + "spans": [ + { + "bbox": [ + 486, + 151, + 545, + 209 + ], + "type": "image", + "image_path": "05a870d5c09eb569cdf29856d0a5c2fcbf320ea4b4df440c07baa87c25716a7e.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 47, + 251, + 111, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 251, + 111, + 262 + ], + "spans": [ + { + "bbox": [ + 47, + 251, + 111, + 262 + ], + "type": "text", + "content": "Contribution" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 46, + 270, + 287, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 270, + 287, + 341 + ], + "spans": [ + { + "bbox": [ + 46, + 270, + 287, + 341 + ], + "type": "text", + "content": "We introduce Neuralizer, a general-purpose neuroimaging model that can solve a broad range of neuroimaging tasks on diverse image modalities (Fig. 2), without the need for task-specific training or fine-tuning. Neuralizer can solve new tasks, unseen during training, using a set of examples of the new task at inference (Fig. 1)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 46, + 342, + 287, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 342, + 287, + 413 + ], + "spans": [ + { + "bbox": [ + 46, + 342, + 287, + 413 + ], + "type": "text", + "content": "Neuralizer involves a convolutional architecture (Fig. 3), that takes as input a context set of examples that define the processing task, and thus does not require prior specification of the tasks. The method enables single-pass generalization during inference and can process any number of reference images in a single pass to inform the prediction." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 46, + 415, + 287, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 415, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 46, + 415, + 287, + 545 + ], + "type": "text", + "content": "As a first method tackling task generalization in neuroimaging, we focus on analyzing the capabilities of such system and presenting general insights, and limit our experiments to 2D. We evaluate our model by comparing the single-pass generalization performance to task-specific baselines conditioned on an equivalent amount of data. We find that Neuralizer outperforms the baselines on tasks where " + }, + { + "bbox": [ + 46, + 415, + 287, + 545 + ], + "type": "inline_equation", + "content": "\\leq 32" + }, + { + "bbox": [ + 46, + 415, + 287, + 545 + ], + "type": "text", + "content": " labeled examples are available, despite never training on the task. When generalizing to new segmentation protocols, Neuralizer matches the performance of baselines trained directly on the dataset." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 47, + 558, + 134, + 570 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 558, + 134, + 570 + ], + "spans": [ + { + "bbox": [ + 47, + 558, + 134, + 570 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 46, + 578, + 287, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 578, + 287, + 638 + ], + "spans": [ + { + "bbox": [ + 46, + 578, + 287, + 638 + ], + "type": "text", + "content": "We give a short introduction to neuroimaging tasks, terminology, and methods. We then provide an overview of fundamental methods for adapting a model to multiple domains, including multi-task learning, few-shot learning, fine-tuning, and data synthesis." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 47, + 647, + 167, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 167, + 659 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 167, + 659 + ], + "type": "text", + "content": "2.1. Neuroimage analysis" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 46, + 666, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 666, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 666, + 287, + 713 + ], + "type": "text", + "content": "Neuroimage analysis employs computational techniques to study the structure and function of the human brain. Common imaging techniques are structural magnetic resonance imaging (MRI), functional MRI, diffusion tensor" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 251, + 545, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 251, + 545, + 323 + ], + "spans": [ + { + "bbox": [ + 304, + 251, + 545, + 323 + ], + "type": "text", + "content": "imaging (DTI), computed tomography (CT), and Positron emission tomography (PET). Each imaging method can create diverse images with different characteristics and contrasts, which are further diversified depending on the properties of the acquisition site [60, 110], device, protocol, imaging sequence [56], and use of contrast agents [9,36]." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 324, + 545, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 324, + 545, + 432 + ], + "spans": [ + { + "bbox": [ + 304, + 324, + 545, + 432 + ], + "type": "text", + "content": "To analyze these images, a variety of processing tasks are most often combined in a processing pipeline. Common processing tasks include anatomical segmentation [12, 18, 22, 23, 31, 82, 98, 103], skull stripping [49, 49, 53, 87, 94, 109], defacing [2, 39], registration [5-7, 10, 20, 28, 45, 47, 54], modality transfer [79, 80, 97], in-painting [41, 68, 69, 78, 107], super-resolution [21, 58, 72, 73, 104], reconstruction, and de-noising [58, 70, 92], bias field removal [35, 59], surface fitting [48] and parcellation [91, 99]." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 433, + 545, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 433, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 304, + 433, + 545, + 612 + ], + "type": "text", + "content": "Multiple toolboxes provide a suite of interoperable software components, most implementing classical optimization strategies. Widely used toolboxes include Freesurfer [31], FSL [51, 95, 106], SPM [34], CIVET [3], BrainSuite [90], HCP pipeline [100], and BrainIAK [57]. Deep-learning-based methods are starting to be included because of their improved accuracy and shorter runtime [13, 49]. While these methods provide solutions for common neuroimaging applications, most are limited to a single task and few modalities. Developers need to manually update the pipelines to include new processing tasks and to support a wider variety of image modalities. This process requires extensive technical expertise and computational resources, often not available to the clinical neuroscientists focusing on scientific questions." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 306, + 623, + 420, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 623, + 420, + 635 + ], + "spans": [ + { + "bbox": [ + 306, + 623, + 420, + 635 + ], + "type": "text", + "content": "2.2. Multi-task learning" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": "Multi-Task Learning (MTL) frameworks solve multiple tasks simultaneously by exploiting similarities between related tasks [17]. MTL can improve performance and reduce computational cost and development time compared to designing task-specific solutions [26, 88]. In neuroimaging, MTL networks were recently proposed for the simultaneous" + } + ] + } + ], + "index": 30 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6218" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": "segmentation and classification of brain tumors by training a single network with separate prediction heads associated with the different tasks [25, 38]. This strategy is challenging to scale as the number of tasks increases, requires prior determination of the set of tasks, and importantly does not enable generalization of the model to new tasks. With Neuralizer, we build on these methods to achieve scalable MTL, without the need for multiple network heads, and importantly with the ability to generalize to new tasks and modalities." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 200, + 126, + 213 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 200, + 126, + 213 + ], + "spans": [ + { + "bbox": [ + 47, + 200, + 126, + 213 + ], + "type": "text", + "content": "2.3. Fine-tuning" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 218, + 289, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 218, + 289, + 386 + ], + "spans": [ + { + "bbox": [ + 46, + 218, + 289, + 386 + ], + "type": "text", + "content": "To tackle problems in the limited data scenarios frequent in medical imaging, neural networks can be pre-trained on a related task with high data availability and then fine-tuned for specific tasks. For example, a common approach involves taking a Res-Net [43] trained on ImageNet [24] and fine-tuning part of the network for a new task [50, 55, 101]. For medical imaging, networks pre-trained on large sets of medical images are available [19], and fine-tuning them to new tasks results in shortened training time and higher accuracy [4, 71]. However, fine-tuning also requires machine learning expertise and computational resources, most often not available in clinical research. Additionally, in scenarios with small datasets, fine-tuning models trained on large vision datasets can be harmful [83]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 393, + 154, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 393, + 154, + 406 + ], + "spans": [ + { + "bbox": [ + 47, + 393, + 154, + 406 + ], + "type": "text", + "content": "2.4. Few-shot learning" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 412, + 289, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 412, + 289, + 663 + ], + "spans": [ + { + "bbox": [ + 46, + 412, + 289, + 663 + ], + "type": "text", + "content": "Few-shot models generate predictions from just a few labeled examples [66, 84, 86, 105], or in the case of zero-shot methods [14], none at all. Many of these methods require training or fine-tuning. In computer vision, several methods pass a query image, along with a set of support images and labels as input to the model [66, 89, 96, 102]. Natural image segmentation methods [67, 112] use single image-label pairs [61, 111] as support or aggregate information from a larger support set [63]. Recent few-shot learning methods in the medical image segmentation setting [11, 29, 30] operate on a specific anatomical region in a single image modality [42, 114]. Similar Prior-Data Fitted Networks (PFNs) are fitted to multiple datasets at once to learn the training and prediction algorithm [77]. During training, this strategy draws a dataset, a set of data points and their labels from it, masks one of the labels and predicts it. The resulting model aims to generalize to new datasets. PFNs have only been applied to low-dimensional and tabular data [46]. Our solution builds on ideas from these methods, but aims to solve a much larger range of diverse image-to-image tasks on neuroimages of many modalities." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 671, + 225, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 225, + 684 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 225, + 684 + ], + "type": "text", + "content": "2.5. Data augmentation and synthesis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "content": "Data augmentation increases the diversity of training data by augmenting or modifying existing data [85, 113]. It" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "text", + "content": "improves model robustness to input variability that may not be available in the original training data. In neuroimaging, arbitrary image modalities can be simulated by synthesis of images without requiring any real data [13, 16, 45, 49, 93]. In meta-learning, data augmentation can further be used to generate entirely new tasks [15, 65, 108]. We use data augmentations and further expand existing methods by developing rich neuroimaging task augmentations for generalization to unseen neuroimaging tasks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 191, + 376, + 204 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 191, + 376, + 204 + ], + "spans": [ + { + "bbox": [ + 306, + 191, + 376, + 204 + ], + "type": "text", + "content": "3. Neuralizer" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 211, + 547, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 211, + 547, + 272 + ], + "spans": [ + { + "bbox": [ + 304, + 211, + 547, + 272 + ], + "type": "text", + "content": "We introduce Neuralizer, a multi-task model for neuroimage analysis tasks. In this section, we first define the training framework and adaptations necessary to operate on a diverse range of tasks and input types. We then introduce the model architecture, training, and inference strategies." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 279, + 473, + 291 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 279, + 473, + 291 + ], + "spans": [ + { + "bbox": [ + 306, + 279, + 473, + 291 + ], + "type": "text", + "content": "3.1. Generalizable multi-task model" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 297, + 545, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 297, + 545, + 346 + ], + "spans": [ + { + "bbox": [ + 304, + 297, + 545, + 346 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 304, + 297, + 545, + 346 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 297, + 545, + 346 + ], + "type": "text", + "content": " represent a set of tasks, with a subset of tasks " + }, + { + "bbox": [ + 304, + 297, + 545, + 346 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{seen}}" + }, + { + "bbox": [ + 304, + 297, + 545, + 346 + ], + "type": "text", + "content": " seen during training. Each task consists of input-output image pairs " + }, + { + "bbox": [ + 304, + 297, + 545, + 346 + ], + "type": "inline_equation", + "content": "(x_{t}, y_{t})" + }, + { + "bbox": [ + 304, + 297, + 545, + 346 + ], + "type": "text", + "content": " from potentially multiple underlying datasets with input and output spaces " + }, + { + "bbox": [ + 304, + 297, + 545, + 346 + ], + "type": "inline_equation", + "content": "x_{t} \\in \\mathcal{X}, y_{t} \\in \\mathcal{Y}" + }, + { + "bbox": [ + 304, + 297, + 545, + 346 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 346, + 545, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 346, + 545, + 429 + ], + "spans": [ + { + "bbox": [ + 304, + 346, + 545, + 429 + ], + "type": "text", + "content": "To enable generalization to unseen tasks, we condition the model on a context set " + }, + { + "bbox": [ + 304, + 346, + 545, + 429 + ], + "type": "inline_equation", + "content": "C_t = \\{(x_{t,i},y_{t,i})\\}_{i=1}^N" + }, + { + "bbox": [ + 304, + 346, + 545, + 429 + ], + "type": "text", + "content": " of input-output image pairs passed to the model alongside the prediction task. The context set defines the desired task, and can vary in size " + }, + { + "bbox": [ + 304, + 346, + 545, + 429 + ], + "type": "inline_equation", + "content": "|C_t| = N" + }, + { + "bbox": [ + 304, + 346, + 545, + 429 + ], + "type": "text", + "content": " and is re-sampled from the underlying task-datasets for each input. Fig. 1 gives an example for a modality transfer task." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 430, + 545, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 430, + 545, + 477 + ], + "spans": [ + { + "bbox": [ + 304, + 430, + 545, + 477 + ], + "type": "text", + "content": "We employ a neural network " + }, + { + "bbox": [ + 304, + 430, + 545, + 477 + ], + "type": "inline_equation", + "content": "g_{\\theta}(x_t, C_t) = y_t" + }, + { + "bbox": [ + 304, + 430, + 545, + 477 + ], + "type": "text", + "content": " with weights " + }, + { + "bbox": [ + 304, + 430, + 545, + 477 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 430, + 545, + 477 + ], + "type": "text", + "content": " that applies the task defined by context set " + }, + { + "bbox": [ + 304, + 430, + 545, + 477 + ], + "type": "inline_equation", + "content": "C_t" + }, + { + "bbox": [ + 304, + 430, + 545, + 477 + ], + "type": "text", + "content": " to the input neuroimage " + }, + { + "bbox": [ + 304, + 430, + 545, + 477 + ], + "type": "inline_equation", + "content": "x_t" + }, + { + "bbox": [ + 304, + 430, + 545, + 477 + ], + "type": "text", + "content": ". We optimize the network using supervised training with the loss" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 315, + 487, + 545, + 511 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 487, + 545, + 511 + ], + "spans": [ + { + "bbox": [ + 315, + 487, + 545, + 511 + ], + "type": "interline_equation", + "content": "\\mathcal {L} \\left(T _ {\\text {s e e n}}; \\theta\\right) = \\mathbb {E} _ {t \\in T _ {\\text {s e e n}}} \\left[ \\mathbb {E} _ {\\left(x _ {t}, y _ {t}, C _ {t}\\right)} \\left[ \\mathcal {L} _ {t} \\left(y _ {t}, g _ {\\theta} \\left(x _ {t}, C _ {t}\\right)\\right) \\right] \\right], \\tag {1}", + "image_path": "5ec41ff2485feb1d3bd110e3c172d8573ef22ce7bfd122da538364c59050bf55.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 512, + 471, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 512, + 471, + 523 + ], + "spans": [ + { + "bbox": [ + 306, + 512, + 471, + 523 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 512, + 471, + 523 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_t" + }, + { + "bbox": [ + 306, + 512, + 471, + 523 + ], + "type": "text", + "content": " is a task-specific loss function." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 532, + 440, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 532, + 440, + 544 + ], + "spans": [ + { + "bbox": [ + 306, + 532, + 440, + 544 + ], + "type": "text", + "content": "3.2. Design for diverse tasks" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 550, + 545, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 550, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 304, + 550, + 545, + 586 + ], + "type": "text", + "content": "To process different tasks with a single model, we carefully select the loss function, neuroimage encodings, and generation of the training set for each task type." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 596, + 547, + 717 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 596, + 547, + 717 + ], + "spans": [ + { + "bbox": [ + 304, + 596, + 547, + 717 + ], + "type": "text", + "content": "Loss function. Neuralizer solves both segmentation tasks (e.g. anatomical segmentation and skull-stripping via a brain mask), more general and image-to-image tasks (e.g. denoising). We use the Soft Dice Loss [76] for segmentation-like tasks, and the pixel-wise Mean Squared Error " + }, + { + "bbox": [ + 304, + 596, + 547, + 717 + ], + "type": "inline_equation", + "content": "MSE(y_{t}, g(x_{t}, C_{t})) = \\frac{1}{2\\sigma^{2}} \\sum_{p} [y_{tp} - g(x_{t}, C_{t})_{p}]^{2}" + }, + { + "bbox": [ + 304, + 596, + 547, + 717 + ], + "type": "text", + "content": " with balancing hyperparameter " + }, + { + "bbox": [ + 304, + 596, + 547, + 717 + ], + "type": "inline_equation", + "content": "\\sigma^{2}" + }, + { + "bbox": [ + 304, + 596, + 547, + 717 + ], + "type": "text", + "content": " for other tasks. As the network optimizes multiple tasks during training, the balance of the loss terms can dramatically affect the optimization and resulting performance." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6219" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 69, + 255, + 228 + ], + "blocks": [ + { + "bbox": [ + 47, + 69, + 255, + 228 + ], + "lines": [ + { + "bbox": [ + 47, + 69, + 255, + 228 + ], + "spans": [ + { + "bbox": [ + 47, + 69, + 255, + 228 + ], + "type": "image", + "image_path": "f98a6eaa912876ee02c28a05c983027b5f7e96ca4889be678ceb5cc3377b2e0a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 235, + 546, + 281 + ], + "lines": [ + { + "bbox": [ + 45, + 235, + 546, + 281 + ], + "spans": [ + { + "bbox": [ + 45, + 235, + 546, + 281 + ], + "type": "text", + "content": "Figure 3. Neuralizer consists of 7 Pairwise-Conv-Avg blocks (right), arranged in a U-Net-like [76, 85] configuration (left). Each Pairwise-Conv-Avg block enables interaction between the input image and the image pairs present in the context set. The block consists of a residual unit, pairwise convolution of each context member with the target, and an averaging of results across the context set to update the representation. The architecture is invariant to context size " + }, + { + "bbox": [ + 45, + 235, + 546, + 281 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 45, + 235, + 546, + 281 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 258, + 69, + 547, + 227 + ], + "blocks": [ + { + "bbox": [ + 258, + 69, + 547, + 227 + ], + "lines": [ + { + "bbox": [ + 258, + 69, + 547, + 227 + ], + "spans": [ + { + "bbox": [ + 258, + 69, + 547, + 227 + ], + "type": "image", + "image_path": "cca2e8bdd235ecf870b8147531fe736c1aeb1d4f8d5074958ed0bc830beffdd5.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 300, + 287, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 300, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 46, + 300, + 287, + 384 + ], + "type": "text", + "content": "Input and output encoding. For Neuralizer to work on both segmentation and image-to-image tasks, we facilitate simultaneous input of multiple image modalities and masks. We design the input space " + }, + { + "bbox": [ + 46, + 300, + 287, + 384 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 46, + 300, + 287, + 384 + ], + "type": "text", + "content": " to accept floating point value images with three channels, and zero-pad any channels unnecessary for a specific task. The output space " + }, + { + "bbox": [ + 46, + 300, + 287, + 384 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}" + }, + { + "bbox": [ + 46, + 300, + 287, + 384 + ], + "type": "text", + "content": " follows the same design but uses only one channel." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 397, + 287, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 397, + 287, + 505 + ], + "spans": [ + { + "bbox": [ + 46, + 397, + 287, + 505 + ], + "type": "text", + "content": "Training dataset creation. At each training iteration, we first sample a task " + }, + { + "bbox": [ + 46, + 397, + 287, + 505 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 397, + 287, + 505 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 46, + 397, + 287, + 505 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{seen}}" + }, + { + "bbox": [ + 46, + 397, + 287, + 505 + ], + "type": "text", + "content": ", selecting the task-specific dataset (Tab. 1). From this dataset, we sample the input image, ground truth output, and image pairs for the context set. To increase the range of images that can be used to condition the trained model, the image modalities and acquisition protocols of entries of the context set can differ from the input image for some tasks. Supplemental section G contains a detailed description of the training data generator." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 514, + 160, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 514, + 160, + 526 + ], + "spans": [ + { + "bbox": [ + 47, + 514, + 160, + 526 + ], + "type": "text", + "content": "3.3. Model architecture" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 533, + 287, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 287, + 592 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 287, + 592 + ], + "type": "text", + "content": "Fig. 3 shows the Neuralizer network architecture, adapted with the concurrently developed [15] - a method that focuses on solving broad segmentation tasks. As the architecture is independent of the task, we omit the task subscript in this section." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": "The input image " + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": " and the image pairs of the context set " + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "inline_equation", + "content": "C_i = (x_i, y_i), i = 1, \\dots, N" + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": " are first passed through an embedding layer consisting of a single " + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": " convolution with learnable kernels " + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "inline_equation", + "content": "e_x, e_C" + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": ", to obtain the representations " + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "inline_equation", + "content": "r_x = x * e_x" + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "inline_equation", + "content": "r_{C_i} = \\operatorname{cat}(x_i, y_i) * e_C" + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "inline_equation", + "content": "*" + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": " is the convolution operator. This combines each context image pair to a joint representation " + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "inline_equation", + "content": "r_{C_i}" + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": " and maps all representations to a uniform channel width " + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": ", which is constant throughout the model. Next, we process the representations using multiple Pairwise-Conv-Avg Blocks (explained below), ar" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 306, + 331, + 545, + 577 + ], + "blocks": [ + { + "bbox": [ + 305, + 299, + 545, + 322 + ], + "lines": [ + { + "bbox": [ + 305, + 299, + 545, + 322 + ], + "spans": [ + { + "bbox": [ + 305, + 299, + 545, + 322 + ], + "type": "text", + "content": "Table 1. Tasks, Modalities, Datasets, and Segmentation classes used in this paper, and involved in training Neuralizer." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 306, + 331, + 545, + 577 + ], + "lines": [ + { + "bbox": [ + 306, + 331, + 545, + 577 + ], + "spans": [ + { + "bbox": [ + 306, + 331, + 545, + 577 + ], + "type": "table", + "html": "
TasksModalities
Binary SegmentationT1-w.
Modality TransferT2-w.
Super ResolutionMRA
Skull StrippingPD
Motion CorrectionFLAIR
Undersampled ReconstructionADC
Denoising & Bias correctionDWI
InpaintingDTI (17 dir.)
DatasetsSegmentation Classes
OASIS 3 [47, 74]Freesufer protocol, 31
BRATS [8, 9, 75]classes [13, 31]
IXI [1]
ATLAS R2.0 [64]Manually-annotated Hammers
Hammers Atlas [40]Atlas, 96 classes [27, 37, 40]
WMH Challenge [56]
ISLES2022 [81]Brainmasks [31, 49]
", + "image_path": "710a2ed9cdd57888092773dd27337213128c15bbfcc6915aac2628ad7a718919.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 603, + 545, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 603, + 545, + 674 + ], + "spans": [ + { + "bbox": [ + 304, + 603, + 545, + 674 + ], + "type": "text", + "content": "ranged as a U-Net-like configuration [76,85] to exploit multiple scales. The output " + }, + { + "bbox": [ + 304, + 603, + 545, + 674 + ], + "type": "inline_equation", + "content": "r_x^{\\mathrm{out}}" + }, + { + "bbox": [ + 304, + 603, + 545, + 674 + ], + "type": "text", + "content": " of the final Pairwise-Conv-Avg Block is processed by a residual unit [43] and a final " + }, + { + "bbox": [ + 304, + 603, + 545, + 674 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 304, + 603, + 545, + 674 + ], + "type": "text", + "content": " conv layer to map to one output channel. All residual units consist of two " + }, + { + "bbox": [ + 304, + 603, + 545, + 674 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 304, + 603, + 545, + 674 + ], + "type": "text", + "content": " conv layers, a shortcut connection, and GELU activation functions [44]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 677, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 546, + 713 + ], + "type": "text", + "content": "Compared to standard CNNs, Neuralizer uses a mechanism to enable knowledge transfer from the context set to the input image. We design the Pairwise-Conv-Avg Block" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6220" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 312 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 312 + ], + "type": "text", + "content": "(Fig. 3, right) to model this interaction. The block maps from representations of the target input " + }, + { + "bbox": [ + 47, + 72, + 289, + 312 + ], + "type": "inline_equation", + "content": "r_x^{\\mathrm{in}}" + }, + { + "bbox": [ + 47, + 72, + 289, + 312 + ], + "type": "text", + "content": " and context pairs " + }, + { + "bbox": [ + 47, + 72, + 289, + 312 + ], + "type": "inline_equation", + "content": "r_{C_i}^{\\mathrm{in}}" + }, + { + "bbox": [ + 47, + 72, + 289, + 312 + ], + "type": "text", + "content": " to output representations " + }, + { + "bbox": [ + 47, + 72, + 289, + 312 + ], + "type": "inline_equation", + "content": "r_x^{\\mathrm{out}}" + }, + { + "bbox": [ + 47, + 72, + 289, + 312 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 72, + 289, + 312 + ], + "type": "inline_equation", + "content": "r_{C_i}^{\\mathrm{out}}" + }, + { + "bbox": [ + 47, + 72, + 289, + 312 + ], + "type": "text", + "content": " of the same size. First, we process each input separately with a residual unit to obtain " + }, + { + "bbox": [ + 47, + 72, + 289, + 312 + ], + "type": "inline_equation", + "content": "r_x^{\\mathrm{int}} = \\mathrm{ResUnit}_x(r_x^{\\mathrm{in}})" + }, + { + "bbox": [ + 47, + 72, + 289, + 312 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 72, + 289, + 312 + ], + "type": "inline_equation", + "content": "r_{C_i}^{\\mathrm{int}} = \\mathrm{ResUnit}_C(r_{C_i}^{\\mathrm{in}})" + }, + { + "bbox": [ + 47, + 72, + 289, + 312 + ], + "type": "text", + "content": ". The residual units, which involve two convolutions, operate on the context representations and have shared weights. Second, we pairwise concatenate the context representations with the target representation on the channel dimension: " + }, + { + "bbox": [ + 47, + 72, + 289, + 312 + ], + "type": "inline_equation", + "content": "p_i = \\mathrm{cat}(r_x^{\\mathrm{int}}, r_{C_i}^{\\mathrm{int}})" + }, + { + "bbox": [ + 47, + 72, + 289, + 312 + ], + "type": "text", + "content": ". We combine the pairwise representations and reduce the channel size back to " + }, + { + "bbox": [ + 47, + 72, + 289, + 312 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 47, + 72, + 289, + 312 + ], + "type": "text", + "content": " using a " + }, + { + "bbox": [ + 47, + 72, + 289, + 312 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 47, + 72, + 289, + 312 + ], + "type": "text", + "content": " convolution with learnable kernel " + }, + { + "bbox": [ + 47, + 72, + 289, + 312 + ], + "type": "inline_equation", + "content": "k_x" + }, + { + "bbox": [ + 47, + 72, + 289, + 312 + ], + "type": "text", + "content": ", and update the target representation by averaging across context members " + }, + { + "bbox": [ + 47, + 72, + 289, + 312 + ], + "type": "inline_equation", + "content": "r_x^{\\mathrm{out}} = r_x^{\\mathrm{int}} + \\frac{1}{N}\\sum_{i=1}^{N}p_i * k_x" + }, + { + "bbox": [ + 47, + 72, + 289, + 312 + ], + "type": "text", + "content": ". The context representations are updated with a separate kernel " + }, + { + "bbox": [ + 47, + 72, + 289, + 312 + ], + "type": "inline_equation", + "content": "r_{C_i}^{\\mathrm{out}} = r_{C_i}^{\\mathrm{int}} + p_i * k_C" + }, + { + "bbox": [ + 47, + 72, + 289, + 312 + ], + "type": "text", + "content": ". We then re-size the outputs of a Pairwise-Conv-Avg Block before feeding them as input the next block. We experimented with attention-based and weighted average approaches but found that they did not lead to an increased generalization to unseen tasks." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 319, + 163, + 331 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 319, + 163, + 331 + ], + "spans": [ + { + "bbox": [ + 47, + 319, + 163, + 331 + ], + "type": "text", + "content": "3.4. Task augmentations" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 337, + 287, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 337, + 287, + 433 + ], + "spans": [ + { + "bbox": [ + 46, + 337, + 287, + 433 + ], + "type": "text", + "content": "To further diversify the training dataset, we employ task augmentations [15], a group of transformations applied at random to the input, output, and context images. The objective is to increase the diversity of tasks to discourage the model from merely memorizing the tasks in the training data. A list of all task augmentations is summarized in Tab. 2, with more detailed descriptions and visual examples in Supplement C." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 441, + 115, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 441, + 115, + 451 + ], + "spans": [ + { + "bbox": [ + 47, + 441, + 115, + 451 + ], + "type": "text", + "content": "3.5. Inference" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 459, + 287, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 459, + 287, + 543 + ], + "spans": [ + { + "bbox": [ + 46, + 459, + 287, + 543 + ], + "type": "text", + "content": "During inference, we supply an input image " + }, + { + "bbox": [ + 46, + 459, + 287, + 543 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 46, + 459, + 287, + 543 + ], + "type": "text", + "content": " and a context set " + }, + { + "bbox": [ + 46, + 459, + 287, + 543 + ], + "type": "inline_equation", + "content": "C_i" + }, + { + "bbox": [ + 46, + 459, + 287, + 543 + ], + "type": "text", + "content": " from the desired task. Given these inputs, a simple feed-forward pass through the model provides the prediction " + }, + { + "bbox": [ + 46, + 459, + 287, + 543 + ], + "type": "inline_equation", + "content": "\\hat{y} = g(x,C)" + }, + { + "bbox": [ + 46, + 459, + 287, + 543 + ], + "type": "text", + "content": ". To further increase accuracy at test-time, we use context-set bootstrapping [15]. We also increase the context set by sampling with replacement from the context set, and add small affine augmentations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 553, + 128, + 566 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 553, + 128, + 566 + ], + "spans": [ + { + "bbox": [ + 47, + 553, + 128, + 566 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 573, + 287, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 573, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 47, + 573, + 287, + 609 + ], + "type": "text", + "content": "We first compare Neuralizer with task-specific networks, which require substantial expertise and compute. We then analyze the effect of the size of the context set, and the" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 48, + 647, + 287, + 720 + ], + "blocks": [ + { + "bbox": [ + 113, + 627, + 221, + 638 + ], + "lines": [ + { + "bbox": [ + 113, + 627, + 221, + 638 + ], + "spans": [ + { + "bbox": [ + 113, + 627, + 221, + 638 + ], + "type": "text", + "content": "Table 2. Task Augmentations" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 647, + 287, + 720 + ], + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 720 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 720 + ], + "type": "table", + "html": "
Task Augmentations
IntensityMappingSyntheticModality
SobelFilterMaskInvert
MaskContourMaskDilation
PermuteChannelsDuplicateChannels
", + "image_path": "4f4e7f6a0c2995295d52c931361db2593a7a3ed70b08e7459af5482c1a44326e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": "multi-task generalization to unseen segmentation protocols and image modalities. For this first method of large-scale multi-task generalization in neuroimaging, we conduct the experiments on 2D image slices." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 127, + 351, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 127, + 351, + 138 + ], + "spans": [ + { + "bbox": [ + 306, + 127, + 351, + 138 + ], + "type": "text", + "content": "4.1. Data" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 145, + 547, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 145, + 547, + 408 + ], + "spans": [ + { + "bbox": [ + 304, + 145, + 547, + 408 + ], + "type": "text", + "content": "To create a diverse dataset encompassing a multitude of different modalities, acquisition protocols, devices, and tasks, we pool neuroimages from the public datasets OASIS3 [47, 74], BRATS [8, 9, 75], Atlas R2.0 [64], Hammers Atlas [40], IXI [1], ISLES2022 [81], and the White Matter Hyperintensities Challenge [56]. We segment all subjects with Synthseg [13, 31]. Based on the segmentation, we affinely align the images to the MNI 152 template space [32, 33], and resample to " + }, + { + "bbox": [ + 304, + 145, + 547, + 408 + ], + "type": "inline_equation", + "content": "1\\mathrm{mm}" + }, + { + "bbox": [ + 304, + 145, + 547, + 408 + ], + "type": "text", + "content": " isometric resolution at a size of " + }, + { + "bbox": [ + 304, + 145, + 547, + 408 + ], + "type": "inline_equation", + "content": "192 \\times 224 \\times 192\\mathrm{mm}" + }, + { + "bbox": [ + 304, + 145, + 547, + 408 + ], + "type": "text", + "content": ". We perform manual quality control of the segmentation and registration by ensuring no segmented areas fall outside of the cropped volume and discard subjects failing this check (4 subjects). We extract a coronal slice of " + }, + { + "bbox": [ + 304, + 145, + 547, + 408 + ], + "type": "inline_equation", + "content": "192 \\times 192\\mathrm{mm}" + }, + { + "bbox": [ + 304, + 145, + 547, + 408 + ], + "type": "text", + "content": ", bisecting the frontal Brain stem, Hippocampus, Thalamus, and Lateral ventricles. We rescale image intensities to the " + }, + { + "bbox": [ + 304, + 145, + 547, + 408 + ], + "type": "inline_equation", + "content": "[0, 1]" + }, + { + "bbox": [ + 304, + 145, + 547, + 408 + ], + "type": "text", + "content": " interval using dataset-specific percentiles. For full head images, we create a brain mask with Synthstrip [31, 49]. The final dataset contains 2,282 subjects with 15,911 images and segmentation masks across 8 modalities. Subjects of the seven original datasets are split into " + }, + { + "bbox": [ + 304, + 145, + 547, + 408 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 304, + 145, + 547, + 408 + ], + "type": "text", + "content": " for training and validation, " + }, + { + "bbox": [ + 304, + 145, + 547, + 408 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 304, + 145, + 547, + 408 + ], + "type": "text", + "content": " test, with a minimum of 15 test subjects per dataset." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 415, + 362, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 415, + 362, + 426 + ], + "spans": [ + { + "bbox": [ + 306, + 415, + 362, + 426 + ], + "type": "text", + "content": "4.2. Models" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 434, + 545, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 434, + 545, + 519 + ], + "spans": [ + { + "bbox": [ + 304, + 434, + 545, + 519 + ], + "type": "text", + "content": "Neuralizer-seen. This Neuralizer model includes all tasks available during training. We use this model to evaluate the performance on unseen scans from tasks and modalities that have been included in the training. The model uses the 4-stage architecture shown in Fig. 3 with 64 channels per layer. During training, the context size " + }, + { + "bbox": [ + 304, + 434, + 545, + 519 + ], + "type": "inline_equation", + "content": "|C_i|" + }, + { + "bbox": [ + 304, + 434, + 545, + 519 + ], + "type": "text", + "content": " is sampled from " + }, + { + "bbox": [ + 304, + 434, + 545, + 519 + ], + "type": "inline_equation", + "content": "\\mathcal{U}_{\\{1,32\\}}" + }, + { + "bbox": [ + 304, + 434, + 545, + 519 + ], + "type": "text", + "content": " at each iteration." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 532, + 547, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 532, + 547, + 591 + ], + "spans": [ + { + "bbox": [ + 304, + 532, + 547, + 591 + ], + "type": "text", + "content": "Neuralizer-unseen. To evaluate Neuralizer performance on tasks and modalities it has not been trained on, we train a family of Neuralizer models where a single task or modality is excluded from the training set. The model architecture of Neuralizer-unseen is identical to Neuralizer-seen." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "content": "Baseline-seen. As no established baseline for multi-task and multi-modality models in neuroimaging can tackle the number of tasks we aim for, we compare Neuralizer to an ensemble of task-specific U-Nets [76, 85]. However, training one model for each task and modality requires overwhelming computational resources. To reduce the computational requirement, we follow previous modality-agnostic models [13, 49] and train each model on multiple input modalities. This lowers the number of models to be" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "text", + "content": "6221" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 69, + 165, + 154 + ], + "blocks": [ + { + "bbox": [ + 50, + 69, + 165, + 154 + ], + "lines": [ + { + "bbox": [ + 50, + 69, + 165, + 154 + ], + "spans": [ + { + "bbox": [ + 50, + 69, + 165, + 154 + ], + "type": "image", + "image_path": "c0fd6b8c4ab958a4866fa6e48d308a57f8a7fbfd10c48d98718dc57123b19619.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 176, + 70, + 291, + 153 + ], + "blocks": [ + { + "bbox": [ + 176, + 70, + 291, + 153 + ], + "lines": [ + { + "bbox": [ + 176, + 70, + 291, + 153 + ], + "spans": [ + { + "bbox": [ + 176, + 70, + 291, + 153 + ], + "type": "image", + "image_path": "2fb8191a3e762802546feea73218571c4caa76075a1acd257d07e3c463b4abb4.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 304, + 70, + 419, + 153 + ], + "blocks": [ + { + "bbox": [ + 304, + 70, + 419, + 153 + ], + "lines": [ + { + "bbox": [ + 304, + 70, + 419, + 153 + ], + "spans": [ + { + "bbox": [ + 304, + 70, + 419, + 153 + ], + "type": "image", + "image_path": "06758ead4dccdb38c713dc105b69754416a18ba9690bc8090a141449b4d41348.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 432, + 70, + 546, + 153 + ], + "blocks": [ + { + "bbox": [ + 432, + 70, + 546, + 153 + ], + "lines": [ + { + "bbox": [ + 432, + 70, + 546, + 153 + ], + "spans": [ + { + "bbox": [ + 432, + 70, + 546, + 153 + ], + "type": "image", + "image_path": "b3174a3fc67059c080ee525a6a4e5a2fd8a44c5325a7bc8c8afa8eae1bf28f61.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 48, + 161, + 165, + 247 + ], + "blocks": [ + { + "bbox": [ + 48, + 161, + 165, + 247 + ], + "lines": [ + { + "bbox": [ + 48, + 161, + 165, + 247 + ], + "spans": [ + { + "bbox": [ + 48, + 161, + 165, + 247 + ], + "type": "image", + "image_path": "d9c8c77962f286661c55d3d0979620d756a1d7b37b834ea90c78fa357d8d0643.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 173, + 161, + 291, + 247 + ], + "blocks": [ + { + "bbox": [ + 173, + 161, + 291, + 247 + ], + "lines": [ + { + "bbox": [ + 173, + 161, + 291, + 247 + ], + "spans": [ + { + "bbox": [ + 173, + 161, + 291, + 247 + ], + "type": "image", + "image_path": "20705a6f1271ea2172df0597c863044496fa17068ac696b8daddf9a8b5d09a51.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 301, + 161, + 419, + 247 + ], + "blocks": [ + { + "bbox": [ + 301, + 161, + 419, + 247 + ], + "lines": [ + { + "bbox": [ + 301, + 161, + 419, + 247 + ], + "spans": [ + { + "bbox": [ + 301, + 161, + 419, + 247 + ], + "type": "image", + "image_path": "c732eb5110f46f99c6fb1d1f34776829913682328cbb8c87f2661694e0ef1c00.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 429, + 161, + 547, + 247 + ], + "blocks": [ + { + "bbox": [ + 429, + 161, + 547, + 247 + ], + "lines": [ + { + "bbox": [ + 429, + 161, + 547, + 247 + ], + "spans": [ + { + "bbox": [ + 429, + 161, + 547, + 247 + ], + "type": "image", + "image_path": "ae87f88a2dd257a524a2cb1e347b10913b005c6aea99e1da76a96975419356fc.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 47, + 255, + 165, + 339 + ], + "blocks": [ + { + "bbox": [ + 47, + 255, + 165, + 339 + ], + "lines": [ + { + "bbox": [ + 47, + 255, + 165, + 339 + ], + "spans": [ + { + "bbox": [ + 47, + 255, + 165, + 339 + ], + "type": "image", + "image_path": "ad495076e3e82ba2851a7bb69018eb31854c3e2d26c32a0eb69559d4bc4668cd.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 346, + 547, + 401 + ], + "lines": [ + { + "bbox": [ + 46, + 346, + 547, + 401 + ], + "spans": [ + { + "bbox": [ + 46, + 346, + 547, + 401 + ], + "type": "text", + "content": "Figure 4. Performance of multi-task Neuralizer and the task-specific baselines on each task, averaged across all modalities in the test set. The tasks being evaluated were included in the training of Neuralizer-seen (orange), held out in Neuralizer-unseen (blue), and specifically trained on by each task-specific baseline (gray). The x-axis is the size of the train/context set, and the y-axis is the Dice/PSNR score. Some points on the x-axis are omitted for better visibility. 'All' refers to all available train data for the task, ranging from 249 to 2,282 subjects depending on the task. The bars denote standard deviation across modalities. We extract results for T1 scans in Supplement D." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 179, + 255, + 291, + 339 + ], + "blocks": [ + { + "bbox": [ + 179, + 255, + 291, + 339 + ], + "lines": [ + { + "bbox": [ + 179, + 255, + 291, + 339 + ], + "spans": [ + { + "bbox": [ + 179, + 255, + 291, + 339 + ], + "type": "image", + "image_path": "de6954e83fcf785cfb55148edc6831cb10ae20987179d155976a63a0f0f80a1f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 301, + 255, + 499, + 339 + ], + "blocks": [ + { + "bbox": [ + 301, + 255, + 499, + 339 + ], + "lines": [ + { + "bbox": [ + 301, + 255, + 499, + 339 + ], + "spans": [ + { + "bbox": [ + 301, + 255, + 499, + 339 + ], + "type": "image", + "image_path": "789c960037e4757f84c39eb1536dca37df9d20e00206bf3bebb3a821046ae5bb.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 423, + 287, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 423, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 46, + 423, + 287, + 483 + ], + "type": "text", + "content": "trained to one per task, segmentation class, and modalitytransfer output modality. To compare Baseline-seen with Neuralizer-unseen given an equal amount of data, we train baselines with training set sizes of " + }, + { + "bbox": [ + 46, + 423, + 287, + 483 + ], + "type": "inline_equation", + "content": "\\{1,2,4,8,16,32,\\mathrm{all}\\}" + }, + { + "bbox": [ + 46, + 423, + 287, + 483 + ], + "type": "text", + "content": " and employ standard data augmentation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 483, + 287, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 483, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 483, + 287, + 567 + ], + "type": "text", + "content": "We use a 4-stage U-Net architecture with one residual block per layer. The channel width is tuned experimentally for each training dataset size. We select 256 channels when all data is available for training, and 64 channels otherwise. Using larger U-Nets resulted in overfitting and lower performance. Supplement H summarizes model parameter counts and inference costs." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 574, + 111, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 574, + 111, + 588 + ], + "spans": [ + { + "bbox": [ + 47, + 574, + 111, + 588 + ], + "type": "text", + "content": "4.3. Training" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 594, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 677 + ], + "type": "text", + "content": "We use supervised training, task-specific loss functions, and weigh the MSE loss by selecting " + }, + { + "bbox": [ + 46, + 594, + 287, + 677 + ], + "type": "inline_equation", + "content": "\\sigma^2 = 0.05" + }, + { + "bbox": [ + 46, + 594, + 287, + 677 + ], + "type": "text", + "content": ", resulting in the loss terms being of similar magnitude. All models are trained with a batch size of 8, a learning rate of " + }, + { + "bbox": [ + 46, + 594, + 287, + 677 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 46, + 594, + 287, + 677 + ], + "type": "text", + "content": ", and the ADAM optimizer [52]. To speed up training, we undersample tasks that the model learns quickly, with sampling weights given in Supplement I." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "content": "In addition to the task augmentations, we use data augmentations via random affine movements, random elastic deformations, and random flips along the sagittal plane." + } + ] + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 305, + 420, + 545, + 571 + ], + "blocks": [ + { + "bbox": [ + 305, + 420, + 545, + 571 + ], + "lines": [ + { + "bbox": [ + 305, + 420, + 545, + 571 + ], + "spans": [ + { + "bbox": [ + 305, + 420, + 545, + 571 + ], + "type": "image", + "image_path": "67d6f6e7811437e39ebb21b71eff8e192b68c9aee78c225c8b8526877df1ae3c.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 578, + 547, + 689 + ], + "lines": [ + { + "bbox": [ + 304, + 578, + 547, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 578, + 547, + 689 + ], + "type": "text", + "content": "Figure 5. Results averaged across tasks, expressed as relative performance compared to the baseline trained on all data. The tasks being evaluated were included in the training of Neuralizer-seen (orange), held out in Neuralizer-unseen (blue), and specifically trained on by each task-specific baseline (gray). The x-axis is the size of the train/context set, and the y-axis is the relative score. Some points on the x-axis are omitted for better visibility. 'All' refers to all available data for the task, ranging from 249 to 2,282 subjects depending on the task. Bars: standard deviation across tasks/modalities." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6222" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 167 + ], + "type": "text", + "content": "For Baseline-seen, we reuse the augmentations but remove those that introduce uncertainty in the desired output. The training time for Neuralizer is 7 days on a single A100 GPU. The training time of the baseline models is capped at 5 days. All models use early stopping, ending the training after 25 epochs exhibiting no decrease in validation loss. The model with the lowest validation loss is used for further evaluation on the test set." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 183, + 121, + 196 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 183, + 121, + 196 + ], + "spans": [ + { + "bbox": [ + 47, + 183, + 121, + 196 + ], + "type": "text", + "content": "4.4. Evaluation" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 205, + 289, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 205, + 289, + 313 + ], + "spans": [ + { + "bbox": [ + 46, + 205, + 289, + 313 + ], + "type": "text", + "content": "We evaluate the Dice coefficient for the segmentation and skull stripping tasks, and the Peak Signal-to-Noise Ratio (PSNR) for the image-to-image tasks on the test set. As the low-data regime is of particular interest, we measure performance as a function of context set size for the Neuralizer models and use training set size as an analog for the U-Net models. We evaluate context sets of up to 32 subjects. Larger context sets are possible but come at a linear cost in memory." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 327, + 241, + 341 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 327, + 241, + 341 + ], + "spans": [ + { + "bbox": [ + 47, + 327, + 241, + 341 + ], + "type": "text", + "content": "4.5. Experiment 1: Baseline Comparison" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 349, + 287, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 349, + 287, + 433 + ], + "spans": [ + { + "bbox": [ + 46, + 349, + 287, + 433 + ], + "type": "text", + "content": "To assess if the proposed multi-task approach is competitive with task-specific models, we evaluate the performance and runtime of Neuralizer-seen, Neuralizer-unseen, and Baseline-seen on the test-set of each task. For Neuralizer-unseen, we withhold image modalities using a leave-one-out strategy during training and evaluate on the unseen modalities at test time." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 456, + 287, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 456, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 46, + 456, + 287, + 624 + ], + "type": "text", + "content": "Results. We display the results by task, averaged across modalities in Fig. 4. We also provide an evaluation of using just the T1 modality in Supplement D, since many task-specific networks in neuroimage analysis literature focus on T1 images. We further aggregate performance across all tasks in Fig. 5, and provide tabular results in Supplement E. Both Neuralizer models outperform most task-specific baselines trained on up to 32 samples. When training the baselines on all available data, the baselines outperform Neuralizer-seen by 2 percentage points in relative performance, and Neuralizer-unseen by 3 percentage points. The loss in performance when generalizing to an unseen modality (between Neuralizer-seen and Neuralizer-unseen) is less than 2 percentage points for all context set sizes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 627, + 287, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 627, + 287, + 687 + ], + "spans": [ + { + "bbox": [ + 46, + 627, + 287, + 687 + ], + "type": "text", + "content": "Training the baseline model to convergence on 32 samples took on average " + }, + { + "bbox": [ + 46, + 627, + 287, + 687 + ], + "type": "inline_equation", + "content": "28.2 \\pm 16.6" + }, + { + "bbox": [ + 46, + 627, + 287, + 687 + ], + "type": "text", + "content": " hours per task, using one A100 GPU. Since Neuralizer only requires inference for a new task, it is orders of magnitude faster, requiring less than 0.1 seconds on a GPU and less than 3 seconds on a CPU." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "We provide qualitative samples of the predictions from Neuralizer-seen model in Supplement A, Figures 6-8." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 72, + 512, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 512, + 85 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 512, + 85 + ], + "type": "text", + "content": "4.6. Experiment 2: Context set size analysis" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 90, + 545, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 90, + 545, + 222 + ], + "spans": [ + { + "bbox": [ + 304, + 90, + 545, + 222 + ], + "type": "text", + "content": "We assess the few-shot setting that is prevalent in neuroimage analysis, where few annotated images are often available for a new task. We evaluate performance as a function of the number of labeled samples. For Neuralizer, we evaluate the model with context-set sizes of " + }, + { + "bbox": [ + 304, + 90, + 545, + 222 + ], + "type": "inline_equation", + "content": "\\{1,2,4,8,16,32\\}" + }, + { + "bbox": [ + 304, + 90, + 545, + 222 + ], + "type": "text", + "content": " unique subjects from the test set. For the baseline, we trained models with reduced training set sizes of the same amount of subjects. To reduce the effect of random training subject selection, we train three separate baselines with " + }, + { + "bbox": [ + 304, + 90, + 545, + 222 + ], + "type": "inline_equation", + "content": "n = 1" + }, + { + "bbox": [ + 304, + 90, + 545, + 222 + ], + "type": "text", + "content": ", two baselines with " + }, + { + "bbox": [ + 304, + 90, + 545, + 222 + ], + "type": "inline_equation", + "content": "n = 2" + }, + { + "bbox": [ + 304, + 90, + 545, + 222 + ], + "type": "text", + "content": ", and average results of models with the same " + }, + { + "bbox": [ + 304, + 90, + 545, + 222 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 90, + 545, + 222 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 230, + 545, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 230, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 304, + 230, + 545, + 327 + ], + "type": "text", + "content": "Results. Tab. 4 and Figs. 4, 5 illustrate the results. For all models, prediction accuracy increases with the availability of labeled data, with diminishing returns. For both Neuralizer models, a context set size of one achieves more than " + }, + { + "bbox": [ + 304, + 230, + 545, + 327 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 304, + 230, + 545, + 327 + ], + "type": "text", + "content": " of the performance attainable with all data. For most tasks, the baseline performs overall worse than both Neuralizer models when " + }, + { + "bbox": [ + 304, + 230, + 545, + 327 + ], + "type": "inline_equation", + "content": "\\leq 32" + }, + { + "bbox": [ + 304, + 230, + 545, + 327 + ], + "type": "text", + "content": " labeled samples are available but achieves the best overall performance on larger datasets." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 333, + 545, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 333, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 305, + 333, + 545, + 358 + ], + "type": "text", + "content": "4.7. Experiment 3: Generalization to a new segmentation protocol" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 363, + 545, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 363, + 545, + 495 + ], + "spans": [ + { + "bbox": [ + 304, + 363, + 545, + 495 + ], + "type": "text", + "content": "The Hammers Atlas dataset [27,37,40] provides an alternative anatomical segmentation protocol to the widely-used Freesurfer segmentation available for most subjects in the dataset. The shape, size, and amount of annotated regions in the protocols differ drastically. A different image acquisition site also leads to differences in visual characteristics. We use the Hammers Atlas dataset to evaluate Neuralizer-unseen by entirely withholding the dataset and its annotations from training. We evaluate the Dice coefficient of the 14 major anatomical segmentation classes present in the center and right half of the coronal slice." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 503, + 545, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 503, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 304, + 503, + 545, + 624 + ], + "type": "text", + "content": "Results. Tab. 3 illustrates the results. Neuralizer-unseen performs similarly to Neuralizer-seen and the baseline, while not requiring lengthy re-training or fine-tuning on the Hammers Atlas dataset, and not having seen the segmentation protocol. All three models achieve a mean Dice coefficient of 0.84. The largest performance difference is in the third ventricle class, where both Neuralizer models outperform the baseline by at least 0.04 Dice. The Freesurfer segmentation protocol included in the training set of the Neuralizer models also contains a third ventricle class." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 634, + 375, + 646 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 634, + 375, + 646 + ], + "spans": [ + { + "bbox": [ + 306, + 634, + 375, + 646 + ], + "type": "text", + "content": "5. Discussion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 654, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 715 + ], + "type": "text", + "content": "Our experiments using modality and segmentation class hold-outs show that Neuralizer can generalize well to unseen neuroimaging tasks. Across all context set sizes, the generalization loss between seen and unseen modalities and segmentation classes is less than 2 percentage points across" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6223" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 70, + 547, + 148 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 547, + 148 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 547, + 148 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 547, + 148 + ], + "type": "table", + "html": "
ModelTask SeenSegmentation Class (Hammers Atlas)Mean (std)
HipPAGSTGMIGFuGStmInsPCGThaCC3VPrGPoGALG
Baseline-seen✓.88.86.93.92.79.87.82.87.90.80.68.83.77.82.84 (.07)
Neuralizer-seen✓.88.86.92.92.76.88.83.85.90.82.73.86.77.81.84 (.07)
Neuralizer-unseenX.88.87.93.91.78.87.82.85.90.81.72.85.78.81.84 (.06)
", + "image_path": "2dc92878a2517af1737d2685b4a700d0b1814d99759c0e49175f1a5eb5548391.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 156, + 546, + 179 + ], + "lines": [ + { + "bbox": [ + 46, + 156, + 546, + 179 + ], + "spans": [ + { + "bbox": [ + 46, + 156, + 546, + 179 + ], + "type": "text", + "content": "Table 3. Segmentation of the Hammers Atlas dataset. For Neuralizer-unseen, this dataset and segmentation protocol is withheld from training. Evaluation of major labels located in the center and right of the coronal slice. See Supplement F for class abbreviations." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 200, + 286, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 200, + 286, + 272 + ], + "spans": [ + { + "bbox": [ + 46, + 200, + 286, + 272 + ], + "type": "text", + "content": "Experiments 1 and 2. On the smaller held-out HammersAtlas segmentation dataset, we find that Neuralizer can generalize to unseen tasks with similar performance. These results show promise that a single Neuralizer model can perform multiple neuroimaging tasks, including generalization to new inference tasks not seen during training." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 274, + 286, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 274, + 286, + 405 + ], + "spans": [ + { + "bbox": [ + 46, + 274, + 286, + 405 + ], + "type": "text", + "content": "In settings with 32 or fewer labeled example images, Neuralizer-unseen outperforms task-specific baselines despite never having seen the task or modality at train time, and taking nearly no effort or compute compared to the baselines which require substantial expertise, manual labor, and compute resources. The performance difference is largest when only one labeled subject is available, but still present at 32 subjects (Fig. 5). Neuralizer provides a performance advantage on smaller datasets likely by exploiting neuroimaging similarities across the many other neuroimaging tasks and datasets available in training." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 406, + 286, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 406, + 286, + 561 + ], + "spans": [ + { + "bbox": [ + 46, + 406, + 286, + 561 + ], + "type": "text", + "content": "When training the baselines on all available data, they can outperform Neuralizer-seen and Neuralizer-unseen by at most 3 percentage points. The inflection point of identical performance between Neuralizer and the baselines is not covered by the range of context set sizes chosen for training and evaluation due to prohibitive computational costs and is an interesting direction of study. When large annotated datasets are available, the baselines performed best on most tasks. However, training task-specific models comes at a significant cost. As a first step in the proposed problem formulation, Neuralizer offers an alternative with near equal performance, while only requiring seconds to infer any task from the context set." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 572, + 105, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 572, + 105, + 585 + ], + "spans": [ + { + "bbox": [ + 47, + 572, + 105, + 585 + ], + "type": "text", + "content": "Limitations" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 592, + 286, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 592, + 286, + 688 + ], + "spans": [ + { + "bbox": [ + 46, + 592, + 286, + 688 + ], + "type": "text", + "content": "We made simplifying assumptions in this first paper demonstrating the potential of multi-task generalization in neuroimaging. The experiments are conducted on 2D data slices. In large part, we did this since running the hundreds of baselines in 3D would be infeasible on our compute cluster. Entire volumetric data also impose a challenging memory requirement on Neuralizer models. To tackle 3D data in the future, we plan to process multiple slices at a time." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "type": "text", + "content": "We affinely aligned the neuroimages of the context set to the target image. Early in Neuralizer development, we tried" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 200, + 545, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 200, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 304, + 200, + 545, + 259 + ], + "type": "text", + "content": "training on non-aligned inputs but found that it deteriorated performance. The need for affine alignment provides an obstacle to adoption. While existing affine-alignment tools are fast and can be employed, we also believe that this requirement can be removed with further development." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 259, + 545, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 259, + 545, + 331 + ], + "spans": [ + { + "bbox": [ + 304, + 259, + 545, + 331 + ], + "type": "text", + "content": "We originally experimented with tumor and lesion segmentation tasks but found this to be a more challenging scenario. Lesions are spatially heterogeneous, making learning from the context set much harder. We excluded tumor and lesion segmentation masks from the experiments, but plan to study this setting in the future." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 331, + 545, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 331, + 545, + 402 + ], + "spans": [ + { + "bbox": [ + 304, + 331, + 545, + 402 + ], + "type": "text", + "content": "While we demonstrate the proposed ideas on a broad range of tasks and modalities, neuroimage analysis can involve more domains, tasks, and populations, like image registration, surface-based tasks, CT image domains, and pediatric data. We plan to extend Neuralizer to tackle these in the future." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 412, + 378, + 425 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 412, + 378, + 425 + ], + "spans": [ + { + "bbox": [ + 306, + 412, + 378, + 425 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 433, + 545, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 433, + 545, + 575 + ], + "spans": [ + { + "bbox": [ + 304, + 433, + 545, + 575 + ], + "type": "text", + "content": "Neuralizer performs accurate rapid single-pass, multitask generalization, and even outperforms task-specific baselines in limited data scenarios. Even when a large amount of annotated data is available, Neuralizer often matches baseline performance despite not training on the data. Neuralizer provides clinical researchers and scientists with a single model to solve a wide range of neuroimaging tasks on images of many modalities, and can be easily adapted to new tasks without the prohibitive requirement of retraining or fine-tuning a task-specific model. We believe this will facilitate neuroscience analyses not currently possible." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 586, + 403, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 586, + 403, + 600 + ], + "spans": [ + { + "bbox": [ + 306, + 586, + 403, + 600 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "This work was performed during Steffen Czolbes's visit to the Athinoula A. Martins Center for Biomedical Imaging, MGH. The collaboration was supported by the EliteForsk Travel Grant of the Danish Ministry of Higher Education and Science. The work was further funded by the Novo Nordisk Foundation (grants no. NNF20OC0062606 and NNF17OC0028360), the Lundbeck Foundation (grant no. R218-2016-883), and NIH grants 1R01AG064027 and 1R01AG070988." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6224" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 91, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 58, + 91, + 135, + 101 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 91, + 135, + 101 + ], + "spans": [ + { + "bbox": [ + 58, + 91, + 135, + 101 + ], + "type": "text", + "content": "[1] IXI Dataset. 4, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 103, + 287, + 146 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 103, + 287, + 146 + ], + "spans": [ + { + "bbox": [ + 58, + 103, + 287, + 146 + ], + "type": "text", + "content": "[2] David Abramian and Anders Eklund. Refacing: Reconstructing anonymized facial features using GANS. International Symposium on Biomedical Imaging, 2019-April:1104-1108,apr 2019.2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 148, + 287, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 148, + 287, + 223 + ], + "spans": [ + { + "bbox": [ + 58, + 148, + 287, + 223 + ], + "type": "text", + "content": "[3] Yasser Ad-Dab'bagh, O Lyttelton, JS Muehlboeck, C Lepage, D Einarson, K Mok, O Ivanov, RD Vincent, J Lerch, and E Fombonne. The CIVET image-processing environment: a fully automated comprehensive pipeline for anatomical neuroimaging research. In Proceedings of the 12th annual meeting of the organization for human brain mapping, 2006. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 225, + 287, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 225, + 287, + 289 + ], + "spans": [ + { + "bbox": [ + 58, + 225, + 287, + 289 + ], + "type": "text", + "content": "[4] Laith Alzubaidi, Muthana Al-Amidie, Ahmed Al-Asadi, Amjad J. Humaidi, Omran Al-Shamma, Mohammed A. Fadhel, Jinglan Zhang, J. Santamaría, and Ye Duan. Novel Transfer Learning Approach for Medical Imaging with Limited Labeled Data. Cancers 2021, Vol. 13, Page 1590, 13(7):1590, mar 2021. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 292, + 287, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 292, + 287, + 346 + ], + "spans": [ + { + "bbox": [ + 58, + 292, + 287, + 346 + ], + "type": "text", + "content": "[5] Vincent Arsigny, Olivier Commowick, Xavier Pennec, and Nicholas Ayache. A log-euclidean framework for statistics on diffeomorphisms. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 924-931. Springer Verlag, 2006. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 58, + 347, + 286, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 347, + 286, + 368 + ], + "spans": [ + { + "bbox": [ + 58, + 347, + 286, + 368 + ], + "type": "text", + "content": "[6] John Ashburner. A fast diffeomorphic image registration algorithm. NeuroImage, 38(1):95-113, oct 2007. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 58, + 370, + 287, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 370, + 287, + 423 + ], + "spans": [ + { + "bbox": [ + 58, + 370, + 287, + 423 + ], + "type": "text", + "content": "[7] B. B. Avants, C. L. Epstein, M. Grossman, and J. C. Gee. Symmetric diffeomorphic image registration with cross-correlation: Evaluating automated labeling of elderly and neurodegenerative brain. Medical Image Analysis, 12(1):26-41, feb 2008. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 58, + 426, + 287, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 426, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 58, + 426, + 287, + 491 + ], + "type": "text", + "content": "[8] Spyridon Bakas, Hamed Akbari, Aristeidis Sotiras, Michel Bilello, Martin Rozycki, Justin S. Kirby, John B. Freymann, Keyvan Farahani, and Christos Davatzikos. Advancing The Cancer Genome Atlas glioma MRI collections with expert segmentation labels and radiomic features. Scientific Data, 4, sep 2017. 4, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 58, + 492, + 287, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 492, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 58, + 492, + 287, + 545 + ], + "type": "text", + "content": "[9] Spyridon Bakas, Bjoern Menze, and Others. Identifying the Best Machine Learning Algorithms for Brain Tumor Segmentation, Progression Assessment, and Overall Survival Prediction in the BRATS Challenge. arXiv, 124, nov 2018. 2, 4, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 548, + 287, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 548, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 53, + 548, + 287, + 601 + ], + "type": "text", + "content": "[10] Guha Balakrishnan, Amy Zhao, Mert R. Sabuncu, John Guttag, and Adrian V. Dalca. VoxelMorph: A Learning Framework for Deformable Medical Image Registration. IEEE Transactions on Medical Imaging, 38(8):1788-1800, feb 2019. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 603, + 287, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 603, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 53, + 603, + 287, + 666 + ], + "type": "text", + "content": "[11] Cheng Bian, Chenglang Yuan, Kai Ma, Shuang Yu, Dong Wei, and Yefeng Zheng. Domain Adaptation Meets Zero-Shot Learning: An Annotation-Efficient Approach to Multi-Modality Medical Image Segmentation. IEEE Transactions on Medical Imaging, 41(5):1043–1056, may 2022. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 670, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 670, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 53, + 670, + 286, + 712 + ], + "type": "text", + "content": "[12] Benjamin Billot, Douglas Greve, Koen Van Leemput, Bruce Fischl, Juan Eugenio Iglesias, and Adrian V Dalca. A learning strategy for contrast-agnostic mri segmentation. arXiv preprint arXiv:2003.01995, 2020. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 312, + 73, + 545, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 73, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 312, + 73, + 545, + 127 + ], + "type": "text", + "content": "[13] Benjamin Billot, Douglas N Greve, Oula Puonti, Axel Thielscher, Koen Van Leemput, Bruce Fischl, Adrian V Dalca, and Juan Eugenio Iglesias. SynthSeg: Domain Randomisation for Segmentation of Brain Scans of any Contrast and Resolution. jul 2021. 1, 2, 3, 4, 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 129, + 545, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 129, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 312, + 129, + 545, + 171 + ], + "type": "text", + "content": "[14] Maxime Bucher, Tuan-Hung VU, Matthieu Cord, and Patrick Pérez. Zero-Shot Semantic Segmentation. Advances in Neural Information Processing Systems, 32, 2019. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 173, + 545, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 173, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 312, + 173, + 545, + 216 + ], + "type": "text", + "content": "[15] Victor Ion Butoi, Jose Javier Gonzalez Ortiz, Tianyu Ma, Mert Sabuncu, John Guttag, and Adrian V. Dalca. Universeg: Universal medical image segmentation. arXiv preprint, 2023. 3, 4, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 217, + 545, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 217, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 312, + 217, + 545, + 281 + ], + "type": "text", + "content": "[16] Víctor M. Campello, Carlos Martín-Isla, Cristian Izquierdo, Steffen E. Petersen, Miguel A.González Ballester, and Karim Lekadir. Combining Multi-Sequence and Synthetic Images for Improved Segmentation of Late Gadolinium Enhancement Cardiac MRI. Lecture Notes in Computer Science, 12009 LNCS:290-299, 2020. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 283, + 545, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 283, + 545, + 314 + ], + "spans": [ + { + "bbox": [ + 312, + 283, + 545, + 314 + ], + "type": "text", + "content": "[17] Rich Caruana, Lorien Pratt, and Sebastian Thrun. Multitask Learning. Machine Learning 1997 28:1, 28(1):41-75, 1997. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 316, + 545, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 316, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 312, + 316, + 545, + 358 + ], + "type": "text", + "content": "[18] Hao Chen, Qi Dou, Lequan Yu, Jing Qin, and Pheng Ann Heng. VoxResNet: Deep voxelwise residual networks for brain segmentation from 3D MR images. NeuroImage, 170:446-455, apr 2018. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 361, + 545, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 361, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 312, + 361, + 545, + 381 + ], + "type": "text", + "content": "[19] Sihong Chen, Kai Ma, and Yefeng Zheng. Med3D: Transfer Learning for 3D Medical Image Analysis. apr 2019. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 383, + 545, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 383, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 312, + 383, + 545, + 415 + ], + "type": "text", + "content": "[20] Steffen Czolbe, Oswin Krause, and Aasa Feragen. Semantic similarity metrics for learned image registration. Proceedings of Machine Learning Research, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 416, + 545, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 416, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 312, + 416, + 545, + 459 + ], + "type": "text", + "content": "[21] Adrian V Dalca, Katherine L Bouman, William T Freeman, Natalia S Rost, Mert R Sabuncu, and Polina Golland. Medical image imputation from image collections. IEEE transactions on medical imaging, 38(2):504-514, 2018. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 460, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 460, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 312, + 460, + 545, + 514 + ], + "type": "text", + "content": "[22] Adrian V Dalca, John Guttag, and Mert R Sabuncu. Anatomical priors in convolutional networks for unsupervised biomedical segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 9290-9299, 2018. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 515, + 545, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 515, + 545, + 591 + ], + "spans": [ + { + "bbox": [ + 312, + 515, + 545, + 591 + ], + "type": "text", + "content": "[23] Adrian V Dalca, Evan Yu, Polina Golland, Bruce Fischl, Mert R Sabuncu, and Juan Eugenio Iglesias. Unsupervised deep learning for bayesian brain mri segmentation. In Medical Image Computing and Computer Assisted Intervention—MICCAI 2019: 22nd International Conference, Shenzhen, China, October 13–17, 2019, Proceedings, Part III 22, pages 356–365. Springer, 2019. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 593, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 593, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 312, + 593, + 545, + 635 + ], + "type": "text", + "content": "[24] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In Conference on Computer Vision and Pattern Recognition, pages 248-255, 2009. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 636, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 636, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 312, + 636, + 545, + 689 + ], + "type": "text", + "content": "[25] Francisco Javier Díaz-Pernas, Mario Martínez-Zarzuela, David González-Ortega, and Míriam Antón-Rodríguez. A Deep Learning Approach for Brain Tumor Classification and Segmentation Using a Multiscale Convolutional Neural Network. Healthcare, 9(2):153, feb 2021. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 692, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 692, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 312, + 692, + 545, + 712 + ], + "type": "text", + "content": "[26] Theodoras Evgeniou and Massimiliano Pontil. Regularized multi-task learning. Proceedings of the Tenth ACM" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6225" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 73, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 72, + 73, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 73, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 72, + 73, + 286, + 95 + ], + "type": "text", + "content": "SIGKDD International Conference on Knowledge Discovery and Data Mining, pages 109-117, 2004. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 96, + 287, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 96, + 287, + 139 + ], + "spans": [ + { + "bbox": [ + 53, + 96, + 287, + 139 + ], + "type": "text", + "content": "[27] Isabelle Faillenot, Rolf A. Heckemann, Maud Frot, and Alexander Hammers. Macroanatomy and 3D probabilistic atlas of the human insula. NeuroImage, 150:88-98, apr 2017. 4, 7, 24" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 141, + 286, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 141, + 286, + 194 + ], + "spans": [ + { + "bbox": [ + 53, + 141, + 286, + 194 + ], + "type": "text", + "content": "[28] Mirza Faisal Beg, Michael I Miller, Alain Trouvétrouv, and Laurent Younes. Computing Large Deformation Metric Mappings via Geodesic Flows of Diffeomorphisms. International Journal of Computer Vision, 61(2):139-157, 2005. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 198, + 286, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 198, + 286, + 251 + ], + "spans": [ + { + "bbox": [ + 53, + 198, + 286, + 251 + ], + "type": "text", + "content": "[29] Ruiwei Feng, Xiangshang Zheng, Tianxiang Gao, Jintai Chen, Wenzhe Wang, Danny Z. Chen, and Jian Wu. Interactive Few-Shot Learning: Limited Supervision, Better Medical Image Segmentation. IEEE Transactions on Medical Imaging, 40(10):2575-2588, oct 2021. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 253, + 286, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 253, + 286, + 296 + ], + "spans": [ + { + "bbox": [ + 53, + 253, + 286, + 296 + ], + "type": "text", + "content": "[30] Abdur R Feyjie, Reza Azad, Marco Pedersoli, Claude Kauffman, Ismail Ben Ayed, and Jose Dolz. Semi-supervised few-shot learning for medical image segmentation. mar 2020. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 298, + 286, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 298, + 286, + 319 + ], + "spans": [ + { + "bbox": [ + 53, + 298, + 286, + 319 + ], + "type": "text", + "content": "[31] B Fischl. FreeSurfer. NeuroImage, 62 (2), 774-781, 2012. 2, 4, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 321, + 286, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 321, + 286, + 364 + ], + "spans": [ + { + "bbox": [ + 53, + 321, + 286, + 364 + ], + "type": "text", + "content": "[32] VS Fonov, AC Evans, RC McKinstry, CR Almli, and DL Collins. Unbiased nonlinear average age-appropriate brain templates from birth to adulthood. NeuroImage, 47:S102, jul 2009. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 366, + 286, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 366, + 286, + 409 + ], + "spans": [ + { + "bbox": [ + 53, + 366, + 286, + 409 + ], + "type": "text", + "content": "[33] Vladimir Fonov, Alan C. Evans, Kelly Botteron, C. Robert Almli, Robert C. McKinstry, and D. Louis Collins. Unbiased average age-appropriate atlases for pediatric studies. NeuroImage, 54(1):313-327, jan 2011. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 411, + 281, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 411, + 281, + 421 + ], + "spans": [ + { + "bbox": [ + 53, + 411, + 281, + 421 + ], + "type": "text", + "content": "[34] Richard SJ Frackowiak. Human Brain Function. 2004. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 423, + 286, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 423, + 286, + 466 + ], + "spans": [ + { + "bbox": [ + 53, + 423, + 286, + 466 + ], + "type": "text", + "content": "[35] Tal Goldfryd, Shiri Gordon, and Tammy Riklin Raviv. Deep semi-supervised bias field correction of mr images. Proceedings - International Symposium on Biomedical Imaging, 2021-April:1836-1840, apr 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 468, + 286, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 468, + 286, + 511 + ], + "spans": [ + { + "bbox": [ + 53, + 468, + 286, + 511 + ], + "type": "text", + "content": "[36] Enhao Gong, John M. Pauly, Max Wintermark, and Greg Zaharchuk. Deep learning enables reduced gadolinium dose for contrast-enhanced brain MRI. Journal of Magnetic Resonance Imaging, 48(2):330-340, aug 2018. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 514, + 286, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 514, + 286, + 567 + ], + "spans": [ + { + "bbox": [ + 53, + 514, + 286, + 567 + ], + "type": "text", + "content": "[37] Ioannis S. Gousias, Daniel Rueckert, Rolf A. Heckemann, Leigh E. Dyet, James P. Boardman, A. David Edwards, and Alexander Hammers. Automatic segmentation of brain MRIs of 2-year-olds into 83 regions of interest. NeuroImage, 40(2):672-684, apr 2008. 4, 7, 24" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 569, + 286, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 569, + 286, + 643 + ], + "spans": [ + { + "bbox": [ + 53, + 569, + 286, + 643 + ], + "type": "text", + "content": "[38] Sachin Gupta, Narinder Singh Punn, Sanjay Kumar Sonbhadra, and Sonali Agarwal. MAG-Net: Multi-task Attention Guided Network for Brain Tumor Segmentation and Classification. Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), 13147 LNCS:3-15, 2021. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 647, + 286, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 647, + 286, + 689 + ], + "spans": [ + { + "bbox": [ + 53, + 647, + 286, + 689 + ], + "type": "text", + "content": "[39] J Hale, Nakeisha Schimke, and John Hale. Quickshear defacing for neuroimages Attack Graph Generation on HPC Clusters View project Quickshear Defacing for Neuroimages. Frontiers in psychiatry, 21, 2011. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 53, + 692, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 692, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 53, + 692, + 286, + 713 + ], + "type": "text", + "content": "[40] Alexander Hammers, Richard Allom, Matthias J. Koepp, Samantha L. Free, Ralph Myers, Louis Lemieux, Tejal N." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 331, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 331, + 73, + 545, + 116 + ], + "type": "text", + "content": "Mitchell, David J. Brooks, and John S. Duncan. Three-dimensional maximum probability atlas of the human brain, with particular reference to the temporal lobe. Human brain mapping, 19(4):224-247, aug 2003. 4, 5, 7, 24" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 118, + 545, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 118, + 545, + 182 + ], + "spans": [ + { + "bbox": [ + 312, + 118, + 545, + 182 + ], + "type": "text", + "content": "[41] Xu Han, Roland Kwitt, Stephen Aylward, Spyridon Bakas, Bjoern Menze, Alexander Asturias, Paul Vespa, John Van Horn, and Marc Niethammer. Brain extraction from normal and pathological images: A joint PCA/Image-Reconstruction approach. NeuroImage, 176:431-445, aug 2018. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 184, + 545, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 184, + 545, + 236 + ], + "spans": [ + { + "bbox": [ + 312, + 184, + 545, + 236 + ], + "type": "text", + "content": "[42] Stine Hansen, Srishti Gautam, Robert Jenssen, and Michael Kampffmeyer. Anomaly Detection-Inspired Few-Shot Medical Image Segmentation Through Self-Supervision With Supervoxels. Medical Image Analysis, 78, mar 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 239, + 545, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 239, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 312, + 239, + 545, + 281 + ], + "type": "text", + "content": "[43] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 3, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 284, + 545, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 284, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 312, + 284, + 545, + 304 + ], + "type": "text", + "content": "[44] Dan Hendrycks and Kevin Gimpel. Gaussian Error Linear Units (GELUs). jun 2016. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 305, + 545, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 305, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 312, + 305, + 545, + 358 + ], + "type": "text", + "content": "[45] Malte Hoffmann, Benjamin Billot, Douglas N. Greve, Juan Eugenio Iglesias, Bruce Fischl, and Adrian V. Dalca. SynthMorph: learning contrast-invariant registration without acquired images. IEEE Transactions on Medical Imaging, apr 2020. 1, 2, 3, 19" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 361, + 545, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 361, + 545, + 413 + ], + "spans": [ + { + "bbox": [ + 312, + 361, + 545, + 413 + ], + "type": "text", + "content": "[46] Noah Hollmann, Samuel Müller, Katharina Eggensperger, and Frank Hutter. TabPFN: A Transformer That Solves Small Tabular Classification Problems in a Second. In NeurIPS 2022 First Table Representation Workshop, jul 2022. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 415, + 545, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 415, + 545, + 458 + ], + "spans": [ + { + "bbox": [ + 312, + 415, + 545, + 458 + ], + "type": "text", + "content": "[47] Andrew Hoopes, Malte Hoffmann, Douglas N Greve, Bruce Fischl, John Guttag, and Adrian V Dalca. Learning the effect of registration hyperparameters with hypermorph. arXiv preprint arXiv:2203.16680, 2022. 2, 4, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 460, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 460, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 312, + 460, + 545, + 514 + ], + "type": "text", + "content": "[48] Andrew Hoopes, Juan Eugenio Iglesias, Bruce Fischl, Douglas Greve, and Adrian V Dalca. TopoFit: Rapid Reconstruction of Topologically-Correct Cortical Surfaces. Proceedings of Machine Learning Research-Under Review, pages 1-13, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 515, + 545, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 515, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 312, + 515, + 545, + 547 + ], + "type": "text", + "content": "[49] Andrew Hoopes, Jocelyn S. Mora, Adrian V. Dalca, Bruce Fischl, and Malte Hoffmann. SynthStrip: Skull-Stripping for Any Brain Image. mar 2022. 2, 3, 4, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 548, + 545, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 548, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 312, + 548, + 545, + 569 + ], + "type": "text", + "content": "[50] Minyoung Huh, Pulkit Agrawal, and Alexei A. Efros. What makes ImageNet good for transfer learning? aug 2016. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 571, + 545, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 571, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 312, + 571, + 545, + 602 + ], + "type": "text", + "content": "[51] Mark Jenkinson, Christian F. Beckmann, Timothy E.J. Behrens, Mark W. Woolrich, and Stephen M. Smith. FSL. NeuroImage, 62(2):782-790, aug 2012. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 312, + 604, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 604, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 312, + 604, + 545, + 635 + ], + "type": "text", + "content": "[52] Diederik P. Kingma and Jimmy Lei Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations, dec 2015. 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 312, + 636, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 636, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 312, + 636, + 545, + 689 + ], + "type": "text", + "content": "[53] Jens Kleesiek, Gregor Urban, Alexander Hubert, Daniel Schwarz, Klaus Maier-Hein, Martin Bendszus, and Armin Biller. Deep MRI brain extraction: A 3D convolutional neural network for skull stripping. NeuroImage, 129:460–469, apr 2016. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 312, + 692, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 692, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 692, + 545, + 713 + ], + "type": "text", + "content": "[54] Arno Klein, Jesper Andersson, Babak A. Ardekani, John Ashburner, Brian Avants, Ming Chang Chiang, Gary E." + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6226" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 72, + 289, + 713 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 72, + 72, + 287, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 72, + 287, + 148 + ], + "spans": [ + { + "bbox": [ + 72, + 72, + 287, + 148 + ], + "type": "text", + "content": "Christensen, D. Louis Collins, James Gee, Pierre Hellier, Joo Hyun Song, Mark Jenkinson, Claude Lepage, Daniel Rueckert, Paul Thompson, Tom Vercauteren, Roger P. Woods, J. John Mann, and Ramin V. Parsey. Evaluation of 14 nonlinear deformation algorithms applied to human brain MRI registration. NeuroImage, 46(3):786-802, jul 2009. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 150, + 288, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 150, + 288, + 205 + ], + "spans": [ + { + "bbox": [ + 53, + 150, + 288, + 205 + ], + "type": "text", + "content": "[55] Simon Kornblith, Jonathon Shlens, and Quoc V Le Google Brain. Do Better ImageNet Models Transfer Better? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2661-2671, 2019. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 206, + 289, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 206, + 289, + 392 + ], + "spans": [ + { + "bbox": [ + 53, + 206, + 289, + 392 + ], + "type": "text", + "content": "[56] Hugo J. Kuijf, Adrià Casamitjana, D. Louis Collins, Mahsa Dadar, Achilleas Georgiou, Mohsen Ghafoorian, Dakai Jin, April Khademi, Jesse Knight, Hongwei Li, Xavier Llado, J. Matthijs Biesbroek, Miguel Luna, Qaiser Mahmood, Richard Mckinley, Alireza Mehrtash, Sebastien Ourselin, Bo Yong Park, Hyunjin Park, Sang Hyun Park, Simon Pezold, Elodie Puybureau, Jeroen De Bresser, Leticia Rittner, Carole H. Sudre, Sergi Valverde, Veronica Vilaplana, Roland Wiest, Yongchao Xu, Ziyue Xu, Guodong Zeng, Jianguo Zhang, Guoyan Zheng, Rutger Heinen, Christopher Chen, Wiesje Van Der Flier, Frederik Barkhof, Max A. Viergever, Geert Jan Biessels, Simon Andermatt, Mariana Bento, Matt Berseth, Mikhail Belyaev, and M. Jorge Cardoso. Standardized Assessment of Automatic Segmentation of White Matter Hyperintensities and Results of the WMH Segmentation Challenge. IEEE Transactions on Medical Imaging, 38(11):2556-2568, nov 2019. 2, 4, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 393, + 289, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 393, + 289, + 547 + ], + "spans": [ + { + "bbox": [ + 53, + 393, + 289, + 547 + ], + "type": "text", + "content": "[57] Manoj Kumar, Michael J. Anderson, James W. Antony, Christopher Baldassano, Paula P. Brooks, Ming Bo Cai, Po-Hsuan Cameron Chen, Cameron T. Ellis, Gregory Henselman-Petrusek, David Huberdeau, J. Benjamin Hutchinson, Y. Peeta Li, Qihong Lu, Jeremy R. Manning, Anne C. Mennen, Samuel A. Nastase, Hugo Richard, Anna C. Schapiro, Nicolas W. Schuck, Michael Shvartsman, Narayanan Sundaram, Daniel Suo, Javier S. Turek, David Turner, Vy A. Vo, Grant Wallace, Yida Wang, Jamal A. Williams, Hejia Zhang, Xia Zhu, Mihai Capota" + }, + { + "bbox": [ + 53, + 393, + 289, + 547 + ], + "type": "inline_equation", + "content": "^{\\text{一}}" + }, + { + "bbox": [ + 53, + 393, + 289, + 547 + ], + "type": "text", + "content": ", Jonathan D. Cohen, Uri Hasson, Kai Li, Peter J. Ramadge, Nicholas B. Turk-Browne, Theodore L. Willke, and Kenneth A. Norman. BrainIAK: The Brain Imaging Analysis Kit. Aperture Neuro, 2021(4), jan 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 548, + 288, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 548, + 288, + 624 + ], + "spans": [ + { + "bbox": [ + 53, + 548, + 288, + 624 + ], + "type": "text", + "content": "[58] Sonia Laguna, Riana Schleicher, Benjamin Billot, Pamela Schaefer, Brenna Mckaig, Joshua N Goldstein, Kevin N Sheth, Matthew S Rosen, W Taylor Kimberly, and Juan Eugenio Iglesias. Super-resolution of portable low-field MRI in real scenarios: integration with denoising and domain adaptation. Medical Imaging with Deep Learning (MIDL), 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 624, + 288, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 624, + 288, + 668 + ], + "spans": [ + { + "bbox": [ + 53, + 624, + 288, + 668 + ], + "type": "text", + "content": "[59] Erik G Learned-Miller and Parvez Ahammad. Joint MRI Bias Removal Using Entropy Minimization Across Images. Advances in Neural Information Processing Systems, 17, 2004. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 669, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 669, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 53, + 669, + 288, + 713 + ], + "type": "text", + "content": "[60] Jaein Lee, Eunsong Kang, Eunjin Jeon, and Heung Il Suk. Meta-modulation Network for Domain Generalization in Multi-site fMRI Classification. Lecture Notes in Computer Science, 12905 LNCS:500-509, 2021. 2" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 72, + 547, + 715 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 312, + 72, + 547, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 72, + 547, + 138 + ], + "spans": [ + { + "bbox": [ + 312, + 72, + 547, + 138 + ], + "type": "text", + "content": "[61] Gen Li, Varun Jampani, Laura Sevilla-Lara, Deqing Sun, Jonghyun Kim, and Joongkyu Kim. Adaptive Prototype Learning and Allocation for Few-Shot Segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8334-8343, 2021. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 312, + 140, + 547, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 140, + 547, + 194 + ], + "spans": [ + { + "bbox": [ + 312, + 140, + 547, + 194 + ], + "type": "text", + "content": "[62] Xiaoxiao Li, Yufeng Gu, Nicha Dvornek, Lawrence H. Staib, Pamela Ventola, and James S. Duncan. Multi-site fMRI analysis using privacy-preserving federated learning and domain adaptation: ABIDE results. Medical Image Analysis, 65:101765, oct 2020. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 312, + 195, + 547, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 195, + 547, + 250 + ], + "spans": [ + { + "bbox": [ + 312, + 195, + 547, + 250 + ], + "type": "text", + "content": "[63] Xiang Li, Tianhan Wei, Yau Pun Chen, Yu-Wing Tai, and Chi-Keung Tang. FSS-1000: A 1000-Class Dataset for Few-Shot Segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2869–2878, 2020. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 251, + 547, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 251, + 547, + 502 + ], + "spans": [ + { + "bbox": [ + 312, + 251, + 547, + 502 + ], + "type": "text", + "content": "[64] Sook-Lei Liew, Bethany P. Lo, Miranda R. Donnelly, Artemis Zavaliangos-Petropulu, Jessica N. Jeong, Giuseppe Barisano, Alexandre Hutton, Julia P. Simon, Julia M. Juliano, Anisha Suri, Zhizhuo Wang, Aisha Abdullah, Jun Kim, Tyler Ard, Nerisa Banaj, Michael R. Borich, Lara A. Boyd, Amy Brodtmann, Cathrin M. Buetefisch, Lei Cao, Jessica M. Cassidy, Valentina Ciullo, Adriana B. Conforto, Steven C. Cramer, Rosalia Dacosta-Aguayo, Ezequiel de la Rosa, Martin Domin, Adrienne N. Dula, Wuwei Feng, Alexandre R. Franco, Fatemeh Geranmayeh, Alexandre Gramfort, Chris M. Gregory, Colleen A. Hanlon, Brenton G. Hordacre, Steven A. Kautz, Mohamed Salah Khlif, Hosung Kim, Jan S. Kirschke, Jingchun Liu, Martin Lotze, Bradley J. MacIntosh, Maria Mataró, Feroze B. Mohamed, Jan E. Nordvik, Gilsoon Park, Amy Pienta, Fabrizio Piras, Shane M. Redman, Kate P. Revill, Mauricio Reyes, Andrew D. Robertson, Na Jin Seo, Surjo R. Soekadar, Gianfranco Spalletta, Alison Sweet, Maria Telenczuk, Gregory Thielman, Lars T. Westlye, Carolee J. Winstein, George F. Wittenberg, Kristin A. Wong, and Chunshui Yu. A large, curated, open-source stroke neuroimaging dataset to improve lesion segmentation algorithms. Scientific Data, 9(1):1-12, jun 2022. 4, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 502, + 545, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 502, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 312, + 502, + 545, + 525 + ], + "type": "text", + "content": "[65] Jialin Liu, Fei Chao, and Chih-Min Lin. Task Augmentation by Rotating for Meta-Learning. feb 2020. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 525, + 547, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 525, + 547, + 580 + ], + "spans": [ + { + "bbox": [ + 312, + 525, + 547, + 580 + ], + "type": "text", + "content": "[66] Ming-Yu Liu, Xun Huang, Arun Mallya, Tero Karras, Timo Aila, Jaakko Lehtinen, and Jan Kautz. Few-Shot Unsupervised Image-to-Image Translation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10551–10560, 2019. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 581, + 547, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 581, + 547, + 635 + ], + "spans": [ + { + "bbox": [ + 312, + 581, + 547, + 635 + ], + "type": "text", + "content": "[67] Weide Liu, Chi Zhang, Guosheng Lin, and Fayao Liu. CR-Net: Cross-Reference Networks for Few-Shot Segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4165–4173, 2020. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 635, + 547, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 635, + 547, + 689 + ], + "spans": [ + { + "bbox": [ + 312, + 635, + 547, + 689 + ], + "type": "text", + "content": "[68] Xiaoxiao Liu, Marc Niethammer, Roland Kwitt, Nikhil Singh, Matt McCormick, and Stephen Aylward. Low-Rank Atlas Image Analyses in the Presence of Pathologies. IEEE Transactions on Medical Imaging, 34(12):2583-2591, dec 2015. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 691, + 547, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 691, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 691, + 547, + 715 + ], + "type": "text", + "content": "[69] Xiaofeng Liu, Fangxu Xing, Chao Yang, C. C.Jay Kuo, Georges El Fakhri, and Jonghye Woo. Symmetric-" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6227" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 72, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 72, + 72, + 287, + 127 + ], + "type": "text", + "content": "Constrained Irregular Structure Inpainting for Brain MRI Registration with Tumor Pathology. Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), 12658 LNCS:80-91, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 129, + 287, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 129, + 287, + 172 + ], + "spans": [ + { + "bbox": [ + 53, + 129, + 287, + 172 + ], + "type": "text", + "content": "[70] Michael Lustig, David L. Donoho, Juan M. Santos, and John M. Pauly. Compressed sensing MRI: A look at how CS can improve on current imaging techniques. IEEE Signal Processing Magazine, 25(2):72-82, 2008. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 173, + 287, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 173, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 53, + 173, + 287, + 228 + ], + "type": "text", + "content": "[71] Kushagra Mahajan, Monika Sharma, and Lovekesh Vig. Meta-DermDiagnosis: Few-Shot Skin Disease Identification Using Meta-Learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 730-731, 2020. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 228, + 287, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 228, + 287, + 271 + ], + "spans": [ + { + "bbox": [ + 54, + 228, + 287, + 271 + ], + "type": "text", + "content": "[72] José V. Manjón, Pierrick Coup, Antonio Buades, Vladimir Fonov, D. Louis Collins, and Montserrat Robles. Non-local MRI upsampling. Medical Image Analysis, 14(6):784-792, dec 2010. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 272, + 288, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 272, + 288, + 316 + ], + "spans": [ + { + "bbox": [ + 53, + 272, + 288, + 316 + ], + "type": "text", + "content": "[73] José V. Manjón, Pierrick Coupe, Antonio Buades, D. Louis Collins, and Montserrat Robles. MRI superresolution using self-similarity and image priors. International Journal of Biomedical Imaging, 2010, 2010. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 317, + 287, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 317, + 287, + 382 + ], + "spans": [ + { + "bbox": [ + 53, + 317, + 287, + 382 + ], + "type": "text", + "content": "[74] Daniel S Marcus, Tracy H Wang, Jamie Parker, John G Csernansky, John C Morris, and Randy L Buckner. Open Access Series of Imaging Studies (OASIS): Cross-Sectional MRI Data in Young, Middle Aged, Nondemented, and Demented Older Adults. Journal of Cognitive Neuroscience, 19:1498-1507, 2007. 4, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 384, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 384, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 53, + 384, + 287, + 635 + ], + "type": "text", + "content": "[75] Bjoern H. Menze, Andras Jakab, Stefan Bauer, Jayashree Kalpathy-Cramer, Keyvan Farahani, Justin Kirby, Yuliya Burren, Nicole Porz, Johannes Slotboom, Roland Wiest, Lerente Lanczi, Elizabeth Gerstner, Marc Andre Weber, Tal Arbel, Brian B. Avants, Nicholas Ayache, Patricia Buendia, D. Louis Collins, Nicolas Cordier, Jason J. Corso, Antonio Criminisi, Tilak Das, Hervé Delingette, Cagatay Demiralp, Christopher R. Durst, Michel Dojat, Senan Doyle, Joana Festa, Florence Forbes, Ezequiel Geremia, Ben Glocker, Polina Golland, Xiaotao Guo, Andac Hamamci, Khan M. Iftekharuddin, Raj Jena, Nigel M. John, Ender Konukoglu, Danial Lashkari, José Antônio Mariz, Raphael Meier, Sérgio Pereira, Doina Precup, Stephen J. Price, Tammy Riklin Raviv, Syed M.S. Reza, Michael Ryan, Duygu Sarikaya, Lawrence Schwartz, Hoo Chang Shin, Jamie Shotton, Carlos A. Silva, Nuno Sousa, Nagesh K. Subbanna, Gabor Szekely, Thomas J. Taylor, Owen M. Thomas, Nicholas J. Tustison, Gozde Unal, Flor Vasseur, Max Wintermark, Dong Hye Ye, Liang Zhao, Binsheng Zhao, Darko Zikic, Marcel Prastawa, Mauricio Reyes, and Koen Van Leemput. The Multimodal Brain Tumor Image Segmentation Benchmark (BRATS). IEEE Transactions on Medical Imaging, 34(10):1993-2024, oct 2015. 4, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 636, + 287, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 636, + 287, + 690 + ], + "spans": [ + { + "bbox": [ + 53, + 636, + 287, + 690 + ], + "type": "text", + "content": "[76] Fausto Milletari, Nassir Navab, and Seyed Ahmad Ahmadi. V-Net: Fully convolutional neural networks for volumetric medical image segmentation. Proceedings - 2016 4th International Conference on 3D Vision, 3DV 2016, pages 565-571, dec 2016. 3, 4, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 691, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 691, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 53, + 691, + 287, + 712 + ], + "type": "text", + "content": "[77] Samuel Müller, Noah Hollmann, Sebastian Pineda Arango, Josif Grabocka, and Frank Hutter. Transformers Can Do" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 331, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 331, + 73, + 545, + 95 + ], + "type": "text", + "content": "Bayesian Inference. In International Conference on Learning Representations, mar 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 96, + 545, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 96, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 312, + 96, + 545, + 149 + ], + "type": "text", + "content": "[78] Bao Nguyen, Adam Feldman, Sarath Bethapudi, Andrew Jennings, and Chris G. Willcocks. Unsupervised region-based anomaly detection in brain mri with adversarial image inpainting. Proceedings - International Symposium on Biomedical Imaging, 2021-April:1127-1131, apr 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 150, + 545, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 150, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 312, + 150, + 545, + 216 + ], + "type": "text", + "content": "[79] Dong Nie, Roger Trullo, Jun Lian, Caroline Petitjean, Su Ruan, Qian Wang, and Dinggang Shen. Medical image synthesis with context-aware generative adversarial networks. Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), 10435 LNCS:417-425, 2017. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 217, + 545, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 217, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 312, + 217, + 545, + 259 + ], + "type": "text", + "content": "[80] Alexander F.I. Osman and Nissren M. Tamam. Deep learning-based convolutional neural network for intramodality brain MRI synthesis. Journal of Applied Clinical Medical Physics, 23(4):e13530, apr 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 261, + 545, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 261, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 312, + 261, + 545, + 371 + ], + "type": "text", + "content": "[81] Moritz Roman Hernandez Petzsche, Ezequiel de la Rosa, Uta Hanning, Roland Wiest, Waldo Enrique Valenzuela Pinilla, Mauricio Reyes, Maria Ines Meyer, Sook-Lei Liew, Florian Kofler, Ivan Ezhov, David Robben, Alexander Hutton, Tassilo Friedrich, Teresa Zarth, Johannes Burkle, The Anh Baran, Bjoern Menze, Gabriel Broocks, Lukas Meyer, Claus Zimmer, Tobias Boeckh-Behrens, Maria Berndt, Benno Ikenberg, Benedikt Wiestler, and Jan S. Kirschke. ISLES 2022: A multi-center magnetic resonance imaging stroke lesion segmentation dataset. jun 2022. 4, 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 371, + 545, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 371, + 545, + 435 + ], + "spans": [ + { + "bbox": [ + 312, + 371, + 545, + 435 + ], + "type": "text", + "content": "[82] Kilian M Pohl, Sylvain Bouix, Motoaki Nakamura, Torsten Rohlfing, Robert W McCarley, Ron Kikinis, W Eric L Grimson, Martha E Shenton, and William M Wells. A hierarchical algorithm for mr brain image parcellation. IEEE transactions on medical imaging, 26(9):1201-1212, 2007. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 437, + 545, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 437, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 312, + 437, + 545, + 480 + ], + "type": "text", + "content": "[83] Maithra Raghu, Chiyuan Zhang, Google Brain, Jon Kleinberg, and Samy Bengio. Transfusion: Understanding Transfer Learning for Medical Imaging. Advances in Neural Information Processing Systems, 32, 2019. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 481, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 481, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 312, + 481, + 545, + 514 + ], + "type": "text", + "content": "[84] Sachin Ravi and Hugo Larochelle. Optimization as a Model for Few-Shot Learning. In International Conference on Learning Representations (ICLR), jul 2017. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 514, + 545, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 514, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 312, + 514, + 545, + 569 + ], + "type": "text", + "content": "[85] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, volume 9351, pages 234-241. Springer Verlag, 2015. 1, 3, 4, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 570, + 545, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 570, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 312, + 570, + 545, + 634 + ], + "type": "text", + "content": "[86] Edgar Schonfeld, Sayna Ebrahimi, Samarth Sinha, Trevor Darrell, and Zeynep Akata. Generalized Zero- and Few-Shot Learning via Aligned Variational Autoencoders. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8247–8255, 2019. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 635, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 635, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 312, + 635, + 545, + 679 + ], + "type": "text", + "content": "[87] F. Segonne, A. M. Dale, E. Busa, M. Glessner, D. Salat, H. K. Hahn, and B. Fischl. A hybrid approach to the skull stripping problem in MRI. NeuroImage, 22(3):1060-1075, jul 2004. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 681, + 545, + 713 + ], + "type": "text", + "content": "[88] Ozan Sener and Vladlen Koltun. Multi-Task Learning as Multi-Objective Optimization. Advances in Neural Information Processing Systems, 31, 2018. 2" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6228" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 286, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 73, + 286, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 73, + 286, + 106 + ], + "spans": [ + { + "bbox": [ + 53, + 73, + 286, + 106 + ], + "type": "text", + "content": "[89] Jun Seo, Young-Hyun Park, Sung Whan Yoon, and Jaekyun Moon. Task-Adaptive Feature Transformer with Semantic Enrichment for Few-Shot Segmentation. feb 2022. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 107, + 286, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 107, + 286, + 140 + ], + "spans": [ + { + "bbox": [ + 53, + 107, + 286, + 140 + ], + "type": "text", + "content": "[90] David W. Shattuck and Richard M. Leahy. BrainSuite: An automated cortical surface identification tool. Medical Image Analysis, 6(2):129-142, jun 2002. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 141, + 286, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 141, + 286, + 184 + ], + "spans": [ + { + "bbox": [ + 53, + 141, + 286, + 184 + ], + "type": "text", + "content": "[91] X. Shen, F. Tokoglu, X. Papademetris, and R. T. Constable. Groupwise whole-brain parcellation from resting-state fMRI data for network node identification. NeuroImage, 82:403-415, nov 2013. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 186, + 286, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 186, + 286, + 241 + ], + "spans": [ + { + "bbox": [ + 53, + 186, + 286, + 241 + ], + "type": "text", + "content": "[92] Nalini M Singh, Juan Eugenio Iglesias, Elfar Adalsteinson, Adrian V Dalca, and Polina Golland. Joint Frequency and Image Space Learning for MRI Reconstruction and Analysis. Journal of Machine Learning for Biomedical Imaging, 2022:1-28, 2022. 2, 25" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 243, + 286, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 243, + 286, + 275 + ], + "spans": [ + { + "bbox": [ + 53, + 243, + 286, + 275 + ], + "type": "text", + "content": "[93] Youssef Skandarani, Nathan Painchaud, Pierre-Marc Jodoin, and Alain Lalande. On the effectiveness of GAN generated cardiac MRIs for segmentation. may 2020. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 277, + 286, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 277, + 286, + 298 + ], + "spans": [ + { + "bbox": [ + 53, + 277, + 286, + 298 + ], + "type": "text", + "content": "[94] Stephen M. Smith. Fast robust automated brain extraction. Human Brain Mapping, 17(3):143-155, nov 2002. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 300, + 286, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 300, + 286, + 387 + ], + "spans": [ + { + "bbox": [ + 53, + 300, + 286, + 387 + ], + "type": "text", + "content": "[95] Stephen M. Smith, Mark Jenkinson, Mark W. Woolrich, Christian F. Beckmann, Timothy E.J. Behrens, Heidi Johansen-Berg, Peter R. Bannister, Marilena De Luca, Ivana Drobnjak, David E. Flitney, Rami K. Niazy, James Saunders, John Vickers, Yongyue Zhang, Nicola De Stefano, J. Michael Brady, and Paul M. Matthews. Advances in functional and structural MR image analysis and implementation as FSL. NeuroImage, 23(1), 2004. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 388, + 286, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 388, + 286, + 422 + ], + "spans": [ + { + "bbox": [ + 53, + 388, + 286, + 422 + ], + "type": "text", + "content": "[96] Jake Snell, Kevin Swersky, and Twitter Richard Zemel. Prototypical Networks for Few-shot Learning. Advances in Neural Information Processing Systems, 30, 2017. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 423, + 286, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 423, + 286, + 488 + ], + "spans": [ + { + "bbox": [ + 53, + 423, + 286, + 488 + ], + "type": "text", + "content": "[97] Haoliang Sun, Ronak Mehta, Hao H. Zhou, Zhichun Huang, Sterling C. Johnson, Vivek Prabhakaran, and Vikas Singh. DUAL-GLOW: Conditional Flow-Based Generative Model for Modality Transfer. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10611-10620, 2019. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 490, + 286, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 490, + 286, + 555 + ], + "spans": [ + { + "bbox": [ + 53, + 490, + 286, + 555 + ], + "type": "text", + "content": "[98] Nathalie Tzourio-Mazoyer, Brigitte Landeau, Dimitri Papathanassiou, Fabrice Crivello, Octave Etard, Nicolas Delcroix, Bernard Mazoyer, and Marc Joliot. Automated anatomical labeling of activations in spm using a macroscopic anatomical parcellation of the mni mri single-subject brain. Neuroimage, 15(1):273-289, 2002. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 557, + 286, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 557, + 286, + 623 + ], + "spans": [ + { + "bbox": [ + 53, + 557, + 286, + 623 + ], + "type": "text", + "content": "[99] N. Tzourio-Mazoyer, B. Landeau, D. Papathanassiou, F. Crivello, O. Etard, N. Delcroix, B. Mazoyer, and M. Joliot. Automated Anatomical Labeling of Activations in SPM Using a Macroscopic Anatomical Parcellation of the MNI MRI Single-Subject Brain. NeuroImage, 15(1):273–289, jan 2002. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 624, + 286, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 286, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 286, + 667 + ], + "type": "text", + "content": "[100] David C. Van Essen, Stephen M. Smith, Deanna M. Barch, Timothy E.J. Behrens, Essa Yacoub, and Kamil Ugurbil. The WU-Minn Human Connectome Project: An overview. NeuroImage, 80:62-79, oct 2013. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 669, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 286, + 712 + ], + "type": "text", + "content": "[101] Oriol Vinyals, Google Deepmind, Charles Blundell, Timothy Lillicrap, Koray Kavukcuoglu, and Daan Wierstra. Matching Networks for One Shot Learning. Advances in Neural Information Processing Systems, 29, 2016. 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[102] Oriol Vinyals, Google Deepmind, Charles Blundell, Timothy Lillicrap, Koray Kavukcuoglu, and Daan Wierstra. Matching Networks for One Shot Learning. Advances in Neural Information Processing Systems, 29, 2016. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 118, + 545, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 149 + ], + "type": "text", + "content": "[103] Christian Wachinger, Martin Reuter, and Tassilo Klein. Deepnat: Deep convolutional neural network for segmenting neuroanatomy. NeuroImage, 170:434-445, 2018. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 151, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 151, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 307, + 151, + 545, + 194 + ], + "type": "text", + "content": "[104] Qi Wang, Julius Steiglechner, Tobias Lindig, Benjamin Bender, Klaus Scheffler, and Gabriele Lohmann. SuperResolution for Ultra High-Field MR Images. In Medical Imaging with Deep Learning (MIDL), 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 195, + 545, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 195, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 307, + 195, + 545, + 227 + ], + "type": "text", + "content": "[105] Yaqing Wang, Quanming Yao, James T. Kwok, and Lionel M. Ni. Generalizing from a Few Examples. ACM Computing Surveys, 53(3), jun 2020. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 228, + 545, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 228, + 545, + 282 + ], + "spans": [ + { + "bbox": [ + 307, + 228, + 545, + 282 + ], + "type": "text", + "content": "[106] Mark W. Woolrich, Saad Jbabdi, Brian Patenaude, Michael Chappell, Salima Makni, Timothy Behrens, Christian Beckmann, Mark Jenkinson, and Stephen M. Smith. Bayesian analysis of neuroimaging data in FSL. NeuroImage, 45(1), 2009. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 284, + 545, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 284, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 307, + 284, + 545, + 327 + ], + "type": "text", + "content": "[107] Xiao Yang, Xu Han, Eunbyung Park, Stephen Aylward, Roland Kwitt, and Marc Niethammer. Registration of Pathological Images. Simulation and synthesis in medical imaging (Workshop), 9968:97, 2016. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 328, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 328, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 307, + 328, + 545, + 392 + ], + "type": "text", + "content": "[108] Huaxiu Yao, Long-Kai Huang, Linjun Zhang, Ying Wei, Li Tian, James Zou, Junzhou Huang, and Zhenhui Li. Improving Generalization in Meta-learning via Task Augmentation. In Proceedings of the 38th International Conference on Machine Learning, PMLR, pages 11887-11897. PMLR, jul 2021. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 394, + 545, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 394, + 545, + 458 + ], + "spans": [ + { + "bbox": [ + 307, + 394, + 545, + 458 + ], + "type": "text", + "content": "[109] Chandan Ganesh Bangalore Yogananda, Benjamin C. Wagner, Gowtham K. Murugesan, Ananth Madhuranthakam, and Joseph A. Maldjian. A deep learning pipeline for automatic skull stripping and brain segmentation. Proceedings - International Symposium on Biomedical Imaging, pages 727-731, apr 2019. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 460, + 545, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 460, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 307, + 460, + 545, + 502 + ], + "type": "text", + "content": "[110] Lin Yuan, Xue Wei, Hui Shen, Ling Li Zeng, and Dewen Hu. Multi-center brain imaging classification using a novel 3d cnn approach. IEEE Access, 6:49925-49934, sep 2018. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 504, + 545, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 504, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 307, + 504, + 545, + 568 + ], + "type": "text", + "content": "[111] Chi Zhang, Guosheng Lin, Fayao Liu, Jiushuang Guo, Qingyao Wu, and Rui Yao. Pyramid Graph Networks With Connection Attentions for Region-Based One-Shot Semantic Segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9587-9595, 2019. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 570, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 570, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 307, + 570, + 545, + 624 + ], + "type": "text", + "content": "[112] Chi Zhang, Guosheng Lin, Fayao Liu, Rui Yao, and Chunhua Shen. CANet: Class-Agnostic Segmentation Networks With Iterative Refinement and Attentive Few-Shot Learning. In Conference on Computer Vision and Pattern Recognition (CVPR), pages 5217-5226, 2019. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 625, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 625, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 625, + 545, + 689 + ], + "type": "text", + "content": "[113] Amy Zhao, Guha Balakrishnan, Frédo Durand, John V Guttag, and Adrian V Dalca. Data Augmentation Using Learned Transformations for One-Shot Medical Image Segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8543-8553, 2019. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 691, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 712 + ], + "type": "text", + "content": "[114] Guoyan Zheng, Chengwen Chu, Daniel L. Belavy, Bulat Ibragimov, Robert Korez, Tomaz Vrtovec, Hugo Hutt," + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6229" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 72, + 72, + 289, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 72, + 289, + 173 + ], + "spans": [ + { + "bbox": [ + 72, + 72, + 289, + 173 + ], + "type": "text", + "content": "Richard Everson, Judith Meakin, Isabel López Andrade, Ben Glocker, Hao Chen, Qi Dou, Pheng Ann Heng, Chunliang Wang, Daniel Forsberg, Aleš Neubert, Jurgen Fripp, Martin Urschler, Darko Stern, Maria Wimmer, Alexey A. Novikov, Hui Cheng, Gabriele Armbrecht, Dieter Felsenberg, and Shuo Li. Evaluation and comparison of 3D intervertebral disc localization and segmentation methods for 3D T2 MR data: A grand challenge. Medical Image Analysis, 35:327-344, jan 2017. 3" + } + ] + } + ], + "index": 0 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "type": "text", + "content": "6230" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Neuro-Modulated Hebbian Learning for Fully Test-Time Adaptation/28e8edf6-185c-4dae-9bfc-b17f1c2f7606_content_list.json b/2023/Neuro-Modulated Hebbian Learning for Fully Test-Time Adaptation/28e8edf6-185c-4dae-9bfc-b17f1c2f7606_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..66bf99ceac862e853cfa028d12632aa27ca3b606 --- /dev/null +++ b/2023/Neuro-Modulated Hebbian Learning for Fully Test-Time Adaptation/28e8edf6-185c-4dae-9bfc-b17f1c2f7606_content_list.json @@ -0,0 +1,1691 @@ +[ + { + "type": "text", + "text": "Neuro-Modulated Hebbian Learning for Fully Test-Time Adaptation", + "text_level": 1, + "bbox": [ + 138, + 130, + 831, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yushun Tang $^{1,2}$ , Ce Zhang $^{1}$ , Heng Xu $^{1}$ , Shuoshuo Chen $^{1}$ , Jie Cheng $^{2}$ , Luziwei Leng $^{2}$ , Qinghai Guo $^{2*}$ , Zhihai He $^{1,3*}$", + "bbox": [ + 256, + 179, + 712, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Department of Electronic and Electrical Engineering, Southern University of Science and Technology, Shenzhen, China", + "bbox": [ + 89, + 217, + 880, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Advanced Computing and Storage Laboratory, Huawei Technologies Co., Ltd., Shenzhen, China", + "bbox": [ + 163, + 234, + 805, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ Pengcheng Laboratory, Shenzhen, China", + "bbox": [ + 346, + 253, + 622, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{tangys2022, zhangc2019, xuh2022, chenss2021}@mail.sustech.edu.cn", + "bbox": [ + 197, + 272, + 769, + 286 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{chengjie8, lengluziwei, guoqinghai}@huawei.com, hezh@sustech.edu.cn", + "bbox": [ + 183, + 289, + 781, + 304 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 339, + 313, + 354 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Fully test-time adaptation aims to adapt the network model based on sequential analysis of input samples during the inference stage to address the cross-domain performance degradation problem of deep neural networks. We take inspiration from the biological plausibility learning where the neuron responses are tuned based on a local synapse-change procedure and activated by competitive lateral inhibition rules. Based on these feed-forward learning rules, we design a soft Hebbian learning process which provides an unsupervised and effective mechanism for online adaptation. We observe that the performance of this feed-forward Hebbian learning for fully test-time adaptation can be significantly improved by incorporating a feedback neuro-modulation layer. It is able to fine-tune the neuron responses based on the external feedback generated by the error back-propagation from the top inference layers. This leads to our proposed neuro-modulated Hebbian learning (NHL) method for fully test-time adaptation. With the unsupervised feed-forward soft Hebbian learning being combined with a learned neuro-modulator to capture feedback from external responses, the source model can be effectively adapted during the testing process. Experimental results on benchmark datasets demonstrate that our proposed method can significantly improve the adaptation performance of network models and outperforms existing state-of-the-art methods.", + "bbox": [ + 75, + 371, + 472, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 792, + 209, + 806 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Although deep neural networks have achieved great success in various machine learning tasks, their performance tends to degrade significantly when there is data shift [27, 55] between the training data in the source do", + "bbox": [ + 75, + 816, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "main and the testing data in the target domain [40]. To address the performance degradation problem, unsupervised domain adaptation (UDA) [16,38,50] has been proposed to fine-tune the model parameters with a large amount of unlabeled testing data in an unsupervised manner. Source-free UDA methods [33, 35, 67] aim to adapt the network model without the need to access the source-domain samples.", + "bbox": [ + 500, + 340, + 890, + 445 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "There are two major categories of source-free UDA methods. The first category needs to access the whole test dataset on the target domain to achieve their adaptation performance [35, 67]. Notice that, in many practical scenarios when we deploy the network model on client devices, the network model does not have access to the whole dataset in the target domain since collecting and constructing the test dataset on the client side is very costly. The second type of method, called fully test-time adaptation, only needs access to live streams of test samples [41, 64, 66], which is able to dynamically adapt the source model on the fly during the testing process. Existing methods for fully test-time adaptation mainly focus on constructing various loss functions to regulate the inference process and adapt the model based on error back-propagation. For example, the TENT method [66] updates the batch normalization module by minimizing an entropy loss. The TTT method [64] updates the feature extractor parameters according to a self-supervised loss on a proxy learning task. The TTT++ method [37] introduces a feature alignment strategy based on online moment matching.", + "bbox": [ + 496, + 446, + 893, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1.1. Challenges in Fully Test-Time UDA", + "text_level": 1, + "bbox": [ + 500, + 771, + 807, + 787 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We recognize that most of the domain variations, such as changes in the visual scenes and image transformations or corruptions, are early layers of features in the semantic hierarchy [66]. They can be effectively captured and modeled by lower layers of the network model. From the perspective of machine learning, early representations through the lower layer play an important role to capture the pos", + "bbox": [ + 496, + 794, + 890, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding authors.", + "bbox": [ + 94, + 887, + 230, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "3728", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "terior distribution of the underlying explanatory factors for the observed input [1]. For instance, in deep neural network models, the early layers of the network tend to respond to corners, edges, or colors. In contrast, deeper layers respond to more class-specific features [72]. In the corruption test-time adaptation scenario, the class-specific features are always the same because the testing datasets are the corruption of the training domain. However, the early layers of models can be failed due to corruption.", + "bbox": [ + 75, + 90, + 472, + 226 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Therefore, the central challenge in fully test-time UDA lies in how to learn useful early layer representations of the test samples without supervision. Motivated by this observation, we propose to explore neurobiology-inspired Hebbian learning for effective early-layer representation learning and fully test-time adaptation. It has been recognized that the learning rule of supervised end-to-end deep neural network training using back-propagation and the learning rules of the early front-end neural processing in neurobiology are unrelated [28]. The responses of neurons in biological neural networks are tuned by local pre-synaptic and post-synaptic activity, along with global variables that measure task performance, rather than the specific activity of other neurons [69].", + "bbox": [ + 75, + 227, + 472, + 439 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/814346d0284596a998a03f03241a53d6d35566a32707afcdebc88afd959575bc.jpg", + "image_caption": [ + "Figure 1. The feature map visualization after the first convolution layer obtained by different learning methods." + ], + "image_footnote": [], + "bbox": [ + 86, + 455, + 473, + 597 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1.2. Hebbian Learning", + "text_level": 1, + "bbox": [ + 76, + 665, + 256, + 683 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Hebbian learning aims to learn useful early layer representations without supervision based on local synaptic plasticity rules, which is able to generate early representations that are as good as those learned by end-to-end supervised training with back-propagation [28, 52]. Drastically different from the current error back-propagation methods which require pseudo-labels or loss functions from the top network layers, Hebbian learning is a pure feed-forward adaptation process and does not require feedback from the distant top network layers. The responses of neurons are tuned based on a local synapse-change procedure and activated by competitive lateral inhibition rules [28]. During the learning process, the strength of synapses undergoes local changes that are proportional to the activity of", + "bbox": [ + 75, + 688, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the pre-synaptic cell and dependent on the activity of the post-synaptic cell. It also introduces local lateral inhibition between neurons within a layer, where the synapses of hidden units with strong responses are pushed toward the patterns that drive them, while those with weaker responses are pushed away from these patterns.", + "bbox": [ + 496, + 90, + 890, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Existing literature has shown that early representations learned by Hebbian learning are as well as backpropagation and even more robust in testing [28, 29, 52]. Figure 1 shows the feature maps learned by different methods. The first row shows the original image. The second row shows the images in the target domain with significant image corruption. The third row shows the feature maps learned by the network model trained in the source domain for these target-domain images. The fourth row shows the feature maps learned by our Hebbian learning method. The last row (\"oracle\") shows the feature maps learned with true labels. We can see that the unsupervised Hebbian learning is able to generate feature maps which are as good as those from supervised learning.", + "bbox": [ + 496, + 186, + 892, + 398 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1.3. Our Major Idea", + "text_level": 1, + "bbox": [ + 500, + 426, + 661, + 445 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we observe that Hebbian learning, although provides a new and effective approach for unsupervised learning of early layer representation of the image, when directly applied to the network model, is not able to achieve satisfactory performance in fully test-time adaptation. First, the original hard decision for competitive learning is not suitable for fully test-time adaptation. Second, the Hebbian learning does not have an effective mechanism to consider external feedback, especially the feedback from the top network layers. We observe that, biologically, the visual processing is realized through hierarchical models considering a bottom-up early representation learning for the sensory input, and a top-down feedback mechanism based on predictive coding [15, 56].", + "bbox": [ + 496, + 455, + 890, + 667 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Motivated by this, in this work, we propose to develop a new approach, called neuro-modulated Hebbian learning (NHL), for fully test-time adaptation. We first incorporate a soft decision rule into the feed-forward Hebbian learning to improve its competitive learning. Second, we learn a neuro-modulator to capture feedback from external responses, which controls which type of feature is consolidated and further processed to minimize the predictive error. During inference, the source model is adapted by the proposed NHL rule for each mini-batch of testing samples during the inference process. Experimental results on benchmark datasets demonstrate that our proposed method can significantly improve the adaptation performance of network models and outperforms existing state-of-the-art methods.", + "bbox": [ + 496, + 674, + 893, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "3729", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1.4. Summary of Major Contributions", + "text_level": 1, + "bbox": [ + 78, + 90, + 374, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To summarize, our major contributions include: (1) we identify that the major challenge in fully test-time adaptation lies in effective unsupervised learning of early layer representations, and explore neurobiology-inspired soft Hebbian learning for effective early layer representation learning and fully test-time adaptation. (2) We develop a new neuro-modulated Hebbian learning method which combines unsupervised feed-forward Hebbian learning of early layer representation with a learned neuro-modulator to capture feedback from external responses. We analyze the optimal property of the proposed NHL algorithm based on free-energy principles [14, 15]. (3) We evaluate our proposed NHL method on benchmark datasets for fully test-time adaptation, demonstrating its significant performance improvement over existing methods.", + "bbox": [ + 76, + 114, + 472, + 340 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 356, + 218, + 372 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we review existing methods related to our work including test-time adaptation, source-free UDA, domain generalization, and unsupervised Hebbian learning.", + "bbox": [ + 76, + 382, + 468, + 429 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1. Test-time Adaptation", + "text_level": 1, + "bbox": [ + 76, + 439, + 276, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Test-time adaptation aims to online adapt the trained model while testing the input samples in the target domain. Sun et al. [64] proposed Test-time Training (TTT) by optimizing a self-supervised loss through a proxy task on the source before adapting to the target domain. The method proposed in TTT++ [37] aims to minimize the distribution shift between the training and testing feature distributions by dynamically matching the moments online. It should be noted that this method requires specific training on the source data, which is not available in fully test-time adaptation scenarios [66]. TENT [66] fine-tuned the scale and bias parameters of the batch normalization layers using an entropy minimization loss during the inference process. DUA [41] adapted the statistics of the batch normalization layer only on a tiny fraction of test data and augmented a small batch of target data to adapt the model. Choi et al. [8] proposed a shift-agnostic weight regularization and an auxiliary task for the alignment between the source and target features. Note that this method requires the source data for computing the source prototypes. The continual test-time adaptation methods [39,68] consider online TTA where target data is continually changing during inference. Instead of using parameters of the pre-trained model, Boudiaf et al. [4] only adapted the model's output by optimizing an objective function based on Laplacian adjusted maximum-likelihood estimation. Besides image classification, test-time adaptation has been successfully applied in various machine learning tasks, such as scene deblurring [7], superresolution [62], human pose estimation [34], image seg", + "bbox": [ + 76, + 463, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "mentation [21], object detection [41], etc.", + "bbox": [ + 500, + 90, + 774, + 106 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Source-free Unsupervised Domain Adaptation", + "text_level": 1, + "bbox": [ + 500, + 116, + 890, + 132 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recently, source-free UDA has emerged as an important research topic. It aims to adapt the source models to unlabeled target domains without accessing the data from the source domain. Among them, SHOT [35] proposed to learn target-specific features based on an information maximization criteria and pseudo-label prediction. 3C-GAN [33] generated target-style images by training a collaborative class conditional GAN module and using a clustering-based regularization scheme. G-SFDA [70] proposed a domain-specific attention mechanism that selectively activates different feature channels for different domains. CPGA [54] proposed to generate avatar prototypes instead of images via contrastive learning. HCL [22] proposed a solution for addressing the lack of source data by introducing both instance-level and category-level historical contrastive learning. DIPE [67] focuses on exploring the domain-invariant parameters of the model, rather than trying to learn domain-invariant representations. SFDA-DE [10] aligned domains by estimating source class-conditioned feature distribution and minimizing a contrastive adaptation loss function. Existing source-free methods are offline. They need to analyze the whole test dataset and update the model for a number of adaptation epochs.", + "bbox": [ + 496, + 138, + 893, + 487 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Domain Generalization", + "text_level": 1, + "bbox": [ + 500, + 496, + 715, + 511 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Domain generalization aims to train a model only on the source data that are generalizable to unseen target domains [74]. Domain generalization methods typically aim to learn domain-invariant representations by aligning the distributions of the source domains [43, 44, 73]. Methods based on data augmentation [53, 65] for regularizing the training process have been studied. Ensemble learning methods [11, 61] generate an ensemble of different model weights from different partitions of the training data. Recently, inference-time optimization without training the network model has been studied for domain generalization. Pandey et al. [49] used a generative model to project the target data onto the source-feature manifold with labels being preserved by solving an optimization problem during the inference stage. Iwasawa et al. [25] proposed a backpropagation-free generalization method by computing distance to a pseudo-prototype representation.", + "bbox": [ + 496, + 520, + 893, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.4. Unsupervised Hebbian Learning", + "text_level": 1, + "bbox": [ + 500, + 786, + 785, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In traditional end-to-end training, back-propagation is usually used to update the weights of deep neural networks. It has been recognized that the learning rule of supervised end-to-end deep neural network training using back-propagation is much different from the learning rules of the early front-end neural processing in neurobiology. In", + "bbox": [ + 496, + 810, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3730", + "bbox": [ + 482, + 945, + 516, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "addition, supervised training of deep neural networks with back-propagation requires a large amount of labeled samples [30]. The tuning of neuron responses in biological neural networks is achieved through a physically local synapse-change procedure. The superficial cerebral cortex exhibits a common connectivity pattern in which neurons are activated through competition via lateral inhibition [2, 12]. This competition leads to the suppression of weakly activated neurons and the amplification of strongly activated ones, a phenomenon known as competitive learning or \"winnertakes-all (WTA)\" learning [57, 59]. Motivated by Hebb's idea [19], several biological plausibility learning rules have been proposed, where changes of the synapse strength depend only on the activities of the pre-synaptic and postsynaptic neurons. The Oja's rule [47] proposed a linear neuron model with constrained Hebbian synaptic modification and derived a new unconstrained learning method. Krotov et al. [28] proposed a family of biologically plausible learning rules that enable the learning of early representations that are comparable to those achieved by end-to-end supervised training with back-propagation. Pogodin et al. [52] presents a family of learning rules which use pre- and post-synaptic firing rates and a global teaching signal. They perform almost the same as the back-propagation method.", + "bbox": [ + 76, + 90, + 472, + 454 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 465, + 171, + 479 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we present our method of neuromodulated Hebbian learning for fully test-time adaptation.", + "bbox": [ + 76, + 491, + 468, + 521 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Problem Formulation", + "text_level": 1, + "bbox": [ + 76, + 527, + 279, + 542 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Suppose that a model $q_{\\theta_s}(y|X_s)$ with parameters $\\theta_s$ has been successfully trained on the source data $\\{X_s\\}$ with labels $\\{Y_s\\}$ where the true distribution is $p_s(y|X_s)$ . It consists of a feature extractor $\\mathbf{F}$ and a classifier $\\mathbf{C}$ . During fully test-time adaptation, we are given the target data $\\{X_t\\}$ with unknown labels $\\{Y_t\\}$ . Our goal is to adapt the trained model $q_{\\theta_s}$ in an unsupervised manner during testing to approximate the true target distribution $p_t(y|X_t)$ . Given a sequence of input sample batches $\\{B_1, B_2, \\dots, B_n\\}$ , the $i$ -th adaptation of the network model can only rely on the $i$ -th batch of test samples $B_i$ .", + "bbox": [ + 76, + 551, + 468, + 717 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We consider this transfer learning on a new target domain as an active inference process. The pre-trained model $q_{\\theta_s}(y|X_s)$ is considered as a prior. Thus, predicting the labels for target samples becomes a Bayesian posterior $q_{\\theta_1}(y|B_1)$ for the first batch, given the prior model $q_{\\theta_s}(y|X_s)$ . After the first adaptation batch, the prior model is updated to $q_{\\theta_1}(y|B_1)$ . This Bayesian inference process is repeated for all subsequent batches and produces a final posterior estimation", + "bbox": [ + 76, + 718, + 468, + 853 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nq _ {\\theta , t} (y \\mid X _ {t}) = q _ {\\theta_ {n}} (y \\mid B _ {n}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 862, + 468, + 878 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "given the prior model $q_{\\theta_{n - 1}}(y|B_{n - 1})$", + "bbox": [ + 76, + 885, + 331, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/731f0308d9923657f8a57da1de826732da748b5d978c09f8a9a18d6ed1a16429.jpg", + "image_caption": [ + "Figure 2. An overview of the proposed NHL method. During inference, the first convolution layer of the source model is fine-tuned by the Hebbian learning rule and the neuromodulator is further fine-tuned by the entropy loss before making a prediction given each mini-batch testing sample. The lock symbol means the classifier is fixed in the test-time adaptation process." + ], + "image_footnote": [], + "bbox": [ + 509, + 85, + 887, + 290 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Overview of Our NHL Method", + "text_level": 1, + "bbox": [ + 500, + 393, + 774, + 409 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As illustrated in Figure 2, our proposed neuro-modulated Hebbian learning consists of two major components: the feed-forward soft Hebbian learning layer and the neuromodulator. The soft Hebbian learning layer aims to learn useful early layer representations without supervision based on local synaptic plasticity and soft competitive learning rules. It is able to generate early representations which are as good as those learned by end-to-end supervised training with labeled samples and back-propagation. During our experiments, we find that this soft Hebbian learning layer can significantly improve the performance of the network model in the target domain. However, we recognize that the feedforward Hebbian learning only is not able to achieve competitive performance as the current state-of-the-art methods for fully test-time adaptation. We find that it lacks the capability to effectively respond to external stimulus, specifically the feedback from the network output layer.", + "bbox": [ + 496, + 416, + 892, + 674 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To address this issue, we propose to design a neuromodulator layer, an intermediate layer or an interface between the soft Hebbian learning layer and the classifier module of the network. This neuro-modulator layer is updated using back-propagation with the entropy loss being evaluated at the network output. It serves as the bridge between two different learning approaches: the feed-forward Hebbian learning and the original error back-propagation. From the ablation studies summarized in Table 5, we can see that both algorithm components, the feed-forward soft Hebbian learning, and the neuro-modulator are contributing significantly to the overall performance improvement.", + "bbox": [ + 496, + 674, + 892, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The proposed NHL method can be also formulated by the free-energy principle in cognitive science [14, 15, 56]. A hierarchical predictive coding model was proposed by Rao", + "bbox": [ + 498, + 854, + 893, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "3731", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and Ballard [56] to learn a hierarchical internal model for human perception by maximizing the posterior probability of generating the observed data. This can be realized by a concurrent process of prediction through a bottom-up feedforward generation process (such as our Hebbian learning layer) combined with a top-down feedback-based optimization process (such as our neuro-modulator). Specifically, given the sensory input $x$ , assume $x$ is generated by environmental causes $\\vartheta$ , denoted as $p(x) = p(x, \\vartheta)$ . The free-energy principle states that the brain encodes the recognition density over sensory causes [14]. Mathematically, it optimizes a generative probabilistic mixture as $q_{1}(x) := q(x, \\vartheta)$ . It has been demonstrated that this process can be achieved by a Hebbian-like learning approach [42, 45].", + "bbox": [ + 75, + 90, + 472, + 301 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "On the other hand, to approximate the true distribution $p_t(y|X_t)$ by a posterior approximation $q_2(y) \\coloneqq q_{\\theta,t}(y|X_t)$ , one can consider the similarity between these two distributions measured by the following Kullback-Leibler (KL) divergence", + "bbox": [ + 76, + 301, + 468, + 378 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {K L} \\left[ q _ {2} (y) | | p _ {t} (y | X _ {t}) \\right] = \\int q _ {2} (y) \\log \\frac {q _ {2} (y)}{p _ {t} (y | X _ {t})} d y. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 388, + 468, + 421 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "According to the analysis in Bogacz [3], this minimization of KL-divergence can be converted to minimization of the free-energy $F$ defined as:", + "bbox": [ + 73, + 430, + 468, + 476 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF = \\int q _ {2} (y) \\log \\frac {q _ {2} (y)}{p _ {t} \\left(X _ {t} , y\\right)} d y. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 168, + 486, + 468, + 518 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The maximization of the free-energy $F$ can be solved by an active inference process if we consider the adaptation for each batch of test samples as a decision-making step [13]. This leads to a feedback-based optimization, just like the neuro-modulation as proposed in our NHL method. In the following sections, we explain the two major components of our NHL method in more detail.", + "bbox": [ + 75, + 527, + 468, + 633 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Feed-Forward Soft Hebbian Learning", + "text_level": 1, + "bbox": [ + 76, + 643, + 405, + 660 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As discussed above, the early feed-forward layer aims to learn useful layer representations of the input without any supervision. An approximate solution to this problem can be reduced to finding the first principal component of the input data [47], known as Oja's rule and its variations. The Oja's rule is itself a variation of the Hebbian rule, plus a normalization condition. In Oja's rule, the plasticity defined for a synaptic weight $w_{ik}$ which connects a pre-synaptic neuron $i$ with input $x_{i}$ to a post-synaptic neuron $k$ with activation $y_{k}$ is:", + "bbox": [ + 75, + 666, + 468, + 816 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta w _ {i k} = \\eta y _ {k} \\left(x _ {i} - y _ {k} w _ {i k}\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 816, + 468, + 834 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "These weights are updated based on a plasticity rule [28]. We recognize that this process lacks the ability to detect different features of the data by different hidden units in the network. To address this problem, we propose to identify a", + "bbox": [ + 75, + 839, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "good set of weights, by considering a generative model to optimize the distribution similarity to the input data.", + "bbox": [ + 498, + 90, + 890, + 121 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, we assume that the target-domain input samples $\\{X_{t}\\}$ are generated by hidden causes $\\vartheta$ with distribution $p_t(x)$ . We define an approximation to $p_t(x)$ by $q(x|\\boldsymbol{w})$ conditioned on the weights $\\boldsymbol{w}$ . Moreover, we use a mixture of exponential functions to define the probability:", + "bbox": [ + 498, + 121, + 893, + 196 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nq (x \\mid \\boldsymbol {w} _ {k}) = \\exp (\\langle \\boldsymbol {w} _ {k}, x \\rangle) / N, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 207, + 890, + 224 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\pmb{w}_k$ is the weight vector corresponds to the postsynaptic neuron $k$ , and $N$ is a normalization factor to ensure that $q$ is a probabilistic measure. In the above generative model, the objective function is given by the KL divergence $\\mathrm{KL}[p_t(x)|q(x|\\pmb{w})]$ . It can be shown that the optimal parameter vector that optimizes the KL-divergence is proportional to the mean of the input distribution [45].", + "bbox": [ + 498, + 234, + 890, + 339 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "On the other hand, (5) corresponds to neural interpretation as the activation function with normalized weights $\\pmb{w}_k / R$ , where $R$ is the norm of the weight vector $\\pmb{w}_k$ . For the network layer with $K$ neurons, we define the output of the $k$ -th neuron to be:", + "bbox": [ + 498, + 340, + 893, + 415 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ny _ {k} = e ^ {\\frac {u _ {k}}{\\tau}} / \\left(\\sum_ {i = 1} ^ {K} e ^ {\\frac {u _ {i}}{\\tau}}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 422, + 890, + 465 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $u_{k}$ is the $k$ -th neuron's weighted input, i.e. $u_{k} = \\sum_{i} w_{ik} \\cdot x_{i}$ . $\\tau$ is a temperature-scaling hyper-parameter. This leads to a new soft Hebbian plasticity rule:", + "bbox": [ + 498, + 474, + 890, + 520 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta w _ {i k} = \\eta y _ {k} \\left(R x _ {i} - u _ {k} w _ {i k}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 598, + 530, + 890, + 547 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\eta$ is the learning rate. It can be shown that, using the plasticity rule in (7) to update the weights, they can converge to the equilibrium which lies in the sphere of radius $R$ . The derivative of the weight vector norm can be expressed as:", + "bbox": [ + 498, + 556, + 890, + 631 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\frac {d \\| \\boldsymbol {w} _ {k} \\|}{d t} = 2 \\boldsymbol {w} _ {k} \\frac {d \\boldsymbol {w} _ {k}}{d t} = 2 \\boldsymbol {w} _ {k} \\eta y _ {k} (R \\boldsymbol {x} - u _ {k} \\boldsymbol {w} _ {k}) \\\\ = 2 \\eta \\boldsymbol {w} _ {k} y _ {k} \\left(R \\boldsymbol {x} - \\boldsymbol {w} _ {k} \\boldsymbol {x} \\boldsymbol {w} _ {k}\\right) = 2 \\eta u _ {k} y _ {k} \\left(R - \\| \\boldsymbol {w} _ {k} \\|\\right). \\tag {8} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 638, + 890, + 700 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With this, we see that the norm of the weight vector convergences to a sphere of radius $R$ . This is because the norm of the weights decreases if $\\| \\pmb{w}_k\\| > R$ and increases if $\\| \\pmb{w}_k\\| < R$ . More details are provided in the Supplemental Materials.", + "bbox": [ + 498, + 702, + 890, + 776 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. The Neuro-Modulation Layer", + "text_level": 1, + "bbox": [ + 500, + 786, + 764, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "From the ablation studies in Section 4, we can see that the above soft Hebbian learning alone does not automatically lead to a perfect posterior estimation for $p_t(y|X_t)$ . It cannot achieve the state-of-the-art performance for fully test-time adaptation. This is because it does not have an effective mechanism to consider external feedback, especially", + "bbox": [ + 498, + 809, + 890, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "3732", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the feedback from the top network layers. Our proposed solution is to incorporate one or more modulating layers to steer the updates of weights to the desired outcome [17,36]. This so-called neuro-modulator has been explored in neuroscience research [5,48,51,63]. Recent research [32] shows that the level of neuro-modulation may change the process of synaptic consolidation, thus ultimately controlling which type of information is consolidated in the upper neural network. Unlike the above neuromodulator-based learning, where the modulator factor is embedded inside the Hebbian rule, we consider such a neuromodulation process in a disentangled way derived from the free-energy principle. It serves as an interface between the feed-forward Hebbian layer and the top decision network.", + "bbox": [ + 75, + 90, + 472, + 301 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As defined in (3), the problem of minimizing the KL-divergence for $q_{2}(y)$ and its true posterior $p_t(y|X_t)$ can be formulated based on the free-energy principle:", + "bbox": [ + 76, + 301, + 468, + 348 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {K L} \\left[ q _ {2} (y) \\mid \\left| p _ {t} (y \\mid X _ {t}) \\right. \\right] = F + \\log P _ {t} \\left(X _ {t}\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 359, + 468, + 376 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $P_{t}(X_{t}) \\coloneqq \\int q_{2}(y)p_{t}(X_{t})dy = p_{t}(X_{t})$ is the normalization term. Note that this term does not depend on $q_{2}(y)$ . Therefore, minimizing the KL-divergence is reduced to minimizing $F$ . To this end, given a batch $B$ of data in the target domain, we rewrite (use $p_t(y,B) = p_t(B|y)p_t(y)$ ) and decompose the free-energy $F_{B}$ for current batch into the following two items:", + "bbox": [ + 75, + 387, + 468, + 494 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nF _ {B} = \\operatorname {K L} \\left[ q _ {2} (y | B) \\mid \\mid p _ {t} (y) \\right] - \\int_ {y} q _ {2} (y | B) \\log p _ {t} (B | y) d y. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 503, + 468, + 549 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The first term in (10) is already minimized through soft Hebbian learning, while minimizing the second term requires the likelihood distribution $p_t(B|y)$ . Since $p_t(B|y) = p_t(y|B)p_t(B) / p_t(y)$ and $q_2(y|B)$ is considered as an approximation of $p_t(y|B)$ , we minimize the entropy of $y$ given the current batch $B$ in a discrete way as:", + "bbox": [ + 75, + 550, + 468, + 642 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\underset {\\boldsymbol {w}} {\\arg \\min } H (y | B) = \\underset {\\boldsymbol {w}} {\\arg \\max } \\sum q _ {2} (y | B) \\log q _ {2} (y | B), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 652, + 468, + 691 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathbf{w}$ are the weights in the layer implementing the neuro-modulator. As in existing deep neural network training, we can use gradient descent to optimize these weights.", + "bbox": [ + 75, + 691, + 468, + 738 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 751, + 209, + 768 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we conduct experiments on multiple test-time adaptation benchmark datasets to evaluate the performance of our proposed NHL method.", + "bbox": [ + 75, + 776, + 468, + 821 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Benchmark Datasets", + "text_level": 1, + "bbox": [ + 76, + 832, + 272, + 847 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate our method among the following popular benchmark datasets for test-time adaptation. (1) CIFAR-10/100C. We choose CIFAR-10/100 [26] with 10/100", + "bbox": [ + 76, + 854, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "classes, a source training set of 50,000, and a target testing set CIFAR-10/100C [20] of 10,000 for small-size image experiments at an accessible scale. (2) ImageNet-C. We choose the ImageNet [58] with 1,000 classes, a source training set of 1.2 million, and a validation set of 50,000 for large-size image experiments. It should be noted that we use a fixed target testing subset of the validation set with 5000 images on which all models are evaluated following the RobustBench protocol [9]. (3) SVHN $\\rightarrow$ MNIST/MNIST-M/USPS. Following TENT for domain adaptation, we choose a SVHN [46] source model and transfer it to MNIST [31] (10,000 test samples), MNIST-M [16] (10,000 test samples) and USPS [23] (2,007 test samples), respectively.", + "bbox": [ + 496, + 90, + 890, + 303 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparison Methods", + "text_level": 1, + "bbox": [ + 500, + 316, + 702, + 333 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare our method against the following fully test-time adaptation methods: (1) Source: the baseline model is trained only on the source data without any fine-tuning during the test process. (2) TTT [64]: it adapts the feature extractor by optimizing a self-supervised loss through a proxy task. However, it requires training the same proxy task on the source domain. (3) NORM [60]: this test-time normalization method updates the batch normalization [24] statistics using the mini-batch samples during the test process. (4) TENT [66]: it fine-tunes scale and bias parameters of the batch normalization layers using an entropy minimization loss during inference. (5) DUA [41]: it adapts the statistics of the batch normalization layer only on a tiny fraction of test data and augments a small batch of target data to adapt the model.", + "bbox": [ + 496, + 340, + 890, + 568 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Implementation Details", + "text_level": 1, + "bbox": [ + 500, + 582, + 718, + 598 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Following the official implementations of TTT $^1$ , TENT $^2$ and DUA $^3$ , we use the ResNet-26 [18], Wide-ResNet-28-10 [71], and Wide-ResNet-40-2 [71] as the backbone networks for the CIFAR-10C dataset. We use the Wide-ResNet-40-2 network for the CIFAR-100C dataset. For the ImageNet-C dataset, we use the ResNet-18 [18] backbone network. We use the pre-trained model weights from the official implementations of TENT or DUA for all backbones based on the RobustBench protocol [9]. For the digit recognition transfer tasks, we use the pre-trained model weights of SVHN from the pytorch-playground [6]. For fair performance comparisons, all methods in each experimental condition share the same architecture and the same pre-trained model parameters. The batch size is set to 128. More implementation details are provided in the Supplemental Materials.", + "bbox": [ + 496, + 606, + 890, + 847 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1 TTT: https://github.com/yueatsprograms/ttt_cifar_release", + "2TENT: https://github.com/DequanWang/tent", + "$^{3}$ DUA: https://github.com/jmiemirza/DUA" + ], + "bbox": [ + 517, + 862, + 823, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "3733", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ab102445d472969611d395fc2e014cbb0353ea965f7de0c6b25860d2ca17fe16.jpg", + "table_caption": [ + "Table 1. Top-1 Classification Error (%) for each corruption in CIFAR-10C at the highest severity (Level 5). For TTT, TENT, and DUA, we use the ResNet-26 (top), WRN-28-10 (middle), and WRN-40-2 (bottom) from their official implementation. The smallest error is shown in bold." + ], + "table_footnote": [], + "table_body": "
MethodsgausshotimpuldefcsglsmtnzmsnwfrstfgbrtcntelspxjpgAvg.
Source67.763.169.955.356.642.250.131.646.339.117.174.634.257.931.749.2
TTT [64]45.641.850.021.846.123.023.929.930.025.112.223.922.647.227.231.4
NORM [60]44.643.749.129.445.226.226.925.827.923.818.334.329.337.032.532.9
TENT [66]39.438.847.919.945.023.220.628.132.124.516.126.732.430.635.530.7
DUA [41]34.932.642.218.740.224.018.423.924.020.912.327.127.226.228.726.8
Ours33.230.638.217.741.220.817.424.027.220.413.521.128.423.728.925.8
Source72.365.772.946.954.334.842.025.141.326.09.346.726.658.530.343.5
NORM [60]28.126.136.312.835.314.212.117.317.415.38.412.623.819.727.320.4
TENT [66]24.823.533.012.031.813.710.815.916.213.77.912.122.017.324.218.6
DUA [41]27.424.635.313.134.914.611.616.817.513.17.614.122.719.326.219.9
Ours23.621.430.911.031.113.010.914.215.513.08.010.321.816.722.417.6
Source28.822.926.29.520.610.69.314.215.317.57.620.914.741.314.718.3
NORM [60]18.716.422.39.122.110.59.713.013.215.47.812.016.415.117.614.6
TENT [66]15.713.218.87.918.19.08.010.410.812.46.710.014.011.414.812.1
DUA [41]15.413.417.38.018.09.17.710.810.812.16.610.913.613.014.312.1
Ours13.412.315.07.516.08.77.79.19.610.16.48.213.39.313.310.7
", + "bbox": [ + 80, + 125, + 888, + 386 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f64dc1dac946b25ae959e3572a75adf705a4daae0cb836bd571c5dfc5b4e288a.jpg", + "table_caption": [ + "Table 2. Top-1 Classification Error (%) for each corruption in CIFAR-100C at the highest severity (Level 5)." + ], + "table_footnote": [], + "table_body": "
MethodsgausshotimpuldefcsglsmtnzmsnwfrstfgbrtcntelspxjpgAvg.
Source65.760.159.132.051.033.632.441.445.251.431.655.540.359.742.446.7
NORM [60]44.744.247.432.446.432.933.039.038.445.330.236.640.637.244.239.5
TENT [66]40.339.941.829.842.331.030.034.535.239.528.033.938.433.441.436.0
DUA [41]42.240.941.030.544.832.229.938.937.243.629.539.239.035.341.237.6
Ours38.437.136.228.441.029.329.732.233.136.126.430.936.230.838.333.6
", + "bbox": [ + 81, + 419, + 888, + 513 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ae4565f114bdb57d9036ce8fd7731db7b835f4a1f9faa8ceeedcd4f647b6ece3.jpg", + "table_caption": [ + "Table 3. Top-1 Classification Error (%) for each corruption in ImageNet-C at the highest severity (Level 5)." + ], + "table_footnote": [], + "table_body": "
MethodsgausshotimpuldefcsglsmtnzmsnwfrstfgbrtcntelspxjpgAvg.
Source98.997.699.293.389.090.282.387.992.099.575.999.565.360.354.085.7
TTT [64]75.576.881.989.682.879.171.383.681.098.359.099.054.753.249.675.7
NORM [60]60.260.759.876.668.767.464.264.666.274.757.088.855.853.052.364.7
TENT [66]59.459.658.772.566.164.962.162.264.968.655.297.954.552.151.762.7
DUA [41]71.972.672.490.280.883.174.776.477.987.362.699.360.858.452.674.7
Ours56.756.756.673.365.761.062.058.663.363.953.177.554.052.051.560.4
", + "bbox": [ + 81, + 539, + 888, + 650 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Performance Results", + "text_level": 1, + "bbox": [ + 76, + 675, + 272, + 690 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The classification errors in the highest severity level corruption test datasets for test-time adaptation are reported in Tables 1, 2 and 3, with results of comparison methods directly cited from their original papers. Table 1 compares the classification error of our proposed method against recent test-time adaptation methods on the CIFAR-10C dataset. Our method performs better than other baselines with the three backbones including ResNet-26, WRN-28-10, and WRN-40-2, indicating the effectiveness of the proposed test-time adaptation method. Table 2 shows the performance comparison results on the CIFAR-100C dataset. Very encouraging results are also obtained on the large-size complicated ImageNet-C dataset, as shown in Table 3. The", + "bbox": [ + 75, + 703, + 470, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "mean adaptation error of the full test dataset on Gaussian noise of CIFAR-10C in the different adaptation stages is shown in Figure 3. We can see that our method outperforms the NORM and TENT methods after the 5-th batch test-time adaptation. This implies our method can reduce error faster given few testing samples. Results for lower severity levels of corruption are provided in Supplemental Materials.", + "bbox": [ + 496, + 676, + 890, + 782 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results for digit recognition from the SVHN to MNIST, MNIST-M, and USPS datasets are also reported in Table 4. All experiments use the same open-source source model. Note that the result for the TENT method is reproduced here since the pre-trained model for the TENT method was not provided and explained in its original paper. We can see that our method achieves the lowest average error when", + "bbox": [ + 496, + 794, + 892, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "3734", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "compared to other test-time adaptation methods. The performance improvement is quite impressive.", + "bbox": [ + 76, + 90, + 468, + 121 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/351ec2ca9b4a63b9a2ffe5f6c30b6e0cc45ea46cd7716d3920917dca7f9d4a94.jpg", + "image_caption": [ + "Figure 3. The mean adaptation error of the full test dataset on Gaussian noise of CIFAR-10C in the different adaptation stages." + ], + "image_footnote": [], + "bbox": [ + 117, + 133, + 424, + 273 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/36d9e9a7c0904e38e8ff74e7c6261842971da2c454228a918bb8f77c71d4a731.jpg", + "table_caption": [ + "Table 4. Top-1 Classification Error (\\%) for test-time adaptation on digit recognition. The asterisk (*) means the implementation with a pytorch-playground [6] pre-trained source model by us." + ], + "table_footnote": [], + "table_body": "
MethodsMNISTMNIST-MUSPSAvg.
NORM* [60]39.652.141.444.4
TENT* [66]45.856.248.350.1
Ours31.247.932.637.2
", + "bbox": [ + 93, + 358, + 454, + 429 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Further Performance Analysis", + "text_level": 1, + "bbox": [ + 76, + 441, + 346, + 458 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5.1 Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 465, + 235, + 479 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conduct the ablation study with test-time adaptation tasks on the ImageNet-C dataset to investigate the contribution of our method. Hebbian learning alone does not automatically lead to a perfect posterior estimation for $p_t(y|X_t)$ due to the lack of global information communication. From Table 5, we can see that feed-forward soft Hebbian Learning plays a significant effect and fine-tuning the feedback neuro-modulator of Block 1/2 also improves the performance. The average error is increased when expanding the neuro-modulator to more blocks. It is because optimizing more parameters with a mini-batch of testing samples is becoming more challenging.", + "bbox": [ + 75, + 489, + 467, + 671 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/31ebc6fbf6531f648fafd57d35492dd4959df2fe4135baab2f9601125096861f.jpg", + "table_caption": [ + "Table 5. Ablation study on ImageNet-C based on ResNet-18 including 4 Blocks at the highest severity (Level 5)." + ], + "table_footnote": [], + "table_body": "
MethodsAvg. Error
Source (without adaptation)85.7
BP Conv185.0
Hebbian Conv167.2
Hebbian Conv1 + Neuromodulator (Block 1)60.5
Hebbian Conv1 + Neuromodulator (Block 1/2)60.4
Hebbian Conv1 + Neuromodulator (Block 1/2/3)61.2
Hebbian Conv1 + Neuromodulator (Block 1/2/3/4)64.8
", + "bbox": [ + 83, + 705, + 464, + 823 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5.2 Feature Visualization", + "text_level": 1, + "bbox": [ + 76, + 845, + 279, + 859 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figure 4 compares the feature distributions on corrupted data obtained by different adaptation methods, including", + "bbox": [ + 76, + 869, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the source model with no adaptation, the TENT method, and our method. We also include the feature distribution for the supervised learning with the labeled sample, denoted by \"Oracle\". We can see that our method is able to learn features that are close to those obtained by supervised learning.", + "bbox": [ + 498, + 90, + 890, + 167 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4662bd51e0d1f89539a89c153c6e65f8fab52d42c3935928e4c945bfbf0aaa83.jpg", + "image_caption": [ + "(a) Source" + ], + "image_footnote": [], + "bbox": [ + 527, + 180, + 684, + 273 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/46ba2820524de6541c93edc9bc923e2a4a588acd681512486780c375117e928a.jpg", + "image_caption": [ + "(b) TENT" + ], + "image_footnote": [], + "bbox": [ + 709, + 181, + 864, + 273 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/afb646411263bd345abc7a06d45d4097e2d58a40dd72594f15e806f4404b2acb.jpg", + "image_caption": [ + "(c) Ours" + ], + "image_footnote": [], + "bbox": [ + 526, + 287, + 681, + 378 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d1de06b416b12b764755c233d5d605f385f8d97d381fea748d1301443820f55e.jpg", + "image_caption": [ + "(d) Oracle", + "Figure 4. Density plots of test-time adapted features distribution on CIFAR-10-C with Gaussian noise (front) and reference features without corruption (back with yellow color). Each horizontal axis corresponds to one channel. The height of each ridge indicates the number of features that take the same value." + ], + "image_footnote": [], + "bbox": [ + 707, + 287, + 861, + 378 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Further Discussions", + "text_level": 1, + "bbox": [ + 500, + 497, + 691, + 512 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Hebbian Learning by competition through lateral inhibition is a feed-forward process that has no gradient. If combining Hebbian Learning with back-propagation, it is limited to propagate the gradient through Hebbian layers to earlier gradient-based layers during training. Therefore, the Hebbian layers can only be designed in the early layers of models without back-propagation and gradient. In addition, although the Hebbian learning rule is commonly used for long-term reinforcement, the Hebb principle does not cover all forms of long-term synaptic plasticity.", + "bbox": [ + 496, + 527, + 890, + 678 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 705, + 617, + 719 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we take inspiration from the biological neuromodulation process to construct a novel neuro-modulated Hebbian Learning (NHL) framework. With the unsupervised feed-forward soft Hebbian learning being combined with a learned neuro-modulator to capture feedback from external responses, the source model can be effectively adapted during the testing process. Experimental results on benchmark datasets demonstrate that our proposed method can significantly reduce the testing error for image classification with corruption, and reach new state-of-the-art performance.", + "bbox": [ + 496, + 734, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "3735", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Yoshua Bengio, Aaron Courville, and Pascal Vincent. Representation learning: A review and new perspectives. IEEE TPAMI, 35(8):1798-1828, 2013. 2", + "[2] Tom Binzegger, Rodney J Douglas, and Kevan AC Martin. A quantitative map of the circuit of cat primary visual cortex. Journal of Neuroscience, 24(39):8441-8453, 2004. 4", + "[3] Rafal Bogacz. A tutorial on the free-energy framework for modelling perception and learning. Journal of Mathematical Psychology, 76:198-211, 2017. 5", + "[4] Malik Boudiaf, Romain Mueller, Ismail Ben Ayed, and Luca Bertinetto. Parameter-free online test-time adaptation. In CVPR, pages 8344-8353, 2022. 3", + "[5] K Bougrova, N Bonacchi, JA Catarino, EE Dewitt, GT Meijer, P Dayan, ZF Mainen, et al. Comparison of neuromodulator signaling in a visual decision-making task. In Annual Meeting of the Society for Neuroscience, 2022. 6", + "[6] Aaron Chen and Gus Smith. pytorch-playground. https://github.com/aaron-xichen/pytorch-playground, 2020. 6, 8", + "[7] Zhixiang Chi, Yang Wang, Yuanhao Yu, and Jin Tang. Test-time fast adaptation for dynamic scene deblurring via meta-auxiliary learning. In CVPR, pages 9137-9146, 2021. 3", + "[8] Sungha Choi, Seunghan Yang, Seokeon Choi, and Sungrack Yun. Improving test-time adaptation via shift-agnostic weight regularization and nearest source prototypes. arXiv preprint arXiv:2207.11707, 2022. 3", + "[9] Francesco Croce, Maksym Andriushchenko, Vikash Sehwag, Edoardo Debenedetti, Nicolas Flammarion, Mung Chiang, Prateek Mittal, and Matthias Hein. Robustbench: a standardized adversarial robustness benchmark. In NeurIPS, 2021. 6", + "[10] Ning Ding, Yixing Xu, Yehui Tang, Chao Xu, Yunhe Wang, and Dacheng Tao. Source-free domain adaptation via distribution estimation. In CVPR, pages 7212-7222, 2022. 3", + "[11] Zhengming Ding and Yun Fu. Deep domain generalization with structured low-rank constraint. IEEE TIP, 27(1):304-313, 2017. 3", + "[12] Rodney J Douglas, Kevan AC Martin, et al. Neuronal circuits of the neocortex. Annual Review of Neuroscience, 27(1):419-451, 2004. 4", + "[13] Karl Friston. Hierarchical models in the brain. PLoS Computational Biology, 4(11):e1000211, 2008. 5", + "[14] Karl Friston. The free-energy principle: a rough guide to the brain? Trends in Cognitive Sciences, 13(7):293-301, 2009. 3, 4, 5", + "[15] Karl Friston, James Kilner, and Lee Harrison. A free energy principle for the brain. Journal of Physiology, Paris, 100(1-3):70-87, 2006. 2, 3, 4", + "[16] Yaroslav Ganin and Victor Lempitsky. Unsupervised domain adaptation by backpropagation. In ICML, pages 1180-1189. PMLR, 2015. 1, 6", + "[17] Wulfram Gerstner, Marco Lehmann, Vasiliki Liakoni, Dane Corneil, and Johann Brea. Eligibility traces and plasticity on behavioral time scales: experimental support of neohoebbian three-factor learning rules. Frontiers in Neural Circuits, 12:53, 2018. 6" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 6", + "[19] Donald Olding Hebb. The organization of behavior: A neuropsychological theory. Psychology Press, 2005. 4", + "[20] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. In ICLR, 2018. 6", + "[21] Minhao Hu, Tao Song, Yujun Gu, Xiangde Luo, Jieneng Chen, Yinan Chen, Ya Zhang, and Shaoting Zhang. Fully test-time adaptation for image segmentation. In MICCAI, pages 251-260. Springer, 2021. 3", + "[22] Jiaxing Huang, Dayan Guan, Aoran Xiao, and Shijian Lu. Model adaptation: Historical contrastive learning for unsupervised domain adaptation without source data. In NeurIPS, volume 34, pages 3635-3649, 2021. 3", + "[23] Jonathan J. Hull. A database for handwritten text recognition research. IEEE TPAMI, 16(5):550-554, 1994. 6", + "[24] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In ICML, pages 448-456. PMLR, 2015. 6", + "[25] Yusuke Iwasawa and Yutaka Matsuo. Test-time classifier adjustment module for model-agnostic domain generalization. In NeurIPS, volume 34, pages 2427-2440, 2021. 3", + "[26] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. Master's thesis, University of Tront, 2009. 6", + "[27] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Communications of the ACM, 60(6):84-90, 2017. 1", + "[28] Dmitry Krotov and John J Hopfield. Unsupervised learning by competing hidden units. Proceedings of the National Academy of Sciences, 116(16):7723-7731, 2019. 2, 4, 5", + "[29] Gabriele Lagani, Fabrizio Falchi, Claudio Gennaro, and Giuseppe Amato. Hebbian semi-supervised learning in a sample efficiency setting. *Neural Networks*, 143:719–731, 2021. 2", + "[30] Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Deep learning. Nature, 521(7553):436-444, 2015. 4", + "[31] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 6", + "[32] Andrew B Lehr, Jannik Luboeinski, and Christian Tetzlaff. Neuromodulator-dependent synaptic tagging and capture retroactively controls neural coding in spiking neural networks. Scientific Reports, 12(1):1-18, 2022. 6", + "[33] Rui Li, Qianfen Jiao, Wenming Cao, Hau-San Wong, and Si Wu. Model adaptation: Unsupervised domain adaptation without source data. In CVPR, pages 9641-9650, 2020. 1, 3", + "[34] Yizhuo Li, Miao Hao, Zonglin Di, Nitesh Bharadwaj Gundavarapu, and Xiaolong Wang. Test-time personalization with a transformer for human pose estimation. In NeurIPS, volume 34, pages 2583-2597, 2021. 3", + "[35] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In ICML, pages 6028-6039. PMLR, 2020. 1, 3" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "3736", + "bbox": [ + 482, + 945, + 516, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[36] Timothy P Lillicrap, Adam Santoro, Luke Marris, Colin J Akerman, and Geoffrey Hinton. Backpropagation and the brain. Nature Reviews Neuroscience, 21(6):335-346, 2020. 6", + "[37] Yuejiang Liu, Parth Kothari, Bastien van Delft, Baptiste Bellot-Gurlet, Taylor Mordan, and Alexandre Alahi. Ttt++: When does self-supervised test-time training fail or thrive? In NeurIPS, volume 34, pages 21808-21820, 2021. 1, 3", + "[38] Mingsheng Long, Yue Cao, Jianmin Wang, and Michael Jordan. Learning transferable features with deep adaptation networks. In ICML, pages 97-105. PMLR, 2015. 1", + "[39] Robert A Marsden, Mario Döbler, and Bin Yang. Gradual test-time adaptation by self-training and style transfer. arXiv preprint arXiv:2208.07736, 2022. 3", + "[40] Muhammad Jehanzeb Mirza, Cornelius Buerkle, Julio Jarquin, Michael Opitz, Fabian Oboril, Kay-Ulrich Scholl, and Horst Bischof. Robustness of object detectors in degrading weather conditions. In ITSC, pages 2719-2724. IEEE, 2021. 1", + "[41] M Jehanzeb Mirza, Jakub Micorek, Horst Possegger, and Horst Bischof. The norm must go on: Dynamic unsupervised domain adaptation by normalization. In CVPR, pages 14765-14775, 2022. 1, 3, 6, 7", + "[42] Timoleon Moraitis, Dmitry Toichkin, Adrien Journé, Yansong Chua, and Qinghai Guo. Softhebb: Bayesian inference in unsupervised hebbian soft winner-take-all networks. *Neuromorphic Computing and Engineering*, 2(4):044017, 2022. 5", + "[43] Saeid Motiian, Marco Piccirilli, Donald A Adjeroh, and Gianfranco Doretto. Unified deep supervised domain adaptation and generalization. In ICCV, pages 5715-5725, 2017. 3", + "[44] Krikamol Muandet, David Balduzzi, and Bernhard Schölkopf. Domain generalization via invariant feature representation. In ICML, pages 10-18. PMLR, 2013. 3", + "[45] Bernhard Nessler, Michael Pfeiffer, Lars Buesing, and Wolfgang Maass. Bayesian computation emerges in generic cortical microcircuits through spike-timing-dependent plasticity. PLoS Computational Biology, 9(4):e1003037, 2013. 5", + "[46] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y Ng. Reading digits in natural images with unsupervised feature learning. 2011. 6", + "[47] Erkki Oja. Simplified neuron model as a principal component analyzer. Journal of Mathematical Biology, 15(3):267-273, 1982. 4, 5", + "[48] John O'Donnell, Douglas Zeppenfeld, Evan McConnell, Salvador Pena, and Maiken Nedergaard. Norepinephrine: a neuromodulator that boosts the function of multiple cell types to optimize cns performance. Neurochemical Research, 37(11):2496-2512, 2012. 6", + "[49] Prashant Pandey, Mrigank Raman, Sumanth Varambally, and Prathosh Ap. Generalization on unseen domains via inference-time label-preserving target projections. In CVPR, pages 12924-12933, 2021. 3", + "[50] Zhongyi Pei, Zhangjie Cao, Mingsheng Long, and Jianmin Wang. Multi-adversarial domain adaptation. In AAAI, volume 32, 2018. 1" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[51] Marina R Picciotto, Michael J Higley, and Yann S Mineur. Acetylcholine as a neuromodulator: cholinergic signaling shapes nervous system function and behavior. *Neuron*, 76(1):116-129, 2012. 6", + "[52] Roman Pogodin and Peter Latham. Kernelized information bottleneck leads to biologically plausible 3-factor hebbian learning in deep networks. In NeurIPS, volume 33, pages 7296-7307, 2020. 2, 4", + "[53] Fengchun Qiao, Long Zhao, and Xi Peng. Learning to learn single domain generalization. In CVPR, pages 12556-12565, 2020. 3", + "[54] Zhen Qiu, Yifan Zhang, Hongbin Lin, Shuaicheng Niu, Yanxia Liu, Qing Du, and Mingkui Tan. Source-free domain adaptation via avatar prototype generation and adaptation. In IJCAI, pages 2921-2927, 2021. 3", + "[55] Joaquin Quinonero-Candela, Masashi Sugiyama, Anton Schwaighofer, and Neil D Lawrence. Dataset shift in machine learning. MIT Press, 2008. 1", + "[56] Rajesh PN Rao and Dana H Ballard. Predictive coding in the visual cortex: a functional interpretation of some extra-classical receptive-field effects. Nature Neuroscience, 2(1):79-87, 1999. 2, 4, 5", + "[57] David E Rumelhart and David Zipser. Feature discovery by competitive learning. Cognitive Science, 9(1):75-112, 1985. 4", + "[58] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. IJCV, 115(3):211-252, 2015. 6", + "[59] Ueli Rutishauser, Rodney J Douglas, and Jean-Jacques Slotine. Collective stability of networks of winner-take-all circuits. Neural Computation, 23(3):735-773, 2011. 4", + "[60] Steffen Schneider, Evgenia Rusak, Luisa Eck, Oliver Bringmann, Wieland Brendel, and Matthias Bethge. Improving robustness against common corruptions by covariate shift adaptation. In NeurIPS, volume 33, pages 11539-11551, 2020. 6, 7, 8", + "[61] Seonguk Seo, Yumin Suh, Dongwan Kim, Geeho Kim, Jongwoo Han, and Bohyung Han. Learning to optimize domain specific normalization for domain generalization. In ECCV, pages 68-83. Springer, 2020. 3", + "[62] Assaf Shocher, Nadav Cohen, and Michal Irani. \"zero-shot\" super-resolution using deep internal learning. In CVPR, pages 3118-3126, 2018. 3", + "[63] Luisa Speranza, Umberto di Porzio, Davide Viggiano, Antonio de Donato, and Floriana Volpicelli. Dopamine: The neuromodulator of long-term synaptic plasticity, reward and movement control. Cells, 10(4):735, 2021. 6", + "[64] Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei A. Efros, and Moritz Hardt. Test-time training with self-supervision for generalization under distribution shifts. In ICML, volume 119, pages 9229-9248. PMLR, 2020. 1, 3, 6, 7", + "[65] Riccardo Volpi and Vittorio Murino. Addressing model vulnerability to distributional shifts over image transformation sets. In ICCV, pages 7980-7989, 2019. 3" + ], + "bbox": [ + 501, + 90, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "3737", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[66] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In ICLR, 2020. 1, 3, 6, 7, 8", + "[67] Fan Wang, Zhongyi Han, Yongshun Gong, and Yilong Yin. Exploring domain-invariant parameters for source free domain adaptation. In CVPR, pages 7151-7160, June 2022. 1, 3", + "[68] Qin Wang, Olga Fink, Luc Van Gool, and Dengxin Dai. Continual test-time domain adaptation. In CVPR, pages 7201-7211, 2022. 3", + "[69] James CR Whittington and Rafal Bogacz. Theories of error back-propagation in the brain. Trends in cognitive sciences, 23(3):235-250, 2019. 2", + "[70] Shiqi Yang, Yaxing Wang, Joost van de Weijer, Luis Herranz, and Shangling Jui. Generalized source-free domain adaptation. In ICCV, pages 8978-8987, October 2021. 3", + "[71] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. In BMVC. British Machine Vision Association, 2016. 6", + "[72] Matthew D Zeiler and Rob Fergus. Visualizing and understanding convolutional networks. In ECCV, pages 818-833. Springer, 2014. 2", + "[73] Shanshan Zhao, Mingming Gong, Tongliang Liu, Huan Fu, and Dacheng Tao. Domain generalization via entropy regularization. In NeurIPS, volume 33, pages 16096-16107, 2020. 3", + "[74] Kaiyang Zhou, Ziwei Liu, Yu Qiao, Tao Xiang, and Chen Change Loy. Domain generalization: A survey. IEEE TPAMI, 2022. 3" + ], + "bbox": [ + 78, + 90, + 467, + 502 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "3738", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Neuro-Modulated Hebbian Learning for Fully Test-Time Adaptation/28e8edf6-185c-4dae-9bfc-b17f1c2f7606_model.json b/2023/Neuro-Modulated Hebbian Learning for Fully Test-Time Adaptation/28e8edf6-185c-4dae-9bfc-b17f1c2f7606_model.json new file mode 100644 index 0000000000000000000000000000000000000000..9260b8d473fe01d97c53ca3425fd1618da3f0c3e --- /dev/null +++ b/2023/Neuro-Modulated Hebbian Learning for Fully Test-Time Adaptation/28e8edf6-185c-4dae-9bfc-b17f1c2f7606_model.json @@ -0,0 +1,2521 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.139, + 0.131, + 0.833, + 0.154 + ], + "angle": 0, + "content": "Neuro-Modulated Hebbian Learning for Fully Test-Time Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.257, + 0.18, + 0.714, + 0.218 + ], + "angle": 0, + "content": "Yushun Tang\\(^{1,2}\\), Ce Zhang\\(^{1}\\), Heng Xu\\(^{1}\\), Shuoshuo Chen\\(^{1}\\), Jie Cheng\\(^{2}\\), Luziwei Leng\\(^{2}\\), Qinghai Guo\\(^{2*}\\), Zhihai He\\(^{1,3*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.218, + 0.882, + 0.234 + ], + "angle": 0, + "content": "1Department of Electronic and Electrical Engineering, Southern University of Science and Technology, Shenzhen, China" + }, + { + "type": "text", + "bbox": [ + 0.164, + 0.235, + 0.806, + 0.252 + ], + "angle": 0, + "content": "\\(^{2}\\)Advanced Computing and Storage Laboratory, Huawei Technologies Co., Ltd., Shenzhen, China" + }, + { + "type": "text", + "bbox": [ + 0.348, + 0.254, + 0.624, + 0.269 + ], + "angle": 0, + "content": "\\(^{3}\\)Pengcheng Laboratory, Shenzhen, China" + }, + { + "type": "text", + "bbox": [ + 0.199, + 0.273, + 0.771, + 0.287 + ], + "angle": 0, + "content": "{tangys2022, zhangc2019, xuh2022, chenss2021}@mail.sustech.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.184, + 0.29, + 0.782, + 0.305 + ], + "angle": 0, + "content": "{chengjie8, lengluziwei, guoqinghai}@huawei.com, hezh@sustech.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.34, + 0.314, + 0.356 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.372, + 0.473, + 0.765 + ], + "angle": 0, + "content": "Fully test-time adaptation aims to adapt the network model based on sequential analysis of input samples during the inference stage to address the cross-domain performance degradation problem of deep neural networks. We take inspiration from the biological plausibility learning where the neuron responses are tuned based on a local synapse-change procedure and activated by competitive lateral inhibition rules. Based on these feed-forward learning rules, we design a soft Hebbian learning process which provides an unsupervised and effective mechanism for online adaptation. We observe that the performance of this feed-forward Hebbian learning for fully test-time adaptation can be significantly improved by incorporating a feedback neuro-modulation layer. It is able to fine-tune the neuron responses based on the external feedback generated by the error back-propagation from the top inference layers. This leads to our proposed neuro-modulated Hebbian learning (NHL) method for fully test-time adaptation. With the unsupervised feed-forward soft Hebbian learning being combined with a learned neuro-modulator to capture feedback from external responses, the source model can be effectively adapted during the testing process. Experimental results on benchmark datasets demonstrate that our proposed method can significantly improve the adaptation performance of network models and outperforms existing state-of-the-art methods." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.793, + 0.21, + 0.808 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.818, + 0.47, + 0.879 + ], + "angle": 0, + "content": "Although deep neural networks have achieved great success in various machine learning tasks, their performance tends to degrade significantly when there is data shift [27, 55] between the training data in the source do" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.341, + 0.892, + 0.446 + ], + "angle": 0, + "content": "main and the testing data in the target domain [40]. To address the performance degradation problem, unsupervised domain adaptation (UDA) [16,38,50] has been proposed to fine-tune the model parameters with a large amount of unlabeled testing data in an unsupervised manner. Source-free UDA methods [33, 35, 67] aim to adapt the network model without the need to access the source-domain samples." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.447, + 0.895, + 0.765 + ], + "angle": 0, + "content": "There are two major categories of source-free UDA methods. The first category needs to access the whole test dataset on the target domain to achieve their adaptation performance [35, 67]. Notice that, in many practical scenarios when we deploy the network model on client devices, the network model does not have access to the whole dataset in the target domain since collecting and constructing the test dataset on the client side is very costly. The second type of method, called fully test-time adaptation, only needs access to live streams of test samples [41, 64, 66], which is able to dynamically adapt the source model on the fly during the testing process. Existing methods for fully test-time adaptation mainly focus on constructing various loss functions to regulate the inference process and adapt the model based on error back-propagation. For example, the TENT method [66] updates the batch normalization module by minimizing an entropy loss. The TTT method [64] updates the feature extractor parameters according to a self-supervised loss on a proxy learning task. The TTT++ method [37] introduces a feature alignment strategy based on online moment matching." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.772, + 0.808, + 0.788 + ], + "angle": 0, + "content": "1.1. Challenges in Fully Test-Time UDA" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.903 + ], + "angle": 0, + "content": "We recognize that most of the domain variations, such as changes in the visual scenes and image transformations or corruptions, are early layers of features in the semantic hierarchy [66]. They can be effectively captured and modeled by lower layers of the network model. From the perspective of machine learning, early representations through the lower layer play an important role to capture the pos" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.888, + 0.232, + 0.901 + ], + "angle": 0, + "content": "* Corresponding authors." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3728" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.227 + ], + "angle": 0, + "content": "terior distribution of the underlying explanatory factors for the observed input [1]. For instance, in deep neural network models, the early layers of the network tend to respond to corners, edges, or colors. In contrast, deeper layers respond to more class-specific features [72]. In the corruption test-time adaptation scenario, the class-specific features are always the same because the testing datasets are the corruption of the training domain. However, the early layers of models can be failed due to corruption." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.228, + 0.473, + 0.44 + ], + "angle": 0, + "content": "Therefore, the central challenge in fully test-time UDA lies in how to learn useful early layer representations of the test samples without supervision. Motivated by this observation, we propose to explore neurobiology-inspired Hebbian learning for effective early-layer representation learning and fully test-time adaptation. It has been recognized that the learning rule of supervised end-to-end deep neural network training using back-propagation and the learning rules of the early front-end neural processing in neurobiology are unrelated [28]. The responses of neurons in biological neural networks are tuned by local pre-synaptic and post-synaptic activity, along with global variables that measure task performance, rather than the specific activity of other neurons [69]." + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.456, + 0.475, + 0.598 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.615, + 0.471, + 0.645 + ], + "angle": 0, + "content": "Figure 1. The feature map visualization after the first convolution layer obtained by different learning methods." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.666, + 0.258, + 0.684 + ], + "angle": 0, + "content": "1.2. Hebbian Learning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Hebbian learning aims to learn useful early layer representations without supervision based on local synaptic plasticity rules, which is able to generate early representations that are as good as those learned by end-to-end supervised training with back-propagation [28, 52]. Drastically different from the current error back-propagation methods which require pseudo-labels or loss functions from the top network layers, Hebbian learning is a pure feed-forward adaptation process and does not require feedback from the distant top network layers. The responses of neurons are tuned based on a local synapse-change procedure and activated by competitive lateral inhibition rules [28]. During the learning process, the strength of synapses undergoes local changes that are proportional to the activity of" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.183 + ], + "angle": 0, + "content": "the pre-synaptic cell and dependent on the activity of the post-synaptic cell. It also introduces local lateral inhibition between neurons within a layer, where the synapses of hidden units with strong responses are pushed toward the patterns that drive them, while those with weaker responses are pushed away from these patterns." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.188, + 0.893, + 0.4 + ], + "angle": 0, + "content": "Existing literature has shown that early representations learned by Hebbian learning are as well as backpropagation and even more robust in testing [28, 29, 52]. Figure 1 shows the feature maps learned by different methods. The first row shows the original image. The second row shows the images in the target domain with significant image corruption. The third row shows the feature maps learned by the network model trained in the source domain for these target-domain images. The fourth row shows the feature maps learned by our Hebbian learning method. The last row (\"oracle\") shows the feature maps learned with true labels. We can see that the unsupervised Hebbian learning is able to generate feature maps which are as good as those from supervised learning." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.427, + 0.662, + 0.446 + ], + "angle": 0, + "content": "1.3. Our Major Idea" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.457, + 0.892, + 0.669 + ], + "angle": 0, + "content": "In this work, we observe that Hebbian learning, although provides a new and effective approach for unsupervised learning of early layer representation of the image, when directly applied to the network model, is not able to achieve satisfactory performance in fully test-time adaptation. First, the original hard decision for competitive learning is not suitable for fully test-time adaptation. Second, the Hebbian learning does not have an effective mechanism to consider external feedback, especially the feedback from the top network layers. We observe that, biologically, the visual processing is realized through hierarchical models considering a bottom-up early representation learning for the sensory input, and a top-down feedback mechanism based on predictive coding [15, 56]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Motivated by this, in this work, we propose to develop a new approach, called neuro-modulated Hebbian learning (NHL), for fully test-time adaptation. We first incorporate a soft decision rule into the feed-forward Hebbian learning to improve its competitive learning. Second, we learn a neuro-modulator to capture feedback from external responses, which controls which type of feature is consolidated and further processed to minimize the predictive error. During inference, the source model is adapted by the proposed NHL rule for each mini-batch of testing samples during the inference process. Experimental results on benchmark datasets demonstrate that our proposed method can significantly improve the adaptation performance of network models and outperforms existing state-of-the-art methods." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3729" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.091, + 0.375, + 0.108 + ], + "angle": 0, + "content": "1.4. Summary of Major Contributions" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.115, + 0.473, + 0.342 + ], + "angle": 0, + "content": "To summarize, our major contributions include: (1) we identify that the major challenge in fully test-time adaptation lies in effective unsupervised learning of early layer representations, and explore neurobiology-inspired soft Hebbian learning for effective early layer representation learning and fully test-time adaptation. (2) We develop a new neuro-modulated Hebbian learning method which combines unsupervised feed-forward Hebbian learning of early layer representation with a learned neuro-modulator to capture feedback from external responses. We analyze the optimal property of the proposed NHL algorithm based on free-energy principles [14, 15]. (3) We evaluate our proposed NHL method on benchmark datasets for fully test-time adaptation, demonstrating its significant performance improvement over existing methods." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.357, + 0.22, + 0.373 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.383, + 0.47, + 0.43 + ], + "angle": 0, + "content": "In this section, we review existing methods related to our work including test-time adaptation, source-free UDA, domain generalization, and unsupervised Hebbian learning." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.44, + 0.277, + 0.456 + ], + "angle": 0, + "content": "2.1. Test-time Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.464, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Test-time adaptation aims to online adapt the trained model while testing the input samples in the target domain. Sun et al. [64] proposed Test-time Training (TTT) by optimizing a self-supervised loss through a proxy task on the source before adapting to the target domain. The method proposed in TTT++ [37] aims to minimize the distribution shift between the training and testing feature distributions by dynamically matching the moments online. It should be noted that this method requires specific training on the source data, which is not available in fully test-time adaptation scenarios [66]. TENT [66] fine-tuned the scale and bias parameters of the batch normalization layers using an entropy minimization loss during the inference process. DUA [41] adapted the statistics of the batch normalization layer only on a tiny fraction of test data and augmented a small batch of target data to adapt the model. Choi et al. [8] proposed a shift-agnostic weight regularization and an auxiliary task for the alignment between the source and target features. Note that this method requires the source data for computing the source prototypes. The continual test-time adaptation methods [39,68] consider online TTA where target data is continually changing during inference. Instead of using parameters of the pre-trained model, Boudiaf et al. [4] only adapted the model's output by optimizing an objective function based on Laplacian adjusted maximum-likelihood estimation. Besides image classification, test-time adaptation has been successfully applied in various machine learning tasks, such as scene deblurring [7], superresolution [62], human pose estimation [34], image seg" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.776, + 0.107 + ], + "angle": 0, + "content": "mentation [21], object detection [41], etc." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.117, + 0.891, + 0.133 + ], + "angle": 0, + "content": "2.2. Source-free Unsupervised Domain Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.14, + 0.895, + 0.488 + ], + "angle": 0, + "content": "Recently, source-free UDA has emerged as an important research topic. It aims to adapt the source models to unlabeled target domains without accessing the data from the source domain. Among them, SHOT [35] proposed to learn target-specific features based on an information maximization criteria and pseudo-label prediction. 3C-GAN [33] generated target-style images by training a collaborative class conditional GAN module and using a clustering-based regularization scheme. G-SFDA [70] proposed a domain-specific attention mechanism that selectively activates different feature channels for different domains. CPGA [54] proposed to generate avatar prototypes instead of images via contrastive learning. HCL [22] proposed a solution for addressing the lack of source data by introducing both instance-level and category-level historical contrastive learning. DIPE [67] focuses on exploring the domain-invariant parameters of the model, rather than trying to learn domain-invariant representations. SFDA-DE [10] aligned domains by estimating source class-conditioned feature distribution and minimizing a contrastive adaptation loss function. Existing source-free methods are offline. They need to analyze the whole test dataset and update the model for a number of adaptation epochs." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.497, + 0.717, + 0.512 + ], + "angle": 0, + "content": "2.3. Domain Generalization" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.521, + 0.895, + 0.779 + ], + "angle": 0, + "content": "Domain generalization aims to train a model only on the source data that are generalizable to unseen target domains [74]. Domain generalization methods typically aim to learn domain-invariant representations by aligning the distributions of the source domains [43, 44, 73]. Methods based on data augmentation [53, 65] for regularizing the training process have been studied. Ensemble learning methods [11, 61] generate an ensemble of different model weights from different partitions of the training data. Recently, inference-time optimization without training the network model has been studied for domain generalization. Pandey et al. [49] used a generative model to project the target data onto the source-feature manifold with labels being preserved by solving an optimization problem during the inference stage. Iwasawa et al. [25] proposed a backpropagation-free generalization method by computing distance to a pseudo-prototype representation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.787, + 0.787, + 0.803 + ], + "angle": 0, + "content": "2.4. Unsupervised Hebbian Learning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.895, + 0.903 + ], + "angle": 0, + "content": "In traditional end-to-end training, back-propagation is usually used to update the weights of deep neural networks. It has been recognized that the learning rule of supervised end-to-end deep neural network training using back-propagation is much different from the learning rules of the early front-end neural processing in neurobiology. In" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.517, + 0.957 + ], + "angle": 0, + "content": "3730" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.455 + ], + "angle": 0, + "content": "addition, supervised training of deep neural networks with back-propagation requires a large amount of labeled samples [30]. The tuning of neuron responses in biological neural networks is achieved through a physically local synapse-change procedure. The superficial cerebral cortex exhibits a common connectivity pattern in which neurons are activated through competition via lateral inhibition [2, 12]. This competition leads to the suppression of weakly activated neurons and the amplification of strongly activated ones, a phenomenon known as competitive learning or \"winnertakes-all (WTA)\" learning [57, 59]. Motivated by Hebb's idea [19], several biological plausibility learning rules have been proposed, where changes of the synapse strength depend only on the activities of the pre-synaptic and postsynaptic neurons. The Oja's rule [47] proposed a linear neuron model with constrained Hebbian synaptic modification and derived a new unconstrained learning method. Krotov et al. [28] proposed a family of biologically plausible learning rules that enable the learning of early representations that are comparable to those achieved by end-to-end supervised training with back-propagation. Pogodin et al. [52] presents a family of learning rules which use pre- and post-synaptic firing rates and a global teaching signal. They perform almost the same as the back-propagation method." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.466, + 0.172, + 0.481 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.492, + 0.47, + 0.522 + ], + "angle": 0, + "content": "In this section, we present our method of neuromodulated Hebbian learning for fully test-time adaptation." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.529, + 0.281, + 0.543 + ], + "angle": 0, + "content": "3.1. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.552, + 0.47, + 0.718 + ], + "angle": 0, + "content": "Suppose that a model \\( q_{\\theta_s}(y|X_s) \\) with parameters \\( \\theta_s \\) has been successfully trained on the source data \\( \\{X_s\\} \\) with labels \\( \\{Y_s\\} \\) where the true distribution is \\( p_s(y|X_s) \\). It consists of a feature extractor \\( \\mathbf{F} \\) and a classifier \\( \\mathbf{C} \\). During fully test-time adaptation, we are given the target data \\( \\{X_t\\} \\) with unknown labels \\( \\{Y_t\\} \\). Our goal is to adapt the trained model \\( q_{\\theta_s} \\) in an unsupervised manner during testing to approximate the true target distribution \\( p_t(y|X_t) \\). Given a sequence of input sample batches \\( \\{B_1, B_2, \\dots, B_n\\} \\), the \\( i \\)-th adaptation of the network model can only rely on the \\( i \\)-th batch of test samples \\( B_i \\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.719, + 0.47, + 0.854 + ], + "angle": 0, + "content": "We consider this transfer learning on a new target domain as an active inference process. The pre-trained model \\( q_{\\theta_s}(y|X_s) \\) is considered as a prior. Thus, predicting the labels for target samples becomes a Bayesian posterior \\( q_{\\theta_1}(y|B_1) \\) for the first batch, given the prior model \\( q_{\\theta_s}(y|X_s) \\). After the first adaptation batch, the prior model is updated to \\( q_{\\theta_1}(y|B_1) \\). This Bayesian inference process is repeated for all subsequent batches and produces a final posterior estimation" + }, + { + "type": "equation", + "bbox": [ + 0.188, + 0.863, + 0.469, + 0.88 + ], + "angle": 0, + "content": "\\[\nq _ {\\theta , t} (y \\mid X _ {t}) = q _ {\\theta_ {n}} (y \\mid B _ {n}), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.886, + 0.332, + 0.902 + ], + "angle": 0, + "content": "given the prior model \\(q_{\\theta_{n - 1}}(y|B_{n - 1})\\)" + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.087, + 0.888, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.301, + 0.892, + 0.385 + ], + "angle": 0, + "content": "Figure 2. An overview of the proposed NHL method. During inference, the first convolution layer of the source model is fine-tuned by the Hebbian learning rule and the neuromodulator is further fine-tuned by the entropy loss before making a prediction given each mini-batch testing sample. The lock symbol means the classifier is fixed in the test-time adaptation process." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.395, + 0.776, + 0.41 + ], + "angle": 0, + "content": "3.2. Overview of Our NHL Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.417, + 0.893, + 0.675 + ], + "angle": 0, + "content": "As illustrated in Figure 2, our proposed neuro-modulated Hebbian learning consists of two major components: the feed-forward soft Hebbian learning layer and the neuromodulator. The soft Hebbian learning layer aims to learn useful early layer representations without supervision based on local synaptic plasticity and soft competitive learning rules. It is able to generate early representations which are as good as those learned by end-to-end supervised training with labeled samples and back-propagation. During our experiments, we find that this soft Hebbian learning layer can significantly improve the performance of the network model in the target domain. However, we recognize that the feedforward Hebbian learning only is not able to achieve competitive performance as the current state-of-the-art methods for fully test-time adaptation. We find that it lacks the capability to effectively respond to external stimulus, specifically the feedback from the network output layer." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.893, + 0.856 + ], + "angle": 0, + "content": "To address this issue, we propose to design a neuromodulator layer, an intermediate layer or an interface between the soft Hebbian learning layer and the classifier module of the network. This neuro-modulator layer is updated using back-propagation with the entropy loss being evaluated at the network output. It serves as the bridge between two different learning approaches: the feed-forward Hebbian learning and the original error back-propagation. From the ablation studies summarized in Table 5, we can see that both algorithm components, the feed-forward soft Hebbian learning, and the neuro-modulator are contributing significantly to the overall performance improvement." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.894, + 0.901 + ], + "angle": 0, + "content": "The proposed NHL method can be also formulated by the free-energy principle in cognitive science [14, 15, 56]. A hierarchical predictive coding model was proposed by Rao" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "3731" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.303 + ], + "angle": 0, + "content": "and Ballard [56] to learn a hierarchical internal model for human perception by maximizing the posterior probability of generating the observed data. This can be realized by a concurrent process of prediction through a bottom-up feedforward generation process (such as our Hebbian learning layer) combined with a top-down feedback-based optimization process (such as our neuro-modulator). Specifically, given the sensory input \\( x \\), assume \\( x \\) is generated by environmental causes \\( \\vartheta \\), denoted as \\( p(x) = p(x, \\vartheta) \\). The free-energy principle states that the brain encodes the recognition density over sensory causes [14]. Mathematically, it optimizes a generative probabilistic mixture as \\( q_{1}(x) := q(x, \\vartheta) \\). It has been demonstrated that this process can be achieved by a Hebbian-like learning approach [42, 45]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.303, + 0.47, + 0.379 + ], + "angle": 0, + "content": "On the other hand, to approximate the true distribution \\( p_t(y|X_t) \\) by a posterior approximation \\( q_2(y) \\coloneqq q_{\\theta,t}(y|X_t) \\), one can consider the similarity between these two distributions measured by the following Kullback-Leibler (KL) divergence" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.389, + 0.47, + 0.422 + ], + "angle": 0, + "content": "\\[\n\\operatorname {K L} \\left[ q _ {2} (y) | | p _ {t} (y | X _ {t}) \\right] = \\int q _ {2} (y) \\log \\frac {q _ {2} (y)}{p _ {t} (y | X _ {t})} d y. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.431, + 0.47, + 0.477 + ], + "angle": 0, + "content": "According to the analysis in Bogacz [3], this minimization of KL-divergence can be converted to minimization of the free-energy \\(F\\) defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.169, + 0.487, + 0.47, + 0.52 + ], + "angle": 0, + "content": "\\[\nF = \\int q _ {2} (y) \\log \\frac {q _ {2} (y)}{p _ {t} \\left(X _ {t} , y\\right)} d y. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.529, + 0.47, + 0.635 + ], + "angle": 0, + "content": "The maximization of the free-energy \\( F \\) can be solved by an active inference process if we consider the adaptation for each batch of test samples as a decision-making step [13]. This leads to a feedback-based optimization, just like the neuro-modulation as proposed in our NHL method. In the following sections, we explain the two major components of our NHL method in more detail." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.644, + 0.406, + 0.661 + ], + "angle": 0, + "content": "3.3. Feed-Forward Soft Hebbian Learning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.667, + 0.47, + 0.817 + ], + "angle": 0, + "content": "As discussed above, the early feed-forward layer aims to learn useful layer representations of the input without any supervision. An approximate solution to this problem can be reduced to finding the first principal component of the input data [47], known as Oja's rule and its variations. The Oja's rule is itself a variation of the Hebbian rule, plus a normalization condition. In Oja's rule, the plasticity defined for a synaptic weight \\( w_{ik} \\) which connects a pre-synaptic neuron \\( i \\) with input \\( x_{i} \\) to a post-synaptic neuron \\( k \\) with activation \\( y_{k} \\) is:" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.818, + 0.469, + 0.835 + ], + "angle": 0, + "content": "\\[\n\\Delta w _ {i k} = \\eta y _ {k} \\left(x _ {i} - y _ {k} w _ {i k}\\right). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.902 + ], + "angle": 0, + "content": "These weights are updated based on a plasticity rule [28]. We recognize that this process lacks the ability to detect different features of the data by different hidden units in the network. To address this problem, we propose to identify a" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "good set of weights, by considering a generative model to optimize the distribution similarity to the input data." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.122, + 0.894, + 0.198 + ], + "angle": 0, + "content": "Specifically, we assume that the target-domain input samples \\(\\{X_{t}\\}\\) are generated by hidden causes \\(\\vartheta\\) with distribution \\(p_t(x)\\). We define an approximation to \\(p_t(x)\\) by \\(q(x|\\boldsymbol{w})\\) conditioned on the weights \\(\\boldsymbol{w}\\). Moreover, we use a mixture of exponential functions to define the probability:" + }, + { + "type": "equation", + "bbox": [ + 0.586, + 0.208, + 0.892, + 0.226 + ], + "angle": 0, + "content": "\\[\nq (x \\mid \\boldsymbol {w} _ {k}) = \\exp (\\langle \\boldsymbol {w} _ {k}, x \\rangle) / N, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.235, + 0.892, + 0.34 + ], + "angle": 0, + "content": "where \\(\\pmb{w}_k\\) is the weight vector corresponds to the postsynaptic neuron \\(k\\), and \\(N\\) is a normalization factor to ensure that \\(q\\) is a probabilistic measure. In the above generative model, the objective function is given by the KL divergence \\(\\mathrm{KL}[p_t(x)|q(x|\\pmb{w})]\\). It can be shown that the optimal parameter vector that optimizes the KL-divergence is proportional to the mean of the input distribution [45]." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.341, + 0.894, + 0.416 + ], + "angle": 0, + "content": "On the other hand, (5) corresponds to neural interpretation as the activation function with normalized weights \\( \\pmb{w}_k / R \\), where \\( R \\) is the norm of the weight vector \\( \\pmb{w}_k \\). For the network layer with \\( K \\) neurons, we define the output of the \\( k \\)-th neuron to be:" + }, + { + "type": "equation", + "bbox": [ + 0.614, + 0.424, + 0.892, + 0.466 + ], + "angle": 0, + "content": "\\[\ny _ {k} = e ^ {\\frac {u _ {k}}{\\tau}} / \\left(\\sum_ {i = 1} ^ {K} e ^ {\\frac {u _ {i}}{\\tau}}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.475, + 0.892, + 0.521 + ], + "angle": 0, + "content": "where \\( u_{k} \\) is the \\( k \\)-th neuron's weighted input, i.e. \\( u_{k} = \\sum_{i} w_{ik} \\cdot x_{i} \\). \\( \\tau \\) is a temperature-scaling hyper-parameter. This leads to a new soft Hebbian plasticity rule:" + }, + { + "type": "equation", + "bbox": [ + 0.599, + 0.531, + 0.892, + 0.548 + ], + "angle": 0, + "content": "\\[\n\\Delta w _ {i k} = \\eta y _ {k} \\left(R x _ {i} - u _ {k} w _ {i k}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.558, + 0.892, + 0.632 + ], + "angle": 0, + "content": "where \\(\\eta\\) is the learning rate. It can be shown that, using the plasticity rule in (7) to update the weights, they can converge to the equilibrium which lies in the sphere of radius \\(R\\). The derivative of the weight vector norm can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.518, + 0.64, + 0.891, + 0.702 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\frac {d \\| \\boldsymbol {w} _ {k} \\|}{d t} = 2 \\boldsymbol {w} _ {k} \\frac {d \\boldsymbol {w} _ {k}}{d t} = 2 \\boldsymbol {w} _ {k} \\eta y _ {k} (R \\boldsymbol {x} - u _ {k} \\boldsymbol {w} _ {k}) \\\\ = 2 \\eta \\boldsymbol {w} _ {k} y _ {k} \\left(R \\boldsymbol {x} - \\boldsymbol {w} _ {k} \\boldsymbol {x} \\boldsymbol {w} _ {k}\\right) = 2 \\eta u _ {k} y _ {k} \\left(R - \\| \\boldsymbol {w} _ {k} \\|\\right). \\tag {8} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.703, + 0.892, + 0.777 + ], + "angle": 0, + "content": "With this, we see that the norm of the weight vector convergences to a sphere of radius \\( R \\). This is because the norm of the weights decreases if \\( \\| \\pmb{w}_k\\| > R \\) and increases if \\( \\| \\pmb{w}_k\\| < R \\). More details are provided in the Supplemental Materials." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.787, + 0.766, + 0.803 + ], + "angle": 0, + "content": "3.4. The Neuro-Modulation Layer" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.81, + 0.892, + 0.902 + ], + "angle": 0, + "content": "From the ablation studies in Section 4, we can see that the above soft Hebbian learning alone does not automatically lead to a perfect posterior estimation for \\( p_t(y|X_t) \\). It cannot achieve the state-of-the-art performance for fully test-time adaptation. This is because it does not have an effective mechanism to consider external feedback, especially" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3732" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.302 + ], + "angle": 0, + "content": "the feedback from the top network layers. Our proposed solution is to incorporate one or more modulating layers to steer the updates of weights to the desired outcome [17,36]. This so-called neuro-modulator has been explored in neuroscience research [5,48,51,63]. Recent research [32] shows that the level of neuro-modulation may change the process of synaptic consolidation, thus ultimately controlling which type of information is consolidated in the upper neural network. Unlike the above neuromodulator-based learning, where the modulator factor is embedded inside the Hebbian rule, we consider such a neuromodulation process in a disentangled way derived from the free-energy principle. It serves as an interface between the feed-forward Hebbian layer and the top decision network." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.303, + 0.47, + 0.349 + ], + "angle": 0, + "content": "As defined in (3), the problem of minimizing the KL-divergence for \\( q_{2}(y) \\) and its true posterior \\( p_t(y|X_t) \\) can be formulated based on the free-energy principle:" + }, + { + "type": "equation", + "bbox": [ + 0.137, + 0.36, + 0.47, + 0.377 + ], + "angle": 0, + "content": "\\[\n\\operatorname {K L} \\left[ q _ {2} (y) \\mid \\left| p _ {t} (y \\mid X _ {t}) \\right. \\right] = F + \\log P _ {t} \\left(X _ {t}\\right), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.388, + 0.47, + 0.495 + ], + "angle": 0, + "content": "where \\( P_{t}(X_{t}) \\coloneqq \\int q_{2}(y)p_{t}(X_{t})dy = p_{t}(X_{t}) \\) is the normalization term. Note that this term does not depend on \\( q_{2}(y) \\). Therefore, minimizing the KL-divergence is reduced to minimizing \\( F \\). To this end, given a batch \\( B \\) of data in the target domain, we rewrite (use \\( p_t(y,B) = p_t(B|y)p_t(y) \\)) and decompose the free-energy \\( F_{B} \\) for current batch into the following two items:" + }, + { + "type": "equation", + "bbox": [ + 0.086, + 0.505, + 0.469, + 0.55 + ], + "angle": 0, + "content": "\\[\nF _ {B} = \\operatorname {K L} \\left[ q _ {2} (y | B) \\mid \\mid p _ {t} (y) \\right] - \\int_ {y} q _ {2} (y | B) \\log p _ {t} (B | y) d y. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.551, + 0.47, + 0.643 + ], + "angle": 0, + "content": "The first term in (10) is already minimized through soft Hebbian learning, while minimizing the second term requires the likelihood distribution \\( p_t(B|y) \\). Since \\( p_t(B|y) = p_t(y|B)p_t(B) / p_t(y) \\) and \\( q_2(y|B) \\) is considered as an approximation of \\( p_t(y|B) \\), we minimize the entropy of \\( y \\) given the current batch \\( B \\) in a discrete way as:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.654, + 0.469, + 0.692 + ], + "angle": 0, + "content": "\\[\n\\underset {\\boldsymbol {w}} {\\arg \\min } H (y | B) = \\underset {\\boldsymbol {w}} {\\arg \\max } \\sum q _ {2} (y | B) \\log q _ {2} (y | B), \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.693, + 0.47, + 0.739 + ], + "angle": 0, + "content": "where \\( \\mathbf{w} \\) are the weights in the layer implementing the neuro-modulator. As in existing deep neural network training, we can use gradient descent to optimize these weights." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.752, + 0.21, + 0.769 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.777, + 0.469, + 0.823 + ], + "angle": 0, + "content": "In this section, we conduct experiments on multiple test-time adaptation benchmark datasets to evaluate the performance of our proposed NHL method." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.833, + 0.274, + 0.848 + ], + "angle": 0, + "content": "4.1. Benchmark Datasets" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.47, + 0.901 + ], + "angle": 0, + "content": "We evaluate our method among the following popular benchmark datasets for test-time adaptation. (1) CIFAR-10/100C. We choose CIFAR-10/100 [26] with 10/100" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.304 + ], + "angle": 0, + "content": "classes, a source training set of 50,000, and a target testing set CIFAR-10/100C [20] of 10,000 for small-size image experiments at an accessible scale. (2) ImageNet-C. We choose the ImageNet [58] with 1,000 classes, a source training set of 1.2 million, and a validation set of 50,000 for large-size image experiments. It should be noted that we use a fixed target testing subset of the validation set with 5000 images on which all models are evaluated following the RobustBench protocol [9]. (3) SVHN \\(\\rightarrow\\) MNIST/MNIST-M/USPS. Following TENT for domain adaptation, we choose a SVHN [46] source model and transfer it to MNIST [31] (10,000 test samples), MNIST-M [16] (10,000 test samples) and USPS [23] (2,007 test samples), respectively." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.318, + 0.704, + 0.334 + ], + "angle": 0, + "content": "4.2. Comparison Methods" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.342, + 0.892, + 0.569 + ], + "angle": 0, + "content": "We compare our method against the following fully test-time adaptation methods: (1) Source: the baseline model is trained only on the source data without any fine-tuning during the test process. (2) TTT [64]: it adapts the feature extractor by optimizing a self-supervised loss through a proxy task. However, it requires training the same proxy task on the source domain. (3) NORM [60]: this test-time normalization method updates the batch normalization [24] statistics using the mini-batch samples during the test process. (4) TENT [66]: it fine-tunes scale and bias parameters of the batch normalization layers using an entropy minimization loss during inference. (5) DUA [41]: it adapts the statistics of the batch normalization layer only on a tiny fraction of test data and augments a small batch of target data to adapt the model." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.583, + 0.719, + 0.599 + ], + "angle": 0, + "content": "4.3. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.607, + 0.892, + 0.848 + ], + "angle": 0, + "content": "Following the official implementations of TTT\\(^1\\), TENT\\(^2\\) and DUA\\(^3\\), we use the ResNet-26 [18], Wide-ResNet-28-10 [71], and Wide-ResNet-40-2 [71] as the backbone networks for the CIFAR-10C dataset. We use the Wide-ResNet-40-2 network for the CIFAR-100C dataset. For the ImageNet-C dataset, we use the ResNet-18 [18] backbone network. We use the pre-trained model weights from the official implementations of TENT or DUA for all backbones based on the RobustBench protocol [9]. For the digit recognition transfer tasks, we use the pre-trained model weights of SVHN from the pytorch-playground [6]. For fair performance comparisons, all methods in each experimental condition share the same architecture and the same pre-trained model parameters. The batch size is set to 128. More implementation details are provided in the Supplemental Materials." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.863, + 0.824, + 0.876 + ], + "angle": 0, + "content": "1 TTT: https://github.com/yueatsprograms/ttt_cifar_release" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.876, + 0.76, + 0.888 + ], + "angle": 0, + "content": "2TENT: https://github.com/DequanWang/tent" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.888, + 0.746, + 0.9 + ], + "angle": 0, + "content": "\\(^{3}\\)DUA: https://github.com/jmiemirza/DUA" + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.863, + 0.824, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3733" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.082, + 0.892, + 0.122 + ], + "angle": 0, + "content": "Table 1. Top-1 Classification Error (%) for each corruption in CIFAR-10C at the highest severity (Level 5). For TTT, TENT, and DUA, we use the ResNet-26 (top), WRN-28-10 (middle), and WRN-40-2 (bottom) from their official implementation. The smallest error is shown in bold." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.125, + 0.89, + 0.387 + ], + "angle": 0, + "content": "
MethodsgausshotimpuldefcsglsmtnzmsnwfrstfgbrtcntelspxjpgAvg.
Source67.763.169.955.356.642.250.131.646.339.117.174.634.257.931.749.2
TTT [64]45.641.850.021.846.123.023.929.930.025.112.223.922.647.227.231.4
NORM [60]44.643.749.129.445.226.226.925.827.923.818.334.329.337.032.532.9
TENT [66]39.438.847.919.945.023.220.628.132.124.516.126.732.430.635.530.7
DUA [41]34.932.642.218.740.224.018.423.924.020.912.327.127.226.228.726.8
Ours33.230.638.217.741.220.817.424.027.220.413.521.128.423.728.925.8
Source72.365.772.946.954.334.842.025.141.326.09.346.726.658.530.343.5
NORM [60]28.126.136.312.835.314.212.117.317.415.38.412.623.819.727.320.4
TENT [66]24.823.533.012.031.813.710.815.916.213.77.912.122.017.324.218.6
DUA [41]27.424.635.313.134.914.611.616.817.513.17.614.122.719.326.219.9
Ours23.621.430.911.031.113.010.914.215.513.08.010.321.816.722.417.6
Source28.822.926.29.520.610.69.314.215.317.57.620.914.741.314.718.3
NORM [60]18.716.422.39.122.110.59.713.013.215.47.812.016.415.117.614.6
TENT [66]15.713.218.87.918.19.08.010.410.812.46.710.014.011.414.812.1
DUA [41]15.413.417.38.018.09.17.710.810.812.16.610.913.613.014.312.1
Ours13.412.315.07.516.08.77.79.19.610.16.48.213.39.313.310.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.162, + 0.404, + 0.806, + 0.418 + ], + "angle": 0, + "content": "Table 2. Top-1 Classification Error (%) for each corruption in CIFAR-100C at the highest severity (Level 5)." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.42, + 0.889, + 0.515 + ], + "angle": 0, + "content": "
MethodsgausshotimpuldefcsglsmtnzmsnwfrstfgbrtcntelspxjpgAvg.
Source65.760.159.132.051.033.632.441.445.251.431.655.540.359.742.446.7
NORM [60]44.744.247.432.446.432.933.039.038.445.330.236.640.637.244.239.5
TENT [66]40.339.941.829.842.331.030.034.535.239.528.033.938.433.441.436.0
DUA [41]42.240.941.030.544.832.229.938.937.243.629.539.239.035.341.237.6
Ours38.437.136.228.441.029.329.732.233.136.126.430.936.230.838.333.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.166, + 0.525, + 0.804, + 0.539 + ], + "angle": 0, + "content": "Table 3. Top-1 Classification Error (%) for each corruption in ImageNet-C at the highest severity (Level 5)." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.54, + 0.889, + 0.651 + ], + "angle": 0, + "content": "
MethodsgausshotimpuldefcsglsmtnzmsnwfrstfgbrtcntelspxjpgAvg.
Source98.997.699.293.389.090.282.387.992.099.575.999.565.360.354.085.7
TTT [64]75.576.881.989.682.879.171.383.681.098.359.099.054.753.249.675.7
NORM [60]60.260.759.876.668.767.464.264.666.274.757.088.855.853.052.364.7
TENT [66]59.459.658.772.566.164.962.162.264.968.655.297.954.552.151.762.7
DUA [41]71.972.672.490.280.883.174.776.477.987.362.699.360.858.452.674.7
Ours56.756.756.673.365.761.062.058.663.363.953.177.554.052.051.560.4
" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.676, + 0.274, + 0.691 + ], + "angle": 0, + "content": "4.4. Performance Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.471, + 0.901 + ], + "angle": 0, + "content": "The classification errors in the highest severity level corruption test datasets for test-time adaptation are reported in Tables 1, 2 and 3, with results of comparison methods directly cited from their original papers. Table 1 compares the classification error of our proposed method against recent test-time adaptation methods on the CIFAR-10C dataset. Our method performs better than other baselines with the three backbones including ResNet-26, WRN-28-10, and WRN-40-2, indicating the effectiveness of the proposed test-time adaptation method. Table 2 shows the performance comparison results on the CIFAR-100C dataset. Very encouraging results are also obtained on the large-size complicated ImageNet-C dataset, as shown in Table 3. The" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.677, + 0.892, + 0.784 + ], + "angle": 0, + "content": "mean adaptation error of the full test dataset on Gaussian noise of CIFAR-10C in the different adaptation stages is shown in Figure 3. We can see that our method outperforms the NORM and TENT methods after the 5-th batch test-time adaptation. This implies our method can reduce error faster given few testing samples. Results for lower severity levels of corruption are provided in Supplemental Materials." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Results for digit recognition from the SVHN to MNIST, MNIST-M, and USPS datasets are also reported in Table 4. All experiments use the same open-source source model. Note that the result for the TENT method is reproduced here since the pre-trained model for the TENT method was not provided and explained in its original paper. We can see that our method achieves the lowest average error when" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3734" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.122 + ], + "angle": 0, + "content": "compared to other test-time adaptation methods. The performance improvement is quite impressive." + }, + { + "type": "image", + "bbox": [ + 0.118, + 0.134, + 0.426, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.279, + 0.468, + 0.307 + ], + "angle": 0, + "content": "Figure 3. The mean adaptation error of the full test dataset on Gaussian noise of CIFAR-10C in the different adaptation stages." + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.316, + 0.468, + 0.358 + ], + "angle": 0, + "content": "Table 4. Top-1 Classification Error (\\%) for test-time adaptation on digit recognition. The asterisk (*) means the implementation with a pytorch-playground [6] pre-trained source model by us." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.359, + 0.455, + 0.43 + ], + "angle": 0, + "content": "
MethodsMNISTMNIST-MUSPSAvg.
NORM* [60]39.652.141.444.4
TENT* [66]45.856.248.350.1
Ours31.247.932.637.2
" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.443, + 0.347, + 0.459 + ], + "angle": 0, + "content": "4.5. Further Performance Analysis" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.466, + 0.236, + 0.481 + ], + "angle": 0, + "content": "4.5.1 Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.491, + 0.468, + 0.672 + ], + "angle": 0, + "content": "We conduct the ablation study with test-time adaptation tasks on the ImageNet-C dataset to investigate the contribution of our method. Hebbian learning alone does not automatically lead to a perfect posterior estimation for \\( p_t(y|X_t) \\) due to the lack of global information communication. From Table 5, we can see that feed-forward soft Hebbian Learning plays a significant effect and fine-tuning the feedback neuro-modulator of Block 1/2 also improves the performance. The average error is increased when expanding the neuro-modulator to more blocks. It is because optimizing more parameters with a mini-batch of testing samples is becoming more challenging." + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.675, + 0.468, + 0.703 + ], + "angle": 0, + "content": "Table 5. Ablation study on ImageNet-C based on ResNet-18 including 4 Blocks at the highest severity (Level 5)." + }, + { + "type": "table", + "bbox": [ + 0.084, + 0.706, + 0.465, + 0.824 + ], + "angle": 0, + "content": "
MethodsAvg. Error
Source (without adaptation)85.7
BP Conv185.0
Hebbian Conv167.2
Hebbian Conv1 + Neuromodulator (Block 1)60.5
Hebbian Conv1 + Neuromodulator (Block 1/2)60.4
Hebbian Conv1 + Neuromodulator (Block 1/2/3)61.2
Hebbian Conv1 + Neuromodulator (Block 1/2/3/4)64.8
" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.847, + 0.28, + 0.86 + ], + "angle": 0, + "content": "4.5.2 Feature Visualization" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Figure 4 compares the feature distributions on corrupted data obtained by different adaptation methods, including" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "the source model with no adaptation, the TENT method, and our method. We also include the feature distribution for the supervised learning with the labeled sample, denoted by \"Oracle\". We can see that our method is able to learn features that are close to those obtained by supervised learning." + }, + { + "type": "image", + "bbox": [ + 0.529, + 0.181, + 0.685, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.577, + 0.276, + 0.635, + 0.286 + ], + "angle": 0, + "content": "(a) Source" + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.182, + 0.865, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.76, + 0.276, + 0.815, + 0.286 + ], + "angle": 0, + "content": "(b) TENT" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.288, + 0.682, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.582, + 0.383, + 0.627, + 0.394 + ], + "angle": 0, + "content": "(c) Ours" + }, + { + "type": "image", + "bbox": [ + 0.709, + 0.289, + 0.862, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.758, + 0.382, + 0.814, + 0.394 + ], + "angle": 0, + "content": "(d) Oracle" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.4, + 0.892, + 0.469 + ], + "angle": 0, + "content": "Figure 4. Density plots of test-time adapted features distribution on CIFAR-10-C with Gaussian noise (front) and reference features without corruption (back with yellow color). Each horizontal axis corresponds to one channel. The height of each ridge indicates the number of features that take the same value." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.498, + 0.692, + 0.513 + ], + "angle": 0, + "content": "5. Further Discussions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.528, + 0.892, + 0.679 + ], + "angle": 0, + "content": "Hebbian Learning by competition through lateral inhibition is a feed-forward process that has no gradient. If combining Hebbian Learning with back-propagation, it is limited to propagate the gradient through Hebbian layers to earlier gradient-based layers during training. Therefore, the Hebbian layers can only be designed in the early layers of models without back-propagation and gradient. In addition, although the Hebbian learning rule is commonly used for long-term reinforcement, the Hebb principle does not cover all forms of long-term synaptic plasticity." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.706, + 0.619, + 0.72 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.9 + ], + "angle": 0, + "content": "In this work, we take inspiration from the biological neuromodulation process to construct a novel neuro-modulated Hebbian Learning (NHL) framework. With the unsupervised feed-forward soft Hebbian learning being combined with a learned neuro-modulator to capture feedback from external responses, the source model can be effectively adapted during the testing process. Experimental results on benchmark datasets demonstrate that our proposed method can significantly reduce the testing error for image classification with corruption, and reach new state-of-the-art performance." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.957 + ], + "angle": 0, + "content": "3735" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.156 + ], + "angle": 0, + "content": "[1] Yoshua Bengio, Aaron Courville, and Pascal Vincent. Representation learning: A review and new perspectives. IEEE TPAMI, 35(8):1798-1828, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.157, + 0.47, + 0.199 + ], + "angle": 0, + "content": "[2] Tom Binzegger, Rodney J Douglas, and Kevan AC Martin. A quantitative map of the circuit of cat primary visual cortex. Journal of Neuroscience, 24(39):8441-8453, 2004. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.2, + 0.471, + 0.242 + ], + "angle": 0, + "content": "[3] Rafal Bogacz. A tutorial on the free-energy framework for modelling perception and learning. Journal of Mathematical Psychology, 76:198-211, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.242, + 0.47, + 0.284 + ], + "angle": 0, + "content": "[4] Malik Boudiaf, Romain Mueller, Ismail Ben Ayed, and Luca Bertinetto. Parameter-free online test-time adaptation. In CVPR, pages 8344-8353, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.285, + 0.471, + 0.34 + ], + "angle": 0, + "content": "[5] K Bougrova, N Bonacchi, JA Catarino, EE Dewitt, GT Meijer, P Dayan, ZF Mainen, et al. Comparison of neuromodulator signaling in a visual decision-making task. In Annual Meeting of the Society for Neuroscience, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.341, + 0.469, + 0.382 + ], + "angle": 0, + "content": "[6] Aaron Chen and Gus Smith. pytorch-playground. https://github.com/aaron-xichen/pytorch-playground, 2020. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.383, + 0.47, + 0.424 + ], + "angle": 0, + "content": "[7] Zhixiang Chi, Yang Wang, Yuanhao Yu, and Jin Tang. Test-time fast adaptation for dynamic scene deblurring via meta-auxiliary learning. In CVPR, pages 9137-9146, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.425, + 0.471, + 0.48 + ], + "angle": 0, + "content": "[8] Sungha Choi, Seunghan Yang, Seokeon Choi, and Sungrack Yun. Improving test-time adaptation via shift-agnostic weight regularization and nearest source prototypes. arXiv preprint arXiv:2207.11707, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.481, + 0.471, + 0.548 + ], + "angle": 0, + "content": "[9] Francesco Croce, Maksym Andriushchenko, Vikash Sehwag, Edoardo Debenedetti, Nicolas Flammarion, Mung Chiang, Prateek Mittal, and Matthias Hein. Robustbench: a standardized adversarial robustness benchmark. In NeurIPS, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.55, + 0.469, + 0.592 + ], + "angle": 0, + "content": "[10] Ning Ding, Yixing Xu, Yehui Tang, Chao Xu, Yunhe Wang, and Dacheng Tao. Source-free domain adaptation via distribution estimation. In CVPR, pages 7212-7222, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.593, + 0.469, + 0.633 + ], + "angle": 0, + "content": "[11] Zhengming Ding and Yun Fu. Deep domain generalization with structured low-rank constraint. IEEE TIP, 27(1):304-313, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.635, + 0.469, + 0.676 + ], + "angle": 0, + "content": "[12] Rodney J Douglas, Kevan AC Martin, et al. Neuronal circuits of the neocortex. Annual Review of Neuroscience, 27(1):419-451, 2004. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.677, + 0.469, + 0.704 + ], + "angle": 0, + "content": "[13] Karl Friston. Hierarchical models in the brain. PLoS Computational Biology, 4(11):e1000211, 2008. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.705, + 0.469, + 0.746 + ], + "angle": 0, + "content": "[14] Karl Friston. The free-energy principle: a rough guide to the brain? Trends in Cognitive Sciences, 13(7):293-301, 2009. 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.748, + 0.469, + 0.789 + ], + "angle": 0, + "content": "[15] Karl Friston, James Kilner, and Lee Harrison. A free energy principle for the brain. Journal of Physiology, Paris, 100(1-3):70-87, 2006. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.469, + 0.83 + ], + "angle": 0, + "content": "[16] Yaroslav Ganin and Victor Lempitsky. Unsupervised domain adaptation by backpropagation. In ICML, pages 1180-1189. PMLR, 2015. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[17] Wulfram Gerstner, Marco Lehmann, Vasiliki Liakoni, Dane Corneil, and Johann Brea. Eligibility traces and plasticity on behavioral time scales: experimental support of neohoebbian three-factor learning rules. Frontiers in Neural Circuits, 12:53, 2018. 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.471, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.163 + ], + "angle": 0, + "content": "[19] Donald Olding Hebb. The organization of behavior: A neuropsychological theory. Psychology Press, 2005. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.892, + 0.203 + ], + "angle": 0, + "content": "[20] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. In ICLR, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.204, + 0.892, + 0.26 + ], + "angle": 0, + "content": "[21] Minhao Hu, Tao Song, Yujun Gu, Xiangde Luo, Jieneng Chen, Yinan Chen, Ya Zhang, and Shaoting Zhang. Fully test-time adaptation for image segmentation. In MICCAI, pages 251-260. Springer, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.261, + 0.892, + 0.315 + ], + "angle": 0, + "content": "[22] Jiaxing Huang, Dayan Guan, Aoran Xiao, and Shijian Lu. Model adaptation: Historical contrastive learning for unsupervised domain adaptation without source data. In NeurIPS, volume 34, pages 3635-3649, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.316, + 0.892, + 0.344 + ], + "angle": 0, + "content": "[23] Jonathan J. Hull. A database for handwritten text recognition research. IEEE TPAMI, 16(5):550-554, 1994. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.344, + 0.892, + 0.385 + ], + "angle": 0, + "content": "[24] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In ICML, pages 448-456. PMLR, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.386, + 0.892, + 0.428 + ], + "angle": 0, + "content": "[25] Yusuke Iwasawa and Yutaka Matsuo. Test-time classifier adjustment module for model-agnostic domain generalization. In NeurIPS, volume 34, pages 2427-2440, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.429, + 0.892, + 0.469 + ], + "angle": 0, + "content": "[26] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. Master's thesis, University of Tront, 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.47, + 0.892, + 0.511 + ], + "angle": 0, + "content": "[27] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Communications of the ACM, 60(6):84-90, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.511, + 0.892, + 0.552 + ], + "angle": 0, + "content": "[28] Dmitry Krotov and John J Hopfield. Unsupervised learning by competing hidden units. Proceedings of the National Academy of Sciences, 116(16):7723-7731, 2019. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.553, + 0.892, + 0.607 + ], + "angle": 0, + "content": "[29] Gabriele Lagani, Fabrizio Falchi, Claudio Gennaro, and Giuseppe Amato. Hebbian semi-supervised learning in a sample efficiency setting. *Neural Networks*, 143:719–731, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.609, + 0.892, + 0.637 + ], + "angle": 0, + "content": "[30] Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Deep learning. Nature, 521(7553):436-444, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.637, + 0.892, + 0.691 + ], + "angle": 0, + "content": "[31] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.692, + 0.892, + 0.748 + ], + "angle": 0, + "content": "[32] Andrew B Lehr, Jannik Luboeinski, and Christian Tetzlaff. Neuromodulator-dependent synaptic tagging and capture retroactively controls neural coding in spiking neural networks. Scientific Reports, 12(1):1-18, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.749, + 0.892, + 0.789 + ], + "angle": 0, + "content": "[33] Rui Li, Qianfen Jiao, Wenming Cao, Hau-San Wong, and Si Wu. Model adaptation: Unsupervised domain adaptation without source data. In CVPR, pages 9641-9650, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.79, + 0.892, + 0.845 + ], + "angle": 0, + "content": "[34] Yizhuo Li, Miao Hao, Zonglin Di, Nitesh Bharadwaj Gundavarapu, and Xiaolong Wang. Test-time personalization with a transformer for human pose estimation. In NeurIPS, volume 34, pages 2583-2597, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[35] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In ICML, pages 6028-6039. PMLR, 2020. 1, 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.517, + 0.956 + ], + "angle": 0, + "content": "3736" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.146 + ], + "angle": 0, + "content": "[36] Timothy P Lillicrap, Adam Santoro, Luke Marris, Colin J Akerman, and Geoffrey Hinton. Backpropagation and the brain. Nature Reviews Neuroscience, 21(6):335-346, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.148, + 0.47, + 0.204 + ], + "angle": 0, + "content": "[37] Yuejiang Liu, Parth Kothari, Bastien van Delft, Baptiste Bellot-Gurlet, Taylor Mordan, and Alexandre Alahi. Ttt++: When does self-supervised test-time training fail or thrive? In NeurIPS, volume 34, pages 21808-21820, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.205, + 0.469, + 0.247 + ], + "angle": 0, + "content": "[38] Mingsheng Long, Yue Cao, Jianmin Wang, and Michael Jordan. Learning transferable features with deep adaptation networks. In ICML, pages 97-105. PMLR, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.247, + 0.469, + 0.287 + ], + "angle": 0, + "content": "[39] Robert A Marsden, Mario Döbler, and Bin Yang. Gradual test-time adaptation by self-training and style transfer. arXiv preprint arXiv:2208.07736, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.288, + 0.469, + 0.355 + ], + "angle": 0, + "content": "[40] Muhammad Jehanzeb Mirza, Cornelius Buerkle, Julio Jarquin, Michael Opitz, Fabian Oboril, Kay-Ulrich Scholl, and Horst Bischof. Robustness of object detectors in degrading weather conditions. In ITSC, pages 2719-2724. IEEE, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.357, + 0.469, + 0.412 + ], + "angle": 0, + "content": "[41] M Jehanzeb Mirza, Jakub Micorek, Horst Possegger, and Horst Bischof. The norm must go on: Dynamic unsupervised domain adaptation by normalization. In CVPR, pages 14765-14775, 2022. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.413, + 0.469, + 0.481 + ], + "angle": 0, + "content": "[42] Timoleon Moraitis, Dmitry Toichkin, Adrien Journé, Yansong Chua, and Qinghai Guo. Softhebb: Bayesian inference in unsupervised hebbian soft winner-take-all networks. *Neuromorphic Computing and Engineering*, 2(4):044017, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.483, + 0.469, + 0.536 + ], + "angle": 0, + "content": "[43] Saeid Motiian, Marco Piccirilli, Donald A Adjeroh, and Gianfranco Doretto. Unified deep supervised domain adaptation and generalization. In ICCV, pages 5715-5725, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.538, + 0.469, + 0.592 + ], + "angle": 0, + "content": "[44] Krikamol Muandet, David Balduzzi, and Bernhard Schölkopf. Domain generalization via invariant feature representation. In ICML, pages 10-18. PMLR, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.594, + 0.469, + 0.65 + ], + "angle": 0, + "content": "[45] Bernhard Nessler, Michael Pfeiffer, Lars Buesing, and Wolfgang Maass. Bayesian computation emerges in generic cortical microcircuits through spike-timing-dependent plasticity. PLoS Computational Biology, 9(4):e1003037, 2013. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.65, + 0.469, + 0.692 + ], + "angle": 0, + "content": "[46] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y Ng. Reading digits in natural images with unsupervised feature learning. 2011. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.692, + 0.469, + 0.733 + ], + "angle": 0, + "content": "[47] Erkki Oja. Simplified neuron model as a principal component analyzer. Journal of Mathematical Biology, 15(3):267-273, 1982. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.734, + 0.469, + 0.802 + ], + "angle": 0, + "content": "[48] John O'Donnell, Douglas Zeppenfeld, Evan McConnell, Salvador Pena, and Maiken Nedergaard. Norepinephrine: a neuromodulator that boosts the function of multiple cell types to optimize cns performance. Neurochemical Research, 37(11):2496-2512, 2012. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.469, + 0.859 + ], + "angle": 0, + "content": "[49] Prashant Pandey, Mrigank Raman, Sumanth Varambally, and Prathosh Ap. Generalization on unseen domains via inference-time label-preserving target projections. In CVPR, pages 12924-12933, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[50] Zhongyi Pei, Zhangjie Cao, Mingsheng Long, and Jianmin Wang. Multi-adversarial domain adaptation. In AAAI, volume 32, 2018. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[51] Marina R Picciotto, Michael J Higley, and Yann S Mineur. Acetylcholine as a neuromodulator: cholinergic signaling shapes nervous system function and behavior. *Neuron*, 76(1):116-129, 2012. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[52] Roman Pogodin and Peter Latham. Kernelized information bottleneck leads to biologically plausible 3-factor hebbian learning in deep networks. In NeurIPS, volume 33, pages 7296-7307, 2020. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.205, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[53] Fengchun Qiao, Long Zhao, and Xi Peng. Learning to learn single domain generalization. In CVPR, pages 12556-12565, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.305 + ], + "angle": 0, + "content": "[54] Zhen Qiu, Yifan Zhang, Hongbin Lin, Shuaicheng Niu, Yanxia Liu, Qing Du, and Mingkui Tan. Source-free domain adaptation via avatar prototype generation and adaptation. In IJCAI, pages 2921-2927, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.306, + 0.892, + 0.347 + ], + "angle": 0, + "content": "[55] Joaquin Quinonero-Candela, Masashi Sugiyama, Anton Schwaighofer, and Neil D Lawrence. Dataset shift in machine learning. MIT Press, 2008. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.349, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[56] Rajesh PN Rao and Dana H Ballard. Predictive coding in the visual cortex: a functional interpretation of some extra-classical receptive-field effects. Nature Neuroscience, 2(1):79-87, 1999. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.405, + 0.892, + 0.446 + ], + "angle": 0, + "content": "[57] David E Rumelhart and David Zipser. Feature discovery by competitive learning. Cognitive Science, 9(1):75-112, 1985. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.448, + 0.892, + 0.516 + ], + "angle": 0, + "content": "[58] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. IJCV, 115(3):211-252, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.518, + 0.892, + 0.56 + ], + "angle": 0, + "content": "[59] Ueli Rutishauser, Rodney J Douglas, and Jean-Jacques Slotine. Collective stability of networks of winner-take-all circuits. Neural Computation, 23(3):735-773, 2011. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.562, + 0.892, + 0.63 + ], + "angle": 0, + "content": "[60] Steffen Schneider, Evgenia Rusak, Luisa Eck, Oliver Bringmann, Wieland Brendel, and Matthias Bethge. Improving robustness against common corruptions by covariate shift adaptation. In NeurIPS, volume 33, pages 11539-11551, 2020. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.632, + 0.892, + 0.688 + ], + "angle": 0, + "content": "[61] Seonguk Seo, Yumin Suh, Dongwan Kim, Geeho Kim, Jongwoo Han, and Bohyung Han. Learning to optimize domain specific normalization for domain generalization. In ECCV, pages 68-83. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.689, + 0.892, + 0.731 + ], + "angle": 0, + "content": "[62] Assaf Shocher, Nadav Cohen, and Michal Irani. \"zero-shot\" super-resolution using deep internal learning. In CVPR, pages 3118-3126, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.732, + 0.892, + 0.787 + ], + "angle": 0, + "content": "[63] Luisa Speranza, Umberto di Porzio, Davide Viggiano, Antonio de Donato, and Floriana Volpicelli. Dopamine: The neuromodulator of long-term synaptic plasticity, reward and movement control. Cells, 10(4):735, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.789, + 0.892, + 0.856 + ], + "angle": 0, + "content": "[64] Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei A. Efros, and Moritz Hardt. Test-time training with self-supervision for generalization under distribution shifts. In ICML, volume 119, pages 9229-9248. PMLR, 2020. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[65] Riccardo Volpi and Vittorio Murino. Addressing model vulnerability to distributional shifts over image transformation sets. In ICCV, pages 7980-7989, 2019. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3737" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.134 + ], + "angle": 0, + "content": "[66] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In ICLR, 2020. 1, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.136, + 0.468, + 0.189 + ], + "angle": 0, + "content": "[67] Fan Wang, Zhongyi Han, Yongshun Gong, and Yilong Yin. Exploring domain-invariant parameters for source free domain adaptation. In CVPR, pages 7151-7160, June 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.192, + 0.468, + 0.233 + ], + "angle": 0, + "content": "[68] Qin Wang, Olga Fink, Luc Van Gool, and Dengxin Dai. Continual test-time domain adaptation. In CVPR, pages 7201-7211, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.235, + 0.468, + 0.275 + ], + "angle": 0, + "content": "[69] James CR Whittington and Rafal Bogacz. Theories of error back-propagation in the brain. Trends in cognitive sciences, 23(3):235-250, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.278, + 0.468, + 0.319 + ], + "angle": 0, + "content": "[70] Shiqi Yang, Yaxing Wang, Joost van de Weijer, Luis Herranz, and Shangling Jui. Generalized source-free domain adaptation. In ICCV, pages 8978-8987, October 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.321, + 0.468, + 0.359 + ], + "angle": 0, + "content": "[71] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. In BMVC. British Machine Vision Association, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.363, + 0.468, + 0.404 + ], + "angle": 0, + "content": "[72] Matthew D Zeiler and Rob Fergus. Visualizing and understanding convolutional networks. In ECCV, pages 818-833. Springer, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.407, + 0.468, + 0.459 + ], + "angle": 0, + "content": "[73] Shanshan Zhao, Mingming Gong, Tongliang Liu, Huan Fu, and Dacheng Tao. Domain generalization via entropy regularization. In NeurIPS, volume 33, pages 16096-16107, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.463, + 0.468, + 0.503 + ], + "angle": 0, + "content": "[74] Kaiyang Zhou, Ziwei Liu, Yu Qiao, Tao Xiang, and Chen Change Loy. Domain generalization: A survey. IEEE TPAMI, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.503 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3738" + } + ] +] \ No newline at end of file diff --git a/2023/Neuro-Modulated Hebbian Learning for Fully Test-Time Adaptation/28e8edf6-185c-4dae-9bfc-b17f1c2f7606_origin.pdf b/2023/Neuro-Modulated Hebbian Learning for Fully Test-Time Adaptation/28e8edf6-185c-4dae-9bfc-b17f1c2f7606_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..12b81ec19a2f6bac1e1b2ea03f9d7dd29b13e1eb --- /dev/null +++ b/2023/Neuro-Modulated Hebbian Learning for Fully Test-Time Adaptation/28e8edf6-185c-4dae-9bfc-b17f1c2f7606_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fe95f732f46cbcdd22210f86b44d003e8380f77a0a67166bc1c5e1b8d440e416 +size 1659519 diff --git a/2023/Neuro-Modulated Hebbian Learning for Fully Test-Time Adaptation/full.md b/2023/Neuro-Modulated Hebbian Learning for Fully Test-Time Adaptation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..342ef18def890410a006e4e2f5418ff0682318c0 --- /dev/null +++ b/2023/Neuro-Modulated Hebbian Learning for Fully Test-Time Adaptation/full.md @@ -0,0 +1,356 @@ +# Neuro-Modulated Hebbian Learning for Fully Test-Time Adaptation + +Yushun Tang $^{1,2}$ , Ce Zhang $^{1}$ , Heng Xu $^{1}$ , Shuoshuo Chen $^{1}$ , Jie Cheng $^{2}$ , Luziwei Leng $^{2}$ , Qinghai Guo $^{2*}$ , Zhihai He $^{1,3*}$ + +1Department of Electronic and Electrical Engineering, Southern University of Science and Technology, Shenzhen, China + +$^{2}$ Advanced Computing and Storage Laboratory, Huawei Technologies Co., Ltd., Shenzhen, China + +$^{3}$ Pengcheng Laboratory, Shenzhen, China + +{tangys2022, zhangc2019, xuh2022, chenss2021}@mail.sustech.edu.cn + +{chengjie8, lengluziwei, guoqinghai}@huawei.com, hezh@sustech.edu.cn + +# Abstract + +Fully test-time adaptation aims to adapt the network model based on sequential analysis of input samples during the inference stage to address the cross-domain performance degradation problem of deep neural networks. We take inspiration from the biological plausibility learning where the neuron responses are tuned based on a local synapse-change procedure and activated by competitive lateral inhibition rules. Based on these feed-forward learning rules, we design a soft Hebbian learning process which provides an unsupervised and effective mechanism for online adaptation. We observe that the performance of this feed-forward Hebbian learning for fully test-time adaptation can be significantly improved by incorporating a feedback neuro-modulation layer. It is able to fine-tune the neuron responses based on the external feedback generated by the error back-propagation from the top inference layers. This leads to our proposed neuro-modulated Hebbian learning (NHL) method for fully test-time adaptation. With the unsupervised feed-forward soft Hebbian learning being combined with a learned neuro-modulator to capture feedback from external responses, the source model can be effectively adapted during the testing process. Experimental results on benchmark datasets demonstrate that our proposed method can significantly improve the adaptation performance of network models and outperforms existing state-of-the-art methods. + +# 1. Introduction + +Although deep neural networks have achieved great success in various machine learning tasks, their performance tends to degrade significantly when there is data shift [27, 55] between the training data in the source do + +main and the testing data in the target domain [40]. To address the performance degradation problem, unsupervised domain adaptation (UDA) [16,38,50] has been proposed to fine-tune the model parameters with a large amount of unlabeled testing data in an unsupervised manner. Source-free UDA methods [33, 35, 67] aim to adapt the network model without the need to access the source-domain samples. + +There are two major categories of source-free UDA methods. The first category needs to access the whole test dataset on the target domain to achieve their adaptation performance [35, 67]. Notice that, in many practical scenarios when we deploy the network model on client devices, the network model does not have access to the whole dataset in the target domain since collecting and constructing the test dataset on the client side is very costly. The second type of method, called fully test-time adaptation, only needs access to live streams of test samples [41, 64, 66], which is able to dynamically adapt the source model on the fly during the testing process. Existing methods for fully test-time adaptation mainly focus on constructing various loss functions to regulate the inference process and adapt the model based on error back-propagation. For example, the TENT method [66] updates the batch normalization module by minimizing an entropy loss. The TTT method [64] updates the feature extractor parameters according to a self-supervised loss on a proxy learning task. The TTT++ method [37] introduces a feature alignment strategy based on online moment matching. + +# 1.1. Challenges in Fully Test-Time UDA + +We recognize that most of the domain variations, such as changes in the visual scenes and image transformations or corruptions, are early layers of features in the semantic hierarchy [66]. They can be effectively captured and modeled by lower layers of the network model. From the perspective of machine learning, early representations through the lower layer play an important role to capture the pos + +terior distribution of the underlying explanatory factors for the observed input [1]. For instance, in deep neural network models, the early layers of the network tend to respond to corners, edges, or colors. In contrast, deeper layers respond to more class-specific features [72]. In the corruption test-time adaptation scenario, the class-specific features are always the same because the testing datasets are the corruption of the training domain. However, the early layers of models can be failed due to corruption. + +Therefore, the central challenge in fully test-time UDA lies in how to learn useful early layer representations of the test samples without supervision. Motivated by this observation, we propose to explore neurobiology-inspired Hebbian learning for effective early-layer representation learning and fully test-time adaptation. It has been recognized that the learning rule of supervised end-to-end deep neural network training using back-propagation and the learning rules of the early front-end neural processing in neurobiology are unrelated [28]. The responses of neurons in biological neural networks are tuned by local pre-synaptic and post-synaptic activity, along with global variables that measure task performance, rather than the specific activity of other neurons [69]. + +![](images/814346d0284596a998a03f03241a53d6d35566a32707afcdebc88afd959575bc.jpg) +Figure 1. The feature map visualization after the first convolution layer obtained by different learning methods. + +# 1.2. Hebbian Learning + +Hebbian learning aims to learn useful early layer representations without supervision based on local synaptic plasticity rules, which is able to generate early representations that are as good as those learned by end-to-end supervised training with back-propagation [28, 52]. Drastically different from the current error back-propagation methods which require pseudo-labels or loss functions from the top network layers, Hebbian learning is a pure feed-forward adaptation process and does not require feedback from the distant top network layers. The responses of neurons are tuned based on a local synapse-change procedure and activated by competitive lateral inhibition rules [28]. During the learning process, the strength of synapses undergoes local changes that are proportional to the activity of + +the pre-synaptic cell and dependent on the activity of the post-synaptic cell. It also introduces local lateral inhibition between neurons within a layer, where the synapses of hidden units with strong responses are pushed toward the patterns that drive them, while those with weaker responses are pushed away from these patterns. + +Existing literature has shown that early representations learned by Hebbian learning are as well as backpropagation and even more robust in testing [28, 29, 52]. Figure 1 shows the feature maps learned by different methods. The first row shows the original image. The second row shows the images in the target domain with significant image corruption. The third row shows the feature maps learned by the network model trained in the source domain for these target-domain images. The fourth row shows the feature maps learned by our Hebbian learning method. The last row ("oracle") shows the feature maps learned with true labels. We can see that the unsupervised Hebbian learning is able to generate feature maps which are as good as those from supervised learning. + +# 1.3. Our Major Idea + +In this work, we observe that Hebbian learning, although provides a new and effective approach for unsupervised learning of early layer representation of the image, when directly applied to the network model, is not able to achieve satisfactory performance in fully test-time adaptation. First, the original hard decision for competitive learning is not suitable for fully test-time adaptation. Second, the Hebbian learning does not have an effective mechanism to consider external feedback, especially the feedback from the top network layers. We observe that, biologically, the visual processing is realized through hierarchical models considering a bottom-up early representation learning for the sensory input, and a top-down feedback mechanism based on predictive coding [15, 56]. + +Motivated by this, in this work, we propose to develop a new approach, called neuro-modulated Hebbian learning (NHL), for fully test-time adaptation. We first incorporate a soft decision rule into the feed-forward Hebbian learning to improve its competitive learning. Second, we learn a neuro-modulator to capture feedback from external responses, which controls which type of feature is consolidated and further processed to minimize the predictive error. During inference, the source model is adapted by the proposed NHL rule for each mini-batch of testing samples during the inference process. Experimental results on benchmark datasets demonstrate that our proposed method can significantly improve the adaptation performance of network models and outperforms existing state-of-the-art methods. + +# 1.4. Summary of Major Contributions + +To summarize, our major contributions include: (1) we identify that the major challenge in fully test-time adaptation lies in effective unsupervised learning of early layer representations, and explore neurobiology-inspired soft Hebbian learning for effective early layer representation learning and fully test-time adaptation. (2) We develop a new neuro-modulated Hebbian learning method which combines unsupervised feed-forward Hebbian learning of early layer representation with a learned neuro-modulator to capture feedback from external responses. We analyze the optimal property of the proposed NHL algorithm based on free-energy principles [14, 15]. (3) We evaluate our proposed NHL method on benchmark datasets for fully test-time adaptation, demonstrating its significant performance improvement over existing methods. + +# 2. Related Work + +In this section, we review existing methods related to our work including test-time adaptation, source-free UDA, domain generalization, and unsupervised Hebbian learning. + +# 2.1. Test-time Adaptation + +Test-time adaptation aims to online adapt the trained model while testing the input samples in the target domain. Sun et al. [64] proposed Test-time Training (TTT) by optimizing a self-supervised loss through a proxy task on the source before adapting to the target domain. The method proposed in TTT++ [37] aims to minimize the distribution shift between the training and testing feature distributions by dynamically matching the moments online. It should be noted that this method requires specific training on the source data, which is not available in fully test-time adaptation scenarios [66]. TENT [66] fine-tuned the scale and bias parameters of the batch normalization layers using an entropy minimization loss during the inference process. DUA [41] adapted the statistics of the batch normalization layer only on a tiny fraction of test data and augmented a small batch of target data to adapt the model. Choi et al. [8] proposed a shift-agnostic weight regularization and an auxiliary task for the alignment between the source and target features. Note that this method requires the source data for computing the source prototypes. The continual test-time adaptation methods [39,68] consider online TTA where target data is continually changing during inference. Instead of using parameters of the pre-trained model, Boudiaf et al. [4] only adapted the model's output by optimizing an objective function based on Laplacian adjusted maximum-likelihood estimation. Besides image classification, test-time adaptation has been successfully applied in various machine learning tasks, such as scene deblurring [7], superresolution [62], human pose estimation [34], image seg + +mentation [21], object detection [41], etc. + +# 2.2. Source-free Unsupervised Domain Adaptation + +Recently, source-free UDA has emerged as an important research topic. It aims to adapt the source models to unlabeled target domains without accessing the data from the source domain. Among them, SHOT [35] proposed to learn target-specific features based on an information maximization criteria and pseudo-label prediction. 3C-GAN [33] generated target-style images by training a collaborative class conditional GAN module and using a clustering-based regularization scheme. G-SFDA [70] proposed a domain-specific attention mechanism that selectively activates different feature channels for different domains. CPGA [54] proposed to generate avatar prototypes instead of images via contrastive learning. HCL [22] proposed a solution for addressing the lack of source data by introducing both instance-level and category-level historical contrastive learning. DIPE [67] focuses on exploring the domain-invariant parameters of the model, rather than trying to learn domain-invariant representations. SFDA-DE [10] aligned domains by estimating source class-conditioned feature distribution and minimizing a contrastive adaptation loss function. Existing source-free methods are offline. They need to analyze the whole test dataset and update the model for a number of adaptation epochs. + +# 2.3. Domain Generalization + +Domain generalization aims to train a model only on the source data that are generalizable to unseen target domains [74]. Domain generalization methods typically aim to learn domain-invariant representations by aligning the distributions of the source domains [43, 44, 73]. Methods based on data augmentation [53, 65] for regularizing the training process have been studied. Ensemble learning methods [11, 61] generate an ensemble of different model weights from different partitions of the training data. Recently, inference-time optimization without training the network model has been studied for domain generalization. Pandey et al. [49] used a generative model to project the target data onto the source-feature manifold with labels being preserved by solving an optimization problem during the inference stage. Iwasawa et al. [25] proposed a backpropagation-free generalization method by computing distance to a pseudo-prototype representation. + +# 2.4. Unsupervised Hebbian Learning + +In traditional end-to-end training, back-propagation is usually used to update the weights of deep neural networks. It has been recognized that the learning rule of supervised end-to-end deep neural network training using back-propagation is much different from the learning rules of the early front-end neural processing in neurobiology. In + +addition, supervised training of deep neural networks with back-propagation requires a large amount of labeled samples [30]. The tuning of neuron responses in biological neural networks is achieved through a physically local synapse-change procedure. The superficial cerebral cortex exhibits a common connectivity pattern in which neurons are activated through competition via lateral inhibition [2, 12]. This competition leads to the suppression of weakly activated neurons and the amplification of strongly activated ones, a phenomenon known as competitive learning or "winnertakes-all (WTA)" learning [57, 59]. Motivated by Hebb's idea [19], several biological plausibility learning rules have been proposed, where changes of the synapse strength depend only on the activities of the pre-synaptic and postsynaptic neurons. The Oja's rule [47] proposed a linear neuron model with constrained Hebbian synaptic modification and derived a new unconstrained learning method. Krotov et al. [28] proposed a family of biologically plausible learning rules that enable the learning of early representations that are comparable to those achieved by end-to-end supervised training with back-propagation. Pogodin et al. [52] presents a family of learning rules which use pre- and post-synaptic firing rates and a global teaching signal. They perform almost the same as the back-propagation method. + +# 3. Method + +In this section, we present our method of neuromodulated Hebbian learning for fully test-time adaptation. + +# 3.1. Problem Formulation + +Suppose that a model $q_{\theta_s}(y|X_s)$ with parameters $\theta_s$ has been successfully trained on the source data $\{X_s\}$ with labels $\{Y_s\}$ where the true distribution is $p_s(y|X_s)$ . It consists of a feature extractor $\mathbf{F}$ and a classifier $\mathbf{C}$ . During fully test-time adaptation, we are given the target data $\{X_t\}$ with unknown labels $\{Y_t\}$ . Our goal is to adapt the trained model $q_{\theta_s}$ in an unsupervised manner during testing to approximate the true target distribution $p_t(y|X_t)$ . Given a sequence of input sample batches $\{B_1, B_2, \dots, B_n\}$ , the $i$ -th adaptation of the network model can only rely on the $i$ -th batch of test samples $B_i$ . + +We consider this transfer learning on a new target domain as an active inference process. The pre-trained model $q_{\theta_s}(y|X_s)$ is considered as a prior. Thus, predicting the labels for target samples becomes a Bayesian posterior $q_{\theta_1}(y|B_1)$ for the first batch, given the prior model $q_{\theta_s}(y|X_s)$ . After the first adaptation batch, the prior model is updated to $q_{\theta_1}(y|B_1)$ . This Bayesian inference process is repeated for all subsequent batches and produces a final posterior estimation + +$$ +q _ {\theta , t} (y \mid X _ {t}) = q _ {\theta_ {n}} (y \mid B _ {n}), \tag {1} +$$ + +given the prior model $q_{\theta_{n - 1}}(y|B_{n - 1})$ + +![](images/731f0308d9923657f8a57da1de826732da748b5d978c09f8a9a18d6ed1a16429.jpg) +Figure 2. An overview of the proposed NHL method. During inference, the first convolution layer of the source model is fine-tuned by the Hebbian learning rule and the neuromodulator is further fine-tuned by the entropy loss before making a prediction given each mini-batch testing sample. The lock symbol means the classifier is fixed in the test-time adaptation process. + +# 3.2. Overview of Our NHL Method + +As illustrated in Figure 2, our proposed neuro-modulated Hebbian learning consists of two major components: the feed-forward soft Hebbian learning layer and the neuromodulator. The soft Hebbian learning layer aims to learn useful early layer representations without supervision based on local synaptic plasticity and soft competitive learning rules. It is able to generate early representations which are as good as those learned by end-to-end supervised training with labeled samples and back-propagation. During our experiments, we find that this soft Hebbian learning layer can significantly improve the performance of the network model in the target domain. However, we recognize that the feedforward Hebbian learning only is not able to achieve competitive performance as the current state-of-the-art methods for fully test-time adaptation. We find that it lacks the capability to effectively respond to external stimulus, specifically the feedback from the network output layer. + +To address this issue, we propose to design a neuromodulator layer, an intermediate layer or an interface between the soft Hebbian learning layer and the classifier module of the network. This neuro-modulator layer is updated using back-propagation with the entropy loss being evaluated at the network output. It serves as the bridge between two different learning approaches: the feed-forward Hebbian learning and the original error back-propagation. From the ablation studies summarized in Table 5, we can see that both algorithm components, the feed-forward soft Hebbian learning, and the neuro-modulator are contributing significantly to the overall performance improvement. + +The proposed NHL method can be also formulated by the free-energy principle in cognitive science [14, 15, 56]. A hierarchical predictive coding model was proposed by Rao + +and Ballard [56] to learn a hierarchical internal model for human perception by maximizing the posterior probability of generating the observed data. This can be realized by a concurrent process of prediction through a bottom-up feedforward generation process (such as our Hebbian learning layer) combined with a top-down feedback-based optimization process (such as our neuro-modulator). Specifically, given the sensory input $x$ , assume $x$ is generated by environmental causes $\vartheta$ , denoted as $p(x) = p(x, \vartheta)$ . The free-energy principle states that the brain encodes the recognition density over sensory causes [14]. Mathematically, it optimizes a generative probabilistic mixture as $q_{1}(x) := q(x, \vartheta)$ . It has been demonstrated that this process can be achieved by a Hebbian-like learning approach [42, 45]. + +On the other hand, to approximate the true distribution $p_t(y|X_t)$ by a posterior approximation $q_2(y) \coloneqq q_{\theta,t}(y|X_t)$ , one can consider the similarity between these two distributions measured by the following Kullback-Leibler (KL) divergence + +$$ +\operatorname {K L} \left[ q _ {2} (y) | | p _ {t} (y | X _ {t}) \right] = \int q _ {2} (y) \log \frac {q _ {2} (y)}{p _ {t} (y | X _ {t})} d y. \tag {2} +$$ + +According to the analysis in Bogacz [3], this minimization of KL-divergence can be converted to minimization of the free-energy $F$ defined as: + +$$ +F = \int q _ {2} (y) \log \frac {q _ {2} (y)}{p _ {t} \left(X _ {t} , y\right)} d y. \tag {3} +$$ + +The maximization of the free-energy $F$ can be solved by an active inference process if we consider the adaptation for each batch of test samples as a decision-making step [13]. This leads to a feedback-based optimization, just like the neuro-modulation as proposed in our NHL method. In the following sections, we explain the two major components of our NHL method in more detail. + +# 3.3. Feed-Forward Soft Hebbian Learning + +As discussed above, the early feed-forward layer aims to learn useful layer representations of the input without any supervision. An approximate solution to this problem can be reduced to finding the first principal component of the input data [47], known as Oja's rule and its variations. The Oja's rule is itself a variation of the Hebbian rule, plus a normalization condition. In Oja's rule, the plasticity defined for a synaptic weight $w_{ik}$ which connects a pre-synaptic neuron $i$ with input $x_{i}$ to a post-synaptic neuron $k$ with activation $y_{k}$ is: + +$$ +\Delta w _ {i k} = \eta y _ {k} \left(x _ {i} - y _ {k} w _ {i k}\right). \tag {4} +$$ + +These weights are updated based on a plasticity rule [28]. We recognize that this process lacks the ability to detect different features of the data by different hidden units in the network. To address this problem, we propose to identify a + +good set of weights, by considering a generative model to optimize the distribution similarity to the input data. + +Specifically, we assume that the target-domain input samples $\{X_{t}\}$ are generated by hidden causes $\vartheta$ with distribution $p_t(x)$ . We define an approximation to $p_t(x)$ by $q(x|\boldsymbol{w})$ conditioned on the weights $\boldsymbol{w}$ . Moreover, we use a mixture of exponential functions to define the probability: + +$$ +q (x \mid \boldsymbol {w} _ {k}) = \exp (\langle \boldsymbol {w} _ {k}, x \rangle) / N, \tag {5} +$$ + +where $\pmb{w}_k$ is the weight vector corresponds to the postsynaptic neuron $k$ , and $N$ is a normalization factor to ensure that $q$ is a probabilistic measure. In the above generative model, the objective function is given by the KL divergence $\mathrm{KL}[p_t(x)|q(x|\pmb{w})]$ . It can be shown that the optimal parameter vector that optimizes the KL-divergence is proportional to the mean of the input distribution [45]. + +On the other hand, (5) corresponds to neural interpretation as the activation function with normalized weights $\pmb{w}_k / R$ , where $R$ is the norm of the weight vector $\pmb{w}_k$ . For the network layer with $K$ neurons, we define the output of the $k$ -th neuron to be: + +$$ +y _ {k} = e ^ {\frac {u _ {k}}{\tau}} / \left(\sum_ {i = 1} ^ {K} e ^ {\frac {u _ {i}}{\tau}}\right), \tag {6} +$$ + +where $u_{k}$ is the $k$ -th neuron's weighted input, i.e. $u_{k} = \sum_{i} w_{ik} \cdot x_{i}$ . $\tau$ is a temperature-scaling hyper-parameter. This leads to a new soft Hebbian plasticity rule: + +$$ +\Delta w _ {i k} = \eta y _ {k} \left(R x _ {i} - u _ {k} w _ {i k}\right), \tag {7} +$$ + +where $\eta$ is the learning rate. It can be shown that, using the plasticity rule in (7) to update the weights, they can converge to the equilibrium which lies in the sphere of radius $R$ . The derivative of the weight vector norm can be expressed as: + +$$ +\begin{array}{l} \frac {d \| \boldsymbol {w} _ {k} \|}{d t} = 2 \boldsymbol {w} _ {k} \frac {d \boldsymbol {w} _ {k}}{d t} = 2 \boldsymbol {w} _ {k} \eta y _ {k} (R \boldsymbol {x} - u _ {k} \boldsymbol {w} _ {k}) \\ = 2 \eta \boldsymbol {w} _ {k} y _ {k} \left(R \boldsymbol {x} - \boldsymbol {w} _ {k} \boldsymbol {x} \boldsymbol {w} _ {k}\right) = 2 \eta u _ {k} y _ {k} \left(R - \| \boldsymbol {w} _ {k} \|\right). \tag {8} \\ \end{array} +$$ + +With this, we see that the norm of the weight vector convergences to a sphere of radius $R$ . This is because the norm of the weights decreases if $\| \pmb{w}_k\| > R$ and increases if $\| \pmb{w}_k\| < R$ . More details are provided in the Supplemental Materials. + +# 3.4. The Neuro-Modulation Layer + +From the ablation studies in Section 4, we can see that the above soft Hebbian learning alone does not automatically lead to a perfect posterior estimation for $p_t(y|X_t)$ . It cannot achieve the state-of-the-art performance for fully test-time adaptation. This is because it does not have an effective mechanism to consider external feedback, especially + +the feedback from the top network layers. Our proposed solution is to incorporate one or more modulating layers to steer the updates of weights to the desired outcome [17,36]. This so-called neuro-modulator has been explored in neuroscience research [5,48,51,63]. Recent research [32] shows that the level of neuro-modulation may change the process of synaptic consolidation, thus ultimately controlling which type of information is consolidated in the upper neural network. Unlike the above neuromodulator-based learning, where the modulator factor is embedded inside the Hebbian rule, we consider such a neuromodulation process in a disentangled way derived from the free-energy principle. It serves as an interface between the feed-forward Hebbian layer and the top decision network. + +As defined in (3), the problem of minimizing the KL-divergence for $q_{2}(y)$ and its true posterior $p_t(y|X_t)$ can be formulated based on the free-energy principle: + +$$ +\operatorname {K L} \left[ q _ {2} (y) \mid \left| p _ {t} (y \mid X _ {t}) \right. \right] = F + \log P _ {t} \left(X _ {t}\right), \tag {9} +$$ + +where $P_{t}(X_{t}) \coloneqq \int q_{2}(y)p_{t}(X_{t})dy = p_{t}(X_{t})$ is the normalization term. Note that this term does not depend on $q_{2}(y)$ . Therefore, minimizing the KL-divergence is reduced to minimizing $F$ . To this end, given a batch $B$ of data in the target domain, we rewrite (use $p_t(y,B) = p_t(B|y)p_t(y)$ ) and decompose the free-energy $F_{B}$ for current batch into the following two items: + +$$ +F _ {B} = \operatorname {K L} \left[ q _ {2} (y | B) \mid \mid p _ {t} (y) \right] - \int_ {y} q _ {2} (y | B) \log p _ {t} (B | y) d y. \tag {10} +$$ + +The first term in (10) is already minimized through soft Hebbian learning, while minimizing the second term requires the likelihood distribution $p_t(B|y)$ . Since $p_t(B|y) = p_t(y|B)p_t(B) / p_t(y)$ and $q_2(y|B)$ is considered as an approximation of $p_t(y|B)$ , we minimize the entropy of $y$ given the current batch $B$ in a discrete way as: + +$$ +\underset {\boldsymbol {w}} {\arg \min } H (y | B) = \underset {\boldsymbol {w}} {\arg \max } \sum q _ {2} (y | B) \log q _ {2} (y | B), \tag {11} +$$ + +where $\mathbf{w}$ are the weights in the layer implementing the neuro-modulator. As in existing deep neural network training, we can use gradient descent to optimize these weights. + +# 4. Experiments + +In this section, we conduct experiments on multiple test-time adaptation benchmark datasets to evaluate the performance of our proposed NHL method. + +# 4.1. Benchmark Datasets + +We evaluate our method among the following popular benchmark datasets for test-time adaptation. (1) CIFAR-10/100C. We choose CIFAR-10/100 [26] with 10/100 + +classes, a source training set of 50,000, and a target testing set CIFAR-10/100C [20] of 10,000 for small-size image experiments at an accessible scale. (2) ImageNet-C. We choose the ImageNet [58] with 1,000 classes, a source training set of 1.2 million, and a validation set of 50,000 for large-size image experiments. It should be noted that we use a fixed target testing subset of the validation set with 5000 images on which all models are evaluated following the RobustBench protocol [9]. (3) SVHN $\rightarrow$ MNIST/MNIST-M/USPS. Following TENT for domain adaptation, we choose a SVHN [46] source model and transfer it to MNIST [31] (10,000 test samples), MNIST-M [16] (10,000 test samples) and USPS [23] (2,007 test samples), respectively. + +# 4.2. Comparison Methods + +We compare our method against the following fully test-time adaptation methods: (1) Source: the baseline model is trained only on the source data without any fine-tuning during the test process. (2) TTT [64]: it adapts the feature extractor by optimizing a self-supervised loss through a proxy task. However, it requires training the same proxy task on the source domain. (3) NORM [60]: this test-time normalization method updates the batch normalization [24] statistics using the mini-batch samples during the test process. (4) TENT [66]: it fine-tunes scale and bias parameters of the batch normalization layers using an entropy minimization loss during inference. (5) DUA [41]: it adapts the statistics of the batch normalization layer only on a tiny fraction of test data and augments a small batch of target data to adapt the model. + +# 4.3. Implementation Details + +Following the official implementations of TTT $^1$ , TENT $^2$ and DUA $^3$ , we use the ResNet-26 [18], Wide-ResNet-28-10 [71], and Wide-ResNet-40-2 [71] as the backbone networks for the CIFAR-10C dataset. We use the Wide-ResNet-40-2 network for the CIFAR-100C dataset. For the ImageNet-C dataset, we use the ResNet-18 [18] backbone network. We use the pre-trained model weights from the official implementations of TENT or DUA for all backbones based on the RobustBench protocol [9]. For the digit recognition transfer tasks, we use the pre-trained model weights of SVHN from the pytorch-playground [6]. For fair performance comparisons, all methods in each experimental condition share the same architecture and the same pre-trained model parameters. The batch size is set to 128. More implementation details are provided in the Supplemental Materials. + +1 TTT: https://github.com/yueatsprograms/ttt_cifar_release +2TENT: https://github.com/DequanWang/tent +$^{3}$ DUA: https://github.com/jmiemirza/DUA + +Table 1. Top-1 Classification Error (%) for each corruption in CIFAR-10C at the highest severity (Level 5). For TTT, TENT, and DUA, we use the ResNet-26 (top), WRN-28-10 (middle), and WRN-40-2 (bottom) from their official implementation. The smallest error is shown in bold. + +
MethodsgausshotimpuldefcsglsmtnzmsnwfrstfgbrtcntelspxjpgAvg.
Source67.763.169.955.356.642.250.131.646.339.117.174.634.257.931.749.2
TTT [64]45.641.850.021.846.123.023.929.930.025.112.223.922.647.227.231.4
NORM [60]44.643.749.129.445.226.226.925.827.923.818.334.329.337.032.532.9
TENT [66]39.438.847.919.945.023.220.628.132.124.516.126.732.430.635.530.7
DUA [41]34.932.642.218.740.224.018.423.924.020.912.327.127.226.228.726.8
Ours33.230.638.217.741.220.817.424.027.220.413.521.128.423.728.925.8
Source72.365.772.946.954.334.842.025.141.326.09.346.726.658.530.343.5
NORM [60]28.126.136.312.835.314.212.117.317.415.38.412.623.819.727.320.4
TENT [66]24.823.533.012.031.813.710.815.916.213.77.912.122.017.324.218.6
DUA [41]27.424.635.313.134.914.611.616.817.513.17.614.122.719.326.219.9
Ours23.621.430.911.031.113.010.914.215.513.08.010.321.816.722.417.6
Source28.822.926.29.520.610.69.314.215.317.57.620.914.741.314.718.3
NORM [60]18.716.422.39.122.110.59.713.013.215.47.812.016.415.117.614.6
TENT [66]15.713.218.87.918.19.08.010.410.812.46.710.014.011.414.812.1
DUA [41]15.413.417.38.018.09.17.710.810.812.16.610.913.613.014.312.1
Ours13.412.315.07.516.08.77.79.19.610.16.48.213.39.313.310.7
+ +Table 2. Top-1 Classification Error (%) for each corruption in CIFAR-100C at the highest severity (Level 5). + +
MethodsgausshotimpuldefcsglsmtnzmsnwfrstfgbrtcntelspxjpgAvg.
Source65.760.159.132.051.033.632.441.445.251.431.655.540.359.742.446.7
NORM [60]44.744.247.432.446.432.933.039.038.445.330.236.640.637.244.239.5
TENT [66]40.339.941.829.842.331.030.034.535.239.528.033.938.433.441.436.0
DUA [41]42.240.941.030.544.832.229.938.937.243.629.539.239.035.341.237.6
Ours38.437.136.228.441.029.329.732.233.136.126.430.936.230.838.333.6
+ +Table 3. Top-1 Classification Error (%) for each corruption in ImageNet-C at the highest severity (Level 5). + +
MethodsgausshotimpuldefcsglsmtnzmsnwfrstfgbrtcntelspxjpgAvg.
Source98.997.699.293.389.090.282.387.992.099.575.999.565.360.354.085.7
TTT [64]75.576.881.989.682.879.171.383.681.098.359.099.054.753.249.675.7
NORM [60]60.260.759.876.668.767.464.264.666.274.757.088.855.853.052.364.7
TENT [66]59.459.658.772.566.164.962.162.264.968.655.297.954.552.151.762.7
DUA [41]71.972.672.490.280.883.174.776.477.987.362.699.360.858.452.674.7
Ours56.756.756.673.365.761.062.058.663.363.953.177.554.052.051.560.4
+ +# 4.4. Performance Results + +The classification errors in the highest severity level corruption test datasets for test-time adaptation are reported in Tables 1, 2 and 3, with results of comparison methods directly cited from their original papers. Table 1 compares the classification error of our proposed method against recent test-time adaptation methods on the CIFAR-10C dataset. Our method performs better than other baselines with the three backbones including ResNet-26, WRN-28-10, and WRN-40-2, indicating the effectiveness of the proposed test-time adaptation method. Table 2 shows the performance comparison results on the CIFAR-100C dataset. Very encouraging results are also obtained on the large-size complicated ImageNet-C dataset, as shown in Table 3. The + +mean adaptation error of the full test dataset on Gaussian noise of CIFAR-10C in the different adaptation stages is shown in Figure 3. We can see that our method outperforms the NORM and TENT methods after the 5-th batch test-time adaptation. This implies our method can reduce error faster given few testing samples. Results for lower severity levels of corruption are provided in Supplemental Materials. + +Results for digit recognition from the SVHN to MNIST, MNIST-M, and USPS datasets are also reported in Table 4. All experiments use the same open-source source model. Note that the result for the TENT method is reproduced here since the pre-trained model for the TENT method was not provided and explained in its original paper. We can see that our method achieves the lowest average error when + +compared to other test-time adaptation methods. The performance improvement is quite impressive. + +![](images/351ec2ca9b4a63b9a2ffe5f6c30b6e0cc45ea46cd7716d3920917dca7f9d4a94.jpg) +Figure 3. The mean adaptation error of the full test dataset on Gaussian noise of CIFAR-10C in the different adaptation stages. + +Table 4. Top-1 Classification Error (\%) for test-time adaptation on digit recognition. The asterisk (*) means the implementation with a pytorch-playground [6] pre-trained source model by us. + +
MethodsMNISTMNIST-MUSPSAvg.
NORM* [60]39.652.141.444.4
TENT* [66]45.856.248.350.1
Ours31.247.932.637.2
+ +# 4.5. Further Performance Analysis + +# 4.5.1 Ablation Study + +We conduct the ablation study with test-time adaptation tasks on the ImageNet-C dataset to investigate the contribution of our method. Hebbian learning alone does not automatically lead to a perfect posterior estimation for $p_t(y|X_t)$ due to the lack of global information communication. From Table 5, we can see that feed-forward soft Hebbian Learning plays a significant effect and fine-tuning the feedback neuro-modulator of Block 1/2 also improves the performance. The average error is increased when expanding the neuro-modulator to more blocks. It is because optimizing more parameters with a mini-batch of testing samples is becoming more challenging. + +Table 5. Ablation study on ImageNet-C based on ResNet-18 including 4 Blocks at the highest severity (Level 5). + +
MethodsAvg. Error
Source (without adaptation)85.7
BP Conv185.0
Hebbian Conv167.2
Hebbian Conv1 + Neuromodulator (Block 1)60.5
Hebbian Conv1 + Neuromodulator (Block 1/2)60.4
Hebbian Conv1 + Neuromodulator (Block 1/2/3)61.2
Hebbian Conv1 + Neuromodulator (Block 1/2/3/4)64.8
+ +# 4.5.2 Feature Visualization + +Figure 4 compares the feature distributions on corrupted data obtained by different adaptation methods, including + +the source model with no adaptation, the TENT method, and our method. We also include the feature distribution for the supervised learning with the labeled sample, denoted by "Oracle". We can see that our method is able to learn features that are close to those obtained by supervised learning. + +![](images/4662bd51e0d1f89539a89c153c6e65f8fab52d42c3935928e4c945bfbf0aaa83.jpg) +(a) Source + +![](images/46ba2820524de6541c93edc9bc923e2a4a588acd681512486780c375117e928a.jpg) +(b) TENT + +![](images/afb646411263bd345abc7a06d45d4097e2d58a40dd72594f15e806f4404b2acb.jpg) +(c) Ours + +![](images/d1de06b416b12b764755c233d5d605f385f8d97d381fea748d1301443820f55e.jpg) +(d) Oracle +Figure 4. Density plots of test-time adapted features distribution on CIFAR-10-C with Gaussian noise (front) and reference features without corruption (back with yellow color). Each horizontal axis corresponds to one channel. The height of each ridge indicates the number of features that take the same value. + +# 5. Further Discussions + +Hebbian Learning by competition through lateral inhibition is a feed-forward process that has no gradient. If combining Hebbian Learning with back-propagation, it is limited to propagate the gradient through Hebbian layers to earlier gradient-based layers during training. Therefore, the Hebbian layers can only be designed in the early layers of models without back-propagation and gradient. In addition, although the Hebbian learning rule is commonly used for long-term reinforcement, the Hebb principle does not cover all forms of long-term synaptic plasticity. + +# 6. Conclusion + +In this work, we take inspiration from the biological neuromodulation process to construct a novel neuro-modulated Hebbian Learning (NHL) framework. With the unsupervised feed-forward soft Hebbian learning being combined with a learned neuro-modulator to capture feedback from external responses, the source model can be effectively adapted during the testing process. Experimental results on benchmark datasets demonstrate that our proposed method can significantly reduce the testing error for image classification with corruption, and reach new state-of-the-art performance. + +# References + +[1] Yoshua Bengio, Aaron Courville, and Pascal Vincent. Representation learning: A review and new perspectives. IEEE TPAMI, 35(8):1798-1828, 2013. 2 +[2] Tom Binzegger, Rodney J Douglas, and Kevan AC Martin. A quantitative map of the circuit of cat primary visual cortex. Journal of Neuroscience, 24(39):8441-8453, 2004. 4 +[3] Rafal Bogacz. A tutorial on the free-energy framework for modelling perception and learning. Journal of Mathematical Psychology, 76:198-211, 2017. 5 +[4] Malik Boudiaf, Romain Mueller, Ismail Ben Ayed, and Luca Bertinetto. Parameter-free online test-time adaptation. In CVPR, pages 8344-8353, 2022. 3 +[5] K Bougrova, N Bonacchi, JA Catarino, EE Dewitt, GT Meijer, P Dayan, ZF Mainen, et al. Comparison of neuromodulator signaling in a visual decision-making task. In Annual Meeting of the Society for Neuroscience, 2022. 6 +[6] Aaron Chen and Gus Smith. pytorch-playground. https://github.com/aaron-xichen/pytorch-playground, 2020. 6, 8 +[7] Zhixiang Chi, Yang Wang, Yuanhao Yu, and Jin Tang. Test-time fast adaptation for dynamic scene deblurring via meta-auxiliary learning. In CVPR, pages 9137-9146, 2021. 3 +[8] Sungha Choi, Seunghan Yang, Seokeon Choi, and Sungrack Yun. Improving test-time adaptation via shift-agnostic weight regularization and nearest source prototypes. arXiv preprint arXiv:2207.11707, 2022. 3 +[9] Francesco Croce, Maksym Andriushchenko, Vikash Sehwag, Edoardo Debenedetti, Nicolas Flammarion, Mung Chiang, Prateek Mittal, and Matthias Hein. Robustbench: a standardized adversarial robustness benchmark. In NeurIPS, 2021. 6 +[10] Ning Ding, Yixing Xu, Yehui Tang, Chao Xu, Yunhe Wang, and Dacheng Tao. Source-free domain adaptation via distribution estimation. In CVPR, pages 7212-7222, 2022. 3 +[11] Zhengming Ding and Yun Fu. Deep domain generalization with structured low-rank constraint. IEEE TIP, 27(1):304-313, 2017. 3 +[12] Rodney J Douglas, Kevan AC Martin, et al. Neuronal circuits of the neocortex. Annual Review of Neuroscience, 27(1):419-451, 2004. 4 +[13] Karl Friston. Hierarchical models in the brain. PLoS Computational Biology, 4(11):e1000211, 2008. 5 +[14] Karl Friston. The free-energy principle: a rough guide to the brain? Trends in Cognitive Sciences, 13(7):293-301, 2009. 3, 4, 5 +[15] Karl Friston, James Kilner, and Lee Harrison. A free energy principle for the brain. Journal of Physiology, Paris, 100(1-3):70-87, 2006. 2, 3, 4 +[16] Yaroslav Ganin and Victor Lempitsky. Unsupervised domain adaptation by backpropagation. In ICML, pages 1180-1189. PMLR, 2015. 1, 6 +[17] Wulfram Gerstner, Marco Lehmann, Vasiliki Liakoni, Dane Corneil, and Johann Brea. Eligibility traces and plasticity on behavioral time scales: experimental support of neohoebbian three-factor learning rules. Frontiers in Neural Circuits, 12:53, 2018. 6 + +[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 6 +[19] Donald Olding Hebb. The organization of behavior: A neuropsychological theory. Psychology Press, 2005. 4 +[20] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. In ICLR, 2018. 6 +[21] Minhao Hu, Tao Song, Yujun Gu, Xiangde Luo, Jieneng Chen, Yinan Chen, Ya Zhang, and Shaoting Zhang. Fully test-time adaptation for image segmentation. In MICCAI, pages 251-260. Springer, 2021. 3 +[22] Jiaxing Huang, Dayan Guan, Aoran Xiao, and Shijian Lu. Model adaptation: Historical contrastive learning for unsupervised domain adaptation without source data. In NeurIPS, volume 34, pages 3635-3649, 2021. 3 +[23] Jonathan J. Hull. A database for handwritten text recognition research. IEEE TPAMI, 16(5):550-554, 1994. 6 +[24] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In ICML, pages 448-456. PMLR, 2015. 6 +[25] Yusuke Iwasawa and Yutaka Matsuo. Test-time classifier adjustment module for model-agnostic domain generalization. In NeurIPS, volume 34, pages 2427-2440, 2021. 3 +[26] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. Master's thesis, University of Tront, 2009. 6 +[27] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Communications of the ACM, 60(6):84-90, 2017. 1 +[28] Dmitry Krotov and John J Hopfield. Unsupervised learning by competing hidden units. Proceedings of the National Academy of Sciences, 116(16):7723-7731, 2019. 2, 4, 5 +[29] Gabriele Lagani, Fabrizio Falchi, Claudio Gennaro, and Giuseppe Amato. Hebbian semi-supervised learning in a sample efficiency setting. *Neural Networks*, 143:719–731, 2021. 2 +[30] Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Deep learning. Nature, 521(7553):436-444, 2015. 4 +[31] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 6 +[32] Andrew B Lehr, Jannik Luboeinski, and Christian Tetzlaff. Neuromodulator-dependent synaptic tagging and capture retroactively controls neural coding in spiking neural networks. Scientific Reports, 12(1):1-18, 2022. 6 +[33] Rui Li, Qianfen Jiao, Wenming Cao, Hau-San Wong, and Si Wu. Model adaptation: Unsupervised domain adaptation without source data. In CVPR, pages 9641-9650, 2020. 1, 3 +[34] Yizhuo Li, Miao Hao, Zonglin Di, Nitesh Bharadwaj Gundavarapu, and Xiaolong Wang. Test-time personalization with a transformer for human pose estimation. In NeurIPS, volume 34, pages 2583-2597, 2021. 3 +[35] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In ICML, pages 6028-6039. PMLR, 2020. 1, 3 + +[36] Timothy P Lillicrap, Adam Santoro, Luke Marris, Colin J Akerman, and Geoffrey Hinton. Backpropagation and the brain. Nature Reviews Neuroscience, 21(6):335-346, 2020. 6 +[37] Yuejiang Liu, Parth Kothari, Bastien van Delft, Baptiste Bellot-Gurlet, Taylor Mordan, and Alexandre Alahi. Ttt++: When does self-supervised test-time training fail or thrive? In NeurIPS, volume 34, pages 21808-21820, 2021. 1, 3 +[38] Mingsheng Long, Yue Cao, Jianmin Wang, and Michael Jordan. Learning transferable features with deep adaptation networks. In ICML, pages 97-105. PMLR, 2015. 1 +[39] Robert A Marsden, Mario Döbler, and Bin Yang. Gradual test-time adaptation by self-training and style transfer. arXiv preprint arXiv:2208.07736, 2022. 3 +[40] Muhammad Jehanzeb Mirza, Cornelius Buerkle, Julio Jarquin, Michael Opitz, Fabian Oboril, Kay-Ulrich Scholl, and Horst Bischof. Robustness of object detectors in degrading weather conditions. In ITSC, pages 2719-2724. IEEE, 2021. 1 +[41] M Jehanzeb Mirza, Jakub Micorek, Horst Possegger, and Horst Bischof. The norm must go on: Dynamic unsupervised domain adaptation by normalization. In CVPR, pages 14765-14775, 2022. 1, 3, 6, 7 +[42] Timoleon Moraitis, Dmitry Toichkin, Adrien Journé, Yansong Chua, and Qinghai Guo. Softhebb: Bayesian inference in unsupervised hebbian soft winner-take-all networks. *Neuromorphic Computing and Engineering*, 2(4):044017, 2022. 5 +[43] Saeid Motiian, Marco Piccirilli, Donald A Adjeroh, and Gianfranco Doretto. Unified deep supervised domain adaptation and generalization. In ICCV, pages 5715-5725, 2017. 3 +[44] Krikamol Muandet, David Balduzzi, and Bernhard Schölkopf. Domain generalization via invariant feature representation. In ICML, pages 10-18. PMLR, 2013. 3 +[45] Bernhard Nessler, Michael Pfeiffer, Lars Buesing, and Wolfgang Maass. Bayesian computation emerges in generic cortical microcircuits through spike-timing-dependent plasticity. PLoS Computational Biology, 9(4):e1003037, 2013. 5 +[46] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y Ng. Reading digits in natural images with unsupervised feature learning. 2011. 6 +[47] Erkki Oja. Simplified neuron model as a principal component analyzer. Journal of Mathematical Biology, 15(3):267-273, 1982. 4, 5 +[48] John O'Donnell, Douglas Zeppenfeld, Evan McConnell, Salvador Pena, and Maiken Nedergaard. Norepinephrine: a neuromodulator that boosts the function of multiple cell types to optimize cns performance. Neurochemical Research, 37(11):2496-2512, 2012. 6 +[49] Prashant Pandey, Mrigank Raman, Sumanth Varambally, and Prathosh Ap. Generalization on unseen domains via inference-time label-preserving target projections. In CVPR, pages 12924-12933, 2021. 3 +[50] Zhongyi Pei, Zhangjie Cao, Mingsheng Long, and Jianmin Wang. Multi-adversarial domain adaptation. In AAAI, volume 32, 2018. 1 + +[51] Marina R Picciotto, Michael J Higley, and Yann S Mineur. Acetylcholine as a neuromodulator: cholinergic signaling shapes nervous system function and behavior. *Neuron*, 76(1):116-129, 2012. 6 +[52] Roman Pogodin and Peter Latham. Kernelized information bottleneck leads to biologically plausible 3-factor hebbian learning in deep networks. In NeurIPS, volume 33, pages 7296-7307, 2020. 2, 4 +[53] Fengchun Qiao, Long Zhao, and Xi Peng. Learning to learn single domain generalization. In CVPR, pages 12556-12565, 2020. 3 +[54] Zhen Qiu, Yifan Zhang, Hongbin Lin, Shuaicheng Niu, Yanxia Liu, Qing Du, and Mingkui Tan. Source-free domain adaptation via avatar prototype generation and adaptation. In IJCAI, pages 2921-2927, 2021. 3 +[55] Joaquin Quinonero-Candela, Masashi Sugiyama, Anton Schwaighofer, and Neil D Lawrence. Dataset shift in machine learning. MIT Press, 2008. 1 +[56] Rajesh PN Rao and Dana H Ballard. Predictive coding in the visual cortex: a functional interpretation of some extra-classical receptive-field effects. Nature Neuroscience, 2(1):79-87, 1999. 2, 4, 5 +[57] David E Rumelhart and David Zipser. Feature discovery by competitive learning. Cognitive Science, 9(1):75-112, 1985. 4 +[58] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. IJCV, 115(3):211-252, 2015. 6 +[59] Ueli Rutishauser, Rodney J Douglas, and Jean-Jacques Slotine. Collective stability of networks of winner-take-all circuits. Neural Computation, 23(3):735-773, 2011. 4 +[60] Steffen Schneider, Evgenia Rusak, Luisa Eck, Oliver Bringmann, Wieland Brendel, and Matthias Bethge. Improving robustness against common corruptions by covariate shift adaptation. In NeurIPS, volume 33, pages 11539-11551, 2020. 6, 7, 8 +[61] Seonguk Seo, Yumin Suh, Dongwan Kim, Geeho Kim, Jongwoo Han, and Bohyung Han. Learning to optimize domain specific normalization for domain generalization. In ECCV, pages 68-83. Springer, 2020. 3 +[62] Assaf Shocher, Nadav Cohen, and Michal Irani. "zero-shot" super-resolution using deep internal learning. In CVPR, pages 3118-3126, 2018. 3 +[63] Luisa Speranza, Umberto di Porzio, Davide Viggiano, Antonio de Donato, and Floriana Volpicelli. Dopamine: The neuromodulator of long-term synaptic plasticity, reward and movement control. Cells, 10(4):735, 2021. 6 +[64] Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei A. Efros, and Moritz Hardt. Test-time training with self-supervision for generalization under distribution shifts. In ICML, volume 119, pages 9229-9248. PMLR, 2020. 1, 3, 6, 7 +[65] Riccardo Volpi and Vittorio Murino. Addressing model vulnerability to distributional shifts over image transformation sets. In ICCV, pages 7980-7989, 2019. 3 + +[66] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In ICLR, 2020. 1, 3, 6, 7, 8 +[67] Fan Wang, Zhongyi Han, Yongshun Gong, and Yilong Yin. Exploring domain-invariant parameters for source free domain adaptation. In CVPR, pages 7151-7160, June 2022. 1, 3 +[68] Qin Wang, Olga Fink, Luc Van Gool, and Dengxin Dai. Continual test-time domain adaptation. In CVPR, pages 7201-7211, 2022. 3 +[69] James CR Whittington and Rafal Bogacz. Theories of error back-propagation in the brain. Trends in cognitive sciences, 23(3):235-250, 2019. 2 +[70] Shiqi Yang, Yaxing Wang, Joost van de Weijer, Luis Herranz, and Shangling Jui. Generalized source-free domain adaptation. In ICCV, pages 8978-8987, October 2021. 3 +[71] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. In BMVC. British Machine Vision Association, 2016. 6 +[72] Matthew D Zeiler and Rob Fergus. Visualizing and understanding convolutional networks. In ECCV, pages 818-833. Springer, 2014. 2 +[73] Shanshan Zhao, Mingming Gong, Tongliang Liu, Huan Fu, and Dacheng Tao. Domain generalization via entropy regularization. In NeurIPS, volume 33, pages 16096-16107, 2020. 3 +[74] Kaiyang Zhou, Ziwei Liu, Yu Qiao, Tao Xiang, and Chen Change Loy. Domain generalization: A survey. IEEE TPAMI, 2022. 3 \ No newline at end of file diff --git a/2023/Neuro-Modulated Hebbian Learning for Fully Test-Time Adaptation/images.zip b/2023/Neuro-Modulated Hebbian Learning for Fully Test-Time Adaptation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..8ffc1445aed4441c74b524e8e7895a564ab167ae --- /dev/null +++ b/2023/Neuro-Modulated Hebbian Learning for Fully Test-Time Adaptation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e67860322bb62166192fd00718f9b4ea9d4152098e6885f8e6f8d7a30430b9ec +size 536496 diff --git a/2023/Neuro-Modulated Hebbian Learning for Fully Test-Time Adaptation/layout.json b/2023/Neuro-Modulated Hebbian Learning for Fully Test-Time Adaptation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5b638b9172bace425279f1fc0b063f13a5ad8eb4 --- /dev/null +++ b/2023/Neuro-Modulated Hebbian Learning for Fully Test-Time Adaptation/layout.json @@ -0,0 +1,9434 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 85, + 103, + 509, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 103, + 509, + 121 + ], + "spans": [ + { + "bbox": [ + 85, + 103, + 509, + 121 + ], + "type": "text", + "content": "Neuro-Modulated Hebbian Learning for Fully Test-Time Adaptation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 157, + 142, + 436, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 142, + 436, + 172 + ], + "spans": [ + { + "bbox": [ + 157, + 142, + 436, + 172 + ], + "type": "text", + "content": "Yushun Tang" + }, + { + "bbox": [ + 157, + 142, + 436, + 172 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 157, + 142, + 436, + 172 + ], + "type": "text", + "content": ", Ce Zhang" + }, + { + "bbox": [ + 157, + 142, + 436, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 157, + 142, + 436, + 172 + ], + "type": "text", + "content": ", Heng Xu" + }, + { + "bbox": [ + 157, + 142, + 436, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 157, + 142, + 436, + 172 + ], + "type": "text", + "content": ", Shuoshuo Chen" + }, + { + "bbox": [ + 157, + 142, + 436, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 157, + 142, + 436, + 172 + ], + "type": "text", + "content": ", Jie Cheng" + }, + { + "bbox": [ + 157, + 142, + 436, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 157, + 142, + 436, + 172 + ], + "type": "text", + "content": ", Luziwei Leng" + }, + { + "bbox": [ + 157, + 142, + 436, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 157, + 142, + 436, + 172 + ], + "type": "text", + "content": ", Qinghai Guo" + }, + { + "bbox": [ + 157, + 142, + 436, + 172 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 157, + 142, + 436, + 172 + ], + "type": "text", + "content": ", Zhihai He" + }, + { + "bbox": [ + 157, + 142, + 436, + 172 + ], + "type": "inline_equation", + "content": "^{1,3*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 172, + 539, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 172, + 539, + 185 + ], + "spans": [ + { + "bbox": [ + 55, + 172, + 539, + 185 + ], + "type": "text", + "content": "1Department of Electronic and Electrical Engineering, Southern University of Science and Technology, Shenzhen, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 100, + 186, + 493, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 186, + 493, + 199 + ], + "spans": [ + { + "bbox": [ + 100, + 186, + 493, + 199 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 100, + 186, + 493, + 199 + ], + "type": "text", + "content": "Advanced Computing and Storage Laboratory, Huawei Technologies Co., Ltd., Shenzhen, China" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 212, + 201, + 381, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 201, + 381, + 213 + ], + "spans": [ + { + "bbox": [ + 212, + 201, + 381, + 213 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 212, + 201, + 381, + 213 + ], + "type": "text", + "content": "Pengcheng Laboratory, Shenzhen, China" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 121, + 216, + 471, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 216, + 471, + 227 + ], + "spans": [ + { + "bbox": [ + 121, + 216, + 471, + 227 + ], + "type": "text", + "content": "{tangys2022, zhangc2019, xuh2022, chenss2021}@mail.sustech.edu.cn" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 112, + 229, + 478, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 112, + 229, + 478, + 241 + ], + "spans": [ + { + "bbox": [ + 112, + 229, + 478, + 241 + ], + "type": "text", + "content": "{chengjie8, lengluziwei, guoqinghai}@huawei.com, hezh@sustech.edu.cn" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 143, + 269, + 192, + 281 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 269, + 192, + 281 + ], + "spans": [ + { + "bbox": [ + 143, + 269, + 192, + 281 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 294, + 289, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 294, + 289, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 294, + 289, + 605 + ], + "type": "text", + "content": "Fully test-time adaptation aims to adapt the network model based on sequential analysis of input samples during the inference stage to address the cross-domain performance degradation problem of deep neural networks. We take inspiration from the biological plausibility learning where the neuron responses are tuned based on a local synapse-change procedure and activated by competitive lateral inhibition rules. Based on these feed-forward learning rules, we design a soft Hebbian learning process which provides an unsupervised and effective mechanism for online adaptation. We observe that the performance of this feed-forward Hebbian learning for fully test-time adaptation can be significantly improved by incorporating a feedback neuro-modulation layer. It is able to fine-tune the neuron responses based on the external feedback generated by the error back-propagation from the top inference layers. This leads to our proposed neuro-modulated Hebbian learning (NHL) method for fully test-time adaptation. With the unsupervised feed-forward soft Hebbian learning being combined with a learned neuro-modulator to capture feedback from external responses, the source model can be effectively adapted during the testing process. Experimental results on benchmark datasets demonstrate that our proposed method can significantly improve the adaptation performance of network models and outperforms existing state-of-the-art methods." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 628, + 128, + 639 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 628, + 128, + 639 + ], + "spans": [ + { + "bbox": [ + 47, + 628, + 128, + 639 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 647, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 647, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 647, + 287, + 696 + ], + "type": "text", + "content": "Although deep neural networks have achieved great success in various machine learning tasks, their performance tends to degrade significantly when there is data shift [27, 55] between the training data in the source do" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 270, + 545, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 270, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 306, + 270, + 545, + 353 + ], + "type": "text", + "content": "main and the testing data in the target domain [40]. To address the performance degradation problem, unsupervised domain adaptation (UDA) [16,38,50] has been proposed to fine-tune the model parameters with a large amount of unlabeled testing data in an unsupervised manner. Source-free UDA methods [33, 35, 67] aim to adapt the network model without the need to access the source-domain samples." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 354, + 547, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 354, + 547, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 547, + 605 + ], + "type": "text", + "content": "There are two major categories of source-free UDA methods. The first category needs to access the whole test dataset on the target domain to achieve their adaptation performance [35, 67]. Notice that, in many practical scenarios when we deploy the network model on client devices, the network model does not have access to the whole dataset in the target domain since collecting and constructing the test dataset on the client side is very costly. The second type of method, called fully test-time adaptation, only needs access to live streams of test samples [41, 64, 66], which is able to dynamically adapt the source model on the fly during the testing process. Existing methods for fully test-time adaptation mainly focus on constructing various loss functions to regulate the inference process and adapt the model based on error back-propagation. For example, the TENT method [66] updates the batch normalization module by minimizing an entropy loss. The TTT method [64] updates the feature extractor parameters according to a self-supervised loss on a proxy learning task. The TTT++ method [37] introduces a feature alignment strategy based on online moment matching." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 611, + 494, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 611, + 494, + 624 + ], + "spans": [ + { + "bbox": [ + 306, + 611, + 494, + 624 + ], + "type": "text", + "content": "1.1. Challenges in Fully Test-Time UDA" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 629, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 715 + ], + "type": "text", + "content": "We recognize that most of the domain variations, such as changes in the visual scenes and image transformations or corruptions, are early layers of features in the semantic hierarchy [66]. They can be effectively captured and modeled by lower layers of the network model. From the perspective of machine learning, early representations through the lower layer play an important role to capture the pos" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 703, + 141, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 703, + 141, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 703, + 141, + 713 + ], + "type": "text", + "content": "* Corresponding authors." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3728" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 179 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 179 + ], + "type": "text", + "content": "terior distribution of the underlying explanatory factors for the observed input [1]. For instance, in deep neural network models, the early layers of the network tend to respond to corners, edges, or colors. In contrast, deeper layers respond to more class-specific features [72]. In the corruption test-time adaptation scenario, the class-specific features are always the same because the testing datasets are the corruption of the training domain. However, the early layers of models can be failed due to corruption." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 180, + 289, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 180, + 289, + 348 + ], + "spans": [ + { + "bbox": [ + 46, + 180, + 289, + 348 + ], + "type": "text", + "content": "Therefore, the central challenge in fully test-time UDA lies in how to learn useful early layer representations of the test samples without supervision. Motivated by this observation, we propose to explore neurobiology-inspired Hebbian learning for effective early-layer representation learning and fully test-time adaptation. It has been recognized that the learning rule of supervised end-to-end deep neural network training using back-propagation and the learning rules of the early front-end neural processing in neurobiology are unrelated [28]. The responses of neurons in biological neural networks are tuned by local pre-synaptic and post-synaptic activity, along with global variables that measure task performance, rather than the specific activity of other neurons [69]." + } + ] + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 53, + 361, + 290, + 473 + ], + "blocks": [ + { + "bbox": [ + 53, + 361, + 290, + 473 + ], + "lines": [ + { + "bbox": [ + 53, + 361, + 290, + 473 + ], + "spans": [ + { + "bbox": [ + 53, + 361, + 290, + 473 + ], + "type": "image", + "image_path": "814346d0284596a998a03f03241a53d6d35566a32707afcdebc88afd959575bc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 487, + 288, + 510 + ], + "lines": [ + { + "bbox": [ + 46, + 487, + 288, + 510 + ], + "spans": [ + { + "bbox": [ + 46, + 487, + 288, + 510 + ], + "type": "text", + "content": "Figure 1. The feature map visualization after the first convolution layer obtained by different learning methods." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 527, + 157, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 527, + 157, + 541 + ], + "spans": [ + { + "bbox": [ + 47, + 527, + 157, + 541 + ], + "type": "text", + "content": "1.2. Hebbian Learning" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 289, + 715 + ], + "type": "text", + "content": "Hebbian learning aims to learn useful early layer representations without supervision based on local synaptic plasticity rules, which is able to generate early representations that are as good as those learned by end-to-end supervised training with back-propagation [28, 52]. Drastically different from the current error back-propagation methods which require pseudo-labels or loss functions from the top network layers, Hebbian learning is a pure feed-forward adaptation process and does not require feedback from the distant top network layers. The responses of neurons are tuned based on a local synapse-change procedure and activated by competitive lateral inhibition rules [28]. During the learning process, the strength of synapses undergoes local changes that are proportional to the activity of" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": "the pre-synaptic cell and dependent on the activity of the post-synaptic cell. It also introduces local lateral inhibition between neurons within a layer, where the synapses of hidden units with strong responses are pushed toward the patterns that drive them, while those with weaker responses are pushed away from these patterns." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 148, + 546, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 148, + 546, + 316 + ], + "spans": [ + { + "bbox": [ + 304, + 148, + 546, + 316 + ], + "type": "text", + "content": "Existing literature has shown that early representations learned by Hebbian learning are as well as backpropagation and even more robust in testing [28, 29, 52]. Figure 1 shows the feature maps learned by different methods. The first row shows the original image. The second row shows the images in the target domain with significant image corruption. The third row shows the feature maps learned by the network model trained in the source domain for these target-domain images. The fourth row shows the feature maps learned by our Hebbian learning method. The last row (\"oracle\") shows the feature maps learned with true labels. We can see that the unsupervised Hebbian learning is able to generate feature maps which are as good as those from supervised learning." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 338, + 405, + 353 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 338, + 405, + 353 + ], + "spans": [ + { + "bbox": [ + 306, + 338, + 405, + 353 + ], + "type": "text", + "content": "1.3. Our Major Idea" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 361, + 545, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 361, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 304, + 361, + 545, + 529 + ], + "type": "text", + "content": "In this work, we observe that Hebbian learning, although provides a new and effective approach for unsupervised learning of early layer representation of the image, when directly applied to the network model, is not able to achieve satisfactory performance in fully test-time adaptation. First, the original hard decision for competitive learning is not suitable for fully test-time adaptation. Second, the Hebbian learning does not have an effective mechanism to consider external feedback, especially the feedback from the top network layers. We observe that, biologically, the visual processing is realized through hierarchical models considering a bottom-up early representation learning for the sensory input, and a top-down feedback mechanism based on predictive coding [15, 56]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 534, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 547, + 713 + ], + "type": "text", + "content": "Motivated by this, in this work, we propose to develop a new approach, called neuro-modulated Hebbian learning (NHL), for fully test-time adaptation. We first incorporate a soft decision rule into the feed-forward Hebbian learning to improve its competitive learning. Second, we learn a neuro-modulator to capture feedback from external responses, which controls which type of feature is consolidated and further processed to minimize the predictive error. During inference, the source model is adapted by the proposed NHL rule for each mini-batch of testing samples during the inference process. Experimental results on benchmark datasets demonstrate that our proposed method can significantly improve the adaptation performance of network models and outperforms existing state-of-the-art methods." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3729" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 229, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 229, + 85 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 229, + 85 + ], + "type": "text", + "content": "1.4. Summary of Major Contributions" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 91, + 289, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 91, + 289, + 270 + ], + "spans": [ + { + "bbox": [ + 47, + 91, + 289, + 270 + ], + "type": "text", + "content": "To summarize, our major contributions include: (1) we identify that the major challenge in fully test-time adaptation lies in effective unsupervised learning of early layer representations, and explore neurobiology-inspired soft Hebbian learning for effective early layer representation learning and fully test-time adaptation. (2) We develop a new neuro-modulated Hebbian learning method which combines unsupervised feed-forward Hebbian learning of early layer representation with a learned neuro-modulator to capture feedback from external responses. We analyze the optimal property of the proposed NHL algorithm based on free-energy principles [14, 15]. (3) We evaluate our proposed NHL method on benchmark datasets for fully test-time adaptation, demonstrating its significant performance improvement over existing methods." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 282, + 134, + 295 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 282, + 134, + 295 + ], + "spans": [ + { + "bbox": [ + 47, + 282, + 134, + 295 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 303, + 287, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 303, + 287, + 340 + ], + "spans": [ + { + "bbox": [ + 47, + 303, + 287, + 340 + ], + "type": "text", + "content": "In this section, we review existing methods related to our work including test-time adaptation, source-free UDA, domain generalization, and unsupervised Hebbian learning." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 348, + 169, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 348, + 169, + 361 + ], + "spans": [ + { + "bbox": [ + 47, + 348, + 169, + 361 + ], + "type": "text", + "content": "2.1. Test-time Adaptation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 367, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 367, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 367, + 289, + 715 + ], + "type": "text", + "content": "Test-time adaptation aims to online adapt the trained model while testing the input samples in the target domain. Sun et al. [64] proposed Test-time Training (TTT) by optimizing a self-supervised loss through a proxy task on the source before adapting to the target domain. The method proposed in TTT++ [37] aims to minimize the distribution shift between the training and testing feature distributions by dynamically matching the moments online. It should be noted that this method requires specific training on the source data, which is not available in fully test-time adaptation scenarios [66]. TENT [66] fine-tuned the scale and bias parameters of the batch normalization layers using an entropy minimization loss during the inference process. DUA [41] adapted the statistics of the batch normalization layer only on a tiny fraction of test data and augmented a small batch of target data to adapt the model. Choi et al. [8] proposed a shift-agnostic weight regularization and an auxiliary task for the alignment between the source and target features. Note that this method requires the source data for computing the source prototypes. The continual test-time adaptation methods [39,68] consider online TTA where target data is continually changing during inference. Instead of using parameters of the pre-trained model, Boudiaf et al. [4] only adapted the model's output by optimizing an objective function based on Laplacian adjusted maximum-likelihood estimation. Besides image classification, test-time adaptation has been successfully applied in various machine learning tasks, such as scene deblurring [7], superresolution [62], human pose estimation [34], image seg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 72, + 474, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 474, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 474, + 84 + ], + "type": "text", + "content": "mentation [21], object detection [41], etc." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 92, + 545, + 105 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 92, + 545, + 105 + ], + "spans": [ + { + "bbox": [ + 306, + 92, + 545, + 105 + ], + "type": "text", + "content": "2.2. Source-free Unsupervised Domain Adaptation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 110, + 547, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 110, + 547, + 386 + ], + "spans": [ + { + "bbox": [ + 304, + 110, + 547, + 386 + ], + "type": "text", + "content": "Recently, source-free UDA has emerged as an important research topic. It aims to adapt the source models to unlabeled target domains without accessing the data from the source domain. Among them, SHOT [35] proposed to learn target-specific features based on an information maximization criteria and pseudo-label prediction. 3C-GAN [33] generated target-style images by training a collaborative class conditional GAN module and using a clustering-based regularization scheme. G-SFDA [70] proposed a domain-specific attention mechanism that selectively activates different feature channels for different domains. CPGA [54] proposed to generate avatar prototypes instead of images via contrastive learning. HCL [22] proposed a solution for addressing the lack of source data by introducing both instance-level and category-level historical contrastive learning. DIPE [67] focuses on exploring the domain-invariant parameters of the model, rather than trying to learn domain-invariant representations. SFDA-DE [10] aligned domains by estimating source class-conditioned feature distribution and minimizing a contrastive adaptation loss function. Existing source-free methods are offline. They need to analyze the whole test dataset and update the model for a number of adaptation epochs." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 393, + 438, + 405 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 393, + 438, + 405 + ], + "spans": [ + { + "bbox": [ + 306, + 393, + 438, + 405 + ], + "type": "text", + "content": "2.3. Domain Generalization" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 412, + 547, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 412, + 547, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 412, + 547, + 616 + ], + "type": "text", + "content": "Domain generalization aims to train a model only on the source data that are generalizable to unseen target domains [74]. Domain generalization methods typically aim to learn domain-invariant representations by aligning the distributions of the source domains [43, 44, 73]. Methods based on data augmentation [53, 65] for regularizing the training process have been studied. Ensemble learning methods [11, 61] generate an ensemble of different model weights from different partitions of the training data. Recently, inference-time optimization without training the network model has been studied for domain generalization. Pandey et al. [49] used a generative model to project the target data onto the source-feature manifold with labels being preserved by solving an optimization problem during the inference stage. Iwasawa et al. [25] proposed a backpropagation-free generalization method by computing distance to a pseudo-prototype representation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 623, + 481, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 623, + 481, + 635 + ], + "spans": [ + { + "bbox": [ + 306, + 623, + 481, + 635 + ], + "type": "text", + "content": "2.4. Unsupervised Hebbian Learning" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "text", + "content": "In traditional end-to-end training, back-propagation is usually used to update the weights of deep neural networks. It has been recognized that the learning rule of supervised end-to-end deep neural network training using back-propagation is much different from the learning rules of the early front-end neural processing in neurobiology. In" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "type": "text", + "content": "3730" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 360 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 360 + ], + "type": "text", + "content": "addition, supervised training of deep neural networks with back-propagation requires a large amount of labeled samples [30]. The tuning of neuron responses in biological neural networks is achieved through a physically local synapse-change procedure. The superficial cerebral cortex exhibits a common connectivity pattern in which neurons are activated through competition via lateral inhibition [2, 12]. This competition leads to the suppression of weakly activated neurons and the amplification of strongly activated ones, a phenomenon known as competitive learning or \"winnertakes-all (WTA)\" learning [57, 59]. Motivated by Hebb's idea [19], several biological plausibility learning rules have been proposed, where changes of the synapse strength depend only on the activities of the pre-synaptic and postsynaptic neurons. The Oja's rule [47] proposed a linear neuron model with constrained Hebbian synaptic modification and derived a new unconstrained learning method. Krotov et al. [28] proposed a family of biologically plausible learning rules that enable the learning of early representations that are comparable to those achieved by end-to-end supervised training with back-propagation. Pogodin et al. [52] presents a family of learning rules which use pre- and post-synaptic firing rates and a global teaching signal. They perform almost the same as the back-propagation method." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 369, + 105, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 369, + 105, + 380 + ], + "spans": [ + { + "bbox": [ + 47, + 369, + 105, + 380 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 389, + 287, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 389, + 287, + 413 + ], + "spans": [ + { + "bbox": [ + 47, + 389, + 287, + 413 + ], + "type": "text", + "content": "In this section, we present our method of neuromodulated Hebbian learning for fully test-time adaptation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 418, + 171, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 418, + 171, + 430 + ], + "spans": [ + { + "bbox": [ + 47, + 418, + 171, + 430 + ], + "type": "text", + "content": "3.1. Problem Formulation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 437, + 287, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 437, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 47, + 437, + 287, + 568 + ], + "type": "text", + "content": "Suppose that a model " + }, + { + "bbox": [ + 47, + 437, + 287, + 568 + ], + "type": "inline_equation", + "content": "q_{\\theta_s}(y|X_s)" + }, + { + "bbox": [ + 47, + 437, + 287, + 568 + ], + "type": "text", + "content": " with parameters " + }, + { + "bbox": [ + 47, + 437, + 287, + 568 + ], + "type": "inline_equation", + "content": "\\theta_s" + }, + { + "bbox": [ + 47, + 437, + 287, + 568 + ], + "type": "text", + "content": " has been successfully trained on the source data " + }, + { + "bbox": [ + 47, + 437, + 287, + 568 + ], + "type": "inline_equation", + "content": "\\{X_s\\}" + }, + { + "bbox": [ + 47, + 437, + 287, + 568 + ], + "type": "text", + "content": " with labels " + }, + { + "bbox": [ + 47, + 437, + 287, + 568 + ], + "type": "inline_equation", + "content": "\\{Y_s\\}" + }, + { + "bbox": [ + 47, + 437, + 287, + 568 + ], + "type": "text", + "content": " where the true distribution is " + }, + { + "bbox": [ + 47, + 437, + 287, + 568 + ], + "type": "inline_equation", + "content": "p_s(y|X_s)" + }, + { + "bbox": [ + 47, + 437, + 287, + 568 + ], + "type": "text", + "content": ". It consists of a feature extractor " + }, + { + "bbox": [ + 47, + 437, + 287, + 568 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 47, + 437, + 287, + 568 + ], + "type": "text", + "content": " and a classifier " + }, + { + "bbox": [ + 47, + 437, + 287, + 568 + ], + "type": "inline_equation", + "content": "\\mathbf{C}" + }, + { + "bbox": [ + 47, + 437, + 287, + 568 + ], + "type": "text", + "content": ". During fully test-time adaptation, we are given the target data " + }, + { + "bbox": [ + 47, + 437, + 287, + 568 + ], + "type": "inline_equation", + "content": "\\{X_t\\}" + }, + { + "bbox": [ + 47, + 437, + 287, + 568 + ], + "type": "text", + "content": " with unknown labels " + }, + { + "bbox": [ + 47, + 437, + 287, + 568 + ], + "type": "inline_equation", + "content": "\\{Y_t\\}" + }, + { + "bbox": [ + 47, + 437, + 287, + 568 + ], + "type": "text", + "content": ". Our goal is to adapt the trained model " + }, + { + "bbox": [ + 47, + 437, + 287, + 568 + ], + "type": "inline_equation", + "content": "q_{\\theta_s}" + }, + { + "bbox": [ + 47, + 437, + 287, + 568 + ], + "type": "text", + "content": " in an unsupervised manner during testing to approximate the true target distribution " + }, + { + "bbox": [ + 47, + 437, + 287, + 568 + ], + "type": "inline_equation", + "content": "p_t(y|X_t)" + }, + { + "bbox": [ + 47, + 437, + 287, + 568 + ], + "type": "text", + "content": ". Given a sequence of input sample batches " + }, + { + "bbox": [ + 47, + 437, + 287, + 568 + ], + "type": "inline_equation", + "content": "\\{B_1, B_2, \\dots, B_n\\}" + }, + { + "bbox": [ + 47, + 437, + 287, + 568 + ], + "type": "text", + "content": ", the " + }, + { + "bbox": [ + 47, + 437, + 287, + 568 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 437, + 287, + 568 + ], + "type": "text", + "content": "-th adaptation of the network model can only rely on the " + }, + { + "bbox": [ + 47, + 437, + 287, + 568 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 437, + 287, + 568 + ], + "type": "text", + "content": "-th batch of test samples " + }, + { + "bbox": [ + 47, + 437, + 287, + 568 + ], + "type": "inline_equation", + "content": "B_i" + }, + { + "bbox": [ + 47, + 437, + 287, + 568 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 569, + 287, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 569, + 287, + 676 + ], + "spans": [ + { + "bbox": [ + 47, + 569, + 287, + 676 + ], + "type": "text", + "content": "We consider this transfer learning on a new target domain as an active inference process. The pre-trained model " + }, + { + "bbox": [ + 47, + 569, + 287, + 676 + ], + "type": "inline_equation", + "content": "q_{\\theta_s}(y|X_s)" + }, + { + "bbox": [ + 47, + 569, + 287, + 676 + ], + "type": "text", + "content": " is considered as a prior. Thus, predicting the labels for target samples becomes a Bayesian posterior " + }, + { + "bbox": [ + 47, + 569, + 287, + 676 + ], + "type": "inline_equation", + "content": "q_{\\theta_1}(y|B_1)" + }, + { + "bbox": [ + 47, + 569, + 287, + 676 + ], + "type": "text", + "content": " for the first batch, given the prior model " + }, + { + "bbox": [ + 47, + 569, + 287, + 676 + ], + "type": "inline_equation", + "content": "q_{\\theta_s}(y|X_s)" + }, + { + "bbox": [ + 47, + 569, + 287, + 676 + ], + "type": "text", + "content": ". After the first adaptation batch, the prior model is updated to " + }, + { + "bbox": [ + 47, + 569, + 287, + 676 + ], + "type": "inline_equation", + "content": "q_{\\theta_1}(y|B_1)" + }, + { + "bbox": [ + 47, + 569, + 287, + 676 + ], + "type": "text", + "content": ". This Bayesian inference process is repeated for all subsequent batches and produces a final posterior estimation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 683, + 287, + 696 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 683, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 115, + 683, + 287, + 696 + ], + "type": "interline_equation", + "content": "q _ {\\theta , t} (y \\mid X _ {t}) = q _ {\\theta_ {n}} (y \\mid B _ {n}), \\tag {1}", + "image_path": "5855731c6942febe11a8e3874379ed0de0b2f90a4cfbc4f7f2615800891f26b7.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 701, + 203, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 203, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 203, + 714 + ], + "type": "text", + "content": "given the prior model " + }, + { + "bbox": [ + 47, + 701, + 203, + 714 + ], + "type": "inline_equation", + "content": "q_{\\theta_{n - 1}}(y|B_{n - 1})" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 312, + 68, + 543, + 230 + ], + "blocks": [ + { + "bbox": [ + 312, + 68, + 543, + 230 + ], + "lines": [ + { + "bbox": [ + 312, + 68, + 543, + 230 + ], + "spans": [ + { + "bbox": [ + 312, + 68, + 543, + 230 + ], + "type": "image", + "image_path": "731f0308d9923657f8a57da1de826732da748b5d978c09f8a9a18d6ed1a16429.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 238, + 545, + 304 + ], + "lines": [ + { + "bbox": [ + 305, + 238, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 305, + 238, + 545, + 304 + ], + "type": "text", + "content": "Figure 2. An overview of the proposed NHL method. During inference, the first convolution layer of the source model is fine-tuned by the Hebbian learning rule and the neuromodulator is further fine-tuned by the entropy loss before making a prediction given each mini-batch testing sample. The lock symbol means the classifier is fixed in the test-time adaptation process." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 312, + 474, + 324 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 312, + 474, + 324 + ], + "spans": [ + { + "bbox": [ + 306, + 312, + 474, + 324 + ], + "type": "text", + "content": "3.2. Overview of Our NHL Method" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 330, + 546, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 330, + 546, + 534 + ], + "spans": [ + { + "bbox": [ + 304, + 330, + 546, + 534 + ], + "type": "text", + "content": "As illustrated in Figure 2, our proposed neuro-modulated Hebbian learning consists of two major components: the feed-forward soft Hebbian learning layer and the neuromodulator. The soft Hebbian learning layer aims to learn useful early layer representations without supervision based on local synaptic plasticity and soft competitive learning rules. It is able to generate early representations which are as good as those learned by end-to-end supervised training with labeled samples and back-propagation. During our experiments, we find that this soft Hebbian learning layer can significantly improve the performance of the network model in the target domain. However, we recognize that the feedforward Hebbian learning only is not able to achieve competitive performance as the current state-of-the-art methods for fully test-time adaptation. We find that it lacks the capability to effectively respond to external stimulus, specifically the feedback from the network output layer." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 534, + 546, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 546, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 546, + 677 + ], + "type": "text", + "content": "To address this issue, we propose to design a neuromodulator layer, an intermediate layer or an interface between the soft Hebbian learning layer and the classifier module of the network. This neuro-modulator layer is updated using back-propagation with the entropy loss being evaluated at the network output. It serves as the bridge between two different learning approaches: the feed-forward Hebbian learning and the original error back-propagation. From the ablation studies summarized in Table 5, we can see that both algorithm components, the feed-forward soft Hebbian learning, and the neuro-modulator are contributing significantly to the overall performance improvement." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "text", + "content": "The proposed NHL method can be also formulated by the free-energy principle in cognitive science [14, 15, 56]. A hierarchical predictive coding model was proposed by Rao" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "3731" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 239 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 239 + ], + "type": "text", + "content": "and Ballard [56] to learn a hierarchical internal model for human perception by maximizing the posterior probability of generating the observed data. This can be realized by a concurrent process of prediction through a bottom-up feedforward generation process (such as our Hebbian learning layer) combined with a top-down feedback-based optimization process (such as our neuro-modulator). Specifically, given the sensory input " + }, + { + "bbox": [ + 46, + 72, + 289, + 239 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 72, + 289, + 239 + ], + "type": "text", + "content": ", assume " + }, + { + "bbox": [ + 46, + 72, + 289, + 239 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 72, + 289, + 239 + ], + "type": "text", + "content": " is generated by environmental causes " + }, + { + "bbox": [ + 46, + 72, + 289, + 239 + ], + "type": "inline_equation", + "content": "\\vartheta" + }, + { + "bbox": [ + 46, + 72, + 289, + 239 + ], + "type": "text", + "content": ", denoted as " + }, + { + "bbox": [ + 46, + 72, + 289, + 239 + ], + "type": "inline_equation", + "content": "p(x) = p(x, \\vartheta)" + }, + { + "bbox": [ + 46, + 72, + 289, + 239 + ], + "type": "text", + "content": ". The free-energy principle states that the brain encodes the recognition density over sensory causes [14]. Mathematically, it optimizes a generative probabilistic mixture as " + }, + { + "bbox": [ + 46, + 72, + 289, + 239 + ], + "type": "inline_equation", + "content": "q_{1}(x) := q(x, \\vartheta)" + }, + { + "bbox": [ + 46, + 72, + 289, + 239 + ], + "type": "text", + "content": ". It has been demonstrated that this process can be achieved by a Hebbian-like learning approach [42, 45]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 239, + 287, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 239, + 287, + 300 + ], + "spans": [ + { + "bbox": [ + 47, + 239, + 287, + 300 + ], + "type": "text", + "content": "On the other hand, to approximate the true distribution " + }, + { + "bbox": [ + 47, + 239, + 287, + 300 + ], + "type": "inline_equation", + "content": "p_t(y|X_t)" + }, + { + "bbox": [ + 47, + 239, + 287, + 300 + ], + "type": "text", + "content": " by a posterior approximation " + }, + { + "bbox": [ + 47, + 239, + 287, + 300 + ], + "type": "inline_equation", + "content": "q_2(y) \\coloneqq q_{\\theta,t}(y|X_t)" + }, + { + "bbox": [ + 47, + 239, + 287, + 300 + ], + "type": "text", + "content": ", one can consider the similarity between these two distributions measured by the following Kullback-Leibler (KL) divergence" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 308, + 287, + 334 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 308, + 287, + 334 + ], + "spans": [ + { + "bbox": [ + 61, + 308, + 287, + 334 + ], + "type": "interline_equation", + "content": "\\operatorname {K L} \\left[ q _ {2} (y) | | p _ {t} (y | X _ {t}) \\right] = \\int q _ {2} (y) \\log \\frac {q _ {2} (y)}{p _ {t} (y | X _ {t})} d y. \\tag {2}", + "image_path": "934fc56e6109d2fb36ea2a8c457521ba77e9d3013fa645774739b88c96854dc1.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 341, + 287, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 341, + 287, + 377 + ], + "spans": [ + { + "bbox": [ + 45, + 341, + 287, + 377 + ], + "type": "text", + "content": "According to the analysis in Bogacz [3], this minimization of KL-divergence can be converted to minimization of the free-energy " + }, + { + "bbox": [ + 45, + 341, + 287, + 377 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 45, + 341, + 287, + 377 + ], + "type": "text", + "content": " defined as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 103, + 385, + 287, + 411 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 103, + 385, + 287, + 411 + ], + "spans": [ + { + "bbox": [ + 103, + 385, + 287, + 411 + ], + "type": "interline_equation", + "content": "F = \\int q _ {2} (y) \\log \\frac {q _ {2} (y)}{p _ {t} \\left(X _ {t} , y\\right)} d y. \\tag {3}", + "image_path": "751947dbb5e3599187e8bb8758092f0ede11c8c6bf556f72fdb4e228258321b6.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 418, + 287, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 418, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 46, + 418, + 287, + 502 + ], + "type": "text", + "content": "The maximization of the free-energy " + }, + { + "bbox": [ + 46, + 418, + 287, + 502 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 46, + 418, + 287, + 502 + ], + "type": "text", + "content": " can be solved by an active inference process if we consider the adaptation for each batch of test samples as a decision-making step [13]. This leads to a feedback-based optimization, just like the neuro-modulation as proposed in our NHL method. In the following sections, we explain the two major components of our NHL method in more detail." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 510, + 248, + 523 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 510, + 248, + 523 + ], + "spans": [ + { + "bbox": [ + 47, + 510, + 248, + 523 + ], + "type": "text", + "content": "3.3. Feed-Forward Soft Hebbian Learning" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 528, + 287, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 528, + 287, + 647 + ], + "spans": [ + { + "bbox": [ + 46, + 528, + 287, + 647 + ], + "type": "text", + "content": "As discussed above, the early feed-forward layer aims to learn useful layer representations of the input without any supervision. An approximate solution to this problem can be reduced to finding the first principal component of the input data [47], known as Oja's rule and its variations. The Oja's rule is itself a variation of the Hebbian rule, plus a normalization condition. In Oja's rule, the plasticity defined for a synaptic weight " + }, + { + "bbox": [ + 46, + 528, + 287, + 647 + ], + "type": "inline_equation", + "content": "w_{ik}" + }, + { + "bbox": [ + 46, + 528, + 287, + 647 + ], + "type": "text", + "content": " which connects a pre-synaptic neuron " + }, + { + "bbox": [ + 46, + 528, + 287, + 647 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 528, + 287, + 647 + ], + "type": "text", + "content": " with input " + }, + { + "bbox": [ + 46, + 528, + 287, + 647 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 46, + 528, + 287, + 647 + ], + "type": "text", + "content": " to a post-synaptic neuron " + }, + { + "bbox": [ + 46, + 528, + 287, + 647 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 528, + 287, + 647 + ], + "type": "text", + "content": " with activation " + }, + { + "bbox": [ + 46, + 528, + 287, + 647 + ], + "type": "inline_equation", + "content": "y_{k}" + }, + { + "bbox": [ + 46, + 528, + 287, + 647 + ], + "type": "text", + "content": " is:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 111, + 647, + 287, + 661 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 647, + 287, + 661 + ], + "spans": [ + { + "bbox": [ + 111, + 647, + 287, + 661 + ], + "type": "interline_equation", + "content": "\\Delta w _ {i k} = \\eta y _ {k} \\left(x _ {i} - y _ {k} w _ {i k}\\right). \\tag {4}", + "image_path": "bf588612299398dee1f92a0eeef4b667b573d1e608b2993081762ecc1841d365.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": "These weights are updated based on a plasticity rule [28]. We recognize that this process lacks the ability to detect different features of the data by different hidden units in the network. To address this problem, we propose to identify a" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": "good set of weights, by considering a generative model to optimize the distribution similarity to the input data." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 96, + 547, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 96, + 547, + 156 + ], + "spans": [ + { + "bbox": [ + 305, + 96, + 547, + 156 + ], + "type": "text", + "content": "Specifically, we assume that the target-domain input samples " + }, + { + "bbox": [ + 305, + 96, + 547, + 156 + ], + "type": "inline_equation", + "content": "\\{X_{t}\\}" + }, + { + "bbox": [ + 305, + 96, + 547, + 156 + ], + "type": "text", + "content": " are generated by hidden causes " + }, + { + "bbox": [ + 305, + 96, + 547, + 156 + ], + "type": "inline_equation", + "content": "\\vartheta" + }, + { + "bbox": [ + 305, + 96, + 547, + 156 + ], + "type": "text", + "content": " with distribution " + }, + { + "bbox": [ + 305, + 96, + 547, + 156 + ], + "type": "inline_equation", + "content": "p_t(x)" + }, + { + "bbox": [ + 305, + 96, + 547, + 156 + ], + "type": "text", + "content": ". We define an approximation to " + }, + { + "bbox": [ + 305, + 96, + 547, + 156 + ], + "type": "inline_equation", + "content": "p_t(x)" + }, + { + "bbox": [ + 305, + 96, + 547, + 156 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 305, + 96, + 547, + 156 + ], + "type": "inline_equation", + "content": "q(x|\\boldsymbol{w})" + }, + { + "bbox": [ + 305, + 96, + 547, + 156 + ], + "type": "text", + "content": " conditioned on the weights " + }, + { + "bbox": [ + 305, + 96, + 547, + 156 + ], + "type": "inline_equation", + "content": "\\boldsymbol{w}" + }, + { + "bbox": [ + 305, + 96, + 547, + 156 + ], + "type": "text", + "content": ". Moreover, we use a mixture of exponential functions to define the probability:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 358, + 164, + 545, + 178 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 164, + 545, + 178 + ], + "spans": [ + { + "bbox": [ + 358, + 164, + 545, + 178 + ], + "type": "interline_equation", + "content": "q (x \\mid \\boldsymbol {w} _ {k}) = \\exp (\\langle \\boldsymbol {w} _ {k}, x \\rangle) / N, \\tag {5}", + "image_path": "5d7b5d7796e15bbd7eecb378586518ed3c06dea1178a8c25f008ff85fb88dff1.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 186, + 545, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 186, + 545, + 269 + ], + "spans": [ + { + "bbox": [ + 305, + 186, + 545, + 269 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 186, + 545, + 269 + ], + "type": "inline_equation", + "content": "\\pmb{w}_k" + }, + { + "bbox": [ + 305, + 186, + 545, + 269 + ], + "type": "text", + "content": " is the weight vector corresponds to the postsynaptic neuron " + }, + { + "bbox": [ + 305, + 186, + 545, + 269 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 305, + 186, + 545, + 269 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 305, + 186, + 545, + 269 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 305, + 186, + 545, + 269 + ], + "type": "text", + "content": " is a normalization factor to ensure that " + }, + { + "bbox": [ + 305, + 186, + 545, + 269 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 305, + 186, + 545, + 269 + ], + "type": "text", + "content": " is a probabilistic measure. In the above generative model, the objective function is given by the KL divergence " + }, + { + "bbox": [ + 305, + 186, + 545, + 269 + ], + "type": "inline_equation", + "content": "\\mathrm{KL}[p_t(x)|q(x|\\pmb{w})]" + }, + { + "bbox": [ + 305, + 186, + 545, + 269 + ], + "type": "text", + "content": ". It can be shown that the optimal parameter vector that optimizes the KL-divergence is proportional to the mean of the input distribution [45]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 270, + 547, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 270, + 547, + 329 + ], + "spans": [ + { + "bbox": [ + 305, + 270, + 547, + 329 + ], + "type": "text", + "content": "On the other hand, (5) corresponds to neural interpretation as the activation function with normalized weights " + }, + { + "bbox": [ + 305, + 270, + 547, + 329 + ], + "type": "inline_equation", + "content": "\\pmb{w}_k / R" + }, + { + "bbox": [ + 305, + 270, + 547, + 329 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 305, + 270, + 547, + 329 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 305, + 270, + 547, + 329 + ], + "type": "text", + "content": " is the norm of the weight vector " + }, + { + "bbox": [ + 305, + 270, + 547, + 329 + ], + "type": "inline_equation", + "content": "\\pmb{w}_k" + }, + { + "bbox": [ + 305, + 270, + 547, + 329 + ], + "type": "text", + "content": ". For the network layer with " + }, + { + "bbox": [ + 305, + 270, + 547, + 329 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 305, + 270, + 547, + 329 + ], + "type": "text", + "content": " neurons, we define the output of the " + }, + { + "bbox": [ + 305, + 270, + 547, + 329 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 305, + 270, + 547, + 329 + ], + "type": "text", + "content": "-th neuron to be:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 375, + 335, + 545, + 369 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 335, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 375, + 335, + 545, + 369 + ], + "type": "interline_equation", + "content": "y _ {k} = e ^ {\\frac {u _ {k}}{\\tau}} / \\left(\\sum_ {i = 1} ^ {K} e ^ {\\frac {u _ {i}}{\\tau}}\\right), \\tag {6}", + "image_path": "a6d782b1798b218b712fc8b14caf8f63bc254accaf38842351b7d24786612fef.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 376, + 545, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 376, + 545, + 412 + ], + "spans": [ + { + "bbox": [ + 305, + 376, + 545, + 412 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 376, + 545, + 412 + ], + "type": "inline_equation", + "content": "u_{k}" + }, + { + "bbox": [ + 305, + 376, + 545, + 412 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 305, + 376, + 545, + 412 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 305, + 376, + 545, + 412 + ], + "type": "text", + "content": "-th neuron's weighted input, i.e. " + }, + { + "bbox": [ + 305, + 376, + 545, + 412 + ], + "type": "inline_equation", + "content": "u_{k} = \\sum_{i} w_{ik} \\cdot x_{i}" + }, + { + "bbox": [ + 305, + 376, + 545, + 412 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 305, + 376, + 545, + 412 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 305, + 376, + 545, + 412 + ], + "type": "text", + "content": " is a temperature-scaling hyper-parameter. This leads to a new soft Hebbian plasticity rule:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 366, + 420, + 545, + 434 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 366, + 420, + 545, + 434 + ], + "spans": [ + { + "bbox": [ + 366, + 420, + 545, + 434 + ], + "type": "interline_equation", + "content": "\\Delta w _ {i k} = \\eta y _ {k} \\left(R x _ {i} - u _ {k} w _ {i k}\\right), \\tag {7}", + "image_path": "428a0c8f7d697395972454f1b575d1bacfc6b27c5279035d8dc0c7a7b1055e8d.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 441, + 545, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 441, + 545, + 500 + ], + "spans": [ + { + "bbox": [ + 305, + 441, + 545, + 500 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 441, + 545, + 500 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 305, + 441, + 545, + 500 + ], + "type": "text", + "content": " is the learning rate. It can be shown that, using the plasticity rule in (7) to update the weights, they can converge to the equilibrium which lies in the sphere of radius " + }, + { + "bbox": [ + 305, + 441, + 545, + 500 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 305, + 441, + 545, + 500 + ], + "type": "text", + "content": ". The derivative of the weight vector norm can be expressed as:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 506, + 545, + 555 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 506, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 317, + 506, + 545, + 555 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\frac {d \\| \\boldsymbol {w} _ {k} \\|}{d t} = 2 \\boldsymbol {w} _ {k} \\frac {d \\boldsymbol {w} _ {k}}{d t} = 2 \\boldsymbol {w} _ {k} \\eta y _ {k} (R \\boldsymbol {x} - u _ {k} \\boldsymbol {w} _ {k}) \\\\ = 2 \\eta \\boldsymbol {w} _ {k} y _ {k} \\left(R \\boldsymbol {x} - \\boldsymbol {w} _ {k} \\boldsymbol {x} \\boldsymbol {w} _ {k}\\right) = 2 \\eta u _ {k} y _ {k} \\left(R - \\| \\boldsymbol {w} _ {k} \\|\\right). \\tag {8} \\\\ \\end{array}", + "image_path": "5b00c3ae72e8c138dee93c45cde79515e5f65aecad8f2f7c81876d12c91ab4e3.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 556, + 545, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 556, + 545, + 615 + ], + "spans": [ + { + "bbox": [ + 305, + 556, + 545, + 615 + ], + "type": "text", + "content": "With this, we see that the norm of the weight vector convergences to a sphere of radius " + }, + { + "bbox": [ + 305, + 556, + 545, + 615 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 305, + 556, + 545, + 615 + ], + "type": "text", + "content": ". This is because the norm of the weights decreases if " + }, + { + "bbox": [ + 305, + 556, + 545, + 615 + ], + "type": "inline_equation", + "content": "\\| \\pmb{w}_k\\| > R" + }, + { + "bbox": [ + 305, + 556, + 545, + 615 + ], + "type": "text", + "content": " and increases if " + }, + { + "bbox": [ + 305, + 556, + 545, + 615 + ], + "type": "inline_equation", + "content": "\\| \\pmb{w}_k\\| < R" + }, + { + "bbox": [ + 305, + 556, + 545, + 615 + ], + "type": "text", + "content": ". More details are provided in the Supplemental Materials." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 623, + 468, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 623, + 468, + 635 + ], + "spans": [ + { + "bbox": [ + 306, + 623, + 468, + 635 + ], + "type": "text", + "content": "3.4. The Neuro-Modulation Layer" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 305, + 641, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 641, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 641, + 545, + 714 + ], + "type": "text", + "content": "From the ablation studies in Section 4, we can see that the above soft Hebbian learning alone does not automatically lead to a perfect posterior estimation for " + }, + { + "bbox": [ + 305, + 641, + 545, + 714 + ], + "type": "inline_equation", + "content": "p_t(y|X_t)" + }, + { + "bbox": [ + 305, + 641, + 545, + 714 + ], + "type": "text", + "content": ". It cannot achieve the state-of-the-art performance for fully test-time adaptation. This is because it does not have an effective mechanism to consider external feedback, especially" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3732" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 239 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 239 + ], + "type": "text", + "content": "the feedback from the top network layers. Our proposed solution is to incorporate one or more modulating layers to steer the updates of weights to the desired outcome [17,36]. This so-called neuro-modulator has been explored in neuroscience research [5,48,51,63]. Recent research [32] shows that the level of neuro-modulation may change the process of synaptic consolidation, thus ultimately controlling which type of information is consolidated in the upper neural network. Unlike the above neuromodulator-based learning, where the modulator factor is embedded inside the Hebbian rule, we consider such a neuromodulation process in a disentangled way derived from the free-energy principle. It serves as an interface between the feed-forward Hebbian layer and the top decision network." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 239, + 287, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 239, + 287, + 276 + ], + "spans": [ + { + "bbox": [ + 47, + 239, + 287, + 276 + ], + "type": "text", + "content": "As defined in (3), the problem of minimizing the KL-divergence for " + }, + { + "bbox": [ + 47, + 239, + 287, + 276 + ], + "type": "inline_equation", + "content": "q_{2}(y)" + }, + { + "bbox": [ + 47, + 239, + 287, + 276 + ], + "type": "text", + "content": " and its true posterior " + }, + { + "bbox": [ + 47, + 239, + 287, + 276 + ], + "type": "inline_equation", + "content": "p_t(y|X_t)" + }, + { + "bbox": [ + 47, + 239, + 287, + 276 + ], + "type": "text", + "content": " can be formulated based on the free-energy principle:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 83, + 285, + 287, + 298 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 285, + 287, + 298 + ], + "spans": [ + { + "bbox": [ + 83, + 285, + 287, + 298 + ], + "type": "interline_equation", + "content": "\\operatorname {K L} \\left[ q _ {2} (y) \\mid \\left| p _ {t} (y \\mid X _ {t}) \\right. \\right] = F + \\log P _ {t} \\left(X _ {t}\\right), \\tag {9}", + "image_path": "0ee28b7d60c6cd6b07755b53e5640f8037fa5aa565e447edf9eec6be1b6320d8.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 307, + 287, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 307, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 46, + 307, + 287, + 392 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 307, + 287, + 392 + ], + "type": "inline_equation", + "content": "P_{t}(X_{t}) \\coloneqq \\int q_{2}(y)p_{t}(X_{t})dy = p_{t}(X_{t})" + }, + { + "bbox": [ + 46, + 307, + 287, + 392 + ], + "type": "text", + "content": " is the normalization term. Note that this term does not depend on " + }, + { + "bbox": [ + 46, + 307, + 287, + 392 + ], + "type": "inline_equation", + "content": "q_{2}(y)" + }, + { + "bbox": [ + 46, + 307, + 287, + 392 + ], + "type": "text", + "content": ". Therefore, minimizing the KL-divergence is reduced to minimizing " + }, + { + "bbox": [ + 46, + 307, + 287, + 392 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 46, + 307, + 287, + 392 + ], + "type": "text", + "content": ". To this end, given a batch " + }, + { + "bbox": [ + 46, + 307, + 287, + 392 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 46, + 307, + 287, + 392 + ], + "type": "text", + "content": " of data in the target domain, we rewrite (use " + }, + { + "bbox": [ + 46, + 307, + 287, + 392 + ], + "type": "inline_equation", + "content": "p_t(y,B) = p_t(B|y)p_t(y)" + }, + { + "bbox": [ + 46, + 307, + 287, + 392 + ], + "type": "text", + "content": ") and decompose the free-energy " + }, + { + "bbox": [ + 46, + 307, + 287, + 392 + ], + "type": "inline_equation", + "content": "F_{B}" + }, + { + "bbox": [ + 46, + 307, + 287, + 392 + ], + "type": "text", + "content": " for current batch into the following two items:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 52, + 399, + 287, + 435 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 399, + 287, + 435 + ], + "spans": [ + { + "bbox": [ + 52, + 399, + 287, + 435 + ], + "type": "interline_equation", + "content": "F _ {B} = \\operatorname {K L} \\left[ q _ {2} (y | B) \\mid \\mid p _ {t} (y) \\right] - \\int_ {y} q _ {2} (y | B) \\log p _ {t} (B | y) d y. \\tag {10}", + "image_path": "0f3ddfa236b845ed59944b6887900fe77b1dd910c88141b94a7720c0af443cc8.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 436, + 287, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 436, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 46, + 436, + 287, + 509 + ], + "type": "text", + "content": "The first term in (10) is already minimized through soft Hebbian learning, while minimizing the second term requires the likelihood distribution " + }, + { + "bbox": [ + 46, + 436, + 287, + 509 + ], + "type": "inline_equation", + "content": "p_t(B|y)" + }, + { + "bbox": [ + 46, + 436, + 287, + 509 + ], + "type": "text", + "content": ". Since " + }, + { + "bbox": [ + 46, + 436, + 287, + 509 + ], + "type": "inline_equation", + "content": "p_t(B|y) = p_t(y|B)p_t(B) / p_t(y)" + }, + { + "bbox": [ + 46, + 436, + 287, + 509 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 436, + 287, + 509 + ], + "type": "inline_equation", + "content": "q_2(y|B)" + }, + { + "bbox": [ + 46, + 436, + 287, + 509 + ], + "type": "text", + "content": " is considered as an approximation of " + }, + { + "bbox": [ + 46, + 436, + 287, + 509 + ], + "type": "inline_equation", + "content": "p_t(y|B)" + }, + { + "bbox": [ + 46, + 436, + 287, + 509 + ], + "type": "text", + "content": ", we minimize the entropy of " + }, + { + "bbox": [ + 46, + 436, + 287, + 509 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 436, + 287, + 509 + ], + "type": "text", + "content": " given the current batch " + }, + { + "bbox": [ + 46, + 436, + 287, + 509 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 46, + 436, + 287, + 509 + ], + "type": "text", + "content": " in a discrete way as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 517, + 287, + 548 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 287, + 548 + ], + "type": "interline_equation", + "content": "\\underset {\\boldsymbol {w}} {\\arg \\min } H (y | B) = \\underset {\\boldsymbol {w}} {\\arg \\max } \\sum q _ {2} (y | B) \\log q _ {2} (y | B), \\tag {11}", + "image_path": "d827a225f81ff49084fd2c2d757b9697f7cd6e9f1360fe5be277b5d6e3495747.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 548, + 287, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 548, + 287, + 585 + ], + "spans": [ + { + "bbox": [ + 46, + 548, + 287, + 585 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 548, + 287, + 585 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 46, + 548, + 287, + 585 + ], + "type": "text", + "content": " are the weights in the layer implementing the neuro-modulator. As in existing deep neural network training, we can use gradient descent to optimize these weights." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 595, + 128, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 595, + 128, + 609 + ], + "spans": [ + { + "bbox": [ + 47, + 595, + 128, + 609 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 615, + 287, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 615, + 287, + 651 + ], + "spans": [ + { + "bbox": [ + 46, + 615, + 287, + 651 + ], + "type": "text", + "content": "In this section, we conduct experiments on multiple test-time adaptation benchmark datasets to evaluate the performance of our proposed NHL method." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 659, + 167, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 659, + 167, + 671 + ], + "spans": [ + { + "bbox": [ + 47, + 659, + 167, + 671 + ], + "type": "text", + "content": "4.1. Benchmark Datasets" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "content": "We evaluate our method among the following popular benchmark datasets for test-time adaptation. (1) CIFAR-10/100C. We choose CIFAR-10/100 [26] with 10/100" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 72, + 545, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 240 + ], + "type": "text", + "content": "classes, a source training set of 50,000, and a target testing set CIFAR-10/100C [20] of 10,000 for small-size image experiments at an accessible scale. (2) ImageNet-C. We choose the ImageNet [58] with 1,000 classes, a source training set of 1.2 million, and a validation set of 50,000 for large-size image experiments. It should be noted that we use a fixed target testing subset of the validation set with 5000 images on which all models are evaluated following the RobustBench protocol [9]. (3) SVHN " + }, + { + "bbox": [ + 304, + 72, + 545, + 240 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 72, + 545, + 240 + ], + "type": "text", + "content": " MNIST/MNIST-M/USPS. Following TENT for domain adaptation, we choose a SVHN [46] source model and transfer it to MNIST [31] (10,000 test samples), MNIST-M [16] (10,000 test samples) and USPS [23] (2,007 test samples), respectively." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 251, + 430, + 264 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 251, + 430, + 264 + ], + "spans": [ + { + "bbox": [ + 306, + 251, + 430, + 264 + ], + "type": "text", + "content": "4.2. Comparison Methods" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 270, + 545, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 270, + 545, + 450 + ], + "spans": [ + { + "bbox": [ + 304, + 270, + 545, + 450 + ], + "type": "text", + "content": "We compare our method against the following fully test-time adaptation methods: (1) Source: the baseline model is trained only on the source data without any fine-tuning during the test process. (2) TTT [64]: it adapts the feature extractor by optimizing a self-supervised loss through a proxy task. However, it requires training the same proxy task on the source domain. (3) NORM [60]: this test-time normalization method updates the batch normalization [24] statistics using the mini-batch samples during the test process. (4) TENT [66]: it fine-tunes scale and bias parameters of the batch normalization layers using an entropy minimization loss during inference. (5) DUA [41]: it adapts the statistics of the batch normalization layer only on a tiny fraction of test data and augments a small batch of target data to adapt the model." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 461, + 440, + 474 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 461, + 440, + 474 + ], + "spans": [ + { + "bbox": [ + 306, + 461, + 440, + 474 + ], + "type": "text", + "content": "4.3. Implementation Details" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 480, + 545, + 671 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 480, + 545, + 671 + ], + "spans": [ + { + "bbox": [ + 304, + 480, + 545, + 671 + ], + "type": "text", + "content": "Following the official implementations of TTT" + }, + { + "bbox": [ + 304, + 480, + 545, + 671 + ], + "type": "inline_equation", + "content": "^1" + }, + { + "bbox": [ + 304, + 480, + 545, + 671 + ], + "type": "text", + "content": ", TENT" + }, + { + "bbox": [ + 304, + 480, + 545, + 671 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 304, + 480, + 545, + 671 + ], + "type": "text", + "content": " and DUA" + }, + { + "bbox": [ + 304, + 480, + 545, + 671 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 304, + 480, + 545, + 671 + ], + "type": "text", + "content": ", we use the ResNet-26 [18], Wide-ResNet-28-10 [71], and Wide-ResNet-40-2 [71] as the backbone networks for the CIFAR-10C dataset. We use the Wide-ResNet-40-2 network for the CIFAR-100C dataset. For the ImageNet-C dataset, we use the ResNet-18 [18] backbone network. We use the pre-trained model weights from the official implementations of TENT or DUA for all backbones based on the RobustBench protocol [9]. For the digit recognition transfer tasks, we use the pre-trained model weights of SVHN from the pytorch-playground [6]. For fair performance comparisons, all methods in each experimental condition share the same architecture and the same pre-trained model parameters. The batch size is set to 128. More implementation details are provided in the Supplemental Materials." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 683, + 504, + 712 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 317, + 683, + 504, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 683, + 504, + 693 + ], + "spans": [ + { + "bbox": [ + 317, + 683, + 504, + 693 + ], + "type": "text", + "content": "1 TTT: https://github.com/yueatsprograms/ttt_cifar_release" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 693, + 465, + 703 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 693, + 465, + 703 + ], + "spans": [ + { + "bbox": [ + 317, + 693, + 465, + 703 + ], + "type": "text", + "content": "2TENT: https://github.com/DequanWang/tent" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 703, + 456, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 703, + 456, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 703, + 456, + 712 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 317, + 703, + 456, + 712 + ], + "type": "text", + "content": "DUA: https://github.com/jmiemirza/DUA" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3733" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 99, + 544, + 306 + ], + "blocks": [ + { + "bbox": [ + 47, + 64, + 545, + 96 + ], + "lines": [ + { + "bbox": [ + 47, + 64, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 47, + 64, + 545, + 96 + ], + "type": "text", + "content": "Table 1. Top-1 Classification Error (%) for each corruption in CIFAR-10C at the highest severity (Level 5). For TTT, TENT, and DUA, we use the ResNet-26 (top), WRN-28-10 (middle), and WRN-40-2 (bottom) from their official implementation. The smallest error is shown in bold." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 99, + 544, + 306 + ], + "lines": [ + { + "bbox": [ + 49, + 99, + 544, + 306 + ], + "spans": [ + { + "bbox": [ + 49, + 99, + 544, + 306 + ], + "type": "table", + "html": "
MethodsgausshotimpuldefcsglsmtnzmsnwfrstfgbrtcntelspxjpgAvg.
Source67.763.169.955.356.642.250.131.646.339.117.174.634.257.931.749.2
TTT [64]45.641.850.021.846.123.023.929.930.025.112.223.922.647.227.231.4
NORM [60]44.643.749.129.445.226.226.925.827.923.818.334.329.337.032.532.9
TENT [66]39.438.847.919.945.023.220.628.132.124.516.126.732.430.635.530.7
DUA [41]34.932.642.218.740.224.018.423.924.020.912.327.127.226.228.726.8
Ours33.230.638.217.741.220.817.424.027.220.413.521.128.423.728.925.8
Source72.365.772.946.954.334.842.025.141.326.09.346.726.658.530.343.5
NORM [60]28.126.136.312.835.314.212.117.317.415.38.412.623.819.727.320.4
TENT [66]24.823.533.012.031.813.710.815.916.213.77.912.122.017.324.218.6
DUA [41]27.424.635.313.134.914.611.616.817.513.17.614.122.719.326.219.9
Ours23.621.430.911.031.113.010.914.215.513.08.010.321.816.722.417.6
Source28.822.926.29.520.610.69.314.215.317.57.620.914.741.314.718.3
NORM [60]18.716.422.39.122.110.59.713.013.215.47.812.016.415.117.614.6
TENT [66]15.713.218.87.918.19.08.010.410.812.46.710.014.011.414.812.1
DUA [41]15.413.417.38.018.09.17.710.810.812.16.610.913.613.014.312.1
Ours13.412.315.07.516.08.77.79.19.610.16.48.213.39.313.310.7
", + "image_path": "ab102445d472969611d395fc2e014cbb0353ea965f7de0c6b25860d2ca17fe16.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 50, + 332, + 544, + 407 + ], + "blocks": [ + { + "bbox": [ + 99, + 319, + 493, + 331 + ], + "lines": [ + { + "bbox": [ + 99, + 319, + 493, + 331 + ], + "spans": [ + { + "bbox": [ + 99, + 319, + 493, + 331 + ], + "type": "text", + "content": "Table 2. Top-1 Classification Error (%) for each corruption in CIFAR-100C at the highest severity (Level 5)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 332, + 544, + 407 + ], + "lines": [ + { + "bbox": [ + 50, + 332, + 544, + 407 + ], + "spans": [ + { + "bbox": [ + 50, + 332, + 544, + 407 + ], + "type": "table", + "html": "
MethodsgausshotimpuldefcsglsmtnzmsnwfrstfgbrtcntelspxjpgAvg.
Source65.760.159.132.051.033.632.441.445.251.431.655.540.359.742.446.7
NORM [60]44.744.247.432.446.432.933.039.038.445.330.236.640.637.244.239.5
TENT [66]40.339.941.829.842.331.030.034.535.239.528.033.938.433.441.436.0
DUA [41]42.240.941.030.544.832.229.938.937.243.629.539.239.035.341.237.6
Ours38.437.136.228.441.029.329.732.233.136.126.430.936.230.838.333.6
", + "image_path": "f64dc1dac946b25ae959e3572a75adf705a4daae0cb836bd571c5dfc5b4e288a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 50, + 427, + 544, + 515 + ], + "blocks": [ + { + "bbox": [ + 101, + 415, + 492, + 426 + ], + "lines": [ + { + "bbox": [ + 101, + 415, + 492, + 426 + ], + "spans": [ + { + "bbox": [ + 101, + 415, + 492, + 426 + ], + "type": "text", + "content": "Table 3. Top-1 Classification Error (%) for each corruption in ImageNet-C at the highest severity (Level 5)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 427, + 544, + 515 + ], + "lines": [ + { + "bbox": [ + 50, + 427, + 544, + 515 + ], + "spans": [ + { + "bbox": [ + 50, + 427, + 544, + 515 + ], + "type": "table", + "html": "
MethodsgausshotimpuldefcsglsmtnzmsnwfrstfgbrtcntelspxjpgAvg.
Source98.997.699.293.389.090.282.387.992.099.575.999.565.360.354.085.7
TTT [64]75.576.881.989.682.879.171.383.681.098.359.099.054.753.249.675.7
NORM [60]60.260.759.876.668.767.464.264.666.274.757.088.855.853.052.364.7
TENT [66]59.459.658.772.566.164.962.162.264.968.655.297.954.552.151.762.7
DUA [41]71.972.672.490.280.883.174.776.477.987.362.699.360.858.452.674.7
Ours56.756.756.673.365.761.062.058.663.363.953.177.554.052.051.560.4
", + "image_path": "ae4565f114bdb57d9036ce8fd7731db7b835f4a1f9faa8ceeedcd4f647b6ece3.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 535, + 167, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 535, + 167, + 547 + ], + "spans": [ + { + "bbox": [ + 47, + 535, + 167, + 547 + ], + "type": "text", + "content": "4.4. Performance Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 557, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 288, + 713 + ], + "type": "text", + "content": "The classification errors in the highest severity level corruption test datasets for test-time adaptation are reported in Tables 1, 2 and 3, with results of comparison methods directly cited from their original papers. Table 1 compares the classification error of our proposed method against recent test-time adaptation methods on the CIFAR-10C dataset. Our method performs better than other baselines with the three backbones including ResNet-26, WRN-28-10, and WRN-40-2, indicating the effectiveness of the proposed test-time adaptation method. Table 2 shows the performance comparison results on the CIFAR-100C dataset. Very encouraging results are also obtained on the large-size complicated ImageNet-C dataset, as shown in Table 3. The" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 536, + 545, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 536, + 545, + 620 + ], + "spans": [ + { + "bbox": [ + 304, + 536, + 545, + 620 + ], + "type": "text", + "content": "mean adaptation error of the full test dataset on Gaussian noise of CIFAR-10C in the different adaptation stages is shown in Figure 3. We can see that our method outperforms the NORM and TENT methods after the 5-th batch test-time adaptation. This implies our method can reduce error faster given few testing samples. Results for lower severity levels of corruption are provided in Supplemental Materials." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 629, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 546, + 713 + ], + "type": "text", + "content": "Results for digit recognition from the SVHN to MNIST, MNIST-M, and USPS datasets are also reported in Table 4. All experiments use the same open-source source model. Note that the result for the TENT method is reproduced here since the pre-trained model for the TENT method was not provided and explained in its original paper. We can see that our method achieves the lowest average error when" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "3734" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "content": "compared to other test-time adaptation methods. The performance improvement is quite impressive." + } + ] + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 72, + 106, + 260, + 217 + ], + "blocks": [ + { + "bbox": [ + 72, + 106, + 260, + 217 + ], + "lines": [ + { + "bbox": [ + 72, + 106, + 260, + 217 + ], + "spans": [ + { + "bbox": [ + 72, + 106, + 260, + 217 + ], + "type": "image", + "image_path": "351ec2ca9b4a63b9a2ffe5f6c30b6e0cc45ea46cd7716d3920917dca7f9d4a94.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 220, + 286, + 243 + ], + "lines": [ + { + "bbox": [ + 47, + 220, + 286, + 243 + ], + "spans": [ + { + "bbox": [ + 47, + 220, + 286, + 243 + ], + "type": "text", + "content": "Figure 3. The mean adaptation error of the full test dataset on Gaussian noise of CIFAR-10C in the different adaptation stages." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 57, + 284, + 278, + 340 + ], + "blocks": [ + { + "bbox": [ + 47, + 250, + 286, + 283 + ], + "lines": [ + { + "bbox": [ + 47, + 250, + 286, + 283 + ], + "spans": [ + { + "bbox": [ + 47, + 250, + 286, + 283 + ], + "type": "text", + "content": "Table 4. Top-1 Classification Error (\\%) for test-time adaptation on digit recognition. The asterisk (*) means the implementation with a pytorch-playground [6] pre-trained source model by us." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 284, + 278, + 340 + ], + "lines": [ + { + "bbox": [ + 57, + 284, + 278, + 340 + ], + "spans": [ + { + "bbox": [ + 57, + 284, + 278, + 340 + ], + "type": "table", + "html": "
MethodsMNISTMNIST-MUSPSAvg.
NORM* [60]39.652.141.444.4
TENT* [66]45.856.248.350.1
Ours31.247.932.637.2
", + "image_path": "36d9e9a7c0904e38e8ff74e7c6261842971da2c454228a918bb8f77c71d4a731.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 350, + 212, + 363 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 350, + 212, + 363 + ], + "spans": [ + { + "bbox": [ + 47, + 350, + 212, + 363 + ], + "type": "text", + "content": "4.5. Further Performance Analysis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 369, + 144, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 369, + 144, + 380 + ], + "spans": [ + { + "bbox": [ + 47, + 369, + 144, + 380 + ], + "type": "text", + "content": "4.5.1 Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 388, + 286, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 388, + 286, + 532 + ], + "spans": [ + { + "bbox": [ + 46, + 388, + 286, + 532 + ], + "type": "text", + "content": "We conduct the ablation study with test-time adaptation tasks on the ImageNet-C dataset to investigate the contribution of our method. Hebbian learning alone does not automatically lead to a perfect posterior estimation for " + }, + { + "bbox": [ + 46, + 388, + 286, + 532 + ], + "type": "inline_equation", + "content": "p_t(y|X_t)" + }, + { + "bbox": [ + 46, + 388, + 286, + 532 + ], + "type": "text", + "content": " due to the lack of global information communication. From Table 5, we can see that feed-forward soft Hebbian Learning plays a significant effect and fine-tuning the feedback neuro-modulator of Block 1/2 also improves the performance. The average error is increased when expanding the neuro-modulator to more blocks. It is because optimizing more parameters with a mini-batch of testing samples is becoming more challenging." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 51, + 559, + 284, + 652 + ], + "blocks": [ + { + "bbox": [ + 47, + 534, + 286, + 556 + ], + "lines": [ + { + "bbox": [ + 47, + 534, + 286, + 556 + ], + "spans": [ + { + "bbox": [ + 47, + 534, + 286, + 556 + ], + "type": "text", + "content": "Table 5. Ablation study on ImageNet-C based on ResNet-18 including 4 Blocks at the highest severity (Level 5)." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 51, + 559, + 284, + 652 + ], + "lines": [ + { + "bbox": [ + 51, + 559, + 284, + 652 + ], + "spans": [ + { + "bbox": [ + 51, + 559, + 284, + 652 + ], + "type": "table", + "html": "
MethodsAvg. Error
Source (without adaptation)85.7
BP Conv185.0
Hebbian Conv167.2
Hebbian Conv1 + Neuromodulator (Block 1)60.5
Hebbian Conv1 + Neuromodulator (Block 1/2)60.4
Hebbian Conv1 + Neuromodulator (Block 1/2/3)61.2
Hebbian Conv1 + Neuromodulator (Block 1/2/3/4)64.8
", + "image_path": "31ebc6fbf6531f648fafd57d35492dd4959df2fe4135baab2f9601125096861f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 670, + 171, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 171, + 681 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 171, + 681 + ], + "type": "text", + "content": "4.5.2 Feature Visualization" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "type": "text", + "content": "Figure 4 compares the feature distributions on corrupted data obtained by different adaptation methods, including" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "text", + "content": "the source model with no adaptation, the TENT method, and our method. We also include the feature distribution for the supervised learning with the labeled sample, denoted by \"Oracle\". We can see that our method is able to learn features that are close to those obtained by supervised learning." + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 323, + 143, + 419, + 217 + ], + "blocks": [ + { + "bbox": [ + 323, + 143, + 419, + 217 + ], + "lines": [ + { + "bbox": [ + 323, + 143, + 419, + 217 + ], + "spans": [ + { + "bbox": [ + 323, + 143, + 419, + 217 + ], + "type": "image", + "image_path": "4662bd51e0d1f89539a89c153c6e65f8fab52d42c3935928e4c945bfbf0aaa83.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 353, + 218, + 388, + 226 + ], + "lines": [ + { + "bbox": [ + 353, + 218, + 388, + 226 + ], + "spans": [ + { + "bbox": [ + 353, + 218, + 388, + 226 + ], + "type": "text", + "content": "(a) Source" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 434, + 144, + 529, + 217 + ], + "blocks": [ + { + "bbox": [ + 434, + 144, + 529, + 217 + ], + "lines": [ + { + "bbox": [ + 434, + 144, + 529, + 217 + ], + "spans": [ + { + "bbox": [ + 434, + 144, + 529, + 217 + ], + "type": "image", + "image_path": "46ba2820524de6541c93edc9bc923e2a4a588acd681512486780c375117e928a.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 465, + 218, + 498, + 226 + ], + "lines": [ + { + "bbox": [ + 465, + 218, + 498, + 226 + ], + "spans": [ + { + "bbox": [ + 465, + 218, + 498, + 226 + ], + "type": "text", + "content": "(b) TENT" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 322, + 228, + 417, + 300 + ], + "blocks": [ + { + "bbox": [ + 322, + 228, + 417, + 300 + ], + "lines": [ + { + "bbox": [ + 322, + 228, + 417, + 300 + ], + "spans": [ + { + "bbox": [ + 322, + 228, + 417, + 300 + ], + "type": "image", + "image_path": "afb646411263bd345abc7a06d45d4097e2d58a40dd72594f15e806f4404b2acb.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 356, + 303, + 383, + 312 + ], + "lines": [ + { + "bbox": [ + 356, + 303, + 383, + 312 + ], + "spans": [ + { + "bbox": [ + 356, + 303, + 383, + 312 + ], + "type": "text", + "content": "(c) Ours" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 433, + 228, + 527, + 300 + ], + "blocks": [ + { + "bbox": [ + 433, + 228, + 527, + 300 + ], + "lines": [ + { + "bbox": [ + 433, + 228, + 527, + 300 + ], + "spans": [ + { + "bbox": [ + 433, + 228, + 527, + 300 + ], + "type": "image", + "image_path": "d1de06b416b12b764755c233d5d605f385f8d97d381fea748d1301443820f55e.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 463, + 302, + 498, + 312 + ], + "lines": [ + { + "bbox": [ + 463, + 302, + 498, + 312 + ], + "spans": [ + { + "bbox": [ + 463, + 302, + 498, + 312 + ], + "type": "text", + "content": "(d) Oracle" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 316, + 545, + 371 + ], + "lines": [ + { + "bbox": [ + 305, + 316, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 305, + 316, + 545, + 371 + ], + "type": "text", + "content": "Figure 4. Density plots of test-time adapted features distribution on CIFAR-10-C with Gaussian noise (front) and reference features without corruption (back with yellow color). Each horizontal axis corresponds to one channel. The height of each ridge indicates the number of features that take the same value." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 394, + 423, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 394, + 423, + 406 + ], + "spans": [ + { + "bbox": [ + 306, + 394, + 423, + 406 + ], + "type": "text", + "content": "5. Further Discussions" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 418, + 545, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 418, + 545, + 537 + ], + "spans": [ + { + "bbox": [ + 304, + 418, + 545, + 537 + ], + "type": "text", + "content": "Hebbian Learning by competition through lateral inhibition is a feed-forward process that has no gradient. If combining Hebbian Learning with back-propagation, it is limited to propagate the gradient through Hebbian layers to earlier gradient-based layers during training. Therefore, the Hebbian layers can only be designed in the early layers of models without back-propagation and gradient. In addition, although the Hebbian learning rule is commonly used for long-term reinforcement, the Hebb principle does not cover all forms of long-term synaptic plasticity." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 559, + 378, + 570 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 559, + 378, + 570 + ], + "spans": [ + { + "bbox": [ + 306, + 559, + 378, + 570 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 582, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 712 + ], + "type": "text", + "content": "In this work, we take inspiration from the biological neuromodulation process to construct a novel neuro-modulated Hebbian Learning (NHL) framework. With the unsupervised feed-forward soft Hebbian learning being combined with a learned neuro-modulator to capture feedback from external responses, the source model can be effectively adapted during the testing process. Experimental results on benchmark datasets demonstrate that our proposed method can significantly reduce the testing error for image classification with corruption, and reach new state-of-the-art performance." + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "3735" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "text", + "content": "[1] Yoshua Bengio, Aaron Courville, and Pascal Vincent. Representation learning: A review and new perspectives. IEEE TPAMI, 35(8):1798-1828, 2013. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 124, + 287, + 157 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 124, + 287, + 157 + ], + "spans": [ + { + "bbox": [ + 53, + 124, + 287, + 157 + ], + "type": "text", + "content": "[2] Tom Binzegger, Rodney J Douglas, and Kevan AC Martin. A quantitative map of the circuit of cat primary visual cortex. Journal of Neuroscience, 24(39):8441-8453, 2004. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 158, + 288, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 158, + 288, + 191 + ], + "spans": [ + { + "bbox": [ + 53, + 158, + 288, + 191 + ], + "type": "text", + "content": "[3] Rafal Bogacz. A tutorial on the free-energy framework for modelling perception and learning. Journal of Mathematical Psychology, 76:198-211, 2017. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 191, + 287, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 191, + 287, + 224 + ], + "spans": [ + { + "bbox": [ + 53, + 191, + 287, + 224 + ], + "type": "text", + "content": "[4] Malik Boudiaf, Romain Mueller, Ismail Ben Ayed, and Luca Bertinetto. Parameter-free online test-time adaptation. In CVPR, pages 8344-8353, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 225, + 288, + 269 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 225, + 288, + 269 + ], + "spans": [ + { + "bbox": [ + 53, + 225, + 288, + 269 + ], + "type": "text", + "content": "[5] K Bougrova, N Bonacchi, JA Catarino, EE Dewitt, GT Meijer, P Dayan, ZF Mainen, et al. Comparison of neuromodulator signaling in a visual decision-making task. In Annual Meeting of the Society for Neuroscience, 2022. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 270, + 287, + 302 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 270, + 287, + 302 + ], + "spans": [ + { + "bbox": [ + 53, + 270, + 287, + 302 + ], + "type": "text", + "content": "[6] Aaron Chen and Gus Smith. pytorch-playground. https://github.com/aaron-xichen/pytorch-playground, 2020. 6, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 303, + 287, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 303, + 287, + 335 + ], + "spans": [ + { + "bbox": [ + 53, + 303, + 287, + 335 + ], + "type": "text", + "content": "[7] Zhixiang Chi, Yang Wang, Yuanhao Yu, and Jin Tang. Test-time fast adaptation for dynamic scene deblurring via meta-auxiliary learning. In CVPR, pages 9137-9146, 2021. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 336, + 288, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 336, + 288, + 380 + ], + "spans": [ + { + "bbox": [ + 53, + 336, + 288, + 380 + ], + "type": "text", + "content": "[8] Sungha Choi, Seunghan Yang, Seokeon Choi, and Sungrack Yun. Improving test-time adaptation via shift-agnostic weight regularization and nearest source prototypes. arXiv preprint arXiv:2207.11707, 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 380, + 288, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 288, + 434 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 288, + 434 + ], + "type": "text", + "content": "[9] Francesco Croce, Maksym Andriushchenko, Vikash Sehwag, Edoardo Debenedetti, Nicolas Flammarion, Mung Chiang, Prateek Mittal, and Matthias Hein. Robustbench: a standardized adversarial robustness benchmark. In NeurIPS, 2021. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 435, + 287, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 435, + 287, + 468 + ], + "spans": [ + { + "bbox": [ + 48, + 435, + 287, + 468 + ], + "type": "text", + "content": "[10] Ning Ding, Yixing Xu, Yehui Tang, Chao Xu, Yunhe Wang, and Dacheng Tao. Source-free domain adaptation via distribution estimation. In CVPR, pages 7212-7222, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 469, + 287, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 469, + 287, + 501 + ], + "spans": [ + { + "bbox": [ + 48, + 469, + 287, + 501 + ], + "type": "text", + "content": "[11] Zhengming Ding and Yun Fu. Deep domain generalization with structured low-rank constraint. IEEE TIP, 27(1):304-313, 2017. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 502, + 287, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 502, + 287, + 535 + ], + "spans": [ + { + "bbox": [ + 48, + 502, + 287, + 535 + ], + "type": "text", + "content": "[12] Rodney J Douglas, Kevan AC Martin, et al. Neuronal circuits of the neocortex. Annual Review of Neuroscience, 27(1):419-451, 2004. 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 536, + 287, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 287, + 557 + ], + "type": "text", + "content": "[13] Karl Friston. Hierarchical models in the brain. PLoS Computational Biology, 4(11):e1000211, 2008. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 558, + 287, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 558, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 558, + 287, + 590 + ], + "type": "text", + "content": "[14] Karl Friston. The free-energy principle: a rough guide to the brain? Trends in Cognitive Sciences, 13(7):293-301, 2009. 3, 4, 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 592, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 287, + 624 + ], + "type": "text", + "content": "[15] Karl Friston, James Kilner, and Lee Harrison. A free energy principle for the brain. Journal of Physiology, Paris, 100(1-3):70-87, 2006. 2, 3, 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 625, + 287, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 287, + 657 + ], + "type": "text", + "content": "[16] Yaroslav Ganin and Victor Lempitsky. Unsupervised domain adaptation by backpropagation. In ICML, pages 1180-1189. PMLR, 2015. 1, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[17] Wulfram Gerstner, Marco Lehmann, Vasiliki Liakoni, Dane Corneil, and Johann Brea. Eligibility traces and plasticity on behavioral time scales: experimental support of neohoebbian three-factor learning rules. Frontiers in Neural Circuits, 12:53, 2018. 6" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 37, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 107, + 545, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 129 + ], + "type": "text", + "content": "[19] Donald Olding Hebb. The organization of behavior: A neuropsychological theory. Psychology Press, 2005. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 129, + 545, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 160 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 160 + ], + "type": "text", + "content": "[20] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. In ICLR, 2018. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 161, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 161, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 307, + 161, + 545, + 205 + ], + "type": "text", + "content": "[21] Minhao Hu, Tao Song, Yujun Gu, Xiangde Luo, Jieneng Chen, Yinan Chen, Ya Zhang, and Shaoting Zhang. Fully test-time adaptation for image segmentation. In MICCAI, pages 251-260. Springer, 2021. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 206, + 545, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 206, + 545, + 249 + ], + "spans": [ + { + "bbox": [ + 307, + 206, + 545, + 249 + ], + "type": "text", + "content": "[22] Jiaxing Huang, Dayan Guan, Aoran Xiao, and Shijian Lu. Model adaptation: Historical contrastive learning for unsupervised domain adaptation without source data. In NeurIPS, volume 34, pages 3635-3649, 2021. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 250, + 545, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 250, + 545, + 272 + ], + "spans": [ + { + "bbox": [ + 307, + 250, + 545, + 272 + ], + "type": "text", + "content": "[23] Jonathan J. Hull. A database for handwritten text recognition research. IEEE TPAMI, 16(5):550-554, 1994. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 272, + 545, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 272, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 307, + 272, + 545, + 304 + ], + "type": "text", + "content": "[24] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In ICML, pages 448-456. PMLR, 2015. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 305, + 545, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 305, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 307, + 305, + 545, + 338 + ], + "type": "text", + "content": "[25] Yusuke Iwasawa and Yutaka Matsuo. Test-time classifier adjustment module for model-agnostic domain generalization. In NeurIPS, volume 34, pages 2427-2440, 2021. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 339, + 545, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 339, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 307, + 339, + 545, + 371 + ], + "type": "text", + "content": "[26] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. Master's thesis, University of Tront, 2009. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 372, + 545, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 372, + 545, + 404 + ], + "spans": [ + { + "bbox": [ + 307, + 372, + 545, + 404 + ], + "type": "text", + "content": "[27] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenet classification with deep convolutional neural networks. Communications of the ACM, 60(6):84-90, 2017. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 404, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 404, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 307, + 404, + 545, + 437 + ], + "type": "text", + "content": "[28] Dmitry Krotov and John J Hopfield. Unsupervised learning by competing hidden units. Proceedings of the National Academy of Sciences, 116(16):7723-7731, 2019. 2, 4, 5" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 437, + 545, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 437, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 307, + 437, + 545, + 480 + ], + "type": "text", + "content": "[29] Gabriele Lagani, Fabrizio Falchi, Claudio Gennaro, and Giuseppe Amato. Hebbian semi-supervised learning in a sample efficiency setting. *Neural Networks*, 143:719–731, 2021. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 482, + 545, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 482, + 545, + 504 + ], + "spans": [ + { + "bbox": [ + 307, + 482, + 545, + 504 + ], + "type": "text", + "content": "[30] Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Deep learning. Nature, 521(7553):436-444, 2015. 4" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 504, + 545, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 504, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 307, + 504, + 545, + 547 + ], + "type": "text", + "content": "[31] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 6" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 548, + 545, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 548, + 545, + 592 + ], + "spans": [ + { + "bbox": [ + 307, + 548, + 545, + 592 + ], + "type": "text", + "content": "[32] Andrew B Lehr, Jannik Luboeinski, and Christian Tetzlaff. Neuromodulator-dependent synaptic tagging and capture retroactively controls neural coding in spiking neural networks. Scientific Reports, 12(1):1-18, 2022. 6" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 307, + 593, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 593, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 307, + 593, + 545, + 624 + ], + "type": "text", + "content": "[33] Rui Li, Qianfen Jiao, Wenming Cao, Hau-San Wong, and Si Wu. Model adaptation: Unsupervised domain adaptation without source data. In CVPR, pages 9641-9650, 2020. 1, 3" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 307, + 625, + 545, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 625, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 307, + 625, + 545, + 669 + ], + "type": "text", + "content": "[34] Yizhuo Li, Miao Hao, Zonglin Di, Nitesh Bharadwaj Gundavarapu, and Xiaolong Wang. Test-time personalization with a transformer for human pose estimation. In NeurIPS, volume 34, pages 2583-2597, 2021. 3" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "text", + "content": "[35] Jian Liang, Dapeng Hu, and Jiashi Feng. Do we really need to access the source data? source hypothesis transfer for unsupervised domain adaptation. In ICML, pages 6028-6039. PMLR, 2020. 1, 3" + } + ] + } + ], + "index": 36 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "type": "text", + "content": "3736" + } + ] + } + ], + "index": 38 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "type": "text", + "content": "[36] Timothy P Lillicrap, Adam Santoro, Luke Marris, Colin J Akerman, and Geoffrey Hinton. Backpropagation and the brain. Nature Reviews Neuroscience, 21(6):335-346, 2020. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 117, + 287, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 117, + 287, + 161 + ], + "spans": [ + { + "bbox": [ + 48, + 117, + 287, + 161 + ], + "type": "text", + "content": "[37] Yuejiang Liu, Parth Kothari, Bastien van Delft, Baptiste Bellot-Gurlet, Taylor Mordan, and Alexandre Alahi. Ttt++: When does self-supervised test-time training fail or thrive? In NeurIPS, volume 34, pages 21808-21820, 2021. 1, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 162, + 287, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 162, + 287, + 195 + ], + "spans": [ + { + "bbox": [ + 48, + 162, + 287, + 195 + ], + "type": "text", + "content": "[38] Mingsheng Long, Yue Cao, Jianmin Wang, and Michael Jordan. Learning transferable features with deep adaptation networks. In ICML, pages 97-105. PMLR, 2015. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 195, + 287, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 195, + 287, + 227 + ], + "spans": [ + { + "bbox": [ + 48, + 195, + 287, + 227 + ], + "type": "text", + "content": "[39] Robert A Marsden, Mario Döbler, and Bin Yang. Gradual test-time adaptation by self-training and style transfer. arXiv preprint arXiv:2208.07736, 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 228, + 287, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 228, + 287, + 281 + ], + "spans": [ + { + "bbox": [ + 48, + 228, + 287, + 281 + ], + "type": "text", + "content": "[40] Muhammad Jehanzeb Mirza, Cornelius Buerkle, Julio Jarquin, Michael Opitz, Fabian Oboril, Kay-Ulrich Scholl, and Horst Bischof. Robustness of object detectors in degrading weather conditions. In ITSC, pages 2719-2724. IEEE, 2021. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 282, + 287, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 282, + 287, + 326 + ], + "spans": [ + { + "bbox": [ + 48, + 282, + 287, + 326 + ], + "type": "text", + "content": "[41] M Jehanzeb Mirza, Jakub Micorek, Horst Possegger, and Horst Bischof. The norm must go on: Dynamic unsupervised domain adaptation by normalization. In CVPR, pages 14765-14775, 2022. 1, 3, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 327, + 287, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 327, + 287, + 380 + ], + "spans": [ + { + "bbox": [ + 48, + 327, + 287, + 380 + ], + "type": "text", + "content": "[42] Timoleon Moraitis, Dmitry Toichkin, Adrien Journé, Yansong Chua, and Qinghai Guo. Softhebb: Bayesian inference in unsupervised hebbian soft winner-take-all networks. *Neuromorphic Computing and Engineering*, 2(4):044017, 2022. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 382, + 287, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 382, + 287, + 424 + ], + "spans": [ + { + "bbox": [ + 48, + 382, + 287, + 424 + ], + "type": "text", + "content": "[43] Saeid Motiian, Marco Piccirilli, Donald A Adjeroh, and Gianfranco Doretto. Unified deep supervised domain adaptation and generalization. In ICCV, pages 5715-5725, 2017. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 426, + 287, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 426, + 287, + 468 + ], + "spans": [ + { + "bbox": [ + 48, + 426, + 287, + 468 + ], + "type": "text", + "content": "[44] Krikamol Muandet, David Balduzzi, and Bernhard Schölkopf. Domain generalization via invariant feature representation. In ICML, pages 10-18. PMLR, 2013. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 470, + 287, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 470, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 48, + 470, + 287, + 514 + ], + "type": "text", + "content": "[45] Bernhard Nessler, Michael Pfeiffer, Lars Buesing, and Wolfgang Maass. Bayesian computation emerges in generic cortical microcircuits through spike-timing-dependent plasticity. PLoS Computational Biology, 9(4):e1003037, 2013. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 514, + 287, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 287, + 548 + ], + "type": "text", + "content": "[46] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y Ng. Reading digits in natural images with unsupervised feature learning. 2011. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 548, + 287, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 287, + 580 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 287, + 580 + ], + "type": "text", + "content": "[47] Erkki Oja. Simplified neuron model as a principal component analyzer. Journal of Mathematical Biology, 15(3):267-273, 1982. 4, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 581, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 581, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 581, + 287, + 635 + ], + "type": "text", + "content": "[48] John O'Donnell, Douglas Zeppenfeld, Evan McConnell, Salvador Pena, and Maiken Nedergaard. Norepinephrine: a neuromodulator that boosts the function of multiple cell types to optimize cns performance. Neurochemical Research, 37(11):2496-2512, 2012. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 635, + 287, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 680 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 680 + ], + "type": "text", + "content": "[49] Prashant Pandey, Mrigank Raman, Sumanth Varambally, and Prathosh Ap. Generalization on unseen domains via inference-time label-preserving target projections. In CVPR, pages 12924-12933, 2021. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 681, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 712 + ], + "type": "text", + "content": "[50] Zhongyi Pei, Zhangjie Cao, Mingsheng Long, and Jianmin Wang. Multi-adversarial domain adaptation. In AAAI, volume 32, 2018. 1" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 307, + 72, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 545, + 116 + ], + "type": "text", + "content": "[51] Marina R Picciotto, Michael J Higley, and Yann S Mineur. Acetylcholine as a neuromodulator: cholinergic signaling shapes nervous system function and behavior. *Neuron*, 76(1):116-129, 2012. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "type": "text", + "content": "[52] Roman Pogodin and Peter Latham. Kernelized information bottleneck leads to biologically plausible 3-factor hebbian learning in deep networks. In NeurIPS, volume 33, pages 7296-7307, 2020. 2, 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 162, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 162, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 162, + 545, + 195 + ], + "type": "text", + "content": "[53] Fengchun Qiao, Long Zhao, and Xi Peng. Learning to learn single domain generalization. In CVPR, pages 12556-12565, 2020. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 197, + 545, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 241 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 241 + ], + "type": "text", + "content": "[54] Zhen Qiu, Yifan Zhang, Hongbin Lin, Shuaicheng Niu, Yanxia Liu, Qing Du, and Mingkui Tan. Source-free domain adaptation via avatar prototype generation and adaptation. In IJCAI, pages 2921-2927, 2021. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 242, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 242, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 242, + 545, + 274 + ], + "type": "text", + "content": "[55] Joaquin Quinonero-Candela, Masashi Sugiyama, Anton Schwaighofer, and Neil D Lawrence. Dataset shift in machine learning. MIT Press, 2008. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 276, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 276, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 276, + 545, + 319 + ], + "type": "text", + "content": "[56] Rajesh PN Rao and Dana H Ballard. Predictive coding in the visual cortex: a functional interpretation of some extra-classical receptive-field effects. Nature Neuroscience, 2(1):79-87, 1999. 2, 4, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 320, + 545, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 320, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 307, + 320, + 545, + 353 + ], + "type": "text", + "content": "[57] David E Rumelhart and David Zipser. Feature discovery by competitive learning. Cognitive Science, 9(1):75-112, 1985. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 354, + 545, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 354, + 545, + 408 + ], + "spans": [ + { + "bbox": [ + 307, + 354, + 545, + 408 + ], + "type": "text", + "content": "[58] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. IJCV, 115(3):211-252, 2015. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 410, + 545, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 410, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 307, + 410, + 545, + 443 + ], + "type": "text", + "content": "[59] Ueli Rutishauser, Rodney J Douglas, and Jean-Jacques Slotine. Collective stability of networks of winner-take-all circuits. Neural Computation, 23(3):735-773, 2011. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 445, + 545, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 445, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 307, + 445, + 545, + 498 + ], + "type": "text", + "content": "[60] Steffen Schneider, Evgenia Rusak, Luisa Eck, Oliver Bringmann, Wieland Brendel, and Matthias Bethge. Improving robustness against common corruptions by covariate shift adaptation. In NeurIPS, volume 33, pages 11539-11551, 2020. 6, 7, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 500, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 500, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 307, + 500, + 545, + 544 + ], + "type": "text", + "content": "[61] Seonguk Seo, Yumin Suh, Dongwan Kim, Geeho Kim, Jongwoo Han, and Bohyung Han. Learning to optimize domain specific normalization for domain generalization. In ECCV, pages 68-83. Springer, 2020. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 545, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 545, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 307, + 545, + 545, + 578 + ], + "type": "text", + "content": "[62] Assaf Shocher, Nadav Cohen, and Michal Irani. \"zero-shot\" super-resolution using deep internal learning. In CVPR, pages 3118-3126, 2018. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 579, + 545, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 579, + 545, + 623 + ], + "spans": [ + { + "bbox": [ + 307, + 579, + 545, + 623 + ], + "type": "text", + "content": "[63] Luisa Speranza, Umberto di Porzio, Davide Viggiano, Antonio de Donato, and Floriana Volpicelli. Dopamine: The neuromodulator of long-term synaptic plasticity, reward and movement control. Cells, 10(4):735, 2021. 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 624, + 545, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 677 + ], + "type": "text", + "content": "[64] Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei A. Efros, and Moritz Hardt. Test-time training with self-supervision for generalization under distribution shifts. In ICML, volume 119, pages 9229-9248. PMLR, 2020. 1, 3, 6, 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "text", + "content": "[65] Riccardo Volpi and Vittorio Murino. Addressing model vulnerability to distributional shifts over image transformation sets. In ICCV, pages 7980-7989, 2019. 3" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3737" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 398 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "type": "text", + "content": "[66] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In ICLR, 2020. 1, 3, 6, 7, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 107, + 286, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 107, + 286, + 149 + ], + "spans": [ + { + "bbox": [ + 49, + 107, + 286, + 149 + ], + "type": "text", + "content": "[67] Fan Wang, Zhongyi Han, Yongshun Gong, and Yilong Yin. Exploring domain-invariant parameters for source free domain adaptation. In CVPR, pages 7151-7160, June 2022. 1, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 152, + 286, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 152, + 286, + 184 + ], + "spans": [ + { + "bbox": [ + 49, + 152, + 286, + 184 + ], + "type": "text", + "content": "[68] Qin Wang, Olga Fink, Luc Van Gool, and Dengxin Dai. Continual test-time domain adaptation. In CVPR, pages 7201-7211, 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 186, + 286, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 186, + 286, + 217 + ], + "spans": [ + { + "bbox": [ + 49, + 186, + 286, + 217 + ], + "type": "text", + "content": "[69] James CR Whittington and Rafal Bogacz. Theories of error back-propagation in the brain. Trends in cognitive sciences, 23(3):235-250, 2019. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 220, + 286, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 220, + 286, + 252 + ], + "spans": [ + { + "bbox": [ + 49, + 220, + 286, + 252 + ], + "type": "text", + "content": "[70] Shiqi Yang, Yaxing Wang, Joost van de Weijer, Luis Herranz, and Shangling Jui. Generalized source-free domain adaptation. In ICCV, pages 8978-8987, October 2021. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 254, + 286, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 254, + 286, + 284 + ], + "spans": [ + { + "bbox": [ + 49, + 254, + 286, + 284 + ], + "type": "text", + "content": "[71] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. In BMVC. British Machine Vision Association, 2016. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 287, + 286, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 287, + 286, + 319 + ], + "spans": [ + { + "bbox": [ + 49, + 287, + 286, + 319 + ], + "type": "text", + "content": "[72] Matthew D Zeiler and Rob Fergus. Visualizing and understanding convolutional networks. In ECCV, pages 818-833. Springer, 2014. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 322, + 286, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 322, + 286, + 363 + ], + "spans": [ + { + "bbox": [ + 49, + 322, + 286, + 363 + ], + "type": "text", + "content": "[73] Shanshan Zhao, Mingming Gong, Tongliang Liu, Huan Fu, and Dacheng Tao. Domain generalization via entropy regularization. In NeurIPS, volume 33, pages 16096-16107, 2020. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 366, + 286, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 366, + 286, + 398 + ], + "spans": [ + { + "bbox": [ + 49, + 366, + 286, + 398 + ], + "type": "text", + "content": "[74] Kaiyang Zhou, Ziwei Liu, Yu Qiao, Tao Xiang, and Chen Change Loy. Domain generalization: A survey. IEEE TPAMI, 2022. 3" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "3738" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Neuron Structure Modeling for Generalizable Remote Physiological Measurement/0c500895-2bd8-4cea-80ae-7bc4fd9b8ba5_content_list.json b/2023/Neuron Structure Modeling for Generalizable Remote Physiological Measurement/0c500895-2bd8-4cea-80ae-7bc4fd9b8ba5_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5705a525bf6e6af8d4d727229b4bfcf0ae6d29e0 --- /dev/null +++ b/2023/Neuron Structure Modeling for Generalizable Remote Physiological Measurement/0c500895-2bd8-4cea-80ae-7bc4fd9b8ba5_content_list.json @@ -0,0 +1,1607 @@ +[ + { + "type": "text", + "text": "Neuron Structure Modeling for Generalizable Remote Physiological Measurement", + "text_level": 1, + "bbox": [ + 143, + 130, + 826, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hao LU $^{1,2}$ , Zitong YU $^{3}$ , Xuesong NIU $^{4}$ , Yingcong CHEN $^{1,2,*}$", + "bbox": [ + 241, + 202, + 733, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1The Hong Kong University of Science & Technology (Guangzhou),", + "bbox": [ + 217, + 220, + 758, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ The Hong Kong University of Science & Technology, $^{3}$ Great Bay University, $^{4}$ Kuaishou Technology hlu585@connect.hkust-gz.edu.cn, zitong.yu@ieee.org, nxsedson@gmail.com, yingcongchen@ust.hk", + "bbox": [ + 88, + 239, + 885, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 325, + 313, + 342 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Remote photoplethysmography (rPPG) technology has drawn increasing attention in recent years. It can extract Blood Volume Pulse (BVP) from facial videos, making many applications like health monitoring and emotional analysis more accessible. However, as the BVP signal is easily affected by environmental changes, existing methods struggle to generalize well for unseen domains. In this paper, we systematically address the domain shift problem in the rPPG measurement task. We show that most domain generalization methods do not work well in this problem, as domain labels are ambiguous in complicated environmental changes. In light of this, we propose a domain-label-free approach called NEuron SStructure modeling (NEST). NEST improves the generalization capacity by maximizing the coverage of feature space during training, which reduces the chance for under-optimized feature activation during inference. Besides, NEST can also enrich and enhance domain invariant features across multi-domain. We create and benchmark a large-scale domain generalization protocol for the rPPG measurement task. Extensive experiments show that our approach outperforms the state-of-the-art methods on both cross-dataset and intra-dataset settings. The codes are available at https://github.com/LuPaoPao/NEST.", + "bbox": [ + 75, + 358, + 473, + 705 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 715, + 209, + 732 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Physiological signals such as heart rate (HR), and heart rate variability (HRV), respiration frequency (RF) are important body indicators that serve not only as vital signs but also track the level of sympathetic activation [17, 33, 54]. Traditional physiological measurements, such as electrocardiograms, heart rate bands, and finger clip devices, have high accuracy. However, they are costly, intrusive, and uncomfortable to wear for a long time. Remote photoplethysmography (rPPG) can extract blood volume pulse (BVP)", + "bbox": [ + 75, + 741, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e9b3e153f156f2e4f7039efef49987157709640e9b430dac9ed9858c5c5d228c.jpg", + "image_caption": [ + "(a) Typical samples from different datasets" + ], + "image_footnote": [], + "bbox": [ + 503, + 325, + 575, + 467 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/74734f2b0862877af9122dbe060ea2681ef4e4059563e49850d9a0741960817f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 581, + 325, + 653, + 465 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d859c2131900ef0d19b9834e4e7bdc4bb0635d3b2f20f29419d29f49dc03ad72.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 656, + 325, + 730, + 465 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1a20cf70a8f365ec011d1ce71fbb4d3e44bf0df7002125b2074baab67bfaf0ba.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 735, + 325, + 808, + 465 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7d06aec4f82f4d151d3d618b1d4b1b395fb4971a43def901c78f1a38d9be854a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 818, + 325, + 890, + 465 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/62d042aca48080d8e790247659b54c3194e234ea51e87f4cbc651b253fe93d81.jpg", + "image_caption": [ + "(b) The performance of different methods on DG protocol", + "Figure 1. (a) Typical samples from different publicly rPPG datasets: VIPL-HR [36], V4V [47], UBFC-rPPG [1], BUAA [68], PURE [52]. (b) The performance of different methods on DG protocol (test on the UBFC-rPPG dataset with training on the VIPL, V4V, PURE, and BUAA)." + ], + "image_footnote": [], + "bbox": [ + 504, + 479, + 870, + 625 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "from face video, which analyzes the periodic changes of the light absorption of the skin caused by heartbeats. Then various physiological indicators (such as HR and HRV) can be calculated based on BVP signals [37, 69, 70]. With nonintrusion and convenience, the rPPG-based physiological measurement method can only use an ordinary camera to monitor physiological indicators and gradually become a research hotspot in the computer vision field [7, 28, 32, 36, 44, 49].", + "bbox": [ + 496, + 714, + 893, + 851 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Traditional rPPG measurement methods include signal blind decomposition [21,34,42] and color space transformation [9,59,63]. These approaches rely on heartbeat-related", + "bbox": [ + 498, + 854, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 94, + 886, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "18589", + "bbox": [ + 480, + 924, + 517, + 936 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "statistical information, only applicable in constrained environments. In recent years, deep learning (DL) based approaches [6, 19, 25, 28, 36, 44, 50, 69, 70] have shown their great potentials in rPPG measurement. By learning dedicated rPPG feature representation, these methods achieve promising performance in much more complicated environments [1, 36, 47, 52, 68].", + "bbox": [ + 75, + 90, + 470, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, deep learning methods suffer from significant performance degradation when applied in real-world scenarios. This is because most training data are captured in lab environments with limited environmental variations. With domain shifts (e.g., different illumination, camera parameters, motions, etc.), these models may struggle to generalize for the unseen testing domain. To validate this, we conduct a cross-dataset evaluation shown in Fig. 1(b). As shown, all DL-based methods do not work well in this evaluation. Furthermore, it is worth noting that DeepPhys [6] and TS-CAN [25] even perform inferior to traditional approach POS [63].", + "bbox": [ + 75, + 199, + 468, + 381 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To improve the performance on the unseen domains, one common practice is to incorporate Domain Generalization (DG) approaches, e.g., encouraging intermediate features to be domain-invariant [12, 18, 30, 38, 57, 62]. However, as shown in Fig. 1 (b) (AD [12] and GroupDRO [38]), the improvements are still quite limited. One reason is that existing DG methods assume that domain labels can be clearly defined (e.g., by data source). Unfortunately, different background scenes, acquisition devices, or even identities could cause very different data distributions in rPPG measurement tasks, and such distribution discrepancy may exist within or cross datasets. In this case, explicitly defining domains is very difficult, and simply treating one dataset as one domain may lead to inferior performance. This is termed as agnostic domain generalization. Besides this, as physiological cues are usually much more subtle than the various noise, the model may overfit in the source domain with the limited training data.", + "bbox": [ + 75, + 383, + 470, + 655 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose the NEural SStructure modeling (NEST), a principled solution to the abovementioned problems. The main idea of NEST is to narrow the under-optimized and redundant feature space, align domain invariant features, and enrich discriminative features. Our intuition is as follows. Neural structure refers to the channel activation degree in each convolution layer, which reveals the discriminative feature combination for the specific sample. As the limited variation in a certain restricted domain, there are some spaces that the model seldomly is optimized. Out-of-distribution (OOD) samples may cause abnormal activation in these spaces, which may lead to performance degeneration. Therefore, we regularize the neural structure to encourage the model to be more well-conditioned to avoid abnormal activation caused by OOD samples. Specifically, we propose the NEural SStructure Coverage Maxi", + "bbox": [ + 75, + 657, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "mization (NEST-CM) that encourages all neural spaces to be optimized during training, reducing the chance of abnormal activation during testing. Secondly, we propose the NEural SStructure Targeted Alignment (NEST-TA) that encourage network suppresses domain variant feature by comparing the samples with the similar physiological information. Thirdly, we propose the NEural SStructure Diversity Maximization (NEST-DM) to enrich discriminative features against unseen noise. It should be noted that our approach does not rely on domain labels, which is more applicable in the rPPG measurement task. To summarize, our contributions are listed as follows:", + "bbox": [ + 496, + 90, + 890, + 271 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We are the first to study the domain shift problem in rPPG measurement, which introduces a new challenge, agnostic domain generalization.", + "2. We propose the NEural SStructure modeling to alleviate domain shift, which is performed by narrowing the under-optimized feature space, and enhancing and enriching domain invariant features.", + "3. We establish a large-scale domain generalization (DG) benchmark for rPPG measurement, which is the first DG protocol in this task. Extensive experiments in this dataset show the superiority of our approach." + ], + "bbox": [ + 496, + 272, + 890, + 438 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 500, + 441, + 637, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Remote Physiological Measurement", + "text_level": 1, + "bbox": [ + 500, + 462, + 808, + 478 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Remote physiological measurement aims to estimate HR and HRV values by analyzing the chrominance changes in the skin. Traditional methods mainly perform color space transformations or signal decomposition to obtain the BVP signal with high SNR [9, 10, 42, 64]. These methods are designed manually for certain lighting scenes with slight motion and can not generalize to the complex environment in the wild. With powerful modeling capacity, the deep learning (DL) model has been successfully used in remote physiological estimation [6, 13, 35, 36, 51, 67, 69]. Among these methods, various model structures are tailored to extract BVP signals from face videos by aggregating spatial and temporal information [6, 37, 51, 69, 70]. Besides, many hand-designed feature maps are used to reduce the computational burden of the network by removing the irrelevant background [13, 28, 36, 45, 50]. Some work utilizes the distribution of the target domain to improve the model performance [19, 26], which is not practical without touching the unseen domain in advance. Differently, we study the model generalization on domain generalization protocol for rPPG measurement.", + "bbox": [ + 496, + 482, + 890, + 797 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Domain generalization", + "text_level": 1, + "bbox": [ + 500, + 805, + 710, + 821 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Domain generalization (DG), i.e., out-of-distribution generalization, aims to generalize to the unseen domain by training a model on multiple or single source domains [62]. The main solution can be categorized into data manipulation, representation learning, and meta-learning. Data", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "18590", + "bbox": [ + 480, + 924, + 517, + 936 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b47254a0e21d43c8e68d78f1162f3a1ef02ef922759cfe210bd36533a3c81e86.jpg", + "image_caption": [ + "Figure 2. An overview of the proposed method. (a) Spatial-temporal map (STMap) is extracted from face video [36]. (b) NEST representation is calculated from the samples of the different domains. (c) The under-optimized space in NEST presentation distribution is narrowed by NEST-CM. (d) NEST-TA align the feature of multi domain at instance-level. (e) The distance of different samples are pulled away by NEST-DM." + ], + "image_footnote": [], + "bbox": [ + 86, + 88, + 624, + 428 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/078712f385fb6ca430b0a2914122470b6d5cb25b4d750dcb141b7ab88e0dfcf8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 625, + 89, + 883, + 428 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The generalization of the model, which can remove irrelevant style information by data generation and data augmentation [48, 71]. The main goal of representation learning is to learn domain-invariant features by alleviating covariate shift across domain [12, 18, 38, 57]. By simulating domain shift on the virtual dataset, meta-learning seeks transferable and shareable knowledge from multiple domain [23, 30]. Unlike common tasks such as classification, segmentation, and detection [15, 20, 22, 24, 40], rPPG measurement tasks do not have explicit boundaries between different domains, i.e., agnostic domain generalization. Recently, some single source domain generalization (SSDG) methods focus on the worst-case formulation without using the domain label [43, 57, 66]. Different from all the methods above, we tackle the DG problem of regression task in view of neuron structure modeling.", + "bbox": [ + 75, + 500, + 472, + 758 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 773, + 169, + 789 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overall Framework", + "text_level": 1, + "bbox": [ + 76, + 800, + 266, + 814 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this work, we propose the NEuron SStructure modeling (NSET), without using domain label, to alleviate the performance degradation caused by domain shift as illustrated in Fig. 2. Specifically, we first compressed the face video into the spatial-temporal map (STMap) as the input of", + "bbox": [ + 75, + 824, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "CNN refer to [36] as shown in Fig. 2 (a). Sequentially, we define an Activated NEuron SStructure (NEST) representation that describes the activated degree of different channels of the network as shown in Fig. 2 (b). To alleviate the domain shift, we model a unified NEST distribution by three regularizations: (1) NEuron SStructure Coverage Maximization (NEST-CM) is used to narrow under-optimized space that is redundant and unstable as shown in Fig. 2 (c); (2) NEuron SStructure Targeted Alignment (NEST-TA) is introduced to align multi-domain feature distribution at instance-level as shown in Fig. 2 (d). (3) Neuron Structure Diversity Maximization (NEST-DM) loss to force the network to explore more features with discriminant ability as shown in Fig. 2 (e).", + "bbox": [ + 496, + 500, + 893, + 712 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. NEuron SStructure Modeling", + "text_level": 1, + "bbox": [ + 500, + 724, + 758, + 741 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "NEuron SStructure Representation. For the rPPG measurement task, the convolution kernels are adaptively activated to extract physiological features against various noise distribution [27, 56, 65]. However, the facial physiological signals are feeble and can be easily affected by various unseen noises, which make the filters activated abnormally. Therefore, the NEuron SStructure (NEST) representation is defined to explore how domain shift affects the model robustness in view of activated feature distribution. Specifically, we firstly average the feature map", + "bbox": [ + 496, + 750, + 890, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "18591", + "bbox": [ + 480, + 925, + 517, + 936 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1ecd02f2e15691a1050d15c1eb02ba96783252b535a36e696afca9c7ebc73031.jpg", + "image_caption": [ + "(a) The expectation of different base projection" + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 433, + 190 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9a25213a89acee181e1b480a154ddb6982550d3d417498e363153b963c0baa2b.jpg", + "image_caption": [ + "(b) The projection distribution of base with highest eigenvalue", + "Figure 3. (a) The expectation of different base projection. (b) The distribution of base with highest eigenvalue. (c) The distribution of base with lowest eigenvalue." + ], + "image_footnote": [], + "bbox": [ + 96, + 203, + 274, + 297 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/87ec75dcb2c77d0928dffdda3336a82164e716a434f56f0e5eb8c799a9ca4953.jpg", + "image_caption": [ + "(c) The projection distribution of base with lowest eigenvalue" + ], + "image_footnote": [], + "bbox": [ + 292, + 203, + 444, + 296 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$m_{j}\\in R^{W\\times H\\times C_{j}}$ over the spatial dimensions in the end of $j$ -th layer to get the preliminary statistical presentation $\\bar{o}_j\\in R^{C_j}$ of the neuron structure, where $C_j$ is the channel number of the $j$ -th layer. Following that, the preliminary presentation $\\bar{o}_j$ is fed into the max-min normalized to get the NEST representation $o_{j}\\in R_{j}^{C}$ of the $j$ -th layer. The full NEST representation of $i$ -th sample $x_{i}$ is defined as $O_{i} = \\{o_{i,1},o_{i,2},\\dots,o_{i,U}\\} \\in R^{(C_{1} + C_{2} + \\ldots +C_{U})}$ , where $U$ is the number of selected layer.", + "bbox": [ + 75, + 378, + 467, + 516 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "NEuron SStructure Coverage Maximization. Existing works show that the OOD samples may activate abnormal neuron structures (feature combination) that seldom be trained with the source data so that the model's performance degrades drastically [31, 39, 57]. To investigate this phenomenon in rPPG measurement task, we analyze the optimized degree of different feature combinations.", + "bbox": [ + 75, + 518, + 467, + 623 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, we decompose the NEST representation of the last layer into orthogonal bases with Singular Value Decomposition (SVD). Here, small eigenvalues indicate that training samples cover a small space in these corresponding bases. Thus, unseen target samples are more likely to lay out of the well-optimized space, causing a larger domain shift. To validate this, we plot the means of source and target samples on these bases in Fig. 3(a). As shown, the mean differences are larger for bases of smaller eigenvalues. To make it clearer, We further visualize the data distribution on the first and last bases in Fig. 3 (b) and (c). As shown, the distribution difference in the last base is much more significant than that of the first one. All these suggest that for the NEST representation, components of small eigenvalues are less optimized during training, and domain shifts are more likely to appear there.", + "bbox": [ + 75, + 626, + 467, + 867 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In sight of this, we propose the NEuron SStructure Coverage Maximization (NEST-CM) regularization that improves", + "bbox": [ + 76, + 869, + 467, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e21bf16425701cf4d040c06fc08e94c142c0593e717508a8e7841ee3e15836b0.jpg", + "image_caption": [ + "Figure 4. The detail of SVD for the $j$ -th layer in CNN." + ], + "image_footnote": [], + "bbox": [ + 501, + 88, + 887, + 152 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the eigenvalues of the last few bases. This could narrow the under-optimized space as shown in Fig. 2 (c). For a training batch with $N$ samples, NEST representation set of $j$ -th layer $B_{j} = \\{o_{1,j}; o_{2,j}; \\ldots; o_{N,j}\\} \\in R^{N * C_{j}}$ is decomposed to $S_{j} \\cdot V_{j} \\cdot D_{j}$ by SVD method as shown in Fig. 4, i.e., $B_{j} = S_{j} \\cdot V_{j} \\cdot D_{j}$ , where $V_{j} \\in R^{N * C_{j}}$ and $N > C_{j}$ . The bases with small eigenvalue values are rare and under-optimized during the training stage. To reduce the occurrence probability of under-optimized NEST caused by domain shift, we stimulate under-optimized bases by maximizing the corresponding eigenvalues:", + "bbox": [ + 498, + 170, + 890, + 335 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {C M} = - \\mathbb {E} _ {j \\in U} \\frac {\\sum_ {i = 1} ^ {C ^ {\\prime}} \\lambda_ {i , j} ^ {\\prime}}{\\sum_ {i = 1} ^ {C} \\lambda_ {i , j}}, \\lambda_ {i, j} ^ {\\prime} < t, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 349, + 890, + 392 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\lambda_{i,j}$ represent the $i$ -th normalized eigenvalue in $j$ -th layer' diagonal matrix $V_{j}$ , and $\\lambda_{i,j}^{\\prime}$ is the eigenvalue less than the threshold $t$ . Considering different channel numbers in each layer, we design an adaptive threshold $t$ to select the relatively under-optimized bases, i.e., $t = \\frac{\\rho}{C_j}$ , where $\\rho$ is a trade-off parameter. Besides, $C_j^\\prime$ is the number of eigenvalues of $j$ -th layer under the threshold, and $U$ is the number of layers.", + "bbox": [ + 498, + 402, + 890, + 527 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "NEuron SStructure Targeted Alignment. Besides the NEST-CM introduced above, we further improve the generalization ability by aligning the NEST representation of different domains. Domain invariant learning has demonstrated its effectiveness in DG problems [12, 18, 30, 38, 57, 62]. Some recent research [11, 16, 29, 60] further shows that aligning domain features within the same labels could lead to better performance. However, labels in rPPG are continuous, and it is hard to find samples with identical labels in a training batch.", + "bbox": [ + 498, + 527, + 890, + 676 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To cope with this, we propose a novel regularization term called NEuron SStructure Alignment (NEST-TA). NEST-TA builds a synthesized prototype $O_{i}^{p}$ with the same physiological information as the given targeted sample $O_{i}$ . Then, the feature within the same labels can be aligned across domains like [11, 16, 29, 60]. Specifically, we select the top $K$ samples closest to the targeted ground truth $y_{i}$ , i.e., $[O_{i,1}', \\dots, O_{i,k}', \\dots, O_{i,K}']$ . The NEST center of the selected samples $\\overline{O}_{i}'$ is calculated by averaging the NEST presentation of selected samples. Then we calculate the direction vectors between the NEST presentation center of the selected samples $\\overline{O}_{i}'$ and the NEST presentation of each sample $O_{i,k}'$ . Following that, the NEST center of the selected samples $\\overline{O}_{i}'$ is calibrated to the synthesized prototype $O_{i}^{p}$ ,", + "bbox": [ + 498, + 679, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "18592", + "bbox": [ + 480, + 924, + 517, + 935 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "which is performed by using the Gaussian distribution function to fuse these direction vectors:", + "bbox": [ + 76, + 90, + 468, + 119 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nO _ {i} ^ {p} = \\overline {{O ^ {\\prime}}} _ {i} + \\sum_ {k = 1} ^ {K} W _ {i, k} * \\frac {y _ {i} - \\bar {y} _ {i} ^ {\\prime}}{y _ {i , k} ^ {\\prime} - \\bar {y} _ {i} ^ {\\prime}} \\left(O _ {i, k} ^ {\\prime} - \\overline {{O ^ {\\prime}}} _ {i}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 125, + 468, + 170 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nW _ {i, k} = \\frac {G (y _ {i , k} - \\overline {{y}} _ {i} ^ {\\prime} , \\sigma)}{\\sum_ {k = 1} ^ {K} G (y _ {i , k} - \\overline {{y}} _ {i} ^ {\\prime} , \\sigma)},\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 170, + 375, + 205 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $G(-, \\sigma)$ is the Gaussian distribution with the mean of zero and the standard deviation of $\\sigma$ , i.e., $G(y_{i,k} - \\overline{y}_i', \\sigma) = \\frac{1}{\\sqrt{2\\pi}} e^{-\\frac{(y_{i,k} - \\overline{y}_i')^2}{2\\sigma^2}}$ , where $\\sigma$ is a hyper-parameter. To align the feature distribution across different domains, NEST-TA uses the cosine similarity $Cos(-, -)$ to reduce the variant features between targeted samples $O_i$ and prototype sample $O_i^p$ :", + "bbox": [ + 76, + 214, + 470, + 330 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {T A} = \\mathbb {E} _ {i \\in N} \\operatorname {C o s} \\left(O _ {i}, O _ {i} ^ {p}\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 338, + 468, + 356 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "NEuron SStructure Diversity Maximization. Existing works elaborate that it is effective to explore more features with discriminant information to improve the generalization [4,30,41,46]. Motivated by this, NEuron SStructure Diversity Maximization (NEST-DM) is introduced to increase the distinction between the NEST presentation of different samples as shown in Fig. 2 (e). By doing this, the network is forced to explore more features with physiological information in a contrastive learning way, which can further improve model discrimination. NEST-DM $\\mathcal{L}_{DM}$ can be represented as:", + "bbox": [ + 76, + 367, + 468, + 532 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {D M} = - \\mathbb {E} _ {i \\in N} \\log \\left(\\frac {e ^ {C o s \\left(O _ {i} , \\widehat {O} _ {i}\\right) / \\tau}}{\\sum_ {j = 1 (i \\neq j)} ^ {N} e ^ {C o s \\left(O _ {i} , \\widehat {O} _ {j}\\right) / \\tau}}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 542, + 468, + 584 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\widehat{O}_i$ is the NEST representation of augmented sample from $O_i$ , and $N$ and $\\tau$ are the batch size and temperature parameter. The augmentation strategies for STMap will be elaborated in section 4.2.", + "bbox": [ + 76, + 595, + 468, + 655 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Model Training. For the rPPG measurement task, we chose two popular losses in this field. Specifically, L1 loss $\\mathcal{L}_{L1}$ and negative Pearson's correlation coefficient $\\mathcal{L}_p$ are used to constrain the predicted heart rate value and the predicted BVP signal respectively [14, 28, 37, 69]. These two losses are also used for augmented data. With the violent confrontation at the beginning of training, $\\mathcal{L}_{L1}$ make network tends to predict the mean of the heart rate value distribution instead of learning the physiological feature like $\\mathcal{L}_p$ . Besides, $\\mathcal{L}_{NSM}$ , $\\mathcal{L}_{NEST-S}$ , and $\\mathcal{L}_{NEST-DM}$ also have meaningless effects on under-trained networks at the beginning.", + "bbox": [ + 76, + 657, + 468, + 837 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To solve the above-illustrated issues, we introduce the adaptation factor $\\gamma = \\frac{2}{1 + \\exp(-10r)}$ and $r = \\frac{\\text{inters}_{\\text{current}}}{\\text{inters}_{\\text{total}}}$ to suppress meaningless constraints at the early training stage, which can accelerate model convergence and improve training stability [12, 15]. The final loss $\\mathcal{L}_{\\text{overall}}$ can be formulated as follows:", + "bbox": [ + 76, + 838, + 468, + 928 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {o v e r a l l}} = k _ {1} \\mathcal {L} _ {p} + \\gamma \\left(k _ {2} \\mathcal {L} _ {L 1} + k _ {3} \\mathcal {L} _ {C M} \\right. \\tag {5} \\\\ + k _ {4} \\mathcal {L} _ {T A} + k _ {5} \\mathcal {L} _ {D M}), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 114, + 890, + 148 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $k_{1} - k_{5}$ are five trade-off parameters.", + "bbox": [ + 500, + 160, + 785, + 176 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 189, + 630, + 205 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 500, + 214, + 601, + 228 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We elaborately select five datasets (including different motion, camera, and lighting conditions) to curate and benchmark a large-scale DG protocol:", + "bbox": [ + 498, + 237, + 890, + 282 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "VIPL-HR [36] have nine scenarios, three RGB cameras, different illumination conditions, and different levels of movement. It is worth mentioning that the BVP signal and video of the dataset do not match in the time dimension, so $\\mathcal{L}_p$ is not used for this dataset. We used the output of HR head as the evaluation result. Besides, we normalize STMap to 30 fps by cubic spline interpolation to solve the problem of the unstable frame rate of video [28, 50, 70].", + "bbox": [ + 498, + 282, + 890, + 402 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "V4V [47] is designed to collect data with the drastic changes of physiological indicators by simulating ten tasks such as a funny joke, 911 emergency call, and odor experience. It is worth mentioning that There are only heart rate labels in the dataset and no BVP signal labels, so $\\mathcal{L}_p$ is not used for this dataset. We used the output of HR head as the evaluation result.", + "bbox": [ + 498, + 402, + 890, + 507 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "BUAA [68] is proposed to evaluate the performance of the algorithm against various illumination. We only use data with illumination greater than or equal to 10 lux because underexposed images require special algorithms that are not considered in this article. We used the output of BVP head as the evaluation for BUAA, PURE, and UBFC-rPPG datasets.", + "bbox": [ + 498, + 508, + 890, + 612 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "PURE [52] contains 60 RGB videos from 10 subjects with six different activities, specifically, sitting still, talking, and four rotating and moving head variations. The BVP signals are down-sampled from 60 to 30 fps with cubic spline interpolation to align with the videos.", + "bbox": [ + 498, + 614, + 890, + 689 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "UBFC-rPPG [1] containing 42 face videos with sunlight and indoor illumination, the ground-truth BVP signals, and HR values were collected by CMS50E.", + "bbox": [ + 498, + 689, + 890, + 734 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 500, + 739, + 715, + 753 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training Details. Our proposed method is performed under the Pytorch framework. For the generation of STMap from Video, we use the FAN1 to detect the 2D landmarks of faces [2], and other steps completely follow [36]. STMap in each dataset is sampled with a time window of 256 with step 10. Following [55, 61], we also design spatial and temporal augmentation for STMap, i.e., shuffle each row of STMap and slide the time window. In addition, color jitter and blur", + "bbox": [ + 498, + 757, + 890, + 878 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "", + "bbox": [ + 517, + 886, + 857, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 478, + 924, + 490, + 936 + ], + "page_idx": 4 + }, + { + "type": "footer", + "text": "18593", + "bbox": [ + 480, + 957, + 517, + 969 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/ea520a6398e833a1c484e35a3c0ddb9ffee1ed75d55f2fb30ec300c7d05799a5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodUBFCPUREBUAAVIPLV4V
MAE↓RMSE↓r↑MAE↓RMSE↓r↑MAE↓RMSE↓r↑MAE↓RMSE↓r↑MAE↓RMSE↓r↑MAE↓RMSE↓r↑MAE↓RMSE↓r↑MAE↓RMSE↓r↑
GREEN [59]8.029.180.3610.3214.270.525.827.990.5612.1818.230.2515.6421.430.06
CHROM [9]7.238.920.519.7912.760.376.098.290.5111.4416.970.2814.9219.220.08
POS [63]7.358.040.499.8213.440.345.047.120.6314.5921.260.1917.6523.220.04
DeepPhys [6]7.828.420.549.3412.560.554.786.740.6912.5619.130.1414.5219.110.14
TS-CAN [25]7.638.250.559.1212.380.574.846.890.6812.3418.940.1614.7719.960.12
Rhythmnet* [36]5.797.910.787.3910.490.773.385.170.848.9712.160.4910.1614.570.34
Dual-GAN* [28]5.557.620.797.2410.270.783.415.230.848.8811.690.5010.0414.440.35
BVPNet* [7]5.437.710.807.2310.250.783.695.480.818.4511.640.5110.0114.350.36
AD*+ [12]5.928.080.767.4210.610.733.495.490.828.4111.710.5310.4714.640.32
GroupDRO*+ [38]5.737.970.787.6910.830.783.415.210.838.3511.670.549.9414.290.36
Coral*+ [53]5.898.040.767.5910.870.723.645.740.808.6811.910.5310.3214.420.32
VREx*+ [18]5.597.680.817.2410.140.783.275.010.868.3711.620.549.8214.160.37
NCDG*+ [57]5.317.560.827.3210.350.773.125.160.858.4711.810.5210.1414.460.34
Baseline* [36]5.797.910.787.3910.490.773.385.170.848.9712.160.4910.1614.570.34
NEST*+ w/o LCM4.927.190.836.919.940.782.794.620.898.2111.560.559.5813.580.38
NEST*+ w/o LTA4.977.280.836.819.770.813.034.940.868.1911.470.559.5213.520.38
NEST*+ w/o LDM4.897.110.846.879.810.792.934.770.888.1411.430.569.4913.490.40
NEST*+4.676.790.866.719.590.812.884.690.897.8611.150.589.2713.790.41
", + "bbox": [ + 76, + 88, + 895, + 407 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. HR estimation results on MSDG protocol. It is worth noting that we have removed the GRU mechanism in Rhythmnet because it is an unfair comparison with using GRU to aggregate the temporal information of multiple clips instead of only one clip; * means that these methods use the STMap as the input of CNN and use the STMap augmentation strategy as mentioned in section 4.2; + means that these methods are based on baseline (Rhythmnet without GRU).", + "bbox": [ + 75, + 411, + 893, + 467 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "augmentation are used for STMap [5]. Then the STMap $x \\in R^{25 \\times 256 \\times 3}$ is resized to $x' \\in R^{64 \\times 256 \\times 3}$ as the input of the network. We use the ResNet-18 with removing the max pooling layer as the feature extractor, and the final full connection layer is used as the HR estimation head. At the same time, we designed a BVP signal regression head containing four blocks, each block involves a Conv., transpose, and two Conv. layers with batch normalization and Relu function. Hyper-parameter $\\sigma$ , $\\rho$ and $\\tau$ are set to 5, 0.1, and 0.2 according to the experimental result, and the trade-off parameter $k_{1} - k_{5}$ are set to 1, 0.1, 0.001, 0.1 and 0.01 based on the scale of losses. Adam optimizer with a learning rate of 0.001 is used for training. The batch size and iterations are set to 1024 and 20000.", + "bbox": [ + 75, + 476, + 472, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Performance Metrics. Following existing methods [6, 28, 37, 50, 70], standard deviation (SD), mean absolute error (MAE), root mean square error (RMSE), and Pearson's correlation coefficient (r) are used to evaluate the HR estimation. MAE, RMSE, and r are used to evaluate the HRV measurement, including low frequency (LF), high frequency (HF), and LF/HF.", + "bbox": [ + 75, + 686, + 470, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Multi-Source Domain Generalization", + "text_level": 1, + "bbox": [ + 76, + 801, + 398, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3.1 HR Estimation", + "text_level": 1, + "bbox": [ + 76, + 821, + 236, + 835 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For remote HR estimation, all five datasets (i.e., UBFC [1], PURE [52], BUAA [68], VIPL [36], and V4V [47]) are used to evaluate the model generalization. Specifically, four datasets are used for training, and then the other dataset is", + "bbox": [ + 75, + 839, + 470, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "used to test. Firstly, we perform the iPhys $^2$ to evaluate three traditional method (i.e., GREEN [59], CHROM [9], and POS [63]) as shown in Tab. 1. Then, we reproduce five popular DL-based methods (i.e., DeepPhys [6], TSCAN [25], Rhythmnet [36], and BVPNet [7], and DualGAN [28]) for comparison on the same setting. Subsequently, we directly applied the four DG methods (i.e., AD [12], GroupDRO [38], Coral [53], VReX [18], and NCDG [57]) to Rhythmnet with using DeepDG $^3$ .", + "bbox": [ + 496, + 476, + 890, + 611 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The performance of traditional methods. As shown in Tab. 1, traditional algorithms show relatively acceptable results on some simple scenes, such as BUAA and UBFC, but very poor results on other datasets. It is difficult for these traditional algorithms to extract the physiological information on the V4V dataset, which shows that the traditional methods cannot deal with the noise caused by the complex environment and non-rigid motions.", + "bbox": [ + 496, + 612, + 890, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The performance of DL-based methods. For DL-based methods, DeepPhys and TS-CAN even perform worse than traditional methods. Rhythmnet, BVPNet, and Dual-GAN perform much better than both other DL-based methods and traditional methods, which suggests that STMap is the more robust input of CNN than the aligned video. Compared with Rhythmnet, BVPNet and Dual-GAN have not significant improvement on the MSDG protocol. Because BVP-", + "bbox": [ + 496, + 732, + 890, + 853 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "2https://github.com/danmcduff/iphys-toolbox", + "bbox": [ + 514, + 862, + 857, + 875 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "3https://github.com/jindongwang/transferlearning/tree/master/code/DeepDG", + "bbox": [ + 503, + 876, + 888, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "18594", + "bbox": [ + 480, + 925, + 519, + 936 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/403e0e3f206a2ec8449726605cdbd6b4b582d08ac2bfaed5d48557206a3acdc0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TargetMethodLF-(u.n)HF-(u.n)LF/HFHR-(bpm)
MAE↓RMSE↓r↑MAE↓RMSE↓r↑MAE↓RMSE↓r↑MAE↓RMSE↓r↑
UBFCGREEN [59]0.23550.28410.09240.23550.28410.09240.66950.95120.04678.01849.17760.3634
CHROM [9]0.22210.28170.06980.22210.28170.06980.67081.05420.10547.22918.92240.5123
POS [63]0.23640.28610.13590.23640.28610.13590.65150.95350.13457.35398.04020.4923
Baseline0.06210.08130.18730.06210.08130.18730.19850.26670.30435.15427.46720.8165
NEST0.05970.07820.20170.05970.07820.20170.21380.28240.31794.74716.88760.8546
PUREGREEN [59]0.25390.30020.03260.25390.30020.03260.65250.89320.041710.324714.26930.4952
CHROM [9]0.20960.27510.10590.20960.27510.07590.54040.82660.11739.791412.75680.3732
POS [63]0.19590.25710.16840.19590.25710.16840.53730.8460.14339.827313.44140.3432
Baseline0.06710.09230.60460.06710.09230.60460.28640.41840.55268.254211.17650.6832
NEST0.06350.08740.64220.06350.08740.64220.22550.35050.57347.688910.47830.7255
BUAAGREEN [59]0.34720.39510.08710.34720.39510.08710.64530.86320.09215.82317.98820.5624
CHROM [9]0.37860.32370.06820.37860.32370.06820.68130.88360.07156.09348.29380.5165
POS [63]0.31980.37620.09620.31980.37620.09620.62750.84240.11275.04077.11980.6374
Baseline0.14510.16810.28910.14510.16810.28910.55640.69040.29143.78526.32370.7492
NEST0.14360.16650.29550.14360.16650.29550.55140.68840.30043.37235.88060.7647
", + "bbox": [ + 86, + 87, + 879, + 359 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Net and Dual-GAN consider the performance of the single domain instead of focusing on solving domain shift. Therefore, Rhythmnet is also selected as a baseline to further analyze the generalization on MSDG protocol.", + "bbox": [ + 75, + 396, + 468, + 455 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The performance of DG methods. To further improve the generalization of the model, we apply the existing popular DG methods to baseline (Rhythmnet) as shown in Tab. 1. There is a slight model degradation of AD [12], Coral [53], and GroupDRO [38] on some target datasets. The reason behind this is that these algorithms employ inaccurate domain labeling, specifically, a dataset often with different acquisition devices, lighting, and scene cannot be considered as the same domain. On the contrary, VREx [18] and NCDG [57] can slightly improve the performance of Rhythmnet without utilizing the domain label. Inaccurate domain division will make it difficult for the DG algorithm to effectively align the feature distribution across different domains, which is termed as agnostic domain generalization.", + "bbox": [ + 75, + 458, + 468, + 681 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Without the domain labels, NEST experimentally manifested significant improvement on MSDG protocol. For the target dataset BUAA, the results without NEST-CM $\\mathcal{L}_{CM}$ show the best performance. Because NEST-CM may cause the model degradation with activating meaningless features at the instance level, especially for the BUAA with the slight domain shift.", + "bbox": [ + 75, + 684, + 468, + 789 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3.2 HRV Estimation", + "text_level": 1, + "bbox": [ + 76, + 814, + 246, + 828 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also conduct experiments for HRV (i.e., LF, HF, and LF/HF) and HR measurement on the PURE, UBFC, and BUAA. The reason for choosing only these three datasets is that VIPL and V4V do not have reliable ground-truth", + "bbox": [ + 75, + 839, + 468, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a4e00464363c20a10874ff9629ffb384160c2b8d350b33772a2f6ca2612677af.jpg", + "table_caption": [ + "Table 2. HRV and HR estimation results on the MSDG protocol." + ], + "table_footnote": [], + "table_body": "
MethodPUREBUAA
MAE↓RMSE↓r↑MAE↓RMSE↓r↑
GREEN [59]10.3214.270.525.827.990.56
CHROM [9]9.7912.760.376.098.290.51
POS [63]9.8213.440.345.047.120.63
Baseline7.5410.130.732.973.280.76
NEST6.079.060.762.562.730.78
", + "bbox": [ + 509, + 393, + 879, + 518 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. HR estimation results on SSDG protocol by training on the UBFC-rPPG.", + "bbox": [ + 500, + 522, + 890, + 550 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "BVP signals, as mentioned in subsection 4.1. We also perform three tradition method (GREEN [59], CHMROM [9], POS [63]) by iPhys for comparison with our method as shown in Tab. 2. With a more robust NEST presentation distribution, our method also can get better performance than baseline and other traditional methods in all metrics.", + "bbox": [ + 498, + 569, + 890, + 659 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Single-Source Domain Generalization", + "text_level": 1, + "bbox": [ + 500, + 689, + 823, + 705 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For a worst-case scenario, the model is trained on only one dataset to generalize to the unseen domains, namely single-source domain generalization (SSDG). Based on this, we select UBFC-rPPG as the source domain and then select PURE and BUAA as the target domain to further evaluate the generalization of NEST as shown in Tab. 3. As we can see, NEST achieves desired performance by training on only one dataset. The reason is that: (1) NEST-CM can avoid the appearance of abnormal NEST; (2) NEST-TA can find the shared and robust feature of the samples within the same label; (3) NEST-DM can distinguish the boundaries between different samples to improve the generalization.", + "bbox": [ + 496, + 719, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "18595", + "bbox": [ + 480, + 925, + 517, + 936 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/768b69aab91735da2d8ea8ccadb35b656a388ba27d494461f8a0ba6129753315.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodSD↓MAE↓RMSE↓r↑
SAMC [58]18.015.921.00.11
POS [63]15.311.517.20.30
CHROM [9]15.111.416.90.28
I3D [3]15.912.015.90.07
DeepPhy [6]13.611.013.80.11
BVPNet [8]7.755.347.850.70
RhythmNet [36]8.115.308.140.76
CVD [37]7.925.027.970.79
Physformer [70]7.744.977.790.78
Dual-GAN [28]7.634.937.680.81
Baseline8.045.218.070.77
NEST7.494.767.510.84
", + "bbox": [ + 86, + 88, + 455, + 291 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. HR estimation results by our method and several state-of-the-art methods on the VIPL-HR database.", + "bbox": [ + 76, + 296, + 467, + 324 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/846f917949be5f611854fab9d3fdc219c7ceb03be288304b79b6f0faadc4342d.jpg", + "image_caption": [ + "Figure 5. Impacts of the hyperparameter (a) $\\rho$ , (b) $\\sigma$ , and (c) $\\tau$ of our proposed method." + ], + "image_footnote": [], + "bbox": [ + 78, + 327, + 467, + 402 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Intra-Dataset Testing on VIPL", + "text_level": 1, + "bbox": [ + 76, + 448, + 346, + 465 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our proposed method can also improve the effectiveness of the model on intra-dataset testing by reducing the harmful effects of abnormal disturbances. Following [28, 36, 37, 70], we test NEST by using a 5-fold cross-validation protocol on VIPL-HR. Our method compared with ten baseline methods, including the traditional methods (SAMC [58], POS [63] and CHROM [9]) and the DL-based methods (I3D [3], DeepPhy [6], RhythmNet [36], CVD [37], BVP-Net [7], Physformer [70], and Dual-GAN [28]). The results of these methods are directly from [28, 70] as shown in Tab. 4.", + "bbox": [ + 75, + 472, + 468, + 637 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown in Tab. 4, the proposed methods outperform all the methods under all metrics. Compared with the baseline, it also gets a noticeable promotion, i.e., 0.4 in MAE and 0.53 in RMSE. It is worth noting that the VIPL dataset has multiple complex scenes and recording devices and cannot be considered as a single domain, termed the agnostic domain problem. This proves that NEST can effectively alleviate the domain shift without domain labels.", + "bbox": [ + 75, + 638, + 468, + 758 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. Further Analysis", + "text_level": 1, + "bbox": [ + 76, + 770, + 243, + 787 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conduct further experiments on MSDG protocol (testing on UBFC with training on the other four datasets). Impact of $\\rho$ in NEST-CM. The hyperparameter $\\rho$ determines how many bases need to be strengthened by NEST-CM $\\mathcal{L}_{CM}$ . As illustrated in Fig. 5 (a), NEST could achieve smaller RMSE when $\\rho = 0.1$ , indicating that the bases with small eigenvalue require more attention and fine-tuning.", + "bbox": [ + 75, + 794, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/efa6d736dca4e50d7da261bf7c761e53794418d131d14c6ef8cc8dcc547bfa9e.jpg", + "image_caption": [ + "Figure 6. The pearson's correlation coefficient of the NEST representation of the samples with different labels for the anchor HR 60." + ], + "image_footnote": [], + "bbox": [ + 506, + 89, + 880, + 220 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact of $\\sigma$ in NEST-TA. The hyperparameter $\\sigma$ tradeoffs the contribution of each selected sample to form the synthesized prototype. As illustrated in Fig. 5, NEST could achieve the best performance with $\\sigma = 5$ , which shows that the appropriate fusion of the nearby sample is conducive to improving model generalization. The specific reason behind this is that: (1) the synthesized prototype based on multi-domain samples can be used to enhance the domain invariant feature by NEST-TA; (2) NEST-TA can effectively aggregate the label relationship into NEST presentation distribution, which is performed by improving the similarity of NEST representation between the adjacent label samples.", + "bbox": [ + 496, + 267, + 890, + 448 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To further clarify the second reason, we calculate the Pearson's correlation coefficient of the NEST representation of the samples with different labels for the anchor HR 60, as shown in Fig. 6. It is obvious that the distribution changes with NEST-TA loss are more smooth across the continuous HR labels, which allows OOD samples with slight domain shifts to predict nearby heart rates around the ground truth HR value instead of other abnormal HR values.", + "bbox": [ + 496, + 450, + 892, + 571 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact of $\\tau$ in NEST-DM. The hyperparameter $\\tau$ determines the pushing strength of the different samples. With the smallest $\\tau$ , NEST-DM $\\mathcal{L}_{DM}$ has the overemphasized constraints to distinguish the NEST representation between different samples, which causes a harmful impact on the discriminant ability of the model. Overall, NEST-DM $\\mathcal{L}_{DM}$ is an effective way to improve the generalization of model.", + "bbox": [ + 496, + 573, + 890, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 695, + 617, + 710 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper presents the NEural SStructure (NEST) modeling, a novel framework that addresses the domain generalization problem in rPPG measurement for the first time. Our framework contains three regularization terms defined in the view of neural structure, which improves the generalization capacity in the different aspects. All three terms do not rely on domain labels, which are especially suitable in rPPG, as defining domains is difficult in this task. To compare the generalization of the existing algorithm and our method, we establish the first DG protocol for this task. We experimentally manifested the substantial improvement of our method on both MSDG and SSDG protocols.", + "bbox": [ + 496, + 719, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "18596", + "bbox": [ + 480, + 925, + 517, + 936 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Serge Bobbia, Richard Macwan, Yannick Benezeth, Alamin Mansouri, and Julien Dubois. Unsupervised skin tissue segmentation for remote photoplethysmography. Pattern Recognition Letters, 124:82-90, 2019.", + "[2] Adrian Bulat and Georgios Tzimiropoulos. How far are we from solving the 2d & 3d face alignment problem? (and a dataset of 230,000 3d facial landmarks). In Proc. ICCV, 2017.", + "[3] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proc. IEEE CVPR, pages 6299-6308, 2017.", + "[4] Beidi Chen, Tri Dao, Eric Winsor, Zhao Song, Atri Rudra, and Christopher Ré. Scatterbrain: Unifying sparse and low-rank attention. NeurIPS, 34:17413-17426, 2021.", + "[5] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, pages 1597-1607. PMLR, 2020.", + "[6] Weixuan Chen and Daniel McDuff. Deep learning: Video-based physiological measurement using convolutional attention networks. In Proc. ECCV, pages 349-365, 2018.", + "[7] Abhijit Das, Hao Lu, Hu Han, Antitza Dantcheva, Shiguang Shan, and Xilin Chen. Bvpnet: Video-to-bvp signal prediction for remote heart rate estimation. In $FG$ , pages 01-08. IEEE, 2021.", + "[8] Abhijit Das, Hao Lu, Hu Han, Antitza Dantcheva, Shiguang Shan, and Xilin Chen. Bvpnet: Video-to-bvp signal prediction for remote heart rate estimation. In $FG$ , pages 01-08. IEEE, 2021.", + "[9] Gerard De Haan and Vincent Jeanne. Robust pulse rate from chrominance-based rppg. IEEE Trans. Biomed. Eng., 60(10):2878-2886, 2013.", + "[10] Gerard De Haan and Arno Van Leest. Improved motion robustness of remote-ppg by using the blood volume pulse signature. Physiol. Meas., 35(9):1913, 2014.", + "[11] Zhijie Deng, Yucen Luo, and Jun Zhu. Cluster alignment with a teacher for unsupervised domain adaptation. In Proc. CVPR, pages 9944–9953, 2019.", + "[12] Yaroslav Ganin and Victor Lempitsky. Unsupervised domain adaptation by backpropagation. In ICML, pages 1180-1189. PMLR, 2015.", + "[13] Ming-Shiang Chen Gee-Sern Hsu, ArulMurugan Ambikapathi. Deep learning with time-frequency representation for pulse estimation. In Proc. IJCB, pages 642-650, 2017.", + "[14] Min Hu, Fei Qian, Dong Guo, Xiaohua Wang, Lei He, and Fuji Ren. Eta-rppgnet: effective time-domain attention network for remote heart rate measurement. IEEE Trans. on Instrum. and Meas., 70:1-12, 2021.", + "[15] Yunpei Jia, Jie Zhang, Shiguang Shan, and Xilin Chen. Single-side domain generalization for face anti-spoofing. In Proc. CVPR, pages 8484-8493, 2020.", + "[16] Xiang Jiang, Qicheng Lao, Stan Matwin, and Mohammad Havaei. Implicit class-conditioned domain alignment for unsupervised domain adaptation. In ICML, pages 4816-4827. PMLR, 2020." + ], + "bbox": [ + 78, + 114, + 468, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[17] Viktor Kessler, Patrick Thiam, Mohammadreza Amirian, and Friedhelm Schwenker. Pain recognition with camera photoplethysmography. In IEEE IPTA, pages 1-5, 2017.", + "[18] David Krueger, Ethan Caballero, Joern-Henrik Jacobsen, Amy Zhang, Jonathan Binas, Dinghuai Zhang, Remi Le Priol, and Aaron Courville. Out-of-distribution generalization via risk extrapolation (rex). In ICML, pages 5815-5826. PMLR, 2021.", + "[19] Eugene Lee, Evan Chen, and Chen-Yi Lee. Meta-rppg: Remote heart rate estimation using a transductive meta-learner. In ECCV, pages 392–409. Springer, 2020.", + "[20] Alexander Lehner, Stefano Gasperini, Alvaro Marcos-Ramiro, Michael Schmidt, Mohammad-Ali Nikouei Mahani, Nassir Navab, Benjamin Busam, and Federico Tombari. 3d-vfield: Adversarial augmentation of point clouds for domain generalization in 3d object detection. In Proc. CVPR, pages 17295-17304, 2022.", + "[21] Magdalena Lewandowska, Jacek Ruminski, Tomasz Kocejko, and Jedrzej Nowak. Measuring pulse rate with a webcam—a non-contact method for evaluating cardiac activity. In FedCSIS, pages 405-410. IEEE, 2011.", + "[22] Daiqing Li, Junlin Yang, Karsten Kreis, Antonio Torralba, and Sanja Fidler. Semantic segmentation with generative models: Semi-supervised learning and strong out-of-domain generalization. In Proc. CVPR, pages 8300-8311, 2021.", + "[23] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy Hospedales. Learning to generalize: Meta-learning for domain generalization. In Proc. AAAI, 2018.", + "[24] Chuang Lin, Zehuan Yuan, Sicheng Zhao, Peize Sun, Changhu Wang, and Jianfei Cai. Domain-invariant disentangled network for generalizable object detection. In Proce. CVPR, pages 8771-8780, 2021.", + "[25] Xin Liu, Josh Fromm, Shwetak Patel, and Daniel McDuff. Multi-task temporal shift attention networks for on-device contactless vitals measurement. NeurIPS, 33:19400-19411, 2020.", + "[26] Xin Liu, Ziheng Jiang, Josh Fromm, Xuhai Xu, Shwetak Patel, and Daniel McDuff. Metaphys: few-shot adaptation for non-contact physiological measurement. In Proc. HIL, pages 154-163, 2021.", + "[27] Zhenhua Liu, Jizheng Xu, Xiulian Peng, and Ruiqin Xiong. Frequency-domain dynamic pruning for convolutional neural networks. NeurIPS, 31, 2018.", + "[28] Hao Lu, Hu Han, and S Kevin Zhou. Dual-gan: Joint bvp and noise modeling for remote physiological measurement. In Proc. IEEE CVPR, pages 12404-12413, 2021.", + "[29] Yawei Luo, Ping Liu, Liang Zheng, Tao Guan, Junqing Yu, and Yi Yang. Category-level adversarial adaptation for semantic segmentation using purified features. IEEE Trans. Pattern Anal. Mach. Intell., 2021.", + "[30] Fangrui Lv, Jian Liang, Shuang Li, Bin Zang, Chi Harold Liu, Ziteng Wang, and Di Liu. Causality inspired representation learning for domain generalization. In Proc. CVPR, pages 8046-8056, 2022.", + "[31] Lei Ma, Felix Juefei-Xu, Fuyuan Zhang, Jiyuan Sun, Minhui Xue, Bo Li, Chunyang Chen, Ting Su, Li Li, Yang Liu, et al. Deep gauge: Multi-granularity testing criteria for deep learning systems. In Proc. ASE, pages 120-131, 2018." + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "18597", + "bbox": [ + 480, + 926, + 517, + 936 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[32] Daniel McDuff. Camera measurement of physiological vital signs. CSUR, 2021.", + "[33] Daniel McDuff. Applications of camera-based physiological measurement beyond healthcare. In CVSM, pages 165-177. Elsevier, 2022.", + "[34] Daniel McDuff, Sarah Gontarek, and Rosalind W Picard. Improvements in remote cardiopulmonary measurement using a five band digital camera. IEEE Trans. Biomed. Eng., 61(10):2593-2601, 2014.", + "[35] Xuesong Niu, Hu Han, Shiguang Shan, and Xilin Chen. Synrhythm: Learning a deep heart rate estimator from general to specific. In Proc. IEEE ICPR, pages 3580-3585, 2018.", + "[36] Xuesong Niu, Shiguang Shan, Hu Han, and Xilin Chen. Rhythmnet: End-to-end heart rate estimation from face via spatial-temporal representation. IEEE Trans. on Image Process., 29:2409-2423, 2020.", + "[37] Xuesong Niu, Zitong Yu, Hu Han, Xiaobai Li, Shiguang Shan, and Guoying Zhao. Video-based remote physiological measurement via cross-verified feature disentangling. In Proc. ECCV, 2020.", + "[38] Giambattista Parascandolo, Alexander Neitz, Antonio Orvieto, Luigi Gresele, and Bernhard Scholkopf. Learning explanations that are hard to vary. *ICLR2022*, 2022.", + "[39] Kexin Pei, Yinzhi Cao, Junfeng Yang, and Suman Jana. Deepxplore: Automated whitebox testing of deep learning systems. In Proc. SOSP, pages 1-18, 2017.", + "[40] Duo Peng, Yinjie Lei, Munawar Hayat, Yulan Guo, and Wen Li. Semantic-aware domain generalized segmentation. In Proc. CVPR, pages 2594-2605, 2022.", + "[41] Vihari Piratla, Praneeth Netrapalli, and Sunita Sarawagi. Efficient domain generalization via common-specific low-rank decomposition. In ICML, pages 7728-7738. PMLR, 2020.", + "[42] Ming-Zher Poh, Daniel J McDuff, and Rosalind W Picard. Non-contact, automated cardiac pulse measurements using video imaging and blind source separation. Opt. Express, 18(10):10762-10774, 2010.", + "[43] Fengchun Qiao, Long Zhao, and Xi Peng. Learning to learn single domain generalization. In Proc. CVPR, pages 12556-12565, 2020.", + "[44] Ying Qiu, Yang Liu, Juan Arteaga-Falconi, Haiwei Dong, and Abdulmotaleb El Saddik. Evm-cnn: Real-time contactless heart rate estimation from facial video. IEEE Trans. Multimedia, 2019.", + "[45] Attila Reiss, Ina Indlekofer, Philip Schmidt, and Kristof Van Laerhoven. Deep ppg: large-scale heart rate estimation with convolutional neural networks. Sensors, 19(14):3079, 2019.", + "[46] Jiahuan Ren, Zhao Zhang, Richang Hong, Mingliang Xu, Hajun Zhang, Mingbo Zhao, and Meng Wang. Robust low-rank convolution network for image denoising. In Proc. ACM MM, pages 6211-6219, 2022.", + "[47] Ambareesh Revanur, Zhihua Li, Umur A Ciftci, Lijun Yin, and László A Jeni. The first vision for vitals (v4v) challenge for non-contact video-based physiological estimation. In Proc. CVPR workshop, pages 2760-2767, 2021.", + "[48] Shiv Shankar, Vihari Piratla, Soumen Chakrabarti, Siddhartha Chaudhuri, Preethi Jyothi, and Sunita Sarawagi." + ], + "bbox": [ + 78, + 90, + 467, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Generalizing across domains via cross-gradient training. ICLR, 2018.", + "[49] Rencheng Song, Huan Chen, Juan Cheng, Chang Li, Yu Liu, and Xun Chen. Pulsegan: Learning to generate realistic pulse waveforms in remote photoplethysmography. IEEE J-BHI, pages 1-1, 2021.", + "[50] Rencheng Song, Senle Zhang, Chang Li, Yunfei Zhang, Juan Cheng, and Xun Chen. Heart rate estimation from facial videos using a spatiotemporal representation with convolutional neural networks. IEEE Trans. Instrum. Meas., 2020.", + "[51] Radim Špetlík, Vojtech Franc, and Jírí Matas. Visual heart rate estimation with convolutional neural network. In Proc. BMVC, pages 3-6, 2018.", + "[52] Ronny Stricker, Steffen Müller, and Horst-Michael Gross. Non-contact video-based pulse rate measurement on a mobile service robot. In Proc. IEEE ISRHIC, pages 1056-1062, 2014.", + "[53] Baochen Sun and Kate Saenko. Deep coral: Correlation alignment for deep domain adaptation. In ECCV, pages 443-450. Springer, 2016.", + "[54] Bo Sun, Qinglan Wei, Liandong Li, Qihua Xu, Jun He, and Lejun Yu. Lstm for dynamic emotion and group emotion recognition in the wild. In Proc. ACM ICMI, pages 451-457, 2016.", + "[55] Zhaodong Sun and Xiaobai Li. Contrast-phys: Unsupervised video-based remote physiological measurement via spatiotemporal contrast. In Proc. ECCV, 2022.", + "[56] Ling Tang, Wen Shen, Zhanpeng Zhou, Yuefeng Chen, and Quanshi Zhang. Defects of convolutional decoder networks in frequency representation. arXiv preprint arXiv:2210.09020, 2022.", + "[57] Chris Xing Tian, Haoliang Li, Xiaofei Xie, Yang Liu, and Shiqi Wang. Neuron coverage-guided domain generalization. IEEE Trans. on Pattern Anal. Mach. Intell., 2022.", + "[58] Sergey Tulyakov, Xavier Alameda-Pineda, Elisa Ricci, Lijun Yin, Jeffrey F Cohn, and Nicu Sebe. Self-adaptive matrix completion for heart rate estimation from face videos under realistic conditions. In Proc. IEEE CVPR, pages 2396-2404, 2016.", + "[59] Wim Verkruysse, Lars O Svaasand, and J Stuart Nelson. Remote plethysmographic imaging using ambient light. Opt. Express, 16(26):21434-21445, 2008.", + "[60] Tan Wad, Qianru Sun, Sugiri Pranata, Karlekar Jayashree, and Hanwang Zhang. Equivalence and invariance inductive bias for learning from insufficient data. In ECCV, pages 241-258. Springer, 2022.", + "[61] Hao Wang, Euijoon Ahn, and Jinman Kim. Self-supervised representation learning framework for remote physiological measurement using spatiotemporal augmentation loss. In ACM MM, 2022.", + "[62] Jindong Wang, Cuiling Lan, Chang Liu, Yidong Ouyang, Tao Qin, Wang Lu, Yiqiang Chen, Wenjun Zeng, and Philip Yu. Generalizing to unseen domains: A survey on domain generalization. IEEE Trans. on Knowl. Data Eng., 2022.", + "[63] Wenjin Wang, Albertus C den Brinker, Sander Stuijk, and Gerard de Haan. Algorithmic principles of remote ppg. IEEE Trans. Biomed. Eng., 64(7):1479-1491, 2017." + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "18598", + "bbox": [ + 480, + 926, + 519, + 936 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[64] Wenjin Wang, Sander Stuijk, and Gerard De Haan. Exploiting spatial redundancy of image sensor for motion robust rppg. IEEE Trans. Biomed. Eng., 62(2):415-425, 2015.", + "[65] Yunhe Wang, Chang Xu, Shan You, Dacheng Tao, and Chao Xu. Cnnpack: Packing convolutional neural networks in the frequency domain. In NeurIPS, volume 29, 2016.", + "[66] Zijian Wang, Yadan Luo, Ruihong Qiu, Zi Huang, and Mahsa Baktashmotlagh. Learning to diversify for single domain generalization. In Proc. CVPR, pages 834-843, 2021.", + "[67] Zhi-Kuan Wang, Ying Kao, and Chiou-Ting Hsu. Vision-based heart rate estimation via a two-stream cnn. In Proc. IEEE ICIP, pages 3327-3331, 2019.", + "[68] Lin Xi, Weihai Chen, Changchen Zhao, Xingming Wu, and Jianhua Wang. Image enhancement for remote photoplethysmography in a low-light environment. In FG, pages 1-7. IEEE, 2020.", + "[69] Zitong Yu, Wei Peng, Xiaobai Li, Xiaopeng Hong, and Guoying Zhao. Remote heart rate measurement from highly compressed facial videos: an end-to-end deep learning solution with video enhancement. In Proc. IEEE ICCV, pages 151-160, 2019.", + "[70] Zitong Yu, Yuming Shen, Jingang Shi, Hengshuang Zhao, Philip Torr, and Guoying Zhao. Physformer: Facial videobased physiological measurement with temporal difference transformer. IEEE CVPR, 2022.", + "[71] Xiangyu Yue, Yang Zhang, Sicheng Zhao, Alberto Sangiovanni-Vincentelli, Kurt Keutzer, and Boqing Gong. Domain randomization and pyramid consistency: Simulation-to-real generalization without accessing target domain data. In Proc.CVPR, pages 2100-2110, 2019." + ], + "bbox": [ + 78, + 90, + 467, + 515 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "18599", + "bbox": [ + 480, + 926, + 517, + 936 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Neuron Structure Modeling for Generalizable Remote Physiological Measurement/0c500895-2bd8-4cea-80ae-7bc4fd9b8ba5_model.json b/2023/Neuron Structure Modeling for Generalizable Remote Physiological Measurement/0c500895-2bd8-4cea-80ae-7bc4fd9b8ba5_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d75656cfd5be1ca2a15dda26eb41086cffb86d33 --- /dev/null +++ b/2023/Neuron Structure Modeling for Generalizable Remote Physiological Measurement/0c500895-2bd8-4cea-80ae-7bc4fd9b8ba5_model.json @@ -0,0 +1,2422 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.144, + 0.131, + 0.828, + 0.175 + ], + "angle": 0, + "content": "Neuron Structure Modeling for Generalizable Remote Physiological Measurement" + }, + { + "type": "text", + "bbox": [ + 0.243, + 0.203, + 0.735, + 0.221 + ], + "angle": 0, + "content": "Hao LU\\(^{1,2}\\), Zitong YU\\(^{3}\\), Xuesong NIU\\(^{4}\\), Yingcong CHEN\\(^{1,2,*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.222, + 0.759, + 0.24 + ], + "angle": 0, + "content": "1The Hong Kong University of Science & Technology (Guangzhou)," + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.24, + 0.887, + 0.293 + ], + "angle": 0, + "content": "\\(^{2}\\)The Hong Kong University of Science & Technology, \\(^{3}\\)Great Bay University, \\(^{4}\\)Kuaishou Technology hlu585@connect.hkust-gz.edu.cn, zitong.yu@ieee.org, nxsedson@gmail.com, yingcongchen@ust.hk" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.327, + 0.314, + 0.343 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.359, + 0.474, + 0.707 + ], + "angle": 0, + "content": "Remote photoplethysmography (rPPG) technology has drawn increasing attention in recent years. It can extract Blood Volume Pulse (BVP) from facial videos, making many applications like health monitoring and emotional analysis more accessible. However, as the BVP signal is easily affected by environmental changes, existing methods struggle to generalize well for unseen domains. In this paper, we systematically address the domain shift problem in the rPPG measurement task. We show that most domain generalization methods do not work well in this problem, as domain labels are ambiguous in complicated environmental changes. In light of this, we propose a domain-label-free approach called NEuron SStructure modeling (NEST). NEST improves the generalization capacity by maximizing the coverage of feature space during training, which reduces the chance for under-optimized feature activation during inference. Besides, NEST can also enrich and enhance domain invariant features across multi-domain. We create and benchmark a large-scale domain generalization protocol for the rPPG measurement task. Extensive experiments show that our approach outperforms the state-of-the-art methods on both cross-dataset and intra-dataset settings. The codes are available at https://github.com/LuPaoPao/NEST." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.717, + 0.21, + 0.733 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.742, + 0.47, + 0.879 + ], + "angle": 0, + "content": "Physiological signals such as heart rate (HR), and heart rate variability (HRV), respiration frequency (RF) are important body indicators that serve not only as vital signs but also track the level of sympathetic activation [17, 33, 54]. Traditional physiological measurements, such as electrocardiograms, heart rate bands, and finger clip devices, have high accuracy. However, they are costly, intrusive, and uncomfortable to wear for a long time. Remote photoplethysmography (rPPG) can extract blood volume pulse (BVP)" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.326, + 0.576, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.582, + 0.326, + 0.655, + 0.467 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.658, + 0.326, + 0.731, + 0.467 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.736, + 0.326, + 0.81, + 0.467 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.819, + 0.326, + 0.892, + 0.467 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.57, + 0.468, + 0.813, + 0.48 + ], + "angle": 0, + "content": "(a) Typical samples from different datasets" + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.48, + 0.871, + 0.626 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.543, + 0.628, + 0.869, + 0.638 + ], + "angle": 0, + "content": "(b) The performance of different methods on DG protocol" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.638, + 0.894, + 0.706 + ], + "angle": 0, + "content": "Figure 1. (a) Typical samples from different publicly rPPG datasets: VIPL-HR [36], V4V [47], UBFC-rPPG [1], BUAA [68], PURE [52]. (b) The performance of different methods on DG protocol (test on the UBFC-rPPG dataset with training on the VIPL, V4V, PURE, and BUAA)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.715, + 0.894, + 0.852 + ], + "angle": 0, + "content": "from face video, which analyzes the periodic changes of the light absorption of the skin caused by heartbeats. Then various physiological indicators (such as HR and HRV) can be calculated based on BVP signals [37, 69, 70]. With nonintrusion and convenience, the rPPG-based physiological measurement method can only use an ordinary camera to monitor physiological indicators and gradually become a research hotspot in the computer vision field [7, 28, 32, 36, 44, 49]." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Traditional rPPG measurement methods include signal blind decomposition [21,34,42] and color space transformation [9,59,63]. These approaches rely on heartbeat-related" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.222, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.925, + 0.519, + 0.937 + ], + "angle": 0, + "content": "18589" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.198 + ], + "angle": 0, + "content": "statistical information, only applicable in constrained environments. In recent years, deep learning (DL) based approaches [6, 19, 25, 28, 36, 44, 50, 69, 70] have shown their great potentials in rPPG measurement. By learning dedicated rPPG feature representation, these methods achieve promising performance in much more complicated environments [1, 36, 47, 52, 68]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.2, + 0.47, + 0.382 + ], + "angle": 0, + "content": "However, deep learning methods suffer from significant performance degradation when applied in real-world scenarios. This is because most training data are captured in lab environments with limited environmental variations. With domain shifts (e.g., different illumination, camera parameters, motions, etc.), these models may struggle to generalize for the unseen testing domain. To validate this, we conduct a cross-dataset evaluation shown in Fig. 1(b). As shown, all DL-based methods do not work well in this evaluation. Furthermore, it is worth noting that DeepPhys [6] and TS-CAN [25] even perform inferior to traditional approach POS [63]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.384, + 0.471, + 0.656 + ], + "angle": 0, + "content": "To improve the performance on the unseen domains, one common practice is to incorporate Domain Generalization (DG) approaches, e.g., encouraging intermediate features to be domain-invariant [12, 18, 30, 38, 57, 62]. However, as shown in Fig. 1 (b) (AD [12] and GroupDRO [38]), the improvements are still quite limited. One reason is that existing DG methods assume that domain labels can be clearly defined (e.g., by data source). Unfortunately, different background scenes, acquisition devices, or even identities could cause very different data distributions in rPPG measurement tasks, and such distribution discrepancy may exist within or cross datasets. In this case, explicitly defining domains is very difficult, and simply treating one dataset as one domain may lead to inferior performance. This is termed as agnostic domain generalization. Besides this, as physiological cues are usually much more subtle than the various noise, the model may overfit in the source domain with the limited training data." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.471, + 0.902 + ], + "angle": 0, + "content": "In this paper, we propose the NEural SStructure modeling (NEST), a principled solution to the abovementioned problems. The main idea of NEST is to narrow the under-optimized and redundant feature space, align domain invariant features, and enrich discriminative features. Our intuition is as follows. Neural structure refers to the channel activation degree in each convolution layer, which reveals the discriminative feature combination for the specific sample. As the limited variation in a certain restricted domain, there are some spaces that the model seldomly is optimized. Out-of-distribution (OOD) samples may cause abnormal activation in these spaces, which may lead to performance degeneration. Therefore, we regularize the neural structure to encourage the model to be more well-conditioned to avoid abnormal activation caused by OOD samples. Specifically, we propose the NEural SStructure Coverage Maxi" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.272 + ], + "angle": 0, + "content": "mization (NEST-CM) that encourages all neural spaces to be optimized during training, reducing the chance of abnormal activation during testing. Secondly, we propose the NEural SStructure Targeted Alignment (NEST-TA) that encourage network suppresses domain variant feature by comparing the samples with the similar physiological information. Thirdly, we propose the NEural SStructure Diversity Maximization (NEST-DM) to enrich discriminative features against unseen noise. It should be noted that our approach does not rely on domain labels, which is more applicable in the rPPG measurement task. To summarize, our contributions are listed as follows:" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.273, + 0.892, + 0.317 + ], + "angle": 0, + "content": "1. We are the first to study the domain shift problem in rPPG measurement, which introduces a new challenge, agnostic domain generalization." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.318, + 0.892, + 0.378 + ], + "angle": 0, + "content": "2. We propose the NEural SStructure modeling to alleviate domain shift, which is performed by narrowing the under-optimized feature space, and enhancing and enriching domain invariant features." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.379, + 0.892, + 0.439 + ], + "angle": 0, + "content": "3. We establish a large-scale domain generalization (DG) benchmark for rPPG measurement, which is the first DG protocol in this task. Extensive experiments in this dataset show the superiority of our approach." + }, + { + "type": "list", + "bbox": [ + 0.498, + 0.273, + 0.892, + 0.439 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.442, + 0.638, + 0.457 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.463, + 0.81, + 0.479 + ], + "angle": 0, + "content": "2.1. Remote Physiological Measurement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.483, + 0.892, + 0.799 + ], + "angle": 0, + "content": "Remote physiological measurement aims to estimate HR and HRV values by analyzing the chrominance changes in the skin. Traditional methods mainly perform color space transformations or signal decomposition to obtain the BVP signal with high SNR [9, 10, 42, 64]. These methods are designed manually for certain lighting scenes with slight motion and can not generalize to the complex environment in the wild. With powerful modeling capacity, the deep learning (DL) model has been successfully used in remote physiological estimation [6, 13, 35, 36, 51, 67, 69]. Among these methods, various model structures are tailored to extract BVP signals from face videos by aggregating spatial and temporal information [6, 37, 51, 69, 70]. Besides, many hand-designed feature maps are used to reduce the computational burden of the network by removing the irrelevant background [13, 28, 36, 45, 50]. Some work utilizes the distribution of the target domain to improve the model performance [19, 26], which is not practical without touching the unseen domain in advance. Differently, we study the model generalization on domain generalization protocol for rPPG measurement." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.806, + 0.712, + 0.822 + ], + "angle": 0, + "content": "2.2. Domain generalization" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Domain generalization (DG), i.e., out-of-distribution generalization, aims to generalize to the unseen domain by training a model on multiple or single source domains [62]. The main solution can be categorized into data manipulation, representation learning, and meta-learning. Data" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.925, + 0.519, + 0.937 + ], + "angle": 0, + "content": "18590" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.089, + 0.625, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.627, + 0.09, + 0.885, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.435, + 0.895, + 0.492 + ], + "angle": 0, + "content": "Figure 2. An overview of the proposed method. (a) Spatial-temporal map (STMap) is extracted from face video [36]. (b) NEST representation is calculated from the samples of the different domains. (c) The under-optimized space in NEST presentation distribution is narrowed by NEST-CM. (d) NEST-TA align the feature of multi domain at instance-level. (e) The distance of different samples are pulled away by NEST-DM." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.5, + 0.473, + 0.759 + ], + "angle": 0, + "content": "The generalization of the model, which can remove irrelevant style information by data generation and data augmentation [48, 71]. The main goal of representation learning is to learn domain-invariant features by alleviating covariate shift across domain [12, 18, 38, 57]. By simulating domain shift on the virtual dataset, meta-learning seeks transferable and shareable knowledge from multiple domain [23, 30]. Unlike common tasks such as classification, segmentation, and detection [15, 20, 22, 24, 40], rPPG measurement tasks do not have explicit boundaries between different domains, i.e., agnostic domain generalization. Recently, some single source domain generalization (SSDG) methods focus on the worst-case formulation without using the domain label [43, 57, 66]. Different from all the methods above, we tackle the DG problem of regression task in view of neuron structure modeling." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.775, + 0.17, + 0.79 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.801, + 0.267, + 0.815 + ], + "angle": 0, + "content": "3.1. Overall Framework" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.471, + 0.903 + ], + "angle": 0, + "content": "In this work, we propose the NEuron SStructure modeling (NSET), without using domain label, to alleviate the performance degradation caused by domain shift as illustrated in Fig. 2. Specifically, we first compressed the face video into the spatial-temporal map (STMap) as the input of" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.5, + 0.895, + 0.713 + ], + "angle": 0, + "content": "CNN refer to [36] as shown in Fig. 2 (a). Sequentially, we define an Activated NEuron SStructure (NEST) representation that describes the activated degree of different channels of the network as shown in Fig. 2 (b). To alleviate the domain shift, we model a unified NEST distribution by three regularizations: (1) NEuron SStructure Coverage Maximization (NEST-CM) is used to narrow under-optimized space that is redundant and unstable as shown in Fig. 2 (c); (2) NEuron SStructure Targeted Alignment (NEST-TA) is introduced to align multi-domain feature distribution at instance-level as shown in Fig. 2 (d). (3) Neuron Structure Diversity Maximization (NEST-DM) loss to force the network to explore more features with discriminant ability as shown in Fig. 2 (e)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.725, + 0.759, + 0.742 + ], + "angle": 0, + "content": "3.2. NEuron SStructure Modeling" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.903 + ], + "angle": 0, + "content": "NEuron SStructure Representation. For the rPPG measurement task, the convolution kernels are adaptively activated to extract physiological features against various noise distribution [27, 56, 65]. However, the facial physiological signals are feeble and can be easily affected by various unseen noises, which make the filters activated abnormally. Therefore, the NEuron SStructure (NEST) representation is defined to explore how domain shift affects the model robustness in view of activated feature distribution. Specifically, we firstly average the feature map" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.926, + 0.518, + 0.938 + ], + "angle": 0, + "content": "18591" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.434, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.147, + 0.192, + 0.378, + 0.201 + ], + "angle": 0, + "content": "(a) The expectation of different base projection" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.204, + 0.275, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.299, + 0.275, + 0.319 + ], + "angle": 0, + "content": "(b) The projection distribution of base with highest eigenvalue" + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.204, + 0.446, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.289, + 0.299, + 0.45, + 0.319 + ], + "angle": 0, + "content": "(c) The projection distribution of base with lowest eigenvalue" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.324, + 0.47, + 0.366 + ], + "angle": 0, + "content": "Figure 3. (a) The expectation of different base projection. (b) The distribution of base with highest eigenvalue. (c) The distribution of base with lowest eigenvalue." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.379, + 0.468, + 0.517 + ], + "angle": 0, + "content": "\\(m_{j}\\in R^{W\\times H\\times C_{j}}\\) over the spatial dimensions in the end of \\(j\\)-th layer to get the preliminary statistical presentation \\(\\bar{o}_j\\in R^{C_j}\\) of the neuron structure, where \\(C_j\\) is the channel number of the \\(j\\)-th layer. Following that, the preliminary presentation \\(\\bar{o}_j\\) is fed into the max-min normalized to get the NEST representation \\(o_{j}\\in R_{j}^{C}\\) of the \\(j\\)-th layer. The full NEST representation of \\(i\\)-th sample \\(x_{i}\\) is defined as \\(O_{i} = \\{o_{i,1},o_{i,2},\\dots,o_{i,U}\\} \\in R^{(C_{1} + C_{2} + \\ldots +C_{U})}\\), where \\(U\\) is the number of selected layer." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.519, + 0.468, + 0.624 + ], + "angle": 0, + "content": "NEuron SStructure Coverage Maximization. Existing works show that the OOD samples may activate abnormal neuron structures (feature combination) that seldom be trained with the source data so that the model's performance degrades drastically [31, 39, 57]. To investigate this phenomenon in rPPG measurement task, we analyze the optimized degree of different feature combinations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.627, + 0.468, + 0.868 + ], + "angle": 0, + "content": "Specifically, we decompose the NEST representation of the last layer into orthogonal bases with Singular Value Decomposition (SVD). Here, small eigenvalues indicate that training samples cover a small space in these corresponding bases. Thus, unseen target samples are more likely to lay out of the well-optimized space, causing a larger domain shift. To validate this, we plot the means of source and target samples on these bases in Fig. 3(a). As shown, the mean differences are larger for bases of smaller eigenvalues. To make it clearer, We further visualize the data distribution on the first and last bases in Fig. 3 (b) and (c). As shown, the distribution difference in the last base is much more significant than that of the first one. All these suggest that for the NEST representation, components of small eigenvalues are less optimized during training, and domain shifts are more likely to appear there." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.468, + 0.901 + ], + "angle": 0, + "content": "In sight of this, we propose the NEuron SStructure Coverage Maximization (NEST-CM) regularization that improves" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.089, + 0.888, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.532, + 0.155, + 0.859, + 0.169 + ], + "angle": 0, + "content": "Figure 4. The detail of SVD for the \\(j\\)-th layer in CNN." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.171, + 0.892, + 0.337 + ], + "angle": 0, + "content": "the eigenvalues of the last few bases. This could narrow the under-optimized space as shown in Fig. 2 (c). For a training batch with \\(N\\) samples, NEST representation set of \\(j\\)-th layer \\(B_{j} = \\{o_{1,j}; o_{2,j}; \\ldots; o_{N,j}\\} \\in R^{N * C_{j}}\\) is decomposed to \\(S_{j} \\cdot V_{j} \\cdot D_{j}\\) by SVD method as shown in Fig. 4, i.e., \\(B_{j} = S_{j} \\cdot V_{j} \\cdot D_{j}\\), where \\(V_{j} \\in R^{N * C_{j}}\\) and \\(N > C_{j}\\). The bases with small eigenvalue values are rare and under-optimized during the training stage. To reduce the occurrence probability of under-optimized NEST caused by domain shift, we stimulate under-optimized bases by maximizing the corresponding eigenvalues:" + }, + { + "type": "equation", + "bbox": [ + 0.573, + 0.351, + 0.892, + 0.393 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {C M} = - \\mathbb {E} _ {j \\in U} \\frac {\\sum_ {i = 1} ^ {C ^ {\\prime}} \\lambda_ {i , j} ^ {\\prime}}{\\sum_ {i = 1} ^ {C} \\lambda_ {i , j}}, \\lambda_ {i, j} ^ {\\prime} < t, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.404, + 0.892, + 0.528 + ], + "angle": 0, + "content": "where \\(\\lambda_{i,j}\\) represent the \\(i\\)-th normalized eigenvalue in \\(j\\)-th layer' diagonal matrix \\(V_{j}\\), and \\(\\lambda_{i,j}^{\\prime}\\) is the eigenvalue less than the threshold \\(t\\). Considering different channel numbers in each layer, we design an adaptive threshold \\(t\\) to select the relatively under-optimized bases, i.e., \\(t = \\frac{\\rho}{C_j}\\), where \\(\\rho\\) is a trade-off parameter. Besides, \\(C_j^\\prime\\) is the number of eigenvalues of \\(j\\)-th layer under the threshold, and \\(U\\) is the number of layers." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.528, + 0.892, + 0.678 + ], + "angle": 0, + "content": "NEuron SStructure Targeted Alignment. Besides the NEST-CM introduced above, we further improve the generalization ability by aligning the NEST representation of different domains. Domain invariant learning has demonstrated its effectiveness in DG problems [12, 18, 30, 38, 57, 62]. Some recent research [11, 16, 29, 60] further shows that aligning domain features within the same labels could lead to better performance. However, labels in rPPG are continuous, and it is hard to find samples with identical labels in a training batch." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.68, + 0.892, + 0.902 + ], + "angle": 0, + "content": "To cope with this, we propose a novel regularization term called NEuron SStructure Alignment (NEST-TA). NEST-TA builds a synthesized prototype \\( O_{i}^{p} \\) with the same physiological information as the given targeted sample \\( O_{i} \\). Then, the feature within the same labels can be aligned across domains like [11, 16, 29, 60]. Specifically, we select the top \\( K \\) samples closest to the targeted ground truth \\( y_{i} \\), i.e., \\( [O_{i,1}', \\dots, O_{i,k}', \\dots, O_{i,K}'] \\). The NEST center of the selected samples \\( \\overline{O}_{i}' \\) is calculated by averaging the NEST presentation of selected samples. Then we calculate the direction vectors between the NEST presentation center of the selected samples \\( \\overline{O}_{i}' \\) and the NEST presentation of each sample \\( O_{i,k}' \\). Following that, the NEST center of the selected samples \\( \\overline{O}_{i}' \\) is calibrated to the synthesized prototype \\( O_{i}^{p} \\)," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.925, + 0.519, + 0.936 + ], + "angle": 0, + "content": "18592" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.12 + ], + "angle": 0, + "content": "which is performed by using the Gaussian distribution function to fuse these direction vectors:" + }, + { + "type": "equation", + "bbox": [ + 0.115, + 0.126, + 0.469, + 0.171 + ], + "angle": 0, + "content": "\\[\nO _ {i} ^ {p} = \\overline {{O ^ {\\prime}}} _ {i} + \\sum_ {k = 1} ^ {K} W _ {i, k} * \\frac {y _ {i} - \\bar {y} _ {i} ^ {\\prime}}{y _ {i , k} ^ {\\prime} - \\bar {y} _ {i} ^ {\\prime}} \\left(O _ {i, k} ^ {\\prime} - \\overline {{O ^ {\\prime}}} _ {i}\\right), \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.165, + 0.171, + 0.377, + 0.206 + ], + "angle": 0, + "content": "\\[\nW _ {i, k} = \\frac {G (y _ {i , k} - \\overline {{y}} _ {i} ^ {\\prime} , \\sigma)}{\\sum_ {k = 1} ^ {K} G (y _ {i , k} - \\overline {{y}} _ {i} ^ {\\prime} , \\sigma)},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.215, + 0.471, + 0.332 + ], + "angle": 0, + "content": "where \\( G(-, \\sigma) \\) is the Gaussian distribution with the mean of zero and the standard deviation of \\( \\sigma \\), i.e., \\( G(y_{i,k} - \\overline{y}_i', \\sigma) = \\frac{1}{\\sqrt{2\\pi}} e^{-\\frac{(y_{i,k} - \\overline{y}_i')^2}{2\\sigma^2}} \\), where \\( \\sigma \\) is a hyper-parameter. To align the feature distribution across different domains, NEST-TA uses the cosine similarity \\( Cos(-, -) \\) to reduce the variant features between targeted samples \\( O_i \\) and prototype sample \\( O_i^p \\):" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.339, + 0.469, + 0.357 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {T A} = \\mathbb {E} _ {i \\in N} \\operatorname {C o s} \\left(O _ {i}, O _ {i} ^ {p}\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.368, + 0.47, + 0.533 + ], + "angle": 0, + "content": "NEuron SStructure Diversity Maximization. Existing works elaborate that it is effective to explore more features with discriminant information to improve the generalization [4,30,41,46]. Motivated by this, NEuron SStructure Diversity Maximization (NEST-DM) is introduced to increase the distinction between the NEST presentation of different samples as shown in Fig. 2 (e). By doing this, the network is forced to explore more features with physiological information in a contrastive learning way, which can further improve model discrimination. NEST-DM \\(\\mathcal{L}_{DM}\\) can be represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.543, + 0.469, + 0.585 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {D M} = - \\mathbb {E} _ {i \\in N} \\log \\left(\\frac {e ^ {C o s \\left(O _ {i} , \\widehat {O} _ {i}\\right) / \\tau}}{\\sum_ {j = 1 (i \\neq j)} ^ {N} e ^ {C o s \\left(O _ {i} , \\widehat {O} _ {j}\\right) / \\tau}}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.596, + 0.469, + 0.656 + ], + "angle": 0, + "content": "where \\(\\widehat{O}_i\\) is the NEST representation of augmented sample from \\(O_i\\), and \\(N\\) and \\(\\tau\\) are the batch size and temperature parameter. The augmentation strategies for STMap will be elaborated in section 4.2." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.658, + 0.469, + 0.838 + ], + "angle": 0, + "content": "Model Training. For the rPPG measurement task, we chose two popular losses in this field. Specifically, L1 loss \\(\\mathcal{L}_{L1}\\) and negative Pearson's correlation coefficient \\(\\mathcal{L}_p\\) are used to constrain the predicted heart rate value and the predicted BVP signal respectively [14, 28, 37, 69]. These two losses are also used for augmented data. With the violent confrontation at the beginning of training, \\(\\mathcal{L}_{L1}\\) make network tends to predict the mean of the heart rate value distribution instead of learning the physiological feature like \\(\\mathcal{L}_p\\). Besides, \\(\\mathcal{L}_{NSM}\\), \\(\\mathcal{L}_{NEST-S}\\), and \\(\\mathcal{L}_{NEST-DM}\\) also have meaningless effects on under-trained networks at the beginning." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.839, + 0.469, + 0.929 + ], + "angle": 0, + "content": "To solve the above-illustrated issues, we introduce the adaptation factor \\(\\gamma = \\frac{2}{1 + \\exp(-10r)}\\) and \\(r = \\frac{\\text{inters}_{\\text{current}}}{\\text{inters}_{\\text{total}}}\\) to suppress meaningless constraints at the early training stage, which can accelerate model convergence and improve training stability [12, 15]. The final loss \\(\\mathcal{L}_{\\text{overall}}\\) can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.115, + 0.892, + 0.15 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {o v e r a l l}} = k _ {1} \\mathcal {L} _ {p} + \\gamma \\left(k _ {2} \\mathcal {L} _ {L 1} + k _ {3} \\mathcal {L} _ {C M} \\right. \\tag {5} \\\\ + k _ {4} \\mathcal {L} _ {T A} + k _ {5} \\mathcal {L} _ {D M}), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.161, + 0.787, + 0.177 + ], + "angle": 0, + "content": "where \\(k_{1} - k_{5}\\) are five trade-off parameters." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.19, + 0.632, + 0.207 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.215, + 0.602, + 0.229 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.238, + 0.892, + 0.283 + ], + "angle": 0, + "content": "We elaborately select five datasets (including different motion, camera, and lighting conditions) to curate and benchmark a large-scale DG protocol:" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.283, + 0.892, + 0.403 + ], + "angle": 0, + "content": "VIPL-HR [36] have nine scenarios, three RGB cameras, different illumination conditions, and different levels of movement. It is worth mentioning that the BVP signal and video of the dataset do not match in the time dimension, so \\(\\mathcal{L}_p\\) is not used for this dataset. We used the output of HR head as the evaluation result. Besides, we normalize STMap to 30 fps by cubic spline interpolation to solve the problem of the unstable frame rate of video [28, 50, 70]." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.404, + 0.892, + 0.508 + ], + "angle": 0, + "content": "V4V [47] is designed to collect data with the drastic changes of physiological indicators by simulating ten tasks such as a funny joke, 911 emergency call, and odor experience. It is worth mentioning that There are only heart rate labels in the dataset and no BVP signal labels, so \\(\\mathcal{L}_p\\) is not used for this dataset. We used the output of HR head as the evaluation result." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.51, + 0.892, + 0.613 + ], + "angle": 0, + "content": "BUAA [68] is proposed to evaluate the performance of the algorithm against various illumination. We only use data with illumination greater than or equal to 10 lux because underexposed images require special algorithms that are not considered in this article. We used the output of BVP head as the evaluation for BUAA, PURE, and UBFC-rPPG datasets." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.615, + 0.892, + 0.69 + ], + "angle": 0, + "content": "PURE [52] contains 60 RGB videos from 10 subjects with six different activities, specifically, sitting still, talking, and four rotating and moving head variations. The BVP signals are down-sampled from 60 to 30 fps with cubic spline interpolation to align with the videos." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.69, + 0.892, + 0.736 + ], + "angle": 0, + "content": "UBFC-rPPG [1] containing 42 face videos with sunlight and indoor illumination, the ground-truth BVP signals, and HR values were collected by CMS50E." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.74, + 0.717, + 0.755 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.758, + 0.892, + 0.879 + ], + "angle": 0, + "content": "Training Details. Our proposed method is performed under the Pytorch framework. For the generation of STMap from Video, we use the FAN1 to detect the 2D landmarks of faces [2], and other steps completely follow [36]. STMap in each dataset is sampled with a time window of 256 with step 10. Following [55, 61], we also design spatial and temporal augmentation for STMap, i.e., shuffle each row of STMap and slide the time window. In addition, color jitter and blur" + }, + { + "type": "page_footnote", + "bbox": [ + 0.518, + 0.887, + 0.858, + 0.9 + ], + "angle": 0, + "content": "" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.925, + 0.491, + 0.937 + ], + "angle": 0, + "content": "5" + }, + { + "type": "footer", + "bbox": [ + 0.481, + 0.958, + 0.519, + 0.97 + ], + "angle": 0, + "content": "18593" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.078, + 0.089, + 0.896, + 0.409 + ], + "angle": 0, + "content": "
MethodUBFCPUREBUAAVIPLV4V
MAE↓RMSE↓r↑MAE↓RMSE↓r↑MAE↓RMSE↓r↑MAE↓RMSE↓r↑MAE↓RMSE↓r↑MAE↓RMSE↓r↑MAE↓RMSE↓r↑MAE↓RMSE↓r↑
GREEN [59]8.029.180.3610.3214.270.525.827.990.5612.1818.230.2515.6421.430.06
CHROM [9]7.238.920.519.7912.760.376.098.290.5111.4416.970.2814.9219.220.08
POS [63]7.358.040.499.8213.440.345.047.120.6314.5921.260.1917.6523.220.04
DeepPhys [6]7.828.420.549.3412.560.554.786.740.6912.5619.130.1414.5219.110.14
TS-CAN [25]7.638.250.559.1212.380.574.846.890.6812.3418.940.1614.7719.960.12
Rhythmnet* [36]5.797.910.787.3910.490.773.385.170.848.9712.160.4910.1614.570.34
Dual-GAN* [28]5.557.620.797.2410.270.783.415.230.848.8811.690.5010.0414.440.35
BVPNet* [7]5.437.710.807.2310.250.783.695.480.818.4511.640.5110.0114.350.36
AD*+ [12]5.928.080.767.4210.610.733.495.490.828.4111.710.5310.4714.640.32
GroupDRO*+ [38]5.737.970.787.6910.830.783.415.210.838.3511.670.549.9414.290.36
Coral*+ [53]5.898.040.767.5910.870.723.645.740.808.6811.910.5310.3214.420.32
VREx*+ [18]5.597.680.817.2410.140.783.275.010.868.3711.620.549.8214.160.37
NCDG*+ [57]5.317.560.827.3210.350.773.125.160.858.4711.810.5210.1414.460.34
Baseline* [36]5.797.910.787.3910.490.773.385.170.848.9712.160.4910.1614.570.34
NEST*+ w/o LCM4.927.190.836.919.940.782.794.620.898.2111.560.559.5813.580.38
NEST*+ w/o LTA4.977.280.836.819.770.813.034.940.868.1911.470.559.5213.520.38
NEST*+ w/o LDM4.897.110.846.879.810.792.934.770.888.1411.430.569.4913.490.40
NEST*+4.676.790.866.719.590.812.884.690.897.8611.150.589.2713.790.41
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.412, + 0.895, + 0.468 + ], + "angle": 0, + "content": "Table 1. HR estimation results on MSDG protocol. It is worth noting that we have removed the GRU mechanism in Rhythmnet because it is an unfair comparison with using GRU to aggregate the temporal information of multiple clips instead of only one clip; * means that these methods use the STMap as the input of CNN and use the STMap augmentation strategy as mentioned in section 4.2; + means that these methods are based on baseline (Rhythmnet without GRU)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.477, + 0.473, + 0.688 + ], + "angle": 0, + "content": "augmentation are used for STMap [5]. Then the STMap \\( x \\in R^{25 \\times 256 \\times 3} \\) is resized to \\( x' \\in R^{64 \\times 256 \\times 3} \\) as the input of the network. We use the ResNet-18 with removing the max pooling layer as the feature extractor, and the final full connection layer is used as the HR estimation head. At the same time, we designed a BVP signal regression head containing four blocks, each block involves a Conv., transpose, and two Conv. layers with batch normalization and Relu function. Hyper-parameter \\( \\sigma \\), \\( \\rho \\) and \\( \\tau \\) are set to 5, 0.1, and 0.2 according to the experimental result, and the trade-off parameter \\( k_{1} - k_{5} \\) are set to 1, 0.1, 0.001, 0.1 and 0.01 based on the scale of losses. Adam optimizer with a learning rate of 0.001 is used for training. The batch size and iterations are set to 1024 and 20000." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.688, + 0.472, + 0.793 + ], + "angle": 0, + "content": "Performance Metrics. Following existing methods [6, 28, 37, 50, 70], standard deviation (SD), mean absolute error (MAE), root mean square error (RMSE), and Pearson's correlation coefficient (r) are used to evaluate the HR estimation. MAE, RMSE, and r are used to evaluate the HRV measurement, including low frequency (LF), high frequency (HF), and LF/HF." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.803, + 0.4, + 0.818 + ], + "angle": 0, + "content": "4.3. Multi-Source Domain Generalization" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.822, + 0.238, + 0.836 + ], + "angle": 0, + "content": "4.3.1 HR Estimation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.472, + 0.901 + ], + "angle": 0, + "content": "For remote HR estimation, all five datasets (i.e., UBFC [1], PURE [52], BUAA [68], VIPL [36], and V4V [47]) are used to evaluate the model generalization. Specifically, four datasets are used for training, and then the other dataset is" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.477, + 0.892, + 0.612 + ], + "angle": 0, + "content": "used to test. Firstly, we perform the iPhys \\(^2\\) to evaluate three traditional method (i.e., GREEN [59], CHROM [9], and POS [63]) as shown in Tab. 1. Then, we reproduce five popular DL-based methods (i.e., DeepPhys [6], TSCAN [25], Rhythmnet [36], and BVPNet [7], and DualGAN [28]) for comparison on the same setting. Subsequently, we directly applied the four DG methods (i.e., AD [12], GroupDRO [38], Coral [53], VReX [18], and NCDG [57]) to Rhythmnet with using DeepDG \\(^3\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.613, + 0.892, + 0.734 + ], + "angle": 0, + "content": "The performance of traditional methods. As shown in Tab. 1, traditional algorithms show relatively acceptable results on some simple scenes, such as BUAA and UBFC, but very poor results on other datasets. It is difficult for these traditional algorithms to extract the physiological information on the V4V dataset, which shows that the traditional methods cannot deal with the noise caused by the complex environment and non-rigid motions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.733, + 0.892, + 0.854 + ], + "angle": 0, + "content": "The performance of DL-based methods. For DL-based methods, DeepPhys and TS-CAN even perform worse than traditional methods. Rhythmnet, BVPNet, and Dual-GAN perform much better than both other DL-based methods and traditional methods, which suggests that STMap is the more robust input of CNN than the aligned video. Compared with Rhythmnet, BVPNet and Dual-GAN have not significant improvement on the MSDG protocol. Because BVP-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.516, + 0.863, + 0.858, + 0.876 + ], + "angle": 0, + "content": "2https://github.com/danmcduff/iphys-toolbox" + }, + { + "type": "page_footnote", + "bbox": [ + 0.504, + 0.877, + 0.89, + 0.9 + ], + "angle": 0, + "content": "3https://github.com/jindongwang/transferlearning/tree/master/code/DeepDG" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.863, + 0.89, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.926, + 0.52, + 0.938 + ], + "angle": 0, + "content": "18594" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.087, + 0.088, + 0.88, + 0.36 + ], + "angle": 0, + "content": "
TargetMethodLF-(u.n)HF-(u.n)LF/HFHR-(bpm)
MAE↓RMSE↓r↑MAE↓RMSE↓r↑MAE↓RMSE↓r↑MAE↓RMSE↓r↑
UBFCGREEN [59]0.23550.28410.09240.23550.28410.09240.66950.95120.04678.01849.17760.3634
CHROM [9]0.22210.28170.06980.22210.28170.06980.67081.05420.10547.22918.92240.5123
POS [63]0.23640.28610.13590.23640.28610.13590.65150.95350.13457.35398.04020.4923
Baseline0.06210.08130.18730.06210.08130.18730.19850.26670.30435.15427.46720.8165
NEST0.05970.07820.20170.05970.07820.20170.21380.28240.31794.74716.88760.8546
PUREGREEN [59]0.25390.30020.03260.25390.30020.03260.65250.89320.041710.324714.26930.4952
CHROM [9]0.20960.27510.10590.20960.27510.07590.54040.82660.11739.791412.75680.3732
POS [63]0.19590.25710.16840.19590.25710.16840.53730.8460.14339.827313.44140.3432
Baseline0.06710.09230.60460.06710.09230.60460.28640.41840.55268.254211.17650.6832
NEST0.06350.08740.64220.06350.08740.64220.22550.35050.57347.688910.47830.7255
BUAAGREEN [59]0.34720.39510.08710.34720.39510.08710.64530.86320.09215.82317.98820.5624
CHROM [9]0.37860.32370.06820.37860.32370.06820.68130.88360.07156.09348.29380.5165
POS [63]0.31980.37620.09620.31980.37620.09620.62750.84240.11275.04077.11980.6374
Baseline0.14510.16810.28910.14510.16810.28910.55640.69040.29143.78526.32370.7492
NEST0.14360.16650.29550.14360.16650.29550.55140.68840.30043.37235.88060.7647
" + }, + { + "type": "table_caption", + "bbox": [ + 0.292, + 0.363, + 0.677, + 0.377 + ], + "angle": 0, + "content": "Table 2. HRV and HR estimation results on the MSDG protocol." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.397, + 0.47, + 0.457 + ], + "angle": 0, + "content": "Net and Dual-GAN consider the performance of the single domain instead of focusing on solving domain shift. Therefore, Rhythmnet is also selected as a baseline to further analyze the generalization on MSDG protocol." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.459, + 0.47, + 0.683 + ], + "angle": 0, + "content": "The performance of DG methods. To further improve the generalization of the model, we apply the existing popular DG methods to baseline (Rhythmnet) as shown in Tab. 1. There is a slight model degradation of AD [12], Coral [53], and GroupDRO [38] on some target datasets. The reason behind this is that these algorithms employ inaccurate domain labeling, specifically, a dataset often with different acquisition devices, lighting, and scene cannot be considered as the same domain. On the contrary, VREx [18] and NCDG [57] can slightly improve the performance of Rhythmnet without utilizing the domain label. Inaccurate domain division will make it difficult for the DG algorithm to effectively align the feature distribution across different domains, which is termed as agnostic domain generalization." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.685, + 0.47, + 0.79 + ], + "angle": 0, + "content": "Without the domain labels, NEST experimentally manifested significant improvement on MSDG protocol. For the target dataset BUAA, the results without NEST-CM \\(\\mathcal{L}_{CM}\\) show the best performance. Because NEST-CM may cause the model degradation with activating meaningless features at the instance level, especially for the BUAA with the slight domain shift." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.815, + 0.248, + 0.829 + ], + "angle": 0, + "content": "4.3.2 HRV Estimation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.903 + ], + "angle": 0, + "content": "We also conduct experiments for HRV (i.e., LF, HF, and LF/HF) and HR measurement on the PURE, UBFC, and BUAA. The reason for choosing only these three datasets is that VIPL and V4V do not have reliable ground-truth" + }, + { + "type": "table", + "bbox": [ + 0.511, + 0.394, + 0.88, + 0.52 + ], + "angle": 0, + "content": "
MethodPUREBUAA
MAE↓RMSE↓r↑MAE↓RMSE↓r↑
GREEN [59]10.3214.270.525.827.990.56
CHROM [9]9.7912.760.376.098.290.51
POS [63]9.8213.440.345.047.120.63
Baseline7.5410.130.732.973.280.76
NEST6.079.060.762.562.730.78
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.523, + 0.892, + 0.551 + ], + "angle": 0, + "content": "Table 3. HR estimation results on SSDG protocol by training on the UBFC-rPPG." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.57, + 0.892, + 0.66 + ], + "angle": 0, + "content": "BVP signals, as mentioned in subsection 4.1. We also perform three tradition method (GREEN [59], CHMROM [9], POS [63]) by iPhys for comparison with our method as shown in Tab. 2. With a more robust NEST presentation distribution, our method also can get better performance than baseline and other traditional methods in all metrics." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.69, + 0.825, + 0.706 + ], + "angle": 0, + "content": "4.4. Single-Source Domain Generalization" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.901 + ], + "angle": 0, + "content": "For a worst-case scenario, the model is trained on only one dataset to generalize to the unseen domains, namely single-source domain generalization (SSDG). Based on this, we select UBFC-rPPG as the source domain and then select PURE and BUAA as the target domain to further evaluate the generalization of NEST as shown in Tab. 3. As we can see, NEST achieves desired performance by training on only one dataset. The reason is that: (1) NEST-CM can avoid the appearance of abnormal NEST; (2) NEST-TA can find the shared and robust feature of the samples within the same label; (3) NEST-DM can distinguish the boundaries between different samples to improve the generalization." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.926, + 0.519, + 0.938 + ], + "angle": 0, + "content": "18595" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.088, + 0.089, + 0.457, + 0.292 + ], + "angle": 0, + "content": "
MethodSD↓MAE↓RMSE↓r↑
SAMC [58]18.015.921.00.11
POS [63]15.311.517.20.30
CHROM [9]15.111.416.90.28
I3D [3]15.912.015.90.07
DeepPhy [6]13.611.013.80.11
BVPNet [8]7.755.347.850.70
RhythmNet [36]8.115.308.140.76
CVD [37]7.925.027.970.79
Physformer [70]7.744.977.790.78
Dual-GAN [28]7.634.937.680.81
Baseline8.045.218.070.77
NEST7.494.767.510.84
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.297, + 0.468, + 0.325 + ], + "angle": 0, + "content": "Table 4. HR estimation results by our method and several state-of-the-art methods on the VIPL-HR database." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.328, + 0.468, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.41, + 0.469, + 0.438 + ], + "angle": 0, + "content": "Figure 5. Impacts of the hyperparameter (a) \\(\\rho\\), (b) \\(\\sigma\\), and (c) \\(\\tau\\) of our proposed method." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.449, + 0.347, + 0.466 + ], + "angle": 0, + "content": "4.5. Intra-Dataset Testing on VIPL" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.473, + 0.469, + 0.638 + ], + "angle": 0, + "content": "Our proposed method can also improve the effectiveness of the model on intra-dataset testing by reducing the harmful effects of abnormal disturbances. Following [28, 36, 37, 70], we test NEST by using a 5-fold cross-validation protocol on VIPL-HR. Our method compared with ten baseline methods, including the traditional methods (SAMC [58], POS [63] and CHROM [9]) and the DL-based methods (I3D [3], DeepPhy [6], RhythmNet [36], CVD [37], BVP-Net [7], Physformer [70], and Dual-GAN [28]). The results of these methods are directly from [28, 70] as shown in Tab. 4." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.64, + 0.469, + 0.759 + ], + "angle": 0, + "content": "As shown in Tab. 4, the proposed methods outperform all the methods under all metrics. Compared with the baseline, it also gets a noticeable promotion, i.e., 0.4 in MAE and 0.53 in RMSE. It is worth noting that the VIPL dataset has multiple complex scenes and recording devices and cannot be considered as a single domain, termed the agnostic domain problem. This proves that NEST can effectively alleviate the domain shift without domain labels." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.771, + 0.245, + 0.788 + ], + "angle": 0, + "content": "4.6. Further Analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.469, + 0.901 + ], + "angle": 0, + "content": "We conduct further experiments on MSDG protocol (testing on UBFC with training on the other four datasets). Impact of \\(\\rho\\) in NEST-CM. The hyperparameter \\(\\rho\\) determines how many bases need to be strengthened by NEST-CM \\(\\mathcal{L}_{CM}\\). As illustrated in Fig. 5 (a), NEST could achieve smaller RMSE when \\(\\rho = 0.1\\), indicating that the bases with small eigenvalue require more attention and fine-tuning." + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.09, + 0.882, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.221, + 0.892, + 0.259 + ], + "angle": 0, + "content": "Figure 6. The pearson's correlation coefficient of the NEST representation of the samples with different labels for the anchor HR 60." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.268, + 0.892, + 0.449 + ], + "angle": 0, + "content": "Impact of \\(\\sigma\\) in NEST-TA. The hyperparameter \\(\\sigma\\) tradeoffs the contribution of each selected sample to form the synthesized prototype. As illustrated in Fig. 5, NEST could achieve the best performance with \\(\\sigma = 5\\), which shows that the appropriate fusion of the nearby sample is conducive to improving model generalization. The specific reason behind this is that: (1) the synthesized prototype based on multi-domain samples can be used to enhance the domain invariant feature by NEST-TA; (2) NEST-TA can effectively aggregate the label relationship into NEST presentation distribution, which is performed by improving the similarity of NEST representation between the adjacent label samples." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.452, + 0.893, + 0.572 + ], + "angle": 0, + "content": "To further clarify the second reason, we calculate the Pearson's correlation coefficient of the NEST representation of the samples with different labels for the anchor HR 60, as shown in Fig. 6. It is obvious that the distribution changes with NEST-TA loss are more smooth across the continuous HR labels, which allows OOD samples with slight domain shifts to predict nearby heart rates around the ground truth HR value instead of other abnormal HR values." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.574, + 0.892, + 0.68 + ], + "angle": 0, + "content": "Impact of \\(\\tau\\) in NEST-DM. The hyperparameter \\(\\tau\\) determines the pushing strength of the different samples. With the smallest \\(\\tau\\), NEST-DM \\(\\mathcal{L}_{DM}\\) has the overemphasized constraints to distinguish the NEST representation between different samples, which causes a harmful impact on the discriminant ability of the model. Overall, NEST-DM \\(\\mathcal{L}_{DM}\\) is an effective way to improve the generalization of model." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.696, + 0.619, + 0.712 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This paper presents the NEural SStructure (NEST) modeling, a novel framework that addresses the domain generalization problem in rPPG measurement for the first time. Our framework contains three regularization terms defined in the view of neural structure, which improves the generalization capacity in the different aspects. All three terms do not rely on domain labels, which are especially suitable in rPPG, as defining domains is difficult in this task. To compare the generalization of the existing algorithm and our method, we establish the first DG protocol for this task. We experimentally manifested the substantial improvement of our method on both MSDG and SSDG protocols." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.926, + 0.519, + 0.938 + ], + "angle": 0, + "content": "18596" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.469, + 0.17 + ], + "angle": 0, + "content": "[1] Serge Bobbia, Richard Macwan, Yannick Benezeth, Alamin Mansouri, and Julien Dubois. Unsupervised skin tissue segmentation for remote photoplethysmography. Pattern Recognition Letters, 124:82-90, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.469, + 0.227 + ], + "angle": 0, + "content": "[2] Adrian Bulat and Georgios Tzimiropoulos. How far are we from solving the 2d & 3d face alignment problem? (and a dataset of 230,000 3d facial landmarks). In Proc. ICCV, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.23, + 0.468, + 0.271 + ], + "angle": 0, + "content": "[3] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proc. IEEE CVPR, pages 6299-6308, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.273, + 0.468, + 0.314 + ], + "angle": 0, + "content": "[4] Beidi Chen, Tri Dao, Eric Winsor, Zhao Song, Atri Rudra, and Christopher Ré. Scatterbrain: Unifying sparse and low-rank attention. NeurIPS, 34:17413-17426, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.316, + 0.468, + 0.37 + ], + "angle": 0, + "content": "[5] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, pages 1597-1607. PMLR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.373, + 0.468, + 0.415 + ], + "angle": 0, + "content": "[6] Weixuan Chen and Daniel McDuff. Deep learning: Video-based physiological measurement using convolutional attention networks. In Proc. ECCV, pages 349-365, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.417, + 0.468, + 0.47 + ], + "angle": 0, + "content": "[7] Abhijit Das, Hao Lu, Hu Han, Antitza Dantcheva, Shiguang Shan, and Xilin Chen. Bvpnet: Video-to-bvp signal prediction for remote heart rate estimation. In \\(FG\\), pages 01-08. IEEE, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.473, + 0.468, + 0.527 + ], + "angle": 0, + "content": "[8] Abhijit Das, Hao Lu, Hu Han, Antitza Dantcheva, Shiguang Shan, and Xilin Chen. Bvpnet: Video-to-bvp signal prediction for remote heart rate estimation. In \\(FG\\), pages 01-08. IEEE, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.53, + 0.468, + 0.571 + ], + "angle": 0, + "content": "[9] Gerard De Haan and Vincent Jeanne. Robust pulse rate from chrominance-based rppg. IEEE Trans. Biomed. Eng., 60(10):2878-2886, 2013." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.573, + 0.469, + 0.614 + ], + "angle": 0, + "content": "[10] Gerard De Haan and Arno Van Leest. Improved motion robustness of remote-ppg by using the blood volume pulse signature. Physiol. Meas., 35(9):1913, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.616, + 0.469, + 0.657 + ], + "angle": 0, + "content": "[11] Zhijie Deng, Yucen Luo, and Jun Zhu. Cluster alignment with a teacher for unsupervised domain adaptation. In Proc. CVPR, pages 9944–9953, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.66, + 0.469, + 0.7 + ], + "angle": 0, + "content": "[12] Yaroslav Ganin and Victor Lempitsky. Unsupervised domain adaptation by backpropagation. In ICML, pages 1180-1189. PMLR, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.702, + 0.469, + 0.744 + ], + "angle": 0, + "content": "[13] Ming-Shiang Chen Gee-Sern Hsu, ArulMurugan Ambikapathi. Deep learning with time-frequency representation for pulse estimation. In Proc. IJCB, pages 642-650, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.746, + 0.469, + 0.8 + ], + "angle": 0, + "content": "[14] Min Hu, Fei Qian, Dong Guo, Xiaohua Wang, Lei He, and Fuji Ren. Eta-rppgnet: effective time-domain attention network for remote heart rate measurement. IEEE Trans. on Instrum. and Meas., 70:1-12, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.469, + 0.844 + ], + "angle": 0, + "content": "[15] Yunpei Jia, Jie Zhang, Shiguang Shan, and Xilin Chen. Single-side domain generalization for face anti-spoofing. In Proc. CVPR, pages 8484-8493, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[16] Xiang Jiang, Qicheng Lao, Stan Matwin, and Mohammad Havaei. Implicit class-conditioned domain alignment for unsupervised domain adaptation. In ICML, pages 4816-4827. PMLR, 2020." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.469, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "[17] Viktor Kessler, Patrick Thiam, Mohammadreza Amirian, and Friedhelm Schwenker. Pain recognition with camera photoplethysmography. In IEEE IPTA, pages 1-5, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.203 + ], + "angle": 0, + "content": "[18] David Krueger, Ethan Caballero, Joern-Henrik Jacobsen, Amy Zhang, Jonathan Binas, Dinghuai Zhang, Remi Le Priol, and Aaron Courville. Out-of-distribution generalization via risk extrapolation (rex). In ICML, pages 5815-5826. PMLR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.205, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[19] Eugene Lee, Evan Chen, and Chen-Yi Lee. Meta-rppg: Remote heart rate estimation using a transductive meta-learner. In ECCV, pages 392–409. Springer, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.247, + 0.892, + 0.327 + ], + "angle": 0, + "content": "[20] Alexander Lehner, Stefano Gasperini, Alvaro Marcos-Ramiro, Michael Schmidt, Mohammad-Ali Nikouei Mahani, Nassir Navab, Benjamin Busam, and Federico Tombari. 3d-vfield: Adversarial augmentation of point clouds for domain generalization in 3d object detection. In Proc. CVPR, pages 17295-17304, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.33, + 0.892, + 0.385 + ], + "angle": 0, + "content": "[21] Magdalena Lewandowska, Jacek Ruminski, Tomasz Kocejko, and Jedrzej Nowak. Measuring pulse rate with a webcam—a non-contact method for evaluating cardiac activity. In FedCSIS, pages 405-410. IEEE, 2011." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.386, + 0.892, + 0.441 + ], + "angle": 0, + "content": "[22] Daiqing Li, Junlin Yang, Karsten Kreis, Antonio Torralba, and Sanja Fidler. Semantic segmentation with generative models: Semi-supervised learning and strong out-of-domain generalization. In Proc. CVPR, pages 8300-8311, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.442, + 0.892, + 0.482 + ], + "angle": 0, + "content": "[23] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy Hospedales. Learning to generalize: Meta-learning for domain generalization. In Proc. AAAI, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.484, + 0.892, + 0.537 + ], + "angle": 0, + "content": "[24] Chuang Lin, Zehuan Yuan, Sicheng Zhao, Peize Sun, Changhu Wang, and Jianfei Cai. Domain-invariant disentangled network for generalizable object detection. In Proce. CVPR, pages 8771-8780, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.539, + 0.892, + 0.592 + ], + "angle": 0, + "content": "[25] Xin Liu, Josh Fromm, Shwetak Patel, and Daniel McDuff. Multi-task temporal shift attention networks for on-device contactless vitals measurement. NeurIPS, 33:19400-19411, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.595, + 0.892, + 0.649 + ], + "angle": 0, + "content": "[26] Xin Liu, Ziheng Jiang, Josh Fromm, Xuhai Xu, Shwetak Patel, and Daniel McDuff. Metaphys: few-shot adaptation for non-contact physiological measurement. In Proc. HIL, pages 154-163, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.651, + 0.892, + 0.691 + ], + "angle": 0, + "content": "[27] Zhenhua Liu, Jizheng Xu, Xiulian Peng, and Ruiqin Xiong. Frequency-domain dynamic pruning for convolutional neural networks. NeurIPS, 31, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.692, + 0.892, + 0.734 + ], + "angle": 0, + "content": "[28] Hao Lu, Hu Han, and S Kevin Zhou. Dual-gan: Joint bvp and noise modeling for remote physiological measurement. In Proc. IEEE CVPR, pages 12404-12413, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.735, + 0.892, + 0.789 + ], + "angle": 0, + "content": "[29] Yawei Luo, Ping Liu, Liang Zheng, Tao Guan, Junqing Yu, and Yi Yang. Category-level adversarial adaptation for semantic segmentation using purified features. IEEE Trans. Pattern Anal. Mach. Intell., 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.79, + 0.892, + 0.845 + ], + "angle": 0, + "content": "[30] Fangrui Lv, Jian Liang, Shuang Li, Bin Zang, Chi Harold Liu, Ziteng Wang, and Di Liu. Causality inspired representation learning for domain generalization. In Proc. CVPR, pages 8046-8056, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[31] Lei Ma, Felix Juefei-Xu, Fuyuan Zhang, Jiyuan Sun, Minhui Xue, Bo Li, Chunyang Chen, Ting Su, Li Li, Yang Liu, et al. Deep gauge: Multi-granularity testing criteria for deep learning systems. In Proc. ASE, pages 120-131, 2018." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.927, + 0.519, + 0.938 + ], + "angle": 0, + "content": "18597" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.12 + ], + "angle": 0, + "content": "[32] Daniel McDuff. Camera measurement of physiological vital signs. CSUR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.468, + 0.162 + ], + "angle": 0, + "content": "[33] Daniel McDuff. Applications of camera-based physiological measurement beyond healthcare. In CVSM, pages 165-177. Elsevier, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.165, + 0.468, + 0.219 + ], + "angle": 0, + "content": "[34] Daniel McDuff, Sarah Gontarek, and Rosalind W Picard. Improvements in remote cardiopulmonary measurement using a five band digital camera. IEEE Trans. Biomed. Eng., 61(10):2593-2601, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.221, + 0.468, + 0.262 + ], + "angle": 0, + "content": "[35] Xuesong Niu, Hu Han, Shiguang Shan, and Xilin Chen. Synrhythm: Learning a deep heart rate estimator from general to specific. In Proc. IEEE ICPR, pages 3580-3585, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.264, + 0.468, + 0.318 + ], + "angle": 0, + "content": "[36] Xuesong Niu, Shiguang Shan, Hu Han, and Xilin Chen. Rhythmnet: End-to-end heart rate estimation from face via spatial-temporal representation. IEEE Trans. on Image Process., 29:2409-2423, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.321, + 0.468, + 0.374 + ], + "angle": 0, + "content": "[37] Xuesong Niu, Zitong Yu, Hu Han, Xiaobai Li, Shiguang Shan, and Guoying Zhao. Video-based remote physiological measurement via cross-verified feature disentangling. In Proc. ECCV, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.377, + 0.468, + 0.417 + ], + "angle": 0, + "content": "[38] Giambattista Parascandolo, Alexander Neitz, Antonio Orvieto, Luigi Gresele, and Bernhard Scholkopf. Learning explanations that are hard to vary. *ICLR2022*, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.42, + 0.468, + 0.461 + ], + "angle": 0, + "content": "[39] Kexin Pei, Yinzhi Cao, Junfeng Yang, and Suman Jana. Deepxplore: Automated whitebox testing of deep learning systems. In Proc. SOSP, pages 1-18, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.463, + 0.468, + 0.503 + ], + "angle": 0, + "content": "[40] Duo Peng, Yinjie Lei, Munawar Hayat, Yulan Guo, and Wen Li. Semantic-aware domain generalized segmentation. In Proc. CVPR, pages 2594-2605, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.505, + 0.468, + 0.546 + ], + "angle": 0, + "content": "[41] Vihari Piratla, Praneeth Netrapalli, and Sunita Sarawagi. Efficient domain generalization via common-specific low-rank decomposition. In ICML, pages 7728-7738. PMLR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.548, + 0.468, + 0.602 + ], + "angle": 0, + "content": "[42] Ming-Zher Poh, Daniel J McDuff, and Rosalind W Picard. Non-contact, automated cardiac pulse measurements using video imaging and blind source separation. Opt. Express, 18(10):10762-10774, 2010." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.605, + 0.468, + 0.644 + ], + "angle": 0, + "content": "[43] Fengchun Qiao, Long Zhao, and Xi Peng. Learning to learn single domain generalization. In Proc. CVPR, pages 12556-12565, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.647, + 0.468, + 0.701 + ], + "angle": 0, + "content": "[44] Ying Qiu, Yang Liu, Juan Arteaga-Falconi, Haiwei Dong, and Abdulmotaleb El Saddik. Evm-cnn: Real-time contactless heart rate estimation from facial video. IEEE Trans. Multimedia, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.704, + 0.468, + 0.758 + ], + "angle": 0, + "content": "[45] Attila Reiss, Ina Indlekofer, Philip Schmidt, and Kristof Van Laerhoven. Deep ppg: large-scale heart rate estimation with convolutional neural networks. Sensors, 19(14):3079, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.468, + 0.814 + ], + "angle": 0, + "content": "[46] Jiahuan Ren, Zhao Zhang, Richang Hong, Mingliang Xu, Hajun Zhang, Mingbo Zhao, and Meng Wang. Robust low-rank convolution network for image denoising. In Proc. ACM MM, pages 6211-6219, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.468, + 0.871 + ], + "angle": 0, + "content": "[47] Ambareesh Revanur, Zhihua Li, Umur A Ciftci, Lijun Yin, and László A Jeni. The first vision for vitals (v4v) challenge for non-contact video-based physiological estimation. In Proc. CVPR workshop, pages 2760-2767, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.874, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[48] Shiv Shankar, Vihari Piratla, Soumen Chakrabarti, Siddhartha Chaudhuri, Preethi Jyothi, and Sunita Sarawagi." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.89, + 0.12 + ], + "angle": 0, + "content": "Generalizing across domains via cross-gradient training. ICLR, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.177 + ], + "angle": 0, + "content": "[49] Rencheng Song, Huan Chen, Juan Cheng, Chang Li, Yu Liu, and Xun Chen. Pulsegan: Learning to generate realistic pulse waveforms in remote photoplethysmography. IEEE J-BHI, pages 1-1, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.233 + ], + "angle": 0, + "content": "[50] Rencheng Song, Senle Zhang, Chang Li, Yunfei Zhang, Juan Cheng, and Xun Chen. Heart rate estimation from facial videos using a spatiotemporal representation with convolutional neural networks. IEEE Trans. Instrum. Meas., 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.235, + 0.892, + 0.276 + ], + "angle": 0, + "content": "[51] Radim Špetlík, Vojtech Franc, and Jírí Matas. Visual heart rate estimation with convolutional neural network. In Proc. BMVC, pages 3-6, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.278, + 0.892, + 0.33 + ], + "angle": 0, + "content": "[52] Ronny Stricker, Steffen Müller, and Horst-Michael Gross. Non-contact video-based pulse rate measurement on a mobile service robot. In Proc. IEEE ISRHIC, pages 1056-1062, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.334, + 0.892, + 0.375 + ], + "angle": 0, + "content": "[53] Baochen Sun and Kate Saenko. Deep coral: Correlation alignment for deep domain adaptation. In ECCV, pages 443-450. Springer, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.378, + 0.892, + 0.431 + ], + "angle": 0, + "content": "[54] Bo Sun, Qinglan Wei, Liandong Li, Qihua Xu, Jun He, and Lejun Yu. Lstm for dynamic emotion and group emotion recognition in the wild. In Proc. ACM ICMI, pages 451-457, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.434, + 0.892, + 0.475 + ], + "angle": 0, + "content": "[55] Zhaodong Sun and Xiaobai Li. Contrast-phys: Unsupervised video-based remote physiological measurement via spatiotemporal contrast. In Proc. ECCV, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.477, + 0.892, + 0.531 + ], + "angle": 0, + "content": "[56] Ling Tang, Wen Shen, Zhanpeng Zhou, Yuefeng Chen, and Quanshi Zhang. Defects of convolutional decoder networks in frequency representation. arXiv preprint arXiv:2210.09020, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.533, + 0.892, + 0.574 + ], + "angle": 0, + "content": "[57] Chris Xing Tian, Haoliang Li, Xiaofei Xie, Yang Liu, and Shiqi Wang. Neuron coverage-guided domain generalization. IEEE Trans. on Pattern Anal. Mach. Intell., 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.577, + 0.892, + 0.643 + ], + "angle": 0, + "content": "[58] Sergey Tulyakov, Xavier Alameda-Pineda, Elisa Ricci, Lijun Yin, Jeffrey F Cohn, and Nicu Sebe. Self-adaptive matrix completion for heart rate estimation from face videos under realistic conditions. In Proc. IEEE CVPR, pages 2396-2404, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.647, + 0.892, + 0.688 + ], + "angle": 0, + "content": "[59] Wim Verkruysse, Lars O Svaasand, and J Stuart Nelson. Remote plethysmographic imaging using ambient light. Opt. Express, 16(26):21434-21445, 2008." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.69, + 0.892, + 0.745 + ], + "angle": 0, + "content": "[60] Tan Wad, Qianru Sun, Sugiri Pranata, Karlekar Jayashree, and Hanwang Zhang. Equivalence and invariance inductive bias for learning from insufficient data. In ECCV, pages 241-258. Springer, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.747, + 0.892, + 0.8 + ], + "angle": 0, + "content": "[61] Hao Wang, Euijoon Ahn, and Jinman Kim. Self-supervised representation learning framework for remote physiological measurement using spatiotemporal augmentation loss. In ACM MM, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.803, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[62] Jindong Wang, Cuiling Lan, Chang Liu, Yidong Ouyang, Tao Qin, Wang Lu, Yiqiang Chen, Wenjun Zeng, and Philip Yu. Generalizing to unseen domains: A survey on domain generalization. IEEE Trans. on Knowl. Data Eng., 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.86, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[63] Wenjin Wang, Albertus C den Brinker, Sander Stuijk, and Gerard de Haan. Algorithmic principles of remote ppg. IEEE Trans. Biomed. Eng., 64(7):1479-1491, 2017." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.927, + 0.52, + 0.938 + ], + "angle": 0, + "content": "18598" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.134 + ], + "angle": 0, + "content": "[64] Wenjin Wang, Sander Stuijk, and Gerard De Haan. Exploiting spatial redundancy of image sensor for motion robust rppg. IEEE Trans. Biomed. Eng., 62(2):415-425, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.136, + 0.468, + 0.177 + ], + "angle": 0, + "content": "[65] Yunhe Wang, Chang Xu, Shan You, Dacheng Tao, and Chao Xu. Cnnpack: Packing convolutional neural networks in the frequency domain. In NeurIPS, volume 29, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.179, + 0.468, + 0.22 + ], + "angle": 0, + "content": "[66] Zijian Wang, Yadan Luo, Ruihong Qiu, Zi Huang, and Mahsa Baktashmotlagh. Learning to diversify for single domain generalization. In Proc. CVPR, pages 834-843, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.222, + 0.468, + 0.262 + ], + "angle": 0, + "content": "[67] Zhi-Kuan Wang, Ying Kao, and Chiou-Ting Hsu. Vision-based heart rate estimation via a two-stream cnn. In Proc. IEEE ICIP, pages 3327-3331, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.264, + 0.468, + 0.317 + ], + "angle": 0, + "content": "[68] Lin Xi, Weihai Chen, Changchen Zhao, Xingming Wu, and Jianhua Wang. Image enhancement for remote photoplethysmography in a low-light environment. In FG, pages 1-7. IEEE, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.321, + 0.468, + 0.388 + ], + "angle": 0, + "content": "[69] Zitong Yu, Wei Peng, Xiaobai Li, Xiaopeng Hong, and Guoying Zhao. Remote heart rate measurement from highly compressed facial videos: an end-to-end deep learning solution with video enhancement. In Proc. IEEE ICCV, pages 151-160, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.391, + 0.468, + 0.445 + ], + "angle": 0, + "content": "[70] Zitong Yu, Yuming Shen, Jingang Shi, Hengshuang Zhao, Philip Torr, and Guoying Zhao. Physformer: Facial videobased physiological measurement with temporal difference transformer. IEEE CVPR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.448, + 0.468, + 0.516 + ], + "angle": 0, + "content": "[71] Xiangyu Yue, Yang Zhang, Sicheng Zhao, Alberto Sangiovanni-Vincentelli, Kurt Keutzer, and Boqing Gong. Domain randomization and pyramid consistency: Simulation-to-real generalization without accessing target domain data. In Proc.CVPR, pages 2100-2110, 2019." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.516 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.927, + 0.519, + 0.938 + ], + "angle": 0, + "content": "18599" + } + ] +] \ No newline at end of file diff --git a/2023/Neuron Structure Modeling for Generalizable Remote Physiological Measurement/0c500895-2bd8-4cea-80ae-7bc4fd9b8ba5_origin.pdf b/2023/Neuron Structure Modeling for Generalizable Remote Physiological Measurement/0c500895-2bd8-4cea-80ae-7bc4fd9b8ba5_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..3a3dfd10044978aed43baa3dbaf4b11249df6fa9 --- /dev/null +++ b/2023/Neuron Structure Modeling for Generalizable Remote Physiological Measurement/0c500895-2bd8-4cea-80ae-7bc4fd9b8ba5_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e70f5d1606c7f97d69c9fff266ba38ea0270e5bacb8ae96888f2b529f2296431 +size 2869016 diff --git a/2023/Neuron Structure Modeling for Generalizable Remote Physiological Measurement/full.md b/2023/Neuron Structure Modeling for Generalizable Remote Physiological Measurement/full.md new file mode 100644 index 0000000000000000000000000000000000000000..68709c1cfc11b3d8142b5d26e1bb81c106d5aa73 --- /dev/null +++ b/2023/Neuron Structure Modeling for Generalizable Remote Physiological Measurement/full.md @@ -0,0 +1,317 @@ +# Neuron Structure Modeling for Generalizable Remote Physiological Measurement + +Hao LU $^{1,2}$ , Zitong YU $^{3}$ , Xuesong NIU $^{4}$ , Yingcong CHEN $^{1,2,*}$ + +1The Hong Kong University of Science & Technology (Guangzhou), + +$^{2}$ The Hong Kong University of Science & Technology, $^{3}$ Great Bay University, $^{4}$ Kuaishou Technology hlu585@connect.hkust-gz.edu.cn, zitong.yu@ieee.org, nxsedson@gmail.com, yingcongchen@ust.hk + +# Abstract + +Remote photoplethysmography (rPPG) technology has drawn increasing attention in recent years. It can extract Blood Volume Pulse (BVP) from facial videos, making many applications like health monitoring and emotional analysis more accessible. However, as the BVP signal is easily affected by environmental changes, existing methods struggle to generalize well for unseen domains. In this paper, we systematically address the domain shift problem in the rPPG measurement task. We show that most domain generalization methods do not work well in this problem, as domain labels are ambiguous in complicated environmental changes. In light of this, we propose a domain-label-free approach called NEuron SStructure modeling (NEST). NEST improves the generalization capacity by maximizing the coverage of feature space during training, which reduces the chance for under-optimized feature activation during inference. Besides, NEST can also enrich and enhance domain invariant features across multi-domain. We create and benchmark a large-scale domain generalization protocol for the rPPG measurement task. Extensive experiments show that our approach outperforms the state-of-the-art methods on both cross-dataset and intra-dataset settings. The codes are available at https://github.com/LuPaoPao/NEST. + +# 1. Introduction + +Physiological signals such as heart rate (HR), and heart rate variability (HRV), respiration frequency (RF) are important body indicators that serve not only as vital signs but also track the level of sympathetic activation [17, 33, 54]. Traditional physiological measurements, such as electrocardiograms, heart rate bands, and finger clip devices, have high accuracy. However, they are costly, intrusive, and uncomfortable to wear for a long time. Remote photoplethysmography (rPPG) can extract blood volume pulse (BVP) + +![](images/e9b3e153f156f2e4f7039efef49987157709640e9b430dac9ed9858c5c5d228c.jpg) +(a) Typical samples from different datasets + +![](images/74734f2b0862877af9122dbe060ea2681ef4e4059563e49850d9a0741960817f.jpg) + +![](images/d859c2131900ef0d19b9834e4e7bdc4bb0635d3b2f20f29419d29f49dc03ad72.jpg) + +![](images/1a20cf70a8f365ec011d1ce71fbb4d3e44bf0df7002125b2074baab67bfaf0ba.jpg) + +![](images/7d06aec4f82f4d151d3d618b1d4b1b395fb4971a43def901c78f1a38d9be854a.jpg) + +![](images/62d042aca48080d8e790247659b54c3194e234ea51e87f4cbc651b253fe93d81.jpg) +(b) The performance of different methods on DG protocol +Figure 1. (a) Typical samples from different publicly rPPG datasets: VIPL-HR [36], V4V [47], UBFC-rPPG [1], BUAA [68], PURE [52]. (b) The performance of different methods on DG protocol (test on the UBFC-rPPG dataset with training on the VIPL, V4V, PURE, and BUAA). + +from face video, which analyzes the periodic changes of the light absorption of the skin caused by heartbeats. Then various physiological indicators (such as HR and HRV) can be calculated based on BVP signals [37, 69, 70]. With nonintrusion and convenience, the rPPG-based physiological measurement method can only use an ordinary camera to monitor physiological indicators and gradually become a research hotspot in the computer vision field [7, 28, 32, 36, 44, 49]. + +Traditional rPPG measurement methods include signal blind decomposition [21,34,42] and color space transformation [9,59,63]. These approaches rely on heartbeat-related + +statistical information, only applicable in constrained environments. In recent years, deep learning (DL) based approaches [6, 19, 25, 28, 36, 44, 50, 69, 70] have shown their great potentials in rPPG measurement. By learning dedicated rPPG feature representation, these methods achieve promising performance in much more complicated environments [1, 36, 47, 52, 68]. + +However, deep learning methods suffer from significant performance degradation when applied in real-world scenarios. This is because most training data are captured in lab environments with limited environmental variations. With domain shifts (e.g., different illumination, camera parameters, motions, etc.), these models may struggle to generalize for the unseen testing domain. To validate this, we conduct a cross-dataset evaluation shown in Fig. 1(b). As shown, all DL-based methods do not work well in this evaluation. Furthermore, it is worth noting that DeepPhys [6] and TS-CAN [25] even perform inferior to traditional approach POS [63]. + +To improve the performance on the unseen domains, one common practice is to incorporate Domain Generalization (DG) approaches, e.g., encouraging intermediate features to be domain-invariant [12, 18, 30, 38, 57, 62]. However, as shown in Fig. 1 (b) (AD [12] and GroupDRO [38]), the improvements are still quite limited. One reason is that existing DG methods assume that domain labels can be clearly defined (e.g., by data source). Unfortunately, different background scenes, acquisition devices, or even identities could cause very different data distributions in rPPG measurement tasks, and such distribution discrepancy may exist within or cross datasets. In this case, explicitly defining domains is very difficult, and simply treating one dataset as one domain may lead to inferior performance. This is termed as agnostic domain generalization. Besides this, as physiological cues are usually much more subtle than the various noise, the model may overfit in the source domain with the limited training data. + +In this paper, we propose the NEural SStructure modeling (NEST), a principled solution to the abovementioned problems. The main idea of NEST is to narrow the under-optimized and redundant feature space, align domain invariant features, and enrich discriminative features. Our intuition is as follows. Neural structure refers to the channel activation degree in each convolution layer, which reveals the discriminative feature combination for the specific sample. As the limited variation in a certain restricted domain, there are some spaces that the model seldomly is optimized. Out-of-distribution (OOD) samples may cause abnormal activation in these spaces, which may lead to performance degeneration. Therefore, we regularize the neural structure to encourage the model to be more well-conditioned to avoid abnormal activation caused by OOD samples. Specifically, we propose the NEural SStructure Coverage Maxi + +mization (NEST-CM) that encourages all neural spaces to be optimized during training, reducing the chance of abnormal activation during testing. Secondly, we propose the NEural SStructure Targeted Alignment (NEST-TA) that encourage network suppresses domain variant feature by comparing the samples with the similar physiological information. Thirdly, we propose the NEural SStructure Diversity Maximization (NEST-DM) to enrich discriminative features against unseen noise. It should be noted that our approach does not rely on domain labels, which is more applicable in the rPPG measurement task. To summarize, our contributions are listed as follows: + +1. We are the first to study the domain shift problem in rPPG measurement, which introduces a new challenge, agnostic domain generalization. +2. We propose the NEural SStructure modeling to alleviate domain shift, which is performed by narrowing the under-optimized feature space, and enhancing and enriching domain invariant features. +3. We establish a large-scale domain generalization (DG) benchmark for rPPG measurement, which is the first DG protocol in this task. Extensive experiments in this dataset show the superiority of our approach. + +# 2. Related work + +# 2.1. Remote Physiological Measurement + +Remote physiological measurement aims to estimate HR and HRV values by analyzing the chrominance changes in the skin. Traditional methods mainly perform color space transformations or signal decomposition to obtain the BVP signal with high SNR [9, 10, 42, 64]. These methods are designed manually for certain lighting scenes with slight motion and can not generalize to the complex environment in the wild. With powerful modeling capacity, the deep learning (DL) model has been successfully used in remote physiological estimation [6, 13, 35, 36, 51, 67, 69]. Among these methods, various model structures are tailored to extract BVP signals from face videos by aggregating spatial and temporal information [6, 37, 51, 69, 70]. Besides, many hand-designed feature maps are used to reduce the computational burden of the network by removing the irrelevant background [13, 28, 36, 45, 50]. Some work utilizes the distribution of the target domain to improve the model performance [19, 26], which is not practical without touching the unseen domain in advance. Differently, we study the model generalization on domain generalization protocol for rPPG measurement. + +# 2.2. Domain generalization + +Domain generalization (DG), i.e., out-of-distribution generalization, aims to generalize to the unseen domain by training a model on multiple or single source domains [62]. The main solution can be categorized into data manipulation, representation learning, and meta-learning. Data + +![](images/b47254a0e21d43c8e68d78f1162f3a1ef02ef922759cfe210bd36533a3c81e86.jpg) +Figure 2. An overview of the proposed method. (a) Spatial-temporal map (STMap) is extracted from face video [36]. (b) NEST representation is calculated from the samples of the different domains. (c) The under-optimized space in NEST presentation distribution is narrowed by NEST-CM. (d) NEST-TA align the feature of multi domain at instance-level. (e) The distance of different samples are pulled away by NEST-DM. + +![](images/078712f385fb6ca430b0a2914122470b6d5cb25b4d750dcb141b7ab88e0dfcf8.jpg) + +The generalization of the model, which can remove irrelevant style information by data generation and data augmentation [48, 71]. The main goal of representation learning is to learn domain-invariant features by alleviating covariate shift across domain [12, 18, 38, 57]. By simulating domain shift on the virtual dataset, meta-learning seeks transferable and shareable knowledge from multiple domain [23, 30]. Unlike common tasks such as classification, segmentation, and detection [15, 20, 22, 24, 40], rPPG measurement tasks do not have explicit boundaries between different domains, i.e., agnostic domain generalization. Recently, some single source domain generalization (SSDG) methods focus on the worst-case formulation without using the domain label [43, 57, 66]. Different from all the methods above, we tackle the DG problem of regression task in view of neuron structure modeling. + +# 3. Method + +# 3.1. Overall Framework + +In this work, we propose the NEuron SStructure modeling (NSET), without using domain label, to alleviate the performance degradation caused by domain shift as illustrated in Fig. 2. Specifically, we first compressed the face video into the spatial-temporal map (STMap) as the input of + +CNN refer to [36] as shown in Fig. 2 (a). Sequentially, we define an Activated NEuron SStructure (NEST) representation that describes the activated degree of different channels of the network as shown in Fig. 2 (b). To alleviate the domain shift, we model a unified NEST distribution by three regularizations: (1) NEuron SStructure Coverage Maximization (NEST-CM) is used to narrow under-optimized space that is redundant and unstable as shown in Fig. 2 (c); (2) NEuron SStructure Targeted Alignment (NEST-TA) is introduced to align multi-domain feature distribution at instance-level as shown in Fig. 2 (d). (3) Neuron Structure Diversity Maximization (NEST-DM) loss to force the network to explore more features with discriminant ability as shown in Fig. 2 (e). + +# 3.2. NEuron SStructure Modeling + +NEuron SStructure Representation. For the rPPG measurement task, the convolution kernels are adaptively activated to extract physiological features against various noise distribution [27, 56, 65]. However, the facial physiological signals are feeble and can be easily affected by various unseen noises, which make the filters activated abnormally. Therefore, the NEuron SStructure (NEST) representation is defined to explore how domain shift affects the model robustness in view of activated feature distribution. Specifically, we firstly average the feature map + +![](images/1ecd02f2e15691a1050d15c1eb02ba96783252b535a36e696afca9c7ebc73031.jpg) +(a) The expectation of different base projection + +![](images/9a25213a89acee181e1b480a154ddb6982550d3d417498e363153b963c0baa2b.jpg) +(b) The projection distribution of base with highest eigenvalue +Figure 3. (a) The expectation of different base projection. (b) The distribution of base with highest eigenvalue. (c) The distribution of base with lowest eigenvalue. + +![](images/87ec75dcb2c77d0928dffdda3336a82164e716a434f56f0e5eb8c799a9ca4953.jpg) +(c) The projection distribution of base with lowest eigenvalue + +$m_{j}\in R^{W\times H\times C_{j}}$ over the spatial dimensions in the end of $j$ -th layer to get the preliminary statistical presentation $\bar{o}_j\in R^{C_j}$ of the neuron structure, where $C_j$ is the channel number of the $j$ -th layer. Following that, the preliminary presentation $\bar{o}_j$ is fed into the max-min normalized to get the NEST representation $o_{j}\in R_{j}^{C}$ of the $j$ -th layer. The full NEST representation of $i$ -th sample $x_{i}$ is defined as $O_{i} = \{o_{i,1},o_{i,2},\dots,o_{i,U}\} \in R^{(C_{1} + C_{2} + \ldots +C_{U})}$ , where $U$ is the number of selected layer. + +NEuron SStructure Coverage Maximization. Existing works show that the OOD samples may activate abnormal neuron structures (feature combination) that seldom be trained with the source data so that the model's performance degrades drastically [31, 39, 57]. To investigate this phenomenon in rPPG measurement task, we analyze the optimized degree of different feature combinations. + +Specifically, we decompose the NEST representation of the last layer into orthogonal bases with Singular Value Decomposition (SVD). Here, small eigenvalues indicate that training samples cover a small space in these corresponding bases. Thus, unseen target samples are more likely to lay out of the well-optimized space, causing a larger domain shift. To validate this, we plot the means of source and target samples on these bases in Fig. 3(a). As shown, the mean differences are larger for bases of smaller eigenvalues. To make it clearer, We further visualize the data distribution on the first and last bases in Fig. 3 (b) and (c). As shown, the distribution difference in the last base is much more significant than that of the first one. All these suggest that for the NEST representation, components of small eigenvalues are less optimized during training, and domain shifts are more likely to appear there. + +In sight of this, we propose the NEuron SStructure Coverage Maximization (NEST-CM) regularization that improves + +![](images/e21bf16425701cf4d040c06fc08e94c142c0593e717508a8e7841ee3e15836b0.jpg) +Figure 4. The detail of SVD for the $j$ -th layer in CNN. + +the eigenvalues of the last few bases. This could narrow the under-optimized space as shown in Fig. 2 (c). For a training batch with $N$ samples, NEST representation set of $j$ -th layer $B_{j} = \{o_{1,j}; o_{2,j}; \ldots; o_{N,j}\} \in R^{N * C_{j}}$ is decomposed to $S_{j} \cdot V_{j} \cdot D_{j}$ by SVD method as shown in Fig. 4, i.e., $B_{j} = S_{j} \cdot V_{j} \cdot D_{j}$ , where $V_{j} \in R^{N * C_{j}}$ and $N > C_{j}$ . The bases with small eigenvalue values are rare and under-optimized during the training stage. To reduce the occurrence probability of under-optimized NEST caused by domain shift, we stimulate under-optimized bases by maximizing the corresponding eigenvalues: + +$$ +\mathcal {L} _ {C M} = - \mathbb {E} _ {j \in U} \frac {\sum_ {i = 1} ^ {C ^ {\prime}} \lambda_ {i , j} ^ {\prime}}{\sum_ {i = 1} ^ {C} \lambda_ {i , j}}, \lambda_ {i, j} ^ {\prime} < t, \tag {1} +$$ + +where $\lambda_{i,j}$ represent the $i$ -th normalized eigenvalue in $j$ -th layer' diagonal matrix $V_{j}$ , and $\lambda_{i,j}^{\prime}$ is the eigenvalue less than the threshold $t$ . Considering different channel numbers in each layer, we design an adaptive threshold $t$ to select the relatively under-optimized bases, i.e., $t = \frac{\rho}{C_j}$ , where $\rho$ is a trade-off parameter. Besides, $C_j^\prime$ is the number of eigenvalues of $j$ -th layer under the threshold, and $U$ is the number of layers. + +NEuron SStructure Targeted Alignment. Besides the NEST-CM introduced above, we further improve the generalization ability by aligning the NEST representation of different domains. Domain invariant learning has demonstrated its effectiveness in DG problems [12, 18, 30, 38, 57, 62]. Some recent research [11, 16, 29, 60] further shows that aligning domain features within the same labels could lead to better performance. However, labels in rPPG are continuous, and it is hard to find samples with identical labels in a training batch. + +To cope with this, we propose a novel regularization term called NEuron SStructure Alignment (NEST-TA). NEST-TA builds a synthesized prototype $O_{i}^{p}$ with the same physiological information as the given targeted sample $O_{i}$ . Then, the feature within the same labels can be aligned across domains like [11, 16, 29, 60]. Specifically, we select the top $K$ samples closest to the targeted ground truth $y_{i}$ , i.e., $[O_{i,1}', \dots, O_{i,k}', \dots, O_{i,K}']$ . The NEST center of the selected samples $\overline{O}_{i}'$ is calculated by averaging the NEST presentation of selected samples. Then we calculate the direction vectors between the NEST presentation center of the selected samples $\overline{O}_{i}'$ and the NEST presentation of each sample $O_{i,k}'$ . Following that, the NEST center of the selected samples $\overline{O}_{i}'$ is calibrated to the synthesized prototype $O_{i}^{p}$ , + +which is performed by using the Gaussian distribution function to fuse these direction vectors: + +$$ +O _ {i} ^ {p} = \overline {{O ^ {\prime}}} _ {i} + \sum_ {k = 1} ^ {K} W _ {i, k} * \frac {y _ {i} - \bar {y} _ {i} ^ {\prime}}{y _ {i , k} ^ {\prime} - \bar {y} _ {i} ^ {\prime}} \left(O _ {i, k} ^ {\prime} - \overline {{O ^ {\prime}}} _ {i}\right), \tag {2} +$$ + +$$ +W _ {i, k} = \frac {G (y _ {i , k} - \overline {{y}} _ {i} ^ {\prime} , \sigma)}{\sum_ {k = 1} ^ {K} G (y _ {i , k} - \overline {{y}} _ {i} ^ {\prime} , \sigma)}, +$$ + +where $G(-, \sigma)$ is the Gaussian distribution with the mean of zero and the standard deviation of $\sigma$ , i.e., $G(y_{i,k} - \overline{y}_i', \sigma) = \frac{1}{\sqrt{2\pi}} e^{-\frac{(y_{i,k} - \overline{y}_i')^2}{2\sigma^2}}$ , where $\sigma$ is a hyper-parameter. To align the feature distribution across different domains, NEST-TA uses the cosine similarity $Cos(-, -)$ to reduce the variant features between targeted samples $O_i$ and prototype sample $O_i^p$ : + +$$ +\mathcal {L} _ {T A} = \mathbb {E} _ {i \in N} \operatorname {C o s} \left(O _ {i}, O _ {i} ^ {p}\right). \tag {3} +$$ + +NEuron SStructure Diversity Maximization. Existing works elaborate that it is effective to explore more features with discriminant information to improve the generalization [4,30,41,46]. Motivated by this, NEuron SStructure Diversity Maximization (NEST-DM) is introduced to increase the distinction between the NEST presentation of different samples as shown in Fig. 2 (e). By doing this, the network is forced to explore more features with physiological information in a contrastive learning way, which can further improve model discrimination. NEST-DM $\mathcal{L}_{DM}$ can be represented as: + +$$ +\mathcal {L} _ {D M} = - \mathbb {E} _ {i \in N} \log \left(\frac {e ^ {C o s \left(O _ {i} , \widehat {O} _ {i}\right) / \tau}}{\sum_ {j = 1 (i \neq j)} ^ {N} e ^ {C o s \left(O _ {i} , \widehat {O} _ {j}\right) / \tau}}\right), \tag {4} +$$ + +where $\widehat{O}_i$ is the NEST representation of augmented sample from $O_i$ , and $N$ and $\tau$ are the batch size and temperature parameter. The augmentation strategies for STMap will be elaborated in section 4.2. + +Model Training. For the rPPG measurement task, we chose two popular losses in this field. Specifically, L1 loss $\mathcal{L}_{L1}$ and negative Pearson's correlation coefficient $\mathcal{L}_p$ are used to constrain the predicted heart rate value and the predicted BVP signal respectively [14, 28, 37, 69]. These two losses are also used for augmented data. With the violent confrontation at the beginning of training, $\mathcal{L}_{L1}$ make network tends to predict the mean of the heart rate value distribution instead of learning the physiological feature like $\mathcal{L}_p$ . Besides, $\mathcal{L}_{NSM}$ , $\mathcal{L}_{NEST-S}$ , and $\mathcal{L}_{NEST-DM}$ also have meaningless effects on under-trained networks at the beginning. + +To solve the above-illustrated issues, we introduce the adaptation factor $\gamma = \frac{2}{1 + \exp(-10r)}$ and $r = \frac{\text{inters}_{\text{current}}}{\text{inters}_{\text{total}}}$ to suppress meaningless constraints at the early training stage, which can accelerate model convergence and improve training stability [12, 15]. The final loss $\mathcal{L}_{\text{overall}}$ can be formulated as follows: + +$$ +\begin{array}{l} \mathcal {L} _ {\text {o v e r a l l}} = k _ {1} \mathcal {L} _ {p} + \gamma \left(k _ {2} \mathcal {L} _ {L 1} + k _ {3} \mathcal {L} _ {C M} \right. \tag {5} \\ + k _ {4} \mathcal {L} _ {T A} + k _ {5} \mathcal {L} _ {D M}), \\ \end{array} +$$ + +where $k_{1} - k_{5}$ are five trade-off parameters. + +# 4. Experiments + +# 4.1. Datasets + +We elaborately select five datasets (including different motion, camera, and lighting conditions) to curate and benchmark a large-scale DG protocol: + +VIPL-HR [36] have nine scenarios, three RGB cameras, different illumination conditions, and different levels of movement. It is worth mentioning that the BVP signal and video of the dataset do not match in the time dimension, so $\mathcal{L}_p$ is not used for this dataset. We used the output of HR head as the evaluation result. Besides, we normalize STMap to 30 fps by cubic spline interpolation to solve the problem of the unstable frame rate of video [28, 50, 70]. + +V4V [47] is designed to collect data with the drastic changes of physiological indicators by simulating ten tasks such as a funny joke, 911 emergency call, and odor experience. It is worth mentioning that There are only heart rate labels in the dataset and no BVP signal labels, so $\mathcal{L}_p$ is not used for this dataset. We used the output of HR head as the evaluation result. + +BUAA [68] is proposed to evaluate the performance of the algorithm against various illumination. We only use data with illumination greater than or equal to 10 lux because underexposed images require special algorithms that are not considered in this article. We used the output of BVP head as the evaluation for BUAA, PURE, and UBFC-rPPG datasets. + +PURE [52] contains 60 RGB videos from 10 subjects with six different activities, specifically, sitting still, talking, and four rotating and moving head variations. The BVP signals are down-sampled from 60 to 30 fps with cubic spline interpolation to align with the videos. + +UBFC-rPPG [1] containing 42 face videos with sunlight and indoor illumination, the ground-truth BVP signals, and HR values were collected by CMS50E. + +# 4.2. Implementation Details + +Training Details. Our proposed method is performed under the Pytorch framework. For the generation of STMap from Video, we use the FAN1 to detect the 2D landmarks of faces [2], and other steps completely follow [36]. STMap in each dataset is sampled with a time window of 256 with step 10. Following [55, 61], we also design spatial and temporal augmentation for STMap, i.e., shuffle each row of STMap and slide the time window. In addition, color jitter and blur + +
MethodUBFCPUREBUAAVIPLV4V
MAE↓RMSE↓r↑MAE↓RMSE↓r↑MAE↓RMSE↓r↑MAE↓RMSE↓r↑MAE↓RMSE↓r↑MAE↓RMSE↓r↑MAE↓RMSE↓r↑MAE↓RMSE↓r↑
GREEN [59]8.029.180.3610.3214.270.525.827.990.5612.1818.230.2515.6421.430.06
CHROM [9]7.238.920.519.7912.760.376.098.290.5111.4416.970.2814.9219.220.08
POS [63]7.358.040.499.8213.440.345.047.120.6314.5921.260.1917.6523.220.04
DeepPhys [6]7.828.420.549.3412.560.554.786.740.6912.5619.130.1414.5219.110.14
TS-CAN [25]7.638.250.559.1212.380.574.846.890.6812.3418.940.1614.7719.960.12
Rhythmnet* [36]5.797.910.787.3910.490.773.385.170.848.9712.160.4910.1614.570.34
Dual-GAN* [28]5.557.620.797.2410.270.783.415.230.848.8811.690.5010.0414.440.35
BVPNet* [7]5.437.710.807.2310.250.783.695.480.818.4511.640.5110.0114.350.36
AD*+ [12]5.928.080.767.4210.610.733.495.490.828.4111.710.5310.4714.640.32
GroupDRO*+ [38]5.737.970.787.6910.830.783.415.210.838.3511.670.549.9414.290.36
Coral*+ [53]5.898.040.767.5910.870.723.645.740.808.6811.910.5310.3214.420.32
VREx*+ [18]5.597.680.817.2410.140.783.275.010.868.3711.620.549.8214.160.37
NCDG*+ [57]5.317.560.827.3210.350.773.125.160.858.4711.810.5210.1414.460.34
Baseline* [36]5.797.910.787.3910.490.773.385.170.848.9712.160.4910.1614.570.34
NEST*+ w/o LCM4.927.190.836.919.940.782.794.620.898.2111.560.559.5813.580.38
NEST*+ w/o LTA4.977.280.836.819.770.813.034.940.868.1911.470.559.5213.520.38
NEST*+ w/o LDM4.897.110.846.879.810.792.934.770.888.1411.430.569.4913.490.40
NEST*+4.676.790.866.719.590.812.884.690.897.8611.150.589.2713.790.41
+ +Table 1. HR estimation results on MSDG protocol. It is worth noting that we have removed the GRU mechanism in Rhythmnet because it is an unfair comparison with using GRU to aggregate the temporal information of multiple clips instead of only one clip; * means that these methods use the STMap as the input of CNN and use the STMap augmentation strategy as mentioned in section 4.2; + means that these methods are based on baseline (Rhythmnet without GRU). + +augmentation are used for STMap [5]. Then the STMap $x \in R^{25 \times 256 \times 3}$ is resized to $x' \in R^{64 \times 256 \times 3}$ as the input of the network. We use the ResNet-18 with removing the max pooling layer as the feature extractor, and the final full connection layer is used as the HR estimation head. At the same time, we designed a BVP signal regression head containing four blocks, each block involves a Conv., transpose, and two Conv. layers with batch normalization and Relu function. Hyper-parameter $\sigma$ , $\rho$ and $\tau$ are set to 5, 0.1, and 0.2 according to the experimental result, and the trade-off parameter $k_{1} - k_{5}$ are set to 1, 0.1, 0.001, 0.1 and 0.01 based on the scale of losses. Adam optimizer with a learning rate of 0.001 is used for training. The batch size and iterations are set to 1024 and 20000. + +Performance Metrics. Following existing methods [6, 28, 37, 50, 70], standard deviation (SD), mean absolute error (MAE), root mean square error (RMSE), and Pearson's correlation coefficient (r) are used to evaluate the HR estimation. MAE, RMSE, and r are used to evaluate the HRV measurement, including low frequency (LF), high frequency (HF), and LF/HF. + +# 4.3. Multi-Source Domain Generalization + +# 4.3.1 HR Estimation + +For remote HR estimation, all five datasets (i.e., UBFC [1], PURE [52], BUAA [68], VIPL [36], and V4V [47]) are used to evaluate the model generalization. Specifically, four datasets are used for training, and then the other dataset is + +used to test. Firstly, we perform the iPhys $^2$ to evaluate three traditional method (i.e., GREEN [59], CHROM [9], and POS [63]) as shown in Tab. 1. Then, we reproduce five popular DL-based methods (i.e., DeepPhys [6], TSCAN [25], Rhythmnet [36], and BVPNet [7], and DualGAN [28]) for comparison on the same setting. Subsequently, we directly applied the four DG methods (i.e., AD [12], GroupDRO [38], Coral [53], VReX [18], and NCDG [57]) to Rhythmnet with using DeepDG $^3$ . + +The performance of traditional methods. As shown in Tab. 1, traditional algorithms show relatively acceptable results on some simple scenes, such as BUAA and UBFC, but very poor results on other datasets. It is difficult for these traditional algorithms to extract the physiological information on the V4V dataset, which shows that the traditional methods cannot deal with the noise caused by the complex environment and non-rigid motions. + +The performance of DL-based methods. For DL-based methods, DeepPhys and TS-CAN even perform worse than traditional methods. Rhythmnet, BVPNet, and Dual-GAN perform much better than both other DL-based methods and traditional methods, which suggests that STMap is the more robust input of CNN than the aligned video. Compared with Rhythmnet, BVPNet and Dual-GAN have not significant improvement on the MSDG protocol. Because BVP- + +
TargetMethodLF-(u.n)HF-(u.n)LF/HFHR-(bpm)
MAE↓RMSE↓r↑MAE↓RMSE↓r↑MAE↓RMSE↓r↑MAE↓RMSE↓r↑
UBFCGREEN [59]0.23550.28410.09240.23550.28410.09240.66950.95120.04678.01849.17760.3634
CHROM [9]0.22210.28170.06980.22210.28170.06980.67081.05420.10547.22918.92240.5123
POS [63]0.23640.28610.13590.23640.28610.13590.65150.95350.13457.35398.04020.4923
Baseline0.06210.08130.18730.06210.08130.18730.19850.26670.30435.15427.46720.8165
NEST0.05970.07820.20170.05970.07820.20170.21380.28240.31794.74716.88760.8546
PUREGREEN [59]0.25390.30020.03260.25390.30020.03260.65250.89320.041710.324714.26930.4952
CHROM [9]0.20960.27510.10590.20960.27510.07590.54040.82660.11739.791412.75680.3732
POS [63]0.19590.25710.16840.19590.25710.16840.53730.8460.14339.827313.44140.3432
Baseline0.06710.09230.60460.06710.09230.60460.28640.41840.55268.254211.17650.6832
NEST0.06350.08740.64220.06350.08740.64220.22550.35050.57347.688910.47830.7255
BUAAGREEN [59]0.34720.39510.08710.34720.39510.08710.64530.86320.09215.82317.98820.5624
CHROM [9]0.37860.32370.06820.37860.32370.06820.68130.88360.07156.09348.29380.5165
POS [63]0.31980.37620.09620.31980.37620.09620.62750.84240.11275.04077.11980.6374
Baseline0.14510.16810.28910.14510.16810.28910.55640.69040.29143.78526.32370.7492
NEST0.14360.16650.29550.14360.16650.29550.55140.68840.30043.37235.88060.7647
+ +Net and Dual-GAN consider the performance of the single domain instead of focusing on solving domain shift. Therefore, Rhythmnet is also selected as a baseline to further analyze the generalization on MSDG protocol. + +The performance of DG methods. To further improve the generalization of the model, we apply the existing popular DG methods to baseline (Rhythmnet) as shown in Tab. 1. There is a slight model degradation of AD [12], Coral [53], and GroupDRO [38] on some target datasets. The reason behind this is that these algorithms employ inaccurate domain labeling, specifically, a dataset often with different acquisition devices, lighting, and scene cannot be considered as the same domain. On the contrary, VREx [18] and NCDG [57] can slightly improve the performance of Rhythmnet without utilizing the domain label. Inaccurate domain division will make it difficult for the DG algorithm to effectively align the feature distribution across different domains, which is termed as agnostic domain generalization. + +Without the domain labels, NEST experimentally manifested significant improvement on MSDG protocol. For the target dataset BUAA, the results without NEST-CM $\mathcal{L}_{CM}$ show the best performance. Because NEST-CM may cause the model degradation with activating meaningless features at the instance level, especially for the BUAA with the slight domain shift. + +# 4.3.2 HRV Estimation + +We also conduct experiments for HRV (i.e., LF, HF, and LF/HF) and HR measurement on the PURE, UBFC, and BUAA. The reason for choosing only these three datasets is that VIPL and V4V do not have reliable ground-truth + +Table 2. HRV and HR estimation results on the MSDG protocol. + +
MethodPUREBUAA
MAE↓RMSE↓r↑MAE↓RMSE↓r↑
GREEN [59]10.3214.270.525.827.990.56
CHROM [9]9.7912.760.376.098.290.51
POS [63]9.8213.440.345.047.120.63
Baseline7.5410.130.732.973.280.76
NEST6.079.060.762.562.730.78
+ +Table 3. HR estimation results on SSDG protocol by training on the UBFC-rPPG. + +BVP signals, as mentioned in subsection 4.1. We also perform three tradition method (GREEN [59], CHMROM [9], POS [63]) by iPhys for comparison with our method as shown in Tab. 2. With a more robust NEST presentation distribution, our method also can get better performance than baseline and other traditional methods in all metrics. + +# 4.4. Single-Source Domain Generalization + +For a worst-case scenario, the model is trained on only one dataset to generalize to the unseen domains, namely single-source domain generalization (SSDG). Based on this, we select UBFC-rPPG as the source domain and then select PURE and BUAA as the target domain to further evaluate the generalization of NEST as shown in Tab. 3. As we can see, NEST achieves desired performance by training on only one dataset. The reason is that: (1) NEST-CM can avoid the appearance of abnormal NEST; (2) NEST-TA can find the shared and robust feature of the samples within the same label; (3) NEST-DM can distinguish the boundaries between different samples to improve the generalization. + +
MethodSD↓MAE↓RMSE↓r↑
SAMC [58]18.015.921.00.11
POS [63]15.311.517.20.30
CHROM [9]15.111.416.90.28
I3D [3]15.912.015.90.07
DeepPhy [6]13.611.013.80.11
BVPNet [8]7.755.347.850.70
RhythmNet [36]8.115.308.140.76
CVD [37]7.925.027.970.79
Physformer [70]7.744.977.790.78
Dual-GAN [28]7.634.937.680.81
Baseline8.045.218.070.77
NEST7.494.767.510.84
+ +Table 4. HR estimation results by our method and several state-of-the-art methods on the VIPL-HR database. + +![](images/846f917949be5f611854fab9d3fdc219c7ceb03be288304b79b6f0faadc4342d.jpg) +Figure 5. Impacts of the hyperparameter (a) $\rho$ , (b) $\sigma$ , and (c) $\tau$ of our proposed method. + +# 4.5. Intra-Dataset Testing on VIPL + +Our proposed method can also improve the effectiveness of the model on intra-dataset testing by reducing the harmful effects of abnormal disturbances. Following [28, 36, 37, 70], we test NEST by using a 5-fold cross-validation protocol on VIPL-HR. Our method compared with ten baseline methods, including the traditional methods (SAMC [58], POS [63] and CHROM [9]) and the DL-based methods (I3D [3], DeepPhy [6], RhythmNet [36], CVD [37], BVP-Net [7], Physformer [70], and Dual-GAN [28]). The results of these methods are directly from [28, 70] as shown in Tab. 4. + +As shown in Tab. 4, the proposed methods outperform all the methods under all metrics. Compared with the baseline, it also gets a noticeable promotion, i.e., 0.4 in MAE and 0.53 in RMSE. It is worth noting that the VIPL dataset has multiple complex scenes and recording devices and cannot be considered as a single domain, termed the agnostic domain problem. This proves that NEST can effectively alleviate the domain shift without domain labels. + +# 4.6. Further Analysis + +We conduct further experiments on MSDG protocol (testing on UBFC with training on the other four datasets). Impact of $\rho$ in NEST-CM. The hyperparameter $\rho$ determines how many bases need to be strengthened by NEST-CM $\mathcal{L}_{CM}$ . As illustrated in Fig. 5 (a), NEST could achieve smaller RMSE when $\rho = 0.1$ , indicating that the bases with small eigenvalue require more attention and fine-tuning. + +![](images/efa6d736dca4e50d7da261bf7c761e53794418d131d14c6ef8cc8dcc547bfa9e.jpg) +Figure 6. The pearson's correlation coefficient of the NEST representation of the samples with different labels for the anchor HR 60. + +Impact of $\sigma$ in NEST-TA. The hyperparameter $\sigma$ tradeoffs the contribution of each selected sample to form the synthesized prototype. As illustrated in Fig. 5, NEST could achieve the best performance with $\sigma = 5$ , which shows that the appropriate fusion of the nearby sample is conducive to improving model generalization. The specific reason behind this is that: (1) the synthesized prototype based on multi-domain samples can be used to enhance the domain invariant feature by NEST-TA; (2) NEST-TA can effectively aggregate the label relationship into NEST presentation distribution, which is performed by improving the similarity of NEST representation between the adjacent label samples. + +To further clarify the second reason, we calculate the Pearson's correlation coefficient of the NEST representation of the samples with different labels for the anchor HR 60, as shown in Fig. 6. It is obvious that the distribution changes with NEST-TA loss are more smooth across the continuous HR labels, which allows OOD samples with slight domain shifts to predict nearby heart rates around the ground truth HR value instead of other abnormal HR values. + +Impact of $\tau$ in NEST-DM. The hyperparameter $\tau$ determines the pushing strength of the different samples. With the smallest $\tau$ , NEST-DM $\mathcal{L}_{DM}$ has the overemphasized constraints to distinguish the NEST representation between different samples, which causes a harmful impact on the discriminant ability of the model. Overall, NEST-DM $\mathcal{L}_{DM}$ is an effective way to improve the generalization of model. + +# 5. Conclusion + +This paper presents the NEural SStructure (NEST) modeling, a novel framework that addresses the domain generalization problem in rPPG measurement for the first time. Our framework contains three regularization terms defined in the view of neural structure, which improves the generalization capacity in the different aspects. All three terms do not rely on domain labels, which are especially suitable in rPPG, as defining domains is difficult in this task. To compare the generalization of the existing algorithm and our method, we establish the first DG protocol for this task. We experimentally manifested the substantial improvement of our method on both MSDG and SSDG protocols. + +# References + +[1] Serge Bobbia, Richard Macwan, Yannick Benezeth, Alamin Mansouri, and Julien Dubois. Unsupervised skin tissue segmentation for remote photoplethysmography. Pattern Recognition Letters, 124:82-90, 2019. +[2] Adrian Bulat and Georgios Tzimiropoulos. How far are we from solving the 2d & 3d face alignment problem? (and a dataset of 230,000 3d facial landmarks). In Proc. ICCV, 2017. +[3] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proc. IEEE CVPR, pages 6299-6308, 2017. +[4] Beidi Chen, Tri Dao, Eric Winsor, Zhao Song, Atri Rudra, and Christopher Ré. Scatterbrain: Unifying sparse and low-rank attention. NeurIPS, 34:17413-17426, 2021. +[5] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, pages 1597-1607. PMLR, 2020. +[6] Weixuan Chen and Daniel McDuff. Deep learning: Video-based physiological measurement using convolutional attention networks. In Proc. ECCV, pages 349-365, 2018. +[7] Abhijit Das, Hao Lu, Hu Han, Antitza Dantcheva, Shiguang Shan, and Xilin Chen. Bvpnet: Video-to-bvp signal prediction for remote heart rate estimation. In $FG$ , pages 01-08. IEEE, 2021. +[8] Abhijit Das, Hao Lu, Hu Han, Antitza Dantcheva, Shiguang Shan, and Xilin Chen. Bvpnet: Video-to-bvp signal prediction for remote heart rate estimation. In $FG$ , pages 01-08. IEEE, 2021. +[9] Gerard De Haan and Vincent Jeanne. Robust pulse rate from chrominance-based rppg. IEEE Trans. Biomed. Eng., 60(10):2878-2886, 2013. +[10] Gerard De Haan and Arno Van Leest. Improved motion robustness of remote-ppg by using the blood volume pulse signature. Physiol. Meas., 35(9):1913, 2014. +[11] Zhijie Deng, Yucen Luo, and Jun Zhu. Cluster alignment with a teacher for unsupervised domain adaptation. In Proc. CVPR, pages 9944–9953, 2019. +[12] Yaroslav Ganin and Victor Lempitsky. Unsupervised domain adaptation by backpropagation. In ICML, pages 1180-1189. PMLR, 2015. +[13] Ming-Shiang Chen Gee-Sern Hsu, ArulMurugan Ambikapathi. Deep learning with time-frequency representation for pulse estimation. In Proc. IJCB, pages 642-650, 2017. +[14] Min Hu, Fei Qian, Dong Guo, Xiaohua Wang, Lei He, and Fuji Ren. Eta-rppgnet: effective time-domain attention network for remote heart rate measurement. IEEE Trans. on Instrum. and Meas., 70:1-12, 2021. +[15] Yunpei Jia, Jie Zhang, Shiguang Shan, and Xilin Chen. Single-side domain generalization for face anti-spoofing. In Proc. CVPR, pages 8484-8493, 2020. +[16] Xiang Jiang, Qicheng Lao, Stan Matwin, and Mohammad Havaei. Implicit class-conditioned domain alignment for unsupervised domain adaptation. In ICML, pages 4816-4827. PMLR, 2020. + +[17] Viktor Kessler, Patrick Thiam, Mohammadreza Amirian, and Friedhelm Schwenker. Pain recognition with camera photoplethysmography. In IEEE IPTA, pages 1-5, 2017. +[18] David Krueger, Ethan Caballero, Joern-Henrik Jacobsen, Amy Zhang, Jonathan Binas, Dinghuai Zhang, Remi Le Priol, and Aaron Courville. Out-of-distribution generalization via risk extrapolation (rex). In ICML, pages 5815-5826. PMLR, 2021. +[19] Eugene Lee, Evan Chen, and Chen-Yi Lee. Meta-rppg: Remote heart rate estimation using a transductive meta-learner. In ECCV, pages 392–409. Springer, 2020. +[20] Alexander Lehner, Stefano Gasperini, Alvaro Marcos-Ramiro, Michael Schmidt, Mohammad-Ali Nikouei Mahani, Nassir Navab, Benjamin Busam, and Federico Tombari. 3d-vfield: Adversarial augmentation of point clouds for domain generalization in 3d object detection. In Proc. CVPR, pages 17295-17304, 2022. +[21] Magdalena Lewandowska, Jacek Ruminski, Tomasz Kocejko, and Jedrzej Nowak. Measuring pulse rate with a webcam—a non-contact method for evaluating cardiac activity. In FedCSIS, pages 405-410. IEEE, 2011. +[22] Daiqing Li, Junlin Yang, Karsten Kreis, Antonio Torralba, and Sanja Fidler. Semantic segmentation with generative models: Semi-supervised learning and strong out-of-domain generalization. In Proc. CVPR, pages 8300-8311, 2021. +[23] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy Hospedales. Learning to generalize: Meta-learning for domain generalization. In Proc. AAAI, 2018. +[24] Chuang Lin, Zehuan Yuan, Sicheng Zhao, Peize Sun, Changhu Wang, and Jianfei Cai. Domain-invariant disentangled network for generalizable object detection. In Proce. CVPR, pages 8771-8780, 2021. +[25] Xin Liu, Josh Fromm, Shwetak Patel, and Daniel McDuff. Multi-task temporal shift attention networks for on-device contactless vitals measurement. NeurIPS, 33:19400-19411, 2020. +[26] Xin Liu, Ziheng Jiang, Josh Fromm, Xuhai Xu, Shwetak Patel, and Daniel McDuff. Metaphys: few-shot adaptation for non-contact physiological measurement. In Proc. HIL, pages 154-163, 2021. +[27] Zhenhua Liu, Jizheng Xu, Xiulian Peng, and Ruiqin Xiong. Frequency-domain dynamic pruning for convolutional neural networks. NeurIPS, 31, 2018. +[28] Hao Lu, Hu Han, and S Kevin Zhou. Dual-gan: Joint bvp and noise modeling for remote physiological measurement. In Proc. IEEE CVPR, pages 12404-12413, 2021. +[29] Yawei Luo, Ping Liu, Liang Zheng, Tao Guan, Junqing Yu, and Yi Yang. Category-level adversarial adaptation for semantic segmentation using purified features. IEEE Trans. Pattern Anal. Mach. Intell., 2021. +[30] Fangrui Lv, Jian Liang, Shuang Li, Bin Zang, Chi Harold Liu, Ziteng Wang, and Di Liu. Causality inspired representation learning for domain generalization. In Proc. CVPR, pages 8046-8056, 2022. +[31] Lei Ma, Felix Juefei-Xu, Fuyuan Zhang, Jiyuan Sun, Minhui Xue, Bo Li, Chunyang Chen, Ting Su, Li Li, Yang Liu, et al. Deep gauge: Multi-granularity testing criteria for deep learning systems. In Proc. ASE, pages 120-131, 2018. + +[32] Daniel McDuff. Camera measurement of physiological vital signs. CSUR, 2021. +[33] Daniel McDuff. Applications of camera-based physiological measurement beyond healthcare. In CVSM, pages 165-177. Elsevier, 2022. +[34] Daniel McDuff, Sarah Gontarek, and Rosalind W Picard. Improvements in remote cardiopulmonary measurement using a five band digital camera. IEEE Trans. Biomed. Eng., 61(10):2593-2601, 2014. +[35] Xuesong Niu, Hu Han, Shiguang Shan, and Xilin Chen. Synrhythm: Learning a deep heart rate estimator from general to specific. In Proc. IEEE ICPR, pages 3580-3585, 2018. +[36] Xuesong Niu, Shiguang Shan, Hu Han, and Xilin Chen. Rhythmnet: End-to-end heart rate estimation from face via spatial-temporal representation. IEEE Trans. on Image Process., 29:2409-2423, 2020. +[37] Xuesong Niu, Zitong Yu, Hu Han, Xiaobai Li, Shiguang Shan, and Guoying Zhao. Video-based remote physiological measurement via cross-verified feature disentangling. In Proc. ECCV, 2020. +[38] Giambattista Parascandolo, Alexander Neitz, Antonio Orvieto, Luigi Gresele, and Bernhard Scholkopf. Learning explanations that are hard to vary. *ICLR2022*, 2022. +[39] Kexin Pei, Yinzhi Cao, Junfeng Yang, and Suman Jana. Deepxplore: Automated whitebox testing of deep learning systems. In Proc. SOSP, pages 1-18, 2017. +[40] Duo Peng, Yinjie Lei, Munawar Hayat, Yulan Guo, and Wen Li. Semantic-aware domain generalized segmentation. In Proc. CVPR, pages 2594-2605, 2022. +[41] Vihari Piratla, Praneeth Netrapalli, and Sunita Sarawagi. Efficient domain generalization via common-specific low-rank decomposition. In ICML, pages 7728-7738. PMLR, 2020. +[42] Ming-Zher Poh, Daniel J McDuff, and Rosalind W Picard. Non-contact, automated cardiac pulse measurements using video imaging and blind source separation. Opt. Express, 18(10):10762-10774, 2010. +[43] Fengchun Qiao, Long Zhao, and Xi Peng. Learning to learn single domain generalization. In Proc. CVPR, pages 12556-12565, 2020. +[44] Ying Qiu, Yang Liu, Juan Arteaga-Falconi, Haiwei Dong, and Abdulmotaleb El Saddik. Evm-cnn: Real-time contactless heart rate estimation from facial video. IEEE Trans. Multimedia, 2019. +[45] Attila Reiss, Ina Indlekofer, Philip Schmidt, and Kristof Van Laerhoven. Deep ppg: large-scale heart rate estimation with convolutional neural networks. Sensors, 19(14):3079, 2019. +[46] Jiahuan Ren, Zhao Zhang, Richang Hong, Mingliang Xu, Hajun Zhang, Mingbo Zhao, and Meng Wang. Robust low-rank convolution network for image denoising. In Proc. ACM MM, pages 6211-6219, 2022. +[47] Ambareesh Revanur, Zhihua Li, Umur A Ciftci, Lijun Yin, and László A Jeni. The first vision for vitals (v4v) challenge for non-contact video-based physiological estimation. In Proc. CVPR workshop, pages 2760-2767, 2021. +[48] Shiv Shankar, Vihari Piratla, Soumen Chakrabarti, Siddhartha Chaudhuri, Preethi Jyothi, and Sunita Sarawagi. + +Generalizing across domains via cross-gradient training. ICLR, 2018. +[49] Rencheng Song, Huan Chen, Juan Cheng, Chang Li, Yu Liu, and Xun Chen. Pulsegan: Learning to generate realistic pulse waveforms in remote photoplethysmography. IEEE J-BHI, pages 1-1, 2021. +[50] Rencheng Song, Senle Zhang, Chang Li, Yunfei Zhang, Juan Cheng, and Xun Chen. Heart rate estimation from facial videos using a spatiotemporal representation with convolutional neural networks. IEEE Trans. Instrum. Meas., 2020. +[51] Radim Špetlík, Vojtech Franc, and Jírí Matas. Visual heart rate estimation with convolutional neural network. In Proc. BMVC, pages 3-6, 2018. +[52] Ronny Stricker, Steffen Müller, and Horst-Michael Gross. Non-contact video-based pulse rate measurement on a mobile service robot. In Proc. IEEE ISRHIC, pages 1056-1062, 2014. +[53] Baochen Sun and Kate Saenko. Deep coral: Correlation alignment for deep domain adaptation. In ECCV, pages 443-450. Springer, 2016. +[54] Bo Sun, Qinglan Wei, Liandong Li, Qihua Xu, Jun He, and Lejun Yu. Lstm for dynamic emotion and group emotion recognition in the wild. In Proc. ACM ICMI, pages 451-457, 2016. +[55] Zhaodong Sun and Xiaobai Li. Contrast-phys: Unsupervised video-based remote physiological measurement via spatiotemporal contrast. In Proc. ECCV, 2022. +[56] Ling Tang, Wen Shen, Zhanpeng Zhou, Yuefeng Chen, and Quanshi Zhang. Defects of convolutional decoder networks in frequency representation. arXiv preprint arXiv:2210.09020, 2022. +[57] Chris Xing Tian, Haoliang Li, Xiaofei Xie, Yang Liu, and Shiqi Wang. Neuron coverage-guided domain generalization. IEEE Trans. on Pattern Anal. Mach. Intell., 2022. +[58] Sergey Tulyakov, Xavier Alameda-Pineda, Elisa Ricci, Lijun Yin, Jeffrey F Cohn, and Nicu Sebe. Self-adaptive matrix completion for heart rate estimation from face videos under realistic conditions. In Proc. IEEE CVPR, pages 2396-2404, 2016. +[59] Wim Verkruysse, Lars O Svaasand, and J Stuart Nelson. Remote plethysmographic imaging using ambient light. Opt. Express, 16(26):21434-21445, 2008. +[60] Tan Wad, Qianru Sun, Sugiri Pranata, Karlekar Jayashree, and Hanwang Zhang. Equivalence and invariance inductive bias for learning from insufficient data. In ECCV, pages 241-258. Springer, 2022. +[61] Hao Wang, Euijoon Ahn, and Jinman Kim. Self-supervised representation learning framework for remote physiological measurement using spatiotemporal augmentation loss. In ACM MM, 2022. +[62] Jindong Wang, Cuiling Lan, Chang Liu, Yidong Ouyang, Tao Qin, Wang Lu, Yiqiang Chen, Wenjun Zeng, and Philip Yu. Generalizing to unseen domains: A survey on domain generalization. IEEE Trans. on Knowl. Data Eng., 2022. +[63] Wenjin Wang, Albertus C den Brinker, Sander Stuijk, and Gerard de Haan. Algorithmic principles of remote ppg. IEEE Trans. Biomed. Eng., 64(7):1479-1491, 2017. + +[64] Wenjin Wang, Sander Stuijk, and Gerard De Haan. Exploiting spatial redundancy of image sensor for motion robust rppg. IEEE Trans. Biomed. Eng., 62(2):415-425, 2015. +[65] Yunhe Wang, Chang Xu, Shan You, Dacheng Tao, and Chao Xu. Cnnpack: Packing convolutional neural networks in the frequency domain. In NeurIPS, volume 29, 2016. +[66] Zijian Wang, Yadan Luo, Ruihong Qiu, Zi Huang, and Mahsa Baktashmotlagh. Learning to diversify for single domain generalization. In Proc. CVPR, pages 834-843, 2021. +[67] Zhi-Kuan Wang, Ying Kao, and Chiou-Ting Hsu. Vision-based heart rate estimation via a two-stream cnn. In Proc. IEEE ICIP, pages 3327-3331, 2019. +[68] Lin Xi, Weihai Chen, Changchen Zhao, Xingming Wu, and Jianhua Wang. Image enhancement for remote photoplethysmography in a low-light environment. In FG, pages 1-7. IEEE, 2020. +[69] Zitong Yu, Wei Peng, Xiaobai Li, Xiaopeng Hong, and Guoying Zhao. Remote heart rate measurement from highly compressed facial videos: an end-to-end deep learning solution with video enhancement. In Proc. IEEE ICCV, pages 151-160, 2019. +[70] Zitong Yu, Yuming Shen, Jingang Shi, Hengshuang Zhao, Philip Torr, and Guoying Zhao. Physformer: Facial videobased physiological measurement with temporal difference transformer. IEEE CVPR, 2022. +[71] Xiangyu Yue, Yang Zhang, Sicheng Zhao, Alberto Sangiovanni-Vincentelli, Kurt Keutzer, and Boqing Gong. Domain randomization and pyramid consistency: Simulation-to-real generalization without accessing target domain data. In Proc.CVPR, pages 2100-2110, 2019. \ No newline at end of file diff --git a/2023/Neuron Structure Modeling for Generalizable Remote Physiological Measurement/images.zip b/2023/Neuron Structure Modeling for Generalizable Remote Physiological Measurement/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..16101cc54f02dec82571c4b87e7e24b0e0c6c14a --- /dev/null +++ b/2023/Neuron Structure Modeling for Generalizable Remote Physiological Measurement/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2999e1666616e1cab204f47d2545c68bfc8b90a38f459ce06639e705645f451f +size 771268 diff --git a/2023/Neuron Structure Modeling for Generalizable Remote Physiological Measurement/layout.json b/2023/Neuron Structure Modeling for Generalizable Remote Physiological Measurement/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..1ec5affd1233536faa51330277124fb8df013665 --- /dev/null +++ b/2023/Neuron Structure Modeling for Generalizable Remote Physiological Measurement/layout.json @@ -0,0 +1,9370 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 88, + 103, + 506, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 103, + 506, + 138 + ], + "spans": [ + { + "bbox": [ + 88, + 103, + 506, + 138 + ], + "type": "text", + "content": "Neuron Structure Modeling for Generalizable Remote Physiological Measurement" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 148, + 160, + 449, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 160, + 449, + 175 + ], + "spans": [ + { + "bbox": [ + 148, + 160, + 449, + 175 + ], + "type": "text", + "content": "Hao LU" + }, + { + "bbox": [ + 148, + 160, + 449, + 175 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 148, + 160, + 449, + 175 + ], + "type": "text", + "content": ", Zitong YU" + }, + { + "bbox": [ + 148, + 160, + 449, + 175 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 148, + 160, + 449, + 175 + ], + "type": "text", + "content": ", Xuesong NIU" + }, + { + "bbox": [ + 148, + 160, + 449, + 175 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 148, + 160, + 449, + 175 + ], + "type": "text", + "content": ", Yingcong CHEN" + }, + { + "bbox": [ + 148, + 160, + 449, + 175 + ], + "type": "inline_equation", + "content": "^{1,2,*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 175, + 464, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 175, + 464, + 190 + ], + "spans": [ + { + "bbox": [ + 133, + 175, + 464, + 190 + ], + "type": "text", + "content": "1The Hong Kong University of Science & Technology (Guangzhou)," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 190, + 542, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 190, + 542, + 232 + ], + "spans": [ + { + "bbox": [ + 54, + 190, + 542, + 232 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 54, + 190, + 542, + 232 + ], + "type": "text", + "content": "The Hong Kong University of Science & Technology, " + }, + { + "bbox": [ + 54, + 190, + 542, + 232 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 54, + 190, + 542, + 232 + ], + "type": "text", + "content": "Great Bay University, " + }, + { + "bbox": [ + 54, + 190, + 542, + 232 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 54, + 190, + 542, + 232 + ], + "type": "text", + "content": "Kuaishou Technology hlu585@connect.hkust-gz.edu.cn, zitong.yu@ieee.org, nxsedson@gmail.com, yingcongchen@ust.hk" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 258, + 192, + 271 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 258, + 192, + 271 + ], + "spans": [ + { + "bbox": [ + 143, + 258, + 192, + 271 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 284, + 290, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 284, + 290, + 559 + ], + "spans": [ + { + "bbox": [ + 46, + 284, + 290, + 559 + ], + "type": "text", + "content": "Remote photoplethysmography (rPPG) technology has drawn increasing attention in recent years. It can extract Blood Volume Pulse (BVP) from facial videos, making many applications like health monitoring and emotional analysis more accessible. However, as the BVP signal is easily affected by environmental changes, existing methods struggle to generalize well for unseen domains. In this paper, we systematically address the domain shift problem in the rPPG measurement task. We show that most domain generalization methods do not work well in this problem, as domain labels are ambiguous in complicated environmental changes. In light of this, we propose a domain-label-free approach called NEuron SStructure modeling (NEST). NEST improves the generalization capacity by maximizing the coverage of feature space during training, which reduces the chance for under-optimized feature activation during inference. Besides, NEST can also enrich and enhance domain invariant features across multi-domain. We create and benchmark a large-scale domain generalization protocol for the rPPG measurement task. Extensive experiments show that our approach outperforms the state-of-the-art methods on both cross-dataset and intra-dataset settings. The codes are available at https://github.com/LuPaoPao/NEST." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 567, + 128, + 580 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 567, + 128, + 580 + ], + "spans": [ + { + "bbox": [ + 47, + 567, + 128, + 580 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 587, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 587, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 587, + 287, + 696 + ], + "type": "text", + "content": "Physiological signals such as heart rate (HR), and heart rate variability (HRV), respiration frequency (RF) are important body indicators that serve not only as vital signs but also track the level of sympathetic activation [17, 33, 54]. Traditional physiological measurements, such as electrocardiograms, heart rate bands, and finger clip devices, have high accuracy. However, they are costly, intrusive, and uncomfortable to wear for a long time. Remote photoplethysmography (rPPG) can extract blood volume pulse (BVP)" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 308, + 258, + 352, + 370 + ], + "blocks": [ + { + "bbox": [ + 308, + 258, + 352, + 370 + ], + "lines": [ + { + "bbox": [ + 308, + 258, + 352, + 370 + ], + "spans": [ + { + "bbox": [ + 308, + 258, + 352, + 370 + ], + "type": "image", + "image_path": "e9b3e153f156f2e4f7039efef49987157709640e9b430dac9ed9858c5c5d228c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 348, + 370, + 497, + 380 + ], + "lines": [ + { + "bbox": [ + 348, + 370, + 497, + 380 + ], + "spans": [ + { + "bbox": [ + 348, + 370, + 497, + 380 + ], + "type": "text", + "content": "(a) Typical samples from different datasets" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 356, + 258, + 400, + 369 + ], + "blocks": [ + { + "bbox": [ + 356, + 258, + 400, + 369 + ], + "lines": [ + { + "bbox": [ + 356, + 258, + 400, + 369 + ], + "spans": [ + { + "bbox": [ + 356, + 258, + 400, + 369 + ], + "type": "image", + "image_path": "74734f2b0862877af9122dbe060ea2681ef4e4059563e49850d9a0741960817f.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 402, + 258, + 447, + 369 + ], + "blocks": [ + { + "bbox": [ + 402, + 258, + 447, + 369 + ], + "lines": [ + { + "bbox": [ + 402, + 258, + 447, + 369 + ], + "spans": [ + { + "bbox": [ + 402, + 258, + 447, + 369 + ], + "type": "image", + "image_path": "d859c2131900ef0d19b9834e4e7bdc4bb0635d3b2f20f29419d29f49dc03ad72.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 450, + 258, + 495, + 369 + ], + "blocks": [ + { + "bbox": [ + 450, + 258, + 495, + 369 + ], + "lines": [ + { + "bbox": [ + 450, + 258, + 495, + 369 + ], + "spans": [ + { + "bbox": [ + 450, + 258, + 495, + 369 + ], + "type": "image", + "image_path": "1a20cf70a8f365ec011d1ce71fbb4d3e44bf0df7002125b2074baab67bfaf0ba.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 501, + 258, + 545, + 369 + ], + "blocks": [ + { + "bbox": [ + 501, + 258, + 545, + 369 + ], + "lines": [ + { + "bbox": [ + 501, + 258, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 501, + 258, + 545, + 369 + ], + "type": "image", + "image_path": "7d06aec4f82f4d151d3d618b1d4b1b395fb4971a43def901c78f1a38d9be854a.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 309, + 380, + 533, + 495 + ], + "blocks": [ + { + "bbox": [ + 309, + 380, + 533, + 495 + ], + "lines": [ + { + "bbox": [ + 309, + 380, + 533, + 495 + ], + "spans": [ + { + "bbox": [ + 309, + 380, + 533, + 495 + ], + "type": "image", + "image_path": "62d042aca48080d8e790247659b54c3194e234ea51e87f4cbc651b253fe93d81.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 332, + 497, + 531, + 505 + ], + "lines": [ + { + "bbox": [ + 332, + 497, + 531, + 505 + ], + "spans": [ + { + "bbox": [ + 332, + 497, + 531, + 505 + ], + "type": "text", + "content": "(b) The performance of different methods on DG protocol" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 505, + 547, + 559 + ], + "lines": [ + { + "bbox": [ + 304, + 505, + 547, + 559 + ], + "spans": [ + { + "bbox": [ + 304, + 505, + 547, + 559 + ], + "type": "text", + "content": "Figure 1. (a) Typical samples from different publicly rPPG datasets: VIPL-HR [36], V4V [47], UBFC-rPPG [1], BUAA [68], PURE [52]. (b) The performance of different methods on DG protocol (test on the UBFC-rPPG dataset with training on the VIPL, V4V, PURE, and BUAA)." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 566, + 547, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 566, + 547, + 674 + ], + "spans": [ + { + "bbox": [ + 304, + 566, + 547, + 674 + ], + "type": "text", + "content": "from face video, which analyzes the periodic changes of the light absorption of the skin caused by heartbeats. Then various physiological indicators (such as HR and HRV) can be calculated based on BVP signals [37, 69, 70]. With nonintrusion and convenience, the rPPG-based physiological measurement method can only use an ordinary camera to monitor physiological indicators and gradually become a research hotspot in the computer vision field [7, 28, 32, 36, 44, 49]." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "text", + "content": "Traditional rPPG measurement methods include signal blind decomposition [21,34,42] and color space transformation [9,59,63]. These approaches rely on heartbeat-related" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 294, + 732, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 732, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 732, + 317, + 742 + ], + "type": "text", + "content": "18589" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 156 + ], + "type": "text", + "content": "statistical information, only applicable in constrained environments. In recent years, deep learning (DL) based approaches [6, 19, 25, 28, 36, 44, 50, 69, 70] have shown their great potentials in rPPG measurement. By learning dedicated rPPG feature representation, these methods achieve promising performance in much more complicated environments [1, 36, 47, 52, 68]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 158, + 287, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 158, + 287, + 302 + ], + "spans": [ + { + "bbox": [ + 46, + 158, + 287, + 302 + ], + "type": "text", + "content": "However, deep learning methods suffer from significant performance degradation when applied in real-world scenarios. This is because most training data are captured in lab environments with limited environmental variations. With domain shifts (e.g., different illumination, camera parameters, motions, etc.), these models may struggle to generalize for the unseen testing domain. To validate this, we conduct a cross-dataset evaluation shown in Fig. 1(b). As shown, all DL-based methods do not work well in this evaluation. Furthermore, it is worth noting that DeepPhys [6] and TS-CAN [25] even perform inferior to traditional approach POS [63]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 304, + 288, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 304, + 288, + 519 + ], + "spans": [ + { + "bbox": [ + 46, + 304, + 288, + 519 + ], + "type": "text", + "content": "To improve the performance on the unseen domains, one common practice is to incorporate Domain Generalization (DG) approaches, e.g., encouraging intermediate features to be domain-invariant [12, 18, 30, 38, 57, 62]. However, as shown in Fig. 1 (b) (AD [12] and GroupDRO [38]), the improvements are still quite limited. One reason is that existing DG methods assume that domain labels can be clearly defined (e.g., by data source). Unfortunately, different background scenes, acquisition devices, or even identities could cause very different data distributions in rPPG measurement tasks, and such distribution discrepancy may exist within or cross datasets. In this case, explicitly defining domains is very difficult, and simply treating one dataset as one domain may lead to inferior performance. This is termed as agnostic domain generalization. Besides this, as physiological cues are usually much more subtle than the various noise, the model may overfit in the source domain with the limited training data." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 521, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 288, + 714 + ], + "type": "text", + "content": "In this paper, we propose the NEural SStructure modeling (NEST), a principled solution to the abovementioned problems. The main idea of NEST is to narrow the under-optimized and redundant feature space, align domain invariant features, and enrich discriminative features. Our intuition is as follows. Neural structure refers to the channel activation degree in each convolution layer, which reveals the discriminative feature combination for the specific sample. As the limited variation in a certain restricted domain, there are some spaces that the model seldomly is optimized. Out-of-distribution (OOD) samples may cause abnormal activation in these spaces, which may lead to performance degeneration. Therefore, we regularize the neural structure to encourage the model to be more well-conditioned to avoid abnormal activation caused by OOD samples. Specifically, we propose the NEural SStructure Coverage Maxi" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 545, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 215 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 215 + ], + "type": "text", + "content": "mization (NEST-CM) that encourages all neural spaces to be optimized during training, reducing the chance of abnormal activation during testing. Secondly, we propose the NEural SStructure Targeted Alignment (NEST-TA) that encourage network suppresses domain variant feature by comparing the samples with the similar physiological information. Thirdly, we propose the NEural SStructure Diversity Maximization (NEST-DM) to enrich discriminative features against unseen noise. It should be noted that our approach does not rely on domain labels, which is more applicable in the rPPG measurement task. To summarize, our contributions are listed as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 216, + 545, + 347 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 304, + 216, + 545, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 216, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 304, + 216, + 545, + 251 + ], + "type": "text", + "content": "1. We are the first to study the domain shift problem in rPPG measurement, which introduces a new challenge, agnostic domain generalization." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 251, + 545, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 251, + 545, + 299 + ], + "spans": [ + { + "bbox": [ + 304, + 251, + 545, + 299 + ], + "type": "text", + "content": "2. We propose the NEural SStructure modeling to alleviate domain shift, which is performed by narrowing the under-optimized feature space, and enhancing and enriching domain invariant features." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 300, + 545, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 300, + 545, + 347 + ], + "spans": [ + { + "bbox": [ + 304, + 300, + 545, + 347 + ], + "type": "text", + "content": "3. We establish a large-scale domain generalization (DG) benchmark for rPPG measurement, which is the first DG protocol in this task. Extensive experiments in this dataset show the superiority of our approach." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 350, + 390, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 350, + 390, + 361 + ], + "spans": [ + { + "bbox": [ + 306, + 350, + 390, + 361 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 366, + 495, + 379 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 366, + 495, + 379 + ], + "spans": [ + { + "bbox": [ + 306, + 366, + 495, + 379 + ], + "type": "text", + "content": "2.1. Remote Physiological Measurement" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 382, + 545, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 382, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 304, + 382, + 545, + 632 + ], + "type": "text", + "content": "Remote physiological measurement aims to estimate HR and HRV values by analyzing the chrominance changes in the skin. Traditional methods mainly perform color space transformations or signal decomposition to obtain the BVP signal with high SNR [9, 10, 42, 64]. These methods are designed manually for certain lighting scenes with slight motion and can not generalize to the complex environment in the wild. With powerful modeling capacity, the deep learning (DL) model has been successfully used in remote physiological estimation [6, 13, 35, 36, 51, 67, 69]. Among these methods, various model structures are tailored to extract BVP signals from face videos by aggregating spatial and temporal information [6, 37, 51, 69, 70]. Besides, many hand-designed feature maps are used to reduce the computational burden of the network by removing the irrelevant background [13, 28, 36, 45, 50]. Some work utilizes the distribution of the target domain to improve the model performance [19, 26], which is not practical without touching the unseen domain in advance. Differently, we study the model generalization on domain generalization protocol for rPPG measurement." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 638, + 435, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 638, + 435, + 651 + ], + "spans": [ + { + "bbox": [ + 306, + 638, + 435, + 651 + ], + "type": "text", + "content": "2.2. Domain generalization" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "Domain generalization (DG), i.e., out-of-distribution generalization, aims to generalize to the unseen domain by training a model on multiple or single source domains [62]. The main solution can be categorized into data manipulation, representation learning, and meta-learning. Data" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 732, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 732, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 732, + 317, + 742 + ], + "type": "text", + "content": "18590" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 70, + 382, + 339 + ], + "blocks": [ + { + "bbox": [ + 53, + 70, + 382, + 339 + ], + "lines": [ + { + "bbox": [ + 53, + 70, + 382, + 339 + ], + "spans": [ + { + "bbox": [ + 53, + 70, + 382, + 339 + ], + "type": "image", + "image_path": "b47254a0e21d43c8e68d78f1162f3a1ef02ef922759cfe210bd36533a3c81e86.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 344, + 547, + 389 + ], + "lines": [ + { + "bbox": [ + 46, + 344, + 547, + 389 + ], + "spans": [ + { + "bbox": [ + 46, + 344, + 547, + 389 + ], + "type": "text", + "content": "Figure 2. An overview of the proposed method. (a) Spatial-temporal map (STMap) is extracted from face video [36]. (b) NEST representation is calculated from the samples of the different domains. (c) The under-optimized space in NEST presentation distribution is narrowed by NEST-CM. (d) NEST-TA align the feature of multi domain at instance-level. (e) The distance of different samples are pulled away by NEST-DM." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 383, + 71, + 541, + 339 + ], + "blocks": [ + { + "bbox": [ + 383, + 71, + 541, + 339 + ], + "lines": [ + { + "bbox": [ + 383, + 71, + 541, + 339 + ], + "spans": [ + { + "bbox": [ + 383, + 71, + 541, + 339 + ], + "type": "image", + "image_path": "078712f385fb6ca430b0a2914122470b6d5cb25b4d750dcb141b7ab88e0dfcf8.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 396, + 289, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 396, + 289, + 601 + ], + "spans": [ + { + "bbox": [ + 46, + 396, + 289, + 601 + ], + "type": "text", + "content": "The generalization of the model, which can remove irrelevant style information by data generation and data augmentation [48, 71]. The main goal of representation learning is to learn domain-invariant features by alleviating covariate shift across domain [12, 18, 38, 57]. By simulating domain shift on the virtual dataset, meta-learning seeks transferable and shareable knowledge from multiple domain [23, 30]. Unlike common tasks such as classification, segmentation, and detection [15, 20, 22, 24, 40], rPPG measurement tasks do not have explicit boundaries between different domains, i.e., agnostic domain generalization. Recently, some single source domain generalization (SSDG) methods focus on the worst-case formulation without using the domain label [43, 57, 66]. Different from all the methods above, we tackle the DG problem of regression task in view of neuron structure modeling." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 613, + 104, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 613, + 104, + 625 + ], + "spans": [ + { + "bbox": [ + 47, + 613, + 104, + 625 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 634, + 163, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 634, + 163, + 645 + ], + "spans": [ + { + "bbox": [ + 47, + 634, + 163, + 645 + ], + "type": "text", + "content": "3.1. Overall Framework" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": "In this work, we propose the NEuron SStructure modeling (NSET), without using domain label, to alleviate the performance degradation caused by domain shift as illustrated in Fig. 2. Specifically, we first compressed the face video into the spatial-temporal map (STMap) as the input of" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 396, + 547, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 396, + 547, + 564 + ], + "spans": [ + { + "bbox": [ + 304, + 396, + 547, + 564 + ], + "type": "text", + "content": "CNN refer to [36] as shown in Fig. 2 (a). Sequentially, we define an Activated NEuron SStructure (NEST) representation that describes the activated degree of different channels of the network as shown in Fig. 2 (b). To alleviate the domain shift, we model a unified NEST distribution by three regularizations: (1) NEuron SStructure Coverage Maximization (NEST-CM) is used to narrow under-optimized space that is redundant and unstable as shown in Fig. 2 (c); (2) NEuron SStructure Targeted Alignment (NEST-TA) is introduced to align multi-domain feature distribution at instance-level as shown in Fig. 2 (d). (3) Neuron Structure Diversity Maximization (NEST-DM) loss to force the network to explore more features with discriminant ability as shown in Fig. 2 (e)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 574, + 464, + 587 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 574, + 464, + 587 + ], + "spans": [ + { + "bbox": [ + 306, + 574, + 464, + 587 + ], + "type": "text", + "content": "3.2. NEuron SStructure Modeling" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 594, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 715 + ], + "type": "text", + "content": "NEuron SStructure Representation. For the rPPG measurement task, the convolution kernels are adaptively activated to extract physiological features against various noise distribution [27, 56, 65]. However, the facial physiological signals are feeble and can be easily affected by various unseen noises, which make the filters activated abnormally. Therefore, the NEuron SStructure (NEST) representation is defined to explore how domain shift affects the model robustness in view of activated feature distribution. Specifically, we firstly average the feature map" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 733, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 733, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 733, + 317, + 742 + ], + "type": "text", + "content": "18591" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 265, + 151 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 265, + 151 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 265, + 151 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 265, + 151 + ], + "type": "image", + "image_path": "1ecd02f2e15691a1050d15c1eb02ba96783252b535a36e696afca9c7ebc73031.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 89, + 152, + 231, + 159 + ], + "lines": [ + { + "bbox": [ + 89, + 152, + 231, + 159 + ], + "spans": [ + { + "bbox": [ + 89, + 152, + 231, + 159 + ], + "type": "text", + "content": "(a) The expectation of different base projection" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 59, + 161, + 168, + 236 + ], + "blocks": [ + { + "bbox": [ + 59, + 161, + 168, + 236 + ], + "lines": [ + { + "bbox": [ + 59, + 161, + 168, + 236 + ], + "spans": [ + { + "bbox": [ + 59, + 161, + 168, + 236 + ], + "type": "image", + "image_path": "9a25213a89acee181e1b480a154ddb6982550d3d417498e363153b963c0baa2b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 69, + 236, + 168, + 252 + ], + "lines": [ + { + "bbox": [ + 69, + 236, + 168, + 252 + ], + "spans": [ + { + "bbox": [ + 69, + 236, + 168, + 252 + ], + "type": "text", + "content": "(b) The projection distribution of base with highest eigenvalue" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 256, + 287, + 289 + ], + "lines": [ + { + "bbox": [ + 46, + 256, + 287, + 289 + ], + "spans": [ + { + "bbox": [ + 46, + 256, + 287, + 289 + ], + "type": "text", + "content": "Figure 3. (a) The expectation of different base projection. (b) The distribution of base with highest eigenvalue. (c) The distribution of base with lowest eigenvalue." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 179, + 161, + 272, + 235 + ], + "blocks": [ + { + "bbox": [ + 179, + 161, + 272, + 235 + ], + "lines": [ + { + "bbox": [ + 179, + 161, + 272, + 235 + ], + "spans": [ + { + "bbox": [ + 179, + 161, + 272, + 235 + ], + "type": "image", + "image_path": "87ec75dcb2c77d0928dffdda3336a82164e716a434f56f0e5eb8c799a9ca4953.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 176, + 236, + 275, + 252 + ], + "lines": [ + { + "bbox": [ + 176, + 236, + 275, + 252 + ], + "spans": [ + { + "bbox": [ + 176, + 236, + 275, + 252 + ], + "type": "text", + "content": "(c) The projection distribution of base with lowest eigenvalue" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 300, + 286, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 300, + 286, + 409 + ], + "spans": [ + { + "bbox": [ + 46, + 300, + 286, + 409 + ], + "type": "inline_equation", + "content": "m_{j}\\in R^{W\\times H\\times C_{j}}" + }, + { + "bbox": [ + 46, + 300, + 286, + 409 + ], + "type": "text", + "content": " over the spatial dimensions in the end of " + }, + { + "bbox": [ + 46, + 300, + 286, + 409 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 300, + 286, + 409 + ], + "type": "text", + "content": "-th layer to get the preliminary statistical presentation " + }, + { + "bbox": [ + 46, + 300, + 286, + 409 + ], + "type": "inline_equation", + "content": "\\bar{o}_j\\in R^{C_j}" + }, + { + "bbox": [ + 46, + 300, + 286, + 409 + ], + "type": "text", + "content": " of the neuron structure, where " + }, + { + "bbox": [ + 46, + 300, + 286, + 409 + ], + "type": "inline_equation", + "content": "C_j" + }, + { + "bbox": [ + 46, + 300, + 286, + 409 + ], + "type": "text", + "content": " is the channel number of the " + }, + { + "bbox": [ + 46, + 300, + 286, + 409 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 300, + 286, + 409 + ], + "type": "text", + "content": "-th layer. Following that, the preliminary presentation " + }, + { + "bbox": [ + 46, + 300, + 286, + 409 + ], + "type": "inline_equation", + "content": "\\bar{o}_j" + }, + { + "bbox": [ + 46, + 300, + 286, + 409 + ], + "type": "text", + "content": " is fed into the max-min normalized to get the NEST representation " + }, + { + "bbox": [ + 46, + 300, + 286, + 409 + ], + "type": "inline_equation", + "content": "o_{j}\\in R_{j}^{C}" + }, + { + "bbox": [ + 46, + 300, + 286, + 409 + ], + "type": "text", + "content": " of the " + }, + { + "bbox": [ + 46, + 300, + 286, + 409 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 300, + 286, + 409 + ], + "type": "text", + "content": "-th layer. The full NEST representation of " + }, + { + "bbox": [ + 46, + 300, + 286, + 409 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 300, + 286, + 409 + ], + "type": "text", + "content": "-th sample " + }, + { + "bbox": [ + 46, + 300, + 286, + 409 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 46, + 300, + 286, + 409 + ], + "type": "text", + "content": " is defined as " + }, + { + "bbox": [ + 46, + 300, + 286, + 409 + ], + "type": "inline_equation", + "content": "O_{i} = \\{o_{i,1},o_{i,2},\\dots,o_{i,U}\\} \\in R^{(C_{1} + C_{2} + \\ldots +C_{U})}" + }, + { + "bbox": [ + 46, + 300, + 286, + 409 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 300, + 286, + 409 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 46, + 300, + 286, + 409 + ], + "type": "text", + "content": " is the number of selected layer." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 411, + 286, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 411, + 286, + 494 + ], + "spans": [ + { + "bbox": [ + 46, + 411, + 286, + 494 + ], + "type": "text", + "content": "NEuron SStructure Coverage Maximization. Existing works show that the OOD samples may activate abnormal neuron structures (feature combination) that seldom be trained with the source data so that the model's performance degrades drastically [31, 39, 57]. To investigate this phenomenon in rPPG measurement task, we analyze the optimized degree of different feature combinations." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 496, + 286, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 496, + 286, + 687 + ], + "spans": [ + { + "bbox": [ + 46, + 496, + 286, + 687 + ], + "type": "text", + "content": "Specifically, we decompose the NEST representation of the last layer into orthogonal bases with Singular Value Decomposition (SVD). Here, small eigenvalues indicate that training samples cover a small space in these corresponding bases. Thus, unseen target samples are more likely to lay out of the well-optimized space, causing a larger domain shift. To validate this, we plot the means of source and target samples on these bases in Fig. 3(a). As shown, the mean differences are larger for bases of smaller eigenvalues. To make it clearer, We further visualize the data distribution on the first and last bases in Fig. 3 (b) and (c). As shown, the distribution difference in the last base is much more significant than that of the first one. All these suggest that for the NEST representation, components of small eigenvalues are less optimized during training, and domain shifts are more likely to appear there." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "type": "text", + "content": "In sight of this, we propose the NEuron SStructure Coverage Maximization (NEST-CM) regularization that improves" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 307, + 70, + 543, + 121 + ], + "blocks": [ + { + "bbox": [ + 307, + 70, + 543, + 121 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 543, + 121 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 543, + 121 + ], + "type": "image", + "image_path": "e21bf16425701cf4d040c06fc08e94c142c0593e717508a8e7841ee3e15836b0.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 325, + 122, + 525, + 133 + ], + "lines": [ + { + "bbox": [ + 325, + 122, + 525, + 133 + ], + "spans": [ + { + "bbox": [ + 325, + 122, + 525, + 133 + ], + "type": "text", + "content": "Figure 4. The detail of SVD for the " + }, + { + "bbox": [ + 325, + 122, + 525, + 133 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 325, + 122, + 525, + 133 + ], + "type": "text", + "content": "-th layer in CNN." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 135, + 545, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 135, + 545, + 266 + ], + "spans": [ + { + "bbox": [ + 305, + 135, + 545, + 266 + ], + "type": "text", + "content": "the eigenvalues of the last few bases. This could narrow the under-optimized space as shown in Fig. 2 (c). For a training batch with " + }, + { + "bbox": [ + 305, + 135, + 545, + 266 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 305, + 135, + 545, + 266 + ], + "type": "text", + "content": " samples, NEST representation set of " + }, + { + "bbox": [ + 305, + 135, + 545, + 266 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 305, + 135, + 545, + 266 + ], + "type": "text", + "content": "-th layer " + }, + { + "bbox": [ + 305, + 135, + 545, + 266 + ], + "type": "inline_equation", + "content": "B_{j} = \\{o_{1,j}; o_{2,j}; \\ldots; o_{N,j}\\} \\in R^{N * C_{j}}" + }, + { + "bbox": [ + 305, + 135, + 545, + 266 + ], + "type": "text", + "content": " is decomposed to " + }, + { + "bbox": [ + 305, + 135, + 545, + 266 + ], + "type": "inline_equation", + "content": "S_{j} \\cdot V_{j} \\cdot D_{j}" + }, + { + "bbox": [ + 305, + 135, + 545, + 266 + ], + "type": "text", + "content": " by SVD method as shown in Fig. 4, i.e., " + }, + { + "bbox": [ + 305, + 135, + 545, + 266 + ], + "type": "inline_equation", + "content": "B_{j} = S_{j} \\cdot V_{j} \\cdot D_{j}" + }, + { + "bbox": [ + 305, + 135, + 545, + 266 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 305, + 135, + 545, + 266 + ], + "type": "inline_equation", + "content": "V_{j} \\in R^{N * C_{j}}" + }, + { + "bbox": [ + 305, + 135, + 545, + 266 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 135, + 545, + 266 + ], + "type": "inline_equation", + "content": "N > C_{j}" + }, + { + "bbox": [ + 305, + 135, + 545, + 266 + ], + "type": "text", + "content": ". The bases with small eigenvalue values are rare and under-optimized during the training stage. To reduce the occurrence probability of under-optimized NEST caused by domain shift, we stimulate under-optimized bases by maximizing the corresponding eigenvalues:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 350, + 277, + 545, + 311 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 277, + 545, + 311 + ], + "spans": [ + { + "bbox": [ + 350, + 277, + 545, + 311 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {C M} = - \\mathbb {E} _ {j \\in U} \\frac {\\sum_ {i = 1} ^ {C ^ {\\prime}} \\lambda_ {i , j} ^ {\\prime}}{\\sum_ {i = 1} ^ {C} \\lambda_ {i , j}}, \\lambda_ {i, j} ^ {\\prime} < t, \\tag {1}", + "image_path": "02a4aece8e271e553bc90578729a98e6cfd10537fd6d697baa978e1173b1edc8.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 319, + 545, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 319, + 545, + 418 + ], + "spans": [ + { + "bbox": [ + 305, + 319, + 545, + 418 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 319, + 545, + 418 + ], + "type": "inline_equation", + "content": "\\lambda_{i,j}" + }, + { + "bbox": [ + 305, + 319, + 545, + 418 + ], + "type": "text", + "content": " represent the " + }, + { + "bbox": [ + 305, + 319, + 545, + 418 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 305, + 319, + 545, + 418 + ], + "type": "text", + "content": "-th normalized eigenvalue in " + }, + { + "bbox": [ + 305, + 319, + 545, + 418 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 305, + 319, + 545, + 418 + ], + "type": "text", + "content": "-th layer' diagonal matrix " + }, + { + "bbox": [ + 305, + 319, + 545, + 418 + ], + "type": "inline_equation", + "content": "V_{j}" + }, + { + "bbox": [ + 305, + 319, + 545, + 418 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 305, + 319, + 545, + 418 + ], + "type": "inline_equation", + "content": "\\lambda_{i,j}^{\\prime}" + }, + { + "bbox": [ + 305, + 319, + 545, + 418 + ], + "type": "text", + "content": " is the eigenvalue less than the threshold " + }, + { + "bbox": [ + 305, + 319, + 545, + 418 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 305, + 319, + 545, + 418 + ], + "type": "text", + "content": ". Considering different channel numbers in each layer, we design an adaptive threshold " + }, + { + "bbox": [ + 305, + 319, + 545, + 418 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 305, + 319, + 545, + 418 + ], + "type": "text", + "content": " to select the relatively under-optimized bases, i.e., " + }, + { + "bbox": [ + 305, + 319, + 545, + 418 + ], + "type": "inline_equation", + "content": "t = \\frac{\\rho}{C_j}" + }, + { + "bbox": [ + 305, + 319, + 545, + 418 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 305, + 319, + 545, + 418 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 305, + 319, + 545, + 418 + ], + "type": "text", + "content": " is a trade-off parameter. Besides, " + }, + { + "bbox": [ + 305, + 319, + 545, + 418 + ], + "type": "inline_equation", + "content": "C_j^\\prime" + }, + { + "bbox": [ + 305, + 319, + 545, + 418 + ], + "type": "text", + "content": " is the number of eigenvalues of " + }, + { + "bbox": [ + 305, + 319, + 545, + 418 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 305, + 319, + 545, + 418 + ], + "type": "text", + "content": "-th layer under the threshold, and " + }, + { + "bbox": [ + 305, + 319, + 545, + 418 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 305, + 319, + 545, + 418 + ], + "type": "text", + "content": " is the number of layers." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 418, + 545, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 418, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 305, + 418, + 545, + 536 + ], + "type": "text", + "content": "NEuron SStructure Targeted Alignment. Besides the NEST-CM introduced above, we further improve the generalization ability by aligning the NEST representation of different domains. Domain invariant learning has demonstrated its effectiveness in DG problems [12, 18, 30, 38, 57, 62]. Some recent research [11, 16, 29, 60] further shows that aligning domain features within the same labels could lead to better performance. However, labels in rPPG are continuous, and it is hard to find samples with identical labels in a training batch." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 538, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 538, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 538, + 545, + 714 + ], + "type": "text", + "content": "To cope with this, we propose a novel regularization term called NEuron SStructure Alignment (NEST-TA). NEST-TA builds a synthesized prototype " + }, + { + "bbox": [ + 305, + 538, + 545, + 714 + ], + "type": "inline_equation", + "content": "O_{i}^{p}" + }, + { + "bbox": [ + 305, + 538, + 545, + 714 + ], + "type": "text", + "content": " with the same physiological information as the given targeted sample " + }, + { + "bbox": [ + 305, + 538, + 545, + 714 + ], + "type": "inline_equation", + "content": "O_{i}" + }, + { + "bbox": [ + 305, + 538, + 545, + 714 + ], + "type": "text", + "content": ". Then, the feature within the same labels can be aligned across domains like [11, 16, 29, 60]. Specifically, we select the top " + }, + { + "bbox": [ + 305, + 538, + 545, + 714 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 305, + 538, + 545, + 714 + ], + "type": "text", + "content": " samples closest to the targeted ground truth " + }, + { + "bbox": [ + 305, + 538, + 545, + 714 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 305, + 538, + 545, + 714 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 305, + 538, + 545, + 714 + ], + "type": "inline_equation", + "content": "[O_{i,1}', \\dots, O_{i,k}', \\dots, O_{i,K}']" + }, + { + "bbox": [ + 305, + 538, + 545, + 714 + ], + "type": "text", + "content": ". The NEST center of the selected samples " + }, + { + "bbox": [ + 305, + 538, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\overline{O}_{i}'" + }, + { + "bbox": [ + 305, + 538, + 545, + 714 + ], + "type": "text", + "content": " is calculated by averaging the NEST presentation of selected samples. Then we calculate the direction vectors between the NEST presentation center of the selected samples " + }, + { + "bbox": [ + 305, + 538, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\overline{O}_{i}'" + }, + { + "bbox": [ + 305, + 538, + 545, + 714 + ], + "type": "text", + "content": " and the NEST presentation of each sample " + }, + { + "bbox": [ + 305, + 538, + 545, + 714 + ], + "type": "inline_equation", + "content": "O_{i,k}'" + }, + { + "bbox": [ + 305, + 538, + 545, + 714 + ], + "type": "text", + "content": ". Following that, the NEST center of the selected samples " + }, + { + "bbox": [ + 305, + 538, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\overline{O}_{i}'" + }, + { + "bbox": [ + 305, + 538, + 545, + 714 + ], + "type": "text", + "content": " is calibrated to the synthesized prototype " + }, + { + "bbox": [ + 305, + 538, + 545, + 714 + ], + "type": "inline_equation", + "content": "O_{i}^{p}" + }, + { + "bbox": [ + 305, + 538, + 545, + 714 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 732, + 317, + 741 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 732, + 317, + 741 + ], + "spans": [ + { + "bbox": [ + 294, + 732, + 317, + 741 + ], + "type": "text", + "content": "18592" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "type": "text", + "content": "which is performed by using the Gaussian distribution function to fuse these direction vectors:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 70, + 99, + 287, + 135 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 99, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 70, + 99, + 287, + 135 + ], + "type": "interline_equation", + "content": "O _ {i} ^ {p} = \\overline {{O ^ {\\prime}}} _ {i} + \\sum_ {k = 1} ^ {K} W _ {i, k} * \\frac {y _ {i} - \\bar {y} _ {i} ^ {\\prime}}{y _ {i , k} ^ {\\prime} - \\bar {y} _ {i} ^ {\\prime}} \\left(O _ {i, k} ^ {\\prime} - \\overline {{O ^ {\\prime}}} _ {i}\\right), \\tag {2}", + "image_path": "6a40056a84791d2689d394427d56a6ea48a6a47c58cc249c89ccb0448b58204d.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 100, + 135, + 230, + 163 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 135, + 230, + 163 + ], + "spans": [ + { + "bbox": [ + 100, + 135, + 230, + 163 + ], + "type": "interline_equation", + "content": "W _ {i, k} = \\frac {G (y _ {i , k} - \\overline {{y}} _ {i} ^ {\\prime} , \\sigma)}{\\sum_ {k = 1} ^ {K} G (y _ {i , k} - \\overline {{y}} _ {i} ^ {\\prime} , \\sigma)},", + "image_path": "1aed7657e19207af0cc5508ce1a58d198573768f61dffdf7fb80cf3a1372f40a.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 170, + 288, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 170, + 288, + 262 + ], + "spans": [ + { + "bbox": [ + 47, + 170, + 288, + 262 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 170, + 288, + 262 + ], + "type": "inline_equation", + "content": "G(-, \\sigma)" + }, + { + "bbox": [ + 47, + 170, + 288, + 262 + ], + "type": "text", + "content": " is the Gaussian distribution with the mean of zero and the standard deviation of " + }, + { + "bbox": [ + 47, + 170, + 288, + 262 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 47, + 170, + 288, + 262 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 47, + 170, + 288, + 262 + ], + "type": "inline_equation", + "content": "G(y_{i,k} - \\overline{y}_i', \\sigma) = \\frac{1}{\\sqrt{2\\pi}} e^{-\\frac{(y_{i,k} - \\overline{y}_i')^2}{2\\sigma^2}}" + }, + { + "bbox": [ + 47, + 170, + 288, + 262 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 170, + 288, + 262 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 47, + 170, + 288, + 262 + ], + "type": "text", + "content": " is a hyper-parameter. To align the feature distribution across different domains, NEST-TA uses the cosine similarity " + }, + { + "bbox": [ + 47, + 170, + 288, + 262 + ], + "type": "inline_equation", + "content": "Cos(-, -)" + }, + { + "bbox": [ + 47, + 170, + 288, + 262 + ], + "type": "text", + "content": " to reduce the variant features between targeted samples " + }, + { + "bbox": [ + 47, + 170, + 288, + 262 + ], + "type": "inline_equation", + "content": "O_i" + }, + { + "bbox": [ + 47, + 170, + 288, + 262 + ], + "type": "text", + "content": " and prototype sample " + }, + { + "bbox": [ + 47, + 170, + 288, + 262 + ], + "type": "inline_equation", + "content": "O_i^p" + }, + { + "bbox": [ + 47, + 170, + 288, + 262 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 110, + 268, + 287, + 282 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 268, + 287, + 282 + ], + "spans": [ + { + "bbox": [ + 110, + 268, + 287, + 282 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {T A} = \\mathbb {E} _ {i \\in N} \\operatorname {C o s} \\left(O _ {i}, O _ {i} ^ {p}\\right). \\tag {3}", + "image_path": "298864e6b7d30cc2dd634c23a54dc893757677638773d5ba4a5f0a43e18d9165.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 291, + 287, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 291, + 287, + 422 + ], + "spans": [ + { + "bbox": [ + 47, + 291, + 287, + 422 + ], + "type": "text", + "content": "NEuron SStructure Diversity Maximization. Existing works elaborate that it is effective to explore more features with discriminant information to improve the generalization [4,30,41,46]. Motivated by this, NEuron SStructure Diversity Maximization (NEST-DM) is introduced to increase the distinction between the NEST presentation of different samples as shown in Fig. 2 (e). By doing this, the network is forced to explore more features with physiological information in a contrastive learning way, which can further improve model discrimination. NEST-DM " + }, + { + "bbox": [ + 47, + 291, + 287, + 422 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{DM}" + }, + { + "bbox": [ + 47, + 291, + 287, + 422 + ], + "type": "text", + "content": " can be represented as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 63, + 430, + 287, + 463 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 430, + 287, + 463 + ], + "spans": [ + { + "bbox": [ + 63, + 430, + 287, + 463 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {D M} = - \\mathbb {E} _ {i \\in N} \\log \\left(\\frac {e ^ {C o s \\left(O _ {i} , \\widehat {O} _ {i}\\right) / \\tau}}{\\sum_ {j = 1 (i \\neq j)} ^ {N} e ^ {C o s \\left(O _ {i} , \\widehat {O} _ {j}\\right) / \\tau}}\\right), \\tag {4}", + "image_path": "171930998cd47f45f8da2b280969c508c52f185a2690925944bf2e8b26f0d396.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 472, + 287, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 472, + 287, + 519 + ], + "spans": [ + { + "bbox": [ + 47, + 472, + 287, + 519 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 472, + 287, + 519 + ], + "type": "inline_equation", + "content": "\\widehat{O}_i" + }, + { + "bbox": [ + 47, + 472, + 287, + 519 + ], + "type": "text", + "content": " is the NEST representation of augmented sample from " + }, + { + "bbox": [ + 47, + 472, + 287, + 519 + ], + "type": "inline_equation", + "content": "O_i" + }, + { + "bbox": [ + 47, + 472, + 287, + 519 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 472, + 287, + 519 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 47, + 472, + 287, + 519 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 472, + 287, + 519 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 47, + 472, + 287, + 519 + ], + "type": "text", + "content": " are the batch size and temperature parameter. The augmentation strategies for STMap will be elaborated in section 4.2." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 521, + 287, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 521, + 287, + 663 + ], + "spans": [ + { + "bbox": [ + 47, + 521, + 287, + 663 + ], + "type": "text", + "content": "Model Training. For the rPPG measurement task, we chose two popular losses in this field. Specifically, L1 loss " + }, + { + "bbox": [ + 47, + 521, + 287, + 663 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{L1}" + }, + { + "bbox": [ + 47, + 521, + 287, + 663 + ], + "type": "text", + "content": " and negative Pearson's correlation coefficient " + }, + { + "bbox": [ + 47, + 521, + 287, + 663 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_p" + }, + { + "bbox": [ + 47, + 521, + 287, + 663 + ], + "type": "text", + "content": " are used to constrain the predicted heart rate value and the predicted BVP signal respectively [14, 28, 37, 69]. These two losses are also used for augmented data. With the violent confrontation at the beginning of training, " + }, + { + "bbox": [ + 47, + 521, + 287, + 663 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{L1}" + }, + { + "bbox": [ + 47, + 521, + 287, + 663 + ], + "type": "text", + "content": " make network tends to predict the mean of the heart rate value distribution instead of learning the physiological feature like " + }, + { + "bbox": [ + 47, + 521, + 287, + 663 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_p" + }, + { + "bbox": [ + 47, + 521, + 287, + 663 + ], + "type": "text", + "content": ". Besides, " + }, + { + "bbox": [ + 47, + 521, + 287, + 663 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{NSM}" + }, + { + "bbox": [ + 47, + 521, + 287, + 663 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 521, + 287, + 663 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{NEST-S}" + }, + { + "bbox": [ + 47, + 521, + 287, + 663 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 521, + 287, + 663 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{NEST-DM}" + }, + { + "bbox": [ + 47, + 521, + 287, + 663 + ], + "type": "text", + "content": " also have meaningless effects on under-trained networks at the beginning." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 664, + 287, + 735 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 664, + 287, + 735 + ], + "spans": [ + { + "bbox": [ + 47, + 664, + 287, + 735 + ], + "type": "text", + "content": "To solve the above-illustrated issues, we introduce the adaptation factor " + }, + { + "bbox": [ + 47, + 664, + 287, + 735 + ], + "type": "inline_equation", + "content": "\\gamma = \\frac{2}{1 + \\exp(-10r)}" + }, + { + "bbox": [ + 47, + 664, + 287, + 735 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 664, + 287, + 735 + ], + "type": "inline_equation", + "content": "r = \\frac{\\text{inters}_{\\text{current}}}{\\text{inters}_{\\text{total}}}" + }, + { + "bbox": [ + 47, + 664, + 287, + 735 + ], + "type": "text", + "content": " to suppress meaningless constraints at the early training stage, which can accelerate model convergence and improve training stability [12, 15]. The final loss " + }, + { + "bbox": [ + 47, + 664, + 287, + 735 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{overall}}" + }, + { + "bbox": [ + 47, + 664, + 287, + 735 + ], + "type": "text", + "content": " can be formulated as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 344, + 91, + 545, + 118 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 91, + 545, + 118 + ], + "spans": [ + { + "bbox": [ + 344, + 91, + 545, + 118 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\text {o v e r a l l}} = k _ {1} \\mathcal {L} _ {p} + \\gamma \\left(k _ {2} \\mathcal {L} _ {L 1} + k _ {3} \\mathcal {L} _ {C M} \\right. \\tag {5} \\\\ + k _ {4} \\mathcal {L} _ {T A} + k _ {5} \\mathcal {L} _ {D M}), \\\\ \\end{array}", + "image_path": "7023a50f3792561cc5e0fa3ce596a8ccdf692f9d8bc87ea79715c584fcfa843f.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 127, + 481, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 127, + 481, + 140 + ], + "spans": [ + { + "bbox": [ + 306, + 127, + 481, + 140 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 127, + 481, + 140 + ], + "type": "inline_equation", + "content": "k_{1} - k_{5}" + }, + { + "bbox": [ + 306, + 127, + 481, + 140 + ], + "type": "text", + "content": " are five trade-off parameters." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 150, + 386, + 163 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 150, + 386, + 163 + ], + "spans": [ + { + "bbox": [ + 306, + 150, + 386, + 163 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 170, + 368, + 181 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 170, + 368, + 181 + ], + "spans": [ + { + "bbox": [ + 306, + 170, + 368, + 181 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 188, + 545, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 188, + 545, + 224 + ], + "spans": [ + { + "bbox": [ + 305, + 188, + 545, + 224 + ], + "type": "text", + "content": "We elaborately select five datasets (including different motion, camera, and lighting conditions) to curate and benchmark a large-scale DG protocol:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 224, + 545, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 224, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 305, + 224, + 545, + 319 + ], + "type": "text", + "content": "VIPL-HR [36] have nine scenarios, three RGB cameras, different illumination conditions, and different levels of movement. It is worth mentioning that the BVP signal and video of the dataset do not match in the time dimension, so " + }, + { + "bbox": [ + 305, + 224, + 545, + 319 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_p" + }, + { + "bbox": [ + 305, + 224, + 545, + 319 + ], + "type": "text", + "content": " is not used for this dataset. We used the output of HR head as the evaluation result. Besides, we normalize STMap to 30 fps by cubic spline interpolation to solve the problem of the unstable frame rate of video [28, 50, 70]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 319, + 545, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 319, + 545, + 402 + ], + "spans": [ + { + "bbox": [ + 305, + 319, + 545, + 402 + ], + "type": "text", + "content": "V4V [47] is designed to collect data with the drastic changes of physiological indicators by simulating ten tasks such as a funny joke, 911 emergency call, and odor experience. It is worth mentioning that There are only heart rate labels in the dataset and no BVP signal labels, so " + }, + { + "bbox": [ + 305, + 319, + 545, + 402 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_p" + }, + { + "bbox": [ + 305, + 319, + 545, + 402 + ], + "type": "text", + "content": " is not used for this dataset. We used the output of HR head as the evaluation result." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 403, + 545, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 403, + 545, + 485 + ], + "spans": [ + { + "bbox": [ + 305, + 403, + 545, + 485 + ], + "type": "text", + "content": "BUAA [68] is proposed to evaluate the performance of the algorithm against various illumination. We only use data with illumination greater than or equal to 10 lux because underexposed images require special algorithms that are not considered in this article. We used the output of BVP head as the evaluation for BUAA, PURE, and UBFC-rPPG datasets." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 487, + 545, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 487, + 545, + 546 + ], + "spans": [ + { + "bbox": [ + 305, + 487, + 545, + 546 + ], + "type": "text", + "content": "PURE [52] contains 60 RGB videos from 10 subjects with six different activities, specifically, sitting still, talking, and four rotating and moving head variations. The BVP signals are down-sampled from 60 to 30 fps with cubic spline interpolation to align with the videos." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 546, + 545, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 546, + 545, + 582 + ], + "spans": [ + { + "bbox": [ + 305, + 546, + 545, + 582 + ], + "type": "text", + "content": "UBFC-rPPG [1] containing 42 face videos with sunlight and indoor illumination, the ground-truth BVP signals, and HR values were collected by CMS50E." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 586, + 438, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 586, + 438, + 597 + ], + "spans": [ + { + "bbox": [ + 306, + 586, + 438, + 597 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 600, + 545, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 600, + 545, + 696 + ], + "spans": [ + { + "bbox": [ + 305, + 600, + 545, + 696 + ], + "type": "text", + "content": "Training Details. Our proposed method is performed under the Pytorch framework. For the generation of STMap from Video, we use the FAN1 to detect the 2D landmarks of faces [2], and other steps completely follow [36]. STMap in each dataset is sampled with a time window of 256 with step 10. Following [55, 61], we also design spatial and temporal augmentation for STMap, i.e., shuffle each row of STMap and slide the time window. In addition, color jitter and blur" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 317, + 702, + 525, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 702, + 525, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 702, + 525, + 712 + ], + "type": "text", + "content": "" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 293, + 732, + 300, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 732, + 300, + 742 + ], + "spans": [ + { + "bbox": [ + 293, + 732, + 300, + 742 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 294, + 758, + 317, + 768 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 758, + 317, + 768 + ], + "spans": [ + { + "bbox": [ + 294, + 758, + 317, + 768 + ], + "type": "text", + "content": "18593" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 70, + 548, + 323 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 548, + 323 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 548, + 323 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 548, + 323 + ], + "type": "table", + "html": "
MethodUBFCPUREBUAAVIPLV4V
MAE↓RMSE↓r↑MAE↓RMSE↓r↑MAE↓RMSE↓r↑MAE↓RMSE↓r↑MAE↓RMSE↓r↑MAE↓RMSE↓r↑MAE↓RMSE↓r↑MAE↓RMSE↓r↑
GREEN [59]8.029.180.3610.3214.270.525.827.990.5612.1818.230.2515.6421.430.06
CHROM [9]7.238.920.519.7912.760.376.098.290.5111.4416.970.2814.9219.220.08
POS [63]7.358.040.499.8213.440.345.047.120.6314.5921.260.1917.6523.220.04
DeepPhys [6]7.828.420.549.3412.560.554.786.740.6912.5619.130.1414.5219.110.14
TS-CAN [25]7.638.250.559.1212.380.574.846.890.6812.3418.940.1614.7719.960.12
Rhythmnet* [36]5.797.910.787.3910.490.773.385.170.848.9712.160.4910.1614.570.34
Dual-GAN* [28]5.557.620.797.2410.270.783.415.230.848.8811.690.5010.0414.440.35
BVPNet* [7]5.437.710.807.2310.250.783.695.480.818.4511.640.5110.0114.350.36
AD*+ [12]5.928.080.767.4210.610.733.495.490.828.4111.710.5310.4714.640.32
GroupDRO*+ [38]5.737.970.787.6910.830.783.415.210.838.3511.670.549.9414.290.36
Coral*+ [53]5.898.040.767.5910.870.723.645.740.808.6811.910.5310.3214.420.32
VREx*+ [18]5.597.680.817.2410.140.783.275.010.868.3711.620.549.8214.160.37
NCDG*+ [57]5.317.560.827.3210.350.773.125.160.858.4711.810.5210.1414.460.34
Baseline* [36]5.797.910.787.3910.490.773.385.170.848.9712.160.4910.1614.570.34
NEST*+ w/o LCM4.927.190.836.919.940.782.794.620.898.2111.560.559.5813.580.38
NEST*+ w/o LTA4.977.280.836.819.770.813.034.940.868.1911.470.559.5213.520.38
NEST*+ w/o LDM4.897.110.846.879.810.792.934.770.888.1411.430.569.4913.490.40
NEST*+4.676.790.866.719.590.812.884.690.897.8611.150.589.2713.790.41
", + "image_path": "ea520a6398e833a1c484e35a3c0ddb9ffee1ed75d55f2fb30ec300c7d05799a5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 326, + 547, + 370 + ], + "lines": [ + { + "bbox": [ + 46, + 326, + 547, + 370 + ], + "spans": [ + { + "bbox": [ + 46, + 326, + 547, + 370 + ], + "type": "text", + "content": "Table 1. HR estimation results on MSDG protocol. It is worth noting that we have removed the GRU mechanism in Rhythmnet because it is an unfair comparison with using GRU to aggregate the temporal information of multiple clips instead of only one clip; * means that these methods use the STMap as the input of CNN and use the STMap augmentation strategy as mentioned in section 4.2; + means that these methods are based on baseline (Rhythmnet without GRU)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 377, + 289, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 377, + 289, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 377, + 289, + 544 + ], + "type": "text", + "content": "augmentation are used for STMap [5]. Then the STMap " + }, + { + "bbox": [ + 46, + 377, + 289, + 544 + ], + "type": "inline_equation", + "content": "x \\in R^{25 \\times 256 \\times 3}" + }, + { + "bbox": [ + 46, + 377, + 289, + 544 + ], + "type": "text", + "content": " is resized to " + }, + { + "bbox": [ + 46, + 377, + 289, + 544 + ], + "type": "inline_equation", + "content": "x' \\in R^{64 \\times 256 \\times 3}" + }, + { + "bbox": [ + 46, + 377, + 289, + 544 + ], + "type": "text", + "content": " as the input of the network. We use the ResNet-18 with removing the max pooling layer as the feature extractor, and the final full connection layer is used as the HR estimation head. At the same time, we designed a BVP signal regression head containing four blocks, each block involves a Conv., transpose, and two Conv. layers with batch normalization and Relu function. Hyper-parameter " + }, + { + "bbox": [ + 46, + 377, + 289, + 544 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 46, + 377, + 289, + 544 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 377, + 289, + 544 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 46, + 377, + 289, + 544 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 377, + 289, + 544 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 377, + 289, + 544 + ], + "type": "text", + "content": " are set to 5, 0.1, and 0.2 according to the experimental result, and the trade-off parameter " + }, + { + "bbox": [ + 46, + 377, + 289, + 544 + ], + "type": "inline_equation", + "content": "k_{1} - k_{5}" + }, + { + "bbox": [ + 46, + 377, + 289, + 544 + ], + "type": "text", + "content": " are set to 1, 0.1, 0.001, 0.1 and 0.01 based on the scale of losses. Adam optimizer with a learning rate of 0.001 is used for training. The batch size and iterations are set to 1024 and 20000." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 544, + 288, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 544, + 288, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 544, + 288, + 628 + ], + "type": "text", + "content": "Performance Metrics. Following existing methods [6, 28, 37, 50, 70], standard deviation (SD), mean absolute error (MAE), root mean square error (RMSE), and Pearson's correlation coefficient (r) are used to evaluate the HR estimation. MAE, RMSE, and r are used to evaluate the HRV measurement, including low frequency (LF), high frequency (HF), and LF/HF." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 635, + 244, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 244, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 244, + 647 + ], + "type": "text", + "content": "4.3. Multi-Source Domain Generalization" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 651, + 145, + 662 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 651, + 145, + 662 + ], + "spans": [ + { + "bbox": [ + 47, + 651, + 145, + 662 + ], + "type": "text", + "content": "4.3.1 HR Estimation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 665, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 713 + ], + "type": "text", + "content": "For remote HR estimation, all five datasets (i.e., UBFC [1], PURE [52], BUAA [68], VIPL [36], and V4V [47]) are used to evaluate the model generalization. Specifically, four datasets are used for training, and then the other dataset is" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 377, + 545, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 377, + 545, + 484 + ], + "spans": [ + { + "bbox": [ + 304, + 377, + 545, + 484 + ], + "type": "text", + "content": "used to test. Firstly, we perform the iPhys " + }, + { + "bbox": [ + 304, + 377, + 545, + 484 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 304, + 377, + 545, + 484 + ], + "type": "text", + "content": " to evaluate three traditional method (i.e., GREEN [59], CHROM [9], and POS [63]) as shown in Tab. 1. Then, we reproduce five popular DL-based methods (i.e., DeepPhys [6], TSCAN [25], Rhythmnet [36], and BVPNet [7], and DualGAN [28]) for comparison on the same setting. Subsequently, we directly applied the four DG methods (i.e., AD [12], GroupDRO [38], Coral [53], VReX [18], and NCDG [57]) to Rhythmnet with using DeepDG " + }, + { + "bbox": [ + 304, + 377, + 545, + 484 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 304, + 377, + 545, + 484 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 485, + 545, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 485, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 485, + 545, + 581 + ], + "type": "text", + "content": "The performance of traditional methods. As shown in Tab. 1, traditional algorithms show relatively acceptable results on some simple scenes, such as BUAA and UBFC, but very poor results on other datasets. It is difficult for these traditional algorithms to extract the physiological information on the V4V dataset, which shows that the traditional methods cannot deal with the noise caused by the complex environment and non-rigid motions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 580, + 545, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 580, + 545, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 580, + 545, + 676 + ], + "type": "text", + "content": "The performance of DL-based methods. For DL-based methods, DeepPhys and TS-CAN even perform worse than traditional methods. Rhythmnet, BVPNet, and Dual-GAN perform much better than both other DL-based methods and traditional methods, which suggests that STMap is the more robust input of CNN than the aligned video. Compared with Rhythmnet, BVPNet and Dual-GAN have not significant improvement on the MSDG protocol. Because BVP-" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 315, + 683, + 525, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 683, + 525, + 693 + ], + "spans": [ + { + "bbox": [ + 315, + 683, + 525, + 693 + ], + "type": "text", + "content": "2https://github.com/danmcduff/iphys-toolbox" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 308, + 694, + 544, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 694, + 544, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 694, + 544, + 712 + ], + "type": "text", + "content": "3https://github.com/jindongwang/transferlearning/tree/master/code/DeepDG" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 294, + 733, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 733, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 733, + 318, + 742 + ], + "type": "text", + "content": "18594" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 53, + 69, + 538, + 285 + ], + "blocks": [ + { + "bbox": [ + 53, + 69, + 538, + 285 + ], + "lines": [ + { + "bbox": [ + 53, + 69, + 538, + 285 + ], + "spans": [ + { + "bbox": [ + 53, + 69, + 538, + 285 + ], + "type": "table", + "html": "
TargetMethodLF-(u.n)HF-(u.n)LF/HFHR-(bpm)
MAE↓RMSE↓r↑MAE↓RMSE↓r↑MAE↓RMSE↓r↑MAE↓RMSE↓r↑
UBFCGREEN [59]0.23550.28410.09240.23550.28410.09240.66950.95120.04678.01849.17760.3634
CHROM [9]0.22210.28170.06980.22210.28170.06980.67081.05420.10547.22918.92240.5123
POS [63]0.23640.28610.13590.23640.28610.13590.65150.95350.13457.35398.04020.4923
Baseline0.06210.08130.18730.06210.08130.18730.19850.26670.30435.15427.46720.8165
NEST0.05970.07820.20170.05970.07820.20170.21380.28240.31794.74716.88760.8546
PUREGREEN [59]0.25390.30020.03260.25390.30020.03260.65250.89320.041710.324714.26930.4952
CHROM [9]0.20960.27510.10590.20960.27510.07590.54040.82660.11739.791412.75680.3732
POS [63]0.19590.25710.16840.19590.25710.16840.53730.8460.14339.827313.44140.3432
Baseline0.06710.09230.60460.06710.09230.60460.28640.41840.55268.254211.17650.6832
NEST0.06350.08740.64220.06350.08740.64220.22550.35050.57347.688910.47830.7255
BUAAGREEN [59]0.34720.39510.08710.34720.39510.08710.64530.86320.09215.82317.98820.5624
CHROM [9]0.37860.32370.06820.37860.32370.06820.68130.88360.07156.09348.29380.5165
POS [63]0.31980.37620.09620.31980.37620.09620.62750.84240.11275.04077.11980.6374
Baseline0.14510.16810.28910.14510.16810.28910.55640.69040.29143.78526.32370.7492
NEST0.14360.16650.29550.14360.16650.29550.55140.68840.30043.37235.88060.7647
", + "image_path": "403e0e3f206a2ec8449726605cdbd6b4b582d08ac2bfaed5d48557206a3acdc0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 314, + 287, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 314, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 314, + 287, + 361 + ], + "type": "text", + "content": "Net and Dual-GAN consider the performance of the single domain instead of focusing on solving domain shift. Therefore, Rhythmnet is also selected as a baseline to further analyze the generalization on MSDG protocol." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 363, + 287, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 363, + 287, + 540 + ], + "spans": [ + { + "bbox": [ + 46, + 363, + 287, + 540 + ], + "type": "text", + "content": "The performance of DG methods. To further improve the generalization of the model, we apply the existing popular DG methods to baseline (Rhythmnet) as shown in Tab. 1. There is a slight model degradation of AD [12], Coral [53], and GroupDRO [38] on some target datasets. The reason behind this is that these algorithms employ inaccurate domain labeling, specifically, a dataset often with different acquisition devices, lighting, and scene cannot be considered as the same domain. On the contrary, VREx [18] and NCDG [57] can slightly improve the performance of Rhythmnet without utilizing the domain label. Inaccurate domain division will make it difficult for the DG algorithm to effectively align the feature distribution across different domains, which is termed as agnostic domain generalization." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 542, + 287, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 542, + 287, + 625 + ], + "spans": [ + { + "bbox": [ + 46, + 542, + 287, + 625 + ], + "type": "text", + "content": "Without the domain labels, NEST experimentally manifested significant improvement on MSDG protocol. For the target dataset BUAA, the results without NEST-CM " + }, + { + "bbox": [ + 46, + 542, + 287, + 625 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{CM}" + }, + { + "bbox": [ + 46, + 542, + 287, + 625 + ], + "type": "text", + "content": " show the best performance. Because NEST-CM may cause the model degradation with activating meaningless features at the instance level, especially for the BUAA with the slight domain shift." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 645, + 151, + 656 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 645, + 151, + 656 + ], + "spans": [ + { + "bbox": [ + 47, + 645, + 151, + 656 + ], + "type": "text", + "content": "4.3.2 HRV Estimation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "type": "text", + "content": "We also conduct experiments for HRV (i.e., LF, HF, and LF/HF) and HR measurement on the PURE, UBFC, and BUAA. The reason for choosing only these three datasets is that VIPL and V4V do not have reliable ground-truth" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 312, + 312, + 538, + 411 + ], + "blocks": [ + { + "bbox": [ + 178, + 287, + 414, + 298 + ], + "lines": [ + { + "bbox": [ + 178, + 287, + 414, + 298 + ], + "spans": [ + { + "bbox": [ + 178, + 287, + 414, + 298 + ], + "type": "text", + "content": "Table 2. HRV and HR estimation results on the MSDG protocol." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 312, + 312, + 538, + 411 + ], + "lines": [ + { + "bbox": [ + 312, + 312, + 538, + 411 + ], + "spans": [ + { + "bbox": [ + 312, + 312, + 538, + 411 + ], + "type": "table", + "html": "
MethodPUREBUAA
MAE↓RMSE↓r↑MAE↓RMSE↓r↑
GREEN [59]10.3214.270.525.827.990.56
CHROM [9]9.7912.760.376.098.290.51
POS [63]9.8213.440.345.047.120.63
Baseline7.5410.130.732.973.280.76
NEST6.079.060.762.562.730.78
", + "image_path": "a4e00464363c20a10874ff9629ffb384160c2b8d350b33772a2f6ca2612677af.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 414, + 545, + 436 + ], + "lines": [ + { + "bbox": [ + 306, + 414, + 545, + 436 + ], + "spans": [ + { + "bbox": [ + 306, + 414, + 545, + 436 + ], + "type": "text", + "content": "Table 3. HR estimation results on SSDG protocol by training on the UBFC-rPPG." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 451, + 545, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 451, + 545, + 522 + ], + "spans": [ + { + "bbox": [ + 305, + 451, + 545, + 522 + ], + "type": "text", + "content": "BVP signals, as mentioned in subsection 4.1. We also perform three tradition method (GREEN [59], CHMROM [9], POS [63]) by iPhys for comparison with our method as shown in Tab. 2. With a more robust NEST presentation distribution, our method also can get better performance than baseline and other traditional methods in all metrics." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 546, + 504, + 559 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 546, + 504, + 559 + ], + "spans": [ + { + "bbox": [ + 306, + 546, + 504, + 559 + ], + "type": "text", + "content": "4.4. Single-Source Domain Generalization" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": "For a worst-case scenario, the model is trained on only one dataset to generalize to the unseen domains, namely single-source domain generalization (SSDG). Based on this, we select UBFC-rPPG as the source domain and then select PURE and BUAA as the target domain to further evaluate the generalization of NEST as shown in Tab. 3. As we can see, NEST achieves desired performance by training on only one dataset. The reason is that: (1) NEST-CM can avoid the appearance of abnormal NEST; (2) NEST-TA can find the shared and robust feature of the samples within the same label; (3) NEST-DM can distinguish the boundaries between different samples to improve the generalization." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 733, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 733, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 733, + 317, + 742 + ], + "type": "text", + "content": "18595" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 53, + 70, + 279, + 231 + ], + "blocks": [ + { + "bbox": [ + 53, + 70, + 279, + 231 + ], + "lines": [ + { + "bbox": [ + 53, + 70, + 279, + 231 + ], + "spans": [ + { + "bbox": [ + 53, + 70, + 279, + 231 + ], + "type": "table", + "html": "
MethodSD↓MAE↓RMSE↓r↑
SAMC [58]18.015.921.00.11
POS [63]15.311.517.20.30
CHROM [9]15.111.416.90.28
I3D [3]15.912.015.90.07
DeepPhy [6]13.611.013.80.11
BVPNet [8]7.755.347.850.70
RhythmNet [36]8.115.308.140.76
CVD [37]7.925.027.970.79
Physformer [70]7.744.977.790.78
Dual-GAN [28]7.634.937.680.81
Baseline8.045.218.070.77
NEST7.494.767.510.84
", + "image_path": "768b69aab91735da2d8ea8ccadb35b656a388ba27d494461f8a0ba6129753315.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 235, + 286, + 257 + ], + "lines": [ + { + "bbox": [ + 47, + 235, + 286, + 257 + ], + "spans": [ + { + "bbox": [ + 47, + 235, + 286, + 257 + ], + "type": "text", + "content": "Table 4. HR estimation results by our method and several state-of-the-art methods on the VIPL-HR database." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 48, + 259, + 286, + 319 + ], + "blocks": [ + { + "bbox": [ + 48, + 259, + 286, + 319 + ], + "lines": [ + { + "bbox": [ + 48, + 259, + 286, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 259, + 286, + 319 + ], + "type": "image", + "image_path": "846f917949be5f611854fab9d3fdc219c7ceb03be288304b79b6f0faadc4342d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 324, + 287, + 346 + ], + "lines": [ + { + "bbox": [ + 47, + 324, + 287, + 346 + ], + "spans": [ + { + "bbox": [ + 47, + 324, + 287, + 346 + ], + "type": "text", + "content": "Figure 5. Impacts of the hyperparameter (a) " + }, + { + "bbox": [ + 47, + 324, + 287, + 346 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 47, + 324, + 287, + 346 + ], + "type": "text", + "content": ", (b) " + }, + { + "bbox": [ + 47, + 324, + 287, + 346 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 47, + 324, + 287, + 346 + ], + "type": "text", + "content": ", and (c) " + }, + { + "bbox": [ + 47, + 324, + 287, + 346 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 47, + 324, + 287, + 346 + ], + "type": "text", + "content": " of our proposed method." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 355, + 212, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 355, + 212, + 369 + ], + "spans": [ + { + "bbox": [ + 47, + 355, + 212, + 369 + ], + "type": "text", + "content": "4.5. Intra-Dataset Testing on VIPL" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 374, + 287, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 374, + 287, + 505 + ], + "spans": [ + { + "bbox": [ + 46, + 374, + 287, + 505 + ], + "type": "text", + "content": "Our proposed method can also improve the effectiveness of the model on intra-dataset testing by reducing the harmful effects of abnormal disturbances. Following [28, 36, 37, 70], we test NEST by using a 5-fold cross-validation protocol on VIPL-HR. Our method compared with ten baseline methods, including the traditional methods (SAMC [58], POS [63] and CHROM [9]) and the DL-based methods (I3D [3], DeepPhy [6], RhythmNet [36], CVD [37], BVP-Net [7], Physformer [70], and Dual-GAN [28]). The results of these methods are directly from [28, 70] as shown in Tab. 4." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 506, + 287, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 506, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 46, + 506, + 287, + 601 + ], + "type": "text", + "content": "As shown in Tab. 4, the proposed methods outperform all the methods under all metrics. Compared with the baseline, it also gets a noticeable promotion, i.e., 0.4 in MAE and 0.53 in RMSE. It is worth noting that the VIPL dataset has multiple complex scenes and recording devices and cannot be considered as a single domain, termed the agnostic domain problem. This proves that NEST can effectively alleviate the domain shift without domain labels." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 610, + 149, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 610, + 149, + 624 + ], + "spans": [ + { + "bbox": [ + 47, + 610, + 149, + 624 + ], + "type": "text", + "content": "4.6. Further Analysis" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "content": "We conduct further experiments on MSDG protocol (testing on UBFC with training on the other four datasets). Impact of " + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "content": " in NEST-CM. The hyperparameter " + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "content": " determines how many bases need to be strengthened by NEST-CM " + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{CM}" + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "content": ". As illustrated in Fig. 5 (a), NEST could achieve smaller RMSE when " + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\rho = 0.1" + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "content": ", indicating that the bases with small eigenvalue require more attention and fine-tuning." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 310, + 71, + 539, + 175 + ], + "blocks": [ + { + "bbox": [ + 310, + 71, + 539, + 175 + ], + "lines": [ + { + "bbox": [ + 310, + 71, + 539, + 175 + ], + "spans": [ + { + "bbox": [ + 310, + 71, + 539, + 175 + ], + "type": "image", + "image_path": "efa6d736dca4e50d7da261bf7c761e53794418d131d14c6ef8cc8dcc547bfa9e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 175, + 545, + 205 + ], + "lines": [ + { + "bbox": [ + 305, + 175, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 305, + 175, + 545, + 205 + ], + "type": "text", + "content": "Figure 6. The pearson's correlation coefficient of the NEST representation of the samples with different labels for the anchor HR 60." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 212, + 545, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 212, + 545, + 355 + ], + "spans": [ + { + "bbox": [ + 304, + 212, + 545, + 355 + ], + "type": "text", + "content": "Impact of " + }, + { + "bbox": [ + 304, + 212, + 545, + 355 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 304, + 212, + 545, + 355 + ], + "type": "text", + "content": " in NEST-TA. The hyperparameter " + }, + { + "bbox": [ + 304, + 212, + 545, + 355 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 304, + 212, + 545, + 355 + ], + "type": "text", + "content": " tradeoffs the contribution of each selected sample to form the synthesized prototype. As illustrated in Fig. 5, NEST could achieve the best performance with " + }, + { + "bbox": [ + 304, + 212, + 545, + 355 + ], + "type": "inline_equation", + "content": "\\sigma = 5" + }, + { + "bbox": [ + 304, + 212, + 545, + 355 + ], + "type": "text", + "content": ", which shows that the appropriate fusion of the nearby sample is conducive to improving model generalization. The specific reason behind this is that: (1) the synthesized prototype based on multi-domain samples can be used to enhance the domain invariant feature by NEST-TA; (2) NEST-TA can effectively aggregate the label relationship into NEST presentation distribution, which is performed by improving the similarity of NEST representation between the adjacent label samples." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 357, + 546, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 357, + 546, + 453 + ], + "spans": [ + { + "bbox": [ + 304, + 357, + 546, + 453 + ], + "type": "text", + "content": "To further clarify the second reason, we calculate the Pearson's correlation coefficient of the NEST representation of the samples with different labels for the anchor HR 60, as shown in Fig. 6. It is obvious that the distribution changes with NEST-TA loss are more smooth across the continuous HR labels, which allows OOD samples with slight domain shifts to predict nearby heart rates around the ground truth HR value instead of other abnormal HR values." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 454, + 545, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 454, + 545, + 538 + ], + "spans": [ + { + "bbox": [ + 304, + 454, + 545, + 538 + ], + "type": "text", + "content": "Impact of " + }, + { + "bbox": [ + 304, + 454, + 545, + 538 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 454, + 545, + 538 + ], + "type": "text", + "content": " in NEST-DM. The hyperparameter " + }, + { + "bbox": [ + 304, + 454, + 545, + 538 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 454, + 545, + 538 + ], + "type": "text", + "content": " determines the pushing strength of the different samples. With the smallest " + }, + { + "bbox": [ + 304, + 454, + 545, + 538 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 454, + 545, + 538 + ], + "type": "text", + "content": ", NEST-DM " + }, + { + "bbox": [ + 304, + 454, + 545, + 538 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{DM}" + }, + { + "bbox": [ + 304, + 454, + 545, + 538 + ], + "type": "text", + "content": " has the overemphasized constraints to distinguish the NEST representation between different samples, which causes a harmful impact on the discriminant ability of the model. Overall, NEST-DM " + }, + { + "bbox": [ + 304, + 454, + 545, + 538 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{DM}" + }, + { + "bbox": [ + 304, + 454, + 545, + 538 + ], + "type": "text", + "content": " is an effective way to improve the generalization of model." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 551, + 378, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 551, + 378, + 563 + ], + "spans": [ + { + "bbox": [ + 306, + 551, + 378, + 563 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": "This paper presents the NEural SStructure (NEST) modeling, a novel framework that addresses the domain generalization problem in rPPG measurement for the first time. Our framework contains three regularization terms defined in the view of neural structure, which improves the generalization capacity in the different aspects. All three terms do not rely on domain labels, which are especially suitable in rPPG, as defining domains is difficult in this task. To compare the generalization of the existing algorithm and our method, we establish the first DG protocol for this task. We experimentally manifested the substantial improvement of our method on both MSDG and SSDG protocols." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 733, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 733, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 733, + 317, + 742 + ], + "type": "text", + "content": "18596" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "text", + "content": "[1] Serge Bobbia, Richard Macwan, Yannick Benezeth, Alamin Mansouri, and Julien Dubois. Unsupervised skin tissue segmentation for remote photoplethysmography. Pattern Recognition Letters, 124:82-90, 2019." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 287, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 287, + 179 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 287, + 179 + ], + "type": "text", + "content": "[2] Adrian Bulat and Georgios Tzimiropoulos. How far are we from solving the 2d & 3d face alignment problem? (and a dataset of 230,000 3d facial landmarks). In Proc. ICCV, 2017." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 182, + 286, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 182, + 286, + 214 + ], + "spans": [ + { + "bbox": [ + 53, + 182, + 286, + 214 + ], + "type": "text", + "content": "[3] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proc. IEEE CVPR, pages 6299-6308, 2017." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 216, + 286, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 216, + 286, + 248 + ], + "spans": [ + { + "bbox": [ + 53, + 216, + 286, + 248 + ], + "type": "text", + "content": "[4] Beidi Chen, Tri Dao, Eric Winsor, Zhao Song, Atri Rudra, and Christopher Ré. Scatterbrain: Unifying sparse and low-rank attention. NeurIPS, 34:17413-17426, 2021." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 250, + 286, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 250, + 286, + 293 + ], + "spans": [ + { + "bbox": [ + 53, + 250, + 286, + 293 + ], + "type": "text", + "content": "[5] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, pages 1597-1607. PMLR, 2020." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 295, + 286, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 295, + 286, + 328 + ], + "spans": [ + { + "bbox": [ + 53, + 295, + 286, + 328 + ], + "type": "text", + "content": "[6] Weixuan Chen and Daniel McDuff. Deep learning: Video-based physiological measurement using convolutional attention networks. In Proc. ECCV, pages 349-365, 2018." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 330, + 286, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 330, + 286, + 372 + ], + "spans": [ + { + "bbox": [ + 53, + 330, + 286, + 372 + ], + "type": "text", + "content": "[7] Abhijit Das, Hao Lu, Hu Han, Antitza Dantcheva, Shiguang Shan, and Xilin Chen. Bvpnet: Video-to-bvp signal prediction for remote heart rate estimation. In " + }, + { + "bbox": [ + 53, + 330, + 286, + 372 + ], + "type": "inline_equation", + "content": "FG" + }, + { + "bbox": [ + 53, + 330, + 286, + 372 + ], + "type": "text", + "content": ", pages 01-08. IEEE, 2021." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 374, + 286, + 417 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 374, + 286, + 417 + ], + "spans": [ + { + "bbox": [ + 53, + 374, + 286, + 417 + ], + "type": "text", + "content": "[8] Abhijit Das, Hao Lu, Hu Han, Antitza Dantcheva, Shiguang Shan, and Xilin Chen. Bvpnet: Video-to-bvp signal prediction for remote heart rate estimation. In " + }, + { + "bbox": [ + 53, + 374, + 286, + 417 + ], + "type": "inline_equation", + "content": "FG" + }, + { + "bbox": [ + 53, + 374, + 286, + 417 + ], + "type": "text", + "content": ", pages 01-08. IEEE, 2021." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 419, + 286, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 419, + 286, + 452 + ], + "spans": [ + { + "bbox": [ + 53, + 419, + 286, + 452 + ], + "type": "text", + "content": "[9] Gerard De Haan and Vincent Jeanne. Robust pulse rate from chrominance-based rppg. IEEE Trans. Biomed. Eng., 60(10):2878-2886, 2013." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 453, + 287, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 453, + 287, + 486 + ], + "spans": [ + { + "bbox": [ + 48, + 453, + 287, + 486 + ], + "type": "text", + "content": "[10] Gerard De Haan and Arno Van Leest. Improved motion robustness of remote-ppg by using the blood volume pulse signature. Physiol. Meas., 35(9):1913, 2014." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 487, + 287, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 487, + 287, + 520 + ], + "spans": [ + { + "bbox": [ + 48, + 487, + 287, + 520 + ], + "type": "text", + "content": "[11] Zhijie Deng, Yucen Luo, and Jun Zhu. Cluster alignment with a teacher for unsupervised domain adaptation. In Proc. CVPR, pages 9944–9953, 2019." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 522, + 287, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 522, + 287, + 554 + ], + "spans": [ + { + "bbox": [ + 48, + 522, + 287, + 554 + ], + "type": "text", + "content": "[12] Yaroslav Ganin and Victor Lempitsky. Unsupervised domain adaptation by backpropagation. In ICML, pages 1180-1189. PMLR, 2015." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 555, + 287, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 555, + 287, + 589 + ], + "spans": [ + { + "bbox": [ + 48, + 555, + 287, + 589 + ], + "type": "text", + "content": "[13] Ming-Shiang Chen Gee-Sern Hsu, ArulMurugan Ambikapathi. Deep learning with time-frequency representation for pulse estimation. In Proc. IJCB, pages 642-650, 2017." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 590, + 287, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 633 + ], + "type": "text", + "content": "[14] Min Hu, Fei Qian, Dong Guo, Xiaohua Wang, Lei He, and Fuji Ren. Eta-rppgnet: effective time-domain attention network for remote heart rate measurement. IEEE Trans. on Instrum. and Meas., 70:1-12, 2021." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 635, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 668 + ], + "type": "text", + "content": "[15] Yunpei Jia, Jie Zhang, Shiguang Shan, and Xilin Chen. Single-side domain generalization for face anti-spoofing. In Proc. CVPR, pages 8484-8493, 2020." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[16] Xiang Jiang, Qicheng Lao, Stan Matwin, and Mohammad Havaei. Implicit class-conditioned domain alignment for unsupervised domain adaptation. In ICML, pages 4816-4827. PMLR, 2020." + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[17] Viktor Kessler, Patrick Thiam, Mohammadreza Amirian, and Friedhelm Schwenker. Pain recognition with camera photoplethysmography. In IEEE IPTA, pages 1-5, 2017." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 107, + 545, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 160 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 160 + ], + "type": "text", + "content": "[18] David Krueger, Ethan Caballero, Joern-Henrik Jacobsen, Amy Zhang, Jonathan Binas, Dinghuai Zhang, Remi Le Priol, and Aaron Courville. Out-of-distribution generalization via risk extrapolation (rex). In ICML, pages 5815-5826. PMLR, 2021." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 162, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 162, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 308, + 162, + 545, + 194 + ], + "type": "text", + "content": "[19] Eugene Lee, Evan Chen, and Chen-Yi Lee. Meta-rppg: Remote heart rate estimation using a transductive meta-learner. In ECCV, pages 392–409. Springer, 2020." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 195, + 545, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 195, + 545, + 258 + ], + "spans": [ + { + "bbox": [ + 308, + 195, + 545, + 258 + ], + "type": "text", + "content": "[20] Alexander Lehner, Stefano Gasperini, Alvaro Marcos-Ramiro, Michael Schmidt, Mohammad-Ali Nikouei Mahani, Nassir Navab, Benjamin Busam, and Federico Tombari. 3d-vfield: Adversarial augmentation of point clouds for domain generalization in 3d object detection. In Proc. CVPR, pages 17295-17304, 2022." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 261, + 545, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 261, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 308, + 261, + 545, + 304 + ], + "type": "text", + "content": "[21] Magdalena Lewandowska, Jacek Ruminski, Tomasz Kocejko, and Jedrzej Nowak. Measuring pulse rate with a webcam—a non-contact method for evaluating cardiac activity. In FedCSIS, pages 405-410. IEEE, 2011." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 305, + 545, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 305, + 545, + 349 + ], + "spans": [ + { + "bbox": [ + 308, + 305, + 545, + 349 + ], + "type": "text", + "content": "[22] Daiqing Li, Junlin Yang, Karsten Kreis, Antonio Torralba, and Sanja Fidler. Semantic segmentation with generative models: Semi-supervised learning and strong out-of-domain generalization. In Proc. CVPR, pages 8300-8311, 2021." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 350, + 545, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 350, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 308, + 350, + 545, + 381 + ], + "type": "text", + "content": "[23] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy Hospedales. Learning to generalize: Meta-learning for domain generalization. In Proc. AAAI, 2018." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 383, + 545, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 383, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 308, + 383, + 545, + 425 + ], + "type": "text", + "content": "[24] Chuang Lin, Zehuan Yuan, Sicheng Zhao, Peize Sun, Changhu Wang, and Jianfei Cai. Domain-invariant disentangled network for generalizable object detection. In Proce. CVPR, pages 8771-8780, 2021." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 426, + 545, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 426, + 545, + 468 + ], + "spans": [ + { + "bbox": [ + 308, + 426, + 545, + 468 + ], + "type": "text", + "content": "[25] Xin Liu, Josh Fromm, Shwetak Patel, and Daniel McDuff. Multi-task temporal shift attention networks for on-device contactless vitals measurement. NeurIPS, 33:19400-19411, 2020." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 471, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 471, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 308, + 471, + 545, + 514 + ], + "type": "text", + "content": "[26] Xin Liu, Ziheng Jiang, Josh Fromm, Xuhai Xu, Shwetak Patel, and Daniel McDuff. Metaphys: few-shot adaptation for non-contact physiological measurement. In Proc. HIL, pages 154-163, 2021." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 515, + 545, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 515, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 308, + 515, + 545, + 547 + ], + "type": "text", + "content": "[27] Zhenhua Liu, Jizheng Xu, Xiulian Peng, and Ruiqin Xiong. Frequency-domain dynamic pruning for convolutional neural networks. NeurIPS, 31, 2018." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 548, + 545, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 548, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 308, + 548, + 545, + 581 + ], + "type": "text", + "content": "[28] Hao Lu, Hu Han, and S Kevin Zhou. Dual-gan: Joint bvp and noise modeling for remote physiological measurement. In Proc. IEEE CVPR, pages 12404-12413, 2021." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 582, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 582, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 582, + 545, + 624 + ], + "type": "text", + "content": "[29] Yawei Luo, Ping Liu, Liang Zheng, Tao Guan, Junqing Yu, and Yi Yang. Category-level adversarial adaptation for semantic segmentation using purified features. IEEE Trans. Pattern Anal. Mach. Intell., 2021." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 625, + 545, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 625, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 308, + 625, + 545, + 669 + ], + "type": "text", + "content": "[30] Fangrui Lv, Jian Liang, Shuang Li, Bin Zang, Chi Harold Liu, Ziteng Wang, and Di Liu. Causality inspired representation learning for domain generalization. In Proc. CVPR, pages 8046-8056, 2022." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "type": "text", + "content": "[31] Lei Ma, Felix Juefei-Xu, Fuyuan Zhang, Jiyuan Sun, Minhui Xue, Bo Li, Chunyang Chen, Ting Su, Li Li, Yang Liu, et al. Deep gauge: Multi-granularity testing criteria for deep learning systems. In Proc. ASE, pages 120-131, 2018." + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 734, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 734, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 734, + 317, + 742 + ], + "type": "text", + "content": "18597" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 286, + 95 + ], + "type": "text", + "content": "[32] Daniel McDuff. Camera measurement of physiological vital signs. CSUR, 2021." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 286, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 286, + 128 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 286, + 128 + ], + "type": "text", + "content": "[33] Daniel McDuff. Applications of camera-based physiological measurement beyond healthcare. In CVSM, pages 165-177. Elsevier, 2022." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 130, + 286, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 130, + 286, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 130, + 286, + 173 + ], + "type": "text", + "content": "[34] Daniel McDuff, Sarah Gontarek, and Rosalind W Picard. Improvements in remote cardiopulmonary measurement using a five band digital camera. IEEE Trans. Biomed. Eng., 61(10):2593-2601, 2014." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 175, + 286, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 286, + 207 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 286, + 207 + ], + "type": "text", + "content": "[35] Xuesong Niu, Hu Han, Shiguang Shan, and Xilin Chen. Synrhythm: Learning a deep heart rate estimator from general to specific. In Proc. IEEE ICPR, pages 3580-3585, 2018." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 209, + 286, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 286, + 251 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 286, + 251 + ], + "type": "text", + "content": "[36] Xuesong Niu, Shiguang Shan, Hu Han, and Xilin Chen. Rhythmnet: End-to-end heart rate estimation from face via spatial-temporal representation. IEEE Trans. on Image Process., 29:2409-2423, 2020." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 254, + 286, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 254, + 286, + 296 + ], + "spans": [ + { + "bbox": [ + 48, + 254, + 286, + 296 + ], + "type": "text", + "content": "[37] Xuesong Niu, Zitong Yu, Hu Han, Xiaobai Li, Shiguang Shan, and Guoying Zhao. Video-based remote physiological measurement via cross-verified feature disentangling. In Proc. ECCV, 2020." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 298, + 286, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 298, + 286, + 330 + ], + "spans": [ + { + "bbox": [ + 48, + 298, + 286, + 330 + ], + "type": "text", + "content": "[38] Giambattista Parascandolo, Alexander Neitz, Antonio Orvieto, Luigi Gresele, and Bernhard Scholkopf. Learning explanations that are hard to vary. *ICLR2022*, 2022." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 332, + 286, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 332, + 286, + 365 + ], + "spans": [ + { + "bbox": [ + 48, + 332, + 286, + 365 + ], + "type": "text", + "content": "[39] Kexin Pei, Yinzhi Cao, Junfeng Yang, and Suman Jana. Deepxplore: Automated whitebox testing of deep learning systems. In Proc. SOSP, pages 1-18, 2017." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 366, + 286, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 366, + 286, + 398 + ], + "spans": [ + { + "bbox": [ + 48, + 366, + 286, + 398 + ], + "type": "text", + "content": "[40] Duo Peng, Yinjie Lei, Munawar Hayat, Yulan Guo, and Wen Li. Semantic-aware domain generalized segmentation. In Proc. CVPR, pages 2594-2605, 2022." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 399, + 286, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 286, + 432 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 286, + 432 + ], + "type": "text", + "content": "[41] Vihari Piratla, Praneeth Netrapalli, and Sunita Sarawagi. Efficient domain generalization via common-specific low-rank decomposition. In ICML, pages 7728-7738. PMLR, 2020." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 434, + 286, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 434, + 286, + 476 + ], + "spans": [ + { + "bbox": [ + 48, + 434, + 286, + 476 + ], + "type": "text", + "content": "[42] Ming-Zher Poh, Daniel J McDuff, and Rosalind W Picard. Non-contact, automated cardiac pulse measurements using video imaging and blind source separation. Opt. Express, 18(10):10762-10774, 2010." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 479, + 286, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 479, + 286, + 510 + ], + "spans": [ + { + "bbox": [ + 48, + 479, + 286, + 510 + ], + "type": "text", + "content": "[43] Fengchun Qiao, Long Zhao, and Xi Peng. Learning to learn single domain generalization. In Proc. CVPR, pages 12556-12565, 2020." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 512, + 286, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 512, + 286, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 512, + 286, + 555 + ], + "type": "text", + "content": "[44] Ying Qiu, Yang Liu, Juan Arteaga-Falconi, Haiwei Dong, and Abdulmotaleb El Saddik. Evm-cnn: Real-time contactless heart rate estimation from facial video. IEEE Trans. Multimedia, 2019." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 557, + 286, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 557, + 286, + 600 + ], + "spans": [ + { + "bbox": [ + 48, + 557, + 286, + 600 + ], + "type": "text", + "content": "[45] Attila Reiss, Ina Indlekofer, Philip Schmidt, and Kristof Van Laerhoven. Deep ppg: large-scale heart rate estimation with convolutional neural networks. Sensors, 19(14):3079, 2019." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 602, + 286, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 286, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 286, + 644 + ], + "type": "text", + "content": "[46] Jiahuan Ren, Zhao Zhang, Richang Hong, Mingliang Xu, Hajun Zhang, Mingbo Zhao, and Meng Wang. Robust low-rank convolution network for image denoising. In Proc. ACM MM, pages 6211-6219, 2022." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 647, + 286, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 286, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 286, + 689 + ], + "type": "text", + "content": "[47] Ambareesh Revanur, Zhihua Li, Umur A Ciftci, Lijun Yin, and László A Jeni. The first vision for vitals (v4v) challenge for non-contact video-based physiological estimation. In Proc. CVPR workshop, pages 2760-2767, 2021." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 692, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 692, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 692, + 286, + 713 + ], + "type": "text", + "content": "[48] Shiv Shankar, Vihari Piratla, Soumen Chakrabarti, Siddhartha Chaudhuri, Preethi Jyothi, and Sunita Sarawagi." + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 327, + 73, + 544, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 544, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 544, + 95 + ], + "type": "text", + "content": "Generalizing across domains via cross-gradient training. ICLR, 2018." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "type": "text", + "content": "[49] Rencheng Song, Huan Chen, Juan Cheng, Chang Li, Yu Liu, and Xun Chen. Pulsegan: Learning to generate realistic pulse waveforms in remote photoplethysmography. IEEE J-BHI, pages 1-1, 2021." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 141, + 545, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 184 + ], + "type": "text", + "content": "[50] Rencheng Song, Senle Zhang, Chang Li, Yunfei Zhang, Juan Cheng, and Xun Chen. Heart rate estimation from facial videos using a spatiotemporal representation with convolutional neural networks. IEEE Trans. Instrum. Meas., 2020." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 186, + 545, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 218 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 218 + ], + "type": "text", + "content": "[51] Radim Špetlík, Vojtech Franc, and Jírí Matas. Visual heart rate estimation with convolutional neural network. In Proc. BMVC, pages 3-6, 2018." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 220, + 545, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 545, + 261 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 545, + 261 + ], + "type": "text", + "content": "[52] Ronny Stricker, Steffen Müller, and Horst-Michael Gross. Non-contact video-based pulse rate measurement on a mobile service robot. In Proc. IEEE ISRHIC, pages 1056-1062, 2014." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 264, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 264, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 307, + 264, + 545, + 297 + ], + "type": "text", + "content": "[53] Baochen Sun and Kate Saenko. Deep coral: Correlation alignment for deep domain adaptation. In ECCV, pages 443-450. Springer, 2016." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 299, + 545, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 299, + 545, + 341 + ], + "spans": [ + { + "bbox": [ + 307, + 299, + 545, + 341 + ], + "type": "text", + "content": "[54] Bo Sun, Qinglan Wei, Liandong Li, Qihua Xu, Jun He, and Lejun Yu. Lstm for dynamic emotion and group emotion recognition in the wild. In Proc. ACM ICMI, pages 451-457, 2016." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 343, + 545, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 343, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 307, + 343, + 545, + 376 + ], + "type": "text", + "content": "[55] Zhaodong Sun and Xiaobai Li. Contrast-phys: Unsupervised video-based remote physiological measurement via spatiotemporal contrast. In Proc. ECCV, 2022." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 377, + 545, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 377, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 307, + 377, + 545, + 420 + ], + "type": "text", + "content": "[56] Ling Tang, Wen Shen, Zhanpeng Zhou, Yuefeng Chen, and Quanshi Zhang. Defects of convolutional decoder networks in frequency representation. arXiv preprint arXiv:2210.09020, 2022." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 422, + 545, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 422, + 545, + 454 + ], + "spans": [ + { + "bbox": [ + 307, + 422, + 545, + 454 + ], + "type": "text", + "content": "[57] Chris Xing Tian, Haoliang Li, Xiaofei Xie, Yang Liu, and Shiqi Wang. Neuron coverage-guided domain generalization. IEEE Trans. on Pattern Anal. Mach. Intell., 2022." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 456, + 545, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 456, + 545, + 509 + ], + "spans": [ + { + "bbox": [ + 307, + 456, + 545, + 509 + ], + "type": "text", + "content": "[58] Sergey Tulyakov, Xavier Alameda-Pineda, Elisa Ricci, Lijun Yin, Jeffrey F Cohn, and Nicu Sebe. Self-adaptive matrix completion for heart rate estimation from face videos under realistic conditions. In Proc. IEEE CVPR, pages 2396-2404, 2016." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 512, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 512, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 307, + 512, + 545, + 544 + ], + "type": "text", + "content": "[59] Wim Verkruysse, Lars O Svaasand, and J Stuart Nelson. Remote plethysmographic imaging using ambient light. Opt. Express, 16(26):21434-21445, 2008." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 546, + 545, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 546, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 307, + 546, + 545, + 590 + ], + "type": "text", + "content": "[60] Tan Wad, Qianru Sun, Sugiri Pranata, Karlekar Jayashree, and Hanwang Zhang. Equivalence and invariance inductive bias for learning from insufficient data. In ECCV, pages 241-258. Springer, 2022." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 591, + 545, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 591, + 545, + 633 + ], + "spans": [ + { + "bbox": [ + 307, + 591, + 545, + 633 + ], + "type": "text", + "content": "[61] Hao Wang, Euijoon Ahn, and Jinman Kim. Self-supervised representation learning framework for remote physiological measurement using spatiotemporal augmentation loss. In ACM MM, 2022." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 635, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 545, + 679 + ], + "type": "text", + "content": "[62] Jindong Wang, Cuiling Lan, Chang Liu, Yidong Ouyang, Tao Qin, Wang Lu, Yiqiang Chen, Wenjun Zeng, and Philip Yu. Generalizing to unseen domains: A survey on domain generalization. IEEE Trans. on Knowl. Data Eng., 2022." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "type": "text", + "content": "[63] Wenjin Wang, Albertus C den Brinker, Sander Stuijk, and Gerard de Haan. Algorithmic principles of remote ppg. IEEE Trans. Biomed. Eng., 64(7):1479-1491, 2017." + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 734, + 318, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 734, + 318, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 734, + 318, + 742 + ], + "type": "text", + "content": "18598" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 408 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "type": "text", + "content": "[64] Wenjin Wang, Sander Stuijk, and Gerard De Haan. Exploiting spatial redundancy of image sensor for motion robust rppg. IEEE Trans. Biomed. Eng., 62(2):415-425, 2015." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 107, + 286, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 107, + 286, + 140 + ], + "spans": [ + { + "bbox": [ + 49, + 107, + 286, + 140 + ], + "type": "text", + "content": "[65] Yunhe Wang, Chang Xu, Shan You, Dacheng Tao, and Chao Xu. Cnnpack: Packing convolutional neural networks in the frequency domain. In NeurIPS, volume 29, 2016." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 141, + 286, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 141, + 286, + 174 + ], + "spans": [ + { + "bbox": [ + 49, + 141, + 286, + 174 + ], + "type": "text", + "content": "[66] Zijian Wang, Yadan Luo, Ruihong Qiu, Zi Huang, and Mahsa Baktashmotlagh. Learning to diversify for single domain generalization. In Proc. CVPR, pages 834-843, 2021." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 175, + 286, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 175, + 286, + 207 + ], + "spans": [ + { + "bbox": [ + 49, + 175, + 286, + 207 + ], + "type": "text", + "content": "[67] Zhi-Kuan Wang, Ying Kao, and Chiou-Ting Hsu. Vision-based heart rate estimation via a two-stream cnn. In Proc. IEEE ICIP, pages 3327-3331, 2019." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 209, + 286, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 209, + 286, + 251 + ], + "spans": [ + { + "bbox": [ + 49, + 209, + 286, + 251 + ], + "type": "text", + "content": "[68] Lin Xi, Weihai Chen, Changchen Zhao, Xingming Wu, and Jianhua Wang. Image enhancement for remote photoplethysmography in a low-light environment. In FG, pages 1-7. IEEE, 2020." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 254, + 286, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 254, + 286, + 307 + ], + "spans": [ + { + "bbox": [ + 49, + 254, + 286, + 307 + ], + "type": "text", + "content": "[69] Zitong Yu, Wei Peng, Xiaobai Li, Xiaopeng Hong, and Guoying Zhao. Remote heart rate measurement from highly compressed facial videos: an end-to-end deep learning solution with video enhancement. In Proc. IEEE ICCV, pages 151-160, 2019." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 309, + 286, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 309, + 286, + 352 + ], + "spans": [ + { + "bbox": [ + 49, + 309, + 286, + 352 + ], + "type": "text", + "content": "[70] Zitong Yu, Yuming Shen, Jingang Shi, Hengshuang Zhao, Philip Torr, and Guoying Zhao. Physformer: Facial videobased physiological measurement with temporal difference transformer. IEEE CVPR, 2022." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 354, + 286, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 354, + 286, + 408 + ], + "spans": [ + { + "bbox": [ + 49, + 354, + 286, + 408 + ], + "type": "text", + "content": "[71] Xiangyu Yue, Yang Zhang, Sicheng Zhao, Alberto Sangiovanni-Vincentelli, Kurt Keutzer, and Boqing Gong. Domain randomization and pyramid consistency: Simulation-to-real generalization without accessing target domain data. In Proc.CVPR, pages 2100-2110, 2019." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 734, + 317, + 742 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 734, + 317, + 742 + ], + "spans": [ + { + "bbox": [ + 294, + 734, + 317, + 742 + ], + "type": "text", + "content": "18599" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/NewsNet_ A Novel Dataset for Hierarchical Temporal Segmentation/d2756edd-2b85-4c25-85cd-99de3bdf0d9a_content_list.json b/2023/NewsNet_ A Novel Dataset for Hierarchical Temporal Segmentation/d2756edd-2b85-4c25-85cd-99de3bdf0d9a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..fa1a15aefb7c173748c39802752bb38ebea28611 --- /dev/null +++ b/2023/NewsNet_ A Novel Dataset for Hierarchical Temporal Segmentation/d2756edd-2b85-4c25-85cd-99de3bdf0d9a_content_list.json @@ -0,0 +1,1474 @@ +[ + { + "type": "text", + "text": "NewsNet: A Novel Dataset for Hierarchical Temporal Segmentation", + "text_level": 1, + "bbox": [ + 142, + 130, + 826, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Haoqian $\\mathbf{W}\\mathbf{u}^{1\\dagger}$ , Keyu Chen $^{1\\dagger}$ , Haozhe Liu $^{2\\dagger}$ , Mingchen Zhuge $^{2\\dagger}$ , Bing Li $^{2\\boxtimes}$ , Ruizhi Qiao $^{1\\boxtimes}$ , Xiujun Shu $^{1}$ , Bei Gan $^{1}$ , Liangsheng Xu $^{1}$ , Bo Ren $^{1}$ , Mengmeng Xu $^{2}$ , Wentian Zhang $^{2}$ , Raghavendra Ramachandra $^{3}$ , Chia-Wen Lin $^{4}$ , Bernard Ghanem $^{2}$", + "bbox": [ + 84, + 179, + 880, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Tencent $^{2}$ AI Initiative, King Abdullah University of Science and Technology (KAUST) \n $^{3}$ Norwegian University of Science and Technology (NTNU) \n $^{4}$ National Tsing Hua University (NTHU)", + "bbox": [ + 132, + 239, + 836, + 295 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/573668c2860b1009046c02f06c69b022fdb789c6a1705bb0ea19027ffd604e94.jpg", + "image_caption": [ + "Figure 1. Examples from the proposed NewsNet. The dataset provides multimodal and hierarchical labels on each video, including 4-level hierarchical temporal information (i.e., event, scene, story, topic) and 3 modality-specific cues (i.e., video, audio, text). A prerequisite motivation behind NewsNet is that long-form videos can be recursively represented into several sub-videos according to the semantics, which is an intrinsic property of the video but rarely explored in the community. To bridge this research gap, this paper innovatively constructs a large-scale dataset to benchmark the algorithms for hierarchical understanding like a human being." + ], + "image_footnote": [], + "bbox": [ + 94, + 308, + 883, + 534 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 267, + 628, + 346, + 645 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Temporal video segmentation is the get-to-go automatic video analysis, which decomposes a long-form video into smaller components for the following-up understanding tasks. Recent works have studied several levels of granularity to segment a video, such as shot, event, and scene. Those segmentations can help compare the semantics in the corresponding scales, but lack a wider view of larger temporal spans, especially when the video is complex and structured. Therefore, we present two abstractive levels of temporal segmentations and study their hierarchy to the existing fine-grained levels. Accordingly, we collect NewsNet, the largest news video dataset consisting of 1,000 videos", + "bbox": [ + 75, + 662, + 472, + 845 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "in over 900 hours, associated with several tasks for hierarchical temporal video segmentation. Each news video is a collection of stories on different topics, represented as aligned audio, visual, and textual data, along with extensive frame-wise annotations in four granularities. We assert that the study on NewsNet can advance the understanding of complex structured video and benefit more areas such as short-video creation, personalized advertisement, digital instruction, and education. Our dataset and code is publicly available at https://github.com/NewsNet-Benchmark/NewsNet.", + "bbox": [ + 500, + 630, + 893, + 796 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 828, + 632, + 844 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Temporal video segmentation is a critical problem in video understanding, which is essential for many video applications such as video classification [10, 15, 37, 58, 59, 62],", + "bbox": [ + 498, + 854, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 47 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Equal Contribution", + "bbox": [ + 101, + 863, + 215, + 875 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Corresponding Authors: Bing Li (bing.li@kaust.edu.sa) and Ruizhi Qiao (ruizhiqiao@tencent.com).", + "bbox": [ + 78, + 876, + 468, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "10669", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/5a983113068e6f41caad4e4ca5069e3d53aa340446276b1ea9f97ec3e0e73408.jpg", + "table_caption": [ + "Table 1. Data comparison between the NewsNet and other related datasets. The NewsNet provides various multimodal data and hierarchical temporal segmentation annotations. Doc: documentary, Ads: advertisement. (Please refer to our project page for more details.)" + ], + "table_footnote": [ + "The number of annotations for Scene and Shot is counted from MovieScene [47], which is a subset of MovieNet." + ], + "table_body": "
Dataset# VideoDuration (hours)Modality# Annotation(s) per VideoSource
TopicStorySceneEvent
AVS [75]197-Visual---14.2Ads
BBC [6]119Visual--49.7-Doc
OVSD [48]2110Visual--28.9-Generic
Kinetics-GEBD [51]54,691152Audio + Visual---4.9Action
MovieNet [23] †1,1002174Text + Audio + Visual--66.0849.1Movie
RAI [7]10-Visual---98.7News
TI-News [35]477244Audio + Visual-55.6-530.4News
NewsNet (Ours)1,000946Text + Audio + Visual8.551.687.9654.4News
", + "bbox": [ + 109, + 128, + 859, + 295 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "captioning [3, 34, 63, 71] and retrieval [5, 16, 31, 52]. Temporal video segmentation aims to group successive video frames into short segments along the temporal dimension. With the explosive growth of long-form videos, it is desirable that temporal video segmentation can convert a video into more meaningful segments for more efficient access to the video. However, it is challenging to develop effective temporal segmentation tools for long-form videos, since this requires a comprehensive understanding of video structure, while long-form videos contain complex content.", + "bbox": [ + 75, + 333, + 472, + 486 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Towards temporal video segmentation, existing works explore shot, event, and scene segmentation tasks, respectively. Shot segmentation [38, 55, 57] divides a video into shots, where a shot consists of consecutive and visually continuous frames captured by a camera without interruption [23]. Yet, shot segmentation only considers low-level visual cues (i.e. visual similarity), lacking semantic understanding. Instead, event segmentation [25, 51, 56] divides a video by detecting the moments of changes such as action/subject changes. To better capture the underlying structure of a video, recent works [12, 47, 65] introduce video scene segmentation which segments a video into scene segments, each comprising successive shots semantically related to the same scene. Scene segmentation enables a coarser and higher-level representation than shot segmentation. However, compared to the rich content of massive long-form videos, scene/event is fine-grained and often lacks a high-level summarization of video content, which is insufficient for capturing the complex semantic structure of many videos and briefly representing video content.", + "bbox": [ + 75, + 488, + 472, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we first explore how to comprehensively represent the complex structure of a long-form video for temporal video segmentation. Humans can hierarchically divide a video into segments of different granularities according to multi-level semantic information (e.g. scene and topic), from the perspective of cognitive science. Natural language processing researchers have widely explored", + "bbox": [ + 75, + 794, + 470, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "topic-level understanding [29, 41] for summarizing documents [8, 42], while little effort has been devoted to long-form videos. Inspired by these observations, besides scene and event, we propose to introduce two higher-level semantics (i.e. story and topic) into temporal video segmentation, to provide a brief and semantic structure representation. As a result, such hierarchical and multi-level understanding brings about scalable video structure representation for temporal video segmentation on long-form videos. That is, a long-form video can be split into finer segments with lower-level semantics (e.g. scene), but also can be summarized into coarser ones yet with higher-level semantics (e.g. topic) by recursively grouping finer segments, which comprehensively represents video structures from coarse to fine.", + "bbox": [ + 496, + 333, + 893, + 547 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, the community lacks high-quality datasets to conduct this research. In particular, as shown in Table 1, most datasets only provide temporal structure annotations with regard to events or scenes. TI-News [35] and MovieScene [47] provide two levels of annotations, but these datasets lack topic-level ones.", + "bbox": [ + 496, + 549, + 893, + 641 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To effectively break this limitation, we build a novel large-scale dataset for hierarchical temporal segmentation, named NewsNet. The unique properties of our NewsNet introduce many advantages. First, it is among the largest datasets in the news domain. We collect over 900 hours of videos from 20 mainstream news platforms. It has a highly diverse distribution of data. Second, we carefully annotated it frame-by-frame with 4 hierarchical levels to ensure its quality can meet our needs. Third, it is multimodal, including textual, visual, and audio information. Due to the nature of the news, the alignment across modalities is accurate, which makes multimodal joint training of models feasible. Finally, the videos in NewsNet provide a complete understanding of public events. Compared with other video datasets [4,23,26], it introduces more objective open-world knowledge (e.g., news introduction) while including subjective factual commentary (e.g., host comments on news", + "bbox": [ + 496, + 643, + 895, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "10670", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "events), making it more amenable to real-life application.", + "bbox": [ + 76, + 90, + 454, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Based on NewsNet, we empirically highlight two promising directions for long-span temporal segmentation: 1) Infusing Multi-Modality knowledge can significantly improve the performance of long-form temporal segmentation; 2) Although story- and topic-level segmentation is challenging, it can be benefited from hierarchical modeling with the event- and scene-level segmentation tasks.", + "bbox": [ + 75, + 107, + 467, + 212 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The main contributions of this paper are as follows:", + "bbox": [ + 96, + 214, + 436, + 229 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel large-scale dataset NewsNet for long-form video structure understanding. This dataset is derived from $900+$ hours of video and annotated with 4 hierarchical levels of semantics.", + "- NewsNet provides dense annotations and multi-modal information, promoting diverse benchmarks: separate/hierarchical temporal video segmentation in scene/story/topic levels, as well as other common tasks like classification, video localization/grounding, and highlight detection.", + "- We formulate a new benchmark, i.e., hierarchical modeling in the temporal segmentation task, which needs a single model to predict segments of multiple hierarchical levels. Based on the empirical study, we bring insights into how hierarchical modeling potentially benefits the temporal video segmentation task, which was almost never discussed." + ], + "bbox": [ + 96, + 243, + 467, + 529 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 547, + 217, + 561 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1. Related Dataset and Benchmark", + "text_level": 1, + "bbox": [ + 76, + 573, + 361, + 588 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To better introduce our work toward hierarchical temporal segmentation, we first review research related to temporal segmentation and video challenges. As shown in Table 1, extensive benchmarks have been proposed to test the performance of temporal video segmentation. More recently, Shou et al. [51] argues that humans can capture the boundary without predefined target classes, and thus proposed Generic Event Boundary Detection (GEBD) as a new benchmark for detecting generic event boundaries on Kinetics-400 [26]. A few works [25, 33, 43-45, 56] have been proposed to attempt to address this challenge, which is a very active sub-field in long-form video understanding [1, 2]. In this paper, as the NewsNet densely annotates all temporal boundaries, including events like GEBD, scenes like MovieScene [47], and stories like TI-News [35]. NewsNet can test the performance of the existing tasks on almost all temporal localization and segmentation tasks. Moreover, with such hierarchical labels, we can further investigate hierarchical modeling in long-form video understanding, which is an important but rarely explored field.", + "bbox": [ + 75, + 598, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As an overview of the existing benchmarks in NLP and CV, the temporal annotations are labeled in one or two levels, only focusing on the global abstraction or local understanding. However, the video is naturally hierarchical, which can be recursively represented into sub-units. Therefore, the topic-, story-, scene- and event-level tasks should not be drawn into individual lines, but work in a harmonious way. From such a motivation, we contribute a new dataset called NewsNet with hierarchical dense temporal annotations to the community. Since the annotation cost for different groups is quite different, the scale for topic-level benchmarks [4,9,23,26] is generally larger than that of the other datasets [6,7,35,47,48,75]. This paper tries our best to reach a satisfactory trade-off between scale and the quality of the annotations. The proposed dataset does not only cover 1000 videos with 946 hours but also achieves 654.4 event-level annotations per video, which is very competitive over the existing baselines. Moreover, to further generalize the promotion, textual, audio, and visual input are all collected into the NewsNet for the potential application in CV, NLP and Multi-modality.", + "bbox": [ + 496, + 90, + 890, + 407 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Scope of Related Tasks", + "text_level": 1, + "bbox": [ + 500, + 421, + 710, + 436 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Since dense annotations and rich modalities are provided by NewsNet, we can conduct various applications on top of the datasets. Below, we show the parts of video tasks, which can be carried on the proposed dataset.", + "bbox": [ + 496, + 445, + 890, + 505 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Common Video Tasks. The NewsNet is a high-quality video dataset, to evaluate the methods in common video tasks. For instance, NewsNet is an ideal source for video classification [10, 15, 37, 58, 59, 62] (including the out-of-distribution detection settings [49, 61, 72, 73] due to the existence of category 'Others' in Topic level). NewsNet holds annotations in four granularities for temporal segmentation. By training with the mentioned annotation, the method can completely split a video into several short segments, which will serve partial segmentation tasks like video grounding [40, 70, 74], captioning [3, 34, 71], retrieval [5, 16, 31, 52] and highlight detection [17, 18, 20, 28, 32, 54]. Meanwhile, we simultaneously provide aligned audio, text, and visual form in NewsNet, which might promote the improvement on the task of video grounding [30, 53, 66, 69], where the goal is to retrieve the corresponding video clip by adopting an agnostic or specific query as input. We also note that, over the last few years, the 'pre-training' and 'fine-tuning' paradigms [21, 24, 39, 46, 50, 65, 67, 68, 77] have become revolutionary in both NLP and CV communities. We believe that such technology will be further improved by introducing such a high-quality multi-modality video dataset.", + "bbox": [ + 496, + 506, + 890, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Temporal Segmentation Approaches. Towards temporal video segmentation, existing works explore shot, scene and story segmentation tasks, respectively. Shot temporal segmentation [6, 11, 38, 55] is employed as a pre-processing to", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "10671", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/9a691742ff130fad6c75852dd974730328a9a52fb26b33769c9723941dc6fb57.jpg", + "image_caption": [ + "Figure 2. Illustration of the four hierarchies. The illustration shows a hierarchical diagram (left) along with an annotation example (right) for the four hierarchies, including Topic, Story, Scene, and Event. In the given sample, a technology (Topic) news video can be divided into several Stories, such as \"studio\" and \"place\". Taking \"place\" as an example, it can be further characterized into different Scenes, where the backgrounds or places are similar. Based on the changes of shots or persons in a scene, the Scene can be divided into atomic units, i.e., Events." + ], + "image_footnote": [], + "bbox": [ + 122, + 85, + 823, + 258 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "locate the transition positions in videos based on the similarity of the frames. For the scene segmentation, LGSS [47] converts the problem as a binary classification task at the shot level and applies a boundary-based model that aggregates adjacent shot features within a predefined sliding window to make a prediction. While SCRL [65] introduces a boundary-free model in a seq-to-seq way and reduces the inductive bias introduced by the predefined sliding window. As for story temporal segmentation, TI-NEWS [35] proposes a multi-modal framework based on the boundary-free model to complete the story segmentation and classification tasks simultaneously.", + "bbox": [ + 75, + 364, + 472, + 547 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. Dataset Summary", + "text_level": 1, + "bbox": [ + 76, + 560, + 251, + 578 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we will describe some details of the NewsNet. First, we will introduce the definition and taxonomy of the four hierarchical levels of semantics and show some annotated examples of them. Then, the statistical information of the proposed dataset will be given to better understand the data distribution of the NewsNet.", + "bbox": [ + 75, + 585, + 468, + 676 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Definition and Taxonomy", + "text_level": 1, + "bbox": [ + 76, + 686, + 305, + 704 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Existing works focus on the shot- or event-level temporal segmentation, while neglecting the long-horizon cases, such as story- and topic-level segmentation. Hence this paper innovatively contributes more uniform benchmarks to the community, which covers all categories of temporal annotation. In order to clarify the meaning of the different labels, we detail them one by one:", + "bbox": [ + 75, + 710, + 468, + 816 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Event: the atom unit in our setting, which is defined by the switching of the shots or the changing of the person.", + "- Scene: a combination of several successive shots that" + ], + "bbox": [ + 94, + 828, + 468, + 898 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "focus on the same place from different angles. It is defined as a taxonomy-free unit.", + "bbox": [ + 529, + 364, + 890, + 395 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Story: a sequence of scenes about a piece of news broadcast in a studio or outdoors. Story-level segments can be categorized into 'Connection', 'Place', 'Studio', 'Animation', 'Interview', and 'Photo'. A topic is composed of several stories.", + "- Topic: it summarizes the content of a long sequence, with the following keywords: 'Health', 'Politics', 'Entertainment', 'Economy', 'Crime', 'Weather', 'Sport', 'Technology', 'Military' and 'Others'. Generally, an individual news video contains several topics." + ], + "bbox": [ + 517, + 404, + 890, + 565 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The above four different temporal annotations are defined in a hierarchical way. For example, a news video can be divided into several Topic-level segments; a Topic is further composed of several Stories; and so on. The difference between Story and Scene is that Story is category-aware while Scene is not, and the categories in Story are related to the data source. For example, Story contains a category of outdoor place footage, which is unique to news data and usually consists of multiple scenes, as illustrated in Fig. 2.", + "bbox": [ + 496, + 574, + 890, + 710 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Collection and Statistics", + "text_level": 1, + "bbox": [ + 500, + 719, + 720, + 734 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As a large-scale dataset, we collect about 2,000 news broadcast videos from more than 20 different TV stations, of which 1,000 were annotated, and all the data are downloaded from the YouTube and Internet. As shown in Fig. 3, we show the distribution of program duration and sources. News is a kind of typical long-form video, ranging from 10 to 110 minutes. Most of the collected videos have a 1-hour duration, performing as a challenging span for temporal segmentation. During the annotation, we found that", + "bbox": [ + 496, + 742, + 890, + 878 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "Place refers to the news reported outside a studio.", + "bbox": [ + 522, + 887, + 787, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "10672", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a3a23823e90a223ecd619c721749632845ad8bda9746e07f72c52d6adba61058.jpg", + "image_caption": [ + "Figure 3. The statistical information of NewsNet. The left histogram shows the distribution of the Top-20 news programs, (i.e., program A to T, by the total number of videos) in different duration, it can be found that the duration of the same program is relatively close. Besides, the two ring charts on the right show the category proportions of Topic and Story." + ], + "image_footnote": [], + "bbox": [ + 84, + 89, + 883, + 239 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "there exists a long-tailed distribution for the categories of the story- and topic-level annotation, leading to a practical challenge. We believe such a benchmark can simulate the real-world scenario effectively, so that the competing methods could be benchmarked in a practical manner. More details such as annotation process, visualization, etc., can be found in the Supplementary Material.", + "bbox": [ + 75, + 318, + 470, + 424 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Benchmark and Protocol Navigation", + "text_level": 1, + "bbox": [ + 76, + 436, + 406, + 454 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we will first introduce two temporal segmentation task settings, i.e., Separate Modeling and Hierarchical Modeling in NewsNet, with their definitions and protocols. Then, based on the potential relationship between different hierarchies, a simple yet effective loss function is introduced to better tackle hierarchical temporal segmentation. At last, we will briefly introduce the other common video tasks that can be conducted with NewsNet.", + "bbox": [ + 75, + 460, + 468, + 583 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Hierarchical Temporal Segmentation", + "text_level": 1, + "bbox": [ + 76, + 590, + 398, + 606 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As long-form video can be divided into sub-video chunks according to its intrinsic properties, many researchers try to tackle the video segmentation problem from different perspectives in the last few years [12, 35, 47, 65]. Without loss of generality, we classify hierarchical temporal segmentation into two categories: Separate Modeling and Hierarchical Modeling. Separate modeling trains a segmentation model for each level segmentation task, while hierarchical modeling trains a model which simultaneously segments a video into multi-level segments.", + "bbox": [ + 75, + 613, + 468, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "NewsNet supports temporal segmentation tasks with various semantic granularity, i.e., Event, Scene, Story, and Topic. Since many methods [12,47,65], take the Event (aka Shot sometimes) as the basic input unit to reduce the redundancy of the video and computational cost, NewsNet draws the video segmentation tasks on Scene, Story, and Topic. Moreover, these three hierarchies can be strictly aligned according to the Event (Shot) dimension, which is more conducive to Hierarchical Modeling and more convenient for", + "bbox": [ + 75, + 763, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "comparison. Note that we also carried out the event-level experiment (see Supplementary Material).", + "bbox": [ + 496, + 319, + 890, + 348 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Data and Modality. In-domain: all the videos are randomly split into training, validation, and testing sets with a ratio of 3:1:1. Cross-domain: training, validation, and testing sets are split according to different news program types/IDs; 5 different split combinations are set to ensure the generalization ability of the algorithm can be evaluated properly and fairly. As for the embeddings used in the experiment, the shot-level visual, audio, and textual features are extracted through ResNet50 [22] (pre-trained on Places365 [76]), public Bert-base model [14], and Cnn14 [27] (pre-trained on AudioSet [19]), respectively.", + "bbox": [ + 496, + 349, + 892, + 517 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1.1 Separate Modeling", + "text_level": 1, + "bbox": [ + 498, + 532, + 687, + 549 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Task Definition. Given a video with the shot labels, the model requires to classify whether a shot is a segmentation point for one of the three semantic levels (Scene, Story and Topic). A segmentation point refers to a shot that lies at the end of a topic/story/scene segment, and its next shot is the start of a new topic/story/scene.", + "bbox": [ + 496, + 556, + 890, + 647 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Backbone. In general, the common paradigm in those segmentation tasks is considering temporal segment to shot-level classification in boundary-base and -free way. We employ the state-of-the-art boundary-free model, SCRL [65], and the boundary-based model, LGSS [47] as our baseline approaches in separate modeling. More details about the protocol can be found in the Supplementary Material.", + "bbox": [ + 496, + 648, + 890, + 753 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1.2 Hierarchical Modeling", + "text_level": 1, + "bbox": [ + 498, + 771, + 712, + 786 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Task Definition. Given a video with the shot labels, the model requires to classify whether a shot is a segmentation point for three semantic levels (Scene, Story, and Topic).", + "bbox": [ + 496, + 794, + 890, + 840 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Backbone. In this setting, as shown in Fig. 4, we focus on two hierarchical modeling paradigms, including Multi-Label and Multi-Head with a novel learning objective called Hierarchical Ranking. More specifically, Multi", + "bbox": [ + 496, + 840, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "10673", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c5c92b2b846c82cf48f62c9b53e5d63526865ee810ae4ddf00fbe99f7eec5954.jpg", + "image_caption": [ + "Figure 4. An illustration of the separate modeling, two hierarchical modeling paradigms, and the proposed loss function. For the baseline, the model is trained individually corresponding to different segmentation tasks. For hierarchical modeling, all the tasks are aggregated into one model. Specifically, B, L, and N represent the batch size, sequence length, and feature dimension, respectively." + ], + "image_footnote": [], + "bbox": [ + 120, + 88, + 851, + 310 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Label works towards the hierarchical labels from the perspective of multi-label optimization, focusing on the aggregation of information into one head for estimation. While, Multi-Head addresses the challenge as a multi-task problem, where the tasks in different levels should be predicted via the individual pathways. Both pipelines are widely employed in modern architectures [13, 36, 60, 64], performing as simple but representative paradigms. Therefore, we adopt two hierarchical modeling paradigms as the baselines. Hierarchical Ranking Loss. To better model the hierarchical temporal video segmentation task and combine the existing methods, we propose a new loss function called Hierarchical Ranking Loss. The motivation is: due to the accurate alignment between the hierarchical levels with the atom unit Event, as shown in Fig. 1, a positive segment boundary of the higher level task must be a positive segment boundary of the lower task. Hence, the segmentation confidence scores for different tasks should be ranked in a non-increasing order from coarse to fine for the paired task.", + "bbox": [ + 75, + 390, + 472, + 676 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Given two tasks in the hierarchical modeling setting, e.g. task $\\mathcal{H}$ and task $\\mathcal{L}$ , where the semantic level of task $\\mathcal{H}$ is higher than that of task $\\mathcal{L}$ , the proposed hierarchical ranking loss function is as follows:", + "bbox": [ + 75, + 676, + 470, + 731 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {P} _ {\\mathcal {H}, \\mathcal {L}} = \\frac {1}{B} \\frac {1}{L} \\cdot \\sum \\left[ \\sigma \\left(F _ {\\mathcal {H}} - \\mathcal {S G} \\left(F _ {\\mathcal {L}}\\right)\\right) \\cdot Y _ {\\mathcal {L}} \\right] \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 733, + 468, + 760 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $B$ and $L$ are the batch size and sequence length respectively; $\\sigma(\\cdot)$ refers to the contraction mapping, like sigmoid function; $F$ stands for the segmentation confidence score of the task; $\\mathcal{SG}$ is the Stop Gradient operation for the computational graph, and $Y_{\\mathcal{L}}$ corresponds to the label of the task $\\mathcal{L}$ .", + "bbox": [ + 75, + 762, + 468, + 851 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "And the final objective function for all three tasks can be formed as follow:", + "bbox": [ + 76, + 852, + 468, + 878 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {P} _ {\\text {T o p i c}, S t o r y} + \\mathcal {P} _ {\\text {T o p i c}, S c e n e} + \\mathcal {P} _ {\\text {S t o r y}, S c e n e} + \\mathcal {L} _ {C E} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 886, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{CE}$ is the cross-entropy loss function.", + "bbox": [ + 503, + 390, + 808, + 406 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Common Video Understanding Tasks", + "text_level": 1, + "bbox": [ + 500, + 412, + 823, + 430 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The NewsNet could also apply other common video understanding tasks, e.g., video classification, localization, and highlight detection. More details can be found in Supplementary Material.", + "bbox": [ + 496, + 436, + 890, + 497 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiment and Analysis", + "text_level": 1, + "bbox": [ + 500, + 508, + 736, + 527 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we mainly focus on the temporal segmentation tasks on the NewsNet and further contribute several benchmarks on the different baselines. In particular, we conduct the empirical study to thoroughly benchmark the effectiveness of model structure, multi-modality information, domain gap, and hierarchical modeling. Note that, due to the limitation on page length, the implementation details and the results for other common video tasks are given in the Supplementary Material.", + "bbox": [ + 496, + 535, + 890, + 672 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/42e4641d12da66610b52c4013beefb8b393b408f2b358637ee8055da19b9383f.jpg", + "table_caption": [ + "Table 2. The performance between boundary-based (B.B.) and boundary-free (B.F) model with visual input on F1 score $\\uparrow$ , Precision $\\uparrow$ , and Recall $\\uparrow$ . Light gray refers to the failure case." + ], + "table_footnote": [], + "table_body": "
TaskModelF1 scorePrecisionRecall
SceneB.B.77.367.790.2
B.F.76.876.177.5
StoryB.B.70.765.976.2
B.F.71.272.370.0
TopicB.B.13.17.0100.0
B.F.62.972.455.6
", + "bbox": [ + 503, + 741, + 890, + 837 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As an overview, based on the large-scale dataset with rich modalities, we conduct several experiments to respond to five critical questions concerned by the community:", + "bbox": [ + 496, + 854, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "10674", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/0158a464405f42f31307fe8efc1d25aa98b00e59b26f329154dae3d4578a75e7.jpg", + "table_caption": [ + "Table 3. In-domain performance by using boundary-free (B.F.) Table 4. Cross-domain setting by using boundary-free (B.F.) model. The bolded values stand for the optimal performances for model. The bolded values stand for the optimal performances for each task." + ], + "table_footnote": [], + "table_body": "
TaskModalityF1 scorePrecisionRecall
SceneV76.876.177.5
A69.866.873.2
T66.756.381.9
V+A+T78.380.975.8
StoryV71.272.370.0
A59.357.661.1
T50.657.445.2
V+A+T75.474.776.2
TopicV62.972.455.6
A58.159.456.9
T39.046.533.5
V+A+T73.274.372.2
", + "bbox": [ + 84, + 143, + 480, + 325 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Can existing methods also perform well on long-horizon temporal segmentation tasks, such as topic-level and story-level segmentation?", + "2. Do the different-level tasks test the ability of the model on viewing different temporal spans?", + "3. Can existing methods well address the semantic gap between low- and high-level granularities?", + "4. Will the temporal segmentation models be benefited from multi-modality information?", + "5. How could the models benefit from Hierarchical Modeling?" + ], + "bbox": [ + 89, + 340, + 467, + 539 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Results against Increased Temporal Span", + "text_level": 1, + "bbox": [ + 76, + 544, + 429, + 561 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To investigate the generalization of the existing methods against the long-horizon cases, we conduct an experiment on two different head architectures, including boundary-free and boundary-based heads, on three tasks: scene-level, story-level, and topic-level temporal segmentation. As shown in Fig. 2, the performance of the boundary-based model has a significant drop on the Topic task. Compared with the other tasks, the Topic segmentation task has more time slices for input, leading to a hard challenge caused by longer time dependencies. The methods can segment scenes with about 76-77 F1 score while degrading to 62.9 on Topic-level segmentation. For the first question, the answer is that both the boundary-free model and the boundary-based model can not generalize well in the story-and topic-level segmentation. Combining the overall performance and stability on different tasks, the next experiments will utilize a boundary-free model as the basic backbone under comprehensive consideration.", + "bbox": [ + 76, + 566, + 467, + 838 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Performance against Cross News Domains", + "text_level": 1, + "bbox": [ + 76, + 847, + 436, + 864 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To investigate the generalization of the existing method, we analyze the performance of the method against the va", + "bbox": [ + 76, + 869, + 467, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/0f610fa3f1bec2006e0a9c17d635ddd7366326e07e544dd1280f00812dda6bba.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TaskModalityAvg. F1 score (std.)Avg. PrecisionAvg. Recall
SceneV72.9 (2.1)70.975.2
A62.7 (4.0)59.766.6
T61.6 (5.0)52.877.0
V+A+T76.0 (2.1)74.477.9
StoryV68.5 (2.6)70.366.9
A55.7 (3.6)53.659.0
T51.1 (3.6)43.465.4
V+A+T72.9 (2.2)73.772.4
TopicV60.6 (4.7)69.853.8
A59.0 (5.2)56.062.9
T49.8 (5.2)45.755.9
V+A+T72.2 (3.6)72.372.5
", + "bbox": [ + 488, + 143, + 861, + 325 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "rieties of different news programs. As shown in Table 3 and 4, the cross-domain setting slightly degrades the performance on all the tasks. Comparing the in-domain results using the full modality in Table 3 and corresponding cross-domain results in Table 4, the F1 score decreases by 2.3 on task Scene, by 2.5 on task Story, and by 1.0 on task Topic. The reason for this result may be that higher-level tasks, such as Topic, has more abstract semantics and the knowledge is easier to transfer. In contrast, task Scene in the lowest level, generally only considers place switching as the segmentation cue. Besides, the category of Scene is an open set, leading to a greater challenge in the cross-domain setting. Hence, temporal segmentation model also suffers from the out-of-distribution case. Based on the property of the collected dataset, we set the videos from different news programs as different domains, and establish a comprehensive benchmark to facilitate the research on such line.", + "bbox": [ + 496, + 340, + 890, + 595 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Performance using Multi-Modalities", + "text_level": 1, + "bbox": [ + 500, + 604, + 815, + 619 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "When it comes to multi-modality, as shown in Table. 4, the model can be benefited from multimodal information among all the tasks. For example, the performance of using only visual modality achieves only 60.6 recall in topic-level segmentation. However, by aggregating the other modalities including audio and text, the performance can be improved significantly (72.2 vs 49.8), which indicates the potential value for addressing long-discrepancy via multimodality technology. Note that, the improvement in Scene segmentation is, to some extent, limited, we believe that Scene is mainly defined by visual input, hence can not obtain adequate complementarity from other modalities.", + "bbox": [ + 496, + 628, + 890, + 809 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Performance based on Hierarchical Modeling", + "text_level": 1, + "bbox": [ + 500, + 816, + 883, + 834 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Here, we discuss the core challenge revealed in this paper: 'How could the model benefit from hierarchical annotations?' In this case, the baseline is set as the model, which can only be accessible to one specific-level annotation. As", + "bbox": [ + 496, + 840, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "10675", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/783532a966441f3a844ebf146adef59bcd4ad5d64b388806442882a4a9034eae.jpg", + "table_caption": [ + "Table 5. The F1 scores of baselines trained with different levels of annotations on full modalities without our hierarchical ranking loss, where blue and orange indicate the in-domain and cross-domain, respectively. Each row refers to the result corresponding to a single task. Hie. Modeling stands for Hierarchical Modeling while Sep. Modeling is Separate Modeling." + ], + "table_footnote": [], + "table_body": "
RecipeBaseline\nSep. ModelingMulti-Label\nHie. ModelingMulti-Head\nHie. Modeling
Scene78.3 / 76.079.1 / 76.579.9 / 76.9
+ Story75.4 / 72.975.4 / 74.774.2 / 74.0
Scene78.3 / 76.079.8 / 76.479.5 / 76.5
+ Topic73.2 / 72.270.5 / 72.870.9 / 73.0
Story75.4 / 72.976.2 / 74.375.4 / 73.9
+ Topic73.2 / 72.277.3 / 73.275.2 / 73.5
Scene78.3 / 76.077.4 / 76.879.8 / 76.8
+ Story75.4 / 72.974.3 / 74.374.5 / 73.7
+ Topic73.2 / 72.274.3 / 72.676.6 / 70.4
", + "bbox": [ + 81, + 184, + 468, + 369 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "listed in Table 5, we consider two simple strategies, including Multi-Label and Multi-Head, to aggregate the hierarchical information. By training with multi-level annotations, a $1.0 \\sim 3.0\\%$ improvement can be achieved, which indicates, the segmentation can be benefited by introducing the hierarchical structures of videos. However, Table 5 also shows directly combining different-level granularities might not result in gains due to the semantic gap between low- and high-level granularities. For example, directly combining scene- and topic-level information by Multi-Label hierarchical modeling or Multi-Head hierarchical modeling does not consistently improve the topic-level segmentation performance. This poses a new and challenging research question, i.e. how to jointly train with various granularities toward hierarchical modeling of long-form videos?", + "bbox": [ + 75, + 402, + 468, + 628 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To address the above question, we introduce a simple but effective loss function, i.e., Hierarchical Ranking Loss introduced in Sec. 4.1.2. As shown in Table 6, by combining with our proposed loss, the F1 score for segmentation can be further improved in most cases. Typically, in Topic-level, Multi-Head coupled with the proposed loss can reach the best performance, i.e., improving Multi-Head case from 70.4 to 73.2 F1 score.", + "bbox": [ + 75, + 637, + 468, + 758 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.5. Other Benchmarks involved in the NewsNet", + "text_level": 1, + "bbox": [ + 76, + 792, + 449, + 809 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In addition to the temporal segmentation tasks, we also conduct extensive common video tasks, including highlight detection, video classification, and video localization, on the proposed dataset. Due to the limited pages, we show this part in the Supplementary Materials.", + "bbox": [ + 75, + 825, + 468, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/14668eeda4db6d409f3ec78f89a265cc0fdc1bf9de6ba134e55840c9a619e32a.jpg", + "table_caption": [ + "Table 6. The F1 scores of the methods with or without hierarchical ranking loss under the in-domain / cross-domain setting on full modalities. Hie. stands for Hierarchical Modeling and Sep. refers to Separate Modeling." + ], + "table_footnote": [], + "table_body": "
MethodSceneStoryTopic
Baseline (Sep.)78.3 / 76.075.4 / 72.973.2 / 72.2
Multi-Label (Hie.)77.4 / 76.874.3 / 74.374.3 / 72.6
Multi-Label w/ Hie. Loss (Hie.)79.6 / 76.974.4 / 73.577.8 / 73.1
Multi-Head (Hie.)79.8 / 76.874.5 / 73.776.6 / 70.4
Multi-Head w/ Hie. Loss (Hie.)80.3 / 76.976.3 / 74.676.5 / 73.2
", + "bbox": [ + 504, + 156, + 890, + 273 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion, Limitation and Future work", + "text_level": 1, + "bbox": [ + 500, + 296, + 864, + 313 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduce NewsNet, a new large-scale dataset for temporal video segmentation. Compared with existing large-scale datasets for temporal video segmentation, NewsNet additionally provides two abstractive levels of temporal segmentation, which have not been taken into account by these datasets. The four-level and hierarchical annotations enable the community to explore how to comprehensively represent the complex structure of long-form videos from coarse to fine. Extensive efforts of human workers are devoted, so as to ensure that annotations are high-quality. We hope that NewsNet and baseline results can facilitate the development of temporal video segmentation in terms of insightful research and practical tools. In addition, as NewsNet provides diverse annotations for dividing a video into segments of different granularities as well as category labels, it can also serve other video understanding tasks such as video grounding, captioning, retrieval, highlight detection, and video classification.", + "bbox": [ + 496, + 325, + 890, + 598 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we only focus on the problem of hierarchical temporal segmentation. We have not explored other multi-modality tasks, such as visual reasoning and video question answering, while our dataset can serve these tasks thanks to its multi-modality nature. We will explore these tasks in our future work to further promote advances in multi-modality-based video understanding and reasoning.", + "bbox": [ + 496, + 601, + 890, + 708 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 732, + 660, + 750 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported by the KAUST Office of Sponsored Research through the Visual Computing Center (VCC) funding, as well as, the SDAIA-KAUST Center of Excellence in Data Science and Artificial Intelligence (SDAIA-KAUST AI). Raghavendra Ramachandra is supported under the SALT and OFFPAD projects funded by the Research Council of Norway, Norway. Chia-Wen Lin is supported by National Science and Technology Council, Taiwan, under project NSTC 111-2634-F002-023.", + "bbox": [ + 496, + 760, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "10676", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] LOVEU21. https://sites.google.com/view/loveucvpr21, 2021.3", + "[2] LOVEU22 program schedule. https://sites.google.com/view/loveucvpr22/program, 2022.3", + "[3] Nayyer Aafaq, Naveed Akhtar, Wei Liu, Syed Zulqarnain Gilani, and Ajmal Mian. Spatio-temporal dynamics and semantic attribute enriched visual encoding for video captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12487-12496, 2019. 2, 3", + "[4] Sami Abu-El-Haija, Nisarg Kothari, Joonseok Lee, Paul Natsev, George Toderici, Balakrishnan Varadarajan, and Sudheendra Vijayanarasimhan. Youtube-8m: A large-scale video classification benchmark. arXiv preprint arXiv:1609.08675, 2016. 2, 3", + "[5] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1728-1738, 2021. 2, 3", + "[6] Lorenzo Baraldi, Costantino Grana, and Rita Cucchiara. A deep siamese network for scene detection in broadcast videos. In Proceedings of the 23rd ACM international conference on Multimedia, pages 1199-1202, 2015. 2, 3", + "[7] Lorenzo Baraldi, Costantino Grana, and Rita Cucchiara. Shot and scene detection via hierarchical clustering for re-using broadcast video. In International conference on computer analysis of images and patterns, pages 801-811. Springer, 2015. 2, 3", + "[8] Joe Barrow, Rajiv Jain, Vlad Morariu, Varun Manjunatha, Douglas W Oard, and Philip Resnik. A joint model for document segmentation and segment labeling. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 313-322, 2020. 2", + "[9] Fabian Caba Heilbron, Victor Escorcia, Bernard Ghanem, and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 961-970, 2015. 3", + "[10] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 1, 3", + "[11] Vasileios T Chasanis, Aristidis C Likas, and Nikolaos P Galatsanos. Scene detection in videos using shot clustering and sequence alignment. IEEE transactions on multimedia, 11(1):89-100, 2008. 3", + "[12] Shixing Chen, Xiaohan Nie, David Fan, Dongqing Zhang, Vimal Bhat, and Raffay Hamid. Shot contrastive self-supervised learning for scene boundary detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9796-9805, 2021. 2, 5", + "[13] Zhao-Min Chen, Xiu-Shen Wei, Peng Wang, and Yanwen Guo. Multi-label image recognition with graph convolutional networks. In Proceedings of the IEEE/CVF Confer-" + ], + "bbox": [ + 78, + 114, + 467, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ence on Computer Vision and Pattern Recognition (CVPR), June 2019. 6", + "[14] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 5", + "[15] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6202-6211, 2019. 1, 3", + "[16] Valentin Gabeur, Chen Sun, Karteek Alahari, and Cordelia Schmid. Multi-modal transformer for video retrieval. In European Conference on Computer Vision, pages 214-229. Springer, 2020. 2, 3", + "[17] Jiyang Gao, Chen Sun, Zhenheng Yang, and Ram Nevatia. Tall: Temporal activity localization via language query. In Proceedings of the IEEE international conference on computer vision, pages 5267-5275, 2017. 3", + "[18] Ana Garcia del Molino and Michael Gygli. PhD-gifs: personalized highlight detection for automatic gif creation. In Proceedings of the 26th ACM international conference on Multimedia, pages 600–608, 2018. 3", + "[19] Jort F Gemmeke, Daniel PW Ellis, Dylan Freedman, Aren Jansen, Wade Lawrence, R Channing Moore, Manoj Plakal, and Marvin Ritter. Audio set: An ontology and human-labeled dataset for audio events. In 2017 IEEE international conference on acoustics, speech and signal processing (ICASSP), pages 776-780. IEEE, 2017. 5", + "[20] Michael Gygli, Yale Song, and Liangliang Cao. Video2gif: Automatic generation of animated gifs from video. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1001-1009, 2016. 3", + "[21] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 3", + "[22] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5", + "[23] Qingqiu Huang, Yu Xiong, Anyi Rao, Jiaze Wang, and Dahua Lin. Movienet: A holistic dataset for movie understanding. In European Conference on Computer Vision, pages 709-727. Springer, 2020. 2, 3", + "[24] Chen Ju, Tengda Han, Kunhao Zheng, Ya Zhang, and Weidi Xie. Prompting visual-language models for efficient video understanding. arXiv preprint arXiv:2112.04478, 2021. 3", + "[25] Hyolim Kang, Jinwoo Kim, Taehyun Kim, and Seon Joo Kim. Uboco: Unsupervised boundary contrastive learning for generic event boundary detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20073-20082, 2022. 2, 3", + "[26] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 2, 3" + ], + "bbox": [ + 503, + 93, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "10677", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Qiuqiang Kong, Yin Cao, Turab Iqbal, Yuxuan Wang, Wenwu Wang, and Mark D Plumbley. Panns: Large-scale pretrained audio neural networks for audio pattern recognition. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 28:2880-2894, 2020. 5", + "[28] Ranjay Krishna, Kenji Hata, Frederic Ren, Li Fei-Fei, and Juan Carlos Niebles. Dense-captioning events in videos. In Proceedings of the IEEE international conference on computer vision, pages 706–715, 2017. 3", + "[29] Kathy Lee, Diana Palsetia, Ramanathan Narayanan, Md Mostofa Ali Patwary, Ankit Agrawal, and Alok Choudhary. Twitter trending topic classification. In 2011 IEEE 11th International Conference on Data Mining Workshops, pages 251-258. IEEE, 2011. 2", + "[30] Jie Lei, Tamara L Berg, and Mohit Bansal. Detecting moments and highlights in videos via natural language queries. Advances in Neural Information Processing Systems, 34:11846-11858, 2021. 3", + "[31] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L Berg, Mohit Bansal, and Jingjing Liu. Less is more: Clipbert for video-and-language learning via sparse sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7331-7341, 2021. 2, 3", + "[32] Jie Lei, Licheng Yu, Tamara L Berg, and Mohit Bansal. Tvr: A large-scale dataset for video-subtitle moment retrieval. In European Conference on Computer Vision, pages 447-463. Springer, 2020. 3", + "[33] Congcong Li, Xinyao Wang, Longyin Wen, Dxiang Hong, Tiejian Luo, and Libo Zhang. End-to-end compressed video representation learning for generic event boundary detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13967-13976, 2022. 3", + "[34] Kevin Lin, Linjie Li, Chung-Ching Lin, Faisal Ahmed, Zhe Gan, Zicheng Liu, Yumao Lu, and Lijuan Wang. Swinbert: End-to-end transformers with sparse attention for video captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17949-17958, 2022. 2, 3", + "[35] Ye Liu, Lingfeng Qiao, Di Yin, Zhuoxuan Jiang, Xinghua Jiang, Deqiang Jiang, and Bo Ren. Os-msl: One stage multimodal sequential link framework for scene segmentation and classification. In Proceedings of the 30th ACM International Conference on Multimedia, pages 6269-6277, 2022. 2, 3, 4, 5", + "[36] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10012-10022, October 2021. 6", + "[37] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swim transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3202-3211, 2022. 1, 3", + "[38] Jakub Lokoc, Gregor Kovalcik, Tomáš Souček, Jaroslav Moravec, and Přemysl Čech. A framework for effective" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "known-item search in video. In Proceedings of the 27th ACM International Conference on Multimedia, pages 1777-1785, 2019. 2, 3", + "[39] Sauradip Nag, Xiatian Zhu, Yi-Zhe Song, and Tao Xiang. Zero-shot temporal action detection via vision-language prompting. arXiv preprint arXiv:2207.08184, 2022. 3", + "[40] Guoshun Nan, Rui Qiao, Yao Xiao, Jun Liu, Sicong Leng, Hao Zhang, and Wei Lu. Interventional video grounding with dual contrastive learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2765-2775, 2021. 3", + "[41] Fahmi Salman Nurfikri, Mohamad Syahrul Mubarok, et al. News topic classification using mutual information and bayesian network. In 2018 6th International Conference on Information and Communication Technology (ICoICT), pages 162-166. IEEE, 2018. 2", + "[42] Sofia Ares Oliveira, Benoit Seguin, and Frederic Kaplan. dhsegment: A generic deep-learning approach for document segmentation. In 2018 16th International Conference on Frontiers in Handwriting Recognition (ICFHR), pages 7-12. IEEE, 2018. 2", + "[43] Alejandro Pardo, Fabian Caba Heilbron, Juan León Alcazar, Ali Thabet, and Bernard Ghanem. Moviecuts: A new dataset and benchmark for cut type recognition. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part VII, pages 668-685, 2022. 3", + "[44] Will Price, Carl Vondrick, and Dima Damen. Unweavenet: Unweaving activity stories. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13770-13779, 2022. 3", + "[45] Zhiwu Qing, Shiwei Zhang, Ziyuan Huang, Yi Xu, Xiang Wang, Mingqian Tang, Changxin Gao, Rong Jin, and Nong Sang. Learning from untrimmed videos: Self-supervised video representation learning with hierarchical consistency. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13821-13831, 2022. 3", + "[46] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 3", + "[47] Anyi Rao, Linning Xu, Yu Xiong, Guodong Xu, Qingqiu Huang, Bolei Zhou, and Dahua Lin. A local-to-global approach to multi-modal movie scene segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10146-10155, 2020. 2, 3, 4, 5", + "[48] Daniel Rotman, Dror Porat, and Gal Ashour. Optimal sequential grouping for robust video scene detection using multiple modalities. International Journal of Semantic Computing, 11(02):193-208, 2017. 2, 3", + "[49] Ellen Rushe and Brian Mac Namee. Anomaly detection in raw audio using deep autoregressive networks. In ICASSP 2019-2019 IEEE International Conference on Acoustics," + ], + "bbox": [ + 503, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10678", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Speech and Signal Processing (ICASSP), pages 3597-3601. IEEE, 2019. 3", + "[50] Weimin Shi, Mingchen Zhuge, Zhong Zhou, Dehong Gao, and Deng-Ping Fan. Qr-clip: Introducing explicit open-world knowledge for location and time reasoning. arXiv preprint arXiv:2302.00952, 2023. 3", + "[51] Mike Zheng Shou, Stan Weixian Lei, Weiyao Wang, Deepti Ghadiyaram, and Matt Feiszli. Generic event boundary detection: A benchmark for event segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8075-8084, 2021. 2, 3", + "[52] Nina Shvetsova, Brian Chen, Andrew Rouditchenko, Samuel Thomas, Brian Kingsbury, Rogerio S Feris, David Harwath, James Glass, and Hilde Kuehne. Everything at once-multimodal fusion transformer for video retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20020-20029, 2022. 2, 3", + "[53] Mattia Soldan, Alejandro Pardo, Juan León Alcazar, Fabian Caba, Chen Zhao, Silvio Giancola, and Bernard Ghanem. Mad: A scalable dataset for language grounding in videos from movie audio descriptions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5026-5035, 2022. 3", + "[54] Yale Song, Miriam Redi, Jordi Vallmitjana, and Alejandro Jaimes. To click or not to click: Automatic selection of beautiful thumbnails from videos. In Proceedings of the 25th ACM international on conference on information and knowledge management, pages 659-668, 2016. 3", + "[55] Tomáš Souček and Jakub Lokoc. Transnet v2: an effective deep network architecture for fast shot transition detection. arXiv preprint arXiv:2008.04838, 2020. 2, 3", + "[56] Jiaqi Tang, Zhaoyang Liu, Chen Qian, Wayne Wu, and Limin Wang. Progressive attention on multi-level dense difference maps for generic event boundary detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3355-3364, 2022. 2, 3", + "[57] Shitao Tang, Litong Feng, Zhanghui Kuang, Yimin Chen, and Wei Zhang. Fast video shot transition localization with deep structured models. In Asian Conference on Computer Vision, pages 577-592. Springer, 2018. 2", + "[58] Du Tran, Lubomir Bourdev, Rob Fergus, Lorenzo Torresani, and Manohar Paluri. Learning spatiotemporal features with 3d convolutional networks. In Proceedings of the IEEE international conference on computer vision, pages 4489-4497, 2015. 1, 3", + "[59] Du Tran, Heng Wang, Lorenzo Torresani, and Matt Feiszli. Video classification with channel-separated convolutional networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5552-5561, 2019. 1, 3", + "[60] Dongkai Wang and Shiliang Zhang. Unsupervised person re-identification via multi-label classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 6", + "[61] Jue Wang and Anoop Cherian. Gods: Generalized one-class discriminative subspaces for anomaly detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8201-8211, 2019. 3" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[62] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaou Tang, and Luc Van Gool. Temporal segment networks: Towards good practices for deep action recognition. In European conference on computer vision, pages 20-36. Springer, 2016. 1, 3", + "[63] Yuxuan Wang, Difei Gao, Licheng Yu, Weixian Lei, Matt Feiszli, and Mike Zheng Shou. Geb+: A benchmark for generic event boundary captioning, grounding and retrieval. In European Conference on Computer Vision, pages 709-725. Springer, 2022. 2", + "[64] Jonatas Wehrmann, Ricardo Cerri, and Rodrigo Barros. Hierarchical multi-label classification networks. In Jennifer Dy and Andreas Krause, editors, Proceedings of the 35th International Conference on Machine Learning, volume 80 of Proceedings of Machine Learning Research, pages 5075-5084. PMLR, 10-15 Jul 2018. 6", + "[65] Haoqian Wu, Keyu Chen, Yanan Luo, Ruizhi Qiao, Bo Ren, Haozhe Liu, Weicheng Xie, and Linlin Shen. Scene consistency representation learning for video scene segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14021-14030, 2022. 2, 3, 4, 5", + "[66] Bo Xiong, Yannis Kalantidis, Deepti Ghadiyaram, and Kristen Grauman. Less is more: Learning highlight detection from video duration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1258-1267, 2019. 3", + "[67] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. Videoclip: Contrastive pre-training for zero-shot video-text understanding. arXiv preprint arXiv:2109.14084, 2021. 3", + "[68] Mengmeng Xu, Juan-Manuel Pérez-Rúa, Victor Escorcia, Brais Martinez, Xiatian Zhu, Li Zhang, Bernard Ghanem, and Tao Xiang. Boundary-sensitive pre-training for temporal localization in videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7220-7230, 2021. 3", + "[69] Minghao Xu, Hang Wang, Bingbing Ni, Riheng Zhu, Zhenbang Sun, and Changhu Wang. Cross-category video highlight detection via set-based learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7970-7979, 2021. 3", + "[70] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Tubedetr: Spatio-temporal video grounding with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16442-16453, 2022. 3", + "[71] Hanhua Ye, Guorong Li, Yuankai Qi, Shuhui Wang, Qingming Huang, and Ming-Hsuan Yang. Hierarchical modular network for video captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17939-17948, 2022. 2, 3", + "[72] Muhammad Zaigham Zaheer, Arif Mahmood, Marcella Astrid, and Seung-Ik Lee. Claws: Clustering assisted weakly supervised learning with normalcy suppression for anomalous event detection. In European Conference on Computer Vision, pages 358-376. Springer, 2020. 3" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "10679", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[73] M Zaigham Zaheer, Arif Mahmood, M Haris Khan, Mattia Segu, Fisher Yu, and Seung-Ik Lee. Generative cooperative learning for unsupervised video anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14744-14754, 2022. 3", + "[74] Runhao Zeng, Haoming Xu, Wenbing Huang, Peihao Chen, Mingkui Tan, and Chuang Gan. Dense regression network for video grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10287-10296, 2020. 3", + "[75] Haoxin Zhang, Zhimin Li, and Qinglin Lu. Better learning shot boundary detection via multi-task. In Proceedings of the 29th ACM International Conference on Multimedia, pages 4730-4734, 2021. 2, 3", + "[76] Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. IEEE transactions on pattern analysis and machine intelligence, 40(6):1452-1464, 2017. 5", + "[77] Mingchen Zhuge, Dehong Gao, Deng-Ping Fan, Linbo Jin, Ben Chen, Haoming Zhou, Minghui Qiu, and Ling Shao. Kaleido-bert: Vision-language pre-training on fashion domain. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12647-12657, 2021. 3" + ], + "bbox": [ + 78, + 90, + 468, + 428 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "10680", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2023/NewsNet_ A Novel Dataset for Hierarchical Temporal Segmentation/d2756edd-2b85-4c25-85cd-99de3bdf0d9a_model.json b/2023/NewsNet_ A Novel Dataset for Hierarchical Temporal Segmentation/d2756edd-2b85-4c25-85cd-99de3bdf0d9a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..bde49c61422361f2b2d95731b8fa94348d3a382c --- /dev/null +++ b/2023/NewsNet_ A Novel Dataset for Hierarchical Temporal Segmentation/d2756edd-2b85-4c25-85cd-99de3bdf0d9a_model.json @@ -0,0 +1,2446 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.048 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.143, + 0.131, + 0.828, + 0.154 + ], + "angle": 0, + "content": "NewsNet: A Novel Dataset for Hierarchical Temporal Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.086, + 0.18, + 0.882, + 0.234 + ], + "angle": 0, + "content": "Haoqian \\(\\mathbf{W}\\mathbf{u}^{1\\dagger}\\), Keyu Chen\\(^{1\\dagger}\\), Haozhe Liu\\(^{2\\dagger}\\), Mingchen Zhuge\\(^{2\\dagger}\\), Bing Li\\(^{2\\boxtimes}\\), Ruizhi Qiao\\(^{1\\boxtimes}\\), Xiujun Shu\\(^{1}\\), Bei Gan\\(^{1}\\), Liangsheng Xu\\(^{1}\\), Bo Ren\\(^{1}\\), Mengmeng Xu\\(^{2}\\), Wentian Zhang\\(^{2}\\), Raghavendra Ramachandra\\(^{3}\\), Chia-Wen Lin\\(^{4}\\), Bernard Ghanem\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.133, + 0.24, + 0.837, + 0.296 + ], + "angle": 0, + "content": "\\(^{1}\\) Tencent \\(^{2}\\) AI Initiative, King Abdullah University of Science and Technology (KAUST) \n\\(^{3}\\) Norwegian University of Science and Technology (NTNU) \n\\(^{4}\\) National Tsing Hua University (NTHU)" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.309, + 0.885, + 0.535 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.545, + 0.893, + 0.616 + ], + "angle": 0, + "content": "Figure 1. Examples from the proposed NewsNet. The dataset provides multimodal and hierarchical labels on each video, including 4-level hierarchical temporal information (i.e., event, scene, story, topic) and 3 modality-specific cues (i.e., video, audio, text). A prerequisite motivation behind NewsNet is that long-form videos can be recursively represented into several sub-videos according to the semantics, which is an intrinsic property of the video but rarely explored in the community. To bridge this research gap, this paper innovatively constructs a large-scale dataset to benchmark the algorithms for hierarchical understanding like a human being." + }, + { + "type": "title", + "bbox": [ + 0.268, + 0.63, + 0.348, + 0.646 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.664, + 0.473, + 0.846 + ], + "angle": 0, + "content": "Temporal video segmentation is the get-to-go automatic video analysis, which decomposes a long-form video into smaller components for the following-up understanding tasks. Recent works have studied several levels of granularity to segment a video, such as shot, event, and scene. Those segmentations can help compare the semantics in the corresponding scales, but lack a wider view of larger temporal spans, especially when the video is complex and structured. Therefore, we present two abstractive levels of temporal segmentations and study their hierarchy to the existing fine-grained levels. Accordingly, we collect NewsNet, the largest news video dataset consisting of 1,000 videos" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.631, + 0.894, + 0.797 + ], + "angle": 0, + "content": "in over 900 hours, associated with several tasks for hierarchical temporal video segmentation. Each news video is a collection of stories on different topics, represented as aligned audio, visual, and textual data, along with extensive frame-wise annotations in four granularities. We assert that the study on NewsNet can advance the understanding of complex structured video and benefit more areas such as short-video creation, personalized advertisement, digital instruction, and education. Our dataset and code is publicly available at https://github.com/NewsNet-Benchmark/NewsNet." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.829, + 0.633, + 0.845 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Temporal video segmentation is a critical problem in video understanding, which is essential for many video applications such as video classification [10, 15, 37, 58, 59, 62]," + }, + { + "type": "page_footnote", + "bbox": [ + 0.102, + 0.864, + 0.216, + 0.876 + ], + "angle": 0, + "content": "† Equal Contribution" + }, + { + "type": "page_footnote", + "bbox": [ + 0.08, + 0.877, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Corresponding Authors: Bing Li (bing.li@kaust.edu.sa) and Ruizhi Qiao (ruizhiqiao@tencent.com)." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.864, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "10669" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.894, + 0.12 + ], + "angle": 0, + "content": "Table 1. Data comparison between the NewsNet and other related datasets. The NewsNet provides various multimodal data and hierarchical temporal segmentation annotations. Doc: documentary, Ads: advertisement. (Please refer to our project page for more details.)" + }, + { + "type": "table", + "bbox": [ + 0.111, + 0.13, + 0.861, + 0.296 + ], + "angle": 0, + "content": "
Dataset# VideoDuration (hours)Modality# Annotation(s) per VideoSource
TopicStorySceneEvent
AVS [75]197-Visual---14.2Ads
BBC [6]119Visual--49.7-Doc
OVSD [48]2110Visual--28.9-Generic
Kinetics-GEBD [51]54,691152Audio + Visual---4.9Action
MovieNet [23] †1,1002174Text + Audio + Visual--66.0849.1Movie
RAI [7]10-Visual---98.7News
TI-News [35]477244Audio + Visual-55.6-530.4News
NewsNet (Ours)1,000946Text + Audio + Visual8.551.687.9654.4News
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.103, + 0.297, + 0.716, + 0.31 + ], + "angle": 0, + "content": "The number of annotations for Scene and Shot is counted from MovieScene [47], which is a subset of MovieNet." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.334, + 0.473, + 0.487 + ], + "angle": 0, + "content": "captioning [3, 34, 63, 71] and retrieval [5, 16, 31, 52]. Temporal video segmentation aims to group successive video frames into short segments along the temporal dimension. With the explosive growth of long-form videos, it is desirable that temporal video segmentation can convert a video into more meaningful segments for more efficient access to the video. However, it is challenging to develop effective temporal segmentation tools for long-form videos, since this requires a comprehensive understanding of video structure, while long-form videos contain complex content." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.489, + 0.473, + 0.793 + ], + "angle": 0, + "content": "Towards temporal video segmentation, existing works explore shot, event, and scene segmentation tasks, respectively. Shot segmentation [38, 55, 57] divides a video into shots, where a shot consists of consecutive and visually continuous frames captured by a camera without interruption [23]. Yet, shot segmentation only considers low-level visual cues (i.e. visual similarity), lacking semantic understanding. Instead, event segmentation [25, 51, 56] divides a video by detecting the moments of changes such as action/subject changes. To better capture the underlying structure of a video, recent works [12, 47, 65] introduce video scene segmentation which segments a video into scene segments, each comprising successive shots semantically related to the same scene. Scene segmentation enables a coarser and higher-level representation than shot segmentation. However, compared to the rich content of massive long-form videos, scene/event is fine-grained and often lacks a high-level summarization of video content, which is insufficient for capturing the complex semantic structure of many videos and briefly representing video content." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.472, + 0.904 + ], + "angle": 0, + "content": "In this work, we first explore how to comprehensively represent the complex structure of a long-form video for temporal video segmentation. Humans can hierarchically divide a video into segments of different granularities according to multi-level semantic information (e.g. scene and topic), from the perspective of cognitive science. Natural language processing researchers have widely explored" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.334, + 0.895, + 0.548 + ], + "angle": 0, + "content": "topic-level understanding [29, 41] for summarizing documents [8, 42], while little effort has been devoted to long-form videos. Inspired by these observations, besides scene and event, we propose to introduce two higher-level semantics (i.e. story and topic) into temporal video segmentation, to provide a brief and semantic structure representation. As a result, such hierarchical and multi-level understanding brings about scalable video structure representation for temporal video segmentation on long-form videos. That is, a long-form video can be split into finer segments with lower-level semantics (e.g. scene), but also can be summarized into coarser ones yet with higher-level semantics (e.g. topic) by recursively grouping finer segments, which comprehensively represents video structures from coarse to fine." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.55, + 0.895, + 0.642 + ], + "angle": 0, + "content": "However, the community lacks high-quality datasets to conduct this research. In particular, as shown in Table 1, most datasets only provide temporal structure annotations with regard to events or scenes. TI-News [35] and MovieScene [47] provide two levels of annotations, but these datasets lack topic-level ones." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.896, + 0.903 + ], + "angle": 0, + "content": "To effectively break this limitation, we build a novel large-scale dataset for hierarchical temporal segmentation, named NewsNet. The unique properties of our NewsNet introduce many advantages. First, it is among the largest datasets in the news domain. We collect over 900 hours of videos from 20 mainstream news platforms. It has a highly diverse distribution of data. Second, we carefully annotated it frame-by-frame with 4 hierarchical levels to ensure its quality can meet our needs. Third, it is multimodal, including textual, visual, and audio information. Due to the nature of the news, the alignment across modalities is accurate, which makes multimodal joint training of models feasible. Finally, the videos in NewsNet provide a complete understanding of public events. Compared with other video datasets [4,23,26], it introduces more objective open-world knowledge (e.g., news introduction) while including subjective factual commentary (e.g., host comments on news" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "10670" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.455, + 0.106 + ], + "angle": 0, + "content": "events), making it more amenable to real-life application." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.108, + 0.468, + 0.213 + ], + "angle": 0, + "content": "Based on NewsNet, we empirically highlight two promising directions for long-span temporal segmentation: 1) Infusing Multi-Modality knowledge can significantly improve the performance of long-form temporal segmentation; 2) Although story- and topic-level segmentation is challenging, it can be benefited from hierarchical modeling with the event- and scene-level segmentation tasks." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.215, + 0.437, + 0.23 + ], + "angle": 0, + "content": "The main contributions of this paper are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.244, + 0.468, + 0.303 + ], + "angle": 0, + "content": "- We propose a novel large-scale dataset NewsNet for long-form video structure understanding. This dataset is derived from \\(900+\\) hours of video and annotated with 4 hierarchical levels of semantics." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.319, + 0.468, + 0.41 + ], + "angle": 0, + "content": "- NewsNet provides dense annotations and multi-modal information, promoting diverse benchmarks: separate/hierarchical temporal video segmentation in scene/story/topic levels, as well as other common tasks like classification, video localization/grounding, and highlight detection." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.425, + 0.468, + 0.53 + ], + "angle": 0, + "content": "- We formulate a new benchmark, i.e., hierarchical modeling in the temporal segmentation task, which needs a single model to predict segments of multiple hierarchical levels. Based on the empirical study, we bring insights into how hierarchical modeling potentially benefits the temporal video segmentation task, which was almost never discussed." + }, + { + "type": "list", + "bbox": [ + 0.097, + 0.244, + 0.468, + 0.53 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.549, + 0.218, + 0.563 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.574, + 0.362, + 0.589 + ], + "angle": 0, + "content": "2.1. Related Dataset and Benchmark" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.47, + 0.901 + ], + "angle": 0, + "content": "To better introduce our work toward hierarchical temporal segmentation, we first review research related to temporal segmentation and video challenges. As shown in Table 1, extensive benchmarks have been proposed to test the performance of temporal video segmentation. More recently, Shou et al. [51] argues that humans can capture the boundary without predefined target classes, and thus proposed Generic Event Boundary Detection (GEBD) as a new benchmark for detecting generic event boundaries on Kinetics-400 [26]. A few works [25, 33, 43-45, 56] have been proposed to attempt to address this challenge, which is a very active sub-field in long-form video understanding [1, 2]. In this paper, as the NewsNet densely annotates all temporal boundaries, including events like GEBD, scenes like MovieScene [47], and stories like TI-News [35]. NewsNet can test the performance of the existing tasks on almost all temporal localization and segmentation tasks. Moreover, with such hierarchical labels, we can further investigate hierarchical modeling in long-form video understanding, which is an important but rarely explored field." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.409 + ], + "angle": 0, + "content": "As an overview of the existing benchmarks in NLP and CV, the temporal annotations are labeled in one or two levels, only focusing on the global abstraction or local understanding. However, the video is naturally hierarchical, which can be recursively represented into sub-units. Therefore, the topic-, story-, scene- and event-level tasks should not be drawn into individual lines, but work in a harmonious way. From such a motivation, we contribute a new dataset called NewsNet with hierarchical dense temporal annotations to the community. Since the annotation cost for different groups is quite different, the scale for topic-level benchmarks [4,9,23,26] is generally larger than that of the other datasets [6,7,35,47,48,75]. This paper tries our best to reach a satisfactory trade-off between scale and the quality of the annotations. The proposed dataset does not only cover 1000 videos with 946 hours but also achieves 654.4 event-level annotations per video, which is very competitive over the existing baselines. Moreover, to further generalize the promotion, textual, audio, and visual input are all collected into the NewsNet for the potential application in CV, NLP and Multi-modality." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.422, + 0.712, + 0.437 + ], + "angle": 0, + "content": "2.2. Scope of Related Tasks" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.446, + 0.892, + 0.506 + ], + "angle": 0, + "content": "Since dense annotations and rich modalities are provided by NewsNet, we can conduct various applications on top of the datasets. Below, we show the parts of video tasks, which can be carried on the proposed dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.507, + 0.892, + 0.84 + ], + "angle": 0, + "content": "Common Video Tasks. The NewsNet is a high-quality video dataset, to evaluate the methods in common video tasks. For instance, NewsNet is an ideal source for video classification [10, 15, 37, 58, 59, 62] (including the out-of-distribution detection settings [49, 61, 72, 73] due to the existence of category 'Others' in Topic level). NewsNet holds annotations in four granularities for temporal segmentation. By training with the mentioned annotation, the method can completely split a video into several short segments, which will serve partial segmentation tasks like video grounding [40, 70, 74], captioning [3, 34, 71], retrieval [5, 16, 31, 52] and highlight detection [17, 18, 20, 28, 32, 54]. Meanwhile, we simultaneously provide aligned audio, text, and visual form in NewsNet, which might promote the improvement on the task of video grounding [30, 53, 66, 69], where the goal is to retrieve the corresponding video clip by adopting an agnostic or specific query as input. We also note that, over the last few years, the 'pre-training' and 'fine-tuning' paradigms [21, 24, 39, 46, 50, 65, 67, 68, 77] have become revolutionary in both NLP and CV communities. We believe that such technology will be further improved by introducing such a high-quality multi-modality video dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Temporal Segmentation Approaches. Towards temporal video segmentation, existing works explore shot, scene and story segmentation tasks, respectively. Shot temporal segmentation [6, 11, 38, 55] is employed as a pre-processing to" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "10671" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.123, + 0.087, + 0.825, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.269, + 0.892, + 0.337 + ], + "angle": 0, + "content": "Figure 2. Illustration of the four hierarchies. The illustration shows a hierarchical diagram (left) along with an annotation example (right) for the four hierarchies, including Topic, Story, Scene, and Event. In the given sample, a technology (Topic) news video can be divided into several Stories, such as \"studio\" and \"place\". Taking \"place\" as an example, it can be further characterized into different Scenes, where the backgrounds or places are similar. Based on the changes of shots or persons in a scene, the Scene can be divided into atomic units, i.e., Events." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.366, + 0.473, + 0.548 + ], + "angle": 0, + "content": "locate the transition positions in videos based on the similarity of the frames. For the scene segmentation, LGSS [47] converts the problem as a binary classification task at the shot level and applies a boundary-based model that aggregates adjacent shot features within a predefined sliding window to make a prediction. While SCRL [65] introduces a boundary-free model in a seq-to-seq way and reduces the inductive bias introduced by the predefined sliding window. As for story temporal segmentation, TI-NEWS [35] proposes a multi-modal framework based on the boundary-free model to complete the story segmentation and classification tasks simultaneously." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.561, + 0.253, + 0.579 + ], + "angle": 0, + "content": "3. Dataset Summary" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.587, + 0.47, + 0.678 + ], + "angle": 0, + "content": "In this section, we will describe some details of the NewsNet. First, we will introduce the definition and taxonomy of the four hierarchical levels of semantics and show some annotated examples of them. Then, the statistical information of the proposed dataset will be given to better understand the data distribution of the NewsNet." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.688, + 0.307, + 0.705 + ], + "angle": 0, + "content": "3.1. Definition and Taxonomy" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.712, + 0.469, + 0.818 + ], + "angle": 0, + "content": "Existing works focus on the shot- or event-level temporal segmentation, while neglecting the long-horizon cases, such as story- and topic-level segmentation. Hence this paper innovatively contributes more uniform benchmarks to the community, which covers all categories of temporal annotation. In order to clarify the meaning of the different labels, we detail them one by one:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.829, + 0.469, + 0.874 + ], + "angle": 0, + "content": "- Event: the atom unit in our setting, which is defined by the switching of the shots or the changing of the person." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.886, + 0.469, + 0.9 + ], + "angle": 0, + "content": "- Scene: a combination of several successive shots that" + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.829, + 0.469, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.531, + 0.366, + 0.892, + 0.396 + ], + "angle": 0, + "content": "focus on the same place from different angles. It is defined as a taxonomy-free unit." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.405, + 0.892, + 0.48 + ], + "angle": 0, + "content": "- Story: a sequence of scenes about a piece of news broadcast in a studio or outdoors. Story-level segments can be categorized into 'Connection', 'Place', 'Studio', 'Animation', 'Interview', and 'Photo'. A topic is composed of several stories." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.491, + 0.892, + 0.566 + ], + "angle": 0, + "content": "- Topic: it summarizes the content of a long sequence, with the following keywords: 'Health', 'Politics', 'Entertainment', 'Economy', 'Crime', 'Weather', 'Sport', 'Technology', 'Military' and 'Others'. Generally, an individual news video contains several topics." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.405, + 0.892, + 0.566 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.575, + 0.892, + 0.711 + ], + "angle": 0, + "content": "The above four different temporal annotations are defined in a hierarchical way. For example, a news video can be divided into several Topic-level segments; a Topic is further composed of several Stories; and so on. The difference between Story and Scene is that Story is category-aware while Scene is not, and the categories in Story are related to the data source. For example, Story contains a category of outdoor place footage, which is unique to news data and usually consists of multiple scenes, as illustrated in Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.72, + 0.722, + 0.735 + ], + "angle": 0, + "content": "3.2. Collection and Statistics" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.743, + 0.892, + 0.879 + ], + "angle": 0, + "content": "As a large-scale dataset, we collect about 2,000 news broadcast videos from more than 20 different TV stations, of which 1,000 were annotated, and all the data are downloaded from the YouTube and Internet. As shown in Fig. 3, we show the distribution of program duration and sources. News is a kind of typical long-form video, ranging from 10 to 110 minutes. Most of the collected videos have a 1-hour duration, performing as a challenging span for temporal segmentation. During the annotation, we found that" + }, + { + "type": "page_footnote", + "bbox": [ + 0.524, + 0.888, + 0.788, + 0.9 + ], + "angle": 0, + "content": "Place refers to the news reported outside a studio." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10672" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.09, + 0.885, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.251, + 0.893, + 0.294 + ], + "angle": 0, + "content": "Figure 3. The statistical information of NewsNet. The left histogram shows the distribution of the Top-20 news programs, (i.e., program A to T, by the total number of videos) in different duration, it can be found that the duration of the same program is relatively close. Besides, the two ring charts on the right show the category proportions of Topic and Story." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.319, + 0.471, + 0.425 + ], + "angle": 0, + "content": "there exists a long-tailed distribution for the categories of the story- and topic-level annotation, leading to a practical challenge. We believe such a benchmark can simulate the real-world scenario effectively, so that the competing methods could be benchmarked in a practical manner. More details such as annotation process, visualization, etc., can be found in the Supplementary Material." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.437, + 0.407, + 0.455 + ], + "angle": 0, + "content": "4. Benchmark and Protocol Navigation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.462, + 0.47, + 0.584 + ], + "angle": 0, + "content": "In this section, we will first introduce two temporal segmentation task settings, i.e., Separate Modeling and Hierarchical Modeling in NewsNet, with their definitions and protocols. Then, based on the potential relationship between different hierarchies, a simple yet effective loss function is introduced to better tackle hierarchical temporal segmentation. At last, we will briefly introduce the other common video tasks that can be conducted with NewsNet." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.591, + 0.399, + 0.607 + ], + "angle": 0, + "content": "4.1. Hierarchical Temporal Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.47, + 0.765 + ], + "angle": 0, + "content": "As long-form video can be divided into sub-video chunks according to its intrinsic properties, many researchers try to tackle the video segmentation problem from different perspectives in the last few years [12, 35, 47, 65]. Without loss of generality, we classify hierarchical temporal segmentation into two categories: Separate Modeling and Hierarchical Modeling. Separate modeling trains a segmentation model for each level segmentation task, while hierarchical modeling trains a model which simultaneously segments a video into multi-level segments." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.902 + ], + "angle": 0, + "content": "NewsNet supports temporal segmentation tasks with various semantic granularity, i.e., Event, Scene, Story, and Topic. Since many methods [12,47,65], take the Event (aka Shot sometimes) as the basic input unit to reduce the redundancy of the video and computational cost, NewsNet draws the video segmentation tasks on Scene, Story, and Topic. Moreover, these three hierarchies can be strictly aligned according to the Event (Shot) dimension, which is more conducive to Hierarchical Modeling and more convenient for" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.32, + 0.892, + 0.349 + ], + "angle": 0, + "content": "comparison. Note that we also carried out the event-level experiment (see Supplementary Material)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.35, + 0.893, + 0.518 + ], + "angle": 0, + "content": "Data and Modality. In-domain: all the videos are randomly split into training, validation, and testing sets with a ratio of 3:1:1. Cross-domain: training, validation, and testing sets are split according to different news program types/IDs; 5 different split combinations are set to ensure the generalization ability of the algorithm can be evaluated properly and fairly. As for the embeddings used in the experiment, the shot-level visual, audio, and textual features are extracted through ResNet50 [22] (pre-trained on Places365 [76]), public Bert-base model [14], and Cnn14 [27] (pre-trained on AudioSet [19]), respectively." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.534, + 0.688, + 0.55 + ], + "angle": 0, + "content": "4.1.1 Separate Modeling" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.557, + 0.892, + 0.648 + ], + "angle": 0, + "content": "Task Definition. Given a video with the shot labels, the model requires to classify whether a shot is a segmentation point for one of the three semantic levels (Scene, Story and Topic). A segmentation point refers to a shot that lies at the end of a topic/story/scene segment, and its next shot is the start of a new topic/story/scene." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.649, + 0.892, + 0.755 + ], + "angle": 0, + "content": "Backbone. In general, the common paradigm in those segmentation tasks is considering temporal segment to shot-level classification in boundary-base and -free way. We employ the state-of-the-art boundary-free model, SCRL [65], and the boundary-based model, LGSS [47] as our baseline approaches in separate modeling. More details about the protocol can be found in the Supplementary Material." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.772, + 0.714, + 0.787 + ], + "angle": 0, + "content": "4.1.2 Hierarchical Modeling" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.841 + ], + "angle": 0, + "content": "Task Definition. Given a video with the shot labels, the model requires to classify whether a shot is a segmentation point for three semantic levels (Scene, Story, and Topic)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.841, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Backbone. In this setting, as shown in Fig. 4, we focus on two hierarchical modeling paradigms, including Multi-Label and Multi-Head with a novel learning objective called Hierarchical Ranking. More specifically, Multi" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10673" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.122, + 0.089, + 0.852, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.322, + 0.895, + 0.366 + ], + "angle": 0, + "content": "Figure 4. An illustration of the separate modeling, two hierarchical modeling paradigms, and the proposed loss function. For the baseline, the model is trained individually corresponding to different segmentation tasks. For hierarchical modeling, all the tasks are aggregated into one model. Specifically, B, L, and N represent the batch size, sequence length, and feature dimension, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.391, + 0.473, + 0.677 + ], + "angle": 0, + "content": "Label works towards the hierarchical labels from the perspective of multi-label optimization, focusing on the aggregation of information into one head for estimation. While, Multi-Head addresses the challenge as a multi-task problem, where the tasks in different levels should be predicted via the individual pathways. Both pipelines are widely employed in modern architectures [13, 36, 60, 64], performing as simple but representative paradigms. Therefore, we adopt two hierarchical modeling paradigms as the baselines. Hierarchical Ranking Loss. To better model the hierarchical temporal video segmentation task and combine the existing methods, we propose a new loss function called Hierarchical Ranking Loss. The motivation is: due to the accurate alignment between the hierarchical levels with the atom unit Event, as shown in Fig. 1, a positive segment boundary of the higher level task must be a positive segment boundary of the lower task. Hence, the segmentation confidence scores for different tasks should be ranked in a non-increasing order from coarse to fine for the paired task." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.677, + 0.472, + 0.732 + ], + "angle": 0, + "content": "Given two tasks in the hierarchical modeling setting, e.g. task \\(\\mathcal{H}\\) and task \\(\\mathcal{L}\\), where the semantic level of task \\(\\mathcal{H}\\) is higher than that of task \\(\\mathcal{L}\\), the proposed hierarchical ranking loss function is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.137, + 0.734, + 0.47, + 0.761 + ], + "angle": 0, + "content": "\\[\n\\mathcal {P} _ {\\mathcal {H}, \\mathcal {L}} = \\frac {1}{B} \\frac {1}{L} \\cdot \\sum \\left[ \\sigma \\left(F _ {\\mathcal {H}} - \\mathcal {S G} \\left(F _ {\\mathcal {L}}\\right)\\right) \\cdot Y _ {\\mathcal {L}} \\right] \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.763, + 0.47, + 0.852 + ], + "angle": 0, + "content": "where \\(B\\) and \\(L\\) are the batch size and sequence length respectively; \\(\\sigma(\\cdot)\\) refers to the contraction mapping, like sigmoid function; \\(F\\) stands for the segmentation confidence score of the task; \\(\\mathcal{SG}\\) is the Stop Gradient operation for the computational graph, and \\(Y_{\\mathcal{L}}\\) corresponds to the label of the task \\(\\mathcal{L}\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.853, + 0.47, + 0.88 + ], + "angle": 0, + "content": "And the final objective function for all three tasks can be formed as follow:" + }, + { + "type": "equation", + "bbox": [ + 0.088, + 0.887, + 0.47, + 0.902 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {P} _ {\\text {T o p i c}, S t o r y} + \\mathcal {P} _ {\\text {T o p i c}, S c e n e} + \\mathcal {P} _ {\\text {S t o r y}, S c e n e} + \\mathcal {L} _ {C E} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.391, + 0.809, + 0.407 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{CE}\\) is the cross-entropy loss function." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.414, + 0.825, + 0.431 + ], + "angle": 0, + "content": "4.2. Common Video Understanding Tasks" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.438, + 0.892, + 0.498 + ], + "angle": 0, + "content": "The NewsNet could also apply other common video understanding tasks, e.g., video classification, localization, and highlight detection. More details can be found in Supplementary Material." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.51, + 0.738, + 0.528 + ], + "angle": 0, + "content": "5. Experiment and Analysis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.536, + 0.892, + 0.673 + ], + "angle": 0, + "content": "In this section, we mainly focus on the temporal segmentation tasks on the NewsNet and further contribute several benchmarks on the different baselines. In particular, we conduct the empirical study to thoroughly benchmark the effectiveness of model structure, multi-modality information, domain gap, and hierarchical modeling. Note that, due to the limitation on page length, the implementation details and the results for other common video tasks are given in the Supplementary Material." + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.685, + 0.894, + 0.73 + ], + "angle": 0, + "content": "Table 2. The performance between boundary-based (B.B.) and boundary-free (B.F) model with visual input on F1 score \\(\\uparrow\\), Precision \\(\\uparrow\\), and Recall \\(\\uparrow\\). Light gray refers to the failure case." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.742, + 0.892, + 0.838 + ], + "angle": 0, + "content": "
TaskModelF1 scorePrecisionRecall
SceneB.B.77.367.790.2
B.F.76.876.177.5
StoryB.B.70.765.976.2
B.F.71.272.370.0
TopicB.B.13.17.0100.0
B.F.62.972.455.6
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.895, + 0.903 + ], + "angle": 0, + "content": "As an overview, based on the large-scale dataset with rich modalities, we conduct several experiments to respond to five critical questions concerned by the community:" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "10674" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.08, + 0.09, + 0.872, + 0.133 + ], + "angle": 0, + "content": "Table 3. In-domain performance by using boundary-free (B.F.) Table 4. Cross-domain setting by using boundary-free (B.F.) model. The bolded values stand for the optimal performances for model. The bolded values stand for the optimal performances for each task." + }, + { + "type": "table", + "bbox": [ + 0.086, + 0.145, + 0.482, + 0.327 + ], + "angle": 0, + "content": "
TaskModalityF1 scorePrecisionRecall
SceneV76.876.177.5
A69.866.873.2
T66.756.381.9
V+A+T78.380.975.8
StoryV71.272.370.0
A59.357.661.1
T50.657.445.2
V+A+T75.474.776.2
TopicV62.972.455.6
A58.159.456.9
T39.046.533.5
V+A+T73.274.372.2
" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.342, + 0.468, + 0.386 + ], + "angle": 0, + "content": "1. Can existing methods also perform well on long-horizon temporal segmentation tasks, such as topic-level and story-level segmentation?" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.394, + 0.468, + 0.425 + ], + "angle": 0, + "content": "2. Do the different-level tasks test the ability of the model on viewing different temporal spans?" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.433, + 0.468, + 0.463 + ], + "angle": 0, + "content": "3. Can existing methods well address the semantic gap between low- and high-level granularities?" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.471, + 0.468, + 0.501 + ], + "angle": 0, + "content": "4. Will the temporal segmentation models be benefited from multi-modality information?" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.51, + 0.468, + 0.54 + ], + "angle": 0, + "content": "5. How could the models benefit from Hierarchical Modeling?" + }, + { + "type": "list", + "bbox": [ + 0.09, + 0.342, + 0.468, + 0.54 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.545, + 0.43, + 0.562 + ], + "angle": 0, + "content": "5.1. Results against Increased Temporal Span" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.568, + 0.468, + 0.839 + ], + "angle": 0, + "content": "To investigate the generalization of the existing methods against the long-horizon cases, we conduct an experiment on two different head architectures, including boundary-free and boundary-based heads, on three tasks: scene-level, story-level, and topic-level temporal segmentation. As shown in Fig. 2, the performance of the boundary-based model has a significant drop on the Topic task. Compared with the other tasks, the Topic segmentation task has more time slices for input, leading to a hard challenge caused by longer time dependencies. The methods can segment scenes with about 76-77 F1 score while degrading to 62.9 on Topic-level segmentation. For the first question, the answer is that both the boundary-free model and the boundary-based model can not generalize well in the story-and topic-level segmentation. Combining the overall performance and stability on different tasks, the next experiments will utilize a boundary-free model as the basic backbone under comprehensive consideration." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.848, + 0.437, + 0.865 + ], + "angle": 0, + "content": "5.2. Performance against Cross News Domains" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.871, + 0.468, + 0.901 + ], + "angle": 0, + "content": "To investigate the generalization of the existing method, we analyze the performance of the method against the va" + }, + { + "type": "table", + "bbox": [ + 0.49, + 0.145, + 0.862, + 0.327 + ], + "angle": 0, + "content": "
TaskModalityAvg. F1 score (std.)Avg. PrecisionAvg. Recall
SceneV72.9 (2.1)70.975.2
A62.7 (4.0)59.766.6
T61.6 (5.0)52.877.0
V+A+T76.0 (2.1)74.477.9
StoryV68.5 (2.6)70.366.9
A55.7 (3.6)53.659.0
T51.1 (3.6)43.465.4
V+A+T72.9 (2.2)73.772.4
TopicV60.6 (4.7)69.853.8
A59.0 (5.2)56.062.9
T49.8 (5.2)45.755.9
V+A+T72.2 (3.6)72.372.5
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.341, + 0.892, + 0.597 + ], + "angle": 0, + "content": "rieties of different news programs. As shown in Table 3 and 4, the cross-domain setting slightly degrades the performance on all the tasks. Comparing the in-domain results using the full modality in Table 3 and corresponding cross-domain results in Table 4, the F1 score decreases by 2.3 on task Scene, by 2.5 on task Story, and by 1.0 on task Topic. The reason for this result may be that higher-level tasks, such as Topic, has more abstract semantics and the knowledge is easier to transfer. In contrast, task Scene in the lowest level, generally only considers place switching as the segmentation cue. Besides, the category of Scene is an open set, leading to a greater challenge in the cross-domain setting. Hence, temporal segmentation model also suffers from the out-of-distribution case. Based on the property of the collected dataset, we set the videos from different news programs as different domains, and establish a comprehensive benchmark to facilitate the research on such line." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.605, + 0.816, + 0.621 + ], + "angle": 0, + "content": "5.3. Performance using Multi-Modalities" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.892, + 0.81 + ], + "angle": 0, + "content": "When it comes to multi-modality, as shown in Table. 4, the model can be benefited from multimodal information among all the tasks. For example, the performance of using only visual modality achieves only 60.6 recall in topic-level segmentation. However, by aggregating the other modalities including audio and text, the performance can be improved significantly (72.2 vs 49.8), which indicates the potential value for addressing long-discrepancy via multimodality technology. Note that, the improvement in Scene segmentation is, to some extent, limited, we believe that Scene is mainly defined by visual input, hence can not obtain adequate complementarity from other modalities." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.818, + 0.885, + 0.835 + ], + "angle": 0, + "content": "5.4. Performance based on Hierarchical Modeling" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.841, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Here, we discuss the core challenge revealed in this paper: 'How could the model benefit from hierarchical annotations?' In this case, the baseline is set as the model, which can only be accessible to one specific-level annotation. As" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "10675" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.472, + 0.175 + ], + "angle": 0, + "content": "Table 5. The F1 scores of baselines trained with different levels of annotations on full modalities without our hierarchical ranking loss, where blue and orange indicate the in-domain and cross-domain, respectively. Each row refers to the result corresponding to a single task. Hie. Modeling stands for Hierarchical Modeling while Sep. Modeling is Separate Modeling." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.185, + 0.469, + 0.37 + ], + "angle": 0, + "content": "
RecipeBaseline\nSep. ModelingMulti-Label\nHie. ModelingMulti-Head\nHie. Modeling
Scene78.3 / 76.079.1 / 76.579.9 / 76.9
+ Story75.4 / 72.975.4 / 74.774.2 / 74.0
Scene78.3 / 76.079.8 / 76.479.5 / 76.5
+ Topic73.2 / 72.270.5 / 72.870.9 / 73.0
Story75.4 / 72.976.2 / 74.375.4 / 73.9
+ Topic73.2 / 72.277.3 / 73.275.2 / 73.5
Scene78.3 / 76.077.4 / 76.879.8 / 76.8
+ Story75.4 / 72.974.3 / 74.374.5 / 73.7
+ Topic73.2 / 72.274.3 / 72.676.6 / 70.4
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.403, + 0.47, + 0.63 + ], + "angle": 0, + "content": "listed in Table 5, we consider two simple strategies, including Multi-Label and Multi-Head, to aggregate the hierarchical information. By training with multi-level annotations, a \\(1.0 \\sim 3.0\\%\\) improvement can be achieved, which indicates, the segmentation can be benefited by introducing the hierarchical structures of videos. However, Table 5 also shows directly combining different-level granularities might not result in gains due to the semantic gap between low- and high-level granularities. For example, directly combining scene- and topic-level information by Multi-Label hierarchical modeling or Multi-Head hierarchical modeling does not consistently improve the topic-level segmentation performance. This poses a new and challenging research question, i.e. how to jointly train with various granularities toward hierarchical modeling of long-form videos?" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.638, + 0.47, + 0.759 + ], + "angle": 0, + "content": "To address the above question, we introduce a simple but effective loss function, i.e., Hierarchical Ranking Loss introduced in Sec. 4.1.2. As shown in Table 6, by combining with our proposed loss, the F1 score for segmentation can be further improved in most cases. Typically, in Topic-level, Multi-Head coupled with the proposed loss can reach the best performance, i.e., improving Multi-Head case from 70.4 to 73.2 F1 score." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.793, + 0.45, + 0.81 + ], + "angle": 0, + "content": "5.5. Other Benchmarks involved in the NewsNet" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.47, + 0.903 + ], + "angle": 0, + "content": "In addition to the temporal segmentation tasks, we also conduct extensive common video tasks, including highlight detection, video classification, and video localization, on the proposed dataset. Due to the limited pages, we show this part in the Supplementary Materials." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.089, + 0.892, + 0.146 + ], + "angle": 0, + "content": "Table 6. The F1 scores of the methods with or without hierarchical ranking loss under the in-domain / cross-domain setting on full modalities. Hie. stands for Hierarchical Modeling and Sep. refers to Separate Modeling." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.157, + 0.892, + 0.274 + ], + "angle": 0, + "content": "
MethodSceneStoryTopic
Baseline (Sep.)78.3 / 76.075.4 / 72.973.2 / 72.2
Multi-Label (Hie.)77.4 / 76.874.3 / 74.374.3 / 72.6
Multi-Label w/ Hie. Loss (Hie.)79.6 / 76.974.4 / 73.577.8 / 73.1
Multi-Head (Hie.)79.8 / 76.874.5 / 73.776.6 / 70.4
Multi-Head w/ Hie. Loss (Hie.)80.3 / 76.976.3 / 74.676.5 / 73.2
" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.297, + 0.866, + 0.314 + ], + "angle": 0, + "content": "6. Conclusion, Limitation and Future work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.326, + 0.892, + 0.599 + ], + "angle": 0, + "content": "We introduce NewsNet, a new large-scale dataset for temporal video segmentation. Compared with existing large-scale datasets for temporal video segmentation, NewsNet additionally provides two abstractive levels of temporal segmentation, which have not been taken into account by these datasets. The four-level and hierarchical annotations enable the community to explore how to comprehensively represent the complex structure of long-form videos from coarse to fine. Extensive efforts of human workers are devoted, so as to ensure that annotations are high-quality. We hope that NewsNet and baseline results can facilitate the development of temporal video segmentation in terms of insightful research and practical tools. In addition, as NewsNet provides diverse annotations for dividing a video into segments of different granularities as well as category labels, it can also serve other video understanding tasks such as video grounding, captioning, retrieval, highlight detection, and video classification." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.602, + 0.892, + 0.709 + ], + "angle": 0, + "content": "In this paper, we only focus on the problem of hierarchical temporal segmentation. We have not explored other multi-modality tasks, such as visual reasoning and video question answering, while our dataset can serve these tasks thanks to its multi-modality nature. We will explore these tasks in our future work to further promote advances in multi-modality-based video understanding and reasoning." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.733, + 0.661, + 0.751 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.761, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This work was supported by the KAUST Office of Sponsored Research through the Visual Computing Center (VCC) funding, as well as, the SDAIA-KAUST Center of Excellence in Data Science and Artificial Intelligence (SDAIA-KAUST AI). Raghavendra Ramachandra is supported under the SALT and OFFPAD projects funded by the Research Council of Norway, Norway. Chia-Wen Lin is supported by National Science and Technology Council, Taiwan, under project NSTC 111-2634-F002-023." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "10676" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.468, + 0.142 + ], + "angle": 0, + "content": "[1] LOVEU21. https://sites.google.com/view/loveucvpr21, 2021.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.144, + 0.468, + 0.183 + ], + "angle": 0, + "content": "[2] LOVEU22 program schedule. https://sites.google.com/view/loveucvpr22/program, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.187, + 0.468, + 0.268 + ], + "angle": 0, + "content": "[3] Nayyer Aafaq, Naveed Akhtar, Wei Liu, Syed Zulqarnain Gilani, and Ajmal Mian. Spatio-temporal dynamics and semantic attribute enriched visual encoding for video captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12487-12496, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.271, + 0.468, + 0.338 + ], + "angle": 0, + "content": "[4] Sami Abu-El-Haija, Nisarg Kothari, Joonseok Lee, Paul Natsev, George Toderici, Balakrishnan Varadarajan, and Sudheendra Vijayanarasimhan. Youtube-8m: A large-scale video classification benchmark. arXiv preprint arXiv:1609.08675, 2016. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.341, + 0.468, + 0.408 + ], + "angle": 0, + "content": "[5] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1728-1738, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.411, + 0.468, + 0.466 + ], + "angle": 0, + "content": "[6] Lorenzo Baraldi, Costantino Grana, and Rita Cucchiara. A deep siamese network for scene detection in broadcast videos. In Proceedings of the 23rd ACM international conference on Multimedia, pages 1199-1202, 2015. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.468, + 0.468, + 0.535 + ], + "angle": 0, + "content": "[7] Lorenzo Baraldi, Costantino Grana, and Rita Cucchiara. Shot and scene detection via hierarchical clustering for re-using broadcast video. In International conference on computer analysis of images and patterns, pages 801-811. Springer, 2015. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.537, + 0.468, + 0.606 + ], + "angle": 0, + "content": "[8] Joe Barrow, Rajiv Jain, Vlad Morariu, Varun Manjunatha, Douglas W Oard, and Philip Resnik. A joint model for document segmentation and segment labeling. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 313-322, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.607, + 0.468, + 0.675 + ], + "angle": 0, + "content": "[9] Fabian Caba Heilbron, Victor Escorcia, Bernard Ghanem, and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 961-970, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.677, + 0.468, + 0.732 + ], + "angle": 0, + "content": "[10] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.734, + 0.468, + 0.788 + ], + "angle": 0, + "content": "[11] Vasileios T Chasanis, Aristidis C Likas, and Nikolaos P Galatsanos. Scene detection in videos using shot clustering and sequence alignment. IEEE transactions on multimedia, 11(1):89-100, 2008. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.468, + 0.858 + ], + "angle": 0, + "content": "[12] Shixing Chen, Xiaohan Nie, David Fan, Dongqing Zhang, Vimal Bhat, and Raffay Hamid. Shot contrastive self-supervised learning for scene boundary detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9796-9805, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[13] Zhao-Min Chen, Xiu-Shen Wei, Peng Wang, and Yanwen Guo. Multi-label image recognition with graph convolutional networks. In Proceedings of the IEEE/CVF Confer-" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.468, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.094, + 0.892, + 0.119 + ], + "angle": 0, + "content": "ence on Computer Vision and Pattern Recognition (CVPR), June 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.175 + ], + "angle": 0, + "content": "[14] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.178, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[15] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6202-6211, 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.234, + 0.892, + 0.288 + ], + "angle": 0, + "content": "[16] Valentin Gabeur, Chen Sun, Karteek Alahari, and Cordelia Schmid. Multi-modal transformer for video retrieval. In European Conference on Computer Vision, pages 214-229. Springer, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.289, + 0.892, + 0.343 + ], + "angle": 0, + "content": "[17] Jiyang Gao, Chen Sun, Zhenheng Yang, and Ram Nevatia. Tall: Temporal activity localization via language query. In Proceedings of the IEEE international conference on computer vision, pages 5267-5275, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.345, + 0.892, + 0.399 + ], + "angle": 0, + "content": "[18] Ana Garcia del Molino and Michael Gygli. PhD-gifs: personalized highlight detection for automatic gif creation. In Proceedings of the 26th ACM international conference on Multimedia, pages 600–608, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.401, + 0.892, + 0.482 + ], + "angle": 0, + "content": "[19] Jort F Gemmeke, Daniel PW Ellis, Dylan Freedman, Aren Jansen, Wade Lawrence, R Channing Moore, Manoj Plakal, and Marvin Ritter. Audio set: An ontology and human-labeled dataset for audio events. In 2017 IEEE international conference on acoustics, speech and signal processing (ICASSP), pages 776-780. IEEE, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.484, + 0.892, + 0.538 + ], + "angle": 0, + "content": "[20] Michael Gygli, Yale Song, and Liangliang Cao. Video2gif: Automatic generation of animated gifs from video. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1001-1009, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.54, + 0.892, + 0.607 + ], + "angle": 0, + "content": "[21] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.609, + 0.892, + 0.663 + ], + "angle": 0, + "content": "[22] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.665, + 0.892, + 0.719 + ], + "angle": 0, + "content": "[23] Qingqiu Huang, Yu Xiong, Anyi Rao, Jiaze Wang, and Dahua Lin. Movienet: A holistic dataset for movie understanding. In European Conference on Computer Vision, pages 709-727. Springer, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.721, + 0.892, + 0.762 + ], + "angle": 0, + "content": "[24] Chen Ju, Tengda Han, Kunhao Zheng, Ya Zhang, and Weidi Xie. Prompting visual-language models for efficient video understanding. arXiv preprint arXiv:2112.04478, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.763, + 0.892, + 0.831 + ], + "angle": 0, + "content": "[25] Hyolim Kang, Jinwoo Kim, Taehyun Kim, and Seon Joo Kim. Uboco: Unsupervised boundary contrastive learning for generic event boundary detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20073-20082, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.833, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[26] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.094, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10677" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[27] Qiuqiang Kong, Yin Cao, Turab Iqbal, Yuxuan Wang, Wenwu Wang, and Mark D Plumbley. Panns: Large-scale pretrained audio neural networks for audio pattern recognition. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 28:2880-2894, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.164, + 0.47, + 0.22 + ], + "angle": 0, + "content": "[28] Ranjay Krishna, Kenji Hata, Frederic Ren, Li Fei-Fei, and Juan Carlos Niebles. Dense-captioning events in videos. In Proceedings of the IEEE international conference on computer vision, pages 706–715, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.221, + 0.469, + 0.289 + ], + "angle": 0, + "content": "[29] Kathy Lee, Diana Palsetia, Ramanathan Narayanan, Md Mostofa Ali Patwary, Ankit Agrawal, and Alok Choudhary. Twitter trending topic classification. In 2011 IEEE 11th International Conference on Data Mining Workshops, pages 251-258. IEEE, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.291, + 0.469, + 0.347 + ], + "angle": 0, + "content": "[30] Jie Lei, Tamara L Berg, and Mohit Bansal. Detecting moments and highlights in videos via natural language queries. Advances in Neural Information Processing Systems, 34:11846-11858, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.349, + 0.469, + 0.418 + ], + "angle": 0, + "content": "[31] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L Berg, Mohit Bansal, and Jingjing Liu. Less is more: Clipbert for video-and-language learning via sparse sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7331-7341, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.42, + 0.469, + 0.475 + ], + "angle": 0, + "content": "[32] Jie Lei, Licheng Yu, Tamara L Berg, and Mohit Bansal. Tvr: A large-scale dataset for video-subtitle moment retrieval. In European Conference on Computer Vision, pages 447-463. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.477, + 0.469, + 0.558 + ], + "angle": 0, + "content": "[33] Congcong Li, Xinyao Wang, Longyin Wen, Dxiang Hong, Tiejian Luo, and Libo Zhang. End-to-end compressed video representation learning for generic event boundary detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13967-13976, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.562, + 0.469, + 0.644 + ], + "angle": 0, + "content": "[34] Kevin Lin, Linjie Li, Chung-Ching Lin, Faisal Ahmed, Zhe Gan, Zicheng Liu, Yumao Lu, and Lijuan Wang. Swinbert: End-to-end transformers with sparse attention for video captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17949-17958, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.647, + 0.469, + 0.728 + ], + "angle": 0, + "content": "[35] Ye Liu, Lingfeng Qiao, Di Yin, Zhuoxuan Jiang, Xinghua Jiang, Deqiang Jiang, and Bo Ren. Os-msl: One stage multimodal sequential link framework for scene segmentation and classification. In Proceedings of the 30th ACM International Conference on Multimedia, pages 6269-6277, 2022. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.731, + 0.469, + 0.813 + ], + "angle": 0, + "content": "[36] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10012-10022, October 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.816, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[37] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swim transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3202-3211, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[38] Jakub Lokoc, Gregor Kovalcik, Tomáš Souček, Jaroslav Moravec, and Přemysl Čech. A framework for effective" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "known-item search in video. In Proceedings of the 27th ACM International Conference on Multimedia, pages 1777-1785, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.136, + 0.892, + 0.178 + ], + "angle": 0, + "content": "[39] Sauradip Nag, Xiatian Zhu, Yi-Zhe Song, and Tao Xiang. Zero-shot temporal action detection via vision-language prompting. arXiv preprint arXiv:2207.08184, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.18, + 0.892, + 0.248 + ], + "angle": 0, + "content": "[40] Guoshun Nan, Rui Qiao, Yao Xiao, Jun Liu, Sicong Leng, Hao Zhang, and Wei Lu. Interventional video grounding with dual contrastive learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2765-2775, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.251, + 0.892, + 0.319 + ], + "angle": 0, + "content": "[41] Fahmi Salman Nurfikri, Mohamad Syahrul Mubarok, et al. News topic classification using mutual information and bayesian network. In 2018 6th International Conference on Information and Communication Technology (ICoICT), pages 162-166. IEEE, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.321, + 0.892, + 0.389 + ], + "angle": 0, + "content": "[42] Sofia Ares Oliveira, Benoit Seguin, and Frederic Kaplan. dhsegment: A generic deep-learning approach for document segmentation. In 2018 16th International Conference on Frontiers in Handwriting Recognition (ICFHR), pages 7-12. IEEE, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.392, + 0.892, + 0.475 + ], + "angle": 0, + "content": "[43] Alejandro Pardo, Fabian Caba Heilbron, Juan León Alcazar, Ali Thabet, and Bernard Ghanem. Moviecuts: A new dataset and benchmark for cut type recognition. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part VII, pages 668-685, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.477, + 0.892, + 0.532 + ], + "angle": 0, + "content": "[44] Will Price, Carl Vondrick, and Dima Damen. Unweavenet: Unweaving activity stories. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13770-13779, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.534, + 0.892, + 0.629 + ], + "angle": 0, + "content": "[45] Zhiwu Qing, Shiwei Zhang, Ziyuan Huang, Yi Xu, Xiang Wang, Mingqian Tang, Changxin Gao, Rong Jin, and Nong Sang. Learning from untrimmed videos: Self-supervised video representation learning with hierarchical consistency. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13821-13831, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.633, + 0.892, + 0.715 + ], + "angle": 0, + "content": "[46] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.718, + 0.892, + 0.799 + ], + "angle": 0, + "content": "[47] Anyi Rao, Linning Xu, Yu Xiong, Guodong Xu, Qingqiu Huang, Bolei Zhou, and Dahua Lin. A local-to-global approach to multi-modal movie scene segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10146-10155, 2020. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.802, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[48] Daniel Rotman, Dror Porat, and Gal Ashour. Optimal sequential grouping for robust video scene detection using multiple modalities. International Journal of Semantic Computing, 11(02):193-208, 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[49] Ellen Rushe and Brian Mac Namee. Anomaly detection in raw audio using deep autoregressive networks. In ICASSP 2019-2019 IEEE International Conference on Acoustics," + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10678" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.468, + 0.119 + ], + "angle": 0, + "content": "Speech and Signal Processing (ICASSP), pages 3597-3601. IEEE, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.469, + 0.176 + ], + "angle": 0, + "content": "[50] Weimin Shi, Mingchen Zhuge, Zhong Zhou, Dehong Gao, and Deng-Ping Fan. Qr-clip: Introducing explicit open-world knowledge for location and time reasoning. arXiv preprint arXiv:2302.00952, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.178, + 0.469, + 0.246 + ], + "angle": 0, + "content": "[51] Mike Zheng Shou, Stan Weixian Lei, Weiyao Wang, Deepti Ghadiyaram, and Matt Feiszli. Generic event boundary detection: A benchmark for event segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8075-8084, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.247, + 0.469, + 0.329 + ], + "angle": 0, + "content": "[52] Nina Shvetsova, Brian Chen, Andrew Rouditchenko, Samuel Thomas, Brian Kingsbury, Rogerio S Feris, David Harwath, James Glass, and Hilde Kuehne. Everything at once-multimodal fusion transformer for video retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20020-20029, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.33, + 0.469, + 0.412 + ], + "angle": 0, + "content": "[53] Mattia Soldan, Alejandro Pardo, Juan León Alcazar, Fabian Caba, Chen Zhao, Silvio Giancola, and Bernard Ghanem. Mad: A scalable dataset for language grounding in videos from movie audio descriptions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5026-5035, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.414, + 0.469, + 0.482 + ], + "angle": 0, + "content": "[54] Yale Song, Miriam Redi, Jordi Vallmitjana, and Alejandro Jaimes. To click or not to click: Automatic selection of beautiful thumbnails from videos. In Proceedings of the 25th ACM international on conference on information and knowledge management, pages 659-668, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.484, + 0.469, + 0.524 + ], + "angle": 0, + "content": "[55] Tomáš Souček and Jakub Lokoc. Transnet v2: an effective deep network architecture for fast shot transition detection. arXiv preprint arXiv:2008.04838, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.526, + 0.469, + 0.594 + ], + "angle": 0, + "content": "[56] Jiaqi Tang, Zhaoyang Liu, Chen Qian, Wayne Wu, and Limin Wang. Progressive attention on multi-level dense difference maps for generic event boundary detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3355-3364, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.595, + 0.469, + 0.65 + ], + "angle": 0, + "content": "[57] Shitao Tang, Litong Feng, Zhanghui Kuang, Yimin Chen, and Wei Zhang. Fast video shot transition localization with deep structured models. In Asian Conference on Computer Vision, pages 577-592. Springer, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.651, + 0.469, + 0.718 + ], + "angle": 0, + "content": "[58] Du Tran, Lubomir Bourdev, Rob Fergus, Lorenzo Torresani, and Manohar Paluri. Learning spatiotemporal features with 3d convolutional networks. In Proceedings of the IEEE international conference on computer vision, pages 4489-4497, 2015. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.721, + 0.469, + 0.788 + ], + "angle": 0, + "content": "[59] Du Tran, Heng Wang, Lorenzo Torresani, and Matt Feiszli. Video classification with channel-separated convolutional networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5552-5561, 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.47, + 0.845 + ], + "angle": 0, + "content": "[60] Dongkai Wang and Shiliang Zhang. Unsupervised person re-identification via multi-label classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[61] Jue Wang and Anoop Cherian. Gods: Generalized one-class discriminative subspaces for anomaly detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8201-8211, 2019. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.161 + ], + "angle": 0, + "content": "[62] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaou Tang, and Luc Van Gool. Temporal segment networks: Towards good practices for deep action recognition. In European conference on computer vision, pages 20-36. Springer, 2016. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[63] Yuxuan Wang, Difei Gao, Licheng Yu, Weixian Lei, Matt Feiszli, and Mike Zheng Shou. Geb+: A benchmark for generic event boundary captioning, grounding and retrieval. In European Conference on Computer Vision, pages 709-725. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.233, + 0.892, + 0.314 + ], + "angle": 0, + "content": "[64] Jonatas Wehrmann, Ricardo Cerri, and Rodrigo Barros. Hierarchical multi-label classification networks. In Jennifer Dy and Andreas Krause, editors, Proceedings of the 35th International Conference on Machine Learning, volume 80 of Proceedings of Machine Learning Research, pages 5075-5084. PMLR, 10-15 Jul 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.316, + 0.892, + 0.398 + ], + "angle": 0, + "content": "[65] Haoqian Wu, Keyu Chen, Yanan Luo, Ruizhi Qiao, Bo Ren, Haozhe Liu, Weicheng Xie, and Linlin Shen. Scene consistency representation learning for video scene segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14021-14030, 2022. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.4, + 0.892, + 0.468 + ], + "angle": 0, + "content": "[66] Bo Xiong, Yannis Kalantidis, Deepti Ghadiyaram, and Kristen Grauman. Less is more: Learning highlight detection from video duration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1258-1267, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.47, + 0.892, + 0.537 + ], + "angle": 0, + "content": "[67] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. Videoclip: Contrastive pre-training for zero-shot video-text understanding. arXiv preprint arXiv:2109.14084, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.539, + 0.892, + 0.621 + ], + "angle": 0, + "content": "[68] Mengmeng Xu, Juan-Manuel Pérez-Rúa, Victor Escorcia, Brais Martinez, Xiatian Zhu, Li Zhang, Bernard Ghanem, and Tao Xiang. Boundary-sensitive pre-training for temporal localization in videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7220-7230, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.623, + 0.892, + 0.691 + ], + "angle": 0, + "content": "[69] Minghao Xu, Hang Wang, Bingbing Ni, Riheng Zhu, Zhenbang Sun, and Changhu Wang. Cross-category video highlight detection via set-based learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7970-7979, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.692, + 0.892, + 0.761 + ], + "angle": 0, + "content": "[70] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Tubedetr: Spatio-temporal video grounding with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16442-16453, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.763, + 0.892, + 0.831 + ], + "angle": 0, + "content": "[71] Hanhua Ye, Guorong Li, Yuankai Qi, Shuhui Wang, Qingming Huang, and Ming-Hsuan Yang. Hierarchical modular network for video captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17939-17948, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[72] Muhammad Zaigham Zaheer, Arif Mahmood, Marcella Astrid, and Seung-Ik Lee. Claws: Clustering assisted weakly supervised learning with normalcy suppression for anomalous event detection. In European Conference on Computer Vision, pages 358-376. Springer, 2020. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10679" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[73] M Zaigham Zaheer, Arif Mahmood, M Haris Khan, Mattia Segu, Fisher Yu, and Seung-Ik Lee. Generative cooperative learning for unsupervised video anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14744-14754, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.47, + 0.232 + ], + "angle": 0, + "content": "[74] Runhao Zeng, Haoming Xu, Wenbing Huang, Peihao Chen, Mingkui Tan, and Chuang Gan. Dense regression network for video grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10287-10296, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.234, + 0.47, + 0.288 + ], + "angle": 0, + "content": "[75] Haoxin Zhang, Zhimin Li, and Qinglin Lu. Better learning shot boundary detection via multi-task. In Proceedings of the 29th ACM International Conference on Multimedia, pages 4730-4734, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.29, + 0.47, + 0.346 + ], + "angle": 0, + "content": "[76] Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. IEEE transactions on pattern analysis and machine intelligence, 40(6):1452-1464, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.348, + 0.47, + 0.429 + ], + "angle": 0, + "content": "[77] Mingchen Zhuge, Dehong Gao, Deng-Ping Fan, Linbo Jin, Ben Chen, Haoming Zhou, Minghui Qiu, and Ling Shao. Kaleido-bert: Vision-language pre-training on fashion domain. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12647-12657, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "10680" + } + ] +] \ No newline at end of file diff --git a/2023/NewsNet_ A Novel Dataset for Hierarchical Temporal Segmentation/d2756edd-2b85-4c25-85cd-99de3bdf0d9a_origin.pdf b/2023/NewsNet_ A Novel Dataset for Hierarchical Temporal Segmentation/d2756edd-2b85-4c25-85cd-99de3bdf0d9a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a08e968771d289ca43a0910a9aec870412205e29 --- /dev/null +++ b/2023/NewsNet_ A Novel Dataset for Hierarchical Temporal Segmentation/d2756edd-2b85-4c25-85cd-99de3bdf0d9a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0016a9342e19bda0617b5c9983f934bbbd811fce417b77968e9c675097606baa +size 4176446 diff --git a/2023/NewsNet_ A Novel Dataset for Hierarchical Temporal Segmentation/full.md b/2023/NewsNet_ A Novel Dataset for Hierarchical Temporal Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..9015e1524e07163f1042bc2c9d20100f7789d9c8 --- /dev/null +++ b/2023/NewsNet_ A Novel Dataset for Hierarchical Temporal Segmentation/full.md @@ -0,0 +1,305 @@ +# NewsNet: A Novel Dataset for Hierarchical Temporal Segmentation + +Haoqian $\mathbf{W}\mathbf{u}^{1\dagger}$ , Keyu Chen $^{1\dagger}$ , Haozhe Liu $^{2\dagger}$ , Mingchen Zhuge $^{2\dagger}$ , Bing Li $^{2\boxtimes}$ , Ruizhi Qiao $^{1\boxtimes}$ , Xiujun Shu $^{1}$ , Bei Gan $^{1}$ , Liangsheng Xu $^{1}$ , Bo Ren $^{1}$ , Mengmeng Xu $^{2}$ , Wentian Zhang $^{2}$ , Raghavendra Ramachandra $^{3}$ , Chia-Wen Lin $^{4}$ , Bernard Ghanem $^{2}$ + +$^{1}$ Tencent $^{2}$ AI Initiative, King Abdullah University of Science and Technology (KAUST) + $^{3}$ Norwegian University of Science and Technology (NTNU) + $^{4}$ National Tsing Hua University (NTHU) + +![](images/573668c2860b1009046c02f06c69b022fdb789c6a1705bb0ea19027ffd604e94.jpg) +Figure 1. Examples from the proposed NewsNet. The dataset provides multimodal and hierarchical labels on each video, including 4-level hierarchical temporal information (i.e., event, scene, story, topic) and 3 modality-specific cues (i.e., video, audio, text). A prerequisite motivation behind NewsNet is that long-form videos can be recursively represented into several sub-videos according to the semantics, which is an intrinsic property of the video but rarely explored in the community. To bridge this research gap, this paper innovatively constructs a large-scale dataset to benchmark the algorithms for hierarchical understanding like a human being. + +# Abstract + +Temporal video segmentation is the get-to-go automatic video analysis, which decomposes a long-form video into smaller components for the following-up understanding tasks. Recent works have studied several levels of granularity to segment a video, such as shot, event, and scene. Those segmentations can help compare the semantics in the corresponding scales, but lack a wider view of larger temporal spans, especially when the video is complex and structured. Therefore, we present two abstractive levels of temporal segmentations and study their hierarchy to the existing fine-grained levels. Accordingly, we collect NewsNet, the largest news video dataset consisting of 1,000 videos + +in over 900 hours, associated with several tasks for hierarchical temporal video segmentation. Each news video is a collection of stories on different topics, represented as aligned audio, visual, and textual data, along with extensive frame-wise annotations in four granularities. We assert that the study on NewsNet can advance the understanding of complex structured video and benefit more areas such as short-video creation, personalized advertisement, digital instruction, and education. Our dataset and code is publicly available at https://github.com/NewsNet-Benchmark/NewsNet. + +# 1. Introduction + +Temporal video segmentation is a critical problem in video understanding, which is essential for many video applications such as video classification [10, 15, 37, 58, 59, 62], + +Table 1. Data comparison between the NewsNet and other related datasets. The NewsNet provides various multimodal data and hierarchical temporal segmentation annotations. Doc: documentary, Ads: advertisement. (Please refer to our project page for more details.) + +
Dataset# VideoDuration (hours)Modality# Annotation(s) per VideoSource
TopicStorySceneEvent
AVS [75]197-Visual---14.2Ads
BBC [6]119Visual--49.7-Doc
OVSD [48]2110Visual--28.9-Generic
Kinetics-GEBD [51]54,691152Audio + Visual---4.9Action
MovieNet [23] †1,1002174Text + Audio + Visual--66.0849.1Movie
RAI [7]10-Visual---98.7News
TI-News [35]477244Audio + Visual-55.6-530.4News
NewsNet (Ours)1,000946Text + Audio + Visual8.551.687.9654.4News
+ +The number of annotations for Scene and Shot is counted from MovieScene [47], which is a subset of MovieNet. + +captioning [3, 34, 63, 71] and retrieval [5, 16, 31, 52]. Temporal video segmentation aims to group successive video frames into short segments along the temporal dimension. With the explosive growth of long-form videos, it is desirable that temporal video segmentation can convert a video into more meaningful segments for more efficient access to the video. However, it is challenging to develop effective temporal segmentation tools for long-form videos, since this requires a comprehensive understanding of video structure, while long-form videos contain complex content. + +Towards temporal video segmentation, existing works explore shot, event, and scene segmentation tasks, respectively. Shot segmentation [38, 55, 57] divides a video into shots, where a shot consists of consecutive and visually continuous frames captured by a camera without interruption [23]. Yet, shot segmentation only considers low-level visual cues (i.e. visual similarity), lacking semantic understanding. Instead, event segmentation [25, 51, 56] divides a video by detecting the moments of changes such as action/subject changes. To better capture the underlying structure of a video, recent works [12, 47, 65] introduce video scene segmentation which segments a video into scene segments, each comprising successive shots semantically related to the same scene. Scene segmentation enables a coarser and higher-level representation than shot segmentation. However, compared to the rich content of massive long-form videos, scene/event is fine-grained and often lacks a high-level summarization of video content, which is insufficient for capturing the complex semantic structure of many videos and briefly representing video content. + +In this work, we first explore how to comprehensively represent the complex structure of a long-form video for temporal video segmentation. Humans can hierarchically divide a video into segments of different granularities according to multi-level semantic information (e.g. scene and topic), from the perspective of cognitive science. Natural language processing researchers have widely explored + +topic-level understanding [29, 41] for summarizing documents [8, 42], while little effort has been devoted to long-form videos. Inspired by these observations, besides scene and event, we propose to introduce two higher-level semantics (i.e. story and topic) into temporal video segmentation, to provide a brief and semantic structure representation. As a result, such hierarchical and multi-level understanding brings about scalable video structure representation for temporal video segmentation on long-form videos. That is, a long-form video can be split into finer segments with lower-level semantics (e.g. scene), but also can be summarized into coarser ones yet with higher-level semantics (e.g. topic) by recursively grouping finer segments, which comprehensively represents video structures from coarse to fine. + +However, the community lacks high-quality datasets to conduct this research. In particular, as shown in Table 1, most datasets only provide temporal structure annotations with regard to events or scenes. TI-News [35] and MovieScene [47] provide two levels of annotations, but these datasets lack topic-level ones. + +To effectively break this limitation, we build a novel large-scale dataset for hierarchical temporal segmentation, named NewsNet. The unique properties of our NewsNet introduce many advantages. First, it is among the largest datasets in the news domain. We collect over 900 hours of videos from 20 mainstream news platforms. It has a highly diverse distribution of data. Second, we carefully annotated it frame-by-frame with 4 hierarchical levels to ensure its quality can meet our needs. Third, it is multimodal, including textual, visual, and audio information. Due to the nature of the news, the alignment across modalities is accurate, which makes multimodal joint training of models feasible. Finally, the videos in NewsNet provide a complete understanding of public events. Compared with other video datasets [4,23,26], it introduces more objective open-world knowledge (e.g., news introduction) while including subjective factual commentary (e.g., host comments on news + +events), making it more amenable to real-life application. + +Based on NewsNet, we empirically highlight two promising directions for long-span temporal segmentation: 1) Infusing Multi-Modality knowledge can significantly improve the performance of long-form temporal segmentation; 2) Although story- and topic-level segmentation is challenging, it can be benefited from hierarchical modeling with the event- and scene-level segmentation tasks. + +The main contributions of this paper are as follows: + +- We propose a novel large-scale dataset NewsNet for long-form video structure understanding. This dataset is derived from $900+$ hours of video and annotated with 4 hierarchical levels of semantics. +- NewsNet provides dense annotations and multi-modal information, promoting diverse benchmarks: separate/hierarchical temporal video segmentation in scene/story/topic levels, as well as other common tasks like classification, video localization/grounding, and highlight detection. +- We formulate a new benchmark, i.e., hierarchical modeling in the temporal segmentation task, which needs a single model to predict segments of multiple hierarchical levels. Based on the empirical study, we bring insights into how hierarchical modeling potentially benefits the temporal video segmentation task, which was almost never discussed. + +# 2. Related Work + +# 2.1. Related Dataset and Benchmark + +To better introduce our work toward hierarchical temporal segmentation, we first review research related to temporal segmentation and video challenges. As shown in Table 1, extensive benchmarks have been proposed to test the performance of temporal video segmentation. More recently, Shou et al. [51] argues that humans can capture the boundary without predefined target classes, and thus proposed Generic Event Boundary Detection (GEBD) as a new benchmark for detecting generic event boundaries on Kinetics-400 [26]. A few works [25, 33, 43-45, 56] have been proposed to attempt to address this challenge, which is a very active sub-field in long-form video understanding [1, 2]. In this paper, as the NewsNet densely annotates all temporal boundaries, including events like GEBD, scenes like MovieScene [47], and stories like TI-News [35]. NewsNet can test the performance of the existing tasks on almost all temporal localization and segmentation tasks. Moreover, with such hierarchical labels, we can further investigate hierarchical modeling in long-form video understanding, which is an important but rarely explored field. + +As an overview of the existing benchmarks in NLP and CV, the temporal annotations are labeled in one or two levels, only focusing on the global abstraction or local understanding. However, the video is naturally hierarchical, which can be recursively represented into sub-units. Therefore, the topic-, story-, scene- and event-level tasks should not be drawn into individual lines, but work in a harmonious way. From such a motivation, we contribute a new dataset called NewsNet with hierarchical dense temporal annotations to the community. Since the annotation cost for different groups is quite different, the scale for topic-level benchmarks [4,9,23,26] is generally larger than that of the other datasets [6,7,35,47,48,75]. This paper tries our best to reach a satisfactory trade-off between scale and the quality of the annotations. The proposed dataset does not only cover 1000 videos with 946 hours but also achieves 654.4 event-level annotations per video, which is very competitive over the existing baselines. Moreover, to further generalize the promotion, textual, audio, and visual input are all collected into the NewsNet for the potential application in CV, NLP and Multi-modality. + +# 2.2. Scope of Related Tasks + +Since dense annotations and rich modalities are provided by NewsNet, we can conduct various applications on top of the datasets. Below, we show the parts of video tasks, which can be carried on the proposed dataset. + +Common Video Tasks. The NewsNet is a high-quality video dataset, to evaluate the methods in common video tasks. For instance, NewsNet is an ideal source for video classification [10, 15, 37, 58, 59, 62] (including the out-of-distribution detection settings [49, 61, 72, 73] due to the existence of category 'Others' in Topic level). NewsNet holds annotations in four granularities for temporal segmentation. By training with the mentioned annotation, the method can completely split a video into several short segments, which will serve partial segmentation tasks like video grounding [40, 70, 74], captioning [3, 34, 71], retrieval [5, 16, 31, 52] and highlight detection [17, 18, 20, 28, 32, 54]. Meanwhile, we simultaneously provide aligned audio, text, and visual form in NewsNet, which might promote the improvement on the task of video grounding [30, 53, 66, 69], where the goal is to retrieve the corresponding video clip by adopting an agnostic or specific query as input. We also note that, over the last few years, the 'pre-training' and 'fine-tuning' paradigms [21, 24, 39, 46, 50, 65, 67, 68, 77] have become revolutionary in both NLP and CV communities. We believe that such technology will be further improved by introducing such a high-quality multi-modality video dataset. + +Temporal Segmentation Approaches. Towards temporal video segmentation, existing works explore shot, scene and story segmentation tasks, respectively. Shot temporal segmentation [6, 11, 38, 55] is employed as a pre-processing to + +![](images/9a691742ff130fad6c75852dd974730328a9a52fb26b33769c9723941dc6fb57.jpg) +Figure 2. Illustration of the four hierarchies. The illustration shows a hierarchical diagram (left) along with an annotation example (right) for the four hierarchies, including Topic, Story, Scene, and Event. In the given sample, a technology (Topic) news video can be divided into several Stories, such as "studio" and "place". Taking "place" as an example, it can be further characterized into different Scenes, where the backgrounds or places are similar. Based on the changes of shots or persons in a scene, the Scene can be divided into atomic units, i.e., Events. + +locate the transition positions in videos based on the similarity of the frames. For the scene segmentation, LGSS [47] converts the problem as a binary classification task at the shot level and applies a boundary-based model that aggregates adjacent shot features within a predefined sliding window to make a prediction. While SCRL [65] introduces a boundary-free model in a seq-to-seq way and reduces the inductive bias introduced by the predefined sliding window. As for story temporal segmentation, TI-NEWS [35] proposes a multi-modal framework based on the boundary-free model to complete the story segmentation and classification tasks simultaneously. + +# 3. Dataset Summary + +In this section, we will describe some details of the NewsNet. First, we will introduce the definition and taxonomy of the four hierarchical levels of semantics and show some annotated examples of them. Then, the statistical information of the proposed dataset will be given to better understand the data distribution of the NewsNet. + +# 3.1. Definition and Taxonomy + +Existing works focus on the shot- or event-level temporal segmentation, while neglecting the long-horizon cases, such as story- and topic-level segmentation. Hence this paper innovatively contributes more uniform benchmarks to the community, which covers all categories of temporal annotation. In order to clarify the meaning of the different labels, we detail them one by one: + +- Event: the atom unit in our setting, which is defined by the switching of the shots or the changing of the person. +- Scene: a combination of several successive shots that + +focus on the same place from different angles. It is defined as a taxonomy-free unit. + +- Story: a sequence of scenes about a piece of news broadcast in a studio or outdoors. Story-level segments can be categorized into 'Connection', 'Place', 'Studio', 'Animation', 'Interview', and 'Photo'. A topic is composed of several stories. +- Topic: it summarizes the content of a long sequence, with the following keywords: 'Health', 'Politics', 'Entertainment', 'Economy', 'Crime', 'Weather', 'Sport', 'Technology', 'Military' and 'Others'. Generally, an individual news video contains several topics. + +The above four different temporal annotations are defined in a hierarchical way. For example, a news video can be divided into several Topic-level segments; a Topic is further composed of several Stories; and so on. The difference between Story and Scene is that Story is category-aware while Scene is not, and the categories in Story are related to the data source. For example, Story contains a category of outdoor place footage, which is unique to news data and usually consists of multiple scenes, as illustrated in Fig. 2. + +# 3.2. Collection and Statistics + +As a large-scale dataset, we collect about 2,000 news broadcast videos from more than 20 different TV stations, of which 1,000 were annotated, and all the data are downloaded from the YouTube and Internet. As shown in Fig. 3, we show the distribution of program duration and sources. News is a kind of typical long-form video, ranging from 10 to 110 minutes. Most of the collected videos have a 1-hour duration, performing as a challenging span for temporal segmentation. During the annotation, we found that + +![](images/a3a23823e90a223ecd619c721749632845ad8bda9746e07f72c52d6adba61058.jpg) +Figure 3. The statistical information of NewsNet. The left histogram shows the distribution of the Top-20 news programs, (i.e., program A to T, by the total number of videos) in different duration, it can be found that the duration of the same program is relatively close. Besides, the two ring charts on the right show the category proportions of Topic and Story. + +there exists a long-tailed distribution for the categories of the story- and topic-level annotation, leading to a practical challenge. We believe such a benchmark can simulate the real-world scenario effectively, so that the competing methods could be benchmarked in a practical manner. More details such as annotation process, visualization, etc., can be found in the Supplementary Material. + +# 4. Benchmark and Protocol Navigation + +In this section, we will first introduce two temporal segmentation task settings, i.e., Separate Modeling and Hierarchical Modeling in NewsNet, with their definitions and protocols. Then, based on the potential relationship between different hierarchies, a simple yet effective loss function is introduced to better tackle hierarchical temporal segmentation. At last, we will briefly introduce the other common video tasks that can be conducted with NewsNet. + +# 4.1. Hierarchical Temporal Segmentation + +As long-form video can be divided into sub-video chunks according to its intrinsic properties, many researchers try to tackle the video segmentation problem from different perspectives in the last few years [12, 35, 47, 65]. Without loss of generality, we classify hierarchical temporal segmentation into two categories: Separate Modeling and Hierarchical Modeling. Separate modeling trains a segmentation model for each level segmentation task, while hierarchical modeling trains a model which simultaneously segments a video into multi-level segments. + +NewsNet supports temporal segmentation tasks with various semantic granularity, i.e., Event, Scene, Story, and Topic. Since many methods [12,47,65], take the Event (aka Shot sometimes) as the basic input unit to reduce the redundancy of the video and computational cost, NewsNet draws the video segmentation tasks on Scene, Story, and Topic. Moreover, these three hierarchies can be strictly aligned according to the Event (Shot) dimension, which is more conducive to Hierarchical Modeling and more convenient for + +comparison. Note that we also carried out the event-level experiment (see Supplementary Material). + +Data and Modality. In-domain: all the videos are randomly split into training, validation, and testing sets with a ratio of 3:1:1. Cross-domain: training, validation, and testing sets are split according to different news program types/IDs; 5 different split combinations are set to ensure the generalization ability of the algorithm can be evaluated properly and fairly. As for the embeddings used in the experiment, the shot-level visual, audio, and textual features are extracted through ResNet50 [22] (pre-trained on Places365 [76]), public Bert-base model [14], and Cnn14 [27] (pre-trained on AudioSet [19]), respectively. + +# 4.1.1 Separate Modeling + +Task Definition. Given a video with the shot labels, the model requires to classify whether a shot is a segmentation point for one of the three semantic levels (Scene, Story and Topic). A segmentation point refers to a shot that lies at the end of a topic/story/scene segment, and its next shot is the start of a new topic/story/scene. + +Backbone. In general, the common paradigm in those segmentation tasks is considering temporal segment to shot-level classification in boundary-base and -free way. We employ the state-of-the-art boundary-free model, SCRL [65], and the boundary-based model, LGSS [47] as our baseline approaches in separate modeling. More details about the protocol can be found in the Supplementary Material. + +# 4.1.2 Hierarchical Modeling + +Task Definition. Given a video with the shot labels, the model requires to classify whether a shot is a segmentation point for three semantic levels (Scene, Story, and Topic). + +Backbone. In this setting, as shown in Fig. 4, we focus on two hierarchical modeling paradigms, including Multi-Label and Multi-Head with a novel learning objective called Hierarchical Ranking. More specifically, Multi + +![](images/c5c92b2b846c82cf48f62c9b53e5d63526865ee810ae4ddf00fbe99f7eec5954.jpg) +Figure 4. An illustration of the separate modeling, two hierarchical modeling paradigms, and the proposed loss function. For the baseline, the model is trained individually corresponding to different segmentation tasks. For hierarchical modeling, all the tasks are aggregated into one model. Specifically, B, L, and N represent the batch size, sequence length, and feature dimension, respectively. + +Label works towards the hierarchical labels from the perspective of multi-label optimization, focusing on the aggregation of information into one head for estimation. While, Multi-Head addresses the challenge as a multi-task problem, where the tasks in different levels should be predicted via the individual pathways. Both pipelines are widely employed in modern architectures [13, 36, 60, 64], performing as simple but representative paradigms. Therefore, we adopt two hierarchical modeling paradigms as the baselines. Hierarchical Ranking Loss. To better model the hierarchical temporal video segmentation task and combine the existing methods, we propose a new loss function called Hierarchical Ranking Loss. The motivation is: due to the accurate alignment between the hierarchical levels with the atom unit Event, as shown in Fig. 1, a positive segment boundary of the higher level task must be a positive segment boundary of the lower task. Hence, the segmentation confidence scores for different tasks should be ranked in a non-increasing order from coarse to fine for the paired task. + +Given two tasks in the hierarchical modeling setting, e.g. task $\mathcal{H}$ and task $\mathcal{L}$ , where the semantic level of task $\mathcal{H}$ is higher than that of task $\mathcal{L}$ , the proposed hierarchical ranking loss function is as follows: + +$$ +\mathcal {P} _ {\mathcal {H}, \mathcal {L}} = \frac {1}{B} \frac {1}{L} \cdot \sum \left[ \sigma \left(F _ {\mathcal {H}} - \mathcal {S G} \left(F _ {\mathcal {L}}\right)\right) \cdot Y _ {\mathcal {L}} \right] \tag {1} +$$ + +where $B$ and $L$ are the batch size and sequence length respectively; $\sigma(\cdot)$ refers to the contraction mapping, like sigmoid function; $F$ stands for the segmentation confidence score of the task; $\mathcal{SG}$ is the Stop Gradient operation for the computational graph, and $Y_{\mathcal{L}}$ corresponds to the label of the task $\mathcal{L}$ . + +And the final objective function for all three tasks can be formed as follow: + +$$ +\mathcal {L} = \mathcal {P} _ {\text {T o p i c}, S t o r y} + \mathcal {P} _ {\text {T o p i c}, S c e n e} + \mathcal {P} _ {\text {S t o r y}, S c e n e} + \mathcal {L} _ {C E} \tag {2} +$$ + +where $\mathcal{L}_{CE}$ is the cross-entropy loss function. + +# 4.2. Common Video Understanding Tasks + +The NewsNet could also apply other common video understanding tasks, e.g., video classification, localization, and highlight detection. More details can be found in Supplementary Material. + +# 5. Experiment and Analysis + +In this section, we mainly focus on the temporal segmentation tasks on the NewsNet and further contribute several benchmarks on the different baselines. In particular, we conduct the empirical study to thoroughly benchmark the effectiveness of model structure, multi-modality information, domain gap, and hierarchical modeling. Note that, due to the limitation on page length, the implementation details and the results for other common video tasks are given in the Supplementary Material. + +Table 2. The performance between boundary-based (B.B.) and boundary-free (B.F) model with visual input on F1 score $\uparrow$ , Precision $\uparrow$ , and Recall $\uparrow$ . Light gray refers to the failure case. + +
TaskModelF1 scorePrecisionRecall
SceneB.B.77.367.790.2
B.F.76.876.177.5
StoryB.B.70.765.976.2
B.F.71.272.370.0
TopicB.B.13.17.0100.0
B.F.62.972.455.6
+ +As an overview, based on the large-scale dataset with rich modalities, we conduct several experiments to respond to five critical questions concerned by the community: + +Table 3. In-domain performance by using boundary-free (B.F.) Table 4. Cross-domain setting by using boundary-free (B.F.) model. The bolded values stand for the optimal performances for model. The bolded values stand for the optimal performances for each task. + +
TaskModalityF1 scorePrecisionRecall
SceneV76.876.177.5
A69.866.873.2
T66.756.381.9
V+A+T78.380.975.8
StoryV71.272.370.0
A59.357.661.1
T50.657.445.2
V+A+T75.474.776.2
TopicV62.972.455.6
A58.159.456.9
T39.046.533.5
V+A+T73.274.372.2
+ +1. Can existing methods also perform well on long-horizon temporal segmentation tasks, such as topic-level and story-level segmentation? +2. Do the different-level tasks test the ability of the model on viewing different temporal spans? +3. Can existing methods well address the semantic gap between low- and high-level granularities? +4. Will the temporal segmentation models be benefited from multi-modality information? +5. How could the models benefit from Hierarchical Modeling? + +# 5.1. Results against Increased Temporal Span + +To investigate the generalization of the existing methods against the long-horizon cases, we conduct an experiment on two different head architectures, including boundary-free and boundary-based heads, on three tasks: scene-level, story-level, and topic-level temporal segmentation. As shown in Fig. 2, the performance of the boundary-based model has a significant drop on the Topic task. Compared with the other tasks, the Topic segmentation task has more time slices for input, leading to a hard challenge caused by longer time dependencies. The methods can segment scenes with about 76-77 F1 score while degrading to 62.9 on Topic-level segmentation. For the first question, the answer is that both the boundary-free model and the boundary-based model can not generalize well in the story-and topic-level segmentation. Combining the overall performance and stability on different tasks, the next experiments will utilize a boundary-free model as the basic backbone under comprehensive consideration. + +# 5.2. Performance against Cross News Domains + +To investigate the generalization of the existing method, we analyze the performance of the method against the va + +
TaskModalityAvg. F1 score (std.)Avg. PrecisionAvg. Recall
SceneV72.9 (2.1)70.975.2
A62.7 (4.0)59.766.6
T61.6 (5.0)52.877.0
V+A+T76.0 (2.1)74.477.9
StoryV68.5 (2.6)70.366.9
A55.7 (3.6)53.659.0
T51.1 (3.6)43.465.4
V+A+T72.9 (2.2)73.772.4
TopicV60.6 (4.7)69.853.8
A59.0 (5.2)56.062.9
T49.8 (5.2)45.755.9
V+A+T72.2 (3.6)72.372.5
+ +rieties of different news programs. As shown in Table 3 and 4, the cross-domain setting slightly degrades the performance on all the tasks. Comparing the in-domain results using the full modality in Table 3 and corresponding cross-domain results in Table 4, the F1 score decreases by 2.3 on task Scene, by 2.5 on task Story, and by 1.0 on task Topic. The reason for this result may be that higher-level tasks, such as Topic, has more abstract semantics and the knowledge is easier to transfer. In contrast, task Scene in the lowest level, generally only considers place switching as the segmentation cue. Besides, the category of Scene is an open set, leading to a greater challenge in the cross-domain setting. Hence, temporal segmentation model also suffers from the out-of-distribution case. Based on the property of the collected dataset, we set the videos from different news programs as different domains, and establish a comprehensive benchmark to facilitate the research on such line. + +# 5.3. Performance using Multi-Modalities + +When it comes to multi-modality, as shown in Table. 4, the model can be benefited from multimodal information among all the tasks. For example, the performance of using only visual modality achieves only 60.6 recall in topic-level segmentation. However, by aggregating the other modalities including audio and text, the performance can be improved significantly (72.2 vs 49.8), which indicates the potential value for addressing long-discrepancy via multimodality technology. Note that, the improvement in Scene segmentation is, to some extent, limited, we believe that Scene is mainly defined by visual input, hence can not obtain adequate complementarity from other modalities. + +# 5.4. Performance based on Hierarchical Modeling + +Here, we discuss the core challenge revealed in this paper: 'How could the model benefit from hierarchical annotations?' In this case, the baseline is set as the model, which can only be accessible to one specific-level annotation. As + +Table 5. The F1 scores of baselines trained with different levels of annotations on full modalities without our hierarchical ranking loss, where blue and orange indicate the in-domain and cross-domain, respectively. Each row refers to the result corresponding to a single task. Hie. Modeling stands for Hierarchical Modeling while Sep. Modeling is Separate Modeling. + +
RecipeBaseline +Sep. ModelingMulti-Label +Hie. ModelingMulti-Head +Hie. Modeling
Scene78.3 / 76.079.1 / 76.579.9 / 76.9
+ Story75.4 / 72.975.4 / 74.774.2 / 74.0
Scene78.3 / 76.079.8 / 76.479.5 / 76.5
+ Topic73.2 / 72.270.5 / 72.870.9 / 73.0
Story75.4 / 72.976.2 / 74.375.4 / 73.9
+ Topic73.2 / 72.277.3 / 73.275.2 / 73.5
Scene78.3 / 76.077.4 / 76.879.8 / 76.8
+ Story75.4 / 72.974.3 / 74.374.5 / 73.7
+ Topic73.2 / 72.274.3 / 72.676.6 / 70.4
+ +listed in Table 5, we consider two simple strategies, including Multi-Label and Multi-Head, to aggregate the hierarchical information. By training with multi-level annotations, a $1.0 \sim 3.0\%$ improvement can be achieved, which indicates, the segmentation can be benefited by introducing the hierarchical structures of videos. However, Table 5 also shows directly combining different-level granularities might not result in gains due to the semantic gap between low- and high-level granularities. For example, directly combining scene- and topic-level information by Multi-Label hierarchical modeling or Multi-Head hierarchical modeling does not consistently improve the topic-level segmentation performance. This poses a new and challenging research question, i.e. how to jointly train with various granularities toward hierarchical modeling of long-form videos? + +To address the above question, we introduce a simple but effective loss function, i.e., Hierarchical Ranking Loss introduced in Sec. 4.1.2. As shown in Table 6, by combining with our proposed loss, the F1 score for segmentation can be further improved in most cases. Typically, in Topic-level, Multi-Head coupled with the proposed loss can reach the best performance, i.e., improving Multi-Head case from 70.4 to 73.2 F1 score. + +# 5.5. Other Benchmarks involved in the NewsNet + +In addition to the temporal segmentation tasks, we also conduct extensive common video tasks, including highlight detection, video classification, and video localization, on the proposed dataset. Due to the limited pages, we show this part in the Supplementary Materials. + +Table 6. The F1 scores of the methods with or without hierarchical ranking loss under the in-domain / cross-domain setting on full modalities. Hie. stands for Hierarchical Modeling and Sep. refers to Separate Modeling. + +
MethodSceneStoryTopic
Baseline (Sep.)78.3 / 76.075.4 / 72.973.2 / 72.2
Multi-Label (Hie.)77.4 / 76.874.3 / 74.374.3 / 72.6
Multi-Label w/ Hie. Loss (Hie.)79.6 / 76.974.4 / 73.577.8 / 73.1
Multi-Head (Hie.)79.8 / 76.874.5 / 73.776.6 / 70.4
Multi-Head w/ Hie. Loss (Hie.)80.3 / 76.976.3 / 74.676.5 / 73.2
+ +# 6. Conclusion, Limitation and Future work + +We introduce NewsNet, a new large-scale dataset for temporal video segmentation. Compared with existing large-scale datasets for temporal video segmentation, NewsNet additionally provides two abstractive levels of temporal segmentation, which have not been taken into account by these datasets. The four-level and hierarchical annotations enable the community to explore how to comprehensively represent the complex structure of long-form videos from coarse to fine. Extensive efforts of human workers are devoted, so as to ensure that annotations are high-quality. We hope that NewsNet and baseline results can facilitate the development of temporal video segmentation in terms of insightful research and practical tools. In addition, as NewsNet provides diverse annotations for dividing a video into segments of different granularities as well as category labels, it can also serve other video understanding tasks such as video grounding, captioning, retrieval, highlight detection, and video classification. + +In this paper, we only focus on the problem of hierarchical temporal segmentation. We have not explored other multi-modality tasks, such as visual reasoning and video question answering, while our dataset can serve these tasks thanks to its multi-modality nature. We will explore these tasks in our future work to further promote advances in multi-modality-based video understanding and reasoning. + +# Acknowledgement + +This work was supported by the KAUST Office of Sponsored Research through the Visual Computing Center (VCC) funding, as well as, the SDAIA-KAUST Center of Excellence in Data Science and Artificial Intelligence (SDAIA-KAUST AI). Raghavendra Ramachandra is supported under the SALT and OFFPAD projects funded by the Research Council of Norway, Norway. Chia-Wen Lin is supported by National Science and Technology Council, Taiwan, under project NSTC 111-2634-F002-023. + +# References + +[1] LOVEU21. https://sites.google.com/view/loveucvpr21, 2021.3 +[2] LOVEU22 program schedule. https://sites.google.com/view/loveucvpr22/program, 2022.3 +[3] Nayyer Aafaq, Naveed Akhtar, Wei Liu, Syed Zulqarnain Gilani, and Ajmal Mian. Spatio-temporal dynamics and semantic attribute enriched visual encoding for video captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12487-12496, 2019. 2, 3 +[4] Sami Abu-El-Haija, Nisarg Kothari, Joonseok Lee, Paul Natsev, George Toderici, Balakrishnan Varadarajan, and Sudheendra Vijayanarasimhan. Youtube-8m: A large-scale video classification benchmark. arXiv preprint arXiv:1609.08675, 2016. 2, 3 +[5] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1728-1738, 2021. 2, 3 +[6] Lorenzo Baraldi, Costantino Grana, and Rita Cucchiara. A deep siamese network for scene detection in broadcast videos. In Proceedings of the 23rd ACM international conference on Multimedia, pages 1199-1202, 2015. 2, 3 +[7] Lorenzo Baraldi, Costantino Grana, and Rita Cucchiara. Shot and scene detection via hierarchical clustering for re-using broadcast video. In International conference on computer analysis of images and patterns, pages 801-811. Springer, 2015. 2, 3 +[8] Joe Barrow, Rajiv Jain, Vlad Morariu, Varun Manjunatha, Douglas W Oard, and Philip Resnik. A joint model for document segmentation and segment labeling. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 313-322, 2020. 2 +[9] Fabian Caba Heilbron, Victor Escorcia, Bernard Ghanem, and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 961-970, 2015. 3 +[10] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 1, 3 +[11] Vasileios T Chasanis, Aristidis C Likas, and Nikolaos P Galatsanos. Scene detection in videos using shot clustering and sequence alignment. IEEE transactions on multimedia, 11(1):89-100, 2008. 3 +[12] Shixing Chen, Xiaohan Nie, David Fan, Dongqing Zhang, Vimal Bhat, and Raffay Hamid. Shot contrastive self-supervised learning for scene boundary detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9796-9805, 2021. 2, 5 +[13] Zhao-Min Chen, Xiu-Shen Wei, Peng Wang, and Yanwen Guo. Multi-label image recognition with graph convolutional networks. In Proceedings of the IEEE/CVF Confer- + +ence on Computer Vision and Pattern Recognition (CVPR), June 2019. 6 +[14] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 5 +[15] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6202-6211, 2019. 1, 3 +[16] Valentin Gabeur, Chen Sun, Karteek Alahari, and Cordelia Schmid. Multi-modal transformer for video retrieval. In European Conference on Computer Vision, pages 214-229. Springer, 2020. 2, 3 +[17] Jiyang Gao, Chen Sun, Zhenheng Yang, and Ram Nevatia. Tall: Temporal activity localization via language query. In Proceedings of the IEEE international conference on computer vision, pages 5267-5275, 2017. 3 +[18] Ana Garcia del Molino and Michael Gygli. PhD-gifs: personalized highlight detection for automatic gif creation. In Proceedings of the 26th ACM international conference on Multimedia, pages 600–608, 2018. 3 +[19] Jort F Gemmeke, Daniel PW Ellis, Dylan Freedman, Aren Jansen, Wade Lawrence, R Channing Moore, Manoj Plakal, and Marvin Ritter. Audio set: An ontology and human-labeled dataset for audio events. In 2017 IEEE international conference on acoustics, speech and signal processing (ICASSP), pages 776-780. IEEE, 2017. 5 +[20] Michael Gygli, Yale Song, and Liangliang Cao. Video2gif: Automatic generation of animated gifs from video. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1001-1009, 2016. 3 +[21] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 3 +[22] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5 +[23] Qingqiu Huang, Yu Xiong, Anyi Rao, Jiaze Wang, and Dahua Lin. Movienet: A holistic dataset for movie understanding. In European Conference on Computer Vision, pages 709-727. Springer, 2020. 2, 3 +[24] Chen Ju, Tengda Han, Kunhao Zheng, Ya Zhang, and Weidi Xie. Prompting visual-language models for efficient video understanding. arXiv preprint arXiv:2112.04478, 2021. 3 +[25] Hyolim Kang, Jinwoo Kim, Taehyun Kim, and Seon Joo Kim. Uboco: Unsupervised boundary contrastive learning for generic event boundary detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20073-20082, 2022. 2, 3 +[26] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 2, 3 + +[27] Qiuqiang Kong, Yin Cao, Turab Iqbal, Yuxuan Wang, Wenwu Wang, and Mark D Plumbley. Panns: Large-scale pretrained audio neural networks for audio pattern recognition. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 28:2880-2894, 2020. 5 +[28] Ranjay Krishna, Kenji Hata, Frederic Ren, Li Fei-Fei, and Juan Carlos Niebles. Dense-captioning events in videos. In Proceedings of the IEEE international conference on computer vision, pages 706–715, 2017. 3 +[29] Kathy Lee, Diana Palsetia, Ramanathan Narayanan, Md Mostofa Ali Patwary, Ankit Agrawal, and Alok Choudhary. Twitter trending topic classification. In 2011 IEEE 11th International Conference on Data Mining Workshops, pages 251-258. IEEE, 2011. 2 +[30] Jie Lei, Tamara L Berg, and Mohit Bansal. Detecting moments and highlights in videos via natural language queries. Advances in Neural Information Processing Systems, 34:11846-11858, 2021. 3 +[31] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L Berg, Mohit Bansal, and Jingjing Liu. Less is more: Clipbert for video-and-language learning via sparse sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7331-7341, 2021. 2, 3 +[32] Jie Lei, Licheng Yu, Tamara L Berg, and Mohit Bansal. Tvr: A large-scale dataset for video-subtitle moment retrieval. In European Conference on Computer Vision, pages 447-463. Springer, 2020. 3 +[33] Congcong Li, Xinyao Wang, Longyin Wen, Dxiang Hong, Tiejian Luo, and Libo Zhang. End-to-end compressed video representation learning for generic event boundary detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13967-13976, 2022. 3 +[34] Kevin Lin, Linjie Li, Chung-Ching Lin, Faisal Ahmed, Zhe Gan, Zicheng Liu, Yumao Lu, and Lijuan Wang. Swinbert: End-to-end transformers with sparse attention for video captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17949-17958, 2022. 2, 3 +[35] Ye Liu, Lingfeng Qiao, Di Yin, Zhuoxuan Jiang, Xinghua Jiang, Deqiang Jiang, and Bo Ren. Os-msl: One stage multimodal sequential link framework for scene segmentation and classification. In Proceedings of the 30th ACM International Conference on Multimedia, pages 6269-6277, 2022. 2, 3, 4, 5 +[36] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10012-10022, October 2021. 6 +[37] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swim transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3202-3211, 2022. 1, 3 +[38] Jakub Lokoc, Gregor Kovalcik, Tomáš Souček, Jaroslav Moravec, and Přemysl Čech. A framework for effective + +known-item search in video. In Proceedings of the 27th ACM International Conference on Multimedia, pages 1777-1785, 2019. 2, 3 +[39] Sauradip Nag, Xiatian Zhu, Yi-Zhe Song, and Tao Xiang. Zero-shot temporal action detection via vision-language prompting. arXiv preprint arXiv:2207.08184, 2022. 3 +[40] Guoshun Nan, Rui Qiao, Yao Xiao, Jun Liu, Sicong Leng, Hao Zhang, and Wei Lu. Interventional video grounding with dual contrastive learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2765-2775, 2021. 3 +[41] Fahmi Salman Nurfikri, Mohamad Syahrul Mubarok, et al. News topic classification using mutual information and bayesian network. In 2018 6th International Conference on Information and Communication Technology (ICoICT), pages 162-166. IEEE, 2018. 2 +[42] Sofia Ares Oliveira, Benoit Seguin, and Frederic Kaplan. dhsegment: A generic deep-learning approach for document segmentation. In 2018 16th International Conference on Frontiers in Handwriting Recognition (ICFHR), pages 7-12. IEEE, 2018. 2 +[43] Alejandro Pardo, Fabian Caba Heilbron, Juan León Alcazar, Ali Thabet, and Bernard Ghanem. Moviecuts: A new dataset and benchmark for cut type recognition. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part VII, pages 668-685, 2022. 3 +[44] Will Price, Carl Vondrick, and Dima Damen. Unweavenet: Unweaving activity stories. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13770-13779, 2022. 3 +[45] Zhiwu Qing, Shiwei Zhang, Ziyuan Huang, Yi Xu, Xiang Wang, Mingqian Tang, Changxin Gao, Rong Jin, and Nong Sang. Learning from untrimmed videos: Self-supervised video representation learning with hierarchical consistency. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13821-13831, 2022. 3 +[46] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 3 +[47] Anyi Rao, Linning Xu, Yu Xiong, Guodong Xu, Qingqiu Huang, Bolei Zhou, and Dahua Lin. A local-to-global approach to multi-modal movie scene segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10146-10155, 2020. 2, 3, 4, 5 +[48] Daniel Rotman, Dror Porat, and Gal Ashour. Optimal sequential grouping for robust video scene detection using multiple modalities. International Journal of Semantic Computing, 11(02):193-208, 2017. 2, 3 +[49] Ellen Rushe and Brian Mac Namee. Anomaly detection in raw audio using deep autoregressive networks. In ICASSP 2019-2019 IEEE International Conference on Acoustics, + +Speech and Signal Processing (ICASSP), pages 3597-3601. IEEE, 2019. 3 +[50] Weimin Shi, Mingchen Zhuge, Zhong Zhou, Dehong Gao, and Deng-Ping Fan. Qr-clip: Introducing explicit open-world knowledge for location and time reasoning. arXiv preprint arXiv:2302.00952, 2023. 3 +[51] Mike Zheng Shou, Stan Weixian Lei, Weiyao Wang, Deepti Ghadiyaram, and Matt Feiszli. Generic event boundary detection: A benchmark for event segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8075-8084, 2021. 2, 3 +[52] Nina Shvetsova, Brian Chen, Andrew Rouditchenko, Samuel Thomas, Brian Kingsbury, Rogerio S Feris, David Harwath, James Glass, and Hilde Kuehne. Everything at once-multimodal fusion transformer for video retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20020-20029, 2022. 2, 3 +[53] Mattia Soldan, Alejandro Pardo, Juan León Alcazar, Fabian Caba, Chen Zhao, Silvio Giancola, and Bernard Ghanem. Mad: A scalable dataset for language grounding in videos from movie audio descriptions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5026-5035, 2022. 3 +[54] Yale Song, Miriam Redi, Jordi Vallmitjana, and Alejandro Jaimes. To click or not to click: Automatic selection of beautiful thumbnails from videos. In Proceedings of the 25th ACM international on conference on information and knowledge management, pages 659-668, 2016. 3 +[55] Tomáš Souček and Jakub Lokoc. Transnet v2: an effective deep network architecture for fast shot transition detection. arXiv preprint arXiv:2008.04838, 2020. 2, 3 +[56] Jiaqi Tang, Zhaoyang Liu, Chen Qian, Wayne Wu, and Limin Wang. Progressive attention on multi-level dense difference maps for generic event boundary detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3355-3364, 2022. 2, 3 +[57] Shitao Tang, Litong Feng, Zhanghui Kuang, Yimin Chen, and Wei Zhang. Fast video shot transition localization with deep structured models. In Asian Conference on Computer Vision, pages 577-592. Springer, 2018. 2 +[58] Du Tran, Lubomir Bourdev, Rob Fergus, Lorenzo Torresani, and Manohar Paluri. Learning spatiotemporal features with 3d convolutional networks. In Proceedings of the IEEE international conference on computer vision, pages 4489-4497, 2015. 1, 3 +[59] Du Tran, Heng Wang, Lorenzo Torresani, and Matt Feiszli. Video classification with channel-separated convolutional networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5552-5561, 2019. 1, 3 +[60] Dongkai Wang and Shiliang Zhang. Unsupervised person re-identification via multi-label classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 6 +[61] Jue Wang and Anoop Cherian. Gods: Generalized one-class discriminative subspaces for anomaly detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8201-8211, 2019. 3 + +[62] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaou Tang, and Luc Van Gool. Temporal segment networks: Towards good practices for deep action recognition. In European conference on computer vision, pages 20-36. Springer, 2016. 1, 3 +[63] Yuxuan Wang, Difei Gao, Licheng Yu, Weixian Lei, Matt Feiszli, and Mike Zheng Shou. Geb+: A benchmark for generic event boundary captioning, grounding and retrieval. In European Conference on Computer Vision, pages 709-725. Springer, 2022. 2 +[64] Jonatas Wehrmann, Ricardo Cerri, and Rodrigo Barros. Hierarchical multi-label classification networks. In Jennifer Dy and Andreas Krause, editors, Proceedings of the 35th International Conference on Machine Learning, volume 80 of Proceedings of Machine Learning Research, pages 5075-5084. PMLR, 10-15 Jul 2018. 6 +[65] Haoqian Wu, Keyu Chen, Yanan Luo, Ruizhi Qiao, Bo Ren, Haozhe Liu, Weicheng Xie, and Linlin Shen. Scene consistency representation learning for video scene segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14021-14030, 2022. 2, 3, 4, 5 +[66] Bo Xiong, Yannis Kalantidis, Deepti Ghadiyaram, and Kristen Grauman. Less is more: Learning highlight detection from video duration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1258-1267, 2019. 3 +[67] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. Videoclip: Contrastive pre-training for zero-shot video-text understanding. arXiv preprint arXiv:2109.14084, 2021. 3 +[68] Mengmeng Xu, Juan-Manuel Pérez-Rúa, Victor Escorcia, Brais Martinez, Xiatian Zhu, Li Zhang, Bernard Ghanem, and Tao Xiang. Boundary-sensitive pre-training for temporal localization in videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7220-7230, 2021. 3 +[69] Minghao Xu, Hang Wang, Bingbing Ni, Riheng Zhu, Zhenbang Sun, and Changhu Wang. Cross-category video highlight detection via set-based learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7970-7979, 2021. 3 +[70] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Tubedetr: Spatio-temporal video grounding with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16442-16453, 2022. 3 +[71] Hanhua Ye, Guorong Li, Yuankai Qi, Shuhui Wang, Qingming Huang, and Ming-Hsuan Yang. Hierarchical modular network for video captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17939-17948, 2022. 2, 3 +[72] Muhammad Zaigham Zaheer, Arif Mahmood, Marcella Astrid, and Seung-Ik Lee. Claws: Clustering assisted weakly supervised learning with normalcy suppression for anomalous event detection. In European Conference on Computer Vision, pages 358-376. Springer, 2020. 3 + +[73] M Zaigham Zaheer, Arif Mahmood, M Haris Khan, Mattia Segu, Fisher Yu, and Seung-Ik Lee. Generative cooperative learning for unsupervised video anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14744-14754, 2022. 3 +[74] Runhao Zeng, Haoming Xu, Wenbing Huang, Peihao Chen, Mingkui Tan, and Chuang Gan. Dense regression network for video grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10287-10296, 2020. 3 +[75] Haoxin Zhang, Zhimin Li, and Qinglin Lu. Better learning shot boundary detection via multi-task. In Proceedings of the 29th ACM International Conference on Multimedia, pages 4730-4734, 2021. 2, 3 +[76] Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. IEEE transactions on pattern analysis and machine intelligence, 40(6):1452-1464, 2017. 5 +[77] Mingchen Zhuge, Dehong Gao, Deng-Ping Fan, Linbo Jin, Ben Chen, Haoming Zhou, Minghui Qiu, and Ling Shao. Kaleido-bert: Vision-language pre-training on fashion domain. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12647-12657, 2021. 3 \ No newline at end of file diff --git a/2023/NewsNet_ A Novel Dataset for Hierarchical Temporal Segmentation/images.zip b/2023/NewsNet_ A Novel Dataset for Hierarchical Temporal Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b6d6d8014203350ce26715d79dc78d1efd043b27 --- /dev/null +++ b/2023/NewsNet_ A Novel Dataset for Hierarchical Temporal Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ee8893b03b44af30e5c63e0d09e2b2b8451f677c88606cd76ab9a846268e905 +size 520330 diff --git a/2023/NewsNet_ A Novel Dataset for Hierarchical Temporal Segmentation/layout.json b/2023/NewsNet_ A Novel Dataset for Hierarchical Temporal Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..538220485fb903e9417936c0259da6522ca8ccf6 --- /dev/null +++ b/2023/NewsNet_ A Novel Dataset for Hierarchical Temporal Segmentation/layout.json @@ -0,0 +1,7984 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 87, + 103, + 506, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 103, + 506, + 121 + ], + "spans": [ + { + "bbox": [ + 87, + 103, + 506, + 121 + ], + "type": "text", + "content": "NewsNet: A Novel Dataset for Hierarchical Temporal Segmentation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 52, + 142, + 539, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 142, + 539, + 185 + ], + "spans": [ + { + "bbox": [ + 52, + 142, + 539, + 185 + ], + "type": "text", + "content": "Haoqian " + }, + { + "bbox": [ + 52, + 142, + 539, + 185 + ], + "type": "inline_equation", + "content": "\\mathbf{W}\\mathbf{u}^{1\\dagger}" + }, + { + "bbox": [ + 52, + 142, + 539, + 185 + ], + "type": "text", + "content": ", Keyu Chen" + }, + { + "bbox": [ + 52, + 142, + 539, + 185 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 52, + 142, + 539, + 185 + ], + "type": "text", + "content": ", Haozhe Liu" + }, + { + "bbox": [ + 52, + 142, + 539, + 185 + ], + "type": "inline_equation", + "content": "^{2\\dagger}" + }, + { + "bbox": [ + 52, + 142, + 539, + 185 + ], + "type": "text", + "content": ", Mingchen Zhuge" + }, + { + "bbox": [ + 52, + 142, + 539, + 185 + ], + "type": "inline_equation", + "content": "^{2\\dagger}" + }, + { + "bbox": [ + 52, + 142, + 539, + 185 + ], + "type": "text", + "content": ", Bing Li" + }, + { + "bbox": [ + 52, + 142, + 539, + 185 + ], + "type": "inline_equation", + "content": "^{2\\boxtimes}" + }, + { + "bbox": [ + 52, + 142, + 539, + 185 + ], + "type": "text", + "content": ", Ruizhi Qiao" + }, + { + "bbox": [ + 52, + 142, + 539, + 185 + ], + "type": "inline_equation", + "content": "^{1\\boxtimes}" + }, + { + "bbox": [ + 52, + 142, + 539, + 185 + ], + "type": "text", + "content": ", Xiujun Shu" + }, + { + "bbox": [ + 52, + 142, + 539, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 52, + 142, + 539, + 185 + ], + "type": "text", + "content": ", Bei Gan" + }, + { + "bbox": [ + 52, + 142, + 539, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 52, + 142, + 539, + 185 + ], + "type": "text", + "content": ", Liangsheng Xu" + }, + { + "bbox": [ + 52, + 142, + 539, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 52, + 142, + 539, + 185 + ], + "type": "text", + "content": ", Bo Ren" + }, + { + "bbox": [ + 52, + 142, + 539, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 52, + 142, + 539, + 185 + ], + "type": "text", + "content": ", Mengmeng Xu" + }, + { + "bbox": [ + 52, + 142, + 539, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 52, + 142, + 539, + 185 + ], + "type": "text", + "content": ", Wentian Zhang" + }, + { + "bbox": [ + 52, + 142, + 539, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 52, + 142, + 539, + 185 + ], + "type": "text", + "content": ", Raghavendra Ramachandra" + }, + { + "bbox": [ + 52, + 142, + 539, + 185 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 52, + 142, + 539, + 185 + ], + "type": "text", + "content": ", Chia-Wen Lin" + }, + { + "bbox": [ + 52, + 142, + 539, + 185 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 52, + 142, + 539, + 185 + ], + "type": "text", + "content": ", Bernard Ghanem" + }, + { + "bbox": [ + 52, + 142, + 539, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 81, + 190, + 512, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 190, + 512, + 234 + ], + "spans": [ + { + "bbox": [ + 81, + 190, + 512, + 234 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 81, + 190, + 512, + 234 + ], + "type": "text", + "content": " Tencent " + }, + { + "bbox": [ + 81, + 190, + 512, + 234 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 81, + 190, + 512, + 234 + ], + "type": "text", + "content": " AI Initiative, King Abdullah University of Science and Technology (KAUST) \n" + }, + { + "bbox": [ + 81, + 190, + 512, + 234 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 81, + 190, + 512, + 234 + ], + "type": "text", + "content": " Norwegian University of Science and Technology (NTNU) \n" + }, + { + "bbox": [ + 81, + 190, + 512, + 234 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 81, + 190, + 512, + 234 + ], + "type": "text", + "content": " National Tsing Hua University (NTHU)" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 58, + 244, + 541, + 423 + ], + "blocks": [ + { + "bbox": [ + 58, + 244, + 541, + 423 + ], + "lines": [ + { + "bbox": [ + 58, + 244, + 541, + 423 + ], + "spans": [ + { + "bbox": [ + 58, + 244, + 541, + 423 + ], + "type": "image", + "image_path": "573668c2860b1009046c02f06c69b022fdb789c6a1705bb0ea19027ffd604e94.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 431, + 546, + 487 + ], + "lines": [ + { + "bbox": [ + 45, + 431, + 546, + 487 + ], + "spans": [ + { + "bbox": [ + 45, + 431, + 546, + 487 + ], + "type": "text", + "content": "Figure 1. Examples from the proposed NewsNet. The dataset provides multimodal and hierarchical labels on each video, including 4-level hierarchical temporal information (i.e., event, scene, story, topic) and 3 modality-specific cues (i.e., video, audio, text). A prerequisite motivation behind NewsNet is that long-form videos can be recursively represented into several sub-videos according to the semantics, which is an intrinsic property of the video but rarely explored in the community. To bridge this research gap, this paper innovatively constructs a large-scale dataset to benchmark the algorithms for hierarchical understanding like a human being." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 164, + 498, + 212, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 498, + 212, + 511 + ], + "spans": [ + { + "bbox": [ + 164, + 498, + 212, + 511 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 525, + 289, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 525, + 289, + 670 + ], + "spans": [ + { + "bbox": [ + 46, + 525, + 289, + 670 + ], + "type": "text", + "content": "Temporal video segmentation is the get-to-go automatic video analysis, which decomposes a long-form video into smaller components for the following-up understanding tasks. Recent works have studied several levels of granularity to segment a video, such as shot, event, and scene. Those segmentations can help compare the semantics in the corresponding scales, but lack a wider view of larger temporal spans, especially when the video is complex and structured. Therefore, we present two abstractive levels of temporal segmentations and study their hierarchy to the existing fine-grained levels. Accordingly, we collect NewsNet, the largest news video dataset consisting of 1,000 videos" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 499, + 547, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 499, + 547, + 631 + ], + "spans": [ + { + "bbox": [ + 306, + 499, + 547, + 631 + ], + "type": "text", + "content": "in over 900 hours, associated with several tasks for hierarchical temporal video segmentation. Each news video is a collection of stories on different topics, represented as aligned audio, visual, and textual data, along with extensive frame-wise annotations in four granularities. We assert that the study on NewsNet can advance the understanding of complex structured video and benefit more areas such as short-video creation, personalized advertisement, digital instruction, and education. Our dataset and code is publicly available at https://github.com/NewsNet-Benchmark/NewsNet." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 656, + 387, + 669 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 656, + 387, + 669 + ], + "spans": [ + { + "bbox": [ + 306, + 656, + 387, + 669 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 547, + 715 + ], + "type": "text", + "content": "Temporal video segmentation is a critical problem in video understanding, which is essential for many video applications such as video classification [10, 15, 37, 58, 59, 62]," + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 38 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 38 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 38 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 684, + 132, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 684, + 132, + 693 + ], + "spans": [ + { + "bbox": [ + 62, + 684, + 132, + 693 + ], + "type": "text", + "content": "† Equal Contribution" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 694, + 287, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 694, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 694, + 287, + 713 + ], + "type": "text", + "content": "Corresponding Authors: Bing Li (bing.li@kaust.edu.sa) and Ruizhi Qiao (ruizhiqiao@tencent.com)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "10669" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 67, + 102, + 526, + 234 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 547, + 95 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 547, + 95 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 547, + 95 + ], + "type": "text", + "content": "Table 1. Data comparison between the NewsNet and other related datasets. The NewsNet provides various multimodal data and hierarchical temporal segmentation annotations. Doc: documentary, Ads: advertisement. (Please refer to our project page for more details.)" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 67, + 102, + 526, + 234 + ], + "lines": [ + { + "bbox": [ + 67, + 102, + 526, + 234 + ], + "spans": [ + { + "bbox": [ + 67, + 102, + 526, + 234 + ], + "type": "table", + "html": "
Dataset# VideoDuration (hours)Modality# Annotation(s) per VideoSource
TopicStorySceneEvent
AVS [75]197-Visual---14.2Ads
BBC [6]119Visual--49.7-Doc
OVSD [48]2110Visual--28.9-Generic
Kinetics-GEBD [51]54,691152Audio + Visual---4.9Action
MovieNet [23] †1,1002174Text + Audio + Visual--66.0849.1Movie
RAI [7]10-Visual---98.7News
TI-News [35]477244Audio + Visual-55.6-530.4News
NewsNet (Ours)1,000946Text + Audio + Visual8.551.687.9654.4News
", + "image_path": "5a983113068e6f41caad4e4ca5069e3d53aa340446276b1ea9f97ec3e0e73408.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 63, + 235, + 438, + 245 + ], + "lines": [ + { + "bbox": [ + 63, + 235, + 438, + 245 + ], + "spans": [ + { + "bbox": [ + 63, + 235, + 438, + 245 + ], + "type": "text", + "content": "The number of annotations for Scene and Shot is counted from MovieScene [47], which is a subset of MovieNet." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 264, + 289, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 264, + 289, + 385 + ], + "spans": [ + { + "bbox": [ + 46, + 264, + 289, + 385 + ], + "type": "text", + "content": "captioning [3, 34, 63, 71] and retrieval [5, 16, 31, 52]. Temporal video segmentation aims to group successive video frames into short segments along the temporal dimension. With the explosive growth of long-form videos, it is desirable that temporal video segmentation can convert a video into more meaningful segments for more efficient access to the video. However, it is challenging to develop effective temporal segmentation tools for long-form videos, since this requires a comprehensive understanding of video structure, while long-form videos contain complex content." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 387, + 289, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 387, + 289, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 387, + 289, + 628 + ], + "type": "text", + "content": "Towards temporal video segmentation, existing works explore shot, event, and scene segmentation tasks, respectively. Shot segmentation [38, 55, 57] divides a video into shots, where a shot consists of consecutive and visually continuous frames captured by a camera without interruption [23]. Yet, shot segmentation only considers low-level visual cues (i.e. visual similarity), lacking semantic understanding. Instead, event segmentation [25, 51, 56] divides a video by detecting the moments of changes such as action/subject changes. To better capture the underlying structure of a video, recent works [12, 47, 65] introduce video scene segmentation which segments a video into scene segments, each comprising successive shots semantically related to the same scene. Scene segmentation enables a coarser and higher-level representation than shot segmentation. However, compared to the rich content of massive long-form videos, scene/event is fine-grained and often lacks a high-level summarization of video content, which is insufficient for capturing the complex semantic structure of many videos and briefly representing video content." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": "In this work, we first explore how to comprehensively represent the complex structure of a long-form video for temporal video segmentation. Humans can hierarchically divide a video into segments of different granularities according to multi-level semantic information (e.g. scene and topic), from the perspective of cognitive science. Natural language processing researchers have widely explored" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 264, + 547, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 264, + 547, + 434 + ], + "spans": [ + { + "bbox": [ + 304, + 264, + 547, + 434 + ], + "type": "text", + "content": "topic-level understanding [29, 41] for summarizing documents [8, 42], while little effort has been devoted to long-form videos. Inspired by these observations, besides scene and event, we propose to introduce two higher-level semantics (i.e. story and topic) into temporal video segmentation, to provide a brief and semantic structure representation. As a result, such hierarchical and multi-level understanding brings about scalable video structure representation for temporal video segmentation on long-form videos. That is, a long-form video can be split into finer segments with lower-level semantics (e.g. scene), but also can be summarized into coarser ones yet with higher-level semantics (e.g. topic) by recursively grouping finer segments, which comprehensively represents video structures from coarse to fine." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 435, + 547, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 435, + 547, + 508 + ], + "spans": [ + { + "bbox": [ + 304, + 435, + 547, + 508 + ], + "type": "text", + "content": "However, the community lacks high-quality datasets to conduct this research. In particular, as shown in Table 1, most datasets only provide temporal structure annotations with regard to events or scenes. TI-News [35] and MovieScene [47] provide two levels of annotations, but these datasets lack topic-level ones." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 510, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 548, + 715 + ], + "type": "text", + "content": "To effectively break this limitation, we build a novel large-scale dataset for hierarchical temporal segmentation, named NewsNet. The unique properties of our NewsNet introduce many advantages. First, it is among the largest datasets in the news domain. We collect over 900 hours of videos from 20 mainstream news platforms. It has a highly diverse distribution of data. Second, we carefully annotated it frame-by-frame with 4 hierarchical levels to ensure its quality can meet our needs. Third, it is multimodal, including textual, visual, and audio information. Due to the nature of the news, the alignment across modalities is accurate, which makes multimodal joint training of models feasible. Finally, the videos in NewsNet provide a complete understanding of public events. Compared with other video datasets [4,23,26], it introduces more objective open-world knowledge (e.g., news introduction) while including subjective factual commentary (e.g., host comments on news" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "10670" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 278, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 278, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 278, + 83 + ], + "type": "text", + "content": "events), making it more amenable to real-life application." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 85, + 286, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 85, + 286, + 168 + ], + "spans": [ + { + "bbox": [ + 46, + 85, + 286, + 168 + ], + "type": "text", + "content": "Based on NewsNet, we empirically highlight two promising directions for long-span temporal segmentation: 1) Infusing Multi-Modality knowledge can significantly improve the performance of long-form temporal segmentation; 2) Although story- and topic-level segmentation is challenging, it can be benefited from hierarchical modeling with the event- and scene-level segmentation tasks." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 59, + 170, + 267, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 170, + 267, + 182 + ], + "spans": [ + { + "bbox": [ + 59, + 170, + 267, + 182 + ], + "type": "text", + "content": "The main contributions of this paper are as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 193, + 286, + 419 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 59, + 193, + 286, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 193, + 286, + 239 + ], + "spans": [ + { + "bbox": [ + 59, + 193, + 286, + 239 + ], + "type": "text", + "content": "- We propose a novel large-scale dataset NewsNet for long-form video structure understanding. This dataset is derived from " + }, + { + "bbox": [ + 59, + 193, + 286, + 239 + ], + "type": "inline_equation", + "content": "900+" + }, + { + "bbox": [ + 59, + 193, + 286, + 239 + ], + "type": "text", + "content": " hours of video and annotated with 4 hierarchical levels of semantics." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 252, + 286, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 252, + 286, + 324 + ], + "spans": [ + { + "bbox": [ + 59, + 252, + 286, + 324 + ], + "type": "text", + "content": "- NewsNet provides dense annotations and multi-modal information, promoting diverse benchmarks: separate/hierarchical temporal video segmentation in scene/story/topic levels, as well as other common tasks like classification, video localization/grounding, and highlight detection." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 336, + 286, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 336, + 286, + 419 + ], + "spans": [ + { + "bbox": [ + 59, + 336, + 286, + 419 + ], + "type": "text", + "content": "- We formulate a new benchmark, i.e., hierarchical modeling in the temporal segmentation task, which needs a single model to predict segments of multiple hierarchical levels. Based on the empirical study, we bring insights into how hierarchical modeling potentially benefits the temporal video segmentation task, which was almost never discussed." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 434, + 133, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 434, + 133, + 445 + ], + "spans": [ + { + "bbox": [ + 47, + 434, + 133, + 445 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 454, + 221, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 454, + 221, + 466 + ], + "spans": [ + { + "bbox": [ + 47, + 454, + 221, + 466 + ], + "type": "text", + "content": "2.1. Related Dataset and Benchmark" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 474, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 287, + 713 + ], + "type": "text", + "content": "To better introduce our work toward hierarchical temporal segmentation, we first review research related to temporal segmentation and video challenges. As shown in Table 1, extensive benchmarks have been proposed to test the performance of temporal video segmentation. More recently, Shou et al. [51] argues that humans can capture the boundary without predefined target classes, and thus proposed Generic Event Boundary Detection (GEBD) as a new benchmark for detecting generic event boundaries on Kinetics-400 [26]. A few works [25, 33, 43-45, 56] have been proposed to attempt to address this challenge, which is a very active sub-field in long-form video understanding [1, 2]. In this paper, as the NewsNet densely annotates all temporal boundaries, including events like GEBD, scenes like MovieScene [47], and stories like TI-News [35]. NewsNet can test the performance of the existing tasks on almost all temporal localization and segmentation tasks. Moreover, with such hierarchical labels, we can further investigate hierarchical modeling in long-form video understanding, which is an important but rarely explored field." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 545, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 323 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 323 + ], + "type": "text", + "content": "As an overview of the existing benchmarks in NLP and CV, the temporal annotations are labeled in one or two levels, only focusing on the global abstraction or local understanding. However, the video is naturally hierarchical, which can be recursively represented into sub-units. Therefore, the topic-, story-, scene- and event-level tasks should not be drawn into individual lines, but work in a harmonious way. From such a motivation, we contribute a new dataset called NewsNet with hierarchical dense temporal annotations to the community. Since the annotation cost for different groups is quite different, the scale for topic-level benchmarks [4,9,23,26] is generally larger than that of the other datasets [6,7,35,47,48,75]. This paper tries our best to reach a satisfactory trade-off between scale and the quality of the annotations. The proposed dataset does not only cover 1000 videos with 946 hours but also achieves 654.4 event-level annotations per video, which is very competitive over the existing baselines. Moreover, to further generalize the promotion, textual, audio, and visual input are all collected into the NewsNet for the potential application in CV, NLP and Multi-modality." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 334, + 435, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 334, + 435, + 346 + ], + "spans": [ + { + "bbox": [ + 306, + 334, + 435, + 346 + ], + "type": "text", + "content": "2.2. Scope of Related Tasks" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 353, + 545, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 353, + 545, + 400 + ], + "spans": [ + { + "bbox": [ + 304, + 353, + 545, + 400 + ], + "type": "text", + "content": "Since dense annotations and rich modalities are provided by NewsNet, we can conduct various applications on top of the datasets. Below, we show the parts of video tasks, which can be carried on the proposed dataset." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 401, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 401, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 401, + 545, + 665 + ], + "type": "text", + "content": "Common Video Tasks. The NewsNet is a high-quality video dataset, to evaluate the methods in common video tasks. For instance, NewsNet is an ideal source for video classification [10, 15, 37, 58, 59, 62] (including the out-of-distribution detection settings [49, 61, 72, 73] due to the existence of category 'Others' in Topic level). NewsNet holds annotations in four granularities for temporal segmentation. By training with the mentioned annotation, the method can completely split a video into several short segments, which will serve partial segmentation tasks like video grounding [40, 70, 74], captioning [3, 34, 71], retrieval [5, 16, 31, 52] and highlight detection [17, 18, 20, 28, 32, 54]. Meanwhile, we simultaneously provide aligned audio, text, and visual form in NewsNet, which might promote the improvement on the task of video grounding [30, 53, 66, 69], where the goal is to retrieve the corresponding video clip by adopting an agnostic or specific query as input. We also note that, over the last few years, the 'pre-training' and 'fine-tuning' paradigms [21, 24, 39, 46, 50, 65, 67, 68, 77] have become revolutionary in both NLP and CV communities. We believe that such technology will be further improved by introducing such a high-quality multi-modality video dataset." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "Temporal Segmentation Approaches. Towards temporal video segmentation, existing works explore shot, scene and story segmentation tasks, respectively. Shot temporal segmentation [6, 11, 38, 55] is employed as a pre-processing to" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10671" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 75, + 68, + 504, + 205 + ], + "blocks": [ + { + "bbox": [ + 75, + 68, + 504, + 205 + ], + "lines": [ + { + "bbox": [ + 75, + 68, + 504, + 205 + ], + "spans": [ + { + "bbox": [ + 75, + 68, + 504, + 205 + ], + "type": "image", + "image_path": "9a691742ff130fad6c75852dd974730328a9a52fb26b33769c9723941dc6fb57.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 213, + 545, + 266 + ], + "lines": [ + { + "bbox": [ + 46, + 213, + 545, + 266 + ], + "spans": [ + { + "bbox": [ + 46, + 213, + 545, + 266 + ], + "type": "text", + "content": "Figure 2. Illustration of the four hierarchies. The illustration shows a hierarchical diagram (left) along with an annotation example (right) for the four hierarchies, including Topic, Story, Scene, and Event. In the given sample, a technology (Topic) news video can be divided into several Stories, such as \"studio\" and \"place\". Taking \"place\" as an example, it can be further characterized into different Scenes, where the backgrounds or places are similar. Based on the changes of shots or persons in a scene, the Scene can be divided into atomic units, i.e., Events." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 289, + 289, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 289, + 289, + 434 + ], + "spans": [ + { + "bbox": [ + 46, + 289, + 289, + 434 + ], + "type": "text", + "content": "locate the transition positions in videos based on the similarity of the frames. For the scene segmentation, LGSS [47] converts the problem as a binary classification task at the shot level and applies a boundary-based model that aggregates adjacent shot features within a predefined sliding window to make a prediction. While SCRL [65] introduces a boundary-free model in a seq-to-seq way and reduces the inductive bias introduced by the predefined sliding window. As for story temporal segmentation, TI-NEWS [35] proposes a multi-modal framework based on the boundary-free model to complete the story segmentation and classification tasks simultaneously." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 444, + 154, + 458 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 444, + 154, + 458 + ], + "spans": [ + { + "bbox": [ + 47, + 444, + 154, + 458 + ], + "type": "text", + "content": "3. Dataset Summary" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 464, + 287, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 464, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 46, + 464, + 287, + 536 + ], + "type": "text", + "content": "In this section, we will describe some details of the NewsNet. First, we will introduce the definition and taxonomy of the four hierarchical levels of semantics and show some annotated examples of them. Then, the statistical information of the proposed dataset will be given to better understand the data distribution of the NewsNet." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 544, + 187, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 544, + 187, + 558 + ], + "spans": [ + { + "bbox": [ + 47, + 544, + 187, + 558 + ], + "type": "text", + "content": "3.1. Definition and Taxonomy" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 563, + 287, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 563, + 287, + 647 + ], + "spans": [ + { + "bbox": [ + 46, + 563, + 287, + 647 + ], + "type": "text", + "content": "Existing works focus on the shot- or event-level temporal segmentation, while neglecting the long-horizon cases, such as story- and topic-level segmentation. Hence this paper innovatively contributes more uniform benchmarks to the community, which covers all categories of temporal annotation. In order to clarify the meaning of the different labels, we detail them one by one:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 58, + 656, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 58, + 656, + 287, + 692 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 656, + 287, + 692 + ], + "spans": [ + { + "bbox": [ + 58, + 656, + 287, + 692 + ], + "type": "text", + "content": "- Event: the atom unit in our setting, which is defined by the switching of the shots or the changing of the person." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 58, + 701, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 701, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 701, + 287, + 712 + ], + "type": "text", + "content": "- Scene: a combination of several successive shots that" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 324, + 289, + 545, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 289, + 545, + 313 + ], + "spans": [ + { + "bbox": [ + 324, + 289, + 545, + 313 + ], + "type": "text", + "content": "focus on the same place from different angles. It is defined as a taxonomy-free unit." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 317, + 320, + 545, + 448 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 317, + 320, + 545, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 320, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 317, + 320, + 545, + 380 + ], + "type": "text", + "content": "- Story: a sequence of scenes about a piece of news broadcast in a studio or outdoors. Story-level segments can be categorized into 'Connection', 'Place', 'Studio', 'Animation', 'Interview', and 'Photo'. A topic is composed of several stories." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 317, + 388, + 545, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 388, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 317, + 388, + 545, + 448 + ], + "type": "text", + "content": "- Topic: it summarizes the content of a long sequence, with the following keywords: 'Health', 'Politics', 'Entertainment', 'Economy', 'Crime', 'Weather', 'Sport', 'Technology', 'Military' and 'Others'. Generally, an individual news video contains several topics." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 304, + 455, + 545, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 455, + 545, + 563 + ], + "spans": [ + { + "bbox": [ + 304, + 455, + 545, + 563 + ], + "type": "text", + "content": "The above four different temporal annotations are defined in a hierarchical way. For example, a news video can be divided into several Topic-level segments; a Topic is further composed of several Stories; and so on. The difference between Story and Scene is that Story is category-aware while Scene is not, and the categories in Story are related to the data source. For example, Story contains a category of outdoor place footage, which is unique to news data and usually consists of multiple scenes, as illustrated in Fig. 2." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 570, + 441, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 570, + 441, + 582 + ], + "spans": [ + { + "bbox": [ + 306, + 570, + 441, + 582 + ], + "type": "text", + "content": "3.2. Collection and Statistics" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 588, + 545, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 588, + 545, + 696 + ], + "spans": [ + { + "bbox": [ + 304, + 588, + 545, + 696 + ], + "type": "text", + "content": "As a large-scale dataset, we collect about 2,000 news broadcast videos from more than 20 different TV stations, of which 1,000 were annotated, and all the data are downloaded from the YouTube and Internet. As shown in Fig. 3, we show the distribution of program duration and sources. News is a kind of typical long-form video, ranging from 10 to 110 minutes. Most of the collected videos have a 1-hour duration, performing as a challenging span for temporal segmentation. During the annotation, we found that" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 320, + 703, + 482, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 703, + 482, + 712 + ], + "spans": [ + { + "bbox": [ + 320, + 703, + 482, + 712 + ], + "type": "text", + "content": "Place refers to the news reported outside a studio." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10672" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 71, + 541, + 190 + ], + "blocks": [ + { + "bbox": [ + 52, + 71, + 541, + 190 + ], + "lines": [ + { + "bbox": [ + 52, + 71, + 541, + 190 + ], + "spans": [ + { + "bbox": [ + 52, + 71, + 541, + 190 + ], + "type": "image", + "image_path": "a3a23823e90a223ecd619c721749632845ad8bda9746e07f72c52d6adba61058.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 198, + 546, + 232 + ], + "lines": [ + { + "bbox": [ + 46, + 198, + 546, + 232 + ], + "spans": [ + { + "bbox": [ + 46, + 198, + 546, + 232 + ], + "type": "text", + "content": "Figure 3. The statistical information of NewsNet. The left histogram shows the distribution of the Top-20 news programs, (i.e., program A to T, by the total number of videos) in different duration, it can be found that the duration of the same program is relatively close. Besides, the two ring charts on the right show the category proportions of Topic and Story." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 252, + 288, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 252, + 288, + 336 + ], + "spans": [ + { + "bbox": [ + 46, + 252, + 288, + 336 + ], + "type": "text", + "content": "there exists a long-tailed distribution for the categories of the story- and topic-level annotation, leading to a practical challenge. We believe such a benchmark can simulate the real-world scenario effectively, so that the competing methods could be benchmarked in a practical manner. More details such as annotation process, visualization, etc., can be found in the Supplementary Material." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 346, + 249, + 360 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 346, + 249, + 360 + ], + "spans": [ + { + "bbox": [ + 47, + 346, + 249, + 360 + ], + "type": "text", + "content": "4. Benchmark and Protocol Navigation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 365, + 287, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 365, + 287, + 462 + ], + "spans": [ + { + "bbox": [ + 46, + 365, + 287, + 462 + ], + "type": "text", + "content": "In this section, we will first introduce two temporal segmentation task settings, i.e., Separate Modeling and Hierarchical Modeling in NewsNet, with their definitions and protocols. Then, based on the potential relationship between different hierarchies, a simple yet effective loss function is introduced to better tackle hierarchical temporal segmentation. At last, we will briefly introduce the other common video tasks that can be conducted with NewsNet." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 468, + 244, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 468, + 244, + 480 + ], + "spans": [ + { + "bbox": [ + 47, + 468, + 244, + 480 + ], + "type": "text", + "content": "4.1. Hierarchical Temporal Segmentation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 486, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 287, + 605 + ], + "type": "text", + "content": "As long-form video can be divided into sub-video chunks according to its intrinsic properties, many researchers try to tackle the video segmentation problem from different perspectives in the last few years [12, 35, 47, 65]. Without loss of generality, we classify hierarchical temporal segmentation into two categories: Separate Modeling and Hierarchical Modeling. Separate modeling trains a segmentation model for each level segmentation task, while hierarchical modeling trains a model which simultaneously segments a video into multi-level segments." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "text", + "content": "NewsNet supports temporal segmentation tasks with various semantic granularity, i.e., Event, Scene, Story, and Topic. Since many methods [12,47,65], take the Event (aka Shot sometimes) as the basic input unit to reduce the redundancy of the video and computational cost, NewsNet draws the video segmentation tasks on Scene, Story, and Topic. Moreover, these three hierarchies can be strictly aligned according to the Event (Shot) dimension, which is more conducive to Hierarchical Modeling and more convenient for" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 253, + 545, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 253, + 545, + 276 + ], + "spans": [ + { + "bbox": [ + 304, + 253, + 545, + 276 + ], + "type": "text", + "content": "comparison. Note that we also carried out the event-level experiment (see Supplementary Material)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 277, + 546, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 277, + 546, + 410 + ], + "spans": [ + { + "bbox": [ + 304, + 277, + 546, + 410 + ], + "type": "text", + "content": "Data and Modality. In-domain: all the videos are randomly split into training, validation, and testing sets with a ratio of 3:1:1. Cross-domain: training, validation, and testing sets are split according to different news program types/IDs; 5 different split combinations are set to ensure the generalization ability of the algorithm can be evaluated properly and fairly. As for the embeddings used in the experiment, the shot-level visual, audio, and textual features are extracted through ResNet50 [22] (pre-trained on Places365 [76]), public Bert-base model [14], and Cnn14 [27] (pre-trained on AudioSet [19]), respectively." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 422, + 421, + 435 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 422, + 421, + 435 + ], + "spans": [ + { + "bbox": [ + 305, + 422, + 421, + 435 + ], + "type": "text", + "content": "4.1.1 Separate Modeling" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 441, + 545, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 441, + 545, + 513 + ], + "spans": [ + { + "bbox": [ + 304, + 441, + 545, + 513 + ], + "type": "text", + "content": "Task Definition. Given a video with the shot labels, the model requires to classify whether a shot is a segmentation point for one of the three semantic levels (Scene, Story and Topic). A segmentation point refers to a shot that lies at the end of a topic/story/scene segment, and its next shot is the start of a new topic/story/scene." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 514, + 545, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 514, + 545, + 597 + ], + "spans": [ + { + "bbox": [ + 304, + 514, + 545, + 597 + ], + "type": "text", + "content": "Backbone. In general, the common paradigm in those segmentation tasks is considering temporal segment to shot-level classification in boundary-base and -free way. We employ the state-of-the-art boundary-free model, SCRL [65], and the boundary-based model, LGSS [47] as our baseline approaches in separate modeling. More details about the protocol can be found in the Supplementary Material." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 611, + 436, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 611, + 436, + 623 + ], + "spans": [ + { + "bbox": [ + 305, + 611, + 436, + 623 + ], + "type": "text", + "content": "4.1.2 Hierarchical Modeling" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 629, + 545, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 666 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 666 + ], + "type": "text", + "content": "Task Definition. Given a video with the shot labels, the model requires to classify whether a shot is a segmentation point for three semantic levels (Scene, Story, and Topic)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "text", + "content": "Backbone. In this setting, as shown in Fig. 4, we focus on two hierarchical modeling paradigms, including Multi-Label and Multi-Head with a novel learning objective called Hierarchical Ranking. More specifically, Multi" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10673" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 74, + 70, + 521, + 246 + ], + "blocks": [ + { + "bbox": [ + 74, + 70, + 521, + 246 + ], + "lines": [ + { + "bbox": [ + 74, + 70, + 521, + 246 + ], + "spans": [ + { + "bbox": [ + 74, + 70, + 521, + 246 + ], + "type": "image", + "image_path": "c5c92b2b846c82cf48f62c9b53e5d63526865ee810ae4ddf00fbe99f7eec5954.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 255, + 547, + 289 + ], + "lines": [ + { + "bbox": [ + 46, + 255, + 547, + 289 + ], + "spans": [ + { + "bbox": [ + 46, + 255, + 547, + 289 + ], + "type": "text", + "content": "Figure 4. An illustration of the separate modeling, two hierarchical modeling paradigms, and the proposed loss function. For the baseline, the model is trained individually corresponding to different segmentation tasks. For hierarchical modeling, all the tasks are aggregated into one model. Specifically, B, L, and N represent the batch size, sequence length, and feature dimension, respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 309, + 289, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 309, + 289, + 536 + ], + "spans": [ + { + "bbox": [ + 46, + 309, + 289, + 536 + ], + "type": "text", + "content": "Label works towards the hierarchical labels from the perspective of multi-label optimization, focusing on the aggregation of information into one head for estimation. While, Multi-Head addresses the challenge as a multi-task problem, where the tasks in different levels should be predicted via the individual pathways. Both pipelines are widely employed in modern architectures [13, 36, 60, 64], performing as simple but representative paradigms. Therefore, we adopt two hierarchical modeling paradigms as the baselines. Hierarchical Ranking Loss. To better model the hierarchical temporal video segmentation task and combine the existing methods, we propose a new loss function called Hierarchical Ranking Loss. The motivation is: due to the accurate alignment between the hierarchical levels with the atom unit Event, as shown in Fig. 1, a positive segment boundary of the higher level task must be a positive segment boundary of the lower task. Hence, the segmentation confidence scores for different tasks should be ranked in a non-increasing order from coarse to fine for the paired task." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 536, + 288, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 536, + 288, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 536, + 288, + 579 + ], + "type": "text", + "content": "Given two tasks in the hierarchical modeling setting, e.g. task " + }, + { + "bbox": [ + 46, + 536, + 288, + 579 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 46, + 536, + 288, + 579 + ], + "type": "text", + "content": " and task " + }, + { + "bbox": [ + 46, + 536, + 288, + 579 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 46, + 536, + 288, + 579 + ], + "type": "text", + "content": ", where the semantic level of task " + }, + { + "bbox": [ + 46, + 536, + 288, + 579 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 46, + 536, + 288, + 579 + ], + "type": "text", + "content": " is higher than that of task " + }, + { + "bbox": [ + 46, + 536, + 288, + 579 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 46, + 536, + 288, + 579 + ], + "type": "text", + "content": ", the proposed hierarchical ranking loss function is as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 83, + 581, + 287, + 602 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 581, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 83, + 581, + 287, + 602 + ], + "type": "interline_equation", + "content": "\\mathcal {P} _ {\\mathcal {H}, \\mathcal {L}} = \\frac {1}{B} \\frac {1}{L} \\cdot \\sum \\left[ \\sigma \\left(F _ {\\mathcal {H}} - \\mathcal {S G} \\left(F _ {\\mathcal {L}}\\right)\\right) \\cdot Y _ {\\mathcal {L}} \\right] \\tag {1}", + "image_path": "a71b0260f2d751698da3c636c0f59bcf2244f958749c0741c80c64cf1141850f.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 604, + 287, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 604, + 287, + 674 + ], + "spans": [ + { + "bbox": [ + 46, + 604, + 287, + 674 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 604, + 287, + 674 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 46, + 604, + 287, + 674 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 604, + 287, + 674 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 46, + 604, + 287, + 674 + ], + "type": "text", + "content": " are the batch size and sequence length respectively; " + }, + { + "bbox": [ + 46, + 604, + 287, + 674 + ], + "type": "inline_equation", + "content": "\\sigma(\\cdot)" + }, + { + "bbox": [ + 46, + 604, + 287, + 674 + ], + "type": "text", + "content": " refers to the contraction mapping, like sigmoid function; " + }, + { + "bbox": [ + 46, + 604, + 287, + 674 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 46, + 604, + 287, + 674 + ], + "type": "text", + "content": " stands for the segmentation confidence score of the task; " + }, + { + "bbox": [ + 46, + 604, + 287, + 674 + ], + "type": "inline_equation", + "content": "\\mathcal{SG}" + }, + { + "bbox": [ + 46, + 604, + 287, + 674 + ], + "type": "text", + "content": " is the Stop Gradient operation for the computational graph, and " + }, + { + "bbox": [ + 46, + 604, + 287, + 674 + ], + "type": "inline_equation", + "content": "Y_{\\mathcal{L}}" + }, + { + "bbox": [ + 46, + 604, + 287, + 674 + ], + "type": "text", + "content": " corresponds to the label of the task " + }, + { + "bbox": [ + 46, + 604, + 287, + 674 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 46, + 604, + 287, + 674 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 675, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 675, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 47, + 675, + 287, + 696 + ], + "type": "text", + "content": "And the final objective function for all three tasks can be formed as follow:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 702, + 287, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 702, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 53, + 702, + 287, + 714 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {P} _ {\\text {T o p i c}, S t o r y} + \\mathcal {P} _ {\\text {T o p i c}, S c e n e} + \\mathcal {P} _ {\\text {S t o r y}, S c e n e} + \\mathcal {L} _ {C E} \\tag {2}", + "image_path": "42b3ff38d312adca47de348422fac2e4aca71e54e02f46a5de5b54da155cb98c.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 308, + 309, + 495, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 309, + 495, + 322 + ], + "spans": [ + { + "bbox": [ + 308, + 309, + 495, + 322 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 308, + 309, + 495, + 322 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{CE}" + }, + { + "bbox": [ + 308, + 309, + 495, + 322 + ], + "type": "text", + "content": " is the cross-entropy loss function." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 327, + 504, + 341 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 327, + 504, + 341 + ], + "spans": [ + { + "bbox": [ + 306, + 327, + 504, + 341 + ], + "type": "text", + "content": "4.2. Common Video Understanding Tasks" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 346, + 545, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 346, + 545, + 394 + ], + "spans": [ + { + "bbox": [ + 304, + 346, + 545, + 394 + ], + "type": "text", + "content": "The NewsNet could also apply other common video understanding tasks, e.g., video classification, localization, and highlight detection. More details can be found in Supplementary Material." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 403, + 451, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 403, + 451, + 418 + ], + "spans": [ + { + "bbox": [ + 306, + 403, + 451, + 418 + ], + "type": "text", + "content": "5. Experiment and Analysis" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 424, + 545, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 424, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 424, + 545, + 533 + ], + "type": "text", + "content": "In this section, we mainly focus on the temporal segmentation tasks on the NewsNet and further contribute several benchmarks on the different baselines. In particular, we conduct the empirical study to thoroughly benchmark the effectiveness of model structure, multi-modality information, domain gap, and hierarchical modeling. Note that, due to the limitation on page length, the implementation details and the results for other common video tasks are given in the Supplementary Material." + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 308, + 587, + 545, + 663 + ], + "blocks": [ + { + "bbox": [ + 304, + 542, + 547, + 578 + ], + "lines": [ + { + "bbox": [ + 304, + 542, + 547, + 578 + ], + "spans": [ + { + "bbox": [ + 304, + 542, + 547, + 578 + ], + "type": "text", + "content": "Table 2. The performance between boundary-based (B.B.) and boundary-free (B.F) model with visual input on F1 score " + }, + { + "bbox": [ + 304, + 542, + 547, + 578 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 304, + 542, + 547, + 578 + ], + "type": "text", + "content": ", Precision " + }, + { + "bbox": [ + 304, + 542, + 547, + 578 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 304, + 542, + 547, + 578 + ], + "type": "text", + "content": ", and Recall " + }, + { + "bbox": [ + 304, + 542, + 547, + 578 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 304, + 542, + 547, + 578 + ], + "type": "text", + "content": ". Light gray refers to the failure case." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 587, + 545, + 663 + ], + "lines": [ + { + "bbox": [ + 308, + 587, + 545, + 663 + ], + "spans": [ + { + "bbox": [ + 308, + 587, + 545, + 663 + ], + "type": "table", + "html": "
TaskModelF1 scorePrecisionRecall
SceneB.B.77.367.790.2
B.F.76.876.177.5
StoryB.B.70.765.976.2
B.F.71.272.370.0
TopicB.B.13.17.0100.0
B.F.62.972.455.6
", + "image_path": "42e4641d12da66610b52c4013beefb8b393b408f2b358637ee8055da19b9383f.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "content": "As an overview, based on the large-scale dataset with rich modalities, we conduct several experiments to respond to five critical questions concerned by the community:" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "10674" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 52, + 114, + 294, + 258 + ], + "blocks": [ + { + "bbox": [ + 48, + 71, + 533, + 105 + ], + "lines": [ + { + "bbox": [ + 48, + 71, + 533, + 105 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 533, + 105 + ], + "type": "text", + "content": "Table 3. In-domain performance by using boundary-free (B.F.) Table 4. Cross-domain setting by using boundary-free (B.F.) model. The bolded values stand for the optimal performances for model. The bolded values stand for the optimal performances for each task." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 52, + 114, + 294, + 258 + ], + "lines": [ + { + "bbox": [ + 52, + 114, + 294, + 258 + ], + "spans": [ + { + "bbox": [ + 52, + 114, + 294, + 258 + ], + "type": "table", + "html": "
TaskModalityF1 scorePrecisionRecall
SceneV76.876.177.5
A69.866.873.2
T66.756.381.9
V+A+T78.380.975.8
StoryV71.272.370.0
A59.357.661.1
T50.657.445.2
V+A+T75.474.776.2
TopicV62.972.455.6
A58.159.456.9
T39.046.533.5
V+A+T73.274.372.2
", + "image_path": "0158a464405f42f31307fe8efc1d25aa98b00e59b26f329154dae3d4578a75e7.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 270, + 286, + 427 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 55, + 270, + 286, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 270, + 286, + 305 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 286, + 305 + ], + "type": "text", + "content": "1. Can existing methods also perform well on long-horizon temporal segmentation tasks, such as topic-level and story-level segmentation?" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 312, + 286, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 312, + 286, + 336 + ], + "spans": [ + { + "bbox": [ + 55, + 312, + 286, + 336 + ], + "type": "text", + "content": "2. Do the different-level tasks test the ability of the model on viewing different temporal spans?" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 342, + 286, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 342, + 286, + 366 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 286, + 366 + ], + "type": "text", + "content": "3. Can existing methods well address the semantic gap between low- and high-level granularities?" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 373, + 286, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 373, + 286, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 373, + 286, + 396 + ], + "type": "text", + "content": "4. Will the temporal segmentation models be benefited from multi-modality information?" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 403, + 286, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 403, + 286, + 427 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 286, + 427 + ], + "type": "text", + "content": "5. How could the models benefit from Hierarchical Modeling?" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 431, + 263, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 431, + 263, + 445 + ], + "spans": [ + { + "bbox": [ + 47, + 431, + 263, + 445 + ], + "type": "text", + "content": "5.1. Results against Increased Temporal Span" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 449, + 286, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 449, + 286, + 664 + ], + "spans": [ + { + "bbox": [ + 47, + 449, + 286, + 664 + ], + "type": "text", + "content": "To investigate the generalization of the existing methods against the long-horizon cases, we conduct an experiment on two different head architectures, including boundary-free and boundary-based heads, on three tasks: scene-level, story-level, and topic-level temporal segmentation. As shown in Fig. 2, the performance of the boundary-based model has a significant drop on the Topic task. Compared with the other tasks, the Topic segmentation task has more time slices for input, leading to a hard challenge caused by longer time dependencies. The methods can segment scenes with about 76-77 F1 score while degrading to 62.9 on Topic-level segmentation. For the first question, the answer is that both the boundary-free model and the boundary-based model can not generalize well in the story-and topic-level segmentation. Combining the overall performance and stability on different tasks, the next experiments will utilize a boundary-free model as the basic backbone under comprehensive consideration." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 671, + 267, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 267, + 685 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 267, + 685 + ], + "type": "text", + "content": "5.2. Performance against Cross News Domains" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "type": "text", + "content": "To investigate the generalization of the existing method, we analyze the performance of the method against the va" + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 299, + 114, + 527, + 258 + ], + "blocks": [ + { + "bbox": [ + 299, + 114, + 527, + 258 + ], + "lines": [ + { + "bbox": [ + 299, + 114, + 527, + 258 + ], + "spans": [ + { + "bbox": [ + 299, + 114, + 527, + 258 + ], + "type": "table", + "html": "
TaskModalityAvg. F1 score (std.)Avg. PrecisionAvg. Recall
SceneV72.9 (2.1)70.975.2
A62.7 (4.0)59.766.6
T61.6 (5.0)52.877.0
V+A+T76.0 (2.1)74.477.9
StoryV68.5 (2.6)70.366.9
A55.7 (3.6)53.659.0
T51.1 (3.6)43.465.4
V+A+T72.9 (2.2)73.772.4
TopicV60.6 (4.7)69.853.8
A59.0 (5.2)56.062.9
T49.8 (5.2)45.755.9
V+A+T72.2 (3.6)72.372.5
", + "image_path": "0f610fa3f1bec2006e0a9c17d635ddd7366326e07e544dd1280f00812dda6bba.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 270, + 545, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 270, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 304, + 270, + 545, + 472 + ], + "type": "text", + "content": "rieties of different news programs. As shown in Table 3 and 4, the cross-domain setting slightly degrades the performance on all the tasks. Comparing the in-domain results using the full modality in Table 3 and corresponding cross-domain results in Table 4, the F1 score decreases by 2.3 on task Scene, by 2.5 on task Story, and by 1.0 on task Topic. The reason for this result may be that higher-level tasks, such as Topic, has more abstract semantics and the knowledge is easier to transfer. In contrast, task Scene in the lowest level, generally only considers place switching as the segmentation cue. Besides, the category of Scene is an open set, leading to a greater challenge in the cross-domain setting. Hence, temporal segmentation model also suffers from the out-of-distribution case. Based on the property of the collected dataset, we set the videos from different news programs as different domains, and establish a comprehensive benchmark to facilitate the research on such line." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 479, + 499, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 479, + 499, + 491 + ], + "spans": [ + { + "bbox": [ + 306, + 479, + 499, + 491 + ], + "type": "text", + "content": "5.3. Performance using Multi-Modalities" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 498, + 545, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 641 + ], + "type": "text", + "content": "When it comes to multi-modality, as shown in Table. 4, the model can be benefited from multimodal information among all the tasks. For example, the performance of using only visual modality achieves only 60.6 recall in topic-level segmentation. However, by aggregating the other modalities including audio and text, the performance can be improved significantly (72.2 vs 49.8), which indicates the potential value for addressing long-discrepancy via multimodality technology. Note that, the improvement in Scene segmentation is, to some extent, limited, we believe that Scene is mainly defined by visual input, hence can not obtain adequate complementarity from other modalities." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 647, + 541, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 647, + 541, + 661 + ], + "spans": [ + { + "bbox": [ + 306, + 647, + 541, + 661 + ], + "type": "text", + "content": "5.4. Performance based on Hierarchical Modeling" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "text", + "content": "Here, we discuss the core challenge revealed in this paper: 'How could the model benefit from hierarchical annotations?' In this case, the baseline is set as the model, which can only be accessible to one specific-level annotation. As" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10675" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 146, + 287, + 293 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 288, + 138 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 288, + 138 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 288, + 138 + ], + "type": "text", + "content": "Table 5. The F1 scores of baselines trained with different levels of annotations on full modalities without our hierarchical ranking loss, where blue and orange indicate the in-domain and cross-domain, respectively. Each row refers to the result corresponding to a single task. Hie. Modeling stands for Hierarchical Modeling while Sep. Modeling is Separate Modeling." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 146, + 287, + 293 + ], + "lines": [ + { + "bbox": [ + 50, + 146, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 50, + 146, + 287, + 293 + ], + "type": "table", + "html": "
RecipeBaseline\nSep. ModelingMulti-Label\nHie. ModelingMulti-Head\nHie. Modeling
Scene78.3 / 76.079.1 / 76.579.9 / 76.9
+ Story75.4 / 72.975.4 / 74.774.2 / 74.0
Scene78.3 / 76.079.8 / 76.479.5 / 76.5
+ Topic73.2 / 72.270.5 / 72.870.9 / 73.0
Story75.4 / 72.976.2 / 74.375.4 / 73.9
+ Topic73.2 / 72.277.3 / 73.275.2 / 73.5
Scene78.3 / 76.077.4 / 76.879.8 / 76.8
+ Story75.4 / 72.974.3 / 74.374.5 / 73.7
+ Topic73.2 / 72.274.3 / 72.676.6 / 70.4
", + "image_path": "783532a966441f3a844ebf146adef59bcd4ad5d64b388806442882a4a9034eae.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 319, + 287, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 319, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 46, + 319, + 287, + 498 + ], + "type": "text", + "content": "listed in Table 5, we consider two simple strategies, including Multi-Label and Multi-Head, to aggregate the hierarchical information. By training with multi-level annotations, a " + }, + { + "bbox": [ + 46, + 319, + 287, + 498 + ], + "type": "inline_equation", + "content": "1.0 \\sim 3.0\\%" + }, + { + "bbox": [ + 46, + 319, + 287, + 498 + ], + "type": "text", + "content": " improvement can be achieved, which indicates, the segmentation can be benefited by introducing the hierarchical structures of videos. However, Table 5 also shows directly combining different-level granularities might not result in gains due to the semantic gap between low- and high-level granularities. For example, directly combining scene- and topic-level information by Multi-Label hierarchical modeling or Multi-Head hierarchical modeling does not consistently improve the topic-level segmentation performance. This poses a new and challenging research question, i.e. how to jointly train with various granularities toward hierarchical modeling of long-form videos?" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 505, + 287, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 505, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 46, + 505, + 287, + 601 + ], + "type": "text", + "content": "To address the above question, we introduce a simple but effective loss function, i.e., Hierarchical Ranking Loss introduced in Sec. 4.1.2. As shown in Table 6, by combining with our proposed loss, the F1 score for segmentation can be further improved in most cases. Typically, in Topic-level, Multi-Head coupled with the proposed loss can reach the best performance, i.e., improving Multi-Head case from 70.4 to 73.2 F1 score." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 628, + 275, + 641 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 628, + 275, + 641 + ], + "spans": [ + { + "bbox": [ + 47, + 628, + 275, + 641 + ], + "type": "text", + "content": "5.5. Other Benchmarks involved in the NewsNet" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 654, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 287, + 715 + ], + "type": "text", + "content": "In addition to the temporal segmentation tasks, we also conduct extensive common video tasks, including highlight detection, video classification, and video localization, on the proposed dataset. Due to the limited pages, we show this part in the Supplementary Materials." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 309, + 124, + 545, + 217 + ], + "blocks": [ + { + "bbox": [ + 305, + 70, + 545, + 115 + ], + "lines": [ + { + "bbox": [ + 305, + 70, + 545, + 115 + ], + "spans": [ + { + "bbox": [ + 305, + 70, + 545, + 115 + ], + "type": "text", + "content": "Table 6. The F1 scores of the methods with or without hierarchical ranking loss under the in-domain / cross-domain setting on full modalities. Hie. stands for Hierarchical Modeling and Sep. refers to Separate Modeling." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 124, + 545, + 217 + ], + "lines": [ + { + "bbox": [ + 309, + 124, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 309, + 124, + 545, + 217 + ], + "type": "table", + "html": "
MethodSceneStoryTopic
Baseline (Sep.)78.3 / 76.075.4 / 72.973.2 / 72.2
Multi-Label (Hie.)77.4 / 76.874.3 / 74.374.3 / 72.6
Multi-Label w/ Hie. Loss (Hie.)79.6 / 76.974.4 / 73.577.8 / 73.1
Multi-Head (Hie.)79.8 / 76.874.5 / 73.776.6 / 70.4
Multi-Head w/ Hie. Loss (Hie.)80.3 / 76.976.3 / 74.676.5 / 73.2
", + "image_path": "14668eeda4db6d409f3ec78f89a265cc0fdc1bf9de6ba134e55840c9a619e32a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 235, + 529, + 248 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 235, + 529, + 248 + ], + "spans": [ + { + "bbox": [ + 306, + 235, + 529, + 248 + ], + "type": "text", + "content": "6. Conclusion, Limitation and Future work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 258, + 545, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 258, + 545, + 474 + ], + "spans": [ + { + "bbox": [ + 304, + 258, + 545, + 474 + ], + "type": "text", + "content": "We introduce NewsNet, a new large-scale dataset for temporal video segmentation. Compared with existing large-scale datasets for temporal video segmentation, NewsNet additionally provides two abstractive levels of temporal segmentation, which have not been taken into account by these datasets. The four-level and hierarchical annotations enable the community to explore how to comprehensively represent the complex structure of long-form videos from coarse to fine. Extensive efforts of human workers are devoted, so as to ensure that annotations are high-quality. We hope that NewsNet and baseline results can facilitate the development of temporal video segmentation in terms of insightful research and practical tools. In addition, as NewsNet provides diverse annotations for dividing a video into segments of different granularities as well as category labels, it can also serve other video understanding tasks such as video grounding, captioning, retrieval, highlight detection, and video classification." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 476, + 545, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 476, + 545, + 561 + ], + "spans": [ + { + "bbox": [ + 304, + 476, + 545, + 561 + ], + "type": "text", + "content": "In this paper, we only focus on the problem of hierarchical temporal segmentation. We have not explored other multi-modality tasks, such as visual reasoning and video question answering, while our dataset can serve these tasks thanks to its multi-modality nature. We will explore these tasks in our future work to further promote advances in multi-modality-based video understanding and reasoning." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 580, + 404, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 580, + 404, + 594 + ], + "spans": [ + { + "bbox": [ + 306, + 580, + 404, + 594 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 602, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 602, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 602, + 545, + 713 + ], + "type": "text", + "content": "This work was supported by the KAUST Office of Sponsored Research through the Visual Computing Center (VCC) funding, as well as, the SDAIA-KAUST Center of Excellence in Data Science and Artificial Intelligence (SDAIA-KAUST AI). Raghavendra Ramachandra is supported under the SALT and OFFPAD projects funded by the Research Council of Norway, Norway. Chia-Wen Lin is supported by National Science and Technology Council, Taiwan, under project NSTC 111-2634-F002-023." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "10676" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 286, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 286, + 112 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 286, + 112 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 286, + 112 + ], + "type": "text", + "content": "[1] LOVEU21. https://sites.google.com/view/loveucvpr21, 2021.3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 114, + 286, + 144 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 114, + 286, + 144 + ], + "spans": [ + { + "bbox": [ + 53, + 114, + 286, + 144 + ], + "type": "text", + "content": "[2] LOVEU22 program schedule. https://sites.google.com/view/loveucvpr22/program, 2022.3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 148, + 286, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 286, + 212 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 286, + 212 + ], + "type": "text", + "content": "[3] Nayyer Aafaq, Naveed Akhtar, Wei Liu, Syed Zulqarnain Gilani, and Ajmal Mian. Spatio-temporal dynamics and semantic attribute enriched visual encoding for video captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12487-12496, 2019. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 214, + 286, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 214, + 286, + 267 + ], + "spans": [ + { + "bbox": [ + 53, + 214, + 286, + 267 + ], + "type": "text", + "content": "[4] Sami Abu-El-Haija, Nisarg Kothari, Joonseok Lee, Paul Natsev, George Toderici, Balakrishnan Varadarajan, and Sudheendra Vijayanarasimhan. Youtube-8m: A large-scale video classification benchmark. arXiv preprint arXiv:1609.08675, 2016. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 270, + 286, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 270, + 286, + 323 + ], + "spans": [ + { + "bbox": [ + 53, + 270, + 286, + 323 + ], + "type": "text", + "content": "[5] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1728-1738, 2021. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 325, + 286, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 325, + 286, + 369 + ], + "spans": [ + { + "bbox": [ + 53, + 325, + 286, + 369 + ], + "type": "text", + "content": "[6] Lorenzo Baraldi, Costantino Grana, and Rita Cucchiara. A deep siamese network for scene detection in broadcast videos. In Proceedings of the 23rd ACM international conference on Multimedia, pages 1199-1202, 2015. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 370, + 286, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 370, + 286, + 423 + ], + "spans": [ + { + "bbox": [ + 53, + 370, + 286, + 423 + ], + "type": "text", + "content": "[7] Lorenzo Baraldi, Costantino Grana, and Rita Cucchiara. Shot and scene detection via hierarchical clustering for re-using broadcast video. In International conference on computer analysis of images and patterns, pages 801-811. Springer, 2015. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 425, + 286, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 425, + 286, + 479 + ], + "spans": [ + { + "bbox": [ + 53, + 425, + 286, + 479 + ], + "type": "text", + "content": "[8] Joe Barrow, Rajiv Jain, Vlad Morariu, Varun Manjunatha, Douglas W Oard, and Philip Resnik. A joint model for document segmentation and segment labeling. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 313-322, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 480, + 286, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 480, + 286, + 534 + ], + "spans": [ + { + "bbox": [ + 53, + 480, + 286, + 534 + ], + "type": "text", + "content": "[9] Fabian Caba Heilbron, Victor Escorcia, Bernard Ghanem, and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 961-970, 2015. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 536, + 286, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 286, + 579 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 286, + 579 + ], + "type": "text", + "content": "[10] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 1, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 581, + 286, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 581, + 286, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 581, + 286, + 624 + ], + "type": "text", + "content": "[11] Vasileios T Chasanis, Aristidis C Likas, and Nikolaos P Galatsanos. Scene detection in videos using shot clustering and sequence alignment. IEEE transactions on multimedia, 11(1):89-100, 2008. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 625, + 286, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 286, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 286, + 679 + ], + "type": "text", + "content": "[12] Shixing Chen, Xiaohan Nie, David Fan, Dongqing Zhang, Vimal Bhat, and Raffay Hamid. Shot contrastive self-supervised learning for scene boundary detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9796-9805, 2021. 2, 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 681, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 286, + 713 + ], + "type": "text", + "content": "[13] Zhao-Min Chen, Xiu-Shen Wei, Peng Wang, and Yanwen Guo. Multi-label image recognition with graph convolutional networks. In Proceedings of the IEEE/CVF Confer-" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 74, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 326, + 74, + 545, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 74, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 326, + 74, + 545, + 94 + ], + "type": "text", + "content": "ence on Computer Vision and Pattern Recognition (CVPR), June 2019. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 96, + 545, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 138 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 138 + ], + "type": "text", + "content": "[14] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 140, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 140, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 308, + 140, + 545, + 183 + ], + "type": "text", + "content": "[15] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6202-6211, 2019. 1, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 185, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 185, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 308, + 185, + 545, + 228 + ], + "type": "text", + "content": "[16] Valentin Gabeur, Chen Sun, Karteek Alahari, and Cordelia Schmid. Multi-modal transformer for video retrieval. In European Conference on Computer Vision, pages 214-229. Springer, 2020. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 228, + 545, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 228, + 545, + 271 + ], + "spans": [ + { + "bbox": [ + 308, + 228, + 545, + 271 + ], + "type": "text", + "content": "[17] Jiyang Gao, Chen Sun, Zhenheng Yang, and Ram Nevatia. Tall: Temporal activity localization via language query. In Proceedings of the IEEE international conference on computer vision, pages 5267-5275, 2017. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 273, + 545, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 273, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 308, + 273, + 545, + 316 + ], + "type": "text", + "content": "[18] Ana Garcia del Molino and Michael Gygli. PhD-gifs: personalized highlight detection for automatic gif creation. In Proceedings of the 26th ACM international conference on Multimedia, pages 600–608, 2018. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 317, + 545, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 317, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 308, + 317, + 545, + 381 + ], + "type": "text", + "content": "[19] Jort F Gemmeke, Daniel PW Ellis, Dylan Freedman, Aren Jansen, Wade Lawrence, R Channing Moore, Manoj Plakal, and Marvin Ritter. Audio set: An ontology and human-labeled dataset for audio events. In 2017 IEEE international conference on acoustics, speech and signal processing (ICASSP), pages 776-780. IEEE, 2017. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 383, + 545, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 383, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 308, + 383, + 545, + 426 + ], + "type": "text", + "content": "[20] Michael Gygli, Yale Song, and Liangliang Cao. Video2gif: Automatic generation of animated gifs from video. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1001-1009, 2016. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 427, + 545, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 427, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 308, + 427, + 545, + 480 + ], + "type": "text", + "content": "[21] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 482, + 545, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 482, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 308, + 482, + 545, + 525 + ], + "type": "text", + "content": "[22] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 526, + 545, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 526, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 308, + 526, + 545, + 569 + ], + "type": "text", + "content": "[23] Qingqiu Huang, Yu Xiong, Anyi Rao, Jiaze Wang, and Dahua Lin. Movienet: A holistic dataset for movie understanding. In European Conference on Computer Vision, pages 709-727. Springer, 2020. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 571, + 545, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 571, + 545, + 603 + ], + "spans": [ + { + "bbox": [ + 308, + 571, + 545, + 603 + ], + "type": "text", + "content": "[24] Chen Ju, Tengda Han, Kunhao Zheng, Ya Zhang, and Weidi Xie. Prompting visual-language models for efficient video understanding. arXiv preprint arXiv:2112.04478, 2021. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 604, + 545, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 604, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 308, + 604, + 545, + 658 + ], + "type": "text", + "content": "[25] Hyolim Kang, Jinwoo Kim, Taehyun Kim, and Seon Joo Kim. Uboco: Unsupervised boundary contrastive learning for generic event boundary detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20073-20082, 2022. 2, 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 659, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 659, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 659, + 545, + 712 + ], + "type": "text", + "content": "[26] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 2, 3" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "10677" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[27] Qiuqiang Kong, Yin Cao, Turab Iqbal, Yuxuan Wang, Wenwu Wang, and Mark D Plumbley. Panns: Large-scale pretrained audio neural networks for audio pattern recognition. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 28:2880-2894, 2020. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 287, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 174 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 174 + ], + "type": "text", + "content": "[28] Ranjay Krishna, Kenji Hata, Frederic Ren, Li Fei-Fei, and Juan Carlos Niebles. Dense-captioning events in videos. In Proceedings of the IEEE international conference on computer vision, pages 706–715, 2017. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 175, + 287, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 287, + 228 + ], + "type": "text", + "content": "[29] Kathy Lee, Diana Palsetia, Ramanathan Narayanan, Md Mostofa Ali Patwary, Ankit Agrawal, and Alok Choudhary. Twitter trending topic classification. In 2011 IEEE 11th International Conference on Data Mining Workshops, pages 251-258. IEEE, 2011. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 230, + 287, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 230, + 287, + 274 + ], + "spans": [ + { + "bbox": [ + 48, + 230, + 287, + 274 + ], + "type": "text", + "content": "[30] Jie Lei, Tamara L Berg, and Mohit Bansal. Detecting moments and highlights in videos via natural language queries. Advances in Neural Information Processing Systems, 34:11846-11858, 2021. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 276, + 287, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 276, + 287, + 331 + ], + "spans": [ + { + "bbox": [ + 48, + 276, + 287, + 331 + ], + "type": "text", + "content": "[31] Jie Lei, Linjie Li, Luowei Zhou, Zhe Gan, Tamara L Berg, Mohit Bansal, and Jingjing Liu. Less is more: Clipbert for video-and-language learning via sparse sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7331-7341, 2021. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 332, + 287, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 332, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 48, + 332, + 287, + 376 + ], + "type": "text", + "content": "[32] Jie Lei, Licheng Yu, Tamara L Berg, and Mohit Bansal. Tvr: A large-scale dataset for video-subtitle moment retrieval. In European Conference on Computer Vision, pages 447-463. Springer, 2020. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 377, + 287, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 377, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 48, + 377, + 287, + 441 + ], + "type": "text", + "content": "[33] Congcong Li, Xinyao Wang, Longyin Wen, Dxiang Hong, Tiejian Luo, and Libo Zhang. End-to-end compressed video representation learning for generic event boundary detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13967-13976, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 445, + 287, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 445, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 48, + 445, + 287, + 510 + ], + "type": "text", + "content": "[34] Kevin Lin, Linjie Li, Chung-Ching Lin, Faisal Ahmed, Zhe Gan, Zicheng Liu, Yumao Lu, and Lijuan Wang. Swinbert: End-to-end transformers with sparse attention for video captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17949-17958, 2022. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 512, + 287, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 512, + 287, + 576 + ], + "spans": [ + { + "bbox": [ + 48, + 512, + 287, + 576 + ], + "type": "text", + "content": "[35] Ye Liu, Lingfeng Qiao, Di Yin, Zhuoxuan Jiang, Xinghua Jiang, Deqiang Jiang, and Bo Ren. Os-msl: One stage multimodal sequential link framework for scene segmentation and classification. In Proceedings of the 30th ACM International Conference on Multimedia, pages 6269-6277, 2022. 2, 3, 4, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 578, + 287, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 287, + 643 + ], + "type": "text", + "content": "[36] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10012-10022, October 2021. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 646, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 646, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 646, + 287, + 689 + ], + "type": "text", + "content": "[37] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swim transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3202-3211, 2022. 1, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 691, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 712 + ], + "type": "text", + "content": "[38] Jakub Lokoc, Gregor Kovalcik, Tomáš Souček, Jaroslav Moravec, and Přemysl Čech. A framework for effective" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "text", + "content": "known-item search in video. In Proceedings of the 27th ACM International Conference on Multimedia, pages 1777-1785, 2019. 2, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 308, + 107, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 107, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 308, + 107, + 545, + 140 + ], + "type": "text", + "content": "[39] Sauradip Nag, Xiatian Zhu, Yi-Zhe Song, and Tao Xiang. Zero-shot temporal action detection via vision-language prompting. arXiv preprint arXiv:2207.08184, 2022. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 142, + 545, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 142, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 308, + 142, + 545, + 196 + ], + "type": "text", + "content": "[40] Guoshun Nan, Rui Qiao, Yao Xiao, Jun Liu, Sicong Leng, Hao Zhang, and Wei Lu. Interventional video grounding with dual contrastive learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2765-2775, 2021. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 198, + 545, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 198, + 545, + 252 + ], + "spans": [ + { + "bbox": [ + 308, + 198, + 545, + 252 + ], + "type": "text", + "content": "[41] Fahmi Salman Nurfikri, Mohamad Syahrul Mubarok, et al. News topic classification using mutual information and bayesian network. In 2018 6th International Conference on Information and Communication Technology (ICoICT), pages 162-166. IEEE, 2018. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 254, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 254, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 308, + 254, + 545, + 308 + ], + "type": "text", + "content": "[42] Sofia Ares Oliveira, Benoit Seguin, and Frederic Kaplan. dhsegment: A generic deep-learning approach for document segmentation. In 2018 16th International Conference on Frontiers in Handwriting Recognition (ICFHR), pages 7-12. IEEE, 2018. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 310, + 545, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 310, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 308, + 310, + 545, + 376 + ], + "type": "text", + "content": "[43] Alejandro Pardo, Fabian Caba Heilbron, Juan León Alcazar, Ali Thabet, and Bernard Ghanem. Moviecuts: A new dataset and benchmark for cut type recognition. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part VII, pages 668-685, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 377, + 545, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 377, + 545, + 421 + ], + "spans": [ + { + "bbox": [ + 308, + 377, + 545, + 421 + ], + "type": "text", + "content": "[44] Will Price, Carl Vondrick, and Dima Damen. Unweavenet: Unweaving activity stories. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13770-13779, 2022. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 422, + 545, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 422, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 308, + 422, + 545, + 498 + ], + "type": "text", + "content": "[45] Zhiwu Qing, Shiwei Zhang, Ziyuan Huang, Yi Xu, Xiang Wang, Mingqian Tang, Changxin Gao, Rong Jin, and Nong Sang. Learning from untrimmed videos: Self-supervised video representation learning with hierarchical consistency. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13821-13831, 2022. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 501, + 545, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 501, + 545, + 566 + ], + "spans": [ + { + "bbox": [ + 308, + 501, + 545, + 566 + ], + "type": "text", + "content": "[46] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 568, + 545, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 568, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 308, + 568, + 545, + 632 + ], + "type": "text", + "content": "[47] Anyi Rao, Linning Xu, Yu Xiong, Guodong Xu, Qingqiu Huang, Bolei Zhou, and Dahua Lin. A local-to-global approach to multi-modal movie scene segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10146-10155, 2020. 2, 3, 4, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 635, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 635, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 308, + 635, + 545, + 679 + ], + "type": "text", + "content": "[48] Daniel Rotman, Dror Porat, and Gal Ashour. Optimal sequential grouping for robust video scene detection using multiple modalities. International Journal of Semantic Computing, 11(02):193-208, 2017. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 681, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 545, + 712 + ], + "type": "text", + "content": "[49] Ellen Rushe and Brian Mac Namee. Anomaly detection in raw audio using deep autoregressive networks. In ICASSP 2019-2019 IEEE International Conference on Acoustics," + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "10678" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 66, + 72, + 286, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 286, + 94 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 286, + 94 + ], + "type": "text", + "content": "Speech and Signal Processing (ICASSP), pages 3597-3601. IEEE, 2019. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 139 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 139 + ], + "type": "text", + "content": "[50] Weimin Shi, Mingchen Zhuge, Zhong Zhou, Dehong Gao, and Deng-Ping Fan. Qr-clip: Introducing explicit open-world knowledge for location and time reasoning. arXiv preprint arXiv:2302.00952, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 140, + 287, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 287, + 194 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 287, + 194 + ], + "type": "text", + "content": "[51] Mike Zheng Shou, Stan Weixian Lei, Weiyao Wang, Deepti Ghadiyaram, and Matt Feiszli. Generic event boundary detection: A benchmark for event segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8075-8084, 2021. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 195, + 287, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 195, + 287, + 260 + ], + "spans": [ + { + "bbox": [ + 48, + 195, + 287, + 260 + ], + "type": "text", + "content": "[52] Nina Shvetsova, Brian Chen, Andrew Rouditchenko, Samuel Thomas, Brian Kingsbury, Rogerio S Feris, David Harwath, James Glass, and Hilde Kuehne. Everything at once-multimodal fusion transformer for video retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20020-20029, 2022. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 261, + 287, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 261, + 287, + 326 + ], + "spans": [ + { + "bbox": [ + 48, + 261, + 287, + 326 + ], + "type": "text", + "content": "[53] Mattia Soldan, Alejandro Pardo, Juan León Alcazar, Fabian Caba, Chen Zhao, Silvio Giancola, and Bernard Ghanem. Mad: A scalable dataset for language grounding in videos from movie audio descriptions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5026-5035, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 327, + 287, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 327, + 287, + 381 + ], + "spans": [ + { + "bbox": [ + 48, + 327, + 287, + 381 + ], + "type": "text", + "content": "[54] Yale Song, Miriam Redi, Jordi Vallmitjana, and Alejandro Jaimes. To click or not to click: Automatic selection of beautiful thumbnails from videos. In Proceedings of the 25th ACM international on conference on information and knowledge management, pages 659-668, 2016. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 383, + 287, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 383, + 287, + 415 + ], + "spans": [ + { + "bbox": [ + 48, + 383, + 287, + 415 + ], + "type": "text", + "content": "[55] Tomáš Souček and Jakub Lokoc. Transnet v2: an effective deep network architecture for fast shot transition detection. arXiv preprint arXiv:2008.04838, 2020. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 416, + 287, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 416, + 287, + 470 + ], + "spans": [ + { + "bbox": [ + 48, + 416, + 287, + 470 + ], + "type": "text", + "content": "[56] Jiaqi Tang, Zhaoyang Liu, Chen Qian, Wayne Wu, and Limin Wang. Progressive attention on multi-level dense difference maps for generic event boundary detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3355-3364, 2022. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 471, + 287, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 471, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 48, + 471, + 287, + 514 + ], + "type": "text", + "content": "[57] Shitao Tang, Litong Feng, Zhanghui Kuang, Yimin Chen, and Wei Zhang. Fast video shot transition localization with deep structured models. In Asian Conference on Computer Vision, pages 577-592. Springer, 2018. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 515, + 287, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 515, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 48, + 515, + 287, + 568 + ], + "type": "text", + "content": "[58] Du Tran, Lubomir Bourdev, Rob Fergus, Lorenzo Torresani, and Manohar Paluri. Learning spatiotemporal features with 3d convolutional networks. In Proceedings of the IEEE international conference on computer vision, pages 4489-4497, 2015. 1, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 571, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 571, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 571, + 287, + 624 + ], + "type": "text", + "content": "[59] Du Tran, Heng Wang, Lorenzo Torresani, and Matt Feiszli. Video classification with channel-separated convolutional networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5552-5561, 2019. 1, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 625, + 287, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 287, + 669 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 287, + 669 + ], + "type": "text", + "content": "[60] Dongkai Wang and Shiliang Zhang. Unsupervised person re-identification via multi-label classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "text", + "content": "[61] Jue Wang and Anoop Cherian. Gods: Generalized one-class discriminative subspaces for anomaly detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8201-8211, 2019. 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "text", + "content": "[62] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaou Tang, and Luc Van Gool. Temporal segment networks: Towards good practices for deep action recognition. In European conference on computer vision, pages 20-36. Springer, 2016. 1, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 129, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 183 + ], + "type": "text", + "content": "[63] Yuxuan Wang, Difei Gao, Licheng Yu, Weixian Lei, Matt Feiszli, and Mike Zheng Shou. Geb+: A benchmark for generic event boundary captioning, grounding and retrieval. In European Conference on Computer Vision, pages 709-725. Springer, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 184, + 545, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 184, + 545, + 248 + ], + "spans": [ + { + "bbox": [ + 308, + 184, + 545, + 248 + ], + "type": "text", + "content": "[64] Jonatas Wehrmann, Ricardo Cerri, and Rodrigo Barros. Hierarchical multi-label classification networks. In Jennifer Dy and Andreas Krause, editors, Proceedings of the 35th International Conference on Machine Learning, volume 80 of Proceedings of Machine Learning Research, pages 5075-5084. PMLR, 10-15 Jul 2018. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 250, + 545, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 250, + 545, + 315 + ], + "spans": [ + { + "bbox": [ + 308, + 250, + 545, + 315 + ], + "type": "text", + "content": "[65] Haoqian Wu, Keyu Chen, Yanan Luo, Ruizhi Qiao, Bo Ren, Haozhe Liu, Weicheng Xie, and Linlin Shen. Scene consistency representation learning for video scene segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14021-14030, 2022. 2, 3, 4, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 316, + 545, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 316, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 308, + 316, + 545, + 370 + ], + "type": "text", + "content": "[66] Bo Xiong, Yannis Kalantidis, Deepti Ghadiyaram, and Kristen Grauman. Less is more: Learning highlight detection from video duration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1258-1267, 2019. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 372, + 545, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 372, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 308, + 372, + 545, + 425 + ], + "type": "text", + "content": "[67] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. Videoclip: Contrastive pre-training for zero-shot video-text understanding. arXiv preprint arXiv:2109.14084, 2021. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 426, + 545, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 426, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 308, + 426, + 545, + 491 + ], + "type": "text", + "content": "[68] Mengmeng Xu, Juan-Manuel Pérez-Rúa, Victor Escorcia, Brais Martinez, Xiatian Zhu, Li Zhang, Bernard Ghanem, and Tao Xiang. Boundary-sensitive pre-training for temporal localization in videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7220-7230, 2021. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 493, + 545, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 493, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 308, + 493, + 545, + 547 + ], + "type": "text", + "content": "[69] Minghao Xu, Hang Wang, Bingbing Ni, Riheng Zhu, Zhenbang Sun, and Changhu Wang. Cross-category video highlight detection via set-based learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7970-7979, 2021. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 548, + 545, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 548, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 308, + 548, + 545, + 602 + ], + "type": "text", + "content": "[70] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Tubedetr: Spatio-temporal video grounding with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16442-16453, 2022. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 604, + 545, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 604, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 308, + 604, + 545, + 658 + ], + "type": "text", + "content": "[71] Hanhua Ye, Guorong Li, Yuankai Qi, Shuhui Wang, Qingming Huang, and Ming-Hsuan Yang. Hierarchical modular network for video captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17939-17948, 2022. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "type": "text", + "content": "[72] Muhammad Zaigham Zaheer, Arif Mahmood, Marcella Astrid, and Seung-Ik Lee. Claws: Clustering assisted weakly supervised learning with normalcy suppression for anomalous event detection. In European Conference on Computer Vision, pages 358-376. Springer, 2020. 3" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10679" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 339 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[73] M Zaigham Zaheer, Arif Mahmood, M Haris Khan, Mattia Segu, Fisher Yu, and Seung-Ik Lee. Generative cooperative learning for unsupervised video anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14744-14754, 2022. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 287, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 183 + ], + "type": "text", + "content": "[74] Runhao Zeng, Haoming Xu, Wenbing Huang, Peihao Chen, Mingkui Tan, and Chuang Gan. Dense regression network for video grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10287-10296, 2020. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 185, + 287, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 185, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 48, + 185, + 287, + 228 + ], + "type": "text", + "content": "[75] Haoxin Zhang, Zhimin Li, and Qinglin Lu. Better learning shot boundary detection via multi-task. In Proceedings of the 29th ACM International Conference on Multimedia, pages 4730-4734, 2021. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 229, + 287, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 229, + 287, + 274 + ], + "spans": [ + { + "bbox": [ + 48, + 229, + 287, + 274 + ], + "type": "text", + "content": "[76] Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. IEEE transactions on pattern analysis and machine intelligence, 40(6):1452-1464, 2017. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 275, + 287, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 275, + 287, + 339 + ], + "spans": [ + { + "bbox": [ + 48, + 275, + 287, + 339 + ], + "type": "text", + "content": "[77] Mingchen Zhuge, Dehong Gao, Deng-Ping Fan, Linbo Jin, Ben Chen, Haoming Zhou, Minghui Qiu, and Ling Shao. Kaleido-bert: Vision-language pre-training on fashion domain. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12647-12657, 2021. 3" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "10680" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Next3D_ Generative Neural Texture Rasterization for 3D-Aware Head Avatars/7b615aeb-a836-44b1-beba-de01bf0423b2_content_list.json b/2023/Next3D_ Generative Neural Texture Rasterization for 3D-Aware Head Avatars/7b615aeb-a836-44b1-beba-de01bf0423b2_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..064ae700bcf4539036f5c4a5b09a872e08ddbff8 --- /dev/null +++ b/2023/Next3D_ Generative Neural Texture Rasterization for 3D-Aware Head Avatars/7b615aeb-a836-44b1-beba-de01bf0423b2_content_list.json @@ -0,0 +1,1670 @@ +[ + { + "type": "text", + "text": "Next3D: Generative Neural Texture Rasterization for 3D-Aware Head Avatars", + "text_level": 1, + "bbox": [ + 91, + 130, + 879, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jingxiang Sun", + "bbox": [ + 125, + 178, + 250, + 200 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xuan Wang $^{2}$", + "bbox": [ + 289, + 179, + 393, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Lizhen Wang $^{14}$", + "bbox": [ + 433, + 180, + 558, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xiaoyu Li3", + "bbox": [ + 598, + 180, + 689, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yong Zhang3", + "bbox": [ + 728, + 180, + 836, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hongwen Zhang", + "bbox": [ + 346, + 200, + 486, + 219 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yebin Liu", + "bbox": [ + 529, + 200, + 616, + 218 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Tsinghua University", + "bbox": [ + 150, + 223, + 321, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Ant Group", + "bbox": [ + 361, + 224, + 457, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ Tencent AI Lab", + "bbox": [ + 496, + 224, + 630, + 241 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4NNKosmos", + "bbox": [ + 671, + 224, + 774, + 241 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c04ab9bd231f8e61758d2d3d2122949aed22cecb230f0433f950fe327204b0b5.jpg", + "image_caption": [ + "Gaze Animation", + "3D-Aware Domain Adaptation", + "Figure 1. Our 3D GAN synthesizes generative, high-quality, and 3D-consistent facial avatars from unstructured 2D images. Unlike current animatable 3D GANs that only modify yaw-pitch head poses and facial expressions, our approach enables fine-grained control over full-head rotations, facial expressions, eye blinks, and gaze directions with strict 3D consistency and a high level of photorealism. Our approach also provides strong 3D priors for downstream tasks such as 3D-aware stylization." + ], + "image_footnote": [], + "bbox": [ + 78, + 276, + 893, + 494 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 595, + 313, + 611 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D-aware generative adversarial networks (GANs) synthesize high-fidelity and multi-view-consistent facial images using only collections of single-view 2D imagery. Towards fine-grained control over facial attributes, recent efforts incorporate 3D Morphable Face Model (3DMM) to describe deformation in generative radiance fields either explicitly or implicitly. Explicit methods provide fine-grained expression control but cannot handle topological changes caused by hair and accessories, while implicit ones can model varied topologies but have limited generalization caused by the unconstrained deformation fields. We propose a novel 3D GAN framework for unsupervised learning of generative, high-quality and 3D-consistent facial avatars from unstructured 2D images. To achieve both deformation accuracy and topological flexibility, we propose a 3D representation called Generative Texture-Rasterized Tri-planes. The proposed representation learns Generative Neural Textures on top of parametric mesh templates", + "bbox": [ + 73, + 628, + 472, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "and then projects them into three orthogonal-viewed feature planes through rasterization, forming a tri-plane feature representation for volume rendering. In this way, we combine both fine-grained expression control of mesh-guided explicit deformation and the flexibility of implicit volumetric representation. We further propose specific modules for modeling mouth interior which is not taken into account by 3DMM. Our method demonstrates state-of-the-art 3D-aware synthesis quality and animation ability through extensive experiments. Furthermore, serving as 3D prior, our animatable 3D representation boosts multiple applications including one-shot facial avatars and 3D-aware stylization. Project page: https://mrtornado24.github.io/Next3D/. Code: https://github.com/MrTornado24/Next3D.", + "bbox": [ + 496, + 597, + 893, + 809 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 843, + 632, + 859 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Animatable portrait synthesis is essential for movie postproduction, visual effects, augmented reality (AR), and vir", + "bbox": [ + 498, + 869, + 892, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "20991", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tual reality (VR) telepresence applications. Efficient animatable portrait generators should be capable of synthesizing diverse high-fidelity portraits with full control of the rigid head pose, facial expressions and gaze directions at a fine-grained level. The main challenges of this task lie in how to model accurate deformation and preserve identity through animation in the generative setting, i.e. training with only unstructured corpus of 2D images.", + "bbox": [ + 75, + 90, + 472, + 212 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Several 2D generative models perform image animation by incorporating the 3D Morphable Face Models (3DMM) [4] into the portrait synthesis [13, 16, 35, 52, 62, 65, 70, 73]. These 2D-based methods achieve photorealism but suffer from shape distortion during large motion due to a lack of geometry constraints. Towards better view consistency, many recent efforts incorporate 3DMM with 3D GANs, learning to synthesize animatable and 3D consistent portraits from only 2D image collections in an unsupervised manner [3, 30, 39, 44, 60, 61, 68, 74]. Bergman et al. [3] propose an explicit surface-driven deformation field for warping radiance fields. While modeling accurate facial deformation, it cannot handle topological changes caused by non-facial components, e.g. hair, glasses, and other accessories. AnifaceGAN [68] builds an implicit 3DMM-conditioned deformation field and constrains animation accuracy by imitation learning. It achieves smooth animation on interpolated expressions, however, struggles to generate reasonable extrapolation due to the under-constrained deformation field. Therefore, The key challenge of this task is modeling deformation in the 3D generative setting for animation accuracy and topological flexibility.", + "bbox": [ + 75, + 214, + 472, + 547 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose a novel 3D GAN framework for unsupervised learning of generative, high-quality, and 3D-consistent facial avatars from unstructured 2D images. Our model splits the whole head into dynamic and static parts, and models them respectively. For dynamic parts, the key insight is to combine both fine-grained expression control of mesh-guided explicit deformation and flexibility of implicit volumetric representation. To this end, we propose a novel representation, Generative Texture-Rasterized Triplanes, which learns the facial deformation through Generative Neural Textures on top of a parametric template mesh and samples them into three orthogonal-viewed and axis-aligned feature planes through standard rasterization, forming a tri-plane feature representation. Such texture-rasterized tri-planes re-form high-dimensional dynamic surface features in a volumetric representation for efficient volume rendering and thus inherit both the accurate control of the mesh-driven deformation and the expressiveness of volumetric representations. Furthermore, we represent static components (body, hair, background, etc.) by another triplane branch, and integrate both through alpha blending.", + "bbox": [ + 75, + 550, + 472, + 867 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Another key insight of our method is to model the mouth interior which is not taken into account by 3DMM. Mouth", + "bbox": [ + 76, + 869, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "interior is crucial for animation quality but often ignored by prior arts. We propose an efficient teeth synthesis module, formed as a style-modulated UNet, to complete the inner mouth features missed by the template mesh. To further regularize the deformation accuracy, we introduce a deformation-aware discriminator which takes as input synthetic renderings, encouraging the alignment of the final outputs with the 2D projection of the expected deformation.", + "bbox": [ + 496, + 90, + 890, + 212 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To summarize, the contributions of our approach are:", + "bbox": [ + 517, + 212, + 870, + 227 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We present an animatable 3D-aware GAN framework for photorealistic portrait synthesis with fine-grained animation, including expressions, eye blinks, gaze direction and full head poses.", + "- We propose Generative Texture-Rasterized Triplanes, an efficient deformable 3D representation that inherits both fine-grained expression control of mesh-guided explicit deformation and flexibility of implicit volumetric representation.", + "- Our learned generative animatable 3D representation can serve as a strong 3D prior and boost the downstream application of 3D-aware one-shot facial avatars. Our model also pushes the frontier of 3D stylization with high-quality out-of-domain facial avatars." + ], + "bbox": [ + 517, + 233, + 890, + 460 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 467, + 640, + 483 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Generative 3D-aware Image Synthesis. Generative adversarial networks [24] have achieved photorealistic synthesis in 2D domain. Building on the success of 2D GANs, many efforts have lifted the image synthesis into 3D with explicit view control. Early voxel-based approaches [19, 28, 41, 42, 67, 82] adopt the 3D CNN generators whose heavy computational burden limits the high-resolution image synthesis. Recent works incorporate more efficient neural scene representations, such as fully implicit networks [6, 8-10, 14, 47, 54, 57, 81], sparse voxel grids [55], multiple planes [79] or a combination of low-resolution feature volume and 2D super-resolution [7, 27, 43, 46, 71, 72, 76, 78]. We leverage the tri-plane representation proposed in [7] and endow it with animation ability by the orthogonal-rasterized Generative Neural Textures.", + "bbox": [ + 496, + 492, + 890, + 718 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Some other current works [3,30,36,44,58,59,61,68,77, 83] focus on the editability of 3D-aware generative models. FENeRF [59] and IDE-3D [58] perform semantic-guided 3D face editing by incorporating semantic-aware radiance fields and GAN inversion. However, they cannot produce continuous and stable editing on videos. Other methods [3, 30,44,61,68] employ 3D priors to achieve animatable image synthesis and the main differences lie on the deformation strategies including linear blend skinning [30,44], surface-driven deformation [3], 3DMM-guided latent decomposition [61] and neural deformation fields [68]. These approaches either don't allow for topology changes or need", + "bbox": [ + 496, + 719, + 890, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "20992", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a1dbff3e376f7c470d28d06fbe8a108f32f1ed7c3003834e4a0de514c95607cb.jpg", + "image_caption": [ + "Figure 2. Our 3D GAN framework consists of two tri-plane branches $T_{uv}$ and $T_{static}$ modeling dynamic and static components. $T_{uv}$ is formed by the orthogonal rasterized Generative Neural Textures which are synthesized by a StyleGAN generator, $G_{texture}$ , on top of deformable template mesh. $T_{static}$ is synthesized by another StyleGAN generator, $G_{static}$ . The mouth synthesis module, $G_{teeth}$ , is presented for completing mouth interior. Blended triplanes are incorporated with hybrid neural renderer consisting of volume rendering and a super-resolution module $G_{superres}$ . For discrimination, synthetic renderings $I_{synthetic}$ is taken into the dual discriminator $D_{dual}$ ." + ], + "image_footnote": [], + "bbox": [ + 81, + 87, + 893, + 280 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "elaborate loss design to ensure the accuracy of deformation. On the contrary, our approach naturally achieves accurate animation with explicit mesh guidance and further allows for topology changes by adapting surface deformation into a continuous volumetric representation.", + "bbox": [ + 75, + 376, + 467, + 450 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Facial animation with 3D morphable face models. Blanz et al. [4] model facial texture and shape as vector spaces, known as the 3D Morphable Model (3DMM). Extensions of 3DMM, such as full-head PCA models [12, 51], blend-shape models [38], are extensively studied and widely used in facial animation tasks [23, 65]. Benefiting from 3DMM, these methods can model the deformation of facial parts accurately and continuously, nevertheless, struggle to represent non-facial areas missed by 3DMM, e.g. hair, teeth, eyes, and body. Moreover, these methods are prone to lacking facial details. To fill the missing areas and complete more realistic facial details, later works [13, 16, 18, 22, 35, 52, 62, 64, 70] apply learned approaches on top of 3DMM renderings. DiscoFaceGAN [13] maps 3DMM parameters into the latent space and decouple them by imitative-contrastive learning. Though efficient and photorealistic animation is achieved, these 2D methods don't model 3D geometry and thus cannot remain strict 3D consistency with large head pose changes. For strict 3D consistency, recent efforts [1, 20, 25, 75, 80] incorporate 3DMM into volumetric representations [31, 32, 40, 49, 56] to achieve view-consistent facial animation. Furthermore, volumetric representations enable the modeling of thin structures such as hair, and also mouth interior thanks to its spatial continuity. While 3DMM have been adopted to animate radiance fields for single-scene scenarios, it is challenging to adapt to the generative setting with the absence of groundtruth supervision.", + "bbox": [ + 75, + 457, + 467, + 864 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Neural scene representations. The neural scene representations can be roughly categorized into implicit and explicit", + "bbox": [ + 75, + 869, + 467, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "surface representations and volumetric representations [63]. The surface can be represented explicitly by point clouds [37, 50], meshes [2, 5, 45, 64], or defined implicitly as a zero level-set of a function [11, 48, 69], like signed distance function, which can be approximated by coordinate-based multi-layer perceptrons (MLPs). On the contrary, volumetric representations [32, 40, 49, 56] store volumetric properties (occupancies, radiance, colors, etc.) instead of the surface of an object. These properties can be stored in explicit voxel grids [32, 49, 56] or the weights of a neural network implicitly [40]. In this work, we propose a hybrid surface-volumetric representation. Specifically, we learn the deformable surface radiance by neural textures [26, 64] on top of the template mesh and rasterize it into three orthogonal-viewed feature planes. Then, the planes are reshaped to a tri-plane representation with decoding into neural radiance fields for volume rendering.", + "bbox": [ + 496, + 376, + 890, + 632 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Approach", + "text_level": 1, + "bbox": [ + 500, + 659, + 607, + 676 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We present an animatable 3D-aware facial generator that equips with fine-grained expression and pose control, photorealistic rendering quality, and high-quality underlying geometry. The proposed method models dynamic and static components by two independent tri-plane branches. Specifically, we propose Generative Texture-rasterized Tri-planes for modeling dynamic facial parts (Sec. 3.1). Furthermore, we propose an efficient mouth synthesis module to complete the mouth interior that is not included in 3DMM (Sec. 3.2). We further adopt another tri-plane branch for the static components (Sec. 3.3). Both tri-planes are blended together for hybrid neural rendering (Sec. 3.4). We introduce an deformation-aware discriminator (Sec. 3.5) and illustrate the training objectives in Sec. 3.6.", + "bbox": [ + 496, + 689, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "20993", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Generative texture-rasterized tri-planes", + "text_level": 1, + "bbox": [ + 76, + 90, + 418, + 107 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "EG3D [7] presents an efficient tri-plane-based hybrid 3D representation to synthesize high-resolution images with multi-view consistency. Nonetheless, EG3D lacks control over facial deformations and thus cannot be directly applied to animation tasks. To this end, we leverage Neural textures [64] to represent deformable facial parts. In general, Neural Textures are a set of learned high-dimensional feature maps that can be interpreted by a neural renderer. We extend it to our generative setting and synthesize the neural textures through a StyleGAN2 CNN generator $G_{\\text{texture}}$ . As shown in Fig. 2, we first sample a latent code $z$ and map it into an intermediate latent space by the mapping network. Our texture generator architecture closely follows StyleGAN2 backbone [34], except producing a $256 \\times 256 \\times 32$ neural texture map, $T$ , instead of a three-channel RGB image. Storing a high-dimensional learned feature vector per texel, $T$ can be rasterized to a view-dependent screen-space feature map given a mesh with uv-texture parameterization and a target view as input. In our case, we use the FLAME template [38] to provide a coarse mesh that can be driven by deformation parameters. Given the predesigned texture mapping function, we employ the standard graphics pipeline to rasterize the neural textures from the texture space into the screen space based on the template mesh. We choose Neural Textures as the deformation method for two reasons. First, compared with other explicit deformation (e.g. linear blend skinning and surface-driven deformation) highly dependent on the accurate underlying geometry, Neural Textures embed high-level features which compensate the imperfect geometry and thus are more suitable for our settings where template meshes are not accurate. Furthermore, unlike implicit deformation methods [13, 61, 68], our explicit mesh-guided deformation alleviates the requirement of elaborate imitation learning while gaining better expression generalization (Fig. 4).", + "bbox": [ + 75, + 114, + 472, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Neural Textures encode surface deformation accurately with mesh guidance but lack generalization to 3D points far from surface. Besides, it also doesn't allow for topological changes. To this end, we propose Generative Texture-Rasterized Tri-planes, $T_{uv}$ which reshapes the rasterized textures into a tri-plane representation. Therefore, we can adapt such surface deformation into a continuous volume, and once a 3D point out of the surface could be projected within the facial part on at least one view (e.g. glasses), its feature can be interpolated by the meaningful surface feature through grid sampling in the tri-plane. Specifically, we rasterize neural textures based on the template mesh into three orthogonal views and place them in three axis-aligned feature planes. In practice, considering the zygomorphy, the rasterization is applied at both the left and right views and the rasterized features are concatenated for one single plane by summation. In this way, Our hybrid surface-volumetric", + "bbox": [ + 75, + 643, + 472, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "representation inherits the best of both worlds: accurate mesh-guided facial deformation of Neural Textures and the continuity and topological flexibility of volumetric representations. See Fig. 5 for the topological-aware animation of portraits with glasses.", + "bbox": [ + 496, + 90, + 890, + 167 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Mouth synthesis module", + "text_level": 1, + "bbox": [ + 498, + 179, + 723, + 195 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/71e47e74911ff83045eae0d1df72dc8e8b6ba1b2fe084fd9c70ff17820e52c0a.jpg", + "image_caption": [ + "Figure 3. The teeth synthesis module consists of a teeth completing module and a neural blending module. The teeth competing module produces mouth interior conditioned on multi-scale teeth exterior features and latent codes $w$ ." + ], + "image_footnote": [], + "bbox": [ + 501, + 215, + 890, + 383 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since the FLAME template doesn't contain inner mouth, we propose a teeth synthesis module, $G_{\\text{teeth}}$ , to complete the missing teeth features in $T_{uv}$ . As shown in Fig. 3, for each feature plane of $T_{uv}$ , we crop the teeth area by the expanded landmarks and resize it into $64 \\times 64$ . Then, the stacked mouth features are processed by $G_{\\text{teeth}}$ which employs a style-modulated UNet [66]. The downsampling process of $G_{\\text{teeth}}$ encodes $f_{\\text{teeth}}$ into multi-scale feature maps which serve as content conditions for the following StyleGAN layers. The output teeth features $f_{\\text{teeth}}'$ are transformed inversely and concatenated with the feature planes of $T_{uv}$ . To eliminate the texture flickering of mouth boundary, we further feed $T_{uv}$ into a shallow-UNet-based neural blending module and obtain $T_{uv}'$ . We conduct a series of ablation studies and prove that the proposed teeth synthesis module brings a remarkable improvement on both the animation accuracy and synthesis quality (Sec. 4.2).", + "bbox": [ + 496, + 455, + 892, + 712 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Modeling static components", + "text_level": 1, + "bbox": [ + 500, + 724, + 751, + 742 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The generative texture-rasterized tri-planes manage to model dynamic faces varying expressions and shapes, though, it is challenging to synthesize static parts like diverse haircut, background and upper body which are not included in the FLAME template. To this end, we model these parts by another tri-plane branch, $T_{\\text{static}}$ , which is generated by a StyleGAN2 CNN generator $G_{\\text{static}}$ sharing the same latent code with $G_{\\text{texture}}$ . The plane features of $T_{uv}^{\\prime}$ and $T_{\\text{static}}$ are blended on each plane by the alpha masks rendered by rasterization. Such design not only benefits the", + "bbox": [ + 496, + 750, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "20994", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "modeling of various-styled static components but also enforces their consistency during facial animations.", + "bbox": [ + 76, + 90, + 468, + 121 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Neural rendering", + "text_level": 1, + "bbox": [ + 76, + 130, + 246, + 147 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given the blended tri-planes, for any point in the 3D space, we project it into each plane and sample the features bi-linearly. Then, the sampled features are aggregated by summation and decoded into volume density $\\sigma$ and feature $f$ by a lightweight decoder. Similar to [7,58], the decoder is a single hidden layered multi-layer perceptron (MLP) with softplus activation. The volume rendering is employed to accumulate $\\sigma$ and $f$ along the rays cast through each pixel to compute a 2D feature image $I_{f}$ . Similar to [7,27,46], we leverage a 2D super-resolution module $G_{superres}$ to interpret the feature image into RGB image $I_{RGB}$ with higher resolution. The super-resolution module consists of three StyleGAN2 synthesis blocks and the noise input is removed for alleviating texture flickers. In our case, $I_{f}$ and $I_{RGB}$ are set to $64\\times 64$ and $512\\times 512$ , respectively.", + "bbox": [ + 75, + 152, + 472, + 380 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Deformation-aware discriminator", + "text_level": 1, + "bbox": [ + 76, + 388, + 372, + 404 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To learn the unsupervised 3D representations, we adopt a 2D convolutional discriminator $D$ to critique the renderings. Inspired by [7], we regularize the first three channels of $I_{f}$ as low-resolution RGB image, which is concatenated with $I_{RGB}$ as the input for the discriminator. However, the discrimination with only image input can only ensure that the deformed images are always in a correct distribution instead of matching the expected deformations. Therefore, we make the discriminator aware of the expression and shape under which the generated image are deformed. For the same purpose, GNARF [3] conditions the discriminator on the FLAME parameters by concatenating them as the input to the mapping network. However, we find empirically that such a conditioning method leads to training instability, consistent with [3]. Instead, we re-render the template mesh under the rendered pose to get the synthetic rendering $I_{synthetic}$ and feed it into $D_{\\text{dual}}$ along with image pairs. Here, we adopt correspondence images inspired by [35]. Such concatenation encourages the final output to align with the synthetic rendering and learn the expected deformation.", + "bbox": [ + 75, + 412, + 472, + 715 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.6. Training objectives", + "text_level": 1, + "bbox": [ + 76, + 723, + 261, + 739 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During training, we use the non-saturating GAN loss with R1 regularization. Moreover, we adopt the density regularization proposed in EG3D [7]. Therefore, the total learning objective is:", + "bbox": [ + 76, + 746, + 468, + 806 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {D _ {\\text {d u a l}, G}} = \\mathbb {E} _ {z \\sim p _ {z}, \\epsilon \\sim p _ {\\epsilon}} [ f (D _ {\\text {d u a l}} (G (z, \\epsilon))) ] + \\\\ \\mathbb {E} _ {I ^ {r} \\sim p _ {I ^ {r}}} [ f (- D _ {\\text {d u a l}} (I ^ {r}) + (1) \\\\ \\left. \\lambda \\| \\bigtriangledown D _ {d u a l} \\left(I ^ {r}\\right) \\| ^ {2} \\right], \\\\ \\mathcal {L} _ {\\text {d e n s i t y}} = \\sum_ {x _ {s} \\in \\mathcal {S}} \\| d (x _ {s}) - d (x _ {s} + \\epsilon) \\| _ {2}, (2) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 815, + 468, + 902 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {D _ {\\text {d u a l}}, G} + \\lambda_ {\\text {d e n s i t y}} \\mathcal {L} _ {\\text {d e n s i t y}}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 575, + 104, + 890, + 119 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $I^r$ is the combination of real images, blurred real images, and the corresponding synthetic renderings, which are sampled from the training set with distribution $p_I$ . We adopt many training hyperparameters from EG3D and StyleGAN2 (learning rates of generator and discriminator, batch size, R1 regularization, etc.). We train our model based on the pretrained model of EG3D [7] and continue to train on 4 3090 GPUs for roughly 4 days. Please refer to the supplemental material for the implementation details.", + "bbox": [ + 496, + 128, + 893, + 266 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 277, + 633, + 295 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we first show qualitative and quantitative comparisons to state-of-the-art 2D / 3D animatable generative facial models (Sec. 4.1), and then discuss the conducted ablation studies of our design choices (Sec. 4.2). Furthermore, we show various applications combining our efficient 3D representation with GAN technologies such as GAN inversion and style transfer (Sec. 4.3).", + "bbox": [ + 496, + 303, + 890, + 409 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We train and test our methods on FFHQ [33]. We augment FFHQ with horizontal flips and use an off-the-shelf pose estimator [15] to label images with the approximated camera extrinsic parameters and constant intrinsics. To support full pose animation, in-plane (roll) rotation is also considered. Furthermore, we use DECA [17] to estimate the FLAME parameters of facial identity $\\beta \\in \\mathbb{R}^{100}$ , jaw pose $\\theta_{jaw} \\in \\mathbb{R}^3$ and expression $\\psi \\in \\mathbb{R}^{50}$ . Since DECA doesn't account for eyeball movement, we additionally adopt an efficient facial detector to detect 2D landmarks of irises and optimize eye poses $\\theta_{eye} \\in \\mathbb{R}^6$ by minimizing the re-projection errors. Based on these FLAME parameters, we produce a template mesh with 5023 vertices and 9976 faces to drive facial deformations.", + "bbox": [ + 496, + 409, + 892, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Comparisons", + "text_level": 1, + "bbox": [ + 500, + 628, + 638, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines. We compare our method against two state-of-the-art methods for animatable 3D-aware image synthesis: 3DFaceShop [61], and AniFaceGAN [68]. Besides, we also select DiscoFaceGAN [13] as a baseline, which generates animatable 2D portraits conditioned on 3DMM parameters. Qualitative comparison. Fig. 4 provides a qualitative comparison against baselines. Overall, as can be seen, our method outperforms all baselines by some margin on both synthesis quality and animation accuracy. Specifically, DiscoFaceGAN [13] suffers from inconsistent identity during animation. Moreover, it cannot generate reasonable mouth interior, e.g. stretched teeth. 3DFaceshop and AnifaceGAN synthesize 3D-consistent images, nevertheless, still struggle to model consistent mouth interior with the driving images. This is because their implicit deformation approaches", + "bbox": [ + 496, + 652, + 890, + 878 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "1 https://mediapipe.dev/", + "bbox": [ + 517, + 887, + 645, + 900 + ], + "page_idx": 4 + }, + { + "type": "footer", + "text": "20995", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ae3815df2f0a2a31171a1005563d372e24693015e2a3b8ec4c1162c9344f8af6.jpg", + "image_caption": [ + "Driving", + "Free View" + ], + "image_footnote": [], + "bbox": [ + 107, + 88, + 779, + 176 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c3608cd3755d3e5d9deab540e36e52453e0462f88afbea3b145385dae3875be7.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 107, + 181, + 890, + 268 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b182aff722264e93d8d2b436eda799d715f52ba236c2938ccf8a6c4ac65f3e9d.jpg", + "image_caption": [ + "AntiFaceGAN" + ], + "image_footnote": [], + "bbox": [ + 107, + 272, + 890, + 359 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/bfa21a81a77e31ee6ec3f66b0aecf1efb8cb7235016333588d1752269d2aa997.jpg", + "image_caption": [ + "3DFaceShop" + ], + "image_footnote": [], + "bbox": [ + 107, + 364, + 890, + 452 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c1cc234d65e8b8316cda7fa7ddc34a56ca315f319037992a7fd82a3751259dca.jpg", + "image_caption": [ + "DiscoFaceGAN" + ], + "image_footnote": [], + "bbox": [ + 107, + 458, + 890, + 546 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d0db2f87db9803a0da906aa002e9b16abc4583b704cc5bf121185b5bdce93ba5.jpg", + "image_caption": [ + "Figure 4. Comparison with the state-of-the-art animatable 3D & 2D image synthesis methods. We extract several frames from a video clip and use the interpreted face model parameters to animate random virtual avatars." + ], + "image_footnote": [], + "bbox": [ + 81, + 609, + 173, + 681 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ec97242d1e2a36766e4e1f9e0ef73896b9644c04e13ea18d6003add53b00cf63.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 178, + 609, + 271, + 681 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/491c44e32c6fe6b05d1ec0d69a8d1361306a65d270634e0d87a13c30bcac7310.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 276, + 609, + 370, + 681 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6b50011c5c4550f1a80fe1cc2524b78f9d08ce245533e4a57c4857ccd52973e4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 374, + 609, + 467, + 681 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/747fabddb71b00197f5cde37368f35b8e1132323cc3564f89b80afa35ca2b01e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 686, + 173, + 758 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8d34245a79b281cadd16e8759c45d5da48d5c693dae5e77dbaa2cfccc22ff6ed.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 178, + 686, + 271, + 758 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a89aef0716d09913f5dfbcc52fe9fcb23ab97eca1bcd2e296b381b9514969887.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 276, + 686, + 370, + 758 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d669a0a9d1c38bc3638dff1a521df2e5a05963cf25436f60a8a981d0686de340.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 374, + 686, + 467, + 758 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/08c3f15f9b04411dead30dee5f32042e70407fdc0a968c30f4bf75a36089db44.jpg", + "image_caption": [ + "Figure 5. High-quality dynamic shapes with topological changes. As can be seen, we model detailed dynamic shapes of eyelids and lips, while keeping glasses unchanged." + ], + "image_footnote": [], + "bbox": [ + 81, + 758, + 173, + 792 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b65e08731f9e60d2cd05d296b57f91eafea40bdbe8789a17e10d50ba13ee3190.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 178, + 758, + 271, + 792 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3df380a8b1750ce5e9429c69f72d2b4f418940a89ae3904bbf0db446acdb7901.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 276, + 758, + 370, + 792 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/83691ef05468f4cc561e3309aed50b804bdf082ea5f96fdd68a623bb8b09c2cc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 374, + 758, + 467, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "are under-constrained, leading to overfit the expression bias (smiling with mouth half-opened) of datasets. Compared to the other methods, our approach not only synthesizes im", + "bbox": [ + 76, + 854, + 467, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/753c4c27868f9bccbff94fbc8131234382397a0fcc33e2e9c3e32e7e724e0471.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
FID↓AED↓APD↓APD*↓ID↑
DiscoFaceGAN (2562) [13]17.10.420.0460.0240.73
AniFaceGAN (2562) [68]20.10.250.0410.0220.82
3DFaceShop (5122) [61]23.70.310.0450.0240.75
Ours (2562)3.70.160.0180.0130.85
Ours (5122)3.90.160.0230.0190.84
", + "bbox": [ + 500, + 609, + 893, + 691 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Quantitative comparison using FID, average expression distance (AED), average pose distance (APD), and identity consistency (ID) for FFHQ. APD* means calculating pose distance with roll fixed.", + "bbox": [ + 498, + 702, + 890, + 756 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ages with higher quality, but also preserves more detailed expressions of the driver images, including mouth interior, eye blinks and eye movements. Furthermore, we are the only method that supports in-plane head rotations. Fig. 5 provides visual examples of the synthesized high-quality geometry. Our approach can model detailed shape deformations (see zoomed eyelids and lips in Fig. 5) with topological awareness, i.e. glasses are kept unchanged.", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "20996", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5eaaba2d514d716f07aed2262464eea28519aee93db48132f33c0c3f89c9d905.jpg", + "image_caption": [ + "Figure 6. Ablation study on model designs. $T_{\\text{static}}$ encourages better identity consistency; $G_{\\text{teeth}}$ benefits realistic mouth interior; the discriminator with $I_{\\text{synthetic}}$ input allows more consistent reconstruction of detailed expressions." + ], + "image_footnote": [], + "bbox": [ + 84, + 87, + 467, + 536 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative evaluation. Tab. 1 demonstrates quantitative results comparing our method against baselines evaluated on several metrics. We measure image quality with Frechet Inception Distance (FID) [29] between the entire FFHQ dataset and $50\\mathrm{k}$ generated images using randomly sampled latent codes, camera poses, and FLAME parameters. Since AntiFaceGAN and DiscoFaceGAN synthesize images on the resolution of $256^2$ , we test FID of them on FFHQ256². Following [3, 39], we evaluate the faithfulness of the animation with the Average Expression Distance (AED), the Average Pose Distance (APD), and identity consistency (ID). For each method, we randomly sample 500 identities and animate each with randomly sampled 20 FLAME parameters of expressions and poses. Then, we estimate the FLAME parameters for these 10000 generated images and the average distances between the driving FLAME parameters and the reconstructed ones. For identity consistency, we randomly sample 2000 poses, 2000 sets", + "bbox": [ + 75, + 628, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/755c28a4c6a7aae096a3b027ec7e7eb8818b79d19897c9ad746aef4705225ddd.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
FID↓AED↓APD↓ID↑
w/o Tstatic6.60.250.0260.63
w/o Gteeth(uv)7.20.370.0420.71
w/o Gteeth(s)8.40.320.0360.79
w/o Isynthetic3.80.180.0250.74
Full3.90.160.0230.84
", + "bbox": [ + 532, + 88, + 859, + 186 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Ablation study on model designs. Modeling static components by $T_{\\text{static}}$ improves identity consistency. The teeth synthesis $G_{\\text{teeth}}$ module benefits both animation and synthesis quality significantly while adding synthetic renderings $I_{\\text{synthetic}}$ into discrimination takes both a step further.", + "bbox": [ + 498, + 198, + 890, + 268 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "of FLAME parameters, and 1000 identities. Then we randomly select two poses and two sets of FLAME parameters for each identity, generating a total of 1000 image pairs. We calculate consistency metric using a pre-trained Arcface model [15] for each image pair and report the average result. Since the other baselines don't support in-plane (roll) head rotation, we further report APD* which only accounts for poses on yaw and pitch. Our method achieves the best performance on all metrics. Note that our model demonstrates significant improvements in FID, bringing animatable 3D GAN to the same level as unconditional 3D GANs (4.7 for EG3D [7]). For AED and APD, we also show superiority against baselines. Note that we still achieve the best pose consistency (0.019) when only considering yaw and pitch.", + "bbox": [ + 498, + 286, + 890, + 498 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Ablation study", + "text_level": 1, + "bbox": [ + 500, + 511, + 648, + 526 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Static tri-planes. As suggested by the grey lines in Fig. 8, this baseline removes the static tri-planes $T_{static}$ and entangles both dynamic and static components in $T_{uv}$ . As illustrated in the first row of Fig. 6, the identities change when varying expressions. Since there are no explicit constraints for identity consistency, the model would be prone to unexpected entanglement between expression and identity. Tab. 2 shows a similar trend where removing $T_{static}$ leads to worse identity consistency.", + "bbox": [ + 496, + 535, + 890, + 671 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Mouth Synthesis. When removing the mouth synthesis module $G_{\\text{teeth}}$ , we consider two altered choices: representing mouth features by $T_{\\text{static}}$ or $T_{uv}$ , named w/o $G_{\\text{teeth}}(s)$ (red lines in Fig. 8) and w/o $G_{\\text{teeth}}(uv)$ (blue lines in Fig. 8), respectively. The first baseline, illustrated in the second row of Fig. 6, suffers from 'hole' artifacts of mouth. This is because inferior teeth move along with the jaw rotations and thus cannot be modeled by static features. The second baseline modeling teeth with $T_{uv}$ also leads to an unreasonable mouth interior since the FLAME template doesn't account for teeth area and the neural textures for teeth would be sampled from other unrelated areas leading to artifacts. Quantitatively, both baselines without $G_{\\text{teeth}}$ show significant degradation in AED and APD.", + "bbox": [ + 496, + 672, + 890, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Deformation-aware discriminator. Tab. 2 demonstrates", + "bbox": [ + 500, + 885, + 890, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "20997", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/33d78dd48624762e1e3900456db4959f77cff16ac5f4f4d46f8d0fdbb74e7458.jpg", + "image_caption": [ + "Source" + ], + "image_footnote": [], + "bbox": [ + 98, + 119, + 290, + 195 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/06f12f87eb4149f216329656db72dcbc8a7d140cf3dfb717936ca9d15eb74d18.jpg", + "image_caption": [ + "Inversion", + "Animation" + ], + "image_footnote": [], + "bbox": [ + 295, + 119, + 580, + 196 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b835e071ba2f83a9cef41d244d51c1b9d4c73d4555dc887a9c0dc7e946a229e5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 583, + 92, + 846, + 195 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/51195a567cff7979293e60e15960e03e8e69678b96fc431a268c2f0b97dd8dcb.jpg", + "image_caption": [ + "Figure 7. Applications of our model. We use PTI [53] to fit 3D-aware avatars for real portraits and animate them with sampled video clips. Furthermore, we leverage StyleGAN-NADA [21] into 3D settings and adapt these avatars into textually-prescribed domains." + ], + "image_footnote": [], + "bbox": [ + 98, + 199, + 290, + 273 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e83a8e6a044c63386ffa5984d7aafe634c56ff2bbf9f3b7b7947288886843380.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 294, + 199, + 580, + 273 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bf09cbe0ec110d786ce9195b4a62a8ee47f138e113020020216865a347b56f3d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 584, + 199, + 870, + 273 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1f35502efc539fa94056ed3f81f193815f4d51436468456966ac2b8da92d4bbc.jpg", + "image_caption": [ + "(a) Default Structure" + ], + "image_footnote": [], + "bbox": [ + 99, + 325, + 251, + 531 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1ec1bb7919d793fe0c6701920479a6b6ccf854feba233e084144bb915b512e1a.jpg", + "image_caption": [ + "(b) Altered Structures", + "Figure 8. Illustrations of the proposed three variants of our model for ablation study." + ], + "image_footnote": [], + "bbox": [ + 256, + 325, + 452, + 532 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "that the deformation-aware discriminator with $I_{\\text{synthetic}}$ input improves both animation accuracy and identity consistency, at the negligible expense of slightly reduced image quality. In Fig. 6, we see that this design shows a better reconstruction of detailed expressions, e.g. eye blinks.", + "bbox": [ + 75, + 594, + 468, + 669 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Applications", + "text_level": 1, + "bbox": [ + 76, + 680, + 210, + 696 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "One-shot portrait animation. Fig. 7 shows the application of our model for one-shot head avatars. The learned generative animatbale 3D representation with expressive latent space can serve as a strong 3D prior for high-fidelity single-view 3D reconstruction and animation. Note that we can generate natural and consistent animations without video data training.", + "bbox": [ + 75, + 703, + 468, + 809 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Animatable 3D-aware stylization. Inspired by IDE-3D [58], we incorporate 2D CLIP-guided style tranfer methods [21] with our animatable 3D representation for 3D-aware portrait stylization. The right three columns of Fig. 7 show examples of text-driven, stylized portrait animation. Specifically, to adapt a pre-trained model through", + "bbox": [ + 75, + 810, + 468, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "only a textual prompt, we optimize the generator with two kinds of CLIP-based guidance [21]. However, leveraging text-guided 2D methods directly into the 3D setting is challenging as it tends to break 3D awareness and deformation awareness inherited in the generator parameters. To this end, we make some necessary modifications to the training framework. Please refer to the supplemental material for details. As shown in Fig. 7, we achieve high-quality 3D-aware stylized portrait synthesis with preserving well properties (i.e. 3D consistency and accurate animation).", + "bbox": [ + 496, + 327, + 892, + 479 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Limitations and future work", + "text_level": 1, + "bbox": [ + 500, + 492, + 761, + 508 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Though our approach enables reasonable extrapolation on some rare expressions, it struggles to model some others with full consistency, such as one-side mouth up, sticking tongue out, etc. Besides, the rendered portraits still suffer from slight texture flickering caused by 2D superresolution. We leave both for future work. Furthermore, our model has the potential to provide a strong 3D prior to accelerating person-specific avatar reconstruction.", + "bbox": [ + 496, + 518, + 890, + 638 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 652, + 617, + 669 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have presented Next3D, a novel animatable 3D representation for unsupervised learning of high-quality and 3D-consistent virtual facial avatars from unstructured 2D images. Our approach has pushed the frontier of photorealistic animatabale 3D-aware image synthesis. Serving as a strong 3D prior, we believe our learned 3D representation will boost a series of downstream applications including 3D-aware one-shot facial avatars and animatable out-of-domain avatar generation.", + "bbox": [ + 496, + 679, + 890, + 815 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 500, + 829, + 666, + 845 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper is supported by National Key RD Program of China (2022YFF0902200) and the NSFC project No.62125107.", + "bbox": [ + 500, + 854, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "20998", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] ShahRukh Athar, Zexiang Xu, Kalyan Sunkavalli, Eli Shechtman, and Zhixin Shu. Rignerf: Fully controllable neural 3d portraits. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20364-20373, 2022. 3", + "[2] Hendrik Baatz, Jonathan Granskog, Marios Papas, Fabrice Rousselle, and Jan Novák. Nerf-tex: Neural reflectance field textures. In Computer Graphics Forum. Wiley Online Library, 2021. 3", + "[3] Alexander W. Bergman, Petr Kellnhofer, Wang Yifan, Eric R. Chan, David B. Lindell, and Gordon Wetzstein. Generative neural articulated radiance fields. In NeurIPS, 2022. 2, 5, 7", + "[4] Volker Blanz and Thomas Vetter. A morphable model for the synthesis of 3d faces. In Proceedings of the 26th Annual Conference on Computer Graphics and Interactive Techniques, SIGGRAPH '99, page 187-194, USA, 1999. ACM Press/Addison-Wesley Publishing Co. 2, 3", + "[5] Andrei Burov, Matthias Nießner, and Justus Thies. Dynamic surface function networks for clothed human bodies. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10754-10764, 2021. 3", + "[6] Shengqu Cai, Anton Obukhov, Dengxin Dai, and Luc Van Gool. Pix2nerf: Unsupervised conditional p-gan for single image to neural radiance fields translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3981-3990, 2022. 2", + "[7] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In Proceedings of the IEEE International Conference on Computer Vision (CVPR), 2021. 2, 4, 5, 7", + "[8] Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2", + "[9] Xu Chen, Tianjian Jiang, Jie Song, Jinlong Yang, Michael J Black, Andreas Geiger, and Otmar Hilliges. gdna: Towards generative detailed neural avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20427-20437, 2022. 2", + "[10] Yuedong Chen, Qianyi Wu, Chuanxia Zheng, Tat-Jen Cham, and Jianfei Cai. Sem2nerf: Converting single-view semantic masks to neural radiance fields. In ECCV, 2022. 2", + "[11] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5939-5948, 2019. 3", + "[12] Hang Dai, Nick Pears, William Smith, and Christian Duncan. Statistical modeling of craniofacial shape and texture. International Journal of Computer Vision, 128(2):547-571, 2020. 3" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Yu Deng, Jiaolong Yang, Dong Chen, Fang Wen, and Tong Xin. Disentangled and controllable face image generation via 3d imitative-contrastive learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2, 3, 4, 5, 6", + "[14] Yu Deng, Jiaolong Yang, Jianfeng Xiang, and Xin Tong. Gram: Generative radiance manifolds for 3d-aware image generation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2", + "[15] Yu Deng, Jiaolong Yang, Sicheng Xu, Dong Chen, Yunde Jia, and Xin Tong. Accurate 3d face reconstruction with weakly-supervised learning: From single image to image set. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 0–0, 2019. 5, 7", + "[16] Michail Christos Doukas, Stefanos Zafeiriou, and Viktoriia Sharmanska. Headgan: One-shot neural head synthesis and editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14398-14407, 2021. 2, 3", + "[17] Yao Feng, Haiwen Feng, Michael J Black, and Timo Bolkart. Learning an animatable detailed 3d face model from inthe-wild images. ACM Transactions on Graphics (ToG), 40(4):1-13, 2021. 5", + "[18] Ohad Fried, Ayush Tewari, Michael Zollhöfer, Adam Finkelstein, Eli Shechtman, Dan B Goldman, Kyle Genova, Zeyu Jin, Christian Theobalt, and Maneesh Agrawala. Text-based editing of talking-head video. ACM Transactions on Graphics (TOG), 38(4):1-14, 2019. 3", + "[19] Matheus Gadelha, Subhransu Maji, and Rui Wang. 3d shape induction from 2d views of multiple objects. In International Conference on 3D Vision (3DV), 2017. 2", + "[20] Guy Gafni, Justus Thies, Michael Zollhofer, and Matthias Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8649-8658, 2021. 3", + "[21] Rinon Gal, Or Patashnik, Haggai Maron, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Stylegan-nada: Clip-guided domain adaptation of image generators. ACM Transactions on Graphics (TOG), 41(4):1-13, 2022. 8", + "[22] Baris Gecer, Binod Bhattacharai, Josef Kittler, and Tae-Kyun Kim. Semi-supervised adversarial learning to generate photorealistic face images of new identities from 3d morphable model. In Proceedings of the European conference on computer vision (ECCV), pages 217-234, 2018. 3", + "[23] Baris Gecer, Stylianos Ploumpis, Irene Kotsia, and Stefanos Zafeiriou. Ganfit: Generative adversarial network fitting for high fidelity 3d face reconstruction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1155-1164, 2019. 3", + "[24] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 2", + "[25] Philip-William Grassal, Malte Prinzler, Titus Leistner, Carsten Rother, Matthias Nießner, and Justus Thies. Neural" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "20999", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "head avatars from monocular rgb videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18653-18664, 2022. 3", + "[26] Artur Grigorev, Karim Iskakov, Anastasia Ianina, Renat Bashirov, Ilya Zakharkin, Alexander Vakhitov, and Victor Lempitsky. Stylepeople: A generative model of fullbody human avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5151-5160, 2021. 3", + "[27] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. Stylenerf: A style-based 3d aware generator for high-resolution image synthesis. In International Conference on Learning Representations, 2022. 2, 5", + "[28] Philipp Henzler, Niloy J Mitra, and Tobias Ritschel. Escaping plato's cave: 3d shape from adversarial rendering. In Proceedings of the IEEE International Conference on Computer Vision (CVPR), 2019. 2", + "[29] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In Advances in Neural Information Processing Systems (NeurIPS), 2017. 7", + "[30] Fangzhou Hong, Zhaoxi Chen, Yushi LAN, Liang Pan, and Ziwei Liu. EVA3d: Compositional 3d human generation from 2d image collections. In International Conference on Learning Representations, 2023. 2", + "[31] Yang Hong, Bo Peng, Haiyao Xiao, Ligang Liu, and Juyong Zhang. Headnerf: A real-time nerf-based parametric head model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20374-20384, 2022. 3", + "[32] Angjoo Kanazawa, Shubham Tulsiani, Alexei A Efros, and Jitendra Malik. Learning category-specific mesh reconstruction from image collections. In Proceedings of the European Conference on Computer Vision (ECCV), pages 371-386, 2018. 3", + "[33] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4401-4410, 2019. 5", + "[34] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2020. 4", + "[35] Hyeongwoo Kim, Pablo Garrido, Ayush Tewari, Weipeng Xu, Justus Thies, Matthias Niessner, Patrick Pérez, Christian Richardt, Michael Zollhöfer, and Christian Theobalt. Deep video portraits. ACM Transactions on Graphics (TOG), 37(4):1-14, 2018. 2, 3, 5", + "[36] Jeong-gi Kwak, Yuanming Li, Dongsik Yoon, Donghyeon Kim, David Han, and Hanseok Ko. Injecting 3d perception of controllable nerf-gan into stylegan for editable portrait image synthesis. In Proceedings of European Conference on Computer Vision (ECCV), 2022. 2", + "[37] Henning Lange and J Nathan Kutz. Fc2t2: The fast continuous convolutional taylor transform with applications in vi" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "sion and graphics. arXiv preprint arXiv:2111.00110, 2021. 3", + "[38] Tianye Li, Timo Bolkart, Michael J Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4d scans. ACM Trans. Graph., 36(6):194-1, 2017. 3, 4", + "[39] Connor Z Lin, David B Lindell, Eric R Chan, and Gordon Wetzstein. 3d gan inversion for controllable portrait image animation. arXiv preprint arXiv:2203.13441, 2022. 2, 7", + "[40] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European Conference on Computer Vision (ECCV), 2020. 3", + "[41] Thu Nguyen-Phuoc, Chuan Li, Lucas Theis, Christian Richardt, and Yong-Liang Yang. Hologan: Unsupervised learning of 3d representations from natural images. In Proceedings of the IEEE International Conference on Computer Vision, 2019. 2", + "[42] Thu H Nguyen-Phuoc, Christian Richardt, Long Mai, Yongliang Yang, and Niloy Mitra. Blockgan: Learning 3d object-aware scene representations from unlabelled images. In Advances in Neural Information Processing Systems (NeurIPS), 2020. 2", + "[43] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2", + "[44] Atsuhiro Noguchi, Xiao Sun, Stephen Lin, and Tatsuya Harada. Unsupervised learning of efficient geometry-aware neural articulated representations. In Proceedings of European Conference on Computer Vision (ECCV), 2022. 2", + "[45] Michael Oechsle, Lars Mescheder, Michael Niemeyer, Thilo Strauss, and Andreas Geiger. Texture fields: Learning texture representations in function space. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4531-4540, 2019. 3", + "[46] Roy Or-El, Xuan Luo, Mengyi Shan, Eli Shechtman, Jeong Joon Park, and Ira Kemelmacher-Shlizerman. StyleSDF: High-Resolution 3D-Consistent Image and Geometry Generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 5", + "[47] Xingang Pan, Xudong Xu, Chen Change Loy, Christian Theobalt, and Bo Dai. A shading-guided generative implicit model for shape-accurate 3d-aware image synthesis. Advances in Neural Information Processing Systems, 34:20002-20013, 2021. 2", + "[48] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3", + "[49] Songyou Peng, Michael Niemeyer, Lars Mescheder, Marc Pollefeys, and Andreas Geiger. Convolutional occupancy networks. In European Conference on Computer Vision, pages 523-540. Springer, 2020. 3" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "21000", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[50] Hanspeter Pfister, Matthias Zwicker, Jeroen Van Baar, and Markus Gross. Surfels: Surface elements as rendering primitives. In Proceedings of the 27th annual conference on Computer graphics and interactive techniques, pages 335-342, 2000. 3", + "[51] Stylianos Ploumpis, Evangelos Ververas, Eimear O'Sullivan, Stylianos Moschoglou, Haoyang Wang, Nick Pears, William AP Smith, Baris Gecer, and Stefanos Zafeiriou. Towards a complete 3d morphable model of the human head. IEEE transactions on pattern analysis and machine intelligence, 43(11):4142-4160, 2020. 3", + "[52] Yurui Ren, Ge Li, Yuanqi Chen, Thomas H Li, and Shan Liu. Pirenderer: Controllable portrait image generation via semantic neural rendering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13759-13768, 2021. 2, 3", + "[53] Daniel Roich, Ron Mokady, Amit H Bermano, and Daniel Cohen-Or. Pivotal tuning for latent-based editing of real images. ACM Trans. Graph., 2021. 8", + "[54] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. Graf: Generative radiance fields for 3d-aware image synthesis. In Advances in Neural Information Processing Systems (NeurIPS), 2020. 2", + "[55] Katja Schwarz, Axel Sauer, Michael Niemeyer, Yiyi Liao, and Andreas Geiger. Voxgraf: Fast 3d-aware image synthesis with sparse voxel grids. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 2", + "[56] Vincent Sitzmann, Justus Thies, Felix Heide, Matthias Nießner, Gordon Wetzstein, and Michael Zollhofer. Deepvoxels: Learning persistent 3d feature embeddings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2437-2446, 2019. 3", + "[57] Ivan Skorokhodov, Sergey Tulyakov, Yiqun Wang, and Peter Wonka. Epigraph: Rethinking training of 3d gans. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 2", + "[58] Jingxiang Sun, Xuan Wang, Yichun Shi, Lizhen Wang, Jue Wang, and Yebin Liu. Ide-3d: Interactive disentangled editing for high-resolution 3d-aware portrait synthesis. ACM Transactions on Graphics (TOG), 41(6):1-10, 2022. 2, 5, 8", + "[59] Jingxiang Sun, Xuan Wang, Yong Zhang, Xiaoyu Li, Qi Zhang, Yebin Liu, and Jue Wang. Fenerf: Face editing in neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2", + "[60] Keqiang Sun, Shangzhe Wu, Zhaoyang Huang, Ning Zhang, Quan Wang, and HongSheng Li. Controllable 3d face synthesis with conditional generative occupancy fields. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 2", + "[61] Junshu Tang, Bo Zhang, Binxin Yang, Ting Zhang, Dong Chen, Lizhuang Ma, and Fang Wen. Explicitly controllable 3d-aware portrait generation. arXiv preprint arXiv:2209.05434, 2022. 2, 4, 5, 6", + "[62] Ayush Tewari, Mohamed Elgharib, Gaurav Bharaj, Florian Bernard, Hans-Peter Seidel, Patrick Pérez, Michael Zollhofer, and Christian Theobalt. Stylerig: Rigging stylegan" + ], + "bbox": [ + 78, + 90, + 468, + 901 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "for 3d control over portrait images. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2, 3", + "[63] Ayush Tewari, Justus Thies, Ben Mildenhall, Pratul Srinivasan, Edgar Tretschk, W Yifan, Christoph Lassner, Vincent Sitzmann, Ricardo Martin-Brualla, Stephen Lombardi, et al. Advances in neural rendering. Computer Graphics Forum, 41(2):703-735, 2022. 3", + "[64] Justus Thies, Michael Zollhöfer, and Matthias Nießner. Deferred neural rendering: Image synthesis using neural textures. ACM Transactions on Graphics (TOG), 38(4):1-12, 2019. 3, 4", + "[65] Justus Thies, Michael Zollhofer, Marc Stamminger, Christian Theobalt, and Matthias Nießner. Face2face: Real-time face capture and reenactment of rgb videos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2387-2395, 2016. 2, 3", + "[66] Lizhen Wang, Zhiyuan Chen, Tao Yu, Chenguang Ma, Liang Li, and Yebin Liu. Faceverse: a fine-grained and detail-controllable 3d face morphable model from a hybrid dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20333-20342, 2022. 4", + "[67] Jiajun Wu, Chengkai Zhang, Tianfan Xue, Bill Freeman, and Josh Tenenbaum. Learning a probabilistic latent space of object shapes via 3d generative-adversarial modeling. Advances in neural information processing systems, 29, 2016. 2", + "[68] Yue Wu, Yu Deng, Jiaolong Yang, Fangyun Wei, Chen Qifeng, and Xin Tong. Anifacegan: Animatable 3d-aware face image generation for video avatars. In Advances in Neural Information Processing Systems, 2022. 2, 4, 5, 6", + "[69] Qiangeng Xu, Weiyue Wang, Duygu Ceylan, Radomir Mech, and Ulrich Neumann. Disn: Deep implicit surface network for high-quality single-view 3d reconstruction. Advances in Neural Information Processing Systems, 32, 2019. 3", + "[70] Sicheng Xu, Jiaolong Yang, Dong Chen, Fang Wen, Yu Deng, Yunde Jia, and Xin Tong. Deep 3d portrait from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7710-7720, 2020. 2, 3", + "[71] Yinghao Xu, Sida Peng, Ceyuan Yang, Yujun Shen, and Bolei Zhou. 3d-aware image synthesis via learning structural and textural representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18430-18439, 2022. 2", + "[72] Yang Xue, Yuheng Li, Krishna Kumar Singh, and Yong Jae Lee. Giraffe hd: A high-resolution 3d-aware generative model. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2", + "[73] Fei Yin, Yong Zhang, Xiaodong Cun, Mingdeng Cao, Yanbo Fan, Xuan Wang, Qingyan Bai, Baoyuan Wu, Jue Wang, and Yujiu Yang. Styleheat: One-shot high-resolution editable talking face generation via pretrained stylegan. In Proceedings of the European conference on computer vision (ECCV), 2022. 2" + ], + "bbox": [ + 503, + 92, + 890, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "21001", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[74] Jianfeng Zhang, Zihang Jiang, Dingdong Yang, Hongyi Xu, Yichun Shi, Guoxian Song, Zhongcong Xu, Xinchao Wang, and Jiashi Feng. Avatargen: a 3d generative model for animatable human avatars. arXiv preprint arXiv:2208.00561, 2022. 2", + "[75] Jingbo Zhang, Xiaoyu Li, Ziyu Wan, Can Wang, and Jing Liao. Fdnerf: Few-shot dynamic neural radiance fields for face reconstruction and expression editing. arXiv preprint arXiv:2208.05751, 2022.3", + "[76] Jichao Zhang, E. Sangineto, Hao Tang, Aliaksandr Siarohin, Zhun Zhong, N. Sebe, and Wei Wang. 3d-aware semantic-guided generative model for human synthesis. In ECCV, 2022. 2", + "[77] Jichao Zhang, Aliaksandr Siarohin, Yahui Liu, Hao Tang, Nicu Sebe, and Wei Wang. Training and tuning generative neural radiance fields for attribute-conditional 3d-aware face generation. arXiv preprint arXiv:2208.12550, 2022. 2", + "[78] Xuanmeng Zhang, Zhedong Zheng, Daiheng Gao, Bang Zhang, Pan Pan, and Yi Yang. Multi-view consistent generative adversarial networks for 3d-aware image synthesis. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2", + "[79] Xiaoming Zhao, Fangchang Ma, David Guera, Zhile Ren, Alexander G. Schwing, and Alex Colburn. Generative multiplane images: Making a 2d gan 3d-aware. In Proceedings of European Conference on Computer Vision (ECCV), 2022. 2", + "[80] Yufeng Zheng, Victoria Fernández Abrevaya, Marcel C Bühler, Xu Chen, Michael J Black, and Otmar Hilliges. Im avatar: Implicit morphable head avatars from videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13545-13555, 2022. 3", + "[81] Peng Zhou, Lingxi Xie, Bingbing Ni, and Qi Tian. Cips-3d: A 3d-aware generator of gans based on conditionally-independent pixel synthesis. arXiv:2110.09788, 2021. 2", + "[82] Jun-Yan Zhu, Zhoutong Zhang, Chengkai Zhang, Jiajun Wu, Antonio Torralba, Josh Tenenbaum, and Bill Freeman. Visual object networks: Image generation with disentangled 3d representations. In Advances in Neural Information Processing Systems (NeurIPS), 2018. 2", + "[83] Peiye Zhuang, Liqian Ma, Oluwasanmi Koyejo, and Alexander Schwing. Controllable radiance fields for dynamic face synthesis. In Proc. 3DV, 2022. 2" + ], + "bbox": [ + 78, + 90, + 470, + 698 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "21002", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2023/Next3D_ Generative Neural Texture Rasterization for 3D-Aware Head Avatars/7b615aeb-a836-44b1-beba-de01bf0423b2_model.json b/2023/Next3D_ Generative Neural Texture Rasterization for 3D-Aware Head Avatars/7b615aeb-a836-44b1-beba-de01bf0423b2_model.json new file mode 100644 index 0000000000000000000000000000000000000000..1e0a284b91a1b9c590b2b7823927adb647a26086 --- /dev/null +++ b/2023/Next3D_ Generative Neural Texture Rasterization for 3D-Aware Head Avatars/7b615aeb-a836-44b1-beba-de01bf0423b2_model.json @@ -0,0 +1,2677 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.131, + 0.88, + 0.152 + ], + "angle": 0, + "content": "Next3D: Generative Neural Texture Rasterization for 3D-Aware Head Avatars" + }, + { + "type": "text", + "bbox": [ + 0.126, + 0.179, + 0.251, + 0.201 + ], + "angle": 0, + "content": "Jingxiang Sun" + }, + { + "type": "text", + "bbox": [ + 0.29, + 0.18, + 0.395, + 0.2 + ], + "angle": 0, + "content": "Xuan Wang\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.434, + 0.181, + 0.559, + 0.2 + ], + "angle": 0, + "content": "Lizhen Wang\\(^{14}\\)" + }, + { + "type": "text", + "bbox": [ + 0.599, + 0.181, + 0.691, + 0.2 + ], + "angle": 0, + "content": "Xiaoyu Li3" + }, + { + "type": "text", + "bbox": [ + 0.73, + 0.181, + 0.837, + 0.2 + ], + "angle": 0, + "content": "Yong Zhang3" + }, + { + "type": "text", + "bbox": [ + 0.348, + 0.202, + 0.488, + 0.22 + ], + "angle": 0, + "content": "Hongwen Zhang" + }, + { + "type": "text", + "bbox": [ + 0.53, + 0.202, + 0.617, + 0.219 + ], + "angle": 0, + "content": "Yebin Liu" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.224, + 0.323, + 0.244 + ], + "angle": 0, + "content": "\\(^{1}\\)Tsinghua University" + }, + { + "type": "text", + "bbox": [ + 0.362, + 0.225, + 0.458, + 0.244 + ], + "angle": 0, + "content": "\\(^{2}\\)Ant Group" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.225, + 0.632, + 0.242 + ], + "angle": 0, + "content": "\\(^{3}\\)Tencent AI Lab" + }, + { + "type": "text", + "bbox": [ + 0.673, + 0.226, + 0.775, + 0.242 + ], + "angle": 0, + "content": "4NNKosmos" + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.277, + 0.895, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.239, + 0.496, + 0.343, + 0.508 + ], + "angle": 0, + "content": "Gaze Animation" + }, + { + "type": "image_caption", + "bbox": [ + 0.591, + 0.496, + 0.783, + 0.51 + ], + "angle": 0, + "content": "3D-Aware Domain Adaptation" + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.527, + 0.895, + 0.584 + ], + "angle": 0, + "content": "Figure 1. Our 3D GAN synthesizes generative, high-quality, and 3D-consistent facial avatars from unstructured 2D images. Unlike current animatable 3D GANs that only modify yaw-pitch head poses and facial expressions, our approach enables fine-grained control over full-head rotations, facial expressions, eye blinks, and gaze directions with strict 3D consistency and a high level of photorealism. Our approach also provides strong 3D priors for downstream tasks such as 3D-aware stylization." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.596, + 0.314, + 0.612 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.629, + 0.473, + 0.903 + ], + "angle": 0, + "content": "3D-aware generative adversarial networks (GANs) synthesize high-fidelity and multi-view-consistent facial images using only collections of single-view 2D imagery. Towards fine-grained control over facial attributes, recent efforts incorporate 3D Morphable Face Model (3DMM) to describe deformation in generative radiance fields either explicitly or implicitly. Explicit methods provide fine-grained expression control but cannot handle topological changes caused by hair and accessories, while implicit ones can model varied topologies but have limited generalization caused by the unconstrained deformation fields. We propose a novel 3D GAN framework for unsupervised learning of generative, high-quality and 3D-consistent facial avatars from unstructured 2D images. To achieve both deformation accuracy and topological flexibility, we propose a 3D representation called Generative Texture-Rasterized Tri-planes. The proposed representation learns Generative Neural Textures on top of parametric mesh templates" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.598, + 0.895, + 0.81 + ], + "angle": 0, + "content": "and then projects them into three orthogonal-viewed feature planes through rasterization, forming a tri-plane feature representation for volume rendering. In this way, we combine both fine-grained expression control of mesh-guided explicit deformation and the flexibility of implicit volumetric representation. We further propose specific modules for modeling mouth interior which is not taken into account by 3DMM. Our method demonstrates state-of-the-art 3D-aware synthesis quality and animation ability through extensive experiments. Furthermore, serving as 3D prior, our animatable 3D representation boosts multiple applications including one-shot facial avatars and 3D-aware stylization. Project page: https://mrtornado24.github.io/Next3D/. Code: https://github.com/MrTornado24/Next3D." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.844, + 0.633, + 0.86 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.893, + 0.903 + ], + "angle": 0, + "content": "Animatable portrait synthesis is essential for movie postproduction, visual effects, augmented reality (AR), and vir" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.958 + ], + "angle": 0, + "content": "20991" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.213 + ], + "angle": 0, + "content": "tual reality (VR) telepresence applications. Efficient animatable portrait generators should be capable of synthesizing diverse high-fidelity portraits with full control of the rigid head pose, facial expressions and gaze directions at a fine-grained level. The main challenges of this task lie in how to model accurate deformation and preserve identity through animation in the generative setting, i.e. training with only unstructured corpus of 2D images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.215, + 0.473, + 0.548 + ], + "angle": 0, + "content": "Several 2D generative models perform image animation by incorporating the 3D Morphable Face Models (3DMM) [4] into the portrait synthesis [13, 16, 35, 52, 62, 65, 70, 73]. These 2D-based methods achieve photorealism but suffer from shape distortion during large motion due to a lack of geometry constraints. Towards better view consistency, many recent efforts incorporate 3DMM with 3D GANs, learning to synthesize animatable and 3D consistent portraits from only 2D image collections in an unsupervised manner [3, 30, 39, 44, 60, 61, 68, 74]. Bergman et al. [3] propose an explicit surface-driven deformation field for warping radiance fields. While modeling accurate facial deformation, it cannot handle topological changes caused by non-facial components, e.g. hair, glasses, and other accessories. AnifaceGAN [68] builds an implicit 3DMM-conditioned deformation field and constrains animation accuracy by imitation learning. It achieves smooth animation on interpolated expressions, however, struggles to generate reasonable extrapolation due to the under-constrained deformation field. Therefore, The key challenge of this task is modeling deformation in the 3D generative setting for animation accuracy and topological flexibility." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.551, + 0.473, + 0.868 + ], + "angle": 0, + "content": "In this paper, we propose a novel 3D GAN framework for unsupervised learning of generative, high-quality, and 3D-consistent facial avatars from unstructured 2D images. Our model splits the whole head into dynamic and static parts, and models them respectively. For dynamic parts, the key insight is to combine both fine-grained expression control of mesh-guided explicit deformation and flexibility of implicit volumetric representation. To this end, we propose a novel representation, Generative Texture-Rasterized Triplanes, which learns the facial deformation through Generative Neural Textures on top of a parametric template mesh and samples them into three orthogonal-viewed and axis-aligned feature planes through standard rasterization, forming a tri-plane feature representation. Such texture-rasterized tri-planes re-form high-dimensional dynamic surface features in a volumetric representation for efficient volume rendering and thus inherit both the accurate control of the mesh-driven deformation and the expressiveness of volumetric representations. Furthermore, we represent static components (body, hair, background, etc.) by another triplane branch, and integrate both through alpha blending." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.472, + 0.902 + ], + "angle": 0, + "content": "Another key insight of our method is to model the mouth interior which is not taken into account by 3DMM. Mouth" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.213 + ], + "angle": 0, + "content": "interior is crucial for animation quality but often ignored by prior arts. We propose an efficient teeth synthesis module, formed as a style-modulated UNet, to complete the inner mouth features missed by the template mesh. To further regularize the deformation accuracy, we introduce a deformation-aware discriminator which takes as input synthetic renderings, encouraging the alignment of the final outputs with the 2D projection of the expected deformation." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.213, + 0.871, + 0.228 + ], + "angle": 0, + "content": "To summarize, the contributions of our approach are:" + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.234, + 0.892, + 0.295 + ], + "angle": 0, + "content": "- We present an animatable 3D-aware GAN framework for photorealistic portrait synthesis with fine-grained animation, including expressions, eye blinks, gaze direction and full head poses." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.302, + 0.892, + 0.378 + ], + "angle": 0, + "content": "- We propose Generative Texture-Rasterized Triplanes, an efficient deformable 3D representation that inherits both fine-grained expression control of mesh-guided explicit deformation and flexibility of implicit volumetric representation." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.385, + 0.892, + 0.461 + ], + "angle": 0, + "content": "- Our learned generative animatable 3D representation can serve as a strong 3D prior and boost the downstream application of 3D-aware one-shot facial avatars. Our model also pushes the frontier of 3D stylization with high-quality out-of-domain facial avatars." + }, + { + "type": "list", + "bbox": [ + 0.519, + 0.234, + 0.892, + 0.461 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.468, + 0.642, + 0.484 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.493, + 0.892, + 0.719 + ], + "angle": 0, + "content": "Generative 3D-aware Image Synthesis. Generative adversarial networks [24] have achieved photorealistic synthesis in 2D domain. Building on the success of 2D GANs, many efforts have lifted the image synthesis into 3D with explicit view control. Early voxel-based approaches [19, 28, 41, 42, 67, 82] adopt the 3D CNN generators whose heavy computational burden limits the high-resolution image synthesis. Recent works incorporate more efficient neural scene representations, such as fully implicit networks [6, 8-10, 14, 47, 54, 57, 81], sparse voxel grids [55], multiple planes [79] or a combination of low-resolution feature volume and 2D super-resolution [7, 27, 43, 46, 71, 72, 76, 78]. We leverage the tri-plane representation proposed in [7] and endow it with animation ability by the orthogonal-rasterized Generative Neural Textures." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Some other current works [3,30,36,44,58,59,61,68,77, 83] focus on the editability of 3D-aware generative models. FENeRF [59] and IDE-3D [58] perform semantic-guided 3D face editing by incorporating semantic-aware radiance fields and GAN inversion. However, they cannot produce continuous and stable editing on videos. Other methods [3, 30,44,61,68] employ 3D priors to achieve animatable image synthesis and the main differences lie on the deformation strategies including linear blend skinning [30,44], surface-driven deformation [3], 3DMM-guided latent decomposition [61] and neural deformation fields [68]. These approaches either don't allow for topology changes or need" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20992" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.088, + 0.895, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.293, + 0.893, + 0.364 + ], + "angle": 0, + "content": "Figure 2. Our 3D GAN framework consists of two tri-plane branches \\( T_{uv} \\) and \\( T_{static} \\) modeling dynamic and static components. \\( T_{uv} \\) is formed by the orthogonal rasterized Generative Neural Textures which are synthesized by a StyleGAN generator, \\( G_{texture} \\), on top of deformable template mesh. \\( T_{static} \\) is synthesized by another StyleGAN generator, \\( G_{static} \\). The mouth synthesis module, \\( G_{teeth} \\), is presented for completing mouth interior. Blended triplanes are incorporated with hybrid neural renderer consisting of volume rendering and a super-resolution module \\( G_{superres} \\). For discrimination, synthetic renderings \\( I_{synthetic} \\) is taken into the dual discriminator \\( D_{dual} \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.377, + 0.468, + 0.452 + ], + "angle": 0, + "content": "elaborate loss design to ensure the accuracy of deformation. On the contrary, our approach naturally achieves accurate animation with explicit mesh guidance and further allows for topology changes by adapting surface deformation into a continuous volumetric representation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.458, + 0.468, + 0.865 + ], + "angle": 0, + "content": "Facial animation with 3D morphable face models. Blanz et al. [4] model facial texture and shape as vector spaces, known as the 3D Morphable Model (3DMM). Extensions of 3DMM, such as full-head PCA models [12, 51], blend-shape models [38], are extensively studied and widely used in facial animation tasks [23, 65]. Benefiting from 3DMM, these methods can model the deformation of facial parts accurately and continuously, nevertheless, struggle to represent non-facial areas missed by 3DMM, e.g. hair, teeth, eyes, and body. Moreover, these methods are prone to lacking facial details. To fill the missing areas and complete more realistic facial details, later works [13, 16, 18, 22, 35, 52, 62, 64, 70] apply learned approaches on top of 3DMM renderings. DiscoFaceGAN [13] maps 3DMM parameters into the latent space and decouple them by imitative-contrastive learning. Though efficient and photorealistic animation is achieved, these 2D methods don't model 3D geometry and thus cannot remain strict 3D consistency with large head pose changes. For strict 3D consistency, recent efforts [1, 20, 25, 75, 80] incorporate 3DMM into volumetric representations [31, 32, 40, 49, 56] to achieve view-consistent facial animation. Furthermore, volumetric representations enable the modeling of thin structures such as hair, and also mouth interior thanks to its spatial continuity. While 3DMM have been adopted to animate radiance fields for single-scene scenarios, it is challenging to adapt to the generative setting with the absence of groundtruth supervision." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Neural scene representations. The neural scene representations can be roughly categorized into implicit and explicit" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.377, + 0.892, + 0.633 + ], + "angle": 0, + "content": "surface representations and volumetric representations [63]. The surface can be represented explicitly by point clouds [37, 50], meshes [2, 5, 45, 64], or defined implicitly as a zero level-set of a function [11, 48, 69], like signed distance function, which can be approximated by coordinate-based multi-layer perceptrons (MLPs). On the contrary, volumetric representations [32, 40, 49, 56] store volumetric properties (occupancies, radiance, colors, etc.) instead of the surface of an object. These properties can be stored in explicit voxel grids [32, 49, 56] or the weights of a neural network implicitly [40]. In this work, we propose a hybrid surface-volumetric representation. Specifically, we learn the deformable surface radiance by neural textures [26, 64] on top of the template mesh and rasterize it into three orthogonal-viewed feature planes. Then, the planes are reshaped to a tri-plane representation with decoding into neural radiance fields for volume rendering." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.66, + 0.608, + 0.677 + ], + "angle": 0, + "content": "3. Approach" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We present an animatable 3D-aware facial generator that equips with fine-grained expression and pose control, photorealistic rendering quality, and high-quality underlying geometry. The proposed method models dynamic and static components by two independent tri-plane branches. Specifically, we propose Generative Texture-rasterized Tri-planes for modeling dynamic facial parts (Sec. 3.1). Furthermore, we propose an efficient mouth synthesis module to complete the mouth interior that is not included in 3DMM (Sec. 3.2). We further adopt another tri-plane branch for the static components (Sec. 3.3). Both tri-planes are blended together for hybrid neural rendering (Sec. 3.4). We introduce an deformation-aware discriminator (Sec. 3.5) and illustrate the training objectives in Sec. 3.6." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20993" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.077, + 0.091, + 0.419, + 0.108 + ], + "angle": 0, + "content": "3.1. Generative texture-rasterized tri-planes" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.115, + 0.473, + 0.644 + ], + "angle": 0, + "content": "EG3D [7] presents an efficient tri-plane-based hybrid 3D representation to synthesize high-resolution images with multi-view consistency. Nonetheless, EG3D lacks control over facial deformations and thus cannot be directly applied to animation tasks. To this end, we leverage Neural textures [64] to represent deformable facial parts. In general, Neural Textures are a set of learned high-dimensional feature maps that can be interpreted by a neural renderer. We extend it to our generative setting and synthesize the neural textures through a StyleGAN2 CNN generator \\( G_{\\text{texture}} \\). As shown in Fig. 2, we first sample a latent code \\( z \\) and map it into an intermediate latent space by the mapping network. Our texture generator architecture closely follows StyleGAN2 backbone [34], except producing a \\( 256 \\times 256 \\times 32 \\) neural texture map, \\( T \\), instead of a three-channel RGB image. Storing a high-dimensional learned feature vector per texel, \\( T \\) can be rasterized to a view-dependent screen-space feature map given a mesh with uv-texture parameterization and a target view as input. In our case, we use the FLAME template [38] to provide a coarse mesh that can be driven by deformation parameters. Given the predesigned texture mapping function, we employ the standard graphics pipeline to rasterize the neural textures from the texture space into the screen space based on the template mesh. We choose Neural Textures as the deformation method for two reasons. First, compared with other explicit deformation (e.g. linear blend skinning and surface-driven deformation) highly dependent on the accurate underlying geometry, Neural Textures embed high-level features which compensate the imperfect geometry and thus are more suitable for our settings where template meshes are not accurate. Furthermore, unlike implicit deformation methods [13, 61, 68], our explicit mesh-guided deformation alleviates the requirement of elaborate imitation learning while gaining better expression generalization (Fig. 4)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.645, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Neural Textures encode surface deformation accurately with mesh guidance but lack generalization to 3D points far from surface. Besides, it also doesn't allow for topological changes. To this end, we propose Generative Texture-Rasterized Tri-planes, \\( T_{uv} \\) which reshapes the rasterized textures into a tri-plane representation. Therefore, we can adapt such surface deformation into a continuous volume, and once a 3D point out of the surface could be projected within the facial part on at least one view (e.g. glasses), its feature can be interpolated by the meaningful surface feature through grid sampling in the tri-plane. Specifically, we rasterize neural textures based on the template mesh into three orthogonal views and place them in three axis-aligned feature planes. In practice, considering the zygomorphy, the rasterization is applied at both the left and right views and the rasterized features are concatenated for one single plane by summation. In this way, Our hybrid surface-volumetric" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "representation inherits the best of both worlds: accurate mesh-guided facial deformation of Neural Textures and the continuity and topological flexibility of volumetric representations. See Fig. 5 for the topological-aware animation of portraits with glasses." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.18, + 0.724, + 0.196 + ], + "angle": 0, + "content": "3.2. Mouth synthesis module" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.217, + 0.892, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.393, + 0.892, + 0.449 + ], + "angle": 0, + "content": "Figure 3. The teeth synthesis module consists of a teeth completing module and a neural blending module. The teeth competing module produces mouth interior conditioned on multi-scale teeth exterior features and latent codes \\( w \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.456, + 0.893, + 0.713 + ], + "angle": 0, + "content": "Since the FLAME template doesn't contain inner mouth, we propose a teeth synthesis module, \\( G_{\\text{teeth}} \\), to complete the missing teeth features in \\( T_{uv} \\). As shown in Fig. 3, for each feature plane of \\( T_{uv} \\), we crop the teeth area by the expanded landmarks and resize it into \\( 64 \\times 64 \\). Then, the stacked mouth features are processed by \\( G_{\\text{teeth}} \\) which employs a style-modulated UNet [66]. The downsampling process of \\( G_{\\text{teeth}} \\) encodes \\( f_{\\text{teeth}} \\) into multi-scale feature maps which serve as content conditions for the following StyleGAN layers. The output teeth features \\( f_{\\text{teeth}}' \\) are transformed inversely and concatenated with the feature planes of \\( T_{uv} \\). To eliminate the texture flickering of mouth boundary, we further feed \\( T_{uv} \\) into a shallow-UNet-based neural blending module and obtain \\( T_{uv}' \\). We conduct a series of ablation studies and prove that the proposed teeth synthesis module brings a remarkable improvement on both the animation accuracy and synthesis quality (Sec. 4.2)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.726, + 0.753, + 0.743 + ], + "angle": 0, + "content": "3.3. Modeling static components" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The generative texture-rasterized tri-planes manage to model dynamic faces varying expressions and shapes, though, it is challenging to synthesize static parts like diverse haircut, background and upper body which are not included in the FLAME template. To this end, we model these parts by another tri-plane branch, \\( T_{\\text{static}} \\), which is generated by a StyleGAN2 CNN generator \\( G_{\\text{static}} \\) sharing the same latent code with \\( G_{\\text{texture}} \\). The plane features of \\( T_{uv}^{\\prime} \\) and \\( T_{\\text{static}} \\) are blended on each plane by the alpha masks rendered by rasterization. Such design not only benefits the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20994" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.122 + ], + "angle": 0, + "content": "modeling of various-styled static components but also enforces their consistency during facial animations." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.131, + 0.248, + 0.148 + ], + "angle": 0, + "content": "3.4. Neural rendering" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.154, + 0.473, + 0.381 + ], + "angle": 0, + "content": "Given the blended tri-planes, for any point in the 3D space, we project it into each plane and sample the features bi-linearly. Then, the sampled features are aggregated by summation and decoded into volume density \\(\\sigma\\) and feature \\(f\\) by a lightweight decoder. Similar to [7,58], the decoder is a single hidden layered multi-layer perceptron (MLP) with softplus activation. The volume rendering is employed to accumulate \\(\\sigma\\) and \\(f\\) along the rays cast through each pixel to compute a 2D feature image \\(I_{f}\\). Similar to [7,27,46], we leverage a 2D super-resolution module \\(G_{superres}\\) to interpret the feature image into RGB image \\(I_{RGB}\\) with higher resolution. The super-resolution module consists of three StyleGAN2 synthesis blocks and the noise input is removed for alleviating texture flickers. In our case, \\(I_{f}\\) and \\(I_{RGB}\\) are set to \\(64\\times 64\\) and \\(512\\times 512\\), respectively." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.39, + 0.373, + 0.405 + ], + "angle": 0, + "content": "3.5. Deformation-aware discriminator" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.413, + 0.473, + 0.716 + ], + "angle": 0, + "content": "To learn the unsupervised 3D representations, we adopt a 2D convolutional discriminator \\( D \\) to critique the renderings. Inspired by [7], we regularize the first three channels of \\( I_{f} \\) as low-resolution RGB image, which is concatenated with \\( I_{RGB} \\) as the input for the discriminator. However, the discrimination with only image input can only ensure that the deformed images are always in a correct distribution instead of matching the expected deformations. Therefore, we make the discriminator aware of the expression and shape under which the generated image are deformed. For the same purpose, GNARF [3] conditions the discriminator on the FLAME parameters by concatenating them as the input to the mapping network. However, we find empirically that such a conditioning method leads to training instability, consistent with [3]. Instead, we re-render the template mesh under the rendered pose to get the synthetic rendering \\( I_{synthetic} \\) and feed it into \\( D_{\\text{dual}} \\) along with image pairs. Here, we adopt correspondence images inspired by [35]. Such concatenation encourages the final output to align with the synthetic rendering and learn the expected deformation." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.724, + 0.262, + 0.74 + ], + "angle": 0, + "content": "3.6. Training objectives" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.747, + 0.47, + 0.807 + ], + "angle": 0, + "content": "During training, we use the non-saturating GAN loss with R1 regularization. Moreover, we adopt the density regularization proposed in EG3D [7]. Therefore, the total learning objective is:" + }, + { + "type": "equation", + "bbox": [ + 0.131, + 0.816, + 0.47, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {D _ {\\text {d u a l}, G}} = \\mathbb {E} _ {z \\sim p _ {z}, \\epsilon \\sim p _ {\\epsilon}} [ f (D _ {\\text {d u a l}} (G (z, \\epsilon))) ] + \\\\ \\mathbb {E} _ {I ^ {r} \\sim p _ {I ^ {r}}} [ f (- D _ {\\text {d u a l}} (I ^ {r}) + (1) \\\\ \\left. \\lambda \\| \\bigtriangledown D _ {d u a l} \\left(I ^ {r}\\right) \\| ^ {2} \\right], \\\\ \\mathcal {L} _ {\\text {d e n s i t y}} = \\sum_ {x _ {s} \\in \\mathcal {S}} \\| d (x _ {s}) - d (x _ {s} + \\epsilon) \\| _ {2}, (2) \\\\ \\end{array}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.576, + 0.105, + 0.892, + 0.12 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {D _ {\\text {d u a l}}, G} + \\lambda_ {\\text {d e n s i t y}} \\mathcal {L} _ {\\text {d e n s i t y}}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.13, + 0.894, + 0.267 + ], + "angle": 0, + "content": "where \\(I^r\\) is the combination of real images, blurred real images, and the corresponding synthetic renderings, which are sampled from the training set with distribution \\(p_I\\). We adopt many training hyperparameters from EG3D and StyleGAN2 (learning rates of generator and discriminator, batch size, R1 regularization, etc.). We train our model based on the pretrained model of EG3D [7] and continue to train on 4 3090 GPUs for roughly 4 days. Please refer to the supplemental material for the implementation details." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.279, + 0.634, + 0.296 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.304, + 0.892, + 0.41 + ], + "angle": 0, + "content": "In this section, we first show qualitative and quantitative comparisons to state-of-the-art 2D / 3D animatable generative facial models (Sec. 4.1), and then discuss the conducted ablation studies of our design choices (Sec. 4.2). Furthermore, we show various applications combining our efficient 3D representation with GAN technologies such as GAN inversion and style transfer (Sec. 4.3)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.41, + 0.893, + 0.621 + ], + "angle": 0, + "content": "Datasets. We train and test our methods on FFHQ [33]. We augment FFHQ with horizontal flips and use an off-the-shelf pose estimator [15] to label images with the approximated camera extrinsic parameters and constant intrinsics. To support full pose animation, in-plane (roll) rotation is also considered. Furthermore, we use DECA [17] to estimate the FLAME parameters of facial identity \\(\\beta \\in \\mathbb{R}^{100}\\), jaw pose \\(\\theta_{jaw} \\in \\mathbb{R}^3\\) and expression \\(\\psi \\in \\mathbb{R}^{50}\\). Since DECA doesn't account for eyeball movement, we additionally adopt an efficient facial detector to detect 2D landmarks of irises and optimize eye poses \\(\\theta_{eye} \\in \\mathbb{R}^6\\) by minimizing the re-projection errors. Based on these FLAME parameters, we produce a template mesh with 5023 vertices and 9976 faces to drive facial deformations." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.63, + 0.64, + 0.645 + ], + "angle": 0, + "content": "4.1. Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.653, + 0.892, + 0.879 + ], + "angle": 0, + "content": "Baselines. We compare our method against two state-of-the-art methods for animatable 3D-aware image synthesis: 3DFaceShop [61], and AniFaceGAN [68]. Besides, we also select DiscoFaceGAN [13] as a baseline, which generates animatable 2D portraits conditioned on 3DMM parameters. Qualitative comparison. Fig. 4 provides a qualitative comparison against baselines. Overall, as can be seen, our method outperforms all baselines by some margin on both synthesis quality and animation accuracy. Specifically, DiscoFaceGAN [13] suffers from inconsistent identity during animation. Moreover, it cannot generate reasonable mouth interior, e.g. stretched teeth. 3DFaceshop and AnifaceGAN synthesize 3D-consistent images, nevertheless, still struggle to model consistent mouth interior with the driving images. This is because their implicit deformation approaches" + }, + { + "type": "page_footnote", + "bbox": [ + 0.518, + 0.888, + 0.646, + 0.901 + ], + "angle": 0, + "content": "1 https://mediapipe.dev/" + }, + { + "type": "footer", + "bbox": [ + 0.479, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20995" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.086, + 0.111, + 0.104, + 0.152 + ], + "angle": 270, + "content": "Driving" + }, + { + "type": "image", + "bbox": [ + 0.108, + 0.089, + 0.78, + 0.178 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.8, + 0.13, + 0.868, + 0.143 + ], + "angle": 0, + "content": "Free View" + }, + { + "type": "image_caption", + "bbox": [ + 0.086, + 0.211, + 0.104, + 0.237 + ], + "angle": 270, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.108, + 0.182, + 0.892, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.087, + 0.282, + 0.104, + 0.351 + ], + "angle": 270, + "content": "AntiFaceGAN" + }, + { + "type": "image", + "bbox": [ + 0.108, + 0.273, + 0.891, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.087, + 0.376, + 0.104, + 0.441 + ], + "angle": 270, + "content": "3DFaceShop" + }, + { + "type": "image", + "bbox": [ + 0.108, + 0.365, + 0.891, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.087, + 0.462, + 0.104, + 0.54 + ], + "angle": 270, + "content": "DiscoFaceGAN" + }, + { + "type": "image", + "bbox": [ + 0.108, + 0.459, + 0.891, + 0.547 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.559, + 0.893, + 0.587 + ], + "angle": 0, + "content": "Figure 4. Comparison with the state-of-the-art animatable 3D & 2D image synthesis methods. We extract several frames from a video clip and use the interpreted face model parameters to animate random virtual avatars." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.611, + 0.174, + 0.683 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.179, + 0.611, + 0.272, + 0.683 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.277, + 0.611, + 0.371, + 0.683 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.375, + 0.611, + 0.468, + 0.683 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.687, + 0.174, + 0.759 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.179, + 0.687, + 0.272, + 0.759 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.277, + 0.687, + 0.371, + 0.759 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.375, + 0.687, + 0.468, + 0.759 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.76, + 0.174, + 0.794 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.179, + 0.76, + 0.272, + 0.794 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.277, + 0.76, + 0.371, + 0.794 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.375, + 0.76, + 0.468, + 0.794 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.805, + 0.468, + 0.847 + ], + "angle": 0, + "content": "Figure 5. High-quality dynamic shapes with topological changes. As can be seen, we model detailed dynamic shapes of eyelids and lips, while keeping glasses unchanged." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.468, + 0.901 + ], + "angle": 0, + "content": "are under-constrained, leading to overfit the expression bias (smiling with mouth half-opened) of datasets. Compared to the other methods, our approach not only synthesizes im" + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.611, + 0.894, + 0.692 + ], + "angle": 0, + "content": "
FID↓AED↓APD↓APD*↓ID↑
DiscoFaceGAN (2562) [13]17.10.420.0460.0240.73
AniFaceGAN (2562) [68]20.10.250.0410.0220.82
3DFaceShop (5122) [61]23.70.310.0450.0240.75
Ours (2562)3.70.160.0180.0130.85
Ours (5122)3.90.160.0230.0190.84
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.703, + 0.892, + 0.757 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison using FID, average expression distance (AED), average pose distance (APD), and identity consistency (ID) for FFHQ. APD* means calculating pose distance with roll fixed." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "ages with higher quality, but also preserves more detailed expressions of the driver images, including mouth interior, eye blinks and eye movements. Furthermore, we are the only method that supports in-plane head rotations. Fig. 5 provides visual examples of the synthesized high-quality geometry. Our approach can model detailed shape deformations (see zoomed eyelids and lips in Fig. 5) with topological awareness, i.e. glasses are kept unchanged." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20996" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.088, + 0.468, + 0.537 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.547, + 0.47, + 0.603 + ], + "angle": 0, + "content": "Figure 6. Ablation study on model designs. \\( T_{\\text{static}} \\) encourages better identity consistency; \\( G_{\\text{teeth}} \\) benefits realistic mouth interior; the discriminator with \\( I_{\\text{synthetic}} \\) input allows more consistent reconstruction of detailed expressions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Quantitative evaluation. Tab. 1 demonstrates quantitative results comparing our method against baselines evaluated on several metrics. We measure image quality with Frechet Inception Distance (FID) [29] between the entire FFHQ dataset and \\(50\\mathrm{k}\\) generated images using randomly sampled latent codes, camera poses, and FLAME parameters. Since AntiFaceGAN and DiscoFaceGAN synthesize images on the resolution of \\(256^2\\), we test FID of them on FFHQ256². Following [3, 39], we evaluate the faithfulness of the animation with the Average Expression Distance (AED), the Average Pose Distance (APD), and identity consistency (ID). For each method, we randomly sample 500 identities and animate each with randomly sampled 20 FLAME parameters of expressions and poses. Then, we estimate the FLAME parameters for these 10000 generated images and the average distances between the driving FLAME parameters and the reconstructed ones. For identity consistency, we randomly sample 2000 poses, 2000 sets" + }, + { + "type": "table", + "bbox": [ + 0.533, + 0.089, + 0.86, + 0.188 + ], + "angle": 0, + "content": "
FID↓AED↓APD↓ID↑
w/o Tstatic6.60.250.0260.63
w/o Gteeth(uv)7.20.370.0420.71
w/o Gteeth(s)8.40.320.0360.79
w/o Isynthetic3.80.180.0250.74
Full3.90.160.0230.84
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.199, + 0.892, + 0.269 + ], + "angle": 0, + "content": "Table 2. Ablation study on model designs. Modeling static components by \\( T_{\\text{static}} \\) improves identity consistency. The teeth synthesis \\( G_{\\text{teeth}} \\) module benefits both animation and synthesis quality significantly while adding synthetic renderings \\( I_{\\text{synthetic}} \\) into discrimination takes both a step further." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.287, + 0.892, + 0.499 + ], + "angle": 0, + "content": "of FLAME parameters, and 1000 identities. Then we randomly select two poses and two sets of FLAME parameters for each identity, generating a total of 1000 image pairs. We calculate consistency metric using a pre-trained Arcface model [15] for each image pair and report the average result. Since the other baselines don't support in-plane (roll) head rotation, we further report APD* which only accounts for poses on yaw and pitch. Our method achieves the best performance on all metrics. Note that our model demonstrates significant improvements in FID, bringing animatable 3D GAN to the same level as unconditional 3D GANs (4.7 for EG3D [7]). For AED and APD, we also show superiority against baselines. Note that we still achieve the best pose consistency (0.019) when only considering yaw and pitch." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.512, + 0.649, + 0.527 + ], + "angle": 0, + "content": "4.2. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.536, + 0.892, + 0.672 + ], + "angle": 0, + "content": "Static tri-planes. As suggested by the grey lines in Fig. 8, this baseline removes the static tri-planes \\( T_{static} \\) and entangles both dynamic and static components in \\( T_{uv} \\). As illustrated in the first row of Fig. 6, the identities change when varying expressions. Since there are no explicit constraints for identity consistency, the model would be prone to unexpected entanglement between expression and identity. Tab. 2 shows a similar trend where removing \\( T_{static} \\) leads to worse identity consistency." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.673, + 0.892, + 0.884 + ], + "angle": 0, + "content": "Mouth Synthesis. When removing the mouth synthesis module \\( G_{\\text{teeth}} \\), we consider two altered choices: representing mouth features by \\( T_{\\text{static}} \\) or \\( T_{uv} \\), named w/o \\( G_{\\text{teeth}}(s) \\) (red lines in Fig. 8) and w/o \\( G_{\\text{teeth}}(uv) \\) (blue lines in Fig. 8), respectively. The first baseline, illustrated in the second row of Fig. 6, suffers from 'hole' artifacts of mouth. This is because inferior teeth move along with the jaw rotations and thus cannot be modeled by static features. The second baseline modeling teeth with \\( T_{uv} \\) also leads to an unreasonable mouth interior since the FLAME template doesn't account for teeth area and the neural textures for teeth would be sampled from other unrelated areas leading to artifacts. Quantitatively, both baselines without \\( G_{\\text{teeth}} \\) show significant degradation in AED and APD." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.886, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Deformation-aware discriminator. Tab. 2 demonstrates" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20997" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.124, + 0.1, + 0.164, + 0.112 + ], + "angle": 0, + "content": "Source" + }, + { + "type": "image_caption", + "bbox": [ + 0.211, + 0.1, + 0.264, + 0.111 + ], + "angle": 0, + "content": "Inversion" + }, + { + "type": "image_caption", + "bbox": [ + 0.405, + 0.1, + 0.465, + 0.111 + ], + "angle": 0, + "content": "Animation" + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.121, + 0.291, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.296, + 0.121, + 0.581, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.584, + 0.093, + 0.848, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.2, + 0.291, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.295, + 0.2, + 0.581, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.586, + 0.2, + 0.871, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.285, + 0.89, + 0.314 + ], + "angle": 0, + "content": "Figure 7. Applications of our model. We use PTI [53] to fit 3D-aware avatars for real portraits and animate them with sampled video clips. Furthermore, we leverage StyleGAN-NADA [21] into 3D settings and adapt these avatars into textually-prescribed domains." + }, + { + "type": "image", + "bbox": [ + 0.1, + 0.327, + 0.253, + 0.532 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.127, + 0.538, + 0.209, + 0.548 + ], + "angle": 0, + "content": "(a) Default Structure" + }, + { + "type": "image", + "bbox": [ + 0.258, + 0.327, + 0.453, + 0.533 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.311, + 0.538, + 0.397, + 0.548 + ], + "angle": 0, + "content": "(b) Altered Structures" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.564, + 0.47, + 0.592 + ], + "angle": 0, + "content": "Figure 8. Illustrations of the proposed three variants of our model for ablation study." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.595, + 0.47, + 0.67 + ], + "angle": 0, + "content": "that the deformation-aware discriminator with \\( I_{\\text{synthetic}} \\) input improves both animation accuracy and identity consistency, at the negligible expense of slightly reduced image quality. In Fig. 6, we see that this design shows a better reconstruction of detailed expressions, e.g. eye blinks." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.681, + 0.212, + 0.697 + ], + "angle": 0, + "content": "4.3. Applications" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.469, + 0.81 + ], + "angle": 0, + "content": "One-shot portrait animation. Fig. 7 shows the application of our model for one-shot head avatars. The learned generative animatbale 3D representation with expressive latent space can serve as a strong 3D prior for high-fidelity single-view 3D reconstruction and animation. Note that we can generate natural and consistent animations without video data training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Animatable 3D-aware stylization. Inspired by IDE-3D [58], we incorporate 2D CLIP-guided style tranfer methods [21] with our animatable 3D representation for 3D-aware portrait stylization. The right three columns of Fig. 7 show examples of text-driven, stylized portrait animation. Specifically, to adapt a pre-trained model through" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.328, + 0.893, + 0.48 + ], + "angle": 0, + "content": "only a textual prompt, we optimize the generator with two kinds of CLIP-based guidance [21]. However, leveraging text-guided 2D methods directly into the 3D setting is challenging as it tends to break 3D awareness and deformation awareness inherited in the generator parameters. To this end, we make some necessary modifications to the training framework. Please refer to the supplemental material for details. As shown in Fig. 7, we achieve high-quality 3D-aware stylized portrait synthesis with preserving well properties (i.e. 3D consistency and accurate animation)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.493, + 0.763, + 0.509 + ], + "angle": 0, + "content": "5. Limitations and future work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.519, + 0.892, + 0.64 + ], + "angle": 0, + "content": "Though our approach enables reasonable extrapolation on some rare expressions, it struggles to model some others with full consistency, such as one-side mouth up, sticking tongue out, etc. Besides, the rendered portraits still suffer from slight texture flickering caused by 2D superresolution. We leave both for future work. Furthermore, our model has the potential to provide a strong 3D prior to accelerating person-specific avatar reconstruction." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.654, + 0.619, + 0.67 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.68, + 0.892, + 0.816 + ], + "angle": 0, + "content": "We have presented Next3D, a novel animatable 3D representation for unsupervised learning of high-quality and 3D-consistent virtual facial avatars from unstructured 2D images. Our approach has pushed the frontier of photorealistic animatabale 3D-aware image synthesis. Serving as a strong 3D prior, we believe our learned 3D representation will boost a series of downstream applications including 3D-aware one-shot facial avatars and animatable out-of-domain avatar generation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.83, + 0.667, + 0.847 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.856, + 0.892, + 0.9 + ], + "angle": 0, + "content": "This paper is supported by National Key RD Program of China (2022YFF0902200) and the NSFC project No.62125107." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20998" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] ShahRukh Athar, Zexiang Xu, Kalyan Sunkavalli, Eli Shechtman, and Zhixin Shu. Rignerf: Fully controllable neural 3d portraits. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20364-20373, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.187, + 0.472, + 0.243 + ], + "angle": 0, + "content": "[2] Hendrik Baatz, Jonathan Granskog, Marios Papas, Fabrice Rousselle, and Jan Novák. Nerf-tex: Neural reflectance field textures. In Computer Graphics Forum. Wiley Online Library, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.246, + 0.472, + 0.299 + ], + "angle": 0, + "content": "[3] Alexander W. Bergman, Petr Kellnhofer, Wang Yifan, Eric R. Chan, David B. Lindell, and Gordon Wetzstein. Generative neural articulated radiance fields. In NeurIPS, 2022. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.302, + 0.472, + 0.372 + ], + "angle": 0, + "content": "[4] Volker Blanz and Thomas Vetter. A morphable model for the synthesis of 3d faces. In Proceedings of the 26th Annual Conference on Computer Graphics and Interactive Techniques, SIGGRAPH '99, page 187-194, USA, 1999. ACM Press/Addison-Wesley Publishing Co. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.374, + 0.472, + 0.43 + ], + "angle": 0, + "content": "[5] Andrei Burov, Matthias Nießner, and Justus Thies. Dynamic surface function networks for clothed human bodies. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10754-10764, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.432, + 0.472, + 0.5 + ], + "angle": 0, + "content": "[6] Shengqu Cai, Anton Obukhov, Dengxin Dai, and Luc Van Gool. Pix2nerf: Unsupervised conditional p-gan for single image to neural radiance fields translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3981-3990, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.503, + 0.472, + 0.599 + ], + "angle": 0, + "content": "[7] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In Proceedings of the IEEE International Conference on Computer Vision (CVPR), 2021. 2, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.602, + 0.472, + 0.671 + ], + "angle": 0, + "content": "[8] Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.673, + 0.472, + 0.742 + ], + "angle": 0, + "content": "[9] Xu Chen, Tianjian Jiang, Jie Song, Jinlong Yang, Michael J Black, Andreas Geiger, and Otmar Hilliges. gdna: Towards generative detailed neural avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20427-20437, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.744, + 0.47, + 0.785 + ], + "angle": 0, + "content": "[10] Yuedong Chen, Qianyi Wu, Chuanxia Zheng, Tat-Jen Cham, and Jianfei Cai. Sem2nerf: Converting single-view semantic masks to neural radiance fields. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.788, + 0.47, + 0.843 + ], + "angle": 0, + "content": "[11] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5939-5948, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.47, + 0.899 + ], + "angle": 0, + "content": "[12] Hang Dai, Nick Pears, William Smith, and Christian Duncan. Statistical modeling of craniofacial shape and texture. International Journal of Computer Vision, 128(2):547-571, 2020. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.472, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.161 + ], + "angle": 0, + "content": "[13] Yu Deng, Jiaolong Yang, Dong Chen, Fang Wen, and Tong Xin. Disentangled and controllable face image generation via 3d imitative-contrastive learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[14] Yu Deng, Jiaolong Yang, Jianfeng Xiang, and Xin Tong. Gram: Generative radiance manifolds for 3d-aware image generation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.892, + 0.302 + ], + "angle": 0, + "content": "[15] Yu Deng, Jiaolong Yang, Sicheng Xu, Dong Chen, Yunde Jia, and Xin Tong. Accurate 3d face reconstruction with weakly-supervised learning: From single image to image set. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 0–0, 2019. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.305, + 0.892, + 0.373 + ], + "angle": 0, + "content": "[16] Michail Christos Doukas, Stefanos Zafeiriou, and Viktoriia Sharmanska. Headgan: One-shot neural head synthesis and editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14398-14407, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.376, + 0.892, + 0.431 + ], + "angle": 0, + "content": "[17] Yao Feng, Haiwen Feng, Michael J Black, and Timo Bolkart. Learning an animatable detailed 3d face model from inthe-wild images. ACM Transactions on Graphics (ToG), 40(4):1-13, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.433, + 0.892, + 0.502 + ], + "angle": 0, + "content": "[18] Ohad Fried, Ayush Tewari, Michael Zollhöfer, Adam Finkelstein, Eli Shechtman, Dan B Goldman, Kyle Genova, Zeyu Jin, Christian Theobalt, and Maneesh Agrawala. Text-based editing of talking-head video. ACM Transactions on Graphics (TOG), 38(4):1-14, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.504, + 0.892, + 0.545 + ], + "angle": 0, + "content": "[19] Matheus Gadelha, Subhransu Maji, and Rui Wang. 3d shape induction from 2d views of multiple objects. In International Conference on 3D Vision (3DV), 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.547, + 0.892, + 0.616 + ], + "angle": 0, + "content": "[20] Guy Gafni, Justus Thies, Michael Zollhofer, and Matthias Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8649-8658, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.618, + 0.892, + 0.673 + ], + "angle": 0, + "content": "[21] Rinon Gal, Or Patashnik, Haggai Maron, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Stylegan-nada: Clip-guided domain adaptation of image generators. ACM Transactions on Graphics (TOG), 41(4):1-13, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.675, + 0.892, + 0.744 + ], + "angle": 0, + "content": "[22] Baris Gecer, Binod Bhattacharai, Josef Kittler, and Tae-Kyun Kim. Semi-supervised adversarial learning to generate photorealistic face images of new identities from 3d morphable model. In Proceedings of the European conference on computer vision (ECCV), pages 217-234, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.746, + 0.892, + 0.815 + ], + "angle": 0, + "content": "[23] Baris Gecer, Stylianos Ploumpis, Irene Kotsia, and Stefanos Zafeiriou. Ganfit: Generative adversarial network fitting for high fidelity 3d face reconstruction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1155-1164, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.817, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[24] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.874, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[25] Philip-William Grassal, Malte Prinzler, Titus Leistner, Carsten Rother, Matthias Nießner, and Justus Thies. Neural" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20999" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.472, + 0.134 + ], + "angle": 0, + "content": "head avatars from monocular rgb videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18653-18664, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.472, + 0.219 + ], + "angle": 0, + "content": "[26] Artur Grigorev, Karim Iskakov, Anastasia Ianina, Renat Bashirov, Ilya Zakharkin, Alexander Vakhitov, and Victor Lempitsky. Stylepeople: A generative model of fullbody human avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5151-5160, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.221, + 0.471, + 0.276 + ], + "angle": 0, + "content": "[27] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. Stylenerf: A style-based 3d aware generator for high-resolution image synthesis. In International Conference on Learning Representations, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.278, + 0.471, + 0.332 + ], + "angle": 0, + "content": "[28] Philipp Henzler, Niloy J Mitra, and Tobias Ritschel. Escaping plato's cave: 3d shape from adversarial rendering. In Proceedings of the IEEE International Conference on Computer Vision (CVPR), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.334, + 0.471, + 0.403 + ], + "angle": 0, + "content": "[29] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In Advances in Neural Information Processing Systems (NeurIPS), 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.405, + 0.471, + 0.46 + ], + "angle": 0, + "content": "[30] Fangzhou Hong, Zhaoxi Chen, Yushi LAN, Liang Pan, and Ziwei Liu. EVA3d: Compositional 3d human generation from 2d image collections. In International Conference on Learning Representations, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.462, + 0.471, + 0.531 + ], + "angle": 0, + "content": "[31] Yang Hong, Bo Peng, Haiyao Xiao, Ligang Liu, and Juyong Zhang. Headnerf: A real-time nerf-based parametric head model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20374-20384, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.533, + 0.471, + 0.601 + ], + "angle": 0, + "content": "[32] Angjoo Kanazawa, Shubham Tulsiani, Alexei A Efros, and Jitendra Malik. Learning category-specific mesh reconstruction from image collections. In Proceedings of the European Conference on Computer Vision (ECCV), pages 371-386, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.603, + 0.471, + 0.659 + ], + "angle": 0, + "content": "[33] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4401-4410, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.661, + 0.471, + 0.729 + ], + "angle": 0, + "content": "[34] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.731, + 0.471, + 0.8 + ], + "angle": 0, + "content": "[35] Hyeongwoo Kim, Pablo Garrido, Ayush Tewari, Weipeng Xu, Justus Thies, Matthias Niessner, Patrick Pérez, Christian Richardt, Michael Zollhöfer, and Christian Theobalt. Deep video portraits. ACM Transactions on Graphics (TOG), 37(4):1-14, 2018. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.471, + 0.871 + ], + "angle": 0, + "content": "[36] Jeong-gi Kwak, Yuanming Li, Dongsik Yoon, Donghyeon Kim, David Han, and Hanseok Ko. Injecting 3d perception of controllable nerf-gan into stylegan for editable portrait image synthesis. In Proceedings of European Conference on Computer Vision (ECCV), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[37] Henning Lange and J Nathan Kutz. Fc2t2: The fast continuous convolutional taylor transform with applications in vi" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "sion and graphics. arXiv preprint arXiv:2111.00110, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.123, + 0.892, + 0.165 + ], + "angle": 0, + "content": "[38] Tianye Li, Timo Bolkart, Michael J Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4d scans. ACM Trans. Graph., 36(6):194-1, 2017. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.167, + 0.892, + 0.209 + ], + "angle": 0, + "content": "[39] Connor Z Lin, David B Lindell, Eric R Chan, and Gordon Wetzstein. 3d gan inversion for controllable portrait image animation. arXiv preprint arXiv:2203.13441, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.211, + 0.892, + 0.28 + ], + "angle": 0, + "content": "[40] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European Conference on Computer Vision (ECCV), 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.283, + 0.892, + 0.352 + ], + "angle": 0, + "content": "[41] Thu Nguyen-Phuoc, Chuan Li, Lucas Theis, Christian Richardt, and Yong-Liang Yang. Hologan: Unsupervised learning of 3d representations from natural images. In Proceedings of the IEEE International Conference on Computer Vision, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.354, + 0.892, + 0.424 + ], + "angle": 0, + "content": "[42] Thu H Nguyen-Phuoc, Christian Richardt, Long Mai, Yongliang Yang, and Niloy Mitra. Blockgan: Learning 3d object-aware scene representations from unlabelled images. In Advances in Neural Information Processing Systems (NeurIPS), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.427, + 0.892, + 0.483 + ], + "angle": 0, + "content": "[43] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.485, + 0.892, + 0.542 + ], + "angle": 0, + "content": "[44] Atsuhiro Noguchi, Xiao Sun, Stephen Lin, and Tatsuya Harada. Unsupervised learning of efficient geometry-aware neural articulated representations. In Proceedings of European Conference on Computer Vision (ECCV), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.543, + 0.892, + 0.613 + ], + "angle": 0, + "content": "[45] Michael Oechsle, Lars Mescheder, Michael Niemeyer, Thilo Strauss, and Andreas Geiger. Texture fields: Learning texture representations in function space. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4531-4540, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.615, + 0.892, + 0.698 + ], + "angle": 0, + "content": "[46] Roy Or-El, Xuan Luo, Mengyi Shan, Eli Shechtman, Jeong Joon Park, and Ira Kemelmacher-Shlizerman. StyleSDF: High-Resolution 3D-Consistent Image and Geometry Generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.701, + 0.892, + 0.77 + ], + "angle": 0, + "content": "[47] Xingang Pan, Xudong Xu, Chen Change Loy, Christian Theobalt, and Bo Dai. A shading-guided generative implicit model for shape-accurate 3d-aware image synthesis. Advances in Neural Information Processing Systems, 34:20002-20013, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.772, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[48] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[49] Songyou Peng, Michael Niemeyer, Lars Mescheder, Marc Pollefeys, and Andreas Geiger. Convolutional occupancy networks. In European Conference on Computer Vision, pages 523-540. Springer, 2020. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "21000" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.16 + ], + "angle": 0, + "content": "[50] Hanspeter Pfister, Matthias Zwicker, Jeroen Van Baar, and Markus Gross. Surfels: Surface elements as rendering primitives. In Proceedings of the 27th annual conference on Computer graphics and interactive techniques, pages 335-342, 2000. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.162, + 0.47, + 0.245 + ], + "angle": 0, + "content": "[51] Stylianos Ploumpis, Evangelos Ververas, Eimear O'Sullivan, Stylianos Moschoglou, Haoyang Wang, Nick Pears, William AP Smith, Baris Gecer, and Stefanos Zafeiriou. Towards a complete 3d morphable model of the human head. IEEE transactions on pattern analysis and machine intelligence, 43(11):4142-4160, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.247, + 0.47, + 0.314 + ], + "angle": 0, + "content": "[52] Yurui Ren, Ge Li, Yuanqi Chen, Thomas H Li, and Shan Liu. Pirenderer: Controllable portrait image generation via semantic neural rendering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13759-13768, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.316, + 0.47, + 0.357 + ], + "angle": 0, + "content": "[53] Daniel Roich, Ron Mokady, Amit H Bermano, and Daniel Cohen-Or. Pivotal tuning for latent-based editing of real images. ACM Trans. Graph., 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.358, + 0.47, + 0.412 + ], + "angle": 0, + "content": "[54] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. Graf: Generative radiance fields for 3d-aware image synthesis. In Advances in Neural Information Processing Systems (NeurIPS), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.414, + 0.47, + 0.468 + ], + "angle": 0, + "content": "[55] Katja Schwarz, Axel Sauer, Michael Niemeyer, Yiyi Liao, and Andreas Geiger. Voxgraf: Fast 3d-aware image synthesis with sparse voxel grids. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.47, + 0.47, + 0.538 + ], + "angle": 0, + "content": "[56] Vincent Sitzmann, Justus Thies, Felix Heide, Matthias Nießner, Gordon Wetzstein, and Michael Zollhofer. Deepvoxels: Learning persistent 3d feature embeddings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2437-2446, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.539, + 0.47, + 0.593 + ], + "angle": 0, + "content": "[57] Ivan Skorokhodov, Sergey Tulyakov, Yiqun Wang, and Peter Wonka. Epigraph: Rethinking training of 3d gans. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.595, + 0.47, + 0.662 + ], + "angle": 0, + "content": "[58] Jingxiang Sun, Xuan Wang, Yichun Shi, Lizhen Wang, Jue Wang, and Yebin Liu. Ide-3d: Interactive disentangled editing for high-resolution 3d-aware portrait synthesis. ACM Transactions on Graphics (TOG), 41(6):1-10, 2022. 2, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.665, + 0.47, + 0.732 + ], + "angle": 0, + "content": "[59] Jingxiang Sun, Xuan Wang, Yong Zhang, Xiaoyu Li, Qi Zhang, Yebin Liu, and Jue Wang. Fenerf: Face editing in neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.734, + 0.47, + 0.802 + ], + "angle": 0, + "content": "[60] Keqiang Sun, Shangzhe Wu, Zhaoyang Huang, Ning Zhang, Quan Wang, and HongSheng Li. Controllable 3d face synthesis with conditional generative occupancy fields. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.47, + 0.858 + ], + "angle": 0, + "content": "[61] Junshu Tang, Bo Zhang, Binxin Yang, Ting Zhang, Dong Chen, Lizhuang Ma, and Fang Wen. Explicitly controllable 3d-aware portrait generation. arXiv preprint arXiv:2209.05434, 2022. 2, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.47, + 0.902 + ], + "angle": 0, + "content": "[62] Ayush Tewari, Mohamed Elgharib, Gaurav Bharaj, Florian Bernard, Hans-Peter Seidel, Patrick Pérez, Michael Zollhofer, and Christian Theobalt. Stylerig: Rigging stylegan" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "for 3d control over portrait images. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.136, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[63] Ayush Tewari, Justus Thies, Ben Mildenhall, Pratul Srinivasan, Edgar Tretschk, W Yifan, Christoph Lassner, Vincent Sitzmann, Ricardo Martin-Brualla, Stephen Lombardi, et al. Advances in neural rendering. Computer Graphics Forum, 41(2):703-735, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.207, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[64] Justus Thies, Michael Zollhöfer, and Matthias Nießner. Deferred neural rendering: Image synthesis using neural textures. ACM Transactions on Graphics (TOG), 38(4):1-12, 2019. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.264, + 0.892, + 0.333 + ], + "angle": 0, + "content": "[65] Justus Thies, Michael Zollhofer, Marc Stamminger, Christian Theobalt, and Matthias Nießner. Face2face: Real-time face capture and reenactment of rgb videos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2387-2395, 2016. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.335, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[66] Lizhen Wang, Zhiyuan Chen, Tao Yu, Chenguang Ma, Liang Li, and Yebin Liu. Faceverse: a fine-grained and detail-controllable 3d face morphable model from a hybrid dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20333-20342, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.42, + 0.892, + 0.488 + ], + "angle": 0, + "content": "[67] Jiajun Wu, Chengkai Zhang, Tianfan Xue, Bill Freeman, and Josh Tenenbaum. Learning a probabilistic latent space of object shapes via 3d generative-adversarial modeling. Advances in neural information processing systems, 29, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.491, + 0.892, + 0.546 + ], + "angle": 0, + "content": "[68] Yue Wu, Yu Deng, Jiaolong Yang, Fangyun Wei, Chen Qifeng, and Xin Tong. Anifacegan: Animatable 3d-aware face image generation for video avatars. In Advances in Neural Information Processing Systems, 2022. 2, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.548, + 0.892, + 0.616 + ], + "angle": 0, + "content": "[69] Qiangeng Xu, Weiyue Wang, Duygu Ceylan, Radomir Mech, and Ulrich Neumann. Disn: Deep implicit surface network for high-quality single-view 3d reconstruction. Advances in Neural Information Processing Systems, 32, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.619, + 0.892, + 0.687 + ], + "angle": 0, + "content": "[70] Sicheng Xu, Jiaolong Yang, Dong Chen, Fang Wen, Yu Deng, Yunde Jia, and Xin Tong. Deep 3d portrait from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7710-7720, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.69, + 0.892, + 0.759 + ], + "angle": 0, + "content": "[71] Yinghao Xu, Sida Peng, Ceyuan Yang, Yujun Shen, and Bolei Zhou. 3d-aware image synthesis via learning structural and textural representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18430-18439, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.761, + 0.892, + 0.815 + ], + "angle": 0, + "content": "[72] Yang Xue, Yuheng Li, Krishna Kumar Singh, and Yong Jae Lee. Giraffe hd: A high-resolution 3d-aware generative model. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.818, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[73] Fei Yin, Yong Zhang, Xiaodong Cun, Mingdeng Cao, Yanbo Fan, Xuan Wang, Qingyan Bai, Baoyuan Wu, Jue Wang, and Yujiu Yang. Styleheat: One-shot high-resolution editable talking face generation via pretrained stylegan. In Proceedings of the European conference on computer vision (ECCV), 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "21001" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.16 + ], + "angle": 0, + "content": "[74] Jianfeng Zhang, Zihang Jiang, Dingdong Yang, Hongyi Xu, Yichun Shi, Guoxian Song, Zhongcong Xu, Xinchao Wang, and Jiashi Feng. Avatargen: a 3d generative model for animatable human avatars. arXiv preprint arXiv:2208.00561, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.472, + 0.218 + ], + "angle": 0, + "content": "[75] Jingbo Zhang, Xiaoyu Li, Ziyu Wan, Can Wang, and Jing Liao. Fdnerf: Few-shot dynamic neural radiance fields for face reconstruction and expression editing. arXiv preprint arXiv:2208.05751, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.22, + 0.47, + 0.274 + ], + "angle": 0, + "content": "[76] Jichao Zhang, E. Sangineto, Hao Tang, Aliaksandr Siarohin, Zhun Zhong, N. Sebe, and Wei Wang. 3d-aware semantic-guided generative model for human synthesis. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.277, + 0.47, + 0.332 + ], + "angle": 0, + "content": "[77] Jichao Zhang, Aliaksandr Siarohin, Yahui Liu, Hao Tang, Nicu Sebe, and Wei Wang. Training and tuning generative neural radiance fields for attribute-conditional 3d-aware face generation. arXiv preprint arXiv:2208.12550, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.334, + 0.47, + 0.403 + ], + "angle": 0, + "content": "[78] Xuanmeng Zhang, Zhedong Zheng, Daiheng Gao, Bang Zhang, Pan Pan, and Yi Yang. Multi-view consistent generative adversarial networks for 3d-aware image synthesis. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.404, + 0.47, + 0.471 + ], + "angle": 0, + "content": "[79] Xiaoming Zhao, Fangchang Ma, David Guera, Zhile Ren, Alexander G. Schwing, and Alex Colburn. Generative multiplane images: Making a 2d gan 3d-aware. In Proceedings of European Conference on Computer Vision (ECCV), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.474, + 0.47, + 0.544 + ], + "angle": 0, + "content": "[80] Yufeng Zheng, Victoria Fernández Abrevaya, Marcel C Bühler, Xu Chen, Michael J Black, and Otmar Hilliges. Im avatar: Implicit morphable head avatars from videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13545-13555, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.545, + 0.47, + 0.587 + ], + "angle": 0, + "content": "[81] Peng Zhou, Lingxi Xie, Bingbing Ni, and Qi Tian. Cips-3d: A 3d-aware generator of gans based on conditionally-independent pixel synthesis. arXiv:2110.09788, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.588, + 0.47, + 0.657 + ], + "angle": 0, + "content": "[82] Jun-Yan Zhu, Zhoutong Zhang, Chengkai Zhang, Jiajun Wu, Antonio Torralba, Josh Tenenbaum, and Bill Freeman. Visual object networks: Image generation with disentangled 3d representations. In Advances in Neural Information Processing Systems (NeurIPS), 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.658, + 0.47, + 0.699 + ], + "angle": 0, + "content": "[83] Peiye Zhuang, Liqian Ma, Oluwasanmi Koyejo, and Alexander Schwing. Controllable radiance fields for dynamic face synthesis. In Proc. 3DV, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.699 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "21002" + } + ] +] \ No newline at end of file diff --git a/2023/Next3D_ Generative Neural Texture Rasterization for 3D-Aware Head Avatars/7b615aeb-a836-44b1-beba-de01bf0423b2_origin.pdf b/2023/Next3D_ Generative Neural Texture Rasterization for 3D-Aware Head Avatars/7b615aeb-a836-44b1-beba-de01bf0423b2_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..091b11e1aefa76bc6cc266b91cec15603e6b329e --- /dev/null +++ b/2023/Next3D_ Generative Neural Texture Rasterization for 3D-Aware Head Avatars/7b615aeb-a836-44b1-beba-de01bf0423b2_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f76425f5b952c8b20c60f88a3fb221e2ab885d2155368f005eb543bff746379 +size 9179706 diff --git a/2023/Next3D_ Generative Neural Texture Rasterization for 3D-Aware Head Avatars/full.md b/2023/Next3D_ Generative Neural Texture Rasterization for 3D-Aware Head Avatars/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c7b53a8a48abff202647c5a22386c9a20de388b2 --- /dev/null +++ b/2023/Next3D_ Generative Neural Texture Rasterization for 3D-Aware Head Avatars/full.md @@ -0,0 +1,337 @@ +# Next3D: Generative Neural Texture Rasterization for 3D-Aware Head Avatars + +Jingxiang Sun + +Xuan Wang $^{2}$ + +Lizhen Wang $^{14}$ + +Xiaoyu Li3 + +Yong Zhang3 + +Hongwen Zhang + +Yebin Liu + +$^{1}$ Tsinghua University + +$^{2}$ Ant Group + +$^{3}$ Tencent AI Lab + +4NNKosmos + +![](images/c04ab9bd231f8e61758d2d3d2122949aed22cecb230f0433f950fe327204b0b5.jpg) +Gaze Animation +3D-Aware Domain Adaptation +Figure 1. Our 3D GAN synthesizes generative, high-quality, and 3D-consistent facial avatars from unstructured 2D images. Unlike current animatable 3D GANs that only modify yaw-pitch head poses and facial expressions, our approach enables fine-grained control over full-head rotations, facial expressions, eye blinks, and gaze directions with strict 3D consistency and a high level of photorealism. Our approach also provides strong 3D priors for downstream tasks such as 3D-aware stylization. + +# Abstract + +3D-aware generative adversarial networks (GANs) synthesize high-fidelity and multi-view-consistent facial images using only collections of single-view 2D imagery. Towards fine-grained control over facial attributes, recent efforts incorporate 3D Morphable Face Model (3DMM) to describe deformation in generative radiance fields either explicitly or implicitly. Explicit methods provide fine-grained expression control but cannot handle topological changes caused by hair and accessories, while implicit ones can model varied topologies but have limited generalization caused by the unconstrained deformation fields. We propose a novel 3D GAN framework for unsupervised learning of generative, high-quality and 3D-consistent facial avatars from unstructured 2D images. To achieve both deformation accuracy and topological flexibility, we propose a 3D representation called Generative Texture-Rasterized Tri-planes. The proposed representation learns Generative Neural Textures on top of parametric mesh templates + +and then projects them into three orthogonal-viewed feature planes through rasterization, forming a tri-plane feature representation for volume rendering. In this way, we combine both fine-grained expression control of mesh-guided explicit deformation and the flexibility of implicit volumetric representation. We further propose specific modules for modeling mouth interior which is not taken into account by 3DMM. Our method demonstrates state-of-the-art 3D-aware synthesis quality and animation ability through extensive experiments. Furthermore, serving as 3D prior, our animatable 3D representation boosts multiple applications including one-shot facial avatars and 3D-aware stylization. Project page: https://mrtornado24.github.io/Next3D/. Code: https://github.com/MrTornado24/Next3D. + +# 1. Introduction + +Animatable portrait synthesis is essential for movie postproduction, visual effects, augmented reality (AR), and vir + +tual reality (VR) telepresence applications. Efficient animatable portrait generators should be capable of synthesizing diverse high-fidelity portraits with full control of the rigid head pose, facial expressions and gaze directions at a fine-grained level. The main challenges of this task lie in how to model accurate deformation and preserve identity through animation in the generative setting, i.e. training with only unstructured corpus of 2D images. + +Several 2D generative models perform image animation by incorporating the 3D Morphable Face Models (3DMM) [4] into the portrait synthesis [13, 16, 35, 52, 62, 65, 70, 73]. These 2D-based methods achieve photorealism but suffer from shape distortion during large motion due to a lack of geometry constraints. Towards better view consistency, many recent efforts incorporate 3DMM with 3D GANs, learning to synthesize animatable and 3D consistent portraits from only 2D image collections in an unsupervised manner [3, 30, 39, 44, 60, 61, 68, 74]. Bergman et al. [3] propose an explicit surface-driven deformation field for warping radiance fields. While modeling accurate facial deformation, it cannot handle topological changes caused by non-facial components, e.g. hair, glasses, and other accessories. AnifaceGAN [68] builds an implicit 3DMM-conditioned deformation field and constrains animation accuracy by imitation learning. It achieves smooth animation on interpolated expressions, however, struggles to generate reasonable extrapolation due to the under-constrained deformation field. Therefore, The key challenge of this task is modeling deformation in the 3D generative setting for animation accuracy and topological flexibility. + +In this paper, we propose a novel 3D GAN framework for unsupervised learning of generative, high-quality, and 3D-consistent facial avatars from unstructured 2D images. Our model splits the whole head into dynamic and static parts, and models them respectively. For dynamic parts, the key insight is to combine both fine-grained expression control of mesh-guided explicit deformation and flexibility of implicit volumetric representation. To this end, we propose a novel representation, Generative Texture-Rasterized Triplanes, which learns the facial deformation through Generative Neural Textures on top of a parametric template mesh and samples them into three orthogonal-viewed and axis-aligned feature planes through standard rasterization, forming a tri-plane feature representation. Such texture-rasterized tri-planes re-form high-dimensional dynamic surface features in a volumetric representation for efficient volume rendering and thus inherit both the accurate control of the mesh-driven deformation and the expressiveness of volumetric representations. Furthermore, we represent static components (body, hair, background, etc.) by another triplane branch, and integrate both through alpha blending. + +Another key insight of our method is to model the mouth interior which is not taken into account by 3DMM. Mouth + +interior is crucial for animation quality but often ignored by prior arts. We propose an efficient teeth synthesis module, formed as a style-modulated UNet, to complete the inner mouth features missed by the template mesh. To further regularize the deformation accuracy, we introduce a deformation-aware discriminator which takes as input synthetic renderings, encouraging the alignment of the final outputs with the 2D projection of the expected deformation. + +To summarize, the contributions of our approach are: + +- We present an animatable 3D-aware GAN framework for photorealistic portrait synthesis with fine-grained animation, including expressions, eye blinks, gaze direction and full head poses. +- We propose Generative Texture-Rasterized Triplanes, an efficient deformable 3D representation that inherits both fine-grained expression control of mesh-guided explicit deformation and flexibility of implicit volumetric representation. +- Our learned generative animatable 3D representation can serve as a strong 3D prior and boost the downstream application of 3D-aware one-shot facial avatars. Our model also pushes the frontier of 3D stylization with high-quality out-of-domain facial avatars. + +# 2. Related Work + +Generative 3D-aware Image Synthesis. Generative adversarial networks [24] have achieved photorealistic synthesis in 2D domain. Building on the success of 2D GANs, many efforts have lifted the image synthesis into 3D with explicit view control. Early voxel-based approaches [19, 28, 41, 42, 67, 82] adopt the 3D CNN generators whose heavy computational burden limits the high-resolution image synthesis. Recent works incorporate more efficient neural scene representations, such as fully implicit networks [6, 8-10, 14, 47, 54, 57, 81], sparse voxel grids [55], multiple planes [79] or a combination of low-resolution feature volume and 2D super-resolution [7, 27, 43, 46, 71, 72, 76, 78]. We leverage the tri-plane representation proposed in [7] and endow it with animation ability by the orthogonal-rasterized Generative Neural Textures. + +Some other current works [3,30,36,44,58,59,61,68,77, 83] focus on the editability of 3D-aware generative models. FENeRF [59] and IDE-3D [58] perform semantic-guided 3D face editing by incorporating semantic-aware radiance fields and GAN inversion. However, they cannot produce continuous and stable editing on videos. Other methods [3, 30,44,61,68] employ 3D priors to achieve animatable image synthesis and the main differences lie on the deformation strategies including linear blend skinning [30,44], surface-driven deformation [3], 3DMM-guided latent decomposition [61] and neural deformation fields [68]. These approaches either don't allow for topology changes or need + +![](images/a1dbff3e376f7c470d28d06fbe8a108f32f1ed7c3003834e4a0de514c95607cb.jpg) +Figure 2. Our 3D GAN framework consists of two tri-plane branches $T_{uv}$ and $T_{static}$ modeling dynamic and static components. $T_{uv}$ is formed by the orthogonal rasterized Generative Neural Textures which are synthesized by a StyleGAN generator, $G_{texture}$ , on top of deformable template mesh. $T_{static}$ is synthesized by another StyleGAN generator, $G_{static}$ . The mouth synthesis module, $G_{teeth}$ , is presented for completing mouth interior. Blended triplanes are incorporated with hybrid neural renderer consisting of volume rendering and a super-resolution module $G_{superres}$ . For discrimination, synthetic renderings $I_{synthetic}$ is taken into the dual discriminator $D_{dual}$ . + +elaborate loss design to ensure the accuracy of deformation. On the contrary, our approach naturally achieves accurate animation with explicit mesh guidance and further allows for topology changes by adapting surface deformation into a continuous volumetric representation. + +Facial animation with 3D morphable face models. Blanz et al. [4] model facial texture and shape as vector spaces, known as the 3D Morphable Model (3DMM). Extensions of 3DMM, such as full-head PCA models [12, 51], blend-shape models [38], are extensively studied and widely used in facial animation tasks [23, 65]. Benefiting from 3DMM, these methods can model the deformation of facial parts accurately and continuously, nevertheless, struggle to represent non-facial areas missed by 3DMM, e.g. hair, teeth, eyes, and body. Moreover, these methods are prone to lacking facial details. To fill the missing areas and complete more realistic facial details, later works [13, 16, 18, 22, 35, 52, 62, 64, 70] apply learned approaches on top of 3DMM renderings. DiscoFaceGAN [13] maps 3DMM parameters into the latent space and decouple them by imitative-contrastive learning. Though efficient and photorealistic animation is achieved, these 2D methods don't model 3D geometry and thus cannot remain strict 3D consistency with large head pose changes. For strict 3D consistency, recent efforts [1, 20, 25, 75, 80] incorporate 3DMM into volumetric representations [31, 32, 40, 49, 56] to achieve view-consistent facial animation. Furthermore, volumetric representations enable the modeling of thin structures such as hair, and also mouth interior thanks to its spatial continuity. While 3DMM have been adopted to animate radiance fields for single-scene scenarios, it is challenging to adapt to the generative setting with the absence of groundtruth supervision. + +Neural scene representations. The neural scene representations can be roughly categorized into implicit and explicit + +surface representations and volumetric representations [63]. The surface can be represented explicitly by point clouds [37, 50], meshes [2, 5, 45, 64], or defined implicitly as a zero level-set of a function [11, 48, 69], like signed distance function, which can be approximated by coordinate-based multi-layer perceptrons (MLPs). On the contrary, volumetric representations [32, 40, 49, 56] store volumetric properties (occupancies, radiance, colors, etc.) instead of the surface of an object. These properties can be stored in explicit voxel grids [32, 49, 56] or the weights of a neural network implicitly [40]. In this work, we propose a hybrid surface-volumetric representation. Specifically, we learn the deformable surface radiance by neural textures [26, 64] on top of the template mesh and rasterize it into three orthogonal-viewed feature planes. Then, the planes are reshaped to a tri-plane representation with decoding into neural radiance fields for volume rendering. + +# 3. Approach + +We present an animatable 3D-aware facial generator that equips with fine-grained expression and pose control, photorealistic rendering quality, and high-quality underlying geometry. The proposed method models dynamic and static components by two independent tri-plane branches. Specifically, we propose Generative Texture-rasterized Tri-planes for modeling dynamic facial parts (Sec. 3.1). Furthermore, we propose an efficient mouth synthesis module to complete the mouth interior that is not included in 3DMM (Sec. 3.2). We further adopt another tri-plane branch for the static components (Sec. 3.3). Both tri-planes are blended together for hybrid neural rendering (Sec. 3.4). We introduce an deformation-aware discriminator (Sec. 3.5) and illustrate the training objectives in Sec. 3.6. + +# 3.1. Generative texture-rasterized tri-planes + +EG3D [7] presents an efficient tri-plane-based hybrid 3D representation to synthesize high-resolution images with multi-view consistency. Nonetheless, EG3D lacks control over facial deformations and thus cannot be directly applied to animation tasks. To this end, we leverage Neural textures [64] to represent deformable facial parts. In general, Neural Textures are a set of learned high-dimensional feature maps that can be interpreted by a neural renderer. We extend it to our generative setting and synthesize the neural textures through a StyleGAN2 CNN generator $G_{\text{texture}}$ . As shown in Fig. 2, we first sample a latent code $z$ and map it into an intermediate latent space by the mapping network. Our texture generator architecture closely follows StyleGAN2 backbone [34], except producing a $256 \times 256 \times 32$ neural texture map, $T$ , instead of a three-channel RGB image. Storing a high-dimensional learned feature vector per texel, $T$ can be rasterized to a view-dependent screen-space feature map given a mesh with uv-texture parameterization and a target view as input. In our case, we use the FLAME template [38] to provide a coarse mesh that can be driven by deformation parameters. Given the predesigned texture mapping function, we employ the standard graphics pipeline to rasterize the neural textures from the texture space into the screen space based on the template mesh. We choose Neural Textures as the deformation method for two reasons. First, compared with other explicit deformation (e.g. linear blend skinning and surface-driven deformation) highly dependent on the accurate underlying geometry, Neural Textures embed high-level features which compensate the imperfect geometry and thus are more suitable for our settings where template meshes are not accurate. Furthermore, unlike implicit deformation methods [13, 61, 68], our explicit mesh-guided deformation alleviates the requirement of elaborate imitation learning while gaining better expression generalization (Fig. 4). + +Neural Textures encode surface deformation accurately with mesh guidance but lack generalization to 3D points far from surface. Besides, it also doesn't allow for topological changes. To this end, we propose Generative Texture-Rasterized Tri-planes, $T_{uv}$ which reshapes the rasterized textures into a tri-plane representation. Therefore, we can adapt such surface deformation into a continuous volume, and once a 3D point out of the surface could be projected within the facial part on at least one view (e.g. glasses), its feature can be interpolated by the meaningful surface feature through grid sampling in the tri-plane. Specifically, we rasterize neural textures based on the template mesh into three orthogonal views and place them in three axis-aligned feature planes. In practice, considering the zygomorphy, the rasterization is applied at both the left and right views and the rasterized features are concatenated for one single plane by summation. In this way, Our hybrid surface-volumetric + +representation inherits the best of both worlds: accurate mesh-guided facial deformation of Neural Textures and the continuity and topological flexibility of volumetric representations. See Fig. 5 for the topological-aware animation of portraits with glasses. + +# 3.2. Mouth synthesis module + +![](images/71e47e74911ff83045eae0d1df72dc8e8b6ba1b2fe084fd9c70ff17820e52c0a.jpg) +Figure 3. The teeth synthesis module consists of a teeth completing module and a neural blending module. The teeth competing module produces mouth interior conditioned on multi-scale teeth exterior features and latent codes $w$ . + +Since the FLAME template doesn't contain inner mouth, we propose a teeth synthesis module, $G_{\text{teeth}}$ , to complete the missing teeth features in $T_{uv}$ . As shown in Fig. 3, for each feature plane of $T_{uv}$ , we crop the teeth area by the expanded landmarks and resize it into $64 \times 64$ . Then, the stacked mouth features are processed by $G_{\text{teeth}}$ which employs a style-modulated UNet [66]. The downsampling process of $G_{\text{teeth}}$ encodes $f_{\text{teeth}}$ into multi-scale feature maps which serve as content conditions for the following StyleGAN layers. The output teeth features $f_{\text{teeth}}'$ are transformed inversely and concatenated with the feature planes of $T_{uv}$ . To eliminate the texture flickering of mouth boundary, we further feed $T_{uv}$ into a shallow-UNet-based neural blending module and obtain $T_{uv}'$ . We conduct a series of ablation studies and prove that the proposed teeth synthesis module brings a remarkable improvement on both the animation accuracy and synthesis quality (Sec. 4.2). + +# 3.3. Modeling static components + +The generative texture-rasterized tri-planes manage to model dynamic faces varying expressions and shapes, though, it is challenging to synthesize static parts like diverse haircut, background and upper body which are not included in the FLAME template. To this end, we model these parts by another tri-plane branch, $T_{\text{static}}$ , which is generated by a StyleGAN2 CNN generator $G_{\text{static}}$ sharing the same latent code with $G_{\text{texture}}$ . The plane features of $T_{uv}^{\prime}$ and $T_{\text{static}}$ are blended on each plane by the alpha masks rendered by rasterization. Such design not only benefits the + +modeling of various-styled static components but also enforces their consistency during facial animations. + +# 3.4. Neural rendering + +Given the blended tri-planes, for any point in the 3D space, we project it into each plane and sample the features bi-linearly. Then, the sampled features are aggregated by summation and decoded into volume density $\sigma$ and feature $f$ by a lightweight decoder. Similar to [7,58], the decoder is a single hidden layered multi-layer perceptron (MLP) with softplus activation. The volume rendering is employed to accumulate $\sigma$ and $f$ along the rays cast through each pixel to compute a 2D feature image $I_{f}$ . Similar to [7,27,46], we leverage a 2D super-resolution module $G_{superres}$ to interpret the feature image into RGB image $I_{RGB}$ with higher resolution. The super-resolution module consists of three StyleGAN2 synthesis blocks and the noise input is removed for alleviating texture flickers. In our case, $I_{f}$ and $I_{RGB}$ are set to $64\times 64$ and $512\times 512$ , respectively. + +# 3.5. Deformation-aware discriminator + +To learn the unsupervised 3D representations, we adopt a 2D convolutional discriminator $D$ to critique the renderings. Inspired by [7], we regularize the first three channels of $I_{f}$ as low-resolution RGB image, which is concatenated with $I_{RGB}$ as the input for the discriminator. However, the discrimination with only image input can only ensure that the deformed images are always in a correct distribution instead of matching the expected deformations. Therefore, we make the discriminator aware of the expression and shape under which the generated image are deformed. For the same purpose, GNARF [3] conditions the discriminator on the FLAME parameters by concatenating them as the input to the mapping network. However, we find empirically that such a conditioning method leads to training instability, consistent with [3]. Instead, we re-render the template mesh under the rendered pose to get the synthetic rendering $I_{synthetic}$ and feed it into $D_{\text{dual}}$ along with image pairs. Here, we adopt correspondence images inspired by [35]. Such concatenation encourages the final output to align with the synthetic rendering and learn the expected deformation. + +# 3.6. Training objectives + +During training, we use the non-saturating GAN loss with R1 regularization. Moreover, we adopt the density regularization proposed in EG3D [7]. Therefore, the total learning objective is: + +$$ +\begin{array}{l} \mathcal {L} _ {D _ {\text {d u a l}, G}} = \mathbb {E} _ {z \sim p _ {z}, \epsilon \sim p _ {\epsilon}} [ f (D _ {\text {d u a l}} (G (z, \epsilon))) ] + \\ \mathbb {E} _ {I ^ {r} \sim p _ {I ^ {r}}} [ f (- D _ {\text {d u a l}} (I ^ {r}) + (1) \\ \left. \lambda \| \bigtriangledown D _ {d u a l} \left(I ^ {r}\right) \| ^ {2} \right], \\ \mathcal {L} _ {\text {d e n s i t y}} = \sum_ {x _ {s} \in \mathcal {S}} \| d (x _ {s}) - d (x _ {s} + \epsilon) \| _ {2}, (2) \\ \end{array} +$$ + +$$ +\mathcal {L} _ {\text {t o t a l}} = \mathcal {L} _ {D _ {\text {d u a l}}, G} + \lambda_ {\text {d e n s i t y}} \mathcal {L} _ {\text {d e n s i t y}}, \tag {3} +$$ + +where $I^r$ is the combination of real images, blurred real images, and the corresponding synthetic renderings, which are sampled from the training set with distribution $p_I$ . We adopt many training hyperparameters from EG3D and StyleGAN2 (learning rates of generator and discriminator, batch size, R1 regularization, etc.). We train our model based on the pretrained model of EG3D [7] and continue to train on 4 3090 GPUs for roughly 4 days. Please refer to the supplemental material for the implementation details. + +# 4. Experiments + +In this section, we first show qualitative and quantitative comparisons to state-of-the-art 2D / 3D animatable generative facial models (Sec. 4.1), and then discuss the conducted ablation studies of our design choices (Sec. 4.2). Furthermore, we show various applications combining our efficient 3D representation with GAN technologies such as GAN inversion and style transfer (Sec. 4.3). + +Datasets. We train and test our methods on FFHQ [33]. We augment FFHQ with horizontal flips and use an off-the-shelf pose estimator [15] to label images with the approximated camera extrinsic parameters and constant intrinsics. To support full pose animation, in-plane (roll) rotation is also considered. Furthermore, we use DECA [17] to estimate the FLAME parameters of facial identity $\beta \in \mathbb{R}^{100}$ , jaw pose $\theta_{jaw} \in \mathbb{R}^3$ and expression $\psi \in \mathbb{R}^{50}$ . Since DECA doesn't account for eyeball movement, we additionally adopt an efficient facial detector to detect 2D landmarks of irises and optimize eye poses $\theta_{eye} \in \mathbb{R}^6$ by minimizing the re-projection errors. Based on these FLAME parameters, we produce a template mesh with 5023 vertices and 9976 faces to drive facial deformations. + +# 4.1. Comparisons + +Baselines. We compare our method against two state-of-the-art methods for animatable 3D-aware image synthesis: 3DFaceShop [61], and AniFaceGAN [68]. Besides, we also select DiscoFaceGAN [13] as a baseline, which generates animatable 2D portraits conditioned on 3DMM parameters. Qualitative comparison. Fig. 4 provides a qualitative comparison against baselines. Overall, as can be seen, our method outperforms all baselines by some margin on both synthesis quality and animation accuracy. Specifically, DiscoFaceGAN [13] suffers from inconsistent identity during animation. Moreover, it cannot generate reasonable mouth interior, e.g. stretched teeth. 3DFaceshop and AnifaceGAN synthesize 3D-consistent images, nevertheless, still struggle to model consistent mouth interior with the driving images. This is because their implicit deformation approaches + +![](images/ae3815df2f0a2a31171a1005563d372e24693015e2a3b8ec4c1162c9344f8af6.jpg) +Driving +Free View + +![](images/c3608cd3755d3e5d9deab540e36e52453e0462f88afbea3b145385dae3875be7.jpg) +Ours + +![](images/b182aff722264e93d8d2b436eda799d715f52ba236c2938ccf8a6c4ac65f3e9d.jpg) +AntiFaceGAN + +![](images/bfa21a81a77e31ee6ec3f66b0aecf1efb8cb7235016333588d1752269d2aa997.jpg) +3DFaceShop + +![](images/c1cc234d65e8b8316cda7fa7ddc34a56ca315f319037992a7fd82a3751259dca.jpg) +DiscoFaceGAN + +![](images/d0db2f87db9803a0da906aa002e9b16abc4583b704cc5bf121185b5bdce93ba5.jpg) +Figure 4. Comparison with the state-of-the-art animatable 3D & 2D image synthesis methods. We extract several frames from a video clip and use the interpreted face model parameters to animate random virtual avatars. + +![](images/ec97242d1e2a36766e4e1f9e0ef73896b9644c04e13ea18d6003add53b00cf63.jpg) + +![](images/491c44e32c6fe6b05d1ec0d69a8d1361306a65d270634e0d87a13c30bcac7310.jpg) + +![](images/6b50011c5c4550f1a80fe1cc2524b78f9d08ce245533e4a57c4857ccd52973e4.jpg) + +![](images/747fabddb71b00197f5cde37368f35b8e1132323cc3564f89b80afa35ca2b01e.jpg) + +![](images/8d34245a79b281cadd16e8759c45d5da48d5c693dae5e77dbaa2cfccc22ff6ed.jpg) + +![](images/a89aef0716d09913f5dfbcc52fe9fcb23ab97eca1bcd2e296b381b9514969887.jpg) + +![](images/d669a0a9d1c38bc3638dff1a521df2e5a05963cf25436f60a8a981d0686de340.jpg) + +![](images/08c3f15f9b04411dead30dee5f32042e70407fdc0a968c30f4bf75a36089db44.jpg) +Figure 5. High-quality dynamic shapes with topological changes. As can be seen, we model detailed dynamic shapes of eyelids and lips, while keeping glasses unchanged. + +![](images/b65e08731f9e60d2cd05d296b57f91eafea40bdbe8789a17e10d50ba13ee3190.jpg) + +![](images/3df380a8b1750ce5e9429c69f72d2b4f418940a89ae3904bbf0db446acdb7901.jpg) + +![](images/83691ef05468f4cc561e3309aed50b804bdf082ea5f96fdd68a623bb8b09c2cc.jpg) + +are under-constrained, leading to overfit the expression bias (smiling with mouth half-opened) of datasets. Compared to the other methods, our approach not only synthesizes im + +
FID↓AED↓APD↓APD*↓ID↑
DiscoFaceGAN (2562) [13]17.10.420.0460.0240.73
AniFaceGAN (2562) [68]20.10.250.0410.0220.82
3DFaceShop (5122) [61]23.70.310.0450.0240.75
Ours (2562)3.70.160.0180.0130.85
Ours (5122)3.90.160.0230.0190.84
+ +Table 1. Quantitative comparison using FID, average expression distance (AED), average pose distance (APD), and identity consistency (ID) for FFHQ. APD* means calculating pose distance with roll fixed. + +ages with higher quality, but also preserves more detailed expressions of the driver images, including mouth interior, eye blinks and eye movements. Furthermore, we are the only method that supports in-plane head rotations. Fig. 5 provides visual examples of the synthesized high-quality geometry. Our approach can model detailed shape deformations (see zoomed eyelids and lips in Fig. 5) with topological awareness, i.e. glasses are kept unchanged. + +![](images/5eaaba2d514d716f07aed2262464eea28519aee93db48132f33c0c3f89c9d905.jpg) +Figure 6. Ablation study on model designs. $T_{\text{static}}$ encourages better identity consistency; $G_{\text{teeth}}$ benefits realistic mouth interior; the discriminator with $I_{\text{synthetic}}$ input allows more consistent reconstruction of detailed expressions. + +Quantitative evaluation. Tab. 1 demonstrates quantitative results comparing our method against baselines evaluated on several metrics. We measure image quality with Frechet Inception Distance (FID) [29] between the entire FFHQ dataset and $50\mathrm{k}$ generated images using randomly sampled latent codes, camera poses, and FLAME parameters. Since AntiFaceGAN and DiscoFaceGAN synthesize images on the resolution of $256^2$ , we test FID of them on FFHQ256². Following [3, 39], we evaluate the faithfulness of the animation with the Average Expression Distance (AED), the Average Pose Distance (APD), and identity consistency (ID). For each method, we randomly sample 500 identities and animate each with randomly sampled 20 FLAME parameters of expressions and poses. Then, we estimate the FLAME parameters for these 10000 generated images and the average distances between the driving FLAME parameters and the reconstructed ones. For identity consistency, we randomly sample 2000 poses, 2000 sets + +
FID↓AED↓APD↓ID↑
w/o Tstatic6.60.250.0260.63
w/o Gteeth(uv)7.20.370.0420.71
w/o Gteeth(s)8.40.320.0360.79
w/o Isynthetic3.80.180.0250.74
Full3.90.160.0230.84
+ +Table 2. Ablation study on model designs. Modeling static components by $T_{\text{static}}$ improves identity consistency. The teeth synthesis $G_{\text{teeth}}$ module benefits both animation and synthesis quality significantly while adding synthetic renderings $I_{\text{synthetic}}$ into discrimination takes both a step further. + +of FLAME parameters, and 1000 identities. Then we randomly select two poses and two sets of FLAME parameters for each identity, generating a total of 1000 image pairs. We calculate consistency metric using a pre-trained Arcface model [15] for each image pair and report the average result. Since the other baselines don't support in-plane (roll) head rotation, we further report APD* which only accounts for poses on yaw and pitch. Our method achieves the best performance on all metrics. Note that our model demonstrates significant improvements in FID, bringing animatable 3D GAN to the same level as unconditional 3D GANs (4.7 for EG3D [7]). For AED and APD, we also show superiority against baselines. Note that we still achieve the best pose consistency (0.019) when only considering yaw and pitch. + +# 4.2. Ablation study + +Static tri-planes. As suggested by the grey lines in Fig. 8, this baseline removes the static tri-planes $T_{static}$ and entangles both dynamic and static components in $T_{uv}$ . As illustrated in the first row of Fig. 6, the identities change when varying expressions. Since there are no explicit constraints for identity consistency, the model would be prone to unexpected entanglement between expression and identity. Tab. 2 shows a similar trend where removing $T_{static}$ leads to worse identity consistency. + +Mouth Synthesis. When removing the mouth synthesis module $G_{\text{teeth}}$ , we consider two altered choices: representing mouth features by $T_{\text{static}}$ or $T_{uv}$ , named w/o $G_{\text{teeth}}(s)$ (red lines in Fig. 8) and w/o $G_{\text{teeth}}(uv)$ (blue lines in Fig. 8), respectively. The first baseline, illustrated in the second row of Fig. 6, suffers from 'hole' artifacts of mouth. This is because inferior teeth move along with the jaw rotations and thus cannot be modeled by static features. The second baseline modeling teeth with $T_{uv}$ also leads to an unreasonable mouth interior since the FLAME template doesn't account for teeth area and the neural textures for teeth would be sampled from other unrelated areas leading to artifacts. Quantitatively, both baselines without $G_{\text{teeth}}$ show significant degradation in AED and APD. + +Deformation-aware discriminator. Tab. 2 demonstrates + +![](images/33d78dd48624762e1e3900456db4959f77cff16ac5f4f4d46f8d0fdbb74e7458.jpg) +Source + +![](images/06f12f87eb4149f216329656db72dcbc8a7d140cf3dfb717936ca9d15eb74d18.jpg) +Inversion +Animation + +![](images/b835e071ba2f83a9cef41d244d51c1b9d4c73d4555dc887a9c0dc7e946a229e5.jpg) + +![](images/51195a567cff7979293e60e15960e03e8e69678b96fc431a268c2f0b97dd8dcb.jpg) +Figure 7. Applications of our model. We use PTI [53] to fit 3D-aware avatars for real portraits and animate them with sampled video clips. Furthermore, we leverage StyleGAN-NADA [21] into 3D settings and adapt these avatars into textually-prescribed domains. + +![](images/e83a8e6a044c63386ffa5984d7aafe634c56ff2bbf9f3b7b7947288886843380.jpg) + +![](images/bf09cbe0ec110d786ce9195b4a62a8ee47f138e113020020216865a347b56f3d.jpg) + +![](images/1f35502efc539fa94056ed3f81f193815f4d51436468456966ac2b8da92d4bbc.jpg) +(a) Default Structure + +![](images/1ec1bb7919d793fe0c6701920479a6b6ccf854feba233e084144bb915b512e1a.jpg) +(b) Altered Structures +Figure 8. Illustrations of the proposed three variants of our model for ablation study. + +that the deformation-aware discriminator with $I_{\text{synthetic}}$ input improves both animation accuracy and identity consistency, at the negligible expense of slightly reduced image quality. In Fig. 6, we see that this design shows a better reconstruction of detailed expressions, e.g. eye blinks. + +# 4.3. Applications + +One-shot portrait animation. Fig. 7 shows the application of our model for one-shot head avatars. The learned generative animatbale 3D representation with expressive latent space can serve as a strong 3D prior for high-fidelity single-view 3D reconstruction and animation. Note that we can generate natural and consistent animations without video data training. + +Animatable 3D-aware stylization. Inspired by IDE-3D [58], we incorporate 2D CLIP-guided style tranfer methods [21] with our animatable 3D representation for 3D-aware portrait stylization. The right three columns of Fig. 7 show examples of text-driven, stylized portrait animation. Specifically, to adapt a pre-trained model through + +only a textual prompt, we optimize the generator with two kinds of CLIP-based guidance [21]. However, leveraging text-guided 2D methods directly into the 3D setting is challenging as it tends to break 3D awareness and deformation awareness inherited in the generator parameters. To this end, we make some necessary modifications to the training framework. Please refer to the supplemental material for details. As shown in Fig. 7, we achieve high-quality 3D-aware stylized portrait synthesis with preserving well properties (i.e. 3D consistency and accurate animation). + +# 5. Limitations and future work + +Though our approach enables reasonable extrapolation on some rare expressions, it struggles to model some others with full consistency, such as one-side mouth up, sticking tongue out, etc. Besides, the rendered portraits still suffer from slight texture flickering caused by 2D superresolution. We leave both for future work. Furthermore, our model has the potential to provide a strong 3D prior to accelerating person-specific avatar reconstruction. + +# 6. Conclusion + +We have presented Next3D, a novel animatable 3D representation for unsupervised learning of high-quality and 3D-consistent virtual facial avatars from unstructured 2D images. Our approach has pushed the frontier of photorealistic animatabale 3D-aware image synthesis. Serving as a strong 3D prior, we believe our learned 3D representation will boost a series of downstream applications including 3D-aware one-shot facial avatars and animatable out-of-domain avatar generation. + +# Acknowledgements + +This paper is supported by National Key RD Program of China (2022YFF0902200) and the NSFC project No.62125107. + +# References + +[1] ShahRukh Athar, Zexiang Xu, Kalyan Sunkavalli, Eli Shechtman, and Zhixin Shu. Rignerf: Fully controllable neural 3d portraits. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20364-20373, 2022. 3 +[2] Hendrik Baatz, Jonathan Granskog, Marios Papas, Fabrice Rousselle, and Jan Novák. Nerf-tex: Neural reflectance field textures. In Computer Graphics Forum. Wiley Online Library, 2021. 3 +[3] Alexander W. Bergman, Petr Kellnhofer, Wang Yifan, Eric R. Chan, David B. Lindell, and Gordon Wetzstein. Generative neural articulated radiance fields. In NeurIPS, 2022. 2, 5, 7 +[4] Volker Blanz and Thomas Vetter. A morphable model for the synthesis of 3d faces. In Proceedings of the 26th Annual Conference on Computer Graphics and Interactive Techniques, SIGGRAPH '99, page 187-194, USA, 1999. ACM Press/Addison-Wesley Publishing Co. 2, 3 +[5] Andrei Burov, Matthias Nießner, and Justus Thies. Dynamic surface function networks for clothed human bodies. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10754-10764, 2021. 3 +[6] Shengqu Cai, Anton Obukhov, Dengxin Dai, and Luc Van Gool. Pix2nerf: Unsupervised conditional p-gan for single image to neural radiance fields translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3981-3990, 2022. 2 +[7] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In Proceedings of the IEEE International Conference on Computer Vision (CVPR), 2021. 2, 4, 5, 7 +[8] Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2 +[9] Xu Chen, Tianjian Jiang, Jie Song, Jinlong Yang, Michael J Black, Andreas Geiger, and Otmar Hilliges. gdna: Towards generative detailed neural avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20427-20437, 2022. 2 +[10] Yuedong Chen, Qianyi Wu, Chuanxia Zheng, Tat-Jen Cham, and Jianfei Cai. Sem2nerf: Converting single-view semantic masks to neural radiance fields. In ECCV, 2022. 2 +[11] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5939-5948, 2019. 3 +[12] Hang Dai, Nick Pears, William Smith, and Christian Duncan. Statistical modeling of craniofacial shape and texture. International Journal of Computer Vision, 128(2):547-571, 2020. 3 + +[13] Yu Deng, Jiaolong Yang, Dong Chen, Fang Wen, and Tong Xin. Disentangled and controllable face image generation via 3d imitative-contrastive learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2, 3, 4, 5, 6 +[14] Yu Deng, Jiaolong Yang, Jianfeng Xiang, and Xin Tong. Gram: Generative radiance manifolds for 3d-aware image generation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2 +[15] Yu Deng, Jiaolong Yang, Sicheng Xu, Dong Chen, Yunde Jia, and Xin Tong. Accurate 3d face reconstruction with weakly-supervised learning: From single image to image set. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 0–0, 2019. 5, 7 +[16] Michail Christos Doukas, Stefanos Zafeiriou, and Viktoriia Sharmanska. Headgan: One-shot neural head synthesis and editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14398-14407, 2021. 2, 3 +[17] Yao Feng, Haiwen Feng, Michael J Black, and Timo Bolkart. Learning an animatable detailed 3d face model from inthe-wild images. ACM Transactions on Graphics (ToG), 40(4):1-13, 2021. 5 +[18] Ohad Fried, Ayush Tewari, Michael Zollhöfer, Adam Finkelstein, Eli Shechtman, Dan B Goldman, Kyle Genova, Zeyu Jin, Christian Theobalt, and Maneesh Agrawala. Text-based editing of talking-head video. ACM Transactions on Graphics (TOG), 38(4):1-14, 2019. 3 +[19] Matheus Gadelha, Subhransu Maji, and Rui Wang. 3d shape induction from 2d views of multiple objects. In International Conference on 3D Vision (3DV), 2017. 2 +[20] Guy Gafni, Justus Thies, Michael Zollhofer, and Matthias Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8649-8658, 2021. 3 +[21] Rinon Gal, Or Patashnik, Haggai Maron, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Stylegan-nada: Clip-guided domain adaptation of image generators. ACM Transactions on Graphics (TOG), 41(4):1-13, 2022. 8 +[22] Baris Gecer, Binod Bhattacharai, Josef Kittler, and Tae-Kyun Kim. Semi-supervised adversarial learning to generate photorealistic face images of new identities from 3d morphable model. In Proceedings of the European conference on computer vision (ECCV), pages 217-234, 2018. 3 +[23] Baris Gecer, Stylianos Ploumpis, Irene Kotsia, and Stefanos Zafeiriou. Ganfit: Generative adversarial network fitting for high fidelity 3d face reconstruction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1155-1164, 2019. 3 +[24] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 2 +[25] Philip-William Grassal, Malte Prinzler, Titus Leistner, Carsten Rother, Matthias Nießner, and Justus Thies. Neural + +head avatars from monocular rgb videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18653-18664, 2022. 3 +[26] Artur Grigorev, Karim Iskakov, Anastasia Ianina, Renat Bashirov, Ilya Zakharkin, Alexander Vakhitov, and Victor Lempitsky. Stylepeople: A generative model of fullbody human avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5151-5160, 2021. 3 +[27] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. Stylenerf: A style-based 3d aware generator for high-resolution image synthesis. In International Conference on Learning Representations, 2022. 2, 5 +[28] Philipp Henzler, Niloy J Mitra, and Tobias Ritschel. Escaping plato's cave: 3d shape from adversarial rendering. In Proceedings of the IEEE International Conference on Computer Vision (CVPR), 2019. 2 +[29] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In Advances in Neural Information Processing Systems (NeurIPS), 2017. 7 +[30] Fangzhou Hong, Zhaoxi Chen, Yushi LAN, Liang Pan, and Ziwei Liu. EVA3d: Compositional 3d human generation from 2d image collections. In International Conference on Learning Representations, 2023. 2 +[31] Yang Hong, Bo Peng, Haiyao Xiao, Ligang Liu, and Juyong Zhang. Headnerf: A real-time nerf-based parametric head model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20374-20384, 2022. 3 +[32] Angjoo Kanazawa, Shubham Tulsiani, Alexei A Efros, and Jitendra Malik. Learning category-specific mesh reconstruction from image collections. In Proceedings of the European Conference on Computer Vision (ECCV), pages 371-386, 2018. 3 +[33] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4401-4410, 2019. 5 +[34] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2020. 4 +[35] Hyeongwoo Kim, Pablo Garrido, Ayush Tewari, Weipeng Xu, Justus Thies, Matthias Niessner, Patrick Pérez, Christian Richardt, Michael Zollhöfer, and Christian Theobalt. Deep video portraits. ACM Transactions on Graphics (TOG), 37(4):1-14, 2018. 2, 3, 5 +[36] Jeong-gi Kwak, Yuanming Li, Dongsik Yoon, Donghyeon Kim, David Han, and Hanseok Ko. Injecting 3d perception of controllable nerf-gan into stylegan for editable portrait image synthesis. In Proceedings of European Conference on Computer Vision (ECCV), 2022. 2 +[37] Henning Lange and J Nathan Kutz. Fc2t2: The fast continuous convolutional taylor transform with applications in vi + +sion and graphics. arXiv preprint arXiv:2111.00110, 2021. 3 +[38] Tianye Li, Timo Bolkart, Michael J Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4d scans. ACM Trans. Graph., 36(6):194-1, 2017. 3, 4 +[39] Connor Z Lin, David B Lindell, Eric R Chan, and Gordon Wetzstein. 3d gan inversion for controllable portrait image animation. arXiv preprint arXiv:2203.13441, 2022. 2, 7 +[40] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European Conference on Computer Vision (ECCV), 2020. 3 +[41] Thu Nguyen-Phuoc, Chuan Li, Lucas Theis, Christian Richardt, and Yong-Liang Yang. Hologan: Unsupervised learning of 3d representations from natural images. In Proceedings of the IEEE International Conference on Computer Vision, 2019. 2 +[42] Thu H Nguyen-Phuoc, Christian Richardt, Long Mai, Yongliang Yang, and Niloy Mitra. Blockgan: Learning 3d object-aware scene representations from unlabelled images. In Advances in Neural Information Processing Systems (NeurIPS), 2020. 2 +[43] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2 +[44] Atsuhiro Noguchi, Xiao Sun, Stephen Lin, and Tatsuya Harada. Unsupervised learning of efficient geometry-aware neural articulated representations. In Proceedings of European Conference on Computer Vision (ECCV), 2022. 2 +[45] Michael Oechsle, Lars Mescheder, Michael Niemeyer, Thilo Strauss, and Andreas Geiger. Texture fields: Learning texture representations in function space. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4531-4540, 2019. 3 +[46] Roy Or-El, Xuan Luo, Mengyi Shan, Eli Shechtman, Jeong Joon Park, and Ira Kemelmacher-Shlizerman. StyleSDF: High-Resolution 3D-Consistent Image and Geometry Generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 5 +[47] Xingang Pan, Xudong Xu, Chen Change Loy, Christian Theobalt, and Bo Dai. A shading-guided generative implicit model for shape-accurate 3d-aware image synthesis. Advances in Neural Information Processing Systems, 34:20002-20013, 2021. 2 +[48] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3 +[49] Songyou Peng, Michael Niemeyer, Lars Mescheder, Marc Pollefeys, and Andreas Geiger. Convolutional occupancy networks. In European Conference on Computer Vision, pages 523-540. Springer, 2020. 3 + +[50] Hanspeter Pfister, Matthias Zwicker, Jeroen Van Baar, and Markus Gross. Surfels: Surface elements as rendering primitives. In Proceedings of the 27th annual conference on Computer graphics and interactive techniques, pages 335-342, 2000. 3 +[51] Stylianos Ploumpis, Evangelos Ververas, Eimear O'Sullivan, Stylianos Moschoglou, Haoyang Wang, Nick Pears, William AP Smith, Baris Gecer, and Stefanos Zafeiriou. Towards a complete 3d morphable model of the human head. IEEE transactions on pattern analysis and machine intelligence, 43(11):4142-4160, 2020. 3 +[52] Yurui Ren, Ge Li, Yuanqi Chen, Thomas H Li, and Shan Liu. Pirenderer: Controllable portrait image generation via semantic neural rendering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13759-13768, 2021. 2, 3 +[53] Daniel Roich, Ron Mokady, Amit H Bermano, and Daniel Cohen-Or. Pivotal tuning for latent-based editing of real images. ACM Trans. Graph., 2021. 8 +[54] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. Graf: Generative radiance fields for 3d-aware image synthesis. In Advances in Neural Information Processing Systems (NeurIPS), 2020. 2 +[55] Katja Schwarz, Axel Sauer, Michael Niemeyer, Yiyi Liao, and Andreas Geiger. Voxgraf: Fast 3d-aware image synthesis with sparse voxel grids. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 2 +[56] Vincent Sitzmann, Justus Thies, Felix Heide, Matthias Nießner, Gordon Wetzstein, and Michael Zollhofer. Deepvoxels: Learning persistent 3d feature embeddings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2437-2446, 2019. 3 +[57] Ivan Skorokhodov, Sergey Tulyakov, Yiqun Wang, and Peter Wonka. Epigraph: Rethinking training of 3d gans. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 2 +[58] Jingxiang Sun, Xuan Wang, Yichun Shi, Lizhen Wang, Jue Wang, and Yebin Liu. Ide-3d: Interactive disentangled editing for high-resolution 3d-aware portrait synthesis. ACM Transactions on Graphics (TOG), 41(6):1-10, 2022. 2, 5, 8 +[59] Jingxiang Sun, Xuan Wang, Yong Zhang, Xiaoyu Li, Qi Zhang, Yebin Liu, and Jue Wang. Fenerf: Face editing in neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2 +[60] Keqiang Sun, Shangzhe Wu, Zhaoyang Huang, Ning Zhang, Quan Wang, and HongSheng Li. Controllable 3d face synthesis with conditional generative occupancy fields. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 2 +[61] Junshu Tang, Bo Zhang, Binxin Yang, Ting Zhang, Dong Chen, Lizhuang Ma, and Fang Wen. Explicitly controllable 3d-aware portrait generation. arXiv preprint arXiv:2209.05434, 2022. 2, 4, 5, 6 +[62] Ayush Tewari, Mohamed Elgharib, Gaurav Bharaj, Florian Bernard, Hans-Peter Seidel, Patrick Pérez, Michael Zollhofer, and Christian Theobalt. Stylerig: Rigging stylegan + +for 3d control over portrait images. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2, 3 +[63] Ayush Tewari, Justus Thies, Ben Mildenhall, Pratul Srinivasan, Edgar Tretschk, W Yifan, Christoph Lassner, Vincent Sitzmann, Ricardo Martin-Brualla, Stephen Lombardi, et al. Advances in neural rendering. Computer Graphics Forum, 41(2):703-735, 2022. 3 +[64] Justus Thies, Michael Zollhöfer, and Matthias Nießner. Deferred neural rendering: Image synthesis using neural textures. ACM Transactions on Graphics (TOG), 38(4):1-12, 2019. 3, 4 +[65] Justus Thies, Michael Zollhofer, Marc Stamminger, Christian Theobalt, and Matthias Nießner. Face2face: Real-time face capture and reenactment of rgb videos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2387-2395, 2016. 2, 3 +[66] Lizhen Wang, Zhiyuan Chen, Tao Yu, Chenguang Ma, Liang Li, and Yebin Liu. Faceverse: a fine-grained and detail-controllable 3d face morphable model from a hybrid dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20333-20342, 2022. 4 +[67] Jiajun Wu, Chengkai Zhang, Tianfan Xue, Bill Freeman, and Josh Tenenbaum. Learning a probabilistic latent space of object shapes via 3d generative-adversarial modeling. Advances in neural information processing systems, 29, 2016. 2 +[68] Yue Wu, Yu Deng, Jiaolong Yang, Fangyun Wei, Chen Qifeng, and Xin Tong. Anifacegan: Animatable 3d-aware face image generation for video avatars. In Advances in Neural Information Processing Systems, 2022. 2, 4, 5, 6 +[69] Qiangeng Xu, Weiyue Wang, Duygu Ceylan, Radomir Mech, and Ulrich Neumann. Disn: Deep implicit surface network for high-quality single-view 3d reconstruction. Advances in Neural Information Processing Systems, 32, 2019. 3 +[70] Sicheng Xu, Jiaolong Yang, Dong Chen, Fang Wen, Yu Deng, Yunde Jia, and Xin Tong. Deep 3d portrait from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7710-7720, 2020. 2, 3 +[71] Yinghao Xu, Sida Peng, Ceyuan Yang, Yujun Shen, and Bolei Zhou. 3d-aware image synthesis via learning structural and textural representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18430-18439, 2022. 2 +[72] Yang Xue, Yuheng Li, Krishna Kumar Singh, and Yong Jae Lee. Giraffe hd: A high-resolution 3d-aware generative model. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2 +[73] Fei Yin, Yong Zhang, Xiaodong Cun, Mingdeng Cao, Yanbo Fan, Xuan Wang, Qingyan Bai, Baoyuan Wu, Jue Wang, and Yujiu Yang. Styleheat: One-shot high-resolution editable talking face generation via pretrained stylegan. In Proceedings of the European conference on computer vision (ECCV), 2022. 2 + +[74] Jianfeng Zhang, Zihang Jiang, Dingdong Yang, Hongyi Xu, Yichun Shi, Guoxian Song, Zhongcong Xu, Xinchao Wang, and Jiashi Feng. Avatargen: a 3d generative model for animatable human avatars. arXiv preprint arXiv:2208.00561, 2022. 2 +[75] Jingbo Zhang, Xiaoyu Li, Ziyu Wan, Can Wang, and Jing Liao. Fdnerf: Few-shot dynamic neural radiance fields for face reconstruction and expression editing. arXiv preprint arXiv:2208.05751, 2022.3 +[76] Jichao Zhang, E. Sangineto, Hao Tang, Aliaksandr Siarohin, Zhun Zhong, N. Sebe, and Wei Wang. 3d-aware semantic-guided generative model for human synthesis. In ECCV, 2022. 2 +[77] Jichao Zhang, Aliaksandr Siarohin, Yahui Liu, Hao Tang, Nicu Sebe, and Wei Wang. Training and tuning generative neural radiance fields for attribute-conditional 3d-aware face generation. arXiv preprint arXiv:2208.12550, 2022. 2 +[78] Xuanmeng Zhang, Zhedong Zheng, Daiheng Gao, Bang Zhang, Pan Pan, and Yi Yang. Multi-view consistent generative adversarial networks for 3d-aware image synthesis. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2 +[79] Xiaoming Zhao, Fangchang Ma, David Guera, Zhile Ren, Alexander G. Schwing, and Alex Colburn. Generative multiplane images: Making a 2d gan 3d-aware. In Proceedings of European Conference on Computer Vision (ECCV), 2022. 2 +[80] Yufeng Zheng, Victoria Fernández Abrevaya, Marcel C Bühler, Xu Chen, Michael J Black, and Otmar Hilliges. Im avatar: Implicit morphable head avatars from videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13545-13555, 2022. 3 +[81] Peng Zhou, Lingxi Xie, Bingbing Ni, and Qi Tian. Cips-3d: A 3d-aware generator of gans based on conditionally-independent pixel synthesis. arXiv:2110.09788, 2021. 2 +[82] Jun-Yan Zhu, Zhoutong Zhang, Chengkai Zhang, Jiajun Wu, Antonio Torralba, Josh Tenenbaum, and Bill Freeman. Visual object networks: Image generation with disentangled 3d representations. In Advances in Neural Information Processing Systems (NeurIPS), 2018. 2 +[83] Peiye Zhuang, Liqian Ma, Oluwasanmi Koyejo, and Alexander Schwing. Controllable radiance fields for dynamic face synthesis. In Proc. 3DV, 2022. 2 \ No newline at end of file diff --git a/2023/Next3D_ Generative Neural Texture Rasterization for 3D-Aware Head Avatars/images.zip b/2023/Next3D_ Generative Neural Texture Rasterization for 3D-Aware Head Avatars/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..0355637bb7c0193e1b1e297c1961171bbcd76eff --- /dev/null +++ b/2023/Next3D_ Generative Neural Texture Rasterization for 3D-Aware Head Avatars/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7b9a9d1f20c0fe1757ff33ad5f5447040b0108780a66a29bf169574a59e1fe25 +size 735915 diff --git a/2023/Next3D_ Generative Neural Texture Rasterization for 3D-Aware Head Avatars/layout.json b/2023/Next3D_ Generative Neural Texture Rasterization for 3D-Aware Head Avatars/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3831054591191a2cf53b914be55587ae7a2dd8c1 --- /dev/null +++ b/2023/Next3D_ Generative Neural Texture Rasterization for 3D-Aware Head Avatars/layout.json @@ -0,0 +1,9845 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 56, + 103, + 538, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 103, + 538, + 120 + ], + "spans": [ + { + "bbox": [ + 56, + 103, + 538, + 120 + ], + "type": "text", + "content": "Next3D: Generative Neural Texture Rasterization for 3D-Aware Head Avatars" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 77, + 141, + 153, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 141, + 153, + 159 + ], + "spans": [ + { + "bbox": [ + 77, + 141, + 153, + 159 + ], + "type": "text", + "content": "Jingxiang Sun" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 177, + 142, + 241, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 177, + 142, + 241, + 158 + ], + "spans": [ + { + "bbox": [ + 177, + 142, + 241, + 158 + ], + "type": "text", + "content": "Xuan Wang" + }, + { + "bbox": [ + 177, + 142, + 241, + 158 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 265, + 143, + 342, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 265, + 143, + 342, + 158 + ], + "spans": [ + { + "bbox": [ + 265, + 143, + 342, + 158 + ], + "type": "text", + "content": "Lizhen Wang" + }, + { + "bbox": [ + 265, + 143, + 342, + 158 + ], + "type": "inline_equation", + "content": "^{14}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 366, + 143, + 422, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 366, + 143, + 422, + 158 + ], + "spans": [ + { + "bbox": [ + 366, + 143, + 422, + 158 + ], + "type": "text", + "content": "Xiaoyu Li3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 446, + 143, + 512, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 446, + 143, + 512, + 158 + ], + "spans": [ + { + "bbox": [ + 446, + 143, + 512, + 158 + ], + "type": "text", + "content": "Yong Zhang3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 212, + 159, + 298, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 159, + 298, + 174 + ], + "spans": [ + { + "bbox": [ + 212, + 159, + 298, + 174 + ], + "type": "text", + "content": "Hongwen Zhang" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 324, + 159, + 377, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 159, + 377, + 173 + ], + "spans": [ + { + "bbox": [ + 324, + 159, + 377, + 173 + ], + "type": "text", + "content": "Yebin Liu" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 92, + 177, + 197, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 177, + 197, + 193 + ], + "spans": [ + { + "bbox": [ + 92, + 177, + 197, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 92, + 177, + 197, + 193 + ], + "type": "text", + "content": "Tsinghua University" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 221, + 178, + 280, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 221, + 178, + 280, + 193 + ], + "spans": [ + { + "bbox": [ + 221, + 178, + 280, + 193 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 221, + 178, + 280, + 193 + ], + "type": "text", + "content": "Ant Group" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 178, + 386, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 178, + 386, + 191 + ], + "spans": [ + { + "bbox": [ + 304, + 178, + 386, + 191 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 304, + 178, + 386, + 191 + ], + "type": "text", + "content": "Tencent AI Lab" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 411, + 178, + 474, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 411, + 178, + 474, + 191 + ], + "spans": [ + { + "bbox": [ + 411, + 178, + 474, + 191 + ], + "type": "text", + "content": "4NNKosmos" + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 48, + 219, + 547, + 392 + ], + "blocks": [ + { + "bbox": [ + 48, + 219, + 547, + 392 + ], + "lines": [ + { + "bbox": [ + 48, + 219, + 547, + 392 + ], + "spans": [ + { + "bbox": [ + 48, + 219, + 547, + 392 + ], + "type": "image", + "image_path": "c04ab9bd231f8e61758d2d3d2122949aed22cecb230f0433f950fe327204b0b5.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 146, + 392, + 209, + 402 + ], + "lines": [ + { + "bbox": [ + 146, + 392, + 209, + 402 + ], + "spans": [ + { + "bbox": [ + 146, + 392, + 209, + 402 + ], + "type": "text", + "content": "Gaze Animation" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 361, + 392, + 479, + 403 + ], + "lines": [ + { + "bbox": [ + 361, + 392, + 479, + 403 + ], + "spans": [ + { + "bbox": [ + 361, + 392, + 479, + 403 + ], + "type": "text", + "content": "3D-Aware Domain Adaptation" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 45, + 417, + 547, + 462 + ], + "lines": [ + { + "bbox": [ + 45, + 417, + 547, + 462 + ], + "spans": [ + { + "bbox": [ + 45, + 417, + 547, + 462 + ], + "type": "text", + "content": "Figure 1. Our 3D GAN synthesizes generative, high-quality, and 3D-consistent facial avatars from unstructured 2D images. Unlike current animatable 3D GANs that only modify yaw-pitch head poses and facial expressions, our approach enables fine-grained control over full-head rotations, facial expressions, eye blinks, and gaze directions with strict 3D consistency and a high level of photorealism. Our approach also provides strong 3D priors for downstream tasks such as 3D-aware stylization." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 143, + 472, + 192, + 484 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 472, + 192, + 484 + ], + "spans": [ + { + "bbox": [ + 143, + 472, + 192, + 484 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 45, + 498, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 498, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 45, + 498, + 289, + 715 + ], + "type": "text", + "content": "3D-aware generative adversarial networks (GANs) synthesize high-fidelity and multi-view-consistent facial images using only collections of single-view 2D imagery. Towards fine-grained control over facial attributes, recent efforts incorporate 3D Morphable Face Model (3DMM) to describe deformation in generative radiance fields either explicitly or implicitly. Explicit methods provide fine-grained expression control but cannot handle topological changes caused by hair and accessories, while implicit ones can model varied topologies but have limited generalization caused by the unconstrained deformation fields. We propose a novel 3D GAN framework for unsupervised learning of generative, high-quality and 3D-consistent facial avatars from unstructured 2D images. To achieve both deformation accuracy and topological flexibility, we propose a 3D representation called Generative Texture-Rasterized Tri-planes. The proposed representation learns Generative Neural Textures on top of parametric mesh templates" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 473, + 547, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 473, + 547, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 473, + 547, + 641 + ], + "type": "text", + "content": "and then projects them into three orthogonal-viewed feature planes through rasterization, forming a tri-plane feature representation for volume rendering. In this way, we combine both fine-grained expression control of mesh-guided explicit deformation and the flexibility of implicit volumetric representation. We further propose specific modules for modeling mouth interior which is not taken into account by 3DMM. Our method demonstrates state-of-the-art 3D-aware synthesis quality and animation ability through extensive experiments. Furthermore, serving as 3D prior, our animatable 3D representation boosts multiple applications including one-shot facial avatars and 3D-aware stylization. Project page: https://mrtornado24.github.io/Next3D/. Code: https://github.com/MrTornado24/Next3D." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 668, + 387, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 668, + 387, + 681 + ], + "spans": [ + { + "bbox": [ + 306, + 668, + 387, + 681 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 305, + 689, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 546, + 715 + ], + "type": "text", + "content": "Animatable portrait synthesis is essential for movie postproduction, visual effects, augmented reality (AR), and vir" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "20991" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "content": "tual reality (VR) telepresence applications. Efficient animatable portrait generators should be capable of synthesizing diverse high-fidelity portraits with full control of the rigid head pose, facial expressions and gaze directions at a fine-grained level. The main challenges of this task lie in how to model accurate deformation and preserve identity through animation in the generative setting, i.e. training with only unstructured corpus of 2D images." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 170, + 289, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 170, + 289, + 434 + ], + "spans": [ + { + "bbox": [ + 46, + 170, + 289, + 434 + ], + "type": "text", + "content": "Several 2D generative models perform image animation by incorporating the 3D Morphable Face Models (3DMM) [4] into the portrait synthesis [13, 16, 35, 52, 62, 65, 70, 73]. These 2D-based methods achieve photorealism but suffer from shape distortion during large motion due to a lack of geometry constraints. Towards better view consistency, many recent efforts incorporate 3DMM with 3D GANs, learning to synthesize animatable and 3D consistent portraits from only 2D image collections in an unsupervised manner [3, 30, 39, 44, 60, 61, 68, 74]. Bergman et al. [3] propose an explicit surface-driven deformation field for warping radiance fields. While modeling accurate facial deformation, it cannot handle topological changes caused by non-facial components, e.g. hair, glasses, and other accessories. AnifaceGAN [68] builds an implicit 3DMM-conditioned deformation field and constrains animation accuracy by imitation learning. It achieves smooth animation on interpolated expressions, however, struggles to generate reasonable extrapolation due to the under-constrained deformation field. Therefore, The key challenge of this task is modeling deformation in the 3D generative setting for animation accuracy and topological flexibility." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 436, + 289, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 436, + 289, + 687 + ], + "spans": [ + { + "bbox": [ + 46, + 436, + 289, + 687 + ], + "type": "text", + "content": "In this paper, we propose a novel 3D GAN framework for unsupervised learning of generative, high-quality, and 3D-consistent facial avatars from unstructured 2D images. Our model splits the whole head into dynamic and static parts, and models them respectively. For dynamic parts, the key insight is to combine both fine-grained expression control of mesh-guided explicit deformation and flexibility of implicit volumetric representation. To this end, we propose a novel representation, Generative Texture-Rasterized Triplanes, which learns the facial deformation through Generative Neural Textures on top of a parametric template mesh and samples them into three orthogonal-viewed and axis-aligned feature planes through standard rasterization, forming a tri-plane feature representation. Such texture-rasterized tri-planes re-form high-dimensional dynamic surface features in a volumetric representation for efficient volume rendering and thus inherit both the accurate control of the mesh-driven deformation and the expressiveness of volumetric representations. Furthermore, we represent static components (body, hair, background, etc.) by another triplane branch, and integrate both through alpha blending." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "content": "Another key insight of our method is to model the mouth interior which is not taken into account by 3DMM. Mouth" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": "interior is crucial for animation quality but often ignored by prior arts. We propose an efficient teeth synthesis module, formed as a style-modulated UNet, to complete the inner mouth features missed by the template mesh. To further regularize the deformation accuracy, we introduce a deformation-aware discriminator which takes as input synthetic renderings, encouraging the alignment of the final outputs with the 2D projection of the expected deformation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 317, + 168, + 533, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 168, + 533, + 180 + ], + "spans": [ + { + "bbox": [ + 317, + 168, + 533, + 180 + ], + "type": "text", + "content": "To summarize, the contributions of our approach are:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 317, + 185, + 545, + 365 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 317, + 185, + 545, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 185, + 545, + 233 + ], + "spans": [ + { + "bbox": [ + 317, + 185, + 545, + 233 + ], + "type": "text", + "content": "- We present an animatable 3D-aware GAN framework for photorealistic portrait synthesis with fine-grained animation, including expressions, eye blinks, gaze direction and full head poses." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 239, + 545, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 239, + 545, + 299 + ], + "spans": [ + { + "bbox": [ + 317, + 239, + 545, + 299 + ], + "type": "text", + "content": "- We propose Generative Texture-Rasterized Triplanes, an efficient deformable 3D representation that inherits both fine-grained expression control of mesh-guided explicit deformation and flexibility of implicit volumetric representation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 317, + 304, + 545, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 304, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 317, + 304, + 545, + 365 + ], + "type": "text", + "content": "- Our learned generative animatable 3D representation can serve as a strong 3D prior and boost the downstream application of 3D-aware one-shot facial avatars. Our model also pushes the frontier of 3D stylization with high-quality out-of-domain facial avatars." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 370, + 392, + 383 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 370, + 392, + 383 + ], + "spans": [ + { + "bbox": [ + 306, + 370, + 392, + 383 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 390, + 545, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 390, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 304, + 390, + 545, + 569 + ], + "type": "text", + "content": "Generative 3D-aware Image Synthesis. Generative adversarial networks [24] have achieved photorealistic synthesis in 2D domain. Building on the success of 2D GANs, many efforts have lifted the image synthesis into 3D with explicit view control. Early voxel-based approaches [19, 28, 41, 42, 67, 82] adopt the 3D CNN generators whose heavy computational burden limits the high-resolution image synthesis. Recent works incorporate more efficient neural scene representations, such as fully implicit networks [6, 8-10, 14, 47, 54, 57, 81], sparse voxel grids [55], multiple planes [79] or a combination of low-resolution feature volume and 2D super-resolution [7, 27, 43, 46, 71, 72, 76, 78]. We leverage the tri-plane representation proposed in [7] and endow it with animation ability by the orthogonal-rasterized Generative Neural Textures." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 570, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 714 + ], + "type": "text", + "content": "Some other current works [3,30,36,44,58,59,61,68,77, 83] focus on the editability of 3D-aware generative models. FENeRF [59] and IDE-3D [58] perform semantic-guided 3D face editing by incorporating semantic-aware radiance fields and GAN inversion. However, they cannot produce continuous and stable editing on videos. Other methods [3, 30,44,61,68] employ 3D priors to achieve animatable image synthesis and the main differences lie on the deformation strategies including linear blend skinning [30,44], surface-driven deformation [3], 3DMM-guided latent decomposition [61] and neural deformation fields [68]. These approaches either don't allow for topology changes or need" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20992" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 69, + 547, + 222 + ], + "blocks": [ + { + "bbox": [ + 50, + 69, + 547, + 222 + ], + "lines": [ + { + "bbox": [ + 50, + 69, + 547, + 222 + ], + "spans": [ + { + "bbox": [ + 50, + 69, + 547, + 222 + ], + "type": "image", + "image_path": "a1dbff3e376f7c470d28d06fbe8a108f32f1ed7c3003834e4a0de514c95607cb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 232, + 546, + 288 + ], + "lines": [ + { + "bbox": [ + 46, + 232, + 546, + 288 + ], + "spans": [ + { + "bbox": [ + 46, + 232, + 546, + 288 + ], + "type": "text", + "content": "Figure 2. Our 3D GAN framework consists of two tri-plane branches " + }, + { + "bbox": [ + 46, + 232, + 546, + 288 + ], + "type": "inline_equation", + "content": "T_{uv}" + }, + { + "bbox": [ + 46, + 232, + 546, + 288 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 232, + 546, + 288 + ], + "type": "inline_equation", + "content": "T_{static}" + }, + { + "bbox": [ + 46, + 232, + 546, + 288 + ], + "type": "text", + "content": " modeling dynamic and static components. " + }, + { + "bbox": [ + 46, + 232, + 546, + 288 + ], + "type": "inline_equation", + "content": "T_{uv}" + }, + { + "bbox": [ + 46, + 232, + 546, + 288 + ], + "type": "text", + "content": " is formed by the orthogonal rasterized Generative Neural Textures which are synthesized by a StyleGAN generator, " + }, + { + "bbox": [ + 46, + 232, + 546, + 288 + ], + "type": "inline_equation", + "content": "G_{texture}" + }, + { + "bbox": [ + 46, + 232, + 546, + 288 + ], + "type": "text", + "content": ", on top of deformable template mesh. " + }, + { + "bbox": [ + 46, + 232, + 546, + 288 + ], + "type": "inline_equation", + "content": "T_{static}" + }, + { + "bbox": [ + 46, + 232, + 546, + 288 + ], + "type": "text", + "content": " is synthesized by another StyleGAN generator, " + }, + { + "bbox": [ + 46, + 232, + 546, + 288 + ], + "type": "inline_equation", + "content": "G_{static}" + }, + { + "bbox": [ + 46, + 232, + 546, + 288 + ], + "type": "text", + "content": ". The mouth synthesis module, " + }, + { + "bbox": [ + 46, + 232, + 546, + 288 + ], + "type": "inline_equation", + "content": "G_{teeth}" + }, + { + "bbox": [ + 46, + 232, + 546, + 288 + ], + "type": "text", + "content": ", is presented for completing mouth interior. Blended triplanes are incorporated with hybrid neural renderer consisting of volume rendering and a super-resolution module " + }, + { + "bbox": [ + 46, + 232, + 546, + 288 + ], + "type": "inline_equation", + "content": "G_{superres}" + }, + { + "bbox": [ + 46, + 232, + 546, + 288 + ], + "type": "text", + "content": ". For discrimination, synthetic renderings " + }, + { + "bbox": [ + 46, + 232, + 546, + 288 + ], + "type": "inline_equation", + "content": "I_{synthetic}" + }, + { + "bbox": [ + 46, + 232, + 546, + 288 + ], + "type": "text", + "content": " is taken into the dual discriminator " + }, + { + "bbox": [ + 46, + 232, + 546, + 288 + ], + "type": "inline_equation", + "content": "D_{dual}" + }, + { + "bbox": [ + 46, + 232, + 546, + 288 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 298, + 286, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 298, + 286, + 357 + ], + "spans": [ + { + "bbox": [ + 46, + 298, + 286, + 357 + ], + "type": "text", + "content": "elaborate loss design to ensure the accuracy of deformation. On the contrary, our approach naturally achieves accurate animation with explicit mesh guidance and further allows for topology changes by adapting surface deformation into a continuous volumetric representation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 362, + 286, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 362, + 286, + 685 + ], + "spans": [ + { + "bbox": [ + 46, + 362, + 286, + 685 + ], + "type": "text", + "content": "Facial animation with 3D morphable face models. Blanz et al. [4] model facial texture and shape as vector spaces, known as the 3D Morphable Model (3DMM). Extensions of 3DMM, such as full-head PCA models [12, 51], blend-shape models [38], are extensively studied and widely used in facial animation tasks [23, 65]. Benefiting from 3DMM, these methods can model the deformation of facial parts accurately and continuously, nevertheless, struggle to represent non-facial areas missed by 3DMM, e.g. hair, teeth, eyes, and body. Moreover, these methods are prone to lacking facial details. To fill the missing areas and complete more realistic facial details, later works [13, 16, 18, 22, 35, 52, 62, 64, 70] apply learned approaches on top of 3DMM renderings. DiscoFaceGAN [13] maps 3DMM parameters into the latent space and decouple them by imitative-contrastive learning. Though efficient and photorealistic animation is achieved, these 2D methods don't model 3D geometry and thus cannot remain strict 3D consistency with large head pose changes. For strict 3D consistency, recent efforts [1, 20, 25, 75, 80] incorporate 3DMM into volumetric representations [31, 32, 40, 49, 56] to achieve view-consistent facial animation. Furthermore, volumetric representations enable the modeling of thin structures such as hair, and also mouth interior thanks to its spatial continuity. While 3DMM have been adopted to animate radiance fields for single-scene scenarios, it is challenging to adapt to the generative setting with the absence of groundtruth supervision." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 689, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 286, + 713 + ], + "type": "text", + "content": "Neural scene representations. The neural scene representations can be roughly categorized into implicit and explicit" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 298, + 545, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 298, + 545, + 501 + ], + "spans": [ + { + "bbox": [ + 304, + 298, + 545, + 501 + ], + "type": "text", + "content": "surface representations and volumetric representations [63]. The surface can be represented explicitly by point clouds [37, 50], meshes [2, 5, 45, 64], or defined implicitly as a zero level-set of a function [11, 48, 69], like signed distance function, which can be approximated by coordinate-based multi-layer perceptrons (MLPs). On the contrary, volumetric representations [32, 40, 49, 56] store volumetric properties (occupancies, radiance, colors, etc.) instead of the surface of an object. These properties can be stored in explicit voxel grids [32, 49, 56] or the weights of a neural network implicitly [40]. In this work, we propose a hybrid surface-volumetric representation. Specifically, we learn the deformable surface radiance by neural textures [26, 64] on top of the template mesh and rasterize it into three orthogonal-viewed feature planes. Then, the planes are reshaped to a tri-plane representation with decoding into neural radiance fields for volume rendering." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 522, + 372, + 536 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 522, + 372, + 536 + ], + "spans": [ + { + "bbox": [ + 306, + 522, + 372, + 536 + ], + "type": "text", + "content": "3. Approach" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "text", + "content": "We present an animatable 3D-aware facial generator that equips with fine-grained expression and pose control, photorealistic rendering quality, and high-quality underlying geometry. The proposed method models dynamic and static components by two independent tri-plane branches. Specifically, we propose Generative Texture-rasterized Tri-planes for modeling dynamic facial parts (Sec. 3.1). Furthermore, we propose an efficient mouth synthesis module to complete the mouth interior that is not included in 3DMM (Sec. 3.2). We further adopt another tri-plane branch for the static components (Sec. 3.3). Both tri-planes are blended together for hybrid neural rendering (Sec. 3.4). We introduce an deformation-aware discriminator (Sec. 3.5) and illustrate the training objectives in Sec. 3.6." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20993" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 256, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 256, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 256, + 85 + ], + "type": "text", + "content": "3.1. Generative texture-rasterized tri-planes" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 91, + 289, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 91, + 289, + 510 + ], + "spans": [ + { + "bbox": [ + 46, + 91, + 289, + 510 + ], + "type": "text", + "content": "EG3D [7] presents an efficient tri-plane-based hybrid 3D representation to synthesize high-resolution images with multi-view consistency. Nonetheless, EG3D lacks control over facial deformations and thus cannot be directly applied to animation tasks. To this end, we leverage Neural textures [64] to represent deformable facial parts. In general, Neural Textures are a set of learned high-dimensional feature maps that can be interpreted by a neural renderer. We extend it to our generative setting and synthesize the neural textures through a StyleGAN2 CNN generator " + }, + { + "bbox": [ + 46, + 91, + 289, + 510 + ], + "type": "inline_equation", + "content": "G_{\\text{texture}}" + }, + { + "bbox": [ + 46, + 91, + 289, + 510 + ], + "type": "text", + "content": ". As shown in Fig. 2, we first sample a latent code " + }, + { + "bbox": [ + 46, + 91, + 289, + 510 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 46, + 91, + 289, + 510 + ], + "type": "text", + "content": " and map it into an intermediate latent space by the mapping network. Our texture generator architecture closely follows StyleGAN2 backbone [34], except producing a " + }, + { + "bbox": [ + 46, + 91, + 289, + 510 + ], + "type": "inline_equation", + "content": "256 \\times 256 \\times 32" + }, + { + "bbox": [ + 46, + 91, + 289, + 510 + ], + "type": "text", + "content": " neural texture map, " + }, + { + "bbox": [ + 46, + 91, + 289, + 510 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 91, + 289, + 510 + ], + "type": "text", + "content": ", instead of a three-channel RGB image. Storing a high-dimensional learned feature vector per texel, " + }, + { + "bbox": [ + 46, + 91, + 289, + 510 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 91, + 289, + 510 + ], + "type": "text", + "content": " can be rasterized to a view-dependent screen-space feature map given a mesh with uv-texture parameterization and a target view as input. In our case, we use the FLAME template [38] to provide a coarse mesh that can be driven by deformation parameters. Given the predesigned texture mapping function, we employ the standard graphics pipeline to rasterize the neural textures from the texture space into the screen space based on the template mesh. We choose Neural Textures as the deformation method for two reasons. First, compared with other explicit deformation (e.g. linear blend skinning and surface-driven deformation) highly dependent on the accurate underlying geometry, Neural Textures embed high-level features which compensate the imperfect geometry and thus are more suitable for our settings where template meshes are not accurate. Furthermore, unlike implicit deformation methods [13, 61, 68], our explicit mesh-guided deformation alleviates the requirement of elaborate imitation learning while gaining better expression generalization (Fig. 4)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "text", + "content": "Neural Textures encode surface deformation accurately with mesh guidance but lack generalization to 3D points far from surface. Besides, it also doesn't allow for topological changes. To this end, we propose Generative Texture-Rasterized Tri-planes, " + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "inline_equation", + "content": "T_{uv}" + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "text", + "content": " which reshapes the rasterized textures into a tri-plane representation. Therefore, we can adapt such surface deformation into a continuous volume, and once a 3D point out of the surface could be projected within the facial part on at least one view (e.g. glasses), its feature can be interpolated by the meaningful surface feature through grid sampling in the tri-plane. Specifically, we rasterize neural textures based on the template mesh into three orthogonal views and place them in three axis-aligned feature planes. In practice, considering the zygomorphy, the rasterization is applied at both the left and right views and the rasterized features are concatenated for one single plane by summation. In this way, Our hybrid surface-volumetric" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": "representation inherits the best of both worlds: accurate mesh-guided facial deformation of Neural Textures and the continuity and topological flexibility of volumetric representations. See Fig. 5 for the topological-aware animation of portraits with glasses." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 305, + 142, + 443, + 155 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 142, + 443, + 155 + ], + "spans": [ + { + "bbox": [ + 305, + 142, + 443, + 155 + ], + "type": "text", + "content": "3.2. Mouth synthesis module" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 307, + 171, + 545, + 304 + ], + "blocks": [ + { + "bbox": [ + 307, + 171, + 545, + 304 + ], + "lines": [ + { + "bbox": [ + 307, + 171, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 307, + 171, + 545, + 304 + ], + "type": "image", + "image_path": "71e47e74911ff83045eae0d1df72dc8e8b6ba1b2fe084fd9c70ff17820e52c0a.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 311, + 545, + 355 + ], + "lines": [ + { + "bbox": [ + 304, + 311, + 545, + 355 + ], + "spans": [ + { + "bbox": [ + 304, + 311, + 545, + 355 + ], + "type": "text", + "content": "Figure 3. The teeth synthesis module consists of a teeth completing module and a neural blending module. The teeth competing module produces mouth interior conditioned on multi-scale teeth exterior features and latent codes " + }, + { + "bbox": [ + 304, + 311, + 545, + 355 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 304, + 311, + 545, + 355 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 361, + 546, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 361, + 546, + 564 + ], + "spans": [ + { + "bbox": [ + 304, + 361, + 546, + 564 + ], + "type": "text", + "content": "Since the FLAME template doesn't contain inner mouth, we propose a teeth synthesis module, " + }, + { + "bbox": [ + 304, + 361, + 546, + 564 + ], + "type": "inline_equation", + "content": "G_{\\text{teeth}}" + }, + { + "bbox": [ + 304, + 361, + 546, + 564 + ], + "type": "text", + "content": ", to complete the missing teeth features in " + }, + { + "bbox": [ + 304, + 361, + 546, + 564 + ], + "type": "inline_equation", + "content": "T_{uv}" + }, + { + "bbox": [ + 304, + 361, + 546, + 564 + ], + "type": "text", + "content": ". As shown in Fig. 3, for each feature plane of " + }, + { + "bbox": [ + 304, + 361, + 546, + 564 + ], + "type": "inline_equation", + "content": "T_{uv}" + }, + { + "bbox": [ + 304, + 361, + 546, + 564 + ], + "type": "text", + "content": ", we crop the teeth area by the expanded landmarks and resize it into " + }, + { + "bbox": [ + 304, + 361, + 546, + 564 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 304, + 361, + 546, + 564 + ], + "type": "text", + "content": ". Then, the stacked mouth features are processed by " + }, + { + "bbox": [ + 304, + 361, + 546, + 564 + ], + "type": "inline_equation", + "content": "G_{\\text{teeth}}" + }, + { + "bbox": [ + 304, + 361, + 546, + 564 + ], + "type": "text", + "content": " which employs a style-modulated UNet [66]. The downsampling process of " + }, + { + "bbox": [ + 304, + 361, + 546, + 564 + ], + "type": "inline_equation", + "content": "G_{\\text{teeth}}" + }, + { + "bbox": [ + 304, + 361, + 546, + 564 + ], + "type": "text", + "content": " encodes " + }, + { + "bbox": [ + 304, + 361, + 546, + 564 + ], + "type": "inline_equation", + "content": "f_{\\text{teeth}}" + }, + { + "bbox": [ + 304, + 361, + 546, + 564 + ], + "type": "text", + "content": " into multi-scale feature maps which serve as content conditions for the following StyleGAN layers. The output teeth features " + }, + { + "bbox": [ + 304, + 361, + 546, + 564 + ], + "type": "inline_equation", + "content": "f_{\\text{teeth}}'" + }, + { + "bbox": [ + 304, + 361, + 546, + 564 + ], + "type": "text", + "content": " are transformed inversely and concatenated with the feature planes of " + }, + { + "bbox": [ + 304, + 361, + 546, + 564 + ], + "type": "inline_equation", + "content": "T_{uv}" + }, + { + "bbox": [ + 304, + 361, + 546, + 564 + ], + "type": "text", + "content": ". To eliminate the texture flickering of mouth boundary, we further feed " + }, + { + "bbox": [ + 304, + 361, + 546, + 564 + ], + "type": "inline_equation", + "content": "T_{uv}" + }, + { + "bbox": [ + 304, + 361, + 546, + 564 + ], + "type": "text", + "content": " into a shallow-UNet-based neural blending module and obtain " + }, + { + "bbox": [ + 304, + 361, + 546, + 564 + ], + "type": "inline_equation", + "content": "T_{uv}'" + }, + { + "bbox": [ + 304, + 361, + 546, + 564 + ], + "type": "text", + "content": ". We conduct a series of ablation studies and prove that the proposed teeth synthesis module brings a remarkable improvement on both the animation accuracy and synthesis quality (Sec. 4.2)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 574, + 460, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 574, + 460, + 588 + ], + "spans": [ + { + "bbox": [ + 306, + 574, + 460, + 588 + ], + "type": "text", + "content": "3.3. Modeling static components" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "The generative texture-rasterized tri-planes manage to model dynamic faces varying expressions and shapes, though, it is challenging to synthesize static parts like diverse haircut, background and upper body which are not included in the FLAME template. To this end, we model these parts by another tri-plane branch, " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "T_{\\text{static}}" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": ", which is generated by a StyleGAN2 CNN generator " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "G_{\\text{static}}" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": " sharing the same latent code with " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "G_{\\text{texture}}" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": ". The plane features of " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "T_{uv}^{\\prime}" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "inline_equation", + "content": "T_{\\text{static}}" + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": " are blended on each plane by the alpha masks rendered by rasterization. Such design not only benefits the" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20994" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "content": "modeling of various-styled static components but also enforces their consistency during facial animations." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 103, + 151, + 117 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 103, + 151, + 117 + ], + "spans": [ + { + "bbox": [ + 47, + 103, + 151, + 117 + ], + "type": "text", + "content": "3.4. Neural rendering" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 121, + 289, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 121, + 289, + 301 + ], + "spans": [ + { + "bbox": [ + 46, + 121, + 289, + 301 + ], + "type": "text", + "content": "Given the blended tri-planes, for any point in the 3D space, we project it into each plane and sample the features bi-linearly. Then, the sampled features are aggregated by summation and decoded into volume density " + }, + { + "bbox": [ + 46, + 121, + 289, + 301 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 46, + 121, + 289, + 301 + ], + "type": "text", + "content": " and feature " + }, + { + "bbox": [ + 46, + 121, + 289, + 301 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 121, + 289, + 301 + ], + "type": "text", + "content": " by a lightweight decoder. Similar to [7,58], the decoder is a single hidden layered multi-layer perceptron (MLP) with softplus activation. The volume rendering is employed to accumulate " + }, + { + "bbox": [ + 46, + 121, + 289, + 301 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 46, + 121, + 289, + 301 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 121, + 289, + 301 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 121, + 289, + 301 + ], + "type": "text", + "content": " along the rays cast through each pixel to compute a 2D feature image " + }, + { + "bbox": [ + 46, + 121, + 289, + 301 + ], + "type": "inline_equation", + "content": "I_{f}" + }, + { + "bbox": [ + 46, + 121, + 289, + 301 + ], + "type": "text", + "content": ". Similar to [7,27,46], we leverage a 2D super-resolution module " + }, + { + "bbox": [ + 46, + 121, + 289, + 301 + ], + "type": "inline_equation", + "content": "G_{superres}" + }, + { + "bbox": [ + 46, + 121, + 289, + 301 + ], + "type": "text", + "content": " to interpret the feature image into RGB image " + }, + { + "bbox": [ + 46, + 121, + 289, + 301 + ], + "type": "inline_equation", + "content": "I_{RGB}" + }, + { + "bbox": [ + 46, + 121, + 289, + 301 + ], + "type": "text", + "content": " with higher resolution. The super-resolution module consists of three StyleGAN2 synthesis blocks and the noise input is removed for alleviating texture flickers. In our case, " + }, + { + "bbox": [ + 46, + 121, + 289, + 301 + ], + "type": "inline_equation", + "content": "I_{f}" + }, + { + "bbox": [ + 46, + 121, + 289, + 301 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 121, + 289, + 301 + ], + "type": "inline_equation", + "content": "I_{RGB}" + }, + { + "bbox": [ + 46, + 121, + 289, + 301 + ], + "type": "text", + "content": " are set to " + }, + { + "bbox": [ + 46, + 121, + 289, + 301 + ], + "type": "inline_equation", + "content": "64\\times 64" + }, + { + "bbox": [ + 46, + 121, + 289, + 301 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 121, + 289, + 301 + ], + "type": "inline_equation", + "content": "512\\times 512" + }, + { + "bbox": [ + 46, + 121, + 289, + 301 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 308, + 228, + 320 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 308, + 228, + 320 + ], + "spans": [ + { + "bbox": [ + 47, + 308, + 228, + 320 + ], + "type": "text", + "content": "3.5. Deformation-aware discriminator" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 327, + 289, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 327, + 289, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 327, + 289, + 567 + ], + "type": "text", + "content": "To learn the unsupervised 3D representations, we adopt a 2D convolutional discriminator " + }, + { + "bbox": [ + 46, + 327, + 289, + 567 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 327, + 289, + 567 + ], + "type": "text", + "content": " to critique the renderings. Inspired by [7], we regularize the first three channels of " + }, + { + "bbox": [ + 46, + 327, + 289, + 567 + ], + "type": "inline_equation", + "content": "I_{f}" + }, + { + "bbox": [ + 46, + 327, + 289, + 567 + ], + "type": "text", + "content": " as low-resolution RGB image, which is concatenated with " + }, + { + "bbox": [ + 46, + 327, + 289, + 567 + ], + "type": "inline_equation", + "content": "I_{RGB}" + }, + { + "bbox": [ + 46, + 327, + 289, + 567 + ], + "type": "text", + "content": " as the input for the discriminator. However, the discrimination with only image input can only ensure that the deformed images are always in a correct distribution instead of matching the expected deformations. Therefore, we make the discriminator aware of the expression and shape under which the generated image are deformed. For the same purpose, GNARF [3] conditions the discriminator on the FLAME parameters by concatenating them as the input to the mapping network. However, we find empirically that such a conditioning method leads to training instability, consistent with [3]. Instead, we re-render the template mesh under the rendered pose to get the synthetic rendering " + }, + { + "bbox": [ + 46, + 327, + 289, + 567 + ], + "type": "inline_equation", + "content": "I_{synthetic}" + }, + { + "bbox": [ + 46, + 327, + 289, + 567 + ], + "type": "text", + "content": " and feed it into " + }, + { + "bbox": [ + 46, + 327, + 289, + 567 + ], + "type": "inline_equation", + "content": "D_{\\text{dual}}" + }, + { + "bbox": [ + 46, + 327, + 289, + 567 + ], + "type": "text", + "content": " along with image pairs. Here, we adopt correspondence images inspired by [35]. Such concatenation encourages the final output to align with the synthetic rendering and learn the expected deformation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 573, + 160, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 573, + 160, + 586 + ], + "spans": [ + { + "bbox": [ + 47, + 573, + 160, + 586 + ], + "type": "text", + "content": "3.6. Training objectives" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 591, + 287, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 591, + 287, + 639 + ], + "spans": [ + { + "bbox": [ + 47, + 591, + 287, + 639 + ], + "type": "text", + "content": "During training, we use the non-saturating GAN loss with R1 regularization. Moreover, we adopt the density regularization proposed in EG3D [7]. Therefore, the total learning objective is:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 80, + 646, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 646, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 80, + 646, + 287, + 715 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {D _ {\\text {d u a l}, G}} = \\mathbb {E} _ {z \\sim p _ {z}, \\epsilon \\sim p _ {\\epsilon}} [ f (D _ {\\text {d u a l}} (G (z, \\epsilon))) ] + \\\\ \\mathbb {E} _ {I ^ {r} \\sim p _ {I ^ {r}}} [ f (- D _ {\\text {d u a l}} (I ^ {r}) + (1) \\\\ \\left. \\lambda \\| \\bigtriangledown D _ {d u a l} \\left(I ^ {r}\\right) \\| ^ {2} \\right], \\\\ \\mathcal {L} _ {\\text {d e n s i t y}} = \\sum_ {x _ {s} \\in \\mathcal {S}} \\| d (x _ {s}) - d (x _ {s} + \\epsilon) \\| _ {2}, (2) \\\\ \\end{array}", + "image_path": "a56f9f1a2c0fdf4f1cb3c56bc7d3d28bb2d1189fff1afa1d3a313a9ea0408e80.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 352, + 83, + 545, + 95 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 352, + 83, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 352, + 83, + 545, + 95 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {D _ {\\text {d u a l}}, G} + \\lambda_ {\\text {d e n s i t y}} \\mathcal {L} _ {\\text {d e n s i t y}}, \\tag {3}", + "image_path": "24c7d6574fd51cff373b1dbd289ac91709b647c35ca26da55bdea18b253b3e67.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 102, + 547, + 211 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 102, + 547, + 211 + ], + "spans": [ + { + "bbox": [ + 304, + 102, + 547, + 211 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 102, + 547, + 211 + ], + "type": "inline_equation", + "content": "I^r" + }, + { + "bbox": [ + 304, + 102, + 547, + 211 + ], + "type": "text", + "content": " is the combination of real images, blurred real images, and the corresponding synthetic renderings, which are sampled from the training set with distribution " + }, + { + "bbox": [ + 304, + 102, + 547, + 211 + ], + "type": "inline_equation", + "content": "p_I" + }, + { + "bbox": [ + 304, + 102, + 547, + 211 + ], + "type": "text", + "content": ". We adopt many training hyperparameters from EG3D and StyleGAN2 (learning rates of generator and discriminator, batch size, R1 regularization, etc.). We train our model based on the pretrained model of EG3D [7] and continue to train on 4 3090 GPUs for roughly 4 days. Please refer to the supplemental material for the implementation details." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 220, + 388, + 234 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 220, + 388, + 234 + ], + "spans": [ + { + "bbox": [ + 306, + 220, + 388, + 234 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 240, + 545, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 240, + 545, + 324 + ], + "spans": [ + { + "bbox": [ + 304, + 240, + 545, + 324 + ], + "type": "text", + "content": "In this section, we first show qualitative and quantitative comparisons to state-of-the-art 2D / 3D animatable generative facial models (Sec. 4.1), and then discuss the conducted ablation studies of our design choices (Sec. 4.2). Furthermore, we show various applications combining our efficient 3D representation with GAN technologies such as GAN inversion and style transfer (Sec. 4.3)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 324, + 546, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 324, + 546, + 491 + ], + "spans": [ + { + "bbox": [ + 304, + 324, + 546, + 491 + ], + "type": "text", + "content": "Datasets. We train and test our methods on FFHQ [33]. We augment FFHQ with horizontal flips and use an off-the-shelf pose estimator [15] to label images with the approximated camera extrinsic parameters and constant intrinsics. To support full pose animation, in-plane (roll) rotation is also considered. Furthermore, we use DECA [17] to estimate the FLAME parameters of facial identity " + }, + { + "bbox": [ + 304, + 324, + 546, + 491 + ], + "type": "inline_equation", + "content": "\\beta \\in \\mathbb{R}^{100}" + }, + { + "bbox": [ + 304, + 324, + 546, + 491 + ], + "type": "text", + "content": ", jaw pose " + }, + { + "bbox": [ + 304, + 324, + 546, + 491 + ], + "type": "inline_equation", + "content": "\\theta_{jaw} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 324, + 546, + 491 + ], + "type": "text", + "content": " and expression " + }, + { + "bbox": [ + 304, + 324, + 546, + 491 + ], + "type": "inline_equation", + "content": "\\psi \\in \\mathbb{R}^{50}" + }, + { + "bbox": [ + 304, + 324, + 546, + 491 + ], + "type": "text", + "content": ". Since DECA doesn't account for eyeball movement, we additionally adopt an efficient facial detector to detect 2D landmarks of irises and optimize eye poses " + }, + { + "bbox": [ + 304, + 324, + 546, + 491 + ], + "type": "inline_equation", + "content": "\\theta_{eye} \\in \\mathbb{R}^6" + }, + { + "bbox": [ + 304, + 324, + 546, + 491 + ], + "type": "text", + "content": " by minimizing the re-projection errors. Based on these FLAME parameters, we produce a template mesh with 5023 vertices and 9976 faces to drive facial deformations." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 498, + 391, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 498, + 391, + 510 + ], + "spans": [ + { + "bbox": [ + 306, + 498, + 391, + 510 + ], + "type": "text", + "content": "4.1. Comparisons" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 517, + 545, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 517, + 545, + 696 + ], + "spans": [ + { + "bbox": [ + 304, + 517, + 545, + 696 + ], + "type": "text", + "content": "Baselines. We compare our method against two state-of-the-art methods for animatable 3D-aware image synthesis: 3DFaceShop [61], and AniFaceGAN [68]. Besides, we also select DiscoFaceGAN [13] as a baseline, which generates animatable 2D portraits conditioned on 3DMM parameters. Qualitative comparison. Fig. 4 provides a qualitative comparison against baselines. Overall, as can be seen, our method outperforms all baselines by some margin on both synthesis quality and animation accuracy. Specifically, DiscoFaceGAN [13] suffers from inconsistent identity during animation. Moreover, it cannot generate reasonable mouth interior, e.g. stretched teeth. 3DFaceshop and AnifaceGAN synthesize 3D-consistent images, nevertheless, still struggle to model consistent mouth interior with the driving images. This is because their implicit deformation approaches" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 317, + 703, + 395, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 703, + 395, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 703, + 395, + 713 + ], + "type": "text", + "content": "1 https://mediapipe.dev/" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "20995" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 66, + 70, + 477, + 140 + ], + "blocks": [ + { + "bbox": [ + 52, + 87, + 63, + 120 + ], + "lines": [ + { + "bbox": [ + 52, + 87, + 63, + 120 + ], + "spans": [ + { + "bbox": [ + 52, + 87, + 63, + 120 + ], + "type": "text", + "content": "Driving" + } + ] + } + ], + "index": 0, + "angle": 270, + "type": "image_caption" + }, + { + "bbox": [ + 66, + 70, + 477, + 140 + ], + "lines": [ + { + "bbox": [ + 66, + 70, + 477, + 140 + ], + "spans": [ + { + "bbox": [ + 66, + 70, + 477, + 140 + ], + "type": "image", + "image_path": "ae3815df2f0a2a31171a1005563d372e24693015e2a3b8ec4c1162c9344f8af6.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 489, + 102, + 531, + 113 + ], + "lines": [ + { + "bbox": [ + 489, + 102, + 531, + 113 + ], + "spans": [ + { + "bbox": [ + 489, + 102, + 531, + 113 + ], + "type": "text", + "content": "Free View" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 66, + 144, + 545, + 213 + ], + "blocks": [ + { + "bbox": [ + 52, + 167, + 63, + 187 + ], + "lines": [ + { + "bbox": [ + 52, + 167, + 63, + 187 + ], + "spans": [ + { + "bbox": [ + 52, + 167, + 63, + 187 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 3, + "angle": 270, + "type": "image_caption" + }, + { + "bbox": [ + 66, + 144, + 545, + 213 + ], + "lines": [ + { + "bbox": [ + 66, + 144, + 545, + 213 + ], + "spans": [ + { + "bbox": [ + 66, + 144, + 545, + 213 + ], + "type": "image", + "image_path": "c3608cd3755d3e5d9deab540e36e52453e0462f88afbea3b145385dae3875be7.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 66, + 216, + 545, + 285 + ], + "blocks": [ + { + "bbox": [ + 53, + 223, + 63, + 277 + ], + "lines": [ + { + "bbox": [ + 53, + 223, + 63, + 277 + ], + "spans": [ + { + "bbox": [ + 53, + 223, + 63, + 277 + ], + "type": "text", + "content": "AntiFaceGAN" + } + ] + } + ], + "index": 5, + "angle": 270, + "type": "image_caption" + }, + { + "bbox": [ + 66, + 216, + 545, + 285 + ], + "lines": [ + { + "bbox": [ + 66, + 216, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 66, + 216, + 545, + 285 + ], + "type": "image", + "image_path": "b182aff722264e93d8d2b436eda799d715f52ba236c2938ccf8a6c4ac65f3e9d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 66, + 289, + 545, + 358 + ], + "blocks": [ + { + "bbox": [ + 53, + 297, + 63, + 349 + ], + "lines": [ + { + "bbox": [ + 53, + 297, + 63, + 349 + ], + "spans": [ + { + "bbox": [ + 53, + 297, + 63, + 349 + ], + "type": "text", + "content": "3DFaceShop" + } + ] + } + ], + "index": 7, + "angle": 270, + "type": "image_caption" + }, + { + "bbox": [ + 66, + 289, + 545, + 358 + ], + "lines": [ + { + "bbox": [ + 66, + 289, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 66, + 289, + 545, + 358 + ], + "type": "image", + "image_path": "bfa21a81a77e31ee6ec3f66b0aecf1efb8cb7235016333588d1752269d2aa997.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 66, + 363, + 545, + 433 + ], + "blocks": [ + { + "bbox": [ + 53, + 365, + 63, + 427 + ], + "lines": [ + { + "bbox": [ + 53, + 365, + 63, + 427 + ], + "spans": [ + { + "bbox": [ + 53, + 365, + 63, + 427 + ], + "type": "text", + "content": "DiscoFaceGAN" + } + ] + } + ], + "index": 9, + "angle": 270, + "type": "image_caption" + }, + { + "bbox": [ + 66, + 363, + 545, + 433 + ], + "lines": [ + { + "bbox": [ + 66, + 363, + 545, + 433 + ], + "spans": [ + { + "bbox": [ + 66, + 363, + 545, + 433 + ], + "type": "image", + "image_path": "c1cc234d65e8b8316cda7fa7ddc34a56ca315f319037992a7fd82a3751259dca.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 50, + 483, + 106, + 540 + ], + "blocks": [ + { + "bbox": [ + 46, + 442, + 546, + 464 + ], + "lines": [ + { + "bbox": [ + 46, + 442, + 546, + 464 + ], + "spans": [ + { + "bbox": [ + 46, + 442, + 546, + 464 + ], + "type": "text", + "content": "Figure 4. Comparison with the state-of-the-art animatable 3D & 2D image synthesis methods. We extract several frames from a video clip and use the interpreted face model parameters to animate random virtual avatars." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 50, + 483, + 106, + 540 + ], + "lines": [ + { + "bbox": [ + 50, + 483, + 106, + 540 + ], + "spans": [ + { + "bbox": [ + 50, + 483, + 106, + 540 + ], + "type": "image", + "image_path": "d0db2f87db9803a0da906aa002e9b16abc4583b704cc5bf121185b5bdce93ba5.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 109, + 483, + 166, + 540 + ], + "blocks": [ + { + "bbox": [ + 109, + 483, + 166, + 540 + ], + "lines": [ + { + "bbox": [ + 109, + 483, + 166, + 540 + ], + "spans": [ + { + "bbox": [ + 109, + 483, + 166, + 540 + ], + "type": "image", + "image_path": "ec97242d1e2a36766e4e1f9e0ef73896b9644c04e13ea18d6003add53b00cf63.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 169, + 483, + 227, + 540 + ], + "blocks": [ + { + "bbox": [ + 169, + 483, + 227, + 540 + ], + "lines": [ + { + "bbox": [ + 169, + 483, + 227, + 540 + ], + "spans": [ + { + "bbox": [ + 169, + 483, + 227, + 540 + ], + "type": "image", + "image_path": "491c44e32c6fe6b05d1ec0d69a8d1361306a65d270634e0d87a13c30bcac7310.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 229, + 483, + 286, + 540 + ], + "blocks": [ + { + "bbox": [ + 229, + 483, + 286, + 540 + ], + "lines": [ + { + "bbox": [ + 229, + 483, + 286, + 540 + ], + "spans": [ + { + "bbox": [ + 229, + 483, + 286, + 540 + ], + "type": "image", + "image_path": "6b50011c5c4550f1a80fe1cc2524b78f9d08ce245533e4a57c4857ccd52973e4.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 50, + 544, + 106, + 601 + ], + "blocks": [ + { + "bbox": [ + 50, + 544, + 106, + 601 + ], + "lines": [ + { + "bbox": [ + 50, + 544, + 106, + 601 + ], + "spans": [ + { + "bbox": [ + 50, + 544, + 106, + 601 + ], + "type": "image", + "image_path": "747fabddb71b00197f5cde37368f35b8e1132323cc3564f89b80afa35ca2b01e.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 109, + 544, + 166, + 601 + ], + "blocks": [ + { + "bbox": [ + 109, + 544, + 166, + 601 + ], + "lines": [ + { + "bbox": [ + 109, + 544, + 166, + 601 + ], + "spans": [ + { + "bbox": [ + 109, + 544, + 166, + 601 + ], + "type": "image", + "image_path": "8d34245a79b281cadd16e8759c45d5da48d5c693dae5e77dbaa2cfccc22ff6ed.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 169, + 544, + 227, + 601 + ], + "blocks": [ + { + "bbox": [ + 169, + 544, + 227, + 601 + ], + "lines": [ + { + "bbox": [ + 169, + 544, + 227, + 601 + ], + "spans": [ + { + "bbox": [ + 169, + 544, + 227, + 601 + ], + "type": "image", + "image_path": "a89aef0716d09913f5dfbcc52fe9fcb23ab97eca1bcd2e296b381b9514969887.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 229, + 544, + 286, + 601 + ], + "blocks": [ + { + "bbox": [ + 229, + 544, + 286, + 601 + ], + "lines": [ + { + "bbox": [ + 229, + 544, + 286, + 601 + ], + "spans": [ + { + "bbox": [ + 229, + 544, + 286, + 601 + ], + "type": "image", + "image_path": "d669a0a9d1c38bc3638dff1a521df2e5a05963cf25436f60a8a981d0686de340.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 50, + 601, + 106, + 628 + ], + "blocks": [ + { + "bbox": [ + 50, + 601, + 106, + 628 + ], + "lines": [ + { + "bbox": [ + 50, + 601, + 106, + 628 + ], + "spans": [ + { + "bbox": [ + 50, + 601, + 106, + 628 + ], + "type": "image", + "image_path": "08c3f15f9b04411dead30dee5f32042e70407fdc0a968c30f4bf75a36089db44.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 637, + 286, + 670 + ], + "lines": [ + { + "bbox": [ + 47, + 637, + 286, + 670 + ], + "spans": [ + { + "bbox": [ + 47, + 637, + 286, + 670 + ], + "type": "text", + "content": "Figure 5. High-quality dynamic shapes with topological changes. As can be seen, we model detailed dynamic shapes of eyelids and lips, while keeping glasses unchanged." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 109, + 601, + 166, + 628 + ], + "blocks": [ + { + "bbox": [ + 109, + 601, + 166, + 628 + ], + "lines": [ + { + "bbox": [ + 109, + 601, + 166, + 628 + ], + "spans": [ + { + "bbox": [ + 109, + 601, + 166, + 628 + ], + "type": "image", + "image_path": "b65e08731f9e60d2cd05d296b57f91eafea40bdbe8789a17e10d50ba13ee3190.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 169, + 601, + 227, + 628 + ], + "blocks": [ + { + "bbox": [ + 169, + 601, + 227, + 628 + ], + "lines": [ + { + "bbox": [ + 169, + 601, + 227, + 628 + ], + "spans": [ + { + "bbox": [ + 169, + 601, + 227, + 628 + ], + "type": "image", + "image_path": "3df380a8b1750ce5e9429c69f72d2b4f418940a89ae3904bbf0db446acdb7901.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 229, + 601, + 286, + 628 + ], + "blocks": [ + { + "bbox": [ + 229, + 601, + 286, + 628 + ], + "lines": [ + { + "bbox": [ + 229, + 601, + 286, + 628 + ], + "spans": [ + { + "bbox": [ + 229, + 601, + 286, + 628 + ], + "type": "image", + "image_path": "83691ef05468f4cc561e3309aed50b804bdf082ea5f96fdd68a623bb8b09c2cc.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "bbox": [ + 47, + 677, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 286, + 713 + ], + "type": "text", + "content": "are under-constrained, leading to overfit the expression bias (smiling with mouth half-opened) of datasets. Compared to the other methods, our approach not only synthesizes im" + } + ] + } + ], + "index": 25 + }, + { + "type": "table", + "bbox": [ + 306, + 483, + 547, + 548 + ], + "blocks": [ + { + "bbox": [ + 306, + 483, + 547, + 548 + ], + "lines": [ + { + "bbox": [ + 306, + 483, + 547, + 548 + ], + "spans": [ + { + "bbox": [ + 306, + 483, + 547, + 548 + ], + "type": "table", + "html": "
FID↓AED↓APD↓APD*↓ID↑
DiscoFaceGAN (2562) [13]17.10.420.0460.0240.73
AniFaceGAN (2562) [68]20.10.250.0410.0220.82
3DFaceShop (5122) [61]23.70.310.0450.0240.75
Ours (2562)3.70.160.0180.0130.85
Ours (5122)3.90.160.0230.0190.84
", + "image_path": "753c4c27868f9bccbff94fbc8131234382397a0fcc33e2e9c3e32e7e724e0471.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "table_body" + } + ], + "index": 26 + }, + { + "bbox": [ + 305, + 556, + 545, + 599 + ], + "lines": [ + { + "bbox": [ + 305, + 556, + 545, + 599 + ], + "spans": [ + { + "bbox": [ + 305, + 556, + 545, + 599 + ], + "type": "text", + "content": "Table 1. Quantitative comparison using FID, average expression distance (AED), average pose distance (APD), and identity consistency (ID) for FFHQ. APD* means calculating pose distance with roll fixed." + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "ages with higher quality, but also preserves more detailed expressions of the driver images, including mouth interior, eye blinks and eye movements. Furthermore, we are the only method that supports in-plane head rotations. Fig. 5 provides visual examples of the synthesized high-quality geometry. Our approach can model detailed shape deformations (see zoomed eyelids and lips in Fig. 5) with topological awareness, i.e. glasses are kept unchanged." + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20996" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 69, + 286, + 425 + ], + "blocks": [ + { + "bbox": [ + 52, + 69, + 286, + 425 + ], + "lines": [ + { + "bbox": [ + 52, + 69, + 286, + 425 + ], + "spans": [ + { + "bbox": [ + 52, + 69, + 286, + 425 + ], + "type": "image", + "image_path": "5eaaba2d514d716f07aed2262464eea28519aee93db48132f33c0c3f89c9d905.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 433, + 287, + 477 + ], + "lines": [ + { + "bbox": [ + 46, + 433, + 287, + 477 + ], + "spans": [ + { + "bbox": [ + 46, + 433, + 287, + 477 + ], + "type": "text", + "content": "Figure 6. Ablation study on model designs. " + }, + { + "bbox": [ + 46, + 433, + 287, + 477 + ], + "type": "inline_equation", + "content": "T_{\\text{static}}" + }, + { + "bbox": [ + 46, + 433, + 287, + 477 + ], + "type": "text", + "content": " encourages better identity consistency; " + }, + { + "bbox": [ + 46, + 433, + 287, + 477 + ], + "type": "inline_equation", + "content": "G_{\\text{teeth}}" + }, + { + "bbox": [ + 46, + 433, + 287, + 477 + ], + "type": "text", + "content": " benefits realistic mouth interior; the discriminator with " + }, + { + "bbox": [ + 46, + 433, + 287, + 477 + ], + "type": "inline_equation", + "content": "I_{\\text{synthetic}}" + }, + { + "bbox": [ + 46, + 433, + 287, + 477 + ], + "type": "text", + "content": " input allows more consistent reconstruction of detailed expressions." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 498, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 713 + ], + "type": "text", + "content": "Quantitative evaluation. Tab. 1 demonstrates quantitative results comparing our method against baselines evaluated on several metrics. We measure image quality with Frechet Inception Distance (FID) [29] between the entire FFHQ dataset and " + }, + { + "bbox": [ + 46, + 498, + 287, + 713 + ], + "type": "inline_equation", + "content": "50\\mathrm{k}" + }, + { + "bbox": [ + 46, + 498, + 287, + 713 + ], + "type": "text", + "content": " generated images using randomly sampled latent codes, camera poses, and FLAME parameters. Since AntiFaceGAN and DiscoFaceGAN synthesize images on the resolution of " + }, + { + "bbox": [ + 46, + 498, + 287, + 713 + ], + "type": "inline_equation", + "content": "256^2" + }, + { + "bbox": [ + 46, + 498, + 287, + 713 + ], + "type": "text", + "content": ", we test FID of them on FFHQ256². Following [3, 39], we evaluate the faithfulness of the animation with the Average Expression Distance (AED), the Average Pose Distance (APD), and identity consistency (ID). For each method, we randomly sample 500 identities and animate each with randomly sampled 20 FLAME parameters of expressions and poses. Then, we estimate the FLAME parameters for these 10000 generated images and the average distances between the driving FLAME parameters and the reconstructed ones. For identity consistency, we randomly sample 2000 poses, 2000 sets" + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 326, + 70, + 526, + 148 + ], + "blocks": [ + { + "bbox": [ + 326, + 70, + 526, + 148 + ], + "lines": [ + { + "bbox": [ + 326, + 70, + 526, + 148 + ], + "spans": [ + { + "bbox": [ + 326, + 70, + 526, + 148 + ], + "type": "table", + "html": "
FID↓AED↓APD↓ID↑
w/o Tstatic6.60.250.0260.63
w/o Gteeth(uv)7.20.370.0420.71
w/o Gteeth(s)8.40.320.0360.79
w/o Isynthetic3.80.180.0250.74
Full3.90.160.0230.84
", + "image_path": "755c28a4c6a7aae096a3b027ec7e7eb8818b79d19897c9ad746aef4705225ddd.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 305, + 157, + 545, + 213 + ], + "lines": [ + { + "bbox": [ + 305, + 157, + 545, + 213 + ], + "spans": [ + { + "bbox": [ + 305, + 157, + 545, + 213 + ], + "type": "text", + "content": "Table 2. Ablation study on model designs. Modeling static components by " + }, + { + "bbox": [ + 305, + 157, + 545, + 213 + ], + "type": "inline_equation", + "content": "T_{\\text{static}}" + }, + { + "bbox": [ + 305, + 157, + 545, + 213 + ], + "type": "text", + "content": " improves identity consistency. The teeth synthesis " + }, + { + "bbox": [ + 305, + 157, + 545, + 213 + ], + "type": "inline_equation", + "content": "G_{\\text{teeth}}" + }, + { + "bbox": [ + 305, + 157, + 545, + 213 + ], + "type": "text", + "content": " module benefits both animation and synthesis quality significantly while adding synthetic renderings " + }, + { + "bbox": [ + 305, + 157, + 545, + 213 + ], + "type": "inline_equation", + "content": "I_{\\text{synthetic}}" + }, + { + "bbox": [ + 305, + 157, + 545, + 213 + ], + "type": "text", + "content": " into discrimination takes both a step further." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 227, + 545, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 227, + 545, + 395 + ], + "spans": [ + { + "bbox": [ + 305, + 227, + 545, + 395 + ], + "type": "text", + "content": "of FLAME parameters, and 1000 identities. Then we randomly select two poses and two sets of FLAME parameters for each identity, generating a total of 1000 image pairs. We calculate consistency metric using a pre-trained Arcface model [15] for each image pair and report the average result. Since the other baselines don't support in-plane (roll) head rotation, we further report APD* which only accounts for poses on yaw and pitch. Our method achieves the best performance on all metrics. Note that our model demonstrates significant improvements in FID, bringing animatable 3D GAN to the same level as unconditional 3D GANs (4.7 for EG3D [7]). For AED and APD, we also show superiority against baselines. Note that we still achieve the best pose consistency (0.019) when only considering yaw and pitch." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 405, + 397, + 417 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 405, + 397, + 417 + ], + "spans": [ + { + "bbox": [ + 306, + 405, + 397, + 417 + ], + "type": "text", + "content": "4.2. Ablation study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 424, + 545, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 424, + 545, + 532 + ], + "spans": [ + { + "bbox": [ + 304, + 424, + 545, + 532 + ], + "type": "text", + "content": "Static tri-planes. As suggested by the grey lines in Fig. 8, this baseline removes the static tri-planes " + }, + { + "bbox": [ + 304, + 424, + 545, + 532 + ], + "type": "inline_equation", + "content": "T_{static}" + }, + { + "bbox": [ + 304, + 424, + 545, + 532 + ], + "type": "text", + "content": " and entangles both dynamic and static components in " + }, + { + "bbox": [ + 304, + 424, + 545, + 532 + ], + "type": "inline_equation", + "content": "T_{uv}" + }, + { + "bbox": [ + 304, + 424, + 545, + 532 + ], + "type": "text", + "content": ". As illustrated in the first row of Fig. 6, the identities change when varying expressions. Since there are no explicit constraints for identity consistency, the model would be prone to unexpected entanglement between expression and identity. Tab. 2 shows a similar trend where removing " + }, + { + "bbox": [ + 304, + 424, + 545, + 532 + ], + "type": "inline_equation", + "content": "T_{static}" + }, + { + "bbox": [ + 304, + 424, + 545, + 532 + ], + "type": "text", + "content": " leads to worse identity consistency." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 533, + 545, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 545, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 545, + 700 + ], + "type": "text", + "content": "Mouth Synthesis. When removing the mouth synthesis module " + }, + { + "bbox": [ + 304, + 533, + 545, + 700 + ], + "type": "inline_equation", + "content": "G_{\\text{teeth}}" + }, + { + "bbox": [ + 304, + 533, + 545, + 700 + ], + "type": "text", + "content": ", we consider two altered choices: representing mouth features by " + }, + { + "bbox": [ + 304, + 533, + 545, + 700 + ], + "type": "inline_equation", + "content": "T_{\\text{static}}" + }, + { + "bbox": [ + 304, + 533, + 545, + 700 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 304, + 533, + 545, + 700 + ], + "type": "inline_equation", + "content": "T_{uv}" + }, + { + "bbox": [ + 304, + 533, + 545, + 700 + ], + "type": "text", + "content": ", named w/o " + }, + { + "bbox": [ + 304, + 533, + 545, + 700 + ], + "type": "inline_equation", + "content": "G_{\\text{teeth}}(s)" + }, + { + "bbox": [ + 304, + 533, + 545, + 700 + ], + "type": "text", + "content": " (red lines in Fig. 8) and w/o " + }, + { + "bbox": [ + 304, + 533, + 545, + 700 + ], + "type": "inline_equation", + "content": "G_{\\text{teeth}}(uv)" + }, + { + "bbox": [ + 304, + 533, + 545, + 700 + ], + "type": "text", + "content": " (blue lines in Fig. 8), respectively. The first baseline, illustrated in the second row of Fig. 6, suffers from 'hole' artifacts of mouth. This is because inferior teeth move along with the jaw rotations and thus cannot be modeled by static features. The second baseline modeling teeth with " + }, + { + "bbox": [ + 304, + 533, + 545, + 700 + ], + "type": "inline_equation", + "content": "T_{uv}" + }, + { + "bbox": [ + 304, + 533, + 545, + 700 + ], + "type": "text", + "content": " also leads to an unreasonable mouth interior since the FLAME template doesn't account for teeth area and the neural textures for teeth would be sampled from other unrelated areas leading to artifacts. Quantitatively, both baselines without " + }, + { + "bbox": [ + 304, + 533, + 545, + 700 + ], + "type": "inline_equation", + "content": "G_{\\text{teeth}}" + }, + { + "bbox": [ + 304, + 533, + 545, + 700 + ], + "type": "text", + "content": " show significant degradation in AED and APD." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 701, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 701, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 306, + 701, + 545, + 712 + ], + "type": "text", + "content": "Deformation-aware discriminator. Tab. 2 demonstrates" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20997" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 95, + 178, + 155 + ], + "blocks": [ + { + "bbox": [ + 75, + 79, + 100, + 88 + ], + "lines": [ + { + "bbox": [ + 75, + 79, + 100, + 88 + ], + "spans": [ + { + "bbox": [ + 75, + 79, + 100, + 88 + ], + "type": "text", + "content": "Source" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 60, + 95, + 178, + 155 + ], + "lines": [ + { + "bbox": [ + 60, + 95, + 178, + 155 + ], + "spans": [ + { + "bbox": [ + 60, + 95, + 178, + 155 + ], + "type": "image", + "image_path": "33d78dd48624762e1e3900456db4959f77cff16ac5f4f4d46f8d0fdbb74e7458.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 181, + 95, + 355, + 156 + ], + "blocks": [ + { + "bbox": [ + 129, + 79, + 161, + 87 + ], + "lines": [ + { + "bbox": [ + 129, + 79, + 161, + 87 + ], + "spans": [ + { + "bbox": [ + 129, + 79, + 161, + 87 + ], + "type": "text", + "content": "Inversion" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 247, + 79, + 284, + 87 + ], + "lines": [ + { + "bbox": [ + 247, + 79, + 284, + 87 + ], + "spans": [ + { + "bbox": [ + 247, + 79, + 284, + 87 + ], + "type": "text", + "content": "Animation" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 181, + 95, + 355, + 156 + ], + "lines": [ + { + "bbox": [ + 181, + 95, + 355, + 156 + ], + "spans": [ + { + "bbox": [ + 181, + 95, + 355, + 156 + ], + "type": "image", + "image_path": "06f12f87eb4149f216329656db72dcbc8a7d140cf3dfb717936ca9d15eb74d18.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 357, + 73, + 518, + 155 + ], + "blocks": [ + { + "bbox": [ + 357, + 73, + 518, + 155 + ], + "lines": [ + { + "bbox": [ + 357, + 73, + 518, + 155 + ], + "spans": [ + { + "bbox": [ + 357, + 73, + 518, + 155 + ], + "type": "image", + "image_path": "b835e071ba2f83a9cef41d244d51c1b9d4c73d4555dc887a9c0dc7e946a229e5.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 60, + 158, + 178, + 217 + ], + "blocks": [ + { + "bbox": [ + 60, + 158, + 178, + 217 + ], + "lines": [ + { + "bbox": [ + 60, + 158, + 178, + 217 + ], + "spans": [ + { + "bbox": [ + 60, + 158, + 178, + 217 + ], + "type": "image", + "image_path": "51195a567cff7979293e60e15960e03e8e69678b96fc431a268c2f0b97dd8dcb.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 225, + 544, + 248 + ], + "lines": [ + { + "bbox": [ + 46, + 225, + 544, + 248 + ], + "spans": [ + { + "bbox": [ + 46, + 225, + 544, + 248 + ], + "type": "text", + "content": "Figure 7. Applications of our model. We use PTI [53] to fit 3D-aware avatars for real portraits and animate them with sampled video clips. Furthermore, we leverage StyleGAN-NADA [21] into 3D settings and adapt these avatars into textually-prescribed domains." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 180, + 158, + 355, + 217 + ], + "blocks": [ + { + "bbox": [ + 180, + 158, + 355, + 217 + ], + "lines": [ + { + "bbox": [ + 180, + 158, + 355, + 217 + ], + "spans": [ + { + "bbox": [ + 180, + 158, + 355, + 217 + ], + "type": "image", + "image_path": "e83a8e6a044c63386ffa5984d7aafe634c56ff2bbf9f3b7b7947288886843380.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 358, + 158, + 533, + 217 + ], + "blocks": [ + { + "bbox": [ + 358, + 158, + 533, + 217 + ], + "lines": [ + { + "bbox": [ + 358, + 158, + 533, + 217 + ], + "spans": [ + { + "bbox": [ + 358, + 158, + 533, + 217 + ], + "type": "image", + "image_path": "bf09cbe0ec110d786ce9195b4a62a8ee47f138e113020020216865a347b56f3d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 61, + 258, + 154, + 421 + ], + "blocks": [ + { + "bbox": [ + 61, + 258, + 154, + 421 + ], + "lines": [ + { + "bbox": [ + 61, + 258, + 154, + 421 + ], + "spans": [ + { + "bbox": [ + 61, + 258, + 154, + 421 + ], + "type": "image", + "image_path": "1f35502efc539fa94056ed3f81f193815f4d51436468456966ac2b8da92d4bbc.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 77, + 426, + 127, + 434 + ], + "lines": [ + { + "bbox": [ + 77, + 426, + 127, + 434 + ], + "spans": [ + { + "bbox": [ + 77, + 426, + 127, + 434 + ], + "type": "text", + "content": "(a) Default Structure" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 157, + 258, + 277, + 422 + ], + "blocks": [ + { + "bbox": [ + 157, + 258, + 277, + 422 + ], + "lines": [ + { + "bbox": [ + 157, + 258, + 277, + 422 + ], + "spans": [ + { + "bbox": [ + 157, + 258, + 277, + 422 + ], + "type": "image", + "image_path": "1ec1bb7919d793fe0c6701920479a6b6ccf854feba233e084144bb915b512e1a.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 190, + 426, + 242, + 434 + ], + "lines": [ + { + "bbox": [ + 190, + 426, + 242, + 434 + ], + "spans": [ + { + "bbox": [ + 190, + 426, + 242, + 434 + ], + "type": "text", + "content": "(b) Altered Structures" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 446, + 287, + 468 + ], + "lines": [ + { + "bbox": [ + 46, + 446, + 287, + 468 + ], + "spans": [ + { + "bbox": [ + 46, + 446, + 287, + 468 + ], + "type": "text", + "content": "Figure 8. Illustrations of the proposed three variants of our model for ablation study." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 471, + 287, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 471, + 287, + 530 + ], + "spans": [ + { + "bbox": [ + 46, + 471, + 287, + 530 + ], + "type": "text", + "content": "that the deformation-aware discriminator with " + }, + { + "bbox": [ + 46, + 471, + 287, + 530 + ], + "type": "inline_equation", + "content": "I_{\\text{synthetic}}" + }, + { + "bbox": [ + 46, + 471, + 287, + 530 + ], + "type": "text", + "content": " input improves both animation accuracy and identity consistency, at the negligible expense of slightly reduced image quality. In Fig. 6, we see that this design shows a better reconstruction of detailed expressions, e.g. eye blinks." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 47, + 539, + 129, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 539, + 129, + 552 + ], + "spans": [ + { + "bbox": [ + 47, + 539, + 129, + 552 + ], + "type": "text", + "content": "4.3. Applications" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 557, + 287, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 287, + 641 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 287, + 641 + ], + "type": "text", + "content": "One-shot portrait animation. Fig. 7 shows the application of our model for one-shot head avatars. The learned generative animatbale 3D representation with expressive latent space can serve as a strong 3D prior for high-fidelity single-view 3D reconstruction and animation. Note that we can generate natural and consistent animations without video data training." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "content": "Animatable 3D-aware stylization. Inspired by IDE-3D [58], we incorporate 2D CLIP-guided style tranfer methods [21] with our animatable 3D representation for 3D-aware portrait stylization. The right three columns of Fig. 7 show examples of text-driven, stylized portrait animation. Specifically, to adapt a pre-trained model through" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 259, + 546, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 259, + 546, + 380 + ], + "spans": [ + { + "bbox": [ + 304, + 259, + 546, + 380 + ], + "type": "text", + "content": "only a textual prompt, we optimize the generator with two kinds of CLIP-based guidance [21]. However, leveraging text-guided 2D methods directly into the 3D setting is challenging as it tends to break 3D awareness and deformation awareness inherited in the generator parameters. To this end, we make some necessary modifications to the training framework. Please refer to the supplemental material for details. As shown in Fig. 7, we achieve high-quality 3D-aware stylized portrait synthesis with preserving well properties (i.e. 3D consistency and accurate animation)." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 390, + 466, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 390, + 466, + 403 + ], + "spans": [ + { + "bbox": [ + 306, + 390, + 466, + 403 + ], + "type": "text", + "content": "5. Limitations and future work" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 411, + 545, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 411, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 304, + 411, + 545, + 506 + ], + "type": "text", + "content": "Though our approach enables reasonable extrapolation on some rare expressions, it struggles to model some others with full consistency, such as one-side mouth up, sticking tongue out, etc. Besides, the rendered portraits still suffer from slight texture flickering caused by 2D superresolution. We leave both for future work. Furthermore, our model has the potential to provide a strong 3D prior to accelerating person-specific avatar reconstruction." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 517, + 378, + 530 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 517, + 378, + 530 + ], + "spans": [ + { + "bbox": [ + 306, + 517, + 378, + 530 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 538, + 545, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 538, + 545, + 646 + ], + "spans": [ + { + "bbox": [ + 304, + 538, + 545, + 646 + ], + "type": "text", + "content": "We have presented Next3D, a novel animatable 3D representation for unsupervised learning of high-quality and 3D-consistent virtual facial avatars from unstructured 2D images. Our approach has pushed the frontier of photorealistic animatabale 3D-aware image synthesis. Serving as a strong 3D prior, we believe our learned 3D representation will boost a series of downstream applications including 3D-aware one-shot facial avatars and animatable out-of-domain avatar generation." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 657, + 408, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 657, + 408, + 670 + ], + "spans": [ + { + "bbox": [ + 306, + 657, + 408, + 670 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 306, + 677, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 677, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 306, + 677, + 545, + 712 + ], + "type": "text", + "content": "This paper is supported by National Key RD Program of China (2022YFF0902200) and the NSFC project No.62125107." + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20998" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] ShahRukh Athar, Zexiang Xu, Kalyan Sunkavalli, Eli Shechtman, and Zhixin Shu. Rignerf: Fully controllable neural 3d portraits. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20364-20373, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 148, + 288, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 288, + 192 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 288, + 192 + ], + "type": "text", + "content": "[2] Hendrik Baatz, Jonathan Granskog, Marios Papas, Fabrice Rousselle, and Jan Novák. Nerf-tex: Neural reflectance field textures. In Computer Graphics Forum. Wiley Online Library, 2021. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 194, + 288, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 194, + 288, + 236 + ], + "spans": [ + { + "bbox": [ + 53, + 194, + 288, + 236 + ], + "type": "text", + "content": "[3] Alexander W. Bergman, Petr Kellnhofer, Wang Yifan, Eric R. Chan, David B. Lindell, and Gordon Wetzstein. Generative neural articulated radiance fields. In NeurIPS, 2022. 2, 5, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 239, + 288, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 239, + 288, + 294 + ], + "spans": [ + { + "bbox": [ + 53, + 239, + 288, + 294 + ], + "type": "text", + "content": "[4] Volker Blanz and Thomas Vetter. A morphable model for the synthesis of 3d faces. In Proceedings of the 26th Annual Conference on Computer Graphics and Interactive Techniques, SIGGRAPH '99, page 187-194, USA, 1999. ACM Press/Addison-Wesley Publishing Co. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 296, + 288, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 296, + 288, + 340 + ], + "spans": [ + { + "bbox": [ + 53, + 296, + 288, + 340 + ], + "type": "text", + "content": "[5] Andrei Burov, Matthias Nießner, and Justus Thies. Dynamic surface function networks for clothed human bodies. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10754-10764, 2021. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 342, + 288, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 342, + 288, + 396 + ], + "spans": [ + { + "bbox": [ + 53, + 342, + 288, + 396 + ], + "type": "text", + "content": "[6] Shengqu Cai, Anton Obukhov, Dengxin Dai, and Luc Van Gool. Pix2nerf: Unsupervised conditional p-gan for single image to neural radiance fields translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3981-3990, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 398, + 288, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 398, + 288, + 474 + ], + "spans": [ + { + "bbox": [ + 53, + 398, + 288, + 474 + ], + "type": "text", + "content": "[7] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In Proceedings of the IEEE International Conference on Computer Vision (CVPR), 2021. 2, 4, 5, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 476, + 288, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 476, + 288, + 531 + ], + "spans": [ + { + "bbox": [ + 53, + 476, + 288, + 531 + ], + "type": "text", + "content": "[8] Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 533, + 288, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 533, + 288, + 587 + ], + "spans": [ + { + "bbox": [ + 53, + 533, + 288, + 587 + ], + "type": "text", + "content": "[9] Xu Chen, Tianjian Jiang, Jie Song, Jinlong Yang, Michael J Black, Andreas Geiger, and Otmar Hilliges. gdna: Towards generative detailed neural avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20427-20437, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 589, + 287, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 589, + 287, + 621 + ], + "spans": [ + { + "bbox": [ + 48, + 589, + 287, + 621 + ], + "type": "text", + "content": "[10] Yuedong Chen, Qianyi Wu, Chuanxia Zheng, Tat-Jen Cham, and Jianfei Cai. Sem2nerf: Converting single-view semantic masks to neural radiance fields. In ECCV, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "type": "text", + "content": "[11] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5939-5948, 2019. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[12] Hang Dai, Nick Pears, William Smith, and Christian Duncan. Statistical modeling of craniofacial shape and texture. International Journal of Computer Vision, 128(2):547-571, 2020. 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "text", + "content": "[13] Yu Deng, Jiaolong Yang, Dong Chen, Fang Wen, and Tong Xin. Disentangled and controllable face image generation via 3d imitative-contrastive learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 129, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 173 + ], + "type": "text", + "content": "[14] Yu Deng, Jiaolong Yang, Jianfeng Xiang, and Xin Tong. Gram: Generative radiance manifolds for 3d-aware image generation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 175, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 239 + ], + "type": "text", + "content": "[15] Yu Deng, Jiaolong Yang, Sicheng Xu, Dong Chen, Yunde Jia, and Xin Tong. Accurate 3d face reconstruction with weakly-supervised learning: From single image to image set. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 0–0, 2019. 5, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 241, + 545, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 241, + 545, + 295 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 545, + 295 + ], + "type": "text", + "content": "[16] Michail Christos Doukas, Stefanos Zafeiriou, and Viktoriia Sharmanska. Headgan: One-shot neural head synthesis and editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14398-14407, 2021. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 297, + 545, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 297, + 545, + 341 + ], + "spans": [ + { + "bbox": [ + 307, + 297, + 545, + 341 + ], + "type": "text", + "content": "[17] Yao Feng, Haiwen Feng, Michael J Black, and Timo Bolkart. Learning an animatable detailed 3d face model from inthe-wild images. ACM Transactions on Graphics (ToG), 40(4):1-13, 2021. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 342, + 545, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 342, + 545, + 397 + ], + "spans": [ + { + "bbox": [ + 307, + 342, + 545, + 397 + ], + "type": "text", + "content": "[18] Ohad Fried, Ayush Tewari, Michael Zollhöfer, Adam Finkelstein, Eli Shechtman, Dan B Goldman, Kyle Genova, Zeyu Jin, Christian Theobalt, and Maneesh Agrawala. Text-based editing of talking-head video. ACM Transactions on Graphics (TOG), 38(4):1-14, 2019. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 399, + 545, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 545, + 431 + ], + "type": "text", + "content": "[19] Matheus Gadelha, Subhransu Maji, and Rui Wang. 3d shape induction from 2d views of multiple objects. In International Conference on 3D Vision (3DV), 2017. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 433, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 433, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 307, + 433, + 545, + 487 + ], + "type": "text", + "content": "[20] Guy Gafni, Justus Thies, Michael Zollhofer, and Matthias Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8649-8658, 2021. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 489, + 545, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 489, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 307, + 489, + 545, + 533 + ], + "type": "text", + "content": "[21] Rinon Gal, Or Patashnik, Haggai Maron, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. Stylegan-nada: Clip-guided domain adaptation of image generators. ACM Transactions on Graphics (TOG), 41(4):1-13, 2022. 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 534, + 545, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 534, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 307, + 534, + 545, + 589 + ], + "type": "text", + "content": "[22] Baris Gecer, Binod Bhattacharai, Josef Kittler, and Tae-Kyun Kim. Semi-supervised adversarial learning to generate photorealistic face images of new identities from 3d morphable model. In Proceedings of the European conference on computer vision (ECCV), pages 217-234, 2018. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 590, + 545, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 590, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 307, + 590, + 545, + 645 + ], + "type": "text", + "content": "[23] Baris Gecer, Stylianos Ploumpis, Irene Kotsia, and Stefanos Zafeiriou. Ganfit: Generative adversarial network fitting for high fidelity 3d face reconstruction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1155-1164, 2019. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 647, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 690 + ], + "type": "text", + "content": "[24] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 692, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 692, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 692, + 545, + 713 + ], + "type": "text", + "content": "[25] Philip-William Grassal, Malte Prinzler, Titus Leistner, Carsten Rother, Matthias Nießner, and Justus Thies. Neural" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20999" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 66, + 72, + 288, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 288, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 288, + 106 + ], + "type": "text", + "content": "head avatars from monocular rgb videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18653-18664, 2022. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 288, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 288, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 288, + 173 + ], + "type": "text", + "content": "[26] Artur Grigorev, Karim Iskakov, Anastasia Ianina, Renat Bashirov, Ilya Zakharkin, Alexander Vakhitov, and Victor Lempitsky. Stylepeople: A generative model of fullbody human avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5151-5160, 2021. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 175, + 288, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 288, + 218 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 288, + 218 + ], + "type": "text", + "content": "[27] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. Stylenerf: A style-based 3d aware generator for high-resolution image synthesis. In International Conference on Learning Representations, 2022. 2, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 220, + 288, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 220, + 288, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 220, + 288, + 262 + ], + "type": "text", + "content": "[28] Philipp Henzler, Niloy J Mitra, and Tobias Ritschel. Escaping plato's cave: 3d shape from adversarial rendering. In Proceedings of the IEEE International Conference on Computer Vision (CVPR), 2019. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 264, + 288, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 264, + 288, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 264, + 288, + 319 + ], + "type": "text", + "content": "[29] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In Advances in Neural Information Processing Systems (NeurIPS), 2017. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 320, + 288, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 320, + 288, + 364 + ], + "spans": [ + { + "bbox": [ + 48, + 320, + 288, + 364 + ], + "type": "text", + "content": "[30] Fangzhou Hong, Zhaoxi Chen, Yushi LAN, Liang Pan, and Ziwei Liu. EVA3d: Compositional 3d human generation from 2d image collections. In International Conference on Learning Representations, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 365, + 288, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 365, + 288, + 420 + ], + "spans": [ + { + "bbox": [ + 48, + 365, + 288, + 420 + ], + "type": "text", + "content": "[31] Yang Hong, Bo Peng, Haiyao Xiao, Ligang Liu, and Juyong Zhang. Headnerf: A real-time nerf-based parametric head model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20374-20384, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 422, + 288, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 422, + 288, + 475 + ], + "spans": [ + { + "bbox": [ + 48, + 422, + 288, + 475 + ], + "type": "text", + "content": "[32] Angjoo Kanazawa, Shubham Tulsiani, Alexei A Efros, and Jitendra Malik. Learning category-specific mesh reconstruction from image collections. In Proceedings of the European Conference on Computer Vision (ECCV), pages 371-386, 2018. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 477, + 288, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 477, + 288, + 521 + ], + "spans": [ + { + "bbox": [ + 48, + 477, + 288, + 521 + ], + "type": "text", + "content": "[33] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4401-4410, 2019. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 523, + 288, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 523, + 288, + 577 + ], + "spans": [ + { + "bbox": [ + 48, + 523, + 288, + 577 + ], + "type": "text", + "content": "[34] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2020. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 578, + 288, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 288, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 288, + 633 + ], + "type": "text", + "content": "[35] Hyeongwoo Kim, Pablo Garrido, Ayush Tewari, Weipeng Xu, Justus Thies, Matthias Niessner, Patrick Pérez, Christian Richardt, Michael Zollhöfer, and Christian Theobalt. Deep video portraits. ACM Transactions on Graphics (TOG), 37(4):1-14, 2018. 2, 3, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 635, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 288, + 689 + ], + "type": "text", + "content": "[36] Jeong-gi Kwak, Yuanming Li, Dongsik Yoon, Donghyeon Kim, David Han, and Hanseok Ko. Injecting 3d perception of controllable nerf-gan into stylegan for editable portrait image synthesis. In Proceedings of European Conference on Computer Vision (ECCV), 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "type": "text", + "content": "[37] Henning Lange and J Nathan Kutz. Fc2t2: The fast continuous convolutional taylor transform with applications in vi" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "sion and graphics. arXiv preprint arXiv:2111.00110, 2021. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 97, + 545, + 130 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 97, + 545, + 130 + ], + "spans": [ + { + "bbox": [ + 307, + 97, + 545, + 130 + ], + "type": "text", + "content": "[38] Tianye Li, Timo Bolkart, Michael J Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4d scans. ACM Trans. Graph., 36(6):194-1, 2017. 3, 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 132, + 545, + 165 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 132, + 545, + 165 + ], + "spans": [ + { + "bbox": [ + 307, + 132, + 545, + 165 + ], + "type": "text", + "content": "[39] Connor Z Lin, David B Lindell, Eric R Chan, and Gordon Wetzstein. 3d gan inversion for controllable portrait image animation. arXiv preprint arXiv:2203.13441, 2022. 2, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 167, + 545, + 221 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 167, + 545, + 221 + ], + "spans": [ + { + "bbox": [ + 307, + 167, + 545, + 221 + ], + "type": "text", + "content": "[40] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In European Conference on Computer Vision (ECCV), 2020. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 224, + 545, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 224, + 545, + 278 + ], + "spans": [ + { + "bbox": [ + 307, + 224, + 545, + 278 + ], + "type": "text", + "content": "[41] Thu Nguyen-Phuoc, Chuan Li, Lucas Theis, Christian Richardt, and Yong-Liang Yang. Hologan: Unsupervised learning of 3d representations from natural images. In Proceedings of the IEEE International Conference on Computer Vision, 2019. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 280, + 545, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 280, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 307, + 280, + 545, + 335 + ], + "type": "text", + "content": "[42] Thu H Nguyen-Phuoc, Christian Richardt, Long Mai, Yongliang Yang, and Niloy Mitra. Blockgan: Learning 3d object-aware scene representations from unlabelled images. In Advances in Neural Information Processing Systems (NeurIPS), 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 338, + 545, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 338, + 545, + 382 + ], + "spans": [ + { + "bbox": [ + 307, + 338, + 545, + 382 + ], + "type": "text", + "content": "[43] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 384, + 545, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 384, + 545, + 429 + ], + "spans": [ + { + "bbox": [ + 307, + 384, + 545, + 429 + ], + "type": "text", + "content": "[44] Atsuhiro Noguchi, Xiao Sun, Stephen Lin, and Tatsuya Harada. Unsupervised learning of efficient geometry-aware neural articulated representations. In Proceedings of European Conference on Computer Vision (ECCV), 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 430, + 545, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 430, + 545, + 485 + ], + "spans": [ + { + "bbox": [ + 307, + 430, + 545, + 485 + ], + "type": "text", + "content": "[45] Michael Oechsle, Lars Mescheder, Michael Niemeyer, Thilo Strauss, and Andreas Geiger. Texture fields: Learning texture representations in function space. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4531-4540, 2019. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 487, + 545, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 487, + 545, + 552 + ], + "spans": [ + { + "bbox": [ + 307, + 487, + 545, + 552 + ], + "type": "text", + "content": "[46] Roy Or-El, Xuan Luo, Mengyi Shan, Eli Shechtman, Jeong Joon Park, and Ira Kemelmacher-Shlizerman. StyleSDF: High-Resolution 3D-Consistent Image and Geometry Generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 555, + 545, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 555, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 307, + 555, + 545, + 609 + ], + "type": "text", + "content": "[47] Xingang Pan, Xudong Xu, Chen Change Loy, Christian Theobalt, and Bo Dai. A shading-guided generative implicit model for shape-accurate 3d-aware image synthesis. Advances in Neural Information Processing Systems, 34:20002-20013, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 611, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 611, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 307, + 611, + 545, + 667 + ], + "type": "text", + "content": "[48] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 669, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 545, + 713 + ], + "type": "text", + "content": "[49] Songyou Peng, Michael Niemeyer, Lars Mescheder, Marc Pollefeys, and Andreas Geiger. Convolutional occupancy networks. In European Conference on Computer Vision, pages 523-540. Springer, 2020. 3" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21000" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "type": "text", + "content": "[50] Hanspeter Pfister, Matthias Zwicker, Jeroen Van Baar, and Markus Gross. Surfels: Surface elements as rendering primitives. In Proceedings of the 27th annual conference on Computer graphics and interactive techniques, pages 335-342, 2000. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 128, + 287, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 128, + 287, + 194 + ], + "spans": [ + { + "bbox": [ + 48, + 128, + 287, + 194 + ], + "type": "text", + "content": "[51] Stylianos Ploumpis, Evangelos Ververas, Eimear O'Sullivan, Stylianos Moschoglou, Haoyang Wang, Nick Pears, William AP Smith, Baris Gecer, and Stefanos Zafeiriou. Towards a complete 3d morphable model of the human head. IEEE transactions on pattern analysis and machine intelligence, 43(11):4142-4160, 2020. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 195, + 287, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 195, + 287, + 248 + ], + "spans": [ + { + "bbox": [ + 48, + 195, + 287, + 248 + ], + "type": "text", + "content": "[52] Yurui Ren, Ge Li, Yuanqi Chen, Thomas H Li, and Shan Liu. Pirenderer: Controllable portrait image generation via semantic neural rendering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13759-13768, 2021. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 250, + 287, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 250, + 287, + 282 + ], + "spans": [ + { + "bbox": [ + 48, + 250, + 287, + 282 + ], + "type": "text", + "content": "[53] Daniel Roich, Ron Mokady, Amit H Bermano, and Daniel Cohen-Or. Pivotal tuning for latent-based editing of real images. ACM Trans. Graph., 2021. 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 283, + 287, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 283, + 287, + 326 + ], + "spans": [ + { + "bbox": [ + 48, + 283, + 287, + 326 + ], + "type": "text", + "content": "[54] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. Graf: Generative radiance fields for 3d-aware image synthesis. In Advances in Neural Information Processing Systems (NeurIPS), 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 327, + 287, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 327, + 287, + 370 + ], + "spans": [ + { + "bbox": [ + 48, + 327, + 287, + 370 + ], + "type": "text", + "content": "[55] Katja Schwarz, Axel Sauer, Michael Niemeyer, Yiyi Liao, and Andreas Geiger. Voxgraf: Fast 3d-aware image synthesis with sparse voxel grids. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 372, + 287, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 372, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 48, + 372, + 287, + 426 + ], + "type": "text", + "content": "[56] Vincent Sitzmann, Justus Thies, Felix Heide, Matthias Nießner, Gordon Wetzstein, and Michael Zollhofer. Deepvoxels: Learning persistent 3d feature embeddings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2437-2446, 2019. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 426, + 287, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 426, + 287, + 469 + ], + "spans": [ + { + "bbox": [ + 48, + 426, + 287, + 469 + ], + "type": "text", + "content": "[57] Ivan Skorokhodov, Sergey Tulyakov, Yiqun Wang, and Peter Wonka. Epigraph: Rethinking training of 3d gans. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 471, + 287, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 471, + 287, + 524 + ], + "spans": [ + { + "bbox": [ + 48, + 471, + 287, + 524 + ], + "type": "text", + "content": "[58] Jingxiang Sun, Xuan Wang, Yichun Shi, Lizhen Wang, Jue Wang, and Yebin Liu. Ide-3d: Interactive disentangled editing for high-resolution 3d-aware portrait synthesis. ACM Transactions on Graphics (TOG), 41(6):1-10, 2022. 2, 5, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 526, + 287, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 526, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 48, + 526, + 287, + 579 + ], + "type": "text", + "content": "[59] Jingxiang Sun, Xuan Wang, Yong Zhang, Xiaoyu Li, Qi Zhang, Yebin Liu, and Jue Wang. Fenerf: Face editing in neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 581, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 581, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 581, + 287, + 635 + ], + "type": "text", + "content": "[60] Keqiang Sun, Shangzhe Wu, Zhaoyang Huang, Ning Zhang, Quan Wang, and HongSheng Li. Controllable 3d face synthesis with conditional generative occupancy fields. In Advances in Neural Information Processing Systems (NeurIPS), 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 636, + 287, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 287, + 679 + ], + "type": "text", + "content": "[61] Junshu Tang, Bo Zhang, Binxin Yang, Ting Zhang, Dong Chen, Lizhuang Ma, and Fang Wen. Explicitly controllable 3d-aware portrait generation. arXiv preprint arXiv:2209.05434, 2022. 2, 4, 5, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 681, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 714 + ], + "type": "text", + "content": "[62] Ayush Tewari, Mohamed Elgharib, Gaurav Bharaj, Florian Bernard, Hans-Peter Seidel, Patrick Pérez, Michael Zollhofer, and Christian Theobalt. Stylerig: Rigging stylegan" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "text", + "content": "for 3d control over portrait images. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 107, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 107, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 308, + 107, + 545, + 162 + ], + "type": "text", + "content": "[63] Ayush Tewari, Justus Thies, Ben Mildenhall, Pratul Srinivasan, Edgar Tretschk, W Yifan, Christoph Lassner, Vincent Sitzmann, Ricardo Martin-Brualla, Stephen Lombardi, et al. Advances in neural rendering. Computer Graphics Forum, 41(2):703-735, 2022. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 163, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 163, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 308, + 163, + 545, + 206 + ], + "type": "text", + "content": "[64] Justus Thies, Michael Zollhöfer, and Matthias Nießner. Deferred neural rendering: Image synthesis using neural textures. ACM Transactions on Graphics (TOG), 38(4):1-12, 2019. 3, 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 209, + 545, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 209, + 545, + 263 + ], + "spans": [ + { + "bbox": [ + 308, + 209, + 545, + 263 + ], + "type": "text", + "content": "[65] Justus Thies, Michael Zollhofer, Marc Stamminger, Christian Theobalt, and Matthias Nießner. Face2face: Real-time face capture and reenactment of rgb videos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2387-2395, 2016. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 265, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 265, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 308, + 265, + 545, + 330 + ], + "type": "text", + "content": "[66] Lizhen Wang, Zhiyuan Chen, Tao Yu, Chenguang Ma, Liang Li, and Yebin Liu. Faceverse: a fine-grained and detail-controllable 3d face morphable model from a hybrid dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20333-20342, 2022. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 332, + 545, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 332, + 545, + 386 + ], + "spans": [ + { + "bbox": [ + 308, + 332, + 545, + 386 + ], + "type": "text", + "content": "[67] Jiajun Wu, Chengkai Zhang, Tianfan Xue, Bill Freeman, and Josh Tenenbaum. Learning a probabilistic latent space of object shapes via 3d generative-adversarial modeling. Advances in neural information processing systems, 29, 2016. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 388, + 545, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 388, + 545, + 432 + ], + "spans": [ + { + "bbox": [ + 308, + 388, + 545, + 432 + ], + "type": "text", + "content": "[68] Yue Wu, Yu Deng, Jiaolong Yang, Fangyun Wei, Chen Qifeng, and Xin Tong. Anifacegan: Animatable 3d-aware face image generation for video avatars. In Advances in Neural Information Processing Systems, 2022. 2, 4, 5, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 434, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 434, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 308, + 434, + 545, + 487 + ], + "type": "text", + "content": "[69] Qiangeng Xu, Weiyue Wang, Duygu Ceylan, Radomir Mech, and Ulrich Neumann. Disn: Deep implicit surface network for high-quality single-view 3d reconstruction. Advances in Neural Information Processing Systems, 32, 2019. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 490, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 490, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 308, + 490, + 545, + 544 + ], + "type": "text", + "content": "[70] Sicheng Xu, Jiaolong Yang, Dong Chen, Fang Wen, Yu Deng, Yunde Jia, and Xin Tong. Deep 3d portrait from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7710-7720, 2020. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 546, + 545, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 546, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 308, + 546, + 545, + 601 + ], + "type": "text", + "content": "[71] Yinghao Xu, Sida Peng, Ceyuan Yang, Yujun Shen, and Bolei Zhou. 3d-aware image synthesis via learning structural and textural representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18430-18439, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 602, + 545, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 602, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 308, + 602, + 545, + 645 + ], + "type": "text", + "content": "[72] Yang Xue, Yuheng Li, Krishna Kumar Singh, and Yong Jae Lee. Giraffe hd: A high-resolution 3d-aware generative model. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 647, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 712 + ], + "type": "text", + "content": "[73] Fei Yin, Yong Zhang, Xiaodong Cun, Mingdeng Cao, Yanbo Fan, Xuan Wang, Qingyan Bai, Baoyuan Wu, Jue Wang, and Yujiu Yang. Styleheat: One-shot high-resolution editable talking face generation via pretrained stylegan. In Proceedings of the European conference on computer vision (ECCV), 2022. 2" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "21001" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 553 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "type": "text", + "content": "[74] Jianfeng Zhang, Zihang Jiang, Dingdong Yang, Hongyi Xu, Yichun Shi, Guoxian Song, Zhongcong Xu, Xinchao Wang, and Jiashi Feng. Avatargen: a 3d generative model for animatable human avatars. arXiv preprint arXiv:2208.00561, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 288, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 288, + 172 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 288, + 172 + ], + "type": "text", + "content": "[75] Jingbo Zhang, Xiaoyu Li, Ziyu Wan, Can Wang, and Jing Liao. Fdnerf: Few-shot dynamic neural radiance fields for face reconstruction and expression editing. arXiv preprint arXiv:2208.05751, 2022.3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 174, + 287, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 174, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 174, + 287, + 217 + ], + "type": "text", + "content": "[76] Jichao Zhang, E. Sangineto, Hao Tang, Aliaksandr Siarohin, Zhun Zhong, N. Sebe, and Wei Wang. 3d-aware semantic-guided generative model for human synthesis. In ECCV, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 219, + 287, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 219, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 219, + 287, + 262 + ], + "type": "text", + "content": "[77] Jichao Zhang, Aliaksandr Siarohin, Yahui Liu, Hao Tang, Nicu Sebe, and Wei Wang. Training and tuning generative neural radiance fields for attribute-conditional 3d-aware face generation. arXiv preprint arXiv:2208.12550, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 264, + 287, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 264, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 264, + 287, + 319 + ], + "type": "text", + "content": "[78] Xuanmeng Zhang, Zhedong Zheng, Daiheng Gao, Bang Zhang, Pan Pan, and Yi Yang. Multi-view consistent generative adversarial networks for 3d-aware image synthesis. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 319, + 287, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 319, + 287, + 373 + ], + "spans": [ + { + "bbox": [ + 48, + 319, + 287, + 373 + ], + "type": "text", + "content": "[79] Xiaoming Zhao, Fangchang Ma, David Guera, Zhile Ren, Alexander G. Schwing, and Alex Colburn. Generative multiplane images: Making a 2d gan 3d-aware. In Proceedings of European Conference on Computer Vision (ECCV), 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 375, + 287, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 375, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 48, + 375, + 287, + 430 + ], + "type": "text", + "content": "[80] Yufeng Zheng, Victoria Fernández Abrevaya, Marcel C Bühler, Xu Chen, Michael J Black, and Otmar Hilliges. Im avatar: Implicit morphable head avatars from videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13545-13555, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 431, + 287, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 431, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 48, + 431, + 287, + 464 + ], + "type": "text", + "content": "[81] Peng Zhou, Lingxi Xie, Bingbing Ni, and Qi Tian. Cips-3d: A 3d-aware generator of gans based on conditionally-independent pixel synthesis. arXiv:2110.09788, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 465, + 287, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 465, + 287, + 520 + ], + "spans": [ + { + "bbox": [ + 48, + 465, + 287, + 520 + ], + "type": "text", + "content": "[82] Jun-Yan Zhu, Zhoutong Zhang, Chengkai Zhang, Jiajun Wu, Antonio Torralba, Josh Tenenbaum, and Bill Freeman. Visual object networks: Image generation with disentangled 3d representations. In Advances in Neural Information Processing Systems (NeurIPS), 2018. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 521, + 287, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 521, + 287, + 553 + ], + "spans": [ + { + "bbox": [ + 48, + 521, + 287, + 553 + ], + "type": "text", + "content": "[83] Peiye Zhuang, Liqian Ma, Oluwasanmi Koyejo, and Alexander Schwing. Controllable radiance fields for dynamic face synthesis. In Proc. 3DV, 2022. 2" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21002" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Nighttime Smartphone Reflective Flare Removal Using Optical Center Symmetry Prior/d34edf55-17a4-4aa7-b0c0-de157905c27a_content_list.json b/2023/Nighttime Smartphone Reflective Flare Removal Using Optical Center Symmetry Prior/d34edf55-17a4-4aa7-b0c0-de157905c27a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9e524a5398c3713baa3cf08c759913e49181b70a --- /dev/null +++ b/2023/Nighttime Smartphone Reflective Flare Removal Using Optical Center Symmetry Prior/d34edf55-17a4-4aa7-b0c0-de157905c27a_content_list.json @@ -0,0 +1,2272 @@ +[ + { + "type": "text", + "text": "Nighttime Smartphone Reflective Flare Removal Using Optical Center Symmetry Prior", + "text_level": 1, + "bbox": [ + 238, + 130, + 733, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yuekun Dai Yihang Luo Shangchen Zhou Chongyi Li* Chen Change Loy S-Lab, Nanyang Technological University", + "bbox": [ + 169, + 203, + 799, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{YDAI005, c200211, s200094, chongyi.li, ccloy}@ntu.edu.sg", + "bbox": [ + 232, + 241, + 736, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://ykdai.github.io/projects/BracketFlare", + "bbox": [ + 281, + 258, + 681, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 287, + 313, + 303 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Reflective flare is a phenomenon that occurs when light reflects inside lenses, causing bright spots or a \"ghosting effect\" in photos, which can impact their quality. Eliminating reflective flare is highly desirable but challenging. Many existing methods rely on manually designed features to detect these bright spots, but they often fail to identify reflective flares created by various types of light and may even mistakenly remove the light sources in scenarios with multiple light sources. To address these challenges, we propose an optical center symmetry prior, which suggests that the reflective flare and light source are always symmetrical around the lens's optical center. This prior helps to locate the reflective flare's proposal region more accurately and can be applied to most smartphone cameras. Building on this prior, we create the first reflective flare removal dataset called BracketFlare, which contains diverse and realistic reflective flare patterns. We use continuous bracketing to capture the reflective flare pattern in the underexposed image and combine it with a normally exposed image to synthesize a pair of flare-corrupted and flare-free images. With the dataset, neural networks can be trained to remove the reflective flares effectively. Extensive experiments demonstrate the effectiveness of our method on both synthetic and real-world datasets.", + "bbox": [ + 76, + 320, + 473, + 681 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 698, + 209, + 714 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Lens flare artifacts can occur when a strong light enters the camera's field of view, causing scattering and reflection within the lenses. Unlike radial-shaped scattering flares, reflective flares appear as groups of bright spots [5, 7] or a \"ghosting effect\" like the examples in Fig. 1. With multiple reflections, the light source creates multiple projections at various positions behind the lens. If the projection is far from the focal plane, the defocused effect can brighten the whole sensor and degrade the photo's quality as shown in Fig. 2. In near-focus situations, images of the light source", + "bbox": [ + 75, + 724, + 470, + 876 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6550e8c0093d203935d90d2312c49717f38ee94b2f8313da514f0a03d6d41531.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 290, + 643, + 382 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6a2d7d55295dd1bda06f1347e5622d3139648b025bc507ac815583c9e9b3b57d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 383, + 642, + 474 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/db046f38b982a1a71dac44a5a8b723976beb85ad2768ba40d05efe3fb1064804.jpg", + "image_caption": [ + "(a) iPhone 13 Pro" + ], + "image_footnote": [], + "bbox": [ + 506, + 476, + 642, + 566 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/aa72b59fb4ae6537452c603ca696e6268b35c3cf36362e84a3deec27e97e23b6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 643, + 290, + 763, + 382 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3f957143ddac45f578f5cee0a8768434c0ea7afb57bdf90a514c705ffc6997db.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 643, + 383, + 763, + 474 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c7cbf74198cd8f53741fb18de3446471d90cfba1288b0541eaeafbce1116a92d.jpg", + "image_caption": [ + "(b) Huawei P40" + ], + "image_footnote": [], + "bbox": [ + 643, + 476, + 763, + 566 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4696c233bba023fd7d15fd38e4db745a7c2491186093e60284533358cc7363f8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 764, + 290, + 883, + 382 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/170f32acdf35ba7a862153536978978c671f1632e48a74b74f0c0c725306ab66.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 764, + 383, + 883, + 474 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ebeaf9b3ad5a75adf7db2d1c95926f6e12a918fdca2b156ad57d4dab3b3376f8.jpg", + "image_caption": [ + "(c) Sony ILCE-6400", + "Figure 1. Reflective flare removal with our proposed dataset and prior. The network trained on our dataset can effectively separate the reflective flare and flare-free images from a flare-corrupted image. Our method can generalize well to different types of smartphones and even cameras with specific types of lenses." + ], + "image_footnote": [], + "bbox": [ + 764, + 476, + 883, + 566 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "can result in a line of unfocused, bright blobs on the captured photo.", + "bbox": [ + 498, + 669, + 890, + 700 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Applying anti-reflective (AR) coating is a common technique for reducing internal reflections [22]. The coating is effective for unfocused spots with low luminance, but cannot mitigate a bright spot with high luminance at the focal plane. Even with a well-designed lens system, a bright reflective spot may remain in the photo, known as the ghosting effect [10]. Reflective flare is common on all types of smartphone lenses, and certain types of smartphones (e.g., the iPhone series) can produce a large area of inverted reflective flare from neon lights or large, luminous standing signs, as shown in Fig. 1. The phenomenon becomes more pronounced when capturing a photo at night with multiple lights in the scene. Numerous bright spots and inverted re", + "bbox": [ + 496, + 704, + 892, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 94, + 886, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "20783", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "flections manifest and propagate in different directions with the lens orientation. Therefore, reflective flare removal algorithms are highly desirable.", + "bbox": [ + 76, + 90, + 468, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Removing reflective flare is a difficult task due to the vast diversity of reflective flares that can be caused by different types of light sources and lenses. Reflective flares often appear as bright spots that can resemble streetlights from a distance, making it challenging to differentiate between reflective flares and other light sources. Current methods for removing flare spots mainly rely on feature detection techniques like Speeded-Up Robust Features [2] (SURF) or Scale-Invariant Feature Transform [12] (SIFT) to extract spot regions [1, 3, 19]. These methods, however, struggle to effectively detect blobs on complex backgrounds and only work for specific spot patterns. Moreover, these methods are ineffective in removing reflective flares caused by lamp tubes and LED matrices. Many learning-based methods and datasets for reflective flare removal have been proposed [5, 15, 18, 22]. Although these datasets contain subsets of reflective flares, they are not sufficient for good generalization. For instance, the reflective flare subset from Wu et al. [22] is captured using the same lens and light source, making it difficult to generalize to different flares in real-world scenarios. Dai et al. [5] propose a synthetic dataset with 10 types of reflective flares, but training on this dataset tends to mislead a network to erroneously remove other tiny bright spots like street lights and bright windows. As a result, it remains challenging for existing algorithms to accurately detect and remove various types of reflective flares while preserving details in other regions.", + "bbox": [ + 76, + 138, + 472, + 547 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To overcome the challenges of detecting and removing reflective flares, we propose a novel optical center symmetry prior, which suggests that the main spot of the reflective flare and the light source are always symmetrical around the lens's optical center in captured images. This prior applies to most smartphone cameras and some professional cameras. With this prior, we can easily detect the possible locations of reflective flares. We also found that the flare spots always have the same patterns as the brightest regions of the light source. Therefore, we can synthesize reflective flares by using the light source's patterns at low exposure. Based on these characteristics of reflective flares, we develop the first reflective flare removal dataset, named BracketFlare. Specifically, we use continuous bracketing to capture the light source's pattern on an underexposed image and select the normally exposed image to create the light source and the natural image. Next, we rotate the light source pattern 180 degrees around the optical center, apply blur and color augmentation, and use it to create synthetic reflective flares. We can add the reflective flare to the normally exposed image to produce paired flare-corrupted and flare-free images.", + "bbox": [ + 76, + 550, + 472, + 867 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Thanks to the optical center symmetry prior, we further propose an end-to-end pipeline for reflective flare re", + "bbox": [ + 76, + 869, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a5c94693cc237f600ed1a2d8c35288496b35992a6debe28279e859ef8a2afb20.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 526, + 90, + 699, + 223 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3699cdd3c26ddf177d851d381fe7068b686880a3c840cac5358bc21d8b3e68fd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 702, + 90, + 874, + 223 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/36baeab3108a00989db2c702a828df4ad8591c55495ee080ab0fbdfefc3dec07.jpg", + "image_caption": [ + "(a) w/o AR coating", + "(b) With AR coating", + "(c) Description of reflective flare formation", + "Figure 2. Schematic diagram of reflective flare formation. Reflective flares are generated by multiple reflections inside the lens. In (c), yellow lines show the generation principle of the aperture-shaped reflective flare. Red lines represent the formation of the bright spot. Since the unfocused reflective flare (yellow line, e.g., (a)) can be effectively alleviated by AR coating. In this work, we mainly focus on removing the focused reflective flare (red line, e.g., (b)) from night images." + ], + "image_footnote": [], + "bbox": [ + 535, + 238, + 880, + 372 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "moral, which can easily apply standard neural networks designed for image restoration. Specifically, to demonstrate the effectiveness of our dataset and pipeline, we train two CNN-based networks [4, 24] and two Transformer-based networks [21, 23] as baselines. These models trained on our dataset effectively suppress the reflective flare and generalize well to real-world scenarios. The contributions of our work are as follows:", + "bbox": [ + 496, + 511, + 890, + 630 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a new optical center symmetry prior that uncovers crucial properties of reflective flares on smartphones' cameras, including their locations of occurrence and patterns.", + "- Based on this prior, we explore a new method for constructing a reflective flare removal dataset called BracketFlare. This dataset includes various types of light sources and diverse indoor and outdoor scenes, providing a solid foundation for this task.", + "- We design an end-to-end reflective flare removal pipeline that incorporates the optical center symmetry prior to both training data generation and network structures. Extensive experiments demonstrate the essential role of our dataset and prior in reflective flare removal, which effectively generalizes to different night scenes" + ], + "bbox": [ + 517, + 640, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "20784", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 89, + 218, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Scattering flare removal. Most flare removal methods mainly focus on removing the scattering flares and glare effects. Based on the glow smooth prior, unsupervised methods [8, 16] are proposed to separate the glare effect from the flare-corrupted images. For supervised flare removal, Wu et al. [22] and Dai et al. [5] have respectively proposed two flare image datasets. These flare images can be added to flare-free images to synthesize the flare-corrupted and flare-free image pairs, making training an end-to-end flare removal network possible. Due to the difficulty in collecting real paired data, Qiao et al. [15] designed a pipeline to train the flare removal network with unpaired data following the idea of Cycle-GAN [27]. Besides, artifacts from underdisplay cameras [6] and nighttime haze [11] share similar attributes with the scattering flares. Thus, the related methods can also alleviate scattering flares artifacts.", + "bbox": [ + 75, + 114, + 472, + 358 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Reflective flare removal. Traditional reflective flare removal is typically accomplished through feature detection and image inpainting. Flare detection in these approaches [1,3,13,19] assumes that reflective flares are bright spots with specific shapes. Such an assumption is fragile given nighttime situations with diverse light sources. Since reflective flare removal can be viewed as a specific kind of reflection removal [26], one can adopt learning-based methods. However, existing datasets for this task, such as those proposed by Wu et al. [22], Dai et al. [5], and Qiao et al. [15], are not specifically designed for reflective flare removal and lack diversity in terms of light sources. These limitations hinder deep learning-based methods from effectively addressing real-world reflective flares, and may even lead to the removal of other tiny bright spots such as street lights.", + "bbox": [ + 75, + 364, + 472, + 609 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Optical Center Symmetry Prior", + "text_level": 1, + "bbox": [ + 76, + 623, + 369, + 642 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Definition for the prior", + "text_level": 1, + "bbox": [ + 76, + 648, + 290, + 666 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "When pointing the camera at a light source, strong light can sometimes create lens flares in photos. In well-designed lens systems with AR coatings, reflective flares appear as bright spots on the photo, rather than generating a light fog. Through our investigation of most smartphones' main cameras, we discover that the bright spots of reflective flares and light sources are always symmetrical around the lens's optical center, as shown in Fig. 3. By aligning the light source and optical center, we can easily locate the reflective flares, eliminating the need to search the entire image for possible occurrences. Besides, since the main bright spots are caused by focused reflective flares, they share the same attributes as the light source. For example, the reflective flares may display clear details for each small LED lamp bead, as shown in Fig. 3(b) and Fig. 3(c).", + "bbox": [ + 75, + 674, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c961b2927515414107657dd5f1d0de02cf21adda64988ff3431b1e591b8197ab.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 535, + 92, + 694, + 214 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d3f2083f0b815bbac4403daedcdb782dfc11fa7a6eb0918022bb53920cd5c358.jpg", + "image_caption": [ + "(b) iPhone 13 Pro" + ], + "image_footnote": [], + "bbox": [ + 696, + 92, + 854, + 214 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a704a6064c451fcb9230ee0b4e7ab48d5d335412ceb79c3aa832e10bc03c07b5.jpg", + "image_caption": [ + "(a) Samsung s7 edge", + "(c) Xiaomi Mi 8", + "Figure 3. Strong light source may project a bright reflective flare on the photo. The reflective flares and corresponding light sources are symmetric around the optical center for most smartphone cameras. The yellow cross represents the optical center." + ], + "image_footnote": [], + "bbox": [ + 535, + 226, + 694, + 349 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/3fcc450a33213d3edb3bf42fbee4882cfb11adef28126d4d57e48aefcd301ede.jpg", + "image_caption": [ + "(d) Huawei Mate30" + ], + "image_footnote": [], + "bbox": [ + 696, + 227, + 854, + 349 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Physical formulation", + "text_level": 1, + "bbox": [ + 500, + 438, + 694, + 454 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To illustrate the formation of optical center symmetry prior, we suppose that an optical ray $\\mathbf{r}$ is emitted from the light source, as shown in Fig. 4. This ray can be described as $\\mathbf{r} = [h\\theta ]^{\\top}$ . The states of this ray after each lens surface can be calculated by multiplying transfer matrices [10, 14]. We define the translation matrix at layer $i$ as $\\mathbf{T_i}$ , and the refraction matrix at spherical surface with radius $R_{i}$ and refractive index $n_i,n_{i + 1}$ as $\\mathbf{F_i}$ . For the reflection at the ith layer with radius $R_{i}$ , the reflection matrix can be stated as $\\mathbf{R_i}$ . These matrices can be stated as:", + "bbox": [ + 496, + 458, + 890, + 589 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {T} _ {\\mathrm {i}} = \\left( \\begin{array}{c c} 1 & d _ {i} \\\\ 0 & 1 \\end{array} \\right), \\mathbf {R} _ {\\mathrm {i}} = \\left( \\begin{array}{c c} 1 & 0 \\\\ \\frac {2}{R _ {i}} & 0 \\end{array} \\right), \\mathbf {F} _ {\\mathrm {i}} = \\left( \\begin{array}{c c} 1 & 0 \\\\ \\frac {n _ {i + 1} - n _ {i}}{n _ {i + 1} R _ {i}} & \\frac {n _ {i}}{n _ {i + 1}} \\end{array} \\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 598, + 890, + 628 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In Fig. 4, $h_1$ represents the position of a real image that focuses on the sensor through the lens' internal reflection. We define the ray's forward propagation matrix as $\\mathbf{A}$ and $\\mathbf{C}$ and the reflection process as $\\mathbf{B}$ . They are defined as:", + "bbox": [ + 496, + 636, + 893, + 698 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} A = \\mathbf {T} _ {4} \\mathbf {F} _ {3} \\mathbf {T} _ {3} \\mathbf {F} _ {2} \\mathbf {T} _ {2} \\mathbf {F} _ {1} \\mathbf {T} _ {1} \\mathbf {F} _ {0} \\mathbf {T} _ {0} \\\\ B = \\mathbf {T} _ {4} \\mathbf {F} _ {3} \\mathbf {T} _ {3} \\mathbf {R} _ {2} ^ {- 1} \\mathbf {T} _ {3} \\mathbf {F} _ {3} ^ {- 1} \\mathbf {T} _ {4} \\mathbf {R} _ {4} \\tag {2} \\\\ C = \\mathrm {T} _ {8} \\mathrm {F} _ {8} \\mathrm {T} _ {7} \\mathrm {F} _ {7} \\mathrm {T} _ {6} \\mathrm {F} _ {5} \\mathrm {T} _ {5} \\mathrm {F} _ {4} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 575, + 707, + 890, + 762 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "With these matrices, the ray after reflection in the focal plane $\\mathbf{r}_1$ and direct imaging $\\mathbf{r}_0$ can be written as:", + "bbox": [ + 496, + 772, + 890, + 804 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\binom {h _ {1}} {\\theta_ {1}} = \\mathbf {C B A} \\binom {h} {\\theta}, \\binom {h _ {0}} {\\theta_ {0}} = \\mathbf {C A} \\binom {h} {\\theta} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 811, + 890, + 845 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For the focused image, $h_0$ and $h_1$ will not vary with the change of the angle of incidence $\\theta$ . Thus, we can deduce from Eq. (3) that both $h_0$ and $h_1$ are linear with $h$ . This", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "footer", + "text": "20785", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1776f2a732e8cc8a1852ed23a184dfb51ca8f571376ec0d63e53bb04ad5d9d4e.jpg", + "image_caption": [ + "Figure 4. Physical explanation of reflective flare formation. In this figure, $h$ and $\\theta$ indicate the height and the angle of incidence of the light source. $d$ and $n$ denote the thickness and refractive index of each layer. $A$ , $B$ , and $C$ represent the transfer matrices for each part of the lens system. The red line represents the imaging after reflection and the grey line shows the path of direct imaging. This figure depicts the situation with reflection in the 3rd and 5th air-lens layer, in real-world situations, the reflections can happen at any place including the lens protector and CMOS layer." + ], + "image_footnote": [], + "bbox": [ + 89, + 92, + 460, + 303 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "means that the line between the light source and the reflective flare will pass through the optical center and the ratio between the $h_0$ and $h_1$ is a constant. We define this ratio as $k$ with $h_1 = kh_0$ . Thus, we can calculate the position of the reflective flare, if this constant $k$ is known. Due to the lens design for most smartphone cameras, the value of $k$ is always -1. It leads to the optical center symmetry prior stated above.", + "bbox": [ + 75, + 448, + 472, + 570 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Experimental statistics", + "text_level": 1, + "bbox": [ + 76, + 579, + 290, + 595 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To demonstrate the effectiveness of our proposed prior in real-world situations, we conducted experiments using the main cameras of three smartphones: the Huawei P40, iPhone 13 Pro, and ZTE Axon 20 5G. In dark scenes, we captured 20 photos using each camera and marked the positions of the light source and reflective flare relative to the center of the image. As shown in Fig. 5, we observed that for all three cameras, the position of the reflective flare is linearly related to the position of the light source, with a slope of 1. However, since the light source may often create overexposed regions, manually marking the position of the light source can still introduce some errors. Considering this effect, we conclude that the results from all three phones align with our proposed prior.", + "bbox": [ + 75, + 603, + 470, + 816 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Data Collection", + "text_level": 1, + "bbox": [ + 76, + 829, + 232, + 844 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Although the optical center prior can guide us in finding the reflective flare, the lack of paired images still presents a challenge for training an effective model for flare re", + "bbox": [ + 75, + 854, + 470, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/daffbbce3c9429e95719b8fc8c4f94d983015c6b41392cc6eb5d0a22d4e86f3f.jpg", + "image_caption": [ + "(a) Statistics of flare-light source relation" + ], + "image_footnote": [], + "bbox": [ + 568, + 90, + 823, + 239 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f52a0286cd58d3af36954897bfcfa2a8a5b7e683db0b4a31953ce885676bfecb.jpg", + "image_caption": [ + "(b) flare-light source relation in x-axis" + ], + "image_footnote": [], + "bbox": [ + 537, + 252, + 691, + 367 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0645d13a2265558b07e6c0aba4684dab33f58b27297d2f51ce2ccf6b4f3e65e9.jpg", + "image_caption": [ + "(c) flare-light source relation in y-axis", + "Figure 5. Experimental statistics of the optical center symmetry prior. We capture images with reflective flares in a dark environment and mark the position of the light source and reflective flares. $x_{f}$ and $x_{l}$ represent the x-coordinate of flare and light source and so as $y_{f}$ and $y_{l}$ . The origin is set to be the center of the image. The relationship between the flare and the light source's absolute values of coordinates are plotted above. From figures (b) and (c), the coordinate relationships are linear and the slope is 1, which confirms our prior." + ], + "image_footnote": [], + "bbox": [ + 691, + 253, + 848, + 368 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "moral. As reflective flares are caused by the lens system, it is physically impossible to capture real flare-free and flare-corrupted image pairs. Moreover, given the huge variety of nighttime lights, it would be tedious to synthesize reflective flares for each type of light source.", + "bbox": [ + 496, + 531, + 893, + 607 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To address the challenge of generating paired images for reflective flare removal, we propose a semi-synthetic dataset, named BracketFlare. Since reflective flares always have the same pattern as the brightest region of the light source, we reduce the exposure time to capture the patterns of the light source's brightest region, which are used to synthesize the reflective flares. Specifically, we mount a professional camera (Sony A7 III) on a tripod and use continuous bracketing to capture five photos of the same static scene with a step of $3\\mathrm{EV}$ as a group. We then select the normally exposed image and the lowest exposure image from each group to compose the pair. To increase the diversity and realism of our dataset, we choose scenes with different light sources that are bright enough to leave reflective flares. Finally, we collect a dataset consisting of 440 pairs of 4K images as training data and 40 images as test data. A general overview of our dataset is depicted in Fig. 6, showing that the dataset contains diverse and commonly seen flare patterns.", + "bbox": [ + 496, + 613, + 892, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "20786", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d4ffd77f063a66e1c48bf33a3ae2f56ed36d52e8e487288f5ebb82b7589f83b5.jpg", + "image_caption": [ + "LED" + ], + "image_footnote": [], + "bbox": [ + 120, + 97, + 305, + 191 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/dd72b55dad2b6b7e0fa9ec8279fbbed67df4aae6fb867c3fea0b3df49442fa2d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 313, + 98, + 640, + 271 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a51b80b24715b9f02cc5f513da312df89b089eed721e1e1e26c48e71abed7eb1.jpg", + "image_caption": [ + "Traffic light" + ], + "image_footnote": [], + "bbox": [ + 120, + 204, + 305, + 299 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b2fb03a0035fe1f685e028675d529839498685b13b239678c0e99b4bd339bd1d.jpg", + "image_caption": [ + "Figure 6. Visualization and distribution of our dataset. It contains different types of light sources in diverse scenes. Based on light source types, we classify them into eight categories with different flare patterns." + ], + "image_footnote": [], + "bbox": [ + 313, + 271, + 643, + 308 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e635f1c97d4e46a6a4c50a6334c2d4513393973f87ef1a69d82250ad493dd8f5.jpg", + "image_caption": [ + "Square overhead light", + "Round overhead light" + ], + "image_footnote": [], + "bbox": [ + 658, + 97, + 843, + 191 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1ed183a0c7625b83b3e23b985d49c22d38ae74f8f818fb04ac6ec1bce4c1b6bb.jpg", + "image_caption": [ + "Light tube", + "Light bulb" + ], + "image_footnote": [], + "bbox": [ + 658, + 205, + 841, + 300 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Proposed Method", + "text_level": 1, + "bbox": [ + 76, + 364, + 250, + 383 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Postprocessing our dataset", + "text_level": 1, + "bbox": [ + 76, + 391, + 316, + 407 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To increase the diversity of our dataset, we generate paired flare-corrupted and flare-free images on the fly with augmentation. As depicted in Fig. 7(a), we first resize the 4K image to $800 \\times 1200$ . To simulate the possible misalignment of the image's center and optical center, we introduce a slight translation with $t \\sim U(0,15)$ pixels to the underexposed light source image. Then, we randomly crop the paired light source image and the normally exposed image with a size of $512 \\times 512$ . Since our captured images are already gamma-encoded, we follow the settings of Wu et al. [22] and linearize them by applying a gamma correction, i.e., $I_g = I^\\gamma$ with $\\gamma$ randomly sampled from [1.8, 2.2]. All our operations are performed in the linear space.", + "bbox": [ + 75, + 415, + 468, + 609 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Synthetic reflective flare. As lenses reflect light of different wavelengths with varying reflectivity, reflective flares often appear in different colors. To replicate this, we use a random color factor to multiply the low-exposed image. To simulate an unfocused situation, we also add random blur. In accordance with the optical center symmetry prior, we rotate the light source image to synthesize the reflective flare. Finally, we obtain the synthetic reflective flare by multiplying the flare image with an opacity.", + "bbox": [ + 75, + 612, + 468, + 747 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Flare-free image. The ISO is fixed for all images in our continuous bracketing. Therefore, our normally exposed images always share a long exposure time and have little noise. To simulate real night scenes, we add Gaussian noise to the images. Finally, we randomly apply a small color offset to the flare-free image. These operations aid in simulating light fog and glare effects in real-world scenarios.", + "bbox": [ + 75, + 750, + 468, + 853 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Flare-corrupted image. To create flare-corrupted images, we add the synthetic reflective flare to the flare-free images. Then, we apply gamma correction to the images with the", + "bbox": [ + 76, + 854, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "gamma value obtained by taking the reciprocal of the previously sampled gamma correction value $\\gamma$ . The resulting gamma-encoded flare-corrupted image, flare-free image, and flare image are used to train the neural networks.", + "bbox": [ + 498, + 367, + 890, + 428 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2. Flare removal pipeline", + "text_level": 1, + "bbox": [ + 500, + 435, + 709, + 452 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As the optical center symmetry prior is a global prior, it is difficult for CNN-based networks to learn it, and even Transformer-based networks struggle to learn such symmetry. Therefore, it is crucial to incorporate this prior into the network with specialized designs.", + "bbox": [ + 496, + 459, + 890, + 534 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To improve the network's ability to learn the optical center symmetry prior, as shown in Fig. 7(b), we apply a gamma correction of $I_{g} = I^{\\gamma}$ to the input flare-corrupted image with $\\gamma = 10$ . This operation extracts the nearly saturated areas of the light source in each channel. We then rotate this nearly saturated image by 180 degrees to obtain a prior image, which is concatenated with the flare-corrupted image to form a six-channel input. This rotated prior image provides an initial approximation of the reflective flare's pattern, which we expect the network to refine and enhance. The network's output is set to six channels, with three channels for the estimated flare-free image and three channels for the estimated flare. Our synthetic data provides supervision for all six channels, which facilitates network training.", + "bbox": [ + 496, + 534, + 892, + 744 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.3. Loss function", + "text_level": 1, + "bbox": [ + 500, + 753, + 638, + 768 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With the supervision of flare image $I_{F}$ and flare-free image $I_{B}$ , our final loss is the combination of background loss $\\mathcal{L}_{bg}$ , flare loss $\\mathcal{L}_{flare}$ , and reconstruction loss $\\mathcal{L}_{rec}$ . It can be written as:", + "bbox": [ + 498, + 777, + 890, + 835 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {r e c}} + \\mathcal {L} _ {\\text {f l a r e}} + \\mathcal {L} _ {\\text {b g}}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 845, + 890, + 863 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We define our flare removal network as $f_{\\theta}$ . It takes flare-corrupted image $I_{input}$ and its prior image $I_{prior}$ as inputs.", + "bbox": [ + 498, + 869, + 890, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "20787", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/944c57a82fcb893c5ae85d5767d22274eb557fc0506493a6ca6e315ab8e84746.jpg", + "image_caption": [ + "(a) Training Data Generation", + "(b) Network Learning", + "Figure 7. The pipeline of our data processing module and flare removal network. The flare image and a flare-free image from our paired dataset are resized and randomly cropped. After data augmentation, they are added to synthesize the corresponding flare-corrupted image. To encode the optical symmetry prior, the input flare-corrupted image is center flipped as the prior image. We also apply gamma correction to extract the light source region. Then, the prior image is concatenated with the flare-corrupted image as the input with 6 channels. Finally, the network predicts an output with 6 channels (a flare-free image with 3 channels and an estimated flare with 3 channels)." + ], + "image_footnote": [], + "bbox": [ + 138, + 95, + 818, + 282 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The flare-free image and flare can be estimated by:", + "bbox": [ + 76, + 392, + 413, + 407 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {I} _ {B}, \\hat {I} _ {F} = f _ {\\theta} \\left(I _ {\\text {i n p u t}}, I _ {\\text {p r i o r}}\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 178, + 417, + 468, + 435 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "With the estimated flare and flare-free images, the $L_{1}$ reconstruction loss can be expressed as:", + "bbox": [ + 76, + 445, + 467, + 474 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r e c}} = \\left| I _ {\\text {i n p u t}} - \\hat {I} _ {F} \\oplus \\hat {I} _ {B} \\right|, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 484, + 468, + 503 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\oplus$ denotes an element-wise addition operation in the linearized gamma-decoded domain.", + "bbox": [ + 75, + 513, + 468, + 542 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For the background loss, it consists of a default $\\mathcal{L}_1$ , a region-aware masked $\\mathcal{L}_{mask}$ term, and a perceptual term $\\mathcal{L}_{vgg}$ with a pre-trained VGG-19 network [17]. Compared with previous works [5, 22, 26], the main difference is the $\\mathcal{L}_{mask}$ . Since the regions of reflective flares are always small, a simple $L_1$ loss may ignore these regions, leading to the local optimum. To encourage the network to focus on restoring the flare-corrupted regions, we calculate the flare's mask $I_M$ ( $>0.2$ ) from the ground truth of flare image $I_F$ . This masked loss can be written as:", + "bbox": [ + 76, + 544, + 468, + 693 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {m a s k}} = \\left| I _ {M} * \\left(\\hat {I} _ {B} - I _ {B}\\right) \\right|. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 705, + 468, + 722 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Overall, the background loss $\\mathcal{L}_{bg}$ is defined as:", + "bbox": [ + 96, + 732, + 406, + 750 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {b g} = w _ {1} \\mathcal {L} _ {1} + w _ {2} \\mathcal {L} _ {v g g} + w _ {3} \\mathcal {L} _ {\\text {m a s k}}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 760, + 468, + 776 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $w_{1}, w_{2}, w_{3}$ are respectively set to 0.5, 0.1, and 20.0 in our experiments. Flare loss $\\mathcal{L}_{flare}$ has the identical form as the background loss.", + "bbox": [ + 75, + 786, + 468, + 830 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6. Experiments", + "text_level": 1, + "bbox": [ + 76, + 845, + 209, + 862 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation on synthetic data. To evaluate our method, we collect 40 paired data using the process mentioned in", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/3a38e003e69dbf0a5a0cde48d88d7d95805c2f14ef6ae736e03d80b17c38f883.jpg", + "table_caption": [ + "Table 1. Percentage of users favoring our results (our dataset + MPRNet†) vs. previous work. Our method outperforms existing methods on both indoor and outdoor images." + ], + "table_footnote": [], + "table_body": "
MethodOutdoorIndoor
Reflection removal [26]3.0%8.5%
Flare spot removal [1]4.5%0.0%
Flare removal (Wu et al.) [22]0.5%3.5%
Flare removal (Dai et al.) [5]1.5%2.5%
Ours90.5%85.5%
", + "bbox": [ + 540, + 436, + 846, + 531 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Sec. 4 to generate a synthetic test dataset, in which ground truth images are available. We adopt PSNR, SSIM [20], and LPIPS [25] as our evaluation metrics.", + "bbox": [ + 496, + 535, + 890, + 580 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation on real data. To show our method's generalization ability, we also collect a dataset with the photos captured by different types of mobile phones, including iPhone 12 pro, iPhone 13 pro, iPhone 14 pro, Huawei mate 30, Huawei nova 5i Pro, Huawei P40, Samsung Galaxy s22, Samsung Galaxy Note9, ZTE Axon 20 5G and Oppo Reno4 SE 5G. For real cases, we cannot obtain the ground truth. Since reflective flares always locate in small regions, they are hard to be evaluated by existing non-reference image quality assessment. These metrics emphasize more on global image quality and are not specially designed for our task. Therefore, we provide more visual results and conduct a user study to evaluate the performance.", + "bbox": [ + 496, + 580, + 890, + 776 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.1. Comparison with different networks", + "text_level": 1, + "bbox": [ + 500, + 786, + 815, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To validate the performance of our pipeline, we train different CNN-based and Transformer-based networks with our data. We use the default network settings of MPR-Net [24], HINet [4], and Uformer [21] and reduce the feature channels of Restormer [23] from 48 to 24 due to GPU memory constraint. In our experiments, these models are", + "bbox": [ + 496, + 809, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "20788", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/db4ff8be7c796b93fd3ee33705ab6a454fcc7a2fcc8dacfa2d2f737fae5c9c89.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 148, + 90, + 256, + 175 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bc22b91a1000dfa5e315e05097c4ed22f7074250f4826bc8596267267b625c27.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 148, + 176, + 254, + 258 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0e28a7387ad25a3188e0f00ef64b7976883921e8a2de6ab19a7ec91ee287bff3.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 148, + 260, + 254, + 343 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/48ffa684854b80cdb98ae74a698c88ff432674d9fb5ebdc08e8337f3afd8ad7a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 259, + 90, + 366, + 175 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fb67e9933b70653afd1d83fcf53e40caca193a156f97dda8482eee51466f8589.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 259, + 176, + 366, + 258 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3bc747c7f2ed860729a02cec77fb2ea8f777368fc3dfe005e630ded9732442d3.jpg", + "image_caption": [ + "Reflection removal" + ], + "image_footnote": [], + "bbox": [ + 259, + 260, + 366, + 343 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9726bc3315bca57d0b557af8ae7ef3e278ce9c2ec6488003a825d2315be3e16d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 370, + 90, + 478, + 175 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b8e74dc2197d334dd06372da48bca80eca1ad9daf9c91c24aa62b0860e72bcf5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 370, + 176, + 478, + 258 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/23a4606430801aba1811b766defc720f4d0cfe5aacb8bf09d3953913150a56bf.jpg", + "image_caption": [ + "Flare spot removal" + ], + "image_footnote": [], + "bbox": [ + 370, + 260, + 478, + 343 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8dceb54be224d6e022f18af3acaf48c64cb605ab720e9d6a6b10ac193112287a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 480, + 90, + 588, + 175 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f8d85319bd51a4fc6bec357cfaa83957e1e1dabd92b0d49ea1e8f58d561908fe.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 480, + 176, + 588, + 258 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/aa92a277056b15d31558666ee5a13aebed60f18d25571335288f8d68157f78ad.jpg", + "image_caption": [ + "Wuet al." + ], + "image_footnote": [], + "bbox": [ + 480, + 260, + 588, + 343 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3e57a9445697106823a328fa71b26f3fcbec8015ddc63e993a976cbdc6cc907a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 591, + 90, + 700, + 175 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bef219f9c204b4842482eac749526d905c5b43bd1e2910b1b2d9cca28882669a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 591, + 176, + 700, + 258 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2f77e8e26b73e42d00f65cf69672053f645be9217313011745a8d08188f41a0a.jpg", + "image_caption": [ + "Dai et al." + ], + "image_footnote": [], + "bbox": [ + 591, + 260, + 700, + 343 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6d52eaf9c87e6835d71acbc3c56c9071d27f2041706dc360668e891a7191e1e2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 704, + 90, + 810, + 175 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c40034e529549cb7a8662358b9d3027e031b9aa817278645b37dd66042d81011.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 704, + 176, + 810, + 258 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0d4dc6c70a282c5d96d04a6178755824f8984eb779e1443017610483f6412173.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 704, + 260, + 810, + 343 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5c103a743c10fdfd45e41cf722277a2849943a8aa1234d06eb2f5957102d60f2.jpg", + "image_caption": [ + "Figure 8. Visual comparison of different flare removal and reflection removal methods on real-world nighttime flare images." + ], + "image_footnote": [], + "bbox": [ + 88, + 382, + 199, + 469 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/eb461777130ef69b9236661f07717fe9f5641feda42819453bcd47a5bc06cc80.jpg", + "image_caption": [ + "Input", + "Figure 9. Visual comparison of different flare removal and reflection removal methods on synthetic nighttime flare images." + ], + "image_footnote": [], + "bbox": [ + 88, + 469, + 199, + 555 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/001f546420f7ac9066da37a053b129c6261a0974128a3560d7eb563f79b64caf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 200, + 382, + 310, + 469 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e80bf37784cbe13dd014e5671dfd7d6b11628fa2aa90fa420518e7095da2e6f4.jpg", + "image_caption": [ + "Reflection removal" + ], + "image_footnote": [], + "bbox": [ + 200, + 469, + 310, + 554 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/002c0ca510c5b7f08626df2b9f24f972d15c42fa0c8e41ef910ca8cb17e55993.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 313, + 382, + 424, + 468 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2333412be9c73797574131e92a19711fb6975819bee7f7b5f2cbb6afa0633335.jpg", + "image_caption": [ + "Flare spot removal" + ], + "image_footnote": [], + "bbox": [ + 313, + 469, + 424, + 555 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bcf4a3be8cda5af200cb84d6ce37360377f7611326afb3a65e7ad0d52eb3a370.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 426, + 382, + 537, + 469 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/42051f6c40f73ce84a6d2e414a9374a777a628d30fb0f39fa21aafb7a7e63770.jpg", + "image_caption": [ + "Wu et al." + ], + "image_footnote": [], + "bbox": [ + 426, + 469, + 537, + 555 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/64448d4dcb9a8363b02eb8bbb01bb05ff2c634ca7a0e18fe5ff8206689265bbc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 539, + 382, + 651, + 469 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dc66efa5ba0e54a7b9d8ed6be7d8a48b367a5e580dc258a5407b3098e500e28f.jpg", + "image_caption": [ + "Dai et al." + ], + "image_footnote": [], + "bbox": [ + 539, + 469, + 651, + 555 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/02197a3b7e9d996222ca1bf773cd4d8819f2d1725d965881acffadee4ebb446f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 651, + 382, + 763, + 469 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/23e47ddc750bc19b22df9a39b5fde190f01e91f312ca9c9ac66d8ea45a26ee81.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 651, + 469, + 763, + 555 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b3768e6c35fca6693740ce3733b1634ae6277b7d476ed831dea1a54412fd90ec.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 764, + 382, + 875, + 469 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/84f091058fd93a70875f2f2fbd32e65868afd22ff8364c01c59e1e62883795a2.jpg", + "image_caption": [ + "Ground truth" + ], + "image_footnote": [], + "bbox": [ + 764, + 469, + 875, + 555 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "trained on an NVIDIA Geforce RTX3090 with a batch size of 2. We use Adam optimizer [9] with a learning rate of 1e-4 and train a model for 500k iterations. Besides, we adopt a MultiStepLR with a milestone of 200k and set the gamma to 0.5. These models are evaluated using the synthetic test dataset as shown in Table 4. From experiments, we find that $\\mathrm{MRPNet}^{\\dagger}$ [24] performs better than other networks. Thus, we treat $\\mathrm{MPRNet}^{\\dagger}$ [24] as our baseline method and conduct the following ablation study.", + "bbox": [ + 75, + 601, + 468, + 737 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.2. Comparison with previous works", + "text_level": 1, + "bbox": [ + 76, + 744, + 367, + 760 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compare our method with different flare removal datasets [5, 22] and reflection removal dataset [26] on both real and synthetic data. We use the pre-trained models of Dai et al. [5] and Zhang et al. [26]. Since Wu et al. [22] do not release their model, we retrain their network using their provided data and code. Besides, we also compare our method with the traditional flare spot removal method [1]. As shown in Table 2, since most of these methods are not specially designed for reflective flare removal, they cannot", + "bbox": [ + 75, + 763, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3343bc6466eb108e02a209aa590239642ca3fcdb556255a42f683840f904247d.jpg", + "table_caption": [ + "Table 2. Quantitative comparison on synthetic test data." + ], + "table_footnote": [], + "table_body": "
MethodPSNRSSIMLPIPSMasked PSNR
Input37.300.9900.02521.68
Reflection removal [26]22.020.8300.07422.88
Flare spot removal [1]38.290.9900.02422.53
Flare removal (Wu et al.) [22]26.130.8950.05521.89
Flare removal (Dai et al.) [5]27.490.9420.05021.85
Ours48.410.9940.00432.09
", + "bbox": [ + 501, + 617, + 888, + 714 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "accurately remove the reflective flare while retaining image details simultaneously. Although Asha et al. [1] can detect the bright spots, as shown in Fig. 8, it may also remove the light source which is quite common at night.", + "bbox": [ + 496, + 717, + 890, + 777 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "User Study. To evaluate these methods in real-world situations, we also conduct a user study with 20 participants with 20 images randomly selected from our real data in Table 1.", + "bbox": [ + 496, + 777, + 890, + 823 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.3. Ablation study", + "text_level": 1, + "bbox": [ + 500, + 832, + 648, + 848 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Loss function. We conduct an ablation study to prove the necessity of each loss function used in our baseline. In our experiments, we train the model without adding $\\mathcal{L}_{\\text {flare }}$ ,", + "bbox": [ + 496, + 854, + 890, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "20789", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6da115c232ba0dd5a5583753908fc58d30228cd49e6f4ecc78904616742b6bf7.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 104, + 95, + 230, + 191 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c313f5f2cbae4b35bb8264909bc35cf7f3cd15657bfd1e7981d3c0519e323514.jpg", + "image_caption": [ + "Output", + "(a) iPhone 14 Pro", + "Figure 10. Visual comparison of flare removal on real-world nighttime flare images that are captured by different smartphones." + ], + "image_footnote": [], + "bbox": [ + 104, + 194, + 230, + 291 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/567933723a794726eedc088fc74a3ce60c96ae3b31f346b59cb140154b6f75c4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 235, + 95, + 359, + 191 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/83949e84d2e25fd226a074c25b0bc9c52674adcaec4c411e702138431b4674bd.jpg", + "image_caption": [ + "(b) iPhone 13 Pro" + ], + "image_footnote": [], + "bbox": [ + 235, + 194, + 359, + 291 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/73d94687b622709963faaf5204d8f06d1bd7de4bbad8617103a4dee80f0fb428.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 364, + 94, + 490, + 191 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/91a35f8607fde00c87901c6623428def7285e5b16b78925bd304ace3c1089ff2.jpg", + "image_caption": [ + "(c) ZTE Axon 20 5G" + ], + "image_footnote": [], + "bbox": [ + 364, + 194, + 488, + 291 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/30b04c581f92f4b186b9d63e840c9919965a29ddfe621646f93c8f427cf16ff1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 493, + 95, + 617, + 191 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a1f0614f23ad77ef1efde7eed260fca04cedade4d55b1bf2ca302a4fb2ddb2b1.jpg", + "image_caption": [ + "(d) Samsung Galaxy S22" + ], + "image_footnote": [], + "bbox": [ + 493, + 194, + 617, + 291 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/908f312e9fa316ab94f27ea10306713f837cb992bb904e1236a89061354a733b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 622, + 95, + 746, + 191 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/11aefdabac495aa6b62bf1e9bd24dfb6144975bb9473fa032e5b30c120eecb74.jpg", + "image_caption": [ + "(e) Huawei Mate 30" + ], + "image_footnote": [], + "bbox": [ + 622, + 194, + 746, + 291 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9e81efc2c6ce08bba636e382b96b40a229eadd92f350ec00531a9b7d88f796c6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 751, + 95, + 875, + 191 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ada74c8871fef56c6b26ff81dd53c6241b3b53ba0e6c0a707477d1fb94a14211.jpg", + "image_caption": [ + "(f) Huawei nova 5i Pro" + ], + "image_footnote": [], + "bbox": [ + 751, + 194, + 875, + 290 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/865ebf85c6f8a9633fafbf17e8c8dee092da66bfae9c21569a47d123fe42ccc3.jpg", + "table_caption": [ + "Table 3. Ablation study for different loss functions." + ], + "table_footnote": [], + "table_body": "
LossPSNRSSIMLPIPSMasked PSNR
w/o Lflare47.360.9930.00430.63
w/o Lrec48.180.9940.00331.33
w/o Lvgg48.050.9940.00431.39
w/o Lmask48.010.9930.00331.03
Full48.410.9940.00432.09
", + "bbox": [ + 107, + 359, + 433, + 453 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d1a31061421c916d5a577f52ef6c73dc5380d1e55ff2fbef1291ffa43b92d2c8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 102, + 460, + 210, + 542 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9f6e83b02c0987ef038de89e275634cb4e7f54fe40cef10f3db875a56c8ef233.jpg", + "image_caption": [ + "Input", + "Figure 11. Visual comparison with the network's performance with and w/o optical center symmetry prior." + ], + "image_footnote": [], + "bbox": [ + 102, + 545, + 210, + 627 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5d4cb1c4d224889639b106a9f480c83f4b034bbfd5dd633babfd46e377104522.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 212, + 460, + 320, + 542 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/71e6c233b8f3e23d10eb766fcc3e24115a711e96484507949edc8265fb27342f.jpg", + "image_caption": [ + "w/o prior" + ], + "image_footnote": [], + "bbox": [ + 212, + 545, + 320, + 627 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4b6147282e274c24cef08102bad81052a62352842084aec46989f5ce7263a4a0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 321, + 460, + 429, + 542 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/da016b0a5fb26f9b2d97e0405b93378898ffdf710a4fcb4ef0f6f48a7b705734.jpg", + "image_caption": [ + "with prior" + ], + "image_footnote": [], + "bbox": [ + 321, + 545, + 429, + 627 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "$\\mathcal{L}_{rec}$ , $\\mathcal{L}_{vgg}$ or $\\mathcal{L}_{mask}$ . We report the results in Table 3. The ablation study of $\\mathcal{L}_{flare}$ proves the significance of bringing supervision for both flare and background images. Besides, when the $L_{mask}$ is removed, it degrades the reconstruction results in the flare-corrupted regions, which verifies the importance of focusing on flare regions.", + "bbox": [ + 76, + 684, + 468, + 775 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Prior. To illustrate the performance of introducing the optical symmetry prior by concatenating a prior image, we train these baseline restoration networks with the original input. Since HINet [4] and MPRNet [24] cannot support input and output with different channels, we concatenate a black image with no information and keep other settings the same. The quantitative results are shown in Table 4, and visual comparisons on real-world images are presented in Fig.", + "bbox": [ + 76, + 779, + 468, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/af210b137e39c1272c22a7a64ce0c52626f94bf1650260228437b7cddb2dcdf9.jpg", + "table_caption": [ + "Table 4. Quantitative comparison for different network structures on synthetic test data. ‘†’ indicates that the network is equipped with the proposed prior." + ], + "table_footnote": [], + "table_body": "
NetworkPSNRSSIMLPIPSMasked PSNR
MPRNet [24]43.600.9920.01127.64
Uformer [21]42.570.9890.00927.77
HINet [4]43.780.9910.01227.73
Restormer [23]42.760.9910.01227.27
MPRNet†48.410.9940.00432.09
Uformer†47.470.9910.00331.57
HINet†48.030.9940.00330.88
Restormer†48.110.9940.00431.07
", + "bbox": [ + 513, + 391, + 870, + 534 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "11. It shows that embedding this prior in the network can boost the performance of quantitative results considerably. In Fig. 11, we observe that this prior helps the network to locate and remove the reflective flare precisely.", + "bbox": [ + 500, + 544, + 890, + 604 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 607, + 617, + 621 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have presented a novel approach to tackle the challenge of removing reflective flares in nighttime photography. Our approach is inspired by the optical center symmetry prior, which suggests that reflective flares and their corresponding light sources are always symmetrical around the optical center of the lens system in smartphone cameras. Based on the observation that the patterns of reflective flares are always similar to the brightest region of the light source, we introduce a dataset named BracketFlare. We further propose a reflective flare removal pipeline that encodes the optical center symmetry prior into the network. Experimental results demonstrate that our proposed methods are effective in removing reflective flares in real-world scenarios and have the potential for wide-ranging applications.", + "bbox": [ + 496, + 625, + 890, + 837 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement: This study is supported under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s).", + "bbox": [ + 498, + 839, + 890, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "20790", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] C S Asha, Sooraj Kumar Bhat, Deepa Nayak, and Chaithra Bhat. Auto removal of bright spot from images captured against flashing light source. In ICDC, 2019. 2, 3, 6, 7", + "[2] Herbert Bay, Andreas Ess, Tinne Tuytelaars, and Luc Van Gool. Speeded-up robust features (surf). Computer Vision and Image Understanding, 110(3):346-359, 2008. 2", + "[3] Floris Chabert. Automated lens flare removal, 2014. Technical Report, Department of Electrical Engineering, Stanford University. 2, 3", + "[4] Liangyu Chen, Xin Lu, Jie Zhang, Xiaojie Chu, and Chengpeng Chen. Hinet: Half instance normalization network for image restoration. In CVPRW, 2021. 2, 6, 8", + "[5] Yuekun Dai, Chongyi Li, Shangchen Zhou, Ruicheng Feng, and Chen Change Loy. Flare7k: A phenomenological nighttime flare removal dataset. In Thirty-sixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2022. 1, 2, 3, 6, 7", + "[6] Ruicheng Feng, Chongyi Li, Huaijin Chen, Shuai Li, Chen Change Loy, and Jinwei Gu. Removing diffraction image artifacts in under-display camera via dynamic skip connection network. In CVPR, 2021. 3", + "[7] Matthias Hullin, Elmar Eisemann, Hans-Peter Seidel, and Sungkil Lee. Physically-based real-time lens flare rendering. In ACM SIGGRAPH, 2011. 1", + "[8] Yeying Jin, Wenhan Yang, and Robby T Tan. Unsupervised night image enhancement: When layer decomposition meets light-effects suppression. In ECCV. Springer, 2022. 3", + "[9] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 7", + "[10] Sungkil Lee and Elmar Eisemann. Practical real-time lensflare rendering. Computer Graphics Forum, 32(4):1-6, 2013. 1, 3", + "[11] Yu Li, Robby T. Tan, and Michael S. Brown. Nighttime haze removal with glow and multiple light colors. In ICCV, 2015. 3", + "[12] David G Lowe. Distinctive image features from scale-invariant keypoints. International Journal of Computer Vision, 60(2):91-110, 2004. 2", + "[13] Andreas Nussberger, Helmut Grabner, and Luc Van Gool. Robust aerial object tracking in images with lens flare. In ICRA, 2015. 3", + "[14] Frank L Pedrotti, Leno M Pedrotti, and Leno S Pedrotti. Introduction to Optics. Cambridge University Press, 2017. 3", + "[15] Xiaotian Qiao, Gerhard P. Hancke, and Rynson W. H. Lau. Light source guided single-image flare removal from unpaired data. In ICCV, 2021. 2, 3", + "[16] Aashish Sharma and Robby T. Tan. Nighttime visibility enhancement by increasing the dynamic range and suppression of light effects. In CVPR, 2021. 3", + "[17] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In International Conference on Learning Representations, 2014. 6", + "[18] Qilin Sun, Ethan Tseng, Qiang Fu, Wolfgang Heidrich, and Felix Heide. Learning rank-1 diffractive optics for single-shot high dynamic range imaging. In CVPR, 2020. 2" + ], + "bbox": [ + 78, + 114, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[19] Patricia Vitoria and Coloma Ballester. Automatic flare spot artifact detection and removal in photographs. Journal of Mathematical Imaging and Vision, 61:515-533, 2019. 2, 3", + "[20] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6", + "[21] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restorationn. In CVPR, 2022. 2, 6, 8", + "[22] Yicheng Wu, Qiurui He, Tianfan Xue, Rahul Garg, Jiawen Chen, Ashok Veeraraghavan, and Jonathan T Barron. How to train neural networks for flare removal. In ICCV, 2021. 1, 2, 3, 5, 6, 7", + "[23] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In CVPR, 2022. 2, 6, 8", + "[24] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Multi-stage progressive image restoration. In CVPR, 2021. 2, 6, 7, 8", + "[25] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6", + "[26] Xuaner Zhang, Ren Ng, and Qifeng Chen. Single image reflection separation with perceptual losses. In CVPR, 2018. 3, 6, 7", + "[27] Junyan Zhu, Taesung Park, Phillip Isola, and Alexei A. Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In ICCV, 2017. 3" + ], + "bbox": [ + 501, + 92, + 892, + 544 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "20791", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + } +] \ No newline at end of file diff --git a/2023/Nighttime Smartphone Reflective Flare Removal Using Optical Center Symmetry Prior/d34edf55-17a4-4aa7-b0c0-de157905c27a_model.json b/2023/Nighttime Smartphone Reflective Flare Removal Using Optical Center Symmetry Prior/d34edf55-17a4-4aa7-b0c0-de157905c27a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..258144ebc92e545bc54ba1346b8d54276ef15007 --- /dev/null +++ b/2023/Nighttime Smartphone Reflective Flare Removal Using Optical Center Symmetry Prior/d34edf55-17a4-4aa7-b0c0-de157905c27a_model.json @@ -0,0 +1,2946 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.24, + 0.131, + 0.734, + 0.177 + ], + "angle": 0, + "content": "Nighttime Smartphone Reflective Flare Removal Using Optical Center Symmetry Prior" + }, + { + "type": "text", + "bbox": [ + 0.17, + 0.204, + 0.8, + 0.24 + ], + "angle": 0, + "content": "Yuekun Dai Yihang Luo Shangchen Zhou Chongyi Li* Chen Change Loy S-Lab, Nanyang Technological University" + }, + { + "type": "text", + "bbox": [ + 0.233, + 0.242, + 0.737, + 0.258 + ], + "angle": 0, + "content": "{YDAI005, c200211, s200094, chongyi.li, ccloy}@ntu.edu.sg" + }, + { + "type": "text", + "bbox": [ + 0.282, + 0.26, + 0.682, + 0.275 + ], + "angle": 0, + "content": "https://ykdai.github.io/projects/BracketFlare" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.288, + 0.314, + 0.304 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.321, + 0.474, + 0.683 + ], + "angle": 0, + "content": "Reflective flare is a phenomenon that occurs when light reflects inside lenses, causing bright spots or a \"ghosting effect\" in photos, which can impact their quality. Eliminating reflective flare is highly desirable but challenging. Many existing methods rely on manually designed features to detect these bright spots, but they often fail to identify reflective flares created by various types of light and may even mistakenly remove the light sources in scenarios with multiple light sources. To address these challenges, we propose an optical center symmetry prior, which suggests that the reflective flare and light source are always symmetrical around the lens's optical center. This prior helps to locate the reflective flare's proposal region more accurately and can be applied to most smartphone cameras. Building on this prior, we create the first reflective flare removal dataset called BracketFlare, which contains diverse and realistic reflective flare patterns. We use continuous bracketing to capture the reflective flare pattern in the underexposed image and combine it with a normally exposed image to synthesize a pair of flare-corrupted and flare-free images. With the dataset, neural networks can be trained to remove the reflective flares effectively. Extensive experiments demonstrate the effectiveness of our method on both synthetic and real-world datasets." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.699, + 0.21, + 0.715 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.725, + 0.471, + 0.877 + ], + "angle": 0, + "content": "Lens flare artifacts can occur when a strong light enters the camera's field of view, causing scattering and reflection within the lenses. Unlike radial-shaped scattering flares, reflective flares appear as groups of bright spots [5, 7] or a \"ghosting effect\" like the examples in Fig. 1. With multiple reflections, the light source creates multiple projections at various positions behind the lens. If the projection is far from the focal plane, the defocused effect can brighten the whole sensor and degrade the photo's quality as shown in Fig. 2. In near-focus situations, images of the light source" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.291, + 0.644, + 0.383 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.384, + 0.643, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.477, + 0.643, + 0.568 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.541, + 0.569, + 0.625, + 0.58 + ], + "angle": 0, + "content": "(a) iPhone 13 Pro" + }, + { + "type": "image", + "bbox": [ + 0.645, + 0.291, + 0.764, + 0.383 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.645, + 0.384, + 0.764, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.645, + 0.477, + 0.764, + 0.568 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.664, + 0.569, + 0.74, + 0.58 + ], + "angle": 0, + "content": "(b) Huawei P40" + }, + { + "type": "image", + "bbox": [ + 0.766, + 0.291, + 0.885, + 0.383 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.766, + 0.384, + 0.885, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.766, + 0.477, + 0.885, + 0.568 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.776, + 0.568, + 0.874, + 0.58 + ], + "angle": 0, + "content": "(c) Sony ILCE-6400" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.588, + 0.892, + 0.658 + ], + "angle": 0, + "content": "Figure 1. Reflective flare removal with our proposed dataset and prior. The network trained on our dataset can effectively separate the reflective flare and flare-free images from a flare-corrupted image. Our method can generalize well to different types of smartphones and even cameras with specific types of lenses." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.67, + 0.891, + 0.701 + ], + "angle": 0, + "content": "can result in a line of unfocused, bright blobs on the captured photo." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Applying anti-reflective (AR) coating is a common technique for reducing internal reflections [22]. The coating is effective for unfocused spots with low luminance, but cannot mitigate a bright spot with high luminance at the focal plane. Even with a well-designed lens system, a bright reflective spot may remain in the photo, known as the ghosting effect [10]. Reflective flare is common on all types of smartphone lenses, and certain types of smartphones (e.g., the iPhone series) can produce a large area of inverted reflective flare from neon lights or large, luminous standing signs, as shown in Fig. 1. The phenomenon becomes more pronounced when capturing a photo at night with multiple lights in the scene. Numerous bright spots and inverted re" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.222, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20783" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.137 + ], + "angle": 0, + "content": "flections manifest and propagate in different directions with the lens orientation. Therefore, reflective flare removal algorithms are highly desirable." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.14, + 0.473, + 0.548 + ], + "angle": 0, + "content": "Removing reflective flare is a difficult task due to the vast diversity of reflective flares that can be caused by different types of light sources and lenses. Reflective flares often appear as bright spots that can resemble streetlights from a distance, making it challenging to differentiate between reflective flares and other light sources. Current methods for removing flare spots mainly rely on feature detection techniques like Speeded-Up Robust Features [2] (SURF) or Scale-Invariant Feature Transform [12] (SIFT) to extract spot regions [1, 3, 19]. These methods, however, struggle to effectively detect blobs on complex backgrounds and only work for specific spot patterns. Moreover, these methods are ineffective in removing reflective flares caused by lamp tubes and LED matrices. Many learning-based methods and datasets for reflective flare removal have been proposed [5, 15, 18, 22]. Although these datasets contain subsets of reflective flares, they are not sufficient for good generalization. For instance, the reflective flare subset from Wu et al. [22] is captured using the same lens and light source, making it difficult to generalize to different flares in real-world scenarios. Dai et al. [5] propose a synthetic dataset with 10 types of reflective flares, but training on this dataset tends to mislead a network to erroneously remove other tiny bright spots like street lights and bright windows. As a result, it remains challenging for existing algorithms to accurately detect and remove various types of reflective flares while preserving details in other regions." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.551, + 0.473, + 0.868 + ], + "angle": 0, + "content": "To overcome the challenges of detecting and removing reflective flares, we propose a novel optical center symmetry prior, which suggests that the main spot of the reflective flare and the light source are always symmetrical around the lens's optical center in captured images. This prior applies to most smartphone cameras and some professional cameras. With this prior, we can easily detect the possible locations of reflective flares. We also found that the flare spots always have the same patterns as the brightest regions of the light source. Therefore, we can synthesize reflective flares by using the light source's patterns at low exposure. Based on these characteristics of reflective flares, we develop the first reflective flare removal dataset, named BracketFlare. Specifically, we use continuous bracketing to capture the light source's pattern on an underexposed image and select the normally exposed image to create the light source and the natural image. Next, we rotate the light source pattern 180 degrees around the optical center, apply blur and color augmentation, and use it to create synthetic reflective flares. We can add the reflective flare to the normally exposed image to produce paired flare-corrupted and flare-free images." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.472, + 0.902 + ], + "angle": 0, + "content": "Thanks to the optical center symmetry prior, we further propose an end-to-end pipeline for reflective flare re" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.092, + 0.7, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.556, + 0.226, + 0.668, + 0.24 + ], + "angle": 0, + "content": "(a) w/o AR coating" + }, + { + "type": "image", + "bbox": [ + 0.704, + 0.092, + 0.875, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.73, + 0.226, + 0.85, + 0.239 + ], + "angle": 0, + "content": "(b) With AR coating" + }, + { + "type": "image", + "bbox": [ + 0.537, + 0.239, + 0.882, + 0.373 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.578, + 0.374, + 0.833, + 0.387 + ], + "angle": 0, + "content": "(c) Description of reflective flare formation" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.394, + 0.892, + 0.505 + ], + "angle": 0, + "content": "Figure 2. Schematic diagram of reflective flare formation. Reflective flares are generated by multiple reflections inside the lens. In (c), yellow lines show the generation principle of the aperture-shaped reflective flare. Red lines represent the formation of the bright spot. Since the unfocused reflective flare (yellow line, e.g., (a)) can be effectively alleviated by AR coating. In this work, we mainly focus on removing the focused reflective flare (red line, e.g., (b)) from night images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.512, + 0.892, + 0.631 + ], + "angle": 0, + "content": "moral, which can easily apply standard neural networks designed for image restoration. Specifically, to demonstrate the effectiveness of our dataset and pipeline, we train two CNN-based networks [4, 24] and two Transformer-based networks [21, 23] as baselines. These models trained on our dataset effectively suppress the reflective flare and generalize well to real-world scenarios. The contributions of our work are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.641, + 0.892, + 0.702 + ], + "angle": 0, + "content": "- We propose a new optical center symmetry prior that uncovers crucial properties of reflective flares on smartphones' cameras, including their locations of occurrence and patterns." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.711, + 0.892, + 0.786 + ], + "angle": 0, + "content": "- Based on this prior, we explore a new method for constructing a reflective flare removal dataset called BracketFlare. This dataset includes various types of light sources and diverse indoor and outdoor scenes, providing a solid foundation for this task." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "- We design an end-to-end reflective flare removal pipeline that incorporates the optical center symmetry prior to both training data generation and network structures. Extensive experiments demonstrate the essential role of our dataset and prior in reflective flare removal, which effectively generalizes to different night scenes" + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.641, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "20784" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.09, + 0.22, + 0.106 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.116, + 0.473, + 0.359 + ], + "angle": 0, + "content": "Scattering flare removal. Most flare removal methods mainly focus on removing the scattering flares and glare effects. Based on the glow smooth prior, unsupervised methods [8, 16] are proposed to separate the glare effect from the flare-corrupted images. For supervised flare removal, Wu et al. [22] and Dai et al. [5] have respectively proposed two flare image datasets. These flare images can be added to flare-free images to synthesize the flare-corrupted and flare-free image pairs, making training an end-to-end flare removal network possible. Due to the difficulty in collecting real paired data, Qiao et al. [15] designed a pipeline to train the flare removal network with unpaired data following the idea of Cycle-GAN [27]. Besides, artifacts from underdisplay cameras [6] and nighttime haze [11] share similar attributes with the scattering flares. Thus, the related methods can also alleviate scattering flares artifacts." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.366, + 0.473, + 0.61 + ], + "angle": 0, + "content": "Reflective flare removal. Traditional reflective flare removal is typically accomplished through feature detection and image inpainting. Flare detection in these approaches [1,3,13,19] assumes that reflective flares are bright spots with specific shapes. Such an assumption is fragile given nighttime situations with diverse light sources. Since reflective flare removal can be viewed as a specific kind of reflection removal [26], one can adopt learning-based methods. However, existing datasets for this task, such as those proposed by Wu et al. [22], Dai et al. [5], and Qiao et al. [15], are not specifically designed for reflective flare removal and lack diversity in terms of light sources. These limitations hinder deep learning-based methods from effectively addressing real-world reflective flares, and may even lead to the removal of other tiny bright spots such as street lights." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.624, + 0.37, + 0.643 + ], + "angle": 0, + "content": "3. Optical Center Symmetry Prior" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.65, + 0.291, + 0.667 + ], + "angle": 0, + "content": "3.1. Definition for the prior" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.473, + 0.903 + ], + "angle": 0, + "content": "When pointing the camera at a light source, strong light can sometimes create lens flares in photos. In well-designed lens systems with AR coatings, reflective flares appear as bright spots on the photo, rather than generating a light fog. Through our investigation of most smartphones' main cameras, we discover that the bright spots of reflective flares and light sources are always symmetrical around the lens's optical center, as shown in Fig. 3. By aligning the light source and optical center, we can easily locate the reflective flares, eliminating the need to search the entire image for possible occurrences. Besides, since the main bright spots are caused by focused reflective flares, they share the same attributes as the light source. For example, the reflective flares may display clear details for each small LED lamp bead, as shown in Fig. 3(b) and Fig. 3(c)." + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.093, + 0.695, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.561, + 0.216, + 0.662, + 0.227 + ], + "angle": 0, + "content": "(a) Samsung s7 edge" + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.093, + 0.855, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.724, + 0.216, + 0.81, + 0.226 + ], + "angle": 0, + "content": "(b) iPhone 13 Pro" + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.227, + 0.695, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.581, + 0.351, + 0.661, + 0.361 + ], + "angle": 0, + "content": "(c) Xiaomi Mi 8" + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.228, + 0.855, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.719, + 0.351, + 0.814, + 0.361 + ], + "angle": 0, + "content": "(d) Huawei Mate30" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.372, + 0.892, + 0.428 + ], + "angle": 0, + "content": "Figure 3. Strong light source may project a bright reflective flare on the photo. The reflective flares and corresponding light sources are symmetric around the optical center for most smartphone cameras. The yellow cross represents the optical center." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.439, + 0.696, + 0.455 + ], + "angle": 0, + "content": "3.2. Physical formulation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.459, + 0.892, + 0.59 + ], + "angle": 0, + "content": "To illustrate the formation of optical center symmetry prior, we suppose that an optical ray \\(\\mathbf{r}\\) is emitted from the light source, as shown in Fig. 4. This ray can be described as \\(\\mathbf{r} = [h\\theta ]^{\\top}\\). The states of this ray after each lens surface can be calculated by multiplying transfer matrices [10, 14]. We define the translation matrix at layer \\(i\\) as \\(\\mathbf{T_i}\\), and the refraction matrix at spherical surface with radius \\(R_{i}\\) and refractive index \\(n_i,n_{i + 1}\\) as \\(\\mathbf{F_i}\\). For the reflection at the ith layer with radius \\(R_{i}\\), the reflection matrix can be stated as \\(\\mathbf{R_i}\\). These matrices can be stated as:" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.599, + 0.892, + 0.63 + ], + "angle": 0, + "content": "\\[\n\\mathbf {T} _ {\\mathrm {i}} = \\left( \\begin{array}{c c} 1 & d _ {i} \\\\ 0 & 1 \\end{array} \\right), \\mathbf {R} _ {\\mathrm {i}} = \\left( \\begin{array}{c c} 1 & 0 \\\\ \\frac {2}{R _ {i}} & 0 \\end{array} \\right), \\mathbf {F} _ {\\mathrm {i}} = \\left( \\begin{array}{c c} 1 & 0 \\\\ \\frac {n _ {i + 1} - n _ {i}}{n _ {i + 1} R _ {i}} & \\frac {n _ {i}}{n _ {i + 1}} \\end{array} \\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.637, + 0.894, + 0.699 + ], + "angle": 0, + "content": "In Fig. 4, \\( h_1 \\) represents the position of a real image that focuses on the sensor through the lens' internal reflection. We define the ray's forward propagation matrix as \\( \\mathbf{A} \\) and \\( \\mathbf{C} \\) and the reflection process as \\( \\mathbf{B} \\). They are defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.576, + 0.708, + 0.891, + 0.763 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} A = \\mathbf {T} _ {4} \\mathbf {F} _ {3} \\mathbf {T} _ {3} \\mathbf {F} _ {2} \\mathbf {T} _ {2} \\mathbf {F} _ {1} \\mathbf {T} _ {1} \\mathbf {F} _ {0} \\mathbf {T} _ {0} \\\\ B = \\mathbf {T} _ {4} \\mathbf {F} _ {3} \\mathbf {T} _ {3} \\mathbf {R} _ {2} ^ {- 1} \\mathbf {T} _ {3} \\mathbf {F} _ {3} ^ {- 1} \\mathbf {T} _ {4} \\mathbf {R} _ {4} \\tag {2} \\\\ C = \\mathrm {T} _ {8} \\mathrm {F} _ {8} \\mathrm {T} _ {7} \\mathrm {F} _ {7} \\mathrm {T} _ {6} \\mathrm {F} _ {5} \\mathrm {T} _ {5} \\mathrm {F} _ {4} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.773, + 0.892, + 0.805 + ], + "angle": 0, + "content": "With these matrices, the ray after reflection in the focal plane \\(\\mathbf{r}_1\\) and direct imaging \\(\\mathbf{r}_0\\) can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.557, + 0.813, + 0.892, + 0.847 + ], + "angle": 0, + "content": "\\[\n\\binom {h _ {1}} {\\theta_ {1}} = \\mathbf {C B A} \\binom {h} {\\theta}, \\binom {h _ {0}} {\\theta_ {0}} = \\mathbf {C A} \\binom {h} {\\theta} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "For the focused image, \\( h_0 \\) and \\( h_1 \\) will not vary with the change of the angle of incidence \\( \\theta \\). Thus, we can deduce from Eq. (3) that both \\( h_0 \\) and \\( h_1 \\) are linear with \\( h \\). This" + }, + { + "type": "footer", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20785" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.093, + 0.462, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.311, + 0.473, + 0.438 + ], + "angle": 0, + "content": "Figure 4. Physical explanation of reflective flare formation. In this figure, \\( h \\) and \\( \\theta \\) indicate the height and the angle of incidence of the light source. \\( d \\) and \\( n \\) denote the thickness and refractive index of each layer. \\( A \\), \\( B \\), and \\( C \\) represent the transfer matrices for each part of the lens system. The red line represents the imaging after reflection and the grey line shows the path of direct imaging. This figure depicts the situation with reflection in the 3rd and 5th air-lens layer, in real-world situations, the reflections can happen at any place including the lens protector and CMOS layer." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.449, + 0.473, + 0.571 + ], + "angle": 0, + "content": "means that the line between the light source and the reflective flare will pass through the optical center and the ratio between the \\( h_0 \\) and \\( h_1 \\) is a constant. We define this ratio as \\( k \\) with \\( h_1 = kh_0 \\). Thus, we can calculate the position of the reflective flare, if this constant \\( k \\) is known. Due to the lens design for most smartphone cameras, the value of \\( k \\) is always -1. It leads to the optical center symmetry prior stated above." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.58, + 0.291, + 0.596 + ], + "angle": 0, + "content": "3.3. Experimental statistics" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.604, + 0.471, + 0.817 + ], + "angle": 0, + "content": "To demonstrate the effectiveness of our proposed prior in real-world situations, we conducted experiments using the main cameras of three smartphones: the Huawei P40, iPhone 13 Pro, and ZTE Axon 20 5G. In dark scenes, we captured 20 photos using each camera and marked the positions of the light source and reflective flare relative to the center of the image. As shown in Fig. 5, we observed that for all three cameras, the position of the reflective flare is linearly related to the position of the light source, with a slope of 1. However, since the light source may often create overexposed regions, manually marking the position of the light source can still introduce some errors. Considering this effect, we conclude that the results from all three phones align with our proposed prior." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.83, + 0.233, + 0.845 + ], + "angle": 0, + "content": "4. Data Collection" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Although the optical center prior can guide us in finding the reflective flare, the lack of paired images still presents a challenge for training an effective model for flare re" + }, + { + "type": "image", + "bbox": [ + 0.569, + 0.092, + 0.825, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.586, + 0.24, + 0.802, + 0.251 + ], + "angle": 0, + "content": "(a) Statistics of flare-light source relation" + }, + { + "type": "image", + "bbox": [ + 0.539, + 0.253, + 0.692, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.567, + 0.369, + 0.677, + 0.391 + ], + "angle": 0, + "content": "(b) flare-light source relation in x-axis" + }, + { + "type": "image", + "bbox": [ + 0.692, + 0.254, + 0.849, + 0.369 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.723, + 0.369, + 0.831, + 0.391 + ], + "angle": 0, + "content": "(c) flare-light source relation in y-axis" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.396, + 0.892, + 0.522 + ], + "angle": 0, + "content": "Figure 5. Experimental statistics of the optical center symmetry prior. We capture images with reflective flares in a dark environment and mark the position of the light source and reflective flares. \\( x_{f} \\) and \\( x_{l} \\) represent the x-coordinate of flare and light source and so as \\( y_{f} \\) and \\( y_{l} \\). The origin is set to be the center of the image. The relationship between the flare and the light source's absolute values of coordinates are plotted above. From figures (b) and (c), the coordinate relationships are linear and the slope is 1, which confirms our prior." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.532, + 0.894, + 0.608 + ], + "angle": 0, + "content": "moral. As reflective flares are caused by the lens system, it is physically impossible to capture real flare-free and flare-corrupted image pairs. Moreover, given the huge variety of nighttime lights, it would be tedious to synthesize reflective flares for each type of light source." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.893, + 0.901 + ], + "angle": 0, + "content": "To address the challenge of generating paired images for reflective flare removal, we propose a semi-synthetic dataset, named BracketFlare. Since reflective flares always have the same pattern as the brightest region of the light source, we reduce the exposure time to capture the patterns of the light source's brightest region, which are used to synthesize the reflective flares. Specifically, we mount a professional camera (Sony A7 III) on a tripod and use continuous bracketing to capture five photos of the same static scene with a step of \\(3\\mathrm{EV}\\) as a group. We then select the normally exposed image and the lowest exposure image from each group to compose the pair. To increase the diversity and realism of our dataset, we choose scenes with different light sources that are bright enough to leave reflective flares. Finally, we collect a dataset consisting of 440 pairs of 4K images as training data and 40 images as test data. A general overview of our dataset is depicted in Fig. 6, showing that the dataset contains diverse and commonly seen flare patterns." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20786" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.122, + 0.098, + 0.307, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.158, + 0.194, + 0.179, + 0.203 + ], + "angle": 0, + "content": "LED" + }, + { + "type": "image", + "bbox": [ + 0.315, + 0.099, + 0.642, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.122, + 0.205, + 0.307, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.144, + 0.301, + 0.191, + 0.31 + ], + "angle": 0, + "content": "Traffic light" + }, + { + "type": "image", + "bbox": [ + 0.315, + 0.272, + 0.645, + 0.309 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.659, + 0.098, + 0.844, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.665, + 0.194, + 0.746, + 0.204 + ], + "angle": 0, + "content": "Square overhead light" + }, + { + "type": "image_caption", + "bbox": [ + 0.756, + 0.194, + 0.837, + 0.203 + ], + "angle": 0, + "content": "Round overhead light" + }, + { + "type": "image", + "bbox": [ + 0.659, + 0.206, + 0.843, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.686, + 0.302, + 0.725, + 0.311 + ], + "angle": 0, + "content": "Light tube" + }, + { + "type": "image_caption", + "bbox": [ + 0.776, + 0.302, + 0.817, + 0.312 + ], + "angle": 0, + "content": "Light bulb" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.32, + 0.892, + 0.349 + ], + "angle": 0, + "content": "Figure 6. Visualization and distribution of our dataset. It contains different types of light sources in diverse scenes. Based on light source types, we classify them into eight categories with different flare patterns." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.366, + 0.251, + 0.384 + ], + "angle": 0, + "content": "5. Proposed Method" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.392, + 0.318, + 0.408 + ], + "angle": 0, + "content": "5.1. Postprocessing our dataset" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.416, + 0.469, + 0.611 + ], + "angle": 0, + "content": "To increase the diversity of our dataset, we generate paired flare-corrupted and flare-free images on the fly with augmentation. As depicted in Fig. 7(a), we first resize the 4K image to \\(800 \\times 1200\\). To simulate the possible misalignment of the image's center and optical center, we introduce a slight translation with \\(t \\sim U(0,15)\\) pixels to the underexposed light source image. Then, we randomly crop the paired light source image and the normally exposed image with a size of \\(512 \\times 512\\). Since our captured images are already gamma-encoded, we follow the settings of Wu et al. [22] and linearize them by applying a gamma correction, i.e., \\(I_g = I^\\gamma\\) with \\(\\gamma\\) randomly sampled from [1.8, 2.2]. All our operations are performed in the linear space." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.613, + 0.469, + 0.748 + ], + "angle": 0, + "content": "Synthetic reflective flare. As lenses reflect light of different wavelengths with varying reflectivity, reflective flares often appear in different colors. To replicate this, we use a random color factor to multiply the low-exposed image. To simulate an unfocused situation, we also add random blur. In accordance with the optical center symmetry prior, we rotate the light source image to synthesize the reflective flare. Finally, we obtain the synthetic reflective flare by multiplying the flare image with an opacity." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.469, + 0.854 + ], + "angle": 0, + "content": "Flare-free image. The ISO is fixed for all images in our continuous bracketing. Therefore, our normally exposed images always share a long exposure time and have little noise. To simulate real night scenes, we add Gaussian noise to the images. Finally, we randomly apply a small color offset to the flare-free image. These operations aid in simulating light fog and glare effects in real-world scenarios." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Flare-corrupted image. To create flare-corrupted images, we add the synthetic reflective flare to the flare-free images. Then, we apply gamma correction to the images with the" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.368, + 0.892, + 0.429 + ], + "angle": 0, + "content": "gamma value obtained by taking the reciprocal of the previously sampled gamma correction value \\(\\gamma\\). The resulting gamma-encoded flare-corrupted image, flare-free image, and flare image are used to train the neural networks." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.436, + 0.71, + 0.453 + ], + "angle": 0, + "content": "5.2. Flare removal pipeline" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.46, + 0.892, + 0.535 + ], + "angle": 0, + "content": "As the optical center symmetry prior is a global prior, it is difficult for CNN-based networks to learn it, and even Transformer-based networks struggle to learn such symmetry. Therefore, it is crucial to incorporate this prior into the network with specialized designs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.535, + 0.893, + 0.746 + ], + "angle": 0, + "content": "To improve the network's ability to learn the optical center symmetry prior, as shown in Fig. 7(b), we apply a gamma correction of \\( I_{g} = I^{\\gamma} \\) to the input flare-corrupted image with \\( \\gamma = 10 \\). This operation extracts the nearly saturated areas of the light source in each channel. We then rotate this nearly saturated image by 180 degrees to obtain a prior image, which is concatenated with the flare-corrupted image to form a six-channel input. This rotated prior image provides an initial approximation of the reflective flare's pattern, which we expect the network to refine and enhance. The network's output is set to six channels, with three channels for the estimated flare-free image and three channels for the estimated flare. Our synthetic data provides supervision for all six channels, which facilitates network training." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.755, + 0.64, + 0.769 + ], + "angle": 0, + "content": "5.3. Loss function" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.778, + 0.892, + 0.837 + ], + "angle": 0, + "content": "With the supervision of flare image \\( I_{F} \\) and flare-free image \\( I_{B} \\), our final loss is the combination of background loss \\( \\mathcal{L}_{bg} \\), flare loss \\( \\mathcal{L}_{flare} \\), and reconstruction loss \\( \\mathcal{L}_{rec} \\). It can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.605, + 0.847, + 0.892, + 0.864 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {r e c}} + \\mathcal {L} _ {\\text {f l a r e}} + \\mathcal {L} _ {\\text {b g}}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.903 + ], + "angle": 0, + "content": "We define our flare removal network as \\( f_{\\theta} \\). It takes flare-corrupted image \\( I_{input} \\) and its prior image \\( I_{prior} \\) as inputs." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20787" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.139, + 0.096, + 0.82, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.251, + 0.288, + 0.404, + 0.3 + ], + "angle": 0, + "content": "(a) Training Data Generation" + }, + { + "type": "image_caption", + "bbox": [ + 0.621, + 0.288, + 0.738, + 0.301 + ], + "angle": 0, + "content": "(b) Network Learning" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.304, + 0.892, + 0.374 + ], + "angle": 0, + "content": "Figure 7. The pipeline of our data processing module and flare removal network. The flare image and a flare-free image from our paired dataset are resized and randomly cropped. After data augmentation, they are added to synthesize the corresponding flare-corrupted image. To encode the optical symmetry prior, the input flare-corrupted image is center flipped as the prior image. We also apply gamma correction to extract the light source region. Then, the prior image is concatenated with the flare-corrupted image as the input with 6 channels. Finally, the network predicts an output with 6 channels (a flare-free image with 3 channels and an estimated flare with 3 channels)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.393, + 0.414, + 0.409 + ], + "angle": 0, + "content": "The flare-free image and flare can be estimated by:" + }, + { + "type": "equation", + "bbox": [ + 0.179, + 0.418, + 0.469, + 0.436 + ], + "angle": 0, + "content": "\\[\n\\hat {I} _ {B}, \\hat {I} _ {F} = f _ {\\theta} \\left(I _ {\\text {i n p u t}}, I _ {\\text {p r i o r}}\\right). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.446, + 0.468, + 0.476 + ], + "angle": 0, + "content": "With the estimated flare and flare-free images, the \\(L_{1}\\) reconstruction loss can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.486, + 0.469, + 0.505 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r e c}} = \\left| I _ {\\text {i n p u t}} - \\hat {I} _ {F} \\oplus \\hat {I} _ {B} \\right|, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.515, + 0.469, + 0.543 + ], + "angle": 0, + "content": "where \\(\\oplus\\) denotes an element-wise addition operation in the linearized gamma-decoded domain." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.545, + 0.47, + 0.694 + ], + "angle": 0, + "content": "For the background loss, it consists of a default \\(\\mathcal{L}_1\\), a region-aware masked \\(\\mathcal{L}_{mask}\\) term, and a perceptual term \\(\\mathcal{L}_{vgg}\\) with a pre-trained VGG-19 network [17]. Compared with previous works [5, 22, 26], the main difference is the \\(\\mathcal{L}_{mask}\\). Since the regions of reflective flares are always small, a simple \\(L_1\\) loss may ignore these regions, leading to the local optimum. To encourage the network to focus on restoring the flare-corrupted regions, we calculate the flare's mask \\(I_M\\) (\\(>0.2\\)) from the ground truth of flare image \\(I_F\\). This masked loss can be written as:" + }, + { + "type": "equation", + "bbox": [ + 0.178, + 0.706, + 0.469, + 0.723 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {m a s k}} = \\left| I _ {M} * \\left(\\hat {I} _ {B} - I _ {B}\\right) \\right|. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.733, + 0.408, + 0.75 + ], + "angle": 0, + "content": "Overall, the background loss \\(\\mathcal{L}_{bg}\\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.149, + 0.761, + 0.469, + 0.777 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {b g} = w _ {1} \\mathcal {L} _ {1} + w _ {2} \\mathcal {L} _ {v g g} + w _ {3} \\mathcal {L} _ {\\text {m a s k}}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.787, + 0.469, + 0.832 + ], + "angle": 0, + "content": "where \\( w_{1}, w_{2}, w_{3} \\) are respectively set to 0.5, 0.1, and 20.0 in our experiments. Flare loss \\( \\mathcal{L}_{flare} \\) has the identical form as the background loss." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.846, + 0.21, + 0.863 + ], + "angle": 0, + "content": "6. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Evaluation on synthetic data. To evaluate our method, we collect 40 paired data using the process mentioned in" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.391, + 0.892, + 0.432 + ], + "angle": 0, + "content": "Table 1. Percentage of users favoring our results (our dataset + MPRNet†) vs. previous work. Our method outperforms existing methods on both indoor and outdoor images." + }, + { + "type": "table", + "bbox": [ + 0.541, + 0.437, + 0.848, + 0.532 + ], + "angle": 0, + "content": "
MethodOutdoorIndoor
Reflection removal [26]3.0%8.5%
Flare spot removal [1]4.5%0.0%
Flare removal (Wu et al.) [22]0.5%3.5%
Flare removal (Dai et al.) [5]1.5%2.5%
Ours90.5%85.5%
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.536, + 0.892, + 0.581 + ], + "angle": 0, + "content": "Sec. 4 to generate a synthetic test dataset, in which ground truth images are available. We adopt PSNR, SSIM [20], and LPIPS [25] as our evaluation metrics." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.581, + 0.892, + 0.777 + ], + "angle": 0, + "content": "Evaluation on real data. To show our method's generalization ability, we also collect a dataset with the photos captured by different types of mobile phones, including iPhone 12 pro, iPhone 13 pro, iPhone 14 pro, Huawei mate 30, Huawei nova 5i Pro, Huawei P40, Samsung Galaxy s22, Samsung Galaxy Note9, ZTE Axon 20 5G and Oppo Reno4 SE 5G. For real cases, we cannot obtain the ground truth. Since reflective flares always locate in small regions, they are hard to be evaluated by existing non-reference image quality assessment. These metrics emphasize more on global image quality and are not specially designed for our task. Therefore, we provide more visual results and conduct a user study to evaluate the performance." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.787, + 0.816, + 0.803 + ], + "angle": 0, + "content": "6.1. Comparison with different networks" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.901 + ], + "angle": 0, + "content": "To validate the performance of our pipeline, we train different CNN-based and Transformer-based networks with our data. We use the default network settings of MPR-Net [24], HINet [4], and Uformer [21] and reduce the feature channels of Restormer [23] from 48 to 24 due to GPU memory constraint. In our experiments, these models are" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20788" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.149, + 0.091, + 0.257, + 0.176 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.149, + 0.177, + 0.256, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.149, + 0.261, + 0.256, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.187, + 0.347, + 0.217, + 0.359 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.26, + 0.092, + 0.367, + 0.176 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.26, + 0.177, + 0.367, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.26, + 0.261, + 0.367, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.267, + 0.347, + 0.364, + 0.358 + ], + "angle": 0, + "content": "Reflection removal" + }, + { + "type": "image", + "bbox": [ + 0.371, + 0.092, + 0.479, + 0.176 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.371, + 0.177, + 0.479, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.371, + 0.261, + 0.479, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.382, + 0.347, + 0.475, + 0.358 + ], + "angle": 0, + "content": "Flare spot removal" + }, + { + "type": "image", + "bbox": [ + 0.481, + 0.092, + 0.589, + 0.176 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.481, + 0.177, + 0.589, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.481, + 0.261, + 0.589, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.517, + 0.347, + 0.563, + 0.358 + ], + "angle": 0, + "content": "Wuet al." + }, + { + "type": "image", + "bbox": [ + 0.593, + 0.092, + 0.701, + 0.176 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.593, + 0.177, + 0.701, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.593, + 0.261, + 0.701, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.627, + 0.347, + 0.673, + 0.358 + ], + "angle": 0, + "content": "Dai et al." + }, + { + "type": "image", + "bbox": [ + 0.705, + 0.092, + 0.812, + 0.176 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.705, + 0.177, + 0.812, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.705, + 0.261, + 0.812, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.746, + 0.347, + 0.772, + 0.358 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.118, + 0.363, + 0.85, + 0.378 + ], + "angle": 0, + "content": "Figure 8. Visual comparison of different flare removal and reflection removal methods on real-world nighttime flare images." + }, + { + "type": "image", + "bbox": [ + 0.089, + 0.383, + 0.2, + 0.47 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.089, + 0.47, + 0.2, + 0.556 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.126, + 0.556, + 0.153, + 0.566 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.202, + 0.383, + 0.312, + 0.47 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.202, + 0.47, + 0.312, + 0.555 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.212, + 0.556, + 0.304, + 0.566 + ], + "angle": 0, + "content": "Reflection removal" + }, + { + "type": "image", + "bbox": [ + 0.315, + 0.383, + 0.425, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.315, + 0.47, + 0.425, + 0.556 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.328, + 0.556, + 0.418, + 0.567 + ], + "angle": 0, + "content": "Flare spot removal" + }, + { + "type": "image", + "bbox": [ + 0.427, + 0.383, + 0.538, + 0.47 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.427, + 0.47, + 0.538, + 0.556 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.458, + 0.557, + 0.501, + 0.566 + ], + "angle": 0, + "content": "Wu et al." + }, + { + "type": "image", + "bbox": [ + 0.54, + 0.383, + 0.652, + 0.47 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.54, + 0.47, + 0.652, + 0.556 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.569, + 0.556, + 0.612, + 0.566 + ], + "angle": 0, + "content": "Dai et al." + }, + { + "type": "image", + "bbox": [ + 0.653, + 0.383, + 0.764, + 0.47 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.653, + 0.47, + 0.764, + 0.556 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.696, + 0.556, + 0.721, + 0.566 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.765, + 0.383, + 0.877, + 0.47 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.765, + 0.47, + 0.877, + 0.556 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.788, + 0.556, + 0.851, + 0.566 + ], + "angle": 0, + "content": "Ground truth" + }, + { + "type": "image_caption", + "bbox": [ + 0.121, + 0.572, + 0.845, + 0.587 + ], + "angle": 0, + "content": "Figure 9. Visual comparison of different flare removal and reflection removal methods on synthetic nighttime flare images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.602, + 0.47, + 0.738 + ], + "angle": 0, + "content": "trained on an NVIDIA Geforce RTX3090 with a batch size of 2. We use Adam optimizer [9] with a learning rate of 1e-4 and train a model for 500k iterations. Besides, we adopt a MultiStepLR with a milestone of 200k and set the gamma to 0.5. These models are evaluated using the synthetic test dataset as shown in Table 4. From experiments, we find that \\(\\mathrm{MRPNet}^{\\dagger}\\) [24] performs better than other networks. Thus, we treat \\(\\mathrm{MPRNet}^{\\dagger}\\) [24] as our baseline method and conduct the following ablation study." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.746, + 0.369, + 0.761 + ], + "angle": 0, + "content": "6.2. Comparison with previous works" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.47, + 0.901 + ], + "angle": 0, + "content": "We compare our method with different flare removal datasets [5, 22] and reflection removal dataset [26] on both real and synthetic data. We use the pre-trained models of Dai et al. [5] and Zhang et al. [26]. Since Wu et al. [22] do not release their model, we retrain their network using their provided data and code. Besides, we also compare our method with the traditional flare spot removal method [1]. As shown in Table 2, since most of these methods are not specially designed for reflective flare removal, they cannot" + }, + { + "type": "table_caption", + "bbox": [ + 0.53, + 0.601, + 0.86, + 0.614 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison on synthetic test data." + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.618, + 0.89, + 0.715 + ], + "angle": 0, + "content": "
MethodPSNRSSIMLPIPSMasked PSNR
Input37.300.9900.02521.68
Reflection removal [26]22.020.8300.07422.88
Flare spot removal [1]38.290.9900.02422.53
Flare removal (Wu et al.) [22]26.130.8950.05521.89
Flare removal (Dai et al.) [5]27.490.9420.05021.85
Ours48.410.9940.00432.09
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.718, + 0.892, + 0.778 + ], + "angle": 0, + "content": "accurately remove the reflective flare while retaining image details simultaneously. Although Asha et al. [1] can detect the bright spots, as shown in Fig. 8, it may also remove the light source which is quite common at night." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.779, + 0.892, + 0.824 + ], + "angle": 0, + "content": "User Study. To evaluate these methods in real-world situations, we also conduct a user study with 20 participants with 20 images randomly selected from our real data in Table 1." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.833, + 0.65, + 0.849 + ], + "angle": 0, + "content": "6.3. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Loss function. We conduct an ablation study to prove the necessity of each loss function used in our baseline. In our experiments, we train the model without adding \\(\\mathcal{L}_{\\text {flare }}\\)," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20789" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.086, + 0.136, + 0.101, + 0.162 + ], + "angle": 270, + "content": "Input" + }, + { + "type": "image_caption", + "bbox": [ + 0.086, + 0.232, + 0.101, + 0.264 + ], + "angle": 270, + "content": "Output" + }, + { + "type": "image", + "bbox": [ + 0.106, + 0.096, + 0.232, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.106, + 0.195, + 0.231, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.126, + 0.293, + 0.21, + 0.304 + ], + "angle": 0, + "content": "(a) iPhone 14 Pro" + }, + { + "type": "image", + "bbox": [ + 0.236, + 0.096, + 0.361, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.236, + 0.195, + 0.361, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.248, + 0.293, + 0.333, + 0.304 + ], + "angle": 0, + "content": "(b) iPhone 13 Pro" + }, + { + "type": "image", + "bbox": [ + 0.366, + 0.095, + 0.491, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.366, + 0.195, + 0.49, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.375, + 0.293, + 0.474, + 0.304 + ], + "angle": 0, + "content": "(c) ZTE Axon 20 5G" + }, + { + "type": "image", + "bbox": [ + 0.495, + 0.096, + 0.619, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.495, + 0.195, + 0.619, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.496, + 0.293, + 0.615, + 0.304 + ], + "angle": 0, + "content": "(d) Samsung Galaxy S22" + }, + { + "type": "image", + "bbox": [ + 0.623, + 0.096, + 0.748, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.623, + 0.195, + 0.748, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.636, + 0.293, + 0.731, + 0.303 + ], + "angle": 0, + "content": "(e) Huawei Mate 30" + }, + { + "type": "image", + "bbox": [ + 0.752, + 0.096, + 0.877, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.752, + 0.195, + 0.877, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.76, + 0.293, + 0.869, + 0.303 + ], + "angle": 0, + "content": "(f) Huawei nova 5i Pro" + }, + { + "type": "image_caption", + "bbox": [ + 0.11, + 0.318, + 0.858, + 0.333 + ], + "angle": 0, + "content": "Figure 10. Visual comparison of flare removal on real-world nighttime flare images that are captured by different smartphones." + }, + { + "type": "table_caption", + "bbox": [ + 0.12, + 0.343, + 0.425, + 0.357 + ], + "angle": 0, + "content": "Table 3. Ablation study for different loss functions." + }, + { + "type": "table", + "bbox": [ + 0.108, + 0.36, + 0.434, + 0.454 + ], + "angle": 0, + "content": "
LossPSNRSSIMLPIPSMasked PSNR
w/o Lflare47.360.9930.00430.63
w/o Lrec48.180.9940.00331.33
w/o Lvgg48.050.9940.00431.39
w/o Lmask48.010.9930.00331.03
Full48.410.9940.00432.09
" + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.461, + 0.211, + 0.544 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.546, + 0.211, + 0.628 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.142, + 0.63, + 0.171, + 0.64 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.214, + 0.461, + 0.321, + 0.544 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.214, + 0.546, + 0.321, + 0.628 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.244, + 0.63, + 0.292, + 0.64 + ], + "angle": 0, + "content": "w/o prior" + }, + { + "type": "image", + "bbox": [ + 0.323, + 0.461, + 0.431, + 0.544 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.323, + 0.546, + 0.431, + 0.628 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.351, + 0.63, + 0.403, + 0.64 + ], + "angle": 0, + "content": "with prior" + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.643, + 0.47, + 0.672 + ], + "angle": 0, + "content": "Figure 11. Visual comparison with the network's performance with and w/o optical center symmetry prior." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.685, + 0.469, + 0.776 + ], + "angle": 0, + "content": "\\(\\mathcal{L}_{rec}\\), \\(\\mathcal{L}_{vgg}\\) or \\(\\mathcal{L}_{mask}\\). We report the results in Table 3. The ablation study of \\(\\mathcal{L}_{flare}\\) proves the significance of bringing supervision for both flare and background images. Besides, when the \\(L_{mask}\\) is removed, it degrades the reconstruction results in the flare-corrupted regions, which verifies the importance of focusing on flare regions." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.78, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Prior. To illustrate the performance of introducing the optical symmetry prior by concatenating a prior image, we train these baseline restoration networks with the original input. Since HINet [4] and MPRNet [24] cannot support input and output with different channels, we concatenate a black image with no information and keep other settings the same. The quantitative results are shown in Table 4, and visual comparisons on real-world images are presented in Fig." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.343, + 0.892, + 0.385 + ], + "angle": 0, + "content": "Table 4. Quantitative comparison for different network structures on synthetic test data. ‘†’ indicates that the network is equipped with the proposed prior." + }, + { + "type": "table", + "bbox": [ + 0.514, + 0.392, + 0.871, + 0.535 + ], + "angle": 0, + "content": "
NetworkPSNRSSIMLPIPSMasked PSNR
MPRNet [24]43.600.9920.01127.64
Uformer [21]42.570.9890.00927.77
HINet [4]43.780.9910.01227.73
Restormer [23]42.760.9910.01227.27
MPRNet†48.410.9940.00432.09
Uformer†47.470.9910.00331.57
HINet†48.030.9940.00330.88
Restormer†48.110.9940.00431.07
" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.545, + 0.892, + 0.606 + ], + "angle": 0, + "content": "11. It shows that embedding this prior in the network can boost the performance of quantitative results considerably. In Fig. 11, we observe that this prior helps the network to locate and remove the reflective flare precisely." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.608, + 0.619, + 0.622 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.626, + 0.892, + 0.838 + ], + "angle": 0, + "content": "We have presented a novel approach to tackle the challenge of removing reflective flares in nighttime photography. Our approach is inspired by the optical center symmetry prior, which suggests that reflective flares and their corresponding light sources are always symmetrical around the optical center of the lens system in smartphone cameras. Based on the observation that the patterns of reflective flares are always similar to the brightest region of the light source, we introduce a dataset named BracketFlare. We further propose a reflective flare removal pipeline that encodes the optical center symmetry prior into the network. Experimental results demonstrate that our proposed methods are effective in removing reflective flares in real-world scenarios and have the potential for wide-ranging applications." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.84, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Acknowledgement: This study is supported under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s)." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20790" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.469, + 0.156 + ], + "angle": 0, + "content": "[1] C S Asha, Sooraj Kumar Bhat, Deepa Nayak, and Chaithra Bhat. Auto removal of bright spot from images captured against flashing light source. In ICDC, 2019. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.157, + 0.469, + 0.199 + ], + "angle": 0, + "content": "[2] Herbert Bay, Andreas Ess, Tinne Tuytelaars, and Luc Van Gool. Speeded-up robust features (surf). Computer Vision and Image Understanding, 110(3):346-359, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.2, + 0.469, + 0.242 + ], + "angle": 0, + "content": "[3] Floris Chabert. Automated lens flare removal, 2014. Technical Report, Department of Electrical Engineering, Stanford University. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.243, + 0.469, + 0.283 + ], + "angle": 0, + "content": "[4] Liangyu Chen, Xin Lu, Jie Zhang, Xiaojie Chu, and Chengpeng Chen. Hinet: Half instance normalization network for image restoration. In CVPRW, 2021. 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.284, + 0.469, + 0.352 + ], + "angle": 0, + "content": "[5] Yuekun Dai, Chongyi Li, Shangchen Zhou, Ruicheng Feng, and Chen Change Loy. Flare7k: A phenomenological nighttime flare removal dataset. In Thirty-sixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2022. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.354, + 0.469, + 0.408 + ], + "angle": 0, + "content": "[6] Ruicheng Feng, Chongyi Li, Huaijin Chen, Shuai Li, Chen Change Loy, and Jinwei Gu. Removing diffraction image artifacts in under-display camera via dynamic skip connection network. In CVPR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.41, + 0.469, + 0.45 + ], + "angle": 0, + "content": "[7] Matthias Hullin, Elmar Eisemann, Hans-Peter Seidel, and Sungkil Lee. Physically-based real-time lens flare rendering. In ACM SIGGRAPH, 2011. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.452, + 0.469, + 0.493 + ], + "angle": 0, + "content": "[8] Yeying Jin, Wenhan Yang, and Robby T Tan. Unsupervised night image enhancement: When layer decomposition meets light-effects suppression. In ECCV. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.494, + 0.469, + 0.522 + ], + "angle": 0, + "content": "[9] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.523, + 0.469, + 0.562 + ], + "angle": 0, + "content": "[10] Sungkil Lee and Elmar Eisemann. Practical real-time lensflare rendering. Computer Graphics Forum, 32(4):1-6, 2013. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.565, + 0.469, + 0.604 + ], + "angle": 0, + "content": "[11] Yu Li, Robby T. Tan, and Michael S. Brown. Nighttime haze removal with glow and multiple light colors. In ICCV, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.607, + 0.469, + 0.647 + ], + "angle": 0, + "content": "[12] David G Lowe. Distinctive image features from scale-invariant keypoints. International Journal of Computer Vision, 60(2):91-110, 2004. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.649, + 0.469, + 0.689 + ], + "angle": 0, + "content": "[13] Andreas Nussberger, Helmut Grabner, and Luc Van Gool. Robust aerial object tracking in images with lens flare. In ICRA, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.691, + 0.469, + 0.718 + ], + "angle": 0, + "content": "[14] Frank L Pedrotti, Leno M Pedrotti, and Leno S Pedrotti. Introduction to Optics. Cambridge University Press, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.72, + 0.469, + 0.76 + ], + "angle": 0, + "content": "[15] Xiaotian Qiao, Gerhard P. Hancke, and Rynson W. H. Lau. Light source guided single-image flare removal from unpaired data. In ICCV, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.469, + 0.802 + ], + "angle": 0, + "content": "[16] Aashish Sharma and Robby T. Tan. Nighttime visibility enhancement by increasing the dynamic range and suppression of light effects. In CVPR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.469, + 0.857 + ], + "angle": 0, + "content": "[17] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In International Conference on Learning Representations, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[18] Qilin Sun, Ethan Tseng, Qiang Fu, Wolfgang Heidrich, and Felix Heide. Learning rank-1 diffractive optics for single-shot high dynamic range imaging. In CVPR, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.134 + ], + "angle": 0, + "content": "[19] Patricia Vitoria and Coloma Ballester. Automatic flare spot artifact detection and removal in photographs. Journal of Mathematical Imaging and Vision, 61:515-533, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[20] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[21] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restorationn. In CVPR, 2022. 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.303 + ], + "angle": 0, + "content": "[22] Yicheng Wu, Qiurui He, Tianfan Xue, Rahul Garg, Jiawen Chen, Ashok Veeraraghavan, and Jonathan T Barron. How to train neural networks for flare removal. In ICCV, 2021. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.305, + 0.892, + 0.359 + ], + "angle": 0, + "content": "[23] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In CVPR, 2022. 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.362, + 0.892, + 0.416 + ], + "angle": 0, + "content": "[24] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Multi-stage progressive image restoration. In CVPR, 2021. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.418, + 0.892, + 0.46 + ], + "angle": 0, + "content": "[25] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.461, + 0.892, + 0.502 + ], + "angle": 0, + "content": "[26] Xuaner Zhang, Ren Ng, and Qifeng Chen. Single image reflection separation with perceptual losses. In CVPR, 2018. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.504, + 0.892, + 0.545 + ], + "angle": 0, + "content": "[27] Junyan Zhu, Taesung Park, Phillip Isola, and Alexei A. Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In ICCV, 2017. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.545 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20791" + } + ] +] \ No newline at end of file diff --git a/2023/Nighttime Smartphone Reflective Flare Removal Using Optical Center Symmetry Prior/d34edf55-17a4-4aa7-b0c0-de157905c27a_origin.pdf b/2023/Nighttime Smartphone Reflective Flare Removal Using Optical Center Symmetry Prior/d34edf55-17a4-4aa7-b0c0-de157905c27a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..fdc196405a52caa921c453b3d8c3153bbfcb4b9b --- /dev/null +++ b/2023/Nighttime Smartphone Reflective Flare Removal Using Optical Center Symmetry Prior/d34edf55-17a4-4aa7-b0c0-de157905c27a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:77f3401cfe396a25c9128db0ffc469b5ee040fa380e89cead3d782fd2a268379 +size 3163498 diff --git a/2023/Nighttime Smartphone Reflective Flare Removal Using Optical Center Symmetry Prior/full.md b/2023/Nighttime Smartphone Reflective Flare Removal Using Optical Center Symmetry Prior/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4bea9f5d6abe3026f944f7e26a8cf2d1ca549109 --- /dev/null +++ b/2023/Nighttime Smartphone Reflective Flare Removal Using Optical Center Symmetry Prior/full.md @@ -0,0 +1,434 @@ +# Nighttime Smartphone Reflective Flare Removal Using Optical Center Symmetry Prior + +Yuekun Dai Yihang Luo Shangchen Zhou Chongyi Li* Chen Change Loy S-Lab, Nanyang Technological University + +{YDAI005, c200211, s200094, chongyi.li, ccloy}@ntu.edu.sg + +https://ykdai.github.io/projects/BracketFlare + +# Abstract + +Reflective flare is a phenomenon that occurs when light reflects inside lenses, causing bright spots or a "ghosting effect" in photos, which can impact their quality. Eliminating reflective flare is highly desirable but challenging. Many existing methods rely on manually designed features to detect these bright spots, but they often fail to identify reflective flares created by various types of light and may even mistakenly remove the light sources in scenarios with multiple light sources. To address these challenges, we propose an optical center symmetry prior, which suggests that the reflective flare and light source are always symmetrical around the lens's optical center. This prior helps to locate the reflective flare's proposal region more accurately and can be applied to most smartphone cameras. Building on this prior, we create the first reflective flare removal dataset called BracketFlare, which contains diverse and realistic reflective flare patterns. We use continuous bracketing to capture the reflective flare pattern in the underexposed image and combine it with a normally exposed image to synthesize a pair of flare-corrupted and flare-free images. With the dataset, neural networks can be trained to remove the reflective flares effectively. Extensive experiments demonstrate the effectiveness of our method on both synthetic and real-world datasets. + +# 1. Introduction + +Lens flare artifacts can occur when a strong light enters the camera's field of view, causing scattering and reflection within the lenses. Unlike radial-shaped scattering flares, reflective flares appear as groups of bright spots [5, 7] or a "ghosting effect" like the examples in Fig. 1. With multiple reflections, the light source creates multiple projections at various positions behind the lens. If the projection is far from the focal plane, the defocused effect can brighten the whole sensor and degrade the photo's quality as shown in Fig. 2. In near-focus situations, images of the light source + +![](images/6550e8c0093d203935d90d2312c49717f38ee94b2f8313da514f0a03d6d41531.jpg) + +![](images/6a2d7d55295dd1bda06f1347e5622d3139648b025bc507ac815583c9e9b3b57d.jpg) + +![](images/db046f38b982a1a71dac44a5a8b723976beb85ad2768ba40d05efe3fb1064804.jpg) +(a) iPhone 13 Pro + +![](images/aa72b59fb4ae6537452c603ca696e6268b35c3cf36362e84a3deec27e97e23b6.jpg) + +![](images/3f957143ddac45f578f5cee0a8768434c0ea7afb57bdf90a514c705ffc6997db.jpg) + +![](images/c7cbf74198cd8f53741fb18de3446471d90cfba1288b0541eaeafbce1116a92d.jpg) +(b) Huawei P40 + +![](images/4696c233bba023fd7d15fd38e4db745a7c2491186093e60284533358cc7363f8.jpg) + +![](images/170f32acdf35ba7a862153536978978c671f1632e48a74b74f0c0c725306ab66.jpg) + +![](images/ebeaf9b3ad5a75adf7db2d1c95926f6e12a918fdca2b156ad57d4dab3b3376f8.jpg) +(c) Sony ILCE-6400 +Figure 1. Reflective flare removal with our proposed dataset and prior. The network trained on our dataset can effectively separate the reflective flare and flare-free images from a flare-corrupted image. Our method can generalize well to different types of smartphones and even cameras with specific types of lenses. + +can result in a line of unfocused, bright blobs on the captured photo. + +Applying anti-reflective (AR) coating is a common technique for reducing internal reflections [22]. The coating is effective for unfocused spots with low luminance, but cannot mitigate a bright spot with high luminance at the focal plane. Even with a well-designed lens system, a bright reflective spot may remain in the photo, known as the ghosting effect [10]. Reflective flare is common on all types of smartphone lenses, and certain types of smartphones (e.g., the iPhone series) can produce a large area of inverted reflective flare from neon lights or large, luminous standing signs, as shown in Fig. 1. The phenomenon becomes more pronounced when capturing a photo at night with multiple lights in the scene. Numerous bright spots and inverted re + +flections manifest and propagate in different directions with the lens orientation. Therefore, reflective flare removal algorithms are highly desirable. + +Removing reflective flare is a difficult task due to the vast diversity of reflective flares that can be caused by different types of light sources and lenses. Reflective flares often appear as bright spots that can resemble streetlights from a distance, making it challenging to differentiate between reflective flares and other light sources. Current methods for removing flare spots mainly rely on feature detection techniques like Speeded-Up Robust Features [2] (SURF) or Scale-Invariant Feature Transform [12] (SIFT) to extract spot regions [1, 3, 19]. These methods, however, struggle to effectively detect blobs on complex backgrounds and only work for specific spot patterns. Moreover, these methods are ineffective in removing reflective flares caused by lamp tubes and LED matrices. Many learning-based methods and datasets for reflective flare removal have been proposed [5, 15, 18, 22]. Although these datasets contain subsets of reflective flares, they are not sufficient for good generalization. For instance, the reflective flare subset from Wu et al. [22] is captured using the same lens and light source, making it difficult to generalize to different flares in real-world scenarios. Dai et al. [5] propose a synthetic dataset with 10 types of reflective flares, but training on this dataset tends to mislead a network to erroneously remove other tiny bright spots like street lights and bright windows. As a result, it remains challenging for existing algorithms to accurately detect and remove various types of reflective flares while preserving details in other regions. + +To overcome the challenges of detecting and removing reflective flares, we propose a novel optical center symmetry prior, which suggests that the main spot of the reflective flare and the light source are always symmetrical around the lens's optical center in captured images. This prior applies to most smartphone cameras and some professional cameras. With this prior, we can easily detect the possible locations of reflective flares. We also found that the flare spots always have the same patterns as the brightest regions of the light source. Therefore, we can synthesize reflective flares by using the light source's patterns at low exposure. Based on these characteristics of reflective flares, we develop the first reflective flare removal dataset, named BracketFlare. Specifically, we use continuous bracketing to capture the light source's pattern on an underexposed image and select the normally exposed image to create the light source and the natural image. Next, we rotate the light source pattern 180 degrees around the optical center, apply blur and color augmentation, and use it to create synthetic reflective flares. We can add the reflective flare to the normally exposed image to produce paired flare-corrupted and flare-free images. + +Thanks to the optical center symmetry prior, we further propose an end-to-end pipeline for reflective flare re + +![](images/a5c94693cc237f600ed1a2d8c35288496b35992a6debe28279e859ef8a2afb20.jpg) + +![](images/3699cdd3c26ddf177d851d381fe7068b686880a3c840cac5358bc21d8b3e68fd.jpg) + +![](images/36baeab3108a00989db2c702a828df4ad8591c55495ee080ab0fbdfefc3dec07.jpg) +(a) w/o AR coating +(b) With AR coating +(c) Description of reflective flare formation +Figure 2. Schematic diagram of reflective flare formation. Reflective flares are generated by multiple reflections inside the lens. In (c), yellow lines show the generation principle of the aperture-shaped reflective flare. Red lines represent the formation of the bright spot. Since the unfocused reflective flare (yellow line, e.g., (a)) can be effectively alleviated by AR coating. In this work, we mainly focus on removing the focused reflective flare (red line, e.g., (b)) from night images. + +moral, which can easily apply standard neural networks designed for image restoration. Specifically, to demonstrate the effectiveness of our dataset and pipeline, we train two CNN-based networks [4, 24] and two Transformer-based networks [21, 23] as baselines. These models trained on our dataset effectively suppress the reflective flare and generalize well to real-world scenarios. The contributions of our work are as follows: + +- We propose a new optical center symmetry prior that uncovers crucial properties of reflective flares on smartphones' cameras, including their locations of occurrence and patterns. +- Based on this prior, we explore a new method for constructing a reflective flare removal dataset called BracketFlare. This dataset includes various types of light sources and diverse indoor and outdoor scenes, providing a solid foundation for this task. +- We design an end-to-end reflective flare removal pipeline that incorporates the optical center symmetry prior to both training data generation and network structures. Extensive experiments demonstrate the essential role of our dataset and prior in reflective flare removal, which effectively generalizes to different night scenes + +# 2. Related Work + +Scattering flare removal. Most flare removal methods mainly focus on removing the scattering flares and glare effects. Based on the glow smooth prior, unsupervised methods [8, 16] are proposed to separate the glare effect from the flare-corrupted images. For supervised flare removal, Wu et al. [22] and Dai et al. [5] have respectively proposed two flare image datasets. These flare images can be added to flare-free images to synthesize the flare-corrupted and flare-free image pairs, making training an end-to-end flare removal network possible. Due to the difficulty in collecting real paired data, Qiao et al. [15] designed a pipeline to train the flare removal network with unpaired data following the idea of Cycle-GAN [27]. Besides, artifacts from underdisplay cameras [6] and nighttime haze [11] share similar attributes with the scattering flares. Thus, the related methods can also alleviate scattering flares artifacts. + +Reflective flare removal. Traditional reflective flare removal is typically accomplished through feature detection and image inpainting. Flare detection in these approaches [1,3,13,19] assumes that reflective flares are bright spots with specific shapes. Such an assumption is fragile given nighttime situations with diverse light sources. Since reflective flare removal can be viewed as a specific kind of reflection removal [26], one can adopt learning-based methods. However, existing datasets for this task, such as those proposed by Wu et al. [22], Dai et al. [5], and Qiao et al. [15], are not specifically designed for reflective flare removal and lack diversity in terms of light sources. These limitations hinder deep learning-based methods from effectively addressing real-world reflective flares, and may even lead to the removal of other tiny bright spots such as street lights. + +# 3. Optical Center Symmetry Prior + +# 3.1. Definition for the prior + +When pointing the camera at a light source, strong light can sometimes create lens flares in photos. In well-designed lens systems with AR coatings, reflective flares appear as bright spots on the photo, rather than generating a light fog. Through our investigation of most smartphones' main cameras, we discover that the bright spots of reflective flares and light sources are always symmetrical around the lens's optical center, as shown in Fig. 3. By aligning the light source and optical center, we can easily locate the reflective flares, eliminating the need to search the entire image for possible occurrences. Besides, since the main bright spots are caused by focused reflective flares, they share the same attributes as the light source. For example, the reflective flares may display clear details for each small LED lamp bead, as shown in Fig. 3(b) and Fig. 3(c). + +![](images/c961b2927515414107657dd5f1d0de02cf21adda64988ff3431b1e591b8197ab.jpg) + +![](images/d3f2083f0b815bbac4403daedcdb782dfc11fa7a6eb0918022bb53920cd5c358.jpg) +(b) iPhone 13 Pro + +![](images/a704a6064c451fcb9230ee0b4e7ab48d5d335412ceb79c3aa832e10bc03c07b5.jpg) +(a) Samsung s7 edge +(c) Xiaomi Mi 8 +Figure 3. Strong light source may project a bright reflective flare on the photo. The reflective flares and corresponding light sources are symmetric around the optical center for most smartphone cameras. The yellow cross represents the optical center. + +![](images/3fcc450a33213d3edb3bf42fbee4882cfb11adef28126d4d57e48aefcd301ede.jpg) +(d) Huawei Mate30 + +# 3.2. Physical formulation + +To illustrate the formation of optical center symmetry prior, we suppose that an optical ray $\mathbf{r}$ is emitted from the light source, as shown in Fig. 4. This ray can be described as $\mathbf{r} = [h\theta ]^{\top}$ . The states of this ray after each lens surface can be calculated by multiplying transfer matrices [10, 14]. We define the translation matrix at layer $i$ as $\mathbf{T_i}$ , and the refraction matrix at spherical surface with radius $R_{i}$ and refractive index $n_i,n_{i + 1}$ as $\mathbf{F_i}$ . For the reflection at the ith layer with radius $R_{i}$ , the reflection matrix can be stated as $\mathbf{R_i}$ . These matrices can be stated as: + +$$ +\mathbf {T} _ {\mathrm {i}} = \left( \begin{array}{c c} 1 & d _ {i} \\ 0 & 1 \end{array} \right), \mathbf {R} _ {\mathrm {i}} = \left( \begin{array}{c c} 1 & 0 \\ \frac {2}{R _ {i}} & 0 \end{array} \right), \mathbf {F} _ {\mathrm {i}} = \left( \begin{array}{c c} 1 & 0 \\ \frac {n _ {i + 1} - n _ {i}}{n _ {i + 1} R _ {i}} & \frac {n _ {i}}{n _ {i + 1}} \end{array} \right) \tag {1} +$$ + +In Fig. 4, $h_1$ represents the position of a real image that focuses on the sensor through the lens' internal reflection. We define the ray's forward propagation matrix as $\mathbf{A}$ and $\mathbf{C}$ and the reflection process as $\mathbf{B}$ . They are defined as: + +$$ +\begin{array}{l} A = \mathbf {T} _ {4} \mathbf {F} _ {3} \mathbf {T} _ {3} \mathbf {F} _ {2} \mathbf {T} _ {2} \mathbf {F} _ {1} \mathbf {T} _ {1} \mathbf {F} _ {0} \mathbf {T} _ {0} \\ B = \mathbf {T} _ {4} \mathbf {F} _ {3} \mathbf {T} _ {3} \mathbf {R} _ {2} ^ {- 1} \mathbf {T} _ {3} \mathbf {F} _ {3} ^ {- 1} \mathbf {T} _ {4} \mathbf {R} _ {4} \tag {2} \\ C = \mathrm {T} _ {8} \mathrm {F} _ {8} \mathrm {T} _ {7} \mathrm {F} _ {7} \mathrm {T} _ {6} \mathrm {F} _ {5} \mathrm {T} _ {5} \mathrm {F} _ {4} \\ \end{array} +$$ + +With these matrices, the ray after reflection in the focal plane $\mathbf{r}_1$ and direct imaging $\mathbf{r}_0$ can be written as: + +$$ +\binom {h _ {1}} {\theta_ {1}} = \mathbf {C B A} \binom {h} {\theta}, \binom {h _ {0}} {\theta_ {0}} = \mathbf {C A} \binom {h} {\theta} \tag {3} +$$ + +For the focused image, $h_0$ and $h_1$ will not vary with the change of the angle of incidence $\theta$ . Thus, we can deduce from Eq. (3) that both $h_0$ and $h_1$ are linear with $h$ . This + +![](images/1776f2a732e8cc8a1852ed23a184dfb51ca8f571376ec0d63e53bb04ad5d9d4e.jpg) +Figure 4. Physical explanation of reflective flare formation. In this figure, $h$ and $\theta$ indicate the height and the angle of incidence of the light source. $d$ and $n$ denote the thickness and refractive index of each layer. $A$ , $B$ , and $C$ represent the transfer matrices for each part of the lens system. The red line represents the imaging after reflection and the grey line shows the path of direct imaging. This figure depicts the situation with reflection in the 3rd and 5th air-lens layer, in real-world situations, the reflections can happen at any place including the lens protector and CMOS layer. + +means that the line between the light source and the reflective flare will pass through the optical center and the ratio between the $h_0$ and $h_1$ is a constant. We define this ratio as $k$ with $h_1 = kh_0$ . Thus, we can calculate the position of the reflective flare, if this constant $k$ is known. Due to the lens design for most smartphone cameras, the value of $k$ is always -1. It leads to the optical center symmetry prior stated above. + +# 3.3. Experimental statistics + +To demonstrate the effectiveness of our proposed prior in real-world situations, we conducted experiments using the main cameras of three smartphones: the Huawei P40, iPhone 13 Pro, and ZTE Axon 20 5G. In dark scenes, we captured 20 photos using each camera and marked the positions of the light source and reflective flare relative to the center of the image. As shown in Fig. 5, we observed that for all three cameras, the position of the reflective flare is linearly related to the position of the light source, with a slope of 1. However, since the light source may often create overexposed regions, manually marking the position of the light source can still introduce some errors. Considering this effect, we conclude that the results from all three phones align with our proposed prior. + +# 4. Data Collection + +Although the optical center prior can guide us in finding the reflective flare, the lack of paired images still presents a challenge for training an effective model for flare re + +![](images/daffbbce3c9429e95719b8fc8c4f94d983015c6b41392cc6eb5d0a22d4e86f3f.jpg) +(a) Statistics of flare-light source relation + +![](images/f52a0286cd58d3af36954897bfcfa2a8a5b7e683db0b4a31953ce885676bfecb.jpg) +(b) flare-light source relation in x-axis + +![](images/0645d13a2265558b07e6c0aba4684dab33f58b27297d2f51ce2ccf6b4f3e65e9.jpg) +(c) flare-light source relation in y-axis +Figure 5. Experimental statistics of the optical center symmetry prior. We capture images with reflective flares in a dark environment and mark the position of the light source and reflective flares. $x_{f}$ and $x_{l}$ represent the x-coordinate of flare and light source and so as $y_{f}$ and $y_{l}$ . The origin is set to be the center of the image. The relationship between the flare and the light source's absolute values of coordinates are plotted above. From figures (b) and (c), the coordinate relationships are linear and the slope is 1, which confirms our prior. + +moral. As reflective flares are caused by the lens system, it is physically impossible to capture real flare-free and flare-corrupted image pairs. Moreover, given the huge variety of nighttime lights, it would be tedious to synthesize reflective flares for each type of light source. + +To address the challenge of generating paired images for reflective flare removal, we propose a semi-synthetic dataset, named BracketFlare. Since reflective flares always have the same pattern as the brightest region of the light source, we reduce the exposure time to capture the patterns of the light source's brightest region, which are used to synthesize the reflective flares. Specifically, we mount a professional camera (Sony A7 III) on a tripod and use continuous bracketing to capture five photos of the same static scene with a step of $3\mathrm{EV}$ as a group. We then select the normally exposed image and the lowest exposure image from each group to compose the pair. To increase the diversity and realism of our dataset, we choose scenes with different light sources that are bright enough to leave reflective flares. Finally, we collect a dataset consisting of 440 pairs of 4K images as training data and 40 images as test data. A general overview of our dataset is depicted in Fig. 6, showing that the dataset contains diverse and commonly seen flare patterns. + +![](images/d4ffd77f063a66e1c48bf33a3ae2f56ed36d52e8e487288f5ebb82b7589f83b5.jpg) +LED + +![](images/dd72b55dad2b6b7e0fa9ec8279fbbed67df4aae6fb867c3fea0b3df49442fa2d.jpg) + +![](images/a51b80b24715b9f02cc5f513da312df89b089eed721e1e1e26c48e71abed7eb1.jpg) +Traffic light + +![](images/b2fb03a0035fe1f685e028675d529839498685b13b239678c0e99b4bd339bd1d.jpg) +Figure 6. Visualization and distribution of our dataset. It contains different types of light sources in diverse scenes. Based on light source types, we classify them into eight categories with different flare patterns. + +![](images/e635f1c97d4e46a6a4c50a6334c2d4513393973f87ef1a69d82250ad493dd8f5.jpg) +Square overhead light +Round overhead light + +![](images/1ed183a0c7625b83b3e23b985d49c22d38ae74f8f818fb04ac6ec1bce4c1b6bb.jpg) +Light tube +Light bulb + +# 5. Proposed Method + +# 5.1. Postprocessing our dataset + +To increase the diversity of our dataset, we generate paired flare-corrupted and flare-free images on the fly with augmentation. As depicted in Fig. 7(a), we first resize the 4K image to $800 \times 1200$ . To simulate the possible misalignment of the image's center and optical center, we introduce a slight translation with $t \sim U(0,15)$ pixels to the underexposed light source image. Then, we randomly crop the paired light source image and the normally exposed image with a size of $512 \times 512$ . Since our captured images are already gamma-encoded, we follow the settings of Wu et al. [22] and linearize them by applying a gamma correction, i.e., $I_g = I^\gamma$ with $\gamma$ randomly sampled from [1.8, 2.2]. All our operations are performed in the linear space. + +Synthetic reflective flare. As lenses reflect light of different wavelengths with varying reflectivity, reflective flares often appear in different colors. To replicate this, we use a random color factor to multiply the low-exposed image. To simulate an unfocused situation, we also add random blur. In accordance with the optical center symmetry prior, we rotate the light source image to synthesize the reflective flare. Finally, we obtain the synthetic reflective flare by multiplying the flare image with an opacity. + +Flare-free image. The ISO is fixed for all images in our continuous bracketing. Therefore, our normally exposed images always share a long exposure time and have little noise. To simulate real night scenes, we add Gaussian noise to the images. Finally, we randomly apply a small color offset to the flare-free image. These operations aid in simulating light fog and glare effects in real-world scenarios. + +Flare-corrupted image. To create flare-corrupted images, we add the synthetic reflective flare to the flare-free images. Then, we apply gamma correction to the images with the + +gamma value obtained by taking the reciprocal of the previously sampled gamma correction value $\gamma$ . The resulting gamma-encoded flare-corrupted image, flare-free image, and flare image are used to train the neural networks. + +# 5.2. Flare removal pipeline + +As the optical center symmetry prior is a global prior, it is difficult for CNN-based networks to learn it, and even Transformer-based networks struggle to learn such symmetry. Therefore, it is crucial to incorporate this prior into the network with specialized designs. + +To improve the network's ability to learn the optical center symmetry prior, as shown in Fig. 7(b), we apply a gamma correction of $I_{g} = I^{\gamma}$ to the input flare-corrupted image with $\gamma = 10$ . This operation extracts the nearly saturated areas of the light source in each channel. We then rotate this nearly saturated image by 180 degrees to obtain a prior image, which is concatenated with the flare-corrupted image to form a six-channel input. This rotated prior image provides an initial approximation of the reflective flare's pattern, which we expect the network to refine and enhance. The network's output is set to six channels, with three channels for the estimated flare-free image and three channels for the estimated flare. Our synthetic data provides supervision for all six channels, which facilitates network training. + +# 5.3. Loss function + +With the supervision of flare image $I_{F}$ and flare-free image $I_{B}$ , our final loss is the combination of background loss $\mathcal{L}_{bg}$ , flare loss $\mathcal{L}_{flare}$ , and reconstruction loss $\mathcal{L}_{rec}$ . It can be written as: + +$$ +\mathcal {L} = \mathcal {L} _ {\text {r e c}} + \mathcal {L} _ {\text {f l a r e}} + \mathcal {L} _ {\text {b g}}. \tag {4} +$$ + +We define our flare removal network as $f_{\theta}$ . It takes flare-corrupted image $I_{input}$ and its prior image $I_{prior}$ as inputs. + +![](images/944c57a82fcb893c5ae85d5767d22274eb557fc0506493a6ca6e315ab8e84746.jpg) +(a) Training Data Generation +(b) Network Learning +Figure 7. The pipeline of our data processing module and flare removal network. The flare image and a flare-free image from our paired dataset are resized and randomly cropped. After data augmentation, they are added to synthesize the corresponding flare-corrupted image. To encode the optical symmetry prior, the input flare-corrupted image is center flipped as the prior image. We also apply gamma correction to extract the light source region. Then, the prior image is concatenated with the flare-corrupted image as the input with 6 channels. Finally, the network predicts an output with 6 channels (a flare-free image with 3 channels and an estimated flare with 3 channels). + +The flare-free image and flare can be estimated by: + +$$ +\hat {I} _ {B}, \hat {I} _ {F} = f _ {\theta} \left(I _ {\text {i n p u t}}, I _ {\text {p r i o r}}\right). \tag {5} +$$ + +With the estimated flare and flare-free images, the $L_{1}$ reconstruction loss can be expressed as: + +$$ +\mathcal {L} _ {\text {r e c}} = \left| I _ {\text {i n p u t}} - \hat {I} _ {F} \oplus \hat {I} _ {B} \right|, \tag {6} +$$ + +where $\oplus$ denotes an element-wise addition operation in the linearized gamma-decoded domain. + +For the background loss, it consists of a default $\mathcal{L}_1$ , a region-aware masked $\mathcal{L}_{mask}$ term, and a perceptual term $\mathcal{L}_{vgg}$ with a pre-trained VGG-19 network [17]. Compared with previous works [5, 22, 26], the main difference is the $\mathcal{L}_{mask}$ . Since the regions of reflective flares are always small, a simple $L_1$ loss may ignore these regions, leading to the local optimum. To encourage the network to focus on restoring the flare-corrupted regions, we calculate the flare's mask $I_M$ ( $>0.2$ ) from the ground truth of flare image $I_F$ . This masked loss can be written as: + +$$ +\mathcal {L} _ {\text {m a s k}} = \left| I _ {M} * \left(\hat {I} _ {B} - I _ {B}\right) \right|. \tag {7} +$$ + +Overall, the background loss $\mathcal{L}_{bg}$ is defined as: + +$$ +\mathcal {L} _ {b g} = w _ {1} \mathcal {L} _ {1} + w _ {2} \mathcal {L} _ {v g g} + w _ {3} \mathcal {L} _ {\text {m a s k}}, \tag {8} +$$ + +where $w_{1}, w_{2}, w_{3}$ are respectively set to 0.5, 0.1, and 20.0 in our experiments. Flare loss $\mathcal{L}_{flare}$ has the identical form as the background loss. + +# 6. Experiments + +Evaluation on synthetic data. To evaluate our method, we collect 40 paired data using the process mentioned in + +Table 1. Percentage of users favoring our results (our dataset + MPRNet†) vs. previous work. Our method outperforms existing methods on both indoor and outdoor images. + +
MethodOutdoorIndoor
Reflection removal [26]3.0%8.5%
Flare spot removal [1]4.5%0.0%
Flare removal (Wu et al.) [22]0.5%3.5%
Flare removal (Dai et al.) [5]1.5%2.5%
Ours90.5%85.5%
+ +Sec. 4 to generate a synthetic test dataset, in which ground truth images are available. We adopt PSNR, SSIM [20], and LPIPS [25] as our evaluation metrics. + +Evaluation on real data. To show our method's generalization ability, we also collect a dataset with the photos captured by different types of mobile phones, including iPhone 12 pro, iPhone 13 pro, iPhone 14 pro, Huawei mate 30, Huawei nova 5i Pro, Huawei P40, Samsung Galaxy s22, Samsung Galaxy Note9, ZTE Axon 20 5G and Oppo Reno4 SE 5G. For real cases, we cannot obtain the ground truth. Since reflective flares always locate in small regions, they are hard to be evaluated by existing non-reference image quality assessment. These metrics emphasize more on global image quality and are not specially designed for our task. Therefore, we provide more visual results and conduct a user study to evaluate the performance. + +# 6.1. Comparison with different networks + +To validate the performance of our pipeline, we train different CNN-based and Transformer-based networks with our data. We use the default network settings of MPR-Net [24], HINet [4], and Uformer [21] and reduce the feature channels of Restormer [23] from 48 to 24 due to GPU memory constraint. In our experiments, these models are + +![](images/db4ff8be7c796b93fd3ee33705ab6a454fcc7a2fcc8dacfa2d2f737fae5c9c89.jpg) + +![](images/bc22b91a1000dfa5e315e05097c4ed22f7074250f4826bc8596267267b625c27.jpg) + +![](images/0e28a7387ad25a3188e0f00ef64b7976883921e8a2de6ab19a7ec91ee287bff3.jpg) +Input + +![](images/48ffa684854b80cdb98ae74a698c88ff432674d9fb5ebdc08e8337f3afd8ad7a.jpg) + +![](images/fb67e9933b70653afd1d83fcf53e40caca193a156f97dda8482eee51466f8589.jpg) + +![](images/3bc747c7f2ed860729a02cec77fb2ea8f777368fc3dfe005e630ded9732442d3.jpg) +Reflection removal + +![](images/9726bc3315bca57d0b557af8ae7ef3e278ce9c2ec6488003a825d2315be3e16d.jpg) + +![](images/b8e74dc2197d334dd06372da48bca80eca1ad9daf9c91c24aa62b0860e72bcf5.jpg) + +![](images/23a4606430801aba1811b766defc720f4d0cfe5aacb8bf09d3953913150a56bf.jpg) +Flare spot removal + +![](images/8dceb54be224d6e022f18af3acaf48c64cb605ab720e9d6a6b10ac193112287a.jpg) + +![](images/f8d85319bd51a4fc6bec357cfaa83957e1e1dabd92b0d49ea1e8f58d561908fe.jpg) + +![](images/aa92a277056b15d31558666ee5a13aebed60f18d25571335288f8d68157f78ad.jpg) +Wuet al. + +![](images/3e57a9445697106823a328fa71b26f3fcbec8015ddc63e993a976cbdc6cc907a.jpg) + +![](images/bef219f9c204b4842482eac749526d905c5b43bd1e2910b1b2d9cca28882669a.jpg) + +![](images/2f77e8e26b73e42d00f65cf69672053f645be9217313011745a8d08188f41a0a.jpg) +Dai et al. + +![](images/6d52eaf9c87e6835d71acbc3c56c9071d27f2041706dc360668e891a7191e1e2.jpg) + +![](images/c40034e529549cb7a8662358b9d3027e031b9aa817278645b37dd66042d81011.jpg) + +![](images/0d4dc6c70a282c5d96d04a6178755824f8984eb779e1443017610483f6412173.jpg) +Ours + +![](images/5c103a743c10fdfd45e41cf722277a2849943a8aa1234d06eb2f5957102d60f2.jpg) +Figure 8. Visual comparison of different flare removal and reflection removal methods on real-world nighttime flare images. + +![](images/eb461777130ef69b9236661f07717fe9f5641feda42819453bcd47a5bc06cc80.jpg) +Input +Figure 9. Visual comparison of different flare removal and reflection removal methods on synthetic nighttime flare images. + +![](images/001f546420f7ac9066da37a053b129c6261a0974128a3560d7eb563f79b64caf.jpg) + +![](images/e80bf37784cbe13dd014e5671dfd7d6b11628fa2aa90fa420518e7095da2e6f4.jpg) +Reflection removal + +![](images/002c0ca510c5b7f08626df2b9f24f972d15c42fa0c8e41ef910ca8cb17e55993.jpg) + +![](images/2333412be9c73797574131e92a19711fb6975819bee7f7b5f2cbb6afa0633335.jpg) +Flare spot removal + +![](images/bcf4a3be8cda5af200cb84d6ce37360377f7611326afb3a65e7ad0d52eb3a370.jpg) + +![](images/42051f6c40f73ce84a6d2e414a9374a777a628d30fb0f39fa21aafb7a7e63770.jpg) +Wu et al. + +![](images/64448d4dcb9a8363b02eb8bbb01bb05ff2c634ca7a0e18fe5ff8206689265bbc.jpg) + +![](images/dc66efa5ba0e54a7b9d8ed6be7d8a48b367a5e580dc258a5407b3098e500e28f.jpg) +Dai et al. + +![](images/02197a3b7e9d996222ca1bf773cd4d8819f2d1725d965881acffadee4ebb446f.jpg) + +![](images/23e47ddc750bc19b22df9a39b5fde190f01e91f312ca9c9ac66d8ea45a26ee81.jpg) +Ours + +![](images/b3768e6c35fca6693740ce3733b1634ae6277b7d476ed831dea1a54412fd90ec.jpg) + +![](images/84f091058fd93a70875f2f2fbd32e65868afd22ff8364c01c59e1e62883795a2.jpg) +Ground truth + +trained on an NVIDIA Geforce RTX3090 with a batch size of 2. We use Adam optimizer [9] with a learning rate of 1e-4 and train a model for 500k iterations. Besides, we adopt a MultiStepLR with a milestone of 200k and set the gamma to 0.5. These models are evaluated using the synthetic test dataset as shown in Table 4. From experiments, we find that $\mathrm{MRPNet}^{\dagger}$ [24] performs better than other networks. Thus, we treat $\mathrm{MPRNet}^{\dagger}$ [24] as our baseline method and conduct the following ablation study. + +# 6.2. Comparison with previous works + +We compare our method with different flare removal datasets [5, 22] and reflection removal dataset [26] on both real and synthetic data. We use the pre-trained models of Dai et al. [5] and Zhang et al. [26]. Since Wu et al. [22] do not release their model, we retrain their network using their provided data and code. Besides, we also compare our method with the traditional flare spot removal method [1]. As shown in Table 2, since most of these methods are not specially designed for reflective flare removal, they cannot + +Table 2. Quantitative comparison on synthetic test data. + +
MethodPSNRSSIMLPIPSMasked PSNR
Input37.300.9900.02521.68
Reflection removal [26]22.020.8300.07422.88
Flare spot removal [1]38.290.9900.02422.53
Flare removal (Wu et al.) [22]26.130.8950.05521.89
Flare removal (Dai et al.) [5]27.490.9420.05021.85
Ours48.410.9940.00432.09
+ +accurately remove the reflective flare while retaining image details simultaneously. Although Asha et al. [1] can detect the bright spots, as shown in Fig. 8, it may also remove the light source which is quite common at night. + +User Study. To evaluate these methods in real-world situations, we also conduct a user study with 20 participants with 20 images randomly selected from our real data in Table 1. + +# 6.3. Ablation study + +Loss function. We conduct an ablation study to prove the necessity of each loss function used in our baseline. In our experiments, we train the model without adding $\mathcal{L}_{\text {flare }}$ , + +![](images/6da115c232ba0dd5a5583753908fc58d30228cd49e6f4ecc78904616742b6bf7.jpg) +Input + +![](images/c313f5f2cbae4b35bb8264909bc35cf7f3cd15657bfd1e7981d3c0519e323514.jpg) +Output +(a) iPhone 14 Pro +Figure 10. Visual comparison of flare removal on real-world nighttime flare images that are captured by different smartphones. + +![](images/567933723a794726eedc088fc74a3ce60c96ae3b31f346b59cb140154b6f75c4.jpg) + +![](images/83949e84d2e25fd226a074c25b0bc9c52674adcaec4c411e702138431b4674bd.jpg) +(b) iPhone 13 Pro + +![](images/73d94687b622709963faaf5204d8f06d1bd7de4bbad8617103a4dee80f0fb428.jpg) + +![](images/91a35f8607fde00c87901c6623428def7285e5b16b78925bd304ace3c1089ff2.jpg) +(c) ZTE Axon 20 5G + +![](images/30b04c581f92f4b186b9d63e840c9919965a29ddfe621646f93c8f427cf16ff1.jpg) + +![](images/a1f0614f23ad77ef1efde7eed260fca04cedade4d55b1bf2ca302a4fb2ddb2b1.jpg) +(d) Samsung Galaxy S22 + +![](images/908f312e9fa316ab94f27ea10306713f837cb992bb904e1236a89061354a733b.jpg) + +![](images/11aefdabac495aa6b62bf1e9bd24dfb6144975bb9473fa032e5b30c120eecb74.jpg) +(e) Huawei Mate 30 + +![](images/9e81efc2c6ce08bba636e382b96b40a229eadd92f350ec00531a9b7d88f796c6.jpg) + +![](images/ada74c8871fef56c6b26ff81dd53c6241b3b53ba0e6c0a707477d1fb94a14211.jpg) +(f) Huawei nova 5i Pro + +Table 3. Ablation study for different loss functions. + +
LossPSNRSSIMLPIPSMasked PSNR
w/o Lflare47.360.9930.00430.63
w/o Lrec48.180.9940.00331.33
w/o Lvgg48.050.9940.00431.39
w/o Lmask48.010.9930.00331.03
Full48.410.9940.00432.09
+ +![](images/d1a31061421c916d5a577f52ef6c73dc5380d1e55ff2fbef1291ffa43b92d2c8.jpg) + +![](images/9f6e83b02c0987ef038de89e275634cb4e7f54fe40cef10f3db875a56c8ef233.jpg) +Input +Figure 11. Visual comparison with the network's performance with and w/o optical center symmetry prior. + +![](images/5d4cb1c4d224889639b106a9f480c83f4b034bbfd5dd633babfd46e377104522.jpg) + +![](images/71e6c233b8f3e23d10eb766fcc3e24115a711e96484507949edc8265fb27342f.jpg) +w/o prior + +![](images/4b6147282e274c24cef08102bad81052a62352842084aec46989f5ce7263a4a0.jpg) + +![](images/da016b0a5fb26f9b2d97e0405b93378898ffdf710a4fcb4ef0f6f48a7b705734.jpg) +with prior + +$\mathcal{L}_{rec}$ , $\mathcal{L}_{vgg}$ or $\mathcal{L}_{mask}$ . We report the results in Table 3. The ablation study of $\mathcal{L}_{flare}$ proves the significance of bringing supervision for both flare and background images. Besides, when the $L_{mask}$ is removed, it degrades the reconstruction results in the flare-corrupted regions, which verifies the importance of focusing on flare regions. + +Prior. To illustrate the performance of introducing the optical symmetry prior by concatenating a prior image, we train these baseline restoration networks with the original input. Since HINet [4] and MPRNet [24] cannot support input and output with different channels, we concatenate a black image with no information and keep other settings the same. The quantitative results are shown in Table 4, and visual comparisons on real-world images are presented in Fig. + +Table 4. Quantitative comparison for different network structures on synthetic test data. ‘†’ indicates that the network is equipped with the proposed prior. + +
NetworkPSNRSSIMLPIPSMasked PSNR
MPRNet [24]43.600.9920.01127.64
Uformer [21]42.570.9890.00927.77
HINet [4]43.780.9910.01227.73
Restormer [23]42.760.9910.01227.27
MPRNet†48.410.9940.00432.09
Uformer†47.470.9910.00331.57
HINet†48.030.9940.00330.88
Restormer†48.110.9940.00431.07
+ +11. It shows that embedding this prior in the network can boost the performance of quantitative results considerably. In Fig. 11, we observe that this prior helps the network to locate and remove the reflective flare precisely. + +# 7. Conclusion + +We have presented a novel approach to tackle the challenge of removing reflective flares in nighttime photography. Our approach is inspired by the optical center symmetry prior, which suggests that reflective flares and their corresponding light sources are always symmetrical around the optical center of the lens system in smartphone cameras. Based on the observation that the patterns of reflective flares are always similar to the brightest region of the light source, we introduce a dataset named BracketFlare. We further propose a reflective flare removal pipeline that encodes the optical center symmetry prior into the network. Experimental results demonstrate that our proposed methods are effective in removing reflective flares in real-world scenarios and have the potential for wide-ranging applications. + +Acknowledgement: This study is supported under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s). + +# References + +[1] C S Asha, Sooraj Kumar Bhat, Deepa Nayak, and Chaithra Bhat. Auto removal of bright spot from images captured against flashing light source. In ICDC, 2019. 2, 3, 6, 7 +[2] Herbert Bay, Andreas Ess, Tinne Tuytelaars, and Luc Van Gool. Speeded-up robust features (surf). Computer Vision and Image Understanding, 110(3):346-359, 2008. 2 +[3] Floris Chabert. Automated lens flare removal, 2014. Technical Report, Department of Electrical Engineering, Stanford University. 2, 3 +[4] Liangyu Chen, Xin Lu, Jie Zhang, Xiaojie Chu, and Chengpeng Chen. Hinet: Half instance normalization network for image restoration. In CVPRW, 2021. 2, 6, 8 +[5] Yuekun Dai, Chongyi Li, Shangchen Zhou, Ruicheng Feng, and Chen Change Loy. Flare7k: A phenomenological nighttime flare removal dataset. In Thirty-sixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2022. 1, 2, 3, 6, 7 +[6] Ruicheng Feng, Chongyi Li, Huaijin Chen, Shuai Li, Chen Change Loy, and Jinwei Gu. Removing diffraction image artifacts in under-display camera via dynamic skip connection network. In CVPR, 2021. 3 +[7] Matthias Hullin, Elmar Eisemann, Hans-Peter Seidel, and Sungkil Lee. Physically-based real-time lens flare rendering. In ACM SIGGRAPH, 2011. 1 +[8] Yeying Jin, Wenhan Yang, and Robby T Tan. Unsupervised night image enhancement: When layer decomposition meets light-effects suppression. In ECCV. Springer, 2022. 3 +[9] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 7 +[10] Sungkil Lee and Elmar Eisemann. Practical real-time lensflare rendering. Computer Graphics Forum, 32(4):1-6, 2013. 1, 3 +[11] Yu Li, Robby T. Tan, and Michael S. Brown. Nighttime haze removal with glow and multiple light colors. In ICCV, 2015. 3 +[12] David G Lowe. Distinctive image features from scale-invariant keypoints. International Journal of Computer Vision, 60(2):91-110, 2004. 2 +[13] Andreas Nussberger, Helmut Grabner, and Luc Van Gool. Robust aerial object tracking in images with lens flare. In ICRA, 2015. 3 +[14] Frank L Pedrotti, Leno M Pedrotti, and Leno S Pedrotti. Introduction to Optics. Cambridge University Press, 2017. 3 +[15] Xiaotian Qiao, Gerhard P. Hancke, and Rynson W. H. Lau. Light source guided single-image flare removal from unpaired data. In ICCV, 2021. 2, 3 +[16] Aashish Sharma and Robby T. Tan. Nighttime visibility enhancement by increasing the dynamic range and suppression of light effects. In CVPR, 2021. 3 +[17] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In International Conference on Learning Representations, 2014. 6 +[18] Qilin Sun, Ethan Tseng, Qiang Fu, Wolfgang Heidrich, and Felix Heide. Learning rank-1 diffractive optics for single-shot high dynamic range imaging. In CVPR, 2020. 2 + +[19] Patricia Vitoria and Coloma Ballester. Automatic flare spot artifact detection and removal in photographs. Journal of Mathematical Imaging and Vision, 61:515-533, 2019. 2, 3 +[20] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6 +[21] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restorationn. In CVPR, 2022. 2, 6, 8 +[22] Yicheng Wu, Qiurui He, Tianfan Xue, Rahul Garg, Jiawen Chen, Ashok Veeraraghavan, and Jonathan T Barron. How to train neural networks for flare removal. In ICCV, 2021. 1, 2, 3, 5, 6, 7 +[23] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In CVPR, 2022. 2, 6, 8 +[24] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Multi-stage progressive image restoration. In CVPR, 2021. 2, 6, 7, 8 +[25] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6 +[26] Xuaner Zhang, Ren Ng, and Qifeng Chen. Single image reflection separation with perceptual losses. In CVPR, 2018. 3, 6, 7 +[27] Junyan Zhu, Taesung Park, Phillip Isola, and Alexei A. Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In ICCV, 2017. 3 \ No newline at end of file diff --git a/2023/Nighttime Smartphone Reflective Flare Removal Using Optical Center Symmetry Prior/images.zip b/2023/Nighttime Smartphone Reflective Flare Removal Using Optical Center Symmetry Prior/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..60a38826f941247388ca6861526da80ae2333f46 --- /dev/null +++ b/2023/Nighttime Smartphone Reflective Flare Removal Using Optical Center Symmetry Prior/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dd5b3ad64be3b0b0f6f2e0faeb95c917275e843755c8c0dd2fb60ac8accc031a +size 798742 diff --git a/2023/Nighttime Smartphone Reflective Flare Removal Using Optical Center Symmetry Prior/layout.json b/2023/Nighttime Smartphone Reflective Flare Removal Using Optical Center Symmetry Prior/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..82ac7f17b21d2d4a20ae44cefea50fefd0fe2b61 --- /dev/null +++ b/2023/Nighttime Smartphone Reflective Flare Removal Using Optical Center Symmetry Prior/layout.json @@ -0,0 +1,11320 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 146, + 103, + 449, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 103, + 449, + 140 + ], + "spans": [ + { + "bbox": [ + 146, + 103, + 449, + 140 + ], + "type": "text", + "content": "Nighttime Smartphone Reflective Flare Removal Using Optical Center Symmetry Prior" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 161, + 489, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 161, + 489, + 190 + ], + "spans": [ + { + "bbox": [ + 104, + 161, + 489, + 190 + ], + "type": "text", + "content": "Yuekun Dai Yihang Luo Shangchen Zhou Chongyi Li* Chen Change Loy S-Lab, Nanyang Technological University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 142, + 191, + 451, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 191, + 451, + 204 + ], + "spans": [ + { + "bbox": [ + 142, + 191, + 451, + 204 + ], + "type": "text", + "content": "{YDAI005, c200211, s200094, chongyi.li, ccloy}@ntu.edu.sg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 172, + 205, + 417, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 172, + 205, + 417, + 217 + ], + "spans": [ + { + "bbox": [ + 172, + 205, + 417, + 217 + ], + "type": "text", + "content": "https://ykdai.github.io/projects/BracketFlare" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 228, + 192, + 240 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 228, + 192, + 240 + ], + "spans": [ + { + "bbox": [ + 143, + 228, + 192, + 240 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 254, + 290, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 254, + 290, + 540 + ], + "spans": [ + { + "bbox": [ + 47, + 254, + 290, + 540 + ], + "type": "text", + "content": "Reflective flare is a phenomenon that occurs when light reflects inside lenses, causing bright spots or a \"ghosting effect\" in photos, which can impact their quality. Eliminating reflective flare is highly desirable but challenging. Many existing methods rely on manually designed features to detect these bright spots, but they often fail to identify reflective flares created by various types of light and may even mistakenly remove the light sources in scenarios with multiple light sources. To address these challenges, we propose an optical center symmetry prior, which suggests that the reflective flare and light source are always symmetrical around the lens's optical center. This prior helps to locate the reflective flare's proposal region more accurately and can be applied to most smartphone cameras. Building on this prior, we create the first reflective flare removal dataset called BracketFlare, which contains diverse and realistic reflective flare patterns. We use continuous bracketing to capture the reflective flare pattern in the underexposed image and combine it with a normally exposed image to synthesize a pair of flare-corrupted and flare-free images. With the dataset, neural networks can be trained to remove the reflective flares effectively. Extensive experiments demonstrate the effectiveness of our method on both synthetic and real-world datasets." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 553, + 128, + 566 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 553, + 128, + 566 + ], + "spans": [ + { + "bbox": [ + 47, + 553, + 128, + 566 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 574, + 288, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 574, + 288, + 694 + ], + "spans": [ + { + "bbox": [ + 46, + 574, + 288, + 694 + ], + "type": "text", + "content": "Lens flare artifacts can occur when a strong light enters the camera's field of view, causing scattering and reflection within the lenses. Unlike radial-shaped scattering flares, reflective flares appear as groups of bright spots [5, 7] or a \"ghosting effect\" like the examples in Fig. 1. With multiple reflections, the light source creates multiple projections at various positions behind the lens. If the projection is far from the focal plane, the defocused effect can brighten the whole sensor and degrade the photo's quality as shown in Fig. 2. In near-focus situations, images of the light source" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 310, + 230, + 394, + 303 + ], + "blocks": [ + { + "bbox": [ + 310, + 230, + 394, + 303 + ], + "lines": [ + { + "bbox": [ + 310, + 230, + 394, + 303 + ], + "spans": [ + { + "bbox": [ + 310, + 230, + 394, + 303 + ], + "type": "image", + "image_path": "6550e8c0093d203935d90d2312c49717f38ee94b2f8313da514f0a03d6d41531.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 309, + 304, + 393, + 376 + ], + "blocks": [ + { + "bbox": [ + 309, + 304, + 393, + 376 + ], + "lines": [ + { + "bbox": [ + 309, + 304, + 393, + 376 + ], + "spans": [ + { + "bbox": [ + 309, + 304, + 393, + 376 + ], + "type": "image", + "image_path": "6a2d7d55295dd1bda06f1347e5622d3139648b025bc507ac815583c9e9b3b57d.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 310, + 377, + 393, + 449 + ], + "blocks": [ + { + "bbox": [ + 310, + 377, + 393, + 449 + ], + "lines": [ + { + "bbox": [ + 310, + 377, + 393, + 449 + ], + "spans": [ + { + "bbox": [ + 310, + 377, + 393, + 449 + ], + "type": "image", + "image_path": "db046f38b982a1a71dac44a5a8b723976beb85ad2768ba40d05efe3fb1064804.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 331, + 450, + 382, + 459 + ], + "lines": [ + { + "bbox": [ + 331, + 450, + 382, + 459 + ], + "spans": [ + { + "bbox": [ + 331, + 450, + 382, + 459 + ], + "type": "text", + "content": "(a) iPhone 13 Pro" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 394, + 230, + 467, + 303 + ], + "blocks": [ + { + "bbox": [ + 394, + 230, + 467, + 303 + ], + "lines": [ + { + "bbox": [ + 394, + 230, + 467, + 303 + ], + "spans": [ + { + "bbox": [ + 394, + 230, + 467, + 303 + ], + "type": "image", + "image_path": "aa72b59fb4ae6537452c603ca696e6268b35c3cf36362e84a3deec27e97e23b6.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 394, + 304, + 467, + 376 + ], + "blocks": [ + { + "bbox": [ + 394, + 304, + 467, + 376 + ], + "lines": [ + { + "bbox": [ + 394, + 304, + 467, + 376 + ], + "spans": [ + { + "bbox": [ + 394, + 304, + 467, + 376 + ], + "type": "image", + "image_path": "3f957143ddac45f578f5cee0a8768434c0ea7afb57bdf90a514c705ffc6997db.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 394, + 377, + 467, + 449 + ], + "blocks": [ + { + "bbox": [ + 394, + 377, + 467, + 449 + ], + "lines": [ + { + "bbox": [ + 394, + 377, + 467, + 449 + ], + "spans": [ + { + "bbox": [ + 394, + 377, + 467, + 449 + ], + "type": "image", + "image_path": "c7cbf74198cd8f53741fb18de3446471d90cfba1288b0541eaeafbce1116a92d.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 406, + 450, + 452, + 459 + ], + "lines": [ + { + "bbox": [ + 406, + 450, + 452, + 459 + ], + "spans": [ + { + "bbox": [ + 406, + 450, + 452, + 459 + ], + "type": "text", + "content": "(b) Huawei P40" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 468, + 230, + 541, + 303 + ], + "blocks": [ + { + "bbox": [ + 468, + 230, + 541, + 303 + ], + "lines": [ + { + "bbox": [ + 468, + 230, + 541, + 303 + ], + "spans": [ + { + "bbox": [ + 468, + 230, + 541, + 303 + ], + "type": "image", + "image_path": "4696c233bba023fd7d15fd38e4db745a7c2491186093e60284533358cc7363f8.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 468, + 304, + 541, + 376 + ], + "blocks": [ + { + "bbox": [ + 468, + 304, + 541, + 376 + ], + "lines": [ + { + "bbox": [ + 468, + 304, + 541, + 376 + ], + "spans": [ + { + "bbox": [ + 468, + 304, + 541, + 376 + ], + "type": "image", + "image_path": "170f32acdf35ba7a862153536978978c671f1632e48a74b74f0c0c725306ab66.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 468, + 377, + 541, + 449 + ], + "blocks": [ + { + "bbox": [ + 468, + 377, + 541, + 449 + ], + "lines": [ + { + "bbox": [ + 468, + 377, + 541, + 449 + ], + "spans": [ + { + "bbox": [ + 468, + 377, + 541, + 449 + ], + "type": "image", + "image_path": "ebeaf9b3ad5a75adf7db2d1c95926f6e12a918fdca2b156ad57d4dab3b3376f8.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 474, + 449, + 534, + 459 + ], + "lines": [ + { + "bbox": [ + 474, + 449, + 534, + 459 + ], + "spans": [ + { + "bbox": [ + 474, + 449, + 534, + 459 + ], + "type": "text", + "content": "(c) Sony ILCE-6400" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 465, + 545, + 521 + ], + "lines": [ + { + "bbox": [ + 305, + 465, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 305, + 465, + 545, + 521 + ], + "type": "text", + "content": "Figure 1. Reflective flare removal with our proposed dataset and prior. The network trained on our dataset can effectively separate the reflective flare and flare-free images from a flare-corrupted image. Our method can generalize well to different types of smartphones and even cameras with specific types of lenses." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 530, + 545, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 530, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 305, + 530, + 545, + 555 + ], + "type": "text", + "content": "can result in a line of unfocused, bright blobs on the captured photo." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 558, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 546, + 713 + ], + "type": "text", + "content": "Applying anti-reflective (AR) coating is a common technique for reducing internal reflections [22]. The coating is effective for unfocused spots with low luminance, but cannot mitigate a bright spot with high luminance at the focal plane. Even with a well-designed lens system, a bright reflective spot may remain in the photo, known as the ghosting effect [10]. Reflective flare is common on all types of smartphone lenses, and certain types of smartphones (e.g., the iPhone series) can produce a large area of inverted reflective flare from neon lights or large, luminous standing signs, as shown in Fig. 1. The phenomenon becomes more pronounced when capturing a photo at night with multiple lights in the scene. Numerous bright spots and inverted re" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20783" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 108 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 108 + ], + "type": "text", + "content": "flections manifest and propagate in different directions with the lens orientation. Therefore, reflective flare removal algorithms are highly desirable." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 110, + 289, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 110, + 289, + 434 + ], + "spans": [ + { + "bbox": [ + 47, + 110, + 289, + 434 + ], + "type": "text", + "content": "Removing reflective flare is a difficult task due to the vast diversity of reflective flares that can be caused by different types of light sources and lenses. Reflective flares often appear as bright spots that can resemble streetlights from a distance, making it challenging to differentiate between reflective flares and other light sources. Current methods for removing flare spots mainly rely on feature detection techniques like Speeded-Up Robust Features [2] (SURF) or Scale-Invariant Feature Transform [12] (SIFT) to extract spot regions [1, 3, 19]. These methods, however, struggle to effectively detect blobs on complex backgrounds and only work for specific spot patterns. Moreover, these methods are ineffective in removing reflective flares caused by lamp tubes and LED matrices. Many learning-based methods and datasets for reflective flare removal have been proposed [5, 15, 18, 22]. Although these datasets contain subsets of reflective flares, they are not sufficient for good generalization. For instance, the reflective flare subset from Wu et al. [22] is captured using the same lens and light source, making it difficult to generalize to different flares in real-world scenarios. Dai et al. [5] propose a synthetic dataset with 10 types of reflective flares, but training on this dataset tends to mislead a network to erroneously remove other tiny bright spots like street lights and bright windows. As a result, it remains challenging for existing algorithms to accurately detect and remove various types of reflective flares while preserving details in other regions." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 436, + 289, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 436, + 289, + 687 + ], + "spans": [ + { + "bbox": [ + 47, + 436, + 289, + 687 + ], + "type": "text", + "content": "To overcome the challenges of detecting and removing reflective flares, we propose a novel optical center symmetry prior, which suggests that the main spot of the reflective flare and the light source are always symmetrical around the lens's optical center in captured images. This prior applies to most smartphone cameras and some professional cameras. With this prior, we can easily detect the possible locations of reflective flares. We also found that the flare spots always have the same patterns as the brightest regions of the light source. Therefore, we can synthesize reflective flares by using the light source's patterns at low exposure. Based on these characteristics of reflective flares, we develop the first reflective flare removal dataset, named BracketFlare. Specifically, we use continuous bracketing to capture the light source's pattern on an underexposed image and select the normally exposed image to create the light source and the natural image. Next, we rotate the light source pattern 180 degrees around the optical center, apply blur and color augmentation, and use it to create synthetic reflective flares. We can add the reflective flare to the normally exposed image to produce paired flare-corrupted and flare-free images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "content": "Thanks to the optical center symmetry prior, we further propose an end-to-end pipeline for reflective flare re" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 322, + 72, + 428, + 177 + ], + "blocks": [ + { + "bbox": [ + 322, + 72, + 428, + 177 + ], + "lines": [ + { + "bbox": [ + 322, + 72, + 428, + 177 + ], + "spans": [ + { + "bbox": [ + 322, + 72, + 428, + 177 + ], + "type": "image", + "image_path": "a5c94693cc237f600ed1a2d8c35288496b35992a6debe28279e859ef8a2afb20.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 430, + 72, + 535, + 177 + ], + "blocks": [ + { + "bbox": [ + 430, + 72, + 535, + 177 + ], + "lines": [ + { + "bbox": [ + 430, + 72, + 535, + 177 + ], + "spans": [ + { + "bbox": [ + 430, + 72, + 535, + 177 + ], + "type": "image", + "image_path": "3699cdd3c26ddf177d851d381fe7068b686880a3c840cac5358bc21d8b3e68fd.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 328, + 189, + 539, + 295 + ], + "blocks": [ + { + "bbox": [ + 340, + 178, + 408, + 190 + ], + "lines": [ + { + "bbox": [ + 340, + 178, + 408, + 190 + ], + "spans": [ + { + "bbox": [ + 340, + 178, + 408, + 190 + ], + "type": "text", + "content": "(a) w/o AR coating" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 446, + 178, + 520, + 189 + ], + "lines": [ + { + "bbox": [ + 446, + 178, + 520, + 189 + ], + "spans": [ + { + "bbox": [ + 446, + 178, + 520, + 189 + ], + "type": "text", + "content": "(b) With AR coating" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 328, + 189, + 539, + 295 + ], + "lines": [ + { + "bbox": [ + 328, + 189, + 539, + 295 + ], + "spans": [ + { + "bbox": [ + 328, + 189, + 539, + 295 + ], + "type": "image", + "image_path": "36baeab3108a00989db2c702a828df4ad8591c55495ee080ab0fbdfefc3dec07.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 353, + 296, + 509, + 306 + ], + "lines": [ + { + "bbox": [ + 353, + 296, + 509, + 306 + ], + "spans": [ + { + "bbox": [ + 353, + 296, + 509, + 306 + ], + "type": "text", + "content": "(c) Description of reflective flare formation" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 312, + 545, + 399 + ], + "lines": [ + { + "bbox": [ + 304, + 312, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 304, + 312, + 545, + 399 + ], + "type": "text", + "content": "Figure 2. Schematic diagram of reflective flare formation. Reflective flares are generated by multiple reflections inside the lens. In (c), yellow lines show the generation principle of the aperture-shaped reflective flare. Red lines represent the formation of the bright spot. Since the unfocused reflective flare (yellow line, e.g., (a)) can be effectively alleviated by AR coating. In this work, we mainly focus on removing the focused reflective flare (red line, e.g., (b)) from night images." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 405, + 545, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 405, + 545, + 499 + ], + "spans": [ + { + "bbox": [ + 304, + 405, + 545, + 499 + ], + "type": "text", + "content": "moral, which can easily apply standard neural networks designed for image restoration. Specifically, to demonstrate the effectiveness of our dataset and pipeline, we train two CNN-based networks [4, 24] and two Transformer-based networks [21, 23] as baselines. These models trained on our dataset effectively suppress the reflective flare and generalize well to real-world scenarios. The contributions of our work are as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 317, + 507, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 317, + 507, + 545, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 507, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 317, + 507, + 545, + 555 + ], + "type": "text", + "content": "- We propose a new optical center symmetry prior that uncovers crucial properties of reflective flares on smartphones' cameras, including their locations of occurrence and patterns." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 317, + 563, + 545, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 563, + 545, + 622 + ], + "spans": [ + { + "bbox": [ + 317, + 563, + 545, + 622 + ], + "type": "text", + "content": "- Based on this prior, we explore a new method for constructing a reflective flare removal dataset called BracketFlare. This dataset includes various types of light sources and diverse indoor and outdoor scenes, providing a solid foundation for this task." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 630, + 545, + 713 + ], + "type": "text", + "content": "- We design an end-to-end reflective flare removal pipeline that incorporates the optical center symmetry prior to both training data generation and network structures. Extensive experiments demonstrate the essential role of our dataset and prior in reflective flare removal, which effectively generalizes to different night scenes" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20784" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 71, + 134, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 71, + 134, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 134, + 83 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 91, + 289, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 91, + 289, + 284 + ], + "spans": [ + { + "bbox": [ + 46, + 91, + 289, + 284 + ], + "type": "text", + "content": "Scattering flare removal. Most flare removal methods mainly focus on removing the scattering flares and glare effects. Based on the glow smooth prior, unsupervised methods [8, 16] are proposed to separate the glare effect from the flare-corrupted images. For supervised flare removal, Wu et al. [22] and Dai et al. [5] have respectively proposed two flare image datasets. These flare images can be added to flare-free images to synthesize the flare-corrupted and flare-free image pairs, making training an end-to-end flare removal network possible. Due to the difficulty in collecting real paired data, Qiao et al. [15] designed a pipeline to train the flare removal network with unpaired data following the idea of Cycle-GAN [27]. Besides, artifacts from underdisplay cameras [6] and nighttime haze [11] share similar attributes with the scattering flares. Thus, the related methods can also alleviate scattering flares artifacts." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 289, + 289, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 289, + 289, + 483 + ], + "spans": [ + { + "bbox": [ + 46, + 289, + 289, + 483 + ], + "type": "text", + "content": "Reflective flare removal. Traditional reflective flare removal is typically accomplished through feature detection and image inpainting. Flare detection in these approaches [1,3,13,19] assumes that reflective flares are bright spots with specific shapes. Such an assumption is fragile given nighttime situations with diverse light sources. Since reflective flare removal can be viewed as a specific kind of reflection removal [26], one can adopt learning-based methods. However, existing datasets for this task, such as those proposed by Wu et al. [22], Dai et al. [5], and Qiao et al. [15], are not specifically designed for reflective flare removal and lack diversity in terms of light sources. These limitations hinder deep learning-based methods from effectively addressing real-world reflective flares, and may even lead to the removal of other tiny bright spots such as street lights." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 494, + 226, + 509 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 494, + 226, + 509 + ], + "spans": [ + { + "bbox": [ + 47, + 494, + 226, + 509 + ], + "type": "text", + "content": "3. Optical Center Symmetry Prior" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 514, + 178, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 514, + 178, + 528 + ], + "spans": [ + { + "bbox": [ + 47, + 514, + 178, + 528 + ], + "type": "text", + "content": "3.1. Definition for the prior" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "type": "text", + "content": "When pointing the camera at a light source, strong light can sometimes create lens flares in photos. In well-designed lens systems with AR coatings, reflective flares appear as bright spots on the photo, rather than generating a light fog. Through our investigation of most smartphones' main cameras, we discover that the bright spots of reflective flares and light sources are always symmetrical around the lens's optical center, as shown in Fig. 3. By aligning the light source and optical center, we can easily locate the reflective flares, eliminating the need to search the entire image for possible occurrences. Besides, since the main bright spots are caused by focused reflective flares, they share the same attributes as the light source. For example, the reflective flares may display clear details for each small LED lamp bead, as shown in Fig. 3(b) and Fig. 3(c)." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 328, + 73, + 425, + 170 + ], + "blocks": [ + { + "bbox": [ + 328, + 73, + 425, + 170 + ], + "lines": [ + { + "bbox": [ + 328, + 73, + 425, + 170 + ], + "spans": [ + { + "bbox": [ + 328, + 73, + 425, + 170 + ], + "type": "image", + "image_path": "c961b2927515414107657dd5f1d0de02cf21adda64988ff3431b1e591b8197ab.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 426, + 73, + 523, + 170 + ], + "blocks": [ + { + "bbox": [ + 426, + 73, + 523, + 170 + ], + "lines": [ + { + "bbox": [ + 426, + 73, + 523, + 170 + ], + "spans": [ + { + "bbox": [ + 426, + 73, + 523, + 170 + ], + "type": "image", + "image_path": "d3f2083f0b815bbac4403daedcdb782dfc11fa7a6eb0918022bb53920cd5c358.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 443, + 171, + 495, + 178 + ], + "lines": [ + { + "bbox": [ + 443, + 171, + 495, + 178 + ], + "spans": [ + { + "bbox": [ + 443, + 171, + 495, + 178 + ], + "type": "text", + "content": "(b) iPhone 13 Pro" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 328, + 179, + 425, + 277 + ], + "blocks": [ + { + "bbox": [ + 343, + 171, + 405, + 179 + ], + "lines": [ + { + "bbox": [ + 343, + 171, + 405, + 179 + ], + "spans": [ + { + "bbox": [ + 343, + 171, + 405, + 179 + ], + "type": "text", + "content": "(a) Samsung s7 edge" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 328, + 179, + 425, + 277 + ], + "lines": [ + { + "bbox": [ + 328, + 179, + 425, + 277 + ], + "spans": [ + { + "bbox": [ + 328, + 179, + 425, + 277 + ], + "type": "image", + "image_path": "a704a6064c451fcb9230ee0b4e7ab48d5d335412ceb79c3aa832e10bc03c07b5.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 355, + 277, + 404, + 285 + ], + "lines": [ + { + "bbox": [ + 355, + 277, + 404, + 285 + ], + "spans": [ + { + "bbox": [ + 355, + 277, + 404, + 285 + ], + "type": "text", + "content": "(c) Xiaomi Mi 8" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 294, + 545, + 338 + ], + "lines": [ + { + "bbox": [ + 304, + 294, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 304, + 294, + 545, + 338 + ], + "type": "text", + "content": "Figure 3. Strong light source may project a bright reflective flare on the photo. The reflective flares and corresponding light sources are symmetric around the optical center for most smartphone cameras. The yellow cross represents the optical center." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 426, + 180, + 523, + 277 + ], + "blocks": [ + { + "bbox": [ + 426, + 180, + 523, + 277 + ], + "lines": [ + { + "bbox": [ + 426, + 180, + 523, + 277 + ], + "spans": [ + { + "bbox": [ + 426, + 180, + 523, + 277 + ], + "type": "image", + "image_path": "3fcc450a33213d3edb3bf42fbee4882cfb11adef28126d4d57e48aefcd301ede.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 440, + 277, + 498, + 285 + ], + "lines": [ + { + "bbox": [ + 440, + 277, + 498, + 285 + ], + "spans": [ + { + "bbox": [ + 440, + 277, + 498, + 285 + ], + "type": "text", + "content": "(d) Huawei Mate30" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 347, + 425, + 360 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 347, + 425, + 360 + ], + "spans": [ + { + "bbox": [ + 306, + 347, + 425, + 360 + ], + "type": "text", + "content": "3.2. Physical formulation" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 363, + 545, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 363, + 545, + 467 + ], + "spans": [ + { + "bbox": [ + 304, + 363, + 545, + 467 + ], + "type": "text", + "content": "To illustrate the formation of optical center symmetry prior, we suppose that an optical ray " + }, + { + "bbox": [ + 304, + 363, + 545, + 467 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 304, + 363, + 545, + 467 + ], + "type": "text", + "content": " is emitted from the light source, as shown in Fig. 4. This ray can be described as " + }, + { + "bbox": [ + 304, + 363, + 545, + 467 + ], + "type": "inline_equation", + "content": "\\mathbf{r} = [h\\theta ]^{\\top}" + }, + { + "bbox": [ + 304, + 363, + 545, + 467 + ], + "type": "text", + "content": ". The states of this ray after each lens surface can be calculated by multiplying transfer matrices [10, 14]. We define the translation matrix at layer " + }, + { + "bbox": [ + 304, + 363, + 545, + 467 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 363, + 545, + 467 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 304, + 363, + 545, + 467 + ], + "type": "inline_equation", + "content": "\\mathbf{T_i}" + }, + { + "bbox": [ + 304, + 363, + 545, + 467 + ], + "type": "text", + "content": ", and the refraction matrix at spherical surface with radius " + }, + { + "bbox": [ + 304, + 363, + 545, + 467 + ], + "type": "inline_equation", + "content": "R_{i}" + }, + { + "bbox": [ + 304, + 363, + 545, + 467 + ], + "type": "text", + "content": " and refractive index " + }, + { + "bbox": [ + 304, + 363, + 545, + 467 + ], + "type": "inline_equation", + "content": "n_i,n_{i + 1}" + }, + { + "bbox": [ + 304, + 363, + 545, + 467 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 304, + 363, + 545, + 467 + ], + "type": "inline_equation", + "content": "\\mathbf{F_i}" + }, + { + "bbox": [ + 304, + 363, + 545, + 467 + ], + "type": "text", + "content": ". For the reflection at the ith layer with radius " + }, + { + "bbox": [ + 304, + 363, + 545, + 467 + ], + "type": "inline_equation", + "content": "R_{i}" + }, + { + "bbox": [ + 304, + 363, + 545, + 467 + ], + "type": "text", + "content": ", the reflection matrix can be stated as " + }, + { + "bbox": [ + 304, + 363, + 545, + 467 + ], + "type": "inline_equation", + "content": "\\mathbf{R_i}" + }, + { + "bbox": [ + 304, + 363, + 545, + 467 + ], + "type": "text", + "content": ". These matrices can be stated as:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 474, + 545, + 498 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 474, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 314, + 474, + 545, + 498 + ], + "type": "interline_equation", + "content": "\\mathbf {T} _ {\\mathrm {i}} = \\left( \\begin{array}{c c} 1 & d _ {i} \\\\ 0 & 1 \\end{array} \\right), \\mathbf {R} _ {\\mathrm {i}} = \\left( \\begin{array}{c c} 1 & 0 \\\\ \\frac {2}{R _ {i}} & 0 \\end{array} \\right), \\mathbf {F} _ {\\mathrm {i}} = \\left( \\begin{array}{c c} 1 & 0 \\\\ \\frac {n _ {i + 1} - n _ {i}}{n _ {i + 1} R _ {i}} & \\frac {n _ {i}}{n _ {i + 1}} \\end{array} \\right) \\tag {1}", + "image_path": "1fc5880942c7d7642920f59725995d87fbdfb0b532b582e3136af9759214054a.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 504, + 547, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 504, + 547, + 553 + ], + "spans": [ + { + "bbox": [ + 304, + 504, + 547, + 553 + ], + "type": "text", + "content": "In Fig. 4, " + }, + { + "bbox": [ + 304, + 504, + 547, + 553 + ], + "type": "inline_equation", + "content": "h_1" + }, + { + "bbox": [ + 304, + 504, + 547, + 553 + ], + "type": "text", + "content": " represents the position of a real image that focuses on the sensor through the lens' internal reflection. We define the ray's forward propagation matrix as " + }, + { + "bbox": [ + 304, + 504, + 547, + 553 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 304, + 504, + 547, + 553 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 504, + 547, + 553 + ], + "type": "inline_equation", + "content": "\\mathbf{C}" + }, + { + "bbox": [ + 304, + 504, + 547, + 553 + ], + "type": "text", + "content": " and the reflection process as " + }, + { + "bbox": [ + 304, + 504, + 547, + 553 + ], + "type": "inline_equation", + "content": "\\mathbf{B}" + }, + { + "bbox": [ + 304, + 504, + 547, + 553 + ], + "type": "text", + "content": ". They are defined as:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 352, + 560, + 545, + 604 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 352, + 560, + 545, + 604 + ], + "spans": [ + { + "bbox": [ + 352, + 560, + 545, + 604 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} A = \\mathbf {T} _ {4} \\mathbf {F} _ {3} \\mathbf {T} _ {3} \\mathbf {F} _ {2} \\mathbf {T} _ {2} \\mathbf {F} _ {1} \\mathbf {T} _ {1} \\mathbf {F} _ {0} \\mathbf {T} _ {0} \\\\ B = \\mathbf {T} _ {4} \\mathbf {F} _ {3} \\mathbf {T} _ {3} \\mathbf {R} _ {2} ^ {- 1} \\mathbf {T} _ {3} \\mathbf {F} _ {3} ^ {- 1} \\mathbf {T} _ {4} \\mathbf {R} _ {4} \\tag {2} \\\\ C = \\mathrm {T} _ {8} \\mathrm {F} _ {8} \\mathrm {T} _ {7} \\mathrm {F} _ {7} \\mathrm {T} _ {6} \\mathrm {F} _ {5} \\mathrm {T} _ {5} \\mathrm {F} _ {4} \\\\ \\end{array}", + "image_path": "d6efeebdc57d2f0a7473708f195661dd6906fa26e95277fcd12f188bafeec9db.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 612, + 545, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 612, + 545, + 637 + ], + "spans": [ + { + "bbox": [ + 304, + 612, + 545, + 637 + ], + "type": "text", + "content": "With these matrices, the ray after reflection in the focal plane " + }, + { + "bbox": [ + 304, + 612, + 545, + 637 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_1" + }, + { + "bbox": [ + 304, + 612, + 545, + 637 + ], + "type": "text", + "content": " and direct imaging " + }, + { + "bbox": [ + 304, + 612, + 545, + 637 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_0" + }, + { + "bbox": [ + 304, + 612, + 545, + 637 + ], + "type": "text", + "content": " can be written as:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 340, + 643, + 545, + 670 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 643, + 545, + 670 + ], + "spans": [ + { + "bbox": [ + 340, + 643, + 545, + 670 + ], + "type": "interline_equation", + "content": "\\binom {h _ {1}} {\\theta_ {1}} = \\mathbf {C B A} \\binom {h} {\\theta}, \\binom {h _ {0}} {\\theta_ {0}} = \\mathbf {C A} \\binom {h} {\\theta} \\tag {3}", + "image_path": "70d1714df96f767e3ce6f9bf03b22490b02626e97e560ded065c3a5a3d36ad49.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "For the focused image, " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "h_0" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "h_1" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " will not vary with the change of the angle of incidence " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": ". Thus, we can deduce from Eq. (3) that both " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "h_0" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "h_1" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " are linear with " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": ". This" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20785" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 73, + 282, + 240 + ], + "blocks": [ + { + "bbox": [ + 55, + 73, + 282, + 240 + ], + "lines": [ + { + "bbox": [ + 55, + 73, + 282, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 73, + 282, + 240 + ], + "type": "image", + "image_path": "1776f2a732e8cc8a1852ed23a184dfb51ca8f571376ec0d63e53bb04ad5d9d4e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 246, + 289, + 346 + ], + "lines": [ + { + "bbox": [ + 46, + 246, + 289, + 346 + ], + "spans": [ + { + "bbox": [ + 46, + 246, + 289, + 346 + ], + "type": "text", + "content": "Figure 4. Physical explanation of reflective flare formation. In this figure, " + }, + { + "bbox": [ + 46, + 246, + 289, + 346 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 46, + 246, + 289, + 346 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 246, + 289, + 346 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 246, + 289, + 346 + ], + "type": "text", + "content": " indicate the height and the angle of incidence of the light source. " + }, + { + "bbox": [ + 46, + 246, + 289, + 346 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 246, + 289, + 346 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 246, + 289, + 346 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 246, + 289, + 346 + ], + "type": "text", + "content": " denote the thickness and refractive index of each layer. " + }, + { + "bbox": [ + 46, + 246, + 289, + 346 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 46, + 246, + 289, + 346 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 246, + 289, + 346 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 46, + 246, + 289, + 346 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 246, + 289, + 346 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 246, + 289, + 346 + ], + "type": "text", + "content": " represent the transfer matrices for each part of the lens system. The red line represents the imaging after reflection and the grey line shows the path of direct imaging. This figure depicts the situation with reflection in the 3rd and 5th air-lens layer, in real-world situations, the reflections can happen at any place including the lens protector and CMOS layer." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 355, + 289, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 355, + 289, + 452 + ], + "spans": [ + { + "bbox": [ + 46, + 355, + 289, + 452 + ], + "type": "text", + "content": "means that the line between the light source and the reflective flare will pass through the optical center and the ratio between the " + }, + { + "bbox": [ + 46, + 355, + 289, + 452 + ], + "type": "inline_equation", + "content": "h_0" + }, + { + "bbox": [ + 46, + 355, + 289, + 452 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 355, + 289, + 452 + ], + "type": "inline_equation", + "content": "h_1" + }, + { + "bbox": [ + 46, + 355, + 289, + 452 + ], + "type": "text", + "content": " is a constant. We define this ratio as " + }, + { + "bbox": [ + 46, + 355, + 289, + 452 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 355, + 289, + 452 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 355, + 289, + 452 + ], + "type": "inline_equation", + "content": "h_1 = kh_0" + }, + { + "bbox": [ + 46, + 355, + 289, + 452 + ], + "type": "text", + "content": ". Thus, we can calculate the position of the reflective flare, if this constant " + }, + { + "bbox": [ + 46, + 355, + 289, + 452 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 355, + 289, + 452 + ], + "type": "text", + "content": " is known. Due to the lens design for most smartphone cameras, the value of " + }, + { + "bbox": [ + 46, + 355, + 289, + 452 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 355, + 289, + 452 + ], + "type": "text", + "content": " is always -1. It leads to the optical center symmetry prior stated above." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 459, + 178, + 472 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 459, + 178, + 472 + ], + "spans": [ + { + "bbox": [ + 47, + 459, + 178, + 472 + ], + "type": "text", + "content": "3.3. Experimental statistics" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 478, + 288, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 478, + 288, + 647 + ], + "spans": [ + { + "bbox": [ + 46, + 478, + 288, + 647 + ], + "type": "text", + "content": "To demonstrate the effectiveness of our proposed prior in real-world situations, we conducted experiments using the main cameras of three smartphones: the Huawei P40, iPhone 13 Pro, and ZTE Axon 20 5G. In dark scenes, we captured 20 photos using each camera and marked the positions of the light source and reflective flare relative to the center of the image. As shown in Fig. 5, we observed that for all three cameras, the position of the reflective flare is linearly related to the position of the light source, with a slope of 1. However, since the light source may often create overexposed regions, manually marking the position of the light source can still introduce some errors. Considering this effect, we conclude that the results from all three phones align with our proposed prior." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 657, + 142, + 669 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 657, + 142, + 669 + ], + "spans": [ + { + "bbox": [ + 47, + 657, + 142, + 669 + ], + "type": "text", + "content": "4. Data Collection" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "content": "Although the optical center prior can guide us in finding the reflective flare, the lack of paired images still presents a challenge for training an effective model for flare re" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 348, + 72, + 504, + 190 + ], + "blocks": [ + { + "bbox": [ + 348, + 72, + 504, + 190 + ], + "lines": [ + { + "bbox": [ + 348, + 72, + 504, + 190 + ], + "spans": [ + { + "bbox": [ + 348, + 72, + 504, + 190 + ], + "type": "image", + "image_path": "daffbbce3c9429e95719b8fc8c4f94d983015c6b41392cc6eb5d0a22d4e86f3f.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 358, + 190, + 490, + 198 + ], + "lines": [ + { + "bbox": [ + 358, + 190, + 490, + 198 + ], + "spans": [ + { + "bbox": [ + 358, + 190, + 490, + 198 + ], + "type": "text", + "content": "(a) Statistics of flare-light source relation" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 329, + 200, + 423, + 291 + ], + "blocks": [ + { + "bbox": [ + 329, + 200, + 423, + 291 + ], + "lines": [ + { + "bbox": [ + 329, + 200, + 423, + 291 + ], + "spans": [ + { + "bbox": [ + 329, + 200, + 423, + 291 + ], + "type": "image", + "image_path": "f52a0286cd58d3af36954897bfcfa2a8a5b7e683db0b4a31953ce885676bfecb.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 347, + 292, + 414, + 309 + ], + "lines": [ + { + "bbox": [ + 347, + 292, + 414, + 309 + ], + "spans": [ + { + "bbox": [ + 347, + 292, + 414, + 309 + ], + "type": "text", + "content": "(b) flare-light source relation in x-axis" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 423, + 201, + 519, + 292 + ], + "blocks": [ + { + "bbox": [ + 423, + 201, + 519, + 292 + ], + "lines": [ + { + "bbox": [ + 423, + 201, + 519, + 292 + ], + "spans": [ + { + "bbox": [ + 423, + 201, + 519, + 292 + ], + "type": "image", + "image_path": "0645d13a2265558b07e6c0aba4684dab33f58b27297d2f51ce2ccf6b4f3e65e9.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 442, + 292, + 508, + 309 + ], + "lines": [ + { + "bbox": [ + 442, + 292, + 508, + 309 + ], + "spans": [ + { + "bbox": [ + 442, + 292, + 508, + 309 + ], + "type": "text", + "content": "(c) flare-light source relation in y-axis" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 313, + 545, + 413 + ], + "lines": [ + { + "bbox": [ + 304, + 313, + 545, + 413 + ], + "spans": [ + { + "bbox": [ + 304, + 313, + 545, + 413 + ], + "type": "text", + "content": "Figure 5. Experimental statistics of the optical center symmetry prior. We capture images with reflective flares in a dark environment and mark the position of the light source and reflective flares. " + }, + { + "bbox": [ + 304, + 313, + 545, + 413 + ], + "type": "inline_equation", + "content": "x_{f}" + }, + { + "bbox": [ + 304, + 313, + 545, + 413 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 313, + 545, + 413 + ], + "type": "inline_equation", + "content": "x_{l}" + }, + { + "bbox": [ + 304, + 313, + 545, + 413 + ], + "type": "text", + "content": " represent the x-coordinate of flare and light source and so as " + }, + { + "bbox": [ + 304, + 313, + 545, + 413 + ], + "type": "inline_equation", + "content": "y_{f}" + }, + { + "bbox": [ + 304, + 313, + 545, + 413 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 313, + 545, + 413 + ], + "type": "inline_equation", + "content": "y_{l}" + }, + { + "bbox": [ + 304, + 313, + 545, + 413 + ], + "type": "text", + "content": ". The origin is set to be the center of the image. The relationship between the flare and the light source's absolute values of coordinates are plotted above. From figures (b) and (c), the coordinate relationships are linear and the slope is 1, which confirms our prior." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 421, + 547, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 421, + 547, + 481 + ], + "spans": [ + { + "bbox": [ + 304, + 421, + 547, + 481 + ], + "type": "text", + "content": "moral. As reflective flares are caused by the lens system, it is physically impossible to capture real flare-free and flare-corrupted image pairs. Moreover, given the huge variety of nighttime lights, it would be tedious to synthesize reflective flares for each type of light source." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 486, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 546, + 713 + ], + "type": "text", + "content": "To address the challenge of generating paired images for reflective flare removal, we propose a semi-synthetic dataset, named BracketFlare. Since reflective flares always have the same pattern as the brightest region of the light source, we reduce the exposure time to capture the patterns of the light source's brightest region, which are used to synthesize the reflective flares. Specifically, we mount a professional camera (Sony A7 III) on a tripod and use continuous bracketing to capture five photos of the same static scene with a step of " + }, + { + "bbox": [ + 304, + 486, + 546, + 713 + ], + "type": "inline_equation", + "content": "3\\mathrm{EV}" + }, + { + "bbox": [ + 304, + 486, + 546, + 713 + ], + "type": "text", + "content": " as a group. We then select the normally exposed image and the lowest exposure image from each group to compose the pair. To increase the diversity and realism of our dataset, we choose scenes with different light sources that are bright enough to leave reflective flares. Finally, we collect a dataset consisting of 440 pairs of 4K images as training data and 40 images as test data. A general overview of our dataset is depicted in Fig. 6, showing that the dataset contains diverse and commonly seen flare patterns." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20786" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 74, + 77, + 187, + 152 + ], + "blocks": [ + { + "bbox": [ + 74, + 77, + 187, + 152 + ], + "lines": [ + { + "bbox": [ + 74, + 77, + 187, + 152 + ], + "spans": [ + { + "bbox": [ + 74, + 77, + 187, + 152 + ], + "type": "image", + "image_path": "d4ffd77f063a66e1c48bf33a3ae2f56ed36d52e8e487288f5ebb82b7589f83b5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 96, + 153, + 109, + 160 + ], + "lines": [ + { + "bbox": [ + 96, + 153, + 109, + 160 + ], + "spans": [ + { + "bbox": [ + 96, + 153, + 109, + 160 + ], + "type": "text", + "content": "LED" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 192, + 78, + 392, + 215 + ], + "blocks": [ + { + "bbox": [ + 192, + 78, + 392, + 215 + ], + "lines": [ + { + "bbox": [ + 192, + 78, + 392, + 215 + ], + "spans": [ + { + "bbox": [ + 192, + 78, + 392, + 215 + ], + "type": "image", + "image_path": "dd72b55dad2b6b7e0fa9ec8279fbbed67df4aae6fb867c3fea0b3df49442fa2d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 74, + 162, + 187, + 237 + ], + "blocks": [ + { + "bbox": [ + 74, + 162, + 187, + 237 + ], + "lines": [ + { + "bbox": [ + 74, + 162, + 187, + 237 + ], + "spans": [ + { + "bbox": [ + 74, + 162, + 187, + 237 + ], + "type": "image", + "image_path": "a51b80b24715b9f02cc5f513da312df89b089eed721e1e1e26c48e71abed7eb1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 88, + 238, + 116, + 245 + ], + "lines": [ + { + "bbox": [ + 88, + 238, + 116, + 245 + ], + "spans": [ + { + "bbox": [ + 88, + 238, + 116, + 245 + ], + "type": "text", + "content": "Traffic light" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 192, + 215, + 394, + 244 + ], + "blocks": [ + { + "bbox": [ + 192, + 215, + 394, + 244 + ], + "lines": [ + { + "bbox": [ + 192, + 215, + 394, + 244 + ], + "spans": [ + { + "bbox": [ + 192, + 215, + 394, + 244 + ], + "type": "image", + "image_path": "b2fb03a0035fe1f685e028675d529839498685b13b239678c0e99b4bd339bd1d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 253, + 545, + 276 + ], + "lines": [ + { + "bbox": [ + 46, + 253, + 545, + 276 + ], + "spans": [ + { + "bbox": [ + 46, + 253, + 545, + 276 + ], + "type": "text", + "content": "Figure 6. Visualization and distribution of our dataset. It contains different types of light sources in diverse scenes. Based on light source types, we classify them into eight categories with different flare patterns." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 403, + 77, + 516, + 152 + ], + "blocks": [ + { + "bbox": [ + 403, + 77, + 516, + 152 + ], + "lines": [ + { + "bbox": [ + 403, + 77, + 516, + 152 + ], + "spans": [ + { + "bbox": [ + 403, + 77, + 516, + 152 + ], + "type": "image", + "image_path": "e635f1c97d4e46a6a4c50a6334c2d4513393973f87ef1a69d82250ad493dd8f5.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 406, + 153, + 456, + 161 + ], + "lines": [ + { + "bbox": [ + 406, + 153, + 456, + 161 + ], + "spans": [ + { + "bbox": [ + 406, + 153, + 456, + 161 + ], + "type": "text", + "content": "Square overhead light" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 462, + 153, + 512, + 160 + ], + "lines": [ + { + "bbox": [ + 462, + 153, + 512, + 160 + ], + "spans": [ + { + "bbox": [ + 462, + 153, + 512, + 160 + ], + "type": "text", + "content": "Round overhead light" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 403, + 163, + 515, + 238 + ], + "blocks": [ + { + "bbox": [ + 403, + 163, + 515, + 238 + ], + "lines": [ + { + "bbox": [ + 403, + 163, + 515, + 238 + ], + "spans": [ + { + "bbox": [ + 403, + 163, + 515, + 238 + ], + "type": "image", + "image_path": "1ed183a0c7625b83b3e23b985d49c22d38ae74f8f818fb04ac6ec1bce4c1b6bb.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 419, + 239, + 443, + 246 + ], + "lines": [ + { + "bbox": [ + 419, + 239, + 443, + 246 + ], + "spans": [ + { + "bbox": [ + 419, + 239, + 443, + 246 + ], + "type": "text", + "content": "Light tube" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 474, + 239, + 500, + 247 + ], + "lines": [ + { + "bbox": [ + 474, + 239, + 500, + 247 + ], + "spans": [ + { + "bbox": [ + 474, + 239, + 500, + 247 + ], + "type": "text", + "content": "Light bulb" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 289, + 153, + 304 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 289, + 153, + 304 + ], + "spans": [ + { + "bbox": [ + 47, + 289, + 153, + 304 + ], + "type": "text", + "content": "5. Proposed Method" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 310, + 194, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 310, + 194, + 323 + ], + "spans": [ + { + "bbox": [ + 47, + 310, + 194, + 323 + ], + "type": "text", + "content": "5.1. Postprocessing our dataset" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 329, + 287, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 329, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 46, + 329, + 287, + 483 + ], + "type": "text", + "content": "To increase the diversity of our dataset, we generate paired flare-corrupted and flare-free images on the fly with augmentation. As depicted in Fig. 7(a), we first resize the 4K image to " + }, + { + "bbox": [ + 46, + 329, + 287, + 483 + ], + "type": "inline_equation", + "content": "800 \\times 1200" + }, + { + "bbox": [ + 46, + 329, + 287, + 483 + ], + "type": "text", + "content": ". To simulate the possible misalignment of the image's center and optical center, we introduce a slight translation with " + }, + { + "bbox": [ + 46, + 329, + 287, + 483 + ], + "type": "inline_equation", + "content": "t \\sim U(0,15)" + }, + { + "bbox": [ + 46, + 329, + 287, + 483 + ], + "type": "text", + "content": " pixels to the underexposed light source image. Then, we randomly crop the paired light source image and the normally exposed image with a size of " + }, + { + "bbox": [ + 46, + 329, + 287, + 483 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 46, + 329, + 287, + 483 + ], + "type": "text", + "content": ". Since our captured images are already gamma-encoded, we follow the settings of Wu et al. [22] and linearize them by applying a gamma correction, i.e., " + }, + { + "bbox": [ + 46, + 329, + 287, + 483 + ], + "type": "inline_equation", + "content": "I_g = I^\\gamma" + }, + { + "bbox": [ + 46, + 329, + 287, + 483 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 329, + 287, + 483 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 329, + 287, + 483 + ], + "type": "text", + "content": " randomly sampled from [1.8, 2.2]. All our operations are performed in the linear space." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 485, + 287, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 485, + 287, + 592 + ], + "spans": [ + { + "bbox": [ + 46, + 485, + 287, + 592 + ], + "type": "text", + "content": "Synthetic reflective flare. As lenses reflect light of different wavelengths with varying reflectivity, reflective flares often appear in different colors. To replicate this, we use a random color factor to multiply the low-exposed image. To simulate an unfocused situation, we also add random blur. In accordance with the optical center symmetry prior, we rotate the light source image to synthesize the reflective flare. Finally, we obtain the synthetic reflective flare by multiplying the flare image with an opacity." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 594, + 287, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 676 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 676 + ], + "type": "text", + "content": "Flare-free image. The ISO is fixed for all images in our continuous bracketing. Therefore, our normally exposed images always share a long exposure time and have little noise. To simulate real night scenes, we add Gaussian noise to the images. Finally, we randomly apply a small color offset to the flare-free image. These operations aid in simulating light fog and glare effects in real-world scenarios." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "content": "Flare-corrupted image. To create flare-corrupted images, we add the synthetic reflective flare to the flare-free images. Then, we apply gamma correction to the images with the" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 291, + 545, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 291, + 545, + 339 + ], + "spans": [ + { + "bbox": [ + 305, + 291, + 545, + 339 + ], + "type": "text", + "content": "gamma value obtained by taking the reciprocal of the previously sampled gamma correction value " + }, + { + "bbox": [ + 305, + 291, + 545, + 339 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 305, + 291, + 545, + 339 + ], + "type": "text", + "content": ". The resulting gamma-encoded flare-corrupted image, flare-free image, and flare image are used to train the neural networks." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 345, + 434, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 345, + 434, + 358 + ], + "spans": [ + { + "bbox": [ + 306, + 345, + 434, + 358 + ], + "type": "text", + "content": "5.2. Flare removal pipeline" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 364, + 545, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 364, + 545, + 423 + ], + "spans": [ + { + "bbox": [ + 304, + 364, + 545, + 423 + ], + "type": "text", + "content": "As the optical center symmetry prior is a global prior, it is difficult for CNN-based networks to learn it, and even Transformer-based networks struggle to learn such symmetry. Therefore, it is crucial to incorporate this prior into the network with specialized designs." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 423, + 546, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 423, + 546, + 590 + ], + "spans": [ + { + "bbox": [ + 304, + 423, + 546, + 590 + ], + "type": "text", + "content": "To improve the network's ability to learn the optical center symmetry prior, as shown in Fig. 7(b), we apply a gamma correction of " + }, + { + "bbox": [ + 304, + 423, + 546, + 590 + ], + "type": "inline_equation", + "content": "I_{g} = I^{\\gamma}" + }, + { + "bbox": [ + 304, + 423, + 546, + 590 + ], + "type": "text", + "content": " to the input flare-corrupted image with " + }, + { + "bbox": [ + 304, + 423, + 546, + 590 + ], + "type": "inline_equation", + "content": "\\gamma = 10" + }, + { + "bbox": [ + 304, + 423, + 546, + 590 + ], + "type": "text", + "content": ". This operation extracts the nearly saturated areas of the light source in each channel. We then rotate this nearly saturated image by 180 degrees to obtain a prior image, which is concatenated with the flare-corrupted image to form a six-channel input. This rotated prior image provides an initial approximation of the reflective flare's pattern, which we expect the network to refine and enhance. The network's output is set to six channels, with three channels for the estimated flare-free image and three channels for the estimated flare. Our synthetic data provides supervision for all six channels, which facilitates network training." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 597, + 391, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 597, + 391, + 609 + ], + "spans": [ + { + "bbox": [ + 306, + 597, + 391, + 609 + ], + "type": "text", + "content": "5.3. Loss function" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 305, + 616, + 545, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 616, + 545, + 662 + ], + "spans": [ + { + "bbox": [ + 305, + 616, + 545, + 662 + ], + "type": "text", + "content": "With the supervision of flare image " + }, + { + "bbox": [ + 305, + 616, + 545, + 662 + ], + "type": "inline_equation", + "content": "I_{F}" + }, + { + "bbox": [ + 305, + 616, + 545, + 662 + ], + "type": "text", + "content": " and flare-free image " + }, + { + "bbox": [ + 305, + 616, + 545, + 662 + ], + "type": "inline_equation", + "content": "I_{B}" + }, + { + "bbox": [ + 305, + 616, + 545, + 662 + ], + "type": "text", + "content": ", our final loss is the combination of background loss " + }, + { + "bbox": [ + 305, + 616, + 545, + 662 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{bg}" + }, + { + "bbox": [ + 305, + 616, + 545, + 662 + ], + "type": "text", + "content": ", flare loss " + }, + { + "bbox": [ + 305, + 616, + 545, + 662 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{flare}" + }, + { + "bbox": [ + 305, + 616, + 545, + 662 + ], + "type": "text", + "content": ", and reconstruction loss " + }, + { + "bbox": [ + 305, + 616, + 545, + 662 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rec}" + }, + { + "bbox": [ + 305, + 616, + 545, + 662 + ], + "type": "text", + "content": ". It can be written as:" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 370, + 670, + 545, + 684 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 670, + 545, + 684 + ], + "spans": [ + { + "bbox": [ + 370, + 670, + 545, + 684 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {r e c}} + \\mathcal {L} _ {\\text {f l a r e}} + \\mathcal {L} _ {\\text {b g}}. \\tag {4}", + "image_path": "8c82194d404b717cead980d2e8ed886780643e2bade5f76872829585b604f00b.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 305, + 689, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 715 + ], + "type": "text", + "content": "We define our flare removal network as " + }, + { + "bbox": [ + 305, + 689, + 545, + 715 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 305, + 689, + 545, + 715 + ], + "type": "text", + "content": ". It takes flare-corrupted image " + }, + { + "bbox": [ + 305, + 689, + 545, + 715 + ], + "type": "inline_equation", + "content": "I_{input}" + }, + { + "bbox": [ + 305, + 689, + 545, + 715 + ], + "type": "text", + "content": " and its prior image " + }, + { + "bbox": [ + 305, + 689, + 545, + 715 + ], + "type": "inline_equation", + "content": "I_{prior}" + }, + { + "bbox": [ + 305, + 689, + 545, + 715 + ], + "type": "text", + "content": " as inputs." + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20787" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 85, + 76, + 501, + 224 + ], + "blocks": [ + { + "bbox": [ + 85, + 76, + 501, + 224 + ], + "lines": [ + { + "bbox": [ + 85, + 76, + 501, + 224 + ], + "spans": [ + { + "bbox": [ + 85, + 76, + 501, + 224 + ], + "type": "image", + "image_path": "944c57a82fcb893c5ae85d5767d22274eb557fc0506493a6ca6e315ab8e84746.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 153, + 228, + 247, + 237 + ], + "lines": [ + { + "bbox": [ + 153, + 228, + 247, + 237 + ], + "spans": [ + { + "bbox": [ + 153, + 228, + 247, + 237 + ], + "type": "text", + "content": "(a) Training Data Generation" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 380, + 228, + 451, + 238 + ], + "lines": [ + { + "bbox": [ + 380, + 228, + 451, + 238 + ], + "spans": [ + { + "bbox": [ + 380, + 228, + 451, + 238 + ], + "type": "text", + "content": "(b) Network Learning" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 240, + 545, + 296 + ], + "lines": [ + { + "bbox": [ + 46, + 240, + 545, + 296 + ], + "spans": [ + { + "bbox": [ + 46, + 240, + 545, + 296 + ], + "type": "text", + "content": "Figure 7. The pipeline of our data processing module and flare removal network. The flare image and a flare-free image from our paired dataset are resized and randomly cropped. After data augmentation, they are added to synthesize the corresponding flare-corrupted image. To encode the optical symmetry prior, the input flare-corrupted image is center flipped as the prior image. We also apply gamma correction to extract the light source region. Then, the prior image is concatenated with the flare-corrupted image as the input with 6 channels. Finally, the network predicts an output with 6 channels (a flare-free image with 3 channels and an estimated flare with 3 channels)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 311, + 253, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 311, + 253, + 323 + ], + "spans": [ + { + "bbox": [ + 47, + 311, + 253, + 323 + ], + "type": "text", + "content": "The flare-free image and flare can be estimated by:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 109, + 331, + 287, + 345 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 109, + 331, + 287, + 345 + ], + "spans": [ + { + "bbox": [ + 109, + 331, + 287, + 345 + ], + "type": "interline_equation", + "content": "\\hat {I} _ {B}, \\hat {I} _ {F} = f _ {\\theta} \\left(I _ {\\text {i n p u t}}, I _ {\\text {p r i o r}}\\right). \\tag {5}", + "image_path": "9def9a5929dfe82cc64aa46c150bce748415dc831ceb717ed5d04ffe7060c3aa.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 353, + 286, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 353, + 286, + 376 + ], + "spans": [ + { + "bbox": [ + 47, + 353, + 286, + 376 + ], + "type": "text", + "content": "With the estimated flare and flare-free images, the " + }, + { + "bbox": [ + 47, + 353, + 286, + 376 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 47, + 353, + 286, + 376 + ], + "type": "text", + "content": " reconstruction loss can be expressed as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 110, + 384, + 287, + 399 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 384, + 287, + 399 + ], + "spans": [ + { + "bbox": [ + 110, + 384, + 287, + 399 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r e c}} = \\left| I _ {\\text {i n p u t}} - \\hat {I} _ {F} \\oplus \\hat {I} _ {B} \\right|, \\tag {6}", + "image_path": "6ceef4345b66d06cd9cff661408aaa8d2816c5aa0c237271c469b092a227e1cd.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 407, + 287, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 407, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 46, + 407, + 287, + 430 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 407, + 287, + 430 + ], + "type": "inline_equation", + "content": "\\oplus" + }, + { + "bbox": [ + 46, + 407, + 287, + 430 + ], + "type": "text", + "content": " denotes an element-wise addition operation in the linearized gamma-decoded domain." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 431, + 287, + 549 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 431, + 287, + 549 + ], + "spans": [ + { + "bbox": [ + 47, + 431, + 287, + 549 + ], + "type": "text", + "content": "For the background loss, it consists of a default " + }, + { + "bbox": [ + 47, + 431, + 287, + 549 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_1" + }, + { + "bbox": [ + 47, + 431, + 287, + 549 + ], + "type": "text", + "content": ", a region-aware masked " + }, + { + "bbox": [ + 47, + 431, + 287, + 549 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{mask}" + }, + { + "bbox": [ + 47, + 431, + 287, + 549 + ], + "type": "text", + "content": " term, and a perceptual term " + }, + { + "bbox": [ + 47, + 431, + 287, + 549 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{vgg}" + }, + { + "bbox": [ + 47, + 431, + 287, + 549 + ], + "type": "text", + "content": " with a pre-trained VGG-19 network [17]. Compared with previous works [5, 22, 26], the main difference is the " + }, + { + "bbox": [ + 47, + 431, + 287, + 549 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{mask}" + }, + { + "bbox": [ + 47, + 431, + 287, + 549 + ], + "type": "text", + "content": ". Since the regions of reflective flares are always small, a simple " + }, + { + "bbox": [ + 47, + 431, + 287, + 549 + ], + "type": "inline_equation", + "content": "L_1" + }, + { + "bbox": [ + 47, + 431, + 287, + 549 + ], + "type": "text", + "content": " loss may ignore these regions, leading to the local optimum. To encourage the network to focus on restoring the flare-corrupted regions, we calculate the flare's mask " + }, + { + "bbox": [ + 47, + 431, + 287, + 549 + ], + "type": "inline_equation", + "content": "I_M" + }, + { + "bbox": [ + 47, + 431, + 287, + 549 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 47, + 431, + 287, + 549 + ], + "type": "inline_equation", + "content": ">0.2" + }, + { + "bbox": [ + 47, + 431, + 287, + 549 + ], + "type": "text", + "content": ") from the ground truth of flare image " + }, + { + "bbox": [ + 47, + 431, + 287, + 549 + ], + "type": "inline_equation", + "content": "I_F" + }, + { + "bbox": [ + 47, + 431, + 287, + 549 + ], + "type": "text", + "content": ". This masked loss can be written as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 108, + 559, + 287, + 572 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 559, + 287, + 572 + ], + "spans": [ + { + "bbox": [ + 108, + 559, + 287, + 572 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {m a s k}} = \\left| I _ {M} * \\left(\\hat {I} _ {B} - I _ {B}\\right) \\right|. \\tag {7}", + "image_path": "9d7ad523bc84f7672c1da988d802671d2d2d523cfc36e2cb70b6cfe50687ab17.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 59, + 580, + 249, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 580, + 249, + 594 + ], + "spans": [ + { + "bbox": [ + 59, + 580, + 249, + 594 + ], + "type": "text", + "content": "Overall, the background loss " + }, + { + "bbox": [ + 59, + 580, + 249, + 594 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{bg}" + }, + { + "bbox": [ + 59, + 580, + 249, + 594 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 91, + 602, + 287, + 615 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 602, + 287, + 615 + ], + "spans": [ + { + "bbox": [ + 91, + 602, + 287, + 615 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {b g} = w _ {1} \\mathcal {L} _ {1} + w _ {2} \\mathcal {L} _ {v g g} + w _ {3} \\mathcal {L} _ {\\text {m a s k}}, \\tag {8}", + "image_path": "18af25f2bb313ad3aa868a71e62ab706fcf6742e04ef35b8d657f7ec16070733.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 623, + 287, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 623, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 46, + 623, + 287, + 658 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 623, + 287, + 658 + ], + "type": "inline_equation", + "content": "w_{1}, w_{2}, w_{3}" + }, + { + "bbox": [ + 46, + 623, + 287, + 658 + ], + "type": "text", + "content": " are respectively set to 0.5, 0.1, and 20.0 in our experiments. Flare loss " + }, + { + "bbox": [ + 46, + 623, + 287, + 658 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{flare}" + }, + { + "bbox": [ + 46, + 623, + 287, + 658 + ], + "type": "text", + "content": " has the identical form as the background loss." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 670, + 128, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 128, + 683 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 128, + 683 + ], + "type": "text", + "content": "6. Experiments" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "Evaluation on synthetic data. To evaluate our method, we collect 40 paired data using the process mentioned in" + } + ] + } + ], + "index": 15 + }, + { + "type": "table", + "bbox": [ + 331, + 346, + 518, + 421 + ], + "blocks": [ + { + "bbox": [ + 305, + 309, + 545, + 342 + ], + "lines": [ + { + "bbox": [ + 305, + 309, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 305, + 309, + 545, + 342 + ], + "type": "text", + "content": "Table 1. Percentage of users favoring our results (our dataset + MPRNet†) vs. previous work. Our method outperforms existing methods on both indoor and outdoor images." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 331, + 346, + 518, + 421 + ], + "lines": [ + { + "bbox": [ + 331, + 346, + 518, + 421 + ], + "spans": [ + { + "bbox": [ + 331, + 346, + 518, + 421 + ], + "type": "table", + "html": "
MethodOutdoorIndoor
Reflection removal [26]3.0%8.5%
Flare spot removal [1]4.5%0.0%
Flare removal (Wu et al.) [22]0.5%3.5%
Flare removal (Dai et al.) [5]1.5%2.5%
Ours90.5%85.5%
", + "image_path": "3a38e003e69dbf0a5a0cde48d88d7d95805c2f14ef6ae736e03d80b17c38f883.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 424, + 545, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 424, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 304, + 424, + 545, + 460 + ], + "type": "text", + "content": "Sec. 4 to generate a synthetic test dataset, in which ground truth images are available. We adopt PSNR, SSIM [20], and LPIPS [25] as our evaluation metrics." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 460, + 545, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 545, + 615 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 545, + 615 + ], + "type": "text", + "content": "Evaluation on real data. To show our method's generalization ability, we also collect a dataset with the photos captured by different types of mobile phones, including iPhone 12 pro, iPhone 13 pro, iPhone 14 pro, Huawei mate 30, Huawei nova 5i Pro, Huawei P40, Samsung Galaxy s22, Samsung Galaxy Note9, ZTE Axon 20 5G and Oppo Reno4 SE 5G. For real cases, we cannot obtain the ground truth. Since reflective flares always locate in small regions, they are hard to be evaluated by existing non-reference image quality assessment. These metrics emphasize more on global image quality and are not specially designed for our task. Therefore, we provide more visual results and conduct a user study to evaluate the performance." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 623, + 499, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 623, + 499, + 635 + ], + "spans": [ + { + "bbox": [ + 306, + 623, + 499, + 635 + ], + "type": "text", + "content": "6.1. Comparison with different networks" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 713 + ], + "type": "text", + "content": "To validate the performance of our pipeline, we train different CNN-based and Transformer-based networks with our data. We use the default network settings of MPR-Net [24], HINet [4], and Uformer [21] and reduce the feature channels of Restormer [23] from 48 to 24 due to GPU memory constraint. In our experiments, these models are" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20788" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 91, + 72, + 157, + 139 + ], + "blocks": [ + { + "bbox": [ + 91, + 72, + 157, + 139 + ], + "lines": [ + { + "bbox": [ + 91, + 72, + 157, + 139 + ], + "spans": [ + { + "bbox": [ + 91, + 72, + 157, + 139 + ], + "type": "image", + "image_path": "db4ff8be7c796b93fd3ee33705ab6a454fcc7a2fcc8dacfa2d2f737fae5c9c89.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 91, + 140, + 156, + 205 + ], + "blocks": [ + { + "bbox": [ + 91, + 140, + 156, + 205 + ], + "lines": [ + { + "bbox": [ + 91, + 140, + 156, + 205 + ], + "spans": [ + { + "bbox": [ + 91, + 140, + 156, + 205 + ], + "type": "image", + "image_path": "bc22b91a1000dfa5e315e05097c4ed22f7074250f4826bc8596267267b625c27.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 91, + 206, + 156, + 272 + ], + "blocks": [ + { + "bbox": [ + 91, + 206, + 156, + 272 + ], + "lines": [ + { + "bbox": [ + 91, + 206, + 156, + 272 + ], + "spans": [ + { + "bbox": [ + 91, + 206, + 156, + 272 + ], + "type": "image", + "image_path": "0e28a7387ad25a3188e0f00ef64b7976883921e8a2de6ab19a7ec91ee287bff3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 114, + 274, + 132, + 284 + ], + "lines": [ + { + "bbox": [ + 114, + 274, + 132, + 284 + ], + "spans": [ + { + "bbox": [ + 114, + 274, + 132, + 284 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 159, + 72, + 224, + 139 + ], + "blocks": [ + { + "bbox": [ + 159, + 72, + 224, + 139 + ], + "lines": [ + { + "bbox": [ + 159, + 72, + 224, + 139 + ], + "spans": [ + { + "bbox": [ + 159, + 72, + 224, + 139 + ], + "type": "image", + "image_path": "48ffa684854b80cdb98ae74a698c88ff432674d9fb5ebdc08e8337f3afd8ad7a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 159, + 140, + 224, + 205 + ], + "blocks": [ + { + "bbox": [ + 159, + 140, + 224, + 205 + ], + "lines": [ + { + "bbox": [ + 159, + 140, + 224, + 205 + ], + "spans": [ + { + "bbox": [ + 159, + 140, + 224, + 205 + ], + "type": "image", + "image_path": "fb67e9933b70653afd1d83fcf53e40caca193a156f97dda8482eee51466f8589.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 159, + 206, + 224, + 272 + ], + "blocks": [ + { + "bbox": [ + 159, + 206, + 224, + 272 + ], + "lines": [ + { + "bbox": [ + 159, + 206, + 224, + 272 + ], + "spans": [ + { + "bbox": [ + 159, + 206, + 224, + 272 + ], + "type": "image", + "image_path": "3bc747c7f2ed860729a02cec77fb2ea8f777368fc3dfe005e630ded9732442d3.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 163, + 274, + 222, + 283 + ], + "lines": [ + { + "bbox": [ + 163, + 274, + 222, + 283 + ], + "spans": [ + { + "bbox": [ + 163, + 274, + 222, + 283 + ], + "type": "text", + "content": "Reflection removal" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 227, + 72, + 293, + 139 + ], + "blocks": [ + { + "bbox": [ + 227, + 72, + 293, + 139 + ], + "lines": [ + { + "bbox": [ + 227, + 72, + 293, + 139 + ], + "spans": [ + { + "bbox": [ + 227, + 72, + 293, + 139 + ], + "type": "image", + "image_path": "9726bc3315bca57d0b557af8ae7ef3e278ce9c2ec6488003a825d2315be3e16d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 227, + 140, + 293, + 205 + ], + "blocks": [ + { + "bbox": [ + 227, + 140, + 293, + 205 + ], + "lines": [ + { + "bbox": [ + 227, + 140, + 293, + 205 + ], + "spans": [ + { + "bbox": [ + 227, + 140, + 293, + 205 + ], + "type": "image", + "image_path": "b8e74dc2197d334dd06372da48bca80eca1ad9daf9c91c24aa62b0860e72bcf5.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 227, + 206, + 293, + 272 + ], + "blocks": [ + { + "bbox": [ + 227, + 206, + 293, + 272 + ], + "lines": [ + { + "bbox": [ + 227, + 206, + 293, + 272 + ], + "spans": [ + { + "bbox": [ + 227, + 206, + 293, + 272 + ], + "type": "image", + "image_path": "23a4606430801aba1811b766defc720f4d0cfe5aacb8bf09d3953913150a56bf.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 233, + 274, + 290, + 283 + ], + "lines": [ + { + "bbox": [ + 233, + 274, + 290, + 283 + ], + "spans": [ + { + "bbox": [ + 233, + 274, + 290, + 283 + ], + "type": "text", + "content": "Flare spot removal" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 294, + 72, + 360, + 139 + ], + "blocks": [ + { + "bbox": [ + 294, + 72, + 360, + 139 + ], + "lines": [ + { + "bbox": [ + 294, + 72, + 360, + 139 + ], + "spans": [ + { + "bbox": [ + 294, + 72, + 360, + 139 + ], + "type": "image", + "image_path": "8dceb54be224d6e022f18af3acaf48c64cb605ab720e9d6a6b10ac193112287a.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 294, + 140, + 360, + 205 + ], + "blocks": [ + { + "bbox": [ + 294, + 140, + 360, + 205 + ], + "lines": [ + { + "bbox": [ + 294, + 140, + 360, + 205 + ], + "spans": [ + { + "bbox": [ + 294, + 140, + 360, + 205 + ], + "type": "image", + "image_path": "f8d85319bd51a4fc6bec357cfaa83957e1e1dabd92b0d49ea1e8f58d561908fe.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 294, + 206, + 360, + 272 + ], + "blocks": [ + { + "bbox": [ + 294, + 206, + 360, + 272 + ], + "lines": [ + { + "bbox": [ + 294, + 206, + 360, + 272 + ], + "spans": [ + { + "bbox": [ + 294, + 206, + 360, + 272 + ], + "type": "image", + "image_path": "aa92a277056b15d31558666ee5a13aebed60f18d25571335288f8d68157f78ad.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 316, + 274, + 344, + 283 + ], + "lines": [ + { + "bbox": [ + 316, + 274, + 344, + 283 + ], + "spans": [ + { + "bbox": [ + 316, + 274, + 344, + 283 + ], + "type": "text", + "content": "Wuet al." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 362, + 72, + 429, + 139 + ], + "blocks": [ + { + "bbox": [ + 362, + 72, + 429, + 139 + ], + "lines": [ + { + "bbox": [ + 362, + 72, + 429, + 139 + ], + "spans": [ + { + "bbox": [ + 362, + 72, + 429, + 139 + ], + "type": "image", + "image_path": "3e57a9445697106823a328fa71b26f3fcbec8015ddc63e993a976cbdc6cc907a.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 362, + 140, + 429, + 205 + ], + "blocks": [ + { + "bbox": [ + 362, + 140, + 429, + 205 + ], + "lines": [ + { + "bbox": [ + 362, + 140, + 429, + 205 + ], + "spans": [ + { + "bbox": [ + 362, + 140, + 429, + 205 + ], + "type": "image", + "image_path": "bef219f9c204b4842482eac749526d905c5b43bd1e2910b1b2d9cca28882669a.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 362, + 206, + 429, + 272 + ], + "blocks": [ + { + "bbox": [ + 362, + 206, + 429, + 272 + ], + "lines": [ + { + "bbox": [ + 362, + 206, + 429, + 272 + ], + "spans": [ + { + "bbox": [ + 362, + 206, + 429, + 272 + ], + "type": "image", + "image_path": "2f77e8e26b73e42d00f65cf69672053f645be9217313011745a8d08188f41a0a.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 383, + 274, + 411, + 283 + ], + "lines": [ + { + "bbox": [ + 383, + 274, + 411, + 283 + ], + "spans": [ + { + "bbox": [ + 383, + 274, + 411, + 283 + ], + "type": "text", + "content": "Dai et al." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 431, + 72, + 496, + 139 + ], + "blocks": [ + { + "bbox": [ + 431, + 72, + 496, + 139 + ], + "lines": [ + { + "bbox": [ + 431, + 72, + 496, + 139 + ], + "spans": [ + { + "bbox": [ + 431, + 72, + 496, + 139 + ], + "type": "image", + "image_path": "6d52eaf9c87e6835d71acbc3c56c9071d27f2041706dc360668e891a7191e1e2.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 431, + 140, + 496, + 205 + ], + "blocks": [ + { + "bbox": [ + 431, + 140, + 496, + 205 + ], + "lines": [ + { + "bbox": [ + 431, + 140, + 496, + 205 + ], + "spans": [ + { + "bbox": [ + 431, + 140, + 496, + 205 + ], + "type": "image", + "image_path": "c40034e529549cb7a8662358b9d3027e031b9aa817278645b37dd66042d81011.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 431, + 206, + 496, + 272 + ], + "blocks": [ + { + "bbox": [ + 431, + 206, + 496, + 272 + ], + "lines": [ + { + "bbox": [ + 431, + 206, + 496, + 272 + ], + "spans": [ + { + "bbox": [ + 431, + 206, + 496, + 272 + ], + "type": "image", + "image_path": "0d4dc6c70a282c5d96d04a6178755824f8984eb779e1443017610483f6412173.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 456, + 274, + 472, + 283 + ], + "lines": [ + { + "bbox": [ + 456, + 274, + 472, + 283 + ], + "spans": [ + { + "bbox": [ + 456, + 274, + 472, + 283 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 54, + 303, + 122, + 372 + ], + "blocks": [ + { + "bbox": [ + 72, + 287, + 520, + 299 + ], + "lines": [ + { + "bbox": [ + 72, + 287, + 520, + 299 + ], + "spans": [ + { + "bbox": [ + 72, + 287, + 520, + 299 + ], + "type": "text", + "content": "Figure 8. Visual comparison of different flare removal and reflection removal methods on real-world nighttime flare images." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 303, + 122, + 372 + ], + "lines": [ + { + "bbox": [ + 54, + 303, + 122, + 372 + ], + "spans": [ + { + "bbox": [ + 54, + 303, + 122, + 372 + ], + "type": "image", + "image_path": "5c103a743c10fdfd45e41cf722277a2849943a8aa1234d06eb2f5957102d60f2.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 54, + 372, + 122, + 440 + ], + "blocks": [ + { + "bbox": [ + 54, + 372, + 122, + 440 + ], + "lines": [ + { + "bbox": [ + 54, + 372, + 122, + 440 + ], + "spans": [ + { + "bbox": [ + 54, + 372, + 122, + 440 + ], + "type": "image", + "image_path": "eb461777130ef69b9236661f07717fe9f5641feda42819453bcd47a5bc06cc80.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 77, + 440, + 93, + 448 + ], + "lines": [ + { + "bbox": [ + 77, + 440, + 93, + 448 + ], + "spans": [ + { + "bbox": [ + 77, + 440, + 93, + 448 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 74, + 453, + 517, + 464 + ], + "lines": [ + { + "bbox": [ + 74, + 453, + 517, + 464 + ], + "spans": [ + { + "bbox": [ + 74, + 453, + 517, + 464 + ], + "type": "text", + "content": "Figure 9. Visual comparison of different flare removal and reflection removal methods on synthetic nighttime flare images." + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 123, + 303, + 190, + 372 + ], + "blocks": [ + { + "bbox": [ + 123, + 303, + 190, + 372 + ], + "lines": [ + { + "bbox": [ + 123, + 303, + 190, + 372 + ], + "spans": [ + { + "bbox": [ + 123, + 303, + 190, + 372 + ], + "type": "image", + "image_path": "001f546420f7ac9066da37a053b129c6261a0974128a3560d7eb563f79b64caf.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 123, + 372, + 190, + 439 + ], + "blocks": [ + { + "bbox": [ + 123, + 372, + 190, + 439 + ], + "lines": [ + { + "bbox": [ + 123, + 372, + 190, + 439 + ], + "spans": [ + { + "bbox": [ + 123, + 372, + 190, + 439 + ], + "type": "image", + "image_path": "e80bf37784cbe13dd014e5671dfd7d6b11628fa2aa90fa420518e7095da2e6f4.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 129, + 440, + 186, + 448 + ], + "lines": [ + { + "bbox": [ + 129, + 440, + 186, + 448 + ], + "spans": [ + { + "bbox": [ + 129, + 440, + 186, + 448 + ], + "type": "text", + "content": "Reflection removal" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 192, + 303, + 260, + 371 + ], + "blocks": [ + { + "bbox": [ + 192, + 303, + 260, + 371 + ], + "lines": [ + { + "bbox": [ + 192, + 303, + 260, + 371 + ], + "spans": [ + { + "bbox": [ + 192, + 303, + 260, + 371 + ], + "type": "image", + "image_path": "002c0ca510c5b7f08626df2b9f24f972d15c42fa0c8e41ef910ca8cb17e55993.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 192, + 372, + 260, + 440 + ], + "blocks": [ + { + "bbox": [ + 192, + 372, + 260, + 440 + ], + "lines": [ + { + "bbox": [ + 192, + 372, + 260, + 440 + ], + "spans": [ + { + "bbox": [ + 192, + 372, + 260, + 440 + ], + "type": "image", + "image_path": "2333412be9c73797574131e92a19711fb6975819bee7f7b5f2cbb6afa0633335.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 200, + 440, + 255, + 449 + ], + "lines": [ + { + "bbox": [ + 200, + 440, + 255, + 449 + ], + "spans": [ + { + "bbox": [ + 200, + 440, + 255, + 449 + ], + "type": "text", + "content": "Flare spot removal" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 261, + 303, + 329, + 372 + ], + "blocks": [ + { + "bbox": [ + 261, + 303, + 329, + 372 + ], + "lines": [ + { + "bbox": [ + 261, + 303, + 329, + 372 + ], + "spans": [ + { + "bbox": [ + 261, + 303, + 329, + 372 + ], + "type": "image", + "image_path": "bcf4a3be8cda5af200cb84d6ce37360377f7611326afb3a65e7ad0d52eb3a370.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 261, + 372, + 329, + 440 + ], + "blocks": [ + { + "bbox": [ + 261, + 372, + 329, + 440 + ], + "lines": [ + { + "bbox": [ + 261, + 372, + 329, + 440 + ], + "spans": [ + { + "bbox": [ + 261, + 372, + 329, + 440 + ], + "type": "image", + "image_path": "42051f6c40f73ce84a6d2e414a9374a777a628d30fb0f39fa21aafb7a7e63770.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 280, + 441, + 306, + 448 + ], + "lines": [ + { + "bbox": [ + 280, + 441, + 306, + 448 + ], + "spans": [ + { + "bbox": [ + 280, + 441, + 306, + 448 + ], + "type": "text", + "content": "Wu et al." + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 330, + 303, + 399, + 372 + ], + "blocks": [ + { + "bbox": [ + 330, + 303, + 399, + 372 + ], + "lines": [ + { + "bbox": [ + 330, + 303, + 399, + 372 + ], + "spans": [ + { + "bbox": [ + 330, + 303, + 399, + 372 + ], + "type": "image", + "image_path": "64448d4dcb9a8363b02eb8bbb01bb05ff2c634ca7a0e18fe5ff8206689265bbc.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 330, + 372, + 399, + 440 + ], + "blocks": [ + { + "bbox": [ + 330, + 372, + 399, + 440 + ], + "lines": [ + { + "bbox": [ + 330, + 372, + 399, + 440 + ], + "spans": [ + { + "bbox": [ + 330, + 372, + 399, + 440 + ], + "type": "image", + "image_path": "dc66efa5ba0e54a7b9d8ed6be7d8a48b367a5e580dc258a5407b3098e500e28f.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 348, + 440, + 374, + 448 + ], + "lines": [ + { + "bbox": [ + 348, + 440, + 374, + 448 + ], + "spans": [ + { + "bbox": [ + 348, + 440, + 374, + 448 + ], + "type": "text", + "content": "Dai et al." + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 399, + 303, + 467, + 372 + ], + "blocks": [ + { + "bbox": [ + 399, + 303, + 467, + 372 + ], + "lines": [ + { + "bbox": [ + 399, + 303, + 467, + 372 + ], + "spans": [ + { + "bbox": [ + 399, + 303, + 467, + 372 + ], + "type": "image", + "image_path": "02197a3b7e9d996222ca1bf773cd4d8819f2d1725d965881acffadee4ebb446f.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 399, + 372, + 467, + 440 + ], + "blocks": [ + { + "bbox": [ + 399, + 372, + 467, + 440 + ], + "lines": [ + { + "bbox": [ + 399, + 372, + 467, + 440 + ], + "spans": [ + { + "bbox": [ + 399, + 372, + 467, + 440 + ], + "type": "image", + "image_path": "23e47ddc750bc19b22df9a39b5fde190f01e91f312ca9c9ac66d8ea45a26ee81.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 425, + 440, + 441, + 448 + ], + "lines": [ + { + "bbox": [ + 425, + 440, + 441, + 448 + ], + "spans": [ + { + "bbox": [ + 425, + 440, + 441, + 448 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_caption" + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 468, + 303, + 536, + 372 + ], + "blocks": [ + { + "bbox": [ + 468, + 303, + 536, + 372 + ], + "lines": [ + { + "bbox": [ + 468, + 303, + 536, + 372 + ], + "spans": [ + { + "bbox": [ + 468, + 303, + 536, + 372 + ], + "type": "image", + "image_path": "b3768e6c35fca6693740ce3733b1634ae6277b7d476ed831dea1a54412fd90ec.jpg" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_body" + } + ], + "index": 43 + }, + { + "type": "image", + "bbox": [ + 468, + 372, + 536, + 440 + ], + "blocks": [ + { + "bbox": [ + 468, + 372, + 536, + 440 + ], + "lines": [ + { + "bbox": [ + 468, + 372, + 536, + 440 + ], + "spans": [ + { + "bbox": [ + 468, + 372, + 536, + 440 + ], + "type": "image", + "image_path": "84f091058fd93a70875f2f2fbd32e65868afd22ff8364c01c59e1e62883795a2.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 482, + 440, + 520, + 448 + ], + "lines": [ + { + "bbox": [ + 482, + 440, + 520, + 448 + ], + "spans": [ + { + "bbox": [ + 482, + 440, + 520, + 448 + ], + "type": "text", + "content": "Ground truth" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_caption" + } + ], + "index": 44 + }, + { + "bbox": [ + 46, + 476, + 287, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 476, + 287, + 584 + ], + "spans": [ + { + "bbox": [ + 46, + 476, + 287, + 584 + ], + "type": "text", + "content": "trained on an NVIDIA Geforce RTX3090 with a batch size of 2. We use Adam optimizer [9] with a learning rate of 1e-4 and train a model for 500k iterations. Besides, we adopt a MultiStepLR with a milestone of 200k and set the gamma to 0.5. These models are evaluated using the synthetic test dataset as shown in Table 4. From experiments, we find that " + }, + { + "bbox": [ + 46, + 476, + 287, + 584 + ], + "type": "inline_equation", + "content": "\\mathrm{MRPNet}^{\\dagger}" + }, + { + "bbox": [ + 46, + 476, + 287, + 584 + ], + "type": "text", + "content": " [24] performs better than other networks. Thus, we treat " + }, + { + "bbox": [ + 46, + 476, + 287, + 584 + ], + "type": "inline_equation", + "content": "\\mathrm{MPRNet}^{\\dagger}" + }, + { + "bbox": [ + 46, + 476, + 287, + 584 + ], + "type": "text", + "content": " [24] as our baseline method and conduct the following ablation study." + } + ] + } + ], + "index": 47 + }, + { + "bbox": [ + 47, + 590, + 225, + 602 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 590, + 225, + 602 + ], + "spans": [ + { + "bbox": [ + 47, + 590, + 225, + 602 + ], + "type": "text", + "content": "6.2. Comparison with previous works" + } + ] + } + ], + "index": 48 + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": "We compare our method with different flare removal datasets [5, 22] and reflection removal dataset [26] on both real and synthetic data. We use the pre-trained models of Dai et al. [5] and Zhang et al. [26]. Since Wu et al. [22] do not release their model, we retrain their network using their provided data and code. Besides, we also compare our method with the traditional flare spot removal method [1]. As shown in Table 2, since most of these methods are not specially designed for reflective flare removal, they cannot" + } + ] + } + ], + "index": 49 + }, + { + "type": "table", + "bbox": [ + 307, + 489, + 544, + 566 + ], + "blocks": [ + { + "bbox": [ + 324, + 475, + 526, + 486 + ], + "lines": [ + { + "bbox": [ + 324, + 475, + 526, + 486 + ], + "spans": [ + { + "bbox": [ + 324, + 475, + 526, + 486 + ], + "type": "text", + "content": "Table 2. Quantitative comparison on synthetic test data." + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 489, + 544, + 566 + ], + "lines": [ + { + "bbox": [ + 307, + 489, + 544, + 566 + ], + "spans": [ + { + "bbox": [ + 307, + 489, + 544, + 566 + ], + "type": "table", + "html": "
MethodPSNRSSIMLPIPSMasked PSNR
Input37.300.9900.02521.68
Reflection removal [26]22.020.8300.07422.88
Flare spot removal [1]38.290.9900.02422.53
Flare removal (Wu et al.) [22]26.130.8950.05521.89
Flare removal (Dai et al.) [5]27.490.9420.05021.85
Ours48.410.9940.00432.09
", + "image_path": "3343bc6466eb108e02a209aa590239642ca3fcdb556255a42f683840f904247d.jpg" + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "table_body" + } + ], + "index": 51 + }, + { + "bbox": [ + 304, + 568, + 545, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 568, + 545, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 568, + 545, + 616 + ], + "type": "text", + "content": "accurately remove the reflective flare while retaining image details simultaneously. Although Asha et al. [1] can detect the bright spots, as shown in Fig. 8, it may also remove the light source which is quite common at night." + } + ] + } + ], + "index": 52 + }, + { + "bbox": [ + 304, + 616, + 545, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 616, + 545, + 652 + ], + "spans": [ + { + "bbox": [ + 304, + 616, + 545, + 652 + ], + "type": "text", + "content": "User Study. To evaluate these methods in real-world situations, we also conduct a user study with 20 participants with 20 images randomly selected from our real data in Table 1." + } + ] + } + ], + "index": 53 + }, + { + "bbox": [ + 306, + 659, + 397, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 659, + 397, + 672 + ], + "spans": [ + { + "bbox": [ + 306, + 659, + 397, + 672 + ], + "type": "text", + "content": "6.3. Ablation study" + } + ] + } + ], + "index": 54 + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "content": "Loss function. We conduct an ablation study to prove the necessity of each loss function used in our baseline. In our experiments, we train the model without adding " + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text {flare }}" + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 55 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20789" + } + ] + } + ], + "index": 56 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 64, + 76, + 141, + 152 + ], + "blocks": [ + { + "bbox": [ + 52, + 107, + 61, + 128 + ], + "lines": [ + { + "bbox": [ + 52, + 107, + 61, + 128 + ], + "spans": [ + { + "bbox": [ + 52, + 107, + 61, + 128 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 0, + "angle": 270, + "type": "image_caption" + }, + { + "bbox": [ + 64, + 76, + 141, + 152 + ], + "lines": [ + { + "bbox": [ + 64, + 76, + 141, + 152 + ], + "spans": [ + { + "bbox": [ + 64, + 76, + 141, + 152 + ], + "type": "image", + "image_path": "6da115c232ba0dd5a5583753908fc58d30228cd49e6f4ecc78904616742b6bf7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 64, + 154, + 141, + 231 + ], + "blocks": [ + { + "bbox": [ + 52, + 183, + 61, + 209 + ], + "lines": [ + { + "bbox": [ + 52, + 183, + 61, + 209 + ], + "spans": [ + { + "bbox": [ + 52, + 183, + 61, + 209 + ], + "type": "text", + "content": "Output" + } + ] + } + ], + "index": 1, + "angle": 270, + "type": "image_caption" + }, + { + "bbox": [ + 64, + 154, + 141, + 231 + ], + "lines": [ + { + "bbox": [ + 64, + 154, + 141, + 231 + ], + "spans": [ + { + "bbox": [ + 64, + 154, + 141, + 231 + ], + "type": "image", + "image_path": "c313f5f2cbae4b35bb8264909bc35cf7f3cd15657bfd1e7981d3c0519e323514.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 77, + 232, + 128, + 240 + ], + "lines": [ + { + "bbox": [ + 77, + 232, + 128, + 240 + ], + "spans": [ + { + "bbox": [ + 77, + 232, + 128, + 240 + ], + "type": "text", + "content": "(a) iPhone 14 Pro" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 67, + 251, + 525, + 263 + ], + "lines": [ + { + "bbox": [ + 67, + 251, + 525, + 263 + ], + "spans": [ + { + "bbox": [ + 67, + 251, + 525, + 263 + ], + "type": "text", + "content": "Figure 10. Visual comparison of flare removal on real-world nighttime flare images that are captured by different smartphones." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 144, + 76, + 220, + 152 + ], + "blocks": [ + { + "bbox": [ + 144, + 76, + 220, + 152 + ], + "lines": [ + { + "bbox": [ + 144, + 76, + 220, + 152 + ], + "spans": [ + { + "bbox": [ + 144, + 76, + 220, + 152 + ], + "type": "image", + "image_path": "567933723a794726eedc088fc74a3ce60c96ae3b31f346b59cb140154b6f75c4.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 144, + 154, + 220, + 231 + ], + "blocks": [ + { + "bbox": [ + 144, + 154, + 220, + 231 + ], + "lines": [ + { + "bbox": [ + 144, + 154, + 220, + 231 + ], + "spans": [ + { + "bbox": [ + 144, + 154, + 220, + 231 + ], + "type": "image", + "image_path": "83949e84d2e25fd226a074c25b0bc9c52674adcaec4c411e702138431b4674bd.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 151, + 232, + 203, + 240 + ], + "lines": [ + { + "bbox": [ + 151, + 232, + 203, + 240 + ], + "spans": [ + { + "bbox": [ + 151, + 232, + 203, + 240 + ], + "type": "text", + "content": "(b) iPhone 13 Pro" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 223, + 75, + 300, + 152 + ], + "blocks": [ + { + "bbox": [ + 223, + 75, + 300, + 152 + ], + "lines": [ + { + "bbox": [ + 223, + 75, + 300, + 152 + ], + "spans": [ + { + "bbox": [ + 223, + 75, + 300, + 152 + ], + "type": "image", + "image_path": "73d94687b622709963faaf5204d8f06d1bd7de4bbad8617103a4dee80f0fb428.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 223, + 154, + 299, + 231 + ], + "blocks": [ + { + "bbox": [ + 223, + 154, + 299, + 231 + ], + "lines": [ + { + "bbox": [ + 223, + 154, + 299, + 231 + ], + "spans": [ + { + "bbox": [ + 223, + 154, + 299, + 231 + ], + "type": "image", + "image_path": "91a35f8607fde00c87901c6623428def7285e5b16b78925bd304ace3c1089ff2.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 229, + 232, + 290, + 240 + ], + "lines": [ + { + "bbox": [ + 229, + 232, + 290, + 240 + ], + "spans": [ + { + "bbox": [ + 229, + 232, + 290, + 240 + ], + "type": "text", + "content": "(c) ZTE Axon 20 5G" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 302, + 76, + 378, + 152 + ], + "blocks": [ + { + "bbox": [ + 302, + 76, + 378, + 152 + ], + "lines": [ + { + "bbox": [ + 302, + 76, + 378, + 152 + ], + "spans": [ + { + "bbox": [ + 302, + 76, + 378, + 152 + ], + "type": "image", + "image_path": "30b04c581f92f4b186b9d63e840c9919965a29ddfe621646f93c8f427cf16ff1.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 302, + 154, + 378, + 231 + ], + "blocks": [ + { + "bbox": [ + 302, + 154, + 378, + 231 + ], + "lines": [ + { + "bbox": [ + 302, + 154, + 378, + 231 + ], + "spans": [ + { + "bbox": [ + 302, + 154, + 378, + 231 + ], + "type": "image", + "image_path": "a1f0614f23ad77ef1efde7eed260fca04cedade4d55b1bf2ca302a4fb2ddb2b1.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 303, + 232, + 376, + 240 + ], + "lines": [ + { + "bbox": [ + 303, + 232, + 376, + 240 + ], + "spans": [ + { + "bbox": [ + 303, + 232, + 376, + 240 + ], + "type": "text", + "content": "(d) Samsung Galaxy S22" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 381, + 76, + 457, + 152 + ], + "blocks": [ + { + "bbox": [ + 381, + 76, + 457, + 152 + ], + "lines": [ + { + "bbox": [ + 381, + 76, + 457, + 152 + ], + "spans": [ + { + "bbox": [ + 381, + 76, + 457, + 152 + ], + "type": "image", + "image_path": "908f312e9fa316ab94f27ea10306713f837cb992bb904e1236a89061354a733b.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 381, + 154, + 457, + 231 + ], + "blocks": [ + { + "bbox": [ + 381, + 154, + 457, + 231 + ], + "lines": [ + { + "bbox": [ + 381, + 154, + 457, + 231 + ], + "spans": [ + { + "bbox": [ + 381, + 154, + 457, + 231 + ], + "type": "image", + "image_path": "11aefdabac495aa6b62bf1e9bd24dfb6144975bb9473fa032e5b30c120eecb74.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 389, + 232, + 447, + 239 + ], + "lines": [ + { + "bbox": [ + 389, + 232, + 447, + 239 + ], + "spans": [ + { + "bbox": [ + 389, + 232, + 447, + 239 + ], + "type": "text", + "content": "(e) Huawei Mate 30" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 460, + 76, + 536, + 152 + ], + "blocks": [ + { + "bbox": [ + 460, + 76, + 536, + 152 + ], + "lines": [ + { + "bbox": [ + 460, + 76, + 536, + 152 + ], + "spans": [ + { + "bbox": [ + 460, + 76, + 536, + 152 + ], + "type": "image", + "image_path": "9e81efc2c6ce08bba636e382b96b40a229eadd92f350ec00531a9b7d88f796c6.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 460, + 154, + 536, + 230 + ], + "blocks": [ + { + "bbox": [ + 460, + 154, + 536, + 230 + ], + "lines": [ + { + "bbox": [ + 460, + 154, + 536, + 230 + ], + "spans": [ + { + "bbox": [ + 460, + 154, + 536, + 230 + ], + "type": "image", + "image_path": "ada74c8871fef56c6b26ff81dd53c6241b3b53ba0e6c0a707477d1fb94a14211.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 465, + 232, + 531, + 239 + ], + "lines": [ + { + "bbox": [ + 465, + 232, + 531, + 239 + ], + "spans": [ + { + "bbox": [ + 465, + 232, + 531, + 239 + ], + "type": "text", + "content": "(f) Huawei nova 5i Pro" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "table", + "bbox": [ + 66, + 285, + 265, + 359 + ], + "blocks": [ + { + "bbox": [ + 73, + 271, + 260, + 282 + ], + "lines": [ + { + "bbox": [ + 73, + 271, + 260, + 282 + ], + "spans": [ + { + "bbox": [ + 73, + 271, + 260, + 282 + ], + "type": "text", + "content": "Table 3. Ablation study for different loss functions." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 66, + 285, + 265, + 359 + ], + "lines": [ + { + "bbox": [ + 66, + 285, + 265, + 359 + ], + "spans": [ + { + "bbox": [ + 66, + 285, + 265, + 359 + ], + "type": "table", + "html": "
LossPSNRSSIMLPIPSMasked PSNR
w/o Lflare47.360.9930.00430.63
w/o Lrec48.180.9940.00331.33
w/o Lvgg48.050.9940.00431.39
w/o Lmask48.010.9930.00331.03
Full48.410.9940.00432.09
", + "image_path": "865ebf85c6f8a9633fafbf17e8c8dee092da66bfae9c21569a47d123fe42ccc3.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "table_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 63, + 365, + 129, + 430 + ], + "blocks": [ + { + "bbox": [ + 63, + 365, + 129, + 430 + ], + "lines": [ + { + "bbox": [ + 63, + 365, + 129, + 430 + ], + "spans": [ + { + "bbox": [ + 63, + 365, + 129, + 430 + ], + "type": "image", + "image_path": "d1a31061421c916d5a577f52ef6c73dc5380d1e55ff2fbef1291ffa43b92d2c8.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 63, + 432, + 129, + 497 + ], + "blocks": [ + { + "bbox": [ + 63, + 432, + 129, + 497 + ], + "lines": [ + { + "bbox": [ + 63, + 432, + 129, + 497 + ], + "spans": [ + { + "bbox": [ + 63, + 432, + 129, + 497 + ], + "type": "image", + "image_path": "9f6e83b02c0987ef038de89e275634cb4e7f54fe40cef10f3db875a56c8ef233.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 86, + 498, + 104, + 506 + ], + "lines": [ + { + "bbox": [ + 86, + 498, + 104, + 506 + ], + "spans": [ + { + "bbox": [ + 86, + 498, + 104, + 506 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 47, + 509, + 287, + 532 + ], + "lines": [ + { + "bbox": [ + 47, + 509, + 287, + 532 + ], + "spans": [ + { + "bbox": [ + 47, + 509, + 287, + 532 + ], + "type": "text", + "content": "Figure 11. Visual comparison with the network's performance with and w/o optical center symmetry prior." + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 130, + 365, + 196, + 430 + ], + "blocks": [ + { + "bbox": [ + 130, + 365, + 196, + 430 + ], + "lines": [ + { + "bbox": [ + 130, + 365, + 196, + 430 + ], + "spans": [ + { + "bbox": [ + 130, + 365, + 196, + 430 + ], + "type": "image", + "image_path": "5d4cb1c4d224889639b106a9f480c83f4b034bbfd5dd633babfd46e377104522.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 130, + 432, + 196, + 497 + ], + "blocks": [ + { + "bbox": [ + 130, + 432, + 196, + 497 + ], + "lines": [ + { + "bbox": [ + 130, + 432, + 196, + 497 + ], + "spans": [ + { + "bbox": [ + 130, + 432, + 196, + 497 + ], + "type": "image", + "image_path": "71e6c233b8f3e23d10eb766fcc3e24115a711e96484507949edc8265fb27342f.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 149, + 498, + 178, + 506 + ], + "lines": [ + { + "bbox": [ + 149, + 498, + 178, + 506 + ], + "spans": [ + { + "bbox": [ + 149, + 498, + 178, + 506 + ], + "type": "text", + "content": "w/o prior" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 197, + 365, + 263, + 430 + ], + "blocks": [ + { + "bbox": [ + 197, + 365, + 263, + 430 + ], + "lines": [ + { + "bbox": [ + 197, + 365, + 263, + 430 + ], + "spans": [ + { + "bbox": [ + 197, + 365, + 263, + 430 + ], + "type": "image", + "image_path": "4b6147282e274c24cef08102bad81052a62352842084aec46989f5ce7263a4a0.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 197, + 432, + 263, + 497 + ], + "blocks": [ + { + "bbox": [ + 197, + 432, + 263, + 497 + ], + "lines": [ + { + "bbox": [ + 197, + 432, + 263, + 497 + ], + "spans": [ + { + "bbox": [ + 197, + 432, + 263, + 497 + ], + "type": "image", + "image_path": "da016b0a5fb26f9b2d97e0405b93378898ffdf710a4fcb4ef0f6f48a7b705734.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 214, + 498, + 246, + 506 + ], + "lines": [ + { + "bbox": [ + 214, + 498, + 246, + 506 + ], + "spans": [ + { + "bbox": [ + 214, + 498, + 246, + 506 + ], + "type": "text", + "content": "with prior" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "bbox": [ + 47, + 542, + 287, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 542, + 287, + 614 + ], + "spans": [ + { + "bbox": [ + 47, + 542, + 287, + 614 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rec}" + }, + { + "bbox": [ + 47, + 542, + 287, + 614 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 542, + 287, + 614 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{vgg}" + }, + { + "bbox": [ + 47, + 542, + 287, + 614 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 47, + 542, + 287, + 614 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{mask}" + }, + { + "bbox": [ + 47, + 542, + 287, + 614 + ], + "type": "text", + "content": ". We report the results in Table 3. The ablation study of " + }, + { + "bbox": [ + 47, + 542, + 287, + 614 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{flare}" + }, + { + "bbox": [ + 47, + 542, + 287, + 614 + ], + "type": "text", + "content": " proves the significance of bringing supervision for both flare and background images. Besides, when the " + }, + { + "bbox": [ + 47, + 542, + 287, + 614 + ], + "type": "inline_equation", + "content": "L_{mask}" + }, + { + "bbox": [ + 47, + 542, + 287, + 614 + ], + "type": "text", + "content": " is removed, it degrades the reconstruction results in the flare-corrupted regions, which verifies the importance of focusing on flare regions." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 47, + 617, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 617, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 617, + 287, + 714 + ], + "type": "text", + "content": "Prior. To illustrate the performance of introducing the optical symmetry prior by concatenating a prior image, we train these baseline restoration networks with the original input. Since HINet [4] and MPRNet [24] cannot support input and output with different channels, we concatenate a black image with no information and keep other settings the same. The quantitative results are shown in Table 4, and visual comparisons on real-world images are presented in Fig." + } + ] + } + ], + "index": 34 + }, + { + "type": "table", + "bbox": [ + 314, + 310, + 533, + 423 + ], + "blocks": [ + { + "bbox": [ + 305, + 271, + 545, + 304 + ], + "lines": [ + { + "bbox": [ + 305, + 271, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 305, + 271, + 545, + 304 + ], + "type": "text", + "content": "Table 4. Quantitative comparison for different network structures on synthetic test data. ‘†’ indicates that the network is equipped with the proposed prior." + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 314, + 310, + 533, + 423 + ], + "lines": [ + { + "bbox": [ + 314, + 310, + 533, + 423 + ], + "spans": [ + { + "bbox": [ + 314, + 310, + 533, + 423 + ], + "type": "table", + "html": "
NetworkPSNRSSIMLPIPSMasked PSNR
MPRNet [24]43.600.9920.01127.64
Uformer [21]42.570.9890.00927.77
HINet [4]43.780.9910.01227.73
Restormer [23]42.760.9910.01227.27
MPRNet†48.410.9940.00432.09
Uformer†47.470.9910.00331.57
HINet†48.030.9940.00330.88
Restormer†48.110.9940.00431.07
", + "image_path": "af210b137e39c1272c22a7a64ce0c52626f94bf1650260228437b7cddb2dcdf9.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "table_body" + } + ], + "index": 36 + }, + { + "bbox": [ + 306, + 431, + 545, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 431, + 545, + 479 + ], + "spans": [ + { + "bbox": [ + 306, + 431, + 545, + 479 + ], + "type": "text", + "content": "11. It shows that embedding this prior in the network can boost the performance of quantitative results considerably. In Fig. 11, we observe that this prior helps the network to locate and remove the reflective flare precisely." + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 306, + 481, + 378, + 492 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 481, + 378, + 492 + ], + "spans": [ + { + "bbox": [ + 306, + 481, + 378, + 492 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 304, + 495, + 545, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 495, + 545, + 663 + ], + "spans": [ + { + "bbox": [ + 304, + 495, + 545, + 663 + ], + "type": "text", + "content": "We have presented a novel approach to tackle the challenge of removing reflective flares in nighttime photography. Our approach is inspired by the optical center symmetry prior, which suggests that reflective flares and their corresponding light sources are always symmetrical around the optical center of the lens system in smartphone cameras. Based on the observation that the patterns of reflective flares are always similar to the brightest region of the light source, we introduce a dataset named BracketFlare. We further propose a reflective flare removal pipeline that encodes the optical center symmetry prior into the network. Experimental results demonstrate that our proposed methods are effective in removing reflective flares in real-world scenarios and have the potential for wide-ranging applications." + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 305, + 665, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 665, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 665, + 545, + 714 + ], + "type": "text", + "content": "Acknowledgement: This study is supported under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative, as well as cash and in-kind contribution from the industry partner(s)." + } + ] + } + ], + "index": 40 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20790" + } + ] + } + ], + "index": 41 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "text", + "content": "[1] C S Asha, Sooraj Kumar Bhat, Deepa Nayak, and Chaithra Bhat. Auto removal of bright spot from images captured against flashing light source. In ICDC, 2019. 2, 3, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 124, + 287, + 157 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 124, + 287, + 157 + ], + "spans": [ + { + "bbox": [ + 53, + 124, + 287, + 157 + ], + "type": "text", + "content": "[2] Herbert Bay, Andreas Ess, Tinne Tuytelaars, and Luc Van Gool. Speeded-up robust features (surf). Computer Vision and Image Understanding, 110(3):346-359, 2008. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 158, + 287, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 158, + 287, + 191 + ], + "spans": [ + { + "bbox": [ + 53, + 158, + 287, + 191 + ], + "type": "text", + "content": "[3] Floris Chabert. Automated lens flare removal, 2014. Technical Report, Department of Electrical Engineering, Stanford University. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 192, + 287, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 192, + 287, + 224 + ], + "spans": [ + { + "bbox": [ + 53, + 192, + 287, + 224 + ], + "type": "text", + "content": "[4] Liangyu Chen, Xin Lu, Jie Zhang, Xiaojie Chu, and Chengpeng Chen. Hinet: Half instance normalization network for image restoration. In CVPRW, 2021. 2, 6, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 224, + 287, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 224, + 287, + 278 + ], + "spans": [ + { + "bbox": [ + 53, + 224, + 287, + 278 + ], + "type": "text", + "content": "[5] Yuekun Dai, Chongyi Li, Shangchen Zhou, Ruicheng Feng, and Chen Change Loy. Flare7k: A phenomenological nighttime flare removal dataset. In Thirty-sixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2022. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 280, + 287, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 280, + 287, + 323 + ], + "spans": [ + { + "bbox": [ + 53, + 280, + 287, + 323 + ], + "type": "text", + "content": "[6] Ruicheng Feng, Chongyi Li, Huaijin Chen, Shuai Li, Chen Change Loy, and Jinwei Gu. Removing diffraction image artifacts in under-display camera via dynamic skip connection network. In CVPR, 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 324, + 287, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 324, + 287, + 356 + ], + "spans": [ + { + "bbox": [ + 53, + 324, + 287, + 356 + ], + "type": "text", + "content": "[7] Matthias Hullin, Elmar Eisemann, Hans-Peter Seidel, and Sungkil Lee. Physically-based real-time lens flare rendering. In ACM SIGGRAPH, 2011. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 357, + 287, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 357, + 287, + 390 + ], + "spans": [ + { + "bbox": [ + 53, + 357, + 287, + 390 + ], + "type": "text", + "content": "[8] Yeying Jin, Wenhan Yang, and Robby T Tan. Unsupervised night image enhancement: When layer decomposition meets light-effects suppression. In ECCV. Springer, 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 391, + 287, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 391, + 287, + 413 + ], + "spans": [ + { + "bbox": [ + 53, + 391, + 287, + 413 + ], + "type": "text", + "content": "[9] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 414, + 287, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 414, + 287, + 445 + ], + "spans": [ + { + "bbox": [ + 48, + 414, + 287, + 445 + ], + "type": "text", + "content": "[10] Sungkil Lee and Elmar Eisemann. Practical real-time lensflare rendering. Computer Graphics Forum, 32(4):1-6, 2013. 1, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 447, + 287, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 447, + 287, + 478 + ], + "spans": [ + { + "bbox": [ + 48, + 447, + 287, + 478 + ], + "type": "text", + "content": "[11] Yu Li, Robby T. Tan, and Michael S. Brown. Nighttime haze removal with glow and multiple light colors. In ICCV, 2015. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 480, + 287, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 480, + 287, + 512 + ], + "spans": [ + { + "bbox": [ + 48, + 480, + 287, + 512 + ], + "type": "text", + "content": "[12] David G Lowe. Distinctive image features from scale-invariant keypoints. International Journal of Computer Vision, 60(2):91-110, 2004. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 514, + 287, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 287, + 545 + ], + "type": "text", + "content": "[13] Andreas Nussberger, Helmut Grabner, and Luc Van Gool. Robust aerial object tracking in images with lens flare. In ICRA, 2015. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 547, + 287, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 547, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 48, + 547, + 287, + 568 + ], + "type": "text", + "content": "[14] Frank L Pedrotti, Leno M Pedrotti, and Leno S Pedrotti. Introduction to Optics. Cambridge University Press, 2017. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 570, + 287, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 570, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 570, + 287, + 601 + ], + "type": "text", + "content": "[15] Xiaotian Qiao, Gerhard P. Hancke, and Rynson W. H. Lau. Light source guided single-image flare removal from unpaired data. In ICCV, 2021. 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 602, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 287, + 635 + ], + "type": "text", + "content": "[16] Aashish Sharma and Robby T. Tan. Nighttime visibility enhancement by increasing the dynamic range and suppression of light effects. In CVPR, 2021. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 636, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 287, + 678 + ], + "type": "text", + "content": "[17] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In International Conference on Learning Representations, 2014. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[18] Qilin Sun, Ethan Tseng, Qiang Fu, Wolfgang Heidrich, and Felix Heide. Learning rank-1 diffractive optics for single-shot high dynamic range imaging. In CVPR, 2020. 2" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 546, + 431 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 546, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 546, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 546, + 106 + ], + "type": "text", + "content": "[19] Patricia Vitoria and Coloma Ballester. Automatic flare spot artifact detection and removal in photographs. Journal of Mathematical Imaging and Vision, 61:515-533, 2019. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "text", + "content": "[20] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "type": "text", + "content": "[21] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restorationn. In CVPR, 2022. 2, 6, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 197, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 239 + ], + "type": "text", + "content": "[22] Yicheng Wu, Qiurui He, Tianfan Xue, Rahul Garg, Jiawen Chen, Ashok Veeraraghavan, and Jonathan T Barron. How to train neural networks for flare removal. In ICCV, 2021. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 241, + 545, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 241, + 545, + 284 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 545, + 284 + ], + "type": "text", + "content": "[23] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In CVPR, 2022. 2, 6, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 286, + 545, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 286, + 545, + 329 + ], + "spans": [ + { + "bbox": [ + 307, + 286, + 545, + 329 + ], + "type": "text", + "content": "[24] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Multi-stage progressive image restoration. In CVPR, 2021. 2, 6, 7, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 331, + 545, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 331, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 307, + 331, + 545, + 364 + ], + "type": "text", + "content": "[25] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 365, + 545, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 545, + 397 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 545, + 397 + ], + "type": "text", + "content": "[26] Xuaner Zhang, Ren Ng, and Qifeng Chen. Single image reflection separation with perceptual losses. In CVPR, 2018. 3, 6, 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 399, + 545, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 545, + 431 + ], + "type": "text", + "content": "[27] Junyan Zhu, Taesung Park, Phillip Isola, and Alexei A. Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In ICCV, 2017. 3" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20791" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/No One Left Behind_ Improving the Worst Categories in Long-Tailed Learning/aa8d6d83-fa42-452c-97e5-1d8dfd5a2154_content_list.json b/2023/No One Left Behind_ Improving the Worst Categories in Long-Tailed Learning/aa8d6d83-fa42-452c-97e5-1d8dfd5a2154_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6c383439cad560dbc82b2b24833d4cdc9a076a88 --- /dev/null +++ b/2023/No One Left Behind_ Improving the Worst Categories in Long-Tailed Learning/aa8d6d83-fa42-452c-97e5-1d8dfd5a2154_content_list.json @@ -0,0 +1,1732 @@ +[ + { + "type": "text", + "text": "No One Left Behind: Improving the Worst Categories in Long-Tailed Learning", + "text_level": 1, + "bbox": [ + 86, + 130, + 880, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yingxiao Du, Jianxin Wu* \nState Key Laboratory for Novel Software Technology, Nanjing University \nNanjing, China, 210023", + "bbox": [ + 192, + 180, + 776, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "duyx@lamda.nju.edu.cn, wujx2001@gmail.com", + "bbox": [ + 299, + 237, + 663, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 285, + 313, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Unlike the case when using a balanced training dataset, the per-class recall (i.e., accuracy) of neural networks trained with an imbalanced dataset are known to vary a lot from category to category. The convention in long-tailed recognition is to manually split all categories into three subsets and report the average accuracy within each subset. We argue that under such an evaluation setting, some categories are inevitably sacrificed. On one hand, focusing on the average accuracy on a balanced test set incurs little penalty even if some worst performing categories have zero accuracy. On the other hand, classes in the \"Few\" subset do not necessarily perform worse than those in the \"Many\" or \"Medium\" subsets. We therefore advocate to focus more on improving the lowest recall among all categories and the harmonic mean of all recall values. Specifically, we propose a simple plug-in method that is applicable to a wide range of methods. By simply re-training the classifier of an existing pre-trained model with our proposed loss function and using an optional ensemble trick that combines the predictions of the two classifiers, we achieve a more uniform distribution of recall values across categories, which leads to a higher harmonic mean accuracy while the (arithmetic) average accuracy is still high. The effectiveness of our method is justified on widely used benchmark datasets.", + "bbox": [ + 76, + 316, + 473, + 681 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 708, + 209, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Various gaps exist when adapting image recognition techniques that are developed in the lab to industrial applications. The most noteworthy one is perhaps the difference between training datasets. Most training datasets used in academic research [6, 16] are balanced with respect to the number of images per class. This should not be taken for granted because datasets used in real-world applications are more likely to be imbalanced. Training deep models", + "bbox": [ + 76, + 733, + 468, + 854 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/97852020da1ac5e4e32d437286bfa6f07d9af53cd05137c767a046d87c3a7cf6.jpg", + "image_caption": [ + "Figure 1. The per-class recall of models trained on the imbalanced CIFAR100 (with imbalance ratio 100). Per-class recall value varies a lot from category to category. Moreover, it is not necessarily true that all categories in the \"Few\" subset have lower accuracy than those in the \"Many\" or \"Medium\" subsets." + ], + "image_footnote": [], + "bbox": [ + 514, + 306, + 890, + 484 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/445666365eb1a53ab7d7b987078d69ee5f9d5b8ffcde149a0e9d9cc739e09372.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodMean AccuracyLowest Recall
BSCE [22]42.243.00
DiVE [12]45.112.00
MiSLAS [30]47.055.00
RIDE [26]48.642.00
PaCo [4]51.245.00
", + "bbox": [ + 535, + 582, + 857, + 683 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Table 1. The lowest per-class recall of various state-of-the-art methods on the imbalanced CIFAR100 (with imbalance ratio 100). Although there are rapid improvements over the mean accuracy, the lowest per-class recall remains very low.", + "bbox": [ + 498, + 693, + 893, + 750 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "on these datasets is not trivial as models are known to perform poorly on such datasets. Long-tailed recognition is a research field that aims at tackling this challenge.", + "bbox": [ + 496, + 777, + 890, + 824 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "By plotting the per-class recall (i.e., per-class accuracy) of the model trained on the imbalanced CIFAR00 dataset (with imbalance ratio 100) in Fig. 1, we find that the recall varies dramatically from category to category. We argue that in real world applications, all categories are equally", + "bbox": [ + 496, + 824, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*J. Wu is the corresponding author. This research was partly supported by the National Natural Science Foundation of China under Grant 62276123 and Grant 61921006.", + "bbox": [ + 75, + 862, + 468, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "15804", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "important and no one should be left behind. We also find that despite the rapid developments in this community, no obvious improvement over the lowest per-class recall is witnessed in the past few years [4, 12, 22, 26, 30], as is shown in Tab. 1. The convention in long-tailed recognition research is to split the classes into three subsets based on the number of training images. The accuracy within each subset is often reported along with the overall accuracy. While this evaluation scheme seems reasonable at the first glance, we argue it is potentially problematic. First of all, computing the average recall within each subset is way too coarse, making it impossible to reflect whether some classes are completely sacrificed but covered up by other \"easy\" tail classes. What's more, we find it is not necessarily true that classes in the \"Few\" category all have lower recall than classes in the other two subsets, as is shown in Fig. 1. Therefore, focusing on improving the mean accuracy alone is not enough, especially in real-world applications.", + "bbox": [ + 75, + 90, + 472, + 363 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose a novel method to make sure that no category is left behind. Specifically, we argue that although mean accuracy is widely used in image classification as an optimization objective, due to the fact that different classes have very different recall in long-tailed recognition, it is not the most suitable objective as it incurs little penalty even if some categories have very small per-class accuracy values (e.g., close to 0). Hence, to improve even the worst-performing categories, we believe the harmonic mean of per-class recall would be a better objective. Since harmonic mean is very sensitive to small numbers, further improving classes that already have high recall brings little benefits to the harmonic mean, which makes sure no single class will be left behind. Also, now all classes are treated equally, forcing us to improve classes that have low recall no matter which subset it belongs to.", + "bbox": [ + 75, + 363, + 472, + 604 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, it is difficult to directly minimize the harmonic mean. We therefore propose a novel loss function that maximizes the geometric mean instead, which can be viewed as a surrogate. Our method serves as a simple plug-in that can be used together with both baseline and various state-of-the-art methods. We also propose an ensemble trick that uses the pre-trained and fine-tuned models together to make predictions during inference with nearly no extra cost. We are able to yield a more uniform distribution of recall across categories (i.e., no one left behind), which achieves higher harmonic mean of the recall while the (arithmetic) mean accuracy remains high, too.", + "bbox": [ + 75, + 604, + 472, + 787 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our work has the following contributions:", + "bbox": [ + 96, + 787, + 457, + 801 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We are the first to emphasize the importance of the correct recognition of all categories in long-tailed recognition.", + "- We propose a novel method that aims at increasing the harmonic mean of per-class recall as well as an ensem" + ], + "bbox": [ + 94, + 813, + 468, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ble trick that combines two existing models together during inference with nearly no extra cost.", + "bbox": [ + 529, + 90, + 890, + 121 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We experimented on three widely used benchmark datasets, which justify the effectiveness of our method in terms of both overall and worst per-class accuracy.", + "bbox": [ + 517, + 133, + 893, + 180 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 194, + 640, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Long-tailed recognition is a research field that aims at training models using imbalanced datasets. Since it is widely believed that the learning procedure is dominated by head classes, most existing works focus explicitly on improving the recognition of tail classes.", + "bbox": [ + 496, + 220, + 893, + 297 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Re-sampling and Re-weighting", + "text_level": 1, + "bbox": [ + 498, + 308, + 774, + 324 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Re-sampling and re-weighting are two classical approaches. Re-sampling methods aim at balancing the training set by either over-sampling the tail classes [3, 9, 21, 24] or under-sampling the head classes [7, 10]. Some methods also transfer statistics from the major classes to minor classes to obtain a balanced dataset [15, 19]. One disadvantage of re-sampling methods is that it may lead to either over-fitting of tail classes or under-fitting of head classes. Re-weighting methods, on the other hand, give each instance a weight based on its true label when computing the loss [2, 22, 23]. The major drawback of re-weighting is that it makes the loss function hard to optimize.", + "bbox": [ + 496, + 332, + 893, + 513 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Two-Stage Decoupling", + "text_level": 1, + "bbox": [ + 498, + 523, + 709, + 540 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Due to the shortcomings of re-sampling and reweighting, various other methods are proposed recently. For example, two-stage methods like [14, 18, 30] propose to decouple the learning of features and classifiers and achieve impressive results. Our method is similar to them in the sense that we also re-train a classifier given an existing model. But our motivation is to design a simple and flexible method that can be applied to a wide variety of existing models, and we propose a novel loss function, while two-stage methods like cRT [14] usually use the conventional cross entropy loss in the fine-tuning stage.", + "bbox": [ + 496, + 547, + 893, + 715 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.3. Hybrid and Multiple Heads", + "text_level": 1, + "bbox": [ + 500, + 724, + 750, + 742 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Hybrid methods, like [26, 31], make use of multiple heads to improve the recognition of different classes, which is similar to our proposed ensemble trick. But these methods require a joint training of multiple heads, together with a complex routing module that dynamically determines the head to use during inference, which increases the complexity of the model and makes it hard to use in practice. Our method, on the other hand, is simpler and requires no extra training nor complex modules to combine the predictions, which is different from them.", + "bbox": [ + 496, + 750, + 893, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "15805", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.4. Miscellaneous", + "text_level": 1, + "bbox": [ + 76, + 90, + 222, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "There are many approaches that use knowledge distillation [12, 18] and contrastive learning [4, 17, 25, 28, 33] and achieve descent results. There are also methods that perform logit adjustment [20, 29], which formulate long-tailed recognition as a label distribution shift problem. Recently, there is also one work that tries to solve the problem by performing some regularization on the classifier's weight [1]. They also visualize the recall of each class in their work and find the model is biased towards common classes, which is similar to our motivation. However, they do not emphasize the importance of correctly recognizing all classes, which is different from us.", + "bbox": [ + 75, + 114, + 472, + 297 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 314, + 168, + 329 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We will present our method in this section. First, we will introduce the notations used. Then we will describe the major framework of our method, which includes our proposed loss function and a simple ensemble trick.", + "bbox": [ + 75, + 340, + 470, + 401 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Notations", + "text_level": 1, + "bbox": [ + 76, + 414, + 187, + 428 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The neural network can be represented as a non-linear function $\\mathcal{F}$ . Given the $i$ -th training image $(x^i, y^i)$ from the imbalanced dataset that has $N$ images, where $x^i \\in \\mathbb{R}^{H \\times W \\times D}$ , a forward propagation through the network yields $\\operatorname{logit} o^i = \\mathcal{F}(x^i)$ . For a multi-class classification problem with $C$ categories, a softmax activation function is usually applied and the result can be represented as $\\tilde{p}^i = \\operatorname{softmax}(o^i)$ , where $\\tilde{p}^i \\in \\mathbb{R}^C$ .", + "bbox": [ + 76, + 438, + 468, + 560 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. A Simple Plug-In Method", + "text_level": 1, + "bbox": [ + 76, + 571, + 310, + 589 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Over the past few years, various methods [4,14,22] have been proposed to tackle the challenges in long-tailed recognition. Although our evaluation results in Tab. 1 reveal that no obvious improvements have been made regarding the worst-performing category, many of their design choices are critical and effective in long-tailed recognition. Therefore, in order to make full use of existing advances, we do not aim at building up a whole new training framework, but rather propose a simple plug-in method that can be applied to a wide range of existing methods.", + "bbox": [ + 75, + 597, + 468, + 747 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our method is simple. Given a model trained using either one of existing methods, we re-initialize the last FC layer and re-train it using our proposed loss function while the backbone is frozen. We borrow this idea from many two-stage decoupling methods [14, 30], which is effective and flexible, making our method widely applicable. However, unlike normal decoupling methods that use the conventional cross entropy loss in the second stage, we propose a novel loss function to improve the worst-performing categories. Our proposed loss function, called GML (Geometric", + "bbox": [ + 75, + 750, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Mean Loss), is defined as", + "bbox": [ + 500, + 90, + 671, + 104 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {G M L}} = - \\frac {1}{C} \\sum_ {c = 1} ^ {C} \\log \\bar {\\boldsymbol {p}} _ {c}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 112, + 890, + 152 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\bar{\\pmb{p}}_c$ is computed as", + "bbox": [ + 500, + 157, + 669, + 174 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {\\boldsymbol {p}} _ {c} = \\frac {1}{N _ {c}} \\sum_ {i = 1} ^ {N _ {c}} \\tilde {\\boldsymbol {p}} _ {y ^ {i}} ^ {i}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 633, + 179, + 890, + 219 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "which is the average of $\\tilde{\\pmb{p}}_{y^i}^i$ across all training samples belonging to class $c$ in this mini-batch, and $N_{c}$ is the number of examples from this category in the mini-batch. Inspired by [22], we perform re-weighting when we compute $\\tilde{\\pmb{p}}^i$ :", + "bbox": [ + 496, + 227, + 893, + 291 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\boldsymbol {p}} _ {j} ^ {i} = \\frac {N _ {j} \\exp \\left(o _ {j} ^ {i}\\right)}{\\sum_ {c = 1} ^ {C} N _ {c} \\exp \\left(o _ {c} ^ {i}\\right)}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 295, + 890, + 335 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Why GML Works?", + "text_level": 1, + "bbox": [ + 500, + 339, + 684, + 356 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The definition of mean accuracy is", + "bbox": [ + 519, + 363, + 751, + 378 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {A c c} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\mathbb {I} (y ^ {i} = \\arg \\max \\tilde {\\boldsymbol {p}} ^ {i}), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 385, + 890, + 425 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbb{1}(\\cdot)$ is the indicator function. On a balanced test set, it is the same as the arithmetic mean of per-class recall. We argue that maximizing the mean accuracy inevitably ignores some categories because arithmetic mean incurs little punishment to categories that have very small recall values. On the other hand, harmonic mean, defined as", + "bbox": [ + 496, + 431, + 893, + 521 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {H M} \\left(x _ {1}, \\dots , x _ {n}\\right) = \\frac {n}{\\frac {1}{x _ {1}} + \\cdots + \\frac {1}{x _ {n}}} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 525, + 890, + 560 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "incurs strong punishment to small values. Therefore, we believe it can be a better alternative to optimize for. However, harmonic mean is defined using reciprocal, making it hard and numeric unstable to be optimized. To this extent, we propose to maximize the geometric mean of per-class recall, defined as", + "bbox": [ + 496, + 564, + 890, + 654 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {G M} \\left(x _ {1}, \\dots , x _ {n}\\right) = \\sqrt [ n ]{\\left| x _ {1} \\times \\dots \\times x _ {n} \\right|}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 660, + 890, + 678 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As an illustration, $\\frac{0.01 + 0.99}{2} = 0.5$ , $\\mathrm{HM}(0.01, 0.99) = 0.02$ and $\\mathrm{GM}(0.01, 0.99) = 0.10$ . Both harmonic and geometric means are heavily affected by the small value (0.01), while the arithmetic mean is much less sensitive to it.", + "bbox": [ + 496, + 683, + 890, + 744 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Specifically, given the per-class recall as $r_1, \\ldots, r_C$ , we transform their geometric mean using a simple logarithm transformation:", + "bbox": [ + 496, + 744, + 890, + 789 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\sqrt {r _ {1} \\dots r _ {C}} = \\exp \\left(\\log \\left(r _ {1} \\dots r _ {C}\\right) ^ {1 / C}\\right) (7) \\\\ = \\exp \\left(\\frac {1}{C} \\log \\left(r _ {1} \\dots r _ {C}\\right)\\right) (8) \\\\ = \\exp \\left(\\frac {1}{C} \\sum_ {i = 1} ^ {C} \\log r _ {i}\\right). (9) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 792, + 890, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "15806", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f099466616702b441e44f2378abea8fe708b6035da967be1caa41402c7ffedb0.jpg", + "image_caption": [ + "Figure 2. Bar plot of per-class recall on the imbalanced CIFAR10 (with imbalance ratio 100) before ('CE') and after ('CE+GML') the fine-tuning. For the fine-tuned model, the recall of the first 5 classes dropped while the recall of the latter 5 classes increased, motivating us to combine the prediction of both models." + ], + "image_footnote": [], + "bbox": [ + 102, + 90, + 444, + 270 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since recall is computed by computing the mean of a series of indicator functions and these indicator functions only take value 0 or 1 and is non-differentiable, we use $\\bar{\\pmb{p}}_c$ as a surrogate for $r_c$ during training, resulting in", + "bbox": [ + 75, + 383, + 468, + 444 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\sqrt {r _ {1} \\dots r _ {C}} \\propto \\exp (- \\mathcal {L} _ {\\mathrm {G M L}}). \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 457, + 468, + 474 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "So, by minimizing $\\mathcal{L}_{\\mathrm{GML}}$ , we are effectively maximizing the geometric mean of per-class recall, thus being able to improve on the worst-performing categories.", + "bbox": [ + 76, + 489, + 468, + 535 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Combining the Strength of Both Worlds", + "text_level": 1, + "bbox": [ + 76, + 547, + 419, + 565 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "By fine-tuning the pre-trained model with our proposed loss function, indeed we are able to obtain a more uniform distribution of recall values across categories. As is shown in Fig. 2, the recall of classes that perform worse before are improved while the recall of classes that perform well before are dropped. Motivated by these dynamics, we then propose a simple ensemble trick that aims at combining the strength of both worlds.", + "bbox": [ + 75, + 571, + 468, + 691 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our idea is simple again. Since we re-initialize the classifier during fine-tuning and the backbone is kept frozen, the old classifier is still usable during inference. Let these two models be $\\mathcal{F}_{\\mathrm{old}}$ and $\\mathcal{F}_{\\mathrm{new}}$ , respectively, we then make predictions by", + "bbox": [ + 75, + 694, + 468, + 770 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\boldsymbol {p}} _ {\\text {e n s e m b l e d}} = \\frac {\\tilde {\\boldsymbol {p}} _ {\\text {n e w}} + \\tilde {\\boldsymbol {p}} _ {\\text {o l d}}}{2}. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 782, + 468, + 811 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In practice, since the calibration of some models can be poor [8,30] (e.g., the baseline using cross entropy), we need to manually calibrate the prediction before combining them. For simplicity, we choose temperature scaling to calibrate the model [8] and use two temperature variables $t_{\\mathrm{new}}$ and", + "bbox": [ + 75, + 824, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$t_{\\mathrm{old}}$ when computing $\\tilde{\\pmb{p}}_{\\mathrm{new}}$ and $\\tilde{\\pmb{p}}_{\\mathrm{old}}$ separately:", + "bbox": [ + 498, + 90, + 807, + 107 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\boldsymbol {p}} _ {\\text {n e w}} = \\operatorname {s o f t m a x} \\left(\\frac {\\boldsymbol {o} _ {\\text {n e w}}}{t _ {\\text {n e w}}}\\right), \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 611, + 117, + 890, + 151 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\boldsymbol {p}} _ {\\text {o l d}} = \\operatorname {s o f t m a x} \\left(\\frac {\\boldsymbol {o} _ {\\text {o l d}}}{t _ {\\text {o l d}}}\\right). \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 154, + 890, + 186 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "One big advantage of our proposed ensemble trick is that it brings nearly no extra cost. Unlike many multi-head [31] or multi-expert [26] methods, we do not need extra training as we simply use the classifier that comes with the pretrained model. During inference, the only extra cost is the fully connected layer to compute $\\tilde{\\pmb{p}}_{\\mathrm{old}}$ , which is negligible. Also, no complex routing rules are needed.", + "bbox": [ + 496, + 198, + 890, + 303 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The two temperature variables are the only two hyperparameters introduced by our method, and they can be easily tuned as no training is required. Plus, as we will later show in the ablation studies, most values work perfectly fine so actually no painful tuning is needed. What's more, this ensemble trick is completely optional, our method works well enough even without it.", + "bbox": [ + 496, + 304, + 890, + 409 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.5. Training Pipeline", + "text_level": 1, + "bbox": [ + 500, + 420, + 669, + 436 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our method is very similar to the two-stage decoupling methods [14, 30]. We design it in this way on purpose because we want it to be as flexible and simple as possible. During experiments, we found that our method is orthogonal to many design choices in state-of-the-art methods and is able to improve the worse-case performance when applied to various different pre-trained models. The overall training pipeline of our method consists of three stages:", + "bbox": [ + 496, + 444, + 890, + 566 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Pre-training stage: Obtain the pre-trained model from scratch using any one of the existing methods.", + "2. Fine-tuning stage: Freeze the backbone and re-train the classifier using our proposed loss function.", + "3. An optional ensemble stage: Calibrate the prediction of two classifiers and combine them additively." + ], + "bbox": [ + 511, + 580, + 890, + 694 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The overall training procedure is summarized in Algorithm 1.", + "bbox": [ + 500, + 708, + 890, + 737 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 753, + 630, + 771 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In order to validate the effectiveness of our method, we have conducted experiments on three widely used benchmark datasets. In this section, we will present our experimental results. First, we will introduce the datasets used. Then we will describe various implementation details and evaluation metrics used. After that, we will compare our method with various other methods. Finally, we will present the results of some ablation studies.", + "bbox": [ + 496, + 779, + 890, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "15807", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/36ad8283f9997de5bbf19f30b2284556c08c3ed1116249dafa37f0705c2c405c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetNumber of Classes# Training Images# Test ImagesImbalance Ratio
CIFAR100-LT [16]10010,84710,000100
ImageNet-LT [6, 19]1,000115,84650,000256
Places-LT [32]36562,50036,500996
", + "bbox": [ + 176, + 88, + 795, + 160 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 2. Statistics of three imbalanced datasets used in our experiments.", + "bbox": [ + 269, + 171, + 696, + 185 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Algorithm 1 The overall training procedure", + "text_level": 1, + "bbox": [ + 78, + 210, + 369, + 226 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Input: Training images $x$ and their labels $y$ .", + "bbox": [ + 80, + 231, + 372, + 244 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: Randomly initialize the network and train it from scratch to obtain the pre-trained model.", + "2: Use the pre-trained model to perform initialization.", + "3: Freeze the backbone and re-initialize the classifier.", + "4: Fine-tune the model using the loss function defined as Eq. (1) for a few epochs.", + "5: During inference, combine the results from both classifiers as defined in Eq. (11)." + ], + "bbox": [ + 86, + 247, + 467, + 364 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 76, + 398, + 179, + 412 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use three widely used long-tailed image recognition datasets: CIFAR100-LT, ImageNet-LT and PlacesLT. Some statistics about these three datasets are summarized in Tab. 2. Since CIFAR100 [16], ImageNet [6] and Places [32] are all balanced datasets, we follow previous works [19, 31] to down-sample their original training set. Details about the construction process can be found in the supplementary material.", + "bbox": [ + 75, + 422, + 468, + 542 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 76, + 554, + 294, + 569 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following previous works [19, 31], we use ResNet-32 [11], ResNeXt-50 [27] and ResNet-152 as the backbone network for CIFAR100-LT, ImageNet-LT and Places-LT, respectively. For the pre-training stage, if there exists released checkpoints, we use them directly. If not, we use the officially released codes and try to reproduce the results in our best effort without changing their original settings. This includes stuffs like hyper-parameters and data augmentations used. For the fine-tuning stage, detailed training settings can be found in the supplementary material.", + "bbox": [ + 75, + 578, + 468, + 729 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Evaluation Metrics and Comparison Methods", + "text_level": 1, + "bbox": [ + 76, + 739, + 465, + 757 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The convention in long-tailed recognition research is to split the classes into three subsets based on the number of training images [14, 19]. Typically, \"Head\" denotes classes that have more than 100 images, \"Few\" denotes classes that have less than 20 images and all other classes are denoted as \"Medium\". The accuracy within each subset is often reported along with the overall accuracy to justify the claim that the performance improvement mainly comes from tail classes. And according to our observation in Sec. 1, we ar", + "bbox": [ + 75, + 763, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "gue that such an evaluation scheme can be problematic and we propose to first compute the recall within each category and then compute their harmonic mean. In our experiments, since our method optimizes geometric mean of per-class recall as a surrogate, we also report the geometric mean. Besides, we report the overall mean accuracy as well.", + "bbox": [ + 496, + 212, + 890, + 301 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We compare our methods with both baseline methods like simple cross entropy loss and various state-of-the-arts methods [4, 12, 22, 26]. Many of these methods have very different training settings and strictly speaking, they are not directly comparable. But since our method is a simple plugin method, the main focus is to quantify the improvements brought by our method compared to its respective baseline result.", + "bbox": [ + 496, + 301, + 890, + 421 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Main Comparisons", + "text_level": 1, + "bbox": [ + 498, + 431, + 683, + 448 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We validate our method on three benchmark datasets and present the results on each dataset here separately.", + "bbox": [ + 498, + 455, + 890, + 484 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4.1 CIFAR100-LT", + "text_level": 1, + "bbox": [ + 500, + 503, + 653, + 516 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Tab. 3 shows the experimental results on the CIFAR100-LT dataset. CIFAR100-LT has three different versions with different imbalance ratios and here we only consider the one with imbalance ratio 100. As we can see from the table, although various different methods have been proposed in recent years and there are rapid improvements on the overall accuracy, the performance of the worst-performing category is still relatively poor. The geometric and harmonic mean of per-class recall also justify our claim. We apply our method to both the baseline method CE and state-of-the-art method PaCo [4] on this dataset. When our method is applied to CE and PaCo, we are able to improve the lowest recall from 0.00 and 5.00 to 6.00 and 9.00 respectively. And if we further apply our proposed ensemble trick, we are able to improve the lowest recall to 15.00, achieving state-of-the-art performance in both geometric mean and harmonic mean while keeping the overall accuracy roughly unchanged.", + "bbox": [ + 496, + 526, + 890, + 782 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "It is worth noting that even the simplest CE+GML outperforms all existing methods in terms of lowest recall, that is, in terms of \"no one left behind\".", + "bbox": [ + 496, + 784, + 890, + 827 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4.2 ImageNet-LT and Places-LT", + "text_level": 1, + "bbox": [ + 500, + 845, + 751, + 861 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Tab. 4 shows the result on ImageNet-LT. The overall observation is similar to that on CIFAR100-LT, except that", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "15808", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/63dc89b835167d8a01e3953089353bcec45bb94620c711dbeaf657141585c94c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsAccuracyGeometric MeanHarmonic MeanLowest Recall
CE38.7421.03*2.05*0.00
BSCE [22]42.2435.1626.243.00
LDAM [2]43.5133.6121.493.00
LADE [13]44.3939.0532.585.00
DiVE [12]45.1137.0825.182.00
MiSLAS [30]47.0540.1630.935.00
RIDE [26]48.6438.7123.862.00
PaCo [4]51.2445.2936.425.00
CE + GML41.0636.5931.266.00
PaCo + GML50.5345.4739.209.00
PaCo + GML (Ensemble)49.8245.7041.0215.00
", + "bbox": [ + 191, + 88, + 781, + 282 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/f0ed7253bcae5d238b8c0fe59b55495a0634b20a79e3f962206162af5056e076.jpg", + "table_caption": [ + "Table 3. Results on the CIFAR100-LT dataset with imbalance ratio 100. Numbers with * are computed by substituting zero elements with a small number $(10^{-3})$ or else the geometric and harmonic mean will all be zero." + ], + "table_footnote": [], + "table_body": "
MethodsAccuracyGeometric MeanHarmonic MeanLowest Recall
CE43.9023.25*1.25*0.00
BSCE [22]50.4842.32*13.74*0.00
cRT [14]49.6441.35*13.82*0.00
DiVE [12]53.6345.49*12.76*0.00
RIDE [26]55.6947.56*17.32*0.00
PaCo [4]58.5351.32*21.79*0.00
CE + GML45.6139.8231.672.00
PaCo + GML55.5750.6743.712.00
PaCo + GML (Ensemble)57.2252.2944.802.00
", + "bbox": [ + 189, + 335, + 779, + 503 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 4. Results on the ImageNet-LT dataset. Numbers with * are computed by substituting zero elements with a small number $(10^{-3})$ or else the geometric and harmonic mean will all be zero.", + "bbox": [ + 75, + 515, + 893, + 544 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "since this dataset is larger and harder, the lowest recall of all current methods are zero, making the harmonic mean of per-class recall very low, even when a small number substitutes zero per-class recalls. Our method, on the other hand, greatly improves the harmonic mean by successfully improving the lowest recall value. All categories have per-class recall higher than 0 when our method is used, even in the simplest CE+GML.", + "bbox": [ + 75, + 570, + 470, + 690 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Tab. 5 shows the result on Places-LT. Current state-of-the-art on this dataset is MiSLAS [30], so we apply our method upon it and the simple baseline CE. Again, our method successfully improves the lowest recall, as well as the harmonic mean of per-class recall. And when the ensemble trick is further applied, we are able to achieve new state-of-the-art results.", + "bbox": [ + 75, + 691, + 470, + 797 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4.3 Universality", + "text_level": 1, + "bbox": [ + 76, + 816, + 217, + 832 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Since we want our method to be a universal plug-in, in this subsection we present the results of the experiment of adding our method on top of various different methods. The experiment is conducted on CIFAR100-LT, and the re", + "bbox": [ + 75, + 839, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "sult is shown in Tab. 6. Note that here we do not perform any ensemble. As we can see, our method is applicable to various different methods, and is able to achieve consistent improvements. However, we do observed that there exists some methods, like DiVE [12], where we failed to achieve obvious improvements. More details about this can be found when we discuss limitations in Sec. 5.", + "bbox": [ + 496, + 570, + 893, + 676 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.5. Ablation Studies", + "text_level": 1, + "bbox": [ + 500, + 684, + 665, + 699 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For ablation studies, if not otherwise specified, all experiments are conducted on CIFAR100-LT (with imbalance ratio 100) under the default training setting.", + "bbox": [ + 496, + 707, + 893, + 755 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.5.1 Effects of Two Temperature Variables in the Ensemble Stage", + "text_level": 1, + "bbox": [ + 496, + 771, + 890, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "When we design our method, we'd like to keep the number of new hyper-parameters as less as possible in order for it to be readily applicable in real-world applications. The only two hyper-parameters introduced by our method are in the optional ensemble stage. Here we study the effects of choosing different values of them. Specifically, we change", + "bbox": [ + 496, + 809, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "15809", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/0af05d479702aa7007b5fb5a0b2db52d53958f555d184536983fb6d936fd1f23.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsAccuracyGeometric MeanHarmonic MeanLowest Recall
CE28.7112.11*0.73*0.00
BSCE [22]37.1829.30*5.64*0.00
PaCo [4]40.4527.88*2.53*0.00
MiSLAS [30]40.4835.5328.973.00
CE + GML36.8229.30*8.00*0.00
MiSLAS + GML39.9035.4129.985.00
MiSLAS + GML (Ensemble)40.3735.9230.515.00
", + "bbox": [ + 179, + 88, + 790, + 229 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f8be096050c70dd74ff4f111237a61b406c566ff698e2b6cce26ffccba210869.jpg", + "table_caption": [ + "Table 5. Results on the Places-LT dataset. Numbers with * are computed by substituting zero elements with a small number $(10^{-3})$ or else the geometric and harmonic mean will all be zero. Note that to plug GML into MiSLAS, since MiSLAS is also a two-stage method, we think it's fairer to use GML together with their proposed label aware smoothing loss." + ], + "table_footnote": [], + "table_body": "
MethodsG-MeanH-MeanLowest Recall
CE21.03*2.05*0.00
CE + GML36.59 (15.56↑)31.26 (29.21↑)6.00 (6.00↑)
BSCE35.1626.243.00
BSCE + GML36.52 (1.36↑)30.88 (4.64↑)7.00 (4.00↑)
MiSLAS40.1630.935.00
MiSLAS + GML40.90 (0.74↑)36.49 (5.56↑)11.00 (6.00↑)
PaCo45.2936.425.00
PaCo + GML45.47 (0.18↑)39.20 (2.78↑)9.00 (4.00↑)
", + "bbox": [ + 81, + 305, + 467, + 444 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/0e4d66c3c185358a04d65da7d9281707fe092b95d00793976269f0e50f093310.jpg", + "table_caption": [ + "Table 6. Results on the CIFAR100-LT dataset with imbalance ratio 100. Our proposed method is applicable to various methods. \"G-Mean\" is short for \"Geometric Mean\" and \"H-Mean\" is short for \"Harmonic Mean\"." + ], + "table_footnote": [], + "table_body": "
toldtnewGeometric MeanHarmonic MeanLowest Recall
1145.7041.0215.00
1245.5838.868.00
1344.6836.095.00
2144.3940.2215.00
2245.6141.0115.00
2346.0640.4011.00
3143.7339.6315.00
", + "bbox": [ + 81, + 529, + 467, + 645 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the value of the temperature variables shown in Eq. (12) and Eq. (13) in the ensemble stage when we apply our method to PaCo [4]. The results are shown in Tab. 7. In our experiments on CIFAR100-LT, we set $t_{\\mathrm{old}} = 1$ and $t_{\\mathrm{new}} = 1$ . And as we can see from the table, the value of $t_{\\mathrm{old}}$ does not affect the result much while $t_{\\mathrm{new}}$ is somewhat important, especially for improving the lowest recall—higher $t_{\\mathrm{new}}$ will hurt it. This is natural because higher $t$ results in a smoother distribution and less information is preserved. Since the newly trained classifier is crucial for improving the lowest recall, we need to preserve more information from it.", + "bbox": [ + 75, + 715, + 468, + 882 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Generally speaking, since these two hyper-parameters", + "bbox": [ + 96, + 885, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/75bb34486dddd91ca2e99ea40ff118d28606a067d8a7d0ed620973de64316e2c.jpg", + "table_caption": [ + "Table 7. Ablation on the effects of temperature variables when we apply our ensemble method upon PaCo [4]." + ], + "table_footnote": [], + "table_body": "
MethodsG-MeanH-MeanLowest Recall
CE + GML36.5931.266.00
(w/o re-weighting)25.455.110.00
(w/o re-weighting, re-sampling)32.3223.054.00
", + "bbox": [ + 504, + 305, + 888, + 371 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2e35b23c511286227c97672c49256cb1cd2c36a2e2f6a3425435f4552f360ab5.jpg", + "table_caption": [ + "Table 8. Ablation on the re-weighting design in GML." + ], + "table_footnote": [], + "table_body": "
MethodsG-MeanH-MeanLowest Recall
CE + GML36.5931.266.00
PaCo + GML45.4739.209.00
GML (train from scratch)36.6330.869.00
", + "bbox": [ + 504, + 411, + 888, + 483 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 9. Training from scratch using GML.", + "bbox": [ + 565, + 492, + 823, + 507 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "are introduced for calibration, for different methods, we may need to tune them respectively because the calibrations of different methods are known to vary a lot in long-tailed recognition problems [30]. But usually this will not cause a headache, because the ensemble stage does not involve any training and thus can be tuned very quickly.", + "bbox": [ + 496, + 535, + 890, + 628 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5.2 Re-weighting in GML", + "text_level": 1, + "bbox": [ + 498, + 647, + 707, + 662 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Eq. (3), re-weighting is applied in our proposed GML, here we ablate this design. The results are shown in Tab. 8. As we can see, the performance will be relatively poor without re-weighting. And just like many two-stage decoupling methods [14], re-sampling during fine-tuning can help, but is inferior to re-weighting.", + "bbox": [ + 496, + 672, + 890, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5.3 Training from Scratch using GML", + "text_level": 1, + "bbox": [ + 498, + 784, + 795, + 800 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Since our main purpose is to propose a simple plug-in method that is widely applicable to various existing methods, we use GML only to re-train the classifier of an existing model. However, GML can also be used to train a model from scratch. The results are shown in Tab. 9. As we can see, the result is comparable to the result when GML", + "bbox": [ + 496, + 810, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "15810", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0c9f97b185c9700a78f472c479867b19bd9e22c76213909cd608fc7f7d58bc69.jpg", + "image_caption": [ + "Figure 3. Bar plot of per-class recall on the imbalanced CIFAR100 (with imbalance ratio 100) before and after the fine-tuning. No ensemble is performed. Our method focuses on improving the recognition of classes that have low recall values, making sure no category is left behind." + ], + "image_footnote": [], + "bbox": [ + 102, + 92, + 444, + 267 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "is attached to CE, but inferior to the result when GML is applied to SOTA methods like PaCo, just as expected.", + "bbox": [ + 75, + 380, + 468, + 412 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5.4 Visualization of the Per-Class Recall", + "text_level": 1, + "bbox": [ + 76, + 430, + 383, + 444 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Fig. 1, we visualize the per-class recall of model trained on the imbalanced dataset, and find they vary a lot from class to class. Since our major motivation is to make sure all classes are equally treated so that no category will be left behind, here we visualize the change of per-class recall after applying our method. We perform visualizing using two datasets, CIFAR10-LT and CIFAR100-LT, both have imbalance ratio 100. The visualization results on CIFAR10-LT and CIFAR100-LT are shown in Fig. 2 and Fig. 3, respectively. Note that here we do not perform the ensemble. As we can clearly see in both figures, our fine-tuned model has a more uniform distribution of per-class recall. These two figures justify the effectiveness of our proposed method.", + "bbox": [ + 75, + 453, + 468, + 650 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Limitations and Broader Impacts", + "text_level": 1, + "bbox": [ + 76, + 664, + 382, + 681 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Although our method is applicable to a wide range of existing methods, we do find that when applied to some methods, the improvement is less obvious compared to others. For example, our framework does not cope well with methods like BSCE [22] and DiVE [12]. Interestingly, although DiVE is based on knowledge distillation, BSCE is used in DiVE as well to train both the teacher and student model so we attribute the failure to the fact that these method all re-weight the loss function. Two-stage decoupling methods like cRT [14] are also known to not perform well when the first stage training involves re-weighting or re-sampling. The author argues it is because re-sampling and re-weighting have negative effects on feature learning and we believe that argument applies to our method as well.", + "bbox": [ + 75, + 688, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Another thing worth noting here is that we do not report the experimental results on iNaturalist2018 [5] because we find it generally not suitable in our settings. iNaturalist2018 is another widely used benchmark dataset in long-tailed recognition. It is large-scale, consisting of 437.5K training images from 8142 classes. But for the test split, it only has 24K images, so each class only has 3 images. Since our major motivation is to make sure no category is left behind, it is generally not meaningful to look at the per-class recall in iNaturalist2018 because the variance of the result can be very large due to the very small number of test images. However, in real-world applications, it is generally rare to have such small number of test images.", + "bbox": [ + 496, + 90, + 890, + 287 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For broader impacts, we believe our research have positive impacts on reducing the bias of model training on an imbalanced dataset. Machine learning models are known to be biased if there is lack of training images from some categories. We believe our work will be helpful to achieve better fairness by making sure all categories are equally treated.", + "bbox": [ + 496, + 290, + 890, + 381 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusions", + "text_level": 1, + "bbox": [ + 500, + 404, + 625, + 419 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we focused on improving some of the worst-performing classes in long-tailed recognition. We found that when trained with an imbalanced dataset, the per-class recall of the model varies a lot from class to class. Previous convention reports the arithmetic mean of per-class recall, but we argued such an evaluation scheme can be problematic. First, arithmetic mean only incurs little penalty to small numbers, making it too coarse to reflect whether there are categories that are left behind (i.e., with close to zero per-class accuracy). Second, it is not necessarily true that classes in the \"Few\" category perform worse than those in \"Many\" or \"Medium\", so focusing on improving \"Few\" accuracy alone is not enough. For these reasons, we argued that we should pay more attention to the harmonic mean of per-class recall and we therefore proposed a novel method that aims at improving the lowest recall and the harmonic mean of recalls. Our method is a simple plug-in that is applicable to a wide range of existing methods. Specifically, our method consists of three stages, in the first stage, we use any existing method to train a model from scratch. Then we re-train the classifier using our proposed novel GML loss function. Finally, we propose a simple ensemble trick that can be used to combine the predictions from the two classifiers with nearly no extra cost. We validated the effectiveness of our method on three widely used benchmark datasets, and witnessed consistent improvements on the harmonic mean of recalls and lowest recall value, while the overall accuracy still remains high. By visualizing the distribution of per-class recall values of the fine-tuned model, we found our model indeed achieved a more balanced distribution.", + "bbox": [ + 496, + 431, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "15811", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Shaden Alshammari, Yu-Xiong Wang, Deva Ramanan, and Shu Kong. Long-tailed recognition via weight balancing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6897-6907, June 2022. 3", + "[2] Kaidi Cao, Colin Wei, Adrien Gaidon, Nikos Arechiga, and Tengyu Ma. Learning imbalanced datasets with label-distribution-aware margin loss. In Advances in Neural Information Processing Systems 32, pages 1567-1578, December 2019. 2, 6", + "[3] Nitesh V Chawla, Kevin W Bowyer, Lawrence O Hall, and W Philip Kegelmeyer. SMOTE: Synthetic minority oversampling technique. Journal of Artificial Intelligence Research, 16:321-357, 2002. 2", + "[4] Jiequan Cui, Zhisheng Zhong, Shu Liu, Bei Yu, and Jiaya Jia. Parametric contrastive learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 715-724, October 2021. 1, 2, 3, 5, 6, 7", + "[5] Yin Cui, Yang Song, Chen Sun, Andrew Howard, and Serge Belongie. Large scale fine-grained categorization and domain-specific transfer learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4109-4118, June 2018. 8", + "[6] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248-255, June 2009. 1, 5", + "[7] Chris Drummond and Robert C. Holte. C4.5, class imbalance, and cost sensitivity: Why under-sampling beats oversampling. In Workshop on Learning from Imbalanced Data Sets II, 2003. 2", + "[8] Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q. Weinberger. On calibration of modern neural networks. In Proceedings of the 34th International Conference on Machine Learning, pages 1321-1330, August 2017. 4", + "[9] Hui Han, Wen-Yuan Wang, and Bing-Huan Mao. Borderline-SMOTE: A new over-sampling method in imbalanced data sets learning. In Advances in Intelligent Computing, pages 878-887, August 2005. 2", + "[10] Haibo He and Edwardo A. Garcia. Learning from imbalanced data. IEEE Transactions on Knowledge and Data Engineering, 21(9):1263-1284, 2009. 2", + "[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, June 2016. 5", + "[12] Yin-Yin He, Jianxin Wu, and Xiu-Shen Wei. Distilling virtual examples for long-tailed recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 235-244, October 2021. 1, 2, 3, 5, 6, 8", + "[13] Youngkyu Hong, Seungju Han, Kwanghee Choi, Seokjun Seo, Beomsu Kim, and Buru Chang. Disentangling label distribution for long-tailed visual recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6626-6636, June 2021. 6" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Bingyi Kang, Saining Xie, Marcus Rohrbach, Zhicheng Yan, Albert Gordo, Jiashi Feng, and Yannis Kalantidis. Decoupling representation and classifier for long-tailed recognition. In International Conference on Learning Representations, April 2020. 2, 3, 4, 5, 6, 7, 8", + "[15] Jaehyung Kim, Jongheon Jeong, and Jinwoo Shin. M2m: Imbalanced classification via major-to-minor translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13896-13905, June 2020. 2", + "[16] Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, University of Toronto, 2009. https://www.cs.toronto.edu/~kriz/learning-features-2009-TR.pdf.1,5", + "[17] Tianhong Li, Peng Cao, Yuan Yuan, Lijie Fan, Yuzhe Yang, Rogerio S. Feris, Piotr Indyk, and Dina Katabi. Targeted supervised contrastive learning for long-tailed recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6918-6928, June 2022. 3", + "[18] Tianhao Li, Limin Wang, and Gangshan Wu. Self supervision to distillation for long-tailed visual recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 630-639, October 2021. 2, 3", + "[19] Ziwei Liu, Zhongqi Miao, Xiaohang Zhan, Jiayun Wang, Boqing Gong, and Stella X. Yu. Large-scale long-tailed recognition in an open world. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2537-2546, June 2019. 2, 5", + "[20] Aditya Krishna Menon, Sadeep Jayasumana, Ankit Singh Rawat, Himanshu Jain, Andreas Veit, and Sanjiv Kumar. Long-tail learning via logit adjustment. In International Conference on Learning Representations, May 2021. 3", + "[21] Seulki Park, Youngkyu Hong, Byeongho Heo, Sangdoo Yun, and Jin Young Choi. The majority can help the minority: Context-rich minority oversampling for long-tailed classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6887-6896, June 2022. 2", + "[22] Jiawei Ren, Cunjun Yu, Shunan Sheng, Xiao Ma, Haiyu Zhao, Shuai Yi, and Hongsheng Li. Balanced meta-softmax for long-tailed visual recognition. In Advances in Neural Information Processing Systems 33, pages 4175–4186, December 2020. 1, 2, 3, 5, 6, 7, 8", + "[23] Dvir Samuel and Gal Chechik. Distributional robustness loss for long-tail learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9495-9504, October 2021. 2", + "[24] Li Shen, Zhouchen Lin, and Qingming Huang. Relay backpropagation for effective learning of deep convolutional neural networks. In Proceedings of the European Conference on Computer Vision, pages 467-482, October 2016. 2", + "[25] Peng Wang, Kai Han, Xiu-Shen Wei, Lei Zhang, and Lei Wang. Contrastive learning based hybrid networks for long-tailed image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 943-952, June 2021. 3" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "15812", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Xudong Wang, Long Lian, Zhongqi Miao, Ziwei Liu, and Stella Yu. Long-tailed recognition by routing diverse distribution-aware experts. In International Conference on Learning Representations, May 2021. 1, 2, 4, 5, 6", + "[27] Saining Xie, Ross Girshick, Piotr Dollar, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1492-1500, July 2017. 5", + "[28] Yuzhe Yang and Zhi Xu. Rethinking the value of labels for improving class-imbalanced learning. In Advances in Neural Information Processing Systems 33, pages 19290-19301, December 2020. 3", + "[29] Songyang Zhang, Zeming Li, Shipeng Yan, Xuming He, and Jian Sun. Distribution alignment: A unified framework for long-tail visual recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2361-2370, June 2021. 3", + "[30] Zhisheng Zhong, Jiequan Cui, Shu Liu, and Jiaya Jia. Improving calibration for long-tailed recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16489-16498, June 2021. 1, 2, 3, 4, 6, 7", + "[31] Boyan Zhou, Quan Cui, Xiu-Shen Wei, and Zhao-Min Chen. BBN: Bilateral-branch network with cumulative learning for long-tailed visual recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9719–9728, June 2020. 2, 4, 5", + "[32] Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 40(6):1452-1464, 2018. 5", + "[33] Jianggang Zhu, Zheng Wang, Jingjing Chen, Yi-Ping Phoebe Chen, and Yu-Gang Jiang. Balanced contrastive learning for long-tailed visual recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6908-6917, June 2022. 3" + ], + "bbox": [ + 78, + 90, + 470, + 612 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "15813", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/No One Left Behind_ Improving the Worst Categories in Long-Tailed Learning/aa8d6d83-fa42-452c-97e5-1d8dfd5a2154_model.json b/2023/No One Left Behind_ Improving the Worst Categories in Long-Tailed Learning/aa8d6d83-fa42-452c-97e5-1d8dfd5a2154_model.json new file mode 100644 index 0000000000000000000000000000000000000000..531307887d10dc1fd3cb5cc51d07076e8a21226b --- /dev/null +++ b/2023/No One Left Behind_ Improving the Worst Categories in Long-Tailed Learning/aa8d6d83-fa42-452c-97e5-1d8dfd5a2154_model.json @@ -0,0 +1,2167 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.088, + 0.131, + 0.882, + 0.153 + ], + "angle": 0, + "content": "No One Left Behind: Improving the Worst Categories in Long-Tailed Learning" + }, + { + "type": "text", + "bbox": [ + 0.193, + 0.181, + 0.777, + 0.234 + ], + "angle": 0, + "content": "Yingxiao Du, Jianxin Wu* \nState Key Laboratory for Novel Software Technology, Nanjing University \nNanjing, China, 210023" + }, + { + "type": "text", + "bbox": [ + 0.3, + 0.238, + 0.665, + 0.251 + ], + "angle": 0, + "content": "duyx@lamda.nju.edu.cn, wujx2001@gmail.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.286, + 0.314, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.318, + 0.474, + 0.683 + ], + "angle": 0, + "content": "Unlike the case when using a balanced training dataset, the per-class recall (i.e., accuracy) of neural networks trained with an imbalanced dataset are known to vary a lot from category to category. The convention in long-tailed recognition is to manually split all categories into three subsets and report the average accuracy within each subset. We argue that under such an evaluation setting, some categories are inevitably sacrificed. On one hand, focusing on the average accuracy on a balanced test set incurs little penalty even if some worst performing categories have zero accuracy. On the other hand, classes in the \"Few\" subset do not necessarily perform worse than those in the \"Many\" or \"Medium\" subsets. We therefore advocate to focus more on improving the lowest recall among all categories and the harmonic mean of all recall values. Specifically, we propose a simple plug-in method that is applicable to a wide range of methods. By simply re-training the classifier of an existing pre-trained model with our proposed loss function and using an optional ensemble trick that combines the predictions of the two classifiers, we achieve a more uniform distribution of recall values across categories, which leads to a higher harmonic mean accuracy while the (arithmetic) average accuracy is still high. The effectiveness of our method is justified on widely used benchmark datasets." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.709, + 0.21, + 0.725 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.734, + 0.47, + 0.856 + ], + "angle": 0, + "content": "Various gaps exist when adapting image recognition techniques that are developed in the lab to industrial applications. The most noteworthy one is perhaps the difference between training datasets. Most training datasets used in academic research [6, 16] are balanced with respect to the number of images per class. This should not be taken for granted because datasets used in real-world applications are more likely to be imbalanced. Training deep models" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.308, + 0.891, + 0.485 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.497, + 0.895, + 0.568 + ], + "angle": 0, + "content": "Figure 1. The per-class recall of models trained on the imbalanced CIFAR100 (with imbalance ratio 100). Per-class recall value varies a lot from category to category. Moreover, it is not necessarily true that all categories in the \"Few\" subset have lower accuracy than those in the \"Many\" or \"Medium\" subsets." + }, + { + "type": "table", + "bbox": [ + 0.536, + 0.583, + 0.858, + 0.684 + ], + "angle": 0, + "content": "
MethodMean AccuracyLowest Recall
BSCE [22]42.243.00
DiVE [12]45.112.00
MiSLAS [30]47.055.00
RIDE [26]48.642.00
PaCo [4]51.245.00
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.694, + 0.895, + 0.751 + ], + "angle": 0, + "content": "Table 1. The lowest per-class recall of various state-of-the-art methods on the imbalanced CIFAR100 (with imbalance ratio 100). Although there are rapid improvements over the mean accuracy, the lowest per-class recall remains very low." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.779, + 0.892, + 0.825 + ], + "angle": 0, + "content": "on these datasets is not trivial as models are known to perform poorly on such datasets. Long-tailed recognition is a research field that aims at tackling this challenge." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.895, + 0.903 + ], + "angle": 0, + "content": "By plotting the per-class recall (i.e., per-class accuracy) of the model trained on the imbalanced CIFAR00 dataset (with imbalance ratio 100) in Fig. 1, we find that the recall varies dramatically from category to category. We argue that in real world applications, all categories are equally" + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.863, + 0.47, + 0.9 + ], + "angle": 0, + "content": "*J. Wu is the corresponding author. This research was partly supported by the National Natural Science Foundation of China under Grant 62276123 and Grant 61921006." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15804" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.364 + ], + "angle": 0, + "content": "important and no one should be left behind. We also find that despite the rapid developments in this community, no obvious improvement over the lowest per-class recall is witnessed in the past few years [4, 12, 22, 26, 30], as is shown in Tab. 1. The convention in long-tailed recognition research is to split the classes into three subsets based on the number of training images. The accuracy within each subset is often reported along with the overall accuracy. While this evaluation scheme seems reasonable at the first glance, we argue it is potentially problematic. First of all, computing the average recall within each subset is way too coarse, making it impossible to reflect whether some classes are completely sacrificed but covered up by other \"easy\" tail classes. What's more, we find it is not necessarily true that classes in the \"Few\" category all have lower recall than classes in the other two subsets, as is shown in Fig. 1. Therefore, focusing on improving the mean accuracy alone is not enough, especially in real-world applications." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.364, + 0.473, + 0.606 + ], + "angle": 0, + "content": "In this paper, we propose a novel method to make sure that no category is left behind. Specifically, we argue that although mean accuracy is widely used in image classification as an optimization objective, due to the fact that different classes have very different recall in long-tailed recognition, it is not the most suitable objective as it incurs little penalty even if some categories have very small per-class accuracy values (e.g., close to 0). Hence, to improve even the worst-performing categories, we believe the harmonic mean of per-class recall would be a better objective. Since harmonic mean is very sensitive to small numbers, further improving classes that already have high recall brings little benefits to the harmonic mean, which makes sure no single class will be left behind. Also, now all classes are treated equally, forcing us to improve classes that have low recall no matter which subset it belongs to." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.606, + 0.473, + 0.788 + ], + "angle": 0, + "content": "However, it is difficult to directly minimize the harmonic mean. We therefore propose a novel loss function that maximizes the geometric mean instead, which can be viewed as a surrogate. Our method serves as a simple plug-in that can be used together with both baseline and various state-of-the-art methods. We also propose an ensemble trick that uses the pre-trained and fine-tuned models together to make predictions during inference with nearly no extra cost. We are able to yield a more uniform distribution of recall across categories (i.e., no one left behind), which achieves higher harmonic mean of the recall while the (arithmetic) mean accuracy remains high, too." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.788, + 0.458, + 0.803 + ], + "angle": 0, + "content": "In summary, our work has the following contributions:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.814, + 0.47, + 0.858 + ], + "angle": 0, + "content": "- We are the first to emphasize the importance of the correct recognition of all categories in long-tailed recognition." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.871, + 0.47, + 0.902 + ], + "angle": 0, + "content": "- We propose a novel method that aims at increasing the harmonic mean of per-class recall as well as an ensem" + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.814, + 0.47, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.531, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "ble trick that combines two existing models together during inference with nearly no extra cost." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.135, + 0.894, + 0.181 + ], + "angle": 0, + "content": "- We experimented on three widely used benchmark datasets, which justify the effectiveness of our method in terms of both overall and worst per-class accuracy." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.195, + 0.642, + 0.212 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.222, + 0.894, + 0.298 + ], + "angle": 0, + "content": "Long-tailed recognition is a research field that aims at training models using imbalanced datasets. Since it is widely believed that the learning procedure is dominated by head classes, most existing works focus explicitly on improving the recognition of tail classes." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.309, + 0.776, + 0.325 + ], + "angle": 0, + "content": "2.1. Re-sampling and Re-weighting" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.333, + 0.895, + 0.514 + ], + "angle": 0, + "content": "Re-sampling and re-weighting are two classical approaches. Re-sampling methods aim at balancing the training set by either over-sampling the tail classes [3, 9, 21, 24] or under-sampling the head classes [7, 10]. Some methods also transfer statistics from the major classes to minor classes to obtain a balanced dataset [15, 19]. One disadvantage of re-sampling methods is that it may lead to either over-fitting of tail classes or under-fitting of head classes. Re-weighting methods, on the other hand, give each instance a weight based on its true label when computing the loss [2, 22, 23]. The major drawback of re-weighting is that it makes the loss function hard to optimize." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.525, + 0.71, + 0.541 + ], + "angle": 0, + "content": "2.2. Two-Stage Decoupling" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.549, + 0.894, + 0.716 + ], + "angle": 0, + "content": "Due to the shortcomings of re-sampling and reweighting, various other methods are proposed recently. For example, two-stage methods like [14, 18, 30] propose to decouple the learning of features and classifiers and achieve impressive results. Our method is similar to them in the sense that we also re-train a classifier given an existing model. But our motivation is to design a simple and flexible method that can be applied to a wide variety of existing models, and we propose a novel loss function, while two-stage methods like cRT [14] usually use the conventional cross entropy loss in the fine-tuning stage." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.726, + 0.75, + 0.743 + ], + "angle": 0, + "content": "2.3. Hybrid and Multiple Heads" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Hybrid methods, like [26, 31], make use of multiple heads to improve the recognition of different classes, which is similar to our proposed ensemble trick. But these methods require a joint training of multiple heads, together with a complex routing module that dynamically determines the head to use during inference, which increases the complexity of the model and makes it hard to use in practice. Our method, on the other hand, is simpler and requires no extra training nor complex modules to combine the predictions, which is different from them." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15805" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.091, + 0.223, + 0.106 + ], + "angle": 0, + "content": "2.4. Miscellaneous" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.115, + 0.473, + 0.298 + ], + "angle": 0, + "content": "There are many approaches that use knowledge distillation [12, 18] and contrastive learning [4, 17, 25, 28, 33] and achieve descent results. There are also methods that perform logit adjustment [20, 29], which formulate long-tailed recognition as a label distribution shift problem. Recently, there is also one work that tries to solve the problem by performing some regularization on the classifier's weight [1]. They also visualize the recall of each class in their work and find the model is biased towards common classes, which is similar to our motivation. However, they do not emphasize the importance of correctly recognizing all classes, which is different from us." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.315, + 0.169, + 0.33 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.341, + 0.471, + 0.402 + ], + "angle": 0, + "content": "We will present our method in this section. First, we will introduce the notations used. Then we will describe the major framework of our method, which includes our proposed loss function and a simple ensemble trick." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.415, + 0.188, + 0.429 + ], + "angle": 0, + "content": "3.1. Notations" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.439, + 0.47, + 0.561 + ], + "angle": 0, + "content": "The neural network can be represented as a non-linear function \\(\\mathcal{F}\\). Given the \\(i\\)-th training image \\((x^i, y^i)\\) from the imbalanced dataset that has \\(N\\) images, where \\(x^i \\in \\mathbb{R}^{H \\times W \\times D}\\), a forward propagation through the network yields \\(\\operatorname{logit} o^i = \\mathcal{F}(x^i)\\). For a multi-class classification problem with \\(C\\) categories, a softmax activation function is usually applied and the result can be represented as \\(\\tilde{p}^i = \\operatorname{softmax}(o^i)\\), where \\(\\tilde{p}^i \\in \\mathbb{R}^C\\)." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.573, + 0.312, + 0.59 + ], + "angle": 0, + "content": "3.2. A Simple Plug-In Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.598, + 0.47, + 0.748 + ], + "angle": 0, + "content": "Over the past few years, various methods [4,14,22] have been proposed to tackle the challenges in long-tailed recognition. Although our evaluation results in Tab. 1 reveal that no obvious improvements have been made regarding the worst-performing category, many of their design choices are critical and effective in long-tailed recognition. Therefore, in order to make full use of existing advances, we do not aim at building up a whole new training framework, but rather propose a simple plug-in method that can be applied to a wide range of existing methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Our method is simple. Given a model trained using either one of existing methods, we re-initialize the last FC layer and re-train it using our proposed loss function while the backbone is frozen. We borrow this idea from many two-stage decoupling methods [14, 30], which is effective and flexible, making our method widely applicable. However, unlike normal decoupling methods that use the conventional cross entropy loss in the second stage, we propose a novel loss function to improve the worst-performing categories. Our proposed loss function, called GML (Geometric" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.673, + 0.106 + ], + "angle": 0, + "content": "Mean Loss), is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.613, + 0.113, + 0.892, + 0.153 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {G M L}} = - \\frac {1}{C} \\sum_ {c = 1} ^ {C} \\log \\bar {\\boldsymbol {p}} _ {c}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.159, + 0.671, + 0.175 + ], + "angle": 0, + "content": "where \\(\\bar{\\pmb{p}}_c\\) is computed as" + }, + { + "type": "equation", + "bbox": [ + 0.634, + 0.18, + 0.892, + 0.22 + ], + "angle": 0, + "content": "\\[\n\\bar {\\boldsymbol {p}} _ {c} = \\frac {1}{N _ {c}} \\sum_ {i = 1} ^ {N _ {c}} \\tilde {\\boldsymbol {p}} _ {y ^ {i}} ^ {i}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.228, + 0.894, + 0.292 + ], + "angle": 0, + "content": "which is the average of \\(\\tilde{\\pmb{p}}_{y^i}^i\\) across all training samples belonging to class \\(c\\) in this mini-batch, and \\(N_{c}\\) is the number of examples from this category in the mini-batch. Inspired by [22], we perform re-weighting when we compute \\(\\tilde{\\pmb{p}}^i\\):" + }, + { + "type": "equation", + "bbox": [ + 0.614, + 0.296, + 0.892, + 0.337 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\boldsymbol {p}} _ {j} ^ {i} = \\frac {N _ {j} \\exp \\left(o _ {j} ^ {i}\\right)}{\\sum_ {c = 1} ^ {C} N _ {c} \\exp \\left(o _ {c} ^ {i}\\right)}. \\tag {3}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.34, + 0.686, + 0.357 + ], + "angle": 0, + "content": "3.3. Why GML Works?" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.364, + 0.752, + 0.38 + ], + "angle": 0, + "content": "The definition of mean accuracy is" + }, + { + "type": "equation", + "bbox": [ + 0.583, + 0.386, + 0.892, + 0.426 + ], + "angle": 0, + "content": "\\[\n\\operatorname {A c c} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\mathbb {I} (y ^ {i} = \\arg \\max \\tilde {\\boldsymbol {p}} ^ {i}), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.432, + 0.894, + 0.522 + ], + "angle": 0, + "content": "where \\(\\mathbb{1}(\\cdot)\\) is the indicator function. On a balanced test set, it is the same as the arithmetic mean of per-class recall. We argue that maximizing the mean accuracy inevitably ignores some categories because arithmetic mean incurs little punishment to categories that have very small recall values. On the other hand, harmonic mean, defined as" + }, + { + "type": "equation", + "bbox": [ + 0.583, + 0.526, + 0.892, + 0.561 + ], + "angle": 0, + "content": "\\[\n\\mathrm {H M} \\left(x _ {1}, \\dots , x _ {n}\\right) = \\frac {n}{\\frac {1}{x _ {1}} + \\cdots + \\frac {1}{x _ {n}}} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.565, + 0.892, + 0.655 + ], + "angle": 0, + "content": "incurs strong punishment to small values. Therefore, we believe it can be a better alternative to optimize for. However, harmonic mean is defined using reciprocal, making it hard and numeric unstable to be optimized. To this extent, we propose to maximize the geometric mean of per-class recall, defined as" + }, + { + "type": "equation", + "bbox": [ + 0.574, + 0.661, + 0.892, + 0.679 + ], + "angle": 0, + "content": "\\[\n\\operatorname {G M} \\left(x _ {1}, \\dots , x _ {n}\\right) = \\sqrt [ n ]{\\left| x _ {1} \\times \\dots \\times x _ {n} \\right|}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.684, + 0.892, + 0.745 + ], + "angle": 0, + "content": "As an illustration, \\(\\frac{0.01 + 0.99}{2} = 0.5\\), \\(\\mathrm{HM}(0.01, 0.99) = 0.02\\) and \\(\\mathrm{GM}(0.01, 0.99) = 0.10\\). Both harmonic and geometric means are heavily affected by the small value (0.01), while the arithmetic mean is much less sensitive to it." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.746, + 0.892, + 0.79 + ], + "angle": 0, + "content": "Specifically, given the per-class recall as \\( r_1, \\ldots, r_C \\), we transform their geometric mean using a simple logarithm transformation:" + }, + { + "type": "equation", + "bbox": [ + 0.575, + 0.794, + 0.892, + 0.899 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\sqrt {r _ {1} \\dots r _ {C}} = \\exp \\left(\\log \\left(r _ {1} \\dots r _ {C}\\right) ^ {1 / C}\\right) (7) \\\\ = \\exp \\left(\\frac {1}{C} \\log \\left(r _ {1} \\dots r _ {C}\\right)\\right) (8) \\\\ = \\exp \\left(\\frac {1}{C} \\sum_ {i = 1} ^ {C} \\log r _ {i}\\right). (9) \\\\ \\end{array}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15806" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.103, + 0.092, + 0.445, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.284, + 0.471, + 0.354 + ], + "angle": 0, + "content": "Figure 2. Bar plot of per-class recall on the imbalanced CIFAR10 (with imbalance ratio 100) before ('CE') and after ('CE+GML') the fine-tuning. For the fine-tuned model, the recall of the first 5 classes dropped while the recall of the latter 5 classes increased, motivating us to combine the prediction of both models." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.384, + 0.47, + 0.445 + ], + "angle": 0, + "content": "Since recall is computed by computing the mean of a series of indicator functions and these indicator functions only take value 0 or 1 and is non-differentiable, we use \\(\\bar{\\pmb{p}}_c\\) as a surrogate for \\(r_c\\) during training, resulting in" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.458, + 0.469, + 0.475 + ], + "angle": 0, + "content": "\\[\n\\sqrt {r _ {1} \\dots r _ {C}} \\propto \\exp (- \\mathcal {L} _ {\\mathrm {G M L}}). \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.49, + 0.47, + 0.536 + ], + "angle": 0, + "content": "So, by minimizing \\(\\mathcal{L}_{\\mathrm{GML}}\\), we are effectively maximizing the geometric mean of per-class recall, thus being able to improve on the worst-performing categories." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.548, + 0.42, + 0.566 + ], + "angle": 0, + "content": "3.4. Combining the Strength of Both Worlds" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.573, + 0.469, + 0.693 + ], + "angle": 0, + "content": "By fine-tuning the pre-trained model with our proposed loss function, indeed we are able to obtain a more uniform distribution of recall values across categories. As is shown in Fig. 2, the recall of classes that perform worse before are improved while the recall of classes that perform well before are dropped. Motivated by these dynamics, we then propose a simple ensemble trick that aims at combining the strength of both worlds." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.695, + 0.469, + 0.771 + ], + "angle": 0, + "content": "Our idea is simple again. Since we re-initialize the classifier during fine-tuning and the backbone is kept frozen, the old classifier is still usable during inference. Let these two models be \\(\\mathcal{F}_{\\mathrm{old}}\\) and \\(\\mathcal{F}_{\\mathrm{new}}\\), respectively, we then make predictions by" + }, + { + "type": "equation", + "bbox": [ + 0.186, + 0.783, + 0.469, + 0.813 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\boldsymbol {p}} _ {\\text {e n s e m b l e d}} = \\frac {\\tilde {\\boldsymbol {p}} _ {\\text {n e w}} + \\tilde {\\boldsymbol {p}} _ {\\text {o l d}}}{2}. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.903 + ], + "angle": 0, + "content": "In practice, since the calibration of some models can be poor [8,30] (e.g., the baseline using cross entropy), we need to manually calibrate the prediction before combining them. For simplicity, we choose temperature scaling to calibrate the model [8] and use two temperature variables \\( t_{\\mathrm{new}} \\) and" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.808, + 0.108 + ], + "angle": 0, + "content": "\\(t_{\\mathrm{old}}\\) when computing \\(\\tilde{\\pmb{p}}_{\\mathrm{new}}\\) and \\(\\tilde{\\pmb{p}}_{\\mathrm{old}}\\) separately:" + }, + { + "type": "equation", + "bbox": [ + 0.612, + 0.118, + 0.891, + 0.152 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\boldsymbol {p}} _ {\\text {n e w}} = \\operatorname {s o f t m a x} \\left(\\frac {\\boldsymbol {o} _ {\\text {n e w}}}{t _ {\\text {n e w}}}\\right), \\tag {12}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.616, + 0.155, + 0.891, + 0.188 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\boldsymbol {p}} _ {\\text {o l d}} = \\operatorname {s o f t m a x} \\left(\\frac {\\boldsymbol {o} _ {\\text {o l d}}}{t _ {\\text {o l d}}}\\right). \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.199, + 0.892, + 0.304 + ], + "angle": 0, + "content": "One big advantage of our proposed ensemble trick is that it brings nearly no extra cost. Unlike many multi-head [31] or multi-expert [26] methods, we do not need extra training as we simply use the classifier that comes with the pretrained model. During inference, the only extra cost is the fully connected layer to compute \\(\\tilde{\\pmb{p}}_{\\mathrm{old}}\\), which is negligible. Also, no complex routing rules are needed." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.305, + 0.892, + 0.41 + ], + "angle": 0, + "content": "The two temperature variables are the only two hyperparameters introduced by our method, and they can be easily tuned as no training is required. Plus, as we will later show in the ablation studies, most values work perfectly fine so actually no painful tuning is needed. What's more, this ensemble trick is completely optional, our method works well enough even without it." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.421, + 0.671, + 0.438 + ], + "angle": 0, + "content": "3.5. Training Pipeline" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.445, + 0.892, + 0.567 + ], + "angle": 0, + "content": "Our method is very similar to the two-stage decoupling methods [14, 30]. We design it in this way on purpose because we want it to be as flexible and simple as possible. During experiments, we found that our method is orthogonal to many design choices in state-of-the-art methods and is able to improve the worse-case performance when applied to various different pre-trained models. The overall training pipeline of our method consists of three stages:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.581, + 0.891, + 0.611 + ], + "angle": 0, + "content": "1. Pre-training stage: Obtain the pre-trained model from scratch using any one of the existing methods." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.623, + 0.89, + 0.653 + ], + "angle": 0, + "content": "2. Fine-tuning stage: Freeze the backbone and re-train the classifier using our proposed loss function." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.665, + 0.891, + 0.695 + ], + "angle": 0, + "content": "3. An optional ensemble stage: Calibrate the prediction of two classifiers and combine them additively." + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.581, + 0.891, + 0.695 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.709, + 0.891, + 0.738 + ], + "angle": 0, + "content": "The overall training procedure is summarized in Algorithm 1." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.755, + 0.631, + 0.772 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.9 + ], + "angle": 0, + "content": "In order to validate the effectiveness of our method, we have conducted experiments on three widely used benchmark datasets. In this section, we will present our experimental results. First, we will introduce the datasets used. Then we will describe various implementation details and evaluation metrics used. After that, we will compare our method with various other methods. Finally, we will present the results of some ablation studies." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15807" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.177, + 0.089, + 0.796, + 0.161 + ], + "angle": 0, + "content": "
DatasetNumber of Classes# Training Images# Test ImagesImbalance Ratio
CIFAR100-LT [16]10010,84710,000100
ImageNet-LT [6, 19]1,000115,84650,000256
Places-LT [32]36562,50036,500996
" + }, + { + "type": "table_caption", + "bbox": [ + 0.271, + 0.172, + 0.697, + 0.186 + ], + "angle": 0, + "content": "Table 2. Statistics of three imbalanced datasets used in our experiments." + }, + { + "type": "title", + "bbox": [ + 0.079, + 0.212, + 0.37, + 0.227 + ], + "angle": 0, + "content": "Algorithm 1 The overall training procedure" + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.232, + 0.373, + 0.246 + ], + "angle": 0, + "content": "Input: Training images \\( x \\) and their labels \\( y \\)." + }, + { + "type": "text", + "bbox": [ + 0.087, + 0.248, + 0.468, + 0.275 + ], + "angle": 0, + "content": "1: Randomly initialize the network and train it from scratch to obtain the pre-trained model." + }, + { + "type": "text", + "bbox": [ + 0.087, + 0.278, + 0.443, + 0.291 + ], + "angle": 0, + "content": "2: Use the pre-trained model to perform initialization." + }, + { + "type": "text", + "bbox": [ + 0.087, + 0.293, + 0.44, + 0.305 + ], + "angle": 0, + "content": "3: Freeze the backbone and re-initialize the classifier." + }, + { + "type": "text", + "bbox": [ + 0.087, + 0.308, + 0.468, + 0.336 + ], + "angle": 0, + "content": "4: Fine-tune the model using the loss function defined as Eq. (1) for a few epochs." + }, + { + "type": "text", + "bbox": [ + 0.087, + 0.338, + 0.468, + 0.366 + ], + "angle": 0, + "content": "5: During inference, combine the results from both classifiers as defined in Eq. (11)." + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.248, + 0.468, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.399, + 0.18, + 0.413 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.423, + 0.469, + 0.543 + ], + "angle": 0, + "content": "We use three widely used long-tailed image recognition datasets: CIFAR100-LT, ImageNet-LT and PlacesLT. Some statistics about these three datasets are summarized in Tab. 2. Since CIFAR100 [16], ImageNet [6] and Places [32] are all balanced datasets, we follow previous works [19, 31] to down-sample their original training set. Details about the construction process can be found in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.555, + 0.295, + 0.57 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.579, + 0.469, + 0.73 + ], + "angle": 0, + "content": "Following previous works [19, 31], we use ResNet-32 [11], ResNeXt-50 [27] and ResNet-152 as the backbone network for CIFAR100-LT, ImageNet-LT and Places-LT, respectively. For the pre-training stage, if there exists released checkpoints, we use them directly. If not, we use the officially released codes and try to reproduce the results in our best effort without changing their original settings. This includes stuffs like hyper-parameters and data augmentations used. For the fine-tuning stage, detailed training settings can be found in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.741, + 0.466, + 0.758 + ], + "angle": 0, + "content": "4.3. Evaluation Metrics and Comparison Methods" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.469, + 0.901 + ], + "angle": 0, + "content": "The convention in long-tailed recognition research is to split the classes into three subsets based on the number of training images [14, 19]. Typically, \"Head\" denotes classes that have more than 100 images, \"Few\" denotes classes that have less than 20 images and all other classes are denoted as \"Medium\". The accuracy within each subset is often reported along with the overall accuracy to justify the claim that the performance improvement mainly comes from tail classes. And according to our observation in Sec. 1, we ar" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.213, + 0.892, + 0.303 + ], + "angle": 0, + "content": "gue that such an evaluation scheme can be problematic and we propose to first compute the recall within each category and then compute their harmonic mean. In our experiments, since our method optimizes geometric mean of per-class recall as a surrogate, we also report the geometric mean. Besides, we report the overall mean accuracy as well." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.303, + 0.892, + 0.422 + ], + "angle": 0, + "content": "We compare our methods with both baseline methods like simple cross entropy loss and various state-of-the-arts methods [4, 12, 22, 26]. Many of these methods have very different training settings and strictly speaking, they are not directly comparable. But since our method is a simple plugin method, the main focus is to quantify the improvements brought by our method compared to its respective baseline result." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.433, + 0.684, + 0.449 + ], + "angle": 0, + "content": "4.4. Main Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.456, + 0.892, + 0.486 + ], + "angle": 0, + "content": "We validate our method on three benchmark datasets and present the results on each dataset here separately." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.504, + 0.655, + 0.517 + ], + "angle": 0, + "content": "4.4.1 CIFAR100-LT" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.527, + 0.892, + 0.784 + ], + "angle": 0, + "content": "Tab. 3 shows the experimental results on the CIFAR100-LT dataset. CIFAR100-LT has three different versions with different imbalance ratios and here we only consider the one with imbalance ratio 100. As we can see from the table, although various different methods have been proposed in recent years and there are rapid improvements on the overall accuracy, the performance of the worst-performing category is still relatively poor. The geometric and harmonic mean of per-class recall also justify our claim. We apply our method to both the baseline method CE and state-of-the-art method PaCo [4] on this dataset. When our method is applied to CE and PaCo, we are able to improve the lowest recall from 0.00 and 5.00 to 6.00 and 9.00 respectively. And if we further apply our proposed ensemble trick, we are able to improve the lowest recall to 15.00, achieving state-of-the-art performance in both geometric mean and harmonic mean while keeping the overall accuracy roughly unchanged." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.785, + 0.892, + 0.828 + ], + "angle": 0, + "content": "It is worth noting that even the simplest CE+GML outperforms all existing methods in terms of lowest recall, that is, in terms of \"no one left behind\"." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.847, + 0.753, + 0.862 + ], + "angle": 0, + "content": "4.4.2 ImageNet-LT and Places-LT" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Tab. 4 shows the result on ImageNet-LT. The overall observation is similar to that on CIFAR100-LT, except that" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15808" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.192, + 0.089, + 0.782, + 0.284 + ], + "angle": 0, + "content": "
MethodsAccuracyGeometric MeanHarmonic MeanLowest Recall
CE38.7421.03*2.05*0.00
BSCE [22]42.2435.1626.243.00
LDAM [2]43.5133.6121.493.00
LADE [13]44.3939.0532.585.00
DiVE [12]45.1137.0825.182.00
MiSLAS [30]47.0540.1630.935.00
RIDE [26]48.6438.7123.862.00
PaCo [4]51.2445.2936.425.00
CE + GML41.0636.5931.266.00
PaCo + GML50.5345.4739.209.00
PaCo + GML (Ensemble)49.8245.7041.0215.00
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.295, + 0.895, + 0.325 + ], + "angle": 0, + "content": "Table 3. Results on the CIFAR100-LT dataset with imbalance ratio 100. Numbers with * are computed by substituting zero elements with a small number \\((10^{-3})\\) or else the geometric and harmonic mean will all be zero." + }, + { + "type": "table", + "bbox": [ + 0.191, + 0.336, + 0.781, + 0.505 + ], + "angle": 0, + "content": "
MethodsAccuracyGeometric MeanHarmonic MeanLowest Recall
CE43.9023.25*1.25*0.00
BSCE [22]50.4842.32*13.74*0.00
cRT [14]49.6441.35*13.82*0.00
DiVE [12]53.6345.49*12.76*0.00
RIDE [26]55.6947.56*17.32*0.00
PaCo [4]58.5351.32*21.79*0.00
CE + GML45.6139.8231.672.00
PaCo + GML55.5750.6743.712.00
PaCo + GML (Ensemble)57.2252.2944.802.00
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.516, + 0.895, + 0.545 + ], + "angle": 0, + "content": "Table 4. Results on the ImageNet-LT dataset. Numbers with * are computed by substituting zero elements with a small number \\((10^{-3})\\) or else the geometric and harmonic mean will all be zero." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.571, + 0.472, + 0.691 + ], + "angle": 0, + "content": "since this dataset is larger and harder, the lowest recall of all current methods are zero, making the harmonic mean of per-class recall very low, even when a small number substitutes zero per-class recalls. Our method, on the other hand, greatly improves the harmonic mean by successfully improving the lowest recall value. All categories have per-class recall higher than 0 when our method is used, even in the simplest CE+GML." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.692, + 0.472, + 0.798 + ], + "angle": 0, + "content": "Tab. 5 shows the result on Places-LT. Current state-of-the-art on this dataset is MiSLAS [30], so we apply our method upon it and the simple baseline CE. Again, our method successfully improves the lowest recall, as well as the harmonic mean of per-class recall. And when the ensemble trick is further applied, we are able to achieve new state-of-the-art results." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.817, + 0.218, + 0.833 + ], + "angle": 0, + "content": "4.4.3 Universality" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Since we want our method to be a universal plug-in, in this subsection we present the results of the experiment of adding our method on top of various different methods. The experiment is conducted on CIFAR100-LT, and the re" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.571, + 0.895, + 0.678 + ], + "angle": 0, + "content": "sult is shown in Tab. 6. Note that here we do not perform any ensemble. As we can see, our method is applicable to various different methods, and is able to achieve consistent improvements. However, we do observed that there exists some methods, like DiVE [12], where we failed to achieve obvious improvements. More details about this can be found when we discuss limitations in Sec. 5." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.685, + 0.666, + 0.7 + ], + "angle": 0, + "content": "4.5. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.708, + 0.894, + 0.756 + ], + "angle": 0, + "content": "For ablation studies, if not otherwise specified, all experiments are conducted on CIFAR100-LT (with imbalance ratio 100) under the default training setting." + }, + { + "type": "title", + "bbox": [ + 0.498, + 0.772, + 0.892, + 0.803 + ], + "angle": 0, + "content": "4.5.1 Effects of Two Temperature Variables in the Ensemble Stage" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.895, + 0.903 + ], + "angle": 0, + "content": "When we design our method, we'd like to keep the number of new hyper-parameters as less as possible in order for it to be readily applicable in real-world applications. The only two hyper-parameters introduced by our method are in the optional ensemble stage. Here we study the effects of choosing different values of them. Specifically, we change" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "15809" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.18, + 0.089, + 0.791, + 0.23 + ], + "angle": 0, + "content": "
MethodsAccuracyGeometric MeanHarmonic MeanLowest Recall
CE28.7112.11*0.73*0.00
BSCE [22]37.1829.30*5.64*0.00
PaCo [4]40.4527.88*2.53*0.00
MiSLAS [30]40.4835.5328.973.00
CE + GML36.8229.30*8.00*0.00
MiSLAS + GML39.9035.4129.985.00
MiSLAS + GML (Ensemble)40.3735.9230.515.00
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.24, + 0.895, + 0.285 + ], + "angle": 0, + "content": "Table 5. Results on the Places-LT dataset. Numbers with * are computed by substituting zero elements with a small number \\((10^{-3})\\) or else the geometric and harmonic mean will all be zero. Note that to plug GML into MiSLAS, since MiSLAS is also a two-stage method, we think it's fairer to use GML together with their proposed label aware smoothing loss." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.306, + 0.468, + 0.445 + ], + "angle": 0, + "content": "
MethodsG-MeanH-MeanLowest Recall
CE21.03*2.05*0.00
CE + GML36.59 (15.56↑)31.26 (29.21↑)6.00 (6.00↑)
BSCE35.1626.243.00
BSCE + GML36.52 (1.36↑)30.88 (4.64↑)7.00 (4.00↑)
MiSLAS40.1630.935.00
MiSLAS + GML40.90 (0.74↑)36.49 (5.56↑)11.00 (6.00↑)
PaCo45.2936.425.00
PaCo + GML45.47 (0.18↑)39.20 (2.78↑)9.00 (4.00↑)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.454, + 0.472, + 0.51 + ], + "angle": 0, + "content": "Table 6. Results on the CIFAR100-LT dataset with imbalance ratio 100. Our proposed method is applicable to various methods. \"G-Mean\" is short for \"Geometric Mean\" and \"H-Mean\" is short for \"Harmonic Mean\"." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.53, + 0.468, + 0.646 + ], + "angle": 0, + "content": "
toldtnewGeometric MeanHarmonic MeanLowest Recall
1145.7041.0215.00
1245.5838.868.00
1344.6836.095.00
2144.3940.2215.00
2245.6141.0115.00
2346.0640.4011.00
3143.7339.6315.00
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.656, + 0.47, + 0.687 + ], + "angle": 0, + "content": "Table 7. Ablation on the effects of temperature variables when we apply our ensemble method upon PaCo [4]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.716, + 0.47, + 0.883 + ], + "angle": 0, + "content": "the value of the temperature variables shown in Eq. (12) and Eq. (13) in the ensemble stage when we apply our method to PaCo [4]. The results are shown in Tab. 7. In our experiments on CIFAR100-LT, we set \\( t_{\\mathrm{old}} = 1 \\) and \\( t_{\\mathrm{new}} = 1 \\). And as we can see from the table, the value of \\( t_{\\mathrm{old}} \\) does not affect the result much while \\( t_{\\mathrm{new}} \\) is somewhat important, especially for improving the lowest recall—higher \\( t_{\\mathrm{new}} \\) will hurt it. This is natural because higher \\( t \\) results in a smoother distribution and less information is preserved. Since the newly trained classifier is crucial for improving the lowest recall, we need to preserve more information from it." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Generally speaking, since these two hyper-parameters" + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.306, + 0.89, + 0.372 + ], + "angle": 0, + "content": "
MethodsG-MeanH-MeanLowest Recall
CE + GML36.5931.266.00
(w/o re-weighting)25.455.110.00
(w/o re-weighting, re-sampling)32.3223.054.00
" + }, + { + "type": "table_caption", + "bbox": [ + 0.534, + 0.382, + 0.857, + 0.397 + ], + "angle": 0, + "content": "Table 8. Ablation on the re-weighting design in GML." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.412, + 0.889, + 0.484 + ], + "angle": 0, + "content": "
MethodsG-MeanH-MeanLowest Recall
CE + GML36.5931.266.00
PaCo + GML45.4739.209.00
GML (train from scratch)36.6330.869.00
" + }, + { + "type": "table_caption", + "bbox": [ + 0.566, + 0.493, + 0.825, + 0.508 + ], + "angle": 0, + "content": "Table 9. Training from scratch using GML." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.536, + 0.892, + 0.629 + ], + "angle": 0, + "content": "are introduced for calibration, for different methods, we may need to tune them respectively because the calibrations of different methods are known to vary a lot in long-tailed recognition problems [30]. But usually this will not cause a headache, because the ensemble stage does not involve any training and thus can be tuned very quickly." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.648, + 0.708, + 0.664 + ], + "angle": 0, + "content": "4.5.2 Re-weighting in GML" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.673, + 0.892, + 0.765 + ], + "angle": 0, + "content": "As shown in Eq. (3), re-weighting is applied in our proposed GML, here we ablate this design. The results are shown in Tab. 8. As we can see, the performance will be relatively poor without re-weighting. And just like many two-stage decoupling methods [14], re-sampling during fine-tuning can help, but is inferior to re-weighting." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.785, + 0.796, + 0.801 + ], + "angle": 0, + "content": "4.5.3 Training from Scratch using GML" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Since our main purpose is to propose a simple plug-in method that is widely applicable to various existing methods, we use GML only to re-train the classifier of an existing model. However, GML can also be used to train a model from scratch. The results are shown in Tab. 9. As we can see, the result is comparable to the result when GML" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15810" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.093, + 0.445, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.284, + 0.472, + 0.356 + ], + "angle": 0, + "content": "Figure 3. Bar plot of per-class recall on the imbalanced CIFAR100 (with imbalance ratio 100) before and after the fine-tuning. No ensemble is performed. Our method focuses on improving the recognition of classes that have low recall values, making sure no category is left behind." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.381, + 0.47, + 0.413 + ], + "angle": 0, + "content": "is attached to CE, but inferior to the result when GML is applied to SOTA methods like PaCo, just as expected." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.431, + 0.384, + 0.445 + ], + "angle": 0, + "content": "4.5.4 Visualization of the Per-Class Recall" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.454, + 0.47, + 0.651 + ], + "angle": 0, + "content": "In Fig. 1, we visualize the per-class recall of model trained on the imbalanced dataset, and find they vary a lot from class to class. Since our major motivation is to make sure all classes are equally treated so that no category will be left behind, here we visualize the change of per-class recall after applying our method. We perform visualizing using two datasets, CIFAR10-LT and CIFAR100-LT, both have imbalance ratio 100. The visualization results on CIFAR10-LT and CIFAR100-LT are shown in Fig. 2 and Fig. 3, respectively. Note that here we do not perform the ensemble. As we can clearly see in both figures, our fine-tuned model has a more uniform distribution of per-class recall. These two figures justify the effectiveness of our proposed method." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.665, + 0.383, + 0.683 + ], + "angle": 0, + "content": "5. Limitations and Broader Impacts" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Although our method is applicable to a wide range of existing methods, we do find that when applied to some methods, the improvement is less obvious compared to others. For example, our framework does not cope well with methods like BSCE [22] and DiVE [12]. Interestingly, although DiVE is based on knowledge distillation, BSCE is used in DiVE as well to train both the teacher and student model so we attribute the failure to the fact that these method all re-weight the loss function. Two-stage decoupling methods like cRT [14] are also known to not perform well when the first stage training involves re-weighting or re-sampling. The author argues it is because re-sampling and re-weighting have negative effects on feature learning and we believe that argument applies to our method as well." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.288 + ], + "angle": 0, + "content": "Another thing worth noting here is that we do not report the experimental results on iNaturalist2018 [5] because we find it generally not suitable in our settings. iNaturalist2018 is another widely used benchmark dataset in long-tailed recognition. It is large-scale, consisting of 437.5K training images from 8142 classes. But for the test split, it only has 24K images, so each class only has 3 images. Since our major motivation is to make sure no category is left behind, it is generally not meaningful to look at the per-class recall in iNaturalist2018 because the variance of the result can be very large due to the very small number of test images. However, in real-world applications, it is generally rare to have such small number of test images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.291, + 0.892, + 0.382 + ], + "angle": 0, + "content": "For broader impacts, we believe our research have positive impacts on reducing the bias of model training on an imbalanced dataset. Machine learning models are known to be biased if there is lack of training images from some categories. We believe our work will be helpful to achieve better fairness by making sure all categories are equally treated." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.405, + 0.627, + 0.42 + ], + "angle": 0, + "content": "6. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.433, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In this paper, we focused on improving some of the worst-performing classes in long-tailed recognition. We found that when trained with an imbalanced dataset, the per-class recall of the model varies a lot from class to class. Previous convention reports the arithmetic mean of per-class recall, but we argued such an evaluation scheme can be problematic. First, arithmetic mean only incurs little penalty to small numbers, making it too coarse to reflect whether there are categories that are left behind (i.e., with close to zero per-class accuracy). Second, it is not necessarily true that classes in the \"Few\" category perform worse than those in \"Many\" or \"Medium\", so focusing on improving \"Few\" accuracy alone is not enough. For these reasons, we argued that we should pay more attention to the harmonic mean of per-class recall and we therefore proposed a novel method that aims at improving the lowest recall and the harmonic mean of recalls. Our method is a simple plug-in that is applicable to a wide range of existing methods. Specifically, our method consists of three stages, in the first stage, we use any existing method to train a model from scratch. Then we re-train the classifier using our proposed novel GML loss function. Finally, we propose a simple ensemble trick that can be used to combine the predictions from the two classifiers with nearly no extra cost. We validated the effectiveness of our method on three widely used benchmark datasets, and witnessed consistent improvements on the harmonic mean of recalls and lowest recall value, while the overall accuracy still remains high. By visualizing the distribution of per-class recall values of the fine-tuned model, we found our model indeed achieved a more balanced distribution." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15811" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Shaden Alshammari, Yu-Xiong Wang, Deva Ramanan, and Shu Kong. Long-tailed recognition via weight balancing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6897-6907, June 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.187, + 0.472, + 0.256 + ], + "angle": 0, + "content": "[2] Kaidi Cao, Colin Wei, Adrien Gaidon, Nikos Arechiga, and Tengyu Ma. Learning imbalanced datasets with label-distribution-aware margin loss. In Advances in Neural Information Processing Systems 32, pages 1567-1578, December 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.259, + 0.471, + 0.313 + ], + "angle": 0, + "content": "[3] Nitesh V Chawla, Kevin W Bowyer, Lawrence O Hall, and W Philip Kegelmeyer. SMOTE: Synthetic minority oversampling technique. Journal of Artificial Intelligence Research, 16:321-357, 2002. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.316, + 0.47, + 0.372 + ], + "angle": 0, + "content": "[4] Jiequan Cui, Zhisheng Zhong, Shu Liu, Bei Yu, and Jiaya Jia. Parametric contrastive learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 715-724, October 2021. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.373, + 0.47, + 0.442 + ], + "angle": 0, + "content": "[5] Yin Cui, Yang Song, Chen Sun, Andrew Howard, and Serge Belongie. Large scale fine-grained categorization and domain-specific transfer learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4109-4118, June 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.444, + 0.469, + 0.5 + ], + "angle": 0, + "content": "[6] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248-255, June 2009. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.502, + 0.469, + 0.556 + ], + "angle": 0, + "content": "[7] Chris Drummond and Robert C. Holte. C4.5, class imbalance, and cost sensitivity: Why under-sampling beats oversampling. In Workshop on Learning from Imbalanced Data Sets II, 2003. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.559, + 0.469, + 0.614 + ], + "angle": 0, + "content": "[8] Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q. Weinberger. On calibration of modern neural networks. In Proceedings of the 34th International Conference on Machine Learning, pages 1321-1330, August 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.616, + 0.469, + 0.672 + ], + "angle": 0, + "content": "[9] Hui Han, Wen-Yuan Wang, and Bing-Huan Mao. Borderline-SMOTE: A new over-sampling method in imbalanced data sets learning. In Advances in Intelligent Computing, pages 878-887, August 2005. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.673, + 0.469, + 0.715 + ], + "angle": 0, + "content": "[10] Haibo He and Edwardo A. Garcia. Learning from imbalanced data. IEEE Transactions on Knowledge and Data Engineering, 21(9):1263-1284, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.717, + 0.469, + 0.772 + ], + "angle": 0, + "content": "[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, June 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.774, + 0.469, + 0.829 + ], + "angle": 0, + "content": "[12] Yin-Yin He, Jianxin Wu, and Xiu-Shen Wei. Distilling virtual examples for long-tailed recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 235-244, October 2021. 1, 2, 3, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[13] Youngkyu Hong, Seungju Han, Kwanghee Choi, Seokjun Seo, Beomsu Kim, and Buru Chang. Disentangling label distribution for long-tailed visual recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6626-6636, June 2021. 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.161 + ], + "angle": 0, + "content": "[14] Bingyi Kang, Saining Xie, Marcus Rohrbach, Zhicheng Yan, Albert Gordo, Jiashi Feng, and Yannis Kalantidis. Decoupling representation and classifier for long-tailed recognition. In International Conference on Learning Representations, April 2020. 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[15] Jaehyung Kim, Jongheon Jeong, and Jinwoo Shin. M2m: Imbalanced classification via major-to-minor translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13896-13905, June 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.234, + 0.892, + 0.29 + ], + "angle": 0, + "content": "[16] Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, University of Toronto, 2009. https://www.cs.toronto.edu/~kriz/learning-features-2009-TR.pdf.1,5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.292, + 0.892, + 0.373 + ], + "angle": 0, + "content": "[17] Tianhong Li, Peng Cao, Yuan Yuan, Lijie Fan, Yuzhe Yang, Rogerio S. Feris, Piotr Indyk, and Dina Katabi. Targeted supervised contrastive learning for long-tailed recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6918-6928, June 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.376, + 0.892, + 0.432 + ], + "angle": 0, + "content": "[18] Tianhao Li, Limin Wang, and Gangshan Wu. Self supervision to distillation for long-tailed visual recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 630-639, October 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.434, + 0.892, + 0.503 + ], + "angle": 0, + "content": "[19] Ziwei Liu, Zhongqi Miao, Xiaohang Zhan, Jiayun Wang, Boqing Gong, and Stella X. Yu. Large-scale long-tailed recognition in an open world. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2537-2546, June 2019. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.505, + 0.892, + 0.56 + ], + "angle": 0, + "content": "[20] Aditya Krishna Menon, Sadeep Jayasumana, Ankit Singh Rawat, Himanshu Jain, Andreas Veit, and Sanjiv Kumar. Long-tail learning via logit adjustment. In International Conference on Learning Representations, May 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.562, + 0.892, + 0.644 + ], + "angle": 0, + "content": "[21] Seulki Park, Youngkyu Hong, Byeongho Heo, Sangdoo Yun, and Jin Young Choi. The majority can help the minority: Context-rich minority oversampling for long-tailed classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6887-6896, June 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.647, + 0.892, + 0.715 + ], + "angle": 0, + "content": "[22] Jiawei Ren, Cunjun Yu, Shunan Sheng, Xiao Ma, Haiyu Zhao, Shuai Yi, and Hongsheng Li. Balanced meta-softmax for long-tailed visual recognition. In Advances in Neural Information Processing Systems 33, pages 4175–4186, December 2020. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.717, + 0.892, + 0.772 + ], + "angle": 0, + "content": "[23] Dvir Samuel and Gal Chechik. Distributional robustness loss for long-tail learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9495-9504, October 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.774, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[24] Li Shen, Zhouchen Lin, and Qingming Huang. Relay backpropagation for effective learning of deep convolutional neural networks. In Proceedings of the European Conference on Computer Vision, pages 467-482, October 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[25] Peng Wang, Kai Han, Xiu-Shen Wei, Lei Zhang, and Lei Wang. Contrastive learning based hybrid networks for long-tailed image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 943-952, June 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15812" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[26] Xudong Wang, Long Lian, Zhongqi Miao, Ziwei Liu, and Stella Yu. Long-tailed recognition by routing diverse distribution-aware experts. In International Conference on Learning Representations, May 2021. 1, 2, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.471, + 0.218 + ], + "angle": 0, + "content": "[27] Saining Xie, Ross Girshick, Piotr Dollar, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1492-1500, July 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.22, + 0.47, + 0.274 + ], + "angle": 0, + "content": "[28] Yuzhe Yang and Zhi Xu. Rethinking the value of labels for improving class-imbalanced learning. In Advances in Neural Information Processing Systems 33, pages 19290-19301, December 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.277, + 0.47, + 0.345 + ], + "angle": 0, + "content": "[29] Songyang Zhang, Zeming Li, Shipeng Yan, Xuming He, and Jian Sun. Distribution alignment: A unified framework for long-tail visual recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2361-2370, June 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.347, + 0.47, + 0.414 + ], + "angle": 0, + "content": "[30] Zhisheng Zhong, Jiequan Cui, Shu Liu, and Jiaya Jia. Improving calibration for long-tailed recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16489-16498, June 2021. 1, 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.418, + 0.469, + 0.485 + ], + "angle": 0, + "content": "[31] Boyan Zhou, Quan Cui, Xiu-Shen Wei, and Zhao-Min Chen. BBN: Bilateral-branch network with cumulative learning for long-tailed visual recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9719–9728, June 2020. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.488, + 0.469, + 0.542 + ], + "angle": 0, + "content": "[32] Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 40(6):1452-1464, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.544, + 0.47, + 0.613 + ], + "angle": 0, + "content": "[33] Jianggang Zhu, Zheng Wang, Jingjing Chen, Yi-Ping Phoebe Chen, and Yu-Gang Jiang. Balanced contrastive learning for long-tailed visual recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6908-6917, June 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.613 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "15813" + } + ] +] \ No newline at end of file diff --git a/2023/No One Left Behind_ Improving the Worst Categories in Long-Tailed Learning/aa8d6d83-fa42-452c-97e5-1d8dfd5a2154_origin.pdf b/2023/No One Left Behind_ Improving the Worst Categories in Long-Tailed Learning/aa8d6d83-fa42-452c-97e5-1d8dfd5a2154_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4fd93973fc4a68144c54d2abd0e27b9c16dcdeec --- /dev/null +++ b/2023/No One Left Behind_ Improving the Worst Categories in Long-Tailed Learning/aa8d6d83-fa42-452c-97e5-1d8dfd5a2154_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:95be62cab89d5bd5c817c3fb2fa34308544424227023c69ad06221fe0b8bf625 +size 489352 diff --git a/2023/No One Left Behind_ Improving the Worst Categories in Long-Tailed Learning/full.md b/2023/No One Left Behind_ Improving the Worst Categories in Long-Tailed Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..de346a3a3c1d3f1c03e4855b2f884efeaa7c0715 --- /dev/null +++ b/2023/No One Left Behind_ Improving the Worst Categories in Long-Tailed Learning/full.md @@ -0,0 +1,335 @@ +# No One Left Behind: Improving the Worst Categories in Long-Tailed Learning + +Yingxiao Du, Jianxin Wu* +State Key Laboratory for Novel Software Technology, Nanjing University +Nanjing, China, 210023 + +duyx@lamda.nju.edu.cn, wujx2001@gmail.com + +# Abstract + +Unlike the case when using a balanced training dataset, the per-class recall (i.e., accuracy) of neural networks trained with an imbalanced dataset are known to vary a lot from category to category. The convention in long-tailed recognition is to manually split all categories into three subsets and report the average accuracy within each subset. We argue that under such an evaluation setting, some categories are inevitably sacrificed. On one hand, focusing on the average accuracy on a balanced test set incurs little penalty even if some worst performing categories have zero accuracy. On the other hand, classes in the "Few" subset do not necessarily perform worse than those in the "Many" or "Medium" subsets. We therefore advocate to focus more on improving the lowest recall among all categories and the harmonic mean of all recall values. Specifically, we propose a simple plug-in method that is applicable to a wide range of methods. By simply re-training the classifier of an existing pre-trained model with our proposed loss function and using an optional ensemble trick that combines the predictions of the two classifiers, we achieve a more uniform distribution of recall values across categories, which leads to a higher harmonic mean accuracy while the (arithmetic) average accuracy is still high. The effectiveness of our method is justified on widely used benchmark datasets. + +# 1. Introduction + +Various gaps exist when adapting image recognition techniques that are developed in the lab to industrial applications. The most noteworthy one is perhaps the difference between training datasets. Most training datasets used in academic research [6, 16] are balanced with respect to the number of images per class. This should not be taken for granted because datasets used in real-world applications are more likely to be imbalanced. Training deep models + +![](images/97852020da1ac5e4e32d437286bfa6f07d9af53cd05137c767a046d87c3a7cf6.jpg) +Figure 1. The per-class recall of models trained on the imbalanced CIFAR100 (with imbalance ratio 100). Per-class recall value varies a lot from category to category. Moreover, it is not necessarily true that all categories in the "Few" subset have lower accuracy than those in the "Many" or "Medium" subsets. + +
MethodMean AccuracyLowest Recall
BSCE [22]42.243.00
DiVE [12]45.112.00
MiSLAS [30]47.055.00
RIDE [26]48.642.00
PaCo [4]51.245.00
+ +Table 1. The lowest per-class recall of various state-of-the-art methods on the imbalanced CIFAR100 (with imbalance ratio 100). Although there are rapid improvements over the mean accuracy, the lowest per-class recall remains very low. + +on these datasets is not trivial as models are known to perform poorly on such datasets. Long-tailed recognition is a research field that aims at tackling this challenge. + +By plotting the per-class recall (i.e., per-class accuracy) of the model trained on the imbalanced CIFAR00 dataset (with imbalance ratio 100) in Fig. 1, we find that the recall varies dramatically from category to category. We argue that in real world applications, all categories are equally + +important and no one should be left behind. We also find that despite the rapid developments in this community, no obvious improvement over the lowest per-class recall is witnessed in the past few years [4, 12, 22, 26, 30], as is shown in Tab. 1. The convention in long-tailed recognition research is to split the classes into three subsets based on the number of training images. The accuracy within each subset is often reported along with the overall accuracy. While this evaluation scheme seems reasonable at the first glance, we argue it is potentially problematic. First of all, computing the average recall within each subset is way too coarse, making it impossible to reflect whether some classes are completely sacrificed but covered up by other "easy" tail classes. What's more, we find it is not necessarily true that classes in the "Few" category all have lower recall than classes in the other two subsets, as is shown in Fig. 1. Therefore, focusing on improving the mean accuracy alone is not enough, especially in real-world applications. + +In this paper, we propose a novel method to make sure that no category is left behind. Specifically, we argue that although mean accuracy is widely used in image classification as an optimization objective, due to the fact that different classes have very different recall in long-tailed recognition, it is not the most suitable objective as it incurs little penalty even if some categories have very small per-class accuracy values (e.g., close to 0). Hence, to improve even the worst-performing categories, we believe the harmonic mean of per-class recall would be a better objective. Since harmonic mean is very sensitive to small numbers, further improving classes that already have high recall brings little benefits to the harmonic mean, which makes sure no single class will be left behind. Also, now all classes are treated equally, forcing us to improve classes that have low recall no matter which subset it belongs to. + +However, it is difficult to directly minimize the harmonic mean. We therefore propose a novel loss function that maximizes the geometric mean instead, which can be viewed as a surrogate. Our method serves as a simple plug-in that can be used together with both baseline and various state-of-the-art methods. We also propose an ensemble trick that uses the pre-trained and fine-tuned models together to make predictions during inference with nearly no extra cost. We are able to yield a more uniform distribution of recall across categories (i.e., no one left behind), which achieves higher harmonic mean of the recall while the (arithmetic) mean accuracy remains high, too. + +In summary, our work has the following contributions: + +- We are the first to emphasize the importance of the correct recognition of all categories in long-tailed recognition. +- We propose a novel method that aims at increasing the harmonic mean of per-class recall as well as an ensem + +ble trick that combines two existing models together during inference with nearly no extra cost. + +- We experimented on three widely used benchmark datasets, which justify the effectiveness of our method in terms of both overall and worst per-class accuracy. + +# 2. Related Work + +Long-tailed recognition is a research field that aims at training models using imbalanced datasets. Since it is widely believed that the learning procedure is dominated by head classes, most existing works focus explicitly on improving the recognition of tail classes. + +# 2.1. Re-sampling and Re-weighting + +Re-sampling and re-weighting are two classical approaches. Re-sampling methods aim at balancing the training set by either over-sampling the tail classes [3, 9, 21, 24] or under-sampling the head classes [7, 10]. Some methods also transfer statistics from the major classes to minor classes to obtain a balanced dataset [15, 19]. One disadvantage of re-sampling methods is that it may lead to either over-fitting of tail classes or under-fitting of head classes. Re-weighting methods, on the other hand, give each instance a weight based on its true label when computing the loss [2, 22, 23]. The major drawback of re-weighting is that it makes the loss function hard to optimize. + +# 2.2. Two-Stage Decoupling + +Due to the shortcomings of re-sampling and reweighting, various other methods are proposed recently. For example, two-stage methods like [14, 18, 30] propose to decouple the learning of features and classifiers and achieve impressive results. Our method is similar to them in the sense that we also re-train a classifier given an existing model. But our motivation is to design a simple and flexible method that can be applied to a wide variety of existing models, and we propose a novel loss function, while two-stage methods like cRT [14] usually use the conventional cross entropy loss in the fine-tuning stage. + +# 2.3. Hybrid and Multiple Heads + +Hybrid methods, like [26, 31], make use of multiple heads to improve the recognition of different classes, which is similar to our proposed ensemble trick. But these methods require a joint training of multiple heads, together with a complex routing module that dynamically determines the head to use during inference, which increases the complexity of the model and makes it hard to use in practice. Our method, on the other hand, is simpler and requires no extra training nor complex modules to combine the predictions, which is different from them. + +# 2.4. Miscellaneous + +There are many approaches that use knowledge distillation [12, 18] and contrastive learning [4, 17, 25, 28, 33] and achieve descent results. There are also methods that perform logit adjustment [20, 29], which formulate long-tailed recognition as a label distribution shift problem. Recently, there is also one work that tries to solve the problem by performing some regularization on the classifier's weight [1]. They also visualize the recall of each class in their work and find the model is biased towards common classes, which is similar to our motivation. However, they do not emphasize the importance of correctly recognizing all classes, which is different from us. + +# 3. Method + +We will present our method in this section. First, we will introduce the notations used. Then we will describe the major framework of our method, which includes our proposed loss function and a simple ensemble trick. + +# 3.1. Notations + +The neural network can be represented as a non-linear function $\mathcal{F}$ . Given the $i$ -th training image $(x^i, y^i)$ from the imbalanced dataset that has $N$ images, where $x^i \in \mathbb{R}^{H \times W \times D}$ , a forward propagation through the network yields $\operatorname{logit} o^i = \mathcal{F}(x^i)$ . For a multi-class classification problem with $C$ categories, a softmax activation function is usually applied and the result can be represented as $\tilde{p}^i = \operatorname{softmax}(o^i)$ , where $\tilde{p}^i \in \mathbb{R}^C$ . + +# 3.2. A Simple Plug-In Method + +Over the past few years, various methods [4,14,22] have been proposed to tackle the challenges in long-tailed recognition. Although our evaluation results in Tab. 1 reveal that no obvious improvements have been made regarding the worst-performing category, many of their design choices are critical and effective in long-tailed recognition. Therefore, in order to make full use of existing advances, we do not aim at building up a whole new training framework, but rather propose a simple plug-in method that can be applied to a wide range of existing methods. + +Our method is simple. Given a model trained using either one of existing methods, we re-initialize the last FC layer and re-train it using our proposed loss function while the backbone is frozen. We borrow this idea from many two-stage decoupling methods [14, 30], which is effective and flexible, making our method widely applicable. However, unlike normal decoupling methods that use the conventional cross entropy loss in the second stage, we propose a novel loss function to improve the worst-performing categories. Our proposed loss function, called GML (Geometric + +Mean Loss), is defined as + +$$ +\mathcal {L} _ {\mathrm {G M L}} = - \frac {1}{C} \sum_ {c = 1} ^ {C} \log \bar {\boldsymbol {p}} _ {c}, \tag {1} +$$ + +where $\bar{\pmb{p}}_c$ is computed as + +$$ +\bar {\boldsymbol {p}} _ {c} = \frac {1}{N _ {c}} \sum_ {i = 1} ^ {N _ {c}} \tilde {\boldsymbol {p}} _ {y ^ {i}} ^ {i}, \tag {2} +$$ + +which is the average of $\tilde{\pmb{p}}_{y^i}^i$ across all training samples belonging to class $c$ in this mini-batch, and $N_{c}$ is the number of examples from this category in the mini-batch. Inspired by [22], we perform re-weighting when we compute $\tilde{\pmb{p}}^i$ : + +$$ +\tilde {\boldsymbol {p}} _ {j} ^ {i} = \frac {N _ {j} \exp \left(o _ {j} ^ {i}\right)}{\sum_ {c = 1} ^ {C} N _ {c} \exp \left(o _ {c} ^ {i}\right)}. \tag {3} +$$ + +# 3.3. Why GML Works? + +The definition of mean accuracy is + +$$ +\operatorname {A c c} = \frac {1}{N} \sum_ {i = 1} ^ {N} \mathbb {I} (y ^ {i} = \arg \max \tilde {\boldsymbol {p}} ^ {i}), \tag {4} +$$ + +where $\mathbb{1}(\cdot)$ is the indicator function. On a balanced test set, it is the same as the arithmetic mean of per-class recall. We argue that maximizing the mean accuracy inevitably ignores some categories because arithmetic mean incurs little punishment to categories that have very small recall values. On the other hand, harmonic mean, defined as + +$$ +\mathrm {H M} \left(x _ {1}, \dots , x _ {n}\right) = \frac {n}{\frac {1}{x _ {1}} + \cdots + \frac {1}{x _ {n}}} \tag {5} +$$ + +incurs strong punishment to small values. Therefore, we believe it can be a better alternative to optimize for. However, harmonic mean is defined using reciprocal, making it hard and numeric unstable to be optimized. To this extent, we propose to maximize the geometric mean of per-class recall, defined as + +$$ +\operatorname {G M} \left(x _ {1}, \dots , x _ {n}\right) = \sqrt [ n ]{\left| x _ {1} \times \dots \times x _ {n} \right|}. \tag {6} +$$ + +As an illustration, $\frac{0.01 + 0.99}{2} = 0.5$ , $\mathrm{HM}(0.01, 0.99) = 0.02$ and $\mathrm{GM}(0.01, 0.99) = 0.10$ . Both harmonic and geometric means are heavily affected by the small value (0.01), while the arithmetic mean is much less sensitive to it. + +Specifically, given the per-class recall as $r_1, \ldots, r_C$ , we transform their geometric mean using a simple logarithm transformation: + +$$ +\begin{array}{l} \sqrt {r _ {1} \dots r _ {C}} = \exp \left(\log \left(r _ {1} \dots r _ {C}\right) ^ {1 / C}\right) (7) \\ = \exp \left(\frac {1}{C} \log \left(r _ {1} \dots r _ {C}\right)\right) (8) \\ = \exp \left(\frac {1}{C} \sum_ {i = 1} ^ {C} \log r _ {i}\right). (9) \\ \end{array} +$$ + +![](images/f099466616702b441e44f2378abea8fe708b6035da967be1caa41402c7ffedb0.jpg) +Figure 2. Bar plot of per-class recall on the imbalanced CIFAR10 (with imbalance ratio 100) before ('CE') and after ('CE+GML') the fine-tuning. For the fine-tuned model, the recall of the first 5 classes dropped while the recall of the latter 5 classes increased, motivating us to combine the prediction of both models. + +Since recall is computed by computing the mean of a series of indicator functions and these indicator functions only take value 0 or 1 and is non-differentiable, we use $\bar{\pmb{p}}_c$ as a surrogate for $r_c$ during training, resulting in + +$$ +\sqrt {r _ {1} \dots r _ {C}} \propto \exp (- \mathcal {L} _ {\mathrm {G M L}}). \tag {10} +$$ + +So, by minimizing $\mathcal{L}_{\mathrm{GML}}$ , we are effectively maximizing the geometric mean of per-class recall, thus being able to improve on the worst-performing categories. + +# 3.4. Combining the Strength of Both Worlds + +By fine-tuning the pre-trained model with our proposed loss function, indeed we are able to obtain a more uniform distribution of recall values across categories. As is shown in Fig. 2, the recall of classes that perform worse before are improved while the recall of classes that perform well before are dropped. Motivated by these dynamics, we then propose a simple ensemble trick that aims at combining the strength of both worlds. + +Our idea is simple again. Since we re-initialize the classifier during fine-tuning and the backbone is kept frozen, the old classifier is still usable during inference. Let these two models be $\mathcal{F}_{\mathrm{old}}$ and $\mathcal{F}_{\mathrm{new}}$ , respectively, we then make predictions by + +$$ +\tilde {\boldsymbol {p}} _ {\text {e n s e m b l e d}} = \frac {\tilde {\boldsymbol {p}} _ {\text {n e w}} + \tilde {\boldsymbol {p}} _ {\text {o l d}}}{2}. \tag {11} +$$ + +In practice, since the calibration of some models can be poor [8,30] (e.g., the baseline using cross entropy), we need to manually calibrate the prediction before combining them. For simplicity, we choose temperature scaling to calibrate the model [8] and use two temperature variables $t_{\mathrm{new}}$ and + +$t_{\mathrm{old}}$ when computing $\tilde{\pmb{p}}_{\mathrm{new}}$ and $\tilde{\pmb{p}}_{\mathrm{old}}$ separately: + +$$ +\tilde {\boldsymbol {p}} _ {\text {n e w}} = \operatorname {s o f t m a x} \left(\frac {\boldsymbol {o} _ {\text {n e w}}}{t _ {\text {n e w}}}\right), \tag {12} +$$ + +$$ +\tilde {\boldsymbol {p}} _ {\text {o l d}} = \operatorname {s o f t m a x} \left(\frac {\boldsymbol {o} _ {\text {o l d}}}{t _ {\text {o l d}}}\right). \tag {13} +$$ + +One big advantage of our proposed ensemble trick is that it brings nearly no extra cost. Unlike many multi-head [31] or multi-expert [26] methods, we do not need extra training as we simply use the classifier that comes with the pretrained model. During inference, the only extra cost is the fully connected layer to compute $\tilde{\pmb{p}}_{\mathrm{old}}$ , which is negligible. Also, no complex routing rules are needed. + +The two temperature variables are the only two hyperparameters introduced by our method, and they can be easily tuned as no training is required. Plus, as we will later show in the ablation studies, most values work perfectly fine so actually no painful tuning is needed. What's more, this ensemble trick is completely optional, our method works well enough even without it. + +# 3.5. Training Pipeline + +Our method is very similar to the two-stage decoupling methods [14, 30]. We design it in this way on purpose because we want it to be as flexible and simple as possible. During experiments, we found that our method is orthogonal to many design choices in state-of-the-art methods and is able to improve the worse-case performance when applied to various different pre-trained models. The overall training pipeline of our method consists of three stages: + +1. Pre-training stage: Obtain the pre-trained model from scratch using any one of the existing methods. +2. Fine-tuning stage: Freeze the backbone and re-train the classifier using our proposed loss function. +3. An optional ensemble stage: Calibrate the prediction of two classifiers and combine them additively. + +The overall training procedure is summarized in Algorithm 1. + +# 4. Experiments + +In order to validate the effectiveness of our method, we have conducted experiments on three widely used benchmark datasets. In this section, we will present our experimental results. First, we will introduce the datasets used. Then we will describe various implementation details and evaluation metrics used. After that, we will compare our method with various other methods. Finally, we will present the results of some ablation studies. + +
DatasetNumber of Classes# Training Images# Test ImagesImbalance Ratio
CIFAR100-LT [16]10010,84710,000100
ImageNet-LT [6, 19]1,000115,84650,000256
Places-LT [32]36562,50036,500996
+ +Table 2. Statistics of three imbalanced datasets used in our experiments. + +# Algorithm 1 The overall training procedure + +Input: Training images $x$ and their labels $y$ . + +1: Randomly initialize the network and train it from scratch to obtain the pre-trained model. +2: Use the pre-trained model to perform initialization. +3: Freeze the backbone and re-initialize the classifier. +4: Fine-tune the model using the loss function defined as Eq. (1) for a few epochs. +5: During inference, combine the results from both classifiers as defined in Eq. (11). + +# 4.1. Datasets + +We use three widely used long-tailed image recognition datasets: CIFAR100-LT, ImageNet-LT and PlacesLT. Some statistics about these three datasets are summarized in Tab. 2. Since CIFAR100 [16], ImageNet [6] and Places [32] are all balanced datasets, we follow previous works [19, 31] to down-sample their original training set. Details about the construction process can be found in the supplementary material. + +# 4.2. Implementation Details + +Following previous works [19, 31], we use ResNet-32 [11], ResNeXt-50 [27] and ResNet-152 as the backbone network for CIFAR100-LT, ImageNet-LT and Places-LT, respectively. For the pre-training stage, if there exists released checkpoints, we use them directly. If not, we use the officially released codes and try to reproduce the results in our best effort without changing their original settings. This includes stuffs like hyper-parameters and data augmentations used. For the fine-tuning stage, detailed training settings can be found in the supplementary material. + +# 4.3. Evaluation Metrics and Comparison Methods + +The convention in long-tailed recognition research is to split the classes into three subsets based on the number of training images [14, 19]. Typically, "Head" denotes classes that have more than 100 images, "Few" denotes classes that have less than 20 images and all other classes are denoted as "Medium". The accuracy within each subset is often reported along with the overall accuracy to justify the claim that the performance improvement mainly comes from tail classes. And according to our observation in Sec. 1, we ar + +gue that such an evaluation scheme can be problematic and we propose to first compute the recall within each category and then compute their harmonic mean. In our experiments, since our method optimizes geometric mean of per-class recall as a surrogate, we also report the geometric mean. Besides, we report the overall mean accuracy as well. + +We compare our methods with both baseline methods like simple cross entropy loss and various state-of-the-arts methods [4, 12, 22, 26]. Many of these methods have very different training settings and strictly speaking, they are not directly comparable. But since our method is a simple plugin method, the main focus is to quantify the improvements brought by our method compared to its respective baseline result. + +# 4.4. Main Comparisons + +We validate our method on three benchmark datasets and present the results on each dataset here separately. + +# 4.4.1 CIFAR100-LT + +Tab. 3 shows the experimental results on the CIFAR100-LT dataset. CIFAR100-LT has three different versions with different imbalance ratios and here we only consider the one with imbalance ratio 100. As we can see from the table, although various different methods have been proposed in recent years and there are rapid improvements on the overall accuracy, the performance of the worst-performing category is still relatively poor. The geometric and harmonic mean of per-class recall also justify our claim. We apply our method to both the baseline method CE and state-of-the-art method PaCo [4] on this dataset. When our method is applied to CE and PaCo, we are able to improve the lowest recall from 0.00 and 5.00 to 6.00 and 9.00 respectively. And if we further apply our proposed ensemble trick, we are able to improve the lowest recall to 15.00, achieving state-of-the-art performance in both geometric mean and harmonic mean while keeping the overall accuracy roughly unchanged. + +It is worth noting that even the simplest CE+GML outperforms all existing methods in terms of lowest recall, that is, in terms of "no one left behind". + +# 4.4.2 ImageNet-LT and Places-LT + +Tab. 4 shows the result on ImageNet-LT. The overall observation is similar to that on CIFAR100-LT, except that + +
MethodsAccuracyGeometric MeanHarmonic MeanLowest Recall
CE38.7421.03*2.05*0.00
BSCE [22]42.2435.1626.243.00
LDAM [2]43.5133.6121.493.00
LADE [13]44.3939.0532.585.00
DiVE [12]45.1137.0825.182.00
MiSLAS [30]47.0540.1630.935.00
RIDE [26]48.6438.7123.862.00
PaCo [4]51.2445.2936.425.00
CE + GML41.0636.5931.266.00
PaCo + GML50.5345.4739.209.00
PaCo + GML (Ensemble)49.8245.7041.0215.00
+ +Table 3. Results on the CIFAR100-LT dataset with imbalance ratio 100. Numbers with * are computed by substituting zero elements with a small number $(10^{-3})$ or else the geometric and harmonic mean will all be zero. + +
MethodsAccuracyGeometric MeanHarmonic MeanLowest Recall
CE43.9023.25*1.25*0.00
BSCE [22]50.4842.32*13.74*0.00
cRT [14]49.6441.35*13.82*0.00
DiVE [12]53.6345.49*12.76*0.00
RIDE [26]55.6947.56*17.32*0.00
PaCo [4]58.5351.32*21.79*0.00
CE + GML45.6139.8231.672.00
PaCo + GML55.5750.6743.712.00
PaCo + GML (Ensemble)57.2252.2944.802.00
+ +Table 4. Results on the ImageNet-LT dataset. Numbers with * are computed by substituting zero elements with a small number $(10^{-3})$ or else the geometric and harmonic mean will all be zero. + +since this dataset is larger and harder, the lowest recall of all current methods are zero, making the harmonic mean of per-class recall very low, even when a small number substitutes zero per-class recalls. Our method, on the other hand, greatly improves the harmonic mean by successfully improving the lowest recall value. All categories have per-class recall higher than 0 when our method is used, even in the simplest CE+GML. + +Tab. 5 shows the result on Places-LT. Current state-of-the-art on this dataset is MiSLAS [30], so we apply our method upon it and the simple baseline CE. Again, our method successfully improves the lowest recall, as well as the harmonic mean of per-class recall. And when the ensemble trick is further applied, we are able to achieve new state-of-the-art results. + +# 4.4.3 Universality + +Since we want our method to be a universal plug-in, in this subsection we present the results of the experiment of adding our method on top of various different methods. The experiment is conducted on CIFAR100-LT, and the re + +sult is shown in Tab. 6. Note that here we do not perform any ensemble. As we can see, our method is applicable to various different methods, and is able to achieve consistent improvements. However, we do observed that there exists some methods, like DiVE [12], where we failed to achieve obvious improvements. More details about this can be found when we discuss limitations in Sec. 5. + +# 4.5. Ablation Studies + +For ablation studies, if not otherwise specified, all experiments are conducted on CIFAR100-LT (with imbalance ratio 100) under the default training setting. + +# 4.5.1 Effects of Two Temperature Variables in the Ensemble Stage + +When we design our method, we'd like to keep the number of new hyper-parameters as less as possible in order for it to be readily applicable in real-world applications. The only two hyper-parameters introduced by our method are in the optional ensemble stage. Here we study the effects of choosing different values of them. Specifically, we change + +
MethodsAccuracyGeometric MeanHarmonic MeanLowest Recall
CE28.7112.11*0.73*0.00
BSCE [22]37.1829.30*5.64*0.00
PaCo [4]40.4527.88*2.53*0.00
MiSLAS [30]40.4835.5328.973.00
CE + GML36.8229.30*8.00*0.00
MiSLAS + GML39.9035.4129.985.00
MiSLAS + GML (Ensemble)40.3735.9230.515.00
+ +Table 5. Results on the Places-LT dataset. Numbers with * are computed by substituting zero elements with a small number $(10^{-3})$ or else the geometric and harmonic mean will all be zero. Note that to plug GML into MiSLAS, since MiSLAS is also a two-stage method, we think it's fairer to use GML together with their proposed label aware smoothing loss. + +
MethodsG-MeanH-MeanLowest Recall
CE21.03*2.05*0.00
CE + GML36.59 (15.56↑)31.26 (29.21↑)6.00 (6.00↑)
BSCE35.1626.243.00
BSCE + GML36.52 (1.36↑)30.88 (4.64↑)7.00 (4.00↑)
MiSLAS40.1630.935.00
MiSLAS + GML40.90 (0.74↑)36.49 (5.56↑)11.00 (6.00↑)
PaCo45.2936.425.00
PaCo + GML45.47 (0.18↑)39.20 (2.78↑)9.00 (4.00↑)
+ +Table 6. Results on the CIFAR100-LT dataset with imbalance ratio 100. Our proposed method is applicable to various methods. "G-Mean" is short for "Geometric Mean" and "H-Mean" is short for "Harmonic Mean". + +
toldtnewGeometric MeanHarmonic MeanLowest Recall
1145.7041.0215.00
1245.5838.868.00
1344.6836.095.00
2144.3940.2215.00
2245.6141.0115.00
2346.0640.4011.00
3143.7339.6315.00
+ +the value of the temperature variables shown in Eq. (12) and Eq. (13) in the ensemble stage when we apply our method to PaCo [4]. The results are shown in Tab. 7. In our experiments on CIFAR100-LT, we set $t_{\mathrm{old}} = 1$ and $t_{\mathrm{new}} = 1$ . And as we can see from the table, the value of $t_{\mathrm{old}}$ does not affect the result much while $t_{\mathrm{new}}$ is somewhat important, especially for improving the lowest recall—higher $t_{\mathrm{new}}$ will hurt it. This is natural because higher $t$ results in a smoother distribution and less information is preserved. Since the newly trained classifier is crucial for improving the lowest recall, we need to preserve more information from it. + +Generally speaking, since these two hyper-parameters + +Table 7. Ablation on the effects of temperature variables when we apply our ensemble method upon PaCo [4]. + +
MethodsG-MeanH-MeanLowest Recall
CE + GML36.5931.266.00
(w/o re-weighting)25.455.110.00
(w/o re-weighting, re-sampling)32.3223.054.00
+ +Table 8. Ablation on the re-weighting design in GML. + +
MethodsG-MeanH-MeanLowest Recall
CE + GML36.5931.266.00
PaCo + GML45.4739.209.00
GML (train from scratch)36.6330.869.00
+ +Table 9. Training from scratch using GML. + +are introduced for calibration, for different methods, we may need to tune them respectively because the calibrations of different methods are known to vary a lot in long-tailed recognition problems [30]. But usually this will not cause a headache, because the ensemble stage does not involve any training and thus can be tuned very quickly. + +# 4.5.2 Re-weighting in GML + +As shown in Eq. (3), re-weighting is applied in our proposed GML, here we ablate this design. The results are shown in Tab. 8. As we can see, the performance will be relatively poor without re-weighting. And just like many two-stage decoupling methods [14], re-sampling during fine-tuning can help, but is inferior to re-weighting. + +# 4.5.3 Training from Scratch using GML + +Since our main purpose is to propose a simple plug-in method that is widely applicable to various existing methods, we use GML only to re-train the classifier of an existing model. However, GML can also be used to train a model from scratch. The results are shown in Tab. 9. As we can see, the result is comparable to the result when GML + +![](images/0c9f97b185c9700a78f472c479867b19bd9e22c76213909cd608fc7f7d58bc69.jpg) +Figure 3. Bar plot of per-class recall on the imbalanced CIFAR100 (with imbalance ratio 100) before and after the fine-tuning. No ensemble is performed. Our method focuses on improving the recognition of classes that have low recall values, making sure no category is left behind. + +is attached to CE, but inferior to the result when GML is applied to SOTA methods like PaCo, just as expected. + +# 4.5.4 Visualization of the Per-Class Recall + +In Fig. 1, we visualize the per-class recall of model trained on the imbalanced dataset, and find they vary a lot from class to class. Since our major motivation is to make sure all classes are equally treated so that no category will be left behind, here we visualize the change of per-class recall after applying our method. We perform visualizing using two datasets, CIFAR10-LT and CIFAR100-LT, both have imbalance ratio 100. The visualization results on CIFAR10-LT and CIFAR100-LT are shown in Fig. 2 and Fig. 3, respectively. Note that here we do not perform the ensemble. As we can clearly see in both figures, our fine-tuned model has a more uniform distribution of per-class recall. These two figures justify the effectiveness of our proposed method. + +# 5. Limitations and Broader Impacts + +Although our method is applicable to a wide range of existing methods, we do find that when applied to some methods, the improvement is less obvious compared to others. For example, our framework does not cope well with methods like BSCE [22] and DiVE [12]. Interestingly, although DiVE is based on knowledge distillation, BSCE is used in DiVE as well to train both the teacher and student model so we attribute the failure to the fact that these method all re-weight the loss function. Two-stage decoupling methods like cRT [14] are also known to not perform well when the first stage training involves re-weighting or re-sampling. The author argues it is because re-sampling and re-weighting have negative effects on feature learning and we believe that argument applies to our method as well. + +Another thing worth noting here is that we do not report the experimental results on iNaturalist2018 [5] because we find it generally not suitable in our settings. iNaturalist2018 is another widely used benchmark dataset in long-tailed recognition. It is large-scale, consisting of 437.5K training images from 8142 classes. But for the test split, it only has 24K images, so each class only has 3 images. Since our major motivation is to make sure no category is left behind, it is generally not meaningful to look at the per-class recall in iNaturalist2018 because the variance of the result can be very large due to the very small number of test images. However, in real-world applications, it is generally rare to have such small number of test images. + +For broader impacts, we believe our research have positive impacts on reducing the bias of model training on an imbalanced dataset. Machine learning models are known to be biased if there is lack of training images from some categories. We believe our work will be helpful to achieve better fairness by making sure all categories are equally treated. + +# 6. Conclusions + +In this paper, we focused on improving some of the worst-performing classes in long-tailed recognition. We found that when trained with an imbalanced dataset, the per-class recall of the model varies a lot from class to class. Previous convention reports the arithmetic mean of per-class recall, but we argued such an evaluation scheme can be problematic. First, arithmetic mean only incurs little penalty to small numbers, making it too coarse to reflect whether there are categories that are left behind (i.e., with close to zero per-class accuracy). Second, it is not necessarily true that classes in the "Few" category perform worse than those in "Many" or "Medium", so focusing on improving "Few" accuracy alone is not enough. For these reasons, we argued that we should pay more attention to the harmonic mean of per-class recall and we therefore proposed a novel method that aims at improving the lowest recall and the harmonic mean of recalls. Our method is a simple plug-in that is applicable to a wide range of existing methods. Specifically, our method consists of three stages, in the first stage, we use any existing method to train a model from scratch. Then we re-train the classifier using our proposed novel GML loss function. Finally, we propose a simple ensemble trick that can be used to combine the predictions from the two classifiers with nearly no extra cost. We validated the effectiveness of our method on three widely used benchmark datasets, and witnessed consistent improvements on the harmonic mean of recalls and lowest recall value, while the overall accuracy still remains high. By visualizing the distribution of per-class recall values of the fine-tuned model, we found our model indeed achieved a more balanced distribution. + +# References + +[1] Shaden Alshammari, Yu-Xiong Wang, Deva Ramanan, and Shu Kong. Long-tailed recognition via weight balancing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6897-6907, June 2022. 3 +[2] Kaidi Cao, Colin Wei, Adrien Gaidon, Nikos Arechiga, and Tengyu Ma. Learning imbalanced datasets with label-distribution-aware margin loss. In Advances in Neural Information Processing Systems 32, pages 1567-1578, December 2019. 2, 6 +[3] Nitesh V Chawla, Kevin W Bowyer, Lawrence O Hall, and W Philip Kegelmeyer. SMOTE: Synthetic minority oversampling technique. Journal of Artificial Intelligence Research, 16:321-357, 2002. 2 +[4] Jiequan Cui, Zhisheng Zhong, Shu Liu, Bei Yu, and Jiaya Jia. Parametric contrastive learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 715-724, October 2021. 1, 2, 3, 5, 6, 7 +[5] Yin Cui, Yang Song, Chen Sun, Andrew Howard, and Serge Belongie. Large scale fine-grained categorization and domain-specific transfer learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4109-4118, June 2018. 8 +[6] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248-255, June 2009. 1, 5 +[7] Chris Drummond and Robert C. Holte. C4.5, class imbalance, and cost sensitivity: Why under-sampling beats oversampling. In Workshop on Learning from Imbalanced Data Sets II, 2003. 2 +[8] Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q. Weinberger. On calibration of modern neural networks. In Proceedings of the 34th International Conference on Machine Learning, pages 1321-1330, August 2017. 4 +[9] Hui Han, Wen-Yuan Wang, and Bing-Huan Mao. Borderline-SMOTE: A new over-sampling method in imbalanced data sets learning. In Advances in Intelligent Computing, pages 878-887, August 2005. 2 +[10] Haibo He and Edwardo A. Garcia. Learning from imbalanced data. IEEE Transactions on Knowledge and Data Engineering, 21(9):1263-1284, 2009. 2 +[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, June 2016. 5 +[12] Yin-Yin He, Jianxin Wu, and Xiu-Shen Wei. Distilling virtual examples for long-tailed recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 235-244, October 2021. 1, 2, 3, 5, 6, 8 +[13] Youngkyu Hong, Seungju Han, Kwanghee Choi, Seokjun Seo, Beomsu Kim, and Buru Chang. Disentangling label distribution for long-tailed visual recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6626-6636, June 2021. 6 + +[14] Bingyi Kang, Saining Xie, Marcus Rohrbach, Zhicheng Yan, Albert Gordo, Jiashi Feng, and Yannis Kalantidis. Decoupling representation and classifier for long-tailed recognition. In International Conference on Learning Representations, April 2020. 2, 3, 4, 5, 6, 7, 8 +[15] Jaehyung Kim, Jongheon Jeong, and Jinwoo Shin. M2m: Imbalanced classification via major-to-minor translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13896-13905, June 2020. 2 +[16] Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, University of Toronto, 2009. https://www.cs.toronto.edu/~kriz/learning-features-2009-TR.pdf.1,5 +[17] Tianhong Li, Peng Cao, Yuan Yuan, Lijie Fan, Yuzhe Yang, Rogerio S. Feris, Piotr Indyk, and Dina Katabi. Targeted supervised contrastive learning for long-tailed recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6918-6928, June 2022. 3 +[18] Tianhao Li, Limin Wang, and Gangshan Wu. Self supervision to distillation for long-tailed visual recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 630-639, October 2021. 2, 3 +[19] Ziwei Liu, Zhongqi Miao, Xiaohang Zhan, Jiayun Wang, Boqing Gong, and Stella X. Yu. Large-scale long-tailed recognition in an open world. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2537-2546, June 2019. 2, 5 +[20] Aditya Krishna Menon, Sadeep Jayasumana, Ankit Singh Rawat, Himanshu Jain, Andreas Veit, and Sanjiv Kumar. Long-tail learning via logit adjustment. In International Conference on Learning Representations, May 2021. 3 +[21] Seulki Park, Youngkyu Hong, Byeongho Heo, Sangdoo Yun, and Jin Young Choi. The majority can help the minority: Context-rich minority oversampling for long-tailed classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6887-6896, June 2022. 2 +[22] Jiawei Ren, Cunjun Yu, Shunan Sheng, Xiao Ma, Haiyu Zhao, Shuai Yi, and Hongsheng Li. Balanced meta-softmax for long-tailed visual recognition. In Advances in Neural Information Processing Systems 33, pages 4175–4186, December 2020. 1, 2, 3, 5, 6, 7, 8 +[23] Dvir Samuel and Gal Chechik. Distributional robustness loss for long-tail learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9495-9504, October 2021. 2 +[24] Li Shen, Zhouchen Lin, and Qingming Huang. Relay backpropagation for effective learning of deep convolutional neural networks. In Proceedings of the European Conference on Computer Vision, pages 467-482, October 2016. 2 +[25] Peng Wang, Kai Han, Xiu-Shen Wei, Lei Zhang, and Lei Wang. Contrastive learning based hybrid networks for long-tailed image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 943-952, June 2021. 3 + +[26] Xudong Wang, Long Lian, Zhongqi Miao, Ziwei Liu, and Stella Yu. Long-tailed recognition by routing diverse distribution-aware experts. In International Conference on Learning Representations, May 2021. 1, 2, 4, 5, 6 +[27] Saining Xie, Ross Girshick, Piotr Dollar, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1492-1500, July 2017. 5 +[28] Yuzhe Yang and Zhi Xu. Rethinking the value of labels for improving class-imbalanced learning. In Advances in Neural Information Processing Systems 33, pages 19290-19301, December 2020. 3 +[29] Songyang Zhang, Zeming Li, Shipeng Yan, Xuming He, and Jian Sun. Distribution alignment: A unified framework for long-tail visual recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2361-2370, June 2021. 3 +[30] Zhisheng Zhong, Jiequan Cui, Shu Liu, and Jiaya Jia. Improving calibration for long-tailed recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16489-16498, June 2021. 1, 2, 3, 4, 6, 7 +[31] Boyan Zhou, Quan Cui, Xiu-Shen Wei, and Zhao-Min Chen. BBN: Bilateral-branch network with cumulative learning for long-tailed visual recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9719–9728, June 2020. 2, 4, 5 +[32] Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 40(6):1452-1464, 2018. 5 +[33] Jianggang Zhu, Zheng Wang, Jingjing Chen, Yi-Ping Phoebe Chen, and Yu-Gang Jiang. Balanced contrastive learning for long-tailed visual recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6908-6917, June 2022. 3 \ No newline at end of file diff --git a/2023/No One Left Behind_ Improving the Worst Categories in Long-Tailed Learning/images.zip b/2023/No One Left Behind_ Improving the Worst Categories in Long-Tailed Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d62c22257e75f8e43c207dec70d7812ddf819206 --- /dev/null +++ b/2023/No One Left Behind_ Improving the Worst Categories in Long-Tailed Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8ec6b303fc38101b3a077366317440f85ef7f24bf4aac34062af7f2518b2d005 +size 371211 diff --git a/2023/No One Left Behind_ Improving the Worst Categories in Long-Tailed Learning/layout.json b/2023/No One Left Behind_ Improving the Worst Categories in Long-Tailed Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d14407f4bc68fe098f969afd15de06aaa46f2e1f --- /dev/null +++ b/2023/No One Left Behind_ Improving the Worst Categories in Long-Tailed Learning/layout.json @@ -0,0 +1,7391 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 53, + 103, + 539, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 103, + 539, + 121 + ], + "spans": [ + { + "bbox": [ + 53, + 103, + 539, + 121 + ], + "type": "text", + "content": "No One Left Behind: Improving the Worst Categories in Long-Tailed Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 118, + 143, + 475, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 143, + 475, + 185 + ], + "spans": [ + { + "bbox": [ + 118, + 143, + 475, + 185 + ], + "type": "text", + "content": "Yingxiao Du, Jianxin Wu* \nState Key Laboratory for Novel Software Technology, Nanjing University \nNanjing, China, 210023" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 183, + 188, + 406, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 183, + 188, + 406, + 198 + ], + "spans": [ + { + "bbox": [ + 183, + 188, + 406, + 198 + ], + "type": "text", + "content": "duyx@lamda.nju.edu.cn, wujx2001@gmail.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 226, + 192, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 226, + 192, + 239 + ], + "spans": [ + { + "bbox": [ + 143, + 226, + 192, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 251, + 290, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 251, + 290, + 540 + ], + "spans": [ + { + "bbox": [ + 47, + 251, + 290, + 540 + ], + "type": "text", + "content": "Unlike the case when using a balanced training dataset, the per-class recall (i.e., accuracy) of neural networks trained with an imbalanced dataset are known to vary a lot from category to category. The convention in long-tailed recognition is to manually split all categories into three subsets and report the average accuracy within each subset. We argue that under such an evaluation setting, some categories are inevitably sacrificed. On one hand, focusing on the average accuracy on a balanced test set incurs little penalty even if some worst performing categories have zero accuracy. On the other hand, classes in the \"Few\" subset do not necessarily perform worse than those in the \"Many\" or \"Medium\" subsets. We therefore advocate to focus more on improving the lowest recall among all categories and the harmonic mean of all recall values. Specifically, we propose a simple plug-in method that is applicable to a wide range of methods. By simply re-training the classifier of an existing pre-trained model with our proposed loss function and using an optional ensemble trick that combines the predictions of the two classifiers, we achieve a more uniform distribution of recall values across categories, which leads to a higher harmonic mean accuracy while the (arithmetic) average accuracy is still high. The effectiveness of our method is justified on widely used benchmark datasets." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 561, + 128, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 561, + 128, + 574 + ], + "spans": [ + { + "bbox": [ + 47, + 561, + 128, + 574 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 581, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 581, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 47, + 581, + 287, + 677 + ], + "type": "text", + "content": "Various gaps exist when adapting image recognition techniques that are developed in the lab to industrial applications. The most noteworthy one is perhaps the difference between training datasets. Most training datasets used in academic research [6, 16] are balanced with respect to the number of images per class. This should not be taken for granted because datasets used in real-world applications are more likely to be imbalanced. Training deep models" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 315, + 243, + 545, + 384 + ], + "blocks": [ + { + "bbox": [ + 315, + 243, + 545, + 384 + ], + "lines": [ + { + "bbox": [ + 315, + 243, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 315, + 243, + 545, + 384 + ], + "type": "image", + "image_path": "97852020da1ac5e4e32d437286bfa6f07d9af53cd05137c767a046d87c3a7cf6.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 393, + 547, + 449 + ], + "lines": [ + { + "bbox": [ + 305, + 393, + 547, + 449 + ], + "spans": [ + { + "bbox": [ + 305, + 393, + 547, + 449 + ], + "type": "text", + "content": "Figure 1. The per-class recall of models trained on the imbalanced CIFAR100 (with imbalance ratio 100). Per-class recall value varies a lot from category to category. Moreover, it is not necessarily true that all categories in the \"Few\" subset have lower accuracy than those in the \"Many\" or \"Medium\" subsets." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 328, + 461, + 525, + 541 + ], + "blocks": [ + { + "bbox": [ + 328, + 461, + 525, + 541 + ], + "lines": [ + { + "bbox": [ + 328, + 461, + 525, + 541 + ], + "spans": [ + { + "bbox": [ + 328, + 461, + 525, + 541 + ], + "type": "table", + "html": "
MethodMean AccuracyLowest Recall
BSCE [22]42.243.00
DiVE [12]45.112.00
MiSLAS [30]47.055.00
RIDE [26]48.642.00
PaCo [4]51.245.00
", + "image_path": "445666365eb1a53ab7d7b987078d69ee5f9d5b8ffcde149a0e9d9cc739e09372.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 549, + 547, + 594 + ], + "lines": [ + { + "bbox": [ + 305, + 549, + 547, + 594 + ], + "spans": [ + { + "bbox": [ + 305, + 549, + 547, + 594 + ], + "type": "text", + "content": "Table 1. The lowest per-class recall of various state-of-the-art methods on the imbalanced CIFAR100 (with imbalance ratio 100). Although there are rapid improvements over the mean accuracy, the lowest per-class recall remains very low." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 616, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 616, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 616, + 545, + 653 + ], + "type": "text", + "content": "on these datasets is not trivial as models are known to perform poorly on such datasets. Long-tailed recognition is a research field that aims at tackling this challenge." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": "By plotting the per-class recall (i.e., per-class accuracy) of the model trained on the imbalanced CIFAR00 dataset (with imbalance ratio 100) in Fig. 1, we find that the recall varies dramatically from category to category. We argue that in real world applications, all categories are equally" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 683, + 287, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 683, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 683, + 287, + 712 + ], + "type": "text", + "content": "*J. Wu is the corresponding author. This research was partly supported by the National Natural Science Foundation of China under Grant 62276123 and Grant 61921006." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15804" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 288 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 288 + ], + "type": "text", + "content": "important and no one should be left behind. We also find that despite the rapid developments in this community, no obvious improvement over the lowest per-class recall is witnessed in the past few years [4, 12, 22, 26, 30], as is shown in Tab. 1. The convention in long-tailed recognition research is to split the classes into three subsets based on the number of training images. The accuracy within each subset is often reported along with the overall accuracy. While this evaluation scheme seems reasonable at the first glance, we argue it is potentially problematic. First of all, computing the average recall within each subset is way too coarse, making it impossible to reflect whether some classes are completely sacrificed but covered up by other \"easy\" tail classes. What's more, we find it is not necessarily true that classes in the \"Few\" category all have lower recall than classes in the other two subsets, as is shown in Fig. 1. Therefore, focusing on improving the mean accuracy alone is not enough, especially in real-world applications." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 288, + 289, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 288, + 289, + 479 + ], + "spans": [ + { + "bbox": [ + 46, + 288, + 289, + 479 + ], + "type": "text", + "content": "In this paper, we propose a novel method to make sure that no category is left behind. Specifically, we argue that although mean accuracy is widely used in image classification as an optimization objective, due to the fact that different classes have very different recall in long-tailed recognition, it is not the most suitable objective as it incurs little penalty even if some categories have very small per-class accuracy values (e.g., close to 0). Hence, to improve even the worst-performing categories, we believe the harmonic mean of per-class recall would be a better objective. Since harmonic mean is very sensitive to small numbers, further improving classes that already have high recall brings little benefits to the harmonic mean, which makes sure no single class will be left behind. Also, now all classes are treated equally, forcing us to improve classes that have low recall no matter which subset it belongs to." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 479, + 289, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 479, + 289, + 624 + ], + "spans": [ + { + "bbox": [ + 46, + 479, + 289, + 624 + ], + "type": "text", + "content": "However, it is difficult to directly minimize the harmonic mean. We therefore propose a novel loss function that maximizes the geometric mean instead, which can be viewed as a surrogate. Our method serves as a simple plug-in that can be used together with both baseline and various state-of-the-art methods. We also propose an ensemble trick that uses the pre-trained and fine-tuned models together to make predictions during inference with nearly no extra cost. We are able to yield a more uniform distribution of recall across categories (i.e., no one left behind), which achieves higher harmonic mean of the recall while the (arithmetic) mean accuracy remains high, too." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 624, + 280, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 624, + 280, + 635 + ], + "spans": [ + { + "bbox": [ + 59, + 624, + 280, + 635 + ], + "type": "text", + "content": "In summary, our work has the following contributions:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 644, + 287, + 714 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 58, + 644, + 287, + 679 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 644, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 58, + 644, + 287, + 679 + ], + "type": "text", + "content": "- We are the first to emphasize the importance of the correct recognition of all categories in long-tailed recognition." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 58, + 689, + 287, + 714 + ], + "type": "text", + "content": "- We propose a novel method that aims at increasing the harmonic mean of per-class recall as well as an ensem" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 324, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 324, + 72, + 545, + 96 + ], + "type": "text", + "content": "ble trick that combines two existing models together during inference with nearly no extra cost." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 317, + 106, + 547, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 106, + 547, + 143 + ], + "spans": [ + { + "bbox": [ + 317, + 106, + 547, + 143 + ], + "type": "text", + "content": "- We experimented on three widely used benchmark datasets, which justify the effectiveness of our method in terms of both overall and worst per-class accuracy." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 154, + 392, + 167 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 154, + 392, + 167 + ], + "spans": [ + { + "bbox": [ + 306, + 154, + 392, + 167 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 175, + 547, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 175, + 547, + 236 + ], + "spans": [ + { + "bbox": [ + 304, + 175, + 547, + 236 + ], + "type": "text", + "content": "Long-tailed recognition is a research field that aims at training models using imbalanced datasets. Since it is widely believed that the learning procedure is dominated by head classes, most existing works focus explicitly on improving the recognition of tail classes." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 244, + 474, + 257 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 244, + 474, + 257 + ], + "spans": [ + { + "bbox": [ + 305, + 244, + 474, + 257 + ], + "type": "text", + "content": "2.1. Re-sampling and Re-weighting" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 263, + 547, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 263, + 547, + 407 + ], + "spans": [ + { + "bbox": [ + 304, + 263, + 547, + 407 + ], + "type": "text", + "content": "Re-sampling and re-weighting are two classical approaches. Re-sampling methods aim at balancing the training set by either over-sampling the tail classes [3, 9, 21, 24] or under-sampling the head classes [7, 10]. Some methods also transfer statistics from the major classes to minor classes to obtain a balanced dataset [15, 19]. One disadvantage of re-sampling methods is that it may lead to either over-fitting of tail classes or under-fitting of head classes. Re-weighting methods, on the other hand, give each instance a weight based on its true label when computing the loss [2, 22, 23]. The major drawback of re-weighting is that it makes the loss function hard to optimize." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 415, + 434, + 428 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 415, + 434, + 428 + ], + "spans": [ + { + "bbox": [ + 305, + 415, + 434, + 428 + ], + "type": "text", + "content": "2.2. Two-Stage Decoupling" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 434, + 547, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 434, + 547, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 434, + 547, + 567 + ], + "type": "text", + "content": "Due to the shortcomings of re-sampling and reweighting, various other methods are proposed recently. For example, two-stage methods like [14, 18, 30] propose to decouple the learning of features and classifiers and achieve impressive results. Our method is similar to them in the sense that we also re-train a classifier given an existing model. But our motivation is to design a simple and flexible method that can be applied to a wide variety of existing models, and we propose a novel loss function, while two-stage methods like cRT [14] usually use the conventional cross entropy loss in the fine-tuning stage." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 574, + 459, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 574, + 459, + 588 + ], + "spans": [ + { + "bbox": [ + 306, + 574, + 459, + 588 + ], + "type": "text", + "content": "2.3. Hybrid and Multiple Heads" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "type": "text", + "content": "Hybrid methods, like [26, 31], make use of multiple heads to improve the recognition of different classes, which is similar to our proposed ensemble trick. But these methods require a joint training of multiple heads, together with a complex routing module that dynamically determines the head to use during inference, which increases the complexity of the model and makes it hard to use in practice. Our method, on the other hand, is simpler and requires no extra training nor complex modules to combine the predictions, which is different from them." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15805" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 136, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 136, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 136, + 83 + ], + "type": "text", + "content": "2.4. Miscellaneous" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 91, + 289, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 91, + 289, + 236 + ], + "spans": [ + { + "bbox": [ + 46, + 91, + 289, + 236 + ], + "type": "text", + "content": "There are many approaches that use knowledge distillation [12, 18] and contrastive learning [4, 17, 25, 28, 33] and achieve descent results. There are also methods that perform logit adjustment [20, 29], which formulate long-tailed recognition as a label distribution shift problem. Recently, there is also one work that tries to solve the problem by performing some regularization on the classifier's weight [1]. They also visualize the recall of each class in their work and find the model is biased towards common classes, which is similar to our motivation. However, they do not emphasize the importance of correctly recognizing all classes, which is different from us." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 249, + 103, + 261 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 249, + 103, + 261 + ], + "spans": [ + { + "bbox": [ + 47, + 249, + 103, + 261 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 270, + 288, + 318 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 270, + 288, + 318 + ], + "spans": [ + { + "bbox": [ + 46, + 270, + 288, + 318 + ], + "type": "text", + "content": "We will present our method in this section. First, we will introduce the notations used. Then we will describe the major framework of our method, which includes our proposed loss function and a simple ensemble trick." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 328, + 115, + 339 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 328, + 115, + 339 + ], + "spans": [ + { + "bbox": [ + 47, + 328, + 115, + 339 + ], + "type": "text", + "content": "3.1. Notations" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 347, + 287, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 347, + 287, + 444 + ], + "spans": [ + { + "bbox": [ + 47, + 347, + 287, + 444 + ], + "type": "text", + "content": "The neural network can be represented as a non-linear function " + }, + { + "bbox": [ + 47, + 347, + 287, + 444 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 47, + 347, + 287, + 444 + ], + "type": "text", + "content": ". Given the " + }, + { + "bbox": [ + 47, + 347, + 287, + 444 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 347, + 287, + 444 + ], + "type": "text", + "content": "-th training image " + }, + { + "bbox": [ + 47, + 347, + 287, + 444 + ], + "type": "inline_equation", + "content": "(x^i, y^i)" + }, + { + "bbox": [ + 47, + 347, + 287, + 444 + ], + "type": "text", + "content": " from the imbalanced dataset that has " + }, + { + "bbox": [ + 47, + 347, + 287, + 444 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 47, + 347, + 287, + 444 + ], + "type": "text", + "content": " images, where " + }, + { + "bbox": [ + 47, + 347, + 287, + 444 + ], + "type": "inline_equation", + "content": "x^i \\in \\mathbb{R}^{H \\times W \\times D}" + }, + { + "bbox": [ + 47, + 347, + 287, + 444 + ], + "type": "text", + "content": ", a forward propagation through the network yields " + }, + { + "bbox": [ + 47, + 347, + 287, + 444 + ], + "type": "inline_equation", + "content": "\\operatorname{logit} o^i = \\mathcal{F}(x^i)" + }, + { + "bbox": [ + 47, + 347, + 287, + 444 + ], + "type": "text", + "content": ". For a multi-class classification problem with " + }, + { + "bbox": [ + 47, + 347, + 287, + 444 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 47, + 347, + 287, + 444 + ], + "type": "text", + "content": " categories, a softmax activation function is usually applied and the result can be represented as " + }, + { + "bbox": [ + 47, + 347, + 287, + 444 + ], + "type": "inline_equation", + "content": "\\tilde{p}^i = \\operatorname{softmax}(o^i)" + }, + { + "bbox": [ + 47, + 347, + 287, + 444 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 347, + 287, + 444 + ], + "type": "inline_equation", + "content": "\\tilde{p}^i \\in \\mathbb{R}^C" + }, + { + "bbox": [ + 47, + 347, + 287, + 444 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 453, + 190, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 453, + 190, + 467 + ], + "spans": [ + { + "bbox": [ + 47, + 453, + 190, + 467 + ], + "type": "text", + "content": "3.2. A Simple Plug-In Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 473, + 287, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 473, + 287, + 592 + ], + "spans": [ + { + "bbox": [ + 46, + 473, + 287, + 592 + ], + "type": "text", + "content": "Over the past few years, various methods [4,14,22] have been proposed to tackle the challenges in long-tailed recognition. Although our evaluation results in Tab. 1 reveal that no obvious improvements have been made regarding the worst-performing category, many of their design choices are critical and effective in long-tailed recognition. Therefore, in order to make full use of existing advances, we do not aim at building up a whole new training framework, but rather propose a simple plug-in method that can be applied to a wide range of existing methods." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": "Our method is simple. Given a model trained using either one of existing methods, we re-initialize the last FC layer and re-train it using our proposed loss function while the backbone is frozen. We borrow this idea from many two-stage decoupling methods [14, 30], which is effective and flexible, making our method widely applicable. However, unlike normal decoupling methods that use the conventional cross entropy loss in the second stage, we propose a novel loss function to improve the worst-performing categories. Our proposed loss function, called GML (Geometric" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 72, + 411, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 411, + 83 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 411, + 83 + ], + "type": "text", + "content": "Mean Loss), is defined as" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 375, + 89, + 545, + 121 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 89, + 545, + 121 + ], + "spans": [ + { + "bbox": [ + 375, + 89, + 545, + 121 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {G M L}} = - \\frac {1}{C} \\sum_ {c = 1} ^ {C} \\log \\bar {\\boldsymbol {p}} _ {c}, \\tag {1}", + "image_path": "2b5e375135e1e16ba256fd7b3b5e445f3fe7babdd25e0690ed40eadc6111c862.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 125, + 410, + 138 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 125, + 410, + 138 + ], + "spans": [ + { + "bbox": [ + 306, + 125, + 410, + 138 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 125, + 410, + 138 + ], + "type": "inline_equation", + "content": "\\bar{\\pmb{p}}_c" + }, + { + "bbox": [ + 306, + 125, + 410, + 138 + ], + "type": "text", + "content": " is computed as" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 388, + 142, + 545, + 174 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 142, + 545, + 174 + ], + "spans": [ + { + "bbox": [ + 388, + 142, + 545, + 174 + ], + "type": "interline_equation", + "content": "\\bar {\\boldsymbol {p}} _ {c} = \\frac {1}{N _ {c}} \\sum_ {i = 1} ^ {N _ {c}} \\tilde {\\boldsymbol {p}} _ {y ^ {i}} ^ {i}, \\tag {2}", + "image_path": "786475ea232bc1550e73d2e394510b642b2b20fed7cdbc10245a7b2ab85fd8fd.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 180, + 547, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 180, + 547, + 231 + ], + "spans": [ + { + "bbox": [ + 304, + 180, + 547, + 231 + ], + "type": "text", + "content": "which is the average of " + }, + { + "bbox": [ + 304, + 180, + 547, + 231 + ], + "type": "inline_equation", + "content": "\\tilde{\\pmb{p}}_{y^i}^i" + }, + { + "bbox": [ + 304, + 180, + 547, + 231 + ], + "type": "text", + "content": " across all training samples belonging to class " + }, + { + "bbox": [ + 304, + 180, + 547, + 231 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 180, + 547, + 231 + ], + "type": "text", + "content": " in this mini-batch, and " + }, + { + "bbox": [ + 304, + 180, + 547, + 231 + ], + "type": "inline_equation", + "content": "N_{c}" + }, + { + "bbox": [ + 304, + 180, + 547, + 231 + ], + "type": "text", + "content": " is the number of examples from this category in the mini-batch. Inspired by [22], we perform re-weighting when we compute " + }, + { + "bbox": [ + 304, + 180, + 547, + 231 + ], + "type": "inline_equation", + "content": "\\tilde{\\pmb{p}}^i" + }, + { + "bbox": [ + 304, + 180, + 547, + 231 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 375, + 234, + 545, + 266 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 234, + 545, + 266 + ], + "spans": [ + { + "bbox": [ + 375, + 234, + 545, + 266 + ], + "type": "interline_equation", + "content": "\\tilde {\\boldsymbol {p}} _ {j} ^ {i} = \\frac {N _ {j} \\exp \\left(o _ {j} ^ {i}\\right)}{\\sum_ {c = 1} ^ {C} N _ {c} \\exp \\left(o _ {c} ^ {i}\\right)}. \\tag {3}", + "image_path": "894179fe545721b27ee52f845db4381dd5e7754dff7b5647e8723712c3123c9e.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 269, + 419, + 282 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 269, + 419, + 282 + ], + "spans": [ + { + "bbox": [ + 306, + 269, + 419, + 282 + ], + "type": "text", + "content": "3.3. Why GML Works?" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 318, + 288, + 460, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 288, + 460, + 300 + ], + "spans": [ + { + "bbox": [ + 318, + 288, + 460, + 300 + ], + "type": "text", + "content": "The definition of mean accuracy is" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 356, + 305, + 545, + 337 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 305, + 545, + 337 + ], + "spans": [ + { + "bbox": [ + 356, + 305, + 545, + 337 + ], + "type": "interline_equation", + "content": "\\operatorname {A c c} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\mathbb {I} (y ^ {i} = \\arg \\max \\tilde {\\boldsymbol {p}} ^ {i}), \\tag {4}", + "image_path": "3d1b59af5891d96029e6fe045ae92a1f9d8c0b774c345d0742bf4b673ff6fa09.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 342, + 547, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 342, + 547, + 413 + ], + "spans": [ + { + "bbox": [ + 304, + 342, + 547, + 413 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 342, + 547, + 413 + ], + "type": "inline_equation", + "content": "\\mathbb{1}(\\cdot)" + }, + { + "bbox": [ + 304, + 342, + 547, + 413 + ], + "type": "text", + "content": " is the indicator function. On a balanced test set, it is the same as the arithmetic mean of per-class recall. We argue that maximizing the mean accuracy inevitably ignores some categories because arithmetic mean incurs little punishment to categories that have very small recall values. On the other hand, harmonic mean, defined as" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 356, + 416, + 545, + 444 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 416, + 545, + 444 + ], + "spans": [ + { + "bbox": [ + 356, + 416, + 545, + 444 + ], + "type": "interline_equation", + "content": "\\mathrm {H M} \\left(x _ {1}, \\dots , x _ {n}\\right) = \\frac {n}{\\frac {1}{x _ {1}} + \\cdots + \\frac {1}{x _ {n}}} \\tag {5}", + "image_path": "6c8187441ac48112048b36e3e4c50cbb44d1ede94a7d0be446b329e10dbe373d.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 447, + 545, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 447, + 545, + 518 + ], + "spans": [ + { + "bbox": [ + 304, + 447, + 545, + 518 + ], + "type": "text", + "content": "incurs strong punishment to small values. Therefore, we believe it can be a better alternative to optimize for. However, harmonic mean is defined using reciprocal, making it hard and numeric unstable to be optimized. To this extent, we propose to maximize the geometric mean of per-class recall, defined as" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 351, + 523, + 545, + 537 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 523, + 545, + 537 + ], + "spans": [ + { + "bbox": [ + 351, + 523, + 545, + 537 + ], + "type": "interline_equation", + "content": "\\operatorname {G M} \\left(x _ {1}, \\dots , x _ {n}\\right) = \\sqrt [ n ]{\\left| x _ {1} \\times \\dots \\times x _ {n} \\right|}. \\tag {6}", + "image_path": "165840b8c189db850c6750732acb27afd0bb642f357145f5b1f7cfe5c65a9557.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 541, + 545, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 541, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 304, + 541, + 545, + 590 + ], + "type": "text", + "content": "As an illustration, " + }, + { + "bbox": [ + 304, + 541, + 545, + 590 + ], + "type": "inline_equation", + "content": "\\frac{0.01 + 0.99}{2} = 0.5" + }, + { + "bbox": [ + 304, + 541, + 545, + 590 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 541, + 545, + 590 + ], + "type": "inline_equation", + "content": "\\mathrm{HM}(0.01, 0.99) = 0.02" + }, + { + "bbox": [ + 304, + 541, + 545, + 590 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 541, + 545, + 590 + ], + "type": "inline_equation", + "content": "\\mathrm{GM}(0.01, 0.99) = 0.10" + }, + { + "bbox": [ + 304, + 541, + 545, + 590 + ], + "type": "text", + "content": ". Both harmonic and geometric means are heavily affected by the small value (0.01), while the arithmetic mean is much less sensitive to it." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 590, + 545, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 590, + 545, + 625 + ], + "spans": [ + { + "bbox": [ + 304, + 590, + 545, + 625 + ], + "type": "text", + "content": "Specifically, given the per-class recall as " + }, + { + "bbox": [ + 304, + 590, + 545, + 625 + ], + "type": "inline_equation", + "content": "r_1, \\ldots, r_C" + }, + { + "bbox": [ + 304, + 590, + 545, + 625 + ], + "type": "text", + "content": ", we transform their geometric mean using a simple logarithm transformation:" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 351, + 628, + 545, + 712 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 628, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 351, + 628, + 545, + 712 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\sqrt {r _ {1} \\dots r _ {C}} = \\exp \\left(\\log \\left(r _ {1} \\dots r _ {C}\\right) ^ {1 / C}\\right) (7) \\\\ = \\exp \\left(\\frac {1}{C} \\log \\left(r _ {1} \\dots r _ {C}\\right)\\right) (8) \\\\ = \\exp \\left(\\frac {1}{C} \\sum_ {i = 1} ^ {C} \\log r _ {i}\\right). (9) \\\\ \\end{array}", + "image_path": "86c5399516dbd8435957f9c085a9e6c1c34c9096bac23546bee4acc3f490dd04.jpg" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15806" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 72, + 272, + 214 + ], + "blocks": [ + { + "bbox": [ + 63, + 72, + 272, + 214 + ], + "lines": [ + { + "bbox": [ + 63, + 72, + 272, + 214 + ], + "spans": [ + { + "bbox": [ + 63, + 72, + 272, + 214 + ], + "type": "image", + "image_path": "f099466616702b441e44f2378abea8fe708b6035da967be1caa41402c7ffedb0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 224, + 288, + 280 + ], + "lines": [ + { + "bbox": [ + 46, + 224, + 288, + 280 + ], + "spans": [ + { + "bbox": [ + 46, + 224, + 288, + 280 + ], + "type": "text", + "content": "Figure 2. Bar plot of per-class recall on the imbalanced CIFAR10 (with imbalance ratio 100) before ('CE') and after ('CE+GML') the fine-tuning. For the fine-tuned model, the recall of the first 5 classes dropped while the recall of the latter 5 classes increased, motivating us to combine the prediction of both models." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 304, + 287, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 304, + 287, + 352 + ], + "spans": [ + { + "bbox": [ + 46, + 304, + 287, + 352 + ], + "type": "text", + "content": "Since recall is computed by computing the mean of a series of indicator functions and these indicator functions only take value 0 or 1 and is non-differentiable, we use " + }, + { + "bbox": [ + 46, + 304, + 287, + 352 + ], + "type": "inline_equation", + "content": "\\bar{\\pmb{p}}_c" + }, + { + "bbox": [ + 46, + 304, + 287, + 352 + ], + "type": "text", + "content": " as a surrogate for " + }, + { + "bbox": [ + 46, + 304, + 287, + 352 + ], + "type": "inline_equation", + "content": "r_c" + }, + { + "bbox": [ + 46, + 304, + 287, + 352 + ], + "type": "text", + "content": " during training, resulting in" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 111, + 362, + 287, + 376 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 362, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 111, + 362, + 287, + 376 + ], + "type": "interline_equation", + "content": "\\sqrt {r _ {1} \\dots r _ {C}} \\propto \\exp (- \\mathcal {L} _ {\\mathrm {G M L}}). \\tag {10}", + "image_path": "e56237784bc2eef5c436e80e61f661f585c7fef935b95038a6c8a85991815ae4.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 388, + 287, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 388, + 287, + 424 + ], + "spans": [ + { + "bbox": [ + 47, + 388, + 287, + 424 + ], + "type": "text", + "content": "So, by minimizing " + }, + { + "bbox": [ + 47, + 388, + 287, + 424 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{GML}}" + }, + { + "bbox": [ + 47, + 388, + 287, + 424 + ], + "type": "text", + "content": ", we are effectively maximizing the geometric mean of per-class recall, thus being able to improve on the worst-performing categories." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 434, + 257, + 448 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 434, + 257, + 448 + ], + "spans": [ + { + "bbox": [ + 47, + 434, + 257, + 448 + ], + "type": "text", + "content": "3.4. Combining the Strength of Both Worlds" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 453, + 287, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 453, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 46, + 453, + 287, + 548 + ], + "type": "text", + "content": "By fine-tuning the pre-trained model with our proposed loss function, indeed we are able to obtain a more uniform distribution of recall values across categories. As is shown in Fig. 2, the recall of classes that perform worse before are improved while the recall of classes that perform well before are dropped. Motivated by these dynamics, we then propose a simple ensemble trick that aims at combining the strength of both worlds." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 550, + 287, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 550, + 287, + 610 + ], + "spans": [ + { + "bbox": [ + 46, + 550, + 287, + 610 + ], + "type": "text", + "content": "Our idea is simple again. Since we re-initialize the classifier during fine-tuning and the backbone is kept frozen, the old classifier is still usable during inference. Let these two models be " + }, + { + "bbox": [ + 46, + 550, + 287, + 610 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{old}}" + }, + { + "bbox": [ + 46, + 550, + 287, + 610 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 550, + 287, + 610 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{new}}" + }, + { + "bbox": [ + 46, + 550, + 287, + 610 + ], + "type": "text", + "content": ", respectively, we then make predictions by" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 113, + 620, + 287, + 643 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 620, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 113, + 620, + 287, + 643 + ], + "type": "interline_equation", + "content": "\\tilde {\\boldsymbol {p}} _ {\\text {e n s e m b l e d}} = \\frac {\\tilde {\\boldsymbol {p}} _ {\\text {n e w}} + \\tilde {\\boldsymbol {p}} _ {\\text {o l d}}}{2}. \\tag {11}", + "image_path": "83c4397a4a07c34b9a15d07f2a864a1cd38a3d9505b6863ec56aa756476e1041.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": "In practice, since the calibration of some models can be poor [8,30] (e.g., the baseline using cross entropy), we need to manually calibrate the prediction before combining them. For simplicity, we choose temperature scaling to calibrate the model [8] and use two temperature variables " + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{new}}" + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": " and" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 72, + 494, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 494, + 85 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 494, + 85 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{old}}" + }, + { + "bbox": [ + 305, + 72, + 494, + 85 + ], + "type": "text", + "content": " when computing " + }, + { + "bbox": [ + 305, + 72, + 494, + 85 + ], + "type": "inline_equation", + "content": "\\tilde{\\pmb{p}}_{\\mathrm{new}}" + }, + { + "bbox": [ + 305, + 72, + 494, + 85 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 72, + 494, + 85 + ], + "type": "inline_equation", + "content": "\\tilde{\\pmb{p}}_{\\mathrm{old}}" + }, + { + "bbox": [ + 305, + 72, + 494, + 85 + ], + "type": "text", + "content": " separately:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 374, + 93, + 545, + 120 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 374, + 93, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 374, + 93, + 545, + 120 + ], + "type": "interline_equation", + "content": "\\tilde {\\boldsymbol {p}} _ {\\text {n e w}} = \\operatorname {s o f t m a x} \\left(\\frac {\\boldsymbol {o} _ {\\text {n e w}}}{t _ {\\text {n e w}}}\\right), \\tag {12}", + "image_path": "c10d3c13dfa7f2ad879cac748c001851aaf9bce95a8a0838f71660342571239e.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 376, + 122, + 545, + 148 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 122, + 545, + 148 + ], + "spans": [ + { + "bbox": [ + 376, + 122, + 545, + 148 + ], + "type": "interline_equation", + "content": "\\tilde {\\boldsymbol {p}} _ {\\text {o l d}} = \\operatorname {s o f t m a x} \\left(\\frac {\\boldsymbol {o} _ {\\text {o l d}}}{t _ {\\text {o l d}}}\\right). \\tag {13}", + "image_path": "330c039b7644aa628ed64c2ac10780d62b83281093fe3e705658450e0d285865.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 157, + 545, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 157, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 304, + 157, + 545, + 240 + ], + "type": "text", + "content": "One big advantage of our proposed ensemble trick is that it brings nearly no extra cost. Unlike many multi-head [31] or multi-expert [26] methods, we do not need extra training as we simply use the classifier that comes with the pretrained model. During inference, the only extra cost is the fully connected layer to compute " + }, + { + "bbox": [ + 304, + 157, + 545, + 240 + ], + "type": "inline_equation", + "content": "\\tilde{\\pmb{p}}_{\\mathrm{old}}" + }, + { + "bbox": [ + 304, + 157, + 545, + 240 + ], + "type": "text", + "content": ", which is negligible. Also, no complex routing rules are needed." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 241, + 545, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 241, + 545, + 324 + ], + "spans": [ + { + "bbox": [ + 304, + 241, + 545, + 324 + ], + "type": "text", + "content": "The two temperature variables are the only two hyperparameters introduced by our method, and they can be easily tuned as no training is required. Plus, as we will later show in the ablation studies, most values work perfectly fine so actually no painful tuning is needed. What's more, this ensemble trick is completely optional, our method works well enough even without it." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 333, + 410, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 333, + 410, + 346 + ], + "spans": [ + { + "bbox": [ + 306, + 333, + 410, + 346 + ], + "type": "text", + "content": "3.5. Training Pipeline" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 352, + 545, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 352, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 352, + 545, + 449 + ], + "type": "text", + "content": "Our method is very similar to the two-stage decoupling methods [14, 30]. We design it in this way on purpose because we want it to be as flexible and simple as possible. During experiments, we found that our method is orthogonal to many design choices in state-of-the-art methods and is able to improve the worse-case performance when applied to various different pre-trained models. The overall training pipeline of our method consists of three stages:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 460, + 545, + 550 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 314, + 460, + 545, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 460, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 314, + 460, + 545, + 483 + ], + "type": "text", + "content": "1. Pre-training stage: Obtain the pre-trained model from scratch using any one of the existing methods." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 493, + 544, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 493, + 544, + 517 + ], + "spans": [ + { + "bbox": [ + 313, + 493, + 544, + 517 + ], + "type": "text", + "content": "2. Fine-tuning stage: Freeze the backbone and re-train the classifier using our proposed loss function." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 526, + 545, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 526, + 545, + 550 + ], + "spans": [ + { + "bbox": [ + 313, + 526, + 545, + 550 + ], + "type": "text", + "content": "3. An optional ensemble stage: Calibrate the prediction of two classifiers and combine them additively." + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 561, + 545, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 561, + 545, + 584 + ], + "spans": [ + { + "bbox": [ + 306, + 561, + 545, + 584 + ], + "type": "text", + "content": "The overall training procedure is summarized in Algorithm 1." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 597, + 386, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 597, + 386, + 611 + ], + "spans": [ + { + "bbox": [ + 306, + 597, + 386, + 611 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 617, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 712 + ], + "type": "text", + "content": "In order to validate the effectiveness of our method, we have conducted experiments on three widely used benchmark datasets. In this section, we will present our experimental results. First, we will introduce the datasets used. Then we will describe various implementation details and evaluation metrics used. After that, we will compare our method with various other methods. Finally, we will present the results of some ablation studies." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15807" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 108, + 70, + 487, + 127 + ], + "blocks": [ + { + "bbox": [ + 108, + 70, + 487, + 127 + ], + "lines": [ + { + "bbox": [ + 108, + 70, + 487, + 127 + ], + "spans": [ + { + "bbox": [ + 108, + 70, + 487, + 127 + ], + "type": "table", + "html": "
DatasetNumber of Classes# Training Images# Test ImagesImbalance Ratio
CIFAR100-LT [16]10010,84710,000100
ImageNet-LT [6, 19]1,000115,84650,000256
Places-LT [32]36562,50036,500996
", + "image_path": "36ad8283f9997de5bbf19f30b2284556c08c3ed1116249dafa37f0705c2c405c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 165, + 136, + 426, + 147 + ], + "lines": [ + { + "bbox": [ + 165, + 136, + 426, + 147 + ], + "spans": [ + { + "bbox": [ + 165, + 136, + 426, + 147 + ], + "type": "text", + "content": "Table 2. Statistics of three imbalanced datasets used in our experiments." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 48, + 167, + 226, + 179 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 167, + 226, + 179 + ], + "spans": [ + { + "bbox": [ + 48, + 167, + 226, + 179 + ], + "type": "text", + "content": "Algorithm 1 The overall training procedure" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 183, + 228, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 183, + 228, + 194 + ], + "spans": [ + { + "bbox": [ + 49, + 183, + 228, + 194 + ], + "type": "text", + "content": "Input: Training images " + }, + { + "bbox": [ + 49, + 183, + 228, + 194 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 49, + 183, + 228, + 194 + ], + "type": "text", + "content": " and their labels " + }, + { + "bbox": [ + 49, + 183, + 228, + 194 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 49, + 183, + 228, + 194 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 196, + 286, + 289 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 53, + 196, + 286, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 196, + 286, + 217 + ], + "spans": [ + { + "bbox": [ + 53, + 196, + 286, + 217 + ], + "type": "text", + "content": "1: Randomly initialize the network and train it from scratch to obtain the pre-trained model." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 220, + 271, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 220, + 271, + 230 + ], + "spans": [ + { + "bbox": [ + 53, + 220, + 271, + 230 + ], + "type": "text", + "content": "2: Use the pre-trained model to perform initialization." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 232, + 269, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 232, + 269, + 241 + ], + "spans": [ + { + "bbox": [ + 53, + 232, + 269, + 241 + ], + "type": "text", + "content": "3: Freeze the backbone and re-initialize the classifier." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 243, + 286, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 243, + 286, + 266 + ], + "spans": [ + { + "bbox": [ + 53, + 243, + 286, + 266 + ], + "type": "text", + "content": "4: Fine-tune the model using the loss function defined as Eq. (1) for a few epochs." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 267, + 286, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 267, + 286, + 289 + ], + "spans": [ + { + "bbox": [ + 53, + 267, + 286, + 289 + ], + "type": "text", + "content": "5: During inference, combine the results from both classifiers as defined in Eq. (11)." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 316, + 110, + 327 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 316, + 110, + 327 + ], + "spans": [ + { + "bbox": [ + 47, + 316, + 110, + 327 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 335, + 287, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 335, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 46, + 335, + 287, + 430 + ], + "type": "text", + "content": "We use three widely used long-tailed image recognition datasets: CIFAR100-LT, ImageNet-LT and PlacesLT. Some statistics about these three datasets are summarized in Tab. 2. Since CIFAR100 [16], ImageNet [6] and Places [32] are all balanced datasets, we follow previous works [19, 31] to down-sample their original training set. Details about the construction process can be found in the supplementary material." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 439, + 180, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 439, + 180, + 451 + ], + "spans": [ + { + "bbox": [ + 47, + 439, + 180, + 451 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 458, + 287, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 458, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 46, + 458, + 287, + 578 + ], + "type": "text", + "content": "Following previous works [19, 31], we use ResNet-32 [11], ResNeXt-50 [27] and ResNet-152 as the backbone network for CIFAR100-LT, ImageNet-LT and Places-LT, respectively. For the pre-training stage, if there exists released checkpoints, we use them directly. If not, we use the officially released codes and try to reproduce the results in our best effort without changing their original settings. This includes stuffs like hyper-parameters and data augmentations used. For the fine-tuning stage, detailed training settings can be found in the supplementary material." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 586, + 285, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 586, + 285, + 600 + ], + "spans": [ + { + "bbox": [ + 47, + 586, + 285, + 600 + ], + "type": "text", + "content": "4.3. Evaluation Metrics and Comparison Methods" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": "The convention in long-tailed recognition research is to split the classes into three subsets based on the number of training images [14, 19]. Typically, \"Head\" denotes classes that have more than 100 images, \"Few\" denotes classes that have less than 20 images and all other classes are denoted as \"Medium\". The accuracy within each subset is often reported along with the overall accuracy to justify the claim that the performance improvement mainly comes from tail classes. And according to our observation in Sec. 1, we ar" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 168, + 545, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 168, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 304, + 168, + 545, + 239 + ], + "type": "text", + "content": "gue that such an evaluation scheme can be problematic and we propose to first compute the recall within each category and then compute their harmonic mean. In our experiments, since our method optimizes geometric mean of per-class recall as a surrogate, we also report the geometric mean. Besides, we report the overall mean accuracy as well." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 239, + 545, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 239, + 545, + 334 + ], + "spans": [ + { + "bbox": [ + 304, + 239, + 545, + 334 + ], + "type": "text", + "content": "We compare our methods with both baseline methods like simple cross entropy loss and various state-of-the-arts methods [4, 12, 22, 26]. Many of these methods have very different training settings and strictly speaking, they are not directly comparable. But since our method is a simple plugin method, the main focus is to quantify the improvements brought by our method compared to its respective baseline result." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 342, + 418, + 355 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 342, + 418, + 355 + ], + "spans": [ + { + "bbox": [ + 305, + 342, + 418, + 355 + ], + "type": "text", + "content": "4.4. Main Comparisons" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 361, + 545, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 361, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 305, + 361, + 545, + 384 + ], + "type": "text", + "content": "We validate our method on three benchmark datasets and present the results on each dataset here separately." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 399, + 400, + 409 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 399, + 400, + 409 + ], + "spans": [ + { + "bbox": [ + 306, + 399, + 400, + 409 + ], + "type": "text", + "content": "4.4.1 CIFAR100-LT" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 417, + 545, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 417, + 545, + 620 + ], + "spans": [ + { + "bbox": [ + 304, + 417, + 545, + 620 + ], + "type": "text", + "content": "Tab. 3 shows the experimental results on the CIFAR100-LT dataset. CIFAR100-LT has three different versions with different imbalance ratios and here we only consider the one with imbalance ratio 100. As we can see from the table, although various different methods have been proposed in recent years and there are rapid improvements on the overall accuracy, the performance of the worst-performing category is still relatively poor. The geometric and harmonic mean of per-class recall also justify our claim. We apply our method to both the baseline method CE and state-of-the-art method PaCo [4] on this dataset. When our method is applied to CE and PaCo, we are able to improve the lowest recall from 0.00 and 5.00 to 6.00 and 9.00 respectively. And if we further apply our proposed ensemble trick, we are able to improve the lowest recall to 15.00, achieving state-of-the-art performance in both geometric mean and harmonic mean while keeping the overall accuracy roughly unchanged." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 621, + 545, + 655 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 621, + 545, + 655 + ], + "spans": [ + { + "bbox": [ + 304, + 621, + 545, + 655 + ], + "type": "text", + "content": "It is worth noting that even the simplest CE+GML outperforms all existing methods in terms of lowest recall, that is, in terms of \"no one left behind\"." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 670, + 460, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 670, + 460, + 682 + ], + "spans": [ + { + "bbox": [ + 306, + 670, + 460, + 682 + ], + "type": "text", + "content": "4.4.2 ImageNet-LT and Places-LT" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "Tab. 4 shows the result on ImageNet-LT. The overall observation is similar to that on CIFAR100-LT, except that" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15808" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 117, + 70, + 478, + 224 + ], + "blocks": [ + { + "bbox": [ + 117, + 70, + 478, + 224 + ], + "lines": [ + { + "bbox": [ + 117, + 70, + 478, + 224 + ], + "spans": [ + { + "bbox": [ + 117, + 70, + 478, + 224 + ], + "type": "table", + "html": "
MethodsAccuracyGeometric MeanHarmonic MeanLowest Recall
CE38.7421.03*2.05*0.00
BSCE [22]42.2435.1626.243.00
LDAM [2]43.5133.6121.493.00
LADE [13]44.3939.0532.585.00
DiVE [12]45.1137.0825.182.00
MiSLAS [30]47.0540.1630.935.00
RIDE [26]48.6438.7123.862.00
PaCo [4]51.2445.2936.425.00
CE + GML41.0636.5931.266.00
PaCo + GML50.5345.4739.209.00
PaCo + GML (Ensemble)49.8245.7041.0215.00
", + "image_path": "63dc89b835167d8a01e3953089353bcec45bb94620c711dbeaf657141585c94c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 116, + 266, + 477, + 399 + ], + "blocks": [ + { + "bbox": [ + 46, + 233, + 547, + 257 + ], + "lines": [ + { + "bbox": [ + 46, + 233, + 547, + 257 + ], + "spans": [ + { + "bbox": [ + 46, + 233, + 547, + 257 + ], + "type": "text", + "content": "Table 3. Results on the CIFAR100-LT dataset with imbalance ratio 100. Numbers with * are computed by substituting zero elements with a small number " + }, + { + "bbox": [ + 46, + 233, + 547, + 257 + ], + "type": "inline_equation", + "content": "(10^{-3})" + }, + { + "bbox": [ + 46, + 233, + 547, + 257 + ], + "type": "text", + "content": " or else the geometric and harmonic mean will all be zero." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 116, + 266, + 477, + 399 + ], + "lines": [ + { + "bbox": [ + 116, + 266, + 477, + 399 + ], + "spans": [ + { + "bbox": [ + 116, + 266, + 477, + 399 + ], + "type": "table", + "html": "
MethodsAccuracyGeometric MeanHarmonic MeanLowest Recall
CE43.9023.25*1.25*0.00
BSCE [22]50.4842.32*13.74*0.00
cRT [14]49.6441.35*13.82*0.00
DiVE [12]53.6345.49*12.76*0.00
RIDE [26]55.6947.56*17.32*0.00
PaCo [4]58.5351.32*21.79*0.00
CE + GML45.6139.8231.672.00
PaCo + GML55.5750.6743.712.00
PaCo + GML (Ensemble)57.2252.2944.802.00
", + "image_path": "f0ed7253bcae5d238b8c0fe59b55495a0634b20a79e3f962206162af5056e076.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 408, + 547, + 431 + ], + "lines": [ + { + "bbox": [ + 46, + 408, + 547, + 431 + ], + "spans": [ + { + "bbox": [ + 46, + 408, + 547, + 431 + ], + "type": "text", + "content": "Table 4. Results on the ImageNet-LT dataset. Numbers with * are computed by substituting zero elements with a small number " + }, + { + "bbox": [ + 46, + 408, + 547, + 431 + ], + "type": "inline_equation", + "content": "(10^{-3})" + }, + { + "bbox": [ + 46, + 408, + 547, + 431 + ], + "type": "text", + "content": " or else the geometric and harmonic mean will all be zero." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 452, + 288, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 452, + 288, + 547 + ], + "spans": [ + { + "bbox": [ + 46, + 452, + 288, + 547 + ], + "type": "text", + "content": "since this dataset is larger and harder, the lowest recall of all current methods are zero, making the harmonic mean of per-class recall very low, even when a small number substitutes zero per-class recalls. Our method, on the other hand, greatly improves the harmonic mean by successfully improving the lowest recall value. All categories have per-class recall higher than 0 when our method is used, even in the simplest CE+GML." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 548, + 288, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 548, + 288, + 632 + ], + "spans": [ + { + "bbox": [ + 46, + 548, + 288, + 632 + ], + "type": "text", + "content": "Tab. 5 shows the result on Places-LT. Current state-of-the-art on this dataset is MiSLAS [30], so we apply our method upon it and the simple baseline CE. Again, our method successfully improves the lowest recall, as well as the harmonic mean of per-class recall. And when the ensemble trick is further applied, we are able to achieve new state-of-the-art results." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 647, + 133, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 133, + 659 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 133, + 659 + ], + "type": "text", + "content": "4.4.3 Universality" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": "Since we want our method to be a universal plug-in, in this subsection we present the results of the experiment of adding our method on top of various different methods. The experiment is conducted on CIFAR100-LT, and the re" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 452, + 547, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 452, + 547, + 536 + ], + "spans": [ + { + "bbox": [ + 304, + 452, + 547, + 536 + ], + "type": "text", + "content": "sult is shown in Tab. 6. Note that here we do not perform any ensemble. As we can see, our method is applicable to various different methods, and is able to achieve consistent improvements. However, we do observed that there exists some methods, like DiVE [12], where we failed to achieve obvious improvements. More details about this can be found when we discuss limitations in Sec. 5." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 542, + 407, + 554 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 542, + 407, + 554 + ], + "spans": [ + { + "bbox": [ + 306, + 542, + 407, + 554 + ], + "type": "text", + "content": "4.5. Ablation Studies" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 560, + 547, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 560, + 547, + 598 + ], + "spans": [ + { + "bbox": [ + 304, + 560, + 547, + 598 + ], + "type": "text", + "content": "For ablation studies, if not otherwise specified, all experiments are conducted on CIFAR100-LT (with imbalance ratio 100) under the default training setting." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 611, + 545, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 611, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 304, + 611, + 545, + 635 + ], + "type": "text", + "content": "4.5.1 Effects of Two Temperature Variables in the Ensemble Stage" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "content": "When we design our method, we'd like to keep the number of new hyper-parameters as less as possible in order for it to be readily applicable in real-world applications. The only two hyper-parameters introduced by our method are in the optional ensemble stage. Here we study the effects of choosing different values of them. Specifically, we change" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "15809" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 110, + 70, + 484, + 182 + ], + "blocks": [ + { + "bbox": [ + 110, + 70, + 484, + 182 + ], + "lines": [ + { + "bbox": [ + 110, + 70, + 484, + 182 + ], + "spans": [ + { + "bbox": [ + 110, + 70, + 484, + 182 + ], + "type": "table", + "html": "
MethodsAccuracyGeometric MeanHarmonic MeanLowest Recall
CE28.7112.11*0.73*0.00
BSCE [22]37.1829.30*5.64*0.00
PaCo [4]40.4527.88*2.53*0.00
MiSLAS [30]40.4835.5328.973.00
CE + GML36.8229.30*8.00*0.00
MiSLAS + GML39.9035.4129.985.00
MiSLAS + GML (Ensemble)40.3735.9230.515.00
", + "image_path": "0af05d479702aa7007b5fb5a0b2db52d53958f555d184536983fb6d936fd1f23.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 242, + 286, + 352 + ], + "blocks": [ + { + "bbox": [ + 46, + 190, + 547, + 225 + ], + "lines": [ + { + "bbox": [ + 46, + 190, + 547, + 225 + ], + "spans": [ + { + "bbox": [ + 46, + 190, + 547, + 225 + ], + "type": "text", + "content": "Table 5. Results on the Places-LT dataset. Numbers with * are computed by substituting zero elements with a small number " + }, + { + "bbox": [ + 46, + 190, + 547, + 225 + ], + "type": "inline_equation", + "content": "(10^{-3})" + }, + { + "bbox": [ + 46, + 190, + 547, + 225 + ], + "type": "text", + "content": " or else the geometric and harmonic mean will all be zero. Note that to plug GML into MiSLAS, since MiSLAS is also a two-stage method, we think it's fairer to use GML together with their proposed label aware smoothing loss." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 242, + 286, + 352 + ], + "lines": [ + { + "bbox": [ + 50, + 242, + 286, + 352 + ], + "spans": [ + { + "bbox": [ + 50, + 242, + 286, + 352 + ], + "type": "table", + "html": "
MethodsG-MeanH-MeanLowest Recall
CE21.03*2.05*0.00
CE + GML36.59 (15.56↑)31.26 (29.21↑)6.00 (6.00↑)
BSCE35.1626.243.00
BSCE + GML36.52 (1.36↑)30.88 (4.64↑)7.00 (4.00↑)
MiSLAS40.1630.935.00
MiSLAS + GML40.90 (0.74↑)36.49 (5.56↑)11.00 (6.00↑)
PaCo45.2936.425.00
PaCo + GML45.47 (0.18↑)39.20 (2.78↑)9.00 (4.00↑)
", + "image_path": "f8be096050c70dd74ff4f111237a61b406c566ff698e2b6cce26ffccba210869.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 50, + 419, + 286, + 511 + ], + "blocks": [ + { + "bbox": [ + 47, + 359, + 288, + 403 + ], + "lines": [ + { + "bbox": [ + 47, + 359, + 288, + 403 + ], + "spans": [ + { + "bbox": [ + 47, + 359, + 288, + 403 + ], + "type": "text", + "content": "Table 6. Results on the CIFAR100-LT dataset with imbalance ratio 100. Our proposed method is applicable to various methods. \"G-Mean\" is short for \"Geometric Mean\" and \"H-Mean\" is short for \"Harmonic Mean\"." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 419, + 286, + 511 + ], + "lines": [ + { + "bbox": [ + 50, + 419, + 286, + 511 + ], + "spans": [ + { + "bbox": [ + 50, + 419, + 286, + 511 + ], + "type": "table", + "html": "
toldtnewGeometric MeanHarmonic MeanLowest Recall
1145.7041.0215.00
1245.5838.868.00
1344.6836.095.00
2144.3940.2215.00
2245.6141.0115.00
2346.0640.4011.00
3143.7339.6315.00
", + "image_path": "0e4d66c3c185358a04d65da7d9281707fe092b95d00793976269f0e50f093310.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 567, + 287, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 567, + 287, + 699 + ], + "spans": [ + { + "bbox": [ + 46, + 567, + 287, + 699 + ], + "type": "text", + "content": "the value of the temperature variables shown in Eq. (12) and Eq. (13) in the ensemble stage when we apply our method to PaCo [4]. The results are shown in Tab. 7. In our experiments on CIFAR100-LT, we set " + }, + { + "bbox": [ + 46, + 567, + 287, + 699 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{old}} = 1" + }, + { + "bbox": [ + 46, + 567, + 287, + 699 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 567, + 287, + 699 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{new}} = 1" + }, + { + "bbox": [ + 46, + 567, + 287, + 699 + ], + "type": "text", + "content": ". And as we can see from the table, the value of " + }, + { + "bbox": [ + 46, + 567, + 287, + 699 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{old}}" + }, + { + "bbox": [ + 46, + 567, + 287, + 699 + ], + "type": "text", + "content": " does not affect the result much while " + }, + { + "bbox": [ + 46, + 567, + 287, + 699 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{new}}" + }, + { + "bbox": [ + 46, + 567, + 287, + 699 + ], + "type": "text", + "content": " is somewhat important, especially for improving the lowest recall—higher " + }, + { + "bbox": [ + 46, + 567, + 287, + 699 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{new}}" + }, + { + "bbox": [ + 46, + 567, + 287, + 699 + ], + "type": "text", + "content": " will hurt it. This is natural because higher " + }, + { + "bbox": [ + 46, + 567, + 287, + 699 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 567, + 287, + 699 + ], + "type": "text", + "content": " results in a smoother distribution and less information is preserved. Since the newly trained classifier is crucial for improving the lowest recall, we need to preserve more information from it." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 59, + 701, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 288, + 714 + ], + "type": "text", + "content": "Generally speaking, since these two hyper-parameters" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 309, + 242, + 544, + 294 + ], + "blocks": [ + { + "bbox": [ + 46, + 519, + 287, + 544 + ], + "lines": [ + { + "bbox": [ + 46, + 519, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 519, + 287, + 544 + ], + "type": "text", + "content": "Table 7. Ablation on the effects of temperature variables when we apply our ensemble method upon PaCo [4]." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 242, + 544, + 294 + ], + "lines": [ + { + "bbox": [ + 309, + 242, + 544, + 294 + ], + "spans": [ + { + "bbox": [ + 309, + 242, + 544, + 294 + ], + "type": "table", + "html": "
MethodsG-MeanH-MeanLowest Recall
CE + GML36.5931.266.00
(w/o re-weighting)25.455.110.00
(w/o re-weighting, re-sampling)32.3223.054.00
", + "image_path": "75bb34486dddd91ca2e99ea40ff118d28606a067d8a7d0ed620973de64316e2c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 309, + 326, + 544, + 383 + ], + "blocks": [ + { + "bbox": [ + 326, + 302, + 524, + 314 + ], + "lines": [ + { + "bbox": [ + 326, + 302, + 524, + 314 + ], + "spans": [ + { + "bbox": [ + 326, + 302, + 524, + 314 + ], + "type": "text", + "content": "Table 8. Ablation on the re-weighting design in GML." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 326, + 544, + 383 + ], + "lines": [ + { + "bbox": [ + 309, + 326, + 544, + 383 + ], + "spans": [ + { + "bbox": [ + 309, + 326, + 544, + 383 + ], + "type": "table", + "html": "
MethodsG-MeanH-MeanLowest Recall
CE + GML36.5931.266.00
PaCo + GML45.4739.209.00
GML (train from scratch)36.6330.869.00
", + "image_path": "2e35b23c511286227c97672c49256cb1cd2c36a2e2f6a3425435f4552f360ab5.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 346, + 390, + 504, + 402 + ], + "lines": [ + { + "bbox": [ + 346, + 390, + 504, + 402 + ], + "spans": [ + { + "bbox": [ + 346, + 390, + 504, + 402 + ], + "type": "text", + "content": "Table 9. Training from scratch using GML." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 424, + 545, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 424, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 304, + 424, + 545, + 498 + ], + "type": "text", + "content": "are introduced for calibration, for different methods, we may need to tune them respectively because the calibrations of different methods are known to vary a lot in long-tailed recognition problems [30]. But usually this will not cause a headache, because the ensemble stage does not involve any training and thus can be tuned very quickly." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 513, + 433, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 513, + 433, + 525 + ], + "spans": [ + { + "bbox": [ + 305, + 513, + 433, + 525 + ], + "type": "text", + "content": "4.5.2 Re-weighting in GML" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 533, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 545, + 605 + ], + "type": "text", + "content": "As shown in Eq. (3), re-weighting is applied in our proposed GML, here we ablate this design. The results are shown in Tab. 8. As we can see, the performance will be relatively poor without re-weighting. And just like many two-stage decoupling methods [14], re-sampling during fine-tuning can help, but is inferior to re-weighting." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 621, + 487, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 621, + 487, + 634 + ], + "spans": [ + { + "bbox": [ + 305, + 621, + 487, + 634 + ], + "type": "text", + "content": "4.5.3 Training from Scratch using GML" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": "Since our main purpose is to propose a simple plug-in method that is widely applicable to various existing methods, we use GML only to re-train the classifier of an existing model. However, GML can also be used to train a model from scratch. The results are shown in Tab. 9. As we can see, the result is comparable to the result when GML" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15810" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 73, + 272, + 212 + ], + "blocks": [ + { + "bbox": [ + 63, + 73, + 272, + 212 + ], + "lines": [ + { + "bbox": [ + 63, + 73, + 272, + 212 + ], + "spans": [ + { + "bbox": [ + 63, + 73, + 272, + 212 + ], + "type": "image", + "image_path": "0c9f97b185c9700a78f472c479867b19bd9e22c76213909cd608fc7f7d58bc69.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 224, + 288, + 281 + ], + "lines": [ + { + "bbox": [ + 46, + 224, + 288, + 281 + ], + "spans": [ + { + "bbox": [ + 46, + 224, + 288, + 281 + ], + "type": "text", + "content": "Figure 3. Bar plot of per-class recall on the imbalanced CIFAR100 (with imbalance ratio 100) before and after the fine-tuning. No ensemble is performed. Our method focuses on improving the recognition of classes that have low recall values, making sure no category is left behind." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 301, + 287, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 301, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 46, + 301, + 287, + 327 + ], + "type": "text", + "content": "is attached to CE, but inferior to the result when GML is applied to SOTA methods like PaCo, just as expected." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 341, + 235, + 352 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 341, + 235, + 352 + ], + "spans": [ + { + "bbox": [ + 47, + 341, + 235, + 352 + ], + "type": "text", + "content": "4.5.4 Visualization of the Per-Class Recall" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 359, + 287, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 359, + 287, + 515 + ], + "spans": [ + { + "bbox": [ + 46, + 359, + 287, + 515 + ], + "type": "text", + "content": "In Fig. 1, we visualize the per-class recall of model trained on the imbalanced dataset, and find they vary a lot from class to class. Since our major motivation is to make sure all classes are equally treated so that no category will be left behind, here we visualize the change of per-class recall after applying our method. We perform visualizing using two datasets, CIFAR10-LT and CIFAR100-LT, both have imbalance ratio 100. The visualization results on CIFAR10-LT and CIFAR100-LT are shown in Fig. 2 and Fig. 3, respectively. Note that here we do not perform the ensemble. As we can clearly see in both figures, our fine-tuned model has a more uniform distribution of per-class recall. These two figures justify the effectiveness of our proposed method." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 526, + 234, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 526, + 234, + 540 + ], + "spans": [ + { + "bbox": [ + 47, + 526, + 234, + 540 + ], + "type": "text", + "content": "5. Limitations and Broader Impacts" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 545, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 288, + 715 + ], + "type": "text", + "content": "Although our method is applicable to a wide range of existing methods, we do find that when applied to some methods, the improvement is less obvious compared to others. For example, our framework does not cope well with methods like BSCE [22] and DiVE [12]. Interestingly, although DiVE is based on knowledge distillation, BSCE is used in DiVE as well to train both the teacher and student model so we attribute the failure to the fact that these method all re-weight the loss function. Two-stage decoupling methods like cRT [14] are also known to not perform well when the first stage training involves re-weighting or re-sampling. The author argues it is because re-sampling and re-weighting have negative effects on feature learning and we believe that argument applies to our method as well." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "text", + "content": "Another thing worth noting here is that we do not report the experimental results on iNaturalist2018 [5] because we find it generally not suitable in our settings. iNaturalist2018 is another widely used benchmark dataset in long-tailed recognition. It is large-scale, consisting of 437.5K training images from 8142 classes. But for the test split, it only has 24K images, so each class only has 3 images. Since our major motivation is to make sure no category is left behind, it is generally not meaningful to look at the per-class recall in iNaturalist2018 because the variance of the result can be very large due to the very small number of test images. However, in real-world applications, it is generally rare to have such small number of test images." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 230, + 545, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 230, + 545, + 302 + ], + "spans": [ + { + "bbox": [ + 304, + 230, + 545, + 302 + ], + "type": "text", + "content": "For broader impacts, we believe our research have positive impacts on reducing the bias of model training on an imbalanced dataset. Machine learning models are known to be biased if there is lack of training images from some categories. We believe our work will be helpful to achieve better fairness by making sure all categories are equally treated." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 320, + 383, + 332 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 320, + 383, + 332 + ], + "spans": [ + { + "bbox": [ + 306, + 320, + 383, + 332 + ], + "type": "text", + "content": "6. Conclusions" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 342, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 342, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 342, + 545, + 713 + ], + "type": "text", + "content": "In this paper, we focused on improving some of the worst-performing classes in long-tailed recognition. We found that when trained with an imbalanced dataset, the per-class recall of the model varies a lot from class to class. Previous convention reports the arithmetic mean of per-class recall, but we argued such an evaluation scheme can be problematic. First, arithmetic mean only incurs little penalty to small numbers, making it too coarse to reflect whether there are categories that are left behind (i.e., with close to zero per-class accuracy). Second, it is not necessarily true that classes in the \"Few\" category perform worse than those in \"Many\" or \"Medium\", so focusing on improving \"Few\" accuracy alone is not enough. For these reasons, we argued that we should pay more attention to the harmonic mean of per-class recall and we therefore proposed a novel method that aims at improving the lowest recall and the harmonic mean of recalls. Our method is a simple plug-in that is applicable to a wide range of existing methods. Specifically, our method consists of three stages, in the first stage, we use any existing method to train a model from scratch. Then we re-train the classifier using our proposed novel GML loss function. Finally, we propose a simple ensemble trick that can be used to combine the predictions from the two classifiers with nearly no extra cost. We validated the effectiveness of our method on three widely used benchmark datasets, and witnessed consistent improvements on the harmonic mean of recalls and lowest recall value, while the overall accuracy still remains high. By visualizing the distribution of per-class recall values of the fine-tuned model, we found our model indeed achieved a more balanced distribution." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15811" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Shaden Alshammari, Yu-Xiong Wang, Deva Ramanan, and Shu Kong. Long-tailed recognition via weight balancing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6897-6907, June 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 148, + 288, + 202 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 288, + 202 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 288, + 202 + ], + "type": "text", + "content": "[2] Kaidi Cao, Colin Wei, Adrien Gaidon, Nikos Arechiga, and Tengyu Ma. Learning imbalanced datasets with label-distribution-aware margin loss. In Advances in Neural Information Processing Systems 32, pages 1567-1578, December 2019. 2, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 205, + 288, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 205, + 288, + 247 + ], + "spans": [ + { + "bbox": [ + 53, + 205, + 288, + 247 + ], + "type": "text", + "content": "[3] Nitesh V Chawla, Kevin W Bowyer, Lawrence O Hall, and W Philip Kegelmeyer. SMOTE: Synthetic minority oversampling technique. Journal of Artificial Intelligence Research, 16:321-357, 2002. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 250, + 287, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 250, + 287, + 294 + ], + "spans": [ + { + "bbox": [ + 53, + 250, + 287, + 294 + ], + "type": "text", + "content": "[4] Jiequan Cui, Zhisheng Zhong, Shu Liu, Bei Yu, and Jiaya Jia. Parametric contrastive learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 715-724, October 2021. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 295, + 287, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 295, + 287, + 350 + ], + "spans": [ + { + "bbox": [ + 53, + 295, + 287, + 350 + ], + "type": "text", + "content": "[5] Yin Cui, Yang Song, Chen Sun, Andrew Howard, and Serge Belongie. Large scale fine-grained categorization and domain-specific transfer learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4109-4118, June 2018. 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 351, + 287, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 351, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 53, + 351, + 287, + 396 + ], + "type": "text", + "content": "[6] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248-255, June 2009. 1, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 397, + 287, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 397, + 287, + 440 + ], + "spans": [ + { + "bbox": [ + 53, + 397, + 287, + 440 + ], + "type": "text", + "content": "[7] Chris Drummond and Robert C. Holte. C4.5, class imbalance, and cost sensitivity: Why under-sampling beats oversampling. In Workshop on Learning from Imbalanced Data Sets II, 2003. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 442, + 287, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 442, + 287, + 486 + ], + "spans": [ + { + "bbox": [ + 53, + 442, + 287, + 486 + ], + "type": "text", + "content": "[8] Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q. Weinberger. On calibration of modern neural networks. In Proceedings of the 34th International Conference on Machine Learning, pages 1321-1330, August 2017. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 487, + 287, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 487, + 287, + 532 + ], + "spans": [ + { + "bbox": [ + 53, + 487, + 287, + 532 + ], + "type": "text", + "content": "[9] Hui Han, Wen-Yuan Wang, and Bing-Huan Mao. Borderline-SMOTE: A new over-sampling method in imbalanced data sets learning. In Advances in Intelligent Computing, pages 878-887, August 2005. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 533, + 287, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 533, + 287, + 566 + ], + "spans": [ + { + "bbox": [ + 48, + 533, + 287, + 566 + ], + "type": "text", + "content": "[10] Haibo He and Edwardo A. Garcia. Learning from imbalanced data. IEEE Transactions on Knowledge and Data Engineering, 21(9):1263-1284, 2009. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 567, + 287, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 287, + 611 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 287, + 611 + ], + "type": "text", + "content": "[11] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 770-778, June 2016. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 613, + 287, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 656 + ], + "type": "text", + "content": "[12] Yin-Yin He, Jianxin Wu, and Xiu-Shen Wei. Distilling virtual examples for long-tailed recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 235-244, October 2021. 1, 2, 3, 5, 6, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "text", + "content": "[13] Youngkyu Hong, Seungju Han, Kwanghee Choi, Seokjun Seo, Beomsu Kim, and Buru Chang. Disentangling label distribution for long-tailed visual recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6626-6636, June 2021. 6" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "text", + "content": "[14] Bingyi Kang, Saining Xie, Marcus Rohrbach, Zhicheng Yan, Albert Gordo, Jiashi Feng, and Yannis Kalantidis. Decoupling representation and classifier for long-tailed recognition. In International Conference on Learning Representations, April 2020. 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 129, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 183 + ], + "type": "text", + "content": "[15] Jaehyung Kim, Jongheon Jeong, and Jinwoo Shin. M2m: Imbalanced classification via major-to-minor translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13896-13905, June 2020. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 185, + 545, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 185, + 545, + 229 + ], + "spans": [ + { + "bbox": [ + 307, + 185, + 545, + 229 + ], + "type": "text", + "content": "[16] Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, University of Toronto, 2009. https://www.cs.toronto.edu/~kriz/learning-features-2009-TR.pdf.1,5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 231, + 545, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 231, + 545, + 295 + ], + "spans": [ + { + "bbox": [ + 308, + 231, + 545, + 295 + ], + "type": "text", + "content": "[17] Tianhong Li, Peng Cao, Yuan Yuan, Lijie Fan, Yuzhe Yang, Rogerio S. Feris, Piotr Indyk, and Dina Katabi. Targeted supervised contrastive learning for long-tailed recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6918-6928, June 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 297, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 297, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 308, + 297, + 545, + 342 + ], + "type": "text", + "content": "[18] Tianhao Li, Limin Wang, and Gangshan Wu. Self supervision to distillation for long-tailed visual recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 630-639, October 2021. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 343, + 545, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 343, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 308, + 343, + 545, + 398 + ], + "type": "text", + "content": "[19] Ziwei Liu, Zhongqi Miao, Xiaohang Zhan, Jiayun Wang, Boqing Gong, and Stella X. Yu. Large-scale long-tailed recognition in an open world. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2537-2546, June 2019. 2, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 399, + 545, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 399, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 308, + 399, + 545, + 443 + ], + "type": "text", + "content": "[20] Aditya Krishna Menon, Sadeep Jayasumana, Ankit Singh Rawat, Himanshu Jain, Andreas Veit, and Sanjiv Kumar. Long-tail learning via logit adjustment. In International Conference on Learning Representations, May 2021. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 445, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 445, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 308, + 445, + 545, + 510 + ], + "type": "text", + "content": "[21] Seulki Park, Youngkyu Hong, Byeongho Heo, Sangdoo Yun, and Jin Young Choi. The majority can help the minority: Context-rich minority oversampling for long-tailed classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6887-6896, June 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 512, + 545, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 512, + 545, + 566 + ], + "spans": [ + { + "bbox": [ + 308, + 512, + 545, + 566 + ], + "type": "text", + "content": "[22] Jiawei Ren, Cunjun Yu, Shunan Sheng, Xiao Ma, Haiyu Zhao, Shuai Yi, and Hongsheng Li. Balanced meta-softmax for long-tailed visual recognition. In Advances in Neural Information Processing Systems 33, pages 4175–4186, December 2020. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 567, + 545, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 567, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 308, + 567, + 545, + 611 + ], + "type": "text", + "content": "[23] Dvir Samuel and Gal Chechik. Distributional robustness loss for long-tail learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9495-9504, October 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 613, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 613, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 308, + 613, + 545, + 657 + ], + "type": "text", + "content": "[24] Li Shen, Zhouchen Lin, and Qingming Huang. Relay backpropagation for effective learning of deep convolutional neural networks. In Proceedings of the European Conference on Computer Vision, pages 467-482, October 2016. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "type": "text", + "content": "[25] Peng Wang, Kai Han, Xiu-Shen Wei, Lei Zhang, and Lei Wang. Contrastive learning based hybrid networks for long-tailed image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 943-952, June 2021. 3" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15812" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 485 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[26] Xudong Wang, Long Lian, Zhongqi Miao, Ziwei Liu, and Stella Yu. Long-tailed recognition by routing diverse distribution-aware experts. In International Conference on Learning Representations, May 2021. 1, 2, 4, 5, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 288, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 288, + 172 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 288, + 172 + ], + "type": "text", + "content": "[27] Saining Xie, Ross Girshick, Piotr Dollar, Zhuowen Tu, and Kaiming He. Aggregated residual transformations for deep neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1492-1500, July 2017. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 174, + 287, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 174, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 49, + 174, + 287, + 217 + ], + "type": "text", + "content": "[28] Yuzhe Yang and Zhi Xu. Rethinking the value of labels for improving class-imbalanced learning. In Advances in Neural Information Processing Systems 33, pages 19290-19301, December 2020. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 219, + 287, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 219, + 287, + 273 + ], + "spans": [ + { + "bbox": [ + 49, + 219, + 287, + 273 + ], + "type": "text", + "content": "[29] Songyang Zhang, Zeming Li, Shipeng Yan, Xuming He, and Jian Sun. Distribution alignment: A unified framework for long-tail visual recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2361-2370, June 2021. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 274, + 287, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 274, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 49, + 274, + 287, + 327 + ], + "type": "text", + "content": "[30] Zhisheng Zhong, Jiequan Cui, Shu Liu, and Jiaya Jia. Improving calibration for long-tailed recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16489-16498, June 2021. 1, 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 331, + 287, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 331, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 48, + 331, + 287, + 384 + ], + "type": "text", + "content": "[31] Boyan Zhou, Quan Cui, Xiu-Shen Wei, and Zhao-Min Chen. BBN: Bilateral-branch network with cumulative learning for long-tailed visual recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9719–9728, June 2020. 2, 4, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 386, + 287, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 386, + 287, + 429 + ], + "spans": [ + { + "bbox": [ + 48, + 386, + 287, + 429 + ], + "type": "text", + "content": "[32] Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 40(6):1452-1464, 2018. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 430, + 287, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 430, + 287, + 485 + ], + "spans": [ + { + "bbox": [ + 48, + 430, + 287, + 485 + ], + "type": "text", + "content": "[33] Jianggang Zhu, Zheng Wang, Jingjing Chen, Yi-Ping Phoebe Chen, and Yu-Gang Jiang. Balanced contrastive learning for long-tailed visual recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6908-6917, June 2022. 3" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "15813" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/NoPe-NeRF_ Optimising Neural Radiance Field With No Pose Prior/2260561b-6c24-4454-9ee0-0e88425e374a_content_list.json b/2023/NoPe-NeRF_ Optimising Neural Radiance Field With No Pose Prior/2260561b-6c24-4454-9ee0-0e88425e374a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..3bbd70a9d474134ac401cd47c383c7337b1c90a5 --- /dev/null +++ b/2023/NoPe-NeRF_ Optimising Neural Radiance Field With No Pose Prior/2260561b-6c24-4454-9ee0-0e88425e374a_content_list.json @@ -0,0 +1,1808 @@ +[ + { + "type": "text", + "text": "NoPe-NeRF: Optimising Neural Radiance Field with No Pose Prior", + "text_level": 1, + "bbox": [ + 145, + 130, + 823, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Wenjing Bian", + "bbox": [ + 205, + 180, + 318, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zirui Wang", + "bbox": [ + 372, + 181, + 465, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kejie Li", + "bbox": [ + 521, + 181, + 589, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jia-Wang Bian", + "bbox": [ + 643, + 181, + 763, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Victor Adrian Prisacariu", + "bbox": [ + 385, + 199, + 581, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Active Vision Lab, University of Oxford", + "bbox": [ + 325, + 217, + 645, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{wenjing, ryan, kejie, jiawang, victor}@robots.ox.ac.uk", + "bbox": [ + 238, + 236, + 723, + 251 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6456383c90eff0c41d60483ff6ee3639f9c2a5a312a21425f09678ffb0312283.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 99, + 285, + 215, + 378 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cc1a091e00edbf3086064344697f8b788c846e7fa1b750e06194dde68cef4400.jpg", + "image_caption": [ + "Figure 1. Novel view synthesis comparison. We propose NoPe-NeRF for joint pose estimation and novel view synthesis. Our method enables more robust pose estimation and renders better novel view synthesis than previous state-of-the-art methods." + ], + "image_footnote": [], + "bbox": [ + 99, + 380, + 215, + 470 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7726bb48c80d2abc808b1dbcafa6d8a0a37b77fd79638ab3242ff39329c4c3a4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 218, + 286, + 483, + 378 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7ce87e153f7b0d9c78f8a1d851b3295017619fdf465ce0762f817e85be6a1128.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 218, + 378, + 483, + 470 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0947429dff822ae0042cec68959abcc2791f07b733fa4ff7c712da433d518cde.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 483, + 286, + 750, + 378 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6c2fe2396fdbe9b3d56187ccb624b9fdfad4f3be627a7be220df5c061ccf6c37.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 483, + 378, + 750, + 470 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/829f40ee63e1d95e7d7a6a16af554c09194ea234ce0181c9338caf0da21fac81.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 753, + 286, + 874, + 378 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/560597951ee9f257e9d09e683c2aa220680a9af6f1788756a94fc6199ae59ce9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 753, + 378, + 874, + 470 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 525, + 313, + 540 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Training a Neural Radiance Field (NeRF) without precomputed camera poses is challenging. Recent advances in this direction demonstrate the possibility of jointly optimising a NeRF and camera poses in forward-facing scenes. However, these methods still face difficulties during dramatic camera movement. We tackle this challenging problem by incorporating undistorted monocular depth priors. These priors are generated by correcting scale and shift parameters during training, with which we are then able to constrain the relative poses between consecutive frames. This constraint is achieved using our proposed novel loss functions. Experiments on real-world indoor and outdoor scenes show that our method can handle challenging camera trajectories and outperforms existing methods in terms of novel view rendering quality and pose estimation accuracy. Our project page is https://nope-nerf.active.vision.", + "bbox": [ + 73, + 556, + 472, + 815 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 844, + 207, + 859 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The photo-realistic reconstruction of a scene from a stream of RGB images requires both accurate 3D geometry", + "bbox": [ + 75, + 869, + 470, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "reconstruction and view-dependent appearance modelling. Recently, Neural Radiance Fields (NeRF) [24] have demonstrated the ability to build high-quality results for generating photo-realistic images from novel viewpoints given a sparse set of images.", + "bbox": [ + 496, + 526, + 890, + 602 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "An important preparation step for NeRF training is the estimation of camera parameters for the input images. A current go-to option is the popular Structure-from-Motion (SfM) library COLMAP [35]. Whilst easy to use, this preprocessing step could be an obstacle to NeRF research and real-world deployments in the long term due to its long processing time and its lack of differentiability. Recent works such as NeRFmm [46], BARF [18] and SC-NeRF [12] propose to simultaneously optimise camera poses and the neural implicit representation to address these issues. Nevertheless, these methods can only handle forward-facing scenes when no initial parameters are supplied, and fail in dramatic camera motions, e.g.a casual handheld captured video.", + "bbox": [ + 496, + 607, + 892, + 804 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This limitation has two key causes. First, all these methods estimate a camera pose for each input image individually without considering relative poses between images. Looking back to the literature of Simultaneous localisation and mapping (SLAM) and visual odometry, pose estimation can significantly benefit from estimating relative poses be", + "bbox": [ + 496, + 809, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "4160", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tween adjacent input frames. Second, the radiance field is known to suffer from shape-radiance ambiguity [55]. Estimating camera parameters jointly with NeRF adds another degree of ambiguity, resulting in slow convergence and unstable optimisation.", + "bbox": [ + 75, + 90, + 468, + 165 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To handle the limitation of large camera motion, we seek help from monocular depth estimation [22, 28, 29, 51]. Our motivation is threefold: First, monocular depth provides strong geometry cues that are beneficial to constraint shape-raliance ambiguity. Second, relative poses between adjacent depth maps can be easily injected into the training pipeline via Chamfer Distance. Third, monocular depth is lightweight to run and does not require camera parameters as input, in contrast to multi-view stereo depth estimation. For simplicity, we use the term mono-depth from now on.", + "bbox": [ + 75, + 167, + 467, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Utilising mono-depth effectively is not straightforward with the presence of scale and shift distortions. In other words, mono-depth maps are not multi-view consistent. Previous works [9, 17, 47] simply take mono-depth into a depth-wise loss along with NeRF training. Instead, we propose a novel and effective way to thoroughly integrate mono-depth into our system. First, we explicitly optimise scale and shift parameters for each mono-depth map during NeRF training by penalising the difference between rendered depth and mono-depth. Since NeRF by itself is trained based on multiview consistency, this step transforms mono-depth maps to undistorted multiview consistent depth maps. We further leverage these multiview consistent depth maps in two loss terms: a) a Chamfer Distance loss between two depth maps of adjacent images, which injects relative pose to our system; and b) a depth-based surface rendering loss, which further improves relative pose estimation.", + "bbox": [ + 75, + 316, + 467, + 571 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, we propose a method to jointly optimise camera poses and a NeRF from a sequence of images with large camera motion. Our system is enabled by three contributions. First, we propose a novel way to integrate monodepth into unposed-NeRF training by explicitly modelling scale and shift distortions. Second, we supply relative poses to the camera-NeRF joint optimisation via an inter-frame loss using undistorted mono-depth maps. Third, we further regularise our relative pose estimation with a depth-based surface rendering loss.", + "bbox": [ + 75, + 574, + 467, + 724 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As a result, our method is able to handle large camera motion, and outperforms state-of-the-art methods by a significant margin in terms of novel view synthesis quality and camera trajectory accuracy.", + "bbox": [ + 75, + 724, + 467, + 785 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 799, + 217, + 814 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Novel View Synthesis. While early Novel View Synthesis (NVS) approaches applied interpolations between pixels [3], later works often rendered images from 3D reconstruction [1, 6]. In recent years, different representations of the 3D scene are used, e.g.meshes [30, 31], Multi-Plane", + "bbox": [ + 75, + 824, + 467, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Images [41, 59], layered depth [42] etc. Among them, NeRF [24] has become a popular scene representation for its photorealistic rendering.", + "bbox": [ + 498, + 90, + 890, + 135 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A number of techniques are proposed to improve NeRF's performance with additional regularisation [13, 26, 54], depth priors [7, 32, 47, 53], surface enhancements [27, 43, 49] or latent codes [40, 44, 52]. Other works [2, 10, 25, 34] have also accelerated NeRF training and rendering. However, most of these approaches require pre-computed camera parameters obtained from SfM algorithms [11, 35].", + "bbox": [ + 496, + 136, + 890, + 242 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "NeRF With Pose Optimisation. Removing camera parameter preprocessing is an active line of research recently. One category of the methods [33, 38, 60] use a SLAM-style pipeline, that either requires RGB-D inputs or relies on accurate camera poses generated from the SLAM tracking system. Another category of works optimises camera poses with the NeRF model directly. We term this type of method as unposed-NeRF in this paper. iNeRF [50] shows that poses for novel view images can be estimated using a reconstructed NeRF model. GNeRF [21] combines Generative Adversarial Networks with NeRF to estimate camera poses but requires a known sampling distribution for poses. More relevant to our work, NeRFmm [46] jointly optimises both camera intrinsic and extrinsic alongside NeRF training. BARF [18] proposes a coarse-to-fine positional encoding strategy for camera poses and NeRF joint optimisation. SCNeRF [12] further parameterises camera distortion and employs a geometric loss to regularise rays. GARF [4] shows that using Gaussian-MLPs makes joint pose and scene optimisation easier and more accurate. Recently, SiNeRF [48] uses SIREN [36] layers and a novel sampling strategy to alleviate the sub-optimality of joint optimisation in NeRFmm. Although showing promising results on the forward-facing dataset like LLFF [23], these approaches face difficulties when handling challenging camera trajectories with large camera motion. We address this issue by closely integrating mono-depth maps with the joint optimisation of camera parameters and NeRF.", + "bbox": [ + 496, + 258, + 890, + 681 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 694, + 589, + 709 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We tackle the challenge of handling large camera motions in unposed-NeRF training. Given a sequence of images, camera intrinsics, and their mono-depth estimations, our method recovers camera poses and optimises a NeRF simultaneously. We assume camera intrinsics are available in the image meta block, and we run an off-the-shelf monodepth network DPT [7] to acquire mono-depth estimations. Without repeating the benefit of mono-depth, we unroll this section around the effective integration of monocular depth into unposed-NeRF training.", + "bbox": [ + 496, + 719, + 890, + 869 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The training is a joint optimisation of the NeRF, camera poses, and distortion parameters of each mono-depth", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "4161", + "bbox": [ + 480, + 944, + 513, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/1d697123e9c97e03383d67eb3111e6dc00d5feeec2e1903c66e745ca1242a200.jpg", + "image_caption": [ + "Figure 2. Method Overview. Our method takes a sequence of images as input to reconstruct NeRF and jointly estimates the camera poses of the frames. We first generate monocular depth maps from a mono-depth estimation network and reconstruct the point clouds. We then optimise NeRF, camera poses, and depth distortion parameters jointly with inter-frame and NeRF losses." + ], + "image_footnote": [], + "bbox": [ + 76, + 87, + 519, + 369 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/429c0ded872daf15aad5f8b8de9aab2f06cbcc399ff69748ba8379356f538e90.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 531, + 87, + 736, + 349 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b907a9e723a3afab4a8fd3a5ab307f4a6530b07b83d50befff349e18407a1dbf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 753, + 88, + 890, + 367 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "map. The distortion parameters are supervised by minimising the discrepancies between the mono-depth maps and depth maps rendered from the NeRF, which are multiview consistent. The undistorted depth maps in return effectively mediate the shape-radiance ambiguity, which eases the training of NeRF and camera poses.", + "bbox": [ + 75, + 449, + 468, + 541 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Specifically, the undistorted depth maps enable two constraints. We constrain global pose estimation by supplying relative pose between adjacent images. This is achieved via a Chamfer-Distance-based correspondence between two point clouds, back-projected from undistorted depth maps. Further, we regularise relative pose estimation with a surface-based photometric consistency where we treat undistorted depth as surface.", + "bbox": [ + 75, + 542, + 468, + 662 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We detail our method in the following sections, starting from NeRF in Sec. 3.1 and unposed-NeRF training in Sec. 3.2, looking into mono-depth distortions in Sec. 3.3, followed by our mono-depth enabled loss terms in Sec. 3.4, and finishing with an overall training pipeline Sec. 3.5.", + "bbox": [ + 75, + 664, + 468, + 742 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. NeRF", + "text_level": 1, + "bbox": [ + 76, + 753, + 158, + 768 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Neural Radiance Field (NeRF) [24] represents a scene as a mapping function $F_{\\Theta}:(\\mathbf{x},\\mathbf{d})\\to (\\mathbf{c},\\sigma)$ that maps a 3D location $\\mathbf{x}\\in \\mathbb{R}^3$ and a viewing direction $\\mathbf{d}\\in \\mathbb{R}^3$ to a radiance colour $\\mathbf{c}\\in \\mathbb{R}^3$ and a volume density value $\\sigma$ . This mapping is usually implemented with a neural network parameterised by $F_{\\Theta}$ . Given $N$ images $\\mathcal{I} = \\{I_i\\mid i = 0\\dots N - 1\\}$ with their camera poses $\\Pi = \\{\\pi_i\\mid i = 0\\dots N - 1\\}$ , NeRF can be optimised by minimising photometric error", + "bbox": [ + 75, + 779, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$\\mathcal{L}_{rgb} = \\sum_{i}^{N}\\| I_{i} - \\hat{I}_{i}\\|_{2}^{2}$ between synthesised images $\\hat{\\mathcal{I}}$ and captured images $\\mathcal{I}$ :", + "bbox": [ + 498, + 446, + 890, + 479 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\Theta^ {*} = \\arg \\min _ {\\Theta} \\mathcal {L} _ {r g b} (\\hat {\\mathcal {I}} \\mid \\mathcal {I}, \\Pi), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 593, + 486, + 890, + 508 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\hat{I}_i$ is rendered by aggregating radiance colour on camera rays $\\mathbf{r}(h) = \\mathbf{o} + h\\mathbf{d}$ between near and far bound $h_n$ and $h_f$ . More concretely, we synthesise $\\hat{I}_i$ with a volumetric rendering function", + "bbox": [ + 498, + 518, + 890, + 580 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {I} _ {i} (\\mathbf {r}) = \\int_ {h _ {n}} ^ {h _ {f}} T (h) \\sigma (\\mathbf {r} (h)) \\mathbf {c} (\\mathbf {r} (h), \\mathbf {d}) d h, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 558, + 585, + 890, + 622 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $T(h) = \\exp(-\\int_{h_n}^h \\sigma(\\mathbf{r}(s)) ds)$ is the accumulated transmittance along a ray. We refer to [24] for further details.", + "bbox": [ + 498, + 630, + 890, + 676 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Joint Optimisation of Poses and NeRF", + "text_level": 1, + "bbox": [ + 500, + 685, + 828, + 700 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Prior works [12, 18, 46] show that it is possible to estimate both camera parameters and a NeRF at the same time by minimising the above photometric error $\\mathcal{L}_{rgb}$ under the same volumetric rendering process in Eq. (2).", + "bbox": [ + 498, + 708, + 890, + 768 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The key lies in conditioning camera ray casting on variable camera parameters $\\Pi$ , as the camera ray $\\mathbf{r}$ is a function of camera pose. Mathematically, this joint optimisation can be formulated as:", + "bbox": [ + 496, + 768, + 890, + 828 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\Theta^ {*}, \\Pi^ {*} = \\arg \\min _ {\\Theta , \\Pi} \\mathcal {L} _ {r g b} (\\hat {\\mathcal {I}}, \\hat {\\Pi} \\mid \\mathcal {I}), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 835, + 890, + 859 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\hat{\\Pi}$ denotes camera parameters that are updated during optimising. Note that the only difference between Eq. (1)", + "bbox": [ + 496, + 869, + 890, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "4162", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and Eq. (3) is that Eq. (3) considers camera parameters as variables.", + "bbox": [ + 75, + 90, + 468, + 119 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In general, the camera parameters $\\Pi$ include camera intrinsics, poses, and lens distortions. We only consider estimating camera poses in this work, e.g., camera pose for frame $I_{i}$ is a transformation $\\mathbf{T}_i = [\\mathbf{R}_i\\mid \\mathbf{t}_i]$ with a rotation $\\mathbf{R}_i\\in \\mathrm{SO}(3)$ and a translation $\\mathbf{t}_i\\in \\mathbb{R}^3$", + "bbox": [ + 76, + 121, + 467, + 196 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Undistortion of Monocular Depth", + "text_level": 1, + "bbox": [ + 76, + 205, + 372, + 222 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With an off-the-shelf monocular depth network, e.g., DPT [28], we generate mono-depth sequence $\\mathcal{D} = \\{D_i\\mid i = 0\\dots N - 1\\}$ from input images. Without surprise, mono-depth maps are not multi-view consistent so we aim to recover a sequence of multi-view consistent depth maps, which are further leveraged in our relative pose loss terms.", + "bbox": [ + 76, + 229, + 467, + 332 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, we consider two linear transformation parameters for each mono-depth map, resulting in a sequence of transformation parameters for all frames $\\Psi = \\{(\\alpha_i,\\beta_i)\\mid i = 0\\ldots N - 1\\}$ , where $\\alpha_{i}$ and $\\beta_{i}$ denote a scale and a shift factor. With multi-view consistent constraint from NeRF, we aim to recover a multi-view consistent depth map $D_{i}^{*}$ for $D_{i}$ :", + "bbox": [ + 76, + 335, + 467, + 441 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nD _ {i} ^ {*} = \\alpha_ {i} D _ {i} + \\beta_ {i}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 452, + 468, + 469 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "by joint optimising $\\alpha_{i}$ and $\\beta_{i}$ along with a NeRF. This joint optimisation is mostly achieved by enforcing the consistency between an undistorted depth map $D_{i}^{*}$ and a NeRF rendered depth map $\\hat{D}_i$ via a depth loss:", + "bbox": [ + 75, + 479, + 468, + 542 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {d e p t h}} = \\sum_ {i} ^ {N} \\left\\| D _ {i} ^ {*} - \\hat {D} _ {i} \\right\\|, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 553, + 468, + 594 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where", + "bbox": [ + 76, + 604, + 122, + 617 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {D} _ {i} (\\mathbf {r}) = \\int_ {h _ {n}} ^ {h _ {f}} T (h) \\sigma (\\mathbf {r} (h)) d h \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 616, + 468, + 651 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "denotes a volumetric rendered depth map from NeRF.", + "bbox": [ + 76, + 657, + 431, + 671 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "It is important to note that both NeRF and mono-depth benefit from Eq. (5). On the one hand, mono-depth provides strong geometry prior for NeRF training, reducing shape-radiance ambiguity. On the other hand, NeRF provides multi-view consistency so we can recover a set of multiview consistent depth maps for relative pose estimations.", + "bbox": [ + 75, + 671, + 468, + 762 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Relative Pose Constraint", + "text_level": 1, + "bbox": [ + 76, + 771, + 303, + 786 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Aforementioned unposed-NeRF methods [12, 18, 46] optimise each camera pose independently, resulting in an overfit to target images with incorrect poses. Penalising incorrect relative poses between frames can help to regularise the joint optimisation towards smooth convergence, especially in a complex camera trajectory. Therefore, we propose two losses that constrain relative poses.", + "bbox": [ + 75, + 794, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Point Cloud Loss. We back-project the undistorted depth maps $\\mathcal{D}^*$ using the known camera intrinsics, to point clouds $\\mathcal{P}^* = \\{P_i^* \\mid i = 0 \\dots N - 1\\}$ and optimise the relative pose between consecutive point clouds by minimising a point cloud loss $\\mathcal{L}_{pc}$ :", + "bbox": [ + 498, + 90, + 890, + 167 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {p c} = \\sum_ {(i, j)} l _ {c d} \\left(P _ {j} ^ {*}, \\mathbf {T} _ {j i} P _ {i} ^ {*}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 601, + 176, + 890, + 210 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{T}_{ji} = \\mathbf{T}_j\\mathbf{T}_i^{-1}$ represents the related pose that transforms point cloud $P_{i}^{*}$ to $P_{j}^{*}$ , tuple $(i,j)$ denotes indices of a consecutive pair of instances, and $l_{cd}$ denotes Chamfer Distance:", + "bbox": [ + 498, + 220, + 890, + 280 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nl _ {c d} \\left(P _ {i}, P _ {j}\\right) = \\sum_ {p _ {i} \\in P _ {i}} \\min _ {p _ {j} \\in P _ {j}} \\| p _ {i} - p _ {j} \\| _ {2} + \\sum_ {p _ {j} \\in P _ {j}} \\min _ {p _ {i} \\in P _ {i}} \\| p _ {i} - p _ {j} \\| _ {2}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 498, + 289, + 893, + 335 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Surface-based Photometric Loss. While the point cloud loss $\\mathcal{L}_{pc}$ offers supervision in terms of 3D-3D matching, we observe that a surface-based photometric error can alleviate incorrect matching. With the photometric consistency assumption, this photometric error penalises the differences in appearance between associated pixels. The association is established by projecting the point cloud $P_{i}^{*}$ onto images $I_{i}$ and $I_{j}$ .", + "bbox": [ + 498, + 354, + 890, + 474 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The surface-based photometric loss can then be defined as:", + "bbox": [ + 500, + 476, + 890, + 503 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {r g b - s} = \\sum_ {(i, j)} \\| I _ {i} \\langle \\mathbf {K} _ {i} P _ {i} ^ {*} \\rangle - I _ {j} \\langle \\mathbf {K} _ {j} \\mathbf {T} _ {j} \\mathbf {T} _ {i} ^ {- 1} P _ {i} ^ {*} \\rangle \\|, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 512, + 890, + 546 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\langle \\cdot \\rangle$ represents the sampling operation on the image and $\\mathbf{K}_i$ denotes a projection matrix for $i_{th}$ camera.", + "bbox": [ + 498, + 556, + 890, + 585 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.5. Overall Training Pipeline", + "text_level": 1, + "bbox": [ + 500, + 595, + 732, + 611 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Assembling all loss terms, we get the overall loss function:", + "bbox": [ + 500, + 618, + 888, + 646 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {r g b} + \\lambda_ {1} \\mathcal {L} _ {\\text {d e p t h}} + \\lambda_ {2} \\mathcal {L} _ {p c} + \\lambda_ {3} \\mathcal {L} _ {r g b - s}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 531, + 657, + 890, + 675 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\lambda_1, \\lambda_2, \\lambda_3$ are the weighting factors for respective loss terms. By minimising the combined of loss $\\mathcal{L}$ :", + "bbox": [ + 498, + 684, + 890, + 714 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Theta^ {*}, \\Pi^ {*}, \\Psi^ {*} = \\arg \\min _ {\\Theta , \\Pi , \\Psi} \\mathcal {L} (\\hat {\\mathcal {I}}, \\hat {\\mathcal {D}}, \\hat {\\Pi}, \\hat {\\Psi} \\mid \\mathcal {I}, \\mathcal {D}), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 722, + 890, + 747 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "our method returns the optimised NeRF parameters $\\Theta$ , camera poses $\\Pi$ , and distortion parameters $\\Psi$ .", + "bbox": [ + 498, + 756, + 890, + 787 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 800, + 632, + 816 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We begin with a description of our experimental setup in Sec. 4.1. In Sec. 4.2, we compare our method with pose-unknown methods. Next, we compare our method with the COLMAP-assisted NeRF baseline in Sec. 4.3. Lastly, we conduct ablation studies in Sec. 4.4.", + "bbox": [ + 498, + 825, + 890, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4163", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/bf97d274716646e4a0f70d55a3c67c74a0405cd3380efc61fc648966d5f7f921.jpg", + "image_caption": [ + "(a) BARF", + "(b) NeRFmm", + "(c) SC-NeRF", + "(d) Ours", + "(e) Ground Truth", + "Figure 3. Qualitative results of novel view synthesis and depth prediction on Tanks and Temples. We visualise the synthesised images and the rendered depth maps (top left of each image) for all methods. NoPe-NeRF is able to recover details for both colour and geometry." + ], + "image_footnote": [], + "bbox": [ + 80, + 90, + 890, + 301 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 76, + 386, + 266, + 402 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We conduct experiments on two datasets Tanks and Temples [15] and ScanNet [5]. Tanks and Temples: we use 8 scenes to evaluate pose accuracy and novel view synthesis quality. We chose scenes captured at both indoor and outdoor locations, with different frame sampling rates and lengths. All images are down-sampled to a resolution of $960 \\times 540$ . For the family scene, we sample 200 images and take 100 frames with odd frame ids as training images and the remaining 100 frames for novel view synthesis, in order to analyse the performance under smooth motion. For the remaining scenes, following NeRF [24], 1/8 of the images in each sequence are held out for novel view synthesis, unless otherwise specified. ScanNet: we select 4 scenes for evaluating pose accuracy, depth accuracy, and novel view synthesis quality. For each scene, we take 80-100 consecutive images and use 1/8 of these images for novel view synthesis. For evaluation, we employ depth maps and poses provided by ScanNet as ground truth. ScanNet images are down-sampled to $648 \\times 484$ . We crop images with dark orders during preprocessing.", + "bbox": [ + 75, + 410, + 473, + 712 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Metrics. We evaluate our proposed method in three aspects. For novel view synthesis, we follow previous methods [12,18,46], and use standard evaluation metrics, including Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM) [45] and Learned Perceptual Image Patch Similarity (LPIPS) [56]. For pose evaluation, We use standard visual odometry metrics [16, 37, 57], including the Absolute Trajectory Error (ATE) and Relative Pose Error (RPE). ATE measures the difference between the estimated camera positions and the ground truth positions. RPE measures the relative pose errors between pairs of images,", + "bbox": [ + 75, + 734, + 468, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "which consists of relative rotation error $(\\mathrm{RPE}_r)$ and relative translation error $(\\mathrm{RPE}_t)$ . The estimated trajectory is aligned with the ground truth using Sim(3) with 7 degrees of freedom. We use standard depth metrics [8,19,20,39] (Abs Rel, Sq Rel, RMSE, RMSE log, $\\delta_1$ , $\\delta_2$ and $\\delta_3$ ) for depth evaluation. For further detail, please refer to the supplementary material. To recover the metric scale, we follow Zhou et al. [58] and match the median value between rendered and ground truth depth maps.", + "bbox": [ + 496, + 386, + 893, + 523 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. Our model architecture is based on NeRF [24] with a few modifications: a) replacing ReLU activation function with Softplus and b) sampling 128 points along each ray uniformly with noise, between a predefined range (0.1, 10). We use 2 separate Adam optimisers [14] for NeRF and other parameters. The initial learning rate for NeRF is 0.001 and for the pose and distortion is 0.0005. Camera rotations are optimised in axis-angle representation $\\phi_i \\in \\mathfrak{so}(3)$ . We first train the model with all losses with constant learning rates until the inter-frame losses converge. Then, we remove the inter-frame losses and depth loss to refine the model with the RGB loss only. We decay the learning rates with different schedulers to refine for 10,000 epochs. We balance the loss terms with $\\lambda_1 = 0.04$ , $\\lambda_2 = 1.0$ and $\\lambda_3 = 1.0$ . For each training step, we randomly sample 1024 pixels (rays) from each input image and 128 samples per ray. More details are provided in the supplementary material.", + "bbox": [ + 496, + 546, + 893, + 819 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Comparing With Pose-Unknown Methods", + "text_level": 1, + "bbox": [ + 498, + 830, + 857, + 847 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We compare our method with pose-unknown baselines, including BARF [18], NeRFmm [46] and SC-NeRF [12].", + "bbox": [ + 498, + 854, + 890, + 883 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "View Synthesis Quality. To obtain the camera poses of", + "bbox": [ + 519, + 885, + 892, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "4164", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8a961b15ff878b2b3cffb7e326c300c3b156cd89ca5cdda08b5d72724dfc57d8.jpg", + "image_caption": [ + "Figure 4. Pose Estimation Comparison. We visualise the trajectory (3D plot) and relative rotation errors $\\mathrm{RPE}_r$ (bottom colour bar) of each method on Ballroom and Museum. The colour bar on the right shows the relative scaling of colour. More results are in the supplementary." + ], + "image_footnote": [], + "bbox": [ + 125, + 94, + 851, + 369 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/f6ae0dbcd5d51550dca1fd54288d435fc59b3c5d583838e580665d4fcdab18f3.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
scenesOursBARFNeRFmmSC-NeRF
PSNR ↑SSIM ↑LPIPS ↓PSNRSSIMLPIPSPSNRSSIMLPIPSPSNRSSIMLPIPS
ScanNet0079_0032.470.840.4132.310.830.4330.590.810.4931.330.820.46
0418_0031.330.790.3431.240.790.3530.000.770.4029.050.750.43
0301_0029.830.770.3629.310.760.3827.840.720.4529.450.770.35
0431_0033.830.910.3932.770.900.4131.440.880.4532.570.900.40
mean31.860.830.3831.410.820.3929.970.800.4530.600.810.41
Tanks and TemplesChurch25.170.730.3923.170.620.5221.640.580.5421.960.600.53
Barn26.350.690.4425.280.640.4823.210.610.5323.260.620.51
Museum26.770.760.3523.580.610.5522.370.610.5324.940.690.45
Family26.010.740.4123.040.610.5623.040.580.5622.600.630.51
Horse27.640.840.2624.090.720.4123.120.700.4325.230.760.37
Ballroom25.330.720.3820.660.500.6020.030.480.5722.640.610.48
Francis29.480.800.3825.850.690.5725.4000.690.5226.460.730.49
Ignatius23.960.610.4721.780.470.6021.160.450.6023.000.550.53
mean26.340.740.3923.420.610.5422.500.590.5423.760.650.48
", + "bbox": [ + 78, + 425, + 915, + 623 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Novel view synthesis results on ScanNet and Tanks and Temples. Each baseline method is trained with its public code under the original settings and evaluated with the same evaluation protocol.", + "bbox": [ + 75, + 633, + 892, + 662 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "test views for rendering, we minimise the photometric error of the synthesised images while keeping the NeRF model fixed, as in NeRFmm [46]. Each test pose is initialised with the learned pose of the training frame that is closest to it. We use the same pre-processing for all baseline approaches, which results in higher accuracy than their original implementations. More details are provided in the supplementary material. Our method outperforms all the baselines by a large margin. The quantitative results are summarised in Tab. 1, and qualitative results are shown in Fig. 3.", + "bbox": [ + 75, + 688, + 468, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We recognised that because the test views, which are sampled from videos, are close to the training views, good results may be obtained due to overfitting to the training images. Therefore, we conduct an additional qualitative", + "bbox": [ + 75, + 839, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "evaluation on more novel views. Specifically, we fit aBezier curve from the estimated training poses and sample interpolated poses for each method to render novel view videos. Sampled results are shown in Fig. 5, and the rendered videos are in the supplementary material. These results show that our method renders photo-realistic images consistently, while other methods generate visible artifacts.", + "bbox": [ + 496, + 688, + 890, + 794 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Camera Pose. Our method significantly outperforms other baselines in all metrics. The quantitative pose evaluation results are shown in Tab. 2. For ScanNet, we use the camera poses provided by the dataset as ground truth. For Tanks and Temples, not every video comes with ground truth poses, so we use COLMAP estimations for reference. Our estimated trajectory is better aligned with the ground", + "bbox": [ + 496, + 795, + 892, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "4165", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/c48842464f4f2980bb416893656b67e390a10b13e5418b50cf66c598c2bf76f9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
scenesOursBARFNeRFmmSC-NeRF
RPEt ↓RPEr ↓ATE↓RPEtRPErATERPEtRPErATERPEtRPErATE
ScanNet0079_000.7520.2040.0231.1100.4800.0621.7060.6360.1002.0640.6640.115
0418_000.4550.1190.0151.3980.5380.0201.4020.4600.0131.5280.5020.016
0301_000.3990.1230.0131.3160.7770.2193.0970.8940.2881.1330.4220.056
0431_001.6250.2740.0696.0240.7540.1686.7990.6240.4964.1100.4990.205
mean0.8080.1800.0302.4620.6370.1173.2510.6540.2242.2090.5220.098
Tanks and TemplesChurch0.0340.0080.0080.1140.0380.0520.6260.1270.0650.8360.1870.108
Barn0.0460.0320.0040.3140.2650.0501.6290.4940.1591.3170.4290.157
Museum0.2070.2020.0203.4421.1280.2634.1341.0510.3468.3391.4910.316
Family0.0470.0150.0011.3710.5910.1152.7430.5370.1201.1710.4990.142
Horse0.1790.0170.0031.3330.3940.0141.3490.4340.0181.3660.4380.019
Ballroom0.0410.0180.0020.5310.2280.0180.4490.1770.0310.3280.1460.012
Francis0.0570.0090.0051.3210.5580.0821.6470.6180.2071.2330.4830.192
Ignatius0.0260.0050.0020.7360.3240.0291.3020.3790.0410.5330.2400.085
mean0.0800.0380.0061.0460.4410.0781.7350.4770.1231.8900.4890.129
", + "bbox": [ + 109, + 88, + 859, + 253 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ef935c6effe7ceb41bb555f6a0a57c110a1fe32d45cb395b995c64f995c42920.jpg", + "table_caption": [ + "Table 2. Pose accuracy on ScanNet and Tanks and Temples. Note that we use COLMAP poses in Tanks and Temples as the \"ground truth\". The unit of $\\mathrm{RPE}_r$ is in degrees, ATE is in the ground truth scale and $\\mathrm{RPE}_t$ is scaled by 100." + ], + "table_footnote": [], + "table_body": "
Abs Rel ↓Sq Rel ↓RMSE ↓RMSE log ↓δ1 ↑δ2 ↑δ3 ↑
Ours0.1410.1370.5680.1760.8280.9700.987
BARF0.3760.6840.9900.4010.4900.7510.884
NeRFmm0.5901.7211.6720.5870.3160.5600.743
SC-NeRF0.4170.6421.0790.4760.3620.6580.832
DPT0.1970.2460.7510.2260.7470.9340.975
", + "bbox": [ + 80, + 316, + 465, + 391 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Depth map evaluation on ScanNet. Our depth estimation is more accurate than baseline models BARF [18], NeRFmm [46] and SC-NeRF [12]. Compared with DPT [58], we show our depth is more accurate after undistortion.", + "bbox": [ + 75, + 402, + 468, + 458 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "truth than other methods, and our estimated rotation is two orders of magnitudes more accurate than others. We visualise the camera trajectories and rotations in Fig. 4.", + "bbox": [ + 75, + 486, + 468, + 531 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Depth. We evaluate the accuracy of the rendered depth maps on ScanNet, which provides the ground-truth depths for evaluation. Our rendered depth maps achieve superior accuracy over the previous alternatives. We also compare with the mono-depth maps estimated by DPT. Our rendered depth maps, after undistortion using multiview consistency in the NeRF optimisation, outperform DPT by a large margin. The results are summarised in Tab. 3, and sampled qualitative results are illustrated in Fig. 3.", + "bbox": [ + 75, + 532, + 468, + 667 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Comparing With COLMAP Assisted NeRF", + "text_level": 1, + "bbox": [ + 76, + 679, + 444, + 696 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We make a comparison of pose estimation accuracy between our method and COLMAP against ground truth poses in ScanNet. We achieve on-par accuracy with COLMAP, as shown in Tab. 4. We further analyse the novel view synthesis quality of the NeRF model trained with our learned poses to COLMAP poses on ScanNet and Tanks and Temples. The original NeRF training contains two stages, finding poses using COLMAP and optimising the scene representation. In order to make our comparison fairer, in this section only, we mimic a similar two-stage training as the original NeRF [24]. In the first stage, we train our method with all losses for camera pose estimation, i.e., mimicking the COLMAP processing. Then, we fix the optimised poses", + "bbox": [ + 75, + 704, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3cbccf2b854b65f90026402f87da8bcfadee477424ca19c84ddcaefc04ba8a80.jpg", + "image_caption": [ + "Figure 5. Sampled frames from rendered novel view videos. For each method, we fit the learned trajectory with aBezier curve and uniformly sample new viewpoints for rendering. Our method generates significantly better results than previous methods, which show visible artifacts. The full rendered videos and details about generating novel views are provided in the supplementary." + ], + "image_footnote": [], + "bbox": [ + 506, + 320, + 890, + 545 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "and train a NeRF model from scratch, using the same settings and loss as the original NeRF. This evaluation enables us to compare our estimated poses to the COLMAP poses indirectly, i.e., in terms of contribution to view synthesis.", + "bbox": [ + 496, + 672, + 890, + 734 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our two-stage method outperforms the COLMAP-assisted NeRF baseline, which indicates a better pose estimation for novel view synthesis. The results are summarised in Tab. 5.", + "bbox": [ + 496, + 734, + 890, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As is commonly known, COLMAP performs poorly in low-texture scenes and sometimes fails to find accurate camera poses. Fig. 6 shows an example of a low-texture scene where COLMAP provides inaccurate pose estimation that causes NeRF to render images with visible artifacts. In contrast, our method renders high-quality images, thanks to robust optimisation of camera pose.", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "4166", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/397900dd0cf7a114fc0a3c7befe428acbef60889cde94ee3a1e5a4279ed4cb8a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
scenesOursCOLMAP
RPEt↓RPEr↓ATE↓RPEtRPErATE
0079_000.7520.2040.0230.6550.2210.012
0418_000.4550.1190.0150.4910.1240.016
0301_000.3990.1230.0130.4140.1360.009
0431_001.6250.2740.0691.2920.2490.051
mean0.8080.1800.0300.7130.1820.022
", + "bbox": [ + 78, + 88, + 465, + 176 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ab5ff0ea66cc70c7dccbe02caef55a0d10dcf96e0f561a4273fb78d68fbec191.jpg", + "image_caption": [ + "(a) COLMAP+NeRF" + ], + "image_footnote": [], + "bbox": [ + 80, + 231, + 205, + 306 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3351115fc0353aa326c5261a0a09a87858d4cf14f3c8461db588f41d0fb2a4da.jpg", + "image_caption": [ + "(b) Ours" + ], + "image_footnote": [], + "bbox": [ + 207, + 231, + 336, + 306 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c9121ae35a3206e6ce21c78be3065f214a6c836303976974215795cbce33c325.jpg", + "image_caption": [ + "(c) Ground Truth", + "Figure 6. COLMAP failure case. On a rotation-dominant sequence with low-texture areas, COLMAP fails to estimate correct poses, which results in artifacts in synthesised images." + ], + "image_footnote": [], + "bbox": [ + 336, + 231, + 465, + 306 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/4ad5d338d02e42e52679909d810d7212f450ca61d5a57066ea1e0e49dfa60b47.jpg", + "table_caption": [ + "Table 4. Comparison of pose accuracy with COLMAP on ScanNet." + ], + "table_footnote": [], + "table_body": "
scenesOursOurs-rCOLMAP+NeRF
PSNR ↑SSIM ↑LPIPS ↓PSNRSSIMLPIPSPSNRSSIMLPIPS
ScanNet0079_0032.470.840.4133.120.850.4031.980.830.43
0418_0031.330.790.3430.490.770.4030.600.780.40
0301_0029.830.770.3630.050.780.3430.010.780.36
0431_0033.830.910.3933.860.910.3933.540.910.39
mean31.860.830.3831.880.830.3831.530.820.40
Tanks and TemplesChurch25.170.730.3926.740.780.3225.720.750.37
Barn26.350.690.4426.580.710.4226.720.710.42
Museum26.770.760.3526.980.770.3627.210.780.34
Family26.010.740.4126.210.750.4026.610.770.39
Horse27.640.840.2628.060.840.2627.020.820.29
Ballroom25.330.720.3825.530.730.3825.470.730.38
Francis29.480.800.3829.730.810.3830.050.810.38
Ignatius23.960.610.4723.980.620.4624.080.610.47
mean26.340.740.3926.730.750.3726.610.750.38
", + "bbox": [ + 78, + 393, + 465, + 561 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Interestingly, this experiment also reveals that the two-stage method shows higher accuracy than the one-stage method. We hypothesise that the joint optimisation (from randomly initialised poses) in the one-stage approach causes the NeRF optimisation to be trapped in a local minimum, potentially due to the bad pose initialisation. The two-stage approach circumvents this issue by re-initialising the NeRF and re-training with well-optimised poses, resulting in higher performance.", + "bbox": [ + 75, + 641, + 467, + 777 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 786, + 228, + 801 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section, we analyse the effectiveness of the parameters and components that have been added to our model. The results of ablation studies are shown in Tab. 6.", + "bbox": [ + 75, + 810, + 467, + 854 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of Distortion Parameters. We find that ignoring depth distortions (i.e., setting scales to 1 and shifts to 0 as constants) leads to a degradation in pose accuracy, as in", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/6ff287b7cffc112f520a5df4b3313677c5479c77e92c52122dbd44fc8a442f28.jpg", + "table_caption": [ + "Table 5. Comparison to NeRF with COLMAP poses. Our two-stage method (Ours-r) outperforms both COLMAP+NeRF and our one-stage method (Ours)." + ], + "table_footnote": [], + "table_body": "
NVSPose
PSNR↑SSIM↑LPIPS↓RPEt↓RPEr↓ATE↓
Ours31.860.830.380.8010.1810.031
Ours w/o α,β31.460.820.391.9290.3210.066
Ours w/o Lpc31.730.820.382.2270.4530.101
Ours w/o Lrgb-s31.050.810.411.8140.4010.156
Ours w/o Ldepth31.200.810.401.4980.3830.089
", + "bbox": [ + 500, + 88, + 888, + 180 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Ablation study results on ScanNet.", + "bbox": [ + 558, + 191, + 831, + 205 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "consistent distortions of depth maps introduce errors to the estimation of relative poses and confuse NeRF for geometry reconstruction.", + "bbox": [ + 496, + 233, + 890, + 276 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of Inter-frame Losses. We observe that the interframe losses are the major contributor to improving relative poses. When removing the pairwise point cloud loss $L_{pc}$ or the surface-based photometric loss $L_{rgb-s}$ , there is less constraint between frames, and thus the pose accuracy becomes lower.", + "bbox": [ + 496, + 277, + 890, + 367 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of NeRF Losses. When the depth loss $L_{depth}$ is removed, the distortions of input depth maps are only optimised locally through the inter-frame losses. We find that this can lead to drift and degradation in pose accuracy.", + "bbox": [ + 496, + 369, + 890, + 430 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Limitations", + "text_level": 1, + "bbox": [ + 500, + 440, + 625, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our proposed method optimises camera pose and the NeRF model jointly and works on challenging scenes where other baselines fail. However, the optimisation of the model is also affected by non-linear distortions and the accuracy of the mono-depth estimation, which we did not consider.", + "bbox": [ + 496, + 464, + 890, + 539 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 554, + 617, + 569 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we present NoPe-NeRF, an end-to-end differentiable model for joint camera pose estimation and novel view synthesis from a sequence of images. We demonstrate that previous approaches have difficulty with complex trajectories. To tackle this challenge, we use mono-depth maps to constrain the relative poses between frames and regularise the geometry of NeRF, which leads to better pose estimation. We show the effectiveness and robustness of NoPe-NeRF on challenging scenes. The improved pose estimation leads to better novel view synthesis quality and geometry reconstruction compared with other approaches. We believe our method is an important step towards applying the unknown-pose NeRF models to large-scale scenes in the future.", + "bbox": [ + 496, + 579, + 890, + 790 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 500, + 806, + 666, + 821 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We thank Theo Costain, Michael Hobley, Shuai Chen and Xinghui Li for their helpful proofreading and discussions. Wenjing Bian is supported by the China Scholarship Council - University of Oxford Scholarship.", + "bbox": [ + 496, + 829, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "4167", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Chris Buehler, Michael Bosse, Leonard McMillan, Steven Gortler, and Michael Cohen. Unstructured lumigraph rendering. In SIGGRAPH, 2001. 2", + "[2] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In ECCV, 2022. 2", + "[3] Shenchang Eric Chen and Lance Williams. View interpolation for image synthesis. In SIGGRAPH, 1993. 2", + "[4] Shin-Fang Chng, Sameera Ramasinghe, Jamie Sherrah, and Simon Lucey. Garf: Gaussian activated radiance fields for high fidelity reconstruction and pose estimation. arXiv eprints, 2022. 2", + "[5] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, 2017. 5", + "[6] Paul E Debevec, Camillo J Taylor, and Jitendra Malik. Modeling and rendering architecture from photographs: A hybrid geometry-and image-based approach. In SIGGRAPH, 1996. 2", + "[7] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised nerf: Fewer views and faster training for free. In CVPR, 2022. 2", + "[8] Huan Fu, Mingming Gong, Chaohui Wang, Kayhan Bat-manghelich, and Dacheng Tao. Deep ordinal regression network for monocular depth estimation. In CVPR, 2018. 5", + "[9] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video. In ICCV, 2021. 2", + "[10] Stephan J Garbin, Marek Kowalski, Matthew Johnson, Jamie Shotton, and Julien Valentin. Fastnerf: High-fidelity neural rendering at 200fps. In ICCV, 2021. 2", + "[11] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. 2003. 2", + "[12] Yoonwoo Jeong, Seokjun Ahn, Christopher Choy, Anima Anandkumar, Minsu Cho, and Jaesik Park. Self-calibrating neural radiance fields. In ICCV, 2021. 1, 2, 3, 4, 5, 7", + "[13] Mijeong Kim, Seonguk Seo, and Bohyung Han. Infonerf: Ray entropy minimization for few-shot neural volume rendering. In CVPR, 2022. 2", + "[14] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In *Yoshua Bengio and Yann LeCun*, editors, *ICLR*, 2015. 5", + "[15] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics, 2017. 5", + "[16] Johannes Kopf, Xuejian Rong, and Jia-Bin Huang. Robust consistent video depth estimation. In CVPR, 2021. 5", + "[17] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In CVPR, 2021. 2", + "[18] Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. Barf: Bundle-adjusting neural radiance fields. In ICCV, 2021. 1, 2, 3, 4, 5, 7" + ], + "bbox": [ + 78, + 116, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[19] Fayao Liu, Chunhua Shen, Guosheng Lin, and Ian Reid. Learning depth from single monocular images using deep convolutional neural fields. TPAMI, 2015. 5", + "[20] Xuan Luo, Jia-Bin Huang, Richard Szeliski, Kevin Matzen, and Johannes Kopf. Consistent video depth estimation. ToG. 5", + "[21] Quan Meng, Anpei Chen, Haimin Luo, Minye Wu, Hao Su, Lan Xu, Xuming He, and Jingyi Yu. Gnerf: Gan-based neural radiance field without posed camera. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. 2", + "[22] S Mahdi H Miangoleh, Sebastian Dille, Long Mai, Sylvain Paris, and Yagiz Aksoy. Boosting monocular depth estimation models to high-resolution via content-adaptive multi-resolution merging. In CVPR, 2021. 2", + "[23] Ben Mildenhall, Pratul P Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM Transactions on Graphics (TOG), 2019. 2", + "[24] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 2021. 1, 2, 3, 5, 7", + "[25] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph. 2", + "[26] Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In CVPR, 2022. 2", + "[27] Michael Oechsle, Songyou Peng, and Andreas Geiger. Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In ICCV, 2021. 2", + "[28] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In ICCV, 2021. 2, 4", + "[29] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. TPAMI, 2020. 2", + "[30] Gernot Riegler and Vladlen Koltun. Free view synthesis. In ECCV, 2020. 2", + "[31] Gernot Riegler and Vladlen Koltun. Stable view synthesis. In CVPR, 2021. 2", + "[32] Barbara Roessle, Jonathan T Barron, Ben Mildenhall, Pratul P Srinivasan, and Matthias Nießner. Dense depth priors for neural radiance fields from sparse input views. In CVPR, 2022. 2", + "[33] Antoni Rosinol, John J Leonard, and Luca Carlone. Nerf-slam: Real-time dense monocular slam with neural radiance fields. arXiv preprint arXiv:2210.13641, 2022. 2", + "[34] Sara Fridovich-Keil and Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In CVPR, 2022. 2", + "[35] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In CVPR, 2016. 1, 2" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "4168", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[36] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. NeurIPS, 2020. 2", + "[37] Jürgen Sturm, Nikolas Engelhard, Felix Endres, Wolfram Burgard, and Daniel Cremers. A benchmark for the evaluation of rgb-d slam systems. In IROS. IEEE, 2012. 5", + "[38] Edgar Sucar, Shikun Liu, Joseph Ortiz, and Andrew Davison. iMAP: Implicit mapping and positioning in real-time. In ICCV, 2021. 2", + "[39] Libo Sun, Jia-Wang Bian, Huangying Zhan, Wei Yin, Ian Reid, and Chunhua Shen. Sc-depthv3: Robust self-supervised monocular depth estimation for dynamic scenes. arXiv preprint arXiv:2211.03660, 2022. 5", + "[40] Alex Trevithick and Bo Yang. Grf: Learning a general radiance field for 3d representation and rendering. In ICCV, 2021. 2", + "[41] Richard Tucker and Noah Snavely. Single-view view synthesis with multiplane images. In CVPR, 2020. 2", + "[42] Shubham Tulsiani, Richard Tucker, and Noah Snavely. Layer-structured 3d scene inference via view synthesis. In ECCV, 2018. 2", + "[43] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. NeurIPS, 2021. 2", + "[44] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P Srinivasan, Howard Zhou, Jonathan T Barron, Ricardo Martin-Brualla, Noah Snively, and Thomas Funkhouser. Ibrnet: Learning multi-view image-based rendering. In CVPR, 2021. 2", + "[45] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 2004. 5", + "[46] Zirui Wang, Shangzhe Wu, Weidi Xie, Min Chen, and Victor Adrian Prisacariu. NeRF---: Neural radiance fields without known camera parameters. arXiv preprint arXiv:2102.07064, 2021. 1, 2, 3, 4, 5, 6, 7", + "[47] Yi Wei, Shaohui Liu, Yongming Rao, Wang Zhao, Jiwen Lu, and Jie Zhou. Nerfingmvs: Guided optimization of neural radiance fields for indoor multi-view stereo. In ICCV, 2021. 2", + "[48] Yitong Xia, Hao Tang, Radu Timofte, and Luc Van Gool. Sinerf: Sinusoidal neural radiance fields for joint pose estimation and scene reconstruction. 2022. 2", + "[49] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. In NeurIPS, 2021. 2", + "[50] Lin Yen-Chen, Pete Florence, Jonathan T Barron, Alberto Rodriguez, Phillip Isola, and Tsung-Yi Lin. inerf: Inverting neural radiance fields for pose estimation. In IROS, 2021. 2", + "[51] Wei Yin, Jianming Zhang, Oliver Wang, Simon Niklaus, Simon Chen, Yifan Liu, and Chunhua Shen. Towards accurate reconstruction of 3d scene shape from a single monocular image. TPAMI, 2022. 2", + "[52] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In CVPR, 2021. 2" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[53] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. NeurIPS, 2022. 2", + "[54] Jian Zhang, Yuanqing Zhang, Huan Fu, Xiaowei Zhou, Bowen Cai, Jinchi Huang, Rongfei Jia, Binqiang Zhao, and Xing Tang. Ray priors through reprojection: Improving neural radiance fields for novel view extrapolation. In CVPR, 2022. 2", + "[55] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492, 2020. 2", + "[56] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 5", + "[57] Zichao Zhang and Davide Scaramuzza. A tutorial on quantitative trajectory evaluation for visual (-inertial) odometry. In IROS. IEEE, 2018. 5", + "[58] Tinghui Zhou, Matthew Brown, Noah Snavely, and David G Lowe. Unsupervised learning of depth and ego-motion from video. In CVPR, 2017. 5, 7", + "[59] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images. 2018. 2", + "[60] Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hujun Bao, Zhaopeng Cui, Martin R. Oswald, and Marc Pollefeys. Nice-slam: Neural implicit scalable encoding for slam. In CVPR, 2022. 2" + ], + "bbox": [ + 501, + 90, + 890, + 487 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "4169", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/NoPe-NeRF_ Optimising Neural Radiance Field With No Pose Prior/2260561b-6c24-4454-9ee0-0e88425e374a_model.json b/2023/NoPe-NeRF_ Optimising Neural Radiance Field With No Pose Prior/2260561b-6c24-4454-9ee0-0e88425e374a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..24d926f1cb8afb0e68a1e5bb7b3e6701f635c536 --- /dev/null +++ b/2023/NoPe-NeRF_ Optimising Neural Radiance Field With No Pose Prior/2260561b-6c24-4454-9ee0-0e88425e374a_model.json @@ -0,0 +1,2497 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.147, + 0.131, + 0.825, + 0.152 + ], + "angle": 0, + "content": "NoPe-NeRF: Optimising Neural Radiance Field with No Pose Prior" + }, + { + "type": "text", + "bbox": [ + 0.207, + 0.181, + 0.319, + 0.2 + ], + "angle": 0, + "content": "Wenjing Bian" + }, + { + "type": "text", + "bbox": [ + 0.373, + 0.182, + 0.466, + 0.199 + ], + "angle": 0, + "content": "Zirui Wang" + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.182, + 0.59, + 0.199 + ], + "angle": 0, + "content": "Kejie Li" + }, + { + "type": "text", + "bbox": [ + 0.645, + 0.182, + 0.764, + 0.199 + ], + "angle": 0, + "content": "Jia-Wang Bian" + }, + { + "type": "text", + "bbox": [ + 0.387, + 0.2, + 0.583, + 0.216 + ], + "angle": 0, + "content": "Victor Adrian Prisacariu" + }, + { + "type": "text", + "bbox": [ + 0.326, + 0.218, + 0.646, + 0.234 + ], + "angle": 0, + "content": "Active Vision Lab, University of Oxford" + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.237, + 0.725, + 0.252 + ], + "angle": 0, + "content": "{wenjing, ryan, kejie, jiawang, victor}@robots.ox.ac.uk" + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.286, + 0.216, + 0.379 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.381, + 0.216, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.219, + 0.287, + 0.484, + 0.379 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.379, + 0.484, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.484, + 0.287, + 0.751, + 0.379 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.484, + 0.379, + 0.751, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.754, + 0.287, + 0.875, + 0.379 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.754, + 0.379, + 0.875, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.485, + 0.893, + 0.514 + ], + "angle": 0, + "content": "Figure 1. Novel view synthesis comparison. We propose NoPe-NeRF for joint pose estimation and novel view synthesis. Our method enables more robust pose estimation and renders better novel view synthesis than previous state-of-the-art methods." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.526, + 0.314, + 0.541 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.558, + 0.473, + 0.816 + ], + "angle": 0, + "content": "Training a Neural Radiance Field (NeRF) without precomputed camera poses is challenging. Recent advances in this direction demonstrate the possibility of jointly optimising a NeRF and camera poses in forward-facing scenes. However, these methods still face difficulties during dramatic camera movement. We tackle this challenging problem by incorporating undistorted monocular depth priors. These priors are generated by correcting scale and shift parameters during training, with which we are then able to constrain the relative poses between consecutive frames. This constraint is achieved using our proposed novel loss functions. Experiments on real-world indoor and outdoor scenes show that our method can handle challenging camera trajectories and outperforms existing methods in terms of novel view rendering quality and pose estimation accuracy. Our project page is https://nope-nerf.active.vision." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.845, + 0.208, + 0.86 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.471, + 0.903 + ], + "angle": 0, + "content": "The photo-realistic reconstruction of a scene from a stream of RGB images requires both accurate 3D geometry" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.527, + 0.892, + 0.603 + ], + "angle": 0, + "content": "reconstruction and view-dependent appearance modelling. Recently, Neural Radiance Fields (NeRF) [24] have demonstrated the ability to build high-quality results for generating photo-realistic images from novel viewpoints given a sparse set of images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.608, + 0.893, + 0.805 + ], + "angle": 0, + "content": "An important preparation step for NeRF training is the estimation of camera parameters for the input images. A current go-to option is the popular Structure-from-Motion (SfM) library COLMAP [35]. Whilst easy to use, this preprocessing step could be an obstacle to NeRF research and real-world deployments in the long term due to its long processing time and its lack of differentiability. Recent works such as NeRFmm [46], BARF [18] and SC-NeRF [12] propose to simultaneously optimise camera poses and the neural implicit representation to address these issues. Nevertheless, these methods can only handle forward-facing scenes when no initial parameters are supplied, and fail in dramatic camera motions, e.g.a casual handheld captured video." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.894, + 0.902 + ], + "angle": 0, + "content": "This limitation has two key causes. First, all these methods estimate a camera pose for each input image individually without considering relative poses between images. Looking back to the literature of Simultaneous localisation and mapping (SLAM) and visual odometry, pose estimation can significantly benefit from estimating relative poses be" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4160" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.166 + ], + "angle": 0, + "content": "tween adjacent input frames. Second, the radiance field is known to suffer from shape-radiance ambiguity [55]. Estimating camera parameters jointly with NeRF adds another degree of ambiguity, resulting in slow convergence and unstable optimisation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.168, + 0.468, + 0.317 + ], + "angle": 0, + "content": "To handle the limitation of large camera motion, we seek help from monocular depth estimation [22, 28, 29, 51]. Our motivation is threefold: First, monocular depth provides strong geometry cues that are beneficial to constraint shape-raliance ambiguity. Second, relative poses between adjacent depth maps can be easily injected into the training pipeline via Chamfer Distance. Third, monocular depth is lightweight to run and does not require camera parameters as input, in contrast to multi-view stereo depth estimation. For simplicity, we use the term mono-depth from now on." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.318, + 0.468, + 0.573 + ], + "angle": 0, + "content": "Utilising mono-depth effectively is not straightforward with the presence of scale and shift distortions. In other words, mono-depth maps are not multi-view consistent. Previous works [9, 17, 47] simply take mono-depth into a depth-wise loss along with NeRF training. Instead, we propose a novel and effective way to thoroughly integrate mono-depth into our system. First, we explicitly optimise scale and shift parameters for each mono-depth map during NeRF training by penalising the difference between rendered depth and mono-depth. Since NeRF by itself is trained based on multiview consistency, this step transforms mono-depth maps to undistorted multiview consistent depth maps. We further leverage these multiview consistent depth maps in two loss terms: a) a Chamfer Distance loss between two depth maps of adjacent images, which injects relative pose to our system; and b) a depth-based surface rendering loss, which further improves relative pose estimation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.575, + 0.468, + 0.725 + ], + "angle": 0, + "content": "In summary, we propose a method to jointly optimise camera poses and a NeRF from a sequence of images with large camera motion. Our system is enabled by three contributions. First, we propose a novel way to integrate monodepth into unposed-NeRF training by explicitly modelling scale and shift distortions. Second, we supply relative poses to the camera-NeRF joint optimisation via an inter-frame loss using undistorted mono-depth maps. Third, we further regularise our relative pose estimation with a depth-based surface rendering loss." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.726, + 0.468, + 0.786 + ], + "angle": 0, + "content": "As a result, our method is able to handle large camera motion, and outperforms state-of-the-art methods by a significant margin in terms of novel view synthesis quality and camera trajectory accuracy." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.8, + 0.218, + 0.815 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.468, + 0.9 + ], + "angle": 0, + "content": "Novel View Synthesis. While early Novel View Synthesis (NVS) approaches applied interpolations between pixels [3], later works often rendered images from 3D reconstruction [1, 6]. In recent years, different representations of the 3D scene are used, e.g.meshes [30, 31], Multi-Plane" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.891, + 0.136 + ], + "angle": 0, + "content": "Images [41, 59], layered depth [42] etc. Among them, NeRF [24] has become a popular scene representation for its photorealistic rendering." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.137, + 0.892, + 0.243 + ], + "angle": 0, + "content": "A number of techniques are proposed to improve NeRF's performance with additional regularisation [13, 26, 54], depth priors [7, 32, 47, 53], surface enhancements [27, 43, 49] or latent codes [40, 44, 52]. Other works [2, 10, 25, 34] have also accelerated NeRF training and rendering. However, most of these approaches require pre-computed camera parameters obtained from SfM algorithms [11, 35]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.26, + 0.892, + 0.682 + ], + "angle": 0, + "content": "NeRF With Pose Optimisation. Removing camera parameter preprocessing is an active line of research recently. One category of the methods [33, 38, 60] use a SLAM-style pipeline, that either requires RGB-D inputs or relies on accurate camera poses generated from the SLAM tracking system. Another category of works optimises camera poses with the NeRF model directly. We term this type of method as unposed-NeRF in this paper. iNeRF [50] shows that poses for novel view images can be estimated using a reconstructed NeRF model. GNeRF [21] combines Generative Adversarial Networks with NeRF to estimate camera poses but requires a known sampling distribution for poses. More relevant to our work, NeRFmm [46] jointly optimises both camera intrinsic and extrinsic alongside NeRF training. BARF [18] proposes a coarse-to-fine positional encoding strategy for camera poses and NeRF joint optimisation. SCNeRF [12] further parameterises camera distortion and employs a geometric loss to regularise rays. GARF [4] shows that using Gaussian-MLPs makes joint pose and scene optimisation easier and more accurate. Recently, SiNeRF [48] uses SIREN [36] layers and a novel sampling strategy to alleviate the sub-optimality of joint optimisation in NeRFmm. Although showing promising results on the forward-facing dataset like LLFF [23], these approaches face difficulties when handling challenging camera trajectories with large camera motion. We address this issue by closely integrating mono-depth maps with the joint optimisation of camera parameters and NeRF." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.695, + 0.59, + 0.71 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.891, + 0.87 + ], + "angle": 0, + "content": "We tackle the challenge of handling large camera motions in unposed-NeRF training. Given a sequence of images, camera intrinsics, and their mono-depth estimations, our method recovers camera poses and optimises a NeRF simultaneously. We assume camera intrinsics are available in the image meta block, and we run an off-the-shelf monodepth network DPT [7] to acquire mono-depth estimations. Without repeating the benefit of mono-depth, we unroll this section around the effective integration of monocular depth into unposed-NeRF training." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.891, + 0.901 + ], + "angle": 0, + "content": "The training is a joint optimisation of the NeRF, camera poses, and distortion parameters of each mono-depth" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.514, + 0.956 + ], + "angle": 0, + "content": "4161" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.088, + 0.52, + 0.371 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.088, + 0.738, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.754, + 0.089, + 0.892, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.382, + 0.895, + 0.424 + ], + "angle": 0, + "content": "Figure 2. Method Overview. Our method takes a sequence of images as input to reconstruct NeRF and jointly estimates the camera poses of the frames. We first generate monocular depth maps from a mono-depth estimation network and reconstruct the point clouds. We then optimise NeRF, camera poses, and depth distortion parameters jointly with inter-frame and NeRF losses." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.45, + 0.47, + 0.542 + ], + "angle": 0, + "content": "map. The distortion parameters are supervised by minimising the discrepancies between the mono-depth maps and depth maps rendered from the NeRF, which are multiview consistent. The undistorted depth maps in return effectively mediate the shape-radiance ambiguity, which eases the training of NeRF and camera poses." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.543, + 0.47, + 0.664 + ], + "angle": 0, + "content": "Specifically, the undistorted depth maps enable two constraints. We constrain global pose estimation by supplying relative pose between adjacent images. This is achieved via a Chamfer-Distance-based correspondence between two point clouds, back-projected from undistorted depth maps. Further, we regularise relative pose estimation with a surface-based photometric consistency where we treat undistorted depth as surface." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.665, + 0.47, + 0.743 + ], + "angle": 0, + "content": "We detail our method in the following sections, starting from NeRF in Sec. 3.1 and unposed-NeRF training in Sec. 3.2, looking into mono-depth distortions in Sec. 3.3, followed by our mono-depth enabled loss terms in Sec. 3.4, and finishing with an overall training pipeline Sec. 3.5." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.755, + 0.16, + 0.77 + ], + "angle": 0, + "content": "3.1. NeRF" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Neural Radiance Field (NeRF) [24] represents a scene as a mapping function \\( F_{\\Theta}:(\\mathbf{x},\\mathbf{d})\\to (\\mathbf{c},\\sigma) \\) that maps a 3D location \\( \\mathbf{x}\\in \\mathbb{R}^3 \\) and a viewing direction \\( \\mathbf{d}\\in \\mathbb{R}^3 \\) to a radiance colour \\( \\mathbf{c}\\in \\mathbb{R}^3 \\) and a volume density value \\( \\sigma \\). This mapping is usually implemented with a neural network parameterised by \\( F_{\\Theta} \\). Given \\( N \\) images \\( \\mathcal{I} = \\{I_i\\mid i = 0\\dots N - 1\\} \\) with their camera poses \\( \\Pi = \\{\\pi_i\\mid i = 0\\dots N - 1\\} \\), NeRF can be optimised by minimising photometric error" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.448, + 0.892, + 0.481 + ], + "angle": 0, + "content": "\\(\\mathcal{L}_{rgb} = \\sum_{i}^{N}\\| I_{i} - \\hat{I}_{i}\\|_{2}^{2}\\) between synthesised images \\(\\hat{\\mathcal{I}}\\) and captured images \\(\\mathcal{I}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.594, + 0.487, + 0.892, + 0.51 + ], + "angle": 0, + "content": "\\[\n\\Theta^ {*} = \\arg \\min _ {\\Theta} \\mathcal {L} _ {r g b} (\\hat {\\mathcal {I}} \\mid \\mathcal {I}, \\Pi), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.519, + 0.892, + 0.581 + ], + "angle": 0, + "content": "where \\(\\hat{I}_i\\) is rendered by aggregating radiance colour on camera rays \\(\\mathbf{r}(h) = \\mathbf{o} + h\\mathbf{d}\\) between near and far bound \\(h_n\\) and \\(h_f\\). More concretely, we synthesise \\(\\hat{I}_i\\) with a volumetric rendering function" + }, + { + "type": "equation", + "bbox": [ + 0.56, + 0.586, + 0.892, + 0.623 + ], + "angle": 0, + "content": "\\[\n\\hat {I} _ {i} (\\mathbf {r}) = \\int_ {h _ {n}} ^ {h _ {f}} T (h) \\sigma (\\mathbf {r} (h)) \\mathbf {c} (\\mathbf {r} (h), \\mathbf {d}) d h, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.631, + 0.892, + 0.678 + ], + "angle": 0, + "content": "where \\( T(h) = \\exp(-\\int_{h_n}^h \\sigma(\\mathbf{r}(s)) ds) \\) is the accumulated transmittance along a ray. We refer to [24] for further details." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.686, + 0.829, + 0.702 + ], + "angle": 0, + "content": "3.2. Joint Optimisation of Poses and NeRF" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.709, + 0.892, + 0.77 + ], + "angle": 0, + "content": "Prior works [12, 18, 46] show that it is possible to estimate both camera parameters and a NeRF at the same time by minimising the above photometric error \\(\\mathcal{L}_{rgb}\\) under the same volumetric rendering process in Eq. (2)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.77, + 0.892, + 0.829 + ], + "angle": 0, + "content": "The key lies in conditioning camera ray casting on variable camera parameters \\(\\Pi\\), as the camera ray \\(\\mathbf{r}\\) is a function of camera pose. Mathematically, this joint optimisation can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.58, + 0.836, + 0.892, + 0.861 + ], + "angle": 0, + "content": "\\[\n\\Theta^ {*}, \\Pi^ {*} = \\arg \\min _ {\\Theta , \\Pi} \\mathcal {L} _ {r g b} (\\hat {\\mathcal {I}}, \\hat {\\Pi} \\mid \\mathcal {I}), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.87, + 0.892, + 0.902 + ], + "angle": 0, + "content": "where \\(\\hat{\\Pi}\\) denotes camera parameters that are updated during optimising. Note that the only difference between Eq. (1)" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4162" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.12 + ], + "angle": 0, + "content": "and Eq. (3) is that Eq. (3) considers camera parameters as variables." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.122, + 0.468, + 0.198 + ], + "angle": 0, + "content": "In general, the camera parameters \\(\\Pi\\) include camera intrinsics, poses, and lens distortions. We only consider estimating camera poses in this work, e.g., camera pose for frame \\(I_{i}\\) is a transformation \\(\\mathbf{T}_i = [\\mathbf{R}_i\\mid \\mathbf{t}_i]\\) with a rotation \\(\\mathbf{R}_i\\in \\mathrm{SO}(3)\\) and a translation \\(\\mathbf{t}_i\\in \\mathbb{R}^3\\)" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.207, + 0.373, + 0.223 + ], + "angle": 0, + "content": "3.3. Undistortion of Monocular Depth" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.23, + 0.468, + 0.333 + ], + "angle": 0, + "content": "With an off-the-shelf monocular depth network, e.g., DPT [28], we generate mono-depth sequence \\(\\mathcal{D} = \\{D_i\\mid i = 0\\dots N - 1\\}\\) from input images. Without surprise, mono-depth maps are not multi-view consistent so we aim to recover a sequence of multi-view consistent depth maps, which are further leveraged in our relative pose loss terms." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.336, + 0.468, + 0.442 + ], + "angle": 0, + "content": "Specifically, we consider two linear transformation parameters for each mono-depth map, resulting in a sequence of transformation parameters for all frames \\(\\Psi = \\{(\\alpha_i,\\beta_i)\\mid i = 0\\ldots N - 1\\}\\), where \\(\\alpha_{i}\\) and \\(\\beta_{i}\\) denote a scale and a shift factor. With multi-view consistent constraint from NeRF, we aim to recover a multi-view consistent depth map \\(D_{i}^{*}\\) for \\(D_{i}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.212, + 0.453, + 0.469, + 0.47 + ], + "angle": 0, + "content": "\\[\nD _ {i} ^ {*} = \\alpha_ {i} D _ {i} + \\beta_ {i}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.481, + 0.469, + 0.543 + ], + "angle": 0, + "content": "by joint optimising \\(\\alpha_{i}\\) and \\(\\beta_{i}\\) along with a NeRF. This joint optimisation is mostly achieved by enforcing the consistency between an undistorted depth map \\(D_{i}^{*}\\) and a NeRF rendered depth map \\(\\hat{D}_i\\) via a depth loss:" + }, + { + "type": "equation", + "bbox": [ + 0.181, + 0.554, + 0.469, + 0.595 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {d e p t h}} = \\sum_ {i} ^ {N} \\left\\| D _ {i} ^ {*} - \\hat {D} _ {i} \\right\\|, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.606, + 0.124, + 0.618 + ], + "angle": 0, + "content": "where" + }, + { + "type": "equation", + "bbox": [ + 0.171, + 0.617, + 0.469, + 0.652 + ], + "angle": 0, + "content": "\\[\n\\hat {D} _ {i} (\\mathbf {r}) = \\int_ {h _ {n}} ^ {h _ {f}} T (h) \\sigma (\\mathbf {r} (h)) d h \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.658, + 0.433, + 0.672 + ], + "angle": 0, + "content": "denotes a volumetric rendered depth map from NeRF." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.672, + 0.469, + 0.763 + ], + "angle": 0, + "content": "It is important to note that both NeRF and mono-depth benefit from Eq. (5). On the one hand, mono-depth provides strong geometry prior for NeRF training, reducing shape-radiance ambiguity. On the other hand, NeRF provides multi-view consistency so we can recover a set of multiview consistent depth maps for relative pose estimations." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.772, + 0.304, + 0.787 + ], + "angle": 0, + "content": "3.4. Relative Pose Constraint" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Aforementioned unposed-NeRF methods [12, 18, 46] optimise each camera pose independently, resulting in an overfit to target images with incorrect poses. Penalising incorrect relative poses between frames can help to regularise the joint optimisation towards smooth convergence, especially in a complex camera trajectory. Therefore, we propose two losses that constrain relative poses." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "Point Cloud Loss. We back-project the undistorted depth maps \\(\\mathcal{D}^*\\) using the known camera intrinsics, to point clouds \\(\\mathcal{P}^* = \\{P_i^* \\mid i = 0 \\dots N - 1\\}\\) and optimise the relative pose between consecutive point clouds by minimising a point cloud loss \\(\\mathcal{L}_{pc}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.602, + 0.177, + 0.892, + 0.212 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {p c} = \\sum_ {(i, j)} l _ {c d} \\left(P _ {j} ^ {*}, \\mathbf {T} _ {j i} P _ {i} ^ {*}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.222, + 0.892, + 0.281 + ], + "angle": 0, + "content": "where \\(\\mathbf{T}_{ji} = \\mathbf{T}_j\\mathbf{T}_i^{-1}\\) represents the related pose that transforms point cloud \\(P_{i}^{*}\\) to \\(P_{j}^{*}\\), tuple \\((i,j)\\) denotes indices of a consecutive pair of instances, and \\(l_{cd}\\) denotes Chamfer Distance:" + }, + { + "type": "equation", + "bbox": [ + 0.499, + 0.29, + 0.894, + 0.337 + ], + "angle": 0, + "content": "\\[\nl _ {c d} \\left(P _ {i}, P _ {j}\\right) = \\sum_ {p _ {i} \\in P _ {i}} \\min _ {p _ {j} \\in P _ {j}} \\| p _ {i} - p _ {j} \\| _ {2} + \\sum_ {p _ {j} \\in P _ {j}} \\min _ {p _ {i} \\in P _ {i}} \\| p _ {i} - p _ {j} \\| _ {2}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.355, + 0.892, + 0.476 + ], + "angle": 0, + "content": "Surface-based Photometric Loss. While the point cloud loss \\(\\mathcal{L}_{pc}\\) offers supervision in terms of 3D-3D matching, we observe that a surface-based photometric error can alleviate incorrect matching. With the photometric consistency assumption, this photometric error penalises the differences in appearance between associated pixels. The association is established by projecting the point cloud \\(P_{i}^{*}\\) onto images \\(I_{i}\\) and \\(I_{j}\\)." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.477, + 0.892, + 0.504 + ], + "angle": 0, + "content": "The surface-based photometric loss can then be defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.513, + 0.892, + 0.547 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {r g b - s} = \\sum_ {(i, j)} \\| I _ {i} \\langle \\mathbf {K} _ {i} P _ {i} ^ {*} \\rangle - I _ {j} \\langle \\mathbf {K} _ {j} \\mathbf {T} _ {j} \\mathbf {T} _ {i} ^ {- 1} P _ {i} ^ {*} \\rangle \\|, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.557, + 0.891, + 0.587 + ], + "angle": 0, + "content": "where \\(\\langle \\cdot \\rangle\\) represents the sampling operation on the image and \\(\\mathbf{K}_i\\) denotes a projection matrix for \\(i_{th}\\) camera." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.596, + 0.733, + 0.612 + ], + "angle": 0, + "content": "3.5. Overall Training Pipeline" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.619, + 0.89, + 0.647 + ], + "angle": 0, + "content": "Assembling all loss terms, we get the overall loss function:" + }, + { + "type": "equation", + "bbox": [ + 0.532, + 0.659, + 0.891, + 0.676 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {r g b} + \\lambda_ {1} \\mathcal {L} _ {\\text {d e p t h}} + \\lambda_ {2} \\mathcal {L} _ {p c} + \\lambda_ {3} \\mathcal {L} _ {r g b - s}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.685, + 0.891, + 0.715 + ], + "angle": 0, + "content": "where \\(\\lambda_1, \\lambda_2, \\lambda_3\\) are the weighting factors for respective loss terms. By minimising the combined of loss \\(\\mathcal{L}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.723, + 0.891, + 0.748 + ], + "angle": 0, + "content": "\\[\n\\Theta^ {*}, \\Pi^ {*}, \\Psi^ {*} = \\arg \\min _ {\\Theta , \\Pi , \\Psi} \\mathcal {L} (\\hat {\\mathcal {I}}, \\hat {\\mathcal {D}}, \\hat {\\Pi}, \\hat {\\Psi} \\mid \\mathcal {I}, \\mathcal {D}), \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.757, + 0.891, + 0.788 + ], + "angle": 0, + "content": "our method returns the optimised NeRF parameters \\(\\Theta\\), camera poses \\(\\Pi\\), and distortion parameters \\(\\Psi\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.801, + 0.633, + 0.817 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.826, + 0.892, + 0.9 + ], + "angle": 0, + "content": "We begin with a description of our experimental setup in Sec. 4.1. In Sec. 4.2, we compare our method with pose-unknown methods. Next, we compare our method with the COLMAP-assisted NeRF baseline in Sec. 4.3. Lastly, we conduct ablation studies in Sec. 4.4." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4163" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.091, + 0.891, + 0.302 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.119, + 0.308, + 0.182, + 0.321 + ], + "angle": 0, + "content": "(a) BARF" + }, + { + "type": "image_caption", + "bbox": [ + 0.278, + 0.308, + 0.362, + 0.321 + ], + "angle": 0, + "content": "(b) NeRFmm" + }, + { + "type": "image_caption", + "bbox": [ + 0.442, + 0.308, + 0.526, + 0.321 + ], + "angle": 0, + "content": "(c) SC-NeRF" + }, + { + "type": "image_caption", + "bbox": [ + 0.621, + 0.309, + 0.674, + 0.321 + ], + "angle": 0, + "content": "(d) Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.748, + 0.309, + 0.856, + 0.321 + ], + "angle": 0, + "content": "(e) Ground Truth" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.333, + 0.893, + 0.361 + ], + "angle": 0, + "content": "Figure 3. Qualitative results of novel view synthesis and depth prediction on Tanks and Temples. We visualise the synthesised images and the rendered depth maps (top left of each image) for all methods. NoPe-NeRF is able to recover details for both colour and geometry." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.387, + 0.267, + 0.403 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.411, + 0.474, + 0.713 + ], + "angle": 0, + "content": "Datasets. We conduct experiments on two datasets Tanks and Temples [15] and ScanNet [5]. Tanks and Temples: we use 8 scenes to evaluate pose accuracy and novel view synthesis quality. We chose scenes captured at both indoor and outdoor locations, with different frame sampling rates and lengths. All images are down-sampled to a resolution of \\(960 \\times 540\\). For the family scene, we sample 200 images and take 100 frames with odd frame ids as training images and the remaining 100 frames for novel view synthesis, in order to analyse the performance under smooth motion. For the remaining scenes, following NeRF [24], 1/8 of the images in each sequence are held out for novel view synthesis, unless otherwise specified. ScanNet: we select 4 scenes for evaluating pose accuracy, depth accuracy, and novel view synthesis quality. For each scene, we take 80-100 consecutive images and use 1/8 of these images for novel view synthesis. For evaluation, we employ depth maps and poses provided by ScanNet as ground truth. ScanNet images are down-sampled to \\(648 \\times 484\\). We crop images with dark orders during preprocessing." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Metrics. We evaluate our proposed method in three aspects. For novel view synthesis, we follow previous methods [12,18,46], and use standard evaluation metrics, including Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM) [45] and Learned Perceptual Image Patch Similarity (LPIPS) [56]. For pose evaluation, We use standard visual odometry metrics [16, 37, 57], including the Absolute Trajectory Error (ATE) and Relative Pose Error (RPE). ATE measures the difference between the estimated camera positions and the ground truth positions. RPE measures the relative pose errors between pairs of images," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.387, + 0.895, + 0.525 + ], + "angle": 0, + "content": "which consists of relative rotation error \\((\\mathrm{RPE}_r)\\) and relative translation error \\((\\mathrm{RPE}_t)\\). The estimated trajectory is aligned with the ground truth using Sim(3) with 7 degrees of freedom. We use standard depth metrics [8,19,20,39] (Abs Rel, Sq Rel, RMSE, RMSE log, \\(\\delta_1\\), \\(\\delta_2\\) and \\(\\delta_3\\)) for depth evaluation. For further detail, please refer to the supplementary material. To recover the metric scale, we follow Zhou et al. [58] and match the median value between rendered and ground truth depth maps." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.547, + 0.895, + 0.82 + ], + "angle": 0, + "content": "Implementation Details. Our model architecture is based on NeRF [24] with a few modifications: a) replacing ReLU activation function with Softplus and b) sampling 128 points along each ray uniformly with noise, between a predefined range (0.1, 10). We use 2 separate Adam optimisers [14] for NeRF and other parameters. The initial learning rate for NeRF is 0.001 and for the pose and distortion is 0.0005. Camera rotations are optimised in axis-angle representation \\(\\phi_i \\in \\mathfrak{so}(3)\\). We first train the model with all losses with constant learning rates until the inter-frame losses converge. Then, we remove the inter-frame losses and depth loss to refine the model with the RGB loss only. We decay the learning rates with different schedulers to refine for 10,000 epochs. We balance the loss terms with \\(\\lambda_1 = 0.04\\), \\(\\lambda_2 = 1.0\\) and \\(\\lambda_3 = 1.0\\). For each training step, we randomly sample 1024 pixels (rays) from each input image and 128 samples per ray. More details are provided in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.831, + 0.859, + 0.848 + ], + "angle": 0, + "content": "4.2. Comparing With Pose-Unknown Methods" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.855, + 0.892, + 0.885 + ], + "angle": 0, + "content": "We compare our method with pose-unknown baselines, including BARF [18], NeRFmm [46] and SC-NeRF [12]." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.886, + 0.893, + 0.902 + ], + "angle": 0, + "content": "View Synthesis Quality. To obtain the camera poses of" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4164" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.127, + 0.095, + 0.852, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.386, + 0.893, + 0.415 + ], + "angle": 0, + "content": "Figure 4. Pose Estimation Comparison. We visualise the trajectory (3D plot) and relative rotation errors \\(\\mathrm{RPE}_r\\) (bottom colour bar) of each method on Ballroom and Museum. The colour bar on the right shows the relative scaling of colour. More results are in the supplementary." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.426, + 0.916, + 0.624 + ], + "angle": 0, + "content": "
scenesOursBARFNeRFmmSC-NeRF
PSNR ↑SSIM ↑LPIPS ↓PSNRSSIMLPIPSPSNRSSIMLPIPSPSNRSSIMLPIPS
ScanNet0079_0032.470.840.4132.310.830.4330.590.810.4931.330.820.46
0418_0031.330.790.3431.240.790.3530.000.770.4029.050.750.43
0301_0029.830.770.3629.310.760.3827.840.720.4529.450.770.35
0431_0033.830.910.3932.770.900.4131.440.880.4532.570.900.40
mean31.860.830.3831.410.820.3929.970.800.4530.600.810.41
Tanks and TemplesChurch25.170.730.3923.170.620.5221.640.580.5421.960.600.53
Barn26.350.690.4425.280.640.4823.210.610.5323.260.620.51
Museum26.770.760.3523.580.610.5522.370.610.5324.940.690.45
Family26.010.740.4123.040.610.5623.040.580.5622.600.630.51
Horse27.640.840.2624.090.720.4123.120.700.4325.230.760.37
Ballroom25.330.720.3820.660.500.6020.030.480.5722.640.610.48
Francis29.480.800.3825.850.690.5725.4000.690.5226.460.730.49
Ignatius23.960.610.4721.780.470.6021.160.450.6023.000.550.53
mean26.340.740.3923.420.610.5422.500.590.5423.760.650.48
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.634, + 0.893, + 0.663 + ], + "angle": 0, + "content": "Table 1. Novel view synthesis results on ScanNet and Tanks and Temples. Each baseline method is trained with its public code under the original settings and evaluated with the same evaluation protocol." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.47, + 0.84 + ], + "angle": 0, + "content": "test views for rendering, we minimise the photometric error of the synthesised images while keeping the NeRF model fixed, as in NeRFmm [46]. Each test pose is initialised with the learned pose of the training frame that is closest to it. We use the same pre-processing for all baseline approaches, which results in higher accuracy than their original implementations. More details are provided in the supplementary material. Our method outperforms all the baselines by a large margin. The quantitative results are summarised in Tab. 1, and qualitative results are shown in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.903 + ], + "angle": 0, + "content": "We recognised that because the test views, which are sampled from videos, are close to the training views, good results may be obtained due to overfitting to the training images. Therefore, we conduct an additional qualitative" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.795 + ], + "angle": 0, + "content": "evaluation on more novel views. Specifically, we fit aBezier curve from the estimated training poses and sample interpolated poses for each method to render novel view videos. Sampled results are shown in Fig. 5, and the rendered videos are in the supplementary material. These results show that our method renders photo-realistic images consistently, while other methods generate visible artifacts." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Camera Pose. Our method significantly outperforms other baselines in all metrics. The quantitative pose evaluation results are shown in Tab. 2. For ScanNet, we use the camera poses provided by the dataset as ground truth. For Tanks and Temples, not every video comes with ground truth poses, so we use COLMAP estimations for reference. Our estimated trajectory is better aligned with the ground" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4165" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.11, + 0.089, + 0.86, + 0.254 + ], + "angle": 0, + "content": "
scenesOursBARFNeRFmmSC-NeRF
RPEt ↓RPEr ↓ATE↓RPEtRPErATERPEtRPErATERPEtRPErATE
ScanNet0079_000.7520.2040.0231.1100.4800.0621.7060.6360.1002.0640.6640.115
0418_000.4550.1190.0151.3980.5380.0201.4020.4600.0131.5280.5020.016
0301_000.3990.1230.0131.3160.7770.2193.0970.8940.2881.1330.4220.056
0431_001.6250.2740.0696.0240.7540.1686.7990.6240.4964.1100.4990.205
mean0.8080.1800.0302.4620.6370.1173.2510.6540.2242.2090.5220.098
Tanks and TemplesChurch0.0340.0080.0080.1140.0380.0520.6260.1270.0650.8360.1870.108
Barn0.0460.0320.0040.3140.2650.0501.6290.4940.1591.3170.4290.157
Museum0.2070.2020.0203.4421.1280.2634.1341.0510.3468.3391.4910.316
Family0.0470.0150.0011.3710.5910.1152.7430.5370.1201.1710.4990.142
Horse0.1790.0170.0031.3330.3940.0141.3490.4340.0181.3660.4380.019
Ballroom0.0410.0180.0020.5310.2280.0180.4490.1770.0310.3280.1460.012
Francis0.0570.0090.0051.3210.5580.0821.6470.6180.2071.2330.4830.192
Ignatius0.0260.0050.0020.7360.3240.0291.3020.3790.0410.5330.2400.085
mean0.0800.0380.0061.0460.4410.0781.7350.4770.1231.8900.4890.129
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.265, + 0.893, + 0.294 + ], + "angle": 0, + "content": "Table 2. Pose accuracy on ScanNet and Tanks and Temples. Note that we use COLMAP poses in Tanks and Temples as the \"ground truth\". The unit of \\(\\mathrm{RPE}_r\\) is in degrees, ATE is in the ground truth scale and \\(\\mathrm{RPE}_t\\) is scaled by 100." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.317, + 0.466, + 0.392 + ], + "angle": 0, + "content": "
Abs Rel ↓Sq Rel ↓RMSE ↓RMSE log ↓δ1 ↑δ2 ↑δ3 ↑
Ours0.1410.1370.5680.1760.8280.9700.987
BARF0.3760.6840.9900.4010.4900.7510.884
NeRFmm0.5901.7211.6720.5870.3160.5600.743
SC-NeRF0.4170.6421.0790.4760.3620.6580.832
DPT0.1970.2460.7510.2260.7470.9340.975
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.403, + 0.47, + 0.459 + ], + "angle": 0, + "content": "Table 3. Depth map evaluation on ScanNet. Our depth estimation is more accurate than baseline models BARF [18], NeRFmm [46] and SC-NeRF [12]. Compared with DPT [58], we show our depth is more accurate after undistortion." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.487, + 0.469, + 0.532 + ], + "angle": 0, + "content": "truth than other methods, and our estimated rotation is two orders of magnitudes more accurate than others. We visualise the camera trajectories and rotations in Fig. 4." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.533, + 0.47, + 0.669 + ], + "angle": 0, + "content": "Depth. We evaluate the accuracy of the rendered depth maps on ScanNet, which provides the ground-truth depths for evaluation. Our rendered depth maps achieve superior accuracy over the previous alternatives. We also compare with the mono-depth maps estimated by DPT. Our rendered depth maps, after undistortion using multiview consistency in the NeRF optimisation, outperform DPT by a large margin. The results are summarised in Tab. 3, and sampled qualitative results are illustrated in Fig. 3." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.68, + 0.445, + 0.697 + ], + "angle": 0, + "content": "4.3. Comparing With COLMAP Assisted NeRF" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.471, + 0.902 + ], + "angle": 0, + "content": "We make a comparison of pose estimation accuracy between our method and COLMAP against ground truth poses in ScanNet. We achieve on-par accuracy with COLMAP, as shown in Tab. 4. We further analyse the novel view synthesis quality of the NeRF model trained with our learned poses to COLMAP poses on ScanNet and Tanks and Temples. The original NeRF training contains two stages, finding poses using COLMAP and optimising the scene representation. In order to make our comparison fairer, in this section only, we mimic a similar two-stage training as the original NeRF [24]. In the first stage, we train our method with all losses for camera pose estimation, i.e., mimicking the COLMAP processing. Then, we fix the optimised poses" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.321, + 0.891, + 0.546 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.563, + 0.892, + 0.646 + ], + "angle": 0, + "content": "Figure 5. Sampled frames from rendered novel view videos. For each method, we fit the learned trajectory with aBezier curve and uniformly sample new viewpoints for rendering. Our method generates significantly better results than previous methods, which show visible artifacts. The full rendered videos and details about generating novel views are provided in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.674, + 0.891, + 0.735 + ], + "angle": 0, + "content": "and train a NeRF model from scratch, using the same settings and loss as the original NeRF. This evaluation enables us to compare our estimated poses to the COLMAP poses indirectly, i.e., in terms of contribution to view synthesis." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.794 + ], + "angle": 0, + "content": "Our two-stage method outperforms the COLMAP-assisted NeRF baseline, which indicates a better pose estimation for novel view synthesis. The results are summarised in Tab. 5." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "As is commonly known, COLMAP performs poorly in low-texture scenes and sometimes fails to find accurate camera poses. Fig. 6 shows an example of a low-texture scene where COLMAP provides inaccurate pose estimation that causes NeRF to render images with visible artifacts. In contrast, our method renders high-quality images, thanks to robust optimisation of camera pose." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4166" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.08, + 0.089, + 0.466, + 0.177 + ], + "angle": 0, + "content": "
scenesOursCOLMAP
RPEt↓RPEr↓ATE↓RPEtRPErATE
0079_000.7520.2040.0230.6550.2210.012
0418_000.4550.1190.0150.4910.1240.016
0301_000.3990.1230.0130.4140.1360.009
0431_001.6250.2740.0691.2920.2490.051
mean0.8080.1800.0300.7130.1820.022
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.188, + 0.47, + 0.215 + ], + "angle": 0, + "content": "Table 4. Comparison of pose accuracy with COLMAP on ScanNet." + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.232, + 0.206, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.093, + 0.31, + 0.199, + 0.321 + ], + "angle": 0, + "content": "(a) COLMAP+NeRF" + }, + { + "type": "image", + "bbox": [ + 0.209, + 0.232, + 0.338, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.253, + 0.31, + 0.296, + 0.321 + ], + "angle": 0, + "content": "(b) Ours" + }, + { + "type": "image", + "bbox": [ + 0.338, + 0.232, + 0.466, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.359, + 0.31, + 0.445, + 0.321 + ], + "angle": 0, + "content": "(c) Ground Truth" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.34, + 0.47, + 0.381 + ], + "angle": 0, + "content": "Figure 6. COLMAP failure case. On a rotation-dominant sequence with low-texture areas, COLMAP fails to estimate correct poses, which results in artifacts in synthesised images." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.395, + 0.466, + 0.563 + ], + "angle": 0, + "content": "
scenesOursOurs-rCOLMAP+NeRF
PSNR ↑SSIM ↑LPIPS ↓PSNRSSIMLPIPSPSNRSSIMLPIPS
ScanNet0079_0032.470.840.4133.120.850.4031.980.830.43
0418_0031.330.790.3430.490.770.4030.600.780.40
0301_0029.830.770.3630.050.780.3430.010.780.36
0431_0033.830.910.3933.860.910.3933.540.910.39
mean31.860.830.3831.880.830.3831.530.820.40
Tanks and TemplesChurch25.170.730.3926.740.780.3225.720.750.37
Barn26.350.690.4426.580.710.4226.720.710.42
Museum26.770.760.3526.980.770.3627.210.780.34
Family26.010.740.4126.210.750.4026.610.770.39
Horse27.640.840.2628.060.840.2627.020.820.29
Ballroom25.330.720.3825.530.730.3825.470.730.38
Francis29.480.800.3829.730.810.3830.050.810.38
Ignatius23.960.610.4723.980.620.4624.080.610.47
mean26.340.740.3926.730.750.3726.610.750.38
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.573, + 0.47, + 0.614 + ], + "angle": 0, + "content": "Table 5. Comparison to NeRF with COLMAP poses. Our two-stage method (Ours-r) outperforms both COLMAP+NeRF and our one-stage method (Ours)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.642, + 0.468, + 0.778 + ], + "angle": 0, + "content": "Interestingly, this experiment also reveals that the two-stage method shows higher accuracy than the one-stage method. We hypothesise that the joint optimisation (from randomly initialised poses) in the one-stage approach causes the NeRF optimisation to be trapped in a local minimum, potentially due to the bad pose initialisation. The two-stage approach circumvents this issue by re-initialising the NeRF and re-training with well-optimised poses, resulting in higher performance." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.787, + 0.229, + 0.803 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.468, + 0.855 + ], + "angle": 0, + "content": "In this section, we analyse the effectiveness of the parameters and components that have been added to our model. The results of ablation studies are shown in Tab. 6." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Effect of Distortion Parameters. We find that ignoring depth distortions (i.e., setting scales to 1 and shifts to 0 as constants) leads to a degradation in pose accuracy, as in" + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.089, + 0.889, + 0.181 + ], + "angle": 0, + "content": "
NVSPose
PSNR↑SSIM↑LPIPS↓RPEt↓RPEr↓ATE↓
Ours31.860.830.380.8010.1810.031
Ours w/o α,β31.460.820.391.9290.3210.066
Ours w/o Lpc31.730.820.382.2270.4530.101
Ours w/o Lrgb-s31.050.810.411.8140.4010.156
Ours w/o Ldepth31.200.810.401.4980.3830.089
" + }, + { + "type": "table_caption", + "bbox": [ + 0.559, + 0.192, + 0.832, + 0.206 + ], + "angle": 0, + "content": "Table 6. Ablation study results on ScanNet." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.234, + 0.892, + 0.277 + ], + "angle": 0, + "content": "consistent distortions of depth maps introduce errors to the estimation of relative poses and confuse NeRF for geometry reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.279, + 0.892, + 0.368 + ], + "angle": 0, + "content": "Effect of Inter-frame Losses. We observe that the interframe losses are the major contributor to improving relative poses. When removing the pairwise point cloud loss \\( L_{pc} \\) or the surface-based photometric loss \\( L_{rgb-s} \\), there is less constraint between frames, and thus the pose accuracy becomes lower." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.37, + 0.892, + 0.431 + ], + "angle": 0, + "content": "Effect of NeRF Losses. When the depth loss \\( L_{depth} \\) is removed, the distortions of input depth maps are only optimised locally through the inter-frame losses. We find that this can lead to drift and degradation in pose accuracy." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.441, + 0.627, + 0.456 + ], + "angle": 0, + "content": "4.5. Limitations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.465, + 0.892, + 0.54 + ], + "angle": 0, + "content": "Our proposed method optimises camera pose and the NeRF model jointly and works on challenging scenes where other baselines fail. However, the optimisation of the model is also affected by non-linear distortions and the accuracy of the mono-depth estimation, which we did not consider." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.555, + 0.619, + 0.57 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.58, + 0.892, + 0.791 + ], + "angle": 0, + "content": "In this work, we present NoPe-NeRF, an end-to-end differentiable model for joint camera pose estimation and novel view synthesis from a sequence of images. We demonstrate that previous approaches have difficulty with complex trajectories. To tackle this challenge, we use mono-depth maps to constrain the relative poses between frames and regularise the geometry of NeRF, which leads to better pose estimation. We show the effectiveness and robustness of NoPe-NeRF on challenging scenes. The improved pose estimation leads to better novel view synthesis quality and geometry reconstruction compared with other approaches. We believe our method is an important step towards applying the unknown-pose NeRF models to large-scale scenes in the future." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.807, + 0.667, + 0.823 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.83, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We thank Theo Costain, Michael Hobley, Shuai Chen and Xinghui Li for their helpful proofreading and discussions. Wenjing Bian is supported by the China Scholarship Council - University of Oxford Scholarship." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4167" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.117, + 0.47, + 0.158 + ], + "angle": 0, + "content": "[1] Chris Buehler, Michael Bosse, Leonard McMillan, Steven Gortler, and Michael Cohen. Unstructured lumigraph rendering. In SIGGRAPH, 2001. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.16, + 0.471, + 0.2 + ], + "angle": 0, + "content": "[2] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.204, + 0.471, + 0.232 + ], + "angle": 0, + "content": "[3] Shenchang Eric Chen and Lance Williams. View interpolation for image synthesis. In SIGGRAPH, 1993. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.234, + 0.471, + 0.289 + ], + "angle": 0, + "content": "[4] Shin-Fang Chng, Sameera Ramasinghe, Jamie Sherrah, and Simon Lucey. Garf: Gaussian activated radiance fields for high fidelity reconstruction and pose estimation. arXiv eprints, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.291, + 0.471, + 0.346 + ], + "angle": 0, + "content": "[5] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.349, + 0.471, + 0.403 + ], + "angle": 0, + "content": "[6] Paul E Debevec, Camillo J Taylor, and Jitendra Malik. Modeling and rendering architecture from photographs: A hybrid geometry-and image-based approach. In SIGGRAPH, 1996. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.406, + 0.471, + 0.448 + ], + "angle": 0, + "content": "[7] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised nerf: Fewer views and faster training for free. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.45, + 0.471, + 0.491 + ], + "angle": 0, + "content": "[8] Huan Fu, Mingming Gong, Chaohui Wang, Kayhan Bat-manghelich, and Dacheng Tao. Deep ordinal regression network for monocular depth estimation. In CVPR, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.494, + 0.471, + 0.535 + ], + "angle": 0, + "content": "[9] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.538, + 0.47, + 0.579 + ], + "angle": 0, + "content": "[10] Stephan J Garbin, Marek Kowalski, Matthew Johnson, Jamie Shotton, and Julien Valentin. Fastnerf: High-fidelity neural rendering at 200fps. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.581, + 0.47, + 0.609 + ], + "angle": 0, + "content": "[11] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. 2003. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.611, + 0.47, + 0.652 + ], + "angle": 0, + "content": "[12] Yoonwoo Jeong, Seokjun Ahn, Christopher Choy, Anima Anandkumar, Minsu Cho, and Jaesik Park. Self-calibrating neural radiance fields. In ICCV, 2021. 1, 2, 3, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.654, + 0.47, + 0.696 + ], + "angle": 0, + "content": "[13] Mijeong Kim, Seonguk Seo, and Bohyung Han. Infonerf: Ray entropy minimization for few-shot neural volume rendering. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.698, + 0.47, + 0.739 + ], + "angle": 0, + "content": "[14] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In *Yoshua Bengio and Yann LeCun*, editors, *ICLR*, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.742, + 0.47, + 0.784 + ], + "angle": 0, + "content": "[15] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.786, + 0.47, + 0.814 + ], + "angle": 0, + "content": "[16] Johannes Kopf, Xuejian Rong, and Jia-Bin Huang. Robust consistent video depth estimation. In CVPR, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.816, + 0.47, + 0.857 + ], + "angle": 0, + "content": "[17] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[18] Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. Barf: Bundle-adjusting neural radiance fields. In ICCV, 2021. 1, 2, 3, 4, 5, 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.117, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[19] Fayao Liu, Chunhua Shen, Guosheng Lin, and Ian Reid. Learning depth from single monocular images using deep convolutional neural fields. TPAMI, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[20] Xuan Luo, Jia-Bin Huang, Richard Szeliski, Kevin Matzen, and Johannes Kopf. Consistent video depth estimation. ToG. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[21] Quan Meng, Anpei Chen, Haimin Luo, Minye Wu, Hao Su, Lan Xu, Xuming He, and Jingyi Yu. Gnerf: Gan-based neural radiance field without posed camera. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.304 + ], + "angle": 0, + "content": "[22] S Mahdi H Miangoleh, Sebastian Dille, Long Mai, Sylvain Paris, and Yagiz Aksoy. Boosting monocular depth estimation models to high-resolution via content-adaptive multi-resolution merging. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.305, + 0.892, + 0.374 + ], + "angle": 0, + "content": "[23] Ben Mildenhall, Pratul P Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM Transactions on Graphics (TOG), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.376, + 0.892, + 0.431 + ], + "angle": 0, + "content": "[24] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 2021. 1, 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.433, + 0.892, + 0.473 + ], + "angle": 0, + "content": "[25] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.475, + 0.892, + 0.531 + ], + "angle": 0, + "content": "[26] Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.532, + 0.892, + 0.573 + ], + "angle": 0, + "content": "[27] Michael Oechsle, Songyou Peng, and Andreas Geiger. Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.575, + 0.892, + 0.614 + ], + "angle": 0, + "content": "[28] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In ICCV, 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.617, + 0.892, + 0.671 + ], + "angle": 0, + "content": "[29] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. TPAMI, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.673, + 0.892, + 0.701 + ], + "angle": 0, + "content": "[30] Gernot Riegler and Vladlen Koltun. Free view synthesis. In ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.703, + 0.892, + 0.73 + ], + "angle": 0, + "content": "[31] Gernot Riegler and Vladlen Koltun. Stable view synthesis. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.732, + 0.892, + 0.785 + ], + "angle": 0, + "content": "[32] Barbara Roessle, Jonathan T Barron, Ben Mildenhall, Pratul P Srinivasan, and Matthias Nießner. Dense depth priors for neural radiance fields from sparse input views. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.788, + 0.892, + 0.829 + ], + "angle": 0, + "content": "[33] Antoni Rosinol, John J Leonard, and Luca Carlone. Nerf-slam: Real-time dense monocular slam with neural radiance fields. arXiv preprint arXiv:2210.13641, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.831, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[34] Sara Fridovich-Keil and Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[35] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In CVPR, 2016. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "4168" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[36] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. NeurIPS, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.134, + 0.47, + 0.176 + ], + "angle": 0, + "content": "[37] Jürgen Sturm, Nikolas Engelhard, Felix Endres, Wolfram Burgard, and Daniel Cremers. A benchmark for the evaluation of rgb-d slam systems. In IROS. IEEE, 2012. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.177, + 0.469, + 0.217 + ], + "angle": 0, + "content": "[38] Edgar Sucar, Shikun Liu, Joseph Ortiz, and Andrew Davison. iMAP: Implicit mapping and positioning in real-time. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.218, + 0.469, + 0.273 + ], + "angle": 0, + "content": "[39] Libo Sun, Jia-Wang Bian, Huangying Zhan, Wei Yin, Ian Reid, and Chunhua Shen. Sc-depthv3: Robust self-supervised monocular depth estimation for dynamic scenes. arXiv preprint arXiv:2211.03660, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.274, + 0.469, + 0.314 + ], + "angle": 0, + "content": "[40] Alex Trevithick and Bo Yang. Grf: Learning a general radiance field for 3d representation and rendering. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.315, + 0.469, + 0.344 + ], + "angle": 0, + "content": "[41] Richard Tucker and Noah Snavely. Single-view view synthesis with multiplane images. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.344, + 0.469, + 0.384 + ], + "angle": 0, + "content": "[42] Shubham Tulsiani, Richard Tucker, and Noah Snavely. Layer-structured 3d scene inference via view synthesis. In ECCV, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.385, + 0.469, + 0.44 + ], + "angle": 0, + "content": "[43] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.441, + 0.469, + 0.509 + ], + "angle": 0, + "content": "[44] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P Srinivasan, Howard Zhou, Jonathan T Barron, Ricardo Martin-Brualla, Noah Snively, and Thomas Funkhouser. Ibrnet: Learning multi-view image-based rendering. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.51, + 0.469, + 0.565 + ], + "angle": 0, + "content": "[45] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 2004. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.566, + 0.469, + 0.622 + ], + "angle": 0, + "content": "[46] Zirui Wang, Shangzhe Wu, Weidi Xie, Min Chen, and Victor Adrian Prisacariu. NeRF---: Neural radiance fields without known camera parameters. arXiv preprint arXiv:2102.07064, 2021. 1, 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.622, + 0.469, + 0.676 + ], + "angle": 0, + "content": "[47] Yi Wei, Shaohui Liu, Yongming Rao, Wang Zhao, Jiwen Lu, and Jie Zhou. Nerfingmvs: Guided optimization of neural radiance fields for indoor multi-view stereo. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.678, + 0.469, + 0.718 + ], + "angle": 0, + "content": "[48] Yitong Xia, Hao Tang, Radu Timofte, and Luc Van Gool. Sinerf: Sinusoidal neural radiance fields for joint pose estimation and scene reconstruction. 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.719, + 0.469, + 0.761 + ], + "angle": 0, + "content": "[49] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. In NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.469, + 0.803 + ], + "angle": 0, + "content": "[50] Lin Yen-Chen, Pete Florence, Jonathan T Barron, Alberto Rodriguez, Phillip Isola, and Tsung-Yi Lin. inerf: Inverting neural radiance fields for pose estimation. In IROS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.469, + 0.858 + ], + "angle": 0, + "content": "[51] Wei Yin, Jianming Zhang, Oliver Wang, Simon Niklaus, Simon Chen, Yifan Liu, and Chunhua Shen. Towards accurate reconstruction of 3d scene shape from a single monocular image. TPAMI, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[52] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In CVPR, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[53] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.217 + ], + "angle": 0, + "content": "[54] Jian Zhang, Yuanqing Zhang, Huan Fu, Xiaowei Zhou, Bowen Cai, Jinchi Huang, Rongfei Jia, Binqiang Zhao, and Xing Tang. Ray priors through reprojection: Improving neural radiance fields for novel view extrapolation. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.218, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[55] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.262, + 0.892, + 0.304 + ], + "angle": 0, + "content": "[56] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.305, + 0.892, + 0.346 + ], + "angle": 0, + "content": "[57] Zichao Zhang and Davide Scaramuzza. A tutorial on quantitative trajectory evaluation for visual (-inertial) odometry. In IROS. IEEE, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.346, + 0.892, + 0.389 + ], + "angle": 0, + "content": "[58] Tinghui Zhou, Matthew Brown, Noah Snavely, and David G Lowe. Unsupervised learning of depth and ego-motion from video. In CVPR, 2017. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.39, + 0.892, + 0.432 + ], + "angle": 0, + "content": "[59] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images. 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.433, + 0.892, + 0.488 + ], + "angle": 0, + "content": "[60] Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hujun Bao, Zhaopeng Cui, Martin R. Oswald, and Marc Pollefeys. Nice-slam: Neural implicit scalable encoding for slam. In CVPR, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.488 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "4169" + } + ] +] \ No newline at end of file diff --git a/2023/NoPe-NeRF_ Optimising Neural Radiance Field With No Pose Prior/2260561b-6c24-4454-9ee0-0e88425e374a_origin.pdf b/2023/NoPe-NeRF_ Optimising Neural Radiance Field With No Pose Prior/2260561b-6c24-4454-9ee0-0e88425e374a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..bebec0d715739ce24f1b030fef0482a50fddab89 --- /dev/null +++ b/2023/NoPe-NeRF_ Optimising Neural Radiance Field With No Pose Prior/2260561b-6c24-4454-9ee0-0e88425e374a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ed20d295e0b583d0bac7ed37ce0efcc85266cdeea42bc52af29f6ff7493d8bd3 +size 10277067 diff --git a/2023/NoPe-NeRF_ Optimising Neural Radiance Field With No Pose Prior/full.md b/2023/NoPe-NeRF_ Optimising Neural Radiance Field With No Pose Prior/full.md new file mode 100644 index 0000000000000000000000000000000000000000..2628a746aaf6445ee785fed90cf5b7ee298b7f48 --- /dev/null +++ b/2023/NoPe-NeRF_ Optimising Neural Radiance Field With No Pose Prior/full.md @@ -0,0 +1,369 @@ +# NoPe-NeRF: Optimising Neural Radiance Field with No Pose Prior + +Wenjing Bian + +Zirui Wang + +Kejie Li + +Jia-Wang Bian + +Victor Adrian Prisacariu + +Active Vision Lab, University of Oxford + +{wenjing, ryan, kejie, jiawang, victor}@robots.ox.ac.uk + +![](images/6456383c90eff0c41d60483ff6ee3639f9c2a5a312a21425f09678ffb0312283.jpg) + +![](images/cc1a091e00edbf3086064344697f8b788c846e7fa1b750e06194dde68cef4400.jpg) +Figure 1. Novel view synthesis comparison. We propose NoPe-NeRF for joint pose estimation and novel view synthesis. Our method enables more robust pose estimation and renders better novel view synthesis than previous state-of-the-art methods. + +![](images/7726bb48c80d2abc808b1dbcafa6d8a0a37b77fd79638ab3242ff39329c4c3a4.jpg) + +![](images/7ce87e153f7b0d9c78f8a1d851b3295017619fdf465ce0762f817e85be6a1128.jpg) + +![](images/0947429dff822ae0042cec68959abcc2791f07b733fa4ff7c712da433d518cde.jpg) + +![](images/6c2fe2396fdbe9b3d56187ccb624b9fdfad4f3be627a7be220df5c061ccf6c37.jpg) + +![](images/829f40ee63e1d95e7d7a6a16af554c09194ea234ce0181c9338caf0da21fac81.jpg) + +![](images/560597951ee9f257e9d09e683c2aa220680a9af6f1788756a94fc6199ae59ce9.jpg) + +# Abstract + +Training a Neural Radiance Field (NeRF) without precomputed camera poses is challenging. Recent advances in this direction demonstrate the possibility of jointly optimising a NeRF and camera poses in forward-facing scenes. However, these methods still face difficulties during dramatic camera movement. We tackle this challenging problem by incorporating undistorted monocular depth priors. These priors are generated by correcting scale and shift parameters during training, with which we are then able to constrain the relative poses between consecutive frames. This constraint is achieved using our proposed novel loss functions. Experiments on real-world indoor and outdoor scenes show that our method can handle challenging camera trajectories and outperforms existing methods in terms of novel view rendering quality and pose estimation accuracy. Our project page is https://nope-nerf.active.vision. + +# 1. Introduction + +The photo-realistic reconstruction of a scene from a stream of RGB images requires both accurate 3D geometry + +reconstruction and view-dependent appearance modelling. Recently, Neural Radiance Fields (NeRF) [24] have demonstrated the ability to build high-quality results for generating photo-realistic images from novel viewpoints given a sparse set of images. + +An important preparation step for NeRF training is the estimation of camera parameters for the input images. A current go-to option is the popular Structure-from-Motion (SfM) library COLMAP [35]. Whilst easy to use, this preprocessing step could be an obstacle to NeRF research and real-world deployments in the long term due to its long processing time and its lack of differentiability. Recent works such as NeRFmm [46], BARF [18] and SC-NeRF [12] propose to simultaneously optimise camera poses and the neural implicit representation to address these issues. Nevertheless, these methods can only handle forward-facing scenes when no initial parameters are supplied, and fail in dramatic camera motions, e.g.a casual handheld captured video. + +This limitation has two key causes. First, all these methods estimate a camera pose for each input image individually without considering relative poses between images. Looking back to the literature of Simultaneous localisation and mapping (SLAM) and visual odometry, pose estimation can significantly benefit from estimating relative poses be + +tween adjacent input frames. Second, the radiance field is known to suffer from shape-radiance ambiguity [55]. Estimating camera parameters jointly with NeRF adds another degree of ambiguity, resulting in slow convergence and unstable optimisation. + +To handle the limitation of large camera motion, we seek help from monocular depth estimation [22, 28, 29, 51]. Our motivation is threefold: First, monocular depth provides strong geometry cues that are beneficial to constraint shape-raliance ambiguity. Second, relative poses between adjacent depth maps can be easily injected into the training pipeline via Chamfer Distance. Third, monocular depth is lightweight to run and does not require camera parameters as input, in contrast to multi-view stereo depth estimation. For simplicity, we use the term mono-depth from now on. + +Utilising mono-depth effectively is not straightforward with the presence of scale and shift distortions. In other words, mono-depth maps are not multi-view consistent. Previous works [9, 17, 47] simply take mono-depth into a depth-wise loss along with NeRF training. Instead, we propose a novel and effective way to thoroughly integrate mono-depth into our system. First, we explicitly optimise scale and shift parameters for each mono-depth map during NeRF training by penalising the difference between rendered depth and mono-depth. Since NeRF by itself is trained based on multiview consistency, this step transforms mono-depth maps to undistorted multiview consistent depth maps. We further leverage these multiview consistent depth maps in two loss terms: a) a Chamfer Distance loss between two depth maps of adjacent images, which injects relative pose to our system; and b) a depth-based surface rendering loss, which further improves relative pose estimation. + +In summary, we propose a method to jointly optimise camera poses and a NeRF from a sequence of images with large camera motion. Our system is enabled by three contributions. First, we propose a novel way to integrate monodepth into unposed-NeRF training by explicitly modelling scale and shift distortions. Second, we supply relative poses to the camera-NeRF joint optimisation via an inter-frame loss using undistorted mono-depth maps. Third, we further regularise our relative pose estimation with a depth-based surface rendering loss. + +As a result, our method is able to handle large camera motion, and outperforms state-of-the-art methods by a significant margin in terms of novel view synthesis quality and camera trajectory accuracy. + +# 2. Related Work + +Novel View Synthesis. While early Novel View Synthesis (NVS) approaches applied interpolations between pixels [3], later works often rendered images from 3D reconstruction [1, 6]. In recent years, different representations of the 3D scene are used, e.g.meshes [30, 31], Multi-Plane + +Images [41, 59], layered depth [42] etc. Among them, NeRF [24] has become a popular scene representation for its photorealistic rendering. + +A number of techniques are proposed to improve NeRF's performance with additional regularisation [13, 26, 54], depth priors [7, 32, 47, 53], surface enhancements [27, 43, 49] or latent codes [40, 44, 52]. Other works [2, 10, 25, 34] have also accelerated NeRF training and rendering. However, most of these approaches require pre-computed camera parameters obtained from SfM algorithms [11, 35]. + +NeRF With Pose Optimisation. Removing camera parameter preprocessing is an active line of research recently. One category of the methods [33, 38, 60] use a SLAM-style pipeline, that either requires RGB-D inputs or relies on accurate camera poses generated from the SLAM tracking system. Another category of works optimises camera poses with the NeRF model directly. We term this type of method as unposed-NeRF in this paper. iNeRF [50] shows that poses for novel view images can be estimated using a reconstructed NeRF model. GNeRF [21] combines Generative Adversarial Networks with NeRF to estimate camera poses but requires a known sampling distribution for poses. More relevant to our work, NeRFmm [46] jointly optimises both camera intrinsic and extrinsic alongside NeRF training. BARF [18] proposes a coarse-to-fine positional encoding strategy for camera poses and NeRF joint optimisation. SCNeRF [12] further parameterises camera distortion and employs a geometric loss to regularise rays. GARF [4] shows that using Gaussian-MLPs makes joint pose and scene optimisation easier and more accurate. Recently, SiNeRF [48] uses SIREN [36] layers and a novel sampling strategy to alleviate the sub-optimality of joint optimisation in NeRFmm. Although showing promising results on the forward-facing dataset like LLFF [23], these approaches face difficulties when handling challenging camera trajectories with large camera motion. We address this issue by closely integrating mono-depth maps with the joint optimisation of camera parameters and NeRF. + +# 3. Method + +We tackle the challenge of handling large camera motions in unposed-NeRF training. Given a sequence of images, camera intrinsics, and their mono-depth estimations, our method recovers camera poses and optimises a NeRF simultaneously. We assume camera intrinsics are available in the image meta block, and we run an off-the-shelf monodepth network DPT [7] to acquire mono-depth estimations. Without repeating the benefit of mono-depth, we unroll this section around the effective integration of monocular depth into unposed-NeRF training. + +The training is a joint optimisation of the NeRF, camera poses, and distortion parameters of each mono-depth + +![](images/1d697123e9c97e03383d67eb3111e6dc00d5feeec2e1903c66e745ca1242a200.jpg) +Figure 2. Method Overview. Our method takes a sequence of images as input to reconstruct NeRF and jointly estimates the camera poses of the frames. We first generate monocular depth maps from a mono-depth estimation network and reconstruct the point clouds. We then optimise NeRF, camera poses, and depth distortion parameters jointly with inter-frame and NeRF losses. + +![](images/429c0ded872daf15aad5f8b8de9aab2f06cbcc399ff69748ba8379356f538e90.jpg) + +![](images/b907a9e723a3afab4a8fd3a5ab307f4a6530b07b83d50befff349e18407a1dbf.jpg) + +map. The distortion parameters are supervised by minimising the discrepancies between the mono-depth maps and depth maps rendered from the NeRF, which are multiview consistent. The undistorted depth maps in return effectively mediate the shape-radiance ambiguity, which eases the training of NeRF and camera poses. + +Specifically, the undistorted depth maps enable two constraints. We constrain global pose estimation by supplying relative pose between adjacent images. This is achieved via a Chamfer-Distance-based correspondence between two point clouds, back-projected from undistorted depth maps. Further, we regularise relative pose estimation with a surface-based photometric consistency where we treat undistorted depth as surface. + +We detail our method in the following sections, starting from NeRF in Sec. 3.1 and unposed-NeRF training in Sec. 3.2, looking into mono-depth distortions in Sec. 3.3, followed by our mono-depth enabled loss terms in Sec. 3.4, and finishing with an overall training pipeline Sec. 3.5. + +# 3.1. NeRF + +Neural Radiance Field (NeRF) [24] represents a scene as a mapping function $F_{\Theta}:(\mathbf{x},\mathbf{d})\to (\mathbf{c},\sigma)$ that maps a 3D location $\mathbf{x}\in \mathbb{R}^3$ and a viewing direction $\mathbf{d}\in \mathbb{R}^3$ to a radiance colour $\mathbf{c}\in \mathbb{R}^3$ and a volume density value $\sigma$ . This mapping is usually implemented with a neural network parameterised by $F_{\Theta}$ . Given $N$ images $\mathcal{I} = \{I_i\mid i = 0\dots N - 1\}$ with their camera poses $\Pi = \{\pi_i\mid i = 0\dots N - 1\}$ , NeRF can be optimised by minimising photometric error + +$\mathcal{L}_{rgb} = \sum_{i}^{N}\| I_{i} - \hat{I}_{i}\|_{2}^{2}$ between synthesised images $\hat{\mathcal{I}}$ and captured images $\mathcal{I}$ : + +$$ +\Theta^ {*} = \arg \min _ {\Theta} \mathcal {L} _ {r g b} (\hat {\mathcal {I}} \mid \mathcal {I}, \Pi), \tag {1} +$$ + +where $\hat{I}_i$ is rendered by aggregating radiance colour on camera rays $\mathbf{r}(h) = \mathbf{o} + h\mathbf{d}$ between near and far bound $h_n$ and $h_f$ . More concretely, we synthesise $\hat{I}_i$ with a volumetric rendering function + +$$ +\hat {I} _ {i} (\mathbf {r}) = \int_ {h _ {n}} ^ {h _ {f}} T (h) \sigma (\mathbf {r} (h)) \mathbf {c} (\mathbf {r} (h), \mathbf {d}) d h, \tag {2} +$$ + +where $T(h) = \exp(-\int_{h_n}^h \sigma(\mathbf{r}(s)) ds)$ is the accumulated transmittance along a ray. We refer to [24] for further details. + +# 3.2. Joint Optimisation of Poses and NeRF + +Prior works [12, 18, 46] show that it is possible to estimate both camera parameters and a NeRF at the same time by minimising the above photometric error $\mathcal{L}_{rgb}$ under the same volumetric rendering process in Eq. (2). + +The key lies in conditioning camera ray casting on variable camera parameters $\Pi$ , as the camera ray $\mathbf{r}$ is a function of camera pose. Mathematically, this joint optimisation can be formulated as: + +$$ +\Theta^ {*}, \Pi^ {*} = \arg \min _ {\Theta , \Pi} \mathcal {L} _ {r g b} (\hat {\mathcal {I}}, \hat {\Pi} \mid \mathcal {I}), \tag {3} +$$ + +where $\hat{\Pi}$ denotes camera parameters that are updated during optimising. Note that the only difference between Eq. (1) + +and Eq. (3) is that Eq. (3) considers camera parameters as variables. + +In general, the camera parameters $\Pi$ include camera intrinsics, poses, and lens distortions. We only consider estimating camera poses in this work, e.g., camera pose for frame $I_{i}$ is a transformation $\mathbf{T}_i = [\mathbf{R}_i\mid \mathbf{t}_i]$ with a rotation $\mathbf{R}_i\in \mathrm{SO}(3)$ and a translation $\mathbf{t}_i\in \mathbb{R}^3$ + +# 3.3. Undistortion of Monocular Depth + +With an off-the-shelf monocular depth network, e.g., DPT [28], we generate mono-depth sequence $\mathcal{D} = \{D_i\mid i = 0\dots N - 1\}$ from input images. Without surprise, mono-depth maps are not multi-view consistent so we aim to recover a sequence of multi-view consistent depth maps, which are further leveraged in our relative pose loss terms. + +Specifically, we consider two linear transformation parameters for each mono-depth map, resulting in a sequence of transformation parameters for all frames $\Psi = \{(\alpha_i,\beta_i)\mid i = 0\ldots N - 1\}$ , where $\alpha_{i}$ and $\beta_{i}$ denote a scale and a shift factor. With multi-view consistent constraint from NeRF, we aim to recover a multi-view consistent depth map $D_{i}^{*}$ for $D_{i}$ : + +$$ +D _ {i} ^ {*} = \alpha_ {i} D _ {i} + \beta_ {i}, \tag {4} +$$ + +by joint optimising $\alpha_{i}$ and $\beta_{i}$ along with a NeRF. This joint optimisation is mostly achieved by enforcing the consistency between an undistorted depth map $D_{i}^{*}$ and a NeRF rendered depth map $\hat{D}_i$ via a depth loss: + +$$ +\mathcal {L} _ {\text {d e p t h}} = \sum_ {i} ^ {N} \left\| D _ {i} ^ {*} - \hat {D} _ {i} \right\|, \tag {5} +$$ + +where + +$$ +\hat {D} _ {i} (\mathbf {r}) = \int_ {h _ {n}} ^ {h _ {f}} T (h) \sigma (\mathbf {r} (h)) d h \tag {6} +$$ + +denotes a volumetric rendered depth map from NeRF. + +It is important to note that both NeRF and mono-depth benefit from Eq. (5). On the one hand, mono-depth provides strong geometry prior for NeRF training, reducing shape-radiance ambiguity. On the other hand, NeRF provides multi-view consistency so we can recover a set of multiview consistent depth maps for relative pose estimations. + +# 3.4. Relative Pose Constraint + +Aforementioned unposed-NeRF methods [12, 18, 46] optimise each camera pose independently, resulting in an overfit to target images with incorrect poses. Penalising incorrect relative poses between frames can help to regularise the joint optimisation towards smooth convergence, especially in a complex camera trajectory. Therefore, we propose two losses that constrain relative poses. + +Point Cloud Loss. We back-project the undistorted depth maps $\mathcal{D}^*$ using the known camera intrinsics, to point clouds $\mathcal{P}^* = \{P_i^* \mid i = 0 \dots N - 1\}$ and optimise the relative pose between consecutive point clouds by minimising a point cloud loss $\mathcal{L}_{pc}$ : + +$$ +\mathcal {L} _ {p c} = \sum_ {(i, j)} l _ {c d} \left(P _ {j} ^ {*}, \mathbf {T} _ {j i} P _ {i} ^ {*}\right), \tag {7} +$$ + +where $\mathbf{T}_{ji} = \mathbf{T}_j\mathbf{T}_i^{-1}$ represents the related pose that transforms point cloud $P_{i}^{*}$ to $P_{j}^{*}$ , tuple $(i,j)$ denotes indices of a consecutive pair of instances, and $l_{cd}$ denotes Chamfer Distance: + +$$ +l _ {c d} \left(P _ {i}, P _ {j}\right) = \sum_ {p _ {i} \in P _ {i}} \min _ {p _ {j} \in P _ {j}} \| p _ {i} - p _ {j} \| _ {2} + \sum_ {p _ {j} \in P _ {j}} \min _ {p _ {i} \in P _ {i}} \| p _ {i} - p _ {j} \| _ {2}. \tag {8} +$$ + +Surface-based Photometric Loss. While the point cloud loss $\mathcal{L}_{pc}$ offers supervision in terms of 3D-3D matching, we observe that a surface-based photometric error can alleviate incorrect matching. With the photometric consistency assumption, this photometric error penalises the differences in appearance between associated pixels. The association is established by projecting the point cloud $P_{i}^{*}$ onto images $I_{i}$ and $I_{j}$ . + +The surface-based photometric loss can then be defined as: + +$$ +\mathcal {L} _ {r g b - s} = \sum_ {(i, j)} \| I _ {i} \langle \mathbf {K} _ {i} P _ {i} ^ {*} \rangle - I _ {j} \langle \mathbf {K} _ {j} \mathbf {T} _ {j} \mathbf {T} _ {i} ^ {- 1} P _ {i} ^ {*} \rangle \|, \tag {9} +$$ + +where $\langle \cdot \rangle$ represents the sampling operation on the image and $\mathbf{K}_i$ denotes a projection matrix for $i_{th}$ camera. + +# 3.5. Overall Training Pipeline + +Assembling all loss terms, we get the overall loss function: + +$$ +\mathcal {L} = \mathcal {L} _ {r g b} + \lambda_ {1} \mathcal {L} _ {\text {d e p t h}} + \lambda_ {2} \mathcal {L} _ {p c} + \lambda_ {3} \mathcal {L} _ {r g b - s}, \tag {10} +$$ + +where $\lambda_1, \lambda_2, \lambda_3$ are the weighting factors for respective loss terms. By minimising the combined of loss $\mathcal{L}$ : + +$$ +\Theta^ {*}, \Pi^ {*}, \Psi^ {*} = \arg \min _ {\Theta , \Pi , \Psi} \mathcal {L} (\hat {\mathcal {I}}, \hat {\mathcal {D}}, \hat {\Pi}, \hat {\Psi} \mid \mathcal {I}, \mathcal {D}), \tag {11} +$$ + +our method returns the optimised NeRF parameters $\Theta$ , camera poses $\Pi$ , and distortion parameters $\Psi$ . + +# 4. Experiments + +We begin with a description of our experimental setup in Sec. 4.1. In Sec. 4.2, we compare our method with pose-unknown methods. Next, we compare our method with the COLMAP-assisted NeRF baseline in Sec. 4.3. Lastly, we conduct ablation studies in Sec. 4.4. + +![](images/bf97d274716646e4a0f70d55a3c67c74a0405cd3380efc61fc648966d5f7f921.jpg) +(a) BARF +(b) NeRFmm +(c) SC-NeRF +(d) Ours +(e) Ground Truth +Figure 3. Qualitative results of novel view synthesis and depth prediction on Tanks and Temples. We visualise the synthesised images and the rendered depth maps (top left of each image) for all methods. NoPe-NeRF is able to recover details for both colour and geometry. + +# 4.1. Experimental Setup + +Datasets. We conduct experiments on two datasets Tanks and Temples [15] and ScanNet [5]. Tanks and Temples: we use 8 scenes to evaluate pose accuracy and novel view synthesis quality. We chose scenes captured at both indoor and outdoor locations, with different frame sampling rates and lengths. All images are down-sampled to a resolution of $960 \times 540$ . For the family scene, we sample 200 images and take 100 frames with odd frame ids as training images and the remaining 100 frames for novel view synthesis, in order to analyse the performance under smooth motion. For the remaining scenes, following NeRF [24], 1/8 of the images in each sequence are held out for novel view synthesis, unless otherwise specified. ScanNet: we select 4 scenes for evaluating pose accuracy, depth accuracy, and novel view synthesis quality. For each scene, we take 80-100 consecutive images and use 1/8 of these images for novel view synthesis. For evaluation, we employ depth maps and poses provided by ScanNet as ground truth. ScanNet images are down-sampled to $648 \times 484$ . We crop images with dark orders during preprocessing. + +Metrics. We evaluate our proposed method in three aspects. For novel view synthesis, we follow previous methods [12,18,46], and use standard evaluation metrics, including Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM) [45] and Learned Perceptual Image Patch Similarity (LPIPS) [56]. For pose evaluation, We use standard visual odometry metrics [16, 37, 57], including the Absolute Trajectory Error (ATE) and Relative Pose Error (RPE). ATE measures the difference between the estimated camera positions and the ground truth positions. RPE measures the relative pose errors between pairs of images, + +which consists of relative rotation error $(\mathrm{RPE}_r)$ and relative translation error $(\mathrm{RPE}_t)$ . The estimated trajectory is aligned with the ground truth using Sim(3) with 7 degrees of freedom. We use standard depth metrics [8,19,20,39] (Abs Rel, Sq Rel, RMSE, RMSE log, $\delta_1$ , $\delta_2$ and $\delta_3$ ) for depth evaluation. For further detail, please refer to the supplementary material. To recover the metric scale, we follow Zhou et al. [58] and match the median value between rendered and ground truth depth maps. + +Implementation Details. Our model architecture is based on NeRF [24] with a few modifications: a) replacing ReLU activation function with Softplus and b) sampling 128 points along each ray uniformly with noise, between a predefined range (0.1, 10). We use 2 separate Adam optimisers [14] for NeRF and other parameters. The initial learning rate for NeRF is 0.001 and for the pose and distortion is 0.0005. Camera rotations are optimised in axis-angle representation $\phi_i \in \mathfrak{so}(3)$ . We first train the model with all losses with constant learning rates until the inter-frame losses converge. Then, we remove the inter-frame losses and depth loss to refine the model with the RGB loss only. We decay the learning rates with different schedulers to refine for 10,000 epochs. We balance the loss terms with $\lambda_1 = 0.04$ , $\lambda_2 = 1.0$ and $\lambda_3 = 1.0$ . For each training step, we randomly sample 1024 pixels (rays) from each input image and 128 samples per ray. More details are provided in the supplementary material. + +# 4.2. Comparing With Pose-Unknown Methods + +We compare our method with pose-unknown baselines, including BARF [18], NeRFmm [46] and SC-NeRF [12]. + +View Synthesis Quality. To obtain the camera poses of + +![](images/8a961b15ff878b2b3cffb7e326c300c3b156cd89ca5cdda08b5d72724dfc57d8.jpg) +Figure 4. Pose Estimation Comparison. We visualise the trajectory (3D plot) and relative rotation errors $\mathrm{RPE}_r$ (bottom colour bar) of each method on Ballroom and Museum. The colour bar on the right shows the relative scaling of colour. More results are in the supplementary. + +
scenesOursBARFNeRFmmSC-NeRF
PSNR ↑SSIM ↑LPIPS ↓PSNRSSIMLPIPSPSNRSSIMLPIPSPSNRSSIMLPIPS
ScanNet0079_0032.470.840.4132.310.830.4330.590.810.4931.330.820.46
0418_0031.330.790.3431.240.790.3530.000.770.4029.050.750.43
0301_0029.830.770.3629.310.760.3827.840.720.4529.450.770.35
0431_0033.830.910.3932.770.900.4131.440.880.4532.570.900.40
mean31.860.830.3831.410.820.3929.970.800.4530.600.810.41
Tanks and TemplesChurch25.170.730.3923.170.620.5221.640.580.5421.960.600.53
Barn26.350.690.4425.280.640.4823.210.610.5323.260.620.51
Museum26.770.760.3523.580.610.5522.370.610.5324.940.690.45
Family26.010.740.4123.040.610.5623.040.580.5622.600.630.51
Horse27.640.840.2624.090.720.4123.120.700.4325.230.760.37
Ballroom25.330.720.3820.660.500.6020.030.480.5722.640.610.48
Francis29.480.800.3825.850.690.5725.4000.690.5226.460.730.49
Ignatius23.960.610.4721.780.470.6021.160.450.6023.000.550.53
mean26.340.740.3923.420.610.5422.500.590.5423.760.650.48
+ +Table 1. Novel view synthesis results on ScanNet and Tanks and Temples. Each baseline method is trained with its public code under the original settings and evaluated with the same evaluation protocol. + +test views for rendering, we minimise the photometric error of the synthesised images while keeping the NeRF model fixed, as in NeRFmm [46]. Each test pose is initialised with the learned pose of the training frame that is closest to it. We use the same pre-processing for all baseline approaches, which results in higher accuracy than their original implementations. More details are provided in the supplementary material. Our method outperforms all the baselines by a large margin. The quantitative results are summarised in Tab. 1, and qualitative results are shown in Fig. 3. + +We recognised that because the test views, which are sampled from videos, are close to the training views, good results may be obtained due to overfitting to the training images. Therefore, we conduct an additional qualitative + +evaluation on more novel views. Specifically, we fit aBezier curve from the estimated training poses and sample interpolated poses for each method to render novel view videos. Sampled results are shown in Fig. 5, and the rendered videos are in the supplementary material. These results show that our method renders photo-realistic images consistently, while other methods generate visible artifacts. + +Camera Pose. Our method significantly outperforms other baselines in all metrics. The quantitative pose evaluation results are shown in Tab. 2. For ScanNet, we use the camera poses provided by the dataset as ground truth. For Tanks and Temples, not every video comes with ground truth poses, so we use COLMAP estimations for reference. Our estimated trajectory is better aligned with the ground + +
scenesOursBARFNeRFmmSC-NeRF
RPEt ↓RPEr ↓ATE↓RPEtRPErATERPEtRPErATERPEtRPErATE
ScanNet0079_000.7520.2040.0231.1100.4800.0621.7060.6360.1002.0640.6640.115
0418_000.4550.1190.0151.3980.5380.0201.4020.4600.0131.5280.5020.016
0301_000.3990.1230.0131.3160.7770.2193.0970.8940.2881.1330.4220.056
0431_001.6250.2740.0696.0240.7540.1686.7990.6240.4964.1100.4990.205
mean0.8080.1800.0302.4620.6370.1173.2510.6540.2242.2090.5220.098
Tanks and TemplesChurch0.0340.0080.0080.1140.0380.0520.6260.1270.0650.8360.1870.108
Barn0.0460.0320.0040.3140.2650.0501.6290.4940.1591.3170.4290.157
Museum0.2070.2020.0203.4421.1280.2634.1341.0510.3468.3391.4910.316
Family0.0470.0150.0011.3710.5910.1152.7430.5370.1201.1710.4990.142
Horse0.1790.0170.0031.3330.3940.0141.3490.4340.0181.3660.4380.019
Ballroom0.0410.0180.0020.5310.2280.0180.4490.1770.0310.3280.1460.012
Francis0.0570.0090.0051.3210.5580.0821.6470.6180.2071.2330.4830.192
Ignatius0.0260.0050.0020.7360.3240.0291.3020.3790.0410.5330.2400.085
mean0.0800.0380.0061.0460.4410.0781.7350.4770.1231.8900.4890.129
+ +Table 2. Pose accuracy on ScanNet and Tanks and Temples. Note that we use COLMAP poses in Tanks and Temples as the "ground truth". The unit of $\mathrm{RPE}_r$ is in degrees, ATE is in the ground truth scale and $\mathrm{RPE}_t$ is scaled by 100. + +
Abs Rel ↓Sq Rel ↓RMSE ↓RMSE log ↓δ1 ↑δ2 ↑δ3 ↑
Ours0.1410.1370.5680.1760.8280.9700.987
BARF0.3760.6840.9900.4010.4900.7510.884
NeRFmm0.5901.7211.6720.5870.3160.5600.743
SC-NeRF0.4170.6421.0790.4760.3620.6580.832
DPT0.1970.2460.7510.2260.7470.9340.975
+ +Table 3. Depth map evaluation on ScanNet. Our depth estimation is more accurate than baseline models BARF [18], NeRFmm [46] and SC-NeRF [12]. Compared with DPT [58], we show our depth is more accurate after undistortion. + +truth than other methods, and our estimated rotation is two orders of magnitudes more accurate than others. We visualise the camera trajectories and rotations in Fig. 4. + +Depth. We evaluate the accuracy of the rendered depth maps on ScanNet, which provides the ground-truth depths for evaluation. Our rendered depth maps achieve superior accuracy over the previous alternatives. We also compare with the mono-depth maps estimated by DPT. Our rendered depth maps, after undistortion using multiview consistency in the NeRF optimisation, outperform DPT by a large margin. The results are summarised in Tab. 3, and sampled qualitative results are illustrated in Fig. 3. + +# 4.3. Comparing With COLMAP Assisted NeRF + +We make a comparison of pose estimation accuracy between our method and COLMAP against ground truth poses in ScanNet. We achieve on-par accuracy with COLMAP, as shown in Tab. 4. We further analyse the novel view synthesis quality of the NeRF model trained with our learned poses to COLMAP poses on ScanNet and Tanks and Temples. The original NeRF training contains two stages, finding poses using COLMAP and optimising the scene representation. In order to make our comparison fairer, in this section only, we mimic a similar two-stage training as the original NeRF [24]. In the first stage, we train our method with all losses for camera pose estimation, i.e., mimicking the COLMAP processing. Then, we fix the optimised poses + +![](images/3cbccf2b854b65f90026402f87da8bcfadee477424ca19c84ddcaefc04ba8a80.jpg) +Figure 5. Sampled frames from rendered novel view videos. For each method, we fit the learned trajectory with aBezier curve and uniformly sample new viewpoints for rendering. Our method generates significantly better results than previous methods, which show visible artifacts. The full rendered videos and details about generating novel views are provided in the supplementary. + +and train a NeRF model from scratch, using the same settings and loss as the original NeRF. This evaluation enables us to compare our estimated poses to the COLMAP poses indirectly, i.e., in terms of contribution to view synthesis. + +Our two-stage method outperforms the COLMAP-assisted NeRF baseline, which indicates a better pose estimation for novel view synthesis. The results are summarised in Tab. 5. + +As is commonly known, COLMAP performs poorly in low-texture scenes and sometimes fails to find accurate camera poses. Fig. 6 shows an example of a low-texture scene where COLMAP provides inaccurate pose estimation that causes NeRF to render images with visible artifacts. In contrast, our method renders high-quality images, thanks to robust optimisation of camera pose. + +
scenesOursCOLMAP
RPEt↓RPEr↓ATE↓RPEtRPErATE
0079_000.7520.2040.0230.6550.2210.012
0418_000.4550.1190.0150.4910.1240.016
0301_000.3990.1230.0130.4140.1360.009
0431_001.6250.2740.0691.2920.2490.051
mean0.8080.1800.0300.7130.1820.022
+ +![](images/ab5ff0ea66cc70c7dccbe02caef55a0d10dcf96e0f561a4273fb78d68fbec191.jpg) +(a) COLMAP+NeRF + +![](images/3351115fc0353aa326c5261a0a09a87858d4cf14f3c8461db588f41d0fb2a4da.jpg) +(b) Ours + +![](images/c9121ae35a3206e6ce21c78be3065f214a6c836303976974215795cbce33c325.jpg) +(c) Ground Truth +Figure 6. COLMAP failure case. On a rotation-dominant sequence with low-texture areas, COLMAP fails to estimate correct poses, which results in artifacts in synthesised images. + +Table 4. Comparison of pose accuracy with COLMAP on ScanNet. + +
scenesOursOurs-rCOLMAP+NeRF
PSNR ↑SSIM ↑LPIPS ↓PSNRSSIMLPIPSPSNRSSIMLPIPS
ScanNet0079_0032.470.840.4133.120.850.4031.980.830.43
0418_0031.330.790.3430.490.770.4030.600.780.40
0301_0029.830.770.3630.050.780.3430.010.780.36
0431_0033.830.910.3933.860.910.3933.540.910.39
mean31.860.830.3831.880.830.3831.530.820.40
Tanks and TemplesChurch25.170.730.3926.740.780.3225.720.750.37
Barn26.350.690.4426.580.710.4226.720.710.42
Museum26.770.760.3526.980.770.3627.210.780.34
Family26.010.740.4126.210.750.4026.610.770.39
Horse27.640.840.2628.060.840.2627.020.820.29
Ballroom25.330.720.3825.530.730.3825.470.730.38
Francis29.480.800.3829.730.810.3830.050.810.38
Ignatius23.960.610.4723.980.620.4624.080.610.47
mean26.340.740.3926.730.750.3726.610.750.38
+ +Interestingly, this experiment also reveals that the two-stage method shows higher accuracy than the one-stage method. We hypothesise that the joint optimisation (from randomly initialised poses) in the one-stage approach causes the NeRF optimisation to be trapped in a local minimum, potentially due to the bad pose initialisation. The two-stage approach circumvents this issue by re-initialising the NeRF and re-training with well-optimised poses, resulting in higher performance. + +# 4.4. Ablation Study + +In this section, we analyse the effectiveness of the parameters and components that have been added to our model. The results of ablation studies are shown in Tab. 6. + +Effect of Distortion Parameters. We find that ignoring depth distortions (i.e., setting scales to 1 and shifts to 0 as constants) leads to a degradation in pose accuracy, as in + +Table 5. Comparison to NeRF with COLMAP poses. Our two-stage method (Ours-r) outperforms both COLMAP+NeRF and our one-stage method (Ours). + +
NVSPose
PSNR↑SSIM↑LPIPS↓RPEt↓RPEr↓ATE↓
Ours31.860.830.380.8010.1810.031
Ours w/o α,β31.460.820.391.9290.3210.066
Ours w/o Lpc31.730.820.382.2270.4530.101
Ours w/o Lrgb-s31.050.810.411.8140.4010.156
Ours w/o Ldepth31.200.810.401.4980.3830.089
+ +Table 6. Ablation study results on ScanNet. + +consistent distortions of depth maps introduce errors to the estimation of relative poses and confuse NeRF for geometry reconstruction. + +Effect of Inter-frame Losses. We observe that the interframe losses are the major contributor to improving relative poses. When removing the pairwise point cloud loss $L_{pc}$ or the surface-based photometric loss $L_{rgb-s}$ , there is less constraint between frames, and thus the pose accuracy becomes lower. + +Effect of NeRF Losses. When the depth loss $L_{depth}$ is removed, the distortions of input depth maps are only optimised locally through the inter-frame losses. We find that this can lead to drift and degradation in pose accuracy. + +# 4.5. Limitations + +Our proposed method optimises camera pose and the NeRF model jointly and works on challenging scenes where other baselines fail. However, the optimisation of the model is also affected by non-linear distortions and the accuracy of the mono-depth estimation, which we did not consider. + +# 5. Conclusion + +In this work, we present NoPe-NeRF, an end-to-end differentiable model for joint camera pose estimation and novel view synthesis from a sequence of images. We demonstrate that previous approaches have difficulty with complex trajectories. To tackle this challenge, we use mono-depth maps to constrain the relative poses between frames and regularise the geometry of NeRF, which leads to better pose estimation. We show the effectiveness and robustness of NoPe-NeRF on challenging scenes. The improved pose estimation leads to better novel view synthesis quality and geometry reconstruction compared with other approaches. We believe our method is an important step towards applying the unknown-pose NeRF models to large-scale scenes in the future. + +# Acknowledgements + +We thank Theo Costain, Michael Hobley, Shuai Chen and Xinghui Li for their helpful proofreading and discussions. Wenjing Bian is supported by the China Scholarship Council - University of Oxford Scholarship. + +# References + +[1] Chris Buehler, Michael Bosse, Leonard McMillan, Steven Gortler, and Michael Cohen. Unstructured lumigraph rendering. In SIGGRAPH, 2001. 2 +[2] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In ECCV, 2022. 2 +[3] Shenchang Eric Chen and Lance Williams. View interpolation for image synthesis. In SIGGRAPH, 1993. 2 +[4] Shin-Fang Chng, Sameera Ramasinghe, Jamie Sherrah, and Simon Lucey. Garf: Gaussian activated radiance fields for high fidelity reconstruction and pose estimation. arXiv eprints, 2022. 2 +[5] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, 2017. 5 +[6] Paul E Debevec, Camillo J Taylor, and Jitendra Malik. Modeling and rendering architecture from photographs: A hybrid geometry-and image-based approach. In SIGGRAPH, 1996. 2 +[7] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised nerf: Fewer views and faster training for free. In CVPR, 2022. 2 +[8] Huan Fu, Mingming Gong, Chaohui Wang, Kayhan Bat-manghelich, and Dacheng Tao. Deep ordinal regression network for monocular depth estimation. In CVPR, 2018. 5 +[9] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video. In ICCV, 2021. 2 +[10] Stephan J Garbin, Marek Kowalski, Matthew Johnson, Jamie Shotton, and Julien Valentin. Fastnerf: High-fidelity neural rendering at 200fps. In ICCV, 2021. 2 +[11] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. 2003. 2 +[12] Yoonwoo Jeong, Seokjun Ahn, Christopher Choy, Anima Anandkumar, Minsu Cho, and Jaesik Park. Self-calibrating neural radiance fields. In ICCV, 2021. 1, 2, 3, 4, 5, 7 +[13] Mijeong Kim, Seonguk Seo, and Bohyung Han. Infonerf: Ray entropy minimization for few-shot neural volume rendering. In CVPR, 2022. 2 +[14] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In *Yoshua Bengio and Yann LeCun*, editors, *ICLR*, 2015. 5 +[15] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics, 2017. 5 +[16] Johannes Kopf, Xuejian Rong, and Jia-Bin Huang. Robust consistent video depth estimation. In CVPR, 2021. 5 +[17] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In CVPR, 2021. 2 +[18] Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. Barf: Bundle-adjusting neural radiance fields. In ICCV, 2021. 1, 2, 3, 4, 5, 7 + +[19] Fayao Liu, Chunhua Shen, Guosheng Lin, and Ian Reid. Learning depth from single monocular images using deep convolutional neural fields. TPAMI, 2015. 5 +[20] Xuan Luo, Jia-Bin Huang, Richard Szeliski, Kevin Matzen, and Johannes Kopf. Consistent video depth estimation. ToG. 5 +[21] Quan Meng, Anpei Chen, Haimin Luo, Minye Wu, Hao Su, Lan Xu, Xuming He, and Jingyi Yu. Gnerf: Gan-based neural radiance field without posed camera. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. 2 +[22] S Mahdi H Miangoleh, Sebastian Dille, Long Mai, Sylvain Paris, and Yagiz Aksoy. Boosting monocular depth estimation models to high-resolution via content-adaptive multi-resolution merging. In CVPR, 2021. 2 +[23] Ben Mildenhall, Pratul P Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM Transactions on Graphics (TOG), 2019. 2 +[24] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 2021. 1, 2, 3, 5, 7 +[25] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph. 2 +[26] Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In CVPR, 2022. 2 +[27] Michael Oechsle, Songyou Peng, and Andreas Geiger. Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In ICCV, 2021. 2 +[28] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In ICCV, 2021. 2, 4 +[29] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. TPAMI, 2020. 2 +[30] Gernot Riegler and Vladlen Koltun. Free view synthesis. In ECCV, 2020. 2 +[31] Gernot Riegler and Vladlen Koltun. Stable view synthesis. In CVPR, 2021. 2 +[32] Barbara Roessle, Jonathan T Barron, Ben Mildenhall, Pratul P Srinivasan, and Matthias Nießner. Dense depth priors for neural radiance fields from sparse input views. In CVPR, 2022. 2 +[33] Antoni Rosinol, John J Leonard, and Luca Carlone. Nerf-slam: Real-time dense monocular slam with neural radiance fields. arXiv preprint arXiv:2210.13641, 2022. 2 +[34] Sara Fridovich-Keil and Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In CVPR, 2022. 2 +[35] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In CVPR, 2016. 1, 2 + +[36] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. NeurIPS, 2020. 2 +[37] Jürgen Sturm, Nikolas Engelhard, Felix Endres, Wolfram Burgard, and Daniel Cremers. A benchmark for the evaluation of rgb-d slam systems. In IROS. IEEE, 2012. 5 +[38] Edgar Sucar, Shikun Liu, Joseph Ortiz, and Andrew Davison. iMAP: Implicit mapping and positioning in real-time. In ICCV, 2021. 2 +[39] Libo Sun, Jia-Wang Bian, Huangying Zhan, Wei Yin, Ian Reid, and Chunhua Shen. Sc-depthv3: Robust self-supervised monocular depth estimation for dynamic scenes. arXiv preprint arXiv:2211.03660, 2022. 5 +[40] Alex Trevithick and Bo Yang. Grf: Learning a general radiance field for 3d representation and rendering. In ICCV, 2021. 2 +[41] Richard Tucker and Noah Snavely. Single-view view synthesis with multiplane images. In CVPR, 2020. 2 +[42] Shubham Tulsiani, Richard Tucker, and Noah Snavely. Layer-structured 3d scene inference via view synthesis. In ECCV, 2018. 2 +[43] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. NeurIPS, 2021. 2 +[44] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P Srinivasan, Howard Zhou, Jonathan T Barron, Ricardo Martin-Brualla, Noah Snively, and Thomas Funkhouser. Ibrnet: Learning multi-view image-based rendering. In CVPR, 2021. 2 +[45] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 2004. 5 +[46] Zirui Wang, Shangzhe Wu, Weidi Xie, Min Chen, and Victor Adrian Prisacariu. NeRF---: Neural radiance fields without known camera parameters. arXiv preprint arXiv:2102.07064, 2021. 1, 2, 3, 4, 5, 6, 7 +[47] Yi Wei, Shaohui Liu, Yongming Rao, Wang Zhao, Jiwen Lu, and Jie Zhou. Nerfingmvs: Guided optimization of neural radiance fields for indoor multi-view stereo. In ICCV, 2021. 2 +[48] Yitong Xia, Hao Tang, Radu Timofte, and Luc Van Gool. Sinerf: Sinusoidal neural radiance fields for joint pose estimation and scene reconstruction. 2022. 2 +[49] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. In NeurIPS, 2021. 2 +[50] Lin Yen-Chen, Pete Florence, Jonathan T Barron, Alberto Rodriguez, Phillip Isola, and Tsung-Yi Lin. inerf: Inverting neural radiance fields for pose estimation. In IROS, 2021. 2 +[51] Wei Yin, Jianming Zhang, Oliver Wang, Simon Niklaus, Simon Chen, Yifan Liu, and Chunhua Shen. Towards accurate reconstruction of 3d scene shape from a single monocular image. TPAMI, 2022. 2 +[52] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In CVPR, 2021. 2 + +[53] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. NeurIPS, 2022. 2 +[54] Jian Zhang, Yuanqing Zhang, Huan Fu, Xiaowei Zhou, Bowen Cai, Jinchi Huang, Rongfei Jia, Binqiang Zhao, and Xing Tang. Ray priors through reprojection: Improving neural radiance fields for novel view extrapolation. In CVPR, 2022. 2 +[55] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492, 2020. 2 +[56] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 5 +[57] Zichao Zhang and Davide Scaramuzza. A tutorial on quantitative trajectory evaluation for visual (-inertial) odometry. In IROS. IEEE, 2018. 5 +[58] Tinghui Zhou, Matthew Brown, Noah Snavely, and David G Lowe. Unsupervised learning of depth and ego-motion from video. In CVPR, 2017. 5, 7 +[59] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images. 2018. 2 +[60] Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hujun Bao, Zhaopeng Cui, Martin R. Oswald, and Marc Pollefeys. Nice-slam: Neural implicit scalable encoding for slam. In CVPR, 2022. 2 \ No newline at end of file diff --git a/2023/NoPe-NeRF_ Optimising Neural Radiance Field With No Pose Prior/images.zip b/2023/NoPe-NeRF_ Optimising Neural Radiance Field With No Pose Prior/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1b737d43183622a35ca4cccbb9afff6fafd21caa --- /dev/null +++ b/2023/NoPe-NeRF_ Optimising Neural Radiance Field With No Pose Prior/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:39ca6c14ccdb17d95abac3c948ead25536218fb0437451297d2d48d69f323490 +size 895811 diff --git a/2023/NoPe-NeRF_ Optimising Neural Radiance Field With No Pose Prior/layout.json b/2023/NoPe-NeRF_ Optimising Neural Radiance Field With No Pose Prior/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4c7e5395db7c8e506dfaece2a9f3ce82c8976391 --- /dev/null +++ b/2023/NoPe-NeRF_ Optimising Neural Radiance Field With No Pose Prior/layout.json @@ -0,0 +1,9218 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 89, + 103, + 504, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 103, + 504, + 120 + ], + "spans": [ + { + "bbox": [ + 89, + 103, + 504, + 120 + ], + "type": "text", + "content": "NoPe-NeRF: Optimising Neural Radiance Field with No Pose Prior" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 126, + 143, + 195, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 143, + 195, + 158 + ], + "spans": [ + { + "bbox": [ + 126, + 143, + 195, + 158 + ], + "type": "text", + "content": "Wenjing Bian" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 228, + 144, + 285, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 228, + 144, + 285, + 157 + ], + "spans": [ + { + "bbox": [ + 228, + 144, + 285, + 157 + ], + "type": "text", + "content": "Zirui Wang" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 319, + 144, + 361, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 144, + 361, + 157 + ], + "spans": [ + { + "bbox": [ + 319, + 144, + 361, + 157 + ], + "type": "text", + "content": "Kejie Li" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 394, + 144, + 467, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 144, + 467, + 157 + ], + "spans": [ + { + "bbox": [ + 394, + 144, + 467, + 157 + ], + "type": "text", + "content": "Jia-Wang Bian" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 236, + 158, + 356, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 158, + 356, + 171 + ], + "spans": [ + { + "bbox": [ + 236, + 158, + 356, + 171 + ], + "type": "text", + "content": "Victor Adrian Prisacariu" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 199, + 172, + 395, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 199, + 172, + 395, + 185 + ], + "spans": [ + { + "bbox": [ + 199, + 172, + 395, + 185 + ], + "type": "text", + "content": "Active Vision Lab, University of Oxford" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 146, + 187, + 443, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 187, + 443, + 199 + ], + "spans": [ + { + "bbox": [ + 146, + 187, + 443, + 199 + ], + "type": "text", + "content": "{wenjing, ryan, kejie, jiawang, victor}@robots.ox.ac.uk" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 61, + 226, + 132, + 300 + ], + "blocks": [ + { + "bbox": [ + 61, + 226, + 132, + 300 + ], + "lines": [ + { + "bbox": [ + 61, + 226, + 132, + 300 + ], + "spans": [ + { + "bbox": [ + 61, + 226, + 132, + 300 + ], + "type": "image", + "image_path": "6456383c90eff0c41d60483ff6ee3639f9c2a5a312a21425f09678ffb0312283.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 61, + 301, + 132, + 373 + ], + "blocks": [ + { + "bbox": [ + 61, + 301, + 132, + 373 + ], + "lines": [ + { + "bbox": [ + 61, + 301, + 132, + 373 + ], + "spans": [ + { + "bbox": [ + 61, + 301, + 132, + 373 + ], + "type": "image", + "image_path": "cc1a091e00edbf3086064344697f8b788c846e7fa1b750e06194dde68cef4400.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 384, + 546, + 407 + ], + "lines": [ + { + "bbox": [ + 46, + 384, + 546, + 407 + ], + "spans": [ + { + "bbox": [ + 46, + 384, + 546, + 407 + ], + "type": "text", + "content": "Figure 1. Novel view synthesis comparison. We propose NoPe-NeRF for joint pose estimation and novel view synthesis. Our method enables more robust pose estimation and renders better novel view synthesis than previous state-of-the-art methods." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 134, + 227, + 296, + 300 + ], + "blocks": [ + { + "bbox": [ + 134, + 227, + 296, + 300 + ], + "lines": [ + { + "bbox": [ + 134, + 227, + 296, + 300 + ], + "spans": [ + { + "bbox": [ + 134, + 227, + 296, + 300 + ], + "type": "image", + "image_path": "7726bb48c80d2abc808b1dbcafa6d8a0a37b77fd79638ab3242ff39329c4c3a4.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 134, + 300, + 296, + 373 + ], + "blocks": [ + { + "bbox": [ + 134, + 300, + 296, + 373 + ], + "lines": [ + { + "bbox": [ + 134, + 300, + 296, + 373 + ], + "spans": [ + { + "bbox": [ + 134, + 300, + 296, + 373 + ], + "type": "image", + "image_path": "7ce87e153f7b0d9c78f8a1d851b3295017619fdf465ce0762f817e85be6a1128.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 296, + 227, + 459, + 300 + ], + "blocks": [ + { + "bbox": [ + 296, + 227, + 459, + 300 + ], + "lines": [ + { + "bbox": [ + 296, + 227, + 459, + 300 + ], + "spans": [ + { + "bbox": [ + 296, + 227, + 459, + 300 + ], + "type": "image", + "image_path": "0947429dff822ae0042cec68959abcc2791f07b733fa4ff7c712da433d518cde.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 296, + 300, + 459, + 373 + ], + "blocks": [ + { + "bbox": [ + 296, + 300, + 459, + 373 + ], + "lines": [ + { + "bbox": [ + 296, + 300, + 459, + 373 + ], + "spans": [ + { + "bbox": [ + 296, + 300, + 459, + 373 + ], + "type": "image", + "image_path": "6c2fe2396fdbe9b3d56187ccb624b9fdfad4f3be627a7be220df5c061ccf6c37.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 461, + 227, + 535, + 300 + ], + "blocks": [ + { + "bbox": [ + 461, + 227, + 535, + 300 + ], + "lines": [ + { + "bbox": [ + 461, + 227, + 535, + 300 + ], + "spans": [ + { + "bbox": [ + 461, + 227, + 535, + 300 + ], + "type": "image", + "image_path": "829f40ee63e1d95e7d7a6a16af554c09194ea234ce0181c9338caf0da21fac81.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 461, + 300, + 535, + 373 + ], + "blocks": [ + { + "bbox": [ + 461, + 300, + 535, + 373 + ], + "lines": [ + { + "bbox": [ + 461, + 300, + 535, + 373 + ], + "spans": [ + { + "bbox": [ + 461, + 300, + 535, + 373 + ], + "type": "image", + "image_path": "560597951ee9f257e9d09e683c2aa220680a9af6f1788756a94fc6199ae59ce9.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 143, + 416, + 192, + 428 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 416, + 192, + 428 + ], + "spans": [ + { + "bbox": [ + 143, + 416, + 192, + 428 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 45, + 441, + 289, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 441, + 289, + 646 + ], + "spans": [ + { + "bbox": [ + 45, + 441, + 289, + 646 + ], + "type": "text", + "content": "Training a Neural Radiance Field (NeRF) without precomputed camera poses is challenging. Recent advances in this direction demonstrate the possibility of jointly optimising a NeRF and camera poses in forward-facing scenes. However, these methods still face difficulties during dramatic camera movement. We tackle this challenging problem by incorporating undistorted monocular depth priors. These priors are generated by correcting scale and shift parameters during training, with which we are then able to constrain the relative poses between consecutive frames. This constraint is achieved using our proposed novel loss functions. Experiments on real-world indoor and outdoor scenes show that our method can handle challenging camera trajectories and outperforms existing methods in terms of novel view rendering quality and pose estimation accuracy. Our project page is https://nope-nerf.active.vision." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 47, + 669, + 127, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 669, + 127, + 681 + ], + "spans": [ + { + "bbox": [ + 47, + 669, + 127, + 681 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "text", + "content": "The photo-realistic reconstruction of a scene from a stream of RGB images requires both accurate 3D geometry" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 417, + 545, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 417, + 545, + 477 + ], + "spans": [ + { + "bbox": [ + 304, + 417, + 545, + 477 + ], + "type": "text", + "content": "reconstruction and view-dependent appearance modelling. Recently, Neural Radiance Fields (NeRF) [24] have demonstrated the ability to build high-quality results for generating photo-realistic images from novel viewpoints given a sparse set of images." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 481, + 546, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 481, + 546, + 637 + ], + "spans": [ + { + "bbox": [ + 304, + 481, + 546, + 637 + ], + "type": "text", + "content": "An important preparation step for NeRF training is the estimation of camera parameters for the input images. A current go-to option is the popular Structure-from-Motion (SfM) library COLMAP [35]. Whilst easy to use, this preprocessing step could be an obstacle to NeRF research and real-world deployments in the long term due to its long processing time and its lack of differentiability. Recent works such as NeRFmm [46], BARF [18] and SC-NeRF [12] propose to simultaneously optimise camera poses and the neural implicit representation to address these issues. Nevertheless, these methods can only handle forward-facing scenes when no initial parameters are supplied, and fail in dramatic camera motions, e.g.a casual handheld captured video." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 641, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 714 + ], + "type": "text", + "content": "This limitation has two key causes. First, all these methods estimate a camera pose for each input image individually without considering relative poses between images. Looking back to the literature of Simultaneous localisation and mapping (SLAM) and visual odometry, pose estimation can significantly benefit from estimating relative poses be" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4160" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "type": "text", + "content": "tween adjacent input frames. Second, the radiance field is known to suffer from shape-radiance ambiguity [55]. Estimating camera parameters jointly with NeRF adds another degree of ambiguity, resulting in slow convergence and unstable optimisation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 133, + 286, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 133, + 286, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 133, + 286, + 251 + ], + "type": "text", + "content": "To handle the limitation of large camera motion, we seek help from monocular depth estimation [22, 28, 29, 51]. Our motivation is threefold: First, monocular depth provides strong geometry cues that are beneficial to constraint shape-raliance ambiguity. Second, relative poses between adjacent depth maps can be easily injected into the training pipeline via Chamfer Distance. Third, monocular depth is lightweight to run and does not require camera parameters as input, in contrast to multi-view stereo depth estimation. For simplicity, we use the term mono-depth from now on." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 251, + 286, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 251, + 286, + 453 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 286, + 453 + ], + "type": "text", + "content": "Utilising mono-depth effectively is not straightforward with the presence of scale and shift distortions. In other words, mono-depth maps are not multi-view consistent. Previous works [9, 17, 47] simply take mono-depth into a depth-wise loss along with NeRF training. Instead, we propose a novel and effective way to thoroughly integrate mono-depth into our system. First, we explicitly optimise scale and shift parameters for each mono-depth map during NeRF training by penalising the difference between rendered depth and mono-depth. Since NeRF by itself is trained based on multiview consistency, this step transforms mono-depth maps to undistorted multiview consistent depth maps. We further leverage these multiview consistent depth maps in two loss terms: a) a Chamfer Distance loss between two depth maps of adjacent images, which injects relative pose to our system; and b) a depth-based surface rendering loss, which further improves relative pose estimation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 455, + 286, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 455, + 286, + 574 + ], + "spans": [ + { + "bbox": [ + 46, + 455, + 286, + 574 + ], + "type": "text", + "content": "In summary, we propose a method to jointly optimise camera poses and a NeRF from a sequence of images with large camera motion. Our system is enabled by three contributions. First, we propose a novel way to integrate monodepth into unposed-NeRF training by explicitly modelling scale and shift distortions. Second, we supply relative poses to the camera-NeRF joint optimisation via an inter-frame loss using undistorted mono-depth maps. Third, we further regularise our relative pose estimation with a depth-based surface rendering loss." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 574, + 286, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 574, + 286, + 622 + ], + "spans": [ + { + "bbox": [ + 46, + 574, + 286, + 622 + ], + "type": "text", + "content": "As a result, our method is able to handle large camera motion, and outperforms state-of-the-art methods by a significant margin in terms of novel view synthesis quality and camera trajectory accuracy." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 633, + 133, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 633, + 133, + 645 + ], + "spans": [ + { + "bbox": [ + 47, + 633, + 133, + 645 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 653, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 286, + 712 + ], + "type": "text", + "content": "Novel View Synthesis. While early Novel View Synthesis (NVS) approaches applied interpolations between pixels [3], later works often rendered images from 3D reconstruction [1, 6]. In recent years, different representations of the 3D scene are used, e.g.meshes [30, 31], Multi-Plane" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 72, + 545, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 107 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 107 + ], + "type": "text", + "content": "Images [41, 59], layered depth [42] etc. Among them, NeRF [24] has become a popular scene representation for its photorealistic rendering." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 108, + 545, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 108, + 545, + 192 + ], + "spans": [ + { + "bbox": [ + 304, + 108, + 545, + 192 + ], + "type": "text", + "content": "A number of techniques are proposed to improve NeRF's performance with additional regularisation [13, 26, 54], depth priors [7, 32, 47, 53], surface enhancements [27, 43, 49] or latent codes [40, 44, 52]. Other works [2, 10, 25, 34] have also accelerated NeRF training and rendering. However, most of these approaches require pre-computed camera parameters obtained from SfM algorithms [11, 35]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 205, + 545, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 205, + 545, + 540 + ], + "spans": [ + { + "bbox": [ + 304, + 205, + 545, + 540 + ], + "type": "text", + "content": "NeRF With Pose Optimisation. Removing camera parameter preprocessing is an active line of research recently. One category of the methods [33, 38, 60] use a SLAM-style pipeline, that either requires RGB-D inputs or relies on accurate camera poses generated from the SLAM tracking system. Another category of works optimises camera poses with the NeRF model directly. We term this type of method as unposed-NeRF in this paper. iNeRF [50] shows that poses for novel view images can be estimated using a reconstructed NeRF model. GNeRF [21] combines Generative Adversarial Networks with NeRF to estimate camera poses but requires a known sampling distribution for poses. More relevant to our work, NeRFmm [46] jointly optimises both camera intrinsic and extrinsic alongside NeRF training. BARF [18] proposes a coarse-to-fine positional encoding strategy for camera poses and NeRF joint optimisation. SCNeRF [12] further parameterises camera distortion and employs a geometric loss to regularise rays. GARF [4] shows that using Gaussian-MLPs makes joint pose and scene optimisation easier and more accurate. Recently, SiNeRF [48] uses SIREN [36] layers and a novel sampling strategy to alleviate the sub-optimality of joint optimisation in NeRFmm. Although showing promising results on the forward-facing dataset like LLFF [23], these approaches face difficulties when handling challenging camera trajectories with large camera motion. We address this issue by closely integrating mono-depth maps with the joint optimisation of camera parameters and NeRF." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 550, + 361, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 550, + 361, + 562 + ], + "spans": [ + { + "bbox": [ + 306, + 550, + 361, + 562 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 570, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 689 + ], + "type": "text", + "content": "We tackle the challenge of handling large camera motions in unposed-NeRF training. Given a sequence of images, camera intrinsics, and their mono-depth estimations, our method recovers camera poses and optimises a NeRF simultaneously. We assume camera intrinsics are available in the image meta block, and we run an off-the-shelf monodepth network DPT [7] to acquire mono-depth estimations. Without repeating the benefit of mono-depth, we unroll this section around the effective integration of monocular depth into unposed-NeRF training." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "The training is a joint optimisation of the NeRF, camera poses, and distortion parameters of each mono-depth" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "type": "text", + "content": "4161" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 69, + 318, + 293 + ], + "blocks": [ + { + "bbox": [ + 47, + 69, + 318, + 293 + ], + "lines": [ + { + "bbox": [ + 47, + 69, + 318, + 293 + ], + "spans": [ + { + "bbox": [ + 47, + 69, + 318, + 293 + ], + "type": "image", + "image_path": "1d697123e9c97e03383d67eb3111e6dc00d5feeec2e1903c66e745ca1242a200.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 302, + 547, + 335 + ], + "lines": [ + { + "bbox": [ + 45, + 302, + 547, + 335 + ], + "spans": [ + { + "bbox": [ + 45, + 302, + 547, + 335 + ], + "type": "text", + "content": "Figure 2. Method Overview. Our method takes a sequence of images as input to reconstruct NeRF and jointly estimates the camera poses of the frames. We first generate monocular depth maps from a mono-depth estimation network and reconstruct the point clouds. We then optimise NeRF, camera poses, and depth distortion parameters jointly with inter-frame and NeRF losses." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 325, + 69, + 451, + 277 + ], + "blocks": [ + { + "bbox": [ + 325, + 69, + 451, + 277 + ], + "lines": [ + { + "bbox": [ + 325, + 69, + 451, + 277 + ], + "spans": [ + { + "bbox": [ + 325, + 69, + 451, + 277 + ], + "type": "image", + "image_path": "429c0ded872daf15aad5f8b8de9aab2f06cbcc399ff69748ba8379356f538e90.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 461, + 70, + 545, + 291 + ], + "blocks": [ + { + "bbox": [ + 461, + 70, + 545, + 291 + ], + "lines": [ + { + "bbox": [ + 461, + 70, + 545, + 291 + ], + "spans": [ + { + "bbox": [ + 461, + 70, + 545, + 291 + ], + "type": "image", + "image_path": "b907a9e723a3afab4a8fd3a5ab307f4a6530b07b83d50befff349e18407a1dbf.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 356, + 287, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 356, + 287, + 429 + ], + "spans": [ + { + "bbox": [ + 46, + 356, + 287, + 429 + ], + "type": "text", + "content": "map. The distortion parameters are supervised by minimising the discrepancies between the mono-depth maps and depth maps rendered from the NeRF, which are multiview consistent. The undistorted depth maps in return effectively mediate the shape-radiance ambiguity, which eases the training of NeRF and camera poses." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 430, + 287, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 430, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 46, + 430, + 287, + 525 + ], + "type": "text", + "content": "Specifically, the undistorted depth maps enable two constraints. We constrain global pose estimation by supplying relative pose between adjacent images. This is achieved via a Chamfer-Distance-based correspondence between two point clouds, back-projected from undistorted depth maps. Further, we regularise relative pose estimation with a surface-based photometric consistency where we treat undistorted depth as surface." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 526, + 287, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 526, + 287, + 588 + ], + "spans": [ + { + "bbox": [ + 46, + 526, + 287, + 588 + ], + "type": "text", + "content": "We detail our method in the following sections, starting from NeRF in Sec. 3.1 and unposed-NeRF training in Sec. 3.2, looking into mono-depth distortions in Sec. 3.3, followed by our mono-depth enabled loss terms in Sec. 3.4, and finishing with an overall training pipeline Sec. 3.5." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 597, + 97, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 597, + 97, + 609 + ], + "spans": [ + { + "bbox": [ + 47, + 597, + 97, + 609 + ], + "type": "text", + "content": "3.1. NeRF" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "content": "Neural Radiance Field (NeRF) [24] represents a scene as a mapping function " + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "inline_equation", + "content": "F_{\\Theta}:(\\mathbf{x},\\mathbf{d})\\to (\\mathbf{c},\\sigma)" + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "content": " that maps a 3D location " + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{x}\\in \\mathbb{R}^3" + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "content": " and a viewing direction " + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{d}\\in \\mathbb{R}^3" + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "content": " to a radiance colour " + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{c}\\in \\mathbb{R}^3" + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "content": " and a volume density value " + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "content": ". This mapping is usually implemented with a neural network parameterised by " + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "inline_equation", + "content": "F_{\\Theta}" + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "content": ". Given " + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "content": " images " + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{I} = \\{I_i\\mid i = 0\\dots N - 1\\}" + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "content": " with their camera poses " + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\Pi = \\{\\pi_i\\mid i = 0\\dots N - 1\\}" + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "content": ", NeRF can be optimised by minimising photometric error" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 354, + 545, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 354, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 305, + 354, + 545, + 380 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rgb} = \\sum_{i}^{N}\\| I_{i} - \\hat{I}_{i}\\|_{2}^{2}" + }, + { + "bbox": [ + 305, + 354, + 545, + 380 + ], + "type": "text", + "content": " between synthesised images " + }, + { + "bbox": [ + 305, + 354, + 545, + 380 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{I}}" + }, + { + "bbox": [ + 305, + 354, + 545, + 380 + ], + "type": "text", + "content": " and captured images " + }, + { + "bbox": [ + 305, + 354, + 545, + 380 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 305, + 354, + 545, + 380 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 363, + 385, + 545, + 403 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 363, + 385, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 363, + 385, + 545, + 403 + ], + "type": "interline_equation", + "content": "\\Theta^ {*} = \\arg \\min _ {\\Theta} \\mathcal {L} _ {r g b} (\\hat {\\mathcal {I}} \\mid \\mathcal {I}, \\Pi), \\tag {1}", + "image_path": "5f78d22ffa01143033972e12d64b5327e59d0433bd89fdb90426c804343450fd.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 411, + 545, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 411, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 305, + 411, + 545, + 460 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 411, + 545, + 460 + ], + "type": "inline_equation", + "content": "\\hat{I}_i" + }, + { + "bbox": [ + 305, + 411, + 545, + 460 + ], + "type": "text", + "content": " is rendered by aggregating radiance colour on camera rays " + }, + { + "bbox": [ + 305, + 411, + 545, + 460 + ], + "type": "inline_equation", + "content": "\\mathbf{r}(h) = \\mathbf{o} + h\\mathbf{d}" + }, + { + "bbox": [ + 305, + 411, + 545, + 460 + ], + "type": "text", + "content": " between near and far bound " + }, + { + "bbox": [ + 305, + 411, + 545, + 460 + ], + "type": "inline_equation", + "content": "h_n" + }, + { + "bbox": [ + 305, + 411, + 545, + 460 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 411, + 545, + 460 + ], + "type": "inline_equation", + "content": "h_f" + }, + { + "bbox": [ + 305, + 411, + 545, + 460 + ], + "type": "text", + "content": ". More concretely, we synthesise " + }, + { + "bbox": [ + 305, + 411, + 545, + 460 + ], + "type": "inline_equation", + "content": "\\hat{I}_i" + }, + { + "bbox": [ + 305, + 411, + 545, + 460 + ], + "type": "text", + "content": " with a volumetric rendering function" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 342, + 464, + 545, + 493 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 464, + 545, + 493 + ], + "spans": [ + { + "bbox": [ + 342, + 464, + 545, + 493 + ], + "type": "interline_equation", + "content": "\\hat {I} _ {i} (\\mathbf {r}) = \\int_ {h _ {n}} ^ {h _ {f}} T (h) \\sigma (\\mathbf {r} (h)) \\mathbf {c} (\\mathbf {r} (h), \\mathbf {d}) d h, \\tag {2}", + "image_path": "dfbc7c4cef2bd20a02db570268eb4cc6cd8e93fb8f2d477bfee671fd3f5d619c.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 499, + 545, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 499, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 305, + 499, + 545, + 536 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 499, + 545, + 536 + ], + "type": "inline_equation", + "content": "T(h) = \\exp(-\\int_{h_n}^h \\sigma(\\mathbf{r}(s)) ds)" + }, + { + "bbox": [ + 305, + 499, + 545, + 536 + ], + "type": "text", + "content": " is the accumulated transmittance along a ray. We refer to [24] for further details." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 543, + 507, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 543, + 507, + 555 + ], + "spans": [ + { + "bbox": [ + 306, + 543, + 507, + 555 + ], + "type": "text", + "content": "3.2. Joint Optimisation of Poses and NeRF" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 561, + 545, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 561, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 305, + 561, + 545, + 609 + ], + "type": "text", + "content": "Prior works [12, 18, 46] show that it is possible to estimate both camera parameters and a NeRF at the same time by minimising the above photometric error " + }, + { + "bbox": [ + 305, + 561, + 545, + 609 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rgb}" + }, + { + "bbox": [ + 305, + 561, + 545, + 609 + ], + "type": "text", + "content": " under the same volumetric rendering process in Eq. (2)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 609, + 545, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 609, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 304, + 609, + 545, + 656 + ], + "type": "text", + "content": "The key lies in conditioning camera ray casting on variable camera parameters " + }, + { + "bbox": [ + 304, + 609, + 545, + 656 + ], + "type": "inline_equation", + "content": "\\Pi" + }, + { + "bbox": [ + 304, + 609, + 545, + 656 + ], + "type": "text", + "content": ", as the camera ray " + }, + { + "bbox": [ + 304, + 609, + 545, + 656 + ], + "type": "inline_equation", + "content": "\\mathbf{r}" + }, + { + "bbox": [ + 304, + 609, + 545, + 656 + ], + "type": "text", + "content": " is a function of camera pose. Mathematically, this joint optimisation can be formulated as:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 354, + 662, + 545, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 662, + 545, + 681 + ], + "spans": [ + { + "bbox": [ + 354, + 662, + 545, + 681 + ], + "type": "interline_equation", + "content": "\\Theta^ {*}, \\Pi^ {*} = \\arg \\min _ {\\Theta , \\Pi} \\mathcal {L} _ {r g b} (\\hat {\\mathcal {I}}, \\hat {\\Pi} \\mid \\mathcal {I}), \\tag {3}", + "image_path": "90cadcbd7cd4e412a254b86615765af65be150cebd4b7c2e47cd08ba2fcef74a.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\hat{\\Pi}" + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "content": " denotes camera parameters that are updated during optimising. Note that the only difference between Eq. (1)" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4162" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 95 + ], + "type": "text", + "content": "and Eq. (3) is that Eq. (3) considers camera parameters as variables." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 96, + 286, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 96, + 286, + 156 + ], + "spans": [ + { + "bbox": [ + 47, + 96, + 286, + 156 + ], + "type": "text", + "content": "In general, the camera parameters " + }, + { + "bbox": [ + 47, + 96, + 286, + 156 + ], + "type": "inline_equation", + "content": "\\Pi" + }, + { + "bbox": [ + 47, + 96, + 286, + 156 + ], + "type": "text", + "content": " include camera intrinsics, poses, and lens distortions. We only consider estimating camera poses in this work, e.g., camera pose for frame " + }, + { + "bbox": [ + 47, + 96, + 286, + 156 + ], + "type": "inline_equation", + "content": "I_{i}" + }, + { + "bbox": [ + 47, + 96, + 286, + 156 + ], + "type": "text", + "content": " is a transformation " + }, + { + "bbox": [ + 47, + 96, + 286, + 156 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_i = [\\mathbf{R}_i\\mid \\mathbf{t}_i]" + }, + { + "bbox": [ + 47, + 96, + 286, + 156 + ], + "type": "text", + "content": " with a rotation " + }, + { + "bbox": [ + 47, + 96, + 286, + 156 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_i\\in \\mathrm{SO}(3)" + }, + { + "bbox": [ + 47, + 96, + 286, + 156 + ], + "type": "text", + "content": " and a translation " + }, + { + "bbox": [ + 47, + 96, + 286, + 156 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_i\\in \\mathbb{R}^3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 163, + 228, + 176 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 163, + 228, + 176 + ], + "spans": [ + { + "bbox": [ + 47, + 163, + 228, + 176 + ], + "type": "text", + "content": "3.3. Undistortion of Monocular Depth" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 182, + 286, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 182, + 286, + 263 + ], + "spans": [ + { + "bbox": [ + 47, + 182, + 286, + 263 + ], + "type": "text", + "content": "With an off-the-shelf monocular depth network, e.g., DPT [28], we generate mono-depth sequence " + }, + { + "bbox": [ + 47, + 182, + 286, + 263 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{D_i\\mid i = 0\\dots N - 1\\}" + }, + { + "bbox": [ + 47, + 182, + 286, + 263 + ], + "type": "text", + "content": " from input images. Without surprise, mono-depth maps are not multi-view consistent so we aim to recover a sequence of multi-view consistent depth maps, which are further leveraged in our relative pose loss terms." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 266, + 286, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 266, + 286, + 350 + ], + "spans": [ + { + "bbox": [ + 47, + 266, + 286, + 350 + ], + "type": "text", + "content": "Specifically, we consider two linear transformation parameters for each mono-depth map, resulting in a sequence of transformation parameters for all frames " + }, + { + "bbox": [ + 47, + 266, + 286, + 350 + ], + "type": "inline_equation", + "content": "\\Psi = \\{(\\alpha_i,\\beta_i)\\mid i = 0\\ldots N - 1\\}" + }, + { + "bbox": [ + 47, + 266, + 286, + 350 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 266, + 286, + 350 + ], + "type": "inline_equation", + "content": "\\alpha_{i}" + }, + { + "bbox": [ + 47, + 266, + 286, + 350 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 266, + 286, + 350 + ], + "type": "inline_equation", + "content": "\\beta_{i}" + }, + { + "bbox": [ + 47, + 266, + 286, + 350 + ], + "type": "text", + "content": " denote a scale and a shift factor. With multi-view consistent constraint from NeRF, we aim to recover a multi-view consistent depth map " + }, + { + "bbox": [ + 47, + 266, + 286, + 350 + ], + "type": "inline_equation", + "content": "D_{i}^{*}" + }, + { + "bbox": [ + 47, + 266, + 286, + 350 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 47, + 266, + 286, + 350 + ], + "type": "inline_equation", + "content": "D_{i}" + }, + { + "bbox": [ + 47, + 266, + 286, + 350 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 129, + 358, + 287, + 372 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 358, + 287, + 372 + ], + "spans": [ + { + "bbox": [ + 129, + 358, + 287, + 372 + ], + "type": "interline_equation", + "content": "D _ {i} ^ {*} = \\alpha_ {i} D _ {i} + \\beta_ {i}, \\tag {4}", + "image_path": "a8c6d8a18f77aea61b39bb9c4514c4040ca7ae79644a483c19b2f38023a1a8b7.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 380, + 287, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 380, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 46, + 380, + 287, + 430 + ], + "type": "text", + "content": "by joint optimising " + }, + { + "bbox": [ + 46, + 380, + 287, + 430 + ], + "type": "inline_equation", + "content": "\\alpha_{i}" + }, + { + "bbox": [ + 46, + 380, + 287, + 430 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 380, + 287, + 430 + ], + "type": "inline_equation", + "content": "\\beta_{i}" + }, + { + "bbox": [ + 46, + 380, + 287, + 430 + ], + "type": "text", + "content": " along with a NeRF. This joint optimisation is mostly achieved by enforcing the consistency between an undistorted depth map " + }, + { + "bbox": [ + 46, + 380, + 287, + 430 + ], + "type": "inline_equation", + "content": "D_{i}^{*}" + }, + { + "bbox": [ + 46, + 380, + 287, + 430 + ], + "type": "text", + "content": " and a NeRF rendered depth map " + }, + { + "bbox": [ + 46, + 380, + 287, + 430 + ], + "type": "inline_equation", + "content": "\\hat{D}_i" + }, + { + "bbox": [ + 46, + 380, + 287, + 430 + ], + "type": "text", + "content": " via a depth loss:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 110, + 438, + 287, + 471 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 438, + 287, + 471 + ], + "spans": [ + { + "bbox": [ + 110, + 438, + 287, + 471 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {d e p t h}} = \\sum_ {i} ^ {N} \\left\\| D _ {i} ^ {*} - \\hat {D} _ {i} \\right\\|, \\tag {5}", + "image_path": "62812bc0506763a76c7d04888e7f4eae423c6fed10ff8fbd2de7df42138902c6.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 479, + 75, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 479, + 75, + 489 + ], + "spans": [ + { + "bbox": [ + 47, + 479, + 75, + 489 + ], + "type": "text", + "content": "where" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 104, + 488, + 287, + 516 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 488, + 287, + 516 + ], + "spans": [ + { + "bbox": [ + 104, + 488, + 287, + 516 + ], + "type": "interline_equation", + "content": "\\hat {D} _ {i} (\\mathbf {r}) = \\int_ {h _ {n}} ^ {h _ {f}} T (h) \\sigma (\\mathbf {r} (h)) d h \\tag {6}", + "image_path": "6bf672af3cacb41ba2ecbba14bd180e0030b3174c9fbf8b4bf7ee40ff46d6796.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 521, + 264, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 521, + 264, + 532 + ], + "spans": [ + { + "bbox": [ + 47, + 521, + 264, + 532 + ], + "type": "text", + "content": "denotes a volumetric rendered depth map from NeRF." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 532, + 287, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 532, + 287, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 532, + 287, + 604 + ], + "type": "text", + "content": "It is important to note that both NeRF and mono-depth benefit from Eq. (5). On the one hand, mono-depth provides strong geometry prior for NeRF training, reducing shape-radiance ambiguity. On the other hand, NeRF provides multi-view consistency so we can recover a set of multiview consistent depth maps for relative pose estimations." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 611, + 186, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 611, + 186, + 623 + ], + "spans": [ + { + "bbox": [ + 47, + 611, + 186, + 623 + ], + "type": "text", + "content": "3.4. Relative Pose Constraint" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "content": "Aforementioned unposed-NeRF methods [12, 18, 46] optimise each camera pose independently, resulting in an overfit to target images with incorrect poses. Penalising incorrect relative poses between frames can help to regularise the joint optimisation towards smooth convergence, especially in a complex camera trajectory. Therefore, we propose two losses that constrain relative poses." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "text", + "content": "Point Cloud Loss. We back-project the undistorted depth maps " + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^*" + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "text", + "content": " using the known camera intrinsics, to point clouds " + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^* = \\{P_i^* \\mid i = 0 \\dots N - 1\\}" + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "text", + "content": " and optimise the relative pose between consecutive point clouds by minimising a point cloud loss " + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{pc}" + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 368, + 140, + 545, + 167 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 368, + 140, + 545, + 167 + ], + "spans": [ + { + "bbox": [ + 368, + 140, + 545, + 167 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {p c} = \\sum_ {(i, j)} l _ {c d} \\left(P _ {j} ^ {*}, \\mathbf {T} _ {j i} P _ {i} ^ {*}\\right), \\tag {7}", + "image_path": "f74e1fbcbecdd152065b1a33835f85a8c8a12310e61e21cd082f2ae142d5dac4.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 175, + 545, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 175, + 545, + 222 + ], + "spans": [ + { + "bbox": [ + 305, + 175, + 545, + 222 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 175, + 545, + 222 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{ji} = \\mathbf{T}_j\\mathbf{T}_i^{-1}" + }, + { + "bbox": [ + 305, + 175, + 545, + 222 + ], + "type": "text", + "content": " represents the related pose that transforms point cloud " + }, + { + "bbox": [ + 305, + 175, + 545, + 222 + ], + "type": "inline_equation", + "content": "P_{i}^{*}" + }, + { + "bbox": [ + 305, + 175, + 545, + 222 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 305, + 175, + 545, + 222 + ], + "type": "inline_equation", + "content": "P_{j}^{*}" + }, + { + "bbox": [ + 305, + 175, + 545, + 222 + ], + "type": "text", + "content": ", tuple " + }, + { + "bbox": [ + 305, + 175, + 545, + 222 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 305, + 175, + 545, + 222 + ], + "type": "text", + "content": " denotes indices of a consecutive pair of instances, and " + }, + { + "bbox": [ + 305, + 175, + 545, + 222 + ], + "type": "inline_equation", + "content": "l_{cd}" + }, + { + "bbox": [ + 305, + 175, + 545, + 222 + ], + "type": "text", + "content": " denotes Chamfer Distance:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 229, + 547, + 266 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 229, + 547, + 266 + ], + "spans": [ + { + "bbox": [ + 305, + 229, + 547, + 266 + ], + "type": "interline_equation", + "content": "l _ {c d} \\left(P _ {i}, P _ {j}\\right) = \\sum_ {p _ {i} \\in P _ {i}} \\min _ {p _ {j} \\in P _ {j}} \\| p _ {i} - p _ {j} \\| _ {2} + \\sum_ {p _ {j} \\in P _ {j}} \\min _ {p _ {i} \\in P _ {i}} \\| p _ {i} - p _ {j} \\| _ {2}. \\tag {8}", + "image_path": "f0c2843291e3d42c1a0a1432ba7bf8d77952f0752a343b1f737aad8106f1ad99.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 281, + 545, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 281, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 305, + 281, + 545, + 376 + ], + "type": "text", + "content": "Surface-based Photometric Loss. While the point cloud loss " + }, + { + "bbox": [ + 305, + 281, + 545, + 376 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{pc}" + }, + { + "bbox": [ + 305, + 281, + 545, + 376 + ], + "type": "text", + "content": " offers supervision in terms of 3D-3D matching, we observe that a surface-based photometric error can alleviate incorrect matching. With the photometric consistency assumption, this photometric error penalises the differences in appearance between associated pixels. The association is established by projecting the point cloud " + }, + { + "bbox": [ + 305, + 281, + 545, + 376 + ], + "type": "inline_equation", + "content": "P_{i}^{*}" + }, + { + "bbox": [ + 305, + 281, + 545, + 376 + ], + "type": "text", + "content": " onto images " + }, + { + "bbox": [ + 305, + 281, + 545, + 376 + ], + "type": "inline_equation", + "content": "I_{i}" + }, + { + "bbox": [ + 305, + 281, + 545, + 376 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 281, + 545, + 376 + ], + "type": "inline_equation", + "content": "I_{j}" + }, + { + "bbox": [ + 305, + 281, + 545, + 376 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 377, + 545, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 377, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 306, + 377, + 545, + 399 + ], + "type": "text", + "content": "The surface-based photometric loss can then be defined as:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 319, + 406, + 545, + 433 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 406, + 545, + 433 + ], + "spans": [ + { + "bbox": [ + 319, + 406, + 545, + 433 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {r g b - s} = \\sum_ {(i, j)} \\| I _ {i} \\langle \\mathbf {K} _ {i} P _ {i} ^ {*} \\rangle - I _ {j} \\langle \\mathbf {K} _ {j} \\mathbf {T} _ {j} \\mathbf {T} _ {i} ^ {- 1} P _ {i} ^ {*} \\rangle \\|, \\tag {9}", + "image_path": "8771e44fc14bc7ccb2c783642d1ff4b54dde7eefa541ec8a162bff70f637b8a1.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 441, + 545, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 441, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 305, + 441, + 545, + 464 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 441, + 545, + 464 + ], + "type": "inline_equation", + "content": "\\langle \\cdot \\rangle" + }, + { + "bbox": [ + 305, + 441, + 545, + 464 + ], + "type": "text", + "content": " represents the sampling operation on the image and " + }, + { + "bbox": [ + 305, + 441, + 545, + 464 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_i" + }, + { + "bbox": [ + 305, + 441, + 545, + 464 + ], + "type": "text", + "content": " denotes a projection matrix for " + }, + { + "bbox": [ + 305, + 441, + 545, + 464 + ], + "type": "inline_equation", + "content": "i_{th}" + }, + { + "bbox": [ + 305, + 441, + 545, + 464 + ], + "type": "text", + "content": " camera." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 472, + 448, + 484 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 472, + 448, + 484 + ], + "spans": [ + { + "bbox": [ + 306, + 472, + 448, + 484 + ], + "type": "text", + "content": "3.5. Overall Training Pipeline" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 490, + 544, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 490, + 544, + 512 + ], + "spans": [ + { + "bbox": [ + 306, + 490, + 544, + 512 + ], + "type": "text", + "content": "Assembling all loss terms, we get the overall loss function:" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 325, + 521, + 545, + 535 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 521, + 545, + 535 + ], + "spans": [ + { + "bbox": [ + 325, + 521, + 545, + 535 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {r g b} + \\lambda_ {1} \\mathcal {L} _ {\\text {d e p t h}} + \\lambda_ {2} \\mathcal {L} _ {p c} + \\lambda_ {3} \\mathcal {L} _ {r g b - s}, \\tag {10}", + "image_path": "f9b8078a97272c7ae178ff4a42f418211f3e8ea76c5b3ac34179bf739b501012.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 305, + 542, + 545, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 542, + 545, + 566 + ], + "spans": [ + { + "bbox": [ + 305, + 542, + 545, + 566 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 542, + 545, + 566 + ], + "type": "inline_equation", + "content": "\\lambda_1, \\lambda_2, \\lambda_3" + }, + { + "bbox": [ + 305, + 542, + 545, + 566 + ], + "type": "text", + "content": " are the weighting factors for respective loss terms. By minimising the combined of loss " + }, + { + "bbox": [ + 305, + 542, + 545, + 566 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 305, + 542, + 545, + 566 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 322, + 572, + 545, + 592 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 572, + 545, + 592 + ], + "spans": [ + { + "bbox": [ + 322, + 572, + 545, + 592 + ], + "type": "interline_equation", + "content": "\\Theta^ {*}, \\Pi^ {*}, \\Psi^ {*} = \\arg \\min _ {\\Theta , \\Pi , \\Psi} \\mathcal {L} (\\hat {\\mathcal {I}}, \\hat {\\mathcal {D}}, \\hat {\\Pi}, \\hat {\\Psi} \\mid \\mathcal {I}, \\mathcal {D}), \\tag {11}", + "image_path": "d4e5a93e2585a1244184522801454979a72b42665c5665c5ede23c8b2aae0c53.jpg" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 305, + 599, + 545, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 599, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 305, + 599, + 545, + 624 + ], + "type": "text", + "content": "our method returns the optimised NeRF parameters " + }, + { + "bbox": [ + 305, + 599, + 545, + 624 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 305, + 599, + 545, + 624 + ], + "type": "text", + "content": ", camera poses " + }, + { + "bbox": [ + 305, + 599, + 545, + 624 + ], + "type": "inline_equation", + "content": "\\Pi" + }, + { + "bbox": [ + 305, + 599, + 545, + 624 + ], + "type": "text", + "content": ", and distortion parameters " + }, + { + "bbox": [ + 305, + 599, + 545, + 624 + ], + "type": "inline_equation", + "content": "\\Psi" + }, + { + "bbox": [ + 305, + 599, + 545, + 624 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 306, + 634, + 387, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 634, + 387, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 634, + 387, + 647 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 305, + 654, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 654, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 305, + 654, + 545, + 712 + ], + "type": "text", + "content": "We begin with a description of our experimental setup in Sec. 4.1. In Sec. 4.2, we compare our method with pose-unknown methods. Next, we compare our method with the COLMAP-assisted NeRF baseline in Sec. 4.3. Lastly, we conduct ablation studies in Sec. 4.4." + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4163" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 72, + 545, + 239 + ], + "blocks": [ + { + "bbox": [ + 49, + 72, + 545, + 239 + ], + "lines": [ + { + "bbox": [ + 49, + 72, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 49, + 72, + 545, + 239 + ], + "type": "image", + "image_path": "bf97d274716646e4a0f70d55a3c67c74a0405cd3380efc61fc648966d5f7f921.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 72, + 243, + 111, + 254 + ], + "lines": [ + { + "bbox": [ + 72, + 243, + 111, + 254 + ], + "spans": [ + { + "bbox": [ + 72, + 243, + 111, + 254 + ], + "type": "text", + "content": "(a) BARF" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 170, + 243, + 221, + 254 + ], + "lines": [ + { + "bbox": [ + 170, + 243, + 221, + 254 + ], + "spans": [ + { + "bbox": [ + 170, + 243, + 221, + 254 + ], + "type": "text", + "content": "(b) NeRFmm" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 270, + 243, + 321, + 254 + ], + "lines": [ + { + "bbox": [ + 270, + 243, + 321, + 254 + ], + "spans": [ + { + "bbox": [ + 270, + 243, + 321, + 254 + ], + "type": "text", + "content": "(c) SC-NeRF" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 380, + 244, + 412, + 254 + ], + "lines": [ + { + "bbox": [ + 380, + 244, + 412, + 254 + ], + "spans": [ + { + "bbox": [ + 380, + 244, + 412, + 254 + ], + "type": "text", + "content": "(d) Ours" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 457, + 244, + 523, + 254 + ], + "lines": [ + { + "bbox": [ + 457, + 244, + 523, + 254 + ], + "spans": [ + { + "bbox": [ + 457, + 244, + 523, + 254 + ], + "type": "text", + "content": "(e) Ground Truth" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 263, + 546, + 285 + ], + "lines": [ + { + "bbox": [ + 46, + 263, + 546, + 285 + ], + "spans": [ + { + "bbox": [ + 46, + 263, + 546, + 285 + ], + "type": "text", + "content": "Figure 3. Qualitative results of novel view synthesis and depth prediction on Tanks and Temples. We visualise the synthesised images and the rendered depth maps (top left of each image) for all methods. NoPe-NeRF is able to recover details for both colour and geometry." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 306, + 163, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 306, + 163, + 319 + ], + "spans": [ + { + "bbox": [ + 47, + 306, + 163, + 319 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 325, + 290, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 325, + 290, + 564 + ], + "spans": [ + { + "bbox": [ + 46, + 325, + 290, + 564 + ], + "type": "text", + "content": "Datasets. We conduct experiments on two datasets Tanks and Temples [15] and ScanNet [5]. Tanks and Temples: we use 8 scenes to evaluate pose accuracy and novel view synthesis quality. We chose scenes captured at both indoor and outdoor locations, with different frame sampling rates and lengths. All images are down-sampled to a resolution of " + }, + { + "bbox": [ + 46, + 325, + 290, + 564 + ], + "type": "inline_equation", + "content": "960 \\times 540" + }, + { + "bbox": [ + 46, + 325, + 290, + 564 + ], + "type": "text", + "content": ". For the family scene, we sample 200 images and take 100 frames with odd frame ids as training images and the remaining 100 frames for novel view synthesis, in order to analyse the performance under smooth motion. For the remaining scenes, following NeRF [24], 1/8 of the images in each sequence are held out for novel view synthesis, unless otherwise specified. ScanNet: we select 4 scenes for evaluating pose accuracy, depth accuracy, and novel view synthesis quality. For each scene, we take 80-100 consecutive images and use 1/8 of these images for novel view synthesis. For evaluation, we employ depth maps and poses provided by ScanNet as ground truth. ScanNet images are down-sampled to " + }, + { + "bbox": [ + 46, + 325, + 290, + 564 + ], + "type": "inline_equation", + "content": "648 \\times 484" + }, + { + "bbox": [ + 46, + 325, + 290, + 564 + ], + "type": "text", + "content": ". We crop images with dark orders during preprocessing." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "text", + "content": "Metrics. We evaluate our proposed method in three aspects. For novel view synthesis, we follow previous methods [12,18,46], and use standard evaluation metrics, including Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM) [45] and Learned Perceptual Image Patch Similarity (LPIPS) [56]. For pose evaluation, We use standard visual odometry metrics [16, 37, 57], including the Absolute Trajectory Error (ATE) and Relative Pose Error (RPE). ATE measures the difference between the estimated camera positions and the ground truth positions. RPE measures the relative pose errors between pairs of images," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 306, + 547, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 306, + 547, + 415 + ], + "spans": [ + { + "bbox": [ + 304, + 306, + 547, + 415 + ], + "type": "text", + "content": "which consists of relative rotation error " + }, + { + "bbox": [ + 304, + 306, + 547, + 415 + ], + "type": "inline_equation", + "content": "(\\mathrm{RPE}_r)" + }, + { + "bbox": [ + 304, + 306, + 547, + 415 + ], + "type": "text", + "content": " and relative translation error " + }, + { + "bbox": [ + 304, + 306, + 547, + 415 + ], + "type": "inline_equation", + "content": "(\\mathrm{RPE}_t)" + }, + { + "bbox": [ + 304, + 306, + 547, + 415 + ], + "type": "text", + "content": ". The estimated trajectory is aligned with the ground truth using Sim(3) with 7 degrees of freedom. We use standard depth metrics [8,19,20,39] (Abs Rel, Sq Rel, RMSE, RMSE log, " + }, + { + "bbox": [ + 304, + 306, + 547, + 415 + ], + "type": "inline_equation", + "content": "\\delta_1" + }, + { + "bbox": [ + 304, + 306, + 547, + 415 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 306, + 547, + 415 + ], + "type": "inline_equation", + "content": "\\delta_2" + }, + { + "bbox": [ + 304, + 306, + 547, + 415 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 306, + 547, + 415 + ], + "type": "inline_equation", + "content": "\\delta_3" + }, + { + "bbox": [ + 304, + 306, + 547, + 415 + ], + "type": "text", + "content": ") for depth evaluation. For further detail, please refer to the supplementary material. To recover the metric scale, we follow Zhou et al. [58] and match the median value between rendered and ground truth depth maps." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 433, + 547, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 433, + 547, + 649 + ], + "spans": [ + { + "bbox": [ + 304, + 433, + 547, + 649 + ], + "type": "text", + "content": "Implementation Details. Our model architecture is based on NeRF [24] with a few modifications: a) replacing ReLU activation function with Softplus and b) sampling 128 points along each ray uniformly with noise, between a predefined range (0.1, 10). We use 2 separate Adam optimisers [14] for NeRF and other parameters. The initial learning rate for NeRF is 0.001 and for the pose and distortion is 0.0005. Camera rotations are optimised in axis-angle representation " + }, + { + "bbox": [ + 304, + 433, + 547, + 649 + ], + "type": "inline_equation", + "content": "\\phi_i \\in \\mathfrak{so}(3)" + }, + { + "bbox": [ + 304, + 433, + 547, + 649 + ], + "type": "text", + "content": ". We first train the model with all losses with constant learning rates until the inter-frame losses converge. Then, we remove the inter-frame losses and depth loss to refine the model with the RGB loss only. We decay the learning rates with different schedulers to refine for 10,000 epochs. We balance the loss terms with " + }, + { + "bbox": [ + 304, + 433, + 547, + 649 + ], + "type": "inline_equation", + "content": "\\lambda_1 = 0.04" + }, + { + "bbox": [ + 304, + 433, + 547, + 649 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 433, + 547, + 649 + ], + "type": "inline_equation", + "content": "\\lambda_2 = 1.0" + }, + { + "bbox": [ + 304, + 433, + 547, + 649 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 433, + 547, + 649 + ], + "type": "inline_equation", + "content": "\\lambda_3 = 1.0" + }, + { + "bbox": [ + 304, + 433, + 547, + 649 + ], + "type": "text", + "content": ". For each training step, we randomly sample 1024 pixels (rays) from each input image and 128 samples per ray. More details are provided in the supplementary material." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 658, + 525, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 658, + 525, + 671 + ], + "spans": [ + { + "bbox": [ + 305, + 658, + 525, + 671 + ], + "type": "text", + "content": "4.2. Comparing With Pose-Unknown Methods" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 677, + 545, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 700 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 700 + ], + "type": "text", + "content": "We compare our method with pose-unknown baselines, including BARF [18], NeRFmm [46] and SC-NeRF [12]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 318, + 701, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 701, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 318, + 701, + 546, + 714 + ], + "type": "text", + "content": "View Synthesis Quality. To obtain the camera poses of" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4164" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 77, + 75, + 521, + 293 + ], + "blocks": [ + { + "bbox": [ + 77, + 75, + 521, + 293 + ], + "lines": [ + { + "bbox": [ + 77, + 75, + 521, + 293 + ], + "spans": [ + { + "bbox": [ + 77, + 75, + 521, + 293 + ], + "type": "image", + "image_path": "8a961b15ff878b2b3cffb7e326c300c3b156cd89ca5cdda08b5d72724dfc57d8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 305, + 546, + 328 + ], + "lines": [ + { + "bbox": [ + 46, + 305, + 546, + 328 + ], + "spans": [ + { + "bbox": [ + 46, + 305, + 546, + 328 + ], + "type": "text", + "content": "Figure 4. Pose Estimation Comparison. We visualise the trajectory (3D plot) and relative rotation errors " + }, + { + "bbox": [ + 46, + 305, + 546, + 328 + ], + "type": "inline_equation", + "content": "\\mathrm{RPE}_r" + }, + { + "bbox": [ + 46, + 305, + 546, + 328 + ], + "type": "text", + "content": " (bottom colour bar) of each method on Ballroom and Museum. The colour bar on the right shows the relative scaling of colour. More results are in the supplementary." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 48, + 337, + 560, + 494 + ], + "blocks": [ + { + "bbox": [ + 48, + 337, + 560, + 494 + ], + "lines": [ + { + "bbox": [ + 48, + 337, + 560, + 494 + ], + "spans": [ + { + "bbox": [ + 48, + 337, + 560, + 494 + ], + "type": "table", + "html": "
scenesOursBARFNeRFmmSC-NeRF
PSNR ↑SSIM ↑LPIPS ↓PSNRSSIMLPIPSPSNRSSIMLPIPSPSNRSSIMLPIPS
ScanNet0079_0032.470.840.4132.310.830.4330.590.810.4931.330.820.46
0418_0031.330.790.3431.240.790.3530.000.770.4029.050.750.43
0301_0029.830.770.3629.310.760.3827.840.720.4529.450.770.35
0431_0033.830.910.3932.770.900.4131.440.880.4532.570.900.40
mean31.860.830.3831.410.820.3929.970.800.4530.600.810.41
Tanks and TemplesChurch25.170.730.3923.170.620.5221.640.580.5421.960.600.53
Barn26.350.690.4425.280.640.4823.210.610.5323.260.620.51
Museum26.770.760.3523.580.610.5522.370.610.5324.940.690.45
Family26.010.740.4123.040.610.5623.040.580.5622.600.630.51
Horse27.640.840.2624.090.720.4123.120.700.4325.230.760.37
Ballroom25.330.720.3820.660.500.6020.030.480.5722.640.610.48
Francis29.480.800.3825.850.690.5725.4000.690.5226.460.730.49
Ignatius23.960.610.4721.780.470.6021.160.450.6023.000.550.53
mean26.340.740.3923.420.610.5422.500.590.5423.760.650.48
", + "image_path": "f6ae0dbcd5d51550dca1fd54288d435fc59b3c5d583838e580665d4fcdab18f3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 502, + 546, + 525 + ], + "lines": [ + { + "bbox": [ + 46, + 502, + 546, + 525 + ], + "spans": [ + { + "bbox": [ + 46, + 502, + 546, + 525 + ], + "type": "text", + "content": "Table 1. Novel view synthesis results on ScanNet and Tanks and Temples. Each baseline method is trained with its public code under the original settings and evaluated with the same evaluation protocol." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 545, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 287, + 665 + ], + "type": "text", + "content": "test views for rendering, we minimise the photometric error of the synthesised images while keeping the NeRF model fixed, as in NeRFmm [46]. Each test pose is initialised with the learned pose of the training frame that is closest to it. We use the same pre-processing for all baseline approaches, which results in higher accuracy than their original implementations. More details are provided in the supplementary material. Our method outperforms all the baselines by a large margin. The quantitative results are summarised in Tab. 1, and qualitative results are shown in Fig. 3." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": "We recognised that because the test views, which are sampled from videos, are close to the training views, good results may be obtained due to overfitting to the training images. Therefore, we conduct an additional qualitative" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 545, + 545, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 629 + ], + "type": "text", + "content": "evaluation on more novel views. Specifically, we fit aBezier curve from the estimated training poses and sample interpolated poses for each method to render novel view videos. Sampled results are shown in Fig. 5, and the rendered videos are in the supplementary material. These results show that our method renders photo-realistic images consistently, while other methods generate visible artifacts." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 630, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 546, + 713 + ], + "type": "text", + "content": "Camera Pose. Our method significantly outperforms other baselines in all metrics. The quantitative pose evaluation results are shown in Tab. 2. For ScanNet, we use the camera poses provided by the dataset as ground truth. For Tanks and Temples, not every video comes with ground truth poses, so we use COLMAP estimations for reference. Our estimated trajectory is better aligned with the ground" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4165" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 67, + 70, + 526, + 201 + ], + "blocks": [ + { + "bbox": [ + 67, + 70, + 526, + 201 + ], + "lines": [ + { + "bbox": [ + 67, + 70, + 526, + 201 + ], + "spans": [ + { + "bbox": [ + 67, + 70, + 526, + 201 + ], + "type": "table", + "html": "
scenesOursBARFNeRFmmSC-NeRF
RPEt ↓RPEr ↓ATE↓RPEtRPErATERPEtRPErATERPEtRPErATE
ScanNet0079_000.7520.2040.0231.1100.4800.0621.7060.6360.1002.0640.6640.115
0418_000.4550.1190.0151.3980.5380.0201.4020.4600.0131.5280.5020.016
0301_000.3990.1230.0131.3160.7770.2193.0970.8940.2881.1330.4220.056
0431_001.6250.2740.0696.0240.7540.1686.7990.6240.4964.1100.4990.205
mean0.8080.1800.0302.4620.6370.1173.2510.6540.2242.2090.5220.098
Tanks and TemplesChurch0.0340.0080.0080.1140.0380.0520.6260.1270.0650.8360.1870.108
Barn0.0460.0320.0040.3140.2650.0501.6290.4940.1591.3170.4290.157
Museum0.2070.2020.0203.4421.1280.2634.1341.0510.3468.3391.4910.316
Family0.0470.0150.0011.3710.5910.1152.7430.5370.1201.1710.4990.142
Horse0.1790.0170.0031.3330.3940.0141.3490.4340.0181.3660.4380.019
Ballroom0.0410.0180.0020.5310.2280.0180.4490.1770.0310.3280.1460.012
Francis0.0570.0090.0051.3210.5580.0821.6470.6180.2071.2330.4830.192
Ignatius0.0260.0050.0020.7360.3240.0291.3020.3790.0410.5330.2400.085
mean0.0800.0380.0061.0460.4410.0781.7350.4770.1231.8900.4890.129
", + "image_path": "c48842464f4f2980bb416893656b67e390a10b13e5418b50cf66c598c2bf76f9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 49, + 251, + 285, + 310 + ], + "blocks": [ + { + "bbox": [ + 46, + 209, + 546, + 232 + ], + "lines": [ + { + "bbox": [ + 46, + 209, + 546, + 232 + ], + "spans": [ + { + "bbox": [ + 46, + 209, + 546, + 232 + ], + "type": "text", + "content": "Table 2. Pose accuracy on ScanNet and Tanks and Temples. Note that we use COLMAP poses in Tanks and Temples as the \"ground truth\". The unit of " + }, + { + "bbox": [ + 46, + 209, + 546, + 232 + ], + "type": "inline_equation", + "content": "\\mathrm{RPE}_r" + }, + { + "bbox": [ + 46, + 209, + 546, + 232 + ], + "type": "text", + "content": " is in degrees, ATE is in the ground truth scale and " + }, + { + "bbox": [ + 46, + 209, + 546, + 232 + ], + "type": "inline_equation", + "content": "\\mathrm{RPE}_t" + }, + { + "bbox": [ + 46, + 209, + 546, + 232 + ], + "type": "text", + "content": " is scaled by 100." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 251, + 285, + 310 + ], + "lines": [ + { + "bbox": [ + 49, + 251, + 285, + 310 + ], + "spans": [ + { + "bbox": [ + 49, + 251, + 285, + 310 + ], + "type": "table", + "html": "
Abs Rel ↓Sq Rel ↓RMSE ↓RMSE log ↓δ1 ↑δ2 ↑δ3 ↑
Ours0.1410.1370.5680.1760.8280.9700.987
BARF0.3760.6840.9900.4010.4900.7510.884
NeRFmm0.5901.7211.6720.5870.3160.5600.743
SC-NeRF0.4170.6421.0790.4760.3620.6580.832
DPT0.1970.2460.7510.2260.7470.9340.975
", + "image_path": "ef935c6effe7ceb41bb555f6a0a57c110a1fe32d45cb395b995c64f995c42920.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 319, + 287, + 363 + ], + "lines": [ + { + "bbox": [ + 46, + 319, + 287, + 363 + ], + "spans": [ + { + "bbox": [ + 46, + 319, + 287, + 363 + ], + "type": "text", + "content": "Table 3. Depth map evaluation on ScanNet. Our depth estimation is more accurate than baseline models BARF [18], NeRFmm [46] and SC-NeRF [12]. Compared with DPT [58], we show our depth is more accurate after undistortion." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 385, + 287, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 385, + 287, + 421 + ], + "spans": [ + { + "bbox": [ + 46, + 385, + 287, + 421 + ], + "type": "text", + "content": "truth than other methods, and our estimated rotation is two orders of magnitudes more accurate than others. We visualise the camera trajectories and rotations in Fig. 4." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 422, + 287, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 422, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 422, + 287, + 529 + ], + "type": "text", + "content": "Depth. We evaluate the accuracy of the rendered depth maps on ScanNet, which provides the ground-truth depths for evaluation. Our rendered depth maps achieve superior accuracy over the previous alternatives. We also compare with the mono-depth maps estimated by DPT. Our rendered depth maps, after undistortion using multiview consistency in the NeRF optimisation, outperform DPT by a large margin. The results are summarised in Tab. 3, and sampled qualitative results are illustrated in Fig. 3." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 538, + 272, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 538, + 272, + 552 + ], + "spans": [ + { + "bbox": [ + 47, + 538, + 272, + 552 + ], + "type": "text", + "content": "4.3. Comparing With COLMAP Assisted NeRF" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "content": "We make a comparison of pose estimation accuracy between our method and COLMAP against ground truth poses in ScanNet. We achieve on-par accuracy with COLMAP, as shown in Tab. 4. We further analyse the novel view synthesis quality of the NeRF model trained with our learned poses to COLMAP poses on ScanNet and Tanks and Temples. The original NeRF training contains two stages, finding poses using COLMAP and optimising the scene representation. In order to make our comparison fairer, in this section only, we mimic a similar two-stage training as the original NeRF [24]. In the first stage, we train our method with all losses for camera pose estimation, i.e., mimicking the COLMAP processing. Then, we fix the optimised poses" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 310, + 254, + 545, + 432 + ], + "blocks": [ + { + "bbox": [ + 310, + 254, + 545, + 432 + ], + "lines": [ + { + "bbox": [ + 310, + 254, + 545, + 432 + ], + "spans": [ + { + "bbox": [ + 310, + 254, + 545, + 432 + ], + "type": "image", + "image_path": "3cbccf2b854b65f90026402f87da8bcfadee477424ca19c84ddcaefc04ba8a80.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 445, + 545, + 511 + ], + "lines": [ + { + "bbox": [ + 304, + 445, + 545, + 511 + ], + "spans": [ + { + "bbox": [ + 304, + 445, + 545, + 511 + ], + "type": "text", + "content": "Figure 5. Sampled frames from rendered novel view videos. For each method, we fit the learned trajectory with aBezier curve and uniformly sample new viewpoints for rendering. Our method generates significantly better results than previous methods, which show visible artifacts. The full rendered videos and details about generating novel views are provided in the supplementary." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 533, + 545, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 545, + 582 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 545, + 582 + ], + "type": "text", + "content": "and train a NeRF model from scratch, using the same settings and loss as the original NeRF. This evaluation enables us to compare our estimated poses to the COLMAP poses indirectly, i.e., in terms of contribution to view synthesis." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 582, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 628 + ], + "type": "text", + "content": "Our two-stage method outperforms the COLMAP-assisted NeRF baseline, which indicates a better pose estimation for novel view synthesis. The results are summarised in Tab. 5." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "As is commonly known, COLMAP performs poorly in low-texture scenes and sometimes fails to find accurate camera poses. Fig. 6 shows an example of a low-texture scene where COLMAP provides inaccurate pose estimation that causes NeRF to render images with visible artifacts. In contrast, our method renders high-quality images, thanks to robust optimisation of camera pose." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "4166" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 70, + 285, + 140 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 285, + 140 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 285, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 285, + 140 + ], + "type": "table", + "html": "
scenesOursCOLMAP
RPEt↓RPEr↓ATE↓RPEtRPErATE
0079_000.7520.2040.0230.6550.2210.012
0418_000.4550.1190.0150.4910.1240.016
0301_000.3990.1230.0130.4140.1360.009
0431_001.6250.2740.0691.2920.2490.051
mean0.8080.1800.0300.7130.1820.022
", + "image_path": "397900dd0cf7a114fc0a3c7befe428acbef60889cde94ee3a1e5a4279ed4cb8a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 49, + 183, + 126, + 243 + ], + "blocks": [ + { + "bbox": [ + 49, + 183, + 126, + 243 + ], + "lines": [ + { + "bbox": [ + 49, + 183, + 126, + 243 + ], + "spans": [ + { + "bbox": [ + 49, + 183, + 126, + 243 + ], + "type": "image", + "image_path": "ab5ff0ea66cc70c7dccbe02caef55a0d10dcf96e0f561a4273fb78d68fbec191.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 56, + 245, + 121, + 254 + ], + "lines": [ + { + "bbox": [ + 56, + 245, + 121, + 254 + ], + "spans": [ + { + "bbox": [ + 56, + 245, + 121, + 254 + ], + "type": "text", + "content": "(a) COLMAP+NeRF" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 127, + 183, + 206, + 243 + ], + "blocks": [ + { + "bbox": [ + 127, + 183, + 206, + 243 + ], + "lines": [ + { + "bbox": [ + 127, + 183, + 206, + 243 + ], + "spans": [ + { + "bbox": [ + 127, + 183, + 206, + 243 + ], + "type": "image", + "image_path": "3351115fc0353aa326c5261a0a09a87858d4cf14f3c8461db588f41d0fb2a4da.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 154, + 245, + 181, + 254 + ], + "lines": [ + { + "bbox": [ + 154, + 245, + 181, + 254 + ], + "spans": [ + { + "bbox": [ + 154, + 245, + 181, + 254 + ], + "type": "text", + "content": "(b) Ours" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 206, + 183, + 285, + 243 + ], + "blocks": [ + { + "bbox": [ + 206, + 183, + 285, + 243 + ], + "lines": [ + { + "bbox": [ + 206, + 183, + 285, + 243 + ], + "spans": [ + { + "bbox": [ + 206, + 183, + 285, + 243 + ], + "type": "image", + "image_path": "c9121ae35a3206e6ce21c78be3065f214a6c836303976974215795cbce33c325.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 219, + 245, + 272, + 254 + ], + "lines": [ + { + "bbox": [ + 219, + 245, + 272, + 254 + ], + "spans": [ + { + "bbox": [ + 219, + 245, + 272, + 254 + ], + "type": "text", + "content": "(c) Ground Truth" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 269, + 287, + 301 + ], + "lines": [ + { + "bbox": [ + 46, + 269, + 287, + 301 + ], + "spans": [ + { + "bbox": [ + 46, + 269, + 287, + 301 + ], + "type": "text", + "content": "Figure 6. COLMAP failure case. On a rotation-dominant sequence with low-texture areas, COLMAP fails to estimate correct poses, which results in artifacts in synthesised images." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 48, + 312, + 285, + 445 + ], + "blocks": [ + { + "bbox": [ + 46, + 148, + 287, + 170 + ], + "lines": [ + { + "bbox": [ + 46, + 148, + 287, + 170 + ], + "spans": [ + { + "bbox": [ + 46, + 148, + 287, + 170 + ], + "type": "text", + "content": "Table 4. Comparison of pose accuracy with COLMAP on ScanNet." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 312, + 285, + 445 + ], + "lines": [ + { + "bbox": [ + 48, + 312, + 285, + 445 + ], + "spans": [ + { + "bbox": [ + 48, + 312, + 285, + 445 + ], + "type": "table", + "html": "
scenesOursOurs-rCOLMAP+NeRF
PSNR ↑SSIM ↑LPIPS ↓PSNRSSIMLPIPSPSNRSSIMLPIPS
ScanNet0079_0032.470.840.4133.120.850.4031.980.830.43
0418_0031.330.790.3430.490.770.4030.600.780.40
0301_0029.830.770.3630.050.780.3430.010.780.36
0431_0033.830.910.3933.860.910.3933.540.910.39
mean31.860.830.3831.880.830.3831.530.820.40
Tanks and TemplesChurch25.170.730.3926.740.780.3225.720.750.37
Barn26.350.690.4426.580.710.4226.720.710.42
Museum26.770.760.3526.980.770.3627.210.780.34
Family26.010.740.4126.210.750.4026.610.770.39
Horse27.640.840.2628.060.840.2627.020.820.29
Ballroom25.330.720.3825.530.730.3825.470.730.38
Francis29.480.800.3829.730.810.3830.050.810.38
Ignatius23.960.610.4723.980.620.4624.080.610.47
mean26.340.740.3926.730.750.3726.610.750.38
", + "image_path": "4ad5d338d02e42e52679909d810d7212f450ca61d5a57066ea1e0e49dfa60b47.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 508, + 286, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 508, + 286, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 508, + 286, + 616 + ], + "type": "text", + "content": "Interestingly, this experiment also reveals that the two-stage method shows higher accuracy than the one-stage method. We hypothesise that the joint optimisation (from randomly initialised poses) in the one-stage approach causes the NeRF optimisation to be trapped in a local minimum, potentially due to the bad pose initialisation. The two-stage approach circumvents this issue by re-initialising the NeRF and re-training with well-optimised poses, resulting in higher performance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 623, + 140, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 623, + 140, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 623, + 140, + 635 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 642, + 286, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 286, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 286, + 677 + ], + "type": "text", + "content": "In this section, we analyse the effectiveness of the parameters and components that have been added to our model. The results of ablation studies are shown in Tab. 6." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "Effect of Distortion Parameters. We find that ignoring depth distortions (i.e., setting scales to 1 and shifts to 0 as constants) leads to a degradation in pose accuracy, as in" + } + ] + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 306, + 70, + 544, + 143 + ], + "blocks": [ + { + "bbox": [ + 46, + 453, + 287, + 486 + ], + "lines": [ + { + "bbox": [ + 46, + 453, + 287, + 486 + ], + "spans": [ + { + "bbox": [ + 46, + 453, + 287, + 486 + ], + "type": "text", + "content": "Table 5. Comparison to NeRF with COLMAP poses. Our two-stage method (Ours-r) outperforms both COLMAP+NeRF and our one-stage method (Ours)." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 306, + 70, + 544, + 143 + ], + "lines": [ + { + "bbox": [ + 306, + 70, + 544, + 143 + ], + "spans": [ + { + "bbox": [ + 306, + 70, + 544, + 143 + ], + "type": "table", + "html": "
NVSPose
PSNR↑SSIM↑LPIPS↓RPEt↓RPEr↓ATE↓
Ours31.860.830.380.8010.1810.031
Ours w/o α,β31.460.820.391.9290.3210.066
Ours w/o Lpc31.730.820.382.2270.4530.101
Ours w/o Lrgb-s31.050.810.411.8140.4010.156
Ours w/o Ldepth31.200.810.401.4980.3830.089
", + "image_path": "6ff287b7cffc112f520a5df4b3313677c5479c77e92c52122dbd44fc8a442f28.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 342, + 152, + 509, + 163 + ], + "lines": [ + { + "bbox": [ + 342, + 152, + 509, + 163 + ], + "spans": [ + { + "bbox": [ + 342, + 152, + 509, + 163 + ], + "type": "text", + "content": "Table 6. Ablation study results on ScanNet." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 185, + 545, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 185, + 545, + 219 + ], + "spans": [ + { + "bbox": [ + 304, + 185, + 545, + 219 + ], + "type": "text", + "content": "consistent distortions of depth maps introduce errors to the estimation of relative poses and confuse NeRF for geometry reconstruction." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 220, + 545, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 220, + 545, + 291 + ], + "spans": [ + { + "bbox": [ + 304, + 220, + 545, + 291 + ], + "type": "text", + "content": "Effect of Inter-frame Losses. We observe that the interframe losses are the major contributor to improving relative poses. When removing the pairwise point cloud loss " + }, + { + "bbox": [ + 304, + 220, + 545, + 291 + ], + "type": "inline_equation", + "content": "L_{pc}" + }, + { + "bbox": [ + 304, + 220, + 545, + 291 + ], + "type": "text", + "content": " or the surface-based photometric loss " + }, + { + "bbox": [ + 304, + 220, + 545, + 291 + ], + "type": "inline_equation", + "content": "L_{rgb-s}" + }, + { + "bbox": [ + 304, + 220, + 545, + 291 + ], + "type": "text", + "content": ", there is less constraint between frames, and thus the pose accuracy becomes lower." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 293, + 545, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 293, + 545, + 341 + ], + "spans": [ + { + "bbox": [ + 304, + 293, + 545, + 341 + ], + "type": "text", + "content": "Effect of NeRF Losses. When the depth loss " + }, + { + "bbox": [ + 304, + 293, + 545, + 341 + ], + "type": "inline_equation", + "content": "L_{depth}" + }, + { + "bbox": [ + 304, + 293, + 545, + 341 + ], + "type": "text", + "content": " is removed, the distortions of input depth maps are only optimised locally through the inter-frame losses. We find that this can lead to drift and degradation in pose accuracy." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 349, + 383, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 349, + 383, + 361 + ], + "spans": [ + { + "bbox": [ + 306, + 349, + 383, + 361 + ], + "type": "text", + "content": "4.5. Limitations" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 368, + 545, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 368, + 545, + 427 + ], + "spans": [ + { + "bbox": [ + 304, + 368, + 545, + 427 + ], + "type": "text", + "content": "Our proposed method optimises camera pose and the NeRF model jointly and works on challenging scenes where other baselines fail. However, the optimisation of the model is also affected by non-linear distortions and the accuracy of the mono-depth estimation, which we did not consider." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 439, + 378, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 439, + 378, + 451 + ], + "spans": [ + { + "bbox": [ + 306, + 439, + 378, + 451 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 459, + 545, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 459, + 545, + 626 + ], + "spans": [ + { + "bbox": [ + 304, + 459, + 545, + 626 + ], + "type": "text", + "content": "In this work, we present NoPe-NeRF, an end-to-end differentiable model for joint camera pose estimation and novel view synthesis from a sequence of images. We demonstrate that previous approaches have difficulty with complex trajectories. To tackle this challenge, we use mono-depth maps to constrain the relative poses between frames and regularise the geometry of NeRF, which leads to better pose estimation. We show the effectiveness and robustness of NoPe-NeRF on challenging scenes. The improved pose estimation leads to better novel view synthesis quality and geometry reconstruction compared with other approaches. We believe our method is an important step towards applying the unknown-pose NeRF models to large-scale scenes in the future." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 639, + 408, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 639, + 408, + 651 + ], + "spans": [ + { + "bbox": [ + 306, + 639, + 408, + 651 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 657, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 657, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 657, + 545, + 713 + ], + "type": "text", + "content": "We thank Theo Costain, Michael Hobley, Shuai Chen and Xinghui Li for their helpful proofreading and discussions. Wenjing Bian is supported by the China Scholarship Council - University of Oxford Scholarship." + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4167" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 92, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 53, + 92, + 287, + 125 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 92, + 287, + 125 + ], + "spans": [ + { + "bbox": [ + 53, + 92, + 287, + 125 + ], + "type": "text", + "content": "[1] Chris Buehler, Michael Bosse, Leonard McMillan, Steven Gortler, and Michael Cohen. Unstructured lumigraph rendering. In SIGGRAPH, 2001. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 126, + 288, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 126, + 288, + 158 + ], + "spans": [ + { + "bbox": [ + 53, + 126, + 288, + 158 + ], + "type": "text", + "content": "[2] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In ECCV, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 161, + 288, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 161, + 288, + 183 + ], + "spans": [ + { + "bbox": [ + 53, + 161, + 288, + 183 + ], + "type": "text", + "content": "[3] Shenchang Eric Chen and Lance Williams. View interpolation for image synthesis. In SIGGRAPH, 1993. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 185, + 288, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 185, + 288, + 228 + ], + "spans": [ + { + "bbox": [ + 53, + 185, + 288, + 228 + ], + "type": "text", + "content": "[4] Shin-Fang Chng, Sameera Ramasinghe, Jamie Sherrah, and Simon Lucey. Garf: Gaussian activated radiance fields for high fidelity reconstruction and pose estimation. arXiv eprints, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 230, + 288, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 230, + 288, + 274 + ], + "spans": [ + { + "bbox": [ + 53, + 230, + 288, + 274 + ], + "type": "text", + "content": "[5] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, 2017. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 276, + 288, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 276, + 288, + 319 + ], + "spans": [ + { + "bbox": [ + 53, + 276, + 288, + 319 + ], + "type": "text", + "content": "[6] Paul E Debevec, Camillo J Taylor, and Jitendra Malik. Modeling and rendering architecture from photographs: A hybrid geometry-and image-based approach. In SIGGRAPH, 1996. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 321, + 288, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 321, + 288, + 354 + ], + "spans": [ + { + "bbox": [ + 53, + 321, + 288, + 354 + ], + "type": "text", + "content": "[7] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised nerf: Fewer views and faster training for free. In CVPR, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 356, + 288, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 356, + 288, + 388 + ], + "spans": [ + { + "bbox": [ + 53, + 356, + 288, + 388 + ], + "type": "text", + "content": "[8] Huan Fu, Mingming Gong, Chaohui Wang, Kayhan Bat-manghelich, and Dacheng Tao. Deep ordinal regression network for monocular depth estimation. In CVPR, 2018. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 391, + 288, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 391, + 288, + 423 + ], + "spans": [ + { + "bbox": [ + 53, + 391, + 288, + 423 + ], + "type": "text", + "content": "[9] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video. In ICCV, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 426, + 287, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 426, + 287, + 458 + ], + "spans": [ + { + "bbox": [ + 48, + 426, + 287, + 458 + ], + "type": "text", + "content": "[10] Stephan J Garbin, Marek Kowalski, Matthew Johnson, Jamie Shotton, and Julien Valentin. Fastnerf: High-fidelity neural rendering at 200fps. In ICCV, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 460, + 287, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 460, + 287, + 482 + ], + "spans": [ + { + "bbox": [ + 48, + 460, + 287, + 482 + ], + "type": "text", + "content": "[11] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. 2003. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 483, + 287, + 516 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 483, + 287, + 516 + ], + "spans": [ + { + "bbox": [ + 48, + 483, + 287, + 516 + ], + "type": "text", + "content": "[12] Yoonwoo Jeong, Seokjun Ahn, Christopher Choy, Anima Anandkumar, Minsu Cho, and Jaesik Park. Self-calibrating neural radiance fields. In ICCV, 2021. 1, 2, 3, 4, 5, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 517, + 287, + 551 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 517, + 287, + 551 + ], + "spans": [ + { + "bbox": [ + 48, + 517, + 287, + 551 + ], + "type": "text", + "content": "[13] Mijeong Kim, Seonguk Seo, and Bohyung Han. Infonerf: Ray entropy minimization for few-shot neural volume rendering. In CVPR, 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 552, + 287, + 585 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 552, + 287, + 585 + ], + "spans": [ + { + "bbox": [ + 48, + 552, + 287, + 585 + ], + "type": "text", + "content": "[14] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In *Yoshua Bengio and Yann LeCun*, editors, *ICLR*, 2015. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 587, + 287, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 587, + 287, + 620 + ], + "spans": [ + { + "bbox": [ + 48, + 587, + 287, + 620 + ], + "type": "text", + "content": "[15] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics, 2017. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 622, + 287, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 622, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 622, + 287, + 644 + ], + "type": "text", + "content": "[16] Johannes Kopf, Xuejian Rong, and Jia-Bin Huang. Robust consistent video depth estimation. In CVPR, 2021. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 646, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 646, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 646, + 287, + 678 + ], + "type": "text", + "content": "[17] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In CVPR, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[18] Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. Barf: Bundle-adjusting neural radiance fields. In ICCV, 2021. 1, 2, 3, 4, 5, 7" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 37, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[19] Fayao Liu, Chunhua Shen, Guosheng Lin, and Ian Reid. Learning depth from single monocular images using deep convolutional neural fields. TPAMI, 2015. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 107, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 139 + ], + "type": "text", + "content": "[20] Xuan Luo, Jia-Bin Huang, Richard Szeliski, Kevin Matzen, and Johannes Kopf. Consistent video depth estimation. ToG. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 141, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 195 + ], + "type": "text", + "content": "[21] Quan Meng, Anpei Chen, Haimin Luo, Minye Wu, Hao Su, Lan Xu, Xuming He, and Jingyi Yu. Gnerf: Gan-based neural radiance field without posed camera. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 197, + 545, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 240 + ], + "type": "text", + "content": "[22] S Mahdi H Miangoleh, Sebastian Dille, Long Mai, Sylvain Paris, and Yagiz Aksoy. Boosting monocular depth estimation models to high-resolution via content-adaptive multi-resolution merging. In CVPR, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 241, + 545, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 241, + 545, + 296 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 545, + 296 + ], + "type": "text", + "content": "[23] Ben Mildenhall, Pratul P Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines. ACM Transactions on Graphics (TOG), 2019. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 297, + 545, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 297, + 545, + 341 + ], + "spans": [ + { + "bbox": [ + 307, + 297, + 545, + 341 + ], + "type": "text", + "content": "[24] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 2021. 1, 2, 3, 5, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 342, + 545, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 342, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 307, + 342, + 545, + 374 + ], + "type": "text", + "content": "[25] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 376, + 545, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 376, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 307, + 376, + 545, + 420 + ], + "type": "text", + "content": "[26] Michael Niemeyer, Jonathan T Barron, Ben Mildenhall, Mehdi SM Sajjadi, Andreas Geiger, and Noha Radwan. Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs. In CVPR, 2022. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 421, + 545, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 421, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 307, + 421, + 545, + 453 + ], + "type": "text", + "content": "[27] Michael Oechsle, Songyou Peng, and Andreas Geiger. Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In ICCV, 2021. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 455, + 545, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 455, + 545, + 486 + ], + "spans": [ + { + "bbox": [ + 307, + 455, + 545, + 486 + ], + "type": "text", + "content": "[28] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In ICCV, 2021. 2, 4" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 488, + 545, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 488, + 545, + 531 + ], + "spans": [ + { + "bbox": [ + 307, + 488, + 545, + 531 + ], + "type": "text", + "content": "[29] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. TPAMI, 2020. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 533, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 533, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 307, + 533, + 545, + 555 + ], + "type": "text", + "content": "[30] Gernot Riegler and Vladlen Koltun. Free view synthesis. In ECCV, 2020. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 556, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 556, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 307, + 556, + 545, + 578 + ], + "type": "text", + "content": "[31] Gernot Riegler and Vladlen Koltun. Stable view synthesis. In CVPR, 2021. 2" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 579, + 545, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 579, + 545, + 621 + ], + "spans": [ + { + "bbox": [ + 307, + 579, + 545, + 621 + ], + "type": "text", + "content": "[32] Barbara Roessle, Jonathan T Barron, Ben Mildenhall, Pratul P Srinivasan, and Matthias Nießner. Dense depth priors for neural radiance fields from sparse input views. In CVPR, 2022. 2" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 307, + 624, + 545, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 656 + ], + "type": "text", + "content": "[33] Antoni Rosinol, John J Leonard, and Luca Carlone. Nerf-slam: Real-time dense monocular slam with neural radiance fields. arXiv preprint arXiv:2210.13641, 2022. 2" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 307, + 658, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 545, + 689 + ], + "type": "text", + "content": "[34] Sara Fridovich-Keil and Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In CVPR, 2022. 2" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "text", + "content": "[35] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In CVPR, 2016. 1, 2" + } + ] + } + ], + "index": 36 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "4168" + } + ] + } + ], + "index": 38 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[36] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. NeurIPS, 2020. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 287, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 287, + 139 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 287, + 139 + ], + "type": "text", + "content": "[37] Jürgen Sturm, Nikolas Engelhard, Felix Endres, Wolfram Burgard, and Daniel Cremers. A benchmark for the evaluation of rgb-d slam systems. In IROS. IEEE, 2012. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 140, + 287, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 287, + 171 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 287, + 171 + ], + "type": "text", + "content": "[38] Edgar Sucar, Shikun Liu, Joseph Ortiz, and Andrew Davison. iMAP: Implicit mapping and positioning in real-time. In ICCV, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 172, + 287, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 172, + 287, + 216 + ], + "spans": [ + { + "bbox": [ + 48, + 172, + 287, + 216 + ], + "type": "text", + "content": "[39] Libo Sun, Jia-Wang Bian, Huangying Zhan, Wei Yin, Ian Reid, and Chunhua Shen. Sc-depthv3: Robust self-supervised monocular depth estimation for dynamic scenes. arXiv preprint arXiv:2211.03660, 2022. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 217, + 287, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 217, + 287, + 248 + ], + "spans": [ + { + "bbox": [ + 48, + 217, + 287, + 248 + ], + "type": "text", + "content": "[40] Alex Trevithick and Bo Yang. Grf: Learning a general radiance field for 3d representation and rendering. In ICCV, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 249, + 287, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 249, + 287, + 272 + ], + "spans": [ + { + "bbox": [ + 48, + 249, + 287, + 272 + ], + "type": "text", + "content": "[41] Richard Tucker and Noah Snavely. Single-view view synthesis with multiplane images. In CVPR, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 272, + 287, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 272, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 48, + 272, + 287, + 304 + ], + "type": "text", + "content": "[42] Shubham Tulsiani, Richard Tucker, and Noah Snavely. Layer-structured 3d scene inference via view synthesis. In ECCV, 2018. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 304, + 287, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 304, + 287, + 348 + ], + "spans": [ + { + "bbox": [ + 48, + 304, + 287, + 348 + ], + "type": "text", + "content": "[43] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. NeurIPS, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 349, + 287, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 349, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 48, + 349, + 287, + 403 + ], + "type": "text", + "content": "[44] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P Srinivasan, Howard Zhou, Jonathan T Barron, Ricardo Martin-Brualla, Noah Snively, and Thomas Funkhouser. Ibrnet: Learning multi-view image-based rendering. In CVPR, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 403, + 287, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 403, + 287, + 447 + ], + "spans": [ + { + "bbox": [ + 48, + 403, + 287, + 447 + ], + "type": "text", + "content": "[45] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 2004. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 448, + 287, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 448, + 287, + 492 + ], + "spans": [ + { + "bbox": [ + 48, + 448, + 287, + 492 + ], + "type": "text", + "content": "[46] Zirui Wang, Shangzhe Wu, Weidi Xie, Min Chen, and Victor Adrian Prisacariu. NeRF---: Neural radiance fields without known camera parameters. arXiv preprint arXiv:2102.07064, 2021. 1, 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 492, + 287, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 492, + 287, + 535 + ], + "spans": [ + { + "bbox": [ + 48, + 492, + 287, + 535 + ], + "type": "text", + "content": "[47] Yi Wei, Shaohui Liu, Yongming Rao, Wang Zhao, Jiwen Lu, and Jie Zhou. Nerfingmvs: Guided optimization of neural radiance fields for indoor multi-view stereo. In ICCV, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 536, + 287, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 287, + 568 + ], + "type": "text", + "content": "[48] Yitong Xia, Hao Tang, Radu Timofte, and Luc Van Gool. Sinerf: Sinusoidal neural radiance fields for joint pose estimation and scene reconstruction. 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 569, + 287, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 569, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 48, + 569, + 287, + 602 + ], + "type": "text", + "content": "[49] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. In NeurIPS, 2021. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 602, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 287, + 635 + ], + "type": "text", + "content": "[50] Lin Yen-Chen, Pete Florence, Jonathan T Barron, Alberto Rodriguez, Phillip Isola, and Tsung-Yi Lin. inerf: Inverting neural radiance fields for pose estimation. In IROS, 2021. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 636, + 287, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 287, + 679 + ], + "type": "text", + "content": "[51] Wei Yin, Jianming Zhang, Oliver Wang, Simon Niklaus, Simon Chen, Yifan Liu, and Chunhua Shen. Towards accurate reconstruction of 3d scene shape from a single monocular image. TPAMI, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 712 + ], + "type": "text", + "content": "[52] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In CVPR, 2021. 2" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 545, + 386 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 307, + 72, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 545, + 116 + ], + "type": "text", + "content": "[53] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. NeurIPS, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 118, + 545, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 171 + ], + "type": "text", + "content": "[54] Jian Zhang, Yuanqing Zhang, Huan Fu, Xiaowei Zhou, Bowen Cai, Jinchi Huang, Rongfei Jia, Binqiang Zhao, and Xing Tang. Ray priors through reprojection: Improving neural radiance fields for novel view extrapolation. In CVPR, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 172, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 172, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 172, + 545, + 206 + ], + "type": "text", + "content": "[55] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492, 2020. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 207, + 545, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 207, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 307, + 207, + 545, + 240 + ], + "type": "text", + "content": "[56] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 241, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 241, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 545, + 274 + ], + "type": "text", + "content": "[57] Zichao Zhang and Davide Scaramuzza. A tutorial on quantitative trajectory evaluation for visual (-inertial) odometry. In IROS. IEEE, 2018. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 274, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 274, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 274, + 545, + 308 + ], + "type": "text", + "content": "[58] Tinghui Zhou, Matthew Brown, Noah Snavely, and David G Lowe. Unsupervised learning of depth and ego-motion from video. In CVPR, 2017. 5, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 308, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 308, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 307, + 308, + 545, + 342 + ], + "type": "text", + "content": "[59] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images. 2018. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 342, + 545, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 342, + 545, + 386 + ], + "spans": [ + { + "bbox": [ + 307, + 342, + 545, + 386 + ], + "type": "text", + "content": "[60] Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hujun Bao, Zhaopeng Cui, Martin R. Oswald, and Marc Pollefeys. Nice-slam: Neural implicit scalable encoding for slam. In CVPR, 2022. 2" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "4169" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Noisy Correspondence Learning With Meta Similarity Correction/0167b343-6cc5-4960-bab8-e11934c107f0_content_list.json b/2023/Noisy Correspondence Learning With Meta Similarity Correction/0167b343-6cc5-4960-bab8-e11934c107f0_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6a42a84c106f107673526e8f7a850cdba61527d0 --- /dev/null +++ b/2023/Noisy Correspondence Learning With Meta Similarity Correction/0167b343-6cc5-4960-bab8-e11934c107f0_content_list.json @@ -0,0 +1,1604 @@ +[ + { + "type": "text", + "text": "Noisy Correspondence Learning with Meta Similarity Correction", + "text_level": 1, + "bbox": [ + 155, + 130, + 815, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Haochen Han $^{1}$ , Kaiyao Miao $^{2}$ , Qinghua Zheng $^{1}$ , Minnan Luo $^{1*}$", + "bbox": [ + 236, + 180, + 738, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ School of Computer Science and Technology, Xi'an Jiaotong University, China", + "bbox": [ + 217, + 199, + 750, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ School of Cyber Science and Engineering, Xi'an Jiaotong University, China", + "bbox": [ + 227, + 217, + 740, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{hhc1997, miaoky814}@stu.xjtu.edu.cn, {qhzheng, minnluo}@xjtu.edu.cn", + "bbox": [ + 186, + 236, + 782, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 286, + 313, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite the success of multimodal learning in cross-modal retrieval task, the remarkable progress relies on the correct correspondence among multimedia data. However, collecting such ideal data is expensive and time-consuming. In practice, most widely used datasets are harvested from the Internet and inevitably contain mismatched pairs. Training on such noisy correspondence datasets causes performance degradation because the cross-modal retrieval methods can wrongly enforce the mismatched data to be similar. To tackle this problem, we propose a Meta Similarity Correction Network (MSCN) to provide reliable similarity scores. We view a binary classification task as the meta-process that encourages the MSCN to learn discrimination from positive and negative meta-data. To further alleviate the influence of noise, we design an effective data purification strategy using meta-data as prior knowledge to remove the noisy samples. Extensive experiments are conducted to demonstrate the strengths of our method in both synthetic and real-world noises, including Flickr30K, MS-COCO, and Conceptual Captions. Our code is publicly available. $^{1}$", + "bbox": [ + 75, + 318, + 473, + 636 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 667, + 209, + 684 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, cross-modal retrieval has drawn much attention with the rapid growth of multimedia data. Given a query sample of specific modality, cross-modal retrieval aims to retrieve relevant samples across different modalities. Existing cross-modal retrieval works [1, 6, 11] usually learn a comparable common space to bridge different modalities, which achieved remarkable progress in many applications, including video-audio retrieval [1, 33], visual question answering [11, 26], and image-text matching [44].", + "bbox": [ + 75, + 694, + 468, + 830 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite the promise, a core assumption in cross-modal retrieval is the correct correspondence among multiple", + "bbox": [ + 76, + 830, + 468, + 862 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6f5725cb576e78f8305df691f3e3a92564c58cd94e044e587fff66d835979a4a.jpg", + "image_caption": [ + "Figure 1. Illustration of noisy correspondence in image-text retrieval. A standard triplet loss is used to enforce the positive pairs to be closer than negatives in the common space. Noisy correspondence is the mismatched pairs but wrongly considered as positive ones, and thus results in model performance degradation." + ], + "image_footnote": [], + "bbox": [ + 501, + 287, + 890, + 402 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "modalities. However, collecting such ideal data is expensive and time-consuming. In practice, most widely used datasets are harvested from the Internet and inevitably contain noisy correspondence [31]. As illustrated in Fig. 1, the cross-modal retrieval method will wrongly enforce the mismatched data to be similar when learning with noisy correspondence, which may significantly affect the retrieval performance. To date, few effort has been made to address this. Huang [15] first researches this issue and proposes the NCR method to train from the noisy image-text pairs robustly. Inspired by the prior success for noisy labels [20], NCR divides the data into clean and noisy partitions and rectifies the correspondence with an adaptive model. However, NCR based on the memorization effect of DNNs [3], which leads to poor performance under high noise ratio.", + "bbox": [ + 496, + 506, + 892, + 733 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To tackle the challenge, we propose a Meta Similarity Correction Network (MSCN) which aims to provide reliable similarity scores for the noisy features from main net. We view a binary classification task as the meta-process: given a multimodal sample, the MSCN will learn to determine whether the modalities correspond to each other or not, where the prediction of MSCN can be naturally regarded as the similarity score. Specifically, a small amount of clean pairs is used to construct positive and negative meta-data, both viewed as meta-knowledge that encourages MSCN to learn the discrimination. Meanwhile, the main net", + "bbox": [ + 496, + 734, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Minnan Luo is the corresponding author.", + "bbox": [ + 94, + 875, + 346, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1https://github.com/hhc1997/MSCN", + "bbox": [ + 96, + 887, + 346, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7517", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "trained by the corrected similarity score from MSCN with a self-adaptive margin to achieve robust learning. This interdependency makes the main net and MSCN benefits each other against noise. However, due to the property of triplet loss, it remains to produce positive loss for noisy pairs even if we employ the ideal similarity scores. To this end, we further propose a meta-knowledge guided data purification strategy to remove samples with potentially wrong correspondence. Extensive experiments are conducted to demonstrate the strengths of our method in both synthetic and real-world noises.", + "bbox": [ + 75, + 90, + 472, + 256 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The main contributions of this work are summarized as follows:", + "bbox": [ + 76, + 258, + 470, + 286 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We pioneer the exploration of meta-learning for noisy correspondence problem, where a meta correction network is proposed to provide reliable similarity scores against noise.", + "- We present a novel meta-process that first considers both positive and negative data as meta-knowledge to encourage the MSCN to learn discrimination from them.", + "- We design an effective data purification strategy using meta-data as prior knowledge to remove the noisy samples." + ], + "bbox": [ + 96, + 289, + 467, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 470, + 218, + 487 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Cross-Modal Retrieval", + "text_level": 1, + "bbox": [ + 76, + 497, + 287, + 512 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Cross-modal retrieval works aim to take one modality of data as query to retrieve the related data in other modalities. Based on the utilization of annotation information, cross-modal retrieval methods can be roughly divided into two categories: 1) Unsupervised Methods [8,22,25,36]. It only uses the co-occurrence information to learn the common representations among the multimodal data. 2) Supervised methods [14,34,38,45]. The extra label information is employed to boost the discrimination of common representations. However, the lack of annotations limited the practicability in the real world.", + "bbox": [ + 75, + 522, + 468, + 686 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our proposed approach falls in the category of unsupervised methods. For example, Wang [36] uses a two-branch neural networks to learn the joint embeddings of multimodal data. Inspired by hard negative mining, $\\mathrm{VSE}++$ [8] uses hard negatives to improve the retrieval performance. To capture fine-grained interplay between modalities, SCAN [18] introduce stacked cross attention to enable attention with context from both image and sentence. Recently, motivated by the powerful learning ability of graph model, GSMN [25] and SGRAF [7] construct graph structure for multimodal data to benefit the learning of fine-grained correspondence. Although existing works have achieved remarkable results, they usually depend on the correct correspondence among multimodal data and cannot", + "bbox": [ + 75, + 688, + 470, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tackle the noise issue. Thus, it is significant to explore how to learn cross-modal retrieval with noisy correspondence, but which is rarely touched in previous studies.", + "bbox": [ + 496, + 90, + 890, + 137 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Noisy Correspondence Learning", + "text_level": 1, + "bbox": [ + 498, + 146, + 784, + 162 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As a newly proposed problem, noisy correspondence is the mismatched pairs but wrongly considered as positive ones. Huang [15] first research this issue and propose the NCR method to robustly train the image-text matching model with noisy correspondence pairs. Inspired by the prior success for noisy labels [20], NCR divides the data into clean and noisy part based on the memorization effect of DNNs, and then rectifies the correspondence with an adaptive model. Qin [29] proposes a uncertainty-based method to achieve efficient learning. Moreover, some works [40, 41] study the partially view-unaligned problem which can be viewed as a generalized noisy correspondence learning. Recently, some works [13, 39] explore a more challenging scenario where noisy labels and noisy correspondence may occur simultaneously. In contrast, our work tackles this challenge in cross-modal retrieval from a meta-learning perspective.", + "bbox": [ + 496, + 169, + 890, + 426 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.3. Meta-Learning", + "text_level": 1, + "bbox": [ + 500, + 436, + 653, + 452 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The objective of meta-learning is to learn at a more advanced level than conventional learning, such as learning the update rule [28], finding easily fine-tuned parameters [9], or adapting to new tasks [19]. Recently, researchers use meta-learning to find model's parameters that robust against noisy labels [21,32,46]. For example, Li [21] optimizes a meta-objective before conventional training to enable the model not overfit to noise. MW-Net [32] use a meta-process to automatically assign weights to the training samples. MLC [46] presents a label correction network trained as a meta manner which generate reliable labels for the noisy training data.", + "bbox": [ + 496, + 460, + 890, + 641 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The most relevant existing method is MLC [46], which first poses the noisy label problem as a meta label correction. However, we argue that our proposed method differs from MLC in two aspects. First, we focuses on the noisy correspondence problem in multimodal data instead of the unimodal classification scenarios. Second, we leverage both positive and negative meta-data to guide model learning the discrimination, which is unexplored in previous meta-learning methods.", + "bbox": [ + 496, + 642, + 890, + 777 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. The Proposed Method", + "text_level": 1, + "bbox": [ + 500, + 791, + 710, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Problem Formulation", + "text_level": 1, + "bbox": [ + 500, + 816, + 702, + 830 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Without losing generality, we use the image-text retrieval task as a proxy to investigate the noisy correspondence problem in cross-modal retrieval. Given a training set $\\mathcal{D}_{train} = \\{(I_i,T_i,y_i)\\}_{i = 1}^N$ , where $(I_{i},T_{i})$ is the $i$ -th image-", + "bbox": [ + 496, + 839, + 890, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7518", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/7714af4820a40f75b9ed577c4b180ace7b837498f5c705b5a0cff94a4f9377aa.jpg", + "image_caption": [ + "Figure 2. Overview of the proposed meta-process. The positive pairs are taken from the original meta-data which contain corresponding image and text, while negative pairs are extracting from different pairs in training set to construct mismatched pairs. Then the MSCN is learned to determine whether the modalities correspond to each other or not." + ], + "image_footnote": [], + "bbox": [ + 89, + 88, + 893, + 290 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "text pair, $y_{i} \\in \\{1,0\\}$ indicates the pair is positive (matched), i.e. $(I_i, T_i)$ , or negative (mismatched), i.e. $(I_i, T_{j \\neq i})$ , and $N$ is the number of the entire training data. The noisy correspondence in bimodal data is the negative pair but wrongly considered as $y_{i} = 1$ . To address this issue, we propose a meta-learning based method to achieve robust training. The details are delineated next.", + "bbox": [ + 75, + 362, + 472, + 468 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Meta-Learning Based Similarity Correction", + "text_level": 1, + "bbox": [ + 76, + 482, + 449, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The Training Objective. To begin with, we describe the standard projection module to enable cross-modal features to be comparable. Let $f(I;W_f)$ and $g(T;W_g)$ be the modal-specific networks to map the visual and textual modalities into the joint embedding space, respectively. Then we compute the similarity feature between bimodal features by the function", + "bbox": [ + 75, + 507, + 468, + 612 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nS (f (I), g (T); W _ {s}) = \\frac {W _ {s} | f (I) - g (T) | ^ {2}}{\\| W _ {s} | f (I) - g (T) | ^ {2} \\| _ {2}}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 627, + 468, + 662 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $W_{s}$ is a learnable parameter matrix to obtain low-dimensional similarity representation. For notation convenience, we collect $\\{W_{f},W_{g},W_{s}\\}$ into $W$ as the parameters of main net $\\mathcal{F} = (f,g,S)$ and denote similarity feature as $\\mathcal{F}_W(I,T)$ . With the presence of noisy correspondence, the model will inevitably overfit to noise if we use the similarity representation directly. To tackle this problem, we propose a meta similarity correction network (MSCN), regarding process of obtaining similarity score as a meta-process, which takes the noisy similarity representation as input and produces the corrected similarity score. Specifically, our MSCN is formulated as a MLP network $\\mathcal{V}_{\\Theta}(\\cdot)$ with parameters $\\Theta$ . The output $s_i = \\mathcal{V}_{\\Theta}(\\mathcal{F}_W(I_i,T_i))$ is activated with Sigmoid function to represent the similarity score located in [0, 1].", + "bbox": [ + 75, + 672, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For the noisy training pairs, the main net learns noisetolerant features guided by the corrected similarity scores produced from MSCN; while the MSCN predicts the reliable similarity score based on features from main net. This interdependency enables the main net and meta net reinforce each other against noise, and optimized via a bi-level optimization. Formally, the optimal parameters $W^{*}$ is calculated by minimizing the cumulative loss over training data:", + "bbox": [ + 496, + 362, + 893, + 498 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nW ^ {*} (\\Theta) = \\underset {W} {\\arg \\min } \\mathbb {E} _ {(I _ {i}, T _ {i}) \\in \\mathcal {D} _ {t r a i n}} l ^ {t r a i n} (I _ {i}, T _ {i}), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 521, + 890, + 547 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $l^{train}(\\cdot)$ is the triplet ranking loss:", + "bbox": [ + 500, + 555, + 772, + 571 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} l ^ {t r a i n} \\left(I _ {i}, T _ {i}\\right) = [ \\hat {\\gamma} - \\mathcal {V} _ {\\Theta} \\left(\\mathcal {F} _ {W} \\left(I _ {i}, T _ {i}\\right)\\right) + \\mathcal {V} _ {\\Theta} \\left(\\mathcal {F} _ {W} \\left(I _ {i}, T _ {i} ^ {-}\\right)\\right) ] _ {+} \\\\ + \\left[ \\hat {\\gamma} - \\mathcal {V} _ {\\Theta} \\left(\\mathcal {F} _ {W} \\left(I _ {i}, T _ {i}\\right)\\right) + \\mathcal {V} _ {\\Theta} \\left(\\mathcal {F} _ {W} \\left(I _ {i} ^ {-}, T _ {i}\\right)\\right) \\right] _ {+,} \\tag {3} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 501, + 589, + 890, + 637 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $[x]_{+} = max(x,0), I_{i}^{-}$ and $T_{i}^{-}$ are the hardest negatives corresponding to given pair $(I_i,T_i)$ similar to VSE++ [8]. To achieve robust cross-modal retrieval, we design a self-adaptive margin $\\hat{\\gamma}$ instead of a fixed value in VSE++. Ideally, the margin should put less value on the noisy pairs and more on the clean pairs. To this end, we adjust the margin according to the similarity score:", + "bbox": [ + 496, + 638, + 890, + 743 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\gamma} = \\frac {1}{1 + \\left(\\frac {s _ {i}}{1 - s _ {i}}\\right) ^ {- \\tau}} \\gamma , \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 750, + 890, + 785 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\gamma$ is the original margin value and $\\tau > 0$ is the parameter. This self-adaptive margin will enable the triplet loss put more attention on the samples with high similarity scores that possibly are clean pairs.", + "bbox": [ + 496, + 792, + 890, + 853 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Meta Training Process. To ensure the reliability of the similarity scores produced by MSCN, we utilize a small", + "bbox": [ + 496, + 869, + 890, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7519", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/8aeb08c230bb42145417fb7269135a2ab216757ab23d687d75dfe124ef12a91d.jpg", + "image_caption": [ + "Figure 3. Flowchart of the proposed bi-level optimization. At each iteration, we alternately update the main net and the meta net (MSCN) following the steps in Eq. (7) - (9)." + ], + "image_footnote": [], + "bbox": [ + 86, + 97, + 890, + 310 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "amount clean meta-data set $\\mathcal{D}_{meta} = \\{(I_i,T_i)\\}_{i = 1}^M$ to guide the training, where $M$ is much smaller than $N$ . While meta-learning has been extensively studied in recent literature [10, 32, 46], it remains challenging to construct the meta-process in our objective. On the one hand, the similarity features from main net can not be used directly to produce loss. On the other hand, the meta-process should enable the output of MSCN to represent the similarity score. To achieve this, we view a binary classification task as the meta-process: given an image-text pair, the MSCN is learned to determine whether the modalities correspond to each other or not. Specifically, the positive pairs are taken from the meta-data which contain corresponding image and text, i.e. $(I_{i},T_{i},y_{i} = 1)$ , while negative pairs are extracting from different pairs in training set to construct mismatched pairs, i.e. $(I_{i},T_{j\\neq i},y_{i} = 0)$ . The binary labels are produced from the data itself, and can be regarded as the ideal similarity score to guide the training of MSCN. The prediction of MSCN measures the probability of a pair being clean which can be naturally equivalent to the similarity score. Formally, in each iteration of training, we randomly construct $M$ mismatched pairs from training set to extend the meta-data set as $\\mathcal{D}_{meta}' = \\{(I_i,T_i,y_i)\\}_{i = 1}^{2M}$ . Then the optimal parameter $\\Theta^{*}$ can be learned by minimizing the following meta-loss calculated on meta-data:", + "bbox": [ + 76, + 371, + 472, + 750 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Theta^ {*} = \\underset {\\Theta} {\\arg \\min } \\mathbb {E} _ {(I _ {i}, T _ {i}) \\in \\mathcal {D} _ {m e t a} ^ {\\prime}} l ^ {m e t a} (I _ {i}, T _ {i}, y _ {i}), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 768, + 468, + 795 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $l^{meta}(\\cdot)$ is cross-entropy loss [35, 37]:", + "bbox": [ + 76, + 804, + 375, + 821 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nl ^ {m e t a} (I _ {i}, T _ {i}, y _ {i}) = - y _ {i} \\cdot l o g \\mathcal {V} _ {\\Theta} \\left(\\mathcal {F} _ {W ^ {*} (\\Theta)} (I _ {i}, T _ {i})\\right). (6)\n$$\n", + "text_format": "latex", + "bbox": [ + 93, + 828, + 468, + 847 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As illustrated in Fig. 2, despite its simplicity, our proposed meta-process can not only enforce the MSCN to learn reliable similarity score but also has the following advantages:", + "bbox": [ + 76, + 854, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "1) previous works only guide the meta net with ground truth data; however, we leverage both positive and negative data as meta-knowledge that encourage the MSCN to learn discrimination from them. 2) We overcome one limitation of meta-learning methods that the lack of meta-data. We extract from different pairs among training set to construct massive negative meta-data. Moreover, it is more suitable to the rule of noisy correspondence, which the irregularity of noisy data causes in various patterns.", + "bbox": [ + 496, + 371, + 893, + 508 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Bi-level Optimization. Motivated by recent works [9, 16, 30, 32], we use a bi-level optimization strategy to guarantee the efficiency for updating our main net and MSCN. As illustrated in Fig. 3, it is practically solved via two nested loops of optimization. Specifically, at each iteration $t$ of training, we sample a mini-batch of training pairs $\\{(I_i,T_i)\\}_{i = 1}^n$ , where $n$ is the size of mini-batch. We first update the main net parameters by taking a gradient descent step toward the direction of triplet loss, which can be formulated as follows:", + "bbox": [ + 498, + 526, + 893, + 676 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {W} ^ {(t)} (\\Theta) = W ^ {(t)} - \\alpha \\sum_ {i = 1} ^ {n} \\left. \\nabla_ {W} l ^ {\\text {t r a i n}} \\left(I _ {i}, T _ {i}\\right) \\right| _ {W ^ {(t)}}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 685, + 892, + 724 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\alpha$ means the learning rate. The updated $\\hat{W}^{(t)}$ can be regard as the approximation of the optimal $W^{*}(\\Theta)$ at $t$ iteration, containing the gradient information of $\\Theta$ . Then we update the parameters of MSCN with main net parameters fixed. Guided by the objective loss in Eq. (5) and a minibatch of meta-data $\\{(I_i,T_i,y_i)\\}_{i = 1}^m$ , where $m$ is the minibatch size. The updated parameters of MSCN is calculated as:", + "bbox": [ + 496, + 734, + 893, + 856 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left. \\Theta^ {(t + 1)} = \\Theta^ {(t)} - \\beta \\frac {1}{m} \\sum_ {i = 1} ^ {m} \\nabla_ {\\Theta} l ^ {m e t a} \\left(I _ {i}, T _ {i}, y _ {i}\\right) \\right| _ {\\Theta^ {(t)}}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 864, + 892, + 905 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7520", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\beta$ is the learning rate for meta-process. Finally, we update the parameters of main net while fixing the MSCN parameters. With the corrected similarity scores produced by the newly updated $\\Theta^{(t + 1)}$ , we can update the main net parameters $W^{(t + 1)}$ by employing the gradient decent as:", + "bbox": [ + 75, + 90, + 468, + 167 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nW ^ {(t + 1)} = W ^ {(t)} - \\alpha \\sum_ {i = 1} ^ {n} \\left. \\nabla_ {W} l ^ {\\text {t r a i n}} \\left(I _ {i}, T _ {i}\\right) \\right| _ {W ^ {(t)}}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 179, + 468, + 218 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The above procedure proceeds in an alternating manner to optimize the main net and MSCN until it exceeds the maximal iteration.", + "bbox": [ + 75, + 229, + 468, + 273 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Meta-Knowledge Guided Data Purification", + "text_level": 1, + "bbox": [ + 76, + 286, + 442, + 301 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As mentioned, our MSCN can provide corrected similarity score to avoid noise. However, due to the property of triplet loss, it remains to produce positive loss for the noisy pairs even if we employ the ideal similarity score (e.g., $\\mathcal{V}(F(I_i,T_i))\\approx 0 < \\mathcal{V}\\left(F(I_i^-,T_i)\\right) + \\hat{\\gamma})$ , which results in the model fitting to the noise. To tackle this issue, we adopt sample selection strategy to remove the noisy data and only use clean samples as training data. Previous works [2, 42] motivated by the memorization effect of DNNs that clean sample has a lower loss during the beginning of training. Unfortunately, the triplet ranking loss is more complex, i.e., affected by both positive and negative samples, which leads to sub-optimal selection results. Based on the observation shown in Fig. 4, the clean and noisy pairs are prone to distinguish from the similarity scores. Gaussian Mixture Model (GMM) is a widely used technique to model mixture distribution [15, 20]. However, we find the similarity distribution exhibits high skew toward 1 and causes a poor approximation for GMM. Therefore, we fit a two-component Beta Mixture Model (BMM) [27] of per-pair similarity score $s_i$ to better distinguish clean and noisy samples:", + "bbox": [ + 75, + 309, + 468, + 627 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\np \\left(s _ {i}\\right) = \\sum_ {k = 1} ^ {K} \\lambda_ {k} \\phi \\left(s _ {i} \\mid \\alpha_ {k}, \\beta_ {k}\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 640, + 468, + 680 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $K = 2$ and $\\lambda_{k}$ is the mixture coefficient, and $\\phi (s|\\alpha_k,\\beta_k)$ denotes the probability density function with parameters $\\alpha_{k},\\beta_{k} > 0$", + "bbox": [ + 75, + 688, + 468, + 733 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use the Expectation-Maximization [5] algorithm to optimize the BMM, which poses a crucial challenge: the initialization of model parameters. Existing methods typically adopt the $K$ -Means algorithm [23] for initializing; however, it is time-consuming (i.e., the time complexity is $O(TKN)$ , where $T$ is the number of iterations) and underperforming. We thus propose a novel initialization approach which takes the meta-data as prior knowledge. Specifically, we fit the two Beta components with positive and negative meta-data to model the distribution of clean and noisy pairs, respectively. Given the similarity scores $\\mathcal{S}_p = \\{s_i\\}_{i=1}^M$", + "bbox": [ + 75, + 734, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/79d0c53fb0b4801da8eba6d1025d4c5f036e16be40e0fe4988bd68563c7e1a9d.jpg", + "image_caption": [ + "Figure 4. Per-sample similarity score distribution, estimated BMM and BMM with meta prior knowledge under $20\\%$ noisy correspondence in Flickr30K after 5 epochs warmup." + ], + "image_footnote": [], + "bbox": [ + 501, + 92, + 885, + 284 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "corresponding to positive meta-data, the parameters of the component modeling clean distribution can be initialled by", + "bbox": [ + 498, + 359, + 890, + 388 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha_ {k} = \\frac {\\left(1 - E \\left(\\mathcal {S} _ {p}\\right)\\right) E \\left(\\mathcal {S} _ {p}\\right) ^ {2}}{V \\left(\\mathcal {S} _ {p}\\right)} - E \\left(\\mathcal {S} _ {p}\\right), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 400, + 890, + 441 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\beta_ {k} = \\frac {\\alpha_ {k} \\left(1 - E \\left(\\mathcal {S} _ {p}\\right)\\right)}{E (\\mathcal {S} _ {p})},\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 438, + 728, + 470 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $E(\\cdot)$ and $V(\\cdot)$ are the mean and variance function, respectively. The initialization of component modeling noisy distribution is analogous but uses similarity scores $S_{n} = \\{s_{i}\\}_{i=1}^{M}$ of negative meta-data. As illustrated in Fig. 4, our meta-knowledge guided initialization method achieves better approximation while with lower time complexity (i.e., $O(KM)$ ).", + "bbox": [ + 500, + 482, + 890, + 587 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, we compute the posterior probability $p_i = p(k|s_i) = p(k)p(s_i|\\alpha_k,\\beta_k) / p(s_i)$ as the probability of $i$ -th simple being clean. The purified training set is defined as:", + "bbox": [ + 498, + 588, + 890, + 646 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {D} _ {\\text {t r a i n}} ^ {\\prime} = \\left\\{\\left(I _ {i}, T _ {i}\\right) \\in \\mathcal {D} _ {\\text {t r a i n}} \\mid p _ {i} > 0. 5 \\right\\}. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 648, + 890, + 665 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "At each iteration, we produce the purified dataset before the training of our meta-learning objective. Moreover, following [12, 15, 20], we maintain two networks $\\{\\mathcal{F}_W^1,\\mathcal{V}_{\\Theta}^1\\}$ and $\\{\\mathcal{F}_W^2,\\mathcal{V}_{\\Theta}^2\\}$ to avoid error accumulation that one network will produce the purified training set to train the other one. The full algorithm is shown in supplementary material.", + "bbox": [ + 498, + 672, + 890, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 500, + 776, + 624, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets and Evaluation Protocol", + "text_level": 1, + "bbox": [ + 500, + 801, + 790, + 815 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We evaluate our method on three standard image-text retrieval datasets. To be specific, the Flickr30K [43] collects 31,783 images with 5 corresponding captions each from Flickr website. Following the split in [15], we use 1,000 image-text pairs for validation, 1,000 image-text", + "bbox": [ + 498, + 824, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7521", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "pairs for testing, and the rest for training. The MS-COCO [24] contains 123,287 images with 5 corresponding captions each. We split the MS-COCO following [15] that 113,287 image-text pairs for training, 5,000 pairs for validation, and 5,000 pairs for testing. Conceptual Captions [31] is a large-scale dataset contains 3.3M images with one corresponding caption each. We employ the same subset following [15], i.e., CC152K. Specifically, CC152K consists of 150,000 pairs from training set for training, 1,000 pairs from validation set for validation, and 1,000 pairs from validation set for testing. Moreover, we use an extra meta-data set from validation containing 3,000 pairs.", + "bbox": [ + 75, + 90, + 472, + 272 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Protocol. We evaluate the retrieval performance with the recall rate at K (R@K) metric. In a nutshell, R@K is the proportion for the relevant items retrieved in the closest K items to the query. In our experiments, we take image and text as queries, respectively, and report R@1, R@5, and R@10 results for a comprehensive evaluation. Following [15], we average the similarity scores from two networks at the inference phase.", + "bbox": [ + 75, + 297, + 470, + 417 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 76, + 431, + 294, + 446 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As a general approach, our method can be easily applied to almost all cross-modal retrieval methods to improve robustness. For fair comparison, we chose the same network backbones with NCR [15], i.e., the projection modules and similarity function. For all datasets, the training processes contain 50 epochs after a 5 epochs warmup. We employ ADAM [17] as the optimizer for both main net and meta net with a batch size of 64. To fit the BMM, we set the stop threshold as $10^{-2}$ and maximum number of iterations as 10 for the EM procedure. Note that, we clamp the similarity scores into a range of $[10^{-4}, 1 - 10^{-4}]$ instead of [0,1] for stable fitting. Moreover, we set the margin parameter as 0.2 and $\\tau = 2$ to calculate the self-adaptive margin. For all datasets, the number of ground-truth meta-data is approximately $2\\%$ of training data, and the learning rate of main net and meta net are initialized with $2 \\times 10^{-4}$ and $1.7 \\times 10^{-5}$ , respectively. For CC152K containing real noise, we employ the validation data as the meta set due to the lack of ground-truth data. For all experiments, we decay the learning rate by 0.1 after 30 epochs.", + "bbox": [ + 75, + 455, + 472, + 758 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Comparison with the State-of-the-Art", + "text_level": 1, + "bbox": [ + 76, + 770, + 403, + 787 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we conduct comparison experiments with 7 state-of-the-art methods that include SCAN [18], VSRN [22], IMRAM [4], SGR, SAF, SGRAF [7], and NCR [15]. As Flickr30K and MS-COCO are well annotated datasets, we generate synthetic noisy correspondence by randomly shuffling the training images and captions for a specific noise ratio, and we conduct experiments with three different", + "bbox": [ + 75, + 795, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ff3b2686aa389e0441a49234352caa85e26771c520baa0eb273df3f873cd9d97.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Image to Text
MethodsR@1R@5R@10SUM
SCAN (ECCV'18)30.555.365.3151.1
VSRN(ICCV'19)32.661.370.5164.4
IMRAM(CVPR'20)33.157.668.1158.8
SAF (AAAI'21)31.759.368.2159.2
SGR (AAAI'21)11.329.739.680.6
SGR* (AAAI'21)35.063.473.3171.7
NCR(NIPS'21)39.564.573.5177.5
MSCN*(Ours)39.765.475.3180.4
MSCN(Ours)40.165.776.6182.4
Text to Image
SCAN (ECCV'18)26.953.064.7144.6
VSRN(ICCV'19)32.559.470.4162.3
IMRAM(CVPR'20)29.056.867.4153.2
SAF (AAAI'21)31.959.067.9158.8
SGR (AAAI'21)13.130.141.684.8
SGR* (AAAI'21)34.963.072.8170.7
NCR(NIPS'21)40.364.673.2178.1
MSCN*(Ours)39.866.175.0180.9
MSCN(Ours)40.667.476.3184.3
", + "bbox": [ + 501, + 89, + 897, + 435 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Image-Text Retrieval Performance on CC152K with real noise, and the best results are highlighted in bold.", + "bbox": [ + 498, + 445, + 890, + 472 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "level of noise ratios, i.e., $20\\%$ , $50\\%$ , and $70\\%$ . For all methods, we choose the best checkpoint on the validation set and report its performance on the test set. Following NCR, we also report two strong baselines based on SGR, i.e., SGR-C and SGR*. In short, SGR-C uses only clean data for training, and SGR* employs a pre-training process while training without hard negatives to improve robustness. To evaluate our method on real-world noisy data, we conduct experiments on the CC152K, which is harvested from the Internet and contains about $3\\% \\sim 20\\%$ noisy correspondence [31].", + "bbox": [ + 496, + 484, + 890, + 652 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Experiments on CC152K. Tab. 1 shows the results on CC152K with real noisy correspondence. On CC152K, our MSCN achieves the state-of-the-art performance in terms of all metrics. Specifically, the sum score of MSCN is $4.9\\%$ and $6.2\\%$ higher than the best baseline in text and image retrieval, respectively. In addition, we also present a more strict comparison (denoted by $\\mathrm{MSCN^{*}}$ ) that trained our method with only 1,000 meta-data (approximately $0.67\\%$ of all data). We can see that our $\\mathrm{MSCN^{*}}$ also achieves a competitive performance. Specifically, all metrics except R@1 of image retrieval are superior to the baselines.", + "bbox": [ + 496, + 670, + 890, + 835 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Experiments on Flickr30K and MS-COCO. Tab. 2 shows the results on Flickr30K and MS-COCO under a range of synthetic noise rates. Following [15], we report the", + "bbox": [ + 498, + 854, + 890, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7522", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/cdccbdd4ebcf945e4b6095c80935d406375adcca92ca7af188b824a7ec4c9011.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Flickr30KMS-COCO
Image to TextText to ImageImage to TextText to Image
NoiseMethodsR@1R@5R@10R@1R@5R@10R@1R@5R@10R@1R@5R@10
20%SCAN59.183.490.436.667.077.566.291.096.445.080.289.3
VSRN58.182.689.340.768.778.225.159.074.817.649.064.1
IMRAM63.086.091.341.471.280.568.692.897.655.785.091.0
SAF51.079.388.038.366.576.267.392.596.653.484.592.4
SGR*62.886.292.244.472.380.467.891.796.252.983.590.1
SGR-C72.890.895.456.482.188.675.495.297.960.188.594.8
NCR75.093.997.558.383.089.077.795.598.262.589.395.3
MSCN77.494.997.659.683.289.278.197.298.864.390.495.8
50%SCAN27.757.668.816.239.349.840.873.584.95.415.121.0
VSRN14.337.650.012.130.039.423.554.769.316.047.865.9
IMRAM9.126.638.22.78.412.721.360.275.922.352.864.3
SAF30.363.675.427.953.765.130.467.882.333.569.082.8
SGR*36.968.180.229.356.267.060.687.493.646.074.279.0
SGR-C69.890.394.850.177.585.271.794.197.757.086.693.7
NCR72.993.096.354.379.886.574.694.697.859.187.894.5
MSCN74.493.296.055.380.486.877.596.298.760.789.194.9
70%SCAN5.619.327.42.28.012.818.143.157.40.31.32.3
VSRN0.82.54.10.51.52.75.115.724.62.58.813.3
IMRAM1.33.13.90.31.22.87.120.033.45.315.222.0
SAF0.52.23.00.20.81.70.11.74.00.61.93.0
SGR*17.942.151.914.631.040.835.771.285.431.665.879.0
SGR-C65.089.394.748.174.581.169.893.697.556.586.093.4
NCR16.138.552.811.029.541.435.469.583.431.566.481.1
MSCN69.089.393.049.273.179.074.494.997.758.887.293.7
", + "bbox": [ + 99, + 89, + 879, + 484 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "results by averaging over 5 folds of 1K test images. From Tab. 2, we can draw the observation that noisy correspondence remarkably affects the performance of cross-modal retrieval methods. The retrieval accuracy will decrease fast with the noise rate increasing. Our MSCN is superior to the baselines in almost all metrics under different noise rates. In the low noise cases, i.e., $20\\%$ and $50\\%$ , our MSCN improves R@1 (sum of image and text) by $3.7\\%$ , $2.2\\%$ , $2.5\\%$ , and $4.5\\%$ in the four valuations compared with NCR. In the high noise case, i.e., $70\\%$ , all baselines can easily overfit to noisy correspondence and results in poor performance. Our MSCN improves R@1 by $52.9\\%$ , $38.2\\%$ , $39.4\\%$ , and $27.3\\%$ in these four valuations compared with NCR. Even compared with the SGR-C trained only on clean data, our MSCN also achieves competitive performance.", + "bbox": [ + 75, + 537, + 472, + 765 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 771, + 232, + 787 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To evaluate the performance of the proposed components (i.e., self-adaptive margin and data purification), we conduct the ablation study on the Flickr30K with $20\\%$ noisy correspondence. Note that for MSCN without $\\mathcal{D}_{train}^{\\prime}$ , we only use a single model to perform the experiment. As shown in Tab. 3, we can see that the performance without $\\hat{\\gamma}$ or $\\mathcal{D}_{train}^{\\prime}$ are worse than the complete MSCN, which", + "bbox": [ + 75, + 794, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "indicates that all components are important to achieve advantageous results. We explain more on the data selection strategy in the supplementary material.", + "bbox": [ + 498, + 537, + 892, + 583 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b0c5cbfc58dad73fe67517f1ad932cd4ae07457f7e2e7d770c126d42c84e34ea.jpg", + "table_caption": [ + "Table 2. Image-Text Retrieval Performance under synthetic noise rates of ${20}\\% ,{50}\\%$ and ${70}\\%$ on Flickr30K and MS-COCO 1K,and the best results are highlighted in bold." + ], + "table_footnote": [], + "table_body": "
MethodImage to Text
MSCNw/o γw/o D'trainR@1R@5R@10
✓77.494.997.6
✓✓75.394.597.2
✓✓75.893.896.2
✓✓✓74.191.594.7
Text to Image
✓59.683.289.2
✓✓58.383.188.9
✓✓55.874.577.4
✓✓✓53.471.272.2
", + "bbox": [ + 509, + 593, + 890, + 761 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Ablation studies on Flickr30K with ${20}\\%$ noise rate.", + "bbox": [ + 514, + 771, + 875, + 786 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Progressive Comparison", + "text_level": 1, + "bbox": [ + 500, + 801, + 723, + 816 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fig. 6 plots the average of recalls on testing set of MSCN and NCR as training proceeds. We show a representative result using Flickr30K with $50\\%$ noisy correspondence. From the result, we can see that our MSCN achieves a more stable performance in overall training process. Although NCR", + "bbox": [ + 496, + 824, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7523", + "bbox": [ + 482, + 944, + 514, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9e2434691c1db0c972440d39d2b10ec8317e5cb19ba4c4225af7f4650834da11.jpg", + "image_caption": [ + "the circular movement of the camera. (0.3444)" + ], + "image_footnote": [], + "bbox": [ + 91, + 99, + 222, + 166 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5416ed9692c3717fe83a5e49a16f9bd801e3f58286f39183058aff68ffcab230.jpg", + "image_caption": [ + "person arrives earlier this month. (0.3485)" + ], + "image_footnote": [], + "bbox": [ + 228, + 108, + 346, + 170 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/dbabc25dd5f7d976e226d17998b27e8e347ed66a3a20f0514f56a9dfb3b00fa8.jpg", + "image_caption": [ + "i expected to learn more about plants. (0.3487)" + ], + "image_footnote": [], + "bbox": [ + 364, + 107, + 475, + 172 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9cbcac1eea18f074e487aea5870ba5834c873e679fedf051122762eab4b98da0.jpg", + "image_caption": [ + "man,chairman to speak at show (0.3488)" + ], + "image_footnote": [], + "bbox": [ + 493, + 109, + 609, + 170 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3c991ea05fe1855eff140559fa47d4adc725baab4d8030b840e375b8c6ab8925.jpg", + "image_caption": [ + "note the many parked cars (0.3511)" + ], + "image_footnote": [], + "bbox": [ + 627, + 107, + 738, + 172 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2c3e2842b25d71a4a0e251b0ff41a8f591dea74df2061f4a2faa062509a58c69.jpg", + "image_caption": [ + "a man using his mobile phone. (0.3518)" + ], + "image_footnote": [], + "bbox": [ + 759, + 109, + 869, + 170 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "does not need an extra meta-set to guide the training, it relies on a clean validation set to choose the best model parameters. The stableness also indicates that our method has better alleviated the interference of noisy correspondence.", + "bbox": [ + 75, + 262, + 468, + 323 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8b2cfa94e9146426a496cfd7be0646a851b640647754760cce2af5d3db163311.jpg", + "image_caption": [ + "Figure 6. Progressive performance comparison of MSCN and NCR on testing set as training proceeds." + ], + "image_footnote": [], + "bbox": [ + 125, + 339, + 421, + 484 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. Detected Noisy Samples", + "text_level": 1, + "bbox": [ + 76, + 537, + 295, + 554 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fig. 5 shows some real-world noisy image-text pairs detected (with low similarity score) by MSCN in Conceptual Captions. We also present the similarity score predicted by our MSCN. For most noisy pairs, the image and text contains completely irrelevant semantic information. Moreover, our MSCN can even find the subtle mistake in the noisy image-text pair. In the last pair, the picture shows a man covering his ears but wrongly caption as using a mobile phone. Although this picture is misleading, our method successfully detects the subtle difference.", + "bbox": [ + 75, + 561, + 468, + 712 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.7. Visualization on Similarity Score", + "text_level": 1, + "bbox": [ + 76, + 719, + 366, + 734 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fig. 7 plots the similarity score distribution predicted by MSCN for clean and noisy training samples. It can be seen that almost all large similarity scores belong to clean pairs, and the noisy pairs' scores are smaller than clean samples, which implies that the trained MSCN can successfully provide corrected similarity scores.", + "bbox": [ + 75, + 743, + 468, + 833 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 845, + 194, + 859 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we explore a meta-learning method to address the problem of learning with noisy correspond-", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/92e2f917921919a070aab04042430702cb4149ba13147bbe56f7cc6ca78364f3.jpg", + "image_caption": [ + "Figure 5. Real-world noisy examples detected by our MSCN.", + "(a)", + "Figure 7. Per-sample similarity distribution on training data of Flickr30K. (a) $20\\%$ noisy correspondence. (b) $50\\%$ noisy correspondence." + ], + "image_footnote": [], + "bbox": [ + 508, + 263, + 691, + 436 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ac5ce36a3925aec0bb9dbc6372691a5b0277f3b6287eddd6b638c5c8d8b6b5ee.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 700, + 263, + 883, + 436 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "dence. Specifically, we propose a meta correction network (MSCN) to provide reliable similarity scores. Our MSCN is trained by a novel meta-process that views both positive and negative data as meta-knowledge to encourage MSCN to learn discrimination from them. To further mitigate noise interference, we design a data purification strategy that uses meta-data as prior knowledge to purify the noisy data efficiently. We conduct comprehensive experiments on three widely used datasets. The results validate the effectiveness of our method in both synthetic and real noises.", + "bbox": [ + 496, + 513, + 890, + 666 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 500, + 679, + 666, + 695 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work is supported by the National Key Research and Development Program of China (2020AAA0108800), National Nature Science Foundation of China (62192781, 62272374, 62202367, 62250009, 62137002, 61937001), Innovative Research Group of the National Natural Science Foundation of China (61721002), Innovation Research Team of Ministry of Education (IRT_17R86), Project of China Knowledge Center for Engineering Science and Technology, and Project of Chinese academy of engineering \"The Online and Offline Mixed Educational Service System for 'The Belt and Road' Training in MOOC China\". We would like to express our gratitude for the support of K. C. Wong Education Foundation.", + "bbox": [ + 496, + 703, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7524", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Relja Arandjelovic and Andrew Zisserman. Objects that sound. In Proceedings of the European conference on computer vision (ECCV), pages 435-451, 2018. 1", + "[2] Eric Arazo, Diego Ortego, Paul Albert, Noel O'Connor, and Kevin McGuinness. Unsupervised label noise modeling and loss correction. In International conference on machine learning, pages 312-321. PMLR, 2019. 5", + "[3] Devansh Arpit, Stanislaw K Jastrzebski, Nicolas Ballas, David Krueger, Emmanuel Bengio, Maxinder S Kanwal, Tegan Maharaj, Asja Fischer, Aaron C Courville, Yoshua Bengio, et al. A closer look at memorization in deep networks. In ICML, 2017. 1", + "[4] Hui Chen, Guiguang Ding, Xudong Liu, Zijia Lin, Ji Liu, and Jungong Han. Imram: Iterative matching with recurrent attention memory for cross-modal image-text retrieval. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12655-12663, 2020. 6", + "[5] Arthur P Dempster, Nan M Laird, and Donald B Rubin. Maximum likelihood from incomplete data via the em algorithm. Journal of the Royal Statistical Society: Series B (Methodological), 39(1):1-22, 1977. 5", + "[6] Cheng Deng, Zhaojia Chen, Xianglong Liu, Xinbo Gao, and Dacheng Tao. Triplet-based deep hashing network for cross-modal retrieval. IEEE Transactions on Image Processing, 27(8):3893-3903, 2018. 1", + "[7] Haiwen Diao, Ying Zhang, Lin Ma, and Hutchuan Lu. Similarity reasoning and filtration for image-text matching. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 1218-1226, 2021. 2, 6", + "[8] Fartash Faghri, David J Fleet, Jamie Ryan Kiros, and Sanja Fidler. Vse++: Improving visual-semantic embeddings with hard negatives. arXiv preprint arXiv:1707.05612, 2017. 2, 3", + "[9] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In International conference on machine learning, pages 1126–1135. PMLR, 2017. 2, 4", + "[10] Luca Franceschi, Paolo Frasconi, Saverio Salzo, Riccardo Grazzi, and Massimiliano Pontil. Bilevel programming for hyperparameter optimization and meta-learning. In International Conference on Machine Learning, pages 1568-1577. PMLR, 2018. 4", + "[11] Dalu Guo, Chang Xu, and Dacheng Tao. Image-question-answer synergistic network for visual dialog. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10434-10443, 2019. 1", + "[12] Bo Han, Quanming Yao, Xingrui Yu, Gang Niu, Miao Xu, Weihua Hu, Ivor Tsang, and Masashi Sugiyama. Co-teaching: Robust training of deep neural networks with extremely noisy labels. Advances in neural information processing systems, 31, 2018. 5", + "[13] Haochen Han, Qinghua Zheng, Minnan Luo, Kaiyao Miao, Feng Tian, and Yan Chen. Noise-tolerant learning for audiovisual action recognition. arXiv preprint arXiv:2205.07611, 2022. 2", + "[14] Peng Hu, Xi Peng, Hongyuan Zhu, Liangli Zhen, and Jie Lin. Learning cross-modal retrieval with noisy labels. In" + ], + "bbox": [ + 78, + 114, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5403-5413, 2021. 2", + "[15] Zhenyu Huang, Guocheng Niu, Xiao Liu, Wenbiao Ding, Xinyan Xiao, Hua Wu, and Xi Peng. Learning with noisy correspondence for cross-modal matching. Advances in Neural Information Processing Systems, 34:29406-29419, 2021. 1, 2, 5, 6", + "[16] Haoxuanye Ji, Le Wang, Sanping Zhou, Wei Tang, Nanning Zheng, and Gang Hua. Meta pairwise relationship distillation for unsupervised person re-identification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3661–3670, 2021. 4", + "[17] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR (Poster), 2015. 6", + "[18] Kuang-Huei Lee, Xi Chen, Gang Hua, Houdong Hu, and Xiaodong He. Stacked cross attention for image-text matching. In Proceedings of the European conference on computer vision (ECCV), pages 201-216, 2018. 2, 6", + "[19] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy Hospedales. Learning to generalize: Meta-learning for domain generalization. In Proceedings of the AAAI conference on artificial intelligence, volume 32, 2018. 2", + "[20] Junnan Li, Richard Socher, and Steven CH Hoi. Dividemix: Learning with noisy labels as semi-supervised learning. In International Conference on Learning Representations, 2019. 1, 2, 5", + "[21] Junnan Li, Yongkang Wong, Qi Zhao, and Mohan S Kankanhalli. Learning to learn from noisy labeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5051-5059, 2019. 2", + "[22] Kunpeng Li, Yulun Zhang, Kai Li, Yuanyuan Li, and Yun Fu. Visual semantic reasoning for image-text matching. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4654-4662, 2019. 2, 6", + "[23] Aristidis Likas, Nikos Vlassis, and Jakob J Verbeek. The global k-means clustering algorithm. Pattern recognition, 36(2):451-461, 2003. 5", + "[24] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 6", + "[25] Chunxiao Liu, Zhendong Mao, Tianzhu Zhang, Hongtao Xie, Bin Wang, and Yongdong Zhang. Graph structured network for image-text matching. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10921-10930, 2020. 2", + "[26] Lin Ma, Zhengdong Lu, and Hang Li. Learning to answer questions from image using convolutional neural network. In Thirtieth AAAI Conference on Artificial Intelligence, 2016. 1", + "[27] Zhanyu Ma and Arne Leijon. Bayesian estimation of beta mixture models with variational inference. IEEE Transactions on Pattern Analysis and Machine Intelligence, 33(11):2160-2173, 2011. 5", + "[28] Nikhil Mishra, Mostafa Rohaninejad, Xi Chen, and Pieter Abbeel. A simple neural attentive meta-learner. In International Conference on Learning Representations, 2018. 2" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7525", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Yang Qin, Dezhong Peng, Xi Peng, Xu Wang, and Peng Hu. Deep evidential learning with noisy correspondence for cross-modal retrieval. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4948-4956, 2022. 2", + "[30] Mengye Ren, Wenyuan Zeng, Bin Yang, and Raquel Urtasun. Learning to reweight examples for robust deep learning. In International conference on machine learning, pages 4334-4343. PMLR, 2018. 4", + "[31] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2556-2565, 2018. 1, 6", + "[32] Jun Shu, Qi Xie, Lixuan Yi, Qian Zhao, Sanping Zhou, Zongben Xu, and Deyu Meng. Meta-weight-net: Learning an explicit mapping for sample weighting. Advances in neural information processing systems, 32, 2019. 2, 4", + "[33] Didac Surís, Amanda Duarte, Amaia Salvador, Jordi Torres, and Xavier Giró-i Nieto. Cross-modal embeddings for video and audio retrieval. In Proceedings of the European Conference on Computer Vision (ECCV) Workshops, pages 0–0, 2018. 1", + "[34] Jun Tang, Ke Wang, and Ling Shao. Supervised matrix factorization hashing for cross-modal retrieval. IEEE Transactions on Image Processing, 25(7):3157-3166, 2016. 2", + "[35] Bokun Wang, Yang Yang, Xing Xu, Alan Hanjalic, and Heng Tao Shen. Adversarial cross-modal retrieval. In Proceedings of the 25th ACM international conference on Multimedia, pages 154-162, 2017. 4", + "[36] Liwei Wang, Yin Li, and Svetlana Lazebnik. Learning deep structure-preserving image-text embeddings. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5005-5013, 2016. 2", + "[37] Yunchao Wei, Yao Zhao, Canyi Lu, Shikui Wei, Luoqi Liu, Zhenfeng Zhu, and Shuicheng Yan. Cross-modal retrieval with cnn visual features: A new baseline. IEEE transactions on cybernetics, 47(2):449-460, 2016. 4", + "[38] Xing Xu, Li He, Huimin Lu, Lianli Gao, and Yanli Ji. Deep adversarial metric learning for cross-modal retrieval. *World Wide Web*, 22(2):657-672, 2019. 2", + "[39] Mouxing Yang, Zhenyu Huang, Peng Hu, Taihao Li, Jiancheng Lv, and Xi Peng. Learning with twin noisy labels for visible-infrared person re-identification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14308-14317, 2022. 2", + "[40] Mouxing Yang, Yunfan Li, Peng Hu, Jinfeng Bai, Jiancheng Lv, and Xi Peng. Robust multi-view clustering with incomplete information. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(1):1055-1069, 2022. 2", + "[41] Mouxing Yang, Yunfan Li, Zhenyu Huang, Zitao Liu, Peng Hu, and Xi Peng. Partially view-aligned representation learning with noise-robust contrastive loss. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1134–1143, 2021. 2", + "[42] Quanming Yao, Hansi Yang, Bo Han, Gang Niu, and James Tin-Yau Kwok. Searching to exploit memorization effect in" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "learning with noisy labels. In International Conference on Machine Learning, pages 10789-10798. PMLR, 2020. 5", + "[43] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics, 2:67-78, 2014. 5", + "[44] Qi Zhang, Zhen Lei, Zhaoxiang Zhang, and Stan Z Li. Context-aware attention network for image-text retrieval. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3536-3545, 2020. 1", + "[45] Liangli Zhen, Peng Hu, Xu Wang, and Dezhong Peng. Deep supervised cross-modal retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10394-10403, 2019. 2", + "[46] Guoqing Zheng, Ahmed Hassan Awadallah, and Susan Dumais. Meta label correction for noisy label learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 11053-11061, 2021. 2, 4" + ], + "bbox": [ + 501, + 92, + 893, + 359 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7526", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Noisy Correspondence Learning With Meta Similarity Correction/0167b343-6cc5-4960-bab8-e11934c107f0_model.json b/2023/Noisy Correspondence Learning With Meta Similarity Correction/0167b343-6cc5-4960-bab8-e11934c107f0_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2880109c019026ab18d23d267e5cd57c676c2255 --- /dev/null +++ b/2023/Noisy Correspondence Learning With Meta Similarity Correction/0167b343-6cc5-4960-bab8-e11934c107f0_model.json @@ -0,0 +1,2211 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.156, + 0.131, + 0.816, + 0.154 + ], + "angle": 0, + "content": "Noisy Correspondence Learning with Meta Similarity Correction" + }, + { + "type": "text", + "bbox": [ + 0.238, + 0.181, + 0.739, + 0.199 + ], + "angle": 0, + "content": "Haochen Han\\(^{1}\\), Kaiyao Miao\\(^{2}\\), Qinghua Zheng\\(^{1}\\), Minnan Luo\\(^{1*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.2, + 0.751, + 0.216 + ], + "angle": 0, + "content": "\\(^{1}\\) School of Computer Science and Technology, Xi'an Jiaotong University, China" + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.218, + 0.741, + 0.234 + ], + "angle": 0, + "content": "\\(^{2}\\) School of Cyber Science and Engineering, Xi'an Jiaotong University, China" + }, + { + "type": "text", + "bbox": [ + 0.187, + 0.237, + 0.783, + 0.252 + ], + "angle": 0, + "content": "{hhc1997, miaoky814}@stu.xjtu.edu.cn, {qhzheng, minnluo}@xjtu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.287, + 0.314, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.319, + 0.474, + 0.637 + ], + "angle": 0, + "content": "Despite the success of multimodal learning in cross-modal retrieval task, the remarkable progress relies on the correct correspondence among multimedia data. However, collecting such ideal data is expensive and time-consuming. In practice, most widely used datasets are harvested from the Internet and inevitably contain mismatched pairs. Training on such noisy correspondence datasets causes performance degradation because the cross-modal retrieval methods can wrongly enforce the mismatched data to be similar. To tackle this problem, we propose a Meta Similarity Correction Network (MSCN) to provide reliable similarity scores. We view a binary classification task as the meta-process that encourages the MSCN to learn discrimination from positive and negative meta-data. To further alleviate the influence of noise, we design an effective data purification strategy using meta-data as prior knowledge to remove the noisy samples. Extensive experiments are conducted to demonstrate the strengths of our method in both synthetic and real-world noises, including Flickr30K, MS-COCO, and Conceptual Captions. Our code is publicly available.\\(^{1}\\)" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.669, + 0.21, + 0.685 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.695, + 0.47, + 0.831 + ], + "angle": 0, + "content": "Recently, cross-modal retrieval has drawn much attention with the rapid growth of multimedia data. Given a query sample of specific modality, cross-modal retrieval aims to retrieve relevant samples across different modalities. Existing cross-modal retrieval works [1, 6, 11] usually learn a comparable common space to bridge different modalities, which achieved remarkable progress in many applications, including video-audio retrieval [1, 33], visual question answering [11, 26], and image-text matching [44]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.832, + 0.47, + 0.863 + ], + "angle": 0, + "content": "Despite the promise, a core assumption in cross-modal retrieval is the correct correspondence among multiple" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.288, + 0.892, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.416, + 0.895, + 0.487 + ], + "angle": 0, + "content": "Figure 1. Illustration of noisy correspondence in image-text retrieval. A standard triplet loss is used to enforce the positive pairs to be closer than negatives in the common space. Noisy correspondence is the mismatched pairs but wrongly considered as positive ones, and thus results in model performance degradation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.507, + 0.893, + 0.734 + ], + "angle": 0, + "content": "modalities. However, collecting such ideal data is expensive and time-consuming. In practice, most widely used datasets are harvested from the Internet and inevitably contain noisy correspondence [31]. As illustrated in Fig. 1, the cross-modal retrieval method will wrongly enforce the mismatched data to be similar when learning with noisy correspondence, which may significantly affect the retrieval performance. To date, few effort has been made to address this. Huang [15] first researches this issue and proposes the NCR method to train from the noisy image-text pairs robustly. Inspired by the prior success for noisy labels [20], NCR divides the data into clean and noisy partitions and rectifies the correspondence with an adaptive model. However, NCR based on the memorization effect of DNNs [3], which leads to poor performance under high noise ratio." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.895, + 0.901 + ], + "angle": 0, + "content": "To tackle the challenge, we propose a Meta Similarity Correction Network (MSCN) which aims to provide reliable similarity scores for the noisy features from main net. We view a binary classification task as the meta-process: given a multimodal sample, the MSCN will learn to determine whether the modalities correspond to each other or not, where the prediction of MSCN can be naturally regarded as the similarity score. Specifically, a small amount of clean pairs is used to construct positive and negative meta-data, both viewed as meta-knowledge that encourages MSCN to learn the discrimination. Meanwhile, the main net" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.875, + 0.348, + 0.888 + ], + "angle": 0, + "content": "*Minnan Luo is the corresponding author." + }, + { + "type": "page_footnote", + "bbox": [ + 0.097, + 0.888, + 0.348, + 0.9 + ], + "angle": 0, + "content": "1https://github.com/hhc1997/MSCN" + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.875, + 0.348, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7517" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.257 + ], + "angle": 0, + "content": "trained by the corrected similarity score from MSCN with a self-adaptive margin to achieve robust learning. This interdependency makes the main net and MSCN benefits each other against noise. However, due to the property of triplet loss, it remains to produce positive loss for noisy pairs even if we employ the ideal similarity scores. To this end, we further propose a meta-knowledge guided data purification strategy to remove samples with potentially wrong correspondence. Extensive experiments are conducted to demonstrate the strengths of our method in both synthetic and real-world noises." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.259, + 0.472, + 0.287 + ], + "angle": 0, + "content": "The main contributions of this work are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.29, + 0.468, + 0.35 + ], + "angle": 0, + "content": "- We pioneer the exploration of meta-learning for noisy correspondence problem, where a meta correction network is proposed to provide reliable similarity scores against noise." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.351, + 0.468, + 0.409 + ], + "angle": 0, + "content": "- We present a novel meta-process that first considers both positive and negative data as meta-knowledge to encourage the MSCN to learn discrimination from them." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.411, + 0.468, + 0.456 + ], + "angle": 0, + "content": "- We design an effective data purification strategy using meta-data as prior knowledge to remove the noisy samples." + }, + { + "type": "list", + "bbox": [ + 0.097, + 0.29, + 0.468, + 0.456 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.472, + 0.22, + 0.488 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.498, + 0.289, + 0.513 + ], + "angle": 0, + "content": "2.1. Cross-Modal Retrieval" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.523, + 0.47, + 0.688 + ], + "angle": 0, + "content": "Cross-modal retrieval works aim to take one modality of data as query to retrieve the related data in other modalities. Based on the utilization of annotation information, cross-modal retrieval methods can be roughly divided into two categories: 1) Unsupervised Methods [8,22,25,36]. It only uses the co-occurrence information to learn the common representations among the multimodal data. 2) Supervised methods [14,34,38,45]. The extra label information is employed to boost the discrimination of common representations. However, the lack of annotations limited the practicability in the real world." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Our proposed approach falls in the category of unsupervised methods. For example, Wang [36] uses a two-branch neural networks to learn the joint embeddings of multimodal data. Inspired by hard negative mining, \\(\\mathrm{VSE}++\\) [8] uses hard negatives to improve the retrieval performance. To capture fine-grained interplay between modalities, SCAN [18] introduce stacked cross attention to enable attention with context from both image and sentence. Recently, motivated by the powerful learning ability of graph model, GSMN [25] and SGRAF [7] construct graph structure for multimodal data to benefit the learning of fine-grained correspondence. Although existing works have achieved remarkable results, they usually depend on the correct correspondence among multimodal data and cannot" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.138 + ], + "angle": 0, + "content": "tackle the noise issue. Thus, it is significant to explore how to learn cross-modal retrieval with noisy correspondence, but which is rarely touched in previous studies." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.147, + 0.785, + 0.164 + ], + "angle": 0, + "content": "2.2. Noisy Correspondence Learning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.17, + 0.892, + 0.428 + ], + "angle": 0, + "content": "As a newly proposed problem, noisy correspondence is the mismatched pairs but wrongly considered as positive ones. Huang [15] first research this issue and propose the NCR method to robustly train the image-text matching model with noisy correspondence pairs. Inspired by the prior success for noisy labels [20], NCR divides the data into clean and noisy part based on the memorization effect of DNNs, and then rectifies the correspondence with an adaptive model. Qin [29] proposes a uncertainty-based method to achieve efficient learning. Moreover, some works [40, 41] study the partially view-unaligned problem which can be viewed as a generalized noisy correspondence learning. Recently, some works [13, 39] explore a more challenging scenario where noisy labels and noisy correspondence may occur simultaneously. In contrast, our work tackles this challenge in cross-modal retrieval from a meta-learning perspective." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.437, + 0.655, + 0.453 + ], + "angle": 0, + "content": "2.3. Meta-Learning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.461, + 0.892, + 0.642 + ], + "angle": 0, + "content": "The objective of meta-learning is to learn at a more advanced level than conventional learning, such as learning the update rule [28], finding easily fine-tuned parameters [9], or adapting to new tasks [19]. Recently, researchers use meta-learning to find model's parameters that robust against noisy labels [21,32,46]. For example, Li [21] optimizes a meta-objective before conventional training to enable the model not overfit to noise. MW-Net [32] use a meta-process to automatically assign weights to the training samples. MLC [46] presents a label correction network trained as a meta manner which generate reliable labels for the noisy training data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.643, + 0.892, + 0.779 + ], + "angle": 0, + "content": "The most relevant existing method is MLC [46], which first poses the noisy label problem as a meta label correction. However, we argue that our proposed method differs from MLC in two aspects. First, we focuses on the noisy correspondence problem in multimodal data instead of the unimodal classification scenarios. Second, we leverage both positive and negative meta-data to guide model learning the discrimination, which is unexplored in previous meta-learning methods." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.792, + 0.712, + 0.809 + ], + "angle": 0, + "content": "3. The Proposed Method" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.817, + 0.704, + 0.832 + ], + "angle": 0, + "content": "3.1. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Without losing generality, we use the image-text retrieval task as a proxy to investigate the noisy correspondence problem in cross-modal retrieval. Given a training set \\(\\mathcal{D}_{train} = \\{(I_i,T_i,y_i)\\}_{i = 1}^N\\), where \\((I_{i},T_{i})\\) is the \\(i\\)-th image-" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7518" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.09, + 0.089, + 0.895, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.302, + 0.895, + 0.345 + ], + "angle": 0, + "content": "Figure 2. Overview of the proposed meta-process. The positive pairs are taken from the original meta-data which contain corresponding image and text, while negative pairs are extracting from different pairs in training set to construct mismatched pairs. Then the MSCN is learned to determine whether the modalities correspond to each other or not." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.363, + 0.473, + 0.469 + ], + "angle": 0, + "content": "text pair, \\( y_{i} \\in \\{1,0\\} \\) indicates the pair is positive (matched), i.e. \\( (I_i, T_i) \\), or negative (mismatched), i.e. \\( (I_i, T_{j \\neq i}) \\), and \\( N \\) is the number of the entire training data. The noisy correspondence in bimodal data is the negative pair but wrongly considered as \\( y_{i} = 1 \\). To address this issue, we propose a meta-learning based method to achieve robust training. The details are delineated next." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.483, + 0.45, + 0.5 + ], + "angle": 0, + "content": "3.2. Meta-Learning Based Similarity Correction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.508, + 0.47, + 0.613 + ], + "angle": 0, + "content": "The Training Objective. To begin with, we describe the standard projection module to enable cross-modal features to be comparable. Let \\( f(I;W_f) \\) and \\( g(T;W_g) \\) be the modal-specific networks to map the visual and textual modalities into the joint embedding space, respectively. Then we compute the similarity feature between bimodal features by the function" + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.628, + 0.47, + 0.663 + ], + "angle": 0, + "content": "\\[\nS (f (I), g (T); W _ {s}) = \\frac {W _ {s} | f (I) - g (T) | ^ {2}}{\\| W _ {s} | f (I) - g (T) | ^ {2} \\| _ {2}}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.674, + 0.473, + 0.903 + ], + "angle": 0, + "content": "where \\( W_{s} \\) is a learnable parameter matrix to obtain low-dimensional similarity representation. For notation convenience, we collect \\( \\{W_{f},W_{g},W_{s}\\} \\) into \\( W \\) as the parameters of main net \\( \\mathcal{F} = (f,g,S) \\) and denote similarity feature as \\( \\mathcal{F}_W(I,T) \\). With the presence of noisy correspondence, the model will inevitably overfit to noise if we use the similarity representation directly. To tackle this problem, we propose a meta similarity correction network (MSCN), regarding process of obtaining similarity score as a meta-process, which takes the noisy similarity representation as input and produces the corrected similarity score. Specifically, our MSCN is formulated as a MLP network \\( \\mathcal{V}_{\\Theta}(\\cdot) \\) with parameters \\( \\Theta \\). The output \\( s_i = \\mathcal{V}_{\\Theta}(\\mathcal{F}_W(I_i,T_i)) \\) is activated with Sigmoid function to represent the similarity score located in [0, 1]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.363, + 0.895, + 0.499 + ], + "angle": 0, + "content": "For the noisy training pairs, the main net learns noisetolerant features guided by the corrected similarity scores produced from MSCN; while the MSCN predicts the reliable similarity score based on features from main net. This interdependency enables the main net and meta net reinforce each other against noise, and optimized via a bi-level optimization. Formally, the optimal parameters \\( W^{*} \\) is calculated by minimizing the cumulative loss over training data:" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.522, + 0.892, + 0.548 + ], + "angle": 0, + "content": "\\[\nW ^ {*} (\\Theta) = \\underset {W} {\\arg \\min } \\mathbb {E} _ {(I _ {i}, T _ {i}) \\in \\mathcal {D} _ {t r a i n}} l ^ {t r a i n} (I _ {i}, T _ {i}), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.556, + 0.774, + 0.573 + ], + "angle": 0, + "content": "where \\( l^{train}(\\cdot) \\) is the triplet ranking loss:" + }, + { + "type": "equation", + "bbox": [ + 0.503, + 0.59, + 0.891, + 0.638 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} l ^ {t r a i n} \\left(I _ {i}, T _ {i}\\right) = [ \\hat {\\gamma} - \\mathcal {V} _ {\\Theta} \\left(\\mathcal {F} _ {W} \\left(I _ {i}, T _ {i}\\right)\\right) + \\mathcal {V} _ {\\Theta} \\left(\\mathcal {F} _ {W} \\left(I _ {i}, T _ {i} ^ {-}\\right)\\right) ] _ {+} \\\\ + \\left[ \\hat {\\gamma} - \\mathcal {V} _ {\\Theta} \\left(\\mathcal {F} _ {W} \\left(I _ {i}, T _ {i}\\right)\\right) + \\mathcal {V} _ {\\Theta} \\left(\\mathcal {F} _ {W} \\left(I _ {i} ^ {-}, T _ {i}\\right)\\right) \\right] _ {+,} \\tag {3} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.639, + 0.892, + 0.744 + ], + "angle": 0, + "content": "where \\([x]_{+} = max(x,0), I_{i}^{-}\\) and \\(T_{i}^{-}\\) are the hardest negatives corresponding to given pair \\((I_i,T_i)\\) similar to VSE++ [8]. To achieve robust cross-modal retrieval, we design a self-adaptive margin \\(\\hat{\\gamma}\\) instead of a fixed value in VSE++. Ideally, the margin should put less value on the noisy pairs and more on the clean pairs. To this end, we adjust the margin according to the similarity score:" + }, + { + "type": "equation", + "bbox": [ + 0.622, + 0.751, + 0.892, + 0.786 + ], + "angle": 0, + "content": "\\[\n\\hat {\\gamma} = \\frac {1}{1 + \\left(\\frac {s _ {i}}{1 - s _ {i}}\\right) ^ {- \\tau}} \\gamma , \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.793, + 0.892, + 0.854 + ], + "angle": 0, + "content": "where \\(\\gamma\\) is the original margin value and \\(\\tau > 0\\) is the parameter. This self-adaptive margin will enable the triplet loss put more attention on the samples with high similarity scores that possibly are clean pairs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Meta Training Process. To ensure the reliability of the similarity scores produced by MSCN, we utilize a small" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7519" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.087, + 0.098, + 0.891, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.325, + 0.895, + 0.357 + ], + "angle": 0, + "content": "Figure 3. Flowchart of the proposed bi-level optimization. At each iteration, we alternately update the main net and the meta net (MSCN) following the steps in Eq. (7) - (9)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.372, + 0.473, + 0.751 + ], + "angle": 0, + "content": "amount clean meta-data set \\(\\mathcal{D}_{meta} = \\{(I_i,T_i)\\}_{i = 1}^M\\) to guide the training, where \\(M\\) is much smaller than \\(N\\). While meta-learning has been extensively studied in recent literature [10, 32, 46], it remains challenging to construct the meta-process in our objective. On the one hand, the similarity features from main net can not be used directly to produce loss. On the other hand, the meta-process should enable the output of MSCN to represent the similarity score. To achieve this, we view a binary classification task as the meta-process: given an image-text pair, the MSCN is learned to determine whether the modalities correspond to each other or not. Specifically, the positive pairs are taken from the meta-data which contain corresponding image and text, i.e. \\((I_{i},T_{i},y_{i} = 1)\\), while negative pairs are extracting from different pairs in training set to construct mismatched pairs, i.e. \\((I_{i},T_{j\\neq i},y_{i} = 0)\\). The binary labels are produced from the data itself, and can be regarded as the ideal similarity score to guide the training of MSCN. The prediction of MSCN measures the probability of a pair being clean which can be naturally equivalent to the similarity score. Formally, in each iteration of training, we randomly construct \\(M\\) mismatched pairs from training set to extend the meta-data set as \\(\\mathcal{D}_{meta}' = \\{(I_i,T_i,y_i)\\}_{i = 1}^{2M}\\). Then the optimal parameter \\(\\Theta^{*}\\) can be learned by minimizing the following meta-loss calculated on meta-data:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.77, + 0.47, + 0.796 + ], + "angle": 0, + "content": "\\[\n\\Theta^ {*} = \\underset {\\Theta} {\\arg \\min } \\mathbb {E} _ {(I _ {i}, T _ {i}) \\in \\mathcal {D} _ {m e t a} ^ {\\prime}} l ^ {m e t a} (I _ {i}, T _ {i}, y _ {i}), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.805, + 0.377, + 0.822 + ], + "angle": 0, + "content": "where \\(l^{meta}(\\cdot)\\) is cross-entropy loss [35, 37]:" + }, + { + "type": "equation", + "bbox": [ + 0.094, + 0.829, + 0.47, + 0.848 + ], + "angle": 0, + "content": "\\[\nl ^ {m e t a} (I _ {i}, T _ {i}, y _ {i}) = - y _ {i} \\cdot l o g \\mathcal {V} _ {\\Theta} \\left(\\mathcal {F} _ {W ^ {*} (\\Theta)} (I _ {i}, T _ {i})\\right). (6)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.471, + 0.903 + ], + "angle": 0, + "content": "As illustrated in Fig. 2, despite its simplicity, our proposed meta-process can not only enforce the MSCN to learn reliable similarity score but also has the following advantages:" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.372, + 0.895, + 0.51 + ], + "angle": 0, + "content": "1) previous works only guide the meta net with ground truth data; however, we leverage both positive and negative data as meta-knowledge that encourage the MSCN to learn discrimination from them. 2) We overcome one limitation of meta-learning methods that the lack of meta-data. We extract from different pairs among training set to construct massive negative meta-data. Moreover, it is more suitable to the rule of noisy correspondence, which the irregularity of noisy data causes in various patterns." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.527, + 0.895, + 0.678 + ], + "angle": 0, + "content": "Bi-level Optimization. Motivated by recent works [9, 16, 30, 32], we use a bi-level optimization strategy to guarantee the efficiency for updating our main net and MSCN. As illustrated in Fig. 3, it is practically solved via two nested loops of optimization. Specifically, at each iteration \\( t \\) of training, we sample a mini-batch of training pairs \\( \\{(I_i,T_i)\\}_{i = 1}^n \\), where \\( n \\) is the size of mini-batch. We first update the main net parameters by taking a gradient descent step toward the direction of triplet loss, which can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.686, + 0.893, + 0.726 + ], + "angle": 0, + "content": "\\[\n\\hat {W} ^ {(t)} (\\Theta) = W ^ {(t)} - \\alpha \\sum_ {i = 1} ^ {n} \\left. \\nabla_ {W} l ^ {\\text {t r a i n}} \\left(I _ {i}, T _ {i}\\right) \\right| _ {W ^ {(t)}}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.736, + 0.895, + 0.857 + ], + "angle": 0, + "content": "where \\(\\alpha\\) means the learning rate. The updated \\(\\hat{W}^{(t)}\\) can be regard as the approximation of the optimal \\(W^{*}(\\Theta)\\) at \\(t\\) iteration, containing the gradient information of \\(\\Theta\\). Then we update the parameters of MSCN with main net parameters fixed. Guided by the objective loss in Eq. (5) and a minibatch of meta-data \\(\\{(I_i,T_i,y_i)\\}_{i = 1}^m\\), where \\(m\\) is the minibatch size. The updated parameters of MSCN is calculated as:" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.865, + 0.893, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\left. \\Theta^ {(t + 1)} = \\Theta^ {(t)} - \\beta \\frac {1}{m} \\sum_ {i = 1} ^ {m} \\nabla_ {\\Theta} l ^ {m e t a} \\left(I _ {i}, T _ {i}, y _ {i}\\right) \\right| _ {\\Theta^ {(t)}}, \\tag {8}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7520" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.168 + ], + "angle": 0, + "content": "where \\(\\beta\\) is the learning rate for meta-process. Finally, we update the parameters of main net while fixing the MSCN parameters. With the corrected similarity scores produced by the newly updated \\(\\Theta^{(t + 1)}\\), we can update the main net parameters \\(W^{(t + 1)}\\) by employing the gradient decent as:" + }, + { + "type": "equation", + "bbox": [ + 0.103, + 0.18, + 0.469, + 0.219 + ], + "angle": 0, + "content": "\\[\nW ^ {(t + 1)} = W ^ {(t)} - \\alpha \\sum_ {i = 1} ^ {n} \\left. \\nabla_ {W} l ^ {\\text {t r a i n}} \\left(I _ {i}, T _ {i}\\right) \\right| _ {W ^ {(t)}}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.231, + 0.469, + 0.275 + ], + "angle": 0, + "content": "The above procedure proceeds in an alternating manner to optimize the main net and MSCN until it exceeds the maximal iteration." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.287, + 0.443, + 0.303 + ], + "angle": 0, + "content": "3.3. Meta-Knowledge Guided Data Purification" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.31, + 0.47, + 0.628 + ], + "angle": 0, + "content": "As mentioned, our MSCN can provide corrected similarity score to avoid noise. However, due to the property of triplet loss, it remains to produce positive loss for the noisy pairs even if we employ the ideal similarity score (e.g., \\(\\mathcal{V}(F(I_i,T_i))\\approx 0 < \\mathcal{V}\\left(F(I_i^-,T_i)\\right) + \\hat{\\gamma})\\), which results in the model fitting to the noise. To tackle this issue, we adopt sample selection strategy to remove the noisy data and only use clean samples as training data. Previous works [2, 42] motivated by the memorization effect of DNNs that clean sample has a lower loss during the beginning of training. Unfortunately, the triplet ranking loss is more complex, i.e., affected by both positive and negative samples, which leads to sub-optimal selection results. Based on the observation shown in Fig. 4, the clean and noisy pairs are prone to distinguish from the similarity scores. Gaussian Mixture Model (GMM) is a widely used technique to model mixture distribution [15, 20]. However, we find the similarity distribution exhibits high skew toward 1 and causes a poor approximation for GMM. Therefore, we fit a two-component Beta Mixture Model (BMM) [27] of per-pair similarity score \\(s_i\\) to better distinguish clean and noisy samples:" + }, + { + "type": "equation", + "bbox": [ + 0.175, + 0.641, + 0.469, + 0.681 + ], + "angle": 0, + "content": "\\[\np \\left(s _ {i}\\right) = \\sum_ {k = 1} ^ {K} \\lambda_ {k} \\phi \\left(s _ {i} \\mid \\alpha_ {k}, \\beta_ {k}\\right), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.469, + 0.734 + ], + "angle": 0, + "content": "where \\(K = 2\\) and \\(\\lambda_{k}\\) is the mixture coefficient, and \\(\\phi (s|\\alpha_k,\\beta_k)\\) denotes the probability density function with parameters \\(\\alpha_{k},\\beta_{k} > 0\\)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.469, + 0.902 + ], + "angle": 0, + "content": "We use the Expectation-Maximization [5] algorithm to optimize the BMM, which poses a crucial challenge: the initialization of model parameters. Existing methods typically adopt the \\(K\\)-Means algorithm [23] for initializing; however, it is time-consuming (i.e., the time complexity is \\(O(TKN)\\), where \\(T\\) is the number of iterations) and underperforming. We thus propose a novel initialization approach which takes the meta-data as prior knowledge. Specifically, we fit the two Beta components with positive and negative meta-data to model the distribution of clean and noisy pairs, respectively. Given the similarity scores \\(\\mathcal{S}_p = \\{s_i\\}_{i=1}^M\\)" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.093, + 0.887, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.294, + 0.892, + 0.337 + ], + "angle": 0, + "content": "Figure 4. Per-sample similarity score distribution, estimated BMM and BMM with meta prior knowledge under \\(20\\%\\) noisy correspondence in Flickr30K after 5 epochs warmup." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.36, + 0.891, + 0.39 + ], + "angle": 0, + "content": "corresponding to positive meta-data, the parameters of the component modeling clean distribution can be initialled by" + }, + { + "type": "equation", + "bbox": [ + 0.567, + 0.401, + 0.891, + 0.442 + ], + "angle": 0, + "content": "\\[\n\\alpha_ {k} = \\frac {\\left(1 - E \\left(\\mathcal {S} _ {p}\\right)\\right) E \\left(\\mathcal {S} _ {p}\\right) ^ {2}}{V \\left(\\mathcal {S} _ {p}\\right)} - E \\left(\\mathcal {S} _ {p}\\right), \\tag {11}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.439, + 0.729, + 0.472 + ], + "angle": 0, + "content": "\\[\n\\beta_ {k} = \\frac {\\alpha_ {k} \\left(1 - E \\left(\\mathcal {S} _ {p}\\right)\\right)}{E (\\mathcal {S} _ {p})},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.483, + 0.892, + 0.588 + ], + "angle": 0, + "content": "where \\( E(\\cdot) \\) and \\( V(\\cdot) \\) are the mean and variance function, respectively. The initialization of component modeling noisy distribution is analogous but uses similarity scores \\( S_{n} = \\{s_{i}\\}_{i=1}^{M} \\) of negative meta-data. As illustrated in Fig. 4, our meta-knowledge guided initialization method achieves better approximation while with lower time complexity (i.e., \\( O(KM) \\))." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.589, + 0.892, + 0.647 + ], + "angle": 0, + "content": "Finally, we compute the posterior probability \\( p_i = p(k|s_i) = p(k)p(s_i|\\alpha_k,\\beta_k) / p(s_i) \\) as the probability of \\( i \\)-th simple being clean. The purified training set is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.564, + 0.649, + 0.891, + 0.666 + ], + "angle": 0, + "content": "\\[\n\\mathcal {D} _ {\\text {t r a i n}} ^ {\\prime} = \\left\\{\\left(I _ {i}, T _ {i}\\right) \\in \\mathcal {D} _ {\\text {t r a i n}} \\mid p _ {i} > 0. 5 \\right\\}. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.673, + 0.892, + 0.764 + ], + "angle": 0, + "content": "At each iteration, we produce the purified dataset before the training of our meta-learning objective. Moreover, following [12, 15, 20], we maintain two networks \\(\\{\\mathcal{F}_W^1,\\mathcal{V}_{\\Theta}^1\\}\\) and \\(\\{\\mathcal{F}_W^2,\\mathcal{V}_{\\Theta}^2\\}\\) to avoid error accumulation that one network will produce the purified training set to train the other one. The full algorithm is shown in supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.777, + 0.625, + 0.794 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.802, + 0.792, + 0.816 + ], + "angle": 0, + "content": "4.1. Datasets and Evaluation Protocol" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Datasets. We evaluate our method on three standard image-text retrieval datasets. To be specific, the Flickr30K [43] collects 31,783 images with 5 corresponding captions each from Flickr website. Following the split in [15], we use 1,000 image-text pairs for validation, 1,000 image-text" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "7521" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.273 + ], + "angle": 0, + "content": "pairs for testing, and the rest for training. The MS-COCO [24] contains 123,287 images with 5 corresponding captions each. We split the MS-COCO following [15] that 113,287 image-text pairs for training, 5,000 pairs for validation, and 5,000 pairs for testing. Conceptual Captions [31] is a large-scale dataset contains 3.3M images with one corresponding caption each. We employ the same subset following [15], i.e., CC152K. Specifically, CC152K consists of 150,000 pairs from training set for training, 1,000 pairs from validation set for validation, and 1,000 pairs from validation set for testing. Moreover, we use an extra meta-data set from validation containing 3,000 pairs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.298, + 0.472, + 0.419 + ], + "angle": 0, + "content": "Evaluation Protocol. We evaluate the retrieval performance with the recall rate at K (R@K) metric. In a nutshell, R@K is the proportion for the relevant items retrieved in the closest K items to the query. In our experiments, we take image and text as queries, respectively, and report R@1, R@5, and R@10 results for a comprehensive evaluation. Following [15], we average the similarity scores from two networks at the inference phase." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.432, + 0.295, + 0.448 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.456, + 0.473, + 0.759 + ], + "angle": 0, + "content": "As a general approach, our method can be easily applied to almost all cross-modal retrieval methods to improve robustness. For fair comparison, we chose the same network backbones with NCR [15], i.e., the projection modules and similarity function. For all datasets, the training processes contain 50 epochs after a 5 epochs warmup. We employ ADAM [17] as the optimizer for both main net and meta net with a batch size of 64. To fit the BMM, we set the stop threshold as \\(10^{-2}\\) and maximum number of iterations as 10 for the EM procedure. Note that, we clamp the similarity scores into a range of \\([10^{-4}, 1 - 10^{-4}]\\) instead of [0,1] for stable fitting. Moreover, we set the margin parameter as 0.2 and \\(\\tau = 2\\) to calculate the self-adaptive margin. For all datasets, the number of ground-truth meta-data is approximately \\(2\\%\\) of training data, and the learning rate of main net and meta net are initialized with \\(2 \\times 10^{-4}\\) and \\(1.7 \\times 10^{-5}\\), respectively. For CC152K containing real noise, we employ the validation data as the meta set due to the lack of ground-truth data. For all experiments, we decay the learning rate by 0.1 after 30 epochs." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.771, + 0.405, + 0.788 + ], + "angle": 0, + "content": "4.3. Comparison with the State-of-the-Art" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.47, + 0.902 + ], + "angle": 0, + "content": "In this section, we conduct comparison experiments with 7 state-of-the-art methods that include SCAN [18], VSRN [22], IMRAM [4], SGR, SAF, SGRAF [7], and NCR [15]. As Flickr30K and MS-COCO are well annotated datasets, we generate synthetic noisy correspondence by randomly shuffling the training images and captions for a specific noise ratio, and we conduct experiments with three different" + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.09, + 0.898, + 0.436 + ], + "angle": 0, + "content": "
Image to Text
MethodsR@1R@5R@10SUM
SCAN (ECCV'18)30.555.365.3151.1
VSRN(ICCV'19)32.661.370.5164.4
IMRAM(CVPR'20)33.157.668.1158.8
SAF (AAAI'21)31.759.368.2159.2
SGR (AAAI'21)11.329.739.680.6
SGR* (AAAI'21)35.063.473.3171.7
NCR(NIPS'21)39.564.573.5177.5
MSCN*(Ours)39.765.475.3180.4
MSCN(Ours)40.165.776.6182.4
Text to Image
SCAN (ECCV'18)26.953.064.7144.6
VSRN(ICCV'19)32.559.470.4162.3
IMRAM(CVPR'20)29.056.867.4153.2
SAF (AAAI'21)31.959.067.9158.8
SGR (AAAI'21)13.130.141.684.8
SGR* (AAAI'21)34.963.072.8170.7
NCR(NIPS'21)40.364.673.2178.1
MSCN*(Ours)39.866.175.0180.9
MSCN(Ours)40.667.476.3184.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.446, + 0.892, + 0.473 + ], + "angle": 0, + "content": "Table 1. Image-Text Retrieval Performance on CC152K with real noise, and the best results are highlighted in bold." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.486, + 0.892, + 0.653 + ], + "angle": 0, + "content": "level of noise ratios, i.e., \\(20\\%\\), \\(50\\%\\), and \\(70\\%\\). For all methods, we choose the best checkpoint on the validation set and report its performance on the test set. Following NCR, we also report two strong baselines based on SGR, i.e., SGR-C and SGR*. In short, SGR-C uses only clean data for training, and SGR* employs a pre-training process while training without hard negatives to improve robustness. To evaluate our method on real-world noisy data, we conduct experiments on the CC152K, which is harvested from the Internet and contains about \\(3\\% \\sim 20\\%\\) noisy correspondence [31]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.671, + 0.892, + 0.837 + ], + "angle": 0, + "content": "Experiments on CC152K. Tab. 1 shows the results on CC152K with real noisy correspondence. On CC152K, our MSCN achieves the state-of-the-art performance in terms of all metrics. Specifically, the sum score of MSCN is \\(4.9\\%\\) and \\(6.2\\%\\) higher than the best baseline in text and image retrieval, respectively. In addition, we also present a more strict comparison (denoted by \\(\\mathrm{MSCN^{*}}\\)) that trained our method with only 1,000 meta-data (approximately \\(0.67\\%\\) of all data). We can see that our \\(\\mathrm{MSCN^{*}}\\) also achieves a competitive performance. Specifically, all metrics except R@1 of image retrieval are superior to the baselines." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Experiments on Flickr30K and MS-COCO. Tab. 2 shows the results on Flickr30K and MS-COCO under a range of synthetic noise rates. Following [15], we report the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7522" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.1, + 0.09, + 0.88, + 0.486 + ], + "angle": 0, + "content": "
Flickr30KMS-COCO
Image to TextText to ImageImage to TextText to Image
NoiseMethodsR@1R@5R@10R@1R@5R@10R@1R@5R@10R@1R@5R@10
20%SCAN59.183.490.436.667.077.566.291.096.445.080.289.3
VSRN58.182.689.340.768.778.225.159.074.817.649.064.1
IMRAM63.086.091.341.471.280.568.692.897.655.785.091.0
SAF51.079.388.038.366.576.267.392.596.653.484.592.4
SGR*62.886.292.244.472.380.467.891.796.252.983.590.1
SGR-C72.890.895.456.482.188.675.495.297.960.188.594.8
NCR75.093.997.558.383.089.077.795.598.262.589.395.3
MSCN77.494.997.659.683.289.278.197.298.864.390.495.8
50%SCAN27.757.668.816.239.349.840.873.584.95.415.121.0
VSRN14.337.650.012.130.039.423.554.769.316.047.865.9
IMRAM9.126.638.22.78.412.721.360.275.922.352.864.3
SAF30.363.675.427.953.765.130.467.882.333.569.082.8
SGR*36.968.180.229.356.267.060.687.493.646.074.279.0
SGR-C69.890.394.850.177.585.271.794.197.757.086.693.7
NCR72.993.096.354.379.886.574.694.697.859.187.894.5
MSCN74.493.296.055.380.486.877.596.298.760.789.194.9
70%SCAN5.619.327.42.28.012.818.143.157.40.31.32.3
VSRN0.82.54.10.51.52.75.115.724.62.58.813.3
IMRAM1.33.13.90.31.22.87.120.033.45.315.222.0
SAF0.52.23.00.20.81.70.11.74.00.61.93.0
SGR*17.942.151.914.631.040.835.771.285.431.665.879.0
SGR-C65.089.394.748.174.581.169.893.697.556.586.093.4
NCR16.138.552.811.029.541.435.469.583.431.566.481.1
MSCN69.089.393.049.273.179.074.494.997.758.887.293.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.497, + 0.893, + 0.527 + ], + "angle": 0, + "content": "Table 2. Image-Text Retrieval Performance under synthetic noise rates of \\( {20}\\% ,{50}\\% \\) and \\( {70}\\% \\) on Flickr30K and MS-COCO 1K,and the best results are highlighted in bold." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.538, + 0.473, + 0.766 + ], + "angle": 0, + "content": "results by averaging over 5 folds of 1K test images. From Tab. 2, we can draw the observation that noisy correspondence remarkably affects the performance of cross-modal retrieval methods. The retrieval accuracy will decrease fast with the noise rate increasing. Our MSCN is superior to the baselines in almost all metrics under different noise rates. In the low noise cases, i.e., \\(20\\%\\) and \\(50\\%\\), our MSCN improves R@1 (sum of image and text) by \\(3.7\\%\\), \\(2.2\\%\\), \\(2.5\\%\\), and \\(4.5\\%\\) in the four valuations compared with NCR. In the high noise case, i.e., \\(70\\%\\), all baselines can easily overfit to noisy correspondence and results in poor performance. Our MSCN improves R@1 by \\(52.9\\%\\), \\(38.2\\%\\), \\(39.4\\%\\), and \\(27.3\\%\\) in these four valuations compared with NCR. Even compared with the SGR-C trained only on clean data, our MSCN also achieves competitive performance." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.772, + 0.233, + 0.789 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.471, + 0.904 + ], + "angle": 0, + "content": "To evaluate the performance of the proposed components (i.e., self-adaptive margin and data purification), we conduct the ablation study on the Flickr30K with \\(20\\%\\) noisy correspondence. Note that for MSCN without \\(\\mathcal{D}_{train}^{\\prime}\\), we only use a single model to perform the experiment. As shown in Tab. 3, we can see that the performance without \\(\\hat{\\gamma}\\) or \\(\\mathcal{D}_{train}^{\\prime}\\) are worse than the complete MSCN, which" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.538, + 0.893, + 0.584 + ], + "angle": 0, + "content": "indicates that all components are important to achieve advantageous results. We explain more on the data selection strategy in the supplementary material." + }, + { + "type": "table", + "bbox": [ + 0.51, + 0.594, + 0.892, + 0.762 + ], + "angle": 0, + "content": "
MethodImage to Text
MSCNw/o γw/o D'trainR@1R@5R@10
✓77.494.997.6
✓✓75.394.597.2
✓✓75.893.896.2
✓✓✓74.191.594.7
Text to Image
✓59.683.289.2
✓✓58.383.188.9
✓✓55.874.577.4
✓✓✓53.471.272.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.516, + 0.772, + 0.877, + 0.787 + ], + "angle": 0, + "content": "Table 3. Ablation studies on Flickr30K with \\( {20}\\% \\) noise rate." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.802, + 0.725, + 0.818 + ], + "angle": 0, + "content": "4.5. Progressive Comparison" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Fig. 6 plots the average of recalls on testing set of MSCN and NCR as training proceeds. We show a representative result using Flickr30K with \\(50\\%\\) noisy correspondence. From the result, we can see that our MSCN achieves a more stable performance in overall training process. Although NCR" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.958 + ], + "angle": 0, + "content": "7523" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.1, + 0.223, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.094, + 0.18, + 0.222, + 0.205 + ], + "angle": 0, + "content": "the circular movement of the camera. (0.3444)" + }, + { + "type": "image", + "bbox": [ + 0.23, + 0.109, + 0.348, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.232, + 0.18, + 0.348, + 0.204 + ], + "angle": 0, + "content": "person arrives earlier this month. (0.3485)" + }, + { + "type": "image", + "bbox": [ + 0.366, + 0.108, + 0.476, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.356, + 0.18, + 0.487, + 0.204 + ], + "angle": 0, + "content": "i expected to learn more about plants. (0.3487)" + }, + { + "type": "image", + "bbox": [ + 0.495, + 0.11, + 0.611, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.491, + 0.181, + 0.615, + 0.205 + ], + "angle": 0, + "content": "man,chairman to speak at show (0.3488)" + }, + { + "type": "image", + "bbox": [ + 0.628, + 0.108, + 0.739, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.625, + 0.18, + 0.743, + 0.204 + ], + "angle": 0, + "content": "note the many parked cars (0.3511)" + }, + { + "type": "image", + "bbox": [ + 0.76, + 0.11, + 0.87, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.752, + 0.18, + 0.88, + 0.205 + ], + "angle": 0, + "content": "a man using his mobile phone. (0.3518)" + }, + { + "type": "image_caption", + "bbox": [ + 0.301, + 0.232, + 0.667, + 0.246 + ], + "angle": 0, + "content": "Figure 5. Real-world noisy examples detected by our MSCN." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.263, + 0.469, + 0.324 + ], + "angle": 0, + "content": "does not need an extra meta-set to guide the training, it relies on a clean validation set to choose the best model parameters. The stableness also indicates that our method has better alleviated the interference of noisy correspondence." + }, + { + "type": "image", + "bbox": [ + 0.127, + 0.34, + 0.422, + 0.486 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.497, + 0.469, + 0.525 + ], + "angle": 0, + "content": "Figure 6. Progressive performance comparison of MSCN and NCR on testing set as training proceeds." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.539, + 0.296, + 0.555 + ], + "angle": 0, + "content": "4.6. Detected Noisy Samples" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.562, + 0.469, + 0.713 + ], + "angle": 0, + "content": "Fig. 5 shows some real-world noisy image-text pairs detected (with low similarity score) by MSCN in Conceptual Captions. We also present the similarity score predicted by our MSCN. For most noisy pairs, the image and text contains completely irrelevant semantic information. Moreover, our MSCN can even find the subtle mistake in the noisy image-text pair. In the last pair, the picture shows a man covering his ears but wrongly caption as using a mobile phone. Although this picture is misleading, our method successfully detects the subtle difference." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.72, + 0.367, + 0.736 + ], + "angle": 0, + "content": "4.7. Visualization on Similarity Score" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.744, + 0.469, + 0.834 + ], + "angle": 0, + "content": "Fig. 7 plots the similarity score distribution predicted by MSCN for clean and noisy training samples. It can be seen that almost all large similarity scores belong to clean pairs, and the noisy pairs' scores are smaller than clean samples, which implies that the trained MSCN can successfully provide corrected similarity scores." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.846, + 0.195, + 0.86 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "In this paper, we explore a meta-learning method to address the problem of learning with noisy correspond-" + }, + { + "type": "image", + "bbox": [ + 0.509, + 0.265, + 0.692, + 0.438 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.59, + 0.446, + 0.608, + 0.457 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.702, + 0.265, + 0.885, + 0.438 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.783, + 0.446, + 0.8, + 0.457 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.46, + 0.892, + 0.501 + ], + "angle": 0, + "content": "Figure 7. Per-sample similarity distribution on training data of Flickr30K. (a) \\(20\\%\\) noisy correspondence. (b) \\(50\\%\\) noisy correspondence." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.515, + 0.892, + 0.667 + ], + "angle": 0, + "content": "dence. Specifically, we propose a meta correction network (MSCN) to provide reliable similarity scores. Our MSCN is trained by a novel meta-process that views both positive and negative data as meta-knowledge to encourage MSCN to learn discrimination from them. To further mitigate noise interference, we design a data purification strategy that uses meta-data as prior knowledge to purify the noisy data efficiently. We conduct comprehensive experiments on three widely used datasets. The results validate the effectiveness of our method in both synthetic and real noises." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.68, + 0.667, + 0.696 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This work is supported by the National Key Research and Development Program of China (2020AAA0108800), National Nature Science Foundation of China (62192781, 62272374, 62202367, 62250009, 62137002, 61937001), Innovative Research Group of the National Natural Science Foundation of China (61721002), Innovation Research Team of Ministry of Education (IRT_17R86), Project of China Knowledge Center for Engineering Science and Technology, and Project of Chinese academy of engineering \"The Online and Offline Mixed Educational Service System for 'The Belt and Road' Training in MOOC China\". We would like to express our gratitude for the support of K. C. Wong Education Foundation." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7524" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.469, + 0.157 + ], + "angle": 0, + "content": "[1] Relja Arandjelovic and Andrew Zisserman. Objects that sound. In Proceedings of the European conference on computer vision (ECCV), pages 435-451, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.159, + 0.469, + 0.213 + ], + "angle": 0, + "content": "[2] Eric Arazo, Diego Ortego, Paul Albert, Noel O'Connor, and Kevin McGuinness. Unsupervised label noise modeling and loss correction. In International conference on machine learning, pages 312-321. PMLR, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.215, + 0.469, + 0.282 + ], + "angle": 0, + "content": "[3] Devansh Arpit, Stanislaw K Jastrzebski, Nicolas Ballas, David Krueger, Emmanuel Bengio, Maxinder S Kanwal, Tegan Maharaj, Asja Fischer, Aaron C Courville, Yoshua Bengio, et al. A closer look at memorization in deep networks. In ICML, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.284, + 0.469, + 0.354 + ], + "angle": 0, + "content": "[4] Hui Chen, Guiguang Ding, Xudong Liu, Zijia Lin, Ji Liu, and Jungong Han. Imram: Iterative matching with recurrent attention memory for cross-modal image-text retrieval. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12655-12663, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.355, + 0.469, + 0.409 + ], + "angle": 0, + "content": "[5] Arthur P Dempster, Nan M Laird, and Donald B Rubin. Maximum likelihood from incomplete data via the em algorithm. Journal of the Royal Statistical Society: Series B (Methodological), 39(1):1-22, 1977. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.41, + 0.469, + 0.464 + ], + "angle": 0, + "content": "[6] Cheng Deng, Zhaojia Chen, Xianglong Liu, Xinbo Gao, and Dacheng Tao. Triplet-based deep hashing network for cross-modal retrieval. IEEE Transactions on Image Processing, 27(8):3893-3903, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.466, + 0.469, + 0.522 + ], + "angle": 0, + "content": "[7] Haiwen Diao, Ying Zhang, Lin Ma, and Hutchuan Lu. Similarity reasoning and filtration for image-text matching. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 1218-1226, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.523, + 0.469, + 0.564 + ], + "angle": 0, + "content": "[8] Fartash Faghri, David J Fleet, Jamie Ryan Kiros, and Sanja Fidler. Vse++: Improving visual-semantic embeddings with hard negatives. arXiv preprint arXiv:1707.05612, 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.565, + 0.469, + 0.619 + ], + "angle": 0, + "content": "[9] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In International conference on machine learning, pages 1126–1135. PMLR, 2017. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.621, + 0.469, + 0.689 + ], + "angle": 0, + "content": "[10] Luca Franceschi, Paolo Frasconi, Saverio Salzo, Riccardo Grazzi, and Massimiliano Pontil. Bilevel programming for hyperparameter optimization and meta-learning. In International Conference on Machine Learning, pages 1568-1577. PMLR, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.691, + 0.469, + 0.746 + ], + "angle": 0, + "content": "[11] Dalu Guo, Chang Xu, and Dacheng Tao. Image-question-answer synergistic network for visual dialog. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10434-10443, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.747, + 0.469, + 0.816 + ], + "angle": 0, + "content": "[12] Bo Han, Quanming Yao, Xingrui Yu, Gang Niu, Miao Xu, Weihua Hu, Ivor Tsang, and Masashi Sugiyama. Co-teaching: Robust training of deep neural networks with extremely noisy labels. Advances in neural information processing systems, 31, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[13] Haochen Han, Qinghua Zheng, Minnan Luo, Kaiyao Miao, Feng Tian, and Yan Chen. Noise-tolerant learning for audiovisual action recognition. arXiv preprint arXiv:2205.07611, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[14] Peng Hu, Xi Peng, Hongyuan Zhu, Liangli Zhen, and Jie Lin. Learning cross-modal retrieval with noisy labels. In" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5403-5413, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[15] Zhenyu Huang, Guocheng Niu, Xiao Liu, Wenbiao Ding, Xinyan Xiao, Hua Wu, and Xi Peng. Learning with noisy correspondence for cross-modal matching. Advances in Neural Information Processing Systems, 34:29406-29419, 2021. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.192, + 0.892, + 0.262 + ], + "angle": 0, + "content": "[16] Haoxuanye Ji, Le Wang, Sanping Zhou, Wei Tang, Nanning Zheng, and Gang Hua. Meta pairwise relationship distillation for unsupervised person re-identification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3661–3670, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.263, + 0.892, + 0.29 + ], + "angle": 0, + "content": "[17] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR (Poster), 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.292, + 0.892, + 0.347 + ], + "angle": 0, + "content": "[18] Kuang-Huei Lee, Xi Chen, Gang Hua, Houdong Hu, and Xiaodong He. Stacked cross attention for image-text matching. In Proceedings of the European conference on computer vision (ECCV), pages 201-216, 2018. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.349, + 0.892, + 0.404 + ], + "angle": 0, + "content": "[19] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy Hospedales. Learning to generalize: Meta-learning for domain generalization. In Proceedings of the AAAI conference on artificial intelligence, volume 32, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.406, + 0.892, + 0.46 + ], + "angle": 0, + "content": "[20] Junnan Li, Richard Socher, and Steven CH Hoi. Dividemix: Learning with noisy labels as semi-supervised learning. In International Conference on Learning Representations, 2019. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.462, + 0.892, + 0.517 + ], + "angle": 0, + "content": "[21] Junnan Li, Yongkang Wong, Qi Zhao, and Mohan S Kankanhalli. Learning to learn from noisy labeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5051-5059, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.519, + 0.892, + 0.574 + ], + "angle": 0, + "content": "[22] Kunpeng Li, Yulun Zhang, Kai Li, Yuanyuan Li, and Yun Fu. Visual semantic reasoning for image-text matching. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4654-4662, 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.576, + 0.892, + 0.616 + ], + "angle": 0, + "content": "[23] Aristidis Likas, Nikos Vlassis, and Jakob J Verbeek. The global k-means clustering algorithm. Pattern recognition, 36(2):451-461, 2003. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.618, + 0.892, + 0.687 + ], + "angle": 0, + "content": "[24] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.689, + 0.892, + 0.758 + ], + "angle": 0, + "content": "[25] Chunxiao Liu, Zhendong Mao, Tianzhu Zhang, Hongtao Xie, Bin Wang, and Yongdong Zhang. Graph structured network for image-text matching. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10921-10930, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.76, + 0.892, + 0.801 + ], + "angle": 0, + "content": "[26] Lin Ma, Zhengdong Lu, and Hang Li. Learning to answer questions from image using convolutional neural network. In Thirtieth AAAI Conference on Artificial Intelligence, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.802, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[27] Zhanyu Ma and Arne Leijon. Bayesian estimation of beta mixture models with variational inference. IEEE Transactions on Pattern Analysis and Machine Intelligence, 33(11):2160-2173, 2011. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.859, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[28] Nikhil Mishra, Mostafa Rohaninejad, Xi Chen, and Pieter Abbeel. A simple neural attentive meta-learner. In International Conference on Learning Representations, 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7525" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.16 + ], + "angle": 0, + "content": "[29] Yang Qin, Dezhong Peng, Xi Peng, Xu Wang, and Peng Hu. Deep evidential learning with noisy correspondence for cross-modal retrieval. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4948-4956, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.161, + 0.47, + 0.217 + ], + "angle": 0, + "content": "[30] Mengye Ren, Wenyuan Zeng, Bin Yang, and Raquel Urtasun. Learning to reweight examples for robust deep learning. In International conference on machine learning, pages 4334-4343. PMLR, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.218, + 0.471, + 0.3 + ], + "angle": 0, + "content": "[31] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2556-2565, 2018. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.301, + 0.471, + 0.357 + ], + "angle": 0, + "content": "[32] Jun Shu, Qi Xie, Lixuan Yi, Qian Zhao, Sanping Zhou, Zongben Xu, and Deyu Meng. Meta-weight-net: Learning an explicit mapping for sample weighting. Advances in neural information processing systems, 32, 2019. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.358, + 0.47, + 0.425 + ], + "angle": 0, + "content": "[33] Didac Surís, Amanda Duarte, Amaia Salvador, Jordi Torres, and Xavier Giró-i Nieto. Cross-modal embeddings for video and audio retrieval. In Proceedings of the European Conference on Computer Vision (ECCV) Workshops, pages 0–0, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.427, + 0.47, + 0.468 + ], + "angle": 0, + "content": "[34] Jun Tang, Ke Wang, and Ling Shao. Supervised matrix factorization hashing for cross-modal retrieval. IEEE Transactions on Image Processing, 25(7):3157-3166, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.469, + 0.47, + 0.524 + ], + "angle": 0, + "content": "[35] Bokun Wang, Yang Yang, Xing Xu, Alan Hanjalic, and Heng Tao Shen. Adversarial cross-modal retrieval. In Proceedings of the 25th ACM international conference on Multimedia, pages 154-162, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.525, + 0.47, + 0.58 + ], + "angle": 0, + "content": "[36] Liwei Wang, Yin Li, and Svetlana Lazebnik. Learning deep structure-preserving image-text embeddings. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5005-5013, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.58, + 0.47, + 0.635 + ], + "angle": 0, + "content": "[37] Yunchao Wei, Yao Zhao, Canyi Lu, Shikui Wei, Luoqi Liu, Zhenfeng Zhu, and Shuicheng Yan. Cross-modal retrieval with cnn visual features: A new baseline. IEEE transactions on cybernetics, 47(2):449-460, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.636, + 0.47, + 0.677 + ], + "angle": 0, + "content": "[38] Xing Xu, Li He, Huimin Lu, Lianli Gao, and Yanli Ji. Deep adversarial metric learning for cross-modal retrieval. *World Wide Web*, 22(2):657-672, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.678, + 0.47, + 0.747 + ], + "angle": 0, + "content": "[39] Mouxing Yang, Zhenyu Huang, Peng Hu, Taihao Li, Jiancheng Lv, and Xi Peng. Learning with twin noisy labels for visible-infrared person re-identification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14308-14317, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.748, + 0.47, + 0.803 + ], + "angle": 0, + "content": "[40] Mouxing Yang, Yunfan Li, Peng Hu, Jinfeng Bai, Jiancheng Lv, and Xi Peng. Robust multi-view clustering with incomplete information. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(1):1055-1069, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.471, + 0.873 + ], + "angle": 0, + "content": "[41] Mouxing Yang, Yunfan Li, Zhenyu Huang, Zitao Liu, Peng Hu, and Xi Peng. Partially view-aligned representation learning with noise-robust contrastive loss. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1134–1143, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[42] Quanming Yao, Hansi Yang, Bo Han, Gang Niu, and James Tin-Yau Kwok. Searching to exploit memorization effect in" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "learning with noisy labels. In International Conference on Machine Learning, pages 10789-10798. PMLR, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.121, + 0.894, + 0.19 + ], + "angle": 0, + "content": "[43] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics, 2:67-78, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.191, + 0.894, + 0.248 + ], + "angle": 0, + "content": "[44] Qi Zhang, Zhen Lei, Zhaoxiang Zhang, and Stan Z Li. Context-aware attention network for image-text retrieval. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3536-3545, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.248, + 0.894, + 0.304 + ], + "angle": 0, + "content": "[45] Liangli Zhen, Peng Hu, Xu Wang, and Dezhong Peng. Deep supervised cross-modal retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10394-10403, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.305, + 0.894, + 0.361 + ], + "angle": 0, + "content": "[46] Guoqing Zheng, Ahmed Hassan Awadallah, and Susan Dumais. Meta label correction for noisy label learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 11053-11061, 2021. 2, 4" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7526" + } + ] +] \ No newline at end of file diff --git a/2023/Noisy Correspondence Learning With Meta Similarity Correction/0167b343-6cc5-4960-bab8-e11934c107f0_origin.pdf b/2023/Noisy Correspondence Learning With Meta Similarity Correction/0167b343-6cc5-4960-bab8-e11934c107f0_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..9bce57d247698ccdfbcc46f31cd4fcd06b7bfa3c --- /dev/null +++ b/2023/Noisy Correspondence Learning With Meta Similarity Correction/0167b343-6cc5-4960-bab8-e11934c107f0_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ab75cdf0e5159b6b3ce5fb75e34f8a7bbc2f52d809ebe9b43452bdbf02746bcb +size 7898674 diff --git a/2023/Noisy Correspondence Learning With Meta Similarity Correction/full.md b/2023/Noisy Correspondence Learning With Meta Similarity Correction/full.md new file mode 100644 index 0000000000000000000000000000000000000000..623dc541cd5ec18f0f084c7ff47d0903374eb4bc --- /dev/null +++ b/2023/Noisy Correspondence Learning With Meta Similarity Correction/full.md @@ -0,0 +1,322 @@ +# Noisy Correspondence Learning with Meta Similarity Correction + +Haochen Han $^{1}$ , Kaiyao Miao $^{2}$ , Qinghua Zheng $^{1}$ , Minnan Luo $^{1*}$ + +$^{1}$ School of Computer Science and Technology, Xi'an Jiaotong University, China + +$^{2}$ School of Cyber Science and Engineering, Xi'an Jiaotong University, China + +{hhc1997, miaoky814}@stu.xjtu.edu.cn, {qhzheng, minnluo}@xjtu.edu.cn + +# Abstract + +Despite the success of multimodal learning in cross-modal retrieval task, the remarkable progress relies on the correct correspondence among multimedia data. However, collecting such ideal data is expensive and time-consuming. In practice, most widely used datasets are harvested from the Internet and inevitably contain mismatched pairs. Training on such noisy correspondence datasets causes performance degradation because the cross-modal retrieval methods can wrongly enforce the mismatched data to be similar. To tackle this problem, we propose a Meta Similarity Correction Network (MSCN) to provide reliable similarity scores. We view a binary classification task as the meta-process that encourages the MSCN to learn discrimination from positive and negative meta-data. To further alleviate the influence of noise, we design an effective data purification strategy using meta-data as prior knowledge to remove the noisy samples. Extensive experiments are conducted to demonstrate the strengths of our method in both synthetic and real-world noises, including Flickr30K, MS-COCO, and Conceptual Captions. Our code is publicly available. $^{1}$ + +# 1. Introduction + +Recently, cross-modal retrieval has drawn much attention with the rapid growth of multimedia data. Given a query sample of specific modality, cross-modal retrieval aims to retrieve relevant samples across different modalities. Existing cross-modal retrieval works [1, 6, 11] usually learn a comparable common space to bridge different modalities, which achieved remarkable progress in many applications, including video-audio retrieval [1, 33], visual question answering [11, 26], and image-text matching [44]. + +Despite the promise, a core assumption in cross-modal retrieval is the correct correspondence among multiple + +![](images/6f5725cb576e78f8305df691f3e3a92564c58cd94e044e587fff66d835979a4a.jpg) +Figure 1. Illustration of noisy correspondence in image-text retrieval. A standard triplet loss is used to enforce the positive pairs to be closer than negatives in the common space. Noisy correspondence is the mismatched pairs but wrongly considered as positive ones, and thus results in model performance degradation. + +modalities. However, collecting such ideal data is expensive and time-consuming. In practice, most widely used datasets are harvested from the Internet and inevitably contain noisy correspondence [31]. As illustrated in Fig. 1, the cross-modal retrieval method will wrongly enforce the mismatched data to be similar when learning with noisy correspondence, which may significantly affect the retrieval performance. To date, few effort has been made to address this. Huang [15] first researches this issue and proposes the NCR method to train from the noisy image-text pairs robustly. Inspired by the prior success for noisy labels [20], NCR divides the data into clean and noisy partitions and rectifies the correspondence with an adaptive model. However, NCR based on the memorization effect of DNNs [3], which leads to poor performance under high noise ratio. + +To tackle the challenge, we propose a Meta Similarity Correction Network (MSCN) which aims to provide reliable similarity scores for the noisy features from main net. We view a binary classification task as the meta-process: given a multimodal sample, the MSCN will learn to determine whether the modalities correspond to each other or not, where the prediction of MSCN can be naturally regarded as the similarity score. Specifically, a small amount of clean pairs is used to construct positive and negative meta-data, both viewed as meta-knowledge that encourages MSCN to learn the discrimination. Meanwhile, the main net + +trained by the corrected similarity score from MSCN with a self-adaptive margin to achieve robust learning. This interdependency makes the main net and MSCN benefits each other against noise. However, due to the property of triplet loss, it remains to produce positive loss for noisy pairs even if we employ the ideal similarity scores. To this end, we further propose a meta-knowledge guided data purification strategy to remove samples with potentially wrong correspondence. Extensive experiments are conducted to demonstrate the strengths of our method in both synthetic and real-world noises. + +The main contributions of this work are summarized as follows: + +- We pioneer the exploration of meta-learning for noisy correspondence problem, where a meta correction network is proposed to provide reliable similarity scores against noise. +- We present a novel meta-process that first considers both positive and negative data as meta-knowledge to encourage the MSCN to learn discrimination from them. +- We design an effective data purification strategy using meta-data as prior knowledge to remove the noisy samples. + +# 2. Related Work + +# 2.1. Cross-Modal Retrieval + +Cross-modal retrieval works aim to take one modality of data as query to retrieve the related data in other modalities. Based on the utilization of annotation information, cross-modal retrieval methods can be roughly divided into two categories: 1) Unsupervised Methods [8,22,25,36]. It only uses the co-occurrence information to learn the common representations among the multimodal data. 2) Supervised methods [14,34,38,45]. The extra label information is employed to boost the discrimination of common representations. However, the lack of annotations limited the practicability in the real world. + +Our proposed approach falls in the category of unsupervised methods. For example, Wang [36] uses a two-branch neural networks to learn the joint embeddings of multimodal data. Inspired by hard negative mining, $\mathrm{VSE}++$ [8] uses hard negatives to improve the retrieval performance. To capture fine-grained interplay between modalities, SCAN [18] introduce stacked cross attention to enable attention with context from both image and sentence. Recently, motivated by the powerful learning ability of graph model, GSMN [25] and SGRAF [7] construct graph structure for multimodal data to benefit the learning of fine-grained correspondence. Although existing works have achieved remarkable results, they usually depend on the correct correspondence among multimodal data and cannot + +tackle the noise issue. Thus, it is significant to explore how to learn cross-modal retrieval with noisy correspondence, but which is rarely touched in previous studies. + +# 2.2. Noisy Correspondence Learning + +As a newly proposed problem, noisy correspondence is the mismatched pairs but wrongly considered as positive ones. Huang [15] first research this issue and propose the NCR method to robustly train the image-text matching model with noisy correspondence pairs. Inspired by the prior success for noisy labels [20], NCR divides the data into clean and noisy part based on the memorization effect of DNNs, and then rectifies the correspondence with an adaptive model. Qin [29] proposes a uncertainty-based method to achieve efficient learning. Moreover, some works [40, 41] study the partially view-unaligned problem which can be viewed as a generalized noisy correspondence learning. Recently, some works [13, 39] explore a more challenging scenario where noisy labels and noisy correspondence may occur simultaneously. In contrast, our work tackles this challenge in cross-modal retrieval from a meta-learning perspective. + +# 2.3. Meta-Learning + +The objective of meta-learning is to learn at a more advanced level than conventional learning, such as learning the update rule [28], finding easily fine-tuned parameters [9], or adapting to new tasks [19]. Recently, researchers use meta-learning to find model's parameters that robust against noisy labels [21,32,46]. For example, Li [21] optimizes a meta-objective before conventional training to enable the model not overfit to noise. MW-Net [32] use a meta-process to automatically assign weights to the training samples. MLC [46] presents a label correction network trained as a meta manner which generate reliable labels for the noisy training data. + +The most relevant existing method is MLC [46], which first poses the noisy label problem as a meta label correction. However, we argue that our proposed method differs from MLC in two aspects. First, we focuses on the noisy correspondence problem in multimodal data instead of the unimodal classification scenarios. Second, we leverage both positive and negative meta-data to guide model learning the discrimination, which is unexplored in previous meta-learning methods. + +# 3. The Proposed Method + +# 3.1. Problem Formulation + +Without losing generality, we use the image-text retrieval task as a proxy to investigate the noisy correspondence problem in cross-modal retrieval. Given a training set $\mathcal{D}_{train} = \{(I_i,T_i,y_i)\}_{i = 1}^N$ , where $(I_{i},T_{i})$ is the $i$ -th image- + +![](images/7714af4820a40f75b9ed577c4b180ace7b837498f5c705b5a0cff94a4f9377aa.jpg) +Figure 2. Overview of the proposed meta-process. The positive pairs are taken from the original meta-data which contain corresponding image and text, while negative pairs are extracting from different pairs in training set to construct mismatched pairs. Then the MSCN is learned to determine whether the modalities correspond to each other or not. + +text pair, $y_{i} \in \{1,0\}$ indicates the pair is positive (matched), i.e. $(I_i, T_i)$ , or negative (mismatched), i.e. $(I_i, T_{j \neq i})$ , and $N$ is the number of the entire training data. The noisy correspondence in bimodal data is the negative pair but wrongly considered as $y_{i} = 1$ . To address this issue, we propose a meta-learning based method to achieve robust training. The details are delineated next. + +# 3.2. Meta-Learning Based Similarity Correction + +The Training Objective. To begin with, we describe the standard projection module to enable cross-modal features to be comparable. Let $f(I;W_f)$ and $g(T;W_g)$ be the modal-specific networks to map the visual and textual modalities into the joint embedding space, respectively. Then we compute the similarity feature between bimodal features by the function + +$$ +S (f (I), g (T); W _ {s}) = \frac {W _ {s} | f (I) - g (T) | ^ {2}}{\| W _ {s} | f (I) - g (T) | ^ {2} \| _ {2}}, \tag {1} +$$ + +where $W_{s}$ is a learnable parameter matrix to obtain low-dimensional similarity representation. For notation convenience, we collect $\{W_{f},W_{g},W_{s}\}$ into $W$ as the parameters of main net $\mathcal{F} = (f,g,S)$ and denote similarity feature as $\mathcal{F}_W(I,T)$ . With the presence of noisy correspondence, the model will inevitably overfit to noise if we use the similarity representation directly. To tackle this problem, we propose a meta similarity correction network (MSCN), regarding process of obtaining similarity score as a meta-process, which takes the noisy similarity representation as input and produces the corrected similarity score. Specifically, our MSCN is formulated as a MLP network $\mathcal{V}_{\Theta}(\cdot)$ with parameters $\Theta$ . The output $s_i = \mathcal{V}_{\Theta}(\mathcal{F}_W(I_i,T_i))$ is activated with Sigmoid function to represent the similarity score located in [0, 1]. + +For the noisy training pairs, the main net learns noisetolerant features guided by the corrected similarity scores produced from MSCN; while the MSCN predicts the reliable similarity score based on features from main net. This interdependency enables the main net and meta net reinforce each other against noise, and optimized via a bi-level optimization. Formally, the optimal parameters $W^{*}$ is calculated by minimizing the cumulative loss over training data: + +$$ +W ^ {*} (\Theta) = \underset {W} {\arg \min } \mathbb {E} _ {(I _ {i}, T _ {i}) \in \mathcal {D} _ {t r a i n}} l ^ {t r a i n} (I _ {i}, T _ {i}), \tag {2} +$$ + +where $l^{train}(\cdot)$ is the triplet ranking loss: + +$$ +\begin{array}{l} l ^ {t r a i n} \left(I _ {i}, T _ {i}\right) = [ \hat {\gamma} - \mathcal {V} _ {\Theta} \left(\mathcal {F} _ {W} \left(I _ {i}, T _ {i}\right)\right) + \mathcal {V} _ {\Theta} \left(\mathcal {F} _ {W} \left(I _ {i}, T _ {i} ^ {-}\right)\right) ] _ {+} \\ + \left[ \hat {\gamma} - \mathcal {V} _ {\Theta} \left(\mathcal {F} _ {W} \left(I _ {i}, T _ {i}\right)\right) + \mathcal {V} _ {\Theta} \left(\mathcal {F} _ {W} \left(I _ {i} ^ {-}, T _ {i}\right)\right) \right] _ {+,} \tag {3} \\ \end{array} +$$ + +where $[x]_{+} = max(x,0), I_{i}^{-}$ and $T_{i}^{-}$ are the hardest negatives corresponding to given pair $(I_i,T_i)$ similar to VSE++ [8]. To achieve robust cross-modal retrieval, we design a self-adaptive margin $\hat{\gamma}$ instead of a fixed value in VSE++. Ideally, the margin should put less value on the noisy pairs and more on the clean pairs. To this end, we adjust the margin according to the similarity score: + +$$ +\hat {\gamma} = \frac {1}{1 + \left(\frac {s _ {i}}{1 - s _ {i}}\right) ^ {- \tau}} \gamma , \tag {4} +$$ + +where $\gamma$ is the original margin value and $\tau > 0$ is the parameter. This self-adaptive margin will enable the triplet loss put more attention on the samples with high similarity scores that possibly are clean pairs. + +Meta Training Process. To ensure the reliability of the similarity scores produced by MSCN, we utilize a small + +![](images/8aeb08c230bb42145417fb7269135a2ab216757ab23d687d75dfe124ef12a91d.jpg) +Figure 3. Flowchart of the proposed bi-level optimization. At each iteration, we alternately update the main net and the meta net (MSCN) following the steps in Eq. (7) - (9). + +amount clean meta-data set $\mathcal{D}_{meta} = \{(I_i,T_i)\}_{i = 1}^M$ to guide the training, where $M$ is much smaller than $N$ . While meta-learning has been extensively studied in recent literature [10, 32, 46], it remains challenging to construct the meta-process in our objective. On the one hand, the similarity features from main net can not be used directly to produce loss. On the other hand, the meta-process should enable the output of MSCN to represent the similarity score. To achieve this, we view a binary classification task as the meta-process: given an image-text pair, the MSCN is learned to determine whether the modalities correspond to each other or not. Specifically, the positive pairs are taken from the meta-data which contain corresponding image and text, i.e. $(I_{i},T_{i},y_{i} = 1)$ , while negative pairs are extracting from different pairs in training set to construct mismatched pairs, i.e. $(I_{i},T_{j\neq i},y_{i} = 0)$ . The binary labels are produced from the data itself, and can be regarded as the ideal similarity score to guide the training of MSCN. The prediction of MSCN measures the probability of a pair being clean which can be naturally equivalent to the similarity score. Formally, in each iteration of training, we randomly construct $M$ mismatched pairs from training set to extend the meta-data set as $\mathcal{D}_{meta}' = \{(I_i,T_i,y_i)\}_{i = 1}^{2M}$ . Then the optimal parameter $\Theta^{*}$ can be learned by minimizing the following meta-loss calculated on meta-data: + +$$ +\Theta^ {*} = \underset {\Theta} {\arg \min } \mathbb {E} _ {(I _ {i}, T _ {i}) \in \mathcal {D} _ {m e t a} ^ {\prime}} l ^ {m e t a} (I _ {i}, T _ {i}, y _ {i}), \tag {5} +$$ + +where $l^{meta}(\cdot)$ is cross-entropy loss [35, 37]: + +$$ +l ^ {m e t a} (I _ {i}, T _ {i}, y _ {i}) = - y _ {i} \cdot l o g \mathcal {V} _ {\Theta} \left(\mathcal {F} _ {W ^ {*} (\Theta)} (I _ {i}, T _ {i})\right). (6) +$$ + +As illustrated in Fig. 2, despite its simplicity, our proposed meta-process can not only enforce the MSCN to learn reliable similarity score but also has the following advantages: + +1) previous works only guide the meta net with ground truth data; however, we leverage both positive and negative data as meta-knowledge that encourage the MSCN to learn discrimination from them. 2) We overcome one limitation of meta-learning methods that the lack of meta-data. We extract from different pairs among training set to construct massive negative meta-data. Moreover, it is more suitable to the rule of noisy correspondence, which the irregularity of noisy data causes in various patterns. + +Bi-level Optimization. Motivated by recent works [9, 16, 30, 32], we use a bi-level optimization strategy to guarantee the efficiency for updating our main net and MSCN. As illustrated in Fig. 3, it is practically solved via two nested loops of optimization. Specifically, at each iteration $t$ of training, we sample a mini-batch of training pairs $\{(I_i,T_i)\}_{i = 1}^n$ , where $n$ is the size of mini-batch. We first update the main net parameters by taking a gradient descent step toward the direction of triplet loss, which can be formulated as follows: + +$$ +\hat {W} ^ {(t)} (\Theta) = W ^ {(t)} - \alpha \sum_ {i = 1} ^ {n} \left. \nabla_ {W} l ^ {\text {t r a i n}} \left(I _ {i}, T _ {i}\right) \right| _ {W ^ {(t)}}, \tag {7} +$$ + +where $\alpha$ means the learning rate. The updated $\hat{W}^{(t)}$ can be regard as the approximation of the optimal $W^{*}(\Theta)$ at $t$ iteration, containing the gradient information of $\Theta$ . Then we update the parameters of MSCN with main net parameters fixed. Guided by the objective loss in Eq. (5) and a minibatch of meta-data $\{(I_i,T_i,y_i)\}_{i = 1}^m$ , where $m$ is the minibatch size. The updated parameters of MSCN is calculated as: + +$$ +\left. \Theta^ {(t + 1)} = \Theta^ {(t)} - \beta \frac {1}{m} \sum_ {i = 1} ^ {m} \nabla_ {\Theta} l ^ {m e t a} \left(I _ {i}, T _ {i}, y _ {i}\right) \right| _ {\Theta^ {(t)}}, \tag {8} +$$ + +where $\beta$ is the learning rate for meta-process. Finally, we update the parameters of main net while fixing the MSCN parameters. With the corrected similarity scores produced by the newly updated $\Theta^{(t + 1)}$ , we can update the main net parameters $W^{(t + 1)}$ by employing the gradient decent as: + +$$ +W ^ {(t + 1)} = W ^ {(t)} - \alpha \sum_ {i = 1} ^ {n} \left. \nabla_ {W} l ^ {\text {t r a i n}} \left(I _ {i}, T _ {i}\right) \right| _ {W ^ {(t)}}. \tag {9} +$$ + +The above procedure proceeds in an alternating manner to optimize the main net and MSCN until it exceeds the maximal iteration. + +# 3.3. Meta-Knowledge Guided Data Purification + +As mentioned, our MSCN can provide corrected similarity score to avoid noise. However, due to the property of triplet loss, it remains to produce positive loss for the noisy pairs even if we employ the ideal similarity score (e.g., $\mathcal{V}(F(I_i,T_i))\approx 0 < \mathcal{V}\left(F(I_i^-,T_i)\right) + \hat{\gamma})$ , which results in the model fitting to the noise. To tackle this issue, we adopt sample selection strategy to remove the noisy data and only use clean samples as training data. Previous works [2, 42] motivated by the memorization effect of DNNs that clean sample has a lower loss during the beginning of training. Unfortunately, the triplet ranking loss is more complex, i.e., affected by both positive and negative samples, which leads to sub-optimal selection results. Based on the observation shown in Fig. 4, the clean and noisy pairs are prone to distinguish from the similarity scores. Gaussian Mixture Model (GMM) is a widely used technique to model mixture distribution [15, 20]. However, we find the similarity distribution exhibits high skew toward 1 and causes a poor approximation for GMM. Therefore, we fit a two-component Beta Mixture Model (BMM) [27] of per-pair similarity score $s_i$ to better distinguish clean and noisy samples: + +$$ +p \left(s _ {i}\right) = \sum_ {k = 1} ^ {K} \lambda_ {k} \phi \left(s _ {i} \mid \alpha_ {k}, \beta_ {k}\right), \tag {10} +$$ + +where $K = 2$ and $\lambda_{k}$ is the mixture coefficient, and $\phi (s|\alpha_k,\beta_k)$ denotes the probability density function with parameters $\alpha_{k},\beta_{k} > 0$ + +We use the Expectation-Maximization [5] algorithm to optimize the BMM, which poses a crucial challenge: the initialization of model parameters. Existing methods typically adopt the $K$ -Means algorithm [23] for initializing; however, it is time-consuming (i.e., the time complexity is $O(TKN)$ , where $T$ is the number of iterations) and underperforming. We thus propose a novel initialization approach which takes the meta-data as prior knowledge. Specifically, we fit the two Beta components with positive and negative meta-data to model the distribution of clean and noisy pairs, respectively. Given the similarity scores $\mathcal{S}_p = \{s_i\}_{i=1}^M$ + +![](images/79d0c53fb0b4801da8eba6d1025d4c5f036e16be40e0fe4988bd68563c7e1a9d.jpg) +Figure 4. Per-sample similarity score distribution, estimated BMM and BMM with meta prior knowledge under $20\%$ noisy correspondence in Flickr30K after 5 epochs warmup. + +corresponding to positive meta-data, the parameters of the component modeling clean distribution can be initialled by + +$$ +\alpha_ {k} = \frac {\left(1 - E \left(\mathcal {S} _ {p}\right)\right) E \left(\mathcal {S} _ {p}\right) ^ {2}}{V \left(\mathcal {S} _ {p}\right)} - E \left(\mathcal {S} _ {p}\right), \tag {11} +$$ + +$$ +\beta_ {k} = \frac {\alpha_ {k} \left(1 - E \left(\mathcal {S} _ {p}\right)\right)}{E (\mathcal {S} _ {p})}, +$$ + +where $E(\cdot)$ and $V(\cdot)$ are the mean and variance function, respectively. The initialization of component modeling noisy distribution is analogous but uses similarity scores $S_{n} = \{s_{i}\}_{i=1}^{M}$ of negative meta-data. As illustrated in Fig. 4, our meta-knowledge guided initialization method achieves better approximation while with lower time complexity (i.e., $O(KM)$ ). + +Finally, we compute the posterior probability $p_i = p(k|s_i) = p(k)p(s_i|\alpha_k,\beta_k) / p(s_i)$ as the probability of $i$ -th simple being clean. The purified training set is defined as: + +$$ +\mathcal {D} _ {\text {t r a i n}} ^ {\prime} = \left\{\left(I _ {i}, T _ {i}\right) \in \mathcal {D} _ {\text {t r a i n}} \mid p _ {i} > 0. 5 \right\}. \tag {12} +$$ + +At each iteration, we produce the purified dataset before the training of our meta-learning objective. Moreover, following [12, 15, 20], we maintain two networks $\{\mathcal{F}_W^1,\mathcal{V}_{\Theta}^1\}$ and $\{\mathcal{F}_W^2,\mathcal{V}_{\Theta}^2\}$ to avoid error accumulation that one network will produce the purified training set to train the other one. The full algorithm is shown in supplementary material. + +# 4. Experiment + +# 4.1. Datasets and Evaluation Protocol + +Datasets. We evaluate our method on three standard image-text retrieval datasets. To be specific, the Flickr30K [43] collects 31,783 images with 5 corresponding captions each from Flickr website. Following the split in [15], we use 1,000 image-text pairs for validation, 1,000 image-text + +pairs for testing, and the rest for training. The MS-COCO [24] contains 123,287 images with 5 corresponding captions each. We split the MS-COCO following [15] that 113,287 image-text pairs for training, 5,000 pairs for validation, and 5,000 pairs for testing. Conceptual Captions [31] is a large-scale dataset contains 3.3M images with one corresponding caption each. We employ the same subset following [15], i.e., CC152K. Specifically, CC152K consists of 150,000 pairs from training set for training, 1,000 pairs from validation set for validation, and 1,000 pairs from validation set for testing. Moreover, we use an extra meta-data set from validation containing 3,000 pairs. + +Evaluation Protocol. We evaluate the retrieval performance with the recall rate at K (R@K) metric. In a nutshell, R@K is the proportion for the relevant items retrieved in the closest K items to the query. In our experiments, we take image and text as queries, respectively, and report R@1, R@5, and R@10 results for a comprehensive evaluation. Following [15], we average the similarity scores from two networks at the inference phase. + +# 4.2. Implementation Details + +As a general approach, our method can be easily applied to almost all cross-modal retrieval methods to improve robustness. For fair comparison, we chose the same network backbones with NCR [15], i.e., the projection modules and similarity function. For all datasets, the training processes contain 50 epochs after a 5 epochs warmup. We employ ADAM [17] as the optimizer for both main net and meta net with a batch size of 64. To fit the BMM, we set the stop threshold as $10^{-2}$ and maximum number of iterations as 10 for the EM procedure. Note that, we clamp the similarity scores into a range of $[10^{-4}, 1 - 10^{-4}]$ instead of [0,1] for stable fitting. Moreover, we set the margin parameter as 0.2 and $\tau = 2$ to calculate the self-adaptive margin. For all datasets, the number of ground-truth meta-data is approximately $2\%$ of training data, and the learning rate of main net and meta net are initialized with $2 \times 10^{-4}$ and $1.7 \times 10^{-5}$ , respectively. For CC152K containing real noise, we employ the validation data as the meta set due to the lack of ground-truth data. For all experiments, we decay the learning rate by 0.1 after 30 epochs. + +# 4.3. Comparison with the State-of-the-Art + +In this section, we conduct comparison experiments with 7 state-of-the-art methods that include SCAN [18], VSRN [22], IMRAM [4], SGR, SAF, SGRAF [7], and NCR [15]. As Flickr30K and MS-COCO are well annotated datasets, we generate synthetic noisy correspondence by randomly shuffling the training images and captions for a specific noise ratio, and we conduct experiments with three different + +
Image to Text
MethodsR@1R@5R@10SUM
SCAN (ECCV'18)30.555.365.3151.1
VSRN(ICCV'19)32.661.370.5164.4
IMRAM(CVPR'20)33.157.668.1158.8
SAF (AAAI'21)31.759.368.2159.2
SGR (AAAI'21)11.329.739.680.6
SGR* (AAAI'21)35.063.473.3171.7
NCR(NIPS'21)39.564.573.5177.5
MSCN*(Ours)39.765.475.3180.4
MSCN(Ours)40.165.776.6182.4
Text to Image
SCAN (ECCV'18)26.953.064.7144.6
VSRN(ICCV'19)32.559.470.4162.3
IMRAM(CVPR'20)29.056.867.4153.2
SAF (AAAI'21)31.959.067.9158.8
SGR (AAAI'21)13.130.141.684.8
SGR* (AAAI'21)34.963.072.8170.7
NCR(NIPS'21)40.364.673.2178.1
MSCN*(Ours)39.866.175.0180.9
MSCN(Ours)40.667.476.3184.3
+ +Table 1. Image-Text Retrieval Performance on CC152K with real noise, and the best results are highlighted in bold. + +level of noise ratios, i.e., $20\%$ , $50\%$ , and $70\%$ . For all methods, we choose the best checkpoint on the validation set and report its performance on the test set. Following NCR, we also report two strong baselines based on SGR, i.e., SGR-C and SGR*. In short, SGR-C uses only clean data for training, and SGR* employs a pre-training process while training without hard negatives to improve robustness. To evaluate our method on real-world noisy data, we conduct experiments on the CC152K, which is harvested from the Internet and contains about $3\% \sim 20\%$ noisy correspondence [31]. + +Experiments on CC152K. Tab. 1 shows the results on CC152K with real noisy correspondence. On CC152K, our MSCN achieves the state-of-the-art performance in terms of all metrics. Specifically, the sum score of MSCN is $4.9\%$ and $6.2\%$ higher than the best baseline in text and image retrieval, respectively. In addition, we also present a more strict comparison (denoted by $\mathrm{MSCN^{*}}$ ) that trained our method with only 1,000 meta-data (approximately $0.67\%$ of all data). We can see that our $\mathrm{MSCN^{*}}$ also achieves a competitive performance. Specifically, all metrics except R@1 of image retrieval are superior to the baselines. + +Experiments on Flickr30K and MS-COCO. Tab. 2 shows the results on Flickr30K and MS-COCO under a range of synthetic noise rates. Following [15], we report the + +
Flickr30KMS-COCO
Image to TextText to ImageImage to TextText to Image
NoiseMethodsR@1R@5R@10R@1R@5R@10R@1R@5R@10R@1R@5R@10
20%SCAN59.183.490.436.667.077.566.291.096.445.080.289.3
VSRN58.182.689.340.768.778.225.159.074.817.649.064.1
IMRAM63.086.091.341.471.280.568.692.897.655.785.091.0
SAF51.079.388.038.366.576.267.392.596.653.484.592.4
SGR*62.886.292.244.472.380.467.891.796.252.983.590.1
SGR-C72.890.895.456.482.188.675.495.297.960.188.594.8
NCR75.093.997.558.383.089.077.795.598.262.589.395.3
MSCN77.494.997.659.683.289.278.197.298.864.390.495.8
50%SCAN27.757.668.816.239.349.840.873.584.95.415.121.0
VSRN14.337.650.012.130.039.423.554.769.316.047.865.9
IMRAM9.126.638.22.78.412.721.360.275.922.352.864.3
SAF30.363.675.427.953.765.130.467.882.333.569.082.8
SGR*36.968.180.229.356.267.060.687.493.646.074.279.0
SGR-C69.890.394.850.177.585.271.794.197.757.086.693.7
NCR72.993.096.354.379.886.574.694.697.859.187.894.5
MSCN74.493.296.055.380.486.877.596.298.760.789.194.9
70%SCAN5.619.327.42.28.012.818.143.157.40.31.32.3
VSRN0.82.54.10.51.52.75.115.724.62.58.813.3
IMRAM1.33.13.90.31.22.87.120.033.45.315.222.0
SAF0.52.23.00.20.81.70.11.74.00.61.93.0
SGR*17.942.151.914.631.040.835.771.285.431.665.879.0
SGR-C65.089.394.748.174.581.169.893.697.556.586.093.4
NCR16.138.552.811.029.541.435.469.583.431.566.481.1
MSCN69.089.393.049.273.179.074.494.997.758.887.293.7
+ +results by averaging over 5 folds of 1K test images. From Tab. 2, we can draw the observation that noisy correspondence remarkably affects the performance of cross-modal retrieval methods. The retrieval accuracy will decrease fast with the noise rate increasing. Our MSCN is superior to the baselines in almost all metrics under different noise rates. In the low noise cases, i.e., $20\%$ and $50\%$ , our MSCN improves R@1 (sum of image and text) by $3.7\%$ , $2.2\%$ , $2.5\%$ , and $4.5\%$ in the four valuations compared with NCR. In the high noise case, i.e., $70\%$ , all baselines can easily overfit to noisy correspondence and results in poor performance. Our MSCN improves R@1 by $52.9\%$ , $38.2\%$ , $39.4\%$ , and $27.3\%$ in these four valuations compared with NCR. Even compared with the SGR-C trained only on clean data, our MSCN also achieves competitive performance. + +# 4.4. Ablation Study + +To evaluate the performance of the proposed components (i.e., self-adaptive margin and data purification), we conduct the ablation study on the Flickr30K with $20\%$ noisy correspondence. Note that for MSCN without $\mathcal{D}_{train}^{\prime}$ , we only use a single model to perform the experiment. As shown in Tab. 3, we can see that the performance without $\hat{\gamma}$ or $\mathcal{D}_{train}^{\prime}$ are worse than the complete MSCN, which + +indicates that all components are important to achieve advantageous results. We explain more on the data selection strategy in the supplementary material. + +Table 2. Image-Text Retrieval Performance under synthetic noise rates of ${20}\% ,{50}\%$ and ${70}\%$ on Flickr30K and MS-COCO 1K,and the best results are highlighted in bold. + +
MethodImage to Text
MSCNw/o γw/o D'trainR@1R@5R@10
✓77.494.997.6
✓✓75.394.597.2
✓✓75.893.896.2
✓✓✓74.191.594.7
Text to Image
✓59.683.289.2
✓✓58.383.188.9
✓✓55.874.577.4
✓✓✓53.471.272.2
+ +Table 3. Ablation studies on Flickr30K with ${20}\%$ noise rate. + +# 4.5. Progressive Comparison + +Fig. 6 plots the average of recalls on testing set of MSCN and NCR as training proceeds. We show a representative result using Flickr30K with $50\%$ noisy correspondence. From the result, we can see that our MSCN achieves a more stable performance in overall training process. Although NCR + +![](images/9e2434691c1db0c972440d39d2b10ec8317e5cb19ba4c4225af7f4650834da11.jpg) +the circular movement of the camera. (0.3444) + +![](images/5416ed9692c3717fe83a5e49a16f9bd801e3f58286f39183058aff68ffcab230.jpg) +person arrives earlier this month. (0.3485) + +![](images/dbabc25dd5f7d976e226d17998b27e8e347ed66a3a20f0514f56a9dfb3b00fa8.jpg) +i expected to learn more about plants. (0.3487) + +![](images/9cbcac1eea18f074e487aea5870ba5834c873e679fedf051122762eab4b98da0.jpg) +man,chairman to speak at show (0.3488) + +![](images/3c991ea05fe1855eff140559fa47d4adc725baab4d8030b840e375b8c6ab8925.jpg) +note the many parked cars (0.3511) + +![](images/2c3e2842b25d71a4a0e251b0ff41a8f591dea74df2061f4a2faa062509a58c69.jpg) +a man using his mobile phone. (0.3518) + +does not need an extra meta-set to guide the training, it relies on a clean validation set to choose the best model parameters. The stableness also indicates that our method has better alleviated the interference of noisy correspondence. + +![](images/8b2cfa94e9146426a496cfd7be0646a851b640647754760cce2af5d3db163311.jpg) +Figure 6. Progressive performance comparison of MSCN and NCR on testing set as training proceeds. + +# 4.6. Detected Noisy Samples + +Fig. 5 shows some real-world noisy image-text pairs detected (with low similarity score) by MSCN in Conceptual Captions. We also present the similarity score predicted by our MSCN. For most noisy pairs, the image and text contains completely irrelevant semantic information. Moreover, our MSCN can even find the subtle mistake in the noisy image-text pair. In the last pair, the picture shows a man covering his ears but wrongly caption as using a mobile phone. Although this picture is misleading, our method successfully detects the subtle difference. + +# 4.7. Visualization on Similarity Score + +Fig. 7 plots the similarity score distribution predicted by MSCN for clean and noisy training samples. It can be seen that almost all large similarity scores belong to clean pairs, and the noisy pairs' scores are smaller than clean samples, which implies that the trained MSCN can successfully provide corrected similarity scores. + +# 5. Conclusion + +In this paper, we explore a meta-learning method to address the problem of learning with noisy correspond- + +![](images/92e2f917921919a070aab04042430702cb4149ba13147bbe56f7cc6ca78364f3.jpg) +Figure 5. Real-world noisy examples detected by our MSCN. +(a) +Figure 7. Per-sample similarity distribution on training data of Flickr30K. (a) $20\%$ noisy correspondence. (b) $50\%$ noisy correspondence. + +![](images/ac5ce36a3925aec0bb9dbc6372691a5b0277f3b6287eddd6b638c5c8d8b6b5ee.jpg) +(b) + +dence. Specifically, we propose a meta correction network (MSCN) to provide reliable similarity scores. Our MSCN is trained by a novel meta-process that views both positive and negative data as meta-knowledge to encourage MSCN to learn discrimination from them. To further mitigate noise interference, we design a data purification strategy that uses meta-data as prior knowledge to purify the noisy data efficiently. We conduct comprehensive experiments on three widely used datasets. The results validate the effectiveness of our method in both synthetic and real noises. + +# Acknowledgements + +This work is supported by the National Key Research and Development Program of China (2020AAA0108800), National Nature Science Foundation of China (62192781, 62272374, 62202367, 62250009, 62137002, 61937001), Innovative Research Group of the National Natural Science Foundation of China (61721002), Innovation Research Team of Ministry of Education (IRT_17R86), Project of China Knowledge Center for Engineering Science and Technology, and Project of Chinese academy of engineering "The Online and Offline Mixed Educational Service System for 'The Belt and Road' Training in MOOC China". We would like to express our gratitude for the support of K. C. Wong Education Foundation. + +# References + +[1] Relja Arandjelovic and Andrew Zisserman. Objects that sound. In Proceedings of the European conference on computer vision (ECCV), pages 435-451, 2018. 1 +[2] Eric Arazo, Diego Ortego, Paul Albert, Noel O'Connor, and Kevin McGuinness. Unsupervised label noise modeling and loss correction. In International conference on machine learning, pages 312-321. PMLR, 2019. 5 +[3] Devansh Arpit, Stanislaw K Jastrzebski, Nicolas Ballas, David Krueger, Emmanuel Bengio, Maxinder S Kanwal, Tegan Maharaj, Asja Fischer, Aaron C Courville, Yoshua Bengio, et al. A closer look at memorization in deep networks. In ICML, 2017. 1 +[4] Hui Chen, Guiguang Ding, Xudong Liu, Zijia Lin, Ji Liu, and Jungong Han. Imram: Iterative matching with recurrent attention memory for cross-modal image-text retrieval. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12655-12663, 2020. 6 +[5] Arthur P Dempster, Nan M Laird, and Donald B Rubin. Maximum likelihood from incomplete data via the em algorithm. Journal of the Royal Statistical Society: Series B (Methodological), 39(1):1-22, 1977. 5 +[6] Cheng Deng, Zhaojia Chen, Xianglong Liu, Xinbo Gao, and Dacheng Tao. Triplet-based deep hashing network for cross-modal retrieval. IEEE Transactions on Image Processing, 27(8):3893-3903, 2018. 1 +[7] Haiwen Diao, Ying Zhang, Lin Ma, and Hutchuan Lu. Similarity reasoning and filtration for image-text matching. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 1218-1226, 2021. 2, 6 +[8] Fartash Faghri, David J Fleet, Jamie Ryan Kiros, and Sanja Fidler. Vse++: Improving visual-semantic embeddings with hard negatives. arXiv preprint arXiv:1707.05612, 2017. 2, 3 +[9] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In International conference on machine learning, pages 1126–1135. PMLR, 2017. 2, 4 +[10] Luca Franceschi, Paolo Frasconi, Saverio Salzo, Riccardo Grazzi, and Massimiliano Pontil. Bilevel programming for hyperparameter optimization and meta-learning. In International Conference on Machine Learning, pages 1568-1577. PMLR, 2018. 4 +[11] Dalu Guo, Chang Xu, and Dacheng Tao. Image-question-answer synergistic network for visual dialog. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10434-10443, 2019. 1 +[12] Bo Han, Quanming Yao, Xingrui Yu, Gang Niu, Miao Xu, Weihua Hu, Ivor Tsang, and Masashi Sugiyama. Co-teaching: Robust training of deep neural networks with extremely noisy labels. Advances in neural information processing systems, 31, 2018. 5 +[13] Haochen Han, Qinghua Zheng, Minnan Luo, Kaiyao Miao, Feng Tian, and Yan Chen. Noise-tolerant learning for audiovisual action recognition. arXiv preprint arXiv:2205.07611, 2022. 2 +[14] Peng Hu, Xi Peng, Hongyuan Zhu, Liangli Zhen, and Jie Lin. Learning cross-modal retrieval with noisy labels. In + +Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5403-5413, 2021. 2 +[15] Zhenyu Huang, Guocheng Niu, Xiao Liu, Wenbiao Ding, Xinyan Xiao, Hua Wu, and Xi Peng. Learning with noisy correspondence for cross-modal matching. Advances in Neural Information Processing Systems, 34:29406-29419, 2021. 1, 2, 5, 6 +[16] Haoxuanye Ji, Le Wang, Sanping Zhou, Wei Tang, Nanning Zheng, and Gang Hua. Meta pairwise relationship distillation for unsupervised person re-identification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3661–3670, 2021. 4 +[17] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR (Poster), 2015. 6 +[18] Kuang-Huei Lee, Xi Chen, Gang Hua, Houdong Hu, and Xiaodong He. Stacked cross attention for image-text matching. In Proceedings of the European conference on computer vision (ECCV), pages 201-216, 2018. 2, 6 +[19] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy Hospedales. Learning to generalize: Meta-learning for domain generalization. In Proceedings of the AAAI conference on artificial intelligence, volume 32, 2018. 2 +[20] Junnan Li, Richard Socher, and Steven CH Hoi. Dividemix: Learning with noisy labels as semi-supervised learning. In International Conference on Learning Representations, 2019. 1, 2, 5 +[21] Junnan Li, Yongkang Wong, Qi Zhao, and Mohan S Kankanhalli. Learning to learn from noisy labeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5051-5059, 2019. 2 +[22] Kunpeng Li, Yulun Zhang, Kai Li, Yuanyuan Li, and Yun Fu. Visual semantic reasoning for image-text matching. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4654-4662, 2019. 2, 6 +[23] Aristidis Likas, Nikos Vlassis, and Jakob J Verbeek. The global k-means clustering algorithm. Pattern recognition, 36(2):451-461, 2003. 5 +[24] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 6 +[25] Chunxiao Liu, Zhendong Mao, Tianzhu Zhang, Hongtao Xie, Bin Wang, and Yongdong Zhang. Graph structured network for image-text matching. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10921-10930, 2020. 2 +[26] Lin Ma, Zhengdong Lu, and Hang Li. Learning to answer questions from image using convolutional neural network. In Thirtieth AAAI Conference on Artificial Intelligence, 2016. 1 +[27] Zhanyu Ma and Arne Leijon. Bayesian estimation of beta mixture models with variational inference. IEEE Transactions on Pattern Analysis and Machine Intelligence, 33(11):2160-2173, 2011. 5 +[28] Nikhil Mishra, Mostafa Rohaninejad, Xi Chen, and Pieter Abbeel. A simple neural attentive meta-learner. In International Conference on Learning Representations, 2018. 2 + +[29] Yang Qin, Dezhong Peng, Xi Peng, Xu Wang, and Peng Hu. Deep evidential learning with noisy correspondence for cross-modal retrieval. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4948-4956, 2022. 2 +[30] Mengye Ren, Wenyuan Zeng, Bin Yang, and Raquel Urtasun. Learning to reweight examples for robust deep learning. In International conference on machine learning, pages 4334-4343. PMLR, 2018. 4 +[31] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2556-2565, 2018. 1, 6 +[32] Jun Shu, Qi Xie, Lixuan Yi, Qian Zhao, Sanping Zhou, Zongben Xu, and Deyu Meng. Meta-weight-net: Learning an explicit mapping for sample weighting. Advances in neural information processing systems, 32, 2019. 2, 4 +[33] Didac Surís, Amanda Duarte, Amaia Salvador, Jordi Torres, and Xavier Giró-i Nieto. Cross-modal embeddings for video and audio retrieval. In Proceedings of the European Conference on Computer Vision (ECCV) Workshops, pages 0–0, 2018. 1 +[34] Jun Tang, Ke Wang, and Ling Shao. Supervised matrix factorization hashing for cross-modal retrieval. IEEE Transactions on Image Processing, 25(7):3157-3166, 2016. 2 +[35] Bokun Wang, Yang Yang, Xing Xu, Alan Hanjalic, and Heng Tao Shen. Adversarial cross-modal retrieval. In Proceedings of the 25th ACM international conference on Multimedia, pages 154-162, 2017. 4 +[36] Liwei Wang, Yin Li, and Svetlana Lazebnik. Learning deep structure-preserving image-text embeddings. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5005-5013, 2016. 2 +[37] Yunchao Wei, Yao Zhao, Canyi Lu, Shikui Wei, Luoqi Liu, Zhenfeng Zhu, and Shuicheng Yan. Cross-modal retrieval with cnn visual features: A new baseline. IEEE transactions on cybernetics, 47(2):449-460, 2016. 4 +[38] Xing Xu, Li He, Huimin Lu, Lianli Gao, and Yanli Ji. Deep adversarial metric learning for cross-modal retrieval. *World Wide Web*, 22(2):657-672, 2019. 2 +[39] Mouxing Yang, Zhenyu Huang, Peng Hu, Taihao Li, Jiancheng Lv, and Xi Peng. Learning with twin noisy labels for visible-infrared person re-identification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14308-14317, 2022. 2 +[40] Mouxing Yang, Yunfan Li, Peng Hu, Jinfeng Bai, Jiancheng Lv, and Xi Peng. Robust multi-view clustering with incomplete information. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(1):1055-1069, 2022. 2 +[41] Mouxing Yang, Yunfan Li, Zhenyu Huang, Zitao Liu, Peng Hu, and Xi Peng. Partially view-aligned representation learning with noise-robust contrastive loss. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1134–1143, 2021. 2 +[42] Quanming Yao, Hansi Yang, Bo Han, Gang Niu, and James Tin-Yau Kwok. Searching to exploit memorization effect in + +learning with noisy labels. In International Conference on Machine Learning, pages 10789-10798. PMLR, 2020. 5 +[43] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics, 2:67-78, 2014. 5 +[44] Qi Zhang, Zhen Lei, Zhaoxiang Zhang, and Stan Z Li. Context-aware attention network for image-text retrieval. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3536-3545, 2020. 1 +[45] Liangli Zhen, Peng Hu, Xu Wang, and Dezhong Peng. Deep supervised cross-modal retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10394-10403, 2019. 2 +[46] Guoqing Zheng, Ahmed Hassan Awadallah, and Susan Dumais. Meta label correction for noisy label learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 11053-11061, 2021. 2, 4 \ No newline at end of file diff --git a/2023/Noisy Correspondence Learning With Meta Similarity Correction/images.zip b/2023/Noisy Correspondence Learning With Meta Similarity Correction/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1af9a7c92b8417fb347d0a8a574c6f75cf4412a4 --- /dev/null +++ b/2023/Noisy Correspondence Learning With Meta Similarity Correction/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4e14db9ede0d2f3d92d0ae5dcfc0d5bd55ca1c324ac2ca3fc7faab5ba237a358 +size 663987 diff --git a/2023/Noisy Correspondence Learning With Meta Similarity Correction/layout.json b/2023/Noisy Correspondence Learning With Meta Similarity Correction/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5250606831fae3e0e6d19358c2d706596def8427 --- /dev/null +++ b/2023/Noisy Correspondence Learning With Meta Similarity Correction/layout.json @@ -0,0 +1,8923 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 95, + 103, + 499, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 103, + 499, + 121 + ], + "spans": [ + { + "bbox": [ + 95, + 103, + 499, + 121 + ], + "type": "text", + "content": "Noisy Correspondence Learning with Meta Similarity Correction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 145, + 143, + 452, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 143, + 452, + 157 + ], + "spans": [ + { + "bbox": [ + 145, + 143, + 452, + 157 + ], + "type": "text", + "content": "Haochen Han" + }, + { + "bbox": [ + 145, + 143, + 452, + 157 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 145, + 143, + 452, + 157 + ], + "type": "text", + "content": ", Kaiyao Miao" + }, + { + "bbox": [ + 145, + 143, + 452, + 157 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 145, + 143, + 452, + 157 + ], + "type": "text", + "content": ", Qinghua Zheng" + }, + { + "bbox": [ + 145, + 143, + 452, + 157 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 145, + 143, + 452, + 157 + ], + "type": "text", + "content": ", Minnan Luo" + }, + { + "bbox": [ + 145, + 143, + 452, + 157 + ], + "type": "inline_equation", + "content": "^{1*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 158, + 459, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 158, + 459, + 171 + ], + "spans": [ + { + "bbox": [ + 133, + 158, + 459, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 133, + 158, + 459, + 171 + ], + "type": "text", + "content": " School of Computer Science and Technology, Xi'an Jiaotong University, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 139, + 172, + 453, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 172, + 453, + 185 + ], + "spans": [ + { + "bbox": [ + 139, + 172, + 453, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 139, + 172, + 453, + 185 + ], + "type": "text", + "content": " School of Cyber Science and Engineering, Xi'an Jiaotong University, China" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 114, + 187, + 479, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 187, + 479, + 199 + ], + "spans": [ + { + "bbox": [ + 114, + 187, + 479, + 199 + ], + "type": "text", + "content": "{hhc1997, miaoky814}@stu.xjtu.edu.cn, {qhzheng, minnluo}@xjtu.edu.cn" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "spans": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 252, + 290, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 252, + 290, + 504 + ], + "spans": [ + { + "bbox": [ + 46, + 252, + 290, + 504 + ], + "type": "text", + "content": "Despite the success of multimodal learning in cross-modal retrieval task, the remarkable progress relies on the correct correspondence among multimedia data. However, collecting such ideal data is expensive and time-consuming. In practice, most widely used datasets are harvested from the Internet and inevitably contain mismatched pairs. Training on such noisy correspondence datasets causes performance degradation because the cross-modal retrieval methods can wrongly enforce the mismatched data to be similar. To tackle this problem, we propose a Meta Similarity Correction Network (MSCN) to provide reliable similarity scores. We view a binary classification task as the meta-process that encourages the MSCN to learn discrimination from positive and negative meta-data. To further alleviate the influence of noise, we design an effective data purification strategy using meta-data as prior knowledge to remove the noisy samples. Extensive experiments are conducted to demonstrate the strengths of our method in both synthetic and real-world noises, including Flickr30K, MS-COCO, and Conceptual Captions. Our code is publicly available." + }, + { + "bbox": [ + 46, + 252, + 290, + 504 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 529, + 128, + 542 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 529, + 128, + 542 + ], + "spans": [ + { + "bbox": [ + 47, + 529, + 128, + 542 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 550, + 287, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 550, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 46, + 550, + 287, + 658 + ], + "type": "text", + "content": "Recently, cross-modal retrieval has drawn much attention with the rapid growth of multimedia data. Given a query sample of specific modality, cross-modal retrieval aims to retrieve relevant samples across different modalities. Existing cross-modal retrieval works [1, 6, 11] usually learn a comparable common space to bridge different modalities, which achieved remarkable progress in many applications, including video-audio retrieval [1, 33], visual question answering [11, 26], and image-text matching [44]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 658, + 287, + 683 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 287, + 683 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 287, + 683 + ], + "type": "text", + "content": "Despite the promise, a core assumption in cross-modal retrieval is the correct correspondence among multiple" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 307, + 228, + 545, + 319 + ], + "blocks": [ + { + "bbox": [ + 307, + 228, + 545, + 319 + ], + "lines": [ + { + "bbox": [ + 307, + 228, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 228, + 545, + 319 + ], + "type": "image", + "image_path": "6f5725cb576e78f8305df691f3e3a92564c58cd94e044e587fff66d835979a4a.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 329, + 547, + 385 + ], + "lines": [ + { + "bbox": [ + 304, + 329, + 547, + 385 + ], + "spans": [ + { + "bbox": [ + 304, + 329, + 547, + 385 + ], + "type": "text", + "content": "Figure 1. Illustration of noisy correspondence in image-text retrieval. A standard triplet loss is used to enforce the positive pairs to be closer than negatives in the common space. Noisy correspondence is the mismatched pairs but wrongly considered as positive ones, and thus results in model performance degradation." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 401, + 546, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 401, + 546, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 401, + 546, + 581 + ], + "type": "text", + "content": "modalities. However, collecting such ideal data is expensive and time-consuming. In practice, most widely used datasets are harvested from the Internet and inevitably contain noisy correspondence [31]. As illustrated in Fig. 1, the cross-modal retrieval method will wrongly enforce the mismatched data to be similar when learning with noisy correspondence, which may significantly affect the retrieval performance. To date, few effort has been made to address this. Huang [15] first researches this issue and proposes the NCR method to train from the noisy image-text pairs robustly. Inspired by the prior success for noisy labels [20], NCR divides the data into clean and noisy partitions and rectifies the correspondence with an adaptive model. However, NCR based on the memorization effect of DNNs [3], which leads to poor performance under high noise ratio." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "text", + "content": "To tackle the challenge, we propose a Meta Similarity Correction Network (MSCN) which aims to provide reliable similarity scores for the noisy features from main net. We view a binary classification task as the meta-process: given a multimodal sample, the MSCN will learn to determine whether the modalities correspond to each other or not, where the prediction of MSCN can be naturally regarded as the similarity score. Specifically, a small amount of clean pairs is used to construct positive and negative meta-data, both viewed as meta-knowledge that encourages MSCN to learn the discrimination. Meanwhile, the main net" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 693, + 212, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 693, + 212, + 703 + ], + "spans": [ + { + "bbox": [ + 58, + 693, + 212, + 703 + ], + "type": "text", + "content": "*Minnan Luo is the corresponding author." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 59, + 703, + 212, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 703, + 212, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 703, + 212, + 712 + ], + "type": "text", + "content": "1https://github.com/hhc1997/MSCN" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7517" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 203 + ], + "type": "text", + "content": "trained by the corrected similarity score from MSCN with a self-adaptive margin to achieve robust learning. This interdependency makes the main net and MSCN benefits each other against noise. However, due to the property of triplet loss, it remains to produce positive loss for noisy pairs even if we employ the ideal similarity scores. To this end, we further propose a meta-knowledge guided data purification strategy to remove samples with potentially wrong correspondence. Extensive experiments are conducted to demonstrate the strengths of our method in both synthetic and real-world noises." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 205, + 288, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 205, + 288, + 227 + ], + "spans": [ + { + "bbox": [ + 47, + 205, + 288, + 227 + ], + "type": "text", + "content": "The main contributions of this work are summarized as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 59, + 229, + 286, + 361 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 59, + 229, + 286, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 229, + 286, + 277 + ], + "spans": [ + { + "bbox": [ + 59, + 229, + 286, + 277 + ], + "type": "text", + "content": "- We pioneer the exploration of meta-learning for noisy correspondence problem, where a meta correction network is proposed to provide reliable similarity scores against noise." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 277, + 286, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 277, + 286, + 323 + ], + "spans": [ + { + "bbox": [ + 59, + 277, + 286, + 323 + ], + "type": "text", + "content": "- We present a novel meta-process that first considers both positive and negative data as meta-knowledge to encourage the MSCN to learn discrimination from them." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 325, + 286, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 325, + 286, + 361 + ], + "spans": [ + { + "bbox": [ + 59, + 325, + 286, + 361 + ], + "type": "text", + "content": "- We design an effective data purification strategy using meta-data as prior knowledge to remove the noisy samples." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 373, + 134, + 386 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 373, + 134, + 386 + ], + "spans": [ + { + "bbox": [ + 47, + 373, + 134, + 386 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 394, + 176, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 394, + 176, + 406 + ], + "spans": [ + { + "bbox": [ + 47, + 394, + 176, + 406 + ], + "type": "text", + "content": "2.1. Cross-Modal Retrieval" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 414, + 287, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 414, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 414, + 287, + 544 + ], + "type": "text", + "content": "Cross-modal retrieval works aim to take one modality of data as query to retrieve the related data in other modalities. Based on the utilization of annotation information, cross-modal retrieval methods can be roughly divided into two categories: 1) Unsupervised Methods [8,22,25,36]. It only uses the co-occurrence information to learn the common representations among the multimodal data. 2) Supervised methods [14,34,38,45]. The extra label information is employed to boost the discrimination of common representations. However, the lack of annotations limited the practicability in the real world." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "content": "Our proposed approach falls in the category of unsupervised methods. For example, Wang [36] uses a two-branch neural networks to learn the joint embeddings of multimodal data. Inspired by hard negative mining, " + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{VSE}++" + }, + { + "bbox": [ + 46, + 545, + 288, + 713 + ], + "type": "text", + "content": " [8] uses hard negatives to improve the retrieval performance. To capture fine-grained interplay between modalities, SCAN [18] introduce stacked cross attention to enable attention with context from both image and sentence. Recently, motivated by the powerful learning ability of graph model, GSMN [25] and SGRAF [7] construct graph structure for multimodal data to benefit the learning of fine-grained correspondence. Although existing works have achieved remarkable results, they usually depend on the correct correspondence among multimodal data and cannot" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "type": "text", + "content": "tackle the noise issue. Thus, it is significant to explore how to learn cross-modal retrieval with noisy correspondence, but which is rarely touched in previous studies." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 116, + 480, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 116, + 480, + 129 + ], + "spans": [ + { + "bbox": [ + 305, + 116, + 480, + 129 + ], + "type": "text", + "content": "2.2. Noisy Correspondence Learning" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 134, + 545, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 134, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 304, + 134, + 545, + 338 + ], + "type": "text", + "content": "As a newly proposed problem, noisy correspondence is the mismatched pairs but wrongly considered as positive ones. Huang [15] first research this issue and propose the NCR method to robustly train the image-text matching model with noisy correspondence pairs. Inspired by the prior success for noisy labels [20], NCR divides the data into clean and noisy part based on the memorization effect of DNNs, and then rectifies the correspondence with an adaptive model. Qin [29] proposes a uncertainty-based method to achieve efficient learning. Moreover, some works [40, 41] study the partially view-unaligned problem which can be viewed as a generalized noisy correspondence learning. Recently, some works [13, 39] explore a more challenging scenario where noisy labels and noisy correspondence may occur simultaneously. In contrast, our work tackles this challenge in cross-modal retrieval from a meta-learning perspective." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 346, + 400, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 346, + 400, + 358 + ], + "spans": [ + { + "bbox": [ + 306, + 346, + 400, + 358 + ], + "type": "text", + "content": "2.3. Meta-Learning" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 365, + 545, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 365, + 545, + 508 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 545, + 508 + ], + "type": "text", + "content": "The objective of meta-learning is to learn at a more advanced level than conventional learning, such as learning the update rule [28], finding easily fine-tuned parameters [9], or adapting to new tasks [19]. Recently, researchers use meta-learning to find model's parameters that robust against noisy labels [21,32,46]. For example, Li [21] optimizes a meta-objective before conventional training to enable the model not overfit to noise. MW-Net [32] use a meta-process to automatically assign weights to the training samples. MLC [46] presents a label correction network trained as a meta manner which generate reliable labels for the noisy training data." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 509, + 545, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 509, + 545, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 509, + 545, + 616 + ], + "type": "text", + "content": "The most relevant existing method is MLC [46], which first poses the noisy label problem as a meta label correction. However, we argue that our proposed method differs from MLC in two aspects. First, we focuses on the noisy correspondence problem in multimodal data instead of the unimodal classification scenarios. Second, we leverage both positive and negative meta-data to guide model learning the discrimination, which is unexplored in previous meta-learning methods." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 627, + 435, + 640 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 627, + 435, + 640 + ], + "spans": [ + { + "bbox": [ + 306, + 627, + 435, + 640 + ], + "type": "text", + "content": "3. The Proposed Method" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 647, + 430, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 647, + 430, + 658 + ], + "spans": [ + { + "bbox": [ + 306, + 647, + 430, + 658 + ], + "type": "text", + "content": "3.1. Problem Formulation" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 665, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 715 + ], + "type": "text", + "content": "Without losing generality, we use the image-text retrieval task as a proxy to investigate the noisy correspondence problem in cross-modal retrieval. Given a training set " + }, + { + "bbox": [ + 304, + 665, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{train} = \\{(I_i,T_i,y_i)\\}_{i = 1}^N" + }, + { + "bbox": [ + 304, + 665, + 545, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 665, + 545, + 715 + ], + "type": "inline_equation", + "content": "(I_{i},T_{i})" + }, + { + "bbox": [ + 304, + 665, + 545, + 715 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 304, + 665, + 545, + 715 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 665, + 545, + 715 + ], + "type": "text", + "content": "-th image-" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7518" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 70, + 547, + 230 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 547, + 230 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 547, + 230 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 547, + 230 + ], + "type": "image", + "image_path": "7714af4820a40f75b9ed577c4b180ace7b837498f5c705b5a0cff94a4f9377aa.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 239, + 547, + 273 + ], + "lines": [ + { + "bbox": [ + 46, + 239, + 547, + 273 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 547, + 273 + ], + "type": "text", + "content": "Figure 2. Overview of the proposed meta-process. The positive pairs are taken from the original meta-data which contain corresponding image and text, while negative pairs are extracting from different pairs in training set to construct mismatched pairs. Then the MSCN is learned to determine whether the modalities correspond to each other or not." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 287, + 289, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 287, + 289, + 371 + ], + "spans": [ + { + "bbox": [ + 46, + 287, + 289, + 371 + ], + "type": "text", + "content": "text pair, " + }, + { + "bbox": [ + 46, + 287, + 289, + 371 + ], + "type": "inline_equation", + "content": "y_{i} \\in \\{1,0\\}" + }, + { + "bbox": [ + 46, + 287, + 289, + 371 + ], + "type": "text", + "content": " indicates the pair is positive (matched), i.e. " + }, + { + "bbox": [ + 46, + 287, + 289, + 371 + ], + "type": "inline_equation", + "content": "(I_i, T_i)" + }, + { + "bbox": [ + 46, + 287, + 289, + 371 + ], + "type": "text", + "content": ", or negative (mismatched), i.e. " + }, + { + "bbox": [ + 46, + 287, + 289, + 371 + ], + "type": "inline_equation", + "content": "(I_i, T_{j \\neq i})" + }, + { + "bbox": [ + 46, + 287, + 289, + 371 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 287, + 289, + 371 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 287, + 289, + 371 + ], + "type": "text", + "content": " is the number of the entire training data. The noisy correspondence in bimodal data is the negative pair but wrongly considered as " + }, + { + "bbox": [ + 46, + 287, + 289, + 371 + ], + "type": "inline_equation", + "content": "y_{i} = 1" + }, + { + "bbox": [ + 46, + 287, + 289, + 371 + ], + "type": "text", + "content": ". To address this issue, we propose a meta-learning based method to achieve robust training. The details are delineated next." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 382, + 275, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 382, + 275, + 396 + ], + "spans": [ + { + "bbox": [ + 47, + 382, + 275, + 396 + ], + "type": "text", + "content": "3.2. Meta-Learning Based Similarity Correction" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 402, + 287, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 402, + 287, + 485 + ], + "spans": [ + { + "bbox": [ + 46, + 402, + 287, + 485 + ], + "type": "text", + "content": "The Training Objective. To begin with, we describe the standard projection module to enable cross-modal features to be comparable. Let " + }, + { + "bbox": [ + 46, + 402, + 287, + 485 + ], + "type": "inline_equation", + "content": "f(I;W_f)" + }, + { + "bbox": [ + 46, + 402, + 287, + 485 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 402, + 287, + 485 + ], + "type": "inline_equation", + "content": "g(T;W_g)" + }, + { + "bbox": [ + 46, + 402, + 287, + 485 + ], + "type": "text", + "content": " be the modal-specific networks to map the visual and textual modalities into the joint embedding space, respectively. Then we compute the similarity feature between bimodal features by the function" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 63, + 497, + 287, + 525 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 497, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 63, + 497, + 287, + 525 + ], + "type": "interline_equation", + "content": "S (f (I), g (T); W _ {s}) = \\frac {W _ {s} | f (I) - g (T) | ^ {2}}{\\| W _ {s} | f (I) - g (T) | ^ {2} \\| _ {2}}, \\tag {1}", + "image_path": "22a38a504f51d1a0ff9f12ae5114d1ade0f4947dac06754e9e8f62621b4492a7.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 533, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 289, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 533, + 289, + 715 + ], + "type": "inline_equation", + "content": "W_{s}" + }, + { + "bbox": [ + 46, + 533, + 289, + 715 + ], + "type": "text", + "content": " is a learnable parameter matrix to obtain low-dimensional similarity representation. For notation convenience, we collect " + }, + { + "bbox": [ + 46, + 533, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\{W_{f},W_{g},W_{s}\\}" + }, + { + "bbox": [ + 46, + 533, + 289, + 715 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 46, + 533, + 289, + 715 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 46, + 533, + 289, + 715 + ], + "type": "text", + "content": " as the parameters of main net " + }, + { + "bbox": [ + 46, + 533, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{F} = (f,g,S)" + }, + { + "bbox": [ + 46, + 533, + 289, + 715 + ], + "type": "text", + "content": " and denote similarity feature as " + }, + { + "bbox": [ + 46, + 533, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_W(I,T)" + }, + { + "bbox": [ + 46, + 533, + 289, + 715 + ], + "type": "text", + "content": ". With the presence of noisy correspondence, the model will inevitably overfit to noise if we use the similarity representation directly. To tackle this problem, we propose a meta similarity correction network (MSCN), regarding process of obtaining similarity score as a meta-process, which takes the noisy similarity representation as input and produces the corrected similarity score. Specifically, our MSCN is formulated as a MLP network " + }, + { + "bbox": [ + 46, + 533, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_{\\Theta}(\\cdot)" + }, + { + "bbox": [ + 46, + 533, + 289, + 715 + ], + "type": "text", + "content": " with parameters " + }, + { + "bbox": [ + 46, + 533, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 46, + 533, + 289, + 715 + ], + "type": "text", + "content": ". The output " + }, + { + "bbox": [ + 46, + 533, + 289, + 715 + ], + "type": "inline_equation", + "content": "s_i = \\mathcal{V}_{\\Theta}(\\mathcal{F}_W(I_i,T_i))" + }, + { + "bbox": [ + 46, + 533, + 289, + 715 + ], + "type": "text", + "content": " is activated with Sigmoid function to represent the similarity score located in [0, 1]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 287, + 547, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 287, + 547, + 395 + ], + "spans": [ + { + "bbox": [ + 304, + 287, + 547, + 395 + ], + "type": "text", + "content": "For the noisy training pairs, the main net learns noisetolerant features guided by the corrected similarity scores produced from MSCN; while the MSCN predicts the reliable similarity score based on features from main net. This interdependency enables the main net and meta net reinforce each other against noise, and optimized via a bi-level optimization. Formally, the optimal parameters " + }, + { + "bbox": [ + 304, + 287, + 547, + 395 + ], + "type": "inline_equation", + "content": "W^{*}" + }, + { + "bbox": [ + 304, + 287, + 547, + 395 + ], + "type": "text", + "content": " is calculated by minimizing the cumulative loss over training data:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 321, + 413, + 545, + 434 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 413, + 545, + 434 + ], + "spans": [ + { + "bbox": [ + 321, + 413, + 545, + 434 + ], + "type": "interline_equation", + "content": "W ^ {*} (\\Theta) = \\underset {W} {\\arg \\min } \\mathbb {E} _ {(I _ {i}, T _ {i}) \\in \\mathcal {D} _ {t r a i n}} l ^ {t r a i n} (I _ {i}, T _ {i}), \\tag {2}", + "image_path": "faa8352b5e7de170ac9a1d67ea68112dd7e6b3423821669c17ea6a9ed590f62b.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 440, + 473, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 440, + 473, + 453 + ], + "spans": [ + { + "bbox": [ + 306, + 440, + 473, + 453 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 440, + 473, + 453 + ], + "type": "inline_equation", + "content": "l^{train}(\\cdot)" + }, + { + "bbox": [ + 306, + 440, + 473, + 453 + ], + "type": "text", + "content": " is the triplet ranking loss:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 307, + 467, + 545, + 505 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 467, + 545, + 505 + ], + "spans": [ + { + "bbox": [ + 307, + 467, + 545, + 505 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} l ^ {t r a i n} \\left(I _ {i}, T _ {i}\\right) = [ \\hat {\\gamma} - \\mathcal {V} _ {\\Theta} \\left(\\mathcal {F} _ {W} \\left(I _ {i}, T _ {i}\\right)\\right) + \\mathcal {V} _ {\\Theta} \\left(\\mathcal {F} _ {W} \\left(I _ {i}, T _ {i} ^ {-}\\right)\\right) ] _ {+} \\\\ + \\left[ \\hat {\\gamma} - \\mathcal {V} _ {\\Theta} \\left(\\mathcal {F} _ {W} \\left(I _ {i}, T _ {i}\\right)\\right) + \\mathcal {V} _ {\\Theta} \\left(\\mathcal {F} _ {W} \\left(I _ {i} ^ {-}, T _ {i}\\right)\\right) \\right] _ {+,} \\tag {3} \\\\ \\end{array}", + "image_path": "302e8744aee5d850a97cd10590973157baeee178aeb6a8cdb5fb089b5b166472.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 506, + 545, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 506, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 304, + 506, + 545, + 589 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 506, + 545, + 589 + ], + "type": "inline_equation", + "content": "[x]_{+} = max(x,0), I_{i}^{-}" + }, + { + "bbox": [ + 304, + 506, + 545, + 589 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 506, + 545, + 589 + ], + "type": "inline_equation", + "content": "T_{i}^{-}" + }, + { + "bbox": [ + 304, + 506, + 545, + 589 + ], + "type": "text", + "content": " are the hardest negatives corresponding to given pair " + }, + { + "bbox": [ + 304, + 506, + 545, + 589 + ], + "type": "inline_equation", + "content": "(I_i,T_i)" + }, + { + "bbox": [ + 304, + 506, + 545, + 589 + ], + "type": "text", + "content": " similar to VSE++ [8]. To achieve robust cross-modal retrieval, we design a self-adaptive margin " + }, + { + "bbox": [ + 304, + 506, + 545, + 589 + ], + "type": "inline_equation", + "content": "\\hat{\\gamma}" + }, + { + "bbox": [ + 304, + 506, + 545, + 589 + ], + "type": "text", + "content": " instead of a fixed value in VSE++. Ideally, the margin should put less value on the noisy pairs and more on the clean pairs. To this end, we adjust the margin according to the similarity score:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 380, + 594, + 545, + 622 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 594, + 545, + 622 + ], + "spans": [ + { + "bbox": [ + 380, + 594, + 545, + 622 + ], + "type": "interline_equation", + "content": "\\hat {\\gamma} = \\frac {1}{1 + \\left(\\frac {s _ {i}}{1 - s _ {i}}\\right) ^ {- \\tau}} \\gamma , \\tag {4}", + "image_path": "0e5aa49eeb8c443c291b3cafeac6640f3cdefc4251b9cf8847bed28779a46fd6.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 628, + 545, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 628, + 545, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 628, + 545, + 676 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 628, + 545, + 676 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 628, + 545, + 676 + ], + "type": "text", + "content": " is the original margin value and " + }, + { + "bbox": [ + 304, + 628, + 545, + 676 + ], + "type": "inline_equation", + "content": "\\tau > 0" + }, + { + "bbox": [ + 304, + 628, + 545, + 676 + ], + "type": "text", + "content": " is the parameter. This self-adaptive margin will enable the triplet loss put more attention on the samples with high similarity scores that possibly are clean pairs." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "content": "Meta Training Process. To ensure the reliability of the similarity scores produced by MSCN, we utilize a small" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7519" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 77, + 545, + 246 + ], + "blocks": [ + { + "bbox": [ + 53, + 77, + 545, + 246 + ], + "lines": [ + { + "bbox": [ + 53, + 77, + 545, + 246 + ], + "spans": [ + { + "bbox": [ + 53, + 77, + 545, + 246 + ], + "type": "image", + "image_path": "8aeb08c230bb42145417fb7269135a2ab216757ab23d687d75dfe124ef12a91d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 257, + 547, + 282 + ], + "lines": [ + { + "bbox": [ + 46, + 257, + 547, + 282 + ], + "spans": [ + { + "bbox": [ + 46, + 257, + 547, + 282 + ], + "type": "text", + "content": "Figure 3. Flowchart of the proposed bi-level optimization. At each iteration, we alternately update the main net and the meta net (MSCN) following the steps in Eq. (7) - (9)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 294, + 289, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 294, + 289, + 594 + ], + "spans": [ + { + "bbox": [ + 47, + 294, + 289, + 594 + ], + "type": "text", + "content": "amount clean meta-data set " + }, + { + "bbox": [ + 47, + 294, + 289, + 594 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{meta} = \\{(I_i,T_i)\\}_{i = 1}^M" + }, + { + "bbox": [ + 47, + 294, + 289, + 594 + ], + "type": "text", + "content": " to guide the training, where " + }, + { + "bbox": [ + 47, + 294, + 289, + 594 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 47, + 294, + 289, + 594 + ], + "type": "text", + "content": " is much smaller than " + }, + { + "bbox": [ + 47, + 294, + 289, + 594 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 47, + 294, + 289, + 594 + ], + "type": "text", + "content": ". While meta-learning has been extensively studied in recent literature [10, 32, 46], it remains challenging to construct the meta-process in our objective. On the one hand, the similarity features from main net can not be used directly to produce loss. On the other hand, the meta-process should enable the output of MSCN to represent the similarity score. To achieve this, we view a binary classification task as the meta-process: given an image-text pair, the MSCN is learned to determine whether the modalities correspond to each other or not. Specifically, the positive pairs are taken from the meta-data which contain corresponding image and text, i.e. " + }, + { + "bbox": [ + 47, + 294, + 289, + 594 + ], + "type": "inline_equation", + "content": "(I_{i},T_{i},y_{i} = 1)" + }, + { + "bbox": [ + 47, + 294, + 289, + 594 + ], + "type": "text", + "content": ", while negative pairs are extracting from different pairs in training set to construct mismatched pairs, i.e. " + }, + { + "bbox": [ + 47, + 294, + 289, + 594 + ], + "type": "inline_equation", + "content": "(I_{i},T_{j\\neq i},y_{i} = 0)" + }, + { + "bbox": [ + 47, + 294, + 289, + 594 + ], + "type": "text", + "content": ". The binary labels are produced from the data itself, and can be regarded as the ideal similarity score to guide the training of MSCN. The prediction of MSCN measures the probability of a pair being clean which can be naturally equivalent to the similarity score. Formally, in each iteration of training, we randomly construct " + }, + { + "bbox": [ + 47, + 294, + 289, + 594 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 47, + 294, + 289, + 594 + ], + "type": "text", + "content": " mismatched pairs from training set to extend the meta-data set as " + }, + { + "bbox": [ + 47, + 294, + 289, + 594 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{meta}' = \\{(I_i,T_i,y_i)\\}_{i = 1}^{2M}" + }, + { + "bbox": [ + 47, + 294, + 289, + 594 + ], + "type": "text", + "content": ". Then the optimal parameter " + }, + { + "bbox": [ + 47, + 294, + 289, + 594 + ], + "type": "inline_equation", + "content": "\\Theta^{*}" + }, + { + "bbox": [ + 47, + 294, + 289, + 594 + ], + "type": "text", + "content": " can be learned by minimizing the following meta-loss calculated on meta-data:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 609, + 287, + 630 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 609, + 287, + 630 + ], + "spans": [ + { + "bbox": [ + 61, + 609, + 287, + 630 + ], + "type": "interline_equation", + "content": "\\Theta^ {*} = \\underset {\\Theta} {\\arg \\min } \\mathbb {E} _ {(I _ {i}, T _ {i}) \\in \\mathcal {D} _ {m e t a} ^ {\\prime}} l ^ {m e t a} (I _ {i}, T _ {i}, y _ {i}), \\tag {5}", + "image_path": "508d5b6ff74b0f28ed7a04613a141eef585094a7f0478fe896aa0e1fb0fbb603.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 637, + 230, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 637, + 230, + 651 + ], + "spans": [ + { + "bbox": [ + 47, + 637, + 230, + 651 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 637, + 230, + 651 + ], + "type": "inline_equation", + "content": "l^{meta}(\\cdot)" + }, + { + "bbox": [ + 47, + 637, + 230, + 651 + ], + "type": "text", + "content": " is cross-entropy loss [35, 37]:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 656, + 287, + 671 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 656, + 287, + 671 + ], + "spans": [ + { + "bbox": [ + 57, + 656, + 287, + 671 + ], + "type": "interline_equation", + "content": "l ^ {m e t a} (I _ {i}, T _ {i}, y _ {i}) = - y _ {i} \\cdot l o g \\mathcal {V} _ {\\Theta} \\left(\\mathcal {F} _ {W ^ {*} (\\Theta)} (I _ {i}, T _ {i})\\right). (6)", + "image_path": "8cfabf5f2d2b96258a29788e80ef674e4c09d60176a53efd9c18ce78836638db.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "type": "text", + "content": "As illustrated in Fig. 2, despite its simplicity, our proposed meta-process can not only enforce the MSCN to learn reliable similarity score but also has the following advantages:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 294, + 547, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 294, + 547, + 403 + ], + "spans": [ + { + "bbox": [ + 304, + 294, + 547, + 403 + ], + "type": "text", + "content": "1) previous works only guide the meta net with ground truth data; however, we leverage both positive and negative data as meta-knowledge that encourage the MSCN to learn discrimination from them. 2) We overcome one limitation of meta-learning methods that the lack of meta-data. We extract from different pairs among training set to construct massive negative meta-data. Moreover, it is more suitable to the rule of noisy correspondence, which the irregularity of noisy data causes in various patterns." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 417, + 547, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 417, + 547, + 536 + ], + "spans": [ + { + "bbox": [ + 305, + 417, + 547, + 536 + ], + "type": "text", + "content": "Bi-level Optimization. Motivated by recent works [9, 16, 30, 32], we use a bi-level optimization strategy to guarantee the efficiency for updating our main net and MSCN. As illustrated in Fig. 3, it is practically solved via two nested loops of optimization. Specifically, at each iteration " + }, + { + "bbox": [ + 305, + 417, + 547, + 536 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 305, + 417, + 547, + 536 + ], + "type": "text", + "content": " of training, we sample a mini-batch of training pairs " + }, + { + "bbox": [ + 305, + 417, + 547, + 536 + ], + "type": "inline_equation", + "content": "\\{(I_i,T_i)\\}_{i = 1}^n" + }, + { + "bbox": [ + 305, + 417, + 547, + 536 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 305, + 417, + 547, + 536 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 305, + 417, + 547, + 536 + ], + "type": "text", + "content": " is the size of mini-batch. We first update the main net parameters by taking a gradient descent step toward the direction of triplet loss, which can be formulated as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 318, + 543, + 546, + 574 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 543, + 546, + 574 + ], + "spans": [ + { + "bbox": [ + 318, + 543, + 546, + 574 + ], + "type": "interline_equation", + "content": "\\hat {W} ^ {(t)} (\\Theta) = W ^ {(t)} - \\alpha \\sum_ {i = 1} ^ {n} \\left. \\nabla_ {W} l ^ {\\text {t r a i n}} \\left(I _ {i}, T _ {i}\\right) \\right| _ {W ^ {(t)}}, \\tag {7}", + "image_path": "6a8d4bd142cf5b2fec336042ed3f267ab854b4bc520158267094a41c10af1a96.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 582, + 547, + 678 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 678 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 678 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 582, + 547, + 678 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 582, + 547, + 678 + ], + "type": "text", + "content": " means the learning rate. The updated " + }, + { + "bbox": [ + 304, + 582, + 547, + 678 + ], + "type": "inline_equation", + "content": "\\hat{W}^{(t)}" + }, + { + "bbox": [ + 304, + 582, + 547, + 678 + ], + "type": "text", + "content": " can be regard as the approximation of the optimal " + }, + { + "bbox": [ + 304, + 582, + 547, + 678 + ], + "type": "inline_equation", + "content": "W^{*}(\\Theta)" + }, + { + "bbox": [ + 304, + 582, + 547, + 678 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 304, + 582, + 547, + 678 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 582, + 547, + 678 + ], + "type": "text", + "content": " iteration, containing the gradient information of " + }, + { + "bbox": [ + 304, + 582, + 547, + 678 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 304, + 582, + 547, + 678 + ], + "type": "text", + "content": ". Then we update the parameters of MSCN with main net parameters fixed. Guided by the objective loss in Eq. (5) and a minibatch of meta-data " + }, + { + "bbox": [ + 304, + 582, + 547, + 678 + ], + "type": "inline_equation", + "content": "\\{(I_i,T_i,y_i)\\}_{i = 1}^m" + }, + { + "bbox": [ + 304, + 582, + 547, + 678 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 582, + 547, + 678 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 582, + 547, + 678 + ], + "type": "text", + "content": " is the minibatch size. The updated parameters of MSCN is calculated as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 315, + 685, + 546, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 685, + 546, + 717 + ], + "spans": [ + { + "bbox": [ + 315, + 685, + 546, + 717 + ], + "type": "interline_equation", + "content": "\\left. \\Theta^ {(t + 1)} = \\Theta^ {(t)} - \\beta \\frac {1}{m} \\sum_ {i = 1} ^ {m} \\nabla_ {\\Theta} l ^ {m e t a} \\left(I _ {i}, T _ {i}, y _ {i}\\right) \\right| _ {\\Theta^ {(t)}}, \\tag {8}", + "image_path": "1de3e33652b1cd7a28348837d43bb730f35e540bb09fdb7dfca5c011f63ea236.jpg" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7520" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "content": " is the learning rate for meta-process. Finally, we update the parameters of main net while fixing the MSCN parameters. With the corrected similarity scores produced by the newly updated " + }, + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "inline_equation", + "content": "\\Theta^{(t + 1)}" + }, + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "content": ", we can update the main net parameters " + }, + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "inline_equation", + "content": "W^{(t + 1)}" + }, + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "content": " by employing the gradient decent as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 63, + 142, + 287, + 173 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 142, + 287, + 173 + ], + "spans": [ + { + "bbox": [ + 63, + 142, + 287, + 173 + ], + "type": "interline_equation", + "content": "W ^ {(t + 1)} = W ^ {(t)} - \\alpha \\sum_ {i = 1} ^ {n} \\left. \\nabla_ {W} l ^ {\\text {t r a i n}} \\left(I _ {i}, T _ {i}\\right) \\right| _ {W ^ {(t)}}. \\tag {9}", + "image_path": "e2e288428e4c64ac32bc8aa3a065940d40468330c11d3cd499d0892721a69ab4.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 182, + 287, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 182, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 46, + 182, + 287, + 217 + ], + "type": "text", + "content": "The above procedure proceeds in an alternating manner to optimize the main net and MSCN until it exceeds the maximal iteration." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 227, + 271, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 227, + 271, + 239 + ], + "spans": [ + { + "bbox": [ + 47, + 227, + 271, + 239 + ], + "type": "text", + "content": "3.3. Meta-Knowledge Guided Data Purification" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 245, + 287, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 245, + 287, + 497 + ], + "spans": [ + { + "bbox": [ + 46, + 245, + 287, + 497 + ], + "type": "text", + "content": "As mentioned, our MSCN can provide corrected similarity score to avoid noise. However, due to the property of triplet loss, it remains to produce positive loss for the noisy pairs even if we employ the ideal similarity score (e.g., " + }, + { + "bbox": [ + 46, + 245, + 287, + 497 + ], + "type": "inline_equation", + "content": "\\mathcal{V}(F(I_i,T_i))\\approx 0 < \\mathcal{V}\\left(F(I_i^-,T_i)\\right) + \\hat{\\gamma})" + }, + { + "bbox": [ + 46, + 245, + 287, + 497 + ], + "type": "text", + "content": ", which results in the model fitting to the noise. To tackle this issue, we adopt sample selection strategy to remove the noisy data and only use clean samples as training data. Previous works [2, 42] motivated by the memorization effect of DNNs that clean sample has a lower loss during the beginning of training. Unfortunately, the triplet ranking loss is more complex, i.e., affected by both positive and negative samples, which leads to sub-optimal selection results. Based on the observation shown in Fig. 4, the clean and noisy pairs are prone to distinguish from the similarity scores. Gaussian Mixture Model (GMM) is a widely used technique to model mixture distribution [15, 20]. However, we find the similarity distribution exhibits high skew toward 1 and causes a poor approximation for GMM. Therefore, we fit a two-component Beta Mixture Model (BMM) [27] of per-pair similarity score " + }, + { + "bbox": [ + 46, + 245, + 287, + 497 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 46, + 245, + 287, + 497 + ], + "type": "text", + "content": " to better distinguish clean and noisy samples:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 107, + 507, + 287, + 539 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 507, + 287, + 539 + ], + "spans": [ + { + "bbox": [ + 107, + 507, + 287, + 539 + ], + "type": "interline_equation", + "content": "p \\left(s _ {i}\\right) = \\sum_ {k = 1} ^ {K} \\lambda_ {k} \\phi \\left(s _ {i} \\mid \\alpha_ {k}, \\beta_ {k}\\right), \\tag {10}", + "image_path": "ea43b46f6e22ff79ea5fe724065109e12904e1e12e4cfb33a1f8f0f0e8e3cf46.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 545, + 287, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 287, + 581 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 545, + 287, + 581 + ], + "type": "inline_equation", + "content": "K = 2" + }, + { + "bbox": [ + 46, + 545, + 287, + 581 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 545, + 287, + 581 + ], + "type": "inline_equation", + "content": "\\lambda_{k}" + }, + { + "bbox": [ + 46, + 545, + 287, + 581 + ], + "type": "text", + "content": " is the mixture coefficient, and " + }, + { + "bbox": [ + 46, + 545, + 287, + 581 + ], + "type": "inline_equation", + "content": "\\phi (s|\\alpha_k,\\beta_k)" + }, + { + "bbox": [ + 46, + 545, + 287, + 581 + ], + "type": "text", + "content": " denotes the probability density function with parameters " + }, + { + "bbox": [ + 46, + 545, + 287, + 581 + ], + "type": "inline_equation", + "content": "\\alpha_{k},\\beta_{k} > 0" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 582, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 714 + ], + "type": "text", + "content": "We use the Expectation-Maximization [5] algorithm to optimize the BMM, which poses a crucial challenge: the initialization of model parameters. Existing methods typically adopt the " + }, + { + "bbox": [ + 46, + 582, + 287, + 714 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 582, + 287, + 714 + ], + "type": "text", + "content": "-Means algorithm [23] for initializing; however, it is time-consuming (i.e., the time complexity is " + }, + { + "bbox": [ + 46, + 582, + 287, + 714 + ], + "type": "inline_equation", + "content": "O(TKN)" + }, + { + "bbox": [ + 46, + 582, + 287, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 582, + 287, + 714 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 582, + 287, + 714 + ], + "type": "text", + "content": " is the number of iterations) and underperforming. We thus propose a novel initialization approach which takes the meta-data as prior knowledge. Specifically, we fit the two Beta components with positive and negative meta-data to model the distribution of clean and noisy pairs, respectively. Given the similarity scores " + }, + { + "bbox": [ + 46, + 582, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{S}_p = \\{s_i\\}_{i=1}^M" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 307, + 73, + 542, + 225 + ], + "blocks": [ + { + "bbox": [ + 307, + 73, + 542, + 225 + ], + "lines": [ + { + "bbox": [ + 307, + 73, + 542, + 225 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 542, + 225 + ], + "type": "image", + "image_path": "79d0c53fb0b4801da8eba6d1025d4c5f036e16be40e0fe4988bd68563c7e1a9d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 232, + 545, + 266 + ], + "lines": [ + { + "bbox": [ + 305, + 232, + 545, + 266 + ], + "spans": [ + { + "bbox": [ + 305, + 232, + 545, + 266 + ], + "type": "text", + "content": "Figure 4. Per-sample similarity score distribution, estimated BMM and BMM with meta prior knowledge under " + }, + { + "bbox": [ + 305, + 232, + 545, + 266 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 305, + 232, + 545, + 266 + ], + "type": "text", + "content": " noisy correspondence in Flickr30K after 5 epochs warmup." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 285, + 545, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 285, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 305, + 285, + 545, + 308 + ], + "type": "text", + "content": "corresponding to positive meta-data, the parameters of the component modeling clean distribution can be initialled by" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 347, + 317, + 545, + 350 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 317, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 347, + 317, + 545, + 350 + ], + "type": "interline_equation", + "content": "\\alpha_ {k} = \\frac {\\left(1 - E \\left(\\mathcal {S} _ {p}\\right)\\right) E \\left(\\mathcal {S} _ {p}\\right) ^ {2}}{V \\left(\\mathcal {S} _ {p}\\right)} - E \\left(\\mathcal {S} _ {p}\\right), \\tag {11}", + "image_path": "c8e3049067f58dc9963b668211981f2016c71ff24e910a5904dab5c6317997d7.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 348, + 347, + 446, + 373 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 347, + 446, + 373 + ], + "spans": [ + { + "bbox": [ + 348, + 347, + 446, + 373 + ], + "type": "interline_equation", + "content": "\\beta_ {k} = \\frac {\\alpha_ {k} \\left(1 - E \\left(\\mathcal {S} _ {p}\\right)\\right)}{E (\\mathcal {S} _ {p})},", + "image_path": "a7313419154be314ba0682521da696bc533864c78b455c9f71fc3df7830748c4.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 382, + 545, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 382, + 545, + 465 + ], + "spans": [ + { + "bbox": [ + 306, + 382, + 545, + 465 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 382, + 545, + 465 + ], + "type": "inline_equation", + "content": "E(\\cdot)" + }, + { + "bbox": [ + 306, + 382, + 545, + 465 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 382, + 545, + 465 + ], + "type": "inline_equation", + "content": "V(\\cdot)" + }, + { + "bbox": [ + 306, + 382, + 545, + 465 + ], + "type": "text", + "content": " are the mean and variance function, respectively. The initialization of component modeling noisy distribution is analogous but uses similarity scores " + }, + { + "bbox": [ + 306, + 382, + 545, + 465 + ], + "type": "inline_equation", + "content": "S_{n} = \\{s_{i}\\}_{i=1}^{M}" + }, + { + "bbox": [ + 306, + 382, + 545, + 465 + ], + "type": "text", + "content": " of negative meta-data. As illustrated in Fig. 4, our meta-knowledge guided initialization method achieves better approximation while with lower time complexity (i.e., " + }, + { + "bbox": [ + 306, + 382, + 545, + 465 + ], + "type": "inline_equation", + "content": "O(KM)" + }, + { + "bbox": [ + 306, + 382, + 545, + 465 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 466, + 545, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 466, + 545, + 512 + ], + "spans": [ + { + "bbox": [ + 305, + 466, + 545, + 512 + ], + "type": "text", + "content": "Finally, we compute the posterior probability " + }, + { + "bbox": [ + 305, + 466, + 545, + 512 + ], + "type": "inline_equation", + "content": "p_i = p(k|s_i) = p(k)p(s_i|\\alpha_k,\\beta_k) / p(s_i)" + }, + { + "bbox": [ + 305, + 466, + 545, + 512 + ], + "type": "text", + "content": " as the probability of " + }, + { + "bbox": [ + 305, + 466, + 545, + 512 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 305, + 466, + 545, + 512 + ], + "type": "text", + "content": "-th simple being clean. The purified training set is defined as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 345, + 514, + 545, + 527 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 514, + 545, + 527 + ], + "spans": [ + { + "bbox": [ + 345, + 514, + 545, + 527 + ], + "type": "interline_equation", + "content": "\\mathcal {D} _ {\\text {t r a i n}} ^ {\\prime} = \\left\\{\\left(I _ {i}, T _ {i}\\right) \\in \\mathcal {D} _ {\\text {t r a i n}} \\mid p _ {i} > 0. 5 \\right\\}. \\tag {12}", + "image_path": "ddb485f263ac27a5920f7d881b6caf2a119d5efc9d085e5e6d8947c56e1d8b12.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 533, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 533, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 305, + 533, + 545, + 605 + ], + "type": "text", + "content": "At each iteration, we produce the purified dataset before the training of our meta-learning objective. Moreover, following [12, 15, 20], we maintain two networks " + }, + { + "bbox": [ + 305, + 533, + 545, + 605 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{F}_W^1,\\mathcal{V}_{\\Theta}^1\\}" + }, + { + "bbox": [ + 305, + 533, + 545, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 533, + 545, + 605 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{F}_W^2,\\mathcal{V}_{\\Theta}^2\\}" + }, + { + "bbox": [ + 305, + 533, + 545, + 605 + ], + "type": "text", + "content": " to avoid error accumulation that one network will produce the purified training set to train the other one. The full algorithm is shown in supplementary material." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 615, + 382, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 615, + 382, + 628 + ], + "spans": [ + { + "bbox": [ + 306, + 615, + 382, + 628 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 635, + 484, + 646 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 635, + 484, + 646 + ], + "spans": [ + { + "bbox": [ + 306, + 635, + 484, + 646 + ], + "type": "text", + "content": "4.1. Datasets and Evaluation Protocol" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 653, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 653, + 545, + 713 + ], + "type": "text", + "content": "Datasets. We evaluate our method on three standard image-text retrieval datasets. To be specific, the Flickr30K [43] collects 31,783 images with 5 corresponding captions each from Flickr website. Following the split in [15], we use 1,000 image-text pairs for validation, 1,000 image-text" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "7521" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": "pairs for testing, and the rest for training. The MS-COCO [24] contains 123,287 images with 5 corresponding captions each. We split the MS-COCO following [15] that 113,287 image-text pairs for training, 5,000 pairs for validation, and 5,000 pairs for testing. Conceptual Captions [31] is a large-scale dataset contains 3.3M images with one corresponding caption each. We employ the same subset following [15], i.e., CC152K. Specifically, CC152K consists of 150,000 pairs from training set for training, 1,000 pairs from validation set for validation, and 1,000 pairs from validation set for testing. Moreover, we use an extra meta-data set from validation containing 3,000 pairs." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 236, + 288, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 236, + 288, + 331 + ], + "spans": [ + { + "bbox": [ + 46, + 236, + 288, + 331 + ], + "type": "text", + "content": "Evaluation Protocol. We evaluate the retrieval performance with the recall rate at K (R@K) metric. In a nutshell, R@K is the proportion for the relevant items retrieved in the closest K items to the query. In our experiments, we take image and text as queries, respectively, and report R@1, R@5, and R@10 results for a comprehensive evaluation. Following [15], we average the similarity scores from two networks at the inference phase." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 342, + 180, + 354 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 342, + 180, + 354 + ], + "spans": [ + { + "bbox": [ + 47, + 342, + 180, + 354 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 361, + 289, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 361, + 289, + 601 + ], + "spans": [ + { + "bbox": [ + 46, + 361, + 289, + 601 + ], + "type": "text", + "content": "As a general approach, our method can be easily applied to almost all cross-modal retrieval methods to improve robustness. For fair comparison, we chose the same network backbones with NCR [15], i.e., the projection modules and similarity function. For all datasets, the training processes contain 50 epochs after a 5 epochs warmup. We employ ADAM [17] as the optimizer for both main net and meta net with a batch size of 64. To fit the BMM, we set the stop threshold as " + }, + { + "bbox": [ + 46, + 361, + 289, + 601 + ], + "type": "inline_equation", + "content": "10^{-2}" + }, + { + "bbox": [ + 46, + 361, + 289, + 601 + ], + "type": "text", + "content": " and maximum number of iterations as 10 for the EM procedure. Note that, we clamp the similarity scores into a range of " + }, + { + "bbox": [ + 46, + 361, + 289, + 601 + ], + "type": "inline_equation", + "content": "[10^{-4}, 1 - 10^{-4}]" + }, + { + "bbox": [ + 46, + 361, + 289, + 601 + ], + "type": "text", + "content": " instead of [0,1] for stable fitting. Moreover, we set the margin parameter as 0.2 and " + }, + { + "bbox": [ + 46, + 361, + 289, + 601 + ], + "type": "inline_equation", + "content": "\\tau = 2" + }, + { + "bbox": [ + 46, + 361, + 289, + 601 + ], + "type": "text", + "content": " to calculate the self-adaptive margin. For all datasets, the number of ground-truth meta-data is approximately " + }, + { + "bbox": [ + 46, + 361, + 289, + 601 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 46, + 361, + 289, + 601 + ], + "type": "text", + "content": " of training data, and the learning rate of main net and meta net are initialized with " + }, + { + "bbox": [ + 46, + 361, + 289, + 601 + ], + "type": "inline_equation", + "content": "2 \\times 10^{-4}" + }, + { + "bbox": [ + 46, + 361, + 289, + 601 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 361, + 289, + 601 + ], + "type": "inline_equation", + "content": "1.7 \\times 10^{-5}" + }, + { + "bbox": [ + 46, + 361, + 289, + 601 + ], + "type": "text", + "content": ", respectively. For CC152K containing real noise, we employ the validation data as the meta set due to the lack of ground-truth data. For all experiments, we decay the learning rate by 0.1 after 30 epochs." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 610, + 247, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 610, + 247, + 624 + ], + "spans": [ + { + "bbox": [ + 47, + 610, + 247, + 624 + ], + "type": "text", + "content": "4.3. Comparison with the State-of-the-Art" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "text", + "content": "In this section, we conduct comparison experiments with 7 state-of-the-art methods that include SCAN [18], VSRN [22], IMRAM [4], SGR, SAF, SGRAF [7], and NCR [15]. As Flickr30K and MS-COCO are well annotated datasets, we generate synthetic noisy correspondence by randomly shuffling the training images and captions for a specific noise ratio, and we conduct experiments with three different" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 307, + 71, + 549, + 345 + ], + "blocks": [ + { + "bbox": [ + 307, + 71, + 549, + 345 + ], + "lines": [ + { + "bbox": [ + 307, + 71, + 549, + 345 + ], + "spans": [ + { + "bbox": [ + 307, + 71, + 549, + 345 + ], + "type": "table", + "html": "
Image to Text
MethodsR@1R@5R@10SUM
SCAN (ECCV'18)30.555.365.3151.1
VSRN(ICCV'19)32.661.370.5164.4
IMRAM(CVPR'20)33.157.668.1158.8
SAF (AAAI'21)31.759.368.2159.2
SGR (AAAI'21)11.329.739.680.6
SGR* (AAAI'21)35.063.473.3171.7
NCR(NIPS'21)39.564.573.5177.5
MSCN*(Ours)39.765.475.3180.4
MSCN(Ours)40.165.776.6182.4
Text to Image
SCAN (ECCV'18)26.953.064.7144.6
VSRN(ICCV'19)32.559.470.4162.3
IMRAM(CVPR'20)29.056.867.4153.2
SAF (AAAI'21)31.959.067.9158.8
SGR (AAAI'21)13.130.141.684.8
SGR* (AAAI'21)34.963.072.8170.7
NCR(NIPS'21)40.364.673.2178.1
MSCN*(Ours)39.866.175.0180.9
MSCN(Ours)40.667.476.3184.3
", + "image_path": "ff3b2686aa389e0441a49234352caa85e26771c520baa0eb273df3f873cd9d97.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 353, + 545, + 374 + ], + "lines": [ + { + "bbox": [ + 305, + 353, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 305, + 353, + 545, + 374 + ], + "type": "text", + "content": "Table 1. Image-Text Retrieval Performance on CC152K with real noise, and the best results are highlighted in bold." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 384, + 545, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 384, + 545, + 517 + ], + "spans": [ + { + "bbox": [ + 304, + 384, + 545, + 517 + ], + "type": "text", + "content": "level of noise ratios, i.e., " + }, + { + "bbox": [ + 304, + 384, + 545, + 517 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 304, + 384, + 545, + 517 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 384, + 545, + 517 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 304, + 384, + 545, + 517 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 384, + 545, + 517 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 304, + 384, + 545, + 517 + ], + "type": "text", + "content": ". For all methods, we choose the best checkpoint on the validation set and report its performance on the test set. Following NCR, we also report two strong baselines based on SGR, i.e., SGR-C and SGR*. In short, SGR-C uses only clean data for training, and SGR* employs a pre-training process while training without hard negatives to improve robustness. To evaluate our method on real-world noisy data, we conduct experiments on the CC152K, which is harvested from the Internet and contains about " + }, + { + "bbox": [ + 304, + 384, + 545, + 517 + ], + "type": "inline_equation", + "content": "3\\% \\sim 20\\%" + }, + { + "bbox": [ + 304, + 384, + 545, + 517 + ], + "type": "text", + "content": " noisy correspondence [31]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 531, + 545, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 531, + 545, + 662 + ], + "spans": [ + { + "bbox": [ + 304, + 531, + 545, + 662 + ], + "type": "text", + "content": "Experiments on CC152K. Tab. 1 shows the results on CC152K with real noisy correspondence. On CC152K, our MSCN achieves the state-of-the-art performance in terms of all metrics. Specifically, the sum score of MSCN is " + }, + { + "bbox": [ + 304, + 531, + 545, + 662 + ], + "type": "inline_equation", + "content": "4.9\\%" + }, + { + "bbox": [ + 304, + 531, + 545, + 662 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 531, + 545, + 662 + ], + "type": "inline_equation", + "content": "6.2\\%" + }, + { + "bbox": [ + 304, + 531, + 545, + 662 + ], + "type": "text", + "content": " higher than the best baseline in text and image retrieval, respectively. In addition, we also present a more strict comparison (denoted by " + }, + { + "bbox": [ + 304, + 531, + 545, + 662 + ], + "type": "inline_equation", + "content": "\\mathrm{MSCN^{*}}" + }, + { + "bbox": [ + 304, + 531, + 545, + 662 + ], + "type": "text", + "content": ") that trained our method with only 1,000 meta-data (approximately " + }, + { + "bbox": [ + 304, + 531, + 545, + 662 + ], + "type": "inline_equation", + "content": "0.67\\%" + }, + { + "bbox": [ + 304, + 531, + 545, + 662 + ], + "type": "text", + "content": " of all data). We can see that our " + }, + { + "bbox": [ + 304, + 531, + 545, + 662 + ], + "type": "inline_equation", + "content": "\\mathrm{MSCN^{*}}" + }, + { + "bbox": [ + 304, + 531, + 545, + 662 + ], + "type": "text", + "content": " also achieves a competitive performance. Specifically, all metrics except R@1 of image retrieval are superior to the baselines." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 677, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 714 + ], + "type": "text", + "content": "Experiments on Flickr30K and MS-COCO. Tab. 2 shows the results on Flickr30K and MS-COCO under a range of synthetic noise rates. Following [15], we report the" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "7522" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 71, + 538, + 384 + ], + "blocks": [ + { + "bbox": [ + 61, + 71, + 538, + 384 + ], + "lines": [ + { + "bbox": [ + 61, + 71, + 538, + 384 + ], + "spans": [ + { + "bbox": [ + 61, + 71, + 538, + 384 + ], + "type": "table", + "html": "
Flickr30KMS-COCO
Image to TextText to ImageImage to TextText to Image
NoiseMethodsR@1R@5R@10R@1R@5R@10R@1R@5R@10R@1R@5R@10
20%SCAN59.183.490.436.667.077.566.291.096.445.080.289.3
VSRN58.182.689.340.768.778.225.159.074.817.649.064.1
IMRAM63.086.091.341.471.280.568.692.897.655.785.091.0
SAF51.079.388.038.366.576.267.392.596.653.484.592.4
SGR*62.886.292.244.472.380.467.891.796.252.983.590.1
SGR-C72.890.895.456.482.188.675.495.297.960.188.594.8
NCR75.093.997.558.383.089.077.795.598.262.589.395.3
MSCN77.494.997.659.683.289.278.197.298.864.390.495.8
50%SCAN27.757.668.816.239.349.840.873.584.95.415.121.0
VSRN14.337.650.012.130.039.423.554.769.316.047.865.9
IMRAM9.126.638.22.78.412.721.360.275.922.352.864.3
SAF30.363.675.427.953.765.130.467.882.333.569.082.8
SGR*36.968.180.229.356.267.060.687.493.646.074.279.0
SGR-C69.890.394.850.177.585.271.794.197.757.086.693.7
NCR72.993.096.354.379.886.574.694.697.859.187.894.5
MSCN74.493.296.055.380.486.877.596.298.760.789.194.9
70%SCAN5.619.327.42.28.012.818.143.157.40.31.32.3
VSRN0.82.54.10.51.52.75.115.724.62.58.813.3
IMRAM1.33.13.90.31.22.87.120.033.45.315.222.0
SAF0.52.23.00.20.81.70.11.74.00.61.93.0
SGR*17.942.151.914.631.040.835.771.285.431.665.879.0
SGR-C65.089.394.748.174.581.169.893.697.556.586.093.4
NCR16.138.552.811.029.541.435.469.583.431.566.481.1
MSCN69.089.393.049.273.179.074.494.997.758.887.293.7
", + "image_path": "cdccbdd4ebcf945e4b6095c80935d406375adcca92ca7af188b824a7ec4c9011.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 426, + 289, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 289, + 606 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 289, + 606 + ], + "type": "text", + "content": "results by averaging over 5 folds of 1K test images. From Tab. 2, we can draw the observation that noisy correspondence remarkably affects the performance of cross-modal retrieval methods. The retrieval accuracy will decrease fast with the noise rate increasing. Our MSCN is superior to the baselines in almost all metrics under different noise rates. In the low noise cases, i.e., " + }, + { + "bbox": [ + 46, + 426, + 289, + 606 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 46, + 426, + 289, + 606 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 426, + 289, + 606 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 46, + 426, + 289, + 606 + ], + "type": "text", + "content": ", our MSCN improves R@1 (sum of image and text) by " + }, + { + "bbox": [ + 46, + 426, + 289, + 606 + ], + "type": "inline_equation", + "content": "3.7\\%" + }, + { + "bbox": [ + 46, + 426, + 289, + 606 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 426, + 289, + 606 + ], + "type": "inline_equation", + "content": "2.2\\%" + }, + { + "bbox": [ + 46, + 426, + 289, + 606 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 426, + 289, + 606 + ], + "type": "inline_equation", + "content": "2.5\\%" + }, + { + "bbox": [ + 46, + 426, + 289, + 606 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 426, + 289, + 606 + ], + "type": "inline_equation", + "content": "4.5\\%" + }, + { + "bbox": [ + 46, + 426, + 289, + 606 + ], + "type": "text", + "content": " in the four valuations compared with NCR. In the high noise case, i.e., " + }, + { + "bbox": [ + 46, + 426, + 289, + 606 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 46, + 426, + 289, + 606 + ], + "type": "text", + "content": ", all baselines can easily overfit to noisy correspondence and results in poor performance. Our MSCN improves R@1 by " + }, + { + "bbox": [ + 46, + 426, + 289, + 606 + ], + "type": "inline_equation", + "content": "52.9\\%" + }, + { + "bbox": [ + 46, + 426, + 289, + 606 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 426, + 289, + 606 + ], + "type": "inline_equation", + "content": "38.2\\%" + }, + { + "bbox": [ + 46, + 426, + 289, + 606 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 426, + 289, + 606 + ], + "type": "inline_equation", + "content": "39.4\\%" + }, + { + "bbox": [ + 46, + 426, + 289, + 606 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 426, + 289, + 606 + ], + "type": "inline_equation", + "content": "27.3\\%" + }, + { + "bbox": [ + 46, + 426, + 289, + 606 + ], + "type": "text", + "content": " in these four valuations compared with NCR. Even compared with the SGR-C trained only on clean data, our MSCN also achieves competitive performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 611, + 142, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 611, + 142, + 624 + ], + "spans": [ + { + "bbox": [ + 47, + 611, + 142, + 624 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": "To evaluate the performance of the proposed components (i.e., self-adaptive margin and data purification), we conduct the ablation study on the Flickr30K with " + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": " noisy correspondence. Note that for MSCN without " + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{train}^{\\prime}" + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": ", we only use a single model to perform the experiment. As shown in Tab. 3, we can see that the performance without " + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\hat{\\gamma}" + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{train}^{\\prime}" + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": " are worse than the complete MSCN, which" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 426, + 546, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 426, + 546, + 462 + ], + "spans": [ + { + "bbox": [ + 305, + 426, + 546, + 462 + ], + "type": "text", + "content": "indicates that all components are important to achieve advantageous results. We explain more on the data selection strategy in the supplementary material." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 312, + 470, + 545, + 603 + ], + "blocks": [ + { + "bbox": [ + 46, + 393, + 546, + 417 + ], + "lines": [ + { + "bbox": [ + 46, + 393, + 546, + 417 + ], + "spans": [ + { + "bbox": [ + 46, + 393, + 546, + 417 + ], + "type": "text", + "content": "Table 2. Image-Text Retrieval Performance under synthetic noise rates of " + }, + { + "bbox": [ + 46, + 393, + 546, + 417 + ], + "type": "inline_equation", + "content": "{20}\\% ,{50}\\%" + }, + { + "bbox": [ + 46, + 393, + 546, + 417 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 393, + 546, + 417 + ], + "type": "inline_equation", + "content": "{70}\\%" + }, + { + "bbox": [ + 46, + 393, + 546, + 417 + ], + "type": "text", + "content": " on Flickr30K and MS-COCO 1K,and the best results are highlighted in bold." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 312, + 470, + 545, + 603 + ], + "lines": [ + { + "bbox": [ + 312, + 470, + 545, + 603 + ], + "spans": [ + { + "bbox": [ + 312, + 470, + 545, + 603 + ], + "type": "table", + "html": "
MethodImage to Text
MSCNw/o γw/o D'trainR@1R@5R@10
✓77.494.997.6
✓✓75.394.597.2
✓✓75.893.896.2
✓✓✓74.191.594.7
Text to Image
✓59.683.289.2
✓✓58.383.188.9
✓✓55.874.577.4
✓✓✓53.471.272.2
", + "image_path": "b0c5cbfc58dad73fe67517f1ad932cd4ae07457f7e2e7d770c126d42c84e34ea.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 315, + 611, + 536, + 623 + ], + "lines": [ + { + "bbox": [ + 315, + 611, + 536, + 623 + ], + "spans": [ + { + "bbox": [ + 315, + 611, + 536, + 623 + ], + "type": "text", + "content": "Table 3. Ablation studies on Flickr30K with " + }, + { + "bbox": [ + 315, + 611, + 536, + 623 + ], + "type": "inline_equation", + "content": "{20}\\%" + }, + { + "bbox": [ + 315, + 611, + 536, + 623 + ], + "type": "text", + "content": " noise rate." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 635, + 443, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 635, + 443, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 635, + 443, + 647 + ], + "type": "text", + "content": "4.5. Progressive Comparison" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": "Fig. 6 plots the average of recalls on testing set of MSCN and NCR as training proceeds. We show a representative result using Flickr30K with " + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": " noisy correspondence. From the result, we can see that our MSCN achieves a more stable performance in overall training process. Although NCR" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "text", + "content": "7523" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 79, + 136, + 132 + ], + "blocks": [ + { + "bbox": [ + 56, + 79, + 136, + 132 + ], + "lines": [ + { + "bbox": [ + 56, + 79, + 136, + 132 + ], + "spans": [ + { + "bbox": [ + 56, + 79, + 136, + 132 + ], + "type": "image", + "image_path": "9e2434691c1db0c972440d39d2b10ec8317e5cb19ba4c4225af7f4650834da11.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 57, + 142, + 135, + 162 + ], + "lines": [ + { + "bbox": [ + 57, + 142, + 135, + 162 + ], + "spans": [ + { + "bbox": [ + 57, + 142, + 135, + 162 + ], + "type": "text", + "content": "the circular movement of the camera. (0.3444)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 140, + 86, + 212, + 135 + ], + "blocks": [ + { + "bbox": [ + 140, + 86, + 212, + 135 + ], + "lines": [ + { + "bbox": [ + 140, + 86, + 212, + 135 + ], + "spans": [ + { + "bbox": [ + 140, + 86, + 212, + 135 + ], + "type": "image", + "image_path": "5416ed9692c3717fe83a5e49a16f9bd801e3f58286f39183058aff68ffcab230.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 141, + 142, + 212, + 161 + ], + "lines": [ + { + "bbox": [ + 141, + 142, + 212, + 161 + ], + "spans": [ + { + "bbox": [ + 141, + 142, + 212, + 161 + ], + "type": "text", + "content": "person arrives earlier this month. (0.3485)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 223, + 85, + 291, + 137 + ], + "blocks": [ + { + "bbox": [ + 223, + 85, + 291, + 137 + ], + "lines": [ + { + "bbox": [ + 223, + 85, + 291, + 137 + ], + "spans": [ + { + "bbox": [ + 223, + 85, + 291, + 137 + ], + "type": "image", + "image_path": "dbabc25dd5f7d976e226d17998b27e8e347ed66a3a20f0514f56a9dfb3b00fa8.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 217, + 142, + 298, + 161 + ], + "lines": [ + { + "bbox": [ + 217, + 142, + 298, + 161 + ], + "spans": [ + { + "bbox": [ + 217, + 142, + 298, + 161 + ], + "type": "text", + "content": "i expected to learn more about plants. (0.3487)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 302, + 87, + 373, + 135 + ], + "blocks": [ + { + "bbox": [ + 302, + 87, + 373, + 135 + ], + "lines": [ + { + "bbox": [ + 302, + 87, + 373, + 135 + ], + "spans": [ + { + "bbox": [ + 302, + 87, + 373, + 135 + ], + "type": "image", + "image_path": "9cbcac1eea18f074e487aea5870ba5834c873e679fedf051122762eab4b98da0.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 300, + 143, + 376, + 162 + ], + "lines": [ + { + "bbox": [ + 300, + 143, + 376, + 162 + ], + "spans": [ + { + "bbox": [ + 300, + 143, + 376, + 162 + ], + "type": "text", + "content": "man,chairman to speak at show (0.3488)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 384, + 85, + 452, + 137 + ], + "blocks": [ + { + "bbox": [ + 384, + 85, + 452, + 137 + ], + "lines": [ + { + "bbox": [ + 384, + 85, + 452, + 137 + ], + "spans": [ + { + "bbox": [ + 384, + 85, + 452, + 137 + ], + "type": "image", + "image_path": "3c991ea05fe1855eff140559fa47d4adc725baab4d8030b840e375b8c6ab8925.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 382, + 142, + 454, + 161 + ], + "lines": [ + { + "bbox": [ + 382, + 142, + 454, + 161 + ], + "spans": [ + { + "bbox": [ + 382, + 142, + 454, + 161 + ], + "type": "text", + "content": "note the many parked cars (0.3511)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 465, + 87, + 532, + 135 + ], + "blocks": [ + { + "bbox": [ + 465, + 87, + 532, + 135 + ], + "lines": [ + { + "bbox": [ + 465, + 87, + 532, + 135 + ], + "spans": [ + { + "bbox": [ + 465, + 87, + 532, + 135 + ], + "type": "image", + "image_path": "2c3e2842b25d71a4a0e251b0ff41a8f591dea74df2061f4a2faa062509a58c69.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 460, + 142, + 538, + 162 + ], + "lines": [ + { + "bbox": [ + 460, + 142, + 538, + 162 + ], + "spans": [ + { + "bbox": [ + 460, + 142, + 538, + 162 + ], + "type": "text", + "content": "a man using his mobile phone. (0.3518)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 208, + 287, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 208, + 287, + 256 + ], + "spans": [ + { + "bbox": [ + 46, + 208, + 287, + 256 + ], + "type": "text", + "content": "does not need an extra meta-set to guide the training, it relies on a clean validation set to choose the best model parameters. The stableness also indicates that our method has better alleviated the interference of noisy correspondence." + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 77, + 269, + 258, + 384 + ], + "blocks": [ + { + "bbox": [ + 77, + 269, + 258, + 384 + ], + "lines": [ + { + "bbox": [ + 77, + 269, + 258, + 384 + ], + "spans": [ + { + "bbox": [ + 77, + 269, + 258, + 384 + ], + "type": "image", + "image_path": "8b2cfa94e9146426a496cfd7be0646a851b640647754760cce2af5d3db163311.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 393, + 287, + 415 + ], + "lines": [ + { + "bbox": [ + 46, + 393, + 287, + 415 + ], + "spans": [ + { + "bbox": [ + 46, + 393, + 287, + 415 + ], + "type": "text", + "content": "Figure 6. Progressive performance comparison of MSCN and NCR on testing set as training proceeds." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 426, + 181, + 439 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 426, + 181, + 439 + ], + "spans": [ + { + "bbox": [ + 47, + 426, + 181, + 439 + ], + "type": "text", + "content": "4.6. Detected Noisy Samples" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 445, + 287, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 445, + 287, + 564 + ], + "spans": [ + { + "bbox": [ + 46, + 445, + 287, + 564 + ], + "type": "text", + "content": "Fig. 5 shows some real-world noisy image-text pairs detected (with low similarity score) by MSCN in Conceptual Captions. We also present the similarity score predicted by our MSCN. For most noisy pairs, the image and text contains completely irrelevant semantic information. Moreover, our MSCN can even find the subtle mistake in the noisy image-text pair. In the last pair, the picture shows a man covering his ears but wrongly caption as using a mobile phone. Although this picture is misleading, our method successfully detects the subtle difference." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 47, + 570, + 224, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 570, + 224, + 582 + ], + "spans": [ + { + "bbox": [ + 47, + 570, + 224, + 582 + ], + "type": "text", + "content": "4.7. Visualization on Similarity Score" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 46, + 589, + 287, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 589, + 287, + 660 + ], + "spans": [ + { + "bbox": [ + 46, + 589, + 287, + 660 + ], + "type": "text", + "content": "Fig. 7 plots the similarity score distribution predicted by MSCN for clean and noisy training samples. It can be seen that almost all large similarity scores belong to clean pairs, and the noisy pairs' scores are smaller than clean samples, which implies that the trained MSCN can successfully provide corrected similarity scores." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 47, + 670, + 119, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 119, + 681 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 119, + 681 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "In this paper, we explore a meta-learning method to address the problem of learning with noisy correspond-" + } + ] + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 311, + 209, + 423, + 346 + ], + "blocks": [ + { + "bbox": [ + 184, + 183, + 408, + 194 + ], + "lines": [ + { + "bbox": [ + 184, + 183, + 408, + 194 + ], + "spans": [ + { + "bbox": [ + 184, + 183, + 408, + 194 + ], + "type": "text", + "content": "Figure 5. Real-world noisy examples detected by our MSCN." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 311, + 209, + 423, + 346 + ], + "lines": [ + { + "bbox": [ + 311, + 209, + 423, + 346 + ], + "spans": [ + { + "bbox": [ + 311, + 209, + 423, + 346 + ], + "type": "image", + "image_path": "92e2f917921919a070aab04042430702cb4149ba13147bbe56f7cc6ca78364f3.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 361, + 353, + 372, + 361 + ], + "lines": [ + { + "bbox": [ + 361, + 353, + 372, + 361 + ], + "spans": [ + { + "bbox": [ + 361, + 353, + 372, + 361 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 364, + 545, + 396 + ], + "lines": [ + { + "bbox": [ + 306, + 364, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 306, + 364, + 545, + 396 + ], + "type": "text", + "content": "Figure 7. Per-sample similarity distribution on training data of Flickr30K. (a) " + }, + { + "bbox": [ + 306, + 364, + 545, + 396 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 306, + 364, + 545, + 396 + ], + "type": "text", + "content": " noisy correspondence. (b) " + }, + { + "bbox": [ + 306, + 364, + 545, + 396 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 306, + 364, + 545, + 396 + ], + "type": "text", + "content": " noisy correspondence." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 429, + 209, + 541, + 346 + ], + "blocks": [ + { + "bbox": [ + 429, + 209, + 541, + 346 + ], + "lines": [ + { + "bbox": [ + 429, + 209, + 541, + 346 + ], + "spans": [ + { + "bbox": [ + 429, + 209, + 541, + 346 + ], + "type": "image", + "image_path": "ac5ce36a3925aec0bb9dbc6372691a5b0277f3b6287eddd6b638c5c8d8b6b5ee.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 479, + 353, + 489, + 361 + ], + "lines": [ + { + "bbox": [ + 479, + 353, + 489, + 361 + ], + "spans": [ + { + "bbox": [ + 479, + 353, + 489, + 361 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 407, + 545, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 407, + 545, + 528 + ], + "spans": [ + { + "bbox": [ + 304, + 407, + 545, + 528 + ], + "type": "text", + "content": "dence. Specifically, we propose a meta correction network (MSCN) to provide reliable similarity scores. Our MSCN is trained by a novel meta-process that views both positive and negative data as meta-knowledge to encourage MSCN to learn discrimination from them. To further mitigate noise interference, we design a data purification strategy that uses meta-data as prior knowledge to purify the noisy data efficiently. We conduct comprehensive experiments on three widely used datasets. The results validate the effectiveness of our method in both synthetic and real noises." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 306, + 538, + 408, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 538, + 408, + 551 + ], + "spans": [ + { + "bbox": [ + 306, + 538, + 408, + 551 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 304, + 557, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 545, + 713 + ], + "type": "text", + "content": "This work is supported by the National Key Research and Development Program of China (2020AAA0108800), National Nature Science Foundation of China (62192781, 62272374, 62202367, 62250009, 62137002, 61937001), Innovative Research Group of the National Natural Science Foundation of China (61721002), Innovation Research Team of Ministry of Education (IRT_17R86), Project of China Knowledge Center for Engineering Science and Technology, and Project of Chinese academy of engineering \"The Online and Offline Mixed Educational Service System for 'The Belt and Road' Training in MOOC China\". We would like to express our gratitude for the support of K. C. Wong Education Foundation." + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "7524" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "type": "text", + "content": "[1] Relja Arandjelovic and Andrew Zisserman. Objects that sound. In Proceedings of the European conference on computer vision (ECCV), pages 435-451, 2018. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 125, + 287, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 287, + 168 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 287, + 168 + ], + "type": "text", + "content": "[2] Eric Arazo, Diego Ortego, Paul Albert, Noel O'Connor, and Kevin McGuinness. Unsupervised label noise modeling and loss correction. In International conference on machine learning, pages 312-321. PMLR, 2019. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 170, + 287, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 170, + 287, + 223 + ], + "spans": [ + { + "bbox": [ + 53, + 170, + 287, + 223 + ], + "type": "text", + "content": "[3] Devansh Arpit, Stanislaw K Jastrzebski, Nicolas Ballas, David Krueger, Emmanuel Bengio, Maxinder S Kanwal, Tegan Maharaj, Asja Fischer, Aaron C Courville, Yoshua Bengio, et al. A closer look at memorization in deep networks. In ICML, 2017. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 224, + 287, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 224, + 287, + 280 + ], + "spans": [ + { + "bbox": [ + 53, + 224, + 287, + 280 + ], + "type": "text", + "content": "[4] Hui Chen, Guiguang Ding, Xudong Liu, Zijia Lin, Ji Liu, and Jungong Han. Imram: Iterative matching with recurrent attention memory for cross-modal image-text retrieval. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12655-12663, 2020. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 281, + 287, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 281, + 287, + 323 + ], + "spans": [ + { + "bbox": [ + 53, + 281, + 287, + 323 + ], + "type": "text", + "content": "[5] Arthur P Dempster, Nan M Laird, and Donald B Rubin. Maximum likelihood from incomplete data via the em algorithm. Journal of the Royal Statistical Society: Series B (Methodological), 39(1):1-22, 1977. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 324, + 287, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 324, + 287, + 367 + ], + "spans": [ + { + "bbox": [ + 53, + 324, + 287, + 367 + ], + "type": "text", + "content": "[6] Cheng Deng, Zhaojia Chen, Xianglong Liu, Xinbo Gao, and Dacheng Tao. Triplet-based deep hashing network for cross-modal retrieval. IEEE Transactions on Image Processing, 27(8):3893-3903, 2018. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 369, + 287, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 369, + 287, + 413 + ], + "spans": [ + { + "bbox": [ + 53, + 369, + 287, + 413 + ], + "type": "text", + "content": "[7] Haiwen Diao, Ying Zhang, Lin Ma, and Hutchuan Lu. Similarity reasoning and filtration for image-text matching. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 1218-1226, 2021. 2, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 414, + 287, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 414, + 287, + 446 + ], + "spans": [ + { + "bbox": [ + 53, + 414, + 287, + 446 + ], + "type": "text", + "content": "[8] Fartash Faghri, David J Fleet, Jamie Ryan Kiros, and Sanja Fidler. Vse++: Improving visual-semantic embeddings with hard negatives. arXiv preprint arXiv:1707.05612, 2017. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 447, + 287, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 447, + 287, + 490 + ], + "spans": [ + { + "bbox": [ + 53, + 447, + 287, + 490 + ], + "type": "text", + "content": "[9] Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In International conference on machine learning, pages 1126–1135. PMLR, 2017. 2, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 491, + 287, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 491, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 48, + 491, + 287, + 545 + ], + "type": "text", + "content": "[10] Luca Franceschi, Paolo Frasconi, Saverio Salzo, Riccardo Grazzi, and Massimiliano Pontil. Bilevel programming for hyperparameter optimization and meta-learning. In International Conference on Machine Learning, pages 1568-1577. PMLR, 2018. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 547, + 287, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 547, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 547, + 287, + 590 + ], + "type": "text", + "content": "[11] Dalu Guo, Chang Xu, and Dacheng Tao. Image-question-answer synergistic network for visual dialog. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10434-10443, 2019. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 591, + 287, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 591, + 287, + 646 + ], + "spans": [ + { + "bbox": [ + 48, + 591, + 287, + 646 + ], + "type": "text", + "content": "[12] Bo Han, Quanming Yao, Xingrui Yu, Gang Niu, Miao Xu, Weihua Hu, Ivor Tsang, and Masashi Sugiyama. Co-teaching: Robust training of deep neural networks with extremely noisy labels. Advances in neural information processing systems, 31, 2018. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "text", + "content": "[13] Haochen Han, Qinghua Zheng, Minnan Luo, Kaiyao Miao, Feng Tian, and Yan Chen. Noise-tolerant learning for audiovisual action recognition. arXiv preprint arXiv:2205.07611, 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[14] Peng Hu, Xi Peng, Hongyuan Zhu, Liangli Zhen, and Jie Lin. Learning cross-modal retrieval with noisy labels. In" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5403-5413, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 96, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 150 + ], + "type": "text", + "content": "[15] Zhenyu Huang, Guocheng Niu, Xiao Liu, Wenbiao Ding, Xinyan Xiao, Hua Wu, and Xi Peng. Learning with noisy correspondence for cross-modal matching. Advances in Neural Information Processing Systems, 34:29406-29419, 2021. 1, 2, 5, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 152, + 545, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 152, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 308, + 152, + 545, + 207 + ], + "type": "text", + "content": "[16] Haoxuanye Ji, Le Wang, Sanping Zhou, Wei Tang, Nanning Zheng, and Gang Hua. Meta pairwise relationship distillation for unsupervised person re-identification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3661–3670, 2021. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 208, + 545, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 208, + 545, + 229 + ], + "spans": [ + { + "bbox": [ + 308, + 208, + 545, + 229 + ], + "type": "text", + "content": "[17] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR (Poster), 2015. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 231, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 231, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 308, + 231, + 545, + 274 + ], + "type": "text", + "content": "[18] Kuang-Huei Lee, Xi Chen, Gang Hua, Houdong Hu, and Xiaodong He. Stacked cross attention for image-text matching. In Proceedings of the European conference on computer vision (ECCV), pages 201-216, 2018. 2, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 276, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 276, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 308, + 276, + 545, + 319 + ], + "type": "text", + "content": "[19] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy Hospedales. Learning to generalize: Meta-learning for domain generalization. In Proceedings of the AAAI conference on artificial intelligence, volume 32, 2018. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 321, + 545, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 321, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 308, + 321, + 545, + 364 + ], + "type": "text", + "content": "[20] Junnan Li, Richard Socher, and Steven CH Hoi. Dividemix: Learning with noisy labels as semi-supervised learning. In International Conference on Learning Representations, 2019. 1, 2, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 365, + 545, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 365, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 308, + 365, + 545, + 409 + ], + "type": "text", + "content": "[21] Junnan Li, Yongkang Wong, Qi Zhao, and Mohan S Kankanhalli. Learning to learn from noisy labeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5051-5059, 2019. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 411, + 545, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 411, + 545, + 454 + ], + "spans": [ + { + "bbox": [ + 308, + 411, + 545, + 454 + ], + "type": "text", + "content": "[22] Kunpeng Li, Yulun Zhang, Kai Li, Yuanyuan Li, and Yun Fu. Visual semantic reasoning for image-text matching. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4654-4662, 2019. 2, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 456, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 456, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 308, + 456, + 545, + 487 + ], + "type": "text", + "content": "[23] Aristidis Likas, Nikos Vlassis, and Jakob J Verbeek. The global k-means clustering algorithm. Pattern recognition, 36(2):451-461, 2003. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 489, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 489, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 308, + 489, + 545, + 544 + ], + "type": "text", + "content": "[24] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 545, + 545, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 545, + 545, + 600 + ], + "spans": [ + { + "bbox": [ + 308, + 545, + 545, + 600 + ], + "type": "text", + "content": "[25] Chunxiao Liu, Zhendong Mao, Tianzhu Zhang, Hongtao Xie, Bin Wang, and Yongdong Zhang. Graph structured network for image-text matching. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10921-10930, 2020. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 601, + 545, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 601, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 308, + 601, + 545, + 634 + ], + "type": "text", + "content": "[26] Lin Ma, Zhengdong Lu, and Hang Li. Learning to answer questions from image using convolutional neural network. In Thirtieth AAAI Conference on Artificial Intelligence, 2016. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 635, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 635, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 308, + 635, + 545, + 678 + ], + "type": "text", + "content": "[27] Zhanyu Ma and Arne Leijon. Bayesian estimation of beta mixture models with variational inference. IEEE Transactions on Pattern Analysis and Machine Intelligence, 33(11):2160-2173, 2011. 5" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "type": "text", + "content": "[28] Nikhil Mishra, Mostafa Rohaninejad, Xi Chen, and Pieter Abbeel. A simple neural attentive meta-learner. In International Conference on Learning Representations, 2018. 2" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7525" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "type": "text", + "content": "[29] Yang Qin, Dezhong Peng, Xi Peng, Xu Wang, and Peng Hu. Deep evidential learning with noisy correspondence for cross-modal retrieval. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4948-4956, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 127, + 287, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 127, + 287, + 171 + ], + "spans": [ + { + "bbox": [ + 48, + 127, + 287, + 171 + ], + "type": "text", + "content": "[30] Mengye Ren, Wenyuan Zeng, Bin Yang, and Raquel Urtasun. Learning to reweight examples for robust deep learning. In International conference on machine learning, pages 4334-4343. PMLR, 2018. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 172, + 288, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 172, + 288, + 237 + ], + "spans": [ + { + "bbox": [ + 48, + 172, + 288, + 237 + ], + "type": "text", + "content": "[31] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2556-2565, 2018. 1, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 238, + 288, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 238, + 288, + 282 + ], + "spans": [ + { + "bbox": [ + 48, + 238, + 288, + 282 + ], + "type": "text", + "content": "[32] Jun Shu, Qi Xie, Lixuan Yi, Qian Zhao, Sanping Zhou, Zongben Xu, and Deyu Meng. Meta-weight-net: Learning an explicit mapping for sample weighting. Advances in neural information processing systems, 32, 2019. 2, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 283, + 287, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 283, + 287, + 336 + ], + "spans": [ + { + "bbox": [ + 48, + 283, + 287, + 336 + ], + "type": "text", + "content": "[33] Didac Surís, Amanda Duarte, Amaia Salvador, Jordi Torres, and Xavier Giró-i Nieto. Cross-modal embeddings for video and audio retrieval. In Proceedings of the European Conference on Computer Vision (ECCV) Workshops, pages 0–0, 2018. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 338, + 287, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 338, + 287, + 370 + ], + "spans": [ + { + "bbox": [ + 48, + 338, + 287, + 370 + ], + "type": "text", + "content": "[34] Jun Tang, Ke Wang, and Ling Shao. Supervised matrix factorization hashing for cross-modal retrieval. IEEE Transactions on Image Processing, 25(7):3157-3166, 2016. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 371, + 287, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 371, + 287, + 415 + ], + "spans": [ + { + "bbox": [ + 48, + 371, + 287, + 415 + ], + "type": "text", + "content": "[35] Bokun Wang, Yang Yang, Xing Xu, Alan Hanjalic, and Heng Tao Shen. Adversarial cross-modal retrieval. In Proceedings of the 25th ACM international conference on Multimedia, pages 154-162, 2017. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 415, + 287, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 287, + 459 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 287, + 459 + ], + "type": "text", + "content": "[36] Liwei Wang, Yin Li, and Svetlana Lazebnik. Learning deep structure-preserving image-text embeddings. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5005-5013, 2016. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 459, + 287, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 459, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 48, + 459, + 287, + 502 + ], + "type": "text", + "content": "[37] Yunchao Wei, Yao Zhao, Canyi Lu, Shikui Wei, Luoqi Liu, Zhenfeng Zhu, and Shuicheng Yan. Cross-modal retrieval with cnn visual features: A new baseline. IEEE transactions on cybernetics, 47(2):449-460, 2016. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 503, + 287, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 503, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 48, + 503, + 287, + 536 + ], + "type": "text", + "content": "[38] Xing Xu, Li He, Huimin Lu, Lianli Gao, and Yanli Ji. Deep adversarial metric learning for cross-modal retrieval. *World Wide Web*, 22(2):657-672, 2019. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 536, + 287, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 287, + 591 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 287, + 591 + ], + "type": "text", + "content": "[39] Mouxing Yang, Zhenyu Huang, Peng Hu, Taihao Li, Jiancheng Lv, and Xi Peng. Learning with twin noisy labels for visible-infrared person re-identification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14308-14317, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 592, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 287, + 635 + ], + "type": "text", + "content": "[40] Mouxing Yang, Yunfan Li, Peng Hu, Jinfeng Bai, Jiancheng Lv, and Xi Peng. Robust multi-view clustering with incomplete information. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(1):1055-1069, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 636, + 288, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 288, + 691 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 288, + 691 + ], + "type": "text", + "content": "[41] Mouxing Yang, Yunfan Li, Zhenyu Huang, Zitao Liu, Peng Hu, and Xi Peng. Partially view-aligned representation learning with noise-robust contrastive loss. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1134–1143, 2021. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "type": "text", + "content": "[42] Quanming Yao, Hansi Yang, Bo Han, Gang Niu, and James Tin-Yau Kwok. Searching to exploit memorization effect in" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 285 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "learning with noisy labels. In International Conference on Machine Learning, pages 10789-10798. PMLR, 2020. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 95, + 547, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 95, + 547, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 95, + 547, + 150 + ], + "type": "text", + "content": "[43] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics, 2:67-78, 2014. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 151, + 547, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 151, + 547, + 196 + ], + "spans": [ + { + "bbox": [ + 307, + 151, + 547, + 196 + ], + "type": "text", + "content": "[44] Qi Zhang, Zhen Lei, Zhaoxiang Zhang, and Stan Z Li. Context-aware attention network for image-text retrieval. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3536-3545, 2020. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 196, + 547, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 196, + 547, + 240 + ], + "spans": [ + { + "bbox": [ + 307, + 196, + 547, + 240 + ], + "type": "text", + "content": "[45] Liangli Zhen, Peng Hu, Xu Wang, and Dezhong Peng. Deep supervised cross-modal retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10394-10403, 2019. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 241, + 547, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 241, + 547, + 285 + ], + "spans": [ + { + "bbox": [ + 308, + 241, + 547, + 285 + ], + "type": "text", + "content": "[46] Guoqing Zheng, Ahmed Hassan Awadallah, and Susan Dumais. Meta label correction for noisy label learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 11053-11061, 2021. 2, 4" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "7526" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/NoisyQuant_ Noisy Bias-Enhanced Post-Training Activation Quantization for Vision Transformers/68808ca4-cf24-4585-b35b-3e2acef953ea_content_list.json b/2023/NoisyQuant_ Noisy Bias-Enhanced Post-Training Activation Quantization for Vision Transformers/68808ca4-cf24-4585-b35b-3e2acef953ea_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..71e4f5575edf40620704c9b80d6eac3c581f27b6 --- /dev/null +++ b/2023/NoisyQuant_ Noisy Bias-Enhanced Post-Training Activation Quantization for Vision Transformers/68808ca4-cf24-4585-b35b-3e2acef953ea_content_list.json @@ -0,0 +1,1553 @@ +[ + { + "type": "text", + "text": "NoisyQuant: Noisy Bias-Enhanced Post-Training Activation Quantization for Vision Transformers", + "text_level": 1, + "bbox": [ + 96, + 128, + 875, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yijiang Liu\\*1, Huanrui Yang\\*2, Zhen Dong?, Kurt Keutzer?, Li Du1, Shanghang Zhang3Nanjing University, 2University of California, Berkeley", + "bbox": [ + 125, + 200, + 846, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ National Key Laboratory for Multimedia Information Processing, School of Computer Science, Peking University liuyijiang@smail.nju.edu.cn, {huanrui, zhendong, keutzer}@berkeley.edu ldu@nju.edu.cn, shanghang@pku.edu.cn", + "bbox": [ + 106, + 239, + 861, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 340, + 313, + 358 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The complicated architecture and high training cost of vision transformers urge the exploration of post-training quantization. However, the heavy-tailed distribution of vision transformer activations hinders the effectiveness of previous post-training quantization methods, even with advanced quantizer designs. Instead of tuning the quantizer to better fit the complicated activation distribution, this paper proposes NoisyQuant, a quantizer-agnostic enhancement for the post-training activation quantization performance of vision transformers. We make a surprising theoretical discovery that for a given quantizer, adding a fixed Uniform noisy bias to the values being quantized can significantly reduce the quantization error under provable conditions. Building on the theoretical insight, NoisyQuant achieves the first success on actively altering the heavy-tailed activation distribution with additive noisy bias to fit a given quantizer. Extensive experiments show NoisyQuant largely improves the post-training quantization performance of vision transformer with minimal computation overhead. For instance, on linear uniform 6-bit activation quantization, NoisyQuant improves SOTA top-1 accuracy on ImageNet by up to $1.7\\%$ , $1.1\\%$ and $0.5\\%$ for ViT, DeiT, and Swin Transformer respectively, achieving on-par or even higher performance than previous nonlinear, mixed-precision quantization.", + "bbox": [ + 75, + 373, + 473, + 736 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 763, + 209, + 780 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Inspired by the success of Self Attention (SA)-based transformer models in Natural Language Processing (NLP) tasks [31], recent researches make significant progress in applying transformer models to the field of computer vision [3, 10, 22, 30]. In the meantime, the typical design of", + "bbox": [ + 75, + 789, + 470, + 867 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "transformer models induces large model sizes, high computational consumption, and long training time. For instance, the widely used DeiT-Base model [30] contains 86M parameters, logs 18G floating-point operations for a single input, and requires 300 epochs of training on the ImageNet dataset. This leads to significant difficulties in hardware deployment. In facing such difficulty, a number of compression and acceleration methods are applied to vision transformer models, including pruning [6, 39], quantization [24, 42], and neural architecture search [5], etc.", + "bbox": [ + 500, + 328, + 895, + 479 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Among these methods, quantization appears as one of the most effective and widely applied ways [11]. Quantization process uses a predefined \"quantizer\" function to convert the continuous representation of weights and activations into a small number of discrete symbols, therefore enabling low-precision representations for straightforward memory savings. For DNN models, the approximation error made by the quantizer inevitably leads to performance drop. A series of work focuses on Quantization-Aware Training (QAT) that finetunes the quantized model at low precision [8, 9, 25, 37, 47]. However, given the high training cost and the complicated computation graph of vision transformer models, retraining the model at low precision could be costly and unstable [42]. Alternatively, Post-Training Quantization (PTQ) is preferable for vision transformers as it eliminates the need for re-training or finetuning the quantized model, instead only adjusts the design of the quantizer based on the full-precision pretrained model and a small set of sampled calibration data [1, 2, 23, 32, 36]. For example, linear quantizers [7, 29, 36, 38] reduce quantization error by shifting, clipping, and scaling the values to be quantized. Nonlinear quantizers [13, 42, 44] further adjust the width and location of each quantization bin to better fit the distribution.", + "bbox": [ + 496, + 486, + 895, + 834 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Unfortunately, though progresses are made in designing better PTQ quantizers, it still appears to be significantly challenging to quantize vision transformer models, especially for activation quantization. Transformer layers produce up", + "bbox": [ + 498, + 839, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution. \nCorresponding Author.", + "bbox": [ + 99, + 875, + 240, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "20321", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/74c4599864a12bc52e56972b056b846857fbe93ef3957a5256dc9eb750e9bdaa.jpg", + "image_caption": [ + "Figure 1. Overview of the NoisyQuant pipeline (orange box) with comparison to EasyQuant [36] (gray box). Histograms of input activations, quantized activations, and layer outputs are illustrated in the pipeline. Starting from the input activation $X$ following a GELU function, where the green histogram is plotted in linear-scale and the gray in log-scale, NoisyQuant adds a fixed NoisyBias $N$ sampled from a Uniform distribution onto $X$ . Adding the noise before quantization flattens the peaks in the activation distribution to make it more friendly to quantization, as visualized in the linear/log-scaled histogram of $X + N$ , and theoretically proved in Sec. 3.2. The noise is removed from the result of the fixed-point activation-weight multiplication with a denoising bias term, which we formulate in Sec. 3.3. The rightmost sub-figures illustrate the layer output distribution of EasyQuant and NoisyQuant. Compared to the output of the floating-point model (yellow), NoisyQuant output follows the distribution closely and achieves up to $60\\%$ output error reduction on some transformer layers, as shown in Sec. 4. The resulting PTQ performance improvement of the entire model is provided in Sec. 5." + ], + "image_footnote": [], + "bbox": [ + 99, + 88, + 869, + 296 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "to millions of activation values with sophisticated distributions. For instance, outputs of the GELU function [15] are asymmetrically distributed, with spikes in the histogram at some values and a long tail in a large range (see top-left of Fig. 1). Some linear projection layers also lead to significantly large activation values that are sparsely distributed in a very long tail [23]. Consequently, low-precision PTQ on vision transformer suffers from performance degradation, even utilizing non-linear or mixed-precision quantizers at the cost of additional data communication and computation overheads [23,43]. No linear uniform PTQ method achieves good performance on vision transformer models.", + "bbox": [ + 75, + 450, + 472, + 632 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This paper provides a brand new perspective in dealing with the complicated vision transformer activation distribution. Instead of adding more tricks in the quantizer design to fit the activation distribution, this work explores the potential of actively and cost-effectively altering the distribution being quantized, making it more friendly to a given quantizer. The pipeline of our proposed method is illustrated in Fig. 1. Specifically, we make a surprising discovery that for any quantizer, the quantization error can be significantly reduced by adding a fixed noisy bias sampled from a Uniform distribution to the activation before quantization. We theoretically prove the condition when the quantization error reduction can be achieved. On this basis, we propose NoisyQuant, a plug-and-play, quantizer-agnostic enhancement on the posttraining activation quantization of vision transformer models. For each layer, we sample a Noisy Bias based on the input activation distribution following our theoretical analysis, and", + "bbox": [ + 75, + 643, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "compute the corresponding denoising bias to retain the correct output of the layer. At inference time, the Noisy Bias is added to the input activation before quantization, and the denoising bias is removed from the layer output. This process significantly reduces the quantization error with minimal computation overhead. NoisyQuant leads to significant improvement in the PTQ performance of state-of-the-art vision transformers. Applying NoisyQuant on top of a uniform linear quantization achieves on-par performance to SOTA mixed-precision, nonlinear PTQ methods [23, 42]. Adding NoisyQuant on top of these nonlinear quantizers achieves further performance gain.", + "bbox": [ + 496, + 450, + 893, + 631 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To the best of our knowledge, this paper makes the following novel contributions:", + "bbox": [ + 498, + 631, + 893, + 661 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Theoretically shows the possibility and proves feasible conditions for reducing the quantization error of heavily-tailed distributions with a fixed additive noisy bias;", + "- Proposes NoisyQuant, a quantizer-agnostic enhancement for PTQ performance on activation quantization. NoisyQuant achieves the first success in actively refining the distribution being quantized to reduce quantization error following the theoretical results on additive noisy bias, with minimal computation overhead;", + "- Demonstrates consistent performance improvement by applying NoisyQuant on top of existing PTQ quantizers. For 6-bit PTQ, NoisyQuant improves the ImageNet top-1 accuracy of uniform linear quantized vision transformer models by up to $1.7\\%$ , and improves SOTA" + ], + "bbox": [ + 517, + 670, + 893, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "20322", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "nonlinear quantizer PTQ4ViT [43] by up to $0.7\\%$ .", + "bbox": [ + 109, + 90, + 439, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 76, + 121, + 215, + 137 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1. Transformer models in computer vision", + "text_level": 1, + "bbox": [ + 76, + 146, + 416, + 162 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Since Self-Attention (SA) [31]-based transformer models have shown impressive performance in Natural Language Processing tasks, researchers transfer the idea into computer vision. Pioneering work on Vision Transformer [10] for the first time builds a totally SA-based model in the field of computer vision and shows its effectiveness. Since then, a variety of transformer-based vision models have emerged with boosted performance, claiming state-of-the-art in multiple computer vision tasks. Notable advances include additional data augmentation and distillation [30], incorporating multi-stage architecture [22, 33], and explore novel attention design [14,22,41]. Transformer models also advance to tasks beyond classification, such as object detection [3], semantic segmentation [46], and image generation [17]. Furthermore, the development of vision transformer brings unification of model architecture and representation across image and text, enabling stronger vision-language connection via image-text-pair contrastive learning [27].", + "bbox": [ + 76, + 170, + 472, + 443 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The rapid development of vision transformer models motivates the exploration of quantization, as quantization brings straightforward memory and computation savings agnostic to model architectures. Compared to CNN models, the use of advanced activation functions like GELU [15] and more complicated computation graphs like the attention mechanism makes vision transformers more sensitive to quantization [43]. This work aims to resolve the difficulty of posttraining vision transformer activation quantization caused by the heavy-tailed activation distribution, which we believe will facilitate the deployment of state-of-the-art transformer models into real-world computer vision applications.", + "bbox": [ + 75, + 443, + 472, + 625 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Post-training quantization methods", + "text_level": 1, + "bbox": [ + 76, + 635, + 385, + 651 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Post-training quantization (PTQ) is a preferable quantization method in scenarios with limited training data or limited computational resources for the expensive quantization-aware fine-tuning process. Previous methods have dedicatedly studied PTQ on CNNs. For instance, EasyQuant [36] presents a quick searching mechanism to determine the proper clipping range for quantization. ZeroQ [2] proposes a distillation mechanism to generate proxy input images, which can leverage inherent statistics of batch normalization layers to conduct PTQ. SQuant [12] calibrates the model for lower quantization error on the basis of the sensitivity obtained by the Hessian spectrum. On vision transformers, Liu et al. [23] presents a linear PTQ method that uses Pearson correlation coefficients and ranking loss to determine scaling factors. PTQ4ViT [43] introduces a nonlinear Twin Uniform Quantization based on vision transformers' acti", + "bbox": [ + 75, + 659, + 472, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "vation distributions, which set different scaling factors for 1) positive and negative activations of GeLU, and 2) small and large values of Softmax, which can reduce the activation quantization error to some extent with the cost of additional computation overhead. [19, 34] reparameterize the LN layer to suppress outliers by scaling down activation values.", + "bbox": [ + 496, + 90, + 893, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unlike previous PTQ methods that analyze the activation distribution and fit the quantizer accordingly, our method takes a novel perspective of actively modifying the distribution being quantized with a sampled noisy bias. We show our proposed method can bring quantization error reduction and performance enhancement for all the PTQ quantizers for vision transformer activation quantization.", + "bbox": [ + 496, + 181, + 893, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 300, + 591, + 316 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we provide a preliminary introduction to the notation and typical formulation of DNN quantization in Sec. 3.1, theoretically analyze the potential of reducing quantization error with pre-sampled additive noisy bias in Sec. 3.2, and formulate NoisyQuant, a noisy bias-enhanced post-training activation quantization method with reduced quantization error in Sec. 3.3.", + "bbox": [ + 496, + 325, + 893, + 431 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary", + "text_level": 1, + "bbox": [ + 500, + 440, + 629, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To start with, we recap the post-training quantization process on DNN models. Given the dominance of fully-connected (FC) linear projection layers in transformer models, here we focus our discussion on the FC layer.", + "bbox": [ + 496, + 463, + 893, + 523 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The computation performed in a FC layer with weight $W \\in \\mathbb{R}^{k \\times m}$ , input activation $X \\in \\mathbb{R}^{m \\times n}$ , and bias $B \\in \\mathbb{R}^{k \\times 1}$ can be formulated as", + "bbox": [ + 498, + 523, + 893, + 568 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nf (X) = W X + B. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 627, + 580, + 890, + 595 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The main computation burden in Eq. (1) lies in the matrix multiplication of $WX$ , which requires a total of $k \\times m \\times n$ multiply-accumulate computations (MACs). Post-training quantization aims to reduce the computation cost of matrix multiplication by converting $X$ and $W$ to fixed-point quantized values, therefore replacing floating-point MACs with much cheaper fixed-point operations [16, 40]. The quantized FC layer can therefore be represented as", + "bbox": [ + 496, + 606, + 893, + 727 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nf _ {q} (X) = q _ {W} (W) q _ {A} (X) + B, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 738, + 890, + 753 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $q_{W}(\\cdot)$ and $q_{A}(\\cdot)$ denotes the quantizer function of weight and activation respectively. Previous research has observed that the heave-tailed activation distribution of $X$ in transformer is causing significant quantization error between $X$ and $q_{A}(X)$ at low precision, leading to significant performance degradation. Previous PTQ methods modify the design of quantizer $q_{A}(\\cdot)$ to mitigate the quantization error. In this paper, we propose an alternative approach, where we modify the distribution of activation $X$ with a", + "bbox": [ + 496, + 763, + 893, + 900 + ], + "page_idx": 2 + }, + { + "type": "footer", + "text": "20323", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "pre-sampled noisy bias before quantization, and prove it to be a plug-and-play enhancement on any quantizer design.", + "bbox": [ + 75, + 90, + 468, + 122 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Theoretical analysis on quantization error", + "text_level": 1, + "bbox": [ + 76, + 130, + 434, + 147 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As previous research mainly blames the heavy-tailed distribution of transformer activation as the main cause for large quantization errors [23, 42], in this section we analyze how the quantization error change as we actively alter the activation distribution being quantized. A straightforward way to alter the input activation distribution is to add a fixed \"Noisy Bias\" sampled from a Uniform random distribution. Denoting the input activation as $X$ and the Noisy Bias as $N$ , where $N$ and $X$ have the same dimensions, here we formulate the quantization error difference between $X$ and $X + N$ for a quantizer $Q$ in Eq. (3)", + "bbox": [ + 75, + 152, + 472, + 320 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} D (X, N) = Q E (X + N) - Q E (X) = \\\\ \\left| \\left| \\left(Q (X + N) - X - N\\right) \\right| \\right| _ {2} ^ {2} - \\left| \\left| \\left(Q (X) - X\\right) \\right| \\right| _ {2} ^ {2}. \\tag {3} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 329, + 468, + 366 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Theorem 1 states the condition where $D(x, N) \\leq 0$ holds for each histogram snapshot $x$ of activation $X$ .", + "bbox": [ + 75, + 376, + 468, + 406 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Theorem 1. Consider a given quantizer $Q$ with quantization bin width $2b$ . For each histogram snapshot of $X$ where all elements $X_{i}$ have the same distance $x$ away from the center of the quantization bin, and for a Noisy Bias $N$ sampled from $N \\sim \\mathcal{U}(-n,n)$ where $x \\leq n \\leq 2b - x$ , we have", + "bbox": [ + 73, + 415, + 468, + 491 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nD (x, N) \\leq 0 \\text {i f} 0 \\leq x \\leq n \\left(1 - \\sqrt {\\frac {n}{3 b}}\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 127, + 501, + 468, + 536 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Proof. Without loss of generality, we consider the quantization bin of quantizer $Q$ which $x$ falls into to be centered at the 0 point, therefore $0 \\leq x \\leq b$ .", + "bbox": [ + 75, + 542, + 468, + 588 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this way, $Q(X_{i}) = b$ for all elements in the histogram snapshot, leading to a quantization error", + "bbox": [ + 75, + 588, + 468, + 619 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nQ E (X) = (b - x) ^ {2}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 628, + 468, + 646 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Consider adding $N \\sim \\mathcal{U}(-n, n)$ . $X + N$ will follow $\\mathcal{U}(x - n, x + n)$ . In the case of $x \\leq n \\leq 2b - x$ , $Q(X_i + N_i) = b$ if $N_i \\in [-x, n]$ , and $Q(X_i + N_i) = -b$ if $N_i \\in [-n, -x)$ for all elements in $X + N$ . Though we only sample a single instance of $N$ from the Uniform distribution, given the large number of elements in the histogram snapshot, the empirical quantization error $QE(X + N)$ can be estimated with an expectation over all $N \\sim \\mathcal{U}(-n, n)$ following the Weak Law of Large Numbers, as", + "bbox": [ + 75, + 655, + 472, + 792 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {E} _ {N} [ Q E (X + N) ] = \\frac {1}{2 n} \\int_ {x - n} ^ {x + n} Q E (x + z) d z \\\\ = \\frac {1}{2 n} \\left[ \\int_ {x - n} ^ {0} (z + b) ^ {2} d z + \\int_ {0} ^ {x + n} (z - b) ^ {2} d z \\right] \\tag {6} \\\\ = x ^ {2} - \\frac {b}{n} x ^ {2} + \\frac {n}{3} - n b + b ^ {2}. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 800, + 468, + 905 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Combining Eq. (5) and Eq. (6), we have", + "bbox": [ + 519, + 90, + 787, + 106 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} D (x, N) = \\mathbf {E} _ {N} [ Q E (X + N) ] - Q E (X) \\\\ = - \\frac {b}{n} x ^ {2} + 2 b x + \\frac {n ^ {2}}{3} - n b. \\tag {7} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 116, + 890, + 167 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We verify this derivation empirically in Sec. 4.", + "bbox": [ + 498, + 176, + 808, + 191 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "It can be observed that given $b$ and $n$ , $D(x, N)$ is a quadratic function with respect to the activation value $x$ . The inequality $D(x, N) \\leq 0, 0 \\leq x \\leq n$ can be easily solved with respect to $x$ as", + "bbox": [ + 496, + 191, + 893, + 252 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n0 \\leq x \\leq n \\left(1 - \\sqrt {\\frac {n}{3 b}}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 263, + 890, + 297 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "which is always feasible given both $b$ and $n$ are positive.", + "bbox": [ + 498, + 306, + 890, + 321 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Theorem 1 indicates that adding Noisy Bias can always reduce the quantization error of elements close to the center of the quantization bin, which is the source of large quantization error in a heavy-tailed distribution.", + "bbox": [ + 496, + 330, + 893, + 391 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In practice, to choose a proper $n$ for a layer in a pretrained model, we can first acquire the empirical distribution of activation $X$ by passing a small amount of calibration data through the pretrained model. With the distribution of $X$ , we can estimate the expected quantization error reduction of the layer numerically as a function of $n$", + "bbox": [ + 496, + 392, + 893, + 482 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} (n) = \\sum_ {x \\in X} [ D (x, N) ]. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 493, + 890, + 526 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Note that $x$ here denotes the distance between each activation value and the corresponding quantization bin center. Though we cannot find a closed-form solution for the minima of $\\mathcal{L}(n)$ , we can easily find a good-enough choice with a linear search over a small range of $n$ .", + "bbox": [ + 498, + 536, + 893, + 611 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. NoisyQuant formulation and pipeline", + "text_level": 1, + "bbox": [ + 500, + 619, + 823, + 637 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given our theoretical analysis on reducing quantization error with Noisy Bias, we propose to apply the Noisy Bias as an enhancement to the PTQ methods of vision transformer models, which we name as \"NoisyQuant\". The pipeline of NoisyQuant is visualized in Fig. 1. Specifically, before performing PTQ and deploying the model, we sample a single Noisy Bias $N \\in \\mathbb{R}^{m \\times 1}$ for each layer from a Uniform distribution, and fix it throughout the inference. The range of $N$ is determined by the calibration objective defined in Eq. (9). At inference time, we add the Noisy Bias $N$ to the input activation $X$ under the broadcasting rule before going through the activation quantizer, therefore actively altering the distribution being quantized to reduce the input quantization error. After the activation-weight multiplication in the linear layer, we remove the impact of $N$ by adjusting the bias term in the linear layer with a denoising bias computed from $N$ , therefore retrieving the correct output. Formally,", + "bbox": [ + 496, + 643, + 893, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "20324", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "NoisyQuant converts the quantized FC computation defined in Eq. (2) into", + "bbox": [ + 75, + 90, + 468, + 121 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nf _ {N q} (X) = q _ {W} (W) q _ {A} (X + N) + \\left(B - q _ {W} (W) N\\right). \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 131, + 468, + 148 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since $N$ is sampled before the deployment of the model and fixed it throughout the inference, the denoising output bias $B' = B - q_W(W)N$ only needs to be computed once before the deployment, without any computation overhead in the inference process. The only computation overhead brought by NoisyQuant at inference time is the on-the-fly summation of $X + N$ , which is negligible comparing to the cost of the matrix multiplications. Both $B'$ and $N$ shall be stored with higher precision. In practice, we store both variables as INT16 to enable integer-only inference.", + "bbox": [ + 75, + 157, + 468, + 309 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Comparing to the output of the floating-point FC layer defined in Eq. (1), the output quantization error induced by the activation quantization defined in Eq. (2) is", + "bbox": [ + 75, + 309, + 468, + 354 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nQ E _ {O} (X) = \\left\\| f _ {q} (X) - f (X) \\right\\| _ {2} ^ {2} = \\left\\| W [ Q (X) - X ] \\right\\| _ {2} ^ {2}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 364, + 468, + 395 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $Q(\\cdot)$ is short of $q_{A}(\\cdot)$ , and we omit the weight quantization $q_{W}(\\cdot)$ as we are focusing on the activation quantization. Similarly, the output quantization error after the application of NoisyQuant can be computed as", + "bbox": [ + 75, + 396, + 468, + 455 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} Q E _ {O} ^ {\\prime} (X) = | | f _ {N q} (X) - f (X) | | _ {2} ^ {2} \\\\ = \\left\\| W [ Q (X + N) - X - N ] \\right\\| _ {2} ^ {2}. \\tag {12} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 465, + 468, + 503 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As we prove in Theorem 1 that the Noisy Bias enables $X + N$ to have a lower quantization error than $X$ , we can achieve lower output quantization error in Eq. (12) than in Eq. (11) when NoisyQuant is applied, therefore improving the PTQ performance of the model.", + "bbox": [ + 75, + 513, + 468, + 588 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Theoretical insight verification", + "text_level": 1, + "bbox": [ + 76, + 601, + 357, + 618 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we provide empirical verification of our theoretical analysis on reducing the activation quantization error with NoisyQuant. We numerically verify Theorem 1 with simulated data in Sec. 4.1, and demonstrate the reduction in both input and output quantization errors under the true inference-time activation distribution in Sec. 4.2.", + "bbox": [ + 75, + 626, + 468, + 717 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Empirical verification of Theorem 1", + "text_level": 1, + "bbox": [ + 76, + 726, + 387, + 742 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here we follow the settings in Theorem 1 to empirically verify its theoretical derivation. Specifically, we set the quantization bin range $b = 1$ , and explore how the quantization error difference induced by the Noisy Bias change with different choice of activation value $x$ and noisy bias range $n$ . For all empirical results we experiment with 10 instances of independently sampled Noisy Bias $N$ , and report the mean and standard derivation of $D(X, N)$ defined in Eq. (3) across the 10 instances. We consider input activation $X$ to be a tensor with 20 dimensions. Given the tens to hundreds", + "bbox": [ + 75, + 750, + 470, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5dfc8a3bd58abeb0ef5849f913c9d72b2732a19dcc8bfd0ef13df63ff8b80773.jpg", + "image_caption": [ + "Figure 2. Verifying Eq. (7) with respect to Noisy Bias range $n$ . Circle indicates the mean and error bar the std for evaluated results." + ], + "image_footnote": [], + "bbox": [ + 524, + 97, + 867, + 227 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c4d1d795130efb6e642d1f738f6648bdfd1e708ad0cad0d6f302718bbe515558.jpg", + "image_caption": [ + "Figure 3. Verifying Eq. (7) with respect to activation value $x$ . Circle indicates the mean and error bar the std for evaluated results." + ], + "image_footnote": [], + "bbox": [ + 521, + 273, + 870, + 419 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "of thousands of activation values in each transformer layer, it is likely to see more than 20 activation elements taking the same value. As we base our theoretical derivation on the Weak Law of Large Numbers, having more elements taking the same value will lead to less variance than the simulation results provided in this section.", + "bbox": [ + 496, + 473, + 890, + 563 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For the first experiment, we fix all elements in $X$ to take value $x = 0.1$ , and alter $n$ in the range of [0.1, 1.9]. We compare the empirical $D(X, N)$ and the theoretical result derived in Eq. (7) in Fig. 2. The evaluated results closely follow the theoretical line. Even with only 20 activation elements the standard deviation across independently sampled Noisy Bias is much smaller than the quantization error benefit brought by adding the Noisy Bias.", + "bbox": [ + 496, + 564, + 892, + 685 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Similarly, we fix $n = 1.4$ and alter activation value $x$ in the range of [0, 0.6]. The results are shown in Fig. 3, where we see the same trend of close mean and small variance as in the previous experiment. Quantization error can be reduced by Noisy Bias when $x < 0.44$ , which follows the numerical computation result of Eq. (4). Both results show the correctness and stability of our derivation of Eq. (7) and the proof of Theorem 1.", + "bbox": [ + 496, + 685, + 890, + 806 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Quantization errors in transformer layers", + "text_level": 1, + "bbox": [ + 500, + 816, + 856, + 832 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As mentioned in Sec. 3.2, NoisyQuant enables the reduction of both input activation quantization error $(QE)$ and output error $(QE_O)$ of a linear layer given a quantizer. In this section, we verify the error reduction on the real acti", + "bbox": [ + 496, + 839, + 893, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "20325", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/d09097d4441909ebb85d52dc87c3a79b69d4cf57a80a76028f2c2c37c5bcc3af.jpg", + "table_caption": [ + "Table 1. Averaged quantization error on different types of layers. The calculation is performed with 5,120 images randomly selected from the ImageNet validation set on the Swin-T model. Layer names are defined as same as in Timm [35]. We compute the input quantization error difference $D$ and compare the output errors between NoisyQuant and EasyQuant." + ], + "table_footnote": [], + "table_body": "
Layer NameEq. (3) D(X)Eq. (11) QEO(X)Eq. (12) QE'(X)Output QE drop
qkv-0.0290.1070.09610%
proj-0.1060.0930.08113%
fc1-0.5270.1260.1232%
fc2-5.1301.0490.85219%
", + "bbox": [ + 112, + 183, + 436, + 292 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4c0a4ca95467480a8818a0f4efbb91c0f57851b30811acad0ea47ec99787d011.jpg", + "image_caption": [ + "Figure 4. Quantization error histogram of fc2 input activation. We apply NoisyQuant on top of the EasyQuant quantizer, and compare the quantization error distribution (red) with EasyQuant (grey)." + ], + "image_footnote": [], + "bbox": [ + 114, + 309, + 428, + 441 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "vation distribution of vision transformer model. Practically, we run evaluation on randomly selected 5120 images of validation set with linear PTQ quantizer EasyQuant [36] as baseline, and calculate $QE$ and $QE_{O}$ with or without adding Noisy Bias. Experiments are ran on different types of layers, namely qkv, proj, fc1, and fc2, respectively. The results are averaged across all layers of each type. As shown in Tab. 1, input quantization error difference $D(X)$ defined in Eq. (3) is consistently lower than 0, which meets our expectation that Noisy Bias reduces input quantization error. This further leads to significant drop in output error between the quantized and floating-point layers for all layer types. Up to $19\\%$ averaged output error drop can be achieved in fc2 layers.", + "bbox": [ + 75, + 505, + 470, + 700 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To further understand the impact NoisyQuant makes, we visualize both the input and output distribution of each layer. We start with highlighting fc2 as it achieves the greatest QE reduction. The fc2 layer takes in the output of GELU activations, which has significantly complicated distribution as introduced in Sec. 1. Fig. 4 visualizes the distribution of input quantization error with or without NoisyQuant. With EasyQuant only, significant amount of activation elements are showing large quantization error, which is most probably owing to bad quantization on densely distributed areas. Instead, NoisyQuant generally makes QE distributed plainly and for most elements near zero, therefore leading to significant QE reduction overall. Subsequently, we plot the", + "bbox": [ + 75, + 704, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/978b7759d6d4bb4eb3539e9a99a82b200347cb51e1dfb99fa21970f16265939b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 87, + 890, + 174 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/716ff5a9eb45fde689295d819729611413002e0751353a84adaa23166f962598.jpg", + "image_caption": [ + "(a) $26\\%$ $QE_{O}$ reduction on qkv layer", + "(b) $25\\%$ $QE_{O}$ reduction on proj layer" + ], + "image_footnote": [], + "bbox": [ + 501, + 189, + 890, + 276 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/87dda685b56e4bca845df8e6e12ebceaa8357c72453f137fc7ba1261466a3c4c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 292, + 890, + 378 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/786681dd849f84ee122486e2fe11d4cdbe44948926dc304f6bcc5e46f425c70b.jpg", + "image_caption": [ + "(c) $36\\%$ $QE_{O}$ reduction on fc1 layer", + "(d) $60\\%$ $QE_{O}$ reduction on fc2 layer", + "Figure 5. Output histogram in selected transformer layers before and after activation quantization. NoisyQuant significantly reduces the output quantization error over EasyQuant baselines." + ], + "image_footnote": [], + "bbox": [ + 501, + 393, + 890, + 479 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "histogram of selected layers' outputs in Fig. 5. NoisyQuant output consistently follows the floating-point layer output distribution closer than that of EasyQuant, especially at values with large magnitudes. This leads to up to $60\\%$ error reduction in the output of some vision transformer layers.", + "bbox": [ + 496, + 555, + 893, + 631 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 500, + 645, + 632, + 662 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This section evaluates the impact of NoisyQuant on the PTQ performance of some commonly used vision transformer architectures. First, we evaluate the performance on image classification tasks with the ViT [10], DeiT [30] and Swin [22] transformer models. We also evaluate the performance of object detection with the DETR model [4]. Finally, we provide ablation study on how NoisyQuant improves the PTQ performance when added to each type of layers in the vision transformer model.", + "bbox": [ + 496, + 670, + 893, + 806 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Experimental settings", + "text_level": 1, + "bbox": [ + 500, + 816, + 702, + 832 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset. The classification tasks are implemented on the ImageNet-2012 classification dataset [28], with 1.2 million training images and 50,000 validation images in 1000 classes. For object detection task, the MSCOCO 2017 dataset [20]", + "bbox": [ + 496, + 839, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "20326", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/de6ec65404be564baab6175275f5587c05c5ddfe8e38a932532bebc1256d7f4e.jpg", + "table_caption": [ + "Table 2. PTQ accuracy of ViT models. Top-1 accuracy on ImageNet is reported in the table. W/A represents weight and activation bit-width, respectively. The default patch size is $16 \\times 16$ and the default input resolution is $224 \\times 224$ . “*” after model name indicates $384 \\times 384$ input. “MP” represents mixed-precision. For our methods, “NoisyQuant-Linear” is implemented on top of the EasyQuant quantizer, whereas “NoisyQuant-PTQ4ViT” is implemented on the nonlinear PTQ4ViT quantizer. All NoisyQuant results are shown with mean±STD of 20 repeated experiments." + ], + "table_footnote": [], + "table_body": "
MethodW/AViT-SViT-BViT-B*
Pretrained32/3281.3984.5486.00
Percentile [18]6/667.7477.6377.60
Liu et al. [23]6 MP-75.26-
PTQ4ViT-Linear [43]6/670.2475.6646.88
EasyQuant [36]6/675.1381.4282.02
NoisyQuant-Linear6/676.86±0.0681.90±0.1183.00±0.09
PTQ4ViT [43]6/678.6381.6583.34
NoisyQuant-PTQ4ViT6/678.65±0.0782.32±0.0983.22±0.10
Percentile [18]8/878.7780.1282.53
Liu et al. [23]8 MP-76.98-
FQ-ViT [21]8/8-83.31-
PTQ4ViT-Linear [43]8/880.4683.8985.35
EasyQuant [36]8/880.7583.8985.53
NoisyQuant-Linear8/880.81±0.0184.10±0.0385.56±0.01
PTQ4ViT [43]8/881.0084.2585.82
NoisyQuant-PTQ4ViT8/881.15±0.0284.22±0.0185.86±0.01
", + "bbox": [ + 94, + 224, + 455, + 448 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/353a350ed42faf754b2ecf58a2820f1006733e0e33007ee09f266661a19e570f.jpg", + "table_caption": [ + "Table 3. PTQ accuracy of DeiT models. Abbreviations are the same as Tab. 2." + ], + "table_footnote": [], + "table_body": "
PretrainedW/ADeiT-SDeiT-BDeiT-B*
Pretrained32/3279.8581.8083.11
Percentile [18]6/670.4973.9978.24
Bit-Split [32]6/674.5876.39-
Liu et al. [23]6/674.5877.02-
Liu et al. [23]6 MP75.1077.47-
PTQ4VIT-Linear [43]6/672.2678.7868.44
EasyQuant [36]6/675.2779.4781.26
NoisyQuant-Linear6/676.37±0.1379.77±0.0881.40±0.06
PTQ4VIT [43]6/676.2880.2581.55
NoisyQuant-PTQ4ViT6/677.43±0.0880.70±0.1181.65±0.04
Percentile [18]8/873.9875.2180.02
Bit-Split [32]8/876.3979.42-
Liu et al. [23]8/877.4780.48-
Liu et al. [23]8 MP78.0981.29-
FQ-ViT [21]8/879.1781.20-
PTQ4VIT-Linear [43]8/877.6580.9482.33
EasyQuant [36]8/878.9881.1982.10
NoisyQuant-Linear8/879.11±0.0281.30±0.0382.23±0.02
PTQ4VIT [43]8/879.4781.4882.97
NoisyQuant-PTQ4ViT8/879.51±0.0181.45±0.0282.49±0.02
", + "bbox": [ + 93, + 503, + 455, + 776 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "is utilized to evaluate the PTQ performance, which contains 118,000 training images and 5,000 validation images. For both tasks we randomly sample 1024 images from the training set as the calibration data for all PTQ methods implemented in our experiments.", + "bbox": [ + 75, + 792, + 470, + 868 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Pretrained model architecture. We perform uniform 6-bit and 8-bit PTQ on different variants of pretrained vision", + "bbox": [ + 75, + 869, + 472, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/0500aafde49fdf2784f43e76016007aebe708d654228d889abab876ed9efbf1f.jpg", + "table_caption": [ + "Table 4. PTQ accuracy of Swin models. Abbreviations are the same as Tab. 2." + ], + "table_footnote": [], + "table_body": "
MethodW/ASwin-TSwin-SSwin-BSwin-B*
Pretrained32/3281.3983.2385.2786.44
Percentile [18]6/677.7580.4181.9082.94
PTQ4VIT-Linear [43]6/678.4581.7483.3585.22
EasyQuant [36]6/679.5182.4584.3085.89
NoisyQuant-Linear6/680.01±0.0682.78±0.0484.57±0.0485.96±0.04
PTQ4VIT [43]6/680.4782.3884.0185.38
NoisyQuant-PTQ4ViT6/680.51±0.0382.86±0.0584.68±0.0686.03±0.07
Percentile [18]8/879.8880.9383.0884.31
FQ-ViT [21]8/880.5182.71--
PTQ4VIT [43]8/880.9682.7584.7986.16
EasyQuant [36]8/880.9583.0085.1086.39
NoisyQuant-Linear8/881.05±0.0383.07±0.0385.11±0.0486.42±0.02
PTQ4VIT [43]8/881.2483.1085.1486.39
NoisyQuant-PTQ4ViT8/881.25±0.0283.13±0.0185.20±0.0386.44±0.01
", + "bbox": [ + 506, + 128, + 888, + 313 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "transformer models. For classification tasks we quantize the model family of ViT [10], DeiT [30], and Swin [22] provided by the Timm library [35], including large-scale models with $384 \\times 384$ input resolution (marked with “*”). For detection, we perform PTQ on the official implementation of DETR [4] with ResNet-50 backbone. We use the pretrained model checkpoint provided by the official source of each model, whose floating-point performances are reported in the results tables and match their original papers.", + "bbox": [ + 496, + 318, + 893, + 455 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Implementation Details. Following the mainstream quantization methods [23, 43], we quantize all the weights and inputs involved in matrix multiplication. More specifically, we quantize weights of qkv-projectors, attention output projectors, MLPs, linear embeddings, and model heads. Besides, we also quantize input activations of linear layers and matrix multiplication operators. Following [23, 26, 43, 45], we keep layer normalization and softmax operations as full precision. Apart from special notes, we perform symmetric layer-wise quantization for activations and symmetric channel-wise quantization for weights. In particular, weights are quantized by absolute MinMax values without clamping. We implement NoisyQuant on top of linear quantizer EasyQuant [36] and nonlinear quantizer PTQ4ViT [43]. After we apply the selected quantizer to determine the quantization bin width and location on calibration data, we decide the parameters of the Noisy Bias distribution through a linear search. We use the empirical activation quantization error as defined in Eq. (9) as the search objective, where we find the Noisy Bias parameter that can minimize the quantization error.", + "bbox": [ + 496, + 455, + 893, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. NoisyQuant performance", + "text_level": 1, + "bbox": [ + 500, + 770, + 732, + 787 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Here we compare NoisyQuant with state-of-the-art PTQ methods, including percentile [18], bit-split [32], Liu et al. [23], FQ-ViT [21], EasyQuant [36], and PTQ4ViT [43], on ImageNet. PTQ performance of different variants of ViT, DeiT, and Swin transformer models are provided in Tab. 2, Tab. 3, and Tab. 4, respectively. Experiments are conducted for both linear and nonlinear quantizers.", + "bbox": [ + 496, + 794, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "20327", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4ba97f894bd5d4ee093f4e1dd889e2034880bca12af4ed2fd2c0da5aafe6acb8.jpg", + "table_caption": [ + "Table 5. PTQ accuracy of DETR models. Mean average precision (mAP) evaluated on the MSCOCO 2017 dataset is reported." + ], + "table_footnote": [], + "table_body": "
ModelMethodW/AmAP
DETR [4] (COCO2017)Pretrained32/3242.0
Percentile [18]8/838.6
Bit-Split [32]8/840.6
Liu et al. [23]8/841.2
EasyQuant [36]8/841.1
NoisyQuant-Linear8/841.4±0.05
", + "bbox": [ + 125, + 128, + 426, + 227 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Linear quantizer. Previous linear PTQ methods suffer from severe performance degradation on vision transformer activation quantization. EasyQuant [36] achieves the best performance among linear quantizers, yet still induce 2-6% of accuracy drop under 6-bit quantization compared to the floating-point baseline. By applying the proposed NoisyQuant method on top of the EasyQuant quantizer, the resulted NoisyQuant-Linear quantizer consistently and significantly improves the PTQ accuracy on all variants of vision transformer models. Under the W6A6 setting, NoisyQuant-Linear achieves performance improvement of 1.73% on ViT-S, 0.98% on ViT-B*, 1.1% on DeiT-S, and 0.5% on SWIN-T over EasyQuant. Under the W8A8 setting, as the performance gap between EasyQuant and floating-point gets smaller, NoisyQuant still appears beneficial in further improving the PTQ performance. It is worth noting that although nonlinear quantizer like PTQ4ViT [43] consistently achieves higher PTQ performance than the linear EasyQuant quantizer, the enhancement brought by NoisyQuant enables the NoisyQuant-Linear quantizer to achieve on-par or even better performance than the nonlinear PTQ4ViT, with much lower overhead in hardware deployment.", + "bbox": [ + 75, + 282, + 472, + 614 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Nonlinear quantizer. As we claim NoisyQuant as a quantizer-agnostic enhancement of PTQ, we further implement NoisyQuant on top of the nonlinear quantizer PTQ4ViT [43], namely NoisyQuant-PTQ4ViT. NoisyQuant-PTQ4ViT outperforms most results of PTQ4ViT, notably achieving for a $1.25\\%$ improvement on DeiT-S, $0.67\\%$ on ViT-S, and $0.67\\%$ on Swin-B in the W6A6 setting. The W6A6 PTQ performance of Swin-S and Swin-B* for the first time improved to within $0.5\\%$ of the floating-point baseline by the NoisyQuant-PTQ4ViT quantizer, which can never be achieved without the help of NoisyQuant.", + "bbox": [ + 75, + 628, + 470, + 796 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Besides classification models, we report the PTQ performance of the DETR object detection model in Tab. 5. NoisyQuant implemented on top of the EasyQuant quantizer also outperforms the EasyQuant baseline, and all previous linear PTQ quantizers including percentile [18], bit-split [32], and Liu et al. [23].", + "bbox": [ + 75, + 809, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f34b2c5db688afa3a050d627d05d15bddb196c281a997b2eeafa2932f8659678.jpg", + "table_caption": [ + "Table 6. Applying NoisyQuant on different layer types. Starting from W6A6 PTQ with the EasyQuant quantizer, we add NoisyQuant to each type of layers in the transformer model." + ], + "table_footnote": [], + "table_body": "
Modelqkv noiseproj noisefc1 noisefc2 noiseTop-1 W6A6
DeiT-S [30]XXXX75.27
✓XXX75.38
X✓XX75.45
XX✓X75.33
XXX✓76.21
✓✓✓✓76.37
Swin-T [22]XXXX79.51
✓XXX79.53
X✓XX79.56
XX✓X79.52
XXX✓79.80
✓✓✓✓80.01
", + "bbox": [ + 545, + 142, + 848, + 347 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. NoisyQuant's impact on different layer types", + "text_level": 1, + "bbox": [ + 498, + 361, + 879, + 378 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As we show the effectiveness of NoisyQuant in improving the PTQ performance of the entire model, here we explore if NoisyQuant is helpful for all types of layers in the vision transformer model. Tab. 6 exhibits the effect of applying NoisyQuant on different types of layers. The checkmark in the table means that we apply NoisyQuant to the input activation of all the layers of that particular type, otherwise the layer is quantized with EasyQuant only. For all layer types, applying NoisyQuant consistently improves the PTQ performance of the entire model. Specifically, 'fc2' layers derive the greatest benefit from NoisyQuant, which corresponds to our previous analysis that the activation distribution after GELU function brings the major challenge to PTQ, which can be resolved with the proposed NoisyQuant method. After adding NoisyQuant to the input activation of all linear layers, the model achieves the maximum performance boost.", + "bbox": [ + 496, + 388, + 893, + 628 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusions", + "text_level": 1, + "bbox": [ + 500, + 651, + 627, + 667 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work proposes NoisyQuant, a noisy bias-enhanced post-training activation quantization method for the complicated activation distribution of vision transformer models. We theoretically prove the quantization error reduction brought by adding noisy bias to the input activation before quantization, and empirically show the effectiveness of NoisyQuant on both linear and nonlinear quantizers on vision transformer models. We hope this work opens up a new direction to reduce PTQ quantization error through actively changing the distribution being quantized, and inspires better Noisy Bias distribution design or generation method based on quantizer and activation characteristics.", + "bbox": [ + 496, + 676, + 893, + 844 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement This work was supported in part by the National Key Research and Development Program of China under Grant 2022YFB4400900. The Berkeley team acknowledges support from Berkeley Deep Drive, Intel Corporation, and Panasonic.", + "bbox": [ + 496, + 844, + 893, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "20328", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Yash Bhalgat, Jinwon Lee, Markus Nagel, Tijmen Blankevoort, and Nojun Kwak. Lsq+: Improving low-bit quantization through learnable offsets and better initialization. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 2978-2985, 2020. 1", + "[2] Yaohui Cai, Zhewei Yao, Zhen Dong, Amir Gholami, Michael W. Mahoney, and Kurt Keutzer. Zeroq: A novel zero shot quantization framework. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13166-13175, 2020. 1, 3", + "[3] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 1, 3", + "[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. ArXiv, abs/2005.12872, 2020. 6, 7, 8", + "[5] Minghao Chen, Houwen Peng, Jianlong Fu, and Haibin Ling. Autoformer: Searching transformers for visual recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12270-12280, 2021. 1", + "[6] Tianlong Chen, Yu Cheng, Zhe Gan, Lu Yuan, Lei Zhang, and Zhangyang Wang. Chasing sparsity in vision transformers: An end-to-end exploration. Advances in Neural Information Processing Systems, 34:19974-19988, 2021. 1", + "[7] Jungwook Choi, Zhuo Wang, Swagath Venkataramani, Pierce I-Jen Chuang, Vijayalakshmi Srinivasan, and Kailash Gopalakrishnan. Pact: Parameterized clipping activation for quantized neural networks. arXiv preprint arXiv:1805.06085, 2018. 1", + "[8] Zhen Dong, Zhewei Yao, Daiyaan Arfeen, Amir Gholami, Michael W Mahoney, and Kurt Keutzer. Hawq-v2: Hessian aware trace-weighted quantization of neural networks. Advances in neural information processing systems, 33:18518-18529, 2020. 1", + "[9] Zhen Dong, Zhewei Yao, Amir Gholami, Michael W Mahoney, and Kurt Keutzer. Hawq: Hessian aware quantization of neural networks with mixed-precision. In Proceedings of the IEEE International Conference on Computer Vision, pages 293–302, 2019. 1", + "[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1, 3, 6, 7", + "[11] Amir Gholami, Sehoon Kim, Zhen Dong, Zhewei Yao, Michael W Mahoney, and Kurt Keutzer. A survey of quantization methods for efficient neural network inference. arXiv preprint arXiv:2103.13630, 2021. 1", + "[12] Cong Guo, Yuxian Qiu, Jingwen Leng, Xiaotian Gao, Chen Zhang, Yunxin Liu, Fan Yang, Yuhao Zhu, and Minyi Guo." + ], + "bbox": [ + 78, + 114, + 472, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Squant: On-the-fly data-free quantization via diagonal hessian approximation. arXiv preprint arXiv:2202.07471, 2022.3", + "[13] Song Han, Huizi Mao, and William J Dally. Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. arXiv preprint arXiv:1510.00149, 2015. 1", + "[14] Ali Hatamizadeh, Hongxu Yin, Jan Kautz, and Pavlo Molchanov. Global context vision transformers. arXiv preprint arXiv:2206.09959, 2022. 3", + "[15] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016. 2, 3", + "[16] Mark Horowitz. 1.1 computing's energy problem (and what we can do about it). In ISSCC, 2014. 3", + "[17] Kwonjoon Lee, Huiwen Chang, Lu Jiang, Han Zhang, Zhuowen Tu, and Ce Liu. Vitgan: Training gans with vision transformers. arXiv preprint arXiv:2107.04589, 2021.3", + "[18] Rundong Li, Yan Wang, Feng Liang, Hongwei Qin, Junjie Yan, and Rui Fan. Fully quantized network for object detection. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2805-2814, 2019. 7, 8", + "[19] Zhikai Li, Junrui Xiao, Lianwei Yang, and Qingyi Gu. Repqvit: Scale reparameterization for post-training quantization of vision transformers. arXiv preprint arXiv:2212.08254, 2022.3", + "[20] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dolkar, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 6", + "[21] Yang Lin, Tianyu Zhang, Peiqin Sun, Zheng Li, and Shuchang Zhou. Fq-vit: Post-training quantization for fully quantized vision transformer. In IJCAI, 2022. 7", + "[22] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 9992-10002, 2021. 1, 3, 6, 7, 8", + "[23] Zhenhua Liu, Yunhe Wang, Kai Han, Siwei Ma, and Wen Gao. Post-training quantization for vision transformer. In NeurIPS, 2021. 1, 2, 3, 4, 7, 8", + "[24] Zhenhua Liu, Yunhe Wang, Kai Han, Wei Zhang, Siwei Ma, and Wen Gao. Post-training quantization for vision transformer. Advances in Neural Information Processing Systems, 34:28092-28103, 2021. 1", + "[25] A. Polino, R. Pascanu, and D. Alistarh. Model compression via distillation and quantization. arXiv preprint arXiv:1802.05668, 2018. 1", + "[26] Gabriele Prato, Ella Charlaix, and Mehdi Rezagholizadeh. Fully quantized transformer for improved translation. ArXiv, abs/1910.10485, 2019. 7", + "[27] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 3", + "[28] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy," + ], + "bbox": [ + 501, + 92, + 893, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "20329", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Aditya Khosla, Michael S. Bernstein, Alexander C. Berg, and Li Fei-Fei. Imagenet large scale visual recognition challenge. International Journal of Computer Vision, 115:211-252, 2015. 6", + "[29] Sheng Shen, Zhen Dong, Jiayu Ye, Linjian Ma, Zhewei Yao, Amir Gholami, Michael W Mahoney, and Kurt Keutzer. Qbert: Hessian based ultra low precision quantization of bert. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 8815-8821, 2020. 1", + "[30] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Herv'e J'egou. Training data-efficient image transformers & distillation through attention. In ICML, 2021. 1, 3, 6, 7, 8", + "[31] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 1, 3", + "[32] Peisong Wang, Qiang Chen, Xiangyu He, and Jian Cheng. Towards accurate post-training network quantization via bitsplit and stitching. In ICML, 2020. 1, 7, 8", + "[33] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. arXiv preprint arXiv:2102.12122, 2021. 3", + "[34] Xiuying Wei, Yunchen Zhang, Xiangguo Zhang, Ruihao Gong, Shanghang Zhang, Qi Zhang, Fengwei Yu, and Xianglong Liu. Outlier suppression: Pushing the limit of low-bit transformer language models. arXiv preprint arXiv:2209.13325, 2022. 3", + "[35] Ross Wightman. Pytorch image models. https://github.com/rwrightman/pytorch-image-models, 2019.6, 7", + "[36] Di Wu, Qi Tang, Yongle Zhao, Ming Zhang, Ying Fu, and Debing Zhang. Easyquant: Post-training quantization via scale optimization. arXiv preprint arXiv:2006.16669, 2020. 1, 2, 3, 6, 7, 8", + "[37] Lirui Xiao, Huanrui Yang, Zhen Dong, Kurt Keutzer, Li Du, and Shanghang Zhang. Csq: Growing mixed-precision quantization scheme with bi-level continuous sparsification. arXiv preprint arXiv:2212.02770, 2022. 1", + "[38] Huanrui Yang, Lin Duan, Yiran Chen, and Hai Li. Bsq: Exploring bit-level sparsity for mixed-precision neural network quantization. arXiv preprint arXiv:2102.10462, 2021. 1", + "[39] Huanrui Yang, Hongxu Yin, Pavlo Molchanov, Hai Li, and Jan Kautz. Nvit: Vision transformer compression and parameter redistribution. arXiv preprint arXiv:2110.04869, 2021. 1", + "[40] Zhewei Yao, Zhen Dong, Zhangcheng Zheng, Amir Gholami, Jiali Yu, Eric Tan, Leyuan Wang, Qijing Huang, Yida Wang, Michael Mahoney, et al. Hawq-v3: Dyadic neural network quantization. In International Conference on Machine Learning, pages 11875-11886. PMLR, 2021. 3", + "[41] Li Yuan, Yunpeng Chen, Tao Wang, Weihao Yu, Yujun Shi, Zihang Jiang, Francis EH Tay, Jiashi Feng, and Shuicheng Yan. Tokens-to-token vit: Training vision transformers from scratch onImagenet. arXiv preprint arXiv:2101.11986, 2021.3" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[42] Zhihang Yuan, Chenhao Xue, Yiqi Chen, Qiang Wu, and Guangyu Sun. Ptg4vit: Post-training quantization framework for vision transformers. arXiv preprint arXiv:2111.12293, 2021. 1, 2, 4", + "[43] Zhihang Yuan, Chenhao Xue, Yi-Qiang Chen, Qiang Wu, and Guangyu Sun. Ptq4vit: Post-training quantization for vision transformers with twin uniform quantization. In ECCV, 2022. 2, 3, 7, 8", + "[44] Dongqing Zhang, Jiaolong Yang, Dongqiangzi Ye, and Gang Hua. Lq-nets: Learned quantization for highly accurate and compact deep neural networks. In Proceedings of the European conference on computer vision (ECCV), pages 365–382, 2018. 1", + "[45] Wei Zhang, Lu Hou, Yichun Yin, Lifeng Shang, Xiao Chen, Xin Jiang, and Qun Liu. Ternarybert: Distillation-aware ultra-low bit bert. In EMNLP, 2020. 7", + "[46] Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xiang, Philip H.S. Torr, and Li Zhang. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In CVPR, 2021. 3", + "[47] Shuchang Zhou, Yuxin Wu, Zekun Ni, Xinyu Zhou, He Wen, and Yuheng Zou. Dorefa-net: Training low bitwidth convolutional neural networks with low bitwidth gradients. arXiv preprint arXiv:1606.06160, 2016. 1" + ], + "bbox": [ + 501, + 92, + 893, + 444 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "20330", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/NoisyQuant_ Noisy Bias-Enhanced Post-Training Activation Quantization for Vision Transformers/68808ca4-cf24-4585-b35b-3e2acef953ea_model.json b/2023/NoisyQuant_ Noisy Bias-Enhanced Post-Training Activation Quantization for Vision Transformers/68808ca4-cf24-4585-b35b-3e2acef953ea_model.json new file mode 100644 index 0000000000000000000000000000000000000000..036f17bb908b29b74581289859bf74d7acb96ad5 --- /dev/null +++ b/2023/NoisyQuant_ Noisy Bias-Enhanced Post-Training Activation Quantization for Vision Transformers/68808ca4-cf24-4585-b35b-3e2acef953ea_model.json @@ -0,0 +1,2156 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.098, + 0.13, + 0.877, + 0.175 + ], + "angle": 0, + "content": "NoisyQuant: Noisy Bias-Enhanced Post-Training Activation Quantization for Vision Transformers" + }, + { + "type": "text", + "bbox": [ + 0.126, + 0.202, + 0.847, + 0.24 + ], + "angle": 0, + "content": "Yijiang Liu\\*1, Huanrui Yang\\*2, Zhen Dong?, Kurt Keutzer?, Li Du1, Shanghang Zhang3Nanjing University, 2University of California, Berkeley" + }, + { + "type": "text", + "bbox": [ + 0.107, + 0.24, + 0.862, + 0.293 + ], + "angle": 0, + "content": "\\(^{3}\\)National Key Laboratory for Multimedia Information Processing, School of Computer Science, Peking University liuyijiang@smail.nju.edu.cn, {huanrui, zhendong, keutzer}@berkeley.edu ldu@nju.edu.cn, shanghang@pku.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.342, + 0.314, + 0.359 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.374, + 0.474, + 0.737 + ], + "angle": 0, + "content": "The complicated architecture and high training cost of vision transformers urge the exploration of post-training quantization. However, the heavy-tailed distribution of vision transformer activations hinders the effectiveness of previous post-training quantization methods, even with advanced quantizer designs. Instead of tuning the quantizer to better fit the complicated activation distribution, this paper proposes NoisyQuant, a quantizer-agnostic enhancement for the post-training activation quantization performance of vision transformers. We make a surprising theoretical discovery that for a given quantizer, adding a fixed Uniform noisy bias to the values being quantized can significantly reduce the quantization error under provable conditions. Building on the theoretical insight, NoisyQuant achieves the first success on actively altering the heavy-tailed activation distribution with additive noisy bias to fit a given quantizer. Extensive experiments show NoisyQuant largely improves the post-training quantization performance of vision transformer with minimal computation overhead. For instance, on linear uniform 6-bit activation quantization, NoisyQuant improves SOTA top-1 accuracy on ImageNet by up to \\(1.7\\%\\), \\(1.1\\%\\) and \\(0.5\\%\\) for ViT, DeiT, and Swin Transformer respectively, achieving on-par or even higher performance than previous nonlinear, mixed-precision quantization." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.765, + 0.21, + 0.781 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.79, + 0.471, + 0.868 + ], + "angle": 0, + "content": "Inspired by the success of Self Attention (SA)-based transformer models in Natural Language Processing (NLP) tasks [31], recent researches make significant progress in applying transformer models to the field of computer vision [3, 10, 22, 30]. In the meantime, the typical design of" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.329, + 0.896, + 0.48 + ], + "angle": 0, + "content": "transformer models induces large model sizes, high computational consumption, and long training time. For instance, the widely used DeiT-Base model [30] contains 86M parameters, logs 18G floating-point operations for a single input, and requires 300 epochs of training on the ImageNet dataset. This leads to significant difficulties in hardware deployment. In facing such difficulty, a number of compression and acceleration methods are applied to vision transformer models, including pruning [6, 39], quantization [24, 42], and neural architecture search [5], etc." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.487, + 0.896, + 0.835 + ], + "angle": 0, + "content": "Among these methods, quantization appears as one of the most effective and widely applied ways [11]. Quantization process uses a predefined \"quantizer\" function to convert the continuous representation of weights and activations into a small number of discrete symbols, therefore enabling low-precision representations for straightforward memory savings. For DNN models, the approximation error made by the quantizer inevitably leads to performance drop. A series of work focuses on Quantization-Aware Training (QAT) that finetunes the quantized model at low precision [8, 9, 25, 37, 47]. However, given the high training cost and the complicated computation graph of vision transformer models, retraining the model at low precision could be costly and unstable [42]. Alternatively, Post-Training Quantization (PTQ) is preferable for vision transformers as it eliminates the need for re-training or finetuning the quantized model, instead only adjusts the design of the quantizer based on the full-precision pretrained model and a small set of sampled calibration data [1, 2, 23, 32, 36]. For example, linear quantizers [7, 29, 36, 38] reduce quantization error by shifting, clipping, and scaling the values to be quantized. Nonlinear quantizers [13, 42, 44] further adjust the width and location of each quantization bin to better fit the distribution." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.84, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Unfortunately, though progresses are made in designing better PTQ quantizers, it still appears to be significantly challenging to quantize vision transformer models, especially for activation quantization. Transformer layers produce up" + }, + { + "type": "page_footnote", + "bbox": [ + 0.1, + 0.875, + 0.241, + 0.901 + ], + "angle": 0, + "content": "* Equal contribution. \nCorresponding Author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.958 + ], + "angle": 0, + "content": "20321" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.089, + 0.87, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.312, + 0.896, + 0.438 + ], + "angle": 0, + "content": "Figure 1. Overview of the NoisyQuant pipeline (orange box) with comparison to EasyQuant [36] (gray box). Histograms of input activations, quantized activations, and layer outputs are illustrated in the pipeline. Starting from the input activation \\( X \\) following a GELU function, where the green histogram is plotted in linear-scale and the gray in log-scale, NoisyQuant adds a fixed NoisyBias \\( N \\) sampled from a Uniform distribution onto \\( X \\). Adding the noise before quantization flattens the peaks in the activation distribution to make it more friendly to quantization, as visualized in the linear/log-scaled histogram of \\( X + N \\), and theoretically proved in Sec. 3.2. The noise is removed from the result of the fixed-point activation-weight multiplication with a denoising bias term, which we formulate in Sec. 3.3. The rightmost sub-figures illustrate the layer output distribution of EasyQuant and NoisyQuant. Compared to the output of the floating-point model (yellow), NoisyQuant output follows the distribution closely and achieves up to \\( 60\\% \\) output error reduction on some transformer layers, as shown in Sec. 4. The resulting PTQ performance improvement of the entire model is provided in Sec. 5." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.451, + 0.473, + 0.633 + ], + "angle": 0, + "content": "to millions of activation values with sophisticated distributions. For instance, outputs of the GELU function [15] are asymmetrically distributed, with spikes in the histogram at some values and a long tail in a large range (see top-left of Fig. 1). Some linear projection layers also lead to significantly large activation values that are sparsely distributed in a very long tail [23]. Consequently, low-precision PTQ on vision transformer suffers from performance degradation, even utilizing non-linear or mixed-precision quantizers at the cost of additional data communication and computation overheads [23,43]. No linear uniform PTQ method achieves good performance on vision transformer models." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.645, + 0.473, + 0.903 + ], + "angle": 0, + "content": "This paper provides a brand new perspective in dealing with the complicated vision transformer activation distribution. Instead of adding more tricks in the quantizer design to fit the activation distribution, this work explores the potential of actively and cost-effectively altering the distribution being quantized, making it more friendly to a given quantizer. The pipeline of our proposed method is illustrated in Fig. 1. Specifically, we make a surprising discovery that for any quantizer, the quantization error can be significantly reduced by adding a fixed noisy bias sampled from a Uniform distribution to the activation before quantization. We theoretically prove the condition when the quantization error reduction can be achieved. On this basis, we propose NoisyQuant, a plug-and-play, quantizer-agnostic enhancement on the posttraining activation quantization of vision transformer models. For each layer, we sample a Noisy Bias based on the input activation distribution following our theoretical analysis, and" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.451, + 0.895, + 0.632 + ], + "angle": 0, + "content": "compute the corresponding denoising bias to retain the correct output of the layer. At inference time, the Noisy Bias is added to the input activation before quantization, and the denoising bias is removed from the layer output. This process significantly reduces the quantization error with minimal computation overhead. NoisyQuant leads to significant improvement in the PTQ performance of state-of-the-art vision transformers. Applying NoisyQuant on top of a uniform linear quantization achieves on-par performance to SOTA mixed-precision, nonlinear PTQ methods [23, 42]. Adding NoisyQuant on top of these nonlinear quantizers achieves further performance gain." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.632, + 0.895, + 0.662 + ], + "angle": 0, + "content": "To the best of our knowledge, this paper makes the following novel contributions:" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.671, + 0.895, + 0.716 + ], + "angle": 0, + "content": "- Theoretically shows the possibility and proves feasible conditions for reducing the quantization error of heavily-tailed distributions with a fixed additive noisy bias;" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.726, + 0.895, + 0.815 + ], + "angle": 0, + "content": "- Proposes NoisyQuant, a quantizer-agnostic enhancement for PTQ performance on activation quantization. NoisyQuant achieves the first success in actively refining the distribution being quantized to reduce quantization error following the theoretical results on additive noisy bias, with minimal computation overhead;" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.826, + 0.895, + 0.901 + ], + "angle": 0, + "content": "- Demonstrates consistent performance improvement by applying NoisyQuant on top of existing PTQ quantizers. For 6-bit PTQ, NoisyQuant improves the ImageNet top-1 accuracy of uniform linear quantized vision transformer models by up to \\(1.7\\%\\), and improves SOTA" + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.671, + 0.895, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20322" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.11, + 0.092, + 0.441, + 0.108 + ], + "angle": 0, + "content": "nonlinear quantizer PTQ4ViT [43] by up to \\(0.7\\%\\)." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.122, + 0.216, + 0.138 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.147, + 0.418, + 0.164 + ], + "angle": 0, + "content": "2.1. Transformer models in computer vision" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.171, + 0.473, + 0.444 + ], + "angle": 0, + "content": "Since Self-Attention (SA) [31]-based transformer models have shown impressive performance in Natural Language Processing tasks, researchers transfer the idea into computer vision. Pioneering work on Vision Transformer [10] for the first time builds a totally SA-based model in the field of computer vision and shows its effectiveness. Since then, a variety of transformer-based vision models have emerged with boosted performance, claiming state-of-the-art in multiple computer vision tasks. Notable advances include additional data augmentation and distillation [30], incorporating multi-stage architecture [22, 33], and explore novel attention design [14,22,41]. Transformer models also advance to tasks beyond classification, such as object detection [3], semantic segmentation [46], and image generation [17]. Furthermore, the development of vision transformer brings unification of model architecture and representation across image and text, enabling stronger vision-language connection via image-text-pair contrastive learning [27]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.444, + 0.473, + 0.626 + ], + "angle": 0, + "content": "The rapid development of vision transformer models motivates the exploration of quantization, as quantization brings straightforward memory and computation savings agnostic to model architectures. Compared to CNN models, the use of advanced activation functions like GELU [15] and more complicated computation graphs like the attention mechanism makes vision transformers more sensitive to quantization [43]. This work aims to resolve the difficulty of posttraining vision transformer activation quantization caused by the heavy-tailed activation distribution, which we believe will facilitate the deployment of state-of-the-art transformer models into real-world computer vision applications." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.636, + 0.387, + 0.652 + ], + "angle": 0, + "content": "2.2. Post-training quantization methods" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.473, + 0.901 + ], + "angle": 0, + "content": "Post-training quantization (PTQ) is a preferable quantization method in scenarios with limited training data or limited computational resources for the expensive quantization-aware fine-tuning process. Previous methods have dedicatedly studied PTQ on CNNs. For instance, EasyQuant [36] presents a quick searching mechanism to determine the proper clipping range for quantization. ZeroQ [2] proposes a distillation mechanism to generate proxy input images, which can leverage inherent statistics of batch normalization layers to conduct PTQ. SQuant [12] calibrates the model for lower quantization error on the basis of the sensitivity obtained by the Hessian spectrum. On vision transformers, Liu et al. [23] presents a linear PTQ method that uses Pearson correlation coefficients and ranking loss to determine scaling factors. PTQ4ViT [43] introduces a nonlinear Twin Uniform Quantization based on vision transformers' acti" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.894, + 0.182 + ], + "angle": 0, + "content": "vation distributions, which set different scaling factors for 1) positive and negative activations of GeLU, and 2) small and large values of Softmax, which can reduce the activation quantization error to some extent with the cost of additional computation overhead. [19, 34] reparameterize the LN layer to suppress outliers by scaling down activation values." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.183, + 0.895, + 0.288 + ], + "angle": 0, + "content": "Unlike previous PTQ methods that analyze the activation distribution and fit the quantizer accordingly, our method takes a novel perspective of actively modifying the distribution being quantized with a sampled noisy bias. We show our proposed method can bring quantization error reduction and performance enhancement for all the PTQ quantizers for vision transformer activation quantization." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.301, + 0.593, + 0.317 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.327, + 0.895, + 0.432 + ], + "angle": 0, + "content": "In this section, we provide a preliminary introduction to the notation and typical formulation of DNN quantization in Sec. 3.1, theoretically analyze the potential of reducing quantization error with pre-sampled additive noisy bias in Sec. 3.2, and formulate NoisyQuant, a noisy bias-enhanced post-training activation quantization method with reduced quantization error in Sec. 3.3." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.441, + 0.63, + 0.457 + ], + "angle": 0, + "content": "3.1. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.464, + 0.895, + 0.525 + ], + "angle": 0, + "content": "To start with, we recap the post-training quantization process on DNN models. Given the dominance of fully-connected (FC) linear projection layers in transformer models, here we focus our discussion on the FC layer." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.525, + 0.894, + 0.569 + ], + "angle": 0, + "content": "The computation performed in a FC layer with weight \\( W \\in \\mathbb{R}^{k \\times m} \\), input activation \\( X \\in \\mathbb{R}^{m \\times n} \\), and bias \\( B \\in \\mathbb{R}^{k \\times 1} \\) can be formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.628, + 0.581, + 0.892, + 0.597 + ], + "angle": 0, + "content": "\\[\nf (X) = W X + B. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.607, + 0.894, + 0.728 + ], + "angle": 0, + "content": "The main computation burden in Eq. (1) lies in the matrix multiplication of \\( WX \\), which requires a total of \\( k \\times m \\times n \\) multiply-accumulate computations (MACs). Post-training quantization aims to reduce the computation cost of matrix multiplication by converting \\( X \\) and \\( W \\) to fixed-point quantized values, therefore replacing floating-point MACs with much cheaper fixed-point operations [16, 40]. The quantized FC layer can therefore be represented as" + }, + { + "type": "equation", + "bbox": [ + 0.593, + 0.739, + 0.892, + 0.755 + ], + "angle": 0, + "content": "\\[\nf _ {q} (X) = q _ {W} (W) q _ {A} (X) + B, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.895, + 0.901 + ], + "angle": 0, + "content": "where \\( q_{W}(\\cdot) \\) and \\( q_{A}(\\cdot) \\) denotes the quantizer function of weight and activation respectively. Previous research has observed that the heave-tailed activation distribution of \\( X \\) in transformer is causing significant quantization error between \\( X \\) and \\( q_{A}(X) \\) at low precision, leading to significant performance degradation. Previous PTQ methods modify the design of quantizer \\( q_{A}(\\cdot) \\) to mitigate the quantization error. In this paper, we propose an alternative approach, where we modify the distribution of activation \\( X \\) with a" + }, + { + "type": "footer", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20323" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.123 + ], + "angle": 0, + "content": "pre-sampled noisy bias before quantization, and prove it to be a plug-and-play enhancement on any quantizer design." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.131, + 0.436, + 0.148 + ], + "angle": 0, + "content": "3.2. Theoretical analysis on quantization error" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.154, + 0.473, + 0.321 + ], + "angle": 0, + "content": "As previous research mainly blames the heavy-tailed distribution of transformer activation as the main cause for large quantization errors [23, 42], in this section we analyze how the quantization error change as we actively alter the activation distribution being quantized. A straightforward way to alter the input activation distribution is to add a fixed \"Noisy Bias\" sampled from a Uniform random distribution. Denoting the input activation as \\( X \\) and the Noisy Bias as \\( N \\), where \\( N \\) and \\( X \\) have the same dimensions, here we formulate the quantization error difference between \\( X \\) and \\( X + N \\) for a quantizer \\( Q \\) in Eq. (3)" + }, + { + "type": "equation", + "bbox": [ + 0.113, + 0.33, + 0.47, + 0.367 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} D (X, N) = Q E (X + N) - Q E (X) = \\\\ \\left| \\left| \\left(Q (X + N) - X - N\\right) \\right| \\right| _ {2} ^ {2} - \\left| \\left| \\left(Q (X) - X\\right) \\right| \\right| _ {2} ^ {2}. \\tag {3} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.377, + 0.469, + 0.407 + ], + "angle": 0, + "content": "Theorem 1 states the condition where \\( D(x, N) \\leq 0 \\) holds for each histogram snapshot \\( x \\) of activation \\( X \\)." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.416, + 0.47, + 0.492 + ], + "angle": 0, + "content": "Theorem 1. Consider a given quantizer \\( Q \\) with quantization bin width \\( 2b \\). For each histogram snapshot of \\( X \\) where all elements \\( X_{i} \\) have the same distance \\( x \\) away from the center of the quantization bin, and for a Noisy Bias \\( N \\) sampled from \\( N \\sim \\mathcal{U}(-n,n) \\) where \\( x \\leq n \\leq 2b - x \\), we have" + }, + { + "type": "equation", + "bbox": [ + 0.128, + 0.502, + 0.47, + 0.537 + ], + "angle": 0, + "content": "\\[\nD (x, N) \\leq 0 \\text {i f} 0 \\leq x \\leq n \\left(1 - \\sqrt {\\frac {n}{3 b}}\\right). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.544, + 0.47, + 0.589 + ], + "angle": 0, + "content": "Proof. Without loss of generality, we consider the quantization bin of quantizer \\( Q \\) which \\( x \\) falls into to be centered at the 0 point, therefore \\( 0 \\leq x \\leq b \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.589, + 0.469, + 0.62 + ], + "angle": 0, + "content": "In this way, \\( Q(X_{i}) = b \\) for all elements in the histogram snapshot, leading to a quantization error" + }, + { + "type": "equation", + "bbox": [ + 0.202, + 0.629, + 0.47, + 0.647 + ], + "angle": 0, + "content": "\\[\nQ E (X) = (b - x) ^ {2}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.656, + 0.473, + 0.793 + ], + "angle": 0, + "content": "Consider adding \\(N \\sim \\mathcal{U}(-n, n)\\). \\(X + N\\) will follow \\(\\mathcal{U}(x - n, x + n)\\). In the case of \\(x \\leq n \\leq 2b - x\\), \\(Q(X_i + N_i) = b\\) if \\(N_i \\in [-x, n]\\), and \\(Q(X_i + N_i) = -b\\) if \\(N_i \\in [-n, -x)\\) for all elements in \\(X + N\\). Though we only sample a single instance of \\(N\\) from the Uniform distribution, given the large number of elements in the histogram snapshot, the empirical quantization error \\(QE(X + N)\\) can be estimated with an expectation over all \\(N \\sim \\mathcal{U}(-n, n)\\) following the Weak Law of Large Numbers, as" + }, + { + "type": "equation", + "bbox": [ + 0.113, + 0.801, + 0.47, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {E} _ {N} [ Q E (X + N) ] = \\frac {1}{2 n} \\int_ {x - n} ^ {x + n} Q E (x + z) d z \\\\ = \\frac {1}{2 n} \\left[ \\int_ {x - n} ^ {0} (z + b) ^ {2} d z + \\int_ {0} ^ {x + n} (z - b) ^ {2} d z \\right] \\tag {6} \\\\ = x ^ {2} - \\frac {b}{n} x ^ {2} + \\frac {n}{3} - n b + b ^ {2}. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.092, + 0.788, + 0.107 + ], + "angle": 0, + "content": "Combining Eq. (5) and Eq. (6), we have" + }, + { + "type": "equation", + "bbox": [ + 0.555, + 0.117, + 0.892, + 0.169 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} D (x, N) = \\mathbf {E} _ {N} [ Q E (X + N) ] - Q E (X) \\\\ = - \\frac {b}{n} x ^ {2} + 2 b x + \\frac {n ^ {2}}{3} - n b. \\tag {7} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.177, + 0.809, + 0.192 + ], + "angle": 0, + "content": "We verify this derivation empirically in Sec. 4." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.192, + 0.894, + 0.253 + ], + "angle": 0, + "content": "It can be observed that given \\( b \\) and \\( n \\), \\( D(x, N) \\) is a quadratic function with respect to the activation value \\( x \\). The inequality \\( D(x, N) \\leq 0, 0 \\leq x \\leq n \\) can be easily solved with respect to \\( x \\) as" + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.264, + 0.892, + 0.299 + ], + "angle": 0, + "content": "\\[\n0 \\leq x \\leq n \\left(1 - \\sqrt {\\frac {n}{3 b}}\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.307, + 0.892, + 0.323 + ], + "angle": 0, + "content": "which is always feasible given both \\( b \\) and \\( n \\) are positive." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.332, + 0.894, + 0.392 + ], + "angle": 0, + "content": "Theorem 1 indicates that adding Noisy Bias can always reduce the quantization error of elements close to the center of the quantization bin, which is the source of large quantization error in a heavy-tailed distribution." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.393, + 0.894, + 0.483 + ], + "angle": 0, + "content": "In practice, to choose a proper \\( n \\) for a layer in a pretrained model, we can first acquire the empirical distribution of activation \\( X \\) by passing a small amount of calibration data through the pretrained model. With the distribution of \\( X \\), we can estimate the expected quantization error reduction of the layer numerically as a function of \\( n \\)" + }, + { + "type": "equation", + "bbox": [ + 0.616, + 0.494, + 0.892, + 0.527 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} (n) = \\sum_ {x \\in X} [ D (x, N) ]. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.537, + 0.894, + 0.612 + ], + "angle": 0, + "content": "Note that \\( x \\) here denotes the distance between each activation value and the corresponding quantization bin center. Though we cannot find a closed-form solution for the minima of \\( \\mathcal{L}(n) \\), we can easily find a good-enough choice with a linear search over a small range of \\( n \\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.621, + 0.824, + 0.638 + ], + "angle": 0, + "content": "3.3. NoisyQuant formulation and pipeline" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Given our theoretical analysis on reducing quantization error with Noisy Bias, we propose to apply the Noisy Bias as an enhancement to the PTQ methods of vision transformer models, which we name as \"NoisyQuant\". The pipeline of NoisyQuant is visualized in Fig. 1. Specifically, before performing PTQ and deploying the model, we sample a single Noisy Bias \\( N \\in \\mathbb{R}^{m \\times 1} \\) for each layer from a Uniform distribution, and fix it throughout the inference. The range of \\( N \\) is determined by the calibration objective defined in Eq. (9). At inference time, we add the Noisy Bias \\( N \\) to the input activation \\( X \\) under the broadcasting rule before going through the activation quantizer, therefore actively altering the distribution being quantized to reduce the input quantization error. After the activation-weight multiplication in the linear layer, we remove the impact of \\( N \\) by adjusting the bias term in the linear layer with a denoising bias computed from \\( N \\), therefore retrieving the correct output. Formally," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20324" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.122 + ], + "angle": 0, + "content": "NoisyQuant converts the quantized FC computation defined in Eq. (2) into" + }, + { + "type": "equation", + "bbox": [ + 0.085, + 0.132, + 0.47, + 0.15 + ], + "angle": 0, + "content": "\\[\nf _ {N q} (X) = q _ {W} (W) q _ {A} (X + N) + \\left(B - q _ {W} (W) N\\right). \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.159, + 0.469, + 0.31 + ], + "angle": 0, + "content": "Since \\( N \\) is sampled before the deployment of the model and fixed it throughout the inference, the denoising output bias \\( B' = B - q_W(W)N \\) only needs to be computed once before the deployment, without any computation overhead in the inference process. The only computation overhead brought by NoisyQuant at inference time is the on-the-fly summation of \\( X + N \\), which is negligible comparing to the cost of the matrix multiplications. Both \\( B' \\) and \\( N \\) shall be stored with higher precision. In practice, we store both variables as INT16 to enable integer-only inference." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.31, + 0.469, + 0.356 + ], + "angle": 0, + "content": "Comparing to the output of the floating-point FC layer defined in Eq. (1), the output quantization error induced by the activation quantization defined in Eq. (2) is" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.365, + 0.47, + 0.396 + ], + "angle": 0, + "content": "\\[\nQ E _ {O} (X) = \\left\\| f _ {q} (X) - f (X) \\right\\| _ {2} ^ {2} = \\left\\| W [ Q (X) - X ] \\right\\| _ {2} ^ {2}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.397, + 0.47, + 0.457 + ], + "angle": 0, + "content": "where \\(Q(\\cdot)\\) is short of \\(q_{A}(\\cdot)\\), and we omit the weight quantization \\(q_{W}(\\cdot)\\) as we are focusing on the activation quantization. Similarly, the output quantization error after the application of NoisyQuant can be computed as" + }, + { + "type": "equation", + "bbox": [ + 0.113, + 0.466, + 0.47, + 0.504 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} Q E _ {O} ^ {\\prime} (X) = | | f _ {N q} (X) - f (X) | | _ {2} ^ {2} \\\\ = \\left\\| W [ Q (X + N) - X - N ] \\right\\| _ {2} ^ {2}. \\tag {12} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.514, + 0.469, + 0.589 + ], + "angle": 0, + "content": "As we prove in Theorem 1 that the Noisy Bias enables \\( X + N \\) to have a lower quantization error than \\( X \\), we can achieve lower output quantization error in Eq. (12) than in Eq. (11) when NoisyQuant is applied, therefore improving the PTQ performance of the model." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.602, + 0.358, + 0.619 + ], + "angle": 0, + "content": "4. Theoretical insight verification" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.627, + 0.47, + 0.718 + ], + "angle": 0, + "content": "In this section, we provide empirical verification of our theoretical analysis on reducing the activation quantization error with NoisyQuant. We numerically verify Theorem 1 with simulated data in Sec. 4.1, and demonstrate the reduction in both input and output quantization errors under the true inference-time activation distribution in Sec. 4.2." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.727, + 0.388, + 0.743 + ], + "angle": 0, + "content": "4.1. Empirical verification of Theorem 1" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Here we follow the settings in Theorem 1 to empirically verify its theoretical derivation. Specifically, we set the quantization bin range \\( b = 1 \\), and explore how the quantization error difference induced by the Noisy Bias change with different choice of activation value \\( x \\) and noisy bias range \\( n \\). For all empirical results we experiment with 10 instances of independently sampled Noisy Bias \\( N \\), and report the mean and standard derivation of \\( D(X, N) \\) defined in Eq. (3) across the 10 instances. We consider input activation \\( X \\) to be a tensor with 20 dimensions. Given the tens to hundreds" + }, + { + "type": "image", + "bbox": [ + 0.525, + 0.098, + 0.868, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.238, + 0.894, + 0.267 + ], + "angle": 0, + "content": "Figure 2. Verifying Eq. (7) with respect to Noisy Bias range \\( n \\). Circle indicates the mean and error bar the std for evaluated results." + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.275, + 0.871, + 0.42 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.432, + 0.892, + 0.459 + ], + "angle": 0, + "content": "Figure 3. Verifying Eq. (7) with respect to activation value \\( x \\). Circle indicates the mean and error bar the std for evaluated results." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.474, + 0.892, + 0.564 + ], + "angle": 0, + "content": "of thousands of activation values in each transformer layer, it is likely to see more than 20 activation elements taking the same value. As we base our theoretical derivation on the Weak Law of Large Numbers, having more elements taking the same value will lead to less variance than the simulation results provided in this section." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.565, + 0.893, + 0.686 + ], + "angle": 0, + "content": "For the first experiment, we fix all elements in \\( X \\) to take value \\( x = 0.1 \\), and alter \\( n \\) in the range of [0.1, 1.9]. We compare the empirical \\( D(X, N) \\) and the theoretical result derived in Eq. (7) in Fig. 2. The evaluated results closely follow the theoretical line. Even with only 20 activation elements the standard deviation across independently sampled Noisy Bias is much smaller than the quantization error benefit brought by adding the Noisy Bias." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.686, + 0.892, + 0.807 + ], + "angle": 0, + "content": "Similarly, we fix \\( n = 1.4 \\) and alter activation value \\( x \\) in the range of [0, 0.6]. The results are shown in Fig. 3, where we see the same trend of close mean and small variance as in the previous experiment. Quantization error can be reduced by Noisy Bias when \\( x < 0.44 \\), which follows the numerical computation result of Eq. (4). Both results show the correctness and stability of our derivation of Eq. (7) and the proof of Theorem 1." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.817, + 0.857, + 0.833 + ], + "angle": 0, + "content": "4.2. Quantization errors in transformer layers" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.895, + 0.901 + ], + "angle": 0, + "content": "As mentioned in Sec. 3.2, NoisyQuant enables the reduction of both input activation quantization error \\((QE)\\) and output error \\((QE_O)\\) of a linear layer given a quantizer. In this section, we verify the error reduction on the real acti" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20325" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.471, + 0.173 + ], + "angle": 0, + "content": "Table 1. Averaged quantization error on different types of layers. The calculation is performed with 5,120 images randomly selected from the ImageNet validation set on the Swin-T model. Layer names are defined as same as in Timm [35]. We compute the input quantization error difference \\( D \\) and compare the output errors between NoisyQuant and EasyQuant." + }, + { + "type": "table", + "bbox": [ + 0.113, + 0.184, + 0.437, + 0.293 + ], + "angle": 0, + "content": "
Layer NameEq. (3) D(X)Eq. (11) QEO(X)Eq. (12) QE'(X)Output QE drop
qkv-0.0290.1070.09610%
proj-0.1060.0930.08113%
fc1-0.5270.1260.1232%
fc2-5.1301.0490.85219%
" + }, + { + "type": "image", + "bbox": [ + 0.116, + 0.31, + 0.429, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.444, + 0.47, + 0.486 + ], + "angle": 0, + "content": "Figure 4. Quantization error histogram of fc2 input activation. We apply NoisyQuant on top of the EasyQuant quantizer, and compare the quantization error distribution (red) with EasyQuant (grey)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.506, + 0.471, + 0.702 + ], + "angle": 0, + "content": "vation distribution of vision transformer model. Practically, we run evaluation on randomly selected 5120 images of validation set with linear PTQ quantizer EasyQuant [36] as baseline, and calculate \\( QE \\) and \\( QE_{O} \\) with or without adding Noisy Bias. Experiments are ran on different types of layers, namely qkv, proj, fc1, and fc2, respectively. The results are averaged across all layers of each type. As shown in Tab. 1, input quantization error difference \\( D(X) \\) defined in Eq. (3) is consistently lower than 0, which meets our expectation that Noisy Bias reduces input quantization error. This further leads to significant drop in output error between the quantized and floating-point layers for all layer types. Up to \\( 19\\% \\) averaged output error drop can be achieved in fc2 layers." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.471, + 0.903 + ], + "angle": 0, + "content": "To further understand the impact NoisyQuant makes, we visualize both the input and output distribution of each layer. We start with highlighting fc2 as it achieves the greatest QE reduction. The fc2 layer takes in the output of GELU activations, which has significantly complicated distribution as introduced in Sec. 1. Fig. 4 visualizes the distribution of input quantization error with or without NoisyQuant. With EasyQuant only, significant amount of activation elements are showing large quantization error, which is most probably owing to bad quantization on densely distributed areas. Instead, NoisyQuant generally makes QE distributed plainly and for most elements near zero, therefore leading to significant QE reduction overall. Subsequently, we plot the" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.088, + 0.892, + 0.175 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.597, + 0.178, + 0.796, + 0.19 + ], + "angle": 0, + "content": "(a) \\(26\\%\\) \\(QE_{O}\\) reduction on qkv layer" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.19, + 0.891, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.596, + 0.279, + 0.797, + 0.292 + ], + "angle": 0, + "content": "(b) \\(25\\%\\) \\(QE_{O}\\) reduction on proj layer" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.293, + 0.891, + 0.379 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.598, + 0.382, + 0.795, + 0.394 + ], + "angle": 0, + "content": "(c) \\(36\\%\\) \\(QE_{O}\\) reduction on fc1 layer" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.395, + 0.891, + 0.481 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.598, + 0.484, + 0.795, + 0.496 + ], + "angle": 0, + "content": "(d) \\(60\\%\\) \\(QE_{O}\\) reduction on fc2 layer" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.499, + 0.891, + 0.542 + ], + "angle": 0, + "content": "Figure 5. Output histogram in selected transformer layers before and after activation quantization. NoisyQuant significantly reduces the output quantization error over EasyQuant baselines." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.556, + 0.894, + 0.632 + ], + "angle": 0, + "content": "histogram of selected layers' outputs in Fig. 5. NoisyQuant output consistently follows the floating-point layer output distribution closer than that of EasyQuant, especially at values with large magnitudes. This leads to up to \\(60\\%\\) error reduction in the output of some vision transformer layers." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.646, + 0.633, + 0.663 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.671, + 0.894, + 0.807 + ], + "angle": 0, + "content": "This section evaluates the impact of NoisyQuant on the PTQ performance of some commonly used vision transformer architectures. First, we evaluate the performance on image classification tasks with the ViT [10], DeiT [30] and Swin [22] transformer models. We also evaluate the performance of object detection with the DETR model [4]. Finally, we provide ablation study on how NoisyQuant improves the PTQ performance when added to each type of layers in the vision transformer model." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.817, + 0.704, + 0.833 + ], + "angle": 0, + "content": "5.1. Experimental settings" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Dataset. The classification tasks are implemented on the ImageNet-2012 classification dataset [28], with 1.2 million training images and 50,000 validation images in 1000 classes. For object detection task, the MSCOCO 2017 dataset [20]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20326" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.473, + 0.215 + ], + "angle": 0, + "content": "Table 2. PTQ accuracy of ViT models. Top-1 accuracy on ImageNet is reported in the table. W/A represents weight and activation bit-width, respectively. The default patch size is \\(16 \\times 16\\) and the default input resolution is \\(224 \\times 224\\). “*” after model name indicates \\(384 \\times 384\\) input. “MP” represents mixed-precision. For our methods, “NoisyQuant-Linear” is implemented on top of the EasyQuant quantizer, whereas “NoisyQuant-PTQ4ViT” is implemented on the nonlinear PTQ4ViT quantizer. All NoisyQuant results are shown with mean±STD of 20 repeated experiments." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.226, + 0.457, + 0.449 + ], + "angle": 0, + "content": "
MethodW/AViT-SViT-BViT-B*
Pretrained32/3281.3984.5486.00
Percentile [18]6/667.7477.6377.60
Liu et al. [23]6 MP-75.26-
PTQ4ViT-Linear [43]6/670.2475.6646.88
EasyQuant [36]6/675.1381.4282.02
NoisyQuant-Linear6/676.86±0.0681.90±0.1183.00±0.09
PTQ4ViT [43]6/678.6381.6583.34
NoisyQuant-PTQ4ViT6/678.65±0.0782.32±0.0983.22±0.10
Percentile [18]8/878.7780.1282.53
Liu et al. [23]8 MP-76.98-
FQ-ViT [21]8/8-83.31-
PTQ4ViT-Linear [43]8/880.4683.8985.35
EasyQuant [36]8/880.7583.8985.53
NoisyQuant-Linear8/880.81±0.0184.10±0.0385.56±0.01
PTQ4ViT [43]8/881.0084.2585.82
NoisyQuant-PTQ4ViT8/881.15±0.0284.22±0.0185.86±0.01
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.465, + 0.47, + 0.494 + ], + "angle": 0, + "content": "Table 3. PTQ accuracy of DeiT models. Abbreviations are the same as Tab. 2." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.505, + 0.456, + 0.777 + ], + "angle": 0, + "content": "
PretrainedW/ADeiT-SDeiT-BDeiT-B*
Pretrained32/3279.8581.8083.11
Percentile [18]6/670.4973.9978.24
Bit-Split [32]6/674.5876.39-
Liu et al. [23]6/674.5877.02-
Liu et al. [23]6 MP75.1077.47-
PTQ4VIT-Linear [43]6/672.2678.7868.44
EasyQuant [36]6/675.2779.4781.26
NoisyQuant-Linear6/676.37±0.1379.77±0.0881.40±0.06
PTQ4VIT [43]6/676.2880.2581.55
NoisyQuant-PTQ4ViT6/677.43±0.0880.70±0.1181.65±0.04
Percentile [18]8/873.9875.2180.02
Bit-Split [32]8/876.3979.42-
Liu et al. [23]8/877.4780.48-
Liu et al. [23]8 MP78.0981.29-
FQ-ViT [21]8/879.1781.20-
PTQ4VIT-Linear [43]8/877.6580.9482.33
EasyQuant [36]8/878.9881.1982.10
NoisyQuant-Linear8/879.11±0.0281.30±0.0382.23±0.02
PTQ4VIT [43]8/879.4781.4882.97
NoisyQuant-PTQ4ViT8/879.51±0.0181.45±0.0282.49±0.02
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.793, + 0.472, + 0.869 + ], + "angle": 0, + "content": "is utilized to evaluate the PTQ performance, which contains 118,000 training images and 5,000 validation images. For both tasks we randomly sample 1024 images from the training set as the calibration data for all PTQ methods implemented in our experiments." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Pretrained model architecture. We perform uniform 6-bit and 8-bit PTQ on different variants of pretrained vision" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.089, + 0.892, + 0.118 + ], + "angle": 0, + "content": "Table 4. PTQ accuracy of Swin models. Abbreviations are the same as Tab. 2." + }, + { + "type": "table", + "bbox": [ + 0.507, + 0.129, + 0.89, + 0.314 + ], + "angle": 0, + "content": "
MethodW/ASwin-TSwin-SSwin-BSwin-B*
Pretrained32/3281.3983.2385.2786.44
Percentile [18]6/677.7580.4181.9082.94
PTQ4VIT-Linear [43]6/678.4581.7483.3585.22
EasyQuant [36]6/679.5182.4584.3085.89
NoisyQuant-Linear6/680.01±0.0682.78±0.0484.57±0.0485.96±0.04
PTQ4VIT [43]6/680.4782.3884.0185.38
NoisyQuant-PTQ4ViT6/680.51±0.0382.86±0.0584.68±0.0686.03±0.07
Percentile [18]8/879.8880.9383.0884.31
FQ-ViT [21]8/880.5182.71--
PTQ4VIT [43]8/880.9682.7584.7986.16
EasyQuant [36]8/880.9583.0085.1086.39
NoisyQuant-Linear8/881.05±0.0383.07±0.0385.11±0.0486.42±0.02
PTQ4VIT [43]8/881.2483.1085.1486.39
NoisyQuant-PTQ4ViT8/881.25±0.0283.13±0.0185.20±0.0386.44±0.01
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.319, + 0.895, + 0.456 + ], + "angle": 0, + "content": "transformer models. For classification tasks we quantize the model family of ViT [10], DeiT [30], and Swin [22] provided by the Timm library [35], including large-scale models with \\(384 \\times 384\\) input resolution (marked with “*”). For detection, we perform PTQ on the official implementation of DETR [4] with ResNet-50 backbone. We use the pretrained model checkpoint provided by the official source of each model, whose floating-point performances are reported in the results tables and match their original papers." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.457, + 0.895, + 0.76 + ], + "angle": 0, + "content": "Implementation Details. Following the mainstream quantization methods [23, 43], we quantize all the weights and inputs involved in matrix multiplication. More specifically, we quantize weights of qkv-projectors, attention output projectors, MLPs, linear embeddings, and model heads. Besides, we also quantize input activations of linear layers and matrix multiplication operators. Following [23, 26, 43, 45], we keep layer normalization and softmax operations as full precision. Apart from special notes, we perform symmetric layer-wise quantization for activations and symmetric channel-wise quantization for weights. In particular, weights are quantized by absolute MinMax values without clamping. We implement NoisyQuant on top of linear quantizer EasyQuant [36] and nonlinear quantizer PTQ4ViT [43]. After we apply the selected quantizer to determine the quantization bin width and location on calibration data, we decide the parameters of the Noisy Bias distribution through a linear search. We use the empirical activation quantization error as defined in Eq. (9) as the search objective, where we find the Noisy Bias parameter that can minimize the quantization error." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.771, + 0.733, + 0.789 + ], + "angle": 0, + "content": "5.2. NoisyQuant performance" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Here we compare NoisyQuant with state-of-the-art PTQ methods, including percentile [18], bit-split [32], Liu et al. [23], FQ-ViT [21], EasyQuant [36], and PTQ4ViT [43], on ImageNet. PTQ performance of different variants of ViT, DeiT, and Swin transformer models are provided in Tab. 2, Tab. 3, and Tab. 4, respectively. Experiments are conducted for both linear and nonlinear quantizers." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20327" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.47, + 0.118 + ], + "angle": 0, + "content": "Table 5. PTQ accuracy of DETR models. Mean average precision (mAP) evaluated on the MSCOCO 2017 dataset is reported." + }, + { + "type": "table", + "bbox": [ + 0.126, + 0.129, + 0.428, + 0.228 + ], + "angle": 0, + "content": "
ModelMethodW/AmAP
DETR [4] (COCO2017)Pretrained32/3242.0
Percentile [18]8/838.6
Bit-Split [32]8/840.6
Liu et al. [23]8/841.2
EasyQuant [36]8/841.1
NoisyQuant-Linear8/841.4±0.05
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.283, + 0.473, + 0.616 + ], + "angle": 0, + "content": "Linear quantizer. Previous linear PTQ methods suffer from severe performance degradation on vision transformer activation quantization. EasyQuant [36] achieves the best performance among linear quantizers, yet still induce 2-6% of accuracy drop under 6-bit quantization compared to the floating-point baseline. By applying the proposed NoisyQuant method on top of the EasyQuant quantizer, the resulted NoisyQuant-Linear quantizer consistently and significantly improves the PTQ accuracy on all variants of vision transformer models. Under the W6A6 setting, NoisyQuant-Linear achieves performance improvement of 1.73% on ViT-S, 0.98% on ViT-B*, 1.1% on DeiT-S, and 0.5% on SWIN-T over EasyQuant. Under the W8A8 setting, as the performance gap between EasyQuant and floating-point gets smaller, NoisyQuant still appears beneficial in further improving the PTQ performance. It is worth noting that although nonlinear quantizer like PTQ4ViT [43] consistently achieves higher PTQ performance than the linear EasyQuant quantizer, the enhancement brought by NoisyQuant enables the NoisyQuant-Linear quantizer to achieve on-par or even better performance than the nonlinear PTQ4ViT, with much lower overhead in hardware deployment." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.63, + 0.472, + 0.797 + ], + "angle": 0, + "content": "Nonlinear quantizer. As we claim NoisyQuant as a quantizer-agnostic enhancement of PTQ, we further implement NoisyQuant on top of the nonlinear quantizer PTQ4ViT [43], namely NoisyQuant-PTQ4ViT. NoisyQuant-PTQ4ViT outperforms most results of PTQ4ViT, notably achieving for a \\(1.25\\%\\) improvement on DeiT-S, \\(0.67\\%\\) on ViT-S, and \\(0.67\\%\\) on Swin-B in the W6A6 setting. The W6A6 PTQ performance of Swin-S and Swin-B* for the first time improved to within \\(0.5\\%\\) of the floating-point baseline by the NoisyQuant-PTQ4ViT quantizer, which can never be achieved without the help of NoisyQuant." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.472, + 0.902 + ], + "angle": 0, + "content": "Besides classification models, we report the PTQ performance of the DETR object detection model in Tab. 5. NoisyQuant implemented on top of the EasyQuant quantizer also outperforms the EasyQuant baseline, and all previous linear PTQ quantizers including percentile [18], bit-split [32], and Liu et al. [23]." + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.089, + 0.895, + 0.133 + ], + "angle": 0, + "content": "Table 6. Applying NoisyQuant on different layer types. Starting from W6A6 PTQ with the EasyQuant quantizer, we add NoisyQuant to each type of layers in the transformer model." + }, + { + "type": "table", + "bbox": [ + 0.546, + 0.143, + 0.849, + 0.348 + ], + "angle": 0, + "content": "
Modelqkv noiseproj noisefc1 noisefc2 noiseTop-1 W6A6
DeiT-S [30]XXXX75.27
✓XXX75.38
X✓XX75.45
XX✓X75.33
XXX✓76.21
✓✓✓✓76.37
Swin-T [22]XXXX79.51
✓XXX79.53
X✓XX79.56
XX✓X79.52
XXX✓79.80
✓✓✓✓80.01
" + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.362, + 0.88, + 0.38 + ], + "angle": 0, + "content": "5.3. NoisyQuant's impact on different layer types" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.389, + 0.895, + 0.63 + ], + "angle": 0, + "content": "As we show the effectiveness of NoisyQuant in improving the PTQ performance of the entire model, here we explore if NoisyQuant is helpful for all types of layers in the vision transformer model. Tab. 6 exhibits the effect of applying NoisyQuant on different types of layers. The checkmark in the table means that we apply NoisyQuant to the input activation of all the layers of that particular type, otherwise the layer is quantized with EasyQuant only. For all layer types, applying NoisyQuant consistently improves the PTQ performance of the entire model. Specifically, 'fc2' layers derive the greatest benefit from NoisyQuant, which corresponds to our previous analysis that the activation distribution after GELU function brings the major challenge to PTQ, which can be resolved with the proposed NoisyQuant method. After adding NoisyQuant to the input activation of all linear layers, the model achieves the maximum performance boost." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.652, + 0.628, + 0.668 + ], + "angle": 0, + "content": "6. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.678, + 0.895, + 0.845 + ], + "angle": 0, + "content": "This work proposes NoisyQuant, a noisy bias-enhanced post-training activation quantization method for the complicated activation distribution of vision transformer models. We theoretically prove the quantization error reduction brought by adding noisy bias to the input activation before quantization, and empirically show the effectiveness of NoisyQuant on both linear and nonlinear quantizers on vision transformer models. We hope this work opens up a new direction to reduce PTQ quantization error through actively changing the distribution being quantized, and inspires better Noisy Bias distribution design or generation method based on quantizer and activation characteristics." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.845, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Acknowledgement This work was supported in part by the National Key Research and Development Program of China under Grant 2022YFB4400900. The Berkeley team acknowledges support from Berkeley Deep Drive, Intel Corporation, and Panasonic." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20328" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.473, + 0.199 + ], + "angle": 0, + "content": "[1] Yash Bhalgat, Jinwon Lee, Markus Nagel, Tijmen Blankevoort, and Nojun Kwak. Lsq+: Improving low-bit quantization through learnable offsets and better initialization. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 2978-2985, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.201, + 0.472, + 0.272 + ], + "angle": 0, + "content": "[2] Yaohui Cai, Zhewei Yao, Zhen Dong, Amir Gholami, Michael W. Mahoney, and Kurt Keutzer. Zeroq: A novel zero shot quantization framework. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13166-13175, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.273, + 0.472, + 0.342 + ], + "angle": 0, + "content": "[3] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.344, + 0.472, + 0.399 + ], + "angle": 0, + "content": "[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. ArXiv, abs/2005.12872, 2020. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.402, + 0.472, + 0.457 + ], + "angle": 0, + "content": "[5] Minghao Chen, Houwen Peng, Jianlong Fu, and Haibin Ling. Autoformer: Searching transformers for visual recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12270-12280, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.459, + 0.472, + 0.514 + ], + "angle": 0, + "content": "[6] Tianlong Chen, Yu Cheng, Zhe Gan, Lu Yuan, Lei Zhang, and Zhangyang Wang. Chasing sparsity in vision transformers: An end-to-end exploration. Advances in Neural Information Processing Systems, 34:19974-19988, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.517, + 0.472, + 0.584 + ], + "angle": 0, + "content": "[7] Jungwook Choi, Zhuo Wang, Swagath Venkataramani, Pierce I-Jen Chuang, Vijayalakshmi Srinivasan, and Kailash Gopalakrishnan. Pact: Parameterized clipping activation for quantized neural networks. arXiv preprint arXiv:1805.06085, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.588, + 0.472, + 0.657 + ], + "angle": 0, + "content": "[8] Zhen Dong, Zhewei Yao, Daiyaan Arfeen, Amir Gholami, Michael W Mahoney, and Kurt Keutzer. Hawq-v2: Hessian aware trace-weighted quantization of neural networks. Advances in neural information processing systems, 33:18518-18529, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.659, + 0.472, + 0.729 + ], + "angle": 0, + "content": "[9] Zhen Dong, Zhewei Yao, Amir Gholami, Michael W Mahoney, and Kurt Keutzer. Hawq: Hessian aware quantization of neural networks with mixed-precision. In Proceedings of the IEEE International Conference on Computer Vision, pages 293–302, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.731, + 0.472, + 0.813 + ], + "angle": 0, + "content": "[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.816, + 0.472, + 0.871 + ], + "angle": 0, + "content": "[11] Amir Gholami, Sehoon Kim, Zhen Dong, Zhewei Yao, Michael W Mahoney, and Kurt Keutzer. A survey of quantization methods for efficient neural network inference. arXiv preprint arXiv:2103.13630, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.873, + 0.472, + 0.902 + ], + "angle": 0, + "content": "[12] Cong Guo, Yuxian Qiu, Jingwen Leng, Xiaotian Gao, Chen Zhang, Yunxin Liu, Fan Yang, Yuhao Zhu, and Minyi Guo." + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.116, + 0.473, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "Squant: On-the-fly data-free quantization via diagonal hessian approximation. arXiv preprint arXiv:2202.07471, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.894, + 0.176 + ], + "angle": 0, + "content": "[13] Song Han, Huizi Mao, and William J Dally. Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. arXiv preprint arXiv:1510.00149, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.22 + ], + "angle": 0, + "content": "[14] Ali Hatamizadeh, Hongxu Yin, Jan Kautz, and Pavlo Molchanov. Global context vision transformers. arXiv preprint arXiv:2206.09959, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.222, + 0.892, + 0.248 + ], + "angle": 0, + "content": "[15] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.25, + 0.892, + 0.277 + ], + "angle": 0, + "content": "[16] Mark Horowitz. 1.1 computing's energy problem (and what we can do about it). In ISSCC, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.279, + 0.894, + 0.331 + ], + "angle": 0, + "content": "[17] Kwonjoon Lee, Huiwen Chang, Lu Jiang, Han Zhang, Zhuowen Tu, and Ce Liu. Vitgan: Training gans with vision transformers. arXiv preprint arXiv:2107.04589, 2021.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.335, + 0.894, + 0.39 + ], + "angle": 0, + "content": "[18] Rundong Li, Yan Wang, Feng Liang, Hongwei Qin, Junjie Yan, and Rui Fan. Fully quantized network for object detection. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2805-2814, 2019. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.392, + 0.894, + 0.445 + ], + "angle": 0, + "content": "[19] Zhikai Li, Junrui Xiao, Lianwei Yang, and Qingyi Gu. Repqvit: Scale reparameterization for post-training quantization of vision transformers. arXiv preprint arXiv:2212.08254, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.448, + 0.892, + 0.503 + ], + "angle": 0, + "content": "[20] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dolkar, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.505, + 0.892, + 0.545 + ], + "angle": 0, + "content": "[21] Yang Lin, Tianyu Zhang, Peiqin Sun, Zheng Li, and Shuchang Zhou. Fq-vit: Post-training quantization for fully quantized vision transformer. In IJCAI, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.548, + 0.892, + 0.617 + ], + "angle": 0, + "content": "[22] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 9992-10002, 2021. 1, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.619, + 0.892, + 0.659 + ], + "angle": 0, + "content": "[23] Zhenhua Liu, Yunhe Wang, Kai Han, Siwei Ma, and Wen Gao. Post-training quantization for vision transformer. In NeurIPS, 2021. 1, 2, 3, 4, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.661, + 0.892, + 0.715 + ], + "angle": 0, + "content": "[24] Zhenhua Liu, Yunhe Wang, Kai Han, Wei Zhang, Siwei Ma, and Wen Gao. Post-training quantization for vision transformer. Advances in Neural Information Processing Systems, 34:28092-28103, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.718, + 0.892, + 0.758 + ], + "angle": 0, + "content": "[25] A. Polino, R. Pascanu, and D. Alistarh. Model compression via distillation and quantization. arXiv preprint arXiv:1802.05668, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.76, + 0.892, + 0.801 + ], + "angle": 0, + "content": "[26] Gabriele Prato, Ella Charlaix, and Mehdi Rezagholizadeh. Fully quantized transformer for improved translation. ArXiv, abs/1910.10485, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.803, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[27] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[28] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy," + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20329" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.472, + 0.147 + ], + "angle": 0, + "content": "Aditya Khosla, Michael S. Bernstein, Alexander C. Berg, and Li Fei-Fei. Imagenet large scale visual recognition challenge. International Journal of Computer Vision, 115:211-252, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.472, + 0.218 + ], + "angle": 0, + "content": "[29] Sheng Shen, Zhen Dong, Jiayu Ye, Linjian Ma, Zhewei Yao, Amir Gholami, Michael W Mahoney, and Kurt Keutzer. Qbert: Hessian based ultra low precision quantization of bert. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 8815-8821, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.219, + 0.472, + 0.273 + ], + "angle": 0, + "content": "[30] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Herv'e J'egou. Training data-efficient image transformers & distillation through attention. In ICML, 2021. 1, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.274, + 0.472, + 0.33 + ], + "angle": 0, + "content": "[31] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.33, + 0.472, + 0.371 + ], + "angle": 0, + "content": "[32] Peisong Wang, Qiang Chen, Xiangyu He, and Jian Cheng. Towards accurate post-training network quantization via bitsplit and stitching. In ICML, 2020. 1, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.372, + 0.472, + 0.439 + ], + "angle": 0, + "content": "[33] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. arXiv preprint arXiv:2102.12122, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.441, + 0.472, + 0.51 + ], + "angle": 0, + "content": "[34] Xiuying Wei, Yunchen Zhang, Xiangguo Zhang, Ruihao Gong, Shanghang Zhang, Qi Zhang, Fengwei Yu, and Xianglong Liu. Outlier suppression: Pushing the limit of low-bit transformer language models. arXiv preprint arXiv:2209.13325, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.511, + 0.472, + 0.552 + ], + "angle": 0, + "content": "[35] Ross Wightman. Pytorch image models. https://github.com/rwrightman/pytorch-image-models, 2019.6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.553, + 0.472, + 0.607 + ], + "angle": 0, + "content": "[36] Di Wu, Qi Tang, Yongle Zhao, Ming Zhang, Ying Fu, and Debing Zhang. Easyquant: Post-training quantization via scale optimization. arXiv preprint arXiv:2006.16669, 2020. 1, 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.608, + 0.472, + 0.664 + ], + "angle": 0, + "content": "[37] Lirui Xiao, Huanrui Yang, Zhen Dong, Kurt Keutzer, Li Du, and Shanghang Zhang. Csq: Growing mixed-precision quantization scheme with bi-level continuous sparsification. arXiv preprint arXiv:2212.02770, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.665, + 0.472, + 0.706 + ], + "angle": 0, + "content": "[38] Huanrui Yang, Lin Duan, Yiran Chen, and Hai Li. Bsq: Exploring bit-level sparsity for mixed-precision neural network quantization. arXiv preprint arXiv:2102.10462, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.707, + 0.472, + 0.761 + ], + "angle": 0, + "content": "[39] Huanrui Yang, Hongxu Yin, Pavlo Molchanov, Hai Li, and Jan Kautz. Nvit: Vision transformer compression and parameter redistribution. arXiv preprint arXiv:2110.04869, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.762, + 0.472, + 0.831 + ], + "angle": 0, + "content": "[40] Zhewei Yao, Zhen Dong, Zhangcheng Zheng, Amir Gholami, Jiali Yu, Eric Tan, Leyuan Wang, Qijing Huang, Yida Wang, Michael Mahoney, et al. Hawq-v3: Dyadic neural network quantization. In International Conference on Machine Learning, pages 11875-11886. PMLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.472, + 0.899 + ], + "angle": 0, + "content": "[41] Li Yuan, Yunpeng Chen, Tao Wang, Weihao Yu, Yujun Shi, Zihang Jiang, Francis EH Tay, Jiashi Feng, and Shuicheng Yan. Tokens-to-token vit: Training vision transformers from scratch onImagenet. arXiv preprint arXiv:2101.11986, 2021.3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.147 + ], + "angle": 0, + "content": "[42] Zhihang Yuan, Chenhao Xue, Yiqi Chen, Qiang Wu, and Guangyu Sun. Ptg4vit: Post-training quantization framework for vision transformers. arXiv preprint arXiv:2111.12293, 2021. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.894, + 0.204 + ], + "angle": 0, + "content": "[43] Zhihang Yuan, Chenhao Xue, Yi-Qiang Chen, Qiang Wu, and Guangyu Sun. Ptq4vit: Post-training quantization for vision transformers with twin uniform quantization. In ECCV, 2022. 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.206, + 0.894, + 0.274 + ], + "angle": 0, + "content": "[44] Dongqing Zhang, Jiaolong Yang, Dongqiangzi Ye, and Gang Hua. Lq-nets: Learned quantization for highly accurate and compact deep neural networks. In Proceedings of the European conference on computer vision (ECCV), pages 365–382, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.276, + 0.893, + 0.317 + ], + "angle": 0, + "content": "[45] Wei Zhang, Lu Hou, Yichun Yin, Lifeng Shang, Xiao Chen, Xin Jiang, and Qun Liu. Ternarybert: Distillation-aware ultra-low bit bert. In EMNLP, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.319, + 0.893, + 0.387 + ], + "angle": 0, + "content": "[46] Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xiang, Philip H.S. Torr, and Li Zhang. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In CVPR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.389, + 0.894, + 0.445 + ], + "angle": 0, + "content": "[47] Shuchang Zhou, Yuxin Wu, Zekun Ni, Xinyu Zhou, He Wen, and Yuheng Zou. Dorefa-net: Training low bitwidth convolutional neural networks with low bitwidth gradients. arXiv preprint arXiv:1606.06160, 2016. 1" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20330" + } + ] +] \ No newline at end of file diff --git a/2023/NoisyQuant_ Noisy Bias-Enhanced Post-Training Activation Quantization for Vision Transformers/68808ca4-cf24-4585-b35b-3e2acef953ea_origin.pdf b/2023/NoisyQuant_ Noisy Bias-Enhanced Post-Training Activation Quantization for Vision Transformers/68808ca4-cf24-4585-b35b-3e2acef953ea_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..48c5e44822c63c91fc4e971eec368d4cc5663bae --- /dev/null +++ b/2023/NoisyQuant_ Noisy Bias-Enhanced Post-Training Activation Quantization for Vision Transformers/68808ca4-cf24-4585-b35b-3e2acef953ea_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:41fb0616b4e31530d449aee0c1fdfe5f2994fef4a18059ab7c5d864e9ae8c7f1 +size 3240022 diff --git a/2023/NoisyQuant_ Noisy Bias-Enhanced Post-Training Activation Quantization for Vision Transformers/full.md b/2023/NoisyQuant_ Noisy Bias-Enhanced Post-Training Activation Quantization for Vision Transformers/full.md new file mode 100644 index 0000000000000000000000000000000000000000..34680eb23268d9173942d5529c5c31d312545a7a --- /dev/null +++ b/2023/NoisyQuant_ Noisy Bias-Enhanced Post-Training Activation Quantization for Vision Transformers/full.md @@ -0,0 +1,319 @@ +# NoisyQuant: Noisy Bias-Enhanced Post-Training Activation Quantization for Vision Transformers + +Yijiang Liu\*1, Huanrui Yang\*2, Zhen Dong?, Kurt Keutzer?, Li Du1, Shanghang Zhang3Nanjing University, 2University of California, Berkeley + +$^{3}$ National Key Laboratory for Multimedia Information Processing, School of Computer Science, Peking University liuyijiang@smail.nju.edu.cn, {huanrui, zhendong, keutzer}@berkeley.edu ldu@nju.edu.cn, shanghang@pku.edu.cn + +# Abstract + +The complicated architecture and high training cost of vision transformers urge the exploration of post-training quantization. However, the heavy-tailed distribution of vision transformer activations hinders the effectiveness of previous post-training quantization methods, even with advanced quantizer designs. Instead of tuning the quantizer to better fit the complicated activation distribution, this paper proposes NoisyQuant, a quantizer-agnostic enhancement for the post-training activation quantization performance of vision transformers. We make a surprising theoretical discovery that for a given quantizer, adding a fixed Uniform noisy bias to the values being quantized can significantly reduce the quantization error under provable conditions. Building on the theoretical insight, NoisyQuant achieves the first success on actively altering the heavy-tailed activation distribution with additive noisy bias to fit a given quantizer. Extensive experiments show NoisyQuant largely improves the post-training quantization performance of vision transformer with minimal computation overhead. For instance, on linear uniform 6-bit activation quantization, NoisyQuant improves SOTA top-1 accuracy on ImageNet by up to $1.7\%$ , $1.1\%$ and $0.5\%$ for ViT, DeiT, and Swin Transformer respectively, achieving on-par or even higher performance than previous nonlinear, mixed-precision quantization. + +# 1. Introduction + +Inspired by the success of Self Attention (SA)-based transformer models in Natural Language Processing (NLP) tasks [31], recent researches make significant progress in applying transformer models to the field of computer vision [3, 10, 22, 30]. In the meantime, the typical design of + +transformer models induces large model sizes, high computational consumption, and long training time. For instance, the widely used DeiT-Base model [30] contains 86M parameters, logs 18G floating-point operations for a single input, and requires 300 epochs of training on the ImageNet dataset. This leads to significant difficulties in hardware deployment. In facing such difficulty, a number of compression and acceleration methods are applied to vision transformer models, including pruning [6, 39], quantization [24, 42], and neural architecture search [5], etc. + +Among these methods, quantization appears as one of the most effective and widely applied ways [11]. Quantization process uses a predefined "quantizer" function to convert the continuous representation of weights and activations into a small number of discrete symbols, therefore enabling low-precision representations for straightforward memory savings. For DNN models, the approximation error made by the quantizer inevitably leads to performance drop. A series of work focuses on Quantization-Aware Training (QAT) that finetunes the quantized model at low precision [8, 9, 25, 37, 47]. However, given the high training cost and the complicated computation graph of vision transformer models, retraining the model at low precision could be costly and unstable [42]. Alternatively, Post-Training Quantization (PTQ) is preferable for vision transformers as it eliminates the need for re-training or finetuning the quantized model, instead only adjusts the design of the quantizer based on the full-precision pretrained model and a small set of sampled calibration data [1, 2, 23, 32, 36]. For example, linear quantizers [7, 29, 36, 38] reduce quantization error by shifting, clipping, and scaling the values to be quantized. Nonlinear quantizers [13, 42, 44] further adjust the width and location of each quantization bin to better fit the distribution. + +Unfortunately, though progresses are made in designing better PTQ quantizers, it still appears to be significantly challenging to quantize vision transformer models, especially for activation quantization. Transformer layers produce up + +![](images/74c4599864a12bc52e56972b056b846857fbe93ef3957a5256dc9eb750e9bdaa.jpg) +Figure 1. Overview of the NoisyQuant pipeline (orange box) with comparison to EasyQuant [36] (gray box). Histograms of input activations, quantized activations, and layer outputs are illustrated in the pipeline. Starting from the input activation $X$ following a GELU function, where the green histogram is plotted in linear-scale and the gray in log-scale, NoisyQuant adds a fixed NoisyBias $N$ sampled from a Uniform distribution onto $X$ . Adding the noise before quantization flattens the peaks in the activation distribution to make it more friendly to quantization, as visualized in the linear/log-scaled histogram of $X + N$ , and theoretically proved in Sec. 3.2. The noise is removed from the result of the fixed-point activation-weight multiplication with a denoising bias term, which we formulate in Sec. 3.3. The rightmost sub-figures illustrate the layer output distribution of EasyQuant and NoisyQuant. Compared to the output of the floating-point model (yellow), NoisyQuant output follows the distribution closely and achieves up to $60\%$ output error reduction on some transformer layers, as shown in Sec. 4. The resulting PTQ performance improvement of the entire model is provided in Sec. 5. + +to millions of activation values with sophisticated distributions. For instance, outputs of the GELU function [15] are asymmetrically distributed, with spikes in the histogram at some values and a long tail in a large range (see top-left of Fig. 1). Some linear projection layers also lead to significantly large activation values that are sparsely distributed in a very long tail [23]. Consequently, low-precision PTQ on vision transformer suffers from performance degradation, even utilizing non-linear or mixed-precision quantizers at the cost of additional data communication and computation overheads [23,43]. No linear uniform PTQ method achieves good performance on vision transformer models. + +This paper provides a brand new perspective in dealing with the complicated vision transformer activation distribution. Instead of adding more tricks in the quantizer design to fit the activation distribution, this work explores the potential of actively and cost-effectively altering the distribution being quantized, making it more friendly to a given quantizer. The pipeline of our proposed method is illustrated in Fig. 1. Specifically, we make a surprising discovery that for any quantizer, the quantization error can be significantly reduced by adding a fixed noisy bias sampled from a Uniform distribution to the activation before quantization. We theoretically prove the condition when the quantization error reduction can be achieved. On this basis, we propose NoisyQuant, a plug-and-play, quantizer-agnostic enhancement on the posttraining activation quantization of vision transformer models. For each layer, we sample a Noisy Bias based on the input activation distribution following our theoretical analysis, and + +compute the corresponding denoising bias to retain the correct output of the layer. At inference time, the Noisy Bias is added to the input activation before quantization, and the denoising bias is removed from the layer output. This process significantly reduces the quantization error with minimal computation overhead. NoisyQuant leads to significant improvement in the PTQ performance of state-of-the-art vision transformers. Applying NoisyQuant on top of a uniform linear quantization achieves on-par performance to SOTA mixed-precision, nonlinear PTQ methods [23, 42]. Adding NoisyQuant on top of these nonlinear quantizers achieves further performance gain. + +To the best of our knowledge, this paper makes the following novel contributions: + +- Theoretically shows the possibility and proves feasible conditions for reducing the quantization error of heavily-tailed distributions with a fixed additive noisy bias; +- Proposes NoisyQuant, a quantizer-agnostic enhancement for PTQ performance on activation quantization. NoisyQuant achieves the first success in actively refining the distribution being quantized to reduce quantization error following the theoretical results on additive noisy bias, with minimal computation overhead; +- Demonstrates consistent performance improvement by applying NoisyQuant on top of existing PTQ quantizers. For 6-bit PTQ, NoisyQuant improves the ImageNet top-1 accuracy of uniform linear quantized vision transformer models by up to $1.7\%$ , and improves SOTA + +nonlinear quantizer PTQ4ViT [43] by up to $0.7\%$ . + +# 2. Related work + +# 2.1. Transformer models in computer vision + +Since Self-Attention (SA) [31]-based transformer models have shown impressive performance in Natural Language Processing tasks, researchers transfer the idea into computer vision. Pioneering work on Vision Transformer [10] for the first time builds a totally SA-based model in the field of computer vision and shows its effectiveness. Since then, a variety of transformer-based vision models have emerged with boosted performance, claiming state-of-the-art in multiple computer vision tasks. Notable advances include additional data augmentation and distillation [30], incorporating multi-stage architecture [22, 33], and explore novel attention design [14,22,41]. Transformer models also advance to tasks beyond classification, such as object detection [3], semantic segmentation [46], and image generation [17]. Furthermore, the development of vision transformer brings unification of model architecture and representation across image and text, enabling stronger vision-language connection via image-text-pair contrastive learning [27]. + +The rapid development of vision transformer models motivates the exploration of quantization, as quantization brings straightforward memory and computation savings agnostic to model architectures. Compared to CNN models, the use of advanced activation functions like GELU [15] and more complicated computation graphs like the attention mechanism makes vision transformers more sensitive to quantization [43]. This work aims to resolve the difficulty of posttraining vision transformer activation quantization caused by the heavy-tailed activation distribution, which we believe will facilitate the deployment of state-of-the-art transformer models into real-world computer vision applications. + +# 2.2. Post-training quantization methods + +Post-training quantization (PTQ) is a preferable quantization method in scenarios with limited training data or limited computational resources for the expensive quantization-aware fine-tuning process. Previous methods have dedicatedly studied PTQ on CNNs. For instance, EasyQuant [36] presents a quick searching mechanism to determine the proper clipping range for quantization. ZeroQ [2] proposes a distillation mechanism to generate proxy input images, which can leverage inherent statistics of batch normalization layers to conduct PTQ. SQuant [12] calibrates the model for lower quantization error on the basis of the sensitivity obtained by the Hessian spectrum. On vision transformers, Liu et al. [23] presents a linear PTQ method that uses Pearson correlation coefficients and ranking loss to determine scaling factors. PTQ4ViT [43] introduces a nonlinear Twin Uniform Quantization based on vision transformers' acti + +vation distributions, which set different scaling factors for 1) positive and negative activations of GeLU, and 2) small and large values of Softmax, which can reduce the activation quantization error to some extent with the cost of additional computation overhead. [19, 34] reparameterize the LN layer to suppress outliers by scaling down activation values. + +Unlike previous PTQ methods that analyze the activation distribution and fit the quantizer accordingly, our method takes a novel perspective of actively modifying the distribution being quantized with a sampled noisy bias. We show our proposed method can bring quantization error reduction and performance enhancement for all the PTQ quantizers for vision transformer activation quantization. + +# 3. Method + +In this section, we provide a preliminary introduction to the notation and typical formulation of DNN quantization in Sec. 3.1, theoretically analyze the potential of reducing quantization error with pre-sampled additive noisy bias in Sec. 3.2, and formulate NoisyQuant, a noisy bias-enhanced post-training activation quantization method with reduced quantization error in Sec. 3.3. + +# 3.1. Preliminary + +To start with, we recap the post-training quantization process on DNN models. Given the dominance of fully-connected (FC) linear projection layers in transformer models, here we focus our discussion on the FC layer. + +The computation performed in a FC layer with weight $W \in \mathbb{R}^{k \times m}$ , input activation $X \in \mathbb{R}^{m \times n}$ , and bias $B \in \mathbb{R}^{k \times 1}$ can be formulated as + +$$ +f (X) = W X + B. \tag {1} +$$ + +The main computation burden in Eq. (1) lies in the matrix multiplication of $WX$ , which requires a total of $k \times m \times n$ multiply-accumulate computations (MACs). Post-training quantization aims to reduce the computation cost of matrix multiplication by converting $X$ and $W$ to fixed-point quantized values, therefore replacing floating-point MACs with much cheaper fixed-point operations [16, 40]. The quantized FC layer can therefore be represented as + +$$ +f _ {q} (X) = q _ {W} (W) q _ {A} (X) + B, \tag {2} +$$ + +where $q_{W}(\cdot)$ and $q_{A}(\cdot)$ denotes the quantizer function of weight and activation respectively. Previous research has observed that the heave-tailed activation distribution of $X$ in transformer is causing significant quantization error between $X$ and $q_{A}(X)$ at low precision, leading to significant performance degradation. Previous PTQ methods modify the design of quantizer $q_{A}(\cdot)$ to mitigate the quantization error. In this paper, we propose an alternative approach, where we modify the distribution of activation $X$ with a + +pre-sampled noisy bias before quantization, and prove it to be a plug-and-play enhancement on any quantizer design. + +# 3.2. Theoretical analysis on quantization error + +As previous research mainly blames the heavy-tailed distribution of transformer activation as the main cause for large quantization errors [23, 42], in this section we analyze how the quantization error change as we actively alter the activation distribution being quantized. A straightforward way to alter the input activation distribution is to add a fixed "Noisy Bias" sampled from a Uniform random distribution. Denoting the input activation as $X$ and the Noisy Bias as $N$ , where $N$ and $X$ have the same dimensions, here we formulate the quantization error difference between $X$ and $X + N$ for a quantizer $Q$ in Eq. (3) + +$$ +\begin{array}{l} D (X, N) = Q E (X + N) - Q E (X) = \\ \left| \left| \left(Q (X + N) - X - N\right) \right| \right| _ {2} ^ {2} - \left| \left| \left(Q (X) - X\right) \right| \right| _ {2} ^ {2}. \tag {3} \\ \end{array} +$$ + +Theorem 1 states the condition where $D(x, N) \leq 0$ holds for each histogram snapshot $x$ of activation $X$ . + +Theorem 1. Consider a given quantizer $Q$ with quantization bin width $2b$ . For each histogram snapshot of $X$ where all elements $X_{i}$ have the same distance $x$ away from the center of the quantization bin, and for a Noisy Bias $N$ sampled from $N \sim \mathcal{U}(-n,n)$ where $x \leq n \leq 2b - x$ , we have + +$$ +D (x, N) \leq 0 \text {i f} 0 \leq x \leq n \left(1 - \sqrt {\frac {n}{3 b}}\right). \tag {4} +$$ + +Proof. Without loss of generality, we consider the quantization bin of quantizer $Q$ which $x$ falls into to be centered at the 0 point, therefore $0 \leq x \leq b$ . + +In this way, $Q(X_{i}) = b$ for all elements in the histogram snapshot, leading to a quantization error + +$$ +Q E (X) = (b - x) ^ {2}. \tag {5} +$$ + +Consider adding $N \sim \mathcal{U}(-n, n)$ . $X + N$ will follow $\mathcal{U}(x - n, x + n)$ . In the case of $x \leq n \leq 2b - x$ , $Q(X_i + N_i) = b$ if $N_i \in [-x, n]$ , and $Q(X_i + N_i) = -b$ if $N_i \in [-n, -x)$ for all elements in $X + N$ . Though we only sample a single instance of $N$ from the Uniform distribution, given the large number of elements in the histogram snapshot, the empirical quantization error $QE(X + N)$ can be estimated with an expectation over all $N \sim \mathcal{U}(-n, n)$ following the Weak Law of Large Numbers, as + +$$ +\begin{array}{l} \mathbf {E} _ {N} [ Q E (X + N) ] = \frac {1}{2 n} \int_ {x - n} ^ {x + n} Q E (x + z) d z \\ = \frac {1}{2 n} \left[ \int_ {x - n} ^ {0} (z + b) ^ {2} d z + \int_ {0} ^ {x + n} (z - b) ^ {2} d z \right] \tag {6} \\ = x ^ {2} - \frac {b}{n} x ^ {2} + \frac {n}{3} - n b + b ^ {2}. \\ \end{array} +$$ + +Combining Eq. (5) and Eq. (6), we have + +$$ +\begin{array}{l} D (x, N) = \mathbf {E} _ {N} [ Q E (X + N) ] - Q E (X) \\ = - \frac {b}{n} x ^ {2} + 2 b x + \frac {n ^ {2}}{3} - n b. \tag {7} \\ \end{array} +$$ + +We verify this derivation empirically in Sec. 4. + +It can be observed that given $b$ and $n$ , $D(x, N)$ is a quadratic function with respect to the activation value $x$ . The inequality $D(x, N) \leq 0, 0 \leq x \leq n$ can be easily solved with respect to $x$ as + +$$ +0 \leq x \leq n \left(1 - \sqrt {\frac {n}{3 b}}\right), \tag {8} +$$ + +which is always feasible given both $b$ and $n$ are positive. + +Theorem 1 indicates that adding Noisy Bias can always reduce the quantization error of elements close to the center of the quantization bin, which is the source of large quantization error in a heavy-tailed distribution. + +In practice, to choose a proper $n$ for a layer in a pretrained model, we can first acquire the empirical distribution of activation $X$ by passing a small amount of calibration data through the pretrained model. With the distribution of $X$ , we can estimate the expected quantization error reduction of the layer numerically as a function of $n$ + +$$ +\mathcal {L} (n) = \sum_ {x \in X} [ D (x, N) ]. \tag {9} +$$ + +Note that $x$ here denotes the distance between each activation value and the corresponding quantization bin center. Though we cannot find a closed-form solution for the minima of $\mathcal{L}(n)$ , we can easily find a good-enough choice with a linear search over a small range of $n$ . + +# 3.3. NoisyQuant formulation and pipeline + +Given our theoretical analysis on reducing quantization error with Noisy Bias, we propose to apply the Noisy Bias as an enhancement to the PTQ methods of vision transformer models, which we name as "NoisyQuant". The pipeline of NoisyQuant is visualized in Fig. 1. Specifically, before performing PTQ and deploying the model, we sample a single Noisy Bias $N \in \mathbb{R}^{m \times 1}$ for each layer from a Uniform distribution, and fix it throughout the inference. The range of $N$ is determined by the calibration objective defined in Eq. (9). At inference time, we add the Noisy Bias $N$ to the input activation $X$ under the broadcasting rule before going through the activation quantizer, therefore actively altering the distribution being quantized to reduce the input quantization error. After the activation-weight multiplication in the linear layer, we remove the impact of $N$ by adjusting the bias term in the linear layer with a denoising bias computed from $N$ , therefore retrieving the correct output. Formally, + +NoisyQuant converts the quantized FC computation defined in Eq. (2) into + +$$ +f _ {N q} (X) = q _ {W} (W) q _ {A} (X + N) + \left(B - q _ {W} (W) N\right). \tag {10} +$$ + +Since $N$ is sampled before the deployment of the model and fixed it throughout the inference, the denoising output bias $B' = B - q_W(W)N$ only needs to be computed once before the deployment, without any computation overhead in the inference process. The only computation overhead brought by NoisyQuant at inference time is the on-the-fly summation of $X + N$ , which is negligible comparing to the cost of the matrix multiplications. Both $B'$ and $N$ shall be stored with higher precision. In practice, we store both variables as INT16 to enable integer-only inference. + +Comparing to the output of the floating-point FC layer defined in Eq. (1), the output quantization error induced by the activation quantization defined in Eq. (2) is + +$$ +Q E _ {O} (X) = \left\| f _ {q} (X) - f (X) \right\| _ {2} ^ {2} = \left\| W [ Q (X) - X ] \right\| _ {2} ^ {2}, \tag {11} +$$ + +where $Q(\cdot)$ is short of $q_{A}(\cdot)$ , and we omit the weight quantization $q_{W}(\cdot)$ as we are focusing on the activation quantization. Similarly, the output quantization error after the application of NoisyQuant can be computed as + +$$ +\begin{array}{l} Q E _ {O} ^ {\prime} (X) = | | f _ {N q} (X) - f (X) | | _ {2} ^ {2} \\ = \left\| W [ Q (X + N) - X - N ] \right\| _ {2} ^ {2}. \tag {12} \\ \end{array} +$$ + +As we prove in Theorem 1 that the Noisy Bias enables $X + N$ to have a lower quantization error than $X$ , we can achieve lower output quantization error in Eq. (12) than in Eq. (11) when NoisyQuant is applied, therefore improving the PTQ performance of the model. + +# 4. Theoretical insight verification + +In this section, we provide empirical verification of our theoretical analysis on reducing the activation quantization error with NoisyQuant. We numerically verify Theorem 1 with simulated data in Sec. 4.1, and demonstrate the reduction in both input and output quantization errors under the true inference-time activation distribution in Sec. 4.2. + +# 4.1. Empirical verification of Theorem 1 + +Here we follow the settings in Theorem 1 to empirically verify its theoretical derivation. Specifically, we set the quantization bin range $b = 1$ , and explore how the quantization error difference induced by the Noisy Bias change with different choice of activation value $x$ and noisy bias range $n$ . For all empirical results we experiment with 10 instances of independently sampled Noisy Bias $N$ , and report the mean and standard derivation of $D(X, N)$ defined in Eq. (3) across the 10 instances. We consider input activation $X$ to be a tensor with 20 dimensions. Given the tens to hundreds + +![](images/5dfc8a3bd58abeb0ef5849f913c9d72b2732a19dcc8bfd0ef13df63ff8b80773.jpg) +Figure 2. Verifying Eq. (7) with respect to Noisy Bias range $n$ . Circle indicates the mean and error bar the std for evaluated results. + +![](images/c4d1d795130efb6e642d1f738f6648bdfd1e708ad0cad0d6f302718bbe515558.jpg) +Figure 3. Verifying Eq. (7) with respect to activation value $x$ . Circle indicates the mean and error bar the std for evaluated results. + +of thousands of activation values in each transformer layer, it is likely to see more than 20 activation elements taking the same value. As we base our theoretical derivation on the Weak Law of Large Numbers, having more elements taking the same value will lead to less variance than the simulation results provided in this section. + +For the first experiment, we fix all elements in $X$ to take value $x = 0.1$ , and alter $n$ in the range of [0.1, 1.9]. We compare the empirical $D(X, N)$ and the theoretical result derived in Eq. (7) in Fig. 2. The evaluated results closely follow the theoretical line. Even with only 20 activation elements the standard deviation across independently sampled Noisy Bias is much smaller than the quantization error benefit brought by adding the Noisy Bias. + +Similarly, we fix $n = 1.4$ and alter activation value $x$ in the range of [0, 0.6]. The results are shown in Fig. 3, where we see the same trend of close mean and small variance as in the previous experiment. Quantization error can be reduced by Noisy Bias when $x < 0.44$ , which follows the numerical computation result of Eq. (4). Both results show the correctness and stability of our derivation of Eq. (7) and the proof of Theorem 1. + +# 4.2. Quantization errors in transformer layers + +As mentioned in Sec. 3.2, NoisyQuant enables the reduction of both input activation quantization error $(QE)$ and output error $(QE_O)$ of a linear layer given a quantizer. In this section, we verify the error reduction on the real acti + +Table 1. Averaged quantization error on different types of layers. The calculation is performed with 5,120 images randomly selected from the ImageNet validation set on the Swin-T model. Layer names are defined as same as in Timm [35]. We compute the input quantization error difference $D$ and compare the output errors between NoisyQuant and EasyQuant. + +
Layer NameEq. (3) D(X)Eq. (11) QEO(X)Eq. (12) QE'(X)Output QE drop
qkv-0.0290.1070.09610%
proj-0.1060.0930.08113%
fc1-0.5270.1260.1232%
fc2-5.1301.0490.85219%
+ +![](images/4c0a4ca95467480a8818a0f4efbb91c0f57851b30811acad0ea47ec99787d011.jpg) +Figure 4. Quantization error histogram of fc2 input activation. We apply NoisyQuant on top of the EasyQuant quantizer, and compare the quantization error distribution (red) with EasyQuant (grey). + +vation distribution of vision transformer model. Practically, we run evaluation on randomly selected 5120 images of validation set with linear PTQ quantizer EasyQuant [36] as baseline, and calculate $QE$ and $QE_{O}$ with or without adding Noisy Bias. Experiments are ran on different types of layers, namely qkv, proj, fc1, and fc2, respectively. The results are averaged across all layers of each type. As shown in Tab. 1, input quantization error difference $D(X)$ defined in Eq. (3) is consistently lower than 0, which meets our expectation that Noisy Bias reduces input quantization error. This further leads to significant drop in output error between the quantized and floating-point layers for all layer types. Up to $19\%$ averaged output error drop can be achieved in fc2 layers. + +To further understand the impact NoisyQuant makes, we visualize both the input and output distribution of each layer. We start with highlighting fc2 as it achieves the greatest QE reduction. The fc2 layer takes in the output of GELU activations, which has significantly complicated distribution as introduced in Sec. 1. Fig. 4 visualizes the distribution of input quantization error with or without NoisyQuant. With EasyQuant only, significant amount of activation elements are showing large quantization error, which is most probably owing to bad quantization on densely distributed areas. Instead, NoisyQuant generally makes QE distributed plainly and for most elements near zero, therefore leading to significant QE reduction overall. Subsequently, we plot the + +![](images/978b7759d6d4bb4eb3539e9a99a82b200347cb51e1dfb99fa21970f16265939b.jpg) + +![](images/716ff5a9eb45fde689295d819729611413002e0751353a84adaa23166f962598.jpg) +(a) $26\%$ $QE_{O}$ reduction on qkv layer +(b) $25\%$ $QE_{O}$ reduction on proj layer + +![](images/87dda685b56e4bca845df8e6e12ebceaa8357c72453f137fc7ba1261466a3c4c.jpg) + +![](images/786681dd849f84ee122486e2fe11d4cdbe44948926dc304f6bcc5e46f425c70b.jpg) +(c) $36\%$ $QE_{O}$ reduction on fc1 layer +(d) $60\%$ $QE_{O}$ reduction on fc2 layer +Figure 5. Output histogram in selected transformer layers before and after activation quantization. NoisyQuant significantly reduces the output quantization error over EasyQuant baselines. + +histogram of selected layers' outputs in Fig. 5. NoisyQuant output consistently follows the floating-point layer output distribution closer than that of EasyQuant, especially at values with large magnitudes. This leads to up to $60\%$ error reduction in the output of some vision transformer layers. + +# 5. Experiments + +This section evaluates the impact of NoisyQuant on the PTQ performance of some commonly used vision transformer architectures. First, we evaluate the performance on image classification tasks with the ViT [10], DeiT [30] and Swin [22] transformer models. We also evaluate the performance of object detection with the DETR model [4]. Finally, we provide ablation study on how NoisyQuant improves the PTQ performance when added to each type of layers in the vision transformer model. + +# 5.1. Experimental settings + +Dataset. The classification tasks are implemented on the ImageNet-2012 classification dataset [28], with 1.2 million training images and 50,000 validation images in 1000 classes. For object detection task, the MSCOCO 2017 dataset [20] + +Table 2. PTQ accuracy of ViT models. Top-1 accuracy on ImageNet is reported in the table. W/A represents weight and activation bit-width, respectively. The default patch size is $16 \times 16$ and the default input resolution is $224 \times 224$ . “*” after model name indicates $384 \times 384$ input. “MP” represents mixed-precision. For our methods, “NoisyQuant-Linear” is implemented on top of the EasyQuant quantizer, whereas “NoisyQuant-PTQ4ViT” is implemented on the nonlinear PTQ4ViT quantizer. All NoisyQuant results are shown with mean±STD of 20 repeated experiments. + +
MethodW/AViT-SViT-BViT-B*
Pretrained32/3281.3984.5486.00
Percentile [18]6/667.7477.6377.60
Liu et al. [23]6 MP-75.26-
PTQ4ViT-Linear [43]6/670.2475.6646.88
EasyQuant [36]6/675.1381.4282.02
NoisyQuant-Linear6/676.86±0.0681.90±0.1183.00±0.09
PTQ4ViT [43]6/678.6381.6583.34
NoisyQuant-PTQ4ViT6/678.65±0.0782.32±0.0983.22±0.10
Percentile [18]8/878.7780.1282.53
Liu et al. [23]8 MP-76.98-
FQ-ViT [21]8/8-83.31-
PTQ4ViT-Linear [43]8/880.4683.8985.35
EasyQuant [36]8/880.7583.8985.53
NoisyQuant-Linear8/880.81±0.0184.10±0.0385.56±0.01
PTQ4ViT [43]8/881.0084.2585.82
NoisyQuant-PTQ4ViT8/881.15±0.0284.22±0.0185.86±0.01
+ +Table 3. PTQ accuracy of DeiT models. Abbreviations are the same as Tab. 2. + +
PretrainedW/ADeiT-SDeiT-BDeiT-B*
Pretrained32/3279.8581.8083.11
Percentile [18]6/670.4973.9978.24
Bit-Split [32]6/674.5876.39-
Liu et al. [23]6/674.5877.02-
Liu et al. [23]6 MP75.1077.47-
PTQ4VIT-Linear [43]6/672.2678.7868.44
EasyQuant [36]6/675.2779.4781.26
NoisyQuant-Linear6/676.37±0.1379.77±0.0881.40±0.06
PTQ4VIT [43]6/676.2880.2581.55
NoisyQuant-PTQ4ViT6/677.43±0.0880.70±0.1181.65±0.04
Percentile [18]8/873.9875.2180.02
Bit-Split [32]8/876.3979.42-
Liu et al. [23]8/877.4780.48-
Liu et al. [23]8 MP78.0981.29-
FQ-ViT [21]8/879.1781.20-
PTQ4VIT-Linear [43]8/877.6580.9482.33
EasyQuant [36]8/878.9881.1982.10
NoisyQuant-Linear8/879.11±0.0281.30±0.0382.23±0.02
PTQ4VIT [43]8/879.4781.4882.97
NoisyQuant-PTQ4ViT8/879.51±0.0181.45±0.0282.49±0.02
+ +is utilized to evaluate the PTQ performance, which contains 118,000 training images and 5,000 validation images. For both tasks we randomly sample 1024 images from the training set as the calibration data for all PTQ methods implemented in our experiments. + +Pretrained model architecture. We perform uniform 6-bit and 8-bit PTQ on different variants of pretrained vision + +Table 4. PTQ accuracy of Swin models. Abbreviations are the same as Tab. 2. + +
MethodW/ASwin-TSwin-SSwin-BSwin-B*
Pretrained32/3281.3983.2385.2786.44
Percentile [18]6/677.7580.4181.9082.94
PTQ4VIT-Linear [43]6/678.4581.7483.3585.22
EasyQuant [36]6/679.5182.4584.3085.89
NoisyQuant-Linear6/680.01±0.0682.78±0.0484.57±0.0485.96±0.04
PTQ4VIT [43]6/680.4782.3884.0185.38
NoisyQuant-PTQ4ViT6/680.51±0.0382.86±0.0584.68±0.0686.03±0.07
Percentile [18]8/879.8880.9383.0884.31
FQ-ViT [21]8/880.5182.71--
PTQ4VIT [43]8/880.9682.7584.7986.16
EasyQuant [36]8/880.9583.0085.1086.39
NoisyQuant-Linear8/881.05±0.0383.07±0.0385.11±0.0486.42±0.02
PTQ4VIT [43]8/881.2483.1085.1486.39
NoisyQuant-PTQ4ViT8/881.25±0.0283.13±0.0185.20±0.0386.44±0.01
+ +transformer models. For classification tasks we quantize the model family of ViT [10], DeiT [30], and Swin [22] provided by the Timm library [35], including large-scale models with $384 \times 384$ input resolution (marked with “*”). For detection, we perform PTQ on the official implementation of DETR [4] with ResNet-50 backbone. We use the pretrained model checkpoint provided by the official source of each model, whose floating-point performances are reported in the results tables and match their original papers. + +Implementation Details. Following the mainstream quantization methods [23, 43], we quantize all the weights and inputs involved in matrix multiplication. More specifically, we quantize weights of qkv-projectors, attention output projectors, MLPs, linear embeddings, and model heads. Besides, we also quantize input activations of linear layers and matrix multiplication operators. Following [23, 26, 43, 45], we keep layer normalization and softmax operations as full precision. Apart from special notes, we perform symmetric layer-wise quantization for activations and symmetric channel-wise quantization for weights. In particular, weights are quantized by absolute MinMax values without clamping. We implement NoisyQuant on top of linear quantizer EasyQuant [36] and nonlinear quantizer PTQ4ViT [43]. After we apply the selected quantizer to determine the quantization bin width and location on calibration data, we decide the parameters of the Noisy Bias distribution through a linear search. We use the empirical activation quantization error as defined in Eq. (9) as the search objective, where we find the Noisy Bias parameter that can minimize the quantization error. + +# 5.2. NoisyQuant performance + +Here we compare NoisyQuant with state-of-the-art PTQ methods, including percentile [18], bit-split [32], Liu et al. [23], FQ-ViT [21], EasyQuant [36], and PTQ4ViT [43], on ImageNet. PTQ performance of different variants of ViT, DeiT, and Swin transformer models are provided in Tab. 2, Tab. 3, and Tab. 4, respectively. Experiments are conducted for both linear and nonlinear quantizers. + +Table 5. PTQ accuracy of DETR models. Mean average precision (mAP) evaluated on the MSCOCO 2017 dataset is reported. + +
ModelMethodW/AmAP
DETR [4] (COCO2017)Pretrained32/3242.0
Percentile [18]8/838.6
Bit-Split [32]8/840.6
Liu et al. [23]8/841.2
EasyQuant [36]8/841.1
NoisyQuant-Linear8/841.4±0.05
+ +Linear quantizer. Previous linear PTQ methods suffer from severe performance degradation on vision transformer activation quantization. EasyQuant [36] achieves the best performance among linear quantizers, yet still induce 2-6% of accuracy drop under 6-bit quantization compared to the floating-point baseline. By applying the proposed NoisyQuant method on top of the EasyQuant quantizer, the resulted NoisyQuant-Linear quantizer consistently and significantly improves the PTQ accuracy on all variants of vision transformer models. Under the W6A6 setting, NoisyQuant-Linear achieves performance improvement of 1.73% on ViT-S, 0.98% on ViT-B*, 1.1% on DeiT-S, and 0.5% on SWIN-T over EasyQuant. Under the W8A8 setting, as the performance gap between EasyQuant and floating-point gets smaller, NoisyQuant still appears beneficial in further improving the PTQ performance. It is worth noting that although nonlinear quantizer like PTQ4ViT [43] consistently achieves higher PTQ performance than the linear EasyQuant quantizer, the enhancement brought by NoisyQuant enables the NoisyQuant-Linear quantizer to achieve on-par or even better performance than the nonlinear PTQ4ViT, with much lower overhead in hardware deployment. + +Nonlinear quantizer. As we claim NoisyQuant as a quantizer-agnostic enhancement of PTQ, we further implement NoisyQuant on top of the nonlinear quantizer PTQ4ViT [43], namely NoisyQuant-PTQ4ViT. NoisyQuant-PTQ4ViT outperforms most results of PTQ4ViT, notably achieving for a $1.25\%$ improvement on DeiT-S, $0.67\%$ on ViT-S, and $0.67\%$ on Swin-B in the W6A6 setting. The W6A6 PTQ performance of Swin-S and Swin-B* for the first time improved to within $0.5\%$ of the floating-point baseline by the NoisyQuant-PTQ4ViT quantizer, which can never be achieved without the help of NoisyQuant. + +Besides classification models, we report the PTQ performance of the DETR object detection model in Tab. 5. NoisyQuant implemented on top of the EasyQuant quantizer also outperforms the EasyQuant baseline, and all previous linear PTQ quantizers including percentile [18], bit-split [32], and Liu et al. [23]. + +Table 6. Applying NoisyQuant on different layer types. Starting from W6A6 PTQ with the EasyQuant quantizer, we add NoisyQuant to each type of layers in the transformer model. + +
Modelqkv noiseproj noisefc1 noisefc2 noiseTop-1 W6A6
DeiT-S [30]XXXX75.27
✓XXX75.38
X✓XX75.45
XX✓X75.33
XXX✓76.21
✓✓✓✓76.37
Swin-T [22]XXXX79.51
✓XXX79.53
X✓XX79.56
XX✓X79.52
XXX✓79.80
✓✓✓✓80.01
+ +# 5.3. NoisyQuant's impact on different layer types + +As we show the effectiveness of NoisyQuant in improving the PTQ performance of the entire model, here we explore if NoisyQuant is helpful for all types of layers in the vision transformer model. Tab. 6 exhibits the effect of applying NoisyQuant on different types of layers. The checkmark in the table means that we apply NoisyQuant to the input activation of all the layers of that particular type, otherwise the layer is quantized with EasyQuant only. For all layer types, applying NoisyQuant consistently improves the PTQ performance of the entire model. Specifically, 'fc2' layers derive the greatest benefit from NoisyQuant, which corresponds to our previous analysis that the activation distribution after GELU function brings the major challenge to PTQ, which can be resolved with the proposed NoisyQuant method. After adding NoisyQuant to the input activation of all linear layers, the model achieves the maximum performance boost. + +# 6. Conclusions + +This work proposes NoisyQuant, a noisy bias-enhanced post-training activation quantization method for the complicated activation distribution of vision transformer models. We theoretically prove the quantization error reduction brought by adding noisy bias to the input activation before quantization, and empirically show the effectiveness of NoisyQuant on both linear and nonlinear quantizers on vision transformer models. We hope this work opens up a new direction to reduce PTQ quantization error through actively changing the distribution being quantized, and inspires better Noisy Bias distribution design or generation method based on quantizer and activation characteristics. + +Acknowledgement This work was supported in part by the National Key Research and Development Program of China under Grant 2022YFB4400900. The Berkeley team acknowledges support from Berkeley Deep Drive, Intel Corporation, and Panasonic. + +# References + +[1] Yash Bhalgat, Jinwon Lee, Markus Nagel, Tijmen Blankevoort, and Nojun Kwak. Lsq+: Improving low-bit quantization through learnable offsets and better initialization. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 2978-2985, 2020. 1 +[2] Yaohui Cai, Zhewei Yao, Zhen Dong, Amir Gholami, Michael W. Mahoney, and Kurt Keutzer. Zeroq: A novel zero shot quantization framework. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13166-13175, 2020. 1, 3 +[3] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 1, 3 +[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. ArXiv, abs/2005.12872, 2020. 6, 7, 8 +[5] Minghao Chen, Houwen Peng, Jianlong Fu, and Haibin Ling. Autoformer: Searching transformers for visual recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12270-12280, 2021. 1 +[6] Tianlong Chen, Yu Cheng, Zhe Gan, Lu Yuan, Lei Zhang, and Zhangyang Wang. Chasing sparsity in vision transformers: An end-to-end exploration. Advances in Neural Information Processing Systems, 34:19974-19988, 2021. 1 +[7] Jungwook Choi, Zhuo Wang, Swagath Venkataramani, Pierce I-Jen Chuang, Vijayalakshmi Srinivasan, and Kailash Gopalakrishnan. Pact: Parameterized clipping activation for quantized neural networks. arXiv preprint arXiv:1805.06085, 2018. 1 +[8] Zhen Dong, Zhewei Yao, Daiyaan Arfeen, Amir Gholami, Michael W Mahoney, and Kurt Keutzer. Hawq-v2: Hessian aware trace-weighted quantization of neural networks. Advances in neural information processing systems, 33:18518-18529, 2020. 1 +[9] Zhen Dong, Zhewei Yao, Amir Gholami, Michael W Mahoney, and Kurt Keutzer. Hawq: Hessian aware quantization of neural networks with mixed-precision. In Proceedings of the IEEE International Conference on Computer Vision, pages 293–302, 2019. 1 +[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1, 3, 6, 7 +[11] Amir Gholami, Sehoon Kim, Zhen Dong, Zhewei Yao, Michael W Mahoney, and Kurt Keutzer. A survey of quantization methods for efficient neural network inference. arXiv preprint arXiv:2103.13630, 2021. 1 +[12] Cong Guo, Yuxian Qiu, Jingwen Leng, Xiaotian Gao, Chen Zhang, Yunxin Liu, Fan Yang, Yuhao Zhu, and Minyi Guo. + +Squant: On-the-fly data-free quantization via diagonal hessian approximation. arXiv preprint arXiv:2202.07471, 2022.3 +[13] Song Han, Huizi Mao, and William J Dally. Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. arXiv preprint arXiv:1510.00149, 2015. 1 +[14] Ali Hatamizadeh, Hongxu Yin, Jan Kautz, and Pavlo Molchanov. Global context vision transformers. arXiv preprint arXiv:2206.09959, 2022. 3 +[15] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016. 2, 3 +[16] Mark Horowitz. 1.1 computing's energy problem (and what we can do about it). In ISSCC, 2014. 3 +[17] Kwonjoon Lee, Huiwen Chang, Lu Jiang, Han Zhang, Zhuowen Tu, and Ce Liu. Vitgan: Training gans with vision transformers. arXiv preprint arXiv:2107.04589, 2021.3 +[18] Rundong Li, Yan Wang, Feng Liang, Hongwei Qin, Junjie Yan, and Rui Fan. Fully quantized network for object detection. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2805-2814, 2019. 7, 8 +[19] Zhikai Li, Junrui Xiao, Lianwei Yang, and Qingyi Gu. Repqvit: Scale reparameterization for post-training quantization of vision transformers. arXiv preprint arXiv:2212.08254, 2022.3 +[20] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dolkar, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 6 +[21] Yang Lin, Tianyu Zhang, Peiqin Sun, Zheng Li, and Shuchang Zhou. Fq-vit: Post-training quantization for fully quantized vision transformer. In IJCAI, 2022. 7 +[22] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 9992-10002, 2021. 1, 3, 6, 7, 8 +[23] Zhenhua Liu, Yunhe Wang, Kai Han, Siwei Ma, and Wen Gao. Post-training quantization for vision transformer. In NeurIPS, 2021. 1, 2, 3, 4, 7, 8 +[24] Zhenhua Liu, Yunhe Wang, Kai Han, Wei Zhang, Siwei Ma, and Wen Gao. Post-training quantization for vision transformer. Advances in Neural Information Processing Systems, 34:28092-28103, 2021. 1 +[25] A. Polino, R. Pascanu, and D. Alistarh. Model compression via distillation and quantization. arXiv preprint arXiv:1802.05668, 2018. 1 +[26] Gabriele Prato, Ella Charlaix, and Mehdi Rezagholizadeh. Fully quantized transformer for improved translation. ArXiv, abs/1910.10485, 2019. 7 +[27] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 3 +[28] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, + +Aditya Khosla, Michael S. Bernstein, Alexander C. Berg, and Li Fei-Fei. Imagenet large scale visual recognition challenge. International Journal of Computer Vision, 115:211-252, 2015. 6 +[29] Sheng Shen, Zhen Dong, Jiayu Ye, Linjian Ma, Zhewei Yao, Amir Gholami, Michael W Mahoney, and Kurt Keutzer. Qbert: Hessian based ultra low precision quantization of bert. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 8815-8821, 2020. 1 +[30] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Herv'e J'egou. Training data-efficient image transformers & distillation through attention. In ICML, 2021. 1, 3, 6, 7, 8 +[31] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 1, 3 +[32] Peisong Wang, Qiang Chen, Xiangyu He, and Jian Cheng. Towards accurate post-training network quantization via bitsplit and stitching. In ICML, 2020. 1, 7, 8 +[33] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. arXiv preprint arXiv:2102.12122, 2021. 3 +[34] Xiuying Wei, Yunchen Zhang, Xiangguo Zhang, Ruihao Gong, Shanghang Zhang, Qi Zhang, Fengwei Yu, and Xianglong Liu. Outlier suppression: Pushing the limit of low-bit transformer language models. arXiv preprint arXiv:2209.13325, 2022. 3 +[35] Ross Wightman. Pytorch image models. https://github.com/rwrightman/pytorch-image-models, 2019.6, 7 +[36] Di Wu, Qi Tang, Yongle Zhao, Ming Zhang, Ying Fu, and Debing Zhang. Easyquant: Post-training quantization via scale optimization. arXiv preprint arXiv:2006.16669, 2020. 1, 2, 3, 6, 7, 8 +[37] Lirui Xiao, Huanrui Yang, Zhen Dong, Kurt Keutzer, Li Du, and Shanghang Zhang. Csq: Growing mixed-precision quantization scheme with bi-level continuous sparsification. arXiv preprint arXiv:2212.02770, 2022. 1 +[38] Huanrui Yang, Lin Duan, Yiran Chen, and Hai Li. Bsq: Exploring bit-level sparsity for mixed-precision neural network quantization. arXiv preprint arXiv:2102.10462, 2021. 1 +[39] Huanrui Yang, Hongxu Yin, Pavlo Molchanov, Hai Li, and Jan Kautz. Nvit: Vision transformer compression and parameter redistribution. arXiv preprint arXiv:2110.04869, 2021. 1 +[40] Zhewei Yao, Zhen Dong, Zhangcheng Zheng, Amir Gholami, Jiali Yu, Eric Tan, Leyuan Wang, Qijing Huang, Yida Wang, Michael Mahoney, et al. Hawq-v3: Dyadic neural network quantization. In International Conference on Machine Learning, pages 11875-11886. PMLR, 2021. 3 +[41] Li Yuan, Yunpeng Chen, Tao Wang, Weihao Yu, Yujun Shi, Zihang Jiang, Francis EH Tay, Jiashi Feng, and Shuicheng Yan. Tokens-to-token vit: Training vision transformers from scratch onImagenet. arXiv preprint arXiv:2101.11986, 2021.3 + +[42] Zhihang Yuan, Chenhao Xue, Yiqi Chen, Qiang Wu, and Guangyu Sun. Ptg4vit: Post-training quantization framework for vision transformers. arXiv preprint arXiv:2111.12293, 2021. 1, 2, 4 +[43] Zhihang Yuan, Chenhao Xue, Yi-Qiang Chen, Qiang Wu, and Guangyu Sun. Ptq4vit: Post-training quantization for vision transformers with twin uniform quantization. In ECCV, 2022. 2, 3, 7, 8 +[44] Dongqing Zhang, Jiaolong Yang, Dongqiangzi Ye, and Gang Hua. Lq-nets: Learned quantization for highly accurate and compact deep neural networks. In Proceedings of the European conference on computer vision (ECCV), pages 365–382, 2018. 1 +[45] Wei Zhang, Lu Hou, Yichun Yin, Lifeng Shang, Xiao Chen, Xin Jiang, and Qun Liu. Ternarybert: Distillation-aware ultra-low bit bert. In EMNLP, 2020. 7 +[46] Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xiang, Philip H.S. Torr, and Li Zhang. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In CVPR, 2021. 3 +[47] Shuchang Zhou, Yuxin Wu, Zekun Ni, Xinyu Zhou, He Wen, and Yuheng Zou. Dorefa-net: Training low bitwidth convolutional neural networks with low bitwidth gradients. arXiv preprint arXiv:1606.06160, 2016. 1 \ No newline at end of file diff --git a/2023/NoisyQuant_ Noisy Bias-Enhanced Post-Training Activation Quantization for Vision Transformers/images.zip b/2023/NoisyQuant_ Noisy Bias-Enhanced Post-Training Activation Quantization for Vision Transformers/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..5d505726a49047b3856f09fb6e023bdd88dd8ac3 --- /dev/null +++ b/2023/NoisyQuant_ Noisy Bias-Enhanced Post-Training Activation Quantization for Vision Transformers/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:93e9a904fe0600250445fb589096d7bfcb9c11a2b4dba238db909324336cc536 +size 568761 diff --git a/2023/NoisyQuant_ Noisy Bias-Enhanced Post-Training Activation Quantization for Vision Transformers/layout.json b/2023/NoisyQuant_ Noisy Bias-Enhanced Post-Training Activation Quantization for Vision Transformers/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5bede216fff34b0287622504c17ba754c56a3374 --- /dev/null +++ b/2023/NoisyQuant_ Noisy Bias-Enhanced Post-Training Activation Quantization for Vision Transformers/layout.json @@ -0,0 +1,9129 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 59, + 102, + 536, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 102, + 536, + 138 + ], + "spans": [ + { + "bbox": [ + 59, + 102, + 536, + 138 + ], + "type": "text", + "content": "NoisyQuant: Noisy Bias-Enhanced Post-Training Activation Quantization for Vision Transformers" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 77, + 159, + 518, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 159, + 518, + 190 + ], + "spans": [ + { + "bbox": [ + 77, + 159, + 518, + 190 + ], + "type": "text", + "content": "Yijiang Liu\\*1, Huanrui Yang\\*2, Zhen Dong?, Kurt Keutzer?, Li Du1, Shanghang Zhang3Nanjing University, 2University of California, Berkeley" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 65, + 190, + 527, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 190, + 527, + 232 + ], + "spans": [ + { + "bbox": [ + 65, + 190, + 527, + 232 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 65, + 190, + 527, + 232 + ], + "type": "text", + "content": "National Key Laboratory for Multimedia Information Processing, School of Computer Science, Peking University liuyijiang@smail.nju.edu.cn, {huanrui, zhendong, keutzer}@berkeley.edu ldu@nju.edu.cn, shanghang@pku.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 270, + 192, + 284 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 270, + 192, + 284 + ], + "spans": [ + { + "bbox": [ + 143, + 270, + 192, + 284 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 296, + 290, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 296, + 290, + 583 + ], + "spans": [ + { + "bbox": [ + 46, + 296, + 290, + 583 + ], + "type": "text", + "content": "The complicated architecture and high training cost of vision transformers urge the exploration of post-training quantization. However, the heavy-tailed distribution of vision transformer activations hinders the effectiveness of previous post-training quantization methods, even with advanced quantizer designs. Instead of tuning the quantizer to better fit the complicated activation distribution, this paper proposes NoisyQuant, a quantizer-agnostic enhancement for the post-training activation quantization performance of vision transformers. We make a surprising theoretical discovery that for a given quantizer, adding a fixed Uniform noisy bias to the values being quantized can significantly reduce the quantization error under provable conditions. Building on the theoretical insight, NoisyQuant achieves the first success on actively altering the heavy-tailed activation distribution with additive noisy bias to fit a given quantizer. Extensive experiments show NoisyQuant largely improves the post-training quantization performance of vision transformer with minimal computation overhead. For instance, on linear uniform 6-bit activation quantization, NoisyQuant improves SOTA top-1 accuracy on ImageNet by up to " + }, + { + "bbox": [ + 46, + 296, + 290, + 583 + ], + "type": "inline_equation", + "content": "1.7\\%" + }, + { + "bbox": [ + 46, + 296, + 290, + 583 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 296, + 290, + 583 + ], + "type": "inline_equation", + "content": "1.1\\%" + }, + { + "bbox": [ + 46, + 296, + 290, + 583 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 296, + 290, + 583 + ], + "type": "inline_equation", + "content": "0.5\\%" + }, + { + "bbox": [ + 46, + 296, + 290, + 583 + ], + "type": "text", + "content": " for ViT, DeiT, and Swin Transformer respectively, achieving on-par or even higher performance than previous nonlinear, mixed-precision quantization." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 605, + 128, + 618 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 605, + 128, + 618 + ], + "spans": [ + { + "bbox": [ + 47, + 605, + 128, + 618 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 625, + 288, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 625, + 288, + 687 + ], + "spans": [ + { + "bbox": [ + 46, + 625, + 288, + 687 + ], + "type": "text", + "content": "Inspired by the success of Self Attention (SA)-based transformer models in Natural Language Processing (NLP) tasks [31], recent researches make significant progress in applying transformer models to the field of computer vision [3, 10, 22, 30]. In the meantime, the typical design of" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 260, + 548, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 260, + 548, + 380 + ], + "spans": [ + { + "bbox": [ + 306, + 260, + 548, + 380 + ], + "type": "text", + "content": "transformer models induces large model sizes, high computational consumption, and long training time. For instance, the widely used DeiT-Base model [30] contains 86M parameters, logs 18G floating-point operations for a single input, and requires 300 epochs of training on the ImageNet dataset. This leads to significant difficulties in hardware deployment. In facing such difficulty, a number of compression and acceleration methods are applied to vision transformer models, including pruning [6, 39], quantization [24, 42], and neural architecture search [5], etc." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 385, + 548, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 385, + 548, + 661 + ], + "spans": [ + { + "bbox": [ + 304, + 385, + 548, + 661 + ], + "type": "text", + "content": "Among these methods, quantization appears as one of the most effective and widely applied ways [11]. Quantization process uses a predefined \"quantizer\" function to convert the continuous representation of weights and activations into a small number of discrete symbols, therefore enabling low-precision representations for straightforward memory savings. For DNN models, the approximation error made by the quantizer inevitably leads to performance drop. A series of work focuses on Quantization-Aware Training (QAT) that finetunes the quantized model at low precision [8, 9, 25, 37, 47]. However, given the high training cost and the complicated computation graph of vision transformer models, retraining the model at low precision could be costly and unstable [42]. Alternatively, Post-Training Quantization (PTQ) is preferable for vision transformers as it eliminates the need for re-training or finetuning the quantized model, instead only adjusts the design of the quantizer based on the full-precision pretrained model and a small set of sampled calibration data [1, 2, 23, 32, 36]. For example, linear quantizers [7, 29, 36, 38] reduce quantization error by shifting, clipping, and scaling the values to be quantized. Nonlinear quantizers [13, 42, 44] further adjust the width and location of each quantization bin to better fit the distribution." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 665, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 665, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 665, + 547, + 715 + ], + "type": "text", + "content": "Unfortunately, though progresses are made in designing better PTQ quantizers, it still appears to be significantly challenging to quantize vision transformer models, especially for activation quantization. Transformer layers produce up" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 693, + 147, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 693, + 147, + 713 + ], + "spans": [ + { + "bbox": [ + 61, + 693, + 147, + 713 + ], + "type": "text", + "content": "* Equal contribution. \nCorresponding Author." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "20321" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 70, + 532, + 235 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 532, + 235 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 532, + 235 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 532, + 235 + ], + "type": "image", + "image_path": "74c4599864a12bc52e56972b056b846857fbe93ef3957a5256dc9eb750e9bdaa.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 247, + 548, + 346 + ], + "lines": [ + { + "bbox": [ + 46, + 247, + 548, + 346 + ], + "spans": [ + { + "bbox": [ + 46, + 247, + 548, + 346 + ], + "type": "text", + "content": "Figure 1. Overview of the NoisyQuant pipeline (orange box) with comparison to EasyQuant [36] (gray box). Histograms of input activations, quantized activations, and layer outputs are illustrated in the pipeline. Starting from the input activation " + }, + { + "bbox": [ + 46, + 247, + 548, + 346 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 46, + 247, + 548, + 346 + ], + "type": "text", + "content": " following a GELU function, where the green histogram is plotted in linear-scale and the gray in log-scale, NoisyQuant adds a fixed NoisyBias " + }, + { + "bbox": [ + 46, + 247, + 548, + 346 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 247, + 548, + 346 + ], + "type": "text", + "content": " sampled from a Uniform distribution onto " + }, + { + "bbox": [ + 46, + 247, + 548, + 346 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 46, + 247, + 548, + 346 + ], + "type": "text", + "content": ". Adding the noise before quantization flattens the peaks in the activation distribution to make it more friendly to quantization, as visualized in the linear/log-scaled histogram of " + }, + { + "bbox": [ + 46, + 247, + 548, + 346 + ], + "type": "inline_equation", + "content": "X + N" + }, + { + "bbox": [ + 46, + 247, + 548, + 346 + ], + "type": "text", + "content": ", and theoretically proved in Sec. 3.2. The noise is removed from the result of the fixed-point activation-weight multiplication with a denoising bias term, which we formulate in Sec. 3.3. The rightmost sub-figures illustrate the layer output distribution of EasyQuant and NoisyQuant. Compared to the output of the floating-point model (yellow), NoisyQuant output follows the distribution closely and achieves up to " + }, + { + "bbox": [ + 46, + 247, + 548, + 346 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 46, + 247, + 548, + 346 + ], + "type": "text", + "content": " output error reduction on some transformer layers, as shown in Sec. 4. The resulting PTQ performance improvement of the entire model is provided in Sec. 5." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 357, + 289, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 357, + 289, + 501 + ], + "spans": [ + { + "bbox": [ + 46, + 357, + 289, + 501 + ], + "type": "text", + "content": "to millions of activation values with sophisticated distributions. For instance, outputs of the GELU function [15] are asymmetrically distributed, with spikes in the histogram at some values and a long tail in a large range (see top-left of Fig. 1). Some linear projection layers also lead to significantly large activation values that are sparsely distributed in a very long tail [23]. Consequently, low-precision PTQ on vision transformer suffers from performance degradation, even utilizing non-linear or mixed-precision quantizers at the cost of additional data communication and computation overheads [23,43]. No linear uniform PTQ method achieves good performance on vision transformer models." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "text", + "content": "This paper provides a brand new perspective in dealing with the complicated vision transformer activation distribution. Instead of adding more tricks in the quantizer design to fit the activation distribution, this work explores the potential of actively and cost-effectively altering the distribution being quantized, making it more friendly to a given quantizer. The pipeline of our proposed method is illustrated in Fig. 1. Specifically, we make a surprising discovery that for any quantizer, the quantization error can be significantly reduced by adding a fixed noisy bias sampled from a Uniform distribution to the activation before quantization. We theoretically prove the condition when the quantization error reduction can be achieved. On this basis, we propose NoisyQuant, a plug-and-play, quantizer-agnostic enhancement on the posttraining activation quantization of vision transformer models. For each layer, we sample a Noisy Bias based on the input activation distribution following our theoretical analysis, and" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 357, + 547, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 357, + 547, + 500 + ], + "spans": [ + { + "bbox": [ + 304, + 357, + 547, + 500 + ], + "type": "text", + "content": "compute the corresponding denoising bias to retain the correct output of the layer. At inference time, the Noisy Bias is added to the input activation before quantization, and the denoising bias is removed from the layer output. This process significantly reduces the quantization error with minimal computation overhead. NoisyQuant leads to significant improvement in the PTQ performance of state-of-the-art vision transformers. Applying NoisyQuant on top of a uniform linear quantization achieves on-par performance to SOTA mixed-precision, nonlinear PTQ methods [23, 42]. Adding NoisyQuant on top of these nonlinear quantizers achieves further performance gain." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 500, + 547, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 500, + 547, + 524 + ], + "spans": [ + { + "bbox": [ + 305, + 500, + 547, + 524 + ], + "type": "text", + "content": "To the best of our knowledge, this paper makes the following novel contributions:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 317, + 531, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 317, + 531, + 547, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 531, + 547, + 567 + ], + "spans": [ + { + "bbox": [ + 317, + 531, + 547, + 567 + ], + "type": "text", + "content": "- Theoretically shows the possibility and proves feasible conditions for reducing the quantization error of heavily-tailed distributions with a fixed additive noisy bias;" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 574, + 547, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 574, + 547, + 645 + ], + "spans": [ + { + "bbox": [ + 317, + 574, + 547, + 645 + ], + "type": "text", + "content": "- Proposes NoisyQuant, a quantizer-agnostic enhancement for PTQ performance on activation quantization. NoisyQuant achieves the first success in actively refining the distribution being quantized to reduce quantization error following the theoretical results on additive noisy bias, with minimal computation overhead;" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 317, + 654, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 654, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 654, + 547, + 713 + ], + "type": "text", + "content": "- Demonstrates consistent performance improvement by applying NoisyQuant on top of existing PTQ quantizers. For 6-bit PTQ, NoisyQuant improves the ImageNet top-1 accuracy of uniform linear quantized vision transformer models by up to " + }, + { + "bbox": [ + 317, + 654, + 547, + 713 + ], + "type": "inline_equation", + "content": "1.7\\%" + }, + { + "bbox": [ + 317, + 654, + 547, + 713 + ], + "type": "text", + "content": ", and improves SOTA" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20322" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 67, + 72, + 269, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 269, + 85 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 269, + 85 + ], + "type": "text", + "content": "nonlinear quantizer PTQ4ViT [43] by up to " + }, + { + "bbox": [ + 67, + 72, + 269, + 85 + ], + "type": "inline_equation", + "content": "0.7\\%" + }, + { + "bbox": [ + 67, + 72, + 269, + 85 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 96, + 132, + 109 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 96, + 132, + 109 + ], + "spans": [ + { + "bbox": [ + 47, + 96, + 132, + 109 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 116, + 255, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 116, + 255, + 129 + ], + "spans": [ + { + "bbox": [ + 47, + 116, + 255, + 129 + ], + "type": "text", + "content": "2.1. Transformer models in computer vision" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 135, + 289, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 135, + 289, + 351 + ], + "spans": [ + { + "bbox": [ + 47, + 135, + 289, + 351 + ], + "type": "text", + "content": "Since Self-Attention (SA) [31]-based transformer models have shown impressive performance in Natural Language Processing tasks, researchers transfer the idea into computer vision. Pioneering work on Vision Transformer [10] for the first time builds a totally SA-based model in the field of computer vision and shows its effectiveness. Since then, a variety of transformer-based vision models have emerged with boosted performance, claiming state-of-the-art in multiple computer vision tasks. Notable advances include additional data augmentation and distillation [30], incorporating multi-stage architecture [22, 33], and explore novel attention design [14,22,41]. Transformer models also advance to tasks beyond classification, such as object detection [3], semantic segmentation [46], and image generation [17]. Furthermore, the development of vision transformer brings unification of model architecture and representation across image and text, enabling stronger vision-language connection via image-text-pair contrastive learning [27]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 351, + 289, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 351, + 289, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 351, + 289, + 495 + ], + "type": "text", + "content": "The rapid development of vision transformer models motivates the exploration of quantization, as quantization brings straightforward memory and computation savings agnostic to model architectures. Compared to CNN models, the use of advanced activation functions like GELU [15] and more complicated computation graphs like the attention mechanism makes vision transformers more sensitive to quantization [43]. This work aims to resolve the difficulty of posttraining vision transformer activation quantization caused by the heavy-tailed activation distribution, which we believe will facilitate the deployment of state-of-the-art transformer models into real-world computer vision applications." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 503, + 236, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 503, + 236, + 516 + ], + "spans": [ + { + "bbox": [ + 47, + 503, + 236, + 516 + ], + "type": "text", + "content": "2.2. Post-training quantization methods" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 522, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 289, + 713 + ], + "type": "text", + "content": "Post-training quantization (PTQ) is a preferable quantization method in scenarios with limited training data or limited computational resources for the expensive quantization-aware fine-tuning process. Previous methods have dedicatedly studied PTQ on CNNs. For instance, EasyQuant [36] presents a quick searching mechanism to determine the proper clipping range for quantization. ZeroQ [2] proposes a distillation mechanism to generate proxy input images, which can leverage inherent statistics of batch normalization layers to conduct PTQ. SQuant [12] calibrates the model for lower quantization error on the basis of the sensitivity obtained by the Hessian spectrum. On vision transformers, Liu et al. [23] presents a linear PTQ method that uses Pearson correlation coefficients and ranking loss to determine scaling factors. PTQ4ViT [43] introduces a nonlinear Twin Uniform Quantization based on vision transformers' acti" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 547, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 144 + ], + "type": "text", + "content": "vation distributions, which set different scaling factors for 1) positive and negative activations of GeLU, and 2) small and large values of Softmax, which can reduce the activation quantization error to some extent with the cost of additional computation overhead. [19, 34] reparameterize the LN layer to suppress outliers by scaling down activation values." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 144, + 547, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 144, + 547, + 228 + ], + "spans": [ + { + "bbox": [ + 304, + 144, + 547, + 228 + ], + "type": "text", + "content": "Unlike previous PTQ methods that analyze the activation distribution and fit the quantizer accordingly, our method takes a novel perspective of actively modifying the distribution being quantized with a sampled noisy bias. We show our proposed method can bring quantization error reduction and performance enhancement for all the PTQ quantizers for vision transformer activation quantization." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 238, + 362, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 238, + 362, + 251 + ], + "spans": [ + { + "bbox": [ + 306, + 238, + 362, + 251 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 258, + 547, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 258, + 547, + 342 + ], + "spans": [ + { + "bbox": [ + 304, + 258, + 547, + 342 + ], + "type": "text", + "content": "In this section, we provide a preliminary introduction to the notation and typical formulation of DNN quantization in Sec. 3.1, theoretically analyze the potential of reducing quantization error with pre-sampled additive noisy bias in Sec. 3.2, and formulate NoisyQuant, a noisy bias-enhanced post-training activation quantization method with reduced quantization error in Sec. 3.3." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 349, + 385, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 349, + 385, + 361 + ], + "spans": [ + { + "bbox": [ + 306, + 349, + 385, + 361 + ], + "type": "text", + "content": "3.1. Preliminary" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 367, + 547, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 367, + 547, + 415 + ], + "spans": [ + { + "bbox": [ + 304, + 367, + 547, + 415 + ], + "type": "text", + "content": "To start with, we recap the post-training quantization process on DNN models. Given the dominance of fully-connected (FC) linear projection layers in transformer models, here we focus our discussion on the FC layer." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 415, + 547, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 415, + 547, + 450 + ], + "spans": [ + { + "bbox": [ + 305, + 415, + 547, + 450 + ], + "type": "text", + "content": "The computation performed in a FC layer with weight " + }, + { + "bbox": [ + 305, + 415, + 547, + 450 + ], + "type": "inline_equation", + "content": "W \\in \\mathbb{R}^{k \\times m}" + }, + { + "bbox": [ + 305, + 415, + 547, + 450 + ], + "type": "text", + "content": ", input activation " + }, + { + "bbox": [ + 305, + 415, + 547, + 450 + ], + "type": "inline_equation", + "content": "X \\in \\mathbb{R}^{m \\times n}" + }, + { + "bbox": [ + 305, + 415, + 547, + 450 + ], + "type": "text", + "content": ", and bias " + }, + { + "bbox": [ + 305, + 415, + 547, + 450 + ], + "type": "inline_equation", + "content": "B \\in \\mathbb{R}^{k \\times 1}" + }, + { + "bbox": [ + 305, + 415, + 547, + 450 + ], + "type": "text", + "content": " can be formulated as" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 384, + 460, + 545, + 472 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 460, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 384, + 460, + 545, + 472 + ], + "type": "interline_equation", + "content": "f (X) = W X + B. \\tag {1}", + "image_path": "34853fb98c16560528aab5d530dd9277313358eeceeb23afda957c6042000e6e.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 480, + 547, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 480, + 547, + 576 + ], + "spans": [ + { + "bbox": [ + 304, + 480, + 547, + 576 + ], + "type": "text", + "content": "The main computation burden in Eq. (1) lies in the matrix multiplication of " + }, + { + "bbox": [ + 304, + 480, + 547, + 576 + ], + "type": "inline_equation", + "content": "WX" + }, + { + "bbox": [ + 304, + 480, + 547, + 576 + ], + "type": "text", + "content": ", which requires a total of " + }, + { + "bbox": [ + 304, + 480, + 547, + 576 + ], + "type": "inline_equation", + "content": "k \\times m \\times n" + }, + { + "bbox": [ + 304, + 480, + 547, + 576 + ], + "type": "text", + "content": " multiply-accumulate computations (MACs). Post-training quantization aims to reduce the computation cost of matrix multiplication by converting " + }, + { + "bbox": [ + 304, + 480, + 547, + 576 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 304, + 480, + 547, + 576 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 480, + 547, + 576 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 304, + 480, + 547, + 576 + ], + "type": "text", + "content": " to fixed-point quantized values, therefore replacing floating-point MACs with much cheaper fixed-point operations [16, 40]. The quantized FC layer can therefore be represented as" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 362, + 585, + 545, + 597 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 585, + 545, + 597 + ], + "spans": [ + { + "bbox": [ + 362, + 585, + 545, + 597 + ], + "type": "interline_equation", + "content": "f _ {q} (X) = q _ {W} (W) q _ {A} (X) + B, \\tag {2}", + "image_path": "08d1d4f6e209d0ef25bf66e299da94ca94b4a06373af2d0383f6950494ecb850.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "inline_equation", + "content": "q_{W}(\\cdot)" + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "inline_equation", + "content": "q_{A}(\\cdot)" + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "content": " denotes the quantizer function of weight and activation respectively. Previous research has observed that the heave-tailed activation distribution of " + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "content": " in transformer is causing significant quantization error between " + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "inline_equation", + "content": "q_{A}(X)" + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "content": " at low precision, leading to significant performance degradation. Previous PTQ methods modify the design of quantizer " + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "inline_equation", + "content": "q_{A}(\\cdot)" + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "content": " to mitigate the quantization error. In this paper, we propose an alternative approach, where we modify the distribution of activation " + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 304, + 605, + 547, + 713 + ], + "type": "text", + "content": " with a" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20323" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 97 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 97 + ], + "type": "text", + "content": "pre-sampled noisy bias before quantization, and prove it to be a plug-and-play enhancement on any quantizer design." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 103, + 266, + 117 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 103, + 266, + 117 + ], + "spans": [ + { + "bbox": [ + 47, + 103, + 266, + 117 + ], + "type": "text", + "content": "3.2. Theoretical analysis on quantization error" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 121, + 289, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 121, + 289, + 254 + ], + "spans": [ + { + "bbox": [ + 46, + 121, + 289, + 254 + ], + "type": "text", + "content": "As previous research mainly blames the heavy-tailed distribution of transformer activation as the main cause for large quantization errors [23, 42], in this section we analyze how the quantization error change as we actively alter the activation distribution being quantized. A straightforward way to alter the input activation distribution is to add a fixed \"Noisy Bias\" sampled from a Uniform random distribution. Denoting the input activation as " + }, + { + "bbox": [ + 46, + 121, + 289, + 254 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 46, + 121, + 289, + 254 + ], + "type": "text", + "content": " and the Noisy Bias as " + }, + { + "bbox": [ + 46, + 121, + 289, + 254 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 121, + 289, + 254 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 121, + 289, + 254 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 121, + 289, + 254 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 121, + 289, + 254 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 46, + 121, + 289, + 254 + ], + "type": "text", + "content": " have the same dimensions, here we formulate the quantization error difference between " + }, + { + "bbox": [ + 46, + 121, + 289, + 254 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 46, + 121, + 289, + 254 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 121, + 289, + 254 + ], + "type": "inline_equation", + "content": "X + N" + }, + { + "bbox": [ + 46, + 121, + 289, + 254 + ], + "type": "text", + "content": " for a quantizer " + }, + { + "bbox": [ + 46, + 121, + 289, + 254 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 46, + 121, + 289, + 254 + ], + "type": "text", + "content": " in Eq. (3)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 261, + 287, + 290 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 261, + 287, + 290 + ], + "spans": [ + { + "bbox": [ + 69, + 261, + 287, + 290 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} D (X, N) = Q E (X + N) - Q E (X) = \\\\ \\left| \\left| \\left(Q (X + N) - X - N\\right) \\right| \\right| _ {2} ^ {2} - \\left| \\left| \\left(Q (X) - X\\right) \\right| \\right| _ {2} ^ {2}. \\tag {3} \\\\ \\end{array}", + "image_path": "befa04b5a90201753b7673434e8d94891c989aec4151a22d4733f9cb2fcade6d.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 298, + 287, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 298, + 287, + 322 + ], + "spans": [ + { + "bbox": [ + 46, + 298, + 287, + 322 + ], + "type": "text", + "content": "Theorem 1 states the condition where " + }, + { + "bbox": [ + 46, + 298, + 287, + 322 + ], + "type": "inline_equation", + "content": "D(x, N) \\leq 0" + }, + { + "bbox": [ + 46, + 298, + 287, + 322 + ], + "type": "text", + "content": " holds for each histogram snapshot " + }, + { + "bbox": [ + 46, + 298, + 287, + 322 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 298, + 287, + 322 + ], + "type": "text", + "content": " of activation " + }, + { + "bbox": [ + 46, + 298, + 287, + 322 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 46, + 298, + 287, + 322 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 45, + 329, + 287, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 329, + 287, + 389 + ], + "spans": [ + { + "bbox": [ + 45, + 329, + 287, + 389 + ], + "type": "text", + "content": "Theorem 1. Consider a given quantizer " + }, + { + "bbox": [ + 45, + 329, + 287, + 389 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 45, + 329, + 287, + 389 + ], + "type": "text", + "content": " with quantization bin width " + }, + { + "bbox": [ + 45, + 329, + 287, + 389 + ], + "type": "inline_equation", + "content": "2b" + }, + { + "bbox": [ + 45, + 329, + 287, + 389 + ], + "type": "text", + "content": ". For each histogram snapshot of " + }, + { + "bbox": [ + 45, + 329, + 287, + 389 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 45, + 329, + 287, + 389 + ], + "type": "text", + "content": " where all elements " + }, + { + "bbox": [ + 45, + 329, + 287, + 389 + ], + "type": "inline_equation", + "content": "X_{i}" + }, + { + "bbox": [ + 45, + 329, + 287, + 389 + ], + "type": "text", + "content": " have the same distance " + }, + { + "bbox": [ + 45, + 329, + 287, + 389 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 45, + 329, + 287, + 389 + ], + "type": "text", + "content": " away from the center of the quantization bin, and for a Noisy Bias " + }, + { + "bbox": [ + 45, + 329, + 287, + 389 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 45, + 329, + 287, + 389 + ], + "type": "text", + "content": " sampled from " + }, + { + "bbox": [ + 45, + 329, + 287, + 389 + ], + "type": "inline_equation", + "content": "N \\sim \\mathcal{U}(-n,n)" + }, + { + "bbox": [ + 45, + 329, + 287, + 389 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 45, + 329, + 287, + 389 + ], + "type": "inline_equation", + "content": "x \\leq n \\leq 2b - x" + }, + { + "bbox": [ + 45, + 329, + 287, + 389 + ], + "type": "text", + "content": ", we have" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 78, + 397, + 287, + 425 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 397, + 287, + 425 + ], + "spans": [ + { + "bbox": [ + 78, + 397, + 287, + 425 + ], + "type": "interline_equation", + "content": "D (x, N) \\leq 0 \\text {i f} 0 \\leq x \\leq n \\left(1 - \\sqrt {\\frac {n}{3 b}}\\right). \\tag {4}", + "image_path": "76efac4b112491b9c15a24e2ffa1dfccb6f9ba13d08aca0c5f5b8173b9302fd3.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 430, + 287, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 430, + 287, + 466 + ], + "spans": [ + { + "bbox": [ + 46, + 430, + 287, + 466 + ], + "type": "text", + "content": "Proof. Without loss of generality, we consider the quantization bin of quantizer " + }, + { + "bbox": [ + 46, + 430, + 287, + 466 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 46, + 430, + 287, + 466 + ], + "type": "text", + "content": " which " + }, + { + "bbox": [ + 46, + 430, + 287, + 466 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 430, + 287, + 466 + ], + "type": "text", + "content": " falls into to be centered at the 0 point, therefore " + }, + { + "bbox": [ + 46, + 430, + 287, + 466 + ], + "type": "inline_equation", + "content": "0 \\leq x \\leq b" + }, + { + "bbox": [ + 46, + 430, + 287, + 466 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 466, + 287, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 466, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 46, + 466, + 287, + 491 + ], + "type": "text", + "content": "In this way, " + }, + { + "bbox": [ + 46, + 466, + 287, + 491 + ], + "type": "inline_equation", + "content": "Q(X_{i}) = b" + }, + { + "bbox": [ + 46, + 466, + 287, + 491 + ], + "type": "text", + "content": " for all elements in the histogram snapshot, leading to a quantization error" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 123, + 498, + 287, + 512 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 498, + 287, + 512 + ], + "spans": [ + { + "bbox": [ + 123, + 498, + 287, + 512 + ], + "type": "interline_equation", + "content": "Q E (X) = (b - x) ^ {2}. \\tag {5}", + "image_path": "5b1d3db52e081e6f11bff998ad715bc79b9e2d46fc737242b932d3b8da3a9747.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 519, + 289, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 519, + 289, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 519, + 289, + 628 + ], + "type": "text", + "content": "Consider adding " + }, + { + "bbox": [ + 46, + 519, + 289, + 628 + ], + "type": "inline_equation", + "content": "N \\sim \\mathcal{U}(-n, n)" + }, + { + "bbox": [ + 46, + 519, + 289, + 628 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 519, + 289, + 628 + ], + "type": "inline_equation", + "content": "X + N" + }, + { + "bbox": [ + 46, + 519, + 289, + 628 + ], + "type": "text", + "content": " will follow " + }, + { + "bbox": [ + 46, + 519, + 289, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{U}(x - n, x + n)" + }, + { + "bbox": [ + 46, + 519, + 289, + 628 + ], + "type": "text", + "content": ". In the case of " + }, + { + "bbox": [ + 46, + 519, + 289, + 628 + ], + "type": "inline_equation", + "content": "x \\leq n \\leq 2b - x" + }, + { + "bbox": [ + 46, + 519, + 289, + 628 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 519, + 289, + 628 + ], + "type": "inline_equation", + "content": "Q(X_i + N_i) = b" + }, + { + "bbox": [ + 46, + 519, + 289, + 628 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 46, + 519, + 289, + 628 + ], + "type": "inline_equation", + "content": "N_i \\in [-x, n]" + }, + { + "bbox": [ + 46, + 519, + 289, + 628 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 519, + 289, + 628 + ], + "type": "inline_equation", + "content": "Q(X_i + N_i) = -b" + }, + { + "bbox": [ + 46, + 519, + 289, + 628 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 46, + 519, + 289, + 628 + ], + "type": "inline_equation", + "content": "N_i \\in [-n, -x)" + }, + { + "bbox": [ + 46, + 519, + 289, + 628 + ], + "type": "text", + "content": " for all elements in " + }, + { + "bbox": [ + 46, + 519, + 289, + 628 + ], + "type": "inline_equation", + "content": "X + N" + }, + { + "bbox": [ + 46, + 519, + 289, + 628 + ], + "type": "text", + "content": ". Though we only sample a single instance of " + }, + { + "bbox": [ + 46, + 519, + 289, + 628 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 519, + 289, + 628 + ], + "type": "text", + "content": " from the Uniform distribution, given the large number of elements in the histogram snapshot, the empirical quantization error " + }, + { + "bbox": [ + 46, + 519, + 289, + 628 + ], + "type": "inline_equation", + "content": "QE(X + N)" + }, + { + "bbox": [ + 46, + 519, + 289, + 628 + ], + "type": "text", + "content": " can be estimated with an expectation over all " + }, + { + "bbox": [ + 46, + 519, + 289, + 628 + ], + "type": "inline_equation", + "content": "N \\sim \\mathcal{U}(-n, n)" + }, + { + "bbox": [ + 46, + 519, + 289, + 628 + ], + "type": "text", + "content": " following the Weak Law of Large Numbers, as" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 69, + 634, + 287, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 634, + 287, + 717 + ], + "spans": [ + { + "bbox": [ + 69, + 634, + 287, + 717 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {E} _ {N} [ Q E (X + N) ] = \\frac {1}{2 n} \\int_ {x - n} ^ {x + n} Q E (x + z) d z \\\\ = \\frac {1}{2 n} \\left[ \\int_ {x - n} ^ {0} (z + b) ^ {2} d z + \\int_ {0} ^ {x + n} (z - b) ^ {2} d z \\right] \\tag {6} \\\\ = x ^ {2} - \\frac {b}{n} x ^ {2} + \\frac {n}{3} - n b + b ^ {2}. \\\\ \\end{array}", + "image_path": "52e0baedce405d368e4675cf324272d080512c4c3664f33a4e65ffa918263fe5.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 318, + 72, + 482, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 72, + 482, + 84 + ], + "spans": [ + { + "bbox": [ + 318, + 72, + 482, + 84 + ], + "type": "text", + "content": "Combining Eq. (5) and Eq. (6), we have" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 339, + 92, + 545, + 133 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 92, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 339, + 92, + 545, + 133 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} D (x, N) = \\mathbf {E} _ {N} [ Q E (X + N) ] - Q E (X) \\\\ = - \\frac {b}{n} x ^ {2} + 2 b x + \\frac {n ^ {2}}{3} - n b. \\tag {7} \\\\ \\end{array}", + "image_path": "3bd8f5d1150f754bed213cca43ab54994fa962b1269dced80086d1ae870c9fee.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 140, + 495, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 140, + 495, + 152 + ], + "spans": [ + { + "bbox": [ + 305, + 140, + 495, + 152 + ], + "type": "text", + "content": "We verify this derivation empirically in Sec. 4." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 152, + 547, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 152, + 547, + 200 + ], + "spans": [ + { + "bbox": [ + 304, + 152, + 547, + 200 + ], + "type": "text", + "content": "It can be observed that given " + }, + { + "bbox": [ + 304, + 152, + 547, + 200 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 304, + 152, + 547, + 200 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 152, + 547, + 200 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 152, + 547, + 200 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 152, + 547, + 200 + ], + "type": "inline_equation", + "content": "D(x, N)" + }, + { + "bbox": [ + 304, + 152, + 547, + 200 + ], + "type": "text", + "content": " is a quadratic function with respect to the activation value " + }, + { + "bbox": [ + 304, + 152, + 547, + 200 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 152, + 547, + 200 + ], + "type": "text", + "content": ". The inequality " + }, + { + "bbox": [ + 304, + 152, + 547, + 200 + ], + "type": "inline_equation", + "content": "D(x, N) \\leq 0, 0 \\leq x \\leq n" + }, + { + "bbox": [ + 304, + 152, + 547, + 200 + ], + "type": "text", + "content": " can be easily solved with respect to " + }, + { + "bbox": [ + 304, + 152, + 547, + 200 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 152, + 547, + 200 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 372, + 209, + 545, + 236 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 209, + 545, + 236 + ], + "spans": [ + { + "bbox": [ + 372, + 209, + 545, + 236 + ], + "type": "interline_equation", + "content": "0 \\leq x \\leq n \\left(1 - \\sqrt {\\frac {n}{3 b}}\\right), \\tag {8}", + "image_path": "18f27130d21f6bff8a90d9daaec6953eb952eea683b2cab8b2e1c63e83c46344.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 243, + 545, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 243, + 545, + 255 + ], + "spans": [ + { + "bbox": [ + 305, + 243, + 545, + 255 + ], + "type": "text", + "content": "which is always feasible given both " + }, + { + "bbox": [ + 305, + 243, + 545, + 255 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 305, + 243, + 545, + 255 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 243, + 545, + 255 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 305, + 243, + 545, + 255 + ], + "type": "text", + "content": " are positive." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 262, + 547, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 262, + 547, + 310 + ], + "spans": [ + { + "bbox": [ + 304, + 262, + 547, + 310 + ], + "type": "text", + "content": "Theorem 1 indicates that adding Noisy Bias can always reduce the quantization error of elements close to the center of the quantization bin, which is the source of large quantization error in a heavy-tailed distribution." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 311, + 547, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 311, + 547, + 382 + ], + "spans": [ + { + "bbox": [ + 304, + 311, + 547, + 382 + ], + "type": "text", + "content": "In practice, to choose a proper " + }, + { + "bbox": [ + 304, + 311, + 547, + 382 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 311, + 547, + 382 + ], + "type": "text", + "content": " for a layer in a pretrained model, we can first acquire the empirical distribution of activation " + }, + { + "bbox": [ + 304, + 311, + 547, + 382 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 304, + 311, + 547, + 382 + ], + "type": "text", + "content": " by passing a small amount of calibration data through the pretrained model. With the distribution of " + }, + { + "bbox": [ + 304, + 311, + 547, + 382 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 304, + 311, + 547, + 382 + ], + "type": "text", + "content": ", we can estimate the expected quantization error reduction of the layer numerically as a function of " + }, + { + "bbox": [ + 304, + 311, + 547, + 382 + ], + "type": "inline_equation", + "content": "n" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 376, + 391, + 545, + 417 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 391, + 545, + 417 + ], + "spans": [ + { + "bbox": [ + 376, + 391, + 545, + 417 + ], + "type": "interline_equation", + "content": "\\mathcal {L} (n) = \\sum_ {x \\in X} [ D (x, N) ]. \\tag {9}", + "image_path": "a00546ed5261ed3d55259d24f1af88669b90021f59a237fef8b5f0c2337ce1fc.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 425, + 547, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 425, + 547, + 484 + ], + "spans": [ + { + "bbox": [ + 305, + 425, + 547, + 484 + ], + "type": "text", + "content": "Note that " + }, + { + "bbox": [ + 305, + 425, + 547, + 484 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 305, + 425, + 547, + 484 + ], + "type": "text", + "content": " here denotes the distance between each activation value and the corresponding quantization bin center. Though we cannot find a closed-form solution for the minima of " + }, + { + "bbox": [ + 305, + 425, + 547, + 484 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(n)" + }, + { + "bbox": [ + 305, + 425, + 547, + 484 + ], + "type": "text", + "content": ", we can easily find a good-enough choice with a linear search over a small range of " + }, + { + "bbox": [ + 305, + 425, + 547, + 484 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 305, + 425, + 547, + 484 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 491, + 504, + 505 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 491, + 504, + 505 + ], + "spans": [ + { + "bbox": [ + 306, + 491, + 504, + 505 + ], + "type": "text", + "content": "3.3. NoisyQuant formulation and pipeline" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 510, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 547, + 714 + ], + "type": "text", + "content": "Given our theoretical analysis on reducing quantization error with Noisy Bias, we propose to apply the Noisy Bias as an enhancement to the PTQ methods of vision transformer models, which we name as \"NoisyQuant\". The pipeline of NoisyQuant is visualized in Fig. 1. Specifically, before performing PTQ and deploying the model, we sample a single Noisy Bias " + }, + { + "bbox": [ + 304, + 510, + 547, + 714 + ], + "type": "inline_equation", + "content": "N \\in \\mathbb{R}^{m \\times 1}" + }, + { + "bbox": [ + 304, + 510, + 547, + 714 + ], + "type": "text", + "content": " for each layer from a Uniform distribution, and fix it throughout the inference. The range of " + }, + { + "bbox": [ + 304, + 510, + 547, + 714 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 510, + 547, + 714 + ], + "type": "text", + "content": " is determined by the calibration objective defined in Eq. (9). At inference time, we add the Noisy Bias " + }, + { + "bbox": [ + 304, + 510, + 547, + 714 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 510, + 547, + 714 + ], + "type": "text", + "content": " to the input activation " + }, + { + "bbox": [ + 304, + 510, + 547, + 714 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 304, + 510, + 547, + 714 + ], + "type": "text", + "content": " under the broadcasting rule before going through the activation quantizer, therefore actively altering the distribution being quantized to reduce the input quantization error. After the activation-weight multiplication in the linear layer, we remove the impact of " + }, + { + "bbox": [ + 304, + 510, + 547, + 714 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 510, + 547, + 714 + ], + "type": "text", + "content": " by adjusting the bias term in the linear layer with a denoising bias computed from " + }, + { + "bbox": [ + 304, + 510, + 547, + 714 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 510, + 547, + 714 + ], + "type": "text", + "content": ", therefore retrieving the correct output. Formally," + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20324" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 96 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 96 + ], + "type": "text", + "content": "NoisyQuant converts the quantized FC computation defined in Eq. (2) into" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 52, + 104, + 287, + 118 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 104, + 287, + 118 + ], + "spans": [ + { + "bbox": [ + 52, + 104, + 287, + 118 + ], + "type": "interline_equation", + "content": "f _ {N q} (X) = q _ {W} (W) q _ {A} (X + N) + \\left(B - q _ {W} (W) N\\right). \\tag {10}", + "image_path": "34cf401a346e520abd58cf883c957e0c0423f3762d477679a985b5be22bad59b.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 125, + 287, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 125, + 287, + 245 + ], + "spans": [ + { + "bbox": [ + 46, + 125, + 287, + 245 + ], + "type": "text", + "content": "Since " + }, + { + "bbox": [ + 46, + 125, + 287, + 245 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 125, + 287, + 245 + ], + "type": "text", + "content": " is sampled before the deployment of the model and fixed it throughout the inference, the denoising output bias " + }, + { + "bbox": [ + 46, + 125, + 287, + 245 + ], + "type": "inline_equation", + "content": "B' = B - q_W(W)N" + }, + { + "bbox": [ + 46, + 125, + 287, + 245 + ], + "type": "text", + "content": " only needs to be computed once before the deployment, without any computation overhead in the inference process. The only computation overhead brought by NoisyQuant at inference time is the on-the-fly summation of " + }, + { + "bbox": [ + 46, + 125, + 287, + 245 + ], + "type": "inline_equation", + "content": "X + N" + }, + { + "bbox": [ + 46, + 125, + 287, + 245 + ], + "type": "text", + "content": ", which is negligible comparing to the cost of the matrix multiplications. Both " + }, + { + "bbox": [ + 46, + 125, + 287, + 245 + ], + "type": "inline_equation", + "content": "B'" + }, + { + "bbox": [ + 46, + 125, + 287, + 245 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 125, + 287, + 245 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 125, + 287, + 245 + ], + "type": "text", + "content": " shall be stored with higher precision. In practice, we store both variables as INT16 to enable integer-only inference." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 245, + 287, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 245, + 287, + 281 + ], + "spans": [ + { + "bbox": [ + 46, + 245, + 287, + 281 + ], + "type": "text", + "content": "Comparing to the output of the floating-point FC layer defined in Eq. (1), the output quantization error induced by the activation quantization defined in Eq. (2) is" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 289, + 287, + 313 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 289, + 287, + 313 + ], + "spans": [ + { + "bbox": [ + 55, + 289, + 287, + 313 + ], + "type": "interline_equation", + "content": "Q E _ {O} (X) = \\left\\| f _ {q} (X) - f (X) \\right\\| _ {2} ^ {2} = \\left\\| W [ Q (X) - X ] \\right\\| _ {2} ^ {2}, \\tag {11}", + "image_path": "73b63c3362f5cbd72ae4b02ac1e8506e7d315c95d736871de62f2847f6df9f68.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 314, + 287, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 314, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 314, + 287, + 361 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 314, + 287, + 361 + ], + "type": "inline_equation", + "content": "Q(\\cdot)" + }, + { + "bbox": [ + 46, + 314, + 287, + 361 + ], + "type": "text", + "content": " is short of " + }, + { + "bbox": [ + 46, + 314, + 287, + 361 + ], + "type": "inline_equation", + "content": "q_{A}(\\cdot)" + }, + { + "bbox": [ + 46, + 314, + 287, + 361 + ], + "type": "text", + "content": ", and we omit the weight quantization " + }, + { + "bbox": [ + 46, + 314, + 287, + 361 + ], + "type": "inline_equation", + "content": "q_{W}(\\cdot)" + }, + { + "bbox": [ + 46, + 314, + 287, + 361 + ], + "type": "text", + "content": " as we are focusing on the activation quantization. Similarly, the output quantization error after the application of NoisyQuant can be computed as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 369, + 287, + 399 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 369, + 287, + 399 + ], + "spans": [ + { + "bbox": [ + 69, + 369, + 287, + 399 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} Q E _ {O} ^ {\\prime} (X) = | | f _ {N q} (X) - f (X) | | _ {2} ^ {2} \\\\ = \\left\\| W [ Q (X + N) - X - N ] \\right\\| _ {2} ^ {2}. \\tag {12} \\\\ \\end{array}", + "image_path": "59642ad2db751718b77c6e53cb545643120ce0b97dd488a78b38be87d9300865.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 407, + 287, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 407, + 287, + 466 + ], + "spans": [ + { + "bbox": [ + 46, + 407, + 287, + 466 + ], + "type": "text", + "content": "As we prove in Theorem 1 that the Noisy Bias enables " + }, + { + "bbox": [ + 46, + 407, + 287, + 466 + ], + "type": "inline_equation", + "content": "X + N" + }, + { + "bbox": [ + 46, + 407, + 287, + 466 + ], + "type": "text", + "content": " to have a lower quantization error than " + }, + { + "bbox": [ + 46, + 407, + 287, + 466 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 46, + 407, + 287, + 466 + ], + "type": "text", + "content": ", we can achieve lower output quantization error in Eq. (12) than in Eq. (11) when NoisyQuant is applied, therefore improving the PTQ performance of the model." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 476, + 219, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 476, + 219, + 490 + ], + "spans": [ + { + "bbox": [ + 47, + 476, + 219, + 490 + ], + "type": "text", + "content": "4. Theoretical insight verification" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 496, + 287, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 496, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 46, + 496, + 287, + 568 + ], + "type": "text", + "content": "In this section, we provide empirical verification of our theoretical analysis on reducing the activation quantization error with NoisyQuant. We numerically verify Theorem 1 with simulated data in Sec. 4.1, and demonstrate the reduction in both input and output quantization errors under the true inference-time activation distribution in Sec. 4.2." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 575, + 237, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 575, + 237, + 588 + ], + "spans": [ + { + "bbox": [ + 47, + 575, + 237, + 588 + ], + "type": "text", + "content": "4.1. Empirical verification of Theorem 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "type": "text", + "content": "Here we follow the settings in Theorem 1 to empirically verify its theoretical derivation. Specifically, we set the quantization bin range " + }, + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "type": "inline_equation", + "content": "b = 1" + }, + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "type": "text", + "content": ", and explore how the quantization error difference induced by the Noisy Bias change with different choice of activation value " + }, + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "type": "text", + "content": " and noisy bias range " + }, + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "type": "text", + "content": ". For all empirical results we experiment with 10 instances of independently sampled Noisy Bias " + }, + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "type": "text", + "content": ", and report the mean and standard derivation of " + }, + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "type": "inline_equation", + "content": "D(X, N)" + }, + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "type": "text", + "content": " defined in Eq. (3) across the 10 instances. We consider input activation " + }, + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 46, + 594, + 288, + 713 + ], + "type": "text", + "content": " to be a tensor with 20 dimensions. Given the tens to hundreds" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 321, + 77, + 531, + 180 + ], + "blocks": [ + { + "bbox": [ + 321, + 77, + 531, + 180 + ], + "lines": [ + { + "bbox": [ + 321, + 77, + 531, + 180 + ], + "spans": [ + { + "bbox": [ + 321, + 77, + 531, + 180 + ], + "type": "image", + "image_path": "5dfc8a3bd58abeb0ef5849f913c9d72b2732a19dcc8bfd0ef13df63ff8b80773.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 188, + 547, + 211 + ], + "lines": [ + { + "bbox": [ + 305, + 188, + 547, + 211 + ], + "spans": [ + { + "bbox": [ + 305, + 188, + 547, + 211 + ], + "type": "text", + "content": "Figure 2. Verifying Eq. (7) with respect to Noisy Bias range " + }, + { + "bbox": [ + 305, + 188, + 547, + 211 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 305, + 188, + 547, + 211 + ], + "type": "text", + "content": ". Circle indicates the mean and error bar the std for evaluated results." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 319, + 217, + 533, + 332 + ], + "blocks": [ + { + "bbox": [ + 319, + 217, + 533, + 332 + ], + "lines": [ + { + "bbox": [ + 319, + 217, + 533, + 332 + ], + "spans": [ + { + "bbox": [ + 319, + 217, + 533, + 332 + ], + "type": "image", + "image_path": "c4d1d795130efb6e642d1f738f6648bdfd1e708ad0cad0d6f302718bbe515558.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 342, + 545, + 363 + ], + "lines": [ + { + "bbox": [ + 305, + 342, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 305, + 342, + 545, + 363 + ], + "type": "text", + "content": "Figure 3. Verifying Eq. (7) with respect to activation value " + }, + { + "bbox": [ + 305, + 342, + 545, + 363 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 305, + 342, + 545, + 363 + ], + "type": "text", + "content": ". Circle indicates the mean and error bar the std for evaluated results." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 375, + 545, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 375, + 545, + 446 + ], + "spans": [ + { + "bbox": [ + 304, + 375, + 545, + 446 + ], + "type": "text", + "content": "of thousands of activation values in each transformer layer, it is likely to see more than 20 activation elements taking the same value. As we base our theoretical derivation on the Weak Law of Large Numbers, having more elements taking the same value will lead to less variance than the simulation results provided in this section." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 447, + 546, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 447, + 546, + 543 + ], + "spans": [ + { + "bbox": [ + 304, + 447, + 546, + 543 + ], + "type": "text", + "content": "For the first experiment, we fix all elements in " + }, + { + "bbox": [ + 304, + 447, + 546, + 543 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 304, + 447, + 546, + 543 + ], + "type": "text", + "content": " to take value " + }, + { + "bbox": [ + 304, + 447, + 546, + 543 + ], + "type": "inline_equation", + "content": "x = 0.1" + }, + { + "bbox": [ + 304, + 447, + 546, + 543 + ], + "type": "text", + "content": ", and alter " + }, + { + "bbox": [ + 304, + 447, + 546, + 543 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 447, + 546, + 543 + ], + "type": "text", + "content": " in the range of [0.1, 1.9]. We compare the empirical " + }, + { + "bbox": [ + 304, + 447, + 546, + 543 + ], + "type": "inline_equation", + "content": "D(X, N)" + }, + { + "bbox": [ + 304, + 447, + 546, + 543 + ], + "type": "text", + "content": " and the theoretical result derived in Eq. (7) in Fig. 2. The evaluated results closely follow the theoretical line. Even with only 20 activation elements the standard deviation across independently sampled Noisy Bias is much smaller than the quantization error benefit brought by adding the Noisy Bias." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 543, + 545, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 543, + 545, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 543, + 545, + 639 + ], + "type": "text", + "content": "Similarly, we fix " + }, + { + "bbox": [ + 304, + 543, + 545, + 639 + ], + "type": "inline_equation", + "content": "n = 1.4" + }, + { + "bbox": [ + 304, + 543, + 545, + 639 + ], + "type": "text", + "content": " and alter activation value " + }, + { + "bbox": [ + 304, + 543, + 545, + 639 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 543, + 545, + 639 + ], + "type": "text", + "content": " in the range of [0, 0.6]. The results are shown in Fig. 3, where we see the same trend of close mean and small variance as in the previous experiment. Quantization error can be reduced by Noisy Bias when " + }, + { + "bbox": [ + 304, + 543, + 545, + 639 + ], + "type": "inline_equation", + "content": "x < 0.44" + }, + { + "bbox": [ + 304, + 543, + 545, + 639 + ], + "type": "text", + "content": ", which follows the numerical computation result of Eq. (4). Both results show the correctness and stability of our derivation of Eq. (7) and the proof of Theorem 1." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 647, + 524, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 647, + 524, + 659 + ], + "spans": [ + { + "bbox": [ + 306, + 647, + 524, + 659 + ], + "type": "text", + "content": "4.2. Quantization errors in transformer layers" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": "As mentioned in Sec. 3.2, NoisyQuant enables the reduction of both input activation quantization error " + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "inline_equation", + "content": "(QE)" + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": " and output error " + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "inline_equation", + "content": "(QE_O)" + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": " of a linear layer given a quantizer. In this section, we verify the error reduction on the real acti" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20325" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 69, + 145, + 267, + 232 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 288, + 137 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 288, + 137 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 288, + 137 + ], + "type": "text", + "content": "Table 1. Averaged quantization error on different types of layers. The calculation is performed with 5,120 images randomly selected from the ImageNet validation set on the Swin-T model. Layer names are defined as same as in Timm [35]. We compute the input quantization error difference " + }, + { + "bbox": [ + 46, + 70, + 288, + 137 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 70, + 288, + 137 + ], + "type": "text", + "content": " and compare the output errors between NoisyQuant and EasyQuant." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 145, + 267, + 232 + ], + "lines": [ + { + "bbox": [ + 69, + 145, + 267, + 232 + ], + "spans": [ + { + "bbox": [ + 69, + 145, + 267, + 232 + ], + "type": "table", + "html": "
Layer NameEq. (3) D(X)Eq. (11) QEO(X)Eq. (12) QE'(X)Output QE drop
qkv-0.0290.1070.09610%
proj-0.1060.0930.08113%
fc1-0.5270.1260.1232%
fc2-5.1301.0490.85219%
", + "image_path": "d09097d4441909ebb85d52dc87c3a79b69d4cf57a80a76028f2c2c37c5bcc3af.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 70, + 245, + 262, + 350 + ], + "blocks": [ + { + "bbox": [ + 70, + 245, + 262, + 350 + ], + "lines": [ + { + "bbox": [ + 70, + 245, + 262, + 350 + ], + "spans": [ + { + "bbox": [ + 70, + 245, + 262, + 350 + ], + "type": "image", + "image_path": "4c0a4ca95467480a8818a0f4efbb91c0f57851b30811acad0ea47ec99787d011.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 351, + 287, + 384 + ], + "lines": [ + { + "bbox": [ + 46, + 351, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 46, + 351, + 287, + 384 + ], + "type": "text", + "content": "Figure 4. Quantization error histogram of fc2 input activation. We apply NoisyQuant on top of the EasyQuant quantizer, and compare the quantization error distribution (red) with EasyQuant (grey)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 400, + 288, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 400, + 288, + 555 + ], + "spans": [ + { + "bbox": [ + 46, + 400, + 288, + 555 + ], + "type": "text", + "content": "vation distribution of vision transformer model. Practically, we run evaluation on randomly selected 5120 images of validation set with linear PTQ quantizer EasyQuant [36] as baseline, and calculate " + }, + { + "bbox": [ + 46, + 400, + 288, + 555 + ], + "type": "inline_equation", + "content": "QE" + }, + { + "bbox": [ + 46, + 400, + 288, + 555 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 400, + 288, + 555 + ], + "type": "inline_equation", + "content": "QE_{O}" + }, + { + "bbox": [ + 46, + 400, + 288, + 555 + ], + "type": "text", + "content": " with or without adding Noisy Bias. Experiments are ran on different types of layers, namely qkv, proj, fc1, and fc2, respectively. The results are averaged across all layers of each type. As shown in Tab. 1, input quantization error difference " + }, + { + "bbox": [ + 46, + 400, + 288, + 555 + ], + "type": "inline_equation", + "content": "D(X)" + }, + { + "bbox": [ + 46, + 400, + 288, + 555 + ], + "type": "text", + "content": " defined in Eq. (3) is consistently lower than 0, which meets our expectation that Noisy Bias reduces input quantization error. This further leads to significant drop in output error between the quantized and floating-point layers for all layer types. Up to " + }, + { + "bbox": [ + 46, + 400, + 288, + 555 + ], + "type": "inline_equation", + "content": "19\\%" + }, + { + "bbox": [ + 46, + 400, + 288, + 555 + ], + "type": "text", + "content": " averaged output error drop can be achieved in fc2 layers." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "content": "To further understand the impact NoisyQuant makes, we visualize both the input and output distribution of each layer. We start with highlighting fc2 as it achieves the greatest QE reduction. The fc2 layer takes in the output of GELU activations, which has significantly complicated distribution as introduced in Sec. 1. Fig. 4 visualizes the distribution of input quantization error with or without NoisyQuant. With EasyQuant only, significant amount of activation elements are showing large quantization error, which is most probably owing to bad quantization on densely distributed areas. Instead, NoisyQuant generally makes QE distributed plainly and for most elements near zero, therefore leading to significant QE reduction overall. Subsequently, we plot the" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 307, + 69, + 545, + 138 + ], + "blocks": [ + { + "bbox": [ + 307, + 69, + 545, + 138 + ], + "lines": [ + { + "bbox": [ + 307, + 69, + 545, + 138 + ], + "spans": [ + { + "bbox": [ + 307, + 69, + 545, + 138 + ], + "type": "image", + "image_path": "978b7759d6d4bb4eb3539e9a99a82b200347cb51e1dfb99fa21970f16265939b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 307, + 150, + 545, + 219 + ], + "blocks": [ + { + "bbox": [ + 365, + 140, + 487, + 150 + ], + "lines": [ + { + "bbox": [ + 365, + 140, + 487, + 150 + ], + "spans": [ + { + "bbox": [ + 365, + 140, + 487, + 150 + ], + "type": "text", + "content": "(a) " + }, + { + "bbox": [ + 365, + 140, + 487, + 150 + ], + "type": "inline_equation", + "content": "26\\%" + }, + { + "bbox": [ + 365, + 140, + 487, + 150 + ], + "type": "inline_equation", + "content": "QE_{O}" + }, + { + "bbox": [ + 365, + 140, + 487, + 150 + ], + "type": "text", + "content": " reduction on qkv layer" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 307, + 150, + 545, + 219 + ], + "lines": [ + { + "bbox": [ + 307, + 150, + 545, + 219 + ], + "spans": [ + { + "bbox": [ + 307, + 150, + 545, + 219 + ], + "type": "image", + "image_path": "716ff5a9eb45fde689295d819729611413002e0751353a84adaa23166f962598.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 364, + 220, + 487, + 231 + ], + "lines": [ + { + "bbox": [ + 364, + 220, + 487, + 231 + ], + "spans": [ + { + "bbox": [ + 364, + 220, + 487, + 231 + ], + "type": "text", + "content": "(b) " + }, + { + "bbox": [ + 364, + 220, + 487, + 231 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 364, + 220, + 487, + 231 + ], + "type": "inline_equation", + "content": "QE_{O}" + }, + { + "bbox": [ + 364, + 220, + 487, + 231 + ], + "type": "text", + "content": " reduction on proj layer" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 307, + 232, + 545, + 300 + ], + "blocks": [ + { + "bbox": [ + 307, + 232, + 545, + 300 + ], + "lines": [ + { + "bbox": [ + 307, + 232, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 307, + 232, + 545, + 300 + ], + "type": "image", + "image_path": "87dda685b56e4bca845df8e6e12ebceaa8357c72453f137fc7ba1261466a3c4c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 307, + 312, + 545, + 380 + ], + "blocks": [ + { + "bbox": [ + 365, + 302, + 486, + 312 + ], + "lines": [ + { + "bbox": [ + 365, + 302, + 486, + 312 + ], + "spans": [ + { + "bbox": [ + 365, + 302, + 486, + 312 + ], + "type": "text", + "content": "(c) " + }, + { + "bbox": [ + 365, + 302, + 486, + 312 + ], + "type": "inline_equation", + "content": "36\\%" + }, + { + "bbox": [ + 365, + 302, + 486, + 312 + ], + "type": "inline_equation", + "content": "QE_{O}" + }, + { + "bbox": [ + 365, + 302, + 486, + 312 + ], + "type": "text", + "content": " reduction on fc1 layer" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 307, + 312, + 545, + 380 + ], + "lines": [ + { + "bbox": [ + 307, + 312, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 307, + 312, + 545, + 380 + ], + "type": "image", + "image_path": "786681dd849f84ee122486e2fe11d4cdbe44948926dc304f6bcc5e46f425c70b.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 365, + 383, + 486, + 392 + ], + "lines": [ + { + "bbox": [ + 365, + 383, + 486, + 392 + ], + "spans": [ + { + "bbox": [ + 365, + 383, + 486, + 392 + ], + "type": "text", + "content": "(d) " + }, + { + "bbox": [ + 365, + 383, + 486, + 392 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 365, + 383, + 486, + 392 + ], + "type": "inline_equation", + "content": "QE_{O}" + }, + { + "bbox": [ + 365, + 383, + 486, + 392 + ], + "type": "text", + "content": " reduction on fc2 layer" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 395, + 545, + 429 + ], + "lines": [ + { + "bbox": [ + 305, + 395, + 545, + 429 + ], + "spans": [ + { + "bbox": [ + 305, + 395, + 545, + 429 + ], + "type": "text", + "content": "Figure 5. Output histogram in selected transformer layers before and after activation quantization. NoisyQuant significantly reduces the output quantization error over EasyQuant baselines." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 440, + 547, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 440, + 547, + 500 + ], + "spans": [ + { + "bbox": [ + 304, + 440, + 547, + 500 + ], + "type": "text", + "content": "histogram of selected layers' outputs in Fig. 5. NoisyQuant output consistently follows the floating-point layer output distribution closer than that of EasyQuant, especially at values with large magnitudes. This leads to up to " + }, + { + "bbox": [ + 304, + 440, + 547, + 500 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 304, + 440, + 547, + 500 + ], + "type": "text", + "content": " error reduction in the output of some vision transformer layers." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 511, + 387, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 511, + 387, + 525 + ], + "spans": [ + { + "bbox": [ + 306, + 511, + 387, + 525 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 531, + 547, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 531, + 547, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 531, + 547, + 639 + ], + "type": "text", + "content": "This section evaluates the impact of NoisyQuant on the PTQ performance of some commonly used vision transformer architectures. First, we evaluate the performance on image classification tasks with the ViT [10], DeiT [30] and Swin [22] transformer models. We also evaluate the performance of object detection with the DETR model [4]. Finally, we provide ablation study on how NoisyQuant improves the PTQ performance when added to each type of layers in the vision transformer model." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 647, + 430, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 647, + 430, + 659 + ], + "spans": [ + { + "bbox": [ + 306, + 647, + 430, + 659 + ], + "type": "text", + "content": "5.1. Experimental settings" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": "Dataset. The classification tasks are implemented on the ImageNet-2012 classification dataset [28], with 1.2 million training images and 50,000 validation images in 1000 classes. For object detection task, the MSCOCO 2017 dataset [20]" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20326" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 178, + 279, + 355 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 289, + 170 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 289, + 170 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 289, + 170 + ], + "type": "text", + "content": "Table 2. PTQ accuracy of ViT models. Top-1 accuracy on ImageNet is reported in the table. W/A represents weight and activation bit-width, respectively. The default patch size is " + }, + { + "bbox": [ + 46, + 70, + 289, + 170 + ], + "type": "inline_equation", + "content": "16 \\times 16" + }, + { + "bbox": [ + 46, + 70, + 289, + 170 + ], + "type": "text", + "content": " and the default input resolution is " + }, + { + "bbox": [ + 46, + 70, + 289, + 170 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 46, + 70, + 289, + 170 + ], + "type": "text", + "content": ". “*” after model name indicates " + }, + { + "bbox": [ + 46, + 70, + 289, + 170 + ], + "type": "inline_equation", + "content": "384 \\times 384" + }, + { + "bbox": [ + 46, + 70, + 289, + 170 + ], + "type": "text", + "content": " input. “MP” represents mixed-precision. For our methods, “NoisyQuant-Linear” is implemented on top of the EasyQuant quantizer, whereas “NoisyQuant-PTQ4ViT” is implemented on the nonlinear PTQ4ViT quantizer. All NoisyQuant results are shown with mean±STD of 20 repeated experiments." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 178, + 279, + 355 + ], + "lines": [ + { + "bbox": [ + 58, + 178, + 279, + 355 + ], + "spans": [ + { + "bbox": [ + 58, + 178, + 279, + 355 + ], + "type": "table", + "html": "
MethodW/AViT-SViT-BViT-B*
Pretrained32/3281.3984.5486.00
Percentile [18]6/667.7477.6377.60
Liu et al. [23]6 MP-75.26-
PTQ4ViT-Linear [43]6/670.2475.6646.88
EasyQuant [36]6/675.1381.4282.02
NoisyQuant-Linear6/676.86±0.0681.90±0.1183.00±0.09
PTQ4ViT [43]6/678.6381.6583.34
NoisyQuant-PTQ4ViT6/678.65±0.0782.32±0.0983.22±0.10
Percentile [18]8/878.7780.1282.53
Liu et al. [23]8 MP-76.98-
FQ-ViT [21]8/8-83.31-
PTQ4ViT-Linear [43]8/880.4683.8985.35
EasyQuant [36]8/880.7583.8985.53
NoisyQuant-Linear8/880.81±0.0184.10±0.0385.56±0.01
PTQ4ViT [43]8/881.0084.2585.82
NoisyQuant-PTQ4ViT8/881.15±0.0284.22±0.0185.86±0.01
", + "image_path": "de6ec65404be564baab6175275f5587c05c5ddfe8e38a932532bebc1256d7f4e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 57, + 399, + 279, + 615 + ], + "blocks": [ + { + "bbox": [ + 46, + 368, + 287, + 391 + ], + "lines": [ + { + "bbox": [ + 46, + 368, + 287, + 391 + ], + "spans": [ + { + "bbox": [ + 46, + 368, + 287, + 391 + ], + "type": "text", + "content": "Table 3. PTQ accuracy of DeiT models. Abbreviations are the same as Tab. 2." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 399, + 279, + 615 + ], + "lines": [ + { + "bbox": [ + 57, + 399, + 279, + 615 + ], + "spans": [ + { + "bbox": [ + 57, + 399, + 279, + 615 + ], + "type": "table", + "html": "
PretrainedW/ADeiT-SDeiT-BDeiT-B*
Pretrained32/3279.8581.8083.11
Percentile [18]6/670.4973.9978.24
Bit-Split [32]6/674.5876.39-
Liu et al. [23]6/674.5877.02-
Liu et al. [23]6 MP75.1077.47-
PTQ4VIT-Linear [43]6/672.2678.7868.44
EasyQuant [36]6/675.2779.4781.26
NoisyQuant-Linear6/676.37±0.1379.77±0.0881.40±0.06
PTQ4VIT [43]6/676.2880.2581.55
NoisyQuant-PTQ4ViT6/677.43±0.0880.70±0.1181.65±0.04
Percentile [18]8/873.9875.2180.02
Bit-Split [32]8/876.3979.42-
Liu et al. [23]8/877.4780.48-
Liu et al. [23]8 MP78.0981.29-
FQ-ViT [21]8/879.1781.20-
PTQ4VIT-Linear [43]8/877.6580.9482.33
EasyQuant [36]8/878.9881.1982.10
NoisyQuant-Linear8/879.11±0.0281.30±0.0382.23±0.02
PTQ4VIT [43]8/879.4781.4882.97
NoisyQuant-PTQ4ViT8/879.51±0.0181.45±0.0282.49±0.02
", + "image_path": "353a350ed42faf754b2ecf58a2820f1006733e0e33007ee09f266661a19e570f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 628, + 288, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 628, + 288, + 688 + ], + "spans": [ + { + "bbox": [ + 46, + 628, + 288, + 688 + ], + "type": "text", + "content": "is utilized to evaluate the PTQ performance, which contains 118,000 training images and 5,000 validation images. For both tasks we randomly sample 1024 images from the training set as the calibration data for all PTQ methods implemented in our experiments." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 689, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 289, + 715 + ], + "type": "text", + "content": "Pretrained model architecture. We perform uniform 6-bit and 8-bit PTQ on different variants of pretrained vision" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 310, + 102, + 544, + 248 + ], + "blocks": [ + { + "bbox": [ + 305, + 70, + 545, + 93 + ], + "lines": [ + { + "bbox": [ + 305, + 70, + 545, + 93 + ], + "spans": [ + { + "bbox": [ + 305, + 70, + 545, + 93 + ], + "type": "text", + "content": "Table 4. PTQ accuracy of Swin models. Abbreviations are the same as Tab. 2." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 310, + 102, + 544, + 248 + ], + "lines": [ + { + "bbox": [ + 310, + 102, + 544, + 248 + ], + "spans": [ + { + "bbox": [ + 310, + 102, + 544, + 248 + ], + "type": "table", + "html": "
MethodW/ASwin-TSwin-SSwin-BSwin-B*
Pretrained32/3281.3983.2385.2786.44
Percentile [18]6/677.7580.4181.9082.94
PTQ4VIT-Linear [43]6/678.4581.7483.3585.22
EasyQuant [36]6/679.5182.4584.3085.89
NoisyQuant-Linear6/680.01±0.0682.78±0.0484.57±0.0485.96±0.04
PTQ4VIT [43]6/680.4782.3884.0185.38
NoisyQuant-PTQ4ViT6/680.51±0.0382.86±0.0584.68±0.0686.03±0.07
Percentile [18]8/879.8880.9383.0884.31
FQ-ViT [21]8/880.5182.71--
PTQ4VIT [43]8/880.9682.7584.7986.16
EasyQuant [36]8/880.9583.0085.1086.39
NoisyQuant-Linear8/881.05±0.0383.07±0.0385.11±0.0486.42±0.02
PTQ4VIT [43]8/881.2483.1085.1486.39
NoisyQuant-PTQ4ViT8/881.25±0.0283.13±0.0185.20±0.0386.44±0.01
", + "image_path": "0500aafde49fdf2784f43e76016007aebe708d654228d889abab876ed9efbf1f.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 252, + 547, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 252, + 547, + 361 + ], + "spans": [ + { + "bbox": [ + 304, + 252, + 547, + 361 + ], + "type": "text", + "content": "transformer models. For classification tasks we quantize the model family of ViT [10], DeiT [30], and Swin [22] provided by the Timm library [35], including large-scale models with " + }, + { + "bbox": [ + 304, + 252, + 547, + 361 + ], + "type": "inline_equation", + "content": "384 \\times 384" + }, + { + "bbox": [ + 304, + 252, + 547, + 361 + ], + "type": "text", + "content": " input resolution (marked with “*”). For detection, we perform PTQ on the official implementation of DETR [4] with ResNet-50 backbone. We use the pretrained model checkpoint provided by the official source of each model, whose floating-point performances are reported in the results tables and match their original papers." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 361, + 547, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 361, + 547, + 601 + ], + "spans": [ + { + "bbox": [ + 304, + 361, + 547, + 601 + ], + "type": "text", + "content": "Implementation Details. Following the mainstream quantization methods [23, 43], we quantize all the weights and inputs involved in matrix multiplication. More specifically, we quantize weights of qkv-projectors, attention output projectors, MLPs, linear embeddings, and model heads. Besides, we also quantize input activations of linear layers and matrix multiplication operators. Following [23, 26, 43, 45], we keep layer normalization and softmax operations as full precision. Apart from special notes, we perform symmetric layer-wise quantization for activations and symmetric channel-wise quantization for weights. In particular, weights are quantized by absolute MinMax values without clamping. We implement NoisyQuant on top of linear quantizer EasyQuant [36] and nonlinear quantizer PTQ4ViT [43]. After we apply the selected quantizer to determine the quantization bin width and location on calibration data, we decide the parameters of the Noisy Bias distribution through a linear search. We use the empirical activation quantization error as defined in Eq. (9) as the search objective, where we find the Noisy Bias parameter that can minimize the quantization error." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 610, + 448, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 610, + 448, + 624 + ], + "spans": [ + { + "bbox": [ + 306, + 610, + 448, + 624 + ], + "type": "text", + "content": "5.2. NoisyQuant performance" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": "Here we compare NoisyQuant with state-of-the-art PTQ methods, including percentile [18], bit-split [32], Liu et al. [23], FQ-ViT [21], EasyQuant [36], and PTQ4ViT [43], on ImageNet. PTQ performance of different variants of ViT, DeiT, and Swin transformer models are provided in Tab. 2, Tab. 3, and Tab. 4, respectively. Experiments are conducted for both linear and nonlinear quantizers." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20327" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 77, + 102, + 261, + 180 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 287, + 93 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 287, + 93 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 287, + 93 + ], + "type": "text", + "content": "Table 5. PTQ accuracy of DETR models. Mean average precision (mAP) evaluated on the MSCOCO 2017 dataset is reported." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 77, + 102, + 261, + 180 + ], + "lines": [ + { + "bbox": [ + 77, + 102, + 261, + 180 + ], + "spans": [ + { + "bbox": [ + 77, + 102, + 261, + 180 + ], + "type": "table", + "html": "
ModelMethodW/AmAP
DETR [4] (COCO2017)Pretrained32/3242.0
Percentile [18]8/838.6
Bit-Split [32]8/840.6
Liu et al. [23]8/841.2
EasyQuant [36]8/841.1
NoisyQuant-Linear8/841.4±0.05
", + "image_path": "4ba97f894bd5d4ee093f4e1dd889e2034880bca12af4ed2fd2c0da5aafe6acb8.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 224, + 289, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 224, + 289, + 487 + ], + "spans": [ + { + "bbox": [ + 46, + 224, + 289, + 487 + ], + "type": "text", + "content": "Linear quantizer. Previous linear PTQ methods suffer from severe performance degradation on vision transformer activation quantization. EasyQuant [36] achieves the best performance among linear quantizers, yet still induce 2-6% of accuracy drop under 6-bit quantization compared to the floating-point baseline. By applying the proposed NoisyQuant method on top of the EasyQuant quantizer, the resulted NoisyQuant-Linear quantizer consistently and significantly improves the PTQ accuracy on all variants of vision transformer models. Under the W6A6 setting, NoisyQuant-Linear achieves performance improvement of 1.73% on ViT-S, 0.98% on ViT-B*, 1.1% on DeiT-S, and 0.5% on SWIN-T over EasyQuant. Under the W8A8 setting, as the performance gap between EasyQuant and floating-point gets smaller, NoisyQuant still appears beneficial in further improving the PTQ performance. It is worth noting that although nonlinear quantizer like PTQ4ViT [43] consistently achieves higher PTQ performance than the linear EasyQuant quantizer, the enhancement brought by NoisyQuant enables the NoisyQuant-Linear quantizer to achieve on-par or even better performance than the nonlinear PTQ4ViT, with much lower overhead in hardware deployment." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 498, + 288, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 288, + 631 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 288, + 631 + ], + "type": "text", + "content": "Nonlinear quantizer. As we claim NoisyQuant as a quantizer-agnostic enhancement of PTQ, we further implement NoisyQuant on top of the nonlinear quantizer PTQ4ViT [43], namely NoisyQuant-PTQ4ViT. NoisyQuant-PTQ4ViT outperforms most results of PTQ4ViT, notably achieving for a " + }, + { + "bbox": [ + 46, + 498, + 288, + 631 + ], + "type": "inline_equation", + "content": "1.25\\%" + }, + { + "bbox": [ + 46, + 498, + 288, + 631 + ], + "type": "text", + "content": " improvement on DeiT-S, " + }, + { + "bbox": [ + 46, + 498, + 288, + 631 + ], + "type": "inline_equation", + "content": "0.67\\%" + }, + { + "bbox": [ + 46, + 498, + 288, + 631 + ], + "type": "text", + "content": " on ViT-S, and " + }, + { + "bbox": [ + 46, + 498, + 288, + 631 + ], + "type": "inline_equation", + "content": "0.67\\%" + }, + { + "bbox": [ + 46, + 498, + 288, + 631 + ], + "type": "text", + "content": " on Swin-B in the W6A6 setting. The W6A6 PTQ performance of Swin-S and Swin-B* for the first time improved to within " + }, + { + "bbox": [ + 46, + 498, + 288, + 631 + ], + "type": "inline_equation", + "content": "0.5\\%" + }, + { + "bbox": [ + 46, + 498, + 288, + 631 + ], + "type": "text", + "content": " of the floating-point baseline by the NoisyQuant-PTQ4ViT quantizer, which can never be achieved without the help of NoisyQuant." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": "Besides classification models, we report the PTQ performance of the DETR object detection model in Tab. 5. NoisyQuant implemented on top of the EasyQuant quantizer also outperforms the EasyQuant baseline, and all previous linear PTQ quantizers including percentile [18], bit-split [32], and Liu et al. [23]." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 334, + 113, + 519, + 275 + ], + "blocks": [ + { + "bbox": [ + 305, + 70, + 547, + 105 + ], + "lines": [ + { + "bbox": [ + 305, + 70, + 547, + 105 + ], + "spans": [ + { + "bbox": [ + 305, + 70, + 547, + 105 + ], + "type": "text", + "content": "Table 6. Applying NoisyQuant on different layer types. Starting from W6A6 PTQ with the EasyQuant quantizer, we add NoisyQuant to each type of layers in the transformer model." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 334, + 113, + 519, + 275 + ], + "lines": [ + { + "bbox": [ + 334, + 113, + 519, + 275 + ], + "spans": [ + { + "bbox": [ + 334, + 113, + 519, + 275 + ], + "type": "table", + "html": "
Modelqkv noiseproj noisefc1 noisefc2 noiseTop-1 W6A6
DeiT-S [30]XXXX75.27
✓XXX75.38
X✓XX75.45
XX✓X75.33
XXX✓76.21
✓✓✓✓76.37
Swin-T [22]XXXX79.51
✓XXX79.53
X✓XX79.56
XX✓X79.52
XXX✓79.80
✓✓✓✓80.01
", + "image_path": "f34b2c5db688afa3a050d627d05d15bddb196c281a997b2eeafa2932f8659678.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 286, + 538, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 286, + 538, + 300 + ], + "spans": [ + { + "bbox": [ + 305, + 286, + 538, + 300 + ], + "type": "text", + "content": "5.3. NoisyQuant's impact on different layer types" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 308, + 547, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 308, + 547, + 498 + ], + "spans": [ + { + "bbox": [ + 304, + 308, + 547, + 498 + ], + "type": "text", + "content": "As we show the effectiveness of NoisyQuant in improving the PTQ performance of the entire model, here we explore if NoisyQuant is helpful for all types of layers in the vision transformer model. Tab. 6 exhibits the effect of applying NoisyQuant on different types of layers. The checkmark in the table means that we apply NoisyQuant to the input activation of all the layers of that particular type, otherwise the layer is quantized with EasyQuant only. For all layer types, applying NoisyQuant consistently improves the PTQ performance of the entire model. Specifically, 'fc2' layers derive the greatest benefit from NoisyQuant, which corresponds to our previous analysis that the activation distribution after GELU function brings the major challenge to PTQ, which can be resolved with the proposed NoisyQuant method. After adding NoisyQuant to the input activation of all linear layers, the model achieves the maximum performance boost." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 516, + 384, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 516, + 384, + 529 + ], + "spans": [ + { + "bbox": [ + 306, + 516, + 384, + 529 + ], + "type": "text", + "content": "6. Conclusions" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 536, + 547, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 536, + 547, + 669 + ], + "spans": [ + { + "bbox": [ + 304, + 536, + 547, + 669 + ], + "type": "text", + "content": "This work proposes NoisyQuant, a noisy bias-enhanced post-training activation quantization method for the complicated activation distribution of vision transformer models. We theoretically prove the quantization error reduction brought by adding noisy bias to the input activation before quantization, and empirically show the effectiveness of NoisyQuant on both linear and nonlinear quantizers on vision transformer models. We hope this work opens up a new direction to reduce PTQ quantization error through actively changing the distribution being quantized, and inspires better Noisy Bias distribution design or generation method based on quantizer and activation characteristics." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 669, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 669, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 669, + 547, + 715 + ], + "type": "text", + "content": "Acknowledgement This work was supported in part by the National Key Research and Development Program of China under Grant 2022YFB4400900. The Berkeley team acknowledges support from Berkeley Deep Drive, Intel Corporation, and Panasonic." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20328" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 289, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 289, + 157 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 289, + 157 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 289, + 157 + ], + "type": "text", + "content": "[1] Yash Bhalgat, Jinwon Lee, Markus Nagel, Tijmen Blankevoort, and Nojun Kwak. Lsq+: Improving low-bit quantization through learnable offsets and better initialization. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 2978-2985, 2020. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 159, + 288, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 159, + 288, + 215 + ], + "spans": [ + { + "bbox": [ + 53, + 159, + 288, + 215 + ], + "type": "text", + "content": "[2] Yaohui Cai, Zhewei Yao, Zhen Dong, Amir Gholami, Michael W. Mahoney, and Kurt Keutzer. Zeroq: A novel zero shot quantization framework. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13166-13175, 2020. 1, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 216, + 288, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 216, + 288, + 270 + ], + "spans": [ + { + "bbox": [ + 53, + 216, + 288, + 270 + ], + "type": "text", + "content": "[3] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 1, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 272, + 288, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 272, + 288, + 316 + ], + "spans": [ + { + "bbox": [ + 53, + 272, + 288, + 316 + ], + "type": "text", + "content": "[4] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. ArXiv, abs/2005.12872, 2020. 6, 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 318, + 288, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 318, + 288, + 361 + ], + "spans": [ + { + "bbox": [ + 53, + 318, + 288, + 361 + ], + "type": "text", + "content": "[5] Minghao Chen, Houwen Peng, Jianlong Fu, and Haibin Ling. Autoformer: Searching transformers for visual recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12270-12280, 2021. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 363, + 288, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 363, + 288, + 407 + ], + "spans": [ + { + "bbox": [ + 53, + 363, + 288, + 407 + ], + "type": "text", + "content": "[6] Tianlong Chen, Yu Cheng, Zhe Gan, Lu Yuan, Lei Zhang, and Zhangyang Wang. Chasing sparsity in vision transformers: An end-to-end exploration. Advances in Neural Information Processing Systems, 34:19974-19988, 2021. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 409, + 288, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 409, + 288, + 462 + ], + "spans": [ + { + "bbox": [ + 53, + 409, + 288, + 462 + ], + "type": "text", + "content": "[7] Jungwook Choi, Zhuo Wang, Swagath Venkataramani, Pierce I-Jen Chuang, Vijayalakshmi Srinivasan, and Kailash Gopalakrishnan. Pact: Parameterized clipping activation for quantized neural networks. arXiv preprint arXiv:1805.06085, 2018. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 465, + 288, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 465, + 288, + 520 + ], + "spans": [ + { + "bbox": [ + 53, + 465, + 288, + 520 + ], + "type": "text", + "content": "[8] Zhen Dong, Zhewei Yao, Daiyaan Arfeen, Amir Gholami, Michael W Mahoney, and Kurt Keutzer. Hawq-v2: Hessian aware trace-weighted quantization of neural networks. Advances in neural information processing systems, 33:18518-18529, 2020. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 521, + 288, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 521, + 288, + 577 + ], + "spans": [ + { + "bbox": [ + 53, + 521, + 288, + 577 + ], + "type": "text", + "content": "[9] Zhen Dong, Zhewei Yao, Amir Gholami, Michael W Mahoney, and Kurt Keutzer. Hawq: Hessian aware quantization of neural networks with mixed-precision. In Proceedings of the IEEE International Conference on Computer Vision, pages 293–302, 2019. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 578, + 288, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 288, + 643 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 288, + 643 + ], + "type": "text", + "content": "[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1, 3, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 646, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 646, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 646, + 288, + 689 + ], + "type": "text", + "content": "[11] Amir Gholami, Sehoon Kim, Zhen Dong, Zhewei Yao, Michael W Mahoney, and Kurt Keutzer. A survey of quantization methods for efficient neural network inference. arXiv preprint arXiv:2103.13630, 2021. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 691, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 714 + ], + "type": "text", + "content": "[12] Cong Guo, Yuxian Qiu, Jingwen Leng, Xiaotian Gao, Chen Zhang, Yunxin Liu, Fan Yang, Yuhao Zhu, and Minyi Guo." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "Squant: On-the-fly data-free quantization via diagonal hessian approximation. arXiv preprint arXiv:2202.07471, 2022.3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 96, + 547, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 547, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 547, + 139 + ], + "type": "text", + "content": "[13] Song Han, Huizi Mao, and William J Dally. Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. arXiv preprint arXiv:1510.00149, 2015. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 141, + 545, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 174 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 174 + ], + "type": "text", + "content": "[14] Ali Hatamizadeh, Hongxu Yin, Jan Kautz, and Pavlo Molchanov. Global context vision transformers. arXiv preprint arXiv:2206.09959, 2022. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 175, + 545, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 196 + ], + "type": "text", + "content": "[15] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 198, + 545, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 198, + 545, + 219 + ], + "spans": [ + { + "bbox": [ + 307, + 198, + 545, + 219 + ], + "type": "text", + "content": "[16] Mark Horowitz. 1.1 computing's energy problem (and what we can do about it). In ISSCC, 2014. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 220, + 547, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 547, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 547, + 262 + ], + "type": "text", + "content": "[17] Kwonjoon Lee, Huiwen Chang, Lu Jiang, Han Zhang, Zhuowen Tu, and Ce Liu. Vitgan: Training gans with vision transformers. arXiv preprint arXiv:2107.04589, 2021.3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 265, + 547, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 265, + 547, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 265, + 547, + 308 + ], + "type": "text", + "content": "[18] Rundong Li, Yan Wang, Feng Liang, Hongwei Qin, Junjie Yan, and Rui Fan. Fully quantized network for object detection. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2805-2814, 2019. 7, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 310, + 547, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 310, + 547, + 352 + ], + "spans": [ + { + "bbox": [ + 307, + 310, + 547, + 352 + ], + "type": "text", + "content": "[19] Zhikai Li, Junrui Xiao, Lianwei Yang, and Qingyi Gu. Repqvit: Scale reparameterization for post-training quantization of vision transformers. arXiv preprint arXiv:2212.08254, 2022.3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 354, + 545, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 354, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 307, + 354, + 545, + 398 + ], + "type": "text", + "content": "[20] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dolkar, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 399, + 545, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 545, + 431 + ], + "type": "text", + "content": "[21] Yang Lin, Tianyu Zhang, Peiqin Sun, Zheng Li, and Shuchang Zhou. Fq-vit: Post-training quantization for fully quantized vision transformer. In IJCAI, 2022. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 434, + 545, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 434, + 545, + 488 + ], + "spans": [ + { + "bbox": [ + 307, + 434, + 545, + 488 + ], + "type": "text", + "content": "[22] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 9992-10002, 2021. 1, 3, 6, 7, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 490, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 490, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 307, + 490, + 545, + 521 + ], + "type": "text", + "content": "[23] Zhenhua Liu, Yunhe Wang, Kai Han, Siwei Ma, and Wen Gao. Post-training quantization for vision transformer. In NeurIPS, 2021. 1, 2, 3, 4, 7, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 523, + 545, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 523, + 545, + 566 + ], + "spans": [ + { + "bbox": [ + 307, + 523, + 545, + 566 + ], + "type": "text", + "content": "[24] Zhenhua Liu, Yunhe Wang, Kai Han, Wei Zhang, Siwei Ma, and Wen Gao. Post-training quantization for vision transformer. Advances in Neural Information Processing Systems, 34:28092-28103, 2021. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 568, + 545, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 568, + 545, + 600 + ], + "spans": [ + { + "bbox": [ + 307, + 568, + 545, + 600 + ], + "type": "text", + "content": "[25] A. Polino, R. Pascanu, and D. Alistarh. Model compression via distillation and quantization. arXiv preprint arXiv:1802.05668, 2018. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 601, + 545, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 601, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 307, + 601, + 545, + 634 + ], + "type": "text", + "content": "[26] Gabriele Prato, Ella Charlaix, and Mehdi Rezagholizadeh. Fully quantized transformer for improved translation. ArXiv, abs/1910.10485, 2019. 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 635, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 545, + 689 + ], + "type": "text", + "content": "[27] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 691, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 714 + ], + "type": "text", + "content": "[28] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy," + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20329" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 66, + 72, + 288, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 288, + 116 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 288, + 116 + ], + "type": "text", + "content": "Aditya Khosla, Michael S. Bernstein, Alexander C. Berg, and Li Fei-Fei. Imagenet large scale visual recognition challenge. International Journal of Computer Vision, 115:211-252, 2015. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 288, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 288, + 172 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 288, + 172 + ], + "type": "text", + "content": "[29] Sheng Shen, Zhen Dong, Jiayu Ye, Linjian Ma, Zhewei Yao, Amir Gholami, Michael W Mahoney, and Kurt Keutzer. Qbert: Hessian based ultra low precision quantization of bert. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 8815-8821, 2020. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 173, + 288, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 173, + 288, + 216 + ], + "spans": [ + { + "bbox": [ + 48, + 173, + 288, + 216 + ], + "type": "text", + "content": "[30] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Herv'e J'egou. Training data-efficient image transformers & distillation through attention. In ICML, 2021. 1, 3, 6, 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 217, + 288, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 217, + 288, + 261 + ], + "spans": [ + { + "bbox": [ + 48, + 217, + 288, + 261 + ], + "type": "text", + "content": "[31] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 1, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 261, + 288, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 261, + 288, + 293 + ], + "spans": [ + { + "bbox": [ + 48, + 261, + 288, + 293 + ], + "type": "text", + "content": "[32] Peisong Wang, Qiang Chen, Xiangyu He, and Jian Cheng. Towards accurate post-training network quantization via bitsplit and stitching. In ICML, 2020. 1, 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 294, + 288, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 294, + 288, + 347 + ], + "spans": [ + { + "bbox": [ + 48, + 294, + 288, + 347 + ], + "type": "text", + "content": "[33] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. arXiv preprint arXiv:2102.12122, 2021. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 349, + 288, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 349, + 288, + 403 + ], + "spans": [ + { + "bbox": [ + 48, + 349, + 288, + 403 + ], + "type": "text", + "content": "[34] Xiuying Wei, Yunchen Zhang, Xiangguo Zhang, Ruihao Gong, Shanghang Zhang, Qi Zhang, Fengwei Yu, and Xianglong Liu. Outlier suppression: Pushing the limit of low-bit transformer language models. arXiv preprint arXiv:2209.13325, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 404, + 288, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 404, + 288, + 437 + ], + "spans": [ + { + "bbox": [ + 48, + 404, + 288, + 437 + ], + "type": "text", + "content": "[35] Ross Wightman. Pytorch image models. https://github.com/rwrightman/pytorch-image-models, 2019.6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 437, + 288, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 437, + 288, + 480 + ], + "spans": [ + { + "bbox": [ + 48, + 437, + 288, + 480 + ], + "type": "text", + "content": "[36] Di Wu, Qi Tang, Yongle Zhao, Ming Zhang, Ying Fu, and Debing Zhang. Easyquant: Post-training quantization via scale optimization. arXiv preprint arXiv:2006.16669, 2020. 1, 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 481, + 288, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 481, + 288, + 525 + ], + "spans": [ + { + "bbox": [ + 48, + 481, + 288, + 525 + ], + "type": "text", + "content": "[37] Lirui Xiao, Huanrui Yang, Zhen Dong, Kurt Keutzer, Li Du, and Shanghang Zhang. Csq: Growing mixed-precision quantization scheme with bi-level continuous sparsification. arXiv preprint arXiv:2212.02770, 2022. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 526, + 288, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 526, + 288, + 559 + ], + "spans": [ + { + "bbox": [ + 48, + 526, + 288, + 559 + ], + "type": "text", + "content": "[38] Huanrui Yang, Lin Duan, Yiran Chen, and Hai Li. Bsq: Exploring bit-level sparsity for mixed-precision neural network quantization. arXiv preprint arXiv:2102.10462, 2021. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 559, + 288, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 288, + 602 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 288, + 602 + ], + "type": "text", + "content": "[39] Huanrui Yang, Hongxu Yin, Pavlo Molchanov, Hai Li, and Jan Kautz. Nvit: Vision transformer compression and parameter redistribution. arXiv preprint arXiv:2110.04869, 2021. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 603, + 288, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 603, + 288, + 658 + ], + "spans": [ + { + "bbox": [ + 48, + 603, + 288, + 658 + ], + "type": "text", + "content": "[40] Zhewei Yao, Zhen Dong, Zhangcheng Zheng, Amir Gholami, Jiali Yu, Eric Tan, Leyuan Wang, Qijing Huang, Yida Wang, Michael Mahoney, et al. Hawq-v3: Dyadic neural network quantization. In International Conference on Machine Learning, pages 11875-11886. PMLR, 2021. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "type": "text", + "content": "[41] Li Yuan, Yunpeng Chen, Tao Wang, Weihao Yu, Yujun Shi, Zihang Jiang, Francis EH Tay, Jiashi Feng, and Shuicheng Yan. Tokens-to-token vit: Training vision transformers from scratch onImagenet. arXiv preprint arXiv:2101.11986, 2021.3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 352 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 307, + 73, + 546, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 546, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 546, + 116 + ], + "type": "text", + "content": "[42] Zhihang Yuan, Chenhao Xue, Yiqi Chen, Qiang Wu, and Guangyu Sun. Ptg4vit: Post-training quantization framework for vision transformers. arXiv preprint arXiv:2111.12293, 2021. 1, 2, 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 547, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 547, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 547, + 161 + ], + "type": "text", + "content": "[43] Zhihang Yuan, Chenhao Xue, Yi-Qiang Chen, Qiang Wu, and Guangyu Sun. Ptq4vit: Post-training quantization for vision transformers with twin uniform quantization. In ECCV, 2022. 2, 3, 7, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 163, + 547, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 547, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 547, + 217 + ], + "type": "text", + "content": "[44] Dongqing Zhang, Jiaolong Yang, Dongqiangzi Ye, and Gang Hua. Lq-nets: Learned quantization for highly accurate and compact deep neural networks. In Proceedings of the European conference on computer vision (ECCV), pages 365–382, 2018. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 218, + 546, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 218, + 546, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 218, + 546, + 251 + ], + "type": "text", + "content": "[45] Wei Zhang, Lu Hou, Yichun Yin, Lifeng Shang, Xiao Chen, Xin Jiang, and Qun Liu. Ternarybert: Distillation-aware ultra-low bit bert. In EMNLP, 2020. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 252, + 546, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 252, + 546, + 306 + ], + "spans": [ + { + "bbox": [ + 307, + 252, + 546, + 306 + ], + "type": "text", + "content": "[46] Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xiang, Philip H.S. Torr, and Li Zhang. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. In CVPR, 2021. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 308, + 547, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 308, + 547, + 352 + ], + "spans": [ + { + "bbox": [ + 307, + 308, + 547, + 352 + ], + "type": "text", + "content": "[47] Shuchang Zhou, Yuxin Wu, Zekun Ni, Xinyu Zhou, He Wen, and Yuheng Zou. Dorefa-net: Training low bitwidth convolutional neural networks with low bitwidth gradients. arXiv preprint arXiv:1606.06160, 2016. 1" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20330" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/NoisyTwins_ Class-Consistent and Diverse Image Generation Through StyleGANs/c323d83e-f322-4c02-8ecc-1d0c86cf0f70_content_list.json b/2023/NoisyTwins_ Class-Consistent and Diverse Image Generation Through StyleGANs/c323d83e-f322-4c02-8ecc-1d0c86cf0f70_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..bf93c065f66486b5acb42abd95d190c82b147008 --- /dev/null +++ b/2023/NoisyTwins_ Class-Consistent and Diverse Image Generation Through StyleGANs/c323d83e-f322-4c02-8ecc-1d0c86cf0f70_content_list.json @@ -0,0 +1,1382 @@ +[ + { + "type": "text", + "text": "NoisyTwins: Class-Consistent and Diverse Image Generation through StyleGANs", + "text_level": 1, + "bbox": [ + 76, + 130, + 893, + 154 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Harsh Rangwani $^{1*}$ Lavish Bansal $^{1,3*}$ Kartik Sharma $^{1,4}$ Tejan Karmali $^{2}$ Varun Jampani $^{2}$ R. Venkatesh Babu $^{1}$", + "bbox": [ + 179, + 179, + 790, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Vision and AI Lab, IISc Bangalore $^{2}$ Google Research $^{3}$ IIT BHU Varanasi $^{4}$ BITS Pilani", + "bbox": [ + 99, + 215, + 870, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 285 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "StyleGANs are at the forefront of controllable image generation as they produce a latent space that is semantically disentangled, making it suitable for image editing and manipulation. However, the performance of StyleGANs severely degrades when trained via class-conditioning on large-scale long-tailed datasets. We find that one reason for degradation is the collapse of latents for each class in the $\\mathcal{W}$ latent space. With NoisyTwins, we first introduce an effective and inexpensive augmentation strategy for class embeddings, which then decorrelates the latents based on self-supervision in the $\\mathcal{W}$ space. This decorrelation mitigates collapse, ensuring that our method preserves intraclass diversity with class-consistency in image generation. We show the effectiveness of our approach on large-scale real-world long-tailed datasets of ImageNet-LT and iNaturalist 2019, where our method outperforms other methods by $\\sim 19\\%$ on FID, establishing a new state-of-the-art.", + "bbox": [ + 75, + 301, + 473, + 559 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 590, + 209, + 606 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "StyleGANs [21, 22] have shown unprecedented success in image generation, particularly on well-curated and articulated datasets (eg. FFHQ for face images, etc.). In addition to generating high fidelity and diverse images, StyleGANs also produce a disentangled latent space, which is extensively used for image editing and manipulation tasks [49]. As a result, StyleGANs are being extensively used in various applications like face-editing [11, 41], video generation [46, 52], face reenactment [3], etc., which are a testament to their usability and generality. However, despite being successful on well-curated datasets, training StyleGANs on in-the-wild and multi-category datasets is still challenging. A large-scale conditional StyleGAN (i.e. StyleGAN-XL) on ImageNet was recently trained successfully by Sauer et al. [39] using the ImageNet pre-trained model through the idea of a projection discriminator [38]. While the StyleGAN-XL uses additional pre-trained mod", + "bbox": [ + 75, + 616, + 472, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/294d72e02624c3cfda3606df3ebfcf6ef336ce375e584f3d0a8017284a84da77.jpg", + "image_caption": [ + "Figure 1. Qualitative Comparison on tail classes (T1-T4) for iNaturalist 2019. We provide sample(s) from real class (with class frequency), generated by StyleGAN2-ADA and after adding proposed NoisyTwins. NoisyTwins achieves remarkable diversity, class-consistency and quality by just using 38 samples on average." + ], + "image_footnote": [], + "bbox": [ + 506, + 270, + 890, + 529 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "els, obtaining such models for distinctive image domains like medical, forensics, and fine-grained data may not be feasible, which limits its generalization across domains.", + "bbox": [ + 496, + 625, + 893, + 671 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we aim to train vanilla class-conditional StyleGAN without any pre-trained models on challenging real-world long-tailed data distributions. As training StyleGAN with augmentations [20, 54] leads to low recall [24] (which measures diversity in the generated images) and mode collapse, particularly for minority (i.e. tail) classes. For investigating this phenomenon further, we take a closer look at the latent $\\mathcal{W}$ space of StyleGAN that is produced by a fully-connected mapping network that takes the conditioning variables $\\mathbf{z}$ (i.e. random noise) and class embedding $\\mathbf{c}$ as inputs. The vectors $\\mathbf{w}$ in $\\mathcal{W}$ space are used for conditioning various layers of the generator (Fig. 2). We find that output vectors $\\mathbf{w}$ from the mapping network hinge on the conditioning variable $\\mathbf{c}$ and become invariant to random conditioning vector $\\mathbf{z}$ . This collapse of latents leads to un", + "bbox": [ + 496, + 674, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal Contribution. Link: rangwani-harsh.github.io/NoisyTwins", + "bbox": [ + 94, + 886, + 442, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "5987", + "bbox": [ + 482, + 944, + 514, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f35637b74c73f94493ea4064e32564bcbdf97cdff0ece6410e380c05a3e53b63.jpg", + "image_caption": [ + "Figure 2. Schematic illustration of $\\mathcal{W}$ space for different GANs. Existing conditioning methods either suffer from mode collapse [20] or lead to class confusion [35] in $\\mathcal{W}$ space. With proposed NoisyTwins, we achieve intra class diversity while avoiding class confusion." + ], + "image_footnote": [], + "bbox": [ + 76, + 89, + 893, + 215 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "stable training and is one of the causes of poor recall (a.k.a. mode collapse) for minority classes. Further, on augmenting StyleGAN with recent conditioning and regularization techniques [17, 35], we find that they either lead to a poor recall for minority classes or lead to class confusion (Fig. 2) instead of mitigating the collapse.", + "bbox": [ + 75, + 262, + 470, + 353 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To mitigate the collapse of $\\mathbf{w}$ in $\\mathcal{W}$ space, we need to ensure that the change in conditioning variable $\\mathbf{z}$ leads to the corresponding change in $\\mathbf{w}$ . Recently in self-supervised learning, several techniques [2, 53] have been introduced to prevent the collapse of learned representations by maximizing the information content in the feature dimensions. Inspired by them we propose NoisyTwins, in which we first generate inexpensive twin augmentations for class embeddings and then use them to decorrelate the $\\mathbf{w}$ variables through self-supervision. The decorrelation ensures that $\\mathbf{w}$ vectors are diverse for each class and the GAN is able to produce intra-class diversity among the generated images.", + "bbox": [ + 75, + 357, + 468, + 539 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We evaluate NoisyTwins on challenging benchmarks of large-scale long-tailed datasets of ImageNet-LT [30] and iNaturalist 2019 [48]. These benchmarks are particularly challenging due to a large number of classes present, which makes GANs prone to class confusion. On the other hand, as these datasets are long-tailed with only a few images per class in tail classes, generating diverse images for those classes is challenging. We observe that existing metrics used in GAN evaluations are not able to capture both class confusion and mode collapse. As a remedy, we propose to use intra-class Frechet Inception Distance (FID) [12] based on features obtained from pre-trained CLIP [34] embeddings as an effective metric to measure the performance of class-conditional GANs in long-tailed data setups. Using NoisyTwins enables StyleGAN to generate diverse and class-consistent images across classes, mitigating the mode collapse and class confusion issues in existing state-of-the-art (SotA) (Fig. 1). Further, with NoisyTwins, we obtain diverse generations for tail classes even with $\\leq 30$ images, which can be attributed to the transfer of knowledge from head classes through shared parameters (Fig. 1 and 6). In summary, we make the following contributions:", + "bbox": [ + 75, + 542, + 470, + 875 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1. We evaluate various recent SotA GAN conditioning", + "bbox": [ + 89, + 885, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "and regularization techniques on the challenging task of long-tailed image generation. We find that all existing methods either suffer from mode collapse or lead to class confusion in generations.", + "bbox": [ + 529, + 262, + 890, + 321 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "2. To mitigate mode collapse and class confusion, we introduce NoisyTwins, an effective and inexpensive augmentation strategy for class embeddings that decorrelates latents in the $\\mathcal{W}$ latent space (Sec. 4).", + "3. We evaluate NoisyTwins on large-scale long-tailed datasets of ImageNet-LT and iNaturalist-2019, where it consistently improves the StyleGAN2 performance $(\\sim 19\\%)$ , achieving a new SotA. Further, our approach can also prevent mode collapse and enhance the performance of few-shot GANs (Sec. 5.3)." + ], + "bbox": [ + 511, + 323, + 890, + 473 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 500, + 489, + 650, + 505 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "StyleGANs. Karras et al. introduced StyleGAN [21] and subsequently improved its image quality in StyleGAN2. StyleGAN could produce high-resolution photorealistic images as demonstrated on a wide variety of category-specific datasets. It introduced a mapping network, which mapped the sampled noise into another latent space, which is more disentangled and semantically coherent, as demonstrated by its downstream usage for image editing and manipulation [1,32,33,42,43]. Further, StyleGAN has been extended to get novel views from images [27, 28, 44], thus making it possible to get 3D information from it. These downstream advances are possible due to the impressive performance of StyleGANs on class-specific datasets (such as faces). However, similar photorealism levels are yet uncommon on multi-class long-tailed datasets (such as ImageNet).", + "bbox": [ + 496, + 518, + 890, + 744 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "GANs for Data Efficiency and Imbalance. Failure of GANs on less data was concurrently reported by Karras et al. [20] and Zhao et al. [54]. The problem is rooted in the overfitting of the discriminator due to less real data. Since then, the proposed solutions for this problem have relied on a) augmenting the data, b) introducing regularizers, and c) architectural modifications. Karras et al. [20] and Zhao et al. [54] relied on differentiable data augmentation before passing images into the discriminator to solve this problem. DeceiveD [15] proposed to introduce label-noise for", + "bbox": [ + 496, + 750, + 893, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "5988", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "discriminator training. LeCamGAN [47] finds that enforcing LeCam divergence as a regularization trick in the discriminator can robustify GAN training under a limited data setting. DynamicD [50] tunes the capacity of the discriminator on-the-fly during training. While these methods can handle the data inefficiency, they are ineffective on class imbalanced long-tailed data distribution [35].", + "bbox": [ + 75, + 90, + 468, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "CBGAN [36] proposed a solution to train the unconditional GAN model on long-tailed data distribution by introducing a signal from the classifier to balance the classes generated by GAN. In a long-tailed class-conditional setting, gSR [35] proposes to regularize the exploding spectral norms of the class-specific parameters of the GAN. Collapse-by-conditioning [40] addresses the limited data in classes by introducing a training regime that transitions from an unconditional to a class-conditioned setting, thus exploiting the shared information across classes during the early stages of the training. However, these methods suffer from either class confusion or poor generated image quality on large datasets, which is resolved by NoisyTwins.", + "bbox": [ + 75, + 196, + 470, + 393 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Self-Supervised Learning for GANs. Ideas from Self-supervised learning have shown their benefits in GAN training. IC-GAN [6] trains GAN conditioned on embeddings on SwAV [5], which led to remarkable improvement in performance on the long-tailed version of ImageNet. InsGen [51] and ReACGAN [16, 17] introduce the auxiliary task of instance discrimination for the discriminator, thereby making the discriminator focus on multiple tasks and thus alleviating discriminator overfitting. While InsGen relies on both noise space and image space augmentations, ReACGAN and ContraGAN follow only image space augmentations. Contrary to these, NoisyTwins performs augmentations in the class-embedding space and contrasts them in the $\\mathcal{W}$ -space of the generator instead of the discriminator.", + "bbox": [ + 75, + 398, + 468, + 609 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminaries", + "text_level": 1, + "bbox": [ + 76, + 625, + 215, + 638 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. StyleGAN", + "text_level": 1, + "bbox": [ + 76, + 648, + 191, + 665 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "StyleGAN [21] is a Generative Adversarial Network comprising of its unique Style Conditioning Based Generator $(\\mathcal{G})$ and discriminator network $(\\mathcal{D})$ trained jointly. We will focus on the architecture of StyleGAN2 [23] as we use it in our experiments, although our work is generally applicable to all StyleGAN architectures. The StyleGAN2 generator is composed of blocks that progressively upsample the features and resolution inspired by Progressive GAN [19], starting from a single root image. The diversity in the images comes from conditioning each block of image generation through conditioning on the latent coming from the mapping network (Fig. 2). The mapping network is a fully connected network that takes in the conditioning variables, the $\\mathbf{z} \\in \\mathbb{R}^d$ coming from a random distribution (e.g., Gaussian, etc.) and class conditioning label $c$ which is converted", + "bbox": [ + 75, + 672, + 470, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "to an embedding $\\mathbf{c} \\in \\mathbb{R}^d$ . The mapping network takes these and outputs vectors $\\mathbf{w}$ in the $\\mathcal{W}$ latent space of StyleGAN, which is found to be semantically disentangled to a high extent [49]. The $\\mathbf{w}$ is then processed through an affine transform and passed to each generator layer for conditioning the image generation process through Adaptive Instance Normalization (AdaIN) [13]. The images from generator $\\mathcal{G}$ , along with real images, are passed to discriminator $\\mathcal{D}$ for training. The training utilizes the non-saturating adversarial losses [9] for $\\mathcal{G}$ and $\\mathcal{D}$ given as:", + "bbox": [ + 496, + 90, + 890, + 242 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\mathcal {D}} \\mathcal {L} _ {\\mathcal {D}} = \\sum_ {i = 1} ^ {m} \\log (\\mathcal {D} (\\mathbf {x} _ {i})) + \\log (1 - \\mathcal {D} (\\mathcal {G} (\\mathbf {z} _ {i}, \\mathbf {c} _ {i}))) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 243, + 890, + 282 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\mathcal {G}} \\mathcal {L} _ {\\mathcal {G}} = \\sum_ {i = 1} ^ {m} - \\log \\left(\\mathcal {D} \\left(\\mathcal {G} \\left(\\mathbf {z} _ {i}, \\mathbf {c} _ {i}\\right)\\right)\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 285, + 890, + 324 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We now describe the issues present in the StyleGANs trained on long-tailed data and their analysis in $\\mathcal{W}$ space.", + "bbox": [ + 496, + 325, + 890, + 356 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Class Confusion and Class-Specific Mode Collapse in Conditional StyleGANs", + "text_level": 1, + "bbox": [ + 498, + 363, + 890, + 395 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To finely analyze the performance of StyleGAN and its variants on long-tailed datasets, we train them on the CIFAR10-LT dataset. In Fig. 3, we plot the qualitative results of generated images and create a t-SNE plot for latents in $\\mathcal{W}$ space for each class. We first train the StyleGAN2 baseline with augmentations (DiffAug) [20, 54]. We find that it leads to mode collapse, specifically for tail classes (Fig. 3). In conjunction with images, we also observe that corresponding t-SNE embeddings are also collapsed near each class's mean in $\\mathcal{W}$ space. Further, recent methods which have proposed the usage of contrastive learning for GANs, improve their data efficiency and prevent discriminator overfitting [14, 16]. We also evaluate them by adding the contrastive conditioning method, which is D2DCE loss-based on ReACGAN [17], to the baseline, where in results, we observe that the network omits to learn tail classes and produces head class images at their place (i.e., class confusion). In Fig. 3, it can be seen that the network confuses semantically similar classes, that is, generating cars (head or majority class) in place of trucks and airplanes (head class) instead of ships. In the $\\mathcal{W}$ space, we find the same number of clusters as the number of classes in the dataset. However, the tail label cluster images also belong to the head classes of cars and airplanes. In a very recent work gSR [35], it has been shown that constraining the spectral norm of $\\mathcal{G}$ embedding parameters can help reduce the mode collapse and lead to stable training. However, we find that constraining the embeddings leads to class confusion, as seen in t-SNE visualization in Fig. 3. We find that this class confusion gets further aggravated when StyleGAN is trained on datasets like ImageNet-LT, which contain a large number of classes, along with a bunch of semantically similar classes (Sec. 5.2). Based on our $\\mathcal{W}$", + "bbox": [ + 496, + 401, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "5989", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a811c8c06599313d1d44d9fb7fc8e236c148611c31d29e93bced5155cb9ad7e8.jpg", + "image_caption": [ + "Figure 3. Comparison of GANs and their $\\mathcal{W}$ space for CIFAR10-LT. We plot the generated images on (left) and generate a t-SNE plot of w latents for generated images in $\\mathcal{W}$ space (right). We find that mode collapse and class confusion in images is linked to the corresponding collapse and confusion in latent $\\mathcal{W}$ space. Our proposed NoisyTwins mitigates both collapse (left) and confusion (right) simultaneously." + ], + "image_footnote": [], + "bbox": [ + 76, + 88, + 496, + 372 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ba2441057d6fd53a9b97ffebb6a7efe48f94bd72b0de4351575b78a6fdf133cc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 527, + 97, + 707, + 217 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/13e9cd0abac8a49b6bc45c0656a00091e549c181cef6566fcfbe5916f0788de5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 714, + 98, + 888, + 215 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4563af6695895049295817ed017d1b32f0eb1173c31bd2a3f989fd79a80fbe3b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 527, + 220, + 707, + 340 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e5e02df03fadf4c2db0c35daccfdd79001539044451f35868abc86705e3f9c49.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 714, + 222, + 888, + 340 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/bc0db085d00bb489b207b957f0775ca2d63ec2d3a70a51d3f01d3982e7f0e3cf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 550, + 344, + 859, + 375 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "space analysis and qualitative results above, we observe that the class confusion and mode collapse of images is tightly coupled with the structure of $\\mathcal{W}$ space. Further, the recent SotA methods are either unable to prevent collapse or suffer from class confusion. Hence, this work aims to develop a technique that mitigates both confusion and collapse.", + "bbox": [ + 75, + 439, + 473, + 530 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Approach", + "text_level": 1, + "bbox": [ + 76, + 542, + 187, + 559 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we present our method NoisyTwins, which introduces noise-based augmentation twins in the conditional embedding space (Sec. 4.1), and then combines it with the Barlow-Twins-based regularizer from the self-supervised learning (SSL) paradigm to resolve the issue of class confusion and mode collapse (Sec. 4.2).", + "bbox": [ + 75, + 566, + 470, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Noise Augmentation in Embedding Space", + "text_level": 1, + "bbox": [ + 76, + 665, + 431, + 683 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As we observed in the previous section that $\\mathbf{w}$ vectors for each sample become insensitive to changes in $\\mathbf{z}$ . This collapse in $\\mathbf{w}$ vectors for each class leads to mode collapse for baselines (Fig. 3). One reason for this could be the fact that $\\mathbf{z}$ is composed of continuous variables, whereas the embedding vectors $\\mathbf{c}$ for each class are discrete. Due to this, the GAN converges to the easy degenerate solution where it generates a single sample for each class, becoming insensitive to changes in $\\mathbf{z}$ . For inducing some continuity in $\\mathbf{c}$ embeddings vectors, we introduce an augmentation strategy where we add i.i.d. noise of small magnitude in each of the variables in $\\mathbf{c}$ . Based on our observation (Fig. 3) and existing works [35], there is a high tendency for mode collapse in tail classes. Hence we add noise in embedding space that is", + "bbox": [ + 75, + 689, + 472, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "proportional to the inverse of the frequency of samples. We provide the mathematical expression of noise augmentation $\\tilde{\\mathbf{c}}$ below:", + "bbox": [ + 496, + 439, + 892, + 481 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathbf {c}} \\sim \\mathbf {c} + \\mathcal {N} (\\mathbf {0}, \\sigma_ {c} \\mathbb {I} _ {d}) \\text {w h e r e} \\sigma_ {c} = \\sigma \\frac {(1 - \\alpha)}{1 - \\alpha^ {n _ {c}}} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 479, + 890, + 508 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here $n_c$ is the frequency of training samples in class $c$ , $\\mathbb{I}$ is the identity matrix of size $d \\times d$ , and $\\alpha, \\sigma$ are hyperparameters. The expression of $\\sigma_c$ is from the effective number of samples [8], which is a softer version of inverse frequency proportionality. In contrast to the image space augmentation, these noise augmentations come for free as there is no significant additional computation overhead. This noise is added in the embedding $\\mathbf{c}$ before passing it to the generator and the discriminator, which ensures that the class embeddings occupy a continuous region in latent space.", + "bbox": [ + 496, + 508, + 892, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Insight: The augmentation equation above (Eq. 3) can be interpreted as approximating the discrete random variable $\\mathbf{c}$ with a Gaussian with finite variance and the embedding parameters $\\mathbf{c}$ being the mean $\\mu_{\\mathbf{c}}$ .", + "bbox": [ + 496, + 659, + 890, + 719 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathbf {c}} \\sim \\mathcal {N} \\left(\\mu_ {\\mathbf {c}}, \\sigma_ {c} \\mathbb {I} _ {d}\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 635, + 719, + 890, + 734 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This leads to the class-embedding input $\\tilde{\\mathbf{c}}$ to mapping network to have a Gaussian distribution, similar in nature to $\\mathbf{z}$ . This noise augmentation strategy alone mitigates the degenerate solution of class-wise mode collapse to a great extent (Table 1) and helps generate diverse latent $\\mathbf{w}$ for each class. Due to the diverse $\\mathbf{w}$ conditioning of the GAN, it leads to diverse image generation.", + "bbox": [ + 496, + 734, + 893, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Invariance in $\\mathcal{W}$ -Space with NoisyTwins", + "text_level": 1, + "bbox": [ + 500, + 847, + 841, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The augmentation strategy introduced in the previous section expands the region for each class in $\\mathcal{W}$ latent space.", + "bbox": [ + 500, + 869, + 892, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "5990", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/efb94e7c8a3ba242f1b8e2fea2e43e6860f58772a173022a1344d604c67a202e.jpg", + "image_caption": [ + "Figure 4. Overview of NoisyTwins. For the $i^{th}$ sample of class $c^i$ , we create twin augmentations $(\\tilde{\\mathbf{c}}_a^i,\\tilde{\\mathbf{c}}_b^i)$ , by sampling from a Gaussian centered at class embedding $(\\mu_{\\mathbf{c}^i})$ . After this, we concatenate them with the same $\\mathbf{z}^i$ and obtain $(\\tilde{\\mathbf{w}}_a^i,\\tilde{\\mathbf{w}}_b^i)$ from the mapping network, which we stack in batches of augmented latents ( $\\tilde{\\mathbf{W}}_{\\mathbf{A}}$ and $\\tilde{\\mathbf{W}}_{\\mathbf{B}}$ ). The twin $(\\tilde{\\mathbf{w}}_a^i,\\tilde{\\mathbf{w}}_b^i)$ vectors are then made invariant to augmentations (similar) in the latent space by minimizing cross-correlation [2,53] between the latents of two augmented batches ( $\\tilde{\\mathbf{W}}_{\\mathbf{A}}$ and $\\tilde{\\mathbf{W}}_{\\mathbf{B}}$ )." + ], + "image_footnote": [], + "bbox": [ + 78, + 88, + 893, + 256 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Although that does lead to diverse image generation as $\\mathbf{w}$ are diverse; however, this does not ensure that these $\\mathbf{w}$ will generate class-consistent outputs for augmentations in embedding $(\\tilde{\\mathbf{c}})$ . To ensure class consistent predictions, we need to ensure invariance in $\\mathbf{w}$ to noise augmentation.", + "bbox": [ + 75, + 330, + 470, + 406 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For enforcing invariance to augmentations, a set of recent works [2, 10, 53] in self-supervised learning make the representations of augmentations similar through regularization. Among them, we focus on Barlow Twins as it does not require a large batch size of samples. Inspired by Barlow twins, we introduce NoisyTwins (Fig. 4), where we generate twin augmentations $\\tilde{\\mathbf{c}}_a$ and $\\tilde{\\mathbf{c}}_b$ of the same class embedding $(\\mu_c)$ and concatenate them to same $\\mathbf{z}$ . After creating a batch of such inputs, they are passed to the mapping network to get batches of augmented latents $(\\tilde{\\mathbf{W}}_A$ and $\\tilde{\\mathbf{W}}_B)$ . These batches are then used to calculate the cross-correlation matrix of latent variables given as:", + "bbox": [ + 75, + 409, + 473, + 592 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {C} _ {j, k} = \\frac {\\sum_ {(\\tilde {\\mathbf {w}} _ {a} , \\tilde {\\mathbf {w}} _ {b}) \\in (\\tilde {\\mathbf {W}} _ {A} , \\tilde {\\mathbf {W}} _ {B})} \\tilde {\\mathbf {w}} _ {a} ^ {j} \\tilde {\\mathbf {w}} _ {b} ^ {k}}{\\sum_ {\\tilde {\\mathbf {w}} _ {a} \\in \\tilde {\\mathbf {W}} _ {A}} \\tilde {\\mathbf {w}} _ {a} ^ {j} \\tilde {\\mathbf {w}} _ {a} ^ {j} \\sum_ {\\tilde {\\mathbf {w}} _ {b} \\in \\tilde {\\mathbf {W}} _ {B}} \\tilde {\\mathbf {w}} _ {b} ^ {k} \\tilde {\\mathbf {w}} _ {b} ^ {k}} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 147, + 598, + 468, + 661 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The matrix $\\mathbf{C}$ is a square matrix of size same as of latents $\\mathbf{w}$ . The final loss based on confusion matrix is given as:", + "bbox": [ + 76, + 667, + 470, + 699 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathcal {N T}} = \\sum_ {j} \\left(1 - \\mathbf {C} _ {j j} ^ {2}\\right) + \\gamma \\sum_ {j \\neq k} \\mathbf {C} _ {j, k} ^ {2} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 707, + 468, + 741 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The first term tries to make the two latents $(\\tilde{\\mathbf{w}}_a$ and $\\tilde{\\mathbf{w}}_b)$ invariant to the noise augmentation applied (i.e. similar), whereas the second term tries to de-correlate the different variables, thus maximizing the information content of the embedding vector [53]. The $\\gamma$ is the hyper-parameter that determines the relative importance of the two terms. This loss is then added to the generator loss term $(\\mathcal{L}_{\\mathcal{G}} + \\lambda L_{\\mathcal{NT}})$ and optimized through backpropagation. The above procedure comprises our proposed method, NoisyTwins (Fig. 4), which we empirically evaluate in the subsequent sections.", + "bbox": [ + 75, + 750, + 472, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experimental Evaluation", + "text_level": 1, + "bbox": [ + 500, + 329, + 736, + 347 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Setup", + "text_level": 1, + "bbox": [ + 500, + 357, + 581, + 375 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets: We primarily apply all methods on long-tailed datasets, as GANs trained on them are more prone to class confusion and mode collapse. We first report on the commonly used CIFAR10-LT dataset with an imbalance factor (i.e. ratio of most to least frequent class) of 100. To show our approach's scalability and real-world application, we test our method on the challenging ImageNet-LT and iNaturalist 2019 datasets. The ImageNet-LT [30] is a long-tailed variant of the 1000 class ImageNet dataset, with a plethora of semantically similar classes (e.g., Dogs, Birds etc.), making it challenging to avoid class confusion. iNaturalist-2019 [48] is a real-world long-tailed dataset composed of 1010 different variants of species, some of which have fine-grained differences in their appearance. For such fine-grained datasets, ImageNet pre-trained discriminators [38, 39] may not be useful, as augmentations used to train the model makes it invariant to fine-grained changes.", + "bbox": [ + 496, + 388, + 893, + 645 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training Configuration: We use StyleGAN2 architecture for all our experiments. All our experiments are performed using PyTorch-StudioGAN implemented by Kang et al. [18], which serves as a base for our framework. We use Path Length regularization (PLR) as it ensures changes in $\\mathcal{W}$ space lead to changes in images, using a delayed PLR on ImageNet-LT following [39]. We use a batch size of 128 for all our experiments, with one G step per D step. Unless stated explicitly, we used the general training setup for StyleGANs from [18], including methods like $R_{1}$ regularization [31], etc. More details on the exact training configuration for each dataset are provided in Appendix.", + "bbox": [ + 496, + 651, + 893, + 834 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Metrics: In this work we use the following metrics for evaluation of our methods:", + "bbox": [ + 496, + 839, + 890, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "a) FID: Fréchet Inception Distance [12] is the Wasserstein-2 Distance between the real and validation data. We use a", + "bbox": [ + 500, + 869, + 893, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5991", + "bbox": [ + 482, + 944, + 513, + 957 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ff8ec5d9635145f686f068454e085ede3c0e1758b811ea2946bf87848554e0be.jpg", + "image_caption": [ + "Figure 5. Choice of Eval. backbone: intra-FID (iFID) of a class based on InceptionV3 backbone (left plot) is not able to capture the mode collapse (increase in iFID in the absence of mode collapse). This is well-captured by $\\mathrm{iFID}_{\\mathrm{CLIP}}$ based on CLIP [34] backbone (right plot, decrease in iFID in the absence of mode collapse)." + ], + "image_footnote": [], + "bbox": [ + 80, + 88, + 467, + 282 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "held-out validation set and $50\\mathrm{k}$ generated samples to evaluate FID in each case. As FID is biased towards ImageNet and can be arbitrarily manipulated, we also report $\\mathrm{FID}_{\\mathrm{CLIP}}$ . b) Precision & Recall: As we aim to mitigate the mode collapse and achieve diverse generations across classes, we use improved Precision & Recall [25] metrics, as poor recall indicates mode collapse [39].", + "bbox": [ + 75, + 383, + 467, + 488 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "d) Intra-Class $\\mathbf{FID}_{\\mathrm{CLIP}}$ ( $\\mathbf{iFID}_{\\mathrm{CLIP}}$ ): The usage of only FID based on Inception-V3 Networks for evaluation of Generative Models has severe limitations, as it has been found that FID can be reduced easily by some fringe features [26]. iFID is computed by taking FID between 5k generated and real samples for the same class. As we want to evaluate both class consistency and diversity, we find that similar limitations exist for intra-class FID (iFID), which has been used to evaluate class-conditional GANs [18]. In Fig. 5, we show the existence of generated images for a particular class (more in Appendix) from models trained on iNaturalist 2019, where iFID is better for the mode collapsed model than the other model generating diverse images. Whereas the $\\mathbf{iFID}_{\\mathrm{CLIP}}$ , based on CLIP backbone can rank the models correctly with the model having mode collapse having high $\\mathbf{iFID}_{\\mathrm{CLIP}}$ . Further, we find that the mean iFID can be deceptive in detecting class confusion and collapse cases, as it sometimes ranks models with high realism better than models generating diversity (See Appendix). Hence, mean $\\mathbf{iFID}_{\\mathrm{CLIP}}$ (ref. to as $\\mathbf{iFID}_{\\mathrm{CLIP}}$ in result section for brevity) can be reliably used to evaluate models for class consistency and diversity.", + "bbox": [ + 75, + 489, + 467, + 821 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines: For evaluating NoisyTwins performance in comparison to other methods, we use the implementations present in StudioGAN [16]. For fairness, we re-run all the baselines on StyleGAN2 in the same hyperparameter setting. We compare our method to the StyleGAN2", + "bbox": [ + 75, + 824, + 467, + 898 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "(SG2) and StyleGAN2 with augmentation (DiffAug [54] and ADA [20]) baselines. We further tried to improve the baseline by incorporating the recent LeCam regularization method; however, it resulted in gains only for the iNaturalist 2019 dataset, where we use LeCam for all experiments. Further on StyleGAN2, we also use contrastive D2D-CE loss conditioning (i.e. ReACGAN) as a baseline. However, the D2D-CE baseline completely ignores learning of tail classes (Fig. 3) for CIFAR10-LT and is expensive to train; hence we do not report results for it for large-scale long-tailed datasets. We also compare our method against the recent SotA group Spectral Normalization (gSR) [35] method, which we implement for StyleGAN2 by constraining the spectral norms of embedding parameters of the generator $(\\mathcal{G})$ as suggested by authors. As a sanity check, we reproduce their results on CIFAR10-LT and find that our implementation matches the reported results correctly. We provide results on all datasets, for both the proposed Noise Augmentation (+ Noise) and the overall proposed NoisyTwins (+NoisyTwins) method.", + "bbox": [ + 496, + 90, + 890, + 393 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Results on Long-Tailed Data Distributions", + "text_level": 1, + "bbox": [ + 498, + 406, + 859, + 422 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "CIFAR10-LT. We applied DiffAug [54] on all baselines, except on gSR, where we found that DiffAug provides inferior results compared to ADA (as also used by authors [35]). It can be observed in Table 2 that the addition of NoisyTwins regularization significantly improves over baseline by $(\\sim 14$ FID) along with providing superior class consistency as shown by improved iFIDCLIP. NoisyTwins is also able to outperform the recent gSR regularization method and achieves improved results for all metrics. Further, NoisyTwins improves FID for StyleGAN2-ADA baseline used by gSR too from 32.08 to 23.02, however the final results are inferior than reported DiffAug baseline results. Further, we observed that despite not producing any tail class images (Fig. 3), the D2D-CE baseline has much superior FID in comparison to baselines. Whereas the proposed iFIDCLIP value is similar for the baseline and D2D-CE model. This clearly demonstrates the superiority of proposed iFIDCLIP in detecting class confusion.", + "bbox": [ + 496, + 430, + 890, + 702 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Real-world, Large-scale Long-Tailed Datasets. We experiment with iNaturalist 2019 and ImageNet-LT. These datasets are particularly challenging as they contain long-tailed imbalances and semantically similar classes, making GANs prone to mode collapse and class confusion. The baselines StyleGAN2 and StyleGAN2-ADA both suffer from mode collapse (Fig. 6), particularly for the tail classes. Whereas for the recent SotA gSR method, we find that although it undergoes less collapse in comparison to baselines, it suffers from class confusion as seen from similar Intra-FIDCLIP in comparison to baselines (Table 1). Compared to that, our method NoisyTwins improves when used with StyleGAN2-ADA significantly, leading to", + "bbox": [ + 496, + 704, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "5992", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/7f2c2981b2ff57d28c890a404f7b9e6ce714d9194ac3f1ebdba1bae7f3751a92.jpg", + "table_caption": [ + "Table 1. Quantitative results on ImageNet-LT and iNaturalist 2019 Datasets. We compare FID(↓), FIDCLIP(↓), iFIDCLIP(↓), Precision(↑) and Recall(↑) with other existing approaches on StyleGAN2 (SG2). We obtain an average ~ 19% relative improvement on FID, ~ 33% on FIDCLIP, and ~ 11% on iFIDCLIP metrics over the previous SotA on ImageNet-LT and iNaturalist 2019 datasets." + ], + "table_footnote": [], + "table_body": "
MethodImageNet-LTiNaturalist 2019
FID(↓)FIDCLIP(↓)iFIDCLIP(↓)Precision(↑)Recall(↑)FID(↓)FIDCLIP(↓)iFIDCLIP(↓)Precision(↑)Recall(↑)
SG2 [23]41.2511.6446.930.500.4819.343.3338.240.740.17
SG2+ADA [20]37.2011.0447.410.540.3814.922.3035.190.750.57
SG2+ADA+gSR [35]24.788.2144.420.630.3515.172.0636.220.740.46
SG2+ADA+Noise (Ours)22.177.1141.200.720.3312.871.3731.430.810.63
+ NoisyTwins (Ours)21.296.4139.740.670.4911.461.1431.500.790.67
", + "bbox": [ + 78, + 133, + 890, + 232 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7404f5e305388745ff6e4780caa40d6c80c7de24d1f9ef8a6fb61b7ff1a13bb0.jpg", + "table_caption": [ + "Table 2. Quantitative results on CIFAR10-LT Dataset. We compare with other existing approaches. We obtain $\\sim 26\\%$ relative improvement over the existing methods on $\\mathrm{FID}_{\\mathrm{CLIP}}$ and iFIDCLIP metrics." + ], + "table_footnote": [], + "table_body": "
MethodFID(↓)FIDCLIP(↓)iFIDCLIP(↓)Precision(↑)Recall(↑)
SG2+DiffAug [54]31.736.2711.590.630.35
SG2+D2D-CE [17]19.974.7711.350.730.42
gSR [35]22.105.549.940.700.29
SG2+DiffAug+Noise (Ours)28.905.2610.650.710.38
+ NoisyTwins(Ours)17.743.557.240.700.51
", + "bbox": [ + 80, + 292, + 604, + 383 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d01ec9abc039336a7c583af412b88ec40c084f42846e413a87a3cc8e73844de4.jpg", + "table_caption": [ + "Table 3. Comparison with SotA approaches on BigGAN. We compare FID(↓) with other existing models on ImageNet-LT (IN-LT) and iNaturalist 2019 (iNat-19)." + ], + "table_footnote": [], + "table_body": "
MethodiNat-19IN-LT
BigGAN [4]14.8528.10
+ gSR [35]13.95-
ICGAN [6]-23.40
StyleGAN2-ADA [20]14.9237.20
+ NoisyTwins (Ours)11.4621.29
", + "bbox": [ + 622, + 292, + 867, + 383 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "a relative improvement of $42.7\\%$ in FID for ImageNet-LT and $23.19\\%$ on the iNaturalist 2019 dataset when added to StyleGAN2-ADA baseline. Further with Noise Augmentation $(+\\mathrm{Noise})$ , we observe generations of high-quality class-consistent images, but it also suffers from mode collapse. This can be observed by the high-precision values in comparison to low-recall values. However, adding NoisyTwins regularization over the noise augmentation improves diversity by improving recall (Table 1).", + "bbox": [ + 75, + 393, + 470, + 530 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fig. 6 presents the generated images of tail classes for various methods on ImageNet-LT, where NoisyTwins generations show remarkable diversity in comparison to others. The presence of diversity for classes with just 5-6 training images demonstrates successful transfer of knowledge from head classes to tail classes, due to shared parameters. Further, to compare with existing SotA reported results, we compare FID of BigGAN models from gSR [35] and Instance Conditioned GAN (ICGAN) [6]. For fairness, we compare FID on the validation set for which we obtained gSR models from authors and re-evaluate them, as they reported FID on a balanced training set. As BigGAN models are more common for class-conditioned generation [18], their baseline performs superior to StyleGAN2-ADA baselines (Table 3). However, the addition of NoisyTwins to the StyleGAN2-ADA method improves it significantly, even outperforming the existing methods of gSR (by $18.44\\%$ ) and ICGAN (by $9.44\\%$ ) based on BigGAN architecture. This shows that NoisyTwins allows the StyleGAN2 baseline to scale to large and diverse long-tailed datasets.", + "bbox": [ + 75, + 531, + 468, + 834 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. NoisyTwins on Few-Shot Datasets", + "text_level": 1, + "bbox": [ + 76, + 845, + 372, + 861 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We now demonstrate the potential of NoisyTwins in another challenging scenario of class-conditional few-shot", + "bbox": [ + 76, + 869, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "image generation from GANs. We perform our experiments using a conditional StyleGAN2-ADA baseline, for which we tune hyper-parameters to obtain a strong baseline. We then apply our method of Noise Augmentation and NoisyTwins over the strong baseline for reporting our results. We use the few-shot dataset of LHI-AnimalFaces [45] and a subset of ImageNet Carnivores [29, 40] to report our results. Table 4 shows the results of these experiments, where we find that our method, NoisyTwins, significantly improves the FID of StyleGAN2 ADA baseline by $(22.2\\%)$ on average for both datasets. Further, combining Noisy Twins with SotA Transitional-cGAN [40] through official code, also leads to effective improvement in FID. These results clearly demonstrate the diverse potential and applicability of our proposed method NoisyTwins.", + "bbox": [ + 496, + 393, + 893, + 623 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6. Analysis", + "text_level": 1, + "bbox": [ + 500, + 641, + 596, + 657 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We perform analysis of NoisyTwins w.r.t. to its hyperparameters, standard deviation $(\\sigma)$ of noise augmentation and regularization strength $(\\lambda)$ . We also compare NoisyTwins objective (ref. Eq. 6) with contrastive objective. Finally, we compare NoisyTwins over Latent Diffusion Models for long-tailed class conditional generation task. We perform ablation experiments on CIFAR10-LT, for which additional details and results are present in Appendix. We also present comparison of NoisyTwins for GAN fine-tuning.", + "bbox": [ + 496, + 667, + 893, + 805 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "How much noise and regularization strength is optimal?", + "text_level": 1, + "bbox": [ + 496, + 809, + 890, + 824 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Fig. 7, we ablate over the noise variance parameter $\\sigma$ for CIFAR10-LT. We find that a moderate value of noise strength 0.75 leads to optimal results. For the strength of NoisyTwins loss $(\\lambda)$ , we find that the algorithm performs similarly on values near 0.01 and is robust to it (Fig. 7).", + "bbox": [ + 496, + 825, + 893, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "5993", + "bbox": [ + 482, + 944, + 514, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/71c2871399f1044f6cba014b9cef28dfd0f65057b005f7d878839c3bc5d097ce.jpg", + "image_caption": [ + "(a) Image Augmentation based Methods (ADA)" + ], + "image_footnote": [], + "bbox": [ + 83, + 89, + 356, + 250 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/81ede3449877dabd306eeb752501b5a369e7735973a56bdf84fc7984f41a5836.jpg", + "image_caption": [ + "(b) Regularizer based Methods (gSR)" + ], + "image_footnote": [], + "bbox": [ + 366, + 90, + 522, + 250 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6d18060b187b405762c6daf9bd05455025eb712304f459eba2dd10d557b74dd1.jpg", + "image_caption": [ + "(c) NoisyTwins (Ours)" + ], + "image_footnote": [], + "bbox": [ + 534, + 90, + 887, + 250 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/2e4dd24d72727438c148f93e6ec702fcb0c6345ff0fee142bc18fa2d30ead438.jpg", + "table_caption": [ + "Table 4. Quantitative results on ImageNet Carnivore and AnimalFace Datasets. Our method improves over both StyleGAN2-ADA (SG2-ADA) baseline and SotA Transitional-cGAN." + ], + "table_footnote": [], + "table_body": "
ImageNet CarnivoreAnimalFace
MethodFID(↓)iFIDCLIP(↓)FID(↓)iFIDCLIP(↓)
SG2 [23]111.8336.3494.0929.94
SG2+ADA [20]22.7712.8520.2511.12
SG2+ADA+Noise (Ours)19.2512.5118.7810.42
+ NoisyTwins (Ours)16.0112.4117.2710.03
FID(↓)FID(↓)
Transitional-cGAN [40]14.6020.53
+ NoisyTwins (Ours)13.6516.15
", + "bbox": [ + 81, + 388, + 483, + 522 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Which type of self-supervision to use with noise augmentation? The goal of our method is to achieve invariance to Noise Augmentation in the $\\mathcal{W}$ latent space. This can be achieved using either contrastive learning-based methods like SimCLR [7] or negative-free method like Barlow Twins [53]. Contrastive loss (SimCLR based) produces FID of 26.23 vs 17.74 by NoisyTwins (BarlowTwins based). We find that contrastive baseline improves over the noise augmentation baseline (28.90) however falls significantly below the NoisyTwins, as the former requires a large batch size to be effective which is expensive for GANs.", + "bbox": [ + 75, + 553, + 468, + 719 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "How does NoisyTwins compare with modern Vision and Language models? For evaluating the effectiveness of modern vision language-based diffusion models, we test the generation of the iNaturalist 2019 dataset by creating the prompt \"a photo of S\" where we replace the class name in place of S. We use the LDM [37] model trained on LAION-400M to perform inference, generating 50 images per class. We obtained an FID of 57.04 in comparison to best FID of 11.46 achieved by NoisyTwins. This clearly demonstrates that for specific use cases like fine-grained generation, GANs are still ahead of general-purpose LDM.", + "bbox": [ + 75, + 734, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/57d1a50ade434231c98ee56695e446f0c59a3da443295139f4ff82d16eb117e0.jpg", + "image_caption": [ + "Figure 6. Qualitative results on ImageNet-LT for tail classes. We find that existing SotA methods for tail classes show collapsed (a) or arbitrary image generation (b). With NoisyTwins, we observe diverse and class-consistent image generation, even for classes having 5-6 images. The tail classes get enhanced diversity by transferring the knowledge from head classes, as they share parameters.", + "Figure 7. Ablation of Hyperparameters. Quantitative comparison on CIFAR10-LT for standard deviation of Noise Augmentation $(\\sigma)$ and strength $(\\lambda)$ of NoisyTwins loss." + ], + "image_footnote": [], + "bbox": [ + 500, + 335, + 890, + 489 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 546, + 619, + 561 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we analyze the performance of StyleGAN2 models on the real-world long-tailed datasets including iNaturalist 2019 and ImageNet-LT. We find that existing works lead to either class confusion or mode collapse in the image space. This phenomenon is rooted in collapse and confusion in the latent $\\mathcal{W}$ space of StyleGAN2. Through our analysis, we deduce that this collapse occurs when the latents become invariant to random conditioning vectors $\\mathbf{z}$ , and collapse for each class. To mitigate this, we introduce inexpensive noise based augmentation for discrete class embeddings. Further, to ensure class consistency, we couple this augmentation technique with BarlowTwins' objective in the latent $\\mathcal{W}$ space which imparts intra-class diversity to latent $\\mathbf{w}$ vectors. The noise augmentation and regularization comprises our proposed NoisyTwins technique, which improves the performance of StyleGAN2 establishing a new SotA on iNaturalist 2019 and ImageNet-LT. The extension of NoisyTwins for conditioning on more-complex attributes for StyleGANs is a good direction for future work. Acknowledgements: This work was supported in part by SERB-STAR Project (STR/2020/000128). Harsh Rangwani is supported by PMRF fellowship.", + "bbox": [ + 496, + 568, + 890, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "5994", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Yuval Alaluf, Omer Tov, Ron Mokady, Rinon Gal, and Amit Bermano. Hyperstyle: Stylegan inversion with hypernetworks for real image editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18511-18521, June 2022. 2", + "[2] Adrien Bardes, Jean Ponce, and Yann LeCun. VI-CReg: Variance-invariance-covariance regularization for self-supervised learning. In International Conference on Learning Representations, 2022. 2, 5", + "[3] Stella Bounareli, Vasileios Argyriou, and Georgios Tzimiropoulos. Finding directions in gan's latent space for neural face reenactment. arXiv preprint arXiv:2202.00046, 2022.1", + "[4] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale GAN training for high fidelity natural image synthesis. In International Conference on Learning Representations, 2019. 7", + "[5] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. 2020. 3", + "[6] Arantxa Casanova, Marlene Careil, Jakob Verbeek, Michal Drozdal, and Adriana Romero-Soriano. Instance-conditioned gan. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 3, 7", + "[7] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. arXiv preprint arXiv:2002.05709, 2020.8", + "[8] Yin Cui, Menglin Jia, Tsung-Yi Lin, Yang Song, and Serge Belongie. Class-balanced loss based on effective number of samples. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9268-9277, 2019. 4", + "[9] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139-144, 2020. 3", + "[10] Jean-Bastien Grill, Florian Strub, Florent Altché, Coretin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 5", + "[11] Erik Härkönen, Aaron Hertzmann, Jaakko Lehtinen, and Sylvain Paris. Ganspace: Discovering interpretable gan controls. Advances in Neural Information Processing Systems, 33:9841-9850, 2020. 1", + "[12] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 2, 5", + "[13] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceed-" + ], + "bbox": [ + 78, + 114, + 472, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ings of the IEEE international conference on computer vision, pages 1501-1510, 2017. 3", + "[14] Jongheon Jeong and Jinwoo Shin. Training GANs with stronger augmentations via contrastive discriminator. In International Conference on Learning Representations, 2021. 3", + "[15] Liming Jiang, Bo Dai, Wayne Wu, and Chen Change Loy. Deceive D: Adaptive Pseudo Augmentation for GAN training with limited data. In NeurIPS, 2021. 2", + "[16] Minguk Kang and Jaesik Park. ContraGAN: Contrastive Learning for Conditional Image Generation. 2020. 3, 6", + "[17] Minguk Kang, Woohyeon Shim, Minsu Cho, and Jaesik Park. Rebooting acgan: Auxiliary classifier gans with stable training. Advances in neural information processing systems, 34:23505-23518, 2021. 2, 3, 7", + "[18] MinGuk Kang, Joonghyuk Shin, and Jaesik Park. StudioGAN: A Taxonomy and Benchmark of GANs for Image Synthesis. 2206.09479 (arXiv), 2022. 5, 6, 7", + "[19] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. arXiv preprint arXiv:1710.10196, 2017. 3", + "[20] Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. In Proc. NeurIPS, 2020. 1, 2, 3, 6, 7, 8", + "[21] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 1, 2, 3", + "[22] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. 1", + "[23] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. In Proc. CVPR, 2020. 3, 7, 8", + "[24] Tuomas Kynkänniemi, Tero Karras, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Improved Precision and Recall Metric for Assessing Generative Models. 2019. 1", + "[25] Tuomas Kynkänniemi, Tero Karras, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Improved precision and recall metric for assessing generative models. Advances in Neural Information Processing Systems, 32, 2019. 6", + "[26] Tuomas Kynkänniemi, Tero Karras, Miika Aittala, Timo Aila, and Jaakko Lehtinen. The role of imagenet classes in fréchet inception distance. CoRR, abs/2203.06026, 2022. 6", + "[27] Thomas Leimkuhler and George Drettakis. Freestylegan: Free-view editable portrait rendering with the camera manifold. 40(6), 2021. 2", + "[28] Feng Liu and Xiaoming Liu. 2d gans meet unsupervised single-view 3d reconstruction. In Computer Vision - ECCV 2022, pages 497-514, 2022. 2", + "[29] Ming-Yu Liu, Xun Huang, Arun Mallya, Tero Karras, Timo Aila, Jaakko Lehtinen, and Jan Kautz. Few-shot unsupervised image-to-image translation. In Proceedings of" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "5995", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the IEEE/CVF international conference on computer vision, pages 10551-10560, 2019. 7", + "[30] Ziwei Liu, Zhongqi Miao, Xiaohang Zhan, Jiayun Wang, Boqing Gong, and Stella X Yu. Large-scale long-tailed recognition in an open world. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2537-2546, 2019. 2, 5", + "[31] Lars Mescheder, Andreas Geiger, and Sebastian Nowozin. Which training methods for gans do actually converge? In International conference on machine learning, pages 3481-3490. PMLR, 2018. 5", + "[32] Rishubh Parihar, Ankit Dhiman, Tejan Karmali, and R. Venkatesh Babu. Everything is there in latent space: Attribute editing and attribute style manipulation by stylegan latent space exploration. In Proceedings of the 30th ACM International Conference on Multimedia, pages 1828-1836, 2022. 2", + "[33] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2085–2094, October 2021. 2", + "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2, 6", + "[35] Harsh Rangwani, Naman Jaswani, Tejan Karmali, Varun Jampani, and R. Venkatesh Babu. Improving gans for long-tailed data through group spectral regularization. In European Conference on Computer Vision, 2022. 2, 3, 4, 6, 7", + "[36] Harsh Rangwani, Konda Reddy Mopuri, and R Venkatesh Babu. Class balancing gan with a classifier in the loop. In Uncertainty in Artificial Intelligence, pages 1618-1627. PMLR, 2021. 3", + "[37] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 8", + "[38] Axel Sauer, Kashyap Chitta, Jens Müller, and Andreas Geiger. Projected gans converge faster. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 1, 5", + "[39] Axel Sauer, Katja Schwarz, and Andreas Geiger. Styleganx1: Scaling stylegan to large diverse datasets. volume abs/2201.00273, 2022. 1, 5, 6", + "[40] Mohamad Shahbazi, Martin Danelljan, Danda Pani Paudel, and Luc Van Gool. Collapse by conditioning: Training class-conditional GANs with limited data. In International Conference on Learning Representations, 2022. 3, 7, 8", + "[41] Yujun Shen, Jinjin Gu, Xiaou Tang, and Bolei Zhou. Interpreting the latent space of gans for semantic face editing. In CVPR, 2020. 1", + "[42] Yujun Shen, Ceyuan Yang, Xiaou Tang, and Bolei Zhou. Interfacegan: Interpreting the disentangled face representation learned by gans. IEEE TPAMI, 2020. 2", + "[43] Yujun Shen and Bolei Zhou. Closed-form factorization of latent semantics in gans. In CVPR, 2021. 2" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[44] Yichun Shi, Divyansh Aggarwal, and Anil K. Jain. Lifting 2d stylegan for 3d-aware face generation. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6254-6262, 2021. 2", + "[45] Zhangzhang Si and Song-Chun Zhu. Learning hybrid image templates (hit) by information projection. IEEE Transactions on pattern analysis and machine intelligence, 34(7):1354-1367, 2011. 7", + "[46] Ivan Skorokhodov, Sergey Tulyakov, and Mohamed Elhoseiny. Stylegan-v: A continuous video generator with the price, image quality and perks of stylegan2, 2021. 1", + "[47] Hung-Yu Tseng, Lu Jiang, Ce Liu, Ming-Hsuan Yang, and Weilong Yang. Regularizing generative adversarial networks under limited data. In CVPR, 2021. 3", + "[48] Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alex Shepard, Hartwig Adam, Pietro Perona, and Serge Belongie. The inaturalist species classification and detection dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8769-8778, 2018. 2, 5", + "[49] Zongze Wu, Dani Lischinski, and Eli Shechtman. Stylespace analysis: Disentangled controls for stylegan image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12863–12872, 2021. 1, 3", + "[50] Ceyuan Yang, Yujun Shen, Yinghao Xu, Deli Zhao, Bo Dai, and Bolei Zhou. Improving gans with a dynamic discriminator. 2022. 3", + "[51] Ceyuan Yang, Yujun Shen, Yinghao Xu, and Bolei Zhou. Data-efficient instance generation from instance discrimination. In A. Beygelzimer, Y. Dauphin, P. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, 2021. 3", + "[52] Sihyun Yu, Jihoon Tack, Sangwoo Mo, Hyunsu Kim, Junho Kim, Jung-Woo Ha, and Jinwoo Shin. Generating videos with dynamics-aware implicit generative adversarial networks. In International Conference on Learning Representations, 2022. 1", + "[53] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow twins: Self-supervised learning via redundancy reduction. In International Conference on Machine Learning, pages 12310-12320. PMLR, 2021. 2, 5, 8", + "[54] Shengyu Zhao, Zhijian Liu, Ji Lin, Jun-Yan Zhu, and Song Han. Differentiable augmentation for data-efficient gan training. In Conference on Neural Information Processing Systems (NeurIPS), 2020. 1, 2, 3, 6, 7" + ], + "bbox": [ + 501, + 92, + 890, + 739 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "5996", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/NoisyTwins_ Class-Consistent and Diverse Image Generation Through StyleGANs/c323d83e-f322-4c02-8ecc-1d0c86cf0f70_model.json b/2023/NoisyTwins_ Class-Consistent and Diverse Image Generation Through StyleGANs/c323d83e-f322-4c02-8ecc-1d0c86cf0f70_model.json new file mode 100644 index 0000000000000000000000000000000000000000..08d56039fda8f9e946197b8cecc2cf0e6d0f8c77 --- /dev/null +++ b/2023/NoisyTwins_ Class-Consistent and Diverse Image Generation Through StyleGANs/c323d83e-f322-4c02-8ecc-1d0c86cf0f70_model.json @@ -0,0 +1,2035 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.131, + 0.895, + 0.155 + ], + "angle": 0, + "content": "NoisyTwins: Class-Consistent and Diverse Image Generation through StyleGANs" + }, + { + "type": "text", + "bbox": [ + 0.18, + 0.18, + 0.791, + 0.217 + ], + "angle": 0, + "content": "Harsh Rangwani\\(^{1*}\\) Lavish Bansal\\(^{1,3*}\\) Kartik Sharma\\(^{1,4}\\) Tejan Karmali\\(^{2}\\) Varun Jampani\\(^{2}\\) R. Venkatesh Babu\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.217, + 0.871, + 0.235 + ], + "angle": 0, + "content": "\\(^{1}\\) Vision and AI Lab, IISc Bangalore \\(^{2}\\) Google Research \\(^{3}\\) IIT BHU Varanasi \\(^{4}\\) BITS Pilani" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.286 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.302, + 0.474, + 0.56 + ], + "angle": 0, + "content": "StyleGANs are at the forefront of controllable image generation as they produce a latent space that is semantically disentangled, making it suitable for image editing and manipulation. However, the performance of StyleGANs severely degrades when trained via class-conditioning on large-scale long-tailed datasets. We find that one reason for degradation is the collapse of latents for each class in the \\(\\mathcal{W}\\) latent space. With NoisyTwins, we first introduce an effective and inexpensive augmentation strategy for class embeddings, which then decorrelates the latents based on self-supervision in the \\(\\mathcal{W}\\) space. This decorrelation mitigates collapse, ensuring that our method preserves intraclass diversity with class-consistency in image generation. We show the effectiveness of our approach on large-scale real-world long-tailed datasets of ImageNet-LT and iNaturalist 2019, where our method outperforms other methods by \\(\\sim 19\\%\\) on FID, establishing a new state-of-the-art." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.591, + 0.21, + 0.607 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.617, + 0.473, + 0.875 + ], + "angle": 0, + "content": "StyleGANs [21, 22] have shown unprecedented success in image generation, particularly on well-curated and articulated datasets (eg. FFHQ for face images, etc.). In addition to generating high fidelity and diverse images, StyleGANs also produce a disentangled latent space, which is extensively used for image editing and manipulation tasks [49]. As a result, StyleGANs are being extensively used in various applications like face-editing [11, 41], video generation [46, 52], face reenactment [3], etc., which are a testament to their usability and generality. However, despite being successful on well-curated datasets, training StyleGANs on in-the-wild and multi-category datasets is still challenging. A large-scale conditional StyleGAN (i.e. StyleGAN-XL) on ImageNet was recently trained successfully by Sauer et al. [39] using the ImageNet pre-trained model through the idea of a projection discriminator [38]. While the StyleGAN-XL uses additional pre-trained mod" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.271, + 0.892, + 0.53 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.54, + 0.895, + 0.612 + ], + "angle": 0, + "content": "Figure 1. Qualitative Comparison on tail classes (T1-T4) for iNaturalist 2019. We provide sample(s) from real class (with class frequency), generated by StyleGAN2-ADA and after adding proposed NoisyTwins. NoisyTwins achieves remarkable diversity, class-consistency and quality by just using 38 samples on average." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.626, + 0.895, + 0.672 + ], + "angle": 0, + "content": "els, obtaining such models for distinctive image domains like medical, forensics, and fine-grained data may not be feasible, which limits its generalization across domains." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.895, + 0.903 + ], + "angle": 0, + "content": "In this work, we aim to train vanilla class-conditional StyleGAN without any pre-trained models on challenging real-world long-tailed data distributions. As training StyleGAN with augmentations [20, 54] leads to low recall [24] (which measures diversity in the generated images) and mode collapse, particularly for minority (i.e. tail) classes. For investigating this phenomenon further, we take a closer look at the latent \\(\\mathcal{W}\\) space of StyleGAN that is produced by a fully-connected mapping network that takes the conditioning variables \\(\\mathbf{z}\\) (i.e. random noise) and class embedding \\(\\mathbf{c}\\) as inputs. The vectors \\(\\mathbf{w}\\) in \\(\\mathcal{W}\\) space are used for conditioning various layers of the generator (Fig. 2). We find that output vectors \\(\\mathbf{w}\\) from the mapping network hinge on the conditioning variable \\(\\mathbf{c}\\) and become invariant to random conditioning vector \\(\\mathbf{z}\\). This collapse of latents leads to un" + }, + { + "type": "page_footnote", + "bbox": [ + 0.095, + 0.887, + 0.444, + 0.901 + ], + "angle": 0, + "content": "*Equal Contribution. Link: rangwani-harsh.github.io/NoisyTwins" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.958 + ], + "angle": 0, + "content": "5987" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.09, + 0.895, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.226, + 0.895, + 0.257 + ], + "angle": 0, + "content": "Figure 2. Schematic illustration of \\(\\mathcal{W}\\) space for different GANs. Existing conditioning methods either suffer from mode collapse [20] or lead to class confusion [35] in \\(\\mathcal{W}\\) space. With proposed NoisyTwins, we achieve intra class diversity while avoiding class confusion." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.263, + 0.471, + 0.354 + ], + "angle": 0, + "content": "stable training and is one of the causes of poor recall (a.k.a. mode collapse) for minority classes. Further, on augmenting StyleGAN with recent conditioning and regularization techniques [17, 35], we find that they either lead to a poor recall for minority classes or lead to class confusion (Fig. 2) instead of mitigating the collapse." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.358, + 0.47, + 0.54 + ], + "angle": 0, + "content": "To mitigate the collapse of \\(\\mathbf{w}\\) in \\(\\mathcal{W}\\) space, we need to ensure that the change in conditioning variable \\(\\mathbf{z}\\) leads to the corresponding change in \\(\\mathbf{w}\\). Recently in self-supervised learning, several techniques [2, 53] have been introduced to prevent the collapse of learned representations by maximizing the information content in the feature dimensions. Inspired by them we propose NoisyTwins, in which we first generate inexpensive twin augmentations for class embeddings and then use them to decorrelate the \\(\\mathbf{w}\\) variables through self-supervision. The decorrelation ensures that \\(\\mathbf{w}\\) vectors are diverse for each class and the GAN is able to produce intra-class diversity among the generated images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.544, + 0.471, + 0.876 + ], + "angle": 0, + "content": "We evaluate NoisyTwins on challenging benchmarks of large-scale long-tailed datasets of ImageNet-LT [30] and iNaturalist 2019 [48]. These benchmarks are particularly challenging due to a large number of classes present, which makes GANs prone to class confusion. On the other hand, as these datasets are long-tailed with only a few images per class in tail classes, generating diverse images for those classes is challenging. We observe that existing metrics used in GAN evaluations are not able to capture both class confusion and mode collapse. As a remedy, we propose to use intra-class Frechet Inception Distance (FID) [12] based on features obtained from pre-trained CLIP [34] embeddings as an effective metric to measure the performance of class-conditional GANs in long-tailed data setups. Using NoisyTwins enables StyleGAN to generate diverse and class-consistent images across classes, mitigating the mode collapse and class confusion issues in existing state-of-the-art (SotA) (Fig. 1). Further, with NoisyTwins, we obtain diverse generations for tail classes even with \\(\\leq 30\\) images, which can be attributed to the transfer of knowledge from head classes through shared parameters (Fig. 1 and 6). In summary, we make the following contributions:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.886, + 0.471, + 0.902 + ], + "angle": 0, + "content": "1. We evaluate various recent SotA GAN conditioning" + }, + { + "type": "text", + "bbox": [ + 0.531, + 0.263, + 0.892, + 0.323 + ], + "angle": 0, + "content": "and regularization techniques on the challenging task of long-tailed image generation. We find that all existing methods either suffer from mode collapse or lead to class confusion in generations." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.324, + 0.892, + 0.383 + ], + "angle": 0, + "content": "2. To mitigate mode collapse and class confusion, we introduce NoisyTwins, an effective and inexpensive augmentation strategy for class embeddings that decorrelates latents in the \\(\\mathcal{W}\\) latent space (Sec. 4)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.384, + 0.892, + 0.474 + ], + "angle": 0, + "content": "3. We evaluate NoisyTwins on large-scale long-tailed datasets of ImageNet-LT and iNaturalist-2019, where it consistently improves the StyleGAN2 performance \\((\\sim 19\\%)\\), achieving a new SotA. Further, our approach can also prevent mode collapse and enhance the performance of few-shot GANs (Sec. 5.3)." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.324, + 0.892, + 0.474 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.49, + 0.651, + 0.506 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.519, + 0.892, + 0.746 + ], + "angle": 0, + "content": "StyleGANs. Karras et al. introduced StyleGAN [21] and subsequently improved its image quality in StyleGAN2. StyleGAN could produce high-resolution photorealistic images as demonstrated on a wide variety of category-specific datasets. It introduced a mapping network, which mapped the sampled noise into another latent space, which is more disentangled and semantically coherent, as demonstrated by its downstream usage for image editing and manipulation [1,32,33,42,43]. Further, StyleGAN has been extended to get novel views from images [27, 28, 44], thus making it possible to get 3D information from it. These downstream advances are possible due to the impressive performance of StyleGANs on class-specific datasets (such as faces). However, similar photorealism levels are yet uncommon on multi-class long-tailed datasets (such as ImageNet)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.901 + ], + "angle": 0, + "content": "GANs for Data Efficiency and Imbalance. Failure of GANs on less data was concurrently reported by Karras et al. [20] and Zhao et al. [54]. The problem is rooted in the overfitting of the discriminator due to less real data. Since then, the proposed solutions for this problem have relied on a) augmenting the data, b) introducing regularizers, and c) architectural modifications. Karras et al. [20] and Zhao et al. [54] relied on differentiable data augmentation before passing images into the discriminator to solve this problem. DeceiveD [15] proposed to introduce label-noise for" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5988" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.47, + 0.198 + ], + "angle": 0, + "content": "discriminator training. LeCamGAN [47] finds that enforcing LeCam divergence as a regularization trick in the discriminator can robustify GAN training under a limited data setting. DynamicD [50] tunes the capacity of the discriminator on-the-fly during training. While these methods can handle the data inefficiency, they are ineffective on class imbalanced long-tailed data distribution [35]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.198, + 0.471, + 0.395 + ], + "angle": 0, + "content": "CBGAN [36] proposed a solution to train the unconditional GAN model on long-tailed data distribution by introducing a signal from the classifier to balance the classes generated by GAN. In a long-tailed class-conditional setting, gSR [35] proposes to regularize the exploding spectral norms of the class-specific parameters of the GAN. Collapse-by-conditioning [40] addresses the limited data in classes by introducing a training regime that transitions from an unconditional to a class-conditioned setting, thus exploiting the shared information across classes during the early stages of the training. However, these methods suffer from either class confusion or poor generated image quality on large datasets, which is resolved by NoisyTwins." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.399, + 0.47, + 0.611 + ], + "angle": 0, + "content": "Self-Supervised Learning for GANs. Ideas from Self-supervised learning have shown their benefits in GAN training. IC-GAN [6] trains GAN conditioned on embeddings on SwAV [5], which led to remarkable improvement in performance on the long-tailed version of ImageNet. InsGen [51] and ReACGAN [16, 17] introduce the auxiliary task of instance discrimination for the discriminator, thereby making the discriminator focus on multiple tasks and thus alleviating discriminator overfitting. While InsGen relies on both noise space and image space augmentations, ReACGAN and ContraGAN follow only image space augmentations. Contrary to these, NoisyTwins performs augmentations in the class-embedding space and contrasts them in the \\(\\mathcal{W}\\)-space of the generator instead of the discriminator." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.625, + 0.216, + 0.64 + ], + "angle": 0, + "content": "3. Preliminaries" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.65, + 0.192, + 0.666 + ], + "angle": 0, + "content": "3.1. StyleGAN" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.674, + 0.471, + 0.901 + ], + "angle": 0, + "content": "StyleGAN [21] is a Generative Adversarial Network comprising of its unique Style Conditioning Based Generator \\((\\mathcal{G})\\) and discriminator network \\((\\mathcal{D})\\) trained jointly. We will focus on the architecture of StyleGAN2 [23] as we use it in our experiments, although our work is generally applicable to all StyleGAN architectures. The StyleGAN2 generator is composed of blocks that progressively upsample the features and resolution inspired by Progressive GAN [19], starting from a single root image. The diversity in the images comes from conditioning each block of image generation through conditioning on the latent coming from the mapping network (Fig. 2). The mapping network is a fully connected network that takes in the conditioning variables, the \\(\\mathbf{z} \\in \\mathbb{R}^d\\) coming from a random distribution (e.g., Gaussian, etc.) and class conditioning label \\(c\\) which is converted" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.091, + 0.892, + 0.243 + ], + "angle": 0, + "content": "to an embedding \\(\\mathbf{c} \\in \\mathbb{R}^d\\). The mapping network takes these and outputs vectors \\(\\mathbf{w}\\) in the \\(\\mathcal{W}\\) latent space of StyleGAN, which is found to be semantically disentangled to a high extent [49]. The \\(\\mathbf{w}\\) is then processed through an affine transform and passed to each generator layer for conditioning the image generation process through Adaptive Instance Normalization (AdaIN) [13]. The images from generator \\(\\mathcal{G}\\), along with real images, are passed to discriminator \\(\\mathcal{D}\\) for training. The training utilizes the non-saturating adversarial losses [9] for \\(\\mathcal{G}\\) and \\(\\mathcal{D}\\) given as:" + }, + { + "type": "equation", + "bbox": [ + 0.513, + 0.244, + 0.892, + 0.283 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\mathcal {D}} \\mathcal {L} _ {\\mathcal {D}} = \\sum_ {i = 1} ^ {m} \\log (\\mathcal {D} (\\mathbf {x} _ {i})) + \\log (1 - \\mathcal {D} (\\mathcal {G} (\\mathbf {z} _ {i}, \\mathbf {c} _ {i}))) \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.286, + 0.892, + 0.325 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\mathcal {G}} \\mathcal {L} _ {\\mathcal {G}} = \\sum_ {i = 1} ^ {m} - \\log \\left(\\mathcal {D} \\left(\\mathcal {G} \\left(\\mathbf {z} _ {i}, \\mathbf {c} _ {i}\\right)\\right)\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.326, + 0.892, + 0.357 + ], + "angle": 0, + "content": "We now describe the issues present in the StyleGANs trained on long-tailed data and their analysis in \\(\\mathcal{W}\\) space." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.364, + 0.892, + 0.396 + ], + "angle": 0, + "content": "3.2. Class Confusion and Class-Specific Mode Collapse in Conditional StyleGANs" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.402, + 0.892, + 0.901 + ], + "angle": 0, + "content": "To finely analyze the performance of StyleGAN and its variants on long-tailed datasets, we train them on the CIFAR10-LT dataset. In Fig. 3, we plot the qualitative results of generated images and create a t-SNE plot for latents in \\(\\mathcal{W}\\) space for each class. We first train the StyleGAN2 baseline with augmentations (DiffAug) [20, 54]. We find that it leads to mode collapse, specifically for tail classes (Fig. 3). In conjunction with images, we also observe that corresponding t-SNE embeddings are also collapsed near each class's mean in \\(\\mathcal{W}\\) space. Further, recent methods which have proposed the usage of contrastive learning for GANs, improve their data efficiency and prevent discriminator overfitting [14, 16]. We also evaluate them by adding the contrastive conditioning method, which is D2DCE loss-based on ReACGAN [17], to the baseline, where in results, we observe that the network omits to learn tail classes and produces head class images at their place (i.e., class confusion). In Fig. 3, it can be seen that the network confuses semantically similar classes, that is, generating cars (head or majority class) in place of trucks and airplanes (head class) instead of ships. In the \\(\\mathcal{W}\\) space, we find the same number of clusters as the number of classes in the dataset. However, the tail label cluster images also belong to the head classes of cars and airplanes. In a very recent work gSR [35], it has been shown that constraining the spectral norm of \\(\\mathcal{G}\\) embedding parameters can help reduce the mode collapse and lead to stable training. However, we find that constraining the embeddings leads to class confusion, as seen in t-SNE visualization in Fig. 3. We find that this class confusion gets further aggravated when StyleGAN is trained on datasets like ImageNet-LT, which contain a large number of classes, along with a bunch of semantically similar classes (Sec. 5.2). Based on our \\(\\mathcal{W}\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5989" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.089, + 0.498, + 0.373 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.529, + 0.098, + 0.709, + 0.218 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.715, + 0.099, + 0.89, + 0.217 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.528, + 0.222, + 0.709, + 0.342 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.715, + 0.223, + 0.89, + 0.342 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.551, + 0.345, + 0.86, + 0.376 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.389, + 0.895, + 0.434 + ], + "angle": 0, + "content": "Figure 3. Comparison of GANs and their \\(\\mathcal{W}\\) space for CIFAR10-LT. We plot the generated images on (left) and generate a t-SNE plot of w latents for generated images in \\(\\mathcal{W}\\) space (right). We find that mode collapse and class confusion in images is linked to the corresponding collapse and confusion in latent \\(\\mathcal{W}\\) space. Our proposed NoisyTwins mitigates both collapse (left) and confusion (right) simultaneously." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.44, + 0.475, + 0.531 + ], + "angle": 0, + "content": "space analysis and qualitative results above, we observe that the class confusion and mode collapse of images is tightly coupled with the structure of \\(\\mathcal{W}\\) space. Further, the recent SotA methods are either unable to prevent collapse or suffer from class confusion. Hence, this work aims to develop a technique that mitigates both confusion and collapse." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.543, + 0.188, + 0.56 + ], + "angle": 0, + "content": "4. Approach" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.568, + 0.471, + 0.659 + ], + "angle": 0, + "content": "In this section, we present our method NoisyTwins, which introduces noise-based augmentation twins in the conditional embedding space (Sec. 4.1), and then combines it with the Barlow-Twins-based regularizer from the self-supervised learning (SSL) paradigm to resolve the issue of class confusion and mode collapse (Sec. 4.2)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.666, + 0.433, + 0.684 + ], + "angle": 0, + "content": "4.1. Noise Augmentation in Embedding Space" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.69, + 0.473, + 0.902 + ], + "angle": 0, + "content": "As we observed in the previous section that \\(\\mathbf{w}\\) vectors for each sample become insensitive to changes in \\(\\mathbf{z}\\). This collapse in \\(\\mathbf{w}\\) vectors for each class leads to mode collapse for baselines (Fig. 3). One reason for this could be the fact that \\(\\mathbf{z}\\) is composed of continuous variables, whereas the embedding vectors \\(\\mathbf{c}\\) for each class are discrete. Due to this, the GAN converges to the easy degenerate solution where it generates a single sample for each class, becoming insensitive to changes in \\(\\mathbf{z}\\). For inducing some continuity in \\(\\mathbf{c}\\) embeddings vectors, we introduce an augmentation strategy where we add i.i.d. noise of small magnitude in each of the variables in \\(\\mathbf{c}\\). Based on our observation (Fig. 3) and existing works [35], there is a high tendency for mode collapse in tail classes. Hence we add noise in embedding space that is" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.44, + 0.893, + 0.482 + ], + "angle": 0, + "content": "proportional to the inverse of the frequency of samples. We provide the mathematical expression of noise augmentation \\(\\tilde{\\mathbf{c}}\\) below:" + }, + { + "type": "equation", + "bbox": [ + 0.548, + 0.481, + 0.892, + 0.51 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathbf {c}} \\sim \\mathbf {c} + \\mathcal {N} (\\mathbf {0}, \\sigma_ {c} \\mathbb {I} _ {d}) \\text {w h e r e} \\sigma_ {c} = \\sigma \\frac {(1 - \\alpha)}{1 - \\alpha^ {n _ {c}}} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.51, + 0.893, + 0.659 + ], + "angle": 0, + "content": "Here \\( n_c \\) is the frequency of training samples in class \\( c \\), \\( \\mathbb{I} \\) is the identity matrix of size \\( d \\times d \\), and \\( \\alpha, \\sigma \\) are hyperparameters. The expression of \\( \\sigma_c \\) is from the effective number of samples [8], which is a softer version of inverse frequency proportionality. In contrast to the image space augmentation, these noise augmentations come for free as there is no significant additional computation overhead. This noise is added in the embedding \\( \\mathbf{c} \\) before passing it to the generator and the discriminator, which ensures that the class embeddings occupy a continuous region in latent space." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.892, + 0.72 + ], + "angle": 0, + "content": "Insight: The augmentation equation above (Eq. 3) can be interpreted as approximating the discrete random variable \\(\\mathbf{c}\\) with a Gaussian with finite variance and the embedding parameters \\(\\mathbf{c}\\) being the mean \\(\\mu_{\\mathbf{c}}\\)." + }, + { + "type": "equation", + "bbox": [ + 0.637, + 0.72, + 0.891, + 0.735 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathbf {c}} \\sim \\mathcal {N} \\left(\\mu_ {\\mathbf {c}}, \\sigma_ {c} \\mathbb {I} _ {d}\\right) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.894, + 0.84 + ], + "angle": 0, + "content": "This leads to the class-embedding input \\(\\tilde{\\mathbf{c}}\\) to mapping network to have a Gaussian distribution, similar in nature to \\(\\mathbf{z}\\). This noise augmentation strategy alone mitigates the degenerate solution of class-wise mode collapse to a great extent (Table 1) and helps generate diverse latent \\(\\mathbf{w}\\) for each class. Due to the diverse \\(\\mathbf{w}\\) conditioning of the GAN, it leads to diverse image generation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.848, + 0.843, + 0.865 + ], + "angle": 0, + "content": "4.2. Invariance in \\(\\mathcal{W}\\)-Space with NoisyTwins" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.893, + 0.902 + ], + "angle": 0, + "content": "The augmentation strategy introduced in the previous section expands the region for each class in \\(\\mathcal{W}\\) latent space." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5990" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.089, + 0.895, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.267, + 0.895, + 0.325 + ], + "angle": 0, + "content": "Figure 4. Overview of NoisyTwins. For the \\(i^{th}\\) sample of class \\(c^i\\), we create twin augmentations \\((\\tilde{\\mathbf{c}}_a^i,\\tilde{\\mathbf{c}}_b^i)\\), by sampling from a Gaussian centered at class embedding \\((\\mu_{\\mathbf{c}^i})\\). After this, we concatenate them with the same \\(\\mathbf{z}^i\\) and obtain \\((\\tilde{\\mathbf{w}}_a^i,\\tilde{\\mathbf{w}}_b^i)\\) from the mapping network, which we stack in batches of augmented latents (\\(\\tilde{\\mathbf{W}}_{\\mathbf{A}}\\) and \\(\\tilde{\\mathbf{W}}_{\\mathbf{B}}\\)). The twin \\((\\tilde{\\mathbf{w}}_a^i,\\tilde{\\mathbf{w}}_b^i)\\) vectors are then made invariant to augmentations (similar) in the latent space by minimizing cross-correlation [2,53] between the latents of two augmented batches (\\(\\tilde{\\mathbf{W}}_{\\mathbf{A}}\\) and \\(\\tilde{\\mathbf{W}}_{\\mathbf{B}}\\))." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.332, + 0.471, + 0.407 + ], + "angle": 0, + "content": "Although that does lead to diverse image generation as \\(\\mathbf{w}\\) are diverse; however, this does not ensure that these \\(\\mathbf{w}\\) will generate class-consistent outputs for augmentations in embedding \\((\\tilde{\\mathbf{c}})\\). To ensure class consistent predictions, we need to ensure invariance in \\(\\mathbf{w}\\) to noise augmentation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.41, + 0.474, + 0.593 + ], + "angle": 0, + "content": "For enforcing invariance to augmentations, a set of recent works [2, 10, 53] in self-supervised learning make the representations of augmentations similar through regularization. Among them, we focus on Barlow Twins as it does not require a large batch size of samples. Inspired by Barlow twins, we introduce NoisyTwins (Fig. 4), where we generate twin augmentations \\(\\tilde{\\mathbf{c}}_a\\) and \\(\\tilde{\\mathbf{c}}_b\\) of the same class embedding \\((\\mu_c)\\) and concatenate them to same \\(\\mathbf{z}\\). After creating a batch of such inputs, they are passed to the mapping network to get batches of augmented latents \\((\\tilde{\\mathbf{W}}_A\\) and \\(\\tilde{\\mathbf{W}}_B)\\). These batches are then used to calculate the cross-correlation matrix of latent variables given as:" + }, + { + "type": "equation", + "bbox": [ + 0.148, + 0.599, + 0.47, + 0.662 + ], + "angle": 0, + "content": "\\[\n\\mathbf {C} _ {j, k} = \\frac {\\sum_ {(\\tilde {\\mathbf {w}} _ {a} , \\tilde {\\mathbf {w}} _ {b}) \\in (\\tilde {\\mathbf {W}} _ {A} , \\tilde {\\mathbf {W}} _ {B})} \\tilde {\\mathbf {w}} _ {a} ^ {j} \\tilde {\\mathbf {w}} _ {b} ^ {k}}{\\sum_ {\\tilde {\\mathbf {w}} _ {a} \\in \\tilde {\\mathbf {W}} _ {A}} \\tilde {\\mathbf {w}} _ {a} ^ {j} \\tilde {\\mathbf {w}} _ {a} ^ {j} \\sum_ {\\tilde {\\mathbf {w}} _ {b} \\in \\tilde {\\mathbf {W}} _ {B}} \\tilde {\\mathbf {w}} _ {b} ^ {k} \\tilde {\\mathbf {w}} _ {b} ^ {k}} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.669, + 0.471, + 0.7 + ], + "angle": 0, + "content": "The matrix \\(\\mathbf{C}\\) is a square matrix of size same as of latents \\(\\mathbf{w}\\). The final loss based on confusion matrix is given as:" + }, + { + "type": "equation", + "bbox": [ + 0.152, + 0.708, + 0.47, + 0.742 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathcal {N T}} = \\sum_ {j} \\left(1 - \\mathbf {C} _ {j j} ^ {2}\\right) + \\gamma \\sum_ {j \\neq k} \\mathbf {C} _ {j, k} ^ {2} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.473, + 0.902 + ], + "angle": 0, + "content": "The first term tries to make the two latents \\((\\tilde{\\mathbf{w}}_a\\) and \\(\\tilde{\\mathbf{w}}_b)\\) invariant to the noise augmentation applied (i.e. similar), whereas the second term tries to de-correlate the different variables, thus maximizing the information content of the embedding vector [53]. The \\(\\gamma\\) is the hyper-parameter that determines the relative importance of the two terms. This loss is then added to the generator loss term \\((\\mathcal{L}_{\\mathcal{G}} + \\lambda L_{\\mathcal{NT}})\\) and optimized through backpropagation. The above procedure comprises our proposed method, NoisyTwins (Fig. 4), which we empirically evaluate in the subsequent sections." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.33, + 0.738, + 0.348 + ], + "angle": 0, + "content": "5. Experimental Evaluation" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.358, + 0.583, + 0.375 + ], + "angle": 0, + "content": "5.1. Setup" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.389, + 0.895, + 0.646 + ], + "angle": 0, + "content": "Datasets: We primarily apply all methods on long-tailed datasets, as GANs trained on them are more prone to class confusion and mode collapse. We first report on the commonly used CIFAR10-LT dataset with an imbalance factor (i.e. ratio of most to least frequent class) of 100. To show our approach's scalability and real-world application, we test our method on the challenging ImageNet-LT and iNaturalist 2019 datasets. The ImageNet-LT [30] is a long-tailed variant of the 1000 class ImageNet dataset, with a plethora of semantically similar classes (e.g., Dogs, Birds etc.), making it challenging to avoid class confusion. iNaturalist-2019 [48] is a real-world long-tailed dataset composed of 1010 different variants of species, some of which have fine-grained differences in their appearance. For such fine-grained datasets, ImageNet pre-trained discriminators [38, 39] may not be useful, as augmentations used to train the model makes it invariant to fine-grained changes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.652, + 0.895, + 0.835 + ], + "angle": 0, + "content": "Training Configuration: We use StyleGAN2 architecture for all our experiments. All our experiments are performed using PyTorch-StudioGAN implemented by Kang et al. [18], which serves as a base for our framework. We use Path Length regularization (PLR) as it ensures changes in \\(\\mathcal{W}\\) space lead to changes in images, using a delayed PLR on ImageNet-LT following [39]. We use a batch size of 128 for all our experiments, with one G step per D step. Unless stated explicitly, we used the general training setup for StyleGANs from [18], including methods like \\(R_{1}\\) regularization [31], etc. More details on the exact training configuration for each dataset are provided in Appendix." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.87 + ], + "angle": 0, + "content": "Metrics: In this work we use the following metrics for evaluation of our methods:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.895, + 0.901 + ], + "angle": 0, + "content": "a) FID: Fréchet Inception Distance [12] is the Wasserstein-2 Distance between the real and validation data. We use a" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.958 + ], + "angle": 0, + "content": "5991" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.468, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.298, + 0.47, + 0.368 + ], + "angle": 0, + "content": "Figure 5. Choice of Eval. backbone: intra-FID (iFID) of a class based on InceptionV3 backbone (left plot) is not able to capture the mode collapse (increase in iFID in the absence of mode collapse). This is well-captured by \\(\\mathrm{iFID}_{\\mathrm{CLIP}}\\) based on CLIP [34] backbone (right plot, decrease in iFID in the absence of mode collapse)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.384, + 0.468, + 0.489 + ], + "angle": 0, + "content": "held-out validation set and \\(50\\mathrm{k}\\) generated samples to evaluate FID in each case. As FID is biased towards ImageNet and can be arbitrarily manipulated, we also report \\(\\mathrm{FID}_{\\mathrm{CLIP}}\\). b) Precision & Recall: As we aim to mitigate the mode collapse and achieve diverse generations across classes, we use improved Precision & Recall [25] metrics, as poor recall indicates mode collapse [39]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.49, + 0.468, + 0.822 + ], + "angle": 0, + "content": "d) Intra-Class \\(\\mathbf{FID}_{\\mathrm{CLIP}}\\) (\\(\\mathbf{iFID}_{\\mathrm{CLIP}}\\)): The usage of only FID based on Inception-V3 Networks for evaluation of Generative Models has severe limitations, as it has been found that FID can be reduced easily by some fringe features [26]. iFID is computed by taking FID between 5k generated and real samples for the same class. As we want to evaluate both class consistency and diversity, we find that similar limitations exist for intra-class FID (iFID), which has been used to evaluate class-conditional GANs [18]. In Fig. 5, we show the existence of generated images for a particular class (more in Appendix) from models trained on iNaturalist 2019, where iFID is better for the mode collapsed model than the other model generating diverse images. Whereas the \\(\\mathbf{iFID}_{\\mathrm{CLIP}}\\), based on CLIP backbone can rank the models correctly with the model having mode collapse having high \\(\\mathbf{iFID}_{\\mathrm{CLIP}}\\). Further, we find that the mean iFID can be deceptive in detecting class confusion and collapse cases, as it sometimes ranks models with high realism better than models generating diversity (See Appendix). Hence, mean \\(\\mathbf{iFID}_{\\mathrm{CLIP}}\\) (ref. to as \\(\\mathbf{iFID}_{\\mathrm{CLIP}}\\) in result section for brevity) can be reliably used to evaluate models for class consistency and diversity." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.468, + 0.9 + ], + "angle": 0, + "content": "Baselines: For evaluating NoisyTwins performance in comparison to other methods, we use the implementations present in StudioGAN [16]. For fairness, we re-run all the baselines on StyleGAN2 in the same hyperparameter setting. We compare our method to the StyleGAN2" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.091, + 0.892, + 0.394 + ], + "angle": 0, + "content": "(SG2) and StyleGAN2 with augmentation (DiffAug [54] and ADA [20]) baselines. We further tried to improve the baseline by incorporating the recent LeCam regularization method; however, it resulted in gains only for the iNaturalist 2019 dataset, where we use LeCam for all experiments. Further on StyleGAN2, we also use contrastive D2D-CE loss conditioning (i.e. ReACGAN) as a baseline. However, the D2D-CE baseline completely ignores learning of tail classes (Fig. 3) for CIFAR10-LT and is expensive to train; hence we do not report results for it for large-scale long-tailed datasets. We also compare our method against the recent SotA group Spectral Normalization (gSR) [35] method, which we implement for StyleGAN2 by constraining the spectral norms of embedding parameters of the generator \\((\\mathcal{G})\\) as suggested by authors. As a sanity check, we reproduce their results on CIFAR10-LT and find that our implementation matches the reported results correctly. We provide results on all datasets, for both the proposed Noise Augmentation (+ Noise) and the overall proposed NoisyTwins (+NoisyTwins) method." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.407, + 0.86, + 0.423 + ], + "angle": 0, + "content": "5.2. Results on Long-Tailed Data Distributions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.431, + 0.892, + 0.703 + ], + "angle": 0, + "content": "CIFAR10-LT. We applied DiffAug [54] on all baselines, except on gSR, where we found that DiffAug provides inferior results compared to ADA (as also used by authors [35]). It can be observed in Table 2 that the addition of NoisyTwins regularization significantly improves over baseline by \\((\\sim 14\\) FID) along with providing superior class consistency as shown by improved iFIDCLIP. NoisyTwins is also able to outperform the recent gSR regularization method and achieves improved results for all metrics. Further, NoisyTwins improves FID for StyleGAN2-ADA baseline used by gSR too from 32.08 to 23.02, however the final results are inferior than reported DiffAug baseline results. Further, we observed that despite not producing any tail class images (Fig. 3), the D2D-CE baseline has much superior FID in comparison to baselines. Whereas the proposed iFIDCLIP value is similar for the baseline and D2D-CE model. This clearly demonstrates the superiority of proposed iFIDCLIP in detecting class confusion." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Real-world, Large-scale Long-Tailed Datasets. We experiment with iNaturalist 2019 and ImageNet-LT. These datasets are particularly challenging as they contain long-tailed imbalances and semantically similar classes, making GANs prone to mode collapse and class confusion. The baselines StyleGAN2 and StyleGAN2-ADA both suffer from mode collapse (Fig. 6), particularly for the tail classes. Whereas for the recent SotA gSR method, we find that although it undergoes less collapse in comparison to baselines, it suffers from class confusion as seen from similar Intra-FIDCLIP in comparison to baselines (Table 1). Compared to that, our method NoisyTwins improves when used with StyleGAN2-ADA significantly, leading to" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5992" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.895, + 0.133 + ], + "angle": 0, + "content": "Table 1. Quantitative results on ImageNet-LT and iNaturalist 2019 Datasets. We compare FID(↓), FIDCLIP(↓), iFIDCLIP(↓), Precision(↑) and Recall(↑) with other existing approaches on StyleGAN2 (SG2). We obtain an average ~ 19% relative improvement on FID, ~ 33% on FIDCLIP, and ~ 11% on iFIDCLIP metrics over the previous SotA on ImageNet-LT and iNaturalist 2019 datasets." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.134, + 0.891, + 0.233 + ], + "angle": 0, + "content": "
MethodImageNet-LTiNaturalist 2019
FID(↓)FIDCLIP(↓)iFIDCLIP(↓)Precision(↑)Recall(↑)FID(↓)FIDCLIP(↓)iFIDCLIP(↓)Precision(↑)Recall(↑)
SG2 [23]41.2511.6446.930.500.4819.343.3338.240.740.17
SG2+ADA [20]37.2011.0447.410.540.3814.922.3035.190.750.57
SG2+ADA+gSR [35]24.788.2144.420.630.3515.172.0636.220.740.46
SG2+ADA+Noise (Ours)22.177.1141.200.720.3312.871.3731.430.810.63
+ NoisyTwins (Ours)21.296.4139.740.670.4911.461.1431.500.790.67
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.239, + 0.571, + 0.281 + ], + "angle": 0, + "content": "Table 2. Quantitative results on CIFAR10-LT Dataset. We compare with other existing approaches. We obtain \\(\\sim 26\\%\\) relative improvement over the existing methods on \\(\\mathrm{FID}_{\\mathrm{CLIP}}\\) and iFIDCLIP metrics." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.293, + 0.605, + 0.384 + ], + "angle": 0, + "content": "
MethodFID(↓)FIDCLIP(↓)iFIDCLIP(↓)Precision(↑)Recall(↑)
SG2+DiffAug [54]31.736.2711.590.630.35
SG2+D2D-CE [17]19.974.7711.350.730.42
gSR [35]22.105.549.940.700.29
SG2+DiffAug+Noise (Ours)28.905.2610.650.710.38
+ NoisyTwins(Ours)17.743.557.240.700.51
" + }, + { + "type": "table_caption", + "bbox": [ + 0.598, + 0.239, + 0.889, + 0.293 + ], + "angle": 0, + "content": "Table 3. Comparison with SotA approaches on BigGAN. We compare FID(↓) with other existing models on ImageNet-LT (IN-LT) and iNaturalist 2019 (iNat-19)." + }, + { + "type": "table", + "bbox": [ + 0.624, + 0.293, + 0.868, + 0.385 + ], + "angle": 0, + "content": "
MethodiNat-19IN-LT
BigGAN [4]14.8528.10
+ gSR [35]13.95-
ICGAN [6]-23.40
StyleGAN2-ADA [20]14.9237.20
+ NoisyTwins (Ours)11.4621.29
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.395, + 0.471, + 0.531 + ], + "angle": 0, + "content": "a relative improvement of \\(42.7\\%\\) in FID for ImageNet-LT and \\(23.19\\%\\) on the iNaturalist 2019 dataset when added to StyleGAN2-ADA baseline. Further with Noise Augmentation \\((+\\mathrm{Noise})\\), we observe generations of high-quality class-consistent images, but it also suffers from mode collapse. This can be observed by the high-precision values in comparison to low-recall values. However, adding NoisyTwins regularization over the noise augmentation improves diversity by improving recall (Table 1)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.532, + 0.47, + 0.835 + ], + "angle": 0, + "content": "Fig. 6 presents the generated images of tail classes for various methods on ImageNet-LT, where NoisyTwins generations show remarkable diversity in comparison to others. The presence of diversity for classes with just 5-6 training images demonstrates successful transfer of knowledge from head classes to tail classes, due to shared parameters. Further, to compare with existing SotA reported results, we compare FID of BigGAN models from gSR [35] and Instance Conditioned GAN (ICGAN) [6]. For fairness, we compare FID on the validation set for which we obtained gSR models from authors and re-evaluate them, as they reported FID on a balanced training set. As BigGAN models are more common for class-conditioned generation [18], their baseline performs superior to StyleGAN2-ADA baselines (Table 3). However, the addition of NoisyTwins to the StyleGAN2-ADA method improves it significantly, even outperforming the existing methods of gSR (by \\(18.44\\%\\) ) and ICGAN (by \\(9.44\\%\\) ) based on BigGAN architecture. This shows that NoisyTwins allows the StyleGAN2 baseline to scale to large and diverse long-tailed datasets." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.847, + 0.373, + 0.862 + ], + "angle": 0, + "content": "5.3. NoisyTwins on Few-Shot Datasets" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.471, + 0.903 + ], + "angle": 0, + "content": "We now demonstrate the potential of NoisyTwins in another challenging scenario of class-conditional few-shot" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.395, + 0.895, + 0.624 + ], + "angle": 0, + "content": "image generation from GANs. We perform our experiments using a conditional StyleGAN2-ADA baseline, for which we tune hyper-parameters to obtain a strong baseline. We then apply our method of Noise Augmentation and NoisyTwins over the strong baseline for reporting our results. We use the few-shot dataset of LHI-AnimalFaces [45] and a subset of ImageNet Carnivores [29, 40] to report our results. Table 4 shows the results of these experiments, where we find that our method, NoisyTwins, significantly improves the FID of StyleGAN2 ADA baseline by \\((22.2\\%)\\) on average for both datasets. Further, combining Noisy Twins with SotA Transitional-cGAN [40] through official code, also leads to effective improvement in FID. These results clearly demonstrate the diverse potential and applicability of our proposed method NoisyTwins." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.642, + 0.597, + 0.659 + ], + "angle": 0, + "content": "6. Analysis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.669, + 0.894, + 0.806 + ], + "angle": 0, + "content": "We perform analysis of NoisyTwins w.r.t. to its hyperparameters, standard deviation \\((\\sigma)\\) of noise augmentation and regularization strength \\((\\lambda)\\). We also compare NoisyTwins objective (ref. Eq. 6) with contrastive objective. Finally, we compare NoisyTwins over Latent Diffusion Models for long-tailed class conditional generation task. We perform ablation experiments on CIFAR10-LT, for which additional details and results are present in Appendix. We also present comparison of NoisyTwins for GAN fine-tuning." + }, + { + "type": "title", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.825 + ], + "angle": 0, + "content": "How much noise and regularization strength is optimal?" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.894, + 0.902 + ], + "angle": 0, + "content": "In Fig. 7, we ablate over the noise variance parameter \\(\\sigma\\) for CIFAR10-LT. We find that a moderate value of noise strength 0.75 leads to optimal results. For the strength of NoisyTwins loss \\((\\lambda)\\), we find that the algorithm performs similarly on values near 0.01 and is robust to it (Fig. 7)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.958 + ], + "angle": 0, + "content": "5993" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.09, + 0.357, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.206, + 0.251, + 0.35, + 0.272 + ], + "angle": 0, + "content": "(a) Image Augmentation based Methods (ADA)" + }, + { + "type": "image", + "bbox": [ + 0.367, + 0.091, + 0.524, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.387, + 0.252, + 0.504, + 0.273 + ], + "angle": 0, + "content": "(b) Regularizer based Methods (gSR)" + }, + { + "type": "image", + "bbox": [ + 0.535, + 0.091, + 0.888, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.657, + 0.252, + 0.771, + 0.263 + ], + "angle": 0, + "content": "(c) NoisyTwins (Ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.288, + 0.895, + 0.331 + ], + "angle": 0, + "content": "Figure 6. Qualitative results on ImageNet-LT for tail classes. We find that existing SotA methods for tail classes show collapsed (a) or arbitrary image generation (b). With NoisyTwins, we observe diverse and class-consistent image generation, even for classes having 5-6 images. The tail classes get enhanced diversity by transferring the knowledge from head classes, as they share parameters." + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.337, + 0.471, + 0.378 + ], + "angle": 0, + "content": "Table 4. Quantitative results on ImageNet Carnivore and AnimalFace Datasets. Our method improves over both StyleGAN2-ADA (SG2-ADA) baseline and SotA Transitional-cGAN." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.39, + 0.485, + 0.523 + ], + "angle": 0, + "content": "
ImageNet CarnivoreAnimalFace
MethodFID(↓)iFIDCLIP(↓)FID(↓)iFIDCLIP(↓)
SG2 [23]111.8336.3494.0929.94
SG2+ADA [20]22.7712.8520.2511.12
SG2+ADA+Noise (Ours)19.2512.5118.7810.42
+ NoisyTwins (Ours)16.0112.4117.2710.03
FID(↓)FID(↓)
Transitional-cGAN [40]14.6020.53
+ NoisyTwins (Ours)13.6516.15
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.554, + 0.47, + 0.72 + ], + "angle": 0, + "content": "Which type of self-supervision to use with noise augmentation? The goal of our method is to achieve invariance to Noise Augmentation in the \\(\\mathcal{W}\\) latent space. This can be achieved using either contrastive learning-based methods like SimCLR [7] or negative-free method like Barlow Twins [53]. Contrastive loss (SimCLR based) produces FID of 26.23 vs 17.74 by NoisyTwins (BarlowTwins based). We find that contrastive baseline improves over the noise augmentation baseline (28.90) however falls significantly below the NoisyTwins, as the former requires a large batch size to be effective which is expensive for GANs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.903 + ], + "angle": 0, + "content": "How does NoisyTwins compare with modern Vision and Language models? For evaluating the effectiveness of modern vision language-based diffusion models, we test the generation of the iNaturalist 2019 dataset by creating the prompt \"a photo of S\" where we replace the class name in place of S. We use the LDM [37] model trained on LAION-400M to perform inference, generating 50 images per class. We obtained an FID of 57.04 in comparison to best FID of 11.46 achieved by NoisyTwins. This clearly demonstrates that for specific use cases like fine-grained generation, GANs are still ahead of general-purpose LDM." + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.337, + 0.892, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.5, + 0.892, + 0.542 + ], + "angle": 0, + "content": "Figure 7. Ablation of Hyperparameters. Quantitative comparison on CIFAR10-LT for standard deviation of Noise Augmentation \\((\\sigma)\\) and strength \\((\\lambda)\\) of NoisyTwins loss." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.547, + 0.62, + 0.563 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.569, + 0.892, + 0.903 + ], + "angle": 0, + "content": "In this work, we analyze the performance of StyleGAN2 models on the real-world long-tailed datasets including iNaturalist 2019 and ImageNet-LT. We find that existing works lead to either class confusion or mode collapse in the image space. This phenomenon is rooted in collapse and confusion in the latent \\(\\mathcal{W}\\) space of StyleGAN2. Through our analysis, we deduce that this collapse occurs when the latents become invariant to random conditioning vectors \\(\\mathbf{z}\\), and collapse for each class. To mitigate this, we introduce inexpensive noise based augmentation for discrete class embeddings. Further, to ensure class consistency, we couple this augmentation technique with BarlowTwins' objective in the latent \\(\\mathcal{W}\\) space which imparts intra-class diversity to latent \\(\\mathbf{w}\\) vectors. The noise augmentation and regularization comprises our proposed NoisyTwins technique, which improves the performance of StyleGAN2 establishing a new SotA on iNaturalist 2019 and ImageNet-LT. The extension of NoisyTwins for conditioning on more-complex attributes for StyleGANs is a good direction for future work. Acknowledgements: This work was supported in part by SERB-STAR Project (STR/2020/000128). Harsh Rangwani is supported by PMRF fellowship." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5994" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.473, + 0.186 + ], + "angle": 0, + "content": "[1] Yuval Alaluf, Omer Tov, Ron Mokady, Rinon Gal, and Amit Bermano. Hyperstyle: Stylegan inversion with hypernetworks for real image editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18511-18521, June 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.187, + 0.472, + 0.243 + ], + "angle": 0, + "content": "[2] Adrien Bardes, Jean Ponce, and Yann LeCun. VI-CReg: Variance-invariance-covariance regularization for self-supervised learning. In International Conference on Learning Representations, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.245, + 0.472, + 0.299 + ], + "angle": 0, + "content": "[3] Stella Bounareli, Vasileios Argyriou, and Georgios Tzimiropoulos. Finding directions in gan's latent space for neural face reenactment. arXiv preprint arXiv:2202.00046, 2022.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.302, + 0.472, + 0.357 + ], + "angle": 0, + "content": "[4] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale GAN training for high fidelity natural image synthesis. In International Conference on Learning Representations, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.359, + 0.472, + 0.413 + ], + "angle": 0, + "content": "[5] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.416, + 0.472, + 0.472 + ], + "angle": 0, + "content": "[6] Arantxa Casanova, Marlene Careil, Jakob Verbeek, Michal Drozdal, and Adriana Romero-Soriano. Instance-conditioned gan. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.474, + 0.472, + 0.528 + ], + "angle": 0, + "content": "[7] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. arXiv preprint arXiv:2002.05709, 2020.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.531, + 0.472, + 0.599 + ], + "angle": 0, + "content": "[8] Yin Cui, Menglin Jia, Tsung-Yi Lin, Yang Song, and Serge Belongie. Class-balanced loss based on effective number of samples. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9268-9277, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.602, + 0.472, + 0.657 + ], + "angle": 0, + "content": "[9] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139-144, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.66, + 0.472, + 0.743 + ], + "angle": 0, + "content": "[10] Jean-Bastien Grill, Florian Strub, Florent Altché, Coretin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.745, + 0.472, + 0.799 + ], + "angle": 0, + "content": "[11] Erik Härkönen, Aaron Hertzmann, Jaakko Lehtinen, and Sylvain Paris. Ganspace: Discovering interpretable gan controls. Advances in Neural Information Processing Systems, 33:9841-9850, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.802, + 0.472, + 0.87 + ], + "angle": 0, + "content": "[12] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.873, + 0.472, + 0.901 + ], + "angle": 0, + "content": "[13] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceed-" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.116, + 0.473, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "ings of the IEEE international conference on computer vision, pages 1501-1510, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.121, + 0.892, + 0.175 + ], + "angle": 0, + "content": "[14] Jongheon Jeong and Jinwoo Shin. Training GANs with stronger augmentations via contrastive discriminator. In International Conference on Learning Representations, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.177, + 0.892, + 0.218 + ], + "angle": 0, + "content": "[15] Liming Jiang, Bo Dai, Wayne Wu, and Chen Change Loy. Deceive D: Adaptive Pseudo Augmentation for GAN training with limited data. In NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.219, + 0.892, + 0.246 + ], + "angle": 0, + "content": "[16] Minguk Kang and Jaesik Park. ContraGAN: Contrastive Learning for Conditional Image Generation. 2020. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.247, + 0.892, + 0.3 + ], + "angle": 0, + "content": "[17] Minguk Kang, Woohyeon Shim, Minsu Cho, and Jaesik Park. Rebooting acgan: Auxiliary classifier gans with stable training. Advances in neural information processing systems, 34:23505-23518, 2021. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.302, + 0.892, + 0.343 + ], + "angle": 0, + "content": "[18] MinGuk Kang, Joonghyuk Shin, and Jaesik Park. StudioGAN: A Taxonomy and Benchmark of GANs for Image Synthesis. 2206.09479 (arXiv), 2022. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.344, + 0.892, + 0.385 + ], + "angle": 0, + "content": "[19] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. arXiv preprint arXiv:1710.10196, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.386, + 0.892, + 0.44 + ], + "angle": 0, + "content": "[20] Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. In Proc. NeurIPS, 2020. 1, 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.442, + 0.892, + 0.509 + ], + "angle": 0, + "content": "[21] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.511, + 0.892, + 0.58 + ], + "angle": 0, + "content": "[22] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.581, + 0.892, + 0.634 + ], + "angle": 0, + "content": "[23] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. In Proc. CVPR, 2020. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.636, + 0.892, + 0.677 + ], + "angle": 0, + "content": "[24] Tuomas Kynkänniemi, Tero Karras, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Improved Precision and Recall Metric for Assessing Generative Models. 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.679, + 0.892, + 0.734 + ], + "angle": 0, + "content": "[25] Tuomas Kynkänniemi, Tero Karras, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Improved precision and recall metric for assessing generative models. Advances in Neural Information Processing Systems, 32, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.735, + 0.892, + 0.775 + ], + "angle": 0, + "content": "[26] Tuomas Kynkänniemi, Tero Karras, Miika Aittala, Timo Aila, and Jaakko Lehtinen. The role of imagenet classes in fréchet inception distance. CoRR, abs/2203.06026, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.776, + 0.892, + 0.816 + ], + "angle": 0, + "content": "[27] Thomas Leimkuhler and George Drettakis. Freestylegan: Free-view editable portrait rendering with the camera manifold. 40(6), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.818, + 0.892, + 0.859 + ], + "angle": 0, + "content": "[28] Feng Liu and Xiaoming Liu. 2d gans meet unsupervised single-view 3d reconstruction. In Computer Vision - ECCV 2022, pages 497-514, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.86, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[29] Ming-Yu Liu, Xun Huang, Arun Mallya, Tero Karras, Timo Aila, Jaakko Lehtinen, and Jan Kautz. Few-shot unsupervised image-to-image translation. In Proceedings of" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5995" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.092, + 0.468, + 0.12 + ], + "angle": 0, + "content": "the IEEE/CVF international conference on computer vision, pages 10551-10560, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.469, + 0.191 + ], + "angle": 0, + "content": "[30] Ziwei Liu, Zhongqi Miao, Xiaohang Zhan, Jiayun Wang, Boqing Gong, and Stella X Yu. Large-scale long-tailed recognition in an open world. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2537-2546, 2019. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.469, + 0.247 + ], + "angle": 0, + "content": "[31] Lars Mescheder, Andreas Geiger, and Sebastian Nowozin. Which training methods for gans do actually converge? In International conference on machine learning, pages 3481-3490. PMLR, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.249, + 0.469, + 0.331 + ], + "angle": 0, + "content": "[32] Rishubh Parihar, Ankit Dhiman, Tejan Karmali, and R. Venkatesh Babu. Everything is there in latent space: Attribute editing and attribute style manipulation by stylegan latent space exploration. In Proceedings of the 30th ACM International Conference on Multimedia, pages 1828-1836, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.333, + 0.469, + 0.403 + ], + "angle": 0, + "content": "[33] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2085–2094, October 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.404, + 0.469, + 0.488 + ], + "angle": 0, + "content": "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.489, + 0.469, + 0.545 + ], + "angle": 0, + "content": "[35] Harsh Rangwani, Naman Jaswani, Tejan Karmali, Varun Jampani, and R. Venkatesh Babu. Improving gans for long-tailed data through group spectral regularization. In European Conference on Computer Vision, 2022. 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.546, + 0.469, + 0.6 + ], + "angle": 0, + "content": "[36] Harsh Rangwani, Konda Reddy Mopuri, and R Venkatesh Babu. Class balancing gan with a classifier in the loop. In Uncertainty in Artificial Intelligence, pages 1618-1627. PMLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.602, + 0.469, + 0.643 + ], + "angle": 0, + "content": "[37] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.645, + 0.469, + 0.686 + ], + "angle": 0, + "content": "[38] Axel Sauer, Kashyap Chitta, Jens Müller, and Andreas Geiger. Projected gans converge faster. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.688, + 0.469, + 0.729 + ], + "angle": 0, + "content": "[39] Axel Sauer, Katja Schwarz, and Andreas Geiger. Styleganx1: Scaling stylegan to large diverse datasets. volume abs/2201.00273, 2022. 1, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.731, + 0.469, + 0.786 + ], + "angle": 0, + "content": "[40] Mohamad Shahbazi, Martin Danelljan, Danda Pani Paudel, and Luc Van Gool. Collapse by conditioning: Training class-conditional GANs with limited data. In International Conference on Learning Representations, 2022. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.788, + 0.469, + 0.828 + ], + "angle": 0, + "content": "[41] Yujun Shen, Jinjin Gu, Xiaou Tang, and Bolei Zhou. Interpreting the latent space of gans for semantic face editing. In CVPR, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.83, + 0.469, + 0.872 + ], + "angle": 0, + "content": "[42] Yujun Shen, Ceyuan Yang, Xiaou Tang, and Bolei Zhou. Interfacegan: Interpreting the disentangled face representation learned by gans. IEEE TPAMI, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[43] Yujun Shen and Bolei Zhou. Closed-form factorization of latent semantics in gans. In CVPR, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.148 + ], + "angle": 0, + "content": "[44] Yichun Shi, Divyansh Aggarwal, and Anil K. Jain. Lifting 2d stylegan for 3d-aware face generation. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6254-6262, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[45] Zhangzhang Si and Song-Chun Zhu. Learning hybrid image templates (hit) by information projection. IEEE Transactions on pattern analysis and machine intelligence, 34(7):1354-1367, 2011. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.206, + 0.892, + 0.248 + ], + "angle": 0, + "content": "[46] Ivan Skorokhodov, Sergey Tulyakov, and Mohamed Elhoseiny. Stylegan-v: A continuous video generator with the price, image quality and perks of stylegan2, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.289 + ], + "angle": 0, + "content": "[47] Hung-Yu Tseng, Lu Jiang, Ce Liu, Ming-Hsuan Yang, and Weilong Yang. Regularizing generative adversarial networks under limited data. In CVPR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.291, + 0.892, + 0.373 + ], + "angle": 0, + "content": "[48] Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alex Shepard, Hartwig Adam, Pietro Perona, and Serge Belongie. The inaturalist species classification and detection dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8769-8778, 2018. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.375, + 0.892, + 0.444 + ], + "angle": 0, + "content": "[49] Zongze Wu, Dani Lischinski, and Eli Shechtman. Stylespace analysis: Disentangled controls for stylegan image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12863–12872, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.446, + 0.892, + 0.487 + ], + "angle": 0, + "content": "[50] Ceyuan Yang, Yujun Shen, Yinghao Xu, Deli Zhao, Bo Dai, and Bolei Zhou. Improving gans with a dynamic discriminator. 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.489, + 0.892, + 0.558 + ], + "angle": 0, + "content": "[51] Ceyuan Yang, Yujun Shen, Yinghao Xu, and Bolei Zhou. Data-efficient instance generation from instance discrimination. In A. Beygelzimer, Y. Dauphin, P. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.559, + 0.892, + 0.628 + ], + "angle": 0, + "content": "[52] Sihyun Yu, Jihoon Tack, Sangwoo Mo, Hyunsu Kim, Junho Kim, Jung-Woo Ha, and Jinwoo Shin. Generating videos with dynamics-aware implicit generative adversarial networks. In International Conference on Learning Representations, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.63, + 0.892, + 0.686 + ], + "angle": 0, + "content": "[53] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow twins: Self-supervised learning via redundancy reduction. In International Conference on Machine Learning, pages 12310-12320. PMLR, 2021. 2, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.686, + 0.892, + 0.741 + ], + "angle": 0, + "content": "[54] Shengyu Zhao, Zhijian Liu, Ji Lin, Jun-Yan Zhu, and Song Han. Differentiable augmentation for data-efficient gan training. In Conference on Neural Information Processing Systems (NeurIPS), 2020. 1, 2, 3, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.741 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5996" + } + ] +] \ No newline at end of file diff --git a/2023/NoisyTwins_ Class-Consistent and Diverse Image Generation Through StyleGANs/c323d83e-f322-4c02-8ecc-1d0c86cf0f70_origin.pdf b/2023/NoisyTwins_ Class-Consistent and Diverse Image Generation Through StyleGANs/c323d83e-f322-4c02-8ecc-1d0c86cf0f70_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..05c47bd50cf7f5f3a526c2054911a34d6e0d2de5 --- /dev/null +++ b/2023/NoisyTwins_ Class-Consistent and Diverse Image Generation Through StyleGANs/c323d83e-f322-4c02-8ecc-1d0c86cf0f70_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:817f7a5a0bf1135359da19272191e22d97b6ff18767b9c067780b3df263a5138 +size 8830570 diff --git a/2023/NoisyTwins_ Class-Consistent and Diverse Image Generation Through StyleGANs/full.md b/2023/NoisyTwins_ Class-Consistent and Diverse Image Generation Through StyleGANs/full.md new file mode 100644 index 0000000000000000000000000000000000000000..60b07084d499798e9863ebab583d8da4086d61e2 --- /dev/null +++ b/2023/NoisyTwins_ Class-Consistent and Diverse Image Generation Through StyleGANs/full.md @@ -0,0 +1,278 @@ +# NoisyTwins: Class-Consistent and Diverse Image Generation through StyleGANs + +Harsh Rangwani $^{1*}$ Lavish Bansal $^{1,3*}$ Kartik Sharma $^{1,4}$ Tejan Karmali $^{2}$ Varun Jampani $^{2}$ R. Venkatesh Babu $^{1}$ + +$^{1}$ Vision and AI Lab, IISc Bangalore $^{2}$ Google Research $^{3}$ IIT BHU Varanasi $^{4}$ BITS Pilani + +# Abstract + +StyleGANs are at the forefront of controllable image generation as they produce a latent space that is semantically disentangled, making it suitable for image editing and manipulation. However, the performance of StyleGANs severely degrades when trained via class-conditioning on large-scale long-tailed datasets. We find that one reason for degradation is the collapse of latents for each class in the $\mathcal{W}$ latent space. With NoisyTwins, we first introduce an effective and inexpensive augmentation strategy for class embeddings, which then decorrelates the latents based on self-supervision in the $\mathcal{W}$ space. This decorrelation mitigates collapse, ensuring that our method preserves intraclass diversity with class-consistency in image generation. We show the effectiveness of our approach on large-scale real-world long-tailed datasets of ImageNet-LT and iNaturalist 2019, where our method outperforms other methods by $\sim 19\%$ on FID, establishing a new state-of-the-art. + +# 1. Introduction + +StyleGANs [21, 22] have shown unprecedented success in image generation, particularly on well-curated and articulated datasets (eg. FFHQ for face images, etc.). In addition to generating high fidelity and diverse images, StyleGANs also produce a disentangled latent space, which is extensively used for image editing and manipulation tasks [49]. As a result, StyleGANs are being extensively used in various applications like face-editing [11, 41], video generation [46, 52], face reenactment [3], etc., which are a testament to their usability and generality. However, despite being successful on well-curated datasets, training StyleGANs on in-the-wild and multi-category datasets is still challenging. A large-scale conditional StyleGAN (i.e. StyleGAN-XL) on ImageNet was recently trained successfully by Sauer et al. [39] using the ImageNet pre-trained model through the idea of a projection discriminator [38]. While the StyleGAN-XL uses additional pre-trained mod + +![](images/294d72e02624c3cfda3606df3ebfcf6ef336ce375e584f3d0a8017284a84da77.jpg) +Figure 1. Qualitative Comparison on tail classes (T1-T4) for iNaturalist 2019. We provide sample(s) from real class (with class frequency), generated by StyleGAN2-ADA and after adding proposed NoisyTwins. NoisyTwins achieves remarkable diversity, class-consistency and quality by just using 38 samples on average. + +els, obtaining such models for distinctive image domains like medical, forensics, and fine-grained data may not be feasible, which limits its generalization across domains. + +In this work, we aim to train vanilla class-conditional StyleGAN without any pre-trained models on challenging real-world long-tailed data distributions. As training StyleGAN with augmentations [20, 54] leads to low recall [24] (which measures diversity in the generated images) and mode collapse, particularly for minority (i.e. tail) classes. For investigating this phenomenon further, we take a closer look at the latent $\mathcal{W}$ space of StyleGAN that is produced by a fully-connected mapping network that takes the conditioning variables $\mathbf{z}$ (i.e. random noise) and class embedding $\mathbf{c}$ as inputs. The vectors $\mathbf{w}$ in $\mathcal{W}$ space are used for conditioning various layers of the generator (Fig. 2). We find that output vectors $\mathbf{w}$ from the mapping network hinge on the conditioning variable $\mathbf{c}$ and become invariant to random conditioning vector $\mathbf{z}$ . This collapse of latents leads to un + +![](images/f35637b74c73f94493ea4064e32564bcbdf97cdff0ece6410e380c05a3e53b63.jpg) +Figure 2. Schematic illustration of $\mathcal{W}$ space for different GANs. Existing conditioning methods either suffer from mode collapse [20] or lead to class confusion [35] in $\mathcal{W}$ space. With proposed NoisyTwins, we achieve intra class diversity while avoiding class confusion. + +stable training and is one of the causes of poor recall (a.k.a. mode collapse) for minority classes. Further, on augmenting StyleGAN with recent conditioning and regularization techniques [17, 35], we find that they either lead to a poor recall for minority classes or lead to class confusion (Fig. 2) instead of mitigating the collapse. + +To mitigate the collapse of $\mathbf{w}$ in $\mathcal{W}$ space, we need to ensure that the change in conditioning variable $\mathbf{z}$ leads to the corresponding change in $\mathbf{w}$ . Recently in self-supervised learning, several techniques [2, 53] have been introduced to prevent the collapse of learned representations by maximizing the information content in the feature dimensions. Inspired by them we propose NoisyTwins, in which we first generate inexpensive twin augmentations for class embeddings and then use them to decorrelate the $\mathbf{w}$ variables through self-supervision. The decorrelation ensures that $\mathbf{w}$ vectors are diverse for each class and the GAN is able to produce intra-class diversity among the generated images. + +We evaluate NoisyTwins on challenging benchmarks of large-scale long-tailed datasets of ImageNet-LT [30] and iNaturalist 2019 [48]. These benchmarks are particularly challenging due to a large number of classes present, which makes GANs prone to class confusion. On the other hand, as these datasets are long-tailed with only a few images per class in tail classes, generating diverse images for those classes is challenging. We observe that existing metrics used in GAN evaluations are not able to capture both class confusion and mode collapse. As a remedy, we propose to use intra-class Frechet Inception Distance (FID) [12] based on features obtained from pre-trained CLIP [34] embeddings as an effective metric to measure the performance of class-conditional GANs in long-tailed data setups. Using NoisyTwins enables StyleGAN to generate diverse and class-consistent images across classes, mitigating the mode collapse and class confusion issues in existing state-of-the-art (SotA) (Fig. 1). Further, with NoisyTwins, we obtain diverse generations for tail classes even with $\leq 30$ images, which can be attributed to the transfer of knowledge from head classes through shared parameters (Fig. 1 and 6). In summary, we make the following contributions: + +1. We evaluate various recent SotA GAN conditioning + +and regularization techniques on the challenging task of long-tailed image generation. We find that all existing methods either suffer from mode collapse or lead to class confusion in generations. + +2. To mitigate mode collapse and class confusion, we introduce NoisyTwins, an effective and inexpensive augmentation strategy for class embeddings that decorrelates latents in the $\mathcal{W}$ latent space (Sec. 4). +3. We evaluate NoisyTwins on large-scale long-tailed datasets of ImageNet-LT and iNaturalist-2019, where it consistently improves the StyleGAN2 performance $(\sim 19\%)$ , achieving a new SotA. Further, our approach can also prevent mode collapse and enhance the performance of few-shot GANs (Sec. 5.3). + +# 2. Related Works + +StyleGANs. Karras et al. introduced StyleGAN [21] and subsequently improved its image quality in StyleGAN2. StyleGAN could produce high-resolution photorealistic images as demonstrated on a wide variety of category-specific datasets. It introduced a mapping network, which mapped the sampled noise into another latent space, which is more disentangled and semantically coherent, as demonstrated by its downstream usage for image editing and manipulation [1,32,33,42,43]. Further, StyleGAN has been extended to get novel views from images [27, 28, 44], thus making it possible to get 3D information from it. These downstream advances are possible due to the impressive performance of StyleGANs on class-specific datasets (such as faces). However, similar photorealism levels are yet uncommon on multi-class long-tailed datasets (such as ImageNet). + +GANs for Data Efficiency and Imbalance. Failure of GANs on less data was concurrently reported by Karras et al. [20] and Zhao et al. [54]. The problem is rooted in the overfitting of the discriminator due to less real data. Since then, the proposed solutions for this problem have relied on a) augmenting the data, b) introducing regularizers, and c) architectural modifications. Karras et al. [20] and Zhao et al. [54] relied on differentiable data augmentation before passing images into the discriminator to solve this problem. DeceiveD [15] proposed to introduce label-noise for + +discriminator training. LeCamGAN [47] finds that enforcing LeCam divergence as a regularization trick in the discriminator can robustify GAN training under a limited data setting. DynamicD [50] tunes the capacity of the discriminator on-the-fly during training. While these methods can handle the data inefficiency, they are ineffective on class imbalanced long-tailed data distribution [35]. + +CBGAN [36] proposed a solution to train the unconditional GAN model on long-tailed data distribution by introducing a signal from the classifier to balance the classes generated by GAN. In a long-tailed class-conditional setting, gSR [35] proposes to regularize the exploding spectral norms of the class-specific parameters of the GAN. Collapse-by-conditioning [40] addresses the limited data in classes by introducing a training regime that transitions from an unconditional to a class-conditioned setting, thus exploiting the shared information across classes during the early stages of the training. However, these methods suffer from either class confusion or poor generated image quality on large datasets, which is resolved by NoisyTwins. + +Self-Supervised Learning for GANs. Ideas from Self-supervised learning have shown their benefits in GAN training. IC-GAN [6] trains GAN conditioned on embeddings on SwAV [5], which led to remarkable improvement in performance on the long-tailed version of ImageNet. InsGen [51] and ReACGAN [16, 17] introduce the auxiliary task of instance discrimination for the discriminator, thereby making the discriminator focus on multiple tasks and thus alleviating discriminator overfitting. While InsGen relies on both noise space and image space augmentations, ReACGAN and ContraGAN follow only image space augmentations. Contrary to these, NoisyTwins performs augmentations in the class-embedding space and contrasts them in the $\mathcal{W}$ -space of the generator instead of the discriminator. + +# 3. Preliminaries + +# 3.1. StyleGAN + +StyleGAN [21] is a Generative Adversarial Network comprising of its unique Style Conditioning Based Generator $(\mathcal{G})$ and discriminator network $(\mathcal{D})$ trained jointly. We will focus on the architecture of StyleGAN2 [23] as we use it in our experiments, although our work is generally applicable to all StyleGAN architectures. The StyleGAN2 generator is composed of blocks that progressively upsample the features and resolution inspired by Progressive GAN [19], starting from a single root image. The diversity in the images comes from conditioning each block of image generation through conditioning on the latent coming from the mapping network (Fig. 2). The mapping network is a fully connected network that takes in the conditioning variables, the $\mathbf{z} \in \mathbb{R}^d$ coming from a random distribution (e.g., Gaussian, etc.) and class conditioning label $c$ which is converted + +to an embedding $\mathbf{c} \in \mathbb{R}^d$ . The mapping network takes these and outputs vectors $\mathbf{w}$ in the $\mathcal{W}$ latent space of StyleGAN, which is found to be semantically disentangled to a high extent [49]. The $\mathbf{w}$ is then processed through an affine transform and passed to each generator layer for conditioning the image generation process through Adaptive Instance Normalization (AdaIN) [13]. The images from generator $\mathcal{G}$ , along with real images, are passed to discriminator $\mathcal{D}$ for training. The training utilizes the non-saturating adversarial losses [9] for $\mathcal{G}$ and $\mathcal{D}$ given as: + +$$ +\min _ {\mathcal {D}} \mathcal {L} _ {\mathcal {D}} = \sum_ {i = 1} ^ {m} \log (\mathcal {D} (\mathbf {x} _ {i})) + \log (1 - \mathcal {D} (\mathcal {G} (\mathbf {z} _ {i}, \mathbf {c} _ {i}))) \tag {1} +$$ + +$$ +\min _ {\mathcal {G}} \mathcal {L} _ {\mathcal {G}} = \sum_ {i = 1} ^ {m} - \log \left(\mathcal {D} \left(\mathcal {G} \left(\mathbf {z} _ {i}, \mathbf {c} _ {i}\right)\right)\right) \tag {2} +$$ + +We now describe the issues present in the StyleGANs trained on long-tailed data and their analysis in $\mathcal{W}$ space. + +# 3.2. Class Confusion and Class-Specific Mode Collapse in Conditional StyleGANs + +To finely analyze the performance of StyleGAN and its variants on long-tailed datasets, we train them on the CIFAR10-LT dataset. In Fig. 3, we plot the qualitative results of generated images and create a t-SNE plot for latents in $\mathcal{W}$ space for each class. We first train the StyleGAN2 baseline with augmentations (DiffAug) [20, 54]. We find that it leads to mode collapse, specifically for tail classes (Fig. 3). In conjunction with images, we also observe that corresponding t-SNE embeddings are also collapsed near each class's mean in $\mathcal{W}$ space. Further, recent methods which have proposed the usage of contrastive learning for GANs, improve their data efficiency and prevent discriminator overfitting [14, 16]. We also evaluate them by adding the contrastive conditioning method, which is D2DCE loss-based on ReACGAN [17], to the baseline, where in results, we observe that the network omits to learn tail classes and produces head class images at their place (i.e., class confusion). In Fig. 3, it can be seen that the network confuses semantically similar classes, that is, generating cars (head or majority class) in place of trucks and airplanes (head class) instead of ships. In the $\mathcal{W}$ space, we find the same number of clusters as the number of classes in the dataset. However, the tail label cluster images also belong to the head classes of cars and airplanes. In a very recent work gSR [35], it has been shown that constraining the spectral norm of $\mathcal{G}$ embedding parameters can help reduce the mode collapse and lead to stable training. However, we find that constraining the embeddings leads to class confusion, as seen in t-SNE visualization in Fig. 3. We find that this class confusion gets further aggravated when StyleGAN is trained on datasets like ImageNet-LT, which contain a large number of classes, along with a bunch of semantically similar classes (Sec. 5.2). Based on our $\mathcal{W}$ + +![](images/a811c8c06599313d1d44d9fb7fc8e236c148611c31d29e93bced5155cb9ad7e8.jpg) +Figure 3. Comparison of GANs and their $\mathcal{W}$ space for CIFAR10-LT. We plot the generated images on (left) and generate a t-SNE plot of w latents for generated images in $\mathcal{W}$ space (right). We find that mode collapse and class confusion in images is linked to the corresponding collapse and confusion in latent $\mathcal{W}$ space. Our proposed NoisyTwins mitigates both collapse (left) and confusion (right) simultaneously. + +![](images/ba2441057d6fd53a9b97ffebb6a7efe48f94bd72b0de4351575b78a6fdf133cc.jpg) + +![](images/13e9cd0abac8a49b6bc45c0656a00091e549c181cef6566fcfbe5916f0788de5.jpg) + +![](images/4563af6695895049295817ed017d1b32f0eb1173c31bd2a3f989fd79a80fbe3b.jpg) + +![](images/e5e02df03fadf4c2db0c35daccfdd79001539044451f35868abc86705e3f9c49.jpg) + +![](images/bc0db085d00bb489b207b957f0775ca2d63ec2d3a70a51d3f01d3982e7f0e3cf.jpg) + +space analysis and qualitative results above, we observe that the class confusion and mode collapse of images is tightly coupled with the structure of $\mathcal{W}$ space. Further, the recent SotA methods are either unable to prevent collapse or suffer from class confusion. Hence, this work aims to develop a technique that mitigates both confusion and collapse. + +# 4. Approach + +In this section, we present our method NoisyTwins, which introduces noise-based augmentation twins in the conditional embedding space (Sec. 4.1), and then combines it with the Barlow-Twins-based regularizer from the self-supervised learning (SSL) paradigm to resolve the issue of class confusion and mode collapse (Sec. 4.2). + +# 4.1. Noise Augmentation in Embedding Space + +As we observed in the previous section that $\mathbf{w}$ vectors for each sample become insensitive to changes in $\mathbf{z}$ . This collapse in $\mathbf{w}$ vectors for each class leads to mode collapse for baselines (Fig. 3). One reason for this could be the fact that $\mathbf{z}$ is composed of continuous variables, whereas the embedding vectors $\mathbf{c}$ for each class are discrete. Due to this, the GAN converges to the easy degenerate solution where it generates a single sample for each class, becoming insensitive to changes in $\mathbf{z}$ . For inducing some continuity in $\mathbf{c}$ embeddings vectors, we introduce an augmentation strategy where we add i.i.d. noise of small magnitude in each of the variables in $\mathbf{c}$ . Based on our observation (Fig. 3) and existing works [35], there is a high tendency for mode collapse in tail classes. Hence we add noise in embedding space that is + +proportional to the inverse of the frequency of samples. We provide the mathematical expression of noise augmentation $\tilde{\mathbf{c}}$ below: + +$$ +\tilde {\mathbf {c}} \sim \mathbf {c} + \mathcal {N} (\mathbf {0}, \sigma_ {c} \mathbb {I} _ {d}) \text {w h e r e} \sigma_ {c} = \sigma \frac {(1 - \alpha)}{1 - \alpha^ {n _ {c}}} \tag {3} +$$ + +Here $n_c$ is the frequency of training samples in class $c$ , $\mathbb{I}$ is the identity matrix of size $d \times d$ , and $\alpha, \sigma$ are hyperparameters. The expression of $\sigma_c$ is from the effective number of samples [8], which is a softer version of inverse frequency proportionality. In contrast to the image space augmentation, these noise augmentations come for free as there is no significant additional computation overhead. This noise is added in the embedding $\mathbf{c}$ before passing it to the generator and the discriminator, which ensures that the class embeddings occupy a continuous region in latent space. + +Insight: The augmentation equation above (Eq. 3) can be interpreted as approximating the discrete random variable $\mathbf{c}$ with a Gaussian with finite variance and the embedding parameters $\mathbf{c}$ being the mean $\mu_{\mathbf{c}}$ . + +$$ +\tilde {\mathbf {c}} \sim \mathcal {N} \left(\mu_ {\mathbf {c}}, \sigma_ {c} \mathbb {I} _ {d}\right) \tag {4} +$$ + +This leads to the class-embedding input $\tilde{\mathbf{c}}$ to mapping network to have a Gaussian distribution, similar in nature to $\mathbf{z}$ . This noise augmentation strategy alone mitigates the degenerate solution of class-wise mode collapse to a great extent (Table 1) and helps generate diverse latent $\mathbf{w}$ for each class. Due to the diverse $\mathbf{w}$ conditioning of the GAN, it leads to diverse image generation. + +# 4.2. Invariance in $\mathcal{W}$ -Space with NoisyTwins + +The augmentation strategy introduced in the previous section expands the region for each class in $\mathcal{W}$ latent space. + +![](images/efb94e7c8a3ba242f1b8e2fea2e43e6860f58772a173022a1344d604c67a202e.jpg) +Figure 4. Overview of NoisyTwins. For the $i^{th}$ sample of class $c^i$ , we create twin augmentations $(\tilde{\mathbf{c}}_a^i,\tilde{\mathbf{c}}_b^i)$ , by sampling from a Gaussian centered at class embedding $(\mu_{\mathbf{c}^i})$ . After this, we concatenate them with the same $\mathbf{z}^i$ and obtain $(\tilde{\mathbf{w}}_a^i,\tilde{\mathbf{w}}_b^i)$ from the mapping network, which we stack in batches of augmented latents ( $\tilde{\mathbf{W}}_{\mathbf{A}}$ and $\tilde{\mathbf{W}}_{\mathbf{B}}$ ). The twin $(\tilde{\mathbf{w}}_a^i,\tilde{\mathbf{w}}_b^i)$ vectors are then made invariant to augmentations (similar) in the latent space by minimizing cross-correlation [2,53] between the latents of two augmented batches ( $\tilde{\mathbf{W}}_{\mathbf{A}}$ and $\tilde{\mathbf{W}}_{\mathbf{B}}$ ). + +Although that does lead to diverse image generation as $\mathbf{w}$ are diverse; however, this does not ensure that these $\mathbf{w}$ will generate class-consistent outputs for augmentations in embedding $(\tilde{\mathbf{c}})$ . To ensure class consistent predictions, we need to ensure invariance in $\mathbf{w}$ to noise augmentation. + +For enforcing invariance to augmentations, a set of recent works [2, 10, 53] in self-supervised learning make the representations of augmentations similar through regularization. Among them, we focus on Barlow Twins as it does not require a large batch size of samples. Inspired by Barlow twins, we introduce NoisyTwins (Fig. 4), where we generate twin augmentations $\tilde{\mathbf{c}}_a$ and $\tilde{\mathbf{c}}_b$ of the same class embedding $(\mu_c)$ and concatenate them to same $\mathbf{z}$ . After creating a batch of such inputs, they are passed to the mapping network to get batches of augmented latents $(\tilde{\mathbf{W}}_A$ and $\tilde{\mathbf{W}}_B)$ . These batches are then used to calculate the cross-correlation matrix of latent variables given as: + +$$ +\mathbf {C} _ {j, k} = \frac {\sum_ {(\tilde {\mathbf {w}} _ {a} , \tilde {\mathbf {w}} _ {b}) \in (\tilde {\mathbf {W}} _ {A} , \tilde {\mathbf {W}} _ {B})} \tilde {\mathbf {w}} _ {a} ^ {j} \tilde {\mathbf {w}} _ {b} ^ {k}}{\sum_ {\tilde {\mathbf {w}} _ {a} \in \tilde {\mathbf {W}} _ {A}} \tilde {\mathbf {w}} _ {a} ^ {j} \tilde {\mathbf {w}} _ {a} ^ {j} \sum_ {\tilde {\mathbf {w}} _ {b} \in \tilde {\mathbf {W}} _ {B}} \tilde {\mathbf {w}} _ {b} ^ {k} \tilde {\mathbf {w}} _ {b} ^ {k}} \tag {5} +$$ + +The matrix $\mathbf{C}$ is a square matrix of size same as of latents $\mathbf{w}$ . The final loss based on confusion matrix is given as: + +$$ +\mathcal {L} _ {\mathcal {N T}} = \sum_ {j} \left(1 - \mathbf {C} _ {j j} ^ {2}\right) + \gamma \sum_ {j \neq k} \mathbf {C} _ {j, k} ^ {2} \tag {6} +$$ + +The first term tries to make the two latents $(\tilde{\mathbf{w}}_a$ and $\tilde{\mathbf{w}}_b)$ invariant to the noise augmentation applied (i.e. similar), whereas the second term tries to de-correlate the different variables, thus maximizing the information content of the embedding vector [53]. The $\gamma$ is the hyper-parameter that determines the relative importance of the two terms. This loss is then added to the generator loss term $(\mathcal{L}_{\mathcal{G}} + \lambda L_{\mathcal{NT}})$ and optimized through backpropagation. The above procedure comprises our proposed method, NoisyTwins (Fig. 4), which we empirically evaluate in the subsequent sections. + +# 5. Experimental Evaluation + +# 5.1. Setup + +Datasets: We primarily apply all methods on long-tailed datasets, as GANs trained on them are more prone to class confusion and mode collapse. We first report on the commonly used CIFAR10-LT dataset with an imbalance factor (i.e. ratio of most to least frequent class) of 100. To show our approach's scalability and real-world application, we test our method on the challenging ImageNet-LT and iNaturalist 2019 datasets. The ImageNet-LT [30] is a long-tailed variant of the 1000 class ImageNet dataset, with a plethora of semantically similar classes (e.g., Dogs, Birds etc.), making it challenging to avoid class confusion. iNaturalist-2019 [48] is a real-world long-tailed dataset composed of 1010 different variants of species, some of which have fine-grained differences in their appearance. For such fine-grained datasets, ImageNet pre-trained discriminators [38, 39] may not be useful, as augmentations used to train the model makes it invariant to fine-grained changes. + +Training Configuration: We use StyleGAN2 architecture for all our experiments. All our experiments are performed using PyTorch-StudioGAN implemented by Kang et al. [18], which serves as a base for our framework. We use Path Length regularization (PLR) as it ensures changes in $\mathcal{W}$ space lead to changes in images, using a delayed PLR on ImageNet-LT following [39]. We use a batch size of 128 for all our experiments, with one G step per D step. Unless stated explicitly, we used the general training setup for StyleGANs from [18], including methods like $R_{1}$ regularization [31], etc. More details on the exact training configuration for each dataset are provided in Appendix. + +Metrics: In this work we use the following metrics for evaluation of our methods: + +a) FID: Fréchet Inception Distance [12] is the Wasserstein-2 Distance between the real and validation data. We use a + +![](images/ff8ec5d9635145f686f068454e085ede3c0e1758b811ea2946bf87848554e0be.jpg) +Figure 5. Choice of Eval. backbone: intra-FID (iFID) of a class based on InceptionV3 backbone (left plot) is not able to capture the mode collapse (increase in iFID in the absence of mode collapse). This is well-captured by $\mathrm{iFID}_{\mathrm{CLIP}}$ based on CLIP [34] backbone (right plot, decrease in iFID in the absence of mode collapse). + +held-out validation set and $50\mathrm{k}$ generated samples to evaluate FID in each case. As FID is biased towards ImageNet and can be arbitrarily manipulated, we also report $\mathrm{FID}_{\mathrm{CLIP}}$ . b) Precision & Recall: As we aim to mitigate the mode collapse and achieve diverse generations across classes, we use improved Precision & Recall [25] metrics, as poor recall indicates mode collapse [39]. + +d) Intra-Class $\mathbf{FID}_{\mathrm{CLIP}}$ ( $\mathbf{iFID}_{\mathrm{CLIP}}$ ): The usage of only FID based on Inception-V3 Networks for evaluation of Generative Models has severe limitations, as it has been found that FID can be reduced easily by some fringe features [26]. iFID is computed by taking FID between 5k generated and real samples for the same class. As we want to evaluate both class consistency and diversity, we find that similar limitations exist for intra-class FID (iFID), which has been used to evaluate class-conditional GANs [18]. In Fig. 5, we show the existence of generated images for a particular class (more in Appendix) from models trained on iNaturalist 2019, where iFID is better for the mode collapsed model than the other model generating diverse images. Whereas the $\mathbf{iFID}_{\mathrm{CLIP}}$ , based on CLIP backbone can rank the models correctly with the model having mode collapse having high $\mathbf{iFID}_{\mathrm{CLIP}}$ . Further, we find that the mean iFID can be deceptive in detecting class confusion and collapse cases, as it sometimes ranks models with high realism better than models generating diversity (See Appendix). Hence, mean $\mathbf{iFID}_{\mathrm{CLIP}}$ (ref. to as $\mathbf{iFID}_{\mathrm{CLIP}}$ in result section for brevity) can be reliably used to evaluate models for class consistency and diversity. + +Baselines: For evaluating NoisyTwins performance in comparison to other methods, we use the implementations present in StudioGAN [16]. For fairness, we re-run all the baselines on StyleGAN2 in the same hyperparameter setting. We compare our method to the StyleGAN2 + +(SG2) and StyleGAN2 with augmentation (DiffAug [54] and ADA [20]) baselines. We further tried to improve the baseline by incorporating the recent LeCam regularization method; however, it resulted in gains only for the iNaturalist 2019 dataset, where we use LeCam for all experiments. Further on StyleGAN2, we also use contrastive D2D-CE loss conditioning (i.e. ReACGAN) as a baseline. However, the D2D-CE baseline completely ignores learning of tail classes (Fig. 3) for CIFAR10-LT and is expensive to train; hence we do not report results for it for large-scale long-tailed datasets. We also compare our method against the recent SotA group Spectral Normalization (gSR) [35] method, which we implement for StyleGAN2 by constraining the spectral norms of embedding parameters of the generator $(\mathcal{G})$ as suggested by authors. As a sanity check, we reproduce their results on CIFAR10-LT and find that our implementation matches the reported results correctly. We provide results on all datasets, for both the proposed Noise Augmentation (+ Noise) and the overall proposed NoisyTwins (+NoisyTwins) method. + +# 5.2. Results on Long-Tailed Data Distributions + +CIFAR10-LT. We applied DiffAug [54] on all baselines, except on gSR, where we found that DiffAug provides inferior results compared to ADA (as also used by authors [35]). It can be observed in Table 2 that the addition of NoisyTwins regularization significantly improves over baseline by $(\sim 14$ FID) along with providing superior class consistency as shown by improved iFIDCLIP. NoisyTwins is also able to outperform the recent gSR regularization method and achieves improved results for all metrics. Further, NoisyTwins improves FID for StyleGAN2-ADA baseline used by gSR too from 32.08 to 23.02, however the final results are inferior than reported DiffAug baseline results. Further, we observed that despite not producing any tail class images (Fig. 3), the D2D-CE baseline has much superior FID in comparison to baselines. Whereas the proposed iFIDCLIP value is similar for the baseline and D2D-CE model. This clearly demonstrates the superiority of proposed iFIDCLIP in detecting class confusion. + +Real-world, Large-scale Long-Tailed Datasets. We experiment with iNaturalist 2019 and ImageNet-LT. These datasets are particularly challenging as they contain long-tailed imbalances and semantically similar classes, making GANs prone to mode collapse and class confusion. The baselines StyleGAN2 and StyleGAN2-ADA both suffer from mode collapse (Fig. 6), particularly for the tail classes. Whereas for the recent SotA gSR method, we find that although it undergoes less collapse in comparison to baselines, it suffers from class confusion as seen from similar Intra-FIDCLIP in comparison to baselines (Table 1). Compared to that, our method NoisyTwins improves when used with StyleGAN2-ADA significantly, leading to + +Table 1. Quantitative results on ImageNet-LT and iNaturalist 2019 Datasets. We compare FID(↓), FIDCLIP(↓), iFIDCLIP(↓), Precision(↑) and Recall(↑) with other existing approaches on StyleGAN2 (SG2). We obtain an average ~ 19% relative improvement on FID, ~ 33% on FIDCLIP, and ~ 11% on iFIDCLIP metrics over the previous SotA on ImageNet-LT and iNaturalist 2019 datasets. + +
MethodImageNet-LTiNaturalist 2019
FID(↓)FIDCLIP(↓)iFIDCLIP(↓)Precision(↑)Recall(↑)FID(↓)FIDCLIP(↓)iFIDCLIP(↓)Precision(↑)Recall(↑)
SG2 [23]41.2511.6446.930.500.4819.343.3338.240.740.17
SG2+ADA [20]37.2011.0447.410.540.3814.922.3035.190.750.57
SG2+ADA+gSR [35]24.788.2144.420.630.3515.172.0636.220.740.46
SG2+ADA+Noise (Ours)22.177.1141.200.720.3312.871.3731.430.810.63
+ NoisyTwins (Ours)21.296.4139.740.670.4911.461.1431.500.790.67
+ +Table 2. Quantitative results on CIFAR10-LT Dataset. We compare with other existing approaches. We obtain $\sim 26\%$ relative improvement over the existing methods on $\mathrm{FID}_{\mathrm{CLIP}}$ and iFIDCLIP metrics. + +
MethodFID(↓)FIDCLIP(↓)iFIDCLIP(↓)Precision(↑)Recall(↑)
SG2+DiffAug [54]31.736.2711.590.630.35
SG2+D2D-CE [17]19.974.7711.350.730.42
gSR [35]22.105.549.940.700.29
SG2+DiffAug+Noise (Ours)28.905.2610.650.710.38
+ NoisyTwins(Ours)17.743.557.240.700.51
+ +Table 3. Comparison with SotA approaches on BigGAN. We compare FID(↓) with other existing models on ImageNet-LT (IN-LT) and iNaturalist 2019 (iNat-19). + +
MethodiNat-19IN-LT
BigGAN [4]14.8528.10
+ gSR [35]13.95-
ICGAN [6]-23.40
StyleGAN2-ADA [20]14.9237.20
+ NoisyTwins (Ours)11.4621.29
+ +a relative improvement of $42.7\%$ in FID for ImageNet-LT and $23.19\%$ on the iNaturalist 2019 dataset when added to StyleGAN2-ADA baseline. Further with Noise Augmentation $(+\mathrm{Noise})$ , we observe generations of high-quality class-consistent images, but it also suffers from mode collapse. This can be observed by the high-precision values in comparison to low-recall values. However, adding NoisyTwins regularization over the noise augmentation improves diversity by improving recall (Table 1). + +Fig. 6 presents the generated images of tail classes for various methods on ImageNet-LT, where NoisyTwins generations show remarkable diversity in comparison to others. The presence of diversity for classes with just 5-6 training images demonstrates successful transfer of knowledge from head classes to tail classes, due to shared parameters. Further, to compare with existing SotA reported results, we compare FID of BigGAN models from gSR [35] and Instance Conditioned GAN (ICGAN) [6]. For fairness, we compare FID on the validation set for which we obtained gSR models from authors and re-evaluate them, as they reported FID on a balanced training set. As BigGAN models are more common for class-conditioned generation [18], their baseline performs superior to StyleGAN2-ADA baselines (Table 3). However, the addition of NoisyTwins to the StyleGAN2-ADA method improves it significantly, even outperforming the existing methods of gSR (by $18.44\%$ ) and ICGAN (by $9.44\%$ ) based on BigGAN architecture. This shows that NoisyTwins allows the StyleGAN2 baseline to scale to large and diverse long-tailed datasets. + +# 5.3. NoisyTwins on Few-Shot Datasets + +We now demonstrate the potential of NoisyTwins in another challenging scenario of class-conditional few-shot + +image generation from GANs. We perform our experiments using a conditional StyleGAN2-ADA baseline, for which we tune hyper-parameters to obtain a strong baseline. We then apply our method of Noise Augmentation and NoisyTwins over the strong baseline for reporting our results. We use the few-shot dataset of LHI-AnimalFaces [45] and a subset of ImageNet Carnivores [29, 40] to report our results. Table 4 shows the results of these experiments, where we find that our method, NoisyTwins, significantly improves the FID of StyleGAN2 ADA baseline by $(22.2\%)$ on average for both datasets. Further, combining Noisy Twins with SotA Transitional-cGAN [40] through official code, also leads to effective improvement in FID. These results clearly demonstrate the diverse potential and applicability of our proposed method NoisyTwins. + +# 6. Analysis + +We perform analysis of NoisyTwins w.r.t. to its hyperparameters, standard deviation $(\sigma)$ of noise augmentation and regularization strength $(\lambda)$ . We also compare NoisyTwins objective (ref. Eq. 6) with contrastive objective. Finally, we compare NoisyTwins over Latent Diffusion Models for long-tailed class conditional generation task. We perform ablation experiments on CIFAR10-LT, for which additional details and results are present in Appendix. We also present comparison of NoisyTwins for GAN fine-tuning. + +# How much noise and regularization strength is optimal? + +In Fig. 7, we ablate over the noise variance parameter $\sigma$ for CIFAR10-LT. We find that a moderate value of noise strength 0.75 leads to optimal results. For the strength of NoisyTwins loss $(\lambda)$ , we find that the algorithm performs similarly on values near 0.01 and is robust to it (Fig. 7). + +![](images/71c2871399f1044f6cba014b9cef28dfd0f65057b005f7d878839c3bc5d097ce.jpg) +(a) Image Augmentation based Methods (ADA) + +![](images/81ede3449877dabd306eeb752501b5a369e7735973a56bdf84fc7984f41a5836.jpg) +(b) Regularizer based Methods (gSR) + +![](images/6d18060b187b405762c6daf9bd05455025eb712304f459eba2dd10d557b74dd1.jpg) +(c) NoisyTwins (Ours) + +Table 4. Quantitative results on ImageNet Carnivore and AnimalFace Datasets. Our method improves over both StyleGAN2-ADA (SG2-ADA) baseline and SotA Transitional-cGAN. + +
ImageNet CarnivoreAnimalFace
MethodFID(↓)iFIDCLIP(↓)FID(↓)iFIDCLIP(↓)
SG2 [23]111.8336.3494.0929.94
SG2+ADA [20]22.7712.8520.2511.12
SG2+ADA+Noise (Ours)19.2512.5118.7810.42
+ NoisyTwins (Ours)16.0112.4117.2710.03
FID(↓)FID(↓)
Transitional-cGAN [40]14.6020.53
+ NoisyTwins (Ours)13.6516.15
+ +Which type of self-supervision to use with noise augmentation? The goal of our method is to achieve invariance to Noise Augmentation in the $\mathcal{W}$ latent space. This can be achieved using either contrastive learning-based methods like SimCLR [7] or negative-free method like Barlow Twins [53]. Contrastive loss (SimCLR based) produces FID of 26.23 vs 17.74 by NoisyTwins (BarlowTwins based). We find that contrastive baseline improves over the noise augmentation baseline (28.90) however falls significantly below the NoisyTwins, as the former requires a large batch size to be effective which is expensive for GANs. + +How does NoisyTwins compare with modern Vision and Language models? For evaluating the effectiveness of modern vision language-based diffusion models, we test the generation of the iNaturalist 2019 dataset by creating the prompt "a photo of S" where we replace the class name in place of S. We use the LDM [37] model trained on LAION-400M to perform inference, generating 50 images per class. We obtained an FID of 57.04 in comparison to best FID of 11.46 achieved by NoisyTwins. This clearly demonstrates that for specific use cases like fine-grained generation, GANs are still ahead of general-purpose LDM. + +![](images/57d1a50ade434231c98ee56695e446f0c59a3da443295139f4ff82d16eb117e0.jpg) +Figure 6. Qualitative results on ImageNet-LT for tail classes. We find that existing SotA methods for tail classes show collapsed (a) or arbitrary image generation (b). With NoisyTwins, we observe diverse and class-consistent image generation, even for classes having 5-6 images. The tail classes get enhanced diversity by transferring the knowledge from head classes, as they share parameters. +Figure 7. Ablation of Hyperparameters. Quantitative comparison on CIFAR10-LT for standard deviation of Noise Augmentation $(\sigma)$ and strength $(\lambda)$ of NoisyTwins loss. + +# 7. Conclusion + +In this work, we analyze the performance of StyleGAN2 models on the real-world long-tailed datasets including iNaturalist 2019 and ImageNet-LT. We find that existing works lead to either class confusion or mode collapse in the image space. This phenomenon is rooted in collapse and confusion in the latent $\mathcal{W}$ space of StyleGAN2. Through our analysis, we deduce that this collapse occurs when the latents become invariant to random conditioning vectors $\mathbf{z}$ , and collapse for each class. To mitigate this, we introduce inexpensive noise based augmentation for discrete class embeddings. Further, to ensure class consistency, we couple this augmentation technique with BarlowTwins' objective in the latent $\mathcal{W}$ space which imparts intra-class diversity to latent $\mathbf{w}$ vectors. The noise augmentation and regularization comprises our proposed NoisyTwins technique, which improves the performance of StyleGAN2 establishing a new SotA on iNaturalist 2019 and ImageNet-LT. The extension of NoisyTwins for conditioning on more-complex attributes for StyleGANs is a good direction for future work. Acknowledgements: This work was supported in part by SERB-STAR Project (STR/2020/000128). Harsh Rangwani is supported by PMRF fellowship. + +# References + +[1] Yuval Alaluf, Omer Tov, Ron Mokady, Rinon Gal, and Amit Bermano. Hyperstyle: Stylegan inversion with hypernetworks for real image editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18511-18521, June 2022. 2 +[2] Adrien Bardes, Jean Ponce, and Yann LeCun. VI-CReg: Variance-invariance-covariance regularization for self-supervised learning. In International Conference on Learning Representations, 2022. 2, 5 +[3] Stella Bounareli, Vasileios Argyriou, and Georgios Tzimiropoulos. Finding directions in gan's latent space for neural face reenactment. arXiv preprint arXiv:2202.00046, 2022.1 +[4] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale GAN training for high fidelity natural image synthesis. In International Conference on Learning Representations, 2019. 7 +[5] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. 2020. 3 +[6] Arantxa Casanova, Marlene Careil, Jakob Verbeek, Michal Drozdal, and Adriana Romero-Soriano. Instance-conditioned gan. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 3, 7 +[7] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. arXiv preprint arXiv:2002.05709, 2020.8 +[8] Yin Cui, Menglin Jia, Tsung-Yi Lin, Yang Song, and Serge Belongie. Class-balanced loss based on effective number of samples. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9268-9277, 2019. 4 +[9] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139-144, 2020. 3 +[10] Jean-Bastien Grill, Florian Strub, Florent Altché, Coretin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 5 +[11] Erik Härkönen, Aaron Hertzmann, Jaakko Lehtinen, and Sylvain Paris. Ganspace: Discovering interpretable gan controls. Advances in Neural Information Processing Systems, 33:9841-9850, 2020. 1 +[12] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 2, 5 +[13] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceed- + +ings of the IEEE international conference on computer vision, pages 1501-1510, 2017. 3 +[14] Jongheon Jeong and Jinwoo Shin. Training GANs with stronger augmentations via contrastive discriminator. In International Conference on Learning Representations, 2021. 3 +[15] Liming Jiang, Bo Dai, Wayne Wu, and Chen Change Loy. Deceive D: Adaptive Pseudo Augmentation for GAN training with limited data. In NeurIPS, 2021. 2 +[16] Minguk Kang and Jaesik Park. ContraGAN: Contrastive Learning for Conditional Image Generation. 2020. 3, 6 +[17] Minguk Kang, Woohyeon Shim, Minsu Cho, and Jaesik Park. Rebooting acgan: Auxiliary classifier gans with stable training. Advances in neural information processing systems, 34:23505-23518, 2021. 2, 3, 7 +[18] MinGuk Kang, Joonghyuk Shin, and Jaesik Park. StudioGAN: A Taxonomy and Benchmark of GANs for Image Synthesis. 2206.09479 (arXiv), 2022. 5, 6, 7 +[19] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. arXiv preprint arXiv:1710.10196, 2017. 3 +[20] Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. In Proc. NeurIPS, 2020. 1, 2, 3, 6, 7, 8 +[21] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 1, 2, 3 +[22] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. 1 +[23] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. In Proc. CVPR, 2020. 3, 7, 8 +[24] Tuomas Kynkänniemi, Tero Karras, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Improved Precision and Recall Metric for Assessing Generative Models. 2019. 1 +[25] Tuomas Kynkänniemi, Tero Karras, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Improved precision and recall metric for assessing generative models. Advances in Neural Information Processing Systems, 32, 2019. 6 +[26] Tuomas Kynkänniemi, Tero Karras, Miika Aittala, Timo Aila, and Jaakko Lehtinen. The role of imagenet classes in fréchet inception distance. CoRR, abs/2203.06026, 2022. 6 +[27] Thomas Leimkuhler and George Drettakis. Freestylegan: Free-view editable portrait rendering with the camera manifold. 40(6), 2021. 2 +[28] Feng Liu and Xiaoming Liu. 2d gans meet unsupervised single-view 3d reconstruction. In Computer Vision - ECCV 2022, pages 497-514, 2022. 2 +[29] Ming-Yu Liu, Xun Huang, Arun Mallya, Tero Karras, Timo Aila, Jaakko Lehtinen, and Jan Kautz. Few-shot unsupervised image-to-image translation. In Proceedings of + +the IEEE/CVF international conference on computer vision, pages 10551-10560, 2019. 7 +[30] Ziwei Liu, Zhongqi Miao, Xiaohang Zhan, Jiayun Wang, Boqing Gong, and Stella X Yu. Large-scale long-tailed recognition in an open world. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2537-2546, 2019. 2, 5 +[31] Lars Mescheder, Andreas Geiger, and Sebastian Nowozin. Which training methods for gans do actually converge? In International conference on machine learning, pages 3481-3490. PMLR, 2018. 5 +[32] Rishubh Parihar, Ankit Dhiman, Tejan Karmali, and R. Venkatesh Babu. Everything is there in latent space: Attribute editing and attribute style manipulation by stylegan latent space exploration. In Proceedings of the 30th ACM International Conference on Multimedia, pages 1828-1836, 2022. 2 +[33] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2085–2094, October 2021. 2 +[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2, 6 +[35] Harsh Rangwani, Naman Jaswani, Tejan Karmali, Varun Jampani, and R. Venkatesh Babu. Improving gans for long-tailed data through group spectral regularization. In European Conference on Computer Vision, 2022. 2, 3, 4, 6, 7 +[36] Harsh Rangwani, Konda Reddy Mopuri, and R Venkatesh Babu. Class balancing gan with a classifier in the loop. In Uncertainty in Artificial Intelligence, pages 1618-1627. PMLR, 2021. 3 +[37] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 8 +[38] Axel Sauer, Kashyap Chitta, Jens Müller, and Andreas Geiger. Projected gans converge faster. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 1, 5 +[39] Axel Sauer, Katja Schwarz, and Andreas Geiger. Styleganx1: Scaling stylegan to large diverse datasets. volume abs/2201.00273, 2022. 1, 5, 6 +[40] Mohamad Shahbazi, Martin Danelljan, Danda Pani Paudel, and Luc Van Gool. Collapse by conditioning: Training class-conditional GANs with limited data. In International Conference on Learning Representations, 2022. 3, 7, 8 +[41] Yujun Shen, Jinjin Gu, Xiaou Tang, and Bolei Zhou. Interpreting the latent space of gans for semantic face editing. In CVPR, 2020. 1 +[42] Yujun Shen, Ceyuan Yang, Xiaou Tang, and Bolei Zhou. Interfacegan: Interpreting the disentangled face representation learned by gans. IEEE TPAMI, 2020. 2 +[43] Yujun Shen and Bolei Zhou. Closed-form factorization of latent semantics in gans. In CVPR, 2021. 2 + +[44] Yichun Shi, Divyansh Aggarwal, and Anil K. Jain. Lifting 2d stylegan for 3d-aware face generation. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6254-6262, 2021. 2 +[45] Zhangzhang Si and Song-Chun Zhu. Learning hybrid image templates (hit) by information projection. IEEE Transactions on pattern analysis and machine intelligence, 34(7):1354-1367, 2011. 7 +[46] Ivan Skorokhodov, Sergey Tulyakov, and Mohamed Elhoseiny. Stylegan-v: A continuous video generator with the price, image quality and perks of stylegan2, 2021. 1 +[47] Hung-Yu Tseng, Lu Jiang, Ce Liu, Ming-Hsuan Yang, and Weilong Yang. Regularizing generative adversarial networks under limited data. In CVPR, 2021. 3 +[48] Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alex Shepard, Hartwig Adam, Pietro Perona, and Serge Belongie. The inaturalist species classification and detection dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8769-8778, 2018. 2, 5 +[49] Zongze Wu, Dani Lischinski, and Eli Shechtman. Stylespace analysis: Disentangled controls for stylegan image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12863–12872, 2021. 1, 3 +[50] Ceyuan Yang, Yujun Shen, Yinghao Xu, Deli Zhao, Bo Dai, and Bolei Zhou. Improving gans with a dynamic discriminator. 2022. 3 +[51] Ceyuan Yang, Yujun Shen, Yinghao Xu, and Bolei Zhou. Data-efficient instance generation from instance discrimination. In A. Beygelzimer, Y. Dauphin, P. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, 2021. 3 +[52] Sihyun Yu, Jihoon Tack, Sangwoo Mo, Hyunsu Kim, Junho Kim, Jung-Woo Ha, and Jinwoo Shin. Generating videos with dynamics-aware implicit generative adversarial networks. In International Conference on Learning Representations, 2022. 1 +[53] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow twins: Self-supervised learning via redundancy reduction. In International Conference on Machine Learning, pages 12310-12320. PMLR, 2021. 2, 5, 8 +[54] Shengyu Zhao, Zhijian Liu, Ji Lin, Jun-Yan Zhu, and Song Han. Differentiable augmentation for data-efficient gan training. In Conference on Neural Information Processing Systems (NeurIPS), 2020. 1, 2, 3, 6, 7 \ No newline at end of file diff --git a/2023/NoisyTwins_ Class-Consistent and Diverse Image Generation Through StyleGANs/images.zip b/2023/NoisyTwins_ Class-Consistent and Diverse Image Generation Through StyleGANs/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..557e46f2492c002fe02b1d9a9306567e01e649e0 --- /dev/null +++ b/2023/NoisyTwins_ Class-Consistent and Diverse Image Generation Through StyleGANs/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ce2b90791a4f7c8770e8842be9639a4f47eb7b291d736c2b6c5ec382c516ee92 +size 671805 diff --git a/2023/NoisyTwins_ Class-Consistent and Diverse Image Generation Through StyleGANs/layout.json b/2023/NoisyTwins_ Class-Consistent and Diverse Image Generation Through StyleGANs/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..badce446e7e72b485583fb9892c6c916fadea00e --- /dev/null +++ b/2023/NoisyTwins_ Class-Consistent and Diverse Image Generation Through StyleGANs/layout.json @@ -0,0 +1,9036 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 47, + 103, + 547, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 103, + 547, + 122 + ], + "spans": [ + { + "bbox": [ + 47, + 103, + 547, + 122 + ], + "type": "text", + "content": "NoisyTwins: Class-Consistent and Diverse Image Generation through StyleGANs" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 110, + 142, + 484, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 142, + 484, + 171 + ], + "spans": [ + { + "bbox": [ + 110, + 142, + 484, + 171 + ], + "type": "text", + "content": "Harsh Rangwani" + }, + { + "bbox": [ + 110, + 142, + 484, + 171 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 110, + 142, + 484, + 171 + ], + "type": "text", + "content": " Lavish Bansal" + }, + { + "bbox": [ + 110, + 142, + 484, + 171 + ], + "type": "inline_equation", + "content": "^{1,3*}" + }, + { + "bbox": [ + 110, + 142, + 484, + 171 + ], + "type": "text", + "content": " Kartik Sharma" + }, + { + "bbox": [ + 110, + 142, + 484, + 171 + ], + "type": "inline_equation", + "content": "^{1,4}" + }, + { + "bbox": [ + 110, + 142, + 484, + 171 + ], + "type": "text", + "content": " Tejan Karmali" + }, + { + "bbox": [ + 110, + 142, + 484, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 110, + 142, + 484, + 171 + ], + "type": "text", + "content": " Varun Jampani" + }, + { + "bbox": [ + 110, + 142, + 484, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 110, + 142, + 484, + 171 + ], + "type": "text", + "content": " R. Venkatesh Babu" + }, + { + "bbox": [ + 110, + 142, + 484, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 171, + 533, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 171, + 533, + 186 + ], + "spans": [ + { + "bbox": [ + 61, + 171, + 533, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 171, + 533, + 186 + ], + "type": "text", + "content": " Vision and AI Lab, IISc Bangalore " + }, + { + "bbox": [ + 61, + 171, + 533, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 171, + 533, + 186 + ], + "type": "text", + "content": " Google Research " + }, + { + "bbox": [ + 61, + 171, + 533, + 186 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 61, + 171, + 533, + 186 + ], + "type": "text", + "content": " IIT BHU Varanasi " + }, + { + "bbox": [ + 61, + 171, + 533, + 186 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 61, + 171, + 533, + 186 + ], + "type": "text", + "content": " BITS Pilani" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 213, + 192, + 226 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 226 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 226 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 239, + 290, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 239, + 290, + 443 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 290, + 443 + ], + "type": "text", + "content": "StyleGANs are at the forefront of controllable image generation as they produce a latent space that is semantically disentangled, making it suitable for image editing and manipulation. However, the performance of StyleGANs severely degrades when trained via class-conditioning on large-scale long-tailed datasets. We find that one reason for degradation is the collapse of latents for each class in the " + }, + { + "bbox": [ + 46, + 239, + 290, + 443 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 46, + 239, + 290, + 443 + ], + "type": "text", + "content": " latent space. With NoisyTwins, we first introduce an effective and inexpensive augmentation strategy for class embeddings, which then decorrelates the latents based on self-supervision in the " + }, + { + "bbox": [ + 46, + 239, + 290, + 443 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 46, + 239, + 290, + 443 + ], + "type": "text", + "content": " space. This decorrelation mitigates collapse, ensuring that our method preserves intraclass diversity with class-consistency in image generation. We show the effectiveness of our approach on large-scale real-world long-tailed datasets of ImageNet-LT and iNaturalist 2019, where our method outperforms other methods by " + }, + { + "bbox": [ + 46, + 239, + 290, + 443 + ], + "type": "inline_equation", + "content": "\\sim 19\\%" + }, + { + "bbox": [ + 46, + 239, + 290, + 443 + ], + "type": "text", + "content": " on FID, establishing a new state-of-the-art." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 468, + 128, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 468, + 128, + 480 + ], + "spans": [ + { + "bbox": [ + 47, + 468, + 128, + 480 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 488, + 289, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 488, + 289, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 488, + 289, + 693 + ], + "type": "text", + "content": "StyleGANs [21, 22] have shown unprecedented success in image generation, particularly on well-curated and articulated datasets (eg. FFHQ for face images, etc.). In addition to generating high fidelity and diverse images, StyleGANs also produce a disentangled latent space, which is extensively used for image editing and manipulation tasks [49]. As a result, StyleGANs are being extensively used in various applications like face-editing [11, 41], video generation [46, 52], face reenactment [3], etc., which are a testament to their usability and generality. However, despite being successful on well-curated datasets, training StyleGANs on in-the-wild and multi-category datasets is still challenging. A large-scale conditional StyleGAN (i.e. StyleGAN-XL) on ImageNet was recently trained successfully by Sauer et al. [39] using the ImageNet pre-trained model through the idea of a projection discriminator [38]. While the StyleGAN-XL uses additional pre-trained mod" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 310, + 214, + 545, + 419 + ], + "blocks": [ + { + "bbox": [ + 310, + 214, + 545, + 419 + ], + "lines": [ + { + "bbox": [ + 310, + 214, + 545, + 419 + ], + "spans": [ + { + "bbox": [ + 310, + 214, + 545, + 419 + ], + "type": "image", + "image_path": "294d72e02624c3cfda3606df3ebfcf6ef336ce375e584f3d0a8017284a84da77.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 427, + 547, + 484 + ], + "lines": [ + { + "bbox": [ + 304, + 427, + 547, + 484 + ], + "spans": [ + { + "bbox": [ + 304, + 427, + 547, + 484 + ], + "type": "text", + "content": "Figure 1. Qualitative Comparison on tail classes (T1-T4) for iNaturalist 2019. We provide sample(s) from real class (with class frequency), generated by StyleGAN2-ADA and after adding proposed NoisyTwins. NoisyTwins achieves remarkable diversity, class-consistency and quality by just using 38 samples on average." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 495, + 547, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 495, + 547, + 532 + ], + "spans": [ + { + "bbox": [ + 304, + 495, + 547, + 532 + ], + "type": "text", + "content": "els, obtaining such models for distinctive image domains like medical, forensics, and fine-grained data may not be feasible, which limits its generalization across domains." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "content": "In this work, we aim to train vanilla class-conditional StyleGAN without any pre-trained models on challenging real-world long-tailed data distributions. As training StyleGAN with augmentations [20, 54] leads to low recall [24] (which measures diversity in the generated images) and mode collapse, particularly for minority (i.e. tail) classes. For investigating this phenomenon further, we take a closer look at the latent " + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "content": " space of StyleGAN that is produced by a fully-connected mapping network that takes the conditioning variables " + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "content": " (i.e. random noise) and class embedding " + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "content": " as inputs. The vectors " + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "content": " space are used for conditioning various layers of the generator (Fig. 2). We find that output vectors " + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "content": " from the mapping network hinge on the conditioning variable " + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "content": " and become invariant to random conditioning vector " + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "content": ". This collapse of latents leads to un" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 271, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 271, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 271, + 713 + ], + "type": "text", + "content": "*Equal Contribution. Link: rangwani-harsh.github.io/NoisyTwins" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "text", + "content": "5987" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 71, + 547, + 171 + ], + "blocks": [ + { + "bbox": [ + 47, + 71, + 547, + 171 + ], + "lines": [ + { + "bbox": [ + 47, + 71, + 547, + 171 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 547, + 171 + ], + "type": "image", + "image_path": "f35637b74c73f94493ea4064e32564bcbdf97cdff0ece6410e380c05a3e53b63.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 178, + 547, + 203 + ], + "lines": [ + { + "bbox": [ + 46, + 178, + 547, + 203 + ], + "spans": [ + { + "bbox": [ + 46, + 178, + 547, + 203 + ], + "type": "text", + "content": "Figure 2. Schematic illustration of " + }, + { + "bbox": [ + 46, + 178, + 547, + 203 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 46, + 178, + 547, + 203 + ], + "type": "text", + "content": " space for different GANs. Existing conditioning methods either suffer from mode collapse [20] or lead to class confusion [35] in " + }, + { + "bbox": [ + 46, + 178, + 547, + 203 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 46, + 178, + 547, + 203 + ], + "type": "text", + "content": " space. With proposed NoisyTwins, we achieve intra class diversity while avoiding class confusion." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 208, + 288, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 208, + 288, + 280 + ], + "spans": [ + { + "bbox": [ + 46, + 208, + 288, + 280 + ], + "type": "text", + "content": "stable training and is one of the causes of poor recall (a.k.a. mode collapse) for minority classes. Further, on augmenting StyleGAN with recent conditioning and regularization techniques [17, 35], we find that they either lead to a poor recall for minority classes or lead to class confusion (Fig. 2) instead of mitigating the collapse." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 283, + 287, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 283, + 287, + 427 + ], + "spans": [ + { + "bbox": [ + 46, + 283, + 287, + 427 + ], + "type": "text", + "content": "To mitigate the collapse of " + }, + { + "bbox": [ + 46, + 283, + 287, + 427 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 46, + 283, + 287, + 427 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 46, + 283, + 287, + 427 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 46, + 283, + 287, + 427 + ], + "type": "text", + "content": " space, we need to ensure that the change in conditioning variable " + }, + { + "bbox": [ + 46, + 283, + 287, + 427 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 46, + 283, + 287, + 427 + ], + "type": "text", + "content": " leads to the corresponding change in " + }, + { + "bbox": [ + 46, + 283, + 287, + 427 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 46, + 283, + 287, + 427 + ], + "type": "text", + "content": ". Recently in self-supervised learning, several techniques [2, 53] have been introduced to prevent the collapse of learned representations by maximizing the information content in the feature dimensions. Inspired by them we propose NoisyTwins, in which we first generate inexpensive twin augmentations for class embeddings and then use them to decorrelate the " + }, + { + "bbox": [ + 46, + 283, + 287, + 427 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 46, + 283, + 287, + 427 + ], + "type": "text", + "content": " variables through self-supervision. The decorrelation ensures that " + }, + { + "bbox": [ + 46, + 283, + 287, + 427 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 46, + 283, + 287, + 427 + ], + "type": "text", + "content": " vectors are diverse for each class and the GAN is able to produce intra-class diversity among the generated images." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 430, + 288, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 430, + 288, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 430, + 288, + 693 + ], + "type": "text", + "content": "We evaluate NoisyTwins on challenging benchmarks of large-scale long-tailed datasets of ImageNet-LT [30] and iNaturalist 2019 [48]. These benchmarks are particularly challenging due to a large number of classes present, which makes GANs prone to class confusion. On the other hand, as these datasets are long-tailed with only a few images per class in tail classes, generating diverse images for those classes is challenging. We observe that existing metrics used in GAN evaluations are not able to capture both class confusion and mode collapse. As a remedy, we propose to use intra-class Frechet Inception Distance (FID) [12] based on features obtained from pre-trained CLIP [34] embeddings as an effective metric to measure the performance of class-conditional GANs in long-tailed data setups. Using NoisyTwins enables StyleGAN to generate diverse and class-consistent images across classes, mitigating the mode collapse and class confusion issues in existing state-of-the-art (SotA) (Fig. 1). Further, with NoisyTwins, we obtain diverse generations for tail classes even with " + }, + { + "bbox": [ + 46, + 430, + 288, + 693 + ], + "type": "inline_equation", + "content": "\\leq 30" + }, + { + "bbox": [ + 46, + 430, + 288, + 693 + ], + "type": "text", + "content": " images, which can be attributed to the transfer of knowledge from head classes through shared parameters (Fig. 1 and 6). In summary, we make the following contributions:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 701, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 288, + 714 + ], + "type": "text", + "content": "1. We evaluate various recent SotA GAN conditioning" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 324, + 208, + 545, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 208, + 545, + 255 + ], + "spans": [ + { + "bbox": [ + 324, + 208, + 545, + 255 + ], + "type": "text", + "content": "and regularization techniques on the challenging task of long-tailed image generation. We find that all existing methods either suffer from mode collapse or lead to class confusion in generations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 256, + 545, + 375 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 313, + 256, + 545, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 256, + 545, + 303 + ], + "spans": [ + { + "bbox": [ + 313, + 256, + 545, + 303 + ], + "type": "text", + "content": "2. To mitigate mode collapse and class confusion, we introduce NoisyTwins, an effective and inexpensive augmentation strategy for class embeddings that decorrelates latents in the " + }, + { + "bbox": [ + 313, + 256, + 545, + 303 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 313, + 256, + 545, + 303 + ], + "type": "text", + "content": " latent space (Sec. 4)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 304, + 545, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 304, + 545, + 375 + ], + "spans": [ + { + "bbox": [ + 313, + 304, + 545, + 375 + ], + "type": "text", + "content": "3. We evaluate NoisyTwins on large-scale long-tailed datasets of ImageNet-LT and iNaturalist-2019, where it consistently improves the StyleGAN2 performance " + }, + { + "bbox": [ + 313, + 304, + 545, + 375 + ], + "type": "inline_equation", + "content": "(\\sim 19\\%)" + }, + { + "bbox": [ + 313, + 304, + 545, + 375 + ], + "type": "text", + "content": ", achieving a new SotA. Further, our approach can also prevent mode collapse and enhance the performance of few-shot GANs (Sec. 5.3)." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 388, + 398, + 400 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 388, + 398, + 400 + ], + "spans": [ + { + "bbox": [ + 306, + 388, + 398, + 400 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 411, + 545, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 411, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 304, + 411, + 545, + 590 + ], + "type": "text", + "content": "StyleGANs. Karras et al. introduced StyleGAN [21] and subsequently improved its image quality in StyleGAN2. StyleGAN could produce high-resolution photorealistic images as demonstrated on a wide variety of category-specific datasets. It introduced a mapping network, which mapped the sampled noise into another latent space, which is more disentangled and semantically coherent, as demonstrated by its downstream usage for image editing and manipulation [1,32,33,42,43]. Further, StyleGAN has been extended to get novel views from images [27, 28, 44], thus making it possible to get 3D information from it. These downstream advances are possible due to the impressive performance of StyleGANs on class-specific datasets (such as faces). However, similar photorealism levels are yet uncommon on multi-class long-tailed datasets (such as ImageNet)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "type": "text", + "content": "GANs for Data Efficiency and Imbalance. Failure of GANs on less data was concurrently reported by Karras et al. [20] and Zhao et al. [54]. The problem is rooted in the overfitting of the discriminator due to less real data. Since then, the proposed solutions for this problem have relied on a) augmenting the data, b) introducing regularizers, and c) architectural modifications. Karras et al. [20] and Zhao et al. [54] relied on differentiable data augmentation before passing images into the discriminator to solve this problem. DeceiveD [15] proposed to introduce label-noise for" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5988" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "content": "discriminator training. LeCamGAN [47] finds that enforcing LeCam divergence as a regularization trick in the discriminator can robustify GAN training under a limited data setting. DynamicD [50] tunes the capacity of the discriminator on-the-fly during training. While these methods can handle the data inefficiency, they are ineffective on class imbalanced long-tailed data distribution [35]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 156, + 288, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 156, + 288, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 156, + 288, + 312 + ], + "type": "text", + "content": "CBGAN [36] proposed a solution to train the unconditional GAN model on long-tailed data distribution by introducing a signal from the classifier to balance the classes generated by GAN. In a long-tailed class-conditional setting, gSR [35] proposes to regularize the exploding spectral norms of the class-specific parameters of the GAN. Collapse-by-conditioning [40] addresses the limited data in classes by introducing a training regime that transitions from an unconditional to a class-conditioned setting, thus exploiting the shared information across classes during the early stages of the training. However, these methods suffer from either class confusion or poor generated image quality on large datasets, which is resolved by NoisyTwins." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 316, + 287, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 316, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 287, + 483 + ], + "type": "text", + "content": "Self-Supervised Learning for GANs. Ideas from Self-supervised learning have shown their benefits in GAN training. IC-GAN [6] trains GAN conditioned on embeddings on SwAV [5], which led to remarkable improvement in performance on the long-tailed version of ImageNet. InsGen [51] and ReACGAN [16, 17] introduce the auxiliary task of instance discrimination for the discriminator, thereby making the discriminator focus on multiple tasks and thus alleviating discriminator overfitting. While InsGen relies on both noise space and image space augmentations, ReACGAN and ContraGAN follow only image space augmentations. Contrary to these, NoisyTwins performs augmentations in the class-embedding space and contrasts them in the " + }, + { + "bbox": [ + 46, + 316, + 287, + 483 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 46, + 316, + 287, + 483 + ], + "type": "text", + "content": "-space of the generator instead of the discriminator." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 495, + 132, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 495, + 132, + 506 + ], + "spans": [ + { + "bbox": [ + 47, + 495, + 132, + 506 + ], + "type": "text", + "content": "3. Preliminaries" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 514, + 117, + 527 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 514, + 117, + 527 + ], + "spans": [ + { + "bbox": [ + 47, + 514, + 117, + 527 + ], + "type": "text", + "content": "3.1. StyleGAN" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 533, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 288, + 713 + ], + "type": "text", + "content": "StyleGAN [21] is a Generative Adversarial Network comprising of its unique Style Conditioning Based Generator " + }, + { + "bbox": [ + 46, + 533, + 288, + 713 + ], + "type": "inline_equation", + "content": "(\\mathcal{G})" + }, + { + "bbox": [ + 46, + 533, + 288, + 713 + ], + "type": "text", + "content": " and discriminator network " + }, + { + "bbox": [ + 46, + 533, + 288, + 713 + ], + "type": "inline_equation", + "content": "(\\mathcal{D})" + }, + { + "bbox": [ + 46, + 533, + 288, + 713 + ], + "type": "text", + "content": " trained jointly. We will focus on the architecture of StyleGAN2 [23] as we use it in our experiments, although our work is generally applicable to all StyleGAN architectures. The StyleGAN2 generator is composed of blocks that progressively upsample the features and resolution inspired by Progressive GAN [19], starting from a single root image. The diversity in the images comes from conditioning each block of image generation through conditioning on the latent coming from the mapping network (Fig. 2). The mapping network is a fully connected network that takes in the conditioning variables, the " + }, + { + "bbox": [ + 46, + 533, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{z} \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 46, + 533, + 288, + 713 + ], + "type": "text", + "content": " coming from a random distribution (e.g., Gaussian, etc.) and class conditioning label " + }, + { + "bbox": [ + 46, + 533, + 288, + 713 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 533, + 288, + 713 + ], + "type": "text", + "content": " which is converted" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "text", + "content": "to an embedding " + }, + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "inline_equation", + "content": "\\mathbf{c} \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "text", + "content": ". The mapping network takes these and outputs vectors " + }, + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "text", + "content": " in the " + }, + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "text", + "content": " latent space of StyleGAN, which is found to be semantically disentangled to a high extent [49]. The " + }, + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "text", + "content": " is then processed through an affine transform and passed to each generator layer for conditioning the image generation process through Adaptive Instance Normalization (AdaIN) [13]. The images from generator " + }, + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "text", + "content": ", along with real images, are passed to discriminator " + }, + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "text", + "content": " for training. The training utilizes the non-saturating adversarial losses [9] for " + }, + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 304, + 72, + 545, + 192 + ], + "type": "text", + "content": " given as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 193, + 545, + 224 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 193, + 545, + 224 + ], + "spans": [ + { + "bbox": [ + 313, + 193, + 545, + 224 + ], + "type": "interline_equation", + "content": "\\min _ {\\mathcal {D}} \\mathcal {L} _ {\\mathcal {D}} = \\sum_ {i = 1} ^ {m} \\log (\\mathcal {D} (\\mathbf {x} _ {i})) + \\log (1 - \\mathcal {D} (\\mathcal {G} (\\mathbf {z} _ {i}, \\mathbf {c} _ {i}))) \\tag {1}", + "image_path": "18e9ce02c56808157e5485b3c73a1faf7ec1092346fe8ab124ffcb8dde9ae7a5.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 315, + 226, + 545, + 257 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 226, + 545, + 257 + ], + "spans": [ + { + "bbox": [ + 315, + 226, + 545, + 257 + ], + "type": "interline_equation", + "content": "\\min _ {\\mathcal {G}} \\mathcal {L} _ {\\mathcal {G}} = \\sum_ {i = 1} ^ {m} - \\log \\left(\\mathcal {D} \\left(\\mathcal {G} \\left(\\mathbf {z} _ {i}, \\mathbf {c} _ {i}\\right)\\right)\\right) \\tag {2}", + "image_path": "8a47ade989368a5b1ddc3ae5d830eb77953af521487b333bda2dc8e84356700d.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 258, + 545, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 258, + 545, + 282 + ], + "spans": [ + { + "bbox": [ + 304, + 258, + 545, + 282 + ], + "type": "text", + "content": "We now describe the issues present in the StyleGANs trained on long-tailed data and their analysis in " + }, + { + "bbox": [ + 304, + 258, + 545, + 282 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 304, + 258, + 545, + 282 + ], + "type": "text", + "content": " space." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 288, + 545, + 313 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 288, + 545, + 313 + ], + "spans": [ + { + "bbox": [ + 305, + 288, + 545, + 313 + ], + "type": "text", + "content": "3.2. Class Confusion and Class-Specific Mode Collapse in Conditional StyleGANs" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 318, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 318, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 318, + 545, + 713 + ], + "type": "text", + "content": "To finely analyze the performance of StyleGAN and its variants on long-tailed datasets, we train them on the CIFAR10-LT dataset. In Fig. 3, we plot the qualitative results of generated images and create a t-SNE plot for latents in " + }, + { + "bbox": [ + 304, + 318, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 304, + 318, + 545, + 713 + ], + "type": "text", + "content": " space for each class. We first train the StyleGAN2 baseline with augmentations (DiffAug) [20, 54]. We find that it leads to mode collapse, specifically for tail classes (Fig. 3). In conjunction with images, we also observe that corresponding t-SNE embeddings are also collapsed near each class's mean in " + }, + { + "bbox": [ + 304, + 318, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 304, + 318, + 545, + 713 + ], + "type": "text", + "content": " space. Further, recent methods which have proposed the usage of contrastive learning for GANs, improve their data efficiency and prevent discriminator overfitting [14, 16]. We also evaluate them by adding the contrastive conditioning method, which is D2DCE loss-based on ReACGAN [17], to the baseline, where in results, we observe that the network omits to learn tail classes and produces head class images at their place (i.e., class confusion). In Fig. 3, it can be seen that the network confuses semantically similar classes, that is, generating cars (head or majority class) in place of trucks and airplanes (head class) instead of ships. In the " + }, + { + "bbox": [ + 304, + 318, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 304, + 318, + 545, + 713 + ], + "type": "text", + "content": " space, we find the same number of clusters as the number of classes in the dataset. However, the tail label cluster images also belong to the head classes of cars and airplanes. In a very recent work gSR [35], it has been shown that constraining the spectral norm of " + }, + { + "bbox": [ + 304, + 318, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 304, + 318, + 545, + 713 + ], + "type": "text", + "content": " embedding parameters can help reduce the mode collapse and lead to stable training. However, we find that constraining the embeddings leads to class confusion, as seen in t-SNE visualization in Fig. 3. We find that this class confusion gets further aggravated when StyleGAN is trained on datasets like ImageNet-LT, which contain a large number of classes, along with a bunch of semantically similar classes (Sec. 5.2). Based on our " + }, + { + "bbox": [ + 304, + 318, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5989" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 70, + 304, + 295 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 304, + 295 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 304, + 295 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 304, + 295 + ], + "type": "image", + "image_path": "a811c8c06599313d1d44d9fb7fc8e236c148611c31d29e93bced5155cb9ad7e8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 308, + 547, + 343 + ], + "lines": [ + { + "bbox": [ + 46, + 308, + 547, + 343 + ], + "spans": [ + { + "bbox": [ + 46, + 308, + 547, + 343 + ], + "type": "text", + "content": "Figure 3. Comparison of GANs and their " + }, + { + "bbox": [ + 46, + 308, + 547, + 343 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 46, + 308, + 547, + 343 + ], + "type": "text", + "content": " space for CIFAR10-LT. We plot the generated images on (left) and generate a t-SNE plot of w latents for generated images in " + }, + { + "bbox": [ + 46, + 308, + 547, + 343 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 46, + 308, + 547, + 343 + ], + "type": "text", + "content": " space (right). We find that mode collapse and class confusion in images is linked to the corresponding collapse and confusion in latent " + }, + { + "bbox": [ + 46, + 308, + 547, + 343 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 46, + 308, + 547, + 343 + ], + "type": "text", + "content": " space. Our proposed NoisyTwins mitigates both collapse (left) and confusion (right) simultaneously." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 323, + 77, + 433, + 172 + ], + "blocks": [ + { + "bbox": [ + 323, + 77, + 433, + 172 + ], + "lines": [ + { + "bbox": [ + 323, + 77, + 433, + 172 + ], + "spans": [ + { + "bbox": [ + 323, + 77, + 433, + 172 + ], + "type": "image", + "image_path": "ba2441057d6fd53a9b97ffebb6a7efe48f94bd72b0de4351575b78a6fdf133cc.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 437, + 78, + 544, + 171 + ], + "blocks": [ + { + "bbox": [ + 437, + 78, + 544, + 171 + ], + "lines": [ + { + "bbox": [ + 437, + 78, + 544, + 171 + ], + "spans": [ + { + "bbox": [ + 437, + 78, + 544, + 171 + ], + "type": "image", + "image_path": "13e9cd0abac8a49b6bc45c0656a00091e549c181cef6566fcfbe5916f0788de5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 323, + 175, + 433, + 270 + ], + "blocks": [ + { + "bbox": [ + 323, + 175, + 433, + 270 + ], + "lines": [ + { + "bbox": [ + 323, + 175, + 433, + 270 + ], + "spans": [ + { + "bbox": [ + 323, + 175, + 433, + 270 + ], + "type": "image", + "image_path": "4563af6695895049295817ed017d1b32f0eb1173c31bd2a3f989fd79a80fbe3b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 437, + 176, + 544, + 270 + ], + "blocks": [ + { + "bbox": [ + 437, + 176, + 544, + 270 + ], + "lines": [ + { + "bbox": [ + 437, + 176, + 544, + 270 + ], + "spans": [ + { + "bbox": [ + 437, + 176, + 544, + 270 + ], + "type": "image", + "image_path": "e5e02df03fadf4c2db0c35daccfdd79001539044451f35868abc86705e3f9c49.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 337, + 273, + 526, + 297 + ], + "blocks": [ + { + "bbox": [ + 337, + 273, + 526, + 297 + ], + "lines": [ + { + "bbox": [ + 337, + 273, + 526, + 297 + ], + "spans": [ + { + "bbox": [ + 337, + 273, + 526, + 297 + ], + "type": "image", + "image_path": "bc0db085d00bb489b207b957f0775ca2d63ec2d3a70a51d3f01d3982e7f0e3cf.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 348, + 290, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 348, + 290, + 420 + ], + "spans": [ + { + "bbox": [ + 46, + 348, + 290, + 420 + ], + "type": "text", + "content": "space analysis and qualitative results above, we observe that the class confusion and mode collapse of images is tightly coupled with the structure of " + }, + { + "bbox": [ + 46, + 348, + 290, + 420 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 46, + 348, + 290, + 420 + ], + "type": "text", + "content": " space. Further, the recent SotA methods are either unable to prevent collapse or suffer from class confusion. Hence, this work aims to develop a technique that mitigates both confusion and collapse." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 430, + 115, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 430, + 115, + 443 + ], + "spans": [ + { + "bbox": [ + 47, + 430, + 115, + 443 + ], + "type": "text", + "content": "4. Approach" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 449, + 288, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 288, + 521 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 288, + 521 + ], + "type": "text", + "content": "In this section, we present our method NoisyTwins, which introduces noise-based augmentation twins in the conditional embedding space (Sec. 4.1), and then combines it with the Barlow-Twins-based regularizer from the self-supervised learning (SSL) paradigm to resolve the issue of class confusion and mode collapse (Sec. 4.2)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 527, + 264, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 527, + 264, + 541 + ], + "spans": [ + { + "bbox": [ + 47, + 527, + 264, + 541 + ], + "type": "text", + "content": "4.1. Noise Augmentation in Embedding Space" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 546, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 546, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 546, + 289, + 714 + ], + "type": "text", + "content": "As we observed in the previous section that " + }, + { + "bbox": [ + 46, + 546, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 46, + 546, + 289, + 714 + ], + "type": "text", + "content": " vectors for each sample become insensitive to changes in " + }, + { + "bbox": [ + 46, + 546, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 46, + 546, + 289, + 714 + ], + "type": "text", + "content": ". This collapse in " + }, + { + "bbox": [ + 46, + 546, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 46, + 546, + 289, + 714 + ], + "type": "text", + "content": " vectors for each class leads to mode collapse for baselines (Fig. 3). One reason for this could be the fact that " + }, + { + "bbox": [ + 46, + 546, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 46, + 546, + 289, + 714 + ], + "type": "text", + "content": " is composed of continuous variables, whereas the embedding vectors " + }, + { + "bbox": [ + 46, + 546, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 46, + 546, + 289, + 714 + ], + "type": "text", + "content": " for each class are discrete. Due to this, the GAN converges to the easy degenerate solution where it generates a single sample for each class, becoming insensitive to changes in " + }, + { + "bbox": [ + 46, + 546, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 46, + 546, + 289, + 714 + ], + "type": "text", + "content": ". For inducing some continuity in " + }, + { + "bbox": [ + 46, + 546, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 46, + 546, + 289, + 714 + ], + "type": "text", + "content": " embeddings vectors, we introduce an augmentation strategy where we add i.i.d. noise of small magnitude in each of the variables in " + }, + { + "bbox": [ + 46, + 546, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 46, + 546, + 289, + 714 + ], + "type": "text", + "content": ". Based on our observation (Fig. 3) and existing works [35], there is a high tendency for mode collapse in tail classes. Hence we add noise in embedding space that is" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 348, + 546, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 348, + 546, + 381 + ], + "spans": [ + { + "bbox": [ + 304, + 348, + 546, + 381 + ], + "type": "text", + "content": "proportional to the inverse of the frequency of samples. We provide the mathematical expression of noise augmentation " + }, + { + "bbox": [ + 304, + 348, + 546, + 381 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{c}}" + }, + { + "bbox": [ + 304, + 348, + 546, + 381 + ], + "type": "text", + "content": " below:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 335, + 380, + 545, + 403 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 380, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 335, + 380, + 545, + 403 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathbf {c}} \\sim \\mathbf {c} + \\mathcal {N} (\\mathbf {0}, \\sigma_ {c} \\mathbb {I} _ {d}) \\text {w h e r e} \\sigma_ {c} = \\sigma \\frac {(1 - \\alpha)}{1 - \\alpha^ {n _ {c}}} \\tag {3}", + "image_path": "88ade6c2a5e08ba0f008a72cbf3bac8bdb485085e39376685ddbe3bb09291de1.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 403, + 546, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 403, + 546, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 403, + 546, + 521 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 304, + 403, + 546, + 521 + ], + "type": "inline_equation", + "content": "n_c" + }, + { + "bbox": [ + 304, + 403, + 546, + 521 + ], + "type": "text", + "content": " is the frequency of training samples in class " + }, + { + "bbox": [ + 304, + 403, + 546, + 521 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 403, + 546, + 521 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 403, + 546, + 521 + ], + "type": "inline_equation", + "content": "\\mathbb{I}" + }, + { + "bbox": [ + 304, + 403, + 546, + 521 + ], + "type": "text", + "content": " is the identity matrix of size " + }, + { + "bbox": [ + 304, + 403, + 546, + 521 + ], + "type": "inline_equation", + "content": "d \\times d" + }, + { + "bbox": [ + 304, + 403, + 546, + 521 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 403, + 546, + 521 + ], + "type": "inline_equation", + "content": "\\alpha, \\sigma" + }, + { + "bbox": [ + 304, + 403, + 546, + 521 + ], + "type": "text", + "content": " are hyperparameters. The expression of " + }, + { + "bbox": [ + 304, + 403, + 546, + 521 + ], + "type": "inline_equation", + "content": "\\sigma_c" + }, + { + "bbox": [ + 304, + 403, + 546, + 521 + ], + "type": "text", + "content": " is from the effective number of samples [8], which is a softer version of inverse frequency proportionality. In contrast to the image space augmentation, these noise augmentations come for free as there is no significant additional computation overhead. This noise is added in the embedding " + }, + { + "bbox": [ + 304, + 403, + 546, + 521 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 304, + 403, + 546, + 521 + ], + "type": "text", + "content": " before passing it to the generator and the discriminator, which ensures that the class embeddings occupy a continuous region in latent space." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 522, + 545, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 545, + 570 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 545, + 570 + ], + "type": "text", + "content": "Insight: The augmentation equation above (Eq. 3) can be interpreted as approximating the discrete random variable " + }, + { + "bbox": [ + 304, + 522, + 545, + 570 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 304, + 522, + 545, + 570 + ], + "type": "text", + "content": " with a Gaussian with finite variance and the embedding parameters " + }, + { + "bbox": [ + 304, + 522, + 545, + 570 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 304, + 522, + 545, + 570 + ], + "type": "text", + "content": " being the mean " + }, + { + "bbox": [ + 304, + 522, + 545, + 570 + ], + "type": "inline_equation", + "content": "\\mu_{\\mathbf{c}}" + }, + { + "bbox": [ + 304, + 522, + 545, + 570 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 389, + 570, + 545, + 582 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 570, + 545, + 582 + ], + "spans": [ + { + "bbox": [ + 389, + 570, + 545, + 582 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathbf {c}} \\sim \\mathcal {N} \\left(\\mu_ {\\mathbf {c}}, \\sigma_ {c} \\mathbb {I} _ {d}\\right) \\tag {4}", + "image_path": "7d6122e89322ad12a5eaf6094af697d72ab58ec14a4bf560d6ede84fd0a11e24.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 582, + 547, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 665 + ], + "type": "text", + "content": "This leads to the class-embedding input " + }, + { + "bbox": [ + 304, + 582, + 547, + 665 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{c}}" + }, + { + "bbox": [ + 304, + 582, + 547, + 665 + ], + "type": "text", + "content": " to mapping network to have a Gaussian distribution, similar in nature to " + }, + { + "bbox": [ + 304, + 582, + 547, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 304, + 582, + 547, + 665 + ], + "type": "text", + "content": ". This noise augmentation strategy alone mitigates the degenerate solution of class-wise mode collapse to a great extent (Table 1) and helps generate diverse latent " + }, + { + "bbox": [ + 304, + 582, + 547, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 304, + 582, + 547, + 665 + ], + "type": "text", + "content": " for each class. Due to the diverse " + }, + { + "bbox": [ + 304, + 582, + 547, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 304, + 582, + 547, + 665 + ], + "type": "text", + "content": " conditioning of the GAN, it leads to diverse image generation." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 671, + 515, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 671, + 515, + 685 + ], + "spans": [ + { + "bbox": [ + 306, + 671, + 515, + 685 + ], + "type": "text", + "content": "4.2. Invariance in " + }, + { + "bbox": [ + 306, + 671, + 515, + 685 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 306, + 671, + 515, + 685 + ], + "type": "text", + "content": "-Space with NoisyTwins" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 689, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 546, + 714 + ], + "type": "text", + "content": "The augmentation strategy introduced in the previous section expands the region for each class in " + }, + { + "bbox": [ + 306, + 689, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 306, + 689, + 546, + 714 + ], + "type": "text", + "content": " latent space." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5990" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 70, + 547, + 203 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 547, + 203 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 547, + 203 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 547, + 203 + ], + "type": "image", + "image_path": "efb94e7c8a3ba242f1b8e2fea2e43e6860f58772a173022a1344d604c67a202e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 211, + 547, + 257 + ], + "lines": [ + { + "bbox": [ + 46, + 211, + 547, + 257 + ], + "spans": [ + { + "bbox": [ + 46, + 211, + 547, + 257 + ], + "type": "text", + "content": "Figure 4. Overview of NoisyTwins. For the " + }, + { + "bbox": [ + 46, + 211, + 547, + 257 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 46, + 211, + 547, + 257 + ], + "type": "text", + "content": " sample of class " + }, + { + "bbox": [ + 46, + 211, + 547, + 257 + ], + "type": "inline_equation", + "content": "c^i" + }, + { + "bbox": [ + 46, + 211, + 547, + 257 + ], + "type": "text", + "content": ", we create twin augmentations " + }, + { + "bbox": [ + 46, + 211, + 547, + 257 + ], + "type": "inline_equation", + "content": "(\\tilde{\\mathbf{c}}_a^i,\\tilde{\\mathbf{c}}_b^i)" + }, + { + "bbox": [ + 46, + 211, + 547, + 257 + ], + "type": "text", + "content": ", by sampling from a Gaussian centered at class embedding " + }, + { + "bbox": [ + 46, + 211, + 547, + 257 + ], + "type": "inline_equation", + "content": "(\\mu_{\\mathbf{c}^i})" + }, + { + "bbox": [ + 46, + 211, + 547, + 257 + ], + "type": "text", + "content": ". After this, we concatenate them with the same " + }, + { + "bbox": [ + 46, + 211, + 547, + 257 + ], + "type": "inline_equation", + "content": "\\mathbf{z}^i" + }, + { + "bbox": [ + 46, + 211, + 547, + 257 + ], + "type": "text", + "content": " and obtain " + }, + { + "bbox": [ + 46, + 211, + 547, + 257 + ], + "type": "inline_equation", + "content": "(\\tilde{\\mathbf{w}}_a^i,\\tilde{\\mathbf{w}}_b^i)" + }, + { + "bbox": [ + 46, + 211, + 547, + 257 + ], + "type": "text", + "content": " from the mapping network, which we stack in batches of augmented latents (" + }, + { + "bbox": [ + 46, + 211, + 547, + 257 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{W}}_{\\mathbf{A}}" + }, + { + "bbox": [ + 46, + 211, + 547, + 257 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 211, + 547, + 257 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{W}}_{\\mathbf{B}}" + }, + { + "bbox": [ + 46, + 211, + 547, + 257 + ], + "type": "text", + "content": "). The twin " + }, + { + "bbox": [ + 46, + 211, + 547, + 257 + ], + "type": "inline_equation", + "content": "(\\tilde{\\mathbf{w}}_a^i,\\tilde{\\mathbf{w}}_b^i)" + }, + { + "bbox": [ + 46, + 211, + 547, + 257 + ], + "type": "text", + "content": " vectors are then made invariant to augmentations (similar) in the latent space by minimizing cross-correlation [2,53] between the latents of two augmented batches (" + }, + { + "bbox": [ + 46, + 211, + 547, + 257 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{W}}_{\\mathbf{A}}" + }, + { + "bbox": [ + 46, + 211, + 547, + 257 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 211, + 547, + 257 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{W}}_{\\mathbf{B}}" + }, + { + "bbox": [ + 46, + 211, + 547, + 257 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 262, + 288, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 262, + 288, + 322 + ], + "spans": [ + { + "bbox": [ + 46, + 262, + 288, + 322 + ], + "type": "text", + "content": "Although that does lead to diverse image generation as " + }, + { + "bbox": [ + 46, + 262, + 288, + 322 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 46, + 262, + 288, + 322 + ], + "type": "text", + "content": " are diverse; however, this does not ensure that these " + }, + { + "bbox": [ + 46, + 262, + 288, + 322 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 46, + 262, + 288, + 322 + ], + "type": "text", + "content": " will generate class-consistent outputs for augmentations in embedding " + }, + { + "bbox": [ + 46, + 262, + 288, + 322 + ], + "type": "inline_equation", + "content": "(\\tilde{\\mathbf{c}})" + }, + { + "bbox": [ + 46, + 262, + 288, + 322 + ], + "type": "text", + "content": ". To ensure class consistent predictions, we need to ensure invariance in " + }, + { + "bbox": [ + 46, + 262, + 288, + 322 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 46, + 262, + 288, + 322 + ], + "type": "text", + "content": " to noise augmentation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 324, + 290, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 324, + 290, + 469 + ], + "spans": [ + { + "bbox": [ + 46, + 324, + 290, + 469 + ], + "type": "text", + "content": "For enforcing invariance to augmentations, a set of recent works [2, 10, 53] in self-supervised learning make the representations of augmentations similar through regularization. Among them, we focus on Barlow Twins as it does not require a large batch size of samples. Inspired by Barlow twins, we introduce NoisyTwins (Fig. 4), where we generate twin augmentations " + }, + { + "bbox": [ + 46, + 324, + 290, + 469 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{c}}_a" + }, + { + "bbox": [ + 46, + 324, + 290, + 469 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 324, + 290, + 469 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{c}}_b" + }, + { + "bbox": [ + 46, + 324, + 290, + 469 + ], + "type": "text", + "content": " of the same class embedding " + }, + { + "bbox": [ + 46, + 324, + 290, + 469 + ], + "type": "inline_equation", + "content": "(\\mu_c)" + }, + { + "bbox": [ + 46, + 324, + 290, + 469 + ], + "type": "text", + "content": " and concatenate them to same " + }, + { + "bbox": [ + 46, + 324, + 290, + 469 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 46, + 324, + 290, + 469 + ], + "type": "text", + "content": ". After creating a batch of such inputs, they are passed to the mapping network to get batches of augmented latents " + }, + { + "bbox": [ + 46, + 324, + 290, + 469 + ], + "type": "inline_equation", + "content": "(\\tilde{\\mathbf{W}}_A" + }, + { + "bbox": [ + 46, + 324, + 290, + 469 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 324, + 290, + 469 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{W}}_B)" + }, + { + "bbox": [ + 46, + 324, + 290, + 469 + ], + "type": "text", + "content": ". These batches are then used to calculate the cross-correlation matrix of latent variables given as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 90, + 474, + 287, + 524 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 474, + 287, + 524 + ], + "spans": [ + { + "bbox": [ + 90, + 474, + 287, + 524 + ], + "type": "interline_equation", + "content": "\\mathbf {C} _ {j, k} = \\frac {\\sum_ {(\\tilde {\\mathbf {w}} _ {a} , \\tilde {\\mathbf {w}} _ {b}) \\in (\\tilde {\\mathbf {W}} _ {A} , \\tilde {\\mathbf {W}} _ {B})} \\tilde {\\mathbf {w}} _ {a} ^ {j} \\tilde {\\mathbf {w}} _ {b} ^ {k}}{\\sum_ {\\tilde {\\mathbf {w}} _ {a} \\in \\tilde {\\mathbf {W}} _ {A}} \\tilde {\\mathbf {w}} _ {a} ^ {j} \\tilde {\\mathbf {w}} _ {a} ^ {j} \\sum_ {\\tilde {\\mathbf {w}} _ {b} \\in \\tilde {\\mathbf {W}} _ {B}} \\tilde {\\mathbf {w}} _ {b} ^ {k} \\tilde {\\mathbf {w}} _ {b} ^ {k}} \\tag {5}", + "image_path": "057afdeeb81b9e09b5baacd1e298e76682a50e8e8f139e03c1a628b131ec1e36.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 529, + 288, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 529, + 288, + 554 + ], + "spans": [ + { + "bbox": [ + 47, + 529, + 288, + 554 + ], + "type": "text", + "content": "The matrix " + }, + { + "bbox": [ + 47, + 529, + 288, + 554 + ], + "type": "inline_equation", + "content": "\\mathbf{C}" + }, + { + "bbox": [ + 47, + 529, + 288, + 554 + ], + "type": "text", + "content": " is a square matrix of size same as of latents " + }, + { + "bbox": [ + 47, + 529, + 288, + 554 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 47, + 529, + 288, + 554 + ], + "type": "text", + "content": ". The final loss based on confusion matrix is given as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 93, + 560, + 287, + 587 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 560, + 287, + 587 + ], + "spans": [ + { + "bbox": [ + 93, + 560, + 287, + 587 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathcal {N T}} = \\sum_ {j} \\left(1 - \\mathbf {C} _ {j j} ^ {2}\\right) + \\gamma \\sum_ {j \\neq k} \\mathbf {C} _ {j, k} ^ {2} \\tag {6}", + "image_path": "e3ba68331b01dfa8fc59b104e88baa97fdc762aed8d1602e6875cfab5b4c99f3.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "type": "text", + "content": "The first term tries to make the two latents " + }, + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "type": "inline_equation", + "content": "(\\tilde{\\mathbf{w}}_a" + }, + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{w}}_b)" + }, + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "type": "text", + "content": " invariant to the noise augmentation applied (i.e. similar), whereas the second term tries to de-correlate the different variables, thus maximizing the information content of the embedding vector [53]. The " + }, + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "type": "text", + "content": " is the hyper-parameter that determines the relative importance of the two terms. This loss is then added to the generator loss term " + }, + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\mathcal{G}} + \\lambda L_{\\mathcal{NT}})" + }, + { + "bbox": [ + 46, + 594, + 289, + 714 + ], + "type": "text", + "content": " and optimized through backpropagation. The above procedure comprises our proposed method, NoisyTwins (Fig. 4), which we empirically evaluate in the subsequent sections." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 261, + 451, + 275 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 261, + 451, + 275 + ], + "spans": [ + { + "bbox": [ + 306, + 261, + 451, + 275 + ], + "type": "text", + "content": "5. Experimental Evaluation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 283, + 356, + 297 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 283, + 356, + 297 + ], + "spans": [ + { + "bbox": [ + 306, + 283, + 356, + 297 + ], + "type": "text", + "content": "5.1. Setup" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 308, + 547, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 308, + 547, + 511 + ], + "spans": [ + { + "bbox": [ + 304, + 308, + 547, + 511 + ], + "type": "text", + "content": "Datasets: We primarily apply all methods on long-tailed datasets, as GANs trained on them are more prone to class confusion and mode collapse. We first report on the commonly used CIFAR10-LT dataset with an imbalance factor (i.e. ratio of most to least frequent class) of 100. To show our approach's scalability and real-world application, we test our method on the challenging ImageNet-LT and iNaturalist 2019 datasets. The ImageNet-LT [30] is a long-tailed variant of the 1000 class ImageNet dataset, with a plethora of semantically similar classes (e.g., Dogs, Birds etc.), making it challenging to avoid class confusion. iNaturalist-2019 [48] is a real-world long-tailed dataset composed of 1010 different variants of species, some of which have fine-grained differences in their appearance. For such fine-grained datasets, ImageNet pre-trained discriminators [38, 39] may not be useful, as augmentations used to train the model makes it invariant to fine-grained changes." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 516, + 547, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 516, + 547, + 661 + ], + "spans": [ + { + "bbox": [ + 304, + 516, + 547, + 661 + ], + "type": "text", + "content": "Training Configuration: We use StyleGAN2 architecture for all our experiments. All our experiments are performed using PyTorch-StudioGAN implemented by Kang et al. [18], which serves as a base for our framework. We use Path Length regularization (PLR) as it ensures changes in " + }, + { + "bbox": [ + 304, + 516, + 547, + 661 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 304, + 516, + 547, + 661 + ], + "type": "text", + "content": " space lead to changes in images, using a delayed PLR on ImageNet-LT following [39]. We use a batch size of 128 for all our experiments, with one G step per D step. Unless stated explicitly, we used the general training setup for StyleGANs from [18], including methods like " + }, + { + "bbox": [ + 304, + 516, + 547, + 661 + ], + "type": "inline_equation", + "content": "R_{1}" + }, + { + "bbox": [ + 304, + 516, + 547, + 661 + ], + "type": "text", + "content": " regularization [31], etc. More details on the exact training configuration for each dataset are provided in Appendix." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 665, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 689 + ], + "type": "text", + "content": "Metrics: In this work we use the following metrics for evaluation of our methods:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 689, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 547, + 713 + ], + "type": "text", + "content": "a) FID: Fréchet Inception Distance [12] is the Wasserstein-2 Distance between the real and validation data. We use a" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 758 + ], + "type": "text", + "content": "5991" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 286, + 224 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 286, + 224 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 286, + 224 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 286, + 224 + ], + "type": "image", + "image_path": "ff8ec5d9635145f686f068454e085ede3c0e1758b811ea2946bf87848554e0be.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 236, + 287, + 291 + ], + "lines": [ + { + "bbox": [ + 46, + 236, + 287, + 291 + ], + "spans": [ + { + "bbox": [ + 46, + 236, + 287, + 291 + ], + "type": "text", + "content": "Figure 5. Choice of Eval. backbone: intra-FID (iFID) of a class based on InceptionV3 backbone (left plot) is not able to capture the mode collapse (increase in iFID in the absence of mode collapse). This is well-captured by " + }, + { + "bbox": [ + 46, + 236, + 287, + 291 + ], + "type": "inline_equation", + "content": "\\mathrm{iFID}_{\\mathrm{CLIP}}" + }, + { + "bbox": [ + 46, + 236, + 287, + 291 + ], + "type": "text", + "content": " based on CLIP [34] backbone (right plot, decrease in iFID in the absence of mode collapse)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 304, + 286, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 304, + 286, + 387 + ], + "spans": [ + { + "bbox": [ + 46, + 304, + 286, + 387 + ], + "type": "text", + "content": "held-out validation set and " + }, + { + "bbox": [ + 46, + 304, + 286, + 387 + ], + "type": "inline_equation", + "content": "50\\mathrm{k}" + }, + { + "bbox": [ + 46, + 304, + 286, + 387 + ], + "type": "text", + "content": " generated samples to evaluate FID in each case. As FID is biased towards ImageNet and can be arbitrarily manipulated, we also report " + }, + { + "bbox": [ + 46, + 304, + 286, + 387 + ], + "type": "inline_equation", + "content": "\\mathrm{FID}_{\\mathrm{CLIP}}" + }, + { + "bbox": [ + 46, + 304, + 286, + 387 + ], + "type": "text", + "content": ". b) Precision & Recall: As we aim to mitigate the mode collapse and achieve diverse generations across classes, we use improved Precision & Recall [25] metrics, as poor recall indicates mode collapse [39]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 388, + 286, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 388, + 286, + 651 + ], + "spans": [ + { + "bbox": [ + 46, + 388, + 286, + 651 + ], + "type": "text", + "content": "d) Intra-Class " + }, + { + "bbox": [ + 46, + 388, + 286, + 651 + ], + "type": "inline_equation", + "content": "\\mathbf{FID}_{\\mathrm{CLIP}}" + }, + { + "bbox": [ + 46, + 388, + 286, + 651 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 46, + 388, + 286, + 651 + ], + "type": "inline_equation", + "content": "\\mathbf{iFID}_{\\mathrm{CLIP}}" + }, + { + "bbox": [ + 46, + 388, + 286, + 651 + ], + "type": "text", + "content": "): The usage of only FID based on Inception-V3 Networks for evaluation of Generative Models has severe limitations, as it has been found that FID can be reduced easily by some fringe features [26]. iFID is computed by taking FID between 5k generated and real samples for the same class. As we want to evaluate both class consistency and diversity, we find that similar limitations exist for intra-class FID (iFID), which has been used to evaluate class-conditional GANs [18]. In Fig. 5, we show the existence of generated images for a particular class (more in Appendix) from models trained on iNaturalist 2019, where iFID is better for the mode collapsed model than the other model generating diverse images. Whereas the " + }, + { + "bbox": [ + 46, + 388, + 286, + 651 + ], + "type": "inline_equation", + "content": "\\mathbf{iFID}_{\\mathrm{CLIP}}" + }, + { + "bbox": [ + 46, + 388, + 286, + 651 + ], + "type": "text", + "content": ", based on CLIP backbone can rank the models correctly with the model having mode collapse having high " + }, + { + "bbox": [ + 46, + 388, + 286, + 651 + ], + "type": "inline_equation", + "content": "\\mathbf{iFID}_{\\mathrm{CLIP}}" + }, + { + "bbox": [ + 46, + 388, + 286, + 651 + ], + "type": "text", + "content": ". Further, we find that the mean iFID can be deceptive in detecting class confusion and collapse cases, as it sometimes ranks models with high realism better than models generating diversity (See Appendix). Hence, mean " + }, + { + "bbox": [ + 46, + 388, + 286, + 651 + ], + "type": "inline_equation", + "content": "\\mathbf{iFID}_{\\mathrm{CLIP}}" + }, + { + "bbox": [ + 46, + 388, + 286, + 651 + ], + "type": "text", + "content": " (ref. to as " + }, + { + "bbox": [ + 46, + 388, + 286, + 651 + ], + "type": "inline_equation", + "content": "\\mathbf{iFID}_{\\mathrm{CLIP}}" + }, + { + "bbox": [ + 46, + 388, + 286, + 651 + ], + "type": "text", + "content": " in result section for brevity) can be reliably used to evaluate models for class consistency and diversity." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 653, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 286, + 712 + ], + "type": "text", + "content": "Baselines: For evaluating NoisyTwins performance in comparison to other methods, we use the implementations present in StudioGAN [16]. For fairness, we re-run all the baselines on StyleGAN2 in the same hyperparameter setting. We compare our method to the StyleGAN2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 312 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 312 + ], + "type": "text", + "content": "(SG2) and StyleGAN2 with augmentation (DiffAug [54] and ADA [20]) baselines. We further tried to improve the baseline by incorporating the recent LeCam regularization method; however, it resulted in gains only for the iNaturalist 2019 dataset, where we use LeCam for all experiments. Further on StyleGAN2, we also use contrastive D2D-CE loss conditioning (i.e. ReACGAN) as a baseline. However, the D2D-CE baseline completely ignores learning of tail classes (Fig. 3) for CIFAR10-LT and is expensive to train; hence we do not report results for it for large-scale long-tailed datasets. We also compare our method against the recent SotA group Spectral Normalization (gSR) [35] method, which we implement for StyleGAN2 by constraining the spectral norms of embedding parameters of the generator " + }, + { + "bbox": [ + 304, + 72, + 545, + 312 + ], + "type": "inline_equation", + "content": "(\\mathcal{G})" + }, + { + "bbox": [ + 304, + 72, + 545, + 312 + ], + "type": "text", + "content": " as suggested by authors. As a sanity check, we reproduce their results on CIFAR10-LT and find that our implementation matches the reported results correctly. We provide results on all datasets, for both the proposed Noise Augmentation (+ Noise) and the overall proposed NoisyTwins (+NoisyTwins) method." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 322, + 526, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 322, + 526, + 335 + ], + "spans": [ + { + "bbox": [ + 305, + 322, + 526, + 335 + ], + "type": "text", + "content": "5.2. Results on Long-Tailed Data Distributions" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 341, + 545, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 341, + 545, + 556 + ], + "spans": [ + { + "bbox": [ + 304, + 341, + 545, + 556 + ], + "type": "text", + "content": "CIFAR10-LT. We applied DiffAug [54] on all baselines, except on gSR, where we found that DiffAug provides inferior results compared to ADA (as also used by authors [35]). It can be observed in Table 2 that the addition of NoisyTwins regularization significantly improves over baseline by " + }, + { + "bbox": [ + 304, + 341, + 545, + 556 + ], + "type": "inline_equation", + "content": "(\\sim 14" + }, + { + "bbox": [ + 304, + 341, + 545, + 556 + ], + "type": "text", + "content": " FID) along with providing superior class consistency as shown by improved iFIDCLIP. NoisyTwins is also able to outperform the recent gSR regularization method and achieves improved results for all metrics. Further, NoisyTwins improves FID for StyleGAN2-ADA baseline used by gSR too from 32.08 to 23.02, however the final results are inferior than reported DiffAug baseline results. Further, we observed that despite not producing any tail class images (Fig. 3), the D2D-CE baseline has much superior FID in comparison to baselines. Whereas the proposed iFIDCLIP value is similar for the baseline and D2D-CE model. This clearly demonstrates the superiority of proposed iFIDCLIP in detecting class confusion." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "content": "Real-world, Large-scale Long-Tailed Datasets. We experiment with iNaturalist 2019 and ImageNet-LT. These datasets are particularly challenging as they contain long-tailed imbalances and semantically similar classes, making GANs prone to mode collapse and class confusion. The baselines StyleGAN2 and StyleGAN2-ADA both suffer from mode collapse (Fig. 6), particularly for the tail classes. Whereas for the recent SotA gSR method, we find that although it undergoes less collapse in comparison to baselines, it suffers from class confusion as seen from similar Intra-FIDCLIP in comparison to baselines (Table 1). Compared to that, our method NoisyTwins improves when used with StyleGAN2-ADA significantly, leading to" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5992" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 106, + 545, + 184 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "type": "text", + "content": "Table 1. Quantitative results on ImageNet-LT and iNaturalist 2019 Datasets. We compare FID(↓), FIDCLIP(↓), iFIDCLIP(↓), Precision(↑) and Recall(↑) with other existing approaches on StyleGAN2 (SG2). We obtain an average ~ 19% relative improvement on FID, ~ 33% on FIDCLIP, and ~ 11% on iFIDCLIP metrics over the previous SotA on ImageNet-LT and iNaturalist 2019 datasets." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 106, + 545, + 184 + ], + "lines": [ + { + "bbox": [ + 48, + 106, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 545, + 184 + ], + "type": "table", + "html": "
MethodImageNet-LTiNaturalist 2019
FID(↓)FIDCLIP(↓)iFIDCLIP(↓)Precision(↑)Recall(↑)FID(↓)FIDCLIP(↓)iFIDCLIP(↓)Precision(↑)Recall(↑)
SG2 [23]41.2511.6446.930.500.4819.343.3338.240.740.17
SG2+ADA [20]37.2011.0447.410.540.3814.922.3035.190.750.57
SG2+ADA+gSR [35]24.788.2144.420.630.3515.172.0636.220.740.46
SG2+ADA+Noise (Ours)22.177.1141.200.720.3312.871.3731.430.810.63
+ NoisyTwins (Ours)21.296.4139.740.670.4911.461.1431.500.790.67
", + "image_path": "7f2c2981b2ff57d28c890a404f7b9e6ce714d9194ac3f1ebdba1bae7f3751a92.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 49, + 232, + 370, + 304 + ], + "blocks": [ + { + "bbox": [ + 47, + 189, + 349, + 222 + ], + "lines": [ + { + "bbox": [ + 47, + 189, + 349, + 222 + ], + "spans": [ + { + "bbox": [ + 47, + 189, + 349, + 222 + ], + "type": "text", + "content": "Table 2. Quantitative results on CIFAR10-LT Dataset. We compare with other existing approaches. We obtain " + }, + { + "bbox": [ + 47, + 189, + 349, + 222 + ], + "type": "inline_equation", + "content": "\\sim 26\\%" + }, + { + "bbox": [ + 47, + 189, + 349, + 222 + ], + "type": "text", + "content": " relative improvement over the existing methods on " + }, + { + "bbox": [ + 47, + 189, + 349, + 222 + ], + "type": "inline_equation", + "content": "\\mathrm{FID}_{\\mathrm{CLIP}}" + }, + { + "bbox": [ + 47, + 189, + 349, + 222 + ], + "type": "text", + "content": " and iFIDCLIP metrics." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 232, + 370, + 304 + ], + "lines": [ + { + "bbox": [ + 49, + 232, + 370, + 304 + ], + "spans": [ + { + "bbox": [ + 49, + 232, + 370, + 304 + ], + "type": "table", + "html": "
MethodFID(↓)FIDCLIP(↓)iFIDCLIP(↓)Precision(↑)Recall(↑)
SG2+DiffAug [54]31.736.2711.590.630.35
SG2+D2D-CE [17]19.974.7711.350.730.42
gSR [35]22.105.549.940.700.29
SG2+DiffAug+Noise (Ours)28.905.2610.650.710.38
+ NoisyTwins(Ours)17.743.557.240.700.51
", + "image_path": "7404f5e305388745ff6e4780caa40d6c80c7de24d1f9ef8a6fb61b7ff1a13bb0.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 381, + 232, + 531, + 304 + ], + "blocks": [ + { + "bbox": [ + 365, + 189, + 544, + 232 + ], + "lines": [ + { + "bbox": [ + 365, + 189, + 544, + 232 + ], + "spans": [ + { + "bbox": [ + 365, + 189, + 544, + 232 + ], + "type": "text", + "content": "Table 3. Comparison with SotA approaches on BigGAN. We compare FID(↓) with other existing models on ImageNet-LT (IN-LT) and iNaturalist 2019 (iNat-19)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 381, + 232, + 531, + 304 + ], + "lines": [ + { + "bbox": [ + 381, + 232, + 531, + 304 + ], + "spans": [ + { + "bbox": [ + 381, + 232, + 531, + 304 + ], + "type": "table", + "html": "
MethodiNat-19IN-LT
BigGAN [4]14.8528.10
+ gSR [35]13.95-
ICGAN [6]-23.40
StyleGAN2-ADA [20]14.9237.20
+ NoisyTwins (Ours)11.4621.29
", + "image_path": "d01ec9abc039336a7c583af412b88ec40c084f42846e413a87a3cc8e73844de4.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 312, + 288, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 312, + 288, + 420 + ], + "spans": [ + { + "bbox": [ + 46, + 312, + 288, + 420 + ], + "type": "text", + "content": "a relative improvement of " + }, + { + "bbox": [ + 46, + 312, + 288, + 420 + ], + "type": "inline_equation", + "content": "42.7\\%" + }, + { + "bbox": [ + 46, + 312, + 288, + 420 + ], + "type": "text", + "content": " in FID for ImageNet-LT and " + }, + { + "bbox": [ + 46, + 312, + 288, + 420 + ], + "type": "inline_equation", + "content": "23.19\\%" + }, + { + "bbox": [ + 46, + 312, + 288, + 420 + ], + "type": "text", + "content": " on the iNaturalist 2019 dataset when added to StyleGAN2-ADA baseline. Further with Noise Augmentation " + }, + { + "bbox": [ + 46, + 312, + 288, + 420 + ], + "type": "inline_equation", + "content": "(+\\mathrm{Noise})" + }, + { + "bbox": [ + 46, + 312, + 288, + 420 + ], + "type": "text", + "content": ", we observe generations of high-quality class-consistent images, but it also suffers from mode collapse. This can be observed by the high-precision values in comparison to low-recall values. However, adding NoisyTwins regularization over the noise augmentation improves diversity by improving recall (Table 1)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 421, + 287, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 421, + 287, + 661 + ], + "spans": [ + { + "bbox": [ + 46, + 421, + 287, + 661 + ], + "type": "text", + "content": "Fig. 6 presents the generated images of tail classes for various methods on ImageNet-LT, where NoisyTwins generations show remarkable diversity in comparison to others. The presence of diversity for classes with just 5-6 training images demonstrates successful transfer of knowledge from head classes to tail classes, due to shared parameters. Further, to compare with existing SotA reported results, we compare FID of BigGAN models from gSR [35] and Instance Conditioned GAN (ICGAN) [6]. For fairness, we compare FID on the validation set for which we obtained gSR models from authors and re-evaluate them, as they reported FID on a balanced training set. As BigGAN models are more common for class-conditioned generation [18], their baseline performs superior to StyleGAN2-ADA baselines (Table 3). However, the addition of NoisyTwins to the StyleGAN2-ADA method improves it significantly, even outperforming the existing methods of gSR (by " + }, + { + "bbox": [ + 46, + 421, + 287, + 661 + ], + "type": "inline_equation", + "content": "18.44\\%" + }, + { + "bbox": [ + 46, + 421, + 287, + 661 + ], + "type": "text", + "content": " ) and ICGAN (by " + }, + { + "bbox": [ + 46, + 421, + 287, + 661 + ], + "type": "inline_equation", + "content": "9.44\\%" + }, + { + "bbox": [ + 46, + 421, + 287, + 661 + ], + "type": "text", + "content": " ) based on BigGAN architecture. This shows that NoisyTwins allows the StyleGAN2 baseline to scale to large and diverse long-tailed datasets." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 670, + 228, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 228, + 682 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 228, + 682 + ], + "type": "text", + "content": "5.3. NoisyTwins on Few-Shot Datasets" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "content": "We now demonstrate the potential of NoisyTwins in another challenging scenario of class-conditional few-shot" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 312, + 547, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 312, + 547, + 494 + ], + "spans": [ + { + "bbox": [ + 304, + 312, + 547, + 494 + ], + "type": "text", + "content": "image generation from GANs. We perform our experiments using a conditional StyleGAN2-ADA baseline, for which we tune hyper-parameters to obtain a strong baseline. We then apply our method of Noise Augmentation and NoisyTwins over the strong baseline for reporting our results. We use the few-shot dataset of LHI-AnimalFaces [45] and a subset of ImageNet Carnivores [29, 40] to report our results. Table 4 shows the results of these experiments, where we find that our method, NoisyTwins, significantly improves the FID of StyleGAN2 ADA baseline by " + }, + { + "bbox": [ + 304, + 312, + 547, + 494 + ], + "type": "inline_equation", + "content": "(22.2\\%)" + }, + { + "bbox": [ + 304, + 312, + 547, + 494 + ], + "type": "text", + "content": " on average for both datasets. Further, combining Noisy Twins with SotA Transitional-cGAN [40] through official code, also leads to effective improvement in FID. These results clearly demonstrate the diverse potential and applicability of our proposed method NoisyTwins." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 508, + 365, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 508, + 365, + 521 + ], + "spans": [ + { + "bbox": [ + 306, + 508, + 365, + 521 + ], + "type": "text", + "content": "6. Analysis" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 529, + 547, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 529, + 547, + 638 + ], + "spans": [ + { + "bbox": [ + 304, + 529, + 547, + 638 + ], + "type": "text", + "content": "We perform analysis of NoisyTwins w.r.t. to its hyperparameters, standard deviation " + }, + { + "bbox": [ + 304, + 529, + 547, + 638 + ], + "type": "inline_equation", + "content": "(\\sigma)" + }, + { + "bbox": [ + 304, + 529, + 547, + 638 + ], + "type": "text", + "content": " of noise augmentation and regularization strength " + }, + { + "bbox": [ + 304, + 529, + 547, + 638 + ], + "type": "inline_equation", + "content": "(\\lambda)" + }, + { + "bbox": [ + 304, + 529, + 547, + 638 + ], + "type": "text", + "content": ". We also compare NoisyTwins objective (ref. Eq. 6) with contrastive objective. Finally, we compare NoisyTwins over Latent Diffusion Models for long-tailed class conditional generation task. We perform ablation experiments on CIFAR10-LT, for which additional details and results are present in Appendix. We also present comparison of NoisyTwins for GAN fine-tuning." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 641, + 545, + 653 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 653 + ], + "type": "text", + "content": "How much noise and regularization strength is optimal?" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 654, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 714 + ], + "type": "text", + "content": "In Fig. 7, we ablate over the noise variance parameter " + }, + { + "bbox": [ + 304, + 654, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 304, + 654, + 547, + 714 + ], + "type": "text", + "content": " for CIFAR10-LT. We find that a moderate value of noise strength 0.75 leads to optimal results. For the strength of NoisyTwins loss " + }, + { + "bbox": [ + 304, + 654, + 547, + 714 + ], + "type": "inline_equation", + "content": "(\\lambda)" + }, + { + "bbox": [ + 304, + 654, + 547, + 714 + ], + "type": "text", + "content": ", we find that the algorithm performs similarly on values near 0.01 and is robust to it (Fig. 7)." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "text", + "content": "5993" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 71, + 218, + 198 + ], + "blocks": [ + { + "bbox": [ + 51, + 71, + 218, + 198 + ], + "lines": [ + { + "bbox": [ + 51, + 71, + 218, + 198 + ], + "spans": [ + { + "bbox": [ + 51, + 71, + 218, + 198 + ], + "type": "image", + "image_path": "71c2871399f1044f6cba014b9cef28dfd0f65057b005f7d878839c3bc5d097ce.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 126, + 198, + 214, + 215 + ], + "lines": [ + { + "bbox": [ + 126, + 198, + 214, + 215 + ], + "spans": [ + { + "bbox": [ + 126, + 198, + 214, + 215 + ], + "type": "text", + "content": "(a) Image Augmentation based Methods (ADA)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 224, + 72, + 320, + 198 + ], + "blocks": [ + { + "bbox": [ + 224, + 72, + 320, + 198 + ], + "lines": [ + { + "bbox": [ + 224, + 72, + 320, + 198 + ], + "spans": [ + { + "bbox": [ + 224, + 72, + 320, + 198 + ], + "type": "image", + "image_path": "81ede3449877dabd306eeb752501b5a369e7735973a56bdf84fc7984f41a5836.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 236, + 199, + 308, + 216 + ], + "lines": [ + { + "bbox": [ + 236, + 199, + 308, + 216 + ], + "spans": [ + { + "bbox": [ + 236, + 199, + 308, + 216 + ], + "type": "text", + "content": "(b) Regularizer based Methods (gSR)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 327, + 72, + 543, + 198 + ], + "blocks": [ + { + "bbox": [ + 327, + 72, + 543, + 198 + ], + "lines": [ + { + "bbox": [ + 327, + 72, + 543, + 198 + ], + "spans": [ + { + "bbox": [ + 327, + 72, + 543, + 198 + ], + "type": "image", + "image_path": "6d18060b187b405762c6daf9bd05455025eb712304f459eba2dd10d557b74dd1.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 402, + 199, + 471, + 208 + ], + "lines": [ + { + "bbox": [ + 402, + 199, + 471, + 208 + ], + "spans": [ + { + "bbox": [ + 402, + 199, + 471, + 208 + ], + "type": "text", + "content": "(c) NoisyTwins (Ours)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 50, + 308, + 296, + 414 + ], + "blocks": [ + { + "bbox": [ + 47, + 266, + 288, + 299 + ], + "lines": [ + { + "bbox": [ + 47, + 266, + 288, + 299 + ], + "spans": [ + { + "bbox": [ + 47, + 266, + 288, + 299 + ], + "type": "text", + "content": "Table 4. Quantitative results on ImageNet Carnivore and AnimalFace Datasets. Our method improves over both StyleGAN2-ADA (SG2-ADA) baseline and SotA Transitional-cGAN." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 308, + 296, + 414 + ], + "lines": [ + { + "bbox": [ + 50, + 308, + 296, + 414 + ], + "spans": [ + { + "bbox": [ + 50, + 308, + 296, + 414 + ], + "type": "table", + "html": "
ImageNet CarnivoreAnimalFace
MethodFID(↓)iFIDCLIP(↓)FID(↓)iFIDCLIP(↓)
SG2 [23]111.8336.3494.0929.94
SG2+ADA [20]22.7712.8520.2511.12
SG2+ADA+Noise (Ours)19.2512.5118.7810.42
+ NoisyTwins (Ours)16.0112.4117.2710.03
FID(↓)FID(↓)
Transitional-cGAN [40]14.6020.53
+ NoisyTwins (Ours)13.6516.15
", + "image_path": "2e4dd24d72727438c148f93e6ec702fcb0c6345ff0fee142bc18fa2d30ead438.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 438, + 287, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 438, + 287, + 570 + ], + "spans": [ + { + "bbox": [ + 46, + 438, + 287, + 570 + ], + "type": "text", + "content": "Which type of self-supervision to use with noise augmentation? The goal of our method is to achieve invariance to Noise Augmentation in the " + }, + { + "bbox": [ + 46, + 438, + 287, + 570 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 46, + 438, + 287, + 570 + ], + "type": "text", + "content": " latent space. This can be achieved using either contrastive learning-based methods like SimCLR [7] or negative-free method like Barlow Twins [53]. Contrastive loss (SimCLR based) produces FID of 26.23 vs 17.74 by NoisyTwins (BarlowTwins based). We find that contrastive baseline improves over the noise augmentation baseline (28.90) however falls significantly below the NoisyTwins, as the former requires a large batch size to be effective which is expensive for GANs." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": "How does NoisyTwins compare with modern Vision and Language models? For evaluating the effectiveness of modern vision language-based diffusion models, we test the generation of the iNaturalist 2019 dataset by creating the prompt \"a photo of S\" where we replace the class name in place of S. We use the LDM [37] model trained on LAION-400M to perform inference, generating 50 images per class. We obtained an FID of 57.04 in comparison to best FID of 11.46 achieved by NoisyTwins. This clearly demonstrates that for specific use cases like fine-grained generation, GANs are still ahead of general-purpose LDM." + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 306, + 266, + 545, + 388 + ], + "blocks": [ + { + "bbox": [ + 46, + 228, + 547, + 262 + ], + "lines": [ + { + "bbox": [ + 46, + 228, + 547, + 262 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 547, + 262 + ], + "type": "text", + "content": "Figure 6. Qualitative results on ImageNet-LT for tail classes. We find that existing SotA methods for tail classes show collapsed (a) or arbitrary image generation (b). With NoisyTwins, we observe diverse and class-consistent image generation, even for classes having 5-6 images. The tail classes get enhanced diversity by transferring the knowledge from head classes, as they share parameters." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 266, + 545, + 388 + ], + "lines": [ + { + "bbox": [ + 306, + 266, + 545, + 388 + ], + "spans": [ + { + "bbox": [ + 306, + 266, + 545, + 388 + ], + "type": "image", + "image_path": "57d1a50ade434231c98ee56695e446f0c59a3da443295139f4ff82d16eb117e0.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 396, + 545, + 429 + ], + "lines": [ + { + "bbox": [ + 305, + 396, + 545, + 429 + ], + "spans": [ + { + "bbox": [ + 305, + 396, + 545, + 429 + ], + "type": "text", + "content": "Figure 7. Ablation of Hyperparameters. Quantitative comparison on CIFAR10-LT for standard deviation of Noise Augmentation " + }, + { + "bbox": [ + 305, + 396, + 545, + 429 + ], + "type": "inline_equation", + "content": "(\\sigma)" + }, + { + "bbox": [ + 305, + 396, + 545, + 429 + ], + "type": "text", + "content": " and strength " + }, + { + "bbox": [ + 305, + 396, + 545, + 429 + ], + "type": "inline_equation", + "content": "(\\lambda)" + }, + { + "bbox": [ + 305, + 396, + 545, + 429 + ], + "type": "text", + "content": " of NoisyTwins loss." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 433, + 379, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 433, + 379, + 445 + ], + "spans": [ + { + "bbox": [ + 306, + 433, + 379, + 445 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 450, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 450, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 545, + 715 + ], + "type": "text", + "content": "In this work, we analyze the performance of StyleGAN2 models on the real-world long-tailed datasets including iNaturalist 2019 and ImageNet-LT. We find that existing works lead to either class confusion or mode collapse in the image space. This phenomenon is rooted in collapse and confusion in the latent " + }, + { + "bbox": [ + 304, + 450, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 304, + 450, + 545, + 715 + ], + "type": "text", + "content": " space of StyleGAN2. Through our analysis, we deduce that this collapse occurs when the latents become invariant to random conditioning vectors " + }, + { + "bbox": [ + 304, + 450, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 304, + 450, + 545, + 715 + ], + "type": "text", + "content": ", and collapse for each class. To mitigate this, we introduce inexpensive noise based augmentation for discrete class embeddings. Further, to ensure class consistency, we couple this augmentation technique with BarlowTwins' objective in the latent " + }, + { + "bbox": [ + 304, + 450, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 304, + 450, + 545, + 715 + ], + "type": "text", + "content": " space which imparts intra-class diversity to latent " + }, + { + "bbox": [ + 304, + 450, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 304, + 450, + 545, + 715 + ], + "type": "text", + "content": " vectors. The noise augmentation and regularization comprises our proposed NoisyTwins technique, which improves the performance of StyleGAN2 establishing a new SotA on iNaturalist 2019 and ImageNet-LT. The extension of NoisyTwins for conditioning on more-complex attributes for StyleGANs is a good direction for future work. Acknowledgements: This work was supported in part by SERB-STAR Project (STR/2020/000128). Harsh Rangwani is supported by PMRF fellowship." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5994" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 289, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 289, + 147 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 289, + 147 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 289, + 147 + ], + "type": "text", + "content": "[1] Yuval Alaluf, Omer Tov, Ron Mokady, Rinon Gal, and Amit Bermano. Hyperstyle: Stylegan inversion with hypernetworks for real image editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18511-18521, June 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 148, + 288, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 288, + 192 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 288, + 192 + ], + "type": "text", + "content": "[2] Adrien Bardes, Jean Ponce, and Yann LeCun. VI-CReg: Variance-invariance-covariance regularization for self-supervised learning. In International Conference on Learning Representations, 2022. 2, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 194, + 288, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 194, + 288, + 236 + ], + "spans": [ + { + "bbox": [ + 53, + 194, + 288, + 236 + ], + "type": "text", + "content": "[3] Stella Bounareli, Vasileios Argyriou, and Georgios Tzimiropoulos. Finding directions in gan's latent space for neural face reenactment. arXiv preprint arXiv:2202.00046, 2022.1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 239, + 288, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 239, + 288, + 282 + ], + "spans": [ + { + "bbox": [ + 53, + 239, + 288, + 282 + ], + "type": "text", + "content": "[4] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale GAN training for high fidelity natural image synthesis. In International Conference on Learning Representations, 2019. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 284, + 288, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 284, + 288, + 327 + ], + "spans": [ + { + "bbox": [ + 53, + 284, + 288, + 327 + ], + "type": "text", + "content": "[5] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. 2020. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 329, + 288, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 329, + 288, + 373 + ], + "spans": [ + { + "bbox": [ + 53, + 329, + 288, + 373 + ], + "type": "text", + "content": "[6] Arantxa Casanova, Marlene Careil, Jakob Verbeek, Michal Drozdal, and Adriana Romero-Soriano. Instance-conditioned gan. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 3, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 375, + 288, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 375, + 288, + 418 + ], + "spans": [ + { + "bbox": [ + 53, + 375, + 288, + 418 + ], + "type": "text", + "content": "[7] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. arXiv preprint arXiv:2002.05709, 2020.8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 420, + 288, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 420, + 288, + 474 + ], + "spans": [ + { + "bbox": [ + 53, + 420, + 288, + 474 + ], + "type": "text", + "content": "[8] Yin Cui, Menglin Jia, Tsung-Yi Lin, Yang Song, and Serge Belongie. Class-balanced loss based on effective number of samples. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9268-9277, 2019. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 476, + 288, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 476, + 288, + 520 + ], + "spans": [ + { + "bbox": [ + 53, + 476, + 288, + 520 + ], + "type": "text", + "content": "[9] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139-144, 2020. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 522, + 288, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 522, + 288, + 588 + ], + "spans": [ + { + "bbox": [ + 48, + 522, + 288, + 588 + ], + "type": "text", + "content": "[10] Jean-Bastien Grill, Florian Strub, Florent Altché, Coretin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 590, + 288, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 288, + 632 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 288, + 632 + ], + "type": "text", + "content": "[11] Erik Härkönen, Aaron Hertzmann, Jaakko Lehtinen, and Sylvain Paris. Ganspace: Discovering interpretable gan controls. Advances in Neural Information Processing Systems, 33:9841-9850, 2020. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 635, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 288, + 689 + ], + "type": "text", + "content": "[12] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 2, 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "type": "text", + "content": "[13] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceed-" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "ings of the IEEE international conference on computer vision, pages 1501-1510, 2017. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 95, + 545, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 95, + 545, + 138 + ], + "spans": [ + { + "bbox": [ + 307, + 95, + 545, + 138 + ], + "type": "text", + "content": "[14] Jongheon Jeong and Jinwoo Shin. Training GANs with stronger augmentations via contrastive discriminator. In International Conference on Learning Representations, 2021. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 140, + 545, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 140, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 140, + 545, + 172 + ], + "type": "text", + "content": "[15] Liming Jiang, Bo Dai, Wayne Wu, and Chen Change Loy. Deceive D: Adaptive Pseudo Augmentation for GAN training with limited data. In NeurIPS, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 173, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 173, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 307, + 173, + 545, + 194 + ], + "type": "text", + "content": "[16] Minguk Kang and Jaesik Park. ContraGAN: Contrastive Learning for Conditional Image Generation. 2020. 3, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 195, + 545, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 195, + 545, + 237 + ], + "spans": [ + { + "bbox": [ + 307, + 195, + 545, + 237 + ], + "type": "text", + "content": "[17] Minguk Kang, Woohyeon Shim, Minsu Cho, and Jaesik Park. Rebooting acgan: Auxiliary classifier gans with stable training. Advances in neural information processing systems, 34:23505-23518, 2021. 2, 3, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 239, + 545, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 239, + 545, + 271 + ], + "spans": [ + { + "bbox": [ + 307, + 239, + 545, + 271 + ], + "type": "text", + "content": "[18] MinGuk Kang, Joonghyuk Shin, and Jaesik Park. StudioGAN: A Taxonomy and Benchmark of GANs for Image Synthesis. 2206.09479 (arXiv), 2022. 5, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 272, + 545, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 272, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 307, + 272, + 545, + 304 + ], + "type": "text", + "content": "[19] Tero Karras, Timo Aila, Samuli Laine, and Jaakko Lehtinen. Progressive growing of gans for improved quality, stability, and variation. arXiv preprint arXiv:1710.10196, 2017. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 305, + 545, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 305, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 307, + 305, + 545, + 348 + ], + "type": "text", + "content": "[20] Tero Karras, Miika Aittala, Janne Hellsten, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Training generative adversarial networks with limited data. In Proc. NeurIPS, 2020. 1, 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 350, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 350, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 307, + 350, + 545, + 403 + ], + "type": "text", + "content": "[21] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 1, 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 404, + 545, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 404, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 307, + 404, + 545, + 459 + ], + "type": "text", + "content": "[22] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 460, + 545, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 460, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 307, + 460, + 545, + 502 + ], + "type": "text", + "content": "[23] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. In Proc. CVPR, 2020. 3, 7, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 503, + 545, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 503, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 307, + 503, + 545, + 536 + ], + "type": "text", + "content": "[24] Tuomas Kynkänniemi, Tero Karras, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Improved Precision and Recall Metric for Assessing Generative Models. 2019. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 537, + 545, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 537, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 307, + 537, + 545, + 581 + ], + "type": "text", + "content": "[25] Tuomas Kynkänniemi, Tero Karras, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Improved precision and recall metric for assessing generative models. Advances in Neural Information Processing Systems, 32, 2019. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 582, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 582, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 307, + 582, + 545, + 613 + ], + "type": "text", + "content": "[26] Tuomas Kynkänniemi, Tero Karras, Miika Aittala, Timo Aila, and Jaakko Lehtinen. The role of imagenet classes in fréchet inception distance. CoRR, abs/2203.06026, 2022. 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 614, + 545, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 614, + 545, + 646 + ], + "spans": [ + { + "bbox": [ + 307, + 614, + 545, + 646 + ], + "type": "text", + "content": "[27] Thomas Leimkuhler and George Drettakis. Freestylegan: Free-view editable portrait rendering with the camera manifold. 40(6), 2021. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 647, + 545, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 680 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 680 + ], + "type": "text", + "content": "[28] Feng Liu and Xiaoming Liu. 2d gans meet unsupervised single-view 3d reconstruction. In Computer Vision - ECCV 2022, pages 497-514, 2022. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 681, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 681, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 681, + 547, + 713 + ], + "type": "text", + "content": "[29] Ming-Yu Liu, Xun Huang, Arun Mallya, Tero Karras, Timo Aila, Jaakko Lehtinen, and Jan Kautz. Few-shot unsupervised image-to-image translation. In Proceedings of" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5995" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 66, + 72, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 286, + 95 + ], + "type": "text", + "content": "the IEEE/CVF international conference on computer vision, pages 10551-10560, 2019. 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 151 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 151 + ], + "type": "text", + "content": "[30] Ziwei Liu, Zhongqi Miao, Xiaohang Zhan, Jiayun Wang, Boqing Gong, and Stella X Yu. Large-scale long-tailed recognition in an open world. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2537-2546, 2019. 2, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 287, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 287, + 195 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 287, + 195 + ], + "type": "text", + "content": "[31] Lars Mescheder, Andreas Geiger, and Sebastian Nowozin. Which training methods for gans do actually converge? In International conference on machine learning, pages 3481-3490. PMLR, 2018. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 197, + 287, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 197, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 197, + 287, + 262 + ], + "type": "text", + "content": "[32] Rishubh Parihar, Ankit Dhiman, Tejan Karmali, and R. Venkatesh Babu. Everything is there in latent space: Attribute editing and attribute style manipulation by stylegan latent space exploration. In Proceedings of the 30th ACM International Conference on Multimedia, pages 1828-1836, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 263, + 287, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 263, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 263, + 287, + 319 + ], + "type": "text", + "content": "[33] Or Patashnik, Zongze Wu, Eli Shechtman, Daniel Cohen-Or, and Dani Lischinski. Styleclip: Text-driven manipulation of stylegan imagery. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2085–2094, October 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 319, + 287, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 319, + 287, + 386 + ], + "spans": [ + { + "bbox": [ + 48, + 319, + 287, + 386 + ], + "type": "text", + "content": "[34] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 2, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 387, + 287, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 387, + 287, + 431 + ], + "spans": [ + { + "bbox": [ + 48, + 387, + 287, + 431 + ], + "type": "text", + "content": "[35] Harsh Rangwani, Naman Jaswani, Tejan Karmali, Varun Jampani, and R. Venkatesh Babu. Improving gans for long-tailed data through group spectral regularization. In European Conference on Computer Vision, 2022. 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 432, + 287, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 432, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 48, + 432, + 287, + 475 + ], + "type": "text", + "content": "[36] Harsh Rangwani, Konda Reddy Mopuri, and R Venkatesh Babu. Class balancing gan with a classifier in the loop. In Uncertainty in Artificial Intelligence, pages 1618-1627. PMLR, 2021. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 476, + 287, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 476, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 48, + 476, + 287, + 509 + ], + "type": "text", + "content": "[37] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2021. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 510, + 287, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 510, + 287, + 543 + ], + "spans": [ + { + "bbox": [ + 48, + 510, + 287, + 543 + ], + "type": "text", + "content": "[38] Axel Sauer, Kashyap Chitta, Jens Müller, and Andreas Geiger. Projected gans converge faster. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 1, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 544, + 287, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 287, + 577 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 287, + 577 + ], + "type": "text", + "content": "[39] Axel Sauer, Katja Schwarz, and Andreas Geiger. Styleganx1: Scaling stylegan to large diverse datasets. volume abs/2201.00273, 2022. 1, 5, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 578, + 287, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 287, + 622 + ], + "type": "text", + "content": "[40] Mohamad Shahbazi, Martin Danelljan, Danda Pani Paudel, and Luc Van Gool. Collapse by conditioning: Training class-conditional GANs with limited data. In International Conference on Learning Representations, 2022. 3, 7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 624, + 287, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 655 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 655 + ], + "type": "text", + "content": "[41] Yujun Shen, Jinjin Gu, Xiaou Tang, and Bolei Zhou. Interpreting the latent space of gans for semantic face editing. In CVPR, 2020. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 657, + 287, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 657, + 287, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 657, + 287, + 690 + ], + "type": "text", + "content": "[42] Yujun Shen, Ceyuan Yang, Xiaou Tang, and Bolei Zhou. Interfacegan: Interpreting the disentangled face representation learned by gans. IEEE TPAMI, 2020. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[43] Yujun Shen and Bolei Zhou. Closed-form factorization of latent semantics in gans. In CVPR, 2021. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 586 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "text", + "content": "[44] Yichun Shi, Divyansh Aggarwal, and Anil K. Jain. Lifting 2d stylegan for 3d-aware face generation. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6254-6262, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "type": "text", + "content": "[45] Zhangzhang Si and Song-Chun Zhu. Learning hybrid image templates (hit) by information projection. IEEE Transactions on pattern analysis and machine intelligence, 34(7):1354-1367, 2011. 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 163, + 545, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 196 + ], + "type": "text", + "content": "[46] Ivan Skorokhodov, Sergey Tulyakov, and Mohamed Elhoseiny. Stylegan-v: A continuous video generator with the price, image quality and perks of stylegan2, 2021. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 197, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 228 + ], + "type": "text", + "content": "[47] Hung-Yu Tseng, Lu Jiang, Ce Liu, Ming-Hsuan Yang, and Weilong Yang. Regularizing generative adversarial networks under limited data. In CVPR, 2021. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 230, + 545, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 230, + 545, + 295 + ], + "spans": [ + { + "bbox": [ + 308, + 230, + 545, + 295 + ], + "type": "text", + "content": "[48] Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alex Shepard, Hartwig Adam, Pietro Perona, and Serge Belongie. The inaturalist species classification and detection dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8769-8778, 2018. 2, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 297, + 545, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 297, + 545, + 351 + ], + "spans": [ + { + "bbox": [ + 308, + 297, + 545, + 351 + ], + "type": "text", + "content": "[49] Zongze Wu, Dani Lischinski, and Eli Shechtman. Stylespace analysis: Disentangled controls for stylegan image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12863–12872, 2021. 1, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 353, + 545, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 353, + 545, + 385 + ], + "spans": [ + { + "bbox": [ + 308, + 353, + 545, + 385 + ], + "type": "text", + "content": "[50] Ceyuan Yang, Yujun Shen, Yinghao Xu, Deli Zhao, Bo Dai, and Bolei Zhou. Improving gans with a dynamic discriminator. 2022. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 387, + 545, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 387, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 308, + 387, + 545, + 441 + ], + "type": "text", + "content": "[51] Ceyuan Yang, Yujun Shen, Yinghao Xu, and Bolei Zhou. Data-efficient instance generation from instance discrimination. In A. Beygelzimer, Y. Dauphin, P. Liang, and J. Wortman Vaughan, editors, Advances in Neural Information Processing Systems, 2021. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 442, + 545, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 442, + 545, + 497 + ], + "spans": [ + { + "bbox": [ + 308, + 442, + 545, + 497 + ], + "type": "text", + "content": "[52] Sihyun Yu, Jihoon Tack, Sangwoo Mo, Hyunsu Kim, Junho Kim, Jung-Woo Ha, and Jinwoo Shin. Generating videos with dynamics-aware implicit generative adversarial networks. In International Conference on Learning Representations, 2022. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 498, + 545, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 498, + 545, + 543 + ], + "spans": [ + { + "bbox": [ + 308, + 498, + 545, + 543 + ], + "type": "text", + "content": "[53] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow twins: Self-supervised learning via redundancy reduction. In International Conference on Machine Learning, pages 12310-12320. PMLR, 2021. 2, 5, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 543, + 545, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 543, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 308, + 543, + 545, + 586 + ], + "type": "text", + "content": "[54] Shengyu Zhao, Zhijian Liu, Ji Lin, Jun-Yan Zhu, and Song Han. Differentiable augmentation for data-efficient gan training. In Conference on Neural Information Processing Systems (NeurIPS), 2020. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5996" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Non-Contrastive Learning Meets Language-Image Pre-Training/fe567c45-5ef2-4da5-8116-0c22fc6e398a_content_list.json b/2023/Non-Contrastive Learning Meets Language-Image Pre-Training/fe567c45-5ef2-4da5-8116-0c22fc6e398a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..7adff2390493e7d16be8c13b2d571b42e110d524 --- /dev/null +++ b/2023/Non-Contrastive Learning Meets Language-Image Pre-Training/fe567c45-5ef2-4da5-8116-0c22fc6e398a_content_list.json @@ -0,0 +1,1785 @@ +[ + { + "type": "text", + "text": "Non-Contrastive Learning Meets Language-Image Pre-Training", + "text_level": 1, + "bbox": [ + 161, + 130, + 808, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jinghao Zhou Li Dong Zhe Gan Lijuan Wang Furu Wei Microsoft", + "bbox": [ + 225, + 180, + 743, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 251, + 313, + 266 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Contrastive language-image pre-training (CLIP) serves as a de-facto standard to align images and texts. Nonetheless, the loose correlation between images and texts of web-crawled data renders the contrastive objective data inefficient and craving for a large training batch size. In this work, we explore the validity of non-contrastive language-image pre-training (nCLIP), and study whether nice properties exhibited in visual self-supervised models can emerge. We empirically observe that the non-contrastive objective benefits representation learning while sufficiently underperforming under zero-shot recognition. Based on the above study, we further introduce xCLIP, a multi-tasking framework combining CLIP and nCLIP, and show that nCLIP aids CLIP in enhancing feature semantics. The synergy between two objectives lets xCLIP enjoy the best of both worlds: superior performance in both zero-shot transfer and representation learning. Systematic evaluation is conducted spanning a wide variety of downstream tasks including zero-shot classification, out-of-domain classification, retrieval, visual representation learning, and textual representation learning, showcasing a consistent performance gain and validating the effectiveness of xCLIP. The code and pre-trained models will be publicly available at https://github.com/shallowtoil/xclip.", + "bbox": [ + 75, + 282, + 473, + 648 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 676, + 209, + 694 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Language-image pre-training which simultaneously learns textual and visual representation from large-scale image-text pairs has revolutionized the field of representation learning [25, 59], vision-language understanding [22], and text-to-image generation [61]. Compared to traditional visual models, language-instilled ones intrinsically inherit the capability of zero-shot or few-shot learning prominently demonstrated by large language models such as GPT-3 [12]. The precursor system, Contrastive Language-Image Pre-Training [59] (CLIP) that explicitly aligns the projected features of two modalities, has demonstrated surprising capabilities of zero-shot, representation learning, and robustness, being applied to a wide range of fields [32, 52, 60, 70].", + "bbox": [ + 75, + 704, + 470, + 902 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a0c87e3957ef3b12e91d888b2fdf06da99582a902e374957c0e45e5be91a978b.jpg", + "image_caption": [ + "Figure 1. Architecture comparison between CLIP and nCLIP. We take base-size encoders for an instance. CLIP discriminates instances within a batch using 512-dim projected embeddings. nCLIP projects each modality into a 32768-dim probability distribution as the pseudo-label to supervise the prediction from the other modality. Darker blocks of nCLIP depict a higher response for cluster distribution, signifying the clusters to which text or image instances may belong. xCLIP is a multi-tasking framework with both CLIP and nCLIP." + ], + "image_footnote": [], + "bbox": [ + 504, + 255, + 893, + 393 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To learn from noisy web-crawled image-text pairs, CLIP adopts a formulation of the contrastive objective, where the image and text within a pair are considered as unique instances and are encouraged to be discriminated from all the other negative instances. However, web-crawled image-text pairs [65,69] are usually loosely correlated in the sense that one caption (image) can match reasonably with multiple images (captions) besides the ground-truth one, as statistically shown in [74]. Hence, it is inaccurate and data-inefficient for representation learning to neglect other sensible matches and overlook the semantics hidden inside the textual description. This is also solidified by the undesirable discriminative ability or transferring performance of the visual encoder pre-trained under the contrastive objective, as suggested in [78]. Mining semantics from plentiful concepts appearing in captions, therefore, solicits further exploration beyond the vanilla contrastive objective.", + "bbox": [ + 496, + 536, + 893, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Previously, some works resort to mining nearest neighbor positive samples via intra-modality similarity [50, 74] but require extra storage for auxiliary modules, i.e., the teacher network [74] or the memory bank [50]. Other works conduct multi-tasking with image-label supervision [57, 78, 80], transforming the captions into tag format or the image tag into the captioning format for unified contrastive", + "bbox": [ + 496, + 795, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "11028", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "learning. Notwithstanding, such practice is still unable to account for the loose assignment within the image-text captioning pairs. Intuitively, a caption for an image usually identifies existing objects within the image, which can be captured by a probabilistic distribution estimating how the text is assigned to one or multiple object clusters [13, 14]. Such estimation depicts the semantic meanings for its contained visual contents, and thus can be leveraged as a pseudo-label to guide the learning of the visual encoder.", + "bbox": [ + 75, + 90, + 470, + 227 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Inspired by the superiority of visual self-supervised learning (SSL) models pre-trained with the non-contrastive objective [14, 15], we explore whether the non-contrastive objective can be used across modalities for pre-training language-image models, and whether nice properties displayed on visual SSL models can be inherited. To this end, we follow the same setup as CLIP except for the objective, and study non-Contrastive Language-Image Pre-Training (nCLIP). For an image-text pair, we use the estimation of the textual (visual) distribution as the target to supervise the visual (textual) distribution, measured by cross-entropy. Additional regularizers are applied to avoid trivial solutions. Schematic comparison between CLIP and nCLIP is shown in Fig. 1. Theoretically, such formulation takes one modality and the cluster centroid that the other modality belongs to as the positive samples to contrast [14], as opposed to direct features of two modalities as in contrastive learning, such that a single image is tolerated to be aligned with multiple captions and concepts. Based on the above formulation, we conduct a systematic study in terms of zero-shot transfer and representation learning. We empirically observe that, while nCLIP demonstrates desirable representation learning capability for both modalities, it underperforms prominently in zero-shot classification and retrieval tasks where models pre-trained with negative samples naturally prevail.", + "bbox": [ + 75, + 231, + 470, + 625 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Seeking for unique strengths of both worlds, we further perform multi-tasking of CLIP and nCLIP, short as xCLIP, and seek synergy between two distinct pre-training objectives. With the bearable add-on of computational resources (e.g., $\\sim 27\\%$ memory-wise and $\\sim 30\\%$ time-wise), xCLIP achieves consistent performance gain compared to CLIP on a wide range of tasks spanning from zero-shot classification, retrieval, linear probing, and fine-tuning, etc. An extensive study with different pre-training datasets, evaluation metrics, and optimization configurations is conducted to validate xCLIP's effectiveness in mining semantics and improving data efficiency. Particularly, the base-size model pre-trained using xCLIP under 35-million publicly available image-text pairs achieves a performance gain of $3.3\\%$ and $1.5\\%$ on an average of 27 classification tasks [59], in terms of zero-shot classification and linear probing accuracy, respectively. Performance gain is also valid in terms of zero-shot retrieval, semi-supervised learning, and fine", + "bbox": [ + 75, + 628, + 468, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tuning, with 3.7 points of R@1 on Flickr30K [58], $1.7\\%$ accuracy on $1\\%$ subset of ImageNet [24], and $0.3\\%$ accuracy on ImageNet, respectively.", + "bbox": [ + 498, + 90, + 890, + 137 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 151, + 640, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Language-image pre-training. Language-image pretraining, aka vision-language pre-training (VLP), learns to jointly cope with vision and language input using multi-modality models. Early practices [22, 49, 67, 68] separately encode modalities, i.e., texts with linear embedding and images with convolutional neural networks and region proposals [63]. Recent works take direct images instead of pre-fetched region features as input, using e.g., dual-encoder architecture for alignment [42, 57, 59], single-encoder architecture for fusion [43, 73], or their combinatorial practices forming an encoder-decoder architecture [48, 80]. Most recent literature aims to expand as many supported transferable downstream tasks as possible via pre-training large-scale foundation models [6, 80]. Out of previous works, CLIP [59] blazes the trail to learn transferable visual features from text supervision, and demonstrates surprising zero-shot recognition results. Follow-up study includes leveraging dense [79], augmentation [47], and uni-modal self-supervised [50, 55, 74] signals for improved performance. In this work, we base our study on the dual-encoder architecture, and pre-train models either with the contrastive objective as CLIP [59], or the non-contrastive objective. Further, we seek their complementarity and study whether synergy between two objectives exists.", + "bbox": [ + 496, + 178, + 890, + 555 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Contrastive learning. The idea of contrastive learning is popularized under visual self-supervised learning (SSL) which aims at learning transferable visual representation from unlabelled images. In common practices [18, 33, 75], each image is considered a single class, attracted to its augmented views while pulled away from views of other images. The contrastive objective is proved robust and effective in a wide range of realms beyond visual understanding, including representation learning for natural language [29], audio [54], structured input [44], robotics [66], and multimodality [21, 59]. However, a caveat of these approaches is the requirement for large batch size [18] or memory bank [33, 75], which lies intrinsic in the formulation of InfoNCE estimator [38] and plagues the pre-training of largesize models due to hardware limitations.", + "bbox": [ + 496, + 575, + 890, + 801 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Non-contrastive learning. Beyond the contrastive paradigm, recent state of the arts from visual SSL manage to relieve the dependency on negative samples. With different subtleties to avert collapsing solutions, these works can optimize the affinity of augmented representations alone", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "11029", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "and are categorized as non-contrastive framework. To avoid model collapsing, common practices include asymmetrical architecture [19, 31], dimension de-correlation [11, 27, 81], and clustering [7, 8, 13–15, 71]. Our approach takes inspiration from the last category since it intrinsically frees the absoluteness of assignment of positive samples. Moreover, rich semantics induced by explicit clustering can automatically group visual concepts from noisy image-text pairs, facilitating models' representation learning.", + "bbox": [ + 75, + 90, + 472, + 227 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Approach", + "text_level": 1, + "bbox": [ + 76, + 241, + 184, + 258 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Framework", + "text_level": 1, + "bbox": [ + 76, + 266, + 205, + 281 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Suppose $\\pmb{g}$ and $\\pmb{h} \\in \\mathbb{R}^{D \\times 1}$ are projected backbone features of two modalities. $D$ is the feature dimension. We start by formulating the objective of dominant contrastive framework CLIP, followed by the introduced non-contrastive framework nCLIP. The final framework xCLIP takes the multi-tasking of two objectives, the complete computational pipeline of which is shown in Algorithm 1.", + "bbox": [ + 75, + 287, + 470, + 396 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Contrastive pre-training [59]. Let $\\pmb{u} = \\pmb{g} / ||\\pmb{g}||$ and $\\pmb{v} = \\pmb{h} / ||\\pmb{h}||$ . $\\pmb{U} = [\\pmb{u}_1, \\pmb{u}_2, \\dots, \\pmb{u}_B] \\in \\mathbb{R}^{D \\times B}$ , $\\pmb{V} = [\\pmb{v}_1, \\pmb{v}_2, \\dots, \\pmb{v}_B] \\in \\mathbb{R}^{D \\times B}$ are concatenated embeddings over the batch. $B$ is the batch size. The contrastive objective is formulated as", + "bbox": [ + 75, + 412, + 468, + 491 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {C L I P}} = \\operatorname {I n f o N C E} (\\boldsymbol {U} ^ {\\mathrm {T}} \\boldsymbol {V} / \\sigma) \\\\ = - \\frac {1}{B} \\sum_ {i \\in B} \\log \\frac {\\exp \\left(\\boldsymbol {u} _ {i} ^ {\\mathrm {T}} \\boldsymbol {v} _ {i} / \\sigma\\right)}{\\sum_ {j \\in B} \\exp \\left(\\boldsymbol {u} _ {i} ^ {\\mathrm {T}} \\boldsymbol {v} _ {j} / \\sigma\\right)}, \\tag {1} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 120, + 500, + 468, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\sigma$ is a trainable parameter controlling the temperature. $\\mathcal{L}_{\\mathrm{CLIP}}$ is symmetrized by setting $g$ and $h$ as projected features of the images and texts by turns and taking the average of two terms. Vanilla InfoNCE can be decoupled into two terms accounting for affinity and variability separately [81]. While both variability terms in the contrastive and non-contrastive objectives require for batch statistics (e.g., $\\mathcal{L}_{\\mathrm{EH}}$ in Eq. (4)), CLIP formulation explicitly maximizes the distance between negative pairs of batch samples. Hence, it degrades the learning performance and data efficiency when models are pre-trained with noisy data where sensible matches occur within negative pairs.", + "bbox": [ + 75, + 577, + 468, + 758 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Non-contrastive pre-training. Let $\\pmb{p} = \\mathrm{softmax}(\\pmb{g})$ and $\\pmb{q} = \\mathrm{softmax}(\\pmb{h})$ . We transform the projected feature into probability distributions, which can be seen as an assignment over semantic clusters (i.e., projection weights). We take $\\pmb{p}$ as the target distribution and learn to estimate the predicted distribution $\\pmb{q}$ by minimizing their cross-entropy:", + "bbox": [ + 75, + 777, + 468, + 869 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {C E}} = - \\boldsymbol {p} ^ {\\mathrm {T}} \\log (\\boldsymbol {q}). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 204, + 878, + 468, + 897 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Note the target and prediction branch are backpropagated at the same time and $\\mathcal{L}_{\\mathrm{CE}}$ is symmetrized as $\\mathcal{L}_{\\mathrm{CLIP}}$ . To avert collapsing solutions, we incorporate entropy regularizers, i.e., entropy minimization [71] and mean entropy maximization [8,9], via minimizing:", + "bbox": [ + 496, + 90, + 890, + 167 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {E H}} = - \\boldsymbol {p} ^ {\\mathrm {T}} \\log (\\boldsymbol {p}), - \\mathcal {L} _ {\\mathrm {H E}} = \\bar {\\boldsymbol {p}} ^ {\\mathrm {T}} \\log (\\bar {\\boldsymbol {p}}), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 176, + 890, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\overline{\\boldsymbol{p}} = \\mathbb{E}(p) \\approx \\frac{1}{B} \\sum_{i=1}^{B} \\boldsymbol{p}_i$ is the average distribution over batch. $\\mathcal{L}_{\\mathrm{EH}}$ and $\\mathcal{L}_{\\mathrm{HE}}$ are symmetrized by taking the average with another term where $\\boldsymbol{p}$ is replaced by $\\boldsymbol{q}$ . $\\mathcal{L}_{\\mathrm{EH}}$ encourages the model to make deterministic predictions and ensures assignment's sharpness. $\\mathcal{L}_{\\mathrm{HE}}$ encourages the model to utilize a full set of projection weights and ensures the assignment's smoothness. We illustrate in Sec. 4.4 and Appendix A that $\\mathcal{L}_{\\mathrm{EH}}$ and $\\mathcal{L}_{\\mathrm{HE}}$ are minimally sufficient to yield non-collapsing solutions. We empirically show in Tab. A1 that the solutions is non-trivial as long as both sharpness and smoothness are guaranteed, e.g., via Sinkhorn algorithm [14] instead of loss regularizers. We opt for Eq. (3) for its simplicity and consistency between pretraining and evaluation. The overall non-contrastive objective is formulated as", + "bbox": [ + 496, + 208, + 892, + 436 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {n C L I P}} = \\mathcal {L} _ {\\mathrm {C E}} + \\lambda_ {1} \\cdot \\mathcal {L} _ {\\mathrm {E H}} - \\lambda_ {2} \\cdot \\mathcal {L} _ {\\mathrm {H E}}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 449, + 890, + 465 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\lambda_{1}$ and $\\lambda_{2}$ controls the weight of the regularization. We empirically find that setting $\\lambda_{1} = \\lambda_{2} - 1 = 0.5$ yields stable training and favorable transfer performances. We detail our findings in Sec. 4.4 and Appendix B.1. During evaluation, we utilize negative cross-entropy $-\\mathcal{L}_{\\mathrm{CE}}$ as the similarity metric for zero-shot transfer experiments. As schematically showcased in Fig. 1, nCLIP does not reply on negative examples but requires a relatively larger projection head and greater output dimension. Comparing nCLIP with CLIP, we empirically find that nCLIP produces more coarse-grained (e.g., zero-shot retrieval in Sec. 4.1) but semantic-rich (e.g., linear probing in Sec. 4.2) projections. Empirical evidence suggests that the model will fail to deliver reasonable zero-shot and retrieval results if pre-trained without negative pairs (e.g., nCLIP's 25.2 vs. CLIP's 73.8 points of R@1 under Flickr30K I→T retrieval as shown in Tab. 4).", + "bbox": [ + 496, + 478, + 892, + 734 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Contrastive meets non-contrastive. Given that two objectives each have their own limitations, we further seek the complementarity between $\\mathcal{L}_{\\mathrm{CLIP}}$ and $\\mathcal{L}_{\\mathrm{nCLIP}}$ , and pre-train the models with both objectives, written as", + "bbox": [ + 496, + 753, + 890, + 814 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {x C L I P}} = \\lambda_ {\\mathrm {C L I P}} \\cdot \\mathcal {L} _ {\\mathrm {C L I P}} + \\lambda_ {\\mathrm {n C L I P}} \\cdot \\mathcal {L} _ {\\mathrm {n C L I P}}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 827, + 890, + 843 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\lambda_{\\mathrm{CLIP}}$ and $\\lambda_{\\mathrm{nCLIP}}$ control the weight of the objectives. During evaluation, we find using the negative cosine metric with CLIP's projection head for zero-shot transfer", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "11030", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "code", + "sub_type": "code", + "code_caption": [ + "Algorithm 1: PyTorch-like pseudo-code of xCLIP." + ], + "code_body": "// enc, proj: encoder & projector", + "guess_lang": "typescript", + "bbox": [ + 98, + 118, + 367, + 132 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "code", + "code_caption": [], + "code_body": "$\\pmb{f}_I,\\pmb {f}_T = \\mathrm{enc}(\\mathrm{img}),\\mathrm{enc}(\\mathrm{text})$", + "guess_lang": "latex", + "bbox": [ + 98, + 133, + 290, + 147 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "code", + "code_caption": [], + "code_body": "$\\pmb{g}_I, \\pmb{g}_T = \\mathrm{proj}_{\\mathrm{CLIP}}(\\pmb{f}_I), \\mathrm{proj}_{\\mathrm{CLIP}}(\\pmb{f}_T)$", + "guess_lang": "latex", + "bbox": [ + 98, + 148, + 351, + 162 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "code", + "code_caption": [], + "code_body": "$h_{I},h_{T} = \\mathrm{proj}_{\\mathrm{nCLIP}}(\\pmb {f}_{I}),\\mathrm{proj}_{\\mathrm{nCLIP}}(\\pmb{f}_{T})$", + "guess_lang": "latex", + "bbox": [ + 98, + 162, + 367, + 178 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "code", + "code_caption": [], + "code_body": "$\\mathcal{L}_{\\mathrm{xCLIP}} = \\lambda_{\\mathrm{CLIP}}\\cdot \\mathcal{L}_{\\mathrm{CLIP}}(\\pmb {g}_I,\\pmb {g}_T,\\sigma)$", + "guess_lang": "latex", + "bbox": [ + 99, + 179, + 339, + 193 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "code", + "code_caption": [], + "code_body": "+λnCLIP·LnCLIP(hH,T,λ1,λ2)", + "guess_lang": "csv", + "bbox": [ + 156, + 194, + 387, + 208 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "code", + "code_caption": [], + "code_body": "return $\\mathcal{L}_{\\mathrm{xCLIP}}$", + "guess_lang": "txt", + "bbox": [ + 98, + 209, + 197, + 223 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "code", + "code_caption": [], + "code_body": "def $\\mathcal{L}_{\\mathrm{CLIP}}(\\pmb {f}_I,\\pmb {f}_T,\\sigma)$ ..", + "guess_lang": "txt", + "bbox": [ + 98, + 228, + 256, + 243 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "code", + "code_caption": [], + "code_body": "// INCE: InfoNCE with Eq. (1)", + "guess_lang": "txt", + "bbox": [ + 122, + 244, + 349, + 256 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "code", + "code_caption": [], + "code_body": "$\\pmb{u}_{I}$ = normalize $(f_{I}, p = 2, \\dim = 1)$", + "guess_lang": "latex", + "bbox": [ + 122, + 257, + 339, + 271 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "code", + "code_caption": [], + "code_body": "$\\pmb{u}_{T} = \\mathrm{normalize}(\\pmb{f}_{T}, \\mathrm{p} = 2, \\mathrm{dim} = 1)$", + "guess_lang": "txt", + "bbox": [ + 122, + 272, + 344, + 286 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "code", + "code_caption": [], + "code_body": "$\\mathcal{L} = \\mathrm{INCE}(\\pmb{U}_I^{\\mathrm{T}}\\pmb{U}_T / \\sigma) + \\mathrm{INCE}(\\pmb{U}_T^{\\mathrm{T}}\\pmb{U}_I / \\sigma)$", + "guess_lang": "latex", + "bbox": [ + 122, + 286, + 406, + 301 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "code", + "code_caption": [], + "code_body": "return $\\mathcal{L} / 2$", + "guess_lang": "txt", + "bbox": [ + 124, + 303, + 207, + 315 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "code", + "code_caption": [], + "code_body": "def $\\mathcal{L}_{\\mathrm{nCLIP}}(f_I, f_T, \\lambda_1, \\lambda_2)$ :", + "guess_lang": "latex", + "bbox": [ + 98, + 325, + 295, + 339 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "code", + "code_caption": [], + "code_body": "$\\pmb{p}_I = \\mathrm{softmax}(\\pmb {f}_I,\\mathrm{dim} = 1)$", + "guess_lang": "latex", + "bbox": [ + 122, + 340, + 292, + 354 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "code", + "code_caption": [], + "code_body": "$\\pmb{p}_T = \\mathrm{softmax}(\\pmb {f}_T,\\mathrm{dim} = 1)$", + "guess_lang": "latex", + "bbox": [ + 122, + 354, + 297, + 369 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "code", + "code_caption": [], + "code_body": "$\\mathcal{L}_{\\mathrm{CE}} = -\\left(\\pmb {p}_I\\cdot \\log (\\pmb {p}_T) + \\pmb {p}_T\\cdot \\log (\\pmb {p}_I)\\right).$ sum $(\\mathrm{dim} = 1)$", + "guess_lang": "latex", + "bbox": [ + 122, + 369, + 450, + 383 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "code", + "code_caption": [], + "code_body": ".mean(dim=0)", + "guess_lang": "txt", + "bbox": [ + 173, + 385, + 267, + 397 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "code", + "code_caption": [], + "code_body": "$\\mathcal{L}_{\\mathrm{EH}} = -(\\pmb {p}_I\\cdot \\log (\\pmb {p}_I) + \\pmb {p}_T\\cdot \\log (\\pmb {p}_T)).$ sum $(\\dim = 1)$", + "guess_lang": "latex", + "bbox": [ + 122, + 398, + 450, + 414 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "code", + "code_caption": [], + "code_body": ".mean(dim=0)", + "guess_lang": "txt", + "bbox": [ + 173, + 415, + 267, + 428 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "code", + "code_caption": [], + "code_body": "$\\overline{\\pmb{p}}_I,\\overline{\\pmb{p}}_T = \\pmb {p}_I.\\mathrm{mean}(\\mathrm{dim} = 0),\\pmb {p}_T.\\mathrm{mean}(\\mathrm{dim} = 0)$", + "guess_lang": "latex", + "bbox": [ + 122, + 430, + 419, + 445 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "code", + "code_caption": [], + "code_body": "$\\mathcal{L}_{\\mathrm{HE}} = -\\left(\\overline{\\boldsymbol{p}}_I\\cdot \\log (\\overline{\\boldsymbol{p}}_I) + \\overline{\\boldsymbol{p}}_T\\cdot \\log (\\overline{\\boldsymbol{p}}_T)\\right).$ sum(dim=1)", + "guess_lang": "latex", + "bbox": [ + 122, + 445, + 450, + 460 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "code", + "code_caption": [], + "code_body": "$\\mathcal{L} = \\mathcal{L}_{\\mathrm{CE}} + \\lambda_1\\cdot \\mathcal{L}_{\\mathrm{EH}} - \\lambda_2\\cdot \\mathcal{L}_{\\mathrm{HE}}$", + "guess_lang": "latex", + "bbox": [ + 122, + 460, + 333, + 474 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "code", + "code_caption": [], + "code_body": "return $\\mathcal{L} / 2$", + "guess_lang": "txt", + "bbox": [ + 122, + 474, + 207, + 488 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "experiments generally leads to better results. We note that it is not immediately clear that the two objectives certainly induce stronger models, since they build qualitatively distinctive latent spaces. Qualitatively, we show in Sec. 4 that nCLIP helps CLIP to encode semantics which intrinsically lacks in CLIP, while CLIP helps nCLIP to be transferable for zero-shot recognition. The synergy between the two objectives prompts consistent performance gains across a wide range of tasks by xCLIP over CLIP.", + "bbox": [ + 75, + 506, + 468, + 642 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Implementation", + "text_level": 1, + "bbox": [ + 76, + 655, + 235, + 671 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Architecture. We train the base-size model with ViTB/16 [26] as the visual encoder. The model configuration of the base-size text encoder follows CLIP [59] with byte-pair encoding (BPE) and a maximum context length of 77. The projection head generating output for the noncontrastive objective is a 2-layer MLP with 4096-dim hidden layers, GELU [36], and BatchNorm [41]. The last layer is of 32768 output dimension and followed by a BatchNorm without affine transformation. The projection head for the contrastive objective is a single linear layer with no bias and a dimension of 512.", + "bbox": [ + 75, + 679, + 468, + 844 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Pre-training data. We train our method with publicly available datasets: COCO [51], Visual Genome [45], SBU", + "bbox": [ + 75, + 869, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Captions [56], Conceptual Caption 3M [65], Conceptual Caption 12M [17], and filtered 14M-size subset of Yahoo Flickr Creative Commons 100M dataset, consisting a total of 35M Image-Text pairs (IT35M). We are also intrigued about the behaviors of models pre-trained on ImagNet-21K [24] (IN21K) dataset taking label names as annotation texts by concatenating them with a sampled prompt, similar to [57, 80]. To notify, ImageNet-21K is a subset of ImageNet-22K with classes of ImageNet-1K excluded for fair downstream evaluation. We detail the models' data scaling behavior in Tab. 8.", + "bbox": [ + 496, + 90, + 890, + 257 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Optimization. We use a batch size of 4096, with all data distributed across 32 V100 GPUs. Models are trained by default with AdamW [53] optimizer, a peak learning rate of $1e^{-3}$ , cosine scheduler, a weight decay of 0.2, a $\\beta_{2}$ for AdamW of 0.98, a $\\epsilon$ for AdamW of $1e^{-6}$ , and automatic mixed-precision for 3 warm-up epochs and a total of 32 epochs. $\\lambda_{1}$ and $\\lambda_{2}$ are set as 0.5 and 1.5, respectively. The two objectives are multi-tasked with $\\lambda_{\\mathrm{nCLIP}}$ being 1 and $\\lambda_{\\mathrm{CLIP}}$ being 0.2. Augmentations and pre-processing of pre-training data follow CLIP with image size randomly cropped and resized within the scale of (0.5, 1.0). Detailed pre-training hyper-parameters are detailed in Appendix F.", + "bbox": [ + 496, + 276, + 890, + 458 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 470, + 632, + 488 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Zero-Shot Transfer", + "text_level": 1, + "bbox": [ + 500, + 496, + 684, + 510 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Classification. We evaluate 27 classification benchmarks with zero-shot classification protocols following [59]. As shown in Tab. 1, nCLIP achieves an average top-1 accuracy of $32.7\\%$ . Though performing $7.6\\%$ worse than CLIP, nCLIP demonstrates descent zero-shot recognition capability without explicitly training with negative examples. xCLIP achieves consistent performance gain across a wide range of datasets, with a $43.6\\%$ average top-1 accuracy, which is $3.3\\%$ higher than CLIP. It indicates that the synergy between contrastive and non-contrastive objectives improves the model's zero-shot classification ability.", + "bbox": [ + 496, + 518, + 890, + 686 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Out-of-domain classification. For out-of-distribution classification, we use 5 datasets following [59]: ImageNet Adversarial [37], ImageNet Rendition [35], ImageNetV2 [62], ImageNet Sketch [72], and ObjectNet [10] with 2 additional datasets: ImageNet-C [76] and Stylized ImageNet [30]. We investigate if the non-contrastive objective inherits the robustness of the contrastive objective to natural distribution shift. As shown in Tab. 2, nCLIP achieves an $18.0\\%$ with CLIP an $23.1\\%$ average top-1 accuracy, revealing a similar trend to in-domain datasets. xCLIP further improves CLIP's accuracy by $3.6\\%$ and achieves an accuracy of $26.7\\%$ across 7 datasets, indicating its effectiveness among out-of-domain datasets.", + "bbox": [ + 496, + 704, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "11031", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/71f73f15e1ad2444795f6c158743ed05fe9700982f3f138686813fc9949abd20.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelFood101CIFAR10CIFAR100BirdmapSUN397CarsAircraftVOC2007DTDPetsCaltech101FlowersMNISTFER2013STL10EuroSATRESISC45GTSRBKITTICountry211PCAMUCF101Kinetics700CLEVRHapticMemesSSTImageNetAverage
CLIP61.279.750.623.756.515.95.846.427.654.771.348.910.537.391.324.539.713.131.69.150.045.032.412.853.049.145.740.3
nCLIP28.479.549.111.357.05.94.551.422.914.665.023.19.913.594.815.121.22.735.45.851.242.028.412.452.750.037.032.7
xCLIP65.883.454.525.159.918.05.852.233.257.173.950.012.339.092.840.043.616.339.89.351.149.835.418.452.550.248.843.6
", + "bbox": [ + 80, + 88, + 888, + 176 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/f14ef0a54a8c331c53fc38bc467b741adf98a852090abf4d9694e039acb38f49.jpg", + "table_caption": [ + "Table 1. Zero-shot classification. We report on a variety of classification benchmarks with ViT-B/16 pre-trained on IT35M. Detailed protocols for each dataset strictly follow CLIP [59]. xCLIP achieves a consistent performance gain compared to CLIP in a wide range of classification datasets. Best results of each column are bolded." + ], + "table_footnote": [], + "table_body": "
ModelIN-AIN-RIN-v2IN-SkeIN-StyIN-CObjNetAverage
CLIP21.343.737.926.54.911.915.523.1
nCLIP21.529.030.817.73.611.512.018.0
xCLIP27.749.140.630.66.114.717.826.7
", + "bbox": [ + 78, + 244, + 467, + 340 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/7b1c37cb1bcf3adc3b9e2026ee31c35ff4cacc0875d214d3406145f4b84296ac.jpg", + "table_caption": [ + "Table 2. Zero-shot out-of-distribution classification. We report on a variety of out-of-distribution classification benchmarks. xCLIP demonstrates stronger robustness on various out-of-domain classification datasets." + ], + "table_footnote": [], + "table_body": "
ModelNUS-WIDEOpenImages
mAPF1@3F1@5mAPF1@10F1@20
CLIP15.133.316.081.113.47.2
nCLIP16.135.616.681.411.66.4
xCLIP15.335.216.881.213.87.4
", + "bbox": [ + 80, + 417, + 460, + 508 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 3. Zero-shot multi-label classification. We report mAP and F1 scores with ViT-B/16 pre-trained on IT35M. nCLIP showcases the best multi-label classification capability.", + "bbox": [ + 76, + 512, + 468, + 555 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Multi-label classification. We evaluate zero-shot multi-label classification on NUS-WIDE [23] and OpenImages [46] following standard protocol [39]. Specifically, we use 81 unseen labels for NUS-WIDE and the most frequent 400 unseen test labels for OpenImages during evaluation, following [40]. As shown in Tab. 3, we observe that nCLIP shows the best mAP compared to CLIP. nCLIP is pre-trained without negative examples, resulting in the model's intrinsic strengths on recall over overlapped concepts, with 16.1 and 81.4 points of mAP on NUS-WIDE and OpenImages, respectively. Comparatively, xCLIP slightly improves CLIP in this respect while lagging behind nCLIP. The experiments demonstrate that the non-contrastive objective fits well for those downstream tasks in demand of high recall.", + "bbox": [ + 75, + 561, + 470, + 789 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Retrieval. We evaluate on 2 retrieval benchmarks: Flickr30K [58] and MSCOCO [51] under zero-shot protocol. We do not use prompt engineering and use the original caption. Compared to zero-shot classification, image-text retrieval requires the model's recognition capability at a fine-grained level. We empirically observe in Tab. 4 that", + "bbox": [ + 75, + 809, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "nCLIP performs drastically worse than CLIP, which is because text captions in retrieval are not mutually exclusive, which is different from label names in classification. Therefore, explicit negative examples during pre-training play an imperative role in zero-shot retrieval. See Appendix E.1 for further discussions. Beyond that, xCLIP achieves a noticeable performance improvement over CLIP, with a gain of $3.7\\%$ R@1 on Flickr30K and $3.9\\%$ R@1 on MSCOCO, illustrating that the additional non-contrastive objective renders extra semantic signals that can be transferred well to fine-grained recognition.", + "bbox": [ + 496, + 250, + 893, + 416 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Representation Learning", + "text_level": 1, + "bbox": [ + 500, + 424, + 728, + 441 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We conduct several evaluation protocols to benchmark representation quality under different objectives for both visual (Sec. 4.2.1) and textual (Sec. 4.2.2) modality.", + "bbox": [ + 496, + 446, + 893, + 494 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2.1 Visual Representation Learning", + "text_level": 1, + "bbox": [ + 498, + 511, + 777, + 527 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Linear probing. We evaluate the quality of visual representation via linear probing protocol, where a linear head is fine-tuned on top of the frozen backbones. We follow the same setup as [20] on the same 27 datasets as standard zero-shot classification. Specifically, we use SGD without momentum as the optimizer, no weight decay, and a total epoch of 100 for all evaluation datasets. We use [CLS] token for classification. Following [82], we sweep a set of different learning rates by adding multiple classification heads over a shared frozen backbone, each with its own optimizer and scheduler. We report the best result across different heads. As shown in Tab. 5, we observe that nCLIP generally leads to comparable linear probing accuracy compared with CLIP, suggesting that non-contrastive objectives are able to derive semantically meaningful embedding spaces. nCLIP achieves $2.5\\%$ higher performance compared to CLIP on ImageNet but relatively lower on average. Beyond that, when combining two objectives together, xCLIP achieves $2.7\\%$ higher performance on ImageNet and $1.5\\%$ higher performance on average across 27 different classification tasks compared to CLIP.", + "bbox": [ + 496, + 534, + 893, + 852 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Fine-tuning & semi-supervised learning. We fine-tune the entire network under the full-data regime (100%) and", + "bbox": [ + 498, + 869, + 893, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "11032", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/62a3ca3f9d758302f4e615ca24563f21726ad224f6a55a04677ab6f8089b29e8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelFlickr30KMSCOCO
I→TT→II→TT→I
R@1R@5R@10R@1R@5R@10R@1R@5R@10R@1R@5R@10
CLIP73.893.397.152.979.387.759.485.291.934.362.774.0
nCLIP26.253.367.920.148.561.121.447.961.613.437.150.3
xCLIP77.595.697.757.382.989.063.387.594.138.466.076.7
", + "bbox": [ + 80, + 88, + 885, + 200 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e67cc5f451ddda81e9a5574cd221547b83fd6d53516bd47fc11710726f94dd36.jpg", + "table_caption": [ + "Table 4. Zero-shot image-to-text retrieval. We report R@1, R@5, and R@10 in both image-to-text (I→T) and text-to-image (T→I) settings with ViT-B/16 pre-trained on IT35M. xCLIP consistently improves CLIP in retrieval." + ], + "table_footnote": [ + "CLIP 83.3 92.0 75.6 54.4 78.8 60.5 33.8 91.1 72.4 79.0 90.0 96.1 96.2 56.5 97.1 94.8 90.3 77.3 69.7 23.2 83.4 78.9 53.9 50.3 55.9 55.8 71.4 |72.7 nCLIP 83.8 93.2 76.3 66.8 79.3 49.3 28.0 90.1 70.6 70.2 90.5 95.0 95.4 54.7 97.4 91.9 88.2 75.7 64.0 22.7 84.2 79.9 51.0 44.6 55.6 55.2 73.9 |71.4 xCLIP 85.3 93.4 77.8 58.4 80.7 62.3 36.7 92.3 74.0 81.5 91.6 97.0 96.8 58.5 98.1 95.3 91.2 80.4 69.3 26.3 83.5 81.1 56.0 49.8 58.5 54.8 74.1 |74.2" + ], + "table_body": "
ModelFood101CIFAR10CIFAR100BirdsnapSUN397CarsAircraftVOC2007DTDPetsCaltech101FlowersMNISTFER2013STL10EuroSATRESISC45GTSRBKITTICountry211PCAMUCF101Kinetics,700CLEVRHaarSoftMnesSSTImageNetAverage
", + "bbox": [ + 81, + 244, + 885, + 297 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/be7ad7fcd7f783a1b487d30fb50a19d47118190062ef8a525897e74e1dc558f1.jpg", + "table_caption": [ + "Table 5. Linear probing. We report on a variety of classification benchmarks with ViT-B/16 pre-trained on IT35M. CLIP achieves consistent performance gains compared to CLIP in a wide range of classification datasets." + ], + "table_footnote": [], + "table_body": "
Modelsm. sup.ft.
1%10%100%
CLIP57.872.482.4
nCLIP55.072.282.4
xCLIP59.573.482.7
", + "bbox": [ + 84, + 393, + 455, + 483 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/3451df9e99b6703712dc756a3b685176edcda3b561f04652455af8c318ede033.jpg", + "table_caption": [ + "Table 6. Fine-tuning (ft.) and semi-supervised learning (sm. sup.). Percentage shows sampling ratio for end-to-end fine-tuning. We report top-1 accuracy on the ImageNet-1K validation set. nCLIP performs comparatively with CLIP. xCLIP induces consistent gains in all sampling ratios, especially for semi-supervised learning." + ], + "table_footnote": [], + "table_body": "
ModelSTS12STS13STS14STS15STS16STS-BSICK-RAverage
CLIP52.354.053.969.958.667.069.360.7
nCLIP51.362.756.971.965.866.964.162.8
xCLIP48.755.655.669.759.566.370.260.8
", + "bbox": [ + 81, + 587, + 460, + 681 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 7. Sentence embedding performance on semantic textual similarity. We report on a variety of NLP understanding benchmarks. xCLIP demonstrates the best capability for various STS tasks. xCLIP has comparable performance with CLIP.", + "bbox": [ + 75, + 685, + 468, + 742 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the semi-supervised learning protocol (1% and 10%) on ImageNet-1K [24]. Results are shown in Tab. 6. For CLIP and xCLIP, we find that fine-tuning with the projection head trained under the contrastive objective yields consistent performance gain, especially under a low-data regime. Specifically, we take the text features of the label names as the classifier's initialized parameters. We opt for this setup by default. We highlight that while nCLIP performs worse than CLIP in terms of zero-shot transfer, the non-contrastive objective also serves as a strong baseline to learn represent", + "bbox": [ + 75, + 750, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tation considering the close gap between CLIP and nCLIP in terms of fine-tuning accuracy. xCLIP achieves an $1.7\\%$ , $1.0\\%$ , and $0.3\\%$ performance gain compared with CLIP under three data ratios of $1\\%$ , $10\\%$ , and $100\\%$ , respectively.", + "bbox": [ + 498, + 393, + 892, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Mask probing. To evaluate how well the visual model is capable of deriving explicit scene layout and object boundaries, we conduct mask probing analysis following [15]. For each attention head from the last layer, we extract the attention map with [CLS] token as the query. We then compute the Jaccard similarity $\\mathcal{J}$ of each head's attention mask to the ground truth and retain the attention mask with the highest similarity. We conduct experiments on Pascal VOC 2012 [28] dataset. With IT35M, nCLIP demonstrates better quality in the generated mask with 43.7 points of $\\mathcal{J}$ , while CLIP reaches 41.2 and xCLIP reaches 41.9 points of $\\mathcal{J}$ . Hence, nCLIP learns better representation for object boundaries compared to CLIP, while xCLIP strikes a balance between nCLIP and CLIP. Detailed results are delayed to Appendix D.1. As a related evaluation, we also study unsupervised segmentation with GroupViT [77] in Appendix D.2.", + "bbox": [ + 496, + 476, + 892, + 719 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.2 Textual Representation Learning", + "text_level": 1, + "bbox": [ + 500, + 739, + 784, + 756 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We follow the setup as [29] and evaluate on 7 STS tasks: STS 2012-2016 [1-5], STS Benchmark [16], and SICK-Relatedness [34]. We directly take [EOS] token without any projection as the input for evaluation, which consistently yields better performance compared to the projected feature for all models. We use cosine distance as the similarity metric. The main goal of sentence embeddings is to cluster semantically similar sentences, and hence, we take STS as one yardstick to benchmark textual representation", + "bbox": [ + 496, + 763, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "11033", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a925e383d2dd8d3f86388e86842db8bb754b99661953b0267001752ccbef842a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DataSizeEp.ZS / LN Accuracy
CLIPnCLIPxCLIP
①12M2536.8 / 68.537.5 / 71.042.4 / 72.2
④14M3226.5 / 75.131.4 / 77.027.1 / 77.2
①②22M3237.9 / 68.234.2 / 72.343.0 / 71.9
①②③35M3245.7 / 71.437.0 / 73.948.8 / 74.1
", + "bbox": [ + 83, + 89, + 460, + 195 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 8. Pre-training data scaling & domain. Ep. denotes training epochs. Data abbreviations are as follows. ①: CC12M. ②: COCO, VG, SBU, and CC3M. ③: YFCC14M. ④: IN21K. Note ①②③ = IT35M.", + "bbox": [ + 76, + 198, + 468, + 255 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d8a2d85afcc2806e3dbf4fcdfa85ee47b8515188eb85100155e54c107714a0d5.jpg", + "image_caption": [ + "Figure 2. Batch size scaling. xCLIP performs better than CLIP with a small batch size (i.e., 1024)." + ], + "image_footnote": [], + "bbox": [ + 86, + 275, + 267, + 414 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/02bb5660dade19153329b4fcde68549a498f3f1a2e08a807bbe25348d5c3a5cc.jpg", + "image_caption": [ + "Figure 3. Training time. xCLIP adds $\\sim 1.3x$ training time cost and performs better with equal seen images." + ], + "image_footnote": [], + "bbox": [ + 281, + 275, + 460, + 414 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "learning. Results are shown in Tab. 7. nCLIP shows better textual semantics compared to CLIP, with $2.1\\%$ higher in terms of the average score with IT35M. The conclusion can also be visually extrapolated from the t-SNE visualization of the word embedding, as detailed in Appendix G. We note that xCLIP performs comparatively with CLIP given that STS is evaluated on [EOS] without projection, thus the backbone representation can be similar if dominant by the contrastive objective.", + "bbox": [ + 75, + 481, + 468, + 618 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Properties", + "text_level": 1, + "bbox": [ + 76, + 628, + 192, + 643 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We analyze several crucial properties to demonstrate the generalizability and effectiveness of our method in practical usage. Experiments are conducted with ViT-B/16 on CC12M for 25 epochs by default.", + "bbox": [ + 75, + 652, + 468, + 713 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Pre-training data scaling. We study the model's performance with data scaling in Tab. 8. xCLIP exhibits desirable scaling behaviors similar to CLIP. The performance improves as data increases. However, we do not observe such a tendency in nCLIP in terms of zero-shot accuracy. For example, nCLIP achieves $37.5\\%$ with (1) but only $34.2\\%$ with (2), the trend of which defies those of CLIP's $(36.8\\%$ vs. $37.9\\%)$ and xCLIP's $(42.4\\%$ vs. $43.0\\%)$ . On an important note, both nCLIP and xCLIP showcase desirable scaling behaviors under the linear probing protocol. We hypothesize that the performance of zero-shot classification depends on", + "bbox": [ + 75, + 734, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "pre-training data quality with the non-contrastive objective, while strong representation can always be ensured before the projection.", + "bbox": [ + 498, + 90, + 890, + 137 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Pre-training with tagging data. We consider using tagging data, e.g., ImageNet-21K [24], as pre-training data to validate method's generalizability. To do that, we take label names as annotation texts by concatenating them with a sampled prompt, similar to [57, 80]. As shown in Tab. 8, nCLIP shows superiority to CLIP when pre-trained with IN21K by performance gain of $4.9\\%$ and $1.9\\%$ in terms of zero-shot and linear probing accuracy, respectively. Compared to nCLIP, we instead notice a performance drop of $4.3\\%$ for zero-shot accuracy when two objectives multitasked, indicating that the contrastive objective stymies the quality of projected probability under tagging data. Note, that an $77.2\\%$ linear probing accuracy achieved by xCLIP, $2.1\\%$ higher than CLIP, is decently strong compared to $79.0\\%$ achieved by the state of the art in SSL, iBOT [82], pre-trained using the same data with 80 epochs.", + "bbox": [ + 496, + 157, + 892, + 398 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Batch size scaling. We investigate the scaling behavior of CLIP and xCLIP with different pre-training batch sizes. As shown in Fig. 2, xCLIP performs stronger with smaller batch sizes than CLIP pre-trained even with larger batch sizes. For example, xCLIP pre-trained with a batch size of 1024 achieves a zero-shot accuracy of $38.5\\%$ (vs. CLIP's $36.8\\%$ with a batch size of 4096). However, we still observe a substantial performance drop as batch size decreases, the behavior of which is different from self-supervised models pre-trained with the non-contrastive objective [14, 15] that are insensitive to batch size. We hypothesize that a large batch size is necessary to approximate $\\mathbb{E}(p)$ with $\\frac{1}{B}\\sum_{i=1}^{B} p_i$ in estimating $\\mathcal{L}_{\\mathrm{HE}}$ , especially when two probabilities are derived from drastically different modalities.", + "bbox": [ + 496, + 420, + 892, + 633 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Computation efficiency. To demonstrate methods' computation efficiency, we show the FLOPs and GPU memory consumption of xCLIP compared to CLIP. xCLIP enforces bearable additional computation cost on top of CLIP with only $1.4\\%$ extra FLOPs and $27\\%$ extra GPU memory. As shown in Fig. 3, xCLIP adds $\\sim 1.3x$ extra time cost but with a $6.0\\%$ performance gain in terms of zero-shot accuracy.", + "bbox": [ + 496, + 654, + 890, + 760 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 770, + 653, + 787 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We show in this section the crucial composing factors of nCLIP and xCLIP. The ablations in the first column (Tabs. 9a to 9c) are conducted with $\\mathcal{L}_{\\mathrm{nCLIP}}$ only. The ablations in the second column (Tabs. 9d to 9f) are conducted with full loss $\\mathcal{L}_{\\mathrm{xCLIP}}$ . Experiments are conducted with ViTB/16 on CC12M for 25 epochs. The default settings are highlighted in cyan.", + "bbox": [ + 496, + 794, + 890, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "11034", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d33b3f35ca89a5998e0025b3f1c35156ef0d40f75d3e5e9673188151deb899bd.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
λ1λ2ZSLN
01NaN
0.3126.869.5
0.5123.767.2
0.21.235.369.0
0.51.537.571.0
", + "bbox": [ + 106, + 87, + 321, + 188 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/6784f194f71fb623d85ab9cce5dc08e81d3b6d9591222d91ef1b2eb50836eae1.jpg", + "table_caption": [ + "(a) Coefficient of entropy regularizer. $\\lambda_1 + 1 = \\lambda_2$ generally leads to good performance." + ], + "table_footnote": [], + "table_body": "
#SL / #TLZSLN
0 / 142.472.2
0 / 241.371.3
1 / 242.471.9
2 / 342.572.3
", + "bbox": [ + 102, + 219, + 323, + 304 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/2939b5a8a5d996f74a8bfbd6ea568b4b5d21430afa23459bf59a18d2badd2540.jpg", + "table_caption": [ + "(d) Shared projection layers. #SL denotes shared hidden layers. #TL denotes total hidden layers." + ], + "table_footnote": [], + "table_body": "
dimZSLNMem
819231.070.61.07×
1638435.170.91.13×
3276837.571.01.27×
6553635.971.01.52×
", + "bbox": [ + 372, + 88, + 589, + 172 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/79a3ae0ecfb7b75e00761d10de9d94190eb1ef1e4e458e918fabbd87476ee93b.jpg", + "table_caption": [ + "(b) Projection dimension. Mem (G) is compared to CLIP with a 512-dim embedding layer." + ], + "table_footnote": [], + "table_body": "
λCLIP: λnCLIPZSLN
1.040.266.5
0.541.770.3
0.242.472.2
0.141.872.6
", + "bbox": [ + 372, + 219, + 594, + 304 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b59d32ba45eda4eebd5e131c8c3662b76500e0369527138d353367002b2fa7a8.jpg", + "table_caption": [ + "(e) Loss ratio between $\\mathcal{L}_{\\mathrm{CLIP}}$ and $\\mathcal{L}_{\\mathrm{nCLIP}}$" + ], + "table_footnote": [], + "table_body": "
archZSLN
vanilla37.571.0
w/ bottleneck layer31.269.0
w/ l2-norm layer35.470.7
w/o last BN layer34.369.9
", + "bbox": [ + 632, + 88, + 875, + 172 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/d487fdea8d1f3b7517a7b427afd88ea3558cbb4950ecf56fa8a7157693a8de3e.jpg", + "table_caption": [ + "(c) Projection architecture of non-contrastive branch." + ], + "table_footnote": [], + "table_body": "
techniqueZSLN
vanilla42.472.2
shared space (Eq. (3))29.468.9
debiased sampling42.272.4
warm-up on LnCLIP41.771.9
", + "bbox": [ + 625, + 219, + 885, + 304 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "(f) Training objective and technique for multitasking.", + "bbox": [ + 625, + 306, + 883, + 330 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 9. Ablation study with ViT-B/16 on ImageNet-1K validation set. We report zero-shot (ZS) and linear probing (LN) accuracy (\\%).", + "bbox": [ + 80, + 333, + 885, + 348 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Entropy regularizer. We study the effects of coefficient for entropy regularizers $\\lambda_{1}$ and $\\lambda_{2}$ in Tab. 9a. The pretraining is unstable without regularizers or with only maximization of the entropy of the mean $(\\mathcal{L}_{\\mathrm{HE}})$ . Specifically, the model will collapse to constant uniform distribution and outputs the same uniform probability distribution despite different inputs. The additional minimization of the mean of the entropy $(\\mathcal{L}_{\\mathrm{EH}})$ stabilizes training, but incurs dimensional collapse, eroding the performances (26.8% and 23.7% vs. 37.5% in terms of zero-shot accuracy). Simultaneously adjusting $\\lambda_{2}$ with constraints $\\lambda_{1} + 1 = \\lambda_{2}$ and $\\lambda_{1} = 0.5$ mitigates the problem and yields the optimal performance, with a 37.5% zero-shot and 71.0% linear-probing accuracy. Details are shown in Appendix B.1.", + "bbox": [ + 75, + 375, + 472, + 588 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Projection head. We study the design of two projection heads for $\\mathcal{L}_{\\mathrm{CLIP}}$ and $\\mathcal{L}_{\\mathrm{nCLIP}}$ . As shown in Tab. 9b, large projection dimension matters for good performance. We opt for a projection dimension of 32768 balancing the performance and speed. As shown in Tab. 9c, the last BN layer is crucial and leads to more stable optimization in our experiments. We draw on the idea of prototypes [9, 15] by introducing bottleneck and 12-norm layers, while they do not yield a performance gain. We further study whether two projections can share intermediate computation in Tab. 9d. While sharing middle hidden layers with slightly deeper projection heads performs better, we opt for 2-layer MLP without shared layers for its simplicity.", + "bbox": [ + 75, + 607, + 468, + 804 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Optimization. We study the effect of hyper-parameters in optimization. In Tab. 9e, we study the optimal loss ratio between $\\mathcal{L}_{\\mathrm{CLIP}}$ and $\\mathcal{L}_{\\mathrm{nCLIP}}$ . We further study whether it's the best route for two objectives to be optimized in separate latent spaces via multi-tasking. Specifically, we con", + "bbox": [ + 75, + 824, + 468, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "sider a bespoke objective Eq. (3) taking account of probability estimation and negative samples simultaneously in a shared latent space. Detailed formulation are shown in Appendix B.2. Optimizing in one latent space yields suboptimal results with a nearly $13.0\\%$ drop in zero-shot accuracy and $3.3\\%$ in linear probing accuracy, suggesting two objectives are intrinsically contradictory. We consider debiased sampling, where each batch is sampled from a single data source, since their semantics may be closer and thus more suitable for optimization. This yields similar results. We also add $\\mathcal{L}_{\\mathrm{nCLIP}}$ after warm-up epochs with the linear scheduler to its base scale, which erodes the performance.", + "bbox": [ + 496, + 375, + 890, + 556 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 569, + 617, + 585 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Aligning images and texts is of overriding significance for vision-language understanding. To conquer the systematic insufficiency of the contrastive objective for acquiring semantics and tackling loose correlations between noisy image-text pairs, we explore the non-contrastive objective for language-image pre-training and unravel its properties. Empirical evidence reveals that the non-contrastive objective induces models to perform favorably in representation learning yet poorly in zero-shot transfer. Observing the distinct mechanisms of the two objectives, we further seek synergy between the two, and introduce a simple multi-tasking framework, xCLIP, that enjoys the best of both worlds: nCLIP aids CLIP mining semantics while CLIP inherits intrinsic strengths for zero-shot recognition. The consistent performance gain of xCLIP over CLIP on a wide variety of downstream tasks consolidates our findings. As potential future work, we may continue to scale up the data size (e.g., LAION400M [64]) as well as the model size (e.g., ViT-L [26]) to verify whether the scaling law applies and the performance improvement endures.", + "bbox": [ + 496, + 594, + 890, + 897 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "11035", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Eneko Agirre, Carmen Banea, Claire Cardie, Daniel Cer, Mona Diab, Aitor Gonzalez-Agirre, Weiwei Guo, Inigo Lopez-Gazpio, Montse Maritxalar, Rada Mihalcea, et al. Semeval-2015 task 2: Semantic textual similarity, english, spanish and pilot on interpretability. In SemEval, 2015. 6", + "[2] Eneko Agirre, Carmen Banea, Claire Cardie, Daniel M Cer, Mona T Diab, Aitor Gonzalez-Agirre, Weiwei Guo, Rada Mihalcea, German Rigau, and Janyce Wiebe. Semeval-2014 task 10: Multilingual semantic textual similarity. In SemEval, 2014. 6", + "[3] Eneko Agirre, Carmen Banea, Daniel Cer, Mona Diab, Aitor Gonzalez Agirre, Rada Mihalcea, German Rigau Claramunt, and Janyce Wiebe. Semeval-2016 task 1: Semantic textual similarity, monolingual and cross-lingual evaluation. In SemEval, 2016. 6", + "[4] Eneko Agirre, Daniel Cer, Mona Diab, and Aitor Gonzalez-Agirre. Semeval-2012 task 6: A pilot on semantic textual similarity. In SemEval, 2012. 6", + "[5] Eneko Agirre, Daniel Cer, Mona Diab, Aitor Gonzalez-Agirre, and Weiwei Guo. *sem 2013 shared task: Semantic textual similarity. In SemEval, 2013. 6", + "[6] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. arXiv preprint arXiv:2204.14198, 2022. 2", + "[7] Elad Amrani and Alex Bronstein. Self-supervised classification network. In ECCV, 2022. 3", + "[8] Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Florian Bordes, Pascal Vincent, Armand Joulin, Michael Rabbat, and Nicolas Ballas. Masked siamese networks for label-efficient learning. In ECCV, 2022. 3", + "[9] Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Armand Joulin, Nicolas Ballas, and Michael Rabbat. Semi-supervised learning of visual features by nonparametrically predicting view assignments with support samples. In ICCV, 2021. 3, 8", + "[10] Andrei Barbu, David Mayo, Julian Alverio, William Luo, Christopher Wang, Dan Gutfreund, Josh Tenenbaum, and Boris Katz. Objectnet: A large-scale bias-controlled dataset for pushing the limits of object recognition models. In NeurIPS, 2019. 4", + "[11] Adrien Bardes, Jean Ponce, and Yann LeCun. Vicreg: Variance-invariance-covariance regularization for self-supervised learning. In ICLR, 2022. 3", + "[12] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. In NeurIPS, 2020. 1", + "[13] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features. In ECCV, 2018. 2, 3", + "[14] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. In NeurIPS, 2020. 2, 3, 7" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 2, 3, 6, 7, 8", + "[16] Daniel Cer, Mona Diab, Eneko Agirre, Inigo Lopez-Gazpio, and Lucia Specia. Semeval-2017 task 1: Semantic textual similarity-multilingual and cross-lingual focused evaluation. In SemEval, 2017. 6", + "[17] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing web-scale image-text pretraining to recognize long-tail visual concepts. In CVPR, 2021. 4", + "[18] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 2", + "[19] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In CVPR, 2021. 3", + "[20] Xinlei Chen*, Saining Xie*, and Kaiming He. An empirical study of training self-supervised vision transformers. In ICCV, 2021. 5", + "[21] Yanbei Chen, Yongqin Xian, Sophia Koepke, Ying Shan, and Zeynep Akata. Distilling audio-visual knowledge by compositional contrastive learning. In CVPR, 2021. 2", + "[22] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Universal image-text representation learning. In ECCV, 2020. 1, 2", + "[23] Tat-Seng Chua, Jinhui Tang, Richang Hong, Haojie Li, Zhiping Luo, and Yan-Tao Zheng. Nus-wide: A real-world web image database from national university of Singapore. In CIVR, 2009. 5", + "[24] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. In CVPR, 2009. 2, 4, 6, 7", + "[25] Karan Desai and Justin Johnson. Virtex: Learning visual representations from textual annotations. In CVPR, 2021. 1", + "[26] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 4, 8", + "[27] Aleksandr Ermolov, Aliaksandr Siarohin, Enver Sangineto, and Nicu Sebe. Whitening for self-supervised representation learning. In ICML, 2021. 3", + "[28] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. IJCV, 2010. 6", + "[29] Tianyu Gao, Xingcheng Yao, and Danqi Chen. SimCSE: Simple contrastive learning of sentence embeddings. In EMNLP, 2021. 2, 6", + "[30] Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A Wichmann, and Wieland Brendel. Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. In ICLR, 2019. 4" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "11036", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[31] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. In NeurIPS, 2020. 3", + "[32] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. In CVPR, 2022. 1", + "[33] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, 2020. 2", + "[34] Matthew Henderson, Rami Al-Rfou, Brian Strope, Yun-Hsuan Sung, László Lukács, Ruiqi Guo, Sanjiv Kumar, Balint Miklos, and Ray Kurzweil. Efficient natural language response suggestion for smart reply. arXiv preprint arXiv:1705.00652, 2017. 6", + "[35] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robustness: A critical analysis of out-of-distribution generalization. In ICCV, 2021. 4", + "[36] Dan Hendrycks and Kevin Gimpel. Bridging nonlinearities and stochastic regularizers with gaussian error linear units. In ICLR, 2017. 4", + "[37] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. In CVPR, 2021. 4", + "[38] R Devon Hjelm, Alex Fedorov, Samuel Lavoie-Marchildon, Karan Grewal, Phil Bachman, Adam Trischler, and Yoshua Bengio. Learning deep representations by mutual information estimation and maximization. In ICLR, 2019. 2", + "[39] Dat Huynh and Ehsan Elhamifar. A shared multi-attention framework for multi-label zero-shot learning. In CVPR, 2020. 5", + "[40] D. Huynh and E. Elhamifar. A shared multi-attention framework for multi-label zero-shot learning. In CVPR, 2020. 5", + "[41] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In ICML, 2015. 4", + "[42] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, 2021. 2", + "[43] Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution or region supervision. In ICML, 2021. 2", + "[44] Thomas Kipf, Elise van der Pol, and Max Welling. Contrastive learning of structured world models. arXiv preprint arXiv:1911.12247, 2019. 2", + "[45] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. IJCV, 2017. 4", + "[46] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4. IJCV, 2020. 5" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[47] Janghyeon Lee, Jongsuk Kim, Hyounguk Shon, Bumsoo Kim, Seung Hwan Kim, Honglak Lee, and Junmo Kim. Uniclip: Unified framework for contrastive language-image pretraining. In NeurIPS, 2022. 2", + "[48] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. In NeurIPS, 2021. 2", + "[49] Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. Visualbert: A simple and performant baseline for vision and language. In ACL, 2020. 2", + "[50] Yangguang Li, Feng Liang, Lichen Zhao, Yufeng Cui, Wanli Ouyang, Jing Shao, Fengwei Yu, and Junjie Yan. Supervision exists everywhere: A data efficient contrastive language-image pre-training paradigm. In ICLR, 2022. 1, 2", + "[51] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 4, 5", + "[52] Xingbin Liu, Jinghao Zhou, Tao Kong, Xianming Lin, and Rongrong Ji. Exploring target representations for masked autoencoders. arXiv preprint arXiv:2209.03917, 2022. 1", + "[53] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In ICLR, 2019. 4", + "[54] Pranay Manocha, Zeyu Jin, Richard Zhang, and Adam Finkelstein. Cdpam: Contrastive learning for perceptual audio similarity. In ICASSP, 2021. 2", + "[55] Norman Mu, Alexander Kirillov, David Wagner, and Saining Xie. Slip: Self-supervision meets language-image pretraining. In ECCV, 2022. 2", + "[56] Vicente Ordonez, Girish Kulkarni, and Tamara Berg. Im2text: Describing images using 1 million captioned photographs. In NeuIPS, 2011. 4", + "[57] Hieu Pham, Zihang Dai, Golnaz Ghiasi, Hanxiao Liu, Adams Wei Yu, Minh-Thang Luong, Mingxing Tan, and Quoc V Le. Combined scaling for zero-shot transfer learning. arXiv preprint arXiv:2111.10050, 2021. 1, 2, 4, 7", + "[58] Bryan A Plummer, Liwei Wang, Chris M Cervantes, Juan C Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In ICCV, 2015. 2, 5", + "[59] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 2, 3, 4, 5", + "[60] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 1", + "[61] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, 2021. 1", + "[62] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. Do imagenet classifiers generalize to imageNet? In ICML, 2019. 4" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "11037", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[63] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In NeurIPS, 2015. 2", + "[64] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. In NeurIPS Workshops, 2021. 8", + "[65] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In ACL, 2018. 1, 4", + "[66] Aravind Srinivas, Michael Laskin, and Pieter Abbeel. Curl: Contrastive unsupervised representations for reinforcement learning. In ICML, 2020. 2", + "[67] Weijie Su, Xizhou Zhu, Yue Cao, Bin Li, Lewei Lu, Furu Wei, and Jifeng Dai. Vl-bert: Pre-training of generic visual-linguistic representations. In ICLR, 2020. 2", + "[68] Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. In EMNLP, 2019. 2", + "[69] Bart Thomee, David A Shamma, Gerald Friedland, Benjamin Elizalde, Karl Ni, Douglas Poland, Damian Borth, and Li-Jia Li. Yfcc100m: The new data in multimedia research. In CACM, 2016. 1", + "[70] Yael Vinker, Ehsan Pajouheshgar, Jessica Y Bo, Roman Christian Bachmann, Amit Haim Bermano, Daniel Cohen-Or, Amir Zamir, and Ariel Shamir. Clipasso: Semantically-aware object sketching. In SIGGRAPH, 2022. 1", + "[71] Feng Wang, Tao Kong, Rufeng Zhang, Huaping Liu, and Hang Li. Self-supervised learning by estimating twin class distributions. arXiv preprint arXiv:2110.07402, 2021. 3", + "[72] Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. Learning robust global representations by penalizing local predictive power. In NeurIPS, 2019. 4", + "[73] Zirui Wang, Jiahui Yu, Adams Wei Yu, Zihang Dai, Yulia Tsvetkov, and Yuan Cao. Simvlm: Simple visual language model pretraining with weak supervision. In ICLR, 2022. 2", + "[74] Bichen Wu, Ruizhe Cheng, Peizhao Zhang, Tianren Gao, Joseph E. Gonzalez, and Peter Vajda. Data efficient language-supervised zero-shot recognition with optimal transport distillation. In ICLR, 2022. 1, 2", + "[75] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In CVPR, 2018. 2", + "[76] Qizhe Xie, Minh-Thang Luong, Eduard Hovy, and Quoc V Le. Self-training with noisy student improves imagenet classification. In CVPR, 2020. 4", + "[77] Jiarui Xu, Shalini De Mello, Sifei Liu, Wonmin Byeon, Thomas Breuel, Jan Kautz, and Xiaolong Wang. Groupvit: Semantic segmentation emerges from text supervision. In CVPR, 2022. 6", + "[78] Jianwei Yang, Chunyuan Li, Pengchuan Zhang, Bin Xiao, Ce Liu, Lu Yuan, and Jianfeng Gao. Unified contrastive learning in image-text-label space. In CVPR, 2022. 1" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[79] Lewei Yao, Runhui Huang, Lu Hou, Guansong Lu, Minzhe Niu, Hang Xu, Xiaodan Liang, Zhenguo Li, Xin Jiang, and Chunjing Xu. Filip: Fine-grained interactive language-image pre-training. In ICLR, 2022. 2", + "[80] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917, 2022. 1, 2, 4, 7", + "[81] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow twins: Self-supervised learning via redundancy reduction. In ICML, 2021. 3", + "[82] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. ibot: Image bert pre-training with online tokenizer. In ICLR, 2022. 5, 7" + ], + "bbox": [ + 501, + 90, + 890, + 290 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11038", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Non-Contrastive Learning Meets Language-Image Pre-Training/fe567c45-5ef2-4da5-8116-0c22fc6e398a_model.json b/2023/Non-Contrastive Learning Meets Language-Image Pre-Training/fe567c45-5ef2-4da5-8116-0c22fc6e398a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5bc01f8a3eb418a384689715bf0399a72071fbd3 --- /dev/null +++ b/2023/Non-Contrastive Learning Meets Language-Image Pre-Training/fe567c45-5ef2-4da5-8116-0c22fc6e398a_model.json @@ -0,0 +1,2609 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.163, + 0.131, + 0.809, + 0.154 + ], + "angle": 0, + "content": "Non-Contrastive Learning Meets Language-Image Pre-Training" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.181, + 0.744, + 0.216 + ], + "angle": 0, + "content": "Jinghao Zhou Li Dong Zhe Gan Lijuan Wang Furu Wei Microsoft" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.252, + 0.314, + 0.267 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.284, + 0.474, + 0.649 + ], + "angle": 0, + "content": "Contrastive language-image pre-training (CLIP) serves as a de-facto standard to align images and texts. Nonetheless, the loose correlation between images and texts of web-crawled data renders the contrastive objective data inefficient and craving for a large training batch size. In this work, we explore the validity of non-contrastive language-image pre-training (nCLIP), and study whether nice properties exhibited in visual self-supervised models can emerge. We empirically observe that the non-contrastive objective benefits representation learning while sufficiently underperforming under zero-shot recognition. Based on the above study, we further introduce xCLIP, a multi-tasking framework combining CLIP and nCLIP, and show that nCLIP aids CLIP in enhancing feature semantics. The synergy between two objectives lets xCLIP enjoy the best of both worlds: superior performance in both zero-shot transfer and representation learning. Systematic evaluation is conducted spanning a wide variety of downstream tasks including zero-shot classification, out-of-domain classification, retrieval, visual representation learning, and textual representation learning, showcasing a consistent performance gain and validating the effectiveness of xCLIP. The code and pre-trained models will be publicly available at https://github.com/shallowtoil/xclip." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.678, + 0.21, + 0.695 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Language-image pre-training which simultaneously learns textual and visual representation from large-scale image-text pairs has revolutionized the field of representation learning [25, 59], vision-language understanding [22], and text-to-image generation [61]. Compared to traditional visual models, language-instilled ones intrinsically inherit the capability of zero-shot or few-shot learning prominently demonstrated by large language models such as GPT-3 [12]. The precursor system, Contrastive Language-Image Pre-Training [59] (CLIP) that explicitly aligns the projected features of two modalities, has demonstrated surprising capabilities of zero-shot, representation learning, and robustness, being applied to a wide range of fields [32, 52, 60, 70]." + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.256, + 0.895, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.4, + 0.895, + 0.525 + ], + "angle": 0, + "content": "Figure 1. Architecture comparison between CLIP and nCLIP. We take base-size encoders for an instance. CLIP discriminates instances within a batch using 512-dim projected embeddings. nCLIP projects each modality into a 32768-dim probability distribution as the pseudo-label to supervise the prediction from the other modality. Darker blocks of nCLIP depict a higher response for cluster distribution, signifying the clusters to which text or image instances may belong. xCLIP is a multi-tasking framework with both CLIP and nCLIP." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.537, + 0.895, + 0.794 + ], + "angle": 0, + "content": "To learn from noisy web-crawled image-text pairs, CLIP adopts a formulation of the contrastive objective, where the image and text within a pair are considered as unique instances and are encouraged to be discriminated from all the other negative instances. However, web-crawled image-text pairs [65,69] are usually loosely correlated in the sense that one caption (image) can match reasonably with multiple images (captions) besides the ground-truth one, as statistically shown in [74]. Hence, it is inaccurate and data-inefficient for representation learning to neglect other sensible matches and overlook the semantics hidden inside the textual description. This is also solidified by the undesirable discriminative ability or transferring performance of the visual encoder pre-trained under the contrastive objective, as suggested in [78]. Mining semantics from plentiful concepts appearing in captions, therefore, solicits further exploration beyond the vanilla contrastive objective." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Previously, some works resort to mining nearest neighbor positive samples via intra-modality similarity [50, 74] but require extra storage for auxiliary modules, i.e., the teacher network [74] or the memory bank [50]. Other works conduct multi-tasking with image-label supervision [57, 78, 80], transforming the captions into tag format or the image tag into the captioning format for unified contrastive" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "11028" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.228 + ], + "angle": 0, + "content": "learning. Notwithstanding, such practice is still unable to account for the loose assignment within the image-text captioning pairs. Intuitively, a caption for an image usually identifies existing objects within the image, which can be captured by a probabilistic distribution estimating how the text is assigned to one or multiple object clusters [13, 14]. Such estimation depicts the semantic meanings for its contained visual contents, and thus can be leveraged as a pseudo-label to guide the learning of the visual encoder." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.232, + 0.471, + 0.626 + ], + "angle": 0, + "content": "Inspired by the superiority of visual self-supervised learning (SSL) models pre-trained with the non-contrastive objective [14, 15], we explore whether the non-contrastive objective can be used across modalities for pre-training language-image models, and whether nice properties displayed on visual SSL models can be inherited. To this end, we follow the same setup as CLIP except for the objective, and study non-Contrastive Language-Image Pre-Training (nCLIP). For an image-text pair, we use the estimation of the textual (visual) distribution as the target to supervise the visual (textual) distribution, measured by cross-entropy. Additional regularizers are applied to avoid trivial solutions. Schematic comparison between CLIP and nCLIP is shown in Fig. 1. Theoretically, such formulation takes one modality and the cluster centroid that the other modality belongs to as the positive samples to contrast [14], as opposed to direct features of two modalities as in contrastive learning, such that a single image is tolerated to be aligned with multiple captions and concepts. Based on the above formulation, we conduct a systematic study in terms of zero-shot transfer and representation learning. We empirically observe that, while nCLIP demonstrates desirable representation learning capability for both modalities, it underperforms prominently in zero-shot classification and retrieval tasks where models pre-trained with negative samples naturally prevail." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Seeking for unique strengths of both worlds, we further perform multi-tasking of CLIP and nCLIP, short as xCLIP, and seek synergy between two distinct pre-training objectives. With the bearable add-on of computational resources (e.g., \\(\\sim 27\\%\\) memory-wise and \\(\\sim 30\\%\\) time-wise), xCLIP achieves consistent performance gain compared to CLIP on a wide range of tasks spanning from zero-shot classification, retrieval, linear probing, and fine-tuning, etc. An extensive study with different pre-training datasets, evaluation metrics, and optimization configurations is conducted to validate xCLIP's effectiveness in mining semantics and improving data efficiency. Particularly, the base-size model pre-trained using xCLIP under 35-million publicly available image-text pairs achieves a performance gain of \\(3.3\\%\\) and \\(1.5\\%\\) on an average of 27 classification tasks [59], in terms of zero-shot classification and linear probing accuracy, respectively. Performance gain is also valid in terms of zero-shot retrieval, semi-supervised learning, and fine" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.138 + ], + "angle": 0, + "content": "tuning, with 3.7 points of R@1 on Flickr30K [58], \\(1.7\\%\\) accuracy on \\(1\\%\\) subset of ImageNet [24], and \\(0.3\\%\\) accuracy on ImageNet, respectively." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.152, + 0.642, + 0.168 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.179, + 0.892, + 0.556 + ], + "angle": 0, + "content": "Language-image pre-training. Language-image pretraining, aka vision-language pre-training (VLP), learns to jointly cope with vision and language input using multi-modality models. Early practices [22, 49, 67, 68] separately encode modalities, i.e., texts with linear embedding and images with convolutional neural networks and region proposals [63]. Recent works take direct images instead of pre-fetched region features as input, using e.g., dual-encoder architecture for alignment [42, 57, 59], single-encoder architecture for fusion [43, 73], or their combinatorial practices forming an encoder-decoder architecture [48, 80]. Most recent literature aims to expand as many supported transferable downstream tasks as possible via pre-training large-scale foundation models [6, 80]. Out of previous works, CLIP [59] blazes the trail to learn transferable visual features from text supervision, and demonstrates surprising zero-shot recognition results. Follow-up study includes leveraging dense [79], augmentation [47], and uni-modal self-supervised [50, 55, 74] signals for improved performance. In this work, we base our study on the dual-encoder architecture, and pre-train models either with the contrastive objective as CLIP [59], or the non-contrastive objective. Further, we seek their complementarity and study whether synergy between two objectives exists." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.577, + 0.892, + 0.803 + ], + "angle": 0, + "content": "Contrastive learning. The idea of contrastive learning is popularized under visual self-supervised learning (SSL) which aims at learning transferable visual representation from unlabelled images. In common practices [18, 33, 75], each image is considered a single class, attracted to its augmented views while pulled away from views of other images. The contrastive objective is proved robust and effective in a wide range of realms beyond visual understanding, including representation learning for natural language [29], audio [54], structured input [44], robotics [66], and multimodality [21, 59]. However, a caveat of these approaches is the requirement for large batch size [18] or memory bank [33, 75], which lies intrinsic in the formulation of InfoNCE estimator [38] and plagues the pre-training of largesize models due to hardware limitations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Non-contrastive learning. Beyond the contrastive paradigm, recent state of the arts from visual SSL manage to relieve the dependency on negative samples. With different subtleties to avert collapsing solutions, these works can optimize the affinity of augmented representations alone" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11029" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.228 + ], + "angle": 0, + "content": "and are categorized as non-contrastive framework. To avoid model collapsing, common practices include asymmetrical architecture [19, 31], dimension de-correlation [11, 27, 81], and clustering [7, 8, 13–15, 71]. Our approach takes inspiration from the last category since it intrinsically frees the absoluteness of assignment of positive samples. Moreover, rich semantics induced by explicit clustering can automatically group visual concepts from noisy image-text pairs, facilitating models' representation learning." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.242, + 0.186, + 0.259 + ], + "angle": 0, + "content": "3. Approach" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.267, + 0.206, + 0.282 + ], + "angle": 0, + "content": "3.1. Framework" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.289, + 0.471, + 0.397 + ], + "angle": 0, + "content": "Suppose \\( \\pmb{g} \\) and \\( \\pmb{h} \\in \\mathbb{R}^{D \\times 1} \\) are projected backbone features of two modalities. \\( D \\) is the feature dimension. We start by formulating the objective of dominant contrastive framework CLIP, followed by the introduced non-contrastive framework nCLIP. The final framework xCLIP takes the multi-tasking of two objectives, the complete computational pipeline of which is shown in Algorithm 1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.414, + 0.47, + 0.492 + ], + "angle": 0, + "content": "Contrastive pre-training [59]. Let \\(\\pmb{u} = \\pmb{g} / ||\\pmb{g}||\\) and \\(\\pmb{v} = \\pmb{h} / ||\\pmb{h}||\\). \\(\\pmb{U} = [\\pmb{u}_1, \\pmb{u}_2, \\dots, \\pmb{u}_B] \\in \\mathbb{R}^{D \\times B}\\), \\(\\pmb{V} = [\\pmb{v}_1, \\pmb{v}_2, \\dots, \\pmb{v}_B] \\in \\mathbb{R}^{D \\times B}\\) are concatenated embeddings over the batch. \\(B\\) is the batch size. The contrastive objective is formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.122, + 0.5, + 0.47, + 0.567 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {C L I P}} = \\operatorname {I n f o N C E} (\\boldsymbol {U} ^ {\\mathrm {T}} \\boldsymbol {V} / \\sigma) \\\\ = - \\frac {1}{B} \\sum_ {i \\in B} \\log \\frac {\\exp \\left(\\boldsymbol {u} _ {i} ^ {\\mathrm {T}} \\boldsymbol {v} _ {i} / \\sigma\\right)}{\\sum_ {j \\in B} \\exp \\left(\\boldsymbol {u} _ {i} ^ {\\mathrm {T}} \\boldsymbol {v} _ {j} / \\sigma\\right)}, \\tag {1} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.578, + 0.47, + 0.76 + ], + "angle": 0, + "content": "where \\(\\sigma\\) is a trainable parameter controlling the temperature. \\(\\mathcal{L}_{\\mathrm{CLIP}}\\) is symmetrized by setting \\(g\\) and \\(h\\) as projected features of the images and texts by turns and taking the average of two terms. Vanilla InfoNCE can be decoupled into two terms accounting for affinity and variability separately [81]. While both variability terms in the contrastive and non-contrastive objectives require for batch statistics (e.g., \\(\\mathcal{L}_{\\mathrm{EH}}\\) in Eq. (4)), CLIP formulation explicitly maximizes the distance between negative pairs of batch samples. Hence, it degrades the learning performance and data efficiency when models are pre-trained with noisy data where sensible matches occur within negative pairs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.778, + 0.47, + 0.87 + ], + "angle": 0, + "content": "Non-contrastive pre-training. Let \\( \\pmb{p} = \\mathrm{softmax}(\\pmb{g}) \\) and \\( \\pmb{q} = \\mathrm{softmax}(\\pmb{h}) \\). We transform the projected feature into probability distributions, which can be seen as an assignment over semantic clusters (i.e., projection weights). We take \\( \\pmb{p} \\) as the target distribution and learn to estimate the predicted distribution \\( \\pmb{q} \\) by minimizing their cross-entropy:" + }, + { + "type": "equation", + "bbox": [ + 0.205, + 0.88, + 0.47, + 0.898 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {C E}} = - \\boldsymbol {p} ^ {\\mathrm {T}} \\log (\\boldsymbol {q}). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "Note the target and prediction branch are backpropagated at the same time and \\(\\mathcal{L}_{\\mathrm{CE}}\\) is symmetrized as \\(\\mathcal{L}_{\\mathrm{CLIP}}\\). To avert collapsing solutions, we incorporate entropy regularizers, i.e., entropy minimization [71] and mean entropy maximization [8,9], via minimizing:" + }, + { + "type": "equation", + "bbox": [ + 0.557, + 0.178, + 0.892, + 0.197 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {E H}} = - \\boldsymbol {p} ^ {\\mathrm {T}} \\log (\\boldsymbol {p}), - \\mathcal {L} _ {\\mathrm {H E}} = \\bar {\\boldsymbol {p}} ^ {\\mathrm {T}} \\log (\\bar {\\boldsymbol {p}}), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.209, + 0.893, + 0.437 + ], + "angle": 0, + "content": "where \\(\\overline{\\boldsymbol{p}} = \\mathbb{E}(p) \\approx \\frac{1}{B} \\sum_{i=1}^{B} \\boldsymbol{p}_i\\) is the average distribution over batch. \\(\\mathcal{L}_{\\mathrm{EH}}\\) and \\(\\mathcal{L}_{\\mathrm{HE}}\\) are symmetrized by taking the average with another term where \\(\\boldsymbol{p}\\) is replaced by \\(\\boldsymbol{q}\\). \\(\\mathcal{L}_{\\mathrm{EH}}\\) encourages the model to make deterministic predictions and ensures assignment's sharpness. \\(\\mathcal{L}_{\\mathrm{HE}}\\) encourages the model to utilize a full set of projection weights and ensures the assignment's smoothness. We illustrate in Sec. 4.4 and Appendix A that \\(\\mathcal{L}_{\\mathrm{EH}}\\) and \\(\\mathcal{L}_{\\mathrm{HE}}\\) are minimally sufficient to yield non-collapsing solutions. We empirically show in Tab. A1 that the solutions is non-trivial as long as both sharpness and smoothness are guaranteed, e.g., via Sinkhorn algorithm [14] instead of loss regularizers. We opt for Eq. (3) for its simplicity and consistency between pretraining and evaluation. The overall non-contrastive objective is formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.561, + 0.45, + 0.892, + 0.467 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {n C L I P}} = \\mathcal {L} _ {\\mathrm {C E}} + \\lambda_ {1} \\cdot \\mathcal {L} _ {\\mathrm {E H}} - \\lambda_ {2} \\cdot \\mathcal {L} _ {\\mathrm {H E}}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.479, + 0.893, + 0.735 + ], + "angle": 0, + "content": "where \\(\\lambda_{1}\\) and \\(\\lambda_{2}\\) controls the weight of the regularization. We empirically find that setting \\(\\lambda_{1} = \\lambda_{2} - 1 = 0.5\\) yields stable training and favorable transfer performances. We detail our findings in Sec. 4.4 and Appendix B.1. During evaluation, we utilize negative cross-entropy \\(-\\mathcal{L}_{\\mathrm{CE}}\\) as the similarity metric for zero-shot transfer experiments. As schematically showcased in Fig. 1, nCLIP does not reply on negative examples but requires a relatively larger projection head and greater output dimension. Comparing nCLIP with CLIP, we empirically find that nCLIP produces more coarse-grained (e.g., zero-shot retrieval in Sec. 4.1) but semantic-rich (e.g., linear probing in Sec. 4.2) projections. Empirical evidence suggests that the model will fail to deliver reasonable zero-shot and retrieval results if pre-trained without negative pairs (e.g., nCLIP's 25.2 vs. CLIP's 73.8 points of R@1 under Flickr30K I→T retrieval as shown in Tab. 4)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.755, + 0.892, + 0.815 + ], + "angle": 0, + "content": "Contrastive meets non-contrastive. Given that two objectives each have their own limitations, we further seek the complementarity between \\(\\mathcal{L}_{\\mathrm{CLIP}}\\) and \\(\\mathcal{L}_{\\mathrm{nCLIP}}\\), and pre-train the models with both objectives, written as" + }, + { + "type": "equation", + "bbox": [ + 0.542, + 0.828, + 0.892, + 0.844 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {x C L I P}} = \\lambda_ {\\mathrm {C L I P}} \\cdot \\mathcal {L} _ {\\mathrm {C L I P}} + \\lambda_ {\\mathrm {n C L I P}} \\cdot \\mathcal {L} _ {\\mathrm {n C L I P}}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where \\(\\lambda_{\\mathrm{CLIP}}\\) and \\(\\lambda_{\\mathrm{nCLIP}}\\) control the weight of the objectives. During evaluation, we find using the negative cosine metric with CLIP's projection head for zero-shot transfer" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11030" + } + ], + [ + { + "type": "code_caption", + "bbox": [ + 0.088, + 0.095, + 0.427, + 0.11 + ], + "angle": 0, + "content": "Algorithm 1: PyTorch-like pseudo-code of xCLIP." + }, + { + "type": "code", + "bbox": [ + 0.099, + 0.119, + 0.368, + 0.133 + ], + "angle": 0, + "content": "// enc, proj: encoder & projector" + }, + { + "type": "code", + "bbox": [ + 0.099, + 0.134, + 0.292, + 0.148 + ], + "angle": 0, + "content": "\\(\\pmb{f}_I,\\pmb {f}_T = \\mathrm{enc}(\\mathrm{img}),\\mathrm{enc}(\\mathrm{text})\\)" + }, + { + "type": "code", + "bbox": [ + 0.099, + 0.149, + 0.352, + 0.164 + ], + "angle": 0, + "content": "\\(\\pmb{g}_I, \\pmb{g}_T = \\mathrm{proj}_{\\mathrm{CLIP}}(\\pmb{f}_I), \\mathrm{proj}_{\\mathrm{CLIP}}(\\pmb{f}_T)\\)" + }, + { + "type": "code", + "bbox": [ + 0.099, + 0.164, + 0.369, + 0.179 + ], + "angle": 0, + "content": "\\(h_{I},h_{T} = \\mathrm{proj}_{\\mathrm{nCLIP}}(\\pmb {f}_{I}),\\mathrm{proj}_{\\mathrm{nCLIP}}(\\pmb{f}_{T})\\)" + }, + { + "type": "code", + "bbox": [ + 0.1, + 0.18, + 0.34, + 0.194 + ], + "angle": 0, + "content": "\\(\\mathcal{L}_{\\mathrm{xCLIP}} = \\lambda_{\\mathrm{CLIP}}\\cdot \\mathcal{L}_{\\mathrm{CLIP}}(\\pmb {g}_I,\\pmb {g}_T,\\sigma)\\)" + }, + { + "type": "code", + "bbox": [ + 0.158, + 0.195, + 0.388, + 0.209 + ], + "angle": 0, + "content": "+λnCLIP·LnCLIP(hH,T,λ1,λ2)" + }, + { + "type": "code", + "bbox": [ + 0.099, + 0.21, + 0.199, + 0.224 + ], + "angle": 0, + "content": "return \\(\\mathcal{L}_{\\mathrm{xCLIP}}\\)" + }, + { + "type": "code", + "bbox": [ + 0.099, + 0.229, + 0.258, + 0.244 + ], + "angle": 0, + "content": "def \\(\\mathcal{L}_{\\mathrm{CLIP}}(\\pmb {f}_I,\\pmb {f}_T,\\sigma)\\) .." + }, + { + "type": "code", + "bbox": [ + 0.124, + 0.245, + 0.351, + 0.257 + ], + "angle": 0, + "content": "// INCE: InfoNCE with Eq. (1)" + }, + { + "type": "code", + "bbox": [ + 0.124, + 0.258, + 0.34, + 0.272 + ], + "angle": 0, + "content": "\\(\\pmb{u}_{I}\\) = normalize \\((f_{I}, p = 2, \\dim = 1)\\)" + }, + { + "type": "code", + "bbox": [ + 0.124, + 0.273, + 0.345, + 0.287 + ], + "angle": 0, + "content": "\\(\\pmb{u}_{T} = \\mathrm{normalize}(\\pmb{f}_{T}, \\mathrm{p} = 2, \\mathrm{dim} = 1)\\)" + }, + { + "type": "code", + "bbox": [ + 0.124, + 0.287, + 0.408, + 0.303 + ], + "angle": 0, + "content": "\\(\\mathcal{L} = \\mathrm{INCE}(\\pmb{U}_I^{\\mathrm{T}}\\pmb{U}_T / \\sigma) + \\mathrm{INCE}(\\pmb{U}_T^{\\mathrm{T}}\\pmb{U}_I / \\sigma)\\)" + }, + { + "type": "code", + "bbox": [ + 0.125, + 0.304, + 0.208, + 0.316 + ], + "angle": 0, + "content": "return \\(\\mathcal{L} / 2\\)" + }, + { + "type": "code", + "bbox": [ + 0.099, + 0.326, + 0.297, + 0.34 + ], + "angle": 0, + "content": "def \\(\\mathcal{L}_{\\mathrm{nCLIP}}(f_I, f_T, \\lambda_1, \\lambda_2)\\):" + }, + { + "type": "code", + "bbox": [ + 0.124, + 0.341, + 0.293, + 0.355 + ], + "angle": 0, + "content": "\\(\\pmb{p}_I = \\mathrm{softmax}(\\pmb {f}_I,\\mathrm{dim} = 1)\\)" + }, + { + "type": "code", + "bbox": [ + 0.124, + 0.355, + 0.298, + 0.37 + ], + "angle": 0, + "content": "\\(\\pmb{p}_T = \\mathrm{softmax}(\\pmb {f}_T,\\mathrm{dim} = 1)\\)" + }, + { + "type": "code", + "bbox": [ + 0.124, + 0.371, + 0.452, + 0.385 + ], + "angle": 0, + "content": "\\(\\mathcal{L}_{\\mathrm{CE}} = -\\left(\\pmb {p}_I\\cdot \\log (\\pmb {p}_T) + \\pmb {p}_T\\cdot \\log (\\pmb {p}_I)\\right).\\) sum \\((\\mathrm{dim} = 1)\\)" + }, + { + "type": "code", + "bbox": [ + 0.174, + 0.386, + 0.268, + 0.398 + ], + "angle": 0, + "content": ".mean(dim=0)" + }, + { + "type": "code", + "bbox": [ + 0.124, + 0.4, + 0.452, + 0.415 + ], + "angle": 0, + "content": "\\(\\mathcal{L}_{\\mathrm{EH}} = -(\\pmb {p}_I\\cdot \\log (\\pmb {p}_I) + \\pmb {p}_T\\cdot \\log (\\pmb {p}_T)).\\) sum \\((\\dim = 1)\\)" + }, + { + "type": "code", + "bbox": [ + 0.174, + 0.416, + 0.268, + 0.429 + ], + "angle": 0, + "content": ".mean(dim=0)" + }, + { + "type": "code", + "bbox": [ + 0.124, + 0.431, + 0.421, + 0.446 + ], + "angle": 0, + "content": "\\(\\overline{\\pmb{p}}_I,\\overline{\\pmb{p}}_T = \\pmb {p}_I.\\mathrm{mean}(\\mathrm{dim} = 0),\\pmb {p}_T.\\mathrm{mean}(\\mathrm{dim} = 0)\\)" + }, + { + "type": "code", + "bbox": [ + 0.124, + 0.446, + 0.452, + 0.461 + ], + "angle": 0, + "content": "\\(\\mathcal{L}_{\\mathrm{HE}} = -\\left(\\overline{\\boldsymbol{p}}_I\\cdot \\log (\\overline{\\boldsymbol{p}}_I) + \\overline{\\boldsymbol{p}}_T\\cdot \\log (\\overline{\\boldsymbol{p}}_T)\\right).\\) sum(dim=1)" + }, + { + "type": "code", + "bbox": [ + 0.124, + 0.462, + 0.334, + 0.475 + ], + "angle": 0, + "content": "\\(\\mathcal{L} = \\mathcal{L}_{\\mathrm{CE}} + \\lambda_1\\cdot \\mathcal{L}_{\\mathrm{EH}} - \\lambda_2\\cdot \\mathcal{L}_{\\mathrm{HE}}\\)" + }, + { + "type": "code", + "bbox": [ + 0.124, + 0.476, + 0.208, + 0.489 + ], + "angle": 0, + "content": "return \\(\\mathcal{L} / 2\\)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.507, + 0.47, + 0.643 + ], + "angle": 0, + "content": "experiments generally leads to better results. We note that it is not immediately clear that the two objectives certainly induce stronger models, since they build qualitatively distinctive latent spaces. Qualitatively, we show in Sec. 4 that nCLIP helps CLIP to encode semantics which intrinsically lacks in CLIP, while CLIP helps nCLIP to be transferable for zero-shot recognition. The synergy between the two objectives prompts consistent performance gains across a wide range of tasks by xCLIP over CLIP." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.656, + 0.236, + 0.672 + ], + "angle": 0, + "content": "3.2. Implementation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.68, + 0.469, + 0.845 + ], + "angle": 0, + "content": "Architecture. We train the base-size model with ViTB/16 [26] as the visual encoder. The model configuration of the base-size text encoder follows CLIP [59] with byte-pair encoding (BPE) and a maximum context length of 77. The projection head generating output for the noncontrastive objective is a 2-layer MLP with 4096-dim hidden layers, GELU [36], and BatchNorm [41]. The last layer is of 32768 output dimension and followed by a BatchNorm without affine transformation. The projection head for the contrastive objective is a single linear layer with no bias and a dimension of 512." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Pre-training data. We train our method with publicly available datasets: COCO [51], Visual Genome [45], SBU" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.258 + ], + "angle": 0, + "content": "Captions [56], Conceptual Caption 3M [65], Conceptual Caption 12M [17], and filtered 14M-size subset of Yahoo Flickr Creative Commons 100M dataset, consisting a total of 35M Image-Text pairs (IT35M). We are also intrigued about the behaviors of models pre-trained on ImagNet-21K [24] (IN21K) dataset taking label names as annotation texts by concatenating them with a sampled prompt, similar to [57, 80]. To notify, ImageNet-21K is a subset of ImageNet-22K with classes of ImageNet-1K excluded for fair downstream evaluation. We detail the models' data scaling behavior in Tab. 8." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.277, + 0.892, + 0.459 + ], + "angle": 0, + "content": "Optimization. We use a batch size of 4096, with all data distributed across 32 V100 GPUs. Models are trained by default with AdamW [53] optimizer, a peak learning rate of \\(1e^{-3}\\), cosine scheduler, a weight decay of 0.2, a \\(\\beta_{2}\\) for AdamW of 0.98, a \\(\\epsilon\\) for AdamW of \\(1e^{-6}\\), and automatic mixed-precision for 3 warm-up epochs and a total of 32 epochs. \\(\\lambda_{1}\\) and \\(\\lambda_{2}\\) are set as 0.5 and 1.5, respectively. The two objectives are multi-tasked with \\(\\lambda_{\\mathrm{nCLIP}}\\) being 1 and \\(\\lambda_{\\mathrm{CLIP}}\\) being 0.2. Augmentations and pre-processing of pre-training data follow CLIP with image size randomly cropped and resized within the scale of (0.5, 1.0). Detailed pre-training hyper-parameters are detailed in Appendix F." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.472, + 0.633, + 0.489 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.497, + 0.685, + 0.511 + ], + "angle": 0, + "content": "4.1. Zero-Shot Transfer" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.52, + 0.892, + 0.687 + ], + "angle": 0, + "content": "Classification. We evaluate 27 classification benchmarks with zero-shot classification protocols following [59]. As shown in Tab. 1, nCLIP achieves an average top-1 accuracy of \\(32.7\\%\\). Though performing \\(7.6\\%\\) worse than CLIP, nCLIP demonstrates descent zero-shot recognition capability without explicitly training with negative examples. xCLIP achieves consistent performance gain across a wide range of datasets, with a \\(43.6\\%\\) average top-1 accuracy, which is \\(3.3\\%\\) higher than CLIP. It indicates that the synergy between contrastive and non-contrastive objectives improves the model's zero-shot classification ability." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Out-of-domain classification. For out-of-distribution classification, we use 5 datasets following [59]: ImageNet Adversarial [37], ImageNet Rendition [35], ImageNetV2 [62], ImageNet Sketch [72], and ObjectNet [10] with 2 additional datasets: ImageNet-C [76] and Stylized ImageNet [30]. We investigate if the non-contrastive objective inherits the robustness of the contrastive objective to natural distribution shift. As shown in Tab. 2, nCLIP achieves an \\(18.0\\%\\) with CLIP an \\(23.1\\%\\) average top-1 accuracy, revealing a similar trend to in-domain datasets. xCLIP further improves CLIP's accuracy by \\(3.6\\%\\) and achieves an accuracy of \\(26.7\\%\\) across 7 datasets, indicating its effectiveness among out-of-domain datasets." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "11031" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.081, + 0.089, + 0.89, + 0.178 + ], + "angle": 0, + "content": "
ModelFood101CIFAR10CIFAR100BirdmapSUN397CarsAircraftVOC2007DTDPetsCaltech101FlowersMNISTFER2013STL10EuroSATRESISC45GTSRBKITTICountry211PCAMUCF101Kinetics700CLEVRHapticMemesSSTImageNetAverage
CLIP61.279.750.623.756.515.95.846.427.654.771.348.910.537.391.324.539.713.131.69.150.045.032.412.853.049.145.740.3
nCLIP28.479.549.111.357.05.94.551.422.914.665.023.19.913.594.815.121.22.735.45.851.242.028.412.452.750.037.032.7
xCLIP65.883.454.525.159.918.05.852.233.257.173.950.012.339.092.840.043.616.339.89.351.149.835.418.452.550.248.843.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.181, + 0.895, + 0.224 + ], + "angle": 0, + "content": "Table 1. Zero-shot classification. We report on a variety of classification benchmarks with ViT-B/16 pre-trained on IT35M. Detailed protocols for each dataset strictly follow CLIP [59]. xCLIP achieves a consistent performance gain compared to CLIP in a wide range of classification datasets. Best results of each column are bolded." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.246, + 0.468, + 0.342 + ], + "angle": 0, + "content": "
ModelIN-AIN-RIN-v2IN-SkeIN-StyIN-CObjNetAverage
CLIP21.343.737.926.54.911.915.523.1
nCLIP21.529.030.817.73.611.512.018.0
xCLIP27.749.140.630.66.114.717.826.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.347, + 0.47, + 0.402 + ], + "angle": 0, + "content": "Table 2. Zero-shot out-of-distribution classification. We report on a variety of out-of-distribution classification benchmarks. xCLIP demonstrates stronger robustness on various out-of-domain classification datasets." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.418, + 0.462, + 0.51 + ], + "angle": 0, + "content": "
ModelNUS-WIDEOpenImages
mAPF1@3F1@5mAPF1@10F1@20
CLIP15.133.316.081.113.47.2
nCLIP16.135.616.681.411.66.4
xCLIP15.335.216.881.213.87.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.513, + 0.47, + 0.556 + ], + "angle": 0, + "content": "Table 3. Zero-shot multi-label classification. We report mAP and F1 scores with ViT-B/16 pre-trained on IT35M. nCLIP showcases the best multi-label classification capability." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.562, + 0.471, + 0.79 + ], + "angle": 0, + "content": "Multi-label classification. We evaluate zero-shot multi-label classification on NUS-WIDE [23] and OpenImages [46] following standard protocol [39]. Specifically, we use 81 unseen labels for NUS-WIDE and the most frequent 400 unseen test labels for OpenImages during evaluation, following [40]. As shown in Tab. 3, we observe that nCLIP shows the best mAP compared to CLIP. nCLIP is pre-trained without negative examples, resulting in the model's intrinsic strengths on recall over overlapped concepts, with 16.1 and 81.4 points of mAP on NUS-WIDE and OpenImages, respectively. Comparatively, xCLIP slightly improves CLIP in this respect while lagging behind nCLIP. The experiments demonstrate that the non-contrastive objective fits well for those downstream tasks in demand of high recall." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Retrieval. We evaluate on 2 retrieval benchmarks: Flickr30K [58] and MSCOCO [51] under zero-shot protocol. We do not use prompt engineering and use the original caption. Compared to zero-shot classification, image-text retrieval requires the model's recognition capability at a fine-grained level. We empirically observe in Tab. 4 that" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.25, + 0.895, + 0.417 + ], + "angle": 0, + "content": "nCLIP performs drastically worse than CLIP, which is because text captions in retrieval are not mutually exclusive, which is different from label names in classification. Therefore, explicit negative examples during pre-training play an imperative role in zero-shot retrieval. See Appendix E.1 for further discussions. Beyond that, xCLIP achieves a noticeable performance improvement over CLIP, with a gain of \\(3.7\\%\\) R@1 on Flickr30K and \\(3.9\\%\\) R@1 on MSCOCO, illustrating that the additional non-contrastive objective renders extra semantic signals that can be transferred well to fine-grained recognition." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.425, + 0.73, + 0.442 + ], + "angle": 0, + "content": "4.2. Representation Learning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.448, + 0.894, + 0.495 + ], + "angle": 0, + "content": "We conduct several evaluation protocols to benchmark representation quality under different objectives for both visual (Sec. 4.2.1) and textual (Sec. 4.2.2) modality." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.512, + 0.779, + 0.528 + ], + "angle": 0, + "content": "4.2.1 Visual Representation Learning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.535, + 0.895, + 0.853 + ], + "angle": 0, + "content": "Linear probing. We evaluate the quality of visual representation via linear probing protocol, where a linear head is fine-tuned on top of the frozen backbones. We follow the same setup as [20] on the same 27 datasets as standard zero-shot classification. Specifically, we use SGD without momentum as the optimizer, no weight decay, and a total epoch of 100 for all evaluation datasets. We use [CLS] token for classification. Following [82], we sweep a set of different learning rates by adding multiple classification heads over a shared frozen backbone, each with its own optimizer and scheduler. We report the best result across different heads. As shown in Tab. 5, we observe that nCLIP generally leads to comparable linear probing accuracy compared with CLIP, suggesting that non-contrastive objectives are able to derive semantically meaningful embedding spaces. nCLIP achieves \\(2.5\\%\\) higher performance compared to CLIP on ImageNet but relatively lower on average. Beyond that, when combining two objectives together, xCLIP achieves \\(2.7\\%\\) higher performance on ImageNet and \\(1.5\\%\\) higher performance on average across 27 different classification tasks compared to CLIP." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Fine-tuning & semi-supervised learning. We fine-tune the entire network under the full-data regime (100%) and" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11032" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.081, + 0.089, + 0.887, + 0.201 + ], + "angle": 0, + "content": "
ModelFlickr30KMSCOCO
I→TT→II→TT→I
R@1R@5R@10R@1R@5R@10R@1R@5R@10R@1R@5R@10
CLIP73.893.397.152.979.387.759.485.291.934.362.774.0
nCLIP26.253.367.920.148.561.121.447.961.613.437.150.3
xCLIP77.595.697.757.382.989.063.387.594.138.466.076.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.205, + 0.893, + 0.234 + ], + "angle": 0, + "content": "Table 4. Zero-shot image-to-text retrieval. We report R@1, R@5, and R@10 in both image-to-text (I→T) and text-to-image (T→I) settings with ViT-B/16 pre-trained on IT35M. xCLIP consistently improves CLIP in retrieval." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.246, + 0.887, + 0.298 + ], + "angle": 0, + "content": "
ModelFood101CIFAR10CIFAR100BirdsnapSUN397CarsAircraftVOC2007DTDPetsCaltech101FlowersMNISTFER2013STL10EuroSATRESISC45GTSRBKITTICountry211PCAMUCF101Kinetics,700CLEVRHaarSoftMnesSSTImageNetAverage
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.086, + 0.299, + 0.884, + 0.337 + ], + "angle": 0, + "content": "CLIP 83.3 92.0 75.6 54.4 78.8 60.5 33.8 91.1 72.4 79.0 90.0 96.1 96.2 56.5 97.1 94.8 90.3 77.3 69.7 23.2 83.4 78.9 53.9 50.3 55.9 55.8 71.4 |72.7 nCLIP 83.8 93.2 76.3 66.8 79.3 49.3 28.0 90.1 70.6 70.2 90.5 95.0 95.4 54.7 97.4 91.9 88.2 75.7 64.0 22.7 84.2 79.9 51.0 44.6 55.6 55.2 73.9 |71.4 xCLIP 85.3 93.4 77.8 58.4 80.7 62.3 36.7 92.3 74.0 81.5 91.6 97.0 96.8 58.5 98.1 95.3 91.2 80.4 69.3 26.3 83.5 81.1 56.0 49.8 58.5 54.8 74.1 |74.2" + }, + { + "type": "table_caption", + "bbox": [ + 0.08, + 0.34, + 0.893, + 0.369 + ], + "angle": 0, + "content": "Table 5. Linear probing. We report on a variety of classification benchmarks with ViT-B/16 pre-trained on IT35M. CLIP achieves consistent performance gains compared to CLIP in a wide range of classification datasets." + }, + { + "type": "table", + "bbox": [ + 0.086, + 0.394, + 0.457, + 0.484 + ], + "angle": 0, + "content": "
Modelsm. sup.ft.
1%10%100%
CLIP57.872.482.4
nCLIP55.072.282.4
xCLIP59.573.482.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.488, + 0.47, + 0.571 + ], + "angle": 0, + "content": "Table 6. Fine-tuning (ft.) and semi-supervised learning (sm. sup.). Percentage shows sampling ratio for end-to-end fine-tuning. We report top-1 accuracy on the ImageNet-1K validation set. nCLIP performs comparatively with CLIP. xCLIP induces consistent gains in all sampling ratios, especially for semi-supervised learning." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.588, + 0.462, + 0.683 + ], + "angle": 0, + "content": "
ModelSTS12STS13STS14STS15STS16STS-BSICK-RAverage
CLIP52.354.053.969.958.667.069.360.7
nCLIP51.362.756.971.965.866.964.162.8
xCLIP48.755.655.669.759.566.370.260.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.686, + 0.47, + 0.743 + ], + "angle": 0, + "content": "Table 7. Sentence embedding performance on semantic textual similarity. We report on a variety of NLP understanding benchmarks. xCLIP demonstrates the best capability for various STS tasks. xCLIP has comparable performance with CLIP." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.903 + ], + "angle": 0, + "content": "the semi-supervised learning protocol (1% and 10%) on ImageNet-1K [24]. Results are shown in Tab. 6. For CLIP and xCLIP, we find that fine-tuning with the projection head trained under the contrastive objective yields consistent performance gain, especially under a low-data regime. Specifically, we take the text features of the label names as the classifier's initialized parameters. We opt for this setup by default. We highlight that while nCLIP performs worse than CLIP in terms of zero-shot transfer, the non-contrastive objective also serves as a strong baseline to learn represent" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.395, + 0.893, + 0.456 + ], + "angle": 0, + "content": "tation considering the close gap between CLIP and nCLIP in terms of fine-tuning accuracy. xCLIP achieves an \\(1.7\\%\\), \\(1.0\\%\\), and \\(0.3\\%\\) performance gain compared with CLIP under three data ratios of \\(1\\%\\), \\(10\\%\\), and \\(100\\%\\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.477, + 0.893, + 0.72 + ], + "angle": 0, + "content": "Mask probing. To evaluate how well the visual model is capable of deriving explicit scene layout and object boundaries, we conduct mask probing analysis following [15]. For each attention head from the last layer, we extract the attention map with [CLS] token as the query. We then compute the Jaccard similarity \\(\\mathcal{J}\\) of each head's attention mask to the ground truth and retain the attention mask with the highest similarity. We conduct experiments on Pascal VOC 2012 [28] dataset. With IT35M, nCLIP demonstrates better quality in the generated mask with 43.7 points of \\(\\mathcal{J}\\), while CLIP reaches 41.2 and xCLIP reaches 41.9 points of \\(\\mathcal{J}\\). Hence, nCLIP learns better representation for object boundaries compared to CLIP, while xCLIP strikes a balance between nCLIP and CLIP. Detailed results are delayed to Appendix D.1. As a related evaluation, we also study unsupervised segmentation with GroupViT [77] in Appendix D.2." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.74, + 0.785, + 0.757 + ], + "angle": 0, + "content": "4.2.2 Textual Representation Learning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.894, + 0.903 + ], + "angle": 0, + "content": "We follow the setup as [29] and evaluate on 7 STS tasks: STS 2012-2016 [1-5], STS Benchmark [16], and SICK-Relatedness [34]. We directly take [EOS] token without any projection as the input for evaluation, which consistently yields better performance compared to the projected feature for all models. We use cosine distance as the similarity metric. The main goal of sentence embeddings is to cluster semantically similar sentences, and hence, we take STS as one yardstick to benchmark textual representation" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11033" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.084, + 0.09, + 0.462, + 0.196 + ], + "angle": 0, + "content": "
DataSizeEp.ZS / LN Accuracy
CLIPnCLIPxCLIP
①12M2536.8 / 68.537.5 / 71.042.4 / 72.2
④14M3226.5 / 75.131.4 / 77.027.1 / 77.2
①②22M3237.9 / 68.234.2 / 72.343.0 / 71.9
①②③35M3245.7 / 71.437.0 / 73.948.8 / 74.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.199, + 0.47, + 0.256 + ], + "angle": 0, + "content": "Table 8. Pre-training data scaling & domain. Ep. denotes training epochs. Data abbreviations are as follows. ①: CC12M. ②: COCO, VG, SBU, and CC3M. ③: YFCC14M. ④: IN21K. Note ①②③ = IT35M." + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.276, + 0.268, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.084, + 0.419, + 0.268, + 0.474 + ], + "angle": 0, + "content": "Figure 2. Batch size scaling. xCLIP performs better than CLIP with a small batch size (i.e., 1024)." + }, + { + "type": "image", + "bbox": [ + 0.282, + 0.276, + 0.462, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.278, + 0.419, + 0.462, + 0.476 + ], + "angle": 0, + "content": "Figure 3. Training time. xCLIP adds \\(\\sim 1.3x\\) training time cost and performs better with equal seen images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.482, + 0.47, + 0.619 + ], + "angle": 0, + "content": "learning. Results are shown in Tab. 7. nCLIP shows better textual semantics compared to CLIP, with \\(2.1\\%\\) higher in terms of the average score with IT35M. The conclusion can also be visually extrapolated from the t-SNE visualization of the word embedding, as detailed in Appendix G. We note that xCLIP performs comparatively with CLIP given that STS is evaluated on [EOS] without projection, thus the backbone representation can be similar if dominant by the contrastive objective." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.629, + 0.194, + 0.645 + ], + "angle": 0, + "content": "4.3. Properties" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.653, + 0.469, + 0.714 + ], + "angle": 0, + "content": "We analyze several crucial properties to demonstrate the generalizability and effectiveness of our method in practical usage. Experiments are conducted with ViT-B/16 on CC12M for 25 epochs by default." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Pre-training data scaling. We study the model's performance with data scaling in Tab. 8. xCLIP exhibits desirable scaling behaviors similar to CLIP. The performance improves as data increases. However, we do not observe such a tendency in nCLIP in terms of zero-shot accuracy. For example, nCLIP achieves \\(37.5\\%\\) with (1) but only \\(34.2\\%\\) with (2), the trend of which defies those of CLIP's \\((36.8\\%\\) vs. \\(37.9\\%)\\) and xCLIP's \\((42.4\\%\\) vs. \\(43.0\\%)\\). On an important note, both nCLIP and xCLIP showcase desirable scaling behaviors under the linear probing protocol. We hypothesize that the performance of zero-shot classification depends on" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.138 + ], + "angle": 0, + "content": "pre-training data quality with the non-contrastive objective, while strong representation can always be ensured before the projection." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.158, + 0.893, + 0.4 + ], + "angle": 0, + "content": "Pre-training with tagging data. We consider using tagging data, e.g., ImageNet-21K [24], as pre-training data to validate method's generalizability. To do that, we take label names as annotation texts by concatenating them with a sampled prompt, similar to [57, 80]. As shown in Tab. 8, nCLIP shows superiority to CLIP when pre-trained with IN21K by performance gain of \\(4.9\\%\\) and \\(1.9\\%\\) in terms of zero-shot and linear probing accuracy, respectively. Compared to nCLIP, we instead notice a performance drop of \\(4.3\\%\\) for zero-shot accuracy when two objectives multitasked, indicating that the contrastive objective stymies the quality of projected probability under tagging data. Note, that an \\(77.2\\%\\) linear probing accuracy achieved by xCLIP, \\(2.1\\%\\) higher than CLIP, is decently strong compared to \\(79.0\\%\\) achieved by the state of the art in SSL, iBOT [82], pre-trained using the same data with 80 epochs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.421, + 0.893, + 0.634 + ], + "angle": 0, + "content": "Batch size scaling. We investigate the scaling behavior of CLIP and xCLIP with different pre-training batch sizes. As shown in Fig. 2, xCLIP performs stronger with smaller batch sizes than CLIP pre-trained even with larger batch sizes. For example, xCLIP pre-trained with a batch size of 1024 achieves a zero-shot accuracy of \\(38.5\\%\\) (vs. CLIP's \\(36.8\\%\\) with a batch size of 4096). However, we still observe a substantial performance drop as batch size decreases, the behavior of which is different from self-supervised models pre-trained with the non-contrastive objective [14, 15] that are insensitive to batch size. We hypothesize that a large batch size is necessary to approximate \\(\\mathbb{E}(p)\\) with \\(\\frac{1}{B}\\sum_{i=1}^{B} p_i\\) in estimating \\(\\mathcal{L}_{\\mathrm{HE}}\\), especially when two probabilities are derived from drastically different modalities." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.655, + 0.892, + 0.761 + ], + "angle": 0, + "content": "Computation efficiency. To demonstrate methods' computation efficiency, we show the FLOPs and GPU memory consumption of xCLIP compared to CLIP. xCLIP enforces bearable additional computation cost on top of CLIP with only \\(1.4\\%\\) extra FLOPs and \\(27\\%\\) extra GPU memory. As shown in Fig. 3, xCLIP adds \\(\\sim 1.3x\\) extra time cost but with a \\(6.0\\%\\) performance gain in terms of zero-shot accuracy." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.771, + 0.655, + 0.788 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.903 + ], + "angle": 0, + "content": "We show in this section the crucial composing factors of nCLIP and xCLIP. The ablations in the first column (Tabs. 9a to 9c) are conducted with \\(\\mathcal{L}_{\\mathrm{nCLIP}}\\) only. The ablations in the second column (Tabs. 9d to 9f) are conducted with full loss \\(\\mathcal{L}_{\\mathrm{xCLIP}}\\). Experiments are conducted with ViTB/16 on CC12M for 25 epochs. The default settings are highlighted in cyan." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11034" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.107, + 0.088, + 0.322, + 0.189 + ], + "angle": 0, + "content": "
λ1λ2ZSLN
01NaN
0.3126.869.5
0.5123.767.2
0.21.235.369.0
0.51.537.571.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.086, + 0.191, + 0.343, + 0.216 + ], + "angle": 0, + "content": "(a) Coefficient of entropy regularizer. \\(\\lambda_1 + 1 = \\lambda_2\\) generally leads to good performance." + }, + { + "type": "table", + "bbox": [ + 0.104, + 0.22, + 0.325, + 0.305 + ], + "angle": 0, + "content": "
#SL / #TLZSLN
0 / 142.472.2
0 / 241.371.3
1 / 242.471.9
2 / 342.572.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.307, + 0.35, + 0.332 + ], + "angle": 0, + "content": "(d) Shared projection layers. #SL denotes shared hidden layers. #TL denotes total hidden layers." + }, + { + "type": "table", + "bbox": [ + 0.373, + 0.089, + 0.591, + 0.174 + ], + "angle": 0, + "content": "
dimZSLNMem
819231.070.61.07×
1638435.170.91.13×
3276837.571.01.27×
6553635.971.01.52×
" + }, + { + "type": "table_caption", + "bbox": [ + 0.35, + 0.19, + 0.62, + 0.216 + ], + "angle": 0, + "content": "(b) Projection dimension. Mem (G) is compared to CLIP with a 512-dim embedding layer." + }, + { + "type": "table", + "bbox": [ + 0.373, + 0.22, + 0.596, + 0.305 + ], + "angle": 0, + "content": "
λCLIP: λnCLIPZSLN
1.040.266.5
0.541.770.3
0.242.472.2
0.141.872.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.367, + 0.307, + 0.604, + 0.32 + ], + "angle": 0, + "content": "(e) Loss ratio between \\(\\mathcal{L}_{\\mathrm{CLIP}}\\) and \\(\\mathcal{L}_{\\mathrm{nCLIP}}\\)" + }, + { + "type": "table", + "bbox": [ + 0.633, + 0.089, + 0.876, + 0.174 + ], + "angle": 0, + "content": "
archZSLN
vanilla37.571.0
w/ bottleneck layer31.269.0
w/ l2-norm layer35.470.7
w/o last BN layer34.369.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.627, + 0.19, + 0.885, + 0.214 + ], + "angle": 0, + "content": "(c) Projection architecture of non-contrastive branch." + }, + { + "type": "table", + "bbox": [ + 0.627, + 0.22, + 0.886, + 0.305 + ], + "angle": 0, + "content": "
techniqueZSLN
vanilla42.472.2
shared space (Eq. (3))29.468.9
debiased sampling42.272.4
warm-up on LnCLIP41.771.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.627, + 0.307, + 0.885, + 0.332 + ], + "angle": 0, + "content": "(f) Training objective and technique for multitasking." + }, + { + "type": "table_caption", + "bbox": [ + 0.081, + 0.334, + 0.886, + 0.349 + ], + "angle": 0, + "content": "Table 9. Ablation study with ViT-B/16 on ImageNet-1K validation set. We report zero-shot (ZS) and linear probing (LN) accuracy (\\%)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.375, + 0.473, + 0.589 + ], + "angle": 0, + "content": "Entropy regularizer. We study the effects of coefficient for entropy regularizers \\(\\lambda_{1}\\) and \\(\\lambda_{2}\\) in Tab. 9a. The pretraining is unstable without regularizers or with only maximization of the entropy of the mean \\((\\mathcal{L}_{\\mathrm{HE}})\\). Specifically, the model will collapse to constant uniform distribution and outputs the same uniform probability distribution despite different inputs. The additional minimization of the mean of the entropy \\((\\mathcal{L}_{\\mathrm{EH}})\\) stabilizes training, but incurs dimensional collapse, eroding the performances (26.8% and 23.7% vs. 37.5% in terms of zero-shot accuracy). Simultaneously adjusting \\(\\lambda_{2}\\) with constraints \\(\\lambda_{1} + 1 = \\lambda_{2}\\) and \\(\\lambda_{1} = 0.5\\) mitigates the problem and yields the optimal performance, with a 37.5% zero-shot and 71.0% linear-probing accuracy. Details are shown in Appendix B.1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.608, + 0.47, + 0.805 + ], + "angle": 0, + "content": "Projection head. We study the design of two projection heads for \\(\\mathcal{L}_{\\mathrm{CLIP}}\\) and \\(\\mathcal{L}_{\\mathrm{nCLIP}}\\). As shown in Tab. 9b, large projection dimension matters for good performance. We opt for a projection dimension of 32768 balancing the performance and speed. As shown in Tab. 9c, the last BN layer is crucial and leads to more stable optimization in our experiments. We draw on the idea of prototypes [9, 15] by introducing bottleneck and 12-norm layers, while they do not yield a performance gain. We further study whether two projections can share intermediate computation in Tab. 9d. While sharing middle hidden layers with slightly deeper projection heads performs better, we opt for 2-layer MLP without shared layers for its simplicity." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Optimization. We study the effect of hyper-parameters in optimization. In Tab. 9e, we study the optimal loss ratio between \\(\\mathcal{L}_{\\mathrm{CLIP}}\\) and \\(\\mathcal{L}_{\\mathrm{nCLIP}}\\). We further study whether it's the best route for two objectives to be optimized in separate latent spaces via multi-tasking. Specifically, we con" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.375, + 0.892, + 0.557 + ], + "angle": 0, + "content": "sider a bespoke objective Eq. (3) taking account of probability estimation and negative samples simultaneously in a shared latent space. Detailed formulation are shown in Appendix B.2. Optimizing in one latent space yields suboptimal results with a nearly \\(13.0\\%\\) drop in zero-shot accuracy and \\(3.3\\%\\) in linear probing accuracy, suggesting two objectives are intrinsically contradictory. We consider debiased sampling, where each batch is sampled from a single data source, since their semantics may be closer and thus more suitable for optimization. This yields similar results. We also add \\(\\mathcal{L}_{\\mathrm{nCLIP}}\\) after warm-up epochs with the linear scheduler to its base scale, which erodes the performance." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.57, + 0.619, + 0.587 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.595, + 0.892, + 0.898 + ], + "angle": 0, + "content": "Aligning images and texts is of overriding significance for vision-language understanding. To conquer the systematic insufficiency of the contrastive objective for acquiring semantics and tackling loose correlations between noisy image-text pairs, we explore the non-contrastive objective for language-image pre-training and unravel its properties. Empirical evidence reveals that the non-contrastive objective induces models to perform favorably in representation learning yet poorly in zero-shot transfer. Observing the distinct mechanisms of the two objectives, we further seek synergy between the two, and introduce a simple multi-tasking framework, xCLIP, that enjoys the best of both worlds: nCLIP aids CLIP mining semantics while CLIP inherits intrinsic strengths for zero-shot recognition. The consistent performance gain of xCLIP over CLIP on a wide variety of downstream tasks consolidates our findings. As potential future work, we may continue to scale up the data size (e.g., LAION400M [64]) as well as the model size (e.g., ViT-L [26]) to verify whether the scaling law applies and the performance improvement endures." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11035" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Eneko Agirre, Carmen Banea, Claire Cardie, Daniel Cer, Mona Diab, Aitor Gonzalez-Agirre, Weiwei Guo, Inigo Lopez-Gazpio, Montse Maritxalar, Rada Mihalcea, et al. Semeval-2015 task 2: Semantic textual similarity, english, spanish and pilot on interpretability. In SemEval, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.186, + 0.472, + 0.253 + ], + "angle": 0, + "content": "[2] Eneko Agirre, Carmen Banea, Claire Cardie, Daniel M Cer, Mona T Diab, Aitor Gonzalez-Agirre, Weiwei Guo, Rada Mihalcea, German Rigau, and Janyce Wiebe. Semeval-2014 task 10: Multilingual semantic textual similarity. In SemEval, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.255, + 0.472, + 0.323 + ], + "angle": 0, + "content": "[3] Eneko Agirre, Carmen Banea, Daniel Cer, Mona Diab, Aitor Gonzalez Agirre, Rada Mihalcea, German Rigau Claramunt, and Janyce Wiebe. Semeval-2016 task 1: Semantic textual similarity, monolingual and cross-lingual evaluation. In SemEval, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.326, + 0.47, + 0.366 + ], + "angle": 0, + "content": "[4] Eneko Agirre, Daniel Cer, Mona Diab, and Aitor Gonzalez-Agirre. Semeval-2012 task 6: A pilot on semantic textual similarity. In SemEval, 2012. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.368, + 0.47, + 0.409 + ], + "angle": 0, + "content": "[5] Eneko Agirre, Daniel Cer, Mona Diab, Aitor Gonzalez-Agirre, and Weiwei Guo. *sem 2013 shared task: Semantic textual similarity. In SemEval, 2013. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.41, + 0.47, + 0.478 + ], + "angle": 0, + "content": "[6] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. arXiv preprint arXiv:2204.14198, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.48, + 0.47, + 0.507 + ], + "angle": 0, + "content": "[7] Elad Amrani and Alex Bronstein. Self-supervised classification network. In ECCV, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.509, + 0.47, + 0.564 + ], + "angle": 0, + "content": "[8] Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Florian Bordes, Pascal Vincent, Armand Joulin, Michael Rabbat, and Nicolas Ballas. Masked siamese networks for label-efficient learning. In ECCV, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.565, + 0.47, + 0.633 + ], + "angle": 0, + "content": "[9] Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Armand Joulin, Nicolas Ballas, and Michael Rabbat. Semi-supervised learning of visual features by nonparametrically predicting view assignments with support samples. In ICCV, 2021. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.635, + 0.47, + 0.703 + ], + "angle": 0, + "content": "[10] Andrei Barbu, David Mayo, Julian Alverio, William Luo, Christopher Wang, Dan Gutfreund, Josh Tenenbaum, and Boris Katz. Objectnet: A large-scale bias-controlled dataset for pushing the limits of object recognition models. In NeurIPS, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.705, + 0.47, + 0.746 + ], + "angle": 0, + "content": "[11] Adrien Bardes, Jean Ponce, and Yann LeCun. Vicreg: Variance-invariance-covariance regularization for self-supervised learning. In ICLR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.747, + 0.47, + 0.802 + ], + "angle": 0, + "content": "[12] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. In NeurIPS, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.47, + 0.843 + ], + "angle": 0, + "content": "[13] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features. In ECCV, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[14] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. In NeurIPS, 2020. 2, 3, 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[15] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[16] Daniel Cer, Mona Diab, Eneko Agirre, Inigo Lopez-Gazpio, and Lucia Specia. Semeval-2017 task 1: Semantic textual similarity-multilingual and cross-lingual focused evaluation. In SemEval, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.208, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[17] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing web-scale image-text pretraining to recognize long-tail visual concepts. In CVPR, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.265, + 0.892, + 0.306 + ], + "angle": 0, + "content": "[18] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.309, + 0.892, + 0.337 + ], + "angle": 0, + "content": "[19] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In CVPR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.339, + 0.892, + 0.379 + ], + "angle": 0, + "content": "[20] Xinlei Chen*, Saining Xie*, and Kaiming He. An empirical study of training self-supervised vision transformers. In ICCV, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.383, + 0.892, + 0.424 + ], + "angle": 0, + "content": "[21] Yanbei Chen, Yongqin Xian, Sophia Koepke, Ying Shan, and Zeynep Akata. Distilling audio-visual knowledge by compositional contrastive learning. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.426, + 0.892, + 0.48 + ], + "angle": 0, + "content": "[22] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Universal image-text representation learning. In ECCV, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.484, + 0.892, + 0.538 + ], + "angle": 0, + "content": "[23] Tat-Seng Chua, Jinhui Tang, Richang Hong, Haojie Li, Zhiping Luo, and Yan-Tao Zheng. Nus-wide: A real-world web image database from national university of Singapore. In CIVR, 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.542, + 0.892, + 0.582 + ], + "angle": 0, + "content": "[24] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. In CVPR, 2009. 2, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.585, + 0.892, + 0.613 + ], + "angle": 0, + "content": "[25] Karan Desai and Justin Johnson. Virtex: Learning visual representations from textual annotations. In CVPR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.615, + 0.892, + 0.696 + ], + "angle": 0, + "content": "[26] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.7, + 0.892, + 0.742 + ], + "angle": 0, + "content": "[27] Aleksandr Ermolov, Aliaksandr Siarohin, Enver Sangineto, and Nicu Sebe. Whitening for self-supervised representation learning. In ICML, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.744, + 0.892, + 0.785 + ], + "angle": 0, + "content": "[28] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. IJCV, 2010. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.788, + 0.892, + 0.828 + ], + "angle": 0, + "content": "[29] Tianyu Gao, Xingcheng Yao, and Danqi Chen. SimCSE: Simple contrastive learning of sentence embeddings. In EMNLP, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.831, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[30] Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A Wichmann, and Wieland Brendel. Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. In ICLR, 2019. 4" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11036" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[31] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. In NeurIPS, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.162, + 0.469, + 0.204 + ], + "angle": 0, + "content": "[32] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. In CVPR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.205, + 0.469, + 0.245 + ], + "angle": 0, + "content": "[33] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.247, + 0.469, + 0.314 + ], + "angle": 0, + "content": "[34] Matthew Henderson, Rami Al-Rfou, Brian Strope, Yun-Hsuan Sung, László Lukács, Ruiqi Guo, Sanjiv Kumar, Balint Miklos, and Ray Kurzweil. Efficient natural language response suggestion for smart reply. arXiv preprint arXiv:1705.00652, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.316, + 0.469, + 0.383 + ], + "angle": 0, + "content": "[35] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robustness: A critical analysis of out-of-distribution generalization. In ICCV, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.385, + 0.469, + 0.425 + ], + "angle": 0, + "content": "[36] Dan Hendrycks and Kevin Gimpel. Bridging nonlinearities and stochastic regularizers with gaussian error linear units. In ICLR, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.427, + 0.469, + 0.468 + ], + "angle": 0, + "content": "[37] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. In CVPR, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.469, + 0.469, + 0.523 + ], + "angle": 0, + "content": "[38] R Devon Hjelm, Alex Fedorov, Samuel Lavoie-Marchildon, Karan Grewal, Phil Bachman, Adam Trischler, and Yoshua Bengio. Learning deep representations by mutual information estimation and maximization. In ICLR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.525, + 0.469, + 0.564 + ], + "angle": 0, + "content": "[39] Dat Huynh and Ehsan Elhamifar. A shared multi-attention framework for multi-label zero-shot learning. In CVPR, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.567, + 0.469, + 0.594 + ], + "angle": 0, + "content": "[40] D. Huynh and E. Elhamifar. A shared multi-attention framework for multi-label zero-shot learning. In CVPR, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.595, + 0.469, + 0.635 + ], + "angle": 0, + "content": "[41] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In ICML, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.637, + 0.469, + 0.691 + ], + "angle": 0, + "content": "[42] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.692, + 0.469, + 0.733 + ], + "angle": 0, + "content": "[43] Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution or region supervision. In ICML, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.735, + 0.469, + 0.774 + ], + "angle": 0, + "content": "[44] Thomas Kipf, Elise van der Pol, and Max Welling. Contrastive learning of structured world models. arXiv preprint arXiv:1911.12247, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.469, + 0.844 + ], + "angle": 0, + "content": "[45] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. IJCV, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[46] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4. IJCV, 2020. 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[47] Janghyeon Lee, Jongsuk Kim, Hyounguk Shon, Bumsoo Kim, Seung Hwan Kim, Honglak Lee, and Junmo Kim. Uniclip: Unified framework for contrastive language-image pretraining. In NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[48] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. In NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.205, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[49] Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. Visualbert: A simple and performant baseline for vision and language. In ACL, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.247, + 0.892, + 0.313 + ], + "angle": 0, + "content": "[50] Yangguang Li, Feng Liang, Lichen Zhao, Yufeng Cui, Wanli Ouyang, Jing Shao, Fengwei Yu, and Junjie Yan. Supervision exists everywhere: A data efficient contrastive language-image pre-training paradigm. In ICLR, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.316, + 0.892, + 0.37 + ], + "angle": 0, + "content": "[51] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.372, + 0.892, + 0.412 + ], + "angle": 0, + "content": "[52] Xingbin Liu, Jinghao Zhou, Tao Kong, Xianming Lin, and Rongrong Ji. Exploring target representations for masked autoencoders. arXiv preprint arXiv:2209.03917, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.414, + 0.892, + 0.44 + ], + "angle": 0, + "content": "[53] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In ICLR, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.442, + 0.892, + 0.482 + ], + "angle": 0, + "content": "[54] Pranay Manocha, Zeyu Jin, Richard Zhang, and Adam Finkelstein. Cdpam: Contrastive learning for perceptual audio similarity. In ICASSP, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.484, + 0.892, + 0.524 + ], + "angle": 0, + "content": "[55] Norman Mu, Alexander Kirillov, David Wagner, and Saining Xie. Slip: Self-supervision meets language-image pretraining. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.526, + 0.892, + 0.566 + ], + "angle": 0, + "content": "[56] Vicente Ordonez, Girish Kulkarni, and Tamara Berg. Im2text: Describing images using 1 million captioned photographs. In NeuIPS, 2011. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.568, + 0.892, + 0.622 + ], + "angle": 0, + "content": "[57] Hieu Pham, Zihang Dai, Golnaz Ghiasi, Hanxiao Liu, Adams Wei Yu, Minh-Thang Luong, Mingxing Tan, and Quoc V Le. Combined scaling for zero-shot transfer learning. arXiv preprint arXiv:2111.10050, 2021. 1, 2, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.624, + 0.892, + 0.69 + ], + "angle": 0, + "content": "[58] Bryan A Plummer, Liwei Wang, Chris M Cervantes, Juan C Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In ICCV, 2015. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.692, + 0.892, + 0.761 + ], + "angle": 0, + "content": "[59] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.762, + 0.892, + 0.815 + ], + "angle": 0, + "content": "[60] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.818, + 0.892, + 0.858 + ], + "angle": 0, + "content": "[61] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.86, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[62] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. Do imagenet classifiers generalize to imageNet? In ICML, 2019. 4" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "11037" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.134 + ], + "angle": 0, + "content": "[63] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In NeurIPS, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.469, + 0.207 + ], + "angle": 0, + "content": "[64] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. In NeurIPS Workshops, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.208, + 0.469, + 0.262 + ], + "angle": 0, + "content": "[65] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In ACL, 2018. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.265, + 0.469, + 0.307 + ], + "angle": 0, + "content": "[66] Aravind Srinivas, Michael Laskin, and Pieter Abbeel. Curl: Contrastive unsupervised representations for reinforcement learning. In ICML, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.309, + 0.469, + 0.351 + ], + "angle": 0, + "content": "[67] Weijie Su, Xizhou Zhu, Yue Cao, Bin Li, Lewei Lu, Furu Wei, and Jifeng Dai. Vl-bert: Pre-training of generic visual-linguistic representations. In ICLR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.353, + 0.469, + 0.394 + ], + "angle": 0, + "content": "[68] Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. In EMNLP, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.396, + 0.469, + 0.451 + ], + "angle": 0, + "content": "[69] Bart Thomee, David A Shamma, Gerald Friedland, Benjamin Elizalde, Karl Ni, Douglas Poland, Damian Borth, and Li-Jia Li. Yfcc100m: The new data in multimedia research. In CACM, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.453, + 0.469, + 0.522 + ], + "angle": 0, + "content": "[70] Yael Vinker, Ehsan Pajouheshgar, Jessica Y Bo, Roman Christian Bachmann, Amit Haim Bermano, Daniel Cohen-Or, Amir Zamir, and Ariel Shamir. Clipasso: Semantically-aware object sketching. In SIGGRAPH, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.525, + 0.469, + 0.568 + ], + "angle": 0, + "content": "[71] Feng Wang, Tao Kong, Rufeng Zhang, Huaping Liu, and Hang Li. Self-supervised learning by estimating twin class distributions. arXiv preprint arXiv:2110.07402, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.569, + 0.469, + 0.611 + ], + "angle": 0, + "content": "[72] Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. Learning robust global representations by penalizing local predictive power. In NeurIPS, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.612, + 0.469, + 0.655 + ], + "angle": 0, + "content": "[73] Zirui Wang, Jiahui Yu, Adams Wei Yu, Zihang Dai, Yulia Tsvetkov, and Yuan Cao. Simvlm: Simple visual language model pretraining with weak supervision. In ICLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.656, + 0.469, + 0.712 + ], + "angle": 0, + "content": "[74] Bichen Wu, Ruizhe Cheng, Peizhao Zhang, Tianren Gao, Joseph E. Gonzalez, and Peter Vajda. Data efficient language-supervised zero-shot recognition with optimal transport distillation. In ICLR, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.714, + 0.469, + 0.755 + ], + "angle": 0, + "content": "[75] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In CVPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.757, + 0.469, + 0.799 + ], + "angle": 0, + "content": "[76] Qizhe Xie, Minh-Thang Luong, Eduard Hovy, and Quoc V Le. Self-training with noisy student improves imagenet classification. In CVPR, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.801, + 0.469, + 0.856 + ], + "angle": 0, + "content": "[77] Jiarui Xu, Shalini De Mello, Sifei Liu, Wonmin Byeon, Thomas Breuel, Jan Kautz, and Xiaolong Wang. Groupvit: Semantic segmentation emerges from text supervision. In CVPR, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.858, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[78] Jianwei Yang, Chunyuan Li, Pengchuan Zhang, Bin Xiao, Ce Liu, Lu Yuan, and Jianfeng Gao. Unified contrastive learning in image-text-label space. In CVPR, 2022. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[79] Lewei Yao, Runhui Huang, Lu Hou, Guansong Lu, Minzhe Niu, Hang Xu, Xiaodan Liang, Zhenguo Li, Xin Jiang, and Chunjing Xu. Filip: Fine-grained interactive language-image pre-training. In ICLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[80] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917, 2022. 1, 2, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.206, + 0.892, + 0.248 + ], + "angle": 0, + "content": "[81] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow twins: Self-supervised learning via redundancy reduction. In ICML, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.291 + ], + "angle": 0, + "content": "[82] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. ibot: Image bert pre-training with online tokenizer. In ICLR, 2022. 5, 7" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "11038" + } + ] +] \ No newline at end of file diff --git a/2023/Non-Contrastive Learning Meets Language-Image Pre-Training/fe567c45-5ef2-4da5-8116-0c22fc6e398a_origin.pdf b/2023/Non-Contrastive Learning Meets Language-Image Pre-Training/fe567c45-5ef2-4da5-8116-0c22fc6e398a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2b53c5706ee0bcd035bf91ed6423b756a0161305 --- /dev/null +++ b/2023/Non-Contrastive Learning Meets Language-Image Pre-Training/fe567c45-5ef2-4da5-8116-0c22fc6e398a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:074b02ee4254815b7e8f1c0d4c5f7b07cb3d82c4842e0d06815add6b43a7bfda +size 458002 diff --git a/2023/Non-Contrastive Learning Meets Language-Image Pre-Training/full.md b/2023/Non-Contrastive Learning Meets Language-Image Pre-Training/full.md new file mode 100644 index 0000000000000000000000000000000000000000..dac199793671ce7a4c7878bc4b5ab3faecfba482 --- /dev/null +++ b/2023/Non-Contrastive Learning Meets Language-Image Pre-Training/full.md @@ -0,0 +1,409 @@ +# Non-Contrastive Learning Meets Language-Image Pre-Training + +Jinghao Zhou Li Dong Zhe Gan Lijuan Wang Furu Wei Microsoft + +# Abstract + +Contrastive language-image pre-training (CLIP) serves as a de-facto standard to align images and texts. Nonetheless, the loose correlation between images and texts of web-crawled data renders the contrastive objective data inefficient and craving for a large training batch size. In this work, we explore the validity of non-contrastive language-image pre-training (nCLIP), and study whether nice properties exhibited in visual self-supervised models can emerge. We empirically observe that the non-contrastive objective benefits representation learning while sufficiently underperforming under zero-shot recognition. Based on the above study, we further introduce xCLIP, a multi-tasking framework combining CLIP and nCLIP, and show that nCLIP aids CLIP in enhancing feature semantics. The synergy between two objectives lets xCLIP enjoy the best of both worlds: superior performance in both zero-shot transfer and representation learning. Systematic evaluation is conducted spanning a wide variety of downstream tasks including zero-shot classification, out-of-domain classification, retrieval, visual representation learning, and textual representation learning, showcasing a consistent performance gain and validating the effectiveness of xCLIP. The code and pre-trained models will be publicly available at https://github.com/shallowtoil/xclip. + +# 1. Introduction + +Language-image pre-training which simultaneously learns textual and visual representation from large-scale image-text pairs has revolutionized the field of representation learning [25, 59], vision-language understanding [22], and text-to-image generation [61]. Compared to traditional visual models, language-instilled ones intrinsically inherit the capability of zero-shot or few-shot learning prominently demonstrated by large language models such as GPT-3 [12]. The precursor system, Contrastive Language-Image Pre-Training [59] (CLIP) that explicitly aligns the projected features of two modalities, has demonstrated surprising capabilities of zero-shot, representation learning, and robustness, being applied to a wide range of fields [32, 52, 60, 70]. + +![](images/a0c87e3957ef3b12e91d888b2fdf06da99582a902e374957c0e45e5be91a978b.jpg) +Figure 1. Architecture comparison between CLIP and nCLIP. We take base-size encoders for an instance. CLIP discriminates instances within a batch using 512-dim projected embeddings. nCLIP projects each modality into a 32768-dim probability distribution as the pseudo-label to supervise the prediction from the other modality. Darker blocks of nCLIP depict a higher response for cluster distribution, signifying the clusters to which text or image instances may belong. xCLIP is a multi-tasking framework with both CLIP and nCLIP. + +To learn from noisy web-crawled image-text pairs, CLIP adopts a formulation of the contrastive objective, where the image and text within a pair are considered as unique instances and are encouraged to be discriminated from all the other negative instances. However, web-crawled image-text pairs [65,69] are usually loosely correlated in the sense that one caption (image) can match reasonably with multiple images (captions) besides the ground-truth one, as statistically shown in [74]. Hence, it is inaccurate and data-inefficient for representation learning to neglect other sensible matches and overlook the semantics hidden inside the textual description. This is also solidified by the undesirable discriminative ability or transferring performance of the visual encoder pre-trained under the contrastive objective, as suggested in [78]. Mining semantics from plentiful concepts appearing in captions, therefore, solicits further exploration beyond the vanilla contrastive objective. + +Previously, some works resort to mining nearest neighbor positive samples via intra-modality similarity [50, 74] but require extra storage for auxiliary modules, i.e., the teacher network [74] or the memory bank [50]. Other works conduct multi-tasking with image-label supervision [57, 78, 80], transforming the captions into tag format or the image tag into the captioning format for unified contrastive + +learning. Notwithstanding, such practice is still unable to account for the loose assignment within the image-text captioning pairs. Intuitively, a caption for an image usually identifies existing objects within the image, which can be captured by a probabilistic distribution estimating how the text is assigned to one or multiple object clusters [13, 14]. Such estimation depicts the semantic meanings for its contained visual contents, and thus can be leveraged as a pseudo-label to guide the learning of the visual encoder. + +Inspired by the superiority of visual self-supervised learning (SSL) models pre-trained with the non-contrastive objective [14, 15], we explore whether the non-contrastive objective can be used across modalities for pre-training language-image models, and whether nice properties displayed on visual SSL models can be inherited. To this end, we follow the same setup as CLIP except for the objective, and study non-Contrastive Language-Image Pre-Training (nCLIP). For an image-text pair, we use the estimation of the textual (visual) distribution as the target to supervise the visual (textual) distribution, measured by cross-entropy. Additional regularizers are applied to avoid trivial solutions. Schematic comparison between CLIP and nCLIP is shown in Fig. 1. Theoretically, such formulation takes one modality and the cluster centroid that the other modality belongs to as the positive samples to contrast [14], as opposed to direct features of two modalities as in contrastive learning, such that a single image is tolerated to be aligned with multiple captions and concepts. Based on the above formulation, we conduct a systematic study in terms of zero-shot transfer and representation learning. We empirically observe that, while nCLIP demonstrates desirable representation learning capability for both modalities, it underperforms prominently in zero-shot classification and retrieval tasks where models pre-trained with negative samples naturally prevail. + +Seeking for unique strengths of both worlds, we further perform multi-tasking of CLIP and nCLIP, short as xCLIP, and seek synergy between two distinct pre-training objectives. With the bearable add-on of computational resources (e.g., $\sim 27\%$ memory-wise and $\sim 30\%$ time-wise), xCLIP achieves consistent performance gain compared to CLIP on a wide range of tasks spanning from zero-shot classification, retrieval, linear probing, and fine-tuning, etc. An extensive study with different pre-training datasets, evaluation metrics, and optimization configurations is conducted to validate xCLIP's effectiveness in mining semantics and improving data efficiency. Particularly, the base-size model pre-trained using xCLIP under 35-million publicly available image-text pairs achieves a performance gain of $3.3\%$ and $1.5\%$ on an average of 27 classification tasks [59], in terms of zero-shot classification and linear probing accuracy, respectively. Performance gain is also valid in terms of zero-shot retrieval, semi-supervised learning, and fine + +tuning, with 3.7 points of R@1 on Flickr30K [58], $1.7\%$ accuracy on $1\%$ subset of ImageNet [24], and $0.3\%$ accuracy on ImageNet, respectively. + +# 2. Related Work + +Language-image pre-training. Language-image pretraining, aka vision-language pre-training (VLP), learns to jointly cope with vision and language input using multi-modality models. Early practices [22, 49, 67, 68] separately encode modalities, i.e., texts with linear embedding and images with convolutional neural networks and region proposals [63]. Recent works take direct images instead of pre-fetched region features as input, using e.g., dual-encoder architecture for alignment [42, 57, 59], single-encoder architecture for fusion [43, 73], or their combinatorial practices forming an encoder-decoder architecture [48, 80]. Most recent literature aims to expand as many supported transferable downstream tasks as possible via pre-training large-scale foundation models [6, 80]. Out of previous works, CLIP [59] blazes the trail to learn transferable visual features from text supervision, and demonstrates surprising zero-shot recognition results. Follow-up study includes leveraging dense [79], augmentation [47], and uni-modal self-supervised [50, 55, 74] signals for improved performance. In this work, we base our study on the dual-encoder architecture, and pre-train models either with the contrastive objective as CLIP [59], or the non-contrastive objective. Further, we seek their complementarity and study whether synergy between two objectives exists. + +Contrastive learning. The idea of contrastive learning is popularized under visual self-supervised learning (SSL) which aims at learning transferable visual representation from unlabelled images. In common practices [18, 33, 75], each image is considered a single class, attracted to its augmented views while pulled away from views of other images. The contrastive objective is proved robust and effective in a wide range of realms beyond visual understanding, including representation learning for natural language [29], audio [54], structured input [44], robotics [66], and multimodality [21, 59]. However, a caveat of these approaches is the requirement for large batch size [18] or memory bank [33, 75], which lies intrinsic in the formulation of InfoNCE estimator [38] and plagues the pre-training of largesize models due to hardware limitations. + +Non-contrastive learning. Beyond the contrastive paradigm, recent state of the arts from visual SSL manage to relieve the dependency on negative samples. With different subtleties to avert collapsing solutions, these works can optimize the affinity of augmented representations alone + +and are categorized as non-contrastive framework. To avoid model collapsing, common practices include asymmetrical architecture [19, 31], dimension de-correlation [11, 27, 81], and clustering [7, 8, 13–15, 71]. Our approach takes inspiration from the last category since it intrinsically frees the absoluteness of assignment of positive samples. Moreover, rich semantics induced by explicit clustering can automatically group visual concepts from noisy image-text pairs, facilitating models' representation learning. + +# 3. Approach + +# 3.1. Framework + +Suppose $\pmb{g}$ and $\pmb{h} \in \mathbb{R}^{D \times 1}$ are projected backbone features of two modalities. $D$ is the feature dimension. We start by formulating the objective of dominant contrastive framework CLIP, followed by the introduced non-contrastive framework nCLIP. The final framework xCLIP takes the multi-tasking of two objectives, the complete computational pipeline of which is shown in Algorithm 1. + +Contrastive pre-training [59]. Let $\pmb{u} = \pmb{g} / ||\pmb{g}||$ and $\pmb{v} = \pmb{h} / ||\pmb{h}||$ . $\pmb{U} = [\pmb{u}_1, \pmb{u}_2, \dots, \pmb{u}_B] \in \mathbb{R}^{D \times B}$ , $\pmb{V} = [\pmb{v}_1, \pmb{v}_2, \dots, \pmb{v}_B] \in \mathbb{R}^{D \times B}$ are concatenated embeddings over the batch. $B$ is the batch size. The contrastive objective is formulated as + +$$ +\begin{array}{l} \mathcal {L} _ {\mathrm {C L I P}} = \operatorname {I n f o N C E} (\boldsymbol {U} ^ {\mathrm {T}} \boldsymbol {V} / \sigma) \\ = - \frac {1}{B} \sum_ {i \in B} \log \frac {\exp \left(\boldsymbol {u} _ {i} ^ {\mathrm {T}} \boldsymbol {v} _ {i} / \sigma\right)}{\sum_ {j \in B} \exp \left(\boldsymbol {u} _ {i} ^ {\mathrm {T}} \boldsymbol {v} _ {j} / \sigma\right)}, \tag {1} \\ \end{array} +$$ + +where $\sigma$ is a trainable parameter controlling the temperature. $\mathcal{L}_{\mathrm{CLIP}}$ is symmetrized by setting $g$ and $h$ as projected features of the images and texts by turns and taking the average of two terms. Vanilla InfoNCE can be decoupled into two terms accounting for affinity and variability separately [81]. While both variability terms in the contrastive and non-contrastive objectives require for batch statistics (e.g., $\mathcal{L}_{\mathrm{EH}}$ in Eq. (4)), CLIP formulation explicitly maximizes the distance between negative pairs of batch samples. Hence, it degrades the learning performance and data efficiency when models are pre-trained with noisy data where sensible matches occur within negative pairs. + +Non-contrastive pre-training. Let $\pmb{p} = \mathrm{softmax}(\pmb{g})$ and $\pmb{q} = \mathrm{softmax}(\pmb{h})$ . We transform the projected feature into probability distributions, which can be seen as an assignment over semantic clusters (i.e., projection weights). We take $\pmb{p}$ as the target distribution and learn to estimate the predicted distribution $\pmb{q}$ by minimizing their cross-entropy: + +$$ +\mathcal {L} _ {\mathrm {C E}} = - \boldsymbol {p} ^ {\mathrm {T}} \log (\boldsymbol {q}). \tag {2} +$$ + +Note the target and prediction branch are backpropagated at the same time and $\mathcal{L}_{\mathrm{CE}}$ is symmetrized as $\mathcal{L}_{\mathrm{CLIP}}$ . To avert collapsing solutions, we incorporate entropy regularizers, i.e., entropy minimization [71] and mean entropy maximization [8,9], via minimizing: + +$$ +\mathcal {L} _ {\mathrm {E H}} = - \boldsymbol {p} ^ {\mathrm {T}} \log (\boldsymbol {p}), - \mathcal {L} _ {\mathrm {H E}} = \bar {\boldsymbol {p}} ^ {\mathrm {T}} \log (\bar {\boldsymbol {p}}), \tag {3} +$$ + +where $\overline{\boldsymbol{p}} = \mathbb{E}(p) \approx \frac{1}{B} \sum_{i=1}^{B} \boldsymbol{p}_i$ is the average distribution over batch. $\mathcal{L}_{\mathrm{EH}}$ and $\mathcal{L}_{\mathrm{HE}}$ are symmetrized by taking the average with another term where $\boldsymbol{p}$ is replaced by $\boldsymbol{q}$ . $\mathcal{L}_{\mathrm{EH}}$ encourages the model to make deterministic predictions and ensures assignment's sharpness. $\mathcal{L}_{\mathrm{HE}}$ encourages the model to utilize a full set of projection weights and ensures the assignment's smoothness. We illustrate in Sec. 4.4 and Appendix A that $\mathcal{L}_{\mathrm{EH}}$ and $\mathcal{L}_{\mathrm{HE}}$ are minimally sufficient to yield non-collapsing solutions. We empirically show in Tab. A1 that the solutions is non-trivial as long as both sharpness and smoothness are guaranteed, e.g., via Sinkhorn algorithm [14] instead of loss regularizers. We opt for Eq. (3) for its simplicity and consistency between pretraining and evaluation. The overall non-contrastive objective is formulated as + +$$ +\mathcal {L} _ {\mathrm {n C L I P}} = \mathcal {L} _ {\mathrm {C E}} + \lambda_ {1} \cdot \mathcal {L} _ {\mathrm {E H}} - \lambda_ {2} \cdot \mathcal {L} _ {\mathrm {H E}}, \tag {4} +$$ + +where $\lambda_{1}$ and $\lambda_{2}$ controls the weight of the regularization. We empirically find that setting $\lambda_{1} = \lambda_{2} - 1 = 0.5$ yields stable training and favorable transfer performances. We detail our findings in Sec. 4.4 and Appendix B.1. During evaluation, we utilize negative cross-entropy $-\mathcal{L}_{\mathrm{CE}}$ as the similarity metric for zero-shot transfer experiments. As schematically showcased in Fig. 1, nCLIP does not reply on negative examples but requires a relatively larger projection head and greater output dimension. Comparing nCLIP with CLIP, we empirically find that nCLIP produces more coarse-grained (e.g., zero-shot retrieval in Sec. 4.1) but semantic-rich (e.g., linear probing in Sec. 4.2) projections. Empirical evidence suggests that the model will fail to deliver reasonable zero-shot and retrieval results if pre-trained without negative pairs (e.g., nCLIP's 25.2 vs. CLIP's 73.8 points of R@1 under Flickr30K I→T retrieval as shown in Tab. 4). + +Contrastive meets non-contrastive. Given that two objectives each have their own limitations, we further seek the complementarity between $\mathcal{L}_{\mathrm{CLIP}}$ and $\mathcal{L}_{\mathrm{nCLIP}}$ , and pre-train the models with both objectives, written as + +$$ +\mathcal {L} _ {\mathrm {x C L I P}} = \lambda_ {\mathrm {C L I P}} \cdot \mathcal {L} _ {\mathrm {C L I P}} + \lambda_ {\mathrm {n C L I P}} \cdot \mathcal {L} _ {\mathrm {n C L I P}}, \tag {5} +$$ + +where $\lambda_{\mathrm{CLIP}}$ and $\lambda_{\mathrm{nCLIP}}$ control the weight of the objectives. During evaluation, we find using the negative cosine metric with CLIP's projection head for zero-shot transfer + +Algorithm 1: PyTorch-like pseudo-code of xCLIP. +```typescript +// enc, proj: encoder & projector +``` + +```latex +$\pmb{f}_I,\pmb {f}_T = \mathrm{enc}(\mathrm{img}),\mathrm{enc}(\mathrm{text})$ +``` + +```latex +$\pmb{g}_I, \pmb{g}_T = \mathrm{proj}_{\mathrm{CLIP}}(\pmb{f}_I), \mathrm{proj}_{\mathrm{CLIP}}(\pmb{f}_T)$ +``` + +```latex +$h_{I},h_{T} = \mathrm{proj}_{\mathrm{nCLIP}}(\pmb {f}_{I}),\mathrm{proj}_{\mathrm{nCLIP}}(\pmb{f}_{T})$ +``` + +```latex +$\mathcal{L}_{\mathrm{xCLIP}} = \lambda_{\mathrm{CLIP}}\cdot \mathcal{L}_{\mathrm{CLIP}}(\pmb {g}_I,\pmb {g}_T,\sigma)$ +``` + +```csv ++λnCLIP·LnCLIP(hH,T,λ1,λ2) +``` + +```txt +return $\mathcal{L}_{\mathrm{xCLIP}}$ +``` + +```txt +def $\mathcal{L}_{\mathrm{CLIP}}(\pmb {f}_I,\pmb {f}_T,\sigma)$ .. +``` + +```txt +// INCE: InfoNCE with Eq. (1) +``` + +```latex +$\pmb{u}_{I}$ = normalize $(f_{I}, p = 2, \dim = 1)$ +``` + +```txt +$\pmb{u}_{T} = \mathrm{normalize}(\pmb{f}_{T}, \mathrm{p} = 2, \mathrm{dim} = 1)$ +``` + +```latex +$\mathcal{L} = \mathrm{INCE}(\pmb{U}_I^{\mathrm{T}}\pmb{U}_T / \sigma) + \mathrm{INCE}(\pmb{U}_T^{\mathrm{T}}\pmb{U}_I / \sigma)$ +``` + +```txt +return $\mathcal{L} / 2$ +``` + +```latex +def $\mathcal{L}_{\mathrm{nCLIP}}(f_I, f_T, \lambda_1, \lambda_2)$ : +``` + +```latex +$\pmb{p}_I = \mathrm{softmax}(\pmb {f}_I,\mathrm{dim} = 1)$ +``` + +```latex +$\pmb{p}_T = \mathrm{softmax}(\pmb {f}_T,\mathrm{dim} = 1)$ +``` + +```latex +$\mathcal{L}_{\mathrm{CE}} = -\left(\pmb {p}_I\cdot \log (\pmb {p}_T) + \pmb {p}_T\cdot \log (\pmb {p}_I)\right).$ sum $(\mathrm{dim} = 1)$ +``` + +```txt +.mean(dim=0) +``` + +```latex +$\mathcal{L}_{\mathrm{EH}} = -(\pmb {p}_I\cdot \log (\pmb {p}_I) + \pmb {p}_T\cdot \log (\pmb {p}_T)).$ sum $(\dim = 1)$ +``` + +```txt +.mean(dim=0) +``` + +```latex +$\overline{\pmb{p}}_I,\overline{\pmb{p}}_T = \pmb {p}_I.\mathrm{mean}(\mathrm{dim} = 0),\pmb {p}_T.\mathrm{mean}(\mathrm{dim} = 0)$ +``` + +```latex +$\mathcal{L}_{\mathrm{HE}} = -\left(\overline{\boldsymbol{p}}_I\cdot \log (\overline{\boldsymbol{p}}_I) + \overline{\boldsymbol{p}}_T\cdot \log (\overline{\boldsymbol{p}}_T)\right).$ sum(dim=1) +``` + +```latex +$\mathcal{L} = \mathcal{L}_{\mathrm{CE}} + \lambda_1\cdot \mathcal{L}_{\mathrm{EH}} - \lambda_2\cdot \mathcal{L}_{\mathrm{HE}}$ +``` + +```txt +return $\mathcal{L} / 2$ +``` + +experiments generally leads to better results. We note that it is not immediately clear that the two objectives certainly induce stronger models, since they build qualitatively distinctive latent spaces. Qualitatively, we show in Sec. 4 that nCLIP helps CLIP to encode semantics which intrinsically lacks in CLIP, while CLIP helps nCLIP to be transferable for zero-shot recognition. The synergy between the two objectives prompts consistent performance gains across a wide range of tasks by xCLIP over CLIP. + +# 3.2. Implementation + +Architecture. We train the base-size model with ViTB/16 [26] as the visual encoder. The model configuration of the base-size text encoder follows CLIP [59] with byte-pair encoding (BPE) and a maximum context length of 77. The projection head generating output for the noncontrastive objective is a 2-layer MLP with 4096-dim hidden layers, GELU [36], and BatchNorm [41]. The last layer is of 32768 output dimension and followed by a BatchNorm without affine transformation. The projection head for the contrastive objective is a single linear layer with no bias and a dimension of 512. + +Pre-training data. We train our method with publicly available datasets: COCO [51], Visual Genome [45], SBU + +Captions [56], Conceptual Caption 3M [65], Conceptual Caption 12M [17], and filtered 14M-size subset of Yahoo Flickr Creative Commons 100M dataset, consisting a total of 35M Image-Text pairs (IT35M). We are also intrigued about the behaviors of models pre-trained on ImagNet-21K [24] (IN21K) dataset taking label names as annotation texts by concatenating them with a sampled prompt, similar to [57, 80]. To notify, ImageNet-21K is a subset of ImageNet-22K with classes of ImageNet-1K excluded for fair downstream evaluation. We detail the models' data scaling behavior in Tab. 8. + +Optimization. We use a batch size of 4096, with all data distributed across 32 V100 GPUs. Models are trained by default with AdamW [53] optimizer, a peak learning rate of $1e^{-3}$ , cosine scheduler, a weight decay of 0.2, a $\beta_{2}$ for AdamW of 0.98, a $\epsilon$ for AdamW of $1e^{-6}$ , and automatic mixed-precision for 3 warm-up epochs and a total of 32 epochs. $\lambda_{1}$ and $\lambda_{2}$ are set as 0.5 and 1.5, respectively. The two objectives are multi-tasked with $\lambda_{\mathrm{nCLIP}}$ being 1 and $\lambda_{\mathrm{CLIP}}$ being 0.2. Augmentations and pre-processing of pre-training data follow CLIP with image size randomly cropped and resized within the scale of (0.5, 1.0). Detailed pre-training hyper-parameters are detailed in Appendix F. + +# 4. Experiments + +# 4.1. Zero-Shot Transfer + +Classification. We evaluate 27 classification benchmarks with zero-shot classification protocols following [59]. As shown in Tab. 1, nCLIP achieves an average top-1 accuracy of $32.7\%$ . Though performing $7.6\%$ worse than CLIP, nCLIP demonstrates descent zero-shot recognition capability without explicitly training with negative examples. xCLIP achieves consistent performance gain across a wide range of datasets, with a $43.6\%$ average top-1 accuracy, which is $3.3\%$ higher than CLIP. It indicates that the synergy between contrastive and non-contrastive objectives improves the model's zero-shot classification ability. + +Out-of-domain classification. For out-of-distribution classification, we use 5 datasets following [59]: ImageNet Adversarial [37], ImageNet Rendition [35], ImageNetV2 [62], ImageNet Sketch [72], and ObjectNet [10] with 2 additional datasets: ImageNet-C [76] and Stylized ImageNet [30]. We investigate if the non-contrastive objective inherits the robustness of the contrastive objective to natural distribution shift. As shown in Tab. 2, nCLIP achieves an $18.0\%$ with CLIP an $23.1\%$ average top-1 accuracy, revealing a similar trend to in-domain datasets. xCLIP further improves CLIP's accuracy by $3.6\%$ and achieves an accuracy of $26.7\%$ across 7 datasets, indicating its effectiveness among out-of-domain datasets. + +
ModelFood101CIFAR10CIFAR100BirdmapSUN397CarsAircraftVOC2007DTDPetsCaltech101FlowersMNISTFER2013STL10EuroSATRESISC45GTSRBKITTICountry211PCAMUCF101Kinetics700CLEVRHapticMemesSSTImageNetAverage
CLIP61.279.750.623.756.515.95.846.427.654.771.348.910.537.391.324.539.713.131.69.150.045.032.412.853.049.145.740.3
nCLIP28.479.549.111.357.05.94.551.422.914.665.023.19.913.594.815.121.22.735.45.851.242.028.412.452.750.037.032.7
xCLIP65.883.454.525.159.918.05.852.233.257.173.950.012.339.092.840.043.616.339.89.351.149.835.418.452.550.248.843.6
+ +Table 1. Zero-shot classification. We report on a variety of classification benchmarks with ViT-B/16 pre-trained on IT35M. Detailed protocols for each dataset strictly follow CLIP [59]. xCLIP achieves a consistent performance gain compared to CLIP in a wide range of classification datasets. Best results of each column are bolded. + +
ModelIN-AIN-RIN-v2IN-SkeIN-StyIN-CObjNetAverage
CLIP21.343.737.926.54.911.915.523.1
nCLIP21.529.030.817.73.611.512.018.0
xCLIP27.749.140.630.66.114.717.826.7
+ +Table 2. Zero-shot out-of-distribution classification. We report on a variety of out-of-distribution classification benchmarks. xCLIP demonstrates stronger robustness on various out-of-domain classification datasets. + +
ModelNUS-WIDEOpenImages
mAPF1@3F1@5mAPF1@10F1@20
CLIP15.133.316.081.113.47.2
nCLIP16.135.616.681.411.66.4
xCLIP15.335.216.881.213.87.4
+ +Table 3. Zero-shot multi-label classification. We report mAP and F1 scores with ViT-B/16 pre-trained on IT35M. nCLIP showcases the best multi-label classification capability. + +Multi-label classification. We evaluate zero-shot multi-label classification on NUS-WIDE [23] and OpenImages [46] following standard protocol [39]. Specifically, we use 81 unseen labels for NUS-WIDE and the most frequent 400 unseen test labels for OpenImages during evaluation, following [40]. As shown in Tab. 3, we observe that nCLIP shows the best mAP compared to CLIP. nCLIP is pre-trained without negative examples, resulting in the model's intrinsic strengths on recall over overlapped concepts, with 16.1 and 81.4 points of mAP on NUS-WIDE and OpenImages, respectively. Comparatively, xCLIP slightly improves CLIP in this respect while lagging behind nCLIP. The experiments demonstrate that the non-contrastive objective fits well for those downstream tasks in demand of high recall. + +Retrieval. We evaluate on 2 retrieval benchmarks: Flickr30K [58] and MSCOCO [51] under zero-shot protocol. We do not use prompt engineering and use the original caption. Compared to zero-shot classification, image-text retrieval requires the model's recognition capability at a fine-grained level. We empirically observe in Tab. 4 that + +nCLIP performs drastically worse than CLIP, which is because text captions in retrieval are not mutually exclusive, which is different from label names in classification. Therefore, explicit negative examples during pre-training play an imperative role in zero-shot retrieval. See Appendix E.1 for further discussions. Beyond that, xCLIP achieves a noticeable performance improvement over CLIP, with a gain of $3.7\%$ R@1 on Flickr30K and $3.9\%$ R@1 on MSCOCO, illustrating that the additional non-contrastive objective renders extra semantic signals that can be transferred well to fine-grained recognition. + +# 4.2. Representation Learning + +We conduct several evaluation protocols to benchmark representation quality under different objectives for both visual (Sec. 4.2.1) and textual (Sec. 4.2.2) modality. + +# 4.2.1 Visual Representation Learning + +Linear probing. We evaluate the quality of visual representation via linear probing protocol, where a linear head is fine-tuned on top of the frozen backbones. We follow the same setup as [20] on the same 27 datasets as standard zero-shot classification. Specifically, we use SGD without momentum as the optimizer, no weight decay, and a total epoch of 100 for all evaluation datasets. We use [CLS] token for classification. Following [82], we sweep a set of different learning rates by adding multiple classification heads over a shared frozen backbone, each with its own optimizer and scheduler. We report the best result across different heads. As shown in Tab. 5, we observe that nCLIP generally leads to comparable linear probing accuracy compared with CLIP, suggesting that non-contrastive objectives are able to derive semantically meaningful embedding spaces. nCLIP achieves $2.5\%$ higher performance compared to CLIP on ImageNet but relatively lower on average. Beyond that, when combining two objectives together, xCLIP achieves $2.7\%$ higher performance on ImageNet and $1.5\%$ higher performance on average across 27 different classification tasks compared to CLIP. + +Fine-tuning & semi-supervised learning. We fine-tune the entire network under the full-data regime (100%) and + +
ModelFlickr30KMSCOCO
I→TT→II→TT→I
R@1R@5R@10R@1R@5R@10R@1R@5R@10R@1R@5R@10
CLIP73.893.397.152.979.387.759.485.291.934.362.774.0
nCLIP26.253.367.920.148.561.121.447.961.613.437.150.3
xCLIP77.595.697.757.382.989.063.387.594.138.466.076.7
+ +Table 4. Zero-shot image-to-text retrieval. We report R@1, R@5, and R@10 in both image-to-text (I→T) and text-to-image (T→I) settings with ViT-B/16 pre-trained on IT35M. xCLIP consistently improves CLIP in retrieval. + +
ModelFood101CIFAR10CIFAR100BirdsnapSUN397CarsAircraftVOC2007DTDPetsCaltech101FlowersMNISTFER2013STL10EuroSATRESISC45GTSRBKITTICountry211PCAMUCF101Kinetics,700CLEVRHaarSoftMnesSSTImageNetAverage
+ +CLIP 83.3 92.0 75.6 54.4 78.8 60.5 33.8 91.1 72.4 79.0 90.0 96.1 96.2 56.5 97.1 94.8 90.3 77.3 69.7 23.2 83.4 78.9 53.9 50.3 55.9 55.8 71.4 |72.7 nCLIP 83.8 93.2 76.3 66.8 79.3 49.3 28.0 90.1 70.6 70.2 90.5 95.0 95.4 54.7 97.4 91.9 88.2 75.7 64.0 22.7 84.2 79.9 51.0 44.6 55.6 55.2 73.9 |71.4 xCLIP 85.3 93.4 77.8 58.4 80.7 62.3 36.7 92.3 74.0 81.5 91.6 97.0 96.8 58.5 98.1 95.3 91.2 80.4 69.3 26.3 83.5 81.1 56.0 49.8 58.5 54.8 74.1 |74.2 + +Table 5. Linear probing. We report on a variety of classification benchmarks with ViT-B/16 pre-trained on IT35M. CLIP achieves consistent performance gains compared to CLIP in a wide range of classification datasets. + +
Modelsm. sup.ft.
1%10%100%
CLIP57.872.482.4
nCLIP55.072.282.4
xCLIP59.573.482.7
+ +Table 6. Fine-tuning (ft.) and semi-supervised learning (sm. sup.). Percentage shows sampling ratio for end-to-end fine-tuning. We report top-1 accuracy on the ImageNet-1K validation set. nCLIP performs comparatively with CLIP. xCLIP induces consistent gains in all sampling ratios, especially for semi-supervised learning. + +
ModelSTS12STS13STS14STS15STS16STS-BSICK-RAverage
CLIP52.354.053.969.958.667.069.360.7
nCLIP51.362.756.971.965.866.964.162.8
xCLIP48.755.655.669.759.566.370.260.8
+ +Table 7. Sentence embedding performance on semantic textual similarity. We report on a variety of NLP understanding benchmarks. xCLIP demonstrates the best capability for various STS tasks. xCLIP has comparable performance with CLIP. + +the semi-supervised learning protocol (1% and 10%) on ImageNet-1K [24]. Results are shown in Tab. 6. For CLIP and xCLIP, we find that fine-tuning with the projection head trained under the contrastive objective yields consistent performance gain, especially under a low-data regime. Specifically, we take the text features of the label names as the classifier's initialized parameters. We opt for this setup by default. We highlight that while nCLIP performs worse than CLIP in terms of zero-shot transfer, the non-contrastive objective also serves as a strong baseline to learn represent + +tation considering the close gap between CLIP and nCLIP in terms of fine-tuning accuracy. xCLIP achieves an $1.7\%$ , $1.0\%$ , and $0.3\%$ performance gain compared with CLIP under three data ratios of $1\%$ , $10\%$ , and $100\%$ , respectively. + +Mask probing. To evaluate how well the visual model is capable of deriving explicit scene layout and object boundaries, we conduct mask probing analysis following [15]. For each attention head from the last layer, we extract the attention map with [CLS] token as the query. We then compute the Jaccard similarity $\mathcal{J}$ of each head's attention mask to the ground truth and retain the attention mask with the highest similarity. We conduct experiments on Pascal VOC 2012 [28] dataset. With IT35M, nCLIP demonstrates better quality in the generated mask with 43.7 points of $\mathcal{J}$ , while CLIP reaches 41.2 and xCLIP reaches 41.9 points of $\mathcal{J}$ . Hence, nCLIP learns better representation for object boundaries compared to CLIP, while xCLIP strikes a balance between nCLIP and CLIP. Detailed results are delayed to Appendix D.1. As a related evaluation, we also study unsupervised segmentation with GroupViT [77] in Appendix D.2. + +# 4.2.2 Textual Representation Learning + +We follow the setup as [29] and evaluate on 7 STS tasks: STS 2012-2016 [1-5], STS Benchmark [16], and SICK-Relatedness [34]. We directly take [EOS] token without any projection as the input for evaluation, which consistently yields better performance compared to the projected feature for all models. We use cosine distance as the similarity metric. The main goal of sentence embeddings is to cluster semantically similar sentences, and hence, we take STS as one yardstick to benchmark textual representation + +
DataSizeEp.ZS / LN Accuracy
CLIPnCLIPxCLIP
①12M2536.8 / 68.537.5 / 71.042.4 / 72.2
④14M3226.5 / 75.131.4 / 77.027.1 / 77.2
①②22M3237.9 / 68.234.2 / 72.343.0 / 71.9
①②③35M3245.7 / 71.437.0 / 73.948.8 / 74.1
+ +Table 8. Pre-training data scaling & domain. Ep. denotes training epochs. Data abbreviations are as follows. ①: CC12M. ②: COCO, VG, SBU, and CC3M. ③: YFCC14M. ④: IN21K. Note ①②③ = IT35M. + +![](images/d8a2d85afcc2806e3dbf4fcdfa85ee47b8515188eb85100155e54c107714a0d5.jpg) +Figure 2. Batch size scaling. xCLIP performs better than CLIP with a small batch size (i.e., 1024). + +![](images/02bb5660dade19153329b4fcde68549a498f3f1a2e08a807bbe25348d5c3a5cc.jpg) +Figure 3. Training time. xCLIP adds $\sim 1.3x$ training time cost and performs better with equal seen images. + +learning. Results are shown in Tab. 7. nCLIP shows better textual semantics compared to CLIP, with $2.1\%$ higher in terms of the average score with IT35M. The conclusion can also be visually extrapolated from the t-SNE visualization of the word embedding, as detailed in Appendix G. We note that xCLIP performs comparatively with CLIP given that STS is evaluated on [EOS] without projection, thus the backbone representation can be similar if dominant by the contrastive objective. + +# 4.3. Properties + +We analyze several crucial properties to demonstrate the generalizability and effectiveness of our method in practical usage. Experiments are conducted with ViT-B/16 on CC12M for 25 epochs by default. + +Pre-training data scaling. We study the model's performance with data scaling in Tab. 8. xCLIP exhibits desirable scaling behaviors similar to CLIP. The performance improves as data increases. However, we do not observe such a tendency in nCLIP in terms of zero-shot accuracy. For example, nCLIP achieves $37.5\%$ with (1) but only $34.2\%$ with (2), the trend of which defies those of CLIP's $(36.8\%$ vs. $37.9\%)$ and xCLIP's $(42.4\%$ vs. $43.0\%)$ . On an important note, both nCLIP and xCLIP showcase desirable scaling behaviors under the linear probing protocol. We hypothesize that the performance of zero-shot classification depends on + +pre-training data quality with the non-contrastive objective, while strong representation can always be ensured before the projection. + +Pre-training with tagging data. We consider using tagging data, e.g., ImageNet-21K [24], as pre-training data to validate method's generalizability. To do that, we take label names as annotation texts by concatenating them with a sampled prompt, similar to [57, 80]. As shown in Tab. 8, nCLIP shows superiority to CLIP when pre-trained with IN21K by performance gain of $4.9\%$ and $1.9\%$ in terms of zero-shot and linear probing accuracy, respectively. Compared to nCLIP, we instead notice a performance drop of $4.3\%$ for zero-shot accuracy when two objectives multitasked, indicating that the contrastive objective stymies the quality of projected probability under tagging data. Note, that an $77.2\%$ linear probing accuracy achieved by xCLIP, $2.1\%$ higher than CLIP, is decently strong compared to $79.0\%$ achieved by the state of the art in SSL, iBOT [82], pre-trained using the same data with 80 epochs. + +Batch size scaling. We investigate the scaling behavior of CLIP and xCLIP with different pre-training batch sizes. As shown in Fig. 2, xCLIP performs stronger with smaller batch sizes than CLIP pre-trained even with larger batch sizes. For example, xCLIP pre-trained with a batch size of 1024 achieves a zero-shot accuracy of $38.5\%$ (vs. CLIP's $36.8\%$ with a batch size of 4096). However, we still observe a substantial performance drop as batch size decreases, the behavior of which is different from self-supervised models pre-trained with the non-contrastive objective [14, 15] that are insensitive to batch size. We hypothesize that a large batch size is necessary to approximate $\mathbb{E}(p)$ with $\frac{1}{B}\sum_{i=1}^{B} p_i$ in estimating $\mathcal{L}_{\mathrm{HE}}$ , especially when two probabilities are derived from drastically different modalities. + +Computation efficiency. To demonstrate methods' computation efficiency, we show the FLOPs and GPU memory consumption of xCLIP compared to CLIP. xCLIP enforces bearable additional computation cost on top of CLIP with only $1.4\%$ extra FLOPs and $27\%$ extra GPU memory. As shown in Fig. 3, xCLIP adds $\sim 1.3x$ extra time cost but with a $6.0\%$ performance gain in terms of zero-shot accuracy. + +# 4.4. Ablation Study + +We show in this section the crucial composing factors of nCLIP and xCLIP. The ablations in the first column (Tabs. 9a to 9c) are conducted with $\mathcal{L}_{\mathrm{nCLIP}}$ only. The ablations in the second column (Tabs. 9d to 9f) are conducted with full loss $\mathcal{L}_{\mathrm{xCLIP}}$ . Experiments are conducted with ViTB/16 on CC12M for 25 epochs. The default settings are highlighted in cyan. + +
λ1λ2ZSLN
01NaN
0.3126.869.5
0.5123.767.2
0.21.235.369.0
0.51.537.571.0
+ +(a) Coefficient of entropy regularizer. $\lambda_1 + 1 = \lambda_2$ generally leads to good performance. + +
#SL / #TLZSLN
0 / 142.472.2
0 / 241.371.3
1 / 242.471.9
2 / 342.572.3
+ +(d) Shared projection layers. #SL denotes shared hidden layers. #TL denotes total hidden layers. + +
dimZSLNMem
819231.070.61.07×
1638435.170.91.13×
3276837.571.01.27×
6553635.971.01.52×
+ +(b) Projection dimension. Mem (G) is compared to CLIP with a 512-dim embedding layer. + +
λCLIP: λnCLIPZSLN
1.040.266.5
0.541.770.3
0.242.472.2
0.141.872.6
+ +(e) Loss ratio between $\mathcal{L}_{\mathrm{CLIP}}$ and $\mathcal{L}_{\mathrm{nCLIP}}$ + +
archZSLN
vanilla37.571.0
w/ bottleneck layer31.269.0
w/ l2-norm layer35.470.7
w/o last BN layer34.369.9
+ +(c) Projection architecture of non-contrastive branch. + +
techniqueZSLN
vanilla42.472.2
shared space (Eq. (3))29.468.9
debiased sampling42.272.4
warm-up on LnCLIP41.771.9
+ +(f) Training objective and technique for multitasking. + +Table 9. Ablation study with ViT-B/16 on ImageNet-1K validation set. We report zero-shot (ZS) and linear probing (LN) accuracy (\%). + +Entropy regularizer. We study the effects of coefficient for entropy regularizers $\lambda_{1}$ and $\lambda_{2}$ in Tab. 9a. The pretraining is unstable without regularizers or with only maximization of the entropy of the mean $(\mathcal{L}_{\mathrm{HE}})$ . Specifically, the model will collapse to constant uniform distribution and outputs the same uniform probability distribution despite different inputs. The additional minimization of the mean of the entropy $(\mathcal{L}_{\mathrm{EH}})$ stabilizes training, but incurs dimensional collapse, eroding the performances (26.8% and 23.7% vs. 37.5% in terms of zero-shot accuracy). Simultaneously adjusting $\lambda_{2}$ with constraints $\lambda_{1} + 1 = \lambda_{2}$ and $\lambda_{1} = 0.5$ mitigates the problem and yields the optimal performance, with a 37.5% zero-shot and 71.0% linear-probing accuracy. Details are shown in Appendix B.1. + +Projection head. We study the design of two projection heads for $\mathcal{L}_{\mathrm{CLIP}}$ and $\mathcal{L}_{\mathrm{nCLIP}}$ . As shown in Tab. 9b, large projection dimension matters for good performance. We opt for a projection dimension of 32768 balancing the performance and speed. As shown in Tab. 9c, the last BN layer is crucial and leads to more stable optimization in our experiments. We draw on the idea of prototypes [9, 15] by introducing bottleneck and 12-norm layers, while they do not yield a performance gain. We further study whether two projections can share intermediate computation in Tab. 9d. While sharing middle hidden layers with slightly deeper projection heads performs better, we opt for 2-layer MLP without shared layers for its simplicity. + +Optimization. We study the effect of hyper-parameters in optimization. In Tab. 9e, we study the optimal loss ratio between $\mathcal{L}_{\mathrm{CLIP}}$ and $\mathcal{L}_{\mathrm{nCLIP}}$ . We further study whether it's the best route for two objectives to be optimized in separate latent spaces via multi-tasking. Specifically, we con + +sider a bespoke objective Eq. (3) taking account of probability estimation and negative samples simultaneously in a shared latent space. Detailed formulation are shown in Appendix B.2. Optimizing in one latent space yields suboptimal results with a nearly $13.0\%$ drop in zero-shot accuracy and $3.3\%$ in linear probing accuracy, suggesting two objectives are intrinsically contradictory. We consider debiased sampling, where each batch is sampled from a single data source, since their semantics may be closer and thus more suitable for optimization. This yields similar results. We also add $\mathcal{L}_{\mathrm{nCLIP}}$ after warm-up epochs with the linear scheduler to its base scale, which erodes the performance. + +# 5. Conclusion + +Aligning images and texts is of overriding significance for vision-language understanding. To conquer the systematic insufficiency of the contrastive objective for acquiring semantics and tackling loose correlations between noisy image-text pairs, we explore the non-contrastive objective for language-image pre-training and unravel its properties. Empirical evidence reveals that the non-contrastive objective induces models to perform favorably in representation learning yet poorly in zero-shot transfer. Observing the distinct mechanisms of the two objectives, we further seek synergy between the two, and introduce a simple multi-tasking framework, xCLIP, that enjoys the best of both worlds: nCLIP aids CLIP mining semantics while CLIP inherits intrinsic strengths for zero-shot recognition. The consistent performance gain of xCLIP over CLIP on a wide variety of downstream tasks consolidates our findings. As potential future work, we may continue to scale up the data size (e.g., LAION400M [64]) as well as the model size (e.g., ViT-L [26]) to verify whether the scaling law applies and the performance improvement endures. + +# References + +[1] Eneko Agirre, Carmen Banea, Claire Cardie, Daniel Cer, Mona Diab, Aitor Gonzalez-Agirre, Weiwei Guo, Inigo Lopez-Gazpio, Montse Maritxalar, Rada Mihalcea, et al. Semeval-2015 task 2: Semantic textual similarity, english, spanish and pilot on interpretability. In SemEval, 2015. 6 +[2] Eneko Agirre, Carmen Banea, Claire Cardie, Daniel M Cer, Mona T Diab, Aitor Gonzalez-Agirre, Weiwei Guo, Rada Mihalcea, German Rigau, and Janyce Wiebe. Semeval-2014 task 10: Multilingual semantic textual similarity. In SemEval, 2014. 6 +[3] Eneko Agirre, Carmen Banea, Daniel Cer, Mona Diab, Aitor Gonzalez Agirre, Rada Mihalcea, German Rigau Claramunt, and Janyce Wiebe. Semeval-2016 task 1: Semantic textual similarity, monolingual and cross-lingual evaluation. In SemEval, 2016. 6 +[4] Eneko Agirre, Daniel Cer, Mona Diab, and Aitor Gonzalez-Agirre. Semeval-2012 task 6: A pilot on semantic textual similarity. In SemEval, 2012. 6 +[5] Eneko Agirre, Daniel Cer, Mona Diab, Aitor Gonzalez-Agirre, and Weiwei Guo. *sem 2013 shared task: Semantic textual similarity. In SemEval, 2013. 6 +[6] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. arXiv preprint arXiv:2204.14198, 2022. 2 +[7] Elad Amrani and Alex Bronstein. Self-supervised classification network. In ECCV, 2022. 3 +[8] Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Florian Bordes, Pascal Vincent, Armand Joulin, Michael Rabbat, and Nicolas Ballas. Masked siamese networks for label-efficient learning. In ECCV, 2022. 3 +[9] Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Armand Joulin, Nicolas Ballas, and Michael Rabbat. Semi-supervised learning of visual features by nonparametrically predicting view assignments with support samples. In ICCV, 2021. 3, 8 +[10] Andrei Barbu, David Mayo, Julian Alverio, William Luo, Christopher Wang, Dan Gutfreund, Josh Tenenbaum, and Boris Katz. Objectnet: A large-scale bias-controlled dataset for pushing the limits of object recognition models. In NeurIPS, 2019. 4 +[11] Adrien Bardes, Jean Ponce, and Yann LeCun. Vicreg: Variance-invariance-covariance regularization for self-supervised learning. In ICLR, 2022. 3 +[12] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. In NeurIPS, 2020. 1 +[13] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features. In ECCV, 2018. 2, 3 +[14] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. In NeurIPS, 2020. 2, 3, 7 + +[15] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 2, 3, 6, 7, 8 +[16] Daniel Cer, Mona Diab, Eneko Agirre, Inigo Lopez-Gazpio, and Lucia Specia. Semeval-2017 task 1: Semantic textual similarity-multilingual and cross-lingual focused evaluation. In SemEval, 2017. 6 +[17] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing web-scale image-text pretraining to recognize long-tail visual concepts. In CVPR, 2021. 4 +[18] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 2 +[19] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In CVPR, 2021. 3 +[20] Xinlei Chen*, Saining Xie*, and Kaiming He. An empirical study of training self-supervised vision transformers. In ICCV, 2021. 5 +[21] Yanbei Chen, Yongqin Xian, Sophia Koepke, Ying Shan, and Zeynep Akata. Distilling audio-visual knowledge by compositional contrastive learning. In CVPR, 2021. 2 +[22] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Universal image-text representation learning. In ECCV, 2020. 1, 2 +[23] Tat-Seng Chua, Jinhui Tang, Richang Hong, Haojie Li, Zhiping Luo, and Yan-Tao Zheng. Nus-wide: A real-world web image database from national university of Singapore. In CIVR, 2009. 5 +[24] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. In CVPR, 2009. 2, 4, 6, 7 +[25] Karan Desai and Justin Johnson. Virtex: Learning visual representations from textual annotations. In CVPR, 2021. 1 +[26] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 4, 8 +[27] Aleksandr Ermolov, Aliaksandr Siarohin, Enver Sangineto, and Nicu Sebe. Whitening for self-supervised representation learning. In ICML, 2021. 3 +[28] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. IJCV, 2010. 6 +[29] Tianyu Gao, Xingcheng Yao, and Danqi Chen. SimCSE: Simple contrastive learning of sentence embeddings. In EMNLP, 2021. 2, 6 +[30] Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A Wichmann, and Wieland Brendel. Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. In ICLR, 2019. 4 + +[31] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. In NeurIPS, 2020. 3 +[32] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. In CVPR, 2022. 1 +[33] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, 2020. 2 +[34] Matthew Henderson, Rami Al-Rfou, Brian Strope, Yun-Hsuan Sung, László Lukács, Ruiqi Guo, Sanjiv Kumar, Balint Miklos, and Ray Kurzweil. Efficient natural language response suggestion for smart reply. arXiv preprint arXiv:1705.00652, 2017. 6 +[35] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robustness: A critical analysis of out-of-distribution generalization. In ICCV, 2021. 4 +[36] Dan Hendrycks and Kevin Gimpel. Bridging nonlinearities and stochastic regularizers with gaussian error linear units. In ICLR, 2017. 4 +[37] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. In CVPR, 2021. 4 +[38] R Devon Hjelm, Alex Fedorov, Samuel Lavoie-Marchildon, Karan Grewal, Phil Bachman, Adam Trischler, and Yoshua Bengio. Learning deep representations by mutual information estimation and maximization. In ICLR, 2019. 2 +[39] Dat Huynh and Ehsan Elhamifar. A shared multi-attention framework for multi-label zero-shot learning. In CVPR, 2020. 5 +[40] D. Huynh and E. Elhamifar. A shared multi-attention framework for multi-label zero-shot learning. In CVPR, 2020. 5 +[41] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In ICML, 2015. 4 +[42] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, 2021. 2 +[43] Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution or region supervision. In ICML, 2021. 2 +[44] Thomas Kipf, Elise van der Pol, and Max Welling. Contrastive learning of structured world models. arXiv preprint arXiv:1911.12247, 2019. 2 +[45] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. IJCV, 2017. 4 +[46] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4. IJCV, 2020. 5 + +[47] Janghyeon Lee, Jongsuk Kim, Hyounguk Shon, Bumsoo Kim, Seung Hwan Kim, Honglak Lee, and Junmo Kim. Uniclip: Unified framework for contrastive language-image pretraining. In NeurIPS, 2022. 2 +[48] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. In NeurIPS, 2021. 2 +[49] Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. Visualbert: A simple and performant baseline for vision and language. In ACL, 2020. 2 +[50] Yangguang Li, Feng Liang, Lichen Zhao, Yufeng Cui, Wanli Ouyang, Jing Shao, Fengwei Yu, and Junjie Yan. Supervision exists everywhere: A data efficient contrastive language-image pre-training paradigm. In ICLR, 2022. 1, 2 +[51] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 4, 5 +[52] Xingbin Liu, Jinghao Zhou, Tao Kong, Xianming Lin, and Rongrong Ji. Exploring target representations for masked autoencoders. arXiv preprint arXiv:2209.03917, 2022. 1 +[53] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In ICLR, 2019. 4 +[54] Pranay Manocha, Zeyu Jin, Richard Zhang, and Adam Finkelstein. Cdpam: Contrastive learning for perceptual audio similarity. In ICASSP, 2021. 2 +[55] Norman Mu, Alexander Kirillov, David Wagner, and Saining Xie. Slip: Self-supervision meets language-image pretraining. In ECCV, 2022. 2 +[56] Vicente Ordonez, Girish Kulkarni, and Tamara Berg. Im2text: Describing images using 1 million captioned photographs. In NeuIPS, 2011. 4 +[57] Hieu Pham, Zihang Dai, Golnaz Ghiasi, Hanxiao Liu, Adams Wei Yu, Minh-Thang Luong, Mingxing Tan, and Quoc V Le. Combined scaling for zero-shot transfer learning. arXiv preprint arXiv:2111.10050, 2021. 1, 2, 4, 7 +[58] Bryan A Plummer, Liwei Wang, Chris M Cervantes, Juan C Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In ICCV, 2015. 2, 5 +[59] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 2, 3, 4, 5 +[60] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 1 +[61] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, 2021. 1 +[62] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. Do imagenet classifiers generalize to imageNet? In ICML, 2019. 4 + +[63] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In NeurIPS, 2015. 2 +[64] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. In NeurIPS Workshops, 2021. 8 +[65] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In ACL, 2018. 1, 4 +[66] Aravind Srinivas, Michael Laskin, and Pieter Abbeel. Curl: Contrastive unsupervised representations for reinforcement learning. In ICML, 2020. 2 +[67] Weijie Su, Xizhou Zhu, Yue Cao, Bin Li, Lewei Lu, Furu Wei, and Jifeng Dai. Vl-bert: Pre-training of generic visual-linguistic representations. In ICLR, 2020. 2 +[68] Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. In EMNLP, 2019. 2 +[69] Bart Thomee, David A Shamma, Gerald Friedland, Benjamin Elizalde, Karl Ni, Douglas Poland, Damian Borth, and Li-Jia Li. Yfcc100m: The new data in multimedia research. In CACM, 2016. 1 +[70] Yael Vinker, Ehsan Pajouheshgar, Jessica Y Bo, Roman Christian Bachmann, Amit Haim Bermano, Daniel Cohen-Or, Amir Zamir, and Ariel Shamir. Clipasso: Semantically-aware object sketching. In SIGGRAPH, 2022. 1 +[71] Feng Wang, Tao Kong, Rufeng Zhang, Huaping Liu, and Hang Li. Self-supervised learning by estimating twin class distributions. arXiv preprint arXiv:2110.07402, 2021. 3 +[72] Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. Learning robust global representations by penalizing local predictive power. In NeurIPS, 2019. 4 +[73] Zirui Wang, Jiahui Yu, Adams Wei Yu, Zihang Dai, Yulia Tsvetkov, and Yuan Cao. Simvlm: Simple visual language model pretraining with weak supervision. In ICLR, 2022. 2 +[74] Bichen Wu, Ruizhe Cheng, Peizhao Zhang, Tianren Gao, Joseph E. Gonzalez, and Peter Vajda. Data efficient language-supervised zero-shot recognition with optimal transport distillation. In ICLR, 2022. 1, 2 +[75] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In CVPR, 2018. 2 +[76] Qizhe Xie, Minh-Thang Luong, Eduard Hovy, and Quoc V Le. Self-training with noisy student improves imagenet classification. In CVPR, 2020. 4 +[77] Jiarui Xu, Shalini De Mello, Sifei Liu, Wonmin Byeon, Thomas Breuel, Jan Kautz, and Xiaolong Wang. Groupvit: Semantic segmentation emerges from text supervision. In CVPR, 2022. 6 +[78] Jianwei Yang, Chunyuan Li, Pengchuan Zhang, Bin Xiao, Ce Liu, Lu Yuan, and Jianfeng Gao. Unified contrastive learning in image-text-label space. In CVPR, 2022. 1 + +[79] Lewei Yao, Runhui Huang, Lu Hou, Guansong Lu, Minzhe Niu, Hang Xu, Xiaodan Liang, Zhenguo Li, Xin Jiang, and Chunjing Xu. Filip: Fine-grained interactive language-image pre-training. In ICLR, 2022. 2 +[80] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917, 2022. 1, 2, 4, 7 +[81] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow twins: Self-supervised learning via redundancy reduction. In ICML, 2021. 3 +[82] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. ibot: Image bert pre-training with online tokenizer. In ICLR, 2022. 5, 7 \ No newline at end of file diff --git a/2023/Non-Contrastive Learning Meets Language-Image Pre-Training/images.zip b/2023/Non-Contrastive Learning Meets Language-Image Pre-Training/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..134f9d443e5f28acdc3d90bda431e00d7b37b713 --- /dev/null +++ b/2023/Non-Contrastive Learning Meets Language-Image Pre-Training/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b2eb49cdf8c7fd22a3521534ff1df8999da38f40744e0de55043dcb1209333c3 +size 376380 diff --git a/2023/Non-Contrastive Learning Meets Language-Image Pre-Training/layout.json b/2023/Non-Contrastive Learning Meets Language-Image Pre-Training/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..de277fd24864986e7265c55af51b753b6dc259dd --- /dev/null +++ b/2023/Non-Contrastive Learning Meets Language-Image Pre-Training/layout.json @@ -0,0 +1,10757 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 99, + 103, + 495, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 103, + 495, + 121 + ], + "spans": [ + { + "bbox": [ + 99, + 103, + 495, + 121 + ], + "type": "text", + "content": "Non-Contrastive Learning Meets Language-Image Pre-Training" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 143, + 455, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 143, + 455, + 171 + ], + "spans": [ + { + "bbox": [ + 138, + 143, + 455, + 171 + ], + "type": "text", + "content": "Jinghao Zhou Li Dong Zhe Gan Lijuan Wang Furu Wei Microsoft" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 199, + 192, + 211 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 199, + 192, + 211 + ], + "spans": [ + { + "bbox": [ + 143, + 199, + 192, + 211 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 224, + 290, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 224, + 290, + 514 + ], + "spans": [ + { + "bbox": [ + 46, + 224, + 290, + 514 + ], + "type": "text", + "content": "Contrastive language-image pre-training (CLIP) serves as a de-facto standard to align images and texts. Nonetheless, the loose correlation between images and texts of web-crawled data renders the contrastive objective data inefficient and craving for a large training batch size. In this work, we explore the validity of non-contrastive language-image pre-training (nCLIP), and study whether nice properties exhibited in visual self-supervised models can emerge. We empirically observe that the non-contrastive objective benefits representation learning while sufficiently underperforming under zero-shot recognition. Based on the above study, we further introduce xCLIP, a multi-tasking framework combining CLIP and nCLIP, and show that nCLIP aids CLIP in enhancing feature semantics. The synergy between two objectives lets xCLIP enjoy the best of both worlds: superior performance in both zero-shot transfer and representation learning. Systematic evaluation is conducted spanning a wide variety of downstream tasks including zero-shot classification, out-of-domain classification, retrieval, visual representation learning, and textual representation learning, showcasing a consistent performance gain and validating the effectiveness of xCLIP. The code and pre-trained models will be publicly available at https://github.com/shallowtoil/xclip." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 536, + 128, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 536, + 128, + 550 + ], + "spans": [ + { + "bbox": [ + 47, + 536, + 128, + 550 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "content": "Language-image pre-training which simultaneously learns textual and visual representation from large-scale image-text pairs has revolutionized the field of representation learning [25, 59], vision-language understanding [22], and text-to-image generation [61]. Compared to traditional visual models, language-instilled ones intrinsically inherit the capability of zero-shot or few-shot learning prominently demonstrated by large language models such as GPT-3 [12]. The precursor system, Contrastive Language-Image Pre-Training [59] (CLIP) that explicitly aligns the projected features of two modalities, has demonstrated surprising capabilities of zero-shot, representation learning, and robustness, being applied to a wide range of fields [32, 52, 60, 70]." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 309, + 202, + 547, + 312 + ], + "blocks": [ + { + "bbox": [ + 309, + 202, + 547, + 312 + ], + "lines": [ + { + "bbox": [ + 309, + 202, + 547, + 312 + ], + "spans": [ + { + "bbox": [ + 309, + 202, + 547, + 312 + ], + "type": "image", + "image_path": "a0c87e3957ef3b12e91d888b2fdf06da99582a902e374957c0e45e5be91a978b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 316, + 547, + 415 + ], + "lines": [ + { + "bbox": [ + 305, + 316, + 547, + 415 + ], + "spans": [ + { + "bbox": [ + 305, + 316, + 547, + 415 + ], + "type": "text", + "content": "Figure 1. Architecture comparison between CLIP and nCLIP. We take base-size encoders for an instance. CLIP discriminates instances within a batch using 512-dim projected embeddings. nCLIP projects each modality into a 32768-dim probability distribution as the pseudo-label to supervise the prediction from the other modality. Darker blocks of nCLIP depict a higher response for cluster distribution, signifying the clusters to which text or image instances may belong. xCLIP is a multi-tasking framework with both CLIP and nCLIP." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 425, + 547, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 425, + 547, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 425, + 547, + 628 + ], + "type": "text", + "content": "To learn from noisy web-crawled image-text pairs, CLIP adopts a formulation of the contrastive objective, where the image and text within a pair are considered as unique instances and are encouraged to be discriminated from all the other negative instances. However, web-crawled image-text pairs [65,69] are usually loosely correlated in the sense that one caption (image) can match reasonably with multiple images (captions) besides the ground-truth one, as statistically shown in [74]. Hence, it is inaccurate and data-inefficient for representation learning to neglect other sensible matches and overlook the semantics hidden inside the textual description. This is also solidified by the undesirable discriminative ability or transferring performance of the visual encoder pre-trained under the contrastive objective, as suggested in [78]. Mining semantics from plentiful concepts appearing in captions, therefore, solicits further exploration beyond the vanilla contrastive objective." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "content": "Previously, some works resort to mining nearest neighbor positive samples via intra-modality similarity [50, 74] but require extra storage for auxiliary modules, i.e., the teacher network [74] or the memory bank [50]. Other works conduct multi-tasking with image-label supervision [57, 78, 80], transforming the captions into tag format or the image tag into the captioning format for unified contrastive" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "11028" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 180 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 180 + ], + "type": "text", + "content": "learning. Notwithstanding, such practice is still unable to account for the loose assignment within the image-text captioning pairs. Intuitively, a caption for an image usually identifies existing objects within the image, which can be captured by a probabilistic distribution estimating how the text is assigned to one or multiple object clusters [13, 14]. Such estimation depicts the semantic meanings for its contained visual contents, and thus can be leveraged as a pseudo-label to guide the learning of the visual encoder." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 183, + 288, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 183, + 288, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 183, + 288, + 495 + ], + "type": "text", + "content": "Inspired by the superiority of visual self-supervised learning (SSL) models pre-trained with the non-contrastive objective [14, 15], we explore whether the non-contrastive objective can be used across modalities for pre-training language-image models, and whether nice properties displayed on visual SSL models can be inherited. To this end, we follow the same setup as CLIP except for the objective, and study non-Contrastive Language-Image Pre-Training (nCLIP). For an image-text pair, we use the estimation of the textual (visual) distribution as the target to supervise the visual (textual) distribution, measured by cross-entropy. Additional regularizers are applied to avoid trivial solutions. Schematic comparison between CLIP and nCLIP is shown in Fig. 1. Theoretically, such formulation takes one modality and the cluster centroid that the other modality belongs to as the positive samples to contrast [14], as opposed to direct features of two modalities as in contrastive learning, such that a single image is tolerated to be aligned with multiple captions and concepts. Based on the above formulation, we conduct a systematic study in terms of zero-shot transfer and representation learning. We empirically observe that, while nCLIP demonstrates desirable representation learning capability for both modalities, it underperforms prominently in zero-shot classification and retrieval tasks where models pre-trained with negative samples naturally prevail." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 498, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 714 + ], + "type": "text", + "content": "Seeking for unique strengths of both worlds, we further perform multi-tasking of CLIP and nCLIP, short as xCLIP, and seek synergy between two distinct pre-training objectives. With the bearable add-on of computational resources (e.g., " + }, + { + "bbox": [ + 46, + 498, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\sim 27\\%" + }, + { + "bbox": [ + 46, + 498, + 287, + 714 + ], + "type": "text", + "content": " memory-wise and " + }, + { + "bbox": [ + 46, + 498, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\sim 30\\%" + }, + { + "bbox": [ + 46, + 498, + 287, + 714 + ], + "type": "text", + "content": " time-wise), xCLIP achieves consistent performance gain compared to CLIP on a wide range of tasks spanning from zero-shot classification, retrieval, linear probing, and fine-tuning, etc. An extensive study with different pre-training datasets, evaluation metrics, and optimization configurations is conducted to validate xCLIP's effectiveness in mining semantics and improving data efficiency. Particularly, the base-size model pre-trained using xCLIP under 35-million publicly available image-text pairs achieves a performance gain of " + }, + { + "bbox": [ + 46, + 498, + 287, + 714 + ], + "type": "inline_equation", + "content": "3.3\\%" + }, + { + "bbox": [ + 46, + 498, + 287, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 498, + 287, + 714 + ], + "type": "inline_equation", + "content": "1.5\\%" + }, + { + "bbox": [ + 46, + 498, + 287, + 714 + ], + "type": "text", + "content": " on an average of 27 classification tasks [59], in terms of zero-shot classification and linear probing accuracy, respectively. Performance gain is also valid in terms of zero-shot retrieval, semi-supervised learning, and fine" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 305, + 72, + 545, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 109 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 109 + ], + "type": "text", + "content": "tuning, with 3.7 points of R@1 on Flickr30K [58], " + }, + { + "bbox": [ + 305, + 72, + 545, + 109 + ], + "type": "inline_equation", + "content": "1.7\\%" + }, + { + "bbox": [ + 305, + 72, + 545, + 109 + ], + "type": "text", + "content": " accuracy on " + }, + { + "bbox": [ + 305, + 72, + 545, + 109 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 305, + 72, + 545, + 109 + ], + "type": "text", + "content": " subset of ImageNet [24], and " + }, + { + "bbox": [ + 305, + 72, + 545, + 109 + ], + "type": "inline_equation", + "content": "0.3\\%" + }, + { + "bbox": [ + 305, + 72, + 545, + 109 + ], + "type": "text", + "content": " accuracy on ImageNet, respectively." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 306, + 120, + 392, + 133 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 120, + 392, + 133 + ], + "spans": [ + { + "bbox": [ + 306, + 120, + 392, + 133 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 141, + 545, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 141, + 545, + 440 + ], + "spans": [ + { + "bbox": [ + 304, + 141, + 545, + 440 + ], + "type": "text", + "content": "Language-image pre-training. Language-image pretraining, aka vision-language pre-training (VLP), learns to jointly cope with vision and language input using multi-modality models. Early practices [22, 49, 67, 68] separately encode modalities, i.e., texts with linear embedding and images with convolutional neural networks and region proposals [63]. Recent works take direct images instead of pre-fetched region features as input, using e.g., dual-encoder architecture for alignment [42, 57, 59], single-encoder architecture for fusion [43, 73], or their combinatorial practices forming an encoder-decoder architecture [48, 80]. Most recent literature aims to expand as many supported transferable downstream tasks as possible via pre-training large-scale foundation models [6, 80]. Out of previous works, CLIP [59] blazes the trail to learn transferable visual features from text supervision, and demonstrates surprising zero-shot recognition results. Follow-up study includes leveraging dense [79], augmentation [47], and uni-modal self-supervised [50, 55, 74] signals for improved performance. In this work, we base our study on the dual-encoder architecture, and pre-train models either with the contrastive objective as CLIP [59], or the non-contrastive objective. Further, we seek their complementarity and study whether synergy between two objectives exists." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 456, + 545, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 456, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 304, + 456, + 545, + 635 + ], + "type": "text", + "content": "Contrastive learning. The idea of contrastive learning is popularized under visual self-supervised learning (SSL) which aims at learning transferable visual representation from unlabelled images. In common practices [18, 33, 75], each image is considered a single class, attracted to its augmented views while pulled away from views of other images. The contrastive objective is proved robust and effective in a wide range of realms beyond visual understanding, including representation learning for natural language [29], audio [54], structured input [44], robotics [66], and multimodality [21, 59]. However, a caveat of these approaches is the requirement for large batch size [18] or memory bank [33, 75], which lies intrinsic in the formulation of InfoNCE estimator [38] and plagues the pre-training of largesize models due to hardware limitations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "Non-contrastive learning. Beyond the contrastive paradigm, recent state of the arts from visual SSL manage to relieve the dependency on negative samples. With different subtleties to avert collapsing solutions, these works can optimize the affinity of augmented representations alone" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11029" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 180 + ], + "type": "text", + "content": "and are categorized as non-contrastive framework. To avoid model collapsing, common practices include asymmetrical architecture [19, 31], dimension de-correlation [11, 27, 81], and clustering [7, 8, 13–15, 71]. Our approach takes inspiration from the last category since it intrinsically frees the absoluteness of assignment of positive samples. Moreover, rich semantics induced by explicit clustering can automatically group visual concepts from noisy image-text pairs, facilitating models' representation learning." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 191, + 113, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 191, + 113, + 205 + ], + "spans": [ + { + "bbox": [ + 47, + 191, + 113, + 205 + ], + "type": "text", + "content": "3. Approach" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 211, + 126, + 223 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 211, + 126, + 223 + ], + "spans": [ + { + "bbox": [ + 47, + 211, + 126, + 223 + ], + "type": "text", + "content": "3.1. Framework" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 228, + 288, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 228, + 288, + 314 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 288, + 314 + ], + "type": "text", + "content": "Suppose " + }, + { + "bbox": [ + 46, + 228, + 288, + 314 + ], + "type": "inline_equation", + "content": "\\pmb{g}" + }, + { + "bbox": [ + 46, + 228, + 288, + 314 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 228, + 288, + 314 + ], + "type": "inline_equation", + "content": "\\pmb{h} \\in \\mathbb{R}^{D \\times 1}" + }, + { + "bbox": [ + 46, + 228, + 288, + 314 + ], + "type": "text", + "content": " are projected backbone features of two modalities. " + }, + { + "bbox": [ + 46, + 228, + 288, + 314 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 228, + 288, + 314 + ], + "type": "text", + "content": " is the feature dimension. We start by formulating the objective of dominant contrastive framework CLIP, followed by the introduced non-contrastive framework nCLIP. The final framework xCLIP takes the multi-tasking of two objectives, the complete computational pipeline of which is shown in Algorithm 1." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 327, + 287, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 327, + 287, + 389 + ], + "spans": [ + { + "bbox": [ + 46, + 327, + 287, + 389 + ], + "type": "text", + "content": "Contrastive pre-training [59]. Let " + }, + { + "bbox": [ + 46, + 327, + 287, + 389 + ], + "type": "inline_equation", + "content": "\\pmb{u} = \\pmb{g} / ||\\pmb{g}||" + }, + { + "bbox": [ + 46, + 327, + 287, + 389 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 327, + 287, + 389 + ], + "type": "inline_equation", + "content": "\\pmb{v} = \\pmb{h} / ||\\pmb{h}||" + }, + { + "bbox": [ + 46, + 327, + 287, + 389 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 327, + 287, + 389 + ], + "type": "inline_equation", + "content": "\\pmb{U} = [\\pmb{u}_1, \\pmb{u}_2, \\dots, \\pmb{u}_B] \\in \\mathbb{R}^{D \\times B}" + }, + { + "bbox": [ + 46, + 327, + 287, + 389 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 327, + 287, + 389 + ], + "type": "inline_equation", + "content": "\\pmb{V} = [\\pmb{v}_1, \\pmb{v}_2, \\dots, \\pmb{v}_B] \\in \\mathbb{R}^{D \\times B}" + }, + { + "bbox": [ + 46, + 327, + 287, + 389 + ], + "type": "text", + "content": " are concatenated embeddings over the batch. " + }, + { + "bbox": [ + 46, + 327, + 287, + 389 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 46, + 327, + 287, + 389 + ], + "type": "text", + "content": " is the batch size. The contrastive objective is formulated as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 74, + 396, + 287, + 449 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 396, + 287, + 449 + ], + "spans": [ + { + "bbox": [ + 74, + 396, + 287, + 449 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\mathrm {C L I P}} = \\operatorname {I n f o N C E} (\\boldsymbol {U} ^ {\\mathrm {T}} \\boldsymbol {V} / \\sigma) \\\\ = - \\frac {1}{B} \\sum_ {i \\in B} \\log \\frac {\\exp \\left(\\boldsymbol {u} _ {i} ^ {\\mathrm {T}} \\boldsymbol {v} _ {i} / \\sigma\\right)}{\\sum_ {j \\in B} \\exp \\left(\\boldsymbol {u} _ {i} ^ {\\mathrm {T}} \\boldsymbol {v} _ {j} / \\sigma\\right)}, \\tag {1} \\\\ \\end{array}", + "image_path": "d4a6d6c457514eddfbdb358c820e2a498acbf2d9340b5b8bd0a1aded92631259.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 457, + 287, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 457, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 46, + 457, + 287, + 601 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 457, + 287, + 601 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 46, + 457, + 287, + 601 + ], + "type": "text", + "content": " is a trainable parameter controlling the temperature. " + }, + { + "bbox": [ + 46, + 457, + 287, + 601 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{CLIP}}" + }, + { + "bbox": [ + 46, + 457, + 287, + 601 + ], + "type": "text", + "content": " is symmetrized by setting " + }, + { + "bbox": [ + 46, + 457, + 287, + 601 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 46, + 457, + 287, + 601 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 457, + 287, + 601 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 46, + 457, + 287, + 601 + ], + "type": "text", + "content": " as projected features of the images and texts by turns and taking the average of two terms. Vanilla InfoNCE can be decoupled into two terms accounting for affinity and variability separately [81]. While both variability terms in the contrastive and non-contrastive objectives require for batch statistics (e.g., " + }, + { + "bbox": [ + 46, + 457, + 287, + 601 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{EH}}" + }, + { + "bbox": [ + 46, + 457, + 287, + 601 + ], + "type": "text", + "content": " in Eq. (4)), CLIP formulation explicitly maximizes the distance between negative pairs of batch samples. Hence, it degrades the learning performance and data efficiency when models are pre-trained with noisy data where sensible matches occur within negative pairs." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 616, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 616, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 616, + 287, + 689 + ], + "type": "text", + "content": "Non-contrastive pre-training. Let " + }, + { + "bbox": [ + 46, + 616, + 287, + 689 + ], + "type": "inline_equation", + "content": "\\pmb{p} = \\mathrm{softmax}(\\pmb{g})" + }, + { + "bbox": [ + 46, + 616, + 287, + 689 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 616, + 287, + 689 + ], + "type": "inline_equation", + "content": "\\pmb{q} = \\mathrm{softmax}(\\pmb{h})" + }, + { + "bbox": [ + 46, + 616, + 287, + 689 + ], + "type": "text", + "content": ". We transform the projected feature into probability distributions, which can be seen as an assignment over semantic clusters (i.e., projection weights). We take " + }, + { + "bbox": [ + 46, + 616, + 287, + 689 + ], + "type": "inline_equation", + "content": "\\pmb{p}" + }, + { + "bbox": [ + 46, + 616, + 287, + 689 + ], + "type": "text", + "content": " as the target distribution and learn to estimate the predicted distribution " + }, + { + "bbox": [ + 46, + 616, + 287, + 689 + ], + "type": "inline_equation", + "content": "\\pmb{q}" + }, + { + "bbox": [ + 46, + 616, + 287, + 689 + ], + "type": "text", + "content": " by minimizing their cross-entropy:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 125, + 696, + 287, + 711 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 696, + 287, + 711 + ], + "spans": [ + { + "bbox": [ + 125, + 696, + 287, + 711 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {C E}} = - \\boldsymbol {p} ^ {\\mathrm {T}} \\log (\\boldsymbol {q}). \\tag {2}", + "image_path": "c66a1226134dcc4c013ef1870b4f4421cdbdc4ac814c70c3694055e14304d310.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": "Note the target and prediction branch are backpropagated at the same time and " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{CE}}" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": " is symmetrized as " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{CLIP}}" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": ". To avert collapsing solutions, we incorporate entropy regularizers, i.e., entropy minimization [71] and mean entropy maximization [8,9], via minimizing:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 340, + 140, + 545, + 156 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 140, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 340, + 140, + 545, + 156 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {E H}} = - \\boldsymbol {p} ^ {\\mathrm {T}} \\log (\\boldsymbol {p}), - \\mathcal {L} _ {\\mathrm {H E}} = \\bar {\\boldsymbol {p}} ^ {\\mathrm {T}} \\log (\\bar {\\boldsymbol {p}}), \\tag {3}", + "image_path": "d269eaf3098695b21027205b1f6f60bd76ea32a2627ae7c80b24243fab475086.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 165, + 546, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 165, + 546, + 346 + ], + "spans": [ + { + "bbox": [ + 304, + 165, + 546, + 346 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 165, + 546, + 346 + ], + "type": "inline_equation", + "content": "\\overline{\\boldsymbol{p}} = \\mathbb{E}(p) \\approx \\frac{1}{B} \\sum_{i=1}^{B} \\boldsymbol{p}_i" + }, + { + "bbox": [ + 304, + 165, + 546, + 346 + ], + "type": "text", + "content": " is the average distribution over batch. " + }, + { + "bbox": [ + 304, + 165, + 546, + 346 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{EH}}" + }, + { + "bbox": [ + 304, + 165, + 546, + 346 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 165, + 546, + 346 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{HE}}" + }, + { + "bbox": [ + 304, + 165, + 546, + 346 + ], + "type": "text", + "content": " are symmetrized by taking the average with another term where " + }, + { + "bbox": [ + 304, + 165, + 546, + 346 + ], + "type": "inline_equation", + "content": "\\boldsymbol{p}" + }, + { + "bbox": [ + 304, + 165, + 546, + 346 + ], + "type": "text", + "content": " is replaced by " + }, + { + "bbox": [ + 304, + 165, + 546, + 346 + ], + "type": "inline_equation", + "content": "\\boldsymbol{q}" + }, + { + "bbox": [ + 304, + 165, + 546, + 346 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 165, + 546, + 346 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{EH}}" + }, + { + "bbox": [ + 304, + 165, + 546, + 346 + ], + "type": "text", + "content": " encourages the model to make deterministic predictions and ensures assignment's sharpness. " + }, + { + "bbox": [ + 304, + 165, + 546, + 346 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{HE}}" + }, + { + "bbox": [ + 304, + 165, + 546, + 346 + ], + "type": "text", + "content": " encourages the model to utilize a full set of projection weights and ensures the assignment's smoothness. We illustrate in Sec. 4.4 and Appendix A that " + }, + { + "bbox": [ + 304, + 165, + 546, + 346 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{EH}}" + }, + { + "bbox": [ + 304, + 165, + 546, + 346 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 165, + 546, + 346 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{HE}}" + }, + { + "bbox": [ + 304, + 165, + 546, + 346 + ], + "type": "text", + "content": " are minimally sufficient to yield non-collapsing solutions. We empirically show in Tab. A1 that the solutions is non-trivial as long as both sharpness and smoothness are guaranteed, e.g., via Sinkhorn algorithm [14] instead of loss regularizers. We opt for Eq. (3) for its simplicity and consistency between pretraining and evaluation. The overall non-contrastive objective is formulated as" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 343, + 356, + 545, + 369 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 356, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 343, + 356, + 545, + 369 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {n C L I P}} = \\mathcal {L} _ {\\mathrm {C E}} + \\lambda_ {1} \\cdot \\mathcal {L} _ {\\mathrm {E H}} - \\lambda_ {2} \\cdot \\mathcal {L} _ {\\mathrm {H E}}, \\tag {4}", + "image_path": "6e4db602382612be81a11dd9025310fcc1c9e3b34d6ba48528e2fefb70f6a96f.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 379, + 546, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 379, + 546, + 582 + ], + "spans": [ + { + "bbox": [ + 304, + 379, + 546, + 582 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 379, + 546, + 582 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 304, + 379, + 546, + 582 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 379, + 546, + 582 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 304, + 379, + 546, + 582 + ], + "type": "text", + "content": " controls the weight of the regularization. We empirically find that setting " + }, + { + "bbox": [ + 304, + 379, + 546, + 582 + ], + "type": "inline_equation", + "content": "\\lambda_{1} = \\lambda_{2} - 1 = 0.5" + }, + { + "bbox": [ + 304, + 379, + 546, + 582 + ], + "type": "text", + "content": " yields stable training and favorable transfer performances. We detail our findings in Sec. 4.4 and Appendix B.1. During evaluation, we utilize negative cross-entropy " + }, + { + "bbox": [ + 304, + 379, + 546, + 582 + ], + "type": "inline_equation", + "content": "-\\mathcal{L}_{\\mathrm{CE}}" + }, + { + "bbox": [ + 304, + 379, + 546, + 582 + ], + "type": "text", + "content": " as the similarity metric for zero-shot transfer experiments. As schematically showcased in Fig. 1, nCLIP does not reply on negative examples but requires a relatively larger projection head and greater output dimension. Comparing nCLIP with CLIP, we empirically find that nCLIP produces more coarse-grained (e.g., zero-shot retrieval in Sec. 4.1) but semantic-rich (e.g., linear probing in Sec. 4.2) projections. Empirical evidence suggests that the model will fail to deliver reasonable zero-shot and retrieval results if pre-trained without negative pairs (e.g., nCLIP's 25.2 vs. CLIP's 73.8 points of R@1 under Flickr30K I→T retrieval as shown in Tab. 4)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 597, + 545, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 597, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 304, + 597, + 545, + 645 + ], + "type": "text", + "content": "Contrastive meets non-contrastive. Given that two objectives each have their own limitations, we further seek the complementarity between " + }, + { + "bbox": [ + 304, + 597, + 545, + 645 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{CLIP}}" + }, + { + "bbox": [ + 304, + 597, + 545, + 645 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 597, + 545, + 645 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{nCLIP}}" + }, + { + "bbox": [ + 304, + 597, + 545, + 645 + ], + "type": "text", + "content": ", and pre-train the models with both objectives, written as" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 331, + 655, + 545, + 668 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 655, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 331, + 655, + 545, + 668 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {x C L I P}} = \\lambda_ {\\mathrm {C L I P}} \\cdot \\mathcal {L} _ {\\mathrm {C L I P}} + \\lambda_ {\\mathrm {n C L I P}} \\cdot \\mathcal {L} _ {\\mathrm {n C L I P}}, \\tag {5}", + "image_path": "f4aaa5fffb13c162e6f0a7d09714073e075f1efdfcd37aaeff90e177d06573f5.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{CLIP}}" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{nCLIP}}" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": " control the weight of the objectives. During evaluation, we find using the negative cosine metric with CLIP's projection head for zero-shot transfer" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "11030" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 60, + 94, + 225, + 105 + ], + "blocks": [ + { + "bbox": [ + 53, + 75, + 261, + 87 + ], + "lines": [ + { + "bbox": [ + 53, + 75, + 261, + 87 + ], + "spans": [ + { + "bbox": [ + 53, + 75, + 261, + 87 + ], + "type": "text", + "content": "Algorithm 1: PyTorch-like pseudo-code of xCLIP." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 60, + 94, + 225, + 105 + ], + "lines": [ + { + "bbox": [ + 60, + 94, + 225, + 105 + ], + "spans": [ + { + "bbox": [ + 60, + 94, + 225, + 105 + ], + "type": "text", + "content": "// enc, proj: encoder & projector" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "code_body" + } + ], + "index": 1, + "sub_type": "code", + "guess_lang": "typescript" + }, + { + "type": "code", + "bbox": [ + 60, + 106, + 178, + 117 + ], + "blocks": [ + { + "bbox": [ + 60, + 106, + 178, + 117 + ], + "lines": [ + { + "bbox": [ + 60, + 106, + 178, + 117 + ], + "spans": [ + { + "bbox": [ + 60, + 106, + 178, + 117 + ], + "type": "inline_equation", + "content": "\\pmb{f}_I,\\pmb {f}_T = \\mathrm{enc}(\\mathrm{img}),\\mathrm{enc}(\\mathrm{text})" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "code_body" + } + ], + "index": 2, + "sub_type": "code", + "guess_lang": "latex" + }, + { + "type": "code", + "bbox": [ + 60, + 118, + 215, + 129 + ], + "blocks": [ + { + "bbox": [ + 60, + 118, + 215, + 129 + ], + "lines": [ + { + "bbox": [ + 60, + 118, + 215, + 129 + ], + "spans": [ + { + "bbox": [ + 60, + 118, + 215, + 129 + ], + "type": "inline_equation", + "content": "\\pmb{g}_I, \\pmb{g}_T = \\mathrm{proj}_{\\mathrm{CLIP}}(\\pmb{f}_I), \\mathrm{proj}_{\\mathrm{CLIP}}(\\pmb{f}_T)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "code_body" + } + ], + "index": 3, + "sub_type": "code", + "guess_lang": "latex" + }, + { + "type": "code", + "bbox": [ + 60, + 129, + 225, + 141 + ], + "blocks": [ + { + "bbox": [ + 60, + 129, + 225, + 141 + ], + "lines": [ + { + "bbox": [ + 60, + 129, + 225, + 141 + ], + "spans": [ + { + "bbox": [ + 60, + 129, + 225, + 141 + ], + "type": "inline_equation", + "content": "h_{I},h_{T} = \\mathrm{proj}_{\\mathrm{nCLIP}}(\\pmb {f}_{I}),\\mathrm{proj}_{\\mathrm{nCLIP}}(\\pmb{f}_{T})" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "code_body" + } + ], + "index": 4, + "sub_type": "code", + "guess_lang": "latex" + }, + { + "type": "code", + "bbox": [ + 61, + 142, + 208, + 153 + ], + "blocks": [ + { + "bbox": [ + 61, + 142, + 208, + 153 + ], + "lines": [ + { + "bbox": [ + 61, + 142, + 208, + 153 + ], + "spans": [ + { + "bbox": [ + 61, + 142, + 208, + 153 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{xCLIP}} = \\lambda_{\\mathrm{CLIP}}\\cdot \\mathcal{L}_{\\mathrm{CLIP}}(\\pmb {g}_I,\\pmb {g}_T,\\sigma)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "code_body" + } + ], + "index": 5, + "sub_type": "code", + "guess_lang": "latex" + }, + { + "type": "code", + "bbox": [ + 96, + 154, + 237, + 165 + ], + "blocks": [ + { + "bbox": [ + 96, + 154, + 237, + 165 + ], + "lines": [ + { + "bbox": [ + 96, + 154, + 237, + 165 + ], + "spans": [ + { + "bbox": [ + 96, + 154, + 237, + 165 + ], + "type": "text", + "content": "+λnCLIP·LnCLIP(hH,T,λ1,λ2)" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "code_body" + } + ], + "index": 6, + "sub_type": "code", + "guess_lang": "csv" + }, + { + "type": "code", + "bbox": [ + 60, + 166, + 121, + 177 + ], + "blocks": [ + { + "bbox": [ + 60, + 166, + 121, + 177 + ], + "lines": [ + { + "bbox": [ + 60, + 166, + 121, + 177 + ], + "spans": [ + { + "bbox": [ + 60, + 166, + 121, + 177 + ], + "type": "text", + "content": "return " + }, + { + "bbox": [ + 60, + 166, + 121, + 177 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{xCLIP}}" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "code_body" + } + ], + "index": 7, + "sub_type": "code", + "guess_lang": "txt" + }, + { + "type": "code", + "bbox": [ + 60, + 181, + 157, + 193 + ], + "blocks": [ + { + "bbox": [ + 60, + 181, + 157, + 193 + ], + "lines": [ + { + "bbox": [ + 60, + 181, + 157, + 193 + ], + "spans": [ + { + "bbox": [ + 60, + 181, + 157, + 193 + ], + "type": "text", + "content": "def " + }, + { + "bbox": [ + 60, + 181, + 157, + 193 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{CLIP}}(\\pmb {f}_I,\\pmb {f}_T,\\sigma)" + }, + { + "bbox": [ + 60, + 181, + 157, + 193 + ], + "type": "text", + "content": " .." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "code_body" + } + ], + "index": 8, + "sub_type": "code", + "guess_lang": "txt" + }, + { + "type": "code", + "bbox": [ + 75, + 194, + 214, + 203 + ], + "blocks": [ + { + "bbox": [ + 75, + 194, + 214, + 203 + ], + "lines": [ + { + "bbox": [ + 75, + 194, + 214, + 203 + ], + "spans": [ + { + "bbox": [ + 75, + 194, + 214, + 203 + ], + "type": "text", + "content": "// INCE: InfoNCE with Eq. (1)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "code_body" + } + ], + "index": 9, + "sub_type": "code", + "guess_lang": "txt" + }, + { + "type": "code", + "bbox": [ + 75, + 204, + 208, + 215 + ], + "blocks": [ + { + "bbox": [ + 75, + 204, + 208, + 215 + ], + "lines": [ + { + "bbox": [ + 75, + 204, + 208, + 215 + ], + "spans": [ + { + "bbox": [ + 75, + 204, + 208, + 215 + ], + "type": "inline_equation", + "content": "\\pmb{u}_{I}" + }, + { + "bbox": [ + 75, + 204, + 208, + 215 + ], + "type": "text", + "content": " = normalize " + }, + { + "bbox": [ + 75, + 204, + 208, + 215 + ], + "type": "inline_equation", + "content": "(f_{I}, p = 2, \\dim = 1)" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "code_body" + } + ], + "index": 10, + "sub_type": "code", + "guess_lang": "latex" + }, + { + "type": "code", + "bbox": [ + 75, + 216, + 211, + 227 + ], + "blocks": [ + { + "bbox": [ + 75, + 216, + 211, + 227 + ], + "lines": [ + { + "bbox": [ + 75, + 216, + 211, + 227 + ], + "spans": [ + { + "bbox": [ + 75, + 216, + 211, + 227 + ], + "type": "inline_equation", + "content": "\\pmb{u}_{T} = \\mathrm{normalize}(\\pmb{f}_{T}, \\mathrm{p} = 2, \\mathrm{dim} = 1)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "code_body" + } + ], + "index": 11, + "sub_type": "code", + "guess_lang": "txt" + }, + { + "type": "code", + "bbox": [ + 75, + 227, + 249, + 239 + ], + "blocks": [ + { + "bbox": [ + 75, + 227, + 249, + 239 + ], + "lines": [ + { + "bbox": [ + 75, + 227, + 249, + 239 + ], + "spans": [ + { + "bbox": [ + 75, + 227, + 249, + 239 + ], + "type": "inline_equation", + "content": "\\mathcal{L} = \\mathrm{INCE}(\\pmb{U}_I^{\\mathrm{T}}\\pmb{U}_T / \\sigma) + \\mathrm{INCE}(\\pmb{U}_T^{\\mathrm{T}}\\pmb{U}_I / \\sigma)" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "code_body" + } + ], + "index": 12, + "sub_type": "code", + "guess_lang": "latex" + }, + { + "type": "code", + "bbox": [ + 76, + 240, + 127, + 250 + ], + "blocks": [ + { + "bbox": [ + 76, + 240, + 127, + 250 + ], + "lines": [ + { + "bbox": [ + 76, + 240, + 127, + 250 + ], + "spans": [ + { + "bbox": [ + 76, + 240, + 127, + 250 + ], + "type": "text", + "content": "return " + }, + { + "bbox": [ + 76, + 240, + 127, + 250 + ], + "type": "inline_equation", + "content": "\\mathcal{L} / 2" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "code_body" + } + ], + "index": 13, + "sub_type": "code", + "guess_lang": "txt" + }, + { + "type": "code", + "bbox": [ + 60, + 258, + 181, + 269 + ], + "blocks": [ + { + "bbox": [ + 60, + 258, + 181, + 269 + ], + "lines": [ + { + "bbox": [ + 60, + 258, + 181, + 269 + ], + "spans": [ + { + "bbox": [ + 60, + 258, + 181, + 269 + ], + "type": "text", + "content": "def " + }, + { + "bbox": [ + 60, + 258, + 181, + 269 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{nCLIP}}(f_I, f_T, \\lambda_1, \\lambda_2)" + }, + { + "bbox": [ + 60, + 258, + 181, + 269 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "code_body" + } + ], + "index": 14, + "sub_type": "code", + "guess_lang": "latex" + }, + { + "type": "code", + "bbox": [ + 75, + 270, + 179, + 281 + ], + "blocks": [ + { + "bbox": [ + 75, + 270, + 179, + 281 + ], + "lines": [ + { + "bbox": [ + 75, + 270, + 179, + 281 + ], + "spans": [ + { + "bbox": [ + 75, + 270, + 179, + 281 + ], + "type": "inline_equation", + "content": "\\pmb{p}_I = \\mathrm{softmax}(\\pmb {f}_I,\\mathrm{dim} = 1)" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "code_body" + } + ], + "index": 15, + "sub_type": "code", + "guess_lang": "latex" + }, + { + "type": "code", + "bbox": [ + 75, + 281, + 182, + 293 + ], + "blocks": [ + { + "bbox": [ + 75, + 281, + 182, + 293 + ], + "lines": [ + { + "bbox": [ + 75, + 281, + 182, + 293 + ], + "spans": [ + { + "bbox": [ + 75, + 281, + 182, + 293 + ], + "type": "inline_equation", + "content": "\\pmb{p}_T = \\mathrm{softmax}(\\pmb {f}_T,\\mathrm{dim} = 1)" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "code_body" + } + ], + "index": 16, + "sub_type": "code", + "guess_lang": "latex" + }, + { + "type": "code", + "bbox": [ + 75, + 293, + 276, + 304 + ], + "blocks": [ + { + "bbox": [ + 75, + 293, + 276, + 304 + ], + "lines": [ + { + "bbox": [ + 75, + 293, + 276, + 304 + ], + "spans": [ + { + "bbox": [ + 75, + 293, + 276, + 304 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{CE}} = -\\left(\\pmb {p}_I\\cdot \\log (\\pmb {p}_T) + \\pmb {p}_T\\cdot \\log (\\pmb {p}_I)\\right)." + }, + { + "bbox": [ + 75, + 293, + 276, + 304 + ], + "type": "text", + "content": " sum " + }, + { + "bbox": [ + 75, + 293, + 276, + 304 + ], + "type": "inline_equation", + "content": "(\\mathrm{dim} = 1)" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "code_body" + } + ], + "index": 17, + "sub_type": "code", + "guess_lang": "latex" + }, + { + "type": "code", + "bbox": [ + 106, + 305, + 164, + 315 + ], + "blocks": [ + { + "bbox": [ + 106, + 305, + 164, + 315 + ], + "lines": [ + { + "bbox": [ + 106, + 305, + 164, + 315 + ], + "spans": [ + { + "bbox": [ + 106, + 305, + 164, + 315 + ], + "type": "text", + "content": ".mean(dim=0)" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "code_body" + } + ], + "index": 18, + "sub_type": "code", + "guess_lang": "txt" + }, + { + "type": "code", + "bbox": [ + 75, + 316, + 276, + 328 + ], + "blocks": [ + { + "bbox": [ + 75, + 316, + 276, + 328 + ], + "lines": [ + { + "bbox": [ + 75, + 316, + 276, + 328 + ], + "spans": [ + { + "bbox": [ + 75, + 316, + 276, + 328 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{EH}} = -(\\pmb {p}_I\\cdot \\log (\\pmb {p}_I) + \\pmb {p}_T\\cdot \\log (\\pmb {p}_T))." + }, + { + "bbox": [ + 75, + 316, + 276, + 328 + ], + "type": "text", + "content": " sum " + }, + { + "bbox": [ + 75, + 316, + 276, + 328 + ], + "type": "inline_equation", + "content": "(\\dim = 1)" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "code_body" + } + ], + "index": 19, + "sub_type": "code", + "guess_lang": "latex" + }, + { + "type": "code", + "bbox": [ + 106, + 329, + 164, + 339 + ], + "blocks": [ + { + "bbox": [ + 106, + 329, + 164, + 339 + ], + "lines": [ + { + "bbox": [ + 106, + 329, + 164, + 339 + ], + "spans": [ + { + "bbox": [ + 106, + 329, + 164, + 339 + ], + "type": "text", + "content": ".mean(dim=0)" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "code_body" + } + ], + "index": 20, + "sub_type": "code", + "guess_lang": "txt" + }, + { + "type": "code", + "bbox": [ + 75, + 341, + 257, + 353 + ], + "blocks": [ + { + "bbox": [ + 75, + 341, + 257, + 353 + ], + "lines": [ + { + "bbox": [ + 75, + 341, + 257, + 353 + ], + "spans": [ + { + "bbox": [ + 75, + 341, + 257, + 353 + ], + "type": "inline_equation", + "content": "\\overline{\\pmb{p}}_I,\\overline{\\pmb{p}}_T = \\pmb {p}_I.\\mathrm{mean}(\\mathrm{dim} = 0),\\pmb {p}_T.\\mathrm{mean}(\\mathrm{dim} = 0)" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "code_body" + } + ], + "index": 21, + "sub_type": "code", + "guess_lang": "latex" + }, + { + "type": "code", + "bbox": [ + 75, + 353, + 276, + 365 + ], + "blocks": [ + { + "bbox": [ + 75, + 353, + 276, + 365 + ], + "lines": [ + { + "bbox": [ + 75, + 353, + 276, + 365 + ], + "spans": [ + { + "bbox": [ + 75, + 353, + 276, + 365 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{HE}} = -\\left(\\overline{\\boldsymbol{p}}_I\\cdot \\log (\\overline{\\boldsymbol{p}}_I) + \\overline{\\boldsymbol{p}}_T\\cdot \\log (\\overline{\\boldsymbol{p}}_T)\\right)." + }, + { + "bbox": [ + 75, + 353, + 276, + 365 + ], + "type": "text", + "content": " sum(dim=1)" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "code_body" + } + ], + "index": 22, + "sub_type": "code", + "guess_lang": "latex" + }, + { + "type": "code", + "bbox": [ + 75, + 365, + 204, + 376 + ], + "blocks": [ + { + "bbox": [ + 75, + 365, + 204, + 376 + ], + "lines": [ + { + "bbox": [ + 75, + 365, + 204, + 376 + ], + "spans": [ + { + "bbox": [ + 75, + 365, + 204, + 376 + ], + "type": "inline_equation", + "content": "\\mathcal{L} = \\mathcal{L}_{\\mathrm{CE}} + \\lambda_1\\cdot \\mathcal{L}_{\\mathrm{EH}} - \\lambda_2\\cdot \\mathcal{L}_{\\mathrm{HE}}" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "code_body" + } + ], + "index": 23, + "sub_type": "code", + "guess_lang": "latex" + }, + { + "type": "code", + "bbox": [ + 75, + 376, + 127, + 387 + ], + "blocks": [ + { + "bbox": [ + 75, + 376, + 127, + 387 + ], + "lines": [ + { + "bbox": [ + 75, + 376, + 127, + 387 + ], + "spans": [ + { + "bbox": [ + 75, + 376, + 127, + 387 + ], + "type": "text", + "content": "return " + }, + { + "bbox": [ + 75, + 376, + 127, + 387 + ], + "type": "inline_equation", + "content": "\\mathcal{L} / 2" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "code_body" + } + ], + "index": 24, + "sub_type": "code", + "guess_lang": "txt" + }, + { + "bbox": [ + 46, + 401, + 287, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 401, + 287, + 509 + ], + "spans": [ + { + "bbox": [ + 46, + 401, + 287, + 509 + ], + "type": "text", + "content": "experiments generally leads to better results. We note that it is not immediately clear that the two objectives certainly induce stronger models, since they build qualitatively distinctive latent spaces. Qualitatively, we show in Sec. 4 that nCLIP helps CLIP to encode semantics which intrinsically lacks in CLIP, while CLIP helps nCLIP to be transferable for zero-shot recognition. The synergy between the two objectives prompts consistent performance gains across a wide range of tasks by xCLIP over CLIP." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 47, + 519, + 144, + 532 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 519, + 144, + 532 + ], + "spans": [ + { + "bbox": [ + 47, + 519, + 144, + 532 + ], + "type": "text", + "content": "3.2. Implementation" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 46, + 538, + 287, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 538, + 287, + 669 + ], + "spans": [ + { + "bbox": [ + 46, + 538, + 287, + 669 + ], + "type": "text", + "content": "Architecture. We train the base-size model with ViTB/16 [26] as the visual encoder. The model configuration of the base-size text encoder follows CLIP [59] with byte-pair encoding (BPE) and a maximum context length of 77. The projection head generating output for the noncontrastive objective is a 2-layer MLP with 4096-dim hidden layers, GELU [36], and BatchNorm [41]. The last layer is of 32768 output dimension and followed by a BatchNorm without affine transformation. The projection head for the contrastive objective is a single linear layer with no bias and a dimension of 512." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": "Pre-training data. We train our method with publicly available datasets: COCO [51], Visual Genome [45], SBU" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "text", + "content": "Captions [56], Conceptual Caption 3M [65], Conceptual Caption 12M [17], and filtered 14M-size subset of Yahoo Flickr Creative Commons 100M dataset, consisting a total of 35M Image-Text pairs (IT35M). We are also intrigued about the behaviors of models pre-trained on ImagNet-21K [24] (IN21K) dataset taking label names as annotation texts by concatenating them with a sampled prompt, similar to [57, 80]. To notify, ImageNet-21K is a subset of ImageNet-22K with classes of ImageNet-1K excluded for fair downstream evaluation. We detail the models' data scaling behavior in Tab. 8." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 304, + 219, + 545, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 219, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 304, + 219, + 545, + 363 + ], + "type": "text", + "content": "Optimization. We use a batch size of 4096, with all data distributed across 32 V100 GPUs. Models are trained by default with AdamW [53] optimizer, a peak learning rate of " + }, + { + "bbox": [ + 304, + 219, + 545, + 363 + ], + "type": "inline_equation", + "content": "1e^{-3}" + }, + { + "bbox": [ + 304, + 219, + 545, + 363 + ], + "type": "text", + "content": ", cosine scheduler, a weight decay of 0.2, a " + }, + { + "bbox": [ + 304, + 219, + 545, + 363 + ], + "type": "inline_equation", + "content": "\\beta_{2}" + }, + { + "bbox": [ + 304, + 219, + 545, + 363 + ], + "type": "text", + "content": " for AdamW of 0.98, a " + }, + { + "bbox": [ + 304, + 219, + 545, + 363 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 304, + 219, + 545, + 363 + ], + "type": "text", + "content": " for AdamW of " + }, + { + "bbox": [ + 304, + 219, + 545, + 363 + ], + "type": "inline_equation", + "content": "1e^{-6}" + }, + { + "bbox": [ + 304, + 219, + 545, + 363 + ], + "type": "text", + "content": ", and automatic mixed-precision for 3 warm-up epochs and a total of 32 epochs. " + }, + { + "bbox": [ + 304, + 219, + 545, + 363 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 304, + 219, + 545, + 363 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 219, + 545, + 363 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 304, + 219, + 545, + 363 + ], + "type": "text", + "content": " are set as 0.5 and 1.5, respectively. The two objectives are multi-tasked with " + }, + { + "bbox": [ + 304, + 219, + 545, + 363 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{nCLIP}}" + }, + { + "bbox": [ + 304, + 219, + 545, + 363 + ], + "type": "text", + "content": " being 1 and " + }, + { + "bbox": [ + 304, + 219, + 545, + 363 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{CLIP}}" + }, + { + "bbox": [ + 304, + 219, + 545, + 363 + ], + "type": "text", + "content": " being 0.2. Augmentations and pre-processing of pre-training data follow CLIP with image size randomly cropped and resized within the scale of (0.5, 1.0). Detailed pre-training hyper-parameters are detailed in Appendix F." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 306, + 373, + 387, + 387 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 373, + 387, + 387 + ], + "spans": [ + { + "bbox": [ + 306, + 373, + 387, + 387 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 306, + 393, + 419, + 404 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 393, + 419, + 404 + ], + "spans": [ + { + "bbox": [ + 306, + 393, + 419, + 404 + ], + "type": "text", + "content": "4.1. Zero-Shot Transfer" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 304, + 411, + 545, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 411, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 411, + 545, + 544 + ], + "type": "text", + "content": "Classification. We evaluate 27 classification benchmarks with zero-shot classification protocols following [59]. As shown in Tab. 1, nCLIP achieves an average top-1 accuracy of " + }, + { + "bbox": [ + 304, + 411, + 545, + 544 + ], + "type": "inline_equation", + "content": "32.7\\%" + }, + { + "bbox": [ + 304, + 411, + 545, + 544 + ], + "type": "text", + "content": ". Though performing " + }, + { + "bbox": [ + 304, + 411, + 545, + 544 + ], + "type": "inline_equation", + "content": "7.6\\%" + }, + { + "bbox": [ + 304, + 411, + 545, + 544 + ], + "type": "text", + "content": " worse than CLIP, nCLIP demonstrates descent zero-shot recognition capability without explicitly training with negative examples. xCLIP achieves consistent performance gain across a wide range of datasets, with a " + }, + { + "bbox": [ + 304, + 411, + 545, + 544 + ], + "type": "inline_equation", + "content": "43.6\\%" + }, + { + "bbox": [ + 304, + 411, + 545, + 544 + ], + "type": "text", + "content": " average top-1 accuracy, which is " + }, + { + "bbox": [ + 304, + 411, + 545, + 544 + ], + "type": "inline_equation", + "content": "3.3\\%" + }, + { + "bbox": [ + 304, + 411, + 545, + 544 + ], + "type": "text", + "content": " higher than CLIP. It indicates that the synergy between contrastive and non-contrastive objectives improves the model's zero-shot classification ability." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "content": "Out-of-domain classification. For out-of-distribution classification, we use 5 datasets following [59]: ImageNet Adversarial [37], ImageNet Rendition [35], ImageNetV2 [62], ImageNet Sketch [72], and ObjectNet [10] with 2 additional datasets: ImageNet-C [76] and Stylized ImageNet [30]. We investigate if the non-contrastive objective inherits the robustness of the contrastive objective to natural distribution shift. As shown in Tab. 2, nCLIP achieves an " + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "inline_equation", + "content": "18.0\\%" + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "content": " with CLIP an " + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "inline_equation", + "content": "23.1\\%" + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "content": " average top-1 accuracy, revealing a similar trend to in-domain datasets. xCLIP further improves CLIP's accuracy by " + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "inline_equation", + "content": "3.6\\%" + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "content": " and achieves an accuracy of " + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "inline_equation", + "content": "26.7\\%" + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "content": " across 7 datasets, indicating its effectiveness among out-of-domain datasets." + } + ] + } + ], + "index": 34 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11031" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 70, + 544, + 140 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 544, + 140 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 544, + 140 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 544, + 140 + ], + "type": "table", + "html": "
ModelFood101CIFAR10CIFAR100BirdmapSUN397CarsAircraftVOC2007DTDPetsCaltech101FlowersMNISTFER2013STL10EuroSATRESISC45GTSRBKITTICountry211PCAMUCF101Kinetics700CLEVRHapticMemesSSTImageNetAverage
CLIP61.279.750.623.756.515.95.846.427.654.771.348.910.537.391.324.539.713.131.69.150.045.032.412.853.049.145.740.3
nCLIP28.479.549.111.357.05.94.551.422.914.665.023.19.913.594.815.121.22.735.45.851.242.028.412.452.750.037.032.7
xCLIP65.883.454.525.159.918.05.852.233.257.173.950.012.339.092.840.043.616.339.89.351.149.835.418.452.550.248.843.6
", + "image_path": "71f73f15e1ad2444795f6c158743ed05fe9700982f3f138686813fc9949abd20.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 48, + 194, + 286, + 270 + ], + "blocks": [ + { + "bbox": [ + 47, + 143, + 547, + 177 + ], + "lines": [ + { + "bbox": [ + 47, + 143, + 547, + 177 + ], + "spans": [ + { + "bbox": [ + 47, + 143, + 547, + 177 + ], + "type": "text", + "content": "Table 1. Zero-shot classification. We report on a variety of classification benchmarks with ViT-B/16 pre-trained on IT35M. Detailed protocols for each dataset strictly follow CLIP [59]. xCLIP achieves a consistent performance gain compared to CLIP in a wide range of classification datasets. Best results of each column are bolded." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 194, + 286, + 270 + ], + "lines": [ + { + "bbox": [ + 48, + 194, + 286, + 270 + ], + "spans": [ + { + "bbox": [ + 48, + 194, + 286, + 270 + ], + "type": "table", + "html": "
ModelIN-AIN-RIN-v2IN-SkeIN-StyIN-CObjNetAverage
CLIP21.343.737.926.54.911.915.523.1
nCLIP21.529.030.817.73.611.512.018.0
xCLIP27.749.140.630.66.114.717.826.7
", + "image_path": "f14ef0a54a8c331c53fc38bc467b741adf98a852090abf4d9694e039acb38f49.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 49, + 331, + 282, + 403 + ], + "blocks": [ + { + "bbox": [ + 47, + 274, + 287, + 318 + ], + "lines": [ + { + "bbox": [ + 47, + 274, + 287, + 318 + ], + "spans": [ + { + "bbox": [ + 47, + 274, + 287, + 318 + ], + "type": "text", + "content": "Table 2. Zero-shot out-of-distribution classification. We report on a variety of out-of-distribution classification benchmarks. xCLIP demonstrates stronger robustness on various out-of-domain classification datasets." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 331, + 282, + 403 + ], + "lines": [ + { + "bbox": [ + 49, + 331, + 282, + 403 + ], + "spans": [ + { + "bbox": [ + 49, + 331, + 282, + 403 + ], + "type": "table", + "html": "
ModelNUS-WIDEOpenImages
mAPF1@3F1@5mAPF1@10F1@20
CLIP15.133.316.081.113.47.2
nCLIP16.135.616.681.411.66.4
xCLIP15.335.216.881.213.87.4
", + "image_path": "7b1c37cb1bcf3adc3b9e2026ee31c35ff4cacc0875d214d3406145f4b84296ac.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 406, + 287, + 440 + ], + "lines": [ + { + "bbox": [ + 47, + 406, + 287, + 440 + ], + "spans": [ + { + "bbox": [ + 47, + 406, + 287, + 440 + ], + "type": "text", + "content": "Table 3. Zero-shot multi-label classification. We report mAP and F1 scores with ViT-B/16 pre-trained on IT35M. nCLIP showcases the best multi-label classification capability." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 445, + 288, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 445, + 288, + 625 + ], + "spans": [ + { + "bbox": [ + 46, + 445, + 288, + 625 + ], + "type": "text", + "content": "Multi-label classification. We evaluate zero-shot multi-label classification on NUS-WIDE [23] and OpenImages [46] following standard protocol [39]. Specifically, we use 81 unseen labels for NUS-WIDE and the most frequent 400 unseen test labels for OpenImages during evaluation, following [40]. As shown in Tab. 3, we observe that nCLIP shows the best mAP compared to CLIP. nCLIP is pre-trained without negative examples, resulting in the model's intrinsic strengths on recall over overlapped concepts, with 16.1 and 81.4 points of mAP on NUS-WIDE and OpenImages, respectively. Comparatively, xCLIP slightly improves CLIP in this respect while lagging behind nCLIP. The experiments demonstrate that the non-contrastive objective fits well for those downstream tasks in demand of high recall." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "text", + "content": "Retrieval. We evaluate on 2 retrieval benchmarks: Flickr30K [58] and MSCOCO [51] under zero-shot protocol. We do not use prompt engineering and use the original caption. Compared to zero-shot classification, image-text retrieval requires the model's recognition capability at a fine-grained level. We empirically observe in Tab. 4 that" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 198, + 547, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 198, + 547, + 330 + ], + "spans": [ + { + "bbox": [ + 304, + 198, + 547, + 330 + ], + "type": "text", + "content": "nCLIP performs drastically worse than CLIP, which is because text captions in retrieval are not mutually exclusive, which is different from label names in classification. Therefore, explicit negative examples during pre-training play an imperative role in zero-shot retrieval. See Appendix E.1 for further discussions. Beyond that, xCLIP achieves a noticeable performance improvement over CLIP, with a gain of " + }, + { + "bbox": [ + 304, + 198, + 547, + 330 + ], + "type": "inline_equation", + "content": "3.7\\%" + }, + { + "bbox": [ + 304, + 198, + 547, + 330 + ], + "type": "text", + "content": " R@1 on Flickr30K and " + }, + { + "bbox": [ + 304, + 198, + 547, + 330 + ], + "type": "inline_equation", + "content": "3.9\\%" + }, + { + "bbox": [ + 304, + 198, + 547, + 330 + ], + "type": "text", + "content": " R@1 on MSCOCO, illustrating that the additional non-contrastive objective renders extra semantic signals that can be transferred well to fine-grained recognition." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 336, + 446, + 350 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 336, + 446, + 350 + ], + "spans": [ + { + "bbox": [ + 306, + 336, + 446, + 350 + ], + "type": "text", + "content": "4.2. Representation Learning" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 354, + 547, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 354, + 547, + 392 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 547, + 392 + ], + "type": "text", + "content": "We conduct several evaluation protocols to benchmark representation quality under different objectives for both visual (Sec. 4.2.1) and textual (Sec. 4.2.2) modality." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 405, + 476, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 405, + 476, + 418 + ], + "spans": [ + { + "bbox": [ + 305, + 405, + 476, + 418 + ], + "type": "text", + "content": "4.2.1 Visual Representation Learning" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 423, + 547, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 423, + 547, + 675 + ], + "spans": [ + { + "bbox": [ + 304, + 423, + 547, + 675 + ], + "type": "text", + "content": "Linear probing. We evaluate the quality of visual representation via linear probing protocol, where a linear head is fine-tuned on top of the frozen backbones. We follow the same setup as [20] on the same 27 datasets as standard zero-shot classification. Specifically, we use SGD without momentum as the optimizer, no weight decay, and a total epoch of 100 for all evaluation datasets. We use [CLS] token for classification. Following [82], we sweep a set of different learning rates by adding multiple classification heads over a shared frozen backbone, each with its own optimizer and scheduler. We report the best result across different heads. As shown in Tab. 5, we observe that nCLIP generally leads to comparable linear probing accuracy compared with CLIP, suggesting that non-contrastive objectives are able to derive semantically meaningful embedding spaces. nCLIP achieves " + }, + { + "bbox": [ + 304, + 423, + 547, + 675 + ], + "type": "inline_equation", + "content": "2.5\\%" + }, + { + "bbox": [ + 304, + 423, + 547, + 675 + ], + "type": "text", + "content": " higher performance compared to CLIP on ImageNet but relatively lower on average. Beyond that, when combining two objectives together, xCLIP achieves " + }, + { + "bbox": [ + 304, + 423, + 547, + 675 + ], + "type": "inline_equation", + "content": "2.7\\%" + }, + { + "bbox": [ + 304, + 423, + 547, + 675 + ], + "type": "text", + "content": " higher performance on ImageNet and " + }, + { + "bbox": [ + 304, + 423, + 547, + 675 + ], + "type": "inline_equation", + "content": "1.5\\%" + }, + { + "bbox": [ + 304, + 423, + 547, + 675 + ], + "type": "text", + "content": " higher performance on average across 27 different classification tasks compared to CLIP." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 689, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 547, + 713 + ], + "type": "text", + "content": "Fine-tuning & semi-supervised learning. We fine-tune the entire network under the full-data regime (100%) and" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11032" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 70, + 542, + 159 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 542, + 159 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 542, + 159 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 542, + 159 + ], + "type": "table", + "html": "
ModelFlickr30KMSCOCO
I→TT→II→TT→I
R@1R@5R@10R@1R@5R@10R@1R@5R@10R@1R@5R@10
CLIP73.893.397.152.979.387.759.485.291.934.362.774.0
nCLIP26.253.367.920.148.561.121.447.961.613.437.150.3
xCLIP77.595.697.757.382.989.063.387.594.138.466.076.7
", + "image_path": "62a3ca3f9d758302f4e615ca24563f21726ad224f6a55a04677ab6f8089b29e8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 194, + 542, + 236 + ], + "blocks": [ + { + "bbox": [ + 47, + 162, + 546, + 185 + ], + "lines": [ + { + "bbox": [ + 47, + 162, + 546, + 185 + ], + "spans": [ + { + "bbox": [ + 47, + 162, + 546, + 185 + ], + "type": "text", + "content": "Table 4. Zero-shot image-to-text retrieval. We report R@1, R@5, and R@10 in both image-to-text (I→T) and text-to-image (T→I) settings with ViT-B/16 pre-trained on IT35M. xCLIP consistently improves CLIP in retrieval." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 194, + 542, + 236 + ], + "lines": [ + { + "bbox": [ + 50, + 194, + 542, + 236 + ], + "spans": [ + { + "bbox": [ + 50, + 194, + 542, + 236 + ], + "type": "table", + "html": "
ModelFood101CIFAR10CIFAR100BirdsnapSUN397CarsAircraftVOC2007DTDPetsCaltech101FlowersMNISTFER2013STL10EuroSATRESISC45GTSRBKITTICountry211PCAMUCF101Kinetics,700CLEVRHaarSoftMnesSSTImageNetAverage
", + "image_path": "e67cc5f451ddda81e9a5574cd221547b83fd6d53516bd47fc11710726f94dd36.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 52, + 236, + 541, + 266 + ], + "lines": [ + { + "bbox": [ + 52, + 236, + 541, + 266 + ], + "spans": [ + { + "bbox": [ + 52, + 236, + 541, + 266 + ], + "type": "text", + "content": "CLIP 83.3 92.0 75.6 54.4 78.8 60.5 33.8 91.1 72.4 79.0 90.0 96.1 96.2 56.5 97.1 94.8 90.3 77.3 69.7 23.2 83.4 78.9 53.9 50.3 55.9 55.8 71.4 |72.7 nCLIP 83.8 93.2 76.3 66.8 79.3 49.3 28.0 90.1 70.6 70.2 90.5 95.0 95.4 54.7 97.4 91.9 88.2 75.7 64.0 22.7 84.2 79.9 51.0 44.6 55.6 55.2 73.9 |71.4 xCLIP 85.3 93.4 77.8 58.4 80.7 62.3 36.7 92.3 74.0 81.5 91.6 97.0 96.8 58.5 98.1 95.3 91.2 80.4 69.3 26.3 83.5 81.1 56.0 49.8 58.5 54.8 74.1 |74.2" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 52, + 312, + 279, + 383 + ], + "blocks": [ + { + "bbox": [ + 48, + 269, + 546, + 292 + ], + "lines": [ + { + "bbox": [ + 48, + 269, + 546, + 292 + ], + "spans": [ + { + "bbox": [ + 48, + 269, + 546, + 292 + ], + "type": "text", + "content": "Table 5. Linear probing. We report on a variety of classification benchmarks with ViT-B/16 pre-trained on IT35M. CLIP achieves consistent performance gains compared to CLIP in a wide range of classification datasets." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 52, + 312, + 279, + 383 + ], + "lines": [ + { + "bbox": [ + 52, + 312, + 279, + 383 + ], + "spans": [ + { + "bbox": [ + 52, + 312, + 279, + 383 + ], + "type": "table", + "html": "
Modelsm. sup.ft.
1%10%100%
CLIP57.872.482.4
nCLIP55.072.282.4
xCLIP59.573.482.7
", + "image_path": "be7ad7fcd7f783a1b487d30fb50a19d47118190062ef8a525897e74e1dc558f1.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 50, + 465, + 282, + 540 + ], + "blocks": [ + { + "bbox": [ + 46, + 386, + 287, + 452 + ], + "lines": [ + { + "bbox": [ + 46, + 386, + 287, + 452 + ], + "spans": [ + { + "bbox": [ + 46, + 386, + 287, + 452 + ], + "type": "text", + "content": "Table 6. Fine-tuning (ft.) and semi-supervised learning (sm. sup.). Percentage shows sampling ratio for end-to-end fine-tuning. We report top-1 accuracy on the ImageNet-1K validation set. nCLIP performs comparatively with CLIP. xCLIP induces consistent gains in all sampling ratios, especially for semi-supervised learning." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 465, + 282, + 540 + ], + "lines": [ + { + "bbox": [ + 50, + 465, + 282, + 540 + ], + "spans": [ + { + "bbox": [ + 50, + 465, + 282, + 540 + ], + "type": "table", + "html": "
ModelSTS12STS13STS14STS15STS16STS-BSICK-RAverage
CLIP52.354.053.969.958.667.069.360.7
nCLIP51.362.756.971.965.866.964.162.8
xCLIP48.755.655.669.759.566.370.260.8
", + "image_path": "3451df9e99b6703712dc756a3b685176edcda3b561f04652455af8c318ede033.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 543, + 287, + 588 + ], + "lines": [ + { + "bbox": [ + 46, + 543, + 287, + 588 + ], + "spans": [ + { + "bbox": [ + 46, + 543, + 287, + 588 + ], + "type": "text", + "content": "Table 7. Sentence embedding performance on semantic textual similarity. We report on a variety of NLP understanding benchmarks. xCLIP demonstrates the best capability for various STS tasks. xCLIP has comparable performance with CLIP." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "content": "the semi-supervised learning protocol (1% and 10%) on ImageNet-1K [24]. Results are shown in Tab. 6. For CLIP and xCLIP, we find that fine-tuning with the projection head trained under the contrastive objective yields consistent performance gain, especially under a low-data regime. Specifically, we take the text features of the label names as the classifier's initialized parameters. We opt for this setup by default. We highlight that while nCLIP performs worse than CLIP in terms of zero-shot transfer, the non-contrastive objective also serves as a strong baseline to learn represent" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 312, + 546, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 312, + 546, + 361 + ], + "spans": [ + { + "bbox": [ + 305, + 312, + 546, + 361 + ], + "type": "text", + "content": "tation considering the close gap between CLIP and nCLIP in terms of fine-tuning accuracy. xCLIP achieves an " + }, + { + "bbox": [ + 305, + 312, + 546, + 361 + ], + "type": "inline_equation", + "content": "1.7\\%" + }, + { + "bbox": [ + 305, + 312, + 546, + 361 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 312, + 546, + 361 + ], + "type": "inline_equation", + "content": "1.0\\%" + }, + { + "bbox": [ + 305, + 312, + 546, + 361 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 305, + 312, + 546, + 361 + ], + "type": "inline_equation", + "content": "0.3\\%" + }, + { + "bbox": [ + 305, + 312, + 546, + 361 + ], + "type": "text", + "content": " performance gain compared with CLIP under three data ratios of " + }, + { + "bbox": [ + 305, + 312, + 546, + 361 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 305, + 312, + 546, + 361 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 312, + 546, + 361 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 305, + 312, + 546, + 361 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 305, + 312, + 546, + 361 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 305, + 312, + 546, + 361 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 377, + 546, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 377, + 546, + 570 + ], + "spans": [ + { + "bbox": [ + 304, + 377, + 546, + 570 + ], + "type": "text", + "content": "Mask probing. To evaluate how well the visual model is capable of deriving explicit scene layout and object boundaries, we conduct mask probing analysis following [15]. For each attention head from the last layer, we extract the attention map with [CLS] token as the query. We then compute the Jaccard similarity " + }, + { + "bbox": [ + 304, + 377, + 546, + 570 + ], + "type": "inline_equation", + "content": "\\mathcal{J}" + }, + { + "bbox": [ + 304, + 377, + 546, + 570 + ], + "type": "text", + "content": " of each head's attention mask to the ground truth and retain the attention mask with the highest similarity. We conduct experiments on Pascal VOC 2012 [28] dataset. With IT35M, nCLIP demonstrates better quality in the generated mask with 43.7 points of " + }, + { + "bbox": [ + 304, + 377, + 546, + 570 + ], + "type": "inline_equation", + "content": "\\mathcal{J}" + }, + { + "bbox": [ + 304, + 377, + 546, + 570 + ], + "type": "text", + "content": ", while CLIP reaches 41.2 and xCLIP reaches 41.9 points of " + }, + { + "bbox": [ + 304, + 377, + 546, + 570 + ], + "type": "inline_equation", + "content": "\\mathcal{J}" + }, + { + "bbox": [ + 304, + 377, + 546, + 570 + ], + "type": "text", + "content": ". Hence, nCLIP learns better representation for object boundaries compared to CLIP, while xCLIP strikes a balance between nCLIP and CLIP. Detailed results are delayed to Appendix D.1. As a related evaluation, we also study unsupervised segmentation with GroupViT [77] in Appendix D.2." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 586, + 480, + 599 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 586, + 480, + 599 + ], + "spans": [ + { + "bbox": [ + 306, + 586, + 480, + 599 + ], + "type": "text", + "content": "4.2.2 Textual Representation Learning" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": "We follow the setup as [29] and evaluate on 7 STS tasks: STS 2012-2016 [1-5], STS Benchmark [16], and SICK-Relatedness [34]. We directly take [EOS] token without any projection as the input for evaluation, which consistently yields better performance compared to the projected feature for all models. We use cosine distance as the similarity metric. The main goal of sentence embeddings is to cluster semantically similar sentences, and hence, we take STS as one yardstick to benchmark textual representation" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11033" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 51, + 71, + 282, + 155 + ], + "blocks": [ + { + "bbox": [ + 51, + 71, + 282, + 155 + ], + "lines": [ + { + "bbox": [ + 51, + 71, + 282, + 155 + ], + "spans": [ + { + "bbox": [ + 51, + 71, + 282, + 155 + ], + "type": "table", + "html": "
DataSizeEp.ZS / LN Accuracy
CLIPnCLIPxCLIP
①12M2536.8 / 68.537.5 / 71.042.4 / 72.2
④14M3226.5 / 75.131.4 / 77.027.1 / 77.2
①②22M3237.9 / 68.234.2 / 72.343.0 / 71.9
①②③35M3245.7 / 71.437.0 / 73.948.8 / 74.1
", + "image_path": "a925e383d2dd8d3f86388e86842db8bb754b99661953b0267001752ccbef842a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 157, + 287, + 202 + ], + "lines": [ + { + "bbox": [ + 47, + 157, + 287, + 202 + ], + "spans": [ + { + "bbox": [ + 47, + 157, + 287, + 202 + ], + "type": "text", + "content": "Table 8. Pre-training data scaling & domain. Ep. denotes training epochs. Data abbreviations are as follows. ①: CC12M. ②: COCO, VG, SBU, and CC3M. ③: YFCC14M. ④: IN21K. Note ①②③ = IT35M." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 53, + 218, + 164, + 328 + ], + "blocks": [ + { + "bbox": [ + 53, + 218, + 164, + 328 + ], + "lines": [ + { + "bbox": [ + 53, + 218, + 164, + 328 + ], + "spans": [ + { + "bbox": [ + 53, + 218, + 164, + 328 + ], + "type": "image", + "image_path": "d8a2d85afcc2806e3dbf4fcdfa85ee47b8515188eb85100155e54c107714a0d5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 51, + 331, + 164, + 375 + ], + "lines": [ + { + "bbox": [ + 51, + 331, + 164, + 375 + ], + "spans": [ + { + "bbox": [ + 51, + 331, + 164, + 375 + ], + "type": "text", + "content": "Figure 2. Batch size scaling. xCLIP performs better than CLIP with a small batch size (i.e., 1024)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 172, + 218, + 282, + 328 + ], + "blocks": [ + { + "bbox": [ + 172, + 218, + 282, + 328 + ], + "lines": [ + { + "bbox": [ + 172, + 218, + 282, + 328 + ], + "spans": [ + { + "bbox": [ + 172, + 218, + 282, + 328 + ], + "type": "image", + "image_path": "02bb5660dade19153329b4fcde68549a498f3f1a2e08a807bbe25348d5c3a5cc.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 170, + 331, + 282, + 376 + ], + "lines": [ + { + "bbox": [ + 170, + 331, + 282, + 376 + ], + "spans": [ + { + "bbox": [ + 170, + 331, + 282, + 376 + ], + "type": "text", + "content": "Figure 3. Training time. xCLIP adds " + }, + { + "bbox": [ + 170, + 331, + 282, + 376 + ], + "type": "inline_equation", + "content": "\\sim 1.3x" + }, + { + "bbox": [ + 170, + 331, + 282, + 376 + ], + "type": "text", + "content": " training time cost and performs better with equal seen images." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 381, + 287, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 381, + 287, + 490 + ], + "spans": [ + { + "bbox": [ + 46, + 381, + 287, + 490 + ], + "type": "text", + "content": "learning. Results are shown in Tab. 7. nCLIP shows better textual semantics compared to CLIP, with " + }, + { + "bbox": [ + 46, + 381, + 287, + 490 + ], + "type": "inline_equation", + "content": "2.1\\%" + }, + { + "bbox": [ + 46, + 381, + 287, + 490 + ], + "type": "text", + "content": " higher in terms of the average score with IT35M. The conclusion can also be visually extrapolated from the t-SNE visualization of the word embedding, as detailed in Appendix G. We note that xCLIP performs comparatively with CLIP given that STS is evaluated on [EOS] without projection, thus the backbone representation can be similar if dominant by the contrastive objective." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 498, + 118, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 498, + 118, + 510 + ], + "spans": [ + { + "bbox": [ + 47, + 498, + 118, + 510 + ], + "type": "text", + "content": "4.3. Properties" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 517, + 287, + 565 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 517, + 287, + 565 + ], + "spans": [ + { + "bbox": [ + 46, + 517, + 287, + 565 + ], + "type": "text", + "content": "We analyze several crucial properties to demonstrate the generalizability and effectiveness of our method in practical usage. Experiments are conducted with ViT-B/16 on CC12M for 25 epochs by default." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "type": "text", + "content": "Pre-training data scaling. We study the model's performance with data scaling in Tab. 8. xCLIP exhibits desirable scaling behaviors similar to CLIP. The performance improves as data increases. However, we do not observe such a tendency in nCLIP in terms of zero-shot accuracy. For example, nCLIP achieves " + }, + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "type": "inline_equation", + "content": "37.5\\%" + }, + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "type": "text", + "content": " with (1) but only " + }, + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "type": "inline_equation", + "content": "34.2\\%" + }, + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "type": "text", + "content": " with (2), the trend of which defies those of CLIP's " + }, + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "type": "inline_equation", + "content": "(36.8\\%" + }, + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "type": "text", + "content": " vs. " + }, + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "type": "inline_equation", + "content": "37.9\\%)" + }, + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "type": "text", + "content": " and xCLIP's " + }, + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "type": "inline_equation", + "content": "(42.4\\%" + }, + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "type": "text", + "content": " vs. " + }, + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "type": "inline_equation", + "content": "43.0\\%)" + }, + { + "bbox": [ + 46, + 582, + 287, + 713 + ], + "type": "text", + "content": ". On an important note, both nCLIP and xCLIP showcase desirable scaling behaviors under the linear probing protocol. We hypothesize that the performance of zero-shot classification depends on" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 72, + 545, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 109 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 109 + ], + "type": "text", + "content": "pre-training data quality with the non-contrastive objective, while strong representation can always be ensured before the projection." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 125, + 546, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 125, + 546, + 316 + ], + "spans": [ + { + "bbox": [ + 304, + 125, + 546, + 316 + ], + "type": "text", + "content": "Pre-training with tagging data. We consider using tagging data, e.g., ImageNet-21K [24], as pre-training data to validate method's generalizability. To do that, we take label names as annotation texts by concatenating them with a sampled prompt, similar to [57, 80]. As shown in Tab. 8, nCLIP shows superiority to CLIP when pre-trained with IN21K by performance gain of " + }, + { + "bbox": [ + 304, + 125, + 546, + 316 + ], + "type": "inline_equation", + "content": "4.9\\%" + }, + { + "bbox": [ + 304, + 125, + 546, + 316 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 125, + 546, + 316 + ], + "type": "inline_equation", + "content": "1.9\\%" + }, + { + "bbox": [ + 304, + 125, + 546, + 316 + ], + "type": "text", + "content": " in terms of zero-shot and linear probing accuracy, respectively. Compared to nCLIP, we instead notice a performance drop of " + }, + { + "bbox": [ + 304, + 125, + 546, + 316 + ], + "type": "inline_equation", + "content": "4.3\\%" + }, + { + "bbox": [ + 304, + 125, + 546, + 316 + ], + "type": "text", + "content": " for zero-shot accuracy when two objectives multitasked, indicating that the contrastive objective stymies the quality of projected probability under tagging data. Note, that an " + }, + { + "bbox": [ + 304, + 125, + 546, + 316 + ], + "type": "inline_equation", + "content": "77.2\\%" + }, + { + "bbox": [ + 304, + 125, + 546, + 316 + ], + "type": "text", + "content": " linear probing accuracy achieved by xCLIP, " + }, + { + "bbox": [ + 304, + 125, + 546, + 316 + ], + "type": "inline_equation", + "content": "2.1\\%" + }, + { + "bbox": [ + 304, + 125, + 546, + 316 + ], + "type": "text", + "content": " higher than CLIP, is decently strong compared to " + }, + { + "bbox": [ + 304, + 125, + 546, + 316 + ], + "type": "inline_equation", + "content": "79.0\\%" + }, + { + "bbox": [ + 304, + 125, + 546, + 316 + ], + "type": "text", + "content": " achieved by the state of the art in SSL, iBOT [82], pre-trained using the same data with 80 epochs." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 333, + 546, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 333, + 546, + 502 + ], + "spans": [ + { + "bbox": [ + 304, + 333, + 546, + 502 + ], + "type": "text", + "content": "Batch size scaling. We investigate the scaling behavior of CLIP and xCLIP with different pre-training batch sizes. As shown in Fig. 2, xCLIP performs stronger with smaller batch sizes than CLIP pre-trained even with larger batch sizes. For example, xCLIP pre-trained with a batch size of 1024 achieves a zero-shot accuracy of " + }, + { + "bbox": [ + 304, + 333, + 546, + 502 + ], + "type": "inline_equation", + "content": "38.5\\%" + }, + { + "bbox": [ + 304, + 333, + 546, + 502 + ], + "type": "text", + "content": " (vs. CLIP's " + }, + { + "bbox": [ + 304, + 333, + 546, + 502 + ], + "type": "inline_equation", + "content": "36.8\\%" + }, + { + "bbox": [ + 304, + 333, + 546, + 502 + ], + "type": "text", + "content": " with a batch size of 4096). However, we still observe a substantial performance drop as batch size decreases, the behavior of which is different from self-supervised models pre-trained with the non-contrastive objective [14, 15] that are insensitive to batch size. We hypothesize that a large batch size is necessary to approximate " + }, + { + "bbox": [ + 304, + 333, + 546, + 502 + ], + "type": "inline_equation", + "content": "\\mathbb{E}(p)" + }, + { + "bbox": [ + 304, + 333, + 546, + 502 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 304, + 333, + 546, + 502 + ], + "type": "inline_equation", + "content": "\\frac{1}{B}\\sum_{i=1}^{B} p_i" + }, + { + "bbox": [ + 304, + 333, + 546, + 502 + ], + "type": "text", + "content": " in estimating " + }, + { + "bbox": [ + 304, + 333, + 546, + 502 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{HE}}" + }, + { + "bbox": [ + 304, + 333, + 546, + 502 + ], + "type": "text", + "content": ", especially when two probabilities are derived from drastically different modalities." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 518, + 545, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 518, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 304, + 518, + 545, + 602 + ], + "type": "text", + "content": "Computation efficiency. To demonstrate methods' computation efficiency, we show the FLOPs and GPU memory consumption of xCLIP compared to CLIP. xCLIP enforces bearable additional computation cost on top of CLIP with only " + }, + { + "bbox": [ + 304, + 518, + 545, + 602 + ], + "type": "inline_equation", + "content": "1.4\\%" + }, + { + "bbox": [ + 304, + 518, + 545, + 602 + ], + "type": "text", + "content": " extra FLOPs and " + }, + { + "bbox": [ + 304, + 518, + 545, + 602 + ], + "type": "inline_equation", + "content": "27\\%" + }, + { + "bbox": [ + 304, + 518, + 545, + 602 + ], + "type": "text", + "content": " extra GPU memory. As shown in Fig. 3, xCLIP adds " + }, + { + "bbox": [ + 304, + 518, + 545, + 602 + ], + "type": "inline_equation", + "content": "\\sim 1.3x" + }, + { + "bbox": [ + 304, + 518, + 545, + 602 + ], + "type": "text", + "content": " extra time cost but with a " + }, + { + "bbox": [ + 304, + 518, + 545, + 602 + ], + "type": "inline_equation", + "content": "6.0\\%" + }, + { + "bbox": [ + 304, + 518, + 545, + 602 + ], + "type": "text", + "content": " performance gain in terms of zero-shot accuracy." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 610, + 400, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 610, + 400, + 624 + ], + "spans": [ + { + "bbox": [ + 306, + 610, + 400, + 624 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 629, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 715 + ], + "type": "text", + "content": "We show in this section the crucial composing factors of nCLIP and xCLIP. The ablations in the first column (Tabs. 9a to 9c) are conducted with " + }, + { + "bbox": [ + 304, + 629, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{nCLIP}}" + }, + { + "bbox": [ + 304, + 629, + 545, + 715 + ], + "type": "text", + "content": " only. The ablations in the second column (Tabs. 9d to 9f) are conducted with full loss " + }, + { + "bbox": [ + 304, + 629, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{xCLIP}}" + }, + { + "bbox": [ + 304, + 629, + 545, + 715 + ], + "type": "text", + "content": ". Experiments are conducted with ViTB/16 on CC12M for 25 epochs. The default settings are highlighted in cyan." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11034" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 65, + 69, + 197, + 149 + ], + "blocks": [ + { + "bbox": [ + 65, + 69, + 197, + 149 + ], + "lines": [ + { + "bbox": [ + 65, + 69, + 197, + 149 + ], + "spans": [ + { + "bbox": [ + 65, + 69, + 197, + 149 + ], + "type": "table", + "html": "
λ1λ2ZSLN
01NaN
0.3126.869.5
0.5123.767.2
0.21.235.369.0
0.51.537.571.0
", + "image_path": "d33b3f35ca89a5998e0025b3f1c35156ef0d40f75d3e5e9673188151deb899bd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 63, + 174, + 198, + 241 + ], + "blocks": [ + { + "bbox": [ + 52, + 151, + 209, + 171 + ], + "lines": [ + { + "bbox": [ + 52, + 151, + 209, + 171 + ], + "spans": [ + { + "bbox": [ + 52, + 151, + 209, + 171 + ], + "type": "text", + "content": "(a) Coefficient of entropy regularizer. " + }, + { + "bbox": [ + 52, + 151, + 209, + 171 + ], + "type": "inline_equation", + "content": "\\lambda_1 + 1 = \\lambda_2" + }, + { + "bbox": [ + 52, + 151, + 209, + 171 + ], + "type": "text", + "content": " generally leads to good performance." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 63, + 174, + 198, + 241 + ], + "lines": [ + { + "bbox": [ + 63, + 174, + 198, + 241 + ], + "spans": [ + { + "bbox": [ + 63, + 174, + 198, + 241 + ], + "type": "table", + "html": "
#SL / #TLZSLN
0 / 142.472.2
0 / 241.371.3
1 / 242.471.9
2 / 342.572.3
", + "image_path": "6784f194f71fb623d85ab9cce5dc08e81d3b6d9591222d91ef1b2eb50836eae1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 228, + 70, + 361, + 137 + ], + "blocks": [ + { + "bbox": [ + 47, + 243, + 214, + 262 + ], + "lines": [ + { + "bbox": [ + 47, + 243, + 214, + 262 + ], + "spans": [ + { + "bbox": [ + 47, + 243, + 214, + 262 + ], + "type": "text", + "content": "(d) Shared projection layers. #SL denotes shared hidden layers. #TL denotes total hidden layers." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 228, + 70, + 361, + 137 + ], + "lines": [ + { + "bbox": [ + 228, + 70, + 361, + 137 + ], + "spans": [ + { + "bbox": [ + 228, + 70, + 361, + 137 + ], + "type": "table", + "html": "
dimZSLNMem
819231.070.61.07×
1638435.170.91.13×
3276837.571.01.27×
6553635.971.01.52×
", + "image_path": "2939b5a8a5d996f74a8bfbd6ea568b4b5d21430afa23459bf59a18d2badd2540.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 228, + 174, + 364, + 241 + ], + "blocks": [ + { + "bbox": [ + 214, + 150, + 379, + 171 + ], + "lines": [ + { + "bbox": [ + 214, + 150, + 379, + 171 + ], + "spans": [ + { + "bbox": [ + 214, + 150, + 379, + 171 + ], + "type": "text", + "content": "(b) Projection dimension. Mem (G) is compared to CLIP with a 512-dim embedding layer." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 228, + 174, + 364, + 241 + ], + "lines": [ + { + "bbox": [ + 228, + 174, + 364, + 241 + ], + "spans": [ + { + "bbox": [ + 228, + 174, + 364, + 241 + ], + "type": "table", + "html": "
λCLIP: λnCLIPZSLN
1.040.266.5
0.541.770.3
0.242.472.2
0.141.872.6
", + "image_path": "79a3ae0ecfb7b75e00761d10de9d94190eb1ef1e4e458e918fabbd87476ee93b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 387, + 70, + 536, + 137 + ], + "blocks": [ + { + "bbox": [ + 224, + 243, + 369, + 253 + ], + "lines": [ + { + "bbox": [ + 224, + 243, + 369, + 253 + ], + "spans": [ + { + "bbox": [ + 224, + 243, + 369, + 253 + ], + "type": "text", + "content": "(e) Loss ratio between " + }, + { + "bbox": [ + 224, + 243, + 369, + 253 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{CLIP}}" + }, + { + "bbox": [ + 224, + 243, + 369, + 253 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 224, + 243, + 369, + 253 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{nCLIP}}" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 387, + 70, + 536, + 137 + ], + "lines": [ + { + "bbox": [ + 387, + 70, + 536, + 137 + ], + "spans": [ + { + "bbox": [ + 387, + 70, + 536, + 137 + ], + "type": "table", + "html": "
archZSLN
vanilla37.571.0
w/ bottleneck layer31.269.0
w/ l2-norm layer35.470.7
w/o last BN layer34.369.9
", + "image_path": "b59d32ba45eda4eebd5e131c8c3662b76500e0369527138d353367002b2fa7a8.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 383, + 174, + 542, + 241 + ], + "blocks": [ + { + "bbox": [ + 383, + 150, + 541, + 169 + ], + "lines": [ + { + "bbox": [ + 383, + 150, + 541, + 169 + ], + "spans": [ + { + "bbox": [ + 383, + 150, + 541, + 169 + ], + "type": "text", + "content": "(c) Projection architecture of non-contrastive branch." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 383, + 174, + 542, + 241 + ], + "lines": [ + { + "bbox": [ + 383, + 174, + 542, + 241 + ], + "spans": [ + { + "bbox": [ + 383, + 174, + 542, + 241 + ], + "type": "table", + "html": "
techniqueZSLN
vanilla42.472.2
shared space (Eq. (3))29.468.9
debiased sampling42.272.4
warm-up on LnCLIP41.771.9
", + "image_path": "d487fdea8d1f3b7517a7b427afd88ea3558cbb4950ecf56fa8a7157693a8de3e.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 383, + 243, + 541, + 262 + ], + "lines": [ + { + "bbox": [ + 383, + 243, + 541, + 262 + ], + "spans": [ + { + "bbox": [ + 383, + 243, + 541, + 262 + ], + "type": "text", + "content": "(f) Training objective and technique for multitasking." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 49, + 264, + 542, + 276 + ], + "lines": [ + { + "bbox": [ + 49, + 264, + 542, + 276 + ], + "spans": [ + { + "bbox": [ + 49, + 264, + 542, + 276 + ], + "type": "text", + "content": "Table 9. Ablation study with ViT-B/16 on ImageNet-1K validation set. We report zero-shot (ZS) and linear probing (LN) accuracy (\\%)." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 297, + 289, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 297, + 289, + 466 + ], + "spans": [ + { + "bbox": [ + 46, + 297, + 289, + 466 + ], + "type": "text", + "content": "Entropy regularizer. We study the effects of coefficient for entropy regularizers " + }, + { + "bbox": [ + 46, + 297, + 289, + 466 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 46, + 297, + 289, + 466 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 297, + 289, + 466 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 46, + 297, + 289, + 466 + ], + "type": "text", + "content": " in Tab. 9a. The pretraining is unstable without regularizers or with only maximization of the entropy of the mean " + }, + { + "bbox": [ + 46, + 297, + 289, + 466 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\mathrm{HE}})" + }, + { + "bbox": [ + 46, + 297, + 289, + 466 + ], + "type": "text", + "content": ". Specifically, the model will collapse to constant uniform distribution and outputs the same uniform probability distribution despite different inputs. The additional minimization of the mean of the entropy " + }, + { + "bbox": [ + 46, + 297, + 289, + 466 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\mathrm{EH}})" + }, + { + "bbox": [ + 46, + 297, + 289, + 466 + ], + "type": "text", + "content": " stabilizes training, but incurs dimensional collapse, eroding the performances (26.8% and 23.7% vs. 37.5% in terms of zero-shot accuracy). Simultaneously adjusting " + }, + { + "bbox": [ + 46, + 297, + 289, + 466 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 46, + 297, + 289, + 466 + ], + "type": "text", + "content": " with constraints " + }, + { + "bbox": [ + 46, + 297, + 289, + 466 + ], + "type": "inline_equation", + "content": "\\lambda_{1} + 1 = \\lambda_{2}" + }, + { + "bbox": [ + 46, + 297, + 289, + 466 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 297, + 289, + 466 + ], + "type": "inline_equation", + "content": "\\lambda_{1} = 0.5" + }, + { + "bbox": [ + 46, + 297, + 289, + 466 + ], + "type": "text", + "content": " mitigates the problem and yields the optimal performance, with a 37.5% zero-shot and 71.0% linear-probing accuracy. Details are shown in Appendix B.1." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 481, + 287, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 481, + 287, + 637 + ], + "spans": [ + { + "bbox": [ + 46, + 481, + 287, + 637 + ], + "type": "text", + "content": "Projection head. We study the design of two projection heads for " + }, + { + "bbox": [ + 46, + 481, + 287, + 637 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{CLIP}}" + }, + { + "bbox": [ + 46, + 481, + 287, + 637 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 481, + 287, + 637 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{nCLIP}}" + }, + { + "bbox": [ + 46, + 481, + 287, + 637 + ], + "type": "text", + "content": ". As shown in Tab. 9b, large projection dimension matters for good performance. We opt for a projection dimension of 32768 balancing the performance and speed. As shown in Tab. 9c, the last BN layer is crucial and leads to more stable optimization in our experiments. We draw on the idea of prototypes [9, 15] by introducing bottleneck and 12-norm layers, while they do not yield a performance gain. We further study whether two projections can share intermediate computation in Tab. 9d. While sharing middle hidden layers with slightly deeper projection heads performs better, we opt for 2-layer MLP without shared layers for its simplicity." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": "Optimization. We study the effect of hyper-parameters in optimization. In Tab. 9e, we study the optimal loss ratio between " + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{CLIP}}" + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{nCLIP}}" + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": ". We further study whether it's the best route for two objectives to be optimized in separate latent spaces via multi-tasking. Specifically, we con" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 297, + 545, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 297, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 304, + 297, + 545, + 441 + ], + "type": "text", + "content": "sider a bespoke objective Eq. (3) taking account of probability estimation and negative samples simultaneously in a shared latent space. Detailed formulation are shown in Appendix B.2. Optimizing in one latent space yields suboptimal results with a nearly " + }, + { + "bbox": [ + 304, + 297, + 545, + 441 + ], + "type": "inline_equation", + "content": "13.0\\%" + }, + { + "bbox": [ + 304, + 297, + 545, + 441 + ], + "type": "text", + "content": " drop in zero-shot accuracy and " + }, + { + "bbox": [ + 304, + 297, + 545, + 441 + ], + "type": "inline_equation", + "content": "3.3\\%" + }, + { + "bbox": [ + 304, + 297, + 545, + 441 + ], + "type": "text", + "content": " in linear probing accuracy, suggesting two objectives are intrinsically contradictory. We consider debiased sampling, where each batch is sampled from a single data source, since their semantics may be closer and thus more suitable for optimization. This yields similar results. We also add " + }, + { + "bbox": [ + 304, + 297, + 545, + 441 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{nCLIP}}" + }, + { + "bbox": [ + 304, + 297, + 545, + 441 + ], + "type": "text", + "content": " after warm-up epochs with the linear scheduler to its base scale, which erodes the performance." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 451, + 378, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 451, + 378, + 464 + ], + "spans": [ + { + "bbox": [ + 306, + 451, + 378, + 464 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 471, + 545, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 471, + 545, + 711 + ], + "spans": [ + { + "bbox": [ + 304, + 471, + 545, + 711 + ], + "type": "text", + "content": "Aligning images and texts is of overriding significance for vision-language understanding. To conquer the systematic insufficiency of the contrastive objective for acquiring semantics and tackling loose correlations between noisy image-text pairs, we explore the non-contrastive objective for language-image pre-training and unravel its properties. Empirical evidence reveals that the non-contrastive objective induces models to perform favorably in representation learning yet poorly in zero-shot transfer. Observing the distinct mechanisms of the two objectives, we further seek synergy between the two, and introduce a simple multi-tasking framework, xCLIP, that enjoys the best of both worlds: nCLIP aids CLIP mining semantics while CLIP inherits intrinsic strengths for zero-shot recognition. The consistent performance gain of xCLIP over CLIP on a wide variety of downstream tasks consolidates our findings. As potential future work, we may continue to scale up the data size (e.g., LAION400M [64]) as well as the model size (e.g., ViT-L [26]) to verify whether the scaling law applies and the performance improvement endures." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11035" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Eneko Agirre, Carmen Banea, Claire Cardie, Daniel Cer, Mona Diab, Aitor Gonzalez-Agirre, Weiwei Guo, Inigo Lopez-Gazpio, Montse Maritxalar, Rada Mihalcea, et al. Semeval-2015 task 2: Semantic textual similarity, english, spanish and pilot on interpretability. In SemEval, 2015. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 147, + 288, + 200 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 147, + 288, + 200 + ], + "spans": [ + { + "bbox": [ + 53, + 147, + 288, + 200 + ], + "type": "text", + "content": "[2] Eneko Agirre, Carmen Banea, Claire Cardie, Daniel M Cer, Mona T Diab, Aitor Gonzalez-Agirre, Weiwei Guo, Rada Mihalcea, German Rigau, and Janyce Wiebe. Semeval-2014 task 10: Multilingual semantic textual similarity. In SemEval, 2014. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 201, + 288, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 201, + 288, + 255 + ], + "spans": [ + { + "bbox": [ + 53, + 201, + 288, + 255 + ], + "type": "text", + "content": "[3] Eneko Agirre, Carmen Banea, Daniel Cer, Mona Diab, Aitor Gonzalez Agirre, Rada Mihalcea, German Rigau Claramunt, and Janyce Wiebe. Semeval-2016 task 1: Semantic textual similarity, monolingual and cross-lingual evaluation. In SemEval, 2016. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 258, + 287, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 258, + 287, + 289 + ], + "spans": [ + { + "bbox": [ + 53, + 258, + 287, + 289 + ], + "type": "text", + "content": "[4] Eneko Agirre, Daniel Cer, Mona Diab, and Aitor Gonzalez-Agirre. Semeval-2012 task 6: A pilot on semantic textual similarity. In SemEval, 2012. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 291, + 287, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 291, + 287, + 323 + ], + "spans": [ + { + "bbox": [ + 53, + 291, + 287, + 323 + ], + "type": "text", + "content": "[5] Eneko Agirre, Daniel Cer, Mona Diab, Aitor Gonzalez-Agirre, and Weiwei Guo. *sem 2013 shared task: Semantic textual similarity. In SemEval, 2013. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 324, + 287, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 324, + 287, + 378 + ], + "spans": [ + { + "bbox": [ + 53, + 324, + 287, + 378 + ], + "type": "text", + "content": "[6] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. arXiv preprint arXiv:2204.14198, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 380, + 287, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 287, + 401 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 287, + 401 + ], + "type": "text", + "content": "[7] Elad Amrani and Alex Bronstein. Self-supervised classification network. In ECCV, 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 403, + 287, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 403, + 287, + 446 + ], + "spans": [ + { + "bbox": [ + 53, + 403, + 287, + 446 + ], + "type": "text", + "content": "[8] Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Florian Bordes, Pascal Vincent, Armand Joulin, Michael Rabbat, and Nicolas Ballas. Masked siamese networks for label-efficient learning. In ECCV, 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 447, + 287, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 447, + 287, + 501 + ], + "spans": [ + { + "bbox": [ + 53, + 447, + 287, + 501 + ], + "type": "text", + "content": "[9] Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Armand Joulin, Nicolas Ballas, and Michael Rabbat. Semi-supervised learning of visual features by nonparametrically predicting view assignments with support samples. In ICCV, 2021. 3, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 502, + 287, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 502, + 287, + 556 + ], + "spans": [ + { + "bbox": [ + 48, + 502, + 287, + 556 + ], + "type": "text", + "content": "[10] Andrei Barbu, David Mayo, Julian Alverio, William Luo, Christopher Wang, Dan Gutfreund, Josh Tenenbaum, and Boris Katz. Objectnet: A large-scale bias-controlled dataset for pushing the limits of object recognition models. In NeurIPS, 2019. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 558, + 287, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 558, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 558, + 287, + 590 + ], + "type": "text", + "content": "[11] Adrien Bardes, Jean Ponce, and Yann LeCun. Vicreg: Variance-invariance-covariance regularization for self-supervised learning. In ICLR, 2022. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 591, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 591, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 591, + 287, + 635 + ], + "type": "text", + "content": "[12] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. In NeurIPS, 2020. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 636, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 287, + 667 + ], + "type": "text", + "content": "[13] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features. In ECCV, 2018. 2, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "type": "text", + "content": "[14] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. In NeurIPS, 2020. 2, 3, 7" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[15] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 118, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 162 + ], + "type": "text", + "content": "[16] Daniel Cer, Mona Diab, Eneko Agirre, Inigo Lopez-Gazpio, and Lucia Specia. Semeval-2017 task 1: Semantic textual similarity-multilingual and cross-lingual focused evaluation. In SemEval, 2017. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 164, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 164, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 164, + 545, + 206 + ], + "type": "text", + "content": "[17] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing web-scale image-text pretraining to recognize long-tail visual concepts. In CVPR, 2021. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 209, + 545, + 242 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 209, + 545, + 242 + ], + "spans": [ + { + "bbox": [ + 307, + 209, + 545, + 242 + ], + "type": "text", + "content": "[18] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 244, + 545, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 244, + 545, + 266 + ], + "spans": [ + { + "bbox": [ + 307, + 244, + 545, + 266 + ], + "type": "text", + "content": "[19] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In CVPR, 2021. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 268, + 545, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 268, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 307, + 268, + 545, + 300 + ], + "type": "text", + "content": "[20] Xinlei Chen*, Saining Xie*, and Kaiming He. An empirical study of training self-supervised vision transformers. In ICCV, 2021. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 303, + 545, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 303, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 307, + 303, + 545, + 335 + ], + "type": "text", + "content": "[21] Yanbei Chen, Yongqin Xian, Sophia Koepke, Ying Shan, and Zeynep Akata. Distilling audio-visual knowledge by compositional contrastive learning. In CVPR, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 337, + 545, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 337, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 307, + 337, + 545, + 380 + ], + "type": "text", + "content": "[22] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Universal image-text representation learning. In ECCV, 2020. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 383, + 545, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 383, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 307, + 383, + 545, + 426 + ], + "type": "text", + "content": "[23] Tat-Seng Chua, Jinhui Tang, Richang Hong, Haojie Li, Zhiping Luo, and Yan-Tao Zheng. Nus-wide: A real-world web image database from national university of Singapore. In CIVR, 2009. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 429, + 545, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 429, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 307, + 429, + 545, + 460 + ], + "type": "text", + "content": "[24] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. In CVPR, 2009. 2, 4, 6, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 463, + 545, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 463, + 545, + 485 + ], + "spans": [ + { + "bbox": [ + 307, + 463, + 545, + 485 + ], + "type": "text", + "content": "[25] Karan Desai and Justin Johnson. Virtex: Learning visual representations from textual annotations. In CVPR, 2021. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 487, + 545, + 551 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 487, + 545, + 551 + ], + "spans": [ + { + "bbox": [ + 307, + 487, + 545, + 551 + ], + "type": "text", + "content": "[26] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 4, 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 554, + 545, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 554, + 545, + 587 + ], + "spans": [ + { + "bbox": [ + 307, + 554, + 545, + 587 + ], + "type": "text", + "content": "[27] Aleksandr Ermolov, Aliaksandr Siarohin, Enver Sangineto, and Nicu Sebe. Whitening for self-supervised representation learning. In ICML, 2021. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 589, + 545, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 589, + 545, + 621 + ], + "spans": [ + { + "bbox": [ + 307, + 589, + 545, + 621 + ], + "type": "text", + "content": "[28] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. IJCV, 2010. 6" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 624, + 545, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 655 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 655 + ], + "type": "text", + "content": "[29] Tianyu Gao, Xingcheng Yao, and Danqi Chen. SimCSE: Simple contrastive learning of sentence embeddings. In EMNLP, 2021. 2, 6" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "type": "text", + "content": "[30] Robert Geirhos, Patricia Rubisch, Claudio Michaelis, Matthias Bethge, Felix A Wichmann, and Wieland Brendel. Imagenet-trained cnns are biased towards texture; increasing shape bias improves accuracy and robustness. In ICLR, 2019. 4" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "11036" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[31] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. In NeurIPS, 2020. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 128, + 287, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 128, + 287, + 161 + ], + "spans": [ + { + "bbox": [ + 48, + 128, + 287, + 161 + ], + "type": "text", + "content": "[32] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. In CVPR, 2022. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 162, + 287, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 162, + 287, + 194 + ], + "spans": [ + { + "bbox": [ + 48, + 162, + 287, + 194 + ], + "type": "text", + "content": "[33] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 195, + 287, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 195, + 287, + 248 + ], + "spans": [ + { + "bbox": [ + 48, + 195, + 287, + 248 + ], + "type": "text", + "content": "[34] Matthew Henderson, Rami Al-Rfou, Brian Strope, Yun-Hsuan Sung, László Lukács, Ruiqi Guo, Sanjiv Kumar, Balint Miklos, and Ray Kurzweil. Efficient natural language response suggestion for smart reply. arXiv preprint arXiv:1705.00652, 2017. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 250, + 287, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 250, + 287, + 303 + ], + "spans": [ + { + "bbox": [ + 48, + 250, + 287, + 303 + ], + "type": "text", + "content": "[35] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robustness: A critical analysis of out-of-distribution generalization. In ICCV, 2021. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 304, + 287, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 304, + 287, + 336 + ], + "spans": [ + { + "bbox": [ + 48, + 304, + 287, + 336 + ], + "type": "text", + "content": "[36] Dan Hendrycks and Kevin Gimpel. Bridging nonlinearities and stochastic regularizers with gaussian error linear units. In ICLR, 2017. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 338, + 287, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 338, + 287, + 370 + ], + "spans": [ + { + "bbox": [ + 48, + 338, + 287, + 370 + ], + "type": "text", + "content": "[37] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. In CVPR, 2021. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 371, + 287, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 371, + 287, + 414 + ], + "spans": [ + { + "bbox": [ + 48, + 371, + 287, + 414 + ], + "type": "text", + "content": "[38] R Devon Hjelm, Alex Fedorov, Samuel Lavoie-Marchildon, Karan Grewal, Phil Bachman, Adam Trischler, and Yoshua Bengio. Learning deep representations by mutual information estimation and maximization. In ICLR, 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 415, + 287, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 287, + 446 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 287, + 446 + ], + "type": "text", + "content": "[39] Dat Huynh and Ehsan Elhamifar. A shared multi-attention framework for multi-label zero-shot learning. In CVPR, 2020. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 449, + 287, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 449, + 287, + 470 + ], + "spans": [ + { + "bbox": [ + 48, + 449, + 287, + 470 + ], + "type": "text", + "content": "[40] D. Huynh and E. Elhamifar. A shared multi-attention framework for multi-label zero-shot learning. In CVPR, 2020. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 471, + 287, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 471, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 48, + 471, + 287, + 502 + ], + "type": "text", + "content": "[41] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In ICML, 2015. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 504, + 287, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 504, + 287, + 547 + ], + "spans": [ + { + "bbox": [ + 48, + 504, + 287, + 547 + ], + "type": "text", + "content": "[42] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 548, + 287, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 287, + 580 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 287, + 580 + ], + "type": "text", + "content": "[43] Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution or region supervision. In ICML, 2021. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 582, + 287, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 582, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 582, + 287, + 613 + ], + "type": "text", + "content": "[44] Thomas Kipf, Elise van der Pol, and Max Welling. Contrastive learning of structured world models. arXiv preprint arXiv:1911.12247, 2019. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 614, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 287, + 668 + ], + "type": "text", + "content": "[45] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. IJCV, 2017. 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[46] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4. IJCV, 2020. 5" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[47] Janghyeon Lee, Jongsuk Kim, Hyounguk Shon, Bumsoo Kim, Seung Hwan Kim, Honglak Lee, and Junmo Kim. Uniclip: Unified framework for contrastive language-image pretraining. In NeurIPS, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "type": "text", + "content": "[48] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. In NeurIPS, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 162, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 162, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 307, + 162, + 545, + 194 + ], + "type": "text", + "content": "[49] Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, and Kai-Wei Chang. Visualbert: A simple and performant baseline for vision and language. In ACL, 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 195, + 545, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 195, + 545, + 247 + ], + "spans": [ + { + "bbox": [ + 307, + 195, + 545, + 247 + ], + "type": "text", + "content": "[50] Yangguang Li, Feng Liang, Lichen Zhao, Yufeng Cui, Wanli Ouyang, Jing Shao, Fengwei Yu, and Junjie Yan. Supervision exists everywhere: A data efficient contrastive language-image pre-training paradigm. In ICLR, 2022. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 250, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 250, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 307, + 250, + 545, + 293 + ], + "type": "text", + "content": "[51] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 4, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 294, + 545, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 294, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 307, + 294, + 545, + 326 + ], + "type": "text", + "content": "[52] Xingbin Liu, Jinghao Zhou, Tao Kong, Xianming Lin, and Rongrong Ji. Exploring target representations for masked autoencoders. arXiv preprint arXiv:2209.03917, 2022. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 327, + 545, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 327, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 307, + 327, + 545, + 348 + ], + "type": "text", + "content": "[53] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In ICLR, 2019. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 350, + 545, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 350, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 307, + 350, + 545, + 381 + ], + "type": "text", + "content": "[54] Pranay Manocha, Zeyu Jin, Richard Zhang, and Adam Finkelstein. Cdpam: Contrastive learning for perceptual audio similarity. In ICASSP, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 383, + 545, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 383, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 307, + 383, + 545, + 415 + ], + "type": "text", + "content": "[55] Norman Mu, Alexander Kirillov, David Wagner, and Saining Xie. Slip: Self-supervision meets language-image pretraining. In ECCV, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 416, + 545, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 416, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 307, + 416, + 545, + 448 + ], + "type": "text", + "content": "[56] Vicente Ordonez, Girish Kulkarni, and Tamara Berg. Im2text: Describing images using 1 million captioned photographs. In NeuIPS, 2011. 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 449, + 545, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 449, + 545, + 492 + ], + "spans": [ + { + "bbox": [ + 307, + 449, + 545, + 492 + ], + "type": "text", + "content": "[57] Hieu Pham, Zihang Dai, Golnaz Ghiasi, Hanxiao Liu, Adams Wei Yu, Minh-Thang Luong, Mingxing Tan, and Quoc V Le. Combined scaling for zero-shot transfer learning. arXiv preprint arXiv:2111.10050, 2021. 1, 2, 4, 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 494, + 545, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 494, + 545, + 546 + ], + "spans": [ + { + "bbox": [ + 307, + 494, + 545, + 546 + ], + "type": "text", + "content": "[58] Bryan A Plummer, Liwei Wang, Chris M Cervantes, Juan C Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In ICCV, 2015. 2, 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 548, + 545, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 548, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 307, + 548, + 545, + 602 + ], + "type": "text", + "content": "[59] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 2, 3, 4, 5" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 603, + 545, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 603, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 307, + 603, + 545, + 645 + ], + "type": "text", + "content": "[60] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 1" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 647, + 545, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 679 + ], + "type": "text", + "content": "[61] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, 2021. 1" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 681, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 681, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 681, + 545, + 712 + ], + "type": "text", + "content": "[62] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. Do imagenet classifiers generalize to imageNet? In ICML, 2019. 4" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11037" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[63] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In NeurIPS, 2015. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 163 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 163 + ], + "type": "text", + "content": "[64] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. In NeurIPS Workshops, 2021. 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 164, + 287, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 164, + 287, + 207 + ], + "spans": [ + { + "bbox": [ + 48, + 164, + 287, + 207 + ], + "type": "text", + "content": "[65] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In ACL, 2018. 1, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 209, + 287, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 287, + 243 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 287, + 243 + ], + "type": "text", + "content": "[66] Aravind Srinivas, Michael Laskin, and Pieter Abbeel. Curl: Contrastive unsupervised representations for reinforcement learning. In ICML, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 244, + 287, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 244, + 287, + 277 + ], + "spans": [ + { + "bbox": [ + 48, + 244, + 287, + 277 + ], + "type": "text", + "content": "[67] Weijie Su, Xizhou Zhu, Yue Cao, Bin Li, Lewei Lu, Furu Wei, and Jifeng Dai. Vl-bert: Pre-training of generic visual-linguistic representations. In ICLR, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 279, + 287, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 279, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 48, + 279, + 287, + 312 + ], + "type": "text", + "content": "[68] Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. In EMNLP, 2019. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 313, + 287, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 313, + 287, + 357 + ], + "spans": [ + { + "bbox": [ + 48, + 313, + 287, + 357 + ], + "type": "text", + "content": "[69] Bart Thomee, David A Shamma, Gerald Friedland, Benjamin Elizalde, Karl Ni, Douglas Poland, Damian Borth, and Li-Jia Li. Yfcc100m: The new data in multimedia research. In CACM, 2016. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 358, + 287, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 358, + 287, + 413 + ], + "spans": [ + { + "bbox": [ + 48, + 358, + 287, + 413 + ], + "type": "text", + "content": "[70] Yael Vinker, Ehsan Pajouheshgar, Jessica Y Bo, Roman Christian Bachmann, Amit Haim Bermano, Daniel Cohen-Or, Amir Zamir, and Ariel Shamir. Clipasso: Semantically-aware object sketching. In SIGGRAPH, 2022. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 415, + 287, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 287, + 449 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 287, + 449 + ], + "type": "text", + "content": "[71] Feng Wang, Tao Kong, Rufeng Zhang, Huaping Liu, and Hang Li. Self-supervised learning by estimating twin class distributions. arXiv preprint arXiv:2110.07402, 2021. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 450, + 287, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 450, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 48, + 450, + 287, + 483 + ], + "type": "text", + "content": "[72] Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. Learning robust global representations by penalizing local predictive power. In NeurIPS, 2019. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 484, + 287, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 484, + 287, + 518 + ], + "spans": [ + { + "bbox": [ + 48, + 484, + 287, + 518 + ], + "type": "text", + "content": "[73] Zirui Wang, Jiahui Yu, Adams Wei Yu, Zihang Dai, Yulia Tsvetkov, and Yuan Cao. Simvlm: Simple visual language model pretraining with weak supervision. In ICLR, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 519, + 287, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 519, + 287, + 563 + ], + "spans": [ + { + "bbox": [ + 48, + 519, + 287, + 563 + ], + "type": "text", + "content": "[74] Bichen Wu, Ruizhe Cheng, Peizhao Zhang, Tianren Gao, Joseph E. Gonzalez, and Peter Vajda. Data efficient language-supervised zero-shot recognition with optimal transport distillation. In ICLR, 2022. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 565, + 287, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 565, + 287, + 597 + ], + "spans": [ + { + "bbox": [ + 48, + 565, + 287, + 597 + ], + "type": "text", + "content": "[75] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In CVPR, 2018. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 599, + 287, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 599, + 287, + 632 + ], + "spans": [ + { + "bbox": [ + 48, + 599, + 287, + 632 + ], + "type": "text", + "content": "[76] Qizhe Xie, Minh-Thang Luong, Eduard Hovy, and Quoc V Le. Self-training with noisy student improves imagenet classification. In CVPR, 2020. 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 634, + 287, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 634, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 48, + 634, + 287, + 677 + ], + "type": "text", + "content": "[77] Jiarui Xu, Shalini De Mello, Sifei Liu, Wonmin Byeon, Thomas Breuel, Jan Kautz, and Xiaolong Wang. Groupvit: Semantic segmentation emerges from text supervision. In CVPR, 2022. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 679, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 679, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 679, + 287, + 713 + ], + "type": "text", + "content": "[78] Jianwei Yang, Chunyuan Li, Pengchuan Zhang, Bin Xiao, Ce Liu, Lu Yuan, and Jianfeng Gao. Unified contrastive learning in image-text-label space. In CVPR, 2022. 1" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 545, + 230 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 307, + 72, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 545, + 116 + ], + "type": "text", + "content": "[79] Lewei Yao, Runhui Huang, Lu Hou, Guansong Lu, Minzhe Niu, Hang Xu, Xiaodan Liang, Zhenguo Li, Xin Jiang, and Chunjing Xu. Filip: Fine-grained interactive language-image pre-training. In ICLR, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "type": "text", + "content": "[80] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917, 2022. 1, 2, 4, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 163, + 545, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 196 + ], + "type": "text", + "content": "[81] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow twins: Self-supervised learning via redundancy reduction. In ICML, 2021. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 197, + 545, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 230 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 230 + ], + "type": "text", + "content": "[82] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. ibot: Image bert pre-training with online tokenizer. In ICLR, 2022. 5, 7" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "11038" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Non-Contrastive Unsupervised Learning of Physiological Signals From Video/f2bb9236-2bf0-41f5-a980-2c135b55dca1_content_list.json b/2023/Non-Contrastive Unsupervised Learning of Physiological Signals From Video/f2bb9236-2bf0-41f5-a980-2c135b55dca1_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9373368081e95a63ebe4acdc7d4a2820591461e9 --- /dev/null +++ b/2023/Non-Contrastive Unsupervised Learning of Physiological Signals From Video/f2bb9236-2bf0-41f5-a980-2c135b55dca1_content_list.json @@ -0,0 +1,1657 @@ +[ + { + "type": "text", + "text": "Non-Contrastive Unsupervised Learning of Physiological Signals from Video", + "text_level": 1, + "bbox": [ + 99, + 130, + 870, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jeremy Speth, Nathan Vance, Patrick Flynn, Adam Czajka \nUniversity of Notre Dame", + "bbox": [ + 251, + 180, + 715, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{jspeth,nvancel,flynn,aczajka}@nd.edu", + "bbox": [ + 318, + 219, + 645, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Subtle periodic signals such as blood volume pulse and respiration can be extracted from RGB video, enabling noncontact health monitoring at low cost. Advancements in remote pulse estimation - or remote photoplethysmography (rPPG) - are currently driven by deep learning solutions. However, modern approaches are trained and evaluated on benchmark datasets with ground truth from contact-PPG sensors. We present the first non-contrastive unsupervised learning framework for signal regression to mitigate the need for labelled video data. With minimal assumptions of periodicity and finite bandwidth, our approach discovers the blood volume pulse directly from unlabelled videos. We find that encouraging sparse power spectra within normal physiological bandlimits and variance over batches of power spectra is sufficient for learning visual features of periodic signals. We perform the first experiments utilizing unlabelled video data not specifically created for rPPG to train robust pulse rate estimators. Given the limited inductive biases and impressive empirical results, the approach is theoretically capable of discovering other periodic signals from video, enabling multiple physiological measurements without the need for ground truth signals.", + "bbox": [ + 75, + 300, + 473, + 633 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 662, + 209, + 679 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Camera-based vitals estimation is a rapidly growing field enabling non-contact health monitoring in a variety of settings [23]. Although many of the signals avoid detection from the human eye, video data in the visible and infrared ranges contain subtle intensity changes caused by physiological oscillations such as blood volume and respiration. Significant remote photoplethysmography (rPPG) research for estimating the cardiac pulse has leveraged supervised deep learning for robust signal extraction [7, 18, 30, 38, 51, 52]. While the number of successful approaches has rapidly increased, the size of benchmark video datasets with simultaneous vitals recordings has remained relatively stagnant.", + "bbox": [ + 75, + 688, + 468, + 869 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Robust deep learning-based systems for deployment require training on larger volumes of video data with di", + "bbox": [ + 75, + 869, + 468, + 900 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "verse skin tones, lighting, camera sensors, and movement. However, collecting simultaneous video and physiological ground truth with contact-PPG or electrocardiograms (ECG) is challenging for several reasons. First, many hours of high quality videos is an unwieldy volume of data. Second, recording a diverse subject population in conditions representative of real-world activities is difficult to conduct in the lab setting. Finally, synchronizing contact measurements with video is technically challenging, and even contact measurements used for ground truth contain noise.", + "bbox": [ + 496, + 268, + 893, + 420 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Fortunately, recent works find that contrastive unsupervised learning for rPPG is a promising solution to the data scarcity problem [13, 42, 45, 50]. We extend this research line into non-contrastive unsupervised learning to discover periodic signals in video data. With end-to-end unsupervised learning, collecting more representative training data to learn powerful visual features is much simpler, since only video is required without associated medical information.", + "bbox": [ + 496, + 421, + 893, + 541 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we show that non-contrastive unsupervised learning is especially simple when regressing rPPG signals. We find weak assumptions of periodicity are sufficient for learning the minuscule visual features corresponding to the blood volume pulse from unlabelled face videos. The loss functions can be computed in the frequency domain over batches without the need for pairwise or triplet comparisons. Figure 1 compares the proposed approach with supervised and contrastive unsupervised learning approaches.", + "bbox": [ + 496, + 541, + 895, + 678 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This work creates opportunities for scaling deep learning models for camera-based vitals and estimating periodic or quasi-periodic signals from unlabelled data beyond rPPG. Our novel contributions are:", + "bbox": [ + 496, + 678, + 893, + 738 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. A general framework for physiological signal estimation via non-contrastive unsupervised learning (SiNC) by leveraging periodic signal priors.", + "2. The first non-contrastive unsupervised learning method for camera-based vitals measurement.", + "3. The first experiments and results of training with non-rPPG-specific video data without ground truth vitals." + ], + "bbox": [ + 511, + 743, + 890, + 864 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Source code to replicate this work is available at https://github.com/CVRL/SiNC-rPPG.", + "bbox": [ + 500, + 869, + 888, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "14464", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/86b7a5e481e3b4bd3a900a457ea1e7bf1d3c05fff191a7fa1c24b0152a24e020.jpg", + "image_caption": [ + "Figure 1. Overview of the SiNC framework for rPPG compared with traditional supervised and unsupervised learning. Supervised and contrastive losses use distance metrics to the ground truth or other samples. Our framework applies the loss directly to the prediction by shaping the frequency spectrum, and encouraging variance over a batch of inputs. Power outside of the bandlimits is penalized to learn invariances to irrelevant frequencies. Power within the bandlimits is encouraged to be sparsely distributed near the peak frequency." + ], + "image_footnote": [], + "bbox": [ + 76, + 85, + 893, + 315 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 412, + 218, + 426 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Remote Photoplethysmography (rPPG)", + "text_level": 1, + "bbox": [ + 76, + 436, + 415, + 454 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Approaches for remote pulse estimation have shifted over the last decade from blind source separation [34, 35], through linear color transformations [9, 10, 46, 47] to training supervised deep learning-based models [7, 16, 18, 20, 29, 30, 38, 51, 52, 55]. While the color transformations generalize well across datasets, deep learning-based models give better accuracy when tested on data from a similar distribution to the training set. To this end, deep learning research has focused on optimizing architectures for extracting robust spatial and temporal features from the limited datasets.", + "bbox": [ + 75, + 460, + 468, + 611 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To get around the data bottleneck, large synthetic datasets have recently been proposed [24, 49]. The SCAMPS dataset [24] contains 2,800 videos of synthetic avatars with corresponding PPG, EKG, respiration, and facial action units. The UCLA-synthetic dataset [49] contains 480 videos, which are shown to improve performance when trained with real data. A strength of synthetic datasets is their ability to cover the broad range of skin tones compared to physical data. Another solution to the lack of physiological data is unsupervised learning, where a large set of videos and periodic priors on the output signal is sufficient [13,42,45,50]. We discuss these methods in Sec. 2.3.", + "bbox": [ + 75, + 611, + 470, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Unsupervised Learning", + "text_level": 1, + "bbox": [ + 76, + 801, + 294, + 818 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Self-supervised learning is progressing for image representation learning, with two main competing classes of approaches: contrastive and non-contrastive (or regularized) learning [3]. Contrastive approaches [5, 6, 27] define criteria for distinguishing whether two or more samples are", + "bbox": [ + 75, + 824, + 468, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the same or different. Non-contrastive methods augment positive pairs, and enforce variance in the predictions over batches to avoid collapse, in which the model's embeddings reside in a small subspace of the feature space [3]. Distillation methods only use positive samples and avoid collapse by applying a moving average and stop-gradient operator [8, 14]. Another approaches maximize information content of embeddings [11, 53].", + "bbox": [ + 496, + 412, + 893, + 536 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.3. Unsupervised Learning for rPPG", + "text_level": 1, + "bbox": [ + 500, + 544, + 790, + 561 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "All existing unsupervised rPPG approaches are contrastive [13,42,45,50], in which the same model is fed pairs of input videos, and the predictions over similar videos are pulled closer, while the predictions from dissimilar videos are repelled. Gideon et al. [13] were the first to train a deep learning model without labels for rPPG using the contrastive framework. The core of their approach is frequency resampling to create negative samples. Although spatially similar, a resampled video clip contains a different underlying pulse rate, so the model must learn to attend to the temporal dynamics. For their distance function between pairs, they calculated the mean square error between power spectral densities of the model's waveform predictions. While their approach learns to estimate the pulse, their formulation with negative samples is imprecise. The resampled frequency for the negative sample is known, so the relative translation of the power spectrum from the anchor sample can be directly computed. Thus, rather than repelling the estimated spectra, it is more accurate to penalize differences from the known spectra. Furthermore, resampling close to the original sampling rate causes overlap in the power spectra, so repelling the pair is inaccurate. Yuzhe et al. [50]", + "bbox": [ + 496, + 568, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "14465", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "incorporated the previously ignored resampling factor for a soft modification to the InfoNCE loss [33] that scales the desired similarity between pairs by their relative sampling rate. A downside is that their learning framework is not end-to-end unsupervised and requires fine-tuning with PPG labels after the self-supervised stage.", + "bbox": [ + 75, + 90, + 468, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Differently from [13], Contrast-Phys [42] and SLF-RPM [45] consider all samples different from the anchor to be negatives. This assumes that the power spectra will vary between subjects or sufficiently long windows for the same subject. This runs into similar issues with negative pairs as Gideon's approach. Different subjects may have the same pulse rate, so punishing the model for predicting similar frequencies is common during training. Furthermore, the Fast Fourier Transform (FFT) does not produce perfectly sparse decompositions, resulting in spectral overlap even if the heart rate differs by several beats per minute (bpm). As an example, the last column of the second row in Fig. 2 shows the nulls of the main lobe are nearly 30 bpm apart.", + "bbox": [ + 75, + 183, + 470, + 380 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 396, + 166, + 411 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We first formulate signal regression from video. A video sample $x_{i} \\in \\mathbb{R}^{T \\times W \\times H \\times C}$ sampled from a dataset $\\mathcal{D}$ consists of $T$ images of size $W \\times H$ pixels across $C$ channels, captured over time. State-of-the-art methods offer models $f$ that regress a waveform $\\mathbb{R}^T \\ni y_i = f(x_i)$ of the same length as the video. Recently, the task has been effectively modeled end-to-end with the models $f$ being spatiotemporal neural networks [16, 21, 38, 51, 52]. While most previous works are supervised and minimize the loss to a contact pulse measurement, we perform non-contrastive learning using only the model's estimated waveform.", + "bbox": [ + 75, + 422, + 468, + 589 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The key realization is that we can place strong priors on the estimated pulse regarding its bandwidth and periodicity. Observed signals outside the desired frequency range are pollutants, so penalizing the model for carrying them through the forward pass results in invariances to such noisy visual features. We find that it is surprisingly easy to impose the desired constraints in the frequency domain. Thus, all waveforms are transformed into their discrete Fourier components with the FFT before computing all losses in our approach, $F = \\mathrm{FFT}(y)$ . The following sections describe the loss functions and augmentations used during training.", + "bbox": [ + 75, + 590, + 468, + 757 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Losses", + "text_level": 1, + "bbox": [ + 76, + 770, + 163, + 784 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "One of the advantages of unsupervised learning for periodic signals is that we can constrain the solution space significantly. For physiological signals such as respiration and blood volume pulse, we know the healthy upper and lower bounds of the frequencies. We also desire the extracted signal to be sparse in the frequency domain, and that our model filters out noise signals present in the video. With these", + "bbox": [ + 75, + 794, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "constraints, we can greatly simplify the problem of finding good features for the desired signal in the data.", + "bbox": [ + 498, + 90, + 890, + 119 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Bandwidth Loss. One of the most powerful constraints we can place on the model is frequency bandlimits. Past unsupervised methods have used the irrelevant power ratio (IPR) as a validation metric [12, 13, 42] for model selection. We find that it is also effective during model training. The IPR penalizes the model for generating signals outside the desired bandlimits. With lower and upper bandlimits of $a$ and $b$ , respectively, our bandwidth loss becomes:", + "bbox": [ + 496, + 121, + 890, + 242 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL _ {b} = \\frac {1}{\\sum_ {i = - \\infty} ^ {\\infty} F _ {i}} \\left[ \\sum_ {i = - \\infty} ^ {a} F _ {i} + \\sum_ {i = b} ^ {\\infty} F _ {i} \\right], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 252, + 890, + 306 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $F_{i}$ is the power in the $i$ th frequency bin of the predicted signal. This simple loss enforces learning of many invariants, such as movement from respiration, talking, or facial expressions which typically occupy low frequencies. In our experiments we specify the limits as $a = 0.66\\mathrm{Hz}$ to $b = 3\\mathrm{Hz}$ , which corresponds to a common pulse rate range from 40 bpm to 180 bpm. The first column of Fig. 2 shows the result of training exclusively with the bandwidth loss $L_{b}$ . The last row shows that the model concentrates signal power between the bandlimits.", + "bbox": [ + 496, + 311, + 890, + 462 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Sparsity Loss. The pulse rate is the most common physiological marker associated with the blood volume pulse. Since we are primarily interested in the frequency, we can further improve our model by preventing wideband predictions. This also reveals the true signal we aim to discover by ignoring visual dynamics that are not strongly periodic.", + "bbox": [ + 496, + 463, + 890, + 554 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We penalize energy within the bandlimits that is not near the spectral peak:", + "bbox": [ + 498, + 554, + 890, + 585 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL _ {s} = \\frac {1}{\\sum_ {i = a} ^ {b} F _ {i}} \\left[ \\sum_ {i = a} ^ {F ^ {*} - \\Delta_ {F}} F _ {i} + \\sum_ {i = F ^ {*} + \\Delta_ {F}} ^ {b} F _ {i} \\right], \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 595, + 890, + 652 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $F^{*} = \\operatorname{argmax}(F)$ and $\\Delta_F$ are the frequencies of the spectral peak and padding around the peak, respectively. For all experiments $\\Delta_F = 6$ beats per minute [32]. Figure 2 shows the result of training only with the sparsity loss in the second column. For the whole dataset, the power spectrum is sparsely distributed in the low frequencies, effectively filtering frequencies higher than $1\\mathrm{Hz}$ .", + "bbox": [ + 496, + 657, + 890, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Variance Loss. One of the risks of non-contrastive methods is the model collapsing into trivial solutions and making predictions independently of the input features. In regularized methods such as VICReg [3], a hinge loss on the variance over a batch of predictions is used to enforce diverse outputs. We use a similar strategy to avoid model collapse, but instead spread the variance in power spectral densities towards a uniform distribution over the desired frequency band. Our variance loss processes a uniform prior", + "bbox": [ + 496, + 765, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "14466", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/73f0f9e58fa4278de59a7cde858e22205745b33283d03d17bac423c199cc6949.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 87, + 504, + 231 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a1027376b11c005af2d90c484d98ba8068fba2129028fd19b37ae6c00ba91c3d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 88, + 696, + 229 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d90f10d35c3bfa89e4e9233727e03315dc60db5cd2ec38029fe33a0228e31ce3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 88, + 890, + 229 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5d18cba2f2c08ba61a85737653aa36e1f6cf9ecdb0dadefbab6505365b097ade.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 233, + 504, + 364 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c788f5df5c793649808495d38e68eba7be9a09354a4e016ad5d362e993e14c85.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 233, + 696, + 364 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c0ff9ba078c287422e89c05207b730daddd53f87b305be09b17a175c76c59c3a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 233, + 888, + 364 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/dbb4c7c8a43c9aac763e5d37eecc40de405f8402d698cc8a554b1abc44ce62e0.jpg", + "image_caption": [ + "Figure 2. Each column shows predictions from models trained with one or all of the losses for 20 epochs on UBFC-rPPG. The first two rows show a sample in the time and frequency domain, respectively. The last row shows the signal power over the validation set computed by taking the sum of normalized power spectral densities from each sample, then dividing the result by the number of validation samples. The bandwidth loss penalizes signal power outside predefined bandlimits (40 to 180 bpm) to constrain the output space. The sparsity loss encourages a narrow spectrum containing strong periodicity. The variance loss encourages diverse power spectra in a batch, preventing the model from collapsing to a narrow bandwidth. When combined, the model estimates periodic signals within the desired bandlimits." + ], + "image_footnote": [ + "Predicted Ground Truth" + ], + "bbox": [ + 81, + 366, + 504, + 500 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9414fa58b7d7d7b0713aa77f30a39dd06842029cbd70f07eb3717f3f7a1955d6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 367, + 696, + 500 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e79366bfc32a63af994c89c0ae115e2d0b9ac9d76a25c2ed3b374c48effc9dd9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 367, + 888, + 500 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "distribution $P$ over $d$ frequencies, and a batch of $n$ spectral densities, $F = [v_{1},\\dots,v_{n}]$ , where each vector is a $d$ -dimensional frequency decomposition of a predicted waveform. We calculate the normalized sum of densities over the batch, $Q$ , and define the variance loss as the squared Wasserstein distance [15] to the uniform prior:", + "bbox": [ + 75, + 641, + 468, + 733 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {v} = \\frac {1}{d} \\sum_ {i = 1} ^ {d} \\left(\\mathrm {C D F} _ {i} (Q) - \\mathrm {C D F} _ {i} (P)\\right) ^ {2}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 744, + 468, + 787 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where CDF is a cumulative distribution function. The third column of Fig. 2 shows the effect of the variance loss during training. For a single sample, wide-band signals containing multiple frequencies are predicted, and the predicted frequencies cover the task's bandwidth. In our experiments we use a batch size of 20 samples. See the suppl. materials for an ablation experiment on the impact of batch size.", + "bbox": [ + 75, + 794, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Combining All Losses. Summarizing, our training loss function is a simple sum of the aforementioned losses:", + "bbox": [ + 500, + 641, + 888, + 672 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL = L _ {b} + L _ {s} + L _ {v}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 627, + 703, + 890, + 718 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While one could weight particular components of the loss more than others, we specifically formulated the losses to scale them between 0 and 1. In our experiments, we find that a simple summation without weighting gives good performance. The combined loss function encourages the model to search over the supported frequencies to discover visual features for a strong periodic signal. Remarkably, we find that this simple framework is sufficient for learning to regress the blood volume in video, as shown in the last column of Fig. 2.", + "bbox": [ + 496, + 750, + 892, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "14467", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Augmentations", + "text_level": 1, + "bbox": [ + 76, + 90, + 230, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Unlike Gideon et al.'s [13] approach, which only applies frequency augmentations, we apply several augmentations to both the spatial and temporal dimensions to learn invariances to noisy visual signals. In fact, we found that without augmentations, models did not converge during training (see the supplementary material).", + "bbox": [ + 76, + 112, + 468, + 204 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Image Intensity Augmentations. Gaussian noise is added to each pixel with zero mean and a standard deviation of 2 on an image scale from 0 to 255. Each clip is darkened or brightened by adding a constant from a Gaussian distribution with mean 0 and standard deviation of 10.", + "bbox": [ + 76, + 205, + 468, + 279 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Spatial Augmentations. We randomly horizontally flip a video clip with $50\\%$ probability. The spatial dimension of a clip are randomly square cropped down to between half the original length and the original length. The cropped clip is then linearly interpolated back to the original dimensions.", + "bbox": [ + 76, + 279, + 468, + 354 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Temporal Augmentations. With the general assumption that the desired signal is strongly periodic and sparsely represented in the Fourier domain, we randomly flip a video clip along the time dimension with a probability of $50\\%$ . Note that the Fourier decomposition of a time-reversed sinusoid is identical to that of the original sinusoid.", + "bbox": [ + 76, + 354, + 468, + 445 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Frequency Augmentations. Perhaps the most important augmentation is frequency resampling [13], where the video is linearly interpolated to a different frame rate. This augmentation is particularly interesting for rPPG, because it transforms the video input and target signal equivalently along the time dimension, making it equivariant. Given the aforementioned transformations that are invariant, $\\tau (\\cdot)\\sim$ $\\mathcal{T}$ , the equivariant frequency resampling operation, $\\phi (\\cdot)\\sim$ $\\Phi$ , and a model $f(\\cdot)$ that infers a waveform from a video we have the following:", + "bbox": [ + 76, + 446, + 468, + 597 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\phi (f (\\tau (x))) = f (\\phi (\\tau (x))). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 608, + 468, + 626 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This is a powerful augmentation, because it allows us to augment the target distribution along with the video input. In our experiments we randomly resample input clips by a factor $c \\sim U(0.6, 1.4)$ . After applying the resampling augmentation, we scale the bandlimits by $c$ , to avoid penalizing the model if the augmentation pushed the underlying pulse frequency outside of the original bandlimits.", + "bbox": [ + 76, + 635, + 468, + 742 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Datasets", + "text_level": 1, + "bbox": [ + 76, + 753, + 171, + 768 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use PURE [41], UBFC-rPPG [4], and DDPM [39] as benchmark rPPG datasets for training and testing, and CelebV-HQ dataset [56] and HKBU-MARs [17] for unsupervised training only.", + "bbox": [ + 76, + 779, + 468, + 840 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Deception Detection and Physiological Monitoring (DDPM) [39, 43] consists of data from 86 subjects attempting to answer questions deceptively. Interviews were recorded at 90 frames-per-second for more than 10 minutes", + "bbox": [ + 76, + 840, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "on average. Natural conversation and head pose changes make it a difficult and less-constrained rPPG dataset.", + "bbox": [ + 498, + 90, + 890, + 119 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "PURE [41] is a benchmark rPPG dataset consisting of 10 subjects recorded over 6 sessions. Each session lasted approximately 1 minute, and raw video was recorded at 30 fps. The 6 sessions for each subject consisted of: (1) steady, (2) talking, (3) slow head translation, (4) fast head translation, (5) small and (6) medium head rotations. Pulse rates are at or close to the subject's resting rate.", + "bbox": [ + 498, + 121, + 890, + 227 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "UBFC-rPPG [4] contains 1-minute long videos from 42 subjects recorded at 30 fps. Subjects played a time-sensitive mathematical game to raise their heart rates, but head motion is limited during the recording.", + "bbox": [ + 498, + 227, + 890, + 287 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "HKBU 3D Mask Attack with Real World Variations (HKBU-MARs) [17] consists of 12 subjects captured over 6 different lighting configurations with 7 different cameras each, resulting in 504 videos lasting 10 seconds each. The diverse lighting and camera sensors make it a valuable dataset for unsupervised training. We use version 2 of HKBU-MARs, which contains videos with both realistic 3D masks and unmasked subjects.", + "bbox": [ + 498, + 287, + 890, + 409 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "High-Quality Celebrity Video Dataset (CelebV-HQ) [56] is a set of processed YouTube videos containing 35,666 face videos from over 15,000 identities. The videos vary dramatically in length, lighting, emotion, motion, skin tones, and camera sensors. The greatest challenge in harnessing online videos is their reduced quality due to compression before upload and by the video provider. Compression is a known challenge for rPPG, since the blood volume pulse is so subtle optically [25, 31, 32, 36].", + "bbox": [ + 498, + 409, + 890, + 546 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Training Details", + "text_level": 1, + "bbox": [ + 500, + 560, + 661, + 578 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Data Preprocessing", + "text_level": 1, + "bbox": [ + 500, + 585, + 686, + 602 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To prepare the video clips for the spatiotemporal deep learning models, we first extract 68 face landmarks with OpenFace [2]. We then define a bounding box in each frame with the minimum and maximum $(x,y)$ locations by extending the crop horizontally by $5\\%$ to ensure that the cheeks and jaw are present. The top and bottom are extended by $30\\%$ and $5\\%$ of the bounding box height, respectively, to include the forehead and jaw. We further extend the shorter of the two axes to the length of the other to form a square. The cropped frames are then resized to $64\\times 64$ pixels with bicubic interpolation. For faster processing of the massive CelebV-HQ [56] dataset, we instead use MediaPipe Face Mesh [22] for landmarking.", + "bbox": [ + 498, + 609, + 890, + 805 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2. Model Architectures", + "text_level": 1, + "bbox": [ + 500, + 816, + 692, + 830 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use a 3D-CNN architecture similar to [38] without temporal dilations [51]. We use a temporal kernel width of 5, and replace default zero-padding by repeating the edges. Zero-padding along the time dimension can result", + "bbox": [ + 498, + 839, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "14468", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/ea9c8daeed9482427269614a2f864cdeb54a3a3e5645696a10b47eee0c218207.jpg", + "table_caption": [ + "Table 1. Intra-dataset pulse rate estimation results. The best results for a given metric/dataset are bolded, and the second-best results are underlined. For a better comparison with results in the literature, this table follows [42]. MAE: Mean Absolute Error; RMSE: Root Mean Square Error; $r$ : Pearson correlation coefficient." + ], + "table_footnote": [ + "* Some methods in the unsupervised row require fine-tuning on labeled data with a linear classifier.", + "† Some supervised methods were trained with identical data augmentations to SiNC for fair comparison." + ], + "table_body": "
TypesMethodsUBFC-rPPGPUREDDPM
MAE (bpm)RMSE (bpm)rMAE (bpm)RMSE (bpm)rMAE (bpm)RMSE (bpm)r
TraditionalGREEN [44]7.5014.410.627.2317.050.6932.7943.090.04
ICA [34]5.1711.760.653.7612.600.8522.2235.770.40
CHROM [9]2.369.230.870.752.231.0013.4828.530.56
2SR [48]6.9018.500.652.443.060.9822.0839.730.22
POS [47]2.119.110.870.804.110.989.0323.070.70
SupervisedHR-CNN [40]---1.842.370.98---
SynRhythm [29]5.596.820.72------
PulseGAN [37]1.192.100.98------
Dual-GAN [21]0.440.670.990.821.310.99---
RPNet [38]†0.53 ± 0.011.78 ± 0.020.99 ± 0.001.15 ± 0.275.77 ± 1.250.96 ± 0.023.46 ± 0.2412.47 ± 0.680.91 ± 0.01
PhysNet [51]†0.55 ± 0.032.03 ± 0.370.99 ± 0.000.99 ± 0.195.22 ± 0.930.97 ± 0.013.96 ± 0.7613.57 ± 1.740.89 ± 0.03
UnsupervisedGideon2021 [13]1.854.280.932.32.90.99---
SLF-RPM [45]*8.399.700.70------
SimPer [50]*4.24--3.89-----
Contrast-Phys [42]0.641.000.991.001.400.999.70 ± 2.9025.02 ± 6.010.58 ± 0.19
SiNC (ours)0.59 ± 0.001.83 ± 0.040.99 ± 0.000.61 ± 0.061.84 ± 0.401.00 ± 0.005.87 ± 0.1117.44 ± 0.160.81 ± 0.00
", + "bbox": [ + 76, + 141, + 893, + 393 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "in edge effects that add artificial frequencies to the predictions. Early experiments showed that temporal dilations caused aliasing and reduced the bandwidth of the model to specific frequencies. Our losses and framework may be applied to any task and architecture with dense predictions along one or more dimensions. However, popular rPPG architectures such as DeepPhys [7] and MTTS-CAN [18] may be ill-suited for the approach, since they consume very few frames, and the number of time points should be large enough to give sufficient frequency resolution with the FFT. In our experiments, we use the AdamW [19] optimizer with a learning rate of 0.0001. We use a clip length of $T = 120$ frames (4 seconds), and we set the input signal's length to achieve a frequency resolution of 0.33 bpm.", + "bbox": [ + 75, + 446, + 472, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Supervised Training", + "text_level": 1, + "bbox": [ + 76, + 665, + 272, + 683 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To properly compare our approach to its supervised counterpart we use the same model architecture and train it with the commonly used negative Pearson loss between the predicted waveform and the contact sensor ground truth [51]. During training we apply all of the same augmentations except time reversal. Models are trained for 200 epochs on PURE and UBFC-rPPG, and for 40 epochs on DDPM. The model from the epoch with the lowest loss on the validation set is selected for testing.", + "bbox": [ + 75, + 688, + 470, + 825 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.4. Unsupervised Training", + "text_level": 1, + "bbox": [ + 76, + 832, + 290, + 849 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Unsupervised models are trained for the same number of epochs as the supervised setting for both PURE and UBFCrPPG, but we train for an additional 40 epochs on DDPM,", + "bbox": [ + 75, + 854, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "since this dataset is considerably more difficult. We set the batch size to 20 samples during training. Contrary to previous unsupervised approaches [13, 42], we leverage validation sets for model selection by selecting the model with the lowest combined bandwidth and sparsity losses. The creation of the dataset splits is described in the next section.", + "bbox": [ + 496, + 446, + 890, + 537 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.5. Evaluation", + "text_level": 1, + "bbox": [ + 500, + 544, + 620, + 559 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Pulse rates are computed as the highest spectral peak between $0.66\\mathrm{Hz}$ and $3\\mathrm{Hz}$ (equivalent to $40\\mathrm{bpm}$ to $180\\mathrm{bpm}$ ) over a 10-second sliding window. The same procedure is applied to the ground truth waveforms for a reliable evaluation [26]. We apply common error metrics such as mean absolute error (MAE), root mean square error (RMSE), and Pearson correlation coefficient $(r)$ between the pulse rates.", + "bbox": [ + 496, + 568, + 890, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We perform 5-fold cross validation for both PURE and UBFC with the same folds as [13], and use the predefined dataset splits from DDPM [38]. Differently from [13], we use 3 of the folds for training, 1 for validation, and the remaining for testing rather than only training and testing partitions. We train 3 models with different initializations, resulting in 15 models trained on PURE and UBFC each, and 3 models trained on DDPM. We present the mean and standard deviation of the errors in the results.", + "bbox": [ + 496, + 674, + 890, + 809 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6. Results", + "text_level": 1, + "bbox": [ + 500, + 821, + 586, + 837 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.1. Within-Dataset Testing", + "text_level": 1, + "bbox": [ + 500, + 847, + 714, + 863 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1 shows the results for models trained and tested on subject-disjoint partitions from the same datasets. For", + "bbox": [ + 498, + 869, + 890, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "14469", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "PURE and UBFC we achieve MAE lower than 1 bpm, performing better or on par with all traditional and supervised learning approaches. For PURE, our approach gives the lowest MAE and a Pearson $r$ of nearly 1. Performance drops on DDPM due to the overall difficulty of the dataset. SiNC outperforms contrastive approaches, only being surpassed by supervised deep learning models.", + "bbox": [ + 75, + 90, + 467, + 196 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In comparison to other unsupervised methods, Contrast-Phys [42] gives the most competitive performance on all but DDPM. Note that our approach gives the lowest MAE on all datasets, but has higher RMSE. We believe this is due their use of harmonic removal as a post-processing step when estimating the pulse rate, which is not described in [42], but can be found in their publicly available code.", + "bbox": [ + 75, + 198, + 467, + 303 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.2. Cross-Dataset Testing", + "text_level": 1, + "bbox": [ + 76, + 311, + 281, + 328 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/29dfe9f8d7e006b6195df6f27aaf951707c5bcd5c6e831d04312067ed14d6eed.jpg", + "table_caption": [ + "Table 2. Cross-dataset pulse rate estimation performance. The top 3 training datasets are common rPPG benchmarks, while HKBU was not designed for rPPG and has no pulse ground truth. Note that testing on UBFC after training on DDPM performs well, since their frequency distributions are similar, while PURE's pulse rates tend to be much lower." + ], + "table_footnote": [], + "table_body": "
Training DatasetTesting DatasetMethodMAE (bpm)r
DDPMUBFCSiNC0.88 ± 0.250.98 ± 0.01
UBFCContrast-Phys1.14 ± 0.380.96 ± 0.03
UBFCPhysNet1.11 ± 0.420.95 ± 0.02
PURESiNC3.12 ± 1.070.88 ± 0.05
PUREContrast-Phys13.02 ± 6.120.19 ± 0.59
PUREPhysNet1.46 ± 0.340.95 ± 0.02
UBFCDDPMSiNC18.53 ± 0.360.38 ± 0.01
DDPMContrast-Phys22.93 ± 1.020.18 ± 0.04
DDPMPhysNet18.58 ± 0.120.40 ± 0.00
PURESiNC4.02 ± 0.060.86 ± 0.00
PUREContrast-Phys19.61 ± 2.010.33 ± 0.06
PUREPhysNet3.81 ± 0.340.87 ± 0.02
PUREUBFCSiNC6.64 ± 1.760.59 ± 0.10
UBFCContrast-Phys10.22 ± 0.380.45 ± 0.04
UBFCPhysNet7.02 ± 3.350.60 ± 0.13
DDPMSiNC24.92 ± 0.650.20 ± 0.00
DDPMContrast-Phys29.63 ± 0.480.03 ± 0.02
DDPMPhysNet28.03 ± 2.200.13 ± 0.05
HKBU (non-rPPG)UBFCSiNC1.08 ± 0.030.95 ± 0.00
PURESiNC2.43 ± 0.200.90 ± 0.02
DDPMSiNC20.34 ± 0.250.19 ± 0.02
", + "bbox": [ + 78, + 444, + 475, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We perform cross-dataset testing to analyze whether the approach is robust to changes to the lighting, camera sensor, pulse rate distribution, and motion. Table 2 shows the results for SiNC and supervised training with the same architecture. We find that the performance is similar for the supervised and unsupervised approaches when transferring to different data sources. Training on PURE exclusively gives relatively poor results when transferring to", + "bbox": [ + 75, + 780, + 467, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "UBFC-rPPG and DDPM, due to the low pulse rate variability within PURE samples and lack of movement. Training on DDPM gives the best results overall, since the dataset is the largest and captures larger subjects' movements compared to other datasets.", + "bbox": [ + 496, + 90, + 890, + 167 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.3. Training with CelebV-HQ Videos", + "text_level": 1, + "bbox": [ + 498, + 178, + 790, + 194 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Given the abundance of face videos publicly available online, we trained a model on the CelebV-HQ dataset [56]. After processing the available videos with MediaPipe and resampling to 30 fps, our unlabeled dataset consisted of 34,029 videos. We trained the model for 23 epochs and manually stopped training due to a plateau in the validation loss. Unfortunately, we found that the model could not converge to the true blood volume pulse. We attribute the failure to poor video quality from compression. Although the videos were downloaded with the highest available quality, they have likely been compressed, removing the pulse signal entirely. See the supplementary material for an assessment of rPPG quality using POS [47] on CelebV-HQ.", + "bbox": [ + 496, + 200, + 890, + 398 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.4. Training with HKBU-MARs Videos", + "text_level": 1, + "bbox": [ + 498, + 410, + 808, + 425 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The HKBU-MARs dataset [17] was designed for face presentation attack detection, but we trained models on the \"real\" video sessions in the dataset. The bottom rows in Table 2 show the results for training on HKBU-MARs, then testing on the benchmark rPPG datasets. Training on HKBU-MARs gives better results when testing on UBFC-rPPG and PURE than all training sets except DDPM, which is an order of magnitude larger. To our knowledge, this is the first successful experiment showing that non-rPPG videos can be used to train robust rPPG models, even if they do not have ground-truth pulse labels.", + "bbox": [ + 496, + 434, + 890, + 599 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.5. Ablation Study on Losses", + "text_level": 1, + "bbox": [ + 500, + 612, + 728, + 628 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8f5d1f9733a14bc67c6d99e08ffdfac85217cfa8685e9b0df127ed636b67cee4.jpg", + "table_caption": [ + "Table 3. Ablation study on the loss functions used during training. Results are shown for models trained and tested on UBFC-rPPG." + ], + "table_footnote": [], + "table_body": "
LossMAE (bpm)RMSE (bpm)r
Lb3.08 ± 1.698.08 ± 3.610.87 ±0.08
Ls45.50 ± 1.2250.04 ± 0.94-0.04 ±0.08
Lv22.89 ± 2.8331.51 ± 2.360.22 ±0.09
Ls + Lv51.24 ± 5.3657.80 ± 7.39-0.04 ±0.09
Lb + Ls9.99 ± 2.5517.14 ± 2.360.51 ±0.14
Lb + Lv4.18 ± 2.888.90 ± 5.240.82 ±0.14
Lb + Ls + Lv0.59 ± 0.001.83 ± 0.040.99 ± 0.00
", + "bbox": [ + 506, + 689, + 880, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We trained models using all combinations of loss components to analyze their contributions. Table 3 shows the results for training and testing on UBFC-rPPG. The bandwidth loss is the most critical for discovering the true blood volume pulse, while the sparsity and variance losses do not", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "14470", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0d213413509265e049b90d0f45060212d169c24cc73034bdab220e757a890931.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 89, + 467, + 172 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ba9286360b5aae1a58396a024019804f765b49d433e1589728c7e53827cd5e2d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 175, + 467, + 258 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2a5daec4bc3fe5d7907ed4f892d1f55781180759b00821dd5d4d397f39cb90d9.jpg", + "image_caption": [ + "Figure 3. Within-dataset waveform predictions on all baseline datasets from end-to-end unsupervised models over an 8-second window. The model predictions are remarkably periodic without any form of filtering. Note that phase is not considered during training, so each model learns its own phase shift." + ], + "image_footnote": [], + "bbox": [ + 80, + 260, + 467, + 378 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "learn the desired signal by themselves. Surprisingly, combining the bandwidth loss with just one of the sparsity or variance losses gives worse performance than just the bandwidth loss. However, when combining all three components, the model achieves impressive results.", + "bbox": [ + 75, + 500, + 468, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Discussion", + "text_level": 1, + "bbox": [ + 76, + 590, + 191, + 607 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7.1. Improvements over Supervised Learning", + "text_level": 1, + "bbox": [ + 76, + 618, + 428, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "It is initially surprising that unsupervised training leads to similar or improved rPPG estimation models compared to those trained in a supervised manner. However, there are several potential benefits to unsupervised training. From a hardware perspective, one of the difficulties in supervised training is aligning the contact pulse waveform with the video frames [54]. The pulse sensor and camera may have a time lag, effectively giving the model an out-of-phase target at training time. Unsupervised training gives the model freedom to learn the phase directly from the video. The contact-PPG signal is also sensitive to motion and may be noisy. Since motion may co-occur at the face and fingertip, the contact signal may misguide the model towards visual features for which they should be invariant.", + "bbox": [ + 75, + 642, + 468, + 853 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "From a physiological perspective, the pulse observed optically at the fingertip with a contact sensor has a different phase than that of the face, since blood propagates along a", + "bbox": [ + 75, + 854, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "different path before reaching the peripheral microvasculature, making alignment nearly impossible without shifting the targets to rPPG estimates from existing methods [38]. Additionally, the morphological shape of the contact-PPG waveform depends on numerous factors such as the wavelength of light (and corresponding tissue penetration depth), external pressure from the oximeter clip, and vasodilation at the measurement site [1,28]. This indicates that the morphology and phase of the target PPG waveform is likely different from the observed rPPG waveform.", + "bbox": [ + 496, + 90, + 890, + 241 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7.2. Why Does It Work?", + "text_level": 1, + "bbox": [ + 500, + 250, + 689, + 265 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The success of SiNC depends on specific properties of the data, model, and how the two interact. Limited model capacity is actually required, since it forces learning features that generalize across inputs. By constraining the model's predictions to have specific periodic properties, the limited-capacity model must find a general set of features to produce a signal that exists in most or all of the training samples, which is the blood volume pulse in our datasets.", + "bbox": [ + 496, + 272, + 890, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As a beneficial side-effect, the model intrinsically learns to ignore common noise factors such as illumination, rigid motion, non-rigid motion (e.g. talking, smiling, etc.), and sensor noise, since they may preside outside the predefined bandlimits or with uniform power spectra. Even if noise exhibits periodic tendencies within the bandlimits for some samples, those features would produce poor signals on other samples. Therefore, end-to-end unsupervised approaches like SiNC are well-suited for periodic problems.", + "bbox": [ + 496, + 393, + 890, + 529 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "8. Conclusions", + "text_level": 1, + "bbox": [ + 500, + 541, + 625, + 556 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We proposed a novel non-contrastive learning approach for end-to-end unsupervised signal regression, with specific experiments on blood volume pulse estimation from face videos. This SiNC framework effectively learns powerful visual features with only loose frequency constraints. We demonstrated this by training accurate rPPG models using non-rPPG data and our simple loss functions. Given the subtlety of the rPPG signal, we believe our work can be extended to other signal regression tasks in the domain of remote vitals estimation.", + "bbox": [ + 496, + 566, + 890, + 715 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This research was sponsored by the Securiport Global Innovation Cell, a division of Securiport LLC. Commercial equipment is identified in this work in order to adequately specify or describe the subject matter. In no case does such identification imply recommendation or endorsement by Securiport LLC, nor does it imply that the equipment identified is necessarily the best available for this purpose. The opinions, findings, and conclusions or recommendations expressed in this publication are those of the authors and do not necessarily reflect the views of our sponsors.", + "bbox": [ + 496, + 734, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "14471", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Pierre Abraham, Mélissa Bourgeau, Maïte Camo, Anne Humeau-Heurtier, Sylvain Durand, Pascal Rousseau, and Guillaume Mahe. Effect of skin temperature on skin endothelial function assessment. Microvascular Research, 88:56-60, 2013. 8", + "[2] T. Baltrusaitis, A. Zadeh, Y. C. Lim, and L. Morency. Openface 2.0: Facial behavior analysis toolkit. In IEEE International Conference on Automatic Face Gesture Recognition (FG), pages 59-66, 2018. 5", + "[3] Adrien Bardes, Jean Ponce, and Yann LeCun. Vicreg: Variance-invariance-covariance regularization for self-supervised learning. In ICLR, 2022. 2, 3", + "[4] Serge Bobbia, Richard Macwan, Yannick Benezeth, Alamin Mansouri, and Julien Dubois. Unsupervised skin tissue segmentation for remote photoplethysmography. Pattern Recognition Letters, 124:82-90, 2019. 5", + "[5] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in Neural Information Processing Systems, 2020-Decem(NeurIPS):1-13, 2020. arXiv: 2006.09882. 2", + "[6] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. 37th International Conference on Machine Learning, ICML 2020, pages 1575-1585, 2020. arXiv: 2002.05709 ISBN: 9781713821120. 2", + "[7] Weixuan Chen and Daniel McDuff. DeepPhys: Video-based physiological measurement using convolutional attention networks. In European Conference on Computer Vision (ECCV), pages 356-373, 2018. 1, 2, 6", + "[8] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 15745–15753, 2021. arXiv: 2011.10566 ISBN: 9781665445092. 2", + "[9] G. De Haan and V. Jeanne. Robust pulse rate from chrominance-based rppg. IEEE Transactions on Biomedical Engineering, 60(10):2878-2886, 2013. 2, 6", + "[10] Gerard De Haan and Arno Van Leest. Improved motion robustness of remote-PPG by using the blood volume pulse signature. Physiological Measurement, 35(9):1913–1926, 2014. 2", + "[11] Aleksandr Ermolov, Aliaksandr Siarohin, Enver Sangineto, and Nicu Sebe. Whitening for self-supervised representation learning. In International Conference on Machine Learning, pages 3015-3024. PMLR, 2021. 2", + "[12] John Gideon and Simon Stent. Estimating heart rate from unlabelled video. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, pages 2743-2749, October 2021. 3", + "[13] John Gideon and Simon Stent. The way to my heart is through contrastive learning: Remote photoplethysmography from unlabelled video. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3995-4004, October 2021. 1, 2, 3, 5, 6" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Jean Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre H. Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Daniel Guo, Mohammad Gheshlaghi Azar, Bilal Piot, Koray Kavukcuoglu, Rémi Munos, and Michal Valko. Bootstrap your own latent a new approach to self-supervised learning. Advances in Neural Information Processing Systems, 2020-Decem(NeurIPS):1-14, 2020. arXiv: 2006.07733. 2", + "[15] Le Hou, Chen-Ping Yu, and Dimitris Samaras. Squared Earth Mover's Distance-based Loss for Training Deep Neural Networks, Apr. 2017. arXiv:1611.05916 [cs]. 4", + "[16] Eugene Lee, Evan Chen, and Chen-Yi Lee. Meta-rppg: Remote heart rate estimation using a transductive meta-learner. In European Conference on Computer Vision (ECCV), 2020. 2, 3", + "[17] Siqi Liu, Baoyao Yang, Pong C. Yuen, and Guoying Zhao. A 3d mask face anti-spoofing database with real world variations. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, June 2016. 5, 7", + "[18] Xin Liu, Josh Fromm, Shwetak Patel, and Daniel McDuff. Multi-task temporal shift attention networks for on-device contactless vitals measurement. In H. Larochelle, M. Ranzato, R. Hadsell, M. F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 19400-19411. Curran Associates, Inc., 2020. 1, 2, 6", + "[19] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019. 6", + "[20] Hao Lu, Hu Han, and S Kevin Zhou. Dual-GAN: Joint BVP and Noise Modeling for Remote Physiological Measurement. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 12404-12413, 2021. 2", + "[21] Hao Lu, Hu Han, and S Kevin Zhou. Dual-gan: Joint bvp and noise modeling for remote physiological measurement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12404-12413, 2021. 3, 6", + "[22] Camillo Lugaresi, Jiuqiang Tang, Hadon Nash, Chris McClanahan, Esha Uboweja, Michael Hays, Fan Zhang, Chuoling Chang, Ming Guang Yong, Juhyun Lee, et al. Mediapipe: A framework for building perception pipelines. arXiv preprint arXiv:1906.08172, 2019. 5", + "[23] Daniel McDuff. Camera measurement of physiological vital signs. ACM Comput. Surv., aug 2022. Just Accepted. 1", + "[24] Daniel McDuff, Miah Wander, Xin Liu, Brian L. Hill, Javier Hernandez, Jonathan Lester, and Tadas Baltrusaitis. SCAMPS: Synthetics for camera measurement of physiological signals. In Thirty-sixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2022. 2", + "[25] Daniel J. McDuff, Ethan B. Blackford, and Justin R. Estepp. The Impact of Video Compression on Remote Cardiac Pulse Measurement Using Imaging Photoplethysmography. IEEE International Conference on Automatic Face and Gesture Recognition Workshops (FG), pages 63-70, 2017. 5", + "[26] Yuriy Mironenko, Konstantin Kalinin, Mikhail Kopeliovich, and Mikhail Petrushan. Remote photoplethysmography:" + ], + "bbox": [ + 501, + 92, + 893, + 902 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "14472", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Rarely considered factors. In IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 1197-1206, June 2020. 6", + "[27] Ishan Misra and Laurens van der Maaten. Self-Supervised Learning of Pretext-Invariant Representations. CVPR, 2020. 2", + "[28] Andreia Moco, Sander Stuijk, Gerard De Haan, Ruikang K. Wang, and Wim Verkruysse. A Model for Waveform Dissimilarities in Dual-Depth Reflectance-PPG. Proceedings of the Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBS), 2018-July:5125-5130, 2018. ISBN: 9781538636466. 8", + "[29] Xuesong Niu, Hu Han, Shiguang Shan, and Xin Chen. VIPL-HR: A multi-modal database for pulse estimation from less-constrained face video. In Asian Conference on Computer Vision (ACCV), pages 562-576, 2018. 2, 6", + "[30] Xuesong Niu, Shiguang Shan, Hu Han, and Xin Chen. RhythmNet: End-to-end heart rate estimation from face via spatial-temporal representation. IEEE Transactions on Image Processing, 29:2409-2423, 2020. 1, 2", + "[31] Ewa Nowara and Daniel McDuff. Combating the Impact of Video Compression on Non-Contact Vital Sign Measurement Using Supervised Learning. In IEEE International Conference on Computer Vision Workshop (ICCVW), pages 1706–1712, Oct. 2019. ISSN: 2473-9944. 5", + "[32] Ewa M. Nowara, Daniel McDuff, and Ashok Veeraraghavan. Systematic analysis of video-based pulse measurement from compressed videos. Biomedical Optics Express, 12(1):494, 2021. 3, 5", + "[33] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding, 2018. 3", + "[34] M. Poh, D. J. McDuff, and R. W. Picard. Advancements in noncontact, multiparameter physiological measurements using a webcam. IEEE Transactions on Biomedical Engineering, 58(1):7-11, 2011. 2, 6", + "[35] Ming-Zher Poh, Daniel J. McDuff, and Rosalind W. Picard. Non-contact, automated cardiac pulse measurements using video imaging and blind source separation. Opt. Express, 18(10):10762-10774, May 2010. 2", + "[36] Michal Rapczynski, Philipp Werner, and Ayoub Al-Hamadi. Effects of video encoding on camera-based heart rate estimation. IEEE Transactions on Biomedical Engineering, 66(12):3360-3370, 2019. Publisher: IEEE. 5", + "[37] Rencheng Song, Huan Chen, Juan Cheng, Chang Li, Yu Liu, and Xun Chen. Pulsegan: Learning to generate realistic pulse waveforms in remote photoplethysmography. IEEE Journal of Biomedical and Health Informatics, 25(5):1373-1384, 2021. 6", + "[38] Jeremy Speth, Nathan Vance, Adam Czajka, Kevin Bowyer, and Patrick Flynn. Unifying frame rate and temporal dilations for improved remote pulse detection. Computer Vision and Image Understanding (CVIU), pages 1056-1062, 2021. 1, 2, 3, 5, 6, 8", + "[39] Jeremy Speth, Nathan Vance, Adam Czajka, Kevin Bowyer, Diane Wright, and Patrick Flynn. Deception detection and remote physiological monitoring: A dataset and baseline ex" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "perimental results. In International Joint Conference on Biometrics (IJCB), pages 4264-4271, 2021. 5", + "[40] Radim Špetlík, Vojtech Franc, and Jírí Matas. Visual heart rate estimation with convolutional neural network. In Proceedings of the British machine vision conference, Newcastle, UK, pages 3-6, 2018. 6", + "[41] Ronny Stricker, Steffen Muller, and Horst Michael Gross. Non-contact video-based pulse rate measurement on a mobile service robot. IEEE International Symposium on Robot and Human Interactive Communication, pages 1056-1062, 2014. 5", + "[42] Zhaodong Sun and Xiaobai Li. Contrast-phys: Unsupervised video-based remote physiological measurement via spatiotemporal contrast. European Conference on Computer Vision (ECCV), 2022. 1, 2, 3, 6, 7", + "[43] Nathan Vance, Jeremy Speth, Siamul Khan, Adam Czajka, Kevin W. Bowyer, Diane Wright, and Patrick Flynn. Deception detection and remote physiological monitoring: A dataset and baseline experimental results. IEEE Transactions on Biometrics, Behavior, and Identity Science (TBIOM), pages 1-1, 2022. 5", + "[44] Wim Verkruysse, Lars O Svaasand, and J Stuart Nelson. Remote plethysmographic imaging using ambient light. Opt. Express, 16(26):21434-21445, Dec 2008. 6", + "[45] Hao Wang, Euijoon Ahn, and Jinman Kim. Self-Supervised Representation Learning Framework for Remote Physiological Measurement Using Spatiotemporal Augmentation Loss. Proceedings of the AAAI Conference on Artificial Intelligence, 36(2):2431-2439, June 2022. 1, 2, 3, 6", + "[46] Wenjin Wang, Albertus C. Den Brinker, and Gerard De Haan. Single-Element Remote-PPG. IEEE Transactions on Biomedical Engineering, 66(7):2032-2043, 2019. 2", + "[47] W. Wang, A. C. den Brinker, S. Stuijk, and G. de Haan. Algorithmic principles of remote PPG. IEEE Transactions on Biomedical Engineering, 64(7):1479-1491, 2017. 2, 6, 7", + "[48] Wenjin Wang, Sander Stuijk, and Gerard De Haan. A Novel Algorithm for Remote Photoplethysmography: Spatial Subspace Rotation. IEEE Transactions on Biomedical Engineering, 63(9):1974-1984, 2016. 6", + "[49] Zhen Wang, Yunhao Ba, Pradyumna Chari, Oyku Deniz Bozkurt, Gianna Brown, Parth Patwa, Niranjan Vaddi, Laleh Jalilian, and Achuta Kadambi. Synthetic generation of face videos with plethysmograph physiology. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20555-20564, 2022. 2", + "[50] Yuzhe Yang, Xin Liu, Jiang Wu, Silviu Borac, Dina Katabi, Ming-Zher Poh, and Daniel McDuff. Simper: Simple self-supervised learning of periodic targets, 2022. 1, 2, 6", + "[51] Zitong Yu, Xiaobai Li, and Guoying Zhao. Remote photoplethysmograph signal measurement from facial videos using spatio-temporal networks. In Proceedings of the British Machine Vision Conference (BMVC), pages 1-12, 2019. 1, 2, 3, 5, 6", + "[52] Zitong Yu, Yuming Shen, Jingang Shi, Hengshuang Zhao, Philip H.S. Torr, and Guoying Zhao. Physformer: Facial video-based physiological measurement with temporal difference transformer. In Proceedings of the IEEE/CVF" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "14473", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Conference on Computer Vision and Pattern Recognition (CVPR), pages 4186-4196, June 2022. 1, 2, 3", + "[53] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow Twins: Self-Supervised Learning via Redundancy Reduction. 2021. arXiv: 2103.03230. 2", + "[54] Qi Zhan, Wenjin Wang, and Gerard de Haan. Analysis of CNN-based remote-PPG to understand limitations and sensitivities. Biomedical Optics Express, 11(3):1268-1283, 2020. 8", + "[55] Yu Zhao, Bochao Zou, Fan Yang, Lin Lu, Abdelkader Nasreddine Belkacem, and Chao Chen. Video-based physiological measurement using 3d central difference convolution attention network. In 2021 IEEE International Joint Conference on Biometrics (IJCB), pages 1-6, 2021. 2", + "[56] Hao Zhu, Wayne Wu, Wentao Zhu, Liming Jiang, Siwei Tang, Li Zhang, Ziwei Liu, and Chen Change Loy. CelebV-HQ: A large-scale video facial attributes dataset. In ECCV, 2022. 5, 7" + ], + "bbox": [ + 78, + 90, + 468, + 345 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "14474", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Non-Contrastive Unsupervised Learning of Physiological Signals From Video/f2bb9236-2bf0-41f5-a980-2c135b55dca1_model.json b/2023/Non-Contrastive Unsupervised Learning of Physiological Signals From Video/f2bb9236-2bf0-41f5-a980-2c135b55dca1_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7ac5f8d00010c4c39ec665c0e8b874bed2420ce8 --- /dev/null +++ b/2023/Non-Contrastive Unsupervised Learning of Physiological Signals From Video/f2bb9236-2bf0-41f5-a980-2c135b55dca1_model.json @@ -0,0 +1,2312 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.1, + 0.131, + 0.872, + 0.153 + ], + "angle": 0, + "content": "Non-Contrastive Unsupervised Learning of Physiological Signals from Video" + }, + { + "type": "text", + "bbox": [ + 0.253, + 0.181, + 0.717, + 0.217 + ], + "angle": 0, + "content": "Jeremy Speth, Nathan Vance, Patrick Flynn, Adam Czajka \nUniversity of Notre Dame" + }, + { + "type": "text", + "bbox": [ + 0.32, + 0.22, + 0.647, + 0.234 + ], + "angle": 0, + "content": "{jspeth,nvancel,flynn,aczajka}@nd.edu" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.301, + 0.474, + 0.635 + ], + "angle": 0, + "content": "Subtle periodic signals such as blood volume pulse and respiration can be extracted from RGB video, enabling noncontact health monitoring at low cost. Advancements in remote pulse estimation - or remote photoplethysmography (rPPG) - are currently driven by deep learning solutions. However, modern approaches are trained and evaluated on benchmark datasets with ground truth from contact-PPG sensors. We present the first non-contrastive unsupervised learning framework for signal regression to mitigate the need for labelled video data. With minimal assumptions of periodicity and finite bandwidth, our approach discovers the blood volume pulse directly from unlabelled videos. We find that encouraging sparse power spectra within normal physiological bandlimits and variance over batches of power spectra is sufficient for learning visual features of periodic signals. We perform the first experiments utilizing unlabelled video data not specifically created for rPPG to train robust pulse rate estimators. Given the limited inductive biases and impressive empirical results, the approach is theoretically capable of discovering other periodic signals from video, enabling multiple physiological measurements without the need for ground truth signals." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.664, + 0.21, + 0.68 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.47, + 0.87 + ], + "angle": 0, + "content": "Camera-based vitals estimation is a rapidly growing field enabling non-contact health monitoring in a variety of settings [23]. Although many of the signals avoid detection from the human eye, video data in the visible and infrared ranges contain subtle intensity changes caused by physiological oscillations such as blood volume and respiration. Significant remote photoplethysmography (rPPG) research for estimating the cardiac pulse has leveraged supervised deep learning for robust signal extraction [7, 18, 30, 38, 51, 52]. While the number of successful approaches has rapidly increased, the size of benchmark video datasets with simultaneous vitals recordings has remained relatively stagnant." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Robust deep learning-based systems for deployment require training on larger volumes of video data with di" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.27, + 0.895, + 0.421 + ], + "angle": 0, + "content": "verse skin tones, lighting, camera sensors, and movement. However, collecting simultaneous video and physiological ground truth with contact-PPG or electrocardiograms (ECG) is challenging for several reasons. First, many hours of high quality videos is an unwieldy volume of data. Second, recording a diverse subject population in conditions representative of real-world activities is difficult to conduct in the lab setting. Finally, synchronizing contact measurements with video is technically challenging, and even contact measurements used for ground truth contain noise." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.422, + 0.895, + 0.542 + ], + "angle": 0, + "content": "Fortunately, recent works find that contrastive unsupervised learning for rPPG is a promising solution to the data scarcity problem [13, 42, 45, 50]. We extend this research line into non-contrastive unsupervised learning to discover periodic signals in video data. With end-to-end unsupervised learning, collecting more representative training data to learn powerful visual features is much simpler, since only video is required without associated medical information." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.542, + 0.896, + 0.679 + ], + "angle": 0, + "content": "In this work, we show that non-contrastive unsupervised learning is especially simple when regressing rPPG signals. We find weak assumptions of periodicity are sufficient for learning the minuscule visual features corresponding to the blood volume pulse from unlabelled face videos. The loss functions can be computed in the frequency domain over batches without the need for pairwise or triplet comparisons. Figure 1 compares the proposed approach with supervised and contrastive unsupervised learning approaches." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.679, + 0.895, + 0.739 + ], + "angle": 0, + "content": "This work creates opportunities for scaling deep learning models for camera-based vitals and estimating periodic or quasi-periodic signals from unlabelled data beyond rPPG. Our novel contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.744, + 0.892, + 0.79 + ], + "angle": 0, + "content": "1. A general framework for physiological signal estimation via non-contrastive unsupervised learning (SiNC) by leveraging periodic signal priors." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.797, + 0.892, + 0.826 + ], + "angle": 0, + "content": "2. The first non-contrastive unsupervised learning method for camera-based vitals measurement." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.835, + 0.892, + 0.865 + ], + "angle": 0, + "content": "3. The first experiments and results of training with non-rPPG-specific video data without ground truth vitals." + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.744, + 0.892, + 0.865 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.89, + 0.901 + ], + "angle": 0, + "content": "Source code to replicate this work is available at https://github.com/CVRL/SiNC-rPPG." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14464" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.087, + 0.895, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.332, + 0.895, + 0.39 + ], + "angle": 0, + "content": "Figure 1. Overview of the SiNC framework for rPPG compared with traditional supervised and unsupervised learning. Supervised and contrastive losses use distance metrics to the ground truth or other samples. Our framework applies the loss directly to the prediction by shaping the frequency spectrum, and encouraging variance over a batch of inputs. Power outside of the bandlimits is penalized to learn invariances to irrelevant frequencies. Power within the bandlimits is encouraged to be sparsely distributed near the peak frequency." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.413, + 0.22, + 0.428 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.438, + 0.416, + 0.455 + ], + "angle": 0, + "content": "2.1. Remote Photoplethysmography (rPPG)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.461, + 0.47, + 0.612 + ], + "angle": 0, + "content": "Approaches for remote pulse estimation have shifted over the last decade from blind source separation [34, 35], through linear color transformations [9, 10, 46, 47] to training supervised deep learning-based models [7, 16, 18, 20, 29, 30, 38, 51, 52, 55]. While the color transformations generalize well across datasets, deep learning-based models give better accuracy when tested on data from a similar distribution to the training set. To this end, deep learning research has focused on optimizing architectures for extracting robust spatial and temporal features from the limited datasets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.612, + 0.471, + 0.793 + ], + "angle": 0, + "content": "To get around the data bottleneck, large synthetic datasets have recently been proposed [24, 49]. The SCAMPS dataset [24] contains 2,800 videos of synthetic avatars with corresponding PPG, EKG, respiration, and facial action units. The UCLA-synthetic dataset [49] contains 480 videos, which are shown to improve performance when trained with real data. A strength of synthetic datasets is their ability to cover the broad range of skin tones compared to physical data. Another solution to the lack of physiological data is unsupervised learning, where a large set of videos and periodic priors on the output signal is sufficient [13,42,45,50]. We discuss these methods in Sec. 2.3." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.802, + 0.295, + 0.819 + ], + "angle": 0, + "content": "2.2. Unsupervised Learning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Self-supervised learning is progressing for image representation learning, with two main competing classes of approaches: contrastive and non-contrastive (or regularized) learning [3]. Contrastive approaches [5, 6, 27] define criteria for distinguishing whether two or more samples are" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.414, + 0.895, + 0.537 + ], + "angle": 0, + "content": "the same or different. Non-contrastive methods augment positive pairs, and enforce variance in the predictions over batches to avoid collapse, in which the model's embeddings reside in a small subspace of the feature space [3]. Distillation methods only use positive samples and avoid collapse by applying a moving average and stop-gradient operator [8, 14]. Another approaches maximize information content of embeddings [11, 53]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.545, + 0.792, + 0.562 + ], + "angle": 0, + "content": "2.3. Unsupervised Learning for rPPG" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.569, + 0.895, + 0.903 + ], + "angle": 0, + "content": "All existing unsupervised rPPG approaches are contrastive [13,42,45,50], in which the same model is fed pairs of input videos, and the predictions over similar videos are pulled closer, while the predictions from dissimilar videos are repelled. Gideon et al. [13] were the first to train a deep learning model without labels for rPPG using the contrastive framework. The core of their approach is frequency resampling to create negative samples. Although spatially similar, a resampled video clip contains a different underlying pulse rate, so the model must learn to attend to the temporal dynamics. For their distance function between pairs, they calculated the mean square error between power spectral densities of the model's waveform predictions. While their approach learns to estimate the pulse, their formulation with negative samples is imprecise. The resampled frequency for the negative sample is known, so the relative translation of the power spectrum from the anchor sample can be directly computed. Thus, rather than repelling the estimated spectra, it is more accurate to penalize differences from the known spectra. Furthermore, resampling close to the original sampling rate causes overlap in the power spectra, so repelling the pair is inaccurate. Yuzhe et al. [50]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14465" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.182 + ], + "angle": 0, + "content": "incorporated the previously ignored resampling factor for a soft modification to the InfoNCE loss [33] that scales the desired similarity between pairs by their relative sampling rate. A downside is that their learning framework is not end-to-end unsupervised and requires fine-tuning with PPG labels after the self-supervised stage." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.184, + 0.471, + 0.381 + ], + "angle": 0, + "content": "Differently from [13], Contrast-Phys [42] and SLF-RPM [45] consider all samples different from the anchor to be negatives. This assumes that the power spectra will vary between subjects or sufficiently long windows for the same subject. This runs into similar issues with negative pairs as Gideon's approach. Different subjects may have the same pulse rate, so punishing the model for predicting similar frequencies is common during training. Furthermore, the Fast Fourier Transform (FFT) does not produce perfectly sparse decompositions, resulting in spectral overlap even if the heart rate differs by several beats per minute (bpm). As an example, the last column of the second row in Fig. 2 shows the nulls of the main lobe are nearly 30 bpm apart." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.397, + 0.168, + 0.412 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.424, + 0.469, + 0.59 + ], + "angle": 0, + "content": "We first formulate signal regression from video. A video sample \\( x_{i} \\in \\mathbb{R}^{T \\times W \\times H \\times C} \\) sampled from a dataset \\( \\mathcal{D} \\) consists of \\( T \\) images of size \\( W \\times H \\) pixels across \\( C \\) channels, captured over time. State-of-the-art methods offer models \\( f \\) that regress a waveform \\( \\mathbb{R}^T \\ni y_i = f(x_i) \\) of the same length as the video. Recently, the task has been effectively modeled end-to-end with the models \\( f \\) being spatiotemporal neural networks [16, 21, 38, 51, 52]. While most previous works are supervised and minimize the loss to a contact pulse measurement, we perform non-contrastive learning using only the model's estimated waveform." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.592, + 0.469, + 0.758 + ], + "angle": 0, + "content": "The key realization is that we can place strong priors on the estimated pulse regarding its bandwidth and periodicity. Observed signals outside the desired frequency range are pollutants, so penalizing the model for carrying them through the forward pass results in invariances to such noisy visual features. We find that it is surprisingly easy to impose the desired constraints in the frequency domain. Thus, all waveforms are transformed into their discrete Fourier components with the FFT before computing all losses in our approach, \\( F = \\mathrm{FFT}(y) \\). The following sections describe the loss functions and augmentations used during training." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.771, + 0.165, + 0.785 + ], + "angle": 0, + "content": "3.1. Losses" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.469, + 0.901 + ], + "angle": 0, + "content": "One of the advantages of unsupervised learning for periodic signals is that we can constrain the solution space significantly. For physiological signals such as respiration and blood volume pulse, we know the healthy upper and lower bounds of the frequencies. We also desire the extracted signal to be sparse in the frequency domain, and that our model filters out noise signals present in the video. With these" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.891, + 0.121 + ], + "angle": 0, + "content": "constraints, we can greatly simplify the problem of finding good features for the desired signal in the data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.122, + 0.892, + 0.243 + ], + "angle": 0, + "content": "Bandwidth Loss. One of the most powerful constraints we can place on the model is frequency bandlimits. Past unsupervised methods have used the irrelevant power ratio (IPR) as a validation metric [12, 13, 42] for model selection. We find that it is also effective during model training. The IPR penalizes the model for generating signals outside the desired bandlimits. With lower and upper bandlimits of \\(a\\) and \\(b\\), respectively, our bandwidth loss becomes:" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.253, + 0.892, + 0.307 + ], + "angle": 0, + "content": "\\[\nL _ {b} = \\frac {1}{\\sum_ {i = - \\infty} ^ {\\infty} F _ {i}} \\left[ \\sum_ {i = - \\infty} ^ {a} F _ {i} + \\sum_ {i = b} ^ {\\infty} F _ {i} \\right], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.313, + 0.892, + 0.463 + ], + "angle": 0, + "content": "where \\( F_{i} \\) is the power in the \\( i \\)th frequency bin of the predicted signal. This simple loss enforces learning of many invariants, such as movement from respiration, talking, or facial expressions which typically occupy low frequencies. In our experiments we specify the limits as \\( a = 0.66\\mathrm{Hz} \\) to \\( b = 3\\mathrm{Hz} \\), which corresponds to a common pulse rate range from 40 bpm to 180 bpm. The first column of Fig. 2 shows the result of training exclusively with the bandwidth loss \\( L_{b} \\). The last row shows that the model concentrates signal power between the bandlimits." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.464, + 0.892, + 0.555 + ], + "angle": 0, + "content": "Sparsity Loss. The pulse rate is the most common physiological marker associated with the blood volume pulse. Since we are primarily interested in the frequency, we can further improve our model by preventing wideband predictions. This also reveals the true signal we aim to discover by ignoring visual dynamics that are not strongly periodic." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.555, + 0.892, + 0.586 + ], + "angle": 0, + "content": "We penalize energy within the bandlimits that is not near the spectral peak:" + }, + { + "type": "equation", + "bbox": [ + 0.555, + 0.596, + 0.891, + 0.653 + ], + "angle": 0, + "content": "\\[\nL _ {s} = \\frac {1}{\\sum_ {i = a} ^ {b} F _ {i}} \\left[ \\sum_ {i = a} ^ {F ^ {*} - \\Delta_ {F}} F _ {i} + \\sum_ {i = F ^ {*} + \\Delta_ {F}} ^ {b} F _ {i} \\right], \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.892, + 0.765 + ], + "angle": 0, + "content": "where \\( F^{*} = \\operatorname{argmax}(F) \\) and \\( \\Delta_F \\) are the frequencies of the spectral peak and padding around the peak, respectively. For all experiments \\( \\Delta_F = 6 \\) beats per minute [32]. Figure 2 shows the result of training only with the sparsity loss in the second column. For the whole dataset, the power spectrum is sparsely distributed in the low frequencies, effectively filtering frequencies higher than \\( 1\\mathrm{Hz} \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.766, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Variance Loss. One of the risks of non-contrastive methods is the model collapsing into trivial solutions and making predictions independently of the input features. In regularized methods such as VICReg [3], a hinge loss on the variance over a batch of predictions is used to enforce diverse outputs. We use a similar strategy to avoid model collapse, but instead spread the variance in power spectral densities towards a uniform distribution over the desired frequency band. Our variance loss processes a uniform prior" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14466" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.083, + 0.088, + 0.506, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.089, + 0.697, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.089, + 0.891, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.234, + 0.505, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.234, + 0.697, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.234, + 0.89, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.367, + 0.505, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.368, + 0.697, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.368, + 0.89, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.391, + 0.501, + 0.612, + 0.517 + ], + "angle": 0, + "content": "Predicted Ground Truth" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.531, + 0.893, + 0.616 + ], + "angle": 0, + "content": "Figure 2. Each column shows predictions from models trained with one or all of the losses for 20 epochs on UBFC-rPPG. The first two rows show a sample in the time and frequency domain, respectively. The last row shows the signal power over the validation set computed by taking the sum of normalized power spectral densities from each sample, then dividing the result by the number of validation samples. The bandwidth loss penalizes signal power outside predefined bandlimits (40 to 180 bpm) to constrain the output space. The sparsity loss encourages a narrow spectrum containing strong periodicity. The variance loss encourages diverse power spectra in a batch, preventing the model from collapsing to a narrow bandwidth. When combined, the model estimates periodic signals within the desired bandlimits." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.642, + 0.47, + 0.734 + ], + "angle": 0, + "content": "distribution \\(P\\) over \\(d\\) frequencies, and a batch of \\(n\\) spectral densities, \\(F = [v_{1},\\dots,v_{n}]\\), where each vector is a \\(d\\)-dimensional frequency decomposition of a predicted waveform. We calculate the normalized sum of densities over the batch, \\(Q\\), and define the variance loss as the squared Wasserstein distance [15] to the uniform prior:" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.746, + 0.469, + 0.788 + ], + "angle": 0, + "content": "\\[\nL _ {v} = \\frac {1}{d} \\sum_ {i = 1} ^ {d} \\left(\\mathrm {C D F} _ {i} (Q) - \\mathrm {C D F} _ {i} (P)\\right) ^ {2}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.47, + 0.902 + ], + "angle": 0, + "content": "where CDF is a cumulative distribution function. The third column of Fig. 2 shows the effect of the variance loss during training. For a single sample, wide-band signals containing multiple frequencies are predicted, and the predicted frequencies cover the task's bandwidth. In our experiments we use a batch size of 20 samples. See the suppl. materials for an ablation experiment on the impact of batch size." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.642, + 0.89, + 0.673 + ], + "angle": 0, + "content": "Combining All Losses. Summarizing, our training loss function is a simple sum of the aforementioned losses:" + }, + { + "type": "equation", + "bbox": [ + 0.628, + 0.704, + 0.891, + 0.719 + ], + "angle": 0, + "content": "\\[\nL = L _ {b} + L _ {s} + L _ {v}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.893, + 0.902 + ], + "angle": 0, + "content": "While one could weight particular components of the loss more than others, we specifically formulated the losses to scale them between 0 and 1. In our experiments, we find that a simple summation without weighting gives good performance. The combined loss function encourages the model to search over the supported frequencies to discover visual features for a strong periodic signal. Remarkably, we find that this simple framework is sufficient for learning to regress the blood volume in video, as shown in the last column of Fig. 2." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14467" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.091, + 0.232, + 0.108 + ], + "angle": 0, + "content": "3.2. Augmentations" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.113, + 0.47, + 0.205 + ], + "angle": 0, + "content": "Unlike Gideon et al.'s [13] approach, which only applies frequency augmentations, we apply several augmentations to both the spatial and temporal dimensions to learn invariances to noisy visual signals. In fact, we found that without augmentations, models did not converge during training (see the supplementary material)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.206, + 0.469, + 0.28 + ], + "angle": 0, + "content": "Image Intensity Augmentations. Gaussian noise is added to each pixel with zero mean and a standard deviation of 2 on an image scale from 0 to 255. Each clip is darkened or brightened by adding a constant from a Gaussian distribution with mean 0 and standard deviation of 10." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.28, + 0.469, + 0.356 + ], + "angle": 0, + "content": "Spatial Augmentations. We randomly horizontally flip a video clip with \\(50\\%\\) probability. The spatial dimension of a clip are randomly square cropped down to between half the original length and the original length. The cropped clip is then linearly interpolated back to the original dimensions." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.356, + 0.469, + 0.446 + ], + "angle": 0, + "content": "Temporal Augmentations. With the general assumption that the desired signal is strongly periodic and sparsely represented in the Fourier domain, we randomly flip a video clip along the time dimension with a probability of \\(50\\%\\). Note that the Fourier decomposition of a time-reversed sinusoid is identical to that of the original sinusoid." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.447, + 0.47, + 0.598 + ], + "angle": 0, + "content": "Frequency Augmentations. Perhaps the most important augmentation is frequency resampling [13], where the video is linearly interpolated to a different frame rate. This augmentation is particularly interesting for rPPG, because it transforms the video input and target signal equivalently along the time dimension, making it equivariant. Given the aforementioned transformations that are invariant, \\(\\tau (\\cdot)\\sim\\) \\(\\mathcal{T}\\) , the equivariant frequency resampling operation, \\(\\phi (\\cdot)\\sim\\) \\(\\Phi\\) , and a model \\(f(\\cdot)\\) that infers a waveform from a video we have the following:" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.609, + 0.469, + 0.627 + ], + "angle": 0, + "content": "\\[\n\\phi (f (\\tau (x))) = f (\\phi (\\tau (x))). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.636, + 0.469, + 0.743 + ], + "angle": 0, + "content": "This is a powerful augmentation, because it allows us to augment the target distribution along with the video input. In our experiments we randomly resample input clips by a factor \\( c \\sim U(0.6, 1.4) \\). After applying the resampling augmentation, we scale the bandlimits by \\( c \\), to avoid penalizing the model if the augmentation pushed the underlying pulse frequency outside of the original bandlimits." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.755, + 0.173, + 0.77 + ], + "angle": 0, + "content": "4. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.78, + 0.469, + 0.841 + ], + "angle": 0, + "content": "We use PURE [41], UBFC-rPPG [4], and DDPM [39] as benchmark rPPG datasets for training and testing, and CelebV-HQ dataset [56] and HKBU-MARs [17] for unsupervised training only." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.841, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Deception Detection and Physiological Monitoring (DDPM) [39, 43] consists of data from 86 subjects attempting to answer questions deceptively. Interviews were recorded at 90 frames-per-second for more than 10 minutes" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.121 + ], + "angle": 0, + "content": "on average. Natural conversation and head pose changes make it a difficult and less-constrained rPPG dataset." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.122, + 0.892, + 0.228 + ], + "angle": 0, + "content": "PURE [41] is a benchmark rPPG dataset consisting of 10 subjects recorded over 6 sessions. Each session lasted approximately 1 minute, and raw video was recorded at 30 fps. The 6 sessions for each subject consisted of: (1) steady, (2) talking, (3) slow head translation, (4) fast head translation, (5) small and (6) medium head rotations. Pulse rates are at or close to the subject's resting rate." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.228, + 0.892, + 0.289 + ], + "angle": 0, + "content": "UBFC-rPPG [4] contains 1-minute long videos from 42 subjects recorded at 30 fps. Subjects played a time-sensitive mathematical game to raise their heart rates, but head motion is limited during the recording." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.289, + 0.892, + 0.41 + ], + "angle": 0, + "content": "HKBU 3D Mask Attack with Real World Variations (HKBU-MARs) [17] consists of 12 subjects captured over 6 different lighting configurations with 7 different cameras each, resulting in 504 videos lasting 10 seconds each. The diverse lighting and camera sensors make it a valuable dataset for unsupervised training. We use version 2 of HKBU-MARs, which contains videos with both realistic 3D masks and unmasked subjects." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.41, + 0.892, + 0.547 + ], + "angle": 0, + "content": "High-Quality Celebrity Video Dataset (CelebV-HQ) [56] is a set of processed YouTube videos containing 35,666 face videos from over 15,000 identities. The videos vary dramatically in length, lighting, emotion, motion, skin tones, and camera sensors. The greatest challenge in harnessing online videos is their reduced quality due to compression before upload and by the video provider. Compression is a known challenge for rPPG, since the blood volume pulse is so subtle optically [25, 31, 32, 36]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.561, + 0.662, + 0.579 + ], + "angle": 0, + "content": "5. Training Details" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.587, + 0.687, + 0.603 + ], + "angle": 0, + "content": "5.1. Data Preprocessing" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.61, + 0.892, + 0.806 + ], + "angle": 0, + "content": "To prepare the video clips for the spatiotemporal deep learning models, we first extract 68 face landmarks with OpenFace [2]. We then define a bounding box in each frame with the minimum and maximum \\((x,y)\\) locations by extending the crop horizontally by \\(5\\%\\) to ensure that the cheeks and jaw are present. The top and bottom are extended by \\(30\\%\\) and \\(5\\%\\) of the bounding box height, respectively, to include the forehead and jaw. We further extend the shorter of the two axes to the length of the other to form a square. The cropped frames are then resized to \\(64\\times 64\\) pixels with bicubic interpolation. For faster processing of the massive CelebV-HQ [56] dataset, we instead use MediaPipe Face Mesh [22] for landmarking." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.817, + 0.694, + 0.832 + ], + "angle": 0, + "content": "5.2. Model Architectures" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We use a 3D-CNN architecture similar to [38] without temporal dilations [51]. We use a temporal kernel width of 5, and replace default zero-padding by repeating the edges. Zero-padding along the time dimension can result" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14468" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.895, + 0.133 + ], + "angle": 0, + "content": "Table 1. Intra-dataset pulse rate estimation results. The best results for a given metric/dataset are bolded, and the second-best results are underlined. For a better comparison with results in the literature, this table follows [42]. MAE: Mean Absolute Error; RMSE: Root Mean Square Error; \\( r \\) : Pearson correlation coefficient." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.142, + 0.895, + 0.395 + ], + "angle": 0, + "content": "
TypesMethodsUBFC-rPPGPUREDDPM
MAE (bpm)RMSE (bpm)rMAE (bpm)RMSE (bpm)rMAE (bpm)RMSE (bpm)r
TraditionalGREEN [44]7.5014.410.627.2317.050.6932.7943.090.04
ICA [34]5.1711.760.653.7612.600.8522.2235.770.40
CHROM [9]2.369.230.870.752.231.0013.4828.530.56
2SR [48]6.9018.500.652.443.060.9822.0839.730.22
POS [47]2.119.110.870.804.110.989.0323.070.70
SupervisedHR-CNN [40]---1.842.370.98---
SynRhythm [29]5.596.820.72------
PulseGAN [37]1.192.100.98------
Dual-GAN [21]0.440.670.990.821.310.99---
RPNet [38]†0.53 ± 0.011.78 ± 0.020.99 ± 0.001.15 ± 0.275.77 ± 1.250.96 ± 0.023.46 ± 0.2412.47 ± 0.680.91 ± 0.01
PhysNet [51]†0.55 ± 0.032.03 ± 0.370.99 ± 0.000.99 ± 0.195.22 ± 0.930.97 ± 0.013.96 ± 0.7613.57 ± 1.740.89 ± 0.03
UnsupervisedGideon2021 [13]1.854.280.932.32.90.99---
SLF-RPM [45]*8.399.700.70------
SimPer [50]*4.24--3.89-----
Contrast-Phys [42]0.641.000.991.001.400.999.70 ± 2.9025.02 ± 6.010.58 ± 0.19
SiNC (ours)0.59 ± 0.001.83 ± 0.040.99 ± 0.000.61 ± 0.061.84 ± 0.401.00 ± 0.005.87 ± 0.1117.44 ± 0.160.81 ± 0.00
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.086, + 0.396, + 0.636, + 0.408 + ], + "angle": 0, + "content": "* Some methods in the unsupervised row require fine-tuning on labeled data with a linear classifier." + }, + { + "type": "table_footnote", + "bbox": [ + 0.086, + 0.407, + 0.636, + 0.421 + ], + "angle": 0, + "content": "† Some supervised methods were trained with identical data augmentations to SiNC for fair comparison." + }, + { + "type": "list", + "bbox": [ + 0.086, + 0.396, + 0.636, + 0.421 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.447, + 0.473, + 0.658 + ], + "angle": 0, + "content": "in edge effects that add artificial frequencies to the predictions. Early experiments showed that temporal dilations caused aliasing and reduced the bandwidth of the model to specific frequencies. Our losses and framework may be applied to any task and architecture with dense predictions along one or more dimensions. However, popular rPPG architectures such as DeepPhys [7] and MTTS-CAN [18] may be ill-suited for the approach, since they consume very few frames, and the number of time points should be large enough to give sufficient frequency resolution with the FFT. In our experiments, we use the AdamW [19] optimizer with a learning rate of 0.0001. We use a clip length of \\( T = 120 \\) frames (4 seconds), and we set the input signal's length to achieve a frequency resolution of 0.33 bpm." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.666, + 0.273, + 0.684 + ], + "angle": 0, + "content": "5.3. Supervised Training" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.471, + 0.826 + ], + "angle": 0, + "content": "To properly compare our approach to its supervised counterpart we use the same model architecture and train it with the commonly used negative Pearson loss between the predicted waveform and the contact sensor ground truth [51]. During training we apply all of the same augmentations except time reversal. Models are trained for 200 epochs on PURE and UBFC-rPPG, and for 40 epochs on DDPM. The model from the epoch with the lowest loss on the validation set is selected for testing." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.833, + 0.292, + 0.85 + ], + "angle": 0, + "content": "5.4. Unsupervised Training" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Unsupervised models are trained for the same number of epochs as the supervised setting for both PURE and UBFCrPPG, but we train for an additional 40 epochs on DDPM," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.447, + 0.892, + 0.538 + ], + "angle": 0, + "content": "since this dataset is considerably more difficult. We set the batch size to 20 samples during training. Contrary to previous unsupervised approaches [13, 42], we leverage validation sets for model selection by selecting the model with the lowest combined bandwidth and sparsity losses. The creation of the dataset splits is described in the next section." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.545, + 0.622, + 0.56 + ], + "angle": 0, + "content": "5.5. Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.569, + 0.892, + 0.674 + ], + "angle": 0, + "content": "Pulse rates are computed as the highest spectral peak between \\(0.66\\mathrm{Hz}\\) and \\(3\\mathrm{Hz}\\) (equivalent to \\(40\\mathrm{bpm}\\) to \\(180\\mathrm{bpm}\\)) over a 10-second sliding window. The same procedure is applied to the ground truth waveforms for a reliable evaluation [26]. We apply common error metrics such as mean absolute error (MAE), root mean square error (RMSE), and Pearson correlation coefficient \\((r)\\) between the pulse rates." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.892, + 0.81 + ], + "angle": 0, + "content": "We perform 5-fold cross validation for both PURE and UBFC with the same folds as [13], and use the predefined dataset splits from DDPM [38]. Differently from [13], we use 3 of the folds for training, 1 for validation, and the remaining for testing rather than only training and testing partitions. We train 3 models with different initializations, resulting in 15 models trained on PURE and UBFC each, and 3 models trained on DDPM. We present the mean and standard deviation of the errors in the results." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.823, + 0.587, + 0.838 + ], + "angle": 0, + "content": "6. Results" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.848, + 0.715, + 0.864 + ], + "angle": 0, + "content": "6.1. Within-Dataset Testing" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Table 1 shows the results for models trained and tested on subject-disjoint partitions from the same datasets. For" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14469" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.468, + 0.198 + ], + "angle": 0, + "content": "PURE and UBFC we achieve MAE lower than 1 bpm, performing better or on par with all traditional and supervised learning approaches. For PURE, our approach gives the lowest MAE and a Pearson \\( r \\) of nearly 1. Performance drops on DDPM due to the overall difficulty of the dataset. SiNC outperforms contrastive approaches, only being surpassed by supervised deep learning models." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.199, + 0.468, + 0.304 + ], + "angle": 0, + "content": "In comparison to other unsupervised methods, Contrast-Phys [42] gives the most competitive performance on all but DDPM. Note that our approach gives the lowest MAE on all datasets, but has higher RMSE. We believe this is due their use of harmonic removal as a post-processing step when estimating the pulse rate, which is not described in [42], but can be found in their publicly available code." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.313, + 0.282, + 0.329 + ], + "angle": 0, + "content": "6.2. Cross-Dataset Testing" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.35, + 0.468, + 0.432 + ], + "angle": 0, + "content": "Table 2. Cross-dataset pulse rate estimation performance. The top 3 training datasets are common rPPG benchmarks, while HKBU was not designed for rPPG and has no pulse ground truth. Note that testing on UBFC after training on DDPM performs well, since their frequency distributions are similar, while PURE's pulse rates tend to be much lower." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.445, + 0.477, + 0.764 + ], + "angle": 0, + "content": "
Training DatasetTesting DatasetMethodMAE (bpm)r
DDPMUBFCSiNC0.88 ± 0.250.98 ± 0.01
UBFCContrast-Phys1.14 ± 0.380.96 ± 0.03
UBFCPhysNet1.11 ± 0.420.95 ± 0.02
PURESiNC3.12 ± 1.070.88 ± 0.05
PUREContrast-Phys13.02 ± 6.120.19 ± 0.59
PUREPhysNet1.46 ± 0.340.95 ± 0.02
UBFCDDPMSiNC18.53 ± 0.360.38 ± 0.01
DDPMContrast-Phys22.93 ± 1.020.18 ± 0.04
DDPMPhysNet18.58 ± 0.120.40 ± 0.00
PURESiNC4.02 ± 0.060.86 ± 0.00
PUREContrast-Phys19.61 ± 2.010.33 ± 0.06
PUREPhysNet3.81 ± 0.340.87 ± 0.02
PUREUBFCSiNC6.64 ± 1.760.59 ± 0.10
UBFCContrast-Phys10.22 ± 0.380.45 ± 0.04
UBFCPhysNet7.02 ± 3.350.60 ± 0.13
DDPMSiNC24.92 ± 0.650.20 ± 0.00
DDPMContrast-Phys29.63 ± 0.480.03 ± 0.02
DDPMPhysNet28.03 ± 2.200.13 ± 0.05
HKBU (non-rPPG)UBFCSiNC1.08 ± 0.030.95 ± 0.00
PURESiNC2.43 ± 0.200.90 ± 0.02
DDPMSiNC20.34 ± 0.250.19 ± 0.02
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.468, + 0.901 + ], + "angle": 0, + "content": "We perform cross-dataset testing to analyze whether the approach is robust to changes to the lighting, camera sensor, pulse rate distribution, and motion. Table 2 shows the results for SiNC and supervised training with the same architecture. We find that the performance is similar for the supervised and unsupervised approaches when transferring to different data sources. Training on PURE exclusively gives relatively poor results when transferring to" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "UBFC-rPPG and DDPM, due to the low pulse rate variability within PURE samples and lack of movement. Training on DDPM gives the best results overall, since the dataset is the largest and captures larger subjects' movements compared to other datasets." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.179, + 0.791, + 0.195 + ], + "angle": 0, + "content": "6.3. Training with CelebV-HQ Videos" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.202, + 0.892, + 0.399 + ], + "angle": 0, + "content": "Given the abundance of face videos publicly available online, we trained a model on the CelebV-HQ dataset [56]. After processing the available videos with MediaPipe and resampling to 30 fps, our unlabeled dataset consisted of 34,029 videos. We trained the model for 23 epochs and manually stopped training due to a plateau in the validation loss. Unfortunately, we found that the model could not converge to the true blood volume pulse. We attribute the failure to poor video quality from compression. Although the videos were downloaded with the highest available quality, they have likely been compressed, removing the pulse signal entirely. See the supplementary material for an assessment of rPPG quality using POS [47] on CelebV-HQ." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.411, + 0.809, + 0.426 + ], + "angle": 0, + "content": "6.4. Training with HKBU-MARs Videos" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.435, + 0.892, + 0.601 + ], + "angle": 0, + "content": "The HKBU-MARs dataset [17] was designed for face presentation attack detection, but we trained models on the \"real\" video sessions in the dataset. The bottom rows in Table 2 show the results for training on HKBU-MARs, then testing on the benchmark rPPG datasets. Training on HKBU-MARs gives better results when testing on UBFC-rPPG and PURE than all training sets except DDPM, which is an order of magnitude larger. To our knowledge, this is the first successful experiment showing that non-rPPG videos can be used to train robust rPPG models, even if they do not have ground-truth pulse labels." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.613, + 0.73, + 0.629 + ], + "angle": 0, + "content": "6.5. Ablation Study on Losses" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.651, + 0.89, + 0.678 + ], + "angle": 0, + "content": "Table 3. Ablation study on the loss functions used during training. Results are shown for models trained and tested on UBFC-rPPG." + }, + { + "type": "table", + "bbox": [ + 0.507, + 0.69, + 0.882, + 0.809 + ], + "angle": 0, + "content": "
LossMAE (bpm)RMSE (bpm)r
Lb3.08 ± 1.698.08 ± 3.610.87 ±0.08
Ls45.50 ± 1.2250.04 ± 0.94-0.04 ±0.08
Lv22.89 ± 2.8331.51 ± 2.360.22 ±0.09
Ls + Lv51.24 ± 5.3657.80 ± 7.39-0.04 ±0.09
Lb + Ls9.99 ± 2.5517.14 ± 2.360.51 ±0.14
Lb + Lv4.18 ± 2.888.90 ± 5.240.82 ±0.14
Lb + Ls + Lv0.59 ± 0.001.83 ± 0.040.99 ± 0.00
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We trained models using all combinations of loss components to analyze their contributions. Table 3 shows the results for training and testing on UBFC-rPPG. The bandwidth loss is the most critical for discovering the true blood volume pulse, while the sparsity and variance losses do not" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14470" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.09, + 0.468, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.176, + 0.468, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.261, + 0.468, + 0.379 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.4, + 0.471, + 0.471 + ], + "angle": 0, + "content": "Figure 3. Within-dataset waveform predictions on all baseline datasets from end-to-end unsupervised models over an 8-second window. The model predictions are remarkably periodic without any form of filtering. Note that phase is not considered during training, so each model learns its own phase shift." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.5, + 0.47, + 0.576 + ], + "angle": 0, + "content": "learn the desired signal by themselves. Surprisingly, combining the bandwidth loss with just one of the sparsity or variance losses gives worse performance than just the bandwidth loss. However, when combining all three components, the model achieves impressive results." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.592, + 0.192, + 0.608 + ], + "angle": 0, + "content": "7. Discussion" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.619, + 0.429, + 0.635 + ], + "angle": 0, + "content": "7.1. Improvements over Supervised Learning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.643, + 0.47, + 0.854 + ], + "angle": 0, + "content": "It is initially surprising that unsupervised training leads to similar or improved rPPG estimation models compared to those trained in a supervised manner. However, there are several potential benefits to unsupervised training. From a hardware perspective, one of the difficulties in supervised training is aligning the contact pulse waveform with the video frames [54]. The pulse sensor and camera may have a time lag, effectively giving the model an out-of-phase target at training time. Unsupervised training gives the model freedom to learn the phase directly from the video. The contact-PPG signal is also sensitive to motion and may be noisy. Since motion may co-occur at the face and fingertip, the contact signal may misguide the model towards visual features for which they should be invariant." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.902 + ], + "angle": 0, + "content": "From a physiological perspective, the pulse observed optically at the fingertip with a contact sensor has a different phase than that of the face, since blood propagates along a" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.242 + ], + "angle": 0, + "content": "different path before reaching the peripheral microvasculature, making alignment nearly impossible without shifting the targets to rPPG estimates from existing methods [38]. Additionally, the morphological shape of the contact-PPG waveform depends on numerous factors such as the wavelength of light (and corresponding tissue penetration depth), external pressure from the oximeter clip, and vasodilation at the measurement site [1,28]. This indicates that the morphology and phase of the target PPG waveform is likely different from the observed rPPG waveform." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.25, + 0.691, + 0.266 + ], + "angle": 0, + "content": "7.2. Why Does It Work?" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.273, + 0.892, + 0.394 + ], + "angle": 0, + "content": "The success of SiNC depends on specific properties of the data, model, and how the two interact. Limited model capacity is actually required, since it forces learning features that generalize across inputs. By constraining the model's predictions to have specific periodic properties, the limited-capacity model must find a general set of features to produce a signal that exists in most or all of the training samples, which is the blood volume pulse in our datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.395, + 0.892, + 0.53 + ], + "angle": 0, + "content": "As a beneficial side-effect, the model intrinsically learns to ignore common noise factors such as illumination, rigid motion, non-rigid motion (e.g. talking, smiling, etc.), and sensor noise, since they may preside outside the predefined bandlimits or with uniform power spectra. Even if noise exhibits periodic tendencies within the bandlimits for some samples, those features would produce poor signals on other samples. Therefore, end-to-end unsupervised approaches like SiNC are well-suited for periodic problems." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.542, + 0.627, + 0.557 + ], + "angle": 0, + "content": "8. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.567, + 0.892, + 0.716 + ], + "angle": 0, + "content": "We proposed a novel non-contrastive learning approach for end-to-end unsupervised signal regression, with specific experiments on blood volume pulse estimation from face videos. This SiNC framework effectively learns powerful visual features with only loose frequency constraints. We demonstrated this by training accurate rPPG models using non-rPPG data and our simple loss functions. Given the subtlety of the rPPG signal, we believe our work can be extended to other signal regression tasks in the domain of remote vitals estimation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgements. This research was sponsored by the Securiport Global Innovation Cell, a division of Securiport LLC. Commercial equipment is identified in this work in order to adequately specify or describe the subject matter. In no case does such identification imply recommendation or endorsement by Securiport LLC, nor does it imply that the equipment identified is necessarily the best available for this purpose. The opinions, findings, and conclusions or recommendations expressed in this publication are those of the authors and do not necessarily reflect the views of our sponsors." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "14471" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.185 + ], + "angle": 0, + "content": "[1] Pierre Abraham, Mélissa Bourgeau, Maïte Camo, Anne Humeau-Heurtier, Sylvain Durand, Pascal Rousseau, and Guillaume Mahe. Effect of skin temperature on skin endothelial function assessment. Microvascular Research, 88:56-60, 2013. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.187, + 0.472, + 0.243 + ], + "angle": 0, + "content": "[2] T. Baltrusaitis, A. Zadeh, Y. C. Lim, and L. Morency. Openface 2.0: Facial behavior analysis toolkit. In IEEE International Conference on Automatic Face Gesture Recognition (FG), pages 59-66, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.245, + 0.471, + 0.286 + ], + "angle": 0, + "content": "[3] Adrien Bardes, Jean Ponce, and Yann LeCun. Vicreg: Variance-invariance-covariance regularization for self-supervised learning. In ICLR, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.288, + 0.471, + 0.343 + ], + "angle": 0, + "content": "[4] Serge Bobbia, Richard Macwan, Yannick Benezeth, Alamin Mansouri, and Julien Dubois. Unsupervised skin tissue segmentation for remote photoplethysmography. Pattern Recognition Letters, 124:82-90, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.345, + 0.471, + 0.415 + ], + "angle": 0, + "content": "[5] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in Neural Information Processing Systems, 2020-Decem(NeurIPS):1-13, 2020. arXiv: 2006.09882. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.417, + 0.471, + 0.485 + ], + "angle": 0, + "content": "[6] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. 37th International Conference on Machine Learning, ICML 2020, pages 1575-1585, 2020. arXiv: 2002.05709 ISBN: 9781713821120. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.488, + 0.471, + 0.543 + ], + "angle": 0, + "content": "[7] Weixuan Chen and Daniel McDuff. DeepPhys: Video-based physiological measurement using convolutional attention networks. In European Conference on Computer Vision (ECCV), pages 356-373, 2018. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.545, + 0.471, + 0.613 + ], + "angle": 0, + "content": "[8] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 15745–15753, 2021. arXiv: 2011.10566 ISBN: 9781665445092. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.616, + 0.471, + 0.657 + ], + "angle": 0, + "content": "[9] G. De Haan and V. Jeanne. Robust pulse rate from chrominance-based rppg. IEEE Transactions on Biomedical Engineering, 60(10):2878-2886, 2013. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.66, + 0.47, + 0.714 + ], + "angle": 0, + "content": "[10] Gerard De Haan and Arno Van Leest. Improved motion robustness of remote-PPG by using the blood volume pulse signature. Physiological Measurement, 35(9):1913–1926, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.717, + 0.47, + 0.772 + ], + "angle": 0, + "content": "[11] Aleksandr Ermolov, Aliaksandr Siarohin, Enver Sangineto, and Nicu Sebe. Whitening for self-supervised representation learning. In International Conference on Machine Learning, pages 3015-3024. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.774, + 0.47, + 0.829 + ], + "angle": 0, + "content": "[12] John Gideon and Simon Stent. Estimating heart rate from unlabelled video. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, pages 2743-2749, October 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[13] John Gideon and Simon Stent. The way to my heart is through contrastive learning: Remote photoplethysmography from unlabelled video. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3995-4004, October 2021. 1, 2, 3, 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.203 + ], + "angle": 0, + "content": "[14] Jean Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre H. Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Daniel Guo, Mohammad Gheshlaghi Azar, Bilal Piot, Koray Kavukcuoglu, Rémi Munos, and Michal Valko. Bootstrap your own latent a new approach to self-supervised learning. Advances in Neural Information Processing Systems, 2020-Decem(NeurIPS):1-14, 2020. arXiv: 2006.07733. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.204, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[15] Le Hou, Chen-Ping Yu, and Dimitris Samaras. Squared Earth Mover's Distance-based Loss for Training Deep Neural Networks, Apr. 2017. arXiv:1611.05916 [cs]. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.247, + 0.892, + 0.3 + ], + "angle": 0, + "content": "[16] Eugene Lee, Evan Chen, and Chen-Yi Lee. Meta-rppg: Remote heart rate estimation using a transductive meta-learner. In European Conference on Computer Vision (ECCV), 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.302, + 0.892, + 0.37 + ], + "angle": 0, + "content": "[17] Siqi Liu, Baoyao Yang, Pong C. Yuen, and Guoying Zhao. A 3d mask face anti-spoofing database with real world variations. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, June 2016. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.372, + 0.892, + 0.453 + ], + "angle": 0, + "content": "[18] Xin Liu, Josh Fromm, Shwetak Patel, and Daniel McDuff. Multi-task temporal shift attention networks for on-device contactless vitals measurement. In H. Larochelle, M. Ranzato, R. Hadsell, M. F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 19400-19411. Curran Associates, Inc., 2020. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.455, + 0.892, + 0.496 + ], + "angle": 0, + "content": "[19] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.498, + 0.892, + 0.552 + ], + "angle": 0, + "content": "[20] Hao Lu, Hu Han, and S Kevin Zhou. Dual-GAN: Joint BVP and Noise Modeling for Remote Physiological Measurement. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 12404-12413, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.553, + 0.892, + 0.621 + ], + "angle": 0, + "content": "[21] Hao Lu, Hu Han, and S Kevin Zhou. Dual-gan: Joint bvp and noise modeling for remote physiological measurement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12404-12413, 2021. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.623, + 0.892, + 0.691 + ], + "angle": 0, + "content": "[22] Camillo Lugaresi, Jiuqiang Tang, Hadon Nash, Chris McClanahan, Esha Uboweja, Michael Hays, Fan Zhang, Chuoling Chang, Ming Guang Yong, Juhyun Lee, et al. Mediapipe: A framework for building perception pipelines. arXiv preprint arXiv:1906.08172, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.692, + 0.892, + 0.72 + ], + "angle": 0, + "content": "[23] Daniel McDuff. Camera measurement of physiological vital signs. ACM Comput. Surv., aug 2022. Just Accepted. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.721, + 0.892, + 0.801 + ], + "angle": 0, + "content": "[24] Daniel McDuff, Miah Wander, Xin Liu, Brian L. Hill, Javier Hernandez, Jonathan Lester, and Tadas Baltrusaitis. SCAMPS: Synthetics for camera measurement of physiological signals. In Thirty-sixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.803, + 0.892, + 0.873 + ], + "angle": 0, + "content": "[25] Daniel J. McDuff, Ethan B. Blackford, and Justin R. Estepp. The Impact of Video Compression on Remote Cardiac Pulse Measurement Using Imaging Photoplethysmography. IEEE International Conference on Automatic Face and Gesture Recognition Workshops (FG), pages 63-70, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.874, + 0.892, + 0.903 + ], + "angle": 0, + "content": "[26] Yuriy Mironenko, Konstantin Kalinin, Mikhail Kopeliovich, and Mikhail Petrushan. Remote photoplethysmography:" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14472" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "Rarely considered factors. In IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 1197-1206, June 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.176 + ], + "angle": 0, + "content": "[27] Ishan Misra and Laurens van der Maaten. Self-Supervised Learning of Pretext-Invariant Representations. CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.471, + 0.261 + ], + "angle": 0, + "content": "[28] Andreia Moco, Sander Stuijk, Gerard De Haan, Ruikang K. Wang, and Wim Verkruysse. A Model for Waveform Dissimilarities in Dual-Depth Reflectance-PPG. Proceedings of the Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBS), 2018-July:5125-5130, 2018. ISBN: 9781538636466. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.264, + 0.47, + 0.319 + ], + "angle": 0, + "content": "[29] Xuesong Niu, Hu Han, Shiguang Shan, and Xin Chen. VIPL-HR: A multi-modal database for pulse estimation from less-constrained face video. In Asian Conference on Computer Vision (ACCV), pages 562-576, 2018. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.321, + 0.47, + 0.375 + ], + "angle": 0, + "content": "[30] Xuesong Niu, Shiguang Shan, Hu Han, and Xin Chen. RhythmNet: End-to-end heart rate estimation from face via spatial-temporal representation. IEEE Transactions on Image Processing, 29:2409-2423, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.377, + 0.47, + 0.446 + ], + "angle": 0, + "content": "[31] Ewa Nowara and Daniel McDuff. Combating the Impact of Video Compression on Non-Contact Vital Sign Measurement Using Supervised Learning. In IEEE International Conference on Computer Vision Workshop (ICCVW), pages 1706–1712, Oct. 2019. ISSN: 2473-9944. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.448, + 0.469, + 0.502 + ], + "angle": 0, + "content": "[32] Ewa M. Nowara, Daniel McDuff, and Ashok Veeraraghavan. Systematic analysis of video-based pulse measurement from compressed videos. Biomedical Optics Express, 12(1):494, 2021. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.504, + 0.469, + 0.544 + ], + "angle": 0, + "content": "[33] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.547, + 0.469, + 0.603 + ], + "angle": 0, + "content": "[34] M. Poh, D. J. McDuff, and R. W. Picard. Advancements in noncontact, multiparameter physiological measurements using a webcam. IEEE Transactions on Biomedical Engineering, 58(1):7-11, 2011. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.604, + 0.469, + 0.659 + ], + "angle": 0, + "content": "[35] Ming-Zher Poh, Daniel J. McDuff, and Rosalind W. Picard. Non-contact, automated cardiac pulse measurements using video imaging and blind source separation. Opt. Express, 18(10):10762-10774, May 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.661, + 0.469, + 0.716 + ], + "angle": 0, + "content": "[36] Michal Rapczynski, Philipp Werner, and Ayoub Al-Hamadi. Effects of video encoding on camera-based heart rate estimation. IEEE Transactions on Biomedical Engineering, 66(12):3360-3370, 2019. Publisher: IEEE. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.718, + 0.469, + 0.786 + ], + "angle": 0, + "content": "[37] Rencheng Song, Huan Chen, Juan Cheng, Chang Li, Yu Liu, and Xun Chen. Pulsegan: Learning to generate realistic pulse waveforms in remote photoplethysmography. IEEE Journal of Biomedical and Health Informatics, 25(5):1373-1384, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.469, + 0.857 + ], + "angle": 0, + "content": "[38] Jeremy Speth, Nathan Vance, Adam Czajka, Kevin Bowyer, and Patrick Flynn. Unifying frame rate and temporal dilations for improved remote pulse detection. Computer Vision and Image Understanding (CVIU), pages 1056-1062, 2021. 1, 2, 3, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[39] Jeremy Speth, Nathan Vance, Adam Czajka, Kevin Bowyer, Diane Wright, and Patrick Flynn. Deception detection and remote physiological monitoring: A dataset and baseline ex" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "perimental results. In International Joint Conference on Biometrics (IJCB), pages 4264-4271, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.177 + ], + "angle": 0, + "content": "[40] Radim Špetlík, Vojtech Franc, and Jírí Matas. Visual heart rate estimation with convolutional neural network. In Proceedings of the British machine vision conference, Newcastle, UK, pages 3-6, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.246 + ], + "angle": 0, + "content": "[41] Ronny Stricker, Steffen Muller, and Horst Michael Gross. Non-contact video-based pulse rate measurement on a mobile service robot. IEEE International Symposium on Robot and Human Interactive Communication, pages 1056-1062, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.304 + ], + "angle": 0, + "content": "[42] Zhaodong Sun and Xiaobai Li. Contrast-phys: Unsupervised video-based remote physiological measurement via spatiotemporal contrast. European Conference on Computer Vision (ECCV), 2022. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.306, + 0.892, + 0.389 + ], + "angle": 0, + "content": "[43] Nathan Vance, Jeremy Speth, Siamul Khan, Adam Czajka, Kevin W. Bowyer, Diane Wright, and Patrick Flynn. Deception detection and remote physiological monitoring: A dataset and baseline experimental results. IEEE Transactions on Biometrics, Behavior, and Identity Science (TBIOM), pages 1-1, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.391, + 0.892, + 0.432 + ], + "angle": 0, + "content": "[44] Wim Verkruysse, Lars O Svaasand, and J Stuart Nelson. Remote plethysmographic imaging using ambient light. Opt. Express, 16(26):21434-21445, Dec 2008. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.434, + 0.892, + 0.503 + ], + "angle": 0, + "content": "[45] Hao Wang, Euijoon Ahn, and Jinman Kim. Self-Supervised Representation Learning Framework for Remote Physiological Measurement Using Spatiotemporal Augmentation Loss. Proceedings of the AAAI Conference on Artificial Intelligence, 36(2):2431-2439, June 2022. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.504, + 0.892, + 0.546 + ], + "angle": 0, + "content": "[46] Wenjin Wang, Albertus C. Den Brinker, and Gerard De Haan. Single-Element Remote-PPG. IEEE Transactions on Biomedical Engineering, 66(7):2032-2043, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.548, + 0.892, + 0.589 + ], + "angle": 0, + "content": "[47] W. Wang, A. C. den Brinker, S. Stuijk, and G. de Haan. Algorithmic principles of remote PPG. IEEE Transactions on Biomedical Engineering, 64(7):1479-1491, 2017. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.591, + 0.892, + 0.646 + ], + "angle": 0, + "content": "[48] Wenjin Wang, Sander Stuijk, and Gerard De Haan. A Novel Algorithm for Remote Photoplethysmography: Spatial Subspace Rotation. IEEE Transactions on Biomedical Engineering, 63(9):1974-1984, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.648, + 0.892, + 0.731 + ], + "angle": 0, + "content": "[49] Zhen Wang, Yunhao Ba, Pradyumna Chari, Oyku Deniz Bozkurt, Gianna Brown, Parth Patwa, Niranjan Vaddi, Laleh Jalilian, and Achuta Kadambi. Synthetic generation of face videos with plethysmograph physiology. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20555-20564, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.732, + 0.892, + 0.773 + ], + "angle": 0, + "content": "[50] Yuzhe Yang, Xin Liu, Jiang Wu, Silviu Borac, Dina Katabi, Ming-Zher Poh, and Daniel McDuff. Simper: Simple self-supervised learning of periodic targets, 2022. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.775, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[51] Zitong Yu, Xiaobai Li, and Guoying Zhao. Remote photoplethysmograph signal measurement from facial videos using spatio-temporal networks. In Proceedings of the British Machine Vision Conference (BMVC), pages 1-12, 2019. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[52] Zitong Yu, Yuming Shen, Jingang Shi, Hengshuang Zhao, Philip H.S. Torr, and Guoying Zhao. Physformer: Facial video-based physiological measurement with temporal difference transformer. In Proceedings of the IEEE/CVF" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "14473" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.469, + 0.12 + ], + "angle": 0, + "content": "Conference on Computer Vision and Pattern Recognition (CVPR), pages 4186-4196, June 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.469, + 0.163 + ], + "angle": 0, + "content": "[53] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow Twins: Self-Supervised Learning via Redundancy Reduction. 2021. arXiv: 2103.03230. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.165, + 0.469, + 0.219 + ], + "angle": 0, + "content": "[54] Qi Zhan, Wenjin Wang, and Gerard de Haan. Analysis of CNN-based remote-PPG to understand limitations and sensitivities. Biomedical Optics Express, 11(3):1268-1283, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.221, + 0.469, + 0.29 + ], + "angle": 0, + "content": "[55] Yu Zhao, Bochao Zou, Fan Yang, Lin Lu, Abdelkader Nasreddine Belkacem, and Chao Chen. Video-based physiological measurement using 3d central difference convolution attention network. In 2021 IEEE International Joint Conference on Biometrics (IJCB), pages 1-6, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.292, + 0.469, + 0.346 + ], + "angle": 0, + "content": "[56] Hao Zhu, Wayne Wu, Wentao Zhu, Liming Jiang, Siwei Tang, Li Zhang, Ziwei Liu, and Chen Change Loy. CelebV-HQ: A large-scale video facial attributes dataset. In ECCV, 2022. 5, 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "14474" + } + ] +] \ No newline at end of file diff --git a/2023/Non-Contrastive Unsupervised Learning of Physiological Signals From Video/f2bb9236-2bf0-41f5-a980-2c135b55dca1_origin.pdf b/2023/Non-Contrastive Unsupervised Learning of Physiological Signals From Video/f2bb9236-2bf0-41f5-a980-2c135b55dca1_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..9c376207882be7733064e9f3c083e01fedd9705a --- /dev/null +++ b/2023/Non-Contrastive Unsupervised Learning of Physiological Signals From Video/f2bb9236-2bf0-41f5-a980-2c135b55dca1_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cd75da0a8e431ad0c83cec5a297b88d1c3627d254c2243eceffd8647c261871b +size 2209288 diff --git a/2023/Non-Contrastive Unsupervised Learning of Physiological Signals From Video/full.md b/2023/Non-Contrastive Unsupervised Learning of Physiological Signals From Video/full.md new file mode 100644 index 0000000000000000000000000000000000000000..eb461b0f035c882a7fb2a5395047730272db6ac1 --- /dev/null +++ b/2023/Non-Contrastive Unsupervised Learning of Physiological Signals From Video/full.md @@ -0,0 +1,325 @@ +# Non-Contrastive Unsupervised Learning of Physiological Signals from Video + +Jeremy Speth, Nathan Vance, Patrick Flynn, Adam Czajka +University of Notre Dame + +{jspeth,nvancel,flynn,aczajka}@nd.edu + +# Abstract + +Subtle periodic signals such as blood volume pulse and respiration can be extracted from RGB video, enabling noncontact health monitoring at low cost. Advancements in remote pulse estimation - or remote photoplethysmography (rPPG) - are currently driven by deep learning solutions. However, modern approaches are trained and evaluated on benchmark datasets with ground truth from contact-PPG sensors. We present the first non-contrastive unsupervised learning framework for signal regression to mitigate the need for labelled video data. With minimal assumptions of periodicity and finite bandwidth, our approach discovers the blood volume pulse directly from unlabelled videos. We find that encouraging sparse power spectra within normal physiological bandlimits and variance over batches of power spectra is sufficient for learning visual features of periodic signals. We perform the first experiments utilizing unlabelled video data not specifically created for rPPG to train robust pulse rate estimators. Given the limited inductive biases and impressive empirical results, the approach is theoretically capable of discovering other periodic signals from video, enabling multiple physiological measurements without the need for ground truth signals. + +# 1. Introduction + +Camera-based vitals estimation is a rapidly growing field enabling non-contact health monitoring in a variety of settings [23]. Although many of the signals avoid detection from the human eye, video data in the visible and infrared ranges contain subtle intensity changes caused by physiological oscillations such as blood volume and respiration. Significant remote photoplethysmography (rPPG) research for estimating the cardiac pulse has leveraged supervised deep learning for robust signal extraction [7, 18, 30, 38, 51, 52]. While the number of successful approaches has rapidly increased, the size of benchmark video datasets with simultaneous vitals recordings has remained relatively stagnant. + +Robust deep learning-based systems for deployment require training on larger volumes of video data with di + +verse skin tones, lighting, camera sensors, and movement. However, collecting simultaneous video and physiological ground truth with contact-PPG or electrocardiograms (ECG) is challenging for several reasons. First, many hours of high quality videos is an unwieldy volume of data. Second, recording a diverse subject population in conditions representative of real-world activities is difficult to conduct in the lab setting. Finally, synchronizing contact measurements with video is technically challenging, and even contact measurements used for ground truth contain noise. + +Fortunately, recent works find that contrastive unsupervised learning for rPPG is a promising solution to the data scarcity problem [13, 42, 45, 50]. We extend this research line into non-contrastive unsupervised learning to discover periodic signals in video data. With end-to-end unsupervised learning, collecting more representative training data to learn powerful visual features is much simpler, since only video is required without associated medical information. + +In this work, we show that non-contrastive unsupervised learning is especially simple when regressing rPPG signals. We find weak assumptions of periodicity are sufficient for learning the minuscule visual features corresponding to the blood volume pulse from unlabelled face videos. The loss functions can be computed in the frequency domain over batches without the need for pairwise or triplet comparisons. Figure 1 compares the proposed approach with supervised and contrastive unsupervised learning approaches. + +This work creates opportunities for scaling deep learning models for camera-based vitals and estimating periodic or quasi-periodic signals from unlabelled data beyond rPPG. Our novel contributions are: + +1. A general framework for physiological signal estimation via non-contrastive unsupervised learning (SiNC) by leveraging periodic signal priors. +2. The first non-contrastive unsupervised learning method for camera-based vitals measurement. +3. The first experiments and results of training with non-rPPG-specific video data without ground truth vitals. + +Source code to replicate this work is available at https://github.com/CVRL/SiNC-rPPG. + +![](images/86b7a5e481e3b4bd3a900a457ea1e7bf1d3c05fff191a7fa1c24b0152a24e020.jpg) +Figure 1. Overview of the SiNC framework for rPPG compared with traditional supervised and unsupervised learning. Supervised and contrastive losses use distance metrics to the ground truth or other samples. Our framework applies the loss directly to the prediction by shaping the frequency spectrum, and encouraging variance over a batch of inputs. Power outside of the bandlimits is penalized to learn invariances to irrelevant frequencies. Power within the bandlimits is encouraged to be sparsely distributed near the peak frequency. + +# 2. Related Work + +# 2.1. Remote Photoplethysmography (rPPG) + +Approaches for remote pulse estimation have shifted over the last decade from blind source separation [34, 35], through linear color transformations [9, 10, 46, 47] to training supervised deep learning-based models [7, 16, 18, 20, 29, 30, 38, 51, 52, 55]. While the color transformations generalize well across datasets, deep learning-based models give better accuracy when tested on data from a similar distribution to the training set. To this end, deep learning research has focused on optimizing architectures for extracting robust spatial and temporal features from the limited datasets. + +To get around the data bottleneck, large synthetic datasets have recently been proposed [24, 49]. The SCAMPS dataset [24] contains 2,800 videos of synthetic avatars with corresponding PPG, EKG, respiration, and facial action units. The UCLA-synthetic dataset [49] contains 480 videos, which are shown to improve performance when trained with real data. A strength of synthetic datasets is their ability to cover the broad range of skin tones compared to physical data. Another solution to the lack of physiological data is unsupervised learning, where a large set of videos and periodic priors on the output signal is sufficient [13,42,45,50]. We discuss these methods in Sec. 2.3. + +# 2.2. Unsupervised Learning + +Self-supervised learning is progressing for image representation learning, with two main competing classes of approaches: contrastive and non-contrastive (or regularized) learning [3]. Contrastive approaches [5, 6, 27] define criteria for distinguishing whether two or more samples are + +the same or different. Non-contrastive methods augment positive pairs, and enforce variance in the predictions over batches to avoid collapse, in which the model's embeddings reside in a small subspace of the feature space [3]. Distillation methods only use positive samples and avoid collapse by applying a moving average and stop-gradient operator [8, 14]. Another approaches maximize information content of embeddings [11, 53]. + +# 2.3. Unsupervised Learning for rPPG + +All existing unsupervised rPPG approaches are contrastive [13,42,45,50], in which the same model is fed pairs of input videos, and the predictions over similar videos are pulled closer, while the predictions from dissimilar videos are repelled. Gideon et al. [13] were the first to train a deep learning model without labels for rPPG using the contrastive framework. The core of their approach is frequency resampling to create negative samples. Although spatially similar, a resampled video clip contains a different underlying pulse rate, so the model must learn to attend to the temporal dynamics. For their distance function between pairs, they calculated the mean square error between power spectral densities of the model's waveform predictions. While their approach learns to estimate the pulse, their formulation with negative samples is imprecise. The resampled frequency for the negative sample is known, so the relative translation of the power spectrum from the anchor sample can be directly computed. Thus, rather than repelling the estimated spectra, it is more accurate to penalize differences from the known spectra. Furthermore, resampling close to the original sampling rate causes overlap in the power spectra, so repelling the pair is inaccurate. Yuzhe et al. [50] + +incorporated the previously ignored resampling factor for a soft modification to the InfoNCE loss [33] that scales the desired similarity between pairs by their relative sampling rate. A downside is that their learning framework is not end-to-end unsupervised and requires fine-tuning with PPG labels after the self-supervised stage. + +Differently from [13], Contrast-Phys [42] and SLF-RPM [45] consider all samples different from the anchor to be negatives. This assumes that the power spectra will vary between subjects or sufficiently long windows for the same subject. This runs into similar issues with negative pairs as Gideon's approach. Different subjects may have the same pulse rate, so punishing the model for predicting similar frequencies is common during training. Furthermore, the Fast Fourier Transform (FFT) does not produce perfectly sparse decompositions, resulting in spectral overlap even if the heart rate differs by several beats per minute (bpm). As an example, the last column of the second row in Fig. 2 shows the nulls of the main lobe are nearly 30 bpm apart. + +# 3. Method + +We first formulate signal regression from video. A video sample $x_{i} \in \mathbb{R}^{T \times W \times H \times C}$ sampled from a dataset $\mathcal{D}$ consists of $T$ images of size $W \times H$ pixels across $C$ channels, captured over time. State-of-the-art methods offer models $f$ that regress a waveform $\mathbb{R}^T \ni y_i = f(x_i)$ of the same length as the video. Recently, the task has been effectively modeled end-to-end with the models $f$ being spatiotemporal neural networks [16, 21, 38, 51, 52]. While most previous works are supervised and minimize the loss to a contact pulse measurement, we perform non-contrastive learning using only the model's estimated waveform. + +The key realization is that we can place strong priors on the estimated pulse regarding its bandwidth and periodicity. Observed signals outside the desired frequency range are pollutants, so penalizing the model for carrying them through the forward pass results in invariances to such noisy visual features. We find that it is surprisingly easy to impose the desired constraints in the frequency domain. Thus, all waveforms are transformed into their discrete Fourier components with the FFT before computing all losses in our approach, $F = \mathrm{FFT}(y)$ . The following sections describe the loss functions and augmentations used during training. + +# 3.1. Losses + +One of the advantages of unsupervised learning for periodic signals is that we can constrain the solution space significantly. For physiological signals such as respiration and blood volume pulse, we know the healthy upper and lower bounds of the frequencies. We also desire the extracted signal to be sparse in the frequency domain, and that our model filters out noise signals present in the video. With these + +constraints, we can greatly simplify the problem of finding good features for the desired signal in the data. + +Bandwidth Loss. One of the most powerful constraints we can place on the model is frequency bandlimits. Past unsupervised methods have used the irrelevant power ratio (IPR) as a validation metric [12, 13, 42] for model selection. We find that it is also effective during model training. The IPR penalizes the model for generating signals outside the desired bandlimits. With lower and upper bandlimits of $a$ and $b$ , respectively, our bandwidth loss becomes: + +$$ +L _ {b} = \frac {1}{\sum_ {i = - \infty} ^ {\infty} F _ {i}} \left[ \sum_ {i = - \infty} ^ {a} F _ {i} + \sum_ {i = b} ^ {\infty} F _ {i} \right], \tag {1} +$$ + +where $F_{i}$ is the power in the $i$ th frequency bin of the predicted signal. This simple loss enforces learning of many invariants, such as movement from respiration, talking, or facial expressions which typically occupy low frequencies. In our experiments we specify the limits as $a = 0.66\mathrm{Hz}$ to $b = 3\mathrm{Hz}$ , which corresponds to a common pulse rate range from 40 bpm to 180 bpm. The first column of Fig. 2 shows the result of training exclusively with the bandwidth loss $L_{b}$ . The last row shows that the model concentrates signal power between the bandlimits. + +Sparsity Loss. The pulse rate is the most common physiological marker associated with the blood volume pulse. Since we are primarily interested in the frequency, we can further improve our model by preventing wideband predictions. This also reveals the true signal we aim to discover by ignoring visual dynamics that are not strongly periodic. + +We penalize energy within the bandlimits that is not near the spectral peak: + +$$ +L _ {s} = \frac {1}{\sum_ {i = a} ^ {b} F _ {i}} \left[ \sum_ {i = a} ^ {F ^ {*} - \Delta_ {F}} F _ {i} + \sum_ {i = F ^ {*} + \Delta_ {F}} ^ {b} F _ {i} \right], \tag {2} +$$ + +where $F^{*} = \operatorname{argmax}(F)$ and $\Delta_F$ are the frequencies of the spectral peak and padding around the peak, respectively. For all experiments $\Delta_F = 6$ beats per minute [32]. Figure 2 shows the result of training only with the sparsity loss in the second column. For the whole dataset, the power spectrum is sparsely distributed in the low frequencies, effectively filtering frequencies higher than $1\mathrm{Hz}$ . + +Variance Loss. One of the risks of non-contrastive methods is the model collapsing into trivial solutions and making predictions independently of the input features. In regularized methods such as VICReg [3], a hinge loss on the variance over a batch of predictions is used to enforce diverse outputs. We use a similar strategy to avoid model collapse, but instead spread the variance in power spectral densities towards a uniform distribution over the desired frequency band. Our variance loss processes a uniform prior + +![](images/73f0f9e58fa4278de59a7cde858e22205745b33283d03d17bac423c199cc6949.jpg) + +![](images/a1027376b11c005af2d90c484d98ba8068fba2129028fd19b37ae6c00ba91c3d.jpg) + +![](images/d90f10d35c3bfa89e4e9233727e03315dc60db5cd2ec38029fe33a0228e31ce3.jpg) + +![](images/5d18cba2f2c08ba61a85737653aa36e1f6cf9ecdb0dadefbab6505365b097ade.jpg) + +![](images/c788f5df5c793649808495d38e68eba7be9a09354a4e016ad5d362e993e14c85.jpg) + +![](images/c0ff9ba078c287422e89c05207b730daddd53f87b305be09b17a175c76c59c3a.jpg) + +Figure 2. Each column shows predictions from models trained with one or all of the losses for 20 epochs on UBFC-rPPG. The first two rows show a sample in the time and frequency domain, respectively. The last row shows the signal power over the validation set computed by taking the sum of normalized power spectral densities from each sample, then dividing the result by the number of validation samples. The bandwidth loss penalizes signal power outside predefined bandlimits (40 to 180 bpm) to constrain the output space. The sparsity loss encourages a narrow spectrum containing strong periodicity. The variance loss encourages diverse power spectra in a batch, preventing the model from collapsing to a narrow bandwidth. When combined, the model estimates periodic signals within the desired bandlimits. +![](images/dbb4c7c8a43c9aac763e5d37eecc40de405f8402d698cc8a554b1abc44ce62e0.jpg) +Predicted Ground Truth + +![](images/9414fa58b7d7d7b0713aa77f30a39dd06842029cbd70f07eb3717f3f7a1955d6.jpg) + +![](images/e79366bfc32a63af994c89c0ae115e2d0b9ac9d76a25c2ed3b374c48effc9dd9.jpg) + +distribution $P$ over $d$ frequencies, and a batch of $n$ spectral densities, $F = [v_{1},\dots,v_{n}]$ , where each vector is a $d$ -dimensional frequency decomposition of a predicted waveform. We calculate the normalized sum of densities over the batch, $Q$ , and define the variance loss as the squared Wasserstein distance [15] to the uniform prior: + +$$ +L _ {v} = \frac {1}{d} \sum_ {i = 1} ^ {d} \left(\mathrm {C D F} _ {i} (Q) - \mathrm {C D F} _ {i} (P)\right) ^ {2}, \tag {3} +$$ + +where CDF is a cumulative distribution function. The third column of Fig. 2 shows the effect of the variance loss during training. For a single sample, wide-band signals containing multiple frequencies are predicted, and the predicted frequencies cover the task's bandwidth. In our experiments we use a batch size of 20 samples. See the suppl. materials for an ablation experiment on the impact of batch size. + +Combining All Losses. Summarizing, our training loss function is a simple sum of the aforementioned losses: + +$$ +L = L _ {b} + L _ {s} + L _ {v}. \tag {4} +$$ + +While one could weight particular components of the loss more than others, we specifically formulated the losses to scale them between 0 and 1. In our experiments, we find that a simple summation without weighting gives good performance. The combined loss function encourages the model to search over the supported frequencies to discover visual features for a strong periodic signal. Remarkably, we find that this simple framework is sufficient for learning to regress the blood volume in video, as shown in the last column of Fig. 2. + +# 3.2. Augmentations + +Unlike Gideon et al.'s [13] approach, which only applies frequency augmentations, we apply several augmentations to both the spatial and temporal dimensions to learn invariances to noisy visual signals. In fact, we found that without augmentations, models did not converge during training (see the supplementary material). + +Image Intensity Augmentations. Gaussian noise is added to each pixel with zero mean and a standard deviation of 2 on an image scale from 0 to 255. Each clip is darkened or brightened by adding a constant from a Gaussian distribution with mean 0 and standard deviation of 10. + +Spatial Augmentations. We randomly horizontally flip a video clip with $50\%$ probability. The spatial dimension of a clip are randomly square cropped down to between half the original length and the original length. The cropped clip is then linearly interpolated back to the original dimensions. + +Temporal Augmentations. With the general assumption that the desired signal is strongly periodic and sparsely represented in the Fourier domain, we randomly flip a video clip along the time dimension with a probability of $50\%$ . Note that the Fourier decomposition of a time-reversed sinusoid is identical to that of the original sinusoid. + +Frequency Augmentations. Perhaps the most important augmentation is frequency resampling [13], where the video is linearly interpolated to a different frame rate. This augmentation is particularly interesting for rPPG, because it transforms the video input and target signal equivalently along the time dimension, making it equivariant. Given the aforementioned transformations that are invariant, $\tau (\cdot)\sim$ $\mathcal{T}$ , the equivariant frequency resampling operation, $\phi (\cdot)\sim$ $\Phi$ , and a model $f(\cdot)$ that infers a waveform from a video we have the following: + +$$ +\phi (f (\tau (x))) = f (\phi (\tau (x))). \tag {5} +$$ + +This is a powerful augmentation, because it allows us to augment the target distribution along with the video input. In our experiments we randomly resample input clips by a factor $c \sim U(0.6, 1.4)$ . After applying the resampling augmentation, we scale the bandlimits by $c$ , to avoid penalizing the model if the augmentation pushed the underlying pulse frequency outside of the original bandlimits. + +# 4. Datasets + +We use PURE [41], UBFC-rPPG [4], and DDPM [39] as benchmark rPPG datasets for training and testing, and CelebV-HQ dataset [56] and HKBU-MARs [17] for unsupervised training only. + +Deception Detection and Physiological Monitoring (DDPM) [39, 43] consists of data from 86 subjects attempting to answer questions deceptively. Interviews were recorded at 90 frames-per-second for more than 10 minutes + +on average. Natural conversation and head pose changes make it a difficult and less-constrained rPPG dataset. + +PURE [41] is a benchmark rPPG dataset consisting of 10 subjects recorded over 6 sessions. Each session lasted approximately 1 minute, and raw video was recorded at 30 fps. The 6 sessions for each subject consisted of: (1) steady, (2) talking, (3) slow head translation, (4) fast head translation, (5) small and (6) medium head rotations. Pulse rates are at or close to the subject's resting rate. + +UBFC-rPPG [4] contains 1-minute long videos from 42 subjects recorded at 30 fps. Subjects played a time-sensitive mathematical game to raise their heart rates, but head motion is limited during the recording. + +HKBU 3D Mask Attack with Real World Variations (HKBU-MARs) [17] consists of 12 subjects captured over 6 different lighting configurations with 7 different cameras each, resulting in 504 videos lasting 10 seconds each. The diverse lighting and camera sensors make it a valuable dataset for unsupervised training. We use version 2 of HKBU-MARs, which contains videos with both realistic 3D masks and unmasked subjects. + +High-Quality Celebrity Video Dataset (CelebV-HQ) [56] is a set of processed YouTube videos containing 35,666 face videos from over 15,000 identities. The videos vary dramatically in length, lighting, emotion, motion, skin tones, and camera sensors. The greatest challenge in harnessing online videos is their reduced quality due to compression before upload and by the video provider. Compression is a known challenge for rPPG, since the blood volume pulse is so subtle optically [25, 31, 32, 36]. + +# 5. Training Details + +# 5.1. Data Preprocessing + +To prepare the video clips for the spatiotemporal deep learning models, we first extract 68 face landmarks with OpenFace [2]. We then define a bounding box in each frame with the minimum and maximum $(x,y)$ locations by extending the crop horizontally by $5\%$ to ensure that the cheeks and jaw are present. The top and bottom are extended by $30\%$ and $5\%$ of the bounding box height, respectively, to include the forehead and jaw. We further extend the shorter of the two axes to the length of the other to form a square. The cropped frames are then resized to $64\times 64$ pixels with bicubic interpolation. For faster processing of the massive CelebV-HQ [56] dataset, we instead use MediaPipe Face Mesh [22] for landmarking. + +# 5.2. Model Architectures + +We use a 3D-CNN architecture similar to [38] without temporal dilations [51]. We use a temporal kernel width of 5, and replace default zero-padding by repeating the edges. Zero-padding along the time dimension can result + +Table 1. Intra-dataset pulse rate estimation results. The best results for a given metric/dataset are bolded, and the second-best results are underlined. For a better comparison with results in the literature, this table follows [42]. MAE: Mean Absolute Error; RMSE: Root Mean Square Error; $r$ : Pearson correlation coefficient. + +
TypesMethodsUBFC-rPPGPUREDDPM
MAE (bpm)RMSE (bpm)rMAE (bpm)RMSE (bpm)rMAE (bpm)RMSE (bpm)r
TraditionalGREEN [44]7.5014.410.627.2317.050.6932.7943.090.04
ICA [34]5.1711.760.653.7612.600.8522.2235.770.40
CHROM [9]2.369.230.870.752.231.0013.4828.530.56
2SR [48]6.9018.500.652.443.060.9822.0839.730.22
POS [47]2.119.110.870.804.110.989.0323.070.70
SupervisedHR-CNN [40]---1.842.370.98---
SynRhythm [29]5.596.820.72------
PulseGAN [37]1.192.100.98------
Dual-GAN [21]0.440.670.990.821.310.99---
RPNet [38]†0.53 ± 0.011.78 ± 0.020.99 ± 0.001.15 ± 0.275.77 ± 1.250.96 ± 0.023.46 ± 0.2412.47 ± 0.680.91 ± 0.01
PhysNet [51]†0.55 ± 0.032.03 ± 0.370.99 ± 0.000.99 ± 0.195.22 ± 0.930.97 ± 0.013.96 ± 0.7613.57 ± 1.740.89 ± 0.03
UnsupervisedGideon2021 [13]1.854.280.932.32.90.99---
SLF-RPM [45]*8.399.700.70------
SimPer [50]*4.24--3.89-----
Contrast-Phys [42]0.641.000.991.001.400.999.70 ± 2.9025.02 ± 6.010.58 ± 0.19
SiNC (ours)0.59 ± 0.001.83 ± 0.040.99 ± 0.000.61 ± 0.061.84 ± 0.401.00 ± 0.005.87 ± 0.1117.44 ± 0.160.81 ± 0.00
+ +* Some methods in the unsupervised row require fine-tuning on labeled data with a linear classifier. +† Some supervised methods were trained with identical data augmentations to SiNC for fair comparison. + +in edge effects that add artificial frequencies to the predictions. Early experiments showed that temporal dilations caused aliasing and reduced the bandwidth of the model to specific frequencies. Our losses and framework may be applied to any task and architecture with dense predictions along one or more dimensions. However, popular rPPG architectures such as DeepPhys [7] and MTTS-CAN [18] may be ill-suited for the approach, since they consume very few frames, and the number of time points should be large enough to give sufficient frequency resolution with the FFT. In our experiments, we use the AdamW [19] optimizer with a learning rate of 0.0001. We use a clip length of $T = 120$ frames (4 seconds), and we set the input signal's length to achieve a frequency resolution of 0.33 bpm. + +# 5.3. Supervised Training + +To properly compare our approach to its supervised counterpart we use the same model architecture and train it with the commonly used negative Pearson loss between the predicted waveform and the contact sensor ground truth [51]. During training we apply all of the same augmentations except time reversal. Models are trained for 200 epochs on PURE and UBFC-rPPG, and for 40 epochs on DDPM. The model from the epoch with the lowest loss on the validation set is selected for testing. + +# 5.4. Unsupervised Training + +Unsupervised models are trained for the same number of epochs as the supervised setting for both PURE and UBFCrPPG, but we train for an additional 40 epochs on DDPM, + +since this dataset is considerably more difficult. We set the batch size to 20 samples during training. Contrary to previous unsupervised approaches [13, 42], we leverage validation sets for model selection by selecting the model with the lowest combined bandwidth and sparsity losses. The creation of the dataset splits is described in the next section. + +# 5.5. Evaluation + +Pulse rates are computed as the highest spectral peak between $0.66\mathrm{Hz}$ and $3\mathrm{Hz}$ (equivalent to $40\mathrm{bpm}$ to $180\mathrm{bpm}$ ) over a 10-second sliding window. The same procedure is applied to the ground truth waveforms for a reliable evaluation [26]. We apply common error metrics such as mean absolute error (MAE), root mean square error (RMSE), and Pearson correlation coefficient $(r)$ between the pulse rates. + +We perform 5-fold cross validation for both PURE and UBFC with the same folds as [13], and use the predefined dataset splits from DDPM [38]. Differently from [13], we use 3 of the folds for training, 1 for validation, and the remaining for testing rather than only training and testing partitions. We train 3 models with different initializations, resulting in 15 models trained on PURE and UBFC each, and 3 models trained on DDPM. We present the mean and standard deviation of the errors in the results. + +# 6. Results + +# 6.1. Within-Dataset Testing + +Table 1 shows the results for models trained and tested on subject-disjoint partitions from the same datasets. For + +PURE and UBFC we achieve MAE lower than 1 bpm, performing better or on par with all traditional and supervised learning approaches. For PURE, our approach gives the lowest MAE and a Pearson $r$ of nearly 1. Performance drops on DDPM due to the overall difficulty of the dataset. SiNC outperforms contrastive approaches, only being surpassed by supervised deep learning models. + +In comparison to other unsupervised methods, Contrast-Phys [42] gives the most competitive performance on all but DDPM. Note that our approach gives the lowest MAE on all datasets, but has higher RMSE. We believe this is due their use of harmonic removal as a post-processing step when estimating the pulse rate, which is not described in [42], but can be found in their publicly available code. + +# 6.2. Cross-Dataset Testing + +Table 2. Cross-dataset pulse rate estimation performance. The top 3 training datasets are common rPPG benchmarks, while HKBU was not designed for rPPG and has no pulse ground truth. Note that testing on UBFC after training on DDPM performs well, since their frequency distributions are similar, while PURE's pulse rates tend to be much lower. + +
Training DatasetTesting DatasetMethodMAE (bpm)r
DDPMUBFCSiNC0.88 ± 0.250.98 ± 0.01
UBFCContrast-Phys1.14 ± 0.380.96 ± 0.03
UBFCPhysNet1.11 ± 0.420.95 ± 0.02
PURESiNC3.12 ± 1.070.88 ± 0.05
PUREContrast-Phys13.02 ± 6.120.19 ± 0.59
PUREPhysNet1.46 ± 0.340.95 ± 0.02
UBFCDDPMSiNC18.53 ± 0.360.38 ± 0.01
DDPMContrast-Phys22.93 ± 1.020.18 ± 0.04
DDPMPhysNet18.58 ± 0.120.40 ± 0.00
PURESiNC4.02 ± 0.060.86 ± 0.00
PUREContrast-Phys19.61 ± 2.010.33 ± 0.06
PUREPhysNet3.81 ± 0.340.87 ± 0.02
PUREUBFCSiNC6.64 ± 1.760.59 ± 0.10
UBFCContrast-Phys10.22 ± 0.380.45 ± 0.04
UBFCPhysNet7.02 ± 3.350.60 ± 0.13
DDPMSiNC24.92 ± 0.650.20 ± 0.00
DDPMContrast-Phys29.63 ± 0.480.03 ± 0.02
DDPMPhysNet28.03 ± 2.200.13 ± 0.05
HKBU (non-rPPG)UBFCSiNC1.08 ± 0.030.95 ± 0.00
PURESiNC2.43 ± 0.200.90 ± 0.02
DDPMSiNC20.34 ± 0.250.19 ± 0.02
+ +We perform cross-dataset testing to analyze whether the approach is robust to changes to the lighting, camera sensor, pulse rate distribution, and motion. Table 2 shows the results for SiNC and supervised training with the same architecture. We find that the performance is similar for the supervised and unsupervised approaches when transferring to different data sources. Training on PURE exclusively gives relatively poor results when transferring to + +UBFC-rPPG and DDPM, due to the low pulse rate variability within PURE samples and lack of movement. Training on DDPM gives the best results overall, since the dataset is the largest and captures larger subjects' movements compared to other datasets. + +# 6.3. Training with CelebV-HQ Videos + +Given the abundance of face videos publicly available online, we trained a model on the CelebV-HQ dataset [56]. After processing the available videos with MediaPipe and resampling to 30 fps, our unlabeled dataset consisted of 34,029 videos. We trained the model for 23 epochs and manually stopped training due to a plateau in the validation loss. Unfortunately, we found that the model could not converge to the true blood volume pulse. We attribute the failure to poor video quality from compression. Although the videos were downloaded with the highest available quality, they have likely been compressed, removing the pulse signal entirely. See the supplementary material for an assessment of rPPG quality using POS [47] on CelebV-HQ. + +# 6.4. Training with HKBU-MARs Videos + +The HKBU-MARs dataset [17] was designed for face presentation attack detection, but we trained models on the "real" video sessions in the dataset. The bottom rows in Table 2 show the results for training on HKBU-MARs, then testing on the benchmark rPPG datasets. Training on HKBU-MARs gives better results when testing on UBFC-rPPG and PURE than all training sets except DDPM, which is an order of magnitude larger. To our knowledge, this is the first successful experiment showing that non-rPPG videos can be used to train robust rPPG models, even if they do not have ground-truth pulse labels. + +# 6.5. Ablation Study on Losses + +Table 3. Ablation study on the loss functions used during training. Results are shown for models trained and tested on UBFC-rPPG. + +
LossMAE (bpm)RMSE (bpm)r
Lb3.08 ± 1.698.08 ± 3.610.87 ±0.08
Ls45.50 ± 1.2250.04 ± 0.94-0.04 ±0.08
Lv22.89 ± 2.8331.51 ± 2.360.22 ±0.09
Ls + Lv51.24 ± 5.3657.80 ± 7.39-0.04 ±0.09
Lb + Ls9.99 ± 2.5517.14 ± 2.360.51 ±0.14
Lb + Lv4.18 ± 2.888.90 ± 5.240.82 ±0.14
Lb + Ls + Lv0.59 ± 0.001.83 ± 0.040.99 ± 0.00
+ +We trained models using all combinations of loss components to analyze their contributions. Table 3 shows the results for training and testing on UBFC-rPPG. The bandwidth loss is the most critical for discovering the true blood volume pulse, while the sparsity and variance losses do not + +![](images/0d213413509265e049b90d0f45060212d169c24cc73034bdab220e757a890931.jpg) + +![](images/ba9286360b5aae1a58396a024019804f765b49d433e1589728c7e53827cd5e2d.jpg) + +![](images/2a5daec4bc3fe5d7907ed4f892d1f55781180759b00821dd5d4d397f39cb90d9.jpg) +Figure 3. Within-dataset waveform predictions on all baseline datasets from end-to-end unsupervised models over an 8-second window. The model predictions are remarkably periodic without any form of filtering. Note that phase is not considered during training, so each model learns its own phase shift. + +learn the desired signal by themselves. Surprisingly, combining the bandwidth loss with just one of the sparsity or variance losses gives worse performance than just the bandwidth loss. However, when combining all three components, the model achieves impressive results. + +# 7. Discussion + +# 7.1. Improvements over Supervised Learning + +It is initially surprising that unsupervised training leads to similar or improved rPPG estimation models compared to those trained in a supervised manner. However, there are several potential benefits to unsupervised training. From a hardware perspective, one of the difficulties in supervised training is aligning the contact pulse waveform with the video frames [54]. The pulse sensor and camera may have a time lag, effectively giving the model an out-of-phase target at training time. Unsupervised training gives the model freedom to learn the phase directly from the video. The contact-PPG signal is also sensitive to motion and may be noisy. Since motion may co-occur at the face and fingertip, the contact signal may misguide the model towards visual features for which they should be invariant. + +From a physiological perspective, the pulse observed optically at the fingertip with a contact sensor has a different phase than that of the face, since blood propagates along a + +different path before reaching the peripheral microvasculature, making alignment nearly impossible without shifting the targets to rPPG estimates from existing methods [38]. Additionally, the morphological shape of the contact-PPG waveform depends on numerous factors such as the wavelength of light (and corresponding tissue penetration depth), external pressure from the oximeter clip, and vasodilation at the measurement site [1,28]. This indicates that the morphology and phase of the target PPG waveform is likely different from the observed rPPG waveform. + +# 7.2. Why Does It Work? + +The success of SiNC depends on specific properties of the data, model, and how the two interact. Limited model capacity is actually required, since it forces learning features that generalize across inputs. By constraining the model's predictions to have specific periodic properties, the limited-capacity model must find a general set of features to produce a signal that exists in most or all of the training samples, which is the blood volume pulse in our datasets. + +As a beneficial side-effect, the model intrinsically learns to ignore common noise factors such as illumination, rigid motion, non-rigid motion (e.g. talking, smiling, etc.), and sensor noise, since they may preside outside the predefined bandlimits or with uniform power spectra. Even if noise exhibits periodic tendencies within the bandlimits for some samples, those features would produce poor signals on other samples. Therefore, end-to-end unsupervised approaches like SiNC are well-suited for periodic problems. + +# 8. Conclusions + +We proposed a novel non-contrastive learning approach for end-to-end unsupervised signal regression, with specific experiments on blood volume pulse estimation from face videos. This SiNC framework effectively learns powerful visual features with only loose frequency constraints. We demonstrated this by training accurate rPPG models using non-rPPG data and our simple loss functions. Given the subtlety of the rPPG signal, we believe our work can be extended to other signal regression tasks in the domain of remote vitals estimation. + +Acknowledgements. This research was sponsored by the Securiport Global Innovation Cell, a division of Securiport LLC. Commercial equipment is identified in this work in order to adequately specify or describe the subject matter. In no case does such identification imply recommendation or endorsement by Securiport LLC, nor does it imply that the equipment identified is necessarily the best available for this purpose. The opinions, findings, and conclusions or recommendations expressed in this publication are those of the authors and do not necessarily reflect the views of our sponsors. + +# References + +[1] Pierre Abraham, Mélissa Bourgeau, Maïte Camo, Anne Humeau-Heurtier, Sylvain Durand, Pascal Rousseau, and Guillaume Mahe. Effect of skin temperature on skin endothelial function assessment. Microvascular Research, 88:56-60, 2013. 8 +[2] T. Baltrusaitis, A. Zadeh, Y. C. Lim, and L. Morency. Openface 2.0: Facial behavior analysis toolkit. In IEEE International Conference on Automatic Face Gesture Recognition (FG), pages 59-66, 2018. 5 +[3] Adrien Bardes, Jean Ponce, and Yann LeCun. Vicreg: Variance-invariance-covariance regularization for self-supervised learning. In ICLR, 2022. 2, 3 +[4] Serge Bobbia, Richard Macwan, Yannick Benezeth, Alamin Mansouri, and Julien Dubois. Unsupervised skin tissue segmentation for remote photoplethysmography. Pattern Recognition Letters, 124:82-90, 2019. 5 +[5] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in Neural Information Processing Systems, 2020-Decem(NeurIPS):1-13, 2020. arXiv: 2006.09882. 2 +[6] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. 37th International Conference on Machine Learning, ICML 2020, pages 1575-1585, 2020. arXiv: 2002.05709 ISBN: 9781713821120. 2 +[7] Weixuan Chen and Daniel McDuff. DeepPhys: Video-based physiological measurement using convolutional attention networks. In European Conference on Computer Vision (ECCV), pages 356-373, 2018. 1, 2, 6 +[8] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 15745–15753, 2021. arXiv: 2011.10566 ISBN: 9781665445092. 2 +[9] G. De Haan and V. Jeanne. Robust pulse rate from chrominance-based rppg. IEEE Transactions on Biomedical Engineering, 60(10):2878-2886, 2013. 2, 6 +[10] Gerard De Haan and Arno Van Leest. Improved motion robustness of remote-PPG by using the blood volume pulse signature. Physiological Measurement, 35(9):1913–1926, 2014. 2 +[11] Aleksandr Ermolov, Aliaksandr Siarohin, Enver Sangineto, and Nicu Sebe. Whitening for self-supervised representation learning. In International Conference on Machine Learning, pages 3015-3024. PMLR, 2021. 2 +[12] John Gideon and Simon Stent. Estimating heart rate from unlabelled video. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, pages 2743-2749, October 2021. 3 +[13] John Gideon and Simon Stent. The way to my heart is through contrastive learning: Remote photoplethysmography from unlabelled video. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3995-4004, October 2021. 1, 2, 3, 5, 6 + +[14] Jean Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre H. Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Daniel Guo, Mohammad Gheshlaghi Azar, Bilal Piot, Koray Kavukcuoglu, Rémi Munos, and Michal Valko. Bootstrap your own latent a new approach to self-supervised learning. Advances in Neural Information Processing Systems, 2020-Decem(NeurIPS):1-14, 2020. arXiv: 2006.07733. 2 +[15] Le Hou, Chen-Ping Yu, and Dimitris Samaras. Squared Earth Mover's Distance-based Loss for Training Deep Neural Networks, Apr. 2017. arXiv:1611.05916 [cs]. 4 +[16] Eugene Lee, Evan Chen, and Chen-Yi Lee. Meta-rppg: Remote heart rate estimation using a transductive meta-learner. In European Conference on Computer Vision (ECCV), 2020. 2, 3 +[17] Siqi Liu, Baoyao Yang, Pong C. Yuen, and Guoying Zhao. A 3d mask face anti-spoofing database with real world variations. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, June 2016. 5, 7 +[18] Xin Liu, Josh Fromm, Shwetak Patel, and Daniel McDuff. Multi-task temporal shift attention networks for on-device contactless vitals measurement. In H. Larochelle, M. Ranzato, R. Hadsell, M. F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 19400-19411. Curran Associates, Inc., 2020. 1, 2, 6 +[19] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019. 6 +[20] Hao Lu, Hu Han, and S Kevin Zhou. Dual-GAN: Joint BVP and Noise Modeling for Remote Physiological Measurement. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 12404-12413, 2021. 2 +[21] Hao Lu, Hu Han, and S Kevin Zhou. Dual-gan: Joint bvp and noise modeling for remote physiological measurement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12404-12413, 2021. 3, 6 +[22] Camillo Lugaresi, Jiuqiang Tang, Hadon Nash, Chris McClanahan, Esha Uboweja, Michael Hays, Fan Zhang, Chuoling Chang, Ming Guang Yong, Juhyun Lee, et al. Mediapipe: A framework for building perception pipelines. arXiv preprint arXiv:1906.08172, 2019. 5 +[23] Daniel McDuff. Camera measurement of physiological vital signs. ACM Comput. Surv., aug 2022. Just Accepted. 1 +[24] Daniel McDuff, Miah Wander, Xin Liu, Brian L. Hill, Javier Hernandez, Jonathan Lester, and Tadas Baltrusaitis. SCAMPS: Synthetics for camera measurement of physiological signals. In Thirty-sixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2022. 2 +[25] Daniel J. McDuff, Ethan B. Blackford, and Justin R. Estepp. The Impact of Video Compression on Remote Cardiac Pulse Measurement Using Imaging Photoplethysmography. IEEE International Conference on Automatic Face and Gesture Recognition Workshops (FG), pages 63-70, 2017. 5 +[26] Yuriy Mironenko, Konstantin Kalinin, Mikhail Kopeliovich, and Mikhail Petrushan. Remote photoplethysmography: + +Rarely considered factors. In IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 1197-1206, June 2020. 6 +[27] Ishan Misra and Laurens van der Maaten. Self-Supervised Learning of Pretext-Invariant Representations. CVPR, 2020. 2 +[28] Andreia Moco, Sander Stuijk, Gerard De Haan, Ruikang K. Wang, and Wim Verkruysse. A Model for Waveform Dissimilarities in Dual-Depth Reflectance-PPG. Proceedings of the Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBS), 2018-July:5125-5130, 2018. ISBN: 9781538636466. 8 +[29] Xuesong Niu, Hu Han, Shiguang Shan, and Xin Chen. VIPL-HR: A multi-modal database for pulse estimation from less-constrained face video. In Asian Conference on Computer Vision (ACCV), pages 562-576, 2018. 2, 6 +[30] Xuesong Niu, Shiguang Shan, Hu Han, and Xin Chen. RhythmNet: End-to-end heart rate estimation from face via spatial-temporal representation. IEEE Transactions on Image Processing, 29:2409-2423, 2020. 1, 2 +[31] Ewa Nowara and Daniel McDuff. Combating the Impact of Video Compression on Non-Contact Vital Sign Measurement Using Supervised Learning. In IEEE International Conference on Computer Vision Workshop (ICCVW), pages 1706–1712, Oct. 2019. ISSN: 2473-9944. 5 +[32] Ewa M. Nowara, Daniel McDuff, and Ashok Veeraraghavan. Systematic analysis of video-based pulse measurement from compressed videos. Biomedical Optics Express, 12(1):494, 2021. 3, 5 +[33] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding, 2018. 3 +[34] M. Poh, D. J. McDuff, and R. W. Picard. Advancements in noncontact, multiparameter physiological measurements using a webcam. IEEE Transactions on Biomedical Engineering, 58(1):7-11, 2011. 2, 6 +[35] Ming-Zher Poh, Daniel J. McDuff, and Rosalind W. Picard. Non-contact, automated cardiac pulse measurements using video imaging and blind source separation. Opt. Express, 18(10):10762-10774, May 2010. 2 +[36] Michal Rapczynski, Philipp Werner, and Ayoub Al-Hamadi. Effects of video encoding on camera-based heart rate estimation. IEEE Transactions on Biomedical Engineering, 66(12):3360-3370, 2019. Publisher: IEEE. 5 +[37] Rencheng Song, Huan Chen, Juan Cheng, Chang Li, Yu Liu, and Xun Chen. Pulsegan: Learning to generate realistic pulse waveforms in remote photoplethysmography. IEEE Journal of Biomedical and Health Informatics, 25(5):1373-1384, 2021. 6 +[38] Jeremy Speth, Nathan Vance, Adam Czajka, Kevin Bowyer, and Patrick Flynn. Unifying frame rate and temporal dilations for improved remote pulse detection. Computer Vision and Image Understanding (CVIU), pages 1056-1062, 2021. 1, 2, 3, 5, 6, 8 +[39] Jeremy Speth, Nathan Vance, Adam Czajka, Kevin Bowyer, Diane Wright, and Patrick Flynn. Deception detection and remote physiological monitoring: A dataset and baseline ex + +perimental results. In International Joint Conference on Biometrics (IJCB), pages 4264-4271, 2021. 5 +[40] Radim Špetlík, Vojtech Franc, and Jírí Matas. Visual heart rate estimation with convolutional neural network. In Proceedings of the British machine vision conference, Newcastle, UK, pages 3-6, 2018. 6 +[41] Ronny Stricker, Steffen Muller, and Horst Michael Gross. Non-contact video-based pulse rate measurement on a mobile service robot. IEEE International Symposium on Robot and Human Interactive Communication, pages 1056-1062, 2014. 5 +[42] Zhaodong Sun and Xiaobai Li. Contrast-phys: Unsupervised video-based remote physiological measurement via spatiotemporal contrast. European Conference on Computer Vision (ECCV), 2022. 1, 2, 3, 6, 7 +[43] Nathan Vance, Jeremy Speth, Siamul Khan, Adam Czajka, Kevin W. Bowyer, Diane Wright, and Patrick Flynn. Deception detection and remote physiological monitoring: A dataset and baseline experimental results. IEEE Transactions on Biometrics, Behavior, and Identity Science (TBIOM), pages 1-1, 2022. 5 +[44] Wim Verkruysse, Lars O Svaasand, and J Stuart Nelson. Remote plethysmographic imaging using ambient light. Opt. Express, 16(26):21434-21445, Dec 2008. 6 +[45] Hao Wang, Euijoon Ahn, and Jinman Kim. Self-Supervised Representation Learning Framework for Remote Physiological Measurement Using Spatiotemporal Augmentation Loss. Proceedings of the AAAI Conference on Artificial Intelligence, 36(2):2431-2439, June 2022. 1, 2, 3, 6 +[46] Wenjin Wang, Albertus C. Den Brinker, and Gerard De Haan. Single-Element Remote-PPG. IEEE Transactions on Biomedical Engineering, 66(7):2032-2043, 2019. 2 +[47] W. Wang, A. C. den Brinker, S. Stuijk, and G. de Haan. Algorithmic principles of remote PPG. IEEE Transactions on Biomedical Engineering, 64(7):1479-1491, 2017. 2, 6, 7 +[48] Wenjin Wang, Sander Stuijk, and Gerard De Haan. A Novel Algorithm for Remote Photoplethysmography: Spatial Subspace Rotation. IEEE Transactions on Biomedical Engineering, 63(9):1974-1984, 2016. 6 +[49] Zhen Wang, Yunhao Ba, Pradyumna Chari, Oyku Deniz Bozkurt, Gianna Brown, Parth Patwa, Niranjan Vaddi, Laleh Jalilian, and Achuta Kadambi. Synthetic generation of face videos with plethysmograph physiology. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20555-20564, 2022. 2 +[50] Yuzhe Yang, Xin Liu, Jiang Wu, Silviu Borac, Dina Katabi, Ming-Zher Poh, and Daniel McDuff. Simper: Simple self-supervised learning of periodic targets, 2022. 1, 2, 6 +[51] Zitong Yu, Xiaobai Li, and Guoying Zhao. Remote photoplethysmograph signal measurement from facial videos using spatio-temporal networks. In Proceedings of the British Machine Vision Conference (BMVC), pages 1-12, 2019. 1, 2, 3, 5, 6 +[52] Zitong Yu, Yuming Shen, Jingang Shi, Hengshuang Zhao, Philip H.S. Torr, and Guoying Zhao. Physformer: Facial video-based physiological measurement with temporal difference transformer. In Proceedings of the IEEE/CVF + +Conference on Computer Vision and Pattern Recognition (CVPR), pages 4186-4196, June 2022. 1, 2, 3 +[53] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow Twins: Self-Supervised Learning via Redundancy Reduction. 2021. arXiv: 2103.03230. 2 +[54] Qi Zhan, Wenjin Wang, and Gerard de Haan. Analysis of CNN-based remote-PPG to understand limitations and sensitivities. Biomedical Optics Express, 11(3):1268-1283, 2020. 8 +[55] Yu Zhao, Bochao Zou, Fan Yang, Lin Lu, Abdelkader Nasreddine Belkacem, and Chao Chen. Video-based physiological measurement using 3d central difference convolution attention network. In 2021 IEEE International Joint Conference on Biometrics (IJCB), pages 1-6, 2021. 2 +[56] Hao Zhu, Wayne Wu, Wentao Zhu, Liming Jiang, Siwei Tang, Li Zhang, Ziwei Liu, and Chen Change Loy. CelebV-HQ: A large-scale video facial attributes dataset. In ECCV, 2022. 5, 7 \ No newline at end of file diff --git a/2023/Non-Contrastive Unsupervised Learning of Physiological Signals From Video/images.zip b/2023/Non-Contrastive Unsupervised Learning of Physiological Signals From Video/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f05c1beb695dc72dd3a1e88b6d39662f4b105aa4 --- /dev/null +++ b/2023/Non-Contrastive Unsupervised Learning of Physiological Signals From Video/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3cdee1c10afbfcac5ded7764ae86525d0026b8eacbbba414e326dd7f2b41d8f0 +size 509311 diff --git a/2023/Non-Contrastive Unsupervised Learning of Physiological Signals From Video/layout.json b/2023/Non-Contrastive Unsupervised Learning of Physiological Signals From Video/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c39e0ac0aedc259b95d1d8292de19b8c05cebd3e --- /dev/null +++ b/2023/Non-Contrastive Unsupervised Learning of Physiological Signals From Video/layout.json @@ -0,0 +1,7984 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 61, + 103, + 533, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 103, + 533, + 121 + ], + "spans": [ + { + "bbox": [ + 61, + 103, + 533, + 121 + ], + "type": "text", + "content": "Non-Contrastive Unsupervised Learning of Physiological Signals from Video" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 154, + 143, + 438, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 154, + 143, + 438, + 171 + ], + "spans": [ + { + "bbox": [ + 154, + 143, + 438, + 171 + ], + "type": "text", + "content": "Jeremy Speth, Nathan Vance, Patrick Flynn, Adam Czajka \nUniversity of Notre Dame" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 195, + 174, + 395, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 195, + 174, + 395, + 185 + ], + "spans": [ + { + "bbox": [ + 195, + 174, + 395, + 185 + ], + "type": "text", + "content": "{jspeth,nvancel,flynn,aczajka}@nd.edu" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 238, + 290, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 238, + 290, + 502 + ], + "spans": [ + { + "bbox": [ + 46, + 238, + 290, + 502 + ], + "type": "text", + "content": "Subtle periodic signals such as blood volume pulse and respiration can be extracted from RGB video, enabling noncontact health monitoring at low cost. Advancements in remote pulse estimation - or remote photoplethysmography (rPPG) - are currently driven by deep learning solutions. However, modern approaches are trained and evaluated on benchmark datasets with ground truth from contact-PPG sensors. We present the first non-contrastive unsupervised learning framework for signal regression to mitigate the need for labelled video data. With minimal assumptions of periodicity and finite bandwidth, our approach discovers the blood volume pulse directly from unlabelled videos. We find that encouraging sparse power spectra within normal physiological bandlimits and variance over batches of power spectra is sufficient for learning visual features of periodic signals. We perform the first experiments utilizing unlabelled video data not specifically created for rPPG to train robust pulse rate estimators. Given the limited inductive biases and impressive empirical results, the approach is theoretically capable of discovering other periodic signals from video, enabling multiple physiological measurements without the need for ground truth signals." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 525, + 128, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 525, + 128, + 538 + ], + "spans": [ + { + "bbox": [ + 47, + 525, + 128, + 538 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 545, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 287, + 689 + ], + "type": "text", + "content": "Camera-based vitals estimation is a rapidly growing field enabling non-contact health monitoring in a variety of settings [23]. Although many of the signals avoid detection from the human eye, video data in the visible and infrared ranges contain subtle intensity changes caused by physiological oscillations such as blood volume and respiration. Significant remote photoplethysmography (rPPG) research for estimating the cardiac pulse has leveraged supervised deep learning for robust signal extraction [7, 18, 30, 38, 51, 52]. While the number of successful approaches has rapidly increased, the size of benchmark video datasets with simultaneous vitals recordings has remained relatively stagnant." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": "Robust deep learning-based systems for deployment require training on larger volumes of video data with di" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 213, + 547, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 213, + 547, + 333 + ], + "spans": [ + { + "bbox": [ + 304, + 213, + 547, + 333 + ], + "type": "text", + "content": "verse skin tones, lighting, camera sensors, and movement. However, collecting simultaneous video and physiological ground truth with contact-PPG or electrocardiograms (ECG) is challenging for several reasons. First, many hours of high quality videos is an unwieldy volume of data. Second, recording a diverse subject population in conditions representative of real-world activities is difficult to conduct in the lab setting. Finally, synchronizing contact measurements with video is technically challenging, and even contact measurements used for ground truth contain noise." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 334, + 547, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 334, + 547, + 429 + ], + "spans": [ + { + "bbox": [ + 304, + 334, + 547, + 429 + ], + "type": "text", + "content": "Fortunately, recent works find that contrastive unsupervised learning for rPPG is a promising solution to the data scarcity problem [13, 42, 45, 50]. We extend this research line into non-contrastive unsupervised learning to discover periodic signals in video data. With end-to-end unsupervised learning, collecting more representative training data to learn powerful visual features is much simpler, since only video is required without associated medical information." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 429, + 548, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 429, + 548, + 537 + ], + "spans": [ + { + "bbox": [ + 304, + 429, + 548, + 537 + ], + "type": "text", + "content": "In this work, we show that non-contrastive unsupervised learning is especially simple when regressing rPPG signals. We find weak assumptions of periodicity are sufficient for learning the minuscule visual features corresponding to the blood volume pulse from unlabelled face videos. The loss functions can be computed in the frequency domain over batches without the need for pairwise or triplet comparisons. Figure 1 compares the proposed approach with supervised and contrastive unsupervised learning approaches." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 537, + 547, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 537, + 547, + 585 + ], + "spans": [ + { + "bbox": [ + 304, + 537, + 547, + 585 + ], + "type": "text", + "content": "This work creates opportunities for scaling deep learning models for camera-based vitals and estimating periodic or quasi-periodic signals from unlabelled data beyond rPPG. Our novel contributions are:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 589, + 545, + 685 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 313, + 589, + 545, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 589, + 545, + 625 + ], + "spans": [ + { + "bbox": [ + 313, + 589, + 545, + 625 + ], + "type": "text", + "content": "1. A general framework for physiological signal estimation via non-contrastive unsupervised learning (SiNC) by leveraging periodic signal priors." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 631, + 545, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 631, + 545, + 654 + ], + "spans": [ + { + "bbox": [ + 313, + 631, + 545, + 654 + ], + "type": "text", + "content": "2. The first non-contrastive unsupervised learning method for camera-based vitals measurement." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 661, + 545, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 661, + 545, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 661, + 545, + 685 + ], + "type": "text", + "content": "3. The first experiments and results of training with non-rPPG-specific video data without ground truth vitals." + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 689, + 544, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 544, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 544, + 713 + ], + "type": "text", + "content": "Source code to replicate this work is available at https://github.com/CVRL/SiNC-rPPG." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14464" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 68, + 547, + 250 + ], + "blocks": [ + { + "bbox": [ + 47, + 68, + 547, + 250 + ], + "lines": [ + { + "bbox": [ + 47, + 68, + 547, + 250 + ], + "spans": [ + { + "bbox": [ + 47, + 68, + 547, + 250 + ], + "type": "image", + "image_path": "86b7a5e481e3b4bd3a900a457ea1e7bf1d3c05fff191a7fa1c24b0152a24e020.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 262, + 547, + 308 + ], + "lines": [ + { + "bbox": [ + 46, + 262, + 547, + 308 + ], + "spans": [ + { + "bbox": [ + 46, + 262, + 547, + 308 + ], + "type": "text", + "content": "Figure 1. Overview of the SiNC framework for rPPG compared with traditional supervised and unsupervised learning. Supervised and contrastive losses use distance metrics to the ground truth or other samples. Our framework applies the loss directly to the prediction by shaping the frequency spectrum, and encouraging variance over a batch of inputs. Power outside of the bandlimits is penalized to learn invariances to irrelevant frequencies. Power within the bandlimits is encouraged to be sparsely distributed near the peak frequency." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 327, + 134, + 338 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 327, + 134, + 338 + ], + "spans": [ + { + "bbox": [ + 47, + 327, + 134, + 338 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 346, + 254, + 360 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 346, + 254, + 360 + ], + "spans": [ + { + "bbox": [ + 47, + 346, + 254, + 360 + ], + "type": "text", + "content": "2.1. Remote Photoplethysmography (rPPG)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 365, + 287, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 365, + 287, + 484 + ], + "spans": [ + { + "bbox": [ + 46, + 365, + 287, + 484 + ], + "type": "text", + "content": "Approaches for remote pulse estimation have shifted over the last decade from blind source separation [34, 35], through linear color transformations [9, 10, 46, 47] to training supervised deep learning-based models [7, 16, 18, 20, 29, 30, 38, 51, 52, 55]. While the color transformations generalize well across datasets, deep learning-based models give better accuracy when tested on data from a similar distribution to the training set. To this end, deep learning research has focused on optimizing architectures for extracting robust spatial and temporal features from the limited datasets." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 484, + 288, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 484, + 288, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 484, + 288, + 628 + ], + "type": "text", + "content": "To get around the data bottleneck, large synthetic datasets have recently been proposed [24, 49]. The SCAMPS dataset [24] contains 2,800 videos of synthetic avatars with corresponding PPG, EKG, respiration, and facial action units. The UCLA-synthetic dataset [49] contains 480 videos, which are shown to improve performance when trained with real data. A strength of synthetic datasets is their ability to cover the broad range of skin tones compared to physical data. Another solution to the lack of physiological data is unsupervised learning, where a large set of videos and periodic priors on the output signal is sufficient [13,42,45,50]. We discuss these methods in Sec. 2.3." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 635, + 180, + 648 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 180, + 648 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 180, + 648 + ], + "type": "text", + "content": "2.2. Unsupervised Learning" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": "Self-supervised learning is progressing for image representation learning, with two main competing classes of approaches: contrastive and non-contrastive (or regularized) learning [3]. Contrastive approaches [5, 6, 27] define criteria for distinguishing whether two or more samples are" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 327, + 547, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 327, + 547, + 425 + ], + "spans": [ + { + "bbox": [ + 304, + 327, + 547, + 425 + ], + "type": "text", + "content": "the same or different. Non-contrastive methods augment positive pairs, and enforce variance in the predictions over batches to avoid collapse, in which the model's embeddings reside in a small subspace of the feature space [3]. Distillation methods only use positive samples and avoid collapse by applying a moving average and stop-gradient operator [8, 14]. Another approaches maximize information content of embeddings [11, 53]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 431, + 484, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 431, + 484, + 445 + ], + "spans": [ + { + "bbox": [ + 306, + 431, + 484, + 445 + ], + "type": "text", + "content": "2.3. Unsupervised Learning for rPPG" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 450, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 450, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 547, + 715 + ], + "type": "text", + "content": "All existing unsupervised rPPG approaches are contrastive [13,42,45,50], in which the same model is fed pairs of input videos, and the predictions over similar videos are pulled closer, while the predictions from dissimilar videos are repelled. Gideon et al. [13] were the first to train a deep learning model without labels for rPPG using the contrastive framework. The core of their approach is frequency resampling to create negative samples. Although spatially similar, a resampled video clip contains a different underlying pulse rate, so the model must learn to attend to the temporal dynamics. For their distance function between pairs, they calculated the mean square error between power spectral densities of the model's waveform predictions. While their approach learns to estimate the pulse, their formulation with negative samples is imprecise. The resampled frequency for the negative sample is known, so the relative translation of the power spectrum from the anchor sample can be directly computed. Thus, rather than repelling the estimated spectra, it is more accurate to penalize differences from the known spectra. Furthermore, resampling close to the original sampling rate causes overlap in the power spectra, so repelling the pair is inaccurate. Yuzhe et al. [50]" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14465" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": "incorporated the previously ignored resampling factor for a soft modification to the InfoNCE loss [33] that scales the desired similarity between pairs by their relative sampling rate. A downside is that their learning framework is not end-to-end unsupervised and requires fine-tuning with PPG labels after the self-supervised stage." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 145, + 288, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 145, + 288, + 301 + ], + "spans": [ + { + "bbox": [ + 46, + 145, + 288, + 301 + ], + "type": "text", + "content": "Differently from [13], Contrast-Phys [42] and SLF-RPM [45] consider all samples different from the anchor to be negatives. This assumes that the power spectra will vary between subjects or sufficiently long windows for the same subject. This runs into similar issues with negative pairs as Gideon's approach. Different subjects may have the same pulse rate, so punishing the model for predicting similar frequencies is common during training. Furthermore, the Fast Fourier Transform (FFT) does not produce perfectly sparse decompositions, resulting in spectral overlap even if the heart rate differs by several beats per minute (bpm). As an example, the last column of the second row in Fig. 2 shows the nulls of the main lobe are nearly 30 bpm apart." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 314, + 102, + 326 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 314, + 102, + 326 + ], + "spans": [ + { + "bbox": [ + 47, + 314, + 102, + 326 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 335, + 287, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 335, + 287, + 467 + ], + "spans": [ + { + "bbox": [ + 46, + 335, + 287, + 467 + ], + "type": "text", + "content": "We first formulate signal regression from video. A video sample " + }, + { + "bbox": [ + 46, + 335, + 287, + 467 + ], + "type": "inline_equation", + "content": "x_{i} \\in \\mathbb{R}^{T \\times W \\times H \\times C}" + }, + { + "bbox": [ + 46, + 335, + 287, + 467 + ], + "type": "text", + "content": " sampled from a dataset " + }, + { + "bbox": [ + 46, + 335, + 287, + 467 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 46, + 335, + 287, + 467 + ], + "type": "text", + "content": " consists of " + }, + { + "bbox": [ + 46, + 335, + 287, + 467 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 335, + 287, + 467 + ], + "type": "text", + "content": " images of size " + }, + { + "bbox": [ + 46, + 335, + 287, + 467 + ], + "type": "inline_equation", + "content": "W \\times H" + }, + { + "bbox": [ + 46, + 335, + 287, + 467 + ], + "type": "text", + "content": " pixels across " + }, + { + "bbox": [ + 46, + 335, + 287, + 467 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 335, + 287, + 467 + ], + "type": "text", + "content": " channels, captured over time. State-of-the-art methods offer models " + }, + { + "bbox": [ + 46, + 335, + 287, + 467 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 335, + 287, + 467 + ], + "type": "text", + "content": " that regress a waveform " + }, + { + "bbox": [ + 46, + 335, + 287, + 467 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^T \\ni y_i = f(x_i)" + }, + { + "bbox": [ + 46, + 335, + 287, + 467 + ], + "type": "text", + "content": " of the same length as the video. Recently, the task has been effectively modeled end-to-end with the models " + }, + { + "bbox": [ + 46, + 335, + 287, + 467 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 335, + 287, + 467 + ], + "type": "text", + "content": " being spatiotemporal neural networks [16, 21, 38, 51, 52]. While most previous works are supervised and minimize the loss to a contact pulse measurement, we perform non-contrastive learning using only the model's estimated waveform." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 468, + 287, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 468, + 287, + 600 + ], + "spans": [ + { + "bbox": [ + 46, + 468, + 287, + 600 + ], + "type": "text", + "content": "The key realization is that we can place strong priors on the estimated pulse regarding its bandwidth and periodicity. Observed signals outside the desired frequency range are pollutants, so penalizing the model for carrying them through the forward pass results in invariances to such noisy visual features. We find that it is surprisingly easy to impose the desired constraints in the frequency domain. Thus, all waveforms are transformed into their discrete Fourier components with the FFT before computing all losses in our approach, " + }, + { + "bbox": [ + 46, + 468, + 287, + 600 + ], + "type": "inline_equation", + "content": "F = \\mathrm{FFT}(y)" + }, + { + "bbox": [ + 46, + 468, + 287, + 600 + ], + "type": "text", + "content": ". The following sections describe the loss functions and augmentations used during training." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 610, + 100, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 610, + 100, + 621 + ], + "spans": [ + { + "bbox": [ + 47, + 610, + 100, + 621 + ], + "type": "text", + "content": "3.1. Losses" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "content": "One of the advantages of unsupervised learning for periodic signals is that we can constrain the solution space significantly. For physiological signals such as respiration and blood volume pulse, we know the healthy upper and lower bounds of the frequencies. We also desire the extracted signal to be sparse in the frequency domain, and that our model filters out noise signals present in the video. With these" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "content": "constraints, we can greatly simplify the problem of finding good features for the desired signal in the data." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 96, + 545, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 96, + 545, + 192 + ], + "spans": [ + { + "bbox": [ + 304, + 96, + 545, + 192 + ], + "type": "text", + "content": "Bandwidth Loss. One of the most powerful constraints we can place on the model is frequency bandlimits. Past unsupervised methods have used the irrelevant power ratio (IPR) as a validation metric [12, 13, 42] for model selection. We find that it is also effective during model training. The IPR penalizes the model for generating signals outside the desired bandlimits. With lower and upper bandlimits of " + }, + { + "bbox": [ + 304, + 96, + 545, + 192 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 304, + 96, + 545, + 192 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 96, + 545, + 192 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 304, + 96, + 545, + 192 + ], + "type": "text", + "content": ", respectively, our bandwidth loss becomes:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 348, + 200, + 545, + 243 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 200, + 545, + 243 + ], + "spans": [ + { + "bbox": [ + 348, + 200, + 545, + 243 + ], + "type": "interline_equation", + "content": "L _ {b} = \\frac {1}{\\sum_ {i = - \\infty} ^ {\\infty} F _ {i}} \\left[ \\sum_ {i = - \\infty} ^ {a} F _ {i} + \\sum_ {i = b} ^ {\\infty} F _ {i} \\right], \\tag {1}", + "image_path": "5b40dd25b617135116b92a4371c323314b5c64ba953d193d3d3a79b70ed082a0.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 247, + 545, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 247, + 545, + 366 + ], + "spans": [ + { + "bbox": [ + 304, + 247, + 545, + 366 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 247, + 545, + 366 + ], + "type": "inline_equation", + "content": "F_{i}" + }, + { + "bbox": [ + 304, + 247, + 545, + 366 + ], + "type": "text", + "content": " is the power in the " + }, + { + "bbox": [ + 304, + 247, + 545, + 366 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 247, + 545, + 366 + ], + "type": "text", + "content": "th frequency bin of the predicted signal. This simple loss enforces learning of many invariants, such as movement from respiration, talking, or facial expressions which typically occupy low frequencies. In our experiments we specify the limits as " + }, + { + "bbox": [ + 304, + 247, + 545, + 366 + ], + "type": "inline_equation", + "content": "a = 0.66\\mathrm{Hz}" + }, + { + "bbox": [ + 304, + 247, + 545, + 366 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 247, + 545, + 366 + ], + "type": "inline_equation", + "content": "b = 3\\mathrm{Hz}" + }, + { + "bbox": [ + 304, + 247, + 545, + 366 + ], + "type": "text", + "content": ", which corresponds to a common pulse rate range from 40 bpm to 180 bpm. The first column of Fig. 2 shows the result of training exclusively with the bandwidth loss " + }, + { + "bbox": [ + 304, + 247, + 545, + 366 + ], + "type": "inline_equation", + "content": "L_{b}" + }, + { + "bbox": [ + 304, + 247, + 545, + 366 + ], + "type": "text", + "content": ". The last row shows that the model concentrates signal power between the bandlimits." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 367, + 545, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 367, + 545, + 439 + ], + "spans": [ + { + "bbox": [ + 304, + 367, + 545, + 439 + ], + "type": "text", + "content": "Sparsity Loss. The pulse rate is the most common physiological marker associated with the blood volume pulse. Since we are primarily interested in the frequency, we can further improve our model by preventing wideband predictions. This also reveals the true signal we aim to discover by ignoring visual dynamics that are not strongly periodic." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 439, + 545, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 439, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 305, + 439, + 545, + 464 + ], + "type": "text", + "content": "We penalize energy within the bandlimits that is not near the spectral peak:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 339, + 472, + 545, + 517 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 472, + 545, + 517 + ], + "spans": [ + { + "bbox": [ + 339, + 472, + 545, + 517 + ], + "type": "interline_equation", + "content": "L _ {s} = \\frac {1}{\\sum_ {i = a} ^ {b} F _ {i}} \\left[ \\sum_ {i = a} ^ {F ^ {*} - \\Delta_ {F}} F _ {i} + \\sum_ {i = F ^ {*} + \\Delta_ {F}} ^ {b} F _ {i} \\right], \\tag {2}", + "image_path": "e00197f37d2c8eebf518392794da5e80d3cfbe23e8a7c05f1b3a46e276b376fa.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 521, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 605 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 521, + 545, + 605 + ], + "type": "inline_equation", + "content": "F^{*} = \\operatorname{argmax}(F)" + }, + { + "bbox": [ + 304, + 521, + 545, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 521, + 545, + 605 + ], + "type": "inline_equation", + "content": "\\Delta_F" + }, + { + "bbox": [ + 304, + 521, + 545, + 605 + ], + "type": "text", + "content": " are the frequencies of the spectral peak and padding around the peak, respectively. For all experiments " + }, + { + "bbox": [ + 304, + 521, + 545, + 605 + ], + "type": "inline_equation", + "content": "\\Delta_F = 6" + }, + { + "bbox": [ + 304, + 521, + 545, + 605 + ], + "type": "text", + "content": " beats per minute [32]. Figure 2 shows the result of training only with the sparsity loss in the second column. For the whole dataset, the power spectrum is sparsely distributed in the low frequencies, effectively filtering frequencies higher than " + }, + { + "bbox": [ + 304, + 521, + 545, + 605 + ], + "type": "inline_equation", + "content": "1\\mathrm{Hz}" + }, + { + "bbox": [ + 304, + 521, + 545, + 605 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 606, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 606, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 606, + 545, + 713 + ], + "type": "text", + "content": "Variance Loss. One of the risks of non-contrastive methods is the model collapsing into trivial solutions and making predictions independently of the input features. In regularized methods such as VICReg [3], a hinge loss on the variance over a batch of predictions is used to enforce diverse outputs. We use a similar strategy to avoid model collapse, but instead spread the variance in power spectral densities towards a uniform distribution over the desired frequency band. Our variance loss processes a uniform prior" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14466" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 69, + 309, + 183 + ], + "blocks": [ + { + "bbox": [ + 50, + 69, + 309, + 183 + ], + "lines": [ + { + "bbox": [ + 50, + 69, + 309, + 183 + ], + "spans": [ + { + "bbox": [ + 50, + 69, + 309, + 183 + ], + "type": "image", + "image_path": "73f0f9e58fa4278de59a7cde858e22205745b33283d03d17bac423c199cc6949.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 310, + 70, + 426, + 182 + ], + "blocks": [ + { + "bbox": [ + 310, + 70, + 426, + 182 + ], + "lines": [ + { + "bbox": [ + 310, + 70, + 426, + 182 + ], + "spans": [ + { + "bbox": [ + 310, + 70, + 426, + 182 + ], + "type": "image", + "image_path": "a1027376b11c005af2d90c484d98ba8068fba2129028fd19b37ae6c00ba91c3d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 427, + 70, + 545, + 182 + ], + "blocks": [ + { + "bbox": [ + 427, + 70, + 545, + 182 + ], + "lines": [ + { + "bbox": [ + 427, + 70, + 545, + 182 + ], + "spans": [ + { + "bbox": [ + 427, + 70, + 545, + 182 + ], + "type": "image", + "image_path": "d90f10d35c3bfa89e4e9233727e03315dc60db5cd2ec38029fe33a0228e31ce3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 49, + 185, + 309, + 289 + ], + "blocks": [ + { + "bbox": [ + 49, + 185, + 309, + 289 + ], + "lines": [ + { + "bbox": [ + 49, + 185, + 309, + 289 + ], + "spans": [ + { + "bbox": [ + 49, + 185, + 309, + 289 + ], + "type": "image", + "image_path": "5d18cba2f2c08ba61a85737653aa36e1f6cf9ecdb0dadefbab6505365b097ade.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 310, + 185, + 426, + 289 + ], + "blocks": [ + { + "bbox": [ + 310, + 185, + 426, + 289 + ], + "lines": [ + { + "bbox": [ + 310, + 185, + 426, + 289 + ], + "spans": [ + { + "bbox": [ + 310, + 185, + 426, + 289 + ], + "type": "image", + "image_path": "c788f5df5c793649808495d38e68eba7be9a09354a4e016ad5d362e993e14c85.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 427, + 185, + 544, + 289 + ], + "blocks": [ + { + "bbox": [ + 427, + 185, + 544, + 289 + ], + "lines": [ + { + "bbox": [ + 427, + 185, + 544, + 289 + ], + "spans": [ + { + "bbox": [ + 427, + 185, + 544, + 289 + ], + "type": "image", + "image_path": "c0ff9ba078c287422e89c05207b730daddd53f87b305be09b17a175c76c59c3a.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 50, + 290, + 309, + 396 + ], + "blocks": [ + { + "bbox": [ + 50, + 290, + 309, + 396 + ], + "lines": [ + { + "bbox": [ + 50, + 290, + 309, + 396 + ], + "spans": [ + { + "bbox": [ + 50, + 290, + 309, + 396 + ], + "type": "image", + "image_path": "dbb4c7c8a43c9aac763e5d37eecc40de405f8402d698cc8a554b1abc44ce62e0.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 239, + 396, + 374, + 409 + ], + "lines": [ + { + "bbox": [ + 239, + 396, + 374, + 409 + ], + "spans": [ + { + "bbox": [ + 239, + 396, + 374, + 409 + ], + "type": "text", + "content": "Predicted Ground Truth" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 46, + 420, + 546, + 487 + ], + "lines": [ + { + "bbox": [ + 46, + 420, + 546, + 487 + ], + "spans": [ + { + "bbox": [ + 46, + 420, + 546, + 487 + ], + "type": "text", + "content": "Figure 2. Each column shows predictions from models trained with one or all of the losses for 20 epochs on UBFC-rPPG. The first two rows show a sample in the time and frequency domain, respectively. The last row shows the signal power over the validation set computed by taking the sum of normalized power spectral densities from each sample, then dividing the result by the number of validation samples. The bandwidth loss penalizes signal power outside predefined bandlimits (40 to 180 bpm) to constrain the output space. The sparsity loss encourages a narrow spectrum containing strong periodicity. The variance loss encourages diverse power spectra in a batch, preventing the model from collapsing to a narrow bandwidth. When combined, the model estimates periodic signals within the desired bandlimits." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 310, + 291, + 426, + 396 + ], + "blocks": [ + { + "bbox": [ + 310, + 291, + 426, + 396 + ], + "lines": [ + { + "bbox": [ + 310, + 291, + 426, + 396 + ], + "spans": [ + { + "bbox": [ + 310, + 291, + 426, + 396 + ], + "type": "image", + "image_path": "9414fa58b7d7d7b0713aa77f30a39dd06842029cbd70f07eb3717f3f7a1955d6.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 427, + 291, + 544, + 396 + ], + "blocks": [ + { + "bbox": [ + 427, + 291, + 544, + 396 + ], + "lines": [ + { + "bbox": [ + 427, + 291, + 544, + 396 + ], + "spans": [ + { + "bbox": [ + 427, + 291, + 544, + 396 + ], + "type": "image", + "image_path": "e79366bfc32a63af994c89c0ae115e2d0b9ac9d76a25c2ed3b374c48effc9dd9.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 508, + 287, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 508, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 508, + 287, + 581 + ], + "type": "text", + "content": "distribution " + }, + { + "bbox": [ + 46, + 508, + 287, + 581 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 508, + 287, + 581 + ], + "type": "text", + "content": " over " + }, + { + "bbox": [ + 46, + 508, + 287, + 581 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 508, + 287, + 581 + ], + "type": "text", + "content": " frequencies, and a batch of " + }, + { + "bbox": [ + 46, + 508, + 287, + 581 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 508, + 287, + 581 + ], + "type": "text", + "content": " spectral densities, " + }, + { + "bbox": [ + 46, + 508, + 287, + 581 + ], + "type": "inline_equation", + "content": "F = [v_{1},\\dots,v_{n}]" + }, + { + "bbox": [ + 46, + 508, + 287, + 581 + ], + "type": "text", + "content": ", where each vector is a " + }, + { + "bbox": [ + 46, + 508, + 287, + 581 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 508, + 287, + 581 + ], + "type": "text", + "content": "-dimensional frequency decomposition of a predicted waveform. We calculate the normalized sum of densities over the batch, " + }, + { + "bbox": [ + 46, + 508, + 287, + 581 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 46, + 508, + 287, + 581 + ], + "type": "text", + "content": ", and define the variance loss as the squared Wasserstein distance [15] to the uniform prior:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 85, + 590, + 287, + 624 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 590, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 85, + 590, + 287, + 624 + ], + "type": "interline_equation", + "content": "L _ {v} = \\frac {1}{d} \\sum_ {i = 1} ^ {d} \\left(\\mathrm {C D F} _ {i} (Q) - \\mathrm {C D F} _ {i} (P)\\right) ^ {2}, \\tag {3}", + "image_path": "9eff7347c750f5ff397fafe9f7dbe53074db83bb7e2a66ac033f7c57c3fa4570.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 629, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 714 + ], + "type": "text", + "content": "where CDF is a cumulative distribution function. The third column of Fig. 2 shows the effect of the variance loss during training. For a single sample, wide-band signals containing multiple frequencies are predicted, and the predicted frequencies cover the task's bandwidth. In our experiments we use a batch size of 20 samples. See the suppl. materials for an ablation experiment on the impact of batch size." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 508, + 544, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 508, + 544, + 533 + ], + "spans": [ + { + "bbox": [ + 306, + 508, + 544, + 533 + ], + "type": "text", + "content": "Combining All Losses. Summarizing, our training loss function is a simple sum of the aforementioned losses:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 384, + 557, + 545, + 569 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 557, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 384, + 557, + 545, + 569 + ], + "type": "interline_equation", + "content": "L = L _ {b} + L _ {s} + L _ {v}. \\tag {4}", + "image_path": "0fa7474513595ff7dcf5dfdf83d5b8035540e33a46faafa4265390a7aef8a83c.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 546, + 714 + ], + "type": "text", + "content": "While one could weight particular components of the loss more than others, we specifically formulated the losses to scale them between 0 and 1. In our experiments, we find that a simple summation without weighting gives good performance. The combined loss function encourages the model to search over the supported frequencies to discover visual features for a strong periodic signal. Remarkably, we find that this simple framework is sufficient for learning to regress the blood volume in video, as shown in the last column of Fig. 2." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14467" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 141, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 141, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 141, + 85 + ], + "type": "text", + "content": "3.2. Augmentations" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 89, + 287, + 162 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 89, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 47, + 89, + 287, + 162 + ], + "type": "text", + "content": "Unlike Gideon et al.'s [13] approach, which only applies frequency augmentations, we apply several augmentations to both the spatial and temporal dimensions to learn invariances to noisy visual signals. In fact, we found that without augmentations, models did not converge during training (see the supplementary material)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 163, + 287, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 163, + 287, + 221 + ], + "spans": [ + { + "bbox": [ + 47, + 163, + 287, + 221 + ], + "type": "text", + "content": "Image Intensity Augmentations. Gaussian noise is added to each pixel with zero mean and a standard deviation of 2 on an image scale from 0 to 255. Each clip is darkened or brightened by adding a constant from a Gaussian distribution with mean 0 and standard deviation of 10." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 221, + 287, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 221, + 287, + 281 + ], + "spans": [ + { + "bbox": [ + 47, + 221, + 287, + 281 + ], + "type": "text", + "content": "Spatial Augmentations. We randomly horizontally flip a video clip with " + }, + { + "bbox": [ + 47, + 221, + 287, + 281 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 47, + 221, + 287, + 281 + ], + "type": "text", + "content": " probability. The spatial dimension of a clip are randomly square cropped down to between half the original length and the original length. The cropped clip is then linearly interpolated back to the original dimensions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 281, + 287, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 281, + 287, + 353 + ], + "spans": [ + { + "bbox": [ + 47, + 281, + 287, + 353 + ], + "type": "text", + "content": "Temporal Augmentations. With the general assumption that the desired signal is strongly periodic and sparsely represented in the Fourier domain, we randomly flip a video clip along the time dimension with a probability of " + }, + { + "bbox": [ + 47, + 281, + 287, + 353 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 47, + 281, + 287, + 353 + ], + "type": "text", + "content": ". Note that the Fourier decomposition of a time-reversed sinusoid is identical to that of the original sinusoid." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 354, + 287, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 354, + 287, + 473 + ], + "spans": [ + { + "bbox": [ + 47, + 354, + 287, + 473 + ], + "type": "text", + "content": "Frequency Augmentations. Perhaps the most important augmentation is frequency resampling [13], where the video is linearly interpolated to a different frame rate. This augmentation is particularly interesting for rPPG, because it transforms the video input and target signal equivalently along the time dimension, making it equivariant. Given the aforementioned transformations that are invariant, " + }, + { + "bbox": [ + 47, + 354, + 287, + 473 + ], + "type": "inline_equation", + "content": "\\tau (\\cdot)\\sim" + }, + { + "bbox": [ + 47, + 354, + 287, + 473 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 47, + 354, + 287, + 473 + ], + "type": "text", + "content": " , the equivariant frequency resampling operation, " + }, + { + "bbox": [ + 47, + 354, + 287, + 473 + ], + "type": "inline_equation", + "content": "\\phi (\\cdot)\\sim" + }, + { + "bbox": [ + 47, + 354, + 287, + 473 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 47, + 354, + 287, + 473 + ], + "type": "text", + "content": " , and a model " + }, + { + "bbox": [ + 47, + 354, + 287, + 473 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 47, + 354, + 287, + 473 + ], + "type": "text", + "content": " that infers a waveform from a video we have the following:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 111, + 482, + 287, + 496 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 482, + 287, + 496 + ], + "spans": [ + { + "bbox": [ + 111, + 482, + 287, + 496 + ], + "type": "interline_equation", + "content": "\\phi (f (\\tau (x))) = f (\\phi (\\tau (x))). \\tag {5}", + "image_path": "38730c91ff5a06177c55c3ce63bbee6138a3050401cad5c7721099f9584f35e7.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 503, + 287, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 503, + 287, + 588 + ], + "spans": [ + { + "bbox": [ + 47, + 503, + 287, + 588 + ], + "type": "text", + "content": "This is a powerful augmentation, because it allows us to augment the target distribution along with the video input. In our experiments we randomly resample input clips by a factor " + }, + { + "bbox": [ + 47, + 503, + 287, + 588 + ], + "type": "inline_equation", + "content": "c \\sim U(0.6, 1.4)" + }, + { + "bbox": [ + 47, + 503, + 287, + 588 + ], + "type": "text", + "content": ". After applying the resampling augmentation, we scale the bandlimits by " + }, + { + "bbox": [ + 47, + 503, + 287, + 588 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 47, + 503, + 287, + 588 + ], + "type": "text", + "content": ", to avoid penalizing the model if the augmentation pushed the underlying pulse frequency outside of the original bandlimits." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 597, + 105, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 597, + 105, + 609 + ], + "spans": [ + { + "bbox": [ + 47, + 597, + 105, + 609 + ], + "type": "text", + "content": "4. Datasets" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 617, + 287, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 617, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 47, + 617, + 287, + 666 + ], + "type": "text", + "content": "We use PURE [41], UBFC-rPPG [4], and DDPM [39] as benchmark rPPG datasets for training and testing, and CelebV-HQ dataset [56] and HKBU-MARs [17] for unsupervised training only." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 666, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 666, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 666, + 287, + 713 + ], + "type": "text", + "content": "Deception Detection and Physiological Monitoring (DDPM) [39, 43] consists of data from 86 subjects attempting to answer questions deceptively. Interviews were recorded at 90 frames-per-second for more than 10 minutes" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "content": "on average. Natural conversation and head pose changes make it a difficult and less-constrained rPPG dataset." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 96, + 545, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 96, + 545, + 180 + ], + "spans": [ + { + "bbox": [ + 305, + 96, + 545, + 180 + ], + "type": "text", + "content": "PURE [41] is a benchmark rPPG dataset consisting of 10 subjects recorded over 6 sessions. Each session lasted approximately 1 minute, and raw video was recorded at 30 fps. The 6 sessions for each subject consisted of: (1) steady, (2) talking, (3) slow head translation, (4) fast head translation, (5) small and (6) medium head rotations. Pulse rates are at or close to the subject's resting rate." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 180, + 545, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 180, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 305, + 180, + 545, + 228 + ], + "type": "text", + "content": "UBFC-rPPG [4] contains 1-minute long videos from 42 subjects recorded at 30 fps. Subjects played a time-sensitive mathematical game to raise their heart rates, but head motion is limited during the recording." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 228, + 545, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 228, + 545, + 324 + ], + "spans": [ + { + "bbox": [ + 305, + 228, + 545, + 324 + ], + "type": "text", + "content": "HKBU 3D Mask Attack with Real World Variations (HKBU-MARs) [17] consists of 12 subjects captured over 6 different lighting configurations with 7 different cameras each, resulting in 504 videos lasting 10 seconds each. The diverse lighting and camera sensors make it a valuable dataset for unsupervised training. We use version 2 of HKBU-MARs, which contains videos with both realistic 3D masks and unmasked subjects." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 324, + 545, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 324, + 545, + 433 + ], + "spans": [ + { + "bbox": [ + 305, + 324, + 545, + 433 + ], + "type": "text", + "content": "High-Quality Celebrity Video Dataset (CelebV-HQ) [56] is a set of processed YouTube videos containing 35,666 face videos from over 15,000 identities. The videos vary dramatically in length, lighting, emotion, motion, skin tones, and camera sensors. The greatest challenge in harnessing online videos is their reduced quality due to compression before upload and by the video provider. Compression is a known challenge for rPPG, since the blood volume pulse is so subtle optically [25, 31, 32, 36]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 444, + 405, + 458 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 444, + 405, + 458 + ], + "spans": [ + { + "bbox": [ + 306, + 444, + 405, + 458 + ], + "type": "text", + "content": "5. Training Details" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 464, + 420, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 464, + 420, + 477 + ], + "spans": [ + { + "bbox": [ + 306, + 464, + 420, + 477 + ], + "type": "text", + "content": "5.1. Data Preprocessing" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 483, + 545, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 483, + 545, + 638 + ], + "spans": [ + { + "bbox": [ + 305, + 483, + 545, + 638 + ], + "type": "text", + "content": "To prepare the video clips for the spatiotemporal deep learning models, we first extract 68 face landmarks with OpenFace [2]. We then define a bounding box in each frame with the minimum and maximum " + }, + { + "bbox": [ + 305, + 483, + 545, + 638 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 305, + 483, + 545, + 638 + ], + "type": "text", + "content": " locations by extending the crop horizontally by " + }, + { + "bbox": [ + 305, + 483, + 545, + 638 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 305, + 483, + 545, + 638 + ], + "type": "text", + "content": " to ensure that the cheeks and jaw are present. The top and bottom are extended by " + }, + { + "bbox": [ + 305, + 483, + 545, + 638 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 305, + 483, + 545, + 638 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 483, + 545, + 638 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 305, + 483, + 545, + 638 + ], + "type": "text", + "content": " of the bounding box height, respectively, to include the forehead and jaw. We further extend the shorter of the two axes to the length of the other to form a square. The cropped frames are then resized to " + }, + { + "bbox": [ + 305, + 483, + 545, + 638 + ], + "type": "inline_equation", + "content": "64\\times 64" + }, + { + "bbox": [ + 305, + 483, + 545, + 638 + ], + "type": "text", + "content": " pixels with bicubic interpolation. For faster processing of the massive CelebV-HQ [56] dataset, we instead use MediaPipe Face Mesh [22] for landmarking." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 647, + 424, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 647, + 424, + 658 + ], + "spans": [ + { + "bbox": [ + 306, + 647, + 424, + 658 + ], + "type": "text", + "content": "5.2. Model Architectures" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 665, + 545, + 713 + ], + "type": "text", + "content": "We use a 3D-CNN architecture similar to [38] without temporal dilations [51]. We use a temporal kernel width of 5, and replace default zero-padding by repeating the edges. Zero-padding along the time dimension can result" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14468" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 112, + 547, + 312 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "type": "text", + "content": "Table 1. Intra-dataset pulse rate estimation results. The best results for a given metric/dataset are bolded, and the second-best results are underlined. For a better comparison with results in the literature, this table follows [42]. MAE: Mean Absolute Error; RMSE: Root Mean Square Error; " + }, + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "type": "text", + "content": " : Pearson correlation coefficient." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 112, + 547, + 312 + ], + "lines": [ + { + "bbox": [ + 47, + 112, + 547, + 312 + ], + "spans": [ + { + "bbox": [ + 47, + 112, + 547, + 312 + ], + "type": "table", + "html": "
TypesMethodsUBFC-rPPGPUREDDPM
MAE (bpm)RMSE (bpm)rMAE (bpm)RMSE (bpm)rMAE (bpm)RMSE (bpm)r
TraditionalGREEN [44]7.5014.410.627.2317.050.6932.7943.090.04
ICA [34]5.1711.760.653.7612.600.8522.2235.770.40
CHROM [9]2.369.230.870.752.231.0013.4828.530.56
2SR [48]6.9018.500.652.443.060.9822.0839.730.22
POS [47]2.119.110.870.804.110.989.0323.070.70
SupervisedHR-CNN [40]---1.842.370.98---
SynRhythm [29]5.596.820.72------
PulseGAN [37]1.192.100.98------
Dual-GAN [21]0.440.670.990.821.310.99---
RPNet [38]†0.53 ± 0.011.78 ± 0.020.99 ± 0.001.15 ± 0.275.77 ± 1.250.96 ± 0.023.46 ± 0.2412.47 ± 0.680.91 ± 0.01
PhysNet [51]†0.55 ± 0.032.03 ± 0.370.99 ± 0.000.99 ± 0.195.22 ± 0.930.97 ± 0.013.96 ± 0.7613.57 ± 1.740.89 ± 0.03
UnsupervisedGideon2021 [13]1.854.280.932.32.90.99---
SLF-RPM [45]*8.399.700.70------
SimPer [50]*4.24--3.89-----
Contrast-Phys [42]0.641.000.991.001.400.999.70 ± 2.9025.02 ± 6.010.58 ± 0.19
SiNC (ours)0.59 ± 0.001.83 ± 0.040.99 ± 0.000.61 ± 0.061.84 ± 0.401.00 ± 0.005.87 ± 0.1117.44 ± 0.160.81 ± 0.00
", + "image_path": "ea9c8daeed9482427269614a2f864cdeb54a3a3e5645696a10b47eee0c218207.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 52, + 313, + 389, + 323 + ], + "lines": [ + { + "bbox": [ + 52, + 313, + 389, + 323 + ], + "spans": [ + { + "bbox": [ + 52, + 313, + 389, + 323 + ], + "type": "text", + "content": "* Some methods in the unsupervised row require fine-tuning on labeled data with a linear classifier." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_footnote" + }, + { + "bbox": [ + 52, + 322, + 389, + 333 + ], + "lines": [ + { + "bbox": [ + 52, + 322, + 389, + 333 + ], + "spans": [ + { + "bbox": [ + 52, + 322, + 389, + 333 + ], + "type": "text", + "content": "† Some supervised methods were trained with identical data augmentations to SiNC for fair comparison." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 354, + 289, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 354, + 289, + 521 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 289, + 521 + ], + "type": "text", + "content": "in edge effects that add artificial frequencies to the predictions. Early experiments showed that temporal dilations caused aliasing and reduced the bandwidth of the model to specific frequencies. Our losses and framework may be applied to any task and architecture with dense predictions along one or more dimensions. However, popular rPPG architectures such as DeepPhys [7] and MTTS-CAN [18] may be ill-suited for the approach, since they consume very few frames, and the number of time points should be large enough to give sufficient frequency resolution with the FFT. In our experiments, we use the AdamW [19] optimizer with a learning rate of 0.0001. We use a clip length of " + }, + { + "bbox": [ + 46, + 354, + 289, + 521 + ], + "type": "inline_equation", + "content": "T = 120" + }, + { + "bbox": [ + 46, + 354, + 289, + 521 + ], + "type": "text", + "content": " frames (4 seconds), and we set the input signal's length to achieve a frequency resolution of 0.33 bpm." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 527, + 167, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 527, + 167, + 541 + ], + "spans": [ + { + "bbox": [ + 47, + 527, + 167, + 541 + ], + "type": "text", + "content": "5.3. Supervised Training" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 545, + 288, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 288, + 654 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 288, + 654 + ], + "type": "text", + "content": "To properly compare our approach to its supervised counterpart we use the same model architecture and train it with the commonly used negative Pearson loss between the predicted waveform and the contact sensor ground truth [51]. During training we apply all of the same augmentations except time reversal. Models are trained for 200 epochs on PURE and UBFC-rPPG, and for 40 epochs on DDPM. The model from the epoch with the lowest loss on the validation set is selected for testing." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 659, + 178, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 659, + 178, + 673 + ], + "spans": [ + { + "bbox": [ + 47, + 659, + 178, + 673 + ], + "type": "text", + "content": "5.4. Unsupervised Training" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": "Unsupervised models are trained for the same number of epochs as the supervised setting for both PURE and UBFCrPPG, but we train for an additional 40 epochs on DDPM," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 354, + 545, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 354, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 545, + 426 + ], + "type": "text", + "content": "since this dataset is considerably more difficult. We set the batch size to 20 samples during training. Contrary to previous unsupervised approaches [13, 42], we leverage validation sets for model selection by selecting the model with the lowest combined bandwidth and sparsity losses. The creation of the dataset splits is described in the next section." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 431, + 380, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 431, + 380, + 443 + ], + "spans": [ + { + "bbox": [ + 306, + 431, + 380, + 443 + ], + "type": "text", + "content": "5.5. Evaluation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 450, + 545, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 450, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 545, + 533 + ], + "type": "text", + "content": "Pulse rates are computed as the highest spectral peak between " + }, + { + "bbox": [ + 304, + 450, + 545, + 533 + ], + "type": "inline_equation", + "content": "0.66\\mathrm{Hz}" + }, + { + "bbox": [ + 304, + 450, + 545, + 533 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 450, + 545, + 533 + ], + "type": "inline_equation", + "content": "3\\mathrm{Hz}" + }, + { + "bbox": [ + 304, + 450, + 545, + 533 + ], + "type": "text", + "content": " (equivalent to " + }, + { + "bbox": [ + 304, + 450, + 545, + 533 + ], + "type": "inline_equation", + "content": "40\\mathrm{bpm}" + }, + { + "bbox": [ + 304, + 450, + 545, + 533 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 450, + 545, + 533 + ], + "type": "inline_equation", + "content": "180\\mathrm{bpm}" + }, + { + "bbox": [ + 304, + 450, + 545, + 533 + ], + "type": "text", + "content": ") over a 10-second sliding window. The same procedure is applied to the ground truth waveforms for a reliable evaluation [26]. We apply common error metrics such as mean absolute error (MAE), root mean square error (RMSE), and Pearson correlation coefficient " + }, + { + "bbox": [ + 304, + 450, + 545, + 533 + ], + "type": "inline_equation", + "content": "(r)" + }, + { + "bbox": [ + 304, + 450, + 545, + 533 + ], + "type": "text", + "content": " between the pulse rates." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 534, + 545, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 545, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 545, + 641 + ], + "type": "text", + "content": "We perform 5-fold cross validation for both PURE and UBFC with the same folds as [13], and use the predefined dataset splits from DDPM [38]. Differently from [13], we use 3 of the folds for training, 1 for validation, and the remaining for testing rather than only training and testing partitions. We train 3 models with different initializations, resulting in 15 models trained on PURE and UBFC each, and 3 models trained on DDPM. We present the mean and standard deviation of the errors in the results." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 651, + 359, + 663 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 651, + 359, + 663 + ], + "spans": [ + { + "bbox": [ + 306, + 651, + 359, + 663 + ], + "type": "text", + "content": "6. Results" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 671, + 437, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 671, + 437, + 684 + ], + "spans": [ + { + "bbox": [ + 306, + 671, + 437, + 684 + ], + "type": "text", + "content": "6.1. Within-Dataset Testing" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "content": "Table 1 shows the results for models trained and tested on subject-disjoint partitions from the same datasets. For" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14469" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 156 + ], + "type": "text", + "content": "PURE and UBFC we achieve MAE lower than 1 bpm, performing better or on par with all traditional and supervised learning approaches. For PURE, our approach gives the lowest MAE and a Pearson " + }, + { + "bbox": [ + 46, + 72, + 286, + 156 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 46, + 72, + 286, + 156 + ], + "type": "text", + "content": " of nearly 1. Performance drops on DDPM due to the overall difficulty of the dataset. SiNC outperforms contrastive approaches, only being surpassed by supervised deep learning models." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 157, + 286, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 157, + 286, + 240 + ], + "spans": [ + { + "bbox": [ + 46, + 157, + 286, + 240 + ], + "type": "text", + "content": "In comparison to other unsupervised methods, Contrast-Phys [42] gives the most competitive performance on all but DDPM. Note that our approach gives the lowest MAE on all datasets, but has higher RMSE. We believe this is due their use of harmonic removal as a post-processing step when estimating the pulse rate, which is not described in [42], but can be found in their publicly available code." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 247, + 172, + 260 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 247, + 172, + 260 + ], + "spans": [ + { + "bbox": [ + 47, + 247, + 172, + 260 + ], + "type": "text", + "content": "6.2. Cross-Dataset Testing" + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 48, + 352, + 291, + 605 + ], + "blocks": [ + { + "bbox": [ + 46, + 277, + 286, + 342 + ], + "lines": [ + { + "bbox": [ + 46, + 277, + 286, + 342 + ], + "spans": [ + { + "bbox": [ + 46, + 277, + 286, + 342 + ], + "type": "text", + "content": "Table 2. Cross-dataset pulse rate estimation performance. The top 3 training datasets are common rPPG benchmarks, while HKBU was not designed for rPPG and has no pulse ground truth. Note that testing on UBFC after training on DDPM performs well, since their frequency distributions are similar, while PURE's pulse rates tend to be much lower." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 352, + 291, + 605 + ], + "lines": [ + { + "bbox": [ + 48, + 352, + 291, + 605 + ], + "spans": [ + { + "bbox": [ + 48, + 352, + 291, + 605 + ], + "type": "table", + "html": "
Training DatasetTesting DatasetMethodMAE (bpm)r
DDPMUBFCSiNC0.88 ± 0.250.98 ± 0.01
UBFCContrast-Phys1.14 ± 0.380.96 ± 0.03
UBFCPhysNet1.11 ± 0.420.95 ± 0.02
PURESiNC3.12 ± 1.070.88 ± 0.05
PUREContrast-Phys13.02 ± 6.120.19 ± 0.59
PUREPhysNet1.46 ± 0.340.95 ± 0.02
UBFCDDPMSiNC18.53 ± 0.360.38 ± 0.01
DDPMContrast-Phys22.93 ± 1.020.18 ± 0.04
DDPMPhysNet18.58 ± 0.120.40 ± 0.00
PURESiNC4.02 ± 0.060.86 ± 0.00
PUREContrast-Phys19.61 ± 2.010.33 ± 0.06
PUREPhysNet3.81 ± 0.340.87 ± 0.02
PUREUBFCSiNC6.64 ± 1.760.59 ± 0.10
UBFCContrast-Phys10.22 ± 0.380.45 ± 0.04
UBFCPhysNet7.02 ± 3.350.60 ± 0.13
DDPMSiNC24.92 ± 0.650.20 ± 0.00
DDPMContrast-Phys29.63 ± 0.480.03 ± 0.02
DDPMPhysNet28.03 ± 2.200.13 ± 0.05
HKBU (non-rPPG)UBFCSiNC1.08 ± 0.030.95 ± 0.00
PURESiNC2.43 ± 0.200.90 ± 0.02
DDPMSiNC20.34 ± 0.250.19 ± 0.02
", + "image_path": "29dfe9f8d7e006b6195df6f27aaf951707c5bcd5c6e831d04312067ed14d6eed.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 618, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 286, + 713 + ], + "type": "text", + "content": "We perform cross-dataset testing to analyze whether the approach is robust to changes to the lighting, camera sensor, pulse rate distribution, and motion. Table 2 shows the results for SiNC and supervised training with the same architecture. We find that the performance is similar for the supervised and unsupervised approaches when transferring to different data sources. Training on PURE exclusively gives relatively poor results when transferring to" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": "UBFC-rPPG and DDPM, due to the low pulse rate variability within PURE samples and lack of movement. Training on DDPM gives the best results overall, since the dataset is the largest and captures larger subjects' movements compared to other datasets." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 141, + 484, + 154 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 141, + 484, + 154 + ], + "spans": [ + { + "bbox": [ + 305, + 141, + 484, + 154 + ], + "type": "text", + "content": "6.3. Training with CelebV-HQ Videos" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 159, + 545, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 159, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 304, + 159, + 545, + 316 + ], + "type": "text", + "content": "Given the abundance of face videos publicly available online, we trained a model on the CelebV-HQ dataset [56]. After processing the available videos with MediaPipe and resampling to 30 fps, our unlabeled dataset consisted of 34,029 videos. We trained the model for 23 epochs and manually stopped training due to a plateau in the validation loss. Unfortunately, we found that the model could not converge to the true blood volume pulse. We attribute the failure to poor video quality from compression. Although the videos were downloaded with the highest available quality, they have likely been compressed, removing the pulse signal entirely. See the supplementary material for an assessment of rPPG quality using POS [47] on CelebV-HQ." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 325, + 495, + 337 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 325, + 495, + 337 + ], + "spans": [ + { + "bbox": [ + 305, + 325, + 495, + 337 + ], + "type": "text", + "content": "6.4. Training with HKBU-MARs Videos" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 344, + 545, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 344, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 304, + 344, + 545, + 475 + ], + "type": "text", + "content": "The HKBU-MARs dataset [17] was designed for face presentation attack detection, but we trained models on the \"real\" video sessions in the dataset. The bottom rows in Table 2 show the results for training on HKBU-MARs, then testing on the benchmark rPPG datasets. Training on HKBU-MARs gives better results when testing on UBFC-rPPG and PURE than all training sets except DDPM, which is an order of magnitude larger. To our knowledge, this is the first successful experiment showing that non-rPPG videos can be used to train robust rPPG models, even if they do not have ground-truth pulse labels." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 485, + 446, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 485, + 446, + 498 + ], + "spans": [ + { + "bbox": [ + 306, + 485, + 446, + 498 + ], + "type": "text", + "content": "6.5. Ablation Study on Losses" + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 310, + 546, + 539, + 640 + ], + "blocks": [ + { + "bbox": [ + 304, + 515, + 544, + 536 + ], + "lines": [ + { + "bbox": [ + 304, + 515, + 544, + 536 + ], + "spans": [ + { + "bbox": [ + 304, + 515, + 544, + 536 + ], + "type": "text", + "content": "Table 3. Ablation study on the loss functions used during training. Results are shown for models trained and tested on UBFC-rPPG." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 310, + 546, + 539, + 640 + ], + "lines": [ + { + "bbox": [ + 310, + 546, + 539, + 640 + ], + "spans": [ + { + "bbox": [ + 310, + 546, + 539, + 640 + ], + "type": "table", + "html": "
LossMAE (bpm)RMSE (bpm)r
Lb3.08 ± 1.698.08 ± 3.610.87 ±0.08
Ls45.50 ± 1.2250.04 ± 0.94-0.04 ±0.08
Lv22.89 ± 2.8331.51 ± 2.360.22 ±0.09
Ls + Lv51.24 ± 5.3657.80 ± 7.39-0.04 ±0.09
Lb + Ls9.99 ± 2.5517.14 ± 2.360.51 ±0.14
Lb + Lv4.18 ± 2.888.90 ± 5.240.82 ±0.14
Lb + Ls + Lv0.59 ± 0.001.83 ± 0.040.99 ± 0.00
", + "image_path": "8f5d1f9733a14bc67c6d99e08ffdfac85217cfa8685e9b0df127ed636b67cee4.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "We trained models using all combinations of loss components to analyze their contributions. Table 3 shows the results for training and testing on UBFC-rPPG. The bandwidth loss is the most critical for discovering the true blood volume pulse, while the sparsity and variance losses do not" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14470" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 286, + 137 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 286, + 137 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 286, + 137 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 286, + 137 + ], + "type": "image", + "image_path": "0d213413509265e049b90d0f45060212d169c24cc73034bdab220e757a890931.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 49, + 139, + 286, + 205 + ], + "blocks": [ + { + "bbox": [ + 49, + 139, + 286, + 205 + ], + "lines": [ + { + "bbox": [ + 49, + 139, + 286, + 205 + ], + "spans": [ + { + "bbox": [ + 49, + 139, + 286, + 205 + ], + "type": "image", + "image_path": "ba9286360b5aae1a58396a024019804f765b49d433e1589728c7e53827cd5e2d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 49, + 206, + 286, + 300 + ], + "blocks": [ + { + "bbox": [ + 49, + 206, + 286, + 300 + ], + "lines": [ + { + "bbox": [ + 49, + 206, + 286, + 300 + ], + "spans": [ + { + "bbox": [ + 49, + 206, + 286, + 300 + ], + "type": "image", + "image_path": "2a5daec4bc3fe5d7907ed4f892d1f55781180759b00821dd5d4d397f39cb90d9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 316, + 288, + 373 + ], + "lines": [ + { + "bbox": [ + 46, + 316, + 288, + 373 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 288, + 373 + ], + "type": "text", + "content": "Figure 3. Within-dataset waveform predictions on all baseline datasets from end-to-end unsupervised models over an 8-second window. The model predictions are remarkably periodic without any form of filtering. Note that phase is not considered during training, so each model learns its own phase shift." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 396, + 287, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 396, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 396, + 287, + 456 + ], + "type": "text", + "content": "learn the desired signal by themselves. Surprisingly, combining the bandwidth loss with just one of the sparsity or variance losses gives worse performance than just the bandwidth loss. However, when combining all three components, the model achieves impressive results." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 468, + 117, + 481 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 468, + 117, + 481 + ], + "spans": [ + { + "bbox": [ + 47, + 468, + 117, + 481 + ], + "type": "text", + "content": "7. Discussion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 490, + 262, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 490, + 262, + 502 + ], + "spans": [ + { + "bbox": [ + 47, + 490, + 262, + 502 + ], + "type": "text", + "content": "7.1. Improvements over Supervised Learning" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 509, + 287, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 509, + 287, + 676 + ], + "spans": [ + { + "bbox": [ + 46, + 509, + 287, + 676 + ], + "type": "text", + "content": "It is initially surprising that unsupervised training leads to similar or improved rPPG estimation models compared to those trained in a supervised manner. However, there are several potential benefits to unsupervised training. From a hardware perspective, one of the difficulties in supervised training is aligning the contact pulse waveform with the video frames [54]. The pulse sensor and camera may have a time lag, effectively giving the model an out-of-phase target at training time. Unsupervised training gives the model freedom to learn the phase directly from the video. The contact-PPG signal is also sensitive to motion and may be noisy. Since motion may co-occur at the face and fingertip, the contact signal may misguide the model towards visual features for which they should be invariant." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "text", + "content": "From a physiological perspective, the pulse observed optically at the fingertip with a contact sensor has a different phase than that of the face, since blood propagates along a" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 191 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 191 + ], + "type": "text", + "content": "different path before reaching the peripheral microvasculature, making alignment nearly impossible without shifting the targets to rPPG estimates from existing methods [38]. Additionally, the morphological shape of the contact-PPG waveform depends on numerous factors such as the wavelength of light (and corresponding tissue penetration depth), external pressure from the oximeter clip, and vasodilation at the measurement site [1,28]. This indicates that the morphology and phase of the target PPG waveform is likely different from the observed rPPG waveform." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 198, + 422, + 210 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 198, + 422, + 210 + ], + "spans": [ + { + "bbox": [ + 306, + 198, + 422, + 210 + ], + "type": "text", + "content": "7.2. Why Does It Work?" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 216, + 545, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 216, + 545, + 312 + ], + "spans": [ + { + "bbox": [ + 304, + 216, + 545, + 312 + ], + "type": "text", + "content": "The success of SiNC depends on specific properties of the data, model, and how the two interact. Limited model capacity is actually required, since it forces learning features that generalize across inputs. By constraining the model's predictions to have specific periodic properties, the limited-capacity model must find a general set of features to produce a signal that exists in most or all of the training samples, which is the blood volume pulse in our datasets." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 312, + 545, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 312, + 545, + 419 + ], + "spans": [ + { + "bbox": [ + 304, + 312, + 545, + 419 + ], + "type": "text", + "content": "As a beneficial side-effect, the model intrinsically learns to ignore common noise factors such as illumination, rigid motion, non-rigid motion (e.g. talking, smiling, etc.), and sensor noise, since they may preside outside the predefined bandlimits or with uniform power spectra. Even if noise exhibits periodic tendencies within the bandlimits for some samples, those features would produce poor signals on other samples. Therefore, end-to-end unsupervised approaches like SiNC are well-suited for periodic problems." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 429, + 383, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 429, + 383, + 441 + ], + "spans": [ + { + "bbox": [ + 306, + 429, + 383, + 441 + ], + "type": "text", + "content": "8. Conclusions" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 449, + 545, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 545, + 567 + ], + "type": "text", + "content": "We proposed a novel non-contrastive learning approach for end-to-end unsupervised signal regression, with specific experiments on blood volume pulse estimation from face videos. This SiNC framework effectively learns powerful visual features with only loose frequency constraints. We demonstrated this by training accurate rPPG models using non-rPPG data and our simple loss functions. Given the subtlety of the rPPG signal, we believe our work can be extended to other signal regression tasks in the domain of remote vitals estimation." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgements. This research was sponsored by the Securiport Global Innovation Cell, a division of Securiport LLC. Commercial equipment is identified in this work in order to adequately specify or describe the subject matter. In no case does such identification imply recommendation or endorsement by Securiport LLC, nor does it imply that the equipment identified is necessarily the best available for this purpose. The opinions, findings, and conclusions or recommendations expressed in this publication are those of the authors and do not necessarily reflect the views of our sponsors." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14471" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 146 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 146 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 146 + ], + "type": "text", + "content": "[1] Pierre Abraham, Mélissa Bourgeau, Maïte Camo, Anne Humeau-Heurtier, Sylvain Durand, Pascal Rousseau, and Guillaume Mahe. Effect of skin temperature on skin endothelial function assessment. Microvascular Research, 88:56-60, 2013. 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 148, + 288, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 288, + 192 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 288, + 192 + ], + "type": "text", + "content": "[2] T. Baltrusaitis, A. Zadeh, Y. C. Lim, and L. Morency. Openface 2.0: Facial behavior analysis toolkit. In IEEE International Conference on Automatic Face Gesture Recognition (FG), pages 59-66, 2018. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 194, + 288, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 194, + 288, + 226 + ], + "spans": [ + { + "bbox": [ + 53, + 194, + 288, + 226 + ], + "type": "text", + "content": "[3] Adrien Bardes, Jean Ponce, and Yann LeCun. Vicreg: Variance-invariance-covariance regularization for self-supervised learning. In ICLR, 2022. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 228, + 288, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 228, + 288, + 271 + ], + "spans": [ + { + "bbox": [ + 53, + 228, + 288, + 271 + ], + "type": "text", + "content": "[4] Serge Bobbia, Richard Macwan, Yannick Benezeth, Alamin Mansouri, and Julien Dubois. Unsupervised skin tissue segmentation for remote photoplethysmography. Pattern Recognition Letters, 124:82-90, 2019. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 273, + 288, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 273, + 288, + 328 + ], + "spans": [ + { + "bbox": [ + 53, + 273, + 288, + 328 + ], + "type": "text", + "content": "[5] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in Neural Information Processing Systems, 2020-Decem(NeurIPS):1-13, 2020. arXiv: 2006.09882. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 330, + 288, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 330, + 288, + 384 + ], + "spans": [ + { + "bbox": [ + 53, + 330, + 288, + 384 + ], + "type": "text", + "content": "[6] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. 37th International Conference on Machine Learning, ICML 2020, pages 1575-1585, 2020. arXiv: 2002.05709 ISBN: 9781713821120. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 386, + 288, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 386, + 288, + 430 + ], + "spans": [ + { + "bbox": [ + 53, + 386, + 288, + 430 + ], + "type": "text", + "content": "[7] Weixuan Chen and Daniel McDuff. DeepPhys: Video-based physiological measurement using convolutional attention networks. In European Conference on Computer Vision (ECCV), pages 356-373, 2018. 1, 2, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 431, + 288, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 431, + 288, + 485 + ], + "spans": [ + { + "bbox": [ + 53, + 431, + 288, + 485 + ], + "type": "text", + "content": "[8] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 15745–15753, 2021. arXiv: 2011.10566 ISBN: 9781665445092. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 487, + 288, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 487, + 288, + 520 + ], + "spans": [ + { + "bbox": [ + 53, + 487, + 288, + 520 + ], + "type": "text", + "content": "[9] G. De Haan and V. Jeanne. Robust pulse rate from chrominance-based rppg. IEEE Transactions on Biomedical Engineering, 60(10):2878-2886, 2013. 2, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 522, + 287, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 522, + 287, + 565 + ], + "spans": [ + { + "bbox": [ + 48, + 522, + 287, + 565 + ], + "type": "text", + "content": "[10] Gerard De Haan and Arno Van Leest. Improved motion robustness of remote-PPG by using the blood volume pulse signature. Physiological Measurement, 35(9):1913–1926, 2014. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 567, + 287, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 287, + 611 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 287, + 611 + ], + "type": "text", + "content": "[11] Aleksandr Ermolov, Aliaksandr Siarohin, Enver Sangineto, and Nicu Sebe. Whitening for self-supervised representation learning. In International Conference on Machine Learning, pages 3015-3024. PMLR, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 613, + 287, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 656 + ], + "type": "text", + "content": "[12] John Gideon and Simon Stent. Estimating heart rate from unlabelled video. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, pages 2743-2749, October 2021. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[13] John Gideon and Simon Stent. The way to my heart is through contrastive learning: Remote photoplethysmography from unlabelled video. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3995-4004, October 2021. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 715 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 160 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 160 + ], + "type": "text", + "content": "[14] Jean Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre H. Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Daniel Guo, Mohammad Gheshlaghi Azar, Bilal Piot, Koray Kavukcuoglu, Rémi Munos, and Michal Valko. Bootstrap your own latent a new approach to self-supervised learning. Advances in Neural Information Processing Systems, 2020-Decem(NeurIPS):1-14, 2020. arXiv: 2006.07733. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 161, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 161, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 307, + 161, + 545, + 194 + ], + "type": "text", + "content": "[15] Le Hou, Chen-Ping Yu, and Dimitris Samaras. Squared Earth Mover's Distance-based Loss for Training Deep Neural Networks, Apr. 2017. arXiv:1611.05916 [cs]. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 195, + 545, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 195, + 545, + 237 + ], + "spans": [ + { + "bbox": [ + 307, + 195, + 545, + 237 + ], + "type": "text", + "content": "[16] Eugene Lee, Evan Chen, and Chen-Yi Lee. Meta-rppg: Remote heart rate estimation using a transductive meta-learner. In European Conference on Computer Vision (ECCV), 2020. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 239, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 239, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 307, + 239, + 545, + 293 + ], + "type": "text", + "content": "[17] Siqi Liu, Baoyao Yang, Pong C. Yuen, and Guoying Zhao. A 3d mask face anti-spoofing database with real world variations. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, June 2016. 5, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 294, + 545, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 294, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 307, + 294, + 545, + 358 + ], + "type": "text", + "content": "[18] Xin Liu, Josh Fromm, Shwetak Patel, and Daniel McDuff. Multi-task temporal shift attention networks for on-device contactless vitals measurement. In H. Larochelle, M. Ranzato, R. Hadsell, M. F. Balcan, and H. Lin, editors, Advances in Neural Information Processing Systems, volume 33, pages 19400-19411. Curran Associates, Inc., 2020. 1, 2, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 360, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 360, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 307, + 360, + 545, + 392 + ], + "type": "text", + "content": "[19] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 394, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 394, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 307, + 394, + 545, + 437 + ], + "type": "text", + "content": "[20] Hao Lu, Hu Han, and S Kevin Zhou. Dual-GAN: Joint BVP and Noise Modeling for Remote Physiological Measurement. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 12404-12413, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 437, + 545, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 437, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 307, + 437, + 545, + 491 + ], + "type": "text", + "content": "[21] Hao Lu, Hu Han, and S Kevin Zhou. Dual-gan: Joint bvp and noise modeling for remote physiological measurement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12404-12413, 2021. 3, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 493, + 545, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 493, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 307, + 493, + 545, + 547 + ], + "type": "text", + "content": "[22] Camillo Lugaresi, Jiuqiang Tang, Hadon Nash, Chris McClanahan, Esha Uboweja, Michael Hays, Fan Zhang, Chuoling Chang, Ming Guang Yong, Juhyun Lee, et al. Mediapipe: A framework for building perception pipelines. arXiv preprint arXiv:1906.08172, 2019. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 548, + 545, + 570 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 548, + 545, + 570 + ], + "spans": [ + { + "bbox": [ + 307, + 548, + 545, + 570 + ], + "type": "text", + "content": "[23] Daniel McDuff. Camera measurement of physiological vital signs. ACM Comput. Surv., aug 2022. Just Accepted. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 571, + 545, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 571, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 307, + 571, + 545, + 634 + ], + "type": "text", + "content": "[24] Daniel McDuff, Miah Wander, Xin Liu, Brian L. Hill, Javier Hernandez, Jonathan Lester, and Tadas Baltrusaitis. SCAMPS: Synthetics for camera measurement of physiological signals. In Thirty-sixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 635, + 545, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 545, + 691 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 545, + 691 + ], + "type": "text", + "content": "[25] Daniel J. McDuff, Ethan B. Blackford, and Justin R. Estepp. The Impact of Video Compression on Remote Cardiac Pulse Measurement Using Imaging Photoplethysmography. IEEE International Conference on Automatic Face and Gesture Recognition Workshops (FG), pages 63-70, 2017. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 692, + 545, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 692, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 307, + 692, + 545, + 715 + ], + "type": "text", + "content": "[26] Yuriy Mironenko, Konstantin Kalinin, Mikhail Kopeliovich, and Mikhail Petrushan. Remote photoplethysmography:" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14472" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 106 + ], + "type": "text", + "content": "Rarely considered factors. In IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 1197-1206, June 2020. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 139 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 139 + ], + "type": "text", + "content": "[27] Ishan Misra and Laurens van der Maaten. Self-Supervised Learning of Pretext-Invariant Representations. CVPR, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 288, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 288, + 206 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 288, + 206 + ], + "type": "text", + "content": "[28] Andreia Moco, Sander Stuijk, Gerard De Haan, Ruikang K. Wang, and Wim Verkruysse. A Model for Waveform Dissimilarities in Dual-Depth Reflectance-PPG. Proceedings of the Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBS), 2018-July:5125-5130, 2018. ISBN: 9781538636466. 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 209, + 287, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 287, + 252 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 287, + 252 + ], + "type": "text", + "content": "[29] Xuesong Niu, Hu Han, Shiguang Shan, and Xin Chen. VIPL-HR: A multi-modal database for pulse estimation from less-constrained face video. In Asian Conference on Computer Vision (ACCV), pages 562-576, 2018. 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 254, + 287, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 254, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 48, + 254, + 287, + 297 + ], + "type": "text", + "content": "[30] Xuesong Niu, Shiguang Shan, Hu Han, and Xin Chen. RhythmNet: End-to-end heart rate estimation from face via spatial-temporal representation. IEEE Transactions on Image Processing, 29:2409-2423, 2020. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 298, + 287, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 298, + 287, + 353 + ], + "spans": [ + { + "bbox": [ + 48, + 298, + 287, + 353 + ], + "type": "text", + "content": "[31] Ewa Nowara and Daniel McDuff. Combating the Impact of Video Compression on Non-Contact Vital Sign Measurement Using Supervised Learning. In IEEE International Conference on Computer Vision Workshop (ICCVW), pages 1706–1712, Oct. 2019. ISSN: 2473-9944. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 354, + 287, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 354, + 287, + 397 + ], + "spans": [ + { + "bbox": [ + 48, + 354, + 287, + 397 + ], + "type": "text", + "content": "[32] Ewa M. Nowara, Daniel McDuff, and Ashok Veeraraghavan. Systematic analysis of video-based pulse measurement from compressed videos. Biomedical Optics Express, 12(1):494, 2021. 3, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 399, + 287, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 287, + 430 + ], + "type": "text", + "content": "[33] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding, 2018. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 433, + 287, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 433, + 287, + 477 + ], + "spans": [ + { + "bbox": [ + 48, + 433, + 287, + 477 + ], + "type": "text", + "content": "[34] M. Poh, D. J. McDuff, and R. W. Picard. Advancements in noncontact, multiparameter physiological measurements using a webcam. IEEE Transactions on Biomedical Engineering, 58(1):7-11, 2011. 2, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 478, + 287, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 478, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 48, + 478, + 287, + 521 + ], + "type": "text", + "content": "[35] Ming-Zher Poh, Daniel J. McDuff, and Rosalind W. Picard. Non-contact, automated cardiac pulse measurements using video imaging and blind source separation. Opt. Express, 18(10):10762-10774, May 2010. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 523, + 287, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 523, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 523, + 287, + 567 + ], + "type": "text", + "content": "[36] Michal Rapczynski, Philipp Werner, and Ayoub Al-Hamadi. Effects of video encoding on camera-based heart rate estimation. IEEE Transactions on Biomedical Engineering, 66(12):3360-3370, 2019. Publisher: IEEE. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 568, + 287, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 287, + 622 + ], + "type": "text", + "content": "[37] Rencheng Song, Huan Chen, Juan Cheng, Chang Li, Yu Liu, and Xun Chen. Pulsegan: Learning to generate realistic pulse waveforms in remote photoplethysmography. IEEE Journal of Biomedical and Health Informatics, 25(5):1373-1384, 2021. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 624, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 678 + ], + "type": "text", + "content": "[38] Jeremy Speth, Nathan Vance, Adam Czajka, Kevin Bowyer, and Patrick Flynn. Unifying frame rate and temporal dilations for improved remote pulse detection. Computer Vision and Image Understanding (CVIU), pages 1056-1062, 2021. 1, 2, 3, 5, 6, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[39] Jeremy Speth, Nathan Vance, Adam Czajka, Kevin Bowyer, Diane Wright, and Patrick Flynn. Deception detection and remote physiological monitoring: A dataset and baseline ex" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "perimental results. In International Joint Conference on Biometrics (IJCB), pages 4264-4271, 2021. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 140 + ], + "type": "text", + "content": "[40] Radim Špetlík, Vojtech Franc, and Jírí Matas. Visual heart rate estimation with convolutional neural network. In Proceedings of the British machine vision conference, Newcastle, UK, pages 3-6, 2018. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 141, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 194 + ], + "type": "text", + "content": "[41] Ronny Stricker, Steffen Muller, and Horst Michael Gross. Non-contact video-based pulse rate measurement on a mobile service robot. IEEE International Symposium on Robot and Human Interactive Communication, pages 1056-1062, 2014. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 197, + 545, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 240 + ], + "type": "text", + "content": "[42] Zhaodong Sun and Xiaobai Li. Contrast-phys: Unsupervised video-based remote physiological measurement via spatiotemporal contrast. European Conference on Computer Vision (ECCV), 2022. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 242, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 242, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 242, + 545, + 308 + ], + "type": "text", + "content": "[43] Nathan Vance, Jeremy Speth, Siamul Khan, Adam Czajka, Kevin W. Bowyer, Diane Wright, and Patrick Flynn. Deception detection and remote physiological monitoring: A dataset and baseline experimental results. IEEE Transactions on Biometrics, Behavior, and Identity Science (TBIOM), pages 1-1, 2022. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 309, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 309, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 307, + 309, + 545, + 342 + ], + "type": "text", + "content": "[44] Wim Verkruysse, Lars O Svaasand, and J Stuart Nelson. Remote plethysmographic imaging using ambient light. Opt. Express, 16(26):21434-21445, Dec 2008. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 343, + 545, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 343, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 307, + 343, + 545, + 398 + ], + "type": "text", + "content": "[45] Hao Wang, Euijoon Ahn, and Jinman Kim. Self-Supervised Representation Learning Framework for Remote Physiological Measurement Using Spatiotemporal Augmentation Loss. Proceedings of the AAAI Conference on Artificial Intelligence, 36(2):2431-2439, June 2022. 1, 2, 3, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 399, + 545, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 545, + 432 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 545, + 432 + ], + "type": "text", + "content": "[46] Wenjin Wang, Albertus C. Den Brinker, and Gerard De Haan. Single-Element Remote-PPG. IEEE Transactions on Biomedical Engineering, 66(7):2032-2043, 2019. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 434, + 545, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 434, + 545, + 466 + ], + "spans": [ + { + "bbox": [ + 307, + 434, + 545, + 466 + ], + "type": "text", + "content": "[47] W. Wang, A. C. den Brinker, S. Stuijk, and G. de Haan. Algorithmic principles of remote PPG. IEEE Transactions on Biomedical Engineering, 64(7):1479-1491, 2017. 2, 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 468, + 545, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 468, + 545, + 511 + ], + "spans": [ + { + "bbox": [ + 307, + 468, + 545, + 511 + ], + "type": "text", + "content": "[48] Wenjin Wang, Sander Stuijk, and Gerard De Haan. A Novel Algorithm for Remote Photoplethysmography: Spatial Subspace Rotation. IEEE Transactions on Biomedical Engineering, 63(9):1974-1984, 2016. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 513, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 513, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 307, + 513, + 545, + 578 + ], + "type": "text", + "content": "[49] Zhen Wang, Yunhao Ba, Pradyumna Chari, Oyku Deniz Bozkurt, Gianna Brown, Parth Patwa, Niranjan Vaddi, Laleh Jalilian, and Achuta Kadambi. Synthetic generation of face videos with plethysmograph physiology. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20555-20564, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 579, + 545, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 579, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 307, + 579, + 545, + 612 + ], + "type": "text", + "content": "[50] Yuzhe Yang, Xin Liu, Jiang Wu, Silviu Borac, Dina Katabi, Ming-Zher Poh, and Daniel McDuff. Simper: Simple self-supervised learning of periodic targets, 2022. 1, 2, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 613, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 545, + 667 + ], + "type": "text", + "content": "[51] Zitong Yu, Xiaobai Li, and Guoying Zhao. Remote photoplethysmograph signal measurement from facial videos using spatio-temporal networks. In Proceedings of the British Machine Vision Conference (BMVC), pages 1-12, 2019. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 669, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 545, + 713 + ], + "type": "text", + "content": "[52] Zitong Yu, Yuming Shen, Jingang Shi, Hengshuang Zhao, Philip H.S. Torr, and Guoying Zhao. Physformer: Facial video-based physiological measurement with temporal difference transformer. In Proceedings of the IEEE/CVF" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14473" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 274 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "text", + "content": "Conference on Computer Vision and Pattern Recognition (CVPR), pages 4186-4196, June 2022. 1, 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 129 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 129 + ], + "type": "text", + "content": "[53] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow Twins: Self-Supervised Learning via Redundancy Reduction. 2021. arXiv: 2103.03230. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 130, + 287, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 130, + 287, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 130, + 287, + 173 + ], + "type": "text", + "content": "[54] Qi Zhan, Wenjin Wang, and Gerard de Haan. Analysis of CNN-based remote-PPG to understand limitations and sensitivities. Biomedical Optics Express, 11(3):1268-1283, 2020. 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 175, + 287, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 287, + 229 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 287, + 229 + ], + "type": "text", + "content": "[55] Yu Zhao, Bochao Zou, Fan Yang, Lin Lu, Abdelkader Nasreddine Belkacem, and Chao Chen. Video-based physiological measurement using 3d central difference convolution attention network. In 2021 IEEE International Joint Conference on Biometrics (IJCB), pages 1-6, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 231, + 287, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 231, + 287, + 274 + ], + "spans": [ + { + "bbox": [ + 48, + 231, + 287, + 274 + ], + "type": "text", + "content": "[56] Hao Zhu, Wayne Wu, Wentao Zhu, Liming Jiang, Siwei Tang, Li Zhang, Ziwei Liu, and Chen Change Loy. CelebV-HQ: A large-scale video facial attributes dataset. In ECCV, 2022. 5, 7" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "14474" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Non-Line-of-Sight Imaging With Signal Superresolution Network/224cfcfa-2657-4e74-b370-35626fe2670d_content_list.json b/2023/Non-Line-of-Sight Imaging With Signal Superresolution Network/224cfcfa-2657-4e74-b370-35626fe2670d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6dd91bc44558ccd86f3833ed136f5acc880c6bdb --- /dev/null +++ b/2023/Non-Line-of-Sight Imaging With Signal Superresolution Network/224cfcfa-2657-4e74-b370-35626fe2670d_content_list.json @@ -0,0 +1,1498 @@ +[ + { + "type": "text", + "text": "Non-line-of-sight Imaging with Signal Superresolution Network", + "text_level": 1, + "bbox": [ + 163, + 130, + 805, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jianyu Wang $^{1}$ , Xintong Liu $^{1}$ , Leping Xiao $^{1}$ , Zuoqiang Shi $^{1,2}$ , Lingyun Qiu $^{1,2}$ , and Xing Fu $^{1}$ $^{1}$ Tsinghua University", + "bbox": [ + 122, + 179, + 843, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Yanqi Lake Beijing Institute of Mathematical Sciences and Applications", + "bbox": [ + 191, + 215, + 772, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Non-line-of-sight (NLOS) imaging aims at reconstructing the location, shape, albedo, and surface normal of the hidden object around the corner with measured transient data. Due to its strong potential in various fields, it has drawn much attention in recent years. However, long exposure time is not always available for applications such as auto-driving, which hinders the practical use of NLOS imaging. Although scanning fewer points can reduce the total measurement time, it also brings the problem of imaging quality degradation. This paper proposes a general learning-based pipeline for increasing imaging quality with only a few scanning points. We tailor a neural network to learn the operator that recovers a high spatial resolution signal. Experiments on synthetic and measured data indicate that the proposed method provides faithful reconstructions of the hidden scene under both confocal and nonconfocal settings. Compared with original measurements, the acquisition of our approach is 16 times faster while maintaining similar reconstruction quality. Besides, the proposed pipeline can be applied directly to existing optical systems and imaging algorithms as a plug-in-and-play module. We believe the proposed pipeline is powerful in increasing the frame rate in NLOS video imaging.", + "bbox": [ + 75, + 301, + 473, + 648 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 676, + 209, + 694 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Non-line-of-sight (NLOS) imaging problem usually employs a high temporal resolution optical system to recover the hidden object around the corner. As shown in Fig. 1a, photons emitted by the laser are collected by the detector after three diffuse reflections: the reflection at the visible wall, the reflection at the hidden object, and the reflection at the visible wall again. Scanning a region on the visible wall can get measured histogram data to recover the hidden scene.", + "bbox": [ + 75, + 703, + 468, + 838 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Due to its potential applications in various fields, such as auto-driving, disaster relief, and remote sensing, the NLOS imaging problem has become an emerging field since it was first proposed by Kirmani et al. [17] in 2009.", + "bbox": [ + 75, + 839, + 470, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/429a37763f5f9637697c2908d9b5c6d7d715ed12096e8194d4c7cf78eb55b456.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 509, + 267, + 883, + 477 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6394929e8689e54aa2c5e7e2876a09f74b7be49627ef25f50cfe9fe009d4f5b7.jpg", + "image_caption": [ + "(b) Scanning points" + ], + "image_footnote": [], + "bbox": [ + 516, + 501, + 691, + 637 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b26ef31d3c80ab0bc1c79602292b638c5999426563c750bf3cf914cfcd211f77.jpg", + "image_caption": [ + "(a) NLOS scenario", + "Low" + ], + "image_footnote": [], + "bbox": [ + 700, + 496, + 777, + 564 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c44ef8d632ba6e3af2af42f6d0c0ef089fbac977395c26c62a0681b607f5b73c.jpg", + "image_caption": [ + "(c) Results", + "Figure 1. A typical NLOS scenario and reconstruction results. (a) An illustration of a typical NLOS scenario. (b) An illustration of the scanning points on the visible wall. Using the proposed pipeline, only the blue circle points are needed to be illuminated, and the signal at the yellow square points can be recovered with the proposed pipeline. (c) Comparisons of the reconstruction results of the pyramid with different signals. The maximum intensity projection of the albedo values along the depth direction is shown in the upper left corner (GT). The results are reconstructed by the original signal with spatial resolution $32 \\times 32$ (Original), the subsampled signal with spatial resolution $8 \\times 8$ (Low), and the signal recovered by the proposed network with spatial resolution $32 \\times 32$ (Ours)." + ], + "image_footnote": [], + "bbox": [ + 782, + 496, + 882, + 638 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Many methods have been proposed to improve the prac", + "bbox": [ + 500, + 885, + 890, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "17420", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cc7ccacf4ab46248265f69b400c2fa39b4250dd6c0c0826939aa02df70a2837d.jpg", + "image_caption": [ + "Figure 2. Flowchart of the proposed pipeline. The high resolution signal is recovered from the low resolution signal using a neural network. The hidden scene is then reconstructed using state-of-the-art imaging algorithms designed for high resolution signal." + ], + "image_footnote": [], + "bbox": [ + 130, + 88, + 843, + 227 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ticability [1, 3, 13, 14, 34, 35, 39] or reconstruction quality [10, 11, 26, 41]. A back-projection method with Laplacian of Gaussian filter (LOG-BP) [18] is introduced to improve the reconstruction quality of the back-projection method [38]. In 2018, O'Toole et al. [31] first apply the light-contransform (LCT) method to image the hidden object with a confocal scanning mode. Young et al. [45] then propose the directional light-cone-transform (D-LCT) method, which simultaneously recovers the albedo and surface normal of the hidden object. From the perspective of wave characteristics, Lindell et al. [22] introduce the fast frequency-wavenumber migration (F-K) method to NLOS. Another remarkable imaging method proposed by Liu et al. [25] employs the phasor field model [9, 33] and provides an efficient solution for fast NLOS imaging under non-confocal scenario [24]. Further improvements of this method using single photon avalanche diode (SPAD) arrays achieve real-time imaging [21, 30]. Deep learning based methods have also been introduced [7, 8, 12, 16, 29]. Recently, Liu et al. [27] propose the signal-object collaborative regularization (SOCR) method, which can improve the resolution of the reconstructed object, especially in cases with high measurement noise.", + "bbox": [ + 75, + 291, + 472, + 638 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, one main obstacle in NLOS imaging is that we need a long exposure time to improve the signal-to-noise ratio (SNR) of the measurement, which is not admissible in many scenarios, such as auto-driving. Several methods have been proposed to reduce the capture time [20, 23]. For a confocal scenario, Isogawa et al. [15] propose a circular scanning mode. In 2021, Metzler et al. [28] consider an extreme case that tracks the objects with a single detection point. By introducing the downsampling operation, Ye et al. [44] reconstruct the object with a subset of the virtual sources. For non-confocal scenarios, most methods employ SPAD arrays to realize fast measurement. Among them, Nam et al. [30] achieve a reconstruction of five frames per second by remapping the data from two $16 \\times 1$ SPAD arrays. Pei et al. [32] realize 20 frames per second reconstruction using a $32 \\times 32$ SPAD array. In 2022, Yang et al. [42] propose a spatial multiplexing method based on the", + "bbox": [ + 75, + 643, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "compressed sensing strategy. Although these methods can reduce the measurement time, they might either cause deterioration in imaging quality or substantially increase expense.", + "bbox": [ + 496, + 292, + 890, + 351 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose a deep learning based pipeline that can reconstruct the hidden object with fine details using only a small number of scanning points. Different measurement densities are shown in Fig. 1b. In the following sections, we term the signal measured at the blue circle points as the \"low resolution signal\" while those measured at both circle and square points as the \"high resolution signal\".", + "bbox": [ + 496, + 352, + 890, + 458 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Instead of designing new algorithms for sparse measurements, we focus on recovering the dense virtual signals, which can be applied to most algorithms as a plug-in-and-play module. The proposed pipeline consists of two steps: the low resolution signal is first mapped into the high resolution signal space by a neural network; the reconstruction is then obtained from the recovered high resolution signal using existing imaging algorithms. The reconstruction results of the pyramid are shown in Fig. 1c. These results are reconstructed by the F-K method. As the comparison shows, the proposed pipeline can provide a high quality reconstruction with only 64 scanning points, which is $6.25\\%$ of the original number.", + "bbox": [ + 496, + 458, + 892, + 654 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To our best knowledge, this is the first work in NLOS to study the signal recovery problem of the low resolution signal sparsely measured on a uniform grid. The proposed method can reduce the exposure time without losing the reconstruction quality. The main contributions of this work are:", + "bbox": [ + 496, + 655, + 890, + 744 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel learning based pipeline for faithful reconstruction of the hidden scene, which only takes $6.25\\%$ of the original measuring time.", + "- We show that the high resolution signal can be recovered accurately with the proposed signal superresolution network, which we abbreviate as \"SSN\".", + "- The effectiveness of the proposed method is validated on public datasets under both confocal and non" + ], + "bbox": [ + 514, + 756, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "17421", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/fd498f1a352f288178a504422cc88beab378c33bdd54c93a0803636122ed287c.jpg", + "image_caption": [ + "Figure 3. The architecture of the signal superresolution network (SSN) in the proposed pipeline. The network is composed of two main branches: a conventional upsampling branch and a learning branch. The final output is the sum of the results from these two branches." + ], + "image_footnote": [], + "bbox": [ + 133, + 90, + 843, + 294 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "confocal settings.", + "bbox": [ + 109, + 361, + 230, + 378 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- We demonstrate the potential application of the proposed method in high frame rate (64 frames per second) NLOS video imaging using public datasets.", + "bbox": [ + 93, + 386, + 470, + 433 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Reconstruction Pipeline", + "text_level": 1, + "bbox": [ + 76, + 445, + 303, + 463 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This work aims to develop a learning based pipeline for NLOS imaging, which can reduce the measurement time while maintaining the reconstruction quality. In this section, we formulate the signal recovery problem and introduce the proposed pipeline.", + "bbox": [ + 75, + 470, + 468, + 547 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1. Formulation of the problem", + "text_level": 1, + "bbox": [ + 76, + 555, + 328, + 571 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Most existing methods that reconstruct the hidden object from low resolution signals $d_{l}$ can be formulated as", + "bbox": [ + 75, + 578, + 468, + 609 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nI (x, y, z) = F _ {l} \\left(d _ {l}\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 207, + 619, + 468, + 638 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "in which $I(x,y,z)$ is the reconstructed object and $F_{l}$ is the imaging algorithm designed for low resolution signal [15, 28, 44]. Since only a few sampling points are scanned, the total exposure time can be reduced. However, the reconstruction quality also decreases with the number of scan points.", + "bbox": [ + 75, + 648, + 468, + 738 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "On the contrary, the algorithms designed for high resolution signal can be expressed as", + "bbox": [ + 75, + 739, + 468, + 768 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nI (x, y, z) = F _ {h} \\left(d _ {h}\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 204, + 781, + 468, + 799 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "in which $F_{h}$ represents the algorithm designed for high resolution signal $d_{h}$ . Many works have been proposed for fast [22, 24, 31, 45] and high resolution [18, 27] reconstruction as discussed above.", + "bbox": [ + 75, + 809, + 468, + 869 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this work, we aim at designing a pipeline that can provide faithful reconstruction with only a small number of", + "bbox": [ + 75, + 869, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "scanning points, and has good compatibility with traditional optical systems as well as existing imaging algorithms. To solve this problem, we introduce a signal recovering operator $\\Phi$ , which maps the low resolution signal $d_{l}$ into the space of high resolution signal $d_{h}$ . The proposed pipeline can be expressed as", + "bbox": [ + 496, + 361, + 890, + 450 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nI (x, y, z) = F _ {h} \\left(\\Phi \\left(d _ {l}\\right)\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 616, + 458, + 890, + 476 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "in which $\\Phi(d_l)$ is of the same size with $d_h$ . Thus, the problem of NLOS imaging from low resolution signal converts to the problem of finding the operator $\\Phi$ which satisfies $\\Phi(d_l) = d_h$ .", + "bbox": [ + 496, + 489, + 892, + 551 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. The choice of recovering operator", + "text_level": 1, + "bbox": [ + 500, + 566, + 795, + 583 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The choice of the operator $\\Phi$ is highly dependent on the physical model of the measurement process. We adopt a simplified version of the widely used three-point transport model [37] as", + "bbox": [ + 496, + 592, + 890, + 654 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\tau (\\mathbf {x} _ {l}, \\mathbf {x} _ {d}, t) = \\int_ {S} T (\\mathbf {x} _ {l}, \\mathbf {x} _ {d}, t, \\mathbf {s}) C (\\mathbf {x} _ {l}, \\mathbf {x} _ {d}, \\mathbf {s}) f (\\mathbf {s}) d A (\\mathbf {s}) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 667, + 890, + 710 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "in which $\\tau$ is the photon intensity detected with the virtual source $\\mathbf{x}_l$ , virtual detector $\\mathbf{x}_d$ and time $t$ . $f(\\mathbf{s})$ represents the albedo of the point $\\mathbf{s}$ on the hidden object $S$ and $A(\\mathbf{s})$ is the corresponding area measure. $T$ is a function that characterizes the arrival time of the photons, which is defined as", + "bbox": [ + 496, + 710, + 890, + 801 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nT (\\mathbf {x} _ {l}, \\mathbf {x} _ {d}, t, \\mathbf {s}) = \\delta (| | \\mathbf {x} _ {l} - \\mathbf {s} | | _ {2} + | | \\mathbf {x} _ {d} - \\mathbf {s} | | _ {2} - c t) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 837, + 890, + 854 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "in which $\\delta$ is the Dirac function and $c$ is the light speed. Besides, the function $C$ characterizes the intensity attenuation", + "bbox": [ + 496, + 869, + 890, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "17422", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "in the physical model, which is expressed as", + "bbox": [ + 76, + 90, + 372, + 106 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} C \\left(\\mathbf {x} _ {l}, \\mathbf {x} _ {d}, \\mathbf {s}\\right) = \\frac {< \\mathbf {s} - \\mathbf {x} _ {l} , \\mathbf {n} (\\mathbf {x} _ {l}) > < \\mathbf {x} _ {l} - \\mathbf {s} , \\mathbf {n} (\\mathbf {s}) >}{\\left| \\left| \\mathbf {x} _ {l} - \\mathbf {s} \\right| \\right| _ {2} ^ {4}} (6) \\\\ \\times \\frac {< \\mathbf {s} - \\mathbf {x} _ {d} , \\mathbf {n} (\\mathbf {x} _ {d}) > < \\mathbf {x} _ {d} - \\mathbf {s} , \\mathbf {n} (\\mathbf {s}) >}{\\left| \\left| \\mathbf {x} _ {d} - \\mathbf {s} \\right| \\right| _ {2} ^ {\\frac {4}}}. (6) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 116, + 468, + 186 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "in which $\\mathbf{n}(\\mathbf{s})$ is the unit normal vector pointing toward the visible wall at point $\\mathbf{s}$ and $<\\cdot, \\cdot>$ is the inner product in $\\mathbb{R}^3$ .", + "bbox": [ + 76, + 196, + 468, + 241 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Suppose that there are two pairs $(\\mathbf{x}_{l_1},\\mathbf{x}_{d_1}),(\\mathbf{x}_{l_2},\\mathbf{x}_{d_2})$ on the visible wall. For a fixed point s on the hidden object, we assume that the contributions of this point to the signal at these two pairs are contained in time bin $t_1$ and $t_2$ respectively. Then from Eq. (5), we know that $t_1$ and $t_2$ are close if", + "bbox": [ + 75, + 243, + 468, + 332 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nD = \\left\\| \\mathbf {x} _ {l _ {1}} - \\mathbf {x} _ {l _ {2}} \\right\\| _ {2} + \\left\\| \\mathbf {x} _ {d _ {1}} - \\mathbf {x} _ {d _ {2}} \\right\\| _ {2} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 335, + 468, + 352 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "is small. We can then regard the signal recovery problem as an interpolation problem, since the points have contributions to neighboring time bins.", + "bbox": [ + 75, + 359, + 468, + 404 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, when the spacing between adjacent detection points is large, conventional interpolation methods will fail due to the violation of the assumption. This phenomenon is very common in practical applications. To solve this problem, Nam et al. [30] employs two SPAD arrays to meet the assumption, but brings additional expense to the system. Another solution is brought up by Yang et al. [42], which combines a micro-mirror device with the compressed sensing strategy to increase detection efficiency. However, this method is sensitive to the geometry of the visible wall, which limits its application.", + "bbox": [ + 75, + 406, + 468, + 571 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Instead of introducing new detection systems, we take advantage of the fast development of deep learning methods and adopt a neural network to learn the nonlinear operator $\\Phi$ . Details about the network will be introduced in the next section. With the employed network, we can obtain a better approximation of the operator $\\Phi$ , as well as a higher reconstruction quality.", + "bbox": [ + 75, + 571, + 468, + 676 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The proposed pipeline is shown in Fig. 2. The low resolution signal $d_{l}$ is first processed by the neural network $\\Phi$ to recover the high resolution signal $\\Phi(d_{l})$ . Then the reconstruction $I$ is obtained using the existing imaging algorithm $F_{h}$ . With the proposed pipeline, we can decrease the exposure time significantly while maintaining the reconstruction quality.", + "bbox": [ + 75, + 678, + 468, + 785 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. Signal Superresolution network", + "text_level": 1, + "bbox": [ + 76, + 799, + 366, + 815 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on the analysis in the last section, we have formulated the signal recovery problem as an interpolation problem and have noticed that conventional interpolation methods may fail when the spacing between adjacent detection pairs gets larger. Besides, the contribution of one point on", + "bbox": [ + 75, + 825, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/515b81d3bb34b1b265dd6414e6a21e2e620637f7f4106cd0fafbab1039c5d4a3.jpg", + "image_caption": [ + "Figure 4. Comparisons of the mean SNR values of the results obtained with the nearest neighbor method and the proposed network on testing sets. The signals recovered by the proposed network have much higher SNR on both testing sets." + ], + "image_footnote": [], + "bbox": [ + 527, + 90, + 828, + 314 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the hidden object is contained in adjacent time bins for adjacent detection pairs. It means that the problem can not be solved by simply applying 2-D image superresolution to each time slice.", + "bbox": [ + 498, + 412, + 890, + 470 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we introduce the neural network which learns the nonlinear signal recovery operator $\\Phi$ and the training dataset. Details about the training process and the ablation study are provided in the supplement.", + "bbox": [ + 498, + 473, + 890, + 532 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Network architecture", + "text_level": 1, + "bbox": [ + 500, + 544, + 700, + 558 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Recently, attention mechanism [4, 36] has become increasingly popular in the deep learning society. However, it has been shown that there is no need to apply the attention mechanism to every part of the network [6].", + "bbox": [ + 498, + 566, + 890, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To learn the operator, we generalize the attention in attention net $(\\mathrm{A}^2\\mathrm{N})$ [6] which has two branches and dropout modules for the attention mechanism. One of the branches uses the conventional method to provide a rough estimation of the high resolution signal, and the other branch employs the network. The final output is the sum of the results from these two branches. The network architecture with two branches has been widely studied in the field of image superresolution [2, 19, 43].", + "bbox": [ + 496, + 628, + 890, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The proposed signal superresolution network (SSN) is illustrated in Fig. 3. The learning branch is composed of several convolution layers and 3-D attention in attention blocks $(3\\mathrm{D - A}^2\\mathrm{B})$ . Each block has a dropout module to learn the weight between the attention branch and the non-attention branch, which can determine whether the attention mechanism is needed in the block. The reconstruction module contains several convolution layers to reconstruct the high resolution signal through the extracted feature. The low di", + "bbox": [ + 496, + 765, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "17423", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "mensional feature is first interpolated to a high dimensional feature using trilinear interpolation, and then the feature is processed by several convolution layers. The output of the convolution layer is passed to the next part of the reconstruction module and the process is repeated.", + "bbox": [ + 75, + 90, + 468, + 166 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The generalization of the network from $\\mathrm{A}^2\\mathrm{N}$ [6] to SSN is not trivial. Based on our analysis, the contribution of one point on the hidden object is contained in adjacent time bins for adjacent detection pairs. Conventional interpolation methods will perform poorly when the spacing between adjacent detection pairs gets larger. Thus, 3-D kernels are needed to extract and recover the information, and the size of the kernels is relevant to the spacing. In addition, the signal is only interpolated in spatial dimensions instead of all three dimensions to make the operator easier to learn.", + "bbox": [ + 75, + 166, + 470, + 316 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. Dataset", + "text_level": 1, + "bbox": [ + 76, + 325, + 171, + 339 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For all experiments in this paper, we use subsets of Fashion-MNIST [40] and ShapeNet [5] to generate synthetic signals for training and testing. The Fashion-MNIST dataset consists of 2-D figures, while the ShapeNet comprises 3-D objects. We choose 5,000 samples from the training set of Fashion-MNIST and the class \"Car\" of ShapeNet respectively to generate our training set. In addition, we choose 1,000 samples from the testing set of Fashion-MNIST and the class \"Car\" of ShapeNet respectively to generate the testing set. Thus, the training set used in this work comprises 10,000 signal pairs, and the testing set comprises 2,000. All the signals are generated using Eq. (4). Besides, we set $N_{x} = N_{y} = 8$ , $N_{t} = 100$ in this paper. For measured signals whose $N_{t}$ is larger than 100, the network can also be applied directly since the network operates locally.", + "bbox": [ + 75, + 349, + 468, + 589 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For the confocal scenario, to match the experimental setting with the Stanford dataset [22], the 2-D figures from FashionMNIST [40] are placed $1\\mathrm{m}$ away from the visible wall, the size of the figures is set as $1\\times 1\\mathrm{m}^2$ . And the 3-D objects from ShapeNet [5] are placed in a $1\\times 1\\times 0.2\\mathrm{m}^3$ (horizontal, vertical, depth) box with a maximum depth of $1.1\\mathrm{m}$ . The illumination region on the visible wall is $2\\times 2\\mathrm{m}^2$ , and the time resolution of the SPAD is $32~\\mathrm{ps}$ .", + "bbox": [ + 75, + 590, + 468, + 710 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For the non-confocal setting, we match the experimental setting with the data provided by phasor field [25]. The 2-D figures from Fashion-MNIST are placed $1\\mathrm{m}$ away from the wall, and the size is $1.25\\times 1.25\\mathrm{m}^2$ . The 3-D objects from ShapeNet are placed in a $1.25\\times 1.25\\times 0.2\\mathrm{m}^3$ (horizontal, vertical, depth) box with a maximum depth of $1.1\\mathrm{m}$ . The illumination region on the visible wall is $1.25\\times 1.25\\mathrm{m}^2$ , and the time resolution of the SPAD is 16 ps.", + "bbox": [ + 75, + 710, + 468, + 832 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Results", + "text_level": 1, + "bbox": [ + 76, + 845, + 161, + 859 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we show the reconstruction results of the proposed pipeline on both synthetic and measured data. For", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c59a5991ae980e8fbc1943f5cb87dc6d2835595d0868a6aea70a94cacea9a769.jpg", + "image_caption": [ + "Figure 5. Recovered signals and reconstruction results of the pyramid. (a) Ground truth of the hidden object. (b) A comparison of the recovered signals (the first arrival time of the original signal is marked by the blue curve in the zoom-in window of each subfigure). (c) Reconstruction results of state-of-the-art methods with different signals." + ], + "image_footnote": [], + "bbox": [ + 498, + 85, + 890, + 777 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "17424", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "each instance, reconstruction results obtained with signals of different spatial resolutions are shown to validate the effectiveness of the proposed pipeline. The first column (Original) shows reconstructions obtained from the original signal with a spatial resolution of $32 \\times 32$ ; results in the second column (Low) are reconstructed using the signal scanned at $8 \\times 8$ points; the third column (Nearest) shows reconstructions of the recovered signal obtained with nearest neighbor interpolation; the last column (Ours) illustrates reconstruction results obtained with the proposed pipeline. Comparisons with other interpolation methods and existing learning based NLOS imaging method are provided in the supplement.", + "bbox": [ + 75, + 90, + 472, + 287 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To illustrate the proposed pipeline can be acted as a plugin-and-play module for most algorithms, we combine state-of-the-art methods with our pipeline. For confocal experiments, the results are reconstructed by F-K [22], LCT [31], LOG-BP [18], D-LCT [45] and SOCR [27] algorithms, while for non-confocal experiments, results are given by LOG-BP, Phasor [24] and SOCR methods. For each instance, the parameters are fixed in each algorithm.", + "bbox": [ + 75, + 287, + 468, + 409 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To assess the results quantitatively, we compute the signal to noise ratio (SNR) of the recovered signals, as well as the peak signal to noise ratio (PSNR) and structural similarity index measure (SSIM) of the reconstructed maximum intensity projections. For measured data, we use the reconstructions obtained from the original signals as references to compute PSNR and SSIM.", + "bbox": [ + 75, + 410, + 468, + 515 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Results of confocal experiment", + "text_level": 1, + "bbox": [ + 76, + 527, + 348, + 542 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Fig. 4, we compare the mean SNR values of the signals interpolated with SSN and the nearest neighbor method. It is shown that the network provides much better results on both datasets. In other words, the network approximates the oracle interpolation operator $\\Phi$ with high accuracy and serves as a crucial part of the proposed pipeline.", + "bbox": [ + 75, + 551, + 468, + 643 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To test the robustness of the proposed pipeline, we use the instance of the pyramid from SOCR [27], whose signal was generated with a different physical model. As shown in Fig. 5b, the first arrival time is correctly recovered by the proposed network, while the nearest neighbor interpolation method fails. Fig. 5c shows that the proposed pipeline provides a signal for high quality reconstructions with better visual quality and higher PSNR and SSIM values.", + "bbox": [ + 75, + 643, + 468, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The proposed pipeline is then tested on measured data from Stanford dataset [22]. This dataset contains signals measured with different hidden objects and exposure time. We choose the instance of the statue to validate the effectiveness of the proposed pipeline. The measurement time of each sampling point is $0.0069\\mathrm{s}$ . The total exposure time of the signal scanned at $32\\times 32$ points is $7.03\\mathrm{s}$ , and it only takes $0.44\\mathrm{s}$ to scan $8\\times 8$ points. As is shown in Fig. 6b, the result of SSN has higher SNR. The first arrival time can", + "bbox": [ + 75, + 763, + 470, + 900 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b052fceca843684b67e7d0eb7163a0c3e5a5e1ab70230ab06882186645278365.jpg", + "image_caption": [ + "Figure 6. Recovered signals and reconstruction results of the statue. (a) Ground truth of the hidden object. (b) A comparison of recovered signals (the first arrival time of the original signal is marked by the blue curve in the zoom-in window of each subfigure). (c) Reconstruction results of state-of-the-art methods with different signals." + ], + "image_footnote": [], + "bbox": [ + 496, + 85, + 890, + 777 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "17425", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "hardly be distinguished in the nearest neighbor interpolated signal, while it is much clearer in the signal recovered by the proposed network. As shown in Fig. 6c, the proposed pipeline results in fewer artifacts in the background. More comparisons of different exposure times are provided in the supplement, which also indicates the outperformance of the proposed pipeline.", + "bbox": [ + 75, + 90, + 472, + 196 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Results of the dynamic scene", + "text_level": 1, + "bbox": [ + 76, + 207, + 333, + 223 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To illustrate the potential application of the proposed pipeline on real-time NLOS imaging, we test it on the captured dynamic scenes from the Stanford dataset. The data is captured over a $2 \\times 2\\mathrm{m}^2$ region on the visible wall, and the person is wearing a retroreflective suit to increase the detection of the indirect signal. It takes $0.25\\mathrm{s}$ to scan $32 \\times 32$ points, which allows for a reconstruction video with 4 frames per second. The reflection model of the retroreflective material [31] is different from the one defined in Eq. (6).", + "bbox": [ + 75, + 231, + 468, + 381 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The results of several frames are shown in Fig. 7b, the complete video of the reconstruction results can be found in the supplement. All the results are reconstructed by the F-K method. Although the training set is generated by a different forward model, the proposed pipeline can still recover signals for high quality reconstructions. The pose of the person can be clearly identified in the reconstructions using the proposed pipeline, while the reconstructions using the nearest neighbor interpolated signal contain more artifacts. The mean PSNR and SSIM of the reconstruction results obtained from the proposed pipeline are 24.46 and 0.63, while those obtained with nearest neighbor interpolation are only 18.68 and 0.37.", + "bbox": [ + 75, + 382, + 470, + 575 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The total capture time of each frame is reduced from 0.25 s to 0.0156 s. This enables us to obtain a high quality reconstruction video with 64 frames per second. Besides, it only takes 0.08 s to recover the high resolution signal with the proposed network, and 0.01s to reconstruct the hidden object with the F-K method on an NVIDIA RTX 3090. Thus, real-time video imaging with a high frame rate can be realized with a delay of a few frames.", + "bbox": [ + 75, + 579, + 468, + 699 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Results of non-confocal experiment", + "text_level": 1, + "bbox": [ + 76, + 710, + 382, + 726 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For the non-confocal setting, we use the instance of \"4\" provided by Phasor [24] to validate the effectiveness of the proposed method. The measured signal is subsampled to $32 \\times 32$ and the illumination region on the visible wall is cropped to $1.25 \\times 1.25\\mathrm{m}^2$ . We use the signal scanned at 8 $\\times 8$ points as the input of our pipeline. It takes 1 s to scan each point on the visible wall.", + "bbox": [ + 75, + 734, + 468, + 838 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As is shown in Fig. 8b, the SNR of the signal recovered by the network is much higher than the one interpolated with the nearest neighbor method. Reconstruction results are shown in Fig. 8c. The proposed pipeline provides re", + "bbox": [ + 75, + 839, + 468, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4cf3f4389b401b9c7c3cf6ed7462d1ca6b347a6759fec99723260566fe74fdd9.jpg", + "image_caption": [ + "(a) GT" + ], + "image_footnote": [], + "bbox": [ + 617, + 88, + 774, + 172 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1ce59b218de5a7c16fda97620e7f7cf32a22da1c216a90d19c35cdeaf93deb5a.jpg", + "image_caption": [ + "(b) Reconstruction results", + "Figure 7. Reconstruction results of the dynamic scene. (a) An illustration of the dynamic scene. (b) Reconstruction results of the dynamic scene. The results are obtained with the F-K method. The PSNR and SSIM values are computed with the reconstructions obtained with the original signals as references. The proposed pipeline provides much better reconstructions in terms of both quantitative criteria and visual quality." + ], + "image_footnote": [], + "bbox": [ + 506, + 191, + 890, + 523 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "constructions with higher quantitative and visual quality.", + "bbox": [ + 500, + 675, + 872, + 691 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Discussion", + "text_level": 1, + "bbox": [ + 500, + 705, + 612, + 720 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We have introduced a novel pipeline for high quality NLOS reconstruction with the signal measured at only a small number of scanning points. In this section, we discuss several factors that affect the reconstruction quality and some feasible solutions to improve the pipeline further.", + "bbox": [ + 496, + 729, + 890, + 806 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Training of the network", + "text_level": 1, + "bbox": [ + 500, + 816, + 718, + 832 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The training of the network employed in the proposed pipeline is supervised. Thus, it is affected by the generation of the training set, including the forward model, experimental setting, and hidden objects. For example, hidden objects", + "bbox": [ + 496, + 839, + 892, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "17426", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "with different materials will lead to various physical models. The networks trained and tested on signals with different models would bring additional errors to the results.", + "bbox": [ + 75, + 90, + 468, + 136 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "However, a network with good generalization ability can avoid degradation. As shown in Fig. 7, although the physical model of the dynamic scenes differs from the one used in the training set, the reconstructed objects are only slightly affected. Thus, the proposed pipeline has strong robustness to the forward model.", + "bbox": [ + 75, + 136, + 468, + 224 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2. Performance of different algorithms", + "text_level": 1, + "bbox": [ + 76, + 234, + 390, + 250 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As the reconstruction results shown in previous sections, the improvement of reconstruction quality of the proposed pipeline varies between different methods. Compared to the reconstruction results obtained from the low resolution signal, fast imaging algorithms like F-K, LCT and D-LCT have larger improvements, while the iterative inversion algorithm SOCR might have similar visual quality in some cases. This phenomenon may come from the fact that the iterative algorithm can make full use of the information contained in the signal, but with a longer reconstruction time.", + "bbox": [ + 75, + 258, + 468, + 409 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Combination with SPAD array for further acceleration", + "text_level": 1, + "bbox": [ + 76, + 417, + 468, + 446 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Since the proposed pipeline can recover the signal faithfully under the non-confocal setting, combining it with a SPAD array can accelerate the measurement process further. However, the existing SPAD array can only focus on a rectangular area on the visible wall, which means that the network has to learn an extrapolation operator. This is much harder than learning an interpolation operator. Thus, it would be more practical if the SPAD array could focus on sparse rectangular points on a relatively large area of the visible wall. This may be seen as a new area for hardware design.", + "bbox": [ + 75, + 455, + 468, + 619 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 633, + 194, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we have studied the signal recovery problem in NLOS imaging, and proposed a two-step learning based pipeline for fast NLOS imaging with only a small number of scanning points. In the first step, the signal of high spatial resolution is recovered with a neural network. In the second step, the hidden objects are obtained from the virtually measured signals at a dense grid. Our pipeline provides high-quality reconstructions under both confocal and non-confocal settings, which takes only $6.25\\%$ of the original measurement time. Besides, the proposed pipeline is compatible with most existing detection systems and imaging algorithms. With the proposed pipeline, the high frame rate imaging video is possible, which is a noteworthy advance in this area. It is expected to combine emerging techniques in the fields of SPAD array for real-time and high quality NLOS reconstructions with higher frame rates.", + "bbox": [ + 75, + 659, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6d02ab8bd732e183ac0ccfaf3814939689e1178c80bfa33a1c3a7f6b82a5e7bc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 660, + 90, + 736, + 169 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/98394bb3ec72c783acb60f1f019e86e01f3c3ac1300f249bb043f42b8cecddc1.jpg", + "image_caption": [ + "(a) GT" + ], + "image_footnote": [], + "bbox": [ + 503, + 186, + 890, + 303 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b80a4457d701df0888351b443cdbef191820a0492924a2217592583be92c4abe.jpg", + "image_caption": [ + "(b) Recovered signals", + "(c) Reconstruction results", + "Figure 8. Recovered signals and reconstruction results of the instance \"4\". (a) Ground truth of the hidden object. (b) A comparison of recovered signals (the first arrival time of the original signal is marked by the blue curve in the zoom-in window of each subfigure). (c) Reconstruction results of state-of-the-art methods with different signals." + ], + "image_footnote": [], + "bbox": [ + 506, + 320, + 887, + 601 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments. The authors would like to thank Zhu-peng Li and Rongqian Wang for fruitful discussions. This work was supported by the National Natural Science Foundation of China (61975087, 12071244, 11971258).", + "bbox": [ + 498, + 750, + 890, + 809 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 500, + 830, + 594, + 845 + ], + "page_idx": 7 + }, + { + "type": "ref_text", + "text": "[1] Byeongjoo Ahn, Akshat Dave, Ashok Veeraraghavan, Ioannis Gkioulekas, and Aswin C Sankaranarayanan. Convolutional approximations to the general non-line-of-sight imag", + "bbox": [ + 506, + 858, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "17427", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ing operator. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7889-7899, 2019. 2", + "[2] Saeed Anwar, Salman Khan, and Nick Barnes. A deep journey into super-resolution: A survey. ACM Computing Surveys (CSUR), 53(3):1-34, 2020. 4", + "[3] Victor Arellano, Diego Gutierrez, and Adrian Jarabo. Fast back-projection for non-line of sight reconstruction. In ACM SIGGRAPH 2017 Posters, SIGGRAPH '17, New York, NY, USA, 2017. Association for Computing Machinery. 2", + "[4] Syed Muhammad Arsalan Bashir, Yi Wang, Mahrukh Khan, and Yilong Niu. A comprehensive review of deep learning-based single image super-resolution. PeerJ Computer Science, 7:e621, 2021. 4", + "[5] Angel X. Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. ShapeNet: An Information-Rich 3D Model Repository. Technical Report arXiv:1512.03012 [cs.GR], Stanford University — Princeton University — Toyota Technological Institute at Chicago, 2015. 5", + "[6] Haoyu Chen, Jinjin Gu, and Zhi Zhang. Attention in attention network for image super-resolution. arXiv preprint arXiv:2104.09497, 2021. 4", + "[7] Wenzheng Chen, Simon Daneau, Fahim Mannan, and Felix Heide. Steady-state non-line-of-sight imaging. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2019. 2", + "[8] Wenzheng Chen, Fangyin Wei, Kiriakos N Kutulakos, Szymon Rusinkiewicz, and Felix Heide. Learned feature embeddings for non-line-of-sight imaging and recognition. ACM Transactions on Graphics (TOG), 39(6):1-18, 2020. 2", + "[9] Justin Dove and Jeffrey H. Shapiro. Paraxial theory of phasor-field imaging. Opt. Express, 27(13):18016-18037, Jun 2019. 2", + "[10] Xiaohua Feng and Liang Gao. Improving non-line-of-sight image reconstruction with weighting factors. Optics letters, 45(14):3921-3924, 2020. 2", + "[11] Xiaohua Feng and Liang Gao. Ultrafast light field tomography for snapshot transient and non-line-of-sight imaging. Nature Communications, 12:2179, 04 2021. 2", + "[12] Javier Grau Chopite, Matthias B. Hullin, Michael Wand, and Julian Iseringhausen. Deep non-line-of-sight reconstruction. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 957-966, 2020. 2", + "[13] Felix Heide, Matthew O'Toole, Kai Zang, David B. Lindell, Steven Diamond, and Gordon Wetzstein. Non-line-of-sight imaging with partial occluders and surface normals. ACM Trans. Graph., 38(3), May 2019. 2", + "[14] Julian Iseringhausen and Matthias B Hullin. Non-line-of-sight reconstruction using efficient transient rendering. ACM Transactions on Graphics (TOG), 39(1):1-14, 2020. 2", + "[15] Mariko Isogawa, Dorian Chan, Ye Yuan, Kris M. Kitani, and Matthew O'Toole. Efficient non-line-of-sight imaging from transient sinograms. In 16th European Conference on Computer Vision (ECCV), 2020. 2, 3", + "[16] Mariko Isogawa, Ye Yuan, Matthew O'Toole, and Kris Kitani. Optical non-line-of-sight physics-based 3d human pose" + ], + "bbox": [ + 78, + 92, + 470, + 902 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "estimation. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7011-7020, 2020. 2", + "[17] Ahmed Kirmani, Tyler Hutchison, James Davis, and Ramesh Raskar. Looking around the corner using transient imaging. In 2009 IEEE 12th International Conference on Computer Vision, pages 159–166, 2009. 1", + "[18] Martin Laurenzis and Andreas Velten. Feature selection and back-projection algorithms for nonlinear-of-sight laser-gated viewing. Journal of Electronic Imaging, 23:063003, 11 2014. 2, 3, 6", + "[19] Y. Li, B. Sixou, and F. Peyrin. A review of the deep learning methods for medical images super resolution problems. IRBM, 42(2):120-133, 2021. 4", + "[20] Zhupeng Li, Xintong Liu, Jianyu Wang, Zuoqiang Shi, Lingyun Qiu, and Xing Fu. Fast non-line-of-sight imaging based on first photon event stamping. Optics letters, 47(8):1928—1931, April 2022. 2", + "[21] Zhengpeng Liao, Deyang Jiang, Xiaochun Liu, Andreas Velten, Yajun Ha, and Xin Lou. FPGA accelerator for real-time non-line-of-sight imaging. IEEE Transactions on Circuits and Systems I: Regular Papers, 2022. 2", + "[22] David B Lindell, Gordon Wetzstein, and Matthew O'Toole. Wave-based non-line-of-sight imaging using fast fk migration. ACM Transactions on Graphics (ToG), 38(4):1-13, 2019. 2, 3, 5, 6", + "[23] Jianjiang Liu, Yijun Zhou, Xin Huang, Zheng-Ping Li, and Feihu Xu. Photon-efficient non-line-of-sight imaging. IEEE Transactions on Computational Imaging, 8:639-650, 2022. 2", + "[24] Xiaochun Liu, Sebastian Bauer, and Andreas Velten. Phasor field diffraction based reconstruction for fast non-line-of-sight imaging systems. Nature Communications, 11:1645, 2020. 2, 3, 6, 7", + "[25] Xiaochun Liu, Ibón Guillén, Marco Manna, Ji Nam, Azer Reza, Toan Le, Adrián Jarabo, Diego Gutiérrez, and Andreas Velten. Non-line-of-sight imaging using phasor-field virtual wave optics. Nature, 572, 2019. 2, 5", + "[26] Xiaochun Liu and Andreas Velten. The role of wigner distribution function in non-line-of-sight imaging. In 2020 IEEE International Conference on Computational Photography (ICCP), pages 1-12. IEEE, 2020. 2", + "[27] Xintong Liu, Jianyu Wang, Zhupeng Li, Zuoqiang Shi, Xing Fu, and Lingyun Qiu. Non-line-of-sight reconstruction with signal-object collaborative regularization. Light: Science & Applications, 08 2021. 2, 3, 6", + "[28] Christopher Metzler, David Lindell, and Gordon Wetzstein. Keyhole imaging: Non-line-of-sight imaging and tracking of moving objects along a single optical path. IEEE Transactions on Computational Imaging, PP:1-1, 12 2020. 2, 3", + "[29] Christopher A. Metzler, Felix Heide, Prasana Rangarajan, Muralidhar Madabhushi Balaji, Aparna Viswanath, Ashok Veeraraghavan, and Richard G. Baraniuk. Deep-inverse correlography: towards real-time high-resolution non-line-of-sight imaging. Optica, 7(1):63-71, Jan 2020. 2", + "[30] Ji Hyun Nam, Eric Brandt, Sebastian Bauer, Xiaochun Liu, Marco Renna, Alberto Tosi, Eftychios Sifakis, and Andreas" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "17428", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Velten. Low-latency time-of-flight non-line-of-sight imaging at 5 frames per second. Nature communications, 12(1):1-10, 2021. 2, 4", + "[31] Matthew O'Toole, David Lindell, and Gordon Wetzstein. Confocal non-line-of-sight imaging based on the light-cone transform. Nature, 555, 2018. 2, 3, 6, 7", + "[32] Chengquan Pei, Anke Zhang, Yue Deng, Feihu Xu, Jiamin Wu, David U-Lei Li, Hui Qiao, Lu Fang, and Qionghai Dai. Dynamic non-line-of-sight imaging system based on the optimization of point spread functions. Opt. Express, 29(20):32349-32364, Sep 2021. 2", + "[33] Syed Azer Reza, Marco La Manna, Sebastian Bauer, and Andreas Velten. Phasor field waves: A huygens-like light transport model for non-line-of-sight imaging applications. Opt. Express, 27(20):29380-29400, Sep 2019. 2", + "[34] Christos Thrampoulidis, Gal Shulkind, Feihu Xu, William T. Freeman, Jeffrey H. Shapiro, Antonio Torralba, Franco N. C. Wong, and Gregory W. Wornell. Exploiting occlusion in non-line-of-sight active imaging. IEEE Transactions on Computational Imaging, 4(3):419-431, 2018. 2", + "[35] Chia-Yin Tsai, Aswin C Sankaranarayanan, and Ioannis Gkioulekas. Beyond volumetric albedo-a surface optimization framework for non-line-of-sight imaging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1545-1555, 2019. 2", + "[36] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Proceedings of the 31st International Conference on Neural Information Processing Systems, NIPS'17, page 6000-6010, Red Hook, NY, USA, 2017. Curran Associates Inc. 4", + "[37] Eric Veach and Leonidas J. Guibas. Optimally combining sampling techniques for monte carlo rendering. In Proceedings of the 22nd Annual Conference on Computer Graphics and Interactive Techniques, SIGGRAPH '95, page 419-428, New York, NY, USA, 1995. Association for Computing Machinery. 3", + "[38] Andreas Velten, Thomas Willwacher, Otkrist Gupta, Ashok Veeraraghavan, Moungi G Bawendi, and Ramesh Raskar. Recovering three-dimensional shape around a corner using ultrafast time-of-flight imaging. Nature communications, 3(1):1-8, 2012. 2", + "[39] Cheng Wu, Jianjiang Liu, Xin Huang, Zheng-Ping Li, Chao Yu, Jun-Tian Ye, Jun Zhang, Qiang Zhang, Xiankang Dou, Vivek Goyal, Feihu Xu, and Jian-Wei Pan. Non-line-of-sight imaging over $1.43\\mathrm{km}$ . Proceedings of the National Academy of Sciences of the United States of America, 118, 03 2021. 2", + "[40] Han Xiao, Kashif Rasul, and Roland Vollgraf. Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms. arXiv preprint arXiv:1708.07747, 2017. 5", + "[41] Shumian Xin, Sotiris Nousias, Kiriakos N. Kutulakos, Aswin C. Sankaranarayanan, Srinivasa G. Narasimhan, and Ioannis Gkioulekas. A theory of fermat paths for non-line-of-sight shape reconstruction. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6793-6802, 2019. 2" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[42] Wenqing Yang, Chao Zhang, Wenjie Jiang, Zexin Zhang, and Baoqing Sun. None-line-of-sight imaging enhanced with spatial multiplexing. Opt. Express, 30(4):5855-5867, Feb 2022. 2, 4", + "[43] Wenming Yang, Xuechen Zhang, Yapeng Tian, Wei Wang, Jing-Hao Xue, and Qingmin Liao. Deep learning for single image super-resolution: A brief review. IEEE Transactions on Multimedia, 21(12):3106-3121, Dec 2019. 4", + "[44] Jun-Tian Ye, Xin Huang, Zheng-Ping Li, and Feihu Xu. Compressed sensing for active non-line-of-sight imaging. Opt. Express, 29(2):1749-1763, Jan 2021. 2, 3", + "[45] Sean I Young, David B Lindell, Bernd Girod, David Taubman, and Gordon Wetzstein. Non-line-of-sight surface reconstruction using the directional light-cone transform. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1407-1416, 2020. 2, 3, 6" + ], + "bbox": [ + 501, + 92, + 890, + 328 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "17429", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Non-Line-of-Sight Imaging With Signal Superresolution Network/224cfcfa-2657-4e74-b370-35626fe2670d_model.json b/2023/Non-Line-of-Sight Imaging With Signal Superresolution Network/224cfcfa-2657-4e74-b370-35626fe2670d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c30253eac6dbf57d4052160911eda7b4d87c156a --- /dev/null +++ b/2023/Non-Line-of-Sight Imaging With Signal Superresolution Network/224cfcfa-2657-4e74-b370-35626fe2670d_model.json @@ -0,0 +1,2112 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.165, + 0.131, + 0.806, + 0.154 + ], + "angle": 0, + "content": "Non-line-of-sight Imaging with Signal Superresolution Network" + }, + { + "type": "text", + "bbox": [ + 0.124, + 0.18, + 0.844, + 0.217 + ], + "angle": 0, + "content": "Jianyu Wang\\(^{1}\\), Xintong Liu\\(^{1}\\), Leping Xiao\\(^{1}\\), Zuoqiang Shi\\(^{1,2}\\), Lingyun Qiu\\(^{1,2}\\), and Xing Fu\\(^{1}\\) \n\\(^{1}\\)Tsinghua University" + }, + { + "type": "text", + "bbox": [ + 0.192, + 0.217, + 0.773, + 0.235 + ], + "angle": 0, + "content": "\\(^{2}\\)Yanqi Lake Beijing Institute of Mathematical Sciences and Applications" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.302, + 0.474, + 0.65 + ], + "angle": 0, + "content": "Non-line-of-sight (NLOS) imaging aims at reconstructing the location, shape, albedo, and surface normal of the hidden object around the corner with measured transient data. Due to its strong potential in various fields, it has drawn much attention in recent years. However, long exposure time is not always available for applications such as auto-driving, which hinders the practical use of NLOS imaging. Although scanning fewer points can reduce the total measurement time, it also brings the problem of imaging quality degradation. This paper proposes a general learning-based pipeline for increasing imaging quality with only a few scanning points. We tailor a neural network to learn the operator that recovers a high spatial resolution signal. Experiments on synthetic and measured data indicate that the proposed method provides faithful reconstructions of the hidden scene under both confocal and nonconfocal settings. Compared with original measurements, the acquisition of our approach is 16 times faster while maintaining similar reconstruction quality. Besides, the proposed pipeline can be applied directly to existing optical systems and imaging algorithms as a plug-in-and-play module. We believe the proposed pipeline is powerful in increasing the frame rate in NLOS video imaging." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.678, + 0.21, + 0.695 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.47, + 0.839 + ], + "angle": 0, + "content": "Non-line-of-sight (NLOS) imaging problem usually employs a high temporal resolution optical system to recover the hidden object around the corner. As shown in Fig. 1a, photons emitted by the laser are collected by the detector after three diffuse reflections: the reflection at the visible wall, the reflection at the hidden object, and the reflection at the visible wall again. Scanning a region on the visible wall can get measured histogram data to recover the hidden scene." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Due to its potential applications in various fields, such as auto-driving, disaster relief, and remote sensing, the NLOS imaging problem has become an emerging field since it was first proposed by Kirmani et al. [17] in 2009." + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.268, + 0.885, + 0.478 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.645, + 0.483, + 0.748, + 0.495 + ], + "angle": 0, + "content": "(a) NLOS scenario" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.502, + 0.692, + 0.638 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.549, + 0.641, + 0.655, + 0.653 + ], + "angle": 0, + "content": "(b) Scanning points" + }, + { + "type": "image", + "bbox": [ + 0.702, + 0.497, + 0.779, + 0.565 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.701, + 0.569, + 0.752, + 0.578 + ], + "angle": 0, + "content": "Low" + }, + { + "type": "image", + "bbox": [ + 0.783, + 0.497, + 0.883, + 0.639 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.762, + 0.644, + 0.822, + 0.656 + ], + "angle": 0, + "content": "(c) Results" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.668, + 0.894, + 0.849 + ], + "angle": 0, + "content": "Figure 1. A typical NLOS scenario and reconstruction results. (a) An illustration of a typical NLOS scenario. (b) An illustration of the scanning points on the visible wall. Using the proposed pipeline, only the blue circle points are needed to be illuminated, and the signal at the yellow square points can be recovered with the proposed pipeline. (c) Comparisons of the reconstruction results of the pyramid with different signals. The maximum intensity projection of the albedo values along the depth direction is shown in the upper left corner (GT). The results are reconstructed by the original signal with spatial resolution \\(32 \\times 32\\) (Original), the subsampled signal with spatial resolution \\(8 \\times 8\\) (Low), and the signal recovered by the proposed network with spatial resolution \\(32 \\times 32\\) (Ours)." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Many methods have been proposed to improve the prac" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "17420" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.131, + 0.089, + 0.844, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.238, + 0.895, + 0.268 + ], + "angle": 0, + "content": "Figure 2. Flowchart of the proposed pipeline. The high resolution signal is recovered from the low resolution signal using a neural network. The hidden scene is then reconstructed using state-of-the-art imaging algorithms designed for high resolution signal." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.292, + 0.473, + 0.64 + ], + "angle": 0, + "content": "ticability [1, 3, 13, 14, 34, 35, 39] or reconstruction quality [10, 11, 26, 41]. A back-projection method with Laplacian of Gaussian filter (LOG-BP) [18] is introduced to improve the reconstruction quality of the back-projection method [38]. In 2018, O'Toole et al. [31] first apply the light-contransform (LCT) method to image the hidden object with a confocal scanning mode. Young et al. [45] then propose the directional light-cone-transform (D-LCT) method, which simultaneously recovers the albedo and surface normal of the hidden object. From the perspective of wave characteristics, Lindell et al. [22] introduce the fast frequency-wavenumber migration (F-K) method to NLOS. Another remarkable imaging method proposed by Liu et al. [25] employs the phasor field model [9, 33] and provides an efficient solution for fast NLOS imaging under non-confocal scenario [24]. Further improvements of this method using single photon avalanche diode (SPAD) arrays achieve real-time imaging [21, 30]. Deep learning based methods have also been introduced [7, 8, 12, 16, 29]. Recently, Liu et al. [27] propose the signal-object collaborative regularization (SOCR) method, which can improve the resolution of the reconstructed object, especially in cases with high measurement noise." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.645, + 0.473, + 0.903 + ], + "angle": 0, + "content": "However, one main obstacle in NLOS imaging is that we need a long exposure time to improve the signal-to-noise ratio (SNR) of the measurement, which is not admissible in many scenarios, such as auto-driving. Several methods have been proposed to reduce the capture time [20, 23]. For a confocal scenario, Isogawa et al. [15] propose a circular scanning mode. In 2021, Metzler et al. [28] consider an extreme case that tracks the objects with a single detection point. By introducing the downsampling operation, Ye et al. [44] reconstruct the object with a subset of the virtual sources. For non-confocal scenarios, most methods employ SPAD arrays to realize fast measurement. Among them, Nam et al. [30] achieve a reconstruction of five frames per second by remapping the data from two \\(16 \\times 1\\) SPAD arrays. Pei et al. [32] realize 20 frames per second reconstruction using a \\(32 \\times 32\\) SPAD array. In 2022, Yang et al. [42] propose a spatial multiplexing method based on the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.293, + 0.892, + 0.352 + ], + "angle": 0, + "content": "compressed sensing strategy. Although these methods can reduce the measurement time, they might either cause deterioration in imaging quality or substantially increase expense." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.353, + 0.892, + 0.459 + ], + "angle": 0, + "content": "In this paper, we propose a deep learning based pipeline that can reconstruct the hidden object with fine details using only a small number of scanning points. Different measurement densities are shown in Fig. 1b. In the following sections, we term the signal measured at the blue circle points as the \"low resolution signal\" while those measured at both circle and square points as the \"high resolution signal\"." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.459, + 0.893, + 0.655 + ], + "angle": 0, + "content": "Instead of designing new algorithms for sparse measurements, we focus on recovering the dense virtual signals, which can be applied to most algorithms as a plug-in-and-play module. The proposed pipeline consists of two steps: the low resolution signal is first mapped into the high resolution signal space by a neural network; the reconstruction is then obtained from the recovered high resolution signal using existing imaging algorithms. The reconstruction results of the pyramid are shown in Fig. 1c. These results are reconstructed by the F-K method. As the comparison shows, the proposed pipeline can provide a high quality reconstruction with only 64 scanning points, which is \\(6.25\\%\\) of the original number." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.656, + 0.892, + 0.746 + ], + "angle": 0, + "content": "To our best knowledge, this is the first work in NLOS to study the signal recovery problem of the low resolution signal sparsely measured on a uniform grid. The proposed method can reduce the exposure time without losing the reconstruction quality. The main contributions of this work are:" + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.757, + 0.892, + 0.803 + ], + "angle": 0, + "content": "- We propose a novel learning based pipeline for faithful reconstruction of the hidden scene, which only takes \\(6.25\\%\\) of the original measuring time." + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.814, + 0.892, + 0.859 + ], + "angle": 0, + "content": "- We show that the high resolution signal can be recovered accurately with the proposed signal superresolution network, which we abbreviate as \"SSN\"." + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "- The effectiveness of the proposed method is validated on public datasets under both confocal and non" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.757, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "17421" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.134, + 0.092, + 0.844, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.307, + 0.895, + 0.337 + ], + "angle": 0, + "content": "Figure 3. The architecture of the signal superresolution network (SSN) in the proposed pipeline. The network is composed of two main branches: a conventional upsampling branch and a learning branch. The final output is the sum of the results from these two branches." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.362, + 0.232, + 0.379 + ], + "angle": 0, + "content": "confocal settings." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.387, + 0.471, + 0.434 + ], + "angle": 0, + "content": "- We demonstrate the potential application of the proposed method in high frame rate (64 frames per second) NLOS video imaging using public datasets." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.446, + 0.304, + 0.464 + ], + "angle": 0, + "content": "2. Reconstruction Pipeline" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.471, + 0.47, + 0.548 + ], + "angle": 0, + "content": "This work aims to develop a learning based pipeline for NLOS imaging, which can reduce the measurement time while maintaining the reconstruction quality. In this section, we formulate the signal recovery problem and introduce the proposed pipeline." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.556, + 0.329, + 0.573 + ], + "angle": 0, + "content": "2.1. Formulation of the problem" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.579, + 0.47, + 0.61 + ], + "angle": 0, + "content": "Most existing methods that reconstruct the hidden object from low resolution signals \\( d_{l} \\) can be formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.208, + 0.621, + 0.47, + 0.639 + ], + "angle": 0, + "content": "\\[\nI (x, y, z) = F _ {l} \\left(d _ {l}\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.649, + 0.47, + 0.739 + ], + "angle": 0, + "content": "in which \\( I(x,y,z) \\) is the reconstructed object and \\( F_{l} \\) is the imaging algorithm designed for low resolution signal [15, 28, 44]. Since only a few sampling points are scanned, the total exposure time can be reduced. However, the reconstruction quality also decreases with the number of scan points." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.74, + 0.469, + 0.77 + ], + "angle": 0, + "content": "On the contrary, the algorithms designed for high resolution signal can be expressed as" + }, + { + "type": "equation", + "bbox": [ + 0.205, + 0.782, + 0.47, + 0.8 + ], + "angle": 0, + "content": "\\[\nI (x, y, z) = F _ {h} \\left(d _ {h}\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.47, + 0.87 + ], + "angle": 0, + "content": "in which \\( F_{h} \\) represents the algorithm designed for high resolution signal \\( d_{h} \\). Many works have been proposed for fast [22, 24, 31, 45] and high resolution [18, 27] reconstruction as discussed above." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.471, + 0.902 + ], + "angle": 0, + "content": "In this work, we aim at designing a pipeline that can provide faithful reconstruction with only a small number of" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.362, + 0.892, + 0.452 + ], + "angle": 0, + "content": "scanning points, and has good compatibility with traditional optical systems as well as existing imaging algorithms. To solve this problem, we introduce a signal recovering operator \\(\\Phi\\), which maps the low resolution signal \\(d_{l}\\) into the space of high resolution signal \\(d_{h}\\). The proposed pipeline can be expressed as" + }, + { + "type": "equation", + "bbox": [ + 0.617, + 0.459, + 0.892, + 0.477 + ], + "angle": 0, + "content": "\\[\nI (x, y, z) = F _ {h} \\left(\\Phi \\left(d _ {l}\\right)\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.49, + 0.893, + 0.552 + ], + "angle": 0, + "content": "in which \\(\\Phi(d_l)\\) is of the same size with \\(d_h\\). Thus, the problem of NLOS imaging from low resolution signal converts to the problem of finding the operator \\(\\Phi\\) which satisfies \\(\\Phi(d_l) = d_h\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.567, + 0.796, + 0.584 + ], + "angle": 0, + "content": "2.2. The choice of recovering operator" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.593, + 0.892, + 0.655 + ], + "angle": 0, + "content": "The choice of the operator \\(\\Phi\\) is highly dependent on the physical model of the measurement process. We adopt a simplified version of the widely used three-point transport model [37] as" + }, + { + "type": "equation", + "bbox": [ + 0.513, + 0.668, + 0.892, + 0.712 + ], + "angle": 0, + "content": "\\[\n\\tau (\\mathbf {x} _ {l}, \\mathbf {x} _ {d}, t) = \\int_ {S} T (\\mathbf {x} _ {l}, \\mathbf {x} _ {d}, t, \\mathbf {s}) C (\\mathbf {x} _ {l}, \\mathbf {x} _ {d}, \\mathbf {s}) f (\\mathbf {s}) d A (\\mathbf {s}) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.712, + 0.892, + 0.802 + ], + "angle": 0, + "content": "in which \\(\\tau\\) is the photon intensity detected with the virtual source \\(\\mathbf{x}_l\\), virtual detector \\(\\mathbf{x}_d\\) and time \\(t\\). \\(f(\\mathbf{s})\\) represents the albedo of the point \\(\\mathbf{s}\\) on the hidden object \\(S\\) and \\(A(\\mathbf{s})\\) is the corresponding area measure. \\(T\\) is a function that characterizes the arrival time of the photons, which is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.838, + 0.892, + 0.855 + ], + "angle": 0, + "content": "\\[\nT (\\mathbf {x} _ {l}, \\mathbf {x} _ {d}, t, \\mathbf {s}) = \\delta (| | \\mathbf {x} _ {l} - \\mathbf {s} | | _ {2} + | | \\mathbf {x} _ {d} - \\mathbf {s} | | _ {2} - c t) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "in which \\(\\delta\\) is the Dirac function and \\(c\\) is the light speed. Besides, the function \\(C\\) characterizes the intensity attenuation" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "17422" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.373, + 0.107 + ], + "angle": 0, + "content": "in the physical model, which is expressed as" + }, + { + "type": "equation", + "bbox": [ + 0.088, + 0.117, + 0.47, + 0.187 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} C \\left(\\mathbf {x} _ {l}, \\mathbf {x} _ {d}, \\mathbf {s}\\right) = \\frac {< \\mathbf {s} - \\mathbf {x} _ {l} , \\mathbf {n} (\\mathbf {x} _ {l}) > < \\mathbf {x} _ {l} - \\mathbf {s} , \\mathbf {n} (\\mathbf {s}) >}{\\left| \\left| \\mathbf {x} _ {l} - \\mathbf {s} \\right| \\right| _ {2} ^ {4}} (6) \\\\ \\times \\frac {< \\mathbf {s} - \\mathbf {x} _ {d} , \\mathbf {n} (\\mathbf {x} _ {d}) > < \\mathbf {x} _ {d} - \\mathbf {s} , \\mathbf {n} (\\mathbf {s}) >}{\\left| \\left| \\mathbf {x} _ {d} - \\mathbf {s} \\right| \\right| _ {2} ^ {\\frac {4}}}. (6) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.198, + 0.469, + 0.242 + ], + "angle": 0, + "content": "in which \\(\\mathbf{n}(\\mathbf{s})\\) is the unit normal vector pointing toward the visible wall at point \\(\\mathbf{s}\\) and \\(<\\cdot, \\cdot>\\) is the inner product in \\(\\mathbb{R}^3\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.244, + 0.469, + 0.333 + ], + "angle": 0, + "content": "Suppose that there are two pairs \\((\\mathbf{x}_{l_1},\\mathbf{x}_{d_1}),(\\mathbf{x}_{l_2},\\mathbf{x}_{d_2})\\) on the visible wall. For a fixed point s on the hidden object, we assume that the contributions of this point to the signal at these two pairs are contained in time bin \\(t_1\\) and \\(t_2\\) respectively. Then from Eq. (5), we know that \\(t_1\\) and \\(t_2\\) are close if" + }, + { + "type": "equation", + "bbox": [ + 0.152, + 0.336, + 0.469, + 0.353 + ], + "angle": 0, + "content": "\\[\nD = \\left\\| \\mathbf {x} _ {l _ {1}} - \\mathbf {x} _ {l _ {2}} \\right\\| _ {2} + \\left\\| \\mathbf {x} _ {d _ {1}} - \\mathbf {x} _ {d _ {2}} \\right\\| _ {2} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.36, + 0.469, + 0.405 + ], + "angle": 0, + "content": "is small. We can then regard the signal recovery problem as an interpolation problem, since the points have contributions to neighboring time bins." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.407, + 0.469, + 0.572 + ], + "angle": 0, + "content": "However, when the spacing between adjacent detection points is large, conventional interpolation methods will fail due to the violation of the assumption. This phenomenon is very common in practical applications. To solve this problem, Nam et al. [30] employs two SPAD arrays to meet the assumption, but brings additional expense to the system. Another solution is brought up by Yang et al. [42], which combines a micro-mirror device with the compressed sensing strategy to increase detection efficiency. However, this method is sensitive to the geometry of the visible wall, which limits its application." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.573, + 0.469, + 0.678 + ], + "angle": 0, + "content": "Instead of introducing new detection systems, we take advantage of the fast development of deep learning methods and adopt a neural network to learn the nonlinear operator \\(\\Phi\\). Details about the network will be introduced in the next section. With the employed network, we can obtain a better approximation of the operator \\(\\Phi\\), as well as a higher reconstruction quality." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.679, + 0.469, + 0.786 + ], + "angle": 0, + "content": "The proposed pipeline is shown in Fig. 2. The low resolution signal \\( d_{l} \\) is first processed by the neural network \\( \\Phi \\) to recover the high resolution signal \\( \\Phi(d_{l}) \\). Then the reconstruction \\( I \\) is obtained using the existing imaging algorithm \\( F_{h} \\). With the proposed pipeline, we can decrease the exposure time significantly while maintaining the reconstruction quality." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.8, + 0.367, + 0.816 + ], + "angle": 0, + "content": "3. Signal Superresolution network" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Based on the analysis in the last section, we have formulated the signal recovery problem as an interpolation problem and have noticed that conventional interpolation methods may fail when the spacing between adjacent detection pairs gets larger. Besides, the contribution of one point on" + }, + { + "type": "image", + "bbox": [ + 0.528, + 0.091, + 0.83, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.33, + 0.892, + 0.386 + ], + "angle": 0, + "content": "Figure 4. Comparisons of the mean SNR values of the results obtained with the nearest neighbor method and the proposed network on testing sets. The signals recovered by the proposed network have much higher SNR on both testing sets." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.413, + 0.892, + 0.472 + ], + "angle": 0, + "content": "the hidden object is contained in adjacent time bins for adjacent detection pairs. It means that the problem can not be solved by simply applying 2-D image superresolution to each time slice." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.474, + 0.892, + 0.534 + ], + "angle": 0, + "content": "In this section, we introduce the neural network which learns the nonlinear signal recovery operator \\(\\Phi\\) and the training dataset. Details about the training process and the ablation study are provided in the supplement." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.545, + 0.702, + 0.559 + ], + "angle": 0, + "content": "3.1. Network architecture" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.568, + 0.892, + 0.629 + ], + "angle": 0, + "content": "Recently, attention mechanism [4, 36] has become increasingly popular in the deep learning society. However, it has been shown that there is no need to apply the attention mechanism to every part of the network [6]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.63, + 0.892, + 0.765 + ], + "angle": 0, + "content": "To learn the operator, we generalize the attention in attention net \\((\\mathrm{A}^2\\mathrm{N})\\) [6] which has two branches and dropout modules for the attention mechanism. One of the branches uses the conventional method to provide a rough estimation of the high resolution signal, and the other branch employs the network. The final output is the sum of the results from these two branches. The network architecture with two branches has been widely studied in the field of image superresolution [2, 19, 43]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.766, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The proposed signal superresolution network (SSN) is illustrated in Fig. 3. The learning branch is composed of several convolution layers and 3-D attention in attention blocks \\((3\\mathrm{D - A}^2\\mathrm{B})\\). Each block has a dropout module to learn the weight between the attention branch and the non-attention branch, which can determine whether the attention mechanism is needed in the block. The reconstruction module contains several convolution layers to reconstruct the high resolution signal through the extracted feature. The low di" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17423" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.167 + ], + "angle": 0, + "content": "mensional feature is first interpolated to a high dimensional feature using trilinear interpolation, and then the feature is processed by several convolution layers. The output of the convolution layer is passed to the next part of the reconstruction module and the process is repeated." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.167, + 0.471, + 0.318 + ], + "angle": 0, + "content": "The generalization of the network from \\(\\mathrm{A}^2\\mathrm{N}\\) [6] to SSN is not trivial. Based on our analysis, the contribution of one point on the hidden object is contained in adjacent time bins for adjacent detection pairs. Conventional interpolation methods will perform poorly when the spacing between adjacent detection pairs gets larger. Thus, 3-D kernels are needed to extract and recover the information, and the size of the kernels is relevant to the spacing. In addition, the signal is only interpolated in spatial dimensions instead of all three dimensions to make the operator easier to learn." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.327, + 0.172, + 0.34 + ], + "angle": 0, + "content": "3.2. Dataset" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.35, + 0.47, + 0.59 + ], + "angle": 0, + "content": "For all experiments in this paper, we use subsets of Fashion-MNIST [40] and ShapeNet [5] to generate synthetic signals for training and testing. The Fashion-MNIST dataset consists of 2-D figures, while the ShapeNet comprises 3-D objects. We choose 5,000 samples from the training set of Fashion-MNIST and the class \"Car\" of ShapeNet respectively to generate our training set. In addition, we choose 1,000 samples from the testing set of Fashion-MNIST and the class \"Car\" of ShapeNet respectively to generate the testing set. Thus, the training set used in this work comprises 10,000 signal pairs, and the testing set comprises 2,000. All the signals are generated using Eq. (4). Besides, we set \\( N_{x} = N_{y} = 8 \\), \\( N_{t} = 100 \\) in this paper. For measured signals whose \\( N_{t} \\) is larger than 100, the network can also be applied directly since the network operates locally." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.591, + 0.47, + 0.712 + ], + "angle": 0, + "content": "For the confocal scenario, to match the experimental setting with the Stanford dataset [22], the 2-D figures from FashionMNIST [40] are placed \\(1\\mathrm{m}\\) away from the visible wall, the size of the figures is set as \\(1\\times 1\\mathrm{m}^2\\). And the 3-D objects from ShapeNet [5] are placed in a \\(1\\times 1\\times 0.2\\mathrm{m}^3\\) (horizontal, vertical, depth) box with a maximum depth of \\(1.1\\mathrm{m}\\). The illumination region on the visible wall is \\(2\\times 2\\mathrm{m}^2\\), and the time resolution of the SPAD is \\(32~\\mathrm{ps}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.712, + 0.469, + 0.833 + ], + "angle": 0, + "content": "For the non-confocal setting, we match the experimental setting with the data provided by phasor field [25]. The 2-D figures from Fashion-MNIST are placed \\(1\\mathrm{m}\\) away from the wall, and the size is \\(1.25\\times 1.25\\mathrm{m}^2\\). The 3-D objects from ShapeNet are placed in a \\(1.25\\times 1.25\\times 0.2\\mathrm{m}^3\\) (horizontal, vertical, depth) box with a maximum depth of \\(1.1\\mathrm{m}\\). The illumination region on the visible wall is \\(1.25\\times 1.25\\mathrm{m}^2\\), and the time resolution of the SPAD is 16 ps." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.846, + 0.163, + 0.86 + ], + "angle": 0, + "content": "4. Results" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "In this section, we show the reconstruction results of the proposed pipeline on both synthetic and measured data. For" + }, + { + "type": "image", + "bbox": [ + 0.499, + 0.087, + 0.892, + 0.778 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.788, + 0.892, + 0.871 + ], + "angle": 0, + "content": "Figure 5. Recovered signals and reconstruction results of the pyramid. (a) Ground truth of the hidden object. (b) A comparison of the recovered signals (the first arrival time of the original signal is marked by the blue curve in the zoom-in window of each subfigure). (c) Reconstruction results of state-of-the-art methods with different signals." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17424" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.288 + ], + "angle": 0, + "content": "each instance, reconstruction results obtained with signals of different spatial resolutions are shown to validate the effectiveness of the proposed pipeline. The first column (Original) shows reconstructions obtained from the original signal with a spatial resolution of \\(32 \\times 32\\); results in the second column (Low) are reconstructed using the signal scanned at \\(8 \\times 8\\) points; the third column (Nearest) shows reconstructions of the recovered signal obtained with nearest neighbor interpolation; the last column (Ours) illustrates reconstruction results obtained with the proposed pipeline. Comparisons with other interpolation methods and existing learning based NLOS imaging method are provided in the supplement." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.289, + 0.47, + 0.41 + ], + "angle": 0, + "content": "To illustrate the proposed pipeline can be acted as a plugin-and-play module for most algorithms, we combine state-of-the-art methods with our pipeline. For confocal experiments, the results are reconstructed by F-K [22], LCT [31], LOG-BP [18], D-LCT [45] and SOCR [27] algorithms, while for non-confocal experiments, results are given by LOG-BP, Phasor [24] and SOCR methods. For each instance, the parameters are fixed in each algorithm." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.411, + 0.47, + 0.516 + ], + "angle": 0, + "content": "To assess the results quantitatively, we compute the signal to noise ratio (SNR) of the recovered signals, as well as the peak signal to noise ratio (PSNR) and structural similarity index measure (SSIM) of the reconstructed maximum intensity projections. For measured data, we use the reconstructions obtained from the original signals as references to compute PSNR and SSIM." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.528, + 0.349, + 0.544 + ], + "angle": 0, + "content": "4.1. Results of confocal experiment" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.552, + 0.47, + 0.644 + ], + "angle": 0, + "content": "In Fig. 4, we compare the mean SNR values of the signals interpolated with SSN and the nearest neighbor method. It is shown that the network provides much better results on both datasets. In other words, the network approximates the oracle interpolation operator \\(\\Phi\\) with high accuracy and serves as a crucial part of the proposed pipeline." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.47, + 0.764 + ], + "angle": 0, + "content": "To test the robustness of the proposed pipeline, we use the instance of the pyramid from SOCR [27], whose signal was generated with a different physical model. As shown in Fig. 5b, the first arrival time is correctly recovered by the proposed network, while the nearest neighbor interpolation method fails. Fig. 5c shows that the proposed pipeline provides a signal for high quality reconstructions with better visual quality and higher PSNR and SSIM values." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.901 + ], + "angle": 0, + "content": "The proposed pipeline is then tested on measured data from Stanford dataset [22]. This dataset contains signals measured with different hidden objects and exposure time. We choose the instance of the statue to validate the effectiveness of the proposed pipeline. The measurement time of each sampling point is \\(0.0069\\mathrm{s}\\). The total exposure time of the signal scanned at \\(32\\times 32\\) points is \\(7.03\\mathrm{s}\\), and it only takes \\(0.44\\mathrm{s}\\) to scan \\(8\\times 8\\) points. As is shown in Fig. 6b, the result of SSN has higher SNR. The first arrival time can" + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.087, + 0.892, + 0.778 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.788, + 0.895, + 0.871 + ], + "angle": 0, + "content": "Figure 6. Recovered signals and reconstruction results of the statue. (a) Ground truth of the hidden object. (b) A comparison of recovered signals (the first arrival time of the original signal is marked by the blue curve in the zoom-in window of each subfigure). (c) Reconstruction results of state-of-the-art methods with different signals." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17425" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.198 + ], + "angle": 0, + "content": "hardly be distinguished in the nearest neighbor interpolated signal, while it is much clearer in the signal recovered by the proposed network. As shown in Fig. 6c, the proposed pipeline results in fewer artifacts in the background. More comparisons of different exposure times are provided in the supplement, which also indicates the outperformance of the proposed pipeline." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.208, + 0.334, + 0.224 + ], + "angle": 0, + "content": "4.2. Results of the dynamic scene" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.232, + 0.47, + 0.382 + ], + "angle": 0, + "content": "To illustrate the potential application of the proposed pipeline on real-time NLOS imaging, we test it on the captured dynamic scenes from the Stanford dataset. The data is captured over a \\(2 \\times 2\\mathrm{m}^2\\) region on the visible wall, and the person is wearing a retroreflective suit to increase the detection of the indirect signal. It takes \\(0.25\\mathrm{s}\\) to scan \\(32 \\times 32\\) points, which allows for a reconstruction video with 4 frames per second. The reflection model of the retroreflective material [31] is different from the one defined in Eq. (6)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.383, + 0.471, + 0.577 + ], + "angle": 0, + "content": "The results of several frames are shown in Fig. 7b, the complete video of the reconstruction results can be found in the supplement. All the results are reconstructed by the F-K method. Although the training set is generated by a different forward model, the proposed pipeline can still recover signals for high quality reconstructions. The pose of the person can be clearly identified in the reconstructions using the proposed pipeline, while the reconstructions using the nearest neighbor interpolated signal contain more artifacts. The mean PSNR and SSIM of the reconstruction results obtained from the proposed pipeline are 24.46 and 0.63, while those obtained with nearest neighbor interpolation are only 18.68 and 0.37." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.58, + 0.47, + 0.7 + ], + "angle": 0, + "content": "The total capture time of each frame is reduced from 0.25 s to 0.0156 s. This enables us to obtain a high quality reconstruction video with 64 frames per second. Besides, it only takes 0.08 s to recover the high resolution signal with the proposed network, and 0.01s to reconstruct the hidden object with the F-K method on an NVIDIA RTX 3090. Thus, real-time video imaging with a high frame rate can be realized with a delay of a few frames." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.711, + 0.383, + 0.727 + ], + "angle": 0, + "content": "4.3. Results of non-confocal experiment" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.469, + 0.839 + ], + "angle": 0, + "content": "For the non-confocal setting, we use the instance of \"4\" provided by Phasor [24] to validate the effectiveness of the proposed method. The measured signal is subsampled to \\(32 \\times 32\\) and the illumination region on the visible wall is cropped to \\(1.25 \\times 1.25\\mathrm{m}^2\\). We use the signal scanned at 8 \\(\\times 8\\) points as the input of our pipeline. It takes 1 s to scan each point on the visible wall." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.902 + ], + "angle": 0, + "content": "As is shown in Fig. 8b, the SNR of the signal recovered by the network is much higher than the one interpolated with the nearest neighbor method. Reconstruction results are shown in Fig. 8c. The proposed pipeline provides re" + }, + { + "type": "image", + "bbox": [ + 0.618, + 0.089, + 0.775, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.678, + 0.175, + 0.716, + 0.187 + ], + "angle": 0, + "content": "(a) GT" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.192, + 0.892, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.629, + 0.527, + 0.766, + 0.539 + ], + "angle": 0, + "content": "(b) Reconstruction results" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.551, + 0.894, + 0.65 + ], + "angle": 0, + "content": "Figure 7. Reconstruction results of the dynamic scene. (a) An illustration of the dynamic scene. (b) Reconstruction results of the dynamic scene. The results are obtained with the F-K method. The PSNR and SSIM values are computed with the reconstructions obtained with the original signals as references. The proposed pipeline provides much better reconstructions in terms of both quantitative criteria and visual quality." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.676, + 0.874, + 0.692 + ], + "angle": 0, + "content": "constructions with higher quantitative and visual quality." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.706, + 0.614, + 0.721 + ], + "angle": 0, + "content": "5. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.731, + 0.892, + 0.807 + ], + "angle": 0, + "content": "We have introduced a novel pipeline for high quality NLOS reconstruction with the signal measured at only a small number of scanning points. In this section, we discuss several factors that affect the reconstruction quality and some feasible solutions to improve the pipeline further." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.817, + 0.72, + 0.833 + ], + "angle": 0, + "content": "5.1. Training of the network" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.893, + 0.901 + ], + "angle": 0, + "content": "The training of the network employed in the proposed pipeline is supervised. Thus, it is affected by the generation of the training set, including the forward model, experimental setting, and hidden objects. For example, hidden objects" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17426" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.137 + ], + "angle": 0, + "content": "with different materials will lead to various physical models. The networks trained and tested on signals with different models would bring additional errors to the results." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.137, + 0.47, + 0.226 + ], + "angle": 0, + "content": "However, a network with good generalization ability can avoid degradation. As shown in Fig. 7, although the physical model of the dynamic scenes differs from the one used in the training set, the reconstructed objects are only slightly affected. Thus, the proposed pipeline has strong robustness to the forward model." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.235, + 0.391, + 0.251 + ], + "angle": 0, + "content": "5.2. Performance of different algorithms" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.259, + 0.469, + 0.41 + ], + "angle": 0, + "content": "As the reconstruction results shown in previous sections, the improvement of reconstruction quality of the proposed pipeline varies between different methods. Compared to the reconstruction results obtained from the low resolution signal, fast imaging algorithms like F-K, LCT and D-LCT have larger improvements, while the iterative inversion algorithm SOCR might have similar visual quality in some cases. This phenomenon may come from the fact that the iterative algorithm can make full use of the information contained in the signal, but with a longer reconstruction time." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.418, + 0.469, + 0.448 + ], + "angle": 0, + "content": "5.3. Combination with SPAD array for further acceleration" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.456, + 0.469, + 0.621 + ], + "angle": 0, + "content": "Since the proposed pipeline can recover the signal faithfully under the non-confocal setting, combining it with a SPAD array can accelerate the measurement process further. However, the existing SPAD array can only focus on a rectangular area on the visible wall, which means that the network has to learn an extrapolation operator. This is much harder than learning an interpolation operator. Thus, it would be more practical if the SPAD array could focus on sparse rectangular points on a relatively large area of the visible wall. This may be seen as a new area for hardware design." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.634, + 0.195, + 0.649 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.469, + 0.901 + ], + "angle": 0, + "content": "In this paper, we have studied the signal recovery problem in NLOS imaging, and proposed a two-step learning based pipeline for fast NLOS imaging with only a small number of scanning points. In the first step, the signal of high spatial resolution is recovered with a neural network. In the second step, the hidden objects are obtained from the virtually measured signals at a dense grid. Our pipeline provides high-quality reconstructions under both confocal and non-confocal settings, which takes only \\(6.25\\%\\) of the original measurement time. Besides, the proposed pipeline is compatible with most existing detection systems and imaging algorithms. With the proposed pipeline, the high frame rate imaging video is possible, which is a noteworthy advance in this area. It is expected to combine emerging techniques in the fields of SPAD array for real-time and high quality NLOS reconstructions with higher frame rates." + }, + { + "type": "image", + "bbox": [ + 0.661, + 0.092, + 0.737, + 0.17 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.678, + 0.174, + 0.716, + 0.185 + ], + "angle": 0, + "content": "(a) GT" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.187, + 0.892, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.639, + 0.308, + 0.754, + 0.319 + ], + "angle": 0, + "content": "(b) Recovered signals" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.321, + 0.888, + 0.602 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.629, + 0.611, + 0.765, + 0.622 + ], + "angle": 0, + "content": "(c) Reconstruction results" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.635, + 0.892, + 0.718 + ], + "angle": 0, + "content": "Figure 8. Recovered signals and reconstruction results of the instance \"4\". (a) Ground truth of the hidden object. (b) A comparison of recovered signals (the first arrival time of the original signal is marked by the blue curve in the zoom-in window of each subfigure). (c) Reconstruction results of state-of-the-art methods with different signals." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.75, + 0.892, + 0.81 + ], + "angle": 0, + "content": "Acknowledgments. The authors would like to thank Zhu-peng Li and Rongqian Wang for fruitful discussions. This work was supported by the National Natural Science Foundation of China (61975087, 12071244, 11971258)." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.832, + 0.596, + 0.847 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.508, + 0.859, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[1] Byeongjoo Ahn, Akshat Dave, Ashok Veeraraghavan, Ioannis Gkioulekas, and Aswin C Sankaranarayanan. Convolutional approximations to the general non-line-of-sight imag" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "17427" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.093, + 0.47, + 0.12 + ], + "angle": 0, + "content": "ing operator. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7889-7899, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.122, + 0.472, + 0.163 + ], + "angle": 0, + "content": "[2] Saeed Anwar, Salman Khan, and Nick Barnes. A deep journey into super-resolution: A survey. ACM Computing Surveys (CSUR), 53(3):1-34, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.165, + 0.471, + 0.22 + ], + "angle": 0, + "content": "[3] Victor Arellano, Diego Gutierrez, and Adrian Jarabo. Fast back-projection for non-line of sight reconstruction. In ACM SIGGRAPH 2017 Posters, SIGGRAPH '17, New York, NY, USA, 2017. Association for Computing Machinery. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.222, + 0.471, + 0.276 + ], + "angle": 0, + "content": "[4] Syed Muhammad Arsalan Bashir, Yi Wang, Mahrukh Khan, and Yilong Niu. A comprehensive review of deep learning-based single image super-resolution. PeerJ Computer Science, 7:e621, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.278, + 0.471, + 0.375 + ], + "angle": 0, + "content": "[5] Angel X. Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. ShapeNet: An Information-Rich 3D Model Repository. Technical Report arXiv:1512.03012 [cs.GR], Stanford University — Princeton University — Toyota Technological Institute at Chicago, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.377, + 0.471, + 0.417 + ], + "angle": 0, + "content": "[6] Haoyu Chen, Jinjin Gu, and Zhi Zhang. Attention in attention network for image super-resolution. arXiv preprint arXiv:2104.09497, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.419, + 0.471, + 0.473 + ], + "angle": 0, + "content": "[7] Wenzheng Chen, Simon Daneau, Fahim Mannan, and Felix Heide. Steady-state non-line-of-sight imaging. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.476, + 0.471, + 0.532 + ], + "angle": 0, + "content": "[8] Wenzheng Chen, Fangyin Wei, Kiriakos N Kutulakos, Szymon Rusinkiewicz, and Felix Heide. Learned feature embeddings for non-line-of-sight imaging and recognition. ACM Transactions on Graphics (TOG), 39(6):1-18, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.533, + 0.471, + 0.573 + ], + "angle": 0, + "content": "[9] Justin Dove and Jeffrey H. Shapiro. Paraxial theory of phasor-field imaging. Opt. Express, 27(13):18016-18037, Jun 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.575, + 0.47, + 0.616 + ], + "angle": 0, + "content": "[10] Xiaohua Feng and Liang Gao. Improving non-line-of-sight image reconstruction with weighting factors. Optics letters, 45(14):3921-3924, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.618, + 0.47, + 0.658 + ], + "angle": 0, + "content": "[11] Xiaohua Feng and Liang Gao. Ultrafast light field tomography for snapshot transient and non-line-of-sight imaging. Nature Communications, 12:2179, 04 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.661, + 0.47, + 0.716 + ], + "angle": 0, + "content": "[12] Javier Grau Chopite, Matthias B. Hullin, Michael Wand, and Julian Iseringhausen. Deep non-line-of-sight reconstruction. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 957-966, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.718, + 0.47, + 0.773 + ], + "angle": 0, + "content": "[13] Felix Heide, Matthew O'Toole, Kai Zang, David B. Lindell, Steven Diamond, and Gordon Wetzstein. Non-line-of-sight imaging with partial occluders and surface normals. ACM Trans. Graph., 38(3), May 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.47, + 0.815 + ], + "angle": 0, + "content": "[14] Julian Iseringhausen and Matthias B Hullin. Non-line-of-sight reconstruction using efficient transient rendering. ACM Transactions on Graphics (TOG), 39(1):1-14, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.47, + 0.872 + ], + "angle": 0, + "content": "[15] Mariko Isogawa, Dorian Chan, Ye Yuan, Kris M. Kitani, and Matthew O'Toole. Efficient non-line-of-sight imaging from transient sinograms. In 16th European Conference on Computer Vision (ECCV), 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.874, + 0.47, + 0.903 + ], + "angle": 0, + "content": "[16] Mariko Isogawa, Ye Yuan, Matthew O'Toole, and Kris Kitani. Optical non-line-of-sight physics-based 3d human pose" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.472, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.133 + ], + "angle": 0, + "content": "estimation. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7011-7020, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.136, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[17] Ahmed Kirmani, Tyler Hutchison, James Davis, and Ramesh Raskar. Looking around the corner using transient imaging. In 2009 IEEE 12th International Conference on Computer Vision, pages 159–166, 2009. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.193, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[18] Martin Laurenzis and Andreas Velten. Feature selection and back-projection algorithms for nonlinear-of-sight laser-gated viewing. Journal of Electronic Imaging, 23:063003, 11 2014. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.29 + ], + "angle": 0, + "content": "[19] Y. Li, B. Sixou, and F. Peyrin. A review of the deep learning methods for medical images super resolution problems. IRBM, 42(2):120-133, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.292, + 0.892, + 0.347 + ], + "angle": 0, + "content": "[20] Zhupeng Li, Xintong Liu, Jianyu Wang, Zuoqiang Shi, Lingyun Qiu, and Xing Fu. Fast non-line-of-sight imaging based on first photon event stamping. Optics letters, 47(8):1928—1931, April 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.349, + 0.892, + 0.404 + ], + "angle": 0, + "content": "[21] Zhengpeng Liao, Deyang Jiang, Xiaochun Liu, Andreas Velten, Yajun Ha, and Xin Lou. FPGA accelerator for real-time non-line-of-sight imaging. IEEE Transactions on Circuits and Systems I: Regular Papers, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.406, + 0.892, + 0.46 + ], + "angle": 0, + "content": "[22] David B Lindell, Gordon Wetzstein, and Matthew O'Toole. Wave-based non-line-of-sight imaging using fast fk migration. ACM Transactions on Graphics (ToG), 38(4):1-13, 2019. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.462, + 0.892, + 0.516 + ], + "angle": 0, + "content": "[23] Jianjiang Liu, Yijun Zhou, Xin Huang, Zheng-Ping Li, and Feihu Xu. Photon-efficient non-line-of-sight imaging. IEEE Transactions on Computational Imaging, 8:639-650, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.519, + 0.892, + 0.573 + ], + "angle": 0, + "content": "[24] Xiaochun Liu, Sebastian Bauer, and Andreas Velten. Phasor field diffraction based reconstruction for fast non-line-of-sight imaging systems. Nature Communications, 11:1645, 2020. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.576, + 0.892, + 0.631 + ], + "angle": 0, + "content": "[25] Xiaochun Liu, Ibón Guillén, Marco Manna, Ji Nam, Azer Reza, Toan Le, Adrián Jarabo, Diego Gutiérrez, and Andreas Velten. Non-line-of-sight imaging using phasor-field virtual wave optics. Nature, 572, 2019. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.633, + 0.892, + 0.688 + ], + "angle": 0, + "content": "[26] Xiaochun Liu and Andreas Velten. The role of wigner distribution function in non-line-of-sight imaging. In 2020 IEEE International Conference on Computational Photography (ICCP), pages 1-12. IEEE, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.689, + 0.892, + 0.745 + ], + "angle": 0, + "content": "[27] Xintong Liu, Jianyu Wang, Zhupeng Li, Zuoqiang Shi, Xing Fu, and Lingyun Qiu. Non-line-of-sight reconstruction with signal-object collaborative regularization. Light: Science & Applications, 08 2021. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.747, + 0.892, + 0.801 + ], + "angle": 0, + "content": "[28] Christopher Metzler, David Lindell, and Gordon Wetzstein. Keyhole imaging: Non-line-of-sight imaging and tracking of moving objects along a single optical path. IEEE Transactions on Computational Imaging, PP:1-1, 12 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.803, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[29] Christopher A. Metzler, Felix Heide, Prasana Rangarajan, Muralidhar Madabhushi Balaji, Aparna Viswanath, Ashok Veeraraghavan, and Richard G. Baraniuk. Deep-inverse correlography: towards real-time high-resolution non-line-of-sight imaging. Optica, 7(1):63-71, Jan 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.874, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[30] Ji Hyun Nam, Eric Brandt, Sebastian Bauer, Xiaochun Liu, Marco Renna, Alberto Tosi, Eftychios Sifakis, and Andreas" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17428" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.133 + ], + "angle": 0, + "content": "Velten. Low-latency time-of-flight non-line-of-sight imaging at 5 frames per second. Nature communications, 12(1):1-10, 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.177 + ], + "angle": 0, + "content": "[31] Matthew O'Toole, David Lindell, and Gordon Wetzstein. Confocal non-line-of-sight imaging based on the light-cone transform. Nature, 555, 2018. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.18, + 0.469, + 0.248 + ], + "angle": 0, + "content": "[32] Chengquan Pei, Anke Zhang, Yue Deng, Feihu Xu, Jiamin Wu, David U-Lei Li, Hui Qiao, Lu Fang, and Qionghai Dai. Dynamic non-line-of-sight imaging system based on the optimization of point spread functions. Opt. Express, 29(20):32349-32364, Sep 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.251, + 0.469, + 0.305 + ], + "angle": 0, + "content": "[33] Syed Azer Reza, Marco La Manna, Sebastian Bauer, and Andreas Velten. Phasor field waves: A huygens-like light transport model for non-line-of-sight imaging applications. Opt. Express, 27(20):29380-29400, Sep 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.308, + 0.469, + 0.376 + ], + "angle": 0, + "content": "[34] Christos Thrampoulidis, Gal Shulkind, Feihu Xu, William T. Freeman, Jeffrey H. Shapiro, Antonio Torralba, Franco N. C. Wong, and Gregory W. Wornell. Exploiting occlusion in non-line-of-sight active imaging. IEEE Transactions on Computational Imaging, 4(3):419-431, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.379, + 0.469, + 0.447 + ], + "angle": 0, + "content": "[35] Chia-Yin Tsai, Aswin C Sankaranarayanan, and Ioannis Gkioulekas. Beyond volumetric albedo-a surface optimization framework for non-line-of-sight imaging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1545-1555, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.45, + 0.469, + 0.531 + ], + "angle": 0, + "content": "[36] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Proceedings of the 31st International Conference on Neural Information Processing Systems, NIPS'17, page 6000-6010, Red Hook, NY, USA, 2017. Curran Associates Inc. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.534, + 0.469, + 0.617 + ], + "angle": 0, + "content": "[37] Eric Veach and Leonidas J. Guibas. Optimally combining sampling techniques for monte carlo rendering. In Proceedings of the 22nd Annual Conference on Computer Graphics and Interactive Techniques, SIGGRAPH '95, page 419-428, New York, NY, USA, 1995. Association for Computing Machinery. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.619, + 0.469, + 0.687 + ], + "angle": 0, + "content": "[38] Andreas Velten, Thomas Willwacher, Otkrist Gupta, Ashok Veeraraghavan, Moungi G Bawendi, and Ramesh Raskar. Recovering three-dimensional shape around a corner using ultrafast time-of-flight imaging. Nature communications, 3(1):1-8, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.69, + 0.469, + 0.759 + ], + "angle": 0, + "content": "[39] Cheng Wu, Jianjiang Liu, Xin Huang, Zheng-Ping Li, Chao Yu, Jun-Tian Ye, Jun Zhang, Qiang Zhang, Xiankang Dou, Vivek Goyal, Feihu Xu, and Jian-Wei Pan. Non-line-of-sight imaging over \\(1.43\\mathrm{km}\\). Proceedings of the National Academy of Sciences of the United States of America, 118, 03 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.469, + 0.814 + ], + "angle": 0, + "content": "[40] Han Xiao, Kashif Rasul, and Roland Vollgraf. Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms. arXiv preprint arXiv:1708.07747, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[41] Shumian Xin, Sotiris Nousias, Kiriakos N. Kutulakos, Aswin C. Sankaranarayanan, Srinivasa G. Narasimhan, and Ioannis Gkioulekas. A theory of fermat paths for non-line-of-sight shape reconstruction. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6793-6802, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[42] Wenqing Yang, Chao Zhang, Wenjie Jiang, Zexin Zhang, and Baoqing Sun. None-line-of-sight imaging enhanced with spatial multiplexing. Opt. Express, 30(4):5855-5867, Feb 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[43] Wenming Yang, Xuechen Zhang, Yapeng Tian, Wei Wang, Jing-Hao Xue, and Qingmin Liao. Deep learning for single image super-resolution: A brief review. IEEE Transactions on Multimedia, 21(12):3106-3121, Dec 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.892, + 0.248 + ], + "angle": 0, + "content": "[44] Jun-Tian Ye, Xin Huang, Zheng-Ping Li, and Feihu Xu. Compressed sensing for active non-line-of-sight imaging. Opt. Express, 29(2):1749-1763, Jan 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.25, + 0.892, + 0.329 + ], + "angle": 0, + "content": "[45] Sean I Young, David B Lindell, Bernd Girod, David Taubman, and Gordon Wetzstein. Non-line-of-sight surface reconstruction using the directional light-cone transform. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1407-1416, 2020. 2, 3, 6" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "17429" + } + ] +] \ No newline at end of file diff --git a/2023/Non-Line-of-Sight Imaging With Signal Superresolution Network/224cfcfa-2657-4e74-b370-35626fe2670d_origin.pdf b/2023/Non-Line-of-Sight Imaging With Signal Superresolution Network/224cfcfa-2657-4e74-b370-35626fe2670d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2837facaa482e5fd0f11d4d5781ff8a147ebf9bb --- /dev/null +++ b/2023/Non-Line-of-Sight Imaging With Signal Superresolution Network/224cfcfa-2657-4e74-b370-35626fe2670d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:10d520fd1efab953b6f84cd8c545a824390d26fcfc48306a772b005646b6973c +size 7073158 diff --git a/2023/Non-Line-of-Sight Imaging With Signal Superresolution Network/full.md b/2023/Non-Line-of-Sight Imaging With Signal Superresolution Network/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5a98c7f29d8828ccbfdfe3a4c1fce4e804230cc9 --- /dev/null +++ b/2023/Non-Line-of-Sight Imaging With Signal Superresolution Network/full.md @@ -0,0 +1,297 @@ +# Non-line-of-sight Imaging with Signal Superresolution Network + +Jianyu Wang $^{1}$ , Xintong Liu $^{1}$ , Leping Xiao $^{1}$ , Zuoqiang Shi $^{1,2}$ , Lingyun Qiu $^{1,2}$ , and Xing Fu $^{1}$ $^{1}$ Tsinghua University + +$^{2}$ Yanqi Lake Beijing Institute of Mathematical Sciences and Applications + +# Abstract + +Non-line-of-sight (NLOS) imaging aims at reconstructing the location, shape, albedo, and surface normal of the hidden object around the corner with measured transient data. Due to its strong potential in various fields, it has drawn much attention in recent years. However, long exposure time is not always available for applications such as auto-driving, which hinders the practical use of NLOS imaging. Although scanning fewer points can reduce the total measurement time, it also brings the problem of imaging quality degradation. This paper proposes a general learning-based pipeline for increasing imaging quality with only a few scanning points. We tailor a neural network to learn the operator that recovers a high spatial resolution signal. Experiments on synthetic and measured data indicate that the proposed method provides faithful reconstructions of the hidden scene under both confocal and nonconfocal settings. Compared with original measurements, the acquisition of our approach is 16 times faster while maintaining similar reconstruction quality. Besides, the proposed pipeline can be applied directly to existing optical systems and imaging algorithms as a plug-in-and-play module. We believe the proposed pipeline is powerful in increasing the frame rate in NLOS video imaging. + +# 1. Introduction + +Non-line-of-sight (NLOS) imaging problem usually employs a high temporal resolution optical system to recover the hidden object around the corner. As shown in Fig. 1a, photons emitted by the laser are collected by the detector after three diffuse reflections: the reflection at the visible wall, the reflection at the hidden object, and the reflection at the visible wall again. Scanning a region on the visible wall can get measured histogram data to recover the hidden scene. + +Due to its potential applications in various fields, such as auto-driving, disaster relief, and remote sensing, the NLOS imaging problem has become an emerging field since it was first proposed by Kirmani et al. [17] in 2009. + +![](images/429a37763f5f9637697c2908d9b5c6d7d715ed12096e8194d4c7cf78eb55b456.jpg) + +![](images/6394929e8689e54aa2c5e7e2876a09f74b7be49627ef25f50cfe9fe009d4f5b7.jpg) +(b) Scanning points + +![](images/b26ef31d3c80ab0bc1c79602292b638c5999426563c750bf3cf914cfcd211f77.jpg) +(a) NLOS scenario +Low + +![](images/c44ef8d632ba6e3af2af42f6d0c0ef089fbac977395c26c62a0681b607f5b73c.jpg) +(c) Results +Figure 1. A typical NLOS scenario and reconstruction results. (a) An illustration of a typical NLOS scenario. (b) An illustration of the scanning points on the visible wall. Using the proposed pipeline, only the blue circle points are needed to be illuminated, and the signal at the yellow square points can be recovered with the proposed pipeline. (c) Comparisons of the reconstruction results of the pyramid with different signals. The maximum intensity projection of the albedo values along the depth direction is shown in the upper left corner (GT). The results are reconstructed by the original signal with spatial resolution $32 \times 32$ (Original), the subsampled signal with spatial resolution $8 \times 8$ (Low), and the signal recovered by the proposed network with spatial resolution $32 \times 32$ (Ours). + +Many methods have been proposed to improve the prac + +![](images/cc7ccacf4ab46248265f69b400c2fa39b4250dd6c0c0826939aa02df70a2837d.jpg) +Figure 2. Flowchart of the proposed pipeline. The high resolution signal is recovered from the low resolution signal using a neural network. The hidden scene is then reconstructed using state-of-the-art imaging algorithms designed for high resolution signal. + +ticability [1, 3, 13, 14, 34, 35, 39] or reconstruction quality [10, 11, 26, 41]. A back-projection method with Laplacian of Gaussian filter (LOG-BP) [18] is introduced to improve the reconstruction quality of the back-projection method [38]. In 2018, O'Toole et al. [31] first apply the light-contransform (LCT) method to image the hidden object with a confocal scanning mode. Young et al. [45] then propose the directional light-cone-transform (D-LCT) method, which simultaneously recovers the albedo and surface normal of the hidden object. From the perspective of wave characteristics, Lindell et al. [22] introduce the fast frequency-wavenumber migration (F-K) method to NLOS. Another remarkable imaging method proposed by Liu et al. [25] employs the phasor field model [9, 33] and provides an efficient solution for fast NLOS imaging under non-confocal scenario [24]. Further improvements of this method using single photon avalanche diode (SPAD) arrays achieve real-time imaging [21, 30]. Deep learning based methods have also been introduced [7, 8, 12, 16, 29]. Recently, Liu et al. [27] propose the signal-object collaborative regularization (SOCR) method, which can improve the resolution of the reconstructed object, especially in cases with high measurement noise. + +However, one main obstacle in NLOS imaging is that we need a long exposure time to improve the signal-to-noise ratio (SNR) of the measurement, which is not admissible in many scenarios, such as auto-driving. Several methods have been proposed to reduce the capture time [20, 23]. For a confocal scenario, Isogawa et al. [15] propose a circular scanning mode. In 2021, Metzler et al. [28] consider an extreme case that tracks the objects with a single detection point. By introducing the downsampling operation, Ye et al. [44] reconstruct the object with a subset of the virtual sources. For non-confocal scenarios, most methods employ SPAD arrays to realize fast measurement. Among them, Nam et al. [30] achieve a reconstruction of five frames per second by remapping the data from two $16 \times 1$ SPAD arrays. Pei et al. [32] realize 20 frames per second reconstruction using a $32 \times 32$ SPAD array. In 2022, Yang et al. [42] propose a spatial multiplexing method based on the + +compressed sensing strategy. Although these methods can reduce the measurement time, they might either cause deterioration in imaging quality or substantially increase expense. + +In this paper, we propose a deep learning based pipeline that can reconstruct the hidden object with fine details using only a small number of scanning points. Different measurement densities are shown in Fig. 1b. In the following sections, we term the signal measured at the blue circle points as the "low resolution signal" while those measured at both circle and square points as the "high resolution signal". + +Instead of designing new algorithms for sparse measurements, we focus on recovering the dense virtual signals, which can be applied to most algorithms as a plug-in-and-play module. The proposed pipeline consists of two steps: the low resolution signal is first mapped into the high resolution signal space by a neural network; the reconstruction is then obtained from the recovered high resolution signal using existing imaging algorithms. The reconstruction results of the pyramid are shown in Fig. 1c. These results are reconstructed by the F-K method. As the comparison shows, the proposed pipeline can provide a high quality reconstruction with only 64 scanning points, which is $6.25\%$ of the original number. + +To our best knowledge, this is the first work in NLOS to study the signal recovery problem of the low resolution signal sparsely measured on a uniform grid. The proposed method can reduce the exposure time without losing the reconstruction quality. The main contributions of this work are: + +- We propose a novel learning based pipeline for faithful reconstruction of the hidden scene, which only takes $6.25\%$ of the original measuring time. +- We show that the high resolution signal can be recovered accurately with the proposed signal superresolution network, which we abbreviate as "SSN". +- The effectiveness of the proposed method is validated on public datasets under both confocal and non + +![](images/fd498f1a352f288178a504422cc88beab378c33bdd54c93a0803636122ed287c.jpg) +Figure 3. The architecture of the signal superresolution network (SSN) in the proposed pipeline. The network is composed of two main branches: a conventional upsampling branch and a learning branch. The final output is the sum of the results from these two branches. + +confocal settings. + +- We demonstrate the potential application of the proposed method in high frame rate (64 frames per second) NLOS video imaging using public datasets. + +# 2. Reconstruction Pipeline + +This work aims to develop a learning based pipeline for NLOS imaging, which can reduce the measurement time while maintaining the reconstruction quality. In this section, we formulate the signal recovery problem and introduce the proposed pipeline. + +# 2.1. Formulation of the problem + +Most existing methods that reconstruct the hidden object from low resolution signals $d_{l}$ can be formulated as + +$$ +I (x, y, z) = F _ {l} \left(d _ {l}\right) \tag {1} +$$ + +in which $I(x,y,z)$ is the reconstructed object and $F_{l}$ is the imaging algorithm designed for low resolution signal [15, 28, 44]. Since only a few sampling points are scanned, the total exposure time can be reduced. However, the reconstruction quality also decreases with the number of scan points. + +On the contrary, the algorithms designed for high resolution signal can be expressed as + +$$ +I (x, y, z) = F _ {h} \left(d _ {h}\right) \tag {2} +$$ + +in which $F_{h}$ represents the algorithm designed for high resolution signal $d_{h}$ . Many works have been proposed for fast [22, 24, 31, 45] and high resolution [18, 27] reconstruction as discussed above. + +In this work, we aim at designing a pipeline that can provide faithful reconstruction with only a small number of + +scanning points, and has good compatibility with traditional optical systems as well as existing imaging algorithms. To solve this problem, we introduce a signal recovering operator $\Phi$ , which maps the low resolution signal $d_{l}$ into the space of high resolution signal $d_{h}$ . The proposed pipeline can be expressed as + +$$ +I (x, y, z) = F _ {h} \left(\Phi \left(d _ {l}\right)\right) \tag {3} +$$ + +in which $\Phi(d_l)$ is of the same size with $d_h$ . Thus, the problem of NLOS imaging from low resolution signal converts to the problem of finding the operator $\Phi$ which satisfies $\Phi(d_l) = d_h$ . + +# 2.2. The choice of recovering operator + +The choice of the operator $\Phi$ is highly dependent on the physical model of the measurement process. We adopt a simplified version of the widely used three-point transport model [37] as + +$$ +\tau (\mathbf {x} _ {l}, \mathbf {x} _ {d}, t) = \int_ {S} T (\mathbf {x} _ {l}, \mathbf {x} _ {d}, t, \mathbf {s}) C (\mathbf {x} _ {l}, \mathbf {x} _ {d}, \mathbf {s}) f (\mathbf {s}) d A (\mathbf {s}) \tag {4} +$$ + +in which $\tau$ is the photon intensity detected with the virtual source $\mathbf{x}_l$ , virtual detector $\mathbf{x}_d$ and time $t$ . $f(\mathbf{s})$ represents the albedo of the point $\mathbf{s}$ on the hidden object $S$ and $A(\mathbf{s})$ is the corresponding area measure. $T$ is a function that characterizes the arrival time of the photons, which is defined as + +$$ +T (\mathbf {x} _ {l}, \mathbf {x} _ {d}, t, \mathbf {s}) = \delta (| | \mathbf {x} _ {l} - \mathbf {s} | | _ {2} + | | \mathbf {x} _ {d} - \mathbf {s} | | _ {2} - c t) \tag {5} +$$ + +in which $\delta$ is the Dirac function and $c$ is the light speed. Besides, the function $C$ characterizes the intensity attenuation + +in the physical model, which is expressed as + +$$ +\begin{array}{l} C \left(\mathbf {x} _ {l}, \mathbf {x} _ {d}, \mathbf {s}\right) = \frac {< \mathbf {s} - \mathbf {x} _ {l} , \mathbf {n} (\mathbf {x} _ {l}) > < \mathbf {x} _ {l} - \mathbf {s} , \mathbf {n} (\mathbf {s}) >}{\left| \left| \mathbf {x} _ {l} - \mathbf {s} \right| \right| _ {2} ^ {4}} (6) \\ \times \frac {< \mathbf {s} - \mathbf {x} _ {d} , \mathbf {n} (\mathbf {x} _ {d}) > < \mathbf {x} _ {d} - \mathbf {s} , \mathbf {n} (\mathbf {s}) >}{\left| \left| \mathbf {x} _ {d} - \mathbf {s} \right| \right| _ {2} ^ {\frac {4}}}. (6) \\ \end{array} +$$ + +in which $\mathbf{n}(\mathbf{s})$ is the unit normal vector pointing toward the visible wall at point $\mathbf{s}$ and $<\cdot, \cdot>$ is the inner product in $\mathbb{R}^3$ . + +Suppose that there are two pairs $(\mathbf{x}_{l_1},\mathbf{x}_{d_1}),(\mathbf{x}_{l_2},\mathbf{x}_{d_2})$ on the visible wall. For a fixed point s on the hidden object, we assume that the contributions of this point to the signal at these two pairs are contained in time bin $t_1$ and $t_2$ respectively. Then from Eq. (5), we know that $t_1$ and $t_2$ are close if + +$$ +D = \left\| \mathbf {x} _ {l _ {1}} - \mathbf {x} _ {l _ {2}} \right\| _ {2} + \left\| \mathbf {x} _ {d _ {1}} - \mathbf {x} _ {d _ {2}} \right\| _ {2} \tag {7} +$$ + +is small. We can then regard the signal recovery problem as an interpolation problem, since the points have contributions to neighboring time bins. + +However, when the spacing between adjacent detection points is large, conventional interpolation methods will fail due to the violation of the assumption. This phenomenon is very common in practical applications. To solve this problem, Nam et al. [30] employs two SPAD arrays to meet the assumption, but brings additional expense to the system. Another solution is brought up by Yang et al. [42], which combines a micro-mirror device with the compressed sensing strategy to increase detection efficiency. However, this method is sensitive to the geometry of the visible wall, which limits its application. + +Instead of introducing new detection systems, we take advantage of the fast development of deep learning methods and adopt a neural network to learn the nonlinear operator $\Phi$ . Details about the network will be introduced in the next section. With the employed network, we can obtain a better approximation of the operator $\Phi$ , as well as a higher reconstruction quality. + +The proposed pipeline is shown in Fig. 2. The low resolution signal $d_{l}$ is first processed by the neural network $\Phi$ to recover the high resolution signal $\Phi(d_{l})$ . Then the reconstruction $I$ is obtained using the existing imaging algorithm $F_{h}$ . With the proposed pipeline, we can decrease the exposure time significantly while maintaining the reconstruction quality. + +# 3. Signal Superresolution network + +Based on the analysis in the last section, we have formulated the signal recovery problem as an interpolation problem and have noticed that conventional interpolation methods may fail when the spacing between adjacent detection pairs gets larger. Besides, the contribution of one point on + +![](images/515b81d3bb34b1b265dd6414e6a21e2e620637f7f4106cd0fafbab1039c5d4a3.jpg) +Figure 4. Comparisons of the mean SNR values of the results obtained with the nearest neighbor method and the proposed network on testing sets. The signals recovered by the proposed network have much higher SNR on both testing sets. + +the hidden object is contained in adjacent time bins for adjacent detection pairs. It means that the problem can not be solved by simply applying 2-D image superresolution to each time slice. + +In this section, we introduce the neural network which learns the nonlinear signal recovery operator $\Phi$ and the training dataset. Details about the training process and the ablation study are provided in the supplement. + +# 3.1. Network architecture + +Recently, attention mechanism [4, 36] has become increasingly popular in the deep learning society. However, it has been shown that there is no need to apply the attention mechanism to every part of the network [6]. + +To learn the operator, we generalize the attention in attention net $(\mathrm{A}^2\mathrm{N})$ [6] which has two branches and dropout modules for the attention mechanism. One of the branches uses the conventional method to provide a rough estimation of the high resolution signal, and the other branch employs the network. The final output is the sum of the results from these two branches. The network architecture with two branches has been widely studied in the field of image superresolution [2, 19, 43]. + +The proposed signal superresolution network (SSN) is illustrated in Fig. 3. The learning branch is composed of several convolution layers and 3-D attention in attention blocks $(3\mathrm{D - A}^2\mathrm{B})$ . Each block has a dropout module to learn the weight between the attention branch and the non-attention branch, which can determine whether the attention mechanism is needed in the block. The reconstruction module contains several convolution layers to reconstruct the high resolution signal through the extracted feature. The low di + +mensional feature is first interpolated to a high dimensional feature using trilinear interpolation, and then the feature is processed by several convolution layers. The output of the convolution layer is passed to the next part of the reconstruction module and the process is repeated. + +The generalization of the network from $\mathrm{A}^2\mathrm{N}$ [6] to SSN is not trivial. Based on our analysis, the contribution of one point on the hidden object is contained in adjacent time bins for adjacent detection pairs. Conventional interpolation methods will perform poorly when the spacing between adjacent detection pairs gets larger. Thus, 3-D kernels are needed to extract and recover the information, and the size of the kernels is relevant to the spacing. In addition, the signal is only interpolated in spatial dimensions instead of all three dimensions to make the operator easier to learn. + +# 3.2. Dataset + +For all experiments in this paper, we use subsets of Fashion-MNIST [40] and ShapeNet [5] to generate synthetic signals for training and testing. The Fashion-MNIST dataset consists of 2-D figures, while the ShapeNet comprises 3-D objects. We choose 5,000 samples from the training set of Fashion-MNIST and the class "Car" of ShapeNet respectively to generate our training set. In addition, we choose 1,000 samples from the testing set of Fashion-MNIST and the class "Car" of ShapeNet respectively to generate the testing set. Thus, the training set used in this work comprises 10,000 signal pairs, and the testing set comprises 2,000. All the signals are generated using Eq. (4). Besides, we set $N_{x} = N_{y} = 8$ , $N_{t} = 100$ in this paper. For measured signals whose $N_{t}$ is larger than 100, the network can also be applied directly since the network operates locally. + +For the confocal scenario, to match the experimental setting with the Stanford dataset [22], the 2-D figures from FashionMNIST [40] are placed $1\mathrm{m}$ away from the visible wall, the size of the figures is set as $1\times 1\mathrm{m}^2$ . And the 3-D objects from ShapeNet [5] are placed in a $1\times 1\times 0.2\mathrm{m}^3$ (horizontal, vertical, depth) box with a maximum depth of $1.1\mathrm{m}$ . The illumination region on the visible wall is $2\times 2\mathrm{m}^2$ , and the time resolution of the SPAD is $32~\mathrm{ps}$ . + +For the non-confocal setting, we match the experimental setting with the data provided by phasor field [25]. The 2-D figures from Fashion-MNIST are placed $1\mathrm{m}$ away from the wall, and the size is $1.25\times 1.25\mathrm{m}^2$ . The 3-D objects from ShapeNet are placed in a $1.25\times 1.25\times 0.2\mathrm{m}^3$ (horizontal, vertical, depth) box with a maximum depth of $1.1\mathrm{m}$ . The illumination region on the visible wall is $1.25\times 1.25\mathrm{m}^2$ , and the time resolution of the SPAD is 16 ps. + +# 4. Results + +In this section, we show the reconstruction results of the proposed pipeline on both synthetic and measured data. For + +![](images/c59a5991ae980e8fbc1943f5cb87dc6d2835595d0868a6aea70a94cacea9a769.jpg) +Figure 5. Recovered signals and reconstruction results of the pyramid. (a) Ground truth of the hidden object. (b) A comparison of the recovered signals (the first arrival time of the original signal is marked by the blue curve in the zoom-in window of each subfigure). (c) Reconstruction results of state-of-the-art methods with different signals. + +each instance, reconstruction results obtained with signals of different spatial resolutions are shown to validate the effectiveness of the proposed pipeline. The first column (Original) shows reconstructions obtained from the original signal with a spatial resolution of $32 \times 32$ ; results in the second column (Low) are reconstructed using the signal scanned at $8 \times 8$ points; the third column (Nearest) shows reconstructions of the recovered signal obtained with nearest neighbor interpolation; the last column (Ours) illustrates reconstruction results obtained with the proposed pipeline. Comparisons with other interpolation methods and existing learning based NLOS imaging method are provided in the supplement. + +To illustrate the proposed pipeline can be acted as a plugin-and-play module for most algorithms, we combine state-of-the-art methods with our pipeline. For confocal experiments, the results are reconstructed by F-K [22], LCT [31], LOG-BP [18], D-LCT [45] and SOCR [27] algorithms, while for non-confocal experiments, results are given by LOG-BP, Phasor [24] and SOCR methods. For each instance, the parameters are fixed in each algorithm. + +To assess the results quantitatively, we compute the signal to noise ratio (SNR) of the recovered signals, as well as the peak signal to noise ratio (PSNR) and structural similarity index measure (SSIM) of the reconstructed maximum intensity projections. For measured data, we use the reconstructions obtained from the original signals as references to compute PSNR and SSIM. + +# 4.1. Results of confocal experiment + +In Fig. 4, we compare the mean SNR values of the signals interpolated with SSN and the nearest neighbor method. It is shown that the network provides much better results on both datasets. In other words, the network approximates the oracle interpolation operator $\Phi$ with high accuracy and serves as a crucial part of the proposed pipeline. + +To test the robustness of the proposed pipeline, we use the instance of the pyramid from SOCR [27], whose signal was generated with a different physical model. As shown in Fig. 5b, the first arrival time is correctly recovered by the proposed network, while the nearest neighbor interpolation method fails. Fig. 5c shows that the proposed pipeline provides a signal for high quality reconstructions with better visual quality and higher PSNR and SSIM values. + +The proposed pipeline is then tested on measured data from Stanford dataset [22]. This dataset contains signals measured with different hidden objects and exposure time. We choose the instance of the statue to validate the effectiveness of the proposed pipeline. The measurement time of each sampling point is $0.0069\mathrm{s}$ . The total exposure time of the signal scanned at $32\times 32$ points is $7.03\mathrm{s}$ , and it only takes $0.44\mathrm{s}$ to scan $8\times 8$ points. As is shown in Fig. 6b, the result of SSN has higher SNR. The first arrival time can + +![](images/b052fceca843684b67e7d0eb7163a0c3e5a5e1ab70230ab06882186645278365.jpg) +Figure 6. Recovered signals and reconstruction results of the statue. (a) Ground truth of the hidden object. (b) A comparison of recovered signals (the first arrival time of the original signal is marked by the blue curve in the zoom-in window of each subfigure). (c) Reconstruction results of state-of-the-art methods with different signals. + +hardly be distinguished in the nearest neighbor interpolated signal, while it is much clearer in the signal recovered by the proposed network. As shown in Fig. 6c, the proposed pipeline results in fewer artifacts in the background. More comparisons of different exposure times are provided in the supplement, which also indicates the outperformance of the proposed pipeline. + +# 4.2. Results of the dynamic scene + +To illustrate the potential application of the proposed pipeline on real-time NLOS imaging, we test it on the captured dynamic scenes from the Stanford dataset. The data is captured over a $2 \times 2\mathrm{m}^2$ region on the visible wall, and the person is wearing a retroreflective suit to increase the detection of the indirect signal. It takes $0.25\mathrm{s}$ to scan $32 \times 32$ points, which allows for a reconstruction video with 4 frames per second. The reflection model of the retroreflective material [31] is different from the one defined in Eq. (6). + +The results of several frames are shown in Fig. 7b, the complete video of the reconstruction results can be found in the supplement. All the results are reconstructed by the F-K method. Although the training set is generated by a different forward model, the proposed pipeline can still recover signals for high quality reconstructions. The pose of the person can be clearly identified in the reconstructions using the proposed pipeline, while the reconstructions using the nearest neighbor interpolated signal contain more artifacts. The mean PSNR and SSIM of the reconstruction results obtained from the proposed pipeline are 24.46 and 0.63, while those obtained with nearest neighbor interpolation are only 18.68 and 0.37. + +The total capture time of each frame is reduced from 0.25 s to 0.0156 s. This enables us to obtain a high quality reconstruction video with 64 frames per second. Besides, it only takes 0.08 s to recover the high resolution signal with the proposed network, and 0.01s to reconstruct the hidden object with the F-K method on an NVIDIA RTX 3090. Thus, real-time video imaging with a high frame rate can be realized with a delay of a few frames. + +# 4.3. Results of non-confocal experiment + +For the non-confocal setting, we use the instance of "4" provided by Phasor [24] to validate the effectiveness of the proposed method. The measured signal is subsampled to $32 \times 32$ and the illumination region on the visible wall is cropped to $1.25 \times 1.25\mathrm{m}^2$ . We use the signal scanned at 8 $\times 8$ points as the input of our pipeline. It takes 1 s to scan each point on the visible wall. + +As is shown in Fig. 8b, the SNR of the signal recovered by the network is much higher than the one interpolated with the nearest neighbor method. Reconstruction results are shown in Fig. 8c. The proposed pipeline provides re + +![](images/4cf3f4389b401b9c7c3cf6ed7462d1ca6b347a6759fec99723260566fe74fdd9.jpg) +(a) GT + +![](images/1ce59b218de5a7c16fda97620e7f7cf32a22da1c216a90d19c35cdeaf93deb5a.jpg) +(b) Reconstruction results +Figure 7. Reconstruction results of the dynamic scene. (a) An illustration of the dynamic scene. (b) Reconstruction results of the dynamic scene. The results are obtained with the F-K method. The PSNR and SSIM values are computed with the reconstructions obtained with the original signals as references. The proposed pipeline provides much better reconstructions in terms of both quantitative criteria and visual quality. + +constructions with higher quantitative and visual quality. + +# 5. Discussion + +We have introduced a novel pipeline for high quality NLOS reconstruction with the signal measured at only a small number of scanning points. In this section, we discuss several factors that affect the reconstruction quality and some feasible solutions to improve the pipeline further. + +# 5.1. Training of the network + +The training of the network employed in the proposed pipeline is supervised. Thus, it is affected by the generation of the training set, including the forward model, experimental setting, and hidden objects. For example, hidden objects + +with different materials will lead to various physical models. The networks trained and tested on signals with different models would bring additional errors to the results. + +However, a network with good generalization ability can avoid degradation. As shown in Fig. 7, although the physical model of the dynamic scenes differs from the one used in the training set, the reconstructed objects are only slightly affected. Thus, the proposed pipeline has strong robustness to the forward model. + +# 5.2. Performance of different algorithms + +As the reconstruction results shown in previous sections, the improvement of reconstruction quality of the proposed pipeline varies between different methods. Compared to the reconstruction results obtained from the low resolution signal, fast imaging algorithms like F-K, LCT and D-LCT have larger improvements, while the iterative inversion algorithm SOCR might have similar visual quality in some cases. This phenomenon may come from the fact that the iterative algorithm can make full use of the information contained in the signal, but with a longer reconstruction time. + +# 5.3. Combination with SPAD array for further acceleration + +Since the proposed pipeline can recover the signal faithfully under the non-confocal setting, combining it with a SPAD array can accelerate the measurement process further. However, the existing SPAD array can only focus on a rectangular area on the visible wall, which means that the network has to learn an extrapolation operator. This is much harder than learning an interpolation operator. Thus, it would be more practical if the SPAD array could focus on sparse rectangular points on a relatively large area of the visible wall. This may be seen as a new area for hardware design. + +# 6. Conclusion + +In this paper, we have studied the signal recovery problem in NLOS imaging, and proposed a two-step learning based pipeline for fast NLOS imaging with only a small number of scanning points. In the first step, the signal of high spatial resolution is recovered with a neural network. In the second step, the hidden objects are obtained from the virtually measured signals at a dense grid. Our pipeline provides high-quality reconstructions under both confocal and non-confocal settings, which takes only $6.25\%$ of the original measurement time. Besides, the proposed pipeline is compatible with most existing detection systems and imaging algorithms. With the proposed pipeline, the high frame rate imaging video is possible, which is a noteworthy advance in this area. It is expected to combine emerging techniques in the fields of SPAD array for real-time and high quality NLOS reconstructions with higher frame rates. + +![](images/6d02ab8bd732e183ac0ccfaf3814939689e1178c80bfa33a1c3a7f6b82a5e7bc.jpg) + +![](images/98394bb3ec72c783acb60f1f019e86e01f3c3ac1300f249bb043f42b8cecddc1.jpg) +(a) GT + +![](images/b80a4457d701df0888351b443cdbef191820a0492924a2217592583be92c4abe.jpg) +(b) Recovered signals +(c) Reconstruction results +Figure 8. Recovered signals and reconstruction results of the instance "4". (a) Ground truth of the hidden object. (b) A comparison of recovered signals (the first arrival time of the original signal is marked by the blue curve in the zoom-in window of each subfigure). (c) Reconstruction results of state-of-the-art methods with different signals. + +Acknowledgments. The authors would like to thank Zhu-peng Li and Rongqian Wang for fruitful discussions. This work was supported by the National Natural Science Foundation of China (61975087, 12071244, 11971258). + +# References + +[1] Byeongjoo Ahn, Akshat Dave, Ashok Veeraraghavan, Ioannis Gkioulekas, and Aswin C Sankaranarayanan. Convolutional approximations to the general non-line-of-sight imag + +ing operator. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7889-7899, 2019. 2 +[2] Saeed Anwar, Salman Khan, and Nick Barnes. A deep journey into super-resolution: A survey. ACM Computing Surveys (CSUR), 53(3):1-34, 2020. 4 +[3] Victor Arellano, Diego Gutierrez, and Adrian Jarabo. Fast back-projection for non-line of sight reconstruction. In ACM SIGGRAPH 2017 Posters, SIGGRAPH '17, New York, NY, USA, 2017. Association for Computing Machinery. 2 +[4] Syed Muhammad Arsalan Bashir, Yi Wang, Mahrukh Khan, and Yilong Niu. A comprehensive review of deep learning-based single image super-resolution. PeerJ Computer Science, 7:e621, 2021. 4 +[5] Angel X. Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. ShapeNet: An Information-Rich 3D Model Repository. Technical Report arXiv:1512.03012 [cs.GR], Stanford University — Princeton University — Toyota Technological Institute at Chicago, 2015. 5 +[6] Haoyu Chen, Jinjin Gu, and Zhi Zhang. Attention in attention network for image super-resolution. arXiv preprint arXiv:2104.09497, 2021. 4 +[7] Wenzheng Chen, Simon Daneau, Fahim Mannan, and Felix Heide. Steady-state non-line-of-sight imaging. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2019. 2 +[8] Wenzheng Chen, Fangyin Wei, Kiriakos N Kutulakos, Szymon Rusinkiewicz, and Felix Heide. Learned feature embeddings for non-line-of-sight imaging and recognition. ACM Transactions on Graphics (TOG), 39(6):1-18, 2020. 2 +[9] Justin Dove and Jeffrey H. Shapiro. Paraxial theory of phasor-field imaging. Opt. Express, 27(13):18016-18037, Jun 2019. 2 +[10] Xiaohua Feng and Liang Gao. Improving non-line-of-sight image reconstruction with weighting factors. Optics letters, 45(14):3921-3924, 2020. 2 +[11] Xiaohua Feng and Liang Gao. Ultrafast light field tomography for snapshot transient and non-line-of-sight imaging. Nature Communications, 12:2179, 04 2021. 2 +[12] Javier Grau Chopite, Matthias B. Hullin, Michael Wand, and Julian Iseringhausen. Deep non-line-of-sight reconstruction. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 957-966, 2020. 2 +[13] Felix Heide, Matthew O'Toole, Kai Zang, David B. Lindell, Steven Diamond, and Gordon Wetzstein. Non-line-of-sight imaging with partial occluders and surface normals. ACM Trans. Graph., 38(3), May 2019. 2 +[14] Julian Iseringhausen and Matthias B Hullin. Non-line-of-sight reconstruction using efficient transient rendering. ACM Transactions on Graphics (TOG), 39(1):1-14, 2020. 2 +[15] Mariko Isogawa, Dorian Chan, Ye Yuan, Kris M. Kitani, and Matthew O'Toole. Efficient non-line-of-sight imaging from transient sinograms. In 16th European Conference on Computer Vision (ECCV), 2020. 2, 3 +[16] Mariko Isogawa, Ye Yuan, Matthew O'Toole, and Kris Kitani. Optical non-line-of-sight physics-based 3d human pose + +estimation. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7011-7020, 2020. 2 +[17] Ahmed Kirmani, Tyler Hutchison, James Davis, and Ramesh Raskar. Looking around the corner using transient imaging. In 2009 IEEE 12th International Conference on Computer Vision, pages 159–166, 2009. 1 +[18] Martin Laurenzis and Andreas Velten. Feature selection and back-projection algorithms for nonlinear-of-sight laser-gated viewing. Journal of Electronic Imaging, 23:063003, 11 2014. 2, 3, 6 +[19] Y. Li, B. Sixou, and F. Peyrin. A review of the deep learning methods for medical images super resolution problems. IRBM, 42(2):120-133, 2021. 4 +[20] Zhupeng Li, Xintong Liu, Jianyu Wang, Zuoqiang Shi, Lingyun Qiu, and Xing Fu. Fast non-line-of-sight imaging based on first photon event stamping. Optics letters, 47(8):1928—1931, April 2022. 2 +[21] Zhengpeng Liao, Deyang Jiang, Xiaochun Liu, Andreas Velten, Yajun Ha, and Xin Lou. FPGA accelerator for real-time non-line-of-sight imaging. IEEE Transactions on Circuits and Systems I: Regular Papers, 2022. 2 +[22] David B Lindell, Gordon Wetzstein, and Matthew O'Toole. Wave-based non-line-of-sight imaging using fast fk migration. ACM Transactions on Graphics (ToG), 38(4):1-13, 2019. 2, 3, 5, 6 +[23] Jianjiang Liu, Yijun Zhou, Xin Huang, Zheng-Ping Li, and Feihu Xu. Photon-efficient non-line-of-sight imaging. IEEE Transactions on Computational Imaging, 8:639-650, 2022. 2 +[24] Xiaochun Liu, Sebastian Bauer, and Andreas Velten. Phasor field diffraction based reconstruction for fast non-line-of-sight imaging systems. Nature Communications, 11:1645, 2020. 2, 3, 6, 7 +[25] Xiaochun Liu, Ibón Guillén, Marco Manna, Ji Nam, Azer Reza, Toan Le, Adrián Jarabo, Diego Gutiérrez, and Andreas Velten. Non-line-of-sight imaging using phasor-field virtual wave optics. Nature, 572, 2019. 2, 5 +[26] Xiaochun Liu and Andreas Velten. The role of wigner distribution function in non-line-of-sight imaging. In 2020 IEEE International Conference on Computational Photography (ICCP), pages 1-12. IEEE, 2020. 2 +[27] Xintong Liu, Jianyu Wang, Zhupeng Li, Zuoqiang Shi, Xing Fu, and Lingyun Qiu. Non-line-of-sight reconstruction with signal-object collaborative regularization. Light: Science & Applications, 08 2021. 2, 3, 6 +[28] Christopher Metzler, David Lindell, and Gordon Wetzstein. Keyhole imaging: Non-line-of-sight imaging and tracking of moving objects along a single optical path. IEEE Transactions on Computational Imaging, PP:1-1, 12 2020. 2, 3 +[29] Christopher A. Metzler, Felix Heide, Prasana Rangarajan, Muralidhar Madabhushi Balaji, Aparna Viswanath, Ashok Veeraraghavan, and Richard G. Baraniuk. Deep-inverse correlography: towards real-time high-resolution non-line-of-sight imaging. Optica, 7(1):63-71, Jan 2020. 2 +[30] Ji Hyun Nam, Eric Brandt, Sebastian Bauer, Xiaochun Liu, Marco Renna, Alberto Tosi, Eftychios Sifakis, and Andreas + +Velten. Low-latency time-of-flight non-line-of-sight imaging at 5 frames per second. Nature communications, 12(1):1-10, 2021. 2, 4 +[31] Matthew O'Toole, David Lindell, and Gordon Wetzstein. Confocal non-line-of-sight imaging based on the light-cone transform. Nature, 555, 2018. 2, 3, 6, 7 +[32] Chengquan Pei, Anke Zhang, Yue Deng, Feihu Xu, Jiamin Wu, David U-Lei Li, Hui Qiao, Lu Fang, and Qionghai Dai. Dynamic non-line-of-sight imaging system based on the optimization of point spread functions. Opt. Express, 29(20):32349-32364, Sep 2021. 2 +[33] Syed Azer Reza, Marco La Manna, Sebastian Bauer, and Andreas Velten. Phasor field waves: A huygens-like light transport model for non-line-of-sight imaging applications. Opt. Express, 27(20):29380-29400, Sep 2019. 2 +[34] Christos Thrampoulidis, Gal Shulkind, Feihu Xu, William T. Freeman, Jeffrey H. Shapiro, Antonio Torralba, Franco N. C. Wong, and Gregory W. Wornell. Exploiting occlusion in non-line-of-sight active imaging. IEEE Transactions on Computational Imaging, 4(3):419-431, 2018. 2 +[35] Chia-Yin Tsai, Aswin C Sankaranarayanan, and Ioannis Gkioulekas. Beyond volumetric albedo-a surface optimization framework for non-line-of-sight imaging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1545-1555, 2019. 2 +[36] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Proceedings of the 31st International Conference on Neural Information Processing Systems, NIPS'17, page 6000-6010, Red Hook, NY, USA, 2017. Curran Associates Inc. 4 +[37] Eric Veach and Leonidas J. Guibas. Optimally combining sampling techniques for monte carlo rendering. In Proceedings of the 22nd Annual Conference on Computer Graphics and Interactive Techniques, SIGGRAPH '95, page 419-428, New York, NY, USA, 1995. Association for Computing Machinery. 3 +[38] Andreas Velten, Thomas Willwacher, Otkrist Gupta, Ashok Veeraraghavan, Moungi G Bawendi, and Ramesh Raskar. Recovering three-dimensional shape around a corner using ultrafast time-of-flight imaging. Nature communications, 3(1):1-8, 2012. 2 +[39] Cheng Wu, Jianjiang Liu, Xin Huang, Zheng-Ping Li, Chao Yu, Jun-Tian Ye, Jun Zhang, Qiang Zhang, Xiankang Dou, Vivek Goyal, Feihu Xu, and Jian-Wei Pan. Non-line-of-sight imaging over $1.43\mathrm{km}$ . Proceedings of the National Academy of Sciences of the United States of America, 118, 03 2021. 2 +[40] Han Xiao, Kashif Rasul, and Roland Vollgraf. Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms. arXiv preprint arXiv:1708.07747, 2017. 5 +[41] Shumian Xin, Sotiris Nousias, Kiriakos N. Kutulakos, Aswin C. Sankaranarayanan, Srinivasa G. Narasimhan, and Ioannis Gkioulekas. A theory of fermat paths for non-line-of-sight shape reconstruction. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6793-6802, 2019. 2 + +[42] Wenqing Yang, Chao Zhang, Wenjie Jiang, Zexin Zhang, and Baoqing Sun. None-line-of-sight imaging enhanced with spatial multiplexing. Opt. Express, 30(4):5855-5867, Feb 2022. 2, 4 +[43] Wenming Yang, Xuechen Zhang, Yapeng Tian, Wei Wang, Jing-Hao Xue, and Qingmin Liao. Deep learning for single image super-resolution: A brief review. IEEE Transactions on Multimedia, 21(12):3106-3121, Dec 2019. 4 +[44] Jun-Tian Ye, Xin Huang, Zheng-Ping Li, and Feihu Xu. Compressed sensing for active non-line-of-sight imaging. Opt. Express, 29(2):1749-1763, Jan 2021. 2, 3 +[45] Sean I Young, David B Lindell, Bernd Girod, David Taubman, and Gordon Wetzstein. Non-line-of-sight surface reconstruction using the directional light-cone transform. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1407-1416, 2020. 2, 3, 6 \ No newline at end of file diff --git a/2023/Non-Line-of-Sight Imaging With Signal Superresolution Network/images.zip b/2023/Non-Line-of-Sight Imaging With Signal Superresolution Network/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..11da812b18bcc3c89ef6ff974089d341de571ca0 --- /dev/null +++ b/2023/Non-Line-of-Sight Imaging With Signal Superresolution Network/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:38aa3943bba3f5e4ec67e06e1146014d59055f448f8029ee37e5fd2d2102f06b +size 613238 diff --git a/2023/Non-Line-of-Sight Imaging With Signal Superresolution Network/layout.json b/2023/Non-Line-of-Sight Imaging With Signal Superresolution Network/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..0ae79d4658cffcf06d7ec4e0390c9165aeb02aa0 --- /dev/null +++ b/2023/Non-Line-of-Sight Imaging With Signal Superresolution Network/layout.json @@ -0,0 +1,8246 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 100, + 103, + 493, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 103, + 493, + 121 + ], + "spans": [ + { + "bbox": [ + 100, + 103, + 493, + 121 + ], + "type": "text", + "content": "Non-line-of-sight Imaging with Signal Superresolution Network" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 75, + 142, + 516, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 142, + 516, + 171 + ], + "spans": [ + { + "bbox": [ + 75, + 142, + 516, + 171 + ], + "type": "text", + "content": "Jianyu Wang" + }, + { + "bbox": [ + 75, + 142, + 516, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 75, + 142, + 516, + 171 + ], + "type": "text", + "content": ", Xintong Liu" + }, + { + "bbox": [ + 75, + 142, + 516, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 75, + 142, + 516, + 171 + ], + "type": "text", + "content": ", Leping Xiao" + }, + { + "bbox": [ + 75, + 142, + 516, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 75, + 142, + 516, + 171 + ], + "type": "text", + "content": ", Zuoqiang Shi" + }, + { + "bbox": [ + 75, + 142, + 516, + 171 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 75, + 142, + 516, + 171 + ], + "type": "text", + "content": ", Lingyun Qiu" + }, + { + "bbox": [ + 75, + 142, + 516, + 171 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 75, + 142, + 516, + 171 + ], + "type": "text", + "content": ", and Xing Fu" + }, + { + "bbox": [ + 75, + 142, + 516, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 75, + 142, + 516, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 75, + 142, + 516, + 171 + ], + "type": "text", + "content": "Tsinghua University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 117, + 171, + 473, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 171, + 473, + 186 + ], + "spans": [ + { + "bbox": [ + 117, + 171, + 473, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 117, + 171, + 473, + 186 + ], + "type": "text", + "content": "Yanqi Lake Beijing Institute of Mathematical Sciences and Applications" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 239, + 290, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 239, + 290, + 514 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 290, + 514 + ], + "type": "text", + "content": "Non-line-of-sight (NLOS) imaging aims at reconstructing the location, shape, albedo, and surface normal of the hidden object around the corner with measured transient data. Due to its strong potential in various fields, it has drawn much attention in recent years. However, long exposure time is not always available for applications such as auto-driving, which hinders the practical use of NLOS imaging. Although scanning fewer points can reduce the total measurement time, it also brings the problem of imaging quality degradation. This paper proposes a general learning-based pipeline for increasing imaging quality with only a few scanning points. We tailor a neural network to learn the operator that recovers a high spatial resolution signal. Experiments on synthetic and measured data indicate that the proposed method provides faithful reconstructions of the hidden scene under both confocal and nonconfocal settings. Compared with original measurements, the acquisition of our approach is 16 times faster while maintaining similar reconstruction quality. Besides, the proposed pipeline can be applied directly to existing optical systems and imaging algorithms as a plug-in-and-play module. We believe the proposed pipeline is powerful in increasing the frame rate in NLOS video imaging." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 536, + 128, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 536, + 128, + 550 + ], + "spans": [ + { + "bbox": [ + 47, + 536, + 128, + 550 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 557, + 287, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 287, + 664 + ], + "type": "text", + "content": "Non-line-of-sight (NLOS) imaging problem usually employs a high temporal resolution optical system to recover the hidden object around the corner. As shown in Fig. 1a, photons emitted by the laser are collected by the detector after three diffuse reflections: the reflection at the visible wall, the reflection at the hidden object, and the reflection at the visible wall again. Scanning a region on the visible wall can get measured histogram data to recover the hidden scene." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": "Due to its potential applications in various fields, such as auto-driving, disaster relief, and remote sensing, the NLOS imaging problem has become an emerging field since it was first proposed by Kirmani et al. [17] in 2009." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 312, + 212, + 541, + 378 + ], + "blocks": [ + { + "bbox": [ + 312, + 212, + 541, + 378 + ], + "lines": [ + { + "bbox": [ + 312, + 212, + 541, + 378 + ], + "spans": [ + { + "bbox": [ + 312, + 212, + 541, + 378 + ], + "type": "image", + "image_path": "429a37763f5f9637697c2908d9b5c6d7d715ed12096e8194d4c7cf78eb55b456.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 316, + 397, + 423, + 505 + ], + "blocks": [ + { + "bbox": [ + 316, + 397, + 423, + 505 + ], + "lines": [ + { + "bbox": [ + 316, + 397, + 423, + 505 + ], + "spans": [ + { + "bbox": [ + 316, + 397, + 423, + 505 + ], + "type": "image", + "image_path": "6394929e8689e54aa2c5e7e2876a09f74b7be49627ef25f50cfe9fe009d4f5b7.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 335, + 507, + 400, + 517 + ], + "lines": [ + { + "bbox": [ + 335, + 507, + 400, + 517 + ], + "spans": [ + { + "bbox": [ + 335, + 507, + 400, + 517 + ], + "type": "text", + "content": "(b) Scanning points" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 429, + 393, + 476, + 447 + ], + "blocks": [ + { + "bbox": [ + 394, + 382, + 457, + 392 + ], + "lines": [ + { + "bbox": [ + 394, + 382, + 457, + 392 + ], + "spans": [ + { + "bbox": [ + 394, + 382, + 457, + 392 + ], + "type": "text", + "content": "(a) NLOS scenario" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 429, + 393, + 476, + 447 + ], + "lines": [ + { + "bbox": [ + 429, + 393, + 476, + 447 + ], + "spans": [ + { + "bbox": [ + 429, + 393, + 476, + 447 + ], + "type": "image", + "image_path": "b26ef31d3c80ab0bc1c79602292b638c5999426563c750bf3cf914cfcd211f77.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 429, + 450, + 460, + 457 + ], + "lines": [ + { + "bbox": [ + 429, + 450, + 460, + 457 + ], + "spans": [ + { + "bbox": [ + 429, + 450, + 460, + 457 + ], + "type": "text", + "content": "Low" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 479, + 393, + 540, + 506 + ], + "blocks": [ + { + "bbox": [ + 479, + 393, + 540, + 506 + ], + "lines": [ + { + "bbox": [ + 479, + 393, + 540, + 506 + ], + "spans": [ + { + "bbox": [ + 479, + 393, + 540, + 506 + ], + "type": "image", + "image_path": "c44ef8d632ba6e3af2af42f6d0c0ef089fbac977395c26c62a0681b607f5b73c.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 466, + 510, + 503, + 519 + ], + "lines": [ + { + "bbox": [ + 466, + 510, + 503, + 519 + ], + "spans": [ + { + "bbox": [ + 466, + 510, + 503, + 519 + ], + "type": "text", + "content": "(c) Results" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 529, + 547, + 672 + ], + "lines": [ + { + "bbox": [ + 304, + 529, + 547, + 672 + ], + "spans": [ + { + "bbox": [ + 304, + 529, + 547, + 672 + ], + "type": "text", + "content": "Figure 1. A typical NLOS scenario and reconstruction results. (a) An illustration of a typical NLOS scenario. (b) An illustration of the scanning points on the visible wall. Using the proposed pipeline, only the blue circle points are needed to be illuminated, and the signal at the yellow square points can be recovered with the proposed pipeline. (c) Comparisons of the reconstruction results of the pyramid with different signals. The maximum intensity projection of the albedo values along the depth direction is shown in the upper left corner (GT). The results are reconstructed by the original signal with spatial resolution " + }, + { + "bbox": [ + 304, + 529, + 547, + 672 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 304, + 529, + 547, + 672 + ], + "type": "text", + "content": " (Original), the subsampled signal with spatial resolution " + }, + { + "bbox": [ + 304, + 529, + 547, + 672 + ], + "type": "inline_equation", + "content": "8 \\times 8" + }, + { + "bbox": [ + 304, + 529, + 547, + 672 + ], + "type": "text", + "content": " (Low), and the signal recovered by the proposed network with spatial resolution " + }, + { + "bbox": [ + 304, + 529, + 547, + 672 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 304, + 529, + 547, + 672 + ], + "type": "text", + "content": " (Ours)." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "type": "text", + "content": "Many methods have been proposed to improve the prac" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "17420" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 80, + 70, + 516, + 180 + ], + "blocks": [ + { + "bbox": [ + 80, + 70, + 516, + 180 + ], + "lines": [ + { + "bbox": [ + 80, + 70, + 516, + 180 + ], + "spans": [ + { + "bbox": [ + 80, + 70, + 516, + 180 + ], + "type": "image", + "image_path": "cc7ccacf4ab46248265f69b400c2fa39b4250dd6c0c0826939aa02df70a2837d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 188, + 547, + 212 + ], + "lines": [ + { + "bbox": [ + 46, + 188, + 547, + 212 + ], + "spans": [ + { + "bbox": [ + 46, + 188, + 547, + 212 + ], + "type": "text", + "content": "Figure 2. Flowchart of the proposed pipeline. The high resolution signal is recovered from the low resolution signal using a neural network. The hidden scene is then reconstructed using state-of-the-art imaging algorithms designed for high resolution signal." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 231, + 289, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 231, + 289, + 506 + ], + "spans": [ + { + "bbox": [ + 46, + 231, + 289, + 506 + ], + "type": "text", + "content": "ticability [1, 3, 13, 14, 34, 35, 39] or reconstruction quality [10, 11, 26, 41]. A back-projection method with Laplacian of Gaussian filter (LOG-BP) [18] is introduced to improve the reconstruction quality of the back-projection method [38]. In 2018, O'Toole et al. [31] first apply the light-contransform (LCT) method to image the hidden object with a confocal scanning mode. Young et al. [45] then propose the directional light-cone-transform (D-LCT) method, which simultaneously recovers the albedo and surface normal of the hidden object. From the perspective of wave characteristics, Lindell et al. [22] introduce the fast frequency-wavenumber migration (F-K) method to NLOS. Another remarkable imaging method proposed by Liu et al. [25] employs the phasor field model [9, 33] and provides an efficient solution for fast NLOS imaging under non-confocal scenario [24]. Further improvements of this method using single photon avalanche diode (SPAD) arrays achieve real-time imaging [21, 30]. Deep learning based methods have also been introduced [7, 8, 12, 16, 29]. Recently, Liu et al. [27] propose the signal-object collaborative regularization (SOCR) method, which can improve the resolution of the reconstructed object, especially in cases with high measurement noise." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "text", + "content": "However, one main obstacle in NLOS imaging is that we need a long exposure time to improve the signal-to-noise ratio (SNR) of the measurement, which is not admissible in many scenarios, such as auto-driving. Several methods have been proposed to reduce the capture time [20, 23]. For a confocal scenario, Isogawa et al. [15] propose a circular scanning mode. In 2021, Metzler et al. [28] consider an extreme case that tracks the objects with a single detection point. By introducing the downsampling operation, Ye et al. [44] reconstruct the object with a subset of the virtual sources. For non-confocal scenarios, most methods employ SPAD arrays to realize fast measurement. Among them, Nam et al. [30] achieve a reconstruction of five frames per second by remapping the data from two " + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "inline_equation", + "content": "16 \\times 1" + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "text", + "content": " SPAD arrays. Pei et al. [32] realize 20 frames per second reconstruction using a " + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "text", + "content": " SPAD array. In 2022, Yang et al. [42] propose a spatial multiplexing method based on the" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 232, + 545, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 232, + 545, + 278 + ], + "spans": [ + { + "bbox": [ + 304, + 232, + 545, + 278 + ], + "type": "text", + "content": "compressed sensing strategy. Although these methods can reduce the measurement time, they might either cause deterioration in imaging quality or substantially increase expense." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 279, + 545, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 279, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 304, + 279, + 545, + 363 + ], + "type": "text", + "content": "In this paper, we propose a deep learning based pipeline that can reconstruct the hidden object with fine details using only a small number of scanning points. Different measurement densities are shown in Fig. 1b. In the following sections, we term the signal measured at the blue circle points as the \"low resolution signal\" while those measured at both circle and square points as the \"high resolution signal\"." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 363, + 546, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 363, + 546, + 518 + ], + "spans": [ + { + "bbox": [ + 304, + 363, + 546, + 518 + ], + "type": "text", + "content": "Instead of designing new algorithms for sparse measurements, we focus on recovering the dense virtual signals, which can be applied to most algorithms as a plug-in-and-play module. The proposed pipeline consists of two steps: the low resolution signal is first mapped into the high resolution signal space by a neural network; the reconstruction is then obtained from the recovered high resolution signal using existing imaging algorithms. The reconstruction results of the pyramid are shown in Fig. 1c. These results are reconstructed by the F-K method. As the comparison shows, the proposed pipeline can provide a high quality reconstruction with only 64 scanning points, which is " + }, + { + "bbox": [ + 304, + 363, + 546, + 518 + ], + "type": "inline_equation", + "content": "6.25\\%" + }, + { + "bbox": [ + 304, + 363, + 546, + 518 + ], + "type": "text", + "content": " of the original number." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 519, + 545, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 519, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 304, + 519, + 545, + 590 + ], + "type": "text", + "content": "To our best knowledge, this is the first work in NLOS to study the signal recovery problem of the low resolution signal sparsely measured on a uniform grid. The proposed method can reduce the exposure time without losing the reconstruction quality. The main contributions of this work are:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 315, + 599, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 315, + 599, + 545, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 599, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 315, + 599, + 545, + 635 + ], + "type": "text", + "content": "- We propose a novel learning based pipeline for faithful reconstruction of the hidden scene, which only takes " + }, + { + "bbox": [ + 315, + 599, + 545, + 635 + ], + "type": "inline_equation", + "content": "6.25\\%" + }, + { + "bbox": [ + 315, + 599, + 545, + 635 + ], + "type": "text", + "content": " of the original measuring time." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 315, + 644, + 545, + 680 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 644, + 545, + 680 + ], + "spans": [ + { + "bbox": [ + 315, + 644, + 545, + 680 + ], + "type": "text", + "content": "- We show that the high resolution signal can be recovered accurately with the proposed signal superresolution network, which we abbreviate as \"SSN\"." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 315, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 315, + 689, + 545, + 713 + ], + "type": "text", + "content": "- The effectiveness of the proposed method is validated on public datasets under both confocal and non" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17421" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 72, + 516, + 233 + ], + "blocks": [ + { + "bbox": [ + 82, + 72, + 516, + 233 + ], + "lines": [ + { + "bbox": [ + 82, + 72, + 516, + 233 + ], + "spans": [ + { + "bbox": [ + 82, + 72, + 516, + 233 + ], + "type": "image", + "image_path": "fd498f1a352f288178a504422cc88beab378c33bdd54c93a0803636122ed287c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 243, + 547, + 266 + ], + "lines": [ + { + "bbox": [ + 46, + 243, + 547, + 266 + ], + "spans": [ + { + "bbox": [ + 46, + 243, + 547, + 266 + ], + "type": "text", + "content": "Figure 3. The architecture of the signal superresolution network (SSN) in the proposed pipeline. The network is composed of two main branches: a conventional upsampling branch and a learning branch. The final output is the sum of the results from these two branches." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 286, + 141, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 286, + 141, + 300 + ], + "spans": [ + { + "bbox": [ + 67, + 286, + 141, + 300 + ], + "type": "text", + "content": "confocal settings." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 306, + 288, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 306, + 288, + 343 + ], + "spans": [ + { + "bbox": [ + 57, + 306, + 288, + 343 + ], + "type": "text", + "content": "- We demonstrate the potential application of the proposed method in high frame rate (64 frames per second) NLOS video imaging using public datasets." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 353, + 186, + 367 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 353, + 186, + 367 + ], + "spans": [ + { + "bbox": [ + 47, + 353, + 186, + 367 + ], + "type": "text", + "content": "2. Reconstruction Pipeline" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 373, + 287, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 373, + 287, + 434 + ], + "spans": [ + { + "bbox": [ + 46, + 373, + 287, + 434 + ], + "type": "text", + "content": "This work aims to develop a learning based pipeline for NLOS imaging, which can reduce the measurement time while maintaining the reconstruction quality. In this section, we formulate the signal recovery problem and introduce the proposed pipeline." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 440, + 201, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 440, + 201, + 453 + ], + "spans": [ + { + "bbox": [ + 47, + 440, + 201, + 453 + ], + "type": "text", + "content": "2.1. Formulation of the problem" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 458, + 287, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 458, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 46, + 458, + 287, + 483 + ], + "type": "text", + "content": "Most existing methods that reconstruct the hidden object from low resolution signals " + }, + { + "bbox": [ + 46, + 458, + 287, + 483 + ], + "type": "inline_equation", + "content": "d_{l}" + }, + { + "bbox": [ + 46, + 458, + 287, + 483 + ], + "type": "text", + "content": " can be formulated as" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 127, + 491, + 287, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 491, + 287, + 506 + ], + "spans": [ + { + "bbox": [ + 127, + 491, + 287, + 506 + ], + "type": "interline_equation", + "content": "I (x, y, z) = F _ {l} \\left(d _ {l}\\right) \\tag {1}", + "image_path": "26ba9da8229777c94c81f94022712bc6486161e0a508717b664376fda2672dad.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 514, + 287, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 514, + 287, + 585 + ], + "spans": [ + { + "bbox": [ + 46, + 514, + 287, + 585 + ], + "type": "text", + "content": "in which " + }, + { + "bbox": [ + 46, + 514, + 287, + 585 + ], + "type": "inline_equation", + "content": "I(x,y,z)" + }, + { + "bbox": [ + 46, + 514, + 287, + 585 + ], + "type": "text", + "content": " is the reconstructed object and " + }, + { + "bbox": [ + 46, + 514, + 287, + 585 + ], + "type": "inline_equation", + "content": "F_{l}" + }, + { + "bbox": [ + 46, + 514, + 287, + 585 + ], + "type": "text", + "content": " is the imaging algorithm designed for low resolution signal [15, 28, 44]. Since only a few sampling points are scanned, the total exposure time can be reduced. However, the reconstruction quality also decreases with the number of scan points." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 586, + 287, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 586, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 46, + 586, + 287, + 609 + ], + "type": "text", + "content": "On the contrary, the algorithms designed for high resolution signal can be expressed as" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 125, + 619, + 287, + 633 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 619, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 125, + 619, + 287, + 633 + ], + "type": "interline_equation", + "content": "I (x, y, z) = F _ {h} \\left(d _ {h}\\right) \\tag {2}", + "image_path": "aa5b3cdc9a85c603b241921817788300bb30fa036fb403922e70cef1b2647779.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 641, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 287, + 689 + ], + "type": "text", + "content": "in which " + }, + { + "bbox": [ + 46, + 641, + 287, + 689 + ], + "type": "inline_equation", + "content": "F_{h}" + }, + { + "bbox": [ + 46, + 641, + 287, + 689 + ], + "type": "text", + "content": " represents the algorithm designed for high resolution signal " + }, + { + "bbox": [ + 46, + 641, + 287, + 689 + ], + "type": "inline_equation", + "content": "d_{h}" + }, + { + "bbox": [ + 46, + 641, + 287, + 689 + ], + "type": "text", + "content": ". Many works have been proposed for fast [22, 24, 31, 45] and high resolution [18, 27] reconstruction as discussed above." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 689, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 288, + 714 + ], + "type": "text", + "content": "In this work, we aim at designing a pipeline that can provide faithful reconstruction with only a small number of" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 286, + 545, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 286, + 545, + 357 + ], + "spans": [ + { + "bbox": [ + 304, + 286, + 545, + 357 + ], + "type": "text", + "content": "scanning points, and has good compatibility with traditional optical systems as well as existing imaging algorithms. To solve this problem, we introduce a signal recovering operator " + }, + { + "bbox": [ + 304, + 286, + 545, + 357 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 304, + 286, + 545, + 357 + ], + "type": "text", + "content": ", which maps the low resolution signal " + }, + { + "bbox": [ + 304, + 286, + 545, + 357 + ], + "type": "inline_equation", + "content": "d_{l}" + }, + { + "bbox": [ + 304, + 286, + 545, + 357 + ], + "type": "text", + "content": " into the space of high resolution signal " + }, + { + "bbox": [ + 304, + 286, + 545, + 357 + ], + "type": "inline_equation", + "content": "d_{h}" + }, + { + "bbox": [ + 304, + 286, + 545, + 357 + ], + "type": "text", + "content": ". The proposed pipeline can be expressed as" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 377, + 363, + 545, + 377 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 377, + 363, + 545, + 377 + ], + "spans": [ + { + "bbox": [ + 377, + 363, + 545, + 377 + ], + "type": "interline_equation", + "content": "I (x, y, z) = F _ {h} \\left(\\Phi \\left(d _ {l}\\right)\\right) \\tag {3}", + "image_path": "4b81253a00738146b82fb23724bece915b6df0791de177455cd2a81be819978a.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 388, + 546, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 388, + 546, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 388, + 546, + 437 + ], + "type": "text", + "content": "in which " + }, + { + "bbox": [ + 304, + 388, + 546, + 437 + ], + "type": "inline_equation", + "content": "\\Phi(d_l)" + }, + { + "bbox": [ + 304, + 388, + 546, + 437 + ], + "type": "text", + "content": " is of the same size with " + }, + { + "bbox": [ + 304, + 388, + 546, + 437 + ], + "type": "inline_equation", + "content": "d_h" + }, + { + "bbox": [ + 304, + 388, + 546, + 437 + ], + "type": "text", + "content": ". Thus, the problem of NLOS imaging from low resolution signal converts to the problem of finding the operator " + }, + { + "bbox": [ + 304, + 388, + 546, + 437 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 304, + 388, + 546, + 437 + ], + "type": "text", + "content": " which satisfies " + }, + { + "bbox": [ + 304, + 388, + 546, + 437 + ], + "type": "inline_equation", + "content": "\\Phi(d_l) = d_h" + }, + { + "bbox": [ + 304, + 388, + 546, + 437 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 449, + 487, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 449, + 487, + 462 + ], + "spans": [ + { + "bbox": [ + 306, + 449, + 487, + 462 + ], + "type": "text", + "content": "2.2. The choice of recovering operator" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 469, + 545, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 469, + 545, + 518 + ], + "spans": [ + { + "bbox": [ + 304, + 469, + 545, + 518 + ], + "type": "text", + "content": "The choice of the operator " + }, + { + "bbox": [ + 304, + 469, + 545, + 518 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 304, + 469, + 545, + 518 + ], + "type": "text", + "content": " is highly dependent on the physical model of the measurement process. We adopt a simplified version of the widely used three-point transport model [37] as" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 529, + 545, + 563 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 529, + 545, + 563 + ], + "spans": [ + { + "bbox": [ + 313, + 529, + 545, + 563 + ], + "type": "interline_equation", + "content": "\\tau (\\mathbf {x} _ {l}, \\mathbf {x} _ {d}, t) = \\int_ {S} T (\\mathbf {x} _ {l}, \\mathbf {x} _ {d}, t, \\mathbf {s}) C (\\mathbf {x} _ {l}, \\mathbf {x} _ {d}, \\mathbf {s}) f (\\mathbf {s}) d A (\\mathbf {s}) \\tag {4}", + "image_path": "98b66ed0efc7a7e116a2bf3323acd98206e4c32634491a041c7dddd03efdef3c.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 563, + 545, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 563, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 304, + 563, + 545, + 635 + ], + "type": "text", + "content": "in which " + }, + { + "bbox": [ + 304, + 563, + 545, + 635 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 563, + 545, + 635 + ], + "type": "text", + "content": " is the photon intensity detected with the virtual source " + }, + { + "bbox": [ + 304, + 563, + 545, + 635 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_l" + }, + { + "bbox": [ + 304, + 563, + 545, + 635 + ], + "type": "text", + "content": ", virtual detector " + }, + { + "bbox": [ + 304, + 563, + 545, + 635 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_d" + }, + { + "bbox": [ + 304, + 563, + 545, + 635 + ], + "type": "text", + "content": " and time " + }, + { + "bbox": [ + 304, + 563, + 545, + 635 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 563, + 545, + 635 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 563, + 545, + 635 + ], + "type": "inline_equation", + "content": "f(\\mathbf{s})" + }, + { + "bbox": [ + 304, + 563, + 545, + 635 + ], + "type": "text", + "content": " represents the albedo of the point " + }, + { + "bbox": [ + 304, + 563, + 545, + 635 + ], + "type": "inline_equation", + "content": "\\mathbf{s}" + }, + { + "bbox": [ + 304, + 563, + 545, + 635 + ], + "type": "text", + "content": " on the hidden object " + }, + { + "bbox": [ + 304, + 563, + 545, + 635 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 563, + 545, + 635 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 563, + 545, + 635 + ], + "type": "inline_equation", + "content": "A(\\mathbf{s})" + }, + { + "bbox": [ + 304, + 563, + 545, + 635 + ], + "type": "text", + "content": " is the corresponding area measure. " + }, + { + "bbox": [ + 304, + 563, + 545, + 635 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 563, + 545, + 635 + ], + "type": "text", + "content": " is a function that characterizes the arrival time of the photons, which is defined as" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 318, + 663, + 545, + 677 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 663, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 318, + 663, + 545, + 677 + ], + "type": "interline_equation", + "content": "T (\\mathbf {x} _ {l}, \\mathbf {x} _ {d}, t, \\mathbf {s}) = \\delta (| | \\mathbf {x} _ {l} - \\mathbf {s} | | _ {2} + | | \\mathbf {x} _ {d} - \\mathbf {s} | | _ {2} - c t) \\tag {5}", + "image_path": "612e7f6a3cfdfe42740b0b6eb6675981b69dc29aedbc2d0f7e7b665dfe1418ad.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "content": "in which " + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "content": " is the Dirac function and " + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "content": " is the light speed. Besides, the function " + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "content": " characterizes the intensity attenuation" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "17422" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 228, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 228, + 84 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 228, + 84 + ], + "type": "text", + "content": "in the physical model, which is expressed as" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 92, + 287, + 148 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 92, + 287, + 148 + ], + "spans": [ + { + "bbox": [ + 53, + 92, + 287, + 148 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} C \\left(\\mathbf {x} _ {l}, \\mathbf {x} _ {d}, \\mathbf {s}\\right) = \\frac {< \\mathbf {s} - \\mathbf {x} _ {l} , \\mathbf {n} (\\mathbf {x} _ {l}) > < \\mathbf {x} _ {l} - \\mathbf {s} , \\mathbf {n} (\\mathbf {s}) >}{\\left| \\left| \\mathbf {x} _ {l} - \\mathbf {s} \\right| \\right| _ {2} ^ {4}} (6) \\\\ \\times \\frac {< \\mathbf {s} - \\mathbf {x} _ {d} , \\mathbf {n} (\\mathbf {x} _ {d}) > < \\mathbf {x} _ {d} - \\mathbf {s} , \\mathbf {n} (\\mathbf {s}) >}{\\left| \\left| \\mathbf {x} _ {d} - \\mathbf {s} \\right| \\right| _ {2} ^ {\\frac {4}}}. (6) \\\\ \\end{array}", + "image_path": "c406b6d79f04138b1e0dce989e5e5fa12b81fbce2bbbd6a556c8e4429c180194.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 156, + 287, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 156, + 287, + 191 + ], + "spans": [ + { + "bbox": [ + 47, + 156, + 287, + 191 + ], + "type": "text", + "content": "in which " + }, + { + "bbox": [ + 47, + 156, + 287, + 191 + ], + "type": "inline_equation", + "content": "\\mathbf{n}(\\mathbf{s})" + }, + { + "bbox": [ + 47, + 156, + 287, + 191 + ], + "type": "text", + "content": " is the unit normal vector pointing toward the visible wall at point " + }, + { + "bbox": [ + 47, + 156, + 287, + 191 + ], + "type": "inline_equation", + "content": "\\mathbf{s}" + }, + { + "bbox": [ + 47, + 156, + 287, + 191 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 156, + 287, + 191 + ], + "type": "inline_equation", + "content": "<\\cdot, \\cdot>" + }, + { + "bbox": [ + 47, + 156, + 287, + 191 + ], + "type": "text", + "content": " is the inner product in " + }, + { + "bbox": [ + 47, + 156, + 287, + 191 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^3" + }, + { + "bbox": [ + 47, + 156, + 287, + 191 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 193, + 287, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 193, + 287, + 263 + ], + "spans": [ + { + "bbox": [ + 46, + 193, + 287, + 263 + ], + "type": "text", + "content": "Suppose that there are two pairs " + }, + { + "bbox": [ + 46, + 193, + 287, + 263 + ], + "type": "inline_equation", + "content": "(\\mathbf{x}_{l_1},\\mathbf{x}_{d_1}),(\\mathbf{x}_{l_2},\\mathbf{x}_{d_2})" + }, + { + "bbox": [ + 46, + 193, + 287, + 263 + ], + "type": "text", + "content": " on the visible wall. For a fixed point s on the hidden object, we assume that the contributions of this point to the signal at these two pairs are contained in time bin " + }, + { + "bbox": [ + 46, + 193, + 287, + 263 + ], + "type": "inline_equation", + "content": "t_1" + }, + { + "bbox": [ + 46, + 193, + 287, + 263 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 193, + 287, + 263 + ], + "type": "inline_equation", + "content": "t_2" + }, + { + "bbox": [ + 46, + 193, + 287, + 263 + ], + "type": "text", + "content": " respectively. Then from Eq. (5), we know that " + }, + { + "bbox": [ + 46, + 193, + 287, + 263 + ], + "type": "inline_equation", + "content": "t_1" + }, + { + "bbox": [ + 46, + 193, + 287, + 263 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 193, + 287, + 263 + ], + "type": "inline_equation", + "content": "t_2" + }, + { + "bbox": [ + 46, + 193, + 287, + 263 + ], + "type": "text", + "content": " are close if" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 93, + 266, + 287, + 279 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 266, + 287, + 279 + ], + "spans": [ + { + "bbox": [ + 93, + 266, + 287, + 279 + ], + "type": "interline_equation", + "content": "D = \\left\\| \\mathbf {x} _ {l _ {1}} - \\mathbf {x} _ {l _ {2}} \\right\\| _ {2} + \\left\\| \\mathbf {x} _ {d _ {1}} - \\mathbf {x} _ {d _ {2}} \\right\\| _ {2} \\tag {7}", + "image_path": "395a2ff6f40a36cee2b877afe412cd5f43463f9787dfab12498c54f40f58d75a.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 285, + 287, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 285, + 287, + 320 + ], + "spans": [ + { + "bbox": [ + 46, + 285, + 287, + 320 + ], + "type": "text", + "content": "is small. We can then regard the signal recovery problem as an interpolation problem, since the points have contributions to neighboring time bins." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 322, + 287, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 322, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 46, + 322, + 287, + 453 + ], + "type": "text", + "content": "However, when the spacing between adjacent detection points is large, conventional interpolation methods will fail due to the violation of the assumption. This phenomenon is very common in practical applications. To solve this problem, Nam et al. [30] employs two SPAD arrays to meet the assumption, but brings additional expense to the system. Another solution is brought up by Yang et al. [42], which combines a micro-mirror device with the compressed sensing strategy to increase detection efficiency. However, this method is sensitive to the geometry of the visible wall, which limits its application." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 453, + 287, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 453, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 46, + 453, + 287, + 536 + ], + "type": "text", + "content": "Instead of introducing new detection systems, we take advantage of the fast development of deep learning methods and adopt a neural network to learn the nonlinear operator " + }, + { + "bbox": [ + 46, + 453, + 287, + 536 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 46, + 453, + 287, + 536 + ], + "type": "text", + "content": ". Details about the network will be introduced in the next section. With the employed network, we can obtain a better approximation of the operator " + }, + { + "bbox": [ + 46, + 453, + 287, + 536 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 46, + 453, + 287, + 536 + ], + "type": "text", + "content": ", as well as a higher reconstruction quality." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 537, + 287, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 537, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 46, + 537, + 287, + 622 + ], + "type": "text", + "content": "The proposed pipeline is shown in Fig. 2. The low resolution signal " + }, + { + "bbox": [ + 46, + 537, + 287, + 622 + ], + "type": "inline_equation", + "content": "d_{l}" + }, + { + "bbox": [ + 46, + 537, + 287, + 622 + ], + "type": "text", + "content": " is first processed by the neural network " + }, + { + "bbox": [ + 46, + 537, + 287, + 622 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 46, + 537, + 287, + 622 + ], + "type": "text", + "content": " to recover the high resolution signal " + }, + { + "bbox": [ + 46, + 537, + 287, + 622 + ], + "type": "inline_equation", + "content": "\\Phi(d_{l})" + }, + { + "bbox": [ + 46, + 537, + 287, + 622 + ], + "type": "text", + "content": ". Then the reconstruction " + }, + { + "bbox": [ + 46, + 537, + 287, + 622 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 537, + 287, + 622 + ], + "type": "text", + "content": " is obtained using the existing imaging algorithm " + }, + { + "bbox": [ + 46, + 537, + 287, + 622 + ], + "type": "inline_equation", + "content": "F_{h}" + }, + { + "bbox": [ + 46, + 537, + 287, + 622 + ], + "type": "text", + "content": ". With the proposed pipeline, we can decrease the exposure time significantly while maintaining the reconstruction quality." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 633, + 224, + 646 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 633, + 224, + 646 + ], + "spans": [ + { + "bbox": [ + 47, + 633, + 224, + 646 + ], + "type": "text", + "content": "3. Signal Superresolution network" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "text", + "content": "Based on the analysis in the last section, we have formulated the signal recovery problem as an interpolation problem and have noticed that conventional interpolation methods may fail when the spacing between adjacent detection pairs gets larger. Besides, the contribution of one point on" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 323, + 72, + 507, + 249 + ], + "blocks": [ + { + "bbox": [ + 323, + 72, + 507, + 249 + ], + "lines": [ + { + "bbox": [ + 323, + 72, + 507, + 249 + ], + "spans": [ + { + "bbox": [ + 323, + 72, + 507, + 249 + ], + "type": "image", + "image_path": "515b81d3bb34b1b265dd6414e6a21e2e620637f7f4106cd0fafbab1039c5d4a3.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 261, + 545, + 305 + ], + "lines": [ + { + "bbox": [ + 305, + 261, + 545, + 305 + ], + "spans": [ + { + "bbox": [ + 305, + 261, + 545, + 305 + ], + "type": "text", + "content": "Figure 4. Comparisons of the mean SNR values of the results obtained with the nearest neighbor method and the proposed network on testing sets. The signals recovered by the proposed network have much higher SNR on both testing sets." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 327, + 545, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 327, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 305, + 327, + 545, + 373 + ], + "type": "text", + "content": "the hidden object is contained in adjacent time bins for adjacent detection pairs. It means that the problem can not be solved by simply applying 2-D image superresolution to each time slice." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 375, + 545, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 375, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 305, + 375, + 545, + 422 + ], + "type": "text", + "content": "In this section, we introduce the neural network which learns the nonlinear signal recovery operator " + }, + { + "bbox": [ + 305, + 375, + 545, + 422 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 305, + 375, + 545, + 422 + ], + "type": "text", + "content": " and the training dataset. Details about the training process and the ablation study are provided in the supplement." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 431, + 429, + 442 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 431, + 429, + 442 + ], + "spans": [ + { + "bbox": [ + 306, + 431, + 429, + 442 + ], + "type": "text", + "content": "3.1. Network architecture" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 449, + 545, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 449, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 305, + 449, + 545, + 498 + ], + "type": "text", + "content": "Recently, attention mechanism [4, 36] has become increasingly popular in the deep learning society. However, it has been shown that there is no need to apply the attention mechanism to every part of the network [6]." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 498, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 605 + ], + "type": "text", + "content": "To learn the operator, we generalize the attention in attention net " + }, + { + "bbox": [ + 304, + 498, + 545, + 605 + ], + "type": "inline_equation", + "content": "(\\mathrm{A}^2\\mathrm{N})" + }, + { + "bbox": [ + 304, + 498, + 545, + 605 + ], + "type": "text", + "content": " [6] which has two branches and dropout modules for the attention mechanism. One of the branches uses the conventional method to provide a rough estimation of the high resolution signal, and the other branch employs the network. The final output is the sum of the results from these two branches. The network architecture with two branches has been widely studied in the field of image superresolution [2, 19, 43]." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 606, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 606, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 606, + 545, + 713 + ], + "type": "text", + "content": "The proposed signal superresolution network (SSN) is illustrated in Fig. 3. The learning branch is composed of several convolution layers and 3-D attention in attention blocks " + }, + { + "bbox": [ + 304, + 606, + 545, + 713 + ], + "type": "inline_equation", + "content": "(3\\mathrm{D - A}^2\\mathrm{B})" + }, + { + "bbox": [ + 304, + 606, + 545, + 713 + ], + "type": "text", + "content": ". Each block has a dropout module to learn the weight between the attention branch and the non-attention branch, which can determine whether the attention mechanism is needed in the block. The reconstruction module contains several convolution layers to reconstruct the high resolution signal through the extracted feature. The low di" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17423" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 132 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 132 + ], + "type": "text", + "content": "mensional feature is first interpolated to a high dimensional feature using trilinear interpolation, and then the feature is processed by several convolution layers. The output of the convolution layer is passed to the next part of the reconstruction module and the process is repeated." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 132, + 288, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 132, + 288, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 132, + 288, + 251 + ], + "type": "text", + "content": "The generalization of the network from " + }, + { + "bbox": [ + 46, + 132, + 288, + 251 + ], + "type": "inline_equation", + "content": "\\mathrm{A}^2\\mathrm{N}" + }, + { + "bbox": [ + 46, + 132, + 288, + 251 + ], + "type": "text", + "content": " [6] to SSN is not trivial. Based on our analysis, the contribution of one point on the hidden object is contained in adjacent time bins for adjacent detection pairs. Conventional interpolation methods will perform poorly when the spacing between adjacent detection pairs gets larger. Thus, 3-D kernels are needed to extract and recover the information, and the size of the kernels is relevant to the spacing. In addition, the signal is only interpolated in spatial dimensions instead of all three dimensions to make the operator easier to learn." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 258, + 105, + 269 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 258, + 105, + 269 + ], + "spans": [ + { + "bbox": [ + 47, + 258, + 105, + 269 + ], + "type": "text", + "content": "3.2. Dataset" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 277, + 287, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 277, + 287, + 467 + ], + "spans": [ + { + "bbox": [ + 46, + 277, + 287, + 467 + ], + "type": "text", + "content": "For all experiments in this paper, we use subsets of Fashion-MNIST [40] and ShapeNet [5] to generate synthetic signals for training and testing. The Fashion-MNIST dataset consists of 2-D figures, while the ShapeNet comprises 3-D objects. We choose 5,000 samples from the training set of Fashion-MNIST and the class \"Car\" of ShapeNet respectively to generate our training set. In addition, we choose 1,000 samples from the testing set of Fashion-MNIST and the class \"Car\" of ShapeNet respectively to generate the testing set. Thus, the training set used in this work comprises 10,000 signal pairs, and the testing set comprises 2,000. All the signals are generated using Eq. (4). Besides, we set " + }, + { + "bbox": [ + 46, + 277, + 287, + 467 + ], + "type": "inline_equation", + "content": "N_{x} = N_{y} = 8" + }, + { + "bbox": [ + 46, + 277, + 287, + 467 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 277, + 287, + 467 + ], + "type": "inline_equation", + "content": "N_{t} = 100" + }, + { + "bbox": [ + 46, + 277, + 287, + 467 + ], + "type": "text", + "content": " in this paper. For measured signals whose " + }, + { + "bbox": [ + 46, + 277, + 287, + 467 + ], + "type": "inline_equation", + "content": "N_{t}" + }, + { + "bbox": [ + 46, + 277, + 287, + 467 + ], + "type": "text", + "content": " is larger than 100, the network can also be applied directly since the network operates locally." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 468, + 287, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 468, + 287, + 563 + ], + "spans": [ + { + "bbox": [ + 46, + 468, + 287, + 563 + ], + "type": "text", + "content": "For the confocal scenario, to match the experimental setting with the Stanford dataset [22], the 2-D figures from FashionMNIST [40] are placed " + }, + { + "bbox": [ + 46, + 468, + 287, + 563 + ], + "type": "inline_equation", + "content": "1\\mathrm{m}" + }, + { + "bbox": [ + 46, + 468, + 287, + 563 + ], + "type": "text", + "content": " away from the visible wall, the size of the figures is set as " + }, + { + "bbox": [ + 46, + 468, + 287, + 563 + ], + "type": "inline_equation", + "content": "1\\times 1\\mathrm{m}^2" + }, + { + "bbox": [ + 46, + 468, + 287, + 563 + ], + "type": "text", + "content": ". And the 3-D objects from ShapeNet [5] are placed in a " + }, + { + "bbox": [ + 46, + 468, + 287, + 563 + ], + "type": "inline_equation", + "content": "1\\times 1\\times 0.2\\mathrm{m}^3" + }, + { + "bbox": [ + 46, + 468, + 287, + 563 + ], + "type": "text", + "content": " (horizontal, vertical, depth) box with a maximum depth of " + }, + { + "bbox": [ + 46, + 468, + 287, + 563 + ], + "type": "inline_equation", + "content": "1.1\\mathrm{m}" + }, + { + "bbox": [ + 46, + 468, + 287, + 563 + ], + "type": "text", + "content": ". The illumination region on the visible wall is " + }, + { + "bbox": [ + 46, + 468, + 287, + 563 + ], + "type": "inline_equation", + "content": "2\\times 2\\mathrm{m}^2" + }, + { + "bbox": [ + 46, + 468, + 287, + 563 + ], + "type": "text", + "content": ", and the time resolution of the SPAD is " + }, + { + "bbox": [ + 46, + 468, + 287, + 563 + ], + "type": "inline_equation", + "content": "32~\\mathrm{ps}" + }, + { + "bbox": [ + 46, + 468, + 287, + 563 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 563, + 287, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 563, + 287, + 659 + ], + "spans": [ + { + "bbox": [ + 46, + 563, + 287, + 659 + ], + "type": "text", + "content": "For the non-confocal setting, we match the experimental setting with the data provided by phasor field [25]. The 2-D figures from Fashion-MNIST are placed " + }, + { + "bbox": [ + 46, + 563, + 287, + 659 + ], + "type": "inline_equation", + "content": "1\\mathrm{m}" + }, + { + "bbox": [ + 46, + 563, + 287, + 659 + ], + "type": "text", + "content": " away from the wall, and the size is " + }, + { + "bbox": [ + 46, + 563, + 287, + 659 + ], + "type": "inline_equation", + "content": "1.25\\times 1.25\\mathrm{m}^2" + }, + { + "bbox": [ + 46, + 563, + 287, + 659 + ], + "type": "text", + "content": ". The 3-D objects from ShapeNet are placed in a " + }, + { + "bbox": [ + 46, + 563, + 287, + 659 + ], + "type": "inline_equation", + "content": "1.25\\times 1.25\\times 0.2\\mathrm{m}^3" + }, + { + "bbox": [ + 46, + 563, + 287, + 659 + ], + "type": "text", + "content": " (horizontal, vertical, depth) box with a maximum depth of " + }, + { + "bbox": [ + 46, + 563, + 287, + 659 + ], + "type": "inline_equation", + "content": "1.1\\mathrm{m}" + }, + { + "bbox": [ + 46, + 563, + 287, + 659 + ], + "type": "text", + "content": ". The illumination region on the visible wall is " + }, + { + "bbox": [ + 46, + 563, + 287, + 659 + ], + "type": "inline_equation", + "content": "1.25\\times 1.25\\mathrm{m}^2" + }, + { + "bbox": [ + 46, + 563, + 287, + 659 + ], + "type": "text", + "content": ", and the time resolution of the SPAD is 16 ps." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 670, + 99, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 99, + 681 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 99, + 681 + ], + "type": "text", + "content": "4. Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "In this section, we show the reconstruction results of the proposed pipeline on both synthetic and measured data. For" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 305, + 68, + 545, + 616 + ], + "blocks": [ + { + "bbox": [ + 305, + 68, + 545, + 616 + ], + "lines": [ + { + "bbox": [ + 305, + 68, + 545, + 616 + ], + "spans": [ + { + "bbox": [ + 305, + 68, + 545, + 616 + ], + "type": "image", + "image_path": "c59a5991ae980e8fbc1943f5cb87dc6d2835595d0868a6aea70a94cacea9a769.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 624, + 545, + 689 + ], + "lines": [ + { + "bbox": [ + 304, + 624, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 624, + 545, + 689 + ], + "type": "text", + "content": "Figure 5. Recovered signals and reconstruction results of the pyramid. (a) Ground truth of the hidden object. (b) A comparison of the recovered signals (the first arrival time of the original signal is marked by the blue curve in the zoom-in window of each subfigure). (c) Reconstruction results of state-of-the-art methods with different signals." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "17424" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": "each instance, reconstruction results obtained with signals of different spatial resolutions are shown to validate the effectiveness of the proposed pipeline. The first column (Original) shows reconstructions obtained from the original signal with a spatial resolution of " + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": "; results in the second column (Low) are reconstructed using the signal scanned at " + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "inline_equation", + "content": "8 \\times 8" + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": " points; the third column (Nearest) shows reconstructions of the recovered signal obtained with nearest neighbor interpolation; the last column (Ours) illustrates reconstruction results obtained with the proposed pipeline. Comparisons with other interpolation methods and existing learning based NLOS imaging method are provided in the supplement." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 228, + 287, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 228, + 287, + 324 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 287, + 324 + ], + "type": "text", + "content": "To illustrate the proposed pipeline can be acted as a plugin-and-play module for most algorithms, we combine state-of-the-art methods with our pipeline. For confocal experiments, the results are reconstructed by F-K [22], LCT [31], LOG-BP [18], D-LCT [45] and SOCR [27] algorithms, while for non-confocal experiments, results are given by LOG-BP, Phasor [24] and SOCR methods. For each instance, the parameters are fixed in each algorithm." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 325, + 287, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 325, + 287, + 408 + ], + "spans": [ + { + "bbox": [ + 46, + 325, + 287, + 408 + ], + "type": "text", + "content": "To assess the results quantitatively, we compute the signal to noise ratio (SNR) of the recovered signals, as well as the peak signal to noise ratio (PSNR) and structural similarity index measure (SSIM) of the reconstructed maximum intensity projections. For measured data, we use the reconstructions obtained from the original signals as references to compute PSNR and SSIM." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 418, + 213, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 418, + 213, + 430 + ], + "spans": [ + { + "bbox": [ + 47, + 418, + 213, + 430 + ], + "type": "text", + "content": "4.1. Results of confocal experiment" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 437, + 287, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 287, + 510 + ], + "type": "text", + "content": "In Fig. 4, we compare the mean SNR values of the signals interpolated with SSN and the nearest neighbor method. It is shown that the network provides much better results on both datasets. In other words, the network approximates the oracle interpolation operator " + }, + { + "bbox": [ + 46, + 437, + 287, + 510 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 46, + 437, + 287, + 510 + ], + "type": "text", + "content": " with high accuracy and serves as a crucial part of the proposed pipeline." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 510, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 287, + 605 + ], + "type": "text", + "content": "To test the robustness of the proposed pipeline, we use the instance of the pyramid from SOCR [27], whose signal was generated with a different physical model. As shown in Fig. 5b, the first arrival time is correctly recovered by the proposed network, while the nearest neighbor interpolation method fails. Fig. 5c shows that the proposed pipeline provides a signal for high quality reconstructions with better visual quality and higher PSNR and SSIM values." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 605, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 713 + ], + "type": "text", + "content": "The proposed pipeline is then tested on measured data from Stanford dataset [22]. This dataset contains signals measured with different hidden objects and exposure time. We choose the instance of the statue to validate the effectiveness of the proposed pipeline. The measurement time of each sampling point is " + }, + { + "bbox": [ + 46, + 605, + 288, + 713 + ], + "type": "inline_equation", + "content": "0.0069\\mathrm{s}" + }, + { + "bbox": [ + 46, + 605, + 288, + 713 + ], + "type": "text", + "content": ". The total exposure time of the signal scanned at " + }, + { + "bbox": [ + 46, + 605, + 288, + 713 + ], + "type": "inline_equation", + "content": "32\\times 32" + }, + { + "bbox": [ + 46, + 605, + 288, + 713 + ], + "type": "text", + "content": " points is " + }, + { + "bbox": [ + 46, + 605, + 288, + 713 + ], + "type": "inline_equation", + "content": "7.03\\mathrm{s}" + }, + { + "bbox": [ + 46, + 605, + 288, + 713 + ], + "type": "text", + "content": ", and it only takes " + }, + { + "bbox": [ + 46, + 605, + 288, + 713 + ], + "type": "inline_equation", + "content": "0.44\\mathrm{s}" + }, + { + "bbox": [ + 46, + 605, + 288, + 713 + ], + "type": "text", + "content": " to scan " + }, + { + "bbox": [ + 46, + 605, + 288, + 713 + ], + "type": "inline_equation", + "content": "8\\times 8" + }, + { + "bbox": [ + 46, + 605, + 288, + 713 + ], + "type": "text", + "content": " points. As is shown in Fig. 6b, the result of SSN has higher SNR. The first arrival time can" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 304, + 68, + 545, + 616 + ], + "blocks": [ + { + "bbox": [ + 304, + 68, + 545, + 616 + ], + "lines": [ + { + "bbox": [ + 304, + 68, + 545, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 68, + 545, + 616 + ], + "type": "image", + "image_path": "b052fceca843684b67e7d0eb7163a0c3e5a5e1ab70230ab06882186645278365.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 624, + 547, + 689 + ], + "lines": [ + { + "bbox": [ + 304, + 624, + 547, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 624, + 547, + 689 + ], + "type": "text", + "content": "Figure 6. Recovered signals and reconstruction results of the statue. (a) Ground truth of the hidden object. (b) A comparison of recovered signals (the first arrival time of the original signal is marked by the blue curve in the zoom-in window of each subfigure). (c) Reconstruction results of state-of-the-art methods with different signals." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17425" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": "hardly be distinguished in the nearest neighbor interpolated signal, while it is much clearer in the signal recovered by the proposed network. As shown in Fig. 6c, the proposed pipeline results in fewer artifacts in the background. More comparisons of different exposure times are provided in the supplement, which also indicates the outperformance of the proposed pipeline." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 164, + 204, + 177 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 164, + 204, + 177 + ], + "spans": [ + { + "bbox": [ + 47, + 164, + 204, + 177 + ], + "type": "text", + "content": "4.2. Results of the dynamic scene" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 183, + 287, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 183, + 287, + 302 + ], + "spans": [ + { + "bbox": [ + 46, + 183, + 287, + 302 + ], + "type": "text", + "content": "To illustrate the potential application of the proposed pipeline on real-time NLOS imaging, we test it on the captured dynamic scenes from the Stanford dataset. The data is captured over a " + }, + { + "bbox": [ + 46, + 183, + 287, + 302 + ], + "type": "inline_equation", + "content": "2 \\times 2\\mathrm{m}^2" + }, + { + "bbox": [ + 46, + 183, + 287, + 302 + ], + "type": "text", + "content": " region on the visible wall, and the person is wearing a retroreflective suit to increase the detection of the indirect signal. It takes " + }, + { + "bbox": [ + 46, + 183, + 287, + 302 + ], + "type": "inline_equation", + "content": "0.25\\mathrm{s}" + }, + { + "bbox": [ + 46, + 183, + 287, + 302 + ], + "type": "text", + "content": " to scan " + }, + { + "bbox": [ + 46, + 183, + 287, + 302 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 46, + 183, + 287, + 302 + ], + "type": "text", + "content": " points, which allows for a reconstruction video with 4 frames per second. The reflection model of the retroreflective material [31] is different from the one defined in Eq. (6)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 303, + 288, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 303, + 288, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 303, + 288, + 456 + ], + "type": "text", + "content": "The results of several frames are shown in Fig. 7b, the complete video of the reconstruction results can be found in the supplement. All the results are reconstructed by the F-K method. Although the training set is generated by a different forward model, the proposed pipeline can still recover signals for high quality reconstructions. The pose of the person can be clearly identified in the reconstructions using the proposed pipeline, while the reconstructions using the nearest neighbor interpolated signal contain more artifacts. The mean PSNR and SSIM of the reconstruction results obtained from the proposed pipeline are 24.46 and 0.63, while those obtained with nearest neighbor interpolation are only 18.68 and 0.37." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 459, + 287, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 459, + 287, + 554 + ], + "spans": [ + { + "bbox": [ + 46, + 459, + 287, + 554 + ], + "type": "text", + "content": "The total capture time of each frame is reduced from 0.25 s to 0.0156 s. This enables us to obtain a high quality reconstruction video with 64 frames per second. Besides, it only takes 0.08 s to recover the high resolution signal with the proposed network, and 0.01s to reconstruct the hidden object with the F-K method on an NVIDIA RTX 3090. Thus, real-time video imaging with a high frame rate can be realized with a delay of a few frames." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 563, + 234, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 563, + 234, + 575 + ], + "spans": [ + { + "bbox": [ + 47, + 563, + 234, + 575 + ], + "type": "text", + "content": "4.3. Results of non-confocal experiment" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 582, + 287, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 664 + ], + "type": "text", + "content": "For the non-confocal setting, we use the instance of \"4\" provided by Phasor [24] to validate the effectiveness of the proposed method. The measured signal is subsampled to " + }, + { + "bbox": [ + 46, + 582, + 287, + 664 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 46, + 582, + 287, + 664 + ], + "type": "text", + "content": " and the illumination region on the visible wall is cropped to " + }, + { + "bbox": [ + 46, + 582, + 287, + 664 + ], + "type": "inline_equation", + "content": "1.25 \\times 1.25\\mathrm{m}^2" + }, + { + "bbox": [ + 46, + 582, + 287, + 664 + ], + "type": "text", + "content": ". We use the signal scanned at 8 " + }, + { + "bbox": [ + 46, + 582, + 287, + 664 + ], + "type": "inline_equation", + "content": "\\times 8" + }, + { + "bbox": [ + 46, + 582, + 287, + 664 + ], + "type": "text", + "content": " points as the input of our pipeline. It takes 1 s to scan each point on the visible wall." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 714 + ], + "type": "text", + "content": "As is shown in Fig. 8b, the SNR of the signal recovered by the network is much higher than the one interpolated with the nearest neighbor method. Reconstruction results are shown in Fig. 8c. The proposed pipeline provides re" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 378, + 70, + 474, + 137 + ], + "blocks": [ + { + "bbox": [ + 378, + 70, + 474, + 137 + ], + "lines": [ + { + "bbox": [ + 378, + 70, + 474, + 137 + ], + "spans": [ + { + "bbox": [ + 378, + 70, + 474, + 137 + ], + "type": "image", + "image_path": "4cf3f4389b401b9c7c3cf6ed7462d1ca6b347a6759fec99723260566fe74fdd9.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 414, + 138, + 438, + 148 + ], + "lines": [ + { + "bbox": [ + 414, + 138, + 438, + 148 + ], + "spans": [ + { + "bbox": [ + 414, + 138, + 438, + 148 + ], + "type": "text", + "content": "(a) GT" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 310, + 152, + 545, + 415 + ], + "blocks": [ + { + "bbox": [ + 310, + 152, + 545, + 415 + ], + "lines": [ + { + "bbox": [ + 310, + 152, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 310, + 152, + 545, + 415 + ], + "type": "image", + "image_path": "1ce59b218de5a7c16fda97620e7f7cf32a22da1c216a90d19c35cdeaf93deb5a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 384, + 417, + 468, + 426 + ], + "lines": [ + { + "bbox": [ + 384, + 417, + 468, + 426 + ], + "spans": [ + { + "bbox": [ + 384, + 417, + 468, + 426 + ], + "type": "text", + "content": "(b) Reconstruction results" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 436, + 547, + 514 + ], + "lines": [ + { + "bbox": [ + 304, + 436, + 547, + 514 + ], + "spans": [ + { + "bbox": [ + 304, + 436, + 547, + 514 + ], + "type": "text", + "content": "Figure 7. Reconstruction results of the dynamic scene. (a) An illustration of the dynamic scene. (b) Reconstruction results of the dynamic scene. The results are obtained with the F-K method. The PSNR and SSIM values are computed with the reconstructions obtained with the original signals as references. The proposed pipeline provides much better reconstructions in terms of both quantitative criteria and visual quality." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 535, + 534, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 535, + 534, + 548 + ], + "spans": [ + { + "bbox": [ + 306, + 535, + 534, + 548 + ], + "type": "text", + "content": "constructions with higher quantitative and visual quality." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 559, + 375, + 571 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 559, + 375, + 571 + ], + "spans": [ + { + "bbox": [ + 306, + 559, + 375, + 571 + ], + "type": "text", + "content": "5. Discussion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 578, + 545, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 578, + 545, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 578, + 545, + 639 + ], + "type": "text", + "content": "We have introduced a novel pipeline for high quality NLOS reconstruction with the signal measured at only a small number of scanning points. In this section, we discuss several factors that affect the reconstruction quality and some feasible solutions to improve the pipeline further." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 647, + 440, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 647, + 440, + 659 + ], + "spans": [ + { + "bbox": [ + 306, + 647, + 440, + 659 + ], + "type": "text", + "content": "5.1. Training of the network" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 665, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 546, + 713 + ], + "type": "text", + "content": "The training of the network employed in the proposed pipeline is supervised. Thus, it is affected by the generation of the training set, including the forward model, experimental setting, and hidden objects. For example, hidden objects" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17426" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 108 + ], + "type": "text", + "content": "with different materials will lead to various physical models. The networks trained and tested on signals with different models would bring additional errors to the results." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 108, + 287, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 108, + 287, + 178 + ], + "spans": [ + { + "bbox": [ + 46, + 108, + 287, + 178 + ], + "type": "text", + "content": "However, a network with good generalization ability can avoid degradation. As shown in Fig. 7, although the physical model of the dynamic scenes differs from the one used in the training set, the reconstructed objects are only slightly affected. Thus, the proposed pipeline has strong robustness to the forward model." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 186, + 239, + 198 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 186, + 239, + 198 + ], + "spans": [ + { + "bbox": [ + 47, + 186, + 239, + 198 + ], + "type": "text", + "content": "5.2. Performance of different algorithms" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 205, + 287, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 205, + 287, + 324 + ], + "spans": [ + { + "bbox": [ + 46, + 205, + 287, + 324 + ], + "type": "text", + "content": "As the reconstruction results shown in previous sections, the improvement of reconstruction quality of the proposed pipeline varies between different methods. Compared to the reconstruction results obtained from the low resolution signal, fast imaging algorithms like F-K, LCT and D-LCT have larger improvements, while the iterative inversion algorithm SOCR might have similar visual quality in some cases. This phenomenon may come from the fact that the iterative algorithm can make full use of the information contained in the signal, but with a longer reconstruction time." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 331, + 287, + 354 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 331, + 287, + 354 + ], + "spans": [ + { + "bbox": [ + 47, + 331, + 287, + 354 + ], + "type": "text", + "content": "5.3. Combination with SPAD array for further acceleration" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 361, + 287, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 361, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 46, + 361, + 287, + 491 + ], + "type": "text", + "content": "Since the proposed pipeline can recover the signal faithfully under the non-confocal setting, combining it with a SPAD array can accelerate the measurement process further. However, the existing SPAD array can only focus on a rectangular area on the visible wall, which means that the network has to learn an extrapolation operator. This is much harder than learning an interpolation operator. Thus, it would be more practical if the SPAD array could focus on sparse rectangular points on a relatively large area of the visible wall. This may be seen as a new area for hardware design." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 502, + 119, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 502, + 119, + 514 + ], + "spans": [ + { + "bbox": [ + 47, + 502, + 119, + 514 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 522, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 287, + 713 + ], + "type": "text", + "content": "In this paper, we have studied the signal recovery problem in NLOS imaging, and proposed a two-step learning based pipeline for fast NLOS imaging with only a small number of scanning points. In the first step, the signal of high spatial resolution is recovered with a neural network. In the second step, the hidden objects are obtained from the virtually measured signals at a dense grid. Our pipeline provides high-quality reconstructions under both confocal and non-confocal settings, which takes only " + }, + { + "bbox": [ + 46, + 522, + 287, + 713 + ], + "type": "inline_equation", + "content": "6.25\\%" + }, + { + "bbox": [ + 46, + 522, + 287, + 713 + ], + "type": "text", + "content": " of the original measurement time. Besides, the proposed pipeline is compatible with most existing detection systems and imaging algorithms. With the proposed pipeline, the high frame rate imaging video is possible, which is a noteworthy advance in this area. It is expected to combine emerging techniques in the fields of SPAD array for real-time and high quality NLOS reconstructions with higher frame rates." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 404, + 72, + 451, + 134 + ], + "blocks": [ + { + "bbox": [ + 404, + 72, + 451, + 134 + ], + "lines": [ + { + "bbox": [ + 404, + 72, + 451, + 134 + ], + "spans": [ + { + "bbox": [ + 404, + 72, + 451, + 134 + ], + "type": "image", + "image_path": "6d02ab8bd732e183ac0ccfaf3814939689e1178c80bfa33a1c3a7f6b82a5e7bc.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 308, + 148, + 545, + 240 + ], + "blocks": [ + { + "bbox": [ + 414, + 137, + 438, + 146 + ], + "lines": [ + { + "bbox": [ + 414, + 137, + 438, + 146 + ], + "spans": [ + { + "bbox": [ + 414, + 137, + 438, + 146 + ], + "type": "text", + "content": "(a) GT" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 308, + 148, + 545, + 240 + ], + "lines": [ + { + "bbox": [ + 308, + 148, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 308, + 148, + 545, + 240 + ], + "type": "image", + "image_path": "98394bb3ec72c783acb60f1f019e86e01f3c3ac1300f249bb043f42b8cecddc1.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 310, + 254, + 543, + 476 + ], + "blocks": [ + { + "bbox": [ + 391, + 243, + 461, + 252 + ], + "lines": [ + { + "bbox": [ + 391, + 243, + 461, + 252 + ], + "spans": [ + { + "bbox": [ + 391, + 243, + 461, + 252 + ], + "type": "text", + "content": "(b) Recovered signals" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 310, + 254, + 543, + 476 + ], + "lines": [ + { + "bbox": [ + 310, + 254, + 543, + 476 + ], + "spans": [ + { + "bbox": [ + 310, + 254, + 543, + 476 + ], + "type": "image", + "image_path": "b80a4457d701df0888351b443cdbef191820a0492924a2217592583be92c4abe.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 384, + 483, + 468, + 492 + ], + "lines": [ + { + "bbox": [ + 384, + 483, + 468, + 492 + ], + "spans": [ + { + "bbox": [ + 384, + 483, + 468, + 492 + ], + "type": "text", + "content": "(c) Reconstruction results" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 502, + 545, + 568 + ], + "lines": [ + { + "bbox": [ + 305, + 502, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 305, + 502, + 545, + 568 + ], + "type": "text", + "content": "Figure 8. Recovered signals and reconstruction results of the instance \"4\". (a) Ground truth of the hidden object. (b) A comparison of recovered signals (the first arrival time of the original signal is marked by the blue curve in the zoom-in window of each subfigure). (c) Reconstruction results of state-of-the-art methods with different signals." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 594, + 545, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 594, + 545, + 641 + ], + "spans": [ + { + "bbox": [ + 305, + 594, + 545, + 641 + ], + "type": "text", + "content": "Acknowledgments. The authors would like to thank Zhu-peng Li and Rongqian Wang for fruitful discussions. This work was supported by the National Natural Science Foundation of China (61975087, 12071244, 11971258)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 658, + 364, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 658, + 364, + 670 + ], + "spans": [ + { + "bbox": [ + 306, + 658, + 364, + 670 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 310, + 680, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 680, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 310, + 680, + 545, + 713 + ], + "type": "text", + "content": "[1] Byeongjoo Ahn, Akshat Dave, Ashok Veeraraghavan, Ioannis Gkioulekas, and Aswin C Sankaranarayanan. Convolutional approximations to the general non-line-of-sight imag" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17427" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 288, + 715 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 66, + 73, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 73, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 73, + 287, + 95 + ], + "type": "text", + "content": "ing operator. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7889-7899, 2019. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 96, + 288, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 96, + 288, + 129 + ], + "spans": [ + { + "bbox": [ + 53, + 96, + 288, + 129 + ], + "type": "text", + "content": "[2] Saeed Anwar, Salman Khan, and Nick Barnes. A deep journey into super-resolution: A survey. ACM Computing Surveys (CSUR), 53(3):1-34, 2020. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 130, + 288, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 130, + 288, + 174 + ], + "spans": [ + { + "bbox": [ + 53, + 130, + 288, + 174 + ], + "type": "text", + "content": "[3] Victor Arellano, Diego Gutierrez, and Adrian Jarabo. Fast back-projection for non-line of sight reconstruction. In ACM SIGGRAPH 2017 Posters, SIGGRAPH '17, New York, NY, USA, 2017. Association for Computing Machinery. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 175, + 288, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 175, + 288, + 218 + ], + "spans": [ + { + "bbox": [ + 53, + 175, + 288, + 218 + ], + "type": "text", + "content": "[4] Syed Muhammad Arsalan Bashir, Yi Wang, Mahrukh Khan, and Yilong Niu. A comprehensive review of deep learning-based single image super-resolution. PeerJ Computer Science, 7:e621, 2021. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 220, + 288, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 220, + 288, + 297 + ], + "spans": [ + { + "bbox": [ + 53, + 220, + 288, + 297 + ], + "type": "text", + "content": "[5] Angel X. Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. ShapeNet: An Information-Rich 3D Model Repository. Technical Report arXiv:1512.03012 [cs.GR], Stanford University — Princeton University — Toyota Technological Institute at Chicago, 2015. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 298, + 288, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 298, + 288, + 330 + ], + "spans": [ + { + "bbox": [ + 53, + 298, + 288, + 330 + ], + "type": "text", + "content": "[6] Haoyu Chen, Jinjin Gu, and Zhi Zhang. Attention in attention network for image super-resolution. arXiv preprint arXiv:2104.09497, 2021. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 331, + 288, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 331, + 288, + 374 + ], + "spans": [ + { + "bbox": [ + 53, + 331, + 288, + 374 + ], + "type": "text", + "content": "[7] Wenzheng Chen, Simon Daneau, Fahim Mannan, and Felix Heide. Steady-state non-line-of-sight imaging. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2019. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 376, + 288, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 376, + 288, + 421 + ], + "spans": [ + { + "bbox": [ + 53, + 376, + 288, + 421 + ], + "type": "text", + "content": "[8] Wenzheng Chen, Fangyin Wei, Kiriakos N Kutulakos, Szymon Rusinkiewicz, and Felix Heide. Learned feature embeddings for non-line-of-sight imaging and recognition. ACM Transactions on Graphics (TOG), 39(6):1-18, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 422, + 288, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 422, + 288, + 453 + ], + "spans": [ + { + "bbox": [ + 53, + 422, + 288, + 453 + ], + "type": "text", + "content": "[9] Justin Dove and Jeffrey H. Shapiro. Paraxial theory of phasor-field imaging. Opt. Express, 27(13):18016-18037, Jun 2019. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 455, + 287, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 455, + 287, + 487 + ], + "spans": [ + { + "bbox": [ + 48, + 455, + 287, + 487 + ], + "type": "text", + "content": "[10] Xiaohua Feng and Liang Gao. Improving non-line-of-sight image reconstruction with weighting factors. Optics letters, 45(14):3921-3924, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 489, + 287, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 489, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 48, + 489, + 287, + 521 + ], + "type": "text", + "content": "[11] Xiaohua Feng and Liang Gao. Ultrafast light field tomography for snapshot transient and non-line-of-sight imaging. Nature Communications, 12:2179, 04 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 523, + 287, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 523, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 523, + 287, + 567 + ], + "type": "text", + "content": "[12] Javier Grau Chopite, Matthias B. Hullin, Michael Wand, and Julian Iseringhausen. Deep non-line-of-sight reconstruction. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 957-966, 2020. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 568, + 287, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 287, + 612 + ], + "type": "text", + "content": "[13] Felix Heide, Matthew O'Toole, Kai Zang, David B. Lindell, Steven Diamond, and Gordon Wetzstein. Non-line-of-sight imaging with partial occluders and surface normals. ACM Trans. Graph., 38(3), May 2019. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 613, + 287, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 645 + ], + "type": "text", + "content": "[14] Julian Iseringhausen and Matthias B Hullin. Non-line-of-sight reconstruction using efficient transient rendering. ACM Transactions on Graphics (TOG), 39(1):1-14, 2020. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 647, + 287, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 690 + ], + "type": "text", + "content": "[15] Mariko Isogawa, Dorian Chan, Ye Yuan, Kris M. Kitani, and Matthew O'Toole. Efficient non-line-of-sight imaging from transient sinograms. In 16th European Conference on Computer Vision (ECCV), 2020. 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 692, + 287, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 692, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 48, + 692, + 287, + 715 + ], + "type": "text", + "content": "[16] Mariko Isogawa, Ye Yuan, Matthew O'Toole, and Kris Kitani. Optical non-line-of-sight physics-based 3d human pose" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 105 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 105 + ], + "type": "text", + "content": "estimation. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7011-7020, 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 107, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 107, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 308, + 107, + 545, + 150 + ], + "type": "text", + "content": "[17] Ahmed Kirmani, Tyler Hutchison, James Davis, and Ramesh Raskar. Looking around the corner using transient imaging. In 2009 IEEE 12th International Conference on Computer Vision, pages 159–166, 2009. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "type": "text", + "content": "[18] Martin Laurenzis and Andreas Velten. Feature selection and back-projection algorithms for nonlinear-of-sight laser-gated viewing. Journal of Electronic Imaging, 23:063003, 11 2014. 2, 3, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 197, + 545, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 229 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 229 + ], + "type": "text", + "content": "[19] Y. Li, B. Sixou, and F. Peyrin. A review of the deep learning methods for medical images super resolution problems. IRBM, 42(2):120-133, 2021. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 231, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 231, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 308, + 231, + 545, + 274 + ], + "type": "text", + "content": "[20] Zhupeng Li, Xintong Liu, Jianyu Wang, Zuoqiang Shi, Lingyun Qiu, and Xing Fu. Fast non-line-of-sight imaging based on first photon event stamping. Optics letters, 47(8):1928—1931, April 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 276, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 276, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 308, + 276, + 545, + 319 + ], + "type": "text", + "content": "[21] Zhengpeng Liao, Deyang Jiang, Xiaochun Liu, Andreas Velten, Yajun Ha, and Xin Lou. FPGA accelerator for real-time non-line-of-sight imaging. IEEE Transactions on Circuits and Systems I: Regular Papers, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 321, + 545, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 321, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 308, + 321, + 545, + 364 + ], + "type": "text", + "content": "[22] David B Lindell, Gordon Wetzstein, and Matthew O'Toole. Wave-based non-line-of-sight imaging using fast fk migration. ACM Transactions on Graphics (ToG), 38(4):1-13, 2019. 2, 3, 5, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 365, + 545, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 365, + 545, + 408 + ], + "spans": [ + { + "bbox": [ + 308, + 365, + 545, + 408 + ], + "type": "text", + "content": "[23] Jianjiang Liu, Yijun Zhou, Xin Huang, Zheng-Ping Li, and Feihu Xu. Photon-efficient non-line-of-sight imaging. IEEE Transactions on Computational Imaging, 8:639-650, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 411, + 545, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 411, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 308, + 411, + 545, + 453 + ], + "type": "text", + "content": "[24] Xiaochun Liu, Sebastian Bauer, and Andreas Velten. Phasor field diffraction based reconstruction for fast non-line-of-sight imaging systems. Nature Communications, 11:1645, 2020. 2, 3, 6, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 456, + 545, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 456, + 545, + 499 + ], + "spans": [ + { + "bbox": [ + 308, + 456, + 545, + 499 + ], + "type": "text", + "content": "[25] Xiaochun Liu, Ibón Guillén, Marco Manna, Ji Nam, Azer Reza, Toan Le, Adrián Jarabo, Diego Gutiérrez, and Andreas Velten. Non-line-of-sight imaging using phasor-field virtual wave optics. Nature, 572, 2019. 2, 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 501, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 501, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 308, + 501, + 545, + 544 + ], + "type": "text", + "content": "[26] Xiaochun Liu and Andreas Velten. The role of wigner distribution function in non-line-of-sight imaging. In 2020 IEEE International Conference on Computational Photography (ICCP), pages 1-12. IEEE, 2020. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 545, + 545, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 545, + 545, + 590 + ], + "spans": [ + { + "bbox": [ + 308, + 545, + 545, + 590 + ], + "type": "text", + "content": "[27] Xintong Liu, Jianyu Wang, Zhupeng Li, Zuoqiang Shi, Xing Fu, and Lingyun Qiu. Non-line-of-sight reconstruction with signal-object collaborative regularization. Light: Science & Applications, 08 2021. 2, 3, 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 591, + 545, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 591, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 308, + 591, + 545, + 634 + ], + "type": "text", + "content": "[28] Christopher Metzler, David Lindell, and Gordon Wetzstein. Keyhole imaging: Non-line-of-sight imaging and tracking of moving objects along a single optical path. IEEE Transactions on Computational Imaging, PP:1-1, 12 2020. 2, 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 635, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 635, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 308, + 635, + 545, + 690 + ], + "type": "text", + "content": "[29] Christopher A. Metzler, Felix Heide, Prasana Rangarajan, Muralidhar Madabhushi Balaji, Aparna Viswanath, Ashok Veeraraghavan, and Richard G. Baraniuk. Deep-inverse correlography: towards real-time high-resolution non-line-of-sight imaging. Optica, 7(1):63-71, Jan 2020. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 692, + 545, + 713 + ], + "type": "text", + "content": "[30] Ji Hyun Nam, Eric Brandt, Sebastian Bauer, Xiaochun Liu, Marco Renna, Alberto Tosi, Eftychios Sifakis, and Andreas" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17428" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 105 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 105 + ], + "type": "text", + "content": "Velten. Low-latency time-of-flight non-line-of-sight imaging at 5 frames per second. Nature communications, 12(1):1-10, 2021. 2, 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 140 + ], + "type": "text", + "content": "[31] Matthew O'Toole, David Lindell, and Gordon Wetzstein. Confocal non-line-of-sight imaging based on the light-cone transform. Nature, 555, 2018. 2, 3, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 142, + 287, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 142, + 287, + 196 + ], + "spans": [ + { + "bbox": [ + 48, + 142, + 287, + 196 + ], + "type": "text", + "content": "[32] Chengquan Pei, Anke Zhang, Yue Deng, Feihu Xu, Jiamin Wu, David U-Lei Li, Hui Qiao, Lu Fang, and Qionghai Dai. Dynamic non-line-of-sight imaging system based on the optimization of point spread functions. Opt. Express, 29(20):32349-32364, Sep 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 198, + 287, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 198, + 287, + 241 + ], + "spans": [ + { + "bbox": [ + 48, + 198, + 287, + 241 + ], + "type": "text", + "content": "[33] Syed Azer Reza, Marco La Manna, Sebastian Bauer, and Andreas Velten. Phasor field waves: A huygens-like light transport model for non-line-of-sight imaging applications. Opt. Express, 27(20):29380-29400, Sep 2019. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 243, + 287, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 243, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 48, + 243, + 287, + 297 + ], + "type": "text", + "content": "[34] Christos Thrampoulidis, Gal Shulkind, Feihu Xu, William T. Freeman, Jeffrey H. Shapiro, Antonio Torralba, Franco N. C. Wong, and Gregory W. Wornell. Exploiting occlusion in non-line-of-sight active imaging. IEEE Transactions on Computational Imaging, 4(3):419-431, 2018. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 300, + 287, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 300, + 287, + 354 + ], + "spans": [ + { + "bbox": [ + 48, + 300, + 287, + 354 + ], + "type": "text", + "content": "[35] Chia-Yin Tsai, Aswin C Sankaranarayanan, and Ioannis Gkioulekas. Beyond volumetric albedo-a surface optimization framework for non-line-of-sight imaging. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1545-1555, 2019. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 356, + 287, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 356, + 287, + 420 + ], + "spans": [ + { + "bbox": [ + 48, + 356, + 287, + 420 + ], + "type": "text", + "content": "[36] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Proceedings of the 31st International Conference on Neural Information Processing Systems, NIPS'17, page 6000-6010, Red Hook, NY, USA, 2017. Curran Associates Inc. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 422, + 287, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 422, + 287, + 488 + ], + "spans": [ + { + "bbox": [ + 48, + 422, + 287, + 488 + ], + "type": "text", + "content": "[37] Eric Veach and Leonidas J. Guibas. Optimally combining sampling techniques for monte carlo rendering. In Proceedings of the 22nd Annual Conference on Computer Graphics and Interactive Techniques, SIGGRAPH '95, page 419-428, New York, NY, USA, 1995. Association for Computing Machinery. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 490, + 287, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 490, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 48, + 490, + 287, + 544 + ], + "type": "text", + "content": "[38] Andreas Velten, Thomas Willwacher, Otkrist Gupta, Ashok Veeraraghavan, Moungi G Bawendi, and Ramesh Raskar. Recovering three-dimensional shape around a corner using ultrafast time-of-flight imaging. Nature communications, 3(1):1-8, 2012. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 546, + 287, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 546, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 546, + 287, + 601 + ], + "type": "text", + "content": "[39] Cheng Wu, Jianjiang Liu, Xin Huang, Zheng-Ping Li, Chao Yu, Jun-Tian Ye, Jun Zhang, Qiang Zhang, Xiankang Dou, Vivek Goyal, Feihu Xu, and Jian-Wei Pan. Non-line-of-sight imaging over " + }, + { + "bbox": [ + 48, + 546, + 287, + 601 + ], + "type": "inline_equation", + "content": "1.43\\mathrm{km}" + }, + { + "bbox": [ + 48, + 546, + 287, + 601 + ], + "type": "text", + "content": ". Proceedings of the National Academy of Sciences of the United States of America, 118, 03 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 602, + 287, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 287, + 644 + ], + "type": "text", + "content": "[40] Han Xiao, Kashif Rasul, and Roland Vollgraf. Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms. arXiv preprint arXiv:1708.07747, 2017. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "type": "text", + "content": "[41] Shumian Xin, Sotiris Nousias, Kiriakos N. Kutulakos, Aswin C. Sankaranarayanan, Srinivasa G. Narasimhan, and Ioannis Gkioulekas. A theory of fermat paths for non-line-of-sight shape reconstruction. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6793-6802, 2019. 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 260 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[42] Wenqing Yang, Chao Zhang, Wenjie Jiang, Zexin Zhang, and Baoqing Sun. None-line-of-sight imaging enhanced with spatial multiplexing. Opt. Express, 30(4):5855-5867, Feb 2022. 2, 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "type": "text", + "content": "[43] Wenming Yang, Xuechen Zhang, Yapeng Tian, Wei Wang, Jing-Hao Xue, and Qingmin Liao. Deep learning for single image super-resolution: A brief review. IEEE Transactions on Multimedia, 21(12):3106-3121, Dec 2019. 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 163, + 545, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 196 + ], + "type": "text", + "content": "[44] Jun-Tian Ye, Xin Huang, Zheng-Ping Li, and Feihu Xu. Compressed sensing for active non-line-of-sight imaging. Opt. Express, 29(2):1749-1763, Jan 2021. 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 198, + 545, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 198, + 545, + 260 + ], + "spans": [ + { + "bbox": [ + 307, + 198, + 545, + 260 + ], + "type": "text", + "content": "[45] Sean I Young, David B Lindell, Bernd Girod, David Taubman, and Gordon Wetzstein. Non-line-of-sight surface reconstruction using the directional light-cone transform. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1407-1416, 2020. 2, 3, 6" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17429" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Normal-Guided Garment UV Prediction for Human Re-Texturing/6ef67559-1f18-4ef9-b45a-a009b6d372b6_content_list.json b/2023/Normal-Guided Garment UV Prediction for Human Re-Texturing/6ef67559-1f18-4ef9-b45a-a009b6d372b6_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..844c8c29cf893695e3f9bebe6ee463e4f96cb11b --- /dev/null +++ b/2023/Normal-Guided Garment UV Prediction for Human Re-Texturing/6ef67559-1f18-4ef9-b45a-a009b6d372b6_content_list.json @@ -0,0 +1,1879 @@ +[ + { + "type": "text", + "text": "Normal-guided Garment UV Prediction for Human Re-texturing", + "text_level": 1, + "bbox": [ + 156, + 130, + 816, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yasamin Jafarian† Tuanfeng Y. Wang‡ Duygu Ceylan‡ Jimei Yang‡ Nathan Carr‡ Yi Zhou‡ Hyun Soo Park† †University of Minnesota #Adobe Research", + "bbox": [ + 184, + 178, + 779, + 247 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5c20f43b3f797941927ae0400a5b25835e05ae78844d68cc2cbecd8a974f899a.jpg", + "image_caption": [ + "Original video", + "Figure 1. Geometry aware texture editing from images and videos. This paper presents a novel approach to predict a geometry aware texture (UV) map of a garment from an image (first row) and video (second row). The predicted UV map preserves isometry between texture space and 3D surface by leveraging 3D surface normals. Further, we ensure temporal consistency in the predicted UV map across frames in a video, resulting in physically plausible human appearance editing. Project website: yasamin.page/normal-guided-uv." + ], + "image_footnote": [], + "bbox": [ + 78, + 258, + 495, + 584 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ca2d491b5c3cae86995bd2e6cf1a5fb4a0436fe19250b2ef1bfd02f890ee691b.jpg", + "image_caption": [ + "Re-textured video" + ], + "image_footnote": [], + "bbox": [ + 501, + 258, + 890, + 584 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 669, + 313, + 684 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Clothes undergo complex geometric deformations, which lead to appearance changes. To edit human videos in a physically plausible way, a texture map must take into account not only the garment transformation induced by the body movements and clothes fitting, but also its 3D fine-grained surface geometry. This poses, however, a new challenge of 3D reconstruction of dynamic clothes from an image or a video. In this paper, we show that it is possible to edit dressed human images and videos without 3D reconstruction. We estimate a geometry aware texture map between the garment region in an image and the texture space, a.k.a, UV map. Our UV map is designed to preserve isometry with respect to the underlying 3D surface by making use of the 3D surface normals predicted from the", + "bbox": [ + 75, + 693, + 472, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "image. Our approach captures the underlying geometry of the garment in a self-supervised way, requiring no ground truth annotation of UV maps and can be readily extended to predict temporally coherent UV maps. We demonstrate that our method outperforms the state-of-the-art human UV map estimation approaches on both real and synthetic data.", + "bbox": [ + 496, + 670, + 893, + 760 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 771, + 632, + 787 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While browsing online clothing shops, have you ever wondered how the appearance of a dress of interest would look on you as if you were in a fitting room given your dress with a similar shape? A key technology to enable generating such visual experiences is photorealistic re-texturing—editing the texture of clothes in response to the subject's movement in the presented images or videos in a geomet-", + "bbox": [ + 496, + 796, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "4627", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "rically and temporally coherent way. Over the past few years, there has been a significant advancement in the image and video editing technologies [4-6, 11-13, 17, 20, 21, 26, 31-33, 38, 46, 54], such as inserting advertising logos on videos of moving cars or applying face makeup on social media. However, such editing approaches designed for rigid or semi-rigid surfaces are not suitable for garments that undergo complex secondary motion with respect to the underlying body. For example, the fine wrinkles of the dress in Figure 1 result in complex warps in texture over time. In this paper, we present a new method to edit the appearance of a garment in a given image or video by taking into account its fine-grained geometric deformation.", + "bbox": [ + 75, + 90, + 472, + 282 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Previous works address photorealistic texture editing in two ways. (1) 3D reconstruction and rendering: these approaches can achieve high-fidelity texture editing given highly accurate 3D geometry. On the other side of the coin, their performance is dictated by the quality of the 3D reconstruction. While the 3D geometry of the garment can be learned from paired human appearance data, e.g., human modeling repositories with 3D meshes and renderings [1], due to the scarcity of such data, it often cannot generalize well on unseen real images and videos. (2) Direct texture mapping: by estimating dense UV map, these methods can bypass the procedure of 3D reconstruction [18, 22, 39, 41, 55]. However, they usually lack of geometry details and only capture the underlying human body, thus, not applicable for editing garments. Moreover, when applied to videos, visual artifacts of editing become more salient since they are not aware of underlying deformation of the garment's 3D geometry [25, 57].", + "bbox": [ + 75, + 285, + 472, + 553 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We design our method to enjoy the advantages of both two approaches: preserving realistic details in UV mapping while circumventing 3D reconstruction. Our key insight is that the fundamental geometric property of isometry can be imposed into UV map estimation via the 3D surface normals predicted from an image. We formulate a geometric relationship between the UV map and surface normals in the form of a set of partial differential equations.", + "bbox": [ + 75, + 555, + 472, + 674 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our method takes as input an image or video, its surface normal prediction, and dense optical flow (for video), and outputs the geometry aware UV map estimate. The UV map is modeled by a multi-layer perceptron that can predict UV coordinates given a pixel location in an image. We note that the UV map is defined up to the choice of a reference coordinate frame. To disambiguate this, we condition the neural network with a pre-defined proxy UV map (e.g., DensePose [18]). We use the isometry constraints as a loss to optimize the UV map. Further, for a video, we leverage the per-frame image feature to correlate the UV coordinates of the pixels across time using optical flow.", + "bbox": [ + 75, + 676, + 472, + 854 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions can be concluded in three aspects: (1) a novel formulation that captures the geometric relationship between the 3D surface normals and the UV map by the", + "bbox": [ + 75, + 856, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "isometry constraint, which eliminates the requirement of 3D reconstruction and ground truth UV map; (2) a neural network design that learns to predict temporally coherent UV map for the frames by correlating per-frame image features; (3) stronger performance compared to existing re-texturing methods and compelling results on a wide range of real-world imagery.", + "bbox": [ + 496, + 90, + 893, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 210, + 642, + 226 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our work lies at the intersection of human UV map prediction from images and neural UV map optimization.", + "bbox": [ + 498, + 236, + 890, + 267 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Human Dense UV Map Estimation", + "text_level": 1, + "bbox": [ + 500, + 277, + 805, + 294 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A seminal work of DensePose [18] learns to predict a UV map of humans presented in an image, which opens a new opportunity to edit the appearance of a person without 3D reconstruction [3,52]. A series of subsequent works [18,28,29,39-42,56,58,60] bring out a number of applications for human tracking. However, due to their representation specific to the body surface, they exhibit fundamental limitations in expressing highly deformable loose clothing such as skirts and dresses.", + "bbox": [ + 496, + 301, + 893, + 434 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address this challenge, recent approaches leverage multitask learning [61] or incorporate geodesic distance to learn UV maps [47]. BodyMap [22] incorporates the Vision Transformers to learn per-pixel image features on a continuous body surface that handles loose clothes, different hairstyles, and occlusion. TemporalUV [55] focuses on handling garments by extrapolating the initial DensePose estimates [18] and leveraging image features obtained from an input video to obtain a UV aligned with the garment boundary. Despite their promise, the visual artifacts persist due to a lack of understanding the underlying 3D geometry. Unlike previous approaches, we design our framework such that the resulting UV map satisfies the fundamental geometric property of isometry, which results in physically plausible re-texturing.", + "bbox": [ + 496, + 435, + 893, + 657 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Neural UV Optimization from Videos.", + "text_level": 1, + "bbox": [ + 500, + 669, + 826, + 686 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Another line of work [24, 25, 34, 44, 57] resorts to a layered UV map, capturing the geometry to some degree by incorporating video decomposition [7, 8, 14, 30, 51] to optimize the UV coordinates of the foreground and background based on the observed motion. Kasten et al. [25] unwrap a video into a set of layered 2D atlases where for each pixel in the video, its corresponding 2D coordinate in each of the atlases is predicted. Ye et al. [57] proposes a global sprite image that can group the distinct motion trajectories because the collective object structure has a consistent appearance throughout time. While preserving the temporal coherency and maintaining some coarse UV deformations related to arm or leg movements during a sequence, these methods fall short of capturing micro deformations like wrinkles in", + "bbox": [ + 496, + 693, + 893, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "4628", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b77126e684d136db183fe79e5196927b37661ad0ec7a8c7b08792e87a983bd24.jpg", + "image_caption": [ + "Figure 2. Texture mapping geometry. We study the mapping between the image space, texture space, and 3D space. A point $\\mathbf{x}$ in image is mapped to the texture space with $\\mathbf{u} = g(\\mathbf{x})$ . The mapping $f(\\mathbf{u}) = \\mathbf{X}$ lifts the texture plane into a 3D garment surface by an isometric warping. We use the orthographic projection model, resulting in the first two elements of $f$ is the inverse of $g$ , i.e., $f^{1:2} = g^{-1}$ . The spatial derivatives of $f$ form a tangent plane on the 3D surface at $\\mathbf{X}$ , resulting in $\\mathbf{N}_{\\mathbf{x}} = f_{u} \\times f_{v}$ where $\\mathbf{N}_{\\mathbf{x}}$ is the 3D surface normal, and $f_{u}$ and $f_{v}$ are the spatial derivatives of $f$ with respect to $u$ and $v$ ." + ], + "image_footnote": [], + "bbox": [ + 78, + 73, + 482, + 306 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the clothing. Furthermore, these techniques cannot represent UV mapping for an image and can only be applied to videos. Unlike these methods, our surface normal conditioned UV map is highly sensitive to small geometric details and 3D surface deformations, which can be optimized not only for a video but also for single images.", + "bbox": [ + 75, + 457, + 468, + 547 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 559, + 168, + 574 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our goal is to obtain a continuous texture mapping, which allows editing the appearance of dynamic garments. We leverage the geometric property of isometry to constrain the UV map in the form of partial differential equations. We solve this partial differential equations by optimizing a neural network to generate a geometry aware UV map.", + "bbox": [ + 75, + 584, + 468, + 675 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Texture Mapping without 3D Reconstruction", + "text_level": 1, + "bbox": [ + 76, + 681, + 457, + 699 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Consider a mapping $g(\\mathbf{x}) = \\mathbf{u}$ that maps a pixel location $\\mathbf{x} = (x, y) \\in \\mathbb{R}^2$ in the image space that belongs to a garment of interest to a point $\\mathbf{u} = (u, v) \\in \\mathbb{R}^2$ in the UV space of the garment as shown in Figure 2. The goal of our work is to find such a mapping $g$ that takes into account the local surface geometry measured by the surface normal predicted at $\\mathbf{x}$ . We denote the predicted 3D surface normal of $\\mathbf{x}$ in the camera space as $\\mathbf{N}_{\\mathbf{x}} \\in \\mathbb{S}^2$ .", + "bbox": [ + 75, + 705, + 468, + 823 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let us define an isometric map from the UV texture map to the 3D surface, $f(\\mathbf{u}) = \\mathbf{X}$ . This is the fundamental property of a non-stretchable cloth texture mapping [10].", + "bbox": [ + 75, + 824, + 468, + 869 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\| f _ {u} \\right\\| = \\left\\| f _ {v} \\right\\| = 1, \\quad f _ {u} ^ {\\mathrm {T}} f _ {v} = 0, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 878, + 468, + 897 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $f_{u}$ and $f_{v}$ are the partial derivatives of $f$ with respect to $u$ and $v$ , respectively. Geometrically, $f_{u}$ and $f_{v}$ are the tangential vectors on the 3D surface where their cross product forms the surface normal:", + "bbox": [ + 496, + 90, + 892, + 148 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathbf {N}} _ {\\mathbf {x}} = f _ {u} \\times f _ {v} = f _ {u} (g (\\mathbf {x})) \\times f _ {v} (g (\\mathbf {x})) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 162, + 890, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\widetilde{\\mathbf{N}}_{\\mathbf{x}} \\in \\mathbb{S}^2$ is the surface normal at $\\mathbf{X}$ corresponding to $\\mathbf{x}$ .", + "bbox": [ + 496, + 195, + 890, + 224 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We can find the UV mapping $g$ by matching the surface normal $\\widetilde{\\mathbf{N}}_{\\mathbf{x}}$ derived by Equation (2) and the surface normal predicted from the image $\\mathbf{N}_{\\mathbf{x}}$ :", + "bbox": [ + 496, + 227, + 890, + 273 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\underset {\\boldsymbol {\\theta} _ {g}, \\boldsymbol {\\theta} _ {f}} {\\operatorname {m i n i m i z e}} \\sum_ {\\mathbf {x}} \\| f _ {u} (g (\\mathbf {x})) \\times f _ {v} (g (\\mathbf {x})) - \\mathbf {N} _ {\\mathbf {x}} \\| ^ {2}, \\\\ s. t. \\quad \\| f _ {u} \\| = \\| f _ {v} \\| = 1, f _ {u} ^ {\\top} f _ {v} = 0, \\tag {3} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 539, + 286, + 890, + 339 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\theta_{g}$ and $\\theta_{f}$ are the parameters of the function $g$ and $f$ , respectively. A key challenge of solving Equation (3) lies in the dependency of $f$ that requires full 3D reconstruction of the surface. Instead, we formulate a new dual problem that can solve Equation (3) effectively without finding $f$ .", + "bbox": [ + 496, + 351, + 890, + 426 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We use two properties to eliminate $f$ from Equation (3). First, we assume orthographic projection, i.e., $(x,y) = (X,Y)$ where $\\mathbf{X} = \\left[ \\begin{array}{ccc}X & Y & Z \\end{array} \\right]^{\\mathrm{T}}$ . This allows us to express the 3D derivatives using the pixel coordinates:", + "bbox": [ + 496, + 426, + 890, + 491 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\ng = \\left(f ^ {1: 2}\\right) ^ {- 1}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 642, + 502, + 890, + 523 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $f^{1:2}$ is the first two elements $(X, Y)$ of $f$ . To keep $f^{1:2}$ bijective, we assume there is no self occlusion in the camera projection of $f$ . Note that $g^{-1}$ is the inverse of $g$ that maps the UV texture map to the pixel coordinate. Second, we derive the derivatives of $g$ by using the inverse function theorem [9]:", + "bbox": [ + 496, + 535, + 890, + 625 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nf ^ {1: 2} \\circ g (\\mathbf {x}) = \\mathbf {x} \\rightarrow \\mathbf {J} _ {g} = \\left(\\mathbf {J} _ {(f ^ {1: 2})}\\right) ^ {- 1}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 637, + 890, + 657 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{J}_g = \\left[g_x\\quad g_y\\right] = \\left[ \\begin{array}{cc}\\frac{\\partial u}{\\partial x} & \\frac{\\partial u}{\\partial y}\\\\ \\frac{\\partial v}{\\partial x} & \\frac{\\partial v}{\\partial y} \\end{array} \\right]$ is the Jacobian matrix of the function $g$", + "bbox": [ + 496, + 670, + 890, + 723 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "With Equation (1) and (5), Equation (2) can be re-written as the following constraints by eliminating $f$ (3D reconstruction):", + "bbox": [ + 496, + 724, + 890, + 768 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\| g _ {x} \\right\\| = \\sqrt {1 + \\frac {\\widetilde {n} _ {x} ^ {2}}{\\widetilde {n} _ {z} ^ {2}}}, \\left\\| g _ {y} \\right\\| = \\sqrt {1 + \\frac {\\widetilde {n} _ {y} ^ {2}}{\\widetilde {n} _ {z} ^ {2}}}, g _ {x} ^ {\\intercal} g _ {y} = \\frac {\\widetilde {n} _ {x} \\widetilde {n} _ {y}}{\\widetilde {n} _ {z} ^ {2}}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 493, + 781, + 890, + 823 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\widetilde{\\mathbf{N}}_{\\mathbf{x}} = \\left[\\widetilde{n}_{x}\\quad \\widetilde{n}_{y}\\quad \\widetilde{n}_{z}\\right]^{\\mathrm{T}}$ . For the derivation of Equation (6), see Supplementary Material.", + "bbox": [ + 496, + 837, + 890, + 869 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Equation (6) is a set of partial differential equations of $g$ that needs to match with the predicted surface normal $\\mathbf{N}_{\\mathbf{x}} =$", + "bbox": [ + 496, + 869, + 890, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "4629", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/855dd2df351d699838881e994ff24750573122b799c8a4c013220ab617dcf353.jpg", + "image_caption": [ + "Figure 3. 2D simplification of UV map estimation. We illustrate the isometric relationship between (a) the surface and (c) the UV map. The length of the curve in XZ axis needs to be preserved when mapping to $g(x)$ , i.e., $|\\mathrm{d}u| = |\\mathrm{d}r| = \\sqrt{\\mathrm{d}x^2 + \\mathrm{d}z^2}$ (arc length preservation). This relationship can be re-written as (b) a partial differential equation using Equation (6) in terms of its surface normal and the inverse of the spatial derivative of $g$ , i.e., $\\left(\\frac{\\mathrm{d}u}{\\mathrm{d}x}\\right)^2 = 1 + \\left(\\frac{\\mathrm{d}n_x}{\\mathrm{d}n_z}\\right)^2$ ." + ], + "image_footnote": [], + "bbox": [ + 94, + 74, + 457, + 344 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\left[ \\begin{array}{lll}n_{x} & n_{y} & n_{z} \\end{array} \\right]^{\\mathsf{T}}$ , leading to a loss function:", + "bbox": [ + 78, + 474, + 367, + 498 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {g e o}} (\\boldsymbol {\\theta} _ {g}) = \\sum_ {\\mathbf {x}} \\left(\\left. \\left(\\frac {\\partial u}{\\partial x} \\right| _ {\\mathbf {x}}\\right) ^ {2} + \\left. \\left(\\frac {\\partial v}{\\partial x} \\right| _ {\\mathbf {x}}\\right) ^ {2} - 1 - \\frac {n _ {x} ^ {2}}{n _ {z} ^ {2}}\\right) ^ {2} \\\\ + \\left(\\left. \\left(\\frac {\\partial u}{\\partial y} \\right| _ {\\mathbf {x}}\\right) ^ {2} + \\left. \\left(\\frac {\\partial v}{\\partial y} \\right| _ {\\mathbf {x}}\\right) ^ {2} - 1 - \\frac {n _ {y} ^ {2}}{n _ {z} ^ {2}}\\right) ^ {2} \\\\ + \\left(\\frac {\\partial u}{\\partial x} \\bigg | _ {\\mathbf {x}} \\frac {\\partial u}{\\partial y} \\bigg | _ {\\mathbf {x}} + \\left. \\frac {\\partial v}{\\partial x} \\right| _ {\\mathbf {x}} \\left. \\frac {\\partial v}{\\partial y} \\right| _ {\\mathbf {x}} - \\frac {n _ {x} n _ {y}}{n _ {z} ^ {2}}\\right) ^ {2}, \\tag {7} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 73, + 507, + 468, + 636 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\left.\\frac{\\partial u}{\\partial x}\\right|_{\\mathbf{x}}$ is the partial derivative of $u$ with respect to $x$ evaluated at $\\mathbf{x}$ .", + "bbox": [ + 75, + 646, + 468, + 676 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Figure 3 illustrates a 2D simplification of UV map estimation. A curve in XZ plane forms a isometric relation with the UV map, $|\\mathrm{du}| = |\\mathrm{dr}| = \\sqrt{\\mathrm{dx}^2 + \\mathrm{dz}^2}$ . This relationship can be re-written as a partial differential equation in terms of the surface normal $(n_x, n_z)$ and the inverse of the spatial derivative of $g(x)$ using Equation (6), i.e., $\\left(\\frac{\\mathrm{du}}{\\mathrm{dx}}\\right)^2 = 1 + \\left(\\frac{\\mathrm{dn}_x}{\\mathrm{dn}_z}\\right)^2$ . We solve these partial differential equations to estimate $g$ .", + "bbox": [ + 75, + 678, + 468, + 808 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Self-supervised Learning of Texture Mapping", + "text_level": 1, + "bbox": [ + 76, + 816, + 464, + 835 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The texture map $g$ is defined up to a bijective function, i.e., there exists an infinite number of $g$ that are equivalent: $g^{-1} \\circ g = (\\mathcal{T} \\circ g)^{-1} \\circ (\\mathcal{T} \\circ g)$ , where $\\mathcal{T}$ is a bijective map (e.g., Euclidean transform). We resolve this ambiguity by", + "bbox": [ + 75, + 840, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "finding $g$ such that $g \\approx g'$ where $g'$ is a pre-defined proxy map of humans:", + "bbox": [ + 498, + 90, + 890, + 121 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {p r o x}} \\left(\\boldsymbol {\\theta} _ {g}\\right) = \\sum_ {\\mathbf {x}} \\left\\| g ^ {\\prime} (\\mathbf {x}) - g (\\mathbf {x}) \\right\\| ^ {2}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 132, + 890, + 162 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In practice, we use an extended DensePose [18] as the pre-defined proxy map. Since DensePose makes predictions only for the human body, we apply an extrapolation method [50] to inpaint the garment regions that are not covered by DensePose.", + "bbox": [ + 496, + 172, + 890, + 247 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Further, we ensure physical plausibility of the visible 3D surfaces, i.e., the texture map should result in the surface normals pointing to $+Z$ direction, by adding the following loss:", + "bbox": [ + 498, + 247, + 890, + 306 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {z}} \\left(\\boldsymbol {\\theta} _ {g}\\right) = \\sum_ {\\mathbf {x}} \\max (0, \\det (\\mathbf {J} _ {g} | _ {\\mathbf {x}})), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 316, + 890, + 347 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\operatorname{det}(\\mathbf{J}_g)$ is the determinant of the Jacobian $\\mathbf{J}_g$ that is equivalent to $\\widetilde{n}_z$ . $\\mathbf{J}_g|_{\\mathbf{x}}$ is the Jacobian matrix of $g$ evaluated at $\\mathbf{x}$ . See Supplementary Material for derivation.", + "bbox": [ + 496, + 358, + 890, + 402 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For a video, we extend the texture map to include the image feature for each pixel, i.e., $g(\\mathbf{x}, \\mathbf{f}_{\\mathbf{x}})$ where $\\mathbf{f}_{\\mathbf{x}}$ is the image feature at $\\mathbf{x}$ . This allows us to generalize the texture map over time. With the extension, we ensure the temporal consistency of the texture map by leveraging optical flow across frames:", + "bbox": [ + 496, + 402, + 890, + 491 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\operatorname {t m p}} \\left(\\boldsymbol {\\theta} _ {g}\\right) = \\sum_ {i, j} \\sum_ {\\mathbf {x} _ {i}} \\left\\| g \\left(\\mathbf {x} _ {i}, \\mathbf {f} _ {\\mathbf {x} _ {i}}\\right) - g \\left(\\mathbf {x} _ {j}, \\mathbf {f} _ {\\mathbf {x} _ {j}}\\right) \\right\\| ^ {2}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 500, + 890, + 534 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{x}_i$ is a point in the $i^{\\mathrm{th}}$ frame. This point is mapped to $\\mathbf{x}_j$ in the $j^{\\mathrm{th}}$ frame, i.e., $\\mathbf{x}_j = W_{i\\rightarrow j}(\\mathbf{x}_i)$ where $W_{i\\rightarrow j}$ is the optical flow from the $i^{\\mathrm{th}}$ to $j^{\\mathrm{th}}$ frames.", + "bbox": [ + 496, + 546, + 890, + 590 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Overall, we optimize the following loss to learn the texture map:", + "bbox": [ + 498, + 590, + 890, + 621 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} \\left(\\boldsymbol {\\theta} _ {g}\\right) = \\mathcal {L} _ {\\text {g e o}} + \\lambda_ {\\text {p r o x}} \\mathcal {L} _ {\\text {p r o x}} + \\lambda_ {\\mathrm {z}} \\mathcal {L} _ {\\mathrm {z}} + \\lambda_ {\\text {t m p}} \\mathcal {L} _ {\\text {t m p}}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 632, + 890, + 650 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\lambda_{\\mathrm{prox}}$ , $\\lambda_{\\mathrm{z}}$ , and $\\lambda_{\\mathrm{tmp}}$ are the weights that determine the relative importance of losses. Note that when a single image is used, $\\lambda_{\\mathrm{tmp}} = 0$ .", + "bbox": [ + 496, + 660, + 890, + 707 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Implementation Details", + "text_level": 1, + "bbox": [ + 500, + 714, + 718, + 729 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We model $g$ using a 12-layer multi-layer perception, with ReLU [2] as an activation function after each layer that takes as input a pixel coordinate with positional encoding, $\\gamma (\\mathbf{x})$ where $\\gamma :\\mathbb{R}^2\\to \\mathbb{R}^{128}$ is Fourier based positional encoding [48]. For videos, we use ResNet [19] to extract per-frame 256 dimensional image feature $\\mathbf{f}_{\\mathbf{x}}$ . Our network design is illustrated in Figure 4 (for image-based UV map prediction) and 5 (for video-based UV map prediction). To make our prediction scale-invariant, we crop the garments region with $256\\times 256$ resolution. We use an offthe-shelf garment segmentation software, Graphonomy [15]", + "bbox": [ + 496, + 737, + 890, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4630", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b36a786f68411016cf1750e89cd1e0a7277c39007ed4d489f150e882959aebbb.jpg", + "image_caption": [ + "Figure 4. Single image framework. For each pixel location on a garment $\\mathbf{x}$ , we predict the UV coordinate $\\mathbf{u}$ using a multi-layer perceptron. We enforce isometry to the UV map by matching with the predicted 3D surface normals. $\\mathcal{L}_{\\mathrm{z}}$ ensures that the predicted surface normals point toward the camera. To disambiguate the frame of reference of the UV map, we use pre-defined proxy map (e.g., DensePose [18] extrapolation [50]) by applying $\\mathcal{L}_{\\mathrm{prox}}$ ." + ], + "image_footnote": [], + "bbox": [ + 81, + 90, + 478, + 270 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "to separate the garment area. We use Adam optimizer [27] with batch size of 2048 and learning rate of $10^{-4}$ . We set $\\lambda_{\\mathrm{prox}} = 0.2$ , $\\lambda_{\\mathrm{z}} = 0.01$ , and $\\lambda_{\\mathrm{tmp}} = 0.3$ . We used an NVIDIA V100 GPU and Intel(R) Xeon(R) CPU E5-2698 v4 @ 2.20GHz, and implemented our approach with Pytorch [43]. Our method takes 20 minutes for a video of 82 frames while Kasten et al. [25], Ye et al. [57], and TemporalUV [55] take 10 hours, 30 minutes, and 23 hours, respectively.", + "bbox": [ + 75, + 382, + 472, + 517 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Evaluation", + "text_level": 1, + "bbox": [ + 76, + 532, + 194, + 547 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate our method both quantitatively and qualitatively on real images as well as synthetic data with ground truth UV map. We also compare with the state-of-the-art methods on human UV map estimation methods and video UV map optimization approaches.", + "bbox": [ + 75, + 558, + 468, + 632 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Datasets We evaluate our method using the following datasets: (1) five synthetic video sequences of simulated dress and T-shirt garments from Santesteban et al. [45] with random texture patterns over 700 frames; (2) ten real videos from Fashion Video dataset [59]; (3) TikTok dataset [23] and various YouTube videos as well as in-the-wild internet images.", + "bbox": [ + 75, + 633, + 468, + 736 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Metric We use five metrics to evaluate our method. (1) UV error: for synthetic data, we report the absolute UV error in the texture space [45]. We use a Procrustes analysis [16] to align the resulting texture map to account for the diambiguity of the reference frame. We report the mean squared error in metric scale by assuming the height of the person in the input is $165\\mathrm{cm}$ , resulting in $0.41\\mathrm{cm} / \\mathrm{UV}$ for dress and $0.27\\mathrm{cm} / \\mathrm{UV}$ for T-shirt (Table 1). (2) Average Precision percentage: we report the Average Precision (AP) percentage metric computed on all the pixels considering a per-pixel prediction as correct if the UV", + "bbox": [ + 75, + 737, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "error is lower than a threshold. We visualize the AP metric for a range of thresholds from 1 to $15\\mathrm{cm}$ and obtain the graph shown in Figure 7.", + "bbox": [ + 496, + 90, + 890, + 135 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(3) Photometric error: we warp the first frame of an input video to the rest of the frames using the UV map estimates. We report the error between the ground truth RGB images and the warped RGB images as reported in Table 1.", + "(4) Geometric error: we report the $\\mathcal{L}_{\\mathrm{geo}}$ to show how the predicted UV map follows the geometric information captured in the surface normal estimates as reported in Table 1. (5) Temporal error: we evaluate the capability of the different approaches in preserving the temporal coherency by reporting the $\\mathcal{L}_{\\mathrm{tmp}}$ error (Table 1)." + ], + "bbox": [ + 496, + 135, + 893, + 282 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "When the ground truth UV is not available, we use the geometric and temporal errors to evaluate our method.", + "bbox": [ + 496, + 282, + 890, + 313 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baseline Methods We compare our method with previous works that fall into two categories: (1) human UV map prediction; (2) UV optimization.", + "bbox": [ + 496, + 316, + 890, + 361 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1) Human UV map prediction: we compare our method with state-of-the-art that focus on predicting UV maps for the naked human body [18] and dressed humans [47, 55]. We also report the performance of the UV map obtained by extrapolating DensePose predictions as discussed in Section 3.3. Our method achieves the best performance as shown in Table 1 and Figure 7. We notice that DensePose [18] performs competently in precision percentage when the threshold error is less than $7\\mathrm{cm}$ . This observation is based on the fact that, for each pixel, DensePose predicts a part label (among 24 parts) and a UV map with respect to that body part. When aligning these predictions with the ground truth, we warp each occupied body part individually, resulting in a more accurate alignment compared to the other methods (including ours) that are represented by only one patch. However, the performance of DensePose [18] is not improved above $7\\mathrm{cm}$ because of limited ability to predict beyond body surface.", + "2) UV optimization: we compare our method with state-of-the-art in predicting the UV map of a dynamic object observed in a video [25, 57]. However, such methods are not tailored for garments that undergo highly nonlinear transformations as the body moves. Hence, as reported in Table 1 and Figure 7, our method surpasses these baselines in the dense UV error, the average precision percentage, the geometric error, and the temporal error.", + "Ablation Study We conduct an ablation study to analyze the impact of the distance (first two terms in Equation (7)) and angle (third term in Equation (7)) constraints (second and third rows of Table 2 and Figure 7). Our final method has the best performance in UV error and photometric error. We also compare the performance of our method without the temporal consistency $(\\mathcal{L}_{\\mathrm{tmp}})$ (fourth row of Table 2). As expected, this term performs quite similarly to ours in UV error but very poorly in photometric error since the consistency between the frames is not enforced. The role of" + ], + "bbox": [ + 496, + 364, + 893, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "4631", + "bbox": [ + 480, + 944, + 513, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c4350c841f8b3e0633b3d05bf94b2b0772b730aaeea84c7f34c68fddcd3f082c.jpg", + "image_caption": [ + "Figure 5. Video framework. We apply temporal coherence in the predicted UV maps using optical flow [35,36,49]. With the optical flow $W_{i\\rightarrow j}$ , we match the UV prediction of $i^{\\mathrm{th}}$ frame with $j^{\\mathrm{th}}$ via $\\mathcal{L}_{\\mathrm{tmp}}$ ." + ], + "image_footnote": [], + "bbox": [ + 78, + 51, + 893, + 191 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/92f945a7d8ecbd7d85bb0a6f15d8de3bcee25b726c6cfdeddda02e15ff4edca8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
GT dress sequences [45]GT T-shirt sequences [45]Real Fashion sequences [59]
MethodUV. error (cm)photo. errorUV. error (cm)photo. errorgeo. error (Lgeo)tmp. error
DensePose [18]17.27±3.7652.43±10.207.48±0.5234.19±7.011.26±0.478.12±2.70
Extrapolated DensePose [18, 50]8.61±0.7618.29±3.245.34±0.6318.86±4.550.52±0.064.62±0.73
HumanGPS [47]11.97±2.0697.41±30.877.53±0.43108.10±34.891.27±0.6951.40±41.66
Kasten et al. [25]7.06±0.6013.07±2.796.64±0.6314.54±4.260.56±0.102.30±0.70
Ye et al. [57]5.56±0.2933.57±-10.275.75±0.2019.22±5.140.71±0.031.65±0.15
Ours3.16±0.287.54±2.043.58±0.2711.28±2.010.07±0.031.50±0.23
", + "bbox": [ + 76, + 244, + 918, + 383 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/742b3a90ab56cdce922dd061a2f718e3f28a18e144d4d4f339410326f6281274.jpg", + "table_caption": [ + "Table 1. Quantitative Results. UV. error (cm), photo. error (RGB difference), geo. error $(\\mathcal{L}_{\\mathrm{geo}})$ , and tmp. error $(\\mathcal{L}_{\\mathrm{tmp}})$ (image space pixel distance) respectively (mean±std)." + ], + "table_footnote": [], + "table_body": "
MethodUV. error (cm)photo. error (RGB)
Ours3.16±0.287.54±2.04
Distance constraint3.44±0.258.29±1.99
Angle constraint5.96±0.337.68±1.37
W/o Ltmp3.22±0.2414.67±4.82
W/o Lprox3.24±0.358.46±2.19
", + "bbox": [ + 89, + 417, + 452, + 513 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/dd8bdaaa327e79b9acf2d672920507703edd830b2347aacd8cc811fa011efeeb.jpg", + "image_caption": [ + "Figure 6. The impact of $\\mathcal{L}_{\\mathrm{prox}}$ in UV prediction and retexturing." + ], + "image_footnote": [], + "bbox": [ + 80, + 553, + 472, + 635 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b797576192909f839a67b3eb1be501094e380a744fb0365a41dbc0d5f7e635b1.jpg", + "image_caption": [ + "Figure 7. Average precision. We compute the average of pixels with a UV error higher than a threshold (1-15cm) on our method, the baseline methods, and our ablation experiments on ground truth dress sequences [45]." + ], + "image_footnote": [], + "bbox": [ + 89, + 648, + 452, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$\\mathcal{L}_{\\mathrm{prox}}$ is the disambiguation of UV maps, i.e., there exist an infinite number of equivalent UV maps that minimize our", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$\\mathcal{L}_{\\mathrm{geo}}$ (PDE). While the result without $\\mathcal{L}_{\\mathrm{prox}}$ are, therefore, quantitatively competitive as summarized in Table 2, such ambiguity can be resolved by finding UV that is closest to the proxy UV as shown in Figure 6, i.e., the retexture without $\\mathcal{L}_{\\mathrm{prox}}$ can result in arbitrary orientation across subjects.", + "bbox": [ + 496, + 419, + 890, + 494 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative Results To show our results qualitatively, we visualize both re-texturing examples and grid UV illustration to depict the performance of each method in preserving high-frequency details. For retexturing, we first obtain an albedo and shading layer from the input image using an intrinsic image decomposition method [53]. After applying a new texture pattern to the albedo layer, we compose it back with the original shading layer. We apply Gamma-correction [37] on the input image, after generating the albedo layer and shading layer, we inverse the Gamma-correction back when synthesizing the re-textured image. We compare our method qualitatively with the baselines as shown in Figure 8 that illustrates the results on Fashion video sequence [59]. Figure 11 shows the performance of our method on videos and images. Our method not only captures the fine-grained surface details but also is temporally coherent across time.", + "bbox": [ + 496, + 500, + 892, + 750 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "User Study We conducted a user study: asking participants $(n = 21)$ to rate realism (1: unrealistic to 10: most realistic) for our method compared to the baselines as summarized in Table 3. Our method receives the highest score from the users.", + "bbox": [ + 496, + 752, + 679, + 898 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/6bc06d65a26229954a48075ac1b5e217432187b6994c480c3dfd9c6d8a535152.jpg", + "table_caption": [ + "Table 2. Ablation study on dress sequences [45]. UV. error (cm), photo. error (RGB difference) (mean±std)." + ], + "table_footnote": [], + "table_body": "
MethodUser score
DensePose4.09±1.94
Proxy DensePose3.23±1.86
HumanGPS2.28±2.12
Kasten et al.5.52±1.88
Ye et al.7.66±1.27
Ours9.57±0.81
", + "bbox": [ + 691, + 768, + 874, + 867 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. User study.", + "bbox": [ + 740, + 869, + 861, + 883 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "4632", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2be33698c31c3356af08f6606fee2bcdcd11f53e86467fa4d6248403c57e307c.jpg", + "image_caption": [ + "Input image" + ], + "image_footnote": [], + "bbox": [ + 99, + 47, + 187, + 183 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/66871bc101a26ecb8b3f4d6653c161726106cb1e107bda0a291c1eb9ee00c803.jpg", + "image_caption": [ + "Ours w/shading" + ], + "image_footnote": [], + "bbox": [ + 187, + 50, + 243, + 183 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d581e1ceaee474d287393955d80892872e26bee9360bdaa30ebf2ab98b1b5ee5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 253, + 47, + 339, + 183 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f0c59451cfbdcc3a274f87dfa7ac1c8bfcd3a89e7cd745ab5158e717513a4d6c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 253, + 185, + 336, + 237 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fc0e6518f4d8a9998b027713f4c5ab93506f51601dd2ba617b250824658b93cd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 254, + 238, + 310, + 300 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e2e21bd3012fc32ac4a0e6975f2a732e9901c2fa48cb1629b7dd1b45d6b3ffba.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 334, + 47, + 419, + 183 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6c0779eccc17662bfbbe9b586320103499e2652b4bc3b2a932791d8815d1514f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 336, + 185, + 419, + 237 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/be2b0b0ca24b073d82d227c6d888f92ed75effcf4792672b89b5ab2959ba482f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 336, + 238, + 411, + 300 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d008d2f38e6805cbc8eb593a5f61d2f85815d4f97cc02fad91d5c3408faaa0c4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 419, + 47, + 506, + 183 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4c39a9ca66fc27ffe3af2a28983ef9b2929a40ecae1af04062e5fc8ba26c5bc0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 419, + 185, + 504, + 237 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/31e635c011324386602f02f02738f49b9170fc9263875e6da0a5d0c72addafe4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 421, + 238, + 486, + 300 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/480086e898fe57b0e7973d780db1ed696bac5753de7f5652f55c4072b874b632.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 47, + 586, + 183 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7770163965a7cd33562533ad2c386eb95d1be5579eda984eac87a87b1eeaf0a5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 185, + 578, + 237 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c7e489fc13df2fd43159c6b68d6d5499c7c239818614422c76276846440c6232.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 238, + 558, + 300 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/45c683ea33a5861ea29c689615bebe4058a8b6a8947ea501827d7a6d1ab4eeb4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 583, + 47, + 669, + 183 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ebd15b28592d60249c8672f975e629a67e09967e203f1a94ef7fa2ae1957f242.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 583, + 185, + 669, + 300 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b259f90d519ef78375436ce17d19d3173236c052786c59e33411daebc71f853f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 669, + 47, + 751, + 183 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/886860ca57c58c90863285f2d7d3a439d83ca6eb8e7adb35d0e5c42fd58eef4d.jpg", + "image_caption": [ + "Retturung w/o shading", + "Grid-UV" + ], + "image_footnote": [], + "bbox": [ + 751, + 47, + 836, + 183 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6aa03086274b31b936d5a616bcab1a8c23dda42eda1a830ecc0d5c3c989134f8.jpg", + "image_caption": [ + "High", + "Ours" + ], + "image_footnote": [], + "bbox": [ + 243, + 306, + 310, + 402 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/252a6aee4920ca0aed033b349cc7d82275495fbd6ddd63209881e0e9da6a85fe.jpg", + "image_caption": [ + "Extrapolated DensePose" + ], + "image_footnote": [], + "bbox": [ + 328, + 308, + 385, + 401 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ec66ecb3b0f900893d82e8cbae0b2513d8b519fe2f2a0187dc490b0157ee9ac9.jpg", + "image_caption": [ + "DensePose" + ], + "image_footnote": [], + "bbox": [ + 421, + 306, + 473, + 402 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3f1dce994bbdac0e06ee8e167c6af5d7aeea9107c369776e4597a42534143776.jpg", + "image_caption": [ + "HumanGPS" + ], + "image_footnote": [], + "bbox": [ + 504, + 310, + 558, + 401 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/52c4afc0f43bc7319342125e37437841ef402d7c1f15ca60ed425f231836d7b9.jpg", + "image_caption": [ + "TemporalUV" + ], + "image_footnote": [], + "bbox": [ + 589, + 310, + 640, + 400 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d3ecf278b26fc169e62faa97d61bd9865174cf75556f3aea47de03ae9fb71706.jpg", + "image_caption": [ + "Kasten et al." + ], + "image_footnote": [], + "bbox": [ + 674, + 310, + 725, + 400 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fba2ca528dfd684478828c0914d5bff2aca6ab427293ed6e6b7c066618b59c0c.jpg", + "image_caption": [ + "Ye et al." + ], + "image_footnote": [], + "bbox": [ + 756, + 309, + 810, + 398 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e4fd2d54886b0f66d565fce9ee3e2eb272487b2d460c1cfe29bf7ba25a57fcb0.jpg", + "image_caption": [ + "Figure 8. Qualitative comparison. We compare ours with our initial UV (extrapolated DensePose), DensePose [18], HumanGPS [47], Kastenn et al. [25], and Ye et al. [57] in image re-texturing and UV map grid visualization.", + "Figure 9. Limitation. Due to lack of 3D reconstructed surface, our method cannot handle folds that introduce texture discontinuity. Our texture map is continuous around the folds, which is physically incorrect." + ], + "image_footnote": [], + "bbox": [ + 81, + 452, + 470, + 648 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Discussion", + "text_level": 1, + "bbox": [ + 76, + 710, + 189, + 726 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "This paper presents a novel approach to predict a high quality UV map by preserving geometric details from images and videos. We leverage the geometric property of isometry encoded in 3D surface normals to optimize the UV map in the form of partial differential equations. We generalize our method to videos by integrating optical flow, resulting in a temporally coherent video editing. Our method produces strong qualitative and quantitative predictions on real-world imagery compared to state-of-the-art UV map estimation.", + "bbox": [ + 75, + 736, + 468, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Limitation As discussed in Section 3.1, our method makes", + "bbox": [ + 76, + 885, + 468, + 898 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "an assumption about projection, i.e., there is one-to-one correspondence between 3D surface geometry and image. However, this assumption does not hold when there is a fold where a region of 3D surface is not visible to the image. This makes a contrast with 3D reconstruction based method where the invisible part of 3D surface can be still mapped to the image via depth reasoning. Figure 9 illustrates this limitation where there are folds in the skirt, resulting in negative surface normal $n_z$ . Due to the folds, the texture must be discontinuous while our method produces continuous texture rendering due to the lack of 3D reconstructed geometry, which is physically incorrect.", + "bbox": [ + 496, + 454, + 890, + 631 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our pipeline is composed of two components: (1) UV prediction (our contribution) and (2) texture map with shading (not our contribution). For the garments with highly contrasted texture, the shading operation is often biased to color con", + "bbox": [ + 496, + 632, + 660, + 792 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/159478e428bec14296451264dcd57ed3cbad387285f19a63561be588e0284499.jpg", + "image_caption": [ + "Image", + "Figure 10. Limitation. Limitations on textured garment." + ], + "image_footnote": [], + "bbox": [ + 663, + 633, + 740, + 744 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4802e97fdfc756fa95ab8753afcafd6f5a35e50bf28dc2ae7ff181112f5568e9.jpg", + "image_caption": [ + "Retexturing w/o shading" + ], + "image_footnote": [], + "bbox": [ + 741, + 633, + 813, + 743 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2438e0cbecb0eb5827c2d9afeef9573c31459cc756ff01e6c3888cf68534bcf2.jpg", + "image_caption": [ + "Retexturing w/shading" + ], + "image_footnote": [], + "bbox": [ + 813, + 633, + 887, + 743 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "trast, resulting in erroneous appearance. Figure 10 illustrates that despite reasonable UV prediction from our method, the resulting appearance is unrealistic near the textured region. Improving the shading operation is beyond the scope of this work, and we leave it as future work.", + "bbox": [ + 496, + 795, + 890, + 867 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Acknowledgement This work was supported by a NSF NRI 2022894 and NSF CAREER 1846031.", + "bbox": [ + 498, + 868, + 890, + 896 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "4633", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cbde4900b39fdffb54508ce11134874a813903016d4d62681fbe926833fe5b8a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 88, + 215, + 273 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a307d949748ced08d0937bc708ca36f9ab6ff6ce04435e747fecf83350b8ee65.jpg", + "image_caption": [ + "Re-textured images" + ], + "image_footnote": [], + "bbox": [ + 218, + 88, + 888, + 458 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/59721a2c8d4eb4d13fcf13229abb841d62e65fb4e76329bfa4f28e6df9dc7f46.jpg", + "image_caption": [ + "Original image" + ], + "image_footnote": [], + "bbox": [ + 78, + 273, + 215, + 457 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0a6c659cd622c567deeba997a3a30f4966357ec89a395d236c3ad9da5b82ddd3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 488, + 338, + 613 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c5317c17f98d04683b39a8dfa7cca28c755a9f7206c2c4f413cb7054b192a6cc.jpg", + "image_caption": [ + "Original video" + ], + "image_footnote": [], + "bbox": [ + 127, + 614, + 302, + 808 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ddac916919564fdf45bfa6f928c0f0a2e0ef28c8ee9e740378d233f12a9ec5d0.jpg", + "image_caption": [ + "(a) Qualitative results of our method on an image.", + "(b) Qualitative results of our method on videos.", + "Figure 11. Results on images and videos. Our method can produce compelling texture editing given images and videos that preserve geometric details." + ], + "image_footnote": [], + "bbox": [ + 343, + 488, + 890, + 823 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "4634", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] https://renderpeople.com/3d-people.2", + "[2] Abien Fred Agarap. Deep learning using rectified linear units (relu). arXiv, 2018. 4", + "[3] Badour AlBahar, Jingwan Lu, Jimei Yang, Zhixin Shu, Eli Shechtman, and Jia-Bin Huang. Pose with Style: Detail-preserving pose-guided image synthesis with conditional stylegan. TOG, 2021. 2", + "[4] Yazeed Alharbi and Peter Wonka. Disentangled image generation through structured noise injection. In CVPR, 2020. 2", + "[5] David Bau, Hendrik Strobelt, William Peebles, Jonas Wulff, Bolei Zhou, Jun-Yan Zhu, and Antonio Torralba. Semantic photo manipulation with a generative image prior. TOG, 2019. 2", + "[6] Bharat Lal Bhatnagar, Garvita Tiwari, Christian Theobalt, and Gerard Pons-Moll. Multi-garment net: Learning to dress 3d people from images. In ICCV, 2019. 2", + "[7] Michael J. Black and P. Anandan. Robust dynamic motion estimation over time. In CVPR, 1991. 2", + "[8] Gabriel J Brostow and Irfan A Essa. Motion based decompositing of video. In ICCV, 1999. 2", + "[9] Felix E. Browder. On the unification of the calculus of variations and the theory of monotone nonlinear operators in banach spaces. PNAS, 1966. 3", + "[10] Edwin Earl Catmull. A subdivision algorithm for computer display of curved surfaces. The University of Utah, 1974. 3", + "[11] Shu-Yu Chen, Wanchao Su, Lin Gao, Shihong Xia, and Hongbo Fu. Deepfacedrawing: Deep generation of face images from sketches. TOG, 2020. 2", + "[12] Anton Cherepkov, Andrey Voynov, and Artem Babenko. Navigating the gan parameter space for semantic image editing. In CVPR, 2021. 2", + "[13] Edo Collins, Raja Bala, Bob Price, and Sabine Süssstrunk. Editing in style: Uncovering the local semantics of GANs. In CVPR, 2020. 2", + "[14] T. Darrell and A. Pentland. Robust estimation of a multilayered motion representation. In Proceedings of the IEEE Workshop on Visual Motion, 1991. 2", + "[15] Ke Gong, Yiming Gao, Xiaodan Liang, Xiaohui Shen, Meng Wang, and Liang Lin. Graphonomy: Universal human parsing via graph transfer learning. In CVPR, 2019. 4", + "[16] J.C. Gower and G.B. Dijksterhuis. Procrustes problems. new york: Oxford university press. Psychometrika, 70, 2005. 5", + "[17] Artur Grigorev, Artem Sevastopolsky, Alexander Vakhitov, and Victor Lempitsky. Coordinate-based texture inpainting for pose-guided human image generation. In CVPR, 2019. 2", + "[18] Riza Alp Güler, Natalia Neverova, and Iasonas Kokkinos. Densesepos: Dense human pose estimation in the wild. In CVPR, 2018. 2, 4, 5, 6, 7", + "[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 4", + "[20] Zhenliang He, Wangmeng Zuo, Meina Kan, Shiguang Shan, and Xilin Chen. Attgan: Facial attribute editing by only changing what you want. TIP, 2019. 2" + ], + "bbox": [ + 78, + 114, + 467, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[21] Xianxu Hou, Xiaokang Zhang, Linlin Shen, Zhihui Lai, and Jun Wan. Guidedstyle: Attribute knowledge guided style manipulation for semantic face editing. arXiv, 2020. 2", + "[22] Anastasia Ianina, Nikolaos Sarafianos, Yuanlu Xu, Ignacio Rocco, and Tony Tung. Bodymap: Learning full-body dense correspondence map. In CVPR, 2022. 2", + "[23] Yasamin Jafarian and Hyun Soo Park. Learning high fidelity depths of dressed humans by watching social media dance videos. In CVPR, 2021. 5", + "[24] Varun Jampani, Raghudeep Gadde, and Peter V. Gehler. Video propagation networks. In CVPR, 2017. 2", + "[25] Yoni Kasten, Dolev Ofri, Oliver Wang, and Tali Dekel. Layered neural atlases for consistent video editing. TOG, 2021. 2, 5, 6, 7", + "[26] Hyunsu Kim, Yunjey Choi, Junho Kim, Sungwoo Yoo, and Youngjung Uh. Exploiting spatial dimensions of latent in gan for real-time image editing. In CVPR, 2021. 2", + "[27] Diederik P. Kingma and Jimmy Lei Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 5", + "[28] Nilesh Kulkarni, Abhinav Gupta, David F. Fouhey, and Shubham Tulsiani. Articulation-aware canonical surface mapping. In CVPR, 2020. 2", + "[29] Nilesh Kulkarni, Abhinav Gupta, David F Fouhey, and Shubham Tulsiani. Articulation-aware canonical surface mapping. In CVPR, 2020. 2", + "[30] M. Pawan Kumar, Philip H. S. Torr, and Andrew Zisserman. Learning layered motion segmentations of video. IJCV, 2005. 2", + "[31] Verica Lazova, Eldar Insafutdinov, and Gerard Pons-Moll. 360-degree textures of people in clothing from a single image. 3DV, 2019. 2", + "[32] Cheng-Han Lee, Ziwei Liu, Lingyun Wu, and Ping Luo. Maskgan: Towards diverse and interactive facial image manipulation. In CVPR, 2020. 2", + "[33] Huan Ling, Karsten Kreis, Daiqing Li, Seung Wook Kim, Antonio Torralba, and Sanja Fidler. Editgan: High-precision semantic image editing. In NeurIPS, 2021. 2", + "[34] Erika Lu, Forrester Cole, Tali Dekel, Weidi Xie, Andrew Zisserman, David Salesin, William T. Freeman, and Michael Rubinstein. Layered neural rendering for retiming people in video. TOG, 2020. 2", + "[35] Bruce D. Lucas and Takeo Kanade. An iterative image registration technique with an application to stereo vision. In IJCAI, 1981. 6", + "[36] Bruce D. Lucas and Takeo Kanade. Optical navigation by the method of differences. In *IJCAI*, 1985. 6", + "[37] Tom McReynolds and David Blythe. Advanced graphics programming using OpenGL. Elsevier, 2005. 6", + "[38] Aymen Mir, Thiemo Alldieck, and Gerard Pons-Moll. Learning to transfer texture from clothing images to 3d humans. In CVPR, 2020. 2", + "[39] Natalia Neverova, David Novotny, Vasil Khalidov, Marc Szafraniec, Patrick Labatut, and Andrea Vedaldi. Continuous surface embeddings. In NeurIPS, 2020. 2", + "[40] Natalia Neverova, David Novotny, and Andrea Vedaldi. Correlated uncertainty for learning dense correspondences from noisy labels. In NeurIPS, 2019. 2" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "4635", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Natalia Neverova, Artsiom Sanakoyeu, David Novotny, Patrick Labatut, and Andrea Vedaldi. Discovering relationships between object categories via universal canonical maps. In CVPR, 2021. 2", + "[42] Natalia Neverova, James Thewlis, Riza Alp Güler, Iasonas Kokkinos, and Andrea Vedaldi. Slim densepose: Thrifty learning from sparse annotations and motion cues. In CVPR, 2019. 2", + "[43] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In NeurIPS. 2019. 5", + "[44] Alex Rav-Acha, Pushmeet Kohli, Carsten Rother, and Andrew Fitzgibbon. Unwrap mosaics: A new representation for video editing. SIGGRAPH, 2008. 2", + "[45] Igor Santesteban, Nils Thuerey, Miguel A Otaduy, and Dan Casas. Self-Supervised Collision Handling via Generative 3D Garment Models for Virtual Try-On. In CVPR, 2021. 5, 6", + "[46] Yujun Shen, Jinjin Gu, Xiaou Tang, and Bolei Zhou. Interpreting the latent space of gans for semantic face editing. In CVPR, 2020. 2", + "[47] Feitong Tan, Danhang Tang, Dou Mingsong, Guo Kaiwen, Rohit Pandey, Cem Keskin, Ruofei Du, Deqing Sun, Sofien Bouaziz, Sean Fanello, Ping Tan, and Yinda Zhang. HumanGPS: Geodesic preserving feature for dense human correspondences. In CVPR, 2021. 2, 5, 6, 7", + "[48] Matthew Tancik, Pratul P. Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan T. Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. In NeurIPS, 2020. 4", + "[49] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In ECCV, 2020. 6", + "[50] Alexandru Telea. An image inpainting technique based on the fast marching method. Journal of Graphics Tools, 2004, 4, 5, 6", + "[51] John Y. A. Wang and Edward H. Adelson. Representing moving images with layers. TIP, 1994. 2", + "[52] Tuanfeng Wang, Duygu Ceylan, Krishna Kumar Singh, and Niloy J. Mitra. Dance in the wild: Monocular human animation with neural dynamic appearance synthesis. In 3DV, 2021. 2", + "[53] Yair Weiss. Deriving intrinsic images from image sequences. In ICCV, 2001. 6", + "[54] Rongliang Wu, Gongjie Zhang, Shijian Lu, and Tao Chen. Cascade ef-gan: Progressive facial expression editing with local focuses. In CVPR, 2020. 2", + "[55] You Xie, Huiqi Mao, Angela Yao, and Nils Thuerey. Temporaluv: Capturing loose clothing with temporally coherent uv coordinates. In CVPR, 2022. 2, 5", + "[56] Haonan Yan, Jiaqi Chen, Xujie Zhang, Shengkai Zhang, Nianhong Jiao, Xiaodan Liang, and Tianxiang Zheng. Ultrapose: Synthesizing dense pose with 1 billion points by human-body decoupling 3d model. In ICCV, 2021. 2" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[57] Vickie Ye, Zhengqi Li, Richard Tucker, Angjoo Kanazawa, and Noah Snavelly. Deformable sprites for unsupervised video decomposition. In CVPR, 2022. 2, 5, 6, 7", + "[58] Zhixuan Yu, Haozheng Yu, Long Sha, Sujoy Ganguly, and Hyun Soo Park. Semi-supervised dense keypoints using unlabeled multiview images. arXiv, 2021. 2", + "[59] Polina Zablotskaia, Aliaksandr Siarohin, Bo Zhao, and Leonid Sigal. Dwnet: Dense warp-based network for pose-guided human video generation. In BMVC, 2019. 5, 6", + "[60] Wang Zeng, Wanli Ouyang, Ping Luo, Wentao Liu, and Xiaogang Wang. 3d human mesh regression with dense correspondence. In CVPR, 2020. 2", + "[61] Tyler Zhu, Per Karlsson, and Chris Bregler. Simpose: Effectively learning densepose and surface normal of people from simulated data. In ECCV, 2020. 2" + ], + "bbox": [ + 501, + 90, + 890, + 299 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "4636", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Normal-Guided Garment UV Prediction for Human Re-Texturing/6ef67559-1f18-4ef9-b45a-a009b6d372b6_model.json b/2023/Normal-Guided Garment UV Prediction for Human Re-Texturing/6ef67559-1f18-4ef9-b45a-a009b6d372b6_model.json new file mode 100644 index 0000000000000000000000000000000000000000..075f337da93d1110aca14129e5aae1c01a0b8425 --- /dev/null +++ b/2023/Normal-Guided Garment UV Prediction for Human Re-Texturing/6ef67559-1f18-4ef9-b45a-a009b6d372b6_model.json @@ -0,0 +1,2750 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.158, + 0.131, + 0.817, + 0.154 + ], + "angle": 0, + "content": "Normal-guided Garment UV Prediction for Human Re-texturing" + }, + { + "type": "text", + "bbox": [ + 0.186, + 0.179, + 0.78, + 0.248 + ], + "angle": 0, + "content": "Yasamin Jafarian† Tuanfeng Y. Wang‡ Duygu Ceylan‡ Jimei Yang‡ Nathan Carr‡ Yi Zhou‡ Hyun Soo Park† †University of Minnesota #Adobe Research" + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.26, + 0.496, + 0.585 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.218, + 0.587, + 0.31, + 0.599 + ], + "angle": 0, + "content": "Original video" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.26, + 0.892, + 0.585 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.657, + 0.586, + 0.768, + 0.598 + ], + "angle": 0, + "content": "Re-textured video" + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.602, + 0.895, + 0.659 + ], + "angle": 0, + "content": "Figure 1. Geometry aware texture editing from images and videos. This paper presents a novel approach to predict a geometry aware texture (UV) map of a garment from an image (first row) and video (second row). The predicted UV map preserves isometry between texture space and 3D surface by leveraging 3D surface normals. Further, we ensure temporal consistency in the predicted UV map across frames in a video, resulting in physically plausible human appearance editing. Project website: yasamin.page/normal-guided-uv." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.67, + 0.314, + 0.685 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.694, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Clothes undergo complex geometric deformations, which lead to appearance changes. To edit human videos in a physically plausible way, a texture map must take into account not only the garment transformation induced by the body movements and clothes fitting, but also its 3D fine-grained surface geometry. This poses, however, a new challenge of 3D reconstruction of dynamic clothes from an image or a video. In this paper, we show that it is possible to edit dressed human images and videos without 3D reconstruction. We estimate a geometry aware texture map between the garment region in an image and the texture space, a.k.a, UV map. Our UV map is designed to preserve isometry with respect to the underlying 3D surface by making use of the 3D surface normals predicted from the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.671, + 0.895, + 0.761 + ], + "angle": 0, + "content": "image. Our approach captures the underlying geometry of the garment in a self-supervised way, requiring no ground truth annotation of UV maps and can be readily extended to predict temporally coherent UV maps. We demonstrate that our method outperforms the state-of-the-art human UV map estimation approaches on both real and synthetic data." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.772, + 0.633, + 0.788 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.797, + 0.895, + 0.903 + ], + "angle": 0, + "content": "While browsing online clothing shops, have you ever wondered how the appearance of a dress of interest would look on you as if you were in a fitting room given your dress with a similar shape? A key technology to enable generating such visual experiences is photorealistic re-texturing—editing the texture of clothes in response to the subject's movement in the presented images or videos in a geomet-" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4627" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.284 + ], + "angle": 0, + "content": "rically and temporally coherent way. Over the past few years, there has been a significant advancement in the image and video editing technologies [4-6, 11-13, 17, 20, 21, 26, 31-33, 38, 46, 54], such as inserting advertising logos on videos of moving cars or applying face makeup on social media. However, such editing approaches designed for rigid or semi-rigid surfaces are not suitable for garments that undergo complex secondary motion with respect to the underlying body. For example, the fine wrinkles of the dress in Figure 1 result in complex warps in texture over time. In this paper, we present a new method to edit the appearance of a garment in a given image or video by taking into account its fine-grained geometric deformation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.286, + 0.473, + 0.554 + ], + "angle": 0, + "content": "Previous works address photorealistic texture editing in two ways. (1) 3D reconstruction and rendering: these approaches can achieve high-fidelity texture editing given highly accurate 3D geometry. On the other side of the coin, their performance is dictated by the quality of the 3D reconstruction. While the 3D geometry of the garment can be learned from paired human appearance data, e.g., human modeling repositories with 3D meshes and renderings [1], due to the scarcity of such data, it often cannot generalize well on unseen real images and videos. (2) Direct texture mapping: by estimating dense UV map, these methods can bypass the procedure of 3D reconstruction [18, 22, 39, 41, 55]. However, they usually lack of geometry details and only capture the underlying human body, thus, not applicable for editing garments. Moreover, when applied to videos, visual artifacts of editing become more salient since they are not aware of underlying deformation of the garment's 3D geometry [25, 57]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.556, + 0.473, + 0.675 + ], + "angle": 0, + "content": "We design our method to enjoy the advantages of both two approaches: preserving realistic details in UV mapping while circumventing 3D reconstruction. Our key insight is that the fundamental geometric property of isometry can be imposed into UV map estimation via the 3D surface normals predicted from an image. We formulate a geometric relationship between the UV map and surface normals in the form of a set of partial differential equations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.677, + 0.473, + 0.855 + ], + "angle": 0, + "content": "Our method takes as input an image or video, its surface normal prediction, and dense optical flow (for video), and outputs the geometry aware UV map estimate. The UV map is modeled by a multi-layer perceptron that can predict UV coordinates given a pixel location in an image. We note that the UV map is defined up to the choice of a reference coordinate frame. To disambiguate this, we condition the neural network with a pre-defined proxy UV map (e.g., DensePose [18]). We use the isometry constraints as a loss to optimize the UV map. Further, for a video, we leverage the per-frame image feature to correlate the UV coordinates of the pixels across time using optical flow." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.857, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Our contributions can be concluded in three aspects: (1) a novel formulation that captures the geometric relationship between the 3D surface normals and the UV map by the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.198 + ], + "angle": 0, + "content": "isometry constraint, which eliminates the requirement of 3D reconstruction and ground truth UV map; (2) a neural network design that learns to predict temporally coherent UV map for the frames by correlating per-frame image features; (3) stronger performance compared to existing re-texturing methods and compelling results on a wide range of real-world imagery." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.211, + 0.643, + 0.227 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.237, + 0.892, + 0.268 + ], + "angle": 0, + "content": "Our work lies at the intersection of human UV map prediction from images and neural UV map optimization." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.279, + 0.806, + 0.295 + ], + "angle": 0, + "content": "2.1. Human Dense UV Map Estimation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.302, + 0.895, + 0.435 + ], + "angle": 0, + "content": "A seminal work of DensePose [18] learns to predict a UV map of humans presented in an image, which opens a new opportunity to edit the appearance of a person without 3D reconstruction [3,52]. A series of subsequent works [18,28,29,39-42,56,58,60] bring out a number of applications for human tracking. However, due to their representation specific to the body surface, they exhibit fundamental limitations in expressing highly deformable loose clothing such as skirts and dresses." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.436, + 0.895, + 0.659 + ], + "angle": 0, + "content": "To address this challenge, recent approaches leverage multitask learning [61] or incorporate geodesic distance to learn UV maps [47]. BodyMap [22] incorporates the Vision Transformers to learn per-pixel image features on a continuous body surface that handles loose clothes, different hairstyles, and occlusion. TemporalUV [55] focuses on handling garments by extrapolating the initial DensePose estimates [18] and leveraging image features obtained from an input video to obtain a UV aligned with the garment boundary. Despite their promise, the visual artifacts persist due to a lack of understanding the underlying 3D geometry. Unlike previous approaches, we design our framework such that the resulting UV map satisfies the fundamental geometric property of isometry, which results in physically plausible re-texturing." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.67, + 0.828, + 0.687 + ], + "angle": 0, + "content": "2.2. Neural UV Optimization from Videos." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.694, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Another line of work [24, 25, 34, 44, 57] resorts to a layered UV map, capturing the geometry to some degree by incorporating video decomposition [7, 8, 14, 30, 51] to optimize the UV coordinates of the foreground and background based on the observed motion. Kasten et al. [25] unwrap a video into a set of layered 2D atlases where for each pixel in the video, its corresponding 2D coordinate in each of the atlases is predicted. Ye et al. [57] proposes a global sprite image that can group the distinct motion trajectories because the collective object structure has a consistent appearance throughout time. While preserving the temporal coherency and maintaining some coarse UV deformations related to arm or leg movements during a sequence, these methods fall short of capturing micro deformations like wrinkles in" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "4628" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.074, + 0.483, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.313, + 0.474, + 0.45 + ], + "angle": 0, + "content": "Figure 2. Texture mapping geometry. We study the mapping between the image space, texture space, and 3D space. A point \\(\\mathbf{x}\\) in image is mapped to the texture space with \\(\\mathbf{u} = g(\\mathbf{x})\\). The mapping \\(f(\\mathbf{u}) = \\mathbf{X}\\) lifts the texture plane into a 3D garment surface by an isometric warping. We use the orthographic projection model, resulting in the first two elements of \\(f\\) is the inverse of \\(g\\), i.e., \\(f^{1:2} = g^{-1}\\). The spatial derivatives of \\(f\\) form a tangent plane on the 3D surface at \\(\\mathbf{X}\\), resulting in \\(\\mathbf{N}_{\\mathbf{x}} = f_{u} \\times f_{v}\\) where \\(\\mathbf{N}_{\\mathbf{x}}\\) is the 3D surface normal, and \\(f_{u}\\) and \\(f_{v}\\) are the spatial derivatives of \\(f\\) with respect to \\(u\\) and \\(v\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.458, + 0.47, + 0.548 + ], + "angle": 0, + "content": "the clothing. Furthermore, these techniques cannot represent UV mapping for an image and can only be applied to videos. Unlike these methods, our surface normal conditioned UV map is highly sensitive to small geometric details and 3D surface deformations, which can be optimized not only for a video but also for single images." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.56, + 0.169, + 0.575 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.585, + 0.47, + 0.676 + ], + "angle": 0, + "content": "Our goal is to obtain a continuous texture mapping, which allows editing the appearance of dynamic garments. We leverage the geometric property of isometry to constrain the UV map in the form of partial differential equations. We solve this partial differential equations by optimizing a neural network to generate a geometry aware UV map." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.683, + 0.459, + 0.7 + ], + "angle": 0, + "content": "3.1. Texture Mapping without 3D Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.706, + 0.469, + 0.824 + ], + "angle": 0, + "content": "Consider a mapping \\( g(\\mathbf{x}) = \\mathbf{u} \\) that maps a pixel location \\( \\mathbf{x} = (x, y) \\in \\mathbb{R}^2 \\) in the image space that belongs to a garment of interest to a point \\( \\mathbf{u} = (u, v) \\in \\mathbb{R}^2 \\) in the UV space of the garment as shown in Figure 2. The goal of our work is to find such a mapping \\( g \\) that takes into account the local surface geometry measured by the surface normal predicted at \\( \\mathbf{x} \\). We denote the predicted 3D surface normal of \\( \\mathbf{x} \\) in the camera space as \\( \\mathbf{N}_{\\mathbf{x}} \\in \\mathbb{S}^2 \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.469, + 0.87 + ], + "angle": 0, + "content": "Let us define an isometric map from the UV texture map to the 3D surface, \\( f(\\mathbf{u}) = \\mathbf{X} \\). This is the fundamental property of a non-stretchable cloth texture mapping [10]." + }, + { + "type": "equation", + "bbox": [ + 0.171, + 0.88, + 0.47, + 0.898 + ], + "angle": 0, + "content": "\\[\n\\left\\| f _ {u} \\right\\| = \\left\\| f _ {v} \\right\\| = 1, \\quad f _ {u} ^ {\\mathrm {T}} f _ {v} = 0, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.893, + 0.15 + ], + "angle": 0, + "content": "where \\( f_{u} \\) and \\( f_{v} \\) are the partial derivatives of \\( f \\) with respect to \\( u \\) and \\( v \\), respectively. Geometrically, \\( f_{u} \\) and \\( f_{v} \\) are the tangential vectors on the 3D surface where their cross product forms the surface normal:" + }, + { + "type": "equation", + "bbox": [ + 0.564, + 0.164, + 0.892, + 0.182 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathbf {N}} _ {\\mathbf {x}} = f _ {u} \\times f _ {v} = f _ {u} (g (\\mathbf {x})) \\times f _ {v} (g (\\mathbf {x})) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.196, + 0.892, + 0.226 + ], + "angle": 0, + "content": "where \\(\\widetilde{\\mathbf{N}}_{\\mathbf{x}} \\in \\mathbb{S}^2\\) is the surface normal at \\(\\mathbf{X}\\) corresponding to \\(\\mathbf{x}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.228, + 0.892, + 0.275 + ], + "angle": 0, + "content": "We can find the UV mapping \\( g \\) by matching the surface normal \\( \\widetilde{\\mathbf{N}}_{\\mathbf{x}} \\) derived by Equation (2) and the surface normal predicted from the image \\( \\mathbf{N}_{\\mathbf{x}} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.54, + 0.287, + 0.892, + 0.34 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\underset {\\boldsymbol {\\theta} _ {g}, \\boldsymbol {\\theta} _ {f}} {\\operatorname {m i n i m i z e}} \\sum_ {\\mathbf {x}} \\| f _ {u} (g (\\mathbf {x})) \\times f _ {v} (g (\\mathbf {x})) - \\mathbf {N} _ {\\mathbf {x}} \\| ^ {2}, \\\\ s. t. \\quad \\| f _ {u} \\| = \\| f _ {v} \\| = 1, f _ {u} ^ {\\top} f _ {v} = 0, \\tag {3} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.352, + 0.892, + 0.427 + ], + "angle": 0, + "content": "where \\(\\theta_{g}\\) and \\(\\theta_{f}\\) are the parameters of the function \\(g\\) and \\(f\\), respectively. A key challenge of solving Equation (3) lies in the dependency of \\(f\\) that requires full 3D reconstruction of the surface. Instead, we formulate a new dual problem that can solve Equation (3) effectively without finding \\(f\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.428, + 0.892, + 0.492 + ], + "angle": 0, + "content": "We use two properties to eliminate \\( f \\) from Equation (3). First, we assume orthographic projection, i.e., \\( (x,y) = (X,Y) \\) where \\( \\mathbf{X} = \\left[ \\begin{array}{ccc}X & Y & Z \\end{array} \\right]^{\\mathrm{T}} \\). This allows us to express the 3D derivatives using the pixel coordinates:" + }, + { + "type": "equation", + "bbox": [ + 0.643, + 0.503, + 0.892, + 0.524 + ], + "angle": 0, + "content": "\\[\ng = \\left(f ^ {1: 2}\\right) ^ {- 1}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.536, + 0.892, + 0.625 + ], + "angle": 0, + "content": "where \\( f^{1:2} \\) is the first two elements \\( (X, Y) \\) of \\( f \\). To keep \\( f^{1:2} \\) bijective, we assume there is no self occlusion in the camera projection of \\( f \\). Note that \\( g^{-1} \\) is the inverse of \\( g \\) that maps the UV texture map to the pixel coordinate. Second, we derive the derivatives of \\( g \\) by using the inverse function theorem [9]:" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.638, + 0.892, + 0.659 + ], + "angle": 0, + "content": "\\[\nf ^ {1: 2} \\circ g (\\mathbf {x}) = \\mathbf {x} \\rightarrow \\mathbf {J} _ {g} = \\left(\\mathbf {J} _ {(f ^ {1: 2})}\\right) ^ {- 1}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.671, + 0.891, + 0.724 + ], + "angle": 0, + "content": "where \\(\\mathbf{J}_g = \\left[g_x\\quad g_y\\right] = \\left[ \\begin{array}{cc}\\frac{\\partial u}{\\partial x} & \\frac{\\partial u}{\\partial y}\\\\ \\frac{\\partial v}{\\partial x} & \\frac{\\partial v}{\\partial y} \\end{array} \\right]\\) is the Jacobian matrix of the function \\(g\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.725, + 0.892, + 0.769 + ], + "angle": 0, + "content": "With Equation (1) and (5), Equation (2) can be re-written as the following constraints by eliminating \\( f \\) (3D reconstruction):" + }, + { + "type": "equation", + "bbox": [ + 0.495, + 0.782, + 0.892, + 0.824 + ], + "angle": 0, + "content": "\\[\n\\left\\| g _ {x} \\right\\| = \\sqrt {1 + \\frac {\\widetilde {n} _ {x} ^ {2}}{\\widetilde {n} _ {z} ^ {2}}}, \\left\\| g _ {y} \\right\\| = \\sqrt {1 + \\frac {\\widetilde {n} _ {y} ^ {2}}{\\widetilde {n} _ {z} ^ {2}}}, g _ {x} ^ {\\intercal} g _ {y} = \\frac {\\widetilde {n} _ {x} \\widetilde {n} _ {y}}{\\widetilde {n} _ {z} ^ {2}}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.838, + 0.892, + 0.87 + ], + "angle": 0, + "content": "where \\(\\widetilde{\\mathbf{N}}_{\\mathbf{x}} = \\left[\\widetilde{n}_{x}\\quad \\widetilde{n}_{y}\\quad \\widetilde{n}_{z}\\right]^{\\mathrm{T}}\\) . For the derivation of Equation (6), see Supplementary Material." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Equation (6) is a set of partial differential equations of \\( g \\) that needs to match with the predicted surface normal \\( \\mathbf{N}_{\\mathbf{x}} = \\)" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4629" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.075, + 0.458, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.351, + 0.473, + 0.472 + ], + "angle": 0, + "content": "Figure 3. 2D simplification of UV map estimation. We illustrate the isometric relationship between (a) the surface and (c) the UV map. The length of the curve in XZ axis needs to be preserved when mapping to \\( g(x) \\), i.e., \\( |\\mathrm{d}u| = |\\mathrm{d}r| = \\sqrt{\\mathrm{d}x^2 + \\mathrm{d}z^2} \\) (arc length preservation). This relationship can be re-written as (b) a partial differential equation using Equation (6) in terms of its surface normal and the inverse of the spatial derivative of \\( g \\), i.e., \\( \\left(\\frac{\\mathrm{d}u}{\\mathrm{d}x}\\right)^2 = 1 + \\left(\\frac{\\mathrm{d}n_x}{\\mathrm{d}n_z}\\right)^2 \\)." + }, + { + "type": "text", + "bbox": [ + 0.079, + 0.476, + 0.369, + 0.499 + ], + "angle": 0, + "content": "\\(\\left[ \\begin{array}{lll}n_{x} & n_{y} & n_{z} \\end{array} \\right]^{\\mathsf{T}}\\) , leading to a loss function:" + }, + { + "type": "equation", + "bbox": [ + 0.075, + 0.508, + 0.469, + 0.637 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {g e o}} (\\boldsymbol {\\theta} _ {g}) = \\sum_ {\\mathbf {x}} \\left(\\left. \\left(\\frac {\\partial u}{\\partial x} \\right| _ {\\mathbf {x}}\\right) ^ {2} + \\left. \\left(\\frac {\\partial v}{\\partial x} \\right| _ {\\mathbf {x}}\\right) ^ {2} - 1 - \\frac {n _ {x} ^ {2}}{n _ {z} ^ {2}}\\right) ^ {2} \\\\ + \\left(\\left. \\left(\\frac {\\partial u}{\\partial y} \\right| _ {\\mathbf {x}}\\right) ^ {2} + \\left. \\left(\\frac {\\partial v}{\\partial y} \\right| _ {\\mathbf {x}}\\right) ^ {2} - 1 - \\frac {n _ {y} ^ {2}}{n _ {z} ^ {2}}\\right) ^ {2} \\\\ + \\left(\\frac {\\partial u}{\\partial x} \\bigg | _ {\\mathbf {x}} \\frac {\\partial u}{\\partial y} \\bigg | _ {\\mathbf {x}} + \\left. \\frac {\\partial v}{\\partial x} \\right| _ {\\mathbf {x}} \\left. \\frac {\\partial v}{\\partial y} \\right| _ {\\mathbf {x}} - \\frac {n _ {x} n _ {y}}{n _ {z} ^ {2}}\\right) ^ {2}, \\tag {7} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.647, + 0.47, + 0.678 + ], + "angle": 0, + "content": "where \\(\\left.\\frac{\\partial u}{\\partial x}\\right|_{\\mathbf{x}}\\) is the partial derivative of \\(u\\) with respect to \\(x\\) evaluated at \\(\\mathbf{x}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.679, + 0.47, + 0.809 + ], + "angle": 0, + "content": "Figure 3 illustrates a 2D simplification of UV map estimation. A curve in XZ plane forms a isometric relation with the UV map, \\( |\\mathrm{du}| = |\\mathrm{dr}| = \\sqrt{\\mathrm{dx}^2 + \\mathrm{dz}^2} \\). This relationship can be re-written as a partial differential equation in terms of the surface normal \\( (n_x, n_z) \\) and the inverse of the spatial derivative of \\( g(x) \\) using Equation (6), i.e., \\( \\left(\\frac{\\mathrm{du}}{\\mathrm{dx}}\\right)^2 = 1 + \\left(\\frac{\\mathrm{dn}_x}{\\mathrm{dn}_z}\\right)^2 \\). We solve these partial differential equations to estimate \\( g \\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.818, + 0.465, + 0.836 + ], + "angle": 0, + "content": "3.2. Self-supervised Learning of Texture Mapping" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.841, + 0.471, + 0.903 + ], + "angle": 0, + "content": "The texture map \\( g \\) is defined up to a bijective function, i.e., there exists an infinite number of \\( g \\) that are equivalent: \\( g^{-1} \\circ g = (\\mathcal{T} \\circ g)^{-1} \\circ (\\mathcal{T} \\circ g) \\), where \\( \\mathcal{T} \\) is a bijective map (e.g., Euclidean transform). We resolve this ambiguity by" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.091, + 0.892, + 0.122 + ], + "angle": 0, + "content": "finding \\( g \\) such that \\( g \\approx g' \\) where \\( g' \\) is a pre-defined proxy map of humans:" + }, + { + "type": "equation", + "bbox": [ + 0.578, + 0.133, + 0.892, + 0.164 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {p r o x}} \\left(\\boldsymbol {\\theta} _ {g}\\right) = \\sum_ {\\mathbf {x}} \\left\\| g ^ {\\prime} (\\mathbf {x}) - g (\\mathbf {x}) \\right\\| ^ {2}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.174, + 0.892, + 0.248 + ], + "angle": 0, + "content": "In practice, we use an extended DensePose [18] as the pre-defined proxy map. Since DensePose makes predictions only for the human body, we apply an extrapolation method [50] to inpaint the garment regions that are not covered by DensePose." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.248, + 0.892, + 0.307 + ], + "angle": 0, + "content": "Further, we ensure physical plausibility of the visible 3D surfaces, i.e., the texture map should result in the surface normals pointing to \\(+Z\\) direction, by adding the following loss:" + }, + { + "type": "equation", + "bbox": [ + 0.58, + 0.317, + 0.892, + 0.348 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {z}} \\left(\\boldsymbol {\\theta} _ {g}\\right) = \\sum_ {\\mathbf {x}} \\max (0, \\det (\\mathbf {J} _ {g} | _ {\\mathbf {x}})), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.359, + 0.892, + 0.403 + ], + "angle": 0, + "content": "where \\(\\operatorname{det}(\\mathbf{J}_g)\\) is the determinant of the Jacobian \\(\\mathbf{J}_g\\) that is equivalent to \\(\\widetilde{n}_z\\). \\(\\mathbf{J}_g|_{\\mathbf{x}}\\) is the Jacobian matrix of \\(g\\) evaluated at \\(\\mathbf{x}\\). See Supplementary Material for derivation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.404, + 0.892, + 0.492 + ], + "angle": 0, + "content": "For a video, we extend the texture map to include the image feature for each pixel, i.e., \\( g(\\mathbf{x}, \\mathbf{f}_{\\mathbf{x}}) \\) where \\( \\mathbf{f}_{\\mathbf{x}} \\) is the image feature at \\( \\mathbf{x} \\). This allows us to generalize the texture map over time. With the extension, we ensure the temporal consistency of the texture map by leveraging optical flow across frames:" + }, + { + "type": "equation", + "bbox": [ + 0.516, + 0.501, + 0.892, + 0.535 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\operatorname {t m p}} \\left(\\boldsymbol {\\theta} _ {g}\\right) = \\sum_ {i, j} \\sum_ {\\mathbf {x} _ {i}} \\left\\| g \\left(\\mathbf {x} _ {i}, \\mathbf {f} _ {\\mathbf {x} _ {i}}\\right) - g \\left(\\mathbf {x} _ {j}, \\mathbf {f} _ {\\mathbf {x} _ {j}}\\right) \\right\\| ^ {2}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.547, + 0.892, + 0.592 + ], + "angle": 0, + "content": "where \\(\\mathbf{x}_i\\) is a point in the \\(i^{\\mathrm{th}}\\) frame. This point is mapped to \\(\\mathbf{x}_j\\) in the \\(j^{\\mathrm{th}}\\) frame, i.e., \\(\\mathbf{x}_j = W_{i\\rightarrow j}(\\mathbf{x}_i)\\) where \\(W_{i\\rightarrow j}\\) is the optical flow from the \\(i^{\\mathrm{th}}\\) to \\(j^{\\mathrm{th}}\\) frames." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.592, + 0.892, + 0.622 + ], + "angle": 0, + "content": "Overall, we optimize the following loss to learn the texture map:" + }, + { + "type": "equation", + "bbox": [ + 0.513, + 0.633, + 0.892, + 0.651 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} \\left(\\boldsymbol {\\theta} _ {g}\\right) = \\mathcal {L} _ {\\text {g e o}} + \\lambda_ {\\text {p r o x}} \\mathcal {L} _ {\\text {p r o x}} + \\lambda_ {\\mathrm {z}} \\mathcal {L} _ {\\mathrm {z}} + \\lambda_ {\\text {t m p}} \\mathcal {L} _ {\\text {t m p}}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.661, + 0.892, + 0.708 + ], + "angle": 0, + "content": "where \\(\\lambda_{\\mathrm{prox}}\\), \\(\\lambda_{\\mathrm{z}}\\), and \\(\\lambda_{\\mathrm{tmp}}\\) are the weights that determine the relative importance of losses. Note that when a single image is used, \\(\\lambda_{\\mathrm{tmp}} = 0\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.715, + 0.719, + 0.731 + ], + "angle": 0, + "content": "3.3. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.738, + 0.892, + 0.903 + ], + "angle": 0, + "content": "We model \\( g \\) using a 12-layer multi-layer perception, with ReLU [2] as an activation function after each layer that takes as input a pixel coordinate with positional encoding, \\( \\gamma (\\mathbf{x}) \\) where \\( \\gamma :\\mathbb{R}^2\\to \\mathbb{R}^{128} \\) is Fourier based positional encoding [48]. For videos, we use ResNet [19] to extract per-frame 256 dimensional image feature \\( \\mathbf{f}_{\\mathbf{x}} \\). Our network design is illustrated in Figure 4 (for image-based UV map prediction) and 5 (for video-based UV map prediction). To make our prediction scale-invariant, we crop the garments region with \\( 256\\times 256 \\) resolution. We use an offthe-shelf garment segmentation software, Graphonomy [15]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4630" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.092, + 0.48, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.277, + 0.473, + 0.374 + ], + "angle": 0, + "content": "Figure 4. Single image framework. For each pixel location on a garment \\(\\mathbf{x}\\), we predict the UV coordinate \\(\\mathbf{u}\\) using a multi-layer perceptron. We enforce isometry to the UV map by matching with the predicted 3D surface normals. \\(\\mathcal{L}_{\\mathrm{z}}\\) ensures that the predicted surface normals point toward the camera. To disambiguate the frame of reference of the UV map, we use pre-defined proxy map (e.g., DensePose [18] extrapolation [50]) by applying \\(\\mathcal{L}_{\\mathrm{prox}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.383, + 0.473, + 0.518 + ], + "angle": 0, + "content": "to separate the garment area. We use Adam optimizer [27] with batch size of 2048 and learning rate of \\(10^{-4}\\). We set \\(\\lambda_{\\mathrm{prox}} = 0.2\\), \\(\\lambda_{\\mathrm{z}} = 0.01\\), and \\(\\lambda_{\\mathrm{tmp}} = 0.3\\). We used an NVIDIA V100 GPU and Intel(R) Xeon(R) CPU E5-2698 v4 @ 2.20GHz, and implemented our approach with Pytorch [43]. Our method takes 20 minutes for a video of 82 frames while Kasten et al. [25], Ye et al. [57], and TemporalUV [55] take 10 hours, 30 minutes, and 23 hours, respectively." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.533, + 0.196, + 0.549 + ], + "angle": 0, + "content": "4. Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.559, + 0.47, + 0.633 + ], + "angle": 0, + "content": "We evaluate our method both quantitatively and qualitatively on real images as well as synthetic data with ground truth UV map. We also compare with the state-of-the-art methods on human UV map estimation methods and video UV map optimization approaches." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.634, + 0.47, + 0.737 + ], + "angle": 0, + "content": "Evaluation Datasets We evaluate our method using the following datasets: (1) five synthetic video sequences of simulated dress and T-shirt garments from Santesteban et al. [45] with random texture patterns over 700 frames; (2) ten real videos from Fashion Video dataset [59]; (3) TikTok dataset [23] and various YouTube videos as well as in-the-wild internet images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.738, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Evaluation Metric We use five metrics to evaluate our method. (1) UV error: for synthetic data, we report the absolute UV error in the texture space [45]. We use a Procrustes analysis [16] to align the resulting texture map to account for the diambiguity of the reference frame. We report the mean squared error in metric scale by assuming the height of the person in the input is \\(165\\mathrm{cm}\\), resulting in \\(0.41\\mathrm{cm} / \\mathrm{UV}\\) for dress and \\(0.27\\mathrm{cm} / \\mathrm{UV}\\) for T-shirt (Table 1). (2) Average Precision percentage: we report the Average Precision (AP) percentage metric computed on all the pixels considering a per-pixel prediction as correct if the UV" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.136 + ], + "angle": 0, + "content": "error is lower than a threshold. We visualize the AP metric for a range of thresholds from 1 to \\(15\\mathrm{cm}\\) and obtain the graph shown in Figure 7." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.136, + 0.894, + 0.195 + ], + "angle": 0, + "content": "(3) Photometric error: we warp the first frame of an input video to the rest of the frames using the UV map estimates. We report the error between the ground truth RGB images and the warped RGB images as reported in Table 1." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.195, + 0.892, + 0.284 + ], + "angle": 0, + "content": "(4) Geometric error: we report the \\(\\mathcal{L}_{\\mathrm{geo}}\\) to show how the predicted UV map follows the geometric information captured in the surface normal estimates as reported in Table 1. (5) Temporal error: we evaluate the capability of the different approaches in preserving the temporal coherency by reporting the \\(\\mathcal{L}_{\\mathrm{tmp}}\\) error (Table 1)." + }, + { + "type": "list", + "bbox": [ + 0.498, + 0.136, + 0.894, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.284, + 0.892, + 0.314 + ], + "angle": 0, + "content": "When the ground truth UV is not available, we use the geometric and temporal errors to evaluate our method." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.317, + 0.892, + 0.362 + ], + "angle": 0, + "content": "Baseline Methods We compare our method with previous works that fall into two categories: (1) human UV map prediction; (2) UV optimization." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.365, + 0.893, + 0.631 + ], + "angle": 0, + "content": "1) Human UV map prediction: we compare our method with state-of-the-art that focus on predicting UV maps for the naked human body [18] and dressed humans [47, 55]. We also report the performance of the UV map obtained by extrapolating DensePose predictions as discussed in Section 3.3. Our method achieves the best performance as shown in Table 1 and Figure 7. We notice that DensePose [18] performs competently in precision percentage when the threshold error is less than \\(7\\mathrm{cm}\\). This observation is based on the fact that, for each pixel, DensePose predicts a part label (among 24 parts) and a UV map with respect to that body part. When aligning these predictions with the ground truth, we warp each occupied body part individually, resulting in a more accurate alignment compared to the other methods (including ours) that are represented by only one patch. However, the performance of DensePose [18] is not improved above \\(7\\mathrm{cm}\\) because of limited ability to predict beyond body surface." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.634, + 0.894, + 0.752 + ], + "angle": 0, + "content": "2) UV optimization: we compare our method with state-of-the-art in predicting the UV map of a dynamic object observed in a video [25, 57]. However, such methods are not tailored for garments that undergo highly nonlinear transformations as the body moves. Hence, as reported in Table 1 and Figure 7, our method surpasses these baselines in the dense UV error, the average precision percentage, the geometric error, and the temporal error." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.753, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Ablation Study We conduct an ablation study to analyze the impact of the distance (first two terms in Equation (7)) and angle (third term in Equation (7)) constraints (second and third rows of Table 2 and Figure 7). Our final method has the best performance in UV error and photometric error. We also compare the performance of our method without the temporal consistency \\((\\mathcal{L}_{\\mathrm{tmp}})\\) (fourth row of Table 2). As expected, this term performs quite similarly to ours in UV error but very poorly in photometric error since the consistency between the frames is not enforced. The role of" + }, + { + "type": "list", + "bbox": [ + 0.498, + 0.365, + 0.895, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "4631" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.053, + 0.895, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.204, + 0.893, + 0.233 + ], + "angle": 0, + "content": "Figure 5. Video framework. We apply temporal coherence in the predicted UV maps using optical flow [35,36,49]. With the optical flow \\( W_{i\\rightarrow j} \\), we match the UV prediction of \\( i^{\\mathrm{th}} \\) frame with \\( j^{\\mathrm{th}} \\) via \\( \\mathcal{L}_{\\mathrm{tmp}} \\)." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.246, + 0.919, + 0.384 + ], + "angle": 0, + "content": "
GT dress sequences [45]GT T-shirt sequences [45]Real Fashion sequences [59]
MethodUV. error (cm)photo. errorUV. error (cm)photo. errorgeo. error (Lgeo)tmp. error
DensePose [18]17.27±3.7652.43±10.207.48±0.5234.19±7.011.26±0.478.12±2.70
Extrapolated DensePose [18, 50]8.61±0.7618.29±3.245.34±0.6318.86±4.550.52±0.064.62±0.73
HumanGPS [47]11.97±2.0697.41±30.877.53±0.43108.10±34.891.27±0.6951.40±41.66
Kasten et al. [25]7.06±0.6013.07±2.796.64±0.6314.54±4.260.56±0.102.30±0.70
Ye et al. [57]5.56±0.2933.57±-10.275.75±0.2019.22±5.140.71±0.031.65±0.15
Ours3.16±0.287.54±2.043.58±0.2711.28±2.010.07±0.031.50±0.23
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.391, + 0.893, + 0.418 + ], + "angle": 0, + "content": "Table 1. Quantitative Results. UV. error (cm), photo. error (RGB difference), geo. error \\((\\mathcal{L}_{\\mathrm{geo}})\\) , and tmp. error \\((\\mathcal{L}_{\\mathrm{tmp}})\\) (image space pixel distance) respectively (mean±std)." + }, + { + "type": "table", + "bbox": [ + 0.091, + 0.419, + 0.454, + 0.515 + ], + "angle": 0, + "content": "
MethodUV. error (cm)photo. error (RGB)
Ours3.16±0.287.54±2.04
Distance constraint3.44±0.258.29±1.99
Angle constraint5.96±0.337.68±1.37
W/o Ltmp3.22±0.2414.67±4.82
W/o Lprox3.24±0.358.46±2.19
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.522, + 0.468, + 0.55 + ], + "angle": 0, + "content": "Table 2. Ablation study on dress sequences [45]. UV. error (cm), photo. error (RGB difference) (mean±std)." + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.554, + 0.473, + 0.636 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.084, + 0.636, + 0.463, + 0.649 + ], + "angle": 0, + "content": "Figure 6. The impact of \\(\\mathcal{L}_{\\mathrm{prox}}\\) in UV prediction and retexturing." + }, + { + "type": "image", + "bbox": [ + 0.091, + 0.65, + 0.454, + 0.818 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.817, + 0.468, + 0.87 + ], + "angle": 0, + "content": "Figure 7. Average precision. We compute the average of pixels with a UV error higher than a threshold (1-15cm) on our method, the baseline methods, and our ablation experiments on ground truth dress sequences [45]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "\\(\\mathcal{L}_{\\mathrm{prox}}\\) is the disambiguation of UV maps, i.e., there exist an infinite number of equivalent UV maps that minimize our" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.42, + 0.892, + 0.496 + ], + "angle": 0, + "content": "\\(\\mathcal{L}_{\\mathrm{geo}}\\) (PDE). While the result without \\(\\mathcal{L}_{\\mathrm{prox}}\\) are, therefore, quantitatively competitive as summarized in Table 2, such ambiguity can be resolved by finding UV that is closest to the proxy UV as shown in Figure 6, i.e., the retexture without \\(\\mathcal{L}_{\\mathrm{prox}}\\) can result in arbitrary orientation across subjects." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.5, + 0.893, + 0.751 + ], + "angle": 0, + "content": "Qualitative Results To show our results qualitatively, we visualize both re-texturing examples and grid UV illustration to depict the performance of each method in preserving high-frequency details. For retexturing, we first obtain an albedo and shading layer from the input image using an intrinsic image decomposition method [53]. After applying a new texture pattern to the albedo layer, we compose it back with the original shading layer. We apply Gamma-correction [37] on the input image, after generating the albedo layer and shading layer, we inverse the Gamma-correction back when synthesizing the re-textured image. We compare our method qualitatively with the baselines as shown in Figure 8 that illustrates the results on Fashion video sequence [59]. Figure 11 shows the performance of our method on videos and images. Our method not only captures the fine-grained surface details but also is temporally coherent across time." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.753, + 0.681, + 0.9 + ], + "angle": 0, + "content": "User Study We conducted a user study: asking participants \\((n = 21)\\) to rate realism (1: unrealistic to 10: most realistic) for our method compared to the baselines as summarized in Table 3. Our method receives the highest score from the users." + }, + { + "type": "table", + "bbox": [ + 0.692, + 0.77, + 0.875, + 0.868 + ], + "angle": 0, + "content": "
MethodUser score
DensePose4.09±1.94
Proxy DensePose3.23±1.86
HumanGPS2.28±2.12
Kasten et al.5.52±1.88
Ye et al.7.66±1.27
Ours9.57±0.81
" + }, + { + "type": "table_caption", + "bbox": [ + 0.741, + 0.871, + 0.862, + 0.885 + ], + "angle": 0, + "content": "Table 3. User study." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4632" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.048, + 0.188, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.104, + 0.186, + 0.145, + 0.213 + ], + "angle": 0, + "content": "Input image" + }, + { + "type": "image", + "bbox": [ + 0.189, + 0.051, + 0.244, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.19, + 0.186, + 0.242, + 0.213 + ], + "angle": 0, + "content": "Ours w/shading" + }, + { + "type": "image", + "bbox": [ + 0.254, + 0.049, + 0.341, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.254, + 0.186, + 0.337, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.255, + 0.239, + 0.311, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.336, + 0.049, + 0.421, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.337, + 0.186, + 0.421, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.337, + 0.239, + 0.412, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.421, + 0.049, + 0.507, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.421, + 0.186, + 0.505, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.423, + 0.239, + 0.488, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.049, + 0.587, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.186, + 0.58, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.239, + 0.56, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.584, + 0.049, + 0.67, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.584, + 0.186, + 0.67, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.67, + 0.049, + 0.753, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.752, + 0.049, + 0.838, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.84, + 0.049, + 0.858, + 0.169 + ], + "angle": 0, + "content": "Retturung w/o shading" + }, + { + "type": "image_caption", + "bbox": [ + 0.843, + 0.211, + 0.862, + 0.256 + ], + "angle": 0, + "content": "Grid-UV" + }, + { + "type": "image_caption", + "bbox": [ + 0.212, + 0.307, + 0.242, + 0.321 + ], + "angle": 0, + "content": "High" + }, + { + "type": "image", + "bbox": [ + 0.244, + 0.308, + 0.311, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.267, + 0.406, + 0.297, + 0.418 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.329, + 0.309, + 0.387, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.32, + 0.405, + 0.395, + 0.425 + ], + "angle": 0, + "content": "Extrapolated DensePose" + }, + { + "type": "image", + "bbox": [ + 0.423, + 0.308, + 0.474, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.418, + 0.407, + 0.482, + 0.418 + ], + "angle": 0, + "content": "DensePose" + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.311, + 0.56, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.496, + 0.407, + 0.566, + 0.418 + ], + "angle": 0, + "content": "HumanGPS" + }, + { + "type": "image", + "bbox": [ + 0.591, + 0.311, + 0.641, + 0.401 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.583, + 0.407, + 0.659, + 0.419 + ], + "angle": 0, + "content": "TemporalUV" + }, + { + "type": "image", + "bbox": [ + 0.675, + 0.311, + 0.726, + 0.401 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.671, + 0.407, + 0.742, + 0.418 + ], + "angle": 0, + "content": "Kasten et al." + }, + { + "type": "image", + "bbox": [ + 0.757, + 0.31, + 0.811, + 0.4 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.761, + 0.406, + 0.808, + 0.418 + ], + "angle": 0, + "content": "Ye et al." + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.425, + 0.892, + 0.452 + ], + "angle": 0, + "content": "Figure 8. Qualitative comparison. We compare ours with our initial UV (extrapolated DensePose), DensePose [18], HumanGPS [47], Kastenn et al. [25], and Ye et al. [57] in image re-texturing and UV map grid visualization." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.453, + 0.472, + 0.65 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.656, + 0.47, + 0.71 + ], + "angle": 0, + "content": "Figure 9. Limitation. Due to lack of 3D reconstructed surface, our method cannot handle folds that introduce texture discontinuity. Our texture map is continuous around the folds, which is physically incorrect." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.712, + 0.191, + 0.727 + ], + "angle": 0, + "content": "5. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.737, + 0.47, + 0.884 + ], + "angle": 0, + "content": "This paper presents a novel approach to predict a high quality UV map by preserving geometric details from images and videos. We leverage the geometric property of isometry encoded in 3D surface normals to optimize the UV map in the form of partial differential equations. We generalize our method to videos by integrating optical flow, resulting in a temporally coherent video editing. Our method produces strong qualitative and quantitative predictions on real-world imagery compared to state-of-the-art UV map estimation." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.886, + 0.469, + 0.9 + ], + "angle": 0, + "content": "Limitation As discussed in Section 3.1, our method makes" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.455, + 0.892, + 0.632 + ], + "angle": 0, + "content": "an assumption about projection, i.e., there is one-to-one correspondence between 3D surface geometry and image. However, this assumption does not hold when there is a fold where a region of 3D surface is not visible to the image. This makes a contrast with 3D reconstruction based method where the invisible part of 3D surface can be still mapped to the image via depth reasoning. Figure 9 illustrates this limitation where there are folds in the skirt, resulting in negative surface normal \\( n_z \\). Due to the folds, the texture must be discontinuous while our method produces continuous texture rendering due to the lack of 3D reconstructed geometry, which is physically incorrect." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.633, + 0.661, + 0.793 + ], + "angle": 0, + "content": "Our pipeline is composed of two components: (1) UV prediction (our contribution) and (2) texture map with shading (not our contribution). For the garments with highly contrasted texture, the shading operation is often biased to color con" + }, + { + "type": "image", + "bbox": [ + 0.665, + 0.635, + 0.741, + 0.746 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.687, + 0.75, + 0.723, + 0.761 + ], + "angle": 0, + "content": "Image" + }, + { + "type": "image", + "bbox": [ + 0.742, + 0.635, + 0.814, + 0.744 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.742, + 0.744, + 0.807, + 0.768 + ], + "angle": 0, + "content": "Retexturing w/o shading" + }, + { + "type": "image", + "bbox": [ + 0.814, + 0.635, + 0.888, + 0.744 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.816, + 0.744, + 0.882, + 0.768 + ], + "angle": 0, + "content": "Retexturing w/shading" + }, + { + "type": "image_caption", + "bbox": [ + 0.692, + 0.77, + 0.891, + 0.796 + ], + "angle": 0, + "content": "Figure 10. Limitation. Limitations on textured garment." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.868 + ], + "angle": 0, + "content": "trast, resulting in erroneous appearance. Figure 10 illustrates that despite reasonable UV prediction from our method, the resulting appearance is unrealistic near the textured region. Improving the shading operation is beyond the scope of this work, and we leave it as future work." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.869, + 0.891, + 0.897 + ], + "angle": 0, + "content": "Acknowledgement This work was supported by a NSF NRI 2022894 and NSF CAREER 1846031." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4633" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.216, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.22, + 0.089, + 0.89, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.275, + 0.216, + 0.458 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.094, + 0.458, + 0.202, + 0.472 + ], + "angle": 0, + "content": "Original image" + }, + { + "type": "image_caption", + "bbox": [ + 0.221, + 0.459, + 0.89, + 0.472 + ], + "angle": 0, + "content": "Re-textured images" + }, + { + "type": "image_caption", + "bbox": [ + 0.355, + 0.476, + 0.618, + 0.488 + ], + "angle": 0, + "content": "(a) Qualitative results of our method on an image." + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.489, + 0.339, + 0.614 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.128, + 0.616, + 0.303, + 0.809 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.174, + 0.81, + 0.267, + 0.824 + ], + "angle": 0, + "content": "Original video" + }, + { + "type": "image", + "bbox": [ + 0.344, + 0.489, + 0.891, + 0.824 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.361, + 0.83, + 0.61, + 0.842 + ], + "angle": 0, + "content": "(b) Qualitative results of our method on videos." + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.854, + 0.893, + 0.882 + ], + "angle": 0, + "content": "Figure 11. Results on images and videos. Our method can produce compelling texture editing given images and videos that preserve geometric details." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4634" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.341, + 0.13 + ], + "angle": 0, + "content": "[1] https://renderpeople.com/3d-people.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.132, + 0.468, + 0.158 + ], + "angle": 0, + "content": "[2] Abien Fred Agarap. Deep learning using rectified linear units (relu). arXiv, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.161, + 0.468, + 0.215 + ], + "angle": 0, + "content": "[3] Badour AlBahar, Jingwan Lu, Jimei Yang, Zhixin Shu, Eli Shechtman, and Jia-Bin Huang. Pose with Style: Detail-preserving pose-guided image synthesis with conditional stylegan. TOG, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.217, + 0.468, + 0.257 + ], + "angle": 0, + "content": "[4] Yazeed Alharbi and Peter Wonka. Disentangled image generation through structured noise injection. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.26, + 0.468, + 0.313 + ], + "angle": 0, + "content": "[5] David Bau, Hendrik Strobelt, William Peebles, Jonas Wulff, Bolei Zhou, Jun-Yan Zhu, and Antonio Torralba. Semantic photo manipulation with a generative image prior. TOG, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.316, + 0.468, + 0.357 + ], + "angle": 0, + "content": "[6] Bharat Lal Bhatnagar, Garvita Tiwari, Christian Theobalt, and Gerard Pons-Moll. Multi-garment net: Learning to dress 3d people from images. In ICCV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.359, + 0.468, + 0.385 + ], + "angle": 0, + "content": "[7] Michael J. Black and P. Anandan. Robust dynamic motion estimation over time. In CVPR, 1991. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.388, + 0.468, + 0.415 + ], + "angle": 0, + "content": "[8] Gabriel J Brostow and Irfan A Essa. Motion based decompositing of video. In ICCV, 1999. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.418, + 0.468, + 0.457 + ], + "angle": 0, + "content": "[9] Felix E. Browder. On the unification of the calculus of variations and the theory of monotone nonlinear operators in banach spaces. PNAS, 1966. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.46, + 0.468, + 0.487 + ], + "angle": 0, + "content": "[10] Edwin Earl Catmull. A subdivision algorithm for computer display of curved surfaces. The University of Utah, 1974. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.489, + 0.468, + 0.53 + ], + "angle": 0, + "content": "[11] Shu-Yu Chen, Wanchao Su, Lin Gao, Shihong Xia, and Hongbo Fu. Deepfacedrawing: Deep generation of face images from sketches. TOG, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.532, + 0.468, + 0.572 + ], + "angle": 0, + "content": "[12] Anton Cherepkov, Andrey Voynov, and Artem Babenko. Navigating the gan parameter space for semantic image editing. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.575, + 0.468, + 0.614 + ], + "angle": 0, + "content": "[13] Edo Collins, Raja Bala, Bob Price, and Sabine Süssstrunk. Editing in style: Uncovering the local semantics of GANs. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.617, + 0.468, + 0.657 + ], + "angle": 0, + "content": "[14] T. Darrell and A. Pentland. Robust estimation of a multilayered motion representation. In Proceedings of the IEEE Workshop on Visual Motion, 1991. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.66, + 0.468, + 0.701 + ], + "angle": 0, + "content": "[15] Ke Gong, Yiming Gao, Xiaodan Liang, Xiaohui Shen, Meng Wang, and Liang Lin. Graphonomy: Universal human parsing via graph transfer learning. In CVPR, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.703, + 0.468, + 0.73 + ], + "angle": 0, + "content": "[16] J.C. Gower and G.B. Dijksterhuis. Procrustes problems. new york: Oxford university press. Psychometrika, 70, 2005. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.732, + 0.468, + 0.772 + ], + "angle": 0, + "content": "[17] Artur Grigorev, Artem Sevastopolsky, Alexander Vakhitov, and Victor Lempitsky. Coordinate-based texture inpainting for pose-guided human image generation. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.468, + 0.814 + ], + "angle": 0, + "content": "[18] Riza Alp Güler, Natalia Neverova, and Iasonas Kokkinos. Densesepos: Dense human pose estimation in the wild. In CVPR, 2018. 2, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.468, + 0.856 + ], + "angle": 0, + "content": "[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.468, + 0.9 + ], + "angle": 0, + "content": "[20] Zhenliang He, Wangmeng Zuo, Meina Kan, Shiguang Shan, and Xilin Chen. Attgan: Facial attribute editing by only changing what you want. TIP, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.468, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.133 + ], + "angle": 0, + "content": "[21] Xianxu Hou, Xiaokang Zhang, Linlin Shen, Zhihui Lai, and Jun Wan. Guidedstyle: Attribute knowledge guided style manipulation for semantic face editing. arXiv, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.175 + ], + "angle": 0, + "content": "[22] Anastasia Ianina, Nikolaos Sarafianos, Yuanlu Xu, Ignacio Rocco, and Tony Tung. Bodymap: Learning full-body dense correspondence map. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.217 + ], + "angle": 0, + "content": "[23] Yasamin Jafarian and Hyun Soo Park. Learning high fidelity depths of dressed humans by watching social media dance videos. In CVPR, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[24] Varun Jampani, Raghudeep Gadde, and Peter V. Gehler. Video propagation networks. In CVPR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.25, + 0.892, + 0.289 + ], + "angle": 0, + "content": "[25] Yoni Kasten, Dolev Ofri, Oliver Wang, and Tali Dekel. Layered neural atlases for consistent video editing. TOG, 2021. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.293, + 0.892, + 0.332 + ], + "angle": 0, + "content": "[26] Hyunsu Kim, Yunjey Choi, Junho Kim, Sungwoo Yoo, and Youngjung Uh. Exploiting spatial dimensions of latent in gan for real-time image editing. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.335, + 0.892, + 0.361 + ], + "angle": 0, + "content": "[27] Diederik P. Kingma and Jimmy Lei Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.364, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[28] Nilesh Kulkarni, Abhinav Gupta, David F. Fouhey, and Shubham Tulsiani. Articulation-aware canonical surface mapping. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.406, + 0.892, + 0.446 + ], + "angle": 0, + "content": "[29] Nilesh Kulkarni, Abhinav Gupta, David F Fouhey, and Shubham Tulsiani. Articulation-aware canonical surface mapping. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.449, + 0.892, + 0.487 + ], + "angle": 0, + "content": "[30] M. Pawan Kumar, Philip H. S. Torr, and Andrew Zisserman. Learning layered motion segmentations of video. IJCV, 2005. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.491, + 0.892, + 0.531 + ], + "angle": 0, + "content": "[31] Verica Lazova, Eldar Insafutdinov, and Gerard Pons-Moll. 360-degree textures of people in clothing from a single image. 3DV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.534, + 0.892, + 0.573 + ], + "angle": 0, + "content": "[32] Cheng-Han Lee, Ziwei Liu, Lingyun Wu, and Ping Luo. Maskgan: Towards diverse and interactive facial image manipulation. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.576, + 0.892, + 0.616 + ], + "angle": 0, + "content": "[33] Huan Ling, Karsten Kreis, Daiqing Li, Seung Wook Kim, Antonio Torralba, and Sanja Fidler. Editgan: High-precision semantic image editing. In NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.619, + 0.892, + 0.671 + ], + "angle": 0, + "content": "[34] Erika Lu, Forrester Cole, Tali Dekel, Weidi Xie, Andrew Zisserman, David Salesin, William T. Freeman, and Michael Rubinstein. Layered neural rendering for retiming people in video. TOG, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.675, + 0.892, + 0.714 + ], + "angle": 0, + "content": "[35] Bruce D. Lucas and Takeo Kanade. An iterative image registration technique with an application to stereo vision. In IJCAI, 1981. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.718, + 0.892, + 0.744 + ], + "angle": 0, + "content": "[36] Bruce D. Lucas and Takeo Kanade. Optical navigation by the method of differences. In *IJCAI*, 1985. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.747, + 0.892, + 0.773 + ], + "angle": 0, + "content": "[37] Tom McReynolds and David Blythe. Advanced graphics programming using OpenGL. Elsevier, 2005. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.776, + 0.892, + 0.814 + ], + "angle": 0, + "content": "[38] Aymen Mir, Thiemo Alldieck, and Gerard Pons-Moll. Learning to transfer texture from clothing images to 3d humans. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.818, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[39] Natalia Neverova, David Novotny, Vasil Khalidov, Marc Szafraniec, Patrick Labatut, and Andrea Vedaldi. Continuous surface embeddings. In NeurIPS, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.86, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[40] Natalia Neverova, David Novotny, and Andrea Vedaldi. Correlated uncertainty for learning dense correspondences from noisy labels. In NeurIPS, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4635" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.146 + ], + "angle": 0, + "content": "[41] Natalia Neverova, Artsiom Sanakoyeu, David Novotny, Patrick Labatut, and Andrea Vedaldi. Discovering relationships between object categories via universal canonical maps. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.147, + 0.47, + 0.2 + ], + "angle": 0, + "content": "[42] Natalia Neverova, James Thewlis, Riza Alp Güler, Iasonas Kokkinos, and Andrea Vedaldi. Slim densepose: Thrifty learning from sparse annotations and motion cues. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.202, + 0.469, + 0.31 + ], + "angle": 0, + "content": "[43] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In NeurIPS. 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.311, + 0.469, + 0.352 + ], + "angle": 0, + "content": "[44] Alex Rav-Acha, Pushmeet Kohli, Carsten Rother, and Andrew Fitzgibbon. Unwrap mosaics: A new representation for video editing. SIGGRAPH, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.353, + 0.469, + 0.405 + ], + "angle": 0, + "content": "[45] Igor Santesteban, Nils Thuerey, Miguel A Otaduy, and Dan Casas. Self-Supervised Collision Handling via Generative 3D Garment Models for Virtual Try-On. In CVPR, 2021. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.407, + 0.469, + 0.447 + ], + "angle": 0, + "content": "[46] Yujun Shen, Jinjin Gu, Xiaou Tang, and Bolei Zhou. Interpreting the latent space of gans for semantic face editing. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.449, + 0.469, + 0.515 + ], + "angle": 0, + "content": "[47] Feitong Tan, Danhang Tang, Dou Mingsong, Guo Kaiwen, Rohit Pandey, Cem Keskin, Ruofei Du, Deqing Sun, Sofien Bouaziz, Sean Fanello, Ping Tan, and Yinda Zhang. HumanGPS: Geodesic preserving feature for dense human correspondences. In CVPR, 2021. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.517, + 0.469, + 0.583 + ], + "angle": 0, + "content": "[48] Matthew Tancik, Pratul P. Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan T. Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. In NeurIPS, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.585, + 0.469, + 0.613 + ], + "angle": 0, + "content": "[49] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In ECCV, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.614, + 0.469, + 0.652 + ], + "angle": 0, + "content": "[50] Alexandru Telea. An image inpainting technique based on the fast marching method. Journal of Graphics Tools, 2004, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.654, + 0.469, + 0.682 + ], + "angle": 0, + "content": "[51] John Y. A. Wang and Edward H. Adelson. Representing moving images with layers. TIP, 1994. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.682, + 0.469, + 0.734 + ], + "angle": 0, + "content": "[52] Tuanfeng Wang, Duygu Ceylan, Krishna Kumar Singh, and Niloy J. Mitra. Dance in the wild: Monocular human animation with neural dynamic appearance synthesis. In 3DV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.736, + 0.469, + 0.763 + ], + "angle": 0, + "content": "[53] Yair Weiss. Deriving intrinsic images from image sequences. In ICCV, 2001. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.764, + 0.469, + 0.804 + ], + "angle": 0, + "content": "[54] Rongliang Wu, Gongjie Zhang, Shijian Lu, and Tao Chen. Cascade ef-gan: Progressive facial expression editing with local focuses. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.805, + 0.469, + 0.845 + ], + "angle": 0, + "content": "[55] You Xie, Huiqi Mao, Angela Yao, and Nils Thuerey. Temporaluv: Capturing loose clothing with temporally coherent uv coordinates. In CVPR, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.847, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[56] Haonan Yan, Jiaqi Chen, Xujie Zhang, Shengkai Zhang, Nianhong Jiao, Xiaodan Liang, and Tianxiang Zheng. Ultrapose: Synthesizing dense pose with 1 billion points by human-body decoupling 3d model. In ICCV, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.133 + ], + "angle": 0, + "content": "[57] Vickie Ye, Zhengqi Li, Richard Tucker, Angjoo Kanazawa, and Noah Snavelly. Deformable sprites for unsupervised video decomposition. In CVPR, 2022. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.134, + 0.892, + 0.175 + ], + "angle": 0, + "content": "[58] Zhixuan Yu, Haozheng Yu, Long Sha, Sujoy Ganguly, and Hyun Soo Park. Semi-supervised dense keypoints using unlabeled multiview images. arXiv, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.177, + 0.892, + 0.217 + ], + "angle": 0, + "content": "[59] Polina Zablotskaia, Aliaksandr Siarohin, Bo Zhao, and Leonid Sigal. Dwnet: Dense warp-based network for pose-guided human video generation. In BMVC, 2019. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.218, + 0.892, + 0.259 + ], + "angle": 0, + "content": "[60] Wang Zeng, Wanli Ouyang, Ping Luo, Wentao Liu, and Xiaogang Wang. 3d human mesh regression with dense correspondence. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.26, + 0.892, + 0.3 + ], + "angle": 0, + "content": "[61] Tyler Zhu, Per Karlsson, and Chris Bregler. Simpose: Effectively learning densepose and surface normal of people from simulated data. In ECCV, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4636" + } + ] +] \ No newline at end of file diff --git a/2023/Normal-Guided Garment UV Prediction for Human Re-Texturing/6ef67559-1f18-4ef9-b45a-a009b6d372b6_origin.pdf b/2023/Normal-Guided Garment UV Prediction for Human Re-Texturing/6ef67559-1f18-4ef9-b45a-a009b6d372b6_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f38c7266e7bec61c0b603bc6dea566cab2149dd2 --- /dev/null +++ b/2023/Normal-Guided Garment UV Prediction for Human Re-Texturing/6ef67559-1f18-4ef9-b45a-a009b6d372b6_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:04ccaf952b456384c97b268171285319c367036ee27c992adf6c89ee6c66210c +size 6080150 diff --git a/2023/Normal-Guided Garment UV Prediction for Human Re-Texturing/full.md b/2023/Normal-Guided Garment UV Prediction for Human Re-Texturing/full.md new file mode 100644 index 0000000000000000000000000000000000000000..0c158b7ae4c5371d457774bbae15e972e8f7dd34 --- /dev/null +++ b/2023/Normal-Guided Garment UV Prediction for Human Re-Texturing/full.md @@ -0,0 +1,388 @@ +# Normal-guided Garment UV Prediction for Human Re-texturing + +Yasamin Jafarian† Tuanfeng Y. Wang‡ Duygu Ceylan‡ Jimei Yang‡ Nathan Carr‡ Yi Zhou‡ Hyun Soo Park† †University of Minnesota #Adobe Research + +![](images/5c20f43b3f797941927ae0400a5b25835e05ae78844d68cc2cbecd8a974f899a.jpg) +Original video +Figure 1. Geometry aware texture editing from images and videos. This paper presents a novel approach to predict a geometry aware texture (UV) map of a garment from an image (first row) and video (second row). The predicted UV map preserves isometry between texture space and 3D surface by leveraging 3D surface normals. Further, we ensure temporal consistency in the predicted UV map across frames in a video, resulting in physically plausible human appearance editing. Project website: yasamin.page/normal-guided-uv. + +![](images/ca2d491b5c3cae86995bd2e6cf1a5fb4a0436fe19250b2ef1bfd02f890ee691b.jpg) +Re-textured video + +# Abstract + +Clothes undergo complex geometric deformations, which lead to appearance changes. To edit human videos in a physically plausible way, a texture map must take into account not only the garment transformation induced by the body movements and clothes fitting, but also its 3D fine-grained surface geometry. This poses, however, a new challenge of 3D reconstruction of dynamic clothes from an image or a video. In this paper, we show that it is possible to edit dressed human images and videos without 3D reconstruction. We estimate a geometry aware texture map between the garment region in an image and the texture space, a.k.a, UV map. Our UV map is designed to preserve isometry with respect to the underlying 3D surface by making use of the 3D surface normals predicted from the + +image. Our approach captures the underlying geometry of the garment in a self-supervised way, requiring no ground truth annotation of UV maps and can be readily extended to predict temporally coherent UV maps. We demonstrate that our method outperforms the state-of-the-art human UV map estimation approaches on both real and synthetic data. + +# 1. Introduction + +While browsing online clothing shops, have you ever wondered how the appearance of a dress of interest would look on you as if you were in a fitting room given your dress with a similar shape? A key technology to enable generating such visual experiences is photorealistic re-texturing—editing the texture of clothes in response to the subject's movement in the presented images or videos in a geomet- + +rically and temporally coherent way. Over the past few years, there has been a significant advancement in the image and video editing technologies [4-6, 11-13, 17, 20, 21, 26, 31-33, 38, 46, 54], such as inserting advertising logos on videos of moving cars or applying face makeup on social media. However, such editing approaches designed for rigid or semi-rigid surfaces are not suitable for garments that undergo complex secondary motion with respect to the underlying body. For example, the fine wrinkles of the dress in Figure 1 result in complex warps in texture over time. In this paper, we present a new method to edit the appearance of a garment in a given image or video by taking into account its fine-grained geometric deformation. + +Previous works address photorealistic texture editing in two ways. (1) 3D reconstruction and rendering: these approaches can achieve high-fidelity texture editing given highly accurate 3D geometry. On the other side of the coin, their performance is dictated by the quality of the 3D reconstruction. While the 3D geometry of the garment can be learned from paired human appearance data, e.g., human modeling repositories with 3D meshes and renderings [1], due to the scarcity of such data, it often cannot generalize well on unseen real images and videos. (2) Direct texture mapping: by estimating dense UV map, these methods can bypass the procedure of 3D reconstruction [18, 22, 39, 41, 55]. However, they usually lack of geometry details and only capture the underlying human body, thus, not applicable for editing garments. Moreover, when applied to videos, visual artifacts of editing become more salient since they are not aware of underlying deformation of the garment's 3D geometry [25, 57]. + +We design our method to enjoy the advantages of both two approaches: preserving realistic details in UV mapping while circumventing 3D reconstruction. Our key insight is that the fundamental geometric property of isometry can be imposed into UV map estimation via the 3D surface normals predicted from an image. We formulate a geometric relationship between the UV map and surface normals in the form of a set of partial differential equations. + +Our method takes as input an image or video, its surface normal prediction, and dense optical flow (for video), and outputs the geometry aware UV map estimate. The UV map is modeled by a multi-layer perceptron that can predict UV coordinates given a pixel location in an image. We note that the UV map is defined up to the choice of a reference coordinate frame. To disambiguate this, we condition the neural network with a pre-defined proxy UV map (e.g., DensePose [18]). We use the isometry constraints as a loss to optimize the UV map. Further, for a video, we leverage the per-frame image feature to correlate the UV coordinates of the pixels across time using optical flow. + +Our contributions can be concluded in three aspects: (1) a novel formulation that captures the geometric relationship between the 3D surface normals and the UV map by the + +isometry constraint, which eliminates the requirement of 3D reconstruction and ground truth UV map; (2) a neural network design that learns to predict temporally coherent UV map for the frames by correlating per-frame image features; (3) stronger performance compared to existing re-texturing methods and compelling results on a wide range of real-world imagery. + +# 2. Related Work + +Our work lies at the intersection of human UV map prediction from images and neural UV map optimization. + +# 2.1. Human Dense UV Map Estimation + +A seminal work of DensePose [18] learns to predict a UV map of humans presented in an image, which opens a new opportunity to edit the appearance of a person without 3D reconstruction [3,52]. A series of subsequent works [18,28,29,39-42,56,58,60] bring out a number of applications for human tracking. However, due to their representation specific to the body surface, they exhibit fundamental limitations in expressing highly deformable loose clothing such as skirts and dresses. + +To address this challenge, recent approaches leverage multitask learning [61] or incorporate geodesic distance to learn UV maps [47]. BodyMap [22] incorporates the Vision Transformers to learn per-pixel image features on a continuous body surface that handles loose clothes, different hairstyles, and occlusion. TemporalUV [55] focuses on handling garments by extrapolating the initial DensePose estimates [18] and leveraging image features obtained from an input video to obtain a UV aligned with the garment boundary. Despite their promise, the visual artifacts persist due to a lack of understanding the underlying 3D geometry. Unlike previous approaches, we design our framework such that the resulting UV map satisfies the fundamental geometric property of isometry, which results in physically plausible re-texturing. + +# 2.2. Neural UV Optimization from Videos. + +Another line of work [24, 25, 34, 44, 57] resorts to a layered UV map, capturing the geometry to some degree by incorporating video decomposition [7, 8, 14, 30, 51] to optimize the UV coordinates of the foreground and background based on the observed motion. Kasten et al. [25] unwrap a video into a set of layered 2D atlases where for each pixel in the video, its corresponding 2D coordinate in each of the atlases is predicted. Ye et al. [57] proposes a global sprite image that can group the distinct motion trajectories because the collective object structure has a consistent appearance throughout time. While preserving the temporal coherency and maintaining some coarse UV deformations related to arm or leg movements during a sequence, these methods fall short of capturing micro deformations like wrinkles in + +![](images/b77126e684d136db183fe79e5196927b37661ad0ec7a8c7b08792e87a983bd24.jpg) +Figure 2. Texture mapping geometry. We study the mapping between the image space, texture space, and 3D space. A point $\mathbf{x}$ in image is mapped to the texture space with $\mathbf{u} = g(\mathbf{x})$ . The mapping $f(\mathbf{u}) = \mathbf{X}$ lifts the texture plane into a 3D garment surface by an isometric warping. We use the orthographic projection model, resulting in the first two elements of $f$ is the inverse of $g$ , i.e., $f^{1:2} = g^{-1}$ . The spatial derivatives of $f$ form a tangent plane on the 3D surface at $\mathbf{X}$ , resulting in $\mathbf{N}_{\mathbf{x}} = f_{u} \times f_{v}$ where $\mathbf{N}_{\mathbf{x}}$ is the 3D surface normal, and $f_{u}$ and $f_{v}$ are the spatial derivatives of $f$ with respect to $u$ and $v$ . + +the clothing. Furthermore, these techniques cannot represent UV mapping for an image and can only be applied to videos. Unlike these methods, our surface normal conditioned UV map is highly sensitive to small geometric details and 3D surface deformations, which can be optimized not only for a video but also for single images. + +# 3. Method + +Our goal is to obtain a continuous texture mapping, which allows editing the appearance of dynamic garments. We leverage the geometric property of isometry to constrain the UV map in the form of partial differential equations. We solve this partial differential equations by optimizing a neural network to generate a geometry aware UV map. + +# 3.1. Texture Mapping without 3D Reconstruction + +Consider a mapping $g(\mathbf{x}) = \mathbf{u}$ that maps a pixel location $\mathbf{x} = (x, y) \in \mathbb{R}^2$ in the image space that belongs to a garment of interest to a point $\mathbf{u} = (u, v) \in \mathbb{R}^2$ in the UV space of the garment as shown in Figure 2. The goal of our work is to find such a mapping $g$ that takes into account the local surface geometry measured by the surface normal predicted at $\mathbf{x}$ . We denote the predicted 3D surface normal of $\mathbf{x}$ in the camera space as $\mathbf{N}_{\mathbf{x}} \in \mathbb{S}^2$ . + +Let us define an isometric map from the UV texture map to the 3D surface, $f(\mathbf{u}) = \mathbf{X}$ . This is the fundamental property of a non-stretchable cloth texture mapping [10]. + +$$ +\left\| f _ {u} \right\| = \left\| f _ {v} \right\| = 1, \quad f _ {u} ^ {\mathrm {T}} f _ {v} = 0, \tag {1} +$$ + +where $f_{u}$ and $f_{v}$ are the partial derivatives of $f$ with respect to $u$ and $v$ , respectively. Geometrically, $f_{u}$ and $f_{v}$ are the tangential vectors on the 3D surface where their cross product forms the surface normal: + +$$ +\tilde {\mathbf {N}} _ {\mathbf {x}} = f _ {u} \times f _ {v} = f _ {u} (g (\mathbf {x})) \times f _ {v} (g (\mathbf {x})) \tag {2} +$$ + +where $\widetilde{\mathbf{N}}_{\mathbf{x}} \in \mathbb{S}^2$ is the surface normal at $\mathbf{X}$ corresponding to $\mathbf{x}$ . + +We can find the UV mapping $g$ by matching the surface normal $\widetilde{\mathbf{N}}_{\mathbf{x}}$ derived by Equation (2) and the surface normal predicted from the image $\mathbf{N}_{\mathbf{x}}$ : + +$$ +\begin{array}{l} \underset {\boldsymbol {\theta} _ {g}, \boldsymbol {\theta} _ {f}} {\operatorname {m i n i m i z e}} \sum_ {\mathbf {x}} \| f _ {u} (g (\mathbf {x})) \times f _ {v} (g (\mathbf {x})) - \mathbf {N} _ {\mathbf {x}} \| ^ {2}, \\ s. t. \quad \| f _ {u} \| = \| f _ {v} \| = 1, f _ {u} ^ {\top} f _ {v} = 0, \tag {3} \\ \end{array} +$$ + +where $\theta_{g}$ and $\theta_{f}$ are the parameters of the function $g$ and $f$ , respectively. A key challenge of solving Equation (3) lies in the dependency of $f$ that requires full 3D reconstruction of the surface. Instead, we formulate a new dual problem that can solve Equation (3) effectively without finding $f$ . + +We use two properties to eliminate $f$ from Equation (3). First, we assume orthographic projection, i.e., $(x,y) = (X,Y)$ where $\mathbf{X} = \left[ \begin{array}{ccc}X & Y & Z \end{array} \right]^{\mathrm{T}}$ . This allows us to express the 3D derivatives using the pixel coordinates: + +$$ +g = \left(f ^ {1: 2}\right) ^ {- 1}, \tag {4} +$$ + +where $f^{1:2}$ is the first two elements $(X, Y)$ of $f$ . To keep $f^{1:2}$ bijective, we assume there is no self occlusion in the camera projection of $f$ . Note that $g^{-1}$ is the inverse of $g$ that maps the UV texture map to the pixel coordinate. Second, we derive the derivatives of $g$ by using the inverse function theorem [9]: + +$$ +f ^ {1: 2} \circ g (\mathbf {x}) = \mathbf {x} \rightarrow \mathbf {J} _ {g} = \left(\mathbf {J} _ {(f ^ {1: 2})}\right) ^ {- 1}, \tag {5} +$$ + +where $\mathbf{J}_g = \left[g_x\quad g_y\right] = \left[ \begin{array}{cc}\frac{\partial u}{\partial x} & \frac{\partial u}{\partial y}\\ \frac{\partial v}{\partial x} & \frac{\partial v}{\partial y} \end{array} \right]$ is the Jacobian matrix of the function $g$ + +With Equation (1) and (5), Equation (2) can be re-written as the following constraints by eliminating $f$ (3D reconstruction): + +$$ +\left\| g _ {x} \right\| = \sqrt {1 + \frac {\widetilde {n} _ {x} ^ {2}}{\widetilde {n} _ {z} ^ {2}}}, \left\| g _ {y} \right\| = \sqrt {1 + \frac {\widetilde {n} _ {y} ^ {2}}{\widetilde {n} _ {z} ^ {2}}}, g _ {x} ^ {\intercal} g _ {y} = \frac {\widetilde {n} _ {x} \widetilde {n} _ {y}}{\widetilde {n} _ {z} ^ {2}}, \tag {6} +$$ + +where $\widetilde{\mathbf{N}}_{\mathbf{x}} = \left[\widetilde{n}_{x}\quad \widetilde{n}_{y}\quad \widetilde{n}_{z}\right]^{\mathrm{T}}$ . For the derivation of Equation (6), see Supplementary Material. + +Equation (6) is a set of partial differential equations of $g$ that needs to match with the predicted surface normal $\mathbf{N}_{\mathbf{x}} =$ + +![](images/855dd2df351d699838881e994ff24750573122b799c8a4c013220ab617dcf353.jpg) +Figure 3. 2D simplification of UV map estimation. We illustrate the isometric relationship between (a) the surface and (c) the UV map. The length of the curve in XZ axis needs to be preserved when mapping to $g(x)$ , i.e., $|\mathrm{d}u| = |\mathrm{d}r| = \sqrt{\mathrm{d}x^2 + \mathrm{d}z^2}$ (arc length preservation). This relationship can be re-written as (b) a partial differential equation using Equation (6) in terms of its surface normal and the inverse of the spatial derivative of $g$ , i.e., $\left(\frac{\mathrm{d}u}{\mathrm{d}x}\right)^2 = 1 + \left(\frac{\mathrm{d}n_x}{\mathrm{d}n_z}\right)^2$ . + +$\left[ \begin{array}{lll}n_{x} & n_{y} & n_{z} \end{array} \right]^{\mathsf{T}}$ , leading to a loss function: + +$$ +\begin{array}{l} \mathcal {L} _ {\mathrm {g e o}} (\boldsymbol {\theta} _ {g}) = \sum_ {\mathbf {x}} \left(\left. \left(\frac {\partial u}{\partial x} \right| _ {\mathbf {x}}\right) ^ {2} + \left. \left(\frac {\partial v}{\partial x} \right| _ {\mathbf {x}}\right) ^ {2} - 1 - \frac {n _ {x} ^ {2}}{n _ {z} ^ {2}}\right) ^ {2} \\ + \left(\left. \left(\frac {\partial u}{\partial y} \right| _ {\mathbf {x}}\right) ^ {2} + \left. \left(\frac {\partial v}{\partial y} \right| _ {\mathbf {x}}\right) ^ {2} - 1 - \frac {n _ {y} ^ {2}}{n _ {z} ^ {2}}\right) ^ {2} \\ + \left(\frac {\partial u}{\partial x} \bigg | _ {\mathbf {x}} \frac {\partial u}{\partial y} \bigg | _ {\mathbf {x}} + \left. \frac {\partial v}{\partial x} \right| _ {\mathbf {x}} \left. \frac {\partial v}{\partial y} \right| _ {\mathbf {x}} - \frac {n _ {x} n _ {y}}{n _ {z} ^ {2}}\right) ^ {2}, \tag {7} \\ \end{array} +$$ + +where $\left.\frac{\partial u}{\partial x}\right|_{\mathbf{x}}$ is the partial derivative of $u$ with respect to $x$ evaluated at $\mathbf{x}$ . + +Figure 3 illustrates a 2D simplification of UV map estimation. A curve in XZ plane forms a isometric relation with the UV map, $|\mathrm{du}| = |\mathrm{dr}| = \sqrt{\mathrm{dx}^2 + \mathrm{dz}^2}$ . This relationship can be re-written as a partial differential equation in terms of the surface normal $(n_x, n_z)$ and the inverse of the spatial derivative of $g(x)$ using Equation (6), i.e., $\left(\frac{\mathrm{du}}{\mathrm{dx}}\right)^2 = 1 + \left(\frac{\mathrm{dn}_x}{\mathrm{dn}_z}\right)^2$ . We solve these partial differential equations to estimate $g$ . + +# 3.2. Self-supervised Learning of Texture Mapping + +The texture map $g$ is defined up to a bijective function, i.e., there exists an infinite number of $g$ that are equivalent: $g^{-1} \circ g = (\mathcal{T} \circ g)^{-1} \circ (\mathcal{T} \circ g)$ , where $\mathcal{T}$ is a bijective map (e.g., Euclidean transform). We resolve this ambiguity by + +finding $g$ such that $g \approx g'$ where $g'$ is a pre-defined proxy map of humans: + +$$ +\mathcal {L} _ {\mathrm {p r o x}} \left(\boldsymbol {\theta} _ {g}\right) = \sum_ {\mathbf {x}} \left\| g ^ {\prime} (\mathbf {x}) - g (\mathbf {x}) \right\| ^ {2}. \tag {8} +$$ + +In practice, we use an extended DensePose [18] as the pre-defined proxy map. Since DensePose makes predictions only for the human body, we apply an extrapolation method [50] to inpaint the garment regions that are not covered by DensePose. + +Further, we ensure physical plausibility of the visible 3D surfaces, i.e., the texture map should result in the surface normals pointing to $+Z$ direction, by adding the following loss: + +$$ +\mathcal {L} _ {\mathrm {z}} \left(\boldsymbol {\theta} _ {g}\right) = \sum_ {\mathbf {x}} \max (0, \det (\mathbf {J} _ {g} | _ {\mathbf {x}})), \tag {9} +$$ + +where $\operatorname{det}(\mathbf{J}_g)$ is the determinant of the Jacobian $\mathbf{J}_g$ that is equivalent to $\widetilde{n}_z$ . $\mathbf{J}_g|_{\mathbf{x}}$ is the Jacobian matrix of $g$ evaluated at $\mathbf{x}$ . See Supplementary Material for derivation. + +For a video, we extend the texture map to include the image feature for each pixel, i.e., $g(\mathbf{x}, \mathbf{f}_{\mathbf{x}})$ where $\mathbf{f}_{\mathbf{x}}$ is the image feature at $\mathbf{x}$ . This allows us to generalize the texture map over time. With the extension, we ensure the temporal consistency of the texture map by leveraging optical flow across frames: + +$$ +\mathcal {L} _ {\operatorname {t m p}} \left(\boldsymbol {\theta} _ {g}\right) = \sum_ {i, j} \sum_ {\mathbf {x} _ {i}} \left\| g \left(\mathbf {x} _ {i}, \mathbf {f} _ {\mathbf {x} _ {i}}\right) - g \left(\mathbf {x} _ {j}, \mathbf {f} _ {\mathbf {x} _ {j}}\right) \right\| ^ {2}, \tag {10} +$$ + +where $\mathbf{x}_i$ is a point in the $i^{\mathrm{th}}$ frame. This point is mapped to $\mathbf{x}_j$ in the $j^{\mathrm{th}}$ frame, i.e., $\mathbf{x}_j = W_{i\rightarrow j}(\mathbf{x}_i)$ where $W_{i\rightarrow j}$ is the optical flow from the $i^{\mathrm{th}}$ to $j^{\mathrm{th}}$ frames. + +Overall, we optimize the following loss to learn the texture map: + +$$ +\mathcal {L} \left(\boldsymbol {\theta} _ {g}\right) = \mathcal {L} _ {\text {g e o}} + \lambda_ {\text {p r o x}} \mathcal {L} _ {\text {p r o x}} + \lambda_ {\mathrm {z}} \mathcal {L} _ {\mathrm {z}} + \lambda_ {\text {t m p}} \mathcal {L} _ {\text {t m p}}, \tag {11} +$$ + +where $\lambda_{\mathrm{prox}}$ , $\lambda_{\mathrm{z}}$ , and $\lambda_{\mathrm{tmp}}$ are the weights that determine the relative importance of losses. Note that when a single image is used, $\lambda_{\mathrm{tmp}} = 0$ . + +# 3.3. Implementation Details + +We model $g$ using a 12-layer multi-layer perception, with ReLU [2] as an activation function after each layer that takes as input a pixel coordinate with positional encoding, $\gamma (\mathbf{x})$ where $\gamma :\mathbb{R}^2\to \mathbb{R}^{128}$ is Fourier based positional encoding [48]. For videos, we use ResNet [19] to extract per-frame 256 dimensional image feature $\mathbf{f}_{\mathbf{x}}$ . Our network design is illustrated in Figure 4 (for image-based UV map prediction) and 5 (for video-based UV map prediction). To make our prediction scale-invariant, we crop the garments region with $256\times 256$ resolution. We use an offthe-shelf garment segmentation software, Graphonomy [15] + +![](images/b36a786f68411016cf1750e89cd1e0a7277c39007ed4d489f150e882959aebbb.jpg) +Figure 4. Single image framework. For each pixel location on a garment $\mathbf{x}$ , we predict the UV coordinate $\mathbf{u}$ using a multi-layer perceptron. We enforce isometry to the UV map by matching with the predicted 3D surface normals. $\mathcal{L}_{\mathrm{z}}$ ensures that the predicted surface normals point toward the camera. To disambiguate the frame of reference of the UV map, we use pre-defined proxy map (e.g., DensePose [18] extrapolation [50]) by applying $\mathcal{L}_{\mathrm{prox}}$ . + +to separate the garment area. We use Adam optimizer [27] with batch size of 2048 and learning rate of $10^{-4}$ . We set $\lambda_{\mathrm{prox}} = 0.2$ , $\lambda_{\mathrm{z}} = 0.01$ , and $\lambda_{\mathrm{tmp}} = 0.3$ . We used an NVIDIA V100 GPU and Intel(R) Xeon(R) CPU E5-2698 v4 @ 2.20GHz, and implemented our approach with Pytorch [43]. Our method takes 20 minutes for a video of 82 frames while Kasten et al. [25], Ye et al. [57], and TemporalUV [55] take 10 hours, 30 minutes, and 23 hours, respectively. + +# 4. Evaluation + +We evaluate our method both quantitatively and qualitatively on real images as well as synthetic data with ground truth UV map. We also compare with the state-of-the-art methods on human UV map estimation methods and video UV map optimization approaches. + +Evaluation Datasets We evaluate our method using the following datasets: (1) five synthetic video sequences of simulated dress and T-shirt garments from Santesteban et al. [45] with random texture patterns over 700 frames; (2) ten real videos from Fashion Video dataset [59]; (3) TikTok dataset [23] and various YouTube videos as well as in-the-wild internet images. + +Evaluation Metric We use five metrics to evaluate our method. (1) UV error: for synthetic data, we report the absolute UV error in the texture space [45]. We use a Procrustes analysis [16] to align the resulting texture map to account for the diambiguity of the reference frame. We report the mean squared error in metric scale by assuming the height of the person in the input is $165\mathrm{cm}$ , resulting in $0.41\mathrm{cm} / \mathrm{UV}$ for dress and $0.27\mathrm{cm} / \mathrm{UV}$ for T-shirt (Table 1). (2) Average Precision percentage: we report the Average Precision (AP) percentage metric computed on all the pixels considering a per-pixel prediction as correct if the UV + +error is lower than a threshold. We visualize the AP metric for a range of thresholds from 1 to $15\mathrm{cm}$ and obtain the graph shown in Figure 7. + +(3) Photometric error: we warp the first frame of an input video to the rest of the frames using the UV map estimates. We report the error between the ground truth RGB images and the warped RGB images as reported in Table 1. +(4) Geometric error: we report the $\mathcal{L}_{\mathrm{geo}}$ to show how the predicted UV map follows the geometric information captured in the surface normal estimates as reported in Table 1. (5) Temporal error: we evaluate the capability of the different approaches in preserving the temporal coherency by reporting the $\mathcal{L}_{\mathrm{tmp}}$ error (Table 1). + +When the ground truth UV is not available, we use the geometric and temporal errors to evaluate our method. + +Baseline Methods We compare our method with previous works that fall into two categories: (1) human UV map prediction; (2) UV optimization. + +1) Human UV map prediction: we compare our method with state-of-the-art that focus on predicting UV maps for the naked human body [18] and dressed humans [47, 55]. We also report the performance of the UV map obtained by extrapolating DensePose predictions as discussed in Section 3.3. Our method achieves the best performance as shown in Table 1 and Figure 7. We notice that DensePose [18] performs competently in precision percentage when the threshold error is less than $7\mathrm{cm}$ . This observation is based on the fact that, for each pixel, DensePose predicts a part label (among 24 parts) and a UV map with respect to that body part. When aligning these predictions with the ground truth, we warp each occupied body part individually, resulting in a more accurate alignment compared to the other methods (including ours) that are represented by only one patch. However, the performance of DensePose [18] is not improved above $7\mathrm{cm}$ because of limited ability to predict beyond body surface. +2) UV optimization: we compare our method with state-of-the-art in predicting the UV map of a dynamic object observed in a video [25, 57]. However, such methods are not tailored for garments that undergo highly nonlinear transformations as the body moves. Hence, as reported in Table 1 and Figure 7, our method surpasses these baselines in the dense UV error, the average precision percentage, the geometric error, and the temporal error. +Ablation Study We conduct an ablation study to analyze the impact of the distance (first two terms in Equation (7)) and angle (third term in Equation (7)) constraints (second and third rows of Table 2 and Figure 7). Our final method has the best performance in UV error and photometric error. We also compare the performance of our method without the temporal consistency $(\mathcal{L}_{\mathrm{tmp}})$ (fourth row of Table 2). As expected, this term performs quite similarly to ours in UV error but very poorly in photometric error since the consistency between the frames is not enforced. The role of + +![](images/c4350c841f8b3e0633b3d05bf94b2b0772b730aaeea84c7f34c68fddcd3f082c.jpg) +Figure 5. Video framework. We apply temporal coherence in the predicted UV maps using optical flow [35,36,49]. With the optical flow $W_{i\rightarrow j}$ , we match the UV prediction of $i^{\mathrm{th}}$ frame with $j^{\mathrm{th}}$ via $\mathcal{L}_{\mathrm{tmp}}$ . + +
GT dress sequences [45]GT T-shirt sequences [45]Real Fashion sequences [59]
MethodUV. error (cm)photo. errorUV. error (cm)photo. errorgeo. error (Lgeo)tmp. error
DensePose [18]17.27±3.7652.43±10.207.48±0.5234.19±7.011.26±0.478.12±2.70
Extrapolated DensePose [18, 50]8.61±0.7618.29±3.245.34±0.6318.86±4.550.52±0.064.62±0.73
HumanGPS [47]11.97±2.0697.41±30.877.53±0.43108.10±34.891.27±0.6951.40±41.66
Kasten et al. [25]7.06±0.6013.07±2.796.64±0.6314.54±4.260.56±0.102.30±0.70
Ye et al. [57]5.56±0.2933.57±-10.275.75±0.2019.22±5.140.71±0.031.65±0.15
Ours3.16±0.287.54±2.043.58±0.2711.28±2.010.07±0.031.50±0.23
+ +Table 1. Quantitative Results. UV. error (cm), photo. error (RGB difference), geo. error $(\mathcal{L}_{\mathrm{geo}})$ , and tmp. error $(\mathcal{L}_{\mathrm{tmp}})$ (image space pixel distance) respectively (mean±std). + +
MethodUV. error (cm)photo. error (RGB)
Ours3.16±0.287.54±2.04
Distance constraint3.44±0.258.29±1.99
Angle constraint5.96±0.337.68±1.37
W/o Ltmp3.22±0.2414.67±4.82
W/o Lprox3.24±0.358.46±2.19
+ +![](images/dd8bdaaa327e79b9acf2d672920507703edd830b2347aacd8cc811fa011efeeb.jpg) +Figure 6. The impact of $\mathcal{L}_{\mathrm{prox}}$ in UV prediction and retexturing. + +![](images/b797576192909f839a67b3eb1be501094e380a744fb0365a41dbc0d5f7e635b1.jpg) +Figure 7. Average precision. We compute the average of pixels with a UV error higher than a threshold (1-15cm) on our method, the baseline methods, and our ablation experiments on ground truth dress sequences [45]. + +$\mathcal{L}_{\mathrm{prox}}$ is the disambiguation of UV maps, i.e., there exist an infinite number of equivalent UV maps that minimize our + +$\mathcal{L}_{\mathrm{geo}}$ (PDE). While the result without $\mathcal{L}_{\mathrm{prox}}$ are, therefore, quantitatively competitive as summarized in Table 2, such ambiguity can be resolved by finding UV that is closest to the proxy UV as shown in Figure 6, i.e., the retexture without $\mathcal{L}_{\mathrm{prox}}$ can result in arbitrary orientation across subjects. + +Qualitative Results To show our results qualitatively, we visualize both re-texturing examples and grid UV illustration to depict the performance of each method in preserving high-frequency details. For retexturing, we first obtain an albedo and shading layer from the input image using an intrinsic image decomposition method [53]. After applying a new texture pattern to the albedo layer, we compose it back with the original shading layer. We apply Gamma-correction [37] on the input image, after generating the albedo layer and shading layer, we inverse the Gamma-correction back when synthesizing the re-textured image. We compare our method qualitatively with the baselines as shown in Figure 8 that illustrates the results on Fashion video sequence [59]. Figure 11 shows the performance of our method on videos and images. Our method not only captures the fine-grained surface details but also is temporally coherent across time. + +User Study We conducted a user study: asking participants $(n = 21)$ to rate realism (1: unrealistic to 10: most realistic) for our method compared to the baselines as summarized in Table 3. Our method receives the highest score from the users. + +Table 2. Ablation study on dress sequences [45]. UV. error (cm), photo. error (RGB difference) (mean±std). + +
MethodUser score
DensePose4.09±1.94
Proxy DensePose3.23±1.86
HumanGPS2.28±2.12
Kasten et al.5.52±1.88
Ye et al.7.66±1.27
Ours9.57±0.81
+ +Table 3. User study. + +![](images/2be33698c31c3356af08f6606fee2bcdcd11f53e86467fa4d6248403c57e307c.jpg) +Input image + +![](images/66871bc101a26ecb8b3f4d6653c161726106cb1e107bda0a291c1eb9ee00c803.jpg) +Ours w/shading + +![](images/d581e1ceaee474d287393955d80892872e26bee9360bdaa30ebf2ab98b1b5ee5.jpg) + +![](images/f0c59451cfbdcc3a274f87dfa7ac1c8bfcd3a89e7cd745ab5158e717513a4d6c.jpg) + +![](images/fc0e6518f4d8a9998b027713f4c5ab93506f51601dd2ba617b250824658b93cd.jpg) + +![](images/e2e21bd3012fc32ac4a0e6975f2a732e9901c2fa48cb1629b7dd1b45d6b3ffba.jpg) + +![](images/6c0779eccc17662bfbbe9b586320103499e2652b4bc3b2a932791d8815d1514f.jpg) + +![](images/be2b0b0ca24b073d82d227c6d888f92ed75effcf4792672b89b5ab2959ba482f.jpg) + +![](images/d008d2f38e6805cbc8eb593a5f61d2f85815d4f97cc02fad91d5c3408faaa0c4.jpg) + +![](images/4c39a9ca66fc27ffe3af2a28983ef9b2929a40ecae1af04062e5fc8ba26c5bc0.jpg) + +![](images/31e635c011324386602f02f02738f49b9170fc9263875e6da0a5d0c72addafe4.jpg) + +![](images/480086e898fe57b0e7973d780db1ed696bac5753de7f5652f55c4072b874b632.jpg) + +![](images/7770163965a7cd33562533ad2c386eb95d1be5579eda984eac87a87b1eeaf0a5.jpg) + +![](images/c7e489fc13df2fd43159c6b68d6d5499c7c239818614422c76276846440c6232.jpg) + +![](images/45c683ea33a5861ea29c689615bebe4058a8b6a8947ea501827d7a6d1ab4eeb4.jpg) + +![](images/ebd15b28592d60249c8672f975e629a67e09967e203f1a94ef7fa2ae1957f242.jpg) + +![](images/b259f90d519ef78375436ce17d19d3173236c052786c59e33411daebc71f853f.jpg) + +![](images/886860ca57c58c90863285f2d7d3a439d83ca6eb8e7adb35d0e5c42fd58eef4d.jpg) +Retturung w/o shading +Grid-UV + +![](images/6aa03086274b31b936d5a616bcab1a8c23dda42eda1a830ecc0d5c3c989134f8.jpg) +High +Ours + +![](images/252a6aee4920ca0aed033b349cc7d82275495fbd6ddd63209881e0e9da6a85fe.jpg) +Extrapolated DensePose + +![](images/ec66ecb3b0f900893d82e8cbae0b2513d8b519fe2f2a0187dc490b0157ee9ac9.jpg) +DensePose + +![](images/3f1dce994bbdac0e06ee8e167c6af5d7aeea9107c369776e4597a42534143776.jpg) +HumanGPS + +![](images/52c4afc0f43bc7319342125e37437841ef402d7c1f15ca60ed425f231836d7b9.jpg) +TemporalUV + +![](images/d3ecf278b26fc169e62faa97d61bd9865174cf75556f3aea47de03ae9fb71706.jpg) +Kasten et al. + +![](images/fba2ca528dfd684478828c0914d5bff2aca6ab427293ed6e6b7c066618b59c0c.jpg) +Ye et al. + +![](images/e4fd2d54886b0f66d565fce9ee3e2eb272487b2d460c1cfe29bf7ba25a57fcb0.jpg) +Figure 8. Qualitative comparison. We compare ours with our initial UV (extrapolated DensePose), DensePose [18], HumanGPS [47], Kastenn et al. [25], and Ye et al. [57] in image re-texturing and UV map grid visualization. +Figure 9. Limitation. Due to lack of 3D reconstructed surface, our method cannot handle folds that introduce texture discontinuity. Our texture map is continuous around the folds, which is physically incorrect. + +# 5. Discussion + +This paper presents a novel approach to predict a high quality UV map by preserving geometric details from images and videos. We leverage the geometric property of isometry encoded in 3D surface normals to optimize the UV map in the form of partial differential equations. We generalize our method to videos by integrating optical flow, resulting in a temporally coherent video editing. Our method produces strong qualitative and quantitative predictions on real-world imagery compared to state-of-the-art UV map estimation. + +Limitation As discussed in Section 3.1, our method makes + +an assumption about projection, i.e., there is one-to-one correspondence between 3D surface geometry and image. However, this assumption does not hold when there is a fold where a region of 3D surface is not visible to the image. This makes a contrast with 3D reconstruction based method where the invisible part of 3D surface can be still mapped to the image via depth reasoning. Figure 9 illustrates this limitation where there are folds in the skirt, resulting in negative surface normal $n_z$ . Due to the folds, the texture must be discontinuous while our method produces continuous texture rendering due to the lack of 3D reconstructed geometry, which is physically incorrect. + +Our pipeline is composed of two components: (1) UV prediction (our contribution) and (2) texture map with shading (not our contribution). For the garments with highly contrasted texture, the shading operation is often biased to color con + +![](images/159478e428bec14296451264dcd57ed3cbad387285f19a63561be588e0284499.jpg) +Image +Figure 10. Limitation. Limitations on textured garment. + +![](images/4802e97fdfc756fa95ab8753afcafd6f5a35e50bf28dc2ae7ff181112f5568e9.jpg) +Retexturing w/o shading + +![](images/2438e0cbecb0eb5827c2d9afeef9573c31459cc756ff01e6c3888cf68534bcf2.jpg) +Retexturing w/shading + +trast, resulting in erroneous appearance. Figure 10 illustrates that despite reasonable UV prediction from our method, the resulting appearance is unrealistic near the textured region. Improving the shading operation is beyond the scope of this work, and we leave it as future work. + +Acknowledgement This work was supported by a NSF NRI 2022894 and NSF CAREER 1846031. + +![](images/cbde4900b39fdffb54508ce11134874a813903016d4d62681fbe926833fe5b8a.jpg) + +![](images/a307d949748ced08d0937bc708ca36f9ab6ff6ce04435e747fecf83350b8ee65.jpg) +Re-textured images + +![](images/59721a2c8d4eb4d13fcf13229abb841d62e65fb4e76329bfa4f28e6df9dc7f46.jpg) +Original image + +![](images/0a6c659cd622c567deeba997a3a30f4966357ec89a395d236c3ad9da5b82ddd3.jpg) + +![](images/c5317c17f98d04683b39a8dfa7cca28c755a9f7206c2c4f413cb7054b192a6cc.jpg) +Original video + +![](images/ddac916919564fdf45bfa6f928c0f0a2e0ef28c8ee9e740378d233f12a9ec5d0.jpg) +(a) Qualitative results of our method on an image. +(b) Qualitative results of our method on videos. +Figure 11. Results on images and videos. Our method can produce compelling texture editing given images and videos that preserve geometric details. + +# References + +[1] https://renderpeople.com/3d-people.2 +[2] Abien Fred Agarap. Deep learning using rectified linear units (relu). arXiv, 2018. 4 +[3] Badour AlBahar, Jingwan Lu, Jimei Yang, Zhixin Shu, Eli Shechtman, and Jia-Bin Huang. Pose with Style: Detail-preserving pose-guided image synthesis with conditional stylegan. TOG, 2021. 2 +[4] Yazeed Alharbi and Peter Wonka. Disentangled image generation through structured noise injection. In CVPR, 2020. 2 +[5] David Bau, Hendrik Strobelt, William Peebles, Jonas Wulff, Bolei Zhou, Jun-Yan Zhu, and Antonio Torralba. Semantic photo manipulation with a generative image prior. TOG, 2019. 2 +[6] Bharat Lal Bhatnagar, Garvita Tiwari, Christian Theobalt, and Gerard Pons-Moll. Multi-garment net: Learning to dress 3d people from images. In ICCV, 2019. 2 +[7] Michael J. Black and P. Anandan. Robust dynamic motion estimation over time. In CVPR, 1991. 2 +[8] Gabriel J Brostow and Irfan A Essa. Motion based decompositing of video. In ICCV, 1999. 2 +[9] Felix E. Browder. On the unification of the calculus of variations and the theory of monotone nonlinear operators in banach spaces. PNAS, 1966. 3 +[10] Edwin Earl Catmull. A subdivision algorithm for computer display of curved surfaces. The University of Utah, 1974. 3 +[11] Shu-Yu Chen, Wanchao Su, Lin Gao, Shihong Xia, and Hongbo Fu. Deepfacedrawing: Deep generation of face images from sketches. TOG, 2020. 2 +[12] Anton Cherepkov, Andrey Voynov, and Artem Babenko. Navigating the gan parameter space for semantic image editing. In CVPR, 2021. 2 +[13] Edo Collins, Raja Bala, Bob Price, and Sabine Süssstrunk. Editing in style: Uncovering the local semantics of GANs. In CVPR, 2020. 2 +[14] T. Darrell and A. Pentland. Robust estimation of a multilayered motion representation. In Proceedings of the IEEE Workshop on Visual Motion, 1991. 2 +[15] Ke Gong, Yiming Gao, Xiaodan Liang, Xiaohui Shen, Meng Wang, and Liang Lin. Graphonomy: Universal human parsing via graph transfer learning. In CVPR, 2019. 4 +[16] J.C. Gower and G.B. Dijksterhuis. Procrustes problems. new york: Oxford university press. Psychometrika, 70, 2005. 5 +[17] Artur Grigorev, Artem Sevastopolsky, Alexander Vakhitov, and Victor Lempitsky. Coordinate-based texture inpainting for pose-guided human image generation. In CVPR, 2019. 2 +[18] Riza Alp Güler, Natalia Neverova, and Iasonas Kokkinos. Densesepos: Dense human pose estimation in the wild. In CVPR, 2018. 2, 4, 5, 6, 7 +[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 4 +[20] Zhenliang He, Wangmeng Zuo, Meina Kan, Shiguang Shan, and Xilin Chen. Attgan: Facial attribute editing by only changing what you want. TIP, 2019. 2 + +[21] Xianxu Hou, Xiaokang Zhang, Linlin Shen, Zhihui Lai, and Jun Wan. Guidedstyle: Attribute knowledge guided style manipulation for semantic face editing. arXiv, 2020. 2 +[22] Anastasia Ianina, Nikolaos Sarafianos, Yuanlu Xu, Ignacio Rocco, and Tony Tung. Bodymap: Learning full-body dense correspondence map. In CVPR, 2022. 2 +[23] Yasamin Jafarian and Hyun Soo Park. Learning high fidelity depths of dressed humans by watching social media dance videos. In CVPR, 2021. 5 +[24] Varun Jampani, Raghudeep Gadde, and Peter V. Gehler. Video propagation networks. In CVPR, 2017. 2 +[25] Yoni Kasten, Dolev Ofri, Oliver Wang, and Tali Dekel. Layered neural atlases for consistent video editing. TOG, 2021. 2, 5, 6, 7 +[26] Hyunsu Kim, Yunjey Choi, Junho Kim, Sungwoo Yoo, and Youngjung Uh. Exploiting spatial dimensions of latent in gan for real-time image editing. In CVPR, 2021. 2 +[27] Diederik P. Kingma and Jimmy Lei Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 5 +[28] Nilesh Kulkarni, Abhinav Gupta, David F. Fouhey, and Shubham Tulsiani. Articulation-aware canonical surface mapping. In CVPR, 2020. 2 +[29] Nilesh Kulkarni, Abhinav Gupta, David F Fouhey, and Shubham Tulsiani. Articulation-aware canonical surface mapping. In CVPR, 2020. 2 +[30] M. Pawan Kumar, Philip H. S. Torr, and Andrew Zisserman. Learning layered motion segmentations of video. IJCV, 2005. 2 +[31] Verica Lazova, Eldar Insafutdinov, and Gerard Pons-Moll. 360-degree textures of people in clothing from a single image. 3DV, 2019. 2 +[32] Cheng-Han Lee, Ziwei Liu, Lingyun Wu, and Ping Luo. Maskgan: Towards diverse and interactive facial image manipulation. In CVPR, 2020. 2 +[33] Huan Ling, Karsten Kreis, Daiqing Li, Seung Wook Kim, Antonio Torralba, and Sanja Fidler. Editgan: High-precision semantic image editing. In NeurIPS, 2021. 2 +[34] Erika Lu, Forrester Cole, Tali Dekel, Weidi Xie, Andrew Zisserman, David Salesin, William T. Freeman, and Michael Rubinstein. Layered neural rendering for retiming people in video. TOG, 2020. 2 +[35] Bruce D. Lucas and Takeo Kanade. An iterative image registration technique with an application to stereo vision. In IJCAI, 1981. 6 +[36] Bruce D. Lucas and Takeo Kanade. Optical navigation by the method of differences. In *IJCAI*, 1985. 6 +[37] Tom McReynolds and David Blythe. Advanced graphics programming using OpenGL. Elsevier, 2005. 6 +[38] Aymen Mir, Thiemo Alldieck, and Gerard Pons-Moll. Learning to transfer texture from clothing images to 3d humans. In CVPR, 2020. 2 +[39] Natalia Neverova, David Novotny, Vasil Khalidov, Marc Szafraniec, Patrick Labatut, and Andrea Vedaldi. Continuous surface embeddings. In NeurIPS, 2020. 2 +[40] Natalia Neverova, David Novotny, and Andrea Vedaldi. Correlated uncertainty for learning dense correspondences from noisy labels. In NeurIPS, 2019. 2 + +[41] Natalia Neverova, Artsiom Sanakoyeu, David Novotny, Patrick Labatut, and Andrea Vedaldi. Discovering relationships between object categories via universal canonical maps. In CVPR, 2021. 2 +[42] Natalia Neverova, James Thewlis, Riza Alp Güler, Iasonas Kokkinos, and Andrea Vedaldi. Slim densepose: Thrifty learning from sparse annotations and motion cues. In CVPR, 2019. 2 +[43] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In NeurIPS. 2019. 5 +[44] Alex Rav-Acha, Pushmeet Kohli, Carsten Rother, and Andrew Fitzgibbon. Unwrap mosaics: A new representation for video editing. SIGGRAPH, 2008. 2 +[45] Igor Santesteban, Nils Thuerey, Miguel A Otaduy, and Dan Casas. Self-Supervised Collision Handling via Generative 3D Garment Models for Virtual Try-On. In CVPR, 2021. 5, 6 +[46] Yujun Shen, Jinjin Gu, Xiaou Tang, and Bolei Zhou. Interpreting the latent space of gans for semantic face editing. In CVPR, 2020. 2 +[47] Feitong Tan, Danhang Tang, Dou Mingsong, Guo Kaiwen, Rohit Pandey, Cem Keskin, Ruofei Du, Deqing Sun, Sofien Bouaziz, Sean Fanello, Ping Tan, and Yinda Zhang. HumanGPS: Geodesic preserving feature for dense human correspondences. In CVPR, 2021. 2, 5, 6, 7 +[48] Matthew Tancik, Pratul P. Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan T. Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. In NeurIPS, 2020. 4 +[49] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In ECCV, 2020. 6 +[50] Alexandru Telea. An image inpainting technique based on the fast marching method. Journal of Graphics Tools, 2004, 4, 5, 6 +[51] John Y. A. Wang and Edward H. Adelson. Representing moving images with layers. TIP, 1994. 2 +[52] Tuanfeng Wang, Duygu Ceylan, Krishna Kumar Singh, and Niloy J. Mitra. Dance in the wild: Monocular human animation with neural dynamic appearance synthesis. In 3DV, 2021. 2 +[53] Yair Weiss. Deriving intrinsic images from image sequences. In ICCV, 2001. 6 +[54] Rongliang Wu, Gongjie Zhang, Shijian Lu, and Tao Chen. Cascade ef-gan: Progressive facial expression editing with local focuses. In CVPR, 2020. 2 +[55] You Xie, Huiqi Mao, Angela Yao, and Nils Thuerey. Temporaluv: Capturing loose clothing with temporally coherent uv coordinates. In CVPR, 2022. 2, 5 +[56] Haonan Yan, Jiaqi Chen, Xujie Zhang, Shengkai Zhang, Nianhong Jiao, Xiaodan Liang, and Tianxiang Zheng. Ultrapose: Synthesizing dense pose with 1 billion points by human-body decoupling 3d model. In ICCV, 2021. 2 + +[57] Vickie Ye, Zhengqi Li, Richard Tucker, Angjoo Kanazawa, and Noah Snavelly. Deformable sprites for unsupervised video decomposition. In CVPR, 2022. 2, 5, 6, 7 +[58] Zhixuan Yu, Haozheng Yu, Long Sha, Sujoy Ganguly, and Hyun Soo Park. Semi-supervised dense keypoints using unlabeled multiview images. arXiv, 2021. 2 +[59] Polina Zablotskaia, Aliaksandr Siarohin, Bo Zhao, and Leonid Sigal. Dwnet: Dense warp-based network for pose-guided human video generation. In BMVC, 2019. 5, 6 +[60] Wang Zeng, Wanli Ouyang, Ping Luo, Wentao Liu, and Xiaogang Wang. 3d human mesh regression with dense correspondence. In CVPR, 2020. 2 +[61] Tyler Zhu, Per Karlsson, and Chris Bregler. Simpose: Effectively learning densepose and surface normal of people from simulated data. In ECCV, 2020. 2 \ No newline at end of file diff --git a/2023/Normal-Guided Garment UV Prediction for Human Re-Texturing/images.zip b/2023/Normal-Guided Garment UV Prediction for Human Re-Texturing/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7b2d9929afdce18c3846e26c2060e7d6d5e19904 --- /dev/null +++ b/2023/Normal-Guided Garment UV Prediction for Human Re-Texturing/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:433f04ff6ae309682638ff6d847250c13adc353b78eb6542d9ee34b77b158da0 +size 1178982 diff --git a/2023/Normal-Guided Garment UV Prediction for Human Re-Texturing/layout.json b/2023/Normal-Guided Garment UV Prediction for Human Re-Texturing/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3476b61171eca9ab14037dde564f661379e64871 --- /dev/null +++ b/2023/Normal-Guided Garment UV Prediction for Human Re-Texturing/layout.json @@ -0,0 +1,11442 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 96, + 103, + 500, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 103, + 500, + 121 + ], + "spans": [ + { + "bbox": [ + 96, + 103, + 500, + 121 + ], + "type": "text", + "content": "Normal-guided Garment UV Prediction for Human Re-texturing" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 113, + 141, + 477, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 141, + 477, + 196 + ], + "spans": [ + { + "bbox": [ + 113, + 141, + 477, + 196 + ], + "type": "text", + "content": "Yasamin Jafarian† Tuanfeng Y. Wang‡ Duygu Ceylan‡ Jimei Yang‡ Nathan Carr‡ Yi Zhou‡ Hyun Soo Park† †University of Minnesota #Adobe Research" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 48, + 205, + 303, + 463 + ], + "blocks": [ + { + "bbox": [ + 48, + 205, + 303, + 463 + ], + "lines": [ + { + "bbox": [ + 48, + 205, + 303, + 463 + ], + "spans": [ + { + "bbox": [ + 48, + 205, + 303, + 463 + ], + "type": "image", + "image_path": "5c20f43b3f797941927ae0400a5b25835e05ae78844d68cc2cbecd8a974f899a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 133, + 464, + 189, + 474 + ], + "lines": [ + { + "bbox": [ + 133, + 464, + 189, + 474 + ], + "spans": [ + { + "bbox": [ + 133, + 464, + 189, + 474 + ], + "type": "text", + "content": "Original video" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 45, + 476, + 547, + 521 + ], + "lines": [ + { + "bbox": [ + 45, + 476, + 547, + 521 + ], + "spans": [ + { + "bbox": [ + 45, + 476, + 547, + 521 + ], + "type": "text", + "content": "Figure 1. Geometry aware texture editing from images and videos. This paper presents a novel approach to predict a geometry aware texture (UV) map of a garment from an image (first row) and video (second row). The predicted UV map preserves isometry between texture space and 3D surface by leveraging 3D surface normals. Further, we ensure temporal consistency in the predicted UV map across frames in a video, resulting in physically plausible human appearance editing. Project website: yasamin.page/normal-guided-uv." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 307, + 205, + 545, + 463 + ], + "blocks": [ + { + "bbox": [ + 307, + 205, + 545, + 463 + ], + "lines": [ + { + "bbox": [ + 307, + 205, + 545, + 463 + ], + "spans": [ + { + "bbox": [ + 307, + 205, + 545, + 463 + ], + "type": "image", + "image_path": "ca2d491b5c3cae86995bd2e6cf1a5fb4a0436fe19250b2ef1bfd02f890ee691b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 402, + 464, + 470, + 473 + ], + "lines": [ + { + "bbox": [ + 402, + 464, + 470, + 473 + ], + "spans": [ + { + "bbox": [ + 402, + 464, + 470, + 473 + ], + "type": "text", + "content": "Re-textured video" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 530, + 192, + 542 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 530, + 192, + 542 + ], + "spans": [ + { + "bbox": [ + 143, + 530, + 192, + 542 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 549, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 549, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 549, + 289, + 715 + ], + "type": "text", + "content": "Clothes undergo complex geometric deformations, which lead to appearance changes. To edit human videos in a physically plausible way, a texture map must take into account not only the garment transformation induced by the body movements and clothes fitting, but also its 3D fine-grained surface geometry. This poses, however, a new challenge of 3D reconstruction of dynamic clothes from an image or a video. In this paper, we show that it is possible to edit dressed human images and videos without 3D reconstruction. We estimate a geometry aware texture map between the garment region in an image and the texture space, a.k.a, UV map. Our UV map is designed to preserve isometry with respect to the underlying 3D surface by making use of the 3D surface normals predicted from the" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 531, + 547, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 531, + 547, + 602 + ], + "spans": [ + { + "bbox": [ + 304, + 531, + 547, + 602 + ], + "type": "text", + "content": "image. Our approach captures the underlying geometry of the garment in a self-supervised way, requiring no ground truth annotation of UV maps and can be readily extended to predict temporally coherent UV maps. We demonstrate that our method outperforms the state-of-the-art human UV map estimation approaches on both real and synthetic data." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 611, + 387, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 611, + 387, + 624 + ], + "spans": [ + { + "bbox": [ + 306, + 611, + 387, + 624 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 631, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 631, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 631, + 547, + 715 + ], + "type": "text", + "content": "While browsing online clothing shops, have you ever wondered how the appearance of a dress of interest would look on you as if you were in a fitting room given your dress with a similar shape? A key technology to enable generating such visual experiences is photorealistic re-texturing—editing the texture of clothes in response to the subject's movement in the presented images or videos in a geomet-" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4627" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 224 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 224 + ], + "type": "text", + "content": "rically and temporally coherent way. Over the past few years, there has been a significant advancement in the image and video editing technologies [4-6, 11-13, 17, 20, 21, 26, 31-33, 38, 46, 54], such as inserting advertising logos on videos of moving cars or applying face makeup on social media. However, such editing approaches designed for rigid or semi-rigid surfaces are not suitable for garments that undergo complex secondary motion with respect to the underlying body. For example, the fine wrinkles of the dress in Figure 1 result in complex warps in texture over time. In this paper, we present a new method to edit the appearance of a garment in a given image or video by taking into account its fine-grained geometric deformation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 226, + 289, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 226, + 289, + 438 + ], + "spans": [ + { + "bbox": [ + 46, + 226, + 289, + 438 + ], + "type": "text", + "content": "Previous works address photorealistic texture editing in two ways. (1) 3D reconstruction and rendering: these approaches can achieve high-fidelity texture editing given highly accurate 3D geometry. On the other side of the coin, their performance is dictated by the quality of the 3D reconstruction. While the 3D geometry of the garment can be learned from paired human appearance data, e.g., human modeling repositories with 3D meshes and renderings [1], due to the scarcity of such data, it often cannot generalize well on unseen real images and videos. (2) Direct texture mapping: by estimating dense UV map, these methods can bypass the procedure of 3D reconstruction [18, 22, 39, 41, 55]. However, they usually lack of geometry details and only capture the underlying human body, thus, not applicable for editing garments. Moreover, when applied to videos, visual artifacts of editing become more salient since they are not aware of underlying deformation of the garment's 3D geometry [25, 57]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 440, + 289, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 440, + 289, + 534 + ], + "spans": [ + { + "bbox": [ + 46, + 440, + 289, + 534 + ], + "type": "text", + "content": "We design our method to enjoy the advantages of both two approaches: preserving realistic details in UV mapping while circumventing 3D reconstruction. Our key insight is that the fundamental geometric property of isometry can be imposed into UV map estimation via the 3D surface normals predicted from an image. We formulate a geometric relationship between the UV map and surface normals in the form of a set of partial differential equations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 536, + 289, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 536, + 289, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 536, + 289, + 677 + ], + "type": "text", + "content": "Our method takes as input an image or video, its surface normal prediction, and dense optical flow (for video), and outputs the geometry aware UV map estimate. The UV map is modeled by a multi-layer perceptron that can predict UV coordinates given a pixel location in an image. We note that the UV map is defined up to the choice of a reference coordinate frame. To disambiguate this, we condition the neural network with a pre-defined proxy UV map (e.g., DensePose [18]). We use the isometry constraints as a loss to optimize the UV map. Further, for a video, we leverage the per-frame image feature to correlate the UV coordinates of the pixels across time using optical flow." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 678, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 678, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 678, + 289, + 715 + ], + "type": "text", + "content": "Our contributions can be concluded in three aspects: (1) a novel formulation that captures the geometric relationship between the 3D surface normals and the UV map by the" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "type": "text", + "content": "isometry constraint, which eliminates the requirement of 3D reconstruction and ground truth UV map; (2) a neural network design that learns to predict temporally coherent UV map for the frames by correlating per-frame image features; (3) stronger performance compared to existing re-texturing methods and compelling results on a wide range of real-world imagery." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 167, + 393, + 179 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 167, + 393, + 179 + ], + "spans": [ + { + "bbox": [ + 306, + 167, + 393, + 179 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 187, + 545, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 187, + 545, + 212 + ], + "spans": [ + { + "bbox": [ + 305, + 187, + 545, + 212 + ], + "type": "text", + "content": "Our work lies at the intersection of human UV map prediction from images and neural UV map optimization." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 220, + 493, + 233 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 220, + 493, + 233 + ], + "spans": [ + { + "bbox": [ + 306, + 220, + 493, + 233 + ], + "type": "text", + "content": "2.1. Human Dense UV Map Estimation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 239, + 547, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 239, + 547, + 344 + ], + "spans": [ + { + "bbox": [ + 304, + 239, + 547, + 344 + ], + "type": "text", + "content": "A seminal work of DensePose [18] learns to predict a UV map of humans presented in an image, which opens a new opportunity to edit the appearance of a person without 3D reconstruction [3,52]. A series of subsequent works [18,28,29,39-42,56,58,60] bring out a number of applications for human tracking. However, due to their representation specific to the body surface, they exhibit fundamental limitations in expressing highly deformable loose clothing such as skirts and dresses." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 345, + 547, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 345, + 547, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 345, + 547, + 521 + ], + "type": "text", + "content": "To address this challenge, recent approaches leverage multitask learning [61] or incorporate geodesic distance to learn UV maps [47]. BodyMap [22] incorporates the Vision Transformers to learn per-pixel image features on a continuous body surface that handles loose clothes, different hairstyles, and occlusion. TemporalUV [55] focuses on handling garments by extrapolating the initial DensePose estimates [18] and leveraging image features obtained from an input video to obtain a UV aligned with the garment boundary. Despite their promise, the visual artifacts persist due to a lack of understanding the underlying 3D geometry. Unlike previous approaches, we design our framework such that the resulting UV map satisfies the fundamental geometric property of isometry, which results in physically plausible re-texturing." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 530, + 506, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 530, + 506, + 544 + ], + "spans": [ + { + "bbox": [ + 306, + 530, + 506, + 544 + ], + "type": "text", + "content": "2.2. Neural UV Optimization from Videos." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 549, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 549, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 549, + 547, + 713 + ], + "type": "text", + "content": "Another line of work [24, 25, 34, 44, 57] resorts to a layered UV map, capturing the geometry to some degree by incorporating video decomposition [7, 8, 14, 30, 51] to optimize the UV coordinates of the foreground and background based on the observed motion. Kasten et al. [25] unwrap a video into a set of layered 2D atlases where for each pixel in the video, its corresponding 2D coordinate in each of the atlases is predicted. Ye et al. [57] proposes a global sprite image that can group the distinct motion trajectories because the collective object structure has a consistent appearance throughout time. While preserving the temporal coherency and maintaining some coarse UV deformations related to arm or leg movements during a sequence, these methods fall short of capturing micro deformations like wrinkles in" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "4628" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 58, + 295, + 243 + ], + "blocks": [ + { + "bbox": [ + 48, + 58, + 295, + 243 + ], + "lines": [ + { + "bbox": [ + 48, + 58, + 295, + 243 + ], + "spans": [ + { + "bbox": [ + 48, + 58, + 295, + 243 + ], + "type": "image", + "image_path": "b77126e684d136db183fe79e5196927b37661ad0ec7a8c7b08792e87a983bd24.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 247, + 290, + 356 + ], + "lines": [ + { + "bbox": [ + 46, + 247, + 290, + 356 + ], + "spans": [ + { + "bbox": [ + 46, + 247, + 290, + 356 + ], + "type": "text", + "content": "Figure 2. Texture mapping geometry. We study the mapping between the image space, texture space, and 3D space. A point " + }, + { + "bbox": [ + 46, + 247, + 290, + 356 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 46, + 247, + 290, + 356 + ], + "type": "text", + "content": " in image is mapped to the texture space with " + }, + { + "bbox": [ + 46, + 247, + 290, + 356 + ], + "type": "inline_equation", + "content": "\\mathbf{u} = g(\\mathbf{x})" + }, + { + "bbox": [ + 46, + 247, + 290, + 356 + ], + "type": "text", + "content": ". The mapping " + }, + { + "bbox": [ + 46, + 247, + 290, + 356 + ], + "type": "inline_equation", + "content": "f(\\mathbf{u}) = \\mathbf{X}" + }, + { + "bbox": [ + 46, + 247, + 290, + 356 + ], + "type": "text", + "content": " lifts the texture plane into a 3D garment surface by an isometric warping. We use the orthographic projection model, resulting in the first two elements of " + }, + { + "bbox": [ + 46, + 247, + 290, + 356 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 247, + 290, + 356 + ], + "type": "text", + "content": " is the inverse of " + }, + { + "bbox": [ + 46, + 247, + 290, + 356 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 46, + 247, + 290, + 356 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 46, + 247, + 290, + 356 + ], + "type": "inline_equation", + "content": "f^{1:2} = g^{-1}" + }, + { + "bbox": [ + 46, + 247, + 290, + 356 + ], + "type": "text", + "content": ". The spatial derivatives of " + }, + { + "bbox": [ + 46, + 247, + 290, + 356 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 247, + 290, + 356 + ], + "type": "text", + "content": " form a tangent plane on the 3D surface at " + }, + { + "bbox": [ + 46, + 247, + 290, + 356 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 46, + 247, + 290, + 356 + ], + "type": "text", + "content": ", resulting in " + }, + { + "bbox": [ + 46, + 247, + 290, + 356 + ], + "type": "inline_equation", + "content": "\\mathbf{N}_{\\mathbf{x}} = f_{u} \\times f_{v}" + }, + { + "bbox": [ + 46, + 247, + 290, + 356 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 46, + 247, + 290, + 356 + ], + "type": "inline_equation", + "content": "\\mathbf{N}_{\\mathbf{x}}" + }, + { + "bbox": [ + 46, + 247, + 290, + 356 + ], + "type": "text", + "content": " is the 3D surface normal, and " + }, + { + "bbox": [ + 46, + 247, + 290, + 356 + ], + "type": "inline_equation", + "content": "f_{u}" + }, + { + "bbox": [ + 46, + 247, + 290, + 356 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 247, + 290, + 356 + ], + "type": "inline_equation", + "content": "f_{v}" + }, + { + "bbox": [ + 46, + 247, + 290, + 356 + ], + "type": "text", + "content": " are the spatial derivatives of " + }, + { + "bbox": [ + 46, + 247, + 290, + 356 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 247, + 290, + 356 + ], + "type": "text", + "content": " with respect to " + }, + { + "bbox": [ + 46, + 247, + 290, + 356 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 46, + 247, + 290, + 356 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 247, + 290, + 356 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 46, + 247, + 290, + 356 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 362, + 287, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 362, + 287, + 434 + ], + "spans": [ + { + "bbox": [ + 46, + 362, + 287, + 434 + ], + "type": "text", + "content": "the clothing. Furthermore, these techniques cannot represent UV mapping for an image and can only be applied to videos. Unlike these methods, our surface normal conditioned UV map is highly sensitive to small geometric details and 3D surface deformations, which can be optimized not only for a video but also for single images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 443, + 103, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 443, + 103, + 455 + ], + "spans": [ + { + "bbox": [ + 47, + 443, + 103, + 455 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 463, + 287, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 463, + 287, + 535 + ], + "spans": [ + { + "bbox": [ + 46, + 463, + 287, + 535 + ], + "type": "text", + "content": "Our goal is to obtain a continuous texture mapping, which allows editing the appearance of dynamic garments. We leverage the geometric property of isometry to constrain the UV map in the form of partial differential equations. We solve this partial differential equations by optimizing a neural network to generate a geometry aware UV map." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 540, + 280, + 554 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 540, + 280, + 554 + ], + "spans": [ + { + "bbox": [ + 47, + 540, + 280, + 554 + ], + "type": "text", + "content": "3.1. Texture Mapping without 3D Reconstruction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 559, + 287, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 559, + 287, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 559, + 287, + 652 + ], + "type": "text", + "content": "Consider a mapping " + }, + { + "bbox": [ + 46, + 559, + 287, + 652 + ], + "type": "inline_equation", + "content": "g(\\mathbf{x}) = \\mathbf{u}" + }, + { + "bbox": [ + 46, + 559, + 287, + 652 + ], + "type": "text", + "content": " that maps a pixel location " + }, + { + "bbox": [ + 46, + 559, + 287, + 652 + ], + "type": "inline_equation", + "content": "\\mathbf{x} = (x, y) \\in \\mathbb{R}^2" + }, + { + "bbox": [ + 46, + 559, + 287, + 652 + ], + "type": "text", + "content": " in the image space that belongs to a garment of interest to a point " + }, + { + "bbox": [ + 46, + 559, + 287, + 652 + ], + "type": "inline_equation", + "content": "\\mathbf{u} = (u, v) \\in \\mathbb{R}^2" + }, + { + "bbox": [ + 46, + 559, + 287, + 652 + ], + "type": "text", + "content": " in the UV space of the garment as shown in Figure 2. The goal of our work is to find such a mapping " + }, + { + "bbox": [ + 46, + 559, + 287, + 652 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 46, + 559, + 287, + 652 + ], + "type": "text", + "content": " that takes into account the local surface geometry measured by the surface normal predicted at " + }, + { + "bbox": [ + 46, + 559, + 287, + 652 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 46, + 559, + 287, + 652 + ], + "type": "text", + "content": ". We denote the predicted 3D surface normal of " + }, + { + "bbox": [ + 46, + 559, + 287, + 652 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 46, + 559, + 287, + 652 + ], + "type": "text", + "content": " in the camera space as " + }, + { + "bbox": [ + 46, + 559, + 287, + 652 + ], + "type": "inline_equation", + "content": "\\mathbf{N}_{\\mathbf{x}} \\in \\mathbb{S}^2" + }, + { + "bbox": [ + 46, + 559, + 287, + 652 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 653, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 689 + ], + "type": "text", + "content": "Let us define an isometric map from the UV texture map to the 3D surface, " + }, + { + "bbox": [ + 46, + 653, + 287, + 689 + ], + "type": "inline_equation", + "content": "f(\\mathbf{u}) = \\mathbf{X}" + }, + { + "bbox": [ + 46, + 653, + 287, + 689 + ], + "type": "text", + "content": ". This is the fundamental property of a non-stretchable cloth texture mapping [10]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 104, + 696, + 287, + 711 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 696, + 287, + 711 + ], + "spans": [ + { + "bbox": [ + 104, + 696, + 287, + 711 + ], + "type": "interline_equation", + "content": "\\left\\| f _ {u} \\right\\| = \\left\\| f _ {v} \\right\\| = 1, \\quad f _ {u} ^ {\\mathrm {T}} f _ {v} = 0, \\tag {1}", + "image_path": "0d4ace9c1e83c6e62dc625273296ab8102b90c66450026366761828c3bb62a06.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 546, + 118 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 546, + 118 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 546, + 118 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 72, + 546, + 118 + ], + "type": "inline_equation", + "content": "f_{u}" + }, + { + "bbox": [ + 304, + 72, + 546, + 118 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 546, + 118 + ], + "type": "inline_equation", + "content": "f_{v}" + }, + { + "bbox": [ + 304, + 72, + 546, + 118 + ], + "type": "text", + "content": " are the partial derivatives of " + }, + { + "bbox": [ + 304, + 72, + 546, + 118 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 72, + 546, + 118 + ], + "type": "text", + "content": " with respect to " + }, + { + "bbox": [ + 304, + 72, + 546, + 118 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 304, + 72, + 546, + 118 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 546, + 118 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 304, + 72, + 546, + 118 + ], + "type": "text", + "content": ", respectively. Geometrically, " + }, + { + "bbox": [ + 304, + 72, + 546, + 118 + ], + "type": "inline_equation", + "content": "f_{u}" + }, + { + "bbox": [ + 304, + 72, + 546, + 118 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 546, + 118 + ], + "type": "inline_equation", + "content": "f_{v}" + }, + { + "bbox": [ + 304, + 72, + 546, + 118 + ], + "type": "text", + "content": " are the tangential vectors on the 3D surface where their cross product forms the surface normal:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 345, + 129, + 545, + 144 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 129, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 345, + 129, + 545, + 144 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathbf {N}} _ {\\mathbf {x}} = f _ {u} \\times f _ {v} = f _ {u} (g (\\mathbf {x})) \\times f _ {v} (g (\\mathbf {x})) \\tag {2}", + "image_path": "94b8373a9edbf7f478f0717a02cf34bedde5010bda1e60bf9c90c355c4308e90.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 155, + 545, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 155, + 545, + 178 + ], + "spans": [ + { + "bbox": [ + 304, + 155, + 545, + 178 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 155, + 545, + 178 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathbf{N}}_{\\mathbf{x}} \\in \\mathbb{S}^2" + }, + { + "bbox": [ + 304, + 155, + 545, + 178 + ], + "type": "text", + "content": " is the surface normal at " + }, + { + "bbox": [ + 304, + 155, + 545, + 178 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 304, + 155, + 545, + 178 + ], + "type": "text", + "content": " corresponding to " + }, + { + "bbox": [ + 304, + 155, + 545, + 178 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 155, + 545, + 178 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 180, + 545, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 180, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 304, + 180, + 545, + 217 + ], + "type": "text", + "content": "We can find the UV mapping " + }, + { + "bbox": [ + 304, + 180, + 545, + 217 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 304, + 180, + 545, + 217 + ], + "type": "text", + "content": " by matching the surface normal " + }, + { + "bbox": [ + 304, + 180, + 545, + 217 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathbf{N}}_{\\mathbf{x}}" + }, + { + "bbox": [ + 304, + 180, + 545, + 217 + ], + "type": "text", + "content": " derived by Equation (2) and the surface normal predicted from the image " + }, + { + "bbox": [ + 304, + 180, + 545, + 217 + ], + "type": "inline_equation", + "content": "\\mathbf{N}_{\\mathbf{x}}" + }, + { + "bbox": [ + 304, + 180, + 545, + 217 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 330, + 227, + 545, + 269 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 227, + 545, + 269 + ], + "spans": [ + { + "bbox": [ + 330, + 227, + 545, + 269 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\underset {\\boldsymbol {\\theta} _ {g}, \\boldsymbol {\\theta} _ {f}} {\\operatorname {m i n i m i z e}} \\sum_ {\\mathbf {x}} \\| f _ {u} (g (\\mathbf {x})) \\times f _ {v} (g (\\mathbf {x})) - \\mathbf {N} _ {\\mathbf {x}} \\| ^ {2}, \\\\ s. t. \\quad \\| f _ {u} \\| = \\| f _ {v} \\| = 1, f _ {u} ^ {\\top} f _ {v} = 0, \\tag {3} \\\\ \\end{array}", + "image_path": "176f517ef5c73bc96f01cefca60912ae5863c1ffe7788c38acfcf0128af64529.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 278, + 545, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 278, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 304, + 278, + 545, + 338 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 278, + 545, + 338 + ], + "type": "inline_equation", + "content": "\\theta_{g}" + }, + { + "bbox": [ + 304, + 278, + 545, + 338 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 278, + 545, + 338 + ], + "type": "inline_equation", + "content": "\\theta_{f}" + }, + { + "bbox": [ + 304, + 278, + 545, + 338 + ], + "type": "text", + "content": " are the parameters of the function " + }, + { + "bbox": [ + 304, + 278, + 545, + 338 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 304, + 278, + 545, + 338 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 278, + 545, + 338 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 278, + 545, + 338 + ], + "type": "text", + "content": ", respectively. A key challenge of solving Equation (3) lies in the dependency of " + }, + { + "bbox": [ + 304, + 278, + 545, + 338 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 278, + 545, + 338 + ], + "type": "text", + "content": " that requires full 3D reconstruction of the surface. Instead, we formulate a new dual problem that can solve Equation (3) effectively without finding " + }, + { + "bbox": [ + 304, + 278, + 545, + 338 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 278, + 545, + 338 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 338, + 545, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 338, + 545, + 389 + ], + "spans": [ + { + "bbox": [ + 304, + 338, + 545, + 389 + ], + "type": "text", + "content": "We use two properties to eliminate " + }, + { + "bbox": [ + 304, + 338, + 545, + 389 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 338, + 545, + 389 + ], + "type": "text", + "content": " from Equation (3). First, we assume orthographic projection, i.e., " + }, + { + "bbox": [ + 304, + 338, + 545, + 389 + ], + "type": "inline_equation", + "content": "(x,y) = (X,Y)" + }, + { + "bbox": [ + 304, + 338, + 545, + 389 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 338, + 545, + 389 + ], + "type": "inline_equation", + "content": "\\mathbf{X} = \\left[ \\begin{array}{ccc}X & Y & Z \\end{array} \\right]^{\\mathrm{T}}" + }, + { + "bbox": [ + 304, + 338, + 545, + 389 + ], + "type": "text", + "content": ". This allows us to express the 3D derivatives using the pixel coordinates:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 393, + 398, + 545, + 415 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 398, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 393, + 398, + 545, + 415 + ], + "type": "interline_equation", + "content": "g = \\left(f ^ {1: 2}\\right) ^ {- 1}, \\tag {4}", + "image_path": "49c0fe1c9ac06e0b52a825f774349ad2cc6a29023fda47efd456f247329d8107.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 424, + 545, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 424, + 545, + 495 + ], + "spans": [ + { + "bbox": [ + 304, + 424, + 545, + 495 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 424, + 545, + 495 + ], + "type": "inline_equation", + "content": "f^{1:2}" + }, + { + "bbox": [ + 304, + 424, + 545, + 495 + ], + "type": "text", + "content": " is the first two elements " + }, + { + "bbox": [ + 304, + 424, + 545, + 495 + ], + "type": "inline_equation", + "content": "(X, Y)" + }, + { + "bbox": [ + 304, + 424, + 545, + 495 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 304, + 424, + 545, + 495 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 424, + 545, + 495 + ], + "type": "text", + "content": ". To keep " + }, + { + "bbox": [ + 304, + 424, + 545, + 495 + ], + "type": "inline_equation", + "content": "f^{1:2}" + }, + { + "bbox": [ + 304, + 424, + 545, + 495 + ], + "type": "text", + "content": " bijective, we assume there is no self occlusion in the camera projection of " + }, + { + "bbox": [ + 304, + 424, + 545, + 495 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 424, + 545, + 495 + ], + "type": "text", + "content": ". Note that " + }, + { + "bbox": [ + 304, + 424, + 545, + 495 + ], + "type": "inline_equation", + "content": "g^{-1}" + }, + { + "bbox": [ + 304, + 424, + 545, + 495 + ], + "type": "text", + "content": " is the inverse of " + }, + { + "bbox": [ + 304, + 424, + 545, + 495 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 304, + 424, + 545, + 495 + ], + "type": "text", + "content": " that maps the UV texture map to the pixel coordinate. Second, we derive the derivatives of " + }, + { + "bbox": [ + 304, + 424, + 545, + 495 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 304, + 424, + 545, + 495 + ], + "type": "text", + "content": " by using the inverse function theorem [9]:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 341, + 505, + 545, + 521 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 505, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 341, + 505, + 545, + 521 + ], + "type": "interline_equation", + "content": "f ^ {1: 2} \\circ g (\\mathbf {x}) = \\mathbf {x} \\rightarrow \\mathbf {J} _ {g} = \\left(\\mathbf {J} _ {(f ^ {1: 2})}\\right) ^ {- 1}, \\tag {5}", + "image_path": "9606be2facbe61ede512161d5532daa0ef29ad075ac1b3b60982ef8aac78bb48.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 531, + 545, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 531, + 545, + 573 + ], + "spans": [ + { + "bbox": [ + 304, + 531, + 545, + 573 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 531, + 545, + 573 + ], + "type": "inline_equation", + "content": "\\mathbf{J}_g = \\left[g_x\\quad g_y\\right] = \\left[ \\begin{array}{cc}\\frac{\\partial u}{\\partial x} & \\frac{\\partial u}{\\partial y}\\\\ \\frac{\\partial v}{\\partial x} & \\frac{\\partial v}{\\partial y} \\end{array} \\right]" + }, + { + "bbox": [ + 304, + 531, + 545, + 573 + ], + "type": "text", + "content": " is the Jacobian matrix of the function " + }, + { + "bbox": [ + 304, + 531, + 545, + 573 + ], + "type": "inline_equation", + "content": "g" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 574, + 545, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 574, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 304, + 574, + 545, + 609 + ], + "type": "text", + "content": "With Equation (1) and (5), Equation (2) can be re-written as the following constraints by eliminating " + }, + { + "bbox": [ + 304, + 574, + 545, + 609 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 574, + 545, + 609 + ], + "type": "text", + "content": " (3D reconstruction):" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 302, + 619, + 545, + 652 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 619, + 545, + 652 + ], + "spans": [ + { + "bbox": [ + 302, + 619, + 545, + 652 + ], + "type": "interline_equation", + "content": "\\left\\| g _ {x} \\right\\| = \\sqrt {1 + \\frac {\\widetilde {n} _ {x} ^ {2}}{\\widetilde {n} _ {z} ^ {2}}}, \\left\\| g _ {y} \\right\\| = \\sqrt {1 + \\frac {\\widetilde {n} _ {y} ^ {2}}{\\widetilde {n} _ {z} ^ {2}}}, g _ {x} ^ {\\intercal} g _ {y} = \\frac {\\widetilde {n} _ {x} \\widetilde {n} _ {y}}{\\widetilde {n} _ {z} ^ {2}}, \\tag {6}", + "image_path": "d7677345f24caa2e1f2494975589dedf7b850956180c487b41db57175d856a51.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 663, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 663, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 663, + 545, + 689 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 663, + 545, + 689 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathbf{N}}_{\\mathbf{x}} = \\left[\\widetilde{n}_{x}\\quad \\widetilde{n}_{y}\\quad \\widetilde{n}_{z}\\right]^{\\mathrm{T}}" + }, + { + "bbox": [ + 304, + 663, + 545, + 689 + ], + "type": "text", + "content": " . For the derivation of Equation (6), see Supplementary Material." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "content": "Equation (6) is a set of partial differential equations of " + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "content": " that needs to match with the predicted surface normal " + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{N}_{\\mathbf{x}} =" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4629" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 59, + 280, + 273 + ], + "blocks": [ + { + "bbox": [ + 58, + 59, + 280, + 273 + ], + "lines": [ + { + "bbox": [ + 58, + 59, + 280, + 273 + ], + "spans": [ + { + "bbox": [ + 58, + 59, + 280, + 273 + ], + "type": "image", + "image_path": "855dd2df351d699838881e994ff24750573122b799c8a4c013220ab617dcf353.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 277, + 289, + 373 + ], + "lines": [ + { + "bbox": [ + 46, + 277, + 289, + 373 + ], + "spans": [ + { + "bbox": [ + 46, + 277, + 289, + 373 + ], + "type": "text", + "content": "Figure 3. 2D simplification of UV map estimation. We illustrate the isometric relationship between (a) the surface and (c) the UV map. The length of the curve in XZ axis needs to be preserved when mapping to " + }, + { + "bbox": [ + 46, + 277, + 289, + 373 + ], + "type": "inline_equation", + "content": "g(x)" + }, + { + "bbox": [ + 46, + 277, + 289, + 373 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 46, + 277, + 289, + 373 + ], + "type": "inline_equation", + "content": "|\\mathrm{d}u| = |\\mathrm{d}r| = \\sqrt{\\mathrm{d}x^2 + \\mathrm{d}z^2}" + }, + { + "bbox": [ + 46, + 277, + 289, + 373 + ], + "type": "text", + "content": " (arc length preservation). This relationship can be re-written as (b) a partial differential equation using Equation (6) in terms of its surface normal and the inverse of the spatial derivative of " + }, + { + "bbox": [ + 46, + 277, + 289, + 373 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 46, + 277, + 289, + 373 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 46, + 277, + 289, + 373 + ], + "type": "inline_equation", + "content": "\\left(\\frac{\\mathrm{d}u}{\\mathrm{d}x}\\right)^2 = 1 + \\left(\\frac{\\mathrm{d}n_x}{\\mathrm{d}n_z}\\right)^2" + }, + { + "bbox": [ + 46, + 277, + 289, + 373 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 376, + 225, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 376, + 225, + 395 + ], + "spans": [ + { + "bbox": [ + 48, + 376, + 225, + 395 + ], + "type": "inline_equation", + "content": "\\left[ \\begin{array}{lll}n_{x} & n_{y} & n_{z} \\end{array} \\right]^{\\mathsf{T}}" + }, + { + "bbox": [ + 48, + 376, + 225, + 395 + ], + "type": "text", + "content": " , leading to a loss function:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 402, + 287, + 504 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 402, + 287, + 504 + ], + "spans": [ + { + "bbox": [ + 45, + 402, + 287, + 504 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\mathrm {g e o}} (\\boldsymbol {\\theta} _ {g}) = \\sum_ {\\mathbf {x}} \\left(\\left. \\left(\\frac {\\partial u}{\\partial x} \\right| _ {\\mathbf {x}}\\right) ^ {2} + \\left. \\left(\\frac {\\partial v}{\\partial x} \\right| _ {\\mathbf {x}}\\right) ^ {2} - 1 - \\frac {n _ {x} ^ {2}}{n _ {z} ^ {2}}\\right) ^ {2} \\\\ + \\left(\\left. \\left(\\frac {\\partial u}{\\partial y} \\right| _ {\\mathbf {x}}\\right) ^ {2} + \\left. \\left(\\frac {\\partial v}{\\partial y} \\right| _ {\\mathbf {x}}\\right) ^ {2} - 1 - \\frac {n _ {y} ^ {2}}{n _ {z} ^ {2}}\\right) ^ {2} \\\\ + \\left(\\frac {\\partial u}{\\partial x} \\bigg | _ {\\mathbf {x}} \\frac {\\partial u}{\\partial y} \\bigg | _ {\\mathbf {x}} + \\left. \\frac {\\partial v}{\\partial x} \\right| _ {\\mathbf {x}} \\left. \\frac {\\partial v}{\\partial y} \\right| _ {\\mathbf {x}} - \\frac {n _ {x} n _ {y}}{n _ {z} ^ {2}}\\right) ^ {2}, \\tag {7} \\\\ \\end{array}", + "image_path": "853af10d6531f6887bfa03a1b24c00089fff68961649f2f87a37a636d83ea25c.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 512, + 287, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 512, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 46, + 512, + 287, + 536 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 512, + 287, + 536 + ], + "type": "inline_equation", + "content": "\\left.\\frac{\\partial u}{\\partial x}\\right|_{\\mathbf{x}}" + }, + { + "bbox": [ + 46, + 512, + 287, + 536 + ], + "type": "text", + "content": " is the partial derivative of " + }, + { + "bbox": [ + 46, + 512, + 287, + 536 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 46, + 512, + 287, + 536 + ], + "type": "text", + "content": " with respect to " + }, + { + "bbox": [ + 46, + 512, + 287, + 536 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 512, + 287, + 536 + ], + "type": "text", + "content": " evaluated at " + }, + { + "bbox": [ + 46, + 512, + 287, + 536 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 46, + 512, + 287, + 536 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 537, + 287, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 537, + 287, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 537, + 287, + 640 + ], + "type": "text", + "content": "Figure 3 illustrates a 2D simplification of UV map estimation. A curve in XZ plane forms a isometric relation with the UV map, " + }, + { + "bbox": [ + 46, + 537, + 287, + 640 + ], + "type": "inline_equation", + "content": "|\\mathrm{du}| = |\\mathrm{dr}| = \\sqrt{\\mathrm{dx}^2 + \\mathrm{dz}^2}" + }, + { + "bbox": [ + 46, + 537, + 287, + 640 + ], + "type": "text", + "content": ". This relationship can be re-written as a partial differential equation in terms of the surface normal " + }, + { + "bbox": [ + 46, + 537, + 287, + 640 + ], + "type": "inline_equation", + "content": "(n_x, n_z)" + }, + { + "bbox": [ + 46, + 537, + 287, + 640 + ], + "type": "text", + "content": " and the inverse of the spatial derivative of " + }, + { + "bbox": [ + 46, + 537, + 287, + 640 + ], + "type": "inline_equation", + "content": "g(x)" + }, + { + "bbox": [ + 46, + 537, + 287, + 640 + ], + "type": "text", + "content": " using Equation (6), i.e., " + }, + { + "bbox": [ + 46, + 537, + 287, + 640 + ], + "type": "inline_equation", + "content": "\\left(\\frac{\\mathrm{du}}{\\mathrm{dx}}\\right)^2 = 1 + \\left(\\frac{\\mathrm{dn}_x}{\\mathrm{dn}_z}\\right)^2" + }, + { + "bbox": [ + 46, + 537, + 287, + 640 + ], + "type": "text", + "content": ". We solve these partial differential equations to estimate " + }, + { + "bbox": [ + 46, + 537, + 287, + 640 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 46, + 537, + 287, + 640 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 647, + 284, + 662 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 284, + 662 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 284, + 662 + ], + "type": "text", + "content": "3.2. Self-supervised Learning of Texture Mapping" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 666, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 666, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 666, + 288, + 715 + ], + "type": "text", + "content": "The texture map " + }, + { + "bbox": [ + 46, + 666, + 288, + 715 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 46, + 666, + 288, + 715 + ], + "type": "text", + "content": " is defined up to a bijective function, i.e., there exists an infinite number of " + }, + { + "bbox": [ + 46, + 666, + 288, + 715 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 46, + 666, + 288, + 715 + ], + "type": "text", + "content": " that are equivalent: " + }, + { + "bbox": [ + 46, + 666, + 288, + 715 + ], + "type": "inline_equation", + "content": "g^{-1} \\circ g = (\\mathcal{T} \\circ g)^{-1} \\circ (\\mathcal{T} \\circ g)" + }, + { + "bbox": [ + 46, + 666, + 288, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 666, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 46, + 666, + 288, + 715 + ], + "type": "text", + "content": " is a bijective map (e.g., Euclidean transform). We resolve this ambiguity by" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": "finding " + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "inline_equation", + "content": "g \\approx g'" + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "inline_equation", + "content": "g'" + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": " is a pre-defined proxy map of humans:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 353, + 105, + 545, + 129 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 105, + 545, + 129 + ], + "spans": [ + { + "bbox": [ + 353, + 105, + 545, + 129 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {p r o x}} \\left(\\boldsymbol {\\theta} _ {g}\\right) = \\sum_ {\\mathbf {x}} \\left\\| g ^ {\\prime} (\\mathbf {x}) - g (\\mathbf {x}) \\right\\| ^ {2}. \\tag {8}", + "image_path": "e37d4f9ca019e6e2f4b2f011ea60fb25fbc745b4143be477393943729fa0ee0a.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 137, + 545, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 137, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 304, + 137, + 545, + 196 + ], + "type": "text", + "content": "In practice, we use an extended DensePose [18] as the pre-defined proxy map. Since DensePose makes predictions only for the human body, we apply an extrapolation method [50] to inpaint the garment regions that are not covered by DensePose." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 196, + 545, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 196, + 545, + 243 + ], + "spans": [ + { + "bbox": [ + 305, + 196, + 545, + 243 + ], + "type": "text", + "content": "Further, we ensure physical plausibility of the visible 3D surfaces, i.e., the texture map should result in the surface normals pointing to " + }, + { + "bbox": [ + 305, + 196, + 545, + 243 + ], + "type": "inline_equation", + "content": "+Z" + }, + { + "bbox": [ + 305, + 196, + 545, + 243 + ], + "type": "text", + "content": " direction, by adding the following loss:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 354, + 251, + 545, + 275 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 251, + 545, + 275 + ], + "spans": [ + { + "bbox": [ + 354, + 251, + 545, + 275 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {z}} \\left(\\boldsymbol {\\theta} _ {g}\\right) = \\sum_ {\\mathbf {x}} \\max (0, \\det (\\mathbf {J} _ {g} | _ {\\mathbf {x}})), \\tag {9}", + "image_path": "f39a906436c0b840a6f33553af5d58e37ebdc680548aa555eed0f26e97e7113e.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 284, + 545, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 284, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 304, + 284, + 545, + 319 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 284, + 545, + 319 + ], + "type": "inline_equation", + "content": "\\operatorname{det}(\\mathbf{J}_g)" + }, + { + "bbox": [ + 304, + 284, + 545, + 319 + ], + "type": "text", + "content": " is the determinant of the Jacobian " + }, + { + "bbox": [ + 304, + 284, + 545, + 319 + ], + "type": "inline_equation", + "content": "\\mathbf{J}_g" + }, + { + "bbox": [ + 304, + 284, + 545, + 319 + ], + "type": "text", + "content": " that is equivalent to " + }, + { + "bbox": [ + 304, + 284, + 545, + 319 + ], + "type": "inline_equation", + "content": "\\widetilde{n}_z" + }, + { + "bbox": [ + 304, + 284, + 545, + 319 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 284, + 545, + 319 + ], + "type": "inline_equation", + "content": "\\mathbf{J}_g|_{\\mathbf{x}}" + }, + { + "bbox": [ + 304, + 284, + 545, + 319 + ], + "type": "text", + "content": " is the Jacobian matrix of " + }, + { + "bbox": [ + 304, + 284, + 545, + 319 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 304, + 284, + 545, + 319 + ], + "type": "text", + "content": " evaluated at " + }, + { + "bbox": [ + 304, + 284, + 545, + 319 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 284, + 545, + 319 + ], + "type": "text", + "content": ". See Supplementary Material for derivation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 319, + 545, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 319, + 545, + 389 + ], + "spans": [ + { + "bbox": [ + 304, + 319, + 545, + 389 + ], + "type": "text", + "content": "For a video, we extend the texture map to include the image feature for each pixel, i.e., " + }, + { + "bbox": [ + 304, + 319, + 545, + 389 + ], + "type": "inline_equation", + "content": "g(\\mathbf{x}, \\mathbf{f}_{\\mathbf{x}})" + }, + { + "bbox": [ + 304, + 319, + 545, + 389 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 319, + 545, + 389 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{\\mathbf{x}}" + }, + { + "bbox": [ + 304, + 319, + 545, + 389 + ], + "type": "text", + "content": " is the image feature at " + }, + { + "bbox": [ + 304, + 319, + 545, + 389 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 319, + 545, + 389 + ], + "type": "text", + "content": ". This allows us to generalize the texture map over time. With the extension, we ensure the temporal consistency of the texture map by leveraging optical flow across frames:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 315, + 396, + 545, + 423 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 396, + 545, + 423 + ], + "spans": [ + { + "bbox": [ + 315, + 396, + 545, + 423 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\operatorname {t m p}} \\left(\\boldsymbol {\\theta} _ {g}\\right) = \\sum_ {i, j} \\sum_ {\\mathbf {x} _ {i}} \\left\\| g \\left(\\mathbf {x} _ {i}, \\mathbf {f} _ {\\mathbf {x} _ {i}}\\right) - g \\left(\\mathbf {x} _ {j}, \\mathbf {f} _ {\\mathbf {x} _ {j}}\\right) \\right\\| ^ {2}, \\tag {10}", + "image_path": "609ada7d54545f7cddb40850f43221d3bff1c6d02674aa9ef5c1bf424ad62597.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 433, + 545, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 433, + 545, + 468 + ], + "spans": [ + { + "bbox": [ + 304, + 433, + 545, + 468 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 433, + 545, + 468 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 304, + 433, + 545, + 468 + ], + "type": "text", + "content": " is a point in the " + }, + { + "bbox": [ + 304, + 433, + 545, + 468 + ], + "type": "inline_equation", + "content": "i^{\\mathrm{th}}" + }, + { + "bbox": [ + 304, + 433, + 545, + 468 + ], + "type": "text", + "content": " frame. This point is mapped to " + }, + { + "bbox": [ + 304, + 433, + 545, + 468 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_j" + }, + { + "bbox": [ + 304, + 433, + 545, + 468 + ], + "type": "text", + "content": " in the " + }, + { + "bbox": [ + 304, + 433, + 545, + 468 + ], + "type": "inline_equation", + "content": "j^{\\mathrm{th}}" + }, + { + "bbox": [ + 304, + 433, + 545, + 468 + ], + "type": "text", + "content": " frame, i.e., " + }, + { + "bbox": [ + 304, + 433, + 545, + 468 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_j = W_{i\\rightarrow j}(\\mathbf{x}_i)" + }, + { + "bbox": [ + 304, + 433, + 545, + 468 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 433, + 545, + 468 + ], + "type": "inline_equation", + "content": "W_{i\\rightarrow j}" + }, + { + "bbox": [ + 304, + 433, + 545, + 468 + ], + "type": "text", + "content": " is the optical flow from the " + }, + { + "bbox": [ + 304, + 433, + 545, + 468 + ], + "type": "inline_equation", + "content": "i^{\\mathrm{th}}" + }, + { + "bbox": [ + 304, + 433, + 545, + 468 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 433, + 545, + 468 + ], + "type": "inline_equation", + "content": "j^{\\mathrm{th}}" + }, + { + "bbox": [ + 304, + 433, + 545, + 468 + ], + "type": "text", + "content": " frames." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 468, + 545, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 468, + 545, + 492 + ], + "spans": [ + { + "bbox": [ + 305, + 468, + 545, + 492 + ], + "type": "text", + "content": "Overall, we optimize the following loss to learn the texture map:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 501, + 545, + 515 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 501, + 545, + 515 + ], + "spans": [ + { + "bbox": [ + 313, + 501, + 545, + 515 + ], + "type": "interline_equation", + "content": "\\mathcal {L} \\left(\\boldsymbol {\\theta} _ {g}\\right) = \\mathcal {L} _ {\\text {g e o}} + \\lambda_ {\\text {p r o x}} \\mathcal {L} _ {\\text {p r o x}} + \\lambda_ {\\mathrm {z}} \\mathcal {L} _ {\\mathrm {z}} + \\lambda_ {\\text {t m p}} \\mathcal {L} _ {\\text {t m p}}, \\tag {11}", + "image_path": "d40d8aa8fa7181828225034481921176970e0825fdd0b200e83eaa94096bac0d.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 523, + 545, + 560 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 523, + 545, + 560 + ], + "spans": [ + { + "bbox": [ + 304, + 523, + 545, + 560 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 523, + 545, + 560 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{prox}}" + }, + { + "bbox": [ + 304, + 523, + 545, + 560 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 523, + 545, + 560 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{z}}" + }, + { + "bbox": [ + 304, + 523, + 545, + 560 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 523, + 545, + 560 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{tmp}}" + }, + { + "bbox": [ + 304, + 523, + 545, + 560 + ], + "type": "text", + "content": " are the weights that determine the relative importance of losses. Note that when a single image is used, " + }, + { + "bbox": [ + 304, + 523, + 545, + 560 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{tmp}} = 0" + }, + { + "bbox": [ + 304, + 523, + 545, + 560 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 566, + 440, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 566, + 440, + 578 + ], + "spans": [ + { + "bbox": [ + 306, + 566, + 440, + 578 + ], + "type": "text", + "content": "3.3. Implementation Details" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 584, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 584, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 584, + 545, + 715 + ], + "type": "text", + "content": "We model " + }, + { + "bbox": [ + 304, + 584, + 545, + 715 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 304, + 584, + 545, + 715 + ], + "type": "text", + "content": " using a 12-layer multi-layer perception, with ReLU [2] as an activation function after each layer that takes as input a pixel coordinate with positional encoding, " + }, + { + "bbox": [ + 304, + 584, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\gamma (\\mathbf{x})" + }, + { + "bbox": [ + 304, + 584, + 545, + 715 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 584, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\gamma :\\mathbb{R}^2\\to \\mathbb{R}^{128}" + }, + { + "bbox": [ + 304, + 584, + 545, + 715 + ], + "type": "text", + "content": " is Fourier based positional encoding [48]. For videos, we use ResNet [19] to extract per-frame 256 dimensional image feature " + }, + { + "bbox": [ + 304, + 584, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{\\mathbf{x}}" + }, + { + "bbox": [ + 304, + 584, + 545, + 715 + ], + "type": "text", + "content": ". Our network design is illustrated in Figure 4 (for image-based UV map prediction) and 5 (for video-based UV map prediction). To make our prediction scale-invariant, we crop the garments region with " + }, + { + "bbox": [ + 304, + 584, + 545, + 715 + ], + "type": "inline_equation", + "content": "256\\times 256" + }, + { + "bbox": [ + 304, + 584, + 545, + 715 + ], + "type": "text", + "content": " resolution. We use an offthe-shelf garment segmentation software, Graphonomy [15]" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4630" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 72, + 293, + 214 + ], + "blocks": [ + { + "bbox": [ + 50, + 72, + 293, + 214 + ], + "lines": [ + { + "bbox": [ + 50, + 72, + 293, + 214 + ], + "spans": [ + { + "bbox": [ + 50, + 72, + 293, + 214 + ], + "type": "image", + "image_path": "b36a786f68411016cf1750e89cd1e0a7277c39007ed4d489f150e882959aebbb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 219, + 289, + 296 + ], + "lines": [ + { + "bbox": [ + 46, + 219, + 289, + 296 + ], + "spans": [ + { + "bbox": [ + 46, + 219, + 289, + 296 + ], + "type": "text", + "content": "Figure 4. Single image framework. For each pixel location on a garment " + }, + { + "bbox": [ + 46, + 219, + 289, + 296 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 46, + 219, + 289, + 296 + ], + "type": "text", + "content": ", we predict the UV coordinate " + }, + { + "bbox": [ + 46, + 219, + 289, + 296 + ], + "type": "inline_equation", + "content": "\\mathbf{u}" + }, + { + "bbox": [ + 46, + 219, + 289, + 296 + ], + "type": "text", + "content": " using a multi-layer perceptron. We enforce isometry to the UV map by matching with the predicted 3D surface normals. " + }, + { + "bbox": [ + 46, + 219, + 289, + 296 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{z}}" + }, + { + "bbox": [ + 46, + 219, + 289, + 296 + ], + "type": "text", + "content": " ensures that the predicted surface normals point toward the camera. To disambiguate the frame of reference of the UV map, we use pre-defined proxy map (e.g., DensePose [18] extrapolation [50]) by applying " + }, + { + "bbox": [ + 46, + 219, + 289, + 296 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{prox}}" + }, + { + "bbox": [ + 46, + 219, + 289, + 296 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 303, + 289, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 303, + 289, + 410 + ], + "spans": [ + { + "bbox": [ + 46, + 303, + 289, + 410 + ], + "type": "text", + "content": "to separate the garment area. We use Adam optimizer [27] with batch size of 2048 and learning rate of " + }, + { + "bbox": [ + 46, + 303, + 289, + 410 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 46, + 303, + 289, + 410 + ], + "type": "text", + "content": ". We set " + }, + { + "bbox": [ + 46, + 303, + 289, + 410 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{prox}} = 0.2" + }, + { + "bbox": [ + 46, + 303, + 289, + 410 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 303, + 289, + 410 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{z}} = 0.01" + }, + { + "bbox": [ + 46, + 303, + 289, + 410 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 303, + 289, + 410 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{tmp}} = 0.3" + }, + { + "bbox": [ + 46, + 303, + 289, + 410 + ], + "type": "text", + "content": ". We used an NVIDIA V100 GPU and Intel(R) Xeon(R) CPU E5-2698 v4 @ 2.20GHz, and implemented our approach with Pytorch [43]. Our method takes 20 minutes for a video of 82 frames while Kasten et al. [25], Ye et al. [57], and TemporalUV [55] take 10 hours, 30 minutes, and 23 hours, respectively." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 422, + 119, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 422, + 119, + 434 + ], + "spans": [ + { + "bbox": [ + 47, + 422, + 119, + 434 + ], + "type": "text", + "content": "4. Evaluation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 442, + 287, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 442, + 287, + 501 + ], + "spans": [ + { + "bbox": [ + 46, + 442, + 287, + 501 + ], + "type": "text", + "content": "We evaluate our method both quantitatively and qualitatively on real images as well as synthetic data with ground truth UV map. We also compare with the state-of-the-art methods on human UV map estimation methods and video UV map optimization approaches." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 502, + 287, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 502, + 287, + 583 + ], + "spans": [ + { + "bbox": [ + 46, + 502, + 287, + 583 + ], + "type": "text", + "content": "Evaluation Datasets We evaluate our method using the following datasets: (1) five synthetic video sequences of simulated dress and T-shirt garments from Santesteban et al. [45] with random texture patterns over 700 frames; (2) ten real videos from Fashion Video dataset [59]; (3) TikTok dataset [23] and various YouTube videos as well as in-the-wild internet images." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 584, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 584, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 584, + 288, + 714 + ], + "type": "text", + "content": "Evaluation Metric We use five metrics to evaluate our method. (1) UV error: for synthetic data, we report the absolute UV error in the texture space [45]. We use a Procrustes analysis [16] to align the resulting texture map to account for the diambiguity of the reference frame. We report the mean squared error in metric scale by assuming the height of the person in the input is " + }, + { + "bbox": [ + 46, + 584, + 288, + 714 + ], + "type": "inline_equation", + "content": "165\\mathrm{cm}" + }, + { + "bbox": [ + 46, + 584, + 288, + 714 + ], + "type": "text", + "content": ", resulting in " + }, + { + "bbox": [ + 46, + 584, + 288, + 714 + ], + "type": "inline_equation", + "content": "0.41\\mathrm{cm} / \\mathrm{UV}" + }, + { + "bbox": [ + 46, + 584, + 288, + 714 + ], + "type": "text", + "content": " for dress and " + }, + { + "bbox": [ + 46, + 584, + 288, + 714 + ], + "type": "inline_equation", + "content": "0.27\\mathrm{cm} / \\mathrm{UV}" + }, + { + "bbox": [ + 46, + 584, + 288, + 714 + ], + "type": "text", + "content": " for T-shirt (Table 1). (2) Average Precision percentage: we report the Average Precision (AP) percentage metric computed on all the pixels considering a per-pixel prediction as correct if the UV" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 107 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 107 + ], + "type": "text", + "content": "error is lower than a threshold. We visualize the AP metric for a range of thresholds from 1 to " + }, + { + "bbox": [ + 304, + 72, + 545, + 107 + ], + "type": "inline_equation", + "content": "15\\mathrm{cm}" + }, + { + "bbox": [ + 304, + 72, + 545, + 107 + ], + "type": "text", + "content": " and obtain the graph shown in Figure 7." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 107, + 547, + 224 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 304, + 107, + 547, + 154 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 107, + 547, + 154 + ], + "spans": [ + { + "bbox": [ + 304, + 107, + 547, + 154 + ], + "type": "text", + "content": "(3) Photometric error: we warp the first frame of an input video to the rest of the frames using the UV map estimates. We report the error between the ground truth RGB images and the warped RGB images as reported in Table 1." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 154, + 545, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 154, + 545, + 224 + ], + "spans": [ + { + "bbox": [ + 304, + 154, + 545, + 224 + ], + "type": "text", + "content": "(4) Geometric error: we report the " + }, + { + "bbox": [ + 304, + 154, + 545, + 224 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{geo}}" + }, + { + "bbox": [ + 304, + 154, + 545, + 224 + ], + "type": "text", + "content": " to show how the predicted UV map follows the geometric information captured in the surface normal estimates as reported in Table 1. (5) Temporal error: we evaluate the capability of the different approaches in preserving the temporal coherency by reporting the " + }, + { + "bbox": [ + 304, + 154, + 545, + 224 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{tmp}}" + }, + { + "bbox": [ + 304, + 154, + 545, + 224 + ], + "type": "text", + "content": " error (Table 1)." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 304, + 224, + 545, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 224, + 545, + 248 + ], + "spans": [ + { + "bbox": [ + 304, + 224, + 545, + 248 + ], + "type": "text", + "content": "When the ground truth UV is not available, we use the geometric and temporal errors to evaluate our method." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 251, + 545, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 251, + 545, + 286 + ], + "spans": [ + { + "bbox": [ + 304, + 251, + 545, + 286 + ], + "type": "text", + "content": "Baseline Methods We compare our method with previous works that fall into two categories: (1) human UV map prediction; (2) UV optimization." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 289, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 304, + 289, + 546, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 289, + 546, + 499 + ], + "spans": [ + { + "bbox": [ + 304, + 289, + 546, + 499 + ], + "type": "text", + "content": "1) Human UV map prediction: we compare our method with state-of-the-art that focus on predicting UV maps for the naked human body [18] and dressed humans [47, 55]. We also report the performance of the UV map obtained by extrapolating DensePose predictions as discussed in Section 3.3. Our method achieves the best performance as shown in Table 1 and Figure 7. We notice that DensePose [18] performs competently in precision percentage when the threshold error is less than " + }, + { + "bbox": [ + 304, + 289, + 546, + 499 + ], + "type": "inline_equation", + "content": "7\\mathrm{cm}" + }, + { + "bbox": [ + 304, + 289, + 546, + 499 + ], + "type": "text", + "content": ". This observation is based on the fact that, for each pixel, DensePose predicts a part label (among 24 parts) and a UV map with respect to that body part. When aligning these predictions with the ground truth, we warp each occupied body part individually, resulting in a more accurate alignment compared to the other methods (including ours) that are represented by only one patch. However, the performance of DensePose [18] is not improved above " + }, + { + "bbox": [ + 304, + 289, + 546, + 499 + ], + "type": "inline_equation", + "content": "7\\mathrm{cm}" + }, + { + "bbox": [ + 304, + 289, + 546, + 499 + ], + "type": "text", + "content": " because of limited ability to predict beyond body surface." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 502, + 547, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 502, + 547, + 595 + ], + "spans": [ + { + "bbox": [ + 304, + 502, + 547, + 595 + ], + "type": "text", + "content": "2) UV optimization: we compare our method with state-of-the-art in predicting the UV map of a dynamic object observed in a video [25, 57]. However, such methods are not tailored for garments that undergo highly nonlinear transformations as the body moves. Hence, as reported in Table 1 and Figure 7, our method surpasses these baselines in the dense UV error, the average precision percentage, the geometric error, and the temporal error." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 596, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 596, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 596, + 547, + 713 + ], + "type": "text", + "content": "Ablation Study We conduct an ablation study to analyze the impact of the distance (first two terms in Equation (7)) and angle (third term in Equation (7)) constraints (second and third rows of Table 2 and Figure 7). Our final method has the best performance in UV error and photometric error. We also compare the performance of our method without the temporal consistency " + }, + { + "bbox": [ + 304, + 596, + 547, + 713 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\mathrm{tmp}})" + }, + { + "bbox": [ + 304, + 596, + 547, + 713 + ], + "type": "text", + "content": " (fourth row of Table 2). As expected, this term performs quite similarly to ours in UV error but very poorly in photometric error since the consistency between the frames is not enforced. The role of" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "type": "text", + "content": "4631" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 41, + 547, + 152 + ], + "blocks": [ + { + "bbox": [ + 48, + 41, + 547, + 152 + ], + "lines": [ + { + "bbox": [ + 48, + 41, + 547, + 152 + ], + "spans": [ + { + "bbox": [ + 48, + 41, + 547, + 152 + ], + "type": "image", + "image_path": "c4350c841f8b3e0633b3d05bf94b2b0772b730aaeea84c7f34c68fddcd3f082c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 161, + 546, + 184 + ], + "lines": [ + { + "bbox": [ + 46, + 161, + 546, + 184 + ], + "spans": [ + { + "bbox": [ + 46, + 161, + 546, + 184 + ], + "type": "text", + "content": "Figure 5. Video framework. We apply temporal coherence in the predicted UV maps using optical flow [35,36,49]. With the optical flow " + }, + { + "bbox": [ + 46, + 161, + 546, + 184 + ], + "type": "inline_equation", + "content": "W_{i\\rightarrow j}" + }, + { + "bbox": [ + 46, + 161, + 546, + 184 + ], + "type": "text", + "content": ", we match the UV prediction of " + }, + { + "bbox": [ + 46, + 161, + 546, + 184 + ], + "type": "inline_equation", + "content": "i^{\\mathrm{th}}" + }, + { + "bbox": [ + 46, + 161, + 546, + 184 + ], + "type": "text", + "content": " frame with " + }, + { + "bbox": [ + 46, + 161, + 546, + 184 + ], + "type": "inline_equation", + "content": "j^{\\mathrm{th}}" + }, + { + "bbox": [ + 46, + 161, + 546, + 184 + ], + "type": "text", + "content": " via " + }, + { + "bbox": [ + 46, + 161, + 546, + 184 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{tmp}}" + }, + { + "bbox": [ + 46, + 161, + 546, + 184 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 47, + 194, + 562, + 304 + ], + "blocks": [ + { + "bbox": [ + 47, + 194, + 562, + 304 + ], + "lines": [ + { + "bbox": [ + 47, + 194, + 562, + 304 + ], + "spans": [ + { + "bbox": [ + 47, + 194, + 562, + 304 + ], + "type": "table", + "html": "
GT dress sequences [45]GT T-shirt sequences [45]Real Fashion sequences [59]
MethodUV. error (cm)photo. errorUV. error (cm)photo. errorgeo. error (Lgeo)tmp. error
DensePose [18]17.27±3.7652.43±10.207.48±0.5234.19±7.011.26±0.478.12±2.70
Extrapolated DensePose [18, 50]8.61±0.7618.29±3.245.34±0.6318.86±4.550.52±0.064.62±0.73
HumanGPS [47]11.97±2.0697.41±30.877.53±0.43108.10±34.891.27±0.6951.40±41.66
Kasten et al. [25]7.06±0.6013.07±2.796.64±0.6314.54±4.260.56±0.102.30±0.70
Ye et al. [57]5.56±0.2933.57±-10.275.75±0.2019.22±5.140.71±0.031.65±0.15
Ours3.16±0.287.54±2.043.58±0.2711.28±2.010.07±0.031.50±0.23
", + "image_path": "92f945a7d8ecbd7d85bb0a6f15d8de3bcee25b726c6cfdeddda02e15ff4edca8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 55, + 331, + 277, + 407 + ], + "blocks": [ + { + "bbox": [ + 46, + 309, + 546, + 331 + ], + "lines": [ + { + "bbox": [ + 46, + 309, + 546, + 331 + ], + "spans": [ + { + "bbox": [ + 46, + 309, + 546, + 331 + ], + "type": "text", + "content": "Table 1. Quantitative Results. UV. error (cm), photo. error (RGB difference), geo. error " + }, + { + "bbox": [ + 46, + 309, + 546, + 331 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\mathrm{geo}})" + }, + { + "bbox": [ + 46, + 309, + 546, + 331 + ], + "type": "text", + "content": " , and tmp. error " + }, + { + "bbox": [ + 46, + 309, + 546, + 331 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\mathrm{tmp}})" + }, + { + "bbox": [ + 46, + 309, + 546, + 331 + ], + "type": "text", + "content": " (image space pixel distance) respectively (mean±std)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 55, + 331, + 277, + 407 + ], + "lines": [ + { + "bbox": [ + 55, + 331, + 277, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 331, + 277, + 407 + ], + "type": "table", + "html": "
MethodUV. error (cm)photo. error (RGB)
Ours3.16±0.287.54±2.04
Distance constraint3.44±0.258.29±1.99
Angle constraint5.96±0.337.68±1.37
W/o Ltmp3.22±0.2414.67±4.82
W/o Lprox3.24±0.358.46±2.19
", + "image_path": "742b3a90ab56cdce922dd061a2f718e3f28a18e144d4d4f339410326f6281274.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 49, + 438, + 289, + 503 + ], + "blocks": [ + { + "bbox": [ + 49, + 438, + 289, + 503 + ], + "lines": [ + { + "bbox": [ + 49, + 438, + 289, + 503 + ], + "spans": [ + { + "bbox": [ + 49, + 438, + 289, + 503 + ], + "type": "image", + "image_path": "dd8bdaaa327e79b9acf2d672920507703edd830b2347aacd8cc811fa011efeeb.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 51, + 503, + 283, + 514 + ], + "lines": [ + { + "bbox": [ + 51, + 503, + 283, + 514 + ], + "spans": [ + { + "bbox": [ + 51, + 503, + 283, + 514 + ], + "type": "text", + "content": "Figure 6. The impact of " + }, + { + "bbox": [ + 51, + 503, + 283, + 514 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{prox}}" + }, + { + "bbox": [ + 51, + 503, + 283, + 514 + ], + "type": "text", + "content": " in UV prediction and retexturing." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 55, + 514, + 277, + 647 + ], + "blocks": [ + { + "bbox": [ + 55, + 514, + 277, + 647 + ], + "lines": [ + { + "bbox": [ + 55, + 514, + 277, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 514, + 277, + 647 + ], + "type": "image", + "image_path": "b797576192909f839a67b3eb1be501094e380a744fb0365a41dbc0d5f7e635b1.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 647, + 286, + 689 + ], + "lines": [ + { + "bbox": [ + 46, + 647, + 286, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 647, + 286, + 689 + ], + "type": "text", + "content": "Figure 7. Average precision. We compute the average of pixels with a UV error higher than a threshold (1-15cm) on our method, the baseline methods, and our ablation experiments on ground truth dress sequences [45]." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{prox}}" + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": " is the disambiguation of UV maps, i.e., there exist an infinite number of equivalent UV maps that minimize our" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 332, + 545, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 332, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 304, + 332, + 545, + 392 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{geo}}" + }, + { + "bbox": [ + 304, + 332, + 545, + 392 + ], + "type": "text", + "content": " (PDE). While the result without " + }, + { + "bbox": [ + 304, + 332, + 545, + 392 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{prox}}" + }, + { + "bbox": [ + 304, + 332, + 545, + 392 + ], + "type": "text", + "content": " are, therefore, quantitatively competitive as summarized in Table 2, such ambiguity can be resolved by finding UV that is closest to the proxy UV as shown in Figure 6, i.e., the retexture without " + }, + { + "bbox": [ + 304, + 332, + 545, + 392 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{prox}}" + }, + { + "bbox": [ + 304, + 332, + 545, + 392 + ], + "type": "text", + "content": " can result in arbitrary orientation across subjects." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 396, + 546, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 396, + 546, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 396, + 546, + 594 + ], + "type": "text", + "content": "Qualitative Results To show our results qualitatively, we visualize both re-texturing examples and grid UV illustration to depict the performance of each method in preserving high-frequency details. For retexturing, we first obtain an albedo and shading layer from the input image using an intrinsic image decomposition method [53]. After applying a new texture pattern to the albedo layer, we compose it back with the original shading layer. We apply Gamma-correction [37] on the input image, after generating the albedo layer and shading layer, we inverse the Gamma-correction back when synthesizing the re-textured image. We compare our method qualitatively with the baselines as shown in Figure 8 that illustrates the results on Fashion video sequence [59]. Figure 11 shows the performance of our method on videos and images. Our method not only captures the fine-grained surface details but also is temporally coherent across time." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 596, + 416, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 596, + 416, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 596, + 416, + 712 + ], + "type": "text", + "content": "User Study We conducted a user study: asking participants " + }, + { + "bbox": [ + 304, + 596, + 416, + 712 + ], + "type": "inline_equation", + "content": "(n = 21)" + }, + { + "bbox": [ + 304, + 596, + 416, + 712 + ], + "type": "text", + "content": " to rate realism (1: unrealistic to 10: most realistic) for our method compared to the baselines as summarized in Table 3. Our method receives the highest score from the users." + } + ] + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 423, + 609, + 535, + 687 + ], + "blocks": [ + { + "bbox": [ + 47, + 413, + 286, + 435 + ], + "lines": [ + { + "bbox": [ + 47, + 413, + 286, + 435 + ], + "spans": [ + { + "bbox": [ + 47, + 413, + 286, + 435 + ], + "type": "text", + "content": "Table 2. Ablation study on dress sequences [45]. UV. error (cm), photo. error (RGB difference) (mean±std)." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 423, + 609, + 535, + 687 + ], + "lines": [ + { + "bbox": [ + 423, + 609, + 535, + 687 + ], + "spans": [ + { + "bbox": [ + 423, + 609, + 535, + 687 + ], + "type": "table", + "html": "
MethodUser score
DensePose4.09±1.94
Proxy DensePose3.23±1.86
HumanGPS2.28±2.12
Kasten et al.5.52±1.88
Ye et al.7.66±1.27
Ours9.57±0.81
", + "image_path": "6bc06d65a26229954a48075ac1b5e217432187b6994c480c3dfd9c6d8a535152.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 453, + 689, + 527, + 700 + ], + "lines": [ + { + "bbox": [ + 453, + 689, + 527, + 700 + ], + "spans": [ + { + "bbox": [ + 453, + 689, + 527, + 700 + ], + "type": "text", + "content": "Table 3. User study." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4632" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 38, + 115, + 145 + ], + "blocks": [ + { + "bbox": [ + 61, + 38, + 115, + 145 + ], + "lines": [ + { + "bbox": [ + 61, + 38, + 115, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 38, + 115, + 145 + ], + "type": "image", + "image_path": "2be33698c31c3356af08f6606fee2bcdcd11f53e86467fa4d6248403c57e307c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 63, + 147, + 88, + 168 + ], + "lines": [ + { + "bbox": [ + 63, + 147, + 88, + 168 + ], + "spans": [ + { + "bbox": [ + 63, + 147, + 88, + 168 + ], + "type": "text", + "content": "Input image" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 115, + 40, + 149, + 145 + ], + "blocks": [ + { + "bbox": [ + 115, + 40, + 149, + 145 + ], + "lines": [ + { + "bbox": [ + 115, + 40, + 149, + 145 + ], + "spans": [ + { + "bbox": [ + 115, + 40, + 149, + 145 + ], + "type": "image", + "image_path": "66871bc101a26ecb8b3f4d6653c161726106cb1e107bda0a291c1eb9ee00c803.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 116, + 147, + 148, + 168 + ], + "lines": [ + { + "bbox": [ + 116, + 147, + 148, + 168 + ], + "spans": [ + { + "bbox": [ + 116, + 147, + 148, + 168 + ], + "type": "text", + "content": "Ours w/shading" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 155, + 38, + 208, + 145 + ], + "blocks": [ + { + "bbox": [ + 155, + 38, + 208, + 145 + ], + "lines": [ + { + "bbox": [ + 155, + 38, + 208, + 145 + ], + "spans": [ + { + "bbox": [ + 155, + 38, + 208, + 145 + ], + "type": "image", + "image_path": "d581e1ceaee474d287393955d80892872e26bee9360bdaa30ebf2ab98b1b5ee5.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 155, + 147, + 206, + 188 + ], + "blocks": [ + { + "bbox": [ + 155, + 147, + 206, + 188 + ], + "lines": [ + { + "bbox": [ + 155, + 147, + 206, + 188 + ], + "spans": [ + { + "bbox": [ + 155, + 147, + 206, + 188 + ], + "type": "image", + "image_path": "f0c59451cfbdcc3a274f87dfa7ac1c8bfcd3a89e7cd745ab5158e717513a4d6c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 156, + 189, + 190, + 238 + ], + "blocks": [ + { + "bbox": [ + 156, + 189, + 190, + 238 + ], + "lines": [ + { + "bbox": [ + 156, + 189, + 190, + 238 + ], + "spans": [ + { + "bbox": [ + 156, + 189, + 190, + 238 + ], + "type": "image", + "image_path": "fc0e6518f4d8a9998b027713f4c5ab93506f51601dd2ba617b250824658b93cd.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 205, + 38, + 257, + 145 + ], + "blocks": [ + { + "bbox": [ + 205, + 38, + 257, + 145 + ], + "lines": [ + { + "bbox": [ + 205, + 38, + 257, + 145 + ], + "spans": [ + { + "bbox": [ + 205, + 38, + 257, + 145 + ], + "type": "image", + "image_path": "e2e21bd3012fc32ac4a0e6975f2a732e9901c2fa48cb1629b7dd1b45d6b3ffba.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 206, + 147, + 257, + 188 + ], + "blocks": [ + { + "bbox": [ + 206, + 147, + 257, + 188 + ], + "lines": [ + { + "bbox": [ + 206, + 147, + 257, + 188 + ], + "spans": [ + { + "bbox": [ + 206, + 147, + 257, + 188 + ], + "type": "image", + "image_path": "6c0779eccc17662bfbbe9b586320103499e2652b4bc3b2a932791d8815d1514f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 206, + 189, + 252, + 238 + ], + "blocks": [ + { + "bbox": [ + 206, + 189, + 252, + 238 + ], + "lines": [ + { + "bbox": [ + 206, + 189, + 252, + 238 + ], + "spans": [ + { + "bbox": [ + 206, + 189, + 252, + 238 + ], + "type": "image", + "image_path": "be2b0b0ca24b073d82d227c6d888f92ed75effcf4792672b89b5ab2959ba482f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 257, + 38, + 310, + 145 + ], + "blocks": [ + { + "bbox": [ + 257, + 38, + 310, + 145 + ], + "lines": [ + { + "bbox": [ + 257, + 38, + 310, + 145 + ], + "spans": [ + { + "bbox": [ + 257, + 38, + 310, + 145 + ], + "type": "image", + "image_path": "d008d2f38e6805cbc8eb593a5f61d2f85815d4f97cc02fad91d5c3408faaa0c4.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 257, + 147, + 309, + 188 + ], + "blocks": [ + { + "bbox": [ + 257, + 147, + 309, + 188 + ], + "lines": [ + { + "bbox": [ + 257, + 147, + 309, + 188 + ], + "spans": [ + { + "bbox": [ + 257, + 147, + 309, + 188 + ], + "type": "image", + "image_path": "4c39a9ca66fc27ffe3af2a28983ef9b2929a40ecae1af04062e5fc8ba26c5bc0.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 258, + 189, + 298, + 238 + ], + "blocks": [ + { + "bbox": [ + 258, + 189, + 298, + 238 + ], + "lines": [ + { + "bbox": [ + 258, + 189, + 298, + 238 + ], + "spans": [ + { + "bbox": [ + 258, + 189, + 298, + 238 + ], + "type": "image", + "image_path": "31e635c011324386602f02f02738f49b9170fc9263875e6da0a5d0c72addafe4.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 309, + 38, + 359, + 145 + ], + "blocks": [ + { + "bbox": [ + 309, + 38, + 359, + 145 + ], + "lines": [ + { + "bbox": [ + 309, + 38, + 359, + 145 + ], + "spans": [ + { + "bbox": [ + 309, + 38, + 359, + 145 + ], + "type": "image", + "image_path": "480086e898fe57b0e7973d780db1ed696bac5753de7f5652f55c4072b874b632.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 309, + 147, + 354, + 188 + ], + "blocks": [ + { + "bbox": [ + 309, + 147, + 354, + 188 + ], + "lines": [ + { + "bbox": [ + 309, + 147, + 354, + 188 + ], + "spans": [ + { + "bbox": [ + 309, + 147, + 354, + 188 + ], + "type": "image", + "image_path": "7770163965a7cd33562533ad2c386eb95d1be5579eda984eac87a87b1eeaf0a5.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 309, + 189, + 342, + 238 + ], + "blocks": [ + { + "bbox": [ + 309, + 189, + 342, + 238 + ], + "lines": [ + { + "bbox": [ + 309, + 189, + 342, + 238 + ], + "spans": [ + { + "bbox": [ + 309, + 189, + 342, + 238 + ], + "type": "image", + "image_path": "c7e489fc13df2fd43159c6b68d6d5499c7c239818614422c76276846440c6232.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 357, + 38, + 410, + 145 + ], + "blocks": [ + { + "bbox": [ + 357, + 38, + 410, + 145 + ], + "lines": [ + { + "bbox": [ + 357, + 38, + 410, + 145 + ], + "spans": [ + { + "bbox": [ + 357, + 38, + 410, + 145 + ], + "type": "image", + "image_path": "45c683ea33a5861ea29c689615bebe4058a8b6a8947ea501827d7a6d1ab4eeb4.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 357, + 147, + 410, + 238 + ], + "blocks": [ + { + "bbox": [ + 357, + 147, + 410, + 238 + ], + "lines": [ + { + "bbox": [ + 357, + 147, + 410, + 238 + ], + "spans": [ + { + "bbox": [ + 357, + 147, + 410, + 238 + ], + "type": "image", + "image_path": "ebd15b28592d60249c8672f975e629a67e09967e203f1a94ef7fa2ae1957f242.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 410, + 38, + 460, + 145 + ], + "blocks": [ + { + "bbox": [ + 410, + 38, + 460, + 145 + ], + "lines": [ + { + "bbox": [ + 410, + 38, + 460, + 145 + ], + "spans": [ + { + "bbox": [ + 410, + 38, + 460, + 145 + ], + "type": "image", + "image_path": "b259f90d519ef78375436ce17d19d3173236c052786c59e33411daebc71f853f.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 460, + 38, + 512, + 145 + ], + "blocks": [ + { + "bbox": [ + 460, + 38, + 512, + 145 + ], + "lines": [ + { + "bbox": [ + 460, + 38, + 512, + 145 + ], + "spans": [ + { + "bbox": [ + 460, + 38, + 512, + 145 + ], + "type": "image", + "image_path": "886860ca57c58c90863285f2d7d3a439d83ca6eb8e7adb35d0e5c42fd58eef4d.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 514, + 38, + 525, + 133 + ], + "lines": [ + { + "bbox": [ + 514, + 38, + 525, + 133 + ], + "spans": [ + { + "bbox": [ + 514, + 38, + 525, + 133 + ], + "type": "text", + "content": "Retturung w/o shading" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 515, + 167, + 527, + 202 + ], + "lines": [ + { + "bbox": [ + 515, + 167, + 527, + 202 + ], + "spans": [ + { + "bbox": [ + 515, + 167, + 527, + 202 + ], + "type": "text", + "content": "Grid-UV" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 149, + 243, + 190, + 319 + ], + "blocks": [ + { + "bbox": [ + 129, + 243, + 148, + 254 + ], + "lines": [ + { + "bbox": [ + 129, + 243, + 148, + 254 + ], + "spans": [ + { + "bbox": [ + 129, + 243, + 148, + 254 + ], + "type": "text", + "content": "High" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 149, + 243, + 190, + 319 + ], + "lines": [ + { + "bbox": [ + 149, + 243, + 190, + 319 + ], + "spans": [ + { + "bbox": [ + 149, + 243, + 190, + 319 + ], + "type": "image", + "image_path": "6aa03086274b31b936d5a616bcab1a8c23dda42eda1a830ecc0d5c3c989134f8.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 163, + 321, + 181, + 331 + ], + "lines": [ + { + "bbox": [ + 163, + 321, + 181, + 331 + ], + "spans": [ + { + "bbox": [ + 163, + 321, + 181, + 331 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 201, + 244, + 236, + 318 + ], + "blocks": [ + { + "bbox": [ + 201, + 244, + 236, + 318 + ], + "lines": [ + { + "bbox": [ + 201, + 244, + 236, + 318 + ], + "spans": [ + { + "bbox": [ + 201, + 244, + 236, + 318 + ], + "type": "image", + "image_path": "252a6aee4920ca0aed033b349cc7d82275495fbd6ddd63209881e0e9da6a85fe.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 195, + 320, + 241, + 336 + ], + "lines": [ + { + "bbox": [ + 195, + 320, + 241, + 336 + ], + "spans": [ + { + "bbox": [ + 195, + 320, + 241, + 336 + ], + "type": "text", + "content": "Extrapolated DensePose" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 258, + 243, + 290, + 319 + ], + "blocks": [ + { + "bbox": [ + 258, + 243, + 290, + 319 + ], + "lines": [ + { + "bbox": [ + 258, + 243, + 290, + 319 + ], + "spans": [ + { + "bbox": [ + 258, + 243, + 290, + 319 + ], + "type": "image", + "image_path": "ec66ecb3b0f900893d82e8cbae0b2513d8b519fe2f2a0187dc490b0157ee9ac9.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 255, + 322, + 294, + 331 + ], + "lines": [ + { + "bbox": [ + 255, + 322, + 294, + 331 + ], + "spans": [ + { + "bbox": [ + 255, + 322, + 294, + 331 + ], + "type": "text", + "content": "DensePose" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 309, + 246, + 342, + 318 + ], + "blocks": [ + { + "bbox": [ + 309, + 246, + 342, + 318 + ], + "lines": [ + { + "bbox": [ + 309, + 246, + 342, + 318 + ], + "spans": [ + { + "bbox": [ + 309, + 246, + 342, + 318 + ], + "type": "image", + "image_path": "3f1dce994bbdac0e06ee8e167c6af5d7aeea9107c369776e4597a42534143776.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 303, + 322, + 346, + 331 + ], + "lines": [ + { + "bbox": [ + 303, + 322, + 346, + 331 + ], + "spans": [ + { + "bbox": [ + 303, + 322, + 346, + 331 + ], + "type": "text", + "content": "HumanGPS" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 361, + 246, + 392, + 317 + ], + "blocks": [ + { + "bbox": [ + 361, + 246, + 392, + 317 + ], + "lines": [ + { + "bbox": [ + 361, + 246, + 392, + 317 + ], + "spans": [ + { + "bbox": [ + 361, + 246, + 392, + 317 + ], + "type": "image", + "image_path": "52c4afc0f43bc7319342125e37437841ef402d7c1f15ca60ed425f231836d7b9.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 356, + 322, + 403, + 331 + ], + "lines": [ + { + "bbox": [ + 356, + 322, + 403, + 331 + ], + "spans": [ + { + "bbox": [ + 356, + 322, + 403, + 331 + ], + "type": "text", + "content": "TemporalUV" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 413, + 246, + 444, + 317 + ], + "blocks": [ + { + "bbox": [ + 413, + 246, + 444, + 317 + ], + "lines": [ + { + "bbox": [ + 413, + 246, + 444, + 317 + ], + "spans": [ + { + "bbox": [ + 413, + 246, + 444, + 317 + ], + "type": "image", + "image_path": "d3ecf278b26fc169e62faa97d61bd9865174cf75556f3aea47de03ae9fb71706.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 410, + 322, + 454, + 331 + ], + "lines": [ + { + "bbox": [ + 410, + 322, + 454, + 331 + ], + "spans": [ + { + "bbox": [ + 410, + 322, + 454, + 331 + ], + "type": "text", + "content": "Kasten et al." + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 463, + 245, + 496, + 316 + ], + "blocks": [ + { + "bbox": [ + 463, + 245, + 496, + 316 + ], + "lines": [ + { + "bbox": [ + 463, + 245, + 496, + 316 + ], + "spans": [ + { + "bbox": [ + 463, + 245, + 496, + 316 + ], + "type": "image", + "image_path": "fba2ca528dfd684478828c0914d5bff2aca6ab427293ed6e6b7c066618b59c0c.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 465, + 321, + 494, + 331 + ], + "lines": [ + { + "bbox": [ + 465, + 321, + 494, + 331 + ], + "spans": [ + { + "bbox": [ + 465, + 321, + 494, + 331 + ], + "type": "text", + "content": "Ye et al." + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 50, + 358, + 288, + 514 + ], + "blocks": [ + { + "bbox": [ + 47, + 336, + 545, + 357 + ], + "lines": [ + { + "bbox": [ + 47, + 336, + 545, + 357 + ], + "spans": [ + { + "bbox": [ + 47, + 336, + 545, + 357 + ], + "type": "text", + "content": "Figure 8. Qualitative comparison. We compare ours with our initial UV (extrapolated DensePose), DensePose [18], HumanGPS [47], Kastenn et al. [25], and Ye et al. [57] in image re-texturing and UV map grid visualization." + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 50, + 358, + 288, + 514 + ], + "lines": [ + { + "bbox": [ + 50, + 358, + 288, + 514 + ], + "spans": [ + { + "bbox": [ + 50, + 358, + 288, + 514 + ], + "type": "image", + "image_path": "e4fd2d54886b0f66d565fce9ee3e2eb272487b2d460c1cfe29bf7ba25a57fcb0.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 519, + 287, + 562 + ], + "lines": [ + { + "bbox": [ + 47, + 519, + 287, + 562 + ], + "spans": [ + { + "bbox": [ + 47, + 519, + 287, + 562 + ], + "type": "text", + "content": "Figure 9. Limitation. Due to lack of 3D reconstructed surface, our method cannot handle folds that introduce texture discontinuity. Our texture map is continuous around the folds, which is physically incorrect." + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + } + ], + "index": 38 + }, + { + "bbox": [ + 47, + 563, + 116, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 563, + 116, + 575 + ], + "spans": [ + { + "bbox": [ + 47, + 563, + 116, + 575 + ], + "type": "text", + "content": "5. Discussion" + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 46, + 583, + 287, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 583, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 583, + 287, + 700 + ], + "type": "text", + "content": "This paper presents a novel approach to predict a high quality UV map by preserving geometric details from images and videos. We leverage the geometric property of isometry encoded in 3D surface normals to optimize the UV map in the form of partial differential equations. We generalize our method to videos by integrating optical flow, resulting in a temporally coherent video editing. Our method produces strong qualitative and quantitative predictions on real-world imagery compared to state-of-the-art UV map estimation." + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 47, + 701, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 287, + 712 + ], + "type": "text", + "content": "Limitation As discussed in Section 3.1, our method makes" + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 304, + 360, + 545, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 360, + 545, + 500 + ], + "spans": [ + { + "bbox": [ + 304, + 360, + 545, + 500 + ], + "type": "text", + "content": "an assumption about projection, i.e., there is one-to-one correspondence between 3D surface geometry and image. However, this assumption does not hold when there is a fold where a region of 3D surface is not visible to the image. This makes a contrast with 3D reconstruction based method where the invisible part of 3D surface can be still mapped to the image via depth reasoning. Figure 9 illustrates this limitation where there are folds in the skirt, resulting in negative surface normal " + }, + { + "bbox": [ + 304, + 360, + 545, + 500 + ], + "type": "inline_equation", + "content": "n_z" + }, + { + "bbox": [ + 304, + 360, + 545, + 500 + ], + "type": "text", + "content": ". Due to the folds, the texture must be discontinuous while our method produces continuous texture rendering due to the lack of 3D reconstructed geometry, which is physically incorrect." + } + ] + } + ], + "index": 43 + }, + { + "bbox": [ + 304, + 501, + 404, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 501, + 404, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 501, + 404, + 628 + ], + "type": "text", + "content": "Our pipeline is composed of two components: (1) UV prediction (our contribution) and (2) texture map with shading (not our contribution). For the garments with highly contrasted texture, the shading operation is often biased to color con" + } + ] + } + ], + "index": 44 + }, + { + "type": "image", + "bbox": [ + 406, + 502, + 453, + 590 + ], + "blocks": [ + { + "bbox": [ + 406, + 502, + 453, + 590 + ], + "lines": [ + { + "bbox": [ + 406, + 502, + 453, + 590 + ], + "spans": [ + { + "bbox": [ + 406, + 502, + 453, + 590 + ], + "type": "image", + "image_path": "159478e428bec14296451264dcd57ed3cbad387285f19a63561be588e0284499.jpg" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 420, + 594, + 442, + 602 + ], + "lines": [ + { + "bbox": [ + 420, + 594, + 442, + 602 + ], + "spans": [ + { + "bbox": [ + 420, + 594, + 442, + 602 + ], + "type": "text", + "content": "Image" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 423, + 609, + 545, + 630 + ], + "lines": [ + { + "bbox": [ + 423, + 609, + 545, + 630 + ], + "spans": [ + { + "bbox": [ + 423, + 609, + 545, + 630 + ], + "type": "text", + "content": "Figure 10. Limitation. Limitations on textured garment." + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_caption" + } + ], + "index": 45 + }, + { + "type": "image", + "bbox": [ + 454, + 502, + 498, + 589 + ], + "blocks": [ + { + "bbox": [ + 454, + 502, + 498, + 589 + ], + "lines": [ + { + "bbox": [ + 454, + 502, + 498, + 589 + ], + "spans": [ + { + "bbox": [ + 454, + 502, + 498, + 589 + ], + "type": "image", + "image_path": "4802e97fdfc756fa95ab8753afcafd6f5a35e50bf28dc2ae7ff181112f5568e9.jpg" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 454, + 589, + 493, + 608 + ], + "lines": [ + { + "bbox": [ + 454, + 589, + 493, + 608 + ], + "spans": [ + { + "bbox": [ + 454, + 589, + 493, + 608 + ], + "type": "text", + "content": "Retexturing w/o shading" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_caption" + } + ], + "index": 47 + }, + { + "type": "image", + "bbox": [ + 498, + 502, + 543, + 589 + ], + "blocks": [ + { + "bbox": [ + 498, + 502, + 543, + 589 + ], + "lines": [ + { + "bbox": [ + 498, + 502, + 543, + 589 + ], + "spans": [ + { + "bbox": [ + 498, + 502, + 543, + 589 + ], + "type": "image", + "image_path": "2438e0cbecb0eb5827c2d9afeef9573c31459cc756ff01e6c3888cf68534bcf2.jpg" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 499, + 589, + 539, + 608 + ], + "lines": [ + { + "bbox": [ + 499, + 589, + 539, + 608 + ], + "spans": [ + { + "bbox": [ + 499, + 589, + 539, + 608 + ], + "type": "text", + "content": "Retexturing w/shading" + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_caption" + } + ], + "index": 49 + }, + { + "bbox": [ + 304, + 630, + 545, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 687 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 687 + ], + "type": "text", + "content": "trast, resulting in erroneous appearance. Figure 10 illustrates that despite reasonable UV prediction from our method, the resulting appearance is unrealistic near the textured region. Improving the shading operation is beyond the scope of this work, and we leave it as future work." + } + ] + } + ], + "index": 52 + }, + { + "bbox": [ + 305, + 688, + 545, + 710 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 688, + 545, + 710 + ], + "spans": [ + { + "bbox": [ + 305, + 688, + 545, + 710 + ], + "type": "text", + "content": "Acknowledgement This work was supported by a NSF NRI 2022894 and NSF CAREER 1846031." + } + ] + } + ], + "index": 53 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4633" + } + ] + } + ], + "index": 54 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 132, + 217 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 132, + 217 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 132, + 217 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 132, + 217 + ], + "type": "image", + "image_path": "cbde4900b39fdffb54508ce11134874a813903016d4d62681fbe926833fe5b8a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 134, + 70, + 544, + 363 + ], + "blocks": [ + { + "bbox": [ + 134, + 70, + 544, + 363 + ], + "lines": [ + { + "bbox": [ + 134, + 70, + 544, + 363 + ], + "spans": [ + { + "bbox": [ + 134, + 70, + 544, + 363 + ], + "type": "image", + "image_path": "a307d949748ced08d0937bc708ca36f9ab6ff6ce04435e747fecf83350b8ee65.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 135, + 363, + 544, + 373 + ], + "lines": [ + { + "bbox": [ + 135, + 363, + 544, + 373 + ], + "spans": [ + { + "bbox": [ + 135, + 363, + 544, + 373 + ], + "type": "text", + "content": "Re-textured images" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 48, + 217, + 132, + 362 + ], + "blocks": [ + { + "bbox": [ + 48, + 217, + 132, + 362 + ], + "lines": [ + { + "bbox": [ + 48, + 217, + 132, + 362 + ], + "spans": [ + { + "bbox": [ + 48, + 217, + 132, + 362 + ], + "type": "image", + "image_path": "59721a2c8d4eb4d13fcf13229abb841d62e65fb4e76329bfa4f28e6df9dc7f46.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 57, + 362, + 123, + 373 + ], + "lines": [ + { + "bbox": [ + 57, + 362, + 123, + 373 + ], + "spans": [ + { + "bbox": [ + 57, + 362, + 123, + 373 + ], + "type": "text", + "content": "Original image" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 49, + 387, + 207, + 486 + ], + "blocks": [ + { + "bbox": [ + 49, + 387, + 207, + 486 + ], + "lines": [ + { + "bbox": [ + 49, + 387, + 207, + 486 + ], + "spans": [ + { + "bbox": [ + 49, + 387, + 207, + 486 + ], + "type": "image", + "image_path": "0a6c659cd622c567deeba997a3a30f4966357ec89a395d236c3ad9da5b82ddd3.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 78, + 487, + 185, + 640 + ], + "blocks": [ + { + "bbox": [ + 78, + 487, + 185, + 640 + ], + "lines": [ + { + "bbox": [ + 78, + 487, + 185, + 640 + ], + "spans": [ + { + "bbox": [ + 78, + 487, + 185, + 640 + ], + "type": "image", + "image_path": "c5317c17f98d04683b39a8dfa7cca28c755a9f7206c2c4f413cb7054b192a6cc.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 106, + 641, + 163, + 652 + ], + "lines": [ + { + "bbox": [ + 106, + 641, + 163, + 652 + ], + "spans": [ + { + "bbox": [ + 106, + 641, + 163, + 652 + ], + "type": "text", + "content": "Original video" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 210, + 387, + 545, + 652 + ], + "blocks": [ + { + "bbox": [ + 217, + 376, + 378, + 386 + ], + "lines": [ + { + "bbox": [ + 217, + 376, + 378, + 386 + ], + "spans": [ + { + "bbox": [ + 217, + 376, + 378, + 386 + ], + "type": "text", + "content": "(a) Qualitative results of our method on an image." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 210, + 387, + 545, + 652 + ], + "lines": [ + { + "bbox": [ + 210, + 387, + 545, + 652 + ], + "spans": [ + { + "bbox": [ + 210, + 387, + 545, + 652 + ], + "type": "image", + "image_path": "ddac916919564fdf45bfa6f928c0f0a2e0ef28c8ee9e740378d233f12a9ec5d0.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 220, + 657, + 373, + 666 + ], + "lines": [ + { + "bbox": [ + 220, + 657, + 373, + 666 + ], + "spans": [ + { + "bbox": [ + 220, + 657, + 373, + 666 + ], + "type": "text", + "content": "(b) Qualitative results of our method on videos." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 676, + 546, + 698 + ], + "lines": [ + { + "bbox": [ + 46, + 676, + 546, + 698 + ], + "spans": [ + { + "bbox": [ + 46, + 676, + 546, + 698 + ], + "type": "text", + "content": "Figure 11. Results on images and videos. Our method can produce compelling texture editing given images and videos that preserve geometric details." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "4634" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 286, + 712 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 53, + 91, + 208, + 102 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 208, + 102 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 208, + 102 + ], + "type": "text", + "content": "[1] https://renderpeople.com/3d-people.2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 104, + 286, + 125 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 104, + 286, + 125 + ], + "spans": [ + { + "bbox": [ + 53, + 104, + 286, + 125 + ], + "type": "text", + "content": "[2] Abien Fred Agarap. Deep learning using rectified linear units (relu). arXiv, 2018. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 127, + 286, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 127, + 286, + 170 + ], + "spans": [ + { + "bbox": [ + 53, + 127, + 286, + 170 + ], + "type": "text", + "content": "[3] Badour AlBahar, Jingwan Lu, Jimei Yang, Zhixin Shu, Eli Shechtman, and Jia-Bin Huang. Pose with Style: Detail-preserving pose-guided image synthesis with conditional stylegan. TOG, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 171, + 286, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 171, + 286, + 203 + ], + "spans": [ + { + "bbox": [ + 53, + 171, + 286, + 203 + ], + "type": "text", + "content": "[4] Yazeed Alharbi and Peter Wonka. Disentangled image generation through structured noise injection. In CVPR, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 205, + 286, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 205, + 286, + 247 + ], + "spans": [ + { + "bbox": [ + 53, + 205, + 286, + 247 + ], + "type": "text", + "content": "[5] David Bau, Hendrik Strobelt, William Peebles, Jonas Wulff, Bolei Zhou, Jun-Yan Zhu, and Antonio Torralba. Semantic photo manipulation with a generative image prior. TOG, 2019. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 250, + 286, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 250, + 286, + 282 + ], + "spans": [ + { + "bbox": [ + 53, + 250, + 286, + 282 + ], + "type": "text", + "content": "[6] Bharat Lal Bhatnagar, Garvita Tiwari, Christian Theobalt, and Gerard Pons-Moll. Multi-garment net: Learning to dress 3d people from images. In ICCV, 2019. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 284, + 286, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 284, + 286, + 304 + ], + "spans": [ + { + "bbox": [ + 53, + 284, + 286, + 304 + ], + "type": "text", + "content": "[7] Michael J. Black and P. Anandan. Robust dynamic motion estimation over time. In CVPR, 1991. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 307, + 286, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 307, + 286, + 328 + ], + "spans": [ + { + "bbox": [ + 53, + 307, + 286, + 328 + ], + "type": "text", + "content": "[8] Gabriel J Brostow and Irfan A Essa. Motion based decompositing of video. In ICCV, 1999. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 331, + 286, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 331, + 286, + 361 + ], + "spans": [ + { + "bbox": [ + 53, + 331, + 286, + 361 + ], + "type": "text", + "content": "[9] Felix E. Browder. On the unification of the calculus of variations and the theory of monotone nonlinear operators in banach spaces. PNAS, 1966. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 364, + 286, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 364, + 286, + 385 + ], + "spans": [ + { + "bbox": [ + 48, + 364, + 286, + 385 + ], + "type": "text", + "content": "[10] Edwin Earl Catmull. A subdivision algorithm for computer display of curved surfaces. The University of Utah, 1974. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 387, + 286, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 387, + 286, + 419 + ], + "spans": [ + { + "bbox": [ + 48, + 387, + 286, + 419 + ], + "type": "text", + "content": "[11] Shu-Yu Chen, Wanchao Su, Lin Gao, Shihong Xia, and Hongbo Fu. Deepfacedrawing: Deep generation of face images from sketches. TOG, 2020. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 421, + 286, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 421, + 286, + 453 + ], + "spans": [ + { + "bbox": [ + 48, + 421, + 286, + 453 + ], + "type": "text", + "content": "[12] Anton Cherepkov, Andrey Voynov, and Artem Babenko. Navigating the gan parameter space for semantic image editing. In CVPR, 2021. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 455, + 286, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 455, + 286, + 486 + ], + "spans": [ + { + "bbox": [ + 48, + 455, + 286, + 486 + ], + "type": "text", + "content": "[13] Edo Collins, Raja Bala, Bob Price, and Sabine Süssstrunk. Editing in style: Uncovering the local semantics of GANs. In CVPR, 2020. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 488, + 286, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 488, + 286, + 520 + ], + "spans": [ + { + "bbox": [ + 48, + 488, + 286, + 520 + ], + "type": "text", + "content": "[14] T. Darrell and A. Pentland. Robust estimation of a multilayered motion representation. In Proceedings of the IEEE Workshop on Visual Motion, 1991. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 522, + 286, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 522, + 286, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 522, + 286, + 555 + ], + "type": "text", + "content": "[15] Ke Gong, Yiming Gao, Xiaodan Liang, Xiaohui Shen, Meng Wang, and Liang Lin. Graphonomy: Universal human parsing via graph transfer learning. In CVPR, 2019. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 556, + 286, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 556, + 286, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 556, + 286, + 578 + ], + "type": "text", + "content": "[16] J.C. Gower and G.B. Dijksterhuis. Procrustes problems. new york: Oxford university press. Psychometrika, 70, 2005. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 579, + 286, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 579, + 286, + 611 + ], + "spans": [ + { + "bbox": [ + 48, + 579, + 286, + 611 + ], + "type": "text", + "content": "[17] Artur Grigorev, Artem Sevastopolsky, Alexander Vakhitov, and Victor Lempitsky. Coordinate-based texture inpainting for pose-guided human image generation. In CVPR, 2019. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 48, + 613, + 286, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 286, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 286, + 644 + ], + "type": "text", + "content": "[18] Riza Alp Güler, Natalia Neverova, and Iasonas Kokkinos. Densesepos: Dense human pose estimation in the wild. In CVPR, 2018. 2, 4, 5, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 48, + 647, + 286, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 286, + 677 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 286, + 677 + ], + "type": "text", + "content": "[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 48, + 681, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 286, + 712 + ], + "type": "text", + "content": "[20] Zhenliang He, Wangmeng Zuo, Meina Kan, Shiguang Shan, and Xilin Chen. Attgan: Facial attribute editing by only changing what you want. TIP, 2019. 2" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 42, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "type": "text", + "content": "[21] Xianxu Hou, Xiaokang Zhang, Linlin Shen, Zhihui Lai, and Jun Wan. Guidedstyle: Attribute knowledge guided style manipulation for semantic face editing. arXiv, 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 107, + 545, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 138 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 138 + ], + "type": "text", + "content": "[22] Anastasia Ianina, Nikolaos Sarafianos, Yuanlu Xu, Ignacio Rocco, and Tony Tung. Bodymap: Learning full-body dense correspondence map. In CVPR, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 141, + 545, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 171 + ], + "type": "text", + "content": "[23] Yasamin Jafarian and Hyun Soo Park. Learning high fidelity depths of dressed humans by watching social media dance videos. In CVPR, 2021. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 175, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 195 + ], + "type": "text", + "content": "[24] Varun Jampani, Raghudeep Gadde, and Peter V. Gehler. Video propagation networks. In CVPR, 2017. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 198, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 198, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 198, + 545, + 228 + ], + "type": "text", + "content": "[25] Yoni Kasten, Dolev Ofri, Oliver Wang, and Tali Dekel. Layered neural atlases for consistent video editing. TOG, 2021. 2, 5, 6, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 232, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 232, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 232, + 545, + 262 + ], + "type": "text", + "content": "[26] Hyunsu Kim, Yunjey Choi, Junho Kim, Sungwoo Yoo, and Youngjung Uh. Exploiting spatial dimensions of latent in gan for real-time image editing. In CVPR, 2021. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 265, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 265, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 265, + 545, + 285 + ], + "type": "text", + "content": "[27] Diederik P. Kingma and Jimmy Lei Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 288, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 288, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 288, + 545, + 319 + ], + "type": "text", + "content": "[28] Nilesh Kulkarni, Abhinav Gupta, David F. Fouhey, and Shubham Tulsiani. Articulation-aware canonical surface mapping. In CVPR, 2020. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 321, + 545, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 321, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 307, + 321, + 545, + 353 + ], + "type": "text", + "content": "[29] Nilesh Kulkarni, Abhinav Gupta, David F Fouhey, and Shubham Tulsiani. Articulation-aware canonical surface mapping. In CVPR, 2020. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 355, + 545, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 355, + 545, + 385 + ], + "spans": [ + { + "bbox": [ + 307, + 355, + 545, + 385 + ], + "type": "text", + "content": "[30] M. Pawan Kumar, Philip H. S. Torr, and Andrew Zisserman. Learning layered motion segmentations of video. IJCV, 2005. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 388, + 545, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 545, + 420 + ], + "type": "text", + "content": "[31] Verica Lazova, Eldar Insafutdinov, and Gerard Pons-Moll. 360-degree textures of people in clothing from a single image. 3DV, 2019. 2" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 422, + 545, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 422, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 307, + 422, + 545, + 453 + ], + "type": "text", + "content": "[32] Cheng-Han Lee, Ziwei Liu, Lingyun Wu, and Ping Luo. Maskgan: Towards diverse and interactive facial image manipulation. In CVPR, 2020. 2" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 307, + 456, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 456, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 307, + 456, + 545, + 487 + ], + "type": "text", + "content": "[33] Huan Ling, Karsten Kreis, Daiqing Li, Seung Wook Kim, Antonio Torralba, and Sanja Fidler. Editgan: High-precision semantic image editing. In NeurIPS, 2021. 2" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 307, + 490, + 545, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 490, + 545, + 531 + ], + "spans": [ + { + "bbox": [ + 307, + 490, + 545, + 531 + ], + "type": "text", + "content": "[34] Erika Lu, Forrester Cole, Tali Dekel, Weidi Xie, Andrew Zisserman, David Salesin, William T. Freeman, and Michael Rubinstein. Layered neural rendering for retiming people in video. TOG, 2020. 2" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 307, + 534, + 545, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 534, + 545, + 565 + ], + "spans": [ + { + "bbox": [ + 307, + 534, + 545, + 565 + ], + "type": "text", + "content": "[35] Bruce D. Lucas and Takeo Kanade. An iterative image registration technique with an application to stereo vision. In IJCAI, 1981. 6" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 307, + 568, + 545, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 568, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 307, + 568, + 545, + 589 + ], + "type": "text", + "content": "[36] Bruce D. Lucas and Takeo Kanade. Optical navigation by the method of differences. In *IJCAI*, 1985. 6" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 307, + 591, + 545, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 591, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 307, + 591, + 545, + 612 + ], + "type": "text", + "content": "[37] Tom McReynolds and David Blythe. Advanced graphics programming using OpenGL. Elsevier, 2005. 6" + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 307, + 614, + 545, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 614, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 307, + 614, + 545, + 644 + ], + "type": "text", + "content": "[38] Aymen Mir, Thiemo Alldieck, and Gerard Pons-Moll. Learning to transfer texture from clothing images to 3d humans. In CVPR, 2020. 2" + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 307, + 647, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 678 + ], + "type": "text", + "content": "[39] Natalia Neverova, David Novotny, Vasil Khalidov, Marc Szafraniec, Patrick Labatut, and Andrea Vedaldi. Continuous surface embeddings. In NeurIPS, 2020. 2" + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 307, + 681, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 681, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 681, + 545, + 712 + ], + "type": "text", + "content": "[40] Natalia Neverova, David Novotny, and Andrea Vedaldi. Correlated uncertainty for learning dense correspondences from noisy labels. In NeurIPS, 2019. 2" + } + ] + } + ], + "index": 41 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4635" + } + ] + } + ], + "index": 43 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 115 + ], + "type": "text", + "content": "[41] Natalia Neverova, Artsiom Sanakoyeu, David Novotny, Patrick Labatut, and Andrea Vedaldi. Discovering relationships between object categories via universal canonical maps. In CVPR, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 116, + 287, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 116, + 287, + 158 + ], + "spans": [ + { + "bbox": [ + 48, + 116, + 287, + 158 + ], + "type": "text", + "content": "[42] Natalia Neverova, James Thewlis, Riza Alp Güler, Iasonas Kokkinos, and Andrea Vedaldi. Slim densepose: Thrifty learning from sparse annotations and motion cues. In CVPR, 2019. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 159, + 287, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 159, + 287, + 245 + ], + "spans": [ + { + "bbox": [ + 48, + 159, + 287, + 245 + ], + "type": "text", + "content": "[43] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In NeurIPS. 2019. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 246, + 287, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 246, + 287, + 278 + ], + "spans": [ + { + "bbox": [ + 48, + 246, + 287, + 278 + ], + "type": "text", + "content": "[44] Alex Rav-Acha, Pushmeet Kohli, Carsten Rother, and Andrew Fitzgibbon. Unwrap mosaics: A new representation for video editing. SIGGRAPH, 2008. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 279, + 287, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 279, + 287, + 320 + ], + "spans": [ + { + "bbox": [ + 48, + 279, + 287, + 320 + ], + "type": "text", + "content": "[45] Igor Santesteban, Nils Thuerey, Miguel A Otaduy, and Dan Casas. Self-Supervised Collision Handling via Generative 3D Garment Models for Virtual Try-On. In CVPR, 2021. 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 322, + 287, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 322, + 287, + 354 + ], + "spans": [ + { + "bbox": [ + 48, + 322, + 287, + 354 + ], + "type": "text", + "content": "[46] Yujun Shen, Jinjin Gu, Xiaou Tang, and Bolei Zhou. Interpreting the latent space of gans for semantic face editing. In CVPR, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 355, + 287, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 355, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 48, + 355, + 287, + 407 + ], + "type": "text", + "content": "[47] Feitong Tan, Danhang Tang, Dou Mingsong, Guo Kaiwen, Rohit Pandey, Cem Keskin, Ruofei Du, Deqing Sun, Sofien Bouaziz, Sean Fanello, Ping Tan, and Yinda Zhang. HumanGPS: Geodesic preserving feature for dense human correspondences. In CVPR, 2021. 2, 5, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 409, + 287, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 409, + 287, + 461 + ], + "spans": [ + { + "bbox": [ + 48, + 409, + 287, + 461 + ], + "type": "text", + "content": "[48] Matthew Tancik, Pratul P. Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan T. Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. In NeurIPS, 2020. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 463, + 287, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 463, + 287, + 485 + ], + "spans": [ + { + "bbox": [ + 48, + 463, + 287, + 485 + ], + "type": "text", + "content": "[49] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. In ECCV, 2020. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 486, + 287, + 516 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 486, + 287, + 516 + ], + "spans": [ + { + "bbox": [ + 48, + 486, + 287, + 516 + ], + "type": "text", + "content": "[50] Alexandru Telea. An image inpainting technique based on the fast marching method. Journal of Graphics Tools, 2004, 4, 5, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 517, + 287, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 517, + 287, + 540 + ], + "spans": [ + { + "bbox": [ + 48, + 517, + 287, + 540 + ], + "type": "text", + "content": "[51] John Y. A. Wang and Edward H. Adelson. Representing moving images with layers. TIP, 1994. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 540, + 287, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 540, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 48, + 540, + 287, + 581 + ], + "type": "text", + "content": "[52] Tuanfeng Wang, Duygu Ceylan, Krishna Kumar Singh, and Niloy J. Mitra. Dance in the wild: Monocular human animation with neural dynamic appearance synthesis. In 3DV, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 582, + 287, + 604 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 582, + 287, + 604 + ], + "spans": [ + { + "bbox": [ + 48, + 582, + 287, + 604 + ], + "type": "text", + "content": "[53] Yair Weiss. Deriving intrinsic images from image sequences. In ICCV, 2001. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 605, + 287, + 636 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 605, + 287, + 636 + ], + "spans": [ + { + "bbox": [ + 48, + 605, + 287, + 636 + ], + "type": "text", + "content": "[54] Rongliang Wu, Gongjie Zhang, Shijian Lu, and Tao Chen. Cascade ef-gan: Progressive facial expression editing with local focuses. In CVPR, 2020. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 637, + 287, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 637, + 287, + 669 + ], + "spans": [ + { + "bbox": [ + 48, + 637, + 287, + 669 + ], + "type": "text", + "content": "[55] You Xie, Huiqi Mao, Angela Yao, and Nils Thuerey. Temporaluv: Capturing loose clothing with temporally coherent uv coordinates. In CVPR, 2022. 2, 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[56] Haonan Yan, Jiaqi Chen, Xujie Zhang, Shengkai Zhang, Nianhong Jiao, Xiaodan Liang, and Tianxiang Zheng. Ultrapose: Synthesizing dense pose with 1 billion points by human-body decoupling 3d model. In ICCV, 2021. 2" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 545, + 237 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 307, + 72, + 545, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 545, + 105 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 545, + 105 + ], + "type": "text", + "content": "[57] Vickie Ye, Zhengqi Li, Richard Tucker, Angjoo Kanazawa, and Noah Snavelly. Deformable sprites for unsupervised video decomposition. In CVPR, 2022. 2, 5, 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 106, + 545, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 106, + 545, + 138 + ], + "spans": [ + { + "bbox": [ + 307, + 106, + 545, + 138 + ], + "type": "text", + "content": "[58] Zhixuan Yu, Haozheng Yu, Long Sha, Sujoy Ganguly, and Hyun Soo Park. Semi-supervised dense keypoints using unlabeled multiview images. arXiv, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 140, + 545, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 140, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 307, + 140, + 545, + 171 + ], + "type": "text", + "content": "[59] Polina Zablotskaia, Aliaksandr Siarohin, Bo Zhao, and Leonid Sigal. Dwnet: Dense warp-based network for pose-guided human video generation. In BMVC, 2019. 5, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 172, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 172, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 307, + 172, + 545, + 205 + ], + "type": "text", + "content": "[60] Wang Zeng, Wanli Ouyang, Ping Luo, Wentao Liu, and Xiaogang Wang. 3d human mesh regression with dense correspondence. In CVPR, 2020. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 205, + 545, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 205, + 545, + 237 + ], + "spans": [ + { + "bbox": [ + 307, + 205, + 545, + 237 + ], + "type": "text", + "content": "[61] Tyler Zhu, Per Karlsson, and Chris Bregler. Simpose: Effectively learning densepose and surface normal of people from simulated data. In ECCV, 2020. 2" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "4636" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Normalizing Flow Based Feature Synthesis for Outlier-Aware Object Detection/44811963-183d-45a5-8a3b-7dd7378c9155_content_list.json b/2023/Normalizing Flow Based Feature Synthesis for Outlier-Aware Object Detection/44811963-183d-45a5-8a3b-7dd7378c9155_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..3f86b5c209d9c16cbd9df7fedebd1a72e4fb39f1 --- /dev/null +++ b/2023/Normalizing Flow Based Feature Synthesis for Outlier-Aware Object Detection/44811963-183d-45a5-8a3b-7dd7378c9155_content_list.json @@ -0,0 +1,1629 @@ +[ + { + "type": "text", + "text": "Normalizing Flow based Feature Synthesis for Outlier-Aware Object Detection", + "text_level": 1, + "bbox": [ + 89, + 130, + 880, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Nishant Kumar1, Sinisa Šegvic2, Abouzar Eslami3, and Stefan Gumhold1 \n1TU Dresden, 2University of Zagreb - FER, 3Carl Zeiss Meditec AG", + "bbox": [ + 194, + 179, + 772, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 251, + 313, + 266 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Real-world deployment of reliable object detectors is crucial for applications such as autonomous driving. However, general-purpose object detectors like Faster R-CNN are prone to providing overconfident predictions for outlier objects. Recent outlier-aware object detection approaches estimate the density of instance-wide features with class-conditional Gaussians and train on synthesized outlier features from their low-likelihood regions. However, this strategy does not guarantee that the synthesized outlier features will have a low likelihood according to the other class-conditional Gaussians. We propose a novel outlier-aware object detection framework that distinguishes outliers from inlier objects by learning the joint data distribution of all inlier classes with an invertible normalizing flow. The appropriate sampling of the flow model ensures that the synthesized outliers have a lower likelihood than inliers of all object classes, thereby modeling a better decision boundary between inlier and outlier objects. Our approach significantly outperforms the state-of-the-art for outlier-aware object detection on both image and video datasets.", + "bbox": [ + 75, + 282, + 472, + 585 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 616, + 209, + 632 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "General purpose object detectors such as Faster R-CNN [42] and Mask R-CNN [17] deliver high performance for inlier images. However, in many real-world scenarios, such as autonomous driving [3], plenty of unknown outliers (OD) naturally occur in an image or a video scene. Due to the co-existence of OD with the labeled inlier (ID) objects in the scene, object detectors confuse outliers with inliers. Therefore, reliable object-detection deployments require detecting such anomalies without degrading the performance of inlier object detection.", + "bbox": [ + 75, + 642, + 468, + 794 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Many outlier detection approaches focus on multi-class image classification task by either performing outlier detection during inference [19, 24, 30, 31, 35, 44] or training on real outlier data [1, 20, 39, 53]. However, such OD inputs are unaware of the decision boundary between inliers and outliers, resulting in an inaccurate model regularization. A popular work [29] proposed a novel training scheme to", + "bbox": [ + 75, + 795, + 470, + 902 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/dda989cc0da06573cc0da40cd95252352f3eb721bd2e219c5b3720f92d6f4f60.jpg", + "image_caption": [ + "Figure 1. We compare distributions of log-likelihoods for inliers (ID), background patches, and synthetic outlier (OD) samples by applying a pre-trained model on Pascal-VOC [11] validation set. The left plot shows the likelihoods of class-conditional Gaussians from the official approach of VOS [10]. The right plot shows the likelihoods recovered with our normalizing flow. We observe that our flow separates the three distributions much better than VOS." + ], + "image_footnote": [], + "bbox": [ + 498, + 234, + 699, + 348 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ae0c5fef7b414ce8919dd718006553cb79002053a945a157f2096ea0be419a3e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 236, + 890, + 347 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "generate synthetic outlier samples in the high-dimensional pixel space using Generative Adversarial Networks for outlier aware training of an image classification model. However, GANs are challenging to optimize and are likely to deliver insufficient coverage [36]. Moreover, the previous work [29] generates an entire image as an outlier sample, whereas in an object detection problem, both inliers and outliers can coexist in the same image space.", + "bbox": [ + 496, + 472, + 893, + 594 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, [10] proposed to model the inlier features as class-conditional Gaussians and then synthesize OD features from the low likelihood region of the modeled distribution. Even though density is more easily estimated in the feature space than in pixel space, the assumption of class-conditional Gaussian may not provide an accurate decision boundary between inliers and outliers. Furthermore, synthesizing a low-likelihood OD sample from the Gaussian for class A does not guarantee that the sample is also of the low likelihood for the Gaussian of class B, as indicated in the left plot of Figure 1. Recent work for object detection in video [9] extracts the background patches for uncertainty regularization by thresholding the dissimilarity with the reference inlier features. However, the extracted background patches may lie far from the inlier-outlier decision boundary, leading to sub-optimal regularization of the model.", + "bbox": [ + 496, + 595, + 895, + 838 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We propose a novel approach for open-set object detection which we call Flow Feature Synthesis (FFS). Our approach trains an invertible normalizing flow to map the data distribution of ID features of all object classes to a latent", + "bbox": [ + 496, + 839, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "5156", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "representation that conforms to a multivariate Gaussian distribution with zero mean and diagonal unit covariance. As a result, it ensures principled estimation of the complex distribution of inlier features from all classes. We synthesize outlier features from low-likelihood regions of the learned distribution. This enables outlier features to be near the ID-OD decision boundary, leading to more robust uncertainty regularization of the object detector. In contrast, using generative adversarial models for the same task would be cumbersome due to its inability to infer density of the generated samples, Furthermore, variational autoencoders can only infer the lower bound of the sample density [27].", + "bbox": [ + 75, + 90, + 472, + 271 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Specifically, FFS optimizes the normalizing flow model by maximum likelihood training on ID features. This training scheme enables the model to estimate the actual data distribution of the available ID features. Next, FFS utilizes the invertibility of the flow model to randomly sample from its latent space and generate synthetic features in the reverse direction of the model. The normalizing flow allows efficient and exact inference of the distribution density in the generated samples. Consequently, our approach can deliver suitable synthetic outlier data in fewer iterations than VOS [10]. It also requires fewer synthetically generated samples than VOS [10] to obtain OD features from the low-likelihood region of the modeled distribution by the flow. Furthermore, we developed our end-to-end trainable FFS framework to be effective on both image and video datasets. In contrast, previous works [9, 10] proposed standalone strategies for each task. Our main contributions are:", + "bbox": [ + 75, + 272, + 472, + 527 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We present a new outlier-aware object detection framework that utilizes Normalizing Flows to model the joint data distribution of inlier features. Invertibility of the flow allows efficient generation of synthetic outliers for effective uncertainty regularization.", + "- By mapping the data distribution of inlier features from all object classes to a multivariate Normal distribution in the flow's latent space, FFS ensures that an outlier sampled using the flow model is OD with reference to all ID classes.", + "- FFS achieves better OD detection performance while training faster than VOS [10] and STUD [9], due to having to generate fewer synthetic samples.", + "- We show that our method achieves state-of-the-art performance in OD object detection while preserving the baseline ID detection performance for image dataset PASCAL-VOC [11] and video datasets such as Youtube-VIS [51] and BDD100K [52]." + ], + "bbox": [ + 96, + 532, + 468, + 804 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 76, + 815, + 215, + 830 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The research in outlier-aware object detection is in the early stages compared to works in vanilla object detection [13, 17, 42] or outlier-aware semantic segmentation [14]. This section overviews prior approaches to outlier", + "bbox": [ + 75, + 839, + 470, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "detection for image classification and proceeds to outlier-aware object detection for images and videos.", + "bbox": [ + 498, + 90, + 890, + 121 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Outlier-aware image classification. Past approaches trained a classification model without knowledge about outliers and conducted OD detection during inference. For example, [19] used simple softmax probability scores during inference to detect OD samples, while other works used Mahalanobis distance [30], rectified activations [46], KL divergence [24] and Gram Matrices [44] instead of softmax scores to detect such samples. On the other hand, ODIN [31] and Generalized ODIN [23] performed perturbations to the test examples to enhance the performance of softmax function for OD detection. Recently, energy function [33, 35, 48] has performed better for distinguishing ID and OD samples than softmax scores.", + "bbox": [ + 496, + 125, + 893, + 319 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Several other methods [20-22, 39, 47] proposed to regularize image classifiers by exposing them with outlier data during self-supervised training while a GAN based approach [29] synthesizes outliers in pixel space and uses it for outlier-aware image classification. However, such synthetic outliers are sampled from imprecise decision boundaries. Additionally, these approaches may not be applied to an object detection task since outliers defined by such methods are in the entire pixel space of the input image. In contrast, an image may contain both inlier and outlier objects simultaneously, and a trustworthy object detection model should detect such scenarios.", + "bbox": [ + 496, + 320, + 893, + 501 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Outlier-aware object detection for images. One of the early open-set object detection works [5] highlighted the importance of adding an extra background class for unknown objects, though object detection inherently rejects unknown object proposals. Several works improve the background detection in open-set conditions [4, 15, 37, 38] by using Bayesian techniques, such as Monte Carlo-Dropout [12]. However, such methods have high latency during inference which could be sub-optimal for real-world applications. Others [16, 43] presented uncertainty regularization for localization regression but did not examine outlier-aware object detection with a focus on classification head. Works like [26] used the background patches as OD samples for model regularization, but the performance might be low as the samples may be far from the decision boundary. VOS [10] applies energy-based regularization [35] to synthetic OD features generated by class-conditional Gaussians. Instead, our method map features from all ID classes to a Gaussian prior with a normalizing flow model, and thus achieves a tighter decision boundary and improved performance.", + "bbox": [ + 496, + 505, + 890, + 821 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Outlier-aware object detection for videos. Several works aim to identify anomalous events on the object and frame level. An object-level approach [8] uses the $k$ -nearest neighbors algorithm to detect anomalous objects while [25] use $k$ -means to cluster inlier features and train $k$ binary clas", + "bbox": [ + 496, + 825, + 893, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "5157", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d3bcad9a4983c52f1d00edd606f84534214c0ba4453530966fec310a8a314428.jpg", + "image_caption": [ + "Figure 2. Overview of the FFS framework. The ID features $l_{ID}(x,\\hat{b})$ are selected for training the normalizing flow with the negative log-likelihood loss $\\mathcal{L}_{nll}$ . We recover outlier features $o_s$ by iterative sampling in the latent space of the normalizing flow. The classification logits of the ID features $l_{ID}(x,\\hat{b})$ and the outlier features $o_s$ are used to calculate the energy scores. The uncertainty loss $\\mathcal{L}_{reg}$ encourages separation between the energy scores of inlier and outlier features. More details are given in Section 3." + ], + "image_footnote": [], + "bbox": [ + 76, + 88, + 893, + 252 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "sifiers so that each cluster can view other clusters as an outlier. A test object is labeled an outlier if the highest classification score is negative. Frame-level approaches [34, 41] predicted future frames that preserve the events of the previous frames by ensuring the optical flow of the predicted frame remains consistent with ground-truth. In contrast, events with a significant difference between prediction and ground-truth are labeled outliers. However, no previous work synthesizes outliers from video frames for model regularization. Recent work [49] introduced self-supervised learning to find the most distinct object in the next video frame given an inlier object in the previous frame, leading to more knowledge about unseen data domains. STUD [9] uses the inliers in the first frame to select the outliers in the following frames for model regularization. However, this may deliver outlier patches that are very far away from the decision boundary and thus lead to inferior model regularization. To summarize, no single outlier-aware object detection framework exists for images and video with an effective regularization that models a precise inlier manifold.", + "bbox": [ + 75, + 321, + 472, + 623 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 645, + 169, + 660 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Outlier-aware object detection problem is much more complex than detecting outliers for an image classification task since a real-world image may consist of both inlier and outlier objects. Let us define an input $x \\in \\mathcal{X}$ with $\\mathcal{X} := (x_1, x_2, \\dots, x_N)$ being the training dataset consisting of $N$ images or video frames, ground-truth class labels for each ID object as $y \\in \\mathcal{Y}$ with $\\mathcal{Y} := \\{1, 2, \\dots, K\\}$ for $K$ classes and the coordinates of ground-truth bounding boxes $b \\in \\mathcal{B}$ . Then, an end-to-end object detector with parameters $\\theta$ treated as a conditional probability distribution estimates bounding boxes as $p_{\\theta}(b|x)$ and the object class as $p_{\\theta}(y|x,b)$ . Figure 2 shows the training and inference scheme of the FFS framework. The backbone network extracts the feature maps at different resolutions, and the Region Proposal Network (RPN) converts these", + "bbox": [ + 75, + 674, + 472, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "multi-resolution feature maps into object proposals. The ROI align module classifies object proposals and provides a vector of bounding boxes, $\\hat{\\pmb{b}}$ , and fixed-size box features, $l(x,\\hat{\\pmb{b}})$ , of both ID (inlier) and background patches. Note that $l(x,\\hat{\\pmb{b}})$ is in a lower dimensional space than $x$ . The box features $l(x,\\hat{\\pmb{b}})$ and bounding boxes $\\hat{\\pmb{b}}$ are used to train the object classification and box prediction heads according to the object detection loss $\\mathcal{L}_{det}$ , given the ground-truth ID classes and box coordinates, respectively. The box features from $l(x,\\hat{\\pmb{b}})$ with the ground-truth class as background are filtered out to keep only ID features $l_{ID}(x,\\hat{\\pmb{b}})$ .", + "bbox": [ + 496, + 321, + 893, + 489 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 498, + 500, + 609, + 513 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our framework FFS is split into three stages: the training procedure, the threshold estimation for OD detection, and the model inference as summarized in Algorithm 1. During the training stage, we compute the standard object detection loss $\\mathcal{L}_{det}$ via the chosen general-purpose object detector. We define maximum-likelihood training objective $\\mathcal{L}_{nll}$ to train our normalizing flow module on inlier features $l_{ID}(x,\\hat{\\pmb{b}})$ . Subsequently, we sample synthetic outliers $o_s$ from the flow model. The details related to the normalizing flow and the procedure to synthesize outlier features are given in Section 3.2. We then compute energy scores for inliers and outliers for the model regularization using $\\mathcal{L}_{reg}$ . Section 3.3 describes the details about model regularization, and Section 3.4 discusses the overall training objective of the FFS framework. We use the trained FFS model for threshold estimation to perform OD detection. We fix the energy-based threshold $\\xi$ such that the model correctly detects $95\\%$ of inlier objects in the validation set. During inference, given an object $x_t$ in a test image, the trained FFS model provides a bounding box for the object as $b_t$ . The decision of whether the object is an inlier or an outlier lies on the energy $E(h(l(x_t,b_t);\\theta))$ . Given an energy threshold $\\xi$ , we assign $x_t$ as an inlier if $E(h(l(x_t,b_t);\\theta)) < \\xi$ and OD if $E(h(l(x_t,b_t);\\theta)) \\geq \\xi$ . We obtain object class labels for predicted inliers with the softmax confidence score", + "bbox": [ + 496, + 523, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "5158", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and bounding box. We label the bounding box for objects detected as outliers as OOD.", + "bbox": [ + 76, + 90, + 468, + 119 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Algorithm 1 FFS: Normalizing Flow based Feature Synthesis for Outlier-Aware Object Detection", + "bbox": [ + 76, + 127, + 467, + 157 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Input: Training data $\\{(x_{i},\\mathbf{b}_{i},y_{i})\\}_{i = 1}^{N}$ , vanilla object detector, normalizing flow $f$ and binary classifier $\\Phi$ with parameters $\\theta ,\\gamma$ and $\\psi$ respectively, the loss weights $\\alpha$ and $\\beta$ .", + "bbox": [ + 76, + 161, + 467, + 208 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Output: Object detector with trained parameter $\\theta$", + "bbox": [ + 76, + 209, + 410, + 223 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "while train do", + "text_level": 1, + "bbox": [ + 78, + 223, + 176, + 236 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Input training data to FFS model and compute $\\mathcal{L}_{\\text{det}}$", + "2. Input the ID features $l_{ID}(x,\\hat{\\pmb{b}})$ to the flow model $f$ and obtain $\\mathcal{L}_{nll}$ using Eq 2." + ], + "bbox": [ + 102, + 237, + 467, + 281 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "if $i \\leq$ iterations then", + "text_level": 1, + "bbox": [ + 102, + 282, + 253, + 295 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. Update $i$ and the parameters $\\theta$ and $\\gamma$ by backpropagation of the total loss in Eq 5 with $\\alpha = 0$ .", + "bbox": [ + 127, + 296, + 467, + 324 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "else", + "text_level": 1, + "bbox": [ + 102, + 325, + 132, + 337 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "4. Generate synthetic features $g_{k}$ by sampling $z_{k}$ from flow's latent space with $g_{k} = f^{-1}(z_{k})$ .", + "5. Select $o_s$ from $p_{\\gamma}(g_k)$ 's low-likelihood region.", + "6. Calculate energy scores $E(h(l_{ID}(x,\\hat{\\boldsymbol{b}});\\theta))$ and $E(h(o_s;\\gamma))$ using the classification head $h$ .", + "7. Compute regularization loss $\\mathcal{L}_{reg}$ using Eq 4.", + "8. Update $i$ and the parameters $\\theta$ , $\\gamma$ and $\\psi$ by backpropagation of the total loss in Eq 5." + ], + "bbox": [ + 127, + 338, + 468, + 460 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "end", + "text_level": 1, + "bbox": [ + 102, + 460, + 132, + 472 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "end", + "text_level": 1, + "bbox": [ + 78, + 477, + 106, + 488 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fix threshold $\\xi$ when $95\\%$ of inliers are correctly detected.", + "bbox": [ + 76, + 491, + 464, + 505 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "while eval do", + "text_level": 1, + "bbox": [ + 78, + 506, + 171, + 518 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Calculate the energy score of a test object.", + "2. Label it as an outlier or an inlier based on $\\xi$ ." + ], + "bbox": [ + 104, + 520, + 408, + 549 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "end", + "text_level": 1, + "bbox": [ + 78, + 550, + 107, + 561 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Normalizing Flow for feature synthesis", + "text_level": 1, + "bbox": [ + 76, + 589, + 410, + 604 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For simplicity, we will denote the ID features $l_{ID}(x,\\hat{\\pmb{b}})$ as $l\\in \\mathcal{L}$ . In our framework, the normalizing flow $f$ with parameters $\\gamma$ is a sequence of $M$ invertible bijective mappings implemented as affine coupling layers. The flow $f:\\mathcal{L}\\rightarrow \\mathcal{Z}$ transforms the complex data distribution of the features $l$ to a multivariate Gaussian in its latent space $z\\in \\mathcal{Z}$ with $p(z) = \\mathcal{N}_d(\\pmb {\\mu},\\pmb {\\Sigma})$ where $d$ is the number of feature vector components, and $\\pmb{\\mu}$ is zero mean with $\\pmb{\\Sigma}$ as the unit variance. The bijective mapping ensures that the input and the latent space share the same dimensionality $d$ such that $f:\\mathbb{R}^d\\to \\mathbb{R}^d$ . Each coupling layer transforms the feature vector components by scaling $s$ and translation $t$ , which are learnable neural networks [27]. Such a transformation is expressive and, at the same time, easily invertible with high efficiency in computing Jacobian determinants.", + "bbox": [ + 75, + 612, + 468, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.1 Maximizing the likelihood of ID features", + "text_level": 1, + "bbox": [ + 76, + 845, + 413, + 859 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given the above formulation of our normalizing flow $f$ , which transforms input ID features $l$ into $z$ such that $z =$", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$f(l)$ , we aim to maximize the log-likelihood of recovering features $l$ with respect to $f$ . To achieve this, we have to compute the Jacobian matrix of $f(l)$ with respect to the feature vector components [27]. Let $f_{i}(l)$ be the components of $f$ in latent space and $l_{j}$ the components of feature space. The entries of the Jacobian matrix are defined as $J_{ij}^{f,l} = \\frac{\\partial f_i(l)}{\\partial l_j}$ where $i,j\\in 1,\\dots,d$ . According to the change of variables formula, the posterior likelihood $p_{\\gamma}(l)$ can be described as:", + "bbox": [ + 496, + 90, + 890, + 215 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\gamma} (l) = p (f (l)) * \\left| \\det J ^ {f, l} \\right| \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 229, + 890, + 250 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To obtain the log-likelihood of the posteriors, we take a logarithm on both sides of the Eq 1. As the training of a neural network requires minimization of a loss function, we construct $f$ by finding $\\gamma$ that maximizes $\\log(p_{\\gamma}(l))$ by minimizing the negative log-likelihood $\\mathcal{L}_{nll}$ objective:", + "bbox": [ + 496, + 253, + 890, + 330 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {n l l} (l; \\gamma) = \\frac {1}{N} \\sum_ {i = 1} ^ {N} - \\log \\left(p _ {\\gamma} \\left(l \\left(x _ {i}, b \\left(x _ {i}\\right)\\right)\\right)\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 335, + 890, + 375 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Normalizing flows are known for stable and exact training according to the maximum likelihood objective in Eq 2.", + "bbox": [ + 498, + 387, + 890, + 417 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.2 Random sampling from the latent space", + "text_level": 1, + "bbox": [ + 500, + 424, + 831, + 439 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We train the flow model $f$ so that its parameters $\\gamma$ learn the distribution of inlier features for a fixed number of training iterations. Next, we randomly sample $k$ samples $z_{k}$ from the latent distribution $p(z)$ and propagate it in the reverse direction of $f$ to generate synthetic samples $g_{k}$ where $g_{k} = f^{-1}(z_{k})$ during active training. As the flow model learned the data distribution $p_{\\gamma}(l)$ via the maximum likelihood training using Eq 2, the samples $g_{k}$ are synthesized from $p_{\\gamma}(l)$ . Such generation procedure discourages mode collapse, common in generative adversarial models. Hence, we retrieve the exact data distribution of the features $l$ in synthesized samples $g_{k}$ . After that, we compute the log-likelihood scores of each of the $k$ samples $g_{k}$ using Eq 1. Reasonably, the synthesized $g_{k}$ contains samples with a high likelihood of being obtained from $p_{\\gamma}(l)$ and should lie well inside the boundary of the inlier distribution $p_{\\gamma}(l)$ . Nevertheless, there should also be samples in $g_{k}$ with lower likelihood scores that lie near or away from the boundary of the distribution $p_{\\gamma}(l)$ .", + "bbox": [ + 496, + 446, + 890, + 734 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.3 Rejection Sampling based Outlier Synthesis", + "text_level": 1, + "bbox": [ + 500, + 739, + 857, + 756 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "It is pivotal to sample useful synthetic outliers for an effective model regularization for OD detection. To achieve this, our first approach involves rejection sampling. We select $s$ samples as outliers $o_{s}$ from the low-likelihood region of the data distribution $p_{\\gamma}(g_k)$ , where $o_{s} \\subset g_{k}$ and $s$ being much smaller than $k$ . The motivation of such an approach is to obtain synthetic outliers from near the decision boundary and outwards in the exterior outlier space. Figure 3 (a)-(b) shows that generating more synthetic samples", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "5159", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$g_{k}$ and selecting a single outlier $o_s$ with the least likelihood (where $s = 1$ ) improves the OD detection. This is because generating more samples $g_{k}$ leads to the selection of a low-likelihood synthetic outlier $o_s$ that genuinely lies in the outlier space, resulting in better model regularization.", + "bbox": [ + 76, + 90, + 468, + 167 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/cd0c84ee971e659ec4a64f8c3e2b52b5cec94e57f931906525543f335e555683.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 83, + 171, + 256, + 268 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/77f6a32b446018a982fd302f9077d67388febd06e155f570eaf7a8e3b12d7673.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 272, + 171, + 450, + 268 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/80b60ce31524e8f895cbb7d6d1c025440d67b2dd6878cf13a20c8fd49dbb8135.jpg", + "image_caption": [ + "Figure 3. Validation of synthetic samples $g_{k}$ and selected outliers $o_{s}$ on OD detection. $s = 1$ in (a)-(b) and $k = 200$ in (c)-(d)." + ], + "image_footnote": [], + "bbox": [ + 83, + 271, + 259, + 367 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6bc2e3adb1301d5d70f5d624e9215af4b53bce074cc1da98c73e9c7d756ee278.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 276, + 271, + 449, + 367 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Eventually, however, the performance peaks as generating an even higher number of synthetic samples $g_{k}$ may lead to the selection of an outlier $o_{s}$ that is too far away in the outlier space, resulting in a model that is unaware of the decision boundary. Hence, by varying $g_{k}$ , our rejection sampling approach can obtain outliers from the decision boundary and from the outlier region away from the boundary. Figure 3 (c)-(d) show the results when the number of selected outliers $o_{s}$ is varied while keeping the number of the generated synthetic samples $g_{k}$ fixed. The selection of more outliers $o_{s}$ leads to a degradation in the OD detection performance as some selected outliers with high log-likelihood scores lie well within the inlier manifold leading to suboptimal model regularization.", + "bbox": [ + 75, + 416, + 468, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.4 Projection Sampling using Langevin Dynamics", + "text_level": 1, + "bbox": [ + 76, + 633, + 455, + 648 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We investigate whether our model can generalize the outlier space by using synthetic outliers sampled only close to the decision boundary while still being effective with OD detection. To accomplish this, we perform projection sampling based on Stochastic Gradient Langevin Dynamics (SGLD) to sample outliers from our flow model. Under the approach, we generate $s$ samples as outliers $o_s$ directly without needing rejection and require a likelihood threshold to define the boundary of the inlier feature distribution. We fix this threshold as the log-likelihood score $\\delta$ of the inlier, which is least likely to be obtained from the inlier data distribution $p_{\\gamma}(l)$ . Next, we propagate the average log-likelihood score of the synthetic outliers $\\log p_{\\gamma}(o_s)$ in the inverse direction of the flow without updating its parameters and obtain the gradients $\\frac{\\partial \\log p_{\\gamma}(o_s)}{\\partial o_s}$ . Subsequently, the outliers $o_s$ are updated as:", + "bbox": [ + 75, + 657, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\no _ {s} := o _ {s} - \\tau \\frac {\\partial \\log p _ {\\gamma} \\left(o _ {s}\\right)}{\\partial o _ {s}} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 87, + 890, + 119 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\tau$ is the step size of the gradient descent. We end this iterative process when the average log-likelihood of the updated $o_s$ is equal to or lower than $\\delta$ . Hence, the updated outliers $o_s$ are precisely projected to the near decision boundary of the inlier manifold where the least-likelihood inlier is located. The results of our projection sampling approach are shown in Figure 5 (f) and discussed in Section 4.3.", + "bbox": [ + 496, + 123, + 893, + 229 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Energy-based model regularization", + "text_level": 1, + "bbox": [ + 498, + 236, + 805, + 252 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We obtained the outliers $o_s$ by sampling from the data distribution $p_{\\gamma}(l)$ , requiring no actual outliers during training. The classification head $h$ now maps inlier features $l$ and outliers $o_s$ to $K + 1$ classification logits as $h(l;\\theta)$ and $h(o_s;\\gamma)$ , respectively. Note that $h(l;\\theta)$ is influenced by all $\\theta$ parameters, whereas $h(o_s;\\gamma)$ depends on the flow parameters $\\gamma$ and the parameters of the classification head. Recently, [10, 35] used energy function rather than softmax probabilities to differentiate inlier and outlier samples better. We follow [10] and apply the energy function on the classification logits $h(l;\\theta)$ and $h(o_s;\\gamma)$ using $E(h(.)) = -T \\cdot \\log \\sum^{K} w_K \\cdot e^{h_K(.)/T}$ where $T$ is a temperature parameter of the energy function and weight $w$ is learned to overcome class imbalances. We then pass the energy scores through a binary classifier $\\Phi$ with parameters $\\psi$ and define the softmax output as $p_l$ when $E(h(l;\\theta))$ is the input and $p_o$ when $E(h(o_s;\\gamma))$ is the input. Finally, we use binary cross-entropy (BCE) based regularization loss $\\mathcal{L}_{reg}$ :", + "bbox": [ + 496, + 258, + 890, + 532 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {r e g} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} - (\\log (p _ {l}) + \\log (1 - p _ {o})) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 549, + 536, + 890, + 575 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The Eq 4 shapes contrasting inlier and outlier energy surfaces by assigning low energy for inliers and high energy for outliers.", + "bbox": [ + 496, + 579, + 890, + 623 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Overall Training Objective", + "text_level": 1, + "bbox": [ + 500, + 631, + 743, + 647 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We describe the standard object detection loss $\\mathcal{L}_{\\text{det}}$ as a combination of losses for object classification given the ground-truth inlier class labels $y$ and bounding box regression given the ground-truth box coordinates $b$ . Based on the formulation of the negative log-likelihood loss $\\mathcal{L}_{\\text{nll}}$ and the regularization loss $\\mathcal{L}_{\\text{reg}}$ from Eq 2 and 4 respectively, we merge all these loss terms in our overall training objective:", + "bbox": [ + 496, + 654, + 890, + 760 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta , \\gamma , \\psi} \\mathbb {E} _ {(x, \\mathbf {b}, y) \\sim \\mathcal {X}, \\mathcal {B}, \\mathcal {Y}} \\left[ \\mathcal {L} _ {d e t} + \\beta \\cdot \\mathcal {L} _ {n l l} + \\alpha \\cdot \\mathcal {L} _ {r e g} \\right], \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 766, + 890, + 789 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\alpha$ and $\\beta$ are the weights of $\\mathcal{L}_{reg}$ and $\\mathcal{L}_{nll}$ respectively. The $\\mathcal{L}_{reg}$ loss acts as an adversary to $\\mathcal{L}_{nll}$ loss as $\\mathcal{L}_{nll}$ tries to maximize the likelihood of obtaining inliers features while $\\mathcal{L}_{reg}$ forces the model to synthesize outliers away from inliers. Thus, such a training scheme encourages the generation of synthetic samples at the decision boundary resulting in outlier awareness of the object detector.", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5160", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 89, + 209, + 107 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This section evaluates our FFS framework for its effectiveness with outlier-aware object detection. Section 4.1 provides the details related to the image and video datasets, the neural network architecture, and the performance metrics. Section 4.2 shows the results obtained with FFS and compares it with other related approaches. Finally, Section 4.3 contains ablation studies to evaluate the performance of our approach in varied experimental settings.", + "bbox": [ + 75, + 114, + 470, + 238 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Implementation details", + "text_level": 1, + "bbox": [ + 76, + 243, + 290, + 258 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We train our FFS model on the publicly available image dataset PASCAL-VOC 2012 [11] containing 20 object categories as ID and use two OD datasets, namely MS-COCO [32] and OpenImages [28] to evaluate the performance. Additionally, we train FFS on two video datasets, namely Berkeley DeepDrive (BDD100K) [52] and Youtube-Video Instance Segmentation (Youtube-VIS) 2021 [51]. For the FFS trained on video datasets, we infer on two OD datasets, namely MS-COCO [32] and nuImages [2]. We use the pre-processed datasets by VOS [10] and STUD [9] so that the object categories in the OD dataset are independent of the object categories in the ID dataset.", + "bbox": [ + 75, + 266, + 468, + 446 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Network architecture. We adopt the Faster R-CNN model [42] for object detection provided in the publicly available Detectron2 framework [50]. We use RegNetX-4.0GF [40] as the backbone architecture due to its superior inlier detection performance. We employ Glow [27] as our normalizing flow model after observing lower performance for other flow architectures (details in Section 4.3).", + "bbox": [ + 75, + 446, + 468, + 551 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metaparameters. We run our experiments using Python 3.8.6 and PyTorch 1.9.0 with each ID dataset trained on four NVIDIA A100-SXM4 GPUs. We start the training of the FFS framework by turning off uncertainty regularization loss $\\mathcal{L}_{reg}$ for a fixed number of iterations. This enables the flow model to learn the data distribution of inlier objects from all object classes. Across datasets, we found the loss weightages $\\beta = 10^{-4}$ and $\\alpha = 0.1$ as reasonable values for stable end-to-end training of the FFS framework. In Section 4.3, we perform an exhaustive meta-parameter study.", + "bbox": [ + 75, + 553, + 468, + 704 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation metrics. We compute the energy scores from $K$ classification logits of the inlier and outlier objects in the validation set. We evaluate these scores according to the standard outlier detection metrics. These metrics are Area Under the Curve Receiver Operating Characteristic (AUROC ↑) and False Positive Rate at 95% True Positive Rate (FPR95 ↓), where True Positive is the correct detection of an inlier. For ID detection, we first compute Intersection over Union (IoU) between predicted bounding boxes and ground-truth boxes. Then, we compute the average precision (AP) for the standard range of IoU thresholds. Finally, the mean of the average precision over all known classes (mAP ↑) measures ID detection performance.", + "bbox": [ + 75, + 704, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparison with the State-of-the-Art", + "text_level": 1, + "bbox": [ + 498, + 90, + 825, + 107 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Tables 1 and 2 compare our method with other approaches on image and video datasets, respectively. We use the results provided by VOS [10] and STUD [9] to show the performance of the compared approaches. Some evaluated approaches (except VOS and STUD) were developed for image-wide OD detection. However, these approaches can also be used to compare an outlier-aware object detection model since classification is one of the object detection tasks. No real outlier datasets were used during training to compare these approaches with our method.", + "bbox": [ + 496, + 113, + 893, + 266 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/abaa89594740f5d812711cc9a94b3bf85a43904f97daa8b9e8451148d697b428.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
IDMethodFPR95 ↓AUROC ↑mAP ↑
PASCAL-VOCGAN [29]OD: MS-COCO / OpenImages48.5
Energy [35]60.93 / 59.9783.67 / 82.67
CSI [47]56.89 / 58.6983.69 / 82.9848.7
VOS [10]59.91 / 57.4181.83 / 82.9548.1
FFS (ours)47.77 / 48.3389.00 / 87.5951.5
44.15 / 45.0889.71 / 88.2951.8
", + "bbox": [ + 501, + 277, + 893, + 383 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/1b70802e88dc2e6adfb1bec3b18e240a2465b02d7bfc32ec95bd8fc514fe799b.jpg", + "table_caption": [ + "Table 1. Main results on image datasets with Faster R-CNN model and RegNetX-4.0GF [40] as the backbone for VOS and FFS." + ], + "table_footnote": [], + "table_body": "
IDODMethodFPR95 ↓AUROC ↑mAP ↑
BDD100KmulImagesGAN [29]83.6570.3931.5
Energy [35]81.6269.4332.0
CSI [47]80.0074.9131.8
STUD [9]79.7576.5532.3
FFS (ours)76.6877.5336.2
Youtube-VISMS-COCOGAN [29]85.7572.9525.5
Energy [35]88.5467.8326.7
CSI [47]82.4371.8124.2
STUD [9]81.1474.8227.2
FFS (ours)83.0676.3727.6
", + "bbox": [ + 501, + 435, + 895, + 597 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Main results on video datasets with Faster R-CNN model and RegNetX-4.0GF [40] as the backbone network.", + "bbox": [ + 498, + 603, + 890, + 632 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Table 1, we report a $7.58\\%$ and a $6.73\\%$ decrease in FPR95 for FFS compared to VOS when evaluated on MS-COCO and OpenImages, respectively. Additionally, we note an increase in AUROC values when compared with VOS and assessed on MS-COCO and OpenImages. We also obtained higher ID detection performance and a faster training time of 2.18 hr compared to 2.43 hr for VOS on the same hardware and batch size. In Table 2, FFS outperforms STUD, where we report a $3.85\\%$ decrease in FPR95 and a $1.28\\%$ increase in AUROC compared to STUD when the ID is BDD100K, and the OD is nuImages. We also register a significant $12\\%$ increase in mAP performance compared to STUD. Our training time is 2.7 hrs compared to 11 hrs by STUD for the BDD100K dataset on the same hardware. FFS also improves STUD results for the Youtube-VIS as ID and MS-COCO as OD while maintaining high ID detection performance. For this dataset, we report a lower training", + "bbox": [ + 496, + 643, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "5161", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "time of 8.2 hrs compared to 11.3 hrs by STUD.", + "bbox": [ + 76, + 90, + 390, + 106 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6d4e7bac6b03af2fe61bfc93a46f91b4b6312e340119b2e2b9a863f1a59678b1.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
IDMethodFPR95 ↓AUROC ↑mAP ↑Time (h)
PASCAL-VOCVOS [10]49.6788.4348.912.37
FFS (ours)43.1289.8451.732.12
Youtube-VISSTUD [9]83.9374.5424.4211.03
FFS (ours)81.9076.3525.108.23
", + "bbox": [ + 78, + 119, + 470, + 185 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Table 3, we also show the performance while using ResNet-50 [18] as an alternative backbone and compare our results with VOS [10] and STUD [9] with MS-COCO as the outlier dataset. The results show that using ResNet-50 as the backbone achieves state-of-the-art ID and OD detection performance with our framework. Overall, Tables 1, 2 and 3 demonstrate the effectiveness of FFS for OD detection for images and videos with lower training time and high ID detection performance.", + "bbox": [ + 75, + 215, + 468, + 351 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 76, + 359, + 240, + 376 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we perform an exhaustive study of meta-parameters that affects the performance of FFS. We also qualitatively compare OD detection results with the VOS approach. For uniformity, we fix the ID dataset as PASCAL-VOC and OD dataset as MS-COCO unless stated otherwise in the experiments.", + "bbox": [ + 75, + 383, + 468, + 474 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "VOS plus rejection sampling: Sampling outliers from a class conditional Gaussian (as performed in VOS [10]) does not ensure that such an outlier will have lower likelihoods for other class Gaussians. Therefore, we compute the log-likelihoods of the outliers using all class Gaussians and reject those with a higher likelihood from at least one other class compared to its generator class. Table 4 shows the results for VOS [10], VOS plus rejection sampling (named as $\\mathrm{VOS + }$ ), and FFS. $\\mathrm{VOS + }$ enhances the OD detection compared to VOS [10] due to selecting more effective outliers, but it is still worse than FFS. Furthermore, $\\mathrm{VOS + }$ 's training time is much higher due to the iterative process of computing log-likelihoods through each Gaussian class.", + "bbox": [ + 75, + 474, + 470, + 672 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b26c8cb9721ed49cfe65ed064c19f4b5141c940f7a7846b820afc9305fa9ff59.jpg", + "table_caption": [ + "Table 3. Performance evaluation with ResNet-50 backbone." + ], + "table_footnote": [], + "table_body": "
MethodFPR95 ↓AUROC ↑mAP (ID)↑Time (h)
VOS [10]OD: MS-COCO / OpenImages
48.22 / 52.7489.04 / 85.9751.502.43
VOS+45.59 / 47.9689.40 / 87.1951.747.38
FFS (ours)44.15 / 45.0889.71 / 88.2951.802.18
", + "bbox": [ + 78, + 683, + 465, + 752 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Normalizing Flow: We conducted a thorough study on several types of normalizing flow models. We selected NICE [6], RealNVP [7], Glow [27] and GIN [45] in the chronological order for this experiment. We fixed the same coupling layers and each model's sub-network $s$ and $t$ configuration for an equivalent comparison. Table 5 shows that Glow outperforms NICE, RealNVP and GIN in terms of", + "bbox": [ + 75, + 794, + 470, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "both FPR95 and AUROC scores. GIN performs best for ID detection but has a worse OD detection performance. As the primary task of FFS is OD detection, we chose Glow as our flow model for all our experiments.", + "bbox": [ + 496, + 90, + 890, + 152 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/459b76aaaec69a7f824a133033a72da8dbf432637c4bbb696e479545ecae411d.jpg", + "table_caption": [ + "Table 4. Comparison of VOS, VOS+, and FFS (ours). The baseline mAP of a closed-set object detector is 51.35." + ], + "table_footnote": [], + "table_body": "
Flow modelFPR95 ↓AUROC ↑mAP ↑Time (h)
NICE [6]48.2388.4651.812.23
RealNVP [7]45.7688.9551.842.28
Glow [27]44.1589.7151.802.18
GIN [45]48.3188.3652.002.25
", + "bbox": [ + 504, + 156, + 888, + 231 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 5 (c) shows the effect of the number of coupling layers $M$ on OD detection performance. $M = 2$ produces the best results as the inlier features fit a suitable number of flow parameters. However, the OD detection performance degrades for $M > 2$ , and the training time significantly increases due to more trainable parameters in the network. Adding more trainable parameters requires more training data for effective training; otherwise, it might lead to overfitting and a more complex optimization, which should explain the degradation in performance. Hence, we fix $M = 2$ , and in Figure 5 (a), we study the effect of the number of fully-connected (fc) layers within $s$ and $t$ sub-networks of each coupling layer. There is a slight improvement in OD detection performance when two fc layers are employed. In Figure 5 (b), we fix the number of coupling and fc layers in each $s$ and $t$ sub-networks as two and adjust the number of neurons in each fc layer. We report a degradation in OD detection performance above 2048 neurons.", + "bbox": [ + 496, + 253, + 890, + 527 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3663be6682f4ead503d2a31c6047e3b08b9739479e7a2434d2cbb2d523e2d058.jpg", + "table_caption": [ + "Table 5. Validation of the Normalizing Flow models." + ], + "table_footnote": [], + "table_body": "
LregFPR95 ↓AUROC ↑mAP ↑Time (h)
CE69.9284.7248.702.27
JSD52.8385.7948.602.25
Hinge51.3286.3446.542.19
BCE (Ours)44.1589.7151.802.18
", + "bbox": [ + 504, + 537, + 883, + 616 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 6. Validation of ${\\mathcal{L}}_{reg}$ on OD detection performance.", + "bbox": [ + 519, + 619, + 867, + 633 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Regularization loss $\\mathcal{L}_{reg}$ and its weightage $\\alpha$ : We studied several regularization losses for our FFS framework. Firstly, we used a simple multi-class cross-entropy loss (CE) on the classification logits without including the energy module. Secondly, we evaluated Jensen-Shannon divergence (JSD) loss, where we minimized the divergence between the distribution of the classification logits of the outlier samples and uniform distribution. In Table 6, we show that our BCE based $\\mathcal{L}_{reg}$ loss outperforms all other loss functions in terms of OD detection performance. The ID detection performance is also higher, even though the training time is comparable. In Figure 5 (e), we demonstrate the effect of changing the weightage $\\alpha$ of $\\mathcal{L}_{reg}$ for PASCAL-VOC as ID and MS-COCO as OD. As $\\alpha$ increases, the OD detection performance improves. However, an even higher weightage degrades the OD detection performance, due to which a careful selection of $\\alpha$ is desirable.", + "bbox": [ + 496, + 643, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "5162", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b29c94f55e9d1a0791c4f3f6784b4b58951ea2015622fd90ba38e61586956595.jpg", + "image_caption": [ + "Figure 4. Visualization of OD detection on MS-COCO and OpenImages with VOS [10] (first and third row) and FFS (ours) (second and fourth row) trained on PASCAL-VOC 2012 images. Ideally, an object in the images should be labeled as OOD with a green bounding box." + ], + "image_footnote": [], + "bbox": [ + 80, + 89, + 888, + 258 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fb1fb7cd1ddcb6f09dee8ec5eac055b7d67f56de64e70e28037eb0264901d6a7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 297, + 272, + 392 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6527735aa9c0794b0a5f91177d5feb75e5df3bdefb56803e7e7fba5358d26c39.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 274, + 296, + 468, + 392 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/510ccb5007d373515b3648b0987f1c7f9a886af0cc282fd332c105efadf2e80a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 76, + 395, + 272, + 489 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1816e992538b3f66e6881e34f3210b99f35528804e6c67e2c0c4b004e2a0355d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 274, + 393, + 465, + 489 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/35e77aa27926f14bb570c3f23553174584f682c6fe4628826bc80358077d63b9.jpg", + "image_caption": [ + "Figure 5. Validation of the flow architecture, loss weightage $\\alpha$ , batch size, and projection sampling. The training time (in hours) is shown above the horizontal x axis in the plots." + ], + "image_footnote": [], + "bbox": [ + 81, + 493, + 269, + 597 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3dd833c7dfb3e5ad291d8ce94de0846ede80cd48a41060d729233d9440fa644b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 277, + 493, + 464, + 598 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Projection Sampling: Figure 5 (f) shows the results after varying synthetic outliers. The OD detection improves as we increase the number of sampled outliers $o_{s}$ but degrades with much higher values of $s$ . The coverage of log-likelihoods from many sampled outliers broadens even though the average log-likelihood is similar to the likelihood threshold $\\delta$ . Hence, some outliers may lie inside, while others exist far outside the decision boundary leading to bad regularization. We report that rejection sampling achieves better OD detection performance when compared to the best results of projection sampling. Since projection sampling enforces the outliers to be close to the decision boundary, it may be insufficient to estimate the entire outlier manifold. In contrast, the rejection sampling synthesize outliers near the decision boundary and in the proper", + "bbox": [ + 75, + 672, + 472, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "outlier manifold, leading to a better model regularization.", + "bbox": [ + 500, + 300, + 877, + 315 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Qualitative analysis: Figure 4 shows the OD detection results on MS-COCO and OpenImages with FFS and VOS trained on PASCAL-VOC. The results show that FFS outperforms VOS in recognizing outlier objects. Additionally, our method's softmax confidence score is lower than VOS when both methods misclassify the outlier objects as inliers.", + "bbox": [ + 496, + 316, + 890, + 406 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 415, + 617, + 429 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We presented a new methodology for outlier-aware object detection that learns a combined data distribution of all inlier object classes. By sampling from the low-likelihood region of a jointly trained normalizing flow model, our approach generates suitable synthetic outlier samples for training the outlier detection head of our compound model. In contrast, previous approaches model inliers with class-conditional Gaussians, resulting in outlier samples that may have a high likelihood for some other class. We report state-of-the-art results for image and video datasets with standard outlier detection metrics while decreasing the model training time. We also show that simple rejection sampling contributes more useful synthetic outliers than projection sampling with a fixed likelihood threshold. In future work, one could validate this finding on other OD detectors. Our idea also extends to developing an outlier-aware instance segmentation model and an active learning scheme to correct the failure modes in weakly-supervised object detectors.", + "bbox": [ + 496, + 440, + 892, + 712 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement: This work primarily received funding from the German Federal Ministry of Education and Research (BMBF) under Software Campus (grant 01IS17044) and was supported by the Center for Scalable Data Analytics and Artificial Intelligence (ScaDS.AI) Dresden/Leipzig, Germany. The work was also partially funded by DFG grant 389792660 as part of TRR 248 - CPEC, Croatian Science Foundation (grant IP-2020-02-5851 ADEPT) and the Cluster of Excellence CeTI (EXC2050/1, grant 390696704). The authors gratefully acknowledge the Center for Information Services and HPC (ZIH) at TU Dresden for providing computing resources.", + "bbox": [ + 496, + 719, + 893, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "5163", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Petra Bevandic, Ivan Kreso, Marin Orsic, and Sinisa Segvic. Dense open-set recognition based on training with noisy negative images. Image Vis. Comput., 2022. 1", + "[2] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuScenes: A multimodal dataset for autonomous driving. In CVPR, 2020. 6", + "[3] Jiwoong Choi, Dayoung Chun, Hyun Kim, and Hyuk-Jae Lee. Gaussian YOLOv3: An accurate and fast object detector using localization uncertainty for autonomous driving. In ICCV, 2019. 1", + "[4] Kumari Deepshikha, Sai Harsha Yelleni, P. K. Srijith, and C. Krishna Mohan. Monte Carlo dropblock for modelling uncertainty in object detection. CoRR, 2021. 2", + "[5] Akshay Raj Dhamija, Manuel Günther, Jonathan Ventura, and Terrance E. Boult. The overlooked elephant of object detection: Open Set. In WACV, 2020. 2", + "[6] Laurent Dinh, David Krueger, and Yoshua Bengio. NICE: Non-linear Independent components estimation. In ICLR, 2015. 7", + "[7] Laurent Dinh, Jascha Sohl-Dickstein, and Samy Bengio. Density estimation using Real NVP. In ICLR, 2017. 7", + "[8] Keval Doshi and Yasin Yilmaz. Any-shot sequential anomaly detection in surveillance videos. In CVPR Workshops, 2020. 2", + "[9] Xuefeng Du, Xin Wang, Gabriel Gozum, and Yixuan Li. Unknown-aware Object Detection: Learning what you don't know from videos in the wild. In CVPR, 2022. 1, 2, 3, 6, 7, 12, 13, 14", + "[10] Xuefeng Du, Zhaoning Wang, Mu Cai, and Yixuan Li. VOS: Learning what you don't know by virtual outlier synthesis. In ICLR, 2022. 1, 2, 5, 6, 7, 8, 11, 12", + "[11] Mark Everingham, Luc Van Gool, Christopher K. I. Williams, John M. Winn, and Andrew Zisserman. The Pascal visual object classes (VOC) challenge. International Journal of Computer Vision, 2010. 1, 2, 6", + "[12] Yarin Gal and Zoubin Ghahramani. Dropout as a Bayesian Approximation: Representing model uncertainty in deep learning. In ICML, 2016. 2", + "[13] Ross Girshick. Fast R-CNN. In ICCV, 2015. 2", + "[14] Matej Grcic, Petra Bevandic, and Sinisa Segvic. Dense open-set recognition with synthetic outliers generated by Real NVP. CoRR, 2020. 2", + "[15] David Hall, Feras Dayoub, John Skinner, Haoyang Zhang, Dimity Miller, Peter Corke, Gustavo Carneiro, Anelia Angelova, and Niko Sunderhauf. Probabilistic Object Detection: Definition and Evaluation. In WACV, 2020. 2", + "[16] Ali Harakeh and Steven L. Waslander. Estimating and evaluating regression predictive uncertainty in deep object detectors. In ICLR, 2021. 2", + "[17] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask R-CNN. In ICCV, 2017. 1, 2", + "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 7" + ], + "bbox": [ + 78, + 114, + 467, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[19] Dan Hendrycks and Kevin Gimpel. A baseline for detecting misclassified and out-of-distribution examples in neural networks. In ICLR, 2017. 1, 2", + "[20] Dan Hendrycks, Mantas Mazeika, and Thomas Dietterich. Deep anomaly detection with outlier exposure. In ICLR, 2019. 1, 2", + "[21] Dan Hendrycks, Mantas Mazeika, Saurav Kadavath, and Dawn Song. Using self-supervised learning can improve model robustness and uncertainty. In NeurIPS, 2019. 2", + "[22] Dan Hendrycks, Norman Mu, Ekin D. Cubuk, Barret Zoph, Justin Gilmer, and Balaji Lakshminarayanan. AugMix: A simple data processing method to improve robustness and uncertainty. In ICLR, 2020. 2", + "[23] Yen-Chang Hsu, Yilin Shen, Hongxia Jin, and Zsolt Kira. Generalized ODIN: Detecting out-of-distribution image without learning from out-of-distribution data. In CVPR, 2020. 2", + "[24] Rui Huang, Andrew Geng, and Yixuan Li. On the importance of gradients for detecting distributional shifts in the wild. In NeurIPS, 2021. 1, 2", + "[25] Radu Tudor Ionescu, Fahad Shahbaz Khan, Mariana-Iuliana Georgescu, and Ling Shao. Object-centric auto-encoders and dummy anomalies for abnormal event detection in video. In CVPR, 2019. 2", + "[26] K. J. Joseph, Salman Khan, Fahad Shahbaz Khan, and Vineeth N. Balasubramanian. Towards open world object detection. In CVPR, 2021. 2", + "[27] Durk P. Kingma and Prafulla Dhariwal. Glow: Generative flow with invertible 1x1 convolutions. In NeurIPS, 2018. 2, 4, 6, 7", + "[28] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The Open Images dataset v4. International Journal of Computer Vision, 2020. 6", + "[29] Kimin Lee, Honglak Lee, Kibok Lee, and Jinwoo Shin. Training confidence-calibrated classifiers for detecting out-of-distribution samples. In ICLR, 2018. 1, 2, 6", + "[30] Kimin Lee, Kibok Lee, Honglak Lee, and Jinwoo Shin. A simple unified framework for detecting out-of-distribution samples and adversarial attacks. In NeurIPS, 2018. 1, 2", + "[31] Shiyu Liang, Yixuan Li, and Rayadurgam Srikant. Enhancing the reliability of out-of-distribution image detection in neural networks. In ICLR, 2018. 1, 2", + "[32] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, Lubomir D. Bourdev, Ross B. Girshick, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollar, and C. Lawrence Zitnick. Microsoft COCO: Common objects in context. In ECCV, 2014. 6", + "[33] Ziqian Lin, Sreya D. Roy, and Yixuan Li. MOOD: Multi-level out-of-distribution detection. In CVPR, 2021. 2", + "[34] Wen Liu, Weixin Luo, Dongze Lian, and Shenghua Gao. Future frame prediction for anomaly detection - A new baseline. In CVPR, 2018. 3", + "[35] Weitang Liu, Xiaoyun Wang, John Owens, and Yixuan Li. Energy-based out-of-distribution detection. In NeurIPS, 2020. 1, 2, 5, 6" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "5164", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[36] Thomas Lucas, Konstantin Shmelkov, Karteek Alahari, Cordelia Schmid, and Jakob Verbeek. Adaptive density estimation for generative models. In NeurIPS, 2019. 1", + "[37] Dimity Miller, Feras Dayoub, Michael Milford, and Niko Sünderhauf. Evaluating merging strategies for sampling-based uncertainty techniques in object detection. In ICRA, 2019. 2", + "[38] Dimity Miller, Lachlan Nicholson, Feras Dayoub, and Niko Sünderhauf. Dropout sampling for robust object detection in open-set conditions. In ICRA, 2018. 2", + "[39] Sina Mohseni, Mandar Pitale, JBS Yadawa, and Zhangyang Wang. Self-supervised learning for generalizable out-of-distribution detection. In AAAI, 2020. 1, 2", + "[40] Ilija Radosavovic, Raj Prateek Kosaraju, Ross B. Girshick, Kaiming He, and Piotr Dólár. Designing network design spaces. In CVPR, 2020. 6", + "[41] Mahdyar Ravanbakhsh, Moin Nabi, Hossein Mousavi, Enver Sangineto, and Nicu Sebe. Plug-and-play CNN for crowd motion analysis: An application in abnormal event detection. In WACV, 2018. 3", + "[42] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster R-CNN: Towards real-time object detection with region proposal networks. In NeurIPS, 2015. 1, 2, 6", + "[43] Tobias Riedlinger, Matthias Rottmann, Marius Schubert, and Hanno Gottschalk. Gradient-based quantification of epistemic uncertainty for deep object detectors. CoRR, 2021. 2", + "[44] Chandramouli Shama Sastry and Sageev Oore. Detecting out-of-distribution examples with Gram Matrices. In ICML, 2020. 1, 2", + "[45] Peter Sorrenson, Carsten Rother, and Ullrich Kothe. Disentanglement by nonlinear ICA with general incompressible-flow networks (GIN). In *ICLR*, 2020. 7", + "[46] Yiyou Sun, Chuan Guo, and Yixuan Li. ReAct: Out-of-distribution detection with rectified activations. In NeurIPS, 2021. 2", + "[47] Jihoon Tack, Sangwoo Mo, Jongheon Jeong, and Jinwoo Shin. CSI: Novelty detection via contrastive learning on distributionally shifted instances. In NeurIPS, 2020. 2, 6", + "[48] Haoran Wang, Weitang Liu, Alex Bocchieri, and Yixuan Li. Can multi-label classification networks know what they don't know? In NeurIPS, 2021. 2", + "[49] Xin Wang, Thomas E. Huang, Benlin Liu, Fisher Yu, Xiaolong Wang, Joseph E. Gonzalez, and Trevor Darrell. Robust object detection via instance-level temporal cycle confusion. In ICCV, 2021. 3", + "[50] Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. Detector2. https://github.com/facebookresearch/detectron2, 2019.6", + "[51] Linjie Yang, Yuchen Fan, and Ning Xu. Video instance segmentation. In ICCV, 2019. 2, 6", + "[52] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darrell. BDD100K: A diverse driving dataset for heterogeneous multitask learning. In CVPR, 2020. 2, 6", + "[53] Jingyang Zhang, Nathan Inkawich, Yiran Chen, and Hai Li. Fine-grained out-of-distribution detection with mixup outlier exposure. CoRR, 2021. 1" + ], + "bbox": [ + 78, + 90, + 467, + 896 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "5165", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Normalizing Flow Based Feature Synthesis for Outlier-Aware Object Detection/44811963-183d-45a5-8a3b-7dd7378c9155_model.json b/2023/Normalizing Flow Based Feature Synthesis for Outlier-Aware Object Detection/44811963-183d-45a5-8a3b-7dd7378c9155_model.json new file mode 100644 index 0000000000000000000000000000000000000000..967ad721c9d518ca7051e16b241cb0d198247348 --- /dev/null +++ b/2023/Normalizing Flow Based Feature Synthesis for Outlier-Aware Object Detection/44811963-183d-45a5-8a3b-7dd7378c9155_model.json @@ -0,0 +1,2299 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.131, + 0.882, + 0.153 + ], + "angle": 0, + "content": "Normalizing Flow based Feature Synthesis for Outlier-Aware Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.195, + 0.18, + 0.773, + 0.218 + ], + "angle": 0, + "content": "Nishant Kumar1, Sinisa Šegvic2, Abouzar Eslami3, and Stefan Gumhold1 \n1TU Dresden, 2University of Zagreb - FER, 3Carl Zeiss Meditec AG" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.252, + 0.314, + 0.267 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.284, + 0.473, + 0.587 + ], + "angle": 0, + "content": "Real-world deployment of reliable object detectors is crucial for applications such as autonomous driving. However, general-purpose object detectors like Faster R-CNN are prone to providing overconfident predictions for outlier objects. Recent outlier-aware object detection approaches estimate the density of instance-wide features with class-conditional Gaussians and train on synthesized outlier features from their low-likelihood regions. However, this strategy does not guarantee that the synthesized outlier features will have a low likelihood according to the other class-conditional Gaussians. We propose a novel outlier-aware object detection framework that distinguishes outliers from inlier objects by learning the joint data distribution of all inlier classes with an invertible normalizing flow. The appropriate sampling of the flow model ensures that the synthesized outliers have a lower likelihood than inliers of all object classes, thereby modeling a better decision boundary between inlier and outlier objects. Our approach significantly outperforms the state-of-the-art for outlier-aware object detection on both image and video datasets." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.617, + 0.21, + 0.633 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.643, + 0.47, + 0.795 + ], + "angle": 0, + "content": "General purpose object detectors such as Faster R-CNN [42] and Mask R-CNN [17] deliver high performance for inlier images. However, in many real-world scenarios, such as autonomous driving [3], plenty of unknown outliers (OD) naturally occur in an image or a video scene. Due to the co-existence of OD with the labeled inlier (ID) objects in the scene, object detectors confuse outliers with inliers. Therefore, reliable object-detection deployments require detecting such anomalies without degrading the performance of inlier object detection." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.471, + 0.904 + ], + "angle": 0, + "content": "Many outlier detection approaches focus on multi-class image classification task by either performing outlier detection during inference [19, 24, 30, 31, 35, 44] or training on real outlier data [1, 20, 39, 53]. However, such OD inputs are unaware of the decision boundary between inliers and outliers, resulting in an inaccurate model regularization. A popular work [29] proposed a novel training scheme to" + }, + { + "type": "image", + "bbox": [ + 0.499, + 0.236, + 0.7, + 0.349 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.695, + 0.237, + 0.892, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.359, + 0.895, + 0.457 + ], + "angle": 0, + "content": "Figure 1. We compare distributions of log-likelihoods for inliers (ID), background patches, and synthetic outlier (OD) samples by applying a pre-trained model on Pascal-VOC [11] validation set. The left plot shows the likelihoods of class-conditional Gaussians from the official approach of VOS [10]. The right plot shows the likelihoods recovered with our normalizing flow. We observe that our flow separates the three distributions much better than VOS." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.473, + 0.895, + 0.595 + ], + "angle": 0, + "content": "generate synthetic outlier samples in the high-dimensional pixel space using Generative Adversarial Networks for outlier aware training of an image classification model. However, GANs are challenging to optimize and are likely to deliver insufficient coverage [36]. Moreover, the previous work [29] generates an entire image as an outlier sample, whereas in an object detection problem, both inliers and outliers can coexist in the same image space." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.597, + 0.896, + 0.839 + ], + "angle": 0, + "content": "Recently, [10] proposed to model the inlier features as class-conditional Gaussians and then synthesize OD features from the low likelihood region of the modeled distribution. Even though density is more easily estimated in the feature space than in pixel space, the assumption of class-conditional Gaussian may not provide an accurate decision boundary between inliers and outliers. Furthermore, synthesizing a low-likelihood OD sample from the Gaussian for class A does not guarantee that the sample is also of the low likelihood for the Gaussian of class B, as indicated in the left plot of Figure 1. Recent work for object detection in video [9] extracts the background patches for uncertainty regularization by thresholding the dissimilarity with the reference inlier features. However, the extracted background patches may lie far from the inlier-outlier decision boundary, leading to sub-optimal regularization of the model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.895, + 0.901 + ], + "angle": 0, + "content": "We propose a novel approach for open-set object detection which we call Flow Feature Synthesis (FFS). Our approach trains an invertible normalizing flow to map the data distribution of ID features of all object classes to a latent" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "5156" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.272 + ], + "angle": 0, + "content": "representation that conforms to a multivariate Gaussian distribution with zero mean and diagonal unit covariance. As a result, it ensures principled estimation of the complex distribution of inlier features from all classes. We synthesize outlier features from low-likelihood regions of the learned distribution. This enables outlier features to be near the ID-OD decision boundary, leading to more robust uncertainty regularization of the object detector. In contrast, using generative adversarial models for the same task would be cumbersome due to its inability to infer density of the generated samples, Furthermore, variational autoencoders can only infer the lower bound of the sample density [27]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.273, + 0.473, + 0.529 + ], + "angle": 0, + "content": "Specifically, FFS optimizes the normalizing flow model by maximum likelihood training on ID features. This training scheme enables the model to estimate the actual data distribution of the available ID features. Next, FFS utilizes the invertibility of the flow model to randomly sample from its latent space and generate synthetic features in the reverse direction of the model. The normalizing flow allows efficient and exact inference of the distribution density in the generated samples. Consequently, our approach can deliver suitable synthetic outlier data in fewer iterations than VOS [10]. It also requires fewer synthetically generated samples than VOS [10] to obtain OD features from the low-likelihood region of the modeled distribution by the flow. Furthermore, we developed our end-to-end trainable FFS framework to be effective on both image and video datasets. In contrast, previous works [9, 10] proposed standalone strategies for each task. Our main contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.533, + 0.47, + 0.607 + ], + "angle": 0, + "content": "- We present a new outlier-aware object detection framework that utilizes Normalizing Flows to model the joint data distribution of inlier features. Invertibility of the flow allows efficient generation of synthetic outliers for effective uncertainty regularization." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.609, + 0.47, + 0.682 + ], + "angle": 0, + "content": "- By mapping the data distribution of inlier features from all object classes to a multivariate Normal distribution in the flow's latent space, FFS ensures that an outlier sampled using the flow model is OD with reference to all ID classes." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.684, + 0.47, + 0.729 + ], + "angle": 0, + "content": "- FFS achieves better OD detection performance while training faster than VOS [10] and STUD [9], due to having to generate fewer synthetic samples." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.73, + 0.47, + 0.805 + ], + "angle": 0, + "content": "- We show that our method achieves state-of-the-art performance in OD object detection while preserving the baseline ID detection performance for image dataset PASCAL-VOC [11] and video datasets such as Youtube-VIS [51] and BDD100K [52]." + }, + { + "type": "list", + "bbox": [ + 0.097, + 0.533, + 0.47, + 0.805 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.816, + 0.216, + 0.831 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.472, + 0.903 + ], + "angle": 0, + "content": "The research in outlier-aware object detection is in the early stages compared to works in vanilla object detection [13, 17, 42] or outlier-aware semantic segmentation [14]. This section overviews prior approaches to outlier" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "detection for image classification and proceeds to outlier-aware object detection for images and videos." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.125, + 0.894, + 0.32 + ], + "angle": 0, + "content": "Outlier-aware image classification. Past approaches trained a classification model without knowledge about outliers and conducted OD detection during inference. For example, [19] used simple softmax probability scores during inference to detect OD samples, while other works used Mahalanobis distance [30], rectified activations [46], KL divergence [24] and Gram Matrices [44] instead of softmax scores to detect such samples. On the other hand, ODIN [31] and Generalized ODIN [23] performed perturbations to the test examples to enhance the performance of softmax function for OD detection. Recently, energy function [33, 35, 48] has performed better for distinguishing ID and OD samples than softmax scores." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.321, + 0.894, + 0.502 + ], + "angle": 0, + "content": "Several other methods [20-22, 39, 47] proposed to regularize image classifiers by exposing them with outlier data during self-supervised training while a GAN based approach [29] synthesizes outliers in pixel space and uses it for outlier-aware image classification. However, such synthetic outliers are sampled from imprecise decision boundaries. Additionally, these approaches may not be applied to an object detection task since outliers defined by such methods are in the entire pixel space of the input image. In contrast, an image may contain both inlier and outlier objects simultaneously, and a trustworthy object detection model should detect such scenarios." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.506, + 0.892, + 0.822 + ], + "angle": 0, + "content": "Outlier-aware object detection for images. One of the early open-set object detection works [5] highlighted the importance of adding an extra background class for unknown objects, though object detection inherently rejects unknown object proposals. Several works improve the background detection in open-set conditions [4, 15, 37, 38] by using Bayesian techniques, such as Monte Carlo-Dropout [12]. However, such methods have high latency during inference which could be sub-optimal for real-world applications. Others [16, 43] presented uncertainty regularization for localization regression but did not examine outlier-aware object detection with a focus on classification head. Works like [26] used the background patches as OD samples for model regularization, but the performance might be low as the samples may be far from the decision boundary. VOS [10] applies energy-based regularization [35] to synthetic OD features generated by class-conditional Gaussians. Instead, our method map features from all ID classes to a Gaussian prior with a normalizing flow model, and thus achieves a tighter decision boundary and improved performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Outlier-aware object detection for videos. Several works aim to identify anomalous events on the object and frame level. An object-level approach [8] uses the \\(k\\)-nearest neighbors algorithm to detect anomalous objects while [25] use \\(k\\)-means to cluster inlier features and train \\(k\\) binary clas" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5157" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.089, + 0.895, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.255, + 0.895, + 0.311 + ], + "angle": 0, + "content": "Figure 2. Overview of the FFS framework. The ID features \\( l_{ID}(x,\\hat{b}) \\) are selected for training the normalizing flow with the negative log-likelihood loss \\( \\mathcal{L}_{nll} \\). We recover outlier features \\( o_s \\) by iterative sampling in the latent space of the normalizing flow. The classification logits of the ID features \\( l_{ID}(x,\\hat{b}) \\) and the outlier features \\( o_s \\) are used to calculate the energy scores. The uncertainty loss \\( \\mathcal{L}_{reg} \\) encourages separation between the energy scores of inlier and outlier features. More details are given in Section 3." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.322, + 0.473, + 0.624 + ], + "angle": 0, + "content": "sifiers so that each cluster can view other clusters as an outlier. A test object is labeled an outlier if the highest classification score is negative. Frame-level approaches [34, 41] predicted future frames that preserve the events of the previous frames by ensuring the optical flow of the predicted frame remains consistent with ground-truth. In contrast, events with a significant difference between prediction and ground-truth are labeled outliers. However, no previous work synthesizes outliers from video frames for model regularization. Recent work [49] introduced self-supervised learning to find the most distinct object in the next video frame given an inlier object in the previous frame, leading to more knowledge about unseen data domains. STUD [9] uses the inliers in the first frame to select the outliers in the following frames for model regularization. However, this may deliver outlier patches that are very far away from the decision boundary and thus lead to inferior model regularization. To summarize, no single outlier-aware object detection framework exists for images and video with an effective regularization that models a precise inlier manifold." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.646, + 0.17, + 0.661 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.473, + 0.901 + ], + "angle": 0, + "content": "Outlier-aware object detection problem is much more complex than detecting outliers for an image classification task since a real-world image may consist of both inlier and outlier objects. Let us define an input \\( x \\in \\mathcal{X} \\) with \\( \\mathcal{X} := (x_1, x_2, \\dots, x_N) \\) being the training dataset consisting of \\( N \\) images or video frames, ground-truth class labels for each ID object as \\( y \\in \\mathcal{Y} \\) with \\( \\mathcal{Y} := \\{1, 2, \\dots, K\\} \\) for \\( K \\) classes and the coordinates of ground-truth bounding boxes \\( b \\in \\mathcal{B} \\). Then, an end-to-end object detector with parameters \\( \\theta \\) treated as a conditional probability distribution estimates bounding boxes as \\( p_{\\theta}(b|x) \\) and the object class as \\( p_{\\theta}(y|x,b) \\). Figure 2 shows the training and inference scheme of the FFS framework. The backbone network extracts the feature maps at different resolutions, and the Region Proposal Network (RPN) converts these" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.322, + 0.895, + 0.49 + ], + "angle": 0, + "content": "multi-resolution feature maps into object proposals. The ROI align module classifies object proposals and provides a vector of bounding boxes, \\(\\hat{\\pmb{b}}\\), and fixed-size box features, \\(l(x,\\hat{\\pmb{b}})\\), of both ID (inlier) and background patches. Note that \\(l(x,\\hat{\\pmb{b}})\\) is in a lower dimensional space than \\(x\\). The box features \\(l(x,\\hat{\\pmb{b}})\\) and bounding boxes \\(\\hat{\\pmb{b}}\\) are used to train the object classification and box prediction heads according to the object detection loss \\(\\mathcal{L}_{det}\\), given the ground-truth ID classes and box coordinates, respectively. The box features from \\(l(x,\\hat{\\pmb{b}})\\) with the ground-truth class as background are filtered out to keep only ID features \\(l_{ID}(x,\\hat{\\pmb{b}})\\)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.5, + 0.611, + 0.515 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.524, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Our framework FFS is split into three stages: the training procedure, the threshold estimation for OD detection, and the model inference as summarized in Algorithm 1. During the training stage, we compute the standard object detection loss \\(\\mathcal{L}_{det}\\) via the chosen general-purpose object detector. We define maximum-likelihood training objective \\(\\mathcal{L}_{nll}\\) to train our normalizing flow module on inlier features \\(l_{ID}(x,\\hat{\\pmb{b}})\\). Subsequently, we sample synthetic outliers \\(o_s\\) from the flow model. The details related to the normalizing flow and the procedure to synthesize outlier features are given in Section 3.2. We then compute energy scores for inliers and outliers for the model regularization using \\(\\mathcal{L}_{reg}\\). Section 3.3 describes the details about model regularization, and Section 3.4 discusses the overall training objective of the FFS framework. We use the trained FFS model for threshold estimation to perform OD detection. We fix the energy-based threshold \\(\\xi\\) such that the model correctly detects \\(95\\%\\) of inlier objects in the validation set. During inference, given an object \\(x_t\\) in a test image, the trained FFS model provides a bounding box for the object as \\(b_t\\). The decision of whether the object is an inlier or an outlier lies on the energy \\(E(h(l(x_t,b_t);\\theta))\\). Given an energy threshold \\(\\xi\\), we assign \\(x_t\\) as an inlier if \\(E(h(l(x_t,b_t);\\theta)) < \\xi\\) and OD if \\(E(h(l(x_t,b_t);\\theta)) \\geq \\xi\\). We obtain object class labels for predicted inliers with the softmax confidence score" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5158" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.121 + ], + "angle": 0, + "content": "and bounding box. We label the bounding box for objects detected as outliers as OOD." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.128, + 0.468, + 0.158 + ], + "angle": 0, + "content": "Algorithm 1 FFS: Normalizing Flow based Feature Synthesis for Outlier-Aware Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.162, + 0.468, + 0.209 + ], + "angle": 0, + "content": "Input: Training data \\(\\{(x_{i},\\mathbf{b}_{i},y_{i})\\}_{i = 1}^{N}\\), vanilla object detector, normalizing flow \\(f\\) and binary classifier \\(\\Phi\\) with parameters \\(\\theta ,\\gamma\\) and \\(\\psi\\) respectively, the loss weights \\(\\alpha\\) and \\(\\beta\\)." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.21, + 0.411, + 0.224 + ], + "angle": 0, + "content": "Output: Object detector with trained parameter \\(\\theta\\)" + }, + { + "type": "title", + "bbox": [ + 0.079, + 0.224, + 0.177, + 0.237 + ], + "angle": 0, + "content": "while train do" + }, + { + "type": "text", + "bbox": [ + 0.105, + 0.238, + 0.466, + 0.251 + ], + "angle": 0, + "content": "1. Input training data to FFS model and compute \\(\\mathcal{L}_{\\text{det}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.252, + 0.468, + 0.282 + ], + "angle": 0, + "content": "2. Input the ID features \\(l_{ID}(x,\\hat{\\pmb{b}})\\) to the flow model \\(f\\) and obtain \\(\\mathcal{L}_{nll}\\) using Eq 2." + }, + { + "type": "list", + "bbox": [ + 0.104, + 0.238, + 0.468, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.104, + 0.283, + 0.254, + 0.296 + ], + "angle": 0, + "content": "if \\(i \\leq\\) iterations then" + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.297, + 0.468, + 0.325 + ], + "angle": 0, + "content": "3. Update \\(i\\) and the parameters \\(\\theta\\) and \\(\\gamma\\) by backpropagation of the total loss in Eq 5 with \\(\\alpha = 0\\)." + }, + { + "type": "title", + "bbox": [ + 0.104, + 0.327, + 0.133, + 0.338 + ], + "angle": 0, + "content": "else" + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.339, + 0.468, + 0.369 + ], + "angle": 0, + "content": "4. Generate synthetic features \\( g_{k} \\) by sampling \\( z_{k} \\) from flow's latent space with \\( g_{k} = f^{-1}(z_{k}) \\)." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.369, + 0.452, + 0.384 + ], + "angle": 0, + "content": "5. Select \\(o_s\\) from \\(p_{\\gamma}(g_k)\\)'s low-likelihood region." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.385, + 0.469, + 0.415 + ], + "angle": 0, + "content": "6. Calculate energy scores \\( E(h(l_{ID}(x,\\hat{\\boldsymbol{b}});\\theta)) \\) and \\( E(h(o_s;\\gamma)) \\) using the classification head \\( h \\)." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.416, + 0.445, + 0.431 + ], + "angle": 0, + "content": "7. Compute regularization loss \\(\\mathcal{L}_{reg}\\) using Eq 4." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.431, + 0.468, + 0.461 + ], + "angle": 0, + "content": "8. Update \\(i\\) and the parameters \\(\\theta\\), \\(\\gamma\\) and \\(\\psi\\) by backpropagation of the total loss in Eq 5." + }, + { + "type": "list", + "bbox": [ + 0.128, + 0.339, + 0.469, + 0.461 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.104, + 0.462, + 0.133, + 0.473 + ], + "angle": 0, + "content": "end" + }, + { + "type": "title", + "bbox": [ + 0.079, + 0.478, + 0.107, + 0.489 + ], + "angle": 0, + "content": "end" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.492, + 0.465, + 0.506 + ], + "angle": 0, + "content": "Fix threshold \\(\\xi\\) when \\(95\\%\\) of inliers are correctly detected." + }, + { + "type": "title", + "bbox": [ + 0.08, + 0.507, + 0.172, + 0.519 + ], + "angle": 0, + "content": "while eval do" + }, + { + "type": "text", + "bbox": [ + 0.105, + 0.521, + 0.399, + 0.535 + ], + "angle": 0, + "content": "1. Calculate the energy score of a test object." + }, + { + "type": "text", + "bbox": [ + 0.105, + 0.536, + 0.41, + 0.55 + ], + "angle": 0, + "content": "2. Label it as an outlier or an inlier based on \\(\\xi\\)." + }, + { + "type": "list", + "bbox": [ + 0.105, + 0.521, + 0.41, + 0.55 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.079, + 0.551, + 0.108, + 0.562 + ], + "angle": 0, + "content": "end" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.59, + 0.411, + 0.606 + ], + "angle": 0, + "content": "3.2. Normalizing Flow for feature synthesis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.613, + 0.47, + 0.84 + ], + "angle": 0, + "content": "For simplicity, we will denote the ID features \\(l_{ID}(x,\\hat{\\pmb{b}})\\) as \\(l\\in \\mathcal{L}\\). In our framework, the normalizing flow \\(f\\) with parameters \\(\\gamma\\) is a sequence of \\(M\\) invertible bijective mappings implemented as affine coupling layers. The flow \\(f:\\mathcal{L}\\rightarrow \\mathcal{Z}\\) transforms the complex data distribution of the features \\(l\\) to a multivariate Gaussian in its latent space \\(z\\in \\mathcal{Z}\\) with \\(p(z) = \\mathcal{N}_d(\\pmb {\\mu},\\pmb {\\Sigma})\\) where \\(d\\) is the number of feature vector components, and \\(\\pmb{\\mu}\\) is zero mean with \\(\\pmb{\\Sigma}\\) as the unit variance. The bijective mapping ensures that the input and the latent space share the same dimensionality \\(d\\) such that \\(f:\\mathbb{R}^d\\to \\mathbb{R}^d\\). Each coupling layer transforms the feature vector components by scaling \\(s\\) and translation \\(t\\), which are learnable neural networks [27]. Such a transformation is expressive and, at the same time, easily invertible with high efficiency in computing Jacobian determinants." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.847, + 0.414, + 0.861 + ], + "angle": 0, + "content": "3.2.1 Maximizing the likelihood of ID features" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Given the above formulation of our normalizing flow \\( f \\), which transforms input ID features \\( l \\) into \\( z \\) such that \\( z = \\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.217 + ], + "angle": 0, + "content": "\\(f(l)\\), we aim to maximize the log-likelihood of recovering features \\(l\\) with respect to \\(f\\). To achieve this, we have to compute the Jacobian matrix of \\(f(l)\\) with respect to the feature vector components [27]. Let \\(f_{i}(l)\\) be the components of \\(f\\) in latent space and \\(l_{j}\\) the components of feature space. The entries of the Jacobian matrix are defined as \\(J_{ij}^{f,l} = \\frac{\\partial f_i(l)}{\\partial l_j}\\) where \\(i,j\\in 1,\\dots,d\\). According to the change of variables formula, the posterior likelihood \\(p_{\\gamma}(l)\\) can be described as:" + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.23, + 0.892, + 0.25 + ], + "angle": 0, + "content": "\\[\np _ {\\gamma} (l) = p (f (l)) * \\left| \\det J ^ {f, l} \\right| \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.255, + 0.892, + 0.331 + ], + "angle": 0, + "content": "To obtain the log-likelihood of the posteriors, we take a logarithm on both sides of the Eq 1. As the training of a neural network requires minimization of a loss function, we construct \\( f \\) by finding \\( \\gamma \\) that maximizes \\( \\log(p_{\\gamma}(l)) \\) by minimizing the negative log-likelihood \\( \\mathcal{L}_{nll} \\) objective:" + }, + { + "type": "equation", + "bbox": [ + 0.551, + 0.336, + 0.892, + 0.376 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {n l l} (l; \\gamma) = \\frac {1}{N} \\sum_ {i = 1} ^ {N} - \\log \\left(p _ {\\gamma} \\left(l \\left(x _ {i}, b \\left(x _ {i}\\right)\\right)\\right)\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.388, + 0.891, + 0.419 + ], + "angle": 0, + "content": "Normalizing flows are known for stable and exact training according to the maximum likelihood objective in Eq 2." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.425, + 0.832, + 0.44 + ], + "angle": 0, + "content": "3.2.2 Random sampling from the latent space" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.448, + 0.892, + 0.736 + ], + "angle": 0, + "content": "We train the flow model \\( f \\) so that its parameters \\( \\gamma \\) learn the distribution of inlier features for a fixed number of training iterations. Next, we randomly sample \\( k \\) samples \\( z_{k} \\) from the latent distribution \\( p(z) \\) and propagate it in the reverse direction of \\( f \\) to generate synthetic samples \\( g_{k} \\) where \\( g_{k} = f^{-1}(z_{k}) \\) during active training. As the flow model learned the data distribution \\( p_{\\gamma}(l) \\) via the maximum likelihood training using Eq 2, the samples \\( g_{k} \\) are synthesized from \\( p_{\\gamma}(l) \\). Such generation procedure discourages mode collapse, common in generative adversarial models. Hence, we retrieve the exact data distribution of the features \\( l \\) in synthesized samples \\( g_{k} \\). After that, we compute the log-likelihood scores of each of the \\( k \\) samples \\( g_{k} \\) using Eq 1. Reasonably, the synthesized \\( g_{k} \\) contains samples with a high likelihood of being obtained from \\( p_{\\gamma}(l) \\) and should lie well inside the boundary of the inlier distribution \\( p_{\\gamma}(l) \\). Nevertheless, there should also be samples in \\( g_{k} \\) with lower likelihood scores that lie near or away from the boundary of the distribution \\( p_{\\gamma}(l) \\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.741, + 0.858, + 0.757 + ], + "angle": 0, + "content": "3.2.3 Rejection Sampling based Outlier Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "It is pivotal to sample useful synthetic outliers for an effective model regularization for OD detection. To achieve this, our first approach involves rejection sampling. We select \\( s \\) samples as outliers \\( o_{s} \\) from the low-likelihood region of the data distribution \\( p_{\\gamma}(g_k) \\), where \\( o_{s} \\subset g_{k} \\) and \\( s \\) being much smaller than \\( k \\). The motivation of such an approach is to obtain synthetic outliers from near the decision boundary and outwards in the exterior outlier space. Figure 3 (a)-(b) shows that generating more synthetic samples" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5159" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.168 + ], + "angle": 0, + "content": "\\(g_{k}\\) and selecting a single outlier \\(o_s\\) with the least likelihood (where \\(s = 1\\)) improves the OD detection. This is because generating more samples \\(g_{k}\\) leads to the selection of a low-likelihood synthetic outlier \\(o_s\\) that genuinely lies in the outlier space, resulting in better model regularization." + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.172, + 0.258, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.172, + 0.452, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.272, + 0.26, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.277, + 0.272, + 0.45, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.378, + 0.468, + 0.406 + ], + "angle": 0, + "content": "Figure 3. Validation of synthetic samples \\( g_{k} \\) and selected outliers \\( o_{s} \\) on OD detection. \\( s = 1 \\) in (a)-(b) and \\( k = 200 \\) in (c)-(d)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.417, + 0.47, + 0.629 + ], + "angle": 0, + "content": "Eventually, however, the performance peaks as generating an even higher number of synthetic samples \\( g_{k} \\) may lead to the selection of an outlier \\( o_{s} \\) that is too far away in the outlier space, resulting in a model that is unaware of the decision boundary. Hence, by varying \\( g_{k} \\), our rejection sampling approach can obtain outliers from the decision boundary and from the outlier region away from the boundary. Figure 3 (c)-(d) show the results when the number of selected outliers \\( o_{s} \\) is varied while keeping the number of the generated synthetic samples \\( g_{k} \\) fixed. The selection of more outliers \\( o_{s} \\) leads to a degradation in the OD detection performance as some selected outliers with high log-likelihood scores lie well within the inlier manifold leading to suboptimal model regularization." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.634, + 0.457, + 0.649 + ], + "angle": 0, + "content": "3.2.4 Projection Sampling using Langevin Dynamics" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.658, + 0.47, + 0.901 + ], + "angle": 0, + "content": "We investigate whether our model can generalize the outlier space by using synthetic outliers sampled only close to the decision boundary while still being effective with OD detection. To accomplish this, we perform projection sampling based on Stochastic Gradient Langevin Dynamics (SGLD) to sample outliers from our flow model. Under the approach, we generate \\(s\\) samples as outliers \\(o_s\\) directly without needing rejection and require a likelihood threshold to define the boundary of the inlier feature distribution. We fix this threshold as the log-likelihood score \\(\\delta\\) of the inlier, which is least likely to be obtained from the inlier data distribution \\(p_{\\gamma}(l)\\). Next, we propagate the average log-likelihood score of the synthetic outliers \\(\\log p_{\\gamma}(o_s)\\) in the inverse direction of the flow without updating its parameters and obtain the gradients \\(\\frac{\\partial \\log p_{\\gamma}(o_s)}{\\partial o_s}\\). Subsequently, the outliers \\(o_s\\) are updated as:" + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.088, + 0.892, + 0.12 + ], + "angle": 0, + "content": "\\[\no _ {s} := o _ {s} - \\tau \\frac {\\partial \\log p _ {\\gamma} \\left(o _ {s}\\right)}{\\partial o _ {s}} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.124, + 0.894, + 0.23 + ], + "angle": 0, + "content": "where \\(\\tau\\) is the step size of the gradient descent. We end this iterative process when the average log-likelihood of the updated \\(o_s\\) is equal to or lower than \\(\\delta\\). Hence, the updated outliers \\(o_s\\) are precisely projected to the near decision boundary of the inlier manifold where the least-likelihood inlier is located. The results of our projection sampling approach are shown in Figure 5 (f) and discussed in Section 4.3." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.237, + 0.807, + 0.253 + ], + "angle": 0, + "content": "3.3. Energy-based model regularization" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.259, + 0.892, + 0.534 + ], + "angle": 0, + "content": "We obtained the outliers \\(o_s\\) by sampling from the data distribution \\(p_{\\gamma}(l)\\), requiring no actual outliers during training. The classification head \\(h\\) now maps inlier features \\(l\\) and outliers \\(o_s\\) to \\(K + 1\\) classification logits as \\(h(l;\\theta)\\) and \\(h(o_s;\\gamma)\\), respectively. Note that \\(h(l;\\theta)\\) is influenced by all \\(\\theta\\) parameters, whereas \\(h(o_s;\\gamma)\\) depends on the flow parameters \\(\\gamma\\) and the parameters of the classification head. Recently, [10, 35] used energy function rather than softmax probabilities to differentiate inlier and outlier samples better. We follow [10] and apply the energy function on the classification logits \\(h(l;\\theta)\\) and \\(h(o_s;\\gamma)\\) using \\(E(h(.)) = -T \\cdot \\log \\sum^{K} w_K \\cdot e^{h_K(.)/T}\\) where \\(T\\) is a temperature parameter of the energy function and weight \\(w\\) is learned to overcome class imbalances. We then pass the energy scores through a binary classifier \\(\\Phi\\) with parameters \\(\\psi\\) and define the softmax output as \\(p_l\\) when \\(E(h(l;\\theta))\\) is the input and \\(p_o\\) when \\(E(h(o_s;\\gamma))\\) is the input. Finally, we use binary cross-entropy (BCE) based regularization loss \\(\\mathcal{L}_{reg}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.55, + 0.537, + 0.892, + 0.577 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {r e g} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} - (\\log (p _ {l}) + \\log (1 - p _ {o})) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.58, + 0.891, + 0.624 + ], + "angle": 0, + "content": "The Eq 4 shapes contrasting inlier and outlier energy surfaces by assigning low energy for inliers and high energy for outliers." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.632, + 0.744, + 0.648 + ], + "angle": 0, + "content": "3.4. Overall Training Objective" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.655, + 0.892, + 0.761 + ], + "angle": 0, + "content": "We describe the standard object detection loss \\(\\mathcal{L}_{\\text{det}}\\) as a combination of losses for object classification given the ground-truth inlier class labels \\(y\\) and bounding box regression given the ground-truth box coordinates \\(b\\). Based on the formulation of the negative log-likelihood loss \\(\\mathcal{L}_{\\text{nll}}\\) and the regularization loss \\(\\mathcal{L}_{\\text{reg}}\\) from Eq 2 and 4 respectively, we merge all these loss terms in our overall training objective:" + }, + { + "type": "equation", + "bbox": [ + 0.516, + 0.767, + 0.892, + 0.79 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta , \\gamma , \\psi} \\mathbb {E} _ {(x, \\mathbf {b}, y) \\sim \\mathcal {X}, \\mathcal {B}, \\mathcal {Y}} \\left[ \\mathcal {L} _ {d e t} + \\beta \\cdot \\mathcal {L} _ {n l l} + \\alpha \\cdot \\mathcal {L} _ {r e g} \\right], \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where \\(\\alpha\\) and \\(\\beta\\) are the weights of \\(\\mathcal{L}_{reg}\\) and \\(\\mathcal{L}_{nll}\\) respectively. The \\(\\mathcal{L}_{reg}\\) loss acts as an adversary to \\(\\mathcal{L}_{nll}\\) loss as \\(\\mathcal{L}_{nll}\\) tries to maximize the likelihood of obtaining inliers features while \\(\\mathcal{L}_{reg}\\) forces the model to synthesize outliers away from inliers. Thus, such a training scheme encourages the generation of synthetic samples at the decision boundary resulting in outlier awareness of the object detector." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5160" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.077, + 0.09, + 0.21, + 0.108 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.115, + 0.471, + 0.239 + ], + "angle": 0, + "content": "This section evaluates our FFS framework for its effectiveness with outlier-aware object detection. Section 4.1 provides the details related to the image and video datasets, the neural network architecture, and the performance metrics. Section 4.2 shows the results obtained with FFS and compares it with other related approaches. Finally, Section 4.3 contains ablation studies to evaluate the performance of our approach in varied experimental settings." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.244, + 0.292, + 0.26 + ], + "angle": 0, + "content": "4.1. Implementation details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.267, + 0.47, + 0.448 + ], + "angle": 0, + "content": "Datasets. We train our FFS model on the publicly available image dataset PASCAL-VOC 2012 [11] containing 20 object categories as ID and use two OD datasets, namely MS-COCO [32] and OpenImages [28] to evaluate the performance. Additionally, we train FFS on two video datasets, namely Berkeley DeepDrive (BDD100K) [52] and Youtube-Video Instance Segmentation (Youtube-VIS) 2021 [51]. For the FFS trained on video datasets, we infer on two OD datasets, namely MS-COCO [32] and nuImages [2]. We use the pre-processed datasets by VOS [10] and STUD [9] so that the object categories in the OD dataset are independent of the object categories in the ID dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.448, + 0.47, + 0.553 + ], + "angle": 0, + "content": "Network architecture. We adopt the Faster R-CNN model [42] for object detection provided in the publicly available Detectron2 framework [50]. We use RegNetX-4.0GF [40] as the backbone architecture due to its superior inlier detection performance. We employ Glow [27] as our normalizing flow model after observing lower performance for other flow architectures (details in Section 4.3)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.554, + 0.47, + 0.705 + ], + "angle": 0, + "content": "Metaparameters. We run our experiments using Python 3.8.6 and PyTorch 1.9.0 with each ID dataset trained on four NVIDIA A100-SXM4 GPUs. We start the training of the FFS framework by turning off uncertainty regularization loss \\(\\mathcal{L}_{reg}\\) for a fixed number of iterations. This enables the flow model to learn the data distribution of inlier objects from all object classes. Across datasets, we found the loss weightages \\(\\beta = 10^{-4}\\) and \\(\\alpha = 0.1\\) as reasonable values for stable end-to-end training of the FFS framework. In Section 4.3, we perform an exhaustive meta-parameter study." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Evaluation metrics. We compute the energy scores from \\( K \\) classification logits of the inlier and outlier objects in the validation set. We evaluate these scores according to the standard outlier detection metrics. These metrics are Area Under the Curve Receiver Operating Characteristic (AUROC ↑) and False Positive Rate at 95% True Positive Rate (FPR95 ↓), where True Positive is the correct detection of an inlier. For ID detection, we first compute Intersection over Union (IoU) between predicted bounding boxes and ground-truth boxes. Then, we compute the average precision (AP) for the standard range of IoU thresholds. Finally, the mean of the average precision over all known classes (mAP ↑) measures ID detection performance." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.091, + 0.826, + 0.108 + ], + "angle": 0, + "content": "4.2. Comparison with the State-of-the-Art" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.114, + 0.895, + 0.267 + ], + "angle": 0, + "content": "Tables 1 and 2 compare our method with other approaches on image and video datasets, respectively. We use the results provided by VOS [10] and STUD [9] to show the performance of the compared approaches. Some evaluated approaches (except VOS and STUD) were developed for image-wide OD detection. However, these approaches can also be used to compare an outlier-aware object detection model since classification is one of the object detection tasks. No real outlier datasets were used during training to compare these approaches with our method." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.279, + 0.895, + 0.384 + ], + "angle": 0, + "content": "
IDMethodFPR95 ↓AUROC ↑mAP ↑
PASCAL-VOCGAN [29]OD: MS-COCO / OpenImages48.5
Energy [35]60.93 / 59.9783.67 / 82.67
CSI [47]56.89 / 58.6983.69 / 82.9848.7
VOS [10]59.91 / 57.4181.83 / 82.9548.1
FFS (ours)47.77 / 48.3389.00 / 87.5951.5
44.15 / 45.0889.71 / 88.2951.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.389, + 0.892, + 0.418 + ], + "angle": 0, + "content": "Table 1. Main results on image datasets with Faster R-CNN model and RegNetX-4.0GF [40] as the backbone for VOS and FFS." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.436, + 0.897, + 0.598 + ], + "angle": 0, + "content": "
IDODMethodFPR95 ↓AUROC ↑mAP ↑
BDD100KmulImagesGAN [29]83.6570.3931.5
Energy [35]81.6269.4332.0
CSI [47]80.0074.9131.8
STUD [9]79.7576.5532.3
FFS (ours)76.6877.5336.2
Youtube-VISMS-COCOGAN [29]85.7572.9525.5
Energy [35]88.5467.8326.7
CSI [47]82.4371.8124.2
STUD [9]81.1474.8227.2
FFS (ours)83.0676.3727.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.604, + 0.892, + 0.633 + ], + "angle": 0, + "content": "Table 2. Main results on video datasets with Faster R-CNN model and RegNetX-4.0GF [40] as the backbone network." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.895, + 0.903 + ], + "angle": 0, + "content": "In Table 1, we report a \\(7.58\\%\\) and a \\(6.73\\%\\) decrease in FPR95 for FFS compared to VOS when evaluated on MS-COCO and OpenImages, respectively. Additionally, we note an increase in AUROC values when compared with VOS and assessed on MS-COCO and OpenImages. We also obtained higher ID detection performance and a faster training time of 2.18 hr compared to 2.43 hr for VOS on the same hardware and batch size. In Table 2, FFS outperforms STUD, where we report a \\(3.85\\%\\) decrease in FPR95 and a \\(1.28\\%\\) increase in AUROC compared to STUD when the ID is BDD100K, and the OD is nuImages. We also register a significant \\(12\\%\\) increase in mAP performance compared to STUD. Our training time is 2.7 hrs compared to 11 hrs by STUD for the BDD100K dataset on the same hardware. FFS also improves STUD results for the Youtube-VIS as ID and MS-COCO as OD while maintaining high ID detection performance. For this dataset, we report a lower training" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "5161" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.078, + 0.092, + 0.391, + 0.107 + ], + "angle": 0, + "content": "time of 8.2 hrs compared to 11.3 hrs by STUD." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.12, + 0.472, + 0.186 + ], + "angle": 0, + "content": "
IDMethodFPR95 ↓AUROC ↑mAP ↑Time (h)
PASCAL-VOCVOS [10]49.6788.4348.912.37
FFS (ours)43.1289.8451.732.12
Youtube-VISSTUD [9]83.9374.5424.4211.03
FFS (ours)81.9076.3525.108.23
" + }, + { + "type": "table_caption", + "bbox": [ + 0.095, + 0.19, + 0.45, + 0.203 + ], + "angle": 0, + "content": "Table 3. Performance evaluation with ResNet-50 backbone." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.216, + 0.47, + 0.352 + ], + "angle": 0, + "content": "In Table 3, we also show the performance while using ResNet-50 [18] as an alternative backbone and compare our results with VOS [10] and STUD [9] with MS-COCO as the outlier dataset. The results show that using ResNet-50 as the backbone achieves state-of-the-art ID and OD detection performance with our framework. Overall, Tables 1, 2 and 3 demonstrate the effectiveness of FFS for OD detection for images and videos with lower training time and high ID detection performance." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.361, + 0.241, + 0.377 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.385, + 0.47, + 0.476 + ], + "angle": 0, + "content": "In this section, we perform an exhaustive study of meta-parameters that affects the performance of FFS. We also qualitatively compare OD detection results with the VOS approach. For uniformity, we fix the ID dataset as PASCAL-VOC and OD dataset as MS-COCO unless stated otherwise in the experiments." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.476, + 0.471, + 0.673 + ], + "angle": 0, + "content": "VOS plus rejection sampling: Sampling outliers from a class conditional Gaussian (as performed in VOS [10]) does not ensure that such an outlier will have lower likelihoods for other class Gaussians. Therefore, we compute the log-likelihoods of the outliers using all class Gaussians and reject those with a higher likelihood from at least one other class compared to its generator class. Table 4 shows the results for VOS [10], VOS plus rejection sampling (named as \\(\\mathrm{VOS + }\\)), and FFS. \\(\\mathrm{VOS + }\\) enhances the OD detection compared to VOS [10] due to selecting more effective outliers, but it is still worse than FFS. Furthermore, \\(\\mathrm{VOS + }\\)'s training time is much higher due to the iterative process of computing log-likelihoods through each Gaussian class." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.684, + 0.466, + 0.753 + ], + "angle": 0, + "content": "
MethodFPR95 ↓AUROC ↑mAP (ID)↑Time (h)
VOS [10]OD: MS-COCO / OpenImages
48.22 / 52.7489.04 / 85.9751.502.43
VOS+45.59 / 47.9689.40 / 87.1951.747.38
FFS (ours)44.15 / 45.0889.71 / 88.2951.802.18
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.755, + 0.469, + 0.784 + ], + "angle": 0, + "content": "Table 4. Comparison of VOS, VOS+, and FFS (ours). The baseline mAP of a closed-set object detector is 51.35." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Normalizing Flow: We conducted a thorough study on several types of normalizing flow models. We selected NICE [6], RealNVP [7], Glow [27] and GIN [45] in the chronological order for this experiment. We fixed the same coupling layers and each model's sub-network \\( s \\) and \\( t \\) configuration for an equivalent comparison. Table 5 shows that Glow outperforms NICE, RealNVP and GIN in terms of" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.153 + ], + "angle": 0, + "content": "both FPR95 and AUROC scores. GIN performs best for ID detection but has a worse OD detection performance. As the primary task of FFS is OD detection, we chose Glow as our flow model for all our experiments." + }, + { + "type": "table", + "bbox": [ + 0.506, + 0.157, + 0.89, + 0.232 + ], + "angle": 0, + "content": "
Flow modelFPR95 ↓AUROC ↑mAP ↑Time (h)
NICE [6]48.2388.4651.812.23
RealNVP [7]45.7688.9551.842.28
Glow [27]44.1589.7151.802.18
GIN [45]48.3188.3652.002.25
" + }, + { + "type": "table_caption", + "bbox": [ + 0.536, + 0.235, + 0.851, + 0.25 + ], + "angle": 0, + "content": "Table 5. Validation of the Normalizing Flow models." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.255, + 0.892, + 0.528 + ], + "angle": 0, + "content": "Figure 5 (c) shows the effect of the number of coupling layers \\( M \\) on OD detection performance. \\( M = 2 \\) produces the best results as the inlier features fit a suitable number of flow parameters. However, the OD detection performance degrades for \\( M > 2 \\), and the training time significantly increases due to more trainable parameters in the network. Adding more trainable parameters requires more training data for effective training; otherwise, it might lead to overfitting and a more complex optimization, which should explain the degradation in performance. Hence, we fix \\( M = 2 \\), and in Figure 5 (a), we study the effect of the number of fully-connected (fc) layers within \\( s \\) and \\( t \\) sub-networks of each coupling layer. There is a slight improvement in OD detection performance when two fc layers are employed. In Figure 5 (b), we fix the number of coupling and fc layers in each \\( s \\) and \\( t \\) sub-networks as two and adjust the number of neurons in each fc layer. We report a degradation in OD detection performance above 2048 neurons." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.538, + 0.885, + 0.617 + ], + "angle": 0, + "content": "
LregFPR95 ↓AUROC ↑mAP ↑Time (h)
CE69.9284.7248.702.27
JSD52.8385.7948.602.25
Hinge51.3286.3446.542.19
BCE (Ours)44.1589.7151.802.18
" + }, + { + "type": "table_caption", + "bbox": [ + 0.521, + 0.62, + 0.869, + 0.634 + ], + "angle": 0, + "content": "Table 6. Validation of \\( {\\mathcal{L}}_{reg} \\) on OD detection performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Regularization loss \\(\\mathcal{L}_{reg}\\) and its weightage \\(\\alpha\\): We studied several regularization losses for our FFS framework. Firstly, we used a simple multi-class cross-entropy loss (CE) on the classification logits without including the energy module. Secondly, we evaluated Jensen-Shannon divergence (JSD) loss, where we minimized the divergence between the distribution of the classification logits of the outlier samples and uniform distribution. In Table 6, we show that our BCE based \\(\\mathcal{L}_{reg}\\) loss outperforms all other loss functions in terms of OD detection performance. The ID detection performance is also higher, even though the training time is comparable. In Figure 5 (e), we demonstrate the effect of changing the weightage \\(\\alpha\\) of \\(\\mathcal{L}_{reg}\\) for PASCAL-VOC as ID and MS-COCO as OD. As \\(\\alpha\\) increases, the OD detection performance improves. However, an even higher weightage degrades the OD detection performance, due to which a careful selection of \\(\\alpha\\) is desirable." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5162" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.09, + 0.89, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.262, + 0.893, + 0.292 + ], + "angle": 0, + "content": "Figure 4. Visualization of OD detection on MS-COCO and OpenImages with VOS [10] (first and third row) and FFS (ours) (second and fourth row) trained on PASCAL-VOC 2012 images. Ideally, an object in the images should be labeled as OOD with a green bounding box." + }, + { + "type": "image", + "bbox": [ + 0.083, + 0.298, + 0.273, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.275, + 0.297, + 0.47, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.077, + 0.396, + 0.274, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.275, + 0.395, + 0.466, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.494, + 0.27, + 0.598 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.279, + 0.494, + 0.465, + 0.599 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.609, + 0.47, + 0.652 + ], + "angle": 0, + "content": "Figure 5. Validation of the flow architecture, loss weightage \\(\\alpha\\), batch size, and projection sampling. The training time (in hours) is shown above the horizontal x axis in the plots." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.674, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Projection Sampling: Figure 5 (f) shows the results after varying synthetic outliers. The OD detection improves as we increase the number of sampled outliers \\( o_{s} \\) but degrades with much higher values of \\( s \\). The coverage of log-likelihoods from many sampled outliers broadens even though the average log-likelihood is similar to the likelihood threshold \\( \\delta \\). Hence, some outliers may lie inside, while others exist far outside the decision boundary leading to bad regularization. We report that rejection sampling achieves better OD detection performance when compared to the best results of projection sampling. Since projection sampling enforces the outliers to be close to the decision boundary, it may be insufficient to estimate the entire outlier manifold. In contrast, the rejection sampling synthesize outliers near the decision boundary and in the proper" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.301, + 0.879, + 0.316 + ], + "angle": 0, + "content": "outlier manifold, leading to a better model regularization." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.317, + 0.892, + 0.407 + ], + "angle": 0, + "content": "Qualitative analysis: Figure 4 shows the OD detection results on MS-COCO and OpenImages with FFS and VOS trained on PASCAL-VOC. The results show that FFS outperforms VOS in recognizing outlier objects. Additionally, our method's softmax confidence score is lower than VOS when both methods misclassify the outlier objects as inliers." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.416, + 0.619, + 0.43 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.441, + 0.893, + 0.713 + ], + "angle": 0, + "content": "We presented a new methodology for outlier-aware object detection that learns a combined data distribution of all inlier object classes. By sampling from the low-likelihood region of a jointly trained normalizing flow model, our approach generates suitable synthetic outlier samples for training the outlier detection head of our compound model. In contrast, previous approaches model inliers with class-conditional Gaussians, resulting in outlier samples that may have a high likelihood for some other class. We report state-of-the-art results for image and video datasets with standard outlier detection metrics while decreasing the model training time. We also show that simple rejection sampling contributes more useful synthetic outliers than projection sampling with a fixed likelihood threshold. In future work, one could validate this finding on other OD detectors. Our idea also extends to developing an outlier-aware instance segmentation model and an active learning scheme to correct the failure modes in weakly-supervised object detectors." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Acknowledgement: This work primarily received funding from the German Federal Ministry of Education and Research (BMBF) under Software Campus (grant 01IS17044) and was supported by the Center for Scalable Data Analytics and Artificial Intelligence (ScaDS.AI) Dresden/Leipzig, Germany. The work was also partially funded by DFG grant 389792660 as part of TRR 248 - CPEC, Croatian Science Foundation (grant IP-2020-02-5851 ADEPT) and the Cluster of Excellence CeTI (EXC2050/1, grant 390696704). The authors gratefully acknowledge the Center for Information Services and HPC (ZIH) at TU Dresden for providing computing resources." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5163" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.468, + 0.157 + ], + "angle": 0, + "content": "[1] Petra Bevandic, Ivan Kreso, Marin Orsic, and Sinisa Segvic. Dense open-set recognition based on training with noisy negative images. Image Vis. Comput., 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.159, + 0.468, + 0.213 + ], + "angle": 0, + "content": "[2] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuScenes: A multimodal dataset for autonomous driving. In CVPR, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.216, + 0.468, + 0.269 + ], + "angle": 0, + "content": "[3] Jiwoong Choi, Dayoung Chun, Hyun Kim, and Hyuk-Jae Lee. Gaussian YOLOv3: An accurate and fast object detector using localization uncertainty for autonomous driving. In ICCV, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.272, + 0.468, + 0.313 + ], + "angle": 0, + "content": "[4] Kumari Deepshikha, Sai Harsha Yelleni, P. K. Srijith, and C. Krishna Mohan. Monte Carlo dropblock for modelling uncertainty in object detection. CoRR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.316, + 0.468, + 0.355 + ], + "angle": 0, + "content": "[5] Akshay Raj Dhamija, Manuel Günther, Jonathan Ventura, and Terrance E. Boult. The overlooked elephant of object detection: Open Set. In WACV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.358, + 0.468, + 0.397 + ], + "angle": 0, + "content": "[6] Laurent Dinh, David Krueger, and Yoshua Bengio. NICE: Non-linear Independent components estimation. In ICLR, 2015. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.401, + 0.468, + 0.428 + ], + "angle": 0, + "content": "[7] Laurent Dinh, Jascha Sohl-Dickstein, and Samy Bengio. Density estimation using Real NVP. In ICLR, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.431, + 0.468, + 0.469 + ], + "angle": 0, + "content": "[8] Keval Doshi and Yasin Yilmaz. Any-shot sequential anomaly detection in surveillance videos. In CVPR Workshops, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.473, + 0.468, + 0.527 + ], + "angle": 0, + "content": "[9] Xuefeng Du, Xin Wang, Gabriel Gozum, and Yixuan Li. Unknown-aware Object Detection: Learning what you don't know from videos in the wild. In CVPR, 2022. 1, 2, 3, 6, 7, 12, 13, 14" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.53, + 0.468, + 0.571 + ], + "angle": 0, + "content": "[10] Xuefeng Du, Zhaoning Wang, Mu Cai, and Yixuan Li. VOS: Learning what you don't know by virtual outlier synthesis. In ICLR, 2022. 1, 2, 5, 6, 7, 8, 11, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.573, + 0.468, + 0.627 + ], + "angle": 0, + "content": "[11] Mark Everingham, Luc Van Gool, Christopher K. I. Williams, John M. Winn, and Andrew Zisserman. The Pascal visual object classes (VOC) challenge. International Journal of Computer Vision, 2010. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.63, + 0.468, + 0.67 + ], + "angle": 0, + "content": "[12] Yarin Gal and Zoubin Ghahramani. Dropout as a Bayesian Approximation: Representing model uncertainty in deep learning. In ICML, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.673, + 0.392, + 0.685 + ], + "angle": 0, + "content": "[13] Ross Girshick. Fast R-CNN. In ICCV, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.688, + 0.468, + 0.728 + ], + "angle": 0, + "content": "[14] Matej Grcic, Petra Bevandic, and Sinisa Segvic. Dense open-set recognition with synthetic outliers generated by Real NVP. CoRR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.731, + 0.468, + 0.785 + ], + "angle": 0, + "content": "[15] David Hall, Feras Dayoub, John Skinner, Haoyang Zhang, Dimity Miller, Peter Corke, Gustavo Carneiro, Anelia Angelova, and Niko Sunderhauf. Probabilistic Object Detection: Definition and Evaluation. In WACV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.788, + 0.468, + 0.828 + ], + "angle": 0, + "content": "[16] Ali Harakeh and Steven L. Waslander. Estimating and evaluating regression predictive uncertainty in deep object detectors. In ICLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.468, + 0.857 + ], + "angle": 0, + "content": "[17] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask R-CNN. In ICCV, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.468, + 0.899 + ], + "angle": 0, + "content": "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.468, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[19] Dan Hendrycks and Kevin Gimpel. A baseline for detecting misclassified and out-of-distribution examples in neural networks. In ICLR, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[20] Dan Hendrycks, Mantas Mazeika, and Thomas Dietterich. Deep anomaly detection with outlier exposure. In ICLR, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.18, + 0.892, + 0.22 + ], + "angle": 0, + "content": "[21] Dan Hendrycks, Mantas Mazeika, Saurav Kadavath, and Dawn Song. Using self-supervised learning can improve model robustness and uncertainty. In NeurIPS, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.224, + 0.892, + 0.278 + ], + "angle": 0, + "content": "[22] Dan Hendrycks, Norman Mu, Ekin D. Cubuk, Barret Zoph, Justin Gilmer, and Balaji Lakshminarayanan. AugMix: A simple data processing method to improve robustness and uncertainty. In ICLR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.281, + 0.892, + 0.335 + ], + "angle": 0, + "content": "[23] Yen-Chang Hsu, Yilin Shen, Hongxia Jin, and Zsolt Kira. Generalized ODIN: Detecting out-of-distribution image without learning from out-of-distribution data. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.338, + 0.892, + 0.379 + ], + "angle": 0, + "content": "[24] Rui Huang, Andrew Geng, and Yixuan Li. On the importance of gradients for detecting distributional shifts in the wild. In NeurIPS, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.382, + 0.892, + 0.436 + ], + "angle": 0, + "content": "[25] Radu Tudor Ionescu, Fahad Shahbaz Khan, Mariana-Iuliana Georgescu, and Ling Shao. Object-centric auto-encoders and dummy anomalies for abnormal event detection in video. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.439, + 0.892, + 0.48 + ], + "angle": 0, + "content": "[26] K. J. Joseph, Salman Khan, Fahad Shahbaz Khan, and Vineeth N. Balasubramanian. Towards open world object detection. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.483, + 0.892, + 0.523 + ], + "angle": 0, + "content": "[27] Durk P. Kingma and Prafulla Dhariwal. Glow: Generative flow with invertible 1x1 convolutions. In NeurIPS, 2018. 2, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.526, + 0.892, + 0.595 + ], + "angle": 0, + "content": "[28] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The Open Images dataset v4. International Journal of Computer Vision, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.598, + 0.892, + 0.639 + ], + "angle": 0, + "content": "[29] Kimin Lee, Honglak Lee, Kibok Lee, and Jinwoo Shin. Training confidence-calibrated classifiers for detecting out-of-distribution samples. In ICLR, 2018. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.641, + 0.892, + 0.683 + ], + "angle": 0, + "content": "[30] Kimin Lee, Kibok Lee, Honglak Lee, and Jinwoo Shin. A simple unified framework for detecting out-of-distribution samples and adversarial attacks. In NeurIPS, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.686, + 0.892, + 0.725 + ], + "angle": 0, + "content": "[31] Shiyu Liang, Yixuan Li, and Rayadurgam Srikant. Enhancing the reliability of out-of-distribution image detection in neural networks. In ICLR, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.728, + 0.892, + 0.783 + ], + "angle": 0, + "content": "[32] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, Lubomir D. Bourdev, Ross B. Girshick, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollar, and C. Lawrence Zitnick. Microsoft COCO: Common objects in context. In ECCV, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.786, + 0.892, + 0.813 + ], + "angle": 0, + "content": "[33] Ziqian Lin, Sreya D. Roy, and Yixuan Li. MOOD: Multi-level out-of-distribution detection. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.816, + 0.892, + 0.856 + ], + "angle": 0, + "content": "[34] Wen Liu, Weixin Luo, Dongze Lian, and Shenghua Gao. Future frame prediction for anomaly detection - A new baseline. In CVPR, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[35] Weitang Liu, Xiaoyun Wang, John Owens, and Yixuan Li. Energy-based out-of-distribution detection. In NeurIPS, 2020. 1, 2, 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5164" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.134 + ], + "angle": 0, + "content": "[36] Thomas Lucas, Konstantin Shmelkov, Karteek Alahari, Cordelia Schmid, and Jakob Verbeek. Adaptive density estimation for generative models. In NeurIPS, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.134, + 0.468, + 0.188 + ], + "angle": 0, + "content": "[37] Dimity Miller, Feras Dayoub, Michael Milford, and Niko Sünderhauf. Evaluating merging strategies for sampling-based uncertainty techniques in object detection. In ICRA, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.189, + 0.468, + 0.231 + ], + "angle": 0, + "content": "[38] Dimity Miller, Lachlan Nicholson, Feras Dayoub, and Niko Sünderhauf. Dropout sampling for robust object detection in open-set conditions. In ICRA, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.232, + 0.468, + 0.272 + ], + "angle": 0, + "content": "[39] Sina Mohseni, Mandar Pitale, JBS Yadawa, and Zhangyang Wang. Self-supervised learning for generalizable out-of-distribution detection. In AAAI, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.273, + 0.468, + 0.314 + ], + "angle": 0, + "content": "[40] Ilija Radosavovic, Raj Prateek Kosaraju, Ross B. Girshick, Kaiming He, and Piotr Dólár. Designing network design spaces. In CVPR, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.315, + 0.468, + 0.369 + ], + "angle": 0, + "content": "[41] Mahdyar Ravanbakhsh, Moin Nabi, Hossein Mousavi, Enver Sangineto, and Nicu Sebe. Plug-and-play CNN for crowd motion analysis: An application in abnormal event detection. In WACV, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.37, + 0.468, + 0.412 + ], + "angle": 0, + "content": "[42] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster R-CNN: Towards real-time object detection with region proposal networks. In NeurIPS, 2015. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.412, + 0.468, + 0.464 + ], + "angle": 0, + "content": "[43] Tobias Riedlinger, Matthias Rottmann, Marius Schubert, and Hanno Gottschalk. Gradient-based quantification of epistemic uncertainty for deep object detectors. CoRR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.467, + 0.468, + 0.508 + ], + "angle": 0, + "content": "[44] Chandramouli Shama Sastry and Sageev Oore. Detecting out-of-distribution examples with Gram Matrices. In ICML, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.509, + 0.468, + 0.55 + ], + "angle": 0, + "content": "[45] Peter Sorrenson, Carsten Rother, and Ullrich Kothe. Disentanglement by nonlinear ICA with general incompressible-flow networks (GIN). In *ICLR*, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.55, + 0.468, + 0.59 + ], + "angle": 0, + "content": "[46] Yiyou Sun, Chuan Guo, and Yixuan Li. ReAct: Out-of-distribution detection with rectified activations. In NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.591, + 0.468, + 0.633 + ], + "angle": 0, + "content": "[47] Jihoon Tack, Sangwoo Mo, Jongheon Jeong, and Jinwoo Shin. CSI: Novelty detection via contrastive learning on distributionally shifted instances. In NeurIPS, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.633, + 0.468, + 0.674 + ], + "angle": 0, + "content": "[48] Haoran Wang, Weitang Liu, Alex Bocchieri, and Yixuan Li. Can multi-label classification networks know what they don't know? In NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.675, + 0.468, + 0.729 + ], + "angle": 0, + "content": "[49] Xin Wang, Thomas E. Huang, Benlin Liu, Fisher Yu, Xiaolong Wang, Joseph E. Gonzalez, and Trevor Darrell. Robust object detection via instance-level temporal cycle confusion. In ICCV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.73, + 0.468, + 0.772 + ], + "angle": 0, + "content": "[50] Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. Detector2. https://github.com/facebookresearch/detectron2, 2019.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.772, + 0.468, + 0.799 + ], + "angle": 0, + "content": "[51] Linjie Yang, Yuchen Fan, and Ning Xu. Video instance segmentation. In ICCV, 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.8, + 0.468, + 0.855 + ], + "angle": 0, + "content": "[52] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darrell. BDD100K: A diverse driving dataset for heterogeneous multitask learning. In CVPR, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.855, + 0.468, + 0.897 + ], + "angle": 0, + "content": "[53] Jingyang Zhang, Nathan Inkawich, Yiran Chen, and Hai Li. Fine-grained out-of-distribution detection with mixup outlier exposure. CoRR, 2021. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.897 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5165" + } + ] +] \ No newline at end of file diff --git a/2023/Normalizing Flow Based Feature Synthesis for Outlier-Aware Object Detection/44811963-183d-45a5-8a3b-7dd7378c9155_origin.pdf b/2023/Normalizing Flow Based Feature Synthesis for Outlier-Aware Object Detection/44811963-183d-45a5-8a3b-7dd7378c9155_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..44c010a070bdfc33e5043784792dac7b43dd19fe --- /dev/null +++ b/2023/Normalizing Flow Based Feature Synthesis for Outlier-Aware Object Detection/44811963-183d-45a5-8a3b-7dd7378c9155_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:82ebbace82a632f84df0ea3e9e5ae58fdd4750e9e6edf0e88c30d7e3a11fdbc8 +size 4783101 diff --git a/2023/Normalizing Flow Based Feature Synthesis for Outlier-Aware Object Detection/full.md b/2023/Normalizing Flow Based Feature Synthesis for Outlier-Aware Object Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..9b3ad75fa252a6c1e806d379949a683c7ab2d8dd --- /dev/null +++ b/2023/Normalizing Flow Based Feature Synthesis for Outlier-Aware Object Detection/full.md @@ -0,0 +1,322 @@ +# Normalizing Flow based Feature Synthesis for Outlier-Aware Object Detection + +Nishant Kumar1, Sinisa Šegvic2, Abouzar Eslami3, and Stefan Gumhold1 +1TU Dresden, 2University of Zagreb - FER, 3Carl Zeiss Meditec AG + +# Abstract + +Real-world deployment of reliable object detectors is crucial for applications such as autonomous driving. However, general-purpose object detectors like Faster R-CNN are prone to providing overconfident predictions for outlier objects. Recent outlier-aware object detection approaches estimate the density of instance-wide features with class-conditional Gaussians and train on synthesized outlier features from their low-likelihood regions. However, this strategy does not guarantee that the synthesized outlier features will have a low likelihood according to the other class-conditional Gaussians. We propose a novel outlier-aware object detection framework that distinguishes outliers from inlier objects by learning the joint data distribution of all inlier classes with an invertible normalizing flow. The appropriate sampling of the flow model ensures that the synthesized outliers have a lower likelihood than inliers of all object classes, thereby modeling a better decision boundary between inlier and outlier objects. Our approach significantly outperforms the state-of-the-art for outlier-aware object detection on both image and video datasets. + +# 1. Introduction + +General purpose object detectors such as Faster R-CNN [42] and Mask R-CNN [17] deliver high performance for inlier images. However, in many real-world scenarios, such as autonomous driving [3], plenty of unknown outliers (OD) naturally occur in an image or a video scene. Due to the co-existence of OD with the labeled inlier (ID) objects in the scene, object detectors confuse outliers with inliers. Therefore, reliable object-detection deployments require detecting such anomalies without degrading the performance of inlier object detection. + +Many outlier detection approaches focus on multi-class image classification task by either performing outlier detection during inference [19, 24, 30, 31, 35, 44] or training on real outlier data [1, 20, 39, 53]. However, such OD inputs are unaware of the decision boundary between inliers and outliers, resulting in an inaccurate model regularization. A popular work [29] proposed a novel training scheme to + +![](images/dda989cc0da06573cc0da40cd95252352f3eb721bd2e219c5b3720f92d6f4f60.jpg) +Figure 1. We compare distributions of log-likelihoods for inliers (ID), background patches, and synthetic outlier (OD) samples by applying a pre-trained model on Pascal-VOC [11] validation set. The left plot shows the likelihoods of class-conditional Gaussians from the official approach of VOS [10]. The right plot shows the likelihoods recovered with our normalizing flow. We observe that our flow separates the three distributions much better than VOS. + +![](images/ae0c5fef7b414ce8919dd718006553cb79002053a945a157f2096ea0be419a3e.jpg) + +generate synthetic outlier samples in the high-dimensional pixel space using Generative Adversarial Networks for outlier aware training of an image classification model. However, GANs are challenging to optimize and are likely to deliver insufficient coverage [36]. Moreover, the previous work [29] generates an entire image as an outlier sample, whereas in an object detection problem, both inliers and outliers can coexist in the same image space. + +Recently, [10] proposed to model the inlier features as class-conditional Gaussians and then synthesize OD features from the low likelihood region of the modeled distribution. Even though density is more easily estimated in the feature space than in pixel space, the assumption of class-conditional Gaussian may not provide an accurate decision boundary between inliers and outliers. Furthermore, synthesizing a low-likelihood OD sample from the Gaussian for class A does not guarantee that the sample is also of the low likelihood for the Gaussian of class B, as indicated in the left plot of Figure 1. Recent work for object detection in video [9] extracts the background patches for uncertainty regularization by thresholding the dissimilarity with the reference inlier features. However, the extracted background patches may lie far from the inlier-outlier decision boundary, leading to sub-optimal regularization of the model. + +We propose a novel approach for open-set object detection which we call Flow Feature Synthesis (FFS). Our approach trains an invertible normalizing flow to map the data distribution of ID features of all object classes to a latent + +representation that conforms to a multivariate Gaussian distribution with zero mean and diagonal unit covariance. As a result, it ensures principled estimation of the complex distribution of inlier features from all classes. We synthesize outlier features from low-likelihood regions of the learned distribution. This enables outlier features to be near the ID-OD decision boundary, leading to more robust uncertainty regularization of the object detector. In contrast, using generative adversarial models for the same task would be cumbersome due to its inability to infer density of the generated samples, Furthermore, variational autoencoders can only infer the lower bound of the sample density [27]. + +Specifically, FFS optimizes the normalizing flow model by maximum likelihood training on ID features. This training scheme enables the model to estimate the actual data distribution of the available ID features. Next, FFS utilizes the invertibility of the flow model to randomly sample from its latent space and generate synthetic features in the reverse direction of the model. The normalizing flow allows efficient and exact inference of the distribution density in the generated samples. Consequently, our approach can deliver suitable synthetic outlier data in fewer iterations than VOS [10]. It also requires fewer synthetically generated samples than VOS [10] to obtain OD features from the low-likelihood region of the modeled distribution by the flow. Furthermore, we developed our end-to-end trainable FFS framework to be effective on both image and video datasets. In contrast, previous works [9, 10] proposed standalone strategies for each task. Our main contributions are: + +- We present a new outlier-aware object detection framework that utilizes Normalizing Flows to model the joint data distribution of inlier features. Invertibility of the flow allows efficient generation of synthetic outliers for effective uncertainty regularization. +- By mapping the data distribution of inlier features from all object classes to a multivariate Normal distribution in the flow's latent space, FFS ensures that an outlier sampled using the flow model is OD with reference to all ID classes. +- FFS achieves better OD detection performance while training faster than VOS [10] and STUD [9], due to having to generate fewer synthetic samples. +- We show that our method achieves state-of-the-art performance in OD object detection while preserving the baseline ID detection performance for image dataset PASCAL-VOC [11] and video datasets such as Youtube-VIS [51] and BDD100K [52]. + +# 2. Related work + +The research in outlier-aware object detection is in the early stages compared to works in vanilla object detection [13, 17, 42] or outlier-aware semantic segmentation [14]. This section overviews prior approaches to outlier + +detection for image classification and proceeds to outlier-aware object detection for images and videos. + +Outlier-aware image classification. Past approaches trained a classification model without knowledge about outliers and conducted OD detection during inference. For example, [19] used simple softmax probability scores during inference to detect OD samples, while other works used Mahalanobis distance [30], rectified activations [46], KL divergence [24] and Gram Matrices [44] instead of softmax scores to detect such samples. On the other hand, ODIN [31] and Generalized ODIN [23] performed perturbations to the test examples to enhance the performance of softmax function for OD detection. Recently, energy function [33, 35, 48] has performed better for distinguishing ID and OD samples than softmax scores. + +Several other methods [20-22, 39, 47] proposed to regularize image classifiers by exposing them with outlier data during self-supervised training while a GAN based approach [29] synthesizes outliers in pixel space and uses it for outlier-aware image classification. However, such synthetic outliers are sampled from imprecise decision boundaries. Additionally, these approaches may not be applied to an object detection task since outliers defined by such methods are in the entire pixel space of the input image. In contrast, an image may contain both inlier and outlier objects simultaneously, and a trustworthy object detection model should detect such scenarios. + +Outlier-aware object detection for images. One of the early open-set object detection works [5] highlighted the importance of adding an extra background class for unknown objects, though object detection inherently rejects unknown object proposals. Several works improve the background detection in open-set conditions [4, 15, 37, 38] by using Bayesian techniques, such as Monte Carlo-Dropout [12]. However, such methods have high latency during inference which could be sub-optimal for real-world applications. Others [16, 43] presented uncertainty regularization for localization regression but did not examine outlier-aware object detection with a focus on classification head. Works like [26] used the background patches as OD samples for model regularization, but the performance might be low as the samples may be far from the decision boundary. VOS [10] applies energy-based regularization [35] to synthetic OD features generated by class-conditional Gaussians. Instead, our method map features from all ID classes to a Gaussian prior with a normalizing flow model, and thus achieves a tighter decision boundary and improved performance. + +Outlier-aware object detection for videos. Several works aim to identify anomalous events on the object and frame level. An object-level approach [8] uses the $k$ -nearest neighbors algorithm to detect anomalous objects while [25] use $k$ -means to cluster inlier features and train $k$ binary clas + +![](images/d3bcad9a4983c52f1d00edd606f84534214c0ba4453530966fec310a8a314428.jpg) +Figure 2. Overview of the FFS framework. The ID features $l_{ID}(x,\hat{b})$ are selected for training the normalizing flow with the negative log-likelihood loss $\mathcal{L}_{nll}$ . We recover outlier features $o_s$ by iterative sampling in the latent space of the normalizing flow. The classification logits of the ID features $l_{ID}(x,\hat{b})$ and the outlier features $o_s$ are used to calculate the energy scores. The uncertainty loss $\mathcal{L}_{reg}$ encourages separation between the energy scores of inlier and outlier features. More details are given in Section 3. + +sifiers so that each cluster can view other clusters as an outlier. A test object is labeled an outlier if the highest classification score is negative. Frame-level approaches [34, 41] predicted future frames that preserve the events of the previous frames by ensuring the optical flow of the predicted frame remains consistent with ground-truth. In contrast, events with a significant difference between prediction and ground-truth are labeled outliers. However, no previous work synthesizes outliers from video frames for model regularization. Recent work [49] introduced self-supervised learning to find the most distinct object in the next video frame given an inlier object in the previous frame, leading to more knowledge about unseen data domains. STUD [9] uses the inliers in the first frame to select the outliers in the following frames for model regularization. However, this may deliver outlier patches that are very far away from the decision boundary and thus lead to inferior model regularization. To summarize, no single outlier-aware object detection framework exists for images and video with an effective regularization that models a precise inlier manifold. + +# 3. Method + +Outlier-aware object detection problem is much more complex than detecting outliers for an image classification task since a real-world image may consist of both inlier and outlier objects. Let us define an input $x \in \mathcal{X}$ with $\mathcal{X} := (x_1, x_2, \dots, x_N)$ being the training dataset consisting of $N$ images or video frames, ground-truth class labels for each ID object as $y \in \mathcal{Y}$ with $\mathcal{Y} := \{1, 2, \dots, K\}$ for $K$ classes and the coordinates of ground-truth bounding boxes $b \in \mathcal{B}$ . Then, an end-to-end object detector with parameters $\theta$ treated as a conditional probability distribution estimates bounding boxes as $p_{\theta}(b|x)$ and the object class as $p_{\theta}(y|x,b)$ . Figure 2 shows the training and inference scheme of the FFS framework. The backbone network extracts the feature maps at different resolutions, and the Region Proposal Network (RPN) converts these + +multi-resolution feature maps into object proposals. The ROI align module classifies object proposals and provides a vector of bounding boxes, $\hat{\pmb{b}}$ , and fixed-size box features, $l(x,\hat{\pmb{b}})$ , of both ID (inlier) and background patches. Note that $l(x,\hat{\pmb{b}})$ is in a lower dimensional space than $x$ . The box features $l(x,\hat{\pmb{b}})$ and bounding boxes $\hat{\pmb{b}}$ are used to train the object classification and box prediction heads according to the object detection loss $\mathcal{L}_{det}$ , given the ground-truth ID classes and box coordinates, respectively. The box features from $l(x,\hat{\pmb{b}})$ with the ground-truth class as background are filtered out to keep only ID features $l_{ID}(x,\hat{\pmb{b}})$ . + +# 3.1. Overview + +Our framework FFS is split into three stages: the training procedure, the threshold estimation for OD detection, and the model inference as summarized in Algorithm 1. During the training stage, we compute the standard object detection loss $\mathcal{L}_{det}$ via the chosen general-purpose object detector. We define maximum-likelihood training objective $\mathcal{L}_{nll}$ to train our normalizing flow module on inlier features $l_{ID}(x,\hat{\pmb{b}})$ . Subsequently, we sample synthetic outliers $o_s$ from the flow model. The details related to the normalizing flow and the procedure to synthesize outlier features are given in Section 3.2. We then compute energy scores for inliers and outliers for the model regularization using $\mathcal{L}_{reg}$ . Section 3.3 describes the details about model regularization, and Section 3.4 discusses the overall training objective of the FFS framework. We use the trained FFS model for threshold estimation to perform OD detection. We fix the energy-based threshold $\xi$ such that the model correctly detects $95\%$ of inlier objects in the validation set. During inference, given an object $x_t$ in a test image, the trained FFS model provides a bounding box for the object as $b_t$ . The decision of whether the object is an inlier or an outlier lies on the energy $E(h(l(x_t,b_t);\theta))$ . Given an energy threshold $\xi$ , we assign $x_t$ as an inlier if $E(h(l(x_t,b_t);\theta)) < \xi$ and OD if $E(h(l(x_t,b_t);\theta)) \geq \xi$ . We obtain object class labels for predicted inliers with the softmax confidence score + +and bounding box. We label the bounding box for objects detected as outliers as OOD. + +Algorithm 1 FFS: Normalizing Flow based Feature Synthesis for Outlier-Aware Object Detection + +Input: Training data $\{(x_{i},\mathbf{b}_{i},y_{i})\}_{i = 1}^{N}$ , vanilla object detector, normalizing flow $f$ and binary classifier $\Phi$ with parameters $\theta ,\gamma$ and $\psi$ respectively, the loss weights $\alpha$ and $\beta$ . + +Output: Object detector with trained parameter $\theta$ + +# while train do + +1. Input training data to FFS model and compute $\mathcal{L}_{\text{det}}$ +2. Input the ID features $l_{ID}(x,\hat{\pmb{b}})$ to the flow model $f$ and obtain $\mathcal{L}_{nll}$ using Eq 2. + +# if $i \leq$ iterations then + +3. Update $i$ and the parameters $\theta$ and $\gamma$ by backpropagation of the total loss in Eq 5 with $\alpha = 0$ . + +# else + +4. Generate synthetic features $g_{k}$ by sampling $z_{k}$ from flow's latent space with $g_{k} = f^{-1}(z_{k})$ . +5. Select $o_s$ from $p_{\gamma}(g_k)$ 's low-likelihood region. +6. Calculate energy scores $E(h(l_{ID}(x,\hat{\boldsymbol{b}});\theta))$ and $E(h(o_s;\gamma))$ using the classification head $h$ . +7. Compute regularization loss $\mathcal{L}_{reg}$ using Eq 4. +8. Update $i$ and the parameters $\theta$ , $\gamma$ and $\psi$ by backpropagation of the total loss in Eq 5. + +# end + +# end + +Fix threshold $\xi$ when $95\%$ of inliers are correctly detected. + +# while eval do + +1. Calculate the energy score of a test object. +2. Label it as an outlier or an inlier based on $\xi$ . + +# end + +# 3.2. Normalizing Flow for feature synthesis + +For simplicity, we will denote the ID features $l_{ID}(x,\hat{\pmb{b}})$ as $l\in \mathcal{L}$ . In our framework, the normalizing flow $f$ with parameters $\gamma$ is a sequence of $M$ invertible bijective mappings implemented as affine coupling layers. The flow $f:\mathcal{L}\rightarrow \mathcal{Z}$ transforms the complex data distribution of the features $l$ to a multivariate Gaussian in its latent space $z\in \mathcal{Z}$ with $p(z) = \mathcal{N}_d(\pmb {\mu},\pmb {\Sigma})$ where $d$ is the number of feature vector components, and $\pmb{\mu}$ is zero mean with $\pmb{\Sigma}$ as the unit variance. The bijective mapping ensures that the input and the latent space share the same dimensionality $d$ such that $f:\mathbb{R}^d\to \mathbb{R}^d$ . Each coupling layer transforms the feature vector components by scaling $s$ and translation $t$ , which are learnable neural networks [27]. Such a transformation is expressive and, at the same time, easily invertible with high efficiency in computing Jacobian determinants. + +# 3.2.1 Maximizing the likelihood of ID features + +Given the above formulation of our normalizing flow $f$ , which transforms input ID features $l$ into $z$ such that $z =$ + +$f(l)$ , we aim to maximize the log-likelihood of recovering features $l$ with respect to $f$ . To achieve this, we have to compute the Jacobian matrix of $f(l)$ with respect to the feature vector components [27]. Let $f_{i}(l)$ be the components of $f$ in latent space and $l_{j}$ the components of feature space. The entries of the Jacobian matrix are defined as $J_{ij}^{f,l} = \frac{\partial f_i(l)}{\partial l_j}$ where $i,j\in 1,\dots,d$ . According to the change of variables formula, the posterior likelihood $p_{\gamma}(l)$ can be described as: + +$$ +p _ {\gamma} (l) = p (f (l)) * \left| \det J ^ {f, l} \right| \tag {1} +$$ + +To obtain the log-likelihood of the posteriors, we take a logarithm on both sides of the Eq 1. As the training of a neural network requires minimization of a loss function, we construct $f$ by finding $\gamma$ that maximizes $\log(p_{\gamma}(l))$ by minimizing the negative log-likelihood $\mathcal{L}_{nll}$ objective: + +$$ +\mathcal {L} _ {n l l} (l; \gamma) = \frac {1}{N} \sum_ {i = 1} ^ {N} - \log \left(p _ {\gamma} \left(l \left(x _ {i}, b \left(x _ {i}\right)\right)\right)\right) \tag {2} +$$ + +Normalizing flows are known for stable and exact training according to the maximum likelihood objective in Eq 2. + +# 3.2.2 Random sampling from the latent space + +We train the flow model $f$ so that its parameters $\gamma$ learn the distribution of inlier features for a fixed number of training iterations. Next, we randomly sample $k$ samples $z_{k}$ from the latent distribution $p(z)$ and propagate it in the reverse direction of $f$ to generate synthetic samples $g_{k}$ where $g_{k} = f^{-1}(z_{k})$ during active training. As the flow model learned the data distribution $p_{\gamma}(l)$ via the maximum likelihood training using Eq 2, the samples $g_{k}$ are synthesized from $p_{\gamma}(l)$ . Such generation procedure discourages mode collapse, common in generative adversarial models. Hence, we retrieve the exact data distribution of the features $l$ in synthesized samples $g_{k}$ . After that, we compute the log-likelihood scores of each of the $k$ samples $g_{k}$ using Eq 1. Reasonably, the synthesized $g_{k}$ contains samples with a high likelihood of being obtained from $p_{\gamma}(l)$ and should lie well inside the boundary of the inlier distribution $p_{\gamma}(l)$ . Nevertheless, there should also be samples in $g_{k}$ with lower likelihood scores that lie near or away from the boundary of the distribution $p_{\gamma}(l)$ . + +# 3.2.3 Rejection Sampling based Outlier Synthesis + +It is pivotal to sample useful synthetic outliers for an effective model regularization for OD detection. To achieve this, our first approach involves rejection sampling. We select $s$ samples as outliers $o_{s}$ from the low-likelihood region of the data distribution $p_{\gamma}(g_k)$ , where $o_{s} \subset g_{k}$ and $s$ being much smaller than $k$ . The motivation of such an approach is to obtain synthetic outliers from near the decision boundary and outwards in the exterior outlier space. Figure 3 (a)-(b) shows that generating more synthetic samples + +$g_{k}$ and selecting a single outlier $o_s$ with the least likelihood (where $s = 1$ ) improves the OD detection. This is because generating more samples $g_{k}$ leads to the selection of a low-likelihood synthetic outlier $o_s$ that genuinely lies in the outlier space, resulting in better model regularization. + +![](images/cd0c84ee971e659ec4a64f8c3e2b52b5cec94e57f931906525543f335e555683.jpg) + +![](images/77f6a32b446018a982fd302f9077d67388febd06e155f570eaf7a8e3b12d7673.jpg) + +![](images/80b60ce31524e8f895cbb7d6d1c025440d67b2dd6878cf13a20c8fd49dbb8135.jpg) +Figure 3. Validation of synthetic samples $g_{k}$ and selected outliers $o_{s}$ on OD detection. $s = 1$ in (a)-(b) and $k = 200$ in (c)-(d). + +![](images/6bc2e3adb1301d5d70f5d624e9215af4b53bce074cc1da98c73e9c7d756ee278.jpg) + +Eventually, however, the performance peaks as generating an even higher number of synthetic samples $g_{k}$ may lead to the selection of an outlier $o_{s}$ that is too far away in the outlier space, resulting in a model that is unaware of the decision boundary. Hence, by varying $g_{k}$ , our rejection sampling approach can obtain outliers from the decision boundary and from the outlier region away from the boundary. Figure 3 (c)-(d) show the results when the number of selected outliers $o_{s}$ is varied while keeping the number of the generated synthetic samples $g_{k}$ fixed. The selection of more outliers $o_{s}$ leads to a degradation in the OD detection performance as some selected outliers with high log-likelihood scores lie well within the inlier manifold leading to suboptimal model regularization. + +# 3.2.4 Projection Sampling using Langevin Dynamics + +We investigate whether our model can generalize the outlier space by using synthetic outliers sampled only close to the decision boundary while still being effective with OD detection. To accomplish this, we perform projection sampling based on Stochastic Gradient Langevin Dynamics (SGLD) to sample outliers from our flow model. Under the approach, we generate $s$ samples as outliers $o_s$ directly without needing rejection and require a likelihood threshold to define the boundary of the inlier feature distribution. We fix this threshold as the log-likelihood score $\delta$ of the inlier, which is least likely to be obtained from the inlier data distribution $p_{\gamma}(l)$ . Next, we propagate the average log-likelihood score of the synthetic outliers $\log p_{\gamma}(o_s)$ in the inverse direction of the flow without updating its parameters and obtain the gradients $\frac{\partial \log p_{\gamma}(o_s)}{\partial o_s}$ . Subsequently, the outliers $o_s$ are updated as: + +$$ +o _ {s} := o _ {s} - \tau \frac {\partial \log p _ {\gamma} \left(o _ {s}\right)}{\partial o _ {s}} \tag {3} +$$ + +where $\tau$ is the step size of the gradient descent. We end this iterative process when the average log-likelihood of the updated $o_s$ is equal to or lower than $\delta$ . Hence, the updated outliers $o_s$ are precisely projected to the near decision boundary of the inlier manifold where the least-likelihood inlier is located. The results of our projection sampling approach are shown in Figure 5 (f) and discussed in Section 4.3. + +# 3.3. Energy-based model regularization + +We obtained the outliers $o_s$ by sampling from the data distribution $p_{\gamma}(l)$ , requiring no actual outliers during training. The classification head $h$ now maps inlier features $l$ and outliers $o_s$ to $K + 1$ classification logits as $h(l;\theta)$ and $h(o_s;\gamma)$ , respectively. Note that $h(l;\theta)$ is influenced by all $\theta$ parameters, whereas $h(o_s;\gamma)$ depends on the flow parameters $\gamma$ and the parameters of the classification head. Recently, [10, 35] used energy function rather than softmax probabilities to differentiate inlier and outlier samples better. We follow [10] and apply the energy function on the classification logits $h(l;\theta)$ and $h(o_s;\gamma)$ using $E(h(.)) = -T \cdot \log \sum^{K} w_K \cdot e^{h_K(.)/T}$ where $T$ is a temperature parameter of the energy function and weight $w$ is learned to overcome class imbalances. We then pass the energy scores through a binary classifier $\Phi$ with parameters $\psi$ and define the softmax output as $p_l$ when $E(h(l;\theta))$ is the input and $p_o$ when $E(h(o_s;\gamma))$ is the input. Finally, we use binary cross-entropy (BCE) based regularization loss $\mathcal{L}_{reg}$ : + +$$ +\mathcal {L} _ {r e g} = \frac {1}{N} \sum_ {i = 1} ^ {N} - (\log (p _ {l}) + \log (1 - p _ {o})) \tag {4} +$$ + +The Eq 4 shapes contrasting inlier and outlier energy surfaces by assigning low energy for inliers and high energy for outliers. + +# 3.4. Overall Training Objective + +We describe the standard object detection loss $\mathcal{L}_{\text{det}}$ as a combination of losses for object classification given the ground-truth inlier class labels $y$ and bounding box regression given the ground-truth box coordinates $b$ . Based on the formulation of the negative log-likelihood loss $\mathcal{L}_{\text{nll}}$ and the regularization loss $\mathcal{L}_{\text{reg}}$ from Eq 2 and 4 respectively, we merge all these loss terms in our overall training objective: + +$$ +\min _ {\theta , \gamma , \psi} \mathbb {E} _ {(x, \mathbf {b}, y) \sim \mathcal {X}, \mathcal {B}, \mathcal {Y}} \left[ \mathcal {L} _ {d e t} + \beta \cdot \mathcal {L} _ {n l l} + \alpha \cdot \mathcal {L} _ {r e g} \right], \tag {5} +$$ + +where $\alpha$ and $\beta$ are the weights of $\mathcal{L}_{reg}$ and $\mathcal{L}_{nll}$ respectively. The $\mathcal{L}_{reg}$ loss acts as an adversary to $\mathcal{L}_{nll}$ loss as $\mathcal{L}_{nll}$ tries to maximize the likelihood of obtaining inliers features while $\mathcal{L}_{reg}$ forces the model to synthesize outliers away from inliers. Thus, such a training scheme encourages the generation of synthetic samples at the decision boundary resulting in outlier awareness of the object detector. + +# 4. Experiments + +This section evaluates our FFS framework for its effectiveness with outlier-aware object detection. Section 4.1 provides the details related to the image and video datasets, the neural network architecture, and the performance metrics. Section 4.2 shows the results obtained with FFS and compares it with other related approaches. Finally, Section 4.3 contains ablation studies to evaluate the performance of our approach in varied experimental settings. + +# 4.1. Implementation details + +Datasets. We train our FFS model on the publicly available image dataset PASCAL-VOC 2012 [11] containing 20 object categories as ID and use two OD datasets, namely MS-COCO [32] and OpenImages [28] to evaluate the performance. Additionally, we train FFS on two video datasets, namely Berkeley DeepDrive (BDD100K) [52] and Youtube-Video Instance Segmentation (Youtube-VIS) 2021 [51]. For the FFS trained on video datasets, we infer on two OD datasets, namely MS-COCO [32] and nuImages [2]. We use the pre-processed datasets by VOS [10] and STUD [9] so that the object categories in the OD dataset are independent of the object categories in the ID dataset. + +Network architecture. We adopt the Faster R-CNN model [42] for object detection provided in the publicly available Detectron2 framework [50]. We use RegNetX-4.0GF [40] as the backbone architecture due to its superior inlier detection performance. We employ Glow [27] as our normalizing flow model after observing lower performance for other flow architectures (details in Section 4.3). + +Metaparameters. We run our experiments using Python 3.8.6 and PyTorch 1.9.0 with each ID dataset trained on four NVIDIA A100-SXM4 GPUs. We start the training of the FFS framework by turning off uncertainty regularization loss $\mathcal{L}_{reg}$ for a fixed number of iterations. This enables the flow model to learn the data distribution of inlier objects from all object classes. Across datasets, we found the loss weightages $\beta = 10^{-4}$ and $\alpha = 0.1$ as reasonable values for stable end-to-end training of the FFS framework. In Section 4.3, we perform an exhaustive meta-parameter study. + +Evaluation metrics. We compute the energy scores from $K$ classification logits of the inlier and outlier objects in the validation set. We evaluate these scores according to the standard outlier detection metrics. These metrics are Area Under the Curve Receiver Operating Characteristic (AUROC ↑) and False Positive Rate at 95% True Positive Rate (FPR95 ↓), where True Positive is the correct detection of an inlier. For ID detection, we first compute Intersection over Union (IoU) between predicted bounding boxes and ground-truth boxes. Then, we compute the average precision (AP) for the standard range of IoU thresholds. Finally, the mean of the average precision over all known classes (mAP ↑) measures ID detection performance. + +# 4.2. Comparison with the State-of-the-Art + +Tables 1 and 2 compare our method with other approaches on image and video datasets, respectively. We use the results provided by VOS [10] and STUD [9] to show the performance of the compared approaches. Some evaluated approaches (except VOS and STUD) were developed for image-wide OD detection. However, these approaches can also be used to compare an outlier-aware object detection model since classification is one of the object detection tasks. No real outlier datasets were used during training to compare these approaches with our method. + +
IDMethodFPR95 ↓AUROC ↑mAP ↑
PASCAL-VOCGAN [29]OD: MS-COCO / OpenImages48.5
Energy [35]60.93 / 59.9783.67 / 82.67
CSI [47]56.89 / 58.6983.69 / 82.9848.7
VOS [10]59.91 / 57.4181.83 / 82.9548.1
FFS (ours)47.77 / 48.3389.00 / 87.5951.5
44.15 / 45.0889.71 / 88.2951.8
+ +Table 1. Main results on image datasets with Faster R-CNN model and RegNetX-4.0GF [40] as the backbone for VOS and FFS. + +
IDODMethodFPR95 ↓AUROC ↑mAP ↑
BDD100KmulImagesGAN [29]83.6570.3931.5
Energy [35]81.6269.4332.0
CSI [47]80.0074.9131.8
STUD [9]79.7576.5532.3
FFS (ours)76.6877.5336.2
Youtube-VISMS-COCOGAN [29]85.7572.9525.5
Energy [35]88.5467.8326.7
CSI [47]82.4371.8124.2
STUD [9]81.1474.8227.2
FFS (ours)83.0676.3727.6
+ +Table 2. Main results on video datasets with Faster R-CNN model and RegNetX-4.0GF [40] as the backbone network. + +In Table 1, we report a $7.58\%$ and a $6.73\%$ decrease in FPR95 for FFS compared to VOS when evaluated on MS-COCO and OpenImages, respectively. Additionally, we note an increase in AUROC values when compared with VOS and assessed on MS-COCO and OpenImages. We also obtained higher ID detection performance and a faster training time of 2.18 hr compared to 2.43 hr for VOS on the same hardware and batch size. In Table 2, FFS outperforms STUD, where we report a $3.85\%$ decrease in FPR95 and a $1.28\%$ increase in AUROC compared to STUD when the ID is BDD100K, and the OD is nuImages. We also register a significant $12\%$ increase in mAP performance compared to STUD. Our training time is 2.7 hrs compared to 11 hrs by STUD for the BDD100K dataset on the same hardware. FFS also improves STUD results for the Youtube-VIS as ID and MS-COCO as OD while maintaining high ID detection performance. For this dataset, we report a lower training + +time of 8.2 hrs compared to 11.3 hrs by STUD. + +
IDMethodFPR95 ↓AUROC ↑mAP ↑Time (h)
PASCAL-VOCVOS [10]49.6788.4348.912.37
FFS (ours)43.1289.8451.732.12
Youtube-VISSTUD [9]83.9374.5424.4211.03
FFS (ours)81.9076.3525.108.23
+ +In Table 3, we also show the performance while using ResNet-50 [18] as an alternative backbone and compare our results with VOS [10] and STUD [9] with MS-COCO as the outlier dataset. The results show that using ResNet-50 as the backbone achieves state-of-the-art ID and OD detection performance with our framework. Overall, Tables 1, 2 and 3 demonstrate the effectiveness of FFS for OD detection for images and videos with lower training time and high ID detection performance. + +# 4.3. Ablation Studies + +In this section, we perform an exhaustive study of meta-parameters that affects the performance of FFS. We also qualitatively compare OD detection results with the VOS approach. For uniformity, we fix the ID dataset as PASCAL-VOC and OD dataset as MS-COCO unless stated otherwise in the experiments. + +VOS plus rejection sampling: Sampling outliers from a class conditional Gaussian (as performed in VOS [10]) does not ensure that such an outlier will have lower likelihoods for other class Gaussians. Therefore, we compute the log-likelihoods of the outliers using all class Gaussians and reject those with a higher likelihood from at least one other class compared to its generator class. Table 4 shows the results for VOS [10], VOS plus rejection sampling (named as $\mathrm{VOS + }$ ), and FFS. $\mathrm{VOS + }$ enhances the OD detection compared to VOS [10] due to selecting more effective outliers, but it is still worse than FFS. Furthermore, $\mathrm{VOS + }$ 's training time is much higher due to the iterative process of computing log-likelihoods through each Gaussian class. + +Table 3. Performance evaluation with ResNet-50 backbone. + +
MethodFPR95 ↓AUROC ↑mAP (ID)↑Time (h)
VOS [10]OD: MS-COCO / OpenImages
48.22 / 52.7489.04 / 85.9751.502.43
VOS+45.59 / 47.9689.40 / 87.1951.747.38
FFS (ours)44.15 / 45.0889.71 / 88.2951.802.18
+ +Normalizing Flow: We conducted a thorough study on several types of normalizing flow models. We selected NICE [6], RealNVP [7], Glow [27] and GIN [45] in the chronological order for this experiment. We fixed the same coupling layers and each model's sub-network $s$ and $t$ configuration for an equivalent comparison. Table 5 shows that Glow outperforms NICE, RealNVP and GIN in terms of + +both FPR95 and AUROC scores. GIN performs best for ID detection but has a worse OD detection performance. As the primary task of FFS is OD detection, we chose Glow as our flow model for all our experiments. + +Table 4. Comparison of VOS, VOS+, and FFS (ours). The baseline mAP of a closed-set object detector is 51.35. + +
Flow modelFPR95 ↓AUROC ↑mAP ↑Time (h)
NICE [6]48.2388.4651.812.23
RealNVP [7]45.7688.9551.842.28
Glow [27]44.1589.7151.802.18
GIN [45]48.3188.3652.002.25
+ +Figure 5 (c) shows the effect of the number of coupling layers $M$ on OD detection performance. $M = 2$ produces the best results as the inlier features fit a suitable number of flow parameters. However, the OD detection performance degrades for $M > 2$ , and the training time significantly increases due to more trainable parameters in the network. Adding more trainable parameters requires more training data for effective training; otherwise, it might lead to overfitting and a more complex optimization, which should explain the degradation in performance. Hence, we fix $M = 2$ , and in Figure 5 (a), we study the effect of the number of fully-connected (fc) layers within $s$ and $t$ sub-networks of each coupling layer. There is a slight improvement in OD detection performance when two fc layers are employed. In Figure 5 (b), we fix the number of coupling and fc layers in each $s$ and $t$ sub-networks as two and adjust the number of neurons in each fc layer. We report a degradation in OD detection performance above 2048 neurons. + +Table 5. Validation of the Normalizing Flow models. + +
LregFPR95 ↓AUROC ↑mAP ↑Time (h)
CE69.9284.7248.702.27
JSD52.8385.7948.602.25
Hinge51.3286.3446.542.19
BCE (Ours)44.1589.7151.802.18
+ +Table 6. Validation of ${\mathcal{L}}_{reg}$ on OD detection performance. + +Regularization loss $\mathcal{L}_{reg}$ and its weightage $\alpha$ : We studied several regularization losses for our FFS framework. Firstly, we used a simple multi-class cross-entropy loss (CE) on the classification logits without including the energy module. Secondly, we evaluated Jensen-Shannon divergence (JSD) loss, where we minimized the divergence between the distribution of the classification logits of the outlier samples and uniform distribution. In Table 6, we show that our BCE based $\mathcal{L}_{reg}$ loss outperforms all other loss functions in terms of OD detection performance. The ID detection performance is also higher, even though the training time is comparable. In Figure 5 (e), we demonstrate the effect of changing the weightage $\alpha$ of $\mathcal{L}_{reg}$ for PASCAL-VOC as ID and MS-COCO as OD. As $\alpha$ increases, the OD detection performance improves. However, an even higher weightage degrades the OD detection performance, due to which a careful selection of $\alpha$ is desirable. + +![](images/b29c94f55e9d1a0791c4f3f6784b4b58951ea2015622fd90ba38e61586956595.jpg) +Figure 4. Visualization of OD detection on MS-COCO and OpenImages with VOS [10] (first and third row) and FFS (ours) (second and fourth row) trained on PASCAL-VOC 2012 images. Ideally, an object in the images should be labeled as OOD with a green bounding box. + +![](images/fb1fb7cd1ddcb6f09dee8ec5eac055b7d67f56de64e70e28037eb0264901d6a7.jpg) + +![](images/6527735aa9c0794b0a5f91177d5feb75e5df3bdefb56803e7e7fba5358d26c39.jpg) + +![](images/510ccb5007d373515b3648b0987f1c7f9a886af0cc282fd332c105efadf2e80a.jpg) + +![](images/1816e992538b3f66e6881e34f3210b99f35528804e6c67e2c0c4b004e2a0355d.jpg) + +![](images/35e77aa27926f14bb570c3f23553174584f682c6fe4628826bc80358077d63b9.jpg) +Figure 5. Validation of the flow architecture, loss weightage $\alpha$ , batch size, and projection sampling. The training time (in hours) is shown above the horizontal x axis in the plots. + +![](images/3dd833c7dfb3e5ad291d8ce94de0846ede80cd48a41060d729233d9440fa644b.jpg) + +Projection Sampling: Figure 5 (f) shows the results after varying synthetic outliers. The OD detection improves as we increase the number of sampled outliers $o_{s}$ but degrades with much higher values of $s$ . The coverage of log-likelihoods from many sampled outliers broadens even though the average log-likelihood is similar to the likelihood threshold $\delta$ . Hence, some outliers may lie inside, while others exist far outside the decision boundary leading to bad regularization. We report that rejection sampling achieves better OD detection performance when compared to the best results of projection sampling. Since projection sampling enforces the outliers to be close to the decision boundary, it may be insufficient to estimate the entire outlier manifold. In contrast, the rejection sampling synthesize outliers near the decision boundary and in the proper + +outlier manifold, leading to a better model regularization. + +Qualitative analysis: Figure 4 shows the OD detection results on MS-COCO and OpenImages with FFS and VOS trained on PASCAL-VOC. The results show that FFS outperforms VOS in recognizing outlier objects. Additionally, our method's softmax confidence score is lower than VOS when both methods misclassify the outlier objects as inliers. + +# 5. Conclusion + +We presented a new methodology for outlier-aware object detection that learns a combined data distribution of all inlier object classes. By sampling from the low-likelihood region of a jointly trained normalizing flow model, our approach generates suitable synthetic outlier samples for training the outlier detection head of our compound model. In contrast, previous approaches model inliers with class-conditional Gaussians, resulting in outlier samples that may have a high likelihood for some other class. We report state-of-the-art results for image and video datasets with standard outlier detection metrics while decreasing the model training time. We also show that simple rejection sampling contributes more useful synthetic outliers than projection sampling with a fixed likelihood threshold. In future work, one could validate this finding on other OD detectors. Our idea also extends to developing an outlier-aware instance segmentation model and an active learning scheme to correct the failure modes in weakly-supervised object detectors. + +Acknowledgement: This work primarily received funding from the German Federal Ministry of Education and Research (BMBF) under Software Campus (grant 01IS17044) and was supported by the Center for Scalable Data Analytics and Artificial Intelligence (ScaDS.AI) Dresden/Leipzig, Germany. The work was also partially funded by DFG grant 389792660 as part of TRR 248 - CPEC, Croatian Science Foundation (grant IP-2020-02-5851 ADEPT) and the Cluster of Excellence CeTI (EXC2050/1, grant 390696704). The authors gratefully acknowledge the Center for Information Services and HPC (ZIH) at TU Dresden for providing computing resources. + +# References + +[1] Petra Bevandic, Ivan Kreso, Marin Orsic, and Sinisa Segvic. Dense open-set recognition based on training with noisy negative images. Image Vis. Comput., 2022. 1 +[2] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuScenes: A multimodal dataset for autonomous driving. In CVPR, 2020. 6 +[3] Jiwoong Choi, Dayoung Chun, Hyun Kim, and Hyuk-Jae Lee. Gaussian YOLOv3: An accurate and fast object detector using localization uncertainty for autonomous driving. In ICCV, 2019. 1 +[4] Kumari Deepshikha, Sai Harsha Yelleni, P. K. Srijith, and C. Krishna Mohan. Monte Carlo dropblock for modelling uncertainty in object detection. CoRR, 2021. 2 +[5] Akshay Raj Dhamija, Manuel Günther, Jonathan Ventura, and Terrance E. Boult. The overlooked elephant of object detection: Open Set. In WACV, 2020. 2 +[6] Laurent Dinh, David Krueger, and Yoshua Bengio. NICE: Non-linear Independent components estimation. In ICLR, 2015. 7 +[7] Laurent Dinh, Jascha Sohl-Dickstein, and Samy Bengio. Density estimation using Real NVP. In ICLR, 2017. 7 +[8] Keval Doshi and Yasin Yilmaz. Any-shot sequential anomaly detection in surveillance videos. In CVPR Workshops, 2020. 2 +[9] Xuefeng Du, Xin Wang, Gabriel Gozum, and Yixuan Li. Unknown-aware Object Detection: Learning what you don't know from videos in the wild. In CVPR, 2022. 1, 2, 3, 6, 7, 12, 13, 14 +[10] Xuefeng Du, Zhaoning Wang, Mu Cai, and Yixuan Li. VOS: Learning what you don't know by virtual outlier synthesis. In ICLR, 2022. 1, 2, 5, 6, 7, 8, 11, 12 +[11] Mark Everingham, Luc Van Gool, Christopher K. I. Williams, John M. Winn, and Andrew Zisserman. The Pascal visual object classes (VOC) challenge. International Journal of Computer Vision, 2010. 1, 2, 6 +[12] Yarin Gal and Zoubin Ghahramani. Dropout as a Bayesian Approximation: Representing model uncertainty in deep learning. In ICML, 2016. 2 +[13] Ross Girshick. Fast R-CNN. In ICCV, 2015. 2 +[14] Matej Grcic, Petra Bevandic, and Sinisa Segvic. Dense open-set recognition with synthetic outliers generated by Real NVP. CoRR, 2020. 2 +[15] David Hall, Feras Dayoub, John Skinner, Haoyang Zhang, Dimity Miller, Peter Corke, Gustavo Carneiro, Anelia Angelova, and Niko Sunderhauf. Probabilistic Object Detection: Definition and Evaluation. In WACV, 2020. 2 +[16] Ali Harakeh and Steven L. Waslander. Estimating and evaluating regression predictive uncertainty in deep object detectors. In ICLR, 2021. 2 +[17] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask R-CNN. In ICCV, 2017. 1, 2 +[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 7 + +[19] Dan Hendrycks and Kevin Gimpel. A baseline for detecting misclassified and out-of-distribution examples in neural networks. In ICLR, 2017. 1, 2 +[20] Dan Hendrycks, Mantas Mazeika, and Thomas Dietterich. Deep anomaly detection with outlier exposure. In ICLR, 2019. 1, 2 +[21] Dan Hendrycks, Mantas Mazeika, Saurav Kadavath, and Dawn Song. Using self-supervised learning can improve model robustness and uncertainty. In NeurIPS, 2019. 2 +[22] Dan Hendrycks, Norman Mu, Ekin D. Cubuk, Barret Zoph, Justin Gilmer, and Balaji Lakshminarayanan. AugMix: A simple data processing method to improve robustness and uncertainty. In ICLR, 2020. 2 +[23] Yen-Chang Hsu, Yilin Shen, Hongxia Jin, and Zsolt Kira. Generalized ODIN: Detecting out-of-distribution image without learning from out-of-distribution data. In CVPR, 2020. 2 +[24] Rui Huang, Andrew Geng, and Yixuan Li. On the importance of gradients for detecting distributional shifts in the wild. In NeurIPS, 2021. 1, 2 +[25] Radu Tudor Ionescu, Fahad Shahbaz Khan, Mariana-Iuliana Georgescu, and Ling Shao. Object-centric auto-encoders and dummy anomalies for abnormal event detection in video. In CVPR, 2019. 2 +[26] K. J. Joseph, Salman Khan, Fahad Shahbaz Khan, and Vineeth N. Balasubramanian. Towards open world object detection. In CVPR, 2021. 2 +[27] Durk P. Kingma and Prafulla Dhariwal. Glow: Generative flow with invertible 1x1 convolutions. In NeurIPS, 2018. 2, 4, 6, 7 +[28] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The Open Images dataset v4. International Journal of Computer Vision, 2020. 6 +[29] Kimin Lee, Honglak Lee, Kibok Lee, and Jinwoo Shin. Training confidence-calibrated classifiers for detecting out-of-distribution samples. In ICLR, 2018. 1, 2, 6 +[30] Kimin Lee, Kibok Lee, Honglak Lee, and Jinwoo Shin. A simple unified framework for detecting out-of-distribution samples and adversarial attacks. In NeurIPS, 2018. 1, 2 +[31] Shiyu Liang, Yixuan Li, and Rayadurgam Srikant. Enhancing the reliability of out-of-distribution image detection in neural networks. In ICLR, 2018. 1, 2 +[32] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, Lubomir D. Bourdev, Ross B. Girshick, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollar, and C. Lawrence Zitnick. Microsoft COCO: Common objects in context. In ECCV, 2014. 6 +[33] Ziqian Lin, Sreya D. Roy, and Yixuan Li. MOOD: Multi-level out-of-distribution detection. In CVPR, 2021. 2 +[34] Wen Liu, Weixin Luo, Dongze Lian, and Shenghua Gao. Future frame prediction for anomaly detection - A new baseline. In CVPR, 2018. 3 +[35] Weitang Liu, Xiaoyun Wang, John Owens, and Yixuan Li. Energy-based out-of-distribution detection. In NeurIPS, 2020. 1, 2, 5, 6 + +[36] Thomas Lucas, Konstantin Shmelkov, Karteek Alahari, Cordelia Schmid, and Jakob Verbeek. Adaptive density estimation for generative models. In NeurIPS, 2019. 1 +[37] Dimity Miller, Feras Dayoub, Michael Milford, and Niko Sünderhauf. Evaluating merging strategies for sampling-based uncertainty techniques in object detection. In ICRA, 2019. 2 +[38] Dimity Miller, Lachlan Nicholson, Feras Dayoub, and Niko Sünderhauf. Dropout sampling for robust object detection in open-set conditions. In ICRA, 2018. 2 +[39] Sina Mohseni, Mandar Pitale, JBS Yadawa, and Zhangyang Wang. Self-supervised learning for generalizable out-of-distribution detection. In AAAI, 2020. 1, 2 +[40] Ilija Radosavovic, Raj Prateek Kosaraju, Ross B. Girshick, Kaiming He, and Piotr Dólár. Designing network design spaces. In CVPR, 2020. 6 +[41] Mahdyar Ravanbakhsh, Moin Nabi, Hossein Mousavi, Enver Sangineto, and Nicu Sebe. Plug-and-play CNN for crowd motion analysis: An application in abnormal event detection. In WACV, 2018. 3 +[42] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster R-CNN: Towards real-time object detection with region proposal networks. In NeurIPS, 2015. 1, 2, 6 +[43] Tobias Riedlinger, Matthias Rottmann, Marius Schubert, and Hanno Gottschalk. Gradient-based quantification of epistemic uncertainty for deep object detectors. CoRR, 2021. 2 +[44] Chandramouli Shama Sastry and Sageev Oore. Detecting out-of-distribution examples with Gram Matrices. In ICML, 2020. 1, 2 +[45] Peter Sorrenson, Carsten Rother, and Ullrich Kothe. Disentanglement by nonlinear ICA with general incompressible-flow networks (GIN). In *ICLR*, 2020. 7 +[46] Yiyou Sun, Chuan Guo, and Yixuan Li. ReAct: Out-of-distribution detection with rectified activations. In NeurIPS, 2021. 2 +[47] Jihoon Tack, Sangwoo Mo, Jongheon Jeong, and Jinwoo Shin. CSI: Novelty detection via contrastive learning on distributionally shifted instances. In NeurIPS, 2020. 2, 6 +[48] Haoran Wang, Weitang Liu, Alex Bocchieri, and Yixuan Li. Can multi-label classification networks know what they don't know? In NeurIPS, 2021. 2 +[49] Xin Wang, Thomas E. Huang, Benlin Liu, Fisher Yu, Xiaolong Wang, Joseph E. Gonzalez, and Trevor Darrell. Robust object detection via instance-level temporal cycle confusion. In ICCV, 2021. 3 +[50] Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. Detector2. https://github.com/facebookresearch/detectron2, 2019.6 +[51] Linjie Yang, Yuchen Fan, and Ning Xu. Video instance segmentation. In ICCV, 2019. 2, 6 +[52] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darrell. BDD100K: A diverse driving dataset for heterogeneous multitask learning. In CVPR, 2020. 2, 6 +[53] Jingyang Zhang, Nathan Inkawich, Yiran Chen, and Hai Li. Fine-grained out-of-distribution detection with mixup outlier exposure. CoRR, 2021. 1 \ No newline at end of file diff --git a/2023/Normalizing Flow Based Feature Synthesis for Outlier-Aware Object Detection/images.zip b/2023/Normalizing Flow Based Feature Synthesis for Outlier-Aware Object Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..5238a52f970492682e3f11db1e12832db76558ef --- /dev/null +++ b/2023/Normalizing Flow Based Feature Synthesis for Outlier-Aware Object Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2bafbef28b58f51e09f5758320de58950794e5972ac063a16a0cb321a4668a41 +size 535562 diff --git a/2023/Normalizing Flow Based Feature Synthesis for Outlier-Aware Object Detection/layout.json b/2023/Normalizing Flow Based Feature Synthesis for Outlier-Aware Object Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..8fc0c3acd819b156c40152e93319bff5b0adde04 --- /dev/null +++ b/2023/Normalizing Flow Based Feature Synthesis for Outlier-Aware Object Detection/layout.json @@ -0,0 +1,11655 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 55, + 103, + 539, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 103, + 539, + 121 + ], + "spans": [ + { + "bbox": [ + 55, + 103, + 539, + 121 + ], + "type": "text", + "content": "Normalizing Flow based Feature Synthesis for Outlier-Aware Object Detection" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 119, + 142, + 473, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 142, + 473, + 172 + ], + "spans": [ + { + "bbox": [ + 119, + 142, + 473, + 172 + ], + "type": "text", + "content": "Nishant Kumar1, Sinisa Šegvic2, Abouzar Eslami3, and Stefan Gumhold1 \n1TU Dresden, 2University of Zagreb - FER, 3Carl Zeiss Meditec AG" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 199, + 192, + 211 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 199, + 192, + 211 + ], + "spans": [ + { + "bbox": [ + 143, + 199, + 192, + 211 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 224, + 289, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 224, + 289, + 464 + ], + "spans": [ + { + "bbox": [ + 46, + 224, + 289, + 464 + ], + "type": "text", + "content": "Real-world deployment of reliable object detectors is crucial for applications such as autonomous driving. However, general-purpose object detectors like Faster R-CNN are prone to providing overconfident predictions for outlier objects. Recent outlier-aware object detection approaches estimate the density of instance-wide features with class-conditional Gaussians and train on synthesized outlier features from their low-likelihood regions. However, this strategy does not guarantee that the synthesized outlier features will have a low likelihood according to the other class-conditional Gaussians. We propose a novel outlier-aware object detection framework that distinguishes outliers from inlier objects by learning the joint data distribution of all inlier classes with an invertible normalizing flow. The appropriate sampling of the flow model ensures that the synthesized outliers have a lower likelihood than inliers of all object classes, thereby modeling a better decision boundary between inlier and outlier objects. Our approach significantly outperforms the state-of-the-art for outlier-aware object detection on both image and video datasets." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 488, + 128, + 501 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 488, + 128, + 501 + ], + "spans": [ + { + "bbox": [ + 47, + 488, + 128, + 501 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 509, + 287, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 509, + 287, + 629 + ], + "spans": [ + { + "bbox": [ + 46, + 509, + 287, + 629 + ], + "type": "text", + "content": "General purpose object detectors such as Faster R-CNN [42] and Mask R-CNN [17] deliver high performance for inlier images. However, in many real-world scenarios, such as autonomous driving [3], plenty of unknown outliers (OD) naturally occur in an image or a video scene. Due to the co-existence of OD with the labeled inlier (ID) objects in the scene, object detectors confuse outliers with inliers. Therefore, reliable object-detection deployments require detecting such anomalies without degrading the performance of inlier object detection." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "content": "Many outlier detection approaches focus on multi-class image classification task by either performing outlier detection during inference [19, 24, 30, 31, 35, 44] or training on real outlier data [1, 20, 39, 53]. However, such OD inputs are unaware of the decision boundary between inliers and outliers, resulting in an inaccurate model regularization. A popular work [29] proposed a novel training scheme to" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 305, + 186, + 428, + 276 + ], + "blocks": [ + { + "bbox": [ + 305, + 186, + 428, + 276 + ], + "lines": [ + { + "bbox": [ + 305, + 186, + 428, + 276 + ], + "spans": [ + { + "bbox": [ + 305, + 186, + 428, + 276 + ], + "type": "image", + "image_path": "dda989cc0da06573cc0da40cd95252352f3eb721bd2e219c5b3720f92d6f4f60.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 284, + 547, + 361 + ], + "lines": [ + { + "bbox": [ + 305, + 284, + 547, + 361 + ], + "spans": [ + { + "bbox": [ + 305, + 284, + 547, + 361 + ], + "type": "text", + "content": "Figure 1. We compare distributions of log-likelihoods for inliers (ID), background patches, and synthetic outlier (OD) samples by applying a pre-trained model on Pascal-VOC [11] validation set. The left plot shows the likelihoods of class-conditional Gaussians from the official approach of VOS [10]. The right plot shows the likelihoods recovered with our normalizing flow. We observe that our flow separates the three distributions much better than VOS." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 425, + 187, + 545, + 275 + ], + "blocks": [ + { + "bbox": [ + 425, + 187, + 545, + 275 + ], + "lines": [ + { + "bbox": [ + 425, + 187, + 545, + 275 + ], + "spans": [ + { + "bbox": [ + 425, + 187, + 545, + 275 + ], + "type": "image", + "image_path": "ae0c5fef7b414ce8919dd718006553cb79002053a945a157f2096ea0be419a3e.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 374, + 547, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 374, + 547, + 471 + ], + "spans": [ + { + "bbox": [ + 304, + 374, + 547, + 471 + ], + "type": "text", + "content": "generate synthetic outlier samples in the high-dimensional pixel space using Generative Adversarial Networks for outlier aware training of an image classification model. However, GANs are challenging to optimize and are likely to deliver insufficient coverage [36]. Moreover, the previous work [29] generates an entire image as an outlier sample, whereas in an object detection problem, both inliers and outliers can coexist in the same image space." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 472, + 548, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 548, + 664 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 548, + 664 + ], + "type": "text", + "content": "Recently, [10] proposed to model the inlier features as class-conditional Gaussians and then synthesize OD features from the low likelihood region of the modeled distribution. Even though density is more easily estimated in the feature space than in pixel space, the assumption of class-conditional Gaussian may not provide an accurate decision boundary between inliers and outliers. Furthermore, synthesizing a low-likelihood OD sample from the Gaussian for class A does not guarantee that the sample is also of the low likelihood for the Gaussian of class B, as indicated in the left plot of Figure 1. Recent work for object detection in video [9] extracts the background patches for uncertainty regularization by thresholding the dissimilarity with the reference inlier features. However, the extracted background patches may lie far from the inlier-outlier decision boundary, leading to sub-optimal regularization of the model." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": "We propose a novel approach for open-set object detection which we call Flow Feature Synthesis (FFS). Our approach trains an invertible normalizing flow to map the data distribution of ID features of all object classes to a latent" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "5156" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 215 + ], + "type": "text", + "content": "representation that conforms to a multivariate Gaussian distribution with zero mean and diagonal unit covariance. As a result, it ensures principled estimation of the complex distribution of inlier features from all classes. We synthesize outlier features from low-likelihood regions of the learned distribution. This enables outlier features to be near the ID-OD decision boundary, leading to more robust uncertainty regularization of the object detector. In contrast, using generative adversarial models for the same task would be cumbersome due to its inability to infer density of the generated samples, Furthermore, variational autoencoders can only infer the lower bound of the sample density [27]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 216, + 289, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 216, + 289, + 418 + ], + "spans": [ + { + "bbox": [ + 46, + 216, + 289, + 418 + ], + "type": "text", + "content": "Specifically, FFS optimizes the normalizing flow model by maximum likelihood training on ID features. This training scheme enables the model to estimate the actual data distribution of the available ID features. Next, FFS utilizes the invertibility of the flow model to randomly sample from its latent space and generate synthetic features in the reverse direction of the model. The normalizing flow allows efficient and exact inference of the distribution density in the generated samples. Consequently, our approach can deliver suitable synthetic outlier data in fewer iterations than VOS [10]. It also requires fewer synthetically generated samples than VOS [10] to obtain OD features from the low-likelihood region of the modeled distribution by the flow. Furthermore, we developed our end-to-end trainable FFS framework to be effective on both image and video datasets. In contrast, previous works [9, 10] proposed standalone strategies for each task. Our main contributions are:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 59, + 422, + 287, + 637 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 59, + 422, + 287, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 422, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 59, + 422, + 287, + 480 + ], + "type": "text", + "content": "- We present a new outlier-aware object detection framework that utilizes Normalizing Flows to model the joint data distribution of inlier features. Invertibility of the flow allows efficient generation of synthetic outliers for effective uncertainty regularization." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 482, + 287, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 482, + 287, + 540 + ], + "spans": [ + { + "bbox": [ + 59, + 482, + 287, + 540 + ], + "type": "text", + "content": "- By mapping the data distribution of inlier features from all object classes to a multivariate Normal distribution in the flow's latent space, FFS ensures that an outlier sampled using the flow model is OD with reference to all ID classes." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 541, + 287, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 541, + 287, + 577 + ], + "spans": [ + { + "bbox": [ + 59, + 541, + 287, + 577 + ], + "type": "text", + "content": "- FFS achieves better OD detection performance while training faster than VOS [10] and STUD [9], due to having to generate fewer synthetic samples." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 578, + 287, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 578, + 287, + 637 + ], + "spans": [ + { + "bbox": [ + 59, + 578, + 287, + 637 + ], + "type": "text", + "content": "- We show that our method achieves state-of-the-art performance in OD object detection while preserving the baseline ID detection performance for image dataset PASCAL-VOC [11] and video datasets such as Youtube-VIS [51] and BDD100K [52]." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 646, + 132, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 646, + 132, + 658 + ], + "spans": [ + { + "bbox": [ + 47, + 646, + 132, + 658 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": "The research in outlier-aware object detection is in the early stages compared to works in vanilla object detection [13, 17, 42] or outlier-aware semantic segmentation [14]. This section overviews prior approaches to outlier" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": "detection for image classification and proceeds to outlier-aware object detection for images and videos." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 99, + 547, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 99, + 547, + 253 + ], + "spans": [ + { + "bbox": [ + 304, + 99, + 547, + 253 + ], + "type": "text", + "content": "Outlier-aware image classification. Past approaches trained a classification model without knowledge about outliers and conducted OD detection during inference. For example, [19] used simple softmax probability scores during inference to detect OD samples, while other works used Mahalanobis distance [30], rectified activations [46], KL divergence [24] and Gram Matrices [44] instead of softmax scores to detect such samples. On the other hand, ODIN [31] and Generalized ODIN [23] performed perturbations to the test examples to enhance the performance of softmax function for OD detection. Recently, energy function [33, 35, 48] has performed better for distinguishing ID and OD samples than softmax scores." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 254, + 547, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 254, + 547, + 397 + ], + "spans": [ + { + "bbox": [ + 304, + 254, + 547, + 397 + ], + "type": "text", + "content": "Several other methods [20-22, 39, 47] proposed to regularize image classifiers by exposing them with outlier data during self-supervised training while a GAN based approach [29] synthesizes outliers in pixel space and uses it for outlier-aware image classification. However, such synthetic outliers are sampled from imprecise decision boundaries. Additionally, these approaches may not be applied to an object detection task since outliers defined by such methods are in the entire pixel space of the input image. In contrast, an image may contain both inlier and outlier objects simultaneously, and a trustworthy object detection model should detect such scenarios." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 400, + 545, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 400, + 545, + 651 + ], + "spans": [ + { + "bbox": [ + 304, + 400, + 545, + 651 + ], + "type": "text", + "content": "Outlier-aware object detection for images. One of the early open-set object detection works [5] highlighted the importance of adding an extra background class for unknown objects, though object detection inherently rejects unknown object proposals. Several works improve the background detection in open-set conditions [4, 15, 37, 38] by using Bayesian techniques, such as Monte Carlo-Dropout [12]. However, such methods have high latency during inference which could be sub-optimal for real-world applications. Others [16, 43] presented uncertainty regularization for localization regression but did not examine outlier-aware object detection with a focus on classification head. Works like [26] used the background patches as OD samples for model regularization, but the performance might be low as the samples may be far from the decision boundary. VOS [10] applies energy-based regularization [35] to synthetic OD features generated by class-conditional Gaussians. Instead, our method map features from all ID classes to a Gaussian prior with a normalizing flow model, and thus achieves a tighter decision boundary and improved performance." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "content": "Outlier-aware object detection for videos. Several works aim to identify anomalous events on the object and frame level. An object-level approach [8] uses the " + }, + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "content": "-nearest neighbors algorithm to detect anomalous objects while [25] use " + }, + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "content": "-means to cluster inlier features and train " + }, + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "content": " binary clas" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5157" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 70, + 547, + 200 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 547, + 200 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 547, + 200 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 547, + 200 + ], + "type": "image", + "image_path": "d3bcad9a4983c52f1d00edd606f84534214c0ba4453530966fec310a8a314428.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 201, + 547, + 246 + ], + "lines": [ + { + "bbox": [ + 46, + 201, + 547, + 246 + ], + "spans": [ + { + "bbox": [ + 46, + 201, + 547, + 246 + ], + "type": "text", + "content": "Figure 2. Overview of the FFS framework. The ID features " + }, + { + "bbox": [ + 46, + 201, + 547, + 246 + ], + "type": "inline_equation", + "content": "l_{ID}(x,\\hat{b})" + }, + { + "bbox": [ + 46, + 201, + 547, + 246 + ], + "type": "text", + "content": " are selected for training the normalizing flow with the negative log-likelihood loss " + }, + { + "bbox": [ + 46, + 201, + 547, + 246 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{nll}" + }, + { + "bbox": [ + 46, + 201, + 547, + 246 + ], + "type": "text", + "content": ". We recover outlier features " + }, + { + "bbox": [ + 46, + 201, + 547, + 246 + ], + "type": "inline_equation", + "content": "o_s" + }, + { + "bbox": [ + 46, + 201, + 547, + 246 + ], + "type": "text", + "content": " by iterative sampling in the latent space of the normalizing flow. The classification logits of the ID features " + }, + { + "bbox": [ + 46, + 201, + 547, + 246 + ], + "type": "inline_equation", + "content": "l_{ID}(x,\\hat{b})" + }, + { + "bbox": [ + 46, + 201, + 547, + 246 + ], + "type": "text", + "content": " and the outlier features " + }, + { + "bbox": [ + 46, + 201, + 547, + 246 + ], + "type": "inline_equation", + "content": "o_s" + }, + { + "bbox": [ + 46, + 201, + 547, + 246 + ], + "type": "text", + "content": " are used to calculate the energy scores. The uncertainty loss " + }, + { + "bbox": [ + 46, + 201, + 547, + 246 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{reg}" + }, + { + "bbox": [ + 46, + 201, + 547, + 246 + ], + "type": "text", + "content": " encourages separation between the energy scores of inlier and outlier features. More details are given in Section 3." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 255, + 289, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 255, + 289, + 494 + ], + "spans": [ + { + "bbox": [ + 46, + 255, + 289, + 494 + ], + "type": "text", + "content": "sifiers so that each cluster can view other clusters as an outlier. A test object is labeled an outlier if the highest classification score is negative. Frame-level approaches [34, 41] predicted future frames that preserve the events of the previous frames by ensuring the optical flow of the predicted frame remains consistent with ground-truth. In contrast, events with a significant difference between prediction and ground-truth are labeled outliers. However, no previous work synthesizes outliers from video frames for model regularization. Recent work [49] introduced self-supervised learning to find the most distinct object in the next video frame given an inlier object in the previous frame, leading to more knowledge about unseen data domains. STUD [9] uses the inliers in the first frame to select the outliers in the following frames for model regularization. However, this may deliver outlier patches that are very far away from the decision boundary and thus lead to inferior model regularization. To summarize, no single outlier-aware object detection framework exists for images and video with an effective regularization that models a precise inlier manifold." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 511, + 104, + 523 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 511, + 104, + 523 + ], + "spans": [ + { + "bbox": [ + 47, + 511, + 104, + 523 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 534, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 289, + 713 + ], + "type": "text", + "content": "Outlier-aware object detection problem is much more complex than detecting outliers for an image classification task since a real-world image may consist of both inlier and outlier objects. Let us define an input " + }, + { + "bbox": [ + 46, + 534, + 289, + 713 + ], + "type": "inline_equation", + "content": "x \\in \\mathcal{X}" + }, + { + "bbox": [ + 46, + 534, + 289, + 713 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 534, + 289, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{X} := (x_1, x_2, \\dots, x_N)" + }, + { + "bbox": [ + 46, + 534, + 289, + 713 + ], + "type": "text", + "content": " being the training dataset consisting of " + }, + { + "bbox": [ + 46, + 534, + 289, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 534, + 289, + 713 + ], + "type": "text", + "content": " images or video frames, ground-truth class labels for each ID object as " + }, + { + "bbox": [ + 46, + 534, + 289, + 713 + ], + "type": "inline_equation", + "content": "y \\in \\mathcal{Y}" + }, + { + "bbox": [ + 46, + 534, + 289, + 713 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 534, + 289, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{Y} := \\{1, 2, \\dots, K\\}" + }, + { + "bbox": [ + 46, + 534, + 289, + 713 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 46, + 534, + 289, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 534, + 289, + 713 + ], + "type": "text", + "content": " classes and the coordinates of ground-truth bounding boxes " + }, + { + "bbox": [ + 46, + 534, + 289, + 713 + ], + "type": "inline_equation", + "content": "b \\in \\mathcal{B}" + }, + { + "bbox": [ + 46, + 534, + 289, + 713 + ], + "type": "text", + "content": ". Then, an end-to-end object detector with parameters " + }, + { + "bbox": [ + 46, + 534, + 289, + 713 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 534, + 289, + 713 + ], + "type": "text", + "content": " treated as a conditional probability distribution estimates bounding boxes as " + }, + { + "bbox": [ + 46, + 534, + 289, + 713 + ], + "type": "inline_equation", + "content": "p_{\\theta}(b|x)" + }, + { + "bbox": [ + 46, + 534, + 289, + 713 + ], + "type": "text", + "content": " and the object class as " + }, + { + "bbox": [ + 46, + 534, + 289, + 713 + ], + "type": "inline_equation", + "content": "p_{\\theta}(y|x,b)" + }, + { + "bbox": [ + 46, + 534, + 289, + 713 + ], + "type": "text", + "content": ". Figure 2 shows the training and inference scheme of the FFS framework. The backbone network extracts the feature maps at different resolutions, and the Region Proposal Network (RPN) converts these" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 255, + 547, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 255, + 547, + 388 + ], + "spans": [ + { + "bbox": [ + 304, + 255, + 547, + 388 + ], + "type": "text", + "content": "multi-resolution feature maps into object proposals. The ROI align module classifies object proposals and provides a vector of bounding boxes, " + }, + { + "bbox": [ + 304, + 255, + 547, + 388 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{b}}" + }, + { + "bbox": [ + 304, + 255, + 547, + 388 + ], + "type": "text", + "content": ", and fixed-size box features, " + }, + { + "bbox": [ + 304, + 255, + 547, + 388 + ], + "type": "inline_equation", + "content": "l(x,\\hat{\\pmb{b}})" + }, + { + "bbox": [ + 304, + 255, + 547, + 388 + ], + "type": "text", + "content": ", of both ID (inlier) and background patches. Note that " + }, + { + "bbox": [ + 304, + 255, + 547, + 388 + ], + "type": "inline_equation", + "content": "l(x,\\hat{\\pmb{b}})" + }, + { + "bbox": [ + 304, + 255, + 547, + 388 + ], + "type": "text", + "content": " is in a lower dimensional space than " + }, + { + "bbox": [ + 304, + 255, + 547, + 388 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 255, + 547, + 388 + ], + "type": "text", + "content": ". The box features " + }, + { + "bbox": [ + 304, + 255, + 547, + 388 + ], + "type": "inline_equation", + "content": "l(x,\\hat{\\pmb{b}})" + }, + { + "bbox": [ + 304, + 255, + 547, + 388 + ], + "type": "text", + "content": " and bounding boxes " + }, + { + "bbox": [ + 304, + 255, + 547, + 388 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{b}}" + }, + { + "bbox": [ + 304, + 255, + 547, + 388 + ], + "type": "text", + "content": " are used to train the object classification and box prediction heads according to the object detection loss " + }, + { + "bbox": [ + 304, + 255, + 547, + 388 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{det}" + }, + { + "bbox": [ + 304, + 255, + 547, + 388 + ], + "type": "text", + "content": ", given the ground-truth ID classes and box coordinates, respectively. The box features from " + }, + { + "bbox": [ + 304, + 255, + 547, + 388 + ], + "type": "inline_equation", + "content": "l(x,\\hat{\\pmb{b}})" + }, + { + "bbox": [ + 304, + 255, + 547, + 388 + ], + "type": "text", + "content": " with the ground-truth class as background are filtered out to keep only ID features " + }, + { + "bbox": [ + 304, + 255, + 547, + 388 + ], + "type": "inline_equation", + "content": "l_{ID}(x,\\hat{\\pmb{b}})" + }, + { + "bbox": [ + 304, + 255, + 547, + 388 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 396, + 373, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 396, + 373, + 407 + ], + "spans": [ + { + "bbox": [ + 305, + 396, + 373, + 407 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 415, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 415, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 415, + 547, + 715 + ], + "type": "text", + "content": "Our framework FFS is split into three stages: the training procedure, the threshold estimation for OD detection, and the model inference as summarized in Algorithm 1. During the training stage, we compute the standard object detection loss " + }, + { + "bbox": [ + 304, + 415, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{det}" + }, + { + "bbox": [ + 304, + 415, + 547, + 715 + ], + "type": "text", + "content": " via the chosen general-purpose object detector. We define maximum-likelihood training objective " + }, + { + "bbox": [ + 304, + 415, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{nll}" + }, + { + "bbox": [ + 304, + 415, + 547, + 715 + ], + "type": "text", + "content": " to train our normalizing flow module on inlier features " + }, + { + "bbox": [ + 304, + 415, + 547, + 715 + ], + "type": "inline_equation", + "content": "l_{ID}(x,\\hat{\\pmb{b}})" + }, + { + "bbox": [ + 304, + 415, + 547, + 715 + ], + "type": "text", + "content": ". Subsequently, we sample synthetic outliers " + }, + { + "bbox": [ + 304, + 415, + 547, + 715 + ], + "type": "inline_equation", + "content": "o_s" + }, + { + "bbox": [ + 304, + 415, + 547, + 715 + ], + "type": "text", + "content": " from the flow model. The details related to the normalizing flow and the procedure to synthesize outlier features are given in Section 3.2. We then compute energy scores for inliers and outliers for the model regularization using " + }, + { + "bbox": [ + 304, + 415, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{reg}" + }, + { + "bbox": [ + 304, + 415, + 547, + 715 + ], + "type": "text", + "content": ". Section 3.3 describes the details about model regularization, and Section 3.4 discusses the overall training objective of the FFS framework. We use the trained FFS model for threshold estimation to perform OD detection. We fix the energy-based threshold " + }, + { + "bbox": [ + 304, + 415, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 304, + 415, + 547, + 715 + ], + "type": "text", + "content": " such that the model correctly detects " + }, + { + "bbox": [ + 304, + 415, + 547, + 715 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 304, + 415, + 547, + 715 + ], + "type": "text", + "content": " of inlier objects in the validation set. During inference, given an object " + }, + { + "bbox": [ + 304, + 415, + 547, + 715 + ], + "type": "inline_equation", + "content": "x_t" + }, + { + "bbox": [ + 304, + 415, + 547, + 715 + ], + "type": "text", + "content": " in a test image, the trained FFS model provides a bounding box for the object as " + }, + { + "bbox": [ + 304, + 415, + 547, + 715 + ], + "type": "inline_equation", + "content": "b_t" + }, + { + "bbox": [ + 304, + 415, + 547, + 715 + ], + "type": "text", + "content": ". The decision of whether the object is an inlier or an outlier lies on the energy " + }, + { + "bbox": [ + 304, + 415, + 547, + 715 + ], + "type": "inline_equation", + "content": "E(h(l(x_t,b_t);\\theta))" + }, + { + "bbox": [ + 304, + 415, + 547, + 715 + ], + "type": "text", + "content": ". Given an energy threshold " + }, + { + "bbox": [ + 304, + 415, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 304, + 415, + 547, + 715 + ], + "type": "text", + "content": ", we assign " + }, + { + "bbox": [ + 304, + 415, + 547, + 715 + ], + "type": "inline_equation", + "content": "x_t" + }, + { + "bbox": [ + 304, + 415, + 547, + 715 + ], + "type": "text", + "content": " as an inlier if " + }, + { + "bbox": [ + 304, + 415, + 547, + 715 + ], + "type": "inline_equation", + "content": "E(h(l(x_t,b_t);\\theta)) < \\xi" + }, + { + "bbox": [ + 304, + 415, + 547, + 715 + ], + "type": "text", + "content": " and OD if " + }, + { + "bbox": [ + 304, + 415, + 547, + 715 + ], + "type": "inline_equation", + "content": "E(h(l(x_t,b_t);\\theta)) \\geq \\xi" + }, + { + "bbox": [ + 304, + 415, + 547, + 715 + ], + "type": "text", + "content": ". We obtain object class labels for predicted inliers with the softmax confidence score" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5158" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "type": "text", + "content": "and bounding box. We label the bounding box for objects detected as outliers as OOD." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 101, + 286, + 125 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 101, + 286, + 125 + ], + "spans": [ + { + "bbox": [ + 47, + 101, + 286, + 125 + ], + "type": "text", + "content": "Algorithm 1 FFS: Normalizing Flow based Feature Synthesis for Outlier-Aware Object Detection" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 128, + 286, + 165 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 128, + 286, + 165 + ], + "spans": [ + { + "bbox": [ + 47, + 128, + 286, + 165 + ], + "type": "text", + "content": "Input: Training data " + }, + { + "bbox": [ + 47, + 128, + 286, + 165 + ], + "type": "inline_equation", + "content": "\\{(x_{i},\\mathbf{b}_{i},y_{i})\\}_{i = 1}^{N}" + }, + { + "bbox": [ + 47, + 128, + 286, + 165 + ], + "type": "text", + "content": ", vanilla object detector, normalizing flow " + }, + { + "bbox": [ + 47, + 128, + 286, + 165 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 47, + 128, + 286, + 165 + ], + "type": "text", + "content": " and binary classifier " + }, + { + "bbox": [ + 47, + 128, + 286, + 165 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 47, + 128, + 286, + 165 + ], + "type": "text", + "content": " with parameters " + }, + { + "bbox": [ + 47, + 128, + 286, + 165 + ], + "type": "inline_equation", + "content": "\\theta ,\\gamma" + }, + { + "bbox": [ + 47, + 128, + 286, + 165 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 128, + 286, + 165 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 47, + 128, + 286, + 165 + ], + "type": "text", + "content": " respectively, the loss weights " + }, + { + "bbox": [ + 47, + 128, + 286, + 165 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 47, + 128, + 286, + 165 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 128, + 286, + 165 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 47, + 128, + 286, + 165 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 166, + 251, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 166, + 251, + 177 + ], + "spans": [ + { + "bbox": [ + 47, + 166, + 251, + 177 + ], + "type": "text", + "content": "Output: Object detector with trained parameter " + }, + { + "bbox": [ + 47, + 166, + 251, + 177 + ], + "type": "inline_equation", + "content": "\\theta" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 177, + 108, + 187 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 177, + 108, + 187 + ], + "spans": [ + { + "bbox": [ + 48, + 177, + 108, + 187 + ], + "type": "text", + "content": "while train do" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 63, + 188, + 286, + 223 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 64, + 188, + 285, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 188, + 285, + 198 + ], + "spans": [ + { + "bbox": [ + 64, + 188, + 285, + 198 + ], + "type": "text", + "content": "1. Input training data to FFS model and compute " + }, + { + "bbox": [ + 64, + 188, + 285, + 198 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{det}}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 63, + 199, + 286, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 199, + 286, + 223 + ], + "spans": [ + { + "bbox": [ + 63, + 199, + 286, + 223 + ], + "type": "text", + "content": "2. Input the ID features " + }, + { + "bbox": [ + 63, + 199, + 286, + 223 + ], + "type": "inline_equation", + "content": "l_{ID}(x,\\hat{\\pmb{b}})" + }, + { + "bbox": [ + 63, + 199, + 286, + 223 + ], + "type": "text", + "content": " to the flow model " + }, + { + "bbox": [ + 63, + 199, + 286, + 223 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 63, + 199, + 286, + 223 + ], + "type": "text", + "content": " and obtain " + }, + { + "bbox": [ + 63, + 199, + 286, + 223 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{nll}" + }, + { + "bbox": [ + 63, + 199, + 286, + 223 + ], + "type": "text", + "content": " using Eq 2." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 63, + 224, + 155, + 234 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 224, + 155, + 234 + ], + "spans": [ + { + "bbox": [ + 63, + 224, + 155, + 234 + ], + "type": "text", + "content": "if " + }, + { + "bbox": [ + 63, + 224, + 155, + 234 + ], + "type": "inline_equation", + "content": "i \\leq" + }, + { + "bbox": [ + 63, + 224, + 155, + 234 + ], + "type": "text", + "content": " iterations then" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 78, + 235, + 286, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 235, + 286, + 257 + ], + "spans": [ + { + "bbox": [ + 78, + 235, + 286, + 257 + ], + "type": "text", + "content": "3. Update " + }, + { + "bbox": [ + 78, + 235, + 286, + 257 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 78, + 235, + 286, + 257 + ], + "type": "text", + "content": " and the parameters " + }, + { + "bbox": [ + 78, + 235, + 286, + 257 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 78, + 235, + 286, + 257 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 78, + 235, + 286, + 257 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 78, + 235, + 286, + 257 + ], + "type": "text", + "content": " by backpropagation of the total loss in Eq 5 with " + }, + { + "bbox": [ + 78, + 235, + 286, + 257 + ], + "type": "inline_equation", + "content": "\\alpha = 0" + }, + { + "bbox": [ + 78, + 235, + 286, + 257 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 63, + 258, + 81, + 267 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 258, + 81, + 267 + ], + "spans": [ + { + "bbox": [ + 63, + 258, + 81, + 267 + ], + "type": "text", + "content": "else" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 78, + 268, + 287, + 365 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 78, + 268, + 286, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 268, + 286, + 292 + ], + "spans": [ + { + "bbox": [ + 78, + 268, + 286, + 292 + ], + "type": "text", + "content": "4. Generate synthetic features " + }, + { + "bbox": [ + 78, + 268, + 286, + 292 + ], + "type": "inline_equation", + "content": "g_{k}" + }, + { + "bbox": [ + 78, + 268, + 286, + 292 + ], + "type": "text", + "content": " by sampling " + }, + { + "bbox": [ + 78, + 268, + 286, + 292 + ], + "type": "inline_equation", + "content": "z_{k}" + }, + { + "bbox": [ + 78, + 268, + 286, + 292 + ], + "type": "text", + "content": " from flow's latent space with " + }, + { + "bbox": [ + 78, + 268, + 286, + 292 + ], + "type": "inline_equation", + "content": "g_{k} = f^{-1}(z_{k})" + }, + { + "bbox": [ + 78, + 268, + 286, + 292 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 78, + 292, + 276, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 292, + 276, + 304 + ], + "spans": [ + { + "bbox": [ + 78, + 292, + 276, + 304 + ], + "type": "text", + "content": "5. Select " + }, + { + "bbox": [ + 78, + 292, + 276, + 304 + ], + "type": "inline_equation", + "content": "o_s" + }, + { + "bbox": [ + 78, + 292, + 276, + 304 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 78, + 292, + 276, + 304 + ], + "type": "inline_equation", + "content": "p_{\\gamma}(g_k)" + }, + { + "bbox": [ + 78, + 292, + 276, + 304 + ], + "type": "text", + "content": "'s low-likelihood region." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 78, + 304, + 287, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 304, + 287, + 328 + ], + "spans": [ + { + "bbox": [ + 78, + 304, + 287, + 328 + ], + "type": "text", + "content": "6. Calculate energy scores " + }, + { + "bbox": [ + 78, + 304, + 287, + 328 + ], + "type": "inline_equation", + "content": "E(h(l_{ID}(x,\\hat{\\boldsymbol{b}});\\theta))" + }, + { + "bbox": [ + 78, + 304, + 287, + 328 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 78, + 304, + 287, + 328 + ], + "type": "inline_equation", + "content": "E(h(o_s;\\gamma))" + }, + { + "bbox": [ + 78, + 304, + 287, + 328 + ], + "type": "text", + "content": " using the classification head " + }, + { + "bbox": [ + 78, + 304, + 287, + 328 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 78, + 304, + 287, + 328 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 78, + 329, + 272, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 329, + 272, + 341 + ], + "spans": [ + { + "bbox": [ + 78, + 329, + 272, + 341 + ], + "type": "text", + "content": "7. Compute regularization loss " + }, + { + "bbox": [ + 78, + 329, + 272, + 341 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{reg}" + }, + { + "bbox": [ + 78, + 329, + 272, + 341 + ], + "type": "text", + "content": " using Eq 4." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 78, + 341, + 286, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 341, + 286, + 365 + ], + "spans": [ + { + "bbox": [ + 78, + 341, + 286, + 365 + ], + "type": "text", + "content": "8. Update " + }, + { + "bbox": [ + 78, + 341, + 286, + 365 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 78, + 341, + 286, + 365 + ], + "type": "text", + "content": " and the parameters " + }, + { + "bbox": [ + 78, + 341, + 286, + 365 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 78, + 341, + 286, + 365 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 78, + 341, + 286, + 365 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 78, + 341, + 286, + 365 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 78, + 341, + 286, + 365 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 78, + 341, + 286, + 365 + ], + "type": "text", + "content": " by backpropagation of the total loss in Eq 5." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 63, + 365, + 81, + 374 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 365, + 81, + 374 + ], + "spans": [ + { + "bbox": [ + 63, + 365, + 81, + 374 + ], + "type": "text", + "content": "end" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 48, + 378, + 65, + 387 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 378, + 65, + 387 + ], + "spans": [ + { + "bbox": [ + 48, + 378, + 65, + 387 + ], + "type": "text", + "content": "end" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 47, + 389, + 284, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 389, + 284, + 400 + ], + "spans": [ + { + "bbox": [ + 47, + 389, + 284, + 400 + ], + "type": "text", + "content": "Fix threshold " + }, + { + "bbox": [ + 47, + 389, + 284, + 400 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 47, + 389, + 284, + 400 + ], + "type": "text", + "content": " when " + }, + { + "bbox": [ + 47, + 389, + 284, + 400 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 47, + 389, + 284, + 400 + ], + "type": "text", + "content": " of inliers are correctly detected." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 48, + 401, + 105, + 411 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 401, + 105, + 411 + ], + "spans": [ + { + "bbox": [ + 48, + 401, + 105, + 411 + ], + "type": "text", + "content": "while eval do" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 64, + 412, + 250, + 435 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 64, + 412, + 244, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 412, + 244, + 423 + ], + "spans": [ + { + "bbox": [ + 64, + 412, + 244, + 423 + ], + "type": "text", + "content": "1. Calculate the energy score of a test object." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 64, + 424, + 250, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 424, + 250, + 435 + ], + "spans": [ + { + "bbox": [ + 64, + 424, + 250, + 435 + ], + "type": "text", + "content": "2. Label it as an outlier or an inlier based on " + }, + { + "bbox": [ + 64, + 424, + 250, + 435 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 64, + 424, + 250, + 435 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 48, + 436, + 66, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 436, + 66, + 445 + ], + "spans": [ + { + "bbox": [ + 48, + 436, + 66, + 445 + ], + "type": "text", + "content": "end" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 47, + 467, + 251, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 467, + 251, + 479 + ], + "spans": [ + { + "bbox": [ + 47, + 467, + 251, + 479 + ], + "type": "text", + "content": "3.2. Normalizing Flow for feature synthesis" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 46, + 485, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 485, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 485, + 287, + 665 + ], + "type": "text", + "content": "For simplicity, we will denote the ID features " + }, + { + "bbox": [ + 46, + 485, + 287, + 665 + ], + "type": "inline_equation", + "content": "l_{ID}(x,\\hat{\\pmb{b}})" + }, + { + "bbox": [ + 46, + 485, + 287, + 665 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 46, + 485, + 287, + 665 + ], + "type": "inline_equation", + "content": "l\\in \\mathcal{L}" + }, + { + "bbox": [ + 46, + 485, + 287, + 665 + ], + "type": "text", + "content": ". In our framework, the normalizing flow " + }, + { + "bbox": [ + 46, + 485, + 287, + 665 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 485, + 287, + 665 + ], + "type": "text", + "content": " with parameters " + }, + { + "bbox": [ + 46, + 485, + 287, + 665 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 485, + 287, + 665 + ], + "type": "text", + "content": " is a sequence of " + }, + { + "bbox": [ + 46, + 485, + 287, + 665 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 485, + 287, + 665 + ], + "type": "text", + "content": " invertible bijective mappings implemented as affine coupling layers. The flow " + }, + { + "bbox": [ + 46, + 485, + 287, + 665 + ], + "type": "inline_equation", + "content": "f:\\mathcal{L}\\rightarrow \\mathcal{Z}" + }, + { + "bbox": [ + 46, + 485, + 287, + 665 + ], + "type": "text", + "content": " transforms the complex data distribution of the features " + }, + { + "bbox": [ + 46, + 485, + 287, + 665 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 46, + 485, + 287, + 665 + ], + "type": "text", + "content": " to a multivariate Gaussian in its latent space " + }, + { + "bbox": [ + 46, + 485, + 287, + 665 + ], + "type": "inline_equation", + "content": "z\\in \\mathcal{Z}" + }, + { + "bbox": [ + 46, + 485, + 287, + 665 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 485, + 287, + 665 + ], + "type": "inline_equation", + "content": "p(z) = \\mathcal{N}_d(\\pmb {\\mu},\\pmb {\\Sigma})" + }, + { + "bbox": [ + 46, + 485, + 287, + 665 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 46, + 485, + 287, + 665 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 485, + 287, + 665 + ], + "type": "text", + "content": " is the number of feature vector components, and " + }, + { + "bbox": [ + 46, + 485, + 287, + 665 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}" + }, + { + "bbox": [ + 46, + 485, + 287, + 665 + ], + "type": "text", + "content": " is zero mean with " + }, + { + "bbox": [ + 46, + 485, + 287, + 665 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}" + }, + { + "bbox": [ + 46, + 485, + 287, + 665 + ], + "type": "text", + "content": " as the unit variance. The bijective mapping ensures that the input and the latent space share the same dimensionality " + }, + { + "bbox": [ + 46, + 485, + 287, + 665 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 485, + 287, + 665 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 46, + 485, + 287, + 665 + ], + "type": "inline_equation", + "content": "f:\\mathbb{R}^d\\to \\mathbb{R}^d" + }, + { + "bbox": [ + 46, + 485, + 287, + 665 + ], + "type": "text", + "content": ". Each coupling layer transforms the feature vector components by scaling " + }, + { + "bbox": [ + 46, + 485, + 287, + 665 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 485, + 287, + 665 + ], + "type": "text", + "content": " and translation " + }, + { + "bbox": [ + 46, + 485, + 287, + 665 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 485, + 287, + 665 + ], + "type": "text", + "content": ", which are learnable neural networks [27]. Such a transformation is expressive and, at the same time, easily invertible with high efficiency in computing Jacobian determinants." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 47, + 670, + 253, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 253, + 681 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 253, + 681 + ], + "type": "text", + "content": "3.2.1 Maximizing the likelihood of ID features" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "Given the above formulation of our normalizing flow " + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": ", which transforms input ID features " + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "inline_equation", + "content": "z =" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 304, + 72, + 545, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 171 + ], + "type": "inline_equation", + "content": "f(l)" + }, + { + "bbox": [ + 304, + 72, + 545, + 171 + ], + "type": "text", + "content": ", we aim to maximize the log-likelihood of recovering features " + }, + { + "bbox": [ + 304, + 72, + 545, + 171 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 72, + 545, + 171 + ], + "type": "text", + "content": " with respect to " + }, + { + "bbox": [ + 304, + 72, + 545, + 171 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 72, + 545, + 171 + ], + "type": "text", + "content": ". To achieve this, we have to compute the Jacobian matrix of " + }, + { + "bbox": [ + 304, + 72, + 545, + 171 + ], + "type": "inline_equation", + "content": "f(l)" + }, + { + "bbox": [ + 304, + 72, + 545, + 171 + ], + "type": "text", + "content": " with respect to the feature vector components [27]. Let " + }, + { + "bbox": [ + 304, + 72, + 545, + 171 + ], + "type": "inline_equation", + "content": "f_{i}(l)" + }, + { + "bbox": [ + 304, + 72, + 545, + 171 + ], + "type": "text", + "content": " be the components of " + }, + { + "bbox": [ + 304, + 72, + 545, + 171 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 72, + 545, + 171 + ], + "type": "text", + "content": " in latent space and " + }, + { + "bbox": [ + 304, + 72, + 545, + 171 + ], + "type": "inline_equation", + "content": "l_{j}" + }, + { + "bbox": [ + 304, + 72, + 545, + 171 + ], + "type": "text", + "content": " the components of feature space. The entries of the Jacobian matrix are defined as " + }, + { + "bbox": [ + 304, + 72, + 545, + 171 + ], + "type": "inline_equation", + "content": "J_{ij}^{f,l} = \\frac{\\partial f_i(l)}{\\partial l_j}" + }, + { + "bbox": [ + 304, + 72, + 545, + 171 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 72, + 545, + 171 + ], + "type": "inline_equation", + "content": "i,j\\in 1,\\dots,d" + }, + { + "bbox": [ + 304, + 72, + 545, + 171 + ], + "type": "text", + "content": ". According to the change of variables formula, the posterior likelihood " + }, + { + "bbox": [ + 304, + 72, + 545, + 171 + ], + "type": "inline_equation", + "content": "p_{\\gamma}(l)" + }, + { + "bbox": [ + 304, + 72, + 545, + 171 + ], + "type": "text", + "content": " can be described as:" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 367, + 182, + 545, + 198 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 182, + 545, + 198 + ], + "spans": [ + { + "bbox": [ + 367, + 182, + 545, + 198 + ], + "type": "interline_equation", + "content": "p _ {\\gamma} (l) = p (f (l)) * \\left| \\det J ^ {f, l} \\right| \\tag {1}", + "image_path": "ebebb94f001d663edce6be2f5742f5a20765eff669b349fdae2decb0d2ad2fc5.jpg" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 304, + 201, + 545, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 201, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 304, + 201, + 545, + 262 + ], + "type": "text", + "content": "To obtain the log-likelihood of the posteriors, we take a logarithm on both sides of the Eq 1. As the training of a neural network requires minimization of a loss function, we construct " + }, + { + "bbox": [ + 304, + 201, + 545, + 262 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 201, + 545, + 262 + ], + "type": "text", + "content": " by finding " + }, + { + "bbox": [ + 304, + 201, + 545, + 262 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 201, + 545, + 262 + ], + "type": "text", + "content": " that maximizes " + }, + { + "bbox": [ + 304, + 201, + 545, + 262 + ], + "type": "inline_equation", + "content": "\\log(p_{\\gamma}(l))" + }, + { + "bbox": [ + 304, + 201, + 545, + 262 + ], + "type": "text", + "content": " by minimizing the negative log-likelihood " + }, + { + "bbox": [ + 304, + 201, + 545, + 262 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{nll}" + }, + { + "bbox": [ + 304, + 201, + 545, + 262 + ], + "type": "text", + "content": " objective:" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 337, + 266, + 545, + 297 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 266, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 337, + 266, + 545, + 297 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {n l l} (l; \\gamma) = \\frac {1}{N} \\sum_ {i = 1} ^ {N} - \\log \\left(p _ {\\gamma} \\left(l \\left(x _ {i}, b \\left(x _ {i}\\right)\\right)\\right)\\right) \\tag {2}", + "image_path": "fab83ab12c4ca0e0745fa873a8097f4b3aa9ebd35b683a25e0c7feb55b66f4ae.jpg" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 305, + 307, + 545, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 307, + 545, + 331 + ], + "spans": [ + { + "bbox": [ + 305, + 307, + 545, + 331 + ], + "type": "text", + "content": "Normalizing flows are known for stable and exact training according to the maximum likelihood objective in Eq 2." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 306, + 336, + 509, + 348 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 336, + 509, + 348 + ], + "spans": [ + { + "bbox": [ + 306, + 336, + 509, + 348 + ], + "type": "text", + "content": "3.2.2 Random sampling from the latent space" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 304, + 354, + 545, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 354, + 545, + 582 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 545, + 582 + ], + "type": "text", + "content": "We train the flow model " + }, + { + "bbox": [ + 304, + 354, + 545, + 582 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 354, + 545, + 582 + ], + "type": "text", + "content": " so that its parameters " + }, + { + "bbox": [ + 304, + 354, + 545, + 582 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 354, + 545, + 582 + ], + "type": "text", + "content": " learn the distribution of inlier features for a fixed number of training iterations. Next, we randomly sample " + }, + { + "bbox": [ + 304, + 354, + 545, + 582 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 354, + 545, + 582 + ], + "type": "text", + "content": " samples " + }, + { + "bbox": [ + 304, + 354, + 545, + 582 + ], + "type": "inline_equation", + "content": "z_{k}" + }, + { + "bbox": [ + 304, + 354, + 545, + 582 + ], + "type": "text", + "content": " from the latent distribution " + }, + { + "bbox": [ + 304, + 354, + 545, + 582 + ], + "type": "inline_equation", + "content": "p(z)" + }, + { + "bbox": [ + 304, + 354, + 545, + 582 + ], + "type": "text", + "content": " and propagate it in the reverse direction of " + }, + { + "bbox": [ + 304, + 354, + 545, + 582 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 354, + 545, + 582 + ], + "type": "text", + "content": " to generate synthetic samples " + }, + { + "bbox": [ + 304, + 354, + 545, + 582 + ], + "type": "inline_equation", + "content": "g_{k}" + }, + { + "bbox": [ + 304, + 354, + 545, + 582 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 354, + 545, + 582 + ], + "type": "inline_equation", + "content": "g_{k} = f^{-1}(z_{k})" + }, + { + "bbox": [ + 304, + 354, + 545, + 582 + ], + "type": "text", + "content": " during active training. As the flow model learned the data distribution " + }, + { + "bbox": [ + 304, + 354, + 545, + 582 + ], + "type": "inline_equation", + "content": "p_{\\gamma}(l)" + }, + { + "bbox": [ + 304, + 354, + 545, + 582 + ], + "type": "text", + "content": " via the maximum likelihood training using Eq 2, the samples " + }, + { + "bbox": [ + 304, + 354, + 545, + 582 + ], + "type": "inline_equation", + "content": "g_{k}" + }, + { + "bbox": [ + 304, + 354, + 545, + 582 + ], + "type": "text", + "content": " are synthesized from " + }, + { + "bbox": [ + 304, + 354, + 545, + 582 + ], + "type": "inline_equation", + "content": "p_{\\gamma}(l)" + }, + { + "bbox": [ + 304, + 354, + 545, + 582 + ], + "type": "text", + "content": ". Such generation procedure discourages mode collapse, common in generative adversarial models. Hence, we retrieve the exact data distribution of the features " + }, + { + "bbox": [ + 304, + 354, + 545, + 582 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 354, + 545, + 582 + ], + "type": "text", + "content": " in synthesized samples " + }, + { + "bbox": [ + 304, + 354, + 545, + 582 + ], + "type": "inline_equation", + "content": "g_{k}" + }, + { + "bbox": [ + 304, + 354, + 545, + 582 + ], + "type": "text", + "content": ". After that, we compute the log-likelihood scores of each of the " + }, + { + "bbox": [ + 304, + 354, + 545, + 582 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 354, + 545, + 582 + ], + "type": "text", + "content": " samples " + }, + { + "bbox": [ + 304, + 354, + 545, + 582 + ], + "type": "inline_equation", + "content": "g_{k}" + }, + { + "bbox": [ + 304, + 354, + 545, + 582 + ], + "type": "text", + "content": " using Eq 1. Reasonably, the synthesized " + }, + { + "bbox": [ + 304, + 354, + 545, + 582 + ], + "type": "inline_equation", + "content": "g_{k}" + }, + { + "bbox": [ + 304, + 354, + 545, + 582 + ], + "type": "text", + "content": " contains samples with a high likelihood of being obtained from " + }, + { + "bbox": [ + 304, + 354, + 545, + 582 + ], + "type": "inline_equation", + "content": "p_{\\gamma}(l)" + }, + { + "bbox": [ + 304, + 354, + 545, + 582 + ], + "type": "text", + "content": " and should lie well inside the boundary of the inlier distribution " + }, + { + "bbox": [ + 304, + 354, + 545, + 582 + ], + "type": "inline_equation", + "content": "p_{\\gamma}(l)" + }, + { + "bbox": [ + 304, + 354, + 545, + 582 + ], + "type": "text", + "content": ". Nevertheless, there should also be samples in " + }, + { + "bbox": [ + 304, + 354, + 545, + 582 + ], + "type": "inline_equation", + "content": "g_{k}" + }, + { + "bbox": [ + 304, + 354, + 545, + 582 + ], + "type": "text", + "content": " with lower likelihood scores that lie near or away from the boundary of the distribution " + }, + { + "bbox": [ + 304, + 354, + 545, + 582 + ], + "type": "inline_equation", + "content": "p_{\\gamma}(l)" + }, + { + "bbox": [ + 304, + 354, + 545, + 582 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 306, + 586, + 525, + 599 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 586, + 525, + 599 + ], + "spans": [ + { + "bbox": [ + 306, + 586, + 525, + 599 + ], + "type": "text", + "content": "3.2.3 Rejection Sampling based Outlier Synthesis" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "It is pivotal to sample useful synthetic outliers for an effective model regularization for OD detection. To achieve this, our first approach involves rejection sampling. We select " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " samples as outliers " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "o_{s}" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " from the low-likelihood region of the data distribution " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "p_{\\gamma}(g_k)" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "o_{s} \\subset g_{k}" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": " being much smaller than " + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": ". The motivation of such an approach is to obtain synthetic outliers from near the decision boundary and outwards in the exterior outlier space. Figure 3 (a)-(b) shows that generating more synthetic samples" + } + ] + } + ], + "index": 37 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5159" + } + ] + } + ], + "index": 38 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "inline_equation", + "content": "g_{k}" + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "text", + "content": " and selecting a single outlier " + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "inline_equation", + "content": "o_s" + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "text", + "content": " with the least likelihood (where " + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "inline_equation", + "content": "s = 1" + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "text", + "content": ") improves the OD detection. This is because generating more samples " + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "inline_equation", + "content": "g_{k}" + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "text", + "content": " leads to the selection of a low-likelihood synthetic outlier " + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "inline_equation", + "content": "o_s" + }, + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "text", + "content": " that genuinely lies in the outlier space, resulting in better model regularization." + } + ] + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 51, + 136, + 157, + 213 + ], + "blocks": [ + { + "bbox": [ + 51, + 136, + 157, + 213 + ], + "lines": [ + { + "bbox": [ + 51, + 136, + 157, + 213 + ], + "spans": [ + { + "bbox": [ + 51, + 136, + 157, + 213 + ], + "type": "image", + "image_path": "cd0c84ee971e659ec4a64f8c3e2b52b5cec94e57f931906525543f335e555683.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 167, + 136, + 276, + 213 + ], + "blocks": [ + { + "bbox": [ + 167, + 136, + 276, + 213 + ], + "lines": [ + { + "bbox": [ + 167, + 136, + 276, + 213 + ], + "spans": [ + { + "bbox": [ + 167, + 136, + 276, + 213 + ], + "type": "image", + "image_path": "77f6a32b446018a982fd302f9077d67388febd06e155f570eaf7a8e3b12d7673.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 51, + 215, + 159, + 291 + ], + "blocks": [ + { + "bbox": [ + 51, + 215, + 159, + 291 + ], + "lines": [ + { + "bbox": [ + 51, + 215, + 159, + 291 + ], + "spans": [ + { + "bbox": [ + 51, + 215, + 159, + 291 + ], + "type": "image", + "image_path": "80b60ce31524e8f895cbb7d6d1c025440d67b2dd6878cf13a20c8fd49dbb8135.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 299, + 286, + 321 + ], + "lines": [ + { + "bbox": [ + 47, + 299, + 286, + 321 + ], + "spans": [ + { + "bbox": [ + 47, + 299, + 286, + 321 + ], + "type": "text", + "content": "Figure 3. Validation of synthetic samples " + }, + { + "bbox": [ + 47, + 299, + 286, + 321 + ], + "type": "inline_equation", + "content": "g_{k}" + }, + { + "bbox": [ + 47, + 299, + 286, + 321 + ], + "type": "text", + "content": " and selected outliers " + }, + { + "bbox": [ + 47, + 299, + 286, + 321 + ], + "type": "inline_equation", + "content": "o_{s}" + }, + { + "bbox": [ + 47, + 299, + 286, + 321 + ], + "type": "text", + "content": " on OD detection. " + }, + { + "bbox": [ + 47, + 299, + 286, + 321 + ], + "type": "inline_equation", + "content": "s = 1" + }, + { + "bbox": [ + 47, + 299, + 286, + 321 + ], + "type": "text", + "content": " in (a)-(b) and " + }, + { + "bbox": [ + 47, + 299, + 286, + 321 + ], + "type": "inline_equation", + "content": "k = 200" + }, + { + "bbox": [ + 47, + 299, + 286, + 321 + ], + "type": "text", + "content": " in (c)-(d)." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 169, + 215, + 275, + 291 + ], + "blocks": [ + { + "bbox": [ + 169, + 215, + 275, + 291 + ], + "lines": [ + { + "bbox": [ + 169, + 215, + 275, + 291 + ], + "spans": [ + { + "bbox": [ + 169, + 215, + 275, + 291 + ], + "type": "image", + "image_path": "6bc2e3adb1301d5d70f5d624e9215af4b53bce074cc1da98c73e9c7d756ee278.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 330, + 287, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 330, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 46, + 330, + 287, + 498 + ], + "type": "text", + "content": "Eventually, however, the performance peaks as generating an even higher number of synthetic samples " + }, + { + "bbox": [ + 46, + 330, + 287, + 498 + ], + "type": "inline_equation", + "content": "g_{k}" + }, + { + "bbox": [ + 46, + 330, + 287, + 498 + ], + "type": "text", + "content": " may lead to the selection of an outlier " + }, + { + "bbox": [ + 46, + 330, + 287, + 498 + ], + "type": "inline_equation", + "content": "o_{s}" + }, + { + "bbox": [ + 46, + 330, + 287, + 498 + ], + "type": "text", + "content": " that is too far away in the outlier space, resulting in a model that is unaware of the decision boundary. Hence, by varying " + }, + { + "bbox": [ + 46, + 330, + 287, + 498 + ], + "type": "inline_equation", + "content": "g_{k}" + }, + { + "bbox": [ + 46, + 330, + 287, + 498 + ], + "type": "text", + "content": ", our rejection sampling approach can obtain outliers from the decision boundary and from the outlier region away from the boundary. Figure 3 (c)-(d) show the results when the number of selected outliers " + }, + { + "bbox": [ + 46, + 330, + 287, + 498 + ], + "type": "inline_equation", + "content": "o_{s}" + }, + { + "bbox": [ + 46, + 330, + 287, + 498 + ], + "type": "text", + "content": " is varied while keeping the number of the generated synthetic samples " + }, + { + "bbox": [ + 46, + 330, + 287, + 498 + ], + "type": "inline_equation", + "content": "g_{k}" + }, + { + "bbox": [ + 46, + 330, + 287, + 498 + ], + "type": "text", + "content": " fixed. The selection of more outliers " + }, + { + "bbox": [ + 46, + 330, + 287, + 498 + ], + "type": "inline_equation", + "content": "o_{s}" + }, + { + "bbox": [ + 46, + 330, + 287, + 498 + ], + "type": "text", + "content": " leads to a degradation in the OD detection performance as some selected outliers with high log-likelihood scores lie well within the inlier manifold leading to suboptimal model regularization." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 502, + 279, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 502, + 279, + 514 + ], + "spans": [ + { + "bbox": [ + 47, + 502, + 279, + 514 + ], + "type": "text", + "content": "3.2.4 Projection Sampling using Langevin Dynamics" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 521, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 287, + 713 + ], + "type": "text", + "content": "We investigate whether our model can generalize the outlier space by using synthetic outliers sampled only close to the decision boundary while still being effective with OD detection. To accomplish this, we perform projection sampling based on Stochastic Gradient Langevin Dynamics (SGLD) to sample outliers from our flow model. Under the approach, we generate " + }, + { + "bbox": [ + 46, + 521, + 287, + 713 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 521, + 287, + 713 + ], + "type": "text", + "content": " samples as outliers " + }, + { + "bbox": [ + 46, + 521, + 287, + 713 + ], + "type": "inline_equation", + "content": "o_s" + }, + { + "bbox": [ + 46, + 521, + 287, + 713 + ], + "type": "text", + "content": " directly without needing rejection and require a likelihood threshold to define the boundary of the inlier feature distribution. We fix this threshold as the log-likelihood score " + }, + { + "bbox": [ + 46, + 521, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 46, + 521, + 287, + 713 + ], + "type": "text", + "content": " of the inlier, which is least likely to be obtained from the inlier data distribution " + }, + { + "bbox": [ + 46, + 521, + 287, + 713 + ], + "type": "inline_equation", + "content": "p_{\\gamma}(l)" + }, + { + "bbox": [ + 46, + 521, + 287, + 713 + ], + "type": "text", + "content": ". Next, we propagate the average log-likelihood score of the synthetic outliers " + }, + { + "bbox": [ + 46, + 521, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\log p_{\\gamma}(o_s)" + }, + { + "bbox": [ + 46, + 521, + 287, + 713 + ], + "type": "text", + "content": " in the inverse direction of the flow without updating its parameters and obtain the gradients " + }, + { + "bbox": [ + 46, + 521, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\frac{\\partial \\log p_{\\gamma}(o_s)}{\\partial o_s}" + }, + { + "bbox": [ + 46, + 521, + 287, + 713 + ], + "type": "text", + "content": ". Subsequently, the outliers " + }, + { + "bbox": [ + 46, + 521, + 287, + 713 + ], + "type": "inline_equation", + "content": "o_s" + }, + { + "bbox": [ + 46, + 521, + 287, + 713 + ], + "type": "text", + "content": " are updated as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 372, + 69, + 545, + 95 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 69, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 372, + 69, + 545, + 95 + ], + "type": "interline_equation", + "content": "o _ {s} := o _ {s} - \\tau \\frac {\\partial \\log p _ {\\gamma} \\left(o _ {s}\\right)}{\\partial o _ {s}} \\tag {3}", + "image_path": "bfc3ed7ec79eb534c6703395813696fd279a093baa52b4579dd559a4ae38bb34.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 98, + 547, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 98, + 547, + 182 + ], + "spans": [ + { + "bbox": [ + 304, + 98, + 547, + 182 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 98, + 547, + 182 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 98, + 547, + 182 + ], + "type": "text", + "content": " is the step size of the gradient descent. We end this iterative process when the average log-likelihood of the updated " + }, + { + "bbox": [ + 304, + 98, + 547, + 182 + ], + "type": "inline_equation", + "content": "o_s" + }, + { + "bbox": [ + 304, + 98, + 547, + 182 + ], + "type": "text", + "content": " is equal to or lower than " + }, + { + "bbox": [ + 304, + 98, + 547, + 182 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 304, + 98, + 547, + 182 + ], + "type": "text", + "content": ". Hence, the updated outliers " + }, + { + "bbox": [ + 304, + 98, + 547, + 182 + ], + "type": "inline_equation", + "content": "o_s" + }, + { + "bbox": [ + 304, + 98, + 547, + 182 + ], + "type": "text", + "content": " are precisely projected to the near decision boundary of the inlier manifold where the least-likelihood inlier is located. The results of our projection sampling approach are shown in Figure 5 (f) and discussed in Section 4.3." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 187, + 493, + 200 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 187, + 493, + 200 + ], + "spans": [ + { + "bbox": [ + 305, + 187, + 493, + 200 + ], + "type": "text", + "content": "3.3. Energy-based model regularization" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 205, + 545, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 205, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 304, + 205, + 545, + 422 + ], + "type": "text", + "content": "We obtained the outliers " + }, + { + "bbox": [ + 304, + 205, + 545, + 422 + ], + "type": "inline_equation", + "content": "o_s" + }, + { + "bbox": [ + 304, + 205, + 545, + 422 + ], + "type": "text", + "content": " by sampling from the data distribution " + }, + { + "bbox": [ + 304, + 205, + 545, + 422 + ], + "type": "inline_equation", + "content": "p_{\\gamma}(l)" + }, + { + "bbox": [ + 304, + 205, + 545, + 422 + ], + "type": "text", + "content": ", requiring no actual outliers during training. The classification head " + }, + { + "bbox": [ + 304, + 205, + 545, + 422 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 304, + 205, + 545, + 422 + ], + "type": "text", + "content": " now maps inlier features " + }, + { + "bbox": [ + 304, + 205, + 545, + 422 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 304, + 205, + 545, + 422 + ], + "type": "text", + "content": " and outliers " + }, + { + "bbox": [ + 304, + 205, + 545, + 422 + ], + "type": "inline_equation", + "content": "o_s" + }, + { + "bbox": [ + 304, + 205, + 545, + 422 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 205, + 545, + 422 + ], + "type": "inline_equation", + "content": "K + 1" + }, + { + "bbox": [ + 304, + 205, + 545, + 422 + ], + "type": "text", + "content": " classification logits as " + }, + { + "bbox": [ + 304, + 205, + 545, + 422 + ], + "type": "inline_equation", + "content": "h(l;\\theta)" + }, + { + "bbox": [ + 304, + 205, + 545, + 422 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 205, + 545, + 422 + ], + "type": "inline_equation", + "content": "h(o_s;\\gamma)" + }, + { + "bbox": [ + 304, + 205, + 545, + 422 + ], + "type": "text", + "content": ", respectively. Note that " + }, + { + "bbox": [ + 304, + 205, + 545, + 422 + ], + "type": "inline_equation", + "content": "h(l;\\theta)" + }, + { + "bbox": [ + 304, + 205, + 545, + 422 + ], + "type": "text", + "content": " is influenced by all " + }, + { + "bbox": [ + 304, + 205, + 545, + 422 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 205, + 545, + 422 + ], + "type": "text", + "content": " parameters, whereas " + }, + { + "bbox": [ + 304, + 205, + 545, + 422 + ], + "type": "inline_equation", + "content": "h(o_s;\\gamma)" + }, + { + "bbox": [ + 304, + 205, + 545, + 422 + ], + "type": "text", + "content": " depends on the flow parameters " + }, + { + "bbox": [ + 304, + 205, + 545, + 422 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 205, + 545, + 422 + ], + "type": "text", + "content": " and the parameters of the classification head. Recently, [10, 35] used energy function rather than softmax probabilities to differentiate inlier and outlier samples better. We follow [10] and apply the energy function on the classification logits " + }, + { + "bbox": [ + 304, + 205, + 545, + 422 + ], + "type": "inline_equation", + "content": "h(l;\\theta)" + }, + { + "bbox": [ + 304, + 205, + 545, + 422 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 205, + 545, + 422 + ], + "type": "inline_equation", + "content": "h(o_s;\\gamma)" + }, + { + "bbox": [ + 304, + 205, + 545, + 422 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 304, + 205, + 545, + 422 + ], + "type": "inline_equation", + "content": "E(h(.)) = -T \\cdot \\log \\sum^{K} w_K \\cdot e^{h_K(.)/T}" + }, + { + "bbox": [ + 304, + 205, + 545, + 422 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 205, + 545, + 422 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 205, + 545, + 422 + ], + "type": "text", + "content": " is a temperature parameter of the energy function and weight " + }, + { + "bbox": [ + 304, + 205, + 545, + 422 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 304, + 205, + 545, + 422 + ], + "type": "text", + "content": " is learned to overcome class imbalances. We then pass the energy scores through a binary classifier " + }, + { + "bbox": [ + 304, + 205, + 545, + 422 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 304, + 205, + 545, + 422 + ], + "type": "text", + "content": " with parameters " + }, + { + "bbox": [ + 304, + 205, + 545, + 422 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 304, + 205, + 545, + 422 + ], + "type": "text", + "content": " and define the softmax output as " + }, + { + "bbox": [ + 304, + 205, + 545, + 422 + ], + "type": "inline_equation", + "content": "p_l" + }, + { + "bbox": [ + 304, + 205, + 545, + 422 + ], + "type": "text", + "content": " when " + }, + { + "bbox": [ + 304, + 205, + 545, + 422 + ], + "type": "inline_equation", + "content": "E(h(l;\\theta))" + }, + { + "bbox": [ + 304, + 205, + 545, + 422 + ], + "type": "text", + "content": " is the input and " + }, + { + "bbox": [ + 304, + 205, + 545, + 422 + ], + "type": "inline_equation", + "content": "p_o" + }, + { + "bbox": [ + 304, + 205, + 545, + 422 + ], + "type": "text", + "content": " when " + }, + { + "bbox": [ + 304, + 205, + 545, + 422 + ], + "type": "inline_equation", + "content": "E(h(o_s;\\gamma))" + }, + { + "bbox": [ + 304, + 205, + 545, + 422 + ], + "type": "text", + "content": " is the input. Finally, we use binary cross-entropy (BCE) based regularization loss " + }, + { + "bbox": [ + 304, + 205, + 545, + 422 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{reg}" + }, + { + "bbox": [ + 304, + 205, + 545, + 422 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 336, + 425, + 545, + 456 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 425, + 545, + 456 + ], + "spans": [ + { + "bbox": [ + 336, + 425, + 545, + 456 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {r e g} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} - (\\log (p _ {l}) + \\log (1 - p _ {o})) \\tag {4}", + "image_path": "a01b28c7e216429d1246b81d570bd7794f77bb78f8e153126b26eb8c7553ad29.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 459, + 545, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 459, + 545, + 494 + ], + "spans": [ + { + "bbox": [ + 304, + 459, + 545, + 494 + ], + "type": "text", + "content": "The Eq 4 shapes contrasting inlier and outlier energy surfaces by assigning low energy for inliers and high energy for outliers." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 500, + 455, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 500, + 455, + 513 + ], + "spans": [ + { + "bbox": [ + 306, + 500, + 455, + 513 + ], + "type": "text", + "content": "3.4. Overall Training Objective" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 518, + 545, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 518, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 304, + 518, + 545, + 602 + ], + "type": "text", + "content": "We describe the standard object detection loss " + }, + { + "bbox": [ + 304, + 518, + 545, + 602 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{det}}" + }, + { + "bbox": [ + 304, + 518, + 545, + 602 + ], + "type": "text", + "content": " as a combination of losses for object classification given the ground-truth inlier class labels " + }, + { + "bbox": [ + 304, + 518, + 545, + 602 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 518, + 545, + 602 + ], + "type": "text", + "content": " and bounding box regression given the ground-truth box coordinates " + }, + { + "bbox": [ + 304, + 518, + 545, + 602 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 304, + 518, + 545, + 602 + ], + "type": "text", + "content": ". Based on the formulation of the negative log-likelihood loss " + }, + { + "bbox": [ + 304, + 518, + 545, + 602 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{nll}}" + }, + { + "bbox": [ + 304, + 518, + 545, + 602 + ], + "type": "text", + "content": " and the regularization loss " + }, + { + "bbox": [ + 304, + 518, + 545, + 602 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{reg}}" + }, + { + "bbox": [ + 304, + 518, + 545, + 602 + ], + "type": "text", + "content": " from Eq 2 and 4 respectively, we merge all these loss terms in our overall training objective:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 315, + 607, + 545, + 625 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 607, + 545, + 625 + ], + "spans": [ + { + "bbox": [ + 315, + 607, + 545, + 625 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta , \\gamma , \\psi} \\mathbb {E} _ {(x, \\mathbf {b}, y) \\sim \\mathcal {X}, \\mathcal {B}, \\mathcal {Y}} \\left[ \\mathcal {L} _ {d e t} + \\beta \\cdot \\mathcal {L} _ {n l l} + \\alpha \\cdot \\mathcal {L} _ {r e g} \\right], \\tag {5}", + "image_path": "df67f9f5785188c485541faaad4a0d282b4242c8e9e314b9c8d05b38ffde39f4.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " are the weights of " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{reg}" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{nll}" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " respectively. The " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{reg}" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " loss acts as an adversary to " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{nll}" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " loss as " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{nll}" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " tries to maximize the likelihood of obtaining inliers features while " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{reg}" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " forces the model to synthesize outliers away from inliers. Thus, such a training scheme encourages the generation of synthetic samples at the decision boundary resulting in outlier awareness of the object detector." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5160" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 71, + 128, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 71, + 128, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 128, + 85 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 91, + 288, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 91, + 288, + 189 + ], + "spans": [ + { + "bbox": [ + 46, + 91, + 288, + 189 + ], + "type": "text", + "content": "This section evaluates our FFS framework for its effectiveness with outlier-aware object detection. Section 4.1 provides the details related to the image and video datasets, the neural network architecture, and the performance metrics. Section 4.2 shows the results obtained with FFS and compares it with other related approaches. Finally, Section 4.3 contains ablation studies to evaluate the performance of our approach in varied experimental settings." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 193, + 178, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 193, + 178, + 205 + ], + "spans": [ + { + "bbox": [ + 47, + 193, + 178, + 205 + ], + "type": "text", + "content": "4.1. Implementation details" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 211, + 287, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 211, + 287, + 354 + ], + "spans": [ + { + "bbox": [ + 46, + 211, + 287, + 354 + ], + "type": "text", + "content": "Datasets. We train our FFS model on the publicly available image dataset PASCAL-VOC 2012 [11] containing 20 object categories as ID and use two OD datasets, namely MS-COCO [32] and OpenImages [28] to evaluate the performance. Additionally, we train FFS on two video datasets, namely Berkeley DeepDrive (BDD100K) [52] and Youtube-Video Instance Segmentation (Youtube-VIS) 2021 [51]. For the FFS trained on video datasets, we infer on two OD datasets, namely MS-COCO [32] and nuImages [2]. We use the pre-processed datasets by VOS [10] and STUD [9] so that the object categories in the OD dataset are independent of the object categories in the ID dataset." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 354, + 287, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 354, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 287, + 437 + ], + "type": "text", + "content": "Network architecture. We adopt the Faster R-CNN model [42] for object detection provided in the publicly available Detectron2 framework [50]. We use RegNetX-4.0GF [40] as the backbone architecture due to its superior inlier detection performance. We employ Glow [27] as our normalizing flow model after observing lower performance for other flow architectures (details in Section 4.3)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 438, + 287, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 438, + 287, + 558 + ], + "spans": [ + { + "bbox": [ + 46, + 438, + 287, + 558 + ], + "type": "text", + "content": "Metaparameters. We run our experiments using Python 3.8.6 and PyTorch 1.9.0 with each ID dataset trained on four NVIDIA A100-SXM4 GPUs. We start the training of the FFS framework by turning off uncertainty regularization loss " + }, + { + "bbox": [ + 46, + 438, + 287, + 558 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{reg}" + }, + { + "bbox": [ + 46, + 438, + 287, + 558 + ], + "type": "text", + "content": " for a fixed number of iterations. This enables the flow model to learn the data distribution of inlier objects from all object classes. Across datasets, we found the loss weightages " + }, + { + "bbox": [ + 46, + 438, + 287, + 558 + ], + "type": "inline_equation", + "content": "\\beta = 10^{-4}" + }, + { + "bbox": [ + 46, + 438, + 287, + 558 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 438, + 287, + 558 + ], + "type": "inline_equation", + "content": "\\alpha = 0.1" + }, + { + "bbox": [ + 46, + 438, + 287, + 558 + ], + "type": "text", + "content": " as reasonable values for stable end-to-end training of the FFS framework. In Section 4.3, we perform an exhaustive meta-parameter study." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "content": "Evaluation metrics. We compute the energy scores from " + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 558, + 288, + 714 + ], + "type": "text", + "content": " classification logits of the inlier and outlier objects in the validation set. We evaluate these scores according to the standard outlier detection metrics. These metrics are Area Under the Curve Receiver Operating Characteristic (AUROC ↑) and False Positive Rate at 95% True Positive Rate (FPR95 ↓), where True Positive is the correct detection of an inlier. For ID detection, we first compute Intersection over Union (IoU) between predicted bounding boxes and ground-truth boxes. Then, we compute the average precision (AP) for the standard range of IoU thresholds. Finally, the mean of the average precision over all known classes (mAP ↑) measures ID detection performance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 72, + 505, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 505, + 85 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 505, + 85 + ], + "type": "text", + "content": "4.2. Comparison with the State-of-the-Art" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 90, + 547, + 211 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 90, + 547, + 211 + ], + "spans": [ + { + "bbox": [ + 304, + 90, + 547, + 211 + ], + "type": "text", + "content": "Tables 1 and 2 compare our method with other approaches on image and video datasets, respectively. We use the results provided by VOS [10] and STUD [9] to show the performance of the compared approaches. Some evaluated approaches (except VOS and STUD) were developed for image-wide OD detection. However, these approaches can also be used to compare an outlier-aware object detection model since classification is one of the object detection tasks. No real outlier datasets were used during training to compare these approaches with our method." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 307, + 220, + 547, + 304 + ], + "blocks": [ + { + "bbox": [ + 307, + 220, + 547, + 304 + ], + "lines": [ + { + "bbox": [ + 307, + 220, + 547, + 304 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 547, + 304 + ], + "type": "table", + "html": "
IDMethodFPR95 ↓AUROC ↑mAP ↑
PASCAL-VOCGAN [29]OD: MS-COCO / OpenImages48.5
Energy [35]60.93 / 59.9783.67 / 82.67
CSI [47]56.89 / 58.6983.69 / 82.9848.7
VOS [10]59.91 / 57.4181.83 / 82.9548.1
FFS (ours)47.77 / 48.3389.00 / 87.5951.5
44.15 / 45.0889.71 / 88.2951.8
", + "image_path": "abaa89594740f5d812711cc9a94b3bf85a43904f97daa8b9e8451148d697b428.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 307, + 345, + 548, + 473 + ], + "blocks": [ + { + "bbox": [ + 305, + 308, + 545, + 331 + ], + "lines": [ + { + "bbox": [ + 305, + 308, + 545, + 331 + ], + "spans": [ + { + "bbox": [ + 305, + 308, + 545, + 331 + ], + "type": "text", + "content": "Table 1. Main results on image datasets with Faster R-CNN model and RegNetX-4.0GF [40] as the backbone for VOS and FFS." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 345, + 548, + 473 + ], + "lines": [ + { + "bbox": [ + 307, + 345, + 548, + 473 + ], + "spans": [ + { + "bbox": [ + 307, + 345, + 548, + 473 + ], + "type": "table", + "html": "
IDODMethodFPR95 ↓AUROC ↑mAP ↑
BDD100KmulImagesGAN [29]83.6570.3931.5
Energy [35]81.6269.4332.0
CSI [47]80.0074.9131.8
STUD [9]79.7576.5532.3
FFS (ours)76.6877.5336.2
Youtube-VISMS-COCOGAN [29]85.7572.9525.5
Energy [35]88.5467.8326.7
CSI [47]82.4371.8124.2
STUD [9]81.1474.8227.2
FFS (ours)83.0676.3727.6
", + "image_path": "1b70802e88dc2e6adfb1bec3b18e240a2465b02d7bfc32ec95bd8fc514fe799b.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 478, + 545, + 501 + ], + "lines": [ + { + "bbox": [ + 305, + 478, + 545, + 501 + ], + "spans": [ + { + "bbox": [ + 305, + 478, + 545, + 501 + ], + "type": "text", + "content": "Table 2. Main results on video datasets with Faster R-CNN model and RegNetX-4.0GF [40] as the backbone network." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 510, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 547, + 715 + ], + "type": "text", + "content": "In Table 1, we report a " + }, + { + "bbox": [ + 304, + 510, + 547, + 715 + ], + "type": "inline_equation", + "content": "7.58\\%" + }, + { + "bbox": [ + 304, + 510, + 547, + 715 + ], + "type": "text", + "content": " and a " + }, + { + "bbox": [ + 304, + 510, + 547, + 715 + ], + "type": "inline_equation", + "content": "6.73\\%" + }, + { + "bbox": [ + 304, + 510, + 547, + 715 + ], + "type": "text", + "content": " decrease in FPR95 for FFS compared to VOS when evaluated on MS-COCO and OpenImages, respectively. Additionally, we note an increase in AUROC values when compared with VOS and assessed on MS-COCO and OpenImages. We also obtained higher ID detection performance and a faster training time of 2.18 hr compared to 2.43 hr for VOS on the same hardware and batch size. In Table 2, FFS outperforms STUD, where we report a " + }, + { + "bbox": [ + 304, + 510, + 547, + 715 + ], + "type": "inline_equation", + "content": "3.85\\%" + }, + { + "bbox": [ + 304, + 510, + 547, + 715 + ], + "type": "text", + "content": " decrease in FPR95 and a " + }, + { + "bbox": [ + 304, + 510, + 547, + 715 + ], + "type": "inline_equation", + "content": "1.28\\%" + }, + { + "bbox": [ + 304, + 510, + 547, + 715 + ], + "type": "text", + "content": " increase in AUROC compared to STUD when the ID is BDD100K, and the OD is nuImages. We also register a significant " + }, + { + "bbox": [ + 304, + 510, + 547, + 715 + ], + "type": "inline_equation", + "content": "12\\%" + }, + { + "bbox": [ + 304, + 510, + 547, + 715 + ], + "type": "text", + "content": " increase in mAP performance compared to STUD. Our training time is 2.7 hrs compared to 11 hrs by STUD for the BDD100K dataset on the same hardware. FFS also improves STUD results for the Youtube-VIS as ID and MS-COCO as OD while maintaining high ID detection performance. For this dataset, we report a lower training" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "5161" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 239, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 239, + 84 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 239, + 84 + ], + "type": "text", + "content": "time of 8.2 hrs compared to 11.3 hrs by STUD." + } + ] + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 48, + 95, + 288, + 147 + ], + "blocks": [ + { + "bbox": [ + 48, + 95, + 288, + 147 + ], + "lines": [ + { + "bbox": [ + 48, + 95, + 288, + 147 + ], + "spans": [ + { + "bbox": [ + 48, + 95, + 288, + 147 + ], + "type": "table", + "html": "
IDMethodFPR95 ↓AUROC ↑mAP ↑Time (h)
PASCAL-VOCVOS [10]49.6788.4348.912.37
FFS (ours)43.1289.8451.732.12
Youtube-VISSTUD [9]83.9374.5424.4211.03
FFS (ours)81.9076.3525.108.23
", + "image_path": "6d4e7bac6b03af2fe61bfc93a46f91b4b6312e340119b2e2b9a863f1a59678b1.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 171, + 287, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 171, + 287, + 278 + ], + "spans": [ + { + "bbox": [ + 46, + 171, + 287, + 278 + ], + "type": "text", + "content": "In Table 3, we also show the performance while using ResNet-50 [18] as an alternative backbone and compare our results with VOS [10] and STUD [9] with MS-COCO as the outlier dataset. The results show that using ResNet-50 as the backbone achieves state-of-the-art ID and OD detection performance with our framework. Overall, Tables 1, 2 and 3 demonstrate the effectiveness of FFS for OD detection for images and videos with lower training time and high ID detection performance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 285, + 147, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 285, + 147, + 298 + ], + "spans": [ + { + "bbox": [ + 47, + 285, + 147, + 298 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 304, + 287, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 304, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 46, + 304, + 287, + 376 + ], + "type": "text", + "content": "In this section, we perform an exhaustive study of meta-parameters that affects the performance of FFS. We also qualitatively compare OD detection results with the VOS approach. For uniformity, we fix the ID dataset as PASCAL-VOC and OD dataset as MS-COCO unless stated otherwise in the experiments." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 376, + 288, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 376, + 288, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 288, + 533 + ], + "type": "text", + "content": "VOS plus rejection sampling: Sampling outliers from a class conditional Gaussian (as performed in VOS [10]) does not ensure that such an outlier will have lower likelihoods for other class Gaussians. Therefore, we compute the log-likelihoods of the outliers using all class Gaussians and reject those with a higher likelihood from at least one other class compared to its generator class. Table 4 shows the results for VOS [10], VOS plus rejection sampling (named as " + }, + { + "bbox": [ + 46, + 376, + 288, + 533 + ], + "type": "inline_equation", + "content": "\\mathrm{VOS + }" + }, + { + "bbox": [ + 46, + 376, + 288, + 533 + ], + "type": "text", + "content": "), and FFS. " + }, + { + "bbox": [ + 46, + 376, + 288, + 533 + ], + "type": "inline_equation", + "content": "\\mathrm{VOS + }" + }, + { + "bbox": [ + 46, + 376, + 288, + 533 + ], + "type": "text", + "content": " enhances the OD detection compared to VOS [10] due to selecting more effective outliers, but it is still worse than FFS. Furthermore, " + }, + { + "bbox": [ + 46, + 376, + 288, + 533 + ], + "type": "inline_equation", + "content": "\\mathrm{VOS + }" + }, + { + "bbox": [ + 46, + 376, + 288, + 533 + ], + "type": "text", + "content": "'s training time is much higher due to the iterative process of computing log-likelihoods through each Gaussian class." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 48, + 541, + 285, + 596 + ], + "blocks": [ + { + "bbox": [ + 58, + 150, + 275, + 160 + ], + "lines": [ + { + "bbox": [ + 58, + 150, + 275, + 160 + ], + "spans": [ + { + "bbox": [ + 58, + 150, + 275, + 160 + ], + "type": "text", + "content": "Table 3. Performance evaluation with ResNet-50 backbone." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 541, + 285, + 596 + ], + "lines": [ + { + "bbox": [ + 48, + 541, + 285, + 596 + ], + "spans": [ + { + "bbox": [ + 48, + 541, + 285, + 596 + ], + "type": "table", + "html": "
MethodFPR95 ↓AUROC ↑mAP (ID)↑Time (h)
VOS [10]OD: MS-COCO / OpenImages
48.22 / 52.7489.04 / 85.9751.502.43
VOS+45.59 / 47.9689.40 / 87.1951.747.38
FFS (ours)44.15 / 45.0889.71 / 88.2951.802.18
", + "image_path": "b26c8cb9721ed49cfe65ed064c19f4b5141c940f7a7846b820afc9305fa9ff59.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "type": "text", + "content": "Normalizing Flow: We conducted a thorough study on several types of normalizing flow models. We selected NICE [6], RealNVP [7], Glow [27] and GIN [45] in the chronological order for this experiment. We fixed the same coupling layers and each model's sub-network " + }, + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "type": "text", + "content": " configuration for an equivalent comparison. Table 5 shows that Glow outperforms NICE, RealNVP and GIN in terms of" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 545, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 121 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 121 + ], + "type": "text", + "content": "both FPR95 and AUROC scores. GIN performs best for ID detection but has a worse OD detection performance. As the primary task of FFS is OD detection, we chose Glow as our flow model for all our experiments." + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 309, + 124, + 544, + 183 + ], + "blocks": [ + { + "bbox": [ + 46, + 597, + 287, + 620 + ], + "lines": [ + { + "bbox": [ + 46, + 597, + 287, + 620 + ], + "spans": [ + { + "bbox": [ + 46, + 597, + 287, + 620 + ], + "type": "text", + "content": "Table 4. Comparison of VOS, VOS+, and FFS (ours). The baseline mAP of a closed-set object detector is 51.35." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 124, + 544, + 183 + ], + "lines": [ + { + "bbox": [ + 309, + 124, + 544, + 183 + ], + "spans": [ + { + "bbox": [ + 309, + 124, + 544, + 183 + ], + "type": "table", + "html": "
Flow modelFPR95 ↓AUROC ↑mAP ↑Time (h)
NICE [6]48.2388.4651.812.23
RealNVP [7]45.7688.9551.842.28
Glow [27]44.1589.7151.802.18
GIN [45]48.3188.3652.002.25
", + "image_path": "459b76aaaec69a7f824a133033a72da8dbf432637c4bbb696e479545ecae411d.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 201, + 545, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 201, + 545, + 418 + ], + "spans": [ + { + "bbox": [ + 304, + 201, + 545, + 418 + ], + "type": "text", + "content": "Figure 5 (c) shows the effect of the number of coupling layers " + }, + { + "bbox": [ + 304, + 201, + 545, + 418 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 201, + 545, + 418 + ], + "type": "text", + "content": " on OD detection performance. " + }, + { + "bbox": [ + 304, + 201, + 545, + 418 + ], + "type": "inline_equation", + "content": "M = 2" + }, + { + "bbox": [ + 304, + 201, + 545, + 418 + ], + "type": "text", + "content": " produces the best results as the inlier features fit a suitable number of flow parameters. However, the OD detection performance degrades for " + }, + { + "bbox": [ + 304, + 201, + 545, + 418 + ], + "type": "inline_equation", + "content": "M > 2" + }, + { + "bbox": [ + 304, + 201, + 545, + 418 + ], + "type": "text", + "content": ", and the training time significantly increases due to more trainable parameters in the network. Adding more trainable parameters requires more training data for effective training; otherwise, it might lead to overfitting and a more complex optimization, which should explain the degradation in performance. Hence, we fix " + }, + { + "bbox": [ + 304, + 201, + 545, + 418 + ], + "type": "inline_equation", + "content": "M = 2" + }, + { + "bbox": [ + 304, + 201, + 545, + 418 + ], + "type": "text", + "content": ", and in Figure 5 (a), we study the effect of the number of fully-connected (fc) layers within " + }, + { + "bbox": [ + 304, + 201, + 545, + 418 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 201, + 545, + 418 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 201, + 545, + 418 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 201, + 545, + 418 + ], + "type": "text", + "content": " sub-networks of each coupling layer. There is a slight improvement in OD detection performance when two fc layers are employed. In Figure 5 (b), we fix the number of coupling and fc layers in each " + }, + { + "bbox": [ + 304, + 201, + 545, + 418 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 201, + 545, + 418 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 201, + 545, + 418 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 201, + 545, + 418 + ], + "type": "text", + "content": " sub-networks as two and adjust the number of neurons in each fc layer. We report a degradation in OD detection performance above 2048 neurons." + } + ] + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 309, + 426, + 541, + 488 + ], + "blocks": [ + { + "bbox": [ + 328, + 186, + 520, + 198 + ], + "lines": [ + { + "bbox": [ + 328, + 186, + 520, + 198 + ], + "spans": [ + { + "bbox": [ + 328, + 186, + 520, + 198 + ], + "type": "text", + "content": "Table 5. Validation of the Normalizing Flow models." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 426, + 541, + 488 + ], + "lines": [ + { + "bbox": [ + 309, + 426, + 541, + 488 + ], + "spans": [ + { + "bbox": [ + 309, + 426, + 541, + 488 + ], + "type": "table", + "html": "
LregFPR95 ↓AUROC ↑mAP ↑Time (h)
CE69.9284.7248.702.27
JSD52.8385.7948.602.25
Hinge51.3286.3446.542.19
BCE (Ours)44.1589.7151.802.18
", + "image_path": "3663be6682f4ead503d2a31c6047e3b08b9739479e7a2434d2cbb2d523e2d058.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 318, + 491, + 531, + 502 + ], + "lines": [ + { + "bbox": [ + 318, + 491, + 531, + 502 + ], + "spans": [ + { + "bbox": [ + 318, + 491, + 531, + 502 + ], + "type": "text", + "content": "Table 6. Validation of " + }, + { + "bbox": [ + 318, + 491, + 531, + 502 + ], + "type": "inline_equation", + "content": "{\\mathcal{L}}_{reg}" + }, + { + "bbox": [ + 318, + 491, + 531, + 502 + ], + "type": "text", + "content": " on OD detection performance." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": "Regularization loss " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{reg}" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": " and its weightage " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": ": We studied several regularization losses for our FFS framework. Firstly, we used a simple multi-class cross-entropy loss (CE) on the classification logits without including the energy module. Secondly, we evaluated Jensen-Shannon divergence (JSD) loss, where we minimized the divergence between the distribution of the classification logits of the outlier samples and uniform distribution. In Table 6, we show that our BCE based " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{reg}" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": " loss outperforms all other loss functions in terms of OD detection performance. The ID detection performance is also higher, even though the training time is comparable. In Figure 5 (e), we demonstrate the effect of changing the weightage " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{reg}" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": " for PASCAL-VOC as ID and MS-COCO as OD. As " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": " increases, the OD detection performance improves. However, an even higher weightage degrades the OD detection performance, due to which a careful selection of " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": " is desirable." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5162" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 71, + 544, + 205 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 544, + 205 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 544, + 205 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 544, + 205 + ], + "type": "image", + "image_path": "b29c94f55e9d1a0791c4f3f6784b4b58951ea2015622fd90ba38e61586956595.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 207, + 546, + 231 + ], + "lines": [ + { + "bbox": [ + 47, + 207, + 546, + 231 + ], + "spans": [ + { + "bbox": [ + 47, + 207, + 546, + 231 + ], + "type": "text", + "content": "Figure 4. Visualization of OD detection on MS-COCO and OpenImages with VOS [10] (first and third row) and FFS (ours) (second and fourth row) trained on PASCAL-VOC 2012 images. Ideally, an object in the images should be labeled as OOD with a green bounding box." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 50, + 236, + 167, + 311 + ], + "blocks": [ + { + "bbox": [ + 50, + 236, + 167, + 311 + ], + "lines": [ + { + "bbox": [ + 50, + 236, + 167, + 311 + ], + "spans": [ + { + "bbox": [ + 50, + 236, + 167, + 311 + ], + "type": "image", + "image_path": "fb1fb7cd1ddcb6f09dee8ec5eac055b7d67f56de64e70e28037eb0264901d6a7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 168, + 235, + 287, + 311 + ], + "blocks": [ + { + "bbox": [ + 168, + 235, + 287, + 311 + ], + "lines": [ + { + "bbox": [ + 168, + 235, + 287, + 311 + ], + "spans": [ + { + "bbox": [ + 168, + 235, + 287, + 311 + ], + "type": "image", + "image_path": "6527735aa9c0794b0a5f91177d5feb75e5df3bdefb56803e7e7fba5358d26c39.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 47, + 313, + 167, + 388 + ], + "blocks": [ + { + "bbox": [ + 47, + 313, + 167, + 388 + ], + "lines": [ + { + "bbox": [ + 47, + 313, + 167, + 388 + ], + "spans": [ + { + "bbox": [ + 47, + 313, + 167, + 388 + ], + "type": "image", + "image_path": "510ccb5007d373515b3648b0987f1c7f9a886af0cc282fd332c105efadf2e80a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 168, + 312, + 285, + 388 + ], + "blocks": [ + { + "bbox": [ + 168, + 312, + 285, + 388 + ], + "lines": [ + { + "bbox": [ + 168, + 312, + 285, + 388 + ], + "spans": [ + { + "bbox": [ + 168, + 312, + 285, + 388 + ], + "type": "image", + "image_path": "1816e992538b3f66e6881e34f3210b99f35528804e6c67e2c0c4b004e2a0355d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 50, + 391, + 165, + 473 + ], + "blocks": [ + { + "bbox": [ + 50, + 391, + 165, + 473 + ], + "lines": [ + { + "bbox": [ + 50, + 391, + 165, + 473 + ], + "spans": [ + { + "bbox": [ + 50, + 391, + 165, + 473 + ], + "type": "image", + "image_path": "35e77aa27926f14bb570c3f23553174584f682c6fe4628826bc80358077d63b9.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 482, + 287, + 516 + ], + "lines": [ + { + "bbox": [ + 46, + 482, + 287, + 516 + ], + "spans": [ + { + "bbox": [ + 46, + 482, + 287, + 516 + ], + "type": "text", + "content": "Figure 5. Validation of the flow architecture, loss weightage " + }, + { + "bbox": [ + 46, + 482, + 287, + 516 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 482, + 287, + 516 + ], + "type": "text", + "content": ", batch size, and projection sampling. The training time (in hours) is shown above the horizontal x axis in the plots." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 170, + 391, + 284, + 474 + ], + "blocks": [ + { + "bbox": [ + 170, + 391, + 284, + 474 + ], + "lines": [ + { + "bbox": [ + 170, + 391, + 284, + 474 + ], + "spans": [ + { + "bbox": [ + 170, + 391, + 284, + 474 + ], + "type": "image", + "image_path": "3dd833c7dfb3e5ad291d8ce94de0846ede80cd48a41060d729233d9440fa644b.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 533, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 289, + 715 + ], + "type": "text", + "content": "Projection Sampling: Figure 5 (f) shows the results after varying synthetic outliers. The OD detection improves as we increase the number of sampled outliers " + }, + { + "bbox": [ + 46, + 533, + 289, + 715 + ], + "type": "inline_equation", + "content": "o_{s}" + }, + { + "bbox": [ + 46, + 533, + 289, + 715 + ], + "type": "text", + "content": " but degrades with much higher values of " + }, + { + "bbox": [ + 46, + 533, + 289, + 715 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 533, + 289, + 715 + ], + "type": "text", + "content": ". The coverage of log-likelihoods from many sampled outliers broadens even though the average log-likelihood is similar to the likelihood threshold " + }, + { + "bbox": [ + 46, + 533, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 46, + 533, + 289, + 715 + ], + "type": "text", + "content": ". Hence, some outliers may lie inside, while others exist far outside the decision boundary leading to bad regularization. We report that rejection sampling achieves better OD detection performance when compared to the best results of projection sampling. Since projection sampling enforces the outliers to be close to the decision boundary, it may be insufficient to estimate the entire outlier manifold. In contrast, the rejection sampling synthesize outliers near the decision boundary and in the proper" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 238, + 537, + 250 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 238, + 537, + 250 + ], + "spans": [ + { + "bbox": [ + 306, + 238, + 537, + 250 + ], + "type": "text", + "content": "outlier manifold, leading to a better model regularization." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 251, + 545, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 251, + 545, + 322 + ], + "spans": [ + { + "bbox": [ + 304, + 251, + 545, + 322 + ], + "type": "text", + "content": "Qualitative analysis: Figure 4 shows the OD detection results on MS-COCO and OpenImages with FFS and VOS trained on PASCAL-VOC. The results show that FFS outperforms VOS in recognizing outlier objects. Additionally, our method's softmax confidence score is lower than VOS when both methods misclassify the outlier objects as inliers." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 329, + 378, + 340 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 329, + 378, + 340 + ], + "spans": [ + { + "bbox": [ + 306, + 329, + 378, + 340 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 349, + 546, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 349, + 546, + 564 + ], + "spans": [ + { + "bbox": [ + 304, + 349, + 546, + 564 + ], + "type": "text", + "content": "We presented a new methodology for outlier-aware object detection that learns a combined data distribution of all inlier object classes. By sampling from the low-likelihood region of a jointly trained normalizing flow model, our approach generates suitable synthetic outlier samples for training the outlier detection head of our compound model. In contrast, previous approaches model inliers with class-conditional Gaussians, resulting in outlier samples that may have a high likelihood for some other class. We report state-of-the-art results for image and video datasets with standard outlier detection metrics while decreasing the model training time. We also show that simple rejection sampling contributes more useful synthetic outliers than projection sampling with a fixed likelihood threshold. In future work, one could validate this finding on other OD detectors. Our idea also extends to developing an outlier-aware instance segmentation model and an active learning scheme to correct the failure modes in weakly-supervised object detectors." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 714 + ], + "type": "text", + "content": "Acknowledgement: This work primarily received funding from the German Federal Ministry of Education and Research (BMBF) under Software Campus (grant 01IS17044) and was supported by the Center for Scalable Data Analytics and Artificial Intelligence (ScaDS.AI) Dresden/Leipzig, Germany. The work was also partially funded by DFG grant 389792660 as part of TRR 248 - CPEC, Croatian Science Foundation (grant IP-2020-02-5851 ADEPT) and the Cluster of Excellence CeTI (EXC2050/1, grant 390696704). The authors gratefully acknowledge the Center for Information Services and HPC (ZIH) at TU Dresden for providing computing resources." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5163" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 286, + 712 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 53, + 91, + 286, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 286, + 124 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 286, + 124 + ], + "type": "text", + "content": "[1] Petra Bevandic, Ivan Kreso, Marin Orsic, and Sinisa Segvic. Dense open-set recognition based on training with noisy negative images. Image Vis. Comput., 2022. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 125, + 286, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 286, + 168 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 286, + 168 + ], + "type": "text", + "content": "[2] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuScenes: A multimodal dataset for autonomous driving. In CVPR, 2020. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 171, + 286, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 171, + 286, + 213 + ], + "spans": [ + { + "bbox": [ + 53, + 171, + 286, + 213 + ], + "type": "text", + "content": "[3] Jiwoong Choi, Dayoung Chun, Hyun Kim, and Hyuk-Jae Lee. Gaussian YOLOv3: An accurate and fast object detector using localization uncertainty for autonomous driving. In ICCV, 2019. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 215, + 286, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 215, + 286, + 247 + ], + "spans": [ + { + "bbox": [ + 53, + 215, + 286, + 247 + ], + "type": "text", + "content": "[4] Kumari Deepshikha, Sai Harsha Yelleni, P. K. Srijith, and C. Krishna Mohan. Monte Carlo dropblock for modelling uncertainty in object detection. CoRR, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 250, + 286, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 250, + 286, + 281 + ], + "spans": [ + { + "bbox": [ + 53, + 250, + 286, + 281 + ], + "type": "text", + "content": "[5] Akshay Raj Dhamija, Manuel Günther, Jonathan Ventura, and Terrance E. Boult. The overlooked elephant of object detection: Open Set. In WACV, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 283, + 286, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 283, + 286, + 314 + ], + "spans": [ + { + "bbox": [ + 53, + 283, + 286, + 314 + ], + "type": "text", + "content": "[6] Laurent Dinh, David Krueger, and Yoshua Bengio. NICE: Non-linear Independent components estimation. In ICLR, 2015. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 317, + 286, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 317, + 286, + 338 + ], + "spans": [ + { + "bbox": [ + 53, + 317, + 286, + 338 + ], + "type": "text", + "content": "[7] Laurent Dinh, Jascha Sohl-Dickstein, and Samy Bengio. Density estimation using Real NVP. In ICLR, 2017. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 341, + 286, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 341, + 286, + 371 + ], + "spans": [ + { + "bbox": [ + 53, + 341, + 286, + 371 + ], + "type": "text", + "content": "[8] Keval Doshi and Yasin Yilmaz. Any-shot sequential anomaly detection in surveillance videos. In CVPR Workshops, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 374, + 286, + 417 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 374, + 286, + 417 + ], + "spans": [ + { + "bbox": [ + 53, + 374, + 286, + 417 + ], + "type": "text", + "content": "[9] Xuefeng Du, Xin Wang, Gabriel Gozum, and Yixuan Li. Unknown-aware Object Detection: Learning what you don't know from videos in the wild. In CVPR, 2022. 1, 2, 3, 6, 7, 12, 13, 14" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 419, + 286, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 419, + 286, + 452 + ], + "spans": [ + { + "bbox": [ + 48, + 419, + 286, + 452 + ], + "type": "text", + "content": "[10] Xuefeng Du, Zhaoning Wang, Mu Cai, and Yixuan Li. VOS: Learning what you don't know by virtual outlier synthesis. In ICLR, 2022. 1, 2, 5, 6, 7, 8, 11, 12" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 453, + 286, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 453, + 286, + 496 + ], + "spans": [ + { + "bbox": [ + 48, + 453, + 286, + 496 + ], + "type": "text", + "content": "[11] Mark Everingham, Luc Van Gool, Christopher K. I. Williams, John M. Winn, and Andrew Zisserman. The Pascal visual object classes (VOC) challenge. International Journal of Computer Vision, 2010. 1, 2, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 498, + 286, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 498, + 286, + 530 + ], + "spans": [ + { + "bbox": [ + 48, + 498, + 286, + 530 + ], + "type": "text", + "content": "[12] Yarin Gal and Zoubin Ghahramani. Dropout as a Bayesian Approximation: Representing model uncertainty in deep learning. In ICML, 2016. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 533, + 239, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 533, + 239, + 542 + ], + "spans": [ + { + "bbox": [ + 48, + 533, + 239, + 542 + ], + "type": "text", + "content": "[13] Ross Girshick. Fast R-CNN. In ICCV, 2015. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 544, + 286, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 286, + 576 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 286, + 576 + ], + "type": "text", + "content": "[14] Matej Grcic, Petra Bevandic, and Sinisa Segvic. Dense open-set recognition with synthetic outliers generated by Real NVP. CoRR, 2020. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 578, + 286, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 286, + 621 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 286, + 621 + ], + "type": "text", + "content": "[15] David Hall, Feras Dayoub, John Skinner, Haoyang Zhang, Dimity Miller, Peter Corke, Gustavo Carneiro, Anelia Angelova, and Niko Sunderhauf. Probabilistic Object Detection: Definition and Evaluation. In WACV, 2020. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 624, + 286, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 286, + 655 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 286, + 655 + ], + "type": "text", + "content": "[16] Ali Harakeh and Steven L. Waslander. Estimating and evaluating regression predictive uncertainty in deep object detectors. In ICLR, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 658, + 286, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 286, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 286, + 678 + ], + "type": "text", + "content": "[17] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask R-CNN. In ICCV, 2017. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 48, + 681, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 286, + 712 + ], + "type": "text", + "content": "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 7" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 37, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[19] Dan Hendrycks and Kevin Gimpel. A baseline for detecting misclassified and out-of-distribution examples in neural networks. In ICLR, 2017. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 108, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 545, + 139 + ], + "type": "text", + "content": "[20] Dan Hendrycks, Mantas Mazeika, and Thomas Dietterich. Deep anomaly detection with outlier exposure. In ICLR, 2019. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 142, + 545, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 142, + 545, + 174 + ], + "spans": [ + { + "bbox": [ + 307, + 142, + 545, + 174 + ], + "type": "text", + "content": "[21] Dan Hendrycks, Mantas Mazeika, Saurav Kadavath, and Dawn Song. Using self-supervised learning can improve model robustness and uncertainty. In NeurIPS, 2019. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 177, + 545, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 177, + 545, + 220 + ], + "spans": [ + { + "bbox": [ + 307, + 177, + 545, + 220 + ], + "type": "text", + "content": "[22] Dan Hendrycks, Norman Mu, Ekin D. Cubuk, Barret Zoph, Justin Gilmer, and Balaji Lakshminarayanan. AugMix: A simple data processing method to improve robustness and uncertainty. In ICLR, 2020. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 222, + 545, + 265 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 222, + 545, + 265 + ], + "spans": [ + { + "bbox": [ + 307, + 222, + 545, + 265 + ], + "type": "text", + "content": "[23] Yen-Chang Hsu, Yilin Shen, Hongxia Jin, and Zsolt Kira. Generalized ODIN: Detecting out-of-distribution image without learning from out-of-distribution data. In CVPR, 2020. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 267, + 545, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 267, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 307, + 267, + 545, + 300 + ], + "type": "text", + "content": "[24] Rui Huang, Andrew Geng, and Yixuan Li. On the importance of gradients for detecting distributional shifts in the wild. In NeurIPS, 2021. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 302, + 545, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 302, + 545, + 345 + ], + "spans": [ + { + "bbox": [ + 307, + 302, + 545, + 345 + ], + "type": "text", + "content": "[25] Radu Tudor Ionescu, Fahad Shahbaz Khan, Mariana-Iuliana Georgescu, and Ling Shao. Object-centric auto-encoders and dummy anomalies for abnormal event detection in video. In CVPR, 2019. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 347, + 545, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 347, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 307, + 347, + 545, + 380 + ], + "type": "text", + "content": "[26] K. J. Joseph, Salman Khan, Fahad Shahbaz Khan, and Vineeth N. Balasubramanian. Towards open world object detection. In CVPR, 2021. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 382, + 545, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 382, + 545, + 414 + ], + "spans": [ + { + "bbox": [ + 307, + 382, + 545, + 414 + ], + "type": "text", + "content": "[27] Durk P. Kingma and Prafulla Dhariwal. Glow: Generative flow with invertible 1x1 convolutions. In NeurIPS, 2018. 2, 4, 6, 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 416, + 545, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 416, + 545, + 471 + ], + "spans": [ + { + "bbox": [ + 307, + 416, + 545, + 471 + ], + "type": "text", + "content": "[28] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The Open Images dataset v4. International Journal of Computer Vision, 2020. 6" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 473, + 545, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 473, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 307, + 473, + 545, + 506 + ], + "type": "text", + "content": "[29] Kimin Lee, Honglak Lee, Kibok Lee, and Jinwoo Shin. Training confidence-calibrated classifiers for detecting out-of-distribution samples. In ICLR, 2018. 1, 2, 6" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 507, + 545, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 507, + 545, + 540 + ], + "spans": [ + { + "bbox": [ + 307, + 507, + 545, + 540 + ], + "type": "text", + "content": "[30] Kimin Lee, Kibok Lee, Honglak Lee, and Jinwoo Shin. A simple unified framework for detecting out-of-distribution samples and adversarial attacks. In NeurIPS, 2018. 1, 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 543, + 545, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 543, + 545, + 574 + ], + "spans": [ + { + "bbox": [ + 307, + 543, + 545, + 574 + ], + "type": "text", + "content": "[31] Shiyu Liang, Yixuan Li, and Rayadurgam Srikant. Enhancing the reliability of out-of-distribution image detection in neural networks. In ICLR, 2018. 1, 2" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 576, + 545, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 576, + 545, + 620 + ], + "spans": [ + { + "bbox": [ + 307, + 576, + 545, + 620 + ], + "type": "text", + "content": "[32] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, Lubomir D. Bourdev, Ross B. Girshick, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollar, and C. Lawrence Zitnick. Microsoft COCO: Common objects in context. In ECCV, 2014. 6" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 307, + 622, + 545, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 622, + 545, + 643 + ], + "spans": [ + { + "bbox": [ + 307, + 622, + 545, + 643 + ], + "type": "text", + "content": "[33] Ziqian Lin, Sreya D. Roy, and Yixuan Li. MOOD: Multi-level out-of-distribution detection. In CVPR, 2021. 2" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 307, + 646, + 545, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 646, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 307, + 646, + 545, + 677 + ], + "type": "text", + "content": "[34] Wen Liu, Weixin Luo, Dongze Lian, and Shenghua Gao. Future frame prediction for anomaly detection - A new baseline. In CVPR, 2018. 3" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "type": "text", + "content": "[35] Weitang Liu, Xiaoyun Wang, John Owens, and Yixuan Li. Energy-based out-of-distribution detection. In NeurIPS, 2020. 1, 2, 5, 6" + } + ] + } + ], + "index": 36 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5164" + } + ] + } + ], + "index": 38 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 710 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "type": "text", + "content": "[36] Thomas Lucas, Konstantin Shmelkov, Karteek Alahari, Cordelia Schmid, and Jakob Verbeek. Adaptive density estimation for generative models. In NeurIPS, 2019. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 286, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 286, + 148 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 286, + 148 + ], + "type": "text", + "content": "[37] Dimity Miller, Feras Dayoub, Michael Milford, and Niko Sünderhauf. Evaluating merging strategies for sampling-based uncertainty techniques in object detection. In ICRA, 2019. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 149, + 286, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 149, + 286, + 182 + ], + "spans": [ + { + "bbox": [ + 48, + 149, + 286, + 182 + ], + "type": "text", + "content": "[38] Dimity Miller, Lachlan Nicholson, Feras Dayoub, and Niko Sünderhauf. Dropout sampling for robust object detection in open-set conditions. In ICRA, 2018. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 183, + 286, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 183, + 286, + 215 + ], + "spans": [ + { + "bbox": [ + 48, + 183, + 286, + 215 + ], + "type": "text", + "content": "[39] Sina Mohseni, Mandar Pitale, JBS Yadawa, and Zhangyang Wang. Self-supervised learning for generalizable out-of-distribution detection. In AAAI, 2020. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 216, + 286, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 216, + 286, + 248 + ], + "spans": [ + { + "bbox": [ + 48, + 216, + 286, + 248 + ], + "type": "text", + "content": "[40] Ilija Radosavovic, Raj Prateek Kosaraju, Ross B. Girshick, Kaiming He, and Piotr Dólár. Designing network design spaces. In CVPR, 2020. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 249, + 286, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 249, + 286, + 292 + ], + "spans": [ + { + "bbox": [ + 48, + 249, + 286, + 292 + ], + "type": "text", + "content": "[41] Mahdyar Ravanbakhsh, Moin Nabi, Hossein Mousavi, Enver Sangineto, and Nicu Sebe. Plug-and-play CNN for crowd motion analysis: An application in abnormal event detection. In WACV, 2018. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 293, + 286, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 293, + 286, + 326 + ], + "spans": [ + { + "bbox": [ + 48, + 293, + 286, + 326 + ], + "type": "text", + "content": "[42] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster R-CNN: Towards real-time object detection with region proposal networks. In NeurIPS, 2015. 1, 2, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 326, + 286, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 326, + 286, + 367 + ], + "spans": [ + { + "bbox": [ + 48, + 326, + 286, + 367 + ], + "type": "text", + "content": "[43] Tobias Riedlinger, Matthias Rottmann, Marius Schubert, and Hanno Gottschalk. Gradient-based quantification of epistemic uncertainty for deep object detectors. CoRR, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 369, + 286, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 369, + 286, + 402 + ], + "spans": [ + { + "bbox": [ + 48, + 369, + 286, + 402 + ], + "type": "text", + "content": "[44] Chandramouli Shama Sastry and Sageev Oore. Detecting out-of-distribution examples with Gram Matrices. In ICML, 2020. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 403, + 286, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 403, + 286, + 435 + ], + "spans": [ + { + "bbox": [ + 48, + 403, + 286, + 435 + ], + "type": "text", + "content": "[45] Peter Sorrenson, Carsten Rother, and Ullrich Kothe. Disentanglement by nonlinear ICA with general incompressible-flow networks (GIN). In *ICLR*, 2020. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 435, + 286, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 435, + 286, + 467 + ], + "spans": [ + { + "bbox": [ + 48, + 435, + 286, + 467 + ], + "type": "text", + "content": "[46] Yiyou Sun, Chuan Guo, and Yixuan Li. ReAct: Out-of-distribution detection with rectified activations. In NeurIPS, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 468, + 286, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 468, + 286, + 501 + ], + "spans": [ + { + "bbox": [ + 48, + 468, + 286, + 501 + ], + "type": "text", + "content": "[47] Jihoon Tack, Sangwoo Mo, Jongheon Jeong, and Jinwoo Shin. CSI: Novelty detection via contrastive learning on distributionally shifted instances. In NeurIPS, 2020. 2, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 501, + 286, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 501, + 286, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 501, + 286, + 533 + ], + "type": "text", + "content": "[48] Haoran Wang, Weitang Liu, Alex Bocchieri, and Yixuan Li. Can multi-label classification networks know what they don't know? In NeurIPS, 2021. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 534, + 286, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 534, + 286, + 577 + ], + "spans": [ + { + "bbox": [ + 48, + 534, + 286, + 577 + ], + "type": "text", + "content": "[49] Xin Wang, Thomas E. Huang, Benlin Liu, Fisher Yu, Xiaolong Wang, Joseph E. Gonzalez, and Trevor Darrell. Robust object detection via instance-level temporal cycle confusion. In ICCV, 2021. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 578, + 286, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 578, + 286, + 611 + ], + "spans": [ + { + "bbox": [ + 48, + 578, + 286, + 611 + ], + "type": "text", + "content": "[50] Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. Detector2. https://github.com/facebookresearch/detectron2, 2019.6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 611, + 286, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 611, + 286, + 632 + ], + "spans": [ + { + "bbox": [ + 48, + 611, + 286, + 632 + ], + "type": "text", + "content": "[51] Linjie Yang, Yuchen Fan, and Ning Xu. Video instance segmentation. In ICCV, 2019. 2, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 633, + 286, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 633, + 286, + 677 + ], + "spans": [ + { + "bbox": [ + 48, + 633, + 286, + 677 + ], + "type": "text", + "content": "[52] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darrell. BDD100K: A diverse driving dataset for heterogeneous multitask learning. In CVPR, 2020. 2, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 677, + 286, + 710 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 677, + 286, + 710 + ], + "spans": [ + { + "bbox": [ + 48, + 677, + 286, + 710 + ], + "type": "text", + "content": "[53] Jingyang Zhang, Nathan Inkawich, Yiran Chen, and Hai Li. Fine-grained out-of-distribution detection with mixup outlier exposure. CoRR, 2021. 1" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5165" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Not All Image Regions Matter_ Masked Vector Quantization for Autoregressive Image Generation/2a88178e-8ce2-4c0f-b3f1-bb99f0452012_content_list.json b/2023/Not All Image Regions Matter_ Masked Vector Quantization for Autoregressive Image Generation/2a88178e-8ce2-4c0f-b3f1-bb99f0452012_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1613e82c9a9615689a1db4d04e8cd5527c609b1b --- /dev/null +++ b/2023/Not All Image Regions Matter_ Masked Vector Quantization for Autoregressive Image Generation/2a88178e-8ce2-4c0f-b3f1-bb99f0452012_content_list.json @@ -0,0 +1,1685 @@ +[ + { + "type": "text", + "text": "Not All Image Regions Matter: Masked Vector Quantization for Autoregressive Image Generation", + "text_level": 1, + "bbox": [ + 148, + 130, + 818, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mengqi Huang $^{1}$ , Zhendong Mao $^{1,3*}$ , Quan Wang $^{2}$ , Yongdong Zhang $^{1,3}$", + "bbox": [ + 223, + 202, + 779, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1University of Science and Technology of China, Hefei, China; 2MOE Key Laboratory of Trustworthy, Distributed Computing and Service, Beijing University of Posts and Telecommunications, Beijing, China; 3Institute of Artificial intelligence, Hefei Comprehensive National Science Center, Hefei, China", + "bbox": [ + 84, + 220, + 919, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "huangmq@mail.ustc.edu.cn, {zdmao, zhyd73}@ustc.edu.cn, wangquan@bupt.edu.cn", + "bbox": [ + 168, + 277, + 831, + 291 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 325, + 313, + 342 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing autoregressive models follow the two-stage generation paradigm that first learns a codebook in the latent space for image reconstruction and then completes the image generation autoregressively based on the learned codebook. However, existing codebook learning simply models all local region information of images without distinguishing their different perceptual importance, which brings redundancy in the learned codebook that not only limits the next stage's autoregressive model's ability to model important structure but also results in high training cost and slow generation speed. In this study, we borrow the idea of importance perception from classical image coding theory and propose a novel two-stage framework, which consists of Masked Quantization VAE (MQ-VAE) and Stackformer, to relieve the model from modeling redundancy. Specifically, MQ-VAE incorporates an adaptive mask module for masking redundant region features before quantization and an adaptive de-mask module for recovering the original grid image feature map to faithfully reconstruct the original images after quantization. Then, Stackformer learns to predict the combination of the next code and its position in the feature map. Comprehensive experiments on various image generation validate our effectiveness and efficiency. Code will be released at https://github.com/CrossmodalGroup/MaskedVectorQuantization.", + "bbox": [ + 75, + 358, + 473, + 750 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 777, + 209, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep generative models of images have received significant improvements over the past few years and broadly fall into two categories: likelihood-based models, which include VAEs [24], flow-based [36], diffusion models [17] and autoregressive models [40], and generative adversarial", + "bbox": [ + 75, + 803, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/99f379c90a21357ab413c94465c158b648298bb77dcaa5cd4d6f9204332a8f4a.jpg", + "image_caption": [ + "(a) Existing methods", + "Figure 1. Illustration of our motivation. (a) Existing works model all local regions without distinguishing their perceptual importance in stage 1, which not only brings redundancy (e.g., the textural regions like the background) in the learned codebook but also make the autoregressive models overly focus on modeling this redundancy and hinder other important structural regions modeling. (b) The codebook learning in our method only includes the important regions, e.g., the structural regions like corners and edges, since other unimportant ones can be restored even if missing, and thus autoregressive model could focus on modeling these important regions in stage 2 and results in better generation quality." + ], + "image_footnote": [], + "bbox": [ + 501, + 327, + 696, + 476 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8136d80f5730ddfd445c48546d49198624ae79e616abb9f2cf724b2afc2ec8d4.jpg", + "image_caption": [ + "(b) Our method" + ], + "image_footnote": [], + "bbox": [ + 697, + 327, + 890, + 476 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "networks (GANs) [14], which use discriminator networks to distinguish samples from generator networks and real examples. Compared with GANs, likelihood-based models' training objective, i.e., the negative log-likelihood (NLL) or its upper bound, incentives learning the full data distribution and allows for detecting overfitting.", + "bbox": [ + 496, + 686, + 893, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Among the likelihood-based models, autoregressive models have recently attracted increasing attention for their impressive modeling ability and scalability. Recent autoregressive image generation [10, 12, 13, 28, 28, 34, 35, 37, 39] follows the two-stage generation paradigm, i.e., the first stage learns a codebook in the latent space for image reconstruction and the second stage completes the image generation in the raster-scan [13] order by autoregressive models", + "bbox": [ + 496, + 779, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Zhendong Mao is the corresponding author.", + "bbox": [ + 94, + 886, + 334, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "2002", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "based on the learned codebook. Since codebook learning in the first stage defines the discrete image representation for the next autoregressive modeling, a high-quality codebook is the key to generate high-quality images. Several recent works focus on improving the codebook learning in the first stage, e.g., VQGAN [13] introduces adversarial loss and perceptual loss. ViT-VQGAN [42] introduces a more expressive transformer backbone. RQ-VAE [28] introduces the residual quantization to reduce the resolution of the latent space. In general, the essence of existing codebook learning is the modeling of all local region information (i.e., an $8 \\times 8$ or $16 \\times 16$ patch) of images in the dataset, without distinguishing their different perceptual importance.", + "bbox": [ + 75, + 90, + 472, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this study, we point out that existing codebook learning exists gaps with classical image coding theory [20, 25, 26], the basic idea of which is to remove redundant information by perceiving the importance of different regions in images. The image coding theory reveals that an ideal image coding method should only encode images' perceptually important regions (i.e., which cannot be restored if missing) while discarding the unimportant ones (i.e., which can be restored by other image regions even if missing). The neglect of considering such perceptual importance in existing works poses problems in two aspects, as illustrated in Figure 1(a): (1) the existence of this large amount of repetitive and redundant information brings redundancy to the learned codebook, which further makes the autoregressive model in the next stage overly focus on modeling this redundancy while overlooking other important regions and finally degrades generation quality. (2) the redundancy makes the autoregressive model need to predict more (redundant) quantized codes to generate images, which significantly increases the training cost and decreases the generating speed. Although the effectiveness and efficiency of image coding theory have been widely validated, how to introduce this idea into codebook learning remains unexplored.", + "bbox": [ + 75, + 292, + 472, + 640 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The key of applying image coding theory to codebook learning is to distinguish important image parts from unimportant ones correctly. Considering that the essential difference between these two sets lies in whether they can be restored if missing, we found that this distinction can be realized through the mask mechanism, i.e., the masked part is important if it cannot be faithfully restored, and otherwise unimportant. Based on the above observation, we thereby propose a novel two-stage generation paradigm upon the mask mechanism to relieve the model from modeling redundant information. Specifically, we first propose a Masked Quantization VAE (MQ-VAE) with two novel modules, i.e., an adaptive mask module for adaptively masking redundant region features before quantization, and an adaptive de-mask module for adaptively recovering the original grid image feature map to faithfully reconstruct original images after quantization. As for the adaptive mask module, it in", + "bbox": [ + 75, + 643, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "corporates a lightweight content-aware scoring network that learns to measure the importance of each image region feature. The features are then ranked by the importance scores and only a subset of high-scored features will be quantized further. As for the adaptive de-mask module, we design a direction-constrained self-attention to encourage the information flow from the unmasked regions to the masked regions while blocking the reverse, which aims to infer the original masked region information based on unmasked ones. Thanks to the adaptive mask and de-mask mechanism, our MQ-VAE removes the negative effects of redundant image regions and also shortens the sequence length to achieve both effectiveness and efficiency.", + "bbox": [ + 496, + 90, + 893, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Moreover, since different images have different important regions, the position of quantized codes in the feature map also dynamically changed. Therefore, we further propose Stackformer for learning to predict the combination of both codes and their corresponding positions. Concretely, the proposed Stackformer stacks a Code-Transformer and a Position-Transformer, where the Code-Transformer learns to predict the next code based on all previous codes and their positions, and the Position-Transformer learns to predict the next code's position based on all previous codes' positions and current code.", + "bbox": [ + 496, + 287, + 893, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "With our method, as shown in Figure 1(b), the codebook learning only includes the important regions, e.g., the structural regions, since unimportant ones like the background can be restored even if missing. And therefore the autoregressive model in the second stage could focus on modeling these important regions and brings better generation quality.", + "bbox": [ + 496, + 455, + 893, + 546 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In a nutshell, we summarize our main contributions as:", + "bbox": [ + 517, + 546, + 883, + 560 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Conceptually, we point out that existing codebook learning ignores distinguishing the perceptual importance of different image regions, which brings redundancy that degrades generation quality and decreases generation speed.", + "bbox": [ + 496, + 561, + 893, + 623 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Technically, (i) we propose MQ-VAE with a novel adaptive mask module to mask redundant region features before quantization and a novel adaptive de-mask module to recover the original feature map after quantization; (ii) we propose a novel Stackformer to predict the combination of both codes and their corresponding positions.", + "bbox": [ + 496, + 623, + 893, + 714 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Experimentally, comprehensive experiments on various generations validate our effectiveness and efficiency, i.e., we achieve $8.1\\%$ , $2.3\\%$ , and $18.6\\%$ FID improvement on un-, class-, and text-conditional state-of-the-art at million-level parameters, and faster generation speed compared to existing autoregressive models.", + "bbox": [ + 496, + 714, + 893, + 806 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 820, + 640, + 835 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Autoregressive Modeling for Image Generation", + "text_level": 1, + "bbox": [ + 500, + 845, + 890, + 863 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Autoregressive models for image generation have recently attracted increasing research attention and have", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2003", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/27735186e03b9dfe29d96d79864a9d9ed4bb3b557ec51166fa8b9afadc72e871.jpg", + "image_caption": [ + "Figure 2. The Illustration of our proposed two-stage generation framework. (a) In the first stage, MQ-VAE adaptively masks the redundant region features to prevent redundant codes while keeping important ones, which ensures that the original images can still be faithfully recovered. Here, $\\frac{10}{16}$ regions are masked and $\\frac{6}{16}$ regions are kept. (b) In the second stage, Stackformer stacks a Code-Transformer and a Position-Transformer to autoregressively predict the next code and its position in the original 2D feature map, respectively. (c) The attention mask of the proposed direction-constrained self-attention in the adaptive de-mask module for inferring masked regions features." + ], + "image_footnote": [], + "bbox": [ + 78, + 92, + 872, + 459 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "shown impressive results [10, 12, 13, 32, 34, 35, 37, 39, 42] among various generation tasks. Early autoregressive models [5, 33, 40] directly optimizing the likelihood of raw image pixels, e.g., Image-GPT [5] trains a transformer [41] to autoregressively predict pixels' cluster centroids, which could only generate images with a maximum resolution of $64 \\times 64$ . [39] presents the Vector Quantized Variational Autoencoder (VQVAE), which learns images' low-dimension discrete representation and models their distribution autoregressively. VQ-VAE2 [35] extends this approach using a hierarchy of discrete representations. VQGAN [13] further improves the perceptual quality of reconstructed images using adversarial [14, 19] and perceptual loss [27]. ViT-VQGAN [42] introduces a more expressive transformer backbone. RQ-VAE [28] uses Residual Quantization [21, 31] to iteratively quantizes a vector and its residuals and represent the vector as a stack of tokens. Although vector quantization has become the fundamental technique for modern visual autoregressive models, the critical removing redundancy in codebook learning has not been explored yet, which becomes a critical bottleneck.", + "bbox": [ + 75, + 573, + 472, + 891 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Masked Modeling", + "text_level": 1, + "bbox": [ + 500, + 573, + 674, + 589 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Masked modeling is popular among both natural language processing and computer vision. BERT [8] randomly masks a portion of the input sequence and trains models to predict the missing content. In the computer vision domain, the ViT [11] studies masked patch prediction for self-supervised learning. BEiT [1] proposes to predict discrete tokens. Most recently, MaskGIT [4] also used the masking strategy for VQ-based image generation. However, our proposed method differs from MaskGIT in two aspects: (1) Our primary motivation for the masking strategy applied in the proposed MQ-VAE in the first stage aims to learn a more compact and effective vector quantization (VQ) itself by masking perceptual unimportant regions, while MaskGIT uses masking strategy in the second stage to better use a learned VQ. (2) The mask in our proposed MQ-VAE is learned and adaptively changed according to different image content, while the mask in MaskGIT is randomly sampled for the mask-and-predict training. In conclusion, to the best of our knowledge, this is the first time that masked modeling has been applied for vector quantization.", + "bbox": [ + 496, + 598, + 890, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "2004", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 76, + 89, + 210, + 107 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We propose a novel two-stage framework with MQ-VAE and Stackformer for autoregressive image generation, as illustrated in Figure 2. MQ-VAE only masks redundant region features to prevent redundant codes and Stackformer stacks two transformers to autoregressively predict the next code and its position. In the following, we will first briefly revisit the formulation of vector quantization and then describe our proposed method in detail.", + "bbox": [ + 75, + 114, + 468, + 234 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Preliminary", + "text_level": 1, + "bbox": [ + 76, + 244, + 205, + 262 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We follow the definition and notations of previous works [13, 28]. Specifically, the codebook $\\mathcal{C} \\coloneqq \\{(k, e(k))\\}_{k \\in [K]}$ is defined as the set of finite pairs of code $k$ and its code embedding $e(k) \\in \\mathbb{R}^{n_z}$ . Here $K$ is the codebook size and $n_z$ is the code dimension. An image $\\mathbf{X} \\in \\mathbb{R}^{H_0 \\times W_0 \\times 3}$ is first encoded into grid features $\\mathbf{Z} = E(\\mathbf{X}) \\in \\mathbb{R}^{H \\times W \\times n_z}$ by the encoder $E$ , where $(H, W) = (H_0 / f, W_0 / f)$ and $f$ is the corresponding downsampling factor. For each vector $z \\in \\mathbb{R}^{n_z}$ in $\\mathbf{Z}$ , it is replaced with the code embedding that has the closest euclidean distance with it in the codebook $\\mathcal{C}$ through the vector quantization operation $\\mathcal{Q}(\\cdot)$ :", + "bbox": [ + 75, + 268, + 470, + 435 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {Q} (z; \\mathcal {C}) = \\arg \\min _ {k \\in [ K ]} | | z - e _ {k} | | _ {2} ^ {2}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 449, + 468, + 474 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $\\mathcal{Q}(z;\\mathcal{C})$ is the quantized code. $z^q = e(\\mathcal{Q}(z;\\mathcal{C}))$ is the quantized vector. By applying $\\mathcal{Q}(\\cdot)$ to each feature vector, we could get the quantized code map $\\mathbf{M}\\in [K]^{H\\times W}$ and the quantized features $\\mathbf{Z}^q\\in \\mathbb{R}^{H\\times W\\times n_z}$ . The original image is reconstructed by the decoder $D$ as $\\tilde{\\mathbf{X}} = D(\\mathbf{Z}^q)$ .", + "bbox": [ + 75, + 479, + 468, + 556 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Stage 1: MQ-VAE", + "text_level": 1, + "bbox": [ + 76, + 566, + 253, + 583 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Existing methods quantize each feature vector of $\\mathbf{Z}$ without distinguishing their different perceptual importance and thus bring redundancy in the learned codebook, which not only degrades the generation quality but also decreases the generation speed. To relieve the model from this redundancy, we propose MQ-VAE with two novel modules, i.e., the adaptive mask module for adaptively masking redundant region features before vector quantization and adaptive de-mask module for adaptively recovering the original grid image feature map after vector quantization.", + "bbox": [ + 75, + 589, + 468, + 739 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Adaptive Mask Module. The encoded grid feature map $\\mathbf{Z} \\in \\mathbb{R}^{H \\times W \\times n_e}$ is first flattened into $\\mathbf{Z} \\in \\mathbb{R}^{L \\times n_e}$ , where $L = H \\times W$ . The proposed adaptive mask module then uses a lightweight scoring network $f_s$ to measure the importance of each region feature $z_l$ in $\\mathbf{Z}$ , which is implemented as a two-layer MLP:", + "bbox": [ + 75, + 741, + 468, + 832 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ns _ {l} = f _ {s} \\left(\\boldsymbol {z} _ {l}\\right), l = 1, \\dots , L. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 847, + 468, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The larger score $s_l$ is, the more important the region feature $z_l$ is. Then the region features are sorted in descending", + "bbox": [ + 75, + 869, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "order according to the predicted scores. The sorted region features and their scores are denoted as $\\{z_l^{\\prime}\\}$ and $\\{s_l^{\\prime}\\}$ respectively, where $l = 1,\\dots ,L$ . To enable the learning of $f_{s}$ , the predicted scores are further multiplied with the normalized region features as modulating factors. We select the top $N$ scoring vectors as the important region features,", + "bbox": [ + 496, + 90, + 890, + 181 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {Z}} = \\left\\{z _ {l} ^ {\\prime \\prime} \\mid z _ {l} ^ {\\prime \\prime} = \\operatorname {L a y e r N o r m} \\left(z _ {l} ^ {\\prime}\\right) * s _ {l} ^ {\\prime} \\right\\}, l = 1, \\dots , N. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 200, + 890, + 219 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {P}} = \\left\\{p _ {z _ {l} ^ {\\prime \\prime}} \\mid p _ {z _ {l} ^ {\\prime \\prime}} \\in \\{0, \\dots , L \\} \\right\\}, l = 1, \\dots , N. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 223, + 890, + 244 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $\\hat{\\mathbf{Z}}$ denotes the selected important region features set, and $\\hat{\\mathbf{P}}$ denotes the corresponding position set that represents the position of each selected region feature in the original 2D feature map. The selected number $N = \\alpha \\times L$ , where $\\alpha$ is a constant fractional value. The mask ratio is defined as $1 - \\alpha$ . This design also enables a flexible trade-off between the image generation speed and image generation quality, which we will discuss in experiments. After obtaining $\\hat{\\mathbf{Z}}$ , we further apply the quantization function $\\mathcal{Q}$ to each of them and obtain the quantized important region features set $\\hat{\\mathbf{Z}}^q$ as well as its code matrix $\\hat{\\mathbf{M}}$ .", + "bbox": [ + 496, + 250, + 890, + 416 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Adaptive De-mask Module. After quantization, we fill the quantized features $\\hat{\\mathbf{Z}}^q$ back into the original 2D feature map according to $\\hat{\\mathbf{P}}$ , while other masked positions are filled with a uniformly initialized learnable mask code embedding, as shown in Figure 2(a). Directly inputting filled grid features to the decoder $D$ could bring sub-optimal reconstruction results since the mask code embedding here only serves as the placeholders that contain little information. Therefore, we further propose the adaptive de-mask module, which applies a novel direction-constrained self-attention to encourage the information flow from unmasked regions to the masked ones while blocking the reverse. Such a design allows the model to utilize the unmasked region features to infer the masked ones while also preventing the masked regions to have negative impacts on the unmasked ones since they are less informative.", + "bbox": [ + 496, + 417, + 890, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our adaptive de-mask module is implemented as $H$ identical sub-modules, where each consists of a direction-constrained self-attention block and a Resnet block. The direction-constrained self-attention is mathematically formed as (Resnet block is omitted for simplicity):", + "bbox": [ + 496, + 659, + 890, + 736 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {q}, \\boldsymbol {k}, \\boldsymbol {v} = \\boldsymbol {W} _ {q} \\hat {\\boldsymbol {Z}} ^ {q, h}, \\boldsymbol {W} _ {k} \\hat {\\boldsymbol {Z}} ^ {q, h}, \\boldsymbol {W} _ {v} \\hat {\\boldsymbol {Z}} ^ {q, h} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 753, + 890, + 772 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {A} = \\left(\\operatorname {S o f t M a x} \\left(\\frac {\\boldsymbol {q} \\boldsymbol {k} ^ {T}}{\\sqrt {n _ {e}}}\\right)\\right) \\odot \\boldsymbol {B} _ {\\boldsymbol {h}} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 775, + 890, + 811 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {Z}} ^ {q, h + 1} = \\boldsymbol {A} \\boldsymbol {v}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 643, + 813, + 890, + 830 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here $h\\in \\{1,\\dots ,H\\}$ and $\\hat{\\mathbf{Z}}^{q,1}$ is the filled quantized grid features. $\\pmb {W}_q,\\pmb {W}_k,\\pmb {W}_v\\in \\mathbb{R}^{n_e\\times n_e}$ are the learnable parameters. $\\pmb{B}_{h}$ is the attention mask at $h$ sub-module. Specifically, since the initial mask code contains little information,", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "2005", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "we define $B_{1} = [b_{l}\\in \\{0,1\\} ,|l = 1,\\dots,L]\\in \\mathbb{R}^{1\\times L}$ to forbid it from attending to other unmasked codes to avoid negative impact, where 0 for the mask position and 1 for the unmasked. Considering that the mask code is updated with more and more information after each sub-module, we propose to synchronously amplify its interaction with other codes step by step through a mask updating mechanism:", + "bbox": [ + 75, + 90, + 472, + 196 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {B} _ {h + 1} = \\sqrt {\\boldsymbol {B} _ {h}}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 217, + 210, + 468, + 229 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the initial 0 in $B_{1}$ is replaced with a small number 0.02, since $\\sqrt{0}$ will always result in 0. Finally, the original image is recovered by the decoder as $\\tilde{\\mathbf{X}} = D(\\hat{\\mathbf{Z}}^{q,H})$ .", + "bbox": [ + 75, + 236, + 468, + 282 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Stage 2: Stackformer", + "text_level": 1, + "bbox": [ + 76, + 291, + 277, + 306 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The perceptual important regions of different images vary. Therefore the positions of quantized codes in the feature map also dynamically change along with the image content. As a result, our proposed MQ-VAE formulates an image as both the code sequence $\\hat{\\mathbf{M}}$ and the code position sequence $\\hat{\\mathbf{P}}$ . To simultaneously learn the combination of the codes and their positions, we propose a novel Stackformer, which stacks a Code-Transformer and a Position-Transformer. The Code-Transformer learns to predict the next code based on all previous steps' codes and their positions, while the Position-Transformer learns to predict the next code's position based on all previous steps' positions and current code. Directly treating the importance descending order sequence $\\hat{\\mathbf{M}}$ and $\\hat{\\mathbf{P}}$ as the inputs are natural, but the dramatic position changes of adjacent code could make the network hard to converge. For example, the position of the first code may be in the upper left corner of the image, while the position of the second code may be in the lower right corner of the image. Therefore, we further propose to use the raster-scan order [13] to rearrange both sequences to deal with the converge problem.", + "bbox": [ + 75, + 315, + 472, + 632 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Mathematically, taking the raster-scan code and code position sequence $(\\overline{\\mathbf{M}},\\overline{\\mathbf{P}}) =$ rearrange $(\\hat{\\mathbf{M}},\\hat{\\mathbf{P}})$ , Stackformer learns $p(\\overline{\\mathbf{M}},\\overline{\\mathbf{P}})$ , which is autoregressively factorized as:", + "bbox": [ + 75, + 633, + 468, + 678 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\np (\\overline {{\\mathbf {M}}}, \\overline {{\\mathbf {P}}}) = \\prod_ {l = 1} ^ {N} p (\\overline {{\\mathbf {M}}} _ {l} | \\overline {{\\mathbf {M}}} _ {< l}, \\overline {{\\mathbf {P}}} _ {< l}) p (\\overline {{\\mathbf {P}}} _ {l} | \\overline {{\\mathbf {M}}} _ {\\leq l}, \\overline {{\\mathbf {P}}} _ {< l}) \\quad (9)\n$$\n", + "text_format": "latex", + "bbox": [ + 93, + 702, + 468, + 744 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Code-Transformer takes the sum of code embeddings $e_{c}(\\cdot)$ and code position embedding $e_{p}(\\cdot)$ as inputs:", + "bbox": [ + 76, + 755, + 468, + 787 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {U} _ {c} = \\boldsymbol {e} _ {c} \\left(\\overline {{\\mathbf {M}}} _ {[ 1: N _ {c} + N ]}\\right) + \\boldsymbol {e} _ {p 1} \\left(\\overline {{\\mathbf {P}}} _ {[ 1: N _ {c} + N ]}\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 800, + 468, + 819 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $N_{c}$ is the condition length. For the unconditional generation, we add a $<\\mathrm{sos}>$ code at the start of the code and code position sequence. For conditioning, we append class or text codes to the start of the code sequence and the same length of $<\\mathrm{sos}>$ code to the code position sequence. We", + "bbox": [ + 75, + 824, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "further add an extra learned absolute position embedding to $\\mathbf{U}_c$ to form the final input, which makes the network aware of the absolute position of the sequence. After processing by Code-Transformer, the output hidden vector $\\mathbf{H}_c$ encodes both code and their position information and is used for the next code prediction. The negative log-likelihood (NLL) loss for code autoregressive training is:", + "bbox": [ + 496, + 90, + 890, + 196 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {c o d e}} = \\mathbb {E} \\left[ - \\log p \\left(\\overline {{\\mathbf {M}}} _ {l} \\mid \\overline {{\\mathbf {M}}} _ {< l}, \\overline {{\\mathbf {P}}} _ {< l} \\right]\\right). \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 210, + 890, + 229 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Position-Transformer takes the sum of CodeTransformer's output hidden vector $\\mathbf{H}_c$ and an extra code embedding as input:", + "bbox": [ + 496, + 233, + 890, + 279 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {U} _ {p} = \\mathbf {H} _ {c} \\left[ N _ {c}: N _ {c} + N - 1 \\right] + \\boldsymbol {e} _ {c} \\left(\\overline {{\\mathbf {M}}} _ {\\left[ N _ {c} + 1: N _ {c} + N \\right]}\\right). \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 508, + 304, + 890, + 323 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here $\\mathbf{U}_p$ is the input for Position-Transformer and the information of current code is included in $e_c(\\overline{\\mathbf{M}}_{[N_c + 1:N_c + N]})$ . The design idea behind this is that when predicting the next code's position, the model should not only be aware of previous steps' codes and their position information but also should be aware of current code information. The negative log-likelihood (NLL) for position autoregressive training is:", + "bbox": [ + 496, + 332, + 890, + 438 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {p o s i t i o n}} = \\mathbb {E} \\left[ - \\log p \\left(\\overline {{\\mathbf {P}}} _ {l} \\mid \\overline {{\\mathbf {M}}} _ {\\leq l}, \\overline {{\\mathbf {P}}} _ {< l}\\right) \\right]. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 450, + 890, + 469 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training & Inference. The total loss for training Stackformer is defined as:", + "bbox": [ + 500, + 474, + 890, + 503 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {c o d e}} + \\mathcal {L} _ {\\text {p o s i t i o n}}. \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 518, + 890, + 536 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The inference procedure is illustrated in Algorithm 1, where we take the unconditional generation as an example and conditional generation can be derived accordingly.", + "bbox": [ + 496, + 541, + 890, + 587 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Algorithm 1 Unconditional sampling of Stackformer.", + "text_level": 1, + "bbox": [ + 500, + 601, + 854, + 617 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Input: The sample step $N$ ;", + "bbox": [ + 500, + 621, + 687, + 636 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The code sequence $M_{\\text{sample}}$ of a single $<\\text{sos}>$ ;", + "bbox": [ + 527, + 637, + 846, + 651 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The code position sequence $P_{\\text{sample}}$ of a single $<\\text{sos}>$ .", + "bbox": [ + 527, + 651, + 888, + 666 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Output: The generated image $\\mathcal{I}$ .", + "bbox": [ + 500, + 666, + 723, + 680 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: for each $n \\in [1, N]$ do", + "2: $H_{c} =$ Code-Transformer(Msample, $P_{\\text{sample}}$ );", + "3: Sample next code $M_{n}$ by $H_{c}$ ;", + "4: $M_{\\text{sample}} = \\operatorname{concat}(M_{\\text{sample}}, M_n)$ ;", + "5: $H_{p} = \\mathrm{Position - Transformer}(H_{c},M_{\\mathrm{sample}}[2:])$", + "6: Mask already sampled positions to avoid conflicts;", + "7: Sample next code position $P_{n}$ by $H_{p}$ ;", + "8: $P_{\\text{sample}} = \\operatorname{concat}(P_{\\text{sample}}, P_n)$ ;", + "9: end for", + "10: Re-map $M_{\\text{sample}}$ to 2D code map according to $P_{\\text{sample}}$ and the rest are filled with the mask code;", + "11: Decode the code map to the image $\\mathcal{I}$", + "12: return The generated image $\\mathcal{I}$" + ], + "bbox": [ + 504, + 681, + 888, + 878 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "2006", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d642cc09f5748e1d217765a7f04920d0f50e2599648c0ccd33a5e27be05d3a92.jpg", + "image_caption": [ + "goldfish, Carassius auratus" + ], + "image_footnote": [ + "common iguna, iguana, Iguania iguna" + ], + "bbox": [ + 78, + 88, + 238, + 210 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/bf756d23b791a541459c31aca5ae0bd7b40c47c40b7268dea3a9b82b6334bc3b.jpg", + "image_caption": [ + "Figure 3. Left: Our unconditional generated images on FFHQ benchmark. Right: Our class-conditional generated images on ImageNet." + ], + "image_footnote": [], + "bbox": [ + 330, + 90, + 565, + 212 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/072aba8ed82ac3f6b31ed191945af875d1a622daa5b2ff4298f6a453a2b6f98d.jpg", + "image_caption": [ + "lorikeet curly-coated retriever" + ], + "image_footnote": [], + "bbox": [ + 648, + 90, + 888, + 210 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 76, + 262, + 200, + 279 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 76, + 286, + 282, + 303 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "**Benchmarks.** We validate our model for unconditional, class-conditional, and text-conditional image generation tasks on FFHQ [22], ImageNet [7], and MS-COCO [30] benchmarks respectively, with $256 \\times 256$ image resolution.", + "bbox": [ + 75, + 309, + 467, + 369 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics. Following previous works [13,28,42], the standard Frechet Inception Distance (FID) [16] is adopted for evaluating the generation and reconstruction quality (denoted as rFID). Inception Score (IS) [2] is also used for class-conditional generation on the ImageNet benchmark. FID and IS are calculated by sampling 50k images. rFID is calculated over the entire test set.", + "bbox": [ + 75, + 369, + 467, + 474 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implemented details. The architecture of MQ-VAE exactly follows [13] except for the proposed mask and demask modules, with the codebook size of $K = 1024$ . For the de-mask module, the sub-module number $H = 8$ . For the Stackformer, we implement two settings: a small version uses 18 transformer encoder blocks for the CodeTransformer and another 6 transformer encoder blocks for the Position-Transformer with a total of 307M parameters, and a base version uses 36 transformer encoder blocks for the Code-Transformer and another 12 transformer encoder blocks for the Position-Transformer with a total of 607M parameters to further demonstrate our scalability. The generation results are reported with a $25\\%$ mask ratio at $32 \\times 32$ resolution feature map using eight RTX-3090 GPUs. Top-k and top-p sampling are used to report the best performance. More details could be found in the supplementary.", + "bbox": [ + 75, + 476, + 467, + 717 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparison with state-of-the-art methods", + "text_level": 1, + "bbox": [ + 76, + 726, + 434, + 742 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Unconditional generation. We first compare with million-level state-of-the-art autoregressive models in Table 1. Our model significantly outperforms other autoregressive models with the same parameters (307M). With more parameters, we further increase the FID from 6.84 to 5.67, which demonstrates our scalability. We also compare with other types of unconditional state-of-the-art and large-scale big models in Table 3, where we also achieve top-level performance. Our qualitative unconditional generation results are shown on the left of Figure 3.", + "bbox": [ + 75, + 750, + 467, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/54235b9c6bb43ce6c5332aaa520b5bc394987c00d70b5abaf7ed83d1e45b4365.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Methods#ParamsFID↓
DCT [32]738M13.06
VQGAN [13](72.1+307)M11.4
RQ-Transformer [28](100+355)M10.38
Mo-VQGAN [44](82.7+307)M8.52
Stackformer(44.4+307)M6.84
Stackformer(44.4+607)M5.67
", + "bbox": [ + 545, + 261, + 846, + 367 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/f8dcdbbfd10bd6b688fe718768bf830fea69c6c28547b1cf1774a23e5446afe6.jpg", + "table_caption": [ + "Table 1. Comparison of autoregressive unconditional generation at million-level parameters on FFHQ [22] benchmark. #Params splits in (VAE + autoregressive model)." + ], + "table_footnote": [], + "table_body": "
Model TypeMethods#ParamsFID↓IS↑
DiffusionADM [9]554M10.94101.0
GANBigGAN [3]164M7.53168.6
GANBigGAN-deep [3]112M6.84203.6
BidirectionMaskGIT [4]227M6.18182.1
AutoregressiveDCT [32]738M36.5n/a
AutoregressiveVQ-GAN† [13]679M17.0376.85
AutoregressiveRQ-Transformer [28]821M13.11104.3
AutoregressiveMo-VQGAN [44]383M7.13138.3
AutoregressiveStackformer651M6.04172.6
", + "bbox": [ + 501, + 440, + 901, + 592 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2fec8a8f2405bd0701686216a132af3b52e9b1a51f9a86ddb5d391fbbfe37cd1.jpg", + "table_caption": [ + "Table 2. Comparison of class-conditional image generation at million-level parameters without rejection sampling on ImageNet [7]. $\\dagger$ denotes the model we train with the same setup with ours." + ], + "table_footnote": [], + "table_body": "
Model TypeMethods#ParamsFID↓
VAEVDVAE [6]115M28.5
DiffusionImageBART [12]3.5B9.57
GANStyleGAN2 [23]-3.8
GANBigGAN [3]164M12.4
AutoregressiveViT-VQGAN [42]2.2B5.3
AutoregressiveStackformer651M5.67
", + "bbox": [ + 516, + 662, + 875, + 768 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Comparison with other types of state-of-the-art generative models and large-scale billion-level parameters autoregressive models on unconditional FFHQ [22] benchmark.", + "bbox": [ + 498, + 782, + 890, + 824 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Class-conditional generation. We first compare with million-level state-of-the-art in Table 2. We achieve the best FID score compared to all types of models including the re", + "bbox": [ + 500, + 854, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "2007", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a07faec2fbeb3939e0a08df57afd59a032a863a4f9c4f0134bec02159a88153f.jpg", + "image_caption": [ + "Figure 4. The visualization of our adaptive mask module which learns to mask unimportant regions on ImageNet [7]. In the importance map, red denotes high scores while blue denotes low scores." + ], + "image_footnote": [], + "bbox": [ + 76, + 88, + 480, + 339 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/24e037fbf3b3c191e4aff33a020cddf85ea302eb6c003376e6ead38253a73968.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 486, + 88, + 890, + 339 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4c09514718ffa612ecdcfc91ad123a493f179776149ad02b443849eaeb6e86b7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Model TypeMethods#ParamsFID↓IS↑
DiffusionImageBART [12]3.5B21.1961.6
AutoregressiveVQVAE2 [35]13.5B3145
AutoregressiveVQ-GAN [13]1.4B15.7878.3
AutoregressiveViT-VQGAN [42]2.2B4.17175.1
AutoregressiveRQ-Transformer [28]3.8B7.55134
AutoregressiveStackformer651M6.04172.6
", + "bbox": [ + 78, + 400, + 478, + 508 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3cb8bce118f68f7f5628312c1b63b0877582f4d68225fd239ecca2329559299d.jpg", + "table_caption": [ + "Table 4. Comparison between our model and large-scale billion-level parameters models of class-conditional generation without rejection sampling on ImageNet [7] benchmark." + ], + "table_footnote": [], + "table_body": "
Model TypeMethodFID↓
GANDMGAN [45]32.64
GANXMCGAN† [43]50.08
GANDFGAN [38]21.42
GANSSA-GAN [29]19.37
GANDSE-GAN [18]15.30
DiffusionVQ-Diffusion [15]19.75
AutoregressiveVQ-GAN† [13]22.28
AutoregressiveStackformer10.08
", + "bbox": [ + 140, + 577, + 408, + 696 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "cent Mo-VQGAN [44] and RQ-Transformer [28]. We also compare our million-level model with existing billion-level big models in Table 4, where we also achieve top performance with fewer parameters and is only inferior to ViT-VQGAN big model. Our qualitative class-conditional generation results are shown on the right of Figure 3.", + "bbox": [ + 75, + 777, + 468, + 868 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Text-conditional generation. We compare with existing text-conditional state-of-the-art without extra web-scale", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/799f74514619ec0893a15b0480df1ea68d6c532a17136acd0a0b1a3e03ddb661.jpg", + "table_caption": [ + "Table 5. Comparison of text-conditional generation on MS-COCO [30] without using extra web-scale data or pre-trained models. † denotes reproduced results under our same experimental setting." + ], + "table_footnote": [], + "table_body": "
mask ratiofmask typerFID↓FID↓usage↑ (%)
0%32-8.113.570.02
0%16-4.4611.463.89
10%16adaptive4.557.8172.34
25%16adaptive4.797.6778.22
25%16random6.1312.2167.48
50%16adaptive5.318.3684.29
50%16random7.85515.6769.04
75%16adaptive7.6211.7187.60
75%16random10.5817.6270.41
", + "bbox": [ + 501, + 400, + 893, + 549 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 6. Ablations of adaptive mask module on FFHQ. Here $f$ is the downsampling factor. The codebook usage is calculated as the percentage of used codes over the entire test dataset.", + "bbox": [ + 498, + 561, + 892, + 604 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "data or pretrained models on MS-COCO [30] for fair comparison in Table.5. We achieve $18.6\\%$ FID improvement.", + "bbox": [ + 498, + 633, + 890, + 664 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ab2290e9ed4f0135bfecae58bf9f3a4ae106fef5e63f31e24865f5a475bfa0e7.jpg", + "image_caption": [ + "Figure 5. The PCA of learned codebook (1024 codebook size)." + ], + "image_footnote": [], + "bbox": [ + 501, + 676, + 890, + 739 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablations", + "text_level": 1, + "bbox": [ + 500, + 785, + 609, + 799 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct ablations on $16 \\times 16$ resolution feature map using four RTX-3090 GPUs for saving computation resources and the experimental trends are all consistent with $32 \\times 32$ resolution feature map of the main results.", + "bbox": [ + 498, + 809, + 890, + 869 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablations of adaptive mask module. As shown in Table 6, our proposed learned adaptive mask mechanism sig-", + "bbox": [ + 498, + 869, + 890, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "2008", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8e5d0ff3c2401bbac153b064e0c2a2cdad968303d2db7b3843890b39e81654ae.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 89, + 267, + 202 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f64343d27867bc681c2c5f8e5ef7b145c4133d0bbdf180b79fd8275f6c796003.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 281, + 89, + 467, + 200 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d5458069c96a0984b11dafb0fee346aa576ad4748396de6b9a3f0bc4cbc7f961.jpg", + "image_caption": [ + "(a) Training and validation curves of VQGAN and our Stackformer", + "(b) The sampling speed of Stackformer and VQGAN according to batch size and mask radio", + "Figure 6. Comparison of training & validation curves and sample speed between VQGAN [13] and Stackformer." + ], + "image_footnote": [], + "bbox": [ + 81, + 220, + 467, + 330 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "nificantly outperforms the random one, which quantitatively validates that the adaptive mask module enables the model to learn perceptually important regions.", + "bbox": [ + 75, + 428, + 468, + 474 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our model with $10\\%$ and $25\\%$ mask radio has only slightly inferior reconstruction compared with VQGAN [13], but achieves a significant improvement in generation quality, which indicates that the effectiveness of focusing autoregressive models on modeling important regions. When we further increase the mask radio to $50\\%$ and $75\\%$ , the final generation quality drops, we believe the reason lies that an improper high mask radio will inevitably mask some important regions that greatly decrease the reconstruction results and hinder autoregressive modeling.", + "bbox": [ + 75, + 474, + 468, + 626 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The redundancy of the existing learned codebook can be verified from two aspects: i) the PCA of the learned codebook in Figure 5, where each point is a code and closer codes have more similar semantics. We show many codes in VQGAN's codebook overlap, which indicates these codes have nearly the same semantics and are thus redundant. The redundancy increase (more overlaps) when VQGAN uses more code to represent images (smaller downsampling factor $f$ ). The redundancy is largely alleviated in our MQ-VAE. ii) in Table 6, a higher codebook usage indicates more \"useful\" codes in the codebook and thus less redundant. VQGAN has a lower usage compared with our MQ-VAE.", + "bbox": [ + 75, + 627, + 468, + 808 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We visualize the training and validation curves of VQGAN and Stackformer in Figure 6(a). Previous autoregressive models [13, 28, 42] always suffer from the overfitting problem while our Stackformer successfully gets rid of it, which indicates the better generalization of our masked discrete representation and our model.", + "bbox": [ + 75, + 810, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b026544233b4ed45424de1733138754e2d99920d7e3712950a5eb5e235a5eb97.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Model settingrFID↓FID↓
VQGAN4.4611.4
VQGAN*4.1711.02
MQ-VAE w/o de-mask7.0210.74
MQ-VAE w/ de-mask (SA)6.569.8
MQ-VAE w/ de-mask (DC-SA w/o mask update)5.848.92
MQ-VAE w/ de-mask (DC-SA w/ mask update)5.318.36
", + "bbox": [ + 504, + 88, + 887, + 186 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 7. Ablations of adaptive de-mask module on FFHQ. SA for self-attention and DC-SA for direction-constrained self-attention. \"VQGAN*\" is the stronger baseline, where the same numbers of SA and Resnet blocks as MQ-VAE's de-mask module are added.", + "bbox": [ + 498, + 196, + 890, + 252 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We compare the sampling speed on a single RTX-1080Ti in Figure 6(b). Compared with VQGAN, our $25\\%$ mask radio model achieves $32.72\\%$ quality improvement and $15.45\\%$ speed improvement, while Our $50\\%$ mask radio model achieves $26.67\\%$ quality improvement and $61.1\\%$ speed improvement. Therefore, our design enables a flexible trade-off between speed and quality.", + "bbox": [ + 496, + 277, + 890, + 383 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Finally, We visualize the learned mask in Figure 4, with a $75\\%$ mask ratio on $32 \\times 32$ resolution feature map, which validates that our proposed adaptive mask mechanism successfully learns to preserve the perceptual important image regions, i.e., the structural and edge regions of objects.", + "bbox": [ + 496, + 383, + 890, + 460 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablations of adaptive de-mask module. In Table 7, we show that MQ-VAE outperforms VQGAN and the stronger baseline (\"VQGAN*\"), which validates our effectiveness. We could conclude that the proposed direction-constrained self-attention and the mask updating mechanism both improve the reconstruction and generation quality.", + "bbox": [ + 496, + 460, + 890, + 551 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 564, + 617, + 579 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this study, we point out that the existing two-stage autoregressive generation paradigm ignores distinguishing the perceptual importance of different image regions, which brings redundancy that not only degrades generation quality but also decreases generation speed. We propose a novel two-stage generation paradigm with MQ-VAE and Stackformer to relieve the model from redundancy. MQ-VAE incorporates the adaptive mask module to mask redundant region features before quantization and the adaptive de-mask module to recover the original feature map after quantization. Stackformer then efficiently predict the combination of both codes and their positions. Comprehensive experiments on various types of image generation tasks validate the effectiveness and efficiency of our method.", + "bbox": [ + 496, + 589, + 890, + 800 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgments", + "text_level": 1, + "bbox": [ + 500, + 815, + 678, + 832 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work is supported by National Natural Science Foundation of China under Grants 62222212 and U19A2057, Science Fund for Creative Research Groups under Grant 62121002.", + "bbox": [ + 496, + 839, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "2009", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Hangbo Bao, Li Dong, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021.3", + "[2] Shane Barratt and Rishi Sharma. A note on the inception score. arXiv preprint arXiv:1801.01973, 2018. 6", + "[3] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. arXiv preprint arXiv:1809.11096, 2018. 6", + "[4] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T Freeman. Maskgit: Masked generative image transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11315-11325, 2022. 3, 6", + "[5] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In International Conference on Machine Learning, pages 1691-1703. PMLR, 2020. 3", + "[6] Rewon Child. Very deep vaes generalize autoregressive models and can outperform them on images. arXiv preprint arXiv:2011.10650, 2020. 6", + "[7] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6, 7", + "[8] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 3", + "[9] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34, 2021. 6", + "[10] Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, et al. Cogview: Mastering text-to-image generation via transformers. Advances in Neural Information Processing Systems, 34, 2021. 1, 3", + "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.3", + "[12] Patrick Esser, Robin Rombach, Andreas Blattmann, and Bjorn Ommer. Imagebart: Bidirectional context with multinomial diffusion for autoregressive image synthesis. Advances in Neural Information Processing Systems, 34, 2021. 1, 3, 6, 7", + "[13] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12873-12883, 2021. 1, 2, 3, 5, 6, 7, 8", + "[14] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 1, 3" + ], + "bbox": [ + 78, + 114, + 467, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. arXiv preprint arXiv:2111.14822, 2021. 7", + "[16] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6", + "[17] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 1", + "[18] Mengqi Huang, Zhendong Mao, Penghui Wang, Quan Wang, and Yongdong Zhang. Dse-gan: Dynamic semantic evolution generative adversarial network for text-to-image generation. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4345-4354, 2022. 7", + "[19] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 3", + "[20] Nikil Jayant, James Johnston, and Robert Safranek. Signal compression based on models of human perception. Proceedings of the IEEE, 81(10):1385-1422, 1993. 2", + "[21] Biing-Hwang Juang and A Gray. Multiple stage vector quantization for speech coding. In ICASSP'82. IEEE International Conference on Acoustics, Speech, and Signal Processing, volume 7, pages 597-600. IEEE, 1982. 3", + "[22] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 6", + "[23] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. 6", + "[24] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 1", + "[25] M Kocher and M Kunt. A contour-texture approach to picture coding. In ICASSP'82. IEEE International Conference on Acoustics, Speech, and Signal Processing, volume 7, pages 436-439. IEEE, 1982. 2", + "[26] Murat Kunt, Athanassios Ikonomopoulos, and Michel Kocher. Second-generation image-coding techniques. Proceedings of the IEEE, 73(4):549-574, 1985. 2", + "[27] Christian Ledig, Lucas Theis, Ferenc Huszár, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photorealistic single image super-resolution using a generative adversarial network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4681-4690, 2017. 3", + "[28] Doyup Lee, Chiheon Kim, Saehoon Kim, Minsu Cho, and Wook-Shin Han. Autoregressive image generation using residual quantization. arXiv preprint arXiv:2203.01941, 2022. 1, 2, 3, 4, 6, 7, 8" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "2010", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Wentong Liao, Kai Hu, Michael Ying Yang, and Bodo Rosenhahn. Text to image generation with semantic-spatial aware gan. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18187-18196, 2022. 7", + "[30] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 6, 7", + "[31] Julieta Martinez, Holger H Hoos, and James J Little. Stacked quantizers for compositional vector compression. arXiv preprint arXiv:1411.2173, 2014. 3", + "[32] Charlie Nash, Jacob Menick, Sander Dieleman, and Peter W Battaglia. Generating images with sparse representations. arXiv preprint arXiv:2103.03841, 2021. 3, 6", + "[33] Niki Parmar, Ashish Vaswani, Jakob Uszkoreit, Lukasz Kaiser, Noam Shazeer, Alexander Ku, and Dustin Tran. Image transformer. In International Conference on Machine Learning, pages 4055-4064. PMLR, 2018. 3", + "[34] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 1, 3", + "[35] Ali Razavi, Aaron Van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. Advances in neural information processing systems, 32, 2019. 1, 3, 7", + "[36] Danilo Rezende and Shakir Mohamed. Variational inference with normalizing flows. In International conference on machine learning, pages 1530-1538. PMLR, 2015. 1", + "[37] Woncheol Shin, Gyubok Lee, Jiyoung Lee, Joonseok Lee, and Edward Choi. Translation-equivariant image quantizer for bi-directional image-text generation. arXiv preprint arXiv:2112.00384, 2021. 1, 3", + "[38] Ming Tao, Hao Tang, Songsong Wu, Nicu Sebe, Xiao-Yuan Jing, Fei Wu, and Bingkun Bao. Df-gan: Deep fusion generative adversarial networks for text-to-image synthesis. arXiv preprint arXiv:2008.05865, 2020. 7", + "[39] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 1, 3", + "[40] Aaron Van Oord, Nal Kalchbrenner, and Koray Kavukcuoglu. Pixel recurrent neural networks. In International conference on machine learning, pages 1747-1756. PMLR, 2016. 1, 3", + "[41] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3", + "[42] Jiahui Yu, Xin Li, Jing Yu Koh, Han Zhang, Ruoming Pang, James Qin, Alexander Ku, Yuanzhong Xu, Jason Baldridge, and Yonghui Wu. Vector-quantized image modeling with improved vqgan. arXiv preprint arXiv:2110.04627, 2021. 2, 3, 6, 7, 8", + "[43] Han Zhang, Jing Yu Koh, Jason Baldridge, Honglak Lee, and Yinfei Yang. Cross-modal contrastive learning for text-to" + ], + "bbox": [ + 78, + 90, + 468, + 902 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "image generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 833-842, 2021. 7", + "[44] Chuanxia Zheng, Long Tung Vuong, Jianfei Cai, and Dinh Phung. Movq: Modulating quantized vectors for high-fidelity image generation. arXiv preprint arXiv:2209.09002, 2022.6,7", + "[45] Minfeng Zhu, Pingbo Pan, Wei Chen, and Yi Yang. Dm-gan: Dynamic memory generative adversarial networks for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5802-5810, 2019. 7" + ], + "bbox": [ + 501, + 92, + 890, + 260 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "2011", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Not All Image Regions Matter_ Masked Vector Quantization for Autoregressive Image Generation/2a88178e-8ce2-4c0f-b3f1-bb99f0452012_model.json b/2023/Not All Image Regions Matter_ Masked Vector Quantization for Autoregressive Image Generation/2a88178e-8ce2-4c0f-b3f1-bb99f0452012_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e7bf63d6d7ebce1b8dc3300b4f6d7ec0c035c4bc --- /dev/null +++ b/2023/Not All Image Regions Matter_ Masked Vector Quantization for Autoregressive Image Generation/2a88178e-8ce2-4c0f-b3f1-bb99f0452012_model.json @@ -0,0 +1,2354 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.15, + 0.131, + 0.82, + 0.177 + ], + "angle": 0, + "content": "Not All Image Regions Matter: Masked Vector Quantization for Autoregressive Image Generation" + }, + { + "type": "text", + "bbox": [ + 0.224, + 0.203, + 0.78, + 0.222 + ], + "angle": 0, + "content": "Mengqi Huang\\(^{1}\\), Zhendong Mao\\(^{1,3*}\\), Quan Wang\\(^{2}\\), Yongdong Zhang\\(^{1,3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.086, + 0.222, + 0.92, + 0.275 + ], + "angle": 0, + "content": "1University of Science and Technology of China, Hefei, China; 2MOE Key Laboratory of Trustworthy, Distributed Computing and Service, Beijing University of Posts and Telecommunications, Beijing, China; 3Institute of Artificial intelligence, Hefei Comprehensive National Science Center, Hefei, China" + }, + { + "type": "text", + "bbox": [ + 0.169, + 0.278, + 0.832, + 0.292 + ], + "angle": 0, + "content": "huangmq@mail.ustc.edu.cn, {zdmao, zhyd73}@ustc.edu.cn, wangquan@bupt.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.327, + 0.314, + 0.343 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.359, + 0.474, + 0.751 + ], + "angle": 0, + "content": "Existing autoregressive models follow the two-stage generation paradigm that first learns a codebook in the latent space for image reconstruction and then completes the image generation autoregressively based on the learned codebook. However, existing codebook learning simply models all local region information of images without distinguishing their different perceptual importance, which brings redundancy in the learned codebook that not only limits the next stage's autoregressive model's ability to model important structure but also results in high training cost and slow generation speed. In this study, we borrow the idea of importance perception from classical image coding theory and propose a novel two-stage framework, which consists of Masked Quantization VAE (MQ-VAE) and Stackformer, to relieve the model from modeling redundancy. Specifically, MQ-VAE incorporates an adaptive mask module for masking redundant region features before quantization and an adaptive de-mask module for recovering the original grid image feature map to faithfully reconstruct the original images after quantization. Then, Stackformer learns to predict the combination of the next code and its position in the feature map. Comprehensive experiments on various image generation validate our effectiveness and efficiency. Code will be released at https://github.com/CrossmodalGroup/MaskedVectorQuantization." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.779, + 0.21, + 0.794 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.804, + 0.47, + 0.88 + ], + "angle": 0, + "content": "Deep generative models of images have received significant improvements over the past few years and broadly fall into two categories: likelihood-based models, which include VAEs [24], flow-based [36], diffusion models [17] and autoregressive models [40], and generative adversarial" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.328, + 0.697, + 0.477 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.552, + 0.479, + 0.653, + 0.491 + ], + "angle": 0, + "content": "(a) Existing methods" + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.328, + 0.892, + 0.477 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.761, + 0.479, + 0.837, + 0.49 + ], + "angle": 0, + "content": "(b) Our method" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.501, + 0.895, + 0.655 + ], + "angle": 0, + "content": "Figure 1. Illustration of our motivation. (a) Existing works model all local regions without distinguishing their perceptual importance in stage 1, which not only brings redundancy (e.g., the textural regions like the background) in the learned codebook but also make the autoregressive models overly focus on modeling this redundancy and hinder other important structural regions modeling. (b) The codebook learning in our method only includes the important regions, e.g., the structural regions like corners and edges, since other unimportant ones can be restored even if missing, and thus autoregressive model could focus on modeling these important regions in stage 2 and results in better generation quality." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.687, + 0.894, + 0.778 + ], + "angle": 0, + "content": "networks (GANs) [14], which use discriminator networks to distinguish samples from generator networks and real examples. Compared with GANs, likelihood-based models' training objective, i.e., the negative log-likelihood (NLL) or its upper bound, incentives learning the full data distribution and allows for detecting overfitting." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Among the likelihood-based models, autoregressive models have recently attracted increasing attention for their impressive modeling ability and scalability. Recent autoregressive image generation [10, 12, 13, 28, 28, 34, 35, 37, 39] follows the two-stage generation paradigm, i.e., the first stage learns a codebook in the latent space for image reconstruction and the second stage completes the image generation in the raster-scan [13] order by autoregressive models" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.335, + 0.9 + ], + "angle": 0, + "content": "*Zhendong Mao is the corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2002" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.289 + ], + "angle": 0, + "content": "based on the learned codebook. Since codebook learning in the first stage defines the discrete image representation for the next autoregressive modeling, a high-quality codebook is the key to generate high-quality images. Several recent works focus on improving the codebook learning in the first stage, e.g., VQGAN [13] introduces adversarial loss and perceptual loss. ViT-VQGAN [42] introduces a more expressive transformer backbone. RQ-VAE [28] introduces the residual quantization to reduce the resolution of the latent space. In general, the essence of existing codebook learning is the modeling of all local region information (i.e., an \\( 8 \\times 8 \\) or \\( 16 \\times 16 \\) patch) of images in the dataset, without distinguishing their different perceptual importance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.293, + 0.473, + 0.641 + ], + "angle": 0, + "content": "In this study, we point out that existing codebook learning exists gaps with classical image coding theory [20, 25, 26], the basic idea of which is to remove redundant information by perceiving the importance of different regions in images. The image coding theory reveals that an ideal image coding method should only encode images' perceptually important regions (i.e., which cannot be restored if missing) while discarding the unimportant ones (i.e., which can be restored by other image regions even if missing). The neglect of considering such perceptual importance in existing works poses problems in two aspects, as illustrated in Figure 1(a): (1) the existence of this large amount of repetitive and redundant information brings redundancy to the learned codebook, which further makes the autoregressive model in the next stage overly focus on modeling this redundancy while overlooking other important regions and finally degrades generation quality. (2) the redundancy makes the autoregressive model need to predict more (redundant) quantized codes to generate images, which significantly increases the training cost and decreases the generating speed. Although the effectiveness and efficiency of image coding theory have been widely validated, how to introduce this idea into codebook learning remains unexplored." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.645, + 0.473, + 0.903 + ], + "angle": 0, + "content": "The key of applying image coding theory to codebook learning is to distinguish important image parts from unimportant ones correctly. Considering that the essential difference between these two sets lies in whether they can be restored if missing, we found that this distinction can be realized through the mask mechanism, i.e., the masked part is important if it cannot be faithfully restored, and otherwise unimportant. Based on the above observation, we thereby propose a novel two-stage generation paradigm upon the mask mechanism to relieve the model from modeling redundant information. Specifically, we first propose a Masked Quantization VAE (MQ-VAE) with two novel modules, i.e., an adaptive mask module for adaptively masking redundant region features before quantization, and an adaptive de-mask module for adaptively recovering the original grid image feature map to faithfully reconstruct original images after quantization. As for the adaptive mask module, it in" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.288 + ], + "angle": 0, + "content": "corporates a lightweight content-aware scoring network that learns to measure the importance of each image region feature. The features are then ranked by the importance scores and only a subset of high-scored features will be quantized further. As for the adaptive de-mask module, we design a direction-constrained self-attention to encourage the information flow from the unmasked regions to the masked regions while blocking the reverse, which aims to infer the original masked region information based on unmasked ones. Thanks to the adaptive mask and de-mask mechanism, our MQ-VAE removes the negative effects of redundant image regions and also shortens the sequence length to achieve both effectiveness and efficiency." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.289, + 0.895, + 0.456 + ], + "angle": 0, + "content": "Moreover, since different images have different important regions, the position of quantized codes in the feature map also dynamically changed. Therefore, we further propose Stackformer for learning to predict the combination of both codes and their corresponding positions. Concretely, the proposed Stackformer stacks a Code-Transformer and a Position-Transformer, where the Code-Transformer learns to predict the next code based on all previous codes and their positions, and the Position-Transformer learns to predict the next code's position based on all previous codes' positions and current code." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.456, + 0.895, + 0.547 + ], + "angle": 0, + "content": "With our method, as shown in Figure 1(b), the codebook learning only includes the important regions, e.g., the structural regions, since unimportant ones like the background can be restored even if missing. And therefore the autoregressive model in the second stage could focus on modeling these important regions and brings better generation quality." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.547, + 0.884, + 0.561 + ], + "angle": 0, + "content": "In a nutshell, we summarize our main contributions as:" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.563, + 0.894, + 0.624 + ], + "angle": 0, + "content": "Conceptually, we point out that existing codebook learning ignores distinguishing the perceptual importance of different image regions, which brings redundancy that degrades generation quality and decreases generation speed." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.624, + 0.894, + 0.715 + ], + "angle": 0, + "content": "Technically, (i) we propose MQ-VAE with a novel adaptive mask module to mask redundant region features before quantization and a novel adaptive de-mask module to recover the original feature map after quantization; (ii) we propose a novel Stackformer to predict the combination of both codes and their corresponding positions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.715, + 0.894, + 0.807 + ], + "angle": 0, + "content": "Experimentally, comprehensive experiments on various generations validate our effectiveness and efficiency, i.e., we achieve \\(8.1\\%\\), \\(2.3\\%\\), and \\(18.6\\%\\) FID improvement on un-, class-, and text-conditional state-of-the-art at million-level parameters, and faster generation speed compared to existing autoregressive models." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.821, + 0.642, + 0.837 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.847, + 0.892, + 0.864 + ], + "angle": 0, + "content": "2.1. Autoregressive Modeling for Image Generation" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Autoregressive models for image generation have recently attracted increasing research attention and have" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2003" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.093, + 0.874, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.478, + 0.896, + 0.551 + ], + "angle": 0, + "content": "Figure 2. The Illustration of our proposed two-stage generation framework. (a) In the first stage, MQ-VAE adaptively masks the redundant region features to prevent redundant codes while keeping important ones, which ensures that the original images can still be faithfully recovered. Here, \\(\\frac{10}{16}\\) regions are masked and \\(\\frac{6}{16}\\) regions are kept. (b) In the second stage, Stackformer stacks a Code-Transformer and a Position-Transformer to autoregressively predict the next code and its position in the original 2D feature map, respectively. (c) The attention mask of the proposed direction-constrained self-attention in the adaptive de-mask module for inferring masked regions features." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.574, + 0.473, + 0.892 + ], + "angle": 0, + "content": "shown impressive results [10, 12, 13, 32, 34, 35, 37, 39, 42] among various generation tasks. Early autoregressive models [5, 33, 40] directly optimizing the likelihood of raw image pixels, e.g., Image-GPT [5] trains a transformer [41] to autoregressively predict pixels' cluster centroids, which could only generate images with a maximum resolution of \\(64 \\times 64\\). [39] presents the Vector Quantized Variational Autoencoder (VQVAE), which learns images' low-dimension discrete representation and models their distribution autoregressively. VQ-VAE2 [35] extends this approach using a hierarchy of discrete representations. VQGAN [13] further improves the perceptual quality of reconstructed images using adversarial [14, 19] and perceptual loss [27]. ViT-VQGAN [42] introduces a more expressive transformer backbone. RQ-VAE [28] uses Residual Quantization [21, 31] to iteratively quantizes a vector and its residuals and represent the vector as a stack of tokens. Although vector quantization has become the fundamental technique for modern visual autoregressive models, the critical removing redundancy in codebook learning has not been explored yet, which becomes a critical bottleneck." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.574, + 0.676, + 0.59 + ], + "angle": 0, + "content": "2.2. Masked Modeling" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Masked modeling is popular among both natural language processing and computer vision. BERT [8] randomly masks a portion of the input sequence and trains models to predict the missing content. In the computer vision domain, the ViT [11] studies masked patch prediction for self-supervised learning. BEiT [1] proposes to predict discrete tokens. Most recently, MaskGIT [4] also used the masking strategy for VQ-based image generation. However, our proposed method differs from MaskGIT in two aspects: (1) Our primary motivation for the masking strategy applied in the proposed MQ-VAE in the first stage aims to learn a more compact and effective vector quantization (VQ) itself by masking perceptual unimportant regions, while MaskGIT uses masking strategy in the second stage to better use a learned VQ. (2) The mask in our proposed MQ-VAE is learned and adaptively changed according to different image content, while the mask in MaskGIT is randomly sampled for the mask-and-predict training. In conclusion, to the best of our knowledge, this is the first time that masked modeling has been applied for vector quantization." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2004" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.09, + 0.212, + 0.108 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.115, + 0.47, + 0.236 + ], + "angle": 0, + "content": "We propose a novel two-stage framework with MQ-VAE and Stackformer for autoregressive image generation, as illustrated in Figure 2. MQ-VAE only masks redundant region features to prevent redundant codes and Stackformer stacks two transformers to autoregressively predict the next code and its position. In the following, we will first briefly revisit the formulation of vector quantization and then describe our proposed method in detail." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.246, + 0.206, + 0.263 + ], + "angle": 0, + "content": "3.1. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.269, + 0.471, + 0.436 + ], + "angle": 0, + "content": "We follow the definition and notations of previous works [13, 28]. Specifically, the codebook \\(\\mathcal{C} \\coloneqq \\{(k, e(k))\\}_{k \\in [K]}\\) is defined as the set of finite pairs of code \\(k\\) and its code embedding \\(e(k) \\in \\mathbb{R}^{n_z}\\). Here \\(K\\) is the codebook size and \\(n_z\\) is the code dimension. An image \\(\\mathbf{X} \\in \\mathbb{R}^{H_0 \\times W_0 \\times 3}\\) is first encoded into grid features \\(\\mathbf{Z} = E(\\mathbf{X}) \\in \\mathbb{R}^{H \\times W \\times n_z}\\) by the encoder \\(E\\), where \\((H, W) = (H_0 / f, W_0 / f)\\) and \\(f\\) is the corresponding downsampling factor. For each vector \\(z \\in \\mathbb{R}^{n_z}\\) in \\(\\mathbf{Z}\\), it is replaced with the code embedding that has the closest euclidean distance with it in the codebook \\(\\mathcal{C}\\) through the vector quantization operation \\(\\mathcal{Q}(\\cdot)\\):" + }, + { + "type": "equation", + "bbox": [ + 0.165, + 0.45, + 0.47, + 0.475 + ], + "angle": 0, + "content": "\\[\n\\mathcal {Q} (z; \\mathcal {C}) = \\arg \\min _ {k \\in [ K ]} | | z - e _ {k} | | _ {2} ^ {2}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.481, + 0.47, + 0.558 + ], + "angle": 0, + "content": "Here, \\(\\mathcal{Q}(z;\\mathcal{C})\\) is the quantized code. \\(z^q = e(\\mathcal{Q}(z;\\mathcal{C}))\\) is the quantized vector. By applying \\(\\mathcal{Q}(\\cdot)\\) to each feature vector, we could get the quantized code map \\(\\mathbf{M}\\in [K]^{H\\times W}\\) and the quantized features \\(\\mathbf{Z}^q\\in \\mathbb{R}^{H\\times W\\times n_z}\\). The original image is reconstructed by the decoder \\(D\\) as \\(\\tilde{\\mathbf{X}} = D(\\mathbf{Z}^q)\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.567, + 0.254, + 0.584 + ], + "angle": 0, + "content": "3.2. Stage 1: MQ-VAE" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.59, + 0.47, + 0.741 + ], + "angle": 0, + "content": "Existing methods quantize each feature vector of \\(\\mathbf{Z}\\) without distinguishing their different perceptual importance and thus bring redundancy in the learned codebook, which not only degrades the generation quality but also decreases the generation speed. To relieve the model from this redundancy, we propose MQ-VAE with two novel modules, i.e., the adaptive mask module for adaptively masking redundant region features before vector quantization and adaptive de-mask module for adaptively recovering the original grid image feature map after vector quantization." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.742, + 0.47, + 0.833 + ], + "angle": 0, + "content": "Adaptive Mask Module. The encoded grid feature map \\(\\mathbf{Z} \\in \\mathbb{R}^{H \\times W \\times n_e}\\) is first flattened into \\(\\mathbf{Z} \\in \\mathbb{R}^{L \\times n_e}\\), where \\(L = H \\times W\\). The proposed adaptive mask module then uses a lightweight scoring network \\(f_s\\) to measure the importance of each region feature \\(z_l\\) in \\(\\mathbf{Z}\\), which is implemented as a two-layer MLP:" + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.848, + 0.469, + 0.865 + ], + "angle": 0, + "content": "\\[\ns _ {l} = f _ {s} \\left(\\boldsymbol {z} _ {l}\\right), l = 1, \\dots , L. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.902 + ], + "angle": 0, + "content": "The larger score \\( s_l \\) is, the more important the region feature \\( z_l \\) is. Then the region features are sorted in descending" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.183 + ], + "angle": 0, + "content": "order according to the predicted scores. The sorted region features and their scores are denoted as \\(\\{z_l^{\\prime}\\}\\) and \\(\\{s_l^{\\prime}\\}\\) respectively, where \\(l = 1,\\dots ,L\\). To enable the learning of \\(f_{s}\\), the predicted scores are further multiplied with the normalized region features as modulating factors. We select the top \\(N\\) scoring vectors as the important region features," + }, + { + "type": "equation", + "bbox": [ + 0.516, + 0.201, + 0.892, + 0.22 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {Z}} = \\left\\{z _ {l} ^ {\\prime \\prime} \\mid z _ {l} ^ {\\prime \\prime} = \\operatorname {L a y e r N o r m} \\left(z _ {l} ^ {\\prime}\\right) * s _ {l} ^ {\\prime} \\right\\}, l = 1, \\dots , N. \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.555, + 0.224, + 0.892, + 0.245 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {P}} = \\left\\{p _ {z _ {l} ^ {\\prime \\prime}} \\mid p _ {z _ {l} ^ {\\prime \\prime}} \\in \\{0, \\dots , L \\} \\right\\}, l = 1, \\dots , N. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.251, + 0.892, + 0.417 + ], + "angle": 0, + "content": "Here, \\(\\hat{\\mathbf{Z}}\\) denotes the selected important region features set, and \\(\\hat{\\mathbf{P}}\\) denotes the corresponding position set that represents the position of each selected region feature in the original 2D feature map. The selected number \\(N = \\alpha \\times L\\), where \\(\\alpha\\) is a constant fractional value. The mask ratio is defined as \\(1 - \\alpha\\). This design also enables a flexible trade-off between the image generation speed and image generation quality, which we will discuss in experiments. After obtaining \\(\\hat{\\mathbf{Z}}\\), we further apply the quantization function \\(\\mathcal{Q}\\) to each of them and obtain the quantized important region features set \\(\\hat{\\mathbf{Z}}^q\\) as well as its code matrix \\(\\hat{\\mathbf{M}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.418, + 0.892, + 0.659 + ], + "angle": 0, + "content": "Adaptive De-mask Module. After quantization, we fill the quantized features \\(\\hat{\\mathbf{Z}}^q\\) back into the original 2D feature map according to \\(\\hat{\\mathbf{P}}\\), while other masked positions are filled with a uniformly initialized learnable mask code embedding, as shown in Figure 2(a). Directly inputting filled grid features to the decoder \\(D\\) could bring sub-optimal reconstruction results since the mask code embedding here only serves as the placeholders that contain little information. Therefore, we further propose the adaptive de-mask module, which applies a novel direction-constrained self-attention to encourage the information flow from unmasked regions to the masked ones while blocking the reverse. Such a design allows the model to utilize the unmasked region features to infer the masked ones while also preventing the masked regions to have negative impacts on the unmasked ones since they are less informative." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.892, + 0.737 + ], + "angle": 0, + "content": "Our adaptive de-mask module is implemented as \\( H \\) identical sub-modules, where each consists of a direction-constrained self-attention block and a Resnet block. The direction-constrained self-attention is mathematically formed as (Resnet block is omitted for simplicity):" + }, + { + "type": "equation", + "bbox": [ + 0.564, + 0.755, + 0.892, + 0.773 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {q}, \\boldsymbol {k}, \\boldsymbol {v} = \\boldsymbol {W} _ {q} \\hat {\\boldsymbol {Z}} ^ {q, h}, \\boldsymbol {W} _ {k} \\hat {\\boldsymbol {Z}} ^ {q, h}, \\boldsymbol {W} _ {v} \\hat {\\boldsymbol {Z}} ^ {q, h} \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.596, + 0.776, + 0.891, + 0.813 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {A} = \\left(\\operatorname {S o f t M a x} \\left(\\frac {\\boldsymbol {q} \\boldsymbol {k} ^ {T}}{\\sqrt {n _ {e}}}\\right)\\right) \\odot \\boldsymbol {B} _ {\\boldsymbol {h}} \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.645, + 0.814, + 0.891, + 0.832 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {Z}} ^ {q, h + 1} = \\boldsymbol {A} \\boldsymbol {v}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Here \\(h\\in \\{1,\\dots ,H\\}\\) and \\(\\hat{\\mathbf{Z}}^{q,1}\\) is the filled quantized grid features. \\(\\pmb {W}_q,\\pmb {W}_k,\\pmb {W}_v\\in \\mathbb{R}^{n_e\\times n_e}\\) are the learnable parameters. \\(\\pmb{B}_{h}\\) is the attention mask at \\(h\\) sub-module. Specifically, since the initial mask code contains little information," + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2005" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.198 + ], + "angle": 0, + "content": "we define \\( B_{1} = [b_{l}\\in \\{0,1\\} ,|l = 1,\\dots,L]\\in \\mathbb{R}^{1\\times L} \\) to forbid it from attending to other unmasked codes to avoid negative impact, where 0 for the mask position and 1 for the unmasked. Considering that the mask code is updated with more and more information after each sub-module, we propose to synchronously amplify its interaction with other codes step by step through a mask updating mechanism:" + }, + { + "type": "equation", + "bbox": [ + 0.218, + 0.211, + 0.47, + 0.23 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {B} _ {h + 1} = \\sqrt {\\boldsymbol {B} _ {h}}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.237, + 0.47, + 0.283 + ], + "angle": 0, + "content": "where the initial 0 in \\( B_{1} \\) is replaced with a small number 0.02, since \\( \\sqrt{0} \\) will always result in 0. Finally, the original image is recovered by the decoder as \\( \\tilde{\\mathbf{X}} = D(\\hat{\\mathbf{Z}}^{q,H}) \\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.292, + 0.279, + 0.308 + ], + "angle": 0, + "content": "3.3. Stage 2: Stackformer" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.316, + 0.473, + 0.633 + ], + "angle": 0, + "content": "The perceptual important regions of different images vary. Therefore the positions of quantized codes in the feature map also dynamically change along with the image content. As a result, our proposed MQ-VAE formulates an image as both the code sequence \\(\\hat{\\mathbf{M}}\\) and the code position sequence \\(\\hat{\\mathbf{P}}\\). To simultaneously learn the combination of the codes and their positions, we propose a novel Stackformer, which stacks a Code-Transformer and a Position-Transformer. The Code-Transformer learns to predict the next code based on all previous steps' codes and their positions, while the Position-Transformer learns to predict the next code's position based on all previous steps' positions and current code. Directly treating the importance descending order sequence \\(\\hat{\\mathbf{M}}\\) and \\(\\hat{\\mathbf{P}}\\) as the inputs are natural, but the dramatic position changes of adjacent code could make the network hard to converge. For example, the position of the first code may be in the upper left corner of the image, while the position of the second code may be in the lower right corner of the image. Therefore, we further propose to use the raster-scan order [13] to rearrange both sequences to deal with the converge problem." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.634, + 0.47, + 0.679 + ], + "angle": 0, + "content": "Mathematically, taking the raster-scan code and code position sequence \\((\\overline{\\mathbf{M}},\\overline{\\mathbf{P}}) =\\) rearrange \\((\\hat{\\mathbf{M}},\\hat{\\mathbf{P}})\\), Stackformer learns \\(p(\\overline{\\mathbf{M}},\\overline{\\mathbf{P}})\\), which is autoregressively factorized as:" + }, + { + "type": "equation", + "bbox": [ + 0.094, + 0.703, + 0.47, + 0.745 + ], + "angle": 0, + "content": "\\[\np (\\overline {{\\mathbf {M}}}, \\overline {{\\mathbf {P}}}) = \\prod_ {l = 1} ^ {N} p (\\overline {{\\mathbf {M}}} _ {l} | \\overline {{\\mathbf {M}}} _ {< l}, \\overline {{\\mathbf {P}}} _ {< l}) p (\\overline {{\\mathbf {P}}} _ {l} | \\overline {{\\mathbf {M}}} _ {\\leq l}, \\overline {{\\mathbf {P}}} _ {< l}) \\quad (9)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.756, + 0.47, + 0.788 + ], + "angle": 0, + "content": "Code-Transformer takes the sum of code embeddings \\( e_{c}(\\cdot) \\) and code position embedding \\( e_{p}(\\cdot) \\) as inputs:" + }, + { + "type": "equation", + "bbox": [ + 0.134, + 0.801, + 0.47, + 0.82 + ], + "angle": 0, + "content": "\\[\n\\mathbf {U} _ {c} = \\boldsymbol {e} _ {c} \\left(\\overline {{\\mathbf {M}}} _ {[ 1: N _ {c} + N ]}\\right) + \\boldsymbol {e} _ {p 1} \\left(\\overline {{\\mathbf {P}}} _ {[ 1: N _ {c} + N ]}\\right), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.901 + ], + "angle": 0, + "content": "where \\( N_{c} \\) is the condition length. For the unconditional generation, we add a \\( <\\mathrm{sos}> \\) code at the start of the code and code position sequence. For conditioning, we append class or text codes to the start of the code sequence and the same length of \\( <\\mathrm{sos}> \\) code to the code position sequence. We" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.198 + ], + "angle": 0, + "content": "further add an extra learned absolute position embedding to \\(\\mathbf{U}_c\\) to form the final input, which makes the network aware of the absolute position of the sequence. After processing by Code-Transformer, the output hidden vector \\(\\mathbf{H}_c\\) encodes both code and their position information and is used for the next code prediction. The negative log-likelihood (NLL) loss for code autoregressive training is:" + }, + { + "type": "equation", + "bbox": [ + 0.575, + 0.211, + 0.892, + 0.23 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {c o d e}} = \\mathbb {E} \\left[ - \\log p \\left(\\overline {{\\mathbf {M}}} _ {l} \\mid \\overline {{\\mathbf {M}}} _ {< l}, \\overline {{\\mathbf {P}}} _ {< l} \\right]\\right). \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.234, + 0.892, + 0.28 + ], + "angle": 0, + "content": "Position-Transformer takes the sum of CodeTransformer's output hidden vector \\(\\mathbf{H}_c\\) and an extra code embedding as input:" + }, + { + "type": "equation", + "bbox": [ + 0.509, + 0.305, + 0.892, + 0.324 + ], + "angle": 0, + "content": "\\[\n\\mathbf {U} _ {p} = \\mathbf {H} _ {c} \\left[ N _ {c}: N _ {c} + N - 1 \\right] + \\boldsymbol {e} _ {c} \\left(\\overline {{\\mathbf {M}}} _ {\\left[ N _ {c} + 1: N _ {c} + N \\right]}\\right). \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.333, + 0.892, + 0.439 + ], + "angle": 0, + "content": "Here \\(\\mathbf{U}_p\\) is the input for Position-Transformer and the information of current code is included in \\(e_c(\\overline{\\mathbf{M}}_{[N_c + 1:N_c + N]})\\). The design idea behind this is that when predicting the next code's position, the model should not only be aware of previous steps' codes and their position information but also should be aware of current code information. The negative log-likelihood (NLL) for position autoregressive training is:" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.452, + 0.892, + 0.47 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {p o s i t i o n}} = \\mathbb {E} \\left[ - \\log p \\left(\\overline {{\\mathbf {P}}} _ {l} \\mid \\overline {{\\mathbf {M}}} _ {\\leq l}, \\overline {{\\mathbf {P}}} _ {< l}\\right) \\right]. \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.476, + 0.892, + 0.504 + ], + "angle": 0, + "content": "Training & Inference. The total loss for training Stackformer is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.624, + 0.52, + 0.892, + 0.537 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {c o d e}} + \\mathcal {L} _ {\\text {p o s i t i o n}}. \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.542, + 0.892, + 0.588 + ], + "angle": 0, + "content": "The inference procedure is illustrated in Algorithm 1, where we take the unconditional generation as an example and conditional generation can be derived accordingly." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.602, + 0.856, + 0.618 + ], + "angle": 0, + "content": "Algorithm 1 Unconditional sampling of Stackformer." + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.622, + 0.688, + 0.637 + ], + "angle": 0, + "content": "Input: The sample step \\(N\\);" + }, + { + "type": "text", + "bbox": [ + 0.528, + 0.638, + 0.847, + 0.652 + ], + "angle": 0, + "content": "The code sequence \\( M_{\\text{sample}} \\) of a single \\( <\\text{sos}> \\);" + }, + { + "type": "text", + "bbox": [ + 0.528, + 0.652, + 0.89, + 0.667 + ], + "angle": 0, + "content": "The code position sequence \\( P_{\\text{sample}} \\) of a single \\( <\\text{sos}> \\)." + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.667, + 0.724, + 0.681 + ], + "angle": 0, + "content": "Output: The generated image \\(\\mathcal{I}\\)." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.682, + 0.682, + 0.697 + ], + "angle": 0, + "content": "1: for each \\( n \\in [1, N] \\) do" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.697, + 0.856, + 0.713 + ], + "angle": 0, + "content": "2: \\( H_{c} = \\) Code-Transformer(Msample, \\( P_{\\text{sample}} \\));" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.713, + 0.743, + 0.727 + ], + "angle": 0, + "content": "3: Sample next code \\( M_{n} \\) by \\( H_{c} \\);" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.728, + 0.783, + 0.743 + ], + "angle": 0, + "content": "4: \\(M_{\\text{sample}} = \\operatorname{concat}(M_{\\text{sample}}, M_n)\\);" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.743, + 0.872, + 0.758 + ], + "angle": 0, + "content": "5: \\(H_{p} = \\mathrm{Position - Transformer}(H_{c},M_{\\mathrm{sample}}[2:])\\)" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.758, + 0.879, + 0.772 + ], + "angle": 0, + "content": "6: Mask already sampled positions to avoid conflicts;" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.773, + 0.795, + 0.788 + ], + "angle": 0, + "content": "7: Sample next code position \\( P_{n} \\) by \\( H_{p} \\);" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.788, + 0.767, + 0.804 + ], + "angle": 0, + "content": "8: \\(P_{\\text{sample}} = \\operatorname{concat}(P_{\\text{sample}}, P_n)\\);" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.805, + 0.583, + 0.816 + ], + "angle": 0, + "content": "9: end for" + }, + { + "type": "text", + "bbox": [ + 0.506, + 0.818, + 0.89, + 0.847 + ], + "angle": 0, + "content": "10: Re-map \\( M_{\\text{sample}} \\) to 2D code map according to \\( P_{\\text{sample}} \\) and the rest are filled with the mask code;" + }, + { + "type": "text", + "bbox": [ + 0.506, + 0.848, + 0.773, + 0.863 + ], + "angle": 0, + "content": "11: Decode the code map to the image \\(\\mathcal{I}\\)" + }, + { + "type": "text", + "bbox": [ + 0.506, + 0.864, + 0.737, + 0.879 + ], + "angle": 0, + "content": "12: return The generated image \\(\\mathcal{I}\\)" + }, + { + "type": "list", + "bbox": [ + 0.506, + 0.682, + 0.89, + 0.879 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2006" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.089, + 0.239, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.245, + 0.113, + 0.326, + 0.132 + ], + "angle": 0, + "content": "goldfish, Carassius auratus" + }, + { + "type": "image_footnote", + "bbox": [ + 0.249, + 0.169, + 0.321, + 0.197 + ], + "angle": 0, + "content": "common iguna, iguana, Iguania iguna" + }, + { + "type": "image", + "bbox": [ + 0.331, + 0.091, + 0.566, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.578, + 0.118, + 0.638, + 0.193 + ], + "angle": 0, + "content": "lorikeet curly-coated retriever" + }, + { + "type": "image", + "bbox": [ + 0.649, + 0.091, + 0.889, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.083, + 0.223, + 0.884, + 0.238 + ], + "angle": 0, + "content": "Figure 3. Left: Our unconditional generated images on FFHQ benchmark. Right: Our class-conditional generated images on ImageNet." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.263, + 0.202, + 0.28 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.287, + 0.284, + 0.304 + ], + "angle": 0, + "content": "4.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.31, + 0.468, + 0.37 + ], + "angle": 0, + "content": "**Benchmarks.** We validate our model for unconditional, class-conditional, and text-conditional image generation tasks on FFHQ [22], ImageNet [7], and MS-COCO [30] benchmarks respectively, with \\(256 \\times 256\\) image resolution." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.371, + 0.468, + 0.476 + ], + "angle": 0, + "content": "Metrics. Following previous works [13,28,42], the standard Frechet Inception Distance (FID) [16] is adopted for evaluating the generation and reconstruction quality (denoted as rFID). Inception Score (IS) [2] is also used for class-conditional generation on the ImageNet benchmark. FID and IS are calculated by sampling 50k images. rFID is calculated over the entire test set." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.477, + 0.468, + 0.718 + ], + "angle": 0, + "content": "Implemented details. The architecture of MQ-VAE exactly follows [13] except for the proposed mask and demask modules, with the codebook size of \\( K = 1024 \\). For the de-mask module, the sub-module number \\( H = 8 \\). For the Stackformer, we implement two settings: a small version uses 18 transformer encoder blocks for the CodeTransformer and another 6 transformer encoder blocks for the Position-Transformer with a total of 307M parameters, and a base version uses 36 transformer encoder blocks for the Code-Transformer and another 12 transformer encoder blocks for the Position-Transformer with a total of 607M parameters to further demonstrate our scalability. The generation results are reported with a \\( 25\\% \\) mask ratio at \\( 32 \\times 32 \\) resolution feature map using eight RTX-3090 GPUs. Top-k and top-p sampling are used to report the best performance. More details could be found in the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.727, + 0.436, + 0.743 + ], + "angle": 0, + "content": "4.2. Comparison with state-of-the-art methods" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Unconditional generation. We first compare with million-level state-of-the-art autoregressive models in Table 1. Our model significantly outperforms other autoregressive models with the same parameters (307M). With more parameters, we further increase the FID from 6.84 to 5.67, which demonstrates our scalability. We also compare with other types of unconditional state-of-the-art and large-scale big models in Table 3, where we also achieve top-level performance. Our qualitative unconditional generation results are shown on the left of Figure 3." + }, + { + "type": "table", + "bbox": [ + 0.547, + 0.262, + 0.847, + 0.368 + ], + "angle": 0, + "content": "
Methods#ParamsFID↓
DCT [32]738M13.06
VQGAN [13](72.1+307)M11.4
RQ-Transformer [28](100+355)M10.38
Mo-VQGAN [44](82.7+307)M8.52
Stackformer(44.4+307)M6.84
Stackformer(44.4+607)M5.67
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.382, + 0.892, + 0.424 + ], + "angle": 0, + "content": "Table 1. Comparison of autoregressive unconditional generation at million-level parameters on FFHQ [22] benchmark. #Params splits in (VAE + autoregressive model)." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.441, + 0.902, + 0.593 + ], + "angle": 0, + "content": "
Model TypeMethods#ParamsFID↓IS↑
DiffusionADM [9]554M10.94101.0
GANBigGAN [3]164M7.53168.6
GANBigGAN-deep [3]112M6.84203.6
BidirectionMaskGIT [4]227M6.18182.1
AutoregressiveDCT [32]738M36.5n/a
AutoregressiveVQ-GAN† [13]679M17.0376.85
AutoregressiveRQ-Transformer [28]821M13.11104.3
AutoregressiveMo-VQGAN [44]383M7.13138.3
AutoregressiveStackformer651M6.04172.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.603, + 0.892, + 0.646 + ], + "angle": 0, + "content": "Table 2. Comparison of class-conditional image generation at million-level parameters without rejection sampling on ImageNet [7]. \\(\\dagger\\) denotes the model we train with the same setup with ours." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.663, + 0.876, + 0.77 + ], + "angle": 0, + "content": "
Model TypeMethods#ParamsFID↓
VAEVDVAE [6]115M28.5
DiffusionImageBART [12]3.5B9.57
GANStyleGAN2 [23]-3.8
GANBigGAN [3]164M12.4
AutoregressiveViT-VQGAN [42]2.2B5.3
AutoregressiveStackformer651M5.67
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.784, + 0.892, + 0.825 + ], + "angle": 0, + "content": "Table 3. Comparison with other types of state-of-the-art generative models and large-scale billion-level parameters autoregressive models on unconditional FFHQ [22] benchmark." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Class-conditional generation. We first compare with million-level state-of-the-art in Table 2. We achieve the best FID score compared to all types of models including the re" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2007" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.089, + 0.482, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.487, + 0.089, + 0.891, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.349, + 0.895, + 0.379 + ], + "angle": 0, + "content": "Figure 4. The visualization of our adaptive mask module which learns to mask unimportant regions on ImageNet [7]. In the importance map, red denotes high scores while blue denotes low scores." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.401, + 0.48, + 0.51 + ], + "angle": 0, + "content": "
Model TypeMethods#ParamsFID↓IS↑
DiffusionImageBART [12]3.5B21.1961.6
AutoregressiveVQVAE2 [35]13.5B3145
AutoregressiveVQ-GAN [13]1.4B15.7878.3
AutoregressiveViT-VQGAN [42]2.2B4.17175.1
AutoregressiveRQ-Transformer [28]3.8B7.55134
AutoregressiveStackformer651M6.04172.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.52, + 0.472, + 0.563 + ], + "angle": 0, + "content": "Table 4. Comparison between our model and large-scale billion-level parameters models of class-conditional generation without rejection sampling on ImageNet [7] benchmark." + }, + { + "type": "table", + "bbox": [ + 0.141, + 0.578, + 0.409, + 0.698 + ], + "angle": 0, + "content": "
Model TypeMethodFID↓
GANDMGAN [45]32.64
GANXMCGAN† [43]50.08
GANDFGAN [38]21.42
GANSSA-GAN [29]19.37
GANDSE-GAN [18]15.30
DiffusionVQ-Diffusion [15]19.75
AutoregressiveVQ-GAN† [13]22.28
AutoregressiveStackformer10.08
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.708, + 0.469, + 0.751 + ], + "angle": 0, + "content": "Table 5. Comparison of text-conditional generation on MS-COCO [30] without using extra web-scale data or pre-trained models. † denotes reproduced results under our same experimental setting." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.779, + 0.469, + 0.869 + ], + "angle": 0, + "content": "cent Mo-VQGAN [44] and RQ-Transformer [28]. We also compare our million-level model with existing billion-level big models in Table 4, where we also achieve top performance with fewer parameters and is only inferior to ViT-VQGAN big model. Our qualitative class-conditional generation results are shown on the right of Figure 3." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Text-conditional generation. We compare with existing text-conditional state-of-the-art without extra web-scale" + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.401, + 0.895, + 0.55 + ], + "angle": 0, + "content": "
mask ratiofmask typerFID↓FID↓usage↑ (%)
0%32-8.113.570.02
0%16-4.4611.463.89
10%16adaptive4.557.8172.34
25%16adaptive4.797.6778.22
25%16random6.1312.2167.48
50%16adaptive5.318.3684.29
50%16random7.85515.6769.04
75%16adaptive7.6211.7187.60
75%16random10.5817.6270.41
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.563, + 0.893, + 0.606 + ], + "angle": 0, + "content": "Table 6. Ablations of adaptive mask module on FFHQ. Here \\( f \\) is the downsampling factor. The codebook usage is calculated as the percentage of used codes over the entire test dataset." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.634, + 0.892, + 0.665 + ], + "angle": 0, + "content": "data or pretrained models on MS-COCO [30] for fair comparison in Table.5. We achieve \\(18.6\\%\\) FID improvement." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.678, + 0.892, + 0.741 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.508, + 0.751, + 0.882, + 0.765 + ], + "angle": 0, + "content": "Figure 5. The PCA of learned codebook (1024 codebook size)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.786, + 0.611, + 0.8 + ], + "angle": 0, + "content": "4.3. Ablations" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.81, + 0.892, + 0.87 + ], + "angle": 0, + "content": "We conduct ablations on \\(16 \\times 16\\) resolution feature map using four RTX-3090 GPUs for saving computation resources and the experimental trends are all consistent with \\(32 \\times 32\\) resolution feature map of the main results." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Ablations of adaptive mask module. As shown in Table 6, our proposed learned adaptive mask mechanism sig-" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2008" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.09, + 0.268, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.282, + 0.09, + 0.468, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.129, + 0.209, + 0.432, + 0.22 + ], + "angle": 0, + "content": "(a) Training and validation curves of VQGAN and our Stackformer" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.221, + 0.468, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.124, + 0.337, + 0.429, + 0.358 + ], + "angle": 0, + "content": "(b) The sampling speed of Stackformer and VQGAN according to batch size and mask radio" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.372, + 0.47, + 0.401 + ], + "angle": 0, + "content": "Figure 6. Comparison of training & validation curves and sample speed between VQGAN [13] and Stackformer." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.429, + 0.469, + 0.476 + ], + "angle": 0, + "content": "nificantly outperforms the random one, which quantitatively validates that the adaptive mask module enables the model to learn perceptually important regions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.476, + 0.469, + 0.627 + ], + "angle": 0, + "content": "Our model with \\(10\\%\\) and \\(25\\%\\) mask radio has only slightly inferior reconstruction compared with VQGAN [13], but achieves a significant improvement in generation quality, which indicates that the effectiveness of focusing autoregressive models on modeling important regions. When we further increase the mask radio to \\(50\\%\\) and \\(75\\%\\), the final generation quality drops, we believe the reason lies that an improper high mask radio will inevitably mask some important regions that greatly decrease the reconstruction results and hinder autoregressive modeling." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.628, + 0.469, + 0.809 + ], + "angle": 0, + "content": "The redundancy of the existing learned codebook can be verified from two aspects: i) the PCA of the learned codebook in Figure 5, where each point is a code and closer codes have more similar semantics. We show many codes in VQGAN's codebook overlap, which indicates these codes have nearly the same semantics and are thus redundant. The redundancy increase (more overlaps) when VQGAN uses more code to represent images (smaller downsampling factor \\( f \\)). The redundancy is largely alleviated in our MQ-VAE. ii) in Table 6, a higher codebook usage indicates more \"useful\" codes in the codebook and thus less redundant. VQGAN has a lower usage compared with our MQ-VAE." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.469, + 0.901 + ], + "angle": 0, + "content": "We visualize the training and validation curves of VQGAN and Stackformer in Figure 6(a). Previous autoregressive models [13, 28, 42] always suffer from the overfitting problem while our Stackformer successfully gets rid of it, which indicates the better generalization of our masked discrete representation and our model." + }, + { + "type": "table", + "bbox": [ + 0.506, + 0.089, + 0.888, + 0.187 + ], + "angle": 0, + "content": "
Model settingrFID↓FID↓
VQGAN4.4611.4
VQGAN*4.1711.02
MQ-VAE w/o de-mask7.0210.74
MQ-VAE w/ de-mask (SA)6.569.8
MQ-VAE w/ de-mask (DC-SA w/o mask update)5.848.92
MQ-VAE w/ de-mask (DC-SA w/ mask update)5.318.36
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.197, + 0.892, + 0.253 + ], + "angle": 0, + "content": "Table 7. Ablations of adaptive de-mask module on FFHQ. SA for self-attention and DC-SA for direction-constrained self-attention. \"VQGAN*\" is the stronger baseline, where the same numbers of SA and Resnet blocks as MQ-VAE's de-mask module are added." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.279, + 0.892, + 0.385 + ], + "angle": 0, + "content": "We compare the sampling speed on a single RTX-1080Ti in Figure 6(b). Compared with VQGAN, our \\(25\\%\\) mask radio model achieves \\(32.72\\%\\) quality improvement and \\(15.45\\%\\) speed improvement, while Our \\(50\\%\\) mask radio model achieves \\(26.67\\%\\) quality improvement and \\(61.1\\%\\) speed improvement. Therefore, our design enables a flexible trade-off between speed and quality." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.385, + 0.892, + 0.461 + ], + "angle": 0, + "content": "Finally, We visualize the learned mask in Figure 4, with a \\(75\\%\\) mask ratio on \\(32 \\times 32\\) resolution feature map, which validates that our proposed adaptive mask mechanism successfully learns to preserve the perceptual important image regions, i.e., the structural and edge regions of objects." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.461, + 0.892, + 0.552 + ], + "angle": 0, + "content": "Ablations of adaptive de-mask module. In Table 7, we show that MQ-VAE outperforms VQGAN and the stronger baseline (\"VQGAN*\"), which validates our effectiveness. We could conclude that the proposed direction-constrained self-attention and the mask updating mechanism both improve the reconstruction and generation quality." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.565, + 0.619, + 0.58 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.59, + 0.892, + 0.801 + ], + "angle": 0, + "content": "In this study, we point out that the existing two-stage autoregressive generation paradigm ignores distinguishing the perceptual importance of different image regions, which brings redundancy that not only degrades generation quality but also decreases generation speed. We propose a novel two-stage generation paradigm with MQ-VAE and Stackformer to relieve the model from redundancy. MQ-VAE incorporates the adaptive mask module to mask redundant region features before quantization and the adaptive de-mask module to recover the original feature map after quantization. Stackformer then efficiently predict the combination of both codes and their positions. Comprehensive experiments on various types of image generation tasks validate the effectiveness and efficiency of our method." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.816, + 0.679, + 0.833 + ], + "angle": 0, + "content": "6. Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.9 + ], + "angle": 0, + "content": "This work is supported by National Natural Science Foundation of China under Grants 62222212 and U19A2057, Science Fund for Creative Research Groups under Grant 62121002." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2009" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.468, + 0.155 + ], + "angle": 0, + "content": "[1] Hangbo Bao, Li Dong, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.158, + 0.468, + 0.185 + ], + "angle": 0, + "content": "[2] Shane Barratt and Rishi Sharma. A note on the inception score. arXiv preprint arXiv:1801.01973, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.187, + 0.468, + 0.227 + ], + "angle": 0, + "content": "[3] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. arXiv preprint arXiv:1809.11096, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.229, + 0.468, + 0.296 + ], + "angle": 0, + "content": "[4] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T Freeman. Maskgit: Masked generative image transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11315-11325, 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.298, + 0.468, + 0.354 + ], + "angle": 0, + "content": "[5] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In International Conference on Machine Learning, pages 1691-1703. PMLR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.355, + 0.468, + 0.395 + ], + "angle": 0, + "content": "[6] Rewon Child. Very deep vaes generalize autoregressive models and can outperform them on images. arXiv preprint arXiv:2011.10650, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.398, + 0.468, + 0.453 + ], + "angle": 0, + "content": "[7] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.455, + 0.468, + 0.508 + ], + "angle": 0, + "content": "[8] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.51, + 0.468, + 0.55 + ], + "angle": 0, + "content": "[9] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.552, + 0.468, + 0.62 + ], + "angle": 0, + "content": "[10] Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, et al. Cogview: Mastering text-to-image generation via transformers. Advances in Neural Information Processing Systems, 34, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.622, + 0.468, + 0.704 + ], + "angle": 0, + "content": "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.706, + 0.468, + 0.773 + ], + "angle": 0, + "content": "[12] Patrick Esser, Robin Rombach, Andreas Blattmann, and Bjorn Ommer. Imagebart: Bidirectional context with multinomial diffusion for autoregressive image synthesis. Advances in Neural Information Processing Systems, 34, 2021. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.468, + 0.843 + ], + "angle": 0, + "content": "[13] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12873-12883, 2021. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[14] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 1, 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.468, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[15] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. arXiv preprint arXiv:2111.14822, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.149, + 0.892, + 0.216 + ], + "angle": 0, + "content": "[16] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.219, + 0.892, + 0.259 + ], + "angle": 0, + "content": "[17] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.261, + 0.892, + 0.328 + ], + "angle": 0, + "content": "[18] Mengqi Huang, Zhendong Mao, Penghui Wang, Quan Wang, and Yongdong Zhang. Dse-gan: Dynamic semantic evolution generative adversarial network for text-to-image generation. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4345-4354, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.33, + 0.892, + 0.397 + ], + "angle": 0, + "content": "[19] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.4, + 0.892, + 0.44 + ], + "angle": 0, + "content": "[20] Nikil Jayant, James Johnston, and Robert Safranek. Signal compression based on models of human perception. Proceedings of the IEEE, 81(10):1385-1422, 1993. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.442, + 0.892, + 0.496 + ], + "angle": 0, + "content": "[21] Biing-Hwang Juang and A Gray. Multiple stage vector quantization for speech coding. In ICASSP'82. IEEE International Conference on Acoustics, Speech, and Signal Processing, volume 7, pages 597-600. IEEE, 1982. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.498, + 0.892, + 0.552 + ], + "angle": 0, + "content": "[22] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.554, + 0.892, + 0.621 + ], + "angle": 0, + "content": "[23] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.623, + 0.892, + 0.649 + ], + "angle": 0, + "content": "[24] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.651, + 0.892, + 0.705 + ], + "angle": 0, + "content": "[25] M Kocher and M Kunt. A contour-texture approach to picture coding. In ICASSP'82. IEEE International Conference on Acoustics, Speech, and Signal Processing, volume 7, pages 436-439. IEEE, 1982. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.707, + 0.892, + 0.747 + ], + "angle": 0, + "content": "[26] Murat Kunt, Athanassios Ikonomopoulos, and Michel Kocher. Second-generation image-coding techniques. Proceedings of the IEEE, 73(4):549-574, 1985. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.749, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[27] Christian Ledig, Lucas Theis, Ferenc Huszár, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photorealistic single image super-resolution using a generative adversarial network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4681-4690, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[28] Doyup Lee, Chiheon Kim, Saehoon Kim, Minsu Cho, and Wook-Shin Han. Autoregressive image generation using residual quantization. arXiv preprint arXiv:2203.01941, 2022. 1, 2, 3, 4, 6, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2010" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[29] Wentong Liao, Kai Hu, Michael Ying Yang, and Bodo Rosenhahn. Text to image generation with semantic-spatial aware gan. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18187-18196, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.47, + 0.233 + ], + "angle": 0, + "content": "[30] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.234, + 0.47, + 0.275 + ], + "angle": 0, + "content": "[31] Julieta Martinez, Holger H Hoos, and James J Little. Stacked quantizers for compositional vector compression. arXiv preprint arXiv:1411.2173, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.277, + 0.47, + 0.318 + ], + "angle": 0, + "content": "[32] Charlie Nash, Jacob Menick, Sander Dieleman, and Peter W Battaglia. Generating images with sparse representations. arXiv preprint arXiv:2103.03841, 2021. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.32, + 0.47, + 0.375 + ], + "angle": 0, + "content": "[33] Niki Parmar, Ashish Vaswani, Jakob Uszkoreit, Lukasz Kaiser, Noam Shazeer, Alexander Ku, and Dustin Tran. Image transformer. In International Conference on Machine Learning, pages 4055-4064. PMLR, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.377, + 0.47, + 0.445 + ], + "angle": 0, + "content": "[34] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.447, + 0.47, + 0.489 + ], + "angle": 0, + "content": "[35] Ali Razavi, Aaron Van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. Advances in neural information processing systems, 32, 2019. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.49, + 0.47, + 0.532 + ], + "angle": 0, + "content": "[36] Danilo Rezende and Shakir Mohamed. Variational inference with normalizing flows. In International conference on machine learning, pages 1530-1538. PMLR, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.533, + 0.47, + 0.587 + ], + "angle": 0, + "content": "[37] Woncheol Shin, Gyubok Lee, Jiyoung Lee, Joonseok Lee, and Edward Choi. Translation-equivariant image quantizer for bi-directional image-text generation. arXiv preprint arXiv:2112.00384, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.59, + 0.47, + 0.645 + ], + "angle": 0, + "content": "[38] Ming Tao, Hao Tang, Songsong Wu, Nicu Sebe, Xiao-Yuan Jing, Fei Wu, and Bingkun Bao. Df-gan: Deep fusion generative adversarial networks for text-to-image synthesis. arXiv preprint arXiv:2008.05865, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.647, + 0.47, + 0.687 + ], + "angle": 0, + "content": "[39] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.689, + 0.47, + 0.744 + ], + "angle": 0, + "content": "[40] Aaron Van Oord, Nal Kalchbrenner, and Koray Kavukcuoglu. Pixel recurrent neural networks. In International conference on machine learning, pages 1747-1756. PMLR, 2016. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.746, + 0.47, + 0.802 + ], + "angle": 0, + "content": "[41] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.47, + 0.871 + ], + "angle": 0, + "content": "[42] Jiahui Yu, Xin Li, Jing Yu Koh, Han Zhang, Ruoming Pang, James Qin, Alexander Ku, Yuanzhong Xu, Jason Baldridge, and Yonghui Wu. Vector-quantized image modeling with improved vqgan. arXiv preprint arXiv:2110.04627, 2021. 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.47, + 0.903 + ], + "angle": 0, + "content": "[43] Han Zhang, Jing Yu Koh, Jason Baldridge, Honglak Lee, and Yinfei Yang. Cross-modal contrastive learning for text-to" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "image generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 833-842, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[44] Chuanxia Zheng, Long Tung Vuong, Jianfei Cai, and Dinh Phung. Movq: Modulating quantized vectors for high-fidelity image generation. arXiv preprint arXiv:2209.09002, 2022.6,7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[45] Minfeng Zhu, Pingbo Pan, Wei Chen, and Yi Yang. Dm-gan: Dynamic memory generative adversarial networks for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5802-5810, 2019. 7" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "2011" + } + ] +] \ No newline at end of file diff --git a/2023/Not All Image Regions Matter_ Masked Vector Quantization for Autoregressive Image Generation/2a88178e-8ce2-4c0f-b3f1-bb99f0452012_origin.pdf b/2023/Not All Image Regions Matter_ Masked Vector Quantization for Autoregressive Image Generation/2a88178e-8ce2-4c0f-b3f1-bb99f0452012_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..78f11327e7ce3f9b1b5d84e10e8426f8472ff275 --- /dev/null +++ b/2023/Not All Image Regions Matter_ Masked Vector Quantization for Autoregressive Image Generation/2a88178e-8ce2-4c0f-b3f1-bb99f0452012_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dc873f62fd8ab0ca683d1c76ae9b3624c5e5f4468c853c6e2d6b69ab8a3e511f +size 5148755 diff --git a/2023/Not All Image Regions Matter_ Masked Vector Quantization for Autoregressive Image Generation/full.md b/2023/Not All Image Regions Matter_ Masked Vector Quantization for Autoregressive Image Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..8103c8197a95dbcd41069be65e680f9094eaf55e --- /dev/null +++ b/2023/Not All Image Regions Matter_ Masked Vector Quantization for Autoregressive Image Generation/full.md @@ -0,0 +1,349 @@ +# Not All Image Regions Matter: Masked Vector Quantization for Autoregressive Image Generation + +Mengqi Huang $^{1}$ , Zhendong Mao $^{1,3*}$ , Quan Wang $^{2}$ , Yongdong Zhang $^{1,3}$ + +1University of Science and Technology of China, Hefei, China; 2MOE Key Laboratory of Trustworthy, Distributed Computing and Service, Beijing University of Posts and Telecommunications, Beijing, China; 3Institute of Artificial intelligence, Hefei Comprehensive National Science Center, Hefei, China + +huangmq@mail.ustc.edu.cn, {zdmao, zhyd73}@ustc.edu.cn, wangquan@bupt.edu.cn + +# Abstract + +Existing autoregressive models follow the two-stage generation paradigm that first learns a codebook in the latent space for image reconstruction and then completes the image generation autoregressively based on the learned codebook. However, existing codebook learning simply models all local region information of images without distinguishing their different perceptual importance, which brings redundancy in the learned codebook that not only limits the next stage's autoregressive model's ability to model important structure but also results in high training cost and slow generation speed. In this study, we borrow the idea of importance perception from classical image coding theory and propose a novel two-stage framework, which consists of Masked Quantization VAE (MQ-VAE) and Stackformer, to relieve the model from modeling redundancy. Specifically, MQ-VAE incorporates an adaptive mask module for masking redundant region features before quantization and an adaptive de-mask module for recovering the original grid image feature map to faithfully reconstruct the original images after quantization. Then, Stackformer learns to predict the combination of the next code and its position in the feature map. Comprehensive experiments on various image generation validate our effectiveness and efficiency. Code will be released at https://github.com/CrossmodalGroup/MaskedVectorQuantization. + +# 1. Introduction + +Deep generative models of images have received significant improvements over the past few years and broadly fall into two categories: likelihood-based models, which include VAEs [24], flow-based [36], diffusion models [17] and autoregressive models [40], and generative adversarial + +![](images/99f379c90a21357ab413c94465c158b648298bb77dcaa5cd4d6f9204332a8f4a.jpg) +(a) Existing methods +Figure 1. Illustration of our motivation. (a) Existing works model all local regions without distinguishing their perceptual importance in stage 1, which not only brings redundancy (e.g., the textural regions like the background) in the learned codebook but also make the autoregressive models overly focus on modeling this redundancy and hinder other important structural regions modeling. (b) The codebook learning in our method only includes the important regions, e.g., the structural regions like corners and edges, since other unimportant ones can be restored even if missing, and thus autoregressive model could focus on modeling these important regions in stage 2 and results in better generation quality. + +![](images/8136d80f5730ddfd445c48546d49198624ae79e616abb9f2cf724b2afc2ec8d4.jpg) +(b) Our method + +networks (GANs) [14], which use discriminator networks to distinguish samples from generator networks and real examples. Compared with GANs, likelihood-based models' training objective, i.e., the negative log-likelihood (NLL) or its upper bound, incentives learning the full data distribution and allows for detecting overfitting. + +Among the likelihood-based models, autoregressive models have recently attracted increasing attention for their impressive modeling ability and scalability. Recent autoregressive image generation [10, 12, 13, 28, 28, 34, 35, 37, 39] follows the two-stage generation paradigm, i.e., the first stage learns a codebook in the latent space for image reconstruction and the second stage completes the image generation in the raster-scan [13] order by autoregressive models + +based on the learned codebook. Since codebook learning in the first stage defines the discrete image representation for the next autoregressive modeling, a high-quality codebook is the key to generate high-quality images. Several recent works focus on improving the codebook learning in the first stage, e.g., VQGAN [13] introduces adversarial loss and perceptual loss. ViT-VQGAN [42] introduces a more expressive transformer backbone. RQ-VAE [28] introduces the residual quantization to reduce the resolution of the latent space. In general, the essence of existing codebook learning is the modeling of all local region information (i.e., an $8 \times 8$ or $16 \times 16$ patch) of images in the dataset, without distinguishing their different perceptual importance. + +In this study, we point out that existing codebook learning exists gaps with classical image coding theory [20, 25, 26], the basic idea of which is to remove redundant information by perceiving the importance of different regions in images. The image coding theory reveals that an ideal image coding method should only encode images' perceptually important regions (i.e., which cannot be restored if missing) while discarding the unimportant ones (i.e., which can be restored by other image regions even if missing). The neglect of considering such perceptual importance in existing works poses problems in two aspects, as illustrated in Figure 1(a): (1) the existence of this large amount of repetitive and redundant information brings redundancy to the learned codebook, which further makes the autoregressive model in the next stage overly focus on modeling this redundancy while overlooking other important regions and finally degrades generation quality. (2) the redundancy makes the autoregressive model need to predict more (redundant) quantized codes to generate images, which significantly increases the training cost and decreases the generating speed. Although the effectiveness and efficiency of image coding theory have been widely validated, how to introduce this idea into codebook learning remains unexplored. + +The key of applying image coding theory to codebook learning is to distinguish important image parts from unimportant ones correctly. Considering that the essential difference between these two sets lies in whether they can be restored if missing, we found that this distinction can be realized through the mask mechanism, i.e., the masked part is important if it cannot be faithfully restored, and otherwise unimportant. Based on the above observation, we thereby propose a novel two-stage generation paradigm upon the mask mechanism to relieve the model from modeling redundant information. Specifically, we first propose a Masked Quantization VAE (MQ-VAE) with two novel modules, i.e., an adaptive mask module for adaptively masking redundant region features before quantization, and an adaptive de-mask module for adaptively recovering the original grid image feature map to faithfully reconstruct original images after quantization. As for the adaptive mask module, it in + +corporates a lightweight content-aware scoring network that learns to measure the importance of each image region feature. The features are then ranked by the importance scores and only a subset of high-scored features will be quantized further. As for the adaptive de-mask module, we design a direction-constrained self-attention to encourage the information flow from the unmasked regions to the masked regions while blocking the reverse, which aims to infer the original masked region information based on unmasked ones. Thanks to the adaptive mask and de-mask mechanism, our MQ-VAE removes the negative effects of redundant image regions and also shortens the sequence length to achieve both effectiveness and efficiency. + +Moreover, since different images have different important regions, the position of quantized codes in the feature map also dynamically changed. Therefore, we further propose Stackformer for learning to predict the combination of both codes and their corresponding positions. Concretely, the proposed Stackformer stacks a Code-Transformer and a Position-Transformer, where the Code-Transformer learns to predict the next code based on all previous codes and their positions, and the Position-Transformer learns to predict the next code's position based on all previous codes' positions and current code. + +With our method, as shown in Figure 1(b), the codebook learning only includes the important regions, e.g., the structural regions, since unimportant ones like the background can be restored even if missing. And therefore the autoregressive model in the second stage could focus on modeling these important regions and brings better generation quality. + +In a nutshell, we summarize our main contributions as: + +Conceptually, we point out that existing codebook learning ignores distinguishing the perceptual importance of different image regions, which brings redundancy that degrades generation quality and decreases generation speed. + +Technically, (i) we propose MQ-VAE with a novel adaptive mask module to mask redundant region features before quantization and a novel adaptive de-mask module to recover the original feature map after quantization; (ii) we propose a novel Stackformer to predict the combination of both codes and their corresponding positions. + +Experimentally, comprehensive experiments on various generations validate our effectiveness and efficiency, i.e., we achieve $8.1\%$ , $2.3\%$ , and $18.6\%$ FID improvement on un-, class-, and text-conditional state-of-the-art at million-level parameters, and faster generation speed compared to existing autoregressive models. + +# 2. Related Work + +# 2.1. Autoregressive Modeling for Image Generation + +Autoregressive models for image generation have recently attracted increasing research attention and have + +![](images/27735186e03b9dfe29d96d79864a9d9ed4bb3b557ec51166fa8b9afadc72e871.jpg) +Figure 2. The Illustration of our proposed two-stage generation framework. (a) In the first stage, MQ-VAE adaptively masks the redundant region features to prevent redundant codes while keeping important ones, which ensures that the original images can still be faithfully recovered. Here, $\frac{10}{16}$ regions are masked and $\frac{6}{16}$ regions are kept. (b) In the second stage, Stackformer stacks a Code-Transformer and a Position-Transformer to autoregressively predict the next code and its position in the original 2D feature map, respectively. (c) The attention mask of the proposed direction-constrained self-attention in the adaptive de-mask module for inferring masked regions features. + +shown impressive results [10, 12, 13, 32, 34, 35, 37, 39, 42] among various generation tasks. Early autoregressive models [5, 33, 40] directly optimizing the likelihood of raw image pixels, e.g., Image-GPT [5] trains a transformer [41] to autoregressively predict pixels' cluster centroids, which could only generate images with a maximum resolution of $64 \times 64$ . [39] presents the Vector Quantized Variational Autoencoder (VQVAE), which learns images' low-dimension discrete representation and models their distribution autoregressively. VQ-VAE2 [35] extends this approach using a hierarchy of discrete representations. VQGAN [13] further improves the perceptual quality of reconstructed images using adversarial [14, 19] and perceptual loss [27]. ViT-VQGAN [42] introduces a more expressive transformer backbone. RQ-VAE [28] uses Residual Quantization [21, 31] to iteratively quantizes a vector and its residuals and represent the vector as a stack of tokens. Although vector quantization has become the fundamental technique for modern visual autoregressive models, the critical removing redundancy in codebook learning has not been explored yet, which becomes a critical bottleneck. + +# 2.2. Masked Modeling + +Masked modeling is popular among both natural language processing and computer vision. BERT [8] randomly masks a portion of the input sequence and trains models to predict the missing content. In the computer vision domain, the ViT [11] studies masked patch prediction for self-supervised learning. BEiT [1] proposes to predict discrete tokens. Most recently, MaskGIT [4] also used the masking strategy for VQ-based image generation. However, our proposed method differs from MaskGIT in two aspects: (1) Our primary motivation for the masking strategy applied in the proposed MQ-VAE in the first stage aims to learn a more compact and effective vector quantization (VQ) itself by masking perceptual unimportant regions, while MaskGIT uses masking strategy in the second stage to better use a learned VQ. (2) The mask in our proposed MQ-VAE is learned and adaptively changed according to different image content, while the mask in MaskGIT is randomly sampled for the mask-and-predict training. In conclusion, to the best of our knowledge, this is the first time that masked modeling has been applied for vector quantization. + +# 3. Methodology + +We propose a novel two-stage framework with MQ-VAE and Stackformer for autoregressive image generation, as illustrated in Figure 2. MQ-VAE only masks redundant region features to prevent redundant codes and Stackformer stacks two transformers to autoregressively predict the next code and its position. In the following, we will first briefly revisit the formulation of vector quantization and then describe our proposed method in detail. + +# 3.1. Preliminary + +We follow the definition and notations of previous works [13, 28]. Specifically, the codebook $\mathcal{C} \coloneqq \{(k, e(k))\}_{k \in [K]}$ is defined as the set of finite pairs of code $k$ and its code embedding $e(k) \in \mathbb{R}^{n_z}$ . Here $K$ is the codebook size and $n_z$ is the code dimension. An image $\mathbf{X} \in \mathbb{R}^{H_0 \times W_0 \times 3}$ is first encoded into grid features $\mathbf{Z} = E(\mathbf{X}) \in \mathbb{R}^{H \times W \times n_z}$ by the encoder $E$ , where $(H, W) = (H_0 / f, W_0 / f)$ and $f$ is the corresponding downsampling factor. For each vector $z \in \mathbb{R}^{n_z}$ in $\mathbf{Z}$ , it is replaced with the code embedding that has the closest euclidean distance with it in the codebook $\mathcal{C}$ through the vector quantization operation $\mathcal{Q}(\cdot)$ : + +$$ +\mathcal {Q} (z; \mathcal {C}) = \arg \min _ {k \in [ K ]} | | z - e _ {k} | | _ {2} ^ {2}. \tag {1} +$$ + +Here, $\mathcal{Q}(z;\mathcal{C})$ is the quantized code. $z^q = e(\mathcal{Q}(z;\mathcal{C}))$ is the quantized vector. By applying $\mathcal{Q}(\cdot)$ to each feature vector, we could get the quantized code map $\mathbf{M}\in [K]^{H\times W}$ and the quantized features $\mathbf{Z}^q\in \mathbb{R}^{H\times W\times n_z}$ . The original image is reconstructed by the decoder $D$ as $\tilde{\mathbf{X}} = D(\mathbf{Z}^q)$ . + +# 3.2. Stage 1: MQ-VAE + +Existing methods quantize each feature vector of $\mathbf{Z}$ without distinguishing their different perceptual importance and thus bring redundancy in the learned codebook, which not only degrades the generation quality but also decreases the generation speed. To relieve the model from this redundancy, we propose MQ-VAE with two novel modules, i.e., the adaptive mask module for adaptively masking redundant region features before vector quantization and adaptive de-mask module for adaptively recovering the original grid image feature map after vector quantization. + +Adaptive Mask Module. The encoded grid feature map $\mathbf{Z} \in \mathbb{R}^{H \times W \times n_e}$ is first flattened into $\mathbf{Z} \in \mathbb{R}^{L \times n_e}$ , where $L = H \times W$ . The proposed adaptive mask module then uses a lightweight scoring network $f_s$ to measure the importance of each region feature $z_l$ in $\mathbf{Z}$ , which is implemented as a two-layer MLP: + +$$ +s _ {l} = f _ {s} \left(\boldsymbol {z} _ {l}\right), l = 1, \dots , L. \tag {2} +$$ + +The larger score $s_l$ is, the more important the region feature $z_l$ is. Then the region features are sorted in descending + +order according to the predicted scores. The sorted region features and their scores are denoted as $\{z_l^{\prime}\}$ and $\{s_l^{\prime}\}$ respectively, where $l = 1,\dots ,L$ . To enable the learning of $f_{s}$ , the predicted scores are further multiplied with the normalized region features as modulating factors. We select the top $N$ scoring vectors as the important region features, + +$$ +\hat {\mathbf {Z}} = \left\{z _ {l} ^ {\prime \prime} \mid z _ {l} ^ {\prime \prime} = \operatorname {L a y e r N o r m} \left(z _ {l} ^ {\prime}\right) * s _ {l} ^ {\prime} \right\}, l = 1, \dots , N. \tag {3} +$$ + +$$ +\hat {\mathbf {P}} = \left\{p _ {z _ {l} ^ {\prime \prime}} \mid p _ {z _ {l} ^ {\prime \prime}} \in \{0, \dots , L \} \right\}, l = 1, \dots , N. \tag {4} +$$ + +Here, $\hat{\mathbf{Z}}$ denotes the selected important region features set, and $\hat{\mathbf{P}}$ denotes the corresponding position set that represents the position of each selected region feature in the original 2D feature map. The selected number $N = \alpha \times L$ , where $\alpha$ is a constant fractional value. The mask ratio is defined as $1 - \alpha$ . This design also enables a flexible trade-off between the image generation speed and image generation quality, which we will discuss in experiments. After obtaining $\hat{\mathbf{Z}}$ , we further apply the quantization function $\mathcal{Q}$ to each of them and obtain the quantized important region features set $\hat{\mathbf{Z}}^q$ as well as its code matrix $\hat{\mathbf{M}}$ . + +Adaptive De-mask Module. After quantization, we fill the quantized features $\hat{\mathbf{Z}}^q$ back into the original 2D feature map according to $\hat{\mathbf{P}}$ , while other masked positions are filled with a uniformly initialized learnable mask code embedding, as shown in Figure 2(a). Directly inputting filled grid features to the decoder $D$ could bring sub-optimal reconstruction results since the mask code embedding here only serves as the placeholders that contain little information. Therefore, we further propose the adaptive de-mask module, which applies a novel direction-constrained self-attention to encourage the information flow from unmasked regions to the masked ones while blocking the reverse. Such a design allows the model to utilize the unmasked region features to infer the masked ones while also preventing the masked regions to have negative impacts on the unmasked ones since they are less informative. + +Our adaptive de-mask module is implemented as $H$ identical sub-modules, where each consists of a direction-constrained self-attention block and a Resnet block. The direction-constrained self-attention is mathematically formed as (Resnet block is omitted for simplicity): + +$$ +\boldsymbol {q}, \boldsymbol {k}, \boldsymbol {v} = \boldsymbol {W} _ {q} \hat {\boldsymbol {Z}} ^ {q, h}, \boldsymbol {W} _ {k} \hat {\boldsymbol {Z}} ^ {q, h}, \boldsymbol {W} _ {v} \hat {\boldsymbol {Z}} ^ {q, h} \tag {5} +$$ + +$$ +\boldsymbol {A} = \left(\operatorname {S o f t M a x} \left(\frac {\boldsymbol {q} \boldsymbol {k} ^ {T}}{\sqrt {n _ {e}}}\right)\right) \odot \boldsymbol {B} _ {\boldsymbol {h}} \tag {6} +$$ + +$$ +\hat {\mathbf {Z}} ^ {q, h + 1} = \boldsymbol {A} \boldsymbol {v}. \tag {7} +$$ + +Here $h\in \{1,\dots ,H\}$ and $\hat{\mathbf{Z}}^{q,1}$ is the filled quantized grid features. $\pmb {W}_q,\pmb {W}_k,\pmb {W}_v\in \mathbb{R}^{n_e\times n_e}$ are the learnable parameters. $\pmb{B}_{h}$ is the attention mask at $h$ sub-module. Specifically, since the initial mask code contains little information, + +we define $B_{1} = [b_{l}\in \{0,1\} ,|l = 1,\dots,L]\in \mathbb{R}^{1\times L}$ to forbid it from attending to other unmasked codes to avoid negative impact, where 0 for the mask position and 1 for the unmasked. Considering that the mask code is updated with more and more information after each sub-module, we propose to synchronously amplify its interaction with other codes step by step through a mask updating mechanism: + +$$ +\boldsymbol {B} _ {h + 1} = \sqrt {\boldsymbol {B} _ {h}}, \tag {8} +$$ + +where the initial 0 in $B_{1}$ is replaced with a small number 0.02, since $\sqrt{0}$ will always result in 0. Finally, the original image is recovered by the decoder as $\tilde{\mathbf{X}} = D(\hat{\mathbf{Z}}^{q,H})$ . + +# 3.3. Stage 2: Stackformer + +The perceptual important regions of different images vary. Therefore the positions of quantized codes in the feature map also dynamically change along with the image content. As a result, our proposed MQ-VAE formulates an image as both the code sequence $\hat{\mathbf{M}}$ and the code position sequence $\hat{\mathbf{P}}$ . To simultaneously learn the combination of the codes and their positions, we propose a novel Stackformer, which stacks a Code-Transformer and a Position-Transformer. The Code-Transformer learns to predict the next code based on all previous steps' codes and their positions, while the Position-Transformer learns to predict the next code's position based on all previous steps' positions and current code. Directly treating the importance descending order sequence $\hat{\mathbf{M}}$ and $\hat{\mathbf{P}}$ as the inputs are natural, but the dramatic position changes of adjacent code could make the network hard to converge. For example, the position of the first code may be in the upper left corner of the image, while the position of the second code may be in the lower right corner of the image. Therefore, we further propose to use the raster-scan order [13] to rearrange both sequences to deal with the converge problem. + +Mathematically, taking the raster-scan code and code position sequence $(\overline{\mathbf{M}},\overline{\mathbf{P}}) =$ rearrange $(\hat{\mathbf{M}},\hat{\mathbf{P}})$ , Stackformer learns $p(\overline{\mathbf{M}},\overline{\mathbf{P}})$ , which is autoregressively factorized as: + +$$ +p (\overline {{\mathbf {M}}}, \overline {{\mathbf {P}}}) = \prod_ {l = 1} ^ {N} p (\overline {{\mathbf {M}}} _ {l} | \overline {{\mathbf {M}}} _ {< l}, \overline {{\mathbf {P}}} _ {< l}) p (\overline {{\mathbf {P}}} _ {l} | \overline {{\mathbf {M}}} _ {\leq l}, \overline {{\mathbf {P}}} _ {< l}) \quad (9) +$$ + +Code-Transformer takes the sum of code embeddings $e_{c}(\cdot)$ and code position embedding $e_{p}(\cdot)$ as inputs: + +$$ +\mathbf {U} _ {c} = \boldsymbol {e} _ {c} \left(\overline {{\mathbf {M}}} _ {[ 1: N _ {c} + N ]}\right) + \boldsymbol {e} _ {p 1} \left(\overline {{\mathbf {P}}} _ {[ 1: N _ {c} + N ]}\right), \tag {10} +$$ + +where $N_{c}$ is the condition length. For the unconditional generation, we add a $<\mathrm{sos}>$ code at the start of the code and code position sequence. For conditioning, we append class or text codes to the start of the code sequence and the same length of $<\mathrm{sos}>$ code to the code position sequence. We + +further add an extra learned absolute position embedding to $\mathbf{U}_c$ to form the final input, which makes the network aware of the absolute position of the sequence. After processing by Code-Transformer, the output hidden vector $\mathbf{H}_c$ encodes both code and their position information and is used for the next code prediction. The negative log-likelihood (NLL) loss for code autoregressive training is: + +$$ +\mathcal {L} _ {\mathrm {c o d e}} = \mathbb {E} \left[ - \log p \left(\overline {{\mathbf {M}}} _ {l} \mid \overline {{\mathbf {M}}} _ {< l}, \overline {{\mathbf {P}}} _ {< l} \right]\right). \tag {11} +$$ + +Position-Transformer takes the sum of CodeTransformer's output hidden vector $\mathbf{H}_c$ and an extra code embedding as input: + +$$ +\mathbf {U} _ {p} = \mathbf {H} _ {c} \left[ N _ {c}: N _ {c} + N - 1 \right] + \boldsymbol {e} _ {c} \left(\overline {{\mathbf {M}}} _ {\left[ N _ {c} + 1: N _ {c} + N \right]}\right). \tag {12} +$$ + +Here $\mathbf{U}_p$ is the input for Position-Transformer and the information of current code is included in $e_c(\overline{\mathbf{M}}_{[N_c + 1:N_c + N]})$ . The design idea behind this is that when predicting the next code's position, the model should not only be aware of previous steps' codes and their position information but also should be aware of current code information. The negative log-likelihood (NLL) for position autoregressive training is: + +$$ +\mathcal {L} _ {\text {p o s i t i o n}} = \mathbb {E} \left[ - \log p \left(\overline {{\mathbf {P}}} _ {l} \mid \overline {{\mathbf {M}}} _ {\leq l}, \overline {{\mathbf {P}}} _ {< l}\right) \right]. \tag {13} +$$ + +Training & Inference. The total loss for training Stackformer is defined as: + +$$ +\mathcal {L} = \mathcal {L} _ {\text {c o d e}} + \mathcal {L} _ {\text {p o s i t i o n}}. \tag {14} +$$ + +The inference procedure is illustrated in Algorithm 1, where we take the unconditional generation as an example and conditional generation can be derived accordingly. + +# Algorithm 1 Unconditional sampling of Stackformer. + +Input: The sample step $N$ ; + +The code sequence $M_{\text{sample}}$ of a single $<\text{sos}>$ ; + +The code position sequence $P_{\text{sample}}$ of a single $<\text{sos}>$ . + +Output: The generated image $\mathcal{I}$ . + +1: for each $n \in [1, N]$ do +2: $H_{c} =$ Code-Transformer(Msample, $P_{\text{sample}}$ ); +3: Sample next code $M_{n}$ by $H_{c}$ ; +4: $M_{\text{sample}} = \operatorname{concat}(M_{\text{sample}}, M_n)$ ; +5: $H_{p} = \mathrm{Position - Transformer}(H_{c},M_{\mathrm{sample}}[2:])$ +6: Mask already sampled positions to avoid conflicts; +7: Sample next code position $P_{n}$ by $H_{p}$ ; +8: $P_{\text{sample}} = \operatorname{concat}(P_{\text{sample}}, P_n)$ ; +9: end for +10: Re-map $M_{\text{sample}}$ to 2D code map according to $P_{\text{sample}}$ and the rest are filled with the mask code; +11: Decode the code map to the image $\mathcal{I}$ +12: return The generated image $\mathcal{I}$ + +goldfish, Carassius auratus +![](images/d642cc09f5748e1d217765a7f04920d0f50e2599648c0ccd33a5e27be05d3a92.jpg) +common iguna, iguana, Iguania iguna + +![](images/bf756d23b791a541459c31aca5ae0bd7b40c47c40b7268dea3a9b82b6334bc3b.jpg) +Figure 3. Left: Our unconditional generated images on FFHQ benchmark. Right: Our class-conditional generated images on ImageNet. + +![](images/072aba8ed82ac3f6b31ed191945af875d1a622daa5b2ff4298f6a453a2b6f98d.jpg) +lorikeet curly-coated retriever + +# 4. Experiment + +# 4.1. Experimental Settings + +**Benchmarks.** We validate our model for unconditional, class-conditional, and text-conditional image generation tasks on FFHQ [22], ImageNet [7], and MS-COCO [30] benchmarks respectively, with $256 \times 256$ image resolution. + +Metrics. Following previous works [13,28,42], the standard Frechet Inception Distance (FID) [16] is adopted for evaluating the generation and reconstruction quality (denoted as rFID). Inception Score (IS) [2] is also used for class-conditional generation on the ImageNet benchmark. FID and IS are calculated by sampling 50k images. rFID is calculated over the entire test set. + +Implemented details. The architecture of MQ-VAE exactly follows [13] except for the proposed mask and demask modules, with the codebook size of $K = 1024$ . For the de-mask module, the sub-module number $H = 8$ . For the Stackformer, we implement two settings: a small version uses 18 transformer encoder blocks for the CodeTransformer and another 6 transformer encoder blocks for the Position-Transformer with a total of 307M parameters, and a base version uses 36 transformer encoder blocks for the Code-Transformer and another 12 transformer encoder blocks for the Position-Transformer with a total of 607M parameters to further demonstrate our scalability. The generation results are reported with a $25\%$ mask ratio at $32 \times 32$ resolution feature map using eight RTX-3090 GPUs. Top-k and top-p sampling are used to report the best performance. More details could be found in the supplementary. + +# 4.2. Comparison with state-of-the-art methods + +Unconditional generation. We first compare with million-level state-of-the-art autoregressive models in Table 1. Our model significantly outperforms other autoregressive models with the same parameters (307M). With more parameters, we further increase the FID from 6.84 to 5.67, which demonstrates our scalability. We also compare with other types of unconditional state-of-the-art and large-scale big models in Table 3, where we also achieve top-level performance. Our qualitative unconditional generation results are shown on the left of Figure 3. + +
Methods#ParamsFID↓
DCT [32]738M13.06
VQGAN [13](72.1+307)M11.4
RQ-Transformer [28](100+355)M10.38
Mo-VQGAN [44](82.7+307)M8.52
Stackformer(44.4+307)M6.84
Stackformer(44.4+607)M5.67
+ +Table 1. Comparison of autoregressive unconditional generation at million-level parameters on FFHQ [22] benchmark. #Params splits in (VAE + autoregressive model). + +
Model TypeMethods#ParamsFID↓IS↑
DiffusionADM [9]554M10.94101.0
GANBigGAN [3]164M7.53168.6
GANBigGAN-deep [3]112M6.84203.6
BidirectionMaskGIT [4]227M6.18182.1
AutoregressiveDCT [32]738M36.5n/a
AutoregressiveVQ-GAN† [13]679M17.0376.85
AutoregressiveRQ-Transformer [28]821M13.11104.3
AutoregressiveMo-VQGAN [44]383M7.13138.3
AutoregressiveStackformer651M6.04172.6
+ +Table 2. Comparison of class-conditional image generation at million-level parameters without rejection sampling on ImageNet [7]. $\dagger$ denotes the model we train with the same setup with ours. + +
Model TypeMethods#ParamsFID↓
VAEVDVAE [6]115M28.5
DiffusionImageBART [12]3.5B9.57
GANStyleGAN2 [23]-3.8
GANBigGAN [3]164M12.4
AutoregressiveViT-VQGAN [42]2.2B5.3
AutoregressiveStackformer651M5.67
+ +Table 3. Comparison with other types of state-of-the-art generative models and large-scale billion-level parameters autoregressive models on unconditional FFHQ [22] benchmark. + +Class-conditional generation. We first compare with million-level state-of-the-art in Table 2. We achieve the best FID score compared to all types of models including the re + +![](images/a07faec2fbeb3939e0a08df57afd59a032a863a4f9c4f0134bec02159a88153f.jpg) +Figure 4. The visualization of our adaptive mask module which learns to mask unimportant regions on ImageNet [7]. In the importance map, red denotes high scores while blue denotes low scores. + +![](images/24e037fbf3b3c191e4aff33a020cddf85ea302eb6c003376e6ead38253a73968.jpg) + +
Model TypeMethods#ParamsFID↓IS↑
DiffusionImageBART [12]3.5B21.1961.6
AutoregressiveVQVAE2 [35]13.5B3145
AutoregressiveVQ-GAN [13]1.4B15.7878.3
AutoregressiveViT-VQGAN [42]2.2B4.17175.1
AutoregressiveRQ-Transformer [28]3.8B7.55134
AutoregressiveStackformer651M6.04172.6
+ +Table 4. Comparison between our model and large-scale billion-level parameters models of class-conditional generation without rejection sampling on ImageNet [7] benchmark. + +
Model TypeMethodFID↓
GANDMGAN [45]32.64
GANXMCGAN† [43]50.08
GANDFGAN [38]21.42
GANSSA-GAN [29]19.37
GANDSE-GAN [18]15.30
DiffusionVQ-Diffusion [15]19.75
AutoregressiveVQ-GAN† [13]22.28
AutoregressiveStackformer10.08
+ +cent Mo-VQGAN [44] and RQ-Transformer [28]. We also compare our million-level model with existing billion-level big models in Table 4, where we also achieve top performance with fewer parameters and is only inferior to ViT-VQGAN big model. Our qualitative class-conditional generation results are shown on the right of Figure 3. + +Text-conditional generation. We compare with existing text-conditional state-of-the-art without extra web-scale + +Table 5. Comparison of text-conditional generation on MS-COCO [30] without using extra web-scale data or pre-trained models. † denotes reproduced results under our same experimental setting. + +
mask ratiofmask typerFID↓FID↓usage↑ (%)
0%32-8.113.570.02
0%16-4.4611.463.89
10%16adaptive4.557.8172.34
25%16adaptive4.797.6778.22
25%16random6.1312.2167.48
50%16adaptive5.318.3684.29
50%16random7.85515.6769.04
75%16adaptive7.6211.7187.60
75%16random10.5817.6270.41
+ +Table 6. Ablations of adaptive mask module on FFHQ. Here $f$ is the downsampling factor. The codebook usage is calculated as the percentage of used codes over the entire test dataset. + +data or pretrained models on MS-COCO [30] for fair comparison in Table.5. We achieve $18.6\%$ FID improvement. + +![](images/ab2290e9ed4f0135bfecae58bf9f3a4ae106fef5e63f31e24865f5a475bfa0e7.jpg) +Figure 5. The PCA of learned codebook (1024 codebook size). + +# 4.3. Ablations + +We conduct ablations on $16 \times 16$ resolution feature map using four RTX-3090 GPUs for saving computation resources and the experimental trends are all consistent with $32 \times 32$ resolution feature map of the main results. + +Ablations of adaptive mask module. As shown in Table 6, our proposed learned adaptive mask mechanism sig- + +![](images/8e5d0ff3c2401bbac153b064e0c2a2cdad968303d2db7b3843890b39e81654ae.jpg) + +![](images/f64343d27867bc681c2c5f8e5ef7b145c4133d0bbdf180b79fd8275f6c796003.jpg) + +![](images/d5458069c96a0984b11dafb0fee346aa576ad4748396de6b9a3f0bc4cbc7f961.jpg) +(a) Training and validation curves of VQGAN and our Stackformer +(b) The sampling speed of Stackformer and VQGAN according to batch size and mask radio +Figure 6. Comparison of training & validation curves and sample speed between VQGAN [13] and Stackformer. + +nificantly outperforms the random one, which quantitatively validates that the adaptive mask module enables the model to learn perceptually important regions. + +Our model with $10\%$ and $25\%$ mask radio has only slightly inferior reconstruction compared with VQGAN [13], but achieves a significant improvement in generation quality, which indicates that the effectiveness of focusing autoregressive models on modeling important regions. When we further increase the mask radio to $50\%$ and $75\%$ , the final generation quality drops, we believe the reason lies that an improper high mask radio will inevitably mask some important regions that greatly decrease the reconstruction results and hinder autoregressive modeling. + +The redundancy of the existing learned codebook can be verified from two aspects: i) the PCA of the learned codebook in Figure 5, where each point is a code and closer codes have more similar semantics. We show many codes in VQGAN's codebook overlap, which indicates these codes have nearly the same semantics and are thus redundant. The redundancy increase (more overlaps) when VQGAN uses more code to represent images (smaller downsampling factor $f$ ). The redundancy is largely alleviated in our MQ-VAE. ii) in Table 6, a higher codebook usage indicates more "useful" codes in the codebook and thus less redundant. VQGAN has a lower usage compared with our MQ-VAE. + +We visualize the training and validation curves of VQGAN and Stackformer in Figure 6(a). Previous autoregressive models [13, 28, 42] always suffer from the overfitting problem while our Stackformer successfully gets rid of it, which indicates the better generalization of our masked discrete representation and our model. + +
Model settingrFID↓FID↓
VQGAN4.4611.4
VQGAN*4.1711.02
MQ-VAE w/o de-mask7.0210.74
MQ-VAE w/ de-mask (SA)6.569.8
MQ-VAE w/ de-mask (DC-SA w/o mask update)5.848.92
MQ-VAE w/ de-mask (DC-SA w/ mask update)5.318.36
+ +Table 7. Ablations of adaptive de-mask module on FFHQ. SA for self-attention and DC-SA for direction-constrained self-attention. "VQGAN*" is the stronger baseline, where the same numbers of SA and Resnet blocks as MQ-VAE's de-mask module are added. + +We compare the sampling speed on a single RTX-1080Ti in Figure 6(b). Compared with VQGAN, our $25\%$ mask radio model achieves $32.72\%$ quality improvement and $15.45\%$ speed improvement, while Our $50\%$ mask radio model achieves $26.67\%$ quality improvement and $61.1\%$ speed improvement. Therefore, our design enables a flexible trade-off between speed and quality. + +Finally, We visualize the learned mask in Figure 4, with a $75\%$ mask ratio on $32 \times 32$ resolution feature map, which validates that our proposed adaptive mask mechanism successfully learns to preserve the perceptual important image regions, i.e., the structural and edge regions of objects. + +Ablations of adaptive de-mask module. In Table 7, we show that MQ-VAE outperforms VQGAN and the stronger baseline ("VQGAN*"), which validates our effectiveness. We could conclude that the proposed direction-constrained self-attention and the mask updating mechanism both improve the reconstruction and generation quality. + +# 5. Conclusion + +In this study, we point out that the existing two-stage autoregressive generation paradigm ignores distinguishing the perceptual importance of different image regions, which brings redundancy that not only degrades generation quality but also decreases generation speed. We propose a novel two-stage generation paradigm with MQ-VAE and Stackformer to relieve the model from redundancy. MQ-VAE incorporates the adaptive mask module to mask redundant region features before quantization and the adaptive de-mask module to recover the original feature map after quantization. Stackformer then efficiently predict the combination of both codes and their positions. Comprehensive experiments on various types of image generation tasks validate the effectiveness and efficiency of our method. + +# 6. Acknowledgments + +This work is supported by National Natural Science Foundation of China under Grants 62222212 and U19A2057, Science Fund for Creative Research Groups under Grant 62121002. + +# References + +[1] Hangbo Bao, Li Dong, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021.3 +[2] Shane Barratt and Rishi Sharma. A note on the inception score. arXiv preprint arXiv:1801.01973, 2018. 6 +[3] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. arXiv preprint arXiv:1809.11096, 2018. 6 +[4] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T Freeman. Maskgit: Masked generative image transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11315-11325, 2022. 3, 6 +[5] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In International Conference on Machine Learning, pages 1691-1703. PMLR, 2020. 3 +[6] Rewon Child. Very deep vaes generalize autoregressive models and can outperform them on images. arXiv preprint arXiv:2011.10650, 2020. 6 +[7] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6, 7 +[8] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 3 +[9] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34, 2021. 6 +[10] Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, et al. Cogview: Mastering text-to-image generation via transformers. Advances in Neural Information Processing Systems, 34, 2021. 1, 3 +[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.3 +[12] Patrick Esser, Robin Rombach, Andreas Blattmann, and Bjorn Ommer. Imagebart: Bidirectional context with multinomial diffusion for autoregressive image synthesis. Advances in Neural Information Processing Systems, 34, 2021. 1, 3, 6, 7 +[13] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12873-12883, 2021. 1, 2, 3, 5, 6, 7, 8 +[14] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 1, 3 + +[15] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. arXiv preprint arXiv:2111.14822, 2021. 7 +[16] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6 +[17] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 1 +[18] Mengqi Huang, Zhendong Mao, Penghui Wang, Quan Wang, and Yongdong Zhang. Dse-gan: Dynamic semantic evolution generative adversarial network for text-to-image generation. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4345-4354, 2022. 7 +[19] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 3 +[20] Nikil Jayant, James Johnston, and Robert Safranek. Signal compression based on models of human perception. Proceedings of the IEEE, 81(10):1385-1422, 1993. 2 +[21] Biing-Hwang Juang and A Gray. Multiple stage vector quantization for speech coding. In ICASSP'82. IEEE International Conference on Acoustics, Speech, and Signal Processing, volume 7, pages 597-600. IEEE, 1982. 3 +[22] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 6 +[23] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. 6 +[24] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 1 +[25] M Kocher and M Kunt. A contour-texture approach to picture coding. In ICASSP'82. IEEE International Conference on Acoustics, Speech, and Signal Processing, volume 7, pages 436-439. IEEE, 1982. 2 +[26] Murat Kunt, Athanassios Ikonomopoulos, and Michel Kocher. Second-generation image-coding techniques. Proceedings of the IEEE, 73(4):549-574, 1985. 2 +[27] Christian Ledig, Lucas Theis, Ferenc Huszár, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photorealistic single image super-resolution using a generative adversarial network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4681-4690, 2017. 3 +[28] Doyup Lee, Chiheon Kim, Saehoon Kim, Minsu Cho, and Wook-Shin Han. Autoregressive image generation using residual quantization. arXiv preprint arXiv:2203.01941, 2022. 1, 2, 3, 4, 6, 7, 8 + +[29] Wentong Liao, Kai Hu, Michael Ying Yang, and Bodo Rosenhahn. Text to image generation with semantic-spatial aware gan. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18187-18196, 2022. 7 +[30] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 6, 7 +[31] Julieta Martinez, Holger H Hoos, and James J Little. Stacked quantizers for compositional vector compression. arXiv preprint arXiv:1411.2173, 2014. 3 +[32] Charlie Nash, Jacob Menick, Sander Dieleman, and Peter W Battaglia. Generating images with sparse representations. arXiv preprint arXiv:2103.03841, 2021. 3, 6 +[33] Niki Parmar, Ashish Vaswani, Jakob Uszkoreit, Lukasz Kaiser, Noam Shazeer, Alexander Ku, and Dustin Tran. Image transformer. In International Conference on Machine Learning, pages 4055-4064. PMLR, 2018. 3 +[34] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 1, 3 +[35] Ali Razavi, Aaron Van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. Advances in neural information processing systems, 32, 2019. 1, 3, 7 +[36] Danilo Rezende and Shakir Mohamed. Variational inference with normalizing flows. In International conference on machine learning, pages 1530-1538. PMLR, 2015. 1 +[37] Woncheol Shin, Gyubok Lee, Jiyoung Lee, Joonseok Lee, and Edward Choi. Translation-equivariant image quantizer for bi-directional image-text generation. arXiv preprint arXiv:2112.00384, 2021. 1, 3 +[38] Ming Tao, Hao Tang, Songsong Wu, Nicu Sebe, Xiao-Yuan Jing, Fei Wu, and Bingkun Bao. Df-gan: Deep fusion generative adversarial networks for text-to-image synthesis. arXiv preprint arXiv:2008.05865, 2020. 7 +[39] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 1, 3 +[40] Aaron Van Oord, Nal Kalchbrenner, and Koray Kavukcuoglu. Pixel recurrent neural networks. In International conference on machine learning, pages 1747-1756. PMLR, 2016. 1, 3 +[41] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3 +[42] Jiahui Yu, Xin Li, Jing Yu Koh, Han Zhang, Ruoming Pang, James Qin, Alexander Ku, Yuanzhong Xu, Jason Baldridge, and Yonghui Wu. Vector-quantized image modeling with improved vqgan. arXiv preprint arXiv:2110.04627, 2021. 2, 3, 6, 7, 8 +[43] Han Zhang, Jing Yu Koh, Jason Baldridge, Honglak Lee, and Yinfei Yang. Cross-modal contrastive learning for text-to + +image generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 833-842, 2021. 7 +[44] Chuanxia Zheng, Long Tung Vuong, Jianfei Cai, and Dinh Phung. Movq: Modulating quantized vectors for high-fidelity image generation. arXiv preprint arXiv:2209.09002, 2022.6,7 +[45] Minfeng Zhu, Pingbo Pan, Wei Chen, and Yi Yang. Dm-gan: Dynamic memory generative adversarial networks for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5802-5810, 2019. 7 \ No newline at end of file diff --git a/2023/Not All Image Regions Matter_ Masked Vector Quantization for Autoregressive Image Generation/images.zip b/2023/Not All Image Regions Matter_ Masked Vector Quantization for Autoregressive Image Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a2bb58e06ae4ad7ad23f25e35a58247a10b01e87 --- /dev/null +++ b/2023/Not All Image Regions Matter_ Masked Vector Quantization for Autoregressive Image Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bd6f345a71cafbfb1791ba958e0c050442c01534b9820f7f33352c84233fdcfe +size 696790 diff --git a/2023/Not All Image Regions Matter_ Masked Vector Quantization for Autoregressive Image Generation/layout.json b/2023/Not All Image Regions Matter_ Masked Vector Quantization for Autoregressive Image Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e1dbcc89be9c7f3d0bc65f605ec85e7f1b2fac41 --- /dev/null +++ b/2023/Not All Image Regions Matter_ Masked Vector Quantization for Autoregressive Image Generation/layout.json @@ -0,0 +1,9806 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 91, + 103, + 501, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 103, + 501, + 140 + ], + "spans": [ + { + "bbox": [ + 91, + 103, + 501, + 140 + ], + "type": "text", + "content": "Not All Image Regions Matter: Masked Vector Quantization for Autoregressive Image Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 137, + 160, + 477, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 160, + 477, + 175 + ], + "spans": [ + { + "bbox": [ + 137, + 160, + 477, + 175 + ], + "type": "text", + "content": "Mengqi Huang" + }, + { + "bbox": [ + 137, + 160, + 477, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 137, + 160, + 477, + 175 + ], + "type": "text", + "content": ", Zhendong Mao" + }, + { + "bbox": [ + 137, + 160, + 477, + 175 + ], + "type": "inline_equation", + "content": "^{1,3*}" + }, + { + "bbox": [ + 137, + 160, + 477, + 175 + ], + "type": "text", + "content": ", Quan Wang" + }, + { + "bbox": [ + 137, + 160, + 477, + 175 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 137, + 160, + 477, + 175 + ], + "type": "text", + "content": ", Yongdong Zhang" + }, + { + "bbox": [ + 137, + 160, + 477, + 175 + ], + "type": "inline_equation", + "content": "^{1,3}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 52, + 175, + 563, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 175, + 563, + 217 + ], + "spans": [ + { + "bbox": [ + 52, + 175, + 563, + 217 + ], + "type": "text", + "content": "1University of Science and Technology of China, Hefei, China; 2MOE Key Laboratory of Trustworthy, Distributed Computing and Service, Beijing University of Posts and Telecommunications, Beijing, China; 3Institute of Artificial intelligence, Hefei Comprehensive National Science Center, Hefei, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 103, + 220, + 509, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 103, + 220, + 509, + 231 + ], + "spans": [ + { + "bbox": [ + 103, + 220, + 509, + 231 + ], + "type": "text", + "content": "huangmq@mail.ustc.edu.cn, {zdmao, zhyd73}@ustc.edu.cn, wangquan@bupt.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 258, + 192, + 271 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 258, + 192, + 271 + ], + "spans": [ + { + "bbox": [ + 143, + 258, + 192, + 271 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 284, + 290, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 284, + 290, + 594 + ], + "spans": [ + { + "bbox": [ + 46, + 284, + 290, + 594 + ], + "type": "text", + "content": "Existing autoregressive models follow the two-stage generation paradigm that first learns a codebook in the latent space for image reconstruction and then completes the image generation autoregressively based on the learned codebook. However, existing codebook learning simply models all local region information of images without distinguishing their different perceptual importance, which brings redundancy in the learned codebook that not only limits the next stage's autoregressive model's ability to model important structure but also results in high training cost and slow generation speed. In this study, we borrow the idea of importance perception from classical image coding theory and propose a novel two-stage framework, which consists of Masked Quantization VAE (MQ-VAE) and Stackformer, to relieve the model from modeling redundancy. Specifically, MQ-VAE incorporates an adaptive mask module for masking redundant region features before quantization and an adaptive de-mask module for recovering the original grid image feature map to faithfully reconstruct the original images after quantization. Then, Stackformer learns to predict the combination of the next code and its position in the feature map. Comprehensive experiments on various image generation validate our effectiveness and efficiency. Code will be released at https://github.com/CrossmodalGroup/MaskedVectorQuantization." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 616, + 128, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 616, + 128, + 628 + ], + "spans": [ + { + "bbox": [ + 47, + 616, + 128, + 628 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 636, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 636, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 636, + 287, + 696 + ], + "type": "text", + "content": "Deep generative models of images have received significant improvements over the past few years and broadly fall into two categories: likelihood-based models, which include VAEs [24], flow-based [36], diffusion models [17] and autoregressive models [40], and generative adversarial" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 307, + 259, + 426, + 377 + ], + "blocks": [ + { + "bbox": [ + 307, + 259, + 426, + 377 + ], + "lines": [ + { + "bbox": [ + 307, + 259, + 426, + 377 + ], + "spans": [ + { + "bbox": [ + 307, + 259, + 426, + 377 + ], + "type": "image", + "image_path": "99f379c90a21357ab413c94465c158b648298bb77dcaa5cd4d6f9204332a8f4a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 337, + 379, + 399, + 388 + ], + "lines": [ + { + "bbox": [ + 337, + 379, + 399, + 388 + ], + "spans": [ + { + "bbox": [ + 337, + 379, + 399, + 388 + ], + "type": "text", + "content": "(a) Existing methods" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 396, + 547, + 518 + ], + "lines": [ + { + "bbox": [ + 304, + 396, + 547, + 518 + ], + "spans": [ + { + "bbox": [ + 304, + 396, + 547, + 518 + ], + "type": "text", + "content": "Figure 1. Illustration of our motivation. (a) Existing works model all local regions without distinguishing their perceptual importance in stage 1, which not only brings redundancy (e.g., the textural regions like the background) in the learned codebook but also make the autoregressive models overly focus on modeling this redundancy and hinder other important structural regions modeling. (b) The codebook learning in our method only includes the important regions, e.g., the structural regions like corners and edges, since other unimportant ones can be restored even if missing, and thus autoregressive model could focus on modeling these important regions in stage 2 and results in better generation quality." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 427, + 259, + 545, + 377 + ], + "blocks": [ + { + "bbox": [ + 427, + 259, + 545, + 377 + ], + "lines": [ + { + "bbox": [ + 427, + 259, + 545, + 377 + ], + "spans": [ + { + "bbox": [ + 427, + 259, + 545, + 377 + ], + "type": "image", + "image_path": "8136d80f5730ddfd445c48546d49198624ae79e616abb9f2cf724b2afc2ec8d4.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 465, + 379, + 512, + 388 + ], + "lines": [ + { + "bbox": [ + 465, + 379, + 512, + 388 + ], + "spans": [ + { + "bbox": [ + 465, + 379, + 512, + 388 + ], + "type": "text", + "content": "(b) Our method" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 544, + 547, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 544, + 547, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 544, + 547, + 616 + ], + "type": "text", + "content": "networks (GANs) [14], which use discriminator networks to distinguish samples from generator networks and real examples. Compared with GANs, likelihood-based models' training objective, i.e., the negative log-likelihood (NLL) or its upper bound, incentives learning the full data distribution and allows for detecting overfitting." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 617, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 547, + 713 + ], + "type": "text", + "content": "Among the likelihood-based models, autoregressive models have recently attracted increasing attention for their impressive modeling ability and scalability. Recent autoregressive image generation [10, 12, 13, 28, 28, 34, 35, 37, 39] follows the two-stage generation paradigm, i.e., the first stage learns a codebook in the latent space for image reconstruction and the second stage completes the image generation in the raster-scan [13] order by autoregressive models" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 205, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 205, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 205, + 712 + ], + "type": "text", + "content": "*Zhendong Mao is the corresponding author." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2002" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": "based on the learned codebook. Since codebook learning in the first stage defines the discrete image representation for the next autoregressive modeling, a high-quality codebook is the key to generate high-quality images. Several recent works focus on improving the codebook learning in the first stage, e.g., VQGAN [13] introduces adversarial loss and perceptual loss. ViT-VQGAN [42] introduces a more expressive transformer backbone. RQ-VAE [28] introduces the residual quantization to reduce the resolution of the latent space. In general, the essence of existing codebook learning is the modeling of all local region information (i.e., an " + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "inline_equation", + "content": "8 \\times 8" + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "inline_equation", + "content": "16 \\times 16" + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": " patch) of images in the dataset, without distinguishing their different perceptual importance." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 232, + 289, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 232, + 289, + 507 + ], + "spans": [ + { + "bbox": [ + 46, + 232, + 289, + 507 + ], + "type": "text", + "content": "In this study, we point out that existing codebook learning exists gaps with classical image coding theory [20, 25, 26], the basic idea of which is to remove redundant information by perceiving the importance of different regions in images. The image coding theory reveals that an ideal image coding method should only encode images' perceptually important regions (i.e., which cannot be restored if missing) while discarding the unimportant ones (i.e., which can be restored by other image regions even if missing). The neglect of considering such perceptual importance in existing works poses problems in two aspects, as illustrated in Figure 1(a): (1) the existence of this large amount of repetitive and redundant information brings redundancy to the learned codebook, which further makes the autoregressive model in the next stage overly focus on modeling this redundancy while overlooking other important regions and finally degrades generation quality. (2) the redundancy makes the autoregressive model need to predict more (redundant) quantized codes to generate images, which significantly increases the training cost and decreases the generating speed. Although the effectiveness and efficiency of image coding theory have been widely validated, how to introduce this idea into codebook learning remains unexplored." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 289, + 715 + ], + "type": "text", + "content": "The key of applying image coding theory to codebook learning is to distinguish important image parts from unimportant ones correctly. Considering that the essential difference between these two sets lies in whether they can be restored if missing, we found that this distinction can be realized through the mask mechanism, i.e., the masked part is important if it cannot be faithfully restored, and otherwise unimportant. Based on the above observation, we thereby propose a novel two-stage generation paradigm upon the mask mechanism to relieve the model from modeling redundant information. Specifically, we first propose a Masked Quantization VAE (MQ-VAE) with two novel modules, i.e., an adaptive mask module for adaptively masking redundant region features before quantization, and an adaptive de-mask module for adaptively recovering the original grid image feature map to faithfully reconstruct original images after quantization. As for the adaptive mask module, it in" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "text", + "content": "corporates a lightweight content-aware scoring network that learns to measure the importance of each image region feature. The features are then ranked by the importance scores and only a subset of high-scored features will be quantized further. As for the adaptive de-mask module, we design a direction-constrained self-attention to encourage the information flow from the unmasked regions to the masked regions while blocking the reverse, which aims to infer the original masked region information based on unmasked ones. Thanks to the adaptive mask and de-mask mechanism, our MQ-VAE removes the negative effects of redundant image regions and also shortens the sequence length to achieve both effectiveness and efficiency." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 228, + 547, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 228, + 547, + 361 + ], + "spans": [ + { + "bbox": [ + 304, + 228, + 547, + 361 + ], + "type": "text", + "content": "Moreover, since different images have different important regions, the position of quantized codes in the feature map also dynamically changed. Therefore, we further propose Stackformer for learning to predict the combination of both codes and their corresponding positions. Concretely, the proposed Stackformer stacks a Code-Transformer and a Position-Transformer, where the Code-Transformer learns to predict the next code based on all previous codes and their positions, and the Position-Transformer learns to predict the next code's position based on all previous codes' positions and current code." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 361, + 547, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 361, + 547, + 433 + ], + "spans": [ + { + "bbox": [ + 304, + 361, + 547, + 433 + ], + "type": "text", + "content": "With our method, as shown in Figure 1(b), the codebook learning only includes the important regions, e.g., the structural regions, since unimportant ones like the background can be restored even if missing. And therefore the autoregressive model in the second stage could focus on modeling these important regions and brings better generation quality." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 317, + 433, + 541, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 433, + 541, + 444 + ], + "spans": [ + { + "bbox": [ + 317, + 433, + 541, + 444 + ], + "type": "text", + "content": "In a nutshell, we summarize our main contributions as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 445, + 547, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 445, + 547, + 494 + ], + "spans": [ + { + "bbox": [ + 304, + 445, + 547, + 494 + ], + "type": "text", + "content": "Conceptually, we point out that existing codebook learning ignores distinguishing the perceptual importance of different image regions, which brings redundancy that degrades generation quality and decreases generation speed." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 494, + 547, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 494, + 547, + 566 + ], + "spans": [ + { + "bbox": [ + 304, + 494, + 547, + 566 + ], + "type": "text", + "content": "Technically, (i) we propose MQ-VAE with a novel adaptive mask module to mask redundant region features before quantization and a novel adaptive de-mask module to recover the original feature map after quantization; (ii) we propose a novel Stackformer to predict the combination of both codes and their corresponding positions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 566, + 547, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 566, + 547, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 566, + 547, + 639 + ], + "type": "text", + "content": "Experimentally, comprehensive experiments on various generations validate our effectiveness and efficiency, i.e., we achieve " + }, + { + "bbox": [ + 304, + 566, + 547, + 639 + ], + "type": "inline_equation", + "content": "8.1\\%" + }, + { + "bbox": [ + 304, + 566, + 547, + 639 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 566, + 547, + 639 + ], + "type": "inline_equation", + "content": "2.3\\%" + }, + { + "bbox": [ + 304, + 566, + 547, + 639 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 566, + 547, + 639 + ], + "type": "inline_equation", + "content": "18.6\\%" + }, + { + "bbox": [ + 304, + 566, + 547, + 639 + ], + "type": "text", + "content": " FID improvement on un-, class-, and text-conditional state-of-the-art at million-level parameters, and faster generation speed compared to existing autoregressive models." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 650, + 392, + 662 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 650, + 392, + 662 + ], + "spans": [ + { + "bbox": [ + 306, + 650, + 392, + 662 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 670, + 545, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 670, + 545, + 684 + ], + "spans": [ + { + "bbox": [ + 306, + 670, + 545, + 684 + ], + "type": "text", + "content": "2.1. Autoregressive Modeling for Image Generation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "Autoregressive models for image generation have recently attracted increasing research attention and have" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2003" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 73, + 534, + 364 + ], + "blocks": [ + { + "bbox": [ + 48, + 73, + 534, + 364 + ], + "lines": [ + { + "bbox": [ + 48, + 73, + 534, + 364 + ], + "spans": [ + { + "bbox": [ + 48, + 73, + 534, + 364 + ], + "type": "image", + "image_path": "27735186e03b9dfe29d96d79864a9d9ed4bb3b557ec51166fa8b9afadc72e871.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 378, + 548, + 436 + ], + "lines": [ + { + "bbox": [ + 46, + 378, + 548, + 436 + ], + "spans": [ + { + "bbox": [ + 46, + 378, + 548, + 436 + ], + "type": "text", + "content": "Figure 2. The Illustration of our proposed two-stage generation framework. (a) In the first stage, MQ-VAE adaptively masks the redundant region features to prevent redundant codes while keeping important ones, which ensures that the original images can still be faithfully recovered. Here, " + }, + { + "bbox": [ + 46, + 378, + 548, + 436 + ], + "type": "inline_equation", + "content": "\\frac{10}{16}" + }, + { + "bbox": [ + 46, + 378, + 548, + 436 + ], + "type": "text", + "content": " regions are masked and " + }, + { + "bbox": [ + 46, + 378, + 548, + 436 + ], + "type": "inline_equation", + "content": "\\frac{6}{16}" + }, + { + "bbox": [ + 46, + 378, + 548, + 436 + ], + "type": "text", + "content": " regions are kept. (b) In the second stage, Stackformer stacks a Code-Transformer and a Position-Transformer to autoregressively predict the next code and its position in the original 2D feature map, respectively. (c) The attention mask of the proposed direction-constrained self-attention in the adaptive de-mask module for inferring masked regions features." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 454, + 289, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 454, + 289, + 706 + ], + "spans": [ + { + "bbox": [ + 46, + 454, + 289, + 706 + ], + "type": "text", + "content": "shown impressive results [10, 12, 13, 32, 34, 35, 37, 39, 42] among various generation tasks. Early autoregressive models [5, 33, 40] directly optimizing the likelihood of raw image pixels, e.g., Image-GPT [5] trains a transformer [41] to autoregressively predict pixels' cluster centroids, which could only generate images with a maximum resolution of " + }, + { + "bbox": [ + 46, + 454, + 289, + 706 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 46, + 454, + 289, + 706 + ], + "type": "text", + "content": ". [39] presents the Vector Quantized Variational Autoencoder (VQVAE), which learns images' low-dimension discrete representation and models their distribution autoregressively. VQ-VAE2 [35] extends this approach using a hierarchy of discrete representations. VQGAN [13] further improves the perceptual quality of reconstructed images using adversarial [14, 19] and perceptual loss [27]. ViT-VQGAN [42] introduces a more expressive transformer backbone. RQ-VAE [28] uses Residual Quantization [21, 31] to iteratively quantizes a vector and its residuals and represent the vector as a stack of tokens. Although vector quantization has become the fundamental technique for modern visual autoregressive models, the critical removing redundancy in codebook learning has not been explored yet, which becomes a critical bottleneck." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 306, + 454, + 413, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 454, + 413, + 467 + ], + "spans": [ + { + "bbox": [ + 306, + 454, + 413, + 467 + ], + "type": "text", + "content": "2.2. Masked Modeling" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 474, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 545, + 715 + ], + "type": "text", + "content": "Masked modeling is popular among both natural language processing and computer vision. BERT [8] randomly masks a portion of the input sequence and trains models to predict the missing content. In the computer vision domain, the ViT [11] studies masked patch prediction for self-supervised learning. BEiT [1] proposes to predict discrete tokens. Most recently, MaskGIT [4] also used the masking strategy for VQ-based image generation. However, our proposed method differs from MaskGIT in two aspects: (1) Our primary motivation for the masking strategy applied in the proposed MQ-VAE in the first stage aims to learn a more compact and effective vector quantization (VQ) itself by masking perceptual unimportant regions, while MaskGIT uses masking strategy in the second stage to better use a learned VQ. (2) The mask in our proposed MQ-VAE is learned and adaptively changed according to different image content, while the mask in MaskGIT is randomly sampled for the mask-and-predict training. In conclusion, to the best of our knowledge, this is the first time that masked modeling has been applied for vector quantization." + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2004" + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 71, + 129, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 71, + 129, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 129, + 85 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 91, + 287, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 91, + 287, + 186 + ], + "spans": [ + { + "bbox": [ + 46, + 91, + 287, + 186 + ], + "type": "text", + "content": "We propose a novel two-stage framework with MQ-VAE and Stackformer for autoregressive image generation, as illustrated in Figure 2. MQ-VAE only masks redundant region features to prevent redundant codes and Stackformer stacks two transformers to autoregressively predict the next code and its position. In the following, we will first briefly revisit the formulation of vector quantization and then describe our proposed method in detail." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 194, + 126, + 208 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 194, + 126, + 208 + ], + "spans": [ + { + "bbox": [ + 47, + 194, + 126, + 208 + ], + "type": "text", + "content": "3.1. Preliminary" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 213, + 288, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 213, + 288, + 345 + ], + "spans": [ + { + "bbox": [ + 46, + 213, + 288, + 345 + ], + "type": "text", + "content": "We follow the definition and notations of previous works [13, 28]. Specifically, the codebook " + }, + { + "bbox": [ + 46, + 213, + 288, + 345 + ], + "type": "inline_equation", + "content": "\\mathcal{C} \\coloneqq \\{(k, e(k))\\}_{k \\in [K]}" + }, + { + "bbox": [ + 46, + 213, + 288, + 345 + ], + "type": "text", + "content": " is defined as the set of finite pairs of code " + }, + { + "bbox": [ + 46, + 213, + 288, + 345 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 213, + 288, + 345 + ], + "type": "text", + "content": " and its code embedding " + }, + { + "bbox": [ + 46, + 213, + 288, + 345 + ], + "type": "inline_equation", + "content": "e(k) \\in \\mathbb{R}^{n_z}" + }, + { + "bbox": [ + 46, + 213, + 288, + 345 + ], + "type": "text", + "content": ". Here " + }, + { + "bbox": [ + 46, + 213, + 288, + 345 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 213, + 288, + 345 + ], + "type": "text", + "content": " is the codebook size and " + }, + { + "bbox": [ + 46, + 213, + 288, + 345 + ], + "type": "inline_equation", + "content": "n_z" + }, + { + "bbox": [ + 46, + 213, + 288, + 345 + ], + "type": "text", + "content": " is the code dimension. An image " + }, + { + "bbox": [ + 46, + 213, + 288, + 345 + ], + "type": "inline_equation", + "content": "\\mathbf{X} \\in \\mathbb{R}^{H_0 \\times W_0 \\times 3}" + }, + { + "bbox": [ + 46, + 213, + 288, + 345 + ], + "type": "text", + "content": " is first encoded into grid features " + }, + { + "bbox": [ + 46, + 213, + 288, + 345 + ], + "type": "inline_equation", + "content": "\\mathbf{Z} = E(\\mathbf{X}) \\in \\mathbb{R}^{H \\times W \\times n_z}" + }, + { + "bbox": [ + 46, + 213, + 288, + 345 + ], + "type": "text", + "content": " by the encoder " + }, + { + "bbox": [ + 46, + 213, + 288, + 345 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 46, + 213, + 288, + 345 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 213, + 288, + 345 + ], + "type": "inline_equation", + "content": "(H, W) = (H_0 / f, W_0 / f)" + }, + { + "bbox": [ + 46, + 213, + 288, + 345 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 213, + 288, + 345 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 213, + 288, + 345 + ], + "type": "text", + "content": " is the corresponding downsampling factor. For each vector " + }, + { + "bbox": [ + 46, + 213, + 288, + 345 + ], + "type": "inline_equation", + "content": "z \\in \\mathbb{R}^{n_z}" + }, + { + "bbox": [ + 46, + 213, + 288, + 345 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 46, + 213, + 288, + 345 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + }, + { + "bbox": [ + 46, + 213, + 288, + 345 + ], + "type": "text", + "content": ", it is replaced with the code embedding that has the closest euclidean distance with it in the codebook " + }, + { + "bbox": [ + 46, + 213, + 288, + 345 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 46, + 213, + 288, + 345 + ], + "type": "text", + "content": " through the vector quantization operation " + }, + { + "bbox": [ + 46, + 213, + 288, + 345 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}(\\cdot)" + }, + { + "bbox": [ + 46, + 213, + 288, + 345 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 100, + 356, + 287, + 376 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 356, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 100, + 356, + 287, + 376 + ], + "type": "interline_equation", + "content": "\\mathcal {Q} (z; \\mathcal {C}) = \\arg \\min _ {k \\in [ K ]} | | z - e _ {k} | | _ {2} ^ {2}. \\tag {1}", + "image_path": "d576feba1d6dbb470cc08cbab96196367197d6d7dd563f029da10cac984c7760.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 380, + 287, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 380, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 46, + 380, + 287, + 441 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 46, + 380, + 287, + 441 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}(z;\\mathcal{C})" + }, + { + "bbox": [ + 46, + 380, + 287, + 441 + ], + "type": "text", + "content": " is the quantized code. " + }, + { + "bbox": [ + 46, + 380, + 287, + 441 + ], + "type": "inline_equation", + "content": "z^q = e(\\mathcal{Q}(z;\\mathcal{C}))" + }, + { + "bbox": [ + 46, + 380, + 287, + 441 + ], + "type": "text", + "content": " is the quantized vector. By applying " + }, + { + "bbox": [ + 46, + 380, + 287, + 441 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}(\\cdot)" + }, + { + "bbox": [ + 46, + 380, + 287, + 441 + ], + "type": "text", + "content": " to each feature vector, we could get the quantized code map " + }, + { + "bbox": [ + 46, + 380, + 287, + 441 + ], + "type": "inline_equation", + "content": "\\mathbf{M}\\in [K]^{H\\times W}" + }, + { + "bbox": [ + 46, + 380, + 287, + 441 + ], + "type": "text", + "content": " and the quantized features " + }, + { + "bbox": [ + 46, + 380, + 287, + 441 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}^q\\in \\mathbb{R}^{H\\times W\\times n_z}" + }, + { + "bbox": [ + 46, + 380, + 287, + 441 + ], + "type": "text", + "content": ". The original image is reconstructed by the decoder " + }, + { + "bbox": [ + 46, + 380, + 287, + 441 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 380, + 287, + 441 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 46, + 380, + 287, + 441 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{X}} = D(\\mathbf{Z}^q)" + }, + { + "bbox": [ + 46, + 380, + 287, + 441 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 449, + 155, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 449, + 155, + 462 + ], + "spans": [ + { + "bbox": [ + 47, + 449, + 155, + 462 + ], + "type": "text", + "content": "3.2. Stage 1: MQ-VAE" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 467, + 287, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 467, + 287, + 586 + ], + "spans": [ + { + "bbox": [ + 46, + 467, + 287, + 586 + ], + "type": "text", + "content": "Existing methods quantize each feature vector of " + }, + { + "bbox": [ + 46, + 467, + 287, + 586 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + }, + { + "bbox": [ + 46, + 467, + 287, + 586 + ], + "type": "text", + "content": " without distinguishing their different perceptual importance and thus bring redundancy in the learned codebook, which not only degrades the generation quality but also decreases the generation speed. To relieve the model from this redundancy, we propose MQ-VAE with two novel modules, i.e., the adaptive mask module for adaptively masking redundant region features before vector quantization and adaptive de-mask module for adaptively recovering the original grid image feature map after vector quantization." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 587, + 287, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 587, + 287, + 659 + ], + "spans": [ + { + "bbox": [ + 46, + 587, + 287, + 659 + ], + "type": "text", + "content": "Adaptive Mask Module. The encoded grid feature map " + }, + { + "bbox": [ + 46, + 587, + 287, + 659 + ], + "type": "inline_equation", + "content": "\\mathbf{Z} \\in \\mathbb{R}^{H \\times W \\times n_e}" + }, + { + "bbox": [ + 46, + 587, + 287, + 659 + ], + "type": "text", + "content": " is first flattened into " + }, + { + "bbox": [ + 46, + 587, + 287, + 659 + ], + "type": "inline_equation", + "content": "\\mathbf{Z} \\in \\mathbb{R}^{L \\times n_e}" + }, + { + "bbox": [ + 46, + 587, + 287, + 659 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 587, + 287, + 659 + ], + "type": "inline_equation", + "content": "L = H \\times W" + }, + { + "bbox": [ + 46, + 587, + 287, + 659 + ], + "type": "text", + "content": ". The proposed adaptive mask module then uses a lightweight scoring network " + }, + { + "bbox": [ + 46, + 587, + 287, + 659 + ], + "type": "inline_equation", + "content": "f_s" + }, + { + "bbox": [ + 46, + 587, + 287, + 659 + ], + "type": "text", + "content": " to measure the importance of each region feature " + }, + { + "bbox": [ + 46, + 587, + 287, + 659 + ], + "type": "inline_equation", + "content": "z_l" + }, + { + "bbox": [ + 46, + 587, + 287, + 659 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 46, + 587, + 287, + 659 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + }, + { + "bbox": [ + 46, + 587, + 287, + 659 + ], + "type": "text", + "content": ", which is implemented as a two-layer MLP:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 116, + 671, + 287, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 671, + 287, + 685 + ], + "spans": [ + { + "bbox": [ + 116, + 671, + 287, + 685 + ], + "type": "interline_equation", + "content": "s _ {l} = f _ {s} \\left(\\boldsymbol {z} _ {l}\\right), l = 1, \\dots , L. \\tag {2}", + "image_path": "fe21b86186437b9ec5ae7dda9e9e1e7dc7a5e297998941e870c86b4bfef3ee40.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": "The larger score " + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "s_l" + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": " is, the more important the region feature " + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "z_l" + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": " is. Then the region features are sorted in descending" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": "order according to the predicted scores. The sorted region features and their scores are denoted as " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "\\{z_l^{\\prime}\\}" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "\\{s_l^{\\prime}\\}" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": " respectively, where " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "l = 1,\\dots ,L" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": ". To enable the learning of " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "f_{s}" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": ", the predicted scores are further multiplied with the normalized region features as modulating factors. We select the top " + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 72, + 545, + 144 + ], + "type": "text", + "content": " scoring vectors as the important region features," + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 315, + 159, + 545, + 174 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 159, + 545, + 174 + ], + "spans": [ + { + "bbox": [ + 315, + 159, + 545, + 174 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {Z}} = \\left\\{z _ {l} ^ {\\prime \\prime} \\mid z _ {l} ^ {\\prime \\prime} = \\operatorname {L a y e r N o r m} \\left(z _ {l} ^ {\\prime}\\right) * s _ {l} ^ {\\prime} \\right\\}, l = 1, \\dots , N. \\tag {3}", + "image_path": "9918861cd779eadedab545bcd8490b0548e2941c67219547903e3503ca7b7f73.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 339, + 177, + 545, + 194 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 177, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 339, + 177, + 545, + 194 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {P}} = \\left\\{p _ {z _ {l} ^ {\\prime \\prime}} \\mid p _ {z _ {l} ^ {\\prime \\prime}} \\in \\{0, \\dots , L \\} \\right\\}, l = 1, \\dots , N. \\tag {4}", + "image_path": "00931bb63a8e3130c0610846a5e89831974b093cd7d76ba7431fa1ae4ccfc53c.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 198, + 545, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 198, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 304, + 198, + 545, + 330 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 304, + 198, + 545, + 330 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{Z}}" + }, + { + "bbox": [ + 304, + 198, + 545, + 330 + ], + "type": "text", + "content": " denotes the selected important region features set, and " + }, + { + "bbox": [ + 304, + 198, + 545, + 330 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{P}}" + }, + { + "bbox": [ + 304, + 198, + 545, + 330 + ], + "type": "text", + "content": " denotes the corresponding position set that represents the position of each selected region feature in the original 2D feature map. The selected number " + }, + { + "bbox": [ + 304, + 198, + 545, + 330 + ], + "type": "inline_equation", + "content": "N = \\alpha \\times L" + }, + { + "bbox": [ + 304, + 198, + 545, + 330 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 198, + 545, + 330 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 198, + 545, + 330 + ], + "type": "text", + "content": " is a constant fractional value. The mask ratio is defined as " + }, + { + "bbox": [ + 304, + 198, + 545, + 330 + ], + "type": "inline_equation", + "content": "1 - \\alpha" + }, + { + "bbox": [ + 304, + 198, + 545, + 330 + ], + "type": "text", + "content": ". This design also enables a flexible trade-off between the image generation speed and image generation quality, which we will discuss in experiments. After obtaining " + }, + { + "bbox": [ + 304, + 198, + 545, + 330 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{Z}}" + }, + { + "bbox": [ + 304, + 198, + 545, + 330 + ], + "type": "text", + "content": ", we further apply the quantization function " + }, + { + "bbox": [ + 304, + 198, + 545, + 330 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 304, + 198, + 545, + 330 + ], + "type": "text", + "content": " to each of them and obtain the quantized important region features set " + }, + { + "bbox": [ + 304, + 198, + 545, + 330 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{Z}}^q" + }, + { + "bbox": [ + 304, + 198, + 545, + 330 + ], + "type": "text", + "content": " as well as its code matrix " + }, + { + "bbox": [ + 304, + 198, + 545, + 330 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{M}}" + }, + { + "bbox": [ + 304, + 198, + 545, + 330 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 331, + 545, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 331, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 331, + 545, + 521 + ], + "type": "text", + "content": "Adaptive De-mask Module. After quantization, we fill the quantized features " + }, + { + "bbox": [ + 304, + 331, + 545, + 521 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{Z}}^q" + }, + { + "bbox": [ + 304, + 331, + 545, + 521 + ], + "type": "text", + "content": " back into the original 2D feature map according to " + }, + { + "bbox": [ + 304, + 331, + 545, + 521 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{P}}" + }, + { + "bbox": [ + 304, + 331, + 545, + 521 + ], + "type": "text", + "content": ", while other masked positions are filled with a uniformly initialized learnable mask code embedding, as shown in Figure 2(a). Directly inputting filled grid features to the decoder " + }, + { + "bbox": [ + 304, + 331, + 545, + 521 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 304, + 331, + 545, + 521 + ], + "type": "text", + "content": " could bring sub-optimal reconstruction results since the mask code embedding here only serves as the placeholders that contain little information. Therefore, we further propose the adaptive de-mask module, which applies a novel direction-constrained self-attention to encourage the information flow from unmasked regions to the masked ones while blocking the reverse. Such a design allows the model to utilize the unmasked region features to infer the masked ones while also preventing the masked regions to have negative impacts on the unmasked ones since they are less informative." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 522, + 545, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 545, + 583 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 545, + 583 + ], + "type": "text", + "content": "Our adaptive de-mask module is implemented as " + }, + { + "bbox": [ + 304, + 522, + 545, + 583 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 304, + 522, + 545, + 583 + ], + "type": "text", + "content": " identical sub-modules, where each consists of a direction-constrained self-attention block and a Resnet block. The direction-constrained self-attention is mathematically formed as (Resnet block is omitted for simplicity):" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 345, + 597, + 545, + 612 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 597, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 345, + 597, + 545, + 612 + ], + "type": "interline_equation", + "content": "\\boldsymbol {q}, \\boldsymbol {k}, \\boldsymbol {v} = \\boldsymbol {W} _ {q} \\hat {\\boldsymbol {Z}} ^ {q, h}, \\boldsymbol {W} _ {k} \\hat {\\boldsymbol {Z}} ^ {q, h}, \\boldsymbol {W} _ {v} \\hat {\\boldsymbol {Z}} ^ {q, h} \\tag {5}", + "image_path": "6012fba49e2b2540f903ad49334fc862dbbde57a0acbb00d39ec6dabb66cf13a.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 364, + 614, + 545, + 643 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 614, + 545, + 643 + ], + "spans": [ + { + "bbox": [ + 364, + 614, + 545, + 643 + ], + "type": "interline_equation", + "content": "\\boldsymbol {A} = \\left(\\operatorname {S o f t M a x} \\left(\\frac {\\boldsymbol {q} \\boldsymbol {k} ^ {T}}{\\sqrt {n _ {e}}}\\right)\\right) \\odot \\boldsymbol {B} _ {\\boldsymbol {h}} \\tag {6}", + "image_path": "50c99489562bbfcff44c291f0f57ebe52332211948ab0e7bd736116712da6d52.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 394, + 644, + 545, + 658 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 644, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 394, + 644, + 545, + 658 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {Z}} ^ {q, h + 1} = \\boldsymbol {A} \\boldsymbol {v}. \\tag {7}", + "image_path": "e6799e4657f9777d3a3966f9cec00e3a1297c2a1cc6eaa5de8e35bd6895b09a1.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "h\\in \\{1,\\dots ,H\\}" + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{Z}}^{q,1}" + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": " is the filled quantized grid features. " + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\pmb {W}_q,\\pmb {W}_k,\\pmb {W}_v\\in \\mathbb{R}^{n_e\\times n_e}" + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": " are the learnable parameters. " + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{B}_{h}" + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": " is the attention mask at " + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": " sub-module. Specifically, since the initial mask code contains little information," + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "2005" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": "we define " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "B_{1} = [b_{l}\\in \\{0,1\\} ,|l = 1,\\dots,L]\\in \\mathbb{R}^{1\\times L}" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": " to forbid it from attending to other unmasked codes to avoid negative impact, where 0 for the mask position and 1 for the unmasked. Considering that the mask code is updated with more and more information after each sub-module, we propose to synchronously amplify its interaction with other codes step by step through a mask updating mechanism:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 133, + 167, + 287, + 182 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 167, + 287, + 182 + ], + "spans": [ + { + "bbox": [ + 133, + 167, + 287, + 182 + ], + "type": "interline_equation", + "content": "\\boldsymbol {B} _ {h + 1} = \\sqrt {\\boldsymbol {B} _ {h}}, \\tag {8}", + "image_path": "3597f349304e1d05e1d8ca6a70a9494f7e39f72a3bd85ccabf700e41969969b9.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 187, + 287, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 187, + 287, + 224 + ], + "spans": [ + { + "bbox": [ + 46, + 187, + 287, + 224 + ], + "type": "text", + "content": "where the initial 0 in " + }, + { + "bbox": [ + 46, + 187, + 287, + 224 + ], + "type": "inline_equation", + "content": "B_{1}" + }, + { + "bbox": [ + 46, + 187, + 287, + 224 + ], + "type": "text", + "content": " is replaced with a small number 0.02, since " + }, + { + "bbox": [ + 46, + 187, + 287, + 224 + ], + "type": "inline_equation", + "content": "\\sqrt{0}" + }, + { + "bbox": [ + 46, + 187, + 287, + 224 + ], + "type": "text", + "content": " will always result in 0. Finally, the original image is recovered by the decoder as " + }, + { + "bbox": [ + 46, + 187, + 287, + 224 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{X}} = D(\\hat{\\mathbf{Z}}^{q,H})" + }, + { + "bbox": [ + 46, + 187, + 287, + 224 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 231, + 170, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 231, + 170, + 243 + ], + "spans": [ + { + "bbox": [ + 47, + 231, + 170, + 243 + ], + "type": "text", + "content": "3.3. Stage 2: Stackformer" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 250, + 289, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 250, + 289, + 501 + ], + "spans": [ + { + "bbox": [ + 46, + 250, + 289, + 501 + ], + "type": "text", + "content": "The perceptual important regions of different images vary. Therefore the positions of quantized codes in the feature map also dynamically change along with the image content. As a result, our proposed MQ-VAE formulates an image as both the code sequence " + }, + { + "bbox": [ + 46, + 250, + 289, + 501 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{M}}" + }, + { + "bbox": [ + 46, + 250, + 289, + 501 + ], + "type": "text", + "content": " and the code position sequence " + }, + { + "bbox": [ + 46, + 250, + 289, + 501 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{P}}" + }, + { + "bbox": [ + 46, + 250, + 289, + 501 + ], + "type": "text", + "content": ". To simultaneously learn the combination of the codes and their positions, we propose a novel Stackformer, which stacks a Code-Transformer and a Position-Transformer. The Code-Transformer learns to predict the next code based on all previous steps' codes and their positions, while the Position-Transformer learns to predict the next code's position based on all previous steps' positions and current code. Directly treating the importance descending order sequence " + }, + { + "bbox": [ + 46, + 250, + 289, + 501 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{M}}" + }, + { + "bbox": [ + 46, + 250, + 289, + 501 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 250, + 289, + 501 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{P}}" + }, + { + "bbox": [ + 46, + 250, + 289, + 501 + ], + "type": "text", + "content": " as the inputs are natural, but the dramatic position changes of adjacent code could make the network hard to converge. For example, the position of the first code may be in the upper left corner of the image, while the position of the second code may be in the lower right corner of the image. Therefore, we further propose to use the raster-scan order [13] to rearrange both sequences to deal with the converge problem." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 502, + 287, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 502, + 287, + 537 + ], + "spans": [ + { + "bbox": [ + 46, + 502, + 287, + 537 + ], + "type": "text", + "content": "Mathematically, taking the raster-scan code and code position sequence " + }, + { + "bbox": [ + 46, + 502, + 287, + 537 + ], + "type": "inline_equation", + "content": "(\\overline{\\mathbf{M}},\\overline{\\mathbf{P}}) =" + }, + { + "bbox": [ + 46, + 502, + 287, + 537 + ], + "type": "text", + "content": " rearrange " + }, + { + "bbox": [ + 46, + 502, + 287, + 537 + ], + "type": "inline_equation", + "content": "(\\hat{\\mathbf{M}},\\hat{\\mathbf{P}})" + }, + { + "bbox": [ + 46, + 502, + 287, + 537 + ], + "type": "text", + "content": ", Stackformer learns " + }, + { + "bbox": [ + 46, + 502, + 287, + 537 + ], + "type": "inline_equation", + "content": "p(\\overline{\\mathbf{M}},\\overline{\\mathbf{P}})" + }, + { + "bbox": [ + 46, + 502, + 287, + 537 + ], + "type": "text", + "content": ", which is autoregressively factorized as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 556, + 287, + 590 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 556, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 57, + 556, + 287, + 590 + ], + "type": "interline_equation", + "content": "p (\\overline {{\\mathbf {M}}}, \\overline {{\\mathbf {P}}}) = \\prod_ {l = 1} ^ {N} p (\\overline {{\\mathbf {M}}} _ {l} | \\overline {{\\mathbf {M}}} _ {< l}, \\overline {{\\mathbf {P}}} _ {< l}) p (\\overline {{\\mathbf {P}}} _ {l} | \\overline {{\\mathbf {M}}} _ {\\leq l}, \\overline {{\\mathbf {P}}} _ {< l}) \\quad (9)", + "image_path": "e76cc48b6d71ce0cc27fb5f8b7191ec4453586c2be1520838fc649ffe7139b4b.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 598, + 287, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 598, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 47, + 598, + 287, + 624 + ], + "type": "text", + "content": "Code-Transformer takes the sum of code embeddings " + }, + { + "bbox": [ + 47, + 598, + 287, + 624 + ], + "type": "inline_equation", + "content": "e_{c}(\\cdot)" + }, + { + "bbox": [ + 47, + 598, + 287, + 624 + ], + "type": "text", + "content": " and code position embedding " + }, + { + "bbox": [ + 47, + 598, + 287, + 624 + ], + "type": "inline_equation", + "content": "e_{p}(\\cdot)" + }, + { + "bbox": [ + 47, + 598, + 287, + 624 + ], + "type": "text", + "content": " as inputs:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 82, + 634, + 287, + 649 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 634, + 287, + 649 + ], + "spans": [ + { + "bbox": [ + 82, + 634, + 287, + 649 + ], + "type": "interline_equation", + "content": "\\mathbf {U} _ {c} = \\boldsymbol {e} _ {c} \\left(\\overline {{\\mathbf {M}}} _ {[ 1: N _ {c} + N ]}\\right) + \\boldsymbol {e} _ {p 1} \\left(\\overline {{\\mathbf {P}}} _ {[ 1: N _ {c} + N ]}\\right), \\tag {10}", + "image_path": "df13325bdcb205fb66b495afb688d89f568befd29ff66f39fe63e685ede42877.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "inline_equation", + "content": "N_{c}" + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": " is the condition length. For the unconditional generation, we add a " + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "inline_equation", + "content": "<\\mathrm{sos}>" + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": " code at the start of the code and code position sequence. For conditioning, we append class or text codes to the start of the code sequence and the same length of " + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "inline_equation", + "content": "<\\mathrm{sos}>" + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": " code to the code position sequence. We" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": "further add an extra learned absolute position embedding to " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "\\mathbf{U}_c" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " to form the final input, which makes the network aware of the absolute position of the sequence. After processing by Code-Transformer, the output hidden vector " + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_c" + }, + { + "bbox": [ + 304, + 72, + 545, + 156 + ], + "type": "text", + "content": " encodes both code and their position information and is used for the next code prediction. The negative log-likelihood (NLL) loss for code autoregressive training is:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 351, + 167, + 545, + 182 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 167, + 545, + 182 + ], + "spans": [ + { + "bbox": [ + 351, + 167, + 545, + 182 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {c o d e}} = \\mathbb {E} \\left[ - \\log p \\left(\\overline {{\\mathbf {M}}} _ {l} \\mid \\overline {{\\mathbf {M}}} _ {< l}, \\overline {{\\mathbf {P}}} _ {< l} \\right]\\right). \\tag {11}", + "image_path": "47d4ba4a029539201cf01f6d46f870d62287d2c46600712a9ac7e3ceebce07da.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 185, + 545, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 185, + 545, + 221 + ], + "spans": [ + { + "bbox": [ + 304, + 185, + 545, + 221 + ], + "type": "text", + "content": "Position-Transformer takes the sum of CodeTransformer's output hidden vector " + }, + { + "bbox": [ + 304, + 185, + 545, + 221 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_c" + }, + { + "bbox": [ + 304, + 185, + 545, + 221 + ], + "type": "text", + "content": " and an extra code embedding as input:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 311, + 241, + 545, + 256 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 241, + 545, + 256 + ], + "spans": [ + { + "bbox": [ + 311, + 241, + 545, + 256 + ], + "type": "interline_equation", + "content": "\\mathbf {U} _ {p} = \\mathbf {H} _ {c} \\left[ N _ {c}: N _ {c} + N - 1 \\right] + \\boldsymbol {e} _ {c} \\left(\\overline {{\\mathbf {M}}} _ {\\left[ N _ {c} + 1: N _ {c} + N \\right]}\\right). \\tag {12}", + "image_path": "3df77c0885f0710e670da74b62e9220d80e6ed9cd148182ef1156ffa50bf9338.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 263, + 545, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 263, + 545, + 347 + ], + "spans": [ + { + "bbox": [ + 304, + 263, + 545, + 347 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 304, + 263, + 545, + 347 + ], + "type": "inline_equation", + "content": "\\mathbf{U}_p" + }, + { + "bbox": [ + 304, + 263, + 545, + 347 + ], + "type": "text", + "content": " is the input for Position-Transformer and the information of current code is included in " + }, + { + "bbox": [ + 304, + 263, + 545, + 347 + ], + "type": "inline_equation", + "content": "e_c(\\overline{\\mathbf{M}}_{[N_c + 1:N_c + N]})" + }, + { + "bbox": [ + 304, + 263, + 545, + 347 + ], + "type": "text", + "content": ". The design idea behind this is that when predicting the next code's position, the model should not only be aware of previous steps' codes and their position information but also should be aware of current code information. The negative log-likelihood (NLL) for position autoregressive training is:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 348, + 357, + 545, + 372 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 357, + 545, + 372 + ], + "spans": [ + { + "bbox": [ + 348, + 357, + 545, + 372 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {p o s i t i o n}} = \\mathbb {E} \\left[ - \\log p \\left(\\overline {{\\mathbf {P}}} _ {l} \\mid \\overline {{\\mathbf {M}}} _ {\\leq l}, \\overline {{\\mathbf {P}}} _ {< l}\\right) \\right]. \\tag {13}", + "image_path": "07180cf980eff3a16694cdee0c521ef644045afcf6500ee51fada746c5bf4bdb.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 376, + 545, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 376, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 306, + 376, + 545, + 399 + ], + "type": "text", + "content": "Training & Inference. The total loss for training Stackformer is defined as:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 381, + 411, + 545, + 425 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 411, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 381, + 411, + 545, + 425 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {c o d e}} + \\mathcal {L} _ {\\text {p o s i t i o n}}. \\tag {14}", + "image_path": "e230cda1625fe2938b9f8a69e5f8599f86b7718e3f2f9a93d58b4e31af0da455.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 429, + 545, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 429, + 545, + 465 + ], + "spans": [ + { + "bbox": [ + 304, + 429, + 545, + 465 + ], + "type": "text", + "content": "The inference procedure is illustrated in Algorithm 1, where we take the unconditional generation as an example and conditional generation can be derived accordingly." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 476, + 523, + 489 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 476, + 523, + 489 + ], + "spans": [ + { + "bbox": [ + 306, + 476, + 523, + 489 + ], + "type": "text", + "content": "Algorithm 1 Unconditional sampling of Stackformer." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 492, + 421, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 492, + 421, + 504 + ], + "spans": [ + { + "bbox": [ + 306, + 492, + 421, + 504 + ], + "type": "text", + "content": "Input: The sample step " + }, + { + "bbox": [ + 306, + 492, + 421, + 504 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 306, + 492, + 421, + 504 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 323, + 505, + 518, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 505, + 518, + 516 + ], + "spans": [ + { + "bbox": [ + 323, + 505, + 518, + 516 + ], + "type": "text", + "content": "The code sequence " + }, + { + "bbox": [ + 323, + 505, + 518, + 516 + ], + "type": "inline_equation", + "content": "M_{\\text{sample}}" + }, + { + "bbox": [ + 323, + 505, + 518, + 516 + ], + "type": "text", + "content": " of a single " + }, + { + "bbox": [ + 323, + 505, + 518, + 516 + ], + "type": "inline_equation", + "content": "<\\text{sos}>" + }, + { + "bbox": [ + 323, + 505, + 518, + 516 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 323, + 516, + 544, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 516, + 544, + 528 + ], + "spans": [ + { + "bbox": [ + 323, + 516, + 544, + 528 + ], + "type": "text", + "content": "The code position sequence " + }, + { + "bbox": [ + 323, + 516, + 544, + 528 + ], + "type": "inline_equation", + "content": "P_{\\text{sample}}" + }, + { + "bbox": [ + 323, + 516, + 544, + 528 + ], + "type": "text", + "content": " of a single " + }, + { + "bbox": [ + 323, + 516, + 544, + 528 + ], + "type": "inline_equation", + "content": "<\\text{sos}>" + }, + { + "bbox": [ + 323, + 516, + 544, + 528 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 528, + 443, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 528, + 443, + 539 + ], + "spans": [ + { + "bbox": [ + 306, + 528, + 443, + 539 + ], + "type": "text", + "content": "Output: The generated image " + }, + { + "bbox": [ + 306, + 528, + 443, + 539 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 306, + 528, + 443, + 539 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 309, + 540, + 544, + 696 + ], + "type": "list", + "angle": 0, + "index": 36, + "blocks": [ + { + "bbox": [ + 312, + 540, + 417, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 540, + 417, + 552 + ], + "spans": [ + { + "bbox": [ + 312, + 540, + 417, + 552 + ], + "type": "text", + "content": "1: for each " + }, + { + "bbox": [ + 312, + 540, + 417, + 552 + ], + "type": "inline_equation", + "content": "n \\in [1, N]" + }, + { + "bbox": [ + 312, + 540, + 417, + 552 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 552, + 523, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 552, + 523, + 564 + ], + "spans": [ + { + "bbox": [ + 312, + 552, + 523, + 564 + ], + "type": "text", + "content": "2: " + }, + { + "bbox": [ + 312, + 552, + 523, + 564 + ], + "type": "inline_equation", + "content": "H_{c} =" + }, + { + "bbox": [ + 312, + 552, + 523, + 564 + ], + "type": "text", + "content": " Code-Transformer(Msample, " + }, + { + "bbox": [ + 312, + 552, + 523, + 564 + ], + "type": "inline_equation", + "content": "P_{\\text{sample}}" + }, + { + "bbox": [ + 312, + 552, + 523, + 564 + ], + "type": "text", + "content": ");" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 564, + 454, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 564, + 454, + 575 + ], + "spans": [ + { + "bbox": [ + 312, + 564, + 454, + 575 + ], + "type": "text", + "content": "3: Sample next code " + }, + { + "bbox": [ + 312, + 564, + 454, + 575 + ], + "type": "inline_equation", + "content": "M_{n}" + }, + { + "bbox": [ + 312, + 564, + 454, + 575 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 312, + 564, + 454, + 575 + ], + "type": "inline_equation", + "content": "H_{c}" + }, + { + "bbox": [ + 312, + 564, + 454, + 575 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 576, + 479, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 576, + 479, + 588 + ], + "spans": [ + { + "bbox": [ + 312, + 576, + 479, + 588 + ], + "type": "text", + "content": "4: " + }, + { + "bbox": [ + 312, + 576, + 479, + 588 + ], + "type": "inline_equation", + "content": "M_{\\text{sample}} = \\operatorname{concat}(M_{\\text{sample}}, M_n)" + }, + { + "bbox": [ + 312, + 576, + 479, + 588 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 312, + 588, + 533, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 588, + 533, + 600 + ], + "spans": [ + { + "bbox": [ + 312, + 588, + 533, + 600 + ], + "type": "text", + "content": "5: " + }, + { + "bbox": [ + 312, + 588, + 533, + 600 + ], + "type": "inline_equation", + "content": "H_{p} = \\mathrm{Position - Transformer}(H_{c},M_{\\mathrm{sample}}[2:])" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 312, + 600, + 537, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 600, + 537, + 611 + ], + "spans": [ + { + "bbox": [ + 312, + 600, + 537, + 611 + ], + "type": "text", + "content": "6: Mask already sampled positions to avoid conflicts;" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 312, + 612, + 486, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 612, + 486, + 624 + ], + "spans": [ + { + "bbox": [ + 312, + 612, + 486, + 624 + ], + "type": "text", + "content": "7: Sample next code position " + }, + { + "bbox": [ + 312, + 612, + 486, + 624 + ], + "type": "inline_equation", + "content": "P_{n}" + }, + { + "bbox": [ + 312, + 612, + 486, + 624 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 312, + 612, + 486, + 624 + ], + "type": "inline_equation", + "content": "H_{p}" + }, + { + "bbox": [ + 312, + 612, + 486, + 624 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 312, + 624, + 469, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 624, + 469, + 636 + ], + "spans": [ + { + "bbox": [ + 312, + 624, + 469, + 636 + ], + "type": "text", + "content": "8: " + }, + { + "bbox": [ + 312, + 624, + 469, + 636 + ], + "type": "inline_equation", + "content": "P_{\\text{sample}} = \\operatorname{concat}(P_{\\text{sample}}, P_n)" + }, + { + "bbox": [ + 312, + 624, + 469, + 636 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 312, + 637, + 356, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 637, + 356, + 646 + ], + "spans": [ + { + "bbox": [ + 312, + 637, + 356, + 646 + ], + "type": "text", + "content": "9: end for" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 309, + 647, + 544, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 647, + 544, + 670 + ], + "spans": [ + { + "bbox": [ + 309, + 647, + 544, + 670 + ], + "type": "text", + "content": "10: Re-map " + }, + { + "bbox": [ + 309, + 647, + 544, + 670 + ], + "type": "inline_equation", + "content": "M_{\\text{sample}}" + }, + { + "bbox": [ + 309, + 647, + 544, + 670 + ], + "type": "text", + "content": " to 2D code map according to " + }, + { + "bbox": [ + 309, + 647, + 544, + 670 + ], + "type": "inline_equation", + "content": "P_{\\text{sample}}" + }, + { + "bbox": [ + 309, + 647, + 544, + 670 + ], + "type": "text", + "content": " and the rest are filled with the mask code;" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 309, + 671, + 473, + 683 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 671, + 473, + 683 + ], + "spans": [ + { + "bbox": [ + 309, + 671, + 473, + 683 + ], + "type": "text", + "content": "11: Decode the code map to the image " + }, + { + "bbox": [ + 309, + 671, + 473, + 683 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 309, + 684, + 451, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 684, + 451, + 696 + ], + "spans": [ + { + "bbox": [ + 309, + 684, + 451, + 696 + ], + "type": "text", + "content": "12: return The generated image " + }, + { + "bbox": [ + 309, + 684, + 451, + 696 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + } + ] + } + ], + "index": 35 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2006" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 70, + 146, + 167 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 146, + 167 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 146, + 167 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 146, + 167 + ], + "type": "image", + "image_path": "d642cc09f5748e1d217765a7f04920d0f50e2599648c0ccd33a5e27be05d3a92.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 149, + 89, + 199, + 104 + ], + "lines": [ + { + "bbox": [ + 149, + 89, + 199, + 104 + ], + "spans": [ + { + "bbox": [ + 149, + 89, + 199, + 104 + ], + "type": "text", + "content": "goldfish, Carassius auratus" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 152, + 133, + 196, + 156 + ], + "lines": [ + { + "bbox": [ + 152, + 133, + 196, + 156 + ], + "spans": [ + { + "bbox": [ + 152, + 133, + 196, + 156 + ], + "type": "text", + "content": "common iguna, iguana, Iguania iguna" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 202, + 72, + 346, + 168 + ], + "blocks": [ + { + "bbox": [ + 202, + 72, + 346, + 168 + ], + "lines": [ + { + "bbox": [ + 202, + 72, + 346, + 168 + ], + "spans": [ + { + "bbox": [ + 202, + 72, + 346, + 168 + ], + "type": "image", + "image_path": "bf756d23b791a541459c31aca5ae0bd7b40c47c40b7268dea3a9b82b6334bc3b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 50, + 176, + 541, + 188 + ], + "lines": [ + { + "bbox": [ + 50, + 176, + 541, + 188 + ], + "spans": [ + { + "bbox": [ + 50, + 176, + 541, + 188 + ], + "type": "text", + "content": "Figure 3. Left: Our unconditional generated images on FFHQ benchmark. Right: Our class-conditional generated images on ImageNet." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 397, + 72, + 544, + 167 + ], + "blocks": [ + { + "bbox": [ + 353, + 93, + 390, + 152 + ], + "lines": [ + { + "bbox": [ + 353, + 93, + 390, + 152 + ], + "spans": [ + { + "bbox": [ + 353, + 93, + 390, + 152 + ], + "type": "text", + "content": "lorikeet curly-coated retriever" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 397, + 72, + 544, + 167 + ], + "lines": [ + { + "bbox": [ + 397, + 72, + 544, + 167 + ], + "spans": [ + { + "bbox": [ + 397, + 72, + 544, + 167 + ], + "type": "image", + "image_path": "072aba8ed82ac3f6b31ed191945af875d1a622daa5b2ff4298f6a453a2b6f98d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 208, + 123, + 221 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 208, + 123, + 221 + ], + "spans": [ + { + "bbox": [ + 47, + 208, + 123, + 221 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 227, + 173, + 240 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 227, + 173, + 240 + ], + "spans": [ + { + "bbox": [ + 47, + 227, + 173, + 240 + ], + "type": "text", + "content": "4.1. Experimental Settings" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 245, + 286, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 245, + 286, + 293 + ], + "spans": [ + { + "bbox": [ + 46, + 245, + 286, + 293 + ], + "type": "text", + "content": "**Benchmarks.** We validate our model for unconditional, class-conditional, and text-conditional image generation tasks on FFHQ [22], ImageNet [7], and MS-COCO [30] benchmarks respectively, with " + }, + { + "bbox": [ + 46, + 245, + 286, + 293 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 46, + 245, + 286, + 293 + ], + "type": "text", + "content": " image resolution." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 293, + 286, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 293, + 286, + 376 + ], + "spans": [ + { + "bbox": [ + 46, + 293, + 286, + 376 + ], + "type": "text", + "content": "Metrics. Following previous works [13,28,42], the standard Frechet Inception Distance (FID) [16] is adopted for evaluating the generation and reconstruction quality (denoted as rFID). Inception Score (IS) [2] is also used for class-conditional generation on the ImageNet benchmark. FID and IS are calculated by sampling 50k images. rFID is calculated over the entire test set." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 377, + 286, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 377, + 286, + 568 + ], + "spans": [ + { + "bbox": [ + 46, + 377, + 286, + 568 + ], + "type": "text", + "content": "Implemented details. The architecture of MQ-VAE exactly follows [13] except for the proposed mask and demask modules, with the codebook size of " + }, + { + "bbox": [ + 46, + 377, + 286, + 568 + ], + "type": "inline_equation", + "content": "K = 1024" + }, + { + "bbox": [ + 46, + 377, + 286, + 568 + ], + "type": "text", + "content": ". For the de-mask module, the sub-module number " + }, + { + "bbox": [ + 46, + 377, + 286, + 568 + ], + "type": "inline_equation", + "content": "H = 8" + }, + { + "bbox": [ + 46, + 377, + 286, + 568 + ], + "type": "text", + "content": ". For the Stackformer, we implement two settings: a small version uses 18 transformer encoder blocks for the CodeTransformer and another 6 transformer encoder blocks for the Position-Transformer with a total of 307M parameters, and a base version uses 36 transformer encoder blocks for the Code-Transformer and another 12 transformer encoder blocks for the Position-Transformer with a total of 607M parameters to further demonstrate our scalability. The generation results are reported with a " + }, + { + "bbox": [ + 46, + 377, + 286, + 568 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 46, + 377, + 286, + 568 + ], + "type": "text", + "content": " mask ratio at " + }, + { + "bbox": [ + 46, + 377, + 286, + 568 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 46, + 377, + 286, + 568 + ], + "type": "text", + "content": " resolution feature map using eight RTX-3090 GPUs. Top-k and top-p sampling are used to report the best performance. More details could be found in the supplementary." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 575, + 266, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 575, + 266, + 588 + ], + "spans": [ + { + "bbox": [ + 47, + 575, + 266, + 588 + ], + "type": "text", + "content": "4.2. Comparison with state-of-the-art methods" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "content": "Unconditional generation. We first compare with million-level state-of-the-art autoregressive models in Table 1. Our model significantly outperforms other autoregressive models with the same parameters (307M). With more parameters, we further increase the FID from 6.84 to 5.67, which demonstrates our scalability. We also compare with other types of unconditional state-of-the-art and large-scale big models in Table 3, where we also achieve top-level performance. Our qualitative unconditional generation results are shown on the left of Figure 3." + } + ] + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 334, + 207, + 518, + 291 + ], + "blocks": [ + { + "bbox": [ + 334, + 207, + 518, + 291 + ], + "lines": [ + { + "bbox": [ + 334, + 207, + 518, + 291 + ], + "spans": [ + { + "bbox": [ + 334, + 207, + 518, + 291 + ], + "type": "table", + "html": "
Methods#ParamsFID↓
DCT [32]738M13.06
VQGAN [13](72.1+307)M11.4
RQ-Transformer [28](100+355)M10.38
Mo-VQGAN [44](82.7+307)M8.52
Stackformer(44.4+307)M6.84
Stackformer(44.4+607)M5.67
", + "image_path": "54235b9c6bb43ce6c5332aaa520b5bc394987c00d70b5abaf7ed83d1e45b4365.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 307, + 349, + 552, + 469 + ], + "blocks": [ + { + "bbox": [ + 305, + 302, + 545, + 335 + ], + "lines": [ + { + "bbox": [ + 305, + 302, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 305, + 302, + 545, + 335 + ], + "type": "text", + "content": "Table 1. Comparison of autoregressive unconditional generation at million-level parameters on FFHQ [22] benchmark. #Params splits in (VAE + autoregressive model)." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 349, + 552, + 469 + ], + "lines": [ + { + "bbox": [ + 307, + 349, + 552, + 469 + ], + "spans": [ + { + "bbox": [ + 307, + 349, + 552, + 469 + ], + "type": "table", + "html": "
Model TypeMethods#ParamsFID↓IS↑
DiffusionADM [9]554M10.94101.0
GANBigGAN [3]164M7.53168.6
GANBigGAN-deep [3]112M6.84203.6
BidirectionMaskGIT [4]227M6.18182.1
AutoregressiveDCT [32]738M36.5n/a
AutoregressiveVQ-GAN† [13]679M17.0376.85
AutoregressiveRQ-Transformer [28]821M13.11104.3
AutoregressiveMo-VQGAN [44]383M7.13138.3
AutoregressiveStackformer651M6.04172.6
", + "image_path": "f8dcdbbfd10bd6b688fe718768bf830fea69c6c28547b1cf1774a23e5446afe6.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_body" + } + ], + "index": 16 + }, + { + "type": "table", + "bbox": [ + 316, + 525, + 536, + 609 + ], + "blocks": [ + { + "bbox": [ + 305, + 477, + 545, + 511 + ], + "lines": [ + { + "bbox": [ + 305, + 477, + 545, + 511 + ], + "spans": [ + { + "bbox": [ + 305, + 477, + 545, + 511 + ], + "type": "text", + "content": "Table 2. Comparison of class-conditional image generation at million-level parameters without rejection sampling on ImageNet [7]. " + }, + { + "bbox": [ + 305, + 477, + 545, + 511 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 305, + 477, + 545, + 511 + ], + "type": "text", + "content": " denotes the model we train with the same setup with ours." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 525, + 536, + 609 + ], + "lines": [ + { + "bbox": [ + 316, + 525, + 536, + 609 + ], + "spans": [ + { + "bbox": [ + 316, + 525, + 536, + 609 + ], + "type": "table", + "html": "
Model TypeMethods#ParamsFID↓
VAEVDVAE [6]115M28.5
DiffusionImageBART [12]3.5B9.57
GANStyleGAN2 [23]-3.8
GANBigGAN [3]164M12.4
AutoregressiveViT-VQGAN [42]2.2B5.3
AutoregressiveStackformer651M5.67
", + "image_path": "2fec8a8f2405bd0701686216a132af3b52e9b1a51f9a86ddb5d391fbbfe37cd1.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "table_body" + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 620, + 545, + 653 + ], + "lines": [ + { + "bbox": [ + 305, + 620, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 305, + 620, + 545, + 653 + ], + "type": "text", + "content": "Table 3. Comparison with other types of state-of-the-art generative models and large-scale billion-level parameters autoregressive models on unconditional FFHQ [22] benchmark." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 677, + 545, + 713 + ], + "type": "text", + "content": "Class-conditional generation. We first compare with million-level state-of-the-art in Table 2. We achieve the best FID score compared to all types of models including the re" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2007" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 70, + 294, + 269 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 294, + 269 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 294, + 269 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 294, + 269 + ], + "type": "image", + "image_path": "a07faec2fbeb3939e0a08df57afd59a032a863a4f9c4f0134bec02159a88153f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 276, + 547, + 300 + ], + "lines": [ + { + "bbox": [ + 46, + 276, + 547, + 300 + ], + "spans": [ + { + "bbox": [ + 46, + 276, + 547, + 300 + ], + "type": "text", + "content": "Figure 4. The visualization of our adaptive mask module which learns to mask unimportant regions on ImageNet [7]. In the importance map, red denotes high scores while blue denotes low scores." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 298, + 70, + 545, + 269 + ], + "blocks": [ + { + "bbox": [ + 298, + 70, + 545, + 269 + ], + "lines": [ + { + "bbox": [ + 298, + 70, + 545, + 269 + ], + "spans": [ + { + "bbox": [ + 298, + 70, + 545, + 269 + ], + "type": "image", + "image_path": "24e037fbf3b3c191e4aff33a020cddf85ea302eb6c003376e6ead38253a73968.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 48, + 317, + 293, + 403 + ], + "blocks": [ + { + "bbox": [ + 48, + 317, + 293, + 403 + ], + "lines": [ + { + "bbox": [ + 48, + 317, + 293, + 403 + ], + "spans": [ + { + "bbox": [ + 48, + 317, + 293, + 403 + ], + "type": "table", + "html": "
Model TypeMethods#ParamsFID↓IS↑
DiffusionImageBART [12]3.5B21.1961.6
AutoregressiveVQVAE2 [35]13.5B3145
AutoregressiveVQ-GAN [13]1.4B15.7878.3
AutoregressiveViT-VQGAN [42]2.2B4.17175.1
AutoregressiveRQ-Transformer [28]3.8B7.55134
AutoregressiveStackformer651M6.04172.6
", + "image_path": "4c09514718ffa612ecdcfc91ad123a493f179776149ad02b443849eaeb6e86b7.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 86, + 457, + 250, + 552 + ], + "blocks": [ + { + "bbox": [ + 46, + 411, + 288, + 445 + ], + "lines": [ + { + "bbox": [ + 46, + 411, + 288, + 445 + ], + "spans": [ + { + "bbox": [ + 46, + 411, + 288, + 445 + ], + "type": "text", + "content": "Table 4. Comparison between our model and large-scale billion-level parameters models of class-conditional generation without rejection sampling on ImageNet [7] benchmark." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 86, + 457, + 250, + 552 + ], + "lines": [ + { + "bbox": [ + 86, + 457, + 250, + 552 + ], + "spans": [ + { + "bbox": [ + 86, + 457, + 250, + 552 + ], + "type": "table", + "html": "
Model TypeMethodFID↓
GANDMGAN [45]32.64
GANXMCGAN† [43]50.08
GANDFGAN [38]21.42
GANSSA-GAN [29]19.37
GANDSE-GAN [18]15.30
DiffusionVQ-Diffusion [15]19.75
AutoregressiveVQ-GAN† [13]22.28
AutoregressiveStackformer10.08
", + "image_path": "3cb8bce118f68f7f5628312c1b63b0877582f4d68225fd239ecca2329559299d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 616, + 287, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 616, + 287, + 688 + ], + "spans": [ + { + "bbox": [ + 46, + 616, + 287, + 688 + ], + "type": "text", + "content": "cent Mo-VQGAN [44] and RQ-Transformer [28]. We also compare our million-level model with existing billion-level big models in Table 4, where we also achieve top performance with fewer parameters and is only inferior to ViT-VQGAN big model. Our qualitative class-conditional generation results are shown on the right of Figure 3." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "Text-conditional generation. We compare with existing text-conditional state-of-the-art without extra web-scale" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 307, + 317, + 547, + 435 + ], + "blocks": [ + { + "bbox": [ + 46, + 560, + 287, + 594 + ], + "lines": [ + { + "bbox": [ + 46, + 560, + 287, + 594 + ], + "spans": [ + { + "bbox": [ + 46, + 560, + 287, + 594 + ], + "type": "text", + "content": "Table 5. Comparison of text-conditional generation on MS-COCO [30] without using extra web-scale data or pre-trained models. † denotes reproduced results under our same experimental setting." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 317, + 547, + 435 + ], + "lines": [ + { + "bbox": [ + 307, + 317, + 547, + 435 + ], + "spans": [ + { + "bbox": [ + 307, + 317, + 547, + 435 + ], + "type": "table", + "html": "
mask ratiofmask typerFID↓FID↓usage↑ (%)
0%32-8.113.570.02
0%16-4.4611.463.89
10%16adaptive4.557.8172.34
25%16adaptive4.797.6778.22
25%16random6.1312.2167.48
50%16adaptive5.318.3684.29
50%16random7.85515.6769.04
75%16adaptive7.6211.7187.60
75%16random10.5817.6270.41
", + "image_path": "799f74514619ec0893a15b0480df1ea68d6c532a17136acd0a0b1a3e03ddb661.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 445, + 546, + 479 + ], + "lines": [ + { + "bbox": [ + 305, + 445, + 546, + 479 + ], + "spans": [ + { + "bbox": [ + 305, + 445, + 546, + 479 + ], + "type": "text", + "content": "Table 6. Ablations of adaptive mask module on FFHQ. Here " + }, + { + "bbox": [ + 305, + 445, + 546, + 479 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 305, + 445, + 546, + 479 + ], + "type": "text", + "content": " is the downsampling factor. The codebook usage is calculated as the percentage of used codes over the entire test dataset." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 502, + 545, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 502, + 545, + 526 + ], + "spans": [ + { + "bbox": [ + 305, + 502, + 545, + 526 + ], + "type": "text", + "content": "data or pretrained models on MS-COCO [30] for fair comparison in Table.5. We achieve " + }, + { + "bbox": [ + 305, + 502, + 545, + 526 + ], + "type": "inline_equation", + "content": "18.6\\%" + }, + { + "bbox": [ + 305, + 502, + 545, + 526 + ], + "type": "text", + "content": " FID improvement." + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 307, + 536, + 545, + 586 + ], + "blocks": [ + { + "bbox": [ + 307, + 536, + 545, + 586 + ], + "lines": [ + { + "bbox": [ + 307, + 536, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 307, + 536, + 545, + 586 + ], + "type": "image", + "image_path": "ab2290e9ed4f0135bfecae58bf9f3a4ae106fef5e63f31e24865f5a475bfa0e7.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 310, + 594, + 539, + 605 + ], + "lines": [ + { + "bbox": [ + 310, + 594, + 539, + 605 + ], + "spans": [ + { + "bbox": [ + 310, + 594, + 539, + 605 + ], + "type": "text", + "content": "Figure 5. The PCA of learned codebook (1024 codebook size)." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 622, + 373, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 622, + 373, + 633 + ], + "spans": [ + { + "bbox": [ + 306, + 622, + 373, + 633 + ], + "type": "text", + "content": "4.3. Ablations" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 641, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 641, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 305, + 641, + 545, + 689 + ], + "type": "text", + "content": "We conduct ablations on " + }, + { + "bbox": [ + 305, + 641, + 545, + 689 + ], + "type": "inline_equation", + "content": "16 \\times 16" + }, + { + "bbox": [ + 305, + 641, + 545, + 689 + ], + "type": "text", + "content": " resolution feature map using four RTX-3090 GPUs for saving computation resources and the experimental trends are all consistent with " + }, + { + "bbox": [ + 305, + 641, + 545, + 689 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 305, + 641, + 545, + 689 + ], + "type": "text", + "content": " resolution feature map of the main results." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "content": "Ablations of adaptive mask module. As shown in Table 6, our proposed learned adaptive mask mechanism sig-" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2008" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 164, + 160 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 164, + 160 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 164, + 160 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 164, + 160 + ], + "type": "image", + "image_path": "8e5d0ff3c2401bbac153b064e0c2a2cdad968303d2db7b3843890b39e81654ae.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 172, + 71, + 286, + 159 + ], + "blocks": [ + { + "bbox": [ + 172, + 71, + 286, + 159 + ], + "lines": [ + { + "bbox": [ + 172, + 71, + 286, + 159 + ], + "spans": [ + { + "bbox": [ + 172, + 71, + 286, + 159 + ], + "type": "image", + "image_path": "f64343d27867bc681c2c5f8e5ef7b145c4133d0bbdf180b79fd8275f6c796003.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 50, + 175, + 286, + 262 + ], + "blocks": [ + { + "bbox": [ + 78, + 165, + 264, + 174 + ], + "lines": [ + { + "bbox": [ + 78, + 165, + 264, + 174 + ], + "spans": [ + { + "bbox": [ + 78, + 165, + 264, + 174 + ], + "type": "text", + "content": "(a) Training and validation curves of VQGAN and our Stackformer" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 50, + 175, + 286, + 262 + ], + "lines": [ + { + "bbox": [ + 50, + 175, + 286, + 262 + ], + "spans": [ + { + "bbox": [ + 50, + 175, + 286, + 262 + ], + "type": "image", + "image_path": "d5458069c96a0984b11dafb0fee346aa576ad4748396de6b9a3f0bc4cbc7f961.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 75, + 266, + 262, + 283 + ], + "lines": [ + { + "bbox": [ + 75, + 266, + 262, + 283 + ], + "spans": [ + { + "bbox": [ + 75, + 266, + 262, + 283 + ], + "type": "text", + "content": "(b) The sampling speed of Stackformer and VQGAN according to batch size and mask radio" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 294, + 287, + 317 + ], + "lines": [ + { + "bbox": [ + 46, + 294, + 287, + 317 + ], + "spans": [ + { + "bbox": [ + 46, + 294, + 287, + 317 + ], + "type": "text", + "content": "Figure 6. Comparison of training & validation curves and sample speed between VQGAN [13] and Stackformer." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 339, + 287, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 339, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 46, + 339, + 287, + 376 + ], + "type": "text", + "content": "nificantly outperforms the random one, which quantitatively validates that the adaptive mask module enables the model to learn perceptually important regions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 376, + 287, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 376, + 287, + 496 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 287, + 496 + ], + "type": "text", + "content": "Our model with " + }, + { + "bbox": [ + 46, + 376, + 287, + 496 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 376, + 287, + 496 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 376, + 287, + 496 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 46, + 376, + 287, + 496 + ], + "type": "text", + "content": " mask radio has only slightly inferior reconstruction compared with VQGAN [13], but achieves a significant improvement in generation quality, which indicates that the effectiveness of focusing autoregressive models on modeling important regions. When we further increase the mask radio to " + }, + { + "bbox": [ + 46, + 376, + 287, + 496 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 46, + 376, + 287, + 496 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 376, + 287, + 496 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 46, + 376, + 287, + 496 + ], + "type": "text", + "content": ", the final generation quality drops, we believe the reason lies that an improper high mask radio will inevitably mask some important regions that greatly decrease the reconstruction results and hinder autoregressive modeling." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 497, + 287, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 497, + 287, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 497, + 287, + 640 + ], + "type": "text", + "content": "The redundancy of the existing learned codebook can be verified from two aspects: i) the PCA of the learned codebook in Figure 5, where each point is a code and closer codes have more similar semantics. We show many codes in VQGAN's codebook overlap, which indicates these codes have nearly the same semantics and are thus redundant. The redundancy increase (more overlaps) when VQGAN uses more code to represent images (smaller downsampling factor " + }, + { + "bbox": [ + 46, + 497, + 287, + 640 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 497, + 287, + 640 + ], + "type": "text", + "content": "). The redundancy is largely alleviated in our MQ-VAE. ii) in Table 6, a higher codebook usage indicates more \"useful\" codes in the codebook and thus less redundant. VQGAN has a lower usage compared with our MQ-VAE." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 713 + ], + "type": "text", + "content": "We visualize the training and validation curves of VQGAN and Stackformer in Figure 6(a). Previous autoregressive models [13, 28, 42] always suffer from the overfitting problem while our Stackformer successfully gets rid of it, which indicates the better generalization of our masked discrete representation and our model." + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 309, + 70, + 543, + 148 + ], + "blocks": [ + { + "bbox": [ + 309, + 70, + 543, + 148 + ], + "lines": [ + { + "bbox": [ + 309, + 70, + 543, + 148 + ], + "spans": [ + { + "bbox": [ + 309, + 70, + 543, + 148 + ], + "type": "table", + "html": "
Model settingrFID↓FID↓
VQGAN4.4611.4
VQGAN*4.1711.02
MQ-VAE w/o de-mask7.0210.74
MQ-VAE w/ de-mask (SA)6.569.8
MQ-VAE w/ de-mask (DC-SA w/o mask update)5.848.92
MQ-VAE w/ de-mask (DC-SA w/ mask update)5.318.36
", + "image_path": "b026544233b4ed45424de1733138754e2d99920d7e3712950a5eb5e235a5eb97.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 156, + 545, + 200 + ], + "lines": [ + { + "bbox": [ + 305, + 156, + 545, + 200 + ], + "spans": [ + { + "bbox": [ + 305, + 156, + 545, + 200 + ], + "type": "text", + "content": "Table 7. Ablations of adaptive de-mask module on FFHQ. SA for self-attention and DC-SA for direction-constrained self-attention. \"VQGAN*\" is the stronger baseline, where the same numbers of SA and Resnet blocks as MQ-VAE's de-mask module are added." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 220, + 545, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 220, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 304, + 220, + 545, + 304 + ], + "type": "text", + "content": "We compare the sampling speed on a single RTX-1080Ti in Figure 6(b). Compared with VQGAN, our " + }, + { + "bbox": [ + 304, + 220, + 545, + 304 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 304, + 220, + 545, + 304 + ], + "type": "text", + "content": " mask radio model achieves " + }, + { + "bbox": [ + 304, + 220, + 545, + 304 + ], + "type": "inline_equation", + "content": "32.72\\%" + }, + { + "bbox": [ + 304, + 220, + 545, + 304 + ], + "type": "text", + "content": " quality improvement and " + }, + { + "bbox": [ + 304, + 220, + 545, + 304 + ], + "type": "inline_equation", + "content": "15.45\\%" + }, + { + "bbox": [ + 304, + 220, + 545, + 304 + ], + "type": "text", + "content": " speed improvement, while Our " + }, + { + "bbox": [ + 304, + 220, + 545, + 304 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 304, + 220, + 545, + 304 + ], + "type": "text", + "content": " mask radio model achieves " + }, + { + "bbox": [ + 304, + 220, + 545, + 304 + ], + "type": "inline_equation", + "content": "26.67\\%" + }, + { + "bbox": [ + 304, + 220, + 545, + 304 + ], + "type": "text", + "content": " quality improvement and " + }, + { + "bbox": [ + 304, + 220, + 545, + 304 + ], + "type": "inline_equation", + "content": "61.1\\%" + }, + { + "bbox": [ + 304, + 220, + 545, + 304 + ], + "type": "text", + "content": " speed improvement. Therefore, our design enables a flexible trade-off between speed and quality." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 304, + 545, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 304, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 304, + 304, + 545, + 365 + ], + "type": "text", + "content": "Finally, We visualize the learned mask in Figure 4, with a " + }, + { + "bbox": [ + 304, + 304, + 545, + 365 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 304, + 304, + 545, + 365 + ], + "type": "text", + "content": " mask ratio on " + }, + { + "bbox": [ + 304, + 304, + 545, + 365 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 304, + 304, + 545, + 365 + ], + "type": "text", + "content": " resolution feature map, which validates that our proposed adaptive mask mechanism successfully learns to preserve the perceptual important image regions, i.e., the structural and edge regions of objects." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 365, + 545, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 365, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 545, + 437 + ], + "type": "text", + "content": "Ablations of adaptive de-mask module. In Table 7, we show that MQ-VAE outperforms VQGAN and the stronger baseline (\"VQGAN*\"), which validates our effectiveness. We could conclude that the proposed direction-constrained self-attention and the mask updating mechanism both improve the reconstruction and generation quality." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 447, + 378, + 459 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 447, + 378, + 459 + ], + "spans": [ + { + "bbox": [ + 306, + 447, + 378, + 459 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 467, + 545, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 467, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 304, + 467, + 545, + 634 + ], + "type": "text", + "content": "In this study, we point out that the existing two-stage autoregressive generation paradigm ignores distinguishing the perceptual importance of different image regions, which brings redundancy that not only degrades generation quality but also decreases generation speed. We propose a novel two-stage generation paradigm with MQ-VAE and Stackformer to relieve the model from redundancy. MQ-VAE incorporates the adaptive mask module to mask redundant region features before quantization and the adaptive de-mask module to recover the original feature map after quantization. Stackformer then efficiently predict the combination of both codes and their positions. Comprehensive experiments on various types of image generation tasks validate the effectiveness and efficiency of our method." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 646, + 415, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 646, + 415, + 659 + ], + "spans": [ + { + "bbox": [ + 306, + 646, + 415, + 659 + ], + "type": "text", + "content": "6. Acknowledgments" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "type": "text", + "content": "This work is supported by National Natural Science Foundation of China under Grants 62222212 and U19A2057, Science Fund for Creative Research Groups under Grant 62121002." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2009" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 286, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 286, + 122 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 286, + 122 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 286, + 122 + ], + "type": "text", + "content": "[1] Hangbo Bao, Li Dong, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021.3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 125, + 286, + 146 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 286, + 146 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 286, + 146 + ], + "type": "text", + "content": "[2] Shane Barratt and Rishi Sharma. A note on the inception score. arXiv preprint arXiv:1801.01973, 2018. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 148, + 286, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 286, + 179 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 286, + 179 + ], + "type": "text", + "content": "[3] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. arXiv preprint arXiv:1809.11096, 2018. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 181, + 286, + 234 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 181, + 286, + 234 + ], + "spans": [ + { + "bbox": [ + 54, + 181, + 286, + 234 + ], + "type": "text", + "content": "[4] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T Freeman. Maskgit: Masked generative image transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11315-11325, 2022. 3, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 236, + 286, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 236, + 286, + 280 + ], + "spans": [ + { + "bbox": [ + 54, + 236, + 286, + 280 + ], + "type": "text", + "content": "[5] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In International Conference on Machine Learning, pages 1691-1703. PMLR, 2020. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 281, + 286, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 281, + 286, + 312 + ], + "spans": [ + { + "bbox": [ + 53, + 281, + 286, + 312 + ], + "type": "text", + "content": "[6] Rewon Child. Very deep vaes generalize autoregressive models and can outperform them on images. arXiv preprint arXiv:2011.10650, 2020. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 315, + 286, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 315, + 286, + 358 + ], + "spans": [ + { + "bbox": [ + 53, + 315, + 286, + 358 + ], + "type": "text", + "content": "[7] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 360, + 286, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 360, + 286, + 402 + ], + "spans": [ + { + "bbox": [ + 53, + 360, + 286, + 402 + ], + "type": "text", + "content": "[8] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 403, + 286, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 403, + 286, + 435 + ], + "spans": [ + { + "bbox": [ + 53, + 403, + 286, + 435 + ], + "type": "text", + "content": "[9] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in Neural Information Processing Systems, 34, 2021. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 437, + 286, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 437, + 286, + 491 + ], + "spans": [ + { + "bbox": [ + 48, + 437, + 286, + 491 + ], + "type": "text", + "content": "[10] Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, et al. Cogview: Mastering text-to-image generation via transformers. Advances in Neural Information Processing Systems, 34, 2021. 1, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 492, + 286, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 492, + 286, + 557 + ], + "spans": [ + { + "bbox": [ + 48, + 492, + 286, + 557 + ], + "type": "text", + "content": "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020.3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 559, + 286, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 286, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 286, + 612 + ], + "type": "text", + "content": "[12] Patrick Esser, Robin Rombach, Andreas Blattmann, and Bjorn Ommer. Imagebart: Bidirectional context with multinomial diffusion for autoregressive image synthesis. Advances in Neural Information Processing Systems, 34, 2021. 1, 3, 6, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 614, + 286, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 286, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 286, + 667 + ], + "type": "text", + "content": "[13] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12873-12883, 2021. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 670, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 286, + 713 + ], + "type": "text", + "content": "[14] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 1, 3" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[15] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. arXiv preprint arXiv:2111.14822, 2021. 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 118, + 545, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 118, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 308, + 118, + 545, + 171 + ], + "type": "text", + "content": "[16] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 173, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 173, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 308, + 173, + 545, + 205 + ], + "type": "text", + "content": "[17] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 206, + 545, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 206, + 545, + 259 + ], + "spans": [ + { + "bbox": [ + 308, + 206, + 545, + 259 + ], + "type": "text", + "content": "[18] Mengqi Huang, Zhendong Mao, Penghui Wang, Quan Wang, and Yongdong Zhang. Dse-gan: Dynamic semantic evolution generative adversarial network for text-to-image generation. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4345-4354, 2022. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 261, + 545, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 261, + 545, + 314 + ], + "spans": [ + { + "bbox": [ + 308, + 261, + 545, + 314 + ], + "type": "text", + "content": "[19] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 316, + 545, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 316, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 308, + 316, + 545, + 348 + ], + "type": "text", + "content": "[20] Nikil Jayant, James Johnston, and Robert Safranek. Signal compression based on models of human perception. Proceedings of the IEEE, 81(10):1385-1422, 1993. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 350, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 350, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 308, + 350, + 545, + 392 + ], + "type": "text", + "content": "[21] Biing-Hwang Juang and A Gray. Multiple stage vector quantization for speech coding. In ICASSP'82. IEEE International Conference on Acoustics, Speech, and Signal Processing, volume 7, pages 597-600. IEEE, 1982. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 394, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 394, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 308, + 394, + 545, + 437 + ], + "type": "text", + "content": "[22] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 438, + 545, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 438, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 308, + 438, + 545, + 491 + ], + "type": "text", + "content": "[23] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 493, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 493, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 308, + 493, + 545, + 514 + ], + "type": "text", + "content": "[24] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 515, + 545, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 515, + 545, + 558 + ], + "spans": [ + { + "bbox": [ + 308, + 515, + 545, + 558 + ], + "type": "text", + "content": "[25] M Kocher and M Kunt. A contour-texture approach to picture coding. In ICASSP'82. IEEE International Conference on Acoustics, Speech, and Signal Processing, volume 7, pages 436-439. IEEE, 1982. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 559, + 545, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 559, + 545, + 591 + ], + "spans": [ + { + "bbox": [ + 308, + 559, + 545, + 591 + ], + "type": "text", + "content": "[26] Murat Kunt, Athanassios Ikonomopoulos, and Michel Kocher. Second-generation image-coding techniques. Proceedings of the IEEE, 73(4):549-574, 1985. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 593, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 593, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 308, + 593, + 545, + 667 + ], + "type": "text", + "content": "[27] Christian Ledig, Lucas Theis, Ferenc Huszár, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photorealistic single image super-resolution using a generative adversarial network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4681-4690, 2017. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "type": "text", + "content": "[28] Doyup Lee, Chiheon Kim, Saehoon Kim, Minsu Cho, and Wook-Shin Han. Autoregressive image generation using residual quantization. arXiv preprint arXiv:2203.01941, 2022. 1, 2, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2010" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 715 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[29] Wentong Liao, Kai Hu, Michael Ying Yang, and Bodo Rosenhahn. Text to image generation with semantic-spatial aware gan. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18187-18196, 2022. 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 287, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 184 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 184 + ], + "type": "text", + "content": "[30] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 185, + 287, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 185, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 185, + 287, + 217 + ], + "type": "text", + "content": "[31] Julieta Martinez, Holger H Hoos, and James J Little. Stacked quantizers for compositional vector compression. arXiv preprint arXiv:1411.2173, 2014. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 219, + 287, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 219, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 48, + 219, + 287, + 251 + ], + "type": "text", + "content": "[32] Charlie Nash, Jacob Menick, Sander Dieleman, and Peter W Battaglia. Generating images with sparse representations. arXiv preprint arXiv:2103.03841, 2021. 3, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 253, + 287, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 253, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 48, + 253, + 287, + 297 + ], + "type": "text", + "content": "[33] Niki Parmar, Ashish Vaswani, Jakob Uszkoreit, Lukasz Kaiser, Noam Shazeer, Alexander Ku, and Dustin Tran. Image transformer. In International Conference on Machine Learning, pages 4055-4064. PMLR, 2018. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 298, + 287, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 298, + 287, + 352 + ], + "spans": [ + { + "bbox": [ + 48, + 298, + 287, + 352 + ], + "type": "text", + "content": "[34] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 1, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 354, + 287, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 354, + 287, + 387 + ], + "spans": [ + { + "bbox": [ + 48, + 354, + 287, + 387 + ], + "type": "text", + "content": "[35] Ali Razavi, Aaron Van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. Advances in neural information processing systems, 32, 2019. 1, 3, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 388, + 287, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 287, + 421 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 287, + 421 + ], + "type": "text", + "content": "[36] Danilo Rezende and Shakir Mohamed. Variational inference with normalizing flows. In International conference on machine learning, pages 1530-1538. PMLR, 2015. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 422, + 287, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 422, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 48, + 422, + 287, + 464 + ], + "type": "text", + "content": "[37] Woncheol Shin, Gyubok Lee, Jiyoung Lee, Joonseok Lee, and Edward Choi. Translation-equivariant image quantizer for bi-directional image-text generation. arXiv preprint arXiv:2112.00384, 2021. 1, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 467, + 287, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 467, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 48, + 467, + 287, + 510 + ], + "type": "text", + "content": "[38] Ming Tao, Hao Tang, Songsong Wu, Nicu Sebe, Xiao-Yuan Jing, Fei Wu, and Bingkun Bao. Df-gan: Deep fusion generative adversarial networks for text-to-image synthesis. arXiv preprint arXiv:2008.05865, 2020. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 512, + 287, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 512, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 48, + 512, + 287, + 544 + ], + "type": "text", + "content": "[39] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 1, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 545, + 287, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 545, + 287, + 589 + ], + "spans": [ + { + "bbox": [ + 48, + 545, + 287, + 589 + ], + "type": "text", + "content": "[40] Aaron Van Oord, Nal Kalchbrenner, and Koray Kavukcuoglu. Pixel recurrent neural networks. In International conference on machine learning, pages 1747-1756. PMLR, 2016. 1, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 590, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 635 + ], + "type": "text", + "content": "[41] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "text", + "content": "[42] Jiahui Yu, Xin Li, Jing Yu Koh, Han Zhang, Ruoming Pang, James Qin, Alexander Ku, Yuanzhong Xu, Jason Baldridge, and Yonghui Wu. Vector-quantized image modeling with improved vqgan. arXiv preprint arXiv:2110.04627, 2021. 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 691, + 287, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 715 + ], + "type": "text", + "content": "[43] Han Zhang, Jing Yu Koh, Jason Baldridge, Honglak Lee, and Yinfei Yang. Cross-modal contrastive learning for text-to" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 206 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "text", + "content": "image generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 833-842, 2021. 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "text", + "content": "[44] Chuanxia Zheng, Long Tung Vuong, Jianfei Cai, and Dinh Phung. Movq: Modulating quantized vectors for high-fidelity image generation. arXiv preprint arXiv:2209.09002, 2022.6,7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 152, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 206 + ], + "type": "text", + "content": "[45] Minfeng Zhu, Pingbo Pan, Wei Chen, and Yi Yang. Dm-gan: Dynamic memory generative adversarial networks for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5802-5810, 2019. 7" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "2011" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Novel Class Discovery for 3D Point Cloud Semantic Segmentation/b9953b69-6d05-422e-9f85-28806509ef58_content_list.json b/2023/Novel Class Discovery for 3D Point Cloud Semantic Segmentation/b9953b69-6d05-422e-9f85-28806509ef58_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..da613c6803fb26d98edb3dc5de899ef5017d264e --- /dev/null +++ b/2023/Novel Class Discovery for 3D Point Cloud Semantic Segmentation/b9953b69-6d05-422e-9f85-28806509ef58_content_list.json @@ -0,0 +1,1268 @@ +[ + { + "type": "text", + "text": "Novel Class Discovery for 3D Point Cloud Semantic Segmentation", + "text_level": 1, + "bbox": [ + 153, + 94, + 815, + 116 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Luigi Riz $^{1}$ Cristiano Saltori $^{2}$ Elisa Ricci $^{1,2}$ Fabio Poiesi $^{1}$", + "bbox": [ + 230, + 143, + 733, + 162 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Fondazione Bruno Kessler $^{2}$ University of Trento", + "bbox": [ + 313, + 170, + 656, + 186 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/db2cf5c1cac45253f849f7e282e42dd304853e5ba2874d4667d21166dc24fdba.jpg", + "image_caption": [ + "Figure 1. Novel Class Discovery for 3D point cloud semantic segmentation seeks to recognise novel classes by clustering unlabelled novel points with similar semantic features and by exploiting only the knowledge of a set of labelled samples corresponding to the base classes." + ], + "image_footnote": [], + "bbox": [ + 78, + 220, + 890, + 330 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 396, + 312, + 412 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Novel class discovery (NCD) for semantic segmentation is the task of learning a model that can segment unlabelled (novel) classes using only the supervision from labelled (base) classes. This problem has recently been pioneered for 2D image data, but no work exists for 3D point cloud data. In fact, the assumptions made for 2D are loosely applicable to 3D in this case. This paper is presented to advance the state of the art on point cloud data analysis in four directions. Firstly, we address the new problem of NCD for point cloud semantic segmentation. Secondly, we show that the transposition of the only existing NCD method for 2D semantic segmentation to 3D data is suboptimal. Thirdly, we present a new method for NCD based on online clustering that exploits uncertainty quantification to produce prototypes for pseudo-labelling the points of the novel classes. Lastly, we introduce a new evaluation protocol to assess the performance of NCD for point cloud semantic segmentation. We thoroughly evaluate our method on SemanticKITTI and SemanticPOSS datasets, showing that it can significantly outperform the baseline. Project page: https://github.com/LuigiRiz/NOPS.", + "bbox": [ + 73, + 439, + 473, + 758 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This project has received funding from the European Union's Horizon Europe research and innovation programme under grant agreement No 101058589. This work was also partially supported by the PRIN project LEGO-AI (Prot. 2020TA3K9N), the EU ISFP PROTECTOR (101034216) project and the EU H2020 MARVEL (957337) project and, it was carried out in the Vision and Learning joint laboratory of FBK and UNITN.", + "bbox": [ + 75, + 827, + 468, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 501, + 381, + 630, + 397 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As humans, we are fairly skilled in organising new visual knowledge (novelty) into homogeneous groups, even when we do not know what we are observing. However, machines cannot perform this task satisfactorily without our supervision. The challenge here is mainly in the formulation of discriminative latent representations of the real world and in the quantification of the uncertainty when the novelty is observed [16,40,41]. The work of Han et al. [16] pioneered the formulation of the Novel Class Discovery (NCD) problem by defining it as the task that aims to classify the samples of an unlabelled dataset into different classes, i.e. the novel samples, by exploiting the knowledge of a set of labelled samples, i.e. the base samples. Note that the classes in the labelled and unlabelled datasets are disjoint.", + "bbox": [ + 496, + 412, + 892, + 625 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "NCD has been explored in the 2D image domain for classification [12, 16, 41] and later for semantic segmentation [40]. In particular, Zhao et al. [40] presented the first approach for NCD for 2D semantic segmentation. Two key assumptions were made by the authors. Firstly, at most one novel class is allowed in each image. Secondly, the new class belongs to a foreground object that can be found via saliency detection (e.g. a man on a bicycle, where the bicycle is the novel class). Thanks to these assumptions, the authors could pool the features of each image into a single latent representation and cluster the representations of the whole dataset to discover clusters of novel classes. We argue that these two are important constraints that are not applicable to generic 3D data, in particular to point clouds captured with LiDAR sensors in real-world city-scale scenarios. One point cloud can contain more than one novel class, and the saliency for 3D data cannot be leveraged in the same way as that for 2D data. Although they are both", + "bbox": [ + 496, + 628, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "9393", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "related to the attraction of human fixations, 3D saliency is more related to the regional importance of 3D surfaces rather than the foreground/background distinction [30]. Our motivation in exploring NCD for the 3D setting is mainly driven by addressing these shortcomings.", + "bbox": [ + 75, + 90, + 470, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we explore the new problem of NCD for 3D point cloud semantic segmentation (see Fig. 1). Given a partially human-annotated dataset, we jointly learn base and novel semantic classes by clustering unlabelled points with similar semantic features. We adapt the method of Zhao et al. [40] (Entropy-based Uncertainty Modeling and Self-training - EUMS) for point cloud data and use it as our baseline. We go beyond their formulation and, inspired by [5], we integrate batch-level (online) clustering in our method and update prototypes during training in order to make clustering computationally tractable. Cluster assignments are then used as training pseudo-labels. We also exploit over-clustering to achieve a higher clustering accuracy as in EUMS. Because point clouds contain multiple semantic classes, we cannot guarantee that all the classes appear in the point clouds within each batch, some will be missing. Therefore, we design a queuing strategy to store important features over training time, which will be used for pseudolabelling in the case of missing classes. We further introduce a strategy for exploiting the pseudo-label uncertainty to promote the creation of reliable prototypes that we then exploit to produce higher-quality pseudo-labels. Lastly, we produce two augmented views of the same point cloud and impose pseudo-label consistency amongst them. We evaluate our approach on SemanticKITTI [3, 4, 13] and SemanticPOSS [24], introducing an evaluation protocol for NCD and point cloud segmentation that can be adopted in future works. We empirically show that our approach largely outperforms our baseline in both datasets. We also perform an extensive ablation study to demonstrate the importance of the different components of our method.", + "bbox": [ + 75, + 167, + 472, + 636 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To summarise, our contributions are:", + "text_level": 1, + "bbox": [ + 76, + 641, + 323, + 655 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We address the new problem of NCD for 3D semantic segmentation;", + "- We show that the transposition of the only existing NCD method for 2D semantic segmentation [40] to 3D data is suboptimal;", + "- We present a new method for NCD based on online clustering and uncertainty estimation, which we name it NOPS (NOvel Point Segmentation);", + "- We introduce a new evaluation protocol to assess the performance of NCD for 3D semantic segmentation." + ], + "bbox": [ + 94, + 657, + 467, + 809 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 76, + 828, + 215, + 843 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Point cloud semantic segmentation can be performed at the point level [26], on range view maps [27], and by voxelising the input points [43]. Point-level networks pro", + "bbox": [ + 75, + 854, + 468, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "cess the input without intermediate representations. Examples of these include PointNet [25], PointNet++ [26], RandLA-Net [17], and KPConv [33]. PointNet [25] and PointNet++ [26] are based on a series of multi-layer perceptron where PointNet++ introduces global and local feature aggregation at multiple scales. RandLA-Net [17] uses random sampling, attentive pooling, and local spatial encoding. KPConv [33] employs flexible and deformable convolutions in a continuous input space. Point-level networks are computationally inefficient when large-scale point clouds are processed. Range view architectures [23] and voxel-based approaches [10] are more computationally efficient than their point-level counterpart. The former requires projecting the input points on a 2D dense map, processing input maps with 2D convolutional filters [27], and reprojecting predictions to the initial 3D space. SqueezeSeg networks [35, 36], 3D-MiniNet [1], RangeNet++ [23], and PolarNet [39] are examples of this category. Although they are more efficient, these approaches tend to lose information during the projection and re-projection phase. The latter includes 3D quantisation-based approaches that discretise the input points into a 3D voxel grid and employ 3D convolutions [43] or 3D sparse convolutions [10, 15] to predict per-voxel classes. VoxelNet [43], SparseConv [14, 15], MinkowskiNet [10], Cylinder3D [44], and $(\\mathrm{AF})^2$ -S3Net [9] are architectures belonging to this category. These approaches tackle point cloud segmentation in the supervised settings, whereas we tackle novel class discovery with labelled base classes and unlabelled novel classes.", + "bbox": [ + 496, + 90, + 890, + 529 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Novel class discovery (NCD) is explored for 2D classification [12, 16, 19, 20, 28, 34, 37, 41, 42] and 2D segmentation [40]. NCD is more complex than standard semi-supervised learning [31, 32, 38]. In semi-supervised learning, labelled and unlabelled samples belong to the same classes, while in NCD, novel and base samples belong to disjoint classes. Han et al. [16] pioneered the NCD problem for 2D image classification. A classification model is pretrained on a set of base classes and used as feature extractor for the novel classes. They then train a classifier for the novel classes using the pseudo-labels produced by the pretrained model. Zhong et al. [41] introduced neighbourhood contrastive learning to generate discriminative representations for clustering. They retrieve and aggregate pseudopositive pairs with contrastive learning, encouraging the model to learn more discriminative representations. Hard negatives are obtained by mixing labelled and unlabelled samples in the feature space. UNO [12] unifies the two previous works by using a unique classification loss function for both base and novel classes, where pseudo-labels are processed together with ground-truth labels. NCD without Forgetting [20] and FRoST [28] further extend NCD to the incremental learning setting. EUMS [40] is the only approach analysing the NCD problem for semantic segmen", + "bbox": [ + 496, + 537, + 890, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "9394", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/1a61d98d4a2dd970aafcd5cb5669a50c29c66ce080f199adbb7a87f68abde09a.jpg", + "image_caption": [ + "Figure 2. Overview of NOPS. We random augment the input point cloud twice and extract point-level features $\\mathcal{F}$ with the shared model $f_{\\xi}$ . $\\mathcal{F}$ are used to obtain pseudo-labels in the online pseudo-labelling. We forward $\\mathcal{F}$ to a novel $f_{n}$ and a base $f_{b}$ segmentation layer to output the novel and base predictions, respectively. We optimise our network by minimising a global objective function based on cross entropy." + ], + "image_footnote": [], + "bbox": [ + 76, + 88, + 893, + 385 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tation. Unlike image classification, the model has to classify each pixel and handle multiple classes in each image. EUMS consists of a multi-stage pipeline using a saliency model to cluster the latent representations of novel classes to produce pseudo-labels. Moreover, entropy-based uncertainty and self-training are used to overcome noisy pseudolabels while improving the model performance on the novel classes. In this work, we tackle the problem of NCD in 3D point cloud semantic segmentation. Unlike previous works, our problem inherits the challenges from the fields of 2D semantic segmentation [7, 8] and 3D point cloud segmentation [10, 23, 29]. From 2D semantic segmentation, it inherits the additional challenges of multiple novel classes in the same image and the strong class unbalance. From 3D point cloud segmentation, we inherit the sparsity of input data, the different density of point cloud regions and the inability to identify foreground and background. The latter are not present in 2D segmentation [40]. Unlike [40] that use K-Means, we formulate clustering as an optimal transport problem to avoid degenerate solutions (i.e. all data points may be assigned to the same label and learn a constant representation) [2, 22]. Lastly, related to EUMS, REAL is proposed for open-world 3D semantic segmentation [6], where both known and unknown points have to be segmented. Unlike NOPS, all the unknown points belong to a single class and it is the task of a human annotator to separately label the novel classes. Then, these labels are used to update the base model by incrementally learning the novel classes.", + "bbox": [ + 75, + 457, + 472, + 883 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Proposed approach", + "text_level": 1, + "bbox": [ + 500, + 455, + 686, + 473 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 500, + 479, + 609, + 496 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given an input point cloud, we produce two augmented views that are processed with the same deep neural network to extract point-level features. These features are used to obtain pseudo-labels in the online pseudo-labelling step through the Sinkhorn-Knopp algorithm [11] (Sec. 3.3). Concurrently, we process the same features with the last network layers to segment novel and base classes. These features are stored in the class-balanced queue to mitigate the problem of batches with missing classes (Sec. 3.4). We exploit pseudo-label values (class probabilities) to filter out uncertain points, thus adding to the queue only high-quality points (Sec. 3.5). Lastly, we train our network by minimising the optimisation objective function through a swapped prediction task based on the computed pseudolabels (Sec. 3.6). Fig. 2 shows the block diagram of NOPS.", + "bbox": [ + 496, + 505, + 890, + 733 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Problem formulation", + "text_level": 1, + "bbox": [ + 500, + 739, + 697, + 755 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let $\\mathrm{X} = \\{\\mathcal{X}\\}$ be a dataset of 3D point clouds captured in different scenes. $\\mathcal{X}$ is a set composed of a base set $\\mathcal{X}_b$ and a novel set $\\mathcal{X}_n$ , s.t. $\\mathcal{X} = \\mathcal{X}_b \\cup \\mathcal{X}_n$ . The semantic categories that can be present in our point clouds are $\\mathcal{C} = \\mathcal{C}_b \\cup \\mathcal{C}_n$ , where $\\mathcal{C}_b$ is the set of base classes and $\\mathcal{C}_n$ is the set of novel classes, s.t. $\\mathcal{C}_b \\cap \\mathcal{C}_n = \\emptyset$ . Each $\\mathcal{X} \\in \\mathrm{X}$ is composed of a finite but unknown number of 3D points $\\mathcal{X} = \\{(\\mathbf{x}, c)\\}$ , where $\\mathbf{x} \\in \\mathbb{R}^3$ is the coordinate of the a point and $c$ is its semantic class. We know the class of the point $(\\mathbf{x}, c)$ , s.t. $\\mathbf{x} \\in \\mathcal{X}_b$ and", + "bbox": [ + 496, + 763, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "9395", + "bbox": [ + 480, + 944, + 514, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$c \\in \\mathcal{C}_b$ , but we do not know the class of the point $(\\mathbf{x}, c)$ , s.t. $x \\in \\mathcal{X}_n$ and $c \\in \\mathcal{C}_n$ . No points in $\\mathcal{X}_n$ belong to one of the base classes $\\mathcal{C}_b$ . As in [16, 40, 41], we assume that the number of classes to discover is known, i.e. $|\\mathcal{C}_n| = C_n$ . We aim to design a computational approach that trains a deep neural network $f_\\Theta$ that can segment all the points of a given point cloud, thus learning to jointly segment base classes $\\mathcal{C}_b$ and novel classes $\\mathcal{C}_n$ . $\\Theta$ are the weights of our deep neural network. $f_\\Theta$ is composed of two heads, $f_\\Theta = f_\\xi \\circ \\{f_b, f_n\\}$ , where $f_b$ is the segmentation head for the base classes, $f_n$ is the segmentation head for the novel classes, $f_\\xi$ is the feature extractor network and $\\circ$ is the composition operator (Fig. 2).", + "bbox": [ + 75, + 90, + 472, + 273 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Online pseudo-labelling", + "text_level": 1, + "bbox": [ + 76, + 284, + 295, + 301 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We formulate pseudo-labelling as the assignment of novel points to the class-prototypes learnt during training [5]. Let $\\mathsf{P} \\in \\mathbb{R}^{D \\times \\rho}$ be the class prototypes, where $D$ is the size of the output features from $f_{\\xi}$ and $\\rho$ is the number of prototypes. Let $\\mathsf{Z} \\in \\mathbb{R}^{D \\times m}$ be the normalised output features extracted from $f_{\\xi}$ , where $m$ is the number of points of the point cloud. $m$ is not known a priori and it can differ across point clouds. We aim to find the assignment $\\mathsf{Q} \\in \\mathbb{R}^{\\rho \\times m}$ s.t. all the points in the batch are equally partitioned across the $\\rho$ prototypes. This equipartition ensures that the feature representations of the points belonging to different novel classes are well-separated, thus preventing the case in which the novel class feature representations collapse into a unique solution. Caron et al. [5] employs an arbitrary large number of prototypes $\\rho$ to effectively organise the feature space produced by $f_{\\xi}$ . They discard $\\mathsf{P}$ after training. In contrast, we learn exactly $\\rho = C_n$ class prototypes and propose to use $\\mathsf{P}$ as the weights for our new class segmentation head $f_{n}$ , which outputs the $C_n$ logits for the new classes. In order to optimise the assignment $\\mathsf{Q}$ , we maximise the similarity between the features of the new points and the learned prototypes as", + "bbox": [ + 75, + 308, + 472, + 642 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\max _ {\\mathbb {Q} \\in \\mathcal {Q}} \\operatorname {T r} \\left(\\mathbb {Q} ^ {\\top} \\mathbb {P} ^ {\\top} \\mathbb {Z}\\right) + \\epsilon H (\\mathbb {Q}) \\rightarrow \\mathbb {Q} ^ {*}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 652, + 468, + 679 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $H$ is the entropy function, $\\epsilon$ is the parameter that determines the smoothness of the assignment and $\\mathbb{Q}^*$ is our sought solution. Asano et al. [2] enforce the equipartitioning constraint by requiring $\\mathbb{Q}$ to belong to a transportation polytope and perform this optimisation on the whole dataset at once (offline). This operation with point cloud data is computationally impractical. Therefore, we formulate the transportation polytope such that the optimisation is performed online, which consists of processing only the points within the batch being processed", + "bbox": [ + 75, + 690, + 468, + 843 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {Q} = \\left\\{\\mathbb {Q} \\in \\mathbb {R} _ {+} ^ {C _ {n} \\times m} | \\mathbb {Q} 1 _ {m} = \\frac {1}{C _ {n}} 1 _ {C _ {n}}, \\mathbb {Q} ^ {\\top} 1 _ {C _ {n}} = \\frac {1}{m} 1 _ {m} \\right\\}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 854, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $1_{\\star}$ represents a vector of ones of dimension $\\star$ . These constraints ensure that each class prototype is selected on average at least $m / C_n$ times in each batch. The solution $\\mathbb{Q}^*$ can take the form of a normalised exponential matrix", + "bbox": [ + 496, + 90, + 890, + 151 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {Q} ^ {*} = \\operatorname {d i a g} (\\alpha) \\exp \\left(\\frac {\\mathrm {P} ^ {\\top} \\mathrm {Z}}{\\epsilon}\\right) \\operatorname {d i a g} (\\beta), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 575, + 160, + 890, + 195 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\alpha$ and $\\beta$ are renormalization vectors that are computed iteratively with the Sinkhorn-Knopp algorithm [11, 21]. We then transpose the optimised soft assignment $\\mathbb{Q}^* \\in \\mathbb{R}_+^{C_n \\times m}$ to obtain the soft pseudo-labels for each of the $m$ novel points being processed within each batch.", + "bbox": [ + 496, + 204, + 890, + 279 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We empirically found that training can be more effective if pseudo-labels are smoother in the first training epochs and peaked in the last training epochs. Therefore, we introduce a linear decay of $\\epsilon$ during training.", + "bbox": [ + 496, + 279, + 890, + 339 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Multi-headed segmentation: A single segmentation head may converge to a suboptimal feature space, thus producing suboptimal prototype solutions. To further improve the segmentation quality, we use multiple novel class segmentation heads to optimise $f_{\\Theta}$ based on different training solutions. Different solutions increase the likelihood of producing a diverse partitioning of the feature space as they regularise with each other (they share the same backbone) [18]. In practise, we concatenate the logits of the base class segmentation head with the outputs of each novel class segmentation head and we separately evaluate their loss for each novel class segmentation head at training time.", + "bbox": [ + 496, + 339, + 892, + 520 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We task our network to over-cluster novel points, using segmentation heads that output $o \\cdot C_n$ logits, where $o$ is the over-clustering factor. Previous studies empirically showed that this is beneficial to learn more informative features [5, 12, 18, 22]. We observed the same and concur that over-clustering can be useful for increasing expressivity of the feature representations. The over-clustering heads are then discarded at inference time.", + "bbox": [ + 496, + 521, + 890, + 641 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Class-balanced queuing", + "text_level": 1, + "bbox": [ + 500, + 650, + 718, + 667 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Soft pseudo-labelling described in Sec. 3.3 produces an equipartite matching between the novel points and the class centroids. However, it is highly likely that batches are sampled with point clouds containing novel classes with different cardinalities when dealing with 3D data. It is also likely that some scenes may contain only a subset of the novel classes. Therefore, enforcing the equipartitioning constraint in each batch of the dataset could affect the learning of less-frequent (long-tail) classes. As a solution, we introduce a queue $\\mathsf{Z}_q$ containing a randomly extracted portion of the features of the novel points from the previous iterations. We use these additional data to mitigate the potential class imbalance that may occur during training. In practise, we compute $\\mathsf{Z} \\gets \\mathsf{Z} \\oplus \\mathsf{Z}_q$ , where $\\oplus$ is the concatenation operator, and execute the Sinkhorn-Knopp algorithm on this", + "bbox": [ + 496, + 674, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "9396", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "augmented version of $\\mathbb{Z}$ . Then, we retain only the pseudolabels for the first $m$ columns of $\\mathbb{Q}^*$ .", + "bbox": [ + 75, + 90, + 468, + 121 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Uncertainty-aware training and queuing", + "text_level": 1, + "bbox": [ + 76, + 135, + 424, + 152 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We propose to carefully select novel points for training $f_{\\Theta}$ with fewer but more reliable pseudo-labels and to build a more effective queue $Z_{q}$ . We perform this selection by applying a threshold to the class probabilities of the novel class pseudo-labels. We found that seeking a fixed threshold for all the novel classes, that is also compatible with the variations of the class probabilities during training, is impractical. Therefore, we employ an adaptive threshold based on the class probabilities within each batch.", + "bbox": [ + 75, + 160, + 468, + 296 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our selection strategy operates as follows. Let $\\tau_{c}$ be the adaptive threshold for the points of the novel class $c\\in \\mathcal{C}_n$ . Firstly, we extract the novel points that have the greatest class probability for the class $c$ . Secondly, we compute $\\tau_{c}$ as the $p$ -th percentile of the class probabilities of these novel points. Lastly, we retain the novel points of class $c$ whose class probability is above the threshold $\\tau_{c}$ . We define this selection strategy as the function", + "bbox": [ + 75, + 297, + 470, + 417 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\phi : \\left(\\mathcal {F} _ {n}, \\hat {\\mathcal {Y}} _ {n}\\right) \\times p \\mapsto \\left(\\bar {\\mathcal {F}} _ {n}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 431, + 468, + 450 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{F}_n$ is the set of feature vectors extracted from $f_{\\xi}$ and $\\hat{\\mathcal{V}}_n$ is the set of class probabilities predicted by the network for these points. $\\bar{\\mathcal{F}}_n$ are both processed by the Sinkhorn-Knopp algorithm to generate our pseudo-labels and added to $\\mathbb{Z}_q$ to make it more effective.", + "bbox": [ + 75, + 464, + 468, + 544 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.6. Optimisation objective", + "text_level": 1, + "bbox": [ + 76, + 556, + 287, + 571 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We optimise $f_{\\Theta}$ by using the weighted Cross Entropy objective based on the labels $\\mathcal{Y}_b$ of the base samples and the pseudo-labels $\\tilde{\\mathcal{Y}}_n$ of the novel samples. We formulate a swapped prediction task based on these pseudo-labels [5]. Secicfically, we begin by generating two different augmentations of $\\mathcal{X}$ that we define as $\\mathcal{X}'$ and $\\mathcal{X}''$ . We use the known one-hot labels for $\\mathcal{Y}_b$ and the predicted soft pseudolabels for $\\tilde{\\mathcal{Y}}_n$ . We predict the novel pseudo-labels $\\tilde{\\mathcal{Y}}_n'$ and $\\tilde{\\mathcal{Y}}_n''$ of the respective point clouds $\\mathcal{X}'$ and $\\mathcal{X}''$ with our approach. Then, we enforce prediction consistency between the swapped pseudo-labels of the two augmentations as", + "bbox": [ + 75, + 580, + 468, + 748 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} (\\mathcal {X}) = \\ell \\left(\\hat {\\mathcal {Y}} ^ {\\prime}, \\tilde {\\mathcal {Y}} ^ {\\prime \\prime}\\right) + \\ell \\left(\\hat {\\mathcal {Y}} ^ {\\prime \\prime}, \\tilde {\\mathcal {Y}} ^ {\\prime}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 760, + 468, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\hat{\\mathcal{Y}}^{\\prime} = \\hat{\\mathcal{Y}}_{b}^{\\prime}\\cup \\hat{\\mathcal{Y}}_{n}^{\\prime}$ (same for $\\hat{\\mathcal{Y}}^{\\prime \\prime}$ ), $\\tilde{\\mathcal{Y}}^{\\prime} = \\mathcal{Y}_{b}^{\\prime}\\cup \\tilde{\\mathcal{Y}}_{n}^{\\prime}$ (same for $\\tilde{\\mathcal{Y}}^{\\prime \\prime}$ ) and $\\ell$ is the weighted Cross Entropy loss. We use separate segmentation heads for base classes and novel classes. The weights of the loss for the base classes are computed based on their occurrence frequency in the training set. The weights of the loss for the novel classes are all set equally as their occurrence frequency in the dataset is unknown.", + "bbox": [ + 75, + 791, + 468, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Adapting NCD for 2D images to 3D", + "text_level": 1, + "bbox": [ + 498, + 89, + 821, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Another contribution of this work is to adapt the method proposed by Zhao et al. [40] for NCD for 2D semantic segmentation (EUMS) to 3D data. Our empirical evaluation (see Sec. 5) shows that the transposition of EUMS to the 3D domain has some limitations. In particular, as described in Sec. 1, EUMS uses two assumptions: I) the novel classes belong to the foreground and II) each image can contain at most one novel class. This allows EUMS to leverage a saliency detection model to produce a foreground mask and a segmentation model pre-trained on the base classes to determine which portion of the image is background. The portion of the image that belongs to both the foreground mask and the background mask is where features are then pooled. EUMS computes a feature representation for each image by average pooling the features of the pixels belonging to the unknown portion. The feature representations of all the images in the dataset are clustered with K-Means by using the number of classes to discover as the target number of clusters. EUMS shows that overclustering and entropy-based modelling can be exploited to improve the results. The affiliation of a point to its cluster is used to produce hard pseudo-labels that are in turn used along with the ground-truth labels to fine-tune the pre-trained model.", + "bbox": [ + 496, + 114, + 890, + 460 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With 3D point clouds, there is no concept of foreground and background (in contrast with I). Our adaptation is designed to discover the classes of all the unlabelled points (in contrast with II). Therefore, given the unlabelled points of each point cloud, we randomly extract a subset of these by setting a ratio (e.g. $30\\%$ ) with upper bound (e.g. 1K) on the number of points to select. We compute and collect their features for all the point clouds in the dataset and apply K-Means on the whole set of features. Note that this clustering step is computationally expensive, and we had to use High Performance Computing to execute it. The subsampling of the points was necessary to fit the data in the RAM (see Sec. 5 for a detailed analysis). Once the cluster prototypes are computed, we produce the hard pseudo-labels. To enrich the set of pseudo-labels, we propagate the pseudo-label of each point to its nearest neighbour in the coordinate space. This allows us to expand the subset of pseudo-labelled randomly selected points. We also implement the other steps of overclustering and entropy-based modelling to boost the results. Lastly, we fine-tune our model with these pseudo-labels. We name our transposition of EUMS as $\\mathrm{EUMS}^{\\dagger}$ and report its block diagram in the Supplementary Material.", + "bbox": [ + 496, + 462, + 892, + 794 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experimental results", + "text_level": 1, + "bbox": [ + 498, + 806, + 699, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Experimental setup", + "text_level": 1, + "bbox": [ + 498, + 832, + 687, + 848 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We evaluate our approach on SemanticKITTI [3, 4,13] and SemanticPOSS [24]. SemanticKITTI [4] consists of 43,552 point cloud acquisitions with point-level annota", + "bbox": [ + 496, + 854, + 890, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "9397", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/6d6d8acc1efd2dc7e11f5e07b74e71c89e2f321b3aa2d0f6f5e775e803adc302.jpg", + "table_caption": [ + "Table 1. SemanticKITTI splits, is defined as KITTI- $n^i$ , where $n$ is the number of novel classes and $i$ is the split index." + ], + "table_footnote": [], + "table_body": "
SplitNovel Classes
KITTI-50building, road, sidewalk, terrain, vegetation
KITTI-51car, fence, other-ground, parking, trunk
KITTI-52motorcycle, other-vehicle, pole, traffic-sign, truck
KITTI-43bicycle, bicyclist, motorcyclist, person
", + "bbox": [ + 80, + 121, + 467, + 204 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/c247e070242704965f06018b3b40fdd8cce5a9b9504ab6a6ef21f197d77de149.jpg", + "table_caption": [ + "Table 2. SemanticPOSS splits, defined as $\\mathrm{POSS} - n^i$ , where $n$ is the number of novel classes and $i$ is the split index." + ], + "table_footnote": [], + "table_body": "
SplitNovel Classes
POSS-40building, car, ground, plants
POSS-31bike, fence, person
POSS-32pole, traffic-sign, trunk
POSS-33cone-stone, rider, trashcan
", + "bbox": [ + 140, + 252, + 406, + 338 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tions of 19 semantic classes. Based on the official benchmark guidelines [4], we use sequence 08 for validation and the other sequences for training. SemanticPOSS [24] consists of 2,988 real-world point cloud acquisitions with point-level annotations of 13 semantic classes. Based on the official benchmark guidelines [24], we use sequence 03 for validation and the other sequences for training.", + "bbox": [ + 75, + 366, + 467, + 473 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Experimental protocol for 3D NCD. Similarly to what proposed by [40] in the 2D domain, we create different splits of each dataset to validate the NCD performance. We create four splits for SemanticKITTI and SemanticPOSS. We refer to these splits as SemanticKITTI- $n^i$ and SemanticPOSS- $n^i$ , where $i$ indexes the split. In each set, the novel classes and the base classes correspond to unlabelled and labelled points, respectively. Tabs. 1 & 2 detail the splits of our datasets. These splits are selected based on their class distribution in the dataset and on the semantic relationship between novel and base classes, e.g. in KITTI-43 the base class motorcycle can be helpful to discover the novel class motorcyclist. We report additional details about the selection process in the Supplementary Material.", + "bbox": [ + 75, + 474, + 467, + 685 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We quantify the performance by using the mean Intersection over Union (mIoU), which is defined as the average IoU across the considered classes [4]. We provide separate mIoU values for the base and novel classes. We also report the overall mIoU computed across all the classes in the dataset for completeness.", + "bbox": [ + 75, + 686, + 467, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. We implement our network based on a MinkowskiUNet-34C network [10]. Point-level features are extracted from the penultimate layer. The segmentation heads are implemented as linear layers, producing output logits for each point in the batched point clouds. We train our network for 10 epochs. We use the SGD optimizer, with momentum 0.9 and weight decay 0.0001. Our learning rate scheduler consists of linear warm-up and co", + "bbox": [ + 75, + 779, + 467, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "sine annealing, with $lr_{max} = 10^{-2}$ and $lr_{min} = 10^{-5}$ . We train with batch size equal to 4. We employ 5 segmentation heads, that are used in synergy with an equal number of over-clustering heads, with $o = 3$ . In $\\phi$ , we set $p = 0.5$ for SemanticKITTI- $n^i$ and $p = 0.3$ for SemanticPOSS- $n^i$ . We adapted the implementation of the Sinkhorn-Knopp algorithm [11] from the code provided by [5], with the introduction of the queue and an in-place normalisation steps. Similarly to [5], we set $n_{sk\\_iters} = 3$ , while we adopt a linear decay for $\\epsilon$ , with $\\epsilon_{start} = 0.3$ , $\\epsilon_{end} = 0.05$ .", + "bbox": [ + 496, + 90, + 890, + 243 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Quantitative analysis", + "text_level": 1, + "bbox": [ + 500, + 252, + 697, + 268 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Segmentation quality. Tabs. 3 & 4 report the quantitative results on SemanticPOSS and SemanticKITTI, respectively. We report the Full supervision setting as our upper bound.", + "bbox": [ + 496, + 276, + 890, + 321 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "On SemanticPOSS, we outperform EUMS† on three out of four splits with an improvement of 18.3 mIoU on POSS-40, 9.0 mIoU on POSS-31 and 0.6 on POSS-32. In these splits, NOPS shows a large improvement on all the novel classes, with the exception of the class fence in POSS-31 and traffic-sign in POSS-33. Differently, we deem the lower performance in POSS-33 is due to the difficulty and scarce presence of these novel classes. The advantage of EUMS† is the clustering on the whole dataset that enables a complete visibility of all the novel classes. On average, NOPS achieves 21.40 mIoU, improving over EUMS† of 6.5 mIoU.", + "bbox": [ + 496, + 321, + 890, + 487 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "On SemanticKITTI, we outperform EUMS† on all the four splits, improving by 12.6 mIoU on KITTI-5⁰, 1.2 mIoU on KITTI-5¹, 4.2mIoU on KITTI-5² and 5.3 mIoU on KITTI-4³. NOPS outperforms EUMS† by a large margin on the majority of the novel classes. Exceptions are the class sidewalk in KITTI-5⁰, car in KITTI-5¹, motorcycle in KITTI-5² and motorcyclist in KITTI-4³. On average, NOPS achieves 22.84 mIoU, improving over EUMS† of 5.8 mIoU. Interestingly, NOPS outperforms also the supervised upper bound on the class trunk in KITTI-5¹.", + "bbox": [ + 496, + 488, + 890, + 638 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Computational time. NOPS outperforms EUMS† in terms of computational time. Firstly, EUMS† requires a pretraining step and a fine-tuning step, i.e. 30 training epochs in total. Then, EUMS† requires a large amount of memory (up to 200 GB memory for KITTI-5⁰) to store the data required for clustering, taking several hours (50 hrs) to complete the training procedure. Differently, NOPS achieves superior performance with 10 training epochs, by using less memory (10 GB max) and a lower computational time (up to 25 hrs for KITTI-5⁰). We run these tests using one GPU Tesla A40-48GB.", + "bbox": [ + 496, + 638, + 890, + 804 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Qualitative analysis", + "text_level": 1, + "bbox": [ + 500, + 816, + 687, + 832 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fig. 3 shows some segmentation results of NOPS and EUMS† on SemanticPOSS and SemaniticKITTI. We can observe that the predictions of the base classes in the two datasets are correct for both the models, with just minor er", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "9398", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2fa24ac208d596a03a813569770bf714ef68bed88adece4d221ecc0c9220606a.jpg", + "table_caption": [ + "Table 3. Novel class discovery results on SemanticPOSS. NOPs outperforms EUMS† on three out of four splits. Full supervision: model trained with labels for base and novel classes. EUMS†: baseline described in Sec. 4. Highlighted values are the novel classes in each split." + ], + "table_footnote": [], + "table_body": "
SplitModelbluebub-carcome.fencegrou.pers.plantspoleridertraf.trafe.trunkNovelmIoU BaseAll
Full supervision43.2071.3033.0032.5044.6078.5061.8073.9030.9054.7026.7011.0019.30--44.72
POSS-40EUMS† [40]25.673.980.5616.4429.4036.7643.8428.4613.1326.7518.183.3416.9117.4421.5220.26
NOPS (Ours)35.4730.351.2413.5224.1369.1444.7042.0719.1947.6524.448.1721.8235.7026.5729.38
POSS-31EUMS† [40]15.1767.9828.0223.9811.8875.0735.9874.4626.9148.5626.005.6023.0521.0139.9635.59
NOPS (Ours)29.3571.3528.7012.213.9478.2456.7874.2118.2938.8823.3113.7423.5130.0238.2436.35
POSS-32EUMS† [40]40.1469.4527.6713.5034.8676.0354.6675.595.2739.227.798.5211.858.3143.9635.74
NOPS (Ours)37.1671.8129.7414.6428.3877.5352.0973.0011.5147.110.5410.2014.798.9544.1736.04
POSS-33EUMS† [40]41.1770.6828.084.3438.2776.6638.2975.3525.7634.3428.310.3624.4013.0144.7037.38
NOPS (Ours)38.5570.3630.910.0029.3876.5055.9871.8417.0331.8726.150.9522.5710.9443.9336.32
AvgEUMS† [40]14.9437.5432.24
NOPS (Ours)21.4038.2334.52
", + "bbox": [ + 76, + 122, + 890, + 301 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/95a843c5a16f677dcdd021709e1ac708350872cc57031ad3edf1ec55d85078ba.jpg", + "table_caption": [ + "Table 4. Novel class discovery results on SemanticKITTI. NOPS outperforms EUMS† on all four splits. Full supervision: model trained with annotations for base and novel classes. EUMS†: baseline described in Sec. 4. Highlighted values are the novel classes in each split." + ], + "table_footnote": [], + "table_body": "
SplitModelbicyclebicyclebicyclecarfencepolepoleothersothersparkpers.poleroadsidewterracetraff.trucktrunkvessels.NovelmIoU BaseAll
Full supervision6.3039.5085.4090.0023.2020.305.703.9018.0028.9031.0040.6090.9074.6062.1020.5062.9046.2083.90--43.89
KITTI-50EUMS† [40]5.2839.9615.7779.209.0316.892.520.0711.3914.4012.6729.1742.5826.100.0510.3047.3737.9238.3524.5721.0823.11
NOPS (Ours)5.5947.7652.6882.6013.7625.551.361.6614.5219.8025.8632.1256.748.0823.8414.2849.4136.1844.1737.1024.7029.62
KITTI-51EUMS† [40]7.5342.4179.9776.778.6219.581.390.5712.0314.1413.9540.7486.3266.4556.2911.9744.7920.9472.4024.2137.0635.62
NOPS (Ours)7.3651.2384.5350.877.2728.931.760.0022.2019.3930.4237.6190.0772.1860.7516.7857.3449.2585.1225.3643.0940.69
KITTI-52EUMS† [40]8.2650.7882.9888.0517.882.752.320.173.1625.4024.9820.2088.3071.0457.858.6327.1638.3676.9512.3842.2236.59
NOPS (Ours)6.7249.2486.3690.7923.682.690.581.8715.4629.4827.9236.3990.2673.3961.2117.8310.3246.1684.2916.5444.8039.72
KITTI-43EUMS† [40]3.952.4780.1087.2116.8114.0214.980.3114.1320.776.8037.5986.7966.5055.2616.2040.6238.3776.157.0543.3935.74
NOPS (Ours)2.3227.8386.0489.8923.0624.472.923.0618.1930.0916.3239.9090.6573.5161.0417.4049.7644.0183.1812.3548.9541.24
AvgEUMS† [40]17.0535.9432.76
NOPS (Ours)22.8440.3837.73
", + "bbox": [ + 76, + 347, + 890, + 512 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "rors at the edges of the objects. NOPS shows better segmentation capabilities also when dealing with the novel classes, being able to properly segment the correct class for the unknown objects. On the larger objects, such as buildings in the scenes, some mixing of labels can be observed. Differently, EUMS nearly fails in correctly recognising the novel objects, resulting in the inclusion of different classes (either novel or base) into the same object, e.g. the facade of the building in POSS-3 $^1$ .", + "bbox": [ + 75, + 539, + 468, + 675 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6. Ablation studies", + "text_level": 1, + "bbox": [ + 76, + 691, + 236, + 707 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We study the main components of NOPS and its behaviour when varying its parameters on SemanticPOSS.", + "bbox": [ + 75, + 717, + 468, + 747 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Method components. Fig. 4 shows the performance on novel and base classes of seven versions of NOPS. The first three versions use the pre-trained model on the base classes, while the last four versions use the model trained from scratch. Each version is defined as follows:", + "bbox": [ + 75, + 750, + 468, + 823 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- P: we use a pre-trained model, and we remove $Z_{q}$ , $\\tau_{c}$ and the over-clustering heads.", + "- OC: P + over-clustering heads.", + "- Q: $\\mathrm{OC} + Z_{q}$ , i.e. our queue without class-balancing.", + "- NP: Q without pre-training." + ], + "bbox": [ + 94, + 825, + 467, + 900 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- NP+: NP + our selection function $\\phi$ on the queue.", + "- NP++: NP + $\\tau_{c}$ on the features used to derive the pseudo-labels.", + "Full: NOPS with all the components activated." + ], + "bbox": [ + 517, + 539, + 888, + 598 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Pre-trained approaches generally underperform their trained-from-scratch counterparts on the novel classes. This is visible in the low performance of P, OC and Q. We have a significant improvement when pre-training is not used (NP), i.e. we achieve $20.26\\mathrm{mIoU}$ . We can see that the queue both with and without pre-training is helpful. When we add the feature selection for the queue and for the training, i.e. $\\mathrm{NP+}$ and $\\mathrm{NP++}$ , we have improvements, i.e. 20.63 mIoU and 20.90 mIoU, respectively. The best performance is achieved with Full. Although we can observe variations on the performance of the base classes, their information is retained by the network when we discover the novel ones.", + "bbox": [ + 496, + 606, + 890, + 787 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Parameter analysis. We study the behaviour of the percentile $p$ in our selection function $\\phi$ , when we apply it to the features both for pseudo-labelling and for the class-balanced queue $Z_{q}$ . Tab. 5 reports the results on each split of SemanticPOSS. For each split, we can observe that the performance depends on the number of points and difficulty of the novel classes. In $\\mathrm{POSS - 4^0}$ and $\\mathrm{POSS - 3^1}$ , lower values", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "9399", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b48fc6ebf247d7fbeeae594b6b55c2030f35ce13f6313749a2d85341d1a3955b.jpg", + "image_caption": [ + "Figure 3. Qualitative comparisons on (top) SemanticPOSS and (bottom) SemanticKITTI. We report results on both base and novel classes. On novel classes, $\\mathrm{EUMS}^{\\dagger}$ fails in recognising novel objects with mixed and cluttered predictions, e.g. the facade of building in $\\mathrm{POSS - 3^{1}}$ . NOPS shows superior segmentation performance on all the novel classes." + ], + "image_footnote": [], + "bbox": [ + 66, + 74, + 890, + 334 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2050eaad008e6e9120781e8a12d54f5512ec7210806b4c4bf3e31183ecff57f8.jpg", + "image_caption": [ + "Figure 4. Ablation study with different components and initialisation strategies on SemanticPOSS. In P, OC and Q, we initialise the model after base pre-training, and use different configurations of the over-clustering heads and of our queue balancing. In NP NP+, NP++ and Full, we begin with Q, we avoid pre-training, and we use $\\phi$ and $\\tau_{c}$ incrementally. See Sec. 6 for definition of methods." + ], + "image_footnote": [], + "bbox": [ + 76, + 404, + 470, + 513 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "of $p$ result in less severe selection. We believe this is related to the class distribution within these splits. This is in line with what observed in Tab. 3. In $\\mathrm{POSS}-3^{2}$ and $\\mathrm{POSS}-3^{3}$ , we notice a different behaviour, a higher value of $p$ provides better results. We relate this to the difficulty of the novel classes in these splits whose noisy pseudo-labels can benefit from a more rigorous selection of the features.", + "bbox": [ + 75, + 630, + 468, + 737 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusions", + "text_level": 1, + "bbox": [ + 76, + 753, + 204, + 768 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We explored the new problem of novel class discovery for 3D point cloud segmentation. Firstly, we transposed the only NCD method for 2D semantic segmentation to 3D point cloud data, and experimentally found that it has several limitations. We discussed that extending 2D NCD approaches to 3D data (point clouds) is not trivial because the assumptions made for 2D data are not easily transferable to 3D. Secondly, we presented NOPS, to tackle NCD for point", + "bbox": [ + 75, + 779, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/949590fa56add36bc0ddbd876574e1b0fa94613365a273a90078be593835ae8f.jpg", + "table_caption": [ + "Table 5. Ablation study showing how different values of $p$ affect the performance on SemanticPOSS. The lower $p$ is, the less severe the selection of the features, resulting in better performances for POSS-4 $^0$ . Differently, POSS-3 $^3$ benefits from an higher value of $p$ , which leads to a more vigorous filtering of the features. POSS-3 $^1$ and POSS-3 $^2$ show the best performances with $p = 0.5$" + ], + "table_footnote": [], + "table_body": "
SplitPercentile p
0.10.30.50.70.9
POSS-4030.8135.7028.7730.9326.69
POSS-3128.3330.0230.4323.3218.91
POSS-328.078.9510.3210.257.76
POSS-3310.5510.9411.6914.3813.42
Avg.19.4421.4020.3019.7216.70
", + "bbox": [ + 532, + 493, + 859, + 609 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "cloud segmentation by using online clustering and exploiting uncertainty quantification to produce pseudo-labels for the novel points. Lastly, we introduced a new evaluation protocol to assess the performance of NCD for point cloud segmentation. Experiments on two different segmentation dataset showed that NOPS outperforms the compared baselines by a large margin. Future research directions could investigate the extension of our method when base annotations are fewer and/or weakly labelled.", + "bbox": [ + 496, + 640, + 890, + 776 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations NOPS limitations include the prior knowledge on the number of novel classes $C_n$ to discover. This could be a limitation when $C_n$ is not known a-priori and novel classes may appear in an incremental manner. We believe that a solution may be to learn novel classes incrementally. Another limitation is the loss we use to handle class unbalancing. More recent techniques to handle this drawback could be further explored.", + "bbox": [ + 496, + 779, + 892, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "9400", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] I. Alonso, L. Riazuelo, L. Montesano, and A.C. Murillo. 3D MiniNet: Learning a 2D representation from point clouds for fast and efficient 3D LiDAR semantic segmentation. In IROS, 2020. 2", + "[2] Y.M. Asano, C. Rupprecht, and A. Vedaldi. Self-labelling via simultaneous clustering and representation learning. In ICLR, 2020. 3, 4", + "[3] J. Behley, M. Garbade, A. Milioto, J. Quenzel, S. Behnke, J. Gall, and C. Stachniss. Towards 3D LiDAR-based semantic scene understanding of 3D point cloud sequences: The SemanticKITTI Dataset. In IJRR, 2021. 2, 5", + "[4] J. Behley, M. Garbade, A. Milioto, J. Quenzel, S. Behnke, C. Stachniss, and J. Gall. SemanticKITTI: A dataset for semantic scene understanding of LiDAR sequences. In CVPR, 2019. 2, 5, 6", + "[5] M. Caron, I. Misra, J. Mairal, P. Goyal, P. Bojanowski, and A. Joulin. Unsupervised learning of visual features by contrasting cluster assignments. In NeurIPS, 2020. 2, 4, 5, 6", + "[6] J. Cen, P. Yun, S. Zhang, J. Cai, D. Luan, M. Tang, M. Liu, and M.Y. Wang. Open-world semantic segmentation for lidar point clouds. In ECCV, 2022. 3", + "[7] L.C. Chen, G. Papandreou, I. Kokkinos, K. Murphy, and A.L. Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. In T-PAMI, 2017. 3", + "[8] L.C. Chen, Y. Zhu, G. Papandreou, F. Schroff, and A. Hartwig. Encoder-decoder with atrous separable convolution for semantic image segmentation. In ECCV, 2018. 3", + "[9] R. Cheng, R. Razani, E. Taghavi, E. Li, and B. Liu. (AF)2-S3Net: Attentive Feature Fusion with Adaptive Feature Selection for Sparse Semantic Segmentation Network. In CVPR, 2021. 2", + "[10] C. Choy, JY. Gwak, and S. Savarese. 4D spatio-temporal convnets: Minkowski convolutional neural networks. In CVPR, 2019. 2, 3, 6", + "[11] M. Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. In NeurIPS, 2013. 3, 4, 6", + "[12] E. Fini, E. Sangineto, S. Lathuiliere, Z. Zhong, M. Nabi, and E. Ricci. A unified objective for novel class discovery. In CVPR, 2021. 1, 2, 4", + "[13] A. Geiger, P. Lenz, and R. Urtasun. Are we ready for Autonomous Driving? The KITTI Vision Benchmark Suite. In CVPR, 2012. 2, 5", + "[14] B. Graham, M. Engelcke, and L. van der Maaten. 3D semantic segmentation with submanifold sparse convolutional networks. In CVPR, 2018. 2", + "[15] B. Graham and L. van der Maaten. Submanifold sparse convolutional networks. arXiv, 2017. 2", + "[16] K. Han, A. Vedaldi, and A. Zisserman. Learning to discover novel visual categories via deep transfer clustering. In CVPR, 2019. 1, 2, 4", + "[17] Q. Hu, B. Yang, L. Xie, S. Rosa, Y. Guo, Z. Wang, N. Trigoni, and A. Markham. Randla-net: Efficient semantic segmentation of large-scale point clouds. In CVPR, 2020. 2" + ], + "bbox": [ + 78, + 116, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[18] X. Ji, J.F. Henriques, and A. Vedaldi. Invariant information clustering for unsupervised image classification and segmentation. In ICCV, 2019. 4", + "[19] X. Jia, K. Han, Y. Zhu, and B. Green. Joint representation learning and novel category discovery on single-and multimodal data. In ICCV, 2021. 2", + "[20] K. Joseph, S. Paul, G. Aggarwal, S. Biswas, P. Rai, K. Han, and V.N. Balasubramanian. Novel class discovery without forgetting. ECCV, 2022. 2", + "[21] G. Mei, F. Poiesi, C. Saltori, J. Zhang, E. Ricci, and N. Sebe. Overlap-guided gaussian mixture models for point cloud registration. In WACV, 2023. 4", + "[22] G. Mei, C. Saltori, F. Poiesi, J. Zhang, E. Ricci, N. Sebe, and Q. Wu. Data augmentation-free unsupervised learning for 3d point cloud understanding. In BMVC, 2022. 3, 4", + "[23] A. Milioto, I. Vizzo, J. Behley, and C. Stachniss. Rangenet++: Fast and accurate lidar semantic segmentation. In IROS, 2019. 2, 3", + "[24] Y. Pan, B. Gao, J. Mei, S. Geng, C. Li, and H. Zhao. SemanticPOSS: A point cloud dataset with large quantity of dynamic instances. In IV, 2020. 2, 5, 6", + "[25] C.R. Qi, H. Su, K. Mo, and L.J. Guibas. Pointnet: Deep learning on point sets for 3D classification and segmentation. In CVPR, 2017. 2", + "[26] C.R. Qi, L. Yi, H. Su, and L.J. Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In NeurIPS, 2017. 2", + "[27] O. Ronneberger, P. Fischer, and T. Brox. U-net: Convolutional networks for biomedical image segmentation. In MIC-CAI, 2015. 2", + "[28] S. Roy, M. Liu, Z. Zhong, N. Sebe, and E. Ricci. Class incremental novel class discovery. In ECCV, 2022. 2", + "[29] C. Saltori, F. Galasso, G. Fiameni, N. Sebe, E. Ricci, and F. Poiesi. Cosmix: Compositional semantic mix for domain adaptation in 3d lidar segmentation. In ECCV, 2022. 3", + "[30] R. Song, W. Zhang, Y. Zhao, Y. Liu, and P.L. Rosin. Mesh saliency: An independent perceptual measure or a derivative of image saliency? In CVPR, 2021. 2", + "[31] N. Souly, C. Spampinato, and M. Shah. Semi supervised semantic segmentation using generative adversarial network. In ICCV, 2017. 2", + "[32] X. Tang, J. Wang, B. Gao, E. Dellandrea, R. Gaizauskas, and L. Chen. Large scale semi-supervised object detection using visual and semantic knowledge transfer. In CVPR, 2016. 2", + "[33] H. Thomas, C.R. Qi, J.-E. Deschaud, B. Marcotegui, F. Goulette, and J. L. Guibas. Kpconv: Flexible and deformable convolution for point clouds. In ICCV, 2019. 2", + "[34] S. Vaze, K. Han, A. Vedaldi, and A. Zisserman. Generalized category discovery. In CVPR, 2022. 2", + "[35] B. Wu, A. Wan, X. Yue, and K. Keutzer. Squeezeseg: Convolutional neural nets with recurrent CRF for real-time road-object segmentation from 3D LiDAR point cloud. In ICRA, 2018. 2", + "[36] B. Wu, X. Zhou, S. Zhao, X. Yue, and K. Keutzer. Squeeze-segv2: Improved model structure and unsupervised domain adaptation for road-object segmentation from a LiDAR point cloud. In ICRA, 2019. 2" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9401", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] M. Yang, Y. Zhu, J. Yu, A. Wu, and C. Deng. Divide and conquer: Compositional experts for generalized novel class discovery. In CVPR, 2022. 2", + "[38] L. Zhang and G.J. Qi. Wcp: Worst-case perturbations for semi-supervised deep learning. In CVPR, 2020. 2", + "[39] Y. Zhang, Z. Zhou, P. David, X. Yue, Z. Xi, B. Gong, and H. Foroosh. Polarnet: An improved grid representation for online LiDAR point clouds semantic segmentation. In CVPR, 2020. 2", + "[40] Y. Zhao, Z. Zhong, N. Sebe, and G.H. Lee. Novel class discovery in semantic segmentation. In CVPR, 2022. 1, 2, 3, 4, 5, 6, 7, 8", + "[41] Z. Zhong, E. Fini, S. Roy, Z. Luo, E. Ricci, and N. Sebe. Neighborhood contrastive learning for novel class discovery. In CVPR, 2021. 1, 2, 4", + "[42] Z. Zhong, L. Zhu, Z. Luo, S. Li, Y. Yang, and N. Sebe. Openmix: Reviving known knowledge for discovering novel visual categories in an open world. In CVPR, 2021. 2", + "[43] Y. Zhou and O. Tuzel. Voxelnet: End-to-end learning for point cloud based 3D object detection. In CVPR, 2018. 2", + "[44] X. Zhu, H. Zhou, T. Wang, F. Hong, Y. Ma, W. Li, H. Li, and D. Lin. Cylindrical and asymmetrical 3D convolution networks for lidar segmentation. In CVPR, 2021. 2" + ], + "bbox": [ + 78, + 90, + 467, + 417 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "9402", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Novel Class Discovery for 3D Point Cloud Semantic Segmentation/b9953b69-6d05-422e-9f85-28806509ef58_model.json b/2023/Novel Class Discovery for 3D Point Cloud Semantic Segmentation/b9953b69-6d05-422e-9f85-28806509ef58_model.json new file mode 100644 index 0000000000000000000000000000000000000000..958bfee0884f3acf44cec7da06b166ef3798f160 --- /dev/null +++ b/2023/Novel Class Discovery for 3D Point Cloud Semantic Segmentation/b9953b69-6d05-422e-9f85-28806509ef58_model.json @@ -0,0 +1,1859 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.154, + 0.095, + 0.816, + 0.117 + ], + "angle": 0, + "content": "Novel Class Discovery for 3D Point Cloud Semantic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.144, + 0.735, + 0.163 + ], + "angle": 0, + "content": "Luigi Riz\\(^{1}\\) Cristiano Saltori\\(^{2}\\) Elisa Ricci\\(^{1,2}\\) Fabio Poiesi\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.314, + 0.171, + 0.658, + 0.187 + ], + "angle": 0, + "content": "\\(^{1}\\)Fondazione Bruno Kessler \\(^{2}\\)University of Trento" + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.221, + 0.891, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.341, + 0.893, + 0.37 + ], + "angle": 0, + "content": "Figure 1. Novel Class Discovery for 3D point cloud semantic segmentation seeks to recognise novel classes by clustering unlabelled novel points with similar semantic features and by exploiting only the knowledge of a set of labelled samples corresponding to the base classes." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.397, + 0.313, + 0.413 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.44, + 0.474, + 0.759 + ], + "angle": 0, + "content": "Novel class discovery (NCD) for semantic segmentation is the task of learning a model that can segment unlabelled (novel) classes using only the supervision from labelled (base) classes. This problem has recently been pioneered for 2D image data, but no work exists for 3D point cloud data. In fact, the assumptions made for 2D are loosely applicable to 3D in this case. This paper is presented to advance the state of the art on point cloud data analysis in four directions. Firstly, we address the new problem of NCD for point cloud semantic segmentation. Secondly, we show that the transposition of the only existing NCD method for 2D semantic segmentation to 3D data is suboptimal. Thirdly, we present a new method for NCD based on online clustering that exploits uncertainty quantification to produce prototypes for pseudo-labelling the points of the novel classes. Lastly, we introduce a new evaluation protocol to assess the performance of NCD for point cloud semantic segmentation. We thoroughly evaluate our method on SemanticKITTI and SemanticPOSS datasets, showing that it can significantly outperform the baseline. Project page: https://github.com/LuigiRiz/NOPS." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.828, + 0.47, + 0.903 + ], + "angle": 0, + "content": "This project has received funding from the European Union's Horizon Europe research and innovation programme under grant agreement No 101058589. This work was also partially supported by the PRIN project LEGO-AI (Prot. 2020TA3K9N), the EU ISFP PROTECTOR (101034216) project and the EU H2020 MARVEL (957337) project and, it was carried out in the Vision and Learning joint laboratory of FBK and UNITN." + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.382, + 0.631, + 0.398 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.413, + 0.893, + 0.625 + ], + "angle": 0, + "content": "As humans, we are fairly skilled in organising new visual knowledge (novelty) into homogeneous groups, even when we do not know what we are observing. However, machines cannot perform this task satisfactorily without our supervision. The challenge here is mainly in the formulation of discriminative latent representations of the real world and in the quantification of the uncertainty when the novelty is observed [16,40,41]. The work of Han et al. [16] pioneered the formulation of the Novel Class Discovery (NCD) problem by defining it as the task that aims to classify the samples of an unlabelled dataset into different classes, i.e. the novel samples, by exploiting the knowledge of a set of labelled samples, i.e. the base samples. Note that the classes in the labelled and unlabelled datasets are disjoint." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.895, + 0.902 + ], + "angle": 0, + "content": "NCD has been explored in the 2D image domain for classification [12, 16, 41] and later for semantic segmentation [40]. In particular, Zhao et al. [40] presented the first approach for NCD for 2D semantic segmentation. Two key assumptions were made by the authors. Firstly, at most one novel class is allowed in each image. Secondly, the new class belongs to a foreground object that can be found via saliency detection (e.g. a man on a bicycle, where the bicycle is the novel class). Thanks to these assumptions, the authors could pool the features of each image into a single latent representation and cluster the representations of the whole dataset to discover clusters of novel classes. We argue that these two are important constraints that are not applicable to generic 3D data, in particular to point clouds captured with LiDAR sensors in real-world city-scale scenarios. One point cloud can contain more than one novel class, and the saliency for 3D data cannot be leveraged in the same way as that for 2D data. Although they are both" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9393" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.168 + ], + "angle": 0, + "content": "related to the attraction of human fixations, 3D saliency is more related to the regional importance of 3D surfaces rather than the foreground/background distinction [30]. Our motivation in exploring NCD for the 3D setting is mainly driven by addressing these shortcomings." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.169, + 0.473, + 0.637 + ], + "angle": 0, + "content": "In this paper, we explore the new problem of NCD for 3D point cloud semantic segmentation (see Fig. 1). Given a partially human-annotated dataset, we jointly learn base and novel semantic classes by clustering unlabelled points with similar semantic features. We adapt the method of Zhao et al. [40] (Entropy-based Uncertainty Modeling and Self-training - EUMS) for point cloud data and use it as our baseline. We go beyond their formulation and, inspired by [5], we integrate batch-level (online) clustering in our method and update prototypes during training in order to make clustering computationally tractable. Cluster assignments are then used as training pseudo-labels. We also exploit over-clustering to achieve a higher clustering accuracy as in EUMS. Because point clouds contain multiple semantic classes, we cannot guarantee that all the classes appear in the point clouds within each batch, some will be missing. Therefore, we design a queuing strategy to store important features over training time, which will be used for pseudolabelling in the case of missing classes. We further introduce a strategy for exploiting the pseudo-label uncertainty to promote the creation of reliable prototypes that we then exploit to produce higher-quality pseudo-labels. Lastly, we produce two augmented views of the same point cloud and impose pseudo-label consistency amongst them. We evaluate our approach on SemanticKITTI [3, 4, 13] and SemanticPOSS [24], introducing an evaluation protocol for NCD and point cloud segmentation that can be adopted in future works. We empirically show that our approach largely outperforms our baseline in both datasets. We also perform an extensive ablation study to demonstrate the importance of the different components of our method." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.642, + 0.324, + 0.656 + ], + "angle": 0, + "content": "To summarise, our contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.659, + 0.468, + 0.688 + ], + "angle": 0, + "content": "- We address the new problem of NCD for 3D semantic segmentation;" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.689, + 0.468, + 0.734 + ], + "angle": 0, + "content": "- We show that the transposition of the only existing NCD method for 2D semantic segmentation [40] to 3D data is suboptimal;" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.735, + 0.468, + 0.779 + ], + "angle": 0, + "content": "- We present a new method for NCD based on online clustering and uncertainty estimation, which we name it NOPS (NOvel Point Segmentation);" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.78, + 0.468, + 0.81 + ], + "angle": 0, + "content": "- We introduce a new evaluation protocol to assess the performance of NCD for 3D semantic segmentation." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.659, + 0.468, + 0.81 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.829, + 0.216, + 0.844 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Point cloud semantic segmentation can be performed at the point level [26], on range view maps [27], and by voxelising the input points [43]. Point-level networks pro" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.53 + ], + "angle": 0, + "content": "cess the input without intermediate representations. Examples of these include PointNet [25], PointNet++ [26], RandLA-Net [17], and KPConv [33]. PointNet [25] and PointNet++ [26] are based on a series of multi-layer perceptron where PointNet++ introduces global and local feature aggregation at multiple scales. RandLA-Net [17] uses random sampling, attentive pooling, and local spatial encoding. KPConv [33] employs flexible and deformable convolutions in a continuous input space. Point-level networks are computationally inefficient when large-scale point clouds are processed. Range view architectures [23] and voxel-based approaches [10] are more computationally efficient than their point-level counterpart. The former requires projecting the input points on a 2D dense map, processing input maps with 2D convolutional filters [27], and reprojecting predictions to the initial 3D space. SqueezeSeg networks [35, 36], 3D-MiniNet [1], RangeNet++ [23], and PolarNet [39] are examples of this category. Although they are more efficient, these approaches tend to lose information during the projection and re-projection phase. The latter includes 3D quantisation-based approaches that discretise the input points into a 3D voxel grid and employ 3D convolutions [43] or 3D sparse convolutions [10, 15] to predict per-voxel classes. VoxelNet [43], SparseConv [14, 15], MinkowskiNet [10], Cylinder3D [44], and \\((\\mathrm{AF})^2\\)-S3Net [9] are architectures belonging to this category. These approaches tackle point cloud segmentation in the supervised settings, whereas we tackle novel class discovery with labelled base classes and unlabelled novel classes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.538, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Novel class discovery (NCD) is explored for 2D classification [12, 16, 19, 20, 28, 34, 37, 41, 42] and 2D segmentation [40]. NCD is more complex than standard semi-supervised learning [31, 32, 38]. In semi-supervised learning, labelled and unlabelled samples belong to the same classes, while in NCD, novel and base samples belong to disjoint classes. Han et al. [16] pioneered the NCD problem for 2D image classification. A classification model is pretrained on a set of base classes and used as feature extractor for the novel classes. They then train a classifier for the novel classes using the pseudo-labels produced by the pretrained model. Zhong et al. [41] introduced neighbourhood contrastive learning to generate discriminative representations for clustering. They retrieve and aggregate pseudopositive pairs with contrastive learning, encouraging the model to learn more discriminative representations. Hard negatives are obtained by mixing labelled and unlabelled samples in the feature space. UNO [12] unifies the two previous works by using a unique classification loss function for both base and novel classes, where pseudo-labels are processed together with ground-truth labels. NCD without Forgetting [20] and FRoST [28] further extend NCD to the incremental learning setting. EUMS [40] is the only approach analysing the NCD problem for semantic segmen" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9394" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.089, + 0.895, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.389, + 0.893, + 0.435 + ], + "angle": 0, + "content": "Figure 2. Overview of NOPS. We random augment the input point cloud twice and extract point-level features \\(\\mathcal{F}\\) with the shared model \\(f_{\\xi}\\). \\(\\mathcal{F}\\) are used to obtain pseudo-labels in the online pseudo-labelling. We forward \\(\\mathcal{F}\\) to a novel \\(f_{n}\\) and a base \\(f_{b}\\) segmentation layer to output the novel and base predictions, respectively. We optimise our network by minimising a global objective function based on cross entropy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.458, + 0.473, + 0.884 + ], + "angle": 0, + "content": "tation. Unlike image classification, the model has to classify each pixel and handle multiple classes in each image. EUMS consists of a multi-stage pipeline using a saliency model to cluster the latent representations of novel classes to produce pseudo-labels. Moreover, entropy-based uncertainty and self-training are used to overcome noisy pseudolabels while improving the model performance on the novel classes. In this work, we tackle the problem of NCD in 3D point cloud semantic segmentation. Unlike previous works, our problem inherits the challenges from the fields of 2D semantic segmentation [7, 8] and 3D point cloud segmentation [10, 23, 29]. From 2D semantic segmentation, it inherits the additional challenges of multiple novel classes in the same image and the strong class unbalance. From 3D point cloud segmentation, we inherit the sparsity of input data, the different density of point cloud regions and the inability to identify foreground and background. The latter are not present in 2D segmentation [40]. Unlike [40] that use K-Means, we formulate clustering as an optimal transport problem to avoid degenerate solutions (i.e. all data points may be assigned to the same label and learn a constant representation) [2, 22]. Lastly, related to EUMS, REAL is proposed for open-world 3D semantic segmentation [6], where both known and unknown points have to be segmented. Unlike NOPS, all the unknown points belong to a single class and it is the task of a human annotator to separately label the novel classes. Then, these labels are used to update the base model by incrementally learning the novel classes." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.457, + 0.687, + 0.474 + ], + "angle": 0, + "content": "3. Proposed approach" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.481, + 0.611, + 0.497 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.506, + 0.892, + 0.734 + ], + "angle": 0, + "content": "Given an input point cloud, we produce two augmented views that are processed with the same deep neural network to extract point-level features. These features are used to obtain pseudo-labels in the online pseudo-labelling step through the Sinkhorn-Knopp algorithm [11] (Sec. 3.3). Concurrently, we process the same features with the last network layers to segment novel and base classes. These features are stored in the class-balanced queue to mitigate the problem of batches with missing classes (Sec. 3.4). We exploit pseudo-label values (class probabilities) to filter out uncertain points, thus adding to the queue only high-quality points (Sec. 3.5). Lastly, we train our network by minimising the optimisation objective function through a swapped prediction task based on the computed pseudolabels (Sec. 3.6). Fig. 2 shows the block diagram of NOPS." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.741, + 0.698, + 0.756 + ], + "angle": 0, + "content": "3.2. Problem formulation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Let \\( \\mathrm{X} = \\{\\mathcal{X}\\} \\) be a dataset of 3D point clouds captured in different scenes. \\( \\mathcal{X} \\) is a set composed of a base set \\( \\mathcal{X}_b \\) and a novel set \\( \\mathcal{X}_n \\), s.t. \\( \\mathcal{X} = \\mathcal{X}_b \\cup \\mathcal{X}_n \\). The semantic categories that can be present in our point clouds are \\( \\mathcal{C} = \\mathcal{C}_b \\cup \\mathcal{C}_n \\), where \\( \\mathcal{C}_b \\) is the set of base classes and \\( \\mathcal{C}_n \\) is the set of novel classes, s.t. \\( \\mathcal{C}_b \\cap \\mathcal{C}_n = \\emptyset \\). Each \\( \\mathcal{X} \\in \\mathrm{X} \\) is composed of a finite but unknown number of 3D points \\( \\mathcal{X} = \\{(\\mathbf{x}, c)\\} \\), where \\( \\mathbf{x} \\in \\mathbb{R}^3 \\) is the coordinate of the a point and \\( c \\) is its semantic class. We know the class of the point \\( (\\mathbf{x}, c) \\), s.t. \\( \\mathbf{x} \\in \\mathcal{X}_b \\) and" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.958 + ], + "angle": 0, + "content": "9395" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.274 + ], + "angle": 0, + "content": "\\(c \\in \\mathcal{C}_b\\), but we do not know the class of the point \\((\\mathbf{x}, c)\\), s.t. \\(x \\in \\mathcal{X}_n\\) and \\(c \\in \\mathcal{C}_n\\). No points in \\(\\mathcal{X}_n\\) belong to one of the base classes \\(\\mathcal{C}_b\\). As in [16, 40, 41], we assume that the number of classes to discover is known, i.e. \\(|\\mathcal{C}_n| = C_n\\). We aim to design a computational approach that trains a deep neural network \\(f_\\Theta\\) that can segment all the points of a given point cloud, thus learning to jointly segment base classes \\(\\mathcal{C}_b\\) and novel classes \\(\\mathcal{C}_n\\). \\(\\Theta\\) are the weights of our deep neural network. \\(f_\\Theta\\) is composed of two heads, \\(f_\\Theta = f_\\xi \\circ \\{f_b, f_n\\}\\), where \\(f_b\\) is the segmentation head for the base classes, \\(f_n\\) is the segmentation head for the novel classes, \\(f_\\xi\\) is the feature extractor network and \\(\\circ\\) is the composition operator (Fig. 2)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.285, + 0.296, + 0.302 + ], + "angle": 0, + "content": "3.3. Online pseudo-labelling" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.309, + 0.473, + 0.643 + ], + "angle": 0, + "content": "We formulate pseudo-labelling as the assignment of novel points to the class-prototypes learnt during training [5]. Let \\(\\mathsf{P} \\in \\mathbb{R}^{D \\times \\rho}\\) be the class prototypes, where \\(D\\) is the size of the output features from \\(f_{\\xi}\\) and \\(\\rho\\) is the number of prototypes. Let \\(\\mathsf{Z} \\in \\mathbb{R}^{D \\times m}\\) be the normalised output features extracted from \\(f_{\\xi}\\), where \\(m\\) is the number of points of the point cloud. \\(m\\) is not known a priori and it can differ across point clouds. We aim to find the assignment \\(\\mathsf{Q} \\in \\mathbb{R}^{\\rho \\times m}\\) s.t. all the points in the batch are equally partitioned across the \\(\\rho\\) prototypes. This equipartition ensures that the feature representations of the points belonging to different novel classes are well-separated, thus preventing the case in which the novel class feature representations collapse into a unique solution. Caron et al. [5] employs an arbitrary large number of prototypes \\(\\rho\\) to effectively organise the feature space produced by \\(f_{\\xi}\\). They discard \\(\\mathsf{P}\\) after training. In contrast, we learn exactly \\(\\rho = C_n\\) class prototypes and propose to use \\(\\mathsf{P}\\) as the weights for our new class segmentation head \\(f_{n}\\), which outputs the \\(C_n\\) logits for the new classes. In order to optimise the assignment \\(\\mathsf{Q}\\), we maximise the similarity between the features of the new points and the learned prototypes as" + }, + { + "type": "equation", + "bbox": [ + 0.158, + 0.654, + 0.47, + 0.68 + ], + "angle": 0, + "content": "\\[\n\\max _ {\\mathbb {Q} \\in \\mathcal {Q}} \\operatorname {T r} \\left(\\mathbb {Q} ^ {\\top} \\mathbb {P} ^ {\\top} \\mathbb {Z}\\right) + \\epsilon H (\\mathbb {Q}) \\rightarrow \\mathbb {Q} ^ {*}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.691, + 0.47, + 0.844 + ], + "angle": 0, + "content": "where \\( H \\) is the entropy function, \\( \\epsilon \\) is the parameter that determines the smoothness of the assignment and \\( \\mathbb{Q}^* \\) is our sought solution. Asano et al. [2] enforce the equipartitioning constraint by requiring \\( \\mathbb{Q} \\) to belong to a transportation polytope and perform this optimisation on the whole dataset at once (offline). This operation with point cloud data is computationally impractical. Therefore, we formulate the transportation polytope such that the optimisation is performed online, which consists of processing only the points within the batch being processed" + }, + { + "type": "equation", + "bbox": [ + 0.086, + 0.855, + 0.47, + 0.901 + ], + "angle": 0, + "content": "\\[\n\\mathcal {Q} = \\left\\{\\mathbb {Q} \\in \\mathbb {R} _ {+} ^ {C _ {n} \\times m} | \\mathbb {Q} 1 _ {m} = \\frac {1}{C _ {n}} 1 _ {C _ {n}}, \\mathbb {Q} ^ {\\top} 1 _ {C _ {n}} = \\frac {1}{m} 1 _ {m} \\right\\}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "where \\(1_{\\star}\\) represents a vector of ones of dimension \\(\\star\\). These constraints ensure that each class prototype is selected on average at least \\(m / C_n\\) times in each batch. The solution \\(\\mathbb{Q}^*\\) can take the form of a normalised exponential matrix" + }, + { + "type": "equation", + "bbox": [ + 0.576, + 0.161, + 0.892, + 0.196 + ], + "angle": 0, + "content": "\\[\n\\mathbb {Q} ^ {*} = \\operatorname {d i a g} (\\alpha) \\exp \\left(\\frac {\\mathrm {P} ^ {\\top} \\mathrm {Z}}{\\epsilon}\\right) \\operatorname {d i a g} (\\beta), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.205, + 0.892, + 0.28 + ], + "angle": 0, + "content": "where \\(\\alpha\\) and \\(\\beta\\) are renormalization vectors that are computed iteratively with the Sinkhorn-Knopp algorithm [11, 21]. We then transpose the optimised soft assignment \\(\\mathbb{Q}^* \\in \\mathbb{R}_+^{C_n \\times m}\\) to obtain the soft pseudo-labels for each of the \\(m\\) novel points being processed within each batch." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.28, + 0.892, + 0.34 + ], + "angle": 0, + "content": "We empirically found that training can be more effective if pseudo-labels are smoother in the first training epochs and peaked in the last training epochs. Therefore, we introduce a linear decay of \\(\\epsilon\\) during training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.34, + 0.893, + 0.521 + ], + "angle": 0, + "content": "Multi-headed segmentation: A single segmentation head may converge to a suboptimal feature space, thus producing suboptimal prototype solutions. To further improve the segmentation quality, we use multiple novel class segmentation heads to optimise \\( f_{\\Theta} \\) based on different training solutions. Different solutions increase the likelihood of producing a diverse partitioning of the feature space as they regularise with each other (they share the same backbone) [18]. In practise, we concatenate the logits of the base class segmentation head with the outputs of each novel class segmentation head and we separately evaluate their loss for each novel class segmentation head at training time." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.522, + 0.892, + 0.642 + ], + "angle": 0, + "content": "We task our network to over-cluster novel points, using segmentation heads that output \\(o \\cdot C_n\\) logits, where \\(o\\) is the over-clustering factor. Previous studies empirically showed that this is beneficial to learn more informative features [5, 12, 18, 22]. We observed the same and concur that over-clustering can be useful for increasing expressivity of the feature representations. The over-clustering heads are then discarded at inference time." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.651, + 0.72, + 0.668 + ], + "angle": 0, + "content": "3.4. Class-balanced queuing" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Soft pseudo-labelling described in Sec. 3.3 produces an equipartite matching between the novel points and the class centroids. However, it is highly likely that batches are sampled with point clouds containing novel classes with different cardinalities when dealing with 3D data. It is also likely that some scenes may contain only a subset of the novel classes. Therefore, enforcing the equipartitioning constraint in each batch of the dataset could affect the learning of less-frequent (long-tail) classes. As a solution, we introduce a queue \\( \\mathsf{Z}_q \\) containing a randomly extracted portion of the features of the novel points from the previous iterations. We use these additional data to mitigate the potential class imbalance that may occur during training. In practise, we compute \\( \\mathsf{Z} \\gets \\mathsf{Z} \\oplus \\mathsf{Z}_q \\), where \\( \\oplus \\) is the concatenation operator, and execute the Sinkhorn-Knopp algorithm on this" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9396" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.122 + ], + "angle": 0, + "content": "augmented version of \\(\\mathbb{Z}\\). Then, we retain only the pseudolabels for the first \\(m\\) columns of \\(\\mathbb{Q}^*\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.136, + 0.426, + 0.154 + ], + "angle": 0, + "content": "3.5. Uncertainty-aware training and queuing" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.161, + 0.47, + 0.297 + ], + "angle": 0, + "content": "We propose to carefully select novel points for training \\( f_{\\Theta} \\) with fewer but more reliable pseudo-labels and to build a more effective queue \\( Z_{q} \\). We perform this selection by applying a threshold to the class probabilities of the novel class pseudo-labels. We found that seeking a fixed threshold for all the novel classes, that is also compatible with the variations of the class probabilities during training, is impractical. Therefore, we employ an adaptive threshold based on the class probabilities within each batch." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.298, + 0.471, + 0.419 + ], + "angle": 0, + "content": "Our selection strategy operates as follows. Let \\(\\tau_{c}\\) be the adaptive threshold for the points of the novel class \\(c\\in \\mathcal{C}_n\\). Firstly, we extract the novel points that have the greatest class probability for the class \\(c\\). Secondly, we compute \\(\\tau_{c}\\) as the \\(p\\)-th percentile of the class probabilities of these novel points. Lastly, we retain the novel points of class \\(c\\) whose class probability is above the threshold \\(\\tau_{c}\\). We define this selection strategy as the function" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.433, + 0.47, + 0.452 + ], + "angle": 0, + "content": "\\[\n\\phi : \\left(\\mathcal {F} _ {n}, \\hat {\\mathcal {Y}} _ {n}\\right) \\times p \\mapsto \\left(\\bar {\\mathcal {F}} _ {n}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.465, + 0.47, + 0.545 + ], + "angle": 0, + "content": "where \\(\\mathcal{F}_n\\) is the set of feature vectors extracted from \\(f_{\\xi}\\) and \\(\\hat{\\mathcal{V}}_n\\) is the set of class probabilities predicted by the network for these points. \\(\\bar{\\mathcal{F}}_n\\) are both processed by the Sinkhorn-Knopp algorithm to generate our pseudo-labels and added to \\(\\mathbb{Z}_q\\) to make it more effective." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.557, + 0.289, + 0.573 + ], + "angle": 0, + "content": "3.6. Optimisation objective" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.581, + 0.469, + 0.749 + ], + "angle": 0, + "content": "We optimise \\( f_{\\Theta} \\) by using the weighted Cross Entropy objective based on the labels \\( \\mathcal{Y}_b \\) of the base samples and the pseudo-labels \\( \\tilde{\\mathcal{Y}}_n \\) of the novel samples. We formulate a swapped prediction task based on these pseudo-labels [5]. Secicfically, we begin by generating two different augmentations of \\( \\mathcal{X} \\) that we define as \\( \\mathcal{X}' \\) and \\( \\mathcal{X}'' \\). We use the known one-hot labels for \\( \\mathcal{Y}_b \\) and the predicted soft pseudolabels for \\( \\tilde{\\mathcal{Y}}_n \\). We predict the novel pseudo-labels \\( \\tilde{\\mathcal{Y}}_n' \\) and \\( \\tilde{\\mathcal{Y}}_n'' \\) of the respective point clouds \\( \\mathcal{X}' \\) and \\( \\mathcal{X}'' \\) with our approach. Then, we enforce prediction consistency between the swapped pseudo-labels of the two augmentations as" + }, + { + "type": "equation", + "bbox": [ + 0.165, + 0.761, + 0.47, + 0.78 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} (\\mathcal {X}) = \\ell \\left(\\hat {\\mathcal {Y}} ^ {\\prime}, \\tilde {\\mathcal {Y}} ^ {\\prime \\prime}\\right) + \\ell \\left(\\hat {\\mathcal {Y}} ^ {\\prime \\prime}, \\tilde {\\mathcal {Y}} ^ {\\prime}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.792, + 0.47, + 0.903 + ], + "angle": 0, + "content": "where \\(\\hat{\\mathcal{Y}}^{\\prime} = \\hat{\\mathcal{Y}}_{b}^{\\prime}\\cup \\hat{\\mathcal{Y}}_{n}^{\\prime}\\) (same for \\(\\hat{\\mathcal{Y}}^{\\prime \\prime}\\)), \\(\\tilde{\\mathcal{Y}}^{\\prime} = \\mathcal{Y}_{b}^{\\prime}\\cup \\tilde{\\mathcal{Y}}_{n}^{\\prime}\\) (same for \\(\\tilde{\\mathcal{Y}}^{\\prime \\prime}\\)) and \\(\\ell\\) is the weighted Cross Entropy loss. We use separate segmentation heads for base classes and novel classes. The weights of the loss for the base classes are computed based on their occurrence frequency in the training set. The weights of the loss for the novel classes are all set equally as their occurrence frequency in the dataset is unknown." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.09, + 0.822, + 0.108 + ], + "angle": 0, + "content": "4. Adapting NCD for 2D images to 3D" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.115, + 0.892, + 0.462 + ], + "angle": 0, + "content": "Another contribution of this work is to adapt the method proposed by Zhao et al. [40] for NCD for 2D semantic segmentation (EUMS) to 3D data. Our empirical evaluation (see Sec. 5) shows that the transposition of EUMS to the 3D domain has some limitations. In particular, as described in Sec. 1, EUMS uses two assumptions: I) the novel classes belong to the foreground and II) each image can contain at most one novel class. This allows EUMS to leverage a saliency detection model to produce a foreground mask and a segmentation model pre-trained on the base classes to determine which portion of the image is background. The portion of the image that belongs to both the foreground mask and the background mask is where features are then pooled. EUMS computes a feature representation for each image by average pooling the features of the pixels belonging to the unknown portion. The feature representations of all the images in the dataset are clustered with K-Means by using the number of classes to discover as the target number of clusters. EUMS shows that overclustering and entropy-based modelling can be exploited to improve the results. The affiliation of a point to its cluster is used to produce hard pseudo-labels that are in turn used along with the ground-truth labels to fine-tune the pre-trained model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.463, + 0.893, + 0.795 + ], + "angle": 0, + "content": "With 3D point clouds, there is no concept of foreground and background (in contrast with I). Our adaptation is designed to discover the classes of all the unlabelled points (in contrast with II). Therefore, given the unlabelled points of each point cloud, we randomly extract a subset of these by setting a ratio (e.g. \\(30\\%\\)) with upper bound (e.g. 1K) on the number of points to select. We compute and collect their features for all the point clouds in the dataset and apply K-Means on the whole set of features. Note that this clustering step is computationally expensive, and we had to use High Performance Computing to execute it. The subsampling of the points was necessary to fit the data in the RAM (see Sec. 5 for a detailed analysis). Once the cluster prototypes are computed, we produce the hard pseudo-labels. To enrich the set of pseudo-labels, we propagate the pseudo-label of each point to its nearest neighbour in the coordinate space. This allows us to expand the subset of pseudo-labelled randomly selected points. We also implement the other steps of overclustering and entropy-based modelling to boost the results. Lastly, we fine-tune our model with these pseudo-labels. We name our transposition of EUMS as \\(\\mathrm{EUMS}^{\\dagger}\\) and report its block diagram in the Supplementary Material." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.808, + 0.7, + 0.825 + ], + "angle": 0, + "content": "5. Experimental results" + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.833, + 0.688, + 0.849 + ], + "angle": 0, + "content": "5.1. Experimental setup" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Datasets. We evaluate our approach on SemanticKITTI [3, 4,13] and SemanticPOSS [24]. SemanticKITTI [4] consists of 43,552 point cloud acquisitions with point-level annota" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9397" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.47, + 0.118 + ], + "angle": 0, + "content": "Table 1. SemanticKITTI splits, is defined as KITTI-\\(n^i\\), where \\(n\\) is the number of novel classes and \\(i\\) is the split index." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.122, + 0.468, + 0.205 + ], + "angle": 0, + "content": "
SplitNovel Classes
KITTI-50building, road, sidewalk, terrain, vegetation
KITTI-51car, fence, other-ground, parking, trunk
KITTI-52motorcycle, other-vehicle, pole, traffic-sign, truck
KITTI-43bicycle, bicyclist, motorcyclist, person
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.221, + 0.468, + 0.249 + ], + "angle": 0, + "content": "Table 2. SemanticPOSS splits, defined as \\( \\mathrm{POSS} - n^i \\), where \\( n \\) is the number of novel classes and \\( i \\) is the split index." + }, + { + "type": "table", + "bbox": [ + 0.141, + 0.253, + 0.408, + 0.339 + ], + "angle": 0, + "content": "
SplitNovel Classes
POSS-40building, car, ground, plants
POSS-31bike, fence, person
POSS-32pole, traffic-sign, trunk
POSS-33cone-stone, rider, trashcan
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.367, + 0.468, + 0.474 + ], + "angle": 0, + "content": "tions of 19 semantic classes. Based on the official benchmark guidelines [4], we use sequence 08 for validation and the other sequences for training. SemanticPOSS [24] consists of 2,988 real-world point cloud acquisitions with point-level annotations of 13 semantic classes. Based on the official benchmark guidelines [24], we use sequence 03 for validation and the other sequences for training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.476, + 0.468, + 0.686 + ], + "angle": 0, + "content": "Experimental protocol for 3D NCD. Similarly to what proposed by [40] in the 2D domain, we create different splits of each dataset to validate the NCD performance. We create four splits for SemanticKITTI and SemanticPOSS. We refer to these splits as SemanticKITTI-\\(n^i\\) and SemanticPOSS-\\(n^i\\), where \\(i\\) indexes the split. In each set, the novel classes and the base classes correspond to unlabelled and labelled points, respectively. Tabs. 1 & 2 detail the splits of our datasets. These splits are selected based on their class distribution in the dataset and on the semantic relationship between novel and base classes, e.g. in KITTI-43 the base class motorcycle can be helpful to discover the novel class motorcyclist. We report additional details about the selection process in the Supplementary Material." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.688, + 0.468, + 0.779 + ], + "angle": 0, + "content": "We quantify the performance by using the mean Intersection over Union (mIoU), which is defined as the average IoU across the considered classes [4]. We provide separate mIoU values for the base and novel classes. We also report the overall mIoU computed across all the classes in the dataset for completeness." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Implementation Details. We implement our network based on a MinkowskiUNet-34C network [10]. Point-level features are extracted from the penultimate layer. The segmentation heads are implemented as linear layers, producing output logits for each point in the batched point clouds. We train our network for 10 epochs. We use the SGD optimizer, with momentum 0.9 and weight decay 0.0001. Our learning rate scheduler consists of linear warm-up and co" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.091, + 0.892, + 0.244 + ], + "angle": 0, + "content": "sine annealing, with \\(lr_{max} = 10^{-2}\\) and \\(lr_{min} = 10^{-5}\\). We train with batch size equal to 4. We employ 5 segmentation heads, that are used in synergy with an equal number of over-clustering heads, with \\(o = 3\\). In \\(\\phi\\), we set \\(p = 0.5\\) for SemanticKITTI-\\(n^i\\) and \\(p = 0.3\\) for SemanticPOSS-\\(n^i\\). We adapted the implementation of the Sinkhorn-Knopp algorithm [11] from the code provided by [5], with the introduction of the queue and an in-place normalisation steps. Similarly to [5], we set \\(n_{sk\\_iters} = 3\\), while we adopt a linear decay for \\(\\epsilon\\), with \\(\\epsilon_{start} = 0.3\\), \\(\\epsilon_{end} = 0.05\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.253, + 0.698, + 0.269 + ], + "angle": 0, + "content": "5.2. Quantitative analysis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.277, + 0.891, + 0.322 + ], + "angle": 0, + "content": "Segmentation quality. Tabs. 3 & 4 report the quantitative results on SemanticPOSS and SemanticKITTI, respectively. We report the Full supervision setting as our upper bound." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.323, + 0.892, + 0.488 + ], + "angle": 0, + "content": "On SemanticPOSS, we outperform EUMS† on three out of four splits with an improvement of 18.3 mIoU on POSS-40, 9.0 mIoU on POSS-31 and 0.6 on POSS-32. In these splits, NOPS shows a large improvement on all the novel classes, with the exception of the class fence in POSS-31 and traffic-sign in POSS-33. Differently, we deem the lower performance in POSS-33 is due to the difficulty and scarce presence of these novel classes. The advantage of EUMS† is the clustering on the whole dataset that enables a complete visibility of all the novel classes. On average, NOPS achieves 21.40 mIoU, improving over EUMS† of 6.5 mIoU." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.489, + 0.892, + 0.639 + ], + "angle": 0, + "content": "On SemanticKITTI, we outperform EUMS† on all the four splits, improving by 12.6 mIoU on KITTI-5⁰, 1.2 mIoU on KITTI-5¹, 4.2mIoU on KITTI-5² and 5.3 mIoU on KITTI-4³. NOPS outperforms EUMS† by a large margin on the majority of the novel classes. Exceptions are the class sidewalk in KITTI-5⁰, car in KITTI-5¹, motorcycle in KITTI-5² and motorcyclist in KITTI-4³. On average, NOPS achieves 22.84 mIoU, improving over EUMS† of 5.8 mIoU. Interestingly, NOPS outperforms also the supervised upper bound on the class trunk in KITTI-5¹." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.64, + 0.892, + 0.805 + ], + "angle": 0, + "content": "Computational time. NOPS outperforms EUMS† in terms of computational time. Firstly, EUMS† requires a pretraining step and a fine-tuning step, i.e. 30 training epochs in total. Then, EUMS† requires a large amount of memory (up to 200 GB memory for KITTI-5⁰) to store the data required for clustering, taking several hours (50 hrs) to complete the training procedure. Differently, NOPS achieves superior performance with 10 training epochs, by using less memory (10 GB max) and a lower computational time (up to 25 hrs for KITTI-5⁰). We run these tests using one GPU Tesla A40-48GB." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.817, + 0.688, + 0.833 + ], + "angle": 0, + "content": "5.3. Qualitative analysis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Fig. 3 shows some segmentation results of NOPS and EUMS† on SemanticPOSS and SemaniticKITTI. We can observe that the predictions of the base classes in the two datasets are correct for both the models, with just minor er" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9398" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.09, + 0.892, + 0.119 + ], + "angle": 0, + "content": "Table 3. Novel class discovery results on SemanticPOSS. NOPs outperforms EUMS† on three out of four splits. Full supervision: model trained with labels for base and novel classes. EUMS†: baseline described in Sec. 4. Highlighted values are the novel classes in each split." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.123, + 0.892, + 0.302 + ], + "angle": 0, + "content": "
SplitModelbluebub-carcome.fencegrou.pers.plantspoleridertraf.trafe.trunkNovelmIoU BaseAll
Full supervision43.2071.3033.0032.5044.6078.5061.8073.9030.9054.7026.7011.0019.30--44.72
POSS-40EUMS† [40]25.673.980.5616.4429.4036.7643.8428.4613.1326.7518.183.3416.9117.4421.5220.26
NOPS (Ours)35.4730.351.2413.5224.1369.1444.7042.0719.1947.6524.448.1721.8235.7026.5729.38
POSS-31EUMS† [40]15.1767.9828.0223.9811.8875.0735.9874.4626.9148.5626.005.6023.0521.0139.9635.59
NOPS (Ours)29.3571.3528.7012.213.9478.2456.7874.2118.2938.8823.3113.7423.5130.0238.2436.35
POSS-32EUMS† [40]40.1469.4527.6713.5034.8676.0354.6675.595.2739.227.798.5211.858.3143.9635.74
NOPS (Ours)37.1671.8129.7414.6428.3877.5352.0973.0011.5147.110.5410.2014.798.9544.1736.04
POSS-33EUMS† [40]41.1770.6828.084.3438.2776.6638.2975.3525.7634.3428.310.3624.4013.0144.7037.38
NOPS (Ours)38.5570.3630.910.0029.3876.5055.9871.8417.0331.8726.150.9522.5710.9443.9336.32
AvgEUMS† [40]14.9437.5432.24
NOPS (Ours)21.4038.2334.52
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.316, + 0.892, + 0.345 + ], + "angle": 0, + "content": "Table 4. Novel class discovery results on SemanticKITTI. NOPS outperforms EUMS† on all four splits. Full supervision: model trained with annotations for base and novel classes. EUMS†: baseline described in Sec. 4. Highlighted values are the novel classes in each split." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.348, + 0.891, + 0.513 + ], + "angle": 0, + "content": "
SplitModelbicyclebicyclebicyclecarfencepolepoleothersothersparkpers.poleroadsidewterracetraff.trucktrunkvessels.NovelmIoU BaseAll
Full supervision6.3039.5085.4090.0023.2020.305.703.9018.0028.9031.0040.6090.9074.6062.1020.5062.9046.2083.90--43.89
KITTI-50EUMS† [40]5.2839.9615.7779.209.0316.892.520.0711.3914.4012.6729.1742.5826.100.0510.3047.3737.9238.3524.5721.0823.11
NOPS (Ours)5.5947.7652.6882.6013.7625.551.361.6614.5219.8025.8632.1256.748.0823.8414.2849.4136.1844.1737.1024.7029.62
KITTI-51EUMS† [40]7.5342.4179.9776.778.6219.581.390.5712.0314.1413.9540.7486.3266.4556.2911.9744.7920.9472.4024.2137.0635.62
NOPS (Ours)7.3651.2384.5350.877.2728.931.760.0022.2019.3930.4237.6190.0772.1860.7516.7857.3449.2585.1225.3643.0940.69
KITTI-52EUMS† [40]8.2650.7882.9888.0517.882.752.320.173.1625.4024.9820.2088.3071.0457.858.6327.1638.3676.9512.3842.2236.59
NOPS (Ours)6.7249.2486.3690.7923.682.690.581.8715.4629.4827.9236.3990.2673.3961.2117.8310.3246.1684.2916.5444.8039.72
KITTI-43EUMS† [40]3.952.4780.1087.2116.8114.0214.980.3114.1320.776.8037.5986.7966.5055.2616.2040.6238.3776.157.0543.3935.74
NOPS (Ours)2.3227.8386.0489.8923.0624.472.923.0618.1930.0916.3239.9090.6573.5161.0417.4049.7644.0183.1812.3548.9541.24
AvgEUMS† [40]17.0535.9432.76
NOPS (Ours)22.8440.3837.73
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.54, + 0.47, + 0.676 + ], + "angle": 0, + "content": "rors at the edges of the objects. NOPS shows better segmentation capabilities also when dealing with the novel classes, being able to properly segment the correct class for the unknown objects. On the larger objects, such as buildings in the scenes, some mixing of labels can be observed. Differently, EUMS nearly fails in correctly recognising the novel objects, resulting in the inclusion of different classes (either novel or base) into the same object, e.g. the facade of the building in POSS-3\\(^1\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.692, + 0.238, + 0.708 + ], + "angle": 0, + "content": "6. Ablation studies" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.718, + 0.469, + 0.748 + ], + "angle": 0, + "content": "We study the main components of NOPS and its behaviour when varying its parameters on SemanticPOSS." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.469, + 0.824 + ], + "angle": 0, + "content": "Method components. Fig. 4 shows the performance on novel and base classes of seven versions of NOPS. The first three versions use the pre-trained model on the base classes, while the last four versions use the model trained from scratch. Each version is defined as follows:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.826, + 0.468, + 0.855 + ], + "angle": 0, + "content": "- P: we use a pre-trained model, and we remove \\( Z_{q} \\), \\( \\tau_{c} \\) and the over-clustering heads." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.856, + 0.318, + 0.871 + ], + "angle": 0, + "content": "- OC: P + over-clustering heads." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.871, + 0.453, + 0.886 + ], + "angle": 0, + "content": "- Q: \\(\\mathrm{OC} + Z_{q}\\), i.e. our queue without class-balancing." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.886, + 0.297, + 0.901 + ], + "angle": 0, + "content": "- NP: Q without pre-training." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.826, + 0.468, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.54, + 0.864, + 0.554 + ], + "angle": 0, + "content": "- NP+: NP + our selection function \\(\\phi\\) on the queue." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.555, + 0.89, + 0.583 + ], + "angle": 0, + "content": "- NP++: NP + \\(\\tau_{c}\\) on the features used to derive the pseudo-labels." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.584, + 0.84, + 0.599 + ], + "angle": 0, + "content": "Full: NOPS with all the components activated." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.54, + 0.89, + 0.599 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.607, + 0.892, + 0.789 + ], + "angle": 0, + "content": "Pre-trained approaches generally underperform their trained-from-scratch counterparts on the novel classes. This is visible in the low performance of P, OC and Q. We have a significant improvement when pre-training is not used (NP), i.e. we achieve \\(20.26\\mathrm{mIoU}\\). We can see that the queue both with and without pre-training is helpful. When we add the feature selection for the queue and for the training, i.e. \\(\\mathrm{NP+}\\) and \\(\\mathrm{NP++}\\), we have improvements, i.e. 20.63 mIoU and 20.90 mIoU, respectively. The best performance is achieved with Full. Although we can observe variations on the performance of the base classes, their information is retained by the network when we discover the novel ones." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Parameter analysis. We study the behaviour of the percentile \\( p \\) in our selection function \\( \\phi \\), when we apply it to the features both for pseudo-labelling and for the class-balanced queue \\( Z_{q} \\). Tab. 5 reports the results on each split of SemanticPOSS. For each split, we can observe that the performance depends on the number of points and difficulty of the novel classes. In \\( \\mathrm{POSS - 4^0} \\) and \\( \\mathrm{POSS - 3^1} \\), lower values" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9399" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.067, + 0.075, + 0.891, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.34, + 0.893, + 0.385 + ], + "angle": 0, + "content": "Figure 3. Qualitative comparisons on (top) SemanticPOSS and (bottom) SemanticKITTI. We report results on both base and novel classes. On novel classes, \\(\\mathrm{EUMS}^{\\dagger}\\) fails in recognising novel objects with mixed and cluttered predictions, e.g. the facade of building in \\(\\mathrm{POSS - 3^{1}}\\). NOPS shows superior segmentation performance on all the novel classes." + }, + { + "type": "image", + "bbox": [ + 0.078, + 0.405, + 0.472, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.519, + 0.471, + 0.603 + ], + "angle": 0, + "content": "Figure 4. Ablation study with different components and initialisation strategies on SemanticPOSS. In P, OC and Q, we initialise the model after base pre-training, and use different configurations of the over-clustering heads and of our queue balancing. In NP NP+, NP++ and Full, we begin with Q, we avoid pre-training, and we use \\(\\phi\\) and \\(\\tau_{c}\\) incrementally. See Sec. 6 for definition of methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.631, + 0.47, + 0.738 + ], + "angle": 0, + "content": "of \\( p \\) result in less severe selection. We believe this is related to the class distribution within these splits. This is in line with what observed in Tab. 3. In \\( \\mathrm{POSS}-3^{2} \\) and \\( \\mathrm{POSS}-3^{3} \\), we notice a different behaviour, a higher value of \\( p \\) provides better results. We relate this to the difficulty of the novel classes in these splits whose noisy pseudo-labels can benefit from a more rigorous selection of the features." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.754, + 0.205, + 0.77 + ], + "angle": 0, + "content": "7. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.903 + ], + "angle": 0, + "content": "We explored the new problem of novel class discovery for 3D point cloud segmentation. Firstly, we transposed the only NCD method for 2D semantic segmentation to 3D point cloud data, and experimentally found that it has several limitations. We discussed that extending 2D NCD approaches to 3D data (point clouds) is not trivial because the assumptions made for 2D data are not easily transferable to 3D. Secondly, we presented NOPS, to tackle NCD for point" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.407, + 0.895, + 0.492 + ], + "angle": 0, + "content": "Table 5. Ablation study showing how different values of \\( p \\) affect the performance on SemanticPOSS. The lower \\( p \\) is, the less severe the selection of the features, resulting in better performances for POSS-4\\(^0\\). Differently, POSS-3\\(^3\\) benefits from an higher value of \\( p \\), which leads to a more vigorous filtering of the features. POSS-3\\(^1\\) and POSS-3\\(^2\\) show the best performances with \\( p = 0.5 \\)" + }, + { + "type": "table", + "bbox": [ + 0.534, + 0.494, + 0.86, + 0.611 + ], + "angle": 0, + "content": "
SplitPercentile p
0.10.30.50.70.9
POSS-4030.8135.7028.7730.9326.69
POSS-3128.3330.0230.4323.3218.91
POSS-328.078.9510.3210.257.76
POSS-3310.5510.9411.6914.3813.42
Avg.19.4421.4020.3019.7216.70
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.641, + 0.892, + 0.777 + ], + "angle": 0, + "content": "cloud segmentation by using online clustering and exploiting uncertainty quantification to produce pseudo-labels for the novel points. Lastly, we introduced a new evaluation protocol to assess the performance of NCD for point cloud segmentation. Experiments on two different segmentation dataset showed that NOPS outperforms the compared baselines by a large margin. Future research directions could investigate the extension of our method when base annotations are fewer and/or weakly labelled." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Limitations NOPS limitations include the prior knowledge on the number of novel classes \\( C_n \\) to discover. This could be a limitation when \\( C_n \\) is not known a-priori and novel classes may appear in an incremental manner. We believe that a solution may be to learn novel classes incrementally. Another limitation is the loss we use to handle class unbalancing. More recent techniques to handle this drawback could be further explored." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9400" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.117, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] I. Alonso, L. Riazuelo, L. Montesano, and A.C. Murillo. 3D MiniNet: Learning a 2D representation from point clouds for fast and efficient 3D LiDAR semantic segmentation. In IROS, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.174, + 0.47, + 0.215 + ], + "angle": 0, + "content": "[2] Y.M. Asano, C. Rupprecht, and A. Vedaldi. Self-labelling via simultaneous clustering and representation learning. In ICLR, 2020. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.218, + 0.47, + 0.272 + ], + "angle": 0, + "content": "[3] J. Behley, M. Garbade, A. Milioto, J. Quenzel, S. Behnke, J. Gall, and C. Stachniss. Towards 3D LiDAR-based semantic scene understanding of 3D point cloud sequences: The SemanticKITTI Dataset. In IJRR, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.276, + 0.47, + 0.33 + ], + "angle": 0, + "content": "[4] J. Behley, M. Garbade, A. Milioto, J. Quenzel, S. Behnke, C. Stachniss, and J. Gall. SemanticKITTI: A dataset for semantic scene understanding of LiDAR sequences. In CVPR, 2019. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.333, + 0.47, + 0.375 + ], + "angle": 0, + "content": "[5] M. Caron, I. Misra, J. Mairal, P. Goyal, P. Bojanowski, and A. Joulin. Unsupervised learning of visual features by contrasting cluster assignments. In NeurIPS, 2020. 2, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.377, + 0.47, + 0.419 + ], + "angle": 0, + "content": "[6] J. Cen, P. Yun, S. Zhang, J. Cai, D. Luan, M. Tang, M. Liu, and M.Y. Wang. Open-world semantic segmentation for lidar point clouds. In ECCV, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.421, + 0.47, + 0.476 + ], + "angle": 0, + "content": "[7] L.C. Chen, G. Papandreou, I. Kokkinos, K. Murphy, and A.L. Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. In T-PAMI, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.479, + 0.47, + 0.52 + ], + "angle": 0, + "content": "[8] L.C. Chen, Y. Zhu, G. Papandreou, F. Schroff, and A. Hartwig. Encoder-decoder with atrous separable convolution for semantic image segmentation. In ECCV, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.523, + 0.47, + 0.577 + ], + "angle": 0, + "content": "[9] R. Cheng, R. Razani, E. Taghavi, E. Li, and B. Liu. (AF)2-S3Net: Attentive Feature Fusion with Adaptive Feature Selection for Sparse Semantic Segmentation Network. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.58, + 0.47, + 0.621 + ], + "angle": 0, + "content": "[10] C. Choy, JY. Gwak, and S. Savarese. 4D spatio-temporal convnets: Minkowski convolutional neural networks. In CVPR, 2019. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.624, + 0.47, + 0.652 + ], + "angle": 0, + "content": "[11] M. Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. In NeurIPS, 2013. 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.654, + 0.47, + 0.695 + ], + "angle": 0, + "content": "[12] E. Fini, E. Sangineto, S. Lathuiliere, Z. Zhong, M. Nabi, and E. Ricci. A unified objective for novel class discovery. In CVPR, 2021. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.698, + 0.47, + 0.739 + ], + "angle": 0, + "content": "[13] A. Geiger, P. Lenz, and R. Urtasun. Are we ready for Autonomous Driving? The KITTI Vision Benchmark Suite. In CVPR, 2012. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.742, + 0.47, + 0.782 + ], + "angle": 0, + "content": "[14] B. Graham, M. Engelcke, and L. van der Maaten. 3D semantic segmentation with submanifold sparse convolutional networks. In CVPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.786, + 0.47, + 0.813 + ], + "angle": 0, + "content": "[15] B. Graham and L. van der Maaten. Submanifold sparse convolutional networks. arXiv, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.816, + 0.47, + 0.856 + ], + "angle": 0, + "content": "[16] K. Han, A. Vedaldi, and A. Zisserman. Learning to discover novel visual categories via deep transfer clustering. In CVPR, 2019. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.859, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[17] Q. Hu, B. Yang, L. Xie, S. Rosa, Y. Guo, Z. Wang, N. Trigoni, and A. Markham. Randla-net: Efficient semantic segmentation of large-scale point clouds. In CVPR, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.117, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[18] X. Ji, J.F. Henriques, and A. Vedaldi. Invariant information clustering for unsupervised image classification and segmentation. In ICCV, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.175 + ], + "angle": 0, + "content": "[19] X. Jia, K. Han, Y. Zhu, and B. Green. Joint representation learning and novel category discovery on single-and multimodal data. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.178, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[20] K. Joseph, S. Paul, G. Aggarwal, S. Biswas, P. Rai, K. Han, and V.N. Balasubramanian. Novel class discovery without forgetting. ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[21] G. Mei, F. Poiesi, C. Saltori, J. Zhang, E. Ricci, and N. Sebe. Overlap-guided gaussian mixture models for point cloud registration. In WACV, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.263, + 0.892, + 0.304 + ], + "angle": 0, + "content": "[22] G. Mei, C. Saltori, F. Poiesi, J. Zhang, E. Ricci, N. Sebe, and Q. Wu. Data augmentation-free unsupervised learning for 3d point cloud understanding. In BMVC, 2022. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.305, + 0.892, + 0.346 + ], + "angle": 0, + "content": "[23] A. Milioto, I. Vizzo, J. Behley, and C. Stachniss. Rangenet++: Fast and accurate lidar semantic segmentation. In IROS, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.349, + 0.892, + 0.389 + ], + "angle": 0, + "content": "[24] Y. Pan, B. Gao, J. Mei, S. Geng, C. Li, and H. Zhao. SemanticPOSS: A point cloud dataset with large quantity of dynamic instances. In IV, 2020. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.391, + 0.892, + 0.431 + ], + "angle": 0, + "content": "[25] C.R. Qi, H. Su, K. Mo, and L.J. Guibas. Pointnet: Deep learning on point sets for 3D classification and segmentation. In CVPR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.433, + 0.892, + 0.473 + ], + "angle": 0, + "content": "[26] C.R. Qi, L. Yi, H. Su, and L.J. Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In NeurIPS, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.476, + 0.892, + 0.517 + ], + "angle": 0, + "content": "[27] O. Ronneberger, P. Fischer, and T. Brox. U-net: Convolutional networks for biomedical image segmentation. In MIC-CAI, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.519, + 0.892, + 0.546 + ], + "angle": 0, + "content": "[28] S. Roy, M. Liu, Z. Zhong, N. Sebe, and E. Ricci. Class incremental novel class discovery. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.548, + 0.892, + 0.589 + ], + "angle": 0, + "content": "[29] C. Saltori, F. Galasso, G. Fiameni, N. Sebe, E. Ricci, and F. Poiesi. Cosmix: Compositional semantic mix for domain adaptation in 3d lidar segmentation. In ECCV, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.591, + 0.892, + 0.631 + ], + "angle": 0, + "content": "[30] R. Song, W. Zhang, Y. Zhao, Y. Liu, and P.L. Rosin. Mesh saliency: An independent perceptual measure or a derivative of image saliency? In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.633, + 0.892, + 0.673 + ], + "angle": 0, + "content": "[31] N. Souly, C. Spampinato, and M. Shah. Semi supervised semantic segmentation using generative adversarial network. In ICCV, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.676, + 0.892, + 0.716 + ], + "angle": 0, + "content": "[32] X. Tang, J. Wang, B. Gao, E. Dellandrea, R. Gaizauskas, and L. Chen. Large scale semi-supervised object detection using visual and semantic knowledge transfer. In CVPR, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.718, + 0.892, + 0.759 + ], + "angle": 0, + "content": "[33] H. Thomas, C.R. Qi, J.-E. Deschaud, B. Marcotegui, F. Goulette, and J. L. Guibas. Kpconv: Flexible and deformable convolution for point clouds. In ICCV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.761, + 0.892, + 0.788 + ], + "angle": 0, + "content": "[34] S. Vaze, K. Han, A. Vedaldi, and A. Zisserman. Generalized category discovery. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.79, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[35] B. Wu, A. Wan, X. Yue, and K. Keutzer. Squeezeseg: Convolutional neural nets with recurrent CRF for real-time road-object segmentation from 3D LiDAR point cloud. In ICRA, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[36] B. Wu, X. Zhou, S. Zhao, X. Yue, and K. Keutzer. Squeeze-segv2: Improved model structure and unsupervised domain adaptation for road-object segmentation from a LiDAR point cloud. In ICRA, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.956 + ], + "angle": 0, + "content": "9401" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.134 + ], + "angle": 0, + "content": "[37] M. Yang, Y. Zhu, J. Yu, A. Wu, and C. Deng. Divide and conquer: Compositional experts for generalized novel class discovery. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.136, + 0.468, + 0.163 + ], + "angle": 0, + "content": "[38] L. Zhang and G.J. Qi. Wcp: Worst-case perturbations for semi-supervised deep learning. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.165, + 0.468, + 0.219 + ], + "angle": 0, + "content": "[39] Y. Zhang, Z. Zhou, P. David, X. Yue, Z. Xi, B. Gong, and H. Foroosh. Polarnet: An improved grid representation for online LiDAR point clouds semantic segmentation. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.221, + 0.468, + 0.262 + ], + "angle": 0, + "content": "[40] Y. Zhao, Z. Zhong, N. Sebe, and G.H. Lee. Novel class discovery in semantic segmentation. In CVPR, 2022. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.264, + 0.468, + 0.304 + ], + "angle": 0, + "content": "[41] Z. Zhong, E. Fini, S. Roy, Z. Luo, E. Ricci, and N. Sebe. Neighborhood contrastive learning for novel class discovery. In CVPR, 2021. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.307, + 0.468, + 0.348 + ], + "angle": 0, + "content": "[42] Z. Zhong, L. Zhu, Z. Luo, S. Li, Y. Yang, and N. Sebe. Openmix: Reviving known knowledge for discovering novel visual categories in an open world. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.35, + 0.468, + 0.377 + ], + "angle": 0, + "content": "[43] Y. Zhou and O. Tuzel. Voxelnet: End-to-end learning for point cloud based 3D object detection. In CVPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.379, + 0.468, + 0.419 + ], + "angle": 0, + "content": "[44] X. Zhu, H. Zhou, T. Wang, F. Hong, Y. Ma, W. Li, H. Li, and D. Lin. Cylindrical and asymmetrical 3D convolution networks for lidar segmentation. In CVPR, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.419 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9402" + } + ] +] \ No newline at end of file diff --git a/2023/Novel Class Discovery for 3D Point Cloud Semantic Segmentation/b9953b69-6d05-422e-9f85-28806509ef58_origin.pdf b/2023/Novel Class Discovery for 3D Point Cloud Semantic Segmentation/b9953b69-6d05-422e-9f85-28806509ef58_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a7d52e572ead31ee73eb01c93c44500e0618a1b9 --- /dev/null +++ b/2023/Novel Class Discovery for 3D Point Cloud Semantic Segmentation/b9953b69-6d05-422e-9f85-28806509ef58_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:64b1fc3500f568e5f9fdad6f3b4f6feec55af5273840c3efecd4e27c8d10df0c +size 6641351 diff --git a/2023/Novel Class Discovery for 3D Point Cloud Semantic Segmentation/full.md b/2023/Novel Class Discovery for 3D Point Cloud Semantic Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..1df389f00a1a5932313cfb0aac6c577d2a320005 --- /dev/null +++ b/2023/Novel Class Discovery for 3D Point Cloud Semantic Segmentation/full.md @@ -0,0 +1,254 @@ +# Novel Class Discovery for 3D Point Cloud Semantic Segmentation + +Luigi Riz $^{1}$ Cristiano Saltori $^{2}$ Elisa Ricci $^{1,2}$ Fabio Poiesi $^{1}$ + +$^{1}$ Fondazione Bruno Kessler $^{2}$ University of Trento + +![](images/db2cf5c1cac45253f849f7e282e42dd304853e5ba2874d4667d21166dc24fdba.jpg) +Figure 1. Novel Class Discovery for 3D point cloud semantic segmentation seeks to recognise novel classes by clustering unlabelled novel points with similar semantic features and by exploiting only the knowledge of a set of labelled samples corresponding to the base classes. + +# Abstract + +Novel class discovery (NCD) for semantic segmentation is the task of learning a model that can segment unlabelled (novel) classes using only the supervision from labelled (base) classes. This problem has recently been pioneered for 2D image data, but no work exists for 3D point cloud data. In fact, the assumptions made for 2D are loosely applicable to 3D in this case. This paper is presented to advance the state of the art on point cloud data analysis in four directions. Firstly, we address the new problem of NCD for point cloud semantic segmentation. Secondly, we show that the transposition of the only existing NCD method for 2D semantic segmentation to 3D data is suboptimal. Thirdly, we present a new method for NCD based on online clustering that exploits uncertainty quantification to produce prototypes for pseudo-labelling the points of the novel classes. Lastly, we introduce a new evaluation protocol to assess the performance of NCD for point cloud semantic segmentation. We thoroughly evaluate our method on SemanticKITTI and SemanticPOSS datasets, showing that it can significantly outperform the baseline. Project page: https://github.com/LuigiRiz/NOPS. + +This project has received funding from the European Union's Horizon Europe research and innovation programme under grant agreement No 101058589. This work was also partially supported by the PRIN project LEGO-AI (Prot. 2020TA3K9N), the EU ISFP PROTECTOR (101034216) project and the EU H2020 MARVEL (957337) project and, it was carried out in the Vision and Learning joint laboratory of FBK and UNITN. + +# 1. Introduction + +As humans, we are fairly skilled in organising new visual knowledge (novelty) into homogeneous groups, even when we do not know what we are observing. However, machines cannot perform this task satisfactorily without our supervision. The challenge here is mainly in the formulation of discriminative latent representations of the real world and in the quantification of the uncertainty when the novelty is observed [16,40,41]. The work of Han et al. [16] pioneered the formulation of the Novel Class Discovery (NCD) problem by defining it as the task that aims to classify the samples of an unlabelled dataset into different classes, i.e. the novel samples, by exploiting the knowledge of a set of labelled samples, i.e. the base samples. Note that the classes in the labelled and unlabelled datasets are disjoint. + +NCD has been explored in the 2D image domain for classification [12, 16, 41] and later for semantic segmentation [40]. In particular, Zhao et al. [40] presented the first approach for NCD for 2D semantic segmentation. Two key assumptions were made by the authors. Firstly, at most one novel class is allowed in each image. Secondly, the new class belongs to a foreground object that can be found via saliency detection (e.g. a man on a bicycle, where the bicycle is the novel class). Thanks to these assumptions, the authors could pool the features of each image into a single latent representation and cluster the representations of the whole dataset to discover clusters of novel classes. We argue that these two are important constraints that are not applicable to generic 3D data, in particular to point clouds captured with LiDAR sensors in real-world city-scale scenarios. One point cloud can contain more than one novel class, and the saliency for 3D data cannot be leveraged in the same way as that for 2D data. Although they are both + +related to the attraction of human fixations, 3D saliency is more related to the regional importance of 3D surfaces rather than the foreground/background distinction [30]. Our motivation in exploring NCD for the 3D setting is mainly driven by addressing these shortcomings. + +In this paper, we explore the new problem of NCD for 3D point cloud semantic segmentation (see Fig. 1). Given a partially human-annotated dataset, we jointly learn base and novel semantic classes by clustering unlabelled points with similar semantic features. We adapt the method of Zhao et al. [40] (Entropy-based Uncertainty Modeling and Self-training - EUMS) for point cloud data and use it as our baseline. We go beyond their formulation and, inspired by [5], we integrate batch-level (online) clustering in our method and update prototypes during training in order to make clustering computationally tractable. Cluster assignments are then used as training pseudo-labels. We also exploit over-clustering to achieve a higher clustering accuracy as in EUMS. Because point clouds contain multiple semantic classes, we cannot guarantee that all the classes appear in the point clouds within each batch, some will be missing. Therefore, we design a queuing strategy to store important features over training time, which will be used for pseudolabelling in the case of missing classes. We further introduce a strategy for exploiting the pseudo-label uncertainty to promote the creation of reliable prototypes that we then exploit to produce higher-quality pseudo-labels. Lastly, we produce two augmented views of the same point cloud and impose pseudo-label consistency amongst them. We evaluate our approach on SemanticKITTI [3, 4, 13] and SemanticPOSS [24], introducing an evaluation protocol for NCD and point cloud segmentation that can be adopted in future works. We empirically show that our approach largely outperforms our baseline in both datasets. We also perform an extensive ablation study to demonstrate the importance of the different components of our method. + +# To summarise, our contributions are: + +- We address the new problem of NCD for 3D semantic segmentation; +- We show that the transposition of the only existing NCD method for 2D semantic segmentation [40] to 3D data is suboptimal; +- We present a new method for NCD based on online clustering and uncertainty estimation, which we name it NOPS (NOvel Point Segmentation); +- We introduce a new evaluation protocol to assess the performance of NCD for 3D semantic segmentation. + +# 2. Related work + +Point cloud semantic segmentation can be performed at the point level [26], on range view maps [27], and by voxelising the input points [43]. Point-level networks pro + +cess the input without intermediate representations. Examples of these include PointNet [25], PointNet++ [26], RandLA-Net [17], and KPConv [33]. PointNet [25] and PointNet++ [26] are based on a series of multi-layer perceptron where PointNet++ introduces global and local feature aggregation at multiple scales. RandLA-Net [17] uses random sampling, attentive pooling, and local spatial encoding. KPConv [33] employs flexible and deformable convolutions in a continuous input space. Point-level networks are computationally inefficient when large-scale point clouds are processed. Range view architectures [23] and voxel-based approaches [10] are more computationally efficient than their point-level counterpart. The former requires projecting the input points on a 2D dense map, processing input maps with 2D convolutional filters [27], and reprojecting predictions to the initial 3D space. SqueezeSeg networks [35, 36], 3D-MiniNet [1], RangeNet++ [23], and PolarNet [39] are examples of this category. Although they are more efficient, these approaches tend to lose information during the projection and re-projection phase. The latter includes 3D quantisation-based approaches that discretise the input points into a 3D voxel grid and employ 3D convolutions [43] or 3D sparse convolutions [10, 15] to predict per-voxel classes. VoxelNet [43], SparseConv [14, 15], MinkowskiNet [10], Cylinder3D [44], and $(\mathrm{AF})^2$ -S3Net [9] are architectures belonging to this category. These approaches tackle point cloud segmentation in the supervised settings, whereas we tackle novel class discovery with labelled base classes and unlabelled novel classes. + +Novel class discovery (NCD) is explored for 2D classification [12, 16, 19, 20, 28, 34, 37, 41, 42] and 2D segmentation [40]. NCD is more complex than standard semi-supervised learning [31, 32, 38]. In semi-supervised learning, labelled and unlabelled samples belong to the same classes, while in NCD, novel and base samples belong to disjoint classes. Han et al. [16] pioneered the NCD problem for 2D image classification. A classification model is pretrained on a set of base classes and used as feature extractor for the novel classes. They then train a classifier for the novel classes using the pseudo-labels produced by the pretrained model. Zhong et al. [41] introduced neighbourhood contrastive learning to generate discriminative representations for clustering. They retrieve and aggregate pseudopositive pairs with contrastive learning, encouraging the model to learn more discriminative representations. Hard negatives are obtained by mixing labelled and unlabelled samples in the feature space. UNO [12] unifies the two previous works by using a unique classification loss function for both base and novel classes, where pseudo-labels are processed together with ground-truth labels. NCD without Forgetting [20] and FRoST [28] further extend NCD to the incremental learning setting. EUMS [40] is the only approach analysing the NCD problem for semantic segmen + +![](images/1a61d98d4a2dd970aafcd5cb5669a50c29c66ce080f199adbb7a87f68abde09a.jpg) +Figure 2. Overview of NOPS. We random augment the input point cloud twice and extract point-level features $\mathcal{F}$ with the shared model $f_{\xi}$ . $\mathcal{F}$ are used to obtain pseudo-labels in the online pseudo-labelling. We forward $\mathcal{F}$ to a novel $f_{n}$ and a base $f_{b}$ segmentation layer to output the novel and base predictions, respectively. We optimise our network by minimising a global objective function based on cross entropy. + +tation. Unlike image classification, the model has to classify each pixel and handle multiple classes in each image. EUMS consists of a multi-stage pipeline using a saliency model to cluster the latent representations of novel classes to produce pseudo-labels. Moreover, entropy-based uncertainty and self-training are used to overcome noisy pseudolabels while improving the model performance on the novel classes. In this work, we tackle the problem of NCD in 3D point cloud semantic segmentation. Unlike previous works, our problem inherits the challenges from the fields of 2D semantic segmentation [7, 8] and 3D point cloud segmentation [10, 23, 29]. From 2D semantic segmentation, it inherits the additional challenges of multiple novel classes in the same image and the strong class unbalance. From 3D point cloud segmentation, we inherit the sparsity of input data, the different density of point cloud regions and the inability to identify foreground and background. The latter are not present in 2D segmentation [40]. Unlike [40] that use K-Means, we formulate clustering as an optimal transport problem to avoid degenerate solutions (i.e. all data points may be assigned to the same label and learn a constant representation) [2, 22]. Lastly, related to EUMS, REAL is proposed for open-world 3D semantic segmentation [6], where both known and unknown points have to be segmented. Unlike NOPS, all the unknown points belong to a single class and it is the task of a human annotator to separately label the novel classes. Then, these labels are used to update the base model by incrementally learning the novel classes. + +# 3. Proposed approach + +# 3.1. Overview + +Given an input point cloud, we produce two augmented views that are processed with the same deep neural network to extract point-level features. These features are used to obtain pseudo-labels in the online pseudo-labelling step through the Sinkhorn-Knopp algorithm [11] (Sec. 3.3). Concurrently, we process the same features with the last network layers to segment novel and base classes. These features are stored in the class-balanced queue to mitigate the problem of batches with missing classes (Sec. 3.4). We exploit pseudo-label values (class probabilities) to filter out uncertain points, thus adding to the queue only high-quality points (Sec. 3.5). Lastly, we train our network by minimising the optimisation objective function through a swapped prediction task based on the computed pseudolabels (Sec. 3.6). Fig. 2 shows the block diagram of NOPS. + +# 3.2. Problem formulation + +Let $\mathrm{X} = \{\mathcal{X}\}$ be a dataset of 3D point clouds captured in different scenes. $\mathcal{X}$ is a set composed of a base set $\mathcal{X}_b$ and a novel set $\mathcal{X}_n$ , s.t. $\mathcal{X} = \mathcal{X}_b \cup \mathcal{X}_n$ . The semantic categories that can be present in our point clouds are $\mathcal{C} = \mathcal{C}_b \cup \mathcal{C}_n$ , where $\mathcal{C}_b$ is the set of base classes and $\mathcal{C}_n$ is the set of novel classes, s.t. $\mathcal{C}_b \cap \mathcal{C}_n = \emptyset$ . Each $\mathcal{X} \in \mathrm{X}$ is composed of a finite but unknown number of 3D points $\mathcal{X} = \{(\mathbf{x}, c)\}$ , where $\mathbf{x} \in \mathbb{R}^3$ is the coordinate of the a point and $c$ is its semantic class. We know the class of the point $(\mathbf{x}, c)$ , s.t. $\mathbf{x} \in \mathcal{X}_b$ and + +$c \in \mathcal{C}_b$ , but we do not know the class of the point $(\mathbf{x}, c)$ , s.t. $x \in \mathcal{X}_n$ and $c \in \mathcal{C}_n$ . No points in $\mathcal{X}_n$ belong to one of the base classes $\mathcal{C}_b$ . As in [16, 40, 41], we assume that the number of classes to discover is known, i.e. $|\mathcal{C}_n| = C_n$ . We aim to design a computational approach that trains a deep neural network $f_\Theta$ that can segment all the points of a given point cloud, thus learning to jointly segment base classes $\mathcal{C}_b$ and novel classes $\mathcal{C}_n$ . $\Theta$ are the weights of our deep neural network. $f_\Theta$ is composed of two heads, $f_\Theta = f_\xi \circ \{f_b, f_n\}$ , where $f_b$ is the segmentation head for the base classes, $f_n$ is the segmentation head for the novel classes, $f_\xi$ is the feature extractor network and $\circ$ is the composition operator (Fig. 2). + +# 3.3. Online pseudo-labelling + +We formulate pseudo-labelling as the assignment of novel points to the class-prototypes learnt during training [5]. Let $\mathsf{P} \in \mathbb{R}^{D \times \rho}$ be the class prototypes, where $D$ is the size of the output features from $f_{\xi}$ and $\rho$ is the number of prototypes. Let $\mathsf{Z} \in \mathbb{R}^{D \times m}$ be the normalised output features extracted from $f_{\xi}$ , where $m$ is the number of points of the point cloud. $m$ is not known a priori and it can differ across point clouds. We aim to find the assignment $\mathsf{Q} \in \mathbb{R}^{\rho \times m}$ s.t. all the points in the batch are equally partitioned across the $\rho$ prototypes. This equipartition ensures that the feature representations of the points belonging to different novel classes are well-separated, thus preventing the case in which the novel class feature representations collapse into a unique solution. Caron et al. [5] employs an arbitrary large number of prototypes $\rho$ to effectively organise the feature space produced by $f_{\xi}$ . They discard $\mathsf{P}$ after training. In contrast, we learn exactly $\rho = C_n$ class prototypes and propose to use $\mathsf{P}$ as the weights for our new class segmentation head $f_{n}$ , which outputs the $C_n$ logits for the new classes. In order to optimise the assignment $\mathsf{Q}$ , we maximise the similarity between the features of the new points and the learned prototypes as + +$$ +\max _ {\mathbb {Q} \in \mathcal {Q}} \operatorname {T r} \left(\mathbb {Q} ^ {\top} \mathbb {P} ^ {\top} \mathbb {Z}\right) + \epsilon H (\mathbb {Q}) \rightarrow \mathbb {Q} ^ {*}, \tag {1} +$$ + +where $H$ is the entropy function, $\epsilon$ is the parameter that determines the smoothness of the assignment and $\mathbb{Q}^*$ is our sought solution. Asano et al. [2] enforce the equipartitioning constraint by requiring $\mathbb{Q}$ to belong to a transportation polytope and perform this optimisation on the whole dataset at once (offline). This operation with point cloud data is computationally impractical. Therefore, we formulate the transportation polytope such that the optimisation is performed online, which consists of processing only the points within the batch being processed + +$$ +\mathcal {Q} = \left\{\mathbb {Q} \in \mathbb {R} _ {+} ^ {C _ {n} \times m} | \mathbb {Q} 1 _ {m} = \frac {1}{C _ {n}} 1 _ {C _ {n}}, \mathbb {Q} ^ {\top} 1 _ {C _ {n}} = \frac {1}{m} 1 _ {m} \right\}, \tag {2} +$$ + +where $1_{\star}$ represents a vector of ones of dimension $\star$ . These constraints ensure that each class prototype is selected on average at least $m / C_n$ times in each batch. The solution $\mathbb{Q}^*$ can take the form of a normalised exponential matrix + +$$ +\mathbb {Q} ^ {*} = \operatorname {d i a g} (\alpha) \exp \left(\frac {\mathrm {P} ^ {\top} \mathrm {Z}}{\epsilon}\right) \operatorname {d i a g} (\beta), \tag {3} +$$ + +where $\alpha$ and $\beta$ are renormalization vectors that are computed iteratively with the Sinkhorn-Knopp algorithm [11, 21]. We then transpose the optimised soft assignment $\mathbb{Q}^* \in \mathbb{R}_+^{C_n \times m}$ to obtain the soft pseudo-labels for each of the $m$ novel points being processed within each batch. + +We empirically found that training can be more effective if pseudo-labels are smoother in the first training epochs and peaked in the last training epochs. Therefore, we introduce a linear decay of $\epsilon$ during training. + +Multi-headed segmentation: A single segmentation head may converge to a suboptimal feature space, thus producing suboptimal prototype solutions. To further improve the segmentation quality, we use multiple novel class segmentation heads to optimise $f_{\Theta}$ based on different training solutions. Different solutions increase the likelihood of producing a diverse partitioning of the feature space as they regularise with each other (they share the same backbone) [18]. In practise, we concatenate the logits of the base class segmentation head with the outputs of each novel class segmentation head and we separately evaluate their loss for each novel class segmentation head at training time. + +We task our network to over-cluster novel points, using segmentation heads that output $o \cdot C_n$ logits, where $o$ is the over-clustering factor. Previous studies empirically showed that this is beneficial to learn more informative features [5, 12, 18, 22]. We observed the same and concur that over-clustering can be useful for increasing expressivity of the feature representations. The over-clustering heads are then discarded at inference time. + +# 3.4. Class-balanced queuing + +Soft pseudo-labelling described in Sec. 3.3 produces an equipartite matching between the novel points and the class centroids. However, it is highly likely that batches are sampled with point clouds containing novel classes with different cardinalities when dealing with 3D data. It is also likely that some scenes may contain only a subset of the novel classes. Therefore, enforcing the equipartitioning constraint in each batch of the dataset could affect the learning of less-frequent (long-tail) classes. As a solution, we introduce a queue $\mathsf{Z}_q$ containing a randomly extracted portion of the features of the novel points from the previous iterations. We use these additional data to mitigate the potential class imbalance that may occur during training. In practise, we compute $\mathsf{Z} \gets \mathsf{Z} \oplus \mathsf{Z}_q$ , where $\oplus$ is the concatenation operator, and execute the Sinkhorn-Knopp algorithm on this + +augmented version of $\mathbb{Z}$ . Then, we retain only the pseudolabels for the first $m$ columns of $\mathbb{Q}^*$ . + +# 3.5. Uncertainty-aware training and queuing + +We propose to carefully select novel points for training $f_{\Theta}$ with fewer but more reliable pseudo-labels and to build a more effective queue $Z_{q}$ . We perform this selection by applying a threshold to the class probabilities of the novel class pseudo-labels. We found that seeking a fixed threshold for all the novel classes, that is also compatible with the variations of the class probabilities during training, is impractical. Therefore, we employ an adaptive threshold based on the class probabilities within each batch. + +Our selection strategy operates as follows. Let $\tau_{c}$ be the adaptive threshold for the points of the novel class $c\in \mathcal{C}_n$ . Firstly, we extract the novel points that have the greatest class probability for the class $c$ . Secondly, we compute $\tau_{c}$ as the $p$ -th percentile of the class probabilities of these novel points. Lastly, we retain the novel points of class $c$ whose class probability is above the threshold $\tau_{c}$ . We define this selection strategy as the function + +$$ +\phi : \left(\mathcal {F} _ {n}, \hat {\mathcal {Y}} _ {n}\right) \times p \mapsto \left(\bar {\mathcal {F}} _ {n}\right), \tag {4} +$$ + +where $\mathcal{F}_n$ is the set of feature vectors extracted from $f_{\xi}$ and $\hat{\mathcal{V}}_n$ is the set of class probabilities predicted by the network for these points. $\bar{\mathcal{F}}_n$ are both processed by the Sinkhorn-Knopp algorithm to generate our pseudo-labels and added to $\mathbb{Z}_q$ to make it more effective. + +# 3.6. Optimisation objective + +We optimise $f_{\Theta}$ by using the weighted Cross Entropy objective based on the labels $\mathcal{Y}_b$ of the base samples and the pseudo-labels $\tilde{\mathcal{Y}}_n$ of the novel samples. We formulate a swapped prediction task based on these pseudo-labels [5]. Secicfically, we begin by generating two different augmentations of $\mathcal{X}$ that we define as $\mathcal{X}'$ and $\mathcal{X}''$ . We use the known one-hot labels for $\mathcal{Y}_b$ and the predicted soft pseudolabels for $\tilde{\mathcal{Y}}_n$ . We predict the novel pseudo-labels $\tilde{\mathcal{Y}}_n'$ and $\tilde{\mathcal{Y}}_n''$ of the respective point clouds $\mathcal{X}'$ and $\mathcal{X}''$ with our approach. Then, we enforce prediction consistency between the swapped pseudo-labels of the two augmentations as + +$$ +\mathcal {L} (\mathcal {X}) = \ell \left(\hat {\mathcal {Y}} ^ {\prime}, \tilde {\mathcal {Y}} ^ {\prime \prime}\right) + \ell \left(\hat {\mathcal {Y}} ^ {\prime \prime}, \tilde {\mathcal {Y}} ^ {\prime}\right), \tag {5} +$$ + +where $\hat{\mathcal{Y}}^{\prime} = \hat{\mathcal{Y}}_{b}^{\prime}\cup \hat{\mathcal{Y}}_{n}^{\prime}$ (same for $\hat{\mathcal{Y}}^{\prime \prime}$ ), $\tilde{\mathcal{Y}}^{\prime} = \mathcal{Y}_{b}^{\prime}\cup \tilde{\mathcal{Y}}_{n}^{\prime}$ (same for $\tilde{\mathcal{Y}}^{\prime \prime}$ ) and $\ell$ is the weighted Cross Entropy loss. We use separate segmentation heads for base classes and novel classes. The weights of the loss for the base classes are computed based on their occurrence frequency in the training set. The weights of the loss for the novel classes are all set equally as their occurrence frequency in the dataset is unknown. + +# 4. Adapting NCD for 2D images to 3D + +Another contribution of this work is to adapt the method proposed by Zhao et al. [40] for NCD for 2D semantic segmentation (EUMS) to 3D data. Our empirical evaluation (see Sec. 5) shows that the transposition of EUMS to the 3D domain has some limitations. In particular, as described in Sec. 1, EUMS uses two assumptions: I) the novel classes belong to the foreground and II) each image can contain at most one novel class. This allows EUMS to leverage a saliency detection model to produce a foreground mask and a segmentation model pre-trained on the base classes to determine which portion of the image is background. The portion of the image that belongs to both the foreground mask and the background mask is where features are then pooled. EUMS computes a feature representation for each image by average pooling the features of the pixels belonging to the unknown portion. The feature representations of all the images in the dataset are clustered with K-Means by using the number of classes to discover as the target number of clusters. EUMS shows that overclustering and entropy-based modelling can be exploited to improve the results. The affiliation of a point to its cluster is used to produce hard pseudo-labels that are in turn used along with the ground-truth labels to fine-tune the pre-trained model. + +With 3D point clouds, there is no concept of foreground and background (in contrast with I). Our adaptation is designed to discover the classes of all the unlabelled points (in contrast with II). Therefore, given the unlabelled points of each point cloud, we randomly extract a subset of these by setting a ratio (e.g. $30\%$ ) with upper bound (e.g. 1K) on the number of points to select. We compute and collect their features for all the point clouds in the dataset and apply K-Means on the whole set of features. Note that this clustering step is computationally expensive, and we had to use High Performance Computing to execute it. The subsampling of the points was necessary to fit the data in the RAM (see Sec. 5 for a detailed analysis). Once the cluster prototypes are computed, we produce the hard pseudo-labels. To enrich the set of pseudo-labels, we propagate the pseudo-label of each point to its nearest neighbour in the coordinate space. This allows us to expand the subset of pseudo-labelled randomly selected points. We also implement the other steps of overclustering and entropy-based modelling to boost the results. Lastly, we fine-tune our model with these pseudo-labels. We name our transposition of EUMS as $\mathrm{EUMS}^{\dagger}$ and report its block diagram in the Supplementary Material. + +# 5. Experimental results + +# 5.1. Experimental setup + +Datasets. We evaluate our approach on SemanticKITTI [3, 4,13] and SemanticPOSS [24]. SemanticKITTI [4] consists of 43,552 point cloud acquisitions with point-level annota + +Table 1. SemanticKITTI splits, is defined as KITTI- $n^i$ , where $n$ is the number of novel classes and $i$ is the split index. + +
SplitNovel Classes
KITTI-50building, road, sidewalk, terrain, vegetation
KITTI-51car, fence, other-ground, parking, trunk
KITTI-52motorcycle, other-vehicle, pole, traffic-sign, truck
KITTI-43bicycle, bicyclist, motorcyclist, person
+ +Table 2. SemanticPOSS splits, defined as $\mathrm{POSS} - n^i$ , where $n$ is the number of novel classes and $i$ is the split index. + +
SplitNovel Classes
POSS-40building, car, ground, plants
POSS-31bike, fence, person
POSS-32pole, traffic-sign, trunk
POSS-33cone-stone, rider, trashcan
+ +tions of 19 semantic classes. Based on the official benchmark guidelines [4], we use sequence 08 for validation and the other sequences for training. SemanticPOSS [24] consists of 2,988 real-world point cloud acquisitions with point-level annotations of 13 semantic classes. Based on the official benchmark guidelines [24], we use sequence 03 for validation and the other sequences for training. + +Experimental protocol for 3D NCD. Similarly to what proposed by [40] in the 2D domain, we create different splits of each dataset to validate the NCD performance. We create four splits for SemanticKITTI and SemanticPOSS. We refer to these splits as SemanticKITTI- $n^i$ and SemanticPOSS- $n^i$ , where $i$ indexes the split. In each set, the novel classes and the base classes correspond to unlabelled and labelled points, respectively. Tabs. 1 & 2 detail the splits of our datasets. These splits are selected based on their class distribution in the dataset and on the semantic relationship between novel and base classes, e.g. in KITTI-43 the base class motorcycle can be helpful to discover the novel class motorcyclist. We report additional details about the selection process in the Supplementary Material. + +We quantify the performance by using the mean Intersection over Union (mIoU), which is defined as the average IoU across the considered classes [4]. We provide separate mIoU values for the base and novel classes. We also report the overall mIoU computed across all the classes in the dataset for completeness. + +Implementation Details. We implement our network based on a MinkowskiUNet-34C network [10]. Point-level features are extracted from the penultimate layer. The segmentation heads are implemented as linear layers, producing output logits for each point in the batched point clouds. We train our network for 10 epochs. We use the SGD optimizer, with momentum 0.9 and weight decay 0.0001. Our learning rate scheduler consists of linear warm-up and co + +sine annealing, with $lr_{max} = 10^{-2}$ and $lr_{min} = 10^{-5}$ . We train with batch size equal to 4. We employ 5 segmentation heads, that are used in synergy with an equal number of over-clustering heads, with $o = 3$ . In $\phi$ , we set $p = 0.5$ for SemanticKITTI- $n^i$ and $p = 0.3$ for SemanticPOSS- $n^i$ . We adapted the implementation of the Sinkhorn-Knopp algorithm [11] from the code provided by [5], with the introduction of the queue and an in-place normalisation steps. Similarly to [5], we set $n_{sk\_iters} = 3$ , while we adopt a linear decay for $\epsilon$ , with $\epsilon_{start} = 0.3$ , $\epsilon_{end} = 0.05$ . + +# 5.2. Quantitative analysis + +Segmentation quality. Tabs. 3 & 4 report the quantitative results on SemanticPOSS and SemanticKITTI, respectively. We report the Full supervision setting as our upper bound. + +On SemanticPOSS, we outperform EUMS† on three out of four splits with an improvement of 18.3 mIoU on POSS-40, 9.0 mIoU on POSS-31 and 0.6 on POSS-32. In these splits, NOPS shows a large improvement on all the novel classes, with the exception of the class fence in POSS-31 and traffic-sign in POSS-33. Differently, we deem the lower performance in POSS-33 is due to the difficulty and scarce presence of these novel classes. The advantage of EUMS† is the clustering on the whole dataset that enables a complete visibility of all the novel classes. On average, NOPS achieves 21.40 mIoU, improving over EUMS† of 6.5 mIoU. + +On SemanticKITTI, we outperform EUMS† on all the four splits, improving by 12.6 mIoU on KITTI-5⁰, 1.2 mIoU on KITTI-5¹, 4.2mIoU on KITTI-5² and 5.3 mIoU on KITTI-4³. NOPS outperforms EUMS† by a large margin on the majority of the novel classes. Exceptions are the class sidewalk in KITTI-5⁰, car in KITTI-5¹, motorcycle in KITTI-5² and motorcyclist in KITTI-4³. On average, NOPS achieves 22.84 mIoU, improving over EUMS† of 5.8 mIoU. Interestingly, NOPS outperforms also the supervised upper bound on the class trunk in KITTI-5¹. + +Computational time. NOPS outperforms EUMS† in terms of computational time. Firstly, EUMS† requires a pretraining step and a fine-tuning step, i.e. 30 training epochs in total. Then, EUMS† requires a large amount of memory (up to 200 GB memory for KITTI-5⁰) to store the data required for clustering, taking several hours (50 hrs) to complete the training procedure. Differently, NOPS achieves superior performance with 10 training epochs, by using less memory (10 GB max) and a lower computational time (up to 25 hrs for KITTI-5⁰). We run these tests using one GPU Tesla A40-48GB. + +# 5.3. Qualitative analysis + +Fig. 3 shows some segmentation results of NOPS and EUMS† on SemanticPOSS and SemaniticKITTI. We can observe that the predictions of the base classes in the two datasets are correct for both the models, with just minor er + +Table 3. Novel class discovery results on SemanticPOSS. NOPs outperforms EUMS† on three out of four splits. Full supervision: model trained with labels for base and novel classes. EUMS†: baseline described in Sec. 4. Highlighted values are the novel classes in each split. + +
SplitModelbluebub-carcome.fencegrou.pers.plantspoleridertraf.trafe.trunkNovelmIoU BaseAll
Full supervision43.2071.3033.0032.5044.6078.5061.8073.9030.9054.7026.7011.0019.30--44.72
POSS-40EUMS† [40]25.673.980.5616.4429.4036.7643.8428.4613.1326.7518.183.3416.9117.4421.5220.26
NOPS (Ours)35.4730.351.2413.5224.1369.1444.7042.0719.1947.6524.448.1721.8235.7026.5729.38
POSS-31EUMS† [40]15.1767.9828.0223.9811.8875.0735.9874.4626.9148.5626.005.6023.0521.0139.9635.59
NOPS (Ours)29.3571.3528.7012.213.9478.2456.7874.2118.2938.8823.3113.7423.5130.0238.2436.35
POSS-32EUMS† [40]40.1469.4527.6713.5034.8676.0354.6675.595.2739.227.798.5211.858.3143.9635.74
NOPS (Ours)37.1671.8129.7414.6428.3877.5352.0973.0011.5147.110.5410.2014.798.9544.1736.04
POSS-33EUMS† [40]41.1770.6828.084.3438.2776.6638.2975.3525.7634.3428.310.3624.4013.0144.7037.38
NOPS (Ours)38.5570.3630.910.0029.3876.5055.9871.8417.0331.8726.150.9522.5710.9443.9336.32
AvgEUMS† [40]14.9437.5432.24
NOPS (Ours)21.4038.2334.52
+ +Table 4. Novel class discovery results on SemanticKITTI. NOPS outperforms EUMS† on all four splits. Full supervision: model trained with annotations for base and novel classes. EUMS†: baseline described in Sec. 4. Highlighted values are the novel classes in each split. + +
SplitModelbicyclebicyclebicyclecarfencepolepoleothersothersparkpers.poleroadsidewterracetraff.trucktrunkvessels.NovelmIoU BaseAll
Full supervision6.3039.5085.4090.0023.2020.305.703.9018.0028.9031.0040.6090.9074.6062.1020.5062.9046.2083.90--43.89
KITTI-50EUMS† [40]5.2839.9615.7779.209.0316.892.520.0711.3914.4012.6729.1742.5826.100.0510.3047.3737.9238.3524.5721.0823.11
NOPS (Ours)5.5947.7652.6882.6013.7625.551.361.6614.5219.8025.8632.1256.748.0823.8414.2849.4136.1844.1737.1024.7029.62
KITTI-51EUMS† [40]7.5342.4179.9776.778.6219.581.390.5712.0314.1413.9540.7486.3266.4556.2911.9744.7920.9472.4024.2137.0635.62
NOPS (Ours)7.3651.2384.5350.877.2728.931.760.0022.2019.3930.4237.6190.0772.1860.7516.7857.3449.2585.1225.3643.0940.69
KITTI-52EUMS† [40]8.2650.7882.9888.0517.882.752.320.173.1625.4024.9820.2088.3071.0457.858.6327.1638.3676.9512.3842.2236.59
NOPS (Ours)6.7249.2486.3690.7923.682.690.581.8715.4629.4827.9236.3990.2673.3961.2117.8310.3246.1684.2916.5444.8039.72
KITTI-43EUMS† [40]3.952.4780.1087.2116.8114.0214.980.3114.1320.776.8037.5986.7966.5055.2616.2040.6238.3776.157.0543.3935.74
NOPS (Ours)2.3227.8386.0489.8923.0624.472.923.0618.1930.0916.3239.9090.6573.5161.0417.4049.7644.0183.1812.3548.9541.24
AvgEUMS† [40]17.0535.9432.76
NOPS (Ours)22.8440.3837.73
+ +rors at the edges of the objects. NOPS shows better segmentation capabilities also when dealing with the novel classes, being able to properly segment the correct class for the unknown objects. On the larger objects, such as buildings in the scenes, some mixing of labels can be observed. Differently, EUMS nearly fails in correctly recognising the novel objects, resulting in the inclusion of different classes (either novel or base) into the same object, e.g. the facade of the building in POSS-3 $^1$ . + +# 6. Ablation studies + +We study the main components of NOPS and its behaviour when varying its parameters on SemanticPOSS. + +Method components. Fig. 4 shows the performance on novel and base classes of seven versions of NOPS. The first three versions use the pre-trained model on the base classes, while the last four versions use the model trained from scratch. Each version is defined as follows: + +- P: we use a pre-trained model, and we remove $Z_{q}$ , $\tau_{c}$ and the over-clustering heads. +- OC: P + over-clustering heads. +- Q: $\mathrm{OC} + Z_{q}$ , i.e. our queue without class-balancing. +- NP: Q without pre-training. + +- NP+: NP + our selection function $\phi$ on the queue. +- NP++: NP + $\tau_{c}$ on the features used to derive the pseudo-labels. +Full: NOPS with all the components activated. + +Pre-trained approaches generally underperform their trained-from-scratch counterparts on the novel classes. This is visible in the low performance of P, OC and Q. We have a significant improvement when pre-training is not used (NP), i.e. we achieve $20.26\mathrm{mIoU}$ . We can see that the queue both with and without pre-training is helpful. When we add the feature selection for the queue and for the training, i.e. $\mathrm{NP+}$ and $\mathrm{NP++}$ , we have improvements, i.e. 20.63 mIoU and 20.90 mIoU, respectively. The best performance is achieved with Full. Although we can observe variations on the performance of the base classes, their information is retained by the network when we discover the novel ones. + +Parameter analysis. We study the behaviour of the percentile $p$ in our selection function $\phi$ , when we apply it to the features both for pseudo-labelling and for the class-balanced queue $Z_{q}$ . Tab. 5 reports the results on each split of SemanticPOSS. For each split, we can observe that the performance depends on the number of points and difficulty of the novel classes. In $\mathrm{POSS - 4^0}$ and $\mathrm{POSS - 3^1}$ , lower values + +![](images/b48fc6ebf247d7fbeeae594b6b55c2030f35ce13f6313749a2d85341d1a3955b.jpg) +Figure 3. Qualitative comparisons on (top) SemanticPOSS and (bottom) SemanticKITTI. We report results on both base and novel classes. On novel classes, $\mathrm{EUMS}^{\dagger}$ fails in recognising novel objects with mixed and cluttered predictions, e.g. the facade of building in $\mathrm{POSS - 3^{1}}$ . NOPS shows superior segmentation performance on all the novel classes. + +![](images/2050eaad008e6e9120781e8a12d54f5512ec7210806b4c4bf3e31183ecff57f8.jpg) +Figure 4. Ablation study with different components and initialisation strategies on SemanticPOSS. In P, OC and Q, we initialise the model after base pre-training, and use different configurations of the over-clustering heads and of our queue balancing. In NP NP+, NP++ and Full, we begin with Q, we avoid pre-training, and we use $\phi$ and $\tau_{c}$ incrementally. See Sec. 6 for definition of methods. + +of $p$ result in less severe selection. We believe this is related to the class distribution within these splits. This is in line with what observed in Tab. 3. In $\mathrm{POSS}-3^{2}$ and $\mathrm{POSS}-3^{3}$ , we notice a different behaviour, a higher value of $p$ provides better results. We relate this to the difficulty of the novel classes in these splits whose noisy pseudo-labels can benefit from a more rigorous selection of the features. + +# 7. Conclusions + +We explored the new problem of novel class discovery for 3D point cloud segmentation. Firstly, we transposed the only NCD method for 2D semantic segmentation to 3D point cloud data, and experimentally found that it has several limitations. We discussed that extending 2D NCD approaches to 3D data (point clouds) is not trivial because the assumptions made for 2D data are not easily transferable to 3D. Secondly, we presented NOPS, to tackle NCD for point + +Table 5. Ablation study showing how different values of $p$ affect the performance on SemanticPOSS. The lower $p$ is, the less severe the selection of the features, resulting in better performances for POSS-4 $^0$ . Differently, POSS-3 $^3$ benefits from an higher value of $p$ , which leads to a more vigorous filtering of the features. POSS-3 $^1$ and POSS-3 $^2$ show the best performances with $p = 0.5$ + +
SplitPercentile p
0.10.30.50.70.9
POSS-4030.8135.7028.7730.9326.69
POSS-3128.3330.0230.4323.3218.91
POSS-328.078.9510.3210.257.76
POSS-3310.5510.9411.6914.3813.42
Avg.19.4421.4020.3019.7216.70
+ +cloud segmentation by using online clustering and exploiting uncertainty quantification to produce pseudo-labels for the novel points. Lastly, we introduced a new evaluation protocol to assess the performance of NCD for point cloud segmentation. Experiments on two different segmentation dataset showed that NOPS outperforms the compared baselines by a large margin. Future research directions could investigate the extension of our method when base annotations are fewer and/or weakly labelled. + +Limitations NOPS limitations include the prior knowledge on the number of novel classes $C_n$ to discover. This could be a limitation when $C_n$ is not known a-priori and novel classes may appear in an incremental manner. We believe that a solution may be to learn novel classes incrementally. Another limitation is the loss we use to handle class unbalancing. More recent techniques to handle this drawback could be further explored. + +# References + +[1] I. Alonso, L. Riazuelo, L. Montesano, and A.C. Murillo. 3D MiniNet: Learning a 2D representation from point clouds for fast and efficient 3D LiDAR semantic segmentation. In IROS, 2020. 2 +[2] Y.M. Asano, C. Rupprecht, and A. Vedaldi. Self-labelling via simultaneous clustering and representation learning. In ICLR, 2020. 3, 4 +[3] J. Behley, M. Garbade, A. Milioto, J. Quenzel, S. Behnke, J. Gall, and C. Stachniss. Towards 3D LiDAR-based semantic scene understanding of 3D point cloud sequences: The SemanticKITTI Dataset. In IJRR, 2021. 2, 5 +[4] J. Behley, M. Garbade, A. Milioto, J. Quenzel, S. Behnke, C. Stachniss, and J. Gall. SemanticKITTI: A dataset for semantic scene understanding of LiDAR sequences. In CVPR, 2019. 2, 5, 6 +[5] M. Caron, I. Misra, J. Mairal, P. Goyal, P. Bojanowski, and A. Joulin. Unsupervised learning of visual features by contrasting cluster assignments. In NeurIPS, 2020. 2, 4, 5, 6 +[6] J. Cen, P. Yun, S. Zhang, J. Cai, D. Luan, M. Tang, M. Liu, and M.Y. Wang. Open-world semantic segmentation for lidar point clouds. In ECCV, 2022. 3 +[7] L.C. Chen, G. Papandreou, I. Kokkinos, K. Murphy, and A.L. Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. In T-PAMI, 2017. 3 +[8] L.C. Chen, Y. Zhu, G. Papandreou, F. Schroff, and A. Hartwig. Encoder-decoder with atrous separable convolution for semantic image segmentation. In ECCV, 2018. 3 +[9] R. Cheng, R. Razani, E. Taghavi, E. Li, and B. Liu. (AF)2-S3Net: Attentive Feature Fusion with Adaptive Feature Selection for Sparse Semantic Segmentation Network. In CVPR, 2021. 2 +[10] C. Choy, JY. Gwak, and S. Savarese. 4D spatio-temporal convnets: Minkowski convolutional neural networks. In CVPR, 2019. 2, 3, 6 +[11] M. Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. In NeurIPS, 2013. 3, 4, 6 +[12] E. Fini, E. Sangineto, S. Lathuiliere, Z. Zhong, M. Nabi, and E. Ricci. A unified objective for novel class discovery. In CVPR, 2021. 1, 2, 4 +[13] A. Geiger, P. Lenz, and R. Urtasun. Are we ready for Autonomous Driving? The KITTI Vision Benchmark Suite. In CVPR, 2012. 2, 5 +[14] B. Graham, M. Engelcke, and L. van der Maaten. 3D semantic segmentation with submanifold sparse convolutional networks. In CVPR, 2018. 2 +[15] B. Graham and L. van der Maaten. Submanifold sparse convolutional networks. arXiv, 2017. 2 +[16] K. Han, A. Vedaldi, and A. Zisserman. Learning to discover novel visual categories via deep transfer clustering. In CVPR, 2019. 1, 2, 4 +[17] Q. Hu, B. Yang, L. Xie, S. Rosa, Y. Guo, Z. Wang, N. Trigoni, and A. Markham. Randla-net: Efficient semantic segmentation of large-scale point clouds. In CVPR, 2020. 2 + +[18] X. Ji, J.F. Henriques, and A. Vedaldi. Invariant information clustering for unsupervised image classification and segmentation. In ICCV, 2019. 4 +[19] X. Jia, K. Han, Y. Zhu, and B. Green. Joint representation learning and novel category discovery on single-and multimodal data. In ICCV, 2021. 2 +[20] K. Joseph, S. Paul, G. Aggarwal, S. Biswas, P. Rai, K. Han, and V.N. Balasubramanian. Novel class discovery without forgetting. ECCV, 2022. 2 +[21] G. Mei, F. Poiesi, C. Saltori, J. Zhang, E. Ricci, and N. Sebe. Overlap-guided gaussian mixture models for point cloud registration. In WACV, 2023. 4 +[22] G. Mei, C. Saltori, F. Poiesi, J. Zhang, E. Ricci, N. Sebe, and Q. Wu. Data augmentation-free unsupervised learning for 3d point cloud understanding. In BMVC, 2022. 3, 4 +[23] A. Milioto, I. Vizzo, J. Behley, and C. Stachniss. Rangenet++: Fast and accurate lidar semantic segmentation. In IROS, 2019. 2, 3 +[24] Y. Pan, B. Gao, J. Mei, S. Geng, C. Li, and H. Zhao. SemanticPOSS: A point cloud dataset with large quantity of dynamic instances. In IV, 2020. 2, 5, 6 +[25] C.R. Qi, H. Su, K. Mo, and L.J. Guibas. Pointnet: Deep learning on point sets for 3D classification and segmentation. In CVPR, 2017. 2 +[26] C.R. Qi, L. Yi, H. Su, and L.J. Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In NeurIPS, 2017. 2 +[27] O. Ronneberger, P. Fischer, and T. Brox. U-net: Convolutional networks for biomedical image segmentation. In MIC-CAI, 2015. 2 +[28] S. Roy, M. Liu, Z. Zhong, N. Sebe, and E. Ricci. Class incremental novel class discovery. In ECCV, 2022. 2 +[29] C. Saltori, F. Galasso, G. Fiameni, N. Sebe, E. Ricci, and F. Poiesi. Cosmix: Compositional semantic mix for domain adaptation in 3d lidar segmentation. In ECCV, 2022. 3 +[30] R. Song, W. Zhang, Y. Zhao, Y. Liu, and P.L. Rosin. Mesh saliency: An independent perceptual measure or a derivative of image saliency? In CVPR, 2021. 2 +[31] N. Souly, C. Spampinato, and M. Shah. Semi supervised semantic segmentation using generative adversarial network. In ICCV, 2017. 2 +[32] X. Tang, J. Wang, B. Gao, E. Dellandrea, R. Gaizauskas, and L. Chen. Large scale semi-supervised object detection using visual and semantic knowledge transfer. In CVPR, 2016. 2 +[33] H. Thomas, C.R. Qi, J.-E. Deschaud, B. Marcotegui, F. Goulette, and J. L. Guibas. Kpconv: Flexible and deformable convolution for point clouds. In ICCV, 2019. 2 +[34] S. Vaze, K. Han, A. Vedaldi, and A. Zisserman. Generalized category discovery. In CVPR, 2022. 2 +[35] B. Wu, A. Wan, X. Yue, and K. Keutzer. Squeezeseg: Convolutional neural nets with recurrent CRF for real-time road-object segmentation from 3D LiDAR point cloud. In ICRA, 2018. 2 +[36] B. Wu, X. Zhou, S. Zhao, X. Yue, and K. Keutzer. Squeeze-segv2: Improved model structure and unsupervised domain adaptation for road-object segmentation from a LiDAR point cloud. In ICRA, 2019. 2 + +[37] M. Yang, Y. Zhu, J. Yu, A. Wu, and C. Deng. Divide and conquer: Compositional experts for generalized novel class discovery. In CVPR, 2022. 2 +[38] L. Zhang and G.J. Qi. Wcp: Worst-case perturbations for semi-supervised deep learning. In CVPR, 2020. 2 +[39] Y. Zhang, Z. Zhou, P. David, X. Yue, Z. Xi, B. Gong, and H. Foroosh. Polarnet: An improved grid representation for online LiDAR point clouds semantic segmentation. In CVPR, 2020. 2 +[40] Y. Zhao, Z. Zhong, N. Sebe, and G.H. Lee. Novel class discovery in semantic segmentation. In CVPR, 2022. 1, 2, 3, 4, 5, 6, 7, 8 +[41] Z. Zhong, E. Fini, S. Roy, Z. Luo, E. Ricci, and N. Sebe. Neighborhood contrastive learning for novel class discovery. In CVPR, 2021. 1, 2, 4 +[42] Z. Zhong, L. Zhu, Z. Luo, S. Li, Y. Yang, and N. Sebe. Openmix: Reviving known knowledge for discovering novel visual categories in an open world. In CVPR, 2021. 2 +[43] Y. Zhou and O. Tuzel. Voxelnet: End-to-end learning for point cloud based 3D object detection. In CVPR, 2018. 2 +[44] X. Zhu, H. Zhou, T. Wang, F. Hong, Y. Ma, W. Li, H. Li, and D. Lin. Cylindrical and asymmetrical 3D convolution networks for lidar segmentation. In CVPR, 2021. 2 \ No newline at end of file diff --git a/2023/Novel Class Discovery for 3D Point Cloud Semantic Segmentation/images.zip b/2023/Novel Class Discovery for 3D Point Cloud Semantic Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..6d023089471e69880fd7200ecd2e5628a4d48c9f --- /dev/null +++ b/2023/Novel Class Discovery for 3D Point Cloud Semantic Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2306a681b4b4c478a0596f4ca28ee2fafbf3c3f24fe73d3aa9ab31553a7ad983 +size 725308 diff --git a/2023/Novel Class Discovery for 3D Point Cloud Semantic Segmentation/layout.json b/2023/Novel Class Discovery for 3D Point Cloud Semantic Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..470ff243f4cf18ad06ceca831debee5e412176e4 --- /dev/null +++ b/2023/Novel Class Discovery for 3D Point Cloud Semantic Segmentation/layout.json @@ -0,0 +1,9011 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 94, + 75, + 499, + 92 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 75, + 499, + 92 + ], + "spans": [ + { + "bbox": [ + 94, + 75, + 499, + 92 + ], + "type": "text", + "content": "Novel Class Discovery for 3D Point Cloud Semantic Segmentation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 141, + 114, + 449, + 129 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 114, + 449, + 129 + ], + "spans": [ + { + "bbox": [ + 141, + 114, + 449, + 129 + ], + "type": "text", + "content": "Luigi Riz" + }, + { + "bbox": [ + 141, + 114, + 449, + 129 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 141, + 114, + 449, + 129 + ], + "type": "text", + "content": " Cristiano Saltori" + }, + { + "bbox": [ + 141, + 114, + 449, + 129 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 141, + 114, + 449, + 129 + ], + "type": "text", + "content": " Elisa Ricci" + }, + { + "bbox": [ + 141, + 114, + 449, + 129 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 141, + 114, + 449, + 129 + ], + "type": "text", + "content": " Fabio Poiesi" + }, + { + "bbox": [ + 141, + 114, + 449, + 129 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 192, + 135, + 402, + 148 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 135, + 402, + 148 + ], + "spans": [ + { + "bbox": [ + 192, + 135, + 402, + 148 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 192, + 135, + 402, + 148 + ], + "type": "text", + "content": "Fondazione Bruno Kessler " + }, + { + "bbox": [ + 192, + 135, + 402, + 148 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 192, + 135, + 402, + 148 + ], + "type": "text", + "content": "University of Trento" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 48, + 175, + 545, + 262 + ], + "blocks": [ + { + "bbox": [ + 48, + 175, + 545, + 262 + ], + "lines": [ + { + "bbox": [ + 48, + 175, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 545, + 262 + ], + "type": "image", + "image_path": "db2cf5c1cac45253f849f7e282e42dd304853e5ba2874d4667d21166dc24fdba.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 270, + 546, + 293 + ], + "lines": [ + { + "bbox": [ + 46, + 270, + 546, + 293 + ], + "spans": [ + { + "bbox": [ + 46, + 270, + 546, + 293 + ], + "type": "text", + "content": "Figure 1. Novel Class Discovery for 3D point cloud semantic segmentation seeks to recognise novel classes by clustering unlabelled novel points with similar semantic features and by exploiting only the knowledge of a set of labelled samples corresponding to the base classes." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 314, + 191, + 327 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 314, + 191, + 327 + ], + "spans": [ + { + "bbox": [ + 143, + 314, + 191, + 327 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 45, + 348, + 290, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 348, + 290, + 601 + ], + "spans": [ + { + "bbox": [ + 45, + 348, + 290, + 601 + ], + "type": "text", + "content": "Novel class discovery (NCD) for semantic segmentation is the task of learning a model that can segment unlabelled (novel) classes using only the supervision from labelled (base) classes. This problem has recently been pioneered for 2D image data, but no work exists for 3D point cloud data. In fact, the assumptions made for 2D are loosely applicable to 3D in this case. This paper is presented to advance the state of the art on point cloud data analysis in four directions. Firstly, we address the new problem of NCD for point cloud semantic segmentation. Secondly, we show that the transposition of the only existing NCD method for 2D semantic segmentation to 3D data is suboptimal. Thirdly, we present a new method for NCD based on online clustering that exploits uncertainty quantification to produce prototypes for pseudo-labelling the points of the novel classes. Lastly, we introduce a new evaluation protocol to assess the performance of NCD for point cloud semantic segmentation. We thoroughly evaluate our method on SemanticKITTI and SemanticPOSS datasets, showing that it can significantly outperform the baseline. Project page: https://github.com/LuigiRiz/NOPS." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 655, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 655, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 655, + 287, + 715 + ], + "type": "text", + "content": "This project has received funding from the European Union's Horizon Europe research and innovation programme under grant agreement No 101058589. This work was also partially supported by the PRIN project LEGO-AI (Prot. 2020TA3K9N), the EU ISFP PROTECTOR (101034216) project and the EU H2020 MARVEL (957337) project and, it was carried out in the Vision and Learning joint laboratory of FBK and UNITN." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 307, + 302, + 386, + 315 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 302, + 386, + 315 + ], + "spans": [ + { + "bbox": [ + 307, + 302, + 386, + 315 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 327, + 546, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 327, + 546, + 495 + ], + "spans": [ + { + "bbox": [ + 304, + 327, + 546, + 495 + ], + "type": "text", + "content": "As humans, we are fairly skilled in organising new visual knowledge (novelty) into homogeneous groups, even when we do not know what we are observing. However, machines cannot perform this task satisfactorily without our supervision. The challenge here is mainly in the formulation of discriminative latent representations of the real world and in the quantification of the uncertainty when the novelty is observed [16,40,41]. The work of Han et al. [16] pioneered the formulation of the Novel Class Discovery (NCD) problem by defining it as the task that aims to classify the samples of an unlabelled dataset into different classes, i.e. the novel samples, by exploiting the knowledge of a set of labelled samples, i.e. the base samples. Note that the classes in the labelled and unlabelled datasets are disjoint." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 498, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 547, + 714 + ], + "type": "text", + "content": "NCD has been explored in the 2D image domain for classification [12, 16, 41] and later for semantic segmentation [40]. In particular, Zhao et al. [40] presented the first approach for NCD for 2D semantic segmentation. Two key assumptions were made by the authors. Firstly, at most one novel class is allowed in each image. Secondly, the new class belongs to a foreground object that can be found via saliency detection (e.g. a man on a bicycle, where the bicycle is the novel class). Thanks to these assumptions, the authors could pool the features of each image into a single latent representation and cluster the representations of the whole dataset to discover clusters of novel classes. We argue that these two are important constraints that are not applicable to generic 3D data, in particular to point clouds captured with LiDAR sensors in real-world city-scale scenarios. One point cloud can contain more than one novel class, and the saliency for 3D data cannot be leveraged in the same way as that for 2D data. Although they are both" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9393" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 133 + ], + "type": "text", + "content": "related to the attraction of human fixations, 3D saliency is more related to the regional importance of 3D surfaces rather than the foreground/background distinction [30]. Our motivation in exploring NCD for the 3D setting is mainly driven by addressing these shortcomings." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 133, + 289, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 133, + 289, + 504 + ], + "spans": [ + { + "bbox": [ + 46, + 133, + 289, + 504 + ], + "type": "text", + "content": "In this paper, we explore the new problem of NCD for 3D point cloud semantic segmentation (see Fig. 1). Given a partially human-annotated dataset, we jointly learn base and novel semantic classes by clustering unlabelled points with similar semantic features. We adapt the method of Zhao et al. [40] (Entropy-based Uncertainty Modeling and Self-training - EUMS) for point cloud data and use it as our baseline. We go beyond their formulation and, inspired by [5], we integrate batch-level (online) clustering in our method and update prototypes during training in order to make clustering computationally tractable. Cluster assignments are then used as training pseudo-labels. We also exploit over-clustering to achieve a higher clustering accuracy as in EUMS. Because point clouds contain multiple semantic classes, we cannot guarantee that all the classes appear in the point clouds within each batch, some will be missing. Therefore, we design a queuing strategy to store important features over training time, which will be used for pseudolabelling in the case of missing classes. We further introduce a strategy for exploiting the pseudo-label uncertainty to promote the creation of reliable prototypes that we then exploit to produce higher-quality pseudo-labels. Lastly, we produce two augmented views of the same point cloud and impose pseudo-label consistency amongst them. We evaluate our approach on SemanticKITTI [3, 4, 13] and SemanticPOSS [24], introducing an evaluation protocol for NCD and point cloud segmentation that can be adopted in future works. We empirically show that our approach largely outperforms our baseline in both datasets. We also perform an extensive ablation study to demonstrate the importance of the different components of our method." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 508, + 198, + 519 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 508, + 198, + 519 + ], + "spans": [ + { + "bbox": [ + 47, + 508, + 198, + 519 + ], + "type": "text", + "content": "To summarise, our contributions are:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 521, + 286, + 641 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 58, + 521, + 286, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 521, + 286, + 544 + ], + "spans": [ + { + "bbox": [ + 58, + 521, + 286, + 544 + ], + "type": "text", + "content": "- We address the new problem of NCD for 3D semantic segmentation;" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 545, + 286, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 545, + 286, + 581 + ], + "spans": [ + { + "bbox": [ + 58, + 545, + 286, + 581 + ], + "type": "text", + "content": "- We show that the transposition of the only existing NCD method for 2D semantic segmentation [40] to 3D data is suboptimal;" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 582, + 286, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 582, + 286, + 616 + ], + "spans": [ + { + "bbox": [ + 58, + 582, + 286, + 616 + ], + "type": "text", + "content": "- We present a new method for NCD based on online clustering and uncertainty estimation, which we name it NOPS (NOvel Point Segmentation);" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 58, + 617, + 286, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 617, + 286, + 641 + ], + "spans": [ + { + "bbox": [ + 58, + 617, + 286, + 641 + ], + "type": "text", + "content": "- We introduce a new evaluation protocol to assess the performance of NCD for 3D semantic segmentation." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 656, + 132, + 668 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 656, + 132, + 668 + ], + "spans": [ + { + "bbox": [ + 47, + 656, + 132, + 668 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "content": "Point cloud semantic segmentation can be performed at the point level [26], on range view maps [27], and by voxelising the input points [43]. Point-level networks pro" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 545, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 419 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 419 + ], + "type": "text", + "content": "cess the input without intermediate representations. Examples of these include PointNet [25], PointNet++ [26], RandLA-Net [17], and KPConv [33]. PointNet [25] and PointNet++ [26] are based on a series of multi-layer perceptron where PointNet++ introduces global and local feature aggregation at multiple scales. RandLA-Net [17] uses random sampling, attentive pooling, and local spatial encoding. KPConv [33] employs flexible and deformable convolutions in a continuous input space. Point-level networks are computationally inefficient when large-scale point clouds are processed. Range view architectures [23] and voxel-based approaches [10] are more computationally efficient than their point-level counterpart. The former requires projecting the input points on a 2D dense map, processing input maps with 2D convolutional filters [27], and reprojecting predictions to the initial 3D space. SqueezeSeg networks [35, 36], 3D-MiniNet [1], RangeNet++ [23], and PolarNet [39] are examples of this category. Although they are more efficient, these approaches tend to lose information during the projection and re-projection phase. The latter includes 3D quantisation-based approaches that discretise the input points into a 3D voxel grid and employ 3D convolutions [43] or 3D sparse convolutions [10, 15] to predict per-voxel classes. VoxelNet [43], SparseConv [14, 15], MinkowskiNet [10], Cylinder3D [44], and " + }, + { + "bbox": [ + 304, + 72, + 545, + 419 + ], + "type": "inline_equation", + "content": "(\\mathrm{AF})^2" + }, + { + "bbox": [ + 304, + 72, + 545, + 419 + ], + "type": "text", + "content": "-S3Net [9] are architectures belonging to this category. These approaches tackle point cloud segmentation in the supervised settings, whereas we tackle novel class discovery with labelled base classes and unlabelled novel classes." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 426, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 545, + 715 + ], + "type": "text", + "content": "Novel class discovery (NCD) is explored for 2D classification [12, 16, 19, 20, 28, 34, 37, 41, 42] and 2D segmentation [40]. NCD is more complex than standard semi-supervised learning [31, 32, 38]. In semi-supervised learning, labelled and unlabelled samples belong to the same classes, while in NCD, novel and base samples belong to disjoint classes. Han et al. [16] pioneered the NCD problem for 2D image classification. A classification model is pretrained on a set of base classes and used as feature extractor for the novel classes. They then train a classifier for the novel classes using the pseudo-labels produced by the pretrained model. Zhong et al. [41] introduced neighbourhood contrastive learning to generate discriminative representations for clustering. They retrieve and aggregate pseudopositive pairs with contrastive learning, encouraging the model to learn more discriminative representations. Hard negatives are obtained by mixing labelled and unlabelled samples in the feature space. UNO [12] unifies the two previous works by using a unique classification loss function for both base and novel classes, where pseudo-labels are processed together with ground-truth labels. NCD without Forgetting [20] and FRoST [28] further extend NCD to the incremental learning setting. EUMS [40] is the only approach analysing the NCD problem for semantic segmen" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9394" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 70, + 547, + 305 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 547, + 305 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 547, + 305 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 547, + 305 + ], + "type": "image", + "image_path": "1a61d98d4a2dd970aafcd5cb5669a50c29c66ce080f199adbb7a87f68abde09a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 308, + 546, + 344 + ], + "lines": [ + { + "bbox": [ + 46, + 308, + 546, + 344 + ], + "spans": [ + { + "bbox": [ + 46, + 308, + 546, + 344 + ], + "type": "text", + "content": "Figure 2. Overview of NOPS. We random augment the input point cloud twice and extract point-level features " + }, + { + "bbox": [ + 46, + 308, + 546, + 344 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 46, + 308, + 546, + 344 + ], + "type": "text", + "content": " with the shared model " + }, + { + "bbox": [ + 46, + 308, + 546, + 344 + ], + "type": "inline_equation", + "content": "f_{\\xi}" + }, + { + "bbox": [ + 46, + 308, + 546, + 344 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 308, + 546, + 344 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 46, + 308, + 546, + 344 + ], + "type": "text", + "content": " are used to obtain pseudo-labels in the online pseudo-labelling. We forward " + }, + { + "bbox": [ + 46, + 308, + 546, + 344 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 46, + 308, + 546, + 344 + ], + "type": "text", + "content": " to a novel " + }, + { + "bbox": [ + 46, + 308, + 546, + 344 + ], + "type": "inline_equation", + "content": "f_{n}" + }, + { + "bbox": [ + 46, + 308, + 546, + 344 + ], + "type": "text", + "content": " and a base " + }, + { + "bbox": [ + 46, + 308, + 546, + 344 + ], + "type": "inline_equation", + "content": "f_{b}" + }, + { + "bbox": [ + 46, + 308, + 546, + 344 + ], + "type": "text", + "content": " segmentation layer to output the novel and base predictions, respectively. We optimise our network by minimising a global objective function based on cross entropy." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 362, + 289, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 362, + 289, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 362, + 289, + 700 + ], + "type": "text", + "content": "tation. Unlike image classification, the model has to classify each pixel and handle multiple classes in each image. EUMS consists of a multi-stage pipeline using a saliency model to cluster the latent representations of novel classes to produce pseudo-labels. Moreover, entropy-based uncertainty and self-training are used to overcome noisy pseudolabels while improving the model performance on the novel classes. In this work, we tackle the problem of NCD in 3D point cloud semantic segmentation. Unlike previous works, our problem inherits the challenges from the fields of 2D semantic segmentation [7, 8] and 3D point cloud segmentation [10, 23, 29]. From 2D semantic segmentation, it inherits the additional challenges of multiple novel classes in the same image and the strong class unbalance. From 3D point cloud segmentation, we inherit the sparsity of input data, the different density of point cloud regions and the inability to identify foreground and background. The latter are not present in 2D segmentation [40]. Unlike [40] that use K-Means, we formulate clustering as an optimal transport problem to avoid degenerate solutions (i.e. all data points may be assigned to the same label and learn a constant representation) [2, 22]. Lastly, related to EUMS, REAL is proposed for open-world 3D semantic segmentation [6], where both known and unknown points have to be segmented. Unlike NOPS, all the unknown points belong to a single class and it is the task of a human annotator to separately label the novel classes. Then, these labels are used to update the base model by incrementally learning the novel classes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 306, + 361, + 420, + 375 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 361, + 420, + 375 + ], + "spans": [ + { + "bbox": [ + 306, + 361, + 420, + 375 + ], + "type": "text", + "content": "3. Proposed approach" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 306, + 380, + 373, + 393 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 380, + 373, + 393 + ], + "spans": [ + { + "bbox": [ + 306, + 380, + 373, + 393 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 400, + 545, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 400, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 400, + 545, + 581 + ], + "type": "text", + "content": "Given an input point cloud, we produce two augmented views that are processed with the same deep neural network to extract point-level features. These features are used to obtain pseudo-labels in the online pseudo-labelling step through the Sinkhorn-Knopp algorithm [11] (Sec. 3.3). Concurrently, we process the same features with the last network layers to segment novel and base classes. These features are stored in the class-balanced queue to mitigate the problem of batches with missing classes (Sec. 3.4). We exploit pseudo-label values (class probabilities) to filter out uncertain points, thus adding to the queue only high-quality points (Sec. 3.5). Lastly, we train our network by minimising the optimisation objective function through a swapped prediction task based on the computed pseudolabels (Sec. 3.6). Fig. 2 shows the block diagram of NOPS." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 586, + 427, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 586, + 427, + 598 + ], + "spans": [ + { + "bbox": [ + 306, + 586, + 427, + 598 + ], + "type": "text", + "content": "3.2. Problem formulation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{X} = \\{\\mathcal{X}\\}" + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": " be a dataset of 3D point clouds captured in different scenes. " + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": " is a set composed of a base set " + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_b" + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": " and a novel set " + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_n" + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": ", s.t. " + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{X} = \\mathcal{X}_b \\cup \\mathcal{X}_n" + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": ". The semantic categories that can be present in our point clouds are " + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{C} = \\mathcal{C}_b \\cup \\mathcal{C}_n" + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_b" + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": " is the set of base classes and " + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_n" + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": " is the set of novel classes, s.t. " + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_b \\cap \\mathcal{C}_n = \\emptyset" + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": ". Each " + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{X} \\in \\mathrm{X}" + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": " is composed of a finite but unknown number of 3D points " + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{X} = \\{(\\mathbf{x}, c)\\}" + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": " is the coordinate of the a point and " + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": " is its semantic class. We know the class of the point " + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "inline_equation", + "content": "(\\mathbf{x}, c)" + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": ", s.t. " + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathcal{X}_b" + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": " and" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 758 + ], + "type": "text", + "content": "9395" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 217 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 217 + ], + "type": "inline_equation", + "content": "c \\in \\mathcal{C}_b" + }, + { + "bbox": [ + 46, + 72, + 289, + 217 + ], + "type": "text", + "content": ", but we do not know the class of the point " + }, + { + "bbox": [ + 46, + 72, + 289, + 217 + ], + "type": "inline_equation", + "content": "(\\mathbf{x}, c)" + }, + { + "bbox": [ + 46, + 72, + 289, + 217 + ], + "type": "text", + "content": ", s.t. " + }, + { + "bbox": [ + 46, + 72, + 289, + 217 + ], + "type": "inline_equation", + "content": "x \\in \\mathcal{X}_n" + }, + { + "bbox": [ + 46, + 72, + 289, + 217 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 72, + 289, + 217 + ], + "type": "inline_equation", + "content": "c \\in \\mathcal{C}_n" + }, + { + "bbox": [ + 46, + 72, + 289, + 217 + ], + "type": "text", + "content": ". No points in " + }, + { + "bbox": [ + 46, + 72, + 289, + 217 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_n" + }, + { + "bbox": [ + 46, + 72, + 289, + 217 + ], + "type": "text", + "content": " belong to one of the base classes " + }, + { + "bbox": [ + 46, + 72, + 289, + 217 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_b" + }, + { + "bbox": [ + 46, + 72, + 289, + 217 + ], + "type": "text", + "content": ". As in [16, 40, 41], we assume that the number of classes to discover is known, i.e. " + }, + { + "bbox": [ + 46, + 72, + 289, + 217 + ], + "type": "inline_equation", + "content": "|\\mathcal{C}_n| = C_n" + }, + { + "bbox": [ + 46, + 72, + 289, + 217 + ], + "type": "text", + "content": ". We aim to design a computational approach that trains a deep neural network " + }, + { + "bbox": [ + 46, + 72, + 289, + 217 + ], + "type": "inline_equation", + "content": "f_\\Theta" + }, + { + "bbox": [ + 46, + 72, + 289, + 217 + ], + "type": "text", + "content": " that can segment all the points of a given point cloud, thus learning to jointly segment base classes " + }, + { + "bbox": [ + 46, + 72, + 289, + 217 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_b" + }, + { + "bbox": [ + 46, + 72, + 289, + 217 + ], + "type": "text", + "content": " and novel classes " + }, + { + "bbox": [ + 46, + 72, + 289, + 217 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_n" + }, + { + "bbox": [ + 46, + 72, + 289, + 217 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 72, + 289, + 217 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 46, + 72, + 289, + 217 + ], + "type": "text", + "content": " are the weights of our deep neural network. " + }, + { + "bbox": [ + 46, + 72, + 289, + 217 + ], + "type": "inline_equation", + "content": "f_\\Theta" + }, + { + "bbox": [ + 46, + 72, + 289, + 217 + ], + "type": "text", + "content": " is composed of two heads, " + }, + { + "bbox": [ + 46, + 72, + 289, + 217 + ], + "type": "inline_equation", + "content": "f_\\Theta = f_\\xi \\circ \\{f_b, f_n\\}" + }, + { + "bbox": [ + 46, + 72, + 289, + 217 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 72, + 289, + 217 + ], + "type": "inline_equation", + "content": "f_b" + }, + { + "bbox": [ + 46, + 72, + 289, + 217 + ], + "type": "text", + "content": " is the segmentation head for the base classes, " + }, + { + "bbox": [ + 46, + 72, + 289, + 217 + ], + "type": "inline_equation", + "content": "f_n" + }, + { + "bbox": [ + 46, + 72, + 289, + 217 + ], + "type": "text", + "content": " is the segmentation head for the novel classes, " + }, + { + "bbox": [ + 46, + 72, + 289, + 217 + ], + "type": "inline_equation", + "content": "f_\\xi" + }, + { + "bbox": [ + 46, + 72, + 289, + 217 + ], + "type": "text", + "content": " is the feature extractor network and " + }, + { + "bbox": [ + 46, + 72, + 289, + 217 + ], + "type": "inline_equation", + "content": "\\circ" + }, + { + "bbox": [ + 46, + 72, + 289, + 217 + ], + "type": "text", + "content": " is the composition operator (Fig. 2)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 225, + 181, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 225, + 181, + 239 + ], + "spans": [ + { + "bbox": [ + 47, + 225, + 181, + 239 + ], + "type": "text", + "content": "3.3. Online pseudo-labelling" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 244, + 289, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 244, + 289, + 509 + ], + "spans": [ + { + "bbox": [ + 46, + 244, + 289, + 509 + ], + "type": "text", + "content": "We formulate pseudo-labelling as the assignment of novel points to the class-prototypes learnt during training [5]. Let " + }, + { + "bbox": [ + 46, + 244, + 289, + 509 + ], + "type": "inline_equation", + "content": "\\mathsf{P} \\in \\mathbb{R}^{D \\times \\rho}" + }, + { + "bbox": [ + 46, + 244, + 289, + 509 + ], + "type": "text", + "content": " be the class prototypes, where " + }, + { + "bbox": [ + 46, + 244, + 289, + 509 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 244, + 289, + 509 + ], + "type": "text", + "content": " is the size of the output features from " + }, + { + "bbox": [ + 46, + 244, + 289, + 509 + ], + "type": "inline_equation", + "content": "f_{\\xi}" + }, + { + "bbox": [ + 46, + 244, + 289, + 509 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 244, + 289, + 509 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 46, + 244, + 289, + 509 + ], + "type": "text", + "content": " is the number of prototypes. Let " + }, + { + "bbox": [ + 46, + 244, + 289, + 509 + ], + "type": "inline_equation", + "content": "\\mathsf{Z} \\in \\mathbb{R}^{D \\times m}" + }, + { + "bbox": [ + 46, + 244, + 289, + 509 + ], + "type": "text", + "content": " be the normalised output features extracted from " + }, + { + "bbox": [ + 46, + 244, + 289, + 509 + ], + "type": "inline_equation", + "content": "f_{\\xi}" + }, + { + "bbox": [ + 46, + 244, + 289, + 509 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 244, + 289, + 509 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 244, + 289, + 509 + ], + "type": "text", + "content": " is the number of points of the point cloud. " + }, + { + "bbox": [ + 46, + 244, + 289, + 509 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 244, + 289, + 509 + ], + "type": "text", + "content": " is not known a priori and it can differ across point clouds. We aim to find the assignment " + }, + { + "bbox": [ + 46, + 244, + 289, + 509 + ], + "type": "inline_equation", + "content": "\\mathsf{Q} \\in \\mathbb{R}^{\\rho \\times m}" + }, + { + "bbox": [ + 46, + 244, + 289, + 509 + ], + "type": "text", + "content": " s.t. all the points in the batch are equally partitioned across the " + }, + { + "bbox": [ + 46, + 244, + 289, + 509 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 46, + 244, + 289, + 509 + ], + "type": "text", + "content": " prototypes. This equipartition ensures that the feature representations of the points belonging to different novel classes are well-separated, thus preventing the case in which the novel class feature representations collapse into a unique solution. Caron et al. [5] employs an arbitrary large number of prototypes " + }, + { + "bbox": [ + 46, + 244, + 289, + 509 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 46, + 244, + 289, + 509 + ], + "type": "text", + "content": " to effectively organise the feature space produced by " + }, + { + "bbox": [ + 46, + 244, + 289, + 509 + ], + "type": "inline_equation", + "content": "f_{\\xi}" + }, + { + "bbox": [ + 46, + 244, + 289, + 509 + ], + "type": "text", + "content": ". They discard " + }, + { + "bbox": [ + 46, + 244, + 289, + 509 + ], + "type": "inline_equation", + "content": "\\mathsf{P}" + }, + { + "bbox": [ + 46, + 244, + 289, + 509 + ], + "type": "text", + "content": " after training. In contrast, we learn exactly " + }, + { + "bbox": [ + 46, + 244, + 289, + 509 + ], + "type": "inline_equation", + "content": "\\rho = C_n" + }, + { + "bbox": [ + 46, + 244, + 289, + 509 + ], + "type": "text", + "content": " class prototypes and propose to use " + }, + { + "bbox": [ + 46, + 244, + 289, + 509 + ], + "type": "inline_equation", + "content": "\\mathsf{P}" + }, + { + "bbox": [ + 46, + 244, + 289, + 509 + ], + "type": "text", + "content": " as the weights for our new class segmentation head " + }, + { + "bbox": [ + 46, + 244, + 289, + 509 + ], + "type": "inline_equation", + "content": "f_{n}" + }, + { + "bbox": [ + 46, + 244, + 289, + 509 + ], + "type": "text", + "content": ", which outputs the " + }, + { + "bbox": [ + 46, + 244, + 289, + 509 + ], + "type": "inline_equation", + "content": "C_n" + }, + { + "bbox": [ + 46, + 244, + 289, + 509 + ], + "type": "text", + "content": " logits for the new classes. In order to optimise the assignment " + }, + { + "bbox": [ + 46, + 244, + 289, + 509 + ], + "type": "inline_equation", + "content": "\\mathsf{Q}" + }, + { + "bbox": [ + 46, + 244, + 289, + 509 + ], + "type": "text", + "content": ", we maximise the similarity between the features of the new points and the learned prototypes as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 96, + 517, + 287, + 538 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 517, + 287, + 538 + ], + "spans": [ + { + "bbox": [ + 96, + 517, + 287, + 538 + ], + "type": "interline_equation", + "content": "\\max _ {\\mathbb {Q} \\in \\mathcal {Q}} \\operatorname {T r} \\left(\\mathbb {Q} ^ {\\top} \\mathbb {P} ^ {\\top} \\mathbb {Z}\\right) + \\epsilon H (\\mathbb {Q}) \\rightarrow \\mathbb {Q} ^ {*}, \\tag {1}", + "image_path": "0bc31cbb6cc63d1e845c45f223b81b742798fcaeaa74a133b4561d85a3e5916f.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 547, + 287, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 547, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 46, + 547, + 287, + 668 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 547, + 287, + 668 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 46, + 547, + 287, + 668 + ], + "type": "text", + "content": " is the entropy function, " + }, + { + "bbox": [ + 46, + 547, + 287, + 668 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 46, + 547, + 287, + 668 + ], + "type": "text", + "content": " is the parameter that determines the smoothness of the assignment and " + }, + { + "bbox": [ + 46, + 547, + 287, + 668 + ], + "type": "inline_equation", + "content": "\\mathbb{Q}^*" + }, + { + "bbox": [ + 46, + 547, + 287, + 668 + ], + "type": "text", + "content": " is our sought solution. Asano et al. [2] enforce the equipartitioning constraint by requiring " + }, + { + "bbox": [ + 46, + 547, + 287, + 668 + ], + "type": "inline_equation", + "content": "\\mathbb{Q}" + }, + { + "bbox": [ + 46, + 547, + 287, + 668 + ], + "type": "text", + "content": " to belong to a transportation polytope and perform this optimisation on the whole dataset at once (offline). This operation with point cloud data is computationally impractical. Therefore, we formulate the transportation polytope such that the optimisation is performed online, which consists of processing only the points within the batch being processed" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 52, + 677, + 287, + 713 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 52, + 677, + 287, + 713 + ], + "type": "interline_equation", + "content": "\\mathcal {Q} = \\left\\{\\mathbb {Q} \\in \\mathbb {R} _ {+} ^ {C _ {n} \\times m} | \\mathbb {Q} 1 _ {m} = \\frac {1}{C _ {n}} 1 _ {C _ {n}}, \\mathbb {Q} ^ {\\top} 1 _ {C _ {n}} = \\frac {1}{m} 1 _ {m} \\right\\}, \\tag {2}", + "image_path": "1b0f357628601b819dd3a11a0ff28da44ad19993806b0db03969070c8904f9e3.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "1_{\\star}" + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": " represents a vector of ones of dimension " + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "\\star" + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": ". These constraints ensure that each class prototype is selected on average at least " + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "m / C_n" + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": " times in each batch. The solution " + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "\\mathbb{Q}^*" + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": " can take the form of a normalised exponential matrix" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 352, + 127, + 545, + 155 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 352, + 127, + 545, + 155 + ], + "spans": [ + { + "bbox": [ + 352, + 127, + 545, + 155 + ], + "type": "interline_equation", + "content": "\\mathbb {Q} ^ {*} = \\operatorname {d i a g} (\\alpha) \\exp \\left(\\frac {\\mathrm {P} ^ {\\top} \\mathrm {Z}}{\\epsilon}\\right) \\operatorname {d i a g} (\\beta), \\tag {3}", + "image_path": "4be12bc29bd454873e37dcc7502b704ee89bb31ad577543d41135b47cd8bb728.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 162, + 545, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 162, + 545, + 221 + ], + "spans": [ + { + "bbox": [ + 304, + 162, + 545, + 221 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 162, + 545, + 221 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 162, + 545, + 221 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 162, + 545, + 221 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 162, + 545, + 221 + ], + "type": "text", + "content": " are renormalization vectors that are computed iteratively with the Sinkhorn-Knopp algorithm [11, 21]. We then transpose the optimised soft assignment " + }, + { + "bbox": [ + 304, + 162, + 545, + 221 + ], + "type": "inline_equation", + "content": "\\mathbb{Q}^* \\in \\mathbb{R}_+^{C_n \\times m}" + }, + { + "bbox": [ + 304, + 162, + 545, + 221 + ], + "type": "text", + "content": " to obtain the soft pseudo-labels for each of the " + }, + { + "bbox": [ + 304, + 162, + 545, + 221 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 162, + 545, + 221 + ], + "type": "text", + "content": " novel points being processed within each batch." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 221, + 545, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 221, + 545, + 269 + ], + "spans": [ + { + "bbox": [ + 304, + 221, + 545, + 269 + ], + "type": "text", + "content": "We empirically found that training can be more effective if pseudo-labels are smoother in the first training epochs and peaked in the last training epochs. Therefore, we introduce a linear decay of " + }, + { + "bbox": [ + 304, + 221, + 545, + 269 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 304, + 221, + 545, + 269 + ], + "type": "text", + "content": " during training." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 269, + 546, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 269, + 546, + 412 + ], + "spans": [ + { + "bbox": [ + 304, + 269, + 546, + 412 + ], + "type": "text", + "content": "Multi-headed segmentation: A single segmentation head may converge to a suboptimal feature space, thus producing suboptimal prototype solutions. To further improve the segmentation quality, we use multiple novel class segmentation heads to optimise " + }, + { + "bbox": [ + 304, + 269, + 546, + 412 + ], + "type": "inline_equation", + "content": "f_{\\Theta}" + }, + { + "bbox": [ + 304, + 269, + 546, + 412 + ], + "type": "text", + "content": " based on different training solutions. Different solutions increase the likelihood of producing a diverse partitioning of the feature space as they regularise with each other (they share the same backbone) [18]. In practise, we concatenate the logits of the base class segmentation head with the outputs of each novel class segmentation head and we separately evaluate their loss for each novel class segmentation head at training time." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 413, + 545, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 413, + 545, + 508 + ], + "spans": [ + { + "bbox": [ + 304, + 413, + 545, + 508 + ], + "type": "text", + "content": "We task our network to over-cluster novel points, using segmentation heads that output " + }, + { + "bbox": [ + 304, + 413, + 545, + 508 + ], + "type": "inline_equation", + "content": "o \\cdot C_n" + }, + { + "bbox": [ + 304, + 413, + 545, + 508 + ], + "type": "text", + "content": " logits, where " + }, + { + "bbox": [ + 304, + 413, + 545, + 508 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 304, + 413, + 545, + 508 + ], + "type": "text", + "content": " is the over-clustering factor. Previous studies empirically showed that this is beneficial to learn more informative features [5, 12, 18, 22]. We observed the same and concur that over-clustering can be useful for increasing expressivity of the feature representations. The over-clustering heads are then discarded at inference time." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 515, + 440, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 515, + 440, + 529 + ], + "spans": [ + { + "bbox": [ + 306, + 515, + 440, + 529 + ], + "type": "text", + "content": "3.4. Class-balanced queuing" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": "Soft pseudo-labelling described in Sec. 3.3 produces an equipartite matching between the novel points and the class centroids. However, it is highly likely that batches are sampled with point clouds containing novel classes with different cardinalities when dealing with 3D data. It is also likely that some scenes may contain only a subset of the novel classes. Therefore, enforcing the equipartitioning constraint in each batch of the dataset could affect the learning of less-frequent (long-tail) classes. As a solution, we introduce a queue " + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathsf{Z}_q" + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": " containing a randomly extracted portion of the features of the novel points from the previous iterations. We use these additional data to mitigate the potential class imbalance that may occur during training. In practise, we compute " + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathsf{Z} \\gets \\mathsf{Z} \\oplus \\mathsf{Z}_q" + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\oplus" + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": " is the concatenation operator, and execute the Sinkhorn-Knopp algorithm on this" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9396" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 96 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 96 + ], + "type": "text", + "content": "augmented version of " + }, + { + "bbox": [ + 46, + 72, + 287, + 96 + ], + "type": "inline_equation", + "content": "\\mathbb{Z}" + }, + { + "bbox": [ + 46, + 72, + 287, + 96 + ], + "type": "text", + "content": ". Then, we retain only the pseudolabels for the first " + }, + { + "bbox": [ + 46, + 72, + 287, + 96 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 72, + 287, + 96 + ], + "type": "text", + "content": " columns of " + }, + { + "bbox": [ + 46, + 72, + 287, + 96 + ], + "type": "inline_equation", + "content": "\\mathbb{Q}^*" + }, + { + "bbox": [ + 46, + 72, + 287, + 96 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 107, + 260, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 107, + 260, + 121 + ], + "spans": [ + { + "bbox": [ + 47, + 107, + 260, + 121 + ], + "type": "text", + "content": "3.5. Uncertainty-aware training and queuing" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 127, + 287, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 127, + 287, + 235 + ], + "spans": [ + { + "bbox": [ + 46, + 127, + 287, + 235 + ], + "type": "text", + "content": "We propose to carefully select novel points for training " + }, + { + "bbox": [ + 46, + 127, + 287, + 235 + ], + "type": "inline_equation", + "content": "f_{\\Theta}" + }, + { + "bbox": [ + 46, + 127, + 287, + 235 + ], + "type": "text", + "content": " with fewer but more reliable pseudo-labels and to build a more effective queue " + }, + { + "bbox": [ + 46, + 127, + 287, + 235 + ], + "type": "inline_equation", + "content": "Z_{q}" + }, + { + "bbox": [ + 46, + 127, + 287, + 235 + ], + "type": "text", + "content": ". We perform this selection by applying a threshold to the class probabilities of the novel class pseudo-labels. We found that seeking a fixed threshold for all the novel classes, that is also compatible with the variations of the class probabilities during training, is impractical. Therefore, we employ an adaptive threshold based on the class probabilities within each batch." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 236, + 288, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 236, + 288, + 331 + ], + "spans": [ + { + "bbox": [ + 46, + 236, + 288, + 331 + ], + "type": "text", + "content": "Our selection strategy operates as follows. Let " + }, + { + "bbox": [ + 46, + 236, + 288, + 331 + ], + "type": "inline_equation", + "content": "\\tau_{c}" + }, + { + "bbox": [ + 46, + 236, + 288, + 331 + ], + "type": "text", + "content": " be the adaptive threshold for the points of the novel class " + }, + { + "bbox": [ + 46, + 236, + 288, + 331 + ], + "type": "inline_equation", + "content": "c\\in \\mathcal{C}_n" + }, + { + "bbox": [ + 46, + 236, + 288, + 331 + ], + "type": "text", + "content": ". Firstly, we extract the novel points that have the greatest class probability for the class " + }, + { + "bbox": [ + 46, + 236, + 288, + 331 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 236, + 288, + 331 + ], + "type": "text", + "content": ". Secondly, we compute " + }, + { + "bbox": [ + 46, + 236, + 288, + 331 + ], + "type": "inline_equation", + "content": "\\tau_{c}" + }, + { + "bbox": [ + 46, + 236, + 288, + 331 + ], + "type": "text", + "content": " as the " + }, + { + "bbox": [ + 46, + 236, + 288, + 331 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 236, + 288, + 331 + ], + "type": "text", + "content": "-th percentile of the class probabilities of these novel points. Lastly, we retain the novel points of class " + }, + { + "bbox": [ + 46, + 236, + 288, + 331 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 236, + 288, + 331 + ], + "type": "text", + "content": " whose class probability is above the threshold " + }, + { + "bbox": [ + 46, + 236, + 288, + 331 + ], + "type": "inline_equation", + "content": "\\tau_{c}" + }, + { + "bbox": [ + 46, + 236, + 288, + 331 + ], + "type": "text", + "content": ". We define this selection strategy as the function" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 111, + 342, + 287, + 357 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 342, + 287, + 357 + ], + "spans": [ + { + "bbox": [ + 111, + 342, + 287, + 357 + ], + "type": "interline_equation", + "content": "\\phi : \\left(\\mathcal {F} _ {n}, \\hat {\\mathcal {Y}} _ {n}\\right) \\times p \\mapsto \\left(\\bar {\\mathcal {F}} _ {n}\\right), \\tag {4}", + "image_path": "6eee8198462f7c727efd24c85af4724a12488a6658fbf8361cb5c07314257203.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 368, + 287, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 368, + 287, + 431 + ], + "spans": [ + { + "bbox": [ + 46, + 368, + 287, + 431 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 368, + 287, + 431 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_n" + }, + { + "bbox": [ + 46, + 368, + 287, + 431 + ], + "type": "text", + "content": " is the set of feature vectors extracted from " + }, + { + "bbox": [ + 46, + 368, + 287, + 431 + ], + "type": "inline_equation", + "content": "f_{\\xi}" + }, + { + "bbox": [ + 46, + 368, + 287, + 431 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 368, + 287, + 431 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{V}}_n" + }, + { + "bbox": [ + 46, + 368, + 287, + 431 + ], + "type": "text", + "content": " is the set of class probabilities predicted by the network for these points. " + }, + { + "bbox": [ + 46, + 368, + 287, + 431 + ], + "type": "inline_equation", + "content": "\\bar{\\mathcal{F}}_n" + }, + { + "bbox": [ + 46, + 368, + 287, + 431 + ], + "type": "text", + "content": " are both processed by the Sinkhorn-Knopp algorithm to generate our pseudo-labels and added to " + }, + { + "bbox": [ + 46, + 368, + 287, + 431 + ], + "type": "inline_equation", + "content": "\\mathbb{Z}_q" + }, + { + "bbox": [ + 46, + 368, + 287, + 431 + ], + "type": "text", + "content": " to make it more effective." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 441, + 176, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 441, + 176, + 453 + ], + "spans": [ + { + "bbox": [ + 47, + 441, + 176, + 453 + ], + "type": "text", + "content": "3.6. Optimisation objective" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 460, + 287, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 460, + 287, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 287, + 593 + ], + "type": "text", + "content": "We optimise " + }, + { + "bbox": [ + 46, + 460, + 287, + 593 + ], + "type": "inline_equation", + "content": "f_{\\Theta}" + }, + { + "bbox": [ + 46, + 460, + 287, + 593 + ], + "type": "text", + "content": " by using the weighted Cross Entropy objective based on the labels " + }, + { + "bbox": [ + 46, + 460, + 287, + 593 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}_b" + }, + { + "bbox": [ + 46, + 460, + 287, + 593 + ], + "type": "text", + "content": " of the base samples and the pseudo-labels " + }, + { + "bbox": [ + 46, + 460, + 287, + 593 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{Y}}_n" + }, + { + "bbox": [ + 46, + 460, + 287, + 593 + ], + "type": "text", + "content": " of the novel samples. We formulate a swapped prediction task based on these pseudo-labels [5]. Secicfically, we begin by generating two different augmentations of " + }, + { + "bbox": [ + 46, + 460, + 287, + 593 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 46, + 460, + 287, + 593 + ], + "type": "text", + "content": " that we define as " + }, + { + "bbox": [ + 46, + 460, + 287, + 593 + ], + "type": "inline_equation", + "content": "\\mathcal{X}'" + }, + { + "bbox": [ + 46, + 460, + 287, + 593 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 460, + 287, + 593 + ], + "type": "inline_equation", + "content": "\\mathcal{X}''" + }, + { + "bbox": [ + 46, + 460, + 287, + 593 + ], + "type": "text", + "content": ". We use the known one-hot labels for " + }, + { + "bbox": [ + 46, + 460, + 287, + 593 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}_b" + }, + { + "bbox": [ + 46, + 460, + 287, + 593 + ], + "type": "text", + "content": " and the predicted soft pseudolabels for " + }, + { + "bbox": [ + 46, + 460, + 287, + 593 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{Y}}_n" + }, + { + "bbox": [ + 46, + 460, + 287, + 593 + ], + "type": "text", + "content": ". We predict the novel pseudo-labels " + }, + { + "bbox": [ + 46, + 460, + 287, + 593 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{Y}}_n'" + }, + { + "bbox": [ + 46, + 460, + 287, + 593 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 460, + 287, + 593 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{Y}}_n''" + }, + { + "bbox": [ + 46, + 460, + 287, + 593 + ], + "type": "text", + "content": " of the respective point clouds " + }, + { + "bbox": [ + 46, + 460, + 287, + 593 + ], + "type": "inline_equation", + "content": "\\mathcal{X}'" + }, + { + "bbox": [ + 46, + 460, + 287, + 593 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 460, + 287, + 593 + ], + "type": "inline_equation", + "content": "\\mathcal{X}''" + }, + { + "bbox": [ + 46, + 460, + 287, + 593 + ], + "type": "text", + "content": " with our approach. Then, we enforce prediction consistency between the swapped pseudo-labels of the two augmentations as" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 100, + 602, + 287, + 617 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 602, + 287, + 617 + ], + "spans": [ + { + "bbox": [ + 100, + 602, + 287, + 617 + ], + "type": "interline_equation", + "content": "\\mathcal {L} (\\mathcal {X}) = \\ell \\left(\\hat {\\mathcal {Y}} ^ {\\prime}, \\tilde {\\mathcal {Y}} ^ {\\prime \\prime}\\right) + \\ell \\left(\\hat {\\mathcal {Y}} ^ {\\prime \\prime}, \\tilde {\\mathcal {Y}} ^ {\\prime}\\right), \\tag {5}", + "image_path": "314a3bda7d4e1246ae4aeebdc9c3d2418bf74caaf9e79675e69a67979041c6ab.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 627, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 627, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 627, + 287, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 627, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{Y}}^{\\prime} = \\hat{\\mathcal{Y}}_{b}^{\\prime}\\cup \\hat{\\mathcal{Y}}_{n}^{\\prime}" + }, + { + "bbox": [ + 46, + 627, + 287, + 715 + ], + "type": "text", + "content": " (same for " + }, + { + "bbox": [ + 46, + 627, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{Y}}^{\\prime \\prime}" + }, + { + "bbox": [ + 46, + 627, + 287, + 715 + ], + "type": "text", + "content": "), " + }, + { + "bbox": [ + 46, + 627, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{Y}}^{\\prime} = \\mathcal{Y}_{b}^{\\prime}\\cup \\tilde{\\mathcal{Y}}_{n}^{\\prime}" + }, + { + "bbox": [ + 46, + 627, + 287, + 715 + ], + "type": "text", + "content": " (same for " + }, + { + "bbox": [ + 46, + 627, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{Y}}^{\\prime \\prime}" + }, + { + "bbox": [ + 46, + 627, + 287, + 715 + ], + "type": "text", + "content": ") and " + }, + { + "bbox": [ + 46, + 627, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 46, + 627, + 287, + 715 + ], + "type": "text", + "content": " is the weighted Cross Entropy loss. We use separate segmentation heads for base classes and novel classes. The weights of the loss for the base classes are computed based on their occurrence frequency in the training set. The weights of the loss for the novel classes are all set equally as their occurrence frequency in the dataset is unknown." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 71, + 503, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 71, + 503, + 85 + ], + "spans": [ + { + "bbox": [ + 305, + 71, + 503, + 85 + ], + "type": "text", + "content": "4. Adapting NCD for 2D images to 3D" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 91, + 545, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 545, + 365 + ], + "type": "text", + "content": "Another contribution of this work is to adapt the method proposed by Zhao et al. [40] for NCD for 2D semantic segmentation (EUMS) to 3D data. Our empirical evaluation (see Sec. 5) shows that the transposition of EUMS to the 3D domain has some limitations. In particular, as described in Sec. 1, EUMS uses two assumptions: I) the novel classes belong to the foreground and II) each image can contain at most one novel class. This allows EUMS to leverage a saliency detection model to produce a foreground mask and a segmentation model pre-trained on the base classes to determine which portion of the image is background. The portion of the image that belongs to both the foreground mask and the background mask is where features are then pooled. EUMS computes a feature representation for each image by average pooling the features of the pixels belonging to the unknown portion. The feature representations of all the images in the dataset are clustered with K-Means by using the number of classes to discover as the target number of clusters. EUMS shows that overclustering and entropy-based modelling can be exploited to improve the results. The affiliation of a point to its cluster is used to produce hard pseudo-labels that are in turn used along with the ground-truth labels to fine-tune the pre-trained model." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 366, + 546, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 366, + 546, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 366, + 546, + 629 + ], + "type": "text", + "content": "With 3D point clouds, there is no concept of foreground and background (in contrast with I). Our adaptation is designed to discover the classes of all the unlabelled points (in contrast with II). Therefore, given the unlabelled points of each point cloud, we randomly extract a subset of these by setting a ratio (e.g. " + }, + { + "bbox": [ + 304, + 366, + 546, + 629 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 304, + 366, + 546, + 629 + ], + "type": "text", + "content": ") with upper bound (e.g. 1K) on the number of points to select. We compute and collect their features for all the point clouds in the dataset and apply K-Means on the whole set of features. Note that this clustering step is computationally expensive, and we had to use High Performance Computing to execute it. The subsampling of the points was necessary to fit the data in the RAM (see Sec. 5 for a detailed analysis). Once the cluster prototypes are computed, we produce the hard pseudo-labels. To enrich the set of pseudo-labels, we propagate the pseudo-label of each point to its nearest neighbour in the coordinate space. This allows us to expand the subset of pseudo-labelled randomly selected points. We also implement the other steps of overclustering and entropy-based modelling to boost the results. Lastly, we fine-tune our model with these pseudo-labels. We name our transposition of EUMS as " + }, + { + "bbox": [ + 304, + 366, + 546, + 629 + ], + "type": "inline_equation", + "content": "\\mathrm{EUMS}^{\\dagger}" + }, + { + "bbox": [ + 304, + 366, + 546, + 629 + ], + "type": "text", + "content": " and report its block diagram in the Supplementary Material." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 639, + 428, + 653 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 639, + 428, + 653 + ], + "spans": [ + { + "bbox": [ + 305, + 639, + 428, + 653 + ], + "type": "text", + "content": "5. Experimental results" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 659, + 421, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 659, + 421, + 672 + ], + "spans": [ + { + "bbox": [ + 305, + 659, + 421, + 672 + ], + "type": "text", + "content": "5.1. Experimental setup" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 714 + ], + "type": "text", + "content": "Datasets. We evaluate our approach on SemanticKITTI [3, 4,13] and SemanticPOSS [24]. SemanticKITTI [4] consists of 43,552 point cloud acquisitions with point-level annota" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9397" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 96, + 286, + 162 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 287, + 93 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 287, + 93 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 287, + 93 + ], + "type": "text", + "content": "Table 1. SemanticKITTI splits, is defined as KITTI-" + }, + { + "bbox": [ + 47, + 70, + 287, + 93 + ], + "type": "inline_equation", + "content": "n^i" + }, + { + "bbox": [ + 47, + 70, + 287, + 93 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 70, + 287, + 93 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 47, + 70, + 287, + 93 + ], + "type": "text", + "content": " is the number of novel classes and " + }, + { + "bbox": [ + 47, + 70, + 287, + 93 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 70, + 287, + 93 + ], + "type": "text", + "content": " is the split index." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 96, + 286, + 162 + ], + "lines": [ + { + "bbox": [ + 49, + 96, + 286, + 162 + ], + "spans": [ + { + "bbox": [ + 49, + 96, + 286, + 162 + ], + "type": "table", + "html": "
SplitNovel Classes
KITTI-50building, road, sidewalk, terrain, vegetation
KITTI-51car, fence, other-ground, parking, trunk
KITTI-52motorcycle, other-vehicle, pole, traffic-sign, truck
KITTI-43bicycle, bicyclist, motorcyclist, person
", + "image_path": "6d6d8acc1efd2dc7e11f5e07b74e71c89e2f321b3aa2d0f6f5e775e803adc302.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 86, + 200, + 249, + 268 + ], + "blocks": [ + { + "bbox": [ + 47, + 175, + 286, + 197 + ], + "lines": [ + { + "bbox": [ + 47, + 175, + 286, + 197 + ], + "spans": [ + { + "bbox": [ + 47, + 175, + 286, + 197 + ], + "type": "text", + "content": "Table 2. SemanticPOSS splits, defined as " + }, + { + "bbox": [ + 47, + 175, + 286, + 197 + ], + "type": "inline_equation", + "content": "\\mathrm{POSS} - n^i" + }, + { + "bbox": [ + 47, + 175, + 286, + 197 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 175, + 286, + 197 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 47, + 175, + 286, + 197 + ], + "type": "text", + "content": " is the number of novel classes and " + }, + { + "bbox": [ + 47, + 175, + 286, + 197 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 175, + 286, + 197 + ], + "type": "text", + "content": " is the split index." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 86, + 200, + 249, + 268 + ], + "lines": [ + { + "bbox": [ + 86, + 200, + 249, + 268 + ], + "spans": [ + { + "bbox": [ + 86, + 200, + 249, + 268 + ], + "type": "table", + "html": "
SplitNovel Classes
POSS-40building, car, ground, plants
POSS-31bike, fence, person
POSS-32pole, traffic-sign, trunk
POSS-33cone-stone, rider, trashcan
", + "image_path": "c247e070242704965f06018b3b40fdd8cce5a9b9504ab6a6ef21f197d77de149.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 290, + 286, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 290, + 286, + 375 + ], + "spans": [ + { + "bbox": [ + 46, + 290, + 286, + 375 + ], + "type": "text", + "content": "tions of 19 semantic classes. Based on the official benchmark guidelines [4], we use sequence 08 for validation and the other sequences for training. SemanticPOSS [24] consists of 2,988 real-world point cloud acquisitions with point-level annotations of 13 semantic classes. Based on the official benchmark guidelines [24], we use sequence 03 for validation and the other sequences for training." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 376, + 286, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 376, + 286, + 543 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 286, + 543 + ], + "type": "text", + "content": "Experimental protocol for 3D NCD. Similarly to what proposed by [40] in the 2D domain, we create different splits of each dataset to validate the NCD performance. We create four splits for SemanticKITTI and SemanticPOSS. We refer to these splits as SemanticKITTI-" + }, + { + "bbox": [ + 46, + 376, + 286, + 543 + ], + "type": "inline_equation", + "content": "n^i" + }, + { + "bbox": [ + 46, + 376, + 286, + 543 + ], + "type": "text", + "content": " and SemanticPOSS-" + }, + { + "bbox": [ + 46, + 376, + 286, + 543 + ], + "type": "inline_equation", + "content": "n^i" + }, + { + "bbox": [ + 46, + 376, + 286, + 543 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 376, + 286, + 543 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 376, + 286, + 543 + ], + "type": "text", + "content": " indexes the split. In each set, the novel classes and the base classes correspond to unlabelled and labelled points, respectively. Tabs. 1 & 2 detail the splits of our datasets. These splits are selected based on their class distribution in the dataset and on the semantic relationship between novel and base classes, e.g. in KITTI-43 the base class motorcycle can be helpful to discover the novel class motorcyclist. We report additional details about the selection process in the Supplementary Material." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 544, + 286, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 544, + 286, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 544, + 286, + 616 + ], + "type": "text", + "content": "We quantify the performance by using the mean Intersection over Union (mIoU), which is defined as the average IoU across the considered classes [4]. We provide separate mIoU values for the base and novel classes. We also report the overall mIoU computed across all the classes in the dataset for completeness." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 286, + 713 + ], + "type": "text", + "content": "Implementation Details. We implement our network based on a MinkowskiUNet-34C network [10]. Point-level features are extracted from the penultimate layer. The segmentation heads are implemented as linear layers, producing output logits for each point in the batched point clouds. We train our network for 10 epochs. We use the SGD optimizer, with momentum 0.9 and weight decay 0.0001. Our learning rate scheduler consists of linear warm-up and co" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "text", + "content": "sine annealing, with " + }, + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "inline_equation", + "content": "lr_{max} = 10^{-2}" + }, + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "inline_equation", + "content": "lr_{min} = 10^{-5}" + }, + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "text", + "content": ". We train with batch size equal to 4. We employ 5 segmentation heads, that are used in synergy with an equal number of over-clustering heads, with " + }, + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "inline_equation", + "content": "o = 3" + }, + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "text", + "content": ". In " + }, + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "text", + "content": ", we set " + }, + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "inline_equation", + "content": "p = 0.5" + }, + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "text", + "content": " for SemanticKITTI-" + }, + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "inline_equation", + "content": "n^i" + }, + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "inline_equation", + "content": "p = 0.3" + }, + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "text", + "content": " for SemanticPOSS-" + }, + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "inline_equation", + "content": "n^i" + }, + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "text", + "content": ". We adapted the implementation of the Sinkhorn-Knopp algorithm [11] from the code provided by [5], with the introduction of the queue and an in-place normalisation steps. Similarly to [5], we set " + }, + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "inline_equation", + "content": "n_{sk\\_iters} = 3" + }, + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "text", + "content": ", while we adopt a linear decay for " + }, + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "inline_equation", + "content": "\\epsilon_{start} = 0.3" + }, + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "inline_equation", + "content": "\\epsilon_{end} = 0.05" + }, + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 200, + 427, + 213 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 200, + 427, + 213 + ], + "spans": [ + { + "bbox": [ + 306, + 200, + 427, + 213 + ], + "type": "text", + "content": "5.2. Quantitative analysis" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 219, + 545, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 219, + 545, + 255 + ], + "spans": [ + { + "bbox": [ + 304, + 219, + 545, + 255 + ], + "type": "text", + "content": "Segmentation quality. Tabs. 3 & 4 report the quantitative results on SemanticPOSS and SemanticKITTI, respectively. We report the Full supervision setting as our upper bound." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 255, + 545, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 255, + 545, + 386 + ], + "spans": [ + { + "bbox": [ + 304, + 255, + 545, + 386 + ], + "type": "text", + "content": "On SemanticPOSS, we outperform EUMS† on three out of four splits with an improvement of 18.3 mIoU on POSS-40, 9.0 mIoU on POSS-31 and 0.6 on POSS-32. In these splits, NOPS shows a large improvement on all the novel classes, with the exception of the class fence in POSS-31 and traffic-sign in POSS-33. Differently, we deem the lower performance in POSS-33 is due to the difficulty and scarce presence of these novel classes. The advantage of EUMS† is the clustering on the whole dataset that enables a complete visibility of all the novel classes. On average, NOPS achieves 21.40 mIoU, improving over EUMS† of 6.5 mIoU." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 387, + 545, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 387, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 304, + 387, + 545, + 506 + ], + "type": "text", + "content": "On SemanticKITTI, we outperform EUMS† on all the four splits, improving by 12.6 mIoU on KITTI-5⁰, 1.2 mIoU on KITTI-5¹, 4.2mIoU on KITTI-5² and 5.3 mIoU on KITTI-4³. NOPS outperforms EUMS† by a large margin on the majority of the novel classes. Exceptions are the class sidewalk in KITTI-5⁰, car in KITTI-5¹, motorcycle in KITTI-5² and motorcyclist in KITTI-4³. On average, NOPS achieves 22.84 mIoU, improving over EUMS† of 5.8 mIoU. Interestingly, NOPS outperforms also the supervised upper bound on the class trunk in KITTI-5¹." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 506, + 545, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 506, + 545, + 637 + ], + "spans": [ + { + "bbox": [ + 304, + 506, + 545, + 637 + ], + "type": "text", + "content": "Computational time. NOPS outperforms EUMS† in terms of computational time. Firstly, EUMS† requires a pretraining step and a fine-tuning step, i.e. 30 training epochs in total. Then, EUMS† requires a large amount of memory (up to 200 GB memory for KITTI-5⁰) to store the data required for clustering, taking several hours (50 hrs) to complete the training procedure. Differently, NOPS achieves superior performance with 10 training epochs, by using less memory (10 GB max) and a lower computational time (up to 25 hrs for KITTI-5⁰). We run these tests using one GPU Tesla A40-48GB." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 647, + 421, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 647, + 421, + 659 + ], + "spans": [ + { + "bbox": [ + 306, + 647, + 421, + 659 + ], + "type": "text", + "content": "5.3. Qualitative analysis" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "Fig. 3 shows some segmentation results of NOPS and EUMS† on SemanticPOSS and SemaniticKITTI. We can observe that the predictions of the base classes in the two datasets are correct for both the models, with just minor er" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9398" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 97, + 545, + 239 + ], + "blocks": [ + { + "bbox": [ + 47, + 71, + 545, + 94 + ], + "lines": [ + { + "bbox": [ + 47, + 71, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 545, + 94 + ], + "type": "text", + "content": "Table 3. Novel class discovery results on SemanticPOSS. NOPs outperforms EUMS† on three out of four splits. Full supervision: model trained with labels for base and novel classes. EUMS†: baseline described in Sec. 4. Highlighted values are the novel classes in each split." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 97, + 545, + 239 + ], + "lines": [ + { + "bbox": [ + 47, + 97, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 47, + 97, + 545, + 239 + ], + "type": "table", + "html": "
SplitModelbluebub-carcome.fencegrou.pers.plantspoleridertraf.trafe.trunkNovelmIoU BaseAll
Full supervision43.2071.3033.0032.5044.6078.5061.8073.9030.9054.7026.7011.0019.30--44.72
POSS-40EUMS† [40]25.673.980.5616.4429.4036.7643.8428.4613.1326.7518.183.3416.9117.4421.5220.26
NOPS (Ours)35.4730.351.2413.5224.1369.1444.7042.0719.1947.6524.448.1721.8235.7026.5729.38
POSS-31EUMS† [40]15.1767.9828.0223.9811.8875.0735.9874.4626.9148.5626.005.6023.0521.0139.9635.59
NOPS (Ours)29.3571.3528.7012.213.9478.2456.7874.2118.2938.8823.3113.7423.5130.0238.2436.35
POSS-32EUMS† [40]40.1469.4527.6713.5034.8676.0354.6675.595.2739.227.798.5211.858.3143.9635.74
NOPS (Ours)37.1671.8129.7414.6428.3877.5352.0973.0011.5147.110.5410.2014.798.9544.1736.04
POSS-33EUMS† [40]41.1770.6828.084.3438.2776.6638.2975.3525.7634.3428.310.3624.4013.0144.7037.38
NOPS (Ours)38.5570.3630.910.0029.3876.5055.9871.8417.0331.8726.150.9522.5710.9443.9336.32
AvgEUMS† [40]14.9437.5432.24
NOPS (Ours)21.4038.2334.52
", + "image_path": "2fa24ac208d596a03a813569770bf714ef68bed88adece4d221ecc0c9220606a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 47, + 275, + 545, + 406 + ], + "blocks": [ + { + "bbox": [ + 47, + 250, + 545, + 273 + ], + "lines": [ + { + "bbox": [ + 47, + 250, + 545, + 273 + ], + "spans": [ + { + "bbox": [ + 47, + 250, + 545, + 273 + ], + "type": "text", + "content": "Table 4. Novel class discovery results on SemanticKITTI. NOPS outperforms EUMS† on all four splits. Full supervision: model trained with annotations for base and novel classes. EUMS†: baseline described in Sec. 4. Highlighted values are the novel classes in each split." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 275, + 545, + 406 + ], + "lines": [ + { + "bbox": [ + 47, + 275, + 545, + 406 + ], + "spans": [ + { + "bbox": [ + 47, + 275, + 545, + 406 + ], + "type": "table", + "html": "
SplitModelbicyclebicyclebicyclecarfencepolepoleothersothersparkpers.poleroadsidewterracetraff.trucktrunkvessels.NovelmIoU BaseAll
Full supervision6.3039.5085.4090.0023.2020.305.703.9018.0028.9031.0040.6090.9074.6062.1020.5062.9046.2083.90--43.89
KITTI-50EUMS† [40]5.2839.9615.7779.209.0316.892.520.0711.3914.4012.6729.1742.5826.100.0510.3047.3737.9238.3524.5721.0823.11
NOPS (Ours)5.5947.7652.6882.6013.7625.551.361.6614.5219.8025.8632.1256.748.0823.8414.2849.4136.1844.1737.1024.7029.62
KITTI-51EUMS† [40]7.5342.4179.9776.778.6219.581.390.5712.0314.1413.9540.7486.3266.4556.2911.9744.7920.9472.4024.2137.0635.62
NOPS (Ours)7.3651.2384.5350.877.2728.931.760.0022.2019.3930.4237.6190.0772.1860.7516.7857.3449.2585.1225.3643.0940.69
KITTI-52EUMS† [40]8.2650.7882.9888.0517.882.752.320.173.1625.4024.9820.2088.3071.0457.858.6327.1638.3676.9512.3842.2236.59
NOPS (Ours)6.7249.2486.3690.7923.682.690.581.8715.4629.4827.9236.3990.2673.3961.2117.8310.3246.1684.2916.5444.8039.72
KITTI-43EUMS† [40]3.952.4780.1087.2116.8114.0214.980.3114.1320.776.8037.5986.7966.5055.2616.2040.6238.3776.157.0543.3935.74
NOPS (Ours)2.3227.8386.0489.8923.0624.472.923.0618.1930.0916.3239.9090.6573.5161.0417.4049.7644.0183.1812.3548.9541.24
AvgEUMS† [40]17.0535.9432.76
NOPS (Ours)22.8440.3837.73
", + "image_path": "95a843c5a16f677dcdd021709e1ac708350872cc57031ad3edf1ec55d85078ba.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 427, + 287, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 427, + 287, + 535 + ], + "spans": [ + { + "bbox": [ + 46, + 427, + 287, + 535 + ], + "type": "text", + "content": "rors at the edges of the objects. NOPS shows better segmentation capabilities also when dealing with the novel classes, being able to properly segment the correct class for the unknown objects. On the larger objects, such as buildings in the scenes, some mixing of labels can be observed. Differently, EUMS nearly fails in correctly recognising the novel objects, resulting in the inclusion of different classes (either novel or base) into the same object, e.g. the facade of the building in POSS-3" + }, + { + "bbox": [ + 46, + 427, + 287, + 535 + ], + "type": "inline_equation", + "content": "^1" + }, + { + "bbox": [ + 46, + 427, + 287, + 535 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 548, + 145, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 548, + 145, + 560 + ], + "spans": [ + { + "bbox": [ + 47, + 548, + 145, + 560 + ], + "type": "text", + "content": "6. Ablation studies" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 568, + 287, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 568, + 287, + 592 + ], + "spans": [ + { + "bbox": [ + 46, + 568, + 287, + 592 + ], + "type": "text", + "content": "We study the main components of NOPS and its behaviour when varying its parameters on SemanticPOSS." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 594, + 287, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 652 + ], + "type": "text", + "content": "Method components. Fig. 4 shows the performance on novel and base classes of seven versions of NOPS. The first three versions use the pre-trained model on the base classes, while the last four versions use the model trained from scratch. Each version is defined as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 58, + 654, + 286, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 58, + 654, + 286, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 654, + 286, + 677 + ], + "spans": [ + { + "bbox": [ + 58, + 654, + 286, + 677 + ], + "type": "text", + "content": "- P: we use a pre-trained model, and we remove " + }, + { + "bbox": [ + 58, + 654, + 286, + 677 + ], + "type": "inline_equation", + "content": "Z_{q}" + }, + { + "bbox": [ + 58, + 654, + 286, + 677 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 58, + 654, + 286, + 677 + ], + "type": "inline_equation", + "content": "\\tau_{c}" + }, + { + "bbox": [ + 58, + 654, + 286, + 677 + ], + "type": "text", + "content": " and the over-clustering heads." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 58, + 677, + 194, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 677, + 194, + 689 + ], + "spans": [ + { + "bbox": [ + 58, + 677, + 194, + 689 + ], + "type": "text", + "content": "- OC: P + over-clustering heads." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 58, + 689, + 277, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 689, + 277, + 701 + ], + "spans": [ + { + "bbox": [ + 58, + 689, + 277, + 701 + ], + "type": "text", + "content": "- Q: " + }, + { + "bbox": [ + 58, + 689, + 277, + 701 + ], + "type": "inline_equation", + "content": "\\mathrm{OC} + Z_{q}" + }, + { + "bbox": [ + 58, + 689, + 277, + 701 + ], + "type": "text", + "content": ", i.e. our queue without class-balancing." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 58, + 701, + 181, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 701, + 181, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 701, + 181, + 713 + ], + "type": "text", + "content": "- NP: Q without pre-training." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 317, + 427, + 544, + 474 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 317, + 427, + 528, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 427, + 528, + 438 + ], + "spans": [ + { + "bbox": [ + 317, + 427, + 528, + 438 + ], + "type": "text", + "content": "- NP+: NP + our selection function " + }, + { + "bbox": [ + 317, + 427, + 528, + 438 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 317, + 427, + 528, + 438 + ], + "type": "text", + "content": " on the queue." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 439, + 544, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 439, + 544, + 461 + ], + "spans": [ + { + "bbox": [ + 317, + 439, + 544, + 461 + ], + "type": "text", + "content": "- NP++: NP + " + }, + { + "bbox": [ + 317, + 439, + 544, + 461 + ], + "type": "inline_equation", + "content": "\\tau_{c}" + }, + { + "bbox": [ + 317, + 439, + 544, + 461 + ], + "type": "text", + "content": " on the features used to derive the pseudo-labels." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 462, + 514, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 462, + 514, + 474 + ], + "spans": [ + { + "bbox": [ + 317, + 462, + 514, + 474 + ], + "type": "text", + "content": "Full: NOPS with all the components activated." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 304, + 480, + 545, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 480, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 304, + 480, + 545, + 624 + ], + "type": "text", + "content": "Pre-trained approaches generally underperform their trained-from-scratch counterparts on the novel classes. This is visible in the low performance of P, OC and Q. We have a significant improvement when pre-training is not used (NP), i.e. we achieve " + }, + { + "bbox": [ + 304, + 480, + 545, + 624 + ], + "type": "inline_equation", + "content": "20.26\\mathrm{mIoU}" + }, + { + "bbox": [ + 304, + 480, + 545, + 624 + ], + "type": "text", + "content": ". We can see that the queue both with and without pre-training is helpful. When we add the feature selection for the queue and for the training, i.e. " + }, + { + "bbox": [ + 304, + 480, + 545, + 624 + ], + "type": "inline_equation", + "content": "\\mathrm{NP+}" + }, + { + "bbox": [ + 304, + 480, + 545, + 624 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 480, + 545, + 624 + ], + "type": "inline_equation", + "content": "\\mathrm{NP++}" + }, + { + "bbox": [ + 304, + 480, + 545, + 624 + ], + "type": "text", + "content": ", we have improvements, i.e. 20.63 mIoU and 20.90 mIoU, respectively. The best performance is achieved with Full. Although we can observe variations on the performance of the base classes, their information is retained by the network when we discover the novel ones." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "Parameter analysis. We study the behaviour of the percentile " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " in our selection function " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": ", when we apply it to the features both for pseudo-labelling and for the class-balanced queue " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "Z_{q}" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": ". Tab. 5 reports the results on each split of SemanticPOSS. For each split, we can observe that the performance depends on the number of points and difficulty of the novel classes. In " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{POSS - 4^0}" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{POSS - 3^1}" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": ", lower values" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9399" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 41, + 59, + 545, + 265 + ], + "blocks": [ + { + "bbox": [ + 41, + 59, + 545, + 265 + ], + "lines": [ + { + "bbox": [ + 41, + 59, + 545, + 265 + ], + "spans": [ + { + "bbox": [ + 41, + 59, + 545, + 265 + ], + "type": "image", + "image_path": "b48fc6ebf247d7fbeeae594b6b55c2030f35ce13f6313749a2d85341d1a3955b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 269, + 546, + 304 + ], + "lines": [ + { + "bbox": [ + 46, + 269, + 546, + 304 + ], + "spans": [ + { + "bbox": [ + 46, + 269, + 546, + 304 + ], + "type": "text", + "content": "Figure 3. Qualitative comparisons on (top) SemanticPOSS and (bottom) SemanticKITTI. We report results on both base and novel classes. On novel classes, " + }, + { + "bbox": [ + 46, + 269, + 546, + 304 + ], + "type": "inline_equation", + "content": "\\mathrm{EUMS}^{\\dagger}" + }, + { + "bbox": [ + 46, + 269, + 546, + 304 + ], + "type": "text", + "content": " fails in recognising novel objects with mixed and cluttered predictions, e.g. the facade of building in " + }, + { + "bbox": [ + 46, + 269, + 546, + 304 + ], + "type": "inline_equation", + "content": "\\mathrm{POSS - 3^{1}}" + }, + { + "bbox": [ + 46, + 269, + 546, + 304 + ], + "type": "text", + "content": ". NOPS shows superior segmentation performance on all the novel classes." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 47, + 320, + 288, + 407 + ], + "blocks": [ + { + "bbox": [ + 47, + 320, + 288, + 407 + ], + "lines": [ + { + "bbox": [ + 47, + 320, + 288, + 407 + ], + "spans": [ + { + "bbox": [ + 47, + 320, + 288, + 407 + ], + "type": "image", + "image_path": "2050eaad008e6e9120781e8a12d54f5512ec7210806b4c4bf3e31183ecff57f8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 411, + 288, + 477 + ], + "lines": [ + { + "bbox": [ + 46, + 411, + 288, + 477 + ], + "spans": [ + { + "bbox": [ + 46, + 411, + 288, + 477 + ], + "type": "text", + "content": "Figure 4. Ablation study with different components and initialisation strategies on SemanticPOSS. In P, OC and Q, we initialise the model after base pre-training, and use different configurations of the over-clustering heads and of our queue balancing. In NP NP+, NP++ and Full, we begin with Q, we avoid pre-training, and we use " + }, + { + "bbox": [ + 46, + 411, + 288, + 477 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 46, + 411, + 288, + 477 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 411, + 288, + 477 + ], + "type": "inline_equation", + "content": "\\tau_{c}" + }, + { + "bbox": [ + 46, + 411, + 288, + 477 + ], + "type": "text", + "content": " incrementally. See Sec. 6 for definition of methods." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 499, + 287, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 499, + 287, + 584 + ], + "spans": [ + { + "bbox": [ + 46, + 499, + 287, + 584 + ], + "type": "text", + "content": "of " + }, + { + "bbox": [ + 46, + 499, + 287, + 584 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 499, + 287, + 584 + ], + "type": "text", + "content": " result in less severe selection. We believe this is related to the class distribution within these splits. This is in line with what observed in Tab. 3. In " + }, + { + "bbox": [ + 46, + 499, + 287, + 584 + ], + "type": "inline_equation", + "content": "\\mathrm{POSS}-3^{2}" + }, + { + "bbox": [ + 46, + 499, + 287, + 584 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 499, + 287, + 584 + ], + "type": "inline_equation", + "content": "\\mathrm{POSS}-3^{3}" + }, + { + "bbox": [ + 46, + 499, + 287, + 584 + ], + "type": "text", + "content": ", we notice a different behaviour, a higher value of " + }, + { + "bbox": [ + 46, + 499, + 287, + 584 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 499, + 287, + 584 + ], + "type": "text", + "content": " provides better results. We relate this to the difficulty of the novel classes in these splits whose noisy pseudo-labels can benefit from a more rigorous selection of the features." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 597, + 125, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 597, + 125, + 609 + ], + "spans": [ + { + "bbox": [ + 47, + 597, + 125, + 609 + ], + "type": "text", + "content": "7. Conclusions" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "content": "We explored the new problem of novel class discovery for 3D point cloud segmentation. Firstly, we transposed the only NCD method for 2D semantic segmentation to 3D point cloud data, and experimentally found that it has several limitations. We discussed that extending 2D NCD approaches to 3D data (point clouds) is not trivial because the assumptions made for 2D data are not easily transferable to 3D. Secondly, we presented NOPS, to tackle NCD for point" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 326, + 391, + 526, + 483 + ], + "blocks": [ + { + "bbox": [ + 305, + 322, + 547, + 389 + ], + "lines": [ + { + "bbox": [ + 305, + 322, + 547, + 389 + ], + "spans": [ + { + "bbox": [ + 305, + 322, + 547, + 389 + ], + "type": "text", + "content": "Table 5. Ablation study showing how different values of " + }, + { + "bbox": [ + 305, + 322, + 547, + 389 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 305, + 322, + 547, + 389 + ], + "type": "text", + "content": " affect the performance on SemanticPOSS. The lower " + }, + { + "bbox": [ + 305, + 322, + 547, + 389 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 305, + 322, + 547, + 389 + ], + "type": "text", + "content": " is, the less severe the selection of the features, resulting in better performances for POSS-4" + }, + { + "bbox": [ + 305, + 322, + 547, + 389 + ], + "type": "inline_equation", + "content": "^0" + }, + { + "bbox": [ + 305, + 322, + 547, + 389 + ], + "type": "text", + "content": ". Differently, POSS-3" + }, + { + "bbox": [ + 305, + 322, + 547, + 389 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 305, + 322, + 547, + 389 + ], + "type": "text", + "content": " benefits from an higher value of " + }, + { + "bbox": [ + 305, + 322, + 547, + 389 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 305, + 322, + 547, + 389 + ], + "type": "text", + "content": ", which leads to a more vigorous filtering of the features. POSS-3" + }, + { + "bbox": [ + 305, + 322, + 547, + 389 + ], + "type": "inline_equation", + "content": "^1" + }, + { + "bbox": [ + 305, + 322, + 547, + 389 + ], + "type": "text", + "content": " and POSS-3" + }, + { + "bbox": [ + 305, + 322, + 547, + 389 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 305, + 322, + 547, + 389 + ], + "type": "text", + "content": " show the best performances with " + }, + { + "bbox": [ + 305, + 322, + 547, + 389 + ], + "type": "inline_equation", + "content": "p = 0.5" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 326, + 391, + 526, + 483 + ], + "lines": [ + { + "bbox": [ + 326, + 391, + 526, + 483 + ], + "spans": [ + { + "bbox": [ + 326, + 391, + 526, + 483 + ], + "type": "table", + "html": "
SplitPercentile p
0.10.30.50.70.9
POSS-4030.8135.7028.7730.9326.69
POSS-3128.3330.0230.4323.3218.91
POSS-328.078.9510.3210.257.76
POSS-3310.5510.9411.6914.3813.42
Avg.19.4421.4020.3019.7216.70
", + "image_path": "949590fa56add36bc0ddbd876574e1b0fa94613365a273a90078be593835ae8f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 507, + 545, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 507, + 545, + 615 + ], + "spans": [ + { + "bbox": [ + 304, + 507, + 545, + 615 + ], + "type": "text", + "content": "cloud segmentation by using online clustering and exploiting uncertainty quantification to produce pseudo-labels for the novel points. Lastly, we introduced a new evaluation protocol to assess the performance of NCD for point cloud segmentation. Experiments on two different segmentation dataset showed that NOPS outperforms the compared baselines by a large margin. Future research directions could investigate the extension of our method when base annotations are fewer and/or weakly labelled." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "type": "text", + "content": "Limitations NOPS limitations include the prior knowledge on the number of novel classes " + }, + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "type": "inline_equation", + "content": "C_n" + }, + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "type": "text", + "content": " to discover. This could be a limitation when " + }, + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "type": "inline_equation", + "content": "C_n" + }, + { + "bbox": [ + 304, + 617, + 546, + 713 + ], + "type": "text", + "content": " is not known a-priori and novel classes may appear in an incremental manner. We believe that a solution may be to learn novel classes incrementally. Another limitation is the loss we use to handle class unbalancing. More recent techniques to handle this drawback could be further explored." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9400" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 92, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 53, + 92, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 92, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 92, + 287, + 135 + ], + "type": "text", + "content": "[1] I. Alonso, L. Riazuelo, L. Montesano, and A.C. Murillo. 3D MiniNet: Learning a 2D representation from point clouds for fast and efficient 3D LiDAR semantic segmentation. In IROS, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 287, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 287, + 170 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 287, + 170 + ], + "type": "text", + "content": "[2] Y.M. Asano, C. Rupprecht, and A. Vedaldi. Self-labelling via simultaneous clustering and representation learning. In ICLR, 2020. 3, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 172, + 287, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 172, + 287, + 215 + ], + "spans": [ + { + "bbox": [ + 53, + 172, + 287, + 215 + ], + "type": "text", + "content": "[3] J. Behley, M. Garbade, A. Milioto, J. Quenzel, S. Behnke, J. Gall, and C. Stachniss. Towards 3D LiDAR-based semantic scene understanding of 3D point cloud sequences: The SemanticKITTI Dataset. In IJRR, 2021. 2, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 218, + 287, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 218, + 287, + 261 + ], + "spans": [ + { + "bbox": [ + 53, + 218, + 287, + 261 + ], + "type": "text", + "content": "[4] J. Behley, M. Garbade, A. Milioto, J. Quenzel, S. Behnke, C. Stachniss, and J. Gall. SemanticKITTI: A dataset for semantic scene understanding of LiDAR sequences. In CVPR, 2019. 2, 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 263, + 287, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 263, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 53, + 263, + 287, + 297 + ], + "type": "text", + "content": "[5] M. Caron, I. Misra, J. Mairal, P. Goyal, P. Bojanowski, and A. Joulin. Unsupervised learning of visual features by contrasting cluster assignments. In NeurIPS, 2020. 2, 4, 5, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 298, + 287, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 298, + 287, + 331 + ], + "spans": [ + { + "bbox": [ + 53, + 298, + 287, + 331 + ], + "type": "text", + "content": "[6] J. Cen, P. Yun, S. Zhang, J. Cai, D. Luan, M. Tang, M. Liu, and M.Y. Wang. Open-world semantic segmentation for lidar point clouds. In ECCV, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 333, + 287, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 333, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 53, + 333, + 287, + 376 + ], + "type": "text", + "content": "[7] L.C. Chen, G. Papandreou, I. Kokkinos, K. Murphy, and A.L. Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. In T-PAMI, 2017. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 379, + 287, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 379, + 287, + 411 + ], + "spans": [ + { + "bbox": [ + 53, + 379, + 287, + 411 + ], + "type": "text", + "content": "[8] L.C. Chen, Y. Zhu, G. Papandreou, F. Schroff, and A. Hartwig. Encoder-decoder with atrous separable convolution for semantic image segmentation. In ECCV, 2018. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 414, + 287, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 414, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 53, + 414, + 287, + 456 + ], + "type": "text", + "content": "[9] R. Cheng, R. Razani, E. Taghavi, E. Li, and B. Liu. (AF)2-S3Net: Attentive Feature Fusion with Adaptive Feature Selection for Sparse Semantic Segmentation Network. In CVPR, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 459, + 287, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 459, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 48, + 459, + 287, + 491 + ], + "type": "text", + "content": "[10] C. Choy, JY. Gwak, and S. Savarese. 4D spatio-temporal convnets: Minkowski convolutional neural networks. In CVPR, 2019. 2, 3, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 494, + 287, + 516 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 494, + 287, + 516 + ], + "spans": [ + { + "bbox": [ + 48, + 494, + 287, + 516 + ], + "type": "text", + "content": "[11] M. Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. In NeurIPS, 2013. 3, 4, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 517, + 287, + 550 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 517, + 287, + 550 + ], + "spans": [ + { + "bbox": [ + 48, + 517, + 287, + 550 + ], + "type": "text", + "content": "[12] E. Fini, E. Sangineto, S. Lathuiliere, Z. Zhong, M. Nabi, and E. Ricci. A unified objective for novel class discovery. In CVPR, 2021. 1, 2, 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 552, + 287, + 585 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 552, + 287, + 585 + ], + "spans": [ + { + "bbox": [ + 48, + 552, + 287, + 585 + ], + "type": "text", + "content": "[13] A. Geiger, P. Lenz, and R. Urtasun. Are we ready for Autonomous Driving? The KITTI Vision Benchmark Suite. In CVPR, 2012. 2, 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 587, + 287, + 619 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 587, + 287, + 619 + ], + "spans": [ + { + "bbox": [ + 48, + 587, + 287, + 619 + ], + "type": "text", + "content": "[14] B. Graham, M. Engelcke, and L. van der Maaten. 3D semantic segmentation with submanifold sparse convolutional networks. In CVPR, 2018. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 622, + 287, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 622, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 48, + 622, + 287, + 643 + ], + "type": "text", + "content": "[15] B. Graham and L. van der Maaten. Submanifold sparse convolutional networks. arXiv, 2017. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 646, + 287, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 646, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 48, + 646, + 287, + 677 + ], + "type": "text", + "content": "[16] K. Han, A. Vedaldi, and A. Zisserman. Learning to discover novel visual categories via deep transfer clustering. In CVPR, 2019. 1, 2, 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[17] Q. Hu, B. Yang, L. Xie, S. Rosa, Y. Guo, Z. Wang, N. Trigoni, and A. Markham. Randla-net: Efficient semantic segmentation of large-scale point clouds. In CVPR, 2020. 2" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 38, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[18] X. Ji, J.F. Henriques, and A. Vedaldi. Invariant information clustering for unsupervised image classification and segmentation. In ICCV, 2019. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 107, + 545, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 138 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 138 + ], + "type": "text", + "content": "[19] X. Jia, K. Han, Y. Zhu, and B. Green. Joint representation learning and novel category discovery on single-and multimodal data. In ICCV, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 140, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 140, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 140, + 545, + 173 + ], + "type": "text", + "content": "[20] K. Joseph, S. Paul, G. Aggarwal, S. Biswas, P. Rai, K. Han, and V.N. Balasubramanian. Novel class discovery without forgetting. ECCV, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 175, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 206 + ], + "type": "text", + "content": "[21] G. Mei, F. Poiesi, C. Saltori, J. Zhang, E. Ricci, and N. Sebe. Overlap-guided gaussian mixture models for point cloud registration. In WACV, 2023. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 208, + 545, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 208, + 545, + 240 + ], + "spans": [ + { + "bbox": [ + 307, + 208, + 545, + 240 + ], + "type": "text", + "content": "[22] G. Mei, C. Saltori, F. Poiesi, J. Zhang, E. Ricci, N. Sebe, and Q. Wu. Data augmentation-free unsupervised learning for 3d point cloud understanding. In BMVC, 2022. 3, 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 241, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 241, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 545, + 274 + ], + "type": "text", + "content": "[23] A. Milioto, I. Vizzo, J. Behley, and C. Stachniss. Rangenet++: Fast and accurate lidar semantic segmentation. In IROS, 2019. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 276, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 276, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 276, + 545, + 308 + ], + "type": "text", + "content": "[24] Y. Pan, B. Gao, J. Mei, S. Geng, C. Li, and H. Zhao. SemanticPOSS: A point cloud dataset with large quantity of dynamic instances. In IV, 2020. 2, 5, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 309, + 545, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 309, + 545, + 341 + ], + "spans": [ + { + "bbox": [ + 307, + 309, + 545, + 341 + ], + "type": "text", + "content": "[25] C.R. Qi, H. Su, K. Mo, and L.J. Guibas. Pointnet: Deep learning on point sets for 3D classification and segmentation. In CVPR, 2017. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 342, + 545, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 342, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 307, + 342, + 545, + 374 + ], + "type": "text", + "content": "[26] C.R. Qi, L. Yi, H. Su, and L.J. Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In NeurIPS, 2017. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 376, + 545, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 376, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 307, + 376, + 545, + 409 + ], + "type": "text", + "content": "[27] O. Ronneberger, P. Fischer, and T. Brox. U-net: Convolutional networks for biomedical image segmentation. In MIC-CAI, 2015. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 411, + 545, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 411, + 545, + 432 + ], + "spans": [ + { + "bbox": [ + 307, + 411, + 545, + 432 + ], + "type": "text", + "content": "[28] S. Roy, M. Liu, Z. Zhong, N. Sebe, and E. Ricci. Class incremental novel class discovery. In ECCV, 2022. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 434, + 545, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 434, + 545, + 466 + ], + "spans": [ + { + "bbox": [ + 307, + 434, + 545, + 466 + ], + "type": "text", + "content": "[29] C. Saltori, F. Galasso, G. Fiameni, N. Sebe, E. Ricci, and F. Poiesi. Cosmix: Compositional semantic mix for domain adaptation in 3d lidar segmentation. In ECCV, 2022. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 468, + 545, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 468, + 545, + 499 + ], + "spans": [ + { + "bbox": [ + 307, + 468, + 545, + 499 + ], + "type": "text", + "content": "[30] R. Song, W. Zhang, Y. Zhao, Y. Liu, and P.L. Rosin. Mesh saliency: An independent perceptual measure or a derivative of image saliency? In CVPR, 2021. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 501, + 545, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 501, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 307, + 501, + 545, + 533 + ], + "type": "text", + "content": "[31] N. Souly, C. Spampinato, and M. Shah. Semi supervised semantic segmentation using generative adversarial network. In ICCV, 2017. 2" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 535, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 535, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 307, + 535, + 545, + 567 + ], + "type": "text", + "content": "[32] X. Tang, J. Wang, B. Gao, E. Dellandrea, R. Gaizauskas, and L. Chen. Large scale semi-supervised object detection using visual and semantic knowledge transfer. In CVPR, 2016. 2" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 307, + 568, + 545, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 568, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 307, + 568, + 545, + 601 + ], + "type": "text", + "content": "[33] H. Thomas, C.R. Qi, J.-E. Deschaud, B. Marcotegui, F. Goulette, and J. L. Guibas. Kpconv: Flexible and deformable convolution for point clouds. In ICCV, 2019. 2" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 307, + 602, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 602, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 307, + 602, + 545, + 624 + ], + "type": "text", + "content": "[34] S. Vaze, K. Han, A. Vedaldi, and A. Zisserman. Generalized category discovery. In CVPR, 2022. 2" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 307, + 625, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 625, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 307, + 625, + 545, + 667 + ], + "type": "text", + "content": "[35] B. Wu, A. Wan, X. Yue, and K. Keutzer. Squeezeseg: Convolutional neural nets with recurrent CRF for real-time road-object segmentation from 3D LiDAR point cloud. In ICRA, 2018. 2" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "text", + "content": "[36] B. Wu, X. Zhou, S. Zhao, X. Yue, and K. Keutzer. Squeeze-segv2: Improved model structure and unsupervised domain adaptation for road-object segmentation from a LiDAR point cloud. In ICRA, 2019. 2" + } + ] + } + ], + "index": 37 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "9401" + } + ] + } + ], + "index": 39 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 331 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 286, + 106 + ], + "type": "text", + "content": "[37] M. Yang, Y. Zhu, J. Yu, A. Wu, and C. Deng. Divide and conquer: Compositional experts for generalized novel class discovery. In CVPR, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 107, + 286, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 107, + 286, + 129 + ], + "spans": [ + { + "bbox": [ + 49, + 107, + 286, + 129 + ], + "type": "text", + "content": "[38] L. Zhang and G.J. Qi. Wcp: Worst-case perturbations for semi-supervised deep learning. In CVPR, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 130, + 286, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 130, + 286, + 173 + ], + "spans": [ + { + "bbox": [ + 49, + 130, + 286, + 173 + ], + "type": "text", + "content": "[39] Y. Zhang, Z. Zhou, P. David, X. Yue, Z. Xi, B. Gong, and H. Foroosh. Polarnet: An improved grid representation for online LiDAR point clouds semantic segmentation. In CVPR, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 175, + 286, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 175, + 286, + 207 + ], + "spans": [ + { + "bbox": [ + 49, + 175, + 286, + 207 + ], + "type": "text", + "content": "[40] Y. Zhao, Z. Zhong, N. Sebe, and G.H. Lee. Novel class discovery in semantic segmentation. In CVPR, 2022. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 209, + 286, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 209, + 286, + 240 + ], + "spans": [ + { + "bbox": [ + 49, + 209, + 286, + 240 + ], + "type": "text", + "content": "[41] Z. Zhong, E. Fini, S. Roy, Z. Luo, E. Ricci, and N. Sebe. Neighborhood contrastive learning for novel class discovery. In CVPR, 2021. 1, 2, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 243, + 286, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 243, + 286, + 275 + ], + "spans": [ + { + "bbox": [ + 49, + 243, + 286, + 275 + ], + "type": "text", + "content": "[42] Z. Zhong, L. Zhu, Z. Luo, S. Li, Y. Yang, and N. Sebe. Openmix: Reviving known knowledge for discovering novel visual categories in an open world. In CVPR, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 277, + 286, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 277, + 286, + 298 + ], + "spans": [ + { + "bbox": [ + 49, + 277, + 286, + 298 + ], + "type": "text", + "content": "[43] Y. Zhou and O. Tuzel. Voxelnet: End-to-end learning for point cloud based 3D object detection. In CVPR, 2018. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 300, + 286, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 300, + 286, + 331 + ], + "spans": [ + { + "bbox": [ + 49, + 300, + 286, + 331 + ], + "type": "text", + "content": "[44] X. Zhu, H. Zhou, T. Wang, F. Hong, Y. Ma, W. Li, H. Li, and D. Lin. Cylindrical and asymmetrical 3D convolution networks for lidar segmentation. In CVPR, 2021. 2" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9402" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Novel-View Acoustic Synthesis/2dea3110-7f39-4b1f-9a6f-a10e0ea0bd92_content_list.json b/2023/Novel-View Acoustic Synthesis/2dea3110-7f39-4b1f-9a6f-a10e0ea0bd92_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5429c3f46462610c83789e098ae2ad1ae781424b --- /dev/null +++ b/2023/Novel-View Acoustic Synthesis/2dea3110-7f39-4b1f-9a6f-a10e0ea0bd92_content_list.json @@ -0,0 +1,1669 @@ +[ + { + "type": "text", + "text": "Novel-View Acoustic Synthesis", + "text_level": 1, + "bbox": [ + 328, + 130, + 640, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Changan Chen $^{1,3}$ Alexander Richard $^{2}$ Roman Shapovalov $^{3}$ Vamsi Krishna Ithapu $^{2}$ Natalia Neverova $^{3}$ Kristen Grauman $^{1,3}$ Andrea Vedaldi $^{3}$ $^{1}$ University of Texas at Austin $^{2}$ Reality Labs Research at Meta $^{3}$ FAIR, Meta AI", + "bbox": [ + 129, + 179, + 833, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 272, + 313, + 287 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We introduce the novel-view acoustic synthesis (NVAS) task: given the sight and sound observed at a source viewpoint, can we synthesize the sound of that scene from an unseen target viewpoint? We propose a neural rendering approach: Visually-Guided Acoustic Synthesis (ViGAS) network that learns to synthesize the sound of an arbitrary point in space by analyzing the input audio-visual cues. To benchmark this task, we collect two first-of-their-kind large-scale multi-view audio-visual datasets, one synthetic and one real. We show that our model successfully reasons about the spatial cues and synthesizes faithful audio on both datasets. To our knowledge, this work represents the very first formulation, dataset, and approach to solve the novel-view acoustic synthesis task, which has exciting potential applications ranging from AR/VR to art and design. Unlocked by this work, we believe that the future of novel-view synthesis is in multi-modal learning from videos.", + "bbox": [ + 75, + 303, + 473, + 561 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 584, + 209, + 599 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Replaying a video recording from a new viewpoint1 has many applications in cinematography, video enhancement, and virtual reality. For example, it can be used to edit a video, simulate a virtual camera, or, given a video of a personal memory, even enable users to experience a treasured moment again—not just on a 2D screen, but in 3D in a virtual or augmented reality, thus ‘reliving’ the moment.", + "bbox": [ + 75, + 604, + 468, + 710 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While the applications are exciting, there are still many unsolved technical challenges. Recent advances in 3D reconstruction and novel-view synthesis (NVS) address the problem of synthesizing new images of a given scene [32, 34, 44]. However, thus far, the view synthesis problem is concerned with creating visuals alone; the output is silent or at best naively adopts the sounds of the original video (from the \"wrong\" viewpoint). Without sound, the emotional and cognitive significance of the replay is severely diminished.", + "bbox": [ + 75, + 710, + 468, + 847 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we address this gap and introduce the new task of novel-view acoustic synthesis (NVAS). The goal of", + "bbox": [ + 76, + 848, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a0b9f4beb4e51b7fc4434efef3758801200bf5222e33d70bd4a94612aaa30df4.jpg", + "image_caption": [ + "Figure 1. Novel-view acoustic synthesis task. Given audio-visual observations from one viewpoint and the relative target viewpoint pose, render the sound received at the target viewpoint. Note that the target is expressed as the desired pose of the microphones; the image at that pose (right) is neither observed nor synthesized." + ], + "image_footnote": [], + "bbox": [ + 503, + 272, + 890, + 452 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "this task is to synthesize the sound in a scene from a new acoustic viewpoint, given only the visual and acoustic input from another source viewpoint in the same scene (Fig. 1).", + "bbox": [ + 498, + 551, + 892, + 597 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "NVAS is very different from the existing NVS task, where the goal is to reconstruct images instead of sounds, and these differences present new challenges. First, the 3D geometry of most real-life scenes changes in a limited manner during the recording. On the contrary, sound changes substantially over time, so the reconstruction target is highly dynamic. Secondly, visual and audio sensors are very different. A camera matrix captures the light in a highly-directional manner, and a single image comprises a large 2D array of pixels. In contrast, sounds are recorded with one or two microphones which are at best weakly-directional, providing only a coarse sampling of the sound field. Thirdly, the frequency of light waves is much higher than that of sound waves; the length of audio waves is thus larger to the point of being comparable to the size of geometric features of the scene, meaning that effects such as diffraction are often dominant, and spatial resolution is low. As a result, techniques that require spatial precision, such as triangulation and segmentation, are not applicable to audio. Lastly, sounds mix together, making it difficult to segment them,", + "bbox": [ + 496, + 598, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1We use \"viewpoint\" to mean a camera or microphone pose.", + "bbox": [ + 94, + 886, + 415, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "6409", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "and they are affected by environmental effects such as reverberation that are distributed and largely unobservable.", + "bbox": [ + 76, + 90, + 468, + 121 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While the NVS and NVAS tasks are indeed very different, we hypothesize that NVAS is an inherently multimodal task. In fact, vision can play an important role in achieving accurate sound synthesis. First, establishing correspondences between sounds and their sources as they appear in images can provide essential cues for resynthesizing the sounds realistically. For instance, human speech is highly directional and sounds very differently if one faces the speaker or their back, which can only be inferred from visual cues. In addition, the environment acoustics also affect the sound one hears as a function of the scene geometry, materials, and emitter/receiver locations. The same source sounds vary differently if it is located in the center of a room, at the corner, or in a corridor, for example. In short, vision provides cues about space and geometry that affect sound, and are difficult to estimate from the sound alone.", + "bbox": [ + 75, + 125, + 470, + 364 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In order to validate our hypothesis, we propose a novel visually-guided acoustic synthesis network that analyzes audio and visual features and synthesizes the audio at a target location. More specifically, the network first takes as input the image observed at the source viewpoint in order to infer global acoustic and geometric properties of the environment along with the bounding box of the active speaker. The network then reasons how the speaker and scene geometry change in 3D based on the relative target pose with a fusion network. We inject the fused features into audio with a gated multi-modal fusion network and model the acoustic changes between viewpoints with a time-domain model.", + "bbox": [ + 75, + 369, + 468, + 549 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In order to conduct our experiments on the new NVAS task, we require suitable training and benchmarking data, of which currently there is none available. To address that, we contribute two new datasets: one real (Replay-NVAS) and one synthetic (SoundSpaces-NVAS). The key feature of these datasets is to record the sight and sound of different scenes from multiple cameras/viewpoints. Replay-NVAS contains video recordings of groups of people performing social activities (e.g., chatting, watching TV, doing yoga, playing instruments) from 8 surrounding viewpoints simultaneously. It contains 37 hours of highly realistic everyday conversation and social interactions in one homelike environment. To our knowledge, Replay-NVAS represents the first large-scale real-world dataset enabling NVAS. This dataset would also greatly benefit many other existing tasks including NVS, active speaker localization, etc. For SoundSpaces-NVAS, we render 1.3K hours of audiovisual data based on the SoundSpaces [7] platform. Using this simulator, one can easily change the scene geometry and the positions of speakers, cameras, and microphones. This data serves as a powerful test bed with clean ground truth for a large collection of home environments, offering a good complement to Replay-NVAS. For both datasets,", + "bbox": [ + 75, + 553, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "we capture binaural audio, which is what humans perceive with two ears. Together the datasets contain 1,337 hours of audio-visual capture, with 1,032 speakers across 121 3D scenes. Datasets are publicly available for future research.", + "bbox": [ + 496, + 90, + 890, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We show that our model outperforms traditional signal processing approaches as well as learning-based baselines, often by a substantial margin, in a quantitative evaluation and a human study. We show qualitative examples where the model predicts acoustic changes according to the viewpoint changes, e.g., left channel becomes louder when the viewpoint changes from left to right. In a nutshell, we present the first work that deals with novel-view acoustic synthesis, and contribute two large-scale datasets along with a novel neural rendering approach for solving the task.", + "bbox": [ + 496, + 152, + 892, + 303 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 311, + 640, + 325 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Novel-view synthesis (NVS). Kickstarted by advances in neural rendering [34, 51], many recent works consider variants of the NVS problem. Most approaches assume dozens of calibrated images for reconstructing a single static scene. Closer to monocular video NVS, authors have considered reducing the number of input views [20, 25, 38, 46, 59] and modelling dynamic scenes [27, 28, 42, 43, 53, 55]. However, none of these works tackle audio.", + "bbox": [ + 496, + 332, + 890, + 452 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Acoustic matching and spatialization. NVAS requires accounting for (1) the environmental acoustics and (2) the geometric configuration of the target microphone(s) (e.g., monaural vs binaural). Modelling environmental acoustics has been addressed extensively by the audio community [4, 26]. Room impulse response (RIR) functions characterize the environment acoustics as a transfer function between the emitter and receiver, accounting for the scene geometry, materials, and emitter/receiver locations. Estimating the direct-to-reverberant ratio and the reverberation time, is sufficient to synthesize simple RIRs that match audio in a plausible manner [12, 16, 22, 30, 37, 58]. These methods do not synthesize for a target viewpoint, rather they resynthesize to match an audio sample. In [47, 48] sound from a moving emitter is spatialized towards a receiver conditioned on the tracked 3D location of the emitter.", + "bbox": [ + 496, + 455, + 892, + 696 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, the vision community explores using visual information to estimate environmental acoustics [6, 8, 50]. However, these works only synthesize acoustics for a given viewpoint rather than a novel viewpoint. In addition, they have only addressed monaural audio, which is more forgiving than binaural because humans do not perceive subtle absolute acoustic properties, but can detect easily inconsistencies in the sounds perceived by the two ears. Recent work spatializes monaural sounds by upmixing them to multiple channels conditioned on the video, where the sound emitters are static [17, 36]. Because the environment, emitter and receiver are static, so are the acoustics. Other work", + "bbox": [ + 496, + 698, + 893, + 878 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "2https://replay-dataset.github.io", + "bbox": [ + 514, + 886, + 779, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "6410", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "predicts impulse responses in simulation either for a single environment [29], or by using few-shot egocentric observations [31], or by using the 3D scene mesh [45]. While simulated results are satisfying, those models' impact on real-world data is unknown, especially for scenarios where human speakers move and interact with each other. Unlike any of the above, we introduce and tackle the NVAS problem, accounting for both acoustics and spatialization, and we propose a model that addresses the problem effectively on both synthetic and real-world data.", + "bbox": [ + 75, + 90, + 472, + 241 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Audio-visual learning. Recent advances in multi-modal video understanding enable new forms of self-supervised cross-modal feature learning from video [2, 24, 35], sound source localization [19, 21, 54], and audio-visual speech enhancement and source separation [1, 14, 33, 40, 60]. All of these existing tasks and datasets only deal with a single viewpoint. We introduce the first audio-visual learning task and dataset that deals with multi-view audio-visual data.", + "bbox": [ + 75, + 244, + 472, + 367 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. The Novel-view Acoustic Synthesis Task", + "text_level": 1, + "bbox": [ + 76, + 381, + 434, + 398 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We introduce a new task, novel-view acoustic synthesis (NVAS). Assuming there are $N$ sound emitters in the scene (emitter $i$ emits sound $C^i$ from location $L^i$ ), given the audio $A_{S}$ and video $V_{S}$ observed at the source viewpoint $S$ , the goal is to synthesize the audio $A_{T}$ at the target viewpoint $T$ , as it would sound from the target location, specified by the relative pose $P_T$ of the target microphone (translation and orientation) with respect to the source view (Fig. 1). Furthermore, we assume that the active sound emitters in the environment are visible in the source camera, but we make no assumptions about the camera at the target location.", + "bbox": [ + 75, + 407, + 472, + 574 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The sound at any point $R$ is a function of the space:", + "bbox": [ + 96, + 574, + 439, + 589 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nA _ {R} = \\mathcal {F} \\left(L ^ {1, \\dots , N}, C ^ {1, \\dots , N}, R \\mid E\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 155, + 590, + 468, + 609 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $R$ is the receiver location ( $S$ or $T$ ) and $E$ is the environment. The emitted sounds $C^i$ are not restricted to speech but can be ambient noise, sounding objects, etc. Our goal here is to learn a transfer function $\\mathcal{T}(\\cdot)$ defined as $A_{T} = \\mathcal{T}(A_{S}, V_{S}, P_{T})$ , where $S, T, L^{1,\\dots,N}, C^{1,\\dots,N}, E$ are not directly given and need to be inferred from $V_{S}$ and $P_{T}$ , which makes the task inherently multi-modal.", + "bbox": [ + 75, + 617, + 468, + 722 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This task is challenging because the goal is to model the sound field of a dynamic scene and capture acoustic changes between viewpoints given one pair of audio-visual measurements. While traditional signal processing methods can be applied, we show in Sec. 6 that they perform poorly. In this work, we present a learning-based rendering approach.", + "bbox": [ + 75, + 723, + 468, + 814 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Datasets", + "text_level": 1, + "bbox": [ + 76, + 829, + 174, + 844 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We introduce two datasets for the NVAS task: live recordings (Sec. 4.1), and simulated audio in scanned real-world environments (Sec. 4.2) (see Fig. 2). The former is", + "bbox": [ + 75, + 854, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2628d579c955d748b0b80d69315ee94ece60d745a3e22cf3d1514ae0be9e9f22.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 519, + 88, + 870, + 164 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/3eea53311068a0871c2aac1bdae156a0a5a63306f53e32a414c01d79fa28ccae.jpg", + "image_caption": [ + "Figure 2. Example source and target views for the two introduced datasets: Replay-NVAS (left) and SoundSpaces-NVAS (right)." + ], + "image_footnote": [], + "bbox": [ + 519, + 165, + 870, + 241 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "real and covers various social scenarios, but offers limited diversity of sound sources, viewpoints and environments, and is noisy. The latter has a realism gap, but allows perfect control over these aforementioned elements.", + "bbox": [ + 496, + 279, + 890, + 339 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Both datasets focus on human speech given its relevance in applications. However, our model design is not specific to speech. For both datasets, we capture binaural audio, which best aligns with human perception. Note that for both datasets, we collect multiple multi-modal views for training and evaluation; during inference the target viewpoint(s) (and in some cases target environment) are withheld.", + "bbox": [ + 496, + 340, + 890, + 445 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. The Replay-NVAS Dataset", + "text_level": 1, + "bbox": [ + 500, + 454, + 741, + 470 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Replay-NVAS contains multi-view captures of acted scenes in apartments. We capture 46 different scenarios (e.g., having a conversation, having dinner, or doing yoga) from 8 different viewpoints. In total, we collect 37 hours of video data, involving 32 participants across all scenarios.", + "bbox": [ + 496, + 477, + 890, + 553 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In each scenario, we invite 2-4 participants to act on a given topic. Each participant wears a near-range microphone, providing a clean recording of their own speech. The scene is captured by 8 DLSR cameras, each augmented with a 3Dio binaural microphone. In this way, the data captures video and audio simultaneously from multiple cameras, resulting in 56 possible source/target viewpoint combinations for each scene. The videos are recorded at 30 FPS and the audio is recorded with a 48k sampling rate. We use a clapper at the beginning of the recording for temporal synchronization. Each scenario lasts $3 - 8\\mathrm{min}$ . We use off-the-shelf software for multi-view camera calibration (see Supp.).", + "bbox": [ + 496, + 553, + 890, + 734 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To construct the dataset, we extract one-second long clips from each video with overlapping windows. We automatically remove silent and noisy clips based on the energy of near-range microphones, which results in 77K/12K/2K clips in total for train/val/test (details in Supp.) During training, for one sample, we randomly select two out of eight viewpoints, one as the source and one as the target.", + "bbox": [ + 496, + 734, + 890, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This dataset is very challenging. It covers a wide range of social activities. It is harrowed by ambient sound, room reverberation, overlapping speech and non-verbal sounds such as clapping and instruments. Participants can move", + "bbox": [ + 496, + 840, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "6411", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "freely in the environment. We believe that this data will be useful to the community beyond the NVAS task as it can be used for benchmarking many other problems, including active speaker localization, source separation, and NVS.", + "bbox": [ + 75, + 90, + 468, + 151 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. The SoundSpaces-NVAS Dataset", + "text_level": 1, + "bbox": [ + 76, + 156, + 362, + 172 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this dataset, we synthesize multi-view audio-visual data of two people having conversations in 3D scenes. In total, we construct 1.3K hours of audio-visual data for a total of 1,000 speakers, 120 3D scenes and 200K viewpoints.", + "bbox": [ + 75, + 174, + 468, + 234 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our goal is to construct audio-visual data with strong spatial and acoustic correspondences across multiple viewpoints, meaning that the visual information should indicate what the audio should sound like, e.g., observing speaker on the left should indicate the left ear is louder and observing speaker at a distance should indicate there is higher reverberation. We use the SoundSpaces 2.0 platform [7], which allows highly realistic audio and visual rendering for arbitrary camera and microphone locations in 3D scans of real-world environments [5, 52, 57]. It accounts for all major real-world acoustics phenomena: direct sounds, early specular/diffuse reflections, reverberation, binaural spatialization, and effects from materials and air absorption.", + "bbox": [ + 75, + 234, + 468, + 430 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We use the Gibson dataset [57] for scene meshes and LibriSpeech [41] for speech samples. As we are simulating two people having conversations, for a given environment, we randomly sample two speaker locations within $3\\mathrm{m}$ and insert two copyright-free mannequins (one male and one female) at these two locations. We then randomly sample four nearby viewpoints facing the center of the two speakers at a height of $1.5\\mathrm{m}$ (Fig. 2, right). For each speaker, we select a speech sample from LibriSpeech with matching gender. We render images at all locations as well as binaural impulse response for all pairs of points between speakers and viewpoints. The received sound is obtained by convolving the binaural impulse response with the speech sample.", + "bbox": [ + 75, + 431, + 468, + 626 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "During training, for one sample, we randomly sample two out of four rendered viewpoints, one as the source and one as the target. We also randomly choose one speaker to be active, simulating what we observe on the real data (i.e., usually only one person speaks at a time).", + "bbox": [ + 75, + 628, + 468, + 703 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5. Visually-Guided Acoustic Synthesis", + "text_level": 1, + "bbox": [ + 76, + 715, + 398, + 733 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We introduce a new method, Visually-Guided Acoustic Synthesis (ViGAS), to address the NVAS problem, taking as input sound and an image and outputting the sound from a different target microphone pose.", + "bbox": [ + 75, + 742, + 468, + 801 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ViGAS consists of five components: ambient sound separation, active speaker localization, visual acoustic network, acoustic synthesis, and temporal alignment. The high-level idea is to separate the observed sound into primary and ambient, extract useful visual information (active speaker and", + "bbox": [ + 75, + 803, + 468, + 877 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "acoustic features), and use this information to guide acoustic synthesis for the primary sound. Temporal alignment is performed during training for better optimization. ViGAS is discussed in detail next and summarised in Fig. 3.", + "bbox": [ + 498, + 90, + 890, + 151 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5.1. Ambient Sound Separation", + "text_level": 1, + "bbox": [ + 498, + 162, + 745, + 178 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ViGAS starts by decomposing the input sound into primary and ambient (traffic, electric noise from a fridge or the A/C, etc.). Ambient sound is important for realism, but it also interferes with learning the model because it can carry significant energy, making the model focus on it rather than on the primary sounds, and its spatial distribution is very different from the primary sounds.", + "bbox": [ + 496, + 185, + 890, + 291 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "By explicitly separating primary and ambient sounds, ViGAS: (1) accounts for the fact that the transfer functions of primary and ambient sounds are very different and thus difficult to model together; (2) avoids wasting representational power on modelling ambient sounds that might be difficult to reconstruct accurately and depend less on the viewpoint; and (3) prevents ambient sounds, which are noise-like and high-energy, from dominating learning and reconstruction. In practice, as we show in Sec. 6, without the ambient sound separation, the model performs poorly.", + "bbox": [ + 496, + 292, + 890, + 441 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The goal of ambient sound separation is thus to construct a function $(A_C, A_N) = \\mathcal{P}(A_S)$ that separates the input sound $A_S$ into primary sound $A_C$ and ambient sound $A_N$ . Existing approaches to this problem are based on signal processing [3, 13] or learning [11, 15]. We find that pretrained speech enhancement models such as Denoiser [11] tend to aggressively remove the noise including the primary sound, which hinders re-synthesis. We thus opt for band-pass filtering, passing frequencies within a certain range and rejecting/attenuating frequencies outside of it, which we found to work well. We cut frequencies below $80\\mathrm{Hz}$ for SoundSpaces-NVAS and $150\\mathrm{Hz}$ for Replay-NVAS.", + "bbox": [ + 496, + 443, + 890, + 623 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5.2. Active Speaker Localization", + "text_level": 1, + "bbox": [ + 498, + 633, + 751, + 650 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Knowing where the emitters of different primary sounds are located in the environment can help to solve the NVAS task. In this paper, we focus on localizing the active speaker, although there can be other important primary sound events like instruments playing, speakers interacting with objects, etc. The goal of active speaker localization is to predict the bounding box of the active speaker in each frame of the video (examples in Fig. 4). The bounding box is in the format of $(y_{\\mathrm{min}}, y_{\\mathrm{max}}, x_{\\mathrm{min}}, x_{\\mathrm{max}})$ and $x, y$ are normalized to $[0,1]$ by the image width and height, respectively.", + "bbox": [ + 496, + 657, + 890, + 809 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "On SoundSpaces-NVAS, this task is relatively easy because of the strong correspondence between the appearance of the speaker and the gender of the speech sample, which enables to easily train a classifier for active speakers. However, this is much harder on Replay-NVAS because cameras record speakers from a distance and from diverse angles,", + "bbox": [ + 496, + 810, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "3https://renderpeople.com/free-3d-people", + "bbox": [ + 93, + 886, + 411, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "6412", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/72393b9726dd6915942889143168835f6ee7c48b10da29be1ad64de8c0608488.jpg", + "image_caption": [ + "Figure 3. Visually Guided Acoustic Synthesis (ViGAS). Given the input audio $A_{S}$ , we first separate out the ambient sound to focus on the sound of interest. We take the source audio and source visual to localize the active speaker on the 2D image. We also extract the visual acoustic features of the environment by running an encoder on the source visual. We concatenate the active speaker feature, source visual features, and the target pose, and fuse these features with a MLP. We feed both the audio stream $A_{C}$ and fused visual feature $V_{C}$ into the acoustic synthesis network, which has $M$ stacked audio-visual fusion blocks. In each block, the audio sequence is processed by dilated conv1d layers and the visual features are processed by conv1d layers. Lastly, the previously separated ambient sound is added back to the waveform. During training, our temporal alignment module shifts the prediction by the amount of delay estimated between the source and the target audio to align the prediction well with the target." + ], + "image_footnote": [], + "bbox": [ + 99, + 89, + 872, + 272 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "meaning that lip motion, the main cue used by speaker localization methods [21, 49, 54], is often not visible. Hence, the model has to rely on other cues to identify the speaker (such as body motion, gender or identity). Furthermore, sometimes people speak or laugh over each other.", + "bbox": [ + 75, + 400, + 468, + 474 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since our focus is not speaker localization, for the Replay-NVAS we assume that this problem is solved by an external module that does audio-visual active speaker localization. To approximate the output of such a module automatically, we rely on the near-range audio recordings. Specifically, we first run an off-the-shelf detection and tracker [10] on the video at 5 FPS and obtain, with some manual refinement, bounding boxes $B_{t}^{i}$ for $i = 1,\\dots ,N$ at each frame $t$ . We manually assign the near-range microphone audio $A_N^i$ to each tracked person. We select the active speaker $D$ based on the maximum energy of each near-range microphone, i.e., $D = \\mathrm{argmax}_i\\left\\{\\sum A_N^i [t:t + \\Delta t]^2\\right\\}$ , where $\\Delta t$ is the time interval we use to calculate the audio energy. We output bounding box $B^{D}$ as the localization feature $V_{L}$ .", + "bbox": [ + 75, + 476, + 472, + 704 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.3. Visual Acoustic Network and Fusion", + "text_level": 1, + "bbox": [ + 76, + 710, + 393, + 726 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The active speaker bounding box $B^{D}$ only disambiguates the active speaker from all visible humans on 2D, which is not enough to indicate where the speaker is in 3D. To infer that, the visual information is also needed. Since there is usually not much movement in one second (the length of the input video clip), the video clip does not provide much extra information compared to a single frame. Thus, we choose the middle frame to represent the clip and extract the visual acoustic features $V_{E}$ from the input RGB image with a pretrained ResNet18 [18] before the average pooling layer to preserve spatial information. To reduce the", + "bbox": [ + 75, + 734, + 472, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "feature size, we feed $V_{E}$ into a 1D convolution with kernel size 1 and output channel size 8. We then flatten the visual features to obtain feature $V_{F}$ .", + "bbox": [ + 496, + 400, + 890, + 444 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The target pose is specified as the translation along $x, y, z$ axes plus difference between orientations of the source \"view\" and the target \"view\" expressed via rotation angles: $+y$ (roll), $+x$ (pitch) and $+z$ (yaw). We encode each angle $\\alpha$ as its sinusoidal value: $(\\sin(\\alpha), \\cos(\\alpha))$ .", + "bbox": [ + 496, + 445, + 890, + 520 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Similarly, the target pose is not enough by itself to indicate where the target viewpoint $T$ is in the 3D space; to infer that, the source view $V_{S}$ is again needed. For example, in top row of Fig 4, for target viewpoint 3, \"two meters to the right and one meter forward\" is not enough to indicate the target location is in the corridor, while the model can reason that based on the source view.", + "bbox": [ + 496, + 520, + 890, + 626 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use a fusion network to predict a latent representation of the scene variables $S, T, L^{D}, E$ (cf. Sec. 3) by first concatenating $[V_{L}, P_{T}, V_{F}]$ and then feeding it through a multilayer perceptron (MLP). See Fig. 3 for the network.", + "bbox": [ + 496, + 626, + 890, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.4. Acoustic Synthesis", + "text_level": 1, + "bbox": [ + 500, + 696, + 679, + 712 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With the separated primary sound $A_{C}$ and the visual acoustic feature $V_{C}$ as input, the goal of the acoustic synthesis module is to transform $A_{C}$ guided by $V_{C}$ . We design the acoustic synthesis network to learn a non-linear transfer function (implicitly) that captures these major acoustic phenomena, including the attenuation of sound in space, the directivity of sound sources (human speech is directional), the reverberation level, the head-related transfer function, as well as the frequency-dependent acoustic phenomena. Training end-to-end makes it possible to capture these subtle and complicated changes in the audio.", + "bbox": [ + 496, + 719, + 890, + 883 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inspired by recent advances in time-domain signal mod", + "bbox": [ + 517, + 885, + 890, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "6413", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "eling [39, 48], we design the network as $M$ stacked synthesis blocks, where each block consists of multiple conv1D layers. We first encode the input audio $A_{C}$ into a latent space, which is then fed into the synthesis block. The key of the synthesis block is a gated multimodal fusion network that injects the visual information into the audio as follows:", + "bbox": [ + 75, + 90, + 468, + 181 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nz = \\tanh \\left(p _ {A} ^ {k} \\left(A _ {F} ^ {k}\\right) + p _ {V} ^ {k} \\left(V _ {C}\\right)\\right) \\odot \\sigma \\left(q _ {A} ^ {k} \\left(A _ {F} ^ {k}\\right) + q _ {V} ^ {k} \\left(V _ {C}\\right)\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 186, + 468, + 205 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\odot$ indicates element-wise multiplication, $\\sigma$ is a logistic sigmoid function, $k = 1,\\dots ,M$ is the layer index and $p,q$ are both learnable 1D convolutions.", + "bbox": [ + 75, + 213, + 468, + 258 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "After passing $z$ through a sinusoidal activation function, the network uses two separate conv1D layers to process the feature, one producing the residual connection $A_F^{k + 1}$ and one producing the skip connection $A_P^{k + 1}$ . All skip connections $A_P^{k + 1}$ are mean pooled and fed into a decoder to produce the output $A_O$ . We add back the separated ambient sound $A_N$ as the target audio estimate: $\\hat{A}_T = A_O + A_N$ . See Supp. for more details on the architecture.", + "bbox": [ + 75, + 258, + 468, + 382 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.5. Temporal Alignment", + "text_level": 1, + "bbox": [ + 76, + 390, + 272, + 406 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In order for the model to learn well, it is important that input and output sounds are temporally aligned. While the Replay-NVAS data is already synchronised based on the clapper sound, due to the finite speed of sound, the sounds emitted from different locations may still arrive at microphones with a delay slightly different from the one of the clapper, causing misalignments that affect training.", + "bbox": [ + 75, + 414, + 468, + 518 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To align source and target audio for training, we find the delay $\\tau$ that maximizes the generalized cross-correlation:", + "bbox": [ + 75, + 518, + 468, + 549 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {R} _ {A _ {S}, A _ {T}} (\\tau) = \\mathbb {E} _ {t} [ h _ {S} (t) \\cdot h _ {T} (t - \\tau) ], \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 147, + 550, + 468, + 566 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $h_S$ and $h_T$ are the feature embedding for $A_S$ and $A_T$ respectively at time $t$ . We use the feature extractor $h$ from the generalized cross-correlation phase transform (GCC-PHAT) algorithm [23], which whitens the audio by dividing by the magnitude of the cross-power spectral density. After computing $\\tau$ , we shift the prediction $A_O$ by $\\tau$ samples to align with the $A_T$ and obtain $A_L$ . Note that alignment is already exact for SoundSpaces-NVAS.", + "bbox": [ + 75, + 571, + 468, + 691 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.6. Loss", + "text_level": 1, + "bbox": [ + 76, + 700, + 148, + 714 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To compute the loss, we first encode the audio with the short-time Fourier transform (STFT), a complex-valued matrix representation of the audio where the $y$ axis represents frequency and the $x$ axis is time. We then compute the magnitude of the STFT, and optimize the L1 loss between the predicted and ground truth magnitudes as follows:", + "bbox": [ + 75, + 723, + 468, + 813 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL = \\left| \\left| \\left| S T F T \\left(A _ {L}\\right) \\right| \\right| _ {2} - \\left| \\left| S T F T \\left(A _ {T} ^ {\\prime}\\right) \\right| \\right| _ {2} \\right|, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 818, + 468, + 835 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $A_T^\\prime$ is the primary sound separated from $A_{T}$ with $\\mathcal{P}(\\cdot)$ . By taking the magnitude, we do not model the exact phase values, which we find hinders learning if being included in the loss. See implementation details in Supp.", + "bbox": [ + 75, + 839, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6. Experiments", + "text_level": 1, + "bbox": [ + 500, + 90, + 633, + 107 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare with several traditional and learning-based baselines and show that ViGAS outperforms them in both a quantitative evaluation and a human subject study.", + "bbox": [ + 496, + 114, + 890, + 161 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation. We measure performance from three aspects: 1. closeness to GT as measured by the magnitude spectrogram distance (Mag). 2. correctness of the spatial sound as measured by the left-right energy ratio error (LRE), i.e., the difference of ratio of energy between left and right channels and 3. correctness of the acoustic properties measured by RT60 error (RTE) [6, 50], i.e., the error in reverberation time decaying by 60dB (RT60). We use a pretrained model [6] to estimate RT60 directly from speech.", + "bbox": [ + 496, + 165, + 890, + 301 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. We consider the following baselines: 1. Input audio. Copying the input to the output. 2. TF Estimator [56] + Nearest Neighbor, i.e. storing the transfer function estimated during training and retrieving the nearest neighbor during test time. We estimate transfer functions with a Wiener filter [56] and index them with the ground-truth locations of the speaker, source viewpoint, and target viewpoint for the single environment setup and their relative pose for the novel environment setup. At test time, this method searches the database to find the nearest transfer function and applies it on the input audio. 3. Digital Signal Processing (DSP) [9] approach that takes the distance, azimuth, and elevation of the sound source, applies an inverse a head-related transfer function (HRTF) to estimate the speech spoken by the speaker and then applies another HRTF to estimate the audio at the target microphone location. This baseline adjusts the loudness of the left and right channels based on where the speaker is in the target view. We supply GT coordinates for SoundSpaces-NVAS and speakers' head positions estimated with triangulation on Replay-NVAS. 4. Visual Acoustic Matching (VAM) [6], recently proposed for a related task of matching acoustics of input audio with a target image. This task only deals with single viewpoint and single-channel audio. We adapt their model with minimal modification by feeding in the image from the source viewpoint and concatenating the position offset of the target microphone at the multimodal fusion step. See Supp. for details.", + "bbox": [ + 496, + 304, + 892, + 729 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.1. Results on SoundSpaces-NVAS", + "text_level": 1, + "bbox": [ + 500, + 739, + 772, + 755 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1 shows the results. For synthetic data, we consider two evaluation setups: 1. single environment: train and test on the same environment and 2. novel environment: train and test on multiple non-overlapping Gibson environments (90/10/20 for train/val/test).", + "bbox": [ + 496, + 763, + 890, + 838 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the single environment setup, our model largely outperforms all baselines as well as our audio-only ablation on all metrics. TF Estimator performs poorly despite being indexed by the ground truth location values because es", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6414", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/9d9fadd1497dc629a2dc0b66f7a6b5ad60c8c4a757a51ad1d010d46044987d73.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SoundSpaces-NVASReplay-NVAS
Single EnvironmentNovel EnvironmentSingle Environment
MagLRERTEMagLRERTEMagLRERTE
Input audio0.2251.4730.0320.2161.4080.0390.1591.4770.046
TF Estimator [56]0.3592.5960.0590.4403.2610.0920.3272.8610.147
DSP [9]0.3023.6440.0440.3003.6890.0470.4631.3000.067
VAM [6]0.2201.1980.0410.2351.1310.0510.1610.9240.070
ViGAS w/o visual0.1730.9730.0310.1811.0070.0360.1460.8770.046
ViGAS0.1590.7820.0290.1750.9710.0340.1420.7160.048
", + "bbox": [ + 122, + 88, + 843, + 246 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "timating a transfer function directly from two audio clips is non-trivial and noisy for low-energy parts of the signal. DSP also performs badly despite having the ground truth 3D coordinates of the sound source. This is because head related transfer functions are typically recorded in anechoic chambers, which does not account for acoustics of different environments, e.g., reverberation. Both traditional approaches perform worse than simply copying the input audio, indicating that learning-based models are needed for this challenging task. The recent model VAM [6] performs much better compared to the traditional approaches but still underperforms our model. There is a significant difference between ViGAS w/o visual and the full model; this shows that the visual knowledge about the speaker location and the environment is important for this task.", + "bbox": [ + 75, + 340, + 470, + 566 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fig. 4 shows an example where given the same input source viewpoint, our model synthesizes audio for three different target viewpoints. The model reasons about how the geometry and speaker locations changes based on the source view and the target pose, and predicts the acoustic difference accordingly. See Supp. video to listen to sounds.", + "bbox": [ + 75, + 571, + 468, + 662 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For the novel environment setup, our model again outperforms all baselines. Compared to ViGAS in the single environment setup, both the magnitude spectrogram distance and the left-right energy ratio error increase. This is expected because for novel (unseen) environments, single images capture limited geometry and acoustic information. The model fails sometime when there is a drastic viewpoint change, e.g., target viewpoint 3 in Fig. 4. This setup requires the model to reason or \"imagine\" the environment based on single audio-visual observation, which poses great challenge for NVAS as well as NVS, where typically synthesis is performed in a fully observed environment.", + "bbox": [ + 75, + 666, + 470, + 848 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablations. Table 2 shows ablations on the model design. To understand if the model uses visual information, we ablate the visual features $V_{F}$ and the active speaker feature", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3e022f4da14aecd133513384353c6e597a6140b228cb7718ebaf13232534fdba.jpg", + "table_caption": [ + "Table 1. Results on SoundSpaces-NVAS and Replay-NVAS. We report the magnitude spectrogram distance (Mag), left-right energy ratio error (LRE), and RT60 error (RTE). Replay-NVAS does not have novel environment setup due to data being collected in a single environment. For all metrics, lower is better. In addition to baselines, we also evaluate ViGAS w/o visual by removing the active speaker localization and visual features. Note that reverberation time is mostly invariant of the receiver location in the same room and thus input audio has low RTE. A good model should preserve this property while synthesizing the desired acoustics for the target viewpoint." + ], + "table_footnote": [], + "table_body": "
ViGASSS-NVASReplay-NVAS
MagLREMagLRE
full model0.1590.7820.1420.716
w/o visual features0.1710.8970.1460.920
w/o ASL0.1610.8140.1430.757
w/o alignment0.1760.7710.1440.706
w/o separation0.1650.8400.1820.859
", + "bbox": [ + 506, + 338, + 883, + 460 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Ablations of the model on both datasets.", + "bbox": [ + 540, + 465, + 846, + 479 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "$V_{L}$ . Removing the active speaker feature leads to less damage on the model performance, because without the explicitly localized active speaker, the model can still implicitly reason about the active speaker location based on the image and audio. If both are removed (\"ViGAS w/o visual\" in Table 1), the performance suffers most.", + "bbox": [ + 498, + 486, + 890, + 575 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To study the effectiveness of the temporal alignment and ambient sound separation modules, we ablate them separately. Removing the temporal alignment leads to higher Mag error and slightly lower LRE. As for ambient sound separation, the results show that optimizing for the high-energy noise-like ambient sound degrades the performance.", + "bbox": [ + 498, + 577, + 890, + 667 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.2. Results on Replay-NVAS", + "text_level": 1, + "bbox": [ + 500, + 680, + 725, + 696 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1 (right) shows the Replay-NVAS results. Compared to SoundSpaces-NVAS, the magnitudes of all errors are smaller because there are less drastic acoustic changes between viewpoints (8 DLSR cameras form a circle around the participants). Traditional approaches like TF Estimator and DSP still perform poorly despite using the 3D coordinates of the camera and the speaker (triangulated from multiple cameras). VAM performs better due to end-to-end learning; however, our model outperforms it. Compared to ViGAS w/o visual, the full model has much lower left-right energy ratio error and slightly higher reverberation time error, showing that the model takes into account the speaker position and viewpoint change for synthesizing the audio.", + "bbox": [ + 498, + 703, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "6415", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bba20e54029ad3b75f1ad6e549e9eaead118c8bcbb197e3a42d4641063558db8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 88, + 284, + 198 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2a3bc9b8f482259d3bc533a3c7fe27659926713b905a5cd64fa54a22e9625ed5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 285, + 88, + 486, + 198 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/dbdcf8565c1173fa0eefc44219c9345796f5258ee97785a72dac4cae43e3738b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 488, + 88, + 687, + 198 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/244a341a624ed1e79c438f2cd502663b30e347453447777ce98071dcbc56d713.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 691, + 88, + 888, + 198 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e3e5bb3bea372403391cf8fc8e1cf880e10e35f98d61924e445ea189442d3301.jpg", + "image_caption": [ + "Figure 4. Qualitative examples. For all binaural audio, we show the left-channel and the right-channel waveforms side-by-side. Row 1: SoundSpaces-NVAS example where given the source viewpoint and input audio, the model synthesizes audio for three different target viewpoints (target views are for reference only). In this case, the active speaker is the male speaker as indicated by the bounding box. For target viewpoint 1, the view rotates about 90 degrees and the male speaker is on the left side and the predicted left channel is louder than the right channel. Viewpoint 2 moves away from the speaker and thus yields lower amplitude compared to the first prediction. For target viewpoint 3, it is completely located outside of the living room, in which case, the sound could only come from the door open on the right (louder right channel) and the reverberation also greatly increases due to the vanishing direct sound. Row 2: Replay-NVAS example where the speaker is located on the left in the source viewpoint which becomes the right and further from the camera in target viewpoint 2, the model also predicts lower amplitude and louder right channel. On the right side, we show an example of the audio-visual speech enhancement for the active speaker. The model enhances the speech to largely match with the near-range audio (target)." + ], + "image_footnote": [], + "bbox": [ + 80, + 199, + 343, + 297 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c600a1a81ad1afc4eaa470f46a004ce4c3f4ebf6124aabeb9a5b3f381b4d0796.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 344, + 199, + 602, + 297 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fdec0cd93464eb57ac3729b5a726b7c8bff74bdfa670c61ebe913fefb85e8165.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 607, + 199, + 887, + 297 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e5b95849ec52eecae0abbc709e477312eb7a0acacc4786c60e3d7f3063c78820.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MagRTE
Input0.2790.376
ViGAS (ours)0.2340.122
", + "bbox": [ + 137, + 457, + 403, + 518 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fig. 4 (row 2, left) shows a qualitative example. In the source viewpoint, the active speaker is on the left, while in the target viewpoint, he is further from the camera and on the right. The model synthesizes an audio waveform that captures the corresponding acoustic change, showing that our model successfully learns from real videos.", + "bbox": [ + 75, + 546, + 468, + 636 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Audio-visual speech enhancement. In some real-world applications, e.g., hearing aid devices, the goal is to obtain the enhanced clean speech of the active speaker. This can be seen as a special case of NVAS, where the target viewpoint is the active speaker. Our model is capable of performing audio-visual speech enhancement without any modification. We simply set the target audio to the near-range audio recording for the active speaker. We show the results in Table 3. Our model obtains cleaner audio compared to the input audio (example in Fig. 4, row 2, right).", + "bbox": [ + 75, + 640, + 468, + 791 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Human subject study. To supplement the quantitative metrics and evaluate how well our synthesized audio captures the acoustic change between viewpoints, we conduct a human subject study. We show participants the image of the target viewpoint $V_{T}$ as well as the audio $A_{T}$ as reference. We provide three audio samples: the input, the prediction of ViGAS, and the prediction of DSP (the most nat", + "bbox": [ + 75, + 795, + 470, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/36dcf95921f1d5f95ad1f7b16addc02b387400ddce88a38c8c616cddc3450b68.jpg", + "table_caption": [ + "Table 3. Speech enhancement on Replay-NVAS." + ], + "table_footnote": [], + "table_body": "
DatasetInputDSPViGAS
SoundSpaces-NVAS24%2%74%
Replay-NVAS43%6%51%
", + "bbox": [ + 527, + 457, + 859, + 518 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Human Study. Participants favor our approach over the two most realistic sounding baselines, (1) copying the input signal, and (2) a digital signal processing baseline.", + "bbox": [ + 496, + 526, + 890, + 568 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "urally sounding baseline) and ask them to select a clip that sounds closest to the target audio. We select 20 examples from SoundSpaces-NVAS and 20 examples from Replay-NVAS and invite 10 participants to perform the study.", + "bbox": [ + 496, + 575, + 890, + 635 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "See Table 4 for the results. On the synthetic dataset SoundSpaces-NVAS, our approach is preferred over the baselines by a large margin. This margin is lower on the real-world Replay-NVAS dataset but is still significant.", + "bbox": [ + 496, + 636, + 890, + 696 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 709, + 617, + 724 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduce the challenging novel-view acoustic synthesis task and a related benchmark in form of both real and synthetic datasets. We propose a neural rendering model that learns to transform the sound from the source viewpoint to the target viewpoint by reasoning about the observed audio and visual stream. Our model surpasses all baselines on both datasets. We believe this research unlocks many potential applications and research in multimodal novel-view synthesis. In the future, we plan to incorporate active-speaker localization model into the approach and let the model jointly learn to localize and synthesize.", + "bbox": [ + 496, + 734, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "6416", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Triantafyllos Afouras, Joon Son Chung, and Andrew Zisserman. The conversation: Deep audio-visual speech enhancement. In INTERSPEECH, 2018. 3", + "[2] Humam Alwassel, Dhruv Mahajan, Bruno Korbar, Lorenzo Torresani, Bernard Ghanem, and Du Tran. Self-supervised learning by cross-modal audio-video clustering. In NeurIPS, 2020. 3", + "[3] M. Berouti, Richard Schwartz, and John Makhoul. Enhancement of speech corrupted by acoustic noise. In IEEE International Conference on Acoustics, Speech, and Signal Processing, 1979. 4", + "[4] J.S. Bradley. Review of objective room acoustics measures and future needs. Applied Acoustics, 2011. 2", + "[5] Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niessner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from rgb-d data in indoor environments. 3DV, 2017. MatterPort3D dataset license available at: http://kaldir.vc.in.tum.de/matterport/MP_TOS.pdf.4", + "[6] Changan Chen, Ruohan Gao, Paul Calamia, and Kristen Grauman. Visual acoustic matching. In CVPR, 2022. 2, 6, 7", + "[7] Changan Chen, Carl Schissler, Sanchit Garg, Philip Kobernik, Alexander Clegg, Paul Calamia, Dhruv Batra, Philip W Robinson, and Kristen Grauman. Soundspaces 2.0: A simulation platform for visual-acoustic learning. In NeurIPS 2022 Datasets and Benchmarks Track, 2022. 2, 4", + "[8] Changan Chen, Wei Sun, David Harwath, and Kristen Grauman. Learning audio-visual dereverberation. In ICASSP, 2023. 2", + "[9] Corey I. Cheng and Gregory H. Wakefield. Introduction to head-related transfer functions (hrtfs): representations of hrtfs in time, frequency, and space. journal of the audio engineering society, 49(4):231-249, april 2001. 6, 7", + "[10] MMTracking Contributors. MMTracking: OpenMMLab video perception toolbox and benchmark. https://github.com/open-mmlab/mmtracking, 2020.5", + "[11] Alexandre Defossez, Gabriel Synnaeve, and Yossi Adi. Real time speech enhancement in the waveform domain. In *Interspeech*, 2020. 4", + "[12] James Eaton, Nikolay Gaubitch, Allistair Moore, and Patrick Naylor. Estimation of room acoustic parameters: The ACE challenge. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 24(10), 2016. 2", + "[13] Yariv Ephraim and Harry L. Van Trees. A signal subspace approach for speech enhancement. IEEE Transactions on Speech and Audio Processing, 1995. 4", + "[14] Ariel Ephrat, Inbar Mosseri, Oran Lang, Tali Dekel, Kevin Wilson, Avinatan Hassidim, William T Freeman, and Michael Rubinstein. Looking to listen at the cocktail party: A speaker-independent audio-visual model for speech separation. In SIGGRAPH, 2018. 3", + "[15] Szu-Wei Fu, Chien-Feng Liao, Yu Tsao, and Shou-De Lin. Metricgan: Generative adversarial networks based black-box metric scores optimization for speech enhancement. In International Conference on Machine Learning (ICML), 2019. 4" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Hannes Gamper and Ivan J Tashev. Blind reverberation time estimation using a convolutional neural network. In 2018 16th International Workshop on Acoustic Signal Enhancement (IWAENC), pages 136-140, 2018. 2", + "[17] Ruohan Gao and Kristen Grauman. 2.5d visual sound. In CVPR, 2019. 2", + "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 5", + "[19] Di Hu, Rui Qian, Minyue Jiang, Xiao Tan, Shilei Wen, Errui Ding, Weiyao Lin, and Dejing Dou. Discriminative sounding objects localization via self-supervised audiovisual matching. In NeurIPS, 2020. 3", + "[20] Ajay Jain, Matthew Tancik, and Pieter Abbeel. Putting nerf on a diet: Semantically consistent few-shot view synthesis. In Proc. ICCV, 2021. 2", + "[21] Hao Jiang, Calvin Murdock, and Vamsi Krishna Ithapu. Ego-centric deep multi-channel audio-visual active speaker localization. In CVPR, 2022. 3, 5", + "[22] Florian Klein, Annika Neidhardt, and Marius Seipel. Real-time estimation of reverberation time for selection of suitable binaural room impulse responses. In Audio for Virtual, Augmented and Mixed Realities: Proceedings of 5th International Conference on Spatial Audio (ICSA), pages 145–150, 2019. 2", + "[23] C. Knapp and G. Carter. The generalized correlation method for estimation of time delay. IEEE Transactions on Acoustics, Speech, and Signal Processing, 1976. 6", + "[24] Bruno Korbar, Du Tran, and Lorenzo Torresani. Cooperative learning of audio and video models from self-supervised synchronization. In NeurIPS, 2018. 3", + "[25] Jonas Kulhanek, Erik Derner, Torsten Sattler, and Robert Babuška. ViewFormer: NeRF-free neural rendering from few images using transformers. In Proc. ECCV, 2022. 2", + "[26] Heinrich Kuttruff. Room Acoustics. Boca Raton, 6th edition, 2016. 2", + "[27] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. arXiv.cs, abs/2011.13084, 2020. 2", + "[28] Lingjie Liu, Marc Habermann, Viktor Rudnev, Kripasindhu Sarkar, Jiatao Gu, and Christian Theobalt. Neural humans: Pose-controlled free-view synthesis of human actors with template-guided neural radiance fields. In arXiv, 2021. 2", + "[29] Andrew Luo, Yilun Du, Michael J Tarr, Joshua B Tenenbaum, Antonio Torralba, and Chuang Gan. Learning neural acoustic fields. In NeurIPS, 2022. 3", + "[30] Wolfgang Mack, Shuwen Deng, and Emanuele AP Habets. Single-channel blind direct-to-reverberation ratio estimation using masking. In INTERSPEECH, pages 5066-5070, 2020. 2", + "[31] Sagnik Majumder, Changan Chen, Ziad Al-Halah, and Kristen Grauman. Few-shot audio-visual learning of environment acoustics. In Thirty-Sixth Conference on Neural Information Processing Systems, 2022. 3", + "[32] Ricardo Martin-Brualla, Noha Radwan, Mehdi S. M. Sajjadi, Jonathan T. Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In Proceedings of the IEEE/CVF" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "6417", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Conference on Computer Vision and Pattern Recognition (CVPR), pages 7210-7219, June 2021. 1", + "[33] Daniel Michelsanti, Zheng-Hua Tan, Shi-Xiong Zhang, Yong Xu, Meng Yu, Dong Yu, and Jesper Jensen. An overview of deep-learning-based audio-visual speech enhancement and separation. In arXiv, 2020. 3", + "[34] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2", + "[35] Pedro Morgado, Yi Li, and Nuno Vasconcelos. Learning representations from audio-visual spatial alignment. In NeurIPS, 2020. 3", + "[36] Pedro Morgado, Nono Vasconcelos, Timothy Langlois, and Oliver Wang. Self-supervised generation of spatial audio for 360 video. In NeurIPS, 2018. 2", + "[37] Prateek Murgai, Mark Rau, and Jean-Marc Jot. Blind estimation of the reverberation fingerprint of unknown acoustic environments. In Audio Engineering Society Convention 143. Audio Engineering Society, 2017. 2", + "[38] Michael Niemeyer, Jonathan T. Barron, Ben Mildenhall, Mehdi S. M. Sajjadi, Andreas Geiger, and Noha Radwan. RegNeRF: Regularizing neural radiance fields for view synthesis from sparse inputs. In Proc. CVPR, 2022. 2", + "[39] Aaron van den Oord, Sander Dieleman, Heiga Zen, Karen Simonyan, Oriol Vinyls, Alex Graves, Nal Kalchbrenner, Andrew Senior, and Koray Kavukcuoglu. Wavenet: A generative model for raw audio, 2016. cite arxiv:1609.03499. 6", + "[40] Andrew Owens and Alexei A Efros. Audio-visual scene analysis with self-supervised multisensory features. In ECCV, 2018. 3", + "[41] Vassil Panayotov, Guoguo Chen, Daniel Povey, and Sanjeev Khudanpur. Librispeech: an asr corpus based on public domain audio books. In 2015 IEEE international conference on acoustics, speech and signal processing (ICASSP), pages 5206-5210. IEEE, 2015. 4", + "[42] Keunhong Park, Utkarsh Sinha, Jonathan T. Barron, Sofien Bouaziz, Dan B. Goldman, Steven M. Seitz, and Ricardo Martin-Brualla. Deformable neural radiance fields. CoRR, abs/2011.12948, 2020. 2", + "[43] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-NeRF: Neural radiance fields for dynamic scenes. arXiv.cs, abs/2011.13961, 2020. 2", + "[44] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10318–10327, June 2021. 1", + "[45] Anton Ratnarajah, Zhenyu Tang, Rohith Chandrashekar Aralikatti, and Dinesh Manocha. Mesh2ir: Neural acoustic impulse response generator for complex 3d scenes. In ACM Multimedia, 2022. 3", + "[46] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common Objects in 3D: Large-scale learning and evaluation of real-life 3D category reconstruction. In Proc. CVPR, 2021. 2" + ], + "bbox": [ + 78, + 92, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[47] Alexander Richard, Peter Dodds, and Vamsi Krishna Ithapu. Deep impulse responses: Estimating and parameterizing filters with deep networks. In IEEE International Conference on Acoustics, Speech and Signal Processing, 2022. 2", + "[48] Alexander Richard, Dejan Markovic, Israel D Gebru, Steven Krenn, Gladstone Butler, Fernando de la Torre, and Yaser Sheikh. Neural synthesis of binaural speech from mono audio. In International Conference on Learning Representations, 2021. 2, 6", + "[49] Joseph Roth, Sourish Chaudhuri, Ondrej Klejch, Radhika Marvin, Andrew Gallagher, Liat Kaver, Sharadh Ramaswamy, Arkadiusz Stopczynski, Cordelia Schmid, Zhonghua Xi, and Caroline Pantofaru. Ava-activespeaker: An audio-visual dataset for active speaker detection. In ICASSP, 2020. 5", + "[50] Nikhil Singh, Jeff Mentch, Jerry Ng, Matthew Beveridge, and Iddo Drori. Image2reverb: Cross-modal reverb impulse response synthesis. In ICCV, 2021. 2, 6", + "[51] Vincent Sitzmann, Michael Zollhöfer, and Gordon Wetzstein. Scene representation networks: Continuous 3D-structure-aware neural scene representations. Proc. NeurIPS, 2019. 2", + "[52] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, et al. The Replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019.4", + "[53] Shih-Yang Su, Frank Yu, Michael Zollhöfer, and Helge Rhodin. A-NeRF: Surface-free human 3d pose refinement via neural rendering. arXiv.cs, abs/2102.06199, 2021. 2", + "[54] Ruijie Tao, Zexu Pan, Rohan Kumar Das, Xinyuan Qian, Mike Zheng Shou, and Haizhou Li. Is someone speaking? exploring long-term temporal features for audio-visual active speaker detection. In Proceedings of the 29th ACM International Conference on Multimedia, page 3927–3935, 2021. 3, 5", + "[55] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Christoph Lassner, and Christian Theobalt. Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a deforming scene from monocular video. arXiv.cs, abs/2012.12247, 2020. 2", + "[56] Norbert Wiener. Extrapolation, interpolation, and smoothing of stationary time series. Report of the Services 19, Research Project DIC-6037 MIT, 1942. 6, 7", + "[57] Fei Xia, Amir R. Zamir, Zhi-Yang He, Alexander Sax, Jitendra Malik, and Silvio Savarese. Gibson Env: real-world perception for embodied agents. In Computer Vision and Pattern Recognition (CVPR), 2018 IEEE Conference on. IEEE, 2018. 4", + "[58] Feifei Xiong, Stefan Goetze, Birger Kollmeier, and Bernd T Meyer. Joint estimation of reverberation time and early-to-late reverberation ratio from single-channel speech signals. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 27(2):255-267, 2018. 2", + "[59] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. PixelNeRF: Neural radiance fields from one or few images. In Proc. CVPR, 2021. 2" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "6418", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "ref_text", + "text": "[60] Hang Zhou, Ziwei Liu, Xudong Xu, Ping Luo, and Xiaogang Wang. Vision-infused deep audio inpainting. In ICCV, 2019. 3", + "bbox": [ + 78, + 90, + 468, + 132 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "6419", + "bbox": [ + 482, + 945, + 516, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Novel-View Acoustic Synthesis/2dea3110-7f39-4b1f-9a6f-a10e0ea0bd92_model.json b/2023/Novel-View Acoustic Synthesis/2dea3110-7f39-4b1f-9a6f-a10e0ea0bd92_model.json new file mode 100644 index 0000000000000000000000000000000000000000..40f5c5d13a6757103ec92e065e343bcc6484b868 --- /dev/null +++ b/2023/Novel-View Acoustic Synthesis/2dea3110-7f39-4b1f-9a6f-a10e0ea0bd92_model.json @@ -0,0 +1,2279 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.329, + 0.131, + 0.642, + 0.154 + ], + "angle": 0, + "content": "Novel-View Acoustic Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.18, + 0.834, + 0.24 + ], + "angle": 0, + "content": "Changan Chen\\(^{1,3}\\) Alexander Richard\\(^{2}\\) Roman Shapovalov\\(^{3}\\) Vamsi Krishna Ithapu\\(^{2}\\) Natalia Neverova\\(^{3}\\) Kristen Grauman\\(^{1,3}\\) Andrea Vedaldi\\(^{3}\\) \n\\(^{1}\\)University of Texas at Austin \\(^{2}\\)Reality Labs Research at Meta \\(^{3}\\)FAIR, Meta AI" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.273, + 0.314, + 0.289 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.304, + 0.474, + 0.562 + ], + "angle": 0, + "content": "We introduce the novel-view acoustic synthesis (NVAS) task: given the sight and sound observed at a source viewpoint, can we synthesize the sound of that scene from an unseen target viewpoint? We propose a neural rendering approach: Visually-Guided Acoustic Synthesis (ViGAS) network that learns to synthesize the sound of an arbitrary point in space by analyzing the input audio-visual cues. To benchmark this task, we collect two first-of-their-kind large-scale multi-view audio-visual datasets, one synthetic and one real. We show that our model successfully reasons about the spatial cues and synthesizes faithful audio on both datasets. To our knowledge, this work represents the very first formulation, dataset, and approach to solve the novel-view acoustic synthesis task, which has exciting potential applications ranging from AR/VR to art and design. Unlocked by this work, we believe that the future of novel-view synthesis is in multi-modal learning from videos." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.585, + 0.21, + 0.601 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.606, + 0.47, + 0.712 + ], + "angle": 0, + "content": "Replaying a video recording from a new viewpoint1 has many applications in cinematography, video enhancement, and virtual reality. For example, it can be used to edit a video, simulate a virtual camera, or, given a video of a personal memory, even enable users to experience a treasured moment again—not just on a 2D screen, but in 3D in a virtual or augmented reality, thus ‘reliving’ the moment." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.712, + 0.47, + 0.848 + ], + "angle": 0, + "content": "While the applications are exciting, there are still many unsolved technical challenges. Recent advances in 3D reconstruction and novel-view synthesis (NVS) address the problem of synthesizing new images of a given scene [32, 34, 44]. However, thus far, the view synthesis problem is concerned with creating visuals alone; the output is silent or at best naively adopts the sounds of the original video (from the \"wrong\" viewpoint). Without sound, the emotional and cognitive significance of the replay is severely diminished." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.849, + 0.47, + 0.879 + ], + "angle": 0, + "content": "In this work, we address this gap and introduce the new task of novel-view acoustic synthesis (NVAS). The goal of" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.273, + 0.891, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.461, + 0.893, + 0.532 + ], + "angle": 0, + "content": "Figure 1. Novel-view acoustic synthesis task. Given audio-visual observations from one viewpoint and the relative target viewpoint pose, render the sound received at the target viewpoint. Note that the target is expressed as the desired pose of the microphones; the image at that pose (right) is neither observed nor synthesized." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.552, + 0.893, + 0.598 + ], + "angle": 0, + "content": "this task is to synthesize the sound in a scene from a new acoustic viewpoint, given only the visual and acoustic input from another source viewpoint in the same scene (Fig. 1)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.895, + 0.903 + ], + "angle": 0, + "content": "NVAS is very different from the existing NVS task, where the goal is to reconstruct images instead of sounds, and these differences present new challenges. First, the 3D geometry of most real-life scenes changes in a limited manner during the recording. On the contrary, sound changes substantially over time, so the reconstruction target is highly dynamic. Secondly, visual and audio sensors are very different. A camera matrix captures the light in a highly-directional manner, and a single image comprises a large 2D array of pixels. In contrast, sounds are recorded with one or two microphones which are at best weakly-directional, providing only a coarse sampling of the sound field. Thirdly, the frequency of light waves is much higher than that of sound waves; the length of audio waves is thus larger to the point of being comparable to the size of geometric features of the scene, meaning that effects such as diffraction are often dominant, and spatial resolution is low. As a result, techniques that require spatial precision, such as triangulation and segmentation, are not applicable to audio. Lastly, sounds mix together, making it difficult to segment them," + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.416, + 0.901 + ], + "angle": 0, + "content": "1We use \"viewpoint\" to mean a camera or microphone pose." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6409" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.122 + ], + "angle": 0, + "content": "and they are affected by environmental effects such as reverberation that are distributed and largely unobservable." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.125, + 0.471, + 0.366 + ], + "angle": 0, + "content": "While the NVS and NVAS tasks are indeed very different, we hypothesize that NVAS is an inherently multimodal task. In fact, vision can play an important role in achieving accurate sound synthesis. First, establishing correspondences between sounds and their sources as they appear in images can provide essential cues for resynthesizing the sounds realistically. For instance, human speech is highly directional and sounds very differently if one faces the speaker or their back, which can only be inferred from visual cues. In addition, the environment acoustics also affect the sound one hears as a function of the scene geometry, materials, and emitter/receiver locations. The same source sounds vary differently if it is located in the center of a room, at the corner, or in a corridor, for example. In short, vision provides cues about space and geometry that affect sound, and are difficult to estimate from the sound alone." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.37, + 0.47, + 0.55 + ], + "angle": 0, + "content": "In order to validate our hypothesis, we propose a novel visually-guided acoustic synthesis network that analyzes audio and visual features and synthesizes the audio at a target location. More specifically, the network first takes as input the image observed at the source viewpoint in order to infer global acoustic and geometric properties of the environment along with the bounding box of the active speaker. The network then reasons how the speaker and scene geometry change in 3D based on the relative target pose with a fusion network. We inject the fused features into audio with a gated multi-modal fusion network and model the acoustic changes between viewpoints with a time-domain model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.554, + 0.471, + 0.902 + ], + "angle": 0, + "content": "In order to conduct our experiments on the new NVAS task, we require suitable training and benchmarking data, of which currently there is none available. To address that, we contribute two new datasets: one real (Replay-NVAS) and one synthetic (SoundSpaces-NVAS). The key feature of these datasets is to record the sight and sound of different scenes from multiple cameras/viewpoints. Replay-NVAS contains video recordings of groups of people performing social activities (e.g., chatting, watching TV, doing yoga, playing instruments) from 8 surrounding viewpoints simultaneously. It contains 37 hours of highly realistic everyday conversation and social interactions in one homelike environment. To our knowledge, Replay-NVAS represents the first large-scale real-world dataset enabling NVAS. This dataset would also greatly benefit many other existing tasks including NVS, active speaker localization, etc. For SoundSpaces-NVAS, we render 1.3K hours of audiovisual data based on the SoundSpaces [7] platform. Using this simulator, one can easily change the scene geometry and the positions of speakers, cameras, and microphones. This data serves as a powerful test bed with clean ground truth for a large collection of home environments, offering a good complement to Replay-NVAS. For both datasets," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "we capture binaural audio, which is what humans perceive with two ears. Together the datasets contain 1,337 hours of audio-visual capture, with 1,032 speakers across 121 3D scenes. Datasets are publicly available for future research." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.153, + 0.893, + 0.304 + ], + "angle": 0, + "content": "We show that our model outperforms traditional signal processing approaches as well as learning-based baselines, often by a substantial margin, in a quantitative evaluation and a human study. We show qualitative examples where the model predicts acoustic changes according to the viewpoint changes, e.g., left channel becomes louder when the viewpoint changes from left to right. In a nutshell, we present the first work that deals with novel-view acoustic synthesis, and contribute two large-scale datasets along with a novel neural rendering approach for solving the task." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.312, + 0.642, + 0.327 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.333, + 0.892, + 0.453 + ], + "angle": 0, + "content": "Novel-view synthesis (NVS). Kickstarted by advances in neural rendering [34, 51], many recent works consider variants of the NVS problem. Most approaches assume dozens of calibrated images for reconstructing a single static scene. Closer to monocular video NVS, authors have considered reducing the number of input views [20, 25, 38, 46, 59] and modelling dynamic scenes [27, 28, 42, 43, 53, 55]. However, none of these works tackle audio." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.457, + 0.893, + 0.698 + ], + "angle": 0, + "content": "Acoustic matching and spatialization. NVAS requires accounting for (1) the environmental acoustics and (2) the geometric configuration of the target microphone(s) (e.g., monaural vs binaural). Modelling environmental acoustics has been addressed extensively by the audio community [4, 26]. Room impulse response (RIR) functions characterize the environment acoustics as a transfer function between the emitter and receiver, accounting for the scene geometry, materials, and emitter/receiver locations. Estimating the direct-to-reverberant ratio and the reverberation time, is sufficient to synthesize simple RIRs that match audio in a plausible manner [12, 16, 22, 30, 37, 58]. These methods do not synthesize for a target viewpoint, rather they resynthesize to match an audio sample. In [47, 48] sound from a moving emitter is spatialized towards a receiver conditioned on the tracked 3D location of the emitter." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.699, + 0.894, + 0.879 + ], + "angle": 0, + "content": "Recently, the vision community explores using visual information to estimate environmental acoustics [6, 8, 50]. However, these works only synthesize acoustics for a given viewpoint rather than a novel viewpoint. In addition, they have only addressed monaural audio, which is more forgiving than binaural because humans do not perceive subtle absolute acoustic properties, but can detect easily inconsistencies in the sounds perceived by the two ears. Recent work spatializes monaural sounds by upmixing them to multiple channels conditioned on the video, where the sound emitters are static [17, 36]. Because the environment, emitter and receiver are static, so are the acoustics. Other work" + }, + { + "type": "page_footnote", + "bbox": [ + 0.516, + 0.887, + 0.781, + 0.9 + ], + "angle": 0, + "content": "2https://replay-dataset.github.io" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6410" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.242 + ], + "angle": 0, + "content": "predicts impulse responses in simulation either for a single environment [29], or by using few-shot egocentric observations [31], or by using the 3D scene mesh [45]. While simulated results are satisfying, those models' impact on real-world data is unknown, especially for scenarios where human speakers move and interact with each other. Unlike any of the above, we introduce and tackle the NVAS problem, accounting for both acoustics and spatialization, and we propose a model that addresses the problem effectively on both synthetic and real-world data." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.246, + 0.473, + 0.368 + ], + "angle": 0, + "content": "Audio-visual learning. Recent advances in multi-modal video understanding enable new forms of self-supervised cross-modal feature learning from video [2, 24, 35], sound source localization [19, 21, 54], and audio-visual speech enhancement and source separation [1, 14, 33, 40, 60]. All of these existing tasks and datasets only deal with a single viewpoint. We introduce the first audio-visual learning task and dataset that deals with multi-view audio-visual data." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.382, + 0.436, + 0.4 + ], + "angle": 0, + "content": "3. The Novel-view Acoustic Synthesis Task" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.408, + 0.473, + 0.575 + ], + "angle": 0, + "content": "We introduce a new task, novel-view acoustic synthesis (NVAS). Assuming there are \\(N\\) sound emitters in the scene (emitter \\(i\\) emits sound \\(C^i\\) from location \\(L^i\\)), given the audio \\(A_{S}\\) and video \\(V_{S}\\) observed at the source viewpoint \\(S\\), the goal is to synthesize the audio \\(A_{T}\\) at the target viewpoint \\(T\\), as it would sound from the target location, specified by the relative pose \\(P_T\\) of the target microphone (translation and orientation) with respect to the source view (Fig. 1). Furthermore, we assume that the active sound emitters in the environment are visible in the source camera, but we make no assumptions about the camera at the target location." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.575, + 0.44, + 0.59 + ], + "angle": 0, + "content": "The sound at any point \\(R\\) is a function of the space:" + }, + { + "type": "equation", + "bbox": [ + 0.156, + 0.592, + 0.47, + 0.611 + ], + "angle": 0, + "content": "\\[\nA _ {R} = \\mathcal {F} \\left(L ^ {1, \\dots , N}, C ^ {1, \\dots , N}, R \\mid E\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.618, + 0.47, + 0.723 + ], + "angle": 0, + "content": "where \\(R\\) is the receiver location (\\(S\\) or \\(T\\)) and \\(E\\) is the environment. The emitted sounds \\(C^i\\) are not restricted to speech but can be ambient noise, sounding objects, etc. Our goal here is to learn a transfer function \\(\\mathcal{T}(\\cdot)\\) defined as \\(A_{T} = \\mathcal{T}(A_{S}, V_{S}, P_{T})\\), where \\(S, T, L^{1,\\dots,N}, C^{1,\\dots,N}, E\\) are not directly given and need to be inferred from \\(V_{S}\\) and \\(P_{T}\\), which makes the task inherently multi-modal." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.724, + 0.47, + 0.815 + ], + "angle": 0, + "content": "This task is challenging because the goal is to model the sound field of a dynamic scene and capture acoustic changes between viewpoints given one pair of audio-visual measurements. While traditional signal processing methods can be applied, we show in Sec. 6 that they perform poorly. In this work, we present a learning-based rendering approach." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.83, + 0.175, + 0.845 + ], + "angle": 0, + "content": "4. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.473, + 0.903 + ], + "angle": 0, + "content": "We introduce two datasets for the NVAS task: live recordings (Sec. 4.1), and simulated audio in scanned real-world environments (Sec. 4.2) (see Fig. 2). The former is" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.089, + 0.872, + 0.165 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.166, + 0.871, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.244, + 0.892, + 0.273 + ], + "angle": 0, + "content": "Figure 2. Example source and target views for the two introduced datasets: Replay-NVAS (left) and SoundSpaces-NVAS (right)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.28, + 0.892, + 0.34 + ], + "angle": 0, + "content": "real and covers various social scenarios, but offers limited diversity of sound sources, viewpoints and environments, and is noisy. The latter has a realism gap, but allows perfect control over these aforementioned elements." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.341, + 0.892, + 0.446 + ], + "angle": 0, + "content": "Both datasets focus on human speech given its relevance in applications. However, our model design is not specific to speech. For both datasets, we capture binaural audio, which best aligns with human perception. Note that for both datasets, we collect multiple multi-modal views for training and evaluation; during inference the target viewpoint(s) (and in some cases target environment) are withheld." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.455, + 0.742, + 0.472 + ], + "angle": 0, + "content": "4.1. The Replay-NVAS Dataset" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.478, + 0.892, + 0.554 + ], + "angle": 0, + "content": "Replay-NVAS contains multi-view captures of acted scenes in apartments. We capture 46 different scenarios (e.g., having a conversation, having dinner, or doing yoga) from 8 different viewpoints. In total, we collect 37 hours of video data, involving 32 participants across all scenarios." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.554, + 0.892, + 0.735 + ], + "angle": 0, + "content": "In each scenario, we invite 2-4 participants to act on a given topic. Each participant wears a near-range microphone, providing a clean recording of their own speech. The scene is captured by 8 DLSR cameras, each augmented with a 3Dio binaural microphone. In this way, the data captures video and audio simultaneously from multiple cameras, resulting in 56 possible source/target viewpoint combinations for each scene. The videos are recorded at 30 FPS and the audio is recorded with a 48k sampling rate. We use a clapper at the beginning of the recording for temporal synchronization. Each scenario lasts \\(3 - 8\\mathrm{min}\\). We use off-the-shelf software for multi-view camera calibration (see Supp.)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.84 + ], + "angle": 0, + "content": "To construct the dataset, we extract one-second long clips from each video with overlapping windows. We automatically remove silent and noisy clips based on the energy of near-range microphones, which results in 77K/12K/2K clips in total for train/val/test (details in Supp.) During training, for one sample, we randomly select two out of eight viewpoints, one as the source and one as the target." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.841, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This dataset is very challenging. It covers a wide range of social activities. It is harrowed by ambient sound, room reverberation, overlapping speech and non-verbal sounds such as clapping and instruments. Participants can move" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "6411" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "freely in the environment. We believe that this data will be useful to the community beyond the NVAS task as it can be used for benchmarking many other problems, including active speaker localization, source separation, and NVS." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.157, + 0.364, + 0.173 + ], + "angle": 0, + "content": "4.2. The SoundSpaces-NVAS Dataset" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.175, + 0.469, + 0.235 + ], + "angle": 0, + "content": "In this dataset, we synthesize multi-view audio-visual data of two people having conversations in 3D scenes. In total, we construct 1.3K hours of audio-visual data for a total of 1,000 speakers, 120 3D scenes and 200K viewpoints." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.236, + 0.469, + 0.431 + ], + "angle": 0, + "content": "Our goal is to construct audio-visual data with strong spatial and acoustic correspondences across multiple viewpoints, meaning that the visual information should indicate what the audio should sound like, e.g., observing speaker on the left should indicate the left ear is louder and observing speaker at a distance should indicate there is higher reverberation. We use the SoundSpaces 2.0 platform [7], which allows highly realistic audio and visual rendering for arbitrary camera and microphone locations in 3D scans of real-world environments [5, 52, 57]. It accounts for all major real-world acoustics phenomena: direct sounds, early specular/diffuse reflections, reverberation, binaural spatialization, and effects from materials and air absorption." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.432, + 0.469, + 0.627 + ], + "angle": 0, + "content": "We use the Gibson dataset [57] for scene meshes and LibriSpeech [41] for speech samples. As we are simulating two people having conversations, for a given environment, we randomly sample two speaker locations within \\(3\\mathrm{m}\\) and insert two copyright-free mannequins (one male and one female) at these two locations. We then randomly sample four nearby viewpoints facing the center of the two speakers at a height of \\(1.5\\mathrm{m}\\) (Fig. 2, right). For each speaker, we select a speech sample from LibriSpeech with matching gender. We render images at all locations as well as binaural impulse response for all pairs of points between speakers and viewpoints. The received sound is obtained by convolving the binaural impulse response with the speech sample." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.469, + 0.704 + ], + "angle": 0, + "content": "During training, for one sample, we randomly sample two out of four rendered viewpoints, one as the source and one as the target. We also randomly choose one speaker to be active, simulating what we observe on the real data (i.e., usually only one person speaks at a time)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.717, + 0.399, + 0.734 + ], + "angle": 0, + "content": "5. Visually-Guided Acoustic Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.743, + 0.469, + 0.803 + ], + "angle": 0, + "content": "We introduce a new method, Visually-Guided Acoustic Synthesis (ViGAS), to address the NVAS problem, taking as input sound and an image and outputting the sound from a different target microphone pose." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.804, + 0.469, + 0.878 + ], + "angle": 0, + "content": "ViGAS consists of five components: ambient sound separation, active speaker localization, visual acoustic network, acoustic synthesis, and temporal alignment. The high-level idea is to separate the observed sound into primary and ambient, extract useful visual information (active speaker and" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "acoustic features), and use this information to guide acoustic synthesis for the primary sound. Temporal alignment is performed during training for better optimization. ViGAS is discussed in detail next and summarised in Fig. 3." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.163, + 0.746, + 0.179 + ], + "angle": 0, + "content": "5.1. Ambient Sound Separation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.186, + 0.892, + 0.292 + ], + "angle": 0, + "content": "ViGAS starts by decomposing the input sound into primary and ambient (traffic, electric noise from a fridge or the A/C, etc.). Ambient sound is important for realism, but it also interferes with learning the model because it can carry significant energy, making the model focus on it rather than on the primary sounds, and its spatial distribution is very different from the primary sounds." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.293, + 0.892, + 0.443 + ], + "angle": 0, + "content": "By explicitly separating primary and ambient sounds, ViGAS: (1) accounts for the fact that the transfer functions of primary and ambient sounds are very different and thus difficult to model together; (2) avoids wasting representational power on modelling ambient sounds that might be difficult to reconstruct accurately and depend less on the viewpoint; and (3) prevents ambient sounds, which are noise-like and high-energy, from dominating learning and reconstruction. In practice, as we show in Sec. 6, without the ambient sound separation, the model performs poorly." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.444, + 0.892, + 0.624 + ], + "angle": 0, + "content": "The goal of ambient sound separation is thus to construct a function \\((A_C, A_N) = \\mathcal{P}(A_S)\\) that separates the input sound \\(A_S\\) into primary sound \\(A_C\\) and ambient sound \\(A_N\\). Existing approaches to this problem are based on signal processing [3, 13] or learning [11, 15]. We find that pretrained speech enhancement models such as Denoiser [11] tend to aggressively remove the noise including the primary sound, which hinders re-synthesis. We thus opt for band-pass filtering, passing frequencies within a certain range and rejecting/attenuating frequencies outside of it, which we found to work well. We cut frequencies below \\(80\\mathrm{Hz}\\) for SoundSpaces-NVAS and \\(150\\mathrm{Hz}\\) for Replay-NVAS." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.635, + 0.753, + 0.651 + ], + "angle": 0, + "content": "5.2. Active Speaker Localization" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.892, + 0.81 + ], + "angle": 0, + "content": "Knowing where the emitters of different primary sounds are located in the environment can help to solve the NVAS task. In this paper, we focus on localizing the active speaker, although there can be other important primary sound events like instruments playing, speakers interacting with objects, etc. The goal of active speaker localization is to predict the bounding box of the active speaker in each frame of the video (examples in Fig. 4). The bounding box is in the format of \\((y_{\\mathrm{min}}, y_{\\mathrm{max}}, x_{\\mathrm{min}}, x_{\\mathrm{max}})\\) and \\(x, y\\) are normalized to \\([0,1]\\) by the image width and height, respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.901 + ], + "angle": 0, + "content": "On SoundSpaces-NVAS, this task is relatively easy because of the strong correspondence between the appearance of the speaker and the gender of the speech sample, which enables to easily train a classifier for active speakers. However, this is much harder on Replay-NVAS because cameras record speakers from a distance and from diverse angles," + }, + { + "type": "page_footnote", + "bbox": [ + 0.094, + 0.887, + 0.412, + 0.9 + ], + "angle": 0, + "content": "3https://renderpeople.com/free-3d-people" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6412" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.09, + 0.873, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.281, + 0.895, + 0.394 + ], + "angle": 0, + "content": "Figure 3. Visually Guided Acoustic Synthesis (ViGAS). Given the input audio \\( A_{S} \\), we first separate out the ambient sound to focus on the sound of interest. We take the source audio and source visual to localize the active speaker on the 2D image. We also extract the visual acoustic features of the environment by running an encoder on the source visual. We concatenate the active speaker feature, source visual features, and the target pose, and fuse these features with a MLP. We feed both the audio stream \\( A_{C} \\) and fused visual feature \\( V_{C} \\) into the acoustic synthesis network, which has \\( M \\) stacked audio-visual fusion blocks. In each block, the audio sequence is processed by dilated conv1d layers and the visual features are processed by conv1d layers. Lastly, the previously separated ambient sound is added back to the waveform. During training, our temporal alignment module shifts the prediction by the amount of delay estimated between the source and the target audio to align the prediction well with the target." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.401, + 0.47, + 0.476 + ], + "angle": 0, + "content": "meaning that lip motion, the main cue used by speaker localization methods [21, 49, 54], is often not visible. Hence, the model has to rely on other cues to identify the speaker (such as body motion, gender or identity). Furthermore, sometimes people speak or laugh over each other." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.477, + 0.473, + 0.705 + ], + "angle": 0, + "content": "Since our focus is not speaker localization, for the Replay-NVAS we assume that this problem is solved by an external module that does audio-visual active speaker localization. To approximate the output of such a module automatically, we rely on the near-range audio recordings. Specifically, we first run an off-the-shelf detection and tracker [10] on the video at 5 FPS and obtain, with some manual refinement, bounding boxes \\( B_{t}^{i} \\) for \\( i = 1,\\dots ,N \\) at each frame \\( t \\). We manually assign the near-range microphone audio \\( A_N^i \\) to each tracked person. We select the active speaker \\( D \\) based on the maximum energy of each near-range microphone, i.e., \\( D = \\mathrm{argmax}_i\\left\\{\\sum A_N^i [t:t + \\Delta t]^2\\right\\} \\), where \\( \\Delta t \\) is the time interval we use to calculate the audio energy. We output bounding box \\( B^{D} \\) as the localization feature \\( V_{L} \\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.712, + 0.395, + 0.727 + ], + "angle": 0, + "content": "5.3. Visual Acoustic Network and Fusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.473, + 0.903 + ], + "angle": 0, + "content": "The active speaker bounding box \\( B^{D} \\) only disambiguates the active speaker from all visible humans on 2D, which is not enough to indicate where the speaker is in 3D. To infer that, the visual information is also needed. Since there is usually not much movement in one second (the length of the input video clip), the video clip does not provide much extra information compared to a single frame. Thus, we choose the middle frame to represent the clip and extract the visual acoustic features \\( V_{E} \\) from the input RGB image with a pretrained ResNet18 [18] before the average pooling layer to preserve spatial information. To reduce the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.401, + 0.892, + 0.445 + ], + "angle": 0, + "content": "feature size, we feed \\( V_{E} \\) into a 1D convolution with kernel size 1 and output channel size 8. We then flatten the visual features to obtain feature \\( V_{F} \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.446, + 0.892, + 0.521 + ], + "angle": 0, + "content": "The target pose is specified as the translation along \\( x, y, z \\) axes plus difference between orientations of the source \"view\" and the target \"view\" expressed via rotation angles: \\( +y \\) (roll), \\( +x \\) (pitch) and \\( +z \\) (yaw). We encode each angle \\( \\alpha \\) as its sinusoidal value: \\( (\\sin(\\alpha), \\cos(\\alpha)) \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.521, + 0.892, + 0.627 + ], + "angle": 0, + "content": "Similarly, the target pose is not enough by itself to indicate where the target viewpoint \\( T \\) is in the 3D space; to infer that, the source view \\( V_{S} \\) is again needed. For example, in top row of Fig 4, for target viewpoint 3, \"two meters to the right and one meter forward\" is not enough to indicate the target location is in the corridor, while the model can reason that based on the source view." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.627, + 0.892, + 0.688 + ], + "angle": 0, + "content": "We use a fusion network to predict a latent representation of the scene variables \\( S, T, L^{D}, E \\) (cf. Sec. 3) by first concatenating \\( [V_{L}, P_{T}, V_{F}] \\) and then feeding it through a multilayer perceptron (MLP). See Fig. 3 for the network." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.697, + 0.681, + 0.713 + ], + "angle": 0, + "content": "5.4. Acoustic Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.885 + ], + "angle": 0, + "content": "With the separated primary sound \\(A_{C}\\) and the visual acoustic feature \\(V_{C}\\) as input, the goal of the acoustic synthesis module is to transform \\(A_{C}\\) guided by \\(V_{C}\\). We design the acoustic synthesis network to learn a non-linear transfer function (implicitly) that captures these major acoustic phenomena, including the attenuation of sound in space, the directivity of sound sources (human speech is directional), the reverberation level, the head-related transfer function, as well as the frequency-dependent acoustic phenomena. Training end-to-end makes it possible to capture these subtle and complicated changes in the audio." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Inspired by recent advances in time-domain signal mod" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6413" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.182 + ], + "angle": 0, + "content": "eling [39, 48], we design the network as \\(M\\) stacked synthesis blocks, where each block consists of multiple conv1D layers. We first encode the input audio \\(A_{C}\\) into a latent space, which is then fed into the synthesis block. The key of the synthesis block is a gated multimodal fusion network that injects the visual information into the audio as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.085, + 0.187, + 0.47, + 0.207 + ], + "angle": 0, + "content": "\\[\nz = \\tanh \\left(p _ {A} ^ {k} \\left(A _ {F} ^ {k}\\right) + p _ {V} ^ {k} \\left(V _ {C}\\right)\\right) \\odot \\sigma \\left(q _ {A} ^ {k} \\left(A _ {F} ^ {k}\\right) + q _ {V} ^ {k} \\left(V _ {C}\\right)\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.214, + 0.469, + 0.259 + ], + "angle": 0, + "content": "where \\(\\odot\\) indicates element-wise multiplication, \\(\\sigma\\) is a logistic sigmoid function, \\(k = 1,\\dots ,M\\) is the layer index and \\(p,q\\) are both learnable 1D convolutions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.26, + 0.47, + 0.383 + ], + "angle": 0, + "content": "After passing \\( z \\) through a sinusoidal activation function, the network uses two separate conv1D layers to process the feature, one producing the residual connection \\( A_F^{k + 1} \\) and one producing the skip connection \\( A_P^{k + 1} \\). All skip connections \\( A_P^{k + 1} \\) are mean pooled and fed into a decoder to produce the output \\( A_O \\). We add back the separated ambient sound \\( A_N \\) as the target audio estimate: \\( \\hat{A}_T = A_O + A_N \\). See Supp. for more details on the architecture." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.391, + 0.274, + 0.407 + ], + "angle": 0, + "content": "5.5. Temporal Alignment" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.415, + 0.469, + 0.52 + ], + "angle": 0, + "content": "In order for the model to learn well, it is important that input and output sounds are temporally aligned. While the Replay-NVAS data is already synchronised based on the clapper sound, due to the finite speed of sound, the sounds emitted from different locations may still arrive at microphones with a delay slightly different from the one of the clapper, causing misalignments that affect training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.52, + 0.469, + 0.55 + ], + "angle": 0, + "content": "To align source and target audio for training, we find the delay \\(\\tau\\) that maximizes the generalized cross-correlation:" + }, + { + "type": "equation", + "bbox": [ + 0.148, + 0.551, + 0.469, + 0.568 + ], + "angle": 0, + "content": "\\[\n\\mathcal {R} _ {A _ {S}, A _ {T}} (\\tau) = \\mathbb {E} _ {t} [ h _ {S} (t) \\cdot h _ {T} (t - \\tau) ], \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.572, + 0.469, + 0.692 + ], + "angle": 0, + "content": "where \\( h_S \\) and \\( h_T \\) are the feature embedding for \\( A_S \\) and \\( A_T \\) respectively at time \\( t \\). We use the feature extractor \\( h \\) from the generalized cross-correlation phase transform (GCC-PHAT) algorithm [23], which whitens the audio by dividing by the magnitude of the cross-power spectral density. After computing \\( \\tau \\), we shift the prediction \\( A_O \\) by \\( \\tau \\) samples to align with the \\( A_T \\) and obtain \\( A_L \\). Note that alignment is already exact for SoundSpaces-NVAS." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.701, + 0.15, + 0.715 + ], + "angle": 0, + "content": "5.6. Loss" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.724, + 0.469, + 0.814 + ], + "angle": 0, + "content": "To compute the loss, we first encode the audio with the short-time Fourier transform (STFT), a complex-valued matrix representation of the audio where the \\(y\\) axis represents frequency and the \\(x\\) axis is time. We then compute the magnitude of the STFT, and optimize the L1 loss between the predicted and ground truth magnitudes as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.819, + 0.469, + 0.837 + ], + "angle": 0, + "content": "\\[\nL = \\left| \\left| \\left| S T F T \\left(A _ {L}\\right) \\right| \\right| _ {2} - \\left| \\left| S T F T \\left(A _ {T} ^ {\\prime}\\right) \\right| \\right| _ {2} \\right|, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.469, + 0.901 + ], + "angle": 0, + "content": "where \\(A_T^\\prime\\) is the primary sound separated from \\(A_{T}\\) with \\(\\mathcal{P}(\\cdot)\\). By taking the magnitude, we do not model the exact phase values, which we find hinders learning if being included in the loss. See implementation details in Supp." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.634, + 0.108 + ], + "angle": 0, + "content": "6. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.116, + 0.892, + 0.162 + ], + "angle": 0, + "content": "We compare with several traditional and learning-based baselines and show that ViGAS outperforms them in both a quantitative evaluation and a human subject study." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.166, + 0.892, + 0.302 + ], + "angle": 0, + "content": "Evaluation. We measure performance from three aspects: 1. closeness to GT as measured by the magnitude spectrogram distance (Mag). 2. correctness of the spatial sound as measured by the left-right energy ratio error (LRE), i.e., the difference of ratio of energy between left and right channels and 3. correctness of the acoustic properties measured by RT60 error (RTE) [6, 50], i.e., the error in reverberation time decaying by 60dB (RT60). We use a pretrained model [6] to estimate RT60 directly from speech." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.305, + 0.893, + 0.73 + ], + "angle": 0, + "content": "Baselines. We consider the following baselines: 1. Input audio. Copying the input to the output. 2. TF Estimator [56] + Nearest Neighbor, i.e. storing the transfer function estimated during training and retrieving the nearest neighbor during test time. We estimate transfer functions with a Wiener filter [56] and index them with the ground-truth locations of the speaker, source viewpoint, and target viewpoint for the single environment setup and their relative pose for the novel environment setup. At test time, this method searches the database to find the nearest transfer function and applies it on the input audio. 3. Digital Signal Processing (DSP) [9] approach that takes the distance, azimuth, and elevation of the sound source, applies an inverse a head-related transfer function (HRTF) to estimate the speech spoken by the speaker and then applies another HRTF to estimate the audio at the target microphone location. This baseline adjusts the loudness of the left and right channels based on where the speaker is in the target view. We supply GT coordinates for SoundSpaces-NVAS and speakers' head positions estimated with triangulation on Replay-NVAS. 4. Visual Acoustic Matching (VAM) [6], recently proposed for a related task of matching acoustics of input audio with a target image. This task only deals with single viewpoint and single-channel audio. We adapt their model with minimal modification by feeding in the image from the source viewpoint and concatenating the position offset of the target microphone at the multimodal fusion step. See Supp. for details." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.74, + 0.774, + 0.756 + ], + "angle": 0, + "content": "6.1. Results on SoundSpaces-NVAS" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.764, + 0.892, + 0.839 + ], + "angle": 0, + "content": "Table 1 shows the results. For synthetic data, we consider two evaluation setups: 1. single environment: train and test on the same environment and 2. novel environment: train and test on multiple non-overlapping Gibson environments (90/10/20 for train/val/test)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In the single environment setup, our model largely outperforms all baselines as well as our audio-only ablation on all metrics. TF Estimator performs poorly despite being indexed by the ground truth location values because es" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6414" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.124, + 0.089, + 0.844, + 0.247 + ], + "angle": 0, + "content": "
SoundSpaces-NVASReplay-NVAS
Single EnvironmentNovel EnvironmentSingle Environment
MagLRERTEMagLRERTEMagLRERTE
Input audio0.2251.4730.0320.2161.4080.0390.1591.4770.046
TF Estimator [56]0.3592.5960.0590.4403.2610.0920.3272.8610.147
DSP [9]0.3023.6440.0440.3003.6890.0470.4631.3000.067
VAM [6]0.2201.1980.0410.2351.1310.0510.1610.9240.070
ViGAS w/o visual0.1730.9730.0310.1811.0070.0360.1460.8770.046
ViGAS0.1590.7820.0290.1750.9710.0340.1420.7160.048
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.254, + 0.892, + 0.325 + ], + "angle": 0, + "content": "Table 1. Results on SoundSpaces-NVAS and Replay-NVAS. We report the magnitude spectrogram distance (Mag), left-right energy ratio error (LRE), and RT60 error (RTE). Replay-NVAS does not have novel environment setup due to data being collected in a single environment. For all metrics, lower is better. In addition to baselines, we also evaluate ViGAS w/o visual by removing the active speaker localization and visual features. Note that reverberation time is mostly invariant of the receiver location in the same room and thus input audio has low RTE. A good model should preserve this property while synthesizing the desired acoustics for the target viewpoint." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.341, + 0.471, + 0.567 + ], + "angle": 0, + "content": "timating a transfer function directly from two audio clips is non-trivial and noisy for low-energy parts of the signal. DSP also performs badly despite having the ground truth 3D coordinates of the sound source. This is because head related transfer functions are typically recorded in anechoic chambers, which does not account for acoustics of different environments, e.g., reverberation. Both traditional approaches perform worse than simply copying the input audio, indicating that learning-based models are needed for this challenging task. The recent model VAM [6] performs much better compared to the traditional approaches but still underperforms our model. There is a significant difference between ViGAS w/o visual and the full model; this shows that the visual knowledge about the speaker location and the environment is important for this task." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.572, + 0.47, + 0.664 + ], + "angle": 0, + "content": "Fig. 4 shows an example where given the same input source viewpoint, our model synthesizes audio for three different target viewpoints. The model reasons about how the geometry and speaker locations changes based on the source view and the target pose, and predicts the acoustic difference accordingly. See Supp. video to listen to sounds." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.667, + 0.471, + 0.849 + ], + "angle": 0, + "content": "For the novel environment setup, our model again outperforms all baselines. Compared to ViGAS in the single environment setup, both the magnitude spectrogram distance and the left-right energy ratio error increase. This is expected because for novel (unseen) environments, single images capture limited geometry and acoustic information. The model fails sometime when there is a drastic viewpoint change, e.g., target viewpoint 3 in Fig. 4. This setup requires the model to reason or \"imagine\" the environment based on single audio-visual observation, which poses great challenge for NVAS as well as NVS, where typically synthesis is performed in a fully observed environment." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Ablations. Table 2 shows ablations on the model design. To understand if the model uses visual information, we ablate the visual features \\( V_{F} \\) and the active speaker feature" + }, + { + "type": "table", + "bbox": [ + 0.507, + 0.339, + 0.885, + 0.461 + ], + "angle": 0, + "content": "
ViGASSS-NVASReplay-NVAS
MagLREMagLRE
full model0.1590.7820.1420.716
w/o visual features0.1710.8970.1460.920
w/o ASL0.1610.8140.1430.757
w/o alignment0.1760.7710.1440.706
w/o separation0.1650.8400.1820.859
" + }, + { + "type": "table_caption", + "bbox": [ + 0.542, + 0.467, + 0.848, + 0.48 + ], + "angle": 0, + "content": "Table 2. Ablations of the model on both datasets." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.487, + 0.892, + 0.577 + ], + "angle": 0, + "content": "\\(V_{L}\\). Removing the active speaker feature leads to less damage on the model performance, because without the explicitly localized active speaker, the model can still implicitly reason about the active speaker location based on the image and audio. If both are removed (\"ViGAS w/o visual\" in Table 1), the performance suffers most." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.578, + 0.892, + 0.669 + ], + "angle": 0, + "content": "To study the effectiveness of the temporal alignment and ambient sound separation modules, we ablate them separately. Removing the temporal alignment leads to higher Mag error and slightly lower LRE. As for ambient sound separation, the results show that optimizing for the high-energy noise-like ambient sound degrades the performance." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.681, + 0.727, + 0.697 + ], + "angle": 0, + "content": "6.2. Results on Replay-NVAS" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.704, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Table 1 (right) shows the Replay-NVAS results. Compared to SoundSpaces-NVAS, the magnitudes of all errors are smaller because there are less drastic acoustic changes between viewpoints (8 DLSR cameras form a circle around the participants). Traditional approaches like TF Estimator and DSP still perform poorly despite using the 3D coordinates of the camera and the speaker (triangulated from multiple cameras). VAM performs better due to end-to-end learning; however, our model outperforms it. Compared to ViGAS w/o visual, the full model has much lower left-right energy ratio error and slightly higher reverberation time error, showing that the model takes into account the speaker position and viewpoint change for synthesizing the audio." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6415" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.285, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.287, + 0.089, + 0.487, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.49, + 0.089, + 0.688, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.692, + 0.089, + 0.889, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.2, + 0.344, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.345, + 0.2, + 0.604, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.609, + 0.2, + 0.888, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.305, + 0.892, + 0.445 + ], + "angle": 0, + "content": "Figure 4. Qualitative examples. For all binaural audio, we show the left-channel and the right-channel waveforms side-by-side. Row 1: SoundSpaces-NVAS example where given the source viewpoint and input audio, the model synthesizes audio for three different target viewpoints (target views are for reference only). In this case, the active speaker is the male speaker as indicated by the bounding box. For target viewpoint 1, the view rotates about 90 degrees and the male speaker is on the left side and the predicted left channel is louder than the right channel. Viewpoint 2 moves away from the speaker and thus yields lower amplitude compared to the first prediction. For target viewpoint 3, it is completely located outside of the living room, in which case, the sound could only come from the door open on the right (louder right channel) and the reverberation also greatly increases due to the vanishing direct sound. Row 2: Replay-NVAS example where the speaker is located on the left in the source viewpoint which becomes the right and further from the camera in target viewpoint 2, the model also predicts lower amplitude and louder right channel. On the right side, we show an example of the audio-visual speech enhancement for the active speaker. The model enhances the speech to largely match with the near-range audio (target)." + }, + { + "type": "table", + "bbox": [ + 0.138, + 0.458, + 0.405, + 0.52 + ], + "angle": 0, + "content": "
MagRTE
Input0.2790.376
ViGAS (ours)0.2340.122
" + }, + { + "type": "table_caption", + "bbox": [ + 0.123, + 0.527, + 0.422, + 0.541 + ], + "angle": 0, + "content": "Table 3. Speech enhancement on Replay-NVAS." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.547, + 0.47, + 0.637 + ], + "angle": 0, + "content": "Fig. 4 (row 2, left) shows a qualitative example. In the source viewpoint, the active speaker is on the left, while in the target viewpoint, he is further from the camera and on the right. The model synthesizes an audio waveform that captures the corresponding acoustic change, showing that our model successfully learns from real videos." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.641, + 0.47, + 0.792 + ], + "angle": 0, + "content": "Audio-visual speech enhancement. In some real-world applications, e.g., hearing aid devices, the goal is to obtain the enhanced clean speech of the active speaker. This can be seen as a special case of NVAS, where the target viewpoint is the active speaker. Our model is capable of performing audio-visual speech enhancement without any modification. We simply set the target audio to the near-range audio recording for the active speaker. We show the results in Table 3. Our model obtains cleaner audio compared to the input audio (example in Fig. 4, row 2, right)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Human subject study. To supplement the quantitative metrics and evaluate how well our synthesized audio captures the acoustic change between viewpoints, we conduct a human subject study. We show participants the image of the target viewpoint \\( V_{T} \\) as well as the audio \\( A_{T} \\) as reference. We provide three audio samples: the input, the prediction of ViGAS, and the prediction of DSP (the most nat" + }, + { + "type": "table", + "bbox": [ + 0.529, + 0.458, + 0.86, + 0.52 + ], + "angle": 0, + "content": "
DatasetInputDSPViGAS
SoundSpaces-NVAS24%2%74%
Replay-NVAS43%6%51%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.527, + 0.892, + 0.569 + ], + "angle": 0, + "content": "Table 4. Human Study. Participants favor our approach over the two most realistic sounding baselines, (1) copying the input signal, and (2) a digital signal processing baseline." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.576, + 0.892, + 0.636 + ], + "angle": 0, + "content": "urally sounding baseline) and ask them to select a clip that sounds closest to the target audio. We select 20 examples from SoundSpaces-NVAS and 20 examples from Replay-NVAS and invite 10 participants to perform the study." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.637, + 0.892, + 0.697 + ], + "angle": 0, + "content": "See Table 4 for the results. On the synthetic dataset SoundSpaces-NVAS, our approach is preferred over the baselines by a large margin. This margin is lower on the real-world Replay-NVAS dataset but is still significant." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.71, + 0.619, + 0.725 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We introduce the challenging novel-view acoustic synthesis task and a related benchmark in form of both real and synthetic datasets. We propose a neural rendering model that learns to transform the sound from the source viewpoint to the target viewpoint by reasoning about the observed audio and visual stream. Our model surpasses all baselines on both datasets. We believe this research unlocks many potential applications and research in multimodal novel-view synthesis. In the future, we plan to incorporate active-speaker localization model into the approach and let the model jointly learn to localize and synthesize." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6416" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.156 + ], + "angle": 0, + "content": "[1] Triantafyllos Afouras, Joon Son Chung, and Andrew Zisserman. The conversation: Deep audio-visual speech enhancement. In INTERSPEECH, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.157, + 0.472, + 0.212 + ], + "angle": 0, + "content": "[2] Humam Alwassel, Dhruv Mahajan, Bruno Korbar, Lorenzo Torresani, Bernard Ghanem, and Du Tran. Self-supervised learning by cross-modal audio-video clustering. In NeurIPS, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.213, + 0.471, + 0.269 + ], + "angle": 0, + "content": "[3] M. Berouti, Richard Schwartz, and John Makhoul. Enhancement of speech corrupted by acoustic noise. In IEEE International Conference on Acoustics, Speech, and Signal Processing, 1979. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.27, + 0.471, + 0.298 + ], + "angle": 0, + "content": "[4] J.S. Bradley. Review of objective room acoustics measures and future needs. Applied Acoustics, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.299, + 0.471, + 0.381 + ], + "angle": 0, + "content": "[5] Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niessner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from rgb-d data in indoor environments. 3DV, 2017. MatterPort3D dataset license available at: http://kaldir.vc.in.tum.de/matterport/MP_TOS.pdf.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.382, + 0.471, + 0.422 + ], + "angle": 0, + "content": "[6] Changan Chen, Ruohan Gao, Paul Calamia, and Kristen Grauman. Visual acoustic matching. In CVPR, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.424, + 0.471, + 0.493 + ], + "angle": 0, + "content": "[7] Changan Chen, Carl Schissler, Sanchit Garg, Philip Kobernik, Alexander Clegg, Paul Calamia, Dhruv Batra, Philip W Robinson, and Kristen Grauman. Soundspaces 2.0: A simulation platform for visual-acoustic learning. In NeurIPS 2022 Datasets and Benchmarks Track, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.494, + 0.471, + 0.535 + ], + "angle": 0, + "content": "[8] Changan Chen, Wei Sun, David Harwath, and Kristen Grauman. Learning audio-visual dereverberation. In ICASSP, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.536, + 0.471, + 0.592 + ], + "angle": 0, + "content": "[9] Corey I. Cheng and Gregory H. Wakefield. Introduction to head-related transfer functions (hrtfs): representations of hrtfs in time, frequency, and space. journal of the audio engineering society, 49(4):231-249, april 2001. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.593, + 0.469, + 0.635 + ], + "angle": 0, + "content": "[10] MMTracking Contributors. MMTracking: OpenMMLab video perception toolbox and benchmark. https://github.com/open-mmlab/mmtracking, 2020.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.635, + 0.469, + 0.676 + ], + "angle": 0, + "content": "[11] Alexandre Defossez, Gabriel Synnaeve, and Yossi Adi. Real time speech enhancement in the waveform domain. In *Interspeech*, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.677, + 0.469, + 0.733 + ], + "angle": 0, + "content": "[12] James Eaton, Nikolay Gaubitch, Allistair Moore, and Patrick Naylor. Estimation of room acoustic parameters: The ACE challenge. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 24(10), 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.734, + 0.469, + 0.775 + ], + "angle": 0, + "content": "[13] Yariv Ephraim and Harry L. Van Trees. A signal subspace approach for speech enhancement. IEEE Transactions on Speech and Audio Processing, 1995. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.469, + 0.844 + ], + "angle": 0, + "content": "[14] Ariel Ephrat, Inbar Mosseri, Oran Lang, Tali Dekel, Kevin Wilson, Avinatan Hassidim, William T Freeman, and Michael Rubinstein. Looking to listen at the cocktail party: A speaker-independent audio-visual model for speech separation. In SIGGRAPH, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[15] Szu-Wei Fu, Chien-Feng Liao, Yu Tsao, and Shou-De Lin. Metricgan: Generative adversarial networks based black-box metric scores optimization for speech enhancement. In International Conference on Machine Learning (ICML), 2019. 4" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.148 + ], + "angle": 0, + "content": "[16] Hannes Gamper and Ivan J Tashev. Blind reverberation time estimation using a convolutional neural network. In 2018 16th International Workshop on Acoustic Signal Enhancement (IWAENC), pages 136-140, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.175 + ], + "angle": 0, + "content": "[17] Ruohan Gao and Kristen Grauman. 2.5d visual sound. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.177, + 0.892, + 0.218 + ], + "angle": 0, + "content": "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.219, + 0.892, + 0.273 + ], + "angle": 0, + "content": "[19] Di Hu, Rui Qian, Minyue Jiang, Xiao Tan, Shilei Wen, Errui Ding, Weiyao Lin, and Dejing Dou. Discriminative sounding objects localization via self-supervised audiovisual matching. In NeurIPS, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.274, + 0.892, + 0.315 + ], + "angle": 0, + "content": "[20] Ajay Jain, Matthew Tancik, and Pieter Abbeel. Putting nerf on a diet: Semantically consistent few-shot view synthesis. In Proc. ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.316, + 0.892, + 0.357 + ], + "angle": 0, + "content": "[21] Hao Jiang, Calvin Murdock, and Vamsi Krishna Ithapu. Ego-centric deep multi-channel audio-visual active speaker localization. In CVPR, 2022. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.358, + 0.892, + 0.44 + ], + "angle": 0, + "content": "[22] Florian Klein, Annika Neidhardt, and Marius Seipel. Real-time estimation of reverberation time for selection of suitable binaural room impulse responses. In Audio for Virtual, Augmented and Mixed Realities: Proceedings of 5th International Conference on Spatial Audio (ICSA), pages 145–150, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.441, + 0.892, + 0.483 + ], + "angle": 0, + "content": "[23] C. Knapp and G. Carter. The generalized correlation method for estimation of time delay. IEEE Transactions on Acoustics, Speech, and Signal Processing, 1976. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.484, + 0.892, + 0.525 + ], + "angle": 0, + "content": "[24] Bruno Korbar, Du Tran, and Lorenzo Torresani. Cooperative learning of audio and video models from self-supervised synchronization. In NeurIPS, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.525, + 0.892, + 0.567 + ], + "angle": 0, + "content": "[25] Jonas Kulhanek, Erik Derner, Torsten Sattler, and Robert Babuška. ViewFormer: NeRF-free neural rendering from few images using transformers. In Proc. ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.567, + 0.892, + 0.593 + ], + "angle": 0, + "content": "[26] Heinrich Kuttruff. Room Acoustics. Boca Raton, 6th edition, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.595, + 0.892, + 0.637 + ], + "angle": 0, + "content": "[27] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. arXiv.cs, abs/2011.13084, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.637, + 0.892, + 0.692 + ], + "angle": 0, + "content": "[28] Lingjie Liu, Marc Habermann, Viktor Rudnev, Kripasindhu Sarkar, Jiatao Gu, and Christian Theobalt. Neural humans: Pose-controlled free-view synthesis of human actors with template-guided neural radiance fields. In arXiv, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.692, + 0.892, + 0.734 + ], + "angle": 0, + "content": "[29] Andrew Luo, Yilun Du, Michael J Tarr, Joshua B Tenenbaum, Antonio Torralba, and Chuang Gan. Learning neural acoustic fields. In NeurIPS, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.735, + 0.892, + 0.789 + ], + "angle": 0, + "content": "[30] Wolfgang Mack, Shuwen Deng, and Emanuele AP Habets. Single-channel blind direct-to-reverberation ratio estimation using masking. In INTERSPEECH, pages 5066-5070, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.79, + 0.892, + 0.845 + ], + "angle": 0, + "content": "[31] Sagnik Majumder, Changan Chen, Ziad Al-Halah, and Kristen Grauman. Few-shot audio-visual learning of environment acoustics. In Thirty-Sixth Conference on Neural Information Processing Systems, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[32] Ricardo Martin-Brualla, Noha Radwan, Mehdi S. M. Sajjadi, Jonathan T. Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In Proceedings of the IEEE/CVF" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6417" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.093, + 0.469, + 0.12 + ], + "angle": 0, + "content": "Conference on Computer Vision and Pattern Recognition (CVPR), pages 7210-7219, June 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.121, + 0.469, + 0.176 + ], + "angle": 0, + "content": "[33] Daniel Michelsanti, Zheng-Hua Tan, Shi-Xiong Zhang, Yong Xu, Meng Yu, Dong Yu, and Jesper Jensen. An overview of deep-learning-based audio-visual speech enhancement and separation. In arXiv, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.177, + 0.469, + 0.232 + ], + "angle": 0, + "content": "[34] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.233, + 0.469, + 0.273 + ], + "angle": 0, + "content": "[35] Pedro Morgado, Yi Li, and Nuno Vasconcelos. Learning representations from audio-visual spatial alignment. In NeurIPS, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.274, + 0.469, + 0.315 + ], + "angle": 0, + "content": "[36] Pedro Morgado, Nono Vasconcelos, Timothy Langlois, and Oliver Wang. Self-supervised generation of spatial audio for 360 video. In NeurIPS, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.316, + 0.469, + 0.372 + ], + "angle": 0, + "content": "[37] Prateek Murgai, Mark Rau, and Jean-Marc Jot. Blind estimation of the reverberation fingerprint of unknown acoustic environments. In Audio Engineering Society Convention 143. Audio Engineering Society, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.372, + 0.469, + 0.427 + ], + "angle": 0, + "content": "[38] Michael Niemeyer, Jonathan T. Barron, Ben Mildenhall, Mehdi S. M. Sajjadi, Andreas Geiger, and Noha Radwan. RegNeRF: Regularizing neural radiance fields for view synthesis from sparse inputs. In Proc. CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.428, + 0.469, + 0.495 + ], + "angle": 0, + "content": "[39] Aaron van den Oord, Sander Dieleman, Heiga Zen, Karen Simonyan, Oriol Vinyls, Alex Graves, Nal Kalchbrenner, Andrew Senior, and Koray Kavukcuoglu. Wavenet: A generative model for raw audio, 2016. cite arxiv:1609.03499. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.497, + 0.469, + 0.537 + ], + "angle": 0, + "content": "[40] Andrew Owens and Alexei A Efros. Audio-visual scene analysis with self-supervised multisensory features. In ECCV, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.539, + 0.469, + 0.607 + ], + "angle": 0, + "content": "[41] Vassil Panayotov, Guoguo Chen, Daniel Povey, and Sanjeev Khudanpur. Librispeech: an asr corpus based on public domain audio books. In 2015 IEEE international conference on acoustics, speech and signal processing (ICASSP), pages 5206-5210. IEEE, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.609, + 0.469, + 0.663 + ], + "angle": 0, + "content": "[42] Keunhong Park, Utkarsh Sinha, Jonathan T. Barron, Sofien Bouaziz, Dan B. Goldman, Steven M. Seitz, and Ricardo Martin-Brualla. Deformable neural radiance fields. CoRR, abs/2011.12948, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.665, + 0.469, + 0.705 + ], + "angle": 0, + "content": "[43] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-NeRF: Neural radiance fields for dynamic scenes. arXiv.cs, abs/2011.13961, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.707, + 0.469, + 0.775 + ], + "angle": 0, + "content": "[44] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10318–10327, June 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.469, + 0.83 + ], + "angle": 0, + "content": "[45] Anton Ratnarajah, Zhenyu Tang, Rohith Chandrashekar Aralikatti, and Dinesh Manocha. Mesh2ir: Neural acoustic impulse response generator for complex 3d scenes. In ACM Multimedia, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.469, + 0.899 + ], + "angle": 0, + "content": "[46] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common Objects in 3D: Large-scale learning and evaluation of real-life 3D category reconstruction. In Proc. CVPR, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.469, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.149 + ], + "angle": 0, + "content": "[47] Alexander Richard, Peter Dodds, and Vamsi Krishna Ithapu. Deep impulse responses: Estimating and parameterizing filters with deep networks. In IEEE International Conference on Acoustics, Speech and Signal Processing, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[48] Alexander Richard, Dejan Markovic, Israel D Gebru, Steven Krenn, Gladstone Butler, Fernando de la Torre, and Yaser Sheikh. Neural synthesis of binaural speech from mono audio. In International Conference on Learning Representations, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.892, + 0.303 + ], + "angle": 0, + "content": "[49] Joseph Roth, Sourish Chaudhuri, Ondrej Klejch, Radhika Marvin, Andrew Gallagher, Liat Kaver, Sharadh Ramaswamy, Arkadiusz Stopczynski, Cordelia Schmid, Zhonghua Xi, and Caroline Pantofaru. Ava-activespeaker: An audio-visual dataset for active speaker detection. In ICASSP, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.305, + 0.892, + 0.347 + ], + "angle": 0, + "content": "[50] Nikhil Singh, Jeff Mentch, Jerry Ng, Matthew Beveridge, and Iddo Drori. Image2reverb: Cross-modal reverb impulse response synthesis. In ICCV, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.349, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[51] Vincent Sitzmann, Michael Zollhöfer, and Gordon Wetzstein. Scene representation networks: Continuous 3D-structure-aware neural scene representations. Proc. NeurIPS, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.405, + 0.892, + 0.473 + ], + "angle": 0, + "content": "[52] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, et al. The Replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.476, + 0.892, + 0.517 + ], + "angle": 0, + "content": "[53] Shih-Yang Su, Frank Yu, Michael Zollhöfer, and Helge Rhodin. A-NeRF: Surface-free human 3d pose refinement via neural rendering. arXiv.cs, abs/2102.06199, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.519, + 0.892, + 0.6 + ], + "angle": 0, + "content": "[54] Ruijie Tao, Zexu Pan, Rohan Kumar Das, Xinyuan Qian, Mike Zheng Shou, and Haizhou Li. Is someone speaking? exploring long-term temporal features for audio-visual active speaker detection. In Proceedings of the 29th ACM International Conference on Multimedia, page 3927–3935, 2021. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.603, + 0.892, + 0.672 + ], + "angle": 0, + "content": "[55] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Christoph Lassner, and Christian Theobalt. Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a deforming scene from monocular video. arXiv.cs, abs/2012.12247, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.675, + 0.892, + 0.716 + ], + "angle": 0, + "content": "[56] Norbert Wiener. Extrapolation, interpolation, and smoothing of stationary time series. Report of the Services 19, Research Project DIC-6037 MIT, 1942. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.718, + 0.892, + 0.785 + ], + "angle": 0, + "content": "[57] Fei Xia, Amir R. Zamir, Zhi-Yang He, Alexander Sax, Jitendra Malik, and Silvio Savarese. Gibson Env: real-world perception for embodied agents. In Computer Vision and Pattern Recognition (CVPR), 2018 IEEE Conference on. IEEE, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.788, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[58] Feifei Xiong, Stefan Goetze, Birger Kollmeier, and Bernd T Meyer. Joint estimation of reverberation time and early-to-late reverberation ratio from single-channel speech signals. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 27(2):255-267, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[59] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. PixelNeRF: Neural radiance fields from one or few images. In Proc. CVPR, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6418" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.092, + 0.47, + 0.133 + ], + "angle": 0, + "content": "[60] Hang Zhou, Ziwei Liu, Xudong Xu, Ping Luo, and Xiaogang Wang. Vision-infused deep audio inpainting. In ICCV, 2019. 3" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.517, + 0.957 + ], + "angle": 0, + "content": "6419" + } + ] +] \ No newline at end of file diff --git a/2023/Novel-View Acoustic Synthesis/2dea3110-7f39-4b1f-9a6f-a10e0ea0bd92_origin.pdf b/2023/Novel-View Acoustic Synthesis/2dea3110-7f39-4b1f-9a6f-a10e0ea0bd92_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..048e91990cc22b11953cdbae03cd0f77701b35bd --- /dev/null +++ b/2023/Novel-View Acoustic Synthesis/2dea3110-7f39-4b1f-9a6f-a10e0ea0bd92_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f00622b34aedad1eeb7de84aa5153697f3173f93e6f36d04a3e079e4e970a7ef +size 3860721 diff --git a/2023/Novel-View Acoustic Synthesis/full.md b/2023/Novel-View Acoustic Synthesis/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e4134b31eca3f3bcdb287ba430fa8c51edde987a --- /dev/null +++ b/2023/Novel-View Acoustic Synthesis/full.md @@ -0,0 +1,317 @@ +# Novel-View Acoustic Synthesis + +Changan Chen $^{1,3}$ Alexander Richard $^{2}$ Roman Shapovalov $^{3}$ Vamsi Krishna Ithapu $^{2}$ Natalia Neverova $^{3}$ Kristen Grauman $^{1,3}$ Andrea Vedaldi $^{3}$ $^{1}$ University of Texas at Austin $^{2}$ Reality Labs Research at Meta $^{3}$ FAIR, Meta AI + +# Abstract + +We introduce the novel-view acoustic synthesis (NVAS) task: given the sight and sound observed at a source viewpoint, can we synthesize the sound of that scene from an unseen target viewpoint? We propose a neural rendering approach: Visually-Guided Acoustic Synthesis (ViGAS) network that learns to synthesize the sound of an arbitrary point in space by analyzing the input audio-visual cues. To benchmark this task, we collect two first-of-their-kind large-scale multi-view audio-visual datasets, one synthetic and one real. We show that our model successfully reasons about the spatial cues and synthesizes faithful audio on both datasets. To our knowledge, this work represents the very first formulation, dataset, and approach to solve the novel-view acoustic synthesis task, which has exciting potential applications ranging from AR/VR to art and design. Unlocked by this work, we believe that the future of novel-view synthesis is in multi-modal learning from videos. + +# 1. Introduction + +Replaying a video recording from a new viewpoint1 has many applications in cinematography, video enhancement, and virtual reality. For example, it can be used to edit a video, simulate a virtual camera, or, given a video of a personal memory, even enable users to experience a treasured moment again—not just on a 2D screen, but in 3D in a virtual or augmented reality, thus ‘reliving’ the moment. + +While the applications are exciting, there are still many unsolved technical challenges. Recent advances in 3D reconstruction and novel-view synthesis (NVS) address the problem of synthesizing new images of a given scene [32, 34, 44]. However, thus far, the view synthesis problem is concerned with creating visuals alone; the output is silent or at best naively adopts the sounds of the original video (from the "wrong" viewpoint). Without sound, the emotional and cognitive significance of the replay is severely diminished. + +In this work, we address this gap and introduce the new task of novel-view acoustic synthesis (NVAS). The goal of + +![](images/a0b9f4beb4e51b7fc4434efef3758801200bf5222e33d70bd4a94612aaa30df4.jpg) +Figure 1. Novel-view acoustic synthesis task. Given audio-visual observations from one viewpoint and the relative target viewpoint pose, render the sound received at the target viewpoint. Note that the target is expressed as the desired pose of the microphones; the image at that pose (right) is neither observed nor synthesized. + +this task is to synthesize the sound in a scene from a new acoustic viewpoint, given only the visual and acoustic input from another source viewpoint in the same scene (Fig. 1). + +NVAS is very different from the existing NVS task, where the goal is to reconstruct images instead of sounds, and these differences present new challenges. First, the 3D geometry of most real-life scenes changes in a limited manner during the recording. On the contrary, sound changes substantially over time, so the reconstruction target is highly dynamic. Secondly, visual and audio sensors are very different. A camera matrix captures the light in a highly-directional manner, and a single image comprises a large 2D array of pixels. In contrast, sounds are recorded with one or two microphones which are at best weakly-directional, providing only a coarse sampling of the sound field. Thirdly, the frequency of light waves is much higher than that of sound waves; the length of audio waves is thus larger to the point of being comparable to the size of geometric features of the scene, meaning that effects such as diffraction are often dominant, and spatial resolution is low. As a result, techniques that require spatial precision, such as triangulation and segmentation, are not applicable to audio. Lastly, sounds mix together, making it difficult to segment them, + +and they are affected by environmental effects such as reverberation that are distributed and largely unobservable. + +While the NVS and NVAS tasks are indeed very different, we hypothesize that NVAS is an inherently multimodal task. In fact, vision can play an important role in achieving accurate sound synthesis. First, establishing correspondences between sounds and their sources as they appear in images can provide essential cues for resynthesizing the sounds realistically. For instance, human speech is highly directional and sounds very differently if one faces the speaker or their back, which can only be inferred from visual cues. In addition, the environment acoustics also affect the sound one hears as a function of the scene geometry, materials, and emitter/receiver locations. The same source sounds vary differently if it is located in the center of a room, at the corner, or in a corridor, for example. In short, vision provides cues about space and geometry that affect sound, and are difficult to estimate from the sound alone. + +In order to validate our hypothesis, we propose a novel visually-guided acoustic synthesis network that analyzes audio and visual features and synthesizes the audio at a target location. More specifically, the network first takes as input the image observed at the source viewpoint in order to infer global acoustic and geometric properties of the environment along with the bounding box of the active speaker. The network then reasons how the speaker and scene geometry change in 3D based on the relative target pose with a fusion network. We inject the fused features into audio with a gated multi-modal fusion network and model the acoustic changes between viewpoints with a time-domain model. + +In order to conduct our experiments on the new NVAS task, we require suitable training and benchmarking data, of which currently there is none available. To address that, we contribute two new datasets: one real (Replay-NVAS) and one synthetic (SoundSpaces-NVAS). The key feature of these datasets is to record the sight and sound of different scenes from multiple cameras/viewpoints. Replay-NVAS contains video recordings of groups of people performing social activities (e.g., chatting, watching TV, doing yoga, playing instruments) from 8 surrounding viewpoints simultaneously. It contains 37 hours of highly realistic everyday conversation and social interactions in one homelike environment. To our knowledge, Replay-NVAS represents the first large-scale real-world dataset enabling NVAS. This dataset would also greatly benefit many other existing tasks including NVS, active speaker localization, etc. For SoundSpaces-NVAS, we render 1.3K hours of audiovisual data based on the SoundSpaces [7] platform. Using this simulator, one can easily change the scene geometry and the positions of speakers, cameras, and microphones. This data serves as a powerful test bed with clean ground truth for a large collection of home environments, offering a good complement to Replay-NVAS. For both datasets, + +we capture binaural audio, which is what humans perceive with two ears. Together the datasets contain 1,337 hours of audio-visual capture, with 1,032 speakers across 121 3D scenes. Datasets are publicly available for future research. + +We show that our model outperforms traditional signal processing approaches as well as learning-based baselines, often by a substantial margin, in a quantitative evaluation and a human study. We show qualitative examples where the model predicts acoustic changes according to the viewpoint changes, e.g., left channel becomes louder when the viewpoint changes from left to right. In a nutshell, we present the first work that deals with novel-view acoustic synthesis, and contribute two large-scale datasets along with a novel neural rendering approach for solving the task. + +# 2. Related Work + +Novel-view synthesis (NVS). Kickstarted by advances in neural rendering [34, 51], many recent works consider variants of the NVS problem. Most approaches assume dozens of calibrated images for reconstructing a single static scene. Closer to monocular video NVS, authors have considered reducing the number of input views [20, 25, 38, 46, 59] and modelling dynamic scenes [27, 28, 42, 43, 53, 55]. However, none of these works tackle audio. + +Acoustic matching and spatialization. NVAS requires accounting for (1) the environmental acoustics and (2) the geometric configuration of the target microphone(s) (e.g., monaural vs binaural). Modelling environmental acoustics has been addressed extensively by the audio community [4, 26]. Room impulse response (RIR) functions characterize the environment acoustics as a transfer function between the emitter and receiver, accounting for the scene geometry, materials, and emitter/receiver locations. Estimating the direct-to-reverberant ratio and the reverberation time, is sufficient to synthesize simple RIRs that match audio in a plausible manner [12, 16, 22, 30, 37, 58]. These methods do not synthesize for a target viewpoint, rather they resynthesize to match an audio sample. In [47, 48] sound from a moving emitter is spatialized towards a receiver conditioned on the tracked 3D location of the emitter. + +Recently, the vision community explores using visual information to estimate environmental acoustics [6, 8, 50]. However, these works only synthesize acoustics for a given viewpoint rather than a novel viewpoint. In addition, they have only addressed monaural audio, which is more forgiving than binaural because humans do not perceive subtle absolute acoustic properties, but can detect easily inconsistencies in the sounds perceived by the two ears. Recent work spatializes monaural sounds by upmixing them to multiple channels conditioned on the video, where the sound emitters are static [17, 36]. Because the environment, emitter and receiver are static, so are the acoustics. Other work + +predicts impulse responses in simulation either for a single environment [29], or by using few-shot egocentric observations [31], or by using the 3D scene mesh [45]. While simulated results are satisfying, those models' impact on real-world data is unknown, especially for scenarios where human speakers move and interact with each other. Unlike any of the above, we introduce and tackle the NVAS problem, accounting for both acoustics and spatialization, and we propose a model that addresses the problem effectively on both synthetic and real-world data. + +Audio-visual learning. Recent advances in multi-modal video understanding enable new forms of self-supervised cross-modal feature learning from video [2, 24, 35], sound source localization [19, 21, 54], and audio-visual speech enhancement and source separation [1, 14, 33, 40, 60]. All of these existing tasks and datasets only deal with a single viewpoint. We introduce the first audio-visual learning task and dataset that deals with multi-view audio-visual data. + +# 3. The Novel-view Acoustic Synthesis Task + +We introduce a new task, novel-view acoustic synthesis (NVAS). Assuming there are $N$ sound emitters in the scene (emitter $i$ emits sound $C^i$ from location $L^i$ ), given the audio $A_{S}$ and video $V_{S}$ observed at the source viewpoint $S$ , the goal is to synthesize the audio $A_{T}$ at the target viewpoint $T$ , as it would sound from the target location, specified by the relative pose $P_T$ of the target microphone (translation and orientation) with respect to the source view (Fig. 1). Furthermore, we assume that the active sound emitters in the environment are visible in the source camera, but we make no assumptions about the camera at the target location. + +The sound at any point $R$ is a function of the space: + +$$ +A _ {R} = \mathcal {F} \left(L ^ {1, \dots , N}, C ^ {1, \dots , N}, R \mid E\right), \tag {1} +$$ + +where $R$ is the receiver location ( $S$ or $T$ ) and $E$ is the environment. The emitted sounds $C^i$ are not restricted to speech but can be ambient noise, sounding objects, etc. Our goal here is to learn a transfer function $\mathcal{T}(\cdot)$ defined as $A_{T} = \mathcal{T}(A_{S}, V_{S}, P_{T})$ , where $S, T, L^{1,\dots,N}, C^{1,\dots,N}, E$ are not directly given and need to be inferred from $V_{S}$ and $P_{T}$ , which makes the task inherently multi-modal. + +This task is challenging because the goal is to model the sound field of a dynamic scene and capture acoustic changes between viewpoints given one pair of audio-visual measurements. While traditional signal processing methods can be applied, we show in Sec. 6 that they perform poorly. In this work, we present a learning-based rendering approach. + +# 4. Datasets + +We introduce two datasets for the NVAS task: live recordings (Sec. 4.1), and simulated audio in scanned real-world environments (Sec. 4.2) (see Fig. 2). The former is + +![](images/2628d579c955d748b0b80d69315ee94ece60d745a3e22cf3d1514ae0be9e9f22.jpg) + +![](images/3eea53311068a0871c2aac1bdae156a0a5a63306f53e32a414c01d79fa28ccae.jpg) +Figure 2. Example source and target views for the two introduced datasets: Replay-NVAS (left) and SoundSpaces-NVAS (right). + +real and covers various social scenarios, but offers limited diversity of sound sources, viewpoints and environments, and is noisy. The latter has a realism gap, but allows perfect control over these aforementioned elements. + +Both datasets focus on human speech given its relevance in applications. However, our model design is not specific to speech. For both datasets, we capture binaural audio, which best aligns with human perception. Note that for both datasets, we collect multiple multi-modal views for training and evaluation; during inference the target viewpoint(s) (and in some cases target environment) are withheld. + +# 4.1. The Replay-NVAS Dataset + +Replay-NVAS contains multi-view captures of acted scenes in apartments. We capture 46 different scenarios (e.g., having a conversation, having dinner, or doing yoga) from 8 different viewpoints. In total, we collect 37 hours of video data, involving 32 participants across all scenarios. + +In each scenario, we invite 2-4 participants to act on a given topic. Each participant wears a near-range microphone, providing a clean recording of their own speech. The scene is captured by 8 DLSR cameras, each augmented with a 3Dio binaural microphone. In this way, the data captures video and audio simultaneously from multiple cameras, resulting in 56 possible source/target viewpoint combinations for each scene. The videos are recorded at 30 FPS and the audio is recorded with a 48k sampling rate. We use a clapper at the beginning of the recording for temporal synchronization. Each scenario lasts $3 - 8\mathrm{min}$ . We use off-the-shelf software for multi-view camera calibration (see Supp.). + +To construct the dataset, we extract one-second long clips from each video with overlapping windows. We automatically remove silent and noisy clips based on the energy of near-range microphones, which results in 77K/12K/2K clips in total for train/val/test (details in Supp.) During training, for one sample, we randomly select two out of eight viewpoints, one as the source and one as the target. + +This dataset is very challenging. It covers a wide range of social activities. It is harrowed by ambient sound, room reverberation, overlapping speech and non-verbal sounds such as clapping and instruments. Participants can move + +freely in the environment. We believe that this data will be useful to the community beyond the NVAS task as it can be used for benchmarking many other problems, including active speaker localization, source separation, and NVS. + +# 4.2. The SoundSpaces-NVAS Dataset + +In this dataset, we synthesize multi-view audio-visual data of two people having conversations in 3D scenes. In total, we construct 1.3K hours of audio-visual data for a total of 1,000 speakers, 120 3D scenes and 200K viewpoints. + +Our goal is to construct audio-visual data with strong spatial and acoustic correspondences across multiple viewpoints, meaning that the visual information should indicate what the audio should sound like, e.g., observing speaker on the left should indicate the left ear is louder and observing speaker at a distance should indicate there is higher reverberation. We use the SoundSpaces 2.0 platform [7], which allows highly realistic audio and visual rendering for arbitrary camera and microphone locations in 3D scans of real-world environments [5, 52, 57]. It accounts for all major real-world acoustics phenomena: direct sounds, early specular/diffuse reflections, reverberation, binaural spatialization, and effects from materials and air absorption. + +We use the Gibson dataset [57] for scene meshes and LibriSpeech [41] for speech samples. As we are simulating two people having conversations, for a given environment, we randomly sample two speaker locations within $3\mathrm{m}$ and insert two copyright-free mannequins (one male and one female) at these two locations. We then randomly sample four nearby viewpoints facing the center of the two speakers at a height of $1.5\mathrm{m}$ (Fig. 2, right). For each speaker, we select a speech sample from LibriSpeech with matching gender. We render images at all locations as well as binaural impulse response for all pairs of points between speakers and viewpoints. The received sound is obtained by convolving the binaural impulse response with the speech sample. + +During training, for one sample, we randomly sample two out of four rendered viewpoints, one as the source and one as the target. We also randomly choose one speaker to be active, simulating what we observe on the real data (i.e., usually only one person speaks at a time). + +# 5. Visually-Guided Acoustic Synthesis + +We introduce a new method, Visually-Guided Acoustic Synthesis (ViGAS), to address the NVAS problem, taking as input sound and an image and outputting the sound from a different target microphone pose. + +ViGAS consists of five components: ambient sound separation, active speaker localization, visual acoustic network, acoustic synthesis, and temporal alignment. The high-level idea is to separate the observed sound into primary and ambient, extract useful visual information (active speaker and + +acoustic features), and use this information to guide acoustic synthesis for the primary sound. Temporal alignment is performed during training for better optimization. ViGAS is discussed in detail next and summarised in Fig. 3. + +# 5.1. Ambient Sound Separation + +ViGAS starts by decomposing the input sound into primary and ambient (traffic, electric noise from a fridge or the A/C, etc.). Ambient sound is important for realism, but it also interferes with learning the model because it can carry significant energy, making the model focus on it rather than on the primary sounds, and its spatial distribution is very different from the primary sounds. + +By explicitly separating primary and ambient sounds, ViGAS: (1) accounts for the fact that the transfer functions of primary and ambient sounds are very different and thus difficult to model together; (2) avoids wasting representational power on modelling ambient sounds that might be difficult to reconstruct accurately and depend less on the viewpoint; and (3) prevents ambient sounds, which are noise-like and high-energy, from dominating learning and reconstruction. In practice, as we show in Sec. 6, without the ambient sound separation, the model performs poorly. + +The goal of ambient sound separation is thus to construct a function $(A_C, A_N) = \mathcal{P}(A_S)$ that separates the input sound $A_S$ into primary sound $A_C$ and ambient sound $A_N$ . Existing approaches to this problem are based on signal processing [3, 13] or learning [11, 15]. We find that pretrained speech enhancement models such as Denoiser [11] tend to aggressively remove the noise including the primary sound, which hinders re-synthesis. We thus opt for band-pass filtering, passing frequencies within a certain range and rejecting/attenuating frequencies outside of it, which we found to work well. We cut frequencies below $80\mathrm{Hz}$ for SoundSpaces-NVAS and $150\mathrm{Hz}$ for Replay-NVAS. + +# 5.2. Active Speaker Localization + +Knowing where the emitters of different primary sounds are located in the environment can help to solve the NVAS task. In this paper, we focus on localizing the active speaker, although there can be other important primary sound events like instruments playing, speakers interacting with objects, etc. The goal of active speaker localization is to predict the bounding box of the active speaker in each frame of the video (examples in Fig. 4). The bounding box is in the format of $(y_{\mathrm{min}}, y_{\mathrm{max}}, x_{\mathrm{min}}, x_{\mathrm{max}})$ and $x, y$ are normalized to $[0,1]$ by the image width and height, respectively. + +On SoundSpaces-NVAS, this task is relatively easy because of the strong correspondence between the appearance of the speaker and the gender of the speech sample, which enables to easily train a classifier for active speakers. However, this is much harder on Replay-NVAS because cameras record speakers from a distance and from diverse angles, + +![](images/72393b9726dd6915942889143168835f6ee7c48b10da29be1ad64de8c0608488.jpg) +Figure 3. Visually Guided Acoustic Synthesis (ViGAS). Given the input audio $A_{S}$ , we first separate out the ambient sound to focus on the sound of interest. We take the source audio and source visual to localize the active speaker on the 2D image. We also extract the visual acoustic features of the environment by running an encoder on the source visual. We concatenate the active speaker feature, source visual features, and the target pose, and fuse these features with a MLP. We feed both the audio stream $A_{C}$ and fused visual feature $V_{C}$ into the acoustic synthesis network, which has $M$ stacked audio-visual fusion blocks. In each block, the audio sequence is processed by dilated conv1d layers and the visual features are processed by conv1d layers. Lastly, the previously separated ambient sound is added back to the waveform. During training, our temporal alignment module shifts the prediction by the amount of delay estimated between the source and the target audio to align the prediction well with the target. + +meaning that lip motion, the main cue used by speaker localization methods [21, 49, 54], is often not visible. Hence, the model has to rely on other cues to identify the speaker (such as body motion, gender or identity). Furthermore, sometimes people speak or laugh over each other. + +Since our focus is not speaker localization, for the Replay-NVAS we assume that this problem is solved by an external module that does audio-visual active speaker localization. To approximate the output of such a module automatically, we rely on the near-range audio recordings. Specifically, we first run an off-the-shelf detection and tracker [10] on the video at 5 FPS and obtain, with some manual refinement, bounding boxes $B_{t}^{i}$ for $i = 1,\dots ,N$ at each frame $t$ . We manually assign the near-range microphone audio $A_N^i$ to each tracked person. We select the active speaker $D$ based on the maximum energy of each near-range microphone, i.e., $D = \mathrm{argmax}_i\left\{\sum A_N^i [t:t + \Delta t]^2\right\}$ , where $\Delta t$ is the time interval we use to calculate the audio energy. We output bounding box $B^{D}$ as the localization feature $V_{L}$ . + +# 5.3. Visual Acoustic Network and Fusion + +The active speaker bounding box $B^{D}$ only disambiguates the active speaker from all visible humans on 2D, which is not enough to indicate where the speaker is in 3D. To infer that, the visual information is also needed. Since there is usually not much movement in one second (the length of the input video clip), the video clip does not provide much extra information compared to a single frame. Thus, we choose the middle frame to represent the clip and extract the visual acoustic features $V_{E}$ from the input RGB image with a pretrained ResNet18 [18] before the average pooling layer to preserve spatial information. To reduce the + +feature size, we feed $V_{E}$ into a 1D convolution with kernel size 1 and output channel size 8. We then flatten the visual features to obtain feature $V_{F}$ . + +The target pose is specified as the translation along $x, y, z$ axes plus difference between orientations of the source "view" and the target "view" expressed via rotation angles: $+y$ (roll), $+x$ (pitch) and $+z$ (yaw). We encode each angle $\alpha$ as its sinusoidal value: $(\sin(\alpha), \cos(\alpha))$ . + +Similarly, the target pose is not enough by itself to indicate where the target viewpoint $T$ is in the 3D space; to infer that, the source view $V_{S}$ is again needed. For example, in top row of Fig 4, for target viewpoint 3, "two meters to the right and one meter forward" is not enough to indicate the target location is in the corridor, while the model can reason that based on the source view. + +We use a fusion network to predict a latent representation of the scene variables $S, T, L^{D}, E$ (cf. Sec. 3) by first concatenating $[V_{L}, P_{T}, V_{F}]$ and then feeding it through a multilayer perceptron (MLP). See Fig. 3 for the network. + +# 5.4. Acoustic Synthesis + +With the separated primary sound $A_{C}$ and the visual acoustic feature $V_{C}$ as input, the goal of the acoustic synthesis module is to transform $A_{C}$ guided by $V_{C}$ . We design the acoustic synthesis network to learn a non-linear transfer function (implicitly) that captures these major acoustic phenomena, including the attenuation of sound in space, the directivity of sound sources (human speech is directional), the reverberation level, the head-related transfer function, as well as the frequency-dependent acoustic phenomena. Training end-to-end makes it possible to capture these subtle and complicated changes in the audio. + +Inspired by recent advances in time-domain signal mod + +eling [39, 48], we design the network as $M$ stacked synthesis blocks, where each block consists of multiple conv1D layers. We first encode the input audio $A_{C}$ into a latent space, which is then fed into the synthesis block. The key of the synthesis block is a gated multimodal fusion network that injects the visual information into the audio as follows: + +$$ +z = \tanh \left(p _ {A} ^ {k} \left(A _ {F} ^ {k}\right) + p _ {V} ^ {k} \left(V _ {C}\right)\right) \odot \sigma \left(q _ {A} ^ {k} \left(A _ {F} ^ {k}\right) + q _ {V} ^ {k} \left(V _ {C}\right)\right), \tag {2} +$$ + +where $\odot$ indicates element-wise multiplication, $\sigma$ is a logistic sigmoid function, $k = 1,\dots ,M$ is the layer index and $p,q$ are both learnable 1D convolutions. + +After passing $z$ through a sinusoidal activation function, the network uses two separate conv1D layers to process the feature, one producing the residual connection $A_F^{k + 1}$ and one producing the skip connection $A_P^{k + 1}$ . All skip connections $A_P^{k + 1}$ are mean pooled and fed into a decoder to produce the output $A_O$ . We add back the separated ambient sound $A_N$ as the target audio estimate: $\hat{A}_T = A_O + A_N$ . See Supp. for more details on the architecture. + +# 5.5. Temporal Alignment + +In order for the model to learn well, it is important that input and output sounds are temporally aligned. While the Replay-NVAS data is already synchronised based on the clapper sound, due to the finite speed of sound, the sounds emitted from different locations may still arrive at microphones with a delay slightly different from the one of the clapper, causing misalignments that affect training. + +To align source and target audio for training, we find the delay $\tau$ that maximizes the generalized cross-correlation: + +$$ +\mathcal {R} _ {A _ {S}, A _ {T}} (\tau) = \mathbb {E} _ {t} [ h _ {S} (t) \cdot h _ {T} (t - \tau) ], \tag {3} +$$ + +where $h_S$ and $h_T$ are the feature embedding for $A_S$ and $A_T$ respectively at time $t$ . We use the feature extractor $h$ from the generalized cross-correlation phase transform (GCC-PHAT) algorithm [23], which whitens the audio by dividing by the magnitude of the cross-power spectral density. After computing $\tau$ , we shift the prediction $A_O$ by $\tau$ samples to align with the $A_T$ and obtain $A_L$ . Note that alignment is already exact for SoundSpaces-NVAS. + +# 5.6. Loss + +To compute the loss, we first encode the audio with the short-time Fourier transform (STFT), a complex-valued matrix representation of the audio where the $y$ axis represents frequency and the $x$ axis is time. We then compute the magnitude of the STFT, and optimize the L1 loss between the predicted and ground truth magnitudes as follows: + +$$ +L = \left| \left| \left| S T F T \left(A _ {L}\right) \right| \right| _ {2} - \left| \left| S T F T \left(A _ {T} ^ {\prime}\right) \right| \right| _ {2} \right|, \tag {4} +$$ + +where $A_T^\prime$ is the primary sound separated from $A_{T}$ with $\mathcal{P}(\cdot)$ . By taking the magnitude, we do not model the exact phase values, which we find hinders learning if being included in the loss. See implementation details in Supp. + +# 6. Experiments + +We compare with several traditional and learning-based baselines and show that ViGAS outperforms them in both a quantitative evaluation and a human subject study. + +Evaluation. We measure performance from three aspects: 1. closeness to GT as measured by the magnitude spectrogram distance (Mag). 2. correctness of the spatial sound as measured by the left-right energy ratio error (LRE), i.e., the difference of ratio of energy between left and right channels and 3. correctness of the acoustic properties measured by RT60 error (RTE) [6, 50], i.e., the error in reverberation time decaying by 60dB (RT60). We use a pretrained model [6] to estimate RT60 directly from speech. + +Baselines. We consider the following baselines: 1. Input audio. Copying the input to the output. 2. TF Estimator [56] + Nearest Neighbor, i.e. storing the transfer function estimated during training and retrieving the nearest neighbor during test time. We estimate transfer functions with a Wiener filter [56] and index them with the ground-truth locations of the speaker, source viewpoint, and target viewpoint for the single environment setup and their relative pose for the novel environment setup. At test time, this method searches the database to find the nearest transfer function and applies it on the input audio. 3. Digital Signal Processing (DSP) [9] approach that takes the distance, azimuth, and elevation of the sound source, applies an inverse a head-related transfer function (HRTF) to estimate the speech spoken by the speaker and then applies another HRTF to estimate the audio at the target microphone location. This baseline adjusts the loudness of the left and right channels based on where the speaker is in the target view. We supply GT coordinates for SoundSpaces-NVAS and speakers' head positions estimated with triangulation on Replay-NVAS. 4. Visual Acoustic Matching (VAM) [6], recently proposed for a related task of matching acoustics of input audio with a target image. This task only deals with single viewpoint and single-channel audio. We adapt their model with minimal modification by feeding in the image from the source viewpoint and concatenating the position offset of the target microphone at the multimodal fusion step. See Supp. for details. + +# 6.1. Results on SoundSpaces-NVAS + +Table 1 shows the results. For synthetic data, we consider two evaluation setups: 1. single environment: train and test on the same environment and 2. novel environment: train and test on multiple non-overlapping Gibson environments (90/10/20 for train/val/test). + +In the single environment setup, our model largely outperforms all baselines as well as our audio-only ablation on all metrics. TF Estimator performs poorly despite being indexed by the ground truth location values because es + +
SoundSpaces-NVASReplay-NVAS
Single EnvironmentNovel EnvironmentSingle Environment
MagLRERTEMagLRERTEMagLRERTE
Input audio0.2251.4730.0320.2161.4080.0390.1591.4770.046
TF Estimator [56]0.3592.5960.0590.4403.2610.0920.3272.8610.147
DSP [9]0.3023.6440.0440.3003.6890.0470.4631.3000.067
VAM [6]0.2201.1980.0410.2351.1310.0510.1610.9240.070
ViGAS w/o visual0.1730.9730.0310.1811.0070.0360.1460.8770.046
ViGAS0.1590.7820.0290.1750.9710.0340.1420.7160.048
+ +timating a transfer function directly from two audio clips is non-trivial and noisy for low-energy parts of the signal. DSP also performs badly despite having the ground truth 3D coordinates of the sound source. This is because head related transfer functions are typically recorded in anechoic chambers, which does not account for acoustics of different environments, e.g., reverberation. Both traditional approaches perform worse than simply copying the input audio, indicating that learning-based models are needed for this challenging task. The recent model VAM [6] performs much better compared to the traditional approaches but still underperforms our model. There is a significant difference between ViGAS w/o visual and the full model; this shows that the visual knowledge about the speaker location and the environment is important for this task. + +Fig. 4 shows an example where given the same input source viewpoint, our model synthesizes audio for three different target viewpoints. The model reasons about how the geometry and speaker locations changes based on the source view and the target pose, and predicts the acoustic difference accordingly. See Supp. video to listen to sounds. + +For the novel environment setup, our model again outperforms all baselines. Compared to ViGAS in the single environment setup, both the magnitude spectrogram distance and the left-right energy ratio error increase. This is expected because for novel (unseen) environments, single images capture limited geometry and acoustic information. The model fails sometime when there is a drastic viewpoint change, e.g., target viewpoint 3 in Fig. 4. This setup requires the model to reason or "imagine" the environment based on single audio-visual observation, which poses great challenge for NVAS as well as NVS, where typically synthesis is performed in a fully observed environment. + +Ablations. Table 2 shows ablations on the model design. To understand if the model uses visual information, we ablate the visual features $V_{F}$ and the active speaker feature + +Table 1. Results on SoundSpaces-NVAS and Replay-NVAS. We report the magnitude spectrogram distance (Mag), left-right energy ratio error (LRE), and RT60 error (RTE). Replay-NVAS does not have novel environment setup due to data being collected in a single environment. For all metrics, lower is better. In addition to baselines, we also evaluate ViGAS w/o visual by removing the active speaker localization and visual features. Note that reverberation time is mostly invariant of the receiver location in the same room and thus input audio has low RTE. A good model should preserve this property while synthesizing the desired acoustics for the target viewpoint. + +
ViGASSS-NVASReplay-NVAS
MagLREMagLRE
full model0.1590.7820.1420.716
w/o visual features0.1710.8970.1460.920
w/o ASL0.1610.8140.1430.757
w/o alignment0.1760.7710.1440.706
w/o separation0.1650.8400.1820.859
+ +Table 2. Ablations of the model on both datasets. + +$V_{L}$ . Removing the active speaker feature leads to less damage on the model performance, because without the explicitly localized active speaker, the model can still implicitly reason about the active speaker location based on the image and audio. If both are removed ("ViGAS w/o visual" in Table 1), the performance suffers most. + +To study the effectiveness of the temporal alignment and ambient sound separation modules, we ablate them separately. Removing the temporal alignment leads to higher Mag error and slightly lower LRE. As for ambient sound separation, the results show that optimizing for the high-energy noise-like ambient sound degrades the performance. + +# 6.2. Results on Replay-NVAS + +Table 1 (right) shows the Replay-NVAS results. Compared to SoundSpaces-NVAS, the magnitudes of all errors are smaller because there are less drastic acoustic changes between viewpoints (8 DLSR cameras form a circle around the participants). Traditional approaches like TF Estimator and DSP still perform poorly despite using the 3D coordinates of the camera and the speaker (triangulated from multiple cameras). VAM performs better due to end-to-end learning; however, our model outperforms it. Compared to ViGAS w/o visual, the full model has much lower left-right energy ratio error and slightly higher reverberation time error, showing that the model takes into account the speaker position and viewpoint change for synthesizing the audio. + +![](images/bba20e54029ad3b75f1ad6e549e9eaead118c8bcbb197e3a42d4641063558db8.jpg) + +![](images/2a3bc9b8f482259d3bc533a3c7fe27659926713b905a5cd64fa54a22e9625ed5.jpg) + +![](images/dbdcf8565c1173fa0eefc44219c9345796f5258ee97785a72dac4cae43e3738b.jpg) + +![](images/244a341a624ed1e79c438f2cd502663b30e347453447777ce98071dcbc56d713.jpg) + +![](images/e3e5bb3bea372403391cf8fc8e1cf880e10e35f98d61924e445ea189442d3301.jpg) +Figure 4. Qualitative examples. For all binaural audio, we show the left-channel and the right-channel waveforms side-by-side. Row 1: SoundSpaces-NVAS example where given the source viewpoint and input audio, the model synthesizes audio for three different target viewpoints (target views are for reference only). In this case, the active speaker is the male speaker as indicated by the bounding box. For target viewpoint 1, the view rotates about 90 degrees and the male speaker is on the left side and the predicted left channel is louder than the right channel. Viewpoint 2 moves away from the speaker and thus yields lower amplitude compared to the first prediction. For target viewpoint 3, it is completely located outside of the living room, in which case, the sound could only come from the door open on the right (louder right channel) and the reverberation also greatly increases due to the vanishing direct sound. Row 2: Replay-NVAS example where the speaker is located on the left in the source viewpoint which becomes the right and further from the camera in target viewpoint 2, the model also predicts lower amplitude and louder right channel. On the right side, we show an example of the audio-visual speech enhancement for the active speaker. The model enhances the speech to largely match with the near-range audio (target). + +![](images/c600a1a81ad1afc4eaa470f46a004ce4c3f4ebf6124aabeb9a5b3f381b4d0796.jpg) + +![](images/fdec0cd93464eb57ac3729b5a726b7c8bff74bdfa670c61ebe913fefb85e8165.jpg) + +
MagRTE
Input0.2790.376
ViGAS (ours)0.2340.122
+ +Fig. 4 (row 2, left) shows a qualitative example. In the source viewpoint, the active speaker is on the left, while in the target viewpoint, he is further from the camera and on the right. The model synthesizes an audio waveform that captures the corresponding acoustic change, showing that our model successfully learns from real videos. + +Audio-visual speech enhancement. In some real-world applications, e.g., hearing aid devices, the goal is to obtain the enhanced clean speech of the active speaker. This can be seen as a special case of NVAS, where the target viewpoint is the active speaker. Our model is capable of performing audio-visual speech enhancement without any modification. We simply set the target audio to the near-range audio recording for the active speaker. We show the results in Table 3. Our model obtains cleaner audio compared to the input audio (example in Fig. 4, row 2, right). + +Human subject study. To supplement the quantitative metrics and evaluate how well our synthesized audio captures the acoustic change between viewpoints, we conduct a human subject study. We show participants the image of the target viewpoint $V_{T}$ as well as the audio $A_{T}$ as reference. We provide three audio samples: the input, the prediction of ViGAS, and the prediction of DSP (the most nat + +Table 3. Speech enhancement on Replay-NVAS. + +
DatasetInputDSPViGAS
SoundSpaces-NVAS24%2%74%
Replay-NVAS43%6%51%
+ +Table 4. Human Study. Participants favor our approach over the two most realistic sounding baselines, (1) copying the input signal, and (2) a digital signal processing baseline. + +urally sounding baseline) and ask them to select a clip that sounds closest to the target audio. We select 20 examples from SoundSpaces-NVAS and 20 examples from Replay-NVAS and invite 10 participants to perform the study. + +See Table 4 for the results. On the synthetic dataset SoundSpaces-NVAS, our approach is preferred over the baselines by a large margin. This margin is lower on the real-world Replay-NVAS dataset but is still significant. + +# 7. Conclusion + +We introduce the challenging novel-view acoustic synthesis task and a related benchmark in form of both real and synthetic datasets. We propose a neural rendering model that learns to transform the sound from the source viewpoint to the target viewpoint by reasoning about the observed audio and visual stream. Our model surpasses all baselines on both datasets. We believe this research unlocks many potential applications and research in multimodal novel-view synthesis. In the future, we plan to incorporate active-speaker localization model into the approach and let the model jointly learn to localize and synthesize. + +# References + +[1] Triantafyllos Afouras, Joon Son Chung, and Andrew Zisserman. The conversation: Deep audio-visual speech enhancement. In INTERSPEECH, 2018. 3 +[2] Humam Alwassel, Dhruv Mahajan, Bruno Korbar, Lorenzo Torresani, Bernard Ghanem, and Du Tran. Self-supervised learning by cross-modal audio-video clustering. In NeurIPS, 2020. 3 +[3] M. Berouti, Richard Schwartz, and John Makhoul. Enhancement of speech corrupted by acoustic noise. In IEEE International Conference on Acoustics, Speech, and Signal Processing, 1979. 4 +[4] J.S. Bradley. Review of objective room acoustics measures and future needs. Applied Acoustics, 2011. 2 +[5] Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niessner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from rgb-d data in indoor environments. 3DV, 2017. MatterPort3D dataset license available at: http://kaldir.vc.in.tum.de/matterport/MP_TOS.pdf.4 +[6] Changan Chen, Ruohan Gao, Paul Calamia, and Kristen Grauman. Visual acoustic matching. In CVPR, 2022. 2, 6, 7 +[7] Changan Chen, Carl Schissler, Sanchit Garg, Philip Kobernik, Alexander Clegg, Paul Calamia, Dhruv Batra, Philip W Robinson, and Kristen Grauman. Soundspaces 2.0: A simulation platform for visual-acoustic learning. In NeurIPS 2022 Datasets and Benchmarks Track, 2022. 2, 4 +[8] Changan Chen, Wei Sun, David Harwath, and Kristen Grauman. Learning audio-visual dereverberation. In ICASSP, 2023. 2 +[9] Corey I. Cheng and Gregory H. Wakefield. Introduction to head-related transfer functions (hrtfs): representations of hrtfs in time, frequency, and space. journal of the audio engineering society, 49(4):231-249, april 2001. 6, 7 +[10] MMTracking Contributors. MMTracking: OpenMMLab video perception toolbox and benchmark. https://github.com/open-mmlab/mmtracking, 2020.5 +[11] Alexandre Defossez, Gabriel Synnaeve, and Yossi Adi. Real time speech enhancement in the waveform domain. In *Interspeech*, 2020. 4 +[12] James Eaton, Nikolay Gaubitch, Allistair Moore, and Patrick Naylor. Estimation of room acoustic parameters: The ACE challenge. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 24(10), 2016. 2 +[13] Yariv Ephraim and Harry L. Van Trees. A signal subspace approach for speech enhancement. IEEE Transactions on Speech and Audio Processing, 1995. 4 +[14] Ariel Ephrat, Inbar Mosseri, Oran Lang, Tali Dekel, Kevin Wilson, Avinatan Hassidim, William T Freeman, and Michael Rubinstein. Looking to listen at the cocktail party: A speaker-independent audio-visual model for speech separation. In SIGGRAPH, 2018. 3 +[15] Szu-Wei Fu, Chien-Feng Liao, Yu Tsao, and Shou-De Lin. Metricgan: Generative adversarial networks based black-box metric scores optimization for speech enhancement. In International Conference on Machine Learning (ICML), 2019. 4 + +[16] Hannes Gamper and Ivan J Tashev. Blind reverberation time estimation using a convolutional neural network. In 2018 16th International Workshop on Acoustic Signal Enhancement (IWAENC), pages 136-140, 2018. 2 +[17] Ruohan Gao and Kristen Grauman. 2.5d visual sound. In CVPR, 2019. 2 +[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 5 +[19] Di Hu, Rui Qian, Minyue Jiang, Xiao Tan, Shilei Wen, Errui Ding, Weiyao Lin, and Dejing Dou. Discriminative sounding objects localization via self-supervised audiovisual matching. In NeurIPS, 2020. 3 +[20] Ajay Jain, Matthew Tancik, and Pieter Abbeel. Putting nerf on a diet: Semantically consistent few-shot view synthesis. In Proc. ICCV, 2021. 2 +[21] Hao Jiang, Calvin Murdock, and Vamsi Krishna Ithapu. Ego-centric deep multi-channel audio-visual active speaker localization. In CVPR, 2022. 3, 5 +[22] Florian Klein, Annika Neidhardt, and Marius Seipel. Real-time estimation of reverberation time for selection of suitable binaural room impulse responses. In Audio for Virtual, Augmented and Mixed Realities: Proceedings of 5th International Conference on Spatial Audio (ICSA), pages 145–150, 2019. 2 +[23] C. Knapp and G. Carter. The generalized correlation method for estimation of time delay. IEEE Transactions on Acoustics, Speech, and Signal Processing, 1976. 6 +[24] Bruno Korbar, Du Tran, and Lorenzo Torresani. Cooperative learning of audio and video models from self-supervised synchronization. In NeurIPS, 2018. 3 +[25] Jonas Kulhanek, Erik Derner, Torsten Sattler, and Robert Babuška. ViewFormer: NeRF-free neural rendering from few images using transformers. In Proc. ECCV, 2022. 2 +[26] Heinrich Kuttruff. Room Acoustics. Boca Raton, 6th edition, 2016. 2 +[27] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. arXiv.cs, abs/2011.13084, 2020. 2 +[28] Lingjie Liu, Marc Habermann, Viktor Rudnev, Kripasindhu Sarkar, Jiatao Gu, and Christian Theobalt. Neural humans: Pose-controlled free-view synthesis of human actors with template-guided neural radiance fields. In arXiv, 2021. 2 +[29] Andrew Luo, Yilun Du, Michael J Tarr, Joshua B Tenenbaum, Antonio Torralba, and Chuang Gan. Learning neural acoustic fields. In NeurIPS, 2022. 3 +[30] Wolfgang Mack, Shuwen Deng, and Emanuele AP Habets. Single-channel blind direct-to-reverberation ratio estimation using masking. In INTERSPEECH, pages 5066-5070, 2020. 2 +[31] Sagnik Majumder, Changan Chen, Ziad Al-Halah, and Kristen Grauman. Few-shot audio-visual learning of environment acoustics. In Thirty-Sixth Conference on Neural Information Processing Systems, 2022. 3 +[32] Ricardo Martin-Brualla, Noha Radwan, Mehdi S. M. Sajjadi, Jonathan T. Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In Proceedings of the IEEE/CVF + +Conference on Computer Vision and Pattern Recognition (CVPR), pages 7210-7219, June 2021. 1 +[33] Daniel Michelsanti, Zheng-Hua Tan, Shi-Xiong Zhang, Yong Xu, Meng Yu, Dong Yu, and Jesper Jensen. An overview of deep-learning-based audio-visual speech enhancement and separation. In arXiv, 2020. 3 +[34] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2 +[35] Pedro Morgado, Yi Li, and Nuno Vasconcelos. Learning representations from audio-visual spatial alignment. In NeurIPS, 2020. 3 +[36] Pedro Morgado, Nono Vasconcelos, Timothy Langlois, and Oliver Wang. Self-supervised generation of spatial audio for 360 video. In NeurIPS, 2018. 2 +[37] Prateek Murgai, Mark Rau, and Jean-Marc Jot. Blind estimation of the reverberation fingerprint of unknown acoustic environments. In Audio Engineering Society Convention 143. Audio Engineering Society, 2017. 2 +[38] Michael Niemeyer, Jonathan T. Barron, Ben Mildenhall, Mehdi S. M. Sajjadi, Andreas Geiger, and Noha Radwan. RegNeRF: Regularizing neural radiance fields for view synthesis from sparse inputs. In Proc. CVPR, 2022. 2 +[39] Aaron van den Oord, Sander Dieleman, Heiga Zen, Karen Simonyan, Oriol Vinyls, Alex Graves, Nal Kalchbrenner, Andrew Senior, and Koray Kavukcuoglu. Wavenet: A generative model for raw audio, 2016. cite arxiv:1609.03499. 6 +[40] Andrew Owens and Alexei A Efros. Audio-visual scene analysis with self-supervised multisensory features. In ECCV, 2018. 3 +[41] Vassil Panayotov, Guoguo Chen, Daniel Povey, and Sanjeev Khudanpur. Librispeech: an asr corpus based on public domain audio books. In 2015 IEEE international conference on acoustics, speech and signal processing (ICASSP), pages 5206-5210. IEEE, 2015. 4 +[42] Keunhong Park, Utkarsh Sinha, Jonathan T. Barron, Sofien Bouaziz, Dan B. Goldman, Steven M. Seitz, and Ricardo Martin-Brualla. Deformable neural radiance fields. CoRR, abs/2011.12948, 2020. 2 +[43] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-NeRF: Neural radiance fields for dynamic scenes. arXiv.cs, abs/2011.13961, 2020. 2 +[44] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10318–10327, June 2021. 1 +[45] Anton Ratnarajah, Zhenyu Tang, Rohith Chandrashekar Aralikatti, and Dinesh Manocha. Mesh2ir: Neural acoustic impulse response generator for complex 3d scenes. In ACM Multimedia, 2022. 3 +[46] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common Objects in 3D: Large-scale learning and evaluation of real-life 3D category reconstruction. In Proc. CVPR, 2021. 2 + +[47] Alexander Richard, Peter Dodds, and Vamsi Krishna Ithapu. Deep impulse responses: Estimating and parameterizing filters with deep networks. In IEEE International Conference on Acoustics, Speech and Signal Processing, 2022. 2 +[48] Alexander Richard, Dejan Markovic, Israel D Gebru, Steven Krenn, Gladstone Butler, Fernando de la Torre, and Yaser Sheikh. Neural synthesis of binaural speech from mono audio. In International Conference on Learning Representations, 2021. 2, 6 +[49] Joseph Roth, Sourish Chaudhuri, Ondrej Klejch, Radhika Marvin, Andrew Gallagher, Liat Kaver, Sharadh Ramaswamy, Arkadiusz Stopczynski, Cordelia Schmid, Zhonghua Xi, and Caroline Pantofaru. Ava-activespeaker: An audio-visual dataset for active speaker detection. In ICASSP, 2020. 5 +[50] Nikhil Singh, Jeff Mentch, Jerry Ng, Matthew Beveridge, and Iddo Drori. Image2reverb: Cross-modal reverb impulse response synthesis. In ICCV, 2021. 2, 6 +[51] Vincent Sitzmann, Michael Zollhöfer, and Gordon Wetzstein. Scene representation networks: Continuous 3D-structure-aware neural scene representations. Proc. NeurIPS, 2019. 2 +[52] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, et al. The Replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019.4 +[53] Shih-Yang Su, Frank Yu, Michael Zollhöfer, and Helge Rhodin. A-NeRF: Surface-free human 3d pose refinement via neural rendering. arXiv.cs, abs/2102.06199, 2021. 2 +[54] Ruijie Tao, Zexu Pan, Rohan Kumar Das, Xinyuan Qian, Mike Zheng Shou, and Haizhou Li. Is someone speaking? exploring long-term temporal features for audio-visual active speaker detection. In Proceedings of the 29th ACM International Conference on Multimedia, page 3927–3935, 2021. 3, 5 +[55] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Christoph Lassner, and Christian Theobalt. Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a deforming scene from monocular video. arXiv.cs, abs/2012.12247, 2020. 2 +[56] Norbert Wiener. Extrapolation, interpolation, and smoothing of stationary time series. Report of the Services 19, Research Project DIC-6037 MIT, 1942. 6, 7 +[57] Fei Xia, Amir R. Zamir, Zhi-Yang He, Alexander Sax, Jitendra Malik, and Silvio Savarese. Gibson Env: real-world perception for embodied agents. In Computer Vision and Pattern Recognition (CVPR), 2018 IEEE Conference on. IEEE, 2018. 4 +[58] Feifei Xiong, Stefan Goetze, Birger Kollmeier, and Bernd T Meyer. Joint estimation of reverberation time and early-to-late reverberation ratio from single-channel speech signals. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 27(2):255-267, 2018. 2 +[59] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. PixelNeRF: Neural radiance fields from one or few images. In Proc. CVPR, 2021. 2 + +[60] Hang Zhou, Ziwei Liu, Xudong Xu, Ping Luo, and Xiaogang Wang. Vision-infused deep audio inpainting. In ICCV, 2019. 3 \ No newline at end of file diff --git a/2023/Novel-View Acoustic Synthesis/images.zip b/2023/Novel-View Acoustic Synthesis/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..bec429f9ec60ca51801fe2931ca50e66524178e2 --- /dev/null +++ b/2023/Novel-View Acoustic Synthesis/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:060810f770434aff2dd4e08fdf2a1766abf089347acc9a68bf7a3093941437ce +size 425065 diff --git a/2023/Novel-View Acoustic Synthesis/layout.json b/2023/Novel-View Acoustic Synthesis/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..62ab94e4cfc6021719ffa4725f96b832c0f97241 --- /dev/null +++ b/2023/Novel-View Acoustic Synthesis/layout.json @@ -0,0 +1,9144 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 201, + 103, + 392, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 103, + 392, + 121 + ], + "spans": [ + { + "bbox": [ + 201, + 103, + 392, + 121 + ], + "type": "text", + "content": "Novel-View Acoustic Synthesis" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 79, + 142, + 510, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 142, + 510, + 190 + ], + "spans": [ + { + "bbox": [ + 79, + 142, + 510, + 190 + ], + "type": "text", + "content": "Changan Chen" + }, + { + "bbox": [ + 79, + 142, + 510, + 190 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 79, + 142, + 510, + 190 + ], + "type": "text", + "content": " Alexander Richard" + }, + { + "bbox": [ + 79, + 142, + 510, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 79, + 142, + 510, + 190 + ], + "type": "text", + "content": " Roman Shapovalov" + }, + { + "bbox": [ + 79, + 142, + 510, + 190 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 79, + 142, + 510, + 190 + ], + "type": "text", + "content": " Vamsi Krishna Ithapu" + }, + { + "bbox": [ + 79, + 142, + 510, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 79, + 142, + 510, + 190 + ], + "type": "text", + "content": " Natalia Neverova" + }, + { + "bbox": [ + 79, + 142, + 510, + 190 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 79, + 142, + 510, + 190 + ], + "type": "text", + "content": " Kristen Grauman" + }, + { + "bbox": [ + 79, + 142, + 510, + 190 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 79, + 142, + 510, + 190 + ], + "type": "text", + "content": " Andrea Vedaldi" + }, + { + "bbox": [ + 79, + 142, + 510, + 190 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 79, + 142, + 510, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 79, + 142, + 510, + 190 + ], + "type": "text", + "content": "University of Texas at Austin " + }, + { + "bbox": [ + 79, + 142, + 510, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 79, + 142, + 510, + 190 + ], + "type": "text", + "content": "Reality Labs Research at Meta " + }, + { + "bbox": [ + 79, + 142, + 510, + 190 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 79, + 142, + 510, + 190 + ], + "type": "text", + "content": "FAIR, Meta AI" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 216, + 192, + 228 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 216, + 192, + 228 + ], + "spans": [ + { + "bbox": [ + 143, + 216, + 192, + 228 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 240, + 290, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 240, + 290, + 445 + ], + "spans": [ + { + "bbox": [ + 46, + 240, + 290, + 445 + ], + "type": "text", + "content": "We introduce the novel-view acoustic synthesis (NVAS) task: given the sight and sound observed at a source viewpoint, can we synthesize the sound of that scene from an unseen target viewpoint? We propose a neural rendering approach: Visually-Guided Acoustic Synthesis (ViGAS) network that learns to synthesize the sound of an arbitrary point in space by analyzing the input audio-visual cues. To benchmark this task, we collect two first-of-their-kind large-scale multi-view audio-visual datasets, one synthetic and one real. We show that our model successfully reasons about the spatial cues and synthesizes faithful audio on both datasets. To our knowledge, this work represents the very first formulation, dataset, and approach to solve the novel-view acoustic synthesis task, which has exciting potential applications ranging from AR/VR to art and design. Unlocked by this work, we believe that the future of novel-view synthesis is in multi-modal learning from videos." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 463, + 128, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 463, + 128, + 475 + ], + "spans": [ + { + "bbox": [ + 47, + 463, + 128, + 475 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 479, + 287, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 479, + 287, + 563 + ], + "spans": [ + { + "bbox": [ + 46, + 479, + 287, + 563 + ], + "type": "text", + "content": "Replaying a video recording from a new viewpoint1 has many applications in cinematography, video enhancement, and virtual reality. For example, it can be used to edit a video, simulate a virtual camera, or, given a video of a personal memory, even enable users to experience a treasured moment again—not just on a 2D screen, but in 3D in a virtual or augmented reality, thus ‘reliving’ the moment." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 563, + 287, + 671 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 563, + 287, + 671 + ], + "spans": [ + { + "bbox": [ + 46, + 563, + 287, + 671 + ], + "type": "text", + "content": "While the applications are exciting, there are still many unsolved technical challenges. Recent advances in 3D reconstruction and novel-view synthesis (NVS) address the problem of synthesizing new images of a given scene [32, 34, 44]. However, thus far, the view synthesis problem is concerned with creating visuals alone; the output is silent or at best naively adopts the sounds of the original video (from the \"wrong\" viewpoint). Without sound, the emotional and cognitive significance of the replay is severely diminished." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 672, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 672, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 47, + 672, + 287, + 696 + ], + "type": "text", + "content": "In this work, we address this gap and introduce the new task of novel-view acoustic synthesis (NVAS). The goal of" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 308, + 216, + 545, + 358 + ], + "blocks": [ + { + "bbox": [ + 308, + 216, + 545, + 358 + ], + "lines": [ + { + "bbox": [ + 308, + 216, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 308, + 216, + 545, + 358 + ], + "type": "image", + "image_path": "a0b9f4beb4e51b7fc4434efef3758801200bf5222e33d70bd4a94612aaa30df4.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 365, + 546, + 421 + ], + "lines": [ + { + "bbox": [ + 305, + 365, + 546, + 421 + ], + "spans": [ + { + "bbox": [ + 305, + 365, + 546, + 421 + ], + "type": "text", + "content": "Figure 1. Novel-view acoustic synthesis task. Given audio-visual observations from one viewpoint and the relative target viewpoint pose, render the sound received at the target viewpoint. Note that the target is expressed as the desired pose of the microphones; the image at that pose (right) is neither observed nor synthesized." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 437, + 546, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 437, + 546, + 473 + ], + "spans": [ + { + "bbox": [ + 305, + 437, + 546, + 473 + ], + "type": "text", + "content": "this task is to synthesize the sound in a scene from a new acoustic viewpoint, given only the visual and acoustic input from another source viewpoint in the same scene (Fig. 1)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 474, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 547, + 715 + ], + "type": "text", + "content": "NVAS is very different from the existing NVS task, where the goal is to reconstruct images instead of sounds, and these differences present new challenges. First, the 3D geometry of most real-life scenes changes in a limited manner during the recording. On the contrary, sound changes substantially over time, so the reconstruction target is highly dynamic. Secondly, visual and audio sensors are very different. A camera matrix captures the light in a highly-directional manner, and a single image comprises a large 2D array of pixels. In contrast, sounds are recorded with one or two microphones which are at best weakly-directional, providing only a coarse sampling of the sound field. Thirdly, the frequency of light waves is much higher than that of sound waves; the length of audio waves is thus larger to the point of being comparable to the size of geometric features of the scene, meaning that effects such as diffraction are often dominant, and spatial resolution is low. As a result, techniques that require spatial precision, such as triangulation and segmentation, are not applicable to audio. Lastly, sounds mix together, making it difficult to segment them," + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 254, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 254, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 254, + 713 + ], + "type": "text", + "content": "1We use \"viewpoint\" to mean a camera or microphone pose." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6409" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 96 + ], + "type": "text", + "content": "and they are affected by environmental effects such as reverberation that are distributed and largely unobservable." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 99, + 288, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 99, + 288, + 289 + ], + "spans": [ + { + "bbox": [ + 46, + 99, + 288, + 289 + ], + "type": "text", + "content": "While the NVS and NVAS tasks are indeed very different, we hypothesize that NVAS is an inherently multimodal task. In fact, vision can play an important role in achieving accurate sound synthesis. First, establishing correspondences between sounds and their sources as they appear in images can provide essential cues for resynthesizing the sounds realistically. For instance, human speech is highly directional and sounds very differently if one faces the speaker or their back, which can only be inferred from visual cues. In addition, the environment acoustics also affect the sound one hears as a function of the scene geometry, materials, and emitter/receiver locations. The same source sounds vary differently if it is located in the center of a room, at the corner, or in a corridor, for example. In short, vision provides cues about space and geometry that affect sound, and are difficult to estimate from the sound alone." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 293, + 287, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 293, + 287, + 435 + ], + "spans": [ + { + "bbox": [ + 46, + 293, + 287, + 435 + ], + "type": "text", + "content": "In order to validate our hypothesis, we propose a novel visually-guided acoustic synthesis network that analyzes audio and visual features and synthesizes the audio at a target location. More specifically, the network first takes as input the image observed at the source viewpoint in order to infer global acoustic and geometric properties of the environment along with the bounding box of the active speaker. The network then reasons how the speaker and scene geometry change in 3D based on the relative target pose with a fusion network. We inject the fused features into audio with a gated multi-modal fusion network and model the acoustic changes between viewpoints with a time-domain model." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 438, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 438, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 438, + 288, + 714 + ], + "type": "text", + "content": "In order to conduct our experiments on the new NVAS task, we require suitable training and benchmarking data, of which currently there is none available. To address that, we contribute two new datasets: one real (Replay-NVAS) and one synthetic (SoundSpaces-NVAS). The key feature of these datasets is to record the sight and sound of different scenes from multiple cameras/viewpoints. Replay-NVAS contains video recordings of groups of people performing social activities (e.g., chatting, watching TV, doing yoga, playing instruments) from 8 surrounding viewpoints simultaneously. It contains 37 hours of highly realistic everyday conversation and social interactions in one homelike environment. To our knowledge, Replay-NVAS represents the first large-scale real-world dataset enabling NVAS. This dataset would also greatly benefit many other existing tasks including NVS, active speaker localization, etc. For SoundSpaces-NVAS, we render 1.3K hours of audiovisual data based on the SoundSpaces [7] platform. Using this simulator, one can easily change the scene geometry and the positions of speakers, cameras, and microphones. This data serves as a powerful test bed with clean ground truth for a large collection of home environments, offering a good complement to Replay-NVAS. For both datasets," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": "we capture binaural audio, which is what humans perceive with two ears. Together the datasets contain 1,337 hours of audio-visual capture, with 1,032 speakers across 121 3D scenes. Datasets are publicly available for future research." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 121, + 546, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 121, + 546, + 240 + ], + "spans": [ + { + "bbox": [ + 304, + 121, + 546, + 240 + ], + "type": "text", + "content": "We show that our model outperforms traditional signal processing approaches as well as learning-based baselines, often by a substantial margin, in a quantitative evaluation and a human study. We show qualitative examples where the model predicts acoustic changes according to the viewpoint changes, e.g., left channel becomes louder when the viewpoint changes from left to right. In a nutshell, we present the first work that deals with novel-view acoustic synthesis, and contribute two large-scale datasets along with a novel neural rendering approach for solving the task." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 247, + 392, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 247, + 392, + 258 + ], + "spans": [ + { + "bbox": [ + 306, + 247, + 392, + 258 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 263, + 545, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 263, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 304, + 263, + 545, + 358 + ], + "type": "text", + "content": "Novel-view synthesis (NVS). Kickstarted by advances in neural rendering [34, 51], many recent works consider variants of the NVS problem. Most approaches assume dozens of calibrated images for reconstructing a single static scene. Closer to monocular video NVS, authors have considered reducing the number of input views [20, 25, 38, 46, 59] and modelling dynamic scenes [27, 28, 42, 43, 53, 55]. However, none of these works tackle audio." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 361, + 546, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 361, + 546, + 552 + ], + "spans": [ + { + "bbox": [ + 304, + 361, + 546, + 552 + ], + "type": "text", + "content": "Acoustic matching and spatialization. NVAS requires accounting for (1) the environmental acoustics and (2) the geometric configuration of the target microphone(s) (e.g., monaural vs binaural). Modelling environmental acoustics has been addressed extensively by the audio community [4, 26]. Room impulse response (RIR) functions characterize the environment acoustics as a transfer function between the emitter and receiver, accounting for the scene geometry, materials, and emitter/receiver locations. Estimating the direct-to-reverberant ratio and the reverberation time, is sufficient to synthesize simple RIRs that match audio in a plausible manner [12, 16, 22, 30, 37, 58]. These methods do not synthesize for a target viewpoint, rather they resynthesize to match an audio sample. In [47, 48] sound from a moving emitter is spatialized towards a receiver conditioned on the tracked 3D location of the emitter." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 553, + 547, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 553, + 547, + 696 + ], + "spans": [ + { + "bbox": [ + 304, + 553, + 547, + 696 + ], + "type": "text", + "content": "Recently, the vision community explores using visual information to estimate environmental acoustics [6, 8, 50]. However, these works only synthesize acoustics for a given viewpoint rather than a novel viewpoint. In addition, they have only addressed monaural audio, which is more forgiving than binaural because humans do not perceive subtle absolute acoustic properties, but can detect easily inconsistencies in the sounds perceived by the two ears. Recent work spatializes monaural sounds by upmixing them to multiple channels conditioned on the video, where the sound emitters are static [17, 36]. Because the environment, emitter and receiver are static, so are the acoustics. Other work" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 315, + 702, + 477, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 702, + 477, + 712 + ], + "spans": [ + { + "bbox": [ + 315, + 702, + 477, + 712 + ], + "type": "text", + "content": "2https://replay-dataset.github.io" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6410" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": "predicts impulse responses in simulation either for a single environment [29], or by using few-shot egocentric observations [31], or by using the 3D scene mesh [45]. While simulated results are satisfying, those models' impact on real-world data is unknown, especially for scenarios where human speakers move and interact with each other. Unlike any of the above, we introduce and tackle the NVAS problem, accounting for both acoustics and spatialization, and we propose a model that addresses the problem effectively on both synthetic and real-world data." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 194, + 289, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 194, + 289, + 291 + ], + "spans": [ + { + "bbox": [ + 46, + 194, + 289, + 291 + ], + "type": "text", + "content": "Audio-visual learning. Recent advances in multi-modal video understanding enable new forms of self-supervised cross-modal feature learning from video [2, 24, 35], sound source localization [19, 21, 54], and audio-visual speech enhancement and source separation [1, 14, 33, 40, 60]. All of these existing tasks and datasets only deal with a single viewpoint. We introduce the first audio-visual learning task and dataset that deals with multi-view audio-visual data." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 302, + 266, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 302, + 266, + 316 + ], + "spans": [ + { + "bbox": [ + 47, + 302, + 266, + 316 + ], + "type": "text", + "content": "3. The Novel-view Acoustic Synthesis Task" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 323, + 289, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 323, + 289, + 455 + ], + "spans": [ + { + "bbox": [ + 46, + 323, + 289, + 455 + ], + "type": "text", + "content": "We introduce a new task, novel-view acoustic synthesis (NVAS). Assuming there are " + }, + { + "bbox": [ + 46, + 323, + 289, + 455 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 323, + 289, + 455 + ], + "type": "text", + "content": " sound emitters in the scene (emitter " + }, + { + "bbox": [ + 46, + 323, + 289, + 455 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 323, + 289, + 455 + ], + "type": "text", + "content": " emits sound " + }, + { + "bbox": [ + 46, + 323, + 289, + 455 + ], + "type": "inline_equation", + "content": "C^i" + }, + { + "bbox": [ + 46, + 323, + 289, + 455 + ], + "type": "text", + "content": " from location " + }, + { + "bbox": [ + 46, + 323, + 289, + 455 + ], + "type": "inline_equation", + "content": "L^i" + }, + { + "bbox": [ + 46, + 323, + 289, + 455 + ], + "type": "text", + "content": "), given the audio " + }, + { + "bbox": [ + 46, + 323, + 289, + 455 + ], + "type": "inline_equation", + "content": "A_{S}" + }, + { + "bbox": [ + 46, + 323, + 289, + 455 + ], + "type": "text", + "content": " and video " + }, + { + "bbox": [ + 46, + 323, + 289, + 455 + ], + "type": "inline_equation", + "content": "V_{S}" + }, + { + "bbox": [ + 46, + 323, + 289, + 455 + ], + "type": "text", + "content": " observed at the source viewpoint " + }, + { + "bbox": [ + 46, + 323, + 289, + 455 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 323, + 289, + 455 + ], + "type": "text", + "content": ", the goal is to synthesize the audio " + }, + { + "bbox": [ + 46, + 323, + 289, + 455 + ], + "type": "inline_equation", + "content": "A_{T}" + }, + { + "bbox": [ + 46, + 323, + 289, + 455 + ], + "type": "text", + "content": " at the target viewpoint " + }, + { + "bbox": [ + 46, + 323, + 289, + 455 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 323, + 289, + 455 + ], + "type": "text", + "content": ", as it would sound from the target location, specified by the relative pose " + }, + { + "bbox": [ + 46, + 323, + 289, + 455 + ], + "type": "inline_equation", + "content": "P_T" + }, + { + "bbox": [ + 46, + 323, + 289, + 455 + ], + "type": "text", + "content": " of the target microphone (translation and orientation) with respect to the source view (Fig. 1). Furthermore, we assume that the active sound emitters in the environment are visible in the source camera, but we make no assumptions about the camera at the target location." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 455, + 269, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 455, + 269, + 467 + ], + "spans": [ + { + "bbox": [ + 59, + 455, + 269, + 467 + ], + "type": "text", + "content": "The sound at any point " + }, + { + "bbox": [ + 59, + 455, + 269, + 467 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 59, + 455, + 269, + 467 + ], + "type": "text", + "content": " is a function of the space:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 95, + 468, + 287, + 483 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 468, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 95, + 468, + 287, + 483 + ], + "type": "interline_equation", + "content": "A _ {R} = \\mathcal {F} \\left(L ^ {1, \\dots , N}, C ^ {1, \\dots , N}, R \\mid E\\right), \\tag {1}", + "image_path": "108bde995dd521707d90043e4b3ee65bc8b4a7cd1a7ceae60297887a2ed1deaf.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 489, + 287, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 489, + 287, + 572 + ], + "spans": [ + { + "bbox": [ + 46, + 489, + 287, + 572 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 489, + 287, + 572 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 46, + 489, + 287, + 572 + ], + "type": "text", + "content": " is the receiver location (" + }, + { + "bbox": [ + 46, + 489, + 287, + 572 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 489, + 287, + 572 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 46, + 489, + 287, + 572 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 489, + 287, + 572 + ], + "type": "text", + "content": ") and " + }, + { + "bbox": [ + 46, + 489, + 287, + 572 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 46, + 489, + 287, + 572 + ], + "type": "text", + "content": " is the environment. The emitted sounds " + }, + { + "bbox": [ + 46, + 489, + 287, + 572 + ], + "type": "inline_equation", + "content": "C^i" + }, + { + "bbox": [ + 46, + 489, + 287, + 572 + ], + "type": "text", + "content": " are not restricted to speech but can be ambient noise, sounding objects, etc. Our goal here is to learn a transfer function " + }, + { + "bbox": [ + 46, + 489, + 287, + 572 + ], + "type": "inline_equation", + "content": "\\mathcal{T}(\\cdot)" + }, + { + "bbox": [ + 46, + 489, + 287, + 572 + ], + "type": "text", + "content": " defined as " + }, + { + "bbox": [ + 46, + 489, + 287, + 572 + ], + "type": "inline_equation", + "content": "A_{T} = \\mathcal{T}(A_{S}, V_{S}, P_{T})" + }, + { + "bbox": [ + 46, + 489, + 287, + 572 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 489, + 287, + 572 + ], + "type": "inline_equation", + "content": "S, T, L^{1,\\dots,N}, C^{1,\\dots,N}, E" + }, + { + "bbox": [ + 46, + 489, + 287, + 572 + ], + "type": "text", + "content": " are not directly given and need to be inferred from " + }, + { + "bbox": [ + 46, + 489, + 287, + 572 + ], + "type": "inline_equation", + "content": "V_{S}" + }, + { + "bbox": [ + 46, + 489, + 287, + 572 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 489, + 287, + 572 + ], + "type": "inline_equation", + "content": "P_{T}" + }, + { + "bbox": [ + 46, + 489, + 287, + 572 + ], + "type": "text", + "content": ", which makes the task inherently multi-modal." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 573, + 287, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 573, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 46, + 573, + 287, + 645 + ], + "type": "text", + "content": "This task is challenging because the goal is to model the sound field of a dynamic scene and capture acoustic changes between viewpoints given one pair of audio-visual measurements. While traditional signal processing methods can be applied, we show in Sec. 6 that they perform poorly. In this work, we present a learning-based rendering approach." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 657, + 107, + 669 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 657, + 107, + 669 + ], + "spans": [ + { + "bbox": [ + 47, + 657, + 107, + 669 + ], + "type": "text", + "content": "4. Datasets" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 677, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 289, + 715 + ], + "type": "text", + "content": "We introduce two datasets for the NVAS task: live recordings (Sec. 4.1), and simulated audio in scanned real-world environments (Sec. 4.2) (see Fig. 2). The former is" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 318, + 70, + 533, + 130 + ], + "blocks": [ + { + "bbox": [ + 318, + 70, + 533, + 130 + ], + "lines": [ + { + "bbox": [ + 318, + 70, + 533, + 130 + ], + "spans": [ + { + "bbox": [ + 318, + 70, + 533, + 130 + ], + "type": "image", + "image_path": "2628d579c955d748b0b80d69315ee94ece60d745a3e22cf3d1514ae0be9e9f22.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 318, + 131, + 533, + 191 + ], + "blocks": [ + { + "bbox": [ + 318, + 131, + 533, + 191 + ], + "lines": [ + { + "bbox": [ + 318, + 131, + 533, + 191 + ], + "spans": [ + { + "bbox": [ + 318, + 131, + 533, + 191 + ], + "type": "image", + "image_path": "3eea53311068a0871c2aac1bdae156a0a5a63306f53e32a414c01d79fa28ccae.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 193, + 545, + 216 + ], + "lines": [ + { + "bbox": [ + 305, + 193, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 305, + 193, + 545, + 216 + ], + "type": "text", + "content": "Figure 2. Example source and target views for the two introduced datasets: Replay-NVAS (left) and SoundSpaces-NVAS (right)." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 221, + 545, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 221, + 545, + 269 + ], + "spans": [ + { + "bbox": [ + 304, + 221, + 545, + 269 + ], + "type": "text", + "content": "real and covers various social scenarios, but offers limited diversity of sound sources, viewpoints and environments, and is noisy. The latter has a realism gap, but allows perfect control over these aforementioned elements." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 270, + 545, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 270, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 304, + 270, + 545, + 353 + ], + "type": "text", + "content": "Both datasets focus on human speech given its relevance in applications. However, our model design is not specific to speech. For both datasets, we capture binaural audio, which best aligns with human perception. Note that for both datasets, we collect multiple multi-modal views for training and evaluation; during inference the target viewpoint(s) (and in some cases target environment) are withheld." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 360, + 454, + 373 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 360, + 454, + 373 + ], + "spans": [ + { + "bbox": [ + 306, + 360, + 454, + 373 + ], + "type": "text", + "content": "4.1. The Replay-NVAS Dataset" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 378, + 545, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 378, + 545, + 438 + ], + "spans": [ + { + "bbox": [ + 304, + 378, + 545, + 438 + ], + "type": "text", + "content": "Replay-NVAS contains multi-view captures of acted scenes in apartments. We capture 46 different scenarios (e.g., having a conversation, having dinner, or doing yoga) from 8 different viewpoints. In total, we collect 37 hours of video data, involving 32 participants across all scenarios." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 438, + 545, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 438, + 545, + 582 + ], + "spans": [ + { + "bbox": [ + 304, + 438, + 545, + 582 + ], + "type": "text", + "content": "In each scenario, we invite 2-4 participants to act on a given topic. Each participant wears a near-range microphone, providing a clean recording of their own speech. The scene is captured by 8 DLSR cameras, each augmented with a 3Dio binaural microphone. In this way, the data captures video and audio simultaneously from multiple cameras, resulting in 56 possible source/target viewpoint combinations for each scene. The videos are recorded at 30 FPS and the audio is recorded with a 48k sampling rate. We use a clapper at the beginning of the recording for temporal synchronization. Each scenario lasts " + }, + { + "bbox": [ + 304, + 438, + 545, + 582 + ], + "type": "inline_equation", + "content": "3 - 8\\mathrm{min}" + }, + { + "bbox": [ + 304, + 438, + 545, + 582 + ], + "type": "text", + "content": ". We use off-the-shelf software for multi-view camera calibration (see Supp.)." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 582, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 665 + ], + "type": "text", + "content": "To construct the dataset, we extract one-second long clips from each video with overlapping windows. We automatically remove silent and noisy clips based on the energy of near-range microphones, which results in 77K/12K/2K clips in total for train/val/test (details in Supp.) During training, for one sample, we randomly select two out of eight viewpoints, one as the source and one as the target." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 545, + 713 + ], + "type": "text", + "content": "This dataset is very challenging. It covers a wide range of social activities. It is harrowed by ambient sound, room reverberation, overlapping speech and non-verbal sounds such as clapping and instruments. Participants can move" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "6411" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 120 + ], + "type": "text", + "content": "freely in the environment. We believe that this data will be useful to the community beyond the NVAS task as it can be used for benchmarking many other problems, including active speaker localization, source separation, and NVS." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 124, + 222, + 137 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 124, + 222, + 137 + ], + "spans": [ + { + "bbox": [ + 47, + 124, + 222, + 137 + ], + "type": "text", + "content": "4.2. The SoundSpaces-NVAS Dataset" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 138, + 287, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 138, + 287, + 186 + ], + "spans": [ + { + "bbox": [ + 46, + 138, + 287, + 186 + ], + "type": "text", + "content": "In this dataset, we synthesize multi-view audio-visual data of two people having conversations in 3D scenes. In total, we construct 1.3K hours of audio-visual data for a total of 1,000 speakers, 120 3D scenes and 200K viewpoints." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 186, + 287, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 186, + 287, + 341 + ], + "spans": [ + { + "bbox": [ + 46, + 186, + 287, + 341 + ], + "type": "text", + "content": "Our goal is to construct audio-visual data with strong spatial and acoustic correspondences across multiple viewpoints, meaning that the visual information should indicate what the audio should sound like, e.g., observing speaker on the left should indicate the left ear is louder and observing speaker at a distance should indicate there is higher reverberation. We use the SoundSpaces 2.0 platform [7], which allows highly realistic audio and visual rendering for arbitrary camera and microphone locations in 3D scans of real-world environments [5, 52, 57]. It accounts for all major real-world acoustics phenomena: direct sounds, early specular/diffuse reflections, reverberation, binaural spatialization, and effects from materials and air absorption." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 342, + 287, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 342, + 287, + 496 + ], + "spans": [ + { + "bbox": [ + 46, + 342, + 287, + 496 + ], + "type": "text", + "content": "We use the Gibson dataset [57] for scene meshes and LibriSpeech [41] for speech samples. As we are simulating two people having conversations, for a given environment, we randomly sample two speaker locations within " + }, + { + "bbox": [ + 46, + 342, + 287, + 496 + ], + "type": "inline_equation", + "content": "3\\mathrm{m}" + }, + { + "bbox": [ + 46, + 342, + 287, + 496 + ], + "type": "text", + "content": " and insert two copyright-free mannequins (one male and one female) at these two locations. We then randomly sample four nearby viewpoints facing the center of the two speakers at a height of " + }, + { + "bbox": [ + 46, + 342, + 287, + 496 + ], + "type": "inline_equation", + "content": "1.5\\mathrm{m}" + }, + { + "bbox": [ + 46, + 342, + 287, + 496 + ], + "type": "text", + "content": " (Fig. 2, right). For each speaker, we select a speech sample from LibriSpeech with matching gender. We render images at all locations as well as binaural impulse response for all pairs of points between speakers and viewpoints. The received sound is obtained by convolving the binaural impulse response with the speech sample." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 498, + 287, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 557 + ], + "type": "text", + "content": "During training, for one sample, we randomly sample two out of four rendered viewpoints, one as the source and one as the target. We also randomly choose one speaker to be active, simulating what we observe on the real data (i.e., usually only one person speaks at a time)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 567, + 244, + 581 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 567, + 244, + 581 + ], + "spans": [ + { + "bbox": [ + 47, + 567, + 244, + 581 + ], + "type": "text", + "content": "5. Visually-Guided Acoustic Synthesis" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 588, + 287, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 588, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 46, + 588, + 287, + 635 + ], + "type": "text", + "content": "We introduce a new method, Visually-Guided Acoustic Synthesis (ViGAS), to address the NVAS problem, taking as input sound and an image and outputting the sound from a different target microphone pose." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 636, + 287, + 695 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 636, + 287, + 695 + ], + "spans": [ + { + "bbox": [ + 46, + 636, + 287, + 695 + ], + "type": "text", + "content": "ViGAS consists of five components: ambient sound separation, active speaker localization, visual acoustic network, acoustic synthesis, and temporal alignment. The high-level idea is to separate the observed sound into primary and ambient, extract useful visual information (active speaker and" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "text", + "content": "acoustic features), and use this information to guide acoustic synthesis for the primary sound. Temporal alignment is performed during training for better optimization. ViGAS is discussed in detail next and summarised in Fig. 3." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 129, + 456, + 141 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 129, + 456, + 141 + ], + "spans": [ + { + "bbox": [ + 305, + 129, + 456, + 141 + ], + "type": "text", + "content": "5.1. Ambient Sound Separation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 147, + 545, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 147, + 545, + 231 + ], + "spans": [ + { + "bbox": [ + 304, + 147, + 545, + 231 + ], + "type": "text", + "content": "ViGAS starts by decomposing the input sound into primary and ambient (traffic, electric noise from a fridge or the A/C, etc.). Ambient sound is important for realism, but it also interferes with learning the model because it can carry significant energy, making the model focus on it rather than on the primary sounds, and its spatial distribution is very different from the primary sounds." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 232, + 545, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 232, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 304, + 232, + 545, + 350 + ], + "type": "text", + "content": "By explicitly separating primary and ambient sounds, ViGAS: (1) accounts for the fact that the transfer functions of primary and ambient sounds are very different and thus difficult to model together; (2) avoids wasting representational power on modelling ambient sounds that might be difficult to reconstruct accurately and depend less on the viewpoint; and (3) prevents ambient sounds, which are noise-like and high-energy, from dominating learning and reconstruction. In practice, as we show in Sec. 6, without the ambient sound separation, the model performs poorly." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 351, + 545, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 351, + 545, + 494 + ], + "spans": [ + { + "bbox": [ + 304, + 351, + 545, + 494 + ], + "type": "text", + "content": "The goal of ambient sound separation is thus to construct a function " + }, + { + "bbox": [ + 304, + 351, + 545, + 494 + ], + "type": "inline_equation", + "content": "(A_C, A_N) = \\mathcal{P}(A_S)" + }, + { + "bbox": [ + 304, + 351, + 545, + 494 + ], + "type": "text", + "content": " that separates the input sound " + }, + { + "bbox": [ + 304, + 351, + 545, + 494 + ], + "type": "inline_equation", + "content": "A_S" + }, + { + "bbox": [ + 304, + 351, + 545, + 494 + ], + "type": "text", + "content": " into primary sound " + }, + { + "bbox": [ + 304, + 351, + 545, + 494 + ], + "type": "inline_equation", + "content": "A_C" + }, + { + "bbox": [ + 304, + 351, + 545, + 494 + ], + "type": "text", + "content": " and ambient sound " + }, + { + "bbox": [ + 304, + 351, + 545, + 494 + ], + "type": "inline_equation", + "content": "A_N" + }, + { + "bbox": [ + 304, + 351, + 545, + 494 + ], + "type": "text", + "content": ". Existing approaches to this problem are based on signal processing [3, 13] or learning [11, 15]. We find that pretrained speech enhancement models such as Denoiser [11] tend to aggressively remove the noise including the primary sound, which hinders re-synthesis. We thus opt for band-pass filtering, passing frequencies within a certain range and rejecting/attenuating frequencies outside of it, which we found to work well. We cut frequencies below " + }, + { + "bbox": [ + 304, + 351, + 545, + 494 + ], + "type": "inline_equation", + "content": "80\\mathrm{Hz}" + }, + { + "bbox": [ + 304, + 351, + 545, + 494 + ], + "type": "text", + "content": " for SoundSpaces-NVAS and " + }, + { + "bbox": [ + 304, + 351, + 545, + 494 + ], + "type": "inline_equation", + "content": "150\\mathrm{Hz}" + }, + { + "bbox": [ + 304, + 351, + 545, + 494 + ], + "type": "text", + "content": " for Replay-NVAS." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 502, + 460, + 515 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 502, + 460, + 515 + ], + "spans": [ + { + "bbox": [ + 305, + 502, + 460, + 515 + ], + "type": "text", + "content": "5.2. Active Speaker Localization" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 521, + 545, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 641 + ], + "type": "text", + "content": "Knowing where the emitters of different primary sounds are located in the environment can help to solve the NVAS task. In this paper, we focus on localizing the active speaker, although there can be other important primary sound events like instruments playing, speakers interacting with objects, etc. The goal of active speaker localization is to predict the bounding box of the active speaker in each frame of the video (examples in Fig. 4). The bounding box is in the format of " + }, + { + "bbox": [ + 304, + 521, + 545, + 641 + ], + "type": "inline_equation", + "content": "(y_{\\mathrm{min}}, y_{\\mathrm{max}}, x_{\\mathrm{min}}, x_{\\mathrm{max}})" + }, + { + "bbox": [ + 304, + 521, + 545, + 641 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 521, + 545, + 641 + ], + "type": "inline_equation", + "content": "x, y" + }, + { + "bbox": [ + 304, + 521, + 545, + 641 + ], + "type": "text", + "content": " are normalized to " + }, + { + "bbox": [ + 304, + 521, + 545, + 641 + ], + "type": "inline_equation", + "content": "[0,1]" + }, + { + "bbox": [ + 304, + 521, + 545, + 641 + ], + "type": "text", + "content": " by the image width and height, respectively." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": "On SoundSpaces-NVAS, this task is relatively easy because of the strong correspondence between the appearance of the speaker and the gender of the speech sample, which enables to easily train a classifier for active speakers. However, this is much harder on Replay-NVAS because cameras record speakers from a distance and from diverse angles," + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 57, + 702, + 252, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 702, + 252, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 702, + 252, + 712 + ], + "type": "text", + "content": "3https://renderpeople.com/free-3d-people" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6412" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 71, + 534, + 216 + ], + "blocks": [ + { + "bbox": [ + 61, + 71, + 534, + 216 + ], + "lines": [ + { + "bbox": [ + 61, + 71, + 534, + 216 + ], + "spans": [ + { + "bbox": [ + 61, + 71, + 534, + 216 + ], + "type": "image", + "image_path": "72393b9726dd6915942889143168835f6ee7c48b10da29be1ad64de8c0608488.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 222, + 547, + 312 + ], + "lines": [ + { + "bbox": [ + 46, + 222, + 547, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 222, + 547, + 312 + ], + "type": "text", + "content": "Figure 3. Visually Guided Acoustic Synthesis (ViGAS). Given the input audio " + }, + { + "bbox": [ + 46, + 222, + 547, + 312 + ], + "type": "inline_equation", + "content": "A_{S}" + }, + { + "bbox": [ + 46, + 222, + 547, + 312 + ], + "type": "text", + "content": ", we first separate out the ambient sound to focus on the sound of interest. We take the source audio and source visual to localize the active speaker on the 2D image. We also extract the visual acoustic features of the environment by running an encoder on the source visual. We concatenate the active speaker feature, source visual features, and the target pose, and fuse these features with a MLP. We feed both the audio stream " + }, + { + "bbox": [ + 46, + 222, + 547, + 312 + ], + "type": "inline_equation", + "content": "A_{C}" + }, + { + "bbox": [ + 46, + 222, + 547, + 312 + ], + "type": "text", + "content": " and fused visual feature " + }, + { + "bbox": [ + 46, + 222, + 547, + 312 + ], + "type": "inline_equation", + "content": "V_{C}" + }, + { + "bbox": [ + 46, + 222, + 547, + 312 + ], + "type": "text", + "content": " into the acoustic synthesis network, which has " + }, + { + "bbox": [ + 46, + 222, + 547, + 312 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 222, + 547, + 312 + ], + "type": "text", + "content": " stacked audio-visual fusion blocks. In each block, the audio sequence is processed by dilated conv1d layers and the visual features are processed by conv1d layers. Lastly, the previously separated ambient sound is added back to the waveform. During training, our temporal alignment module shifts the prediction by the amount of delay estimated between the source and the target audio to align the prediction well with the target." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 317, + 287, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 317, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 46, + 317, + 287, + 376 + ], + "type": "text", + "content": "meaning that lip motion, the main cue used by speaker localization methods [21, 49, 54], is often not visible. Hence, the model has to rely on other cues to identify the speaker (such as body motion, gender or identity). Furthermore, sometimes people speak or laugh over each other." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 377, + 289, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 377, + 289, + 558 + ], + "spans": [ + { + "bbox": [ + 46, + 377, + 289, + 558 + ], + "type": "text", + "content": "Since our focus is not speaker localization, for the Replay-NVAS we assume that this problem is solved by an external module that does audio-visual active speaker localization. To approximate the output of such a module automatically, we rely on the near-range audio recordings. Specifically, we first run an off-the-shelf detection and tracker [10] on the video at 5 FPS and obtain, with some manual refinement, bounding boxes " + }, + { + "bbox": [ + 46, + 377, + 289, + 558 + ], + "type": "inline_equation", + "content": "B_{t}^{i}" + }, + { + "bbox": [ + 46, + 377, + 289, + 558 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 46, + 377, + 289, + 558 + ], + "type": "inline_equation", + "content": "i = 1,\\dots ,N" + }, + { + "bbox": [ + 46, + 377, + 289, + 558 + ], + "type": "text", + "content": " at each frame " + }, + { + "bbox": [ + 46, + 377, + 289, + 558 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 377, + 289, + 558 + ], + "type": "text", + "content": ". We manually assign the near-range microphone audio " + }, + { + "bbox": [ + 46, + 377, + 289, + 558 + ], + "type": "inline_equation", + "content": "A_N^i" + }, + { + "bbox": [ + 46, + 377, + 289, + 558 + ], + "type": "text", + "content": " to each tracked person. We select the active speaker " + }, + { + "bbox": [ + 46, + 377, + 289, + 558 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 377, + 289, + 558 + ], + "type": "text", + "content": " based on the maximum energy of each near-range microphone, i.e., " + }, + { + "bbox": [ + 46, + 377, + 289, + 558 + ], + "type": "inline_equation", + "content": "D = \\mathrm{argmax}_i\\left\\{\\sum A_N^i [t:t + \\Delta t]^2\\right\\}" + }, + { + "bbox": [ + 46, + 377, + 289, + 558 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 377, + 289, + 558 + ], + "type": "inline_equation", + "content": "\\Delta t" + }, + { + "bbox": [ + 46, + 377, + 289, + 558 + ], + "type": "text", + "content": " is the time interval we use to calculate the audio energy. We output bounding box " + }, + { + "bbox": [ + 46, + 377, + 289, + 558 + ], + "type": "inline_equation", + "content": "B^{D}" + }, + { + "bbox": [ + 46, + 377, + 289, + 558 + ], + "type": "text", + "content": " as the localization feature " + }, + { + "bbox": [ + 46, + 377, + 289, + 558 + ], + "type": "inline_equation", + "content": "V_{L}" + }, + { + "bbox": [ + 46, + 377, + 289, + 558 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 563, + 241, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 563, + 241, + 575 + ], + "spans": [ + { + "bbox": [ + 47, + 563, + 241, + 575 + ], + "type": "text", + "content": "5.3. Visual Acoustic Network and Fusion" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": "The active speaker bounding box " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "B^{D}" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": " only disambiguates the active speaker from all visible humans on 2D, which is not enough to indicate where the speaker is in 3D. To infer that, the visual information is also needed. Since there is usually not much movement in one second (the length of the input video clip), the video clip does not provide much extra information compared to a single frame. Thus, we choose the middle frame to represent the clip and extract the visual acoustic features " + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "inline_equation", + "content": "V_{E}" + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": " from the input RGB image with a pretrained ResNet18 [18] before the average pooling layer to preserve spatial information. To reduce the" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 317, + 545, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 317, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 304, + 317, + 545, + 352 + ], + "type": "text", + "content": "feature size, we feed " + }, + { + "bbox": [ + 304, + 317, + 545, + 352 + ], + "type": "inline_equation", + "content": "V_{E}" + }, + { + "bbox": [ + 304, + 317, + 545, + 352 + ], + "type": "text", + "content": " into a 1D convolution with kernel size 1 and output channel size 8. We then flatten the visual features to obtain feature " + }, + { + "bbox": [ + 304, + 317, + 545, + 352 + ], + "type": "inline_equation", + "content": "V_{F}" + }, + { + "bbox": [ + 304, + 317, + 545, + 352 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 353, + 545, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 353, + 545, + 412 + ], + "spans": [ + { + "bbox": [ + 304, + 353, + 545, + 412 + ], + "type": "text", + "content": "The target pose is specified as the translation along " + }, + { + "bbox": [ + 304, + 353, + 545, + 412 + ], + "type": "inline_equation", + "content": "x, y, z" + }, + { + "bbox": [ + 304, + 353, + 545, + 412 + ], + "type": "text", + "content": " axes plus difference between orientations of the source \"view\" and the target \"view\" expressed via rotation angles: " + }, + { + "bbox": [ + 304, + 353, + 545, + 412 + ], + "type": "inline_equation", + "content": "+y" + }, + { + "bbox": [ + 304, + 353, + 545, + 412 + ], + "type": "text", + "content": " (roll), " + }, + { + "bbox": [ + 304, + 353, + 545, + 412 + ], + "type": "inline_equation", + "content": "+x" + }, + { + "bbox": [ + 304, + 353, + 545, + 412 + ], + "type": "text", + "content": " (pitch) and " + }, + { + "bbox": [ + 304, + 353, + 545, + 412 + ], + "type": "inline_equation", + "content": "+z" + }, + { + "bbox": [ + 304, + 353, + 545, + 412 + ], + "type": "text", + "content": " (yaw). We encode each angle " + }, + { + "bbox": [ + 304, + 353, + 545, + 412 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 353, + 545, + 412 + ], + "type": "text", + "content": " as its sinusoidal value: " + }, + { + "bbox": [ + 304, + 353, + 545, + 412 + ], + "type": "inline_equation", + "content": "(\\sin(\\alpha), \\cos(\\alpha))" + }, + { + "bbox": [ + 304, + 353, + 545, + 412 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 412, + 545, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 412, + 545, + 496 + ], + "spans": [ + { + "bbox": [ + 304, + 412, + 545, + 496 + ], + "type": "text", + "content": "Similarly, the target pose is not enough by itself to indicate where the target viewpoint " + }, + { + "bbox": [ + 304, + 412, + 545, + 496 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 412, + 545, + 496 + ], + "type": "text", + "content": " is in the 3D space; to infer that, the source view " + }, + { + "bbox": [ + 304, + 412, + 545, + 496 + ], + "type": "inline_equation", + "content": "V_{S}" + }, + { + "bbox": [ + 304, + 412, + 545, + 496 + ], + "type": "text", + "content": " is again needed. For example, in top row of Fig 4, for target viewpoint 3, \"two meters to the right and one meter forward\" is not enough to indicate the target location is in the corridor, while the model can reason that based on the source view." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 496, + 545, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 496, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 496, + 545, + 544 + ], + "type": "text", + "content": "We use a fusion network to predict a latent representation of the scene variables " + }, + { + "bbox": [ + 304, + 496, + 545, + 544 + ], + "type": "inline_equation", + "content": "S, T, L^{D}, E" + }, + { + "bbox": [ + 304, + 496, + 545, + 544 + ], + "type": "text", + "content": " (cf. Sec. 3) by first concatenating " + }, + { + "bbox": [ + 304, + 496, + 545, + 544 + ], + "type": "inline_equation", + "content": "[V_{L}, P_{T}, V_{F}]" + }, + { + "bbox": [ + 304, + 496, + 545, + 544 + ], + "type": "text", + "content": " and then feeding it through a multilayer perceptron (MLP). See Fig. 3 for the network." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 552, + 416, + 564 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 552, + 416, + 564 + ], + "spans": [ + { + "bbox": [ + 306, + 552, + 416, + 564 + ], + "type": "text", + "content": "5.4. Acoustic Synthesis" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 570, + 545, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 700 + ], + "type": "text", + "content": "With the separated primary sound " + }, + { + "bbox": [ + 304, + 570, + 545, + 700 + ], + "type": "inline_equation", + "content": "A_{C}" + }, + { + "bbox": [ + 304, + 570, + 545, + 700 + ], + "type": "text", + "content": " and the visual acoustic feature " + }, + { + "bbox": [ + 304, + 570, + 545, + 700 + ], + "type": "inline_equation", + "content": "V_{C}" + }, + { + "bbox": [ + 304, + 570, + 545, + 700 + ], + "type": "text", + "content": " as input, the goal of the acoustic synthesis module is to transform " + }, + { + "bbox": [ + 304, + 570, + 545, + 700 + ], + "type": "inline_equation", + "content": "A_{C}" + }, + { + "bbox": [ + 304, + 570, + 545, + 700 + ], + "type": "text", + "content": " guided by " + }, + { + "bbox": [ + 304, + 570, + 545, + 700 + ], + "type": "inline_equation", + "content": "V_{C}" + }, + { + "bbox": [ + 304, + 570, + 545, + 700 + ], + "type": "text", + "content": ". We design the acoustic synthesis network to learn a non-linear transfer function (implicitly) that captures these major acoustic phenomena, including the attenuation of sound in space, the directivity of sound sources (human speech is directional), the reverberation level, the head-related transfer function, as well as the frequency-dependent acoustic phenomena. Training end-to-end makes it possible to capture these subtle and complicated changes in the audio." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "type": "text", + "content": "Inspired by recent advances in time-domain signal mod" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6413" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": "eling [39, 48], we design the network as " + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": " stacked synthesis blocks, where each block consists of multiple conv1D layers. We first encode the input audio " + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "inline_equation", + "content": "A_{C}" + }, + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": " into a latent space, which is then fed into the synthesis block. The key of the synthesis block is a gated multimodal fusion network that injects the visual information into the audio as follows:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 52, + 148, + 287, + 163 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 148, + 287, + 163 + ], + "spans": [ + { + "bbox": [ + 52, + 148, + 287, + 163 + ], + "type": "interline_equation", + "content": "z = \\tanh \\left(p _ {A} ^ {k} \\left(A _ {F} ^ {k}\\right) + p _ {V} ^ {k} \\left(V _ {C}\\right)\\right) \\odot \\sigma \\left(q _ {A} ^ {k} \\left(A _ {F} ^ {k}\\right) + q _ {V} ^ {k} \\left(V _ {C}\\right)\\right), \\tag {2}", + "image_path": "19fe12bf7371893373af3a2a54eab35977000fe1c2657b08fb43f43cca46760c.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 169, + 287, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 169, + 287, + 205 + ], + "spans": [ + { + "bbox": [ + 46, + 169, + 287, + 205 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 169, + 287, + 205 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 46, + 169, + 287, + 205 + ], + "type": "text", + "content": " indicates element-wise multiplication, " + }, + { + "bbox": [ + 46, + 169, + 287, + 205 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 46, + 169, + 287, + 205 + ], + "type": "text", + "content": " is a logistic sigmoid function, " + }, + { + "bbox": [ + 46, + 169, + 287, + 205 + ], + "type": "inline_equation", + "content": "k = 1,\\dots ,M" + }, + { + "bbox": [ + 46, + 169, + 287, + 205 + ], + "type": "text", + "content": " is the layer index and " + }, + { + "bbox": [ + 46, + 169, + 287, + 205 + ], + "type": "inline_equation", + "content": "p,q" + }, + { + "bbox": [ + 46, + 169, + 287, + 205 + ], + "type": "text", + "content": " are both learnable 1D convolutions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 205, + 287, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 205, + 287, + 303 + ], + "spans": [ + { + "bbox": [ + 46, + 205, + 287, + 303 + ], + "type": "text", + "content": "After passing " + }, + { + "bbox": [ + 46, + 205, + 287, + 303 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 46, + 205, + 287, + 303 + ], + "type": "text", + "content": " through a sinusoidal activation function, the network uses two separate conv1D layers to process the feature, one producing the residual connection " + }, + { + "bbox": [ + 46, + 205, + 287, + 303 + ], + "type": "inline_equation", + "content": "A_F^{k + 1}" + }, + { + "bbox": [ + 46, + 205, + 287, + 303 + ], + "type": "text", + "content": " and one producing the skip connection " + }, + { + "bbox": [ + 46, + 205, + 287, + 303 + ], + "type": "inline_equation", + "content": "A_P^{k + 1}" + }, + { + "bbox": [ + 46, + 205, + 287, + 303 + ], + "type": "text", + "content": ". All skip connections " + }, + { + "bbox": [ + 46, + 205, + 287, + 303 + ], + "type": "inline_equation", + "content": "A_P^{k + 1}" + }, + { + "bbox": [ + 46, + 205, + 287, + 303 + ], + "type": "text", + "content": " are mean pooled and fed into a decoder to produce the output " + }, + { + "bbox": [ + 46, + 205, + 287, + 303 + ], + "type": "inline_equation", + "content": "A_O" + }, + { + "bbox": [ + 46, + 205, + 287, + 303 + ], + "type": "text", + "content": ". We add back the separated ambient sound " + }, + { + "bbox": [ + 46, + 205, + 287, + 303 + ], + "type": "inline_equation", + "content": "A_N" + }, + { + "bbox": [ + 46, + 205, + 287, + 303 + ], + "type": "text", + "content": " as the target audio estimate: " + }, + { + "bbox": [ + 46, + 205, + 287, + 303 + ], + "type": "inline_equation", + "content": "\\hat{A}_T = A_O + A_N" + }, + { + "bbox": [ + 46, + 205, + 287, + 303 + ], + "type": "text", + "content": ". See Supp. for more details on the architecture." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 309, + 167, + 322 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 309, + 167, + 322 + ], + "spans": [ + { + "bbox": [ + 47, + 309, + 167, + 322 + ], + "type": "text", + "content": "5.5. Temporal Alignment" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 328, + 287, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 328, + 287, + 411 + ], + "spans": [ + { + "bbox": [ + 46, + 328, + 287, + 411 + ], + "type": "text", + "content": "In order for the model to learn well, it is important that input and output sounds are temporally aligned. While the Replay-NVAS data is already synchronised based on the clapper sound, due to the finite speed of sound, the sounds emitted from different locations may still arrive at microphones with a delay slightly different from the one of the clapper, causing misalignments that affect training." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 411, + 287, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 411, + 287, + 435 + ], + "spans": [ + { + "bbox": [ + 46, + 411, + 287, + 435 + ], + "type": "text", + "content": "To align source and target audio for training, we find the delay " + }, + { + "bbox": [ + 46, + 411, + 287, + 435 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 411, + 287, + 435 + ], + "type": "text", + "content": " that maximizes the generalized cross-correlation:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 90, + 436, + 287, + 449 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 436, + 287, + 449 + ], + "spans": [ + { + "bbox": [ + 90, + 436, + 287, + 449 + ], + "type": "interline_equation", + "content": "\\mathcal {R} _ {A _ {S}, A _ {T}} (\\tau) = \\mathbb {E} _ {t} [ h _ {S} (t) \\cdot h _ {T} (t - \\tau) ], \\tag {3}", + "image_path": "c96d3e66091764c6d3f80802af13fa65e4be081c4bfb87bd447cdb7a181d4a56.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 453, + 287, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 453, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 46, + 453, + 287, + 548 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 453, + 287, + 548 + ], + "type": "inline_equation", + "content": "h_S" + }, + { + "bbox": [ + 46, + 453, + 287, + 548 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 453, + 287, + 548 + ], + "type": "inline_equation", + "content": "h_T" + }, + { + "bbox": [ + 46, + 453, + 287, + 548 + ], + "type": "text", + "content": " are the feature embedding for " + }, + { + "bbox": [ + 46, + 453, + 287, + 548 + ], + "type": "inline_equation", + "content": "A_S" + }, + { + "bbox": [ + 46, + 453, + 287, + 548 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 453, + 287, + 548 + ], + "type": "inline_equation", + "content": "A_T" + }, + { + "bbox": [ + 46, + 453, + 287, + 548 + ], + "type": "text", + "content": " respectively at time " + }, + { + "bbox": [ + 46, + 453, + 287, + 548 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 453, + 287, + 548 + ], + "type": "text", + "content": ". We use the feature extractor " + }, + { + "bbox": [ + 46, + 453, + 287, + 548 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 46, + 453, + 287, + 548 + ], + "type": "text", + "content": " from the generalized cross-correlation phase transform (GCC-PHAT) algorithm [23], which whitens the audio by dividing by the magnitude of the cross-power spectral density. After computing " + }, + { + "bbox": [ + 46, + 453, + 287, + 548 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 453, + 287, + 548 + ], + "type": "text", + "content": ", we shift the prediction " + }, + { + "bbox": [ + 46, + 453, + 287, + 548 + ], + "type": "inline_equation", + "content": "A_O" + }, + { + "bbox": [ + 46, + 453, + 287, + 548 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 46, + 453, + 287, + 548 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 453, + 287, + 548 + ], + "type": "text", + "content": " samples to align with the " + }, + { + "bbox": [ + 46, + 453, + 287, + 548 + ], + "type": "inline_equation", + "content": "A_T" + }, + { + "bbox": [ + 46, + 453, + 287, + 548 + ], + "type": "text", + "content": " and obtain " + }, + { + "bbox": [ + 46, + 453, + 287, + 548 + ], + "type": "inline_equation", + "content": "A_L" + }, + { + "bbox": [ + 46, + 453, + 287, + 548 + ], + "type": "text", + "content": ". Note that alignment is already exact for SoundSpaces-NVAS." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 555, + 91, + 566 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 555, + 91, + 566 + ], + "spans": [ + { + "bbox": [ + 47, + 555, + 91, + 566 + ], + "type": "text", + "content": "5.6. Loss" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 573, + 287, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 573, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 46, + 573, + 287, + 644 + ], + "type": "text", + "content": "To compute the loss, we first encode the audio with the short-time Fourier transform (STFT), a complex-valued matrix representation of the audio where the " + }, + { + "bbox": [ + 46, + 573, + 287, + 644 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 573, + 287, + 644 + ], + "type": "text", + "content": " axis represents frequency and the " + }, + { + "bbox": [ + 46, + 573, + 287, + 644 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 573, + 287, + 644 + ], + "type": "text", + "content": " axis is time. We then compute the magnitude of the STFT, and optimize the L1 loss between the predicted and ground truth magnitudes as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 85, + 648, + 287, + 662 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 648, + 287, + 662 + ], + "spans": [ + { + "bbox": [ + 85, + 648, + 287, + 662 + ], + "type": "interline_equation", + "content": "L = \\left| \\left| \\left| S T F T \\left(A _ {L}\\right) \\right| \\right| _ {2} - \\left| \\left| S T F T \\left(A _ {T} ^ {\\prime}\\right) \\right| \\right| _ {2} \\right|, \\tag {4}", + "image_path": "d9388813821635d52da8d84c65c397aaedf6717cf99de0d59408c54a66485f8c.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "inline_equation", + "content": "A_T^\\prime" + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": " is the primary sound separated from " + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "inline_equation", + "content": "A_{T}" + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{P}(\\cdot)" + }, + { + "bbox": [ + 46, + 665, + 287, + 713 + ], + "type": "text", + "content": ". By taking the magnitude, we do not model the exact phase values, which we find hinders learning if being included in the loss. See implementation details in Supp." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 72, + 388, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 388, + 85 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 388, + 85 + ], + "type": "text", + "content": "6. Experiments" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 91, + 545, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 545, + 128 + ], + "type": "text", + "content": "We compare with several traditional and learning-based baselines and show that ViGAS outperforms them in both a quantitative evaluation and a human subject study." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 131, + 545, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 131, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 304, + 131, + 545, + 239 + ], + "type": "text", + "content": "Evaluation. We measure performance from three aspects: 1. closeness to GT as measured by the magnitude spectrogram distance (Mag). 2. correctness of the spatial sound as measured by the left-right energy ratio error (LRE), i.e., the difference of ratio of energy between left and right channels and 3. correctness of the acoustic properties measured by RT60 error (RTE) [6, 50], i.e., the error in reverberation time decaying by 60dB (RT60). We use a pretrained model [6] to estimate RT60 directly from speech." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 241, + 546, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 241, + 546, + 578 + ], + "spans": [ + { + "bbox": [ + 304, + 241, + 546, + 578 + ], + "type": "text", + "content": "Baselines. We consider the following baselines: 1. Input audio. Copying the input to the output. 2. TF Estimator [56] + Nearest Neighbor, i.e. storing the transfer function estimated during training and retrieving the nearest neighbor during test time. We estimate transfer functions with a Wiener filter [56] and index them with the ground-truth locations of the speaker, source viewpoint, and target viewpoint for the single environment setup and their relative pose for the novel environment setup. At test time, this method searches the database to find the nearest transfer function and applies it on the input audio. 3. Digital Signal Processing (DSP) [9] approach that takes the distance, azimuth, and elevation of the sound source, applies an inverse a head-related transfer function (HRTF) to estimate the speech spoken by the speaker and then applies another HRTF to estimate the audio at the target microphone location. This baseline adjusts the loudness of the left and right channels based on where the speaker is in the target view. We supply GT coordinates for SoundSpaces-NVAS and speakers' head positions estimated with triangulation on Replay-NVAS. 4. Visual Acoustic Matching (VAM) [6], recently proposed for a related task of matching acoustics of input audio with a target image. This task only deals with single viewpoint and single-channel audio. We adapt their model with minimal modification by feeding in the image from the source viewpoint and concatenating the position offset of the target microphone at the multimodal fusion step. See Supp. for details." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 586, + 473, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 586, + 473, + 598 + ], + "spans": [ + { + "bbox": [ + 306, + 586, + 473, + 598 + ], + "type": "text", + "content": "6.1. Results on SoundSpaces-NVAS" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 605, + 545, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 664 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 664 + ], + "type": "text", + "content": "Table 1 shows the results. For synthetic data, we consider two evaluation setups: 1. single environment: train and test on the same environment and 2. novel environment: train and test on multiple non-overlapping Gibson environments (90/10/20 for train/val/test)." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "In the single environment setup, our model largely outperforms all baselines as well as our audio-only ablation on all metrics. TF Estimator performs poorly despite being indexed by the ground truth location values because es" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6414" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 75, + 70, + 516, + 195 + ], + "blocks": [ + { + "bbox": [ + 75, + 70, + 516, + 195 + ], + "lines": [ + { + "bbox": [ + 75, + 70, + 516, + 195 + ], + "spans": [ + { + "bbox": [ + 75, + 70, + 516, + 195 + ], + "type": "table", + "html": "
SoundSpaces-NVASReplay-NVAS
Single EnvironmentNovel EnvironmentSingle Environment
MagLRERTEMagLRERTEMagLRERTE
Input audio0.2251.4730.0320.2161.4080.0390.1591.4770.046
TF Estimator [56]0.3592.5960.0590.4403.2610.0920.3272.8610.147
DSP [9]0.3023.6440.0440.3003.6890.0470.4631.3000.067
VAM [6]0.2201.1980.0410.2351.1310.0510.1610.9240.070
ViGAS w/o visual0.1730.9730.0310.1811.0070.0360.1460.8770.046
ViGAS0.1590.7820.0290.1750.9710.0340.1420.7160.048
", + "image_path": "9d9fadd1497dc629a2dc0b66f7a6b5ad60c8c4a757a51ad1d010d46044987d73.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 270, + 288, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 270, + 288, + 449 + ], + "spans": [ + { + "bbox": [ + 46, + 270, + 288, + 449 + ], + "type": "text", + "content": "timating a transfer function directly from two audio clips is non-trivial and noisy for low-energy parts of the signal. DSP also performs badly despite having the ground truth 3D coordinates of the sound source. This is because head related transfer functions are typically recorded in anechoic chambers, which does not account for acoustics of different environments, e.g., reverberation. Both traditional approaches perform worse than simply copying the input audio, indicating that learning-based models are needed for this challenging task. The recent model VAM [6] performs much better compared to the traditional approaches but still underperforms our model. There is a significant difference between ViGAS w/o visual and the full model; this shows that the visual knowledge about the speaker location and the environment is important for this task." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 453, + 287, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 453, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 46, + 453, + 287, + 525 + ], + "type": "text", + "content": "Fig. 4 shows an example where given the same input source viewpoint, our model synthesizes audio for three different target viewpoints. The model reasons about how the geometry and speaker locations changes based on the source view and the target pose, and predicts the acoustic difference accordingly. See Supp. video to listen to sounds." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 528, + 288, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 528, + 288, + 672 + ], + "spans": [ + { + "bbox": [ + 46, + 528, + 288, + 672 + ], + "type": "text", + "content": "For the novel environment setup, our model again outperforms all baselines. Compared to ViGAS in the single environment setup, both the magnitude spectrogram distance and the left-right energy ratio error increase. This is expected because for novel (unseen) environments, single images capture limited geometry and acoustic information. The model fails sometime when there is a drastic viewpoint change, e.g., target viewpoint 3 in Fig. 4. This setup requires the model to reason or \"imagine\" the environment based on single audio-visual observation, which poses great challenge for NVAS as well as NVS, where typically synthesis is performed in a fully observed environment." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "Ablations. Table 2 shows ablations on the model design. To understand if the model uses visual information, we ablate the visual features " + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "V_{F}" + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": " and the active speaker feature" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 310, + 268, + 541, + 365 + ], + "blocks": [ + { + "bbox": [ + 46, + 201, + 545, + 257 + ], + "lines": [ + { + "bbox": [ + 46, + 201, + 545, + 257 + ], + "spans": [ + { + "bbox": [ + 46, + 201, + 545, + 257 + ], + "type": "text", + "content": "Table 1. Results on SoundSpaces-NVAS and Replay-NVAS. We report the magnitude spectrogram distance (Mag), left-right energy ratio error (LRE), and RT60 error (RTE). Replay-NVAS does not have novel environment setup due to data being collected in a single environment. For all metrics, lower is better. In addition to baselines, we also evaluate ViGAS w/o visual by removing the active speaker localization and visual features. Note that reverberation time is mostly invariant of the receiver location in the same room and thus input audio has low RTE. A good model should preserve this property while synthesizing the desired acoustics for the target viewpoint." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 310, + 268, + 541, + 365 + ], + "lines": [ + { + "bbox": [ + 310, + 268, + 541, + 365 + ], + "spans": [ + { + "bbox": [ + 310, + 268, + 541, + 365 + ], + "type": "table", + "html": "
ViGASSS-NVASReplay-NVAS
MagLREMagLRE
full model0.1590.7820.1420.716
w/o visual features0.1710.8970.1460.920
w/o ASL0.1610.8140.1430.757
w/o alignment0.1760.7710.1440.706
w/o separation0.1650.8400.1820.859
", + "image_path": "3e022f4da14aecd133513384353c6e597a6140b228cb7718ebaf13232534fdba.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 331, + 369, + 518, + 380 + ], + "lines": [ + { + "bbox": [ + 331, + 369, + 518, + 380 + ], + "spans": [ + { + "bbox": [ + 331, + 369, + 518, + 380 + ], + "type": "text", + "content": "Table 2. Ablations of the model on both datasets." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 385, + 545, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 385, + 545, + 456 + ], + "spans": [ + { + "bbox": [ + 305, + 385, + 545, + 456 + ], + "type": "inline_equation", + "content": "V_{L}" + }, + { + "bbox": [ + 305, + 385, + 545, + 456 + ], + "type": "text", + "content": ". Removing the active speaker feature leads to less damage on the model performance, because without the explicitly localized active speaker, the model can still implicitly reason about the active speaker location based on the image and audio. If both are removed (\"ViGAS w/o visual\" in Table 1), the performance suffers most." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 457, + 545, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 457, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 305, + 457, + 545, + 529 + ], + "type": "text", + "content": "To study the effectiveness of the temporal alignment and ambient sound separation modules, we ablate them separately. Removing the temporal alignment leads to higher Mag error and slightly lower LRE. As for ambient sound separation, the results show that optimizing for the high-energy noise-like ambient sound degrades the performance." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 539, + 444, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 539, + 444, + 552 + ], + "spans": [ + { + "bbox": [ + 306, + 539, + 444, + 552 + ], + "type": "text", + "content": "6.2. Results on Replay-NVAS" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 557, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 557, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 557, + 545, + 713 + ], + "type": "text", + "content": "Table 1 (right) shows the Replay-NVAS results. Compared to SoundSpaces-NVAS, the magnitudes of all errors are smaller because there are less drastic acoustic changes between viewpoints (8 DLSR cameras form a circle around the participants). Traditional approaches like TF Estimator and DSP still perform poorly despite using the 3D coordinates of the camera and the speaker (triangulated from multiple cameras). VAM performs better due to end-to-end learning; however, our model outperforms it. Compared to ViGAS w/o visual, the full model has much lower left-right energy ratio error and slightly higher reverberation time error, showing that the model takes into account the speaker position and viewpoint change for synthesizing the audio." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6415" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 174, + 157 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 174, + 157 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 174, + 157 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 174, + 157 + ], + "type": "image", + "image_path": "bba20e54029ad3b75f1ad6e549e9eaead118c8bcbb197e3a42d4641063558db8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 175, + 70, + 298, + 157 + ], + "blocks": [ + { + "bbox": [ + 175, + 70, + 298, + 157 + ], + "lines": [ + { + "bbox": [ + 175, + 70, + 298, + 157 + ], + "spans": [ + { + "bbox": [ + 175, + 70, + 298, + 157 + ], + "type": "image", + "image_path": "2a3bc9b8f482259d3bc533a3c7fe27659926713b905a5cd64fa54a22e9625ed5.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 299, + 70, + 421, + 157 + ], + "blocks": [ + { + "bbox": [ + 299, + 70, + 421, + 157 + ], + "lines": [ + { + "bbox": [ + 299, + 70, + 421, + 157 + ], + "spans": [ + { + "bbox": [ + 299, + 70, + 421, + 157 + ], + "type": "image", + "image_path": "dbdcf8565c1173fa0eefc44219c9345796f5258ee97785a72dac4cae43e3738b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 423, + 70, + 544, + 157 + ], + "blocks": [ + { + "bbox": [ + 423, + 70, + 544, + 157 + ], + "lines": [ + { + "bbox": [ + 423, + 70, + 544, + 157 + ], + "spans": [ + { + "bbox": [ + 423, + 70, + 544, + 157 + ], + "type": "image", + "image_path": "244a341a624ed1e79c438f2cd502663b30e347453447777ce98071dcbc56d713.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 49, + 158, + 210, + 236 + ], + "blocks": [ + { + "bbox": [ + 49, + 158, + 210, + 236 + ], + "lines": [ + { + "bbox": [ + 49, + 158, + 210, + 236 + ], + "spans": [ + { + "bbox": [ + 49, + 158, + 210, + 236 + ], + "type": "image", + "image_path": "e3e5bb3bea372403391cf8fc8e1cf880e10e35f98d61924e445ea189442d3301.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 241, + 545, + 352 + ], + "lines": [ + { + "bbox": [ + 46, + 241, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 46, + 241, + 545, + 352 + ], + "type": "text", + "content": "Figure 4. Qualitative examples. For all binaural audio, we show the left-channel and the right-channel waveforms side-by-side. Row 1: SoundSpaces-NVAS example where given the source viewpoint and input audio, the model synthesizes audio for three different target viewpoints (target views are for reference only). In this case, the active speaker is the male speaker as indicated by the bounding box. For target viewpoint 1, the view rotates about 90 degrees and the male speaker is on the left side and the predicted left channel is louder than the right channel. Viewpoint 2 moves away from the speaker and thus yields lower amplitude compared to the first prediction. For target viewpoint 3, it is completely located outside of the living room, in which case, the sound could only come from the door open on the right (louder right channel) and the reverberation also greatly increases due to the vanishing direct sound. Row 2: Replay-NVAS example where the speaker is located on the left in the source viewpoint which becomes the right and further from the camera in target viewpoint 2, the model also predicts lower amplitude and louder right channel. On the right side, we show an example of the audio-visual speech enhancement for the active speaker. The model enhances the speech to largely match with the near-range audio (target)." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 211, + 158, + 369, + 236 + ], + "blocks": [ + { + "bbox": [ + 211, + 158, + 369, + 236 + ], + "lines": [ + { + "bbox": [ + 211, + 158, + 369, + 236 + ], + "spans": [ + { + "bbox": [ + 211, + 158, + 369, + 236 + ], + "type": "image", + "image_path": "c600a1a81ad1afc4eaa470f46a004ce4c3f4ebf6124aabeb9a5b3f381b4d0796.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 372, + 158, + 543, + 236 + ], + "blocks": [ + { + "bbox": [ + 372, + 158, + 543, + 236 + ], + "lines": [ + { + "bbox": [ + 372, + 158, + 543, + 236 + ], + "spans": [ + { + "bbox": [ + 372, + 158, + 543, + 236 + ], + "type": "image", + "image_path": "fdec0cd93464eb57ac3729b5a726b7c8bff74bdfa670c61ebe913fefb85e8165.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 84, + 362, + 247, + 411 + ], + "blocks": [ + { + "bbox": [ + 84, + 362, + 247, + 411 + ], + "lines": [ + { + "bbox": [ + 84, + 362, + 247, + 411 + ], + "spans": [ + { + "bbox": [ + 84, + 362, + 247, + 411 + ], + "type": "table", + "html": "
MagRTE
Input0.2790.376
ViGAS (ours)0.2340.122
", + "image_path": "e5b95849ec52eecae0abbc709e477312eb7a0acacc4786c60e3d7f3063c78820.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 433, + 287, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 433, + 287, + 504 + ], + "spans": [ + { + "bbox": [ + 46, + 433, + 287, + 504 + ], + "type": "text", + "content": "Fig. 4 (row 2, left) shows a qualitative example. In the source viewpoint, the active speaker is on the left, while in the target viewpoint, he is further from the camera and on the right. The model synthesizes an audio waveform that captures the corresponding acoustic change, showing that our model successfully learns from real videos." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 507, + 287, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 507, + 287, + 627 + ], + "spans": [ + { + "bbox": [ + 46, + 507, + 287, + 627 + ], + "type": "text", + "content": "Audio-visual speech enhancement. In some real-world applications, e.g., hearing aid devices, the goal is to obtain the enhanced clean speech of the active speaker. This can be seen as a special case of NVAS, where the target viewpoint is the active speaker. Our model is capable of performing audio-visual speech enhancement without any modification. We simply set the target audio to the near-range audio recording for the active speaker. We show the results in Table 3. Our model obtains cleaner audio compared to the input audio (example in Fig. 4, row 2, right)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 630, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 288, + 713 + ], + "type": "text", + "content": "Human subject study. To supplement the quantitative metrics and evaluate how well our synthesized audio captures the acoustic change between viewpoints, we conduct a human subject study. We show participants the image of the target viewpoint " + }, + { + "bbox": [ + 46, + 630, + 288, + 713 + ], + "type": "inline_equation", + "content": "V_{T}" + }, + { + "bbox": [ + 46, + 630, + 288, + 713 + ], + "type": "text", + "content": " as well as the audio " + }, + { + "bbox": [ + 46, + 630, + 288, + 713 + ], + "type": "inline_equation", + "content": "A_{T}" + }, + { + "bbox": [ + 46, + 630, + 288, + 713 + ], + "type": "text", + "content": " as reference. We provide three audio samples: the input, the prediction of ViGAS, and the prediction of DSP (the most nat" + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 323, + 362, + 526, + 411 + ], + "blocks": [ + { + "bbox": [ + 75, + 417, + 258, + 428 + ], + "lines": [ + { + "bbox": [ + 75, + 417, + 258, + 428 + ], + "spans": [ + { + "bbox": [ + 75, + 417, + 258, + 428 + ], + "type": "text", + "content": "Table 3. Speech enhancement on Replay-NVAS." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 323, + 362, + 526, + 411 + ], + "lines": [ + { + "bbox": [ + 323, + 362, + 526, + 411 + ], + "spans": [ + { + "bbox": [ + 323, + 362, + 526, + 411 + ], + "type": "table", + "html": "
DatasetInputDSPViGAS
SoundSpaces-NVAS24%2%74%
Replay-NVAS43%6%51%
", + "image_path": "36dcf95921f1d5f95ad1f7b16addc02b387400ddce88a38c8c616cddc3450b68.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 417, + 545, + 450 + ], + "lines": [ + { + "bbox": [ + 304, + 417, + 545, + 450 + ], + "spans": [ + { + "bbox": [ + 304, + 417, + 545, + 450 + ], + "type": "text", + "content": "Table 4. Human Study. Participants favor our approach over the two most realistic sounding baselines, (1) copying the input signal, and (2) a digital signal processing baseline." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 456, + 545, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 456, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 304, + 456, + 545, + 503 + ], + "type": "text", + "content": "urally sounding baseline) and ask them to select a clip that sounds closest to the target audio. We select 20 examples from SoundSpaces-NVAS and 20 examples from Replay-NVAS and invite 10 participants to perform the study." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 504, + 545, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 504, + 545, + 552 + ], + "spans": [ + { + "bbox": [ + 304, + 504, + 545, + 552 + ], + "type": "text", + "content": "See Table 4 for the results. On the synthetic dataset SoundSpaces-NVAS, our approach is preferred over the baselines by a large margin. This margin is lower on the real-world Replay-NVAS dataset but is still significant." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 562, + 378, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 562, + 378, + 574 + ], + "spans": [ + { + "bbox": [ + 306, + 562, + 378, + 574 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": "We introduce the challenging novel-view acoustic synthesis task and a related benchmark in form of both real and synthetic datasets. We propose a neural rendering model that learns to transform the sound from the source viewpoint to the target viewpoint by reasoning about the observed audio and visual stream. Our model surpasses all baselines on both datasets. We believe this research unlocks many potential applications and research in multimodal novel-view synthesis. In the future, we plan to incorporate active-speaker localization model into the approach and let the model jointly learn to localize and synthesize." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6416" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "text", + "content": "[1] Triantafyllos Afouras, Joon Son Chung, and Andrew Zisserman. The conversation: Deep audio-visual speech enhancement. In INTERSPEECH, 2018. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 124, + 288, + 167 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 124, + 288, + 167 + ], + "spans": [ + { + "bbox": [ + 53, + 124, + 288, + 167 + ], + "type": "text", + "content": "[2] Humam Alwassel, Dhruv Mahajan, Bruno Korbar, Lorenzo Torresani, Bernard Ghanem, and Du Tran. Self-supervised learning by cross-modal audio-video clustering. In NeurIPS, 2020. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 168, + 288, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 168, + 288, + 213 + ], + "spans": [ + { + "bbox": [ + 53, + 168, + 288, + 213 + ], + "type": "text", + "content": "[3] M. Berouti, Richard Schwartz, and John Makhoul. Enhancement of speech corrupted by acoustic noise. In IEEE International Conference on Acoustics, Speech, and Signal Processing, 1979. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 213, + 288, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 213, + 288, + 236 + ], + "spans": [ + { + "bbox": [ + 53, + 213, + 288, + 236 + ], + "type": "text", + "content": "[4] J.S. Bradley. Review of objective room acoustics measures and future needs. Applied Acoustics, 2011. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 236, + 288, + 301 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 236, + 288, + 301 + ], + "spans": [ + { + "bbox": [ + 53, + 236, + 288, + 301 + ], + "type": "text", + "content": "[5] Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niessner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from rgb-d data in indoor environments. 3DV, 2017. MatterPort3D dataset license available at: http://kaldir.vc.in.tum.de/matterport/MP_TOS.pdf.4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 302, + 288, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 302, + 288, + 334 + ], + "spans": [ + { + "bbox": [ + 53, + 302, + 288, + 334 + ], + "type": "text", + "content": "[6] Changan Chen, Ruohan Gao, Paul Calamia, and Kristen Grauman. Visual acoustic matching. In CVPR, 2022. 2, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 335, + 288, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 335, + 288, + 390 + ], + "spans": [ + { + "bbox": [ + 53, + 335, + 288, + 390 + ], + "type": "text", + "content": "[7] Changan Chen, Carl Schissler, Sanchit Garg, Philip Kobernik, Alexander Clegg, Paul Calamia, Dhruv Batra, Philip W Robinson, and Kristen Grauman. Soundspaces 2.0: A simulation platform for visual-acoustic learning. In NeurIPS 2022 Datasets and Benchmarks Track, 2022. 2, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 391, + 288, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 391, + 288, + 423 + ], + "spans": [ + { + "bbox": [ + 53, + 391, + 288, + 423 + ], + "type": "text", + "content": "[8] Changan Chen, Wei Sun, David Harwath, and Kristen Grauman. Learning audio-visual dereverberation. In ICASSP, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 424, + 288, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 424, + 288, + 468 + ], + "spans": [ + { + "bbox": [ + 53, + 424, + 288, + 468 + ], + "type": "text", + "content": "[9] Corey I. Cheng and Gregory H. Wakefield. Introduction to head-related transfer functions (hrtfs): representations of hrtfs in time, frequency, and space. journal of the audio engineering society, 49(4):231-249, april 2001. 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 469, + 287, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 469, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 48, + 469, + 287, + 502 + ], + "type": "text", + "content": "[10] MMTracking Contributors. MMTracking: OpenMMLab video perception toolbox and benchmark. https://github.com/open-mmlab/mmtracking, 2020.5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 502, + 287, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 502, + 287, + 535 + ], + "spans": [ + { + "bbox": [ + 48, + 502, + 287, + 535 + ], + "type": "text", + "content": "[11] Alexandre Defossez, Gabriel Synnaeve, and Yossi Adi. Real time speech enhancement in the waveform domain. In *Interspeech*, 2020. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 536, + 287, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 287, + 580 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 287, + 580 + ], + "type": "text", + "content": "[12] James Eaton, Nikolay Gaubitch, Allistair Moore, and Patrick Naylor. Estimation of room acoustic parameters: The ACE challenge. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 24(10), 2016. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 581, + 287, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 581, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 581, + 287, + 613 + ], + "type": "text", + "content": "[13] Yariv Ephraim and Harry L. Van Trees. A signal subspace approach for speech enhancement. IEEE Transactions on Speech and Audio Processing, 1995. 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 614, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 287, + 668 + ], + "type": "text", + "content": "[14] Ariel Ephrat, Inbar Mosseri, Oran Lang, Tali Dekel, Kevin Wilson, Avinatan Hassidim, William T Freeman, and Michael Rubinstein. Looking to listen at the cocktail party: A speaker-independent audio-visual model for speech separation. In SIGGRAPH, 2018. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 287, + 713 + ], + "type": "text", + "content": "[15] Szu-Wei Fu, Chien-Feng Liao, Yu Tsao, and Shou-De Lin. Metricgan: Generative adversarial networks based black-box metric scores optimization for speech enhancement. In International Conference on Machine Learning (ICML), 2019. 4" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "text", + "content": "[16] Hannes Gamper and Ivan J Tashev. Blind reverberation time estimation using a convolutional neural network. In 2018 16th International Workshop on Acoustic Signal Enhancement (IWAENC), pages 136-140, 2018. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 118, + 545, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 138 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 138 + ], + "type": "text", + "content": "[17] Ruohan Gao and Kristen Grauman. 2.5d visual sound. In CVPR, 2019. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 140, + 545, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 140, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 308, + 140, + 545, + 172 + ], + "type": "text", + "content": "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 173, + 545, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 173, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 308, + 173, + 545, + 216 + ], + "type": "text", + "content": "[19] Di Hu, Rui Qian, Minyue Jiang, Xiao Tan, Shilei Wen, Errui Ding, Weiyao Lin, and Dejing Dou. Discriminative sounding objects localization via self-supervised audiovisual matching. In NeurIPS, 2020. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 217, + 545, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 217, + 545, + 249 + ], + "spans": [ + { + "bbox": [ + 308, + 217, + 545, + 249 + ], + "type": "text", + "content": "[20] Ajay Jain, Matthew Tancik, and Pieter Abbeel. Putting nerf on a diet: Semantically consistent few-shot view synthesis. In Proc. ICCV, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 250, + 545, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 250, + 545, + 282 + ], + "spans": [ + { + "bbox": [ + 308, + 250, + 545, + 282 + ], + "type": "text", + "content": "[21] Hao Jiang, Calvin Murdock, and Vamsi Krishna Ithapu. Ego-centric deep multi-channel audio-visual active speaker localization. In CVPR, 2022. 3, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 283, + 545, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 283, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 308, + 283, + 545, + 348 + ], + "type": "text", + "content": "[22] Florian Klein, Annika Neidhardt, and Marius Seipel. Real-time estimation of reverberation time for selection of suitable binaural room impulse responses. In Audio for Virtual, Augmented and Mixed Realities: Proceedings of 5th International Conference on Spatial Audio (ICSA), pages 145–150, 2019. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 349, + 545, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 349, + 545, + 382 + ], + "spans": [ + { + "bbox": [ + 308, + 349, + 545, + 382 + ], + "type": "text", + "content": "[23] C. Knapp and G. Carter. The generalized correlation method for estimation of time delay. IEEE Transactions on Acoustics, Speech, and Signal Processing, 1976. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 383, + 545, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 383, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 308, + 383, + 545, + 415 + ], + "type": "text", + "content": "[24] Bruno Korbar, Du Tran, and Lorenzo Torresani. Cooperative learning of audio and video models from self-supervised synchronization. In NeurIPS, 2018. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 415, + 545, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 415, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 308, + 415, + 545, + 449 + ], + "type": "text", + "content": "[25] Jonas Kulhanek, Erik Derner, Torsten Sattler, and Robert Babuška. ViewFormer: NeRF-free neural rendering from few images using transformers. In Proc. ECCV, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 449, + 545, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 449, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 308, + 449, + 545, + 469 + ], + "type": "text", + "content": "[26] Heinrich Kuttruff. Room Acoustics. Boca Raton, 6th edition, 2016. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 471, + 545, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 471, + 545, + 504 + ], + "spans": [ + { + "bbox": [ + 308, + 471, + 545, + 504 + ], + "type": "text", + "content": "[27] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. arXiv.cs, abs/2011.13084, 2020. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 504, + 545, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 504, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 308, + 504, + 545, + 548 + ], + "type": "text", + "content": "[28] Lingjie Liu, Marc Habermann, Viktor Rudnev, Kripasindhu Sarkar, Jiatao Gu, and Christian Theobalt. Neural humans: Pose-controlled free-view synthesis of human actors with template-guided neural radiance fields. In arXiv, 2021. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 548, + 545, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 548, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 308, + 548, + 545, + 581 + ], + "type": "text", + "content": "[29] Andrew Luo, Yilun Du, Michael J Tarr, Joshua B Tenenbaum, Antonio Torralba, and Chuang Gan. Learning neural acoustic fields. In NeurIPS, 2022. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 582, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 582, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 582, + 545, + 624 + ], + "type": "text", + "content": "[30] Wolfgang Mack, Shuwen Deng, and Emanuele AP Habets. Single-channel blind direct-to-reverberation ratio estimation using masking. In INTERSPEECH, pages 5066-5070, 2020. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 625, + 545, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 625, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 308, + 625, + 545, + 669 + ], + "type": "text", + "content": "[31] Sagnik Majumder, Changan Chen, Ziad Al-Halah, and Kristen Grauman. Few-shot audio-visual learning of environment acoustics. In Thirty-Sixth Conference on Neural Information Processing Systems, 2022. 3" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "text", + "content": "[32] Ricardo Martin-Brualla, Noha Radwan, Mehdi S. M. Sajjadi, Jonathan T. Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In Proceedings of the IEEE/CVF" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6417" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "type": "text", + "content": "Conference on Computer Vision and Pattern Recognition (CVPR), pages 7210-7219, June 2021. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 95, + 287, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 95, + 287, + 139 + ], + "spans": [ + { + "bbox": [ + 48, + 95, + 287, + 139 + ], + "type": "text", + "content": "[33] Daniel Michelsanti, Zheng-Hua Tan, Shi-Xiong Zhang, Yong Xu, Meng Yu, Dong Yu, and Jesper Jensen. An overview of deep-learning-based audio-visual speech enhancement and separation. In arXiv, 2020. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 140, + 287, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 287, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 287, + 183 + ], + "type": "text", + "content": "[34] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 184, + 287, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 184, + 287, + 216 + ], + "spans": [ + { + "bbox": [ + 48, + 184, + 287, + 216 + ], + "type": "text", + "content": "[35] Pedro Morgado, Yi Li, and Nuno Vasconcelos. Learning representations from audio-visual spatial alignment. In NeurIPS, 2020. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 217, + 287, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 217, + 287, + 249 + ], + "spans": [ + { + "bbox": [ + 48, + 217, + 287, + 249 + ], + "type": "text", + "content": "[36] Pedro Morgado, Nono Vasconcelos, Timothy Langlois, and Oliver Wang. Self-supervised generation of spatial audio for 360 video. In NeurIPS, 2018. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 250, + 287, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 250, + 287, + 294 + ], + "spans": [ + { + "bbox": [ + 48, + 250, + 287, + 294 + ], + "type": "text", + "content": "[37] Prateek Murgai, Mark Rau, and Jean-Marc Jot. Blind estimation of the reverberation fingerprint of unknown acoustic environments. In Audio Engineering Society Convention 143. Audio Engineering Society, 2017. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 294, + 287, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 294, + 287, + 338 + ], + "spans": [ + { + "bbox": [ + 48, + 294, + 287, + 338 + ], + "type": "text", + "content": "[38] Michael Niemeyer, Jonathan T. Barron, Ben Mildenhall, Mehdi S. M. Sajjadi, Andreas Geiger, and Noha Radwan. RegNeRF: Regularizing neural radiance fields for view synthesis from sparse inputs. In Proc. CVPR, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 338, + 287, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 338, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 48, + 338, + 287, + 392 + ], + "type": "text", + "content": "[39] Aaron van den Oord, Sander Dieleman, Heiga Zen, Karen Simonyan, Oriol Vinyls, Alex Graves, Nal Kalchbrenner, Andrew Senior, and Koray Kavukcuoglu. Wavenet: A generative model for raw audio, 2016. cite arxiv:1609.03499. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 393, + 287, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 393, + 287, + 425 + ], + "spans": [ + { + "bbox": [ + 48, + 393, + 287, + 425 + ], + "type": "text", + "content": "[40] Andrew Owens and Alexei A Efros. Audio-visual scene analysis with self-supervised multisensory features. In ECCV, 2018. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 426, + 287, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 426, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 48, + 426, + 287, + 480 + ], + "type": "text", + "content": "[41] Vassil Panayotov, Guoguo Chen, Daniel Povey, and Sanjeev Khudanpur. Librispeech: an asr corpus based on public domain audio books. In 2015 IEEE international conference on acoustics, speech and signal processing (ICASSP), pages 5206-5210. IEEE, 2015. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 482, + 287, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 482, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 48, + 482, + 287, + 525 + ], + "type": "text", + "content": "[42] Keunhong Park, Utkarsh Sinha, Jonathan T. Barron, Sofien Bouaziz, Dan B. Goldman, Steven M. Seitz, and Ricardo Martin-Brualla. Deformable neural radiance fields. CoRR, abs/2011.12948, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 526, + 287, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 526, + 287, + 558 + ], + "spans": [ + { + "bbox": [ + 48, + 526, + 287, + 558 + ], + "type": "text", + "content": "[43] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-NeRF: Neural radiance fields for dynamic scenes. arXiv.cs, abs/2011.13961, 2020. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 559, + 287, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 287, + 613 + ], + "type": "text", + "content": "[44] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10318–10327, June 2021. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 614, + 287, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 287, + 657 + ], + "type": "text", + "content": "[45] Anton Ratnarajah, Zhenyu Tang, Rohith Chandrashekar Aralikatti, and Dinesh Manocha. Mesh2ir: Neural acoustic impulse response generator for complex 3d scenes. In ACM Multimedia, 2022. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[46] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common Objects in 3D: Large-scale learning and evaluation of real-life 3D category reconstruction. In Proc. CVPR, 2021. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 118 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 118 + ], + "type": "text", + "content": "[47] Alexander Richard, Peter Dodds, and Vamsi Krishna Ithapu. Deep impulse responses: Estimating and parameterizing filters with deep networks. In IEEE International Conference on Acoustics, Speech and Signal Processing, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 118, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 173 + ], + "type": "text", + "content": "[48] Alexander Richard, Dejan Markovic, Israel D Gebru, Steven Krenn, Gladstone Butler, Fernando de la Torre, and Yaser Sheikh. Neural synthesis of binaural speech from mono audio. In International Conference on Learning Representations, 2021. 2, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 175, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 239 + ], + "type": "text", + "content": "[49] Joseph Roth, Sourish Chaudhuri, Ondrej Klejch, Radhika Marvin, Andrew Gallagher, Liat Kaver, Sharadh Ramaswamy, Arkadiusz Stopczynski, Cordelia Schmid, Zhonghua Xi, and Caroline Pantofaru. Ava-activespeaker: An audio-visual dataset for active speaker detection. In ICASSP, 2020. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 241, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 241, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 545, + 274 + ], + "type": "text", + "content": "[50] Nikhil Singh, Jeff Mentch, Jerry Ng, Matthew Beveridge, and Iddo Drori. Image2reverb: Cross-modal reverb impulse response synthesis. In ICCV, 2021. 2, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 276, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 276, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 276, + 545, + 319 + ], + "type": "text", + "content": "[51] Vincent Sitzmann, Michael Zollhöfer, and Gordon Wetzstein. Scene representation networks: Continuous 3D-structure-aware neural scene representations. Proc. NeurIPS, 2019. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 320, + 545, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 320, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 307, + 320, + 545, + 374 + ], + "type": "text", + "content": "[52] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, et al. The Replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019.4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 376, + 545, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 376, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 307, + 376, + 545, + 409 + ], + "type": "text", + "content": "[53] Shih-Yang Su, Frank Yu, Michael Zollhöfer, and Helge Rhodin. A-NeRF: Surface-free human 3d pose refinement via neural rendering. arXiv.cs, abs/2102.06199, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 411, + 545, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 411, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 307, + 411, + 545, + 475 + ], + "type": "text", + "content": "[54] Ruijie Tao, Zexu Pan, Rohan Kumar Das, Xinyuan Qian, Mike Zheng Shou, and Haizhou Li. Is someone speaking? exploring long-term temporal features for audio-visual active speaker detection. In Proceedings of the 29th ACM International Conference on Multimedia, page 3927–3935, 2021. 3, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 477, + 545, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 477, + 545, + 532 + ], + "spans": [ + { + "bbox": [ + 307, + 477, + 545, + 532 + ], + "type": "text", + "content": "[55] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Christoph Lassner, and Christian Theobalt. Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a deforming scene from monocular video. arXiv.cs, abs/2012.12247, 2020. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 534, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 534, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 307, + 534, + 545, + 567 + ], + "type": "text", + "content": "[56] Norbert Wiener. Extrapolation, interpolation, and smoothing of stationary time series. Report of the Services 19, Research Project DIC-6037 MIT, 1942. 6, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 568, + 545, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 568, + 545, + 621 + ], + "spans": [ + { + "bbox": [ + 307, + 568, + 545, + 621 + ], + "type": "text", + "content": "[57] Fei Xia, Amir R. Zamir, Zhi-Yang He, Alexander Sax, Jitendra Malik, and Silvio Savarese. Gibson Env: real-world perception for embodied agents. In Computer Vision and Pattern Recognition (CVPR), 2018 IEEE Conference on. IEEE, 2018. 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 624, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 678 + ], + "type": "text", + "content": "[58] Feifei Xiong, Stefan Goetze, Birger Kollmeier, and Bernd T Meyer. Joint estimation of reverberation time and early-to-late reverberation ratio from single-channel speech signals. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 27(2):255-267, 2018. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "type": "text", + "content": "[59] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. PixelNeRF: Neural radiance fields from one or few images. In Proc. CVPR, 2021. 2" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6418" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "type": "text", + "content": "[60] Hang Zhou, Ziwei Liu, Xudong Xu, Ping Luo, and Xiaogang Wang. Vision-infused deep audio inpainting. In ICCV, 2019. 3" + } + ] + } + ], + "index": 0 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 316, + 757 + ], + "type": "text", + "content": "6419" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/OCELOT_ Overlapped Cell on Tissue Dataset for Histopathology/052d2272-757d-4e0e-b0c0-b44903da46a8_content_list.json b/2023/OCELOT_ Overlapped Cell on Tissue Dataset for Histopathology/052d2272-757d-4e0e-b0c0-b44903da46a8_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..bef64d44074be6d076ec99666032f5de480e5af0 --- /dev/null +++ b/2023/OCELOT_ Overlapped Cell on Tissue Dataset for Histopathology/052d2272-757d-4e0e-b0c0-b44903da46a8_content_list.json @@ -0,0 +1,1492 @@ +[ + { + "type": "text", + "text": "OCELOT: Overlapped Cell on Tissue Dataset for Histopathology", + "text_level": 1, + "bbox": [ + 156, + 130, + 815, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jeongun Ryu* Aaron Valero Puche* JaeWoong Shin* Seonwook Park Biagio Brattoli Jinhee Lee Wonkyung Jung Soo Ick Cho Kyunghyun Paeng Chan-Young Ock Donggeun Yoo Sérgio Pereira Lunit Inc.", + "bbox": [ + 86, + 180, + 880, + 236 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{rjw0205, aaron.valero, jwoong.shin, spark, biagio, jinhee.lee, wkjung, sooickcho, khpaeng, ock.chanyoung, dgyoo, sergio}@lunit.io", + "bbox": [ + 184, + 242, + 772, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 297, + 312, + 314 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Cell detection is a fundamental task in computational pathology that can be used for extracting high-level medical information from whole-slide images. For accurate cell detection, pathologists often zoom out to understand the tissue-level structures and zoom in to classify cells based on their morphology and the surrounding context. However, there is a lack of efforts to reflect such behaviors by pathologists in the cell detection models, mainly due to the lack of datasets containing both cell and tissue annotations with overlapping regions. To overcome this limitation, we propose and publicly release OCELOT, a dataset purposely dedicated to the study of cell-tissue relationships for cell detection in histopathology. OCELOT provides overlapping cell and tissue annotations on images acquired from multiple organs. Within this setting, we also propose multi-task learning approaches that benefit from learning both cell and tissue tasks simultaneously. When compared against a model trained only for the cell detection task, our proposed approaches improve cell detection performance on 3 datasets: proposed OCELOT, public TIGER, and internal CARP datasets. On the OCELOT test set in particular, we show up to 6.79 improvement in F1-score. We believe the contributions of this paper, including the release of the OCELOT dataset at https://lunit-io.github.io/research/publications/ocelot are a crucial starting point toward the important research direction of incorporating cell-tissue relationships in computation pathology.", + "bbox": [ + 76, + 330, + 473, + 738 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 748, + 209, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Computational Pathology (CPATH) [3] is a branch of digital pathology that develops methodologies for the analysis of digitized patient specimens, such as Whole-Slide-Images (WSIs). Cell detection in histology images [26, 37, 49] is one of the most important tasks in CPATH. It allows the quantification and analysis of different cell types, which can lead to better prognosis evaluation [34, 40] and patient", + "bbox": [ + 75, + 772, + 470, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "treatment planning while maintaining medical interpretability [13]. Since it has the potential to impact human lives, high-performance cell detection models are essential in real-world applications and need to be investigated.", + "bbox": [ + 500, + 300, + 890, + 359 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To better locate and classify cells, detailed morphological characteristics such as color and shape are crucial. Consequently, cell detection datasets are typically collected at high magnification but small Field-of-View (FoV). However, this can make the cell detection model overly rely on appearance details, without understanding the broader context [42]. This context can help cell detection by providing information about how cells are arranged and grouped together to form high-level tissue structures. In practice, expert annotators (pathologists) first zoom out to understand these broad tissue structures. Next, they zoom in to better classify individual cells while taking into account the context information, as depicted in Fig. 1.", + "bbox": [ + 496, + 364, + 892, + 560 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The behavior of pathologists can be transferred to deep learning, for instance, through a multi-task strategy combining cell detection tasks at high magnification and tissue segmentation at low magnification. This type of approach would allow the model to share knowledge across different tasks and FoVs. However, to train such an approach, a combined dataset with cell-tissue overlapping regions is required; unfortunately, most existing datasets only target a single task, either cell detection [17,25] or tissue segmentation [7, 11].", + "bbox": [ + 496, + 564, + 893, + 715 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we introduce a new research direction: studying cell-tissue relationships for cell detection. First, we publish the OCELOT dataset, which contains cell and tissue annotations in small and large FoV patches, respectively, with overlapping regions. Additionally, the data is collected from WSIs of multiple organs. This can provide the necessary data for researchers to study cell-tissue relationships and their effect on cell detection. Second, we introduce simple multi-task learning approaches for cell detection that can benefit from cell-tissue relationships and demonstrate their advantages over 3 different datasets. These approaches consistently show better cell detection performance compared to", + "bbox": [ + 496, + 719, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*: Equal contribution", + "bbox": [ + 94, + 887, + 210, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "23902", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1a09049a2141ea141e91630bf8e9c6d20f719777a882510d8717a46ee7200df8.jpg", + "image_caption": [ + "Figure 1. Behavior of pathologists and cell detection models. Typical cell detection models infer their predictions only by looking at a limited FoV region. Failure cases can occur when cells are difficult to be classified solely by their morphology, i.e., while disregarding the larger architecture of the tissues (context). In the above example, some groups of tumor cells are misclassified as background cells due to their morphology; these tumor cells are smaller and rounder than the nearby ones. Note that the large size and irregular shape are representative characteristics of tumor cells. Pathologists overcome these limitations by further understanding the tissue context of the larger FoV region. Misclassified cells can be easily corrected by understanding that such regions are cancer areas." + ], + "image_footnote": [], + "bbox": [ + 163, + 80, + 803, + 311 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the cell-only baseline, i.e., a model trained only with small FoV patches with the corresponding cell annotations. We hope that our proposed OCELOT dataset and methods will encourage the CPATH community to learn how to reflect cell-tissue relationships better to improve cell detection.", + "bbox": [ + 75, + 422, + 468, + 497 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are 3-fold and summarized as follows,", + "bbox": [ + 96, + 500, + 467, + 512 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- The first work that exploits cell-tissue relationships for better cell detection, to the best of our knowledge.", + "- We release OCELOT, a dataset with overlapping cell and tissue annotations based on Hematoxylin and Eosin (H&E) stained WSIs of multiple organs.", + "- We introduce several approaches that boost cell detection performance via multi-task learning, and empirically confirm that these methods generalize well across different datasets and histological stainings." + ], + "bbox": [ + 93, + 523, + 468, + 672 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 686, + 218, + 700 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Datasets for Cell detection and Tissue segmentation Tasks", + "text_level": 1, + "bbox": [ + 76, + 710, + 467, + 739 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In recent years, numerous datasets have been released for tackling cell detection. Some of those works only target a single organ [17,32], while others consider multiple ones [4, 15,25,44,45]. The availability of these datasets enables the CPATH community to push forward the development and improvement of cell detection models [17,26,29,30,37,49]. In addition, tissue segmentation datasets have also been proposed for prostate [7,23,33], colorectal [50], brain [1], and multiple organs [20]. Some examples of tissue segmentation works can be found in [8,9,28,36,47,52]. The dataset", + "bbox": [ + 75, + 750, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/7bd920886542aaefcbb9dc8a92f50fc597a648bc31761a2ba2e7629b06a84432.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetTissue Area# CellOrgans
OCELOT4.267cm2114.7Kmultiple
TIGER2.536cm250.8Kbreast
", + "bbox": [ + 555, + 419, + 836, + 477 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Table 1. Dataset comparison in terms of physical tissue annotated area and total cell counting per dataset.", + "bbox": [ + 500, + 484, + 890, + 513 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "in [11] is composed of a cell detection subset and a tissue segmentation subset. However, the subsets are annotated independently and from different patient groups, and, therefore, there are no overlaps between the cell and tissue data. Because of the lack of overlapping data in the aforementioned datasets, it is difficult to build an end-to-end framework to learn cell-tissue relationships by jointly training on the cell and tissue tasks.", + "bbox": [ + 496, + 532, + 890, + 654 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The recently released TIGER dataset [2] contains both cell and tissue annotations to study tumor-infiltrating lymphocytes [38] in H&E breast cancer WSIs. All the cell-annotated areas exist inside the tissue-annotated area, however, this work does not propose nor initiate any effort toward the integration of both cell and tissue tasks.", + "bbox": [ + 496, + 656, + 890, + 746 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To further promote the development of methods that leverage the cell-tissue relationship for the task of cell detection, we propose OCELOT, which is designed to capture the hierarchical relationship between cells and tissues, especially in the tumor environment. OCELOT contains roughly two times more cell and tissue annotations than TIGER (see Tab. 1). Additionally, the data was collected from multiple organs to enable the investigation of the generalizability of cell-tissue relationships over various cancer types. In the end, we utilized both OCELOT and TIGER to reveal the", + "bbox": [ + 496, + 750, + 892, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "23903", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "power of cell-tissue relationship for cell detection in Sec. 6.", + "bbox": [ + 76, + 90, + 467, + 106 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Leveraging Large Field of View", + "text_level": 1, + "bbox": [ + 76, + 114, + 354, + 132 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Some studies [6, 19, 22, 39, 42, 43] extract a large FoV region as an additional input to improve detection/segmentation performance on smaller FoV regions. [22] proposes a dual pathway 3D CNN for brain lesion segmentation, where each pathway receives both small and large center-shared FoV regions as input. Similar studies are also conducted in the CPATH domain for tissue segmentation [19, 39, 42, 43] and cell detection [6]. To fuse different FoV patches, [42] introduces a weighting mechanism and [43] proposes a multi-scale merging block composed of convolution and concatenation. Nevertheless, no inter-task relationship is considered in the previously mentioned methods. In contrast, our models take advantage of the large contextual information while learning the cell-tissue relationship via multi-task objectives at different FoVs.", + "bbox": [ + 75, + 138, + 468, + 364 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Leveraging Cell-Tissue Relationships", + "text_level": 1, + "bbox": [ + 76, + 375, + 398, + 391 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To our understanding, there is no study that considers cell-tissue relationships for cell detection or tissue segmentation tasks. On the other hand, a few efforts have attempted to link tissue and cell for image classification using graph-based methods [35,46,51]. Such studies represent the tissue structure as a graph of detected cells, based on the medical knowledge that cells form tissues. [35] explicitly considers the cell-tissue relationship for the task of breast cancer subtyping, by an interaction between tissue-level and cell-level graphs with a cell-to-tissue hierarchy. However, these methods treat the cell/tissue graph generation as a pre-processing step, by using the inference output of independently pretrained cell detection and tissue segmentation models. In contrast, we directly target the improvement of cell detection by utilizing the cell-tissue relationship.", + "bbox": [ + 75, + 398, + 468, + 625 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. OCELOT", + "text_level": 1, + "bbox": [ + 76, + 638, + 187, + 654 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we introduce OCELOT, a histopathology dataset specifically built to enable the development of methods that leverage cell and tissue relationships. Each sample of the OCELOT dataset $D$ is composed of six components,", + "bbox": [ + 75, + 665, + 468, + 726 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {D} = \\left\\{\\left(x _ {s}, y _ {s} ^ {c}, x _ {l}, y _ {l} ^ {t}, c _ {x}, c _ {y}\\right) _ {i} \\right\\} _ {i = 1} ^ {N} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 165, + 737, + 468, + 760 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $x_{s}, x_{l}$ are the small and large FoV patches extracted from the WSI, $y_{s}^{c}, y_{l}^{t}$ refer to the corresponding cell and tissue annotations, respectively, and $c_{x}, c_{y}$ are the relative coordinates of the center of $x_{s}$ within $x_{l}$ . We drop the sample index $i$ for simplicity. Fig. 2 shows the visualization of a sample in OCELOT. More details about the dataset including data collection and statistics can be found in the following sub-sections. The dataset is publicly available at https://lunit-io.github.io/research/publications/ocelot/.", + "bbox": [ + 75, + 763, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/044543b55eab3dcae11256d4a6b2215ccf6b5d0ff53f96f269c8f0129b460040.jpg", + "image_caption": [ + "$(x_{l},y_{l}^{t})$", + "Figure 2. A sample from the OCELOT dataset. Each sample of the dataset consists of two input patches and the corresponding annotations. Left shows the large FoV patch $x_{l}$ with tissue segmentation annotation $y_{l}^{t}$ , where green denotes the cancer area. Right shows the small FoV patch $x_{s}$ with cell point annotation $y_{s}^{c}$ , where blue and yellow dots denote tumor and background cells, respectively. The red box indicates the size and location of the $x_{s}$ with respect to the $x_{l}$ ." + ], + "image_footnote": [], + "bbox": [ + 506, + 90, + 694, + 236 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a3eec8298136c7b2992ab60c079c30314fdb1beb34c7ee0d51adf921b4119281.jpg", + "image_caption": [ + "$(x_{s},y_{s}^{c})$" + ], + "image_footnote": [], + "bbox": [ + 697, + 90, + 887, + 236 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Data Collection", + "text_level": 1, + "bbox": [ + 500, + 390, + 655, + 404 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We collect 306 TCGA [21] WSIs from a total of 6 different organs: kidney, head-and-neck, prostate, stomach, endometrium, and bladder. From each of the WSIs, we select 1 to 3 large Regions of Interest (ROIs) for the tissue segmentation task. Finally, for the cell detection task, we randomly choose a smaller ROI that is fully contained within the larger tissue ROI. As a result, OCELOT includes 673 paired patches from 6 organs. The numbers of WSIs and pairs of patches per organ are detailed in Tab. 2.", + "bbox": [ + 496, + 415, + 890, + 551 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Some natural image datasets, such as ImageNet [12] or Pascal VOC [14], include thousands of annotated images. However, annotating histopathology images is more challenging and expensive due to the scarcity of expert pathologists [27]. Furthermore, acquiring dense annotations for cell detection and tissue segmentation is especially time-demanding compared to higher-level tasks such as image classification. Nonetheless, in Tab. 1, we observe that OCELOT is roughly double the size of the recent TIGER dataset with respect to the annotated tissue area and the number of annotated cells.", + "bbox": [ + 496, + 554, + 890, + 719 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Patch configuration. Cell detection tasks benefit from fine-grained spatial information to better capture detailed cell properties (e.g. border, shape, color, and opacity). In contrast, tissue segmentation requires a larger context to enable a better understanding of the overall structural information. Therefore, we define the FoV sizes of $x_{s}$ (cell detection) and $x_{l}$ (tissue segmentation) as $1024 \\times 1024$ and $4096 \\times 4096$ pixels, respectively, at a resolution of 0.2 Micronsper-Pixel (MPP). Finally, the large FoV patches and tissue annotations $(x_{l}, y_{l}^{t})$ are down-sampled by a factor of 4, resulting in a size of $1024 \\times 1024$ pixels.", + "bbox": [ + 496, + 734, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "23904", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/ab63a2ffa026aff9c33966182be63ae809e31045b6930c57f34565e5c8513985.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Organs# Slides# Patch Pairs
TrainValTestTrainValTest
Kidney4815181254141
Head-neck135627910
Prostate261210501716
Stomach1565361212
Endometrium381313862925
Bladder351414822926
Total1756566406137130
", + "bbox": [ + 125, + 88, + 419, + 241 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 2. Dataset size per organ and data subset.", + "bbox": [ + 122, + 244, + 421, + 258 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Annotation. All cell-tissue pairs of patches are annotated by board-certified pathologists. Cells are labeled as points, with associated 2D coordinates and class labels. We denote the annotations in a given cell-level patch, $x_{s}$ , as $y_{s}^{c}$ , and consider two classes: Tumor Cell (TC) and Background Cell $(BC)^{1}$ . TC and BC class ratios are $35.01\\%$ and $64.99\\%$ , respectively. Regarding the tissue patches, $x_{l}$ , pathologists annotate the pixel-wise segmentation maps $y_{l}^{t}$ with either Cancer Area (CA) or Background (BG) labels. A minority of pixels where the tissue class was uncertain were labeled as Unknown (UNK). BG, CA, and UNK class ratios are $55.77\\%$ , $40.17\\%$ , and $4.06\\%$ , respectively. The amount of annotated cells and tissue pixels, per data split, can be found in the supplementary material. The detection of TC and BC has clinical relevance. For example, tumor purity [5], computed as the tumor/non-tumor cell ratio in a WSI, has a correlation with cancer prognosis [16, 31, 48].", + "bbox": [ + 75, + 273, + 472, + 530 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Dataset splits. The dataset is divided into three subsets: training, validation, and test, following a 6:2:2 ratio. To prevent information leaking among the data subsets, we randomly split the dataset per WSI, so that different patches from the same WSI are not included in multiple subsets. We maintain consistent cancer-type ratios in each subset.", + "bbox": [ + 75, + 536, + 468, + 626 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Empirical Analysis", + "text_level": 1, + "bbox": [ + 76, + 640, + 261, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The motivation for considering the cell-tissue relationships for the development of cell detection models stem from the biological and hierarchical arrangement of cells and tissues. These insights are further corroborated by two main empirical observations described in the following Sec. 4.1 and Sec. 4.2.", + "bbox": [ + 75, + 666, + 468, + 755 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Interrelation between cell and tissue classes", + "text_level": 1, + "bbox": [ + 76, + 767, + 446, + 782 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We empirically observe the interrelation between specific cell and tissue classes by counting the amount of each annotated cell type within each tissue region, as observed in Tab. 3. Indeed, we verify that in OCELOT, around $93\\%$ of $TC$ are located within $CA$ and $85\\%$ of $BC$ are found outside", + "bbox": [ + 75, + 790, + 468, + 866 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/27dace7f68fb2fca893cbdbb3f21d718329e026e7910dd3b5dde9bfa519f75cb.jpg", + "image_caption": [ + "(a) OCELOT" + ], + "image_footnote": [], + "bbox": [ + 540, + 89, + 679, + 167 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4b308743775d930a5836697e1e3b36a52a7f59ff4d754c4f42ad13db63504812.jpg", + "image_caption": [ + "(b) TIGER" + ], + "image_footnote": [], + "bbox": [ + 710, + 90, + 849, + 167 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/51b4047f3b0b71fc24b1ddc92ae16e247a35ce760e3b4de01cae095d88e5b070.jpg", + "image_caption": [ + "Table 3. Cell counts based on the tissue class. Each value stands for the number of cells located inside the tissue area. $TC$ , $BC$ , $LC$ , $CA$ , and $ST$ stand for Tumor Cell, Background Cell, Lymphocyte Cell, Cancer Area tissue, and Stroma tissue, respectively.", + "(a) OCELOT", + "Table 4. Pixel ratio among tissue classes. $CA$ , and $ST$ stand for Cancer Area tissue, and Stroma tissue, respectively." + ], + "image_footnote": [], + "bbox": [ + 545, + 267, + 676, + 315 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b7943a72b1d8be75727bb468f4fb0b766f3779c56b5f3e7b0523926a3a2965ef.jpg", + "image_caption": [ + "(b) TIGER" + ], + "image_footnote": [], + "bbox": [ + 712, + 268, + 843, + 314 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "of the CA tissue. Note that CA is not the majority tissue class (Tab. 4), therefore, we conclude that there is, in fact, a relationship between the cell and tissue classes. We observe a similar phenomenon in the TIGER dataset when considering the $LC$ and $ST$ classes (Tab. 3 and Tab. 4).", + "bbox": [ + 496, + 383, + 890, + 459 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In practice, pathologists classify cells by taking into account such interrelationships, since it is difficult to classify isolated cells, without considering the larger context of the tissues. As depicted in Fig. 1, pathologists first need to visualize the structure of the tissues at a larger FoV. Then, they zoom in and consider the previously observed context along with the fine-grained details of each individual cell and nearby neighborhood, thus considering the cell-tissue dependencies. Inspired by the behavior of pathologists, we expect a cell detection model to also benefit from understanding the tissue structure from a broader viewpoint.", + "bbox": [ + 496, + 460, + 892, + 626 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Tissue-label Leaking Model", + "text_level": 1, + "bbox": [ + 500, + 635, + 750, + 652 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the previous section, we observe a strong relationship between certain cell and tissue classes. To further validate the hypothesis that a cell detection model can leverage information from the tissue structure, we design an exploratory experiment where the tissue annotation is provided as an extra input to a cell detection model. To this end, we first crop the corresponding cell patch region from the tissue annotation $y_{l}^{t}$ , and upsample it to match the size of the cell patch, $x_{s}$ ; the cropped tissue annotation is denoted as $y_{s}^{t}$ . Finally, we concatenate $x_{s}$ and $y_{s}^{t}$ at the channels dimension and use this data to train a cell detection model. We denote this model as a Tissue-label leaking model and illustrate it in Fig. 3. Note that this model is not appropriate for real-world scenarios as the tissue labels are unknown at inference time, and is explored for the purpose of empirical analysis.", + "bbox": [ + 496, + 659, + 890, + 883 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "When we compare the performance of the tissue-label", + "bbox": [ + 519, + 885, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "${}^{1}{BC}$ includes any of the following cell categories: lymphocyte, macrophage, fibroblast, endothelial, or other remaining cell types.", + "bbox": [ + 76, + 875, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "23905", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/750b945d175481a119d51943a76a362517ec62889727e91356ab12eb8440775e.jpg", + "image_caption": [ + "Figure 3. Tissue label leaking model. This model receives the cell patch $x_{s}$ along with the corresponding tissue labels as input. The region corresponding to the cell patch in the tissue patch annotation is cropped, upsampled, and concatenated to the cell patch. $\\oplus$ denotes channel-wise concatenation." + ], + "image_footnote": [], + "bbox": [ + 101, + 89, + 439, + 311 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "leaking model with the standard cell detection model on the OCELOT dataset, we observe a significant improvement in terms of mean F1-score $^2$ performance of $+7.69$ and $+9.76$ in the validation and test sets, respectively. Detailed results can be found in Tab. 5. Taking these results into consideration, we conclude that there is significant room to improve the cell detection model, which can be achieved by combining the tasks of cell detection and tissue segmentation.", + "bbox": [ + 75, + 407, + 468, + 530 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Method", + "text_level": 1, + "bbox": [ + 76, + 544, + 168, + 558 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we propose to utilize cell-tissue relationships through multi-task learning. First, we propose a set of approaches inspired by the tissue-label leaking model described in Sec. 4.2. These models replace the annotated tissue labels with predictions from an auxiliary tissue segmentation branch. Second, we design a bi-directional information-sharing approach that shares features in both tissue-to-cell and cell-to-tissue directions. The proposed approaches are described in Sec. 5.2 and Sec. 5.3", + "bbox": [ + 75, + 569, + 468, + 705 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Preliminary", + "text_level": 1, + "bbox": [ + 76, + 715, + 207, + 729 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To deal with cell-tissue sample pairs, i.e., $(x_{s},y_{s}^{c})$ and $(x_{l},y_{l}^{t})$ , we build a dual-branch architecture containing separate networks for the cell and tissue tasks. Similarly to [41], we define cell detection as a segmentation task. Specifically, the cell labels are provided as a segmentation map by drawing a fixed-radius circle centered on each cell point an", + "bbox": [ + 75, + 738, + 468, + 829 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/69d0e324baae725b04e3cb68e530637fb67360348e2ae6a44b463074665dca7c.jpg", + "image_caption": [ + "Figure 4. Tissue-prediction injection model injects the tissue segmentation prediction into 1 out of 4 locations of the cell branch: (a) input, (b) after encoder, (c) after ASPP, and (d) after decoder." + ], + "image_footnote": [], + "bbox": [ + 504, + 90, + 890, + 273 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "notation and filled with the corresponding class label. At inference time, we find local peaks within the predicted cell probability maps and output them as point predictions. More details are provided in the supplementary material. Treating cell detection as a segmentation task enables us to use the same architecture for both cell and tissue branches, which largely simplifies the training and tuning of the model and reduces the range of possible decisions, such as neural network architecture, or hyper-parameters. We use DeepLabV3+ [10] as a base architecture for both branches and single-task models.", + "bbox": [ + 496, + 345, + 892, + 512 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2. Tissue-prediction Injection Models", + "text_level": 1, + "bbox": [ + 500, + 523, + 803, + 540 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "These models are a simple and practical extension of the tissue-label leaking model, where we inject the predicted tissue probabilities into the cell detection branch instead of leaking the tissue labels. We consider only one injection point in the cell detection branch, but explore four possible alternatives: (a) at the input (Pred-to-input), (b) after the encoder (Pred-to-inter-1), (c) after the ASPP module (Pred-to-inter-2), and (d) after the decoder (Pred-to-output). Fig. 4 depicts this family of models, denominated as Tissue-prediction injection. Since the tissue and cell patches represent different regions, we need to align the content between the tissue and cell feature maps before concatenation. Therefore, we crop the cell corresponding region from the tissue predictions, upsample them, and, finally, concatenate them in the channel dimensions of the feature maps of the cell branch, as illustrated in Fig. 5a.", + "bbox": [ + 496, + 547, + 890, + 790 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.3. Cell-Tissue Feature Sharing Model", + "text_level": 1, + "bbox": [ + 500, + 800, + 803, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Tissue-prediction injection models share the tissue prediction in a single location and direction, i.e., tissue-to-cell. To enable a more diverse and flexible cell-tissue information flow, we also explore bi-directional feature map sharing from cell-to-tissue (Fig. 5, left) and tissue-to-cell (Fig. 5,", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "$^{2}$ True positive (TP), false positive (FP), and false negative (FN) counts are determined following [41]. If a detected cell is within a valid distance $(\\approx 3\\mu m)$ from an annotated cell and the cell class matches, it is counted as a TP, otherwise an FP. If an annotated cell is not detected, it is counted as an FN. Then, the mean F1 score across classes is computed.", + "bbox": [ + 75, + 838, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "23906", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6c2472a0393278468d9b7427a30da9677432ee38f69c90d1706abdcaf23c9418.jpg", + "image_caption": [ + "(a) Tissue to Cell" + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 261, + 167 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0d45b9b009457d65f00de40d4c95623489bbb49bf540608eac69a9b3dfa5797a.jpg", + "image_caption": [ + "(b) Cell to Tissue" + ], + "image_footnote": [], + "bbox": [ + 277, + 88, + 460, + 167 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/708c0a899a6963e3a3c6d719e5cfda1e936012867f1aa85c7165d6ca25edb7c1.jpg", + "image_caption": [ + "Figure 5. Information is shared between cell and tissue branches via channel-wise concatenation preceded by a shallow convolutional layer with $3 \\times 3$ kernel size. Cropping and upsampling (in Fig. 5a) or downsampling and zero-padding (in Fig. 5b) is applied to match the patch sizes and pixel-alignment between two feature maps from different FoVs. The cell and tissue feature maps are represented in orange and blue, respectively. The red contour denotes the cell patch-associated region in the tissue patch, and the gray regions represent zero padding.", + "Figure 6. Cell-Tissue Feature Sharing Model has two branches for tissue segmentation and cell detection. Information exchange occurs multiple times between the two branches, indicated by the red vertical arrows. Details regarding the information exchange procedures are described in Fig. 5." + ], + "image_footnote": [], + "bbox": [ + 81, + 343, + 467, + 510 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "right). Considering these two operations, we conduct an architecture search procedure to find the optimal feature map sharing configuration between both branches. To limit the search space, we constrain it to only 3 positions in the architecture: after the encoder, after the ASPP module, or after the decoder. Furthermore, we also exclusively allow the branches to inject feature maps at the same depth or position. Finally, we consider only the best-performing model among the $4^{3}$ candidates, which is presented in Fig. 6. We name this model as cell-tissue feature sharing.", + "bbox": [ + 75, + 614, + 468, + 765 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6. Experiments and Results", + "text_level": 1, + "bbox": [ + 76, + 782, + 310, + 800 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We validate the hypothesis that incorporating cell-tissue relationships within a cell detection model is beneficial by evaluating the proposed models on OCELOT and 2 other datasets. First, we describe the additional datasets and implementation details. Then, we show how the proposed multi-task learning methods can improve the performance", + "bbox": [ + 75, + 810, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "of a cell detection task. Lastly, we conduct an ablation study to investigate how important using large FoV patches $\\{x_{l}\\}$ and corresponding tissue segmentation labels $\\{y_{l}^{t}\\}$ are in enhancing cell-tissue relationships.", + "bbox": [ + 496, + 90, + 890, + 151 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.1. Other Datasets", + "text_level": 1, + "bbox": [ + 500, + 161, + 651, + 175 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "TIGER. As mentioned in Sec. 2.1, TIGER is a public dataset, based on H&E stained images, which includes both cell and tissue annotations in overlapping patches. However, the patch sizes and amount of overlap between cell and tissue annotations highly vary among samples. Therefore, a preprocessing step is necessary to generate paired cell-tissue patches with consistent FoVs. After pre-processing, we obtain 9,888 paired patches of size $512 \\times 512$ and $128 \\times 128$ pixels for the tissue segmentation and cell detection tasks, respectively, with an MPP of approximately $0.5\\mu m / px$ . Most samples in TIGER include small regions with annotated cells, making it necessary to extract smaller patches. This explains why the absolute number of patches is larger than in OCELOT or CARP, although the actual amount of annotated cells and tissue area are smaller than the latter datasets, as shown in Tab. 1. In TIGER, one cell class (lymphocyte cells) and seven tissue classes are annotated. More details about the pre-processing and this dataset are provided in the supplementary material.", + "bbox": [ + 496, + 184, + 890, + 470 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "CARP. This is an internal lung cancer dataset of PD-L1 IHC-stained WSIs, containing 6,480 paired patches extracted from 1,012 WSIs. The patch sizes, resolution, annotation protocol, and general configuration are similar to OCELOT, as described in Sec. 3.1. Two cell classes are annotated: PD-L1 positive tumor cells $(TC+)$ , and PD-L1 negative tumor cells $(TC-$ ). The tissue classes are the same as in OCELOT (BG, CA, and UNK). CARP is a real-world and large-scale dataset, with 809.1K annotated cells and 4.1080 mm $^2$ tissue area, which is approximately 10 and 20 times more than OCELOT and TIGER, respectively. Moreover, the stain type is different from OCELOT and TIGER, allowing us to validate the effectiveness of the cell-tissue relationships across different staining methods.", + "bbox": [ + 496, + 474, + 892, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.2. Implementation details", + "text_level": 1, + "bbox": [ + 500, + 696, + 714, + 712 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As previously mentioned in Sec. 5.1, both cell and tissue branches are based on DeepLabV3+ [10], with a ResNet-34 [18] encoder. The models are trained for 300, 150, and 100 epochs in OCELOT, TIGER, and CARP datasets, respectively. We use the Adam optimizer [24], and, for each experiment, we tune the learning rate within a $[5 \\times 10^{-5}, 2 \\times 10^{-3}]$ range. The model at the epoch with the best validation set performance is chosen and used for evaluation on the test set. All experiments are repeated 5 times, and we report the mean and $95\\%$ confidence interval of the performance metrics. For more implementation details, please refer to the supplementary material.", + "bbox": [ + 496, + 719, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "23907", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/9220e1754061a1d4841b3cea96d5e5d4e00d2235917aa5f5102af2cb0b249a23.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodOCELOTTIGERCARP
ValTestValTestValTest
Cell-only68.87±1.7664.44±1.8263.89±1.3953.82±1.2378.48±0.6970.96±1.47
Pred-to-input73.36±0.5969.65±3.9366.00±2.0053.29±1.3079.46±0.7972.98±0.82
Pred-to-inter-172.74±0.5070.54±2.2066.19±1.0255.87±1.7879.74±0.8073.05±0.69
Pred-to-inter-272.68±1.5871.23±0.9665.43±1.1454.75±2.2579.87±0.7873.14±1.53
Pred-to-output66.85±5.6265.05±3.7263.02±0.1653.32±0.4278.92±0.6072.61±0.95
Feature-sharing72.30±0.7368.91±2.5265.64±1.0755.10±2.1879.38±0.7473.00±0.33
Tissue-label leaking76.56±0.8074.20±0.9169.71±0.6161.66±1.1680.13±1.0472.97±0.49
", + "bbox": [ + 183, + 88, + 782, + 255 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7bde9d73b82cbee8dfc74afec089f270e976264f5024391b13e4f414a48e3e47.jpg", + "table_caption": [ + "Table 5. Cell detection mean F1 scores per model. With the exception of the Tissue-label Leaking model, the highest score is written in bold and the second highest score is underlined." + ], + "table_footnote": [], + "table_body": "
Methodpopulation (130)kidney (41)endometrium (25)bladder (26)prostate (16)stomach (12)head-neck (10)
Cell-only64.44±1.8264.12±3.4166.88±3.7159.42±5.6165.46±1.6466.19±8.9359.24±5.04
Pred-to-input69.65±3.9363.37±3.2773.31±6.3563.61±4.5568.36±8.6169.34±7.1575.28±0.82
Pred-to-inter-170.54±2.2066.62±7.3073.05±4.1464.35±3.5970.20±2.5671.55±5.2174.50±1.52
Pred-to-inter-271.23±0.9668.94±5.6575.15±2.7064.94±2.3468.35±8.8370.29±0.8173.74±1.72
Pred-to-output65.05±3.7263.38±4.4868.21±4.1159.81±3.9664.18±8.4167.51±6.3960.88±6.77
Feature-sharing68.91±2.5264.97±5.3371.28±5.3063.21±5.3166.25±9.5269.79±2.2573.88±4.39
Tissue-label leaking74.20±0.9175.75±1.9772.71±1.1271.18±0.9374.56±1.2575.24±1.7175.78±0.46
", + "bbox": [ + 133, + 308, + 838, + 470 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.3. Main Results", + "text_level": 1, + "bbox": [ + 76, + 535, + 215, + 550 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The cell detection results obtained by the Cell-only baseline, the proposed approaches, and the Tissue-label leaking model are shown in Tab. 5. The Cell-only baseline is a simple cell detection model that only receives the small FoV patches as input. In other words, it only considers the cell branch in Fig. 4 and neither tissue annotations nor large FoV patches are leveraged. The Tissue-label leaking model, described in Sec. 5.2, receives the tissue annotations as input and serves as an exploratory experiment to obtain insight regarding how much the cell detection task can benefit from leveraging tissue annotation.", + "bbox": [ + 75, + 561, + 468, + 728 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "From Tab. 5, we observe that all the cell-tissue multi-task learning-based approaches, except for the Pred-to-output model, outperform the Cell-only baseline across all datasets. These improvements imply that cell detection on small FoV patches benefits from learning cell-tissue interrelationships from both large FoVs and tissue annotations. We hypothesize that the reason for the low performance of the Pred-to-output model is that the injection of tissue prediction to the cell branch happens too late. Therefore, there is a lack of network capacity for fusing cell and tissue information.", + "bbox": [ + 75, + 729, + 470, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Furthermore, in Tab. 6, we report mean F1 scores per or", + "bbox": [ + 96, + 885, + 468, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/59f0770bdd69a5c03d15e4774b5a106d1444be1af784293b8657c7616af3697d.jpg", + "table_caption": [ + "Table 6. Per-organ cell detection mean F1 scores in the OCELOT test set. Population mean F1 scores are computed from all the patches in the test set. In parentheses, we indicate the number of samples of each organ subset. With the exception of the Tissue-label Leaking model, the highest score is written in bold and the second highest score is underlined." + ], + "table_footnote": [], + "table_body": "
Further utilizePred-to-inter-2Feature-sharing
Tissue labelLarge FoVValTestValTest
--68.87±1.7664.45±1.8268.87±1.7664.45±1.82
✓-70.21±1.4964.99±1.7770.36±1.3466.58±2.02
-✓68.78±0.9465.56±3.6669.61±1.5966.22±1.53
✓✓72.68±1.5871.23±0.9672.30±0.7368.91±2.52
", + "bbox": [ + 504, + 532, + 890, + 626 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 7. Dataset ablation study. In all experiments, the model architecture is fixed and only input/output changes. The first row corresponds to the Cell-only baseline.", + "bbox": [ + 496, + 636, + 892, + 679 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "gan in the OCELOT test set. Except for the Pred-to-output, all approaches show improvement compared to Cell-only baseline in most organs. Especially, our best-performing model, Pred-to-inter-2, demonstrates superior performance over Cell-only in all 6 organs with a significant margin. Such a result shows that considering cell-tissue relationships can generally help the cell detection task across various organs.", + "bbox": [ + 496, + 685, + 890, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.4. Dataset Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 801, + 715, + 816 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Two data components in OCELOT contribute to better leveraging cell-tissue relationships for the task of cell detection: (1) the large FoV patches $x_{l}$ , and (2) the corresponding tissue annotations $y_{l}^{t}$ that allows us to have a tissue segmentation objective and a multi-task learning setting. To verify", + "bbox": [ + 496, + 824, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "23908", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a661531a0b3009d2bc0a111da695b1eb932502a0c9c63f3295fe923ab328e9d6.jpg", + "image_caption": [ + "Figure 7. Qualitative comparison between the Pred-to-inter-2 and Cell-only models. Pred-to-inter-2 shows enhanced detection results aided by understanding the tissue environment from a broader context (blue: tumor cells, yellow: background cells, green: cancer area)." + ], + "image_footnote": [], + "bbox": [ + 127, + 85, + 844, + 359 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the effectiveness of each data component, we conduct an ablation study where each component is evaluated separately while keeping the model architecture fixed. First, if only the tissue segmentation label is provided without a large FoV patch, the tissue branch takes $x_{s}$ as an input and it is trained with both the tissue and the cell detection objectives. Second, if the large FoV patch is given without the corresponding tissue label, the tissue branch takes $x_{l}$ as an input and the model is only trained with the cell detection objective.", + "bbox": [ + 75, + 407, + 470, + 542 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The results of the ablation study are shown in Tab. 7. We observe that each of the individual components on their own lead to improved performance. Hence, we can conclude that a large FoV patch and tissue segmentation label can contribute to better cell detection by providing a broader context and encoding the cell-tissue relationship, respectively. Moreover, when utilized simultaneously, we observe that both components work synergically, leading to an even better performance improvement. We also found that the tissue branch generates better tissue predictions when trained with a large FoV. Thus, providing accurate tissue information to the cell branch is important for boosting performance. Please, refer to the supplementary material for a detailed tissue performance comparison. Finally, this study shows evidence that the patch configuration defined in OCELOT is suitable for cell-tissue co-training, with a smaller FoV for cell detection $(x_{s},y_{s}^{c})$ and a larger FoV for tissue segmentation $(x_{l},y_{l}^{t})$ .", + "bbox": [ + 75, + 544, + 470, + 819 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.5. Qualitative Analysis", + "text_level": 1, + "bbox": [ + 76, + 830, + 269, + 847 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section, we visualize the cell predictions of the Cell-only and Pred-to-inter-2 models. We select the Pred-to-inter-2 model, as it is the overall best-performing model in", + "bbox": [ + 76, + 854, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "our experiments. In general, when compared to other background cells, tumor cells have the following characteristics: large size and irregular shape. However, cancer is heterogeneous and this is not always the case. Indeed, in Fig. 7, most of the cells are small and have a regular round shape. Based on these appearances, and without a larger context, those cells can be easily misclassified as background cells, which is the case of the Cell-only model. On the other hand, Pred-to-inter-2 shows a more accurate prediction by correctly understanding the cancer area in large FoV regions. This implies that Pred-to-inter-2 indeed considers both the morphology of cells and the tissue context, while Cell-only relies on the cells' morphology alone.", + "bbox": [ + 496, + 407, + 893, + 604 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 628, + 619, + 645 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In histopathology, there is a hierarchical organization and interrelationship between cells and tissues. Hence, we hypothesize that it can be leveraged to improve the performance of cell detection tasks. Throughout this paper, we have shown evidence that such cell-tissue relationships exist across multiple setups. Indeed, we observe improvements when utilizing large FoVs with corresponding tissue annotations together with cell annotations in a multi-task learning approach, using simple neural architecture designs. Although the cell detection task clearly benefits from the multitask learning framework, the improvement on the tissue segmentation task was not yet investigated, and we leave this topic as a future direction. We hope this work, especially the OCELOT dataset, initiates new and promising research directions in CPATH to further study the cell-tissue relationships and improve cell detection and tissue segmentation.", + "bbox": [ + 496, + 657, + 893, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "23909", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] MICCAI 2014 Grand Challenges. https://wiki.cancerimagingarchive.net/display/Public/MICCAI+2014+Grand+Challenges, 2014. [Online; accessed Nov-2022]. 2", + "[2] TIGER: Grand Challenge. https://tiger.grand-challenge.org, 2022. [Online; accessed Nov-2022]. 2", + "[3] Esther Abels, Liron Pantanowitz, Famke Aeffner, Mark D Zarella, Jeroen van der Laak, Marilyn M Bui, Venkata NP Vemuri, Anil V Parwani, Jeff Gibbs, Emmanuel Agosto-Arroyo, et al. Computational pathology definitions, best practices, and recommendations for regulatory guidance: a white paper from the digital pathology association. The Journal of Pathology, 249(3):286-294, 2019. 1", + "[4] Mohamed Amgad, Lamees A Atteya, Hagar Hussein, Kareem Hosny Mohammed, Ehab Hafiz, Maha AT Elsebaie, Ahmed M Alhusseiny, Mohamed Atef AlMoslemany, Abdelmagid M Elmatboly, Philip A Pappalardo, et al. Nuclei: A scalable crowdsourcing approach and dataset for nucleus classification and segmentation in breast cancer. *Giga* Science, 11, 2022. 2", + "[5] Vahid Azimi, Young Hwan Chang, Guillaume Thibault, Jaclyn Smith, Takahiro Tsujikawa, Benjamin Kukull, Braden Jensen, Christopher Corless, Adam Margolin, and Joe W Gray. Breast cancer histopathology image analysis pipeline for tumor purity estimation. In 2017 IEEE 14th International Symposium on Biomedical Imaging (ISBI 2017), pages 1137-1140. IEEE, 2017. 4", + "[6] Tian Bai, Jiayu Xu, and Fuyong Xing. Multi-field of view aggregation and context encoding for single-stage nucleus recognition. In MICCAI, pages 382-392. Springer, 2020. 3", + "[7] Wouter Bulten, Kimmo Kartasalo, Po-Hsuan Cameron Chen, Peter Ström, Hans Pinckaers, Kunal Nagpal, Yuannan Cai, David F Steiner, Hester van Boven, Robert Vink, et al. Artificial intelligence for diagnosis and gleason grading of prostate cancer: the panda challenge. Nature Medicine, 28(1):154-163, 2022. 1, 2", + "[8] Lyndon Chan, Mahdi S Hosseini, Corwyn Rowsell, Konstantinos N Plataniotis, and Savvas Damaskinos. Histosegnet: Semantic segmentation of histological tissue type in whole slide images. In ICCV, pages 10662-10671, 2019. 2", + "[9] Hao Chen, Xiaojuan Qi, Lequan Yu, and Pheng-Ann Heng. Dcan: deep contour-aware networks for accurate gland segmentation. In CVPR, pages 2487–2496, 2016. 2", + "[10] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In ECCV, pages 801-818, 2018. 5, 6", + "[11] Qian Da, Xiaodi Huang, Zhongyu Li, Yanfei Zuo, Chenbin Zhang, Jingxin Liu, Wen Chen, Jiahui Li, Dou Xu, Zhiqiang Hu, et al. Digestpath: a benchmark dataset with challenge review for the pathological detection and segmentation of digestive-system. Medical Image Analysis, page 102485, 2022. 1, 2", + "[12] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248–255. IEEE, 2009. 3" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] James A Diao, Jason K Wang, Wan Fung Chui, Victoria Mountain, Sai Chowdary Gullapally, Ramprakash Srinivasan, Richard N Mitchell, Benjamin Glass, Sara Hoffman, Sudha K Rao, et al. Human-interpretable image features derived from densely mapped cancer pathology slides predict diverse molecular phenotypes. Nature communications, 12(1):1-15, 2021. 1", + "[14] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. *IJCV*, 88(2):303–338, 2010. 3", + "[15] Jevgenij Gamper, Navid Alemi Koohbanani, Ksenija Benes, Simon Graham, Mostafa Jahanifar, Syed Ali Khurram, Ayesha Azam, Katherine Hewitt, and Nasir Rajpoot. Pan-nuke dataset extension, insights and baselines. arXiv preprint arXiv:2003.10778, 2020. 2", + "[16] Zhe Gong, Jieyun Zhang, and Weijian Guo. Tumor purity as a prognosis and immunotherapy relevant feature in gastric cancer. Cancer Medicine, 9(23):9052-9063, 2020. 4", + "[17] Simon Graham, Quoc Dang Vu, Shan E Ahmed Raza, Ayesha Azam, Yee Wah Tsang, Jin Tae Kwak, and Nasir Rajpoot. Hover-net: Simultaneous segmentation and classification of nuclei in multi-tissue histology images. Medical Image Analysis, 58:101563, 2019. 1, 2", + "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. pages 770-778, 2016. 6", + "[19] David Joon Ho, Dig VK Yarlagadda, Timothy M D'Alfonso, Matthew G Hanna, Anne Grabenstetter, Peter Ntiamoah, Edi Brogi, Lee K Tan, and Thomas J Fuchs. Deep multimagnification networks for multi-class breast cancer image segmentation. Computerized Medical Imaging and Graphics, 88:101866, 2021. 3", + "[20] Mahdi S Hosseini, Lyndon Chan, Gabriel Tse, Michael Tang, Jun Deng, Sajad Norouzi, Corwyn Rowsell, Konstantinos N Plataniotis, and Savvas Damaskinos. Atlas of digital pathology: A generalized hierarchical histological tissue type-annotated database for deep learning. In CVPR, pages 11747–11756, 2019. 2", + "[21] Carolyn Hutter and Jean Claude Zenklusen. The cancer genome atlas: creating lasting value beyond its data. Cell, 173(2):283-285, 2018. 3", + "[22] Konstantinos Kamnitsas, Christian Ledig, Virginia FJ Newcombe, Joanna P Simpson, Andrew D Kane, David K Menon, Daniel Rueckert, and Ben Glocker. Efficient multi-scale 3d cnn with fully connected crf for accurate brain lesion segmentation. Medical Image Analysis, 36:61-78, 2017. 3", + "[23] Davood Karimi, Guy Nir, Ladan Fazli, Peter C Black, Larry Goldenberg, and Septimiu E Salcudean. Deep learning-based gleason grading of prostate cancer from histopathology images—role of multiscale decision aggregation and data augmentation. IEEE journal of Biomedical and Health Informatics, 24(5):1413-1426, 2019. 2", + "[24] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 6", + "[25] Neeraj Kumar, Ruchika Verma, Deepak Anand, Yanning Zhou, Omer Fahri Onder, Efstratios Tsougenis, Hao Chen, Pheng-Ann Heng, Jiahui Li, Zhiqiang Hu, et al. A multi" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "23910", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "organ nucleus segmentation challenge. IEEE transactions on Medical Imaging, 39(5):1380-1391, 2019. 1, 2", + "[26] Shyam Lal, Devikalyan Das, Kumar Alabhya, Anirudh Kanfade, Aman Kumar, and Jyoti Kini. Nucleisegnet: robust deep learning architecture for the nuclei segmentation of liver cancer histopathology images. Computers in Biology and Medicine, 128:104075, 2021. 1, 2", + "[27] Chunggi Lee, Seonwook Park, Heon Song, Jeongun Ryu, Sanghoon Kim, Haejoon Kim, Sérgio Pereira, and Donggeun Yoo. Interactive multi-class tiny-object detection. In CVPR, pages 14136-14145, 2022. 3", + "[28] Wenqi Li, Siyamalan Manivannan, Shazia Akbar, Jianguo Zhang, Emanuele Trucco, and Stephen J McKenna. Gland segmentation in colon histology images using hand-crafted features and convolutional neural networks. In 2016 IEEE 13th international symposium on biomedical imaging (ISBI), pages 1405-1408. IEEE, 2016. 2", + "[29] Xieli Li, Yuanyuan Wang, Qisheng Tang, Zhen Fan, and Jinhua Yu. Dual u-net for the segmentation of overlapping glioma nuclei. IEEE Access, 7:84040-84052, 2019. 2", + "[30] Dongnan Liu, Donghao Zhang, Yang Song, Chaoyi Zhang, Fan Zhang, Lauren O'Donnell, and Weidong Cai. Nuclei segmentation via a deep panoptic model with semantic feature fusion. In *IJCAI*, pages 861-868, 2019. 2", + "[31] Yihao Mao, Qingyang Feng, Peng Zheng, Liangliang Yang, Tianyu Liu, Yuqiu Xu, Dexiang Zhu, Wenju Chang, Meiling Ji, Li Ren, et al. Low tumor purity is associated with poor prognosis, heavy mutation burden, and intense immune phenotype in colon cancer. Cancer Management and Research, 10:3569, 2018. 4", + "[32] Peter Naylor, Marick Laé, Fabien Reyal, and Thomas Walter. Segmentation of nuclei in histopathology images by deep regression of the distance map. IEEE transactions on Medical Imaging, 38(2):448-459, 2018. 2", + "[33] Guy Nir, Soheil Hor, Davood Karimi, Ladan Fazli, Brian F Skinnider, Peyman Tavassoli, Dmitry Turbin, Carlos F Villamil, Gang Wang, R Storey Wilson, et al. Automatic grading of prostate cancer in digitized histopathology images: Learning from multiple experts. Medical Image Analysis, 50:167-180, 2018. 2", + "[34] Sehnoon Park, Chan-Young Ock, Hyojin Kim, Sergio Pereira, Seonwook Park, Minuk Ma, Sangjoon Choi, Seokhwi Kim, Seunghwan Shin, Brian Jaehong Aum, et al. Artificial intelligence-powered spatial analysis of tumor-infiltrating lymphocytes as complementary biomarker for immune checkpoint inhibition in non-small-cell lung cancer. Journal of Clinical Oncology, 40(17):1916, 2022. 1", + "[35] Pushpak Pati, Guillaume Jaume, Lauren Alisha Fernandes, Antonio Foncubierta-Rodriguez, Florinda Feroce, Anna Maria Anniciello, Giosue Scognamiglio, Nadia Brancati, Daniel Riccio, Maurizio Di Bonito, et al. Hact-net: A hierarchical cell-to-tissue graph neural network for histopathological image classification. In Uncertainty for Safe Utilization of Machine Learning in Medical Imaging, and Graphs in Biomedical Image Analysis, pages 208-219. Springer, 2020. 3", + "[36] Ziniu Qian, Kailu Li, Maode Lai, Eric I Chang, Bingzheng Wei, Yubo Fan, Yan Xu, et al. Transformer based multiple" + ], + "bbox": [ + 78, + 92, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "instance learning for weakly supervised histopathology image segmentation. arXiv preprint arXiv:2205.08878, 2022. 2", + "[37] Hui Qu, Pengxiang Wu, Qiaoying Huang, Jingru Yi, Zhennan Yan, Kang Li, Gregory M Riedlinger, Subhajyoti De, Shaoting Zhang, and Dimitris N Metaxas. Weakly supervised deep nuclei segmentation using partial points annotation in histopathology images. IEEE transactions on Medical Imaging, 39(11):3655-3666, 2020. 1, 2", + "[38] Roberto Salgado, Carsten Denkert, S Demaria, N Sirtaine, F Klauschen, Giancarlo Pruneri, S Wienert, Gert Van den Eynden, Frederick L Baehner, Frederique Pénault-Llorca, et al. The evaluation of tumor-infiltrating lymphocytes (tils) in breast cancer: recommendations by an international tils working group 2014. Annals of Oncology, 26(2):259-271, 2015. 2", + "[39] Rüdiger Schmitz, Frederic Madesta, Maximilian Nielsen, Jenny Krause, Stefan Steurer, René Werner, and Thomas Rösch. Multi-scale fully convolutional neural networks for histopathology image segmentation: from nuclear aberrations to the global tissue architecture. Medical Image Analysis, 70:101996, 2021. 3", + "[40] Peng Sun, Jiehua He, Xue Chao, Keming Chen, Yuanyuan Xu, Qitao Huang, Jingping Yun, Mei Li, Rongzhen Luo, Jinbo Kuang, et al. A computational tumor-infiltrating lymphocyte assessment method comparable with visual reporting guidelines for triple-negative breast cancer. EBioMedicine, 70:103492, 2021. 1", + "[41] Zaneta Swiderska-Chadaj, Hans Pinckaers, Mart van Rijthoven, Maschenka Balkenhol, Margarita Melnikova, Oscar Geessink, Quirine Manson, Mark Sherman, Antonio Polonia, Jeremy Parry, et al. Learning to detect lymphocytes in immunohistochemistry with deep learning. Medical Image Analysis, 58:101547, 2019. 5", + "[42] Hiroki Tokunaga, Yuki Teramoto, Akihiko Yoshizawa, and Ryoma Bise. Adaptive weighting multi-field-of-view cnn for semantic segmentation in pathology. In CVPR, pages 12597-12606, 2019. 1, 3", + "[43] Mart Van Rijthoven, Maschenka Balkenhol, Karina Silina, Jeroen Van Der Laak, and Francesco Ciompi. Hooknet: Multi-resolution convolutional neural networks for semantic segmentation in histopathology whole-slide images. Medical Image Analysis, 68:101890, 2021. 3", + "[44] Ruchika Verma, Neeraj Kumar, Abhijeet Patil, Nikhil Cherian Kurian, Swapnil Rane, Simon Graham, Quoc Dang Vu, Mieke Zwager, Shan E Ahmed Raza, Nasir Rajpoot, et al. Monusac2020: A multi-organ nuclei segmentation and classification challenge. IEEE Transactions on Medical Imaging, 40(12):3413-3423, 2021. 2", + "[45] Quoc Dang Vu, Simon Graham, Tahsin Kurc, Minh Nguyen Nhat To, Muhammad Shaban, Talha Qaiser, Navid Alemi Koohbanani, Syed Ali Khurram, Jayashree Kalpathy-Cramer, Tianhao Zhao, et al. Methods for segmentation and classification of digital microscopy tissue images. Frontiers in Bioengineering and Biotechnology, page 53, 2019. 2", + "[46] Jingwen Wang, Richard J Chen, Ming Y Lu, Alexander Baras, and Faisal Mahmood. Weakly supervised prostate" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "23911", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tma classification via graph convolutional networks. In 2020 IEEE 17th International Symposium on Biomedical Imaging (ISBI), pages 239-243. IEEE, 2020. 3", + "[47] Yan Xu, Zhipeng Jia, Liang-Bo Wang, Yuqing Ai, Fang Zhang, Maode Lai, Eric I Chang, et al. Large scale tissue histopathology image classification, segmentation, and visualization via deep convolutional activation features. BMC bioinformatics, 18(1):1-17, 2017. 2", + "[48] Chuanbao Zhang, Wen Cheng, Xiufang Ren, Zheng Wang, Xing Liu, Guanzhang Li, Sheng Han, Tao Jiang, and Anhua Wu. Tumor purity as an underlying key factor in gliomatumor purity in glioma. Clinical Cancer Research, 23(20):6279-6291, 2017. 4", + "[49] Bingchao Zhao, Xin Chen, Zhi Li, Zhiwen Yu, Su Yao, Lixu Yan, Yuqian Wang, Zaiyi Liu, Changhong Liang, and Chu Han. Triple u-net: Hematoxylin-aware nuclei segmentation with progressive dense feature aggregation. Medical Image Analysis, 65:101786, 2020. 1, 2", + "[50] Ke Zhao, Zhenhui Li, Su Yao, Yingyi Wang, Xiaomei Wu, Zeyan Xu, Lin Wu, Yanqi Huang, Changhong Liang, and Zaiyi Liu. Artificial intelligence quantified tumour-stroma ratio is an independent predictor for overall survival in resectable colorectal cancer. EBioMedicine, 61:103054, 2020. 2", + "[51] Yanning Zhou, Simon Graham, Navid Alemi Koohbanani, Muhammad Shaban, Pheng-Ann Heng, and Nasir Rajpoot. Cgc-net: Cell graph convolutional network for grading of colorectal cancer histology images. In ICCV Workshops, pages 0–0, 2019. 3", + "[52] Chuang Zhu, Ke Mei, Ting Peng, Yihao Luo, Jun Liu, Ying Wang, and Mulan Jin. Multi-level colonoscopy malignant tissue detection with adversarial cac-unet. Neurocomputing, 438:165-183, 2021. 2" + ], + "bbox": [ + 78, + 90, + 470, + 555 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "23912", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/OCELOT_ Overlapped Cell on Tissue Dataset for Histopathology/052d2272-757d-4e0e-b0c0-b44903da46a8_model.json b/2023/OCELOT_ Overlapped Cell on Tissue Dataset for Histopathology/052d2272-757d-4e0e-b0c0-b44903da46a8_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ef7df7b32e0dcc9e4320e5c87c1589990155ffe2 --- /dev/null +++ b/2023/OCELOT_ Overlapped Cell on Tissue Dataset for Histopathology/052d2272-757d-4e0e-b0c0-b44903da46a8_model.json @@ -0,0 +1,2191 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.157, + 0.131, + 0.816, + 0.154 + ], + "angle": 0, + "content": "OCELOT: Overlapped Cell on Tissue Dataset for Histopathology" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.181, + 0.882, + 0.237 + ], + "angle": 0, + "content": "Jeongun Ryu* Aaron Valero Puche* JaeWoong Shin* Seonwook Park Biagio Brattoli Jinhee Lee Wonkyung Jung Soo Ick Cho Kyunghyun Paeng Chan-Young Ock Donggeun Yoo Sérgio Pereira Lunit Inc." + }, + { + "type": "text", + "bbox": [ + 0.186, + 0.243, + 0.774, + 0.275 + ], + "angle": 0, + "content": "{rjw0205, aaron.valero, jwoong.shin, spark, biagio, jinhee.lee, wkjung, sooickcho, khpaeng, ock.chanyoung, dgyoo, sergio}@lunit.io" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.299, + 0.313, + 0.315 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.331, + 0.474, + 0.739 + ], + "angle": 0, + "content": "Cell detection is a fundamental task in computational pathology that can be used for extracting high-level medical information from whole-slide images. For accurate cell detection, pathologists often zoom out to understand the tissue-level structures and zoom in to classify cells based on their morphology and the surrounding context. However, there is a lack of efforts to reflect such behaviors by pathologists in the cell detection models, mainly due to the lack of datasets containing both cell and tissue annotations with overlapping regions. To overcome this limitation, we propose and publicly release OCELOT, a dataset purposely dedicated to the study of cell-tissue relationships for cell detection in histopathology. OCELOT provides overlapping cell and tissue annotations on images acquired from multiple organs. Within this setting, we also propose multi-task learning approaches that benefit from learning both cell and tissue tasks simultaneously. When compared against a model trained only for the cell detection task, our proposed approaches improve cell detection performance on 3 datasets: proposed OCELOT, public TIGER, and internal CARP datasets. On the OCELOT test set in particular, we show up to 6.79 improvement in F1-score. We believe the contributions of this paper, including the release of the OCELOT dataset at https://lunit-io.github.io/research/publications/ocelot are a crucial starting point toward the important research direction of incorporating cell-tissue relationships in computation pathology." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.749, + 0.21, + 0.765 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.773, + 0.471, + 0.88 + ], + "angle": 0, + "content": "Computational Pathology (CPATH) [3] is a branch of digital pathology that develops methodologies for the analysis of digitized patient specimens, such as Whole-Slide-Images (WSIs). Cell detection in histology images [26, 37, 49] is one of the most important tasks in CPATH. It allows the quantification and analysis of different cell types, which can lead to better prognosis evaluation [34, 40] and patient" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.301, + 0.892, + 0.361 + ], + "angle": 0, + "content": "treatment planning while maintaining medical interpretability [13]. Since it has the potential to impact human lives, high-performance cell detection models are essential in real-world applications and need to be investigated." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.365, + 0.893, + 0.561 + ], + "angle": 0, + "content": "To better locate and classify cells, detailed morphological characteristics such as color and shape are crucial. Consequently, cell detection datasets are typically collected at high magnification but small Field-of-View (FoV). However, this can make the cell detection model overly rely on appearance details, without understanding the broader context [42]. This context can help cell detection by providing information about how cells are arranged and grouped together to form high-level tissue structures. In practice, expert annotators (pathologists) first zoom out to understand these broad tissue structures. Next, they zoom in to better classify individual cells while taking into account the context information, as depicted in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.565, + 0.895, + 0.717 + ], + "angle": 0, + "content": "The behavior of pathologists can be transferred to deep learning, for instance, through a multi-task strategy combining cell detection tasks at high magnification and tissue segmentation at low magnification. This type of approach would allow the model to share knowledge across different tasks and FoVs. However, to train such an approach, a combined dataset with cell-tissue overlapping regions is required; unfortunately, most existing datasets only target a single task, either cell detection [17,25] or tissue segmentation [7, 11]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.901 + ], + "angle": 0, + "content": "In this paper, we introduce a new research direction: studying cell-tissue relationships for cell detection. First, we publish the OCELOT dataset, which contains cell and tissue annotations in small and large FoV patches, respectively, with overlapping regions. Additionally, the data is collected from WSIs of multiple organs. This can provide the necessary data for researchers to study cell-tissue relationships and their effect on cell detection. Second, we introduce simple multi-task learning approaches for cell detection that can benefit from cell-tissue relationships and demonstrate their advantages over 3 different datasets. These approaches consistently show better cell detection performance compared to" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.888, + 0.212, + 0.9 + ], + "angle": 0, + "content": "*: Equal contribution" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23902" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.164, + 0.081, + 0.805, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.317, + 0.895, + 0.402 + ], + "angle": 0, + "content": "Figure 1. Behavior of pathologists and cell detection models. Typical cell detection models infer their predictions only by looking at a limited FoV region. Failure cases can occur when cells are difficult to be classified solely by their morphology, i.e., while disregarding the larger architecture of the tissues (context). In the above example, some groups of tumor cells are misclassified as background cells due to their morphology; these tumor cells are smaller and rounder than the nearby ones. Note that the large size and irregular shape are representative characteristics of tumor cells. Pathologists overcome these limitations by further understanding the tissue context of the larger FoV region. Misclassified cells can be easily corrected by understanding that such regions are cancer areas." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.423, + 0.47, + 0.498 + ], + "angle": 0, + "content": "the cell-only baseline, i.e., a model trained only with small FoV patches with the corresponding cell annotations. We hope that our proposed OCELOT dataset and methods will encourage the CPATH community to learn how to reflect cell-tissue relationships better to improve cell detection." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.5, + 0.468, + 0.513 + ], + "angle": 0, + "content": "Our contributions are 3-fold and summarized as follows," + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.524, + 0.469, + 0.554 + ], + "angle": 0, + "content": "- The first work that exploits cell-tissue relationships for better cell detection, to the best of our knowledge." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.561, + 0.469, + 0.606 + ], + "angle": 0, + "content": "- We release OCELOT, a dataset with overlapping cell and tissue annotations based on Hematoxylin and Eosin (H&E) stained WSIs of multiple organs." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.613, + 0.469, + 0.673 + ], + "angle": 0, + "content": "- We introduce several approaches that boost cell detection performance via multi-task learning, and empirically confirm that these methods generalize well across different datasets and histological stainings." + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.524, + 0.469, + 0.673 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.687, + 0.219, + 0.702 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.712, + 0.468, + 0.741 + ], + "angle": 0, + "content": "2.1. Datasets for Cell detection and Tissue segmentation Tasks" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.47, + 0.901 + ], + "angle": 0, + "content": "In recent years, numerous datasets have been released for tackling cell detection. Some of those works only target a single organ [17,32], while others consider multiple ones [4, 15,25,44,45]. The availability of these datasets enables the CPATH community to push forward the development and improvement of cell detection models [17,26,29,30,37,49]. In addition, tissue segmentation datasets have also been proposed for prostate [7,23,33], colorectal [50], brain [1], and multiple organs [20]. Some examples of tissue segmentation works can be found in [8,9,28,36,47,52]. The dataset" + }, + { + "type": "table", + "bbox": [ + 0.557, + 0.42, + 0.837, + 0.478 + ], + "angle": 0, + "content": "
DatasetTissue Area# CellOrgans
OCELOT4.267cm2114.7Kmultiple
TIGER2.536cm250.8Kbreast
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.486, + 0.892, + 0.514 + ], + "angle": 0, + "content": "Table 1. Dataset comparison in terms of physical tissue annotated area and total cell counting per dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.534, + 0.892, + 0.655 + ], + "angle": 0, + "content": "in [11] is composed of a cell detection subset and a tissue segmentation subset. However, the subsets are annotated independently and from different patient groups, and, therefore, there are no overlaps between the cell and tissue data. Because of the lack of overlapping data in the aforementioned datasets, it is difficult to build an end-to-end framework to learn cell-tissue relationships by jointly training on the cell and tissue tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.657, + 0.892, + 0.747 + ], + "angle": 0, + "content": "The recently released TIGER dataset [2] contains both cell and tissue annotations to study tumor-infiltrating lymphocytes [38] in H&E breast cancer WSIs. All the cell-annotated areas exist inside the tissue-annotated area, however, this work does not propose nor initiate any effort toward the integration of both cell and tissue tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.893, + 0.9 + ], + "angle": 0, + "content": "To further promote the development of methods that leverage the cell-tissue relationship for the task of cell detection, we propose OCELOT, which is designed to capture the hierarchical relationship between cells and tissues, especially in the tumor environment. OCELOT contains roughly two times more cell and tissue annotations than TIGER (see Tab. 1). Additionally, the data was collected from multiple organs to enable the investigation of the generalizability of cell-tissue relationships over various cancer types. In the end, we utilized both OCELOT and TIGER to reveal the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23903" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.468, + 0.107 + ], + "angle": 0, + "content": "power of cell-tissue relationship for cell detection in Sec. 6." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.116, + 0.356, + 0.133 + ], + "angle": 0, + "content": "2.2. Leveraging Large Field of View" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.14, + 0.47, + 0.366 + ], + "angle": 0, + "content": "Some studies [6, 19, 22, 39, 42, 43] extract a large FoV region as an additional input to improve detection/segmentation performance on smaller FoV regions. [22] proposes a dual pathway 3D CNN for brain lesion segmentation, where each pathway receives both small and large center-shared FoV regions as input. Similar studies are also conducted in the CPATH domain for tissue segmentation [19, 39, 42, 43] and cell detection [6]. To fuse different FoV patches, [42] introduces a weighting mechanism and [43] proposes a multi-scale merging block composed of convolution and concatenation. Nevertheless, no inter-task relationship is considered in the previously mentioned methods. In contrast, our models take advantage of the large contextual information while learning the cell-tissue relationship via multi-task objectives at different FoVs." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.376, + 0.399, + 0.392 + ], + "angle": 0, + "content": "2.3. Leveraging Cell-Tissue Relationships" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.4, + 0.47, + 0.626 + ], + "angle": 0, + "content": "To our understanding, there is no study that considers cell-tissue relationships for cell detection or tissue segmentation tasks. On the other hand, a few efforts have attempted to link tissue and cell for image classification using graph-based methods [35,46,51]. Such studies represent the tissue structure as a graph of detected cells, based on the medical knowledge that cells form tissues. [35] explicitly considers the cell-tissue relationship for the task of breast cancer subtyping, by an interaction between tissue-level and cell-level graphs with a cell-to-tissue hierarchy. However, these methods treat the cell/tissue graph generation as a pre-processing step, by using the inference output of independently pretrained cell detection and tissue segmentation models. In contrast, we directly target the improvement of cell detection by utilizing the cell-tissue relationship." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.64, + 0.188, + 0.655 + ], + "angle": 0, + "content": "3. OCELOT" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.666, + 0.469, + 0.727 + ], + "angle": 0, + "content": "In this section, we introduce OCELOT, a histopathology dataset specifically built to enable the development of methods that leverage cell and tissue relationships. Each sample of the OCELOT dataset \\(D\\) is composed of six components," + }, + { + "type": "equation", + "bbox": [ + 0.166, + 0.738, + 0.469, + 0.761 + ], + "angle": 0, + "content": "\\[\n\\mathcal {D} = \\left\\{\\left(x _ {s}, y _ {s} ^ {c}, x _ {l}, y _ {l} ^ {t}, c _ {x}, c _ {y}\\right) _ {i} \\right\\} _ {i = 1} ^ {N} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.47, + 0.901 + ], + "angle": 0, + "content": "where \\(x_{s}, x_{l}\\) are the small and large FoV patches extracted from the WSI, \\(y_{s}^{c}, y_{l}^{t}\\) refer to the corresponding cell and tissue annotations, respectively, and \\(c_{x}, c_{y}\\) are the relative coordinates of the center of \\(x_{s}\\) within \\(x_{l}\\). We drop the sample index \\(i\\) for simplicity. Fig. 2 shows the visualization of a sample in OCELOT. More details about the dataset including data collection and statistics can be found in the following sub-sections. The dataset is publicly available at https://lunit-io.github.io/research/publications/ocelot/." + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.092, + 0.695, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.581, + 0.239, + 0.621, + 0.253 + ], + "angle": 0, + "content": "\\((x_{l},y_{l}^{t})\\)" + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.092, + 0.888, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.773, + 0.239, + 0.815, + 0.252 + ], + "angle": 0, + "content": "\\((x_{s},y_{s}^{c})\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.261, + 0.892, + 0.372 + ], + "angle": 0, + "content": "Figure 2. A sample from the OCELOT dataset. Each sample of the dataset consists of two input patches and the corresponding annotations. Left shows the large FoV patch \\( x_{l} \\) with tissue segmentation annotation \\( y_{l}^{t} \\), where green denotes the cancer area. Right shows the small FoV patch \\( x_{s} \\) with cell point annotation \\( y_{s}^{c} \\), where blue and yellow dots denote tumor and background cells, respectively. The red box indicates the size and location of the \\( x_{s} \\) with respect to the \\( x_{l} \\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.391, + 0.656, + 0.405 + ], + "angle": 0, + "content": "3.1. Data Collection" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.416, + 0.892, + 0.552 + ], + "angle": 0, + "content": "We collect 306 TCGA [21] WSIs from a total of 6 different organs: kidney, head-and-neck, prostate, stomach, endometrium, and bladder. From each of the WSIs, we select 1 to 3 large Regions of Interest (ROIs) for the tissue segmentation task. Finally, for the cell detection task, we randomly choose a smaller ROI that is fully contained within the larger tissue ROI. As a result, OCELOT includes 673 paired patches from 6 organs. The numbers of WSIs and pairs of patches per organ are detailed in Tab. 2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.555, + 0.892, + 0.72 + ], + "angle": 0, + "content": "Some natural image datasets, such as ImageNet [12] or Pascal VOC [14], include thousands of annotated images. However, annotating histopathology images is more challenging and expensive due to the scarcity of expert pathologists [27]. Furthermore, acquiring dense annotations for cell detection and tissue segmentation is especially time-demanding compared to higher-level tasks such as image classification. Nonetheless, in Tab. 1, we observe that OCELOT is roughly double the size of the recent TIGER dataset with respect to the annotated tissue area and the number of annotated cells." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Patch configuration. Cell detection tasks benefit from fine-grained spatial information to better capture detailed cell properties (e.g. border, shape, color, and opacity). In contrast, tissue segmentation requires a larger context to enable a better understanding of the overall structural information. Therefore, we define the FoV sizes of \\( x_{s} \\) (cell detection) and \\( x_{l} \\) (tissue segmentation) as \\( 1024 \\times 1024 \\) and \\( 4096 \\times 4096 \\) pixels, respectively, at a resolution of 0.2 Micronsper-Pixel (MPP). Finally, the large FoV patches and tissue annotations \\( (x_{l}, y_{l}^{t}) \\) are down-sampled by a factor of 4, resulting in a size of \\( 1024 \\times 1024 \\) pixels." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23904" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.127, + 0.089, + 0.42, + 0.242 + ], + "angle": 0, + "content": "
Organs# Slides# Patch Pairs
TrainValTestTrainValTest
Kidney4815181254141
Head-neck135627910
Prostate261210501716
Stomach1565361212
Endometrium381313862925
Bladder351414822926
Total1756566406137130
" + }, + { + "type": "table_caption", + "bbox": [ + 0.124, + 0.246, + 0.423, + 0.26 + ], + "angle": 0, + "content": "Table 2. Dataset size per organ and data subset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.274, + 0.473, + 0.531 + ], + "angle": 0, + "content": "Annotation. All cell-tissue pairs of patches are annotated by board-certified pathologists. Cells are labeled as points, with associated 2D coordinates and class labels. We denote the annotations in a given cell-level patch, \\( x_{s} \\), as \\( y_{s}^{c} \\), and consider two classes: Tumor Cell (TC) and Background Cell \\( (BC)^{1} \\). TC and BC class ratios are \\( 35.01\\% \\) and \\( 64.99\\% \\), respectively. Regarding the tissue patches, \\( x_{l} \\), pathologists annotate the pixel-wise segmentation maps \\( y_{l}^{t} \\) with either Cancer Area (CA) or Background (BG) labels. A minority of pixels where the tissue class was uncertain were labeled as Unknown (UNK). BG, CA, and UNK class ratios are \\( 55.77\\% \\), \\( 40.17\\% \\), and \\( 4.06\\% \\), respectively. The amount of annotated cells and tissue pixels, per data split, can be found in the supplementary material. The detection of TC and BC has clinical relevance. For example, tumor purity [5], computed as the tumor/non-tumor cell ratio in a WSI, has a correlation with cancer prognosis [16, 31, 48]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.537, + 0.47, + 0.627 + ], + "angle": 0, + "content": "Dataset splits. The dataset is divided into three subsets: training, validation, and test, following a 6:2:2 ratio. To prevent information leaking among the data subsets, we randomly split the dataset per WSI, so that different patches from the same WSI are not included in multiple subsets. We maintain consistent cancer-type ratios in each subset." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.641, + 0.262, + 0.659 + ], + "angle": 0, + "content": "4. Empirical Analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.667, + 0.469, + 0.756 + ], + "angle": 0, + "content": "The motivation for considering the cell-tissue relationships for the development of cell detection models stem from the biological and hierarchical arrangement of cells and tissues. These insights are further corroborated by two main empirical observations described in the following Sec. 4.1 and Sec. 4.2." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.768, + 0.447, + 0.784 + ], + "angle": 0, + "content": "4.1. Interrelation between cell and tissue classes" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.791, + 0.47, + 0.867 + ], + "angle": 0, + "content": "We empirically observe the interrelation between specific cell and tissue classes by counting the amount of each annotated cell type within each tissue region, as observed in Tab. 3. Indeed, we verify that in OCELOT, around \\(93\\%\\) of \\(TC\\) are located within \\(CA\\) and \\(85\\%\\) of \\(BC\\) are found outside" + }, + { + "type": "image", + "bbox": [ + 0.541, + 0.09, + 0.681, + 0.169 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.569, + 0.171, + 0.652, + 0.184 + ], + "angle": 0, + "content": "(a) OCELOT" + }, + { + "type": "image", + "bbox": [ + 0.711, + 0.091, + 0.85, + 0.169 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.747, + 0.171, + 0.815, + 0.184 + ], + "angle": 0, + "content": "(b) TIGER" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.197, + 0.894, + 0.253 + ], + "angle": 0, + "content": "Table 3. Cell counts based on the tissue class. Each value stands for the number of cells located inside the tissue area. \\(TC\\), \\(BC\\), \\(LC\\), \\(CA\\), and \\(ST\\) stand for Tumor Cell, Background Cell, Lymphocyte Cell, Cancer Area tissue, and Stroma tissue, respectively." + }, + { + "type": "image", + "bbox": [ + 0.546, + 0.268, + 0.677, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.566, + 0.317, + 0.657, + 0.332 + ], + "angle": 0, + "content": "(a) OCELOT" + }, + { + "type": "image", + "bbox": [ + 0.713, + 0.269, + 0.844, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.743, + 0.317, + 0.816, + 0.332 + ], + "angle": 0, + "content": "(b) TIGER" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.344, + 0.893, + 0.373 + ], + "angle": 0, + "content": "Table 4. Pixel ratio among tissue classes. \\(CA\\), and \\(ST\\) stand for Cancer Area tissue, and Stroma tissue, respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.385, + 0.892, + 0.46 + ], + "angle": 0, + "content": "of the CA tissue. Note that CA is not the majority tissue class (Tab. 4), therefore, we conclude that there is, in fact, a relationship between the cell and tissue classes. We observe a similar phenomenon in the TIGER dataset when considering the \\( LC \\) and \\( ST \\) classes (Tab. 3 and Tab. 4)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.461, + 0.893, + 0.627 + ], + "angle": 0, + "content": "In practice, pathologists classify cells by taking into account such interrelationships, since it is difficult to classify isolated cells, without considering the larger context of the tissues. As depicted in Fig. 1, pathologists first need to visualize the structure of the tissues at a larger FoV. Then, they zoom in and consider the previously observed context along with the fine-grained details of each individual cell and nearby neighborhood, thus considering the cell-tissue dependencies. Inspired by the behavior of pathologists, we expect a cell detection model to also benefit from understanding the tissue structure from a broader viewpoint." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.636, + 0.75, + 0.653 + ], + "angle": 0, + "content": "4.2. Tissue-label Leaking Model" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.892, + 0.885 + ], + "angle": 0, + "content": "In the previous section, we observe a strong relationship between certain cell and tissue classes. To further validate the hypothesis that a cell detection model can leverage information from the tissue structure, we design an exploratory experiment where the tissue annotation is provided as an extra input to a cell detection model. To this end, we first crop the corresponding cell patch region from the tissue annotation \\( y_{l}^{t} \\), and upsample it to match the size of the cell patch, \\( x_{s} \\); the cropped tissue annotation is denoted as \\( y_{s}^{t} \\). Finally, we concatenate \\( x_{s} \\) and \\( y_{s}^{t} \\) at the channels dimension and use this data to train a cell detection model. We denote this model as a Tissue-label leaking model and illustrate it in Fig. 3. Note that this model is not appropriate for real-world scenarios as the tissue labels are unknown at inference time, and is explored for the purpose of empirical analysis." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.886, + 0.892, + 0.901 + ], + "angle": 0, + "content": "When we compare the performance of the tissue-label" + }, + { + "type": "page_footnote", + "bbox": [ + 0.077, + 0.875, + 0.469, + 0.901 + ], + "angle": 0, + "content": "\\( {}^{1}{BC} \\) includes any of the following cell categories: lymphocyte, macrophage, fibroblast, endothelial, or other remaining cell types." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23905" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.09, + 0.441, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.323, + 0.471, + 0.392 + ], + "angle": 0, + "content": "Figure 3. Tissue label leaking model. This model receives the cell patch \\( x_{s} \\) along with the corresponding tissue labels as input. The region corresponding to the cell patch in the tissue patch annotation is cropped, upsampled, and concatenated to the cell patch. \\( \\oplus \\) denotes channel-wise concatenation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.409, + 0.47, + 0.531 + ], + "angle": 0, + "content": "leaking model with the standard cell detection model on the OCELOT dataset, we observe a significant improvement in terms of mean F1-score\\(^2\\) performance of \\(+7.69\\) and \\(+9.76\\) in the validation and test sets, respectively. Detailed results can be found in Tab. 5. Taking these results into consideration, we conclude that there is significant room to improve the cell detection model, which can be achieved by combining the tasks of cell detection and tissue segmentation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.545, + 0.169, + 0.559 + ], + "angle": 0, + "content": "5. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.57, + 0.47, + 0.706 + ], + "angle": 0, + "content": "In this section, we propose to utilize cell-tissue relationships through multi-task learning. First, we propose a set of approaches inspired by the tissue-label leaking model described in Sec. 4.2. These models replace the annotated tissue labels with predictions from an auxiliary tissue segmentation branch. Second, we design a bi-directional information-sharing approach that shares features in both tissue-to-cell and cell-to-tissue directions. The proposed approaches are described in Sec. 5.2 and Sec. 5.3" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.716, + 0.208, + 0.731 + ], + "angle": 0, + "content": "5.1. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.739, + 0.469, + 0.83 + ], + "angle": 0, + "content": "To deal with cell-tissue sample pairs, i.e., \\((x_{s},y_{s}^{c})\\) and \\((x_{l},y_{l}^{t})\\), we build a dual-branch architecture containing separate networks for the cell and tissue tasks. Similarly to [41], we define cell detection as a segmentation task. Specifically, the cell labels are provided as a segmentation map by drawing a fixed-radius circle centered on each cell point an" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.091, + 0.891, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.288, + 0.893, + 0.33 + ], + "angle": 0, + "content": "Figure 4. Tissue-prediction injection model injects the tissue segmentation prediction into 1 out of 4 locations of the cell branch: (a) input, (b) after encoder, (c) after ASPP, and (d) after decoder." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.347, + 0.893, + 0.513 + ], + "angle": 0, + "content": "notation and filled with the corresponding class label. At inference time, we find local peaks within the predicted cell probability maps and output them as point predictions. More details are provided in the supplementary material. Treating cell detection as a segmentation task enables us to use the same architecture for both cell and tissue branches, which largely simplifies the training and tuning of the model and reduces the range of possible decisions, such as neural network architecture, or hyper-parameters. We use DeepLabV3+ [10] as a base architecture for both branches and single-task models." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.524, + 0.804, + 0.541 + ], + "angle": 0, + "content": "5.2. Tissue-prediction Injection Models" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.549, + 0.892, + 0.791 + ], + "angle": 0, + "content": "These models are a simple and practical extension of the tissue-label leaking model, where we inject the predicted tissue probabilities into the cell detection branch instead of leaking the tissue labels. We consider only one injection point in the cell detection branch, but explore four possible alternatives: (a) at the input (Pred-to-input), (b) after the encoder (Pred-to-inter-1), (c) after the ASPP module (Pred-to-inter-2), and (d) after the decoder (Pred-to-output). Fig. 4 depicts this family of models, denominated as Tissue-prediction injection. Since the tissue and cell patches represent different regions, we need to align the content between the tissue and cell feature maps before concatenation. Therefore, we crop the cell corresponding region from the tissue predictions, upsample them, and, finally, concatenate them in the channel dimensions of the feature maps of the cell branch, as illustrated in Fig. 5a." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.801, + 0.805, + 0.817 + ], + "angle": 0, + "content": "5.3. Cell-Tissue Feature Sharing Model" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Tissue-prediction injection models share the tissue prediction in a single location and direction, i.e., tissue-to-cell. To enable a more diverse and flexible cell-tissue information flow, we also explore bi-directional feature map sharing from cell-to-tissue (Fig. 5, left) and tissue-to-cell (Fig. 5," + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.839, + 0.47, + 0.901 + ], + "angle": 0, + "content": "\\(^{2}\\)True positive (TP), false positive (FP), and false negative (FN) counts are determined following [41]. If a detected cell is within a valid distance \\((\\approx 3\\mu m)\\) from an annotated cell and the cell class matches, it is counted as a TP, otherwise an FP. If an annotated cell is not detected, it is counted as an FN. Then, the mean F1 score across classes is computed." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "23906" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.089, + 0.262, + 0.169 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.131, + 0.171, + 0.223, + 0.183 + ], + "angle": 0, + "content": "(a) Tissue to Cell" + }, + { + "type": "image", + "bbox": [ + 0.278, + 0.089, + 0.462, + 0.169 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.324, + 0.171, + 0.418, + 0.184 + ], + "angle": 0, + "content": "(b) Cell to Tissue" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.196, + 0.473, + 0.321 + ], + "angle": 0, + "content": "Figure 5. Information is shared between cell and tissue branches via channel-wise concatenation preceded by a shallow convolutional layer with \\(3 \\times 3\\) kernel size. Cropping and upsampling (in Fig. 5a) or downsampling and zero-padding (in Fig. 5b) is applied to match the patch sizes and pixel-alignment between two feature maps from different FoVs. The cell and tissue feature maps are represented in orange and blue, respectively. The red contour denotes the cell patch-associated region in the tissue patch, and the gray regions represent zero padding." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.344, + 0.468, + 0.511 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.529, + 0.47, + 0.6 + ], + "angle": 0, + "content": "Figure 6. Cell-Tissue Feature Sharing Model has two branches for tissue segmentation and cell detection. Information exchange occurs multiple times between the two branches, indicated by the red vertical arrows. Details regarding the information exchange procedures are described in Fig. 5." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.615, + 0.47, + 0.766 + ], + "angle": 0, + "content": "right). Considering these two operations, we conduct an architecture search procedure to find the optimal feature map sharing configuration between both branches. To limit the search space, we constrain it to only 3 positions in the architecture: after the encoder, after the ASPP module, or after the decoder. Furthermore, we also exclusively allow the branches to inject feature maps at the same depth or position. Finally, we consider only the best-performing model among the \\(4^{3}\\) candidates, which is presented in Fig. 6. We name this model as cell-tissue feature sharing." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.784, + 0.312, + 0.801 + ], + "angle": 0, + "content": "6. Experiments and Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.471, + 0.903 + ], + "angle": 0, + "content": "We validate the hypothesis that incorporating cell-tissue relationships within a cell detection model is beneficial by evaluating the proposed models on OCELOT and 2 other datasets. First, we describe the additional datasets and implementation details. Then, we show how the proposed multi-task learning methods can improve the performance" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "of a cell detection task. Lastly, we conduct an ablation study to investigate how important using large FoV patches \\(\\{x_{l}\\}\\) and corresponding tissue segmentation labels \\(\\{y_{l}^{t}\\}\\) are in enhancing cell-tissue relationships." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.162, + 0.653, + 0.176 + ], + "angle": 0, + "content": "6.1. Other Datasets" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.185, + 0.892, + 0.472 + ], + "angle": 0, + "content": "TIGER. As mentioned in Sec. 2.1, TIGER is a public dataset, based on H&E stained images, which includes both cell and tissue annotations in overlapping patches. However, the patch sizes and amount of overlap between cell and tissue annotations highly vary among samples. Therefore, a preprocessing step is necessary to generate paired cell-tissue patches with consistent FoVs. After pre-processing, we obtain 9,888 paired patches of size \\(512 \\times 512\\) and \\(128 \\times 128\\) pixels for the tissue segmentation and cell detection tasks, respectively, with an MPP of approximately \\(0.5\\mu m / px\\). Most samples in TIGER include small regions with annotated cells, making it necessary to extract smaller patches. This explains why the absolute number of patches is larger than in OCELOT or CARP, although the actual amount of annotated cells and tissue area are smaller than the latter datasets, as shown in Tab. 1. In TIGER, one cell class (lymphocyte cells) and seven tissue classes are annotated. More details about the pre-processing and this dataset are provided in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.476, + 0.893, + 0.688 + ], + "angle": 0, + "content": "CARP. This is an internal lung cancer dataset of PD-L1 IHC-stained WSIs, containing 6,480 paired patches extracted from 1,012 WSIs. The patch sizes, resolution, annotation protocol, and general configuration are similar to OCELOT, as described in Sec. 3.1. Two cell classes are annotated: PD-L1 positive tumor cells \\((TC+)\\), and PD-L1 negative tumor cells \\((TC-\\) ). The tissue classes are the same as in OCELOT (BG, CA, and UNK). CARP is a real-world and large-scale dataset, with 809.1K annotated cells and 4.1080 mm\\(^2\\) tissue area, which is approximately 10 and 20 times more than OCELOT and TIGER, respectively. Moreover, the stain type is different from OCELOT and TIGER, allowing us to validate the effectiveness of the cell-tissue relationships across different staining methods." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.697, + 0.715, + 0.713 + ], + "angle": 0, + "content": "6.2. Implementation details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.901 + ], + "angle": 0, + "content": "As previously mentioned in Sec. 5.1, both cell and tissue branches are based on DeepLabV3+ [10], with a ResNet-34 [18] encoder. The models are trained for 300, 150, and 100 epochs in OCELOT, TIGER, and CARP datasets, respectively. We use the Adam optimizer [24], and, for each experiment, we tune the learning rate within a \\([5 \\times 10^{-5}, 2 \\times 10^{-3}]\\) range. The model at the epoch with the best validation set performance is chosen and used for evaluation on the test set. All experiments are repeated 5 times, and we report the mean and \\(95\\%\\) confidence interval of the performance metrics. For more implementation details, please refer to the supplementary material." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23907" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.184, + 0.089, + 0.784, + 0.256 + ], + "angle": 0, + "content": "
MethodOCELOTTIGERCARP
ValTestValTestValTest
Cell-only68.87±1.7664.44±1.8263.89±1.3953.82±1.2378.48±0.6970.96±1.47
Pred-to-input73.36±0.5969.65±3.9366.00±2.0053.29±1.3079.46±0.7972.98±0.82
Pred-to-inter-172.74±0.5070.54±2.2066.19±1.0255.87±1.7879.74±0.8073.05±0.69
Pred-to-inter-272.68±1.5871.23±0.9665.43±1.1454.75±2.2579.87±0.7873.14±1.53
Pred-to-output66.85±5.6265.05±3.7263.02±0.1653.32±0.4278.92±0.6072.61±0.95
Feature-sharing72.30±0.7368.91±2.5265.64±1.0755.10±2.1879.38±0.7473.00±0.33
Tissue-label leaking76.56±0.8074.20±0.9169.71±0.6161.66±1.1680.13±1.0472.97±0.49
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.265, + 0.894, + 0.297 + ], + "angle": 0, + "content": "Table 5. Cell detection mean F1 scores per model. With the exception of the Tissue-label Leaking model, the highest score is written in bold and the second highest score is underlined." + }, + { + "type": "table", + "bbox": [ + 0.134, + 0.309, + 0.839, + 0.471 + ], + "angle": 0, + "content": "
Methodpopulation (130)kidney (41)endometrium (25)bladder (26)prostate (16)stomach (12)head-neck (10)
Cell-only64.44±1.8264.12±3.4166.88±3.7159.42±5.6165.46±1.6466.19±8.9359.24±5.04
Pred-to-input69.65±3.9363.37±3.2773.31±6.3563.61±4.5568.36±8.6169.34±7.1575.28±0.82
Pred-to-inter-170.54±2.2066.62±7.3073.05±4.1464.35±3.5970.20±2.5671.55±5.2174.50±1.52
Pred-to-inter-271.23±0.9668.94±5.6575.15±2.7064.94±2.3468.35±8.8370.29±0.8173.74±1.72
Pred-to-output65.05±3.7263.38±4.4868.21±4.1159.81±3.9664.18±8.4167.51±6.3960.88±6.77
Feature-sharing68.91±2.5264.97±5.3371.28±5.3063.21±5.3166.25±9.5269.79±2.2573.88±4.39
Tissue-label leaking74.20±0.9175.75±1.9772.71±1.1271.18±0.9374.56±1.2575.24±1.7175.78±0.46
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.48, + 0.893, + 0.524 + ], + "angle": 0, + "content": "Table 6. Per-organ cell detection mean F1 scores in the OCELOT test set. Population mean F1 scores are computed from all the patches in the test set. In parentheses, we indicate the number of samples of each organ subset. With the exception of the Tissue-label Leaking model, the highest score is written in bold and the second highest score is underlined." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.536, + 0.216, + 0.551 + ], + "angle": 0, + "content": "6.3. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.562, + 0.47, + 0.729 + ], + "angle": 0, + "content": "The cell detection results obtained by the Cell-only baseline, the proposed approaches, and the Tissue-label leaking model are shown in Tab. 5. The Cell-only baseline is a simple cell detection model that only receives the small FoV patches as input. In other words, it only considers the cell branch in Fig. 4 and neither tissue annotations nor large FoV patches are leveraged. The Tissue-label leaking model, described in Sec. 5.2, receives the tissue annotations as input and serves as an exploratory experiment to obtain insight regarding how much the cell detection task can benefit from leveraging tissue annotation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.731, + 0.471, + 0.884 + ], + "angle": 0, + "content": "From Tab. 5, we observe that all the cell-tissue multi-task learning-based approaches, except for the Pred-to-output model, outperform the Cell-only baseline across all datasets. These improvements imply that cell detection on small FoV patches benefits from learning cell-tissue interrelationships from both large FoVs and tissue annotations. We hypothesize that the reason for the low performance of the Pred-to-output model is that the injection of tissue prediction to the cell branch happens too late. Therefore, there is a lack of network capacity for fusing cell and tissue information." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Furthermore, in Tab. 6, we report mean F1 scores per or" + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.534, + 0.891, + 0.627 + ], + "angle": 0, + "content": "
Further utilizePred-to-inter-2Feature-sharing
Tissue labelLarge FoVValTestValTest
--68.87±1.7664.45±1.8268.87±1.7664.45±1.82
✓-70.21±1.4964.99±1.7770.36±1.3466.58±2.02
-✓68.78±0.9465.56±3.6669.61±1.5966.22±1.53
✓✓72.68±1.5871.23±0.9672.30±0.7368.91±2.52
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.637, + 0.893, + 0.68 + ], + "angle": 0, + "content": "Table 7. Dataset ablation study. In all experiments, the model architecture is fixed and only input/output changes. The first row corresponds to the Cell-only baseline." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.686, + 0.892, + 0.793 + ], + "angle": 0, + "content": "gan in the OCELOT test set. Except for the Pred-to-output, all approaches show improvement compared to Cell-only baseline in most organs. Especially, our best-performing model, Pred-to-inter-2, demonstrates superior performance over Cell-only in all 6 organs with a significant margin. Such a result shows that considering cell-tissue relationships can generally help the cell detection task across various organs." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.802, + 0.717, + 0.818 + ], + "angle": 0, + "content": "6.4. Dataset Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Two data components in OCELOT contribute to better leveraging cell-tissue relationships for the task of cell detection: (1) the large FoV patches \\( x_{l} \\), and (2) the corresponding tissue annotations \\( y_{l}^{t} \\) that allows us to have a tissue segmentation objective and a multi-task learning setting. To verify" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23908" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.129, + 0.087, + 0.845, + 0.36 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.367, + 0.895, + 0.397 + ], + "angle": 0, + "content": "Figure 7. Qualitative comparison between the Pred-to-inter-2 and Cell-only models. Pred-to-inter-2 shows enhanced detection results aided by understanding the tissue environment from a broader context (blue: tumor cells, yellow: background cells, green: cancer area)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.408, + 0.471, + 0.544 + ], + "angle": 0, + "content": "the effectiveness of each data component, we conduct an ablation study where each component is evaluated separately while keeping the model architecture fixed. First, if only the tissue segmentation label is provided without a large FoV patch, the tissue branch takes \\( x_{s} \\) as an input and it is trained with both the tissue and the cell detection objectives. Second, if the large FoV patch is given without the corresponding tissue label, the tissue branch takes \\( x_{l} \\) as an input and the model is only trained with the cell detection objective." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.545, + 0.471, + 0.82 + ], + "angle": 0, + "content": "The results of the ablation study are shown in Tab. 7. We observe that each of the individual components on their own lead to improved performance. Hence, we can conclude that a large FoV patch and tissue segmentation label can contribute to better cell detection by providing a broader context and encoding the cell-tissue relationship, respectively. Moreover, when utilized simultaneously, we observe that both components work synergically, leading to an even better performance improvement. We also found that the tissue branch generates better tissue predictions when trained with a large FoV. Thus, providing accurate tissue information to the cell branch is important for boosting performance. Please, refer to the supplementary material for a detailed tissue performance comparison. Finally, this study shows evidence that the patch configuration defined in OCELOT is suitable for cell-tissue co-training, with a smaller FoV for cell detection \\((x_{s},y_{s}^{c})\\) and a larger FoV for tissue segmentation \\((x_{l},y_{l}^{t})\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.831, + 0.27, + 0.848 + ], + "angle": 0, + "content": "6.5. Qualitative Analysis" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.471, + 0.903 + ], + "angle": 0, + "content": "In this section, we visualize the cell predictions of the Cell-only and Pred-to-inter-2 models. We select the Pred-to-inter-2 model, as it is the overall best-performing model in" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.408, + 0.895, + 0.605 + ], + "angle": 0, + "content": "our experiments. In general, when compared to other background cells, tumor cells have the following characteristics: large size and irregular shape. However, cancer is heterogeneous and this is not always the case. Indeed, in Fig. 7, most of the cells are small and have a regular round shape. Based on these appearances, and without a larger context, those cells can be easily misclassified as background cells, which is the case of the Cell-only model. On the other hand, Pred-to-inter-2 shows a more accurate prediction by correctly understanding the cancer area in large FoV regions. This implies that Pred-to-inter-2 indeed considers both the morphology of cells and the tissue context, while Cell-only relies on the cells' morphology alone." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.63, + 0.62, + 0.646 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.895, + 0.903 + ], + "angle": 0, + "content": "In histopathology, there is a hierarchical organization and interrelationship between cells and tissues. Hence, we hypothesize that it can be leveraged to improve the performance of cell detection tasks. Throughout this paper, we have shown evidence that such cell-tissue relationships exist across multiple setups. Indeed, we observe improvements when utilizing large FoVs with corresponding tissue annotations together with cell annotations in a multi-task learning approach, using simple neural architecture designs. Although the cell detection task clearly benefits from the multitask learning framework, the improvement on the tissue segmentation task was not yet investigated, and we leave this topic as a future direction. We hope this work, especially the OCELOT dataset, initiates new and promising research directions in CPATH to further study the cell-tissue relationships and improve cell detection and tissue segmentation." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23909" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.17 + ], + "angle": 0, + "content": "[1] MICCAI 2014 Grand Challenges. https://wiki.cancerimagingarchive.net/display/Public/MICCAI+2014+Grand+Challenges, 2014. [Online; accessed Nov-2022]. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.172, + 0.468, + 0.199 + ], + "angle": 0, + "content": "[2] TIGER: Grand Challenge. https://tiger.grand-challenge.org, 2022. [Online; accessed Nov-2022]. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.201, + 0.471, + 0.297 + ], + "angle": 0, + "content": "[3] Esther Abels, Liron Pantanowitz, Famke Aeffner, Mark D Zarella, Jeroen van der Laak, Marilyn M Bui, Venkata NP Vemuri, Anil V Parwani, Jeff Gibbs, Emmanuel Agosto-Arroyo, et al. Computational pathology definitions, best practices, and recommendations for regulatory guidance: a white paper from the digital pathology association. The Journal of Pathology, 249(3):286-294, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.298, + 0.47, + 0.394 + ], + "angle": 0, + "content": "[4] Mohamed Amgad, Lamees A Atteya, Hagar Hussein, Kareem Hosny Mohammed, Ehab Hafiz, Maha AT Elsebaie, Ahmed M Alhusseiny, Mohamed Atef AlMoslemany, Abdelmagid M Elmatboly, Philip A Pappalardo, et al. Nuclei: A scalable crowdsourcing approach and dataset for nucleus classification and segmentation in breast cancer. *Giga* Science, 11, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.396, + 0.471, + 0.491 + ], + "angle": 0, + "content": "[5] Vahid Azimi, Young Hwan Chang, Guillaume Thibault, Jaclyn Smith, Takahiro Tsujikawa, Benjamin Kukull, Braden Jensen, Christopher Corless, Adam Margolin, and Joe W Gray. Breast cancer histopathology image analysis pipeline for tumor purity estimation. In 2017 IEEE 14th International Symposium on Biomedical Imaging (ISBI 2017), pages 1137-1140. IEEE, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.494, + 0.468, + 0.535 + ], + "angle": 0, + "content": "[6] Tian Bai, Jiayu Xu, and Fuyong Xing. Multi-field of view aggregation and context encoding for single-stage nucleus recognition. In MICCAI, pages 382-392. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.537, + 0.468, + 0.619 + ], + "angle": 0, + "content": "[7] Wouter Bulten, Kimmo Kartasalo, Po-Hsuan Cameron Chen, Peter Ström, Hans Pinckaers, Kunal Nagpal, Yuannan Cai, David F Steiner, Hester van Boven, Robert Vink, et al. Artificial intelligence for diagnosis and gleason grading of prostate cancer: the panda challenge. Nature Medicine, 28(1):154-163, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.621, + 0.468, + 0.675 + ], + "angle": 0, + "content": "[8] Lyndon Chan, Mahdi S Hosseini, Corwyn Rowsell, Konstantinos N Plataniotis, and Savvas Damaskinos. Histosegnet: Semantic segmentation of histological tissue type in whole slide images. In ICCV, pages 10662-10671, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.677, + 0.468, + 0.718 + ], + "angle": 0, + "content": "[9] Hao Chen, Xiaojuan Qi, Lequan Yu, and Pheng-Ann Heng. Dcan: deep contour-aware networks for accurate gland segmentation. In CVPR, pages 2487–2496, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.72, + 0.468, + 0.774 + ], + "angle": 0, + "content": "[10] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In ECCV, pages 801-818, 2018. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.468, + 0.857 + ], + "angle": 0, + "content": "[11] Qian Da, Xiaodi Huang, Zhongyu Li, Yanfei Zuo, Chenbin Zhang, Jingxin Liu, Wen Chen, Jiahui Li, Dou Xu, Zhiqiang Hu, et al. Digestpath: a benchmark dataset with challenge review for the pathological detection and segmentation of digestive-system. Medical Image Analysis, page 102485, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[12] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248–255. IEEE, 2009. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.188 + ], + "angle": 0, + "content": "[13] James A Diao, Jason K Wang, Wan Fung Chui, Victoria Mountain, Sai Chowdary Gullapally, Ramprakash Srinivasan, Richard N Mitchell, Benjamin Glass, Sara Hoffman, Sudha K Rao, et al. Human-interpretable image features derived from densely mapped cancer pathology slides predict diverse molecular phenotypes. Nature communications, 12(1):1-15, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.191, + 0.892, + 0.233 + ], + "angle": 0, + "content": "[14] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. *IJCV*, 88(2):303–338, 2010. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.235, + 0.892, + 0.302 + ], + "angle": 0, + "content": "[15] Jevgenij Gamper, Navid Alemi Koohbanani, Ksenija Benes, Simon Graham, Mostafa Jahanifar, Syed Ali Khurram, Ayesha Azam, Katherine Hewitt, and Nasir Rajpoot. Pan-nuke dataset extension, insights and baselines. arXiv preprint arXiv:2003.10778, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.304, + 0.892, + 0.346 + ], + "angle": 0, + "content": "[16] Zhe Gong, Jieyun Zhang, and Weijian Guo. Tumor purity as a prognosis and immunotherapy relevant feature in gastric cancer. Cancer Medicine, 9(23):9052-9063, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.349, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[17] Simon Graham, Quoc Dang Vu, Shan E Ahmed Raza, Ayesha Azam, Yee Wah Tsang, Jin Tae Kwak, and Nasir Rajpoot. Hover-net: Simultaneous segmentation and classification of nuclei in multi-tissue histology images. Medical Image Analysis, 58:101563, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.42, + 0.892, + 0.459 + ], + "angle": 0, + "content": "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. pages 770-778, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.462, + 0.892, + 0.544 + ], + "angle": 0, + "content": "[19] David Joon Ho, Dig VK Yarlagadda, Timothy M D'Alfonso, Matthew G Hanna, Anne Grabenstetter, Peter Ntiamoah, Edi Brogi, Lee K Tan, and Thomas J Fuchs. Deep multimagnification networks for multi-class breast cancer image segmentation. Computerized Medical Imaging and Graphics, 88:101866, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.547, + 0.892, + 0.628 + ], + "angle": 0, + "content": "[20] Mahdi S Hosseini, Lyndon Chan, Gabriel Tse, Michael Tang, Jun Deng, Sajad Norouzi, Corwyn Rowsell, Konstantinos N Plataniotis, and Savvas Damaskinos. Atlas of digital pathology: A generalized hierarchical histological tissue type-annotated database for deep learning. In CVPR, pages 11747–11756, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.632, + 0.892, + 0.672 + ], + "angle": 0, + "content": "[21] Carolyn Hutter and Jean Claude Zenklusen. The cancer genome atlas: creating lasting value beyond its data. Cell, 173(2):283-285, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.675, + 0.892, + 0.743 + ], + "angle": 0, + "content": "[22] Konstantinos Kamnitsas, Christian Ledig, Virginia FJ Newcombe, Joanna P Simpson, Andrew D Kane, David K Menon, Daniel Rueckert, and Ben Glocker. Efficient multi-scale 3d cnn with fully connected crf for accurate brain lesion segmentation. Medical Image Analysis, 36:61-78, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.746, + 0.892, + 0.828 + ], + "angle": 0, + "content": "[23] Davood Karimi, Guy Nir, Ladan Fazli, Peter C Black, Larry Goldenberg, and Septimiu E Salcudean. Deep learning-based gleason grading of prostate cancer from histopathology images—role of multiscale decision aggregation and data augmentation. IEEE journal of Biomedical and Health Informatics, 24(5):1413-1426, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.831, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[24] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.859, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[25] Neeraj Kumar, Ruchika Verma, Deepak Anand, Yanning Zhou, Omer Fahri Onder, Efstratios Tsougenis, Hao Chen, Pheng-Ann Heng, Jiahui Li, Zhiqiang Hu, et al. A multi" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "23910" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.093, + 0.469, + 0.12 + ], + "angle": 0, + "content": "organ nucleus segmentation challenge. IEEE transactions on Medical Imaging, 39(5):1380-1391, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.189 + ], + "angle": 0, + "content": "[26] Shyam Lal, Devikalyan Das, Kumar Alabhya, Anirudh Kanfade, Aman Kumar, and Jyoti Kini. Nucleisegnet: robust deep learning architecture for the nuclei segmentation of liver cancer histopathology images. Computers in Biology and Medicine, 128:104075, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.191, + 0.469, + 0.245 + ], + "angle": 0, + "content": "[27] Chunggi Lee, Seonwook Park, Heon Song, Jeongun Ryu, Sanghoon Kim, Haejoon Kim, Sérgio Pereira, and Donggeun Yoo. Interactive multi-class tiny-object detection. In CVPR, pages 14136-14145, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.247, + 0.469, + 0.329 + ], + "angle": 0, + "content": "[28] Wenqi Li, Siyamalan Manivannan, Shazia Akbar, Jianguo Zhang, Emanuele Trucco, and Stephen J McKenna. Gland segmentation in colon histology images using hand-crafted features and convolutional neural networks. In 2016 IEEE 13th international symposium on biomedical imaging (ISBI), pages 1405-1408. IEEE, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.331, + 0.469, + 0.371 + ], + "angle": 0, + "content": "[29] Xieli Li, Yuanyuan Wang, Qisheng Tang, Zhen Fan, and Jinhua Yu. Dual u-net for the segmentation of overlapping glioma nuclei. IEEE Access, 7:84040-84052, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.373, + 0.469, + 0.427 + ], + "angle": 0, + "content": "[30] Dongnan Liu, Donghao Zhang, Yang Song, Chaoyi Zhang, Fan Zhang, Lauren O'Donnell, and Weidong Cai. Nuclei segmentation via a deep panoptic model with semantic feature fusion. In *IJCAI*, pages 861-868, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.428, + 0.469, + 0.509 + ], + "angle": 0, + "content": "[31] Yihao Mao, Qingyang Feng, Peng Zheng, Liangliang Yang, Tianyu Liu, Yuqiu Xu, Dexiang Zhu, Wenju Chang, Meiling Ji, Li Ren, et al. Low tumor purity is associated with poor prognosis, heavy mutation burden, and intense immune phenotype in colon cancer. Cancer Management and Research, 10:3569, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.512, + 0.469, + 0.566 + ], + "angle": 0, + "content": "[32] Peter Naylor, Marick Laé, Fabien Reyal, and Thomas Walter. Segmentation of nuclei in histopathology images by deep regression of the distance map. IEEE transactions on Medical Imaging, 38(2):448-459, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.568, + 0.469, + 0.649 + ], + "angle": 0, + "content": "[33] Guy Nir, Soheil Hor, Davood Karimi, Ladan Fazli, Brian F Skinnider, Peyman Tavassoli, Dmitry Turbin, Carlos F Villamil, Gang Wang, R Storey Wilson, et al. Automatic grading of prostate cancer in digitized histopathology images: Learning from multiple experts. Medical Image Analysis, 50:167-180, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.651, + 0.469, + 0.747 + ], + "angle": 0, + "content": "[34] Sehnoon Park, Chan-Young Ock, Hyojin Kim, Sergio Pereira, Seonwook Park, Minuk Ma, Sangjoon Choi, Seokhwi Kim, Seunghwan Shin, Brian Jaehong Aum, et al. Artificial intelligence-powered spatial analysis of tumor-infiltrating lymphocytes as complementary biomarker for immune checkpoint inhibition in non-small-cell lung cancer. Journal of Clinical Oncology, 40(17):1916, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.749, + 0.469, + 0.872 + ], + "angle": 0, + "content": "[35] Pushpak Pati, Guillaume Jaume, Lauren Alisha Fernandes, Antonio Foncubierta-Rodriguez, Florinda Feroce, Anna Maria Anniciello, Giosue Scognamiglio, Nadia Brancati, Daniel Riccio, Maurizio Di Bonito, et al. Hact-net: A hierarchical cell-to-tissue graph neural network for histopathological image classification. In Uncertainty for Safe Utilization of Machine Learning in Medical Imaging, and Graphs in Biomedical Image Analysis, pages 208-219. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.874, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[36] Ziniu Qian, Kailu Li, Maode Lai, Eric I Chang, Bingzheng Wei, Yubo Fan, Yan Xu, et al. Transformer based multiple" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.133 + ], + "angle": 0, + "content": "instance learning for weakly supervised histopathology image segmentation. arXiv preprint arXiv:2205.08878, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.135, + 0.892, + 0.217 + ], + "angle": 0, + "content": "[37] Hui Qu, Pengxiang Wu, Qiaoying Huang, Jingru Yi, Zhennan Yan, Kang Li, Gregory M Riedlinger, Subhajyoti De, Shaoting Zhang, and Dimitris N Metaxas. Weakly supervised deep nuclei segmentation using partial points annotation in histopathology images. IEEE transactions on Medical Imaging, 39(11):3655-3666, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.219, + 0.892, + 0.314 + ], + "angle": 0, + "content": "[38] Roberto Salgado, Carsten Denkert, S Demaria, N Sirtaine, F Klauschen, Giancarlo Pruneri, S Wienert, Gert Van den Eynden, Frederick L Baehner, Frederique Pénault-Llorca, et al. The evaluation of tumor-infiltrating lymphocytes (tils) in breast cancer: recommendations by an international tils working group 2014. Annals of Oncology, 26(2):259-271, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.316, + 0.892, + 0.398 + ], + "angle": 0, + "content": "[39] Rüdiger Schmitz, Frederic Madesta, Maximilian Nielsen, Jenny Krause, Stefan Steurer, René Werner, and Thomas Rösch. Multi-scale fully convolutional neural networks for histopathology image segmentation: from nuclear aberrations to the global tissue architecture. Medical Image Analysis, 70:101996, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.4, + 0.892, + 0.481 + ], + "angle": 0, + "content": "[40] Peng Sun, Jiehua He, Xue Chao, Keming Chen, Yuanyuan Xu, Qitao Huang, Jingping Yun, Mei Li, Rongzhen Luo, Jinbo Kuang, et al. A computational tumor-infiltrating lymphocyte assessment method comparable with visual reporting guidelines for triple-negative breast cancer. EBioMedicine, 70:103492, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.484, + 0.892, + 0.565 + ], + "angle": 0, + "content": "[41] Zaneta Swiderska-Chadaj, Hans Pinckaers, Mart van Rijthoven, Maschenka Balkenhol, Margarita Melnikova, Oscar Geessink, Quirine Manson, Mark Sherman, Antonio Polonia, Jeremy Parry, et al. Learning to detect lymphocytes in immunohistochemistry with deep learning. Medical Image Analysis, 58:101547, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.567, + 0.892, + 0.621 + ], + "angle": 0, + "content": "[42] Hiroki Tokunaga, Yuki Teramoto, Akihiko Yoshizawa, and Ryoma Bise. Adaptive weighting multi-field-of-view cnn for semantic segmentation in pathology. In CVPR, pages 12597-12606, 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.623, + 0.892, + 0.691 + ], + "angle": 0, + "content": "[43] Mart Van Rijthoven, Maschenka Balkenhol, Karina Silina, Jeroen Van Der Laak, and Francesco Ciompi. Hooknet: Multi-resolution convolutional neural networks for semantic segmentation in histopathology whole-slide images. Medical Image Analysis, 68:101890, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.693, + 0.892, + 0.775 + ], + "angle": 0, + "content": "[44] Ruchika Verma, Neeraj Kumar, Abhijeet Patil, Nikhil Cherian Kurian, Swapnil Rane, Simon Graham, Quoc Dang Vu, Mieke Zwager, Shan E Ahmed Raza, Nasir Rajpoot, et al. Monusac2020: A multi-organ nuclei segmentation and classification challenge. IEEE Transactions on Medical Imaging, 40(12):3413-3423, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.777, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[45] Quoc Dang Vu, Simon Graham, Tahsin Kurc, Minh Nguyen Nhat To, Muhammad Shaban, Talha Qaiser, Navid Alemi Koohbanani, Syed Ali Khurram, Jayashree Kalpathy-Cramer, Tianhao Zhao, et al. Methods for segmentation and classification of digital microscopy tissue images. Frontiers in Bioengineering and Biotechnology, page 53, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[46] Jingwen Wang, Richard J Chen, Ming Y Lu, Alexander Baras, and Faisal Mahmood. Weakly supervised prostate" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23911" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.469, + 0.134 + ], + "angle": 0, + "content": "tma classification via graph convolutional networks. In 2020 IEEE 17th International Symposium on Biomedical Imaging (ISBI), pages 239-243. IEEE, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.205 + ], + "angle": 0, + "content": "[47] Yan Xu, Zhipeng Jia, Liang-Bo Wang, Yuqing Ai, Fang Zhang, Maode Lai, Eric I Chang, et al. Large scale tissue histopathology image classification, segmentation, and visualization via deep convolutional activation features. BMC bioinformatics, 18(1):1-17, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.471, + 0.274 + ], + "angle": 0, + "content": "[48] Chuanbao Zhang, Wen Cheng, Xiufang Ren, Zheng Wang, Xing Liu, Guanzhang Li, Sheng Han, Tao Jiang, and Anhua Wu. Tumor purity as an underlying key factor in gliomatumor purity in glioma. Clinical Cancer Research, 23(20):6279-6291, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.277, + 0.47, + 0.345 + ], + "angle": 0, + "content": "[49] Bingchao Zhao, Xin Chen, Zhi Li, Zhiwen Yu, Su Yao, Lixu Yan, Yuqian Wang, Zaiyi Liu, Changhong Liang, and Chu Han. Triple u-net: Hematoxylin-aware nuclei segmentation with progressive dense feature aggregation. Medical Image Analysis, 65:101786, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.347, + 0.47, + 0.428 + ], + "angle": 0, + "content": "[50] Ke Zhao, Zhenhui Li, Su Yao, Yingyi Wang, Xiaomei Wu, Zeyan Xu, Lin Wu, Yanqi Huang, Changhong Liang, and Zaiyi Liu. Artificial intelligence quantified tumour-stroma ratio is an independent predictor for overall survival in resectable colorectal cancer. EBioMedicine, 61:103054, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.431, + 0.469, + 0.499 + ], + "angle": 0, + "content": "[51] Yanning Zhou, Simon Graham, Navid Alemi Koohbanani, Muhammad Shaban, Pheng-Ann Heng, and Nasir Rajpoot. Cgc-net: Cell graph convolutional network for grading of colorectal cancer histology images. In ICCV Workshops, pages 0–0, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.501, + 0.469, + 0.556 + ], + "angle": 0, + "content": "[52] Chuang Zhu, Ke Mei, Ting Peng, Yihao Luo, Jun Liu, Ying Wang, and Mulan Jin. Multi-level colonoscopy malignant tissue detection with adversarial cac-unet. Neurocomputing, 438:165-183, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.556 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23912" + } + ] +] \ No newline at end of file diff --git a/2023/OCELOT_ Overlapped Cell on Tissue Dataset for Histopathology/052d2272-757d-4e0e-b0c0-b44903da46a8_origin.pdf b/2023/OCELOT_ Overlapped Cell on Tissue Dataset for Histopathology/052d2272-757d-4e0e-b0c0-b44903da46a8_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..cea6b9731f3642a7738701b9b90b08fe65b4d55a --- /dev/null +++ b/2023/OCELOT_ Overlapped Cell on Tissue Dataset for Histopathology/052d2272-757d-4e0e-b0c0-b44903da46a8_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e3ce56564d1cbd0b23173417434c5ff6b5811e0db60051b3d03cd8b24ed607a4 +size 1557262 diff --git a/2023/OCELOT_ Overlapped Cell on Tissue Dataset for Histopathology/full.md b/2023/OCELOT_ Overlapped Cell on Tissue Dataset for Histopathology/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c651403e3cf09bfd0dddd6a763f38c11c730806b --- /dev/null +++ b/2023/OCELOT_ Overlapped Cell on Tissue Dataset for Histopathology/full.md @@ -0,0 +1,282 @@ +# OCELOT: Overlapped Cell on Tissue Dataset for Histopathology + +Jeongun Ryu* Aaron Valero Puche* JaeWoong Shin* Seonwook Park Biagio Brattoli Jinhee Lee Wonkyung Jung Soo Ick Cho Kyunghyun Paeng Chan-Young Ock Donggeun Yoo Sérgio Pereira Lunit Inc. + +{rjw0205, aaron.valero, jwoong.shin, spark, biagio, jinhee.lee, wkjung, sooickcho, khpaeng, ock.chanyoung, dgyoo, sergio}@lunit.io + +# Abstract + +Cell detection is a fundamental task in computational pathology that can be used for extracting high-level medical information from whole-slide images. For accurate cell detection, pathologists often zoom out to understand the tissue-level structures and zoom in to classify cells based on their morphology and the surrounding context. However, there is a lack of efforts to reflect such behaviors by pathologists in the cell detection models, mainly due to the lack of datasets containing both cell and tissue annotations with overlapping regions. To overcome this limitation, we propose and publicly release OCELOT, a dataset purposely dedicated to the study of cell-tissue relationships for cell detection in histopathology. OCELOT provides overlapping cell and tissue annotations on images acquired from multiple organs. Within this setting, we also propose multi-task learning approaches that benefit from learning both cell and tissue tasks simultaneously. When compared against a model trained only for the cell detection task, our proposed approaches improve cell detection performance on 3 datasets: proposed OCELOT, public TIGER, and internal CARP datasets. On the OCELOT test set in particular, we show up to 6.79 improvement in F1-score. We believe the contributions of this paper, including the release of the OCELOT dataset at https://lunit-io.github.io/research/publications/ocelot are a crucial starting point toward the important research direction of incorporating cell-tissue relationships in computation pathology. + +# 1. Introduction + +Computational Pathology (CPATH) [3] is a branch of digital pathology that develops methodologies for the analysis of digitized patient specimens, such as Whole-Slide-Images (WSIs). Cell detection in histology images [26, 37, 49] is one of the most important tasks in CPATH. It allows the quantification and analysis of different cell types, which can lead to better prognosis evaluation [34, 40] and patient + +treatment planning while maintaining medical interpretability [13]. Since it has the potential to impact human lives, high-performance cell detection models are essential in real-world applications and need to be investigated. + +To better locate and classify cells, detailed morphological characteristics such as color and shape are crucial. Consequently, cell detection datasets are typically collected at high magnification but small Field-of-View (FoV). However, this can make the cell detection model overly rely on appearance details, without understanding the broader context [42]. This context can help cell detection by providing information about how cells are arranged and grouped together to form high-level tissue structures. In practice, expert annotators (pathologists) first zoom out to understand these broad tissue structures. Next, they zoom in to better classify individual cells while taking into account the context information, as depicted in Fig. 1. + +The behavior of pathologists can be transferred to deep learning, for instance, through a multi-task strategy combining cell detection tasks at high magnification and tissue segmentation at low magnification. This type of approach would allow the model to share knowledge across different tasks and FoVs. However, to train such an approach, a combined dataset with cell-tissue overlapping regions is required; unfortunately, most existing datasets only target a single task, either cell detection [17,25] or tissue segmentation [7, 11]. + +In this paper, we introduce a new research direction: studying cell-tissue relationships for cell detection. First, we publish the OCELOT dataset, which contains cell and tissue annotations in small and large FoV patches, respectively, with overlapping regions. Additionally, the data is collected from WSIs of multiple organs. This can provide the necessary data for researchers to study cell-tissue relationships and their effect on cell detection. Second, we introduce simple multi-task learning approaches for cell detection that can benefit from cell-tissue relationships and demonstrate their advantages over 3 different datasets. These approaches consistently show better cell detection performance compared to + +![](images/1a09049a2141ea141e91630bf8e9c6d20f719777a882510d8717a46ee7200df8.jpg) +Figure 1. Behavior of pathologists and cell detection models. Typical cell detection models infer their predictions only by looking at a limited FoV region. Failure cases can occur when cells are difficult to be classified solely by their morphology, i.e., while disregarding the larger architecture of the tissues (context). In the above example, some groups of tumor cells are misclassified as background cells due to their morphology; these tumor cells are smaller and rounder than the nearby ones. Note that the large size and irregular shape are representative characteristics of tumor cells. Pathologists overcome these limitations by further understanding the tissue context of the larger FoV region. Misclassified cells can be easily corrected by understanding that such regions are cancer areas. + +the cell-only baseline, i.e., a model trained only with small FoV patches with the corresponding cell annotations. We hope that our proposed OCELOT dataset and methods will encourage the CPATH community to learn how to reflect cell-tissue relationships better to improve cell detection. + +Our contributions are 3-fold and summarized as follows, + +- The first work that exploits cell-tissue relationships for better cell detection, to the best of our knowledge. +- We release OCELOT, a dataset with overlapping cell and tissue annotations based on Hematoxylin and Eosin (H&E) stained WSIs of multiple organs. +- We introduce several approaches that boost cell detection performance via multi-task learning, and empirically confirm that these methods generalize well across different datasets and histological stainings. + +# 2. Related Work + +# 2.1. Datasets for Cell detection and Tissue segmentation Tasks + +In recent years, numerous datasets have been released for tackling cell detection. Some of those works only target a single organ [17,32], while others consider multiple ones [4, 15,25,44,45]. The availability of these datasets enables the CPATH community to push forward the development and improvement of cell detection models [17,26,29,30,37,49]. In addition, tissue segmentation datasets have also been proposed for prostate [7,23,33], colorectal [50], brain [1], and multiple organs [20]. Some examples of tissue segmentation works can be found in [8,9,28,36,47,52]. The dataset + +
DatasetTissue Area# CellOrgans
OCELOT4.267cm2114.7Kmultiple
TIGER2.536cm250.8Kbreast
+ +Table 1. Dataset comparison in terms of physical tissue annotated area and total cell counting per dataset. + +in [11] is composed of a cell detection subset and a tissue segmentation subset. However, the subsets are annotated independently and from different patient groups, and, therefore, there are no overlaps between the cell and tissue data. Because of the lack of overlapping data in the aforementioned datasets, it is difficult to build an end-to-end framework to learn cell-tissue relationships by jointly training on the cell and tissue tasks. + +The recently released TIGER dataset [2] contains both cell and tissue annotations to study tumor-infiltrating lymphocytes [38] in H&E breast cancer WSIs. All the cell-annotated areas exist inside the tissue-annotated area, however, this work does not propose nor initiate any effort toward the integration of both cell and tissue tasks. + +To further promote the development of methods that leverage the cell-tissue relationship for the task of cell detection, we propose OCELOT, which is designed to capture the hierarchical relationship between cells and tissues, especially in the tumor environment. OCELOT contains roughly two times more cell and tissue annotations than TIGER (see Tab. 1). Additionally, the data was collected from multiple organs to enable the investigation of the generalizability of cell-tissue relationships over various cancer types. In the end, we utilized both OCELOT and TIGER to reveal the + +power of cell-tissue relationship for cell detection in Sec. 6. + +# 2.2. Leveraging Large Field of View + +Some studies [6, 19, 22, 39, 42, 43] extract a large FoV region as an additional input to improve detection/segmentation performance on smaller FoV regions. [22] proposes a dual pathway 3D CNN for brain lesion segmentation, where each pathway receives both small and large center-shared FoV regions as input. Similar studies are also conducted in the CPATH domain for tissue segmentation [19, 39, 42, 43] and cell detection [6]. To fuse different FoV patches, [42] introduces a weighting mechanism and [43] proposes a multi-scale merging block composed of convolution and concatenation. Nevertheless, no inter-task relationship is considered in the previously mentioned methods. In contrast, our models take advantage of the large contextual information while learning the cell-tissue relationship via multi-task objectives at different FoVs. + +# 2.3. Leveraging Cell-Tissue Relationships + +To our understanding, there is no study that considers cell-tissue relationships for cell detection or tissue segmentation tasks. On the other hand, a few efforts have attempted to link tissue and cell for image classification using graph-based methods [35,46,51]. Such studies represent the tissue structure as a graph of detected cells, based on the medical knowledge that cells form tissues. [35] explicitly considers the cell-tissue relationship for the task of breast cancer subtyping, by an interaction between tissue-level and cell-level graphs with a cell-to-tissue hierarchy. However, these methods treat the cell/tissue graph generation as a pre-processing step, by using the inference output of independently pretrained cell detection and tissue segmentation models. In contrast, we directly target the improvement of cell detection by utilizing the cell-tissue relationship. + +# 3. OCELOT + +In this section, we introduce OCELOT, a histopathology dataset specifically built to enable the development of methods that leverage cell and tissue relationships. Each sample of the OCELOT dataset $D$ is composed of six components, + +$$ +\mathcal {D} = \left\{\left(x _ {s}, y _ {s} ^ {c}, x _ {l}, y _ {l} ^ {t}, c _ {x}, c _ {y}\right) _ {i} \right\} _ {i = 1} ^ {N} \tag {1} +$$ + +where $x_{s}, x_{l}$ are the small and large FoV patches extracted from the WSI, $y_{s}^{c}, y_{l}^{t}$ refer to the corresponding cell and tissue annotations, respectively, and $c_{x}, c_{y}$ are the relative coordinates of the center of $x_{s}$ within $x_{l}$ . We drop the sample index $i$ for simplicity. Fig. 2 shows the visualization of a sample in OCELOT. More details about the dataset including data collection and statistics can be found in the following sub-sections. The dataset is publicly available at https://lunit-io.github.io/research/publications/ocelot/. + +![](images/044543b55eab3dcae11256d4a6b2215ccf6b5d0ff53f96f269c8f0129b460040.jpg) +$(x_{l},y_{l}^{t})$ +Figure 2. A sample from the OCELOT dataset. Each sample of the dataset consists of two input patches and the corresponding annotations. Left shows the large FoV patch $x_{l}$ with tissue segmentation annotation $y_{l}^{t}$ , where green denotes the cancer area. Right shows the small FoV patch $x_{s}$ with cell point annotation $y_{s}^{c}$ , where blue and yellow dots denote tumor and background cells, respectively. The red box indicates the size and location of the $x_{s}$ with respect to the $x_{l}$ . + +![](images/a3eec8298136c7b2992ab60c079c30314fdb1beb34c7ee0d51adf921b4119281.jpg) +$(x_{s},y_{s}^{c})$ + +# 3.1. Data Collection + +We collect 306 TCGA [21] WSIs from a total of 6 different organs: kidney, head-and-neck, prostate, stomach, endometrium, and bladder. From each of the WSIs, we select 1 to 3 large Regions of Interest (ROIs) for the tissue segmentation task. Finally, for the cell detection task, we randomly choose a smaller ROI that is fully contained within the larger tissue ROI. As a result, OCELOT includes 673 paired patches from 6 organs. The numbers of WSIs and pairs of patches per organ are detailed in Tab. 2. + +Some natural image datasets, such as ImageNet [12] or Pascal VOC [14], include thousands of annotated images. However, annotating histopathology images is more challenging and expensive due to the scarcity of expert pathologists [27]. Furthermore, acquiring dense annotations for cell detection and tissue segmentation is especially time-demanding compared to higher-level tasks such as image classification. Nonetheless, in Tab. 1, we observe that OCELOT is roughly double the size of the recent TIGER dataset with respect to the annotated tissue area and the number of annotated cells. + +Patch configuration. Cell detection tasks benefit from fine-grained spatial information to better capture detailed cell properties (e.g. border, shape, color, and opacity). In contrast, tissue segmentation requires a larger context to enable a better understanding of the overall structural information. Therefore, we define the FoV sizes of $x_{s}$ (cell detection) and $x_{l}$ (tissue segmentation) as $1024 \times 1024$ and $4096 \times 4096$ pixels, respectively, at a resolution of 0.2 Micronsper-Pixel (MPP). Finally, the large FoV patches and tissue annotations $(x_{l}, y_{l}^{t})$ are down-sampled by a factor of 4, resulting in a size of $1024 \times 1024$ pixels. + +
Organs# Slides# Patch Pairs
TrainValTestTrainValTest
Kidney4815181254141
Head-neck135627910
Prostate261210501716
Stomach1565361212
Endometrium381313862925
Bladder351414822926
Total1756566406137130
+ +Table 2. Dataset size per organ and data subset. + +Annotation. All cell-tissue pairs of patches are annotated by board-certified pathologists. Cells are labeled as points, with associated 2D coordinates and class labels. We denote the annotations in a given cell-level patch, $x_{s}$ , as $y_{s}^{c}$ , and consider two classes: Tumor Cell (TC) and Background Cell $(BC)^{1}$ . TC and BC class ratios are $35.01\%$ and $64.99\%$ , respectively. Regarding the tissue patches, $x_{l}$ , pathologists annotate the pixel-wise segmentation maps $y_{l}^{t}$ with either Cancer Area (CA) or Background (BG) labels. A minority of pixels where the tissue class was uncertain were labeled as Unknown (UNK). BG, CA, and UNK class ratios are $55.77\%$ , $40.17\%$ , and $4.06\%$ , respectively. The amount of annotated cells and tissue pixels, per data split, can be found in the supplementary material. The detection of TC and BC has clinical relevance. For example, tumor purity [5], computed as the tumor/non-tumor cell ratio in a WSI, has a correlation with cancer prognosis [16, 31, 48]. + +Dataset splits. The dataset is divided into three subsets: training, validation, and test, following a 6:2:2 ratio. To prevent information leaking among the data subsets, we randomly split the dataset per WSI, so that different patches from the same WSI are not included in multiple subsets. We maintain consistent cancer-type ratios in each subset. + +# 4. Empirical Analysis + +The motivation for considering the cell-tissue relationships for the development of cell detection models stem from the biological and hierarchical arrangement of cells and tissues. These insights are further corroborated by two main empirical observations described in the following Sec. 4.1 and Sec. 4.2. + +# 4.1. Interrelation between cell and tissue classes + +We empirically observe the interrelation between specific cell and tissue classes by counting the amount of each annotated cell type within each tissue region, as observed in Tab. 3. Indeed, we verify that in OCELOT, around $93\%$ of $TC$ are located within $CA$ and $85\%$ of $BC$ are found outside + +![](images/27dace7f68fb2fca893cbdbb3f21d718329e026e7910dd3b5dde9bfa519f75cb.jpg) +(a) OCELOT + +![](images/4b308743775d930a5836697e1e3b36a52a7f59ff4d754c4f42ad13db63504812.jpg) +(b) TIGER + +![](images/51b4047f3b0b71fc24b1ddc92ae16e247a35ce760e3b4de01cae095d88e5b070.jpg) +Table 3. Cell counts based on the tissue class. Each value stands for the number of cells located inside the tissue area. $TC$ , $BC$ , $LC$ , $CA$ , and $ST$ stand for Tumor Cell, Background Cell, Lymphocyte Cell, Cancer Area tissue, and Stroma tissue, respectively. +(a) OCELOT +Table 4. Pixel ratio among tissue classes. $CA$ , and $ST$ stand for Cancer Area tissue, and Stroma tissue, respectively. + +![](images/b7943a72b1d8be75727bb468f4fb0b766f3779c56b5f3e7b0523926a3a2965ef.jpg) +(b) TIGER + +of the CA tissue. Note that CA is not the majority tissue class (Tab. 4), therefore, we conclude that there is, in fact, a relationship between the cell and tissue classes. We observe a similar phenomenon in the TIGER dataset when considering the $LC$ and $ST$ classes (Tab. 3 and Tab. 4). + +In practice, pathologists classify cells by taking into account such interrelationships, since it is difficult to classify isolated cells, without considering the larger context of the tissues. As depicted in Fig. 1, pathologists first need to visualize the structure of the tissues at a larger FoV. Then, they zoom in and consider the previously observed context along with the fine-grained details of each individual cell and nearby neighborhood, thus considering the cell-tissue dependencies. Inspired by the behavior of pathologists, we expect a cell detection model to also benefit from understanding the tissue structure from a broader viewpoint. + +# 4.2. Tissue-label Leaking Model + +In the previous section, we observe a strong relationship between certain cell and tissue classes. To further validate the hypothesis that a cell detection model can leverage information from the tissue structure, we design an exploratory experiment where the tissue annotation is provided as an extra input to a cell detection model. To this end, we first crop the corresponding cell patch region from the tissue annotation $y_{l}^{t}$ , and upsample it to match the size of the cell patch, $x_{s}$ ; the cropped tissue annotation is denoted as $y_{s}^{t}$ . Finally, we concatenate $x_{s}$ and $y_{s}^{t}$ at the channels dimension and use this data to train a cell detection model. We denote this model as a Tissue-label leaking model and illustrate it in Fig. 3. Note that this model is not appropriate for real-world scenarios as the tissue labels are unknown at inference time, and is explored for the purpose of empirical analysis. + +When we compare the performance of the tissue-label + +![](images/750b945d175481a119d51943a76a362517ec62889727e91356ab12eb8440775e.jpg) +Figure 3. Tissue label leaking model. This model receives the cell patch $x_{s}$ along with the corresponding tissue labels as input. The region corresponding to the cell patch in the tissue patch annotation is cropped, upsampled, and concatenated to the cell patch. $\oplus$ denotes channel-wise concatenation. + +leaking model with the standard cell detection model on the OCELOT dataset, we observe a significant improvement in terms of mean F1-score $^2$ performance of $+7.69$ and $+9.76$ in the validation and test sets, respectively. Detailed results can be found in Tab. 5. Taking these results into consideration, we conclude that there is significant room to improve the cell detection model, which can be achieved by combining the tasks of cell detection and tissue segmentation. + +# 5. Method + +In this section, we propose to utilize cell-tissue relationships through multi-task learning. First, we propose a set of approaches inspired by the tissue-label leaking model described in Sec. 4.2. These models replace the annotated tissue labels with predictions from an auxiliary tissue segmentation branch. Second, we design a bi-directional information-sharing approach that shares features in both tissue-to-cell and cell-to-tissue directions. The proposed approaches are described in Sec. 5.2 and Sec. 5.3 + +# 5.1. Preliminary + +To deal with cell-tissue sample pairs, i.e., $(x_{s},y_{s}^{c})$ and $(x_{l},y_{l}^{t})$ , we build a dual-branch architecture containing separate networks for the cell and tissue tasks. Similarly to [41], we define cell detection as a segmentation task. Specifically, the cell labels are provided as a segmentation map by drawing a fixed-radius circle centered on each cell point an + +![](images/69d0e324baae725b04e3cb68e530637fb67360348e2ae6a44b463074665dca7c.jpg) +Figure 4. Tissue-prediction injection model injects the tissue segmentation prediction into 1 out of 4 locations of the cell branch: (a) input, (b) after encoder, (c) after ASPP, and (d) after decoder. + +notation and filled with the corresponding class label. At inference time, we find local peaks within the predicted cell probability maps and output them as point predictions. More details are provided in the supplementary material. Treating cell detection as a segmentation task enables us to use the same architecture for both cell and tissue branches, which largely simplifies the training and tuning of the model and reduces the range of possible decisions, such as neural network architecture, or hyper-parameters. We use DeepLabV3+ [10] as a base architecture for both branches and single-task models. + +# 5.2. Tissue-prediction Injection Models + +These models are a simple and practical extension of the tissue-label leaking model, where we inject the predicted tissue probabilities into the cell detection branch instead of leaking the tissue labels. We consider only one injection point in the cell detection branch, but explore four possible alternatives: (a) at the input (Pred-to-input), (b) after the encoder (Pred-to-inter-1), (c) after the ASPP module (Pred-to-inter-2), and (d) after the decoder (Pred-to-output). Fig. 4 depicts this family of models, denominated as Tissue-prediction injection. Since the tissue and cell patches represent different regions, we need to align the content between the tissue and cell feature maps before concatenation. Therefore, we crop the cell corresponding region from the tissue predictions, upsample them, and, finally, concatenate them in the channel dimensions of the feature maps of the cell branch, as illustrated in Fig. 5a. + +# 5.3. Cell-Tissue Feature Sharing Model + +Tissue-prediction injection models share the tissue prediction in a single location and direction, i.e., tissue-to-cell. To enable a more diverse and flexible cell-tissue information flow, we also explore bi-directional feature map sharing from cell-to-tissue (Fig. 5, left) and tissue-to-cell (Fig. 5, + +![](images/6c2472a0393278468d9b7427a30da9677432ee38f69c90d1706abdcaf23c9418.jpg) +(a) Tissue to Cell + +![](images/0d45b9b009457d65f00de40d4c95623489bbb49bf540608eac69a9b3dfa5797a.jpg) +(b) Cell to Tissue + +![](images/708c0a899a6963e3a3c6d719e5cfda1e936012867f1aa85c7165d6ca25edb7c1.jpg) +Figure 5. Information is shared between cell and tissue branches via channel-wise concatenation preceded by a shallow convolutional layer with $3 \times 3$ kernel size. Cropping and upsampling (in Fig. 5a) or downsampling and zero-padding (in Fig. 5b) is applied to match the patch sizes and pixel-alignment between two feature maps from different FoVs. The cell and tissue feature maps are represented in orange and blue, respectively. The red contour denotes the cell patch-associated region in the tissue patch, and the gray regions represent zero padding. +Figure 6. Cell-Tissue Feature Sharing Model has two branches for tissue segmentation and cell detection. Information exchange occurs multiple times between the two branches, indicated by the red vertical arrows. Details regarding the information exchange procedures are described in Fig. 5. + +right). Considering these two operations, we conduct an architecture search procedure to find the optimal feature map sharing configuration between both branches. To limit the search space, we constrain it to only 3 positions in the architecture: after the encoder, after the ASPP module, or after the decoder. Furthermore, we also exclusively allow the branches to inject feature maps at the same depth or position. Finally, we consider only the best-performing model among the $4^{3}$ candidates, which is presented in Fig. 6. We name this model as cell-tissue feature sharing. + +# 6. Experiments and Results + +We validate the hypothesis that incorporating cell-tissue relationships within a cell detection model is beneficial by evaluating the proposed models on OCELOT and 2 other datasets. First, we describe the additional datasets and implementation details. Then, we show how the proposed multi-task learning methods can improve the performance + +of a cell detection task. Lastly, we conduct an ablation study to investigate how important using large FoV patches $\{x_{l}\}$ and corresponding tissue segmentation labels $\{y_{l}^{t}\}$ are in enhancing cell-tissue relationships. + +# 6.1. Other Datasets + +TIGER. As mentioned in Sec. 2.1, TIGER is a public dataset, based on H&E stained images, which includes both cell and tissue annotations in overlapping patches. However, the patch sizes and amount of overlap between cell and tissue annotations highly vary among samples. Therefore, a preprocessing step is necessary to generate paired cell-tissue patches with consistent FoVs. After pre-processing, we obtain 9,888 paired patches of size $512 \times 512$ and $128 \times 128$ pixels for the tissue segmentation and cell detection tasks, respectively, with an MPP of approximately $0.5\mu m / px$ . Most samples in TIGER include small regions with annotated cells, making it necessary to extract smaller patches. This explains why the absolute number of patches is larger than in OCELOT or CARP, although the actual amount of annotated cells and tissue area are smaller than the latter datasets, as shown in Tab. 1. In TIGER, one cell class (lymphocyte cells) and seven tissue classes are annotated. More details about the pre-processing and this dataset are provided in the supplementary material. + +CARP. This is an internal lung cancer dataset of PD-L1 IHC-stained WSIs, containing 6,480 paired patches extracted from 1,012 WSIs. The patch sizes, resolution, annotation protocol, and general configuration are similar to OCELOT, as described in Sec. 3.1. Two cell classes are annotated: PD-L1 positive tumor cells $(TC+)$ , and PD-L1 negative tumor cells $(TC-$ ). The tissue classes are the same as in OCELOT (BG, CA, and UNK). CARP is a real-world and large-scale dataset, with 809.1K annotated cells and 4.1080 mm $^2$ tissue area, which is approximately 10 and 20 times more than OCELOT and TIGER, respectively. Moreover, the stain type is different from OCELOT and TIGER, allowing us to validate the effectiveness of the cell-tissue relationships across different staining methods. + +# 6.2. Implementation details + +As previously mentioned in Sec. 5.1, both cell and tissue branches are based on DeepLabV3+ [10], with a ResNet-34 [18] encoder. The models are trained for 300, 150, and 100 epochs in OCELOT, TIGER, and CARP datasets, respectively. We use the Adam optimizer [24], and, for each experiment, we tune the learning rate within a $[5 \times 10^{-5}, 2 \times 10^{-3}]$ range. The model at the epoch with the best validation set performance is chosen and used for evaluation on the test set. All experiments are repeated 5 times, and we report the mean and $95\%$ confidence interval of the performance metrics. For more implementation details, please refer to the supplementary material. + +
MethodOCELOTTIGERCARP
ValTestValTestValTest
Cell-only68.87±1.7664.44±1.8263.89±1.3953.82±1.2378.48±0.6970.96±1.47
Pred-to-input73.36±0.5969.65±3.9366.00±2.0053.29±1.3079.46±0.7972.98±0.82
Pred-to-inter-172.74±0.5070.54±2.2066.19±1.0255.87±1.7879.74±0.8073.05±0.69
Pred-to-inter-272.68±1.5871.23±0.9665.43±1.1454.75±2.2579.87±0.7873.14±1.53
Pred-to-output66.85±5.6265.05±3.7263.02±0.1653.32±0.4278.92±0.6072.61±0.95
Feature-sharing72.30±0.7368.91±2.5265.64±1.0755.10±2.1879.38±0.7473.00±0.33
Tissue-label leaking76.56±0.8074.20±0.9169.71±0.6161.66±1.1680.13±1.0472.97±0.49
+ +Table 5. Cell detection mean F1 scores per model. With the exception of the Tissue-label Leaking model, the highest score is written in bold and the second highest score is underlined. + +
Methodpopulation (130)kidney (41)endometrium (25)bladder (26)prostate (16)stomach (12)head-neck (10)
Cell-only64.44±1.8264.12±3.4166.88±3.7159.42±5.6165.46±1.6466.19±8.9359.24±5.04
Pred-to-input69.65±3.9363.37±3.2773.31±6.3563.61±4.5568.36±8.6169.34±7.1575.28±0.82
Pred-to-inter-170.54±2.2066.62±7.3073.05±4.1464.35±3.5970.20±2.5671.55±5.2174.50±1.52
Pred-to-inter-271.23±0.9668.94±5.6575.15±2.7064.94±2.3468.35±8.8370.29±0.8173.74±1.72
Pred-to-output65.05±3.7263.38±4.4868.21±4.1159.81±3.9664.18±8.4167.51±6.3960.88±6.77
Feature-sharing68.91±2.5264.97±5.3371.28±5.3063.21±5.3166.25±9.5269.79±2.2573.88±4.39
Tissue-label leaking74.20±0.9175.75±1.9772.71±1.1271.18±0.9374.56±1.2575.24±1.7175.78±0.46
+ +# 6.3. Main Results + +The cell detection results obtained by the Cell-only baseline, the proposed approaches, and the Tissue-label leaking model are shown in Tab. 5. The Cell-only baseline is a simple cell detection model that only receives the small FoV patches as input. In other words, it only considers the cell branch in Fig. 4 and neither tissue annotations nor large FoV patches are leveraged. The Tissue-label leaking model, described in Sec. 5.2, receives the tissue annotations as input and serves as an exploratory experiment to obtain insight regarding how much the cell detection task can benefit from leveraging tissue annotation. + +From Tab. 5, we observe that all the cell-tissue multi-task learning-based approaches, except for the Pred-to-output model, outperform the Cell-only baseline across all datasets. These improvements imply that cell detection on small FoV patches benefits from learning cell-tissue interrelationships from both large FoVs and tissue annotations. We hypothesize that the reason for the low performance of the Pred-to-output model is that the injection of tissue prediction to the cell branch happens too late. Therefore, there is a lack of network capacity for fusing cell and tissue information. + +Furthermore, in Tab. 6, we report mean F1 scores per or + +Table 6. Per-organ cell detection mean F1 scores in the OCELOT test set. Population mean F1 scores are computed from all the patches in the test set. In parentheses, we indicate the number of samples of each organ subset. With the exception of the Tissue-label Leaking model, the highest score is written in bold and the second highest score is underlined. + +
Further utilizePred-to-inter-2Feature-sharing
Tissue labelLarge FoVValTestValTest
--68.87±1.7664.45±1.8268.87±1.7664.45±1.82
✓-70.21±1.4964.99±1.7770.36±1.3466.58±2.02
-✓68.78±0.9465.56±3.6669.61±1.5966.22±1.53
✓✓72.68±1.5871.23±0.9672.30±0.7368.91±2.52
+ +Table 7. Dataset ablation study. In all experiments, the model architecture is fixed and only input/output changes. The first row corresponds to the Cell-only baseline. + +gan in the OCELOT test set. Except for the Pred-to-output, all approaches show improvement compared to Cell-only baseline in most organs. Especially, our best-performing model, Pred-to-inter-2, demonstrates superior performance over Cell-only in all 6 organs with a significant margin. Such a result shows that considering cell-tissue relationships can generally help the cell detection task across various organs. + +# 6.4. Dataset Ablation Study + +Two data components in OCELOT contribute to better leveraging cell-tissue relationships for the task of cell detection: (1) the large FoV patches $x_{l}$ , and (2) the corresponding tissue annotations $y_{l}^{t}$ that allows us to have a tissue segmentation objective and a multi-task learning setting. To verify + +![](images/a661531a0b3009d2bc0a111da695b1eb932502a0c9c63f3295fe923ab328e9d6.jpg) +Figure 7. Qualitative comparison between the Pred-to-inter-2 and Cell-only models. Pred-to-inter-2 shows enhanced detection results aided by understanding the tissue environment from a broader context (blue: tumor cells, yellow: background cells, green: cancer area). + +the effectiveness of each data component, we conduct an ablation study where each component is evaluated separately while keeping the model architecture fixed. First, if only the tissue segmentation label is provided without a large FoV patch, the tissue branch takes $x_{s}$ as an input and it is trained with both the tissue and the cell detection objectives. Second, if the large FoV patch is given without the corresponding tissue label, the tissue branch takes $x_{l}$ as an input and the model is only trained with the cell detection objective. + +The results of the ablation study are shown in Tab. 7. We observe that each of the individual components on their own lead to improved performance. Hence, we can conclude that a large FoV patch and tissue segmentation label can contribute to better cell detection by providing a broader context and encoding the cell-tissue relationship, respectively. Moreover, when utilized simultaneously, we observe that both components work synergically, leading to an even better performance improvement. We also found that the tissue branch generates better tissue predictions when trained with a large FoV. Thus, providing accurate tissue information to the cell branch is important for boosting performance. Please, refer to the supplementary material for a detailed tissue performance comparison. Finally, this study shows evidence that the patch configuration defined in OCELOT is suitable for cell-tissue co-training, with a smaller FoV for cell detection $(x_{s},y_{s}^{c})$ and a larger FoV for tissue segmentation $(x_{l},y_{l}^{t})$ . + +# 6.5. Qualitative Analysis + +In this section, we visualize the cell predictions of the Cell-only and Pred-to-inter-2 models. We select the Pred-to-inter-2 model, as it is the overall best-performing model in + +our experiments. In general, when compared to other background cells, tumor cells have the following characteristics: large size and irregular shape. However, cancer is heterogeneous and this is not always the case. Indeed, in Fig. 7, most of the cells are small and have a regular round shape. Based on these appearances, and without a larger context, those cells can be easily misclassified as background cells, which is the case of the Cell-only model. On the other hand, Pred-to-inter-2 shows a more accurate prediction by correctly understanding the cancer area in large FoV regions. This implies that Pred-to-inter-2 indeed considers both the morphology of cells and the tissue context, while Cell-only relies on the cells' morphology alone. + +# 7. Conclusion + +In histopathology, there is a hierarchical organization and interrelationship between cells and tissues. Hence, we hypothesize that it can be leveraged to improve the performance of cell detection tasks. Throughout this paper, we have shown evidence that such cell-tissue relationships exist across multiple setups. Indeed, we observe improvements when utilizing large FoVs with corresponding tissue annotations together with cell annotations in a multi-task learning approach, using simple neural architecture designs. Although the cell detection task clearly benefits from the multitask learning framework, the improvement on the tissue segmentation task was not yet investigated, and we leave this topic as a future direction. We hope this work, especially the OCELOT dataset, initiates new and promising research directions in CPATH to further study the cell-tissue relationships and improve cell detection and tissue segmentation. + +# References + +[1] MICCAI 2014 Grand Challenges. https://wiki.cancerimagingarchive.net/display/Public/MICCAI+2014+Grand+Challenges, 2014. [Online; accessed Nov-2022]. 2 +[2] TIGER: Grand Challenge. https://tiger.grand-challenge.org, 2022. [Online; accessed Nov-2022]. 2 +[3] Esther Abels, Liron Pantanowitz, Famke Aeffner, Mark D Zarella, Jeroen van der Laak, Marilyn M Bui, Venkata NP Vemuri, Anil V Parwani, Jeff Gibbs, Emmanuel Agosto-Arroyo, et al. Computational pathology definitions, best practices, and recommendations for regulatory guidance: a white paper from the digital pathology association. The Journal of Pathology, 249(3):286-294, 2019. 1 +[4] Mohamed Amgad, Lamees A Atteya, Hagar Hussein, Kareem Hosny Mohammed, Ehab Hafiz, Maha AT Elsebaie, Ahmed M Alhusseiny, Mohamed Atef AlMoslemany, Abdelmagid M Elmatboly, Philip A Pappalardo, et al. Nuclei: A scalable crowdsourcing approach and dataset for nucleus classification and segmentation in breast cancer. *Giga* Science, 11, 2022. 2 +[5] Vahid Azimi, Young Hwan Chang, Guillaume Thibault, Jaclyn Smith, Takahiro Tsujikawa, Benjamin Kukull, Braden Jensen, Christopher Corless, Adam Margolin, and Joe W Gray. Breast cancer histopathology image analysis pipeline for tumor purity estimation. In 2017 IEEE 14th International Symposium on Biomedical Imaging (ISBI 2017), pages 1137-1140. IEEE, 2017. 4 +[6] Tian Bai, Jiayu Xu, and Fuyong Xing. Multi-field of view aggregation and context encoding for single-stage nucleus recognition. In MICCAI, pages 382-392. Springer, 2020. 3 +[7] Wouter Bulten, Kimmo Kartasalo, Po-Hsuan Cameron Chen, Peter Ström, Hans Pinckaers, Kunal Nagpal, Yuannan Cai, David F Steiner, Hester van Boven, Robert Vink, et al. Artificial intelligence for diagnosis and gleason grading of prostate cancer: the panda challenge. Nature Medicine, 28(1):154-163, 2022. 1, 2 +[8] Lyndon Chan, Mahdi S Hosseini, Corwyn Rowsell, Konstantinos N Plataniotis, and Savvas Damaskinos. Histosegnet: Semantic segmentation of histological tissue type in whole slide images. In ICCV, pages 10662-10671, 2019. 2 +[9] Hao Chen, Xiaojuan Qi, Lequan Yu, and Pheng-Ann Heng. Dcan: deep contour-aware networks for accurate gland segmentation. In CVPR, pages 2487–2496, 2016. 2 +[10] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In ECCV, pages 801-818, 2018. 5, 6 +[11] Qian Da, Xiaodi Huang, Zhongyu Li, Yanfei Zuo, Chenbin Zhang, Jingxin Liu, Wen Chen, Jiahui Li, Dou Xu, Zhiqiang Hu, et al. Digestpath: a benchmark dataset with challenge review for the pathological detection and segmentation of digestive-system. Medical Image Analysis, page 102485, 2022. 1, 2 +[12] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248–255. IEEE, 2009. 3 + +[13] James A Diao, Jason K Wang, Wan Fung Chui, Victoria Mountain, Sai Chowdary Gullapally, Ramprakash Srinivasan, Richard N Mitchell, Benjamin Glass, Sara Hoffman, Sudha K Rao, et al. Human-interpretable image features derived from densely mapped cancer pathology slides predict diverse molecular phenotypes. Nature communications, 12(1):1-15, 2021. 1 +[14] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. *IJCV*, 88(2):303–338, 2010. 3 +[15] Jevgenij Gamper, Navid Alemi Koohbanani, Ksenija Benes, Simon Graham, Mostafa Jahanifar, Syed Ali Khurram, Ayesha Azam, Katherine Hewitt, and Nasir Rajpoot. Pan-nuke dataset extension, insights and baselines. arXiv preprint arXiv:2003.10778, 2020. 2 +[16] Zhe Gong, Jieyun Zhang, and Weijian Guo. Tumor purity as a prognosis and immunotherapy relevant feature in gastric cancer. Cancer Medicine, 9(23):9052-9063, 2020. 4 +[17] Simon Graham, Quoc Dang Vu, Shan E Ahmed Raza, Ayesha Azam, Yee Wah Tsang, Jin Tae Kwak, and Nasir Rajpoot. Hover-net: Simultaneous segmentation and classification of nuclei in multi-tissue histology images. Medical Image Analysis, 58:101563, 2019. 1, 2 +[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. pages 770-778, 2016. 6 +[19] David Joon Ho, Dig VK Yarlagadda, Timothy M D'Alfonso, Matthew G Hanna, Anne Grabenstetter, Peter Ntiamoah, Edi Brogi, Lee K Tan, and Thomas J Fuchs. Deep multimagnification networks for multi-class breast cancer image segmentation. Computerized Medical Imaging and Graphics, 88:101866, 2021. 3 +[20] Mahdi S Hosseini, Lyndon Chan, Gabriel Tse, Michael Tang, Jun Deng, Sajad Norouzi, Corwyn Rowsell, Konstantinos N Plataniotis, and Savvas Damaskinos. Atlas of digital pathology: A generalized hierarchical histological tissue type-annotated database for deep learning. In CVPR, pages 11747–11756, 2019. 2 +[21] Carolyn Hutter and Jean Claude Zenklusen. The cancer genome atlas: creating lasting value beyond its data. Cell, 173(2):283-285, 2018. 3 +[22] Konstantinos Kamnitsas, Christian Ledig, Virginia FJ Newcombe, Joanna P Simpson, Andrew D Kane, David K Menon, Daniel Rueckert, and Ben Glocker. Efficient multi-scale 3d cnn with fully connected crf for accurate brain lesion segmentation. Medical Image Analysis, 36:61-78, 2017. 3 +[23] Davood Karimi, Guy Nir, Ladan Fazli, Peter C Black, Larry Goldenberg, and Septimiu E Salcudean. Deep learning-based gleason grading of prostate cancer from histopathology images—role of multiscale decision aggregation and data augmentation. IEEE journal of Biomedical and Health Informatics, 24(5):1413-1426, 2019. 2 +[24] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 6 +[25] Neeraj Kumar, Ruchika Verma, Deepak Anand, Yanning Zhou, Omer Fahri Onder, Efstratios Tsougenis, Hao Chen, Pheng-Ann Heng, Jiahui Li, Zhiqiang Hu, et al. A multi + +organ nucleus segmentation challenge. IEEE transactions on Medical Imaging, 39(5):1380-1391, 2019. 1, 2 +[26] Shyam Lal, Devikalyan Das, Kumar Alabhya, Anirudh Kanfade, Aman Kumar, and Jyoti Kini. Nucleisegnet: robust deep learning architecture for the nuclei segmentation of liver cancer histopathology images. Computers in Biology and Medicine, 128:104075, 2021. 1, 2 +[27] Chunggi Lee, Seonwook Park, Heon Song, Jeongun Ryu, Sanghoon Kim, Haejoon Kim, Sérgio Pereira, and Donggeun Yoo. Interactive multi-class tiny-object detection. In CVPR, pages 14136-14145, 2022. 3 +[28] Wenqi Li, Siyamalan Manivannan, Shazia Akbar, Jianguo Zhang, Emanuele Trucco, and Stephen J McKenna. Gland segmentation in colon histology images using hand-crafted features and convolutional neural networks. In 2016 IEEE 13th international symposium on biomedical imaging (ISBI), pages 1405-1408. IEEE, 2016. 2 +[29] Xieli Li, Yuanyuan Wang, Qisheng Tang, Zhen Fan, and Jinhua Yu. Dual u-net for the segmentation of overlapping glioma nuclei. IEEE Access, 7:84040-84052, 2019. 2 +[30] Dongnan Liu, Donghao Zhang, Yang Song, Chaoyi Zhang, Fan Zhang, Lauren O'Donnell, and Weidong Cai. Nuclei segmentation via a deep panoptic model with semantic feature fusion. In *IJCAI*, pages 861-868, 2019. 2 +[31] Yihao Mao, Qingyang Feng, Peng Zheng, Liangliang Yang, Tianyu Liu, Yuqiu Xu, Dexiang Zhu, Wenju Chang, Meiling Ji, Li Ren, et al. Low tumor purity is associated with poor prognosis, heavy mutation burden, and intense immune phenotype in colon cancer. Cancer Management and Research, 10:3569, 2018. 4 +[32] Peter Naylor, Marick Laé, Fabien Reyal, and Thomas Walter. Segmentation of nuclei in histopathology images by deep regression of the distance map. IEEE transactions on Medical Imaging, 38(2):448-459, 2018. 2 +[33] Guy Nir, Soheil Hor, Davood Karimi, Ladan Fazli, Brian F Skinnider, Peyman Tavassoli, Dmitry Turbin, Carlos F Villamil, Gang Wang, R Storey Wilson, et al. Automatic grading of prostate cancer in digitized histopathology images: Learning from multiple experts. Medical Image Analysis, 50:167-180, 2018. 2 +[34] Sehnoon Park, Chan-Young Ock, Hyojin Kim, Sergio Pereira, Seonwook Park, Minuk Ma, Sangjoon Choi, Seokhwi Kim, Seunghwan Shin, Brian Jaehong Aum, et al. Artificial intelligence-powered spatial analysis of tumor-infiltrating lymphocytes as complementary biomarker for immune checkpoint inhibition in non-small-cell lung cancer. Journal of Clinical Oncology, 40(17):1916, 2022. 1 +[35] Pushpak Pati, Guillaume Jaume, Lauren Alisha Fernandes, Antonio Foncubierta-Rodriguez, Florinda Feroce, Anna Maria Anniciello, Giosue Scognamiglio, Nadia Brancati, Daniel Riccio, Maurizio Di Bonito, et al. Hact-net: A hierarchical cell-to-tissue graph neural network for histopathological image classification. In Uncertainty for Safe Utilization of Machine Learning in Medical Imaging, and Graphs in Biomedical Image Analysis, pages 208-219. Springer, 2020. 3 +[36] Ziniu Qian, Kailu Li, Maode Lai, Eric I Chang, Bingzheng Wei, Yubo Fan, Yan Xu, et al. Transformer based multiple + +instance learning for weakly supervised histopathology image segmentation. arXiv preprint arXiv:2205.08878, 2022. 2 +[37] Hui Qu, Pengxiang Wu, Qiaoying Huang, Jingru Yi, Zhennan Yan, Kang Li, Gregory M Riedlinger, Subhajyoti De, Shaoting Zhang, and Dimitris N Metaxas. Weakly supervised deep nuclei segmentation using partial points annotation in histopathology images. IEEE transactions on Medical Imaging, 39(11):3655-3666, 2020. 1, 2 +[38] Roberto Salgado, Carsten Denkert, S Demaria, N Sirtaine, F Klauschen, Giancarlo Pruneri, S Wienert, Gert Van den Eynden, Frederick L Baehner, Frederique Pénault-Llorca, et al. The evaluation of tumor-infiltrating lymphocytes (tils) in breast cancer: recommendations by an international tils working group 2014. Annals of Oncology, 26(2):259-271, 2015. 2 +[39] Rüdiger Schmitz, Frederic Madesta, Maximilian Nielsen, Jenny Krause, Stefan Steurer, René Werner, and Thomas Rösch. Multi-scale fully convolutional neural networks for histopathology image segmentation: from nuclear aberrations to the global tissue architecture. Medical Image Analysis, 70:101996, 2021. 3 +[40] Peng Sun, Jiehua He, Xue Chao, Keming Chen, Yuanyuan Xu, Qitao Huang, Jingping Yun, Mei Li, Rongzhen Luo, Jinbo Kuang, et al. A computational tumor-infiltrating lymphocyte assessment method comparable with visual reporting guidelines for triple-negative breast cancer. EBioMedicine, 70:103492, 2021. 1 +[41] Zaneta Swiderska-Chadaj, Hans Pinckaers, Mart van Rijthoven, Maschenka Balkenhol, Margarita Melnikova, Oscar Geessink, Quirine Manson, Mark Sherman, Antonio Polonia, Jeremy Parry, et al. Learning to detect lymphocytes in immunohistochemistry with deep learning. Medical Image Analysis, 58:101547, 2019. 5 +[42] Hiroki Tokunaga, Yuki Teramoto, Akihiko Yoshizawa, and Ryoma Bise. Adaptive weighting multi-field-of-view cnn for semantic segmentation in pathology. In CVPR, pages 12597-12606, 2019. 1, 3 +[43] Mart Van Rijthoven, Maschenka Balkenhol, Karina Silina, Jeroen Van Der Laak, and Francesco Ciompi. Hooknet: Multi-resolution convolutional neural networks for semantic segmentation in histopathology whole-slide images. Medical Image Analysis, 68:101890, 2021. 3 +[44] Ruchika Verma, Neeraj Kumar, Abhijeet Patil, Nikhil Cherian Kurian, Swapnil Rane, Simon Graham, Quoc Dang Vu, Mieke Zwager, Shan E Ahmed Raza, Nasir Rajpoot, et al. Monusac2020: A multi-organ nuclei segmentation and classification challenge. IEEE Transactions on Medical Imaging, 40(12):3413-3423, 2021. 2 +[45] Quoc Dang Vu, Simon Graham, Tahsin Kurc, Minh Nguyen Nhat To, Muhammad Shaban, Talha Qaiser, Navid Alemi Koohbanani, Syed Ali Khurram, Jayashree Kalpathy-Cramer, Tianhao Zhao, et al. Methods for segmentation and classification of digital microscopy tissue images. Frontiers in Bioengineering and Biotechnology, page 53, 2019. 2 +[46] Jingwen Wang, Richard J Chen, Ming Y Lu, Alexander Baras, and Faisal Mahmood. Weakly supervised prostate + +tma classification via graph convolutional networks. In 2020 IEEE 17th International Symposium on Biomedical Imaging (ISBI), pages 239-243. IEEE, 2020. 3 +[47] Yan Xu, Zhipeng Jia, Liang-Bo Wang, Yuqing Ai, Fang Zhang, Maode Lai, Eric I Chang, et al. Large scale tissue histopathology image classification, segmentation, and visualization via deep convolutional activation features. BMC bioinformatics, 18(1):1-17, 2017. 2 +[48] Chuanbao Zhang, Wen Cheng, Xiufang Ren, Zheng Wang, Xing Liu, Guanzhang Li, Sheng Han, Tao Jiang, and Anhua Wu. Tumor purity as an underlying key factor in gliomatumor purity in glioma. Clinical Cancer Research, 23(20):6279-6291, 2017. 4 +[49] Bingchao Zhao, Xin Chen, Zhi Li, Zhiwen Yu, Su Yao, Lixu Yan, Yuqian Wang, Zaiyi Liu, Changhong Liang, and Chu Han. Triple u-net: Hematoxylin-aware nuclei segmentation with progressive dense feature aggregation. Medical Image Analysis, 65:101786, 2020. 1, 2 +[50] Ke Zhao, Zhenhui Li, Su Yao, Yingyi Wang, Xiaomei Wu, Zeyan Xu, Lin Wu, Yanqi Huang, Changhong Liang, and Zaiyi Liu. Artificial intelligence quantified tumour-stroma ratio is an independent predictor for overall survival in resectable colorectal cancer. EBioMedicine, 61:103054, 2020. 2 +[51] Yanning Zhou, Simon Graham, Navid Alemi Koohbanani, Muhammad Shaban, Pheng-Ann Heng, and Nasir Rajpoot. Cgc-net: Cell graph convolutional network for grading of colorectal cancer histology images. In ICCV Workshops, pages 0–0, 2019. 3 +[52] Chuang Zhu, Ke Mei, Ting Peng, Yihao Luo, Jun Liu, Ying Wang, and Mulan Jin. Multi-level colonoscopy malignant tissue detection with adversarial cac-unet. Neurocomputing, 438:165-183, 2021. 2 \ No newline at end of file diff --git a/2023/OCELOT_ Overlapped Cell on Tissue Dataset for Histopathology/images.zip b/2023/OCELOT_ Overlapped Cell on Tissue Dataset for Histopathology/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..3496efe5d292777b91780c31733401b1113a52b1 --- /dev/null +++ b/2023/OCELOT_ Overlapped Cell on Tissue Dataset for Histopathology/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:eb1e10914d99e3a13aeedc34dbeb72dd506787c150fe1aa3d2d3a2a6d993d079 +size 645097 diff --git a/2023/OCELOT_ Overlapped Cell on Tissue Dataset for Histopathology/layout.json b/2023/OCELOT_ Overlapped Cell on Tissue Dataset for Histopathology/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d2ca9bef8fbed158285a9d193f7e49b72ea931e8 --- /dev/null +++ b/2023/OCELOT_ Overlapped Cell on Tissue Dataset for Histopathology/layout.json @@ -0,0 +1,8236 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 96, + 103, + 499, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 103, + 499, + 121 + ], + "spans": [ + { + "bbox": [ + 96, + 103, + 499, + 121 + ], + "type": "text", + "content": "OCELOT: Overlapped Cell on Tissue Dataset for Histopathology" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 143, + 539, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 143, + 539, + 187 + ], + "spans": [ + { + "bbox": [ + 53, + 143, + 539, + 187 + ], + "type": "text", + "content": "Jeongun Ryu* Aaron Valero Puche* JaeWoong Shin* Seonwook Park Biagio Brattoli Jinhee Lee Wonkyung Jung Soo Ick Cho Kyunghyun Paeng Chan-Young Ock Donggeun Yoo Sérgio Pereira Lunit Inc." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 113, + 192, + 473, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 192, + 473, + 217 + ], + "spans": [ + { + "bbox": [ + 113, + 192, + 473, + 217 + ], + "type": "text", + "content": "{rjw0205, aaron.valero, jwoong.shin, spark, biagio, jinhee.lee, wkjung, sooickcho, khpaeng, ock.chanyoung, dgyoo, sergio}@lunit.io" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 236, + 191, + 249 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 236, + 191, + 249 + ], + "spans": [ + { + "bbox": [ + 143, + 236, + 191, + 249 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 262, + 290, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 262, + 290, + 585 + ], + "spans": [ + { + "bbox": [ + 47, + 262, + 290, + 585 + ], + "type": "text", + "content": "Cell detection is a fundamental task in computational pathology that can be used for extracting high-level medical information from whole-slide images. For accurate cell detection, pathologists often zoom out to understand the tissue-level structures and zoom in to classify cells based on their morphology and the surrounding context. However, there is a lack of efforts to reflect such behaviors by pathologists in the cell detection models, mainly due to the lack of datasets containing both cell and tissue annotations with overlapping regions. To overcome this limitation, we propose and publicly release OCELOT, a dataset purposely dedicated to the study of cell-tissue relationships for cell detection in histopathology. OCELOT provides overlapping cell and tissue annotations on images acquired from multiple organs. Within this setting, we also propose multi-task learning approaches that benefit from learning both cell and tissue tasks simultaneously. When compared against a model trained only for the cell detection task, our proposed approaches improve cell detection performance on 3 datasets: proposed OCELOT, public TIGER, and internal CARP datasets. On the OCELOT test set in particular, we show up to 6.79 improvement in F1-score. We believe the contributions of this paper, including the release of the OCELOT dataset at https://lunit-io.github.io/research/publications/ocelot are a crucial starting point toward the important research direction of incorporating cell-tissue relationships in computation pathology." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 593, + 128, + 605 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 593, + 128, + 605 + ], + "spans": [ + { + "bbox": [ + 47, + 593, + 128, + 605 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 612, + 288, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 612, + 288, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 612, + 288, + 696 + ], + "type": "text", + "content": "Computational Pathology (CPATH) [3] is a branch of digital pathology that develops methodologies for the analysis of digitized patient specimens, such as Whole-Slide-Images (WSIs). Cell detection in histology images [26, 37, 49] is one of the most important tasks in CPATH. It allows the quantification and analysis of different cell types, which can lead to better prognosis evaluation [34, 40] and patient" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 238, + 545, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 238, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 306, + 238, + 545, + 285 + ], + "type": "text", + "content": "treatment planning while maintaining medical interpretability [13]. Since it has the potential to impact human lives, high-performance cell detection models are essential in real-world applications and need to be investigated." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 289, + 546, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 289, + 546, + 444 + ], + "spans": [ + { + "bbox": [ + 304, + 289, + 546, + 444 + ], + "type": "text", + "content": "To better locate and classify cells, detailed morphological characteristics such as color and shape are crucial. Consequently, cell detection datasets are typically collected at high magnification but small Field-of-View (FoV). However, this can make the cell detection model overly rely on appearance details, without understanding the broader context [42]. This context can help cell detection by providing information about how cells are arranged and grouped together to form high-level tissue structures. In practice, expert annotators (pathologists) first zoom out to understand these broad tissue structures. Next, they zoom in to better classify individual cells while taking into account the context information, as depicted in Fig. 1." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 447, + 547, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 447, + 547, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 447, + 547, + 567 + ], + "type": "text", + "content": "The behavior of pathologists can be transferred to deep learning, for instance, through a multi-task strategy combining cell detection tasks at high magnification and tissue segmentation at low magnification. This type of approach would allow the model to share knowledge across different tasks and FoVs. However, to train such an approach, a combined dataset with cell-tissue overlapping regions is required; unfortunately, most existing datasets only target a single task, either cell detection [17,25] or tissue segmentation [7, 11]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": "In this paper, we introduce a new research direction: studying cell-tissue relationships for cell detection. First, we publish the OCELOT dataset, which contains cell and tissue annotations in small and large FoV patches, respectively, with overlapping regions. Additionally, the data is collected from WSIs of multiple organs. This can provide the necessary data for researchers to study cell-tissue relationships and their effect on cell detection. Second, we introduce simple multi-task learning approaches for cell detection that can benefit from cell-tissue relationships and demonstrate their advantages over 3 different datasets. These approaches consistently show better cell detection performance compared to" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 703, + 129, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 703, + 129, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 703, + 129, + 712 + ], + "type": "text", + "content": "*: Equal contribution" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23902" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 100, + 64, + 492, + 247 + ], + "blocks": [ + { + "bbox": [ + 100, + 64, + 492, + 247 + ], + "lines": [ + { + "bbox": [ + 100, + 64, + 492, + 247 + ], + "spans": [ + { + "bbox": [ + 100, + 64, + 492, + 247 + ], + "type": "image", + "image_path": "1a09049a2141ea141e91630bf8e9c6d20f719777a882510d8717a46ee7200df8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 251, + 547, + 318 + ], + "lines": [ + { + "bbox": [ + 46, + 251, + 547, + 318 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 547, + 318 + ], + "type": "text", + "content": "Figure 1. Behavior of pathologists and cell detection models. Typical cell detection models infer their predictions only by looking at a limited FoV region. Failure cases can occur when cells are difficult to be classified solely by their morphology, i.e., while disregarding the larger architecture of the tissues (context). In the above example, some groups of tumor cells are misclassified as background cells due to their morphology; these tumor cells are smaller and rounder than the nearby ones. Note that the large size and irregular shape are representative characteristics of tumor cells. Pathologists overcome these limitations by further understanding the tissue context of the larger FoV region. Misclassified cells can be easily corrected by understanding that such regions are cancer areas." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 335, + 287, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 335, + 287, + 394 + ], + "spans": [ + { + "bbox": [ + 46, + 335, + 287, + 394 + ], + "type": "text", + "content": "the cell-only baseline, i.e., a model trained only with small FoV patches with the corresponding cell annotations. We hope that our proposed OCELOT dataset and methods will encourage the CPATH community to learn how to reflect cell-tissue relationships better to improve cell detection." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 396, + 286, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 396, + 286, + 406 + ], + "spans": [ + { + "bbox": [ + 59, + 396, + 286, + 406 + ], + "type": "text", + "content": "Our contributions are 3-fold and summarized as follows," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 415, + 287, + 533 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 57, + 415, + 287, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 415, + 287, + 438 + ], + "spans": [ + { + "bbox": [ + 57, + 415, + 287, + 438 + ], + "type": "text", + "content": "- The first work that exploits cell-tissue relationships for better cell detection, to the best of our knowledge." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 444, + 287, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 444, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 57, + 444, + 287, + 479 + ], + "type": "text", + "content": "- We release OCELOT, a dataset with overlapping cell and tissue annotations based on Hematoxylin and Eosin (H&E) stained WSIs of multiple organs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 485, + 287, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 485, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 57, + 485, + 287, + 533 + ], + "type": "text", + "content": "- We introduce several approaches that boost cell detection performance via multi-task learning, and empirically confirm that these methods generalize well across different datasets and histological stainings." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 544, + 134, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 544, + 134, + 555 + ], + "spans": [ + { + "bbox": [ + 47, + 544, + 134, + 555 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 563, + 286, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 563, + 286, + 586 + ], + "spans": [ + { + "bbox": [ + 47, + 563, + 286, + 586 + ], + "type": "text", + "content": "2.1. Datasets for Cell detection and Tissue segmentation Tasks" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 713 + ], + "type": "text", + "content": "In recent years, numerous datasets have been released for tackling cell detection. Some of those works only target a single organ [17,32], while others consider multiple ones [4, 15,25,44,45]. The availability of these datasets enables the CPATH community to push forward the development and improvement of cell detection models [17,26,29,30,37,49]. In addition, tissue segmentation datasets have also been proposed for prostate [7,23,33], colorectal [50], brain [1], and multiple organs [20]. Some examples of tissue segmentation works can be found in [8,9,28,36,47,52]. The dataset" + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 340, + 332, + 512, + 378 + ], + "blocks": [ + { + "bbox": [ + 340, + 332, + 512, + 378 + ], + "lines": [ + { + "bbox": [ + 340, + 332, + 512, + 378 + ], + "spans": [ + { + "bbox": [ + 340, + 332, + 512, + 378 + ], + "type": "table", + "html": "
DatasetTissue Area# CellOrgans
OCELOT4.267cm2114.7Kmultiple
TIGER2.536cm250.8Kbreast
", + "image_path": "7bd920886542aaefcbb9dc8a92f50fc597a648bc31761a2ba2e7629b06a84432.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 384, + 545, + 407 + ], + "lines": [ + { + "bbox": [ + 306, + 384, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 306, + 384, + 545, + 407 + ], + "type": "text", + "content": "Table 1. Dataset comparison in terms of physical tissue annotated area and total cell counting per dataset." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 422, + 545, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 422, + 545, + 518 + ], + "spans": [ + { + "bbox": [ + 304, + 422, + 545, + 518 + ], + "type": "text", + "content": "in [11] is composed of a cell detection subset and a tissue segmentation subset. However, the subsets are annotated independently and from different patient groups, and, therefore, there are no overlaps between the cell and tissue data. Because of the lack of overlapping data in the aforementioned datasets, it is difficult to build an end-to-end framework to learn cell-tissue relationships by jointly training on the cell and tissue tasks." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 520, + 545, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 520, + 545, + 591 + ], + "spans": [ + { + "bbox": [ + 304, + 520, + 545, + 591 + ], + "type": "text", + "content": "The recently released TIGER dataset [2] contains both cell and tissue annotations to study tumor-infiltrating lymphocytes [38] in H&E breast cancer WSIs. All the cell-annotated areas exist inside the tissue-annotated area, however, this work does not propose nor initiate any effort toward the integration of both cell and tissue tasks." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 594, + 546, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 546, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 546, + 712 + ], + "type": "text", + "content": "To further promote the development of methods that leverage the cell-tissue relationship for the task of cell detection, we propose OCELOT, which is designed to capture the hierarchical relationship between cells and tissues, especially in the tumor environment. OCELOT contains roughly two times more cell and tissue annotations than TIGER (see Tab. 1). Additionally, the data was collected from multiple organs to enable the investigation of the generalizability of cell-tissue relationships over various cancer types. In the end, we utilized both OCELOT and TIGER to reveal the" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23903" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 286, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 286, + 84 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 286, + 84 + ], + "type": "text", + "content": "power of cell-tissue relationship for cell detection in Sec. 6." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 91, + 217, + 105 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 91, + 217, + 105 + ], + "spans": [ + { + "bbox": [ + 47, + 91, + 217, + 105 + ], + "type": "text", + "content": "2.2. Leveraging Large Field of View" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 110, + 287, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 110, + 287, + 289 + ], + "spans": [ + { + "bbox": [ + 46, + 110, + 287, + 289 + ], + "type": "text", + "content": "Some studies [6, 19, 22, 39, 42, 43] extract a large FoV region as an additional input to improve detection/segmentation performance on smaller FoV regions. [22] proposes a dual pathway 3D CNN for brain lesion segmentation, where each pathway receives both small and large center-shared FoV regions as input. Similar studies are also conducted in the CPATH domain for tissue segmentation [19, 39, 42, 43] and cell detection [6]. To fuse different FoV patches, [42] introduces a weighting mechanism and [43] proposes a multi-scale merging block composed of convolution and concatenation. Nevertheless, no inter-task relationship is considered in the previously mentioned methods. In contrast, our models take advantage of the large contextual information while learning the cell-tissue relationship via multi-task objectives at different FoVs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 297, + 244, + 310 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 297, + 244, + 310 + ], + "spans": [ + { + "bbox": [ + 47, + 297, + 244, + 310 + ], + "type": "text", + "content": "2.3. Leveraging Cell-Tissue Relationships" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 316, + 287, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 316, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 287, + 495 + ], + "type": "text", + "content": "To our understanding, there is no study that considers cell-tissue relationships for cell detection or tissue segmentation tasks. On the other hand, a few efforts have attempted to link tissue and cell for image classification using graph-based methods [35,46,51]. Such studies represent the tissue structure as a graph of detected cells, based on the medical knowledge that cells form tissues. [35] explicitly considers the cell-tissue relationship for the task of breast cancer subtyping, by an interaction between tissue-level and cell-level graphs with a cell-to-tissue hierarchy. However, these methods treat the cell/tissue graph generation as a pre-processing step, by using the inference output of independently pretrained cell detection and tissue segmentation models. In contrast, we directly target the improvement of cell detection by utilizing the cell-tissue relationship." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 506, + 115, + 518 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 506, + 115, + 518 + ], + "spans": [ + { + "bbox": [ + 47, + 506, + 115, + 518 + ], + "type": "text", + "content": "3. OCELOT" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 527, + 287, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 527, + 287, + 575 + ], + "spans": [ + { + "bbox": [ + 46, + 527, + 287, + 575 + ], + "type": "text", + "content": "In this section, we introduce OCELOT, a histopathology dataset specifically built to enable the development of methods that leverage cell and tissue relationships. Each sample of the OCELOT dataset " + }, + { + "bbox": [ + 46, + 527, + 287, + 575 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 527, + 287, + 575 + ], + "type": "text", + "content": " is composed of six components," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 101, + 584, + 287, + 602 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 584, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 101, + 584, + 287, + 602 + ], + "type": "interline_equation", + "content": "\\mathcal {D} = \\left\\{\\left(x _ {s}, y _ {s} ^ {c}, x _ {l}, y _ {l} ^ {t}, c _ {x}, c _ {y}\\right) _ {i} \\right\\} _ {i = 1} ^ {N} \\tag {1}", + "image_path": "a391c3c1e5397c0db9f8ef5cd051d52de55eaffe3502ace4d67eb9a55e903382.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "x_{s}, x_{l}" + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": " are the small and large FoV patches extracted from the WSI, " + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "y_{s}^{c}, y_{l}^{t}" + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": " refer to the corresponding cell and tissue annotations, respectively, and " + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "c_{x}, c_{y}" + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": " are the relative coordinates of the center of " + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "x_{s}" + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": " within " + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "x_{l}" + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": ". We drop the sample index " + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": " for simplicity. Fig. 2 shows the visualization of a sample in OCELOT. More details about the dataset including data collection and statistics can be found in the following sub-sections. The dataset is publicly available at https://lunit-io.github.io/research/publications/ocelot/." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 310, + 72, + 425, + 187 + ], + "blocks": [ + { + "bbox": [ + 310, + 72, + 425, + 187 + ], + "lines": [ + { + "bbox": [ + 310, + 72, + 425, + 187 + ], + "spans": [ + { + "bbox": [ + 310, + 72, + 425, + 187 + ], + "type": "image", + "image_path": "044543b55eab3dcae11256d4a6b2215ccf6b5d0ff53f96f269c8f0129b460040.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 355, + 189, + 380, + 200 + ], + "lines": [ + { + "bbox": [ + 355, + 189, + 380, + 200 + ], + "spans": [ + { + "bbox": [ + 355, + 189, + 380, + 200 + ], + "type": "inline_equation", + "content": "(x_{l},y_{l}^{t})" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 206, + 545, + 294 + ], + "lines": [ + { + "bbox": [ + 304, + 206, + 545, + 294 + ], + "spans": [ + { + "bbox": [ + 304, + 206, + 545, + 294 + ], + "type": "text", + "content": "Figure 2. A sample from the OCELOT dataset. Each sample of the dataset consists of two input patches and the corresponding annotations. Left shows the large FoV patch " + }, + { + "bbox": [ + 304, + 206, + 545, + 294 + ], + "type": "inline_equation", + "content": "x_{l}" + }, + { + "bbox": [ + 304, + 206, + 545, + 294 + ], + "type": "text", + "content": " with tissue segmentation annotation " + }, + { + "bbox": [ + 304, + 206, + 545, + 294 + ], + "type": "inline_equation", + "content": "y_{l}^{t}" + }, + { + "bbox": [ + 304, + 206, + 545, + 294 + ], + "type": "text", + "content": ", where green denotes the cancer area. Right shows the small FoV patch " + }, + { + "bbox": [ + 304, + 206, + 545, + 294 + ], + "type": "inline_equation", + "content": "x_{s}" + }, + { + "bbox": [ + 304, + 206, + 545, + 294 + ], + "type": "text", + "content": " with cell point annotation " + }, + { + "bbox": [ + 304, + 206, + 545, + 294 + ], + "type": "inline_equation", + "content": "y_{s}^{c}" + }, + { + "bbox": [ + 304, + 206, + 545, + 294 + ], + "type": "text", + "content": ", where blue and yellow dots denote tumor and background cells, respectively. The red box indicates the size and location of the " + }, + { + "bbox": [ + 304, + 206, + 545, + 294 + ], + "type": "inline_equation", + "content": "x_{s}" + }, + { + "bbox": [ + 304, + 206, + 545, + 294 + ], + "type": "text", + "content": " with respect to the " + }, + { + "bbox": [ + 304, + 206, + 545, + 294 + ], + "type": "inline_equation", + "content": "x_{l}" + }, + { + "bbox": [ + 304, + 206, + 545, + 294 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 427, + 72, + 543, + 187 + ], + "blocks": [ + { + "bbox": [ + 427, + 72, + 543, + 187 + ], + "lines": [ + { + "bbox": [ + 427, + 72, + 543, + 187 + ], + "spans": [ + { + "bbox": [ + 427, + 72, + 543, + 187 + ], + "type": "image", + "image_path": "a3eec8298136c7b2992ab60c079c30314fdb1beb34c7ee0d51adf921b4119281.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 473, + 189, + 498, + 199 + ], + "lines": [ + { + "bbox": [ + 473, + 189, + 498, + 199 + ], + "spans": [ + { + "bbox": [ + 473, + 189, + 498, + 199 + ], + "type": "inline_equation", + "content": "(x_{s},y_{s}^{c})" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 309, + 401, + 320 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 309, + 401, + 320 + ], + "spans": [ + { + "bbox": [ + 306, + 309, + 401, + 320 + ], + "type": "text", + "content": "3.1. Data Collection" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 329, + 545, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 329, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 329, + 545, + 437 + ], + "type": "text", + "content": "We collect 306 TCGA [21] WSIs from a total of 6 different organs: kidney, head-and-neck, prostate, stomach, endometrium, and bladder. From each of the WSIs, we select 1 to 3 large Regions of Interest (ROIs) for the tissue segmentation task. Finally, for the cell detection task, we randomly choose a smaller ROI that is fully contained within the larger tissue ROI. As a result, OCELOT includes 673 paired patches from 6 organs. The numbers of WSIs and pairs of patches per organ are detailed in Tab. 2." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 439, + 545, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 439, + 545, + 570 + ], + "spans": [ + { + "bbox": [ + 304, + 439, + 545, + 570 + ], + "type": "text", + "content": "Some natural image datasets, such as ImageNet [12] or Pascal VOC [14], include thousands of annotated images. However, annotating histopathology images is more challenging and expensive due to the scarcity of expert pathologists [27]. Furthermore, acquiring dense annotations for cell detection and tissue segmentation is especially time-demanding compared to higher-level tasks such as image classification. Nonetheless, in Tab. 1, we observe that OCELOT is roughly double the size of the recent TIGER dataset with respect to the annotated tissue area and the number of annotated cells." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": "Patch configuration. Cell detection tasks benefit from fine-grained spatial information to better capture detailed cell properties (e.g. border, shape, color, and opacity). In contrast, tissue segmentation requires a larger context to enable a better understanding of the overall structural information. Therefore, we define the FoV sizes of " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "x_{s}" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " (cell detection) and " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "x_{l}" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " (tissue segmentation) as " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "4096 \\times 4096" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " pixels, respectively, at a resolution of 0.2 Micronsper-Pixel (MPP). Finally, the large FoV patches and tissue annotations " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "(x_{l}, y_{l}^{t})" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " are down-sampled by a factor of 4, resulting in a size of " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " pixels." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23904" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 77, + 70, + 257, + 191 + ], + "blocks": [ + { + "bbox": [ + 77, + 70, + 257, + 191 + ], + "lines": [ + { + "bbox": [ + 77, + 70, + 257, + 191 + ], + "spans": [ + { + "bbox": [ + 77, + 70, + 257, + 191 + ], + "type": "table", + "html": "
Organs# Slides# Patch Pairs
TrainValTestTrainValTest
Kidney4815181254141
Head-neck135627910
Prostate261210501716
Stomach1565361212
Endometrium381313862925
Bladder351414822926
Total1756566406137130
", + "image_path": "ab63a2ffa026aff9c33966182be63ae809e31045b6930c57f34565e5c8513985.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 75, + 194, + 258, + 205 + ], + "lines": [ + { + "bbox": [ + 75, + 194, + 258, + 205 + ], + "spans": [ + { + "bbox": [ + 75, + 194, + 258, + 205 + ], + "type": "text", + "content": "Table 2. Dataset size per organ and data subset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 217, + 289, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 217, + 289, + 420 + ], + "spans": [ + { + "bbox": [ + 46, + 217, + 289, + 420 + ], + "type": "text", + "content": "Annotation. All cell-tissue pairs of patches are annotated by board-certified pathologists. Cells are labeled as points, with associated 2D coordinates and class labels. We denote the annotations in a given cell-level patch, " + }, + { + "bbox": [ + 46, + 217, + 289, + 420 + ], + "type": "inline_equation", + "content": "x_{s}" + }, + { + "bbox": [ + 46, + 217, + 289, + 420 + ], + "type": "text", + "content": ", as " + }, + { + "bbox": [ + 46, + 217, + 289, + 420 + ], + "type": "inline_equation", + "content": "y_{s}^{c}" + }, + { + "bbox": [ + 46, + 217, + 289, + 420 + ], + "type": "text", + "content": ", and consider two classes: Tumor Cell (TC) and Background Cell " + }, + { + "bbox": [ + 46, + 217, + 289, + 420 + ], + "type": "inline_equation", + "content": "(BC)^{1}" + }, + { + "bbox": [ + 46, + 217, + 289, + 420 + ], + "type": "text", + "content": ". TC and BC class ratios are " + }, + { + "bbox": [ + 46, + 217, + 289, + 420 + ], + "type": "inline_equation", + "content": "35.01\\%" + }, + { + "bbox": [ + 46, + 217, + 289, + 420 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 217, + 289, + 420 + ], + "type": "inline_equation", + "content": "64.99\\%" + }, + { + "bbox": [ + 46, + 217, + 289, + 420 + ], + "type": "text", + "content": ", respectively. Regarding the tissue patches, " + }, + { + "bbox": [ + 46, + 217, + 289, + 420 + ], + "type": "inline_equation", + "content": "x_{l}" + }, + { + "bbox": [ + 46, + 217, + 289, + 420 + ], + "type": "text", + "content": ", pathologists annotate the pixel-wise segmentation maps " + }, + { + "bbox": [ + 46, + 217, + 289, + 420 + ], + "type": "inline_equation", + "content": "y_{l}^{t}" + }, + { + "bbox": [ + 46, + 217, + 289, + 420 + ], + "type": "text", + "content": " with either Cancer Area (CA) or Background (BG) labels. A minority of pixels where the tissue class was uncertain were labeled as Unknown (UNK). BG, CA, and UNK class ratios are " + }, + { + "bbox": [ + 46, + 217, + 289, + 420 + ], + "type": "inline_equation", + "content": "55.77\\%" + }, + { + "bbox": [ + 46, + 217, + 289, + 420 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 217, + 289, + 420 + ], + "type": "inline_equation", + "content": "40.17\\%" + }, + { + "bbox": [ + 46, + 217, + 289, + 420 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 217, + 289, + 420 + ], + "type": "inline_equation", + "content": "4.06\\%" + }, + { + "bbox": [ + 46, + 217, + 289, + 420 + ], + "type": "text", + "content": ", respectively. The amount of annotated cells and tissue pixels, per data split, can be found in the supplementary material. The detection of TC and BC has clinical relevance. For example, tumor purity [5], computed as the tumor/non-tumor cell ratio in a WSI, has a correlation with cancer prognosis [16, 31, 48]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 425, + 287, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 425, + 287, + 496 + ], + "spans": [ + { + "bbox": [ + 46, + 425, + 287, + 496 + ], + "type": "text", + "content": "Dataset splits. The dataset is divided into three subsets: training, validation, and test, following a 6:2:2 ratio. To prevent information leaking among the data subsets, we randomly split the dataset per WSI, so that different patches from the same WSI are not included in multiple subsets. We maintain consistent cancer-type ratios in each subset." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 507, + 160, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 507, + 160, + 521 + ], + "spans": [ + { + "bbox": [ + 47, + 507, + 160, + 521 + ], + "type": "text", + "content": "4. Empirical Analysis" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 528, + 287, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 528, + 287, + 598 + ], + "spans": [ + { + "bbox": [ + 46, + 528, + 287, + 598 + ], + "type": "text", + "content": "The motivation for considering the cell-tissue relationships for the development of cell detection models stem from the biological and hierarchical arrangement of cells and tissues. These insights are further corroborated by two main empirical observations described in the following Sec. 4.1 and Sec. 4.2." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 608, + 273, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 608, + 273, + 620 + ], + "spans": [ + { + "bbox": [ + 47, + 608, + 273, + 620 + ], + "type": "text", + "content": "4.1. Interrelation between cell and tissue classes" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 626, + 287, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 626, + 287, + 686 + ], + "spans": [ + { + "bbox": [ + 46, + 626, + 287, + 686 + ], + "type": "text", + "content": "We empirically observe the interrelation between specific cell and tissue classes by counting the amount of each annotated cell type within each tissue region, as observed in Tab. 3. Indeed, we verify that in OCELOT, around " + }, + { + "bbox": [ + 46, + 626, + 287, + 686 + ], + "type": "inline_equation", + "content": "93\\%" + }, + { + "bbox": [ + 46, + 626, + 287, + 686 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 46, + 626, + 287, + 686 + ], + "type": "inline_equation", + "content": "TC" + }, + { + "bbox": [ + 46, + 626, + 287, + 686 + ], + "type": "text", + "content": " are located within " + }, + { + "bbox": [ + 46, + 626, + 287, + 686 + ], + "type": "inline_equation", + "content": "CA" + }, + { + "bbox": [ + 46, + 626, + 287, + 686 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 626, + 287, + 686 + ], + "type": "inline_equation", + "content": "85\\%" + }, + { + "bbox": [ + 46, + 626, + 287, + 686 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 46, + 626, + 287, + 686 + ], + "type": "inline_equation", + "content": "BC" + }, + { + "bbox": [ + 46, + 626, + 287, + 686 + ], + "type": "text", + "content": " are found outside" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 331, + 71, + 416, + 133 + ], + "blocks": [ + { + "bbox": [ + 331, + 71, + 416, + 133 + ], + "lines": [ + { + "bbox": [ + 331, + 71, + 416, + 133 + ], + "spans": [ + { + "bbox": [ + 331, + 71, + 416, + 133 + ], + "type": "image", + "image_path": "27dace7f68fb2fca893cbdbb3f21d718329e026e7910dd3b5dde9bfa519f75cb.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 348, + 135, + 399, + 145 + ], + "lines": [ + { + "bbox": [ + 348, + 135, + 399, + 145 + ], + "spans": [ + { + "bbox": [ + 348, + 135, + 399, + 145 + ], + "type": "text", + "content": "(a) OCELOT" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 435, + 72, + 520, + 133 + ], + "blocks": [ + { + "bbox": [ + 435, + 72, + 520, + 133 + ], + "lines": [ + { + "bbox": [ + 435, + 72, + 520, + 133 + ], + "spans": [ + { + "bbox": [ + 435, + 72, + 520, + 133 + ], + "type": "image", + "image_path": "4b308743775d930a5836697e1e3b36a52a7f59ff4d754c4f42ad13db63504812.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 457, + 135, + 498, + 145 + ], + "lines": [ + { + "bbox": [ + 457, + 135, + 498, + 145 + ], + "spans": [ + { + "bbox": [ + 457, + 135, + 498, + 145 + ], + "type": "text", + "content": "(b) TIGER" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 334, + 212, + 414, + 250 + ], + "blocks": [ + { + "bbox": [ + 305, + 156, + 547, + 200 + ], + "lines": [ + { + "bbox": [ + 305, + 156, + 547, + 200 + ], + "spans": [ + { + "bbox": [ + 305, + 156, + 547, + 200 + ], + "type": "text", + "content": "Table 3. Cell counts based on the tissue class. Each value stands for the number of cells located inside the tissue area. " + }, + { + "bbox": [ + 305, + 156, + 547, + 200 + ], + "type": "inline_equation", + "content": "TC" + }, + { + "bbox": [ + 305, + 156, + 547, + 200 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 156, + 547, + 200 + ], + "type": "inline_equation", + "content": "BC" + }, + { + "bbox": [ + 305, + 156, + 547, + 200 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 156, + 547, + 200 + ], + "type": "inline_equation", + "content": "LC" + }, + { + "bbox": [ + 305, + 156, + 547, + 200 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 156, + 547, + 200 + ], + "type": "inline_equation", + "content": "CA" + }, + { + "bbox": [ + 305, + 156, + 547, + 200 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 305, + 156, + 547, + 200 + ], + "type": "inline_equation", + "content": "ST" + }, + { + "bbox": [ + 305, + 156, + 547, + 200 + ], + "type": "text", + "content": " stand for Tumor Cell, Background Cell, Lymphocyte Cell, Cancer Area tissue, and Stroma tissue, respectively." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 334, + 212, + 414, + 250 + ], + "lines": [ + { + "bbox": [ + 334, + 212, + 414, + 250 + ], + "spans": [ + { + "bbox": [ + 334, + 212, + 414, + 250 + ], + "type": "image", + "image_path": "51b4047f3b0b71fc24b1ddc92ae16e247a35ce760e3b4de01cae095d88e5b070.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 346, + 251, + 402, + 262 + ], + "lines": [ + { + "bbox": [ + 346, + 251, + 402, + 262 + ], + "spans": [ + { + "bbox": [ + 346, + 251, + 402, + 262 + ], + "type": "text", + "content": "(a) OCELOT" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 272, + 546, + 295 + ], + "lines": [ + { + "bbox": [ + 305, + 272, + 546, + 295 + ], + "spans": [ + { + "bbox": [ + 305, + 272, + 546, + 295 + ], + "type": "text", + "content": "Table 4. Pixel ratio among tissue classes. " + }, + { + "bbox": [ + 305, + 272, + 546, + 295 + ], + "type": "inline_equation", + "content": "CA" + }, + { + "bbox": [ + 305, + 272, + 546, + 295 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 305, + 272, + 546, + 295 + ], + "type": "inline_equation", + "content": "ST" + }, + { + "bbox": [ + 305, + 272, + 546, + 295 + ], + "type": "text", + "content": " stand for Cancer Area tissue, and Stroma tissue, respectively." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 436, + 213, + 516, + 249 + ], + "blocks": [ + { + "bbox": [ + 436, + 213, + 516, + 249 + ], + "lines": [ + { + "bbox": [ + 436, + 213, + 516, + 249 + ], + "spans": [ + { + "bbox": [ + 436, + 213, + 516, + 249 + ], + "type": "image", + "image_path": "b7943a72b1d8be75727bb468f4fb0b766f3779c56b5f3e7b0523926a3a2965ef.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 454, + 251, + 499, + 262 + ], + "lines": [ + { + "bbox": [ + 454, + 251, + 499, + 262 + ], + "spans": [ + { + "bbox": [ + 454, + 251, + 499, + 262 + ], + "type": "text", + "content": "(b) TIGER" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 304, + 545, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 304, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 304, + 304, + 545, + 364 + ], + "type": "text", + "content": "of the CA tissue. Note that CA is not the majority tissue class (Tab. 4), therefore, we conclude that there is, in fact, a relationship between the cell and tissue classes. We observe a similar phenomenon in the TIGER dataset when considering the " + }, + { + "bbox": [ + 304, + 304, + 545, + 364 + ], + "type": "inline_equation", + "content": "LC" + }, + { + "bbox": [ + 304, + 304, + 545, + 364 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 304, + 545, + 364 + ], + "type": "inline_equation", + "content": "ST" + }, + { + "bbox": [ + 304, + 304, + 545, + 364 + ], + "type": "text", + "content": " classes (Tab. 3 and Tab. 4)." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 365, + 546, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 365, + 546, + 496 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 546, + 496 + ], + "type": "text", + "content": "In practice, pathologists classify cells by taking into account such interrelationships, since it is difficult to classify isolated cells, without considering the larger context of the tissues. As depicted in Fig. 1, pathologists first need to visualize the structure of the tissues at a larger FoV. Then, they zoom in and consider the previously observed context along with the fine-grained details of each individual cell and nearby neighborhood, thus considering the cell-tissue dependencies. Inspired by the behavior of pathologists, we expect a cell detection model to also benefit from understanding the tissue structure from a broader viewpoint." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 503, + 459, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 503, + 459, + 517 + ], + "spans": [ + { + "bbox": [ + 306, + 503, + 459, + 517 + ], + "type": "text", + "content": "4.2. Tissue-label Leaking Model" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 522, + 545, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 545, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 545, + 700 + ], + "type": "text", + "content": "In the previous section, we observe a strong relationship between certain cell and tissue classes. To further validate the hypothesis that a cell detection model can leverage information from the tissue structure, we design an exploratory experiment where the tissue annotation is provided as an extra input to a cell detection model. To this end, we first crop the corresponding cell patch region from the tissue annotation " + }, + { + "bbox": [ + 304, + 522, + 545, + 700 + ], + "type": "inline_equation", + "content": "y_{l}^{t}" + }, + { + "bbox": [ + 304, + 522, + 545, + 700 + ], + "type": "text", + "content": ", and upsample it to match the size of the cell patch, " + }, + { + "bbox": [ + 304, + 522, + 545, + 700 + ], + "type": "inline_equation", + "content": "x_{s}" + }, + { + "bbox": [ + 304, + 522, + 545, + 700 + ], + "type": "text", + "content": "; the cropped tissue annotation is denoted as " + }, + { + "bbox": [ + 304, + 522, + 545, + 700 + ], + "type": "inline_equation", + "content": "y_{s}^{t}" + }, + { + "bbox": [ + 304, + 522, + 545, + 700 + ], + "type": "text", + "content": ". Finally, we concatenate " + }, + { + "bbox": [ + 304, + 522, + 545, + 700 + ], + "type": "inline_equation", + "content": "x_{s}" + }, + { + "bbox": [ + 304, + 522, + 545, + 700 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 522, + 545, + 700 + ], + "type": "inline_equation", + "content": "y_{s}^{t}" + }, + { + "bbox": [ + 304, + 522, + 545, + 700 + ], + "type": "text", + "content": " at the channels dimension and use this data to train a cell detection model. We denote this model as a Tissue-label leaking model and illustrate it in Fig. 3. Note that this model is not appropriate for real-world scenarios as the tissue labels are unknown at inference time, and is explored for the purpose of empirical analysis." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 318, + 701, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 701, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 318, + 701, + 545, + 713 + ], + "type": "text", + "content": "When we compare the performance of the tissue-label" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "type": "inline_equation", + "content": "{}^{1}{BC}" + }, + { + "bbox": [ + 47, + 693, + 287, + 713 + ], + "type": "text", + "content": " includes any of the following cell categories: lymphocyte, macrophage, fibroblast, endothelial, or other remaining cell types." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23905" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 71, + 269, + 247 + ], + "blocks": [ + { + "bbox": [ + 62, + 71, + 269, + 247 + ], + "lines": [ + { + "bbox": [ + 62, + 71, + 269, + 247 + ], + "spans": [ + { + "bbox": [ + 62, + 71, + 269, + 247 + ], + "type": "image", + "image_path": "750b945d175481a119d51943a76a362517ec62889727e91356ab12eb8440775e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 255, + 288, + 310 + ], + "lines": [ + { + "bbox": [ + 46, + 255, + 288, + 310 + ], + "spans": [ + { + "bbox": [ + 46, + 255, + 288, + 310 + ], + "type": "text", + "content": "Figure 3. Tissue label leaking model. This model receives the cell patch " + }, + { + "bbox": [ + 46, + 255, + 288, + 310 + ], + "type": "inline_equation", + "content": "x_{s}" + }, + { + "bbox": [ + 46, + 255, + 288, + 310 + ], + "type": "text", + "content": " along with the corresponding tissue labels as input. The region corresponding to the cell patch in the tissue patch annotation is cropped, upsampled, and concatenated to the cell patch. " + }, + { + "bbox": [ + 46, + 255, + 288, + 310 + ], + "type": "inline_equation", + "content": "\\oplus" + }, + { + "bbox": [ + 46, + 255, + 288, + 310 + ], + "type": "text", + "content": " denotes channel-wise concatenation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 323, + 287, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 323, + 287, + 420 + ], + "spans": [ + { + "bbox": [ + 46, + 323, + 287, + 420 + ], + "type": "text", + "content": "leaking model with the standard cell detection model on the OCELOT dataset, we observe a significant improvement in terms of mean F1-score" + }, + { + "bbox": [ + 46, + 323, + 287, + 420 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 46, + 323, + 287, + 420 + ], + "type": "text", + "content": " performance of " + }, + { + "bbox": [ + 46, + 323, + 287, + 420 + ], + "type": "inline_equation", + "content": "+7.69" + }, + { + "bbox": [ + 46, + 323, + 287, + 420 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 323, + 287, + 420 + ], + "type": "inline_equation", + "content": "+9.76" + }, + { + "bbox": [ + 46, + 323, + 287, + 420 + ], + "type": "text", + "content": " in the validation and test sets, respectively. Detailed results can be found in Tab. 5. Taking these results into consideration, we conclude that there is significant room to improve the cell detection model, which can be achieved by combining the tasks of cell detection and tissue segmentation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 431, + 103, + 442 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 431, + 103, + 442 + ], + "spans": [ + { + "bbox": [ + 47, + 431, + 103, + 442 + ], + "type": "text", + "content": "5. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 451, + 287, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 451, + 287, + 559 + ], + "spans": [ + { + "bbox": [ + 46, + 451, + 287, + 559 + ], + "type": "text", + "content": "In this section, we propose to utilize cell-tissue relationships through multi-task learning. First, we propose a set of approaches inspired by the tissue-label leaking model described in Sec. 4.2. These models replace the annotated tissue labels with predictions from an auxiliary tissue segmentation branch. Second, we design a bi-directional information-sharing approach that shares features in both tissue-to-cell and cell-to-tissue directions. The proposed approaches are described in Sec. 5.2 and Sec. 5.3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 567, + 127, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 567, + 127, + 578 + ], + "spans": [ + { + "bbox": [ + 47, + 567, + 127, + 578 + ], + "type": "text", + "content": "5.1. Preliminary" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 585, + 287, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 585, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 46, + 585, + 287, + 657 + ], + "type": "text", + "content": "To deal with cell-tissue sample pairs, i.e., " + }, + { + "bbox": [ + 46, + 585, + 287, + 657 + ], + "type": "inline_equation", + "content": "(x_{s},y_{s}^{c})" + }, + { + "bbox": [ + 46, + 585, + 287, + 657 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 585, + 287, + 657 + ], + "type": "inline_equation", + "content": "(x_{l},y_{l}^{t})" + }, + { + "bbox": [ + 46, + 585, + 287, + 657 + ], + "type": "text", + "content": ", we build a dual-branch architecture containing separate networks for the cell and tissue tasks. Similarly to [41], we define cell detection as a segmentation task. Specifically, the cell labels are provided as a segmentation map by drawing a fixed-radius circle centered on each cell point an" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 309, + 72, + 545, + 217 + ], + "blocks": [ + { + "bbox": [ + 309, + 72, + 545, + 217 + ], + "lines": [ + { + "bbox": [ + 309, + 72, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 309, + 72, + 545, + 217 + ], + "type": "image", + "image_path": "69d0e324baae725b04e3cb68e530637fb67360348e2ae6a44b463074665dca7c.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 228, + 546, + 261 + ], + "lines": [ + { + "bbox": [ + 305, + 228, + 546, + 261 + ], + "spans": [ + { + "bbox": [ + 305, + 228, + 546, + 261 + ], + "type": "text", + "content": "Figure 4. Tissue-prediction injection model injects the tissue segmentation prediction into 1 out of 4 locations of the cell branch: (a) input, (b) after encoder, (c) after ASPP, and (d) after decoder." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 274, + 546, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 274, + 546, + 406 + ], + "spans": [ + { + "bbox": [ + 304, + 274, + 546, + 406 + ], + "type": "text", + "content": "notation and filled with the corresponding class label. At inference time, we find local peaks within the predicted cell probability maps and output them as point predictions. More details are provided in the supplementary material. Treating cell detection as a segmentation task enables us to use the same architecture for both cell and tissue branches, which largely simplifies the training and tuning of the model and reduces the range of possible decisions, such as neural network architecture, or hyper-parameters. We use DeepLabV3+ [10] as a base architecture for both branches and single-task models." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 415, + 492, + 428 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 415, + 492, + 428 + ], + "spans": [ + { + "bbox": [ + 306, + 415, + 492, + 428 + ], + "type": "text", + "content": "5.2. Tissue-prediction Injection Models" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 434, + 545, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 434, + 545, + 626 + ], + "spans": [ + { + "bbox": [ + 304, + 434, + 545, + 626 + ], + "type": "text", + "content": "These models are a simple and practical extension of the tissue-label leaking model, where we inject the predicted tissue probabilities into the cell detection branch instead of leaking the tissue labels. We consider only one injection point in the cell detection branch, but explore four possible alternatives: (a) at the input (Pred-to-input), (b) after the encoder (Pred-to-inter-1), (c) after the ASPP module (Pred-to-inter-2), and (d) after the decoder (Pred-to-output). Fig. 4 depicts this family of models, denominated as Tissue-prediction injection. Since the tissue and cell patches represent different regions, we need to align the content between the tissue and cell feature maps before concatenation. Therefore, we crop the cell corresponding region from the tissue predictions, upsample them, and, finally, concatenate them in the channel dimensions of the feature maps of the cell branch, as illustrated in Fig. 5a." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 634, + 492, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 634, + 492, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 634, + 492, + 647 + ], + "type": "text", + "content": "5.3. Cell-Tissue Feature Sharing Model" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "Tissue-prediction injection models share the tissue prediction in a single location and direction, i.e., tissue-to-cell. To enable a more diverse and flexible cell-tissue information flow, we also explore bi-directional feature map sharing from cell-to-tissue (Fig. 5, left) and tissue-to-cell (Fig. 5," + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 46, + 664, + 287, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 664, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 664, + 287, + 713 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 46, + 664, + 287, + 713 + ], + "type": "text", + "content": "True positive (TP), false positive (FP), and false negative (FN) counts are determined following [41]. If a detected cell is within a valid distance " + }, + { + "bbox": [ + 46, + 664, + 287, + 713 + ], + "type": "inline_equation", + "content": "(\\approx 3\\mu m)" + }, + { + "bbox": [ + 46, + 664, + 287, + 713 + ], + "type": "text", + "content": " from an annotated cell and the cell class matches, it is counted as a TP, otherwise an FP. If an annotated cell is not detected, it is counted as an FN. Then, the mean F1 score across classes is computed." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "23906" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 160, + 133 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 160, + 133 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 160, + 133 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 160, + 133 + ], + "type": "image", + "image_path": "6c2472a0393278468d9b7427a30da9677432ee38f69c90d1706abdcaf23c9418.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 80, + 135, + 136, + 144 + ], + "lines": [ + { + "bbox": [ + 80, + 135, + 136, + 144 + ], + "spans": [ + { + "bbox": [ + 80, + 135, + 136, + 144 + ], + "type": "text", + "content": "(a) Tissue to Cell" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 170, + 70, + 282, + 133 + ], + "blocks": [ + { + "bbox": [ + 170, + 70, + 282, + 133 + ], + "lines": [ + { + "bbox": [ + 170, + 70, + 282, + 133 + ], + "spans": [ + { + "bbox": [ + 170, + 70, + 282, + 133 + ], + "type": "image", + "image_path": "0d45b9b009457d65f00de40d4c95623489bbb49bf540608eac69a9b3dfa5797a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 198, + 135, + 255, + 145 + ], + "lines": [ + { + "bbox": [ + 198, + 135, + 255, + 145 + ], + "spans": [ + { + "bbox": [ + 198, + 135, + 255, + 145 + ], + "type": "text", + "content": "(b) Cell to Tissue" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 50, + 272, + 286, + 404 + ], + "blocks": [ + { + "bbox": [ + 46, + 155, + 289, + 254 + ], + "lines": [ + { + "bbox": [ + 46, + 155, + 289, + 254 + ], + "spans": [ + { + "bbox": [ + 46, + 155, + 289, + 254 + ], + "type": "text", + "content": "Figure 5. Information is shared between cell and tissue branches via channel-wise concatenation preceded by a shallow convolutional layer with " + }, + { + "bbox": [ + 46, + 155, + 289, + 254 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 46, + 155, + 289, + 254 + ], + "type": "text", + "content": " kernel size. Cropping and upsampling (in Fig. 5a) or downsampling and zero-padding (in Fig. 5b) is applied to match the patch sizes and pixel-alignment between two feature maps from different FoVs. The cell and tissue feature maps are represented in orange and blue, respectively. The red contour denotes the cell patch-associated region in the tissue patch, and the gray regions represent zero padding." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 50, + 272, + 286, + 404 + ], + "lines": [ + { + "bbox": [ + 50, + 272, + 286, + 404 + ], + "spans": [ + { + "bbox": [ + 50, + 272, + 286, + 404 + ], + "type": "image", + "image_path": "708c0a899a6963e3a3c6d719e5cfda1e936012867f1aa85c7165d6ca25edb7c1.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 418, + 287, + 475 + ], + "lines": [ + { + "bbox": [ + 46, + 418, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 46, + 418, + 287, + 475 + ], + "type": "text", + "content": "Figure 6. Cell-Tissue Feature Sharing Model has two branches for tissue segmentation and cell detection. Information exchange occurs multiple times between the two branches, indicated by the red vertical arrows. Details regarding the information exchange procedures are described in Fig. 5." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 487, + 287, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 487, + 287, + 606 + ], + "spans": [ + { + "bbox": [ + 46, + 487, + 287, + 606 + ], + "type": "text", + "content": "right). Considering these two operations, we conduct an architecture search procedure to find the optimal feature map sharing configuration between both branches. To limit the search space, we constrain it to only 3 positions in the architecture: after the encoder, after the ASPP module, or after the decoder. Furthermore, we also exclusively allow the branches to inject feature maps at the same depth or position. Finally, we consider only the best-performing model among the " + }, + { + "bbox": [ + 46, + 487, + 287, + 606 + ], + "type": "inline_equation", + "content": "4^{3}" + }, + { + "bbox": [ + 46, + 487, + 287, + 606 + ], + "type": "text", + "content": " candidates, which is presented in Fig. 6. We name this model as cell-tissue feature sharing." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 620, + 190, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 620, + 190, + 634 + ], + "spans": [ + { + "bbox": [ + 47, + 620, + 190, + 634 + ], + "type": "text", + "content": "6. Experiments and Results" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "type": "text", + "content": "We validate the hypothesis that incorporating cell-tissue relationships within a cell detection model is beneficial by evaluating the proposed models on OCELOT and 2 other datasets. First, we describe the additional datasets and implementation details. Then, we show how the proposed multi-task learning methods can improve the performance" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": "of a cell detection task. Lastly, we conduct an ablation study to investigate how important using large FoV patches " + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "\\{x_{l}\\}" + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": " and corresponding tissue segmentation labels " + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "\\{y_{l}^{t}\\}" + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": " are in enhancing cell-tissue relationships." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 128, + 399, + 139 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 128, + 399, + 139 + ], + "spans": [ + { + "bbox": [ + 306, + 128, + 399, + 139 + ], + "type": "text", + "content": "6.1. Other Datasets" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 146, + 545, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 146, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 304, + 146, + 545, + 373 + ], + "type": "text", + "content": "TIGER. As mentioned in Sec. 2.1, TIGER is a public dataset, based on H&E stained images, which includes both cell and tissue annotations in overlapping patches. However, the patch sizes and amount of overlap between cell and tissue annotations highly vary among samples. Therefore, a preprocessing step is necessary to generate paired cell-tissue patches with consistent FoVs. After pre-processing, we obtain 9,888 paired patches of size " + }, + { + "bbox": [ + 304, + 146, + 545, + 373 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 304, + 146, + 545, + 373 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 146, + 545, + 373 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 304, + 146, + 545, + 373 + ], + "type": "text", + "content": " pixels for the tissue segmentation and cell detection tasks, respectively, with an MPP of approximately " + }, + { + "bbox": [ + 304, + 146, + 545, + 373 + ], + "type": "inline_equation", + "content": "0.5\\mu m / px" + }, + { + "bbox": [ + 304, + 146, + 545, + 373 + ], + "type": "text", + "content": ". Most samples in TIGER include small regions with annotated cells, making it necessary to extract smaller patches. This explains why the absolute number of patches is larger than in OCELOT or CARP, although the actual amount of annotated cells and tissue area are smaller than the latter datasets, as shown in Tab. 1. In TIGER, one cell class (lymphocyte cells) and seven tissue classes are annotated. More details about the pre-processing and this dataset are provided in the supplementary material." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 376, + 546, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 376, + 546, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 376, + 546, + 544 + ], + "type": "text", + "content": "CARP. This is an internal lung cancer dataset of PD-L1 IHC-stained WSIs, containing 6,480 paired patches extracted from 1,012 WSIs. The patch sizes, resolution, annotation protocol, and general configuration are similar to OCELOT, as described in Sec. 3.1. Two cell classes are annotated: PD-L1 positive tumor cells " + }, + { + "bbox": [ + 304, + 376, + 546, + 544 + ], + "type": "inline_equation", + "content": "(TC+)" + }, + { + "bbox": [ + 304, + 376, + 546, + 544 + ], + "type": "text", + "content": ", and PD-L1 negative tumor cells " + }, + { + "bbox": [ + 304, + 376, + 546, + 544 + ], + "type": "inline_equation", + "content": "(TC-" + }, + { + "bbox": [ + 304, + 376, + 546, + 544 + ], + "type": "text", + "content": " ). The tissue classes are the same as in OCELOT (BG, CA, and UNK). CARP is a real-world and large-scale dataset, with 809.1K annotated cells and 4.1080 mm" + }, + { + "bbox": [ + 304, + 376, + 546, + 544 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 304, + 376, + 546, + 544 + ], + "type": "text", + "content": " tissue area, which is approximately 10 and 20 times more than OCELOT and TIGER, respectively. Moreover, the stain type is different from OCELOT and TIGER, allowing us to validate the effectiveness of the cell-tissue relationships across different staining methods." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 552, + 437, + 564 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 552, + 437, + 564 + ], + "spans": [ + { + "bbox": [ + 306, + 552, + 437, + 564 + ], + "type": "text", + "content": "6.2. Implementation details" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": "As previously mentioned in Sec. 5.1, both cell and tissue branches are based on DeepLabV3+ [10], with a ResNet-34 [18] encoder. The models are trained for 300, 150, and 100 epochs in OCELOT, TIGER, and CARP datasets, respectively. We use the Adam optimizer [24], and, for each experiment, we tune the learning rate within a " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "[5 \\times 10^{-5}, 2 \\times 10^{-3}]" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": " range. The model at the epoch with the best validation set performance is chosen and used for evaluation on the test set. All experiments are repeated 5 times, and we report the mean and " + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": " confidence interval of the performance metrics. For more implementation details, please refer to the supplementary material." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23907" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 112, + 70, + 479, + 202 + ], + "blocks": [ + { + "bbox": [ + 112, + 70, + 479, + 202 + ], + "lines": [ + { + "bbox": [ + 112, + 70, + 479, + 202 + ], + "spans": [ + { + "bbox": [ + 112, + 70, + 479, + 202 + ], + "type": "table", + "html": "
MethodOCELOTTIGERCARP
ValTestValTestValTest
Cell-only68.87±1.7664.44±1.8263.89±1.3953.82±1.2378.48±0.6970.96±1.47
Pred-to-input73.36±0.5969.65±3.9366.00±2.0053.29±1.3079.46±0.7972.98±0.82
Pred-to-inter-172.74±0.5070.54±2.2066.19±1.0255.87±1.7879.74±0.8073.05±0.69
Pred-to-inter-272.68±1.5871.23±0.9665.43±1.1454.75±2.2579.87±0.7873.14±1.53
Pred-to-output66.85±5.6265.05±3.7263.02±0.1653.32±0.4278.92±0.6072.61±0.95
Feature-sharing72.30±0.7368.91±2.5265.64±1.0755.10±2.1879.38±0.7473.00±0.33
Tissue-label leaking76.56±0.8074.20±0.9169.71±0.6161.66±1.1680.13±1.0472.97±0.49
", + "image_path": "9220e1754061a1d4841b3cea96d5e5d4e00d2235917aa5f5102af2cb0b249a23.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 82, + 244, + 513, + 373 + ], + "blocks": [ + { + "bbox": [ + 46, + 209, + 547, + 235 + ], + "lines": [ + { + "bbox": [ + 46, + 209, + 547, + 235 + ], + "spans": [ + { + "bbox": [ + 46, + 209, + 547, + 235 + ], + "type": "text", + "content": "Table 5. Cell detection mean F1 scores per model. With the exception of the Tissue-label Leaking model, the highest score is written in bold and the second highest score is underlined." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 82, + 244, + 513, + 373 + ], + "lines": [ + { + "bbox": [ + 82, + 244, + 513, + 373 + ], + "spans": [ + { + "bbox": [ + 82, + 244, + 513, + 373 + ], + "type": "table", + "html": "
Methodpopulation (130)kidney (41)endometrium (25)bladder (26)prostate (16)stomach (12)head-neck (10)
Cell-only64.44±1.8264.12±3.4166.88±3.7159.42±5.6165.46±1.6466.19±8.9359.24±5.04
Pred-to-input69.65±3.9363.37±3.2773.31±6.3563.61±4.5568.36±8.6169.34±7.1575.28±0.82
Pred-to-inter-170.54±2.2066.62±7.3073.05±4.1464.35±3.5970.20±2.5671.55±5.2174.50±1.52
Pred-to-inter-271.23±0.9668.94±5.6575.15±2.7064.94±2.3468.35±8.8370.29±0.8173.74±1.72
Pred-to-output65.05±3.7263.38±4.4868.21±4.1159.81±3.9664.18±8.4167.51±6.3960.88±6.77
Feature-sharing68.91±2.5264.97±5.3371.28±5.3063.21±5.3166.25±9.5269.79±2.2573.88±4.39
Tissue-label leaking74.20±0.9175.75±1.9772.71±1.1271.18±0.9374.56±1.2575.24±1.7175.78±0.46
", + "image_path": "7bde9d73b82cbee8dfc74afec089f270e976264f5024391b13e4f414a48e3e47.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 424, + 132, + 436 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 424, + 132, + 436 + ], + "spans": [ + { + "bbox": [ + 47, + 424, + 132, + 436 + ], + "type": "text", + "content": "6.3. Main Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 445, + 287, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 445, + 287, + 577 + ], + "spans": [ + { + "bbox": [ + 46, + 445, + 287, + 577 + ], + "type": "text", + "content": "The cell detection results obtained by the Cell-only baseline, the proposed approaches, and the Tissue-label leaking model are shown in Tab. 5. The Cell-only baseline is a simple cell detection model that only receives the small FoV patches as input. In other words, it only considers the cell branch in Fig. 4 and neither tissue annotations nor large FoV patches are leveraged. The Tissue-label leaking model, described in Sec. 5.2, receives the tissue annotations as input and serves as an exploratory experiment to obtain insight regarding how much the cell detection task can benefit from leveraging tissue annotation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 578, + 288, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 578, + 288, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 578, + 288, + 700 + ], + "type": "text", + "content": "From Tab. 5, we observe that all the cell-tissue multi-task learning-based approaches, except for the Pred-to-output model, outperform the Cell-only baseline across all datasets. These improvements imply that cell detection on small FoV patches benefits from learning cell-tissue interrelationships from both large FoVs and tissue annotations. We hypothesize that the reason for the low performance of the Pred-to-output model is that the injection of tissue prediction to the cell branch happens too late. Therefore, there is a lack of network capacity for fusing cell and tissue information." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 59, + 701, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 287, + 714 + ], + "type": "text", + "content": "Furthermore, in Tab. 6, we report mean F1 scores per or" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 309, + 422, + 545, + 496 + ], + "blocks": [ + { + "bbox": [ + 46, + 380, + 546, + 415 + ], + "lines": [ + { + "bbox": [ + 46, + 380, + 546, + 415 + ], + "spans": [ + { + "bbox": [ + 46, + 380, + 546, + 415 + ], + "type": "text", + "content": "Table 6. Per-organ cell detection mean F1 scores in the OCELOT test set. Population mean F1 scores are computed from all the patches in the test set. In parentheses, we indicate the number of samples of each organ subset. With the exception of the Tissue-label Leaking model, the highest score is written in bold and the second highest score is underlined." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 422, + 545, + 496 + ], + "lines": [ + { + "bbox": [ + 309, + 422, + 545, + 496 + ], + "spans": [ + { + "bbox": [ + 309, + 422, + 545, + 496 + ], + "type": "table", + "html": "
Further utilizePred-to-inter-2Feature-sharing
Tissue labelLarge FoVValTestValTest
--68.87±1.7664.45±1.8268.87±1.7664.45±1.82
✓-70.21±1.4964.99±1.7770.36±1.3466.58±2.02
-✓68.78±0.9465.56±3.6669.61±1.5966.22±1.53
✓✓72.68±1.5871.23±0.9672.30±0.7368.91±2.52
", + "image_path": "59f0770bdd69a5c03d15e4774b5a106d1444be1af784293b8657c7616af3697d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 504, + 546, + 538 + ], + "lines": [ + { + "bbox": [ + 304, + 504, + 546, + 538 + ], + "spans": [ + { + "bbox": [ + 304, + 504, + 546, + 538 + ], + "type": "text", + "content": "Table 7. Dataset ablation study. In all experiments, the model architecture is fixed and only input/output changes. The first row corresponds to the Cell-only baseline." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 543, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 543, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 543, + 545, + 628 + ], + "type": "text", + "content": "gan in the OCELOT test set. Except for the Pred-to-output, all approaches show improvement compared to Cell-only baseline in most organs. Especially, our best-performing model, Pred-to-inter-2, demonstrates superior performance over Cell-only in all 6 organs with a significant margin. Such a result shows that considering cell-tissue relationships can generally help the cell detection task across various organs." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 635, + 438, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 635, + 438, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 635, + 438, + 647 + ], + "type": "text", + "content": "6.4. Dataset Ablation Study" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": "Two data components in OCELOT contribute to better leveraging cell-tissue relationships for the task of cell detection: (1) the large FoV patches " + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "inline_equation", + "content": "x_{l}" + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": ", and (2) the corresponding tissue annotations " + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "inline_equation", + "content": "y_{l}^{t}" + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": " that allows us to have a tissue segmentation objective and a multi-task learning setting. To verify" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23908" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 78, + 68, + 517, + 285 + ], + "blocks": [ + { + "bbox": [ + 78, + 68, + 517, + 285 + ], + "lines": [ + { + "bbox": [ + 78, + 68, + 517, + 285 + ], + "spans": [ + { + "bbox": [ + 78, + 68, + 517, + 285 + ], + "type": "image", + "image_path": "a661531a0b3009d2bc0a111da695b1eb932502a0c9c63f3295fe923ab328e9d6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 290, + 547, + 314 + ], + "lines": [ + { + "bbox": [ + 46, + 290, + 547, + 314 + ], + "spans": [ + { + "bbox": [ + 46, + 290, + 547, + 314 + ], + "type": "text", + "content": "Figure 7. Qualitative comparison between the Pred-to-inter-2 and Cell-only models. Pred-to-inter-2 shows enhanced detection results aided by understanding the tissue environment from a broader context (blue: tumor cells, yellow: background cells, green: cancer area)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 323, + 288, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 323, + 288, + 430 + ], + "spans": [ + { + "bbox": [ + 46, + 323, + 288, + 430 + ], + "type": "text", + "content": "the effectiveness of each data component, we conduct an ablation study where each component is evaluated separately while keeping the model architecture fixed. First, if only the tissue segmentation label is provided without a large FoV patch, the tissue branch takes " + }, + { + "bbox": [ + 46, + 323, + 288, + 430 + ], + "type": "inline_equation", + "content": "x_{s}" + }, + { + "bbox": [ + 46, + 323, + 288, + 430 + ], + "type": "text", + "content": " as an input and it is trained with both the tissue and the cell detection objectives. Second, if the large FoV patch is given without the corresponding tissue label, the tissue branch takes " + }, + { + "bbox": [ + 46, + 323, + 288, + 430 + ], + "type": "inline_equation", + "content": "x_{l}" + }, + { + "bbox": [ + 46, + 323, + 288, + 430 + ], + "type": "text", + "content": " as an input and the model is only trained with the cell detection objective." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 431, + 288, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 431, + 288, + 649 + ], + "spans": [ + { + "bbox": [ + 46, + 431, + 288, + 649 + ], + "type": "text", + "content": "The results of the ablation study are shown in Tab. 7. We observe that each of the individual components on their own lead to improved performance. Hence, we can conclude that a large FoV patch and tissue segmentation label can contribute to better cell detection by providing a broader context and encoding the cell-tissue relationship, respectively. Moreover, when utilized simultaneously, we observe that both components work synergically, leading to an even better performance improvement. We also found that the tissue branch generates better tissue predictions when trained with a large FoV. Thus, providing accurate tissue information to the cell branch is important for boosting performance. Please, refer to the supplementary material for a detailed tissue performance comparison. Finally, this study shows evidence that the patch configuration defined in OCELOT is suitable for cell-tissue co-training, with a smaller FoV for cell detection " + }, + { + "bbox": [ + 46, + 431, + 288, + 649 + ], + "type": "inline_equation", + "content": "(x_{s},y_{s}^{c})" + }, + { + "bbox": [ + 46, + 431, + 288, + 649 + ], + "type": "text", + "content": " and a larger FoV for tissue segmentation " + }, + { + "bbox": [ + 46, + 431, + 288, + 649 + ], + "type": "inline_equation", + "content": "(x_{l},y_{l}^{t})" + }, + { + "bbox": [ + 46, + 431, + 288, + 649 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 658, + 165, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 165, + 671 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 165, + 671 + ], + "type": "text", + "content": "6.5. Qualitative Analysis" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "type": "text", + "content": "In this section, we visualize the cell predictions of the Cell-only and Pred-to-inter-2 models. We select the Pred-to-inter-2 model, as it is the overall best-performing model in" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 323, + 547, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 323, + 547, + 479 + ], + "spans": [ + { + "bbox": [ + 304, + 323, + 547, + 479 + ], + "type": "text", + "content": "our experiments. In general, when compared to other background cells, tumor cells have the following characteristics: large size and irregular shape. However, cancer is heterogeneous and this is not always the case. Indeed, in Fig. 7, most of the cells are small and have a regular round shape. Based on these appearances, and without a larger context, those cells can be easily misclassified as background cells, which is the case of the Cell-only model. On the other hand, Pred-to-inter-2 shows a more accurate prediction by correctly understanding the cancer area in large FoV regions. This implies that Pred-to-inter-2 indeed considers both the morphology of cells and the tissue context, while Cell-only relies on the cells' morphology alone." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 498, + 379, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 498, + 379, + 511 + ], + "spans": [ + { + "bbox": [ + 306, + 498, + 379, + 511 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 521, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 547, + 715 + ], + "type": "text", + "content": "In histopathology, there is a hierarchical organization and interrelationship between cells and tissues. Hence, we hypothesize that it can be leveraged to improve the performance of cell detection tasks. Throughout this paper, we have shown evidence that such cell-tissue relationships exist across multiple setups. Indeed, we observe improvements when utilizing large FoVs with corresponding tissue annotations together with cell annotations in a multi-task learning approach, using simple neural architecture designs. Although the cell detection task clearly benefits from the multitask learning framework, the improvement on the tissue segmentation task was not yet investigated, and we leave this topic as a future direction. We hope this work, especially the OCELOT dataset, initiates new and promising research directions in CPATH to further study the cell-tissue relationships and improve cell detection and tissue segmentation." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23909" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 134 + ], + "type": "text", + "content": "[1] MICCAI 2014 Grand Challenges. https://wiki.cancerimagingarchive.net/display/Public/MICCAI+2014+Grand+Challenges, 2014. [Online; accessed Nov-2022]. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 136, + 286, + 157 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 136, + 286, + 157 + ], + "spans": [ + { + "bbox": [ + 53, + 136, + 286, + 157 + ], + "type": "text", + "content": "[2] TIGER: Grand Challenge. https://tiger.grand-challenge.org, 2022. [Online; accessed Nov-2022]. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 159, + 288, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 159, + 288, + 235 + ], + "spans": [ + { + "bbox": [ + 54, + 159, + 288, + 235 + ], + "type": "text", + "content": "[3] Esther Abels, Liron Pantanowitz, Famke Aeffner, Mark D Zarella, Jeroen van der Laak, Marilyn M Bui, Venkata NP Vemuri, Anil V Parwani, Jeff Gibbs, Emmanuel Agosto-Arroyo, et al. Computational pathology definitions, best practices, and recommendations for regulatory guidance: a white paper from the digital pathology association. The Journal of Pathology, 249(3):286-294, 2019. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 236, + 287, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 236, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 53, + 236, + 287, + 312 + ], + "type": "text", + "content": "[4] Mohamed Amgad, Lamees A Atteya, Hagar Hussein, Kareem Hosny Mohammed, Ehab Hafiz, Maha AT Elsebaie, Ahmed M Alhusseiny, Mohamed Atef AlMoslemany, Abdelmagid M Elmatboly, Philip A Pappalardo, et al. Nuclei: A scalable crowdsourcing approach and dataset for nucleus classification and segmentation in breast cancer. *Giga* Science, 11, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 313, + 288, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 313, + 288, + 388 + ], + "spans": [ + { + "bbox": [ + 54, + 313, + 288, + 388 + ], + "type": "text", + "content": "[5] Vahid Azimi, Young Hwan Chang, Guillaume Thibault, Jaclyn Smith, Takahiro Tsujikawa, Benjamin Kukull, Braden Jensen, Christopher Corless, Adam Margolin, and Joe W Gray. Breast cancer histopathology image analysis pipeline for tumor purity estimation. In 2017 IEEE 14th International Symposium on Biomedical Imaging (ISBI 2017), pages 1137-1140. IEEE, 2017. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 391, + 286, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 391, + 286, + 423 + ], + "spans": [ + { + "bbox": [ + 53, + 391, + 286, + 423 + ], + "type": "text", + "content": "[6] Tian Bai, Jiayu Xu, and Fuyong Xing. Multi-field of view aggregation and context encoding for single-stage nucleus recognition. In MICCAI, pages 382-392. Springer, 2020. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 425, + 286, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 425, + 286, + 490 + ], + "spans": [ + { + "bbox": [ + 53, + 425, + 286, + 490 + ], + "type": "text", + "content": "[7] Wouter Bulten, Kimmo Kartasalo, Po-Hsuan Cameron Chen, Peter Ström, Hans Pinckaers, Kunal Nagpal, Yuannan Cai, David F Steiner, Hester van Boven, Robert Vink, et al. Artificial intelligence for diagnosis and gleason grading of prostate cancer: the panda challenge. Nature Medicine, 28(1):154-163, 2022. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 491, + 286, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 491, + 286, + 534 + ], + "spans": [ + { + "bbox": [ + 53, + 491, + 286, + 534 + ], + "type": "text", + "content": "[8] Lyndon Chan, Mahdi S Hosseini, Corwyn Rowsell, Konstantinos N Plataniotis, and Savvas Damaskinos. Histosegnet: Semantic segmentation of histological tissue type in whole slide images. In ICCV, pages 10662-10671, 2019. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 536, + 286, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 536, + 286, + 568 + ], + "spans": [ + { + "bbox": [ + 53, + 536, + 286, + 568 + ], + "type": "text", + "content": "[9] Hao Chen, Xiaojuan Qi, Lequan Yu, and Pheng-Ann Heng. Dcan: deep contour-aware networks for accurate gland segmentation. In CVPR, pages 2487–2496, 2016. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 570, + 286, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 570, + 286, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 570, + 286, + 613 + ], + "type": "text", + "content": "[10] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In ECCV, pages 801-818, 2018. 5, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 614, + 286, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 286, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 286, + 678 + ], + "type": "text", + "content": "[11] Qian Da, Xiaodi Huang, Zhongyu Li, Yanfei Zuo, Chenbin Zhang, Jingxin Liu, Wen Chen, Jiahui Li, Dou Xu, Zhiqiang Hu, et al. Digestpath: a benchmark dataset with challenge review for the pathological detection and segmentation of digestive-system. Medical Image Analysis, page 102485, 2022. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 680, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 286, + 713 + ], + "type": "text", + "content": "[12] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248–255. IEEE, 2009. 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 148 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 148 + ], + "type": "text", + "content": "[13] James A Diao, Jason K Wang, Wan Fung Chui, Victoria Mountain, Sai Chowdary Gullapally, Ramprakash Srinivasan, Richard N Mitchell, Benjamin Glass, Sara Hoffman, Sudha K Rao, et al. Human-interpretable image features derived from densely mapped cancer pathology slides predict diverse molecular phenotypes. Nature communications, 12(1):1-15, 2021. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 151, + 545, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 151, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 151, + 545, + 184 + ], + "type": "text", + "content": "[14] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. *IJCV*, 88(2):303–338, 2010. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 186, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 186, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 308, + 186, + 545, + 239 + ], + "type": "text", + "content": "[15] Jevgenij Gamper, Navid Alemi Koohbanani, Ksenija Benes, Simon Graham, Mostafa Jahanifar, Syed Ali Khurram, Ayesha Azam, Katherine Hewitt, and Nasir Rajpoot. Pan-nuke dataset extension, insights and baselines. arXiv preprint arXiv:2003.10778, 2020. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 240, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 240, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 308, + 240, + 545, + 274 + ], + "type": "text", + "content": "[16] Zhe Gong, Jieyun Zhang, and Weijian Guo. Tumor purity as a prognosis and immunotherapy relevant feature in gastric cancer. Cancer Medicine, 9(23):9052-9063, 2020. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 276, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 276, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 308, + 276, + 545, + 330 + ], + "type": "text", + "content": "[17] Simon Graham, Quoc Dang Vu, Shan E Ahmed Raza, Ayesha Azam, Yee Wah Tsang, Jin Tae Kwak, and Nasir Rajpoot. Hover-net: Simultaneous segmentation and classification of nuclei in multi-tissue histology images. Medical Image Analysis, 58:101563, 2019. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 332, + 545, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 332, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 308, + 332, + 545, + 363 + ], + "type": "text", + "content": "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. pages 770-778, 2016. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 365, + 545, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 365, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 308, + 365, + 545, + 430 + ], + "type": "text", + "content": "[19] David Joon Ho, Dig VK Yarlagadda, Timothy M D'Alfonso, Matthew G Hanna, Anne Grabenstetter, Peter Ntiamoah, Edi Brogi, Lee K Tan, and Thomas J Fuchs. Deep multimagnification networks for multi-class breast cancer image segmentation. Computerized Medical Imaging and Graphics, 88:101866, 2021. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 433, + 545, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 433, + 545, + 497 + ], + "spans": [ + { + "bbox": [ + 308, + 433, + 545, + 497 + ], + "type": "text", + "content": "[20] Mahdi S Hosseini, Lyndon Chan, Gabriel Tse, Michael Tang, Jun Deng, Sajad Norouzi, Corwyn Rowsell, Konstantinos N Plataniotis, and Savvas Damaskinos. Atlas of digital pathology: A generalized hierarchical histological tissue type-annotated database for deep learning. In CVPR, pages 11747–11756, 2019. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 500, + 545, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 500, + 545, + 532 + ], + "spans": [ + { + "bbox": [ + 308, + 500, + 545, + 532 + ], + "type": "text", + "content": "[21] Carolyn Hutter and Jean Claude Zenklusen. The cancer genome atlas: creating lasting value beyond its data. Cell, 173(2):283-285, 2018. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 534, + 545, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 534, + 545, + 588 + ], + "spans": [ + { + "bbox": [ + 308, + 534, + 545, + 588 + ], + "type": "text", + "content": "[22] Konstantinos Kamnitsas, Christian Ledig, Virginia FJ Newcombe, Joanna P Simpson, Andrew D Kane, David K Menon, Daniel Rueckert, and Ben Glocker. Efficient multi-scale 3d cnn with fully connected crf for accurate brain lesion segmentation. Medical Image Analysis, 36:61-78, 2017. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 590, + 545, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 590, + 545, + 655 + ], + "spans": [ + { + "bbox": [ + 308, + 590, + 545, + 655 + ], + "type": "text", + "content": "[23] Davood Karimi, Guy Nir, Ladan Fazli, Peter C Black, Larry Goldenberg, and Septimiu E Salcudean. Deep learning-based gleason grading of prostate cancer from histopathology images—role of multiscale decision aggregation and data augmentation. IEEE journal of Biomedical and Health Informatics, 24(5):1413-1426, 2019. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 658, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 678 + ], + "type": "text", + "content": "[24] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 680, + 545, + 713 + ], + "type": "text", + "content": "[25] Neeraj Kumar, Ruchika Verma, Deepak Anand, Yanning Zhou, Omer Fahri Onder, Efstratios Tsougenis, Hao Chen, Pheng-Ann Heng, Jiahui Li, Zhiqiang Hu, et al. A multi" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "23910" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 73, + 287, + 95 + ], + "type": "text", + "content": "organ nucleus segmentation challenge. IEEE transactions on Medical Imaging, 39(5):1380-1391, 2019. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 149 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 149 + ], + "type": "text", + "content": "[26] Shyam Lal, Devikalyan Das, Kumar Alabhya, Anirudh Kanfade, Aman Kumar, and Jyoti Kini. Nucleisegnet: robust deep learning architecture for the nuclei segmentation of liver cancer histopathology images. Computers in Biology and Medicine, 128:104075, 2021. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 151, + 287, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 151, + 287, + 194 + ], + "spans": [ + { + "bbox": [ + 48, + 151, + 287, + 194 + ], + "type": "text", + "content": "[27] Chunggi Lee, Seonwook Park, Heon Song, Jeongun Ryu, Sanghoon Kim, Haejoon Kim, Sérgio Pereira, and Donggeun Yoo. Interactive multi-class tiny-object detection. In CVPR, pages 14136-14145, 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 195, + 287, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 195, + 287, + 260 + ], + "spans": [ + { + "bbox": [ + 48, + 195, + 287, + 260 + ], + "type": "text", + "content": "[28] Wenqi Li, Siyamalan Manivannan, Shazia Akbar, Jianguo Zhang, Emanuele Trucco, and Stephen J McKenna. Gland segmentation in colon histology images using hand-crafted features and convolutional neural networks. In 2016 IEEE 13th international symposium on biomedical imaging (ISBI), pages 1405-1408. IEEE, 2016. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 262, + 287, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 262, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 48, + 262, + 287, + 293 + ], + "type": "text", + "content": "[29] Xieli Li, Yuanyuan Wang, Qisheng Tang, Zhen Fan, and Jinhua Yu. Dual u-net for the segmentation of overlapping glioma nuclei. IEEE Access, 7:84040-84052, 2019. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 295, + 287, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 295, + 287, + 338 + ], + "spans": [ + { + "bbox": [ + 48, + 295, + 287, + 338 + ], + "type": "text", + "content": "[30] Dongnan Liu, Donghao Zhang, Yang Song, Chaoyi Zhang, Fan Zhang, Lauren O'Donnell, and Weidong Cai. Nuclei segmentation via a deep panoptic model with semantic feature fusion. In *IJCAI*, pages 861-868, 2019. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 338, + 287, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 338, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 48, + 338, + 287, + 403 + ], + "type": "text", + "content": "[31] Yihao Mao, Qingyang Feng, Peng Zheng, Liangliang Yang, Tianyu Liu, Yuqiu Xu, Dexiang Zhu, Wenju Chang, Meiling Ji, Li Ren, et al. Low tumor purity is associated with poor prognosis, heavy mutation burden, and intense immune phenotype in colon cancer. Cancer Management and Research, 10:3569, 2018. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 405, + 287, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 405, + 287, + 448 + ], + "spans": [ + { + "bbox": [ + 48, + 405, + 287, + 448 + ], + "type": "text", + "content": "[32] Peter Naylor, Marick Laé, Fabien Reyal, and Thomas Walter. Segmentation of nuclei in histopathology images by deep regression of the distance map. IEEE transactions on Medical Imaging, 38(2):448-459, 2018. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 449, + 287, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 449, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 48, + 449, + 287, + 514 + ], + "type": "text", + "content": "[33] Guy Nir, Soheil Hor, Davood Karimi, Ladan Fazli, Brian F Skinnider, Peyman Tavassoli, Dmitry Turbin, Carlos F Villamil, Gang Wang, R Storey Wilson, et al. Automatic grading of prostate cancer in digitized histopathology images: Learning from multiple experts. Medical Image Analysis, 50:167-180, 2018. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 515, + 287, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 515, + 287, + 591 + ], + "spans": [ + { + "bbox": [ + 48, + 515, + 287, + 591 + ], + "type": "text", + "content": "[34] Sehnoon Park, Chan-Young Ock, Hyojin Kim, Sergio Pereira, Seonwook Park, Minuk Ma, Sangjoon Choi, Seokhwi Kim, Seunghwan Shin, Brian Jaehong Aum, et al. Artificial intelligence-powered spatial analysis of tumor-infiltrating lymphocytes as complementary biomarker for immune checkpoint inhibition in non-small-cell lung cancer. Journal of Clinical Oncology, 40(17):1916, 2022. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 593, + 287, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 593, + 287, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 593, + 287, + 690 + ], + "type": "text", + "content": "[35] Pushpak Pati, Guillaume Jaume, Lauren Alisha Fernandes, Antonio Foncubierta-Rodriguez, Florinda Feroce, Anna Maria Anniciello, Giosue Scognamiglio, Nadia Brancati, Daniel Riccio, Maurizio Di Bonito, et al. Hact-net: A hierarchical cell-to-tissue graph neural network for histopathological image classification. In Uncertainty for Safe Utilization of Machine Learning in Medical Imaging, and Graphs in Biomedical Image Analysis, pages 208-219. Springer, 2020. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "type": "text", + "content": "[36] Ziniu Qian, Kailu Li, Maode Lai, Eric I Chang, Bingzheng Wei, Yubo Fan, Yan Xu, et al. Transformer based multiple" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 105 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 105 + ], + "type": "text", + "content": "instance learning for weakly supervised histopathology image segmentation. arXiv preprint arXiv:2205.08878, 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 308, + 106, + 545, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 106, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 308, + 106, + 545, + 171 + ], + "type": "text", + "content": "[37] Hui Qu, Pengxiang Wu, Qiaoying Huang, Jingru Yi, Zhennan Yan, Kang Li, Gregory M Riedlinger, Subhajyoti De, Shaoting Zhang, and Dimitris N Metaxas. Weakly supervised deep nuclei segmentation using partial points annotation in histopathology images. IEEE transactions on Medical Imaging, 39(11):3655-3666, 2020. 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 173, + 545, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 173, + 545, + 248 + ], + "spans": [ + { + "bbox": [ + 308, + 173, + 545, + 248 + ], + "type": "text", + "content": "[38] Roberto Salgado, Carsten Denkert, S Demaria, N Sirtaine, F Klauschen, Giancarlo Pruneri, S Wienert, Gert Van den Eynden, Frederick L Baehner, Frederique Pénault-Llorca, et al. The evaluation of tumor-infiltrating lymphocytes (tils) in breast cancer: recommendations by an international tils working group 2014. Annals of Oncology, 26(2):259-271, 2015. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 250, + 545, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 250, + 545, + 315 + ], + "spans": [ + { + "bbox": [ + 308, + 250, + 545, + 315 + ], + "type": "text", + "content": "[39] Rüdiger Schmitz, Frederic Madesta, Maximilian Nielsen, Jenny Krause, Stefan Steurer, René Werner, and Thomas Rösch. Multi-scale fully convolutional neural networks for histopathology image segmentation: from nuclear aberrations to the global tissue architecture. Medical Image Analysis, 70:101996, 2021. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 316, + 545, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 316, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 308, + 316, + 545, + 380 + ], + "type": "text", + "content": "[40] Peng Sun, Jiehua He, Xue Chao, Keming Chen, Yuanyuan Xu, Qitao Huang, Jingping Yun, Mei Li, Rongzhen Luo, Jinbo Kuang, et al. A computational tumor-infiltrating lymphocyte assessment method comparable with visual reporting guidelines for triple-negative breast cancer. EBioMedicine, 70:103492, 2021. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 383, + 545, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 383, + 545, + 447 + ], + "spans": [ + { + "bbox": [ + 308, + 383, + 545, + 447 + ], + "type": "text", + "content": "[41] Zaneta Swiderska-Chadaj, Hans Pinckaers, Mart van Rijthoven, Maschenka Balkenhol, Margarita Melnikova, Oscar Geessink, Quirine Manson, Mark Sherman, Antonio Polonia, Jeremy Parry, et al. Learning to detect lymphocytes in immunohistochemistry with deep learning. Medical Image Analysis, 58:101547, 2019. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 449, + 545, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 449, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 308, + 449, + 545, + 491 + ], + "type": "text", + "content": "[42] Hiroki Tokunaga, Yuki Teramoto, Akihiko Yoshizawa, and Ryoma Bise. Adaptive weighting multi-field-of-view cnn for semantic segmentation in pathology. In CVPR, pages 12597-12606, 2019. 1, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 493, + 545, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 493, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 308, + 493, + 545, + 547 + ], + "type": "text", + "content": "[43] Mart Van Rijthoven, Maschenka Balkenhol, Karina Silina, Jeroen Van Der Laak, and Francesco Ciompi. Hooknet: Multi-resolution convolutional neural networks for semantic segmentation in histopathology whole-slide images. Medical Image Analysis, 68:101890, 2021. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 548, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 548, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 308, + 548, + 545, + 613 + ], + "type": "text", + "content": "[44] Ruchika Verma, Neeraj Kumar, Abhijeet Patil, Nikhil Cherian Kurian, Swapnil Rane, Simon Graham, Quoc Dang Vu, Mieke Zwager, Shan E Ahmed Raza, Nasir Rajpoot, et al. Monusac2020: A multi-organ nuclei segmentation and classification challenge. IEEE Transactions on Medical Imaging, 40(12):3413-3423, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 615, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 615, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 615, + 545, + 689 + ], + "type": "text", + "content": "[45] Quoc Dang Vu, Simon Graham, Tahsin Kurc, Minh Nguyen Nhat To, Muhammad Shaban, Talha Qaiser, Navid Alemi Koohbanani, Syed Ali Khurram, Jayashree Kalpathy-Cramer, Tianhao Zhao, et al. Methods for segmentation and classification of digital microscopy tissue images. Frontiers in Bioengineering and Biotechnology, page 53, 2019. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "text", + "content": "[46] Jingwen Wang, Richard J Chen, Ming Y Lu, Alexander Baras, and Faisal Mahmood. Weakly supervised prostate" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23911" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 440 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "text", + "content": "tma classification via graph convolutional networks. In 2020 IEEE 17th International Symposium on Biomedical Imaging (ISBI), pages 239-243. IEEE, 2020. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "type": "text", + "content": "[47] Yan Xu, Zhipeng Jia, Liang-Bo Wang, Yuqing Ai, Fang Zhang, Maode Lai, Eric I Chang, et al. Large scale tissue histopathology image classification, segmentation, and visualization via deep convolutional activation features. BMC bioinformatics, 18(1):1-17, 2017. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 288, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 288, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 288, + 217 + ], + "type": "text", + "content": "[48] Chuanbao Zhang, Wen Cheng, Xiufang Ren, Zheng Wang, Xing Liu, Guanzhang Li, Sheng Han, Tao Jiang, and Anhua Wu. Tumor purity as an underlying key factor in gliomatumor purity in glioma. Clinical Cancer Research, 23(20):6279-6291, 2017. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 219, + 287, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 219, + 287, + 273 + ], + "spans": [ + { + "bbox": [ + 48, + 219, + 287, + 273 + ], + "type": "text", + "content": "[49] Bingchao Zhao, Xin Chen, Zhi Li, Zhiwen Yu, Su Yao, Lixu Yan, Yuqian Wang, Zaiyi Liu, Changhong Liang, and Chu Han. Triple u-net: Hematoxylin-aware nuclei segmentation with progressive dense feature aggregation. Medical Image Analysis, 65:101786, 2020. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 274, + 287, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 274, + 287, + 338 + ], + "spans": [ + { + "bbox": [ + 48, + 274, + 287, + 338 + ], + "type": "text", + "content": "[50] Ke Zhao, Zhenhui Li, Su Yao, Yingyi Wang, Xiaomei Wu, Zeyan Xu, Lin Wu, Yanqi Huang, Changhong Liang, and Zaiyi Liu. Artificial intelligence quantified tumour-stroma ratio is an independent predictor for overall survival in resectable colorectal cancer. EBioMedicine, 61:103054, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 341, + 287, + 395 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 341, + 287, + 395 + ], + "spans": [ + { + "bbox": [ + 48, + 341, + 287, + 395 + ], + "type": "text", + "content": "[51] Yanning Zhou, Simon Graham, Navid Alemi Koohbanani, Muhammad Shaban, Pheng-Ann Heng, and Nasir Rajpoot. Cgc-net: Cell graph convolutional network for grading of colorectal cancer histology images. In ICCV Workshops, pages 0–0, 2019. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 396, + 287, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 396, + 287, + 440 + ], + "spans": [ + { + "bbox": [ + 48, + 396, + 287, + 440 + ], + "type": "text", + "content": "[52] Chuang Zhu, Ke Mei, Ting Peng, Yihao Luo, Jun Liu, Ying Wang, and Mulan Jin. Multi-level colonoscopy malignant tissue detection with adversarial cac-unet. Neurocomputing, 438:165-183, 2021. 2" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23912" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/OCTET_ Object-Aware Counterfactual Explanations/03950ee8-92a4-474c-99c2-49607c7368f1_content_list.json b/2023/OCTET_ Object-Aware Counterfactual Explanations/03950ee8-92a4-474c-99c2-49607c7368f1_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..915e57514200add80ad7c01d19383b29cd318dd1 --- /dev/null +++ b/2023/OCTET_ Object-Aware Counterfactual Explanations/03950ee8-92a4-474c-99c2-49607c7368f1_content_list.json @@ -0,0 +1,1357 @@ +[ + { + "type": "text", + "text": "OCTET: Object-aware Counterfactual Explanations", + "text_level": 1, + "bbox": [ + 218, + 130, + 751, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mehdi Zemni1, Mickaël Chen1, Éloi Zablocki1, Hédi Ben-Younes1, Patrick Pérez1, Matthieu Cord1,2 \n1 Valeo.ai, Paris, France \n2 Sorbonne Université, Paris, France", + "bbox": [ + 76, + 179, + 898, + 219 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 253, + 313, + 271 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Nowadays, deep vision models are being widely deployed in safety-critical applications, e.g., autonomous driving, and explainability of such models is becoming a pressing concern. Among explanation methods, counterfactual explanations aim to find minimal and interpretable changes to the input image that would also change the output of the model to be explained. Such explanations point end-users at the main factors that impact the decision of the model. However, previous methods struggle to explain decision models trained on images with many objects, e.g., urban scenes, which are more difficult to work with but also arguably more critical to explain. In this work, we propose to tackle this issue with an object-centric framework for counterfactual explanation generation. Our method, inspired by recent generative modeling works, encodes the query image into a latent space that is structured in a way to ease object-level manipulations. Doing so, it provides the end-user with control over which search directions (e.g., spatial displacement of objects, style modification, etc.) are to be explored during the counterfactual generation. We conduct a set of experiments on counterfactual explanation benchmarks for driving scenes, and we show that our method can be adapted beyond classification, e.g., to explain semantic segmentation models. To complete our analysis, we design and run a user study that measures the usefulness of counterfactual explanations in understanding a decision model. Code is available at https://github.com/valeoaai/OCTET.", + "bbox": [ + 73, + 286, + 473, + 710 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 738, + 209, + 753 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep learning models are now being widely deployed, notably in safety-critical applications such as autonomous driving. In such contexts, their black-box nature is a major concern, and explainability methods have been developed to improve their trustworthiness. Among them, counterfactual explanations have recently emerged to provide insights into a model's decision [7, 52, 55]. Given a decision model and an input query, a counterfactual explanation is a data point that differs minimally but meaningfully from the query in", + "bbox": [ + 75, + 763, + 470, + 902 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d33776a7a391bf2261a43e245a6f91dce3948c346fb8e8cc627ebdcddf3c8507.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 503, + 253, + 694, + 342 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e47858e4c9cc444c0f953f84b0949b80c1da045a5b34ded8728f5484e20ac5d0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 696, + 255, + 887, + 342 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d73cc6e01394921ba2748a6a2fcc1da5b8ef572cb6da23bf7142c222f05c3999.jpg", + "image_caption": [ + "Figure 1. Counterfactual explanations generated by OCTET. Given a classifier that predicts whether or not it is possible to go left, and a query image (top left), OCTET produces a counterfactual explanation where the most influential features that led to the decision are changed (top right). On the bottom row, we show that OCTET can also operate under different settings that result in different focused explanations. We report the prediction made by the decision model at the top left of each image." + ], + "image_footnote": [], + "bbox": [ + 503, + 343, + 694, + 421 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9b87e53b767850201cdde52254fad8d6e04c50f1d066d5192fd8d202da78a202.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 343, + 887, + 421 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "a way that changes the output decision of the model. By looking at the differences between the query and the explanation, a user is able to infer — by contrast — which elements were essential for the model to come to its decision. However, most counterfactual methods have only shown results for explaining classifiers trained on single-object images such as face portraits [4, 28, 29, 42, 46]. Aside from the technical difficulties of scaling up the resolution of the images, explaining decision models trained on scenes composed of many objects also present the challenge that those decisions are often multi-factorial. In autonomous driving, for example, most decisions have to take into account the position of all other road users, as well as the layout of the road and its markings, the traffic light and signs, the overall visibility, and many other factors.", + "bbox": [ + 496, + 579, + 893, + 806 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we present a new framework, dubbed OCTET for Object-aware Counterfactual ExplanaTions, to generate counterfactual examples for autonomous driving. We leverage recent advances in unsupervised compositional generative modeling [14] to provide a flexible explanation method. Exploiting such a model as our backbone, we can", + "bbox": [ + 496, + 810, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 808, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "15062", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "assess the contribution of each object of the scene independently and look for explanations in relation to their positions, their styles, or combinations of both.", + "bbox": [ + 75, + 90, + 468, + 135 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To validate our claims, extensive experiments are conducted for self-driving action decision models trained with the BDD100k dataset [60] and its BDD-OIA extension [59]. Fig. 1 shows counterfactual explanations found by OCTET: given a query image and a visual decision system trained to assess if the ego vehicle is allowed to go left or not, OCTET proposes a counterfactual example with cars parked on the left side that are moved closer. When inspecting specific items (bottom row of the figure), OCTET finds that moving the yellow car to the left, or adding a double line marking on the road, are ways to change the model's decision. These explanations highlight that the absence of such elements on the left side of the road heavily influenced the model.", + "bbox": [ + 75, + 136, + 467, + 330 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To sum up, our contributions are as follows:", + "bbox": [ + 76, + 333, + 372, + 347 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We tackle the problem of building counterfactual explanations for visual decision models operating on complex compositional scenes. We specifically target autonomous driving visual scenarios.", + "- Our method is also a tool to investigate the role of specific objects in a model's decision, empowering the user with control over which type of explanation to look for.", + "- We thoroughly evaluate the realism of our counterfactual images, the minimality and meaningfulness of changes, and compare against previous reference strategies. Beyond explaining classifiers, we also demonstrate the versatility of our method by addressing explanations for a segmentation network.", + "- Finally, we conduct a user-centered study to assess the usefulness of our explanations in a practical case. As standard evaluation benchmarks for counterfactual explanations are lacking a concrete way to measure the interpretability of the explanations, our user-centered study is a key element to validate the presented pipeline." + ], + "bbox": [ + 86, + 348, + 467, + 633 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Background and Related Work", + "text_level": 1, + "bbox": [ + 76, + 648, + 362, + 665 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Local explanations. The overwhelming majority of deep learning based models are designed without explainability in mind, with only a few notable exceptions [8, 63]. This fact has prompted an interest in post-hoc explanations of already trained models that can be useful to analyze corner cases, understand failures, and find biases [15, 40]. In safety-critical applications, e.g., autonomous driving or medical imaging, explanations are especially needed for liability purposes and to foster end-user trust [45, 49, 61, 64]. Post-hoc explanations are global if they provide a holistic view of the main decision factors driving the model [18, 21, 32, 33], or they are local if they target the understanding of the model behavior on a specific input [40, 44]. Historically, the vast majority of local explanation methods are attribution-based: they generate saliency maps high-", + "bbox": [ + 75, + 674, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "lighting pixels or regions influencing the most the model's decision [5, 16, 39, 44, 47, 53, 56, 62, 66]. In the case of urban scenes, a saliency method would for instance point at the traffic light, or the presence of a pedestrian to explain why a driving model stops [31, 43]. However, saliency explanations may be misleading as they can be independent of the model at hand and merely act as edge detectors [3]. Besides, saliency methods are by nature restricted to show where the important regions are and they cannot indicate what in these regions was deemed decisive. For example, beyond highlighting things, it would be useful to know if the color of the light (red or green?) or the orientation of the pedestrian's body (towards or away from the car?) are taken into account in the model's decision. Counterfactual explanations are a step in that direction.", + "bbox": [ + 496, + 90, + 890, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Counterfactual explanations. For a given decision model and query input, a counterfactual explanation is defined as a minimally but meaningfully modified version of the query that changes the decision of the model [55]. The notions of 'minimal' and 'meaningful' are crucial here. Adversarial attacks [19, 35, 48] for instance are minimal but not meaningful perturbations: they are imperceptible by design and thus not informative for the end-user [6, 17, 38]. Conversely, if not minimal, the explanation might exhibit unnecessary changes that would hinder its interpretability. While originally developed for models operating on tabular data [55], counterfactuals have been recently considered to explain deep computer vision models [20, 27, 42, 46].", + "bbox": [ + 496, + 324, + 890, + 520 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Seminal work [23] explored the use of natural language to produce descriptions of elements that, in case of presence, change the image classifications. Later, retrieval-based methods aimed to explain decisions by mining the dataset for counterfactual explanations [20, 23, 51, 58]. While obtained explanations can be meaningful, the difference to the query can only be as minimal as the dataset granularity allows, limiting overall interpretability.", + "bbox": [ + 496, + 521, + 890, + 642 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "More recently, generator-based counterfactual explanations have been proposed [4, 27-29, 42]. They exploit the power of deep generative models that map a latent space to the data distribution. By optimizing for small manipulations in the latent space, these methods ensure for both the minimality and the meaningfulness of changes. The properties of the generative backbone and the way the latent space is structured have a crucial impact on the resulting counterfactual explanations. DIVE [42], for instance, exploits the disentangled latent space of a $\\beta$ -TCVAE [10] to generate diverse counterfactual explanations. However, the low visual quality of its counterfactuals hinders interpretability. DIME [28] and DVCE [4] instead build on recently popularized denoising diffusion probabilistic models [12], achieving better-looking results but losing the more curated control on the diversity. These works can explain classifiers working on 'simple' images, e.g., face portraits [34] or", + "bbox": [ + 496, + 643, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "15063", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3fe7bee604b8e640f8ff7e16ac9095d4843e05e9bd9339eaefa744bb09677681.jpg", + "image_caption": [ + "Figure 2. Overview of OCTET. From the query image (top left), a trained encoder $E$ provides initial blob parameters. The blobs are refined in optimization stage ① using Eq. 3 to better encode the query. The refined blob parameters are again modified in optimization stage ② using Eq. 1 to obtain a counterfactual example. The parameters of the encoder $E$ and generator $G$ are frozen during the optimization, as is $M$ the decision model to be explained. Losses are illustrated with dashed arrows whose directions represent the gradient flow." + ], + "image_footnote": [], + "bbox": [ + 78, + 88, + 890, + 340 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "featuring a single and centered object [11, 57].", + "bbox": [ + 76, + 438, + 383, + 453 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Scaling up to classifiers operating over more complex images such as crowded urban scenes remains a major challenge. To date, only STEEX [27] proposes a method able to explain decision models dealing with scenes composed of many objects. It employs a segmentation-to-image generation backbone [68], which helps for generating high-quality complex images but comes with two important limitations. Firstly, it requires annotated semantic segmentation masks of images to work, which is not always available or is costly for the user to produce. Secondly, it imposes that the explanation keeps the same spatial and semantic layout as the query, thereby severely restricting the search space. It is, for instance, not able to provide insights about the importance of the position of objects. By contrast, OCTET does not impose a fixed spatial layout. Indeed, the structural layout of the scene is an intermediate representation of the generator, inferred from images only. Therefore, we do not use additional annotations, and the spatial layout can be optimized by back-propagation.", + "bbox": [ + 75, + 453, + 473, + 741 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Object-aware Counterfactual Explanations", + "text_level": 1, + "bbox": [ + 76, + 753, + 465, + 772 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We present here OCTET, a method to produce counterfactual examples for compositional scenes, based on a generative architecture. We first formally describe our approach and objective in Sec. 3.1. Then, we instantiate the generative backbone in Sec. 3.2 and the counterfactual objective in Sec. 3.3. Lastly, we explain how we invert the generator to recover the latent code for the query image in Sec. 3.4. The pipeline of OCTET is shown in Fig. 2.", + "bbox": [ + 75, + 780, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Goal and Notations", + "text_level": 1, + "bbox": [ + 500, + 436, + 684, + 452 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a decision model $M$ , a query image $x^{q}$ , and a target decision $y \\neq M(x^{q})$ , a counterfactual explanation is an image $x$ that is close to the query image $x^{q}$ but modified so that the decision $M(x)$ of the model changes to $y$ . Generative counterfactual methods, such as ours, employ a trained generator $G$ [27, 28, 42]. We denote its latent space $\\mathcal{Z}$ , with $z^{q} \\in \\mathcal{Z}$ the latent code that produces the best approximation of $x^{q}$ when passed through $G$ , i.e., so that $G(z^{q}) \\approx x^{q}$ . To explain differentiable deep vision models, the definition of a counterfactual explanation stated here-above can be translated into the following optimization problem:", + "bbox": [ + 496, + 460, + 893, + 628 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\underset {z \\in \\mathcal {Z}} {\\operatorname {a r g m i n}} L _ {\\text {d e c i s i o n}} \\left(M (G (z)), y\\right) + \\lambda_ {\\text {d i s t}} L _ {\\text {d i s t}} \\left(z, z ^ {q}\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 638, + 890, + 661 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The role of the first term $L_{\\mathrm{decision}}$ is to push the decision of the model $M$ for the generated image $G(z)$ to the target class $y$ ; The second term $L_{\\mathrm{dist}}$ is a distance term in the latent space of $G$ , which ensures that the query $x^q$ and the explanation $G(z)$ remain similar; $\\lambda_{\\mathrm{dist}}$ is an hyper-parameter that controls the relative weight between the two terms. With optimal $z$ , $G(z)$ then recovers a counterfactual image for the query $x^q$ and the decision model $M$ . Since the choice of $L_{\\mathrm{decision}}$ is constrained by the nature of the model to explain, e.g., a classification loss if $M$ is a classifier, we only have to specify $G$ and $L_{\\mathrm{dist}}$ to define our counterfactual method.", + "bbox": [ + 496, + 671, + 893, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Compositional generative backbone", + "text_level": 1, + "bbox": [ + 500, + 847, + 810, + 864 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this work, we focus on explaining decision models that operate on compositional visual scenes and we look for", + "bbox": [ + 500, + 869, + 893, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "15064", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "a generative model capable of modeling such scenes. However, most generative models are designed for images containing centered, isolated objects, and they do not learn a latent space that can easily be disentangled into per-object representations. Instead, we propose that our generative backbone should be object-based by design [14, 37]. Such methods are fully differentiable and trained without supervision to generate images in a compositional fashion.", + "bbox": [ + 75, + 90, + 468, + 210 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We choose to build on BlobGAN [14], a compositional generative model that shows good performances and has the advantage of allowing not only edition but also insertion and removal of objects in a differentiable way. In BlobGAN, images are generated in two steps. First, a layout network maps a random Gaussian noise vector to an intermediary representation consisting of an ordered set of $K$ 'blobs'. Each blob $k \\in \\{1 \\dots K\\}$ is defined by spatial parameters $\\phi_k$ and a style feature vector $\\psi_k$ . The spatial parameters $\\phi_k$ consist of five values that are its center coordinates, scale, aspect ratio, and rotation angle. In practice, the layout network is a neural network that outputs vectors $\\phi_k$ and $\\psi_k$ for each blob. Then, in the second stage, the generator network is tasked to transform the blob parameters into an image. A rendering layer draws the blobs on a canvas as ellipses whose shapes and positions are given directly by $\\phi_k$ . The output of the rendering layer is a map that associates to each 2D spatial location a feature vector computed from the style vectors of the blobs present at that position. This feature map is transformed by convolutional layers into an image that roughly follows the spatial organization of the canvas.", + "bbox": [ + 75, + 212, + 470, + 530 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The layout network and the generator network are trained end-to-end against a discriminator with standard GAN losses only. In particular, we do not use any annotations for the number, size, class, or location of objects. They are discovered by the generator during training. Note also that scale parameters can take non-positive values. In that case, the object is considered absent and the corresponding blob is not rendered on the canvas in the generator network.", + "bbox": [ + 75, + 531, + 468, + 650 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Setting $L_{\\mathrm{dist}}$ for compositional scenes", + "text_level": 1, + "bbox": [ + 76, + 662, + 390, + 680 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Using a compositional generative backbone allows us to design a meaningful, disentangled, per-object distance for images. Indeed, to be interpretable with respect to the query image, a counterfactual explanation should display sparse changes in terms of the number of moved, added, removed, or modified objects. Therefore, we design $L_{\\mathrm{dist}}$ such that it is split into per-object distances where each object can be addressed independently. Moreover, we propose that modifications of the spatial position of an object should be orthogonal to those made to its shape or colors. Since in a well-trained BlobGAN each blob is associated with a distinct element of the scene (see Appendix), such a distance can be naturally defined in the space of blob parameters.", + "bbox": [ + 75, + 688, + 468, + 883 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Formally, let us denote $\\phi$ and $\\psi$ the concatenation across", + "bbox": [ + 96, + 885, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "all blobs $k \\in \\{1 \\dots K\\}$ of spatial parameters $\\phi_k$ and style vectors $\\psi_k$ , and $z = (\\phi, \\psi)$ the complete latent representation that the generator network $G$ can decode into an image. Accordingly, the latent code for the query image $x^q$ reads $z^q = (\\phi^q, \\psi^q)$ . Then, OCTET implements $L_{\\mathrm{dist}}$ as follows:", + "bbox": [ + 498, + 90, + 890, + 167 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {d i s t}} (z, z ^ {q}) = \\sum_ {k = 1} ^ {K} \\| \\phi_ {k} - \\phi_ {k} ^ {q} \\| _ {1} + \\| \\psi_ {k} - \\psi_ {k} ^ {q} \\| _ {1}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 539, + 180, + 890, + 220 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We use the $L_{1}$ distance here to promote sparsity in the number of modified blobs and the type of edits applied to each blob. If the user wishes to inspect the role of the style of objects, their spatial positions, or both, with a counterfactual explanation, she can instead optimize Eq. 2 only on the corresponding features of the blobs.", + "bbox": [ + 496, + 234, + 890, + 325 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In order to specify Eq. 2, and thus the final counterfactual objective (Eq. 1), it remains to be solved how to obtain the blob-wise latent codes $\\phi_k^q$ and $\\psi_k^q$ associated with the query image $x^q$ . We discuss this matter in the next section.", + "bbox": [ + 496, + 325, + 890, + 386 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Inverting the query image", + "text_level": 1, + "bbox": [ + 500, + 398, + 735, + 415 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Obtaining the latent code corresponding to the query image, a process called inversion, is not straightforward when using a GAN-based generator. To produce a counterfactual example, such inversion needs to find latent codes $(\\phi^q,\\psi^q)$ so that $G(\\phi^q,\\psi^q)$ recovers $x^{q}$ . For the downstream counterfactual optimization to make sense, the decision of the model for the query image should be preserved for the reconstructed image, i.e., $M(G(\\phi^q,\\psi^q)) = M(x^q)$ .", + "bbox": [ + 496, + 422, + 890, + 544 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To achieve those objectives, we follow best practices for image inversion [1,2,41,50,67]. We first learn an encoder $E$ to predict blob parameters from input images. Generating an image $G(E(x^q))$ with parameters predicted this way yields a coarse reconstruction of $x^q$ . To improve the inversion, we start from predicted blob parameters $E(x^q)$ and further optimize them with the following objective:", + "bbox": [ + 496, + 545, + 890, + 650 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\phi^ {q}, \\psi^ {q} = \\underset {\\phi , \\psi} {\\arg \\min } L _ {\\mathrm {L P I P S}} (G (\\phi , \\psi), x ^ {q})) + L _ {2} (G (\\phi , \\psi), x ^ {q}) \\\\ + L _ {2} \\left(f _ {M} \\left(x ^ {q}\\right), f _ {M} (G (\\phi , \\psi))\\right) \\\\ + L _ {2} \\left(\\left(\\phi , \\psi\\right), E \\left(x ^ {q}\\right)\\right). \\tag {3} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 664, + 890, + 720 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The first two terms ensure that the image $G(\\phi, \\psi)$ generated from the blob parameters reconstructs the query $x^q$ using a perceptual $L_{\\mathrm{LPIPS}}$ loss and the $L_2$ loss. The third term aims to preserve the features that are important to the decision model $M$ by adding a $L_2$ loss on intermediate features $f_M(x)$ of $M$ . Finally, the last term encourages the parameters to stay close to the reasonably good solution given by the encoder. This inversion method concludes the specification of the optimization problem of Eq. 1 in OCTET and of the pipeline depicted in Fig. 2. More details on the inversion and the training of the encoder are in the appendix.", + "bbox": [ + 496, + 734, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "15065", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f31e535168e4f9d9249280d5abad0182018c73380052c848ce4fd719bd7a5e83.jpg", + "image_caption": [ + "Figure 3. Counterfactual explanations on driving scenes. Explanations are generated for binary classifiers predicting the capacity of the ego-car to respectively Move forward, Stop, turn Left, or Right. We indicate at the top left of each image the output decision of the model for that image. For each row, counterfactuals target the opposite decision to that obtained for the query image. OCTET finds interpretable, sparse, and meaningful semantic and spatial modifications to the query image. OCTET style-only only optimizes on the style features, it is analog to STEEX [27] that can only search for changes in object appearance to flip the decision." + ], + "image_footnote": [], + "bbox": [ + 78, + 88, + 282, + 414 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/662c96748894516838b570468917ecaf4e572d0a0221a528bbdad3e1b5ac202e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 282, + 88, + 491, + 412 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/513ce641086b18691d286a66acf1d2e978ae0552214c7927bacd25144a220cc5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 491, + 88, + 689, + 412 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/868051f78f2c7cfebb90c4a0a9c6249426f0c8d5c0395886a94835af47fdab24.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 689, + 89, + 888, + 412 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 513, + 209, + 532 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental protocol and training details", + "text_level": 1, + "bbox": [ + 76, + 539, + 439, + 556 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Data and decision model. Our method is designed to build counterfactual examples to explain decision models trained on compositional images. To evaluate our method, we follow previous work [27] and use the BDD100k dataset [60], which contains 100k images of diverse driving scenes that we resize to a resolution of $512 \\times 256$ . In addition, we also use BDD-OIA [59], a 20K scene extension of BDD100k that is annotated with 4 binary attributes labeled Move Forward, Stop, Turn Left, and Turn Right. Each label is defined as the capacity in the given situation of the ego-vehicle to perform the corresponding action. Note that by this definition, being allowed to Stop and to Move Forward are not mutually exclusive. The decision models being explained in the experiments are multi-label binary DenseNet121 [26] classifiers trained on BDD-OIA with the four aforementioned labels.", + "bbox": [ + 75, + 563, + 472, + 805 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baseline. We compare our method against STEEX [27], the only counterfactual explanation method that performs well on compositional images like driving scenes. Note that in STEEX, the conservation of the semantic layout is hardcoded in the architecture: it is less flexible as an explanation method than ours. Moreover, the authors use semantic seg", + "bbox": [ + 75, + 809, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "mentation annotations of BDD100k to train their generative backbone, while we do not need any additional annotation. We use the pre-trained model from the official release.", + "bbox": [ + 496, + 516, + 890, + 561 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Technical setup. For our method, we train the generative backbone (BlobGAN) on the training set of BDD100k, with a number $K = 40$ of blobs. We adapt the architecture of BlobGAN to generate rectangular $512 \\times 256$ images, and we decrease the size of the feature vectors of the BlobGAN to compensate for the increased number of blobs. Throughout our experiments, we evaluate two main variants of our method: OCTET where both spatial and style features are optimized altogether and OCTET style-only where only the style vector $\\psi$ is optimized while the spatial vector $\\phi$ is left equal to $\\phi^q$ . This setting only assesses the contribution of style features, and is most similar to the STEEX baseline [27] that finds explanations while conserving the semantic layout of the query. More details in the appendix.", + "bbox": [ + 496, + 565, + 892, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Quality of the explanations", + "text_level": 1, + "bbox": [ + 500, + 786, + 743, + 803 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Counterfactual explanations are defined as images that are similar to the query image but effectively change the decision of the target model. To verify that our explanations fit the criteria, we evaluate them with the following metrics: 1) the FID [24] which measures if they are realistic images, and 2) a perceptual distance LPIPS [65] that mea", + "bbox": [ + 496, + 809, + 892, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "15066", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/06746fddb885cfc21d5ade7ca54e393a7c56d79008c40646fcfdcb216c87ec72.jpg", + "image_caption": [ + "Figure 4. FID vs. LPIPS, labeled with Success rate for OCTET and STEEX [27] explaining a decision model trained on the 4 labels of BDD-OIA. Data points are obtained with different values of $\\lambda_{\\mathrm{dist}}$ yielding different trade-offs. STEEX is directly comparable to OCTET style-only as they operate in similar conditions. See Sec. 4.2." + ], + "image_footnote": [], + "bbox": [ + 76, + 85, + 895, + 273 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "sures the change magnitude between them and the query image. We also compute the success rate corresponding to the percentage of counterfactual examples that are indeed classified in the target class by the decision model. We compute these metrics for different values of $\\lambda_{\\mathrm{dist}}$ that lead to different trade-offs and plot the scores in Fig. 4. As STEEX explanations are constrained to style features, we compare it to our OCTET style-only setting. Firstly, in terms of FID, we can observe that both versions of OCTET obtain lower FID than STEEX: OCTET produces explanations of better visual quality, even though it does not use any annotations as opposed to STEEX, which needs semantic maps, and handles spatial explanations that are more challenging to generate. Secondly, in terms of LPIPS, for style-based explanations OCTET style-only stays closer to the query than STEEX. This can be explained as STEEX latent representation is semantic-based, while ours is instance-based. Our method is therefore able to seek finer-grained modifications, while STEEX tends to change the style for an entire semantic class even when only one instance is important to the decision model. As for OCTET, because LPIPS disproportionately weights structural modifications, the wide range of attainable values indicates that many degrees of changes can be obtained depending on the choice of $\\lambda_{\\mathrm{dist}}$ . Some settings tend to move objects a lot, while other settings attain values similar to 'style-only' edits. Finally, the very high achievable Success Rates displayed by all three models confirm that they will propose explanations in almost any situation.", + "bbox": [ + 75, + 339, + 472, + 762 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We display examples of OCTET's counterfactual explanations in Fig. 3. OCTET finds sparse spatial and/or semantic modifications to objects that strongly influence the output decision. For example, we can see with OCTET that the distance with the car in front plays a role when deciding for 'Forward' (1st row) or 'Stop' (2nd row). On the other hand, OCTET style-only shows that for those same decisions, the brake lights also are taken into account. For going 'Left' class (3rd row), the absence of the continuous line seems", + "bbox": [ + 75, + 763, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/02ecf044eeeb09d18676375f0f91775f46ec942e42550eda0fa0ac116ee09fc0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 503, + 335, + 702, + 412 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/cf2a1f9c734e9004d3040b14da5d2538c94a2698debb9edb5b6896c23da7e16d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 704, + 337, + 888, + 411 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0ba562000555d27052b79a88c49fcfa6b01570a01e619c233aa7dac408cc5c17.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 503, + 412, + 702, + 503 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4cc55b71d51124381d24eff6d9917f9818338f0f7b6c629b64984ab2af6408de.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 704, + 412, + 888, + 503 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e0695404d0381ce301bcafbc942476d2f8f1bc6bb2920e79b4047bef45f7313c.jpg", + "image_caption": [ + "Figure 5. Targeted counterfactual explanations. For each query image (top row), we target different blobs (2nd and 3rd row) to inspect the role of corresponding objects. The output of the decision model is reported on the top right corner for each image." + ], + "image_footnote": [], + "bbox": [ + 516, + 484, + 702, + 559 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/382776078b54f5ced820578e6c6683388632bf6d795c6dd4e5da08983f492584.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 704, + 484, + 888, + 559 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "crucial. On the 3rd and 4th row, we can see that OCTET mixes both style (brake lights) and spatial (proximity) explanations. On the same tasks, STEEX changes the style of the whole class, and not just the necessary instance.", + "bbox": [ + 496, + 656, + 893, + 717 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Targeted counterfactuals", + "text_level": 1, + "bbox": [ + 500, + 726, + 728, + 741 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Models that operate on compositional scenes can have their decisions influenced by many different factors, and we have shown that OCTET can expose those factors in its counterfactual examples. However, the user may want to test some particular hypothesis dealing with specific objects or areas of the scene. Accordingly, the user can decide to optimize Eq. 2 only on selected blobs parameters. Finding the blob corresponding to a specific object in the image is discussed in Appendix. For instance, in the situation shown at the top left of Fig. 5, the car at the front seems to have", + "bbox": [ + 496, + 750, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "15067", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/26199a254b14a14d894dd2752db6136ce1d86e516264b0c238cdd47b2550a2b5.jpg", + "image_caption": [ + "Figure 6. Counterfactual explanations for a segmentation model. The segmentation model predicts the driveable area with primary and secondary lanes shown in red and blue respectively. The first row shows query images and segmentations by the model. In the second row, we add or remove parts of the original masks to build target masks (changes are shown with white dotted lines). The third row displays counterfactual explanations for the target masks and predictions on the new images. Green ellipses are manually added to highlight details." + ], + "image_footnote": [], + "bbox": [ + 80, + 88, + 893, + 320 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "its brake lights on, but the decision is still not to 'Stop'. Targeting the front car (2nd row left image) reveals that the decision would have been 'Stop' if the car was closer, and the brake light more clearly contrasted. OCTET can also target blobs that are currently inactive, corresponding to no visible object in the image. In this example, doing so for car blobs shows that if another vehicle also had brake lights on, the model would also decide to 'Stop' (3rd row left image). This finding is unexpected, and could warrant further investigation of the decision model. We show in Fig. 5 and Fig. 1 some other results we obtain by targeting specific blobs.", + "bbox": [ + 75, + 417, + 472, + 585 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Explaining a semantic segmentation model", + "text_level": 1, + "bbox": [ + 76, + 599, + 441, + 617 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For complex tasks, like autonomous driving, decision models usually rely on multiple neural networks that address different subtasks (e.g., traffic sign recognition, depth estimation, motion forecasting). In this context, it is important to explain models beyond classifiers [25,54], a question not addressed in previous counterfactual methods. Here, we show the versatility of our method as we extend it to explain a semantic segmentation network. To do so, we use the same optimization problem described in Eq. 1 and Sec. 3 but we change $L_{\\text{decision}}$ to an image segmentation loss.", + "bbox": [ + 75, + 626, + 468, + 777 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Fig. 6, we show counterfactual explanations for a driveable area segmentation model. More precisely, OCTET explains outputs of a DeepLabv3 [9] trained to segment images from BDD100k into classes of driveable area: the primary lane, the secondary side lanes, and the background. We define the target masks by manually modifying predictions made on the query images as illustrated in the second row. We see in the obtained counterfactuals, in the", + "bbox": [ + 75, + 779, + 470, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "third row, changes in road markings and vehicles' positions. They confirm that the segmentation model will consider as driveable a lane behind broken white lines (col. 3) but not behind a double yellow line (col. 2), especially if, in addition, the lane is hatched (col. 3). If there is a car on a drivable lane, the model can detect that the occupied portion is non-driveable while considering that the rest of the lane remains accessible. With these examples, OCTET provides explanations for both large semantic regions, e.g., when we add or remove an entire region in the target mask, or small subparts, e.g., when we slightly crop or extend a region.", + "bbox": [ + 496, + 417, + 893, + 585 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. User-study to assess explanations", + "text_level": 1, + "bbox": [ + 500, + 599, + 790, + 617 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Metrics presented in Sec. 4.2 measure how well the produced examples match the definition of counterfactual explanations. We want here to assess how useful the explanations are for a user to better understand a decision model. However, designing such an evaluation is challenging. Indeed, it has to avoid the many biases that appear [13, 36]. Moreover, we usually do not have access to the inner workings of a model, and therefore we cannot compare users' insights on its behavior against a ground truth. To overcome these challenges, we take inspiration from recent work on the usefulness of saliency maps as explanations [15] and we adapt its evaluation protocol to counterfactual explanations.", + "bbox": [ + 496, + 626, + 890, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The goal is to measure if accessing counterfactual explanations on some examples helps users to better understand the model. Grounded within previous literature [16,22,30], we instantiate understand as 'being able to predict the model's output for new instances', a concept known as simulatability. In particular, our user-study has two phases:", + "bbox": [ + 496, + 809, + 893, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "15068", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d274907b67a0c6707ceab17b4b4752903e4a17788ba6607b774777ed6c7d34d3.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Cohort sizeReplicationBias Detection
Group w/ OCTET2070%65%
Group control2052.5%0%
", + "bbox": [ + 76, + 88, + 467, + 145 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 1. Results of the user-study. The performance of users who had access to explanations ('w/ OCTET') in the observation phase is compared to that of users who didn't ('control'). 'Replication' measures the accuracy of the users in predicting the model's output on new samples and 'bias detection' is the proportion of users who found the spurious correlation in the model.", + "bbox": [ + 75, + 155, + 468, + 238 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Observation phase: participants are shown a series of triplets, each being composed of (1) a query image, (2) the model's decision on that image, and (3), a counterfactual explanation for the decision.", + "2. Questionnaire phase: participants are presented with new images and are asked to guess the output of the model on these images. We stress that during the questionnaire, no counterfactual images are shown." + ], + "bbox": [ + 80, + 267, + 467, + 387 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To measure the impact that counterfactual explanations of the first phase have on the users' ability to predict the model output on new samples (second phase), we use a control group. This control group, composed of different participants, undertakes the exact same tasks except that counterfactual explanations are not shown during the first phase. The decision model used in this study is a classifier trained on BDD-OIA [59] to predict the binary 'Turn Right' label. Besides, the decision model is trained, by design, with a flaw: the presence of a car on any side of the road can impact the decision, not only obstacles on the right side. As discussed further below, this bias is introduced to study if users are able to identify model defects thanks to counterfactuals. More details are provided in the appendix.", + "bbox": [ + 75, + 390, + 467, + 599 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Replication score. We first measure the 'replication score': the proportion of correct match between the user's answer and the prediction of the decision model, averaged over participants of the group. We report these results in Tab. 1 and observe that participants in the group accessing counterfactuals during the observation phase can successfully guess the decision model's prediction $70\\%$ of the time while the control group's performance is $52.5\\%$ , barely above random. A statistical unpaired t-test, with as null hypothesis that the counterfactual explanations have no effect, confirms the significance of our result with a p-value of 0.0028. This demonstrates the usefulness of OCTET in understanding the decision model as participants with explanations could better anticipate the model's behavior on new instances.", + "bbox": [ + 75, + 607, + 467, + 816 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Bias detection. Another promise of counterfactual explanations is to enable bias detection. We then want to assess whether, by looking at explanations, users can tell if and how a decision model is biased. In that direction, previous works dealing with human face datasets train biased models", + "bbox": [ + 75, + 824, + 467, + 898 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "by artificially confounding attributes, e.g., smile and age, in the training set. Then, they develop bias detection benchmarks that measure, by means of an oracle neural network, how well attributes remain confounded in the counterfactual explanations [28, 42, 46]. However, a serious limitation of that benchmark is that it keeps end-users out of the evaluation loop and it cannot tell whether or not a user will be able to uncover the spurious correlations by looking at the counterfactual explanation. Moreover, detecting the issue with the model can be especially difficult when the bias is not obvious or its presence not suspected in the first place. Instead of an automated evaluation, at the end of the questionnaire, we asked the participants to describe in free text what they had understood from the model. The goal is to prompt them to mention the issue we embedded in the decision model, i.e., that the presence of a car on the left-hand side of the image factors in the 'Turn Right' decision. The 'bias detection' score simply counts, for each group, the proportion of participants that used the word 'left' in their answers. In Tab. 1, we see that $65\\%$ of the participants that had access to the explanations from OCTET identify the spurious correlation learned by the model, while none of the users that did not have the explanations (control group) mentions it. This clearly shows that our method is useful to pinpoint issues with the decision model. Also, the fact that not every participant detected it confirms our suspicions against automated bias detection benchmarks: the presence of the information in the counterfactual examples is not sufficient to ensure that the issue will be found in practice.", + "bbox": [ + 496, + 90, + 890, + 529 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 545, + 617, + 560 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we presented OCTET, a generative method to produce counterfactual explanations for deep vision models working on complex scenes. Thanks to an object-centric representation of the scene, OCTET is able to process complex compositional scenes and to assess the impact of both the spatial positions and the style features of each object. Our tool allows in-depth exploration of the different explaining factors by letting the user adjust on the fly the weights of the different search directions independently. We evaluated our method thoroughly on real driving scenes from BDD100k, with an improved evaluation benchmark compared to previous works, including a human-centered study that confirmed the usefulness of the approach.", + "bbox": [ + 496, + 571, + 890, + 767 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 500, + 784, + 666, + 801 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported in part by the ANR grants VISA DEEP (ANR-20-CHIA-0022) and MultiTrans (ANR-21-CE23-0032). Authors would like to thank the voluntary participants of the user-study, as well as Remi Cadene, Julien Colin, Thomas Fel, and Guillaume Jeanneret for helpful discussions.", + "bbox": [ + 496, + 809, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "15069", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan: How to embed images into the stylegan latent space? In ICCV, 2019. 4", + "[2] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan++: How to edit the embedded images? In CVPR, 2020. 4", + "[3] Julius Adebayo, Justin Gilmer, Michael Muelly, Ian J. Goodfellow, Moritz Hardt, and Been Kim. Sanity checks for saliency maps. In NeurIPS, 2018. 2", + "[4] Maximilian Augustin, Valentyn Boreiko, Francesco Croce, and Matthias Hein. Diffusion visual counterfactual explanations. In NeurIPS, 2022. 1, 2", + "[5] Mariusz Bojarski, Anna Choromanska, Krzysztof Choromanski, Bernhard Firner, Larry J Ackel, Urs Muller, Phil Yeres, and Karol Zieba. Visualbackprop: Efficient visualization of cnns for autonomous driving. In ICRA, 2018. 2", + "[6] Kieran Browne and Ben Swift. Semantics and explanation: why counterfactual explanations produce adversarial examples in deep neural networks. CoRR, abs/2012.10076, 2020. 2", + "[7] Chun-Hao Chang, Elliot Creager, Anna Goldenberg, and David Duvenaud. Explaining image classifiers by counterfactual generation. In ICLR, 2019. 1", + "[8] Chaofan Chen, Oscar Li, Daniel Tao, Alina Barnett, Cynthia Rudin, and Jonathan Su. This looks like that: Deep learning for interpretable image recognition. In NeurIPS, 2019. 2", + "[9] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. CoRR, abs/1706.05587, 2017. 7", + "[10] Ricky T. Q. Chen, Xuechen Li, Roger Grosse, and David Duvenaud. Isolating sources of disentanglement in variational autoencoders. In NeurIPS, 2018. 2", + "[11] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 3", + "[12] Prafulla Dhariwal and Alexander Quinn Nichol. Diffusion models beat gans on image synthesis. In NeurIPS, 2021. 2", + "[13] Finale Doshi-Velez and Been Kim. Towards a rigorous science of interpretable machine learning. arXiv preprint arXiv:1702.08608, 2017. 7", + "[14] Dave Epstein, Taesung Park, Richard Zhang, Eli Shechtman, and Alexei A. Efros. Blobgan: Spatially disentangled scene representations. In ECCV, 2022. 1, 4", + "[15] Thomas Fel, Julien Colin, Rémi Cadène, and Thomas Serre. What I cannot predict, I do not understand: A human-centered evaluation framework for explainability methods. In NeurIPS, 2022. 2, 7", + "[16] Ruth C. Fong and Andrea Vedaldi. Interpretable explanations of black boxes by meaningful perturbation. In ICCV, 2017. 2, 7", + "[17] Timo Freiesleben. Counterfactual explanations & adversarial examples - common grounds, essential differences, and potential transfers. CoRR, abs/2009.05487, 2020. 2", + "[18] Nicholas Frosst and Geoffrey E. Hinton. Distilling a neural network into a soft decision tree. In Workshop on Comprehensibility and Explanation in AI and ML @AI*IA, 2017. 2" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[19] Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. In ICLR, 2015. 2", + "[20] Yash Goyal, Ziyan Wu, Jan Ernst, Dhruv Batra, Devi Parikh, and Stefan Lee. Counterfactual visual explanations. In ICML, 2019. 2", + "[21] Michael Harradon, Jeff Druce, and Brian E. Ruttenberg. Causal learning and explanation of deep neural networks via autoencoded activations. CoRR, 2018. 2", + "[22] Peter Hase and Mohit Bansal. Evaluating explainable AI: which algorithmic explanations help users predict model behavior? In ACL, 2020. 7", + "[23] Lisa Anne Hendricks, Ronghang Hu, Trevor Darrell, and Zeynep Akata. Grounding visual explanations. In ECCV, 2018. 2", + "[24] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. 5", + "[25] Lukas Hoyer, Maurizio Munoz, Prateek Katiyar, Anna Khoreva, and Volker Fischer. Grid saliency for context explanations of semantic segmentation. In NeurIPS, 2019. 7", + "[26] Gao Huang, Zhuang Liu, Laurens van der Maaten, and Kilian Q. Weinberger. Densely connected convolutional networks. In CVPR, 2017. 5", + "[27] Paul Jacob, Éloi Zablocki, Hedi Ben-Younes, Mickaël Chen, Patrick Pérez, and Matthieu Cord. STEEX: steering counterfactual explanations with semantics. In ECCV, 2022. 2, 3, 5, 6", + "[28] Guillaume Jeanneret, Loic Simon, and Frédéric Jurie. Diffusion models for counterfactual explanations. CoRR, abs/2203.15636, 2022. 1, 2, 3, 8", + "[29] Saeed Khorram and Fuxin Li. Cycle-consistent counterfactuals by latent transformations. In CVPR, 2022. 1, 2", + "[30] Been Kim, Oluwasanmi Koyejo, and Rajiv Khanna. Examples are not enough, learn to criticize! criticism for interpretability. In NeurIPS, 2016. 7", + "[31] Jinkyu Kim and John F. Canny. Interpretable learning for self-driving cars by visualizing causal attention. In ICCV, 2017. 2", + "[32] Oran Lang, Yossi Gandelsman, Michal Yarom, Yoav Wald, Gal Elidan, Avinatan Hassidim, William T. Freeman, Phillip Isola, Amir Globerson, Michal Irani, and Inbar Mosseri. Explaining in style: Training a GAN to explain a classifier in stylespace. In ICCV, 2021. 2", + "[33] Zhiheng Li and Chenliang Xu. Discover the Unknown Biased Attribute of an Image Classifier. In ICCV, 2021. 2", + "[34] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In ICCV, 2015. 2", + "[35] Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, and Pascal Frossard. Deepfool: A simple and accurate method to fool deep neural networks. In CVPR, 2016. 2", + "[36] Giang Nguyen, Daeyoung Kim, and Anh Nguyen. The effectiveness of feature attribution methods and its correlation with automatic evaluation scores. In NeurIPS, 2021. 7" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "15070", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] Michael Niemeyer and Andreas Geiger. GIRAFFE: representing scenes as compositional generative neural feature fields. In CVPR, 2021. 4", + "[38] Martin Pawelczyk, Shalmali Joshi, Chirag Agarwal, Sohini Upadhyay, and Himabindu Lakkaraju. On the connections between counterfactual explanations and adversarial examples. CoRR, abs/2106.09992, 2021. 2", + "[39] Sylvestre-Alvise Rebuffi, Ruth Fong, Xu Ji, and Andrea Vedaldi. There and back again: Revisiting backpropagation saliency methods. In CVPR, 2020. 2", + "[40] Marco Túlio Ribeiro, Sameer Singh, and Carlos Guestrin. \"why should I trust you?\": Explaining the predictions of any classifier. In SIGKDD, 2016. 2", + "[41] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: A stylegan encoder for image-to-image translation. In CVPR, 2021. 4", + "[42] Pau Rodríguez, Massimo Caccia, Alexandre Lacoste, Lee Zamparo, Issam H. Laradji, Laurent Charlin, and David Vázquez. Beyond trivial counterfactual explanations with diverse valuable explanations. In ICCV, 2021. 1, 2, 3, 8", + "[43] Axel Sauer, Nikolay Savinov, and Andreas Geiger. Conditional affordance learning for driving in urban environments. In CoRL, 2018. 2", + "[44] Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In ICCV, 2017. 2", + "[45] Yuan Shen, Shanduojiao Jiang, Yanlin Chen, Eileen Yang, Xilun Jin, Yuliang Fan, and Katie Driggs Campbell. To explain or not to explain: A study on the necessity of explanations for autonomous vehicles. CoRR, 2020. 2", + "[46] Sumedha Singla, Brian Pollack, Junxiang Chen, and Kayhan Batmanghelich. Explanation by progressive exaggeration. In ICLR, 2020. 1, 2, 8", + "[47] Mukund Sundararajan, Ankur Taly, and Qiqi Yan. Axiomatic attribution for deep networks. In ICML, 2017. 2", + "[48] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian J. Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In *ICLR*, 2014. 2", + "[49] Erico Tjoa and Cuntai Guan. A survey on explainable artificial intelligence (XAI): towards medical XAI. CoRR, abs/1907.07374, 2019. 2", + "[50] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for stylegan image manipulation. ACM Transactions on Graphics (TOG), 2021. 4", + "[51] Simon Vandenhende, Dhruv Kumar Mahajan, Filip Radenovic, and Deepti Ghadiyaram. Making heads or tails: Towards semantically consistent visual counterfactuals. In ECCV, 2022. 2", + "[52] Sahil Verma, John P. Dickerson, and Keegan Hines. Counterfactual explanations for machine learning: A review. CoRR, abs/2010.10596, 2020. 1" + ], + "bbox": [ + 78, + 90, + 468, + 852 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[53] Tom Vermeire, Dieter Brughmans, Sofie Goethals, Raphael Mazzine Barbossa de Oliveira, and David Martens. Explainable image classification with evidence counterfactual. Pattern Anal. Appl., 2022. 2", + "[54] Kira Vinogradova, Alexandr Dibrov, and Gene Myers. Towards interpretable semantic segmentation via gradient-weighted class activation mapping (student abstract). In AAAI, 2020. 7", + "[55] Sandra Wachter, Brent Mittelstadt, and Chris Russell. Counterfactual explanations without opening the black box: Automated decisions and the gdpr. Harvard Journal of Law & Technology, 2017. 1, 2", + "[56] Jörg Wagner, Jan Mathias Kohler, Tobias Gindele, Leon Hetzel, Jakob Thaddäus Wiedemer, and Sven Behnke. Interpretable and fine-grained visual explanations for convolutional neural networks. In CVPR, 2019. 2", + "[57] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 3", + "[58] Pei Wang and Nuno Vasconcelos. SCOUT: self-aware discriminant counterfactual explanations. In CVPR, 2020. 2", + "[59] Yiran Xu, Xiaoyin Yang, Lihang Gong, Hsuan-Chu Lin, TzYing Wu, Yunsheng Li, and Nuno Vasconcelos. Explainable object-induced action decision for autonomous vehicles. In CVPR, 2020. 2, 5, 8", + "[60] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darryll. BDD100K: A diverse driving dataset for heterogeneous multitask learning. In CVPR, 2020. 2, 5", + "[61] Éloi Zablocki, Hédi Ben-Younes, Patrick Pérez, and Matthieu Cord. Explainability of deep vision-based autonomous driving systems: Review and challenges. Int. J. Comput. Vis., 2022. 2", + "[62] Matthew D. Zeiler and Rob Fergus. Visualizing and understanding convolutional networks. In ECCV, 2014. 2", + "[63] Quanshi Zhang, Ying Nian Wu, and Song-Chun Zhu. Interpretable convolutional neural networks. In CVPR, 2018. 2", + "[64] Qiaoning Zhang, X. Jessie Yang, and Lionel Peter Robert. Expectations and trust in automated vehicles. In CHI, 2020. 2", + "[65] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 5", + "[66] Bolei Zhou, Aditya Khosla, Ågata Lapedriza, Aude Oliva, and Antonio Torralba. Object detectors emerge in deep scene cnns. In ICLR, 2015. 2", + "[67] Jiapeng Zhu, Yujun Shen, Deli Zhao, and Bolei Zhou. Indomain gan inversion for real image editing. In ECCV, 2020. 4", + "[68] Peihao Zhu, Rameen Abdal, Yipeng Qin, and Peter Wonka. SEAN: image synthesis with semantic region-adaptive normalization. In CVPR, 2020. 3" + ], + "bbox": [ + 501, + 90, + 890, + 852 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "15071", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/OCTET_ Object-Aware Counterfactual Explanations/03950ee8-92a4-474c-99c2-49607c7368f1_model.json b/2023/OCTET_ Object-Aware Counterfactual Explanations/03950ee8-92a4-474c-99c2-49607c7368f1_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2aa387d217fe2ad5513ec461f9933ca82195c326 --- /dev/null +++ b/2023/OCTET_ Object-Aware Counterfactual Explanations/03950ee8-92a4-474c-99c2-49607c7368f1_model.json @@ -0,0 +1,2101 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.809, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.219, + 0.131, + 0.753, + 0.154 + ], + "angle": 0, + "content": "OCTET: Object-aware Counterfactual Explanations" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.18, + 0.899, + 0.22 + ], + "angle": 0, + "content": "Mehdi Zemni1, Mickaël Chen1, Éloi Zablocki1, Hédi Ben-Younes1, Patrick Pérez1, Matthieu Cord1,2 \n1 Valeo.ai, Paris, France \n2 Sorbonne Université, Paris, France" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.255, + 0.314, + 0.272 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.074, + 0.287, + 0.474, + 0.712 + ], + "angle": 0, + "content": "Nowadays, deep vision models are being widely deployed in safety-critical applications, e.g., autonomous driving, and explainability of such models is becoming a pressing concern. Among explanation methods, counterfactual explanations aim to find minimal and interpretable changes to the input image that would also change the output of the model to be explained. Such explanations point end-users at the main factors that impact the decision of the model. However, previous methods struggle to explain decision models trained on images with many objects, e.g., urban scenes, which are more difficult to work with but also arguably more critical to explain. In this work, we propose to tackle this issue with an object-centric framework for counterfactual explanation generation. Our method, inspired by recent generative modeling works, encodes the query image into a latent space that is structured in a way to ease object-level manipulations. Doing so, it provides the end-user with control over which search directions (e.g., spatial displacement of objects, style modification, etc.) are to be explored during the counterfactual generation. We conduct a set of experiments on counterfactual explanation benchmarks for driving scenes, and we show that our method can be adapted beyond classification, e.g., to explain semantic segmentation models. To complete our analysis, we design and run a user study that measures the usefulness of counterfactual explanations in understanding a decision model. Code is available at https://github.com/valeoaai/OCTET." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.739, + 0.21, + 0.755 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Deep learning models are now being widely deployed, notably in safety-critical applications such as autonomous driving. In such contexts, their black-box nature is a major concern, and explainability methods have been developed to improve their trustworthiness. Among them, counterfactual explanations have recently emerged to provide insights into a model's decision [7, 52, 55]. Given a decision model and an input query, a counterfactual explanation is a data point that differs minimally but meaningfully from the query in" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.255, + 0.695, + 0.343 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.256, + 0.888, + 0.343 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.344, + 0.695, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.344, + 0.888, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.434, + 0.895, + 0.546 + ], + "angle": 0, + "content": "Figure 1. Counterfactual explanations generated by OCTET. Given a classifier that predicts whether or not it is possible to go left, and a query image (top left), OCTET produces a counterfactual explanation where the most influential features that led to the decision are changed (top right). On the bottom row, we show that OCTET can also operate under different settings that result in different focused explanations. We report the prediction made by the decision model at the top left of each image." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.58, + 0.895, + 0.807 + ], + "angle": 0, + "content": "a way that changes the output decision of the model. By looking at the differences between the query and the explanation, a user is able to infer — by contrast — which elements were essential for the model to come to its decision. However, most counterfactual methods have only shown results for explaining classifiers trained on single-object images such as face portraits [4, 28, 29, 42, 46]. Aside from the technical difficulties of scaling up the resolution of the images, explaining decision models trained on scenes composed of many objects also present the challenge that those decisions are often multi-factorial. In autonomous driving, for example, most decisions have to take into account the position of all other road users, as well as the layout of the road and its markings, the traffic light and signs, the overall visibility, and many other factors." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.895, + 0.903 + ], + "angle": 0, + "content": "In this paper, we present a new framework, dubbed OCTET for Object-aware Counterfactual ExplanaTions, to generate counterfactual examples for autonomous driving. We leverage recent advances in unsupervised compositional generative modeling [14] to provide a flexible explanation method. Exploiting such a model as our backbone, we can" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "15062" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.136 + ], + "angle": 0, + "content": "assess the contribution of each object of the scene independently and look for explanations in relation to their positions, their styles, or combinations of both." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.137, + 0.468, + 0.332 + ], + "angle": 0, + "content": "To validate our claims, extensive experiments are conducted for self-driving action decision models trained with the BDD100k dataset [60] and its BDD-OIA extension [59]. Fig. 1 shows counterfactual explanations found by OCTET: given a query image and a visual decision system trained to assess if the ego vehicle is allowed to go left or not, OCTET proposes a counterfactual example with cars parked on the left side that are moved closer. When inspecting specific items (bottom row of the figure), OCTET finds that moving the yellow car to the left, or adding a double line marking on the road, are ways to change the model's decision. These explanations highlight that the absence of such elements on the left side of the road heavily influenced the model." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.334, + 0.373, + 0.348 + ], + "angle": 0, + "content": "To sum up, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.349, + 0.468, + 0.408 + ], + "angle": 0, + "content": "- We tackle the problem of building counterfactual explanations for visual decision models operating on complex compositional scenes. We specifically target autonomous driving visual scenarios." + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.41, + 0.468, + 0.453 + ], + "angle": 0, + "content": "- Our method is also a tool to investigate the role of specific objects in a model's decision, empowering the user with control over which type of explanation to look for." + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.455, + 0.468, + 0.543 + ], + "angle": 0, + "content": "- We thoroughly evaluate the realism of our counterfactual images, the minimality and meaningfulness of changes, and compare against previous reference strategies. Beyond explaining classifiers, we also demonstrate the versatility of our method by addressing explanations for a segmentation network." + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.545, + 0.468, + 0.635 + ], + "angle": 0, + "content": "- Finally, we conduct a user-centered study to assess the usefulness of our explanations in a practical case. As standard evaluation benchmarks for counterfactual explanations are lacking a concrete way to measure the interpretability of the explanations, our user-centered study is a key element to validate the presented pipeline." + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.349, + 0.468, + 0.635 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.649, + 0.364, + 0.666 + ], + "angle": 0, + "content": "2. Background and Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Local explanations. The overwhelming majority of deep learning based models are designed without explainability in mind, with only a few notable exceptions [8, 63]. This fact has prompted an interest in post-hoc explanations of already trained models that can be useful to analyze corner cases, understand failures, and find biases [15, 40]. In safety-critical applications, e.g., autonomous driving or medical imaging, explanations are especially needed for liability purposes and to foster end-user trust [45, 49, 61, 64]. Post-hoc explanations are global if they provide a holistic view of the main decision factors driving the model [18, 21, 32, 33], or they are local if they target the understanding of the model behavior on a specific input [40, 44]. Historically, the vast majority of local explanation methods are attribution-based: they generate saliency maps high-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.318 + ], + "angle": 0, + "content": "lighting pixels or regions influencing the most the model's decision [5, 16, 39, 44, 47, 53, 56, 62, 66]. In the case of urban scenes, a saliency method would for instance point at the traffic light, or the presence of a pedestrian to explain why a driving model stops [31, 43]. However, saliency explanations may be misleading as they can be independent of the model at hand and merely act as edge detectors [3]. Besides, saliency methods are by nature restricted to show where the important regions are and they cannot indicate what in these regions was deemed decisive. For example, beyond highlighting things, it would be useful to know if the color of the light (red or green?) or the orientation of the pedestrian's body (towards or away from the car?) are taken into account in the model's decision. Counterfactual explanations are a step in that direction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.325, + 0.892, + 0.521 + ], + "angle": 0, + "content": "Counterfactual explanations. For a given decision model and query input, a counterfactual explanation is defined as a minimally but meaningfully modified version of the query that changes the decision of the model [55]. The notions of 'minimal' and 'meaningful' are crucial here. Adversarial attacks [19, 35, 48] for instance are minimal but not meaningful perturbations: they are imperceptible by design and thus not informative for the end-user [6, 17, 38]. Conversely, if not minimal, the explanation might exhibit unnecessary changes that would hinder its interpretability. While originally developed for models operating on tabular data [55], counterfactuals have been recently considered to explain deep computer vision models [20, 27, 42, 46]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.522, + 0.892, + 0.643 + ], + "angle": 0, + "content": "Seminal work [23] explored the use of natural language to produce descriptions of elements that, in case of presence, change the image classifications. Later, retrieval-based methods aimed to explain decisions by mining the dataset for counterfactual explanations [20, 23, 51, 58]. While obtained explanations can be meaningful, the difference to the query can only be as minimal as the dataset granularity allows, limiting overall interpretability." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.645, + 0.892, + 0.901 + ], + "angle": 0, + "content": "More recently, generator-based counterfactual explanations have been proposed [4, 27-29, 42]. They exploit the power of deep generative models that map a latent space to the data distribution. By optimizing for small manipulations in the latent space, these methods ensure for both the minimality and the meaningfulness of changes. The properties of the generative backbone and the way the latent space is structured have a crucial impact on the resulting counterfactual explanations. DIVE [42], for instance, exploits the disentangled latent space of a \\(\\beta\\)-TCVAE [10] to generate diverse counterfactual explanations. However, the low visual quality of its counterfactuals hinders interpretability. DIME [28] and DVCE [4] instead build on recently popularized denoising diffusion probabilistic models [12], achieving better-looking results but losing the more curated control on the diversity. These works can explain classifiers working on 'simple' images, e.g., face portraits [34] or" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15063" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.089, + 0.891, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.356, + 0.895, + 0.414 + ], + "angle": 0, + "content": "Figure 2. Overview of OCTET. From the query image (top left), a trained encoder \\( E \\) provides initial blob parameters. The blobs are refined in optimization stage ① using Eq. 3 to better encode the query. The refined blob parameters are again modified in optimization stage ② using Eq. 1 to obtain a counterfactual example. The parameters of the encoder \\( E \\) and generator \\( G \\) are frozen during the optimization, as is \\( M \\) the decision model to be explained. Losses are illustrated with dashed arrows whose directions represent the gradient flow." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.439, + 0.384, + 0.454 + ], + "angle": 0, + "content": "featuring a single and centered object [11, 57]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.454, + 0.474, + 0.742 + ], + "angle": 0, + "content": "Scaling up to classifiers operating over more complex images such as crowded urban scenes remains a major challenge. To date, only STEEX [27] proposes a method able to explain decision models dealing with scenes composed of many objects. It employs a segmentation-to-image generation backbone [68], which helps for generating high-quality complex images but comes with two important limitations. Firstly, it requires annotated semantic segmentation masks of images to work, which is not always available or is costly for the user to produce. Secondly, it imposes that the explanation keeps the same spatial and semantic layout as the query, thereby severely restricting the search space. It is, for instance, not able to provide insights about the importance of the position of objects. By contrast, OCTET does not impose a fixed spatial layout. Indeed, the structural layout of the scene is an intermediate representation of the generator, inferred from images only. Therefore, we do not use additional annotations, and the spatial layout can be optimized by back-propagation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.755, + 0.466, + 0.773 + ], + "angle": 0, + "content": "3. Object-aware Counterfactual Explanations" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.471, + 0.903 + ], + "angle": 0, + "content": "We present here OCTET, a method to produce counterfactual examples for compositional scenes, based on a generative architecture. We first formally describe our approach and objective in Sec. 3.1. Then, we instantiate the generative backbone in Sec. 3.2 and the counterfactual objective in Sec. 3.3. Lastly, we explain how we invert the generator to recover the latent code for the query image in Sec. 3.4. The pipeline of OCTET is shown in Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.438, + 0.686, + 0.453 + ], + "angle": 0, + "content": "3.1. Goal and Notations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.461, + 0.895, + 0.629 + ], + "angle": 0, + "content": "Given a decision model \\( M \\), a query image \\( x^{q} \\), and a target decision \\( y \\neq M(x^{q}) \\), a counterfactual explanation is an image \\( x \\) that is close to the query image \\( x^{q} \\) but modified so that the decision \\( M(x) \\) of the model changes to \\( y \\). Generative counterfactual methods, such as ours, employ a trained generator \\( G \\) [27, 28, 42]. We denote its latent space \\( \\mathcal{Z} \\), with \\( z^{q} \\in \\mathcal{Z} \\) the latent code that produces the best approximation of \\( x^{q} \\) when passed through \\( G \\), i.e., so that \\( G(z^{q}) \\approx x^{q} \\). To explain differentiable deep vision models, the definition of a counterfactual explanation stated here-above can be translated into the following optimization problem:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.639, + 0.892, + 0.662 + ], + "angle": 0, + "content": "\\[\n\\underset {z \\in \\mathcal {Z}} {\\operatorname {a r g m i n}} L _ {\\text {d e c i s i o n}} \\left(M (G (z)), y\\right) + \\lambda_ {\\text {d i s t}} L _ {\\text {d i s t}} \\left(z, z ^ {q}\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.672, + 0.895, + 0.84 + ], + "angle": 0, + "content": "The role of the first term \\( L_{\\mathrm{decision}} \\) is to push the decision of the model \\( M \\) for the generated image \\( G(z) \\) to the target class \\( y \\); The second term \\( L_{\\mathrm{dist}} \\) is a distance term in the latent space of \\( G \\), which ensures that the query \\( x^q \\) and the explanation \\( G(z) \\) remain similar; \\( \\lambda_{\\mathrm{dist}} \\) is an hyper-parameter that controls the relative weight between the two terms. With optimal \\( z \\), \\( G(z) \\) then recovers a counterfactual image for the query \\( x^q \\) and the decision model \\( M \\). Since the choice of \\( L_{\\mathrm{decision}} \\) is constrained by the nature of the model to explain, e.g., a classification loss if \\( M \\) is a classifier, we only have to specify \\( G \\) and \\( L_{\\mathrm{dist}} \\) to define our counterfactual method." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.848, + 0.812, + 0.865 + ], + "angle": 0, + "content": "3.2. Compositional generative backbone" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.894, + 0.902 + ], + "angle": 0, + "content": "In this work, we focus on explaining decision models that operate on compositional visual scenes and we look for" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "15064" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.212 + ], + "angle": 0, + "content": "a generative model capable of modeling such scenes. However, most generative models are designed for images containing centered, isolated objects, and they do not learn a latent space that can easily be disentangled into per-object representations. Instead, we propose that our generative backbone should be object-based by design [14, 37]. Such methods are fully differentiable and trained without supervision to generate images in a compositional fashion." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.213, + 0.471, + 0.531 + ], + "angle": 0, + "content": "We choose to build on BlobGAN [14], a compositional generative model that shows good performances and has the advantage of allowing not only edition but also insertion and removal of objects in a differentiable way. In BlobGAN, images are generated in two steps. First, a layout network maps a random Gaussian noise vector to an intermediary representation consisting of an ordered set of \\( K \\) 'blobs'. Each blob \\( k \\in \\{1 \\dots K\\} \\) is defined by spatial parameters \\( \\phi_k \\) and a style feature vector \\( \\psi_k \\). The spatial parameters \\( \\phi_k \\) consist of five values that are its center coordinates, scale, aspect ratio, and rotation angle. In practice, the layout network is a neural network that outputs vectors \\( \\phi_k \\) and \\( \\psi_k \\) for each blob. Then, in the second stage, the generator network is tasked to transform the blob parameters into an image. A rendering layer draws the blobs on a canvas as ellipses whose shapes and positions are given directly by \\( \\phi_k \\). The output of the rendering layer is a map that associates to each 2D spatial location a feature vector computed from the style vectors of the blobs present at that position. This feature map is transformed by convolutional layers into an image that roughly follows the spatial organization of the canvas." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.532, + 0.47, + 0.651 + ], + "angle": 0, + "content": "The layout network and the generator network are trained end-to-end against a discriminator with standard GAN losses only. In particular, we do not use any annotations for the number, size, class, or location of objects. They are discovered by the generator during training. Note also that scale parameters can take non-positive values. In that case, the object is considered absent and the corresponding blob is not rendered on the canvas in the generator network." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.664, + 0.392, + 0.681 + ], + "angle": 0, + "content": "3.3. Setting \\(L_{\\mathrm{dist}}\\) for compositional scenes" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.469, + 0.884 + ], + "angle": 0, + "content": "Using a compositional generative backbone allows us to design a meaningful, disentangled, per-object distance for images. Indeed, to be interpretable with respect to the query image, a counterfactual explanation should display sparse changes in terms of the number of moved, added, removed, or modified objects. Therefore, we design \\( L_{\\mathrm{dist}} \\) such that it is split into per-object distances where each object can be addressed independently. Moreover, we propose that modifications of the spatial position of an object should be orthogonal to those made to its shape or colors. Since in a well-trained BlobGAN each blob is associated with a distinct element of the scene (see Appendix), such a distance can be naturally defined in the space of blob parameters." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Formally, let us denote \\(\\phi\\) and \\(\\psi\\) the concatenation across" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "all blobs \\( k \\in \\{1 \\dots K\\} \\) of spatial parameters \\( \\phi_k \\) and style vectors \\( \\psi_k \\), and \\( z = (\\phi, \\psi) \\) the complete latent representation that the generator network \\( G \\) can decode into an image. Accordingly, the latent code for the query image \\( x^q \\) reads \\( z^q = (\\phi^q, \\psi^q) \\). Then, OCTET implements \\( L_{\\mathrm{dist}} \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.54, + 0.181, + 0.892, + 0.222 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {d i s t}} (z, z ^ {q}) = \\sum_ {k = 1} ^ {K} \\| \\phi_ {k} - \\phi_ {k} ^ {q} \\| _ {1} + \\| \\psi_ {k} - \\psi_ {k} ^ {q} \\| _ {1}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.235, + 0.892, + 0.326 + ], + "angle": 0, + "content": "We use the \\( L_{1} \\) distance here to promote sparsity in the number of modified blobs and the type of edits applied to each blob. If the user wishes to inspect the role of the style of objects, their spatial positions, or both, with a counterfactual explanation, she can instead optimize Eq. 2 only on the corresponding features of the blobs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.327, + 0.892, + 0.387 + ], + "angle": 0, + "content": "In order to specify Eq. 2, and thus the final counterfactual objective (Eq. 1), it remains to be solved how to obtain the blob-wise latent codes \\(\\phi_k^q\\) and \\(\\psi_k^q\\) associated with the query image \\(x^q\\). We discuss this matter in the next section." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.4, + 0.736, + 0.416 + ], + "angle": 0, + "content": "3.4. Inverting the query image" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.424, + 0.892, + 0.545 + ], + "angle": 0, + "content": "Obtaining the latent code corresponding to the query image, a process called inversion, is not straightforward when using a GAN-based generator. To produce a counterfactual example, such inversion needs to find latent codes \\((\\phi^q,\\psi^q)\\) so that \\(G(\\phi^q,\\psi^q)\\) recovers \\(x^{q}\\). For the downstream counterfactual optimization to make sense, the decision of the model for the query image should be preserved for the reconstructed image, i.e., \\(M(G(\\phi^q,\\psi^q)) = M(x^q)\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.546, + 0.892, + 0.651 + ], + "angle": 0, + "content": "To achieve those objectives, we follow best practices for image inversion [1,2,41,50,67]. We first learn an encoder \\( E \\) to predict blob parameters from input images. Generating an image \\( G(E(x^q)) \\) with parameters predicted this way yields a coarse reconstruction of \\( x^q \\). To improve the inversion, we start from predicted blob parameters \\( E(x^q) \\) and further optimize them with the following objective:" + }, + { + "type": "equation", + "bbox": [ + 0.501, + 0.665, + 0.891, + 0.721 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\phi^ {q}, \\psi^ {q} = \\underset {\\phi , \\psi} {\\arg \\min } L _ {\\mathrm {L P I P S}} (G (\\phi , \\psi), x ^ {q})) + L _ {2} (G (\\phi , \\psi), x ^ {q}) \\\\ + L _ {2} \\left(f _ {M} \\left(x ^ {q}\\right), f _ {M} (G (\\phi , \\psi))\\right) \\\\ + L _ {2} \\left(\\left(\\phi , \\psi\\right), E \\left(x ^ {q}\\right)\\right). \\tag {3} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The first two terms ensure that the image \\( G(\\phi, \\psi) \\) generated from the blob parameters reconstructs the query \\( x^q \\) using a perceptual \\( L_{\\mathrm{LPIPS}} \\) loss and the \\( L_2 \\) loss. The third term aims to preserve the features that are important to the decision model \\( M \\) by adding a \\( L_2 \\) loss on intermediate features \\( f_M(x) \\) of \\( M \\). Finally, the last term encourages the parameters to stay close to the reasonably good solution given by the encoder. This inversion method concludes the specification of the optimization problem of Eq. 1 in OCTET and of the pipeline depicted in Fig. 2. More details on the inversion and the training of the encoder are in the appendix." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15065" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.079, + 0.089, + 0.283, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.089, + 0.492, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.492, + 0.089, + 0.691, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.691, + 0.09, + 0.89, + 0.413 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.42, + 0.893, + 0.492 + ], + "angle": 0, + "content": "Figure 3. Counterfactual explanations on driving scenes. Explanations are generated for binary classifiers predicting the capacity of the ego-car to respectively Move forward, Stop, turn Left, or Right. We indicate at the top left of each image the output decision of the model for that image. For each row, counterfactuals target the opposite decision to that obtained for the query image. OCTET finds interpretable, sparse, and meaningful semantic and spatial modifications to the query image. OCTET style-only only optimizes on the style features, it is analog to STEEX [27] that can only search for changes in object appearance to flip the decision." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.515, + 0.21, + 0.533 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.54, + 0.441, + 0.557 + ], + "angle": 0, + "content": "4.1. Experimental protocol and training details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.564, + 0.473, + 0.806 + ], + "angle": 0, + "content": "Data and decision model. Our method is designed to build counterfactual examples to explain decision models trained on compositional images. To evaluate our method, we follow previous work [27] and use the BDD100k dataset [60], which contains 100k images of diverse driving scenes that we resize to a resolution of \\(512 \\times 256\\). In addition, we also use BDD-OIA [59], a 20K scene extension of BDD100k that is annotated with 4 binary attributes labeled Move Forward, Stop, Turn Left, and Turn Right. Each label is defined as the capacity in the given situation of the ego-vehicle to perform the corresponding action. Note that by this definition, being allowed to Stop and to Move Forward are not mutually exclusive. The decision models being explained in the experiments are multi-label binary DenseNet121 [26] classifiers trained on BDD-OIA with the four aforementioned labels." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Baseline. We compare our method against STEEX [27], the only counterfactual explanation method that performs well on compositional images like driving scenes. Note that in STEEX, the conservation of the semantic layout is hardcoded in the architecture: it is less flexible as an explanation method than ours. Moreover, the authors use semantic seg" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.517, + 0.892, + 0.562 + ], + "angle": 0, + "content": "mentation annotations of BDD100k to train their generative backbone, while we do not need any additional annotation. We use the pre-trained model from the official release." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.566, + 0.893, + 0.779 + ], + "angle": 0, + "content": "Technical setup. For our method, we train the generative backbone (BlobGAN) on the training set of BDD100k, with a number \\( K = 40 \\) of blobs. We adapt the architecture of BlobGAN to generate rectangular \\( 512 \\times 256 \\) images, and we decrease the size of the feature vectors of the BlobGAN to compensate for the increased number of blobs. Throughout our experiments, we evaluate two main variants of our method: OCTET where both spatial and style features are optimized altogether and OCTET style-only where only the style vector \\( \\psi \\) is optimized while the spatial vector \\( \\phi \\) is left equal to \\( \\phi^q \\). This setting only assesses the contribution of style features, and is most similar to the STEEX baseline [27] that finds explanations while conserving the semantic layout of the query. More details in the appendix." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.787, + 0.744, + 0.804 + ], + "angle": 0, + "content": "4.2. Quality of the explanations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.893, + 0.903 + ], + "angle": 0, + "content": "Counterfactual explanations are defined as images that are similar to the query image but effectively change the decision of the target model. To verify that our explanations fit the criteria, we evaluate them with the following metrics: 1) the FID [24] which measures if they are realistic images, and 2) a perceptual distance LPIPS [65] that mea" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15066" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.077, + 0.087, + 0.896, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.278, + 0.895, + 0.32 + ], + "angle": 0, + "content": "Figure 4. FID vs. LPIPS, labeled with Success rate for OCTET and STEEX [27] explaining a decision model trained on the 4 labels of BDD-OIA. Data points are obtained with different values of \\(\\lambda_{\\mathrm{dist}}\\) yielding different trade-offs. STEEX is directly comparable to OCTET style-only as they operate in similar conditions. See Sec. 4.2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.34, + 0.473, + 0.763 + ], + "angle": 0, + "content": "sures the change magnitude between them and the query image. We also compute the success rate corresponding to the percentage of counterfactual examples that are indeed classified in the target class by the decision model. We compute these metrics for different values of \\(\\lambda_{\\mathrm{dist}}\\) that lead to different trade-offs and plot the scores in Fig. 4. As STEEX explanations are constrained to style features, we compare it to our OCTET style-only setting. Firstly, in terms of FID, we can observe that both versions of OCTET obtain lower FID than STEEX: OCTET produces explanations of better visual quality, even though it does not use any annotations as opposed to STEEX, which needs semantic maps, and handles spatial explanations that are more challenging to generate. Secondly, in terms of LPIPS, for style-based explanations OCTET style-only stays closer to the query than STEEX. This can be explained as STEEX latent representation is semantic-based, while ours is instance-based. Our method is therefore able to seek finer-grained modifications, while STEEX tends to change the style for an entire semantic class even when only one instance is important to the decision model. As for OCTET, because LPIPS disproportionately weights structural modifications, the wide range of attainable values indicates that many degrees of changes can be obtained depending on the choice of \\(\\lambda_{\\mathrm{dist}}\\). Some settings tend to move objects a lot, while other settings attain values similar to 'style-only' edits. Finally, the very high achievable Success Rates displayed by all three models confirm that they will propose explanations in almost any situation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.47, + 0.901 + ], + "angle": 0, + "content": "We display examples of OCTET's counterfactual explanations in Fig. 3. OCTET finds sparse spatial and/or semantic modifications to objects that strongly influence the output decision. For example, we can see with OCTET that the distance with the car in front plays a role when deciding for 'Forward' (1st row) or 'Stop' (2nd row). On the other hand, OCTET style-only shows that for those same decisions, the brake lights also are taken into account. For going 'Left' class (3rd row), the absence of the continuous line seems" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.337, + 0.703, + 0.413 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.705, + 0.338, + 0.89, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.414, + 0.703, + 0.505 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.705, + 0.414, + 0.89, + 0.505 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.486, + 0.703, + 0.56 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.705, + 0.486, + 0.89, + 0.56 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.575, + 0.892, + 0.631 + ], + "angle": 0, + "content": "Figure 5. Targeted counterfactual explanations. For each query image (top row), we target different blobs (2nd and 3rd row) to inspect the role of corresponding objects. The output of the decision model is reported on the top right corner for each image." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.657, + 0.894, + 0.718 + ], + "angle": 0, + "content": "crucial. On the 3rd and 4th row, we can see that OCTET mixes both style (brake lights) and spatial (proximity) explanations. On the same tasks, STEEX changes the style of the whole class, and not just the necessary instance." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.727, + 0.729, + 0.742 + ], + "angle": 0, + "content": "4.3. Targeted counterfactuals" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Models that operate on compositional scenes can have their decisions influenced by many different factors, and we have shown that OCTET can expose those factors in its counterfactual examples. However, the user may want to test some particular hypothesis dealing with specific objects or areas of the scene. Accordingly, the user can decide to optimize Eq. 2 only on selected blobs parameters. Finding the blob corresponding to a specific object in the image is discussed in Appendix. For instance, in the situation shown at the top left of Fig. 5, the car at the front seems to have" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15067" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.894, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.336, + 0.893, + 0.395 + ], + "angle": 0, + "content": "Figure 6. Counterfactual explanations for a segmentation model. The segmentation model predicts the driveable area with primary and secondary lanes shown in red and blue respectively. The first row shows query images and segmentations by the model. In the second row, we add or remove parts of the original masks to build target masks (changes are shown with white dotted lines). The third row displays counterfactual explanations for the target masks and predictions on the new images. Green ellipses are manually added to highlight details." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.419, + 0.473, + 0.587 + ], + "angle": 0, + "content": "its brake lights on, but the decision is still not to 'Stop'. Targeting the front car (2nd row left image) reveals that the decision would have been 'Stop' if the car was closer, and the brake light more clearly contrasted. OCTET can also target blobs that are currently inactive, corresponding to no visible object in the image. In this example, doing so for car blobs shows that if another vehicle also had brake lights on, the model would also decide to 'Stop' (3rd row left image). This finding is unexpected, and could warrant further investigation of the decision model. We show in Fig. 5 and Fig. 1 some other results we obtain by targeting specific blobs." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.601, + 0.442, + 0.618 + ], + "angle": 0, + "content": "4.4. Explaining a semantic segmentation model" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.627, + 0.47, + 0.779 + ], + "angle": 0, + "content": "For complex tasks, like autonomous driving, decision models usually rely on multiple neural networks that address different subtasks (e.g., traffic sign recognition, depth estimation, motion forecasting). In this context, it is important to explain models beyond classifiers [25,54], a question not addressed in previous counterfactual methods. Here, we show the versatility of our method as we extend it to explain a semantic segmentation network. To do so, we use the same optimization problem described in Eq. 1 and Sec. 3 but we change \\( L_{\\text{decision}} \\) to an image segmentation loss." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.901 + ], + "angle": 0, + "content": "In Fig. 6, we show counterfactual explanations for a driveable area segmentation model. More precisely, OCTET explains outputs of a DeepLabv3 [9] trained to segment images from BDD100k into classes of driveable area: the primary lane, the secondary side lanes, and the background. We define the target masks by manually modifying predictions made on the query images as illustrated in the second row. We see in the obtained counterfactuals, in the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.419, + 0.895, + 0.587 + ], + "angle": 0, + "content": "third row, changes in road markings and vehicles' positions. They confirm that the segmentation model will consider as driveable a lane behind broken white lines (col. 3) but not behind a double yellow line (col. 2), especially if, in addition, the lane is hatched (col. 3). If there is a car on a drivable lane, the model can detect that the occupied portion is non-driveable while considering that the rest of the lane remains accessible. With these examples, OCTET provides explanations for both large semantic regions, e.g., when we add or remove an entire region in the target mask, or small subparts, e.g., when we slightly crop or extend a region." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.601, + 0.792, + 0.618 + ], + "angle": 0, + "content": "4.5. User-study to assess explanations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.627, + 0.892, + 0.809 + ], + "angle": 0, + "content": "Metrics presented in Sec. 4.2 measure how well the produced examples match the definition of counterfactual explanations. We want here to assess how useful the explanations are for a user to better understand a decision model. However, designing such an evaluation is challenging. Indeed, it has to avoid the many biases that appear [13, 36]. Moreover, we usually do not have access to the inner workings of a model, and therefore we cannot compare users' insights on its behavior against a ground truth. To overcome these challenges, we take inspiration from recent work on the usefulness of saliency maps as explanations [15] and we adapt its evaluation protocol to counterfactual explanations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.895, + 0.902 + ], + "angle": 0, + "content": "The goal is to measure if accessing counterfactual explanations on some examples helps users to better understand the model. Grounded within previous literature [16,22,30], we instantiate understand as 'being able to predict the model's output for new instances', a concept known as simulatability. In particular, our user-study has two phases:" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15068" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.078, + 0.089, + 0.468, + 0.146 + ], + "angle": 0, + "content": "
Cohort sizeReplicationBias Detection
Group w/ OCTET2070%65%
Group control2052.5%0%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.156, + 0.47, + 0.239 + ], + "angle": 0, + "content": "Table 1. Results of the user-study. The performance of users who had access to explanations ('w/ OCTET') in the observation phase is compared to that of users who didn't ('control'). 'Replication' measures the accuracy of the users in predicting the model's output on new samples and 'bias detection' is the proportion of users who found the spurious correlation in the model." + }, + { + "type": "text", + "bbox": [ + 0.082, + 0.268, + 0.468, + 0.327 + ], + "angle": 0, + "content": "1. Observation phase: participants are shown a series of triplets, each being composed of (1) a query image, (2) the model's decision on that image, and (3), a counterfactual explanation for the decision." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.329, + 0.468, + 0.388 + ], + "angle": 0, + "content": "2. Questionnaire phase: participants are presented with new images and are asked to guess the output of the model on these images. We stress that during the questionnaire, no counterfactual images are shown." + }, + { + "type": "list", + "bbox": [ + 0.081, + 0.268, + 0.468, + 0.388 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.391, + 0.468, + 0.601 + ], + "angle": 0, + "content": "To measure the impact that counterfactual explanations of the first phase have on the users' ability to predict the model output on new samples (second phase), we use a control group. This control group, composed of different participants, undertakes the exact same tasks except that counterfactual explanations are not shown during the first phase. The decision model used in this study is a classifier trained on BDD-OIA [59] to predict the binary 'Turn Right' label. Besides, the decision model is trained, by design, with a flaw: the presence of a car on any side of the road can impact the decision, not only obstacles on the right side. As discussed further below, this bias is introduced to study if users are able to identify model defects thanks to counterfactuals. More details are provided in the appendix." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.608, + 0.468, + 0.818 + ], + "angle": 0, + "content": "Replication score. We first measure the 'replication score': the proportion of correct match between the user's answer and the prediction of the decision model, averaged over participants of the group. We report these results in Tab. 1 and observe that participants in the group accessing counterfactuals during the observation phase can successfully guess the decision model's prediction \\(70\\%\\) of the time while the control group's performance is \\(52.5\\%\\), barely above random. A statistical unpaired t-test, with as null hypothesis that the counterfactual explanations have no effect, confirms the significance of our result with a p-value of 0.0028. This demonstrates the usefulness of OCTET in understanding the decision model as participants with explanations could better anticipate the model's behavior on new instances." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.468, + 0.9 + ], + "angle": 0, + "content": "Bias detection. Another promise of counterfactual explanations is to enable bias detection. We then want to assess whether, by looking at explanations, users can tell if and how a decision model is biased. In that direction, previous works dealing with human face datasets train biased models" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.53 + ], + "angle": 0, + "content": "by artificially confounding attributes, e.g., smile and age, in the training set. Then, they develop bias detection benchmarks that measure, by means of an oracle neural network, how well attributes remain confounded in the counterfactual explanations [28, 42, 46]. However, a serious limitation of that benchmark is that it keeps end-users out of the evaluation loop and it cannot tell whether or not a user will be able to uncover the spurious correlations by looking at the counterfactual explanation. Moreover, detecting the issue with the model can be especially difficult when the bias is not obvious or its presence not suspected in the first place. Instead of an automated evaluation, at the end of the questionnaire, we asked the participants to describe in free text what they had understood from the model. The goal is to prompt them to mention the issue we embedded in the decision model, i.e., that the presence of a car on the left-hand side of the image factors in the 'Turn Right' decision. The 'bias detection' score simply counts, for each group, the proportion of participants that used the word 'left' in their answers. In Tab. 1, we see that \\(65\\%\\) of the participants that had access to the explanations from OCTET identify the spurious correlation learned by the model, while none of the users that did not have the explanations (control group) mentions it. This clearly shows that our method is useful to pinpoint issues with the decision model. Also, the fact that not every participant detected it confirms our suspicions against automated bias detection benchmarks: the presence of the information in the counterfactual examples is not sufficient to ensure that the issue will be found in practice." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.546, + 0.618, + 0.561 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.572, + 0.892, + 0.768 + ], + "angle": 0, + "content": "In this work, we presented OCTET, a generative method to produce counterfactual explanations for deep vision models working on complex scenes. Thanks to an object-centric representation of the scene, OCTET is able to process complex compositional scenes and to assess the impact of both the spatial positions and the style features of each object. Our tool allows in-depth exploration of the different explaining factors by letting the user adjust on the fly the weights of the different search directions independently. We evaluated our method thoroughly on real driving scenes from BDD100k, with an improved evaluation benchmark compared to previous works, including a human-centered study that confirmed the usefulness of the approach." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.785, + 0.667, + 0.802 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.9 + ], + "angle": 0, + "content": "This work was supported in part by the ANR grants VISA DEEP (ANR-20-CHIA-0022) and MultiTrans (ANR-21-CE23-0032). Authors would like to thank the voluntary participants of the user-study, as well as Remi Cadene, Julien Colin, Thomas Fel, and Guillaume Jeanneret for helpful discussions." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15069" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.156 + ], + "angle": 0, + "content": "[1] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan: How to embed images into the stylegan latent space? In ICCV, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.157, + 0.47, + 0.198 + ], + "angle": 0, + "content": "[2] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan++: How to edit the embedded images? In CVPR, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.2, + 0.471, + 0.242 + ], + "angle": 0, + "content": "[3] Julius Adebayo, Justin Gilmer, Michael Muelly, Ian J. Goodfellow, Moritz Hardt, and Been Kim. Sanity checks for saliency maps. In NeurIPS, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.243, + 0.47, + 0.282 + ], + "angle": 0, + "content": "[4] Maximilian Augustin, Valentyn Boreiko, Francesco Croce, and Matthias Hein. Diffusion visual counterfactual explanations. In NeurIPS, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.284, + 0.471, + 0.34 + ], + "angle": 0, + "content": "[5] Mariusz Bojarski, Anna Choromanska, Krzysztof Choromanski, Bernhard Firner, Larry J Ackel, Urs Muller, Phil Yeres, and Karol Zieba. Visualbackprop: Efficient visualization of cnns for autonomous driving. In ICRA, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.341, + 0.47, + 0.394 + ], + "angle": 0, + "content": "[6] Kieran Browne and Ben Swift. Semantics and explanation: why counterfactual explanations produce adversarial examples in deep neural networks. CoRR, abs/2012.10076, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.396, + 0.471, + 0.437 + ], + "angle": 0, + "content": "[7] Chun-Hao Chang, Elliot Creager, Anna Goldenberg, and David Duvenaud. Explaining image classifiers by counterfactual generation. In ICLR, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.438, + 0.471, + 0.479 + ], + "angle": 0, + "content": "[8] Chaofan Chen, Oscar Li, Daniel Tao, Alina Barnett, Cynthia Rudin, and Jonathan Su. This looks like that: Deep learning for interpretable image recognition. In NeurIPS, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.48, + 0.471, + 0.522 + ], + "angle": 0, + "content": "[9] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. CoRR, abs/1706.05587, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.523, + 0.47, + 0.563 + ], + "angle": 0, + "content": "[10] Ricky T. Q. Chen, Xuechen Li, Roger Grosse, and David Duvenaud. Isolating sources of disentanglement in variational autoencoders. In NeurIPS, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.565, + 0.47, + 0.605 + ], + "angle": 0, + "content": "[11] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.607, + 0.47, + 0.634 + ], + "angle": 0, + "content": "[12] Prafulla Dhariwal and Alexander Quinn Nichol. Diffusion models beat gans on image synthesis. In NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.635, + 0.47, + 0.675 + ], + "angle": 0, + "content": "[13] Finale Doshi-Velez and Been Kim. Towards a rigorous science of interpretable machine learning. arXiv preprint arXiv:1702.08608, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.677, + 0.47, + 0.718 + ], + "angle": 0, + "content": "[14] Dave Epstein, Taesung Park, Richard Zhang, Eli Shechtman, and Alexei A. Efros. Blobgan: Spatially disentangled scene representations. In ECCV, 2022. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.72, + 0.47, + 0.773 + ], + "angle": 0, + "content": "[15] Thomas Fel, Julien Colin, Rémi Cadène, and Thomas Serre. What I cannot predict, I do not understand: A human-centered evaluation framework for explainability methods. In NeurIPS, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.47, + 0.815 + ], + "angle": 0, + "content": "[16] Ruth C. Fong and Andrea Vedaldi. Interpretable explanations of black boxes by meaningful perturbation. In ICCV, 2017. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.47, + 0.859 + ], + "angle": 0, + "content": "[17] Timo Freiesleben. Counterfactual explanations & adversarial examples - common grounds, essential differences, and potential transfers. CoRR, abs/2009.05487, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[18] Nicholas Frosst and Geoffrey E. Hinton. Distilling a neural network into a soft decision tree. In Workshop on Comprehensibility and Explanation in AI and ML @AI*IA, 2017. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[19] Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. In ICLR, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.892, + 0.177 + ], + "angle": 0, + "content": "[20] Yash Goyal, Ziyan Wu, Jan Ernst, Dhruv Batra, Devi Parikh, and Stefan Lee. Counterfactual visual explanations. In ICML, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.221 + ], + "angle": 0, + "content": "[21] Michael Harradon, Jeff Druce, and Brian E. Ruttenberg. Causal learning and explanation of deep neural networks via autoencoded activations. CoRR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.223, + 0.892, + 0.264 + ], + "angle": 0, + "content": "[22] Peter Hase and Mohit Bansal. Evaluating explainable AI: which algorithmic explanations help users predict model behavior? In ACL, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.266, + 0.892, + 0.307 + ], + "angle": 0, + "content": "[23] Lisa Anne Hendricks, Ronghang Hu, Trevor Darrell, and Zeynep Akata. Grounding visual explanations. In ECCV, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.31, + 0.892, + 0.365 + ], + "angle": 0, + "content": "[24] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.368, + 0.892, + 0.409 + ], + "angle": 0, + "content": "[25] Lukas Hoyer, Maurizio Munoz, Prateek Katiyar, Anna Khoreva, and Volker Fischer. Grid saliency for context explanations of semantic segmentation. In NeurIPS, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.411, + 0.892, + 0.451 + ], + "angle": 0, + "content": "[26] Gao Huang, Zhuang Liu, Laurens van der Maaten, and Kilian Q. Weinberger. Densely connected convolutional networks. In CVPR, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.453, + 0.892, + 0.508 + ], + "angle": 0, + "content": "[27] Paul Jacob, Éloi Zablocki, Hedi Ben-Younes, Mickaël Chen, Patrick Pérez, and Matthieu Cord. STEEX: steering counterfactual explanations with semantics. In ECCV, 2022. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.511, + 0.892, + 0.552 + ], + "angle": 0, + "content": "[28] Guillaume Jeanneret, Loic Simon, and Frédéric Jurie. Diffusion models for counterfactual explanations. CoRR, abs/2203.15636, 2022. 1, 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.555, + 0.892, + 0.583 + ], + "angle": 0, + "content": "[29] Saeed Khorram and Fuxin Li. Cycle-consistent counterfactuals by latent transformations. In CVPR, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.585, + 0.892, + 0.626 + ], + "angle": 0, + "content": "[30] Been Kim, Oluwasanmi Koyejo, and Rajiv Khanna. Examples are not enough, learn to criticize! criticism for interpretability. In NeurIPS, 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.628, + 0.892, + 0.668 + ], + "angle": 0, + "content": "[31] Jinkyu Kim and John F. Canny. Interpretable learning for self-driving cars by visualizing causal attention. In ICCV, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.671, + 0.892, + 0.741 + ], + "angle": 0, + "content": "[32] Oran Lang, Yossi Gandelsman, Michal Yarom, Yoav Wald, Gal Elidan, Avinatan Hassidim, William T. Freeman, Phillip Isola, Amir Globerson, Michal Irani, and Inbar Mosseri. Explaining in style: Training a GAN to explain a classifier in stylespace. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.743, + 0.892, + 0.771 + ], + "angle": 0, + "content": "[33] Zhiheng Li and Chenliang Xu. Discover the Unknown Biased Attribute of an Image Classifier. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.773, + 0.892, + 0.813 + ], + "angle": 0, + "content": "[34] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In ICCV, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.816, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[35] Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, and Pascal Frossard. Deepfool: A simple and accurate method to fool deep neural networks. In CVPR, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[36] Giang Nguyen, Daeyoung Kim, and Anh Nguyen. The effectiveness of feature attribution methods and its correlation with automatic evaluation scores. In NeurIPS, 2021. 7" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15070" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.133 + ], + "angle": 0, + "content": "[37] Michael Niemeyer and Andreas Geiger. GIRAFFE: representing scenes as compositional generative neural feature fields. In CVPR, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.134, + 0.47, + 0.19 + ], + "angle": 0, + "content": "[38] Martin Pawelczyk, Shalmali Joshi, Chirag Agarwal, Sohini Upadhyay, and Himabindu Lakkaraju. On the connections between counterfactual explanations and adversarial examples. CoRR, abs/2106.09992, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.191, + 0.469, + 0.233 + ], + "angle": 0, + "content": "[39] Sylvestre-Alvise Rebuffi, Ruth Fong, Xu Ji, and Andrea Vedaldi. There and back again: Revisiting backpropagation saliency methods. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.234, + 0.469, + 0.274 + ], + "angle": 0, + "content": "[40] Marco Túlio Ribeiro, Sameer Singh, and Carlos Guestrin. \"why should I trust you?\": Explaining the predictions of any classifier. In SIGKDD, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.276, + 0.469, + 0.33 + ], + "angle": 0, + "content": "[41] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: A stylegan encoder for image-to-image translation. In CVPR, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.332, + 0.469, + 0.387 + ], + "angle": 0, + "content": "[42] Pau Rodríguez, Massimo Caccia, Alexandre Lacoste, Lee Zamparo, Issam H. Laradji, Laurent Charlin, and David Vázquez. Beyond trivial counterfactual explanations with diverse valuable explanations. In ICCV, 2021. 1, 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.388, + 0.469, + 0.429 + ], + "angle": 0, + "content": "[43] Axel Sauer, Nikolay Savinov, and Andreas Geiger. Conditional affordance learning for driving in urban environments. In CoRL, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.431, + 0.469, + 0.486 + ], + "angle": 0, + "content": "[44] Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In ICCV, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.488, + 0.469, + 0.542 + ], + "angle": 0, + "content": "[45] Yuan Shen, Shanduojiao Jiang, Yanlin Chen, Eileen Yang, Xilun Jin, Yuliang Fan, and Katie Driggs Campbell. To explain or not to explain: A study on the necessity of explanations for autonomous vehicles. CoRR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.543, + 0.469, + 0.584 + ], + "angle": 0, + "content": "[46] Sumedha Singla, Brian Pollack, Junxiang Chen, and Kayhan Batmanghelich. Explanation by progressive exaggeration. In ICLR, 2020. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.585, + 0.469, + 0.613 + ], + "angle": 0, + "content": "[47] Mukund Sundararajan, Ankur Taly, and Qiqi Yan. Axiomatic attribution for deep networks. In ICML, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.614, + 0.469, + 0.655 + ], + "angle": 0, + "content": "[48] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian J. Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In *ICLR*, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.657, + 0.469, + 0.697 + ], + "angle": 0, + "content": "[49] Erico Tjoa and Cuntai Guan. A survey on explainable artificial intelligence (XAI): towards medical XAI. CoRR, abs/1907.07374, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.699, + 0.469, + 0.753 + ], + "angle": 0, + "content": "[50] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for stylegan image manipulation. ACM Transactions on Graphics (TOG), 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.755, + 0.469, + 0.81 + ], + "angle": 0, + "content": "[51] Simon Vandenhende, Dhruv Kumar Mahajan, Filip Radenovic, and Deepti Ghadiyaram. Making heads or tails: Towards semantically consistent visual counterfactuals. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.811, + 0.469, + 0.853 + ], + "angle": 0, + "content": "[52] Sahil Verma, John P. Dickerson, and Keegan Hines. Counterfactual explanations for machine learning: A review. CoRR, abs/2010.10596, 2020. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.853 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[53] Tom Vermeire, Dieter Brughmans, Sofie Goethals, Raphael Mazzine Barbossa de Oliveira, and David Martens. Explainable image classification with evidence counterfactual. Pattern Anal. Appl., 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[54] Kira Vinogradova, Alexandr Dibrov, and Gene Myers. Towards interpretable semantic segmentation via gradient-weighted class activation mapping (student abstract). In AAAI, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.206, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[55] Sandra Wachter, Brent Mittelstadt, and Chris Russell. Counterfactual explanations without opening the black box: Automated decisions and the gdpr. Harvard Journal of Law & Technology, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.263, + 0.892, + 0.318 + ], + "angle": 0, + "content": "[56] Jörg Wagner, Jan Mathias Kohler, Tobias Gindele, Leon Hetzel, Jakob Thaddäus Wiedemer, and Sven Behnke. Interpretable and fine-grained visual explanations for convolutional neural networks. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.321, + 0.892, + 0.361 + ], + "angle": 0, + "content": "[57] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.364, + 0.892, + 0.392 + ], + "angle": 0, + "content": "[58] Pei Wang and Nuno Vasconcelos. SCOUT: self-aware discriminant counterfactual explanations. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.394, + 0.892, + 0.449 + ], + "angle": 0, + "content": "[59] Yiran Xu, Xiaoyin Yang, Lihang Gong, Hsuan-Chu Lin, TzYing Wu, Yunsheng Li, and Nuno Vasconcelos. Explainable object-induced action decision for autonomous vehicles. In CVPR, 2020. 2, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.451, + 0.892, + 0.506 + ], + "angle": 0, + "content": "[60] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darryll. BDD100K: A diverse driving dataset for heterogeneous multitask learning. In CVPR, 2020. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.508, + 0.892, + 0.563 + ], + "angle": 0, + "content": "[61] Éloi Zablocki, Hédi Ben-Younes, Patrick Pérez, and Matthieu Cord. Explainability of deep vision-based autonomous driving systems: Review and challenges. Int. J. Comput. Vis., 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.565, + 0.892, + 0.593 + ], + "angle": 0, + "content": "[62] Matthew D. Zeiler and Rob Fergus. Visualizing and understanding convolutional networks. In ECCV, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.595, + 0.892, + 0.635 + ], + "angle": 0, + "content": "[63] Quanshi Zhang, Ying Nian Wu, and Song-Chun Zhu. Interpretable convolutional neural networks. In CVPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.638, + 0.892, + 0.678 + ], + "angle": 0, + "content": "[64] Qiaoning Zhang, X. Jessie Yang, and Lionel Peter Robert. Expectations and trust in automated vehicles. In CHI, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.681, + 0.892, + 0.723 + ], + "angle": 0, + "content": "[65] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.725, + 0.892, + 0.765 + ], + "angle": 0, + "content": "[66] Bolei Zhou, Aditya Khosla, Ågata Lapedriza, Aude Oliva, and Antonio Torralba. Object detectors emerge in deep scene cnns. In ICLR, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.768, + 0.892, + 0.808 + ], + "angle": 0, + "content": "[67] Jiapeng Zhu, Yujun Shen, Deli Zhao, and Bolei Zhou. Indomain gan inversion for real image editing. In ECCV, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.811, + 0.892, + 0.853 + ], + "angle": 0, + "content": "[68] Peihao Zhu, Rameen Abdal, Yipeng Qin, and Peter Wonka. SEAN: image synthesis with semantic region-adaptive normalization. In CVPR, 2020. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.853 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15071" + } + ] +] \ No newline at end of file diff --git a/2023/OCTET_ Object-Aware Counterfactual Explanations/03950ee8-92a4-474c-99c2-49607c7368f1_origin.pdf b/2023/OCTET_ Object-Aware Counterfactual Explanations/03950ee8-92a4-474c-99c2-49607c7368f1_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..55082abba9afd30a1a22378264db70780d4c70b3 --- /dev/null +++ b/2023/OCTET_ Object-Aware Counterfactual Explanations/03950ee8-92a4-474c-99c2-49607c7368f1_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:861bed5cd9df44727e99fafd2e776e227b682e51db16570eac2f74a34ea8080a +size 7873958 diff --git a/2023/OCTET_ Object-Aware Counterfactual Explanations/full.md b/2023/OCTET_ Object-Aware Counterfactual Explanations/full.md new file mode 100644 index 0000000000000000000000000000000000000000..8813853ec972b0f10daec0df3c59c2068a1312a2 --- /dev/null +++ b/2023/OCTET_ Object-Aware Counterfactual Explanations/full.md @@ -0,0 +1,275 @@ +# OCTET: Object-aware Counterfactual Explanations + +Mehdi Zemni1, Mickaël Chen1, Éloi Zablocki1, Hédi Ben-Younes1, Patrick Pérez1, Matthieu Cord1,2 +1 Valeo.ai, Paris, France +2 Sorbonne Université, Paris, France + +# Abstract + +Nowadays, deep vision models are being widely deployed in safety-critical applications, e.g., autonomous driving, and explainability of such models is becoming a pressing concern. Among explanation methods, counterfactual explanations aim to find minimal and interpretable changes to the input image that would also change the output of the model to be explained. Such explanations point end-users at the main factors that impact the decision of the model. However, previous methods struggle to explain decision models trained on images with many objects, e.g., urban scenes, which are more difficult to work with but also arguably more critical to explain. In this work, we propose to tackle this issue with an object-centric framework for counterfactual explanation generation. Our method, inspired by recent generative modeling works, encodes the query image into a latent space that is structured in a way to ease object-level manipulations. Doing so, it provides the end-user with control over which search directions (e.g., spatial displacement of objects, style modification, etc.) are to be explored during the counterfactual generation. We conduct a set of experiments on counterfactual explanation benchmarks for driving scenes, and we show that our method can be adapted beyond classification, e.g., to explain semantic segmentation models. To complete our analysis, we design and run a user study that measures the usefulness of counterfactual explanations in understanding a decision model. Code is available at https://github.com/valeoaai/OCTET. + +# 1. Introduction + +Deep learning models are now being widely deployed, notably in safety-critical applications such as autonomous driving. In such contexts, their black-box nature is a major concern, and explainability methods have been developed to improve their trustworthiness. Among them, counterfactual explanations have recently emerged to provide insights into a model's decision [7, 52, 55]. Given a decision model and an input query, a counterfactual explanation is a data point that differs minimally but meaningfully from the query in + +![](images/d33776a7a391bf2261a43e245a6f91dce3948c346fb8e8cc627ebdcddf3c8507.jpg) + +![](images/e47858e4c9cc444c0f953f84b0949b80c1da045a5b34ded8728f5484e20ac5d0.jpg) + +![](images/d73cc6e01394921ba2748a6a2fcc1da5b8ef572cb6da23bf7142c222f05c3999.jpg) +Figure 1. Counterfactual explanations generated by OCTET. Given a classifier that predicts whether or not it is possible to go left, and a query image (top left), OCTET produces a counterfactual explanation where the most influential features that led to the decision are changed (top right). On the bottom row, we show that OCTET can also operate under different settings that result in different focused explanations. We report the prediction made by the decision model at the top left of each image. + +![](images/9b87e53b767850201cdde52254fad8d6e04c50f1d066d5192fd8d202da78a202.jpg) + +a way that changes the output decision of the model. By looking at the differences between the query and the explanation, a user is able to infer — by contrast — which elements were essential for the model to come to its decision. However, most counterfactual methods have only shown results for explaining classifiers trained on single-object images such as face portraits [4, 28, 29, 42, 46]. Aside from the technical difficulties of scaling up the resolution of the images, explaining decision models trained on scenes composed of many objects also present the challenge that those decisions are often multi-factorial. In autonomous driving, for example, most decisions have to take into account the position of all other road users, as well as the layout of the road and its markings, the traffic light and signs, the overall visibility, and many other factors. + +In this paper, we present a new framework, dubbed OCTET for Object-aware Counterfactual ExplanaTions, to generate counterfactual examples for autonomous driving. We leverage recent advances in unsupervised compositional generative modeling [14] to provide a flexible explanation method. Exploiting such a model as our backbone, we can + +assess the contribution of each object of the scene independently and look for explanations in relation to their positions, their styles, or combinations of both. + +To validate our claims, extensive experiments are conducted for self-driving action decision models trained with the BDD100k dataset [60] and its BDD-OIA extension [59]. Fig. 1 shows counterfactual explanations found by OCTET: given a query image and a visual decision system trained to assess if the ego vehicle is allowed to go left or not, OCTET proposes a counterfactual example with cars parked on the left side that are moved closer. When inspecting specific items (bottom row of the figure), OCTET finds that moving the yellow car to the left, or adding a double line marking on the road, are ways to change the model's decision. These explanations highlight that the absence of such elements on the left side of the road heavily influenced the model. + +To sum up, our contributions are as follows: + +- We tackle the problem of building counterfactual explanations for visual decision models operating on complex compositional scenes. We specifically target autonomous driving visual scenarios. +- Our method is also a tool to investigate the role of specific objects in a model's decision, empowering the user with control over which type of explanation to look for. +- We thoroughly evaluate the realism of our counterfactual images, the minimality and meaningfulness of changes, and compare against previous reference strategies. Beyond explaining classifiers, we also demonstrate the versatility of our method by addressing explanations for a segmentation network. +- Finally, we conduct a user-centered study to assess the usefulness of our explanations in a practical case. As standard evaluation benchmarks for counterfactual explanations are lacking a concrete way to measure the interpretability of the explanations, our user-centered study is a key element to validate the presented pipeline. + +# 2. Background and Related Work + +Local explanations. The overwhelming majority of deep learning based models are designed without explainability in mind, with only a few notable exceptions [8, 63]. This fact has prompted an interest in post-hoc explanations of already trained models that can be useful to analyze corner cases, understand failures, and find biases [15, 40]. In safety-critical applications, e.g., autonomous driving or medical imaging, explanations are especially needed for liability purposes and to foster end-user trust [45, 49, 61, 64]. Post-hoc explanations are global if they provide a holistic view of the main decision factors driving the model [18, 21, 32, 33], or they are local if they target the understanding of the model behavior on a specific input [40, 44]. Historically, the vast majority of local explanation methods are attribution-based: they generate saliency maps high- + +lighting pixels or regions influencing the most the model's decision [5, 16, 39, 44, 47, 53, 56, 62, 66]. In the case of urban scenes, a saliency method would for instance point at the traffic light, or the presence of a pedestrian to explain why a driving model stops [31, 43]. However, saliency explanations may be misleading as they can be independent of the model at hand and merely act as edge detectors [3]. Besides, saliency methods are by nature restricted to show where the important regions are and they cannot indicate what in these regions was deemed decisive. For example, beyond highlighting things, it would be useful to know if the color of the light (red or green?) or the orientation of the pedestrian's body (towards or away from the car?) are taken into account in the model's decision. Counterfactual explanations are a step in that direction. + +Counterfactual explanations. For a given decision model and query input, a counterfactual explanation is defined as a minimally but meaningfully modified version of the query that changes the decision of the model [55]. The notions of 'minimal' and 'meaningful' are crucial here. Adversarial attacks [19, 35, 48] for instance are minimal but not meaningful perturbations: they are imperceptible by design and thus not informative for the end-user [6, 17, 38]. Conversely, if not minimal, the explanation might exhibit unnecessary changes that would hinder its interpretability. While originally developed for models operating on tabular data [55], counterfactuals have been recently considered to explain deep computer vision models [20, 27, 42, 46]. + +Seminal work [23] explored the use of natural language to produce descriptions of elements that, in case of presence, change the image classifications. Later, retrieval-based methods aimed to explain decisions by mining the dataset for counterfactual explanations [20, 23, 51, 58]. While obtained explanations can be meaningful, the difference to the query can only be as minimal as the dataset granularity allows, limiting overall interpretability. + +More recently, generator-based counterfactual explanations have been proposed [4, 27-29, 42]. They exploit the power of deep generative models that map a latent space to the data distribution. By optimizing for small manipulations in the latent space, these methods ensure for both the minimality and the meaningfulness of changes. The properties of the generative backbone and the way the latent space is structured have a crucial impact on the resulting counterfactual explanations. DIVE [42], for instance, exploits the disentangled latent space of a $\beta$ -TCVAE [10] to generate diverse counterfactual explanations. However, the low visual quality of its counterfactuals hinders interpretability. DIME [28] and DVCE [4] instead build on recently popularized denoising diffusion probabilistic models [12], achieving better-looking results but losing the more curated control on the diversity. These works can explain classifiers working on 'simple' images, e.g., face portraits [34] or + +![](images/3fe7bee604b8e640f8ff7e16ac9095d4843e05e9bd9339eaefa744bb09677681.jpg) +Figure 2. Overview of OCTET. From the query image (top left), a trained encoder $E$ provides initial blob parameters. The blobs are refined in optimization stage ① using Eq. 3 to better encode the query. The refined blob parameters are again modified in optimization stage ② using Eq. 1 to obtain a counterfactual example. The parameters of the encoder $E$ and generator $G$ are frozen during the optimization, as is $M$ the decision model to be explained. Losses are illustrated with dashed arrows whose directions represent the gradient flow. + +featuring a single and centered object [11, 57]. + +Scaling up to classifiers operating over more complex images such as crowded urban scenes remains a major challenge. To date, only STEEX [27] proposes a method able to explain decision models dealing with scenes composed of many objects. It employs a segmentation-to-image generation backbone [68], which helps for generating high-quality complex images but comes with two important limitations. Firstly, it requires annotated semantic segmentation masks of images to work, which is not always available or is costly for the user to produce. Secondly, it imposes that the explanation keeps the same spatial and semantic layout as the query, thereby severely restricting the search space. It is, for instance, not able to provide insights about the importance of the position of objects. By contrast, OCTET does not impose a fixed spatial layout. Indeed, the structural layout of the scene is an intermediate representation of the generator, inferred from images only. Therefore, we do not use additional annotations, and the spatial layout can be optimized by back-propagation. + +# 3. Object-aware Counterfactual Explanations + +We present here OCTET, a method to produce counterfactual examples for compositional scenes, based on a generative architecture. We first formally describe our approach and objective in Sec. 3.1. Then, we instantiate the generative backbone in Sec. 3.2 and the counterfactual objective in Sec. 3.3. Lastly, we explain how we invert the generator to recover the latent code for the query image in Sec. 3.4. The pipeline of OCTET is shown in Fig. 2. + +# 3.1. Goal and Notations + +Given a decision model $M$ , a query image $x^{q}$ , and a target decision $y \neq M(x^{q})$ , a counterfactual explanation is an image $x$ that is close to the query image $x^{q}$ but modified so that the decision $M(x)$ of the model changes to $y$ . Generative counterfactual methods, such as ours, employ a trained generator $G$ [27, 28, 42]. We denote its latent space $\mathcal{Z}$ , with $z^{q} \in \mathcal{Z}$ the latent code that produces the best approximation of $x^{q}$ when passed through $G$ , i.e., so that $G(z^{q}) \approx x^{q}$ . To explain differentiable deep vision models, the definition of a counterfactual explanation stated here-above can be translated into the following optimization problem: + +$$ +\underset {z \in \mathcal {Z}} {\operatorname {a r g m i n}} L _ {\text {d e c i s i o n}} \left(M (G (z)), y\right) + \lambda_ {\text {d i s t}} L _ {\text {d i s t}} \left(z, z ^ {q}\right). \tag {1} +$$ + +The role of the first term $L_{\mathrm{decision}}$ is to push the decision of the model $M$ for the generated image $G(z)$ to the target class $y$ ; The second term $L_{\mathrm{dist}}$ is a distance term in the latent space of $G$ , which ensures that the query $x^q$ and the explanation $G(z)$ remain similar; $\lambda_{\mathrm{dist}}$ is an hyper-parameter that controls the relative weight between the two terms. With optimal $z$ , $G(z)$ then recovers a counterfactual image for the query $x^q$ and the decision model $M$ . Since the choice of $L_{\mathrm{decision}}$ is constrained by the nature of the model to explain, e.g., a classification loss if $M$ is a classifier, we only have to specify $G$ and $L_{\mathrm{dist}}$ to define our counterfactual method. + +# 3.2. Compositional generative backbone + +In this work, we focus on explaining decision models that operate on compositional visual scenes and we look for + +a generative model capable of modeling such scenes. However, most generative models are designed for images containing centered, isolated objects, and they do not learn a latent space that can easily be disentangled into per-object representations. Instead, we propose that our generative backbone should be object-based by design [14, 37]. Such methods are fully differentiable and trained without supervision to generate images in a compositional fashion. + +We choose to build on BlobGAN [14], a compositional generative model that shows good performances and has the advantage of allowing not only edition but also insertion and removal of objects in a differentiable way. In BlobGAN, images are generated in two steps. First, a layout network maps a random Gaussian noise vector to an intermediary representation consisting of an ordered set of $K$ 'blobs'. Each blob $k \in \{1 \dots K\}$ is defined by spatial parameters $\phi_k$ and a style feature vector $\psi_k$ . The spatial parameters $\phi_k$ consist of five values that are its center coordinates, scale, aspect ratio, and rotation angle. In practice, the layout network is a neural network that outputs vectors $\phi_k$ and $\psi_k$ for each blob. Then, in the second stage, the generator network is tasked to transform the blob parameters into an image. A rendering layer draws the blobs on a canvas as ellipses whose shapes and positions are given directly by $\phi_k$ . The output of the rendering layer is a map that associates to each 2D spatial location a feature vector computed from the style vectors of the blobs present at that position. This feature map is transformed by convolutional layers into an image that roughly follows the spatial organization of the canvas. + +The layout network and the generator network are trained end-to-end against a discriminator with standard GAN losses only. In particular, we do not use any annotations for the number, size, class, or location of objects. They are discovered by the generator during training. Note also that scale parameters can take non-positive values. In that case, the object is considered absent and the corresponding blob is not rendered on the canvas in the generator network. + +# 3.3. Setting $L_{\mathrm{dist}}$ for compositional scenes + +Using a compositional generative backbone allows us to design a meaningful, disentangled, per-object distance for images. Indeed, to be interpretable with respect to the query image, a counterfactual explanation should display sparse changes in terms of the number of moved, added, removed, or modified objects. Therefore, we design $L_{\mathrm{dist}}$ such that it is split into per-object distances where each object can be addressed independently. Moreover, we propose that modifications of the spatial position of an object should be orthogonal to those made to its shape or colors. Since in a well-trained BlobGAN each blob is associated with a distinct element of the scene (see Appendix), such a distance can be naturally defined in the space of blob parameters. + +Formally, let us denote $\phi$ and $\psi$ the concatenation across + +all blobs $k \in \{1 \dots K\}$ of spatial parameters $\phi_k$ and style vectors $\psi_k$ , and $z = (\phi, \psi)$ the complete latent representation that the generator network $G$ can decode into an image. Accordingly, the latent code for the query image $x^q$ reads $z^q = (\phi^q, \psi^q)$ . Then, OCTET implements $L_{\mathrm{dist}}$ as follows: + +$$ +L _ {\text {d i s t}} (z, z ^ {q}) = \sum_ {k = 1} ^ {K} \| \phi_ {k} - \phi_ {k} ^ {q} \| _ {1} + \| \psi_ {k} - \psi_ {k} ^ {q} \| _ {1}. \tag {2} +$$ + +We use the $L_{1}$ distance here to promote sparsity in the number of modified blobs and the type of edits applied to each blob. If the user wishes to inspect the role of the style of objects, their spatial positions, or both, with a counterfactual explanation, she can instead optimize Eq. 2 only on the corresponding features of the blobs. + +In order to specify Eq. 2, and thus the final counterfactual objective (Eq. 1), it remains to be solved how to obtain the blob-wise latent codes $\phi_k^q$ and $\psi_k^q$ associated with the query image $x^q$ . We discuss this matter in the next section. + +# 3.4. Inverting the query image + +Obtaining the latent code corresponding to the query image, a process called inversion, is not straightforward when using a GAN-based generator. To produce a counterfactual example, such inversion needs to find latent codes $(\phi^q,\psi^q)$ so that $G(\phi^q,\psi^q)$ recovers $x^{q}$ . For the downstream counterfactual optimization to make sense, the decision of the model for the query image should be preserved for the reconstructed image, i.e., $M(G(\phi^q,\psi^q)) = M(x^q)$ . + +To achieve those objectives, we follow best practices for image inversion [1,2,41,50,67]. We first learn an encoder $E$ to predict blob parameters from input images. Generating an image $G(E(x^q))$ with parameters predicted this way yields a coarse reconstruction of $x^q$ . To improve the inversion, we start from predicted blob parameters $E(x^q)$ and further optimize them with the following objective: + +$$ +\begin{array}{l} \phi^ {q}, \psi^ {q} = \underset {\phi , \psi} {\arg \min } L _ {\mathrm {L P I P S}} (G (\phi , \psi), x ^ {q})) + L _ {2} (G (\phi , \psi), x ^ {q}) \\ + L _ {2} \left(f _ {M} \left(x ^ {q}\right), f _ {M} (G (\phi , \psi))\right) \\ + L _ {2} \left(\left(\phi , \psi\right), E \left(x ^ {q}\right)\right). \tag {3} \\ \end{array} +$$ + +The first two terms ensure that the image $G(\phi, \psi)$ generated from the blob parameters reconstructs the query $x^q$ using a perceptual $L_{\mathrm{LPIPS}}$ loss and the $L_2$ loss. The third term aims to preserve the features that are important to the decision model $M$ by adding a $L_2$ loss on intermediate features $f_M(x)$ of $M$ . Finally, the last term encourages the parameters to stay close to the reasonably good solution given by the encoder. This inversion method concludes the specification of the optimization problem of Eq. 1 in OCTET and of the pipeline depicted in Fig. 2. More details on the inversion and the training of the encoder are in the appendix. + +![](images/f31e535168e4f9d9249280d5abad0182018c73380052c848ce4fd719bd7a5e83.jpg) +Figure 3. Counterfactual explanations on driving scenes. Explanations are generated for binary classifiers predicting the capacity of the ego-car to respectively Move forward, Stop, turn Left, or Right. We indicate at the top left of each image the output decision of the model for that image. For each row, counterfactuals target the opposite decision to that obtained for the query image. OCTET finds interpretable, sparse, and meaningful semantic and spatial modifications to the query image. OCTET style-only only optimizes on the style features, it is analog to STEEX [27] that can only search for changes in object appearance to flip the decision. + +![](images/662c96748894516838b570468917ecaf4e572d0a0221a528bbdad3e1b5ac202e.jpg) + +![](images/513ce641086b18691d286a66acf1d2e978ae0552214c7927bacd25144a220cc5.jpg) + +![](images/868051f78f2c7cfebb90c4a0a9c6249426f0c8d5c0395886a94835af47fdab24.jpg) + +# 4. Experiments + +# 4.1. Experimental protocol and training details + +Data and decision model. Our method is designed to build counterfactual examples to explain decision models trained on compositional images. To evaluate our method, we follow previous work [27] and use the BDD100k dataset [60], which contains 100k images of diverse driving scenes that we resize to a resolution of $512 \times 256$ . In addition, we also use BDD-OIA [59], a 20K scene extension of BDD100k that is annotated with 4 binary attributes labeled Move Forward, Stop, Turn Left, and Turn Right. Each label is defined as the capacity in the given situation of the ego-vehicle to perform the corresponding action. Note that by this definition, being allowed to Stop and to Move Forward are not mutually exclusive. The decision models being explained in the experiments are multi-label binary DenseNet121 [26] classifiers trained on BDD-OIA with the four aforementioned labels. + +Baseline. We compare our method against STEEX [27], the only counterfactual explanation method that performs well on compositional images like driving scenes. Note that in STEEX, the conservation of the semantic layout is hardcoded in the architecture: it is less flexible as an explanation method than ours. Moreover, the authors use semantic seg + +mentation annotations of BDD100k to train their generative backbone, while we do not need any additional annotation. We use the pre-trained model from the official release. + +Technical setup. For our method, we train the generative backbone (BlobGAN) on the training set of BDD100k, with a number $K = 40$ of blobs. We adapt the architecture of BlobGAN to generate rectangular $512 \times 256$ images, and we decrease the size of the feature vectors of the BlobGAN to compensate for the increased number of blobs. Throughout our experiments, we evaluate two main variants of our method: OCTET where both spatial and style features are optimized altogether and OCTET style-only where only the style vector $\psi$ is optimized while the spatial vector $\phi$ is left equal to $\phi^q$ . This setting only assesses the contribution of style features, and is most similar to the STEEX baseline [27] that finds explanations while conserving the semantic layout of the query. More details in the appendix. + +# 4.2. Quality of the explanations + +Counterfactual explanations are defined as images that are similar to the query image but effectively change the decision of the target model. To verify that our explanations fit the criteria, we evaluate them with the following metrics: 1) the FID [24] which measures if they are realistic images, and 2) a perceptual distance LPIPS [65] that mea + +![](images/06746fddb885cfc21d5ade7ca54e393a7c56d79008c40646fcfdcb216c87ec72.jpg) +Figure 4. FID vs. LPIPS, labeled with Success rate for OCTET and STEEX [27] explaining a decision model trained on the 4 labels of BDD-OIA. Data points are obtained with different values of $\lambda_{\mathrm{dist}}$ yielding different trade-offs. STEEX is directly comparable to OCTET style-only as they operate in similar conditions. See Sec. 4.2. + +sures the change magnitude between them and the query image. We also compute the success rate corresponding to the percentage of counterfactual examples that are indeed classified in the target class by the decision model. We compute these metrics for different values of $\lambda_{\mathrm{dist}}$ that lead to different trade-offs and plot the scores in Fig. 4. As STEEX explanations are constrained to style features, we compare it to our OCTET style-only setting. Firstly, in terms of FID, we can observe that both versions of OCTET obtain lower FID than STEEX: OCTET produces explanations of better visual quality, even though it does not use any annotations as opposed to STEEX, which needs semantic maps, and handles spatial explanations that are more challenging to generate. Secondly, in terms of LPIPS, for style-based explanations OCTET style-only stays closer to the query than STEEX. This can be explained as STEEX latent representation is semantic-based, while ours is instance-based. Our method is therefore able to seek finer-grained modifications, while STEEX tends to change the style for an entire semantic class even when only one instance is important to the decision model. As for OCTET, because LPIPS disproportionately weights structural modifications, the wide range of attainable values indicates that many degrees of changes can be obtained depending on the choice of $\lambda_{\mathrm{dist}}$ . Some settings tend to move objects a lot, while other settings attain values similar to 'style-only' edits. Finally, the very high achievable Success Rates displayed by all three models confirm that they will propose explanations in almost any situation. + +We display examples of OCTET's counterfactual explanations in Fig. 3. OCTET finds sparse spatial and/or semantic modifications to objects that strongly influence the output decision. For example, we can see with OCTET that the distance with the car in front plays a role when deciding for 'Forward' (1st row) or 'Stop' (2nd row). On the other hand, OCTET style-only shows that for those same decisions, the brake lights also are taken into account. For going 'Left' class (3rd row), the absence of the continuous line seems + +![](images/02ecf044eeeb09d18676375f0f91775f46ec942e42550eda0fa0ac116ee09fc0.jpg) + +![](images/cf2a1f9c734e9004d3040b14da5d2538c94a2698debb9edb5b6896c23da7e16d.jpg) + +![](images/0ba562000555d27052b79a88c49fcfa6b01570a01e619c233aa7dac408cc5c17.jpg) + +![](images/4cc55b71d51124381d24eff6d9917f9818338f0f7b6c629b64984ab2af6408de.jpg) + +![](images/e0695404d0381ce301bcafbc942476d2f8f1bc6bb2920e79b4047bef45f7313c.jpg) +Figure 5. Targeted counterfactual explanations. For each query image (top row), we target different blobs (2nd and 3rd row) to inspect the role of corresponding objects. The output of the decision model is reported on the top right corner for each image. + +![](images/382776078b54f5ced820578e6c6683388632bf6d795c6dd4e5da08983f492584.jpg) + +crucial. On the 3rd and 4th row, we can see that OCTET mixes both style (brake lights) and spatial (proximity) explanations. On the same tasks, STEEX changes the style of the whole class, and not just the necessary instance. + +# 4.3. Targeted counterfactuals + +Models that operate on compositional scenes can have their decisions influenced by many different factors, and we have shown that OCTET can expose those factors in its counterfactual examples. However, the user may want to test some particular hypothesis dealing with specific objects or areas of the scene. Accordingly, the user can decide to optimize Eq. 2 only on selected blobs parameters. Finding the blob corresponding to a specific object in the image is discussed in Appendix. For instance, in the situation shown at the top left of Fig. 5, the car at the front seems to have + +![](images/26199a254b14a14d894dd2752db6136ce1d86e516264b0c238cdd47b2550a2b5.jpg) +Figure 6. Counterfactual explanations for a segmentation model. The segmentation model predicts the driveable area with primary and secondary lanes shown in red and blue respectively. The first row shows query images and segmentations by the model. In the second row, we add or remove parts of the original masks to build target masks (changes are shown with white dotted lines). The third row displays counterfactual explanations for the target masks and predictions on the new images. Green ellipses are manually added to highlight details. + +its brake lights on, but the decision is still not to 'Stop'. Targeting the front car (2nd row left image) reveals that the decision would have been 'Stop' if the car was closer, and the brake light more clearly contrasted. OCTET can also target blobs that are currently inactive, corresponding to no visible object in the image. In this example, doing so for car blobs shows that if another vehicle also had brake lights on, the model would also decide to 'Stop' (3rd row left image). This finding is unexpected, and could warrant further investigation of the decision model. We show in Fig. 5 and Fig. 1 some other results we obtain by targeting specific blobs. + +# 4.4. Explaining a semantic segmentation model + +For complex tasks, like autonomous driving, decision models usually rely on multiple neural networks that address different subtasks (e.g., traffic sign recognition, depth estimation, motion forecasting). In this context, it is important to explain models beyond classifiers [25,54], a question not addressed in previous counterfactual methods. Here, we show the versatility of our method as we extend it to explain a semantic segmentation network. To do so, we use the same optimization problem described in Eq. 1 and Sec. 3 but we change $L_{\text{decision}}$ to an image segmentation loss. + +In Fig. 6, we show counterfactual explanations for a driveable area segmentation model. More precisely, OCTET explains outputs of a DeepLabv3 [9] trained to segment images from BDD100k into classes of driveable area: the primary lane, the secondary side lanes, and the background. We define the target masks by manually modifying predictions made on the query images as illustrated in the second row. We see in the obtained counterfactuals, in the + +third row, changes in road markings and vehicles' positions. They confirm that the segmentation model will consider as driveable a lane behind broken white lines (col. 3) but not behind a double yellow line (col. 2), especially if, in addition, the lane is hatched (col. 3). If there is a car on a drivable lane, the model can detect that the occupied portion is non-driveable while considering that the rest of the lane remains accessible. With these examples, OCTET provides explanations for both large semantic regions, e.g., when we add or remove an entire region in the target mask, or small subparts, e.g., when we slightly crop or extend a region. + +# 4.5. User-study to assess explanations + +Metrics presented in Sec. 4.2 measure how well the produced examples match the definition of counterfactual explanations. We want here to assess how useful the explanations are for a user to better understand a decision model. However, designing such an evaluation is challenging. Indeed, it has to avoid the many biases that appear [13, 36]. Moreover, we usually do not have access to the inner workings of a model, and therefore we cannot compare users' insights on its behavior against a ground truth. To overcome these challenges, we take inspiration from recent work on the usefulness of saliency maps as explanations [15] and we adapt its evaluation protocol to counterfactual explanations. + +The goal is to measure if accessing counterfactual explanations on some examples helps users to better understand the model. Grounded within previous literature [16,22,30], we instantiate understand as 'being able to predict the model's output for new instances', a concept known as simulatability. In particular, our user-study has two phases: + +
Cohort sizeReplicationBias Detection
Group w/ OCTET2070%65%
Group control2052.5%0%
+ +Table 1. Results of the user-study. The performance of users who had access to explanations ('w/ OCTET') in the observation phase is compared to that of users who didn't ('control'). 'Replication' measures the accuracy of the users in predicting the model's output on new samples and 'bias detection' is the proportion of users who found the spurious correlation in the model. + +1. Observation phase: participants are shown a series of triplets, each being composed of (1) a query image, (2) the model's decision on that image, and (3), a counterfactual explanation for the decision. +2. Questionnaire phase: participants are presented with new images and are asked to guess the output of the model on these images. We stress that during the questionnaire, no counterfactual images are shown. + +To measure the impact that counterfactual explanations of the first phase have on the users' ability to predict the model output on new samples (second phase), we use a control group. This control group, composed of different participants, undertakes the exact same tasks except that counterfactual explanations are not shown during the first phase. The decision model used in this study is a classifier trained on BDD-OIA [59] to predict the binary 'Turn Right' label. Besides, the decision model is trained, by design, with a flaw: the presence of a car on any side of the road can impact the decision, not only obstacles on the right side. As discussed further below, this bias is introduced to study if users are able to identify model defects thanks to counterfactuals. More details are provided in the appendix. + +Replication score. We first measure the 'replication score': the proportion of correct match between the user's answer and the prediction of the decision model, averaged over participants of the group. We report these results in Tab. 1 and observe that participants in the group accessing counterfactuals during the observation phase can successfully guess the decision model's prediction $70\%$ of the time while the control group's performance is $52.5\%$ , barely above random. A statistical unpaired t-test, with as null hypothesis that the counterfactual explanations have no effect, confirms the significance of our result with a p-value of 0.0028. This demonstrates the usefulness of OCTET in understanding the decision model as participants with explanations could better anticipate the model's behavior on new instances. + +Bias detection. Another promise of counterfactual explanations is to enable bias detection. We then want to assess whether, by looking at explanations, users can tell if and how a decision model is biased. In that direction, previous works dealing with human face datasets train biased models + +by artificially confounding attributes, e.g., smile and age, in the training set. Then, they develop bias detection benchmarks that measure, by means of an oracle neural network, how well attributes remain confounded in the counterfactual explanations [28, 42, 46]. However, a serious limitation of that benchmark is that it keeps end-users out of the evaluation loop and it cannot tell whether or not a user will be able to uncover the spurious correlations by looking at the counterfactual explanation. Moreover, detecting the issue with the model can be especially difficult when the bias is not obvious or its presence not suspected in the first place. Instead of an automated evaluation, at the end of the questionnaire, we asked the participants to describe in free text what they had understood from the model. The goal is to prompt them to mention the issue we embedded in the decision model, i.e., that the presence of a car on the left-hand side of the image factors in the 'Turn Right' decision. The 'bias detection' score simply counts, for each group, the proportion of participants that used the word 'left' in their answers. In Tab. 1, we see that $65\%$ of the participants that had access to the explanations from OCTET identify the spurious correlation learned by the model, while none of the users that did not have the explanations (control group) mentions it. This clearly shows that our method is useful to pinpoint issues with the decision model. Also, the fact that not every participant detected it confirms our suspicions against automated bias detection benchmarks: the presence of the information in the counterfactual examples is not sufficient to ensure that the issue will be found in practice. + +# 5. Conclusion + +In this work, we presented OCTET, a generative method to produce counterfactual explanations for deep vision models working on complex scenes. Thanks to an object-centric representation of the scene, OCTET is able to process complex compositional scenes and to assess the impact of both the spatial positions and the style features of each object. Our tool allows in-depth exploration of the different explaining factors by letting the user adjust on the fly the weights of the different search directions independently. We evaluated our method thoroughly on real driving scenes from BDD100k, with an improved evaluation benchmark compared to previous works, including a human-centered study that confirmed the usefulness of the approach. + +# Acknowledgements + +This work was supported in part by the ANR grants VISA DEEP (ANR-20-CHIA-0022) and MultiTrans (ANR-21-CE23-0032). Authors would like to thank the voluntary participants of the user-study, as well as Remi Cadene, Julien Colin, Thomas Fel, and Guillaume Jeanneret for helpful discussions. + +# References + +[1] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan: How to embed images into the stylegan latent space? In ICCV, 2019. 4 +[2] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan++: How to edit the embedded images? In CVPR, 2020. 4 +[3] Julius Adebayo, Justin Gilmer, Michael Muelly, Ian J. Goodfellow, Moritz Hardt, and Been Kim. Sanity checks for saliency maps. In NeurIPS, 2018. 2 +[4] Maximilian Augustin, Valentyn Boreiko, Francesco Croce, and Matthias Hein. Diffusion visual counterfactual explanations. In NeurIPS, 2022. 1, 2 +[5] Mariusz Bojarski, Anna Choromanska, Krzysztof Choromanski, Bernhard Firner, Larry J Ackel, Urs Muller, Phil Yeres, and Karol Zieba. Visualbackprop: Efficient visualization of cnns for autonomous driving. In ICRA, 2018. 2 +[6] Kieran Browne and Ben Swift. Semantics and explanation: why counterfactual explanations produce adversarial examples in deep neural networks. CoRR, abs/2012.10076, 2020. 2 +[7] Chun-Hao Chang, Elliot Creager, Anna Goldenberg, and David Duvenaud. Explaining image classifiers by counterfactual generation. In ICLR, 2019. 1 +[8] Chaofan Chen, Oscar Li, Daniel Tao, Alina Barnett, Cynthia Rudin, and Jonathan Su. This looks like that: Deep learning for interpretable image recognition. In NeurIPS, 2019. 2 +[9] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. CoRR, abs/1706.05587, 2017. 7 +[10] Ricky T. Q. Chen, Xuechen Li, Roger Grosse, and David Duvenaud. Isolating sources of disentanglement in variational autoencoders. In NeurIPS, 2018. 2 +[11] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 3 +[12] Prafulla Dhariwal and Alexander Quinn Nichol. Diffusion models beat gans on image synthesis. In NeurIPS, 2021. 2 +[13] Finale Doshi-Velez and Been Kim. Towards a rigorous science of interpretable machine learning. arXiv preprint arXiv:1702.08608, 2017. 7 +[14] Dave Epstein, Taesung Park, Richard Zhang, Eli Shechtman, and Alexei A. Efros. Blobgan: Spatially disentangled scene representations. In ECCV, 2022. 1, 4 +[15] Thomas Fel, Julien Colin, Rémi Cadène, and Thomas Serre. What I cannot predict, I do not understand: A human-centered evaluation framework for explainability methods. In NeurIPS, 2022. 2, 7 +[16] Ruth C. Fong and Andrea Vedaldi. Interpretable explanations of black boxes by meaningful perturbation. In ICCV, 2017. 2, 7 +[17] Timo Freiesleben. Counterfactual explanations & adversarial examples - common grounds, essential differences, and potential transfers. CoRR, abs/2009.05487, 2020. 2 +[18] Nicholas Frosst and Geoffrey E. Hinton. Distilling a neural network into a soft decision tree. In Workshop on Comprehensibility and Explanation in AI and ML @AI*IA, 2017. 2 + +[19] Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. In ICLR, 2015. 2 +[20] Yash Goyal, Ziyan Wu, Jan Ernst, Dhruv Batra, Devi Parikh, and Stefan Lee. Counterfactual visual explanations. In ICML, 2019. 2 +[21] Michael Harradon, Jeff Druce, and Brian E. Ruttenberg. Causal learning and explanation of deep neural networks via autoencoded activations. CoRR, 2018. 2 +[22] Peter Hase and Mohit Bansal. Evaluating explainable AI: which algorithmic explanations help users predict model behavior? In ACL, 2020. 7 +[23] Lisa Anne Hendricks, Ronghang Hu, Trevor Darrell, and Zeynep Akata. Grounding visual explanations. In ECCV, 2018. 2 +[24] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. 5 +[25] Lukas Hoyer, Maurizio Munoz, Prateek Katiyar, Anna Khoreva, and Volker Fischer. Grid saliency for context explanations of semantic segmentation. In NeurIPS, 2019. 7 +[26] Gao Huang, Zhuang Liu, Laurens van der Maaten, and Kilian Q. Weinberger. Densely connected convolutional networks. In CVPR, 2017. 5 +[27] Paul Jacob, Éloi Zablocki, Hedi Ben-Younes, Mickaël Chen, Patrick Pérez, and Matthieu Cord. STEEX: steering counterfactual explanations with semantics. In ECCV, 2022. 2, 3, 5, 6 +[28] Guillaume Jeanneret, Loic Simon, and Frédéric Jurie. Diffusion models for counterfactual explanations. CoRR, abs/2203.15636, 2022. 1, 2, 3, 8 +[29] Saeed Khorram and Fuxin Li. Cycle-consistent counterfactuals by latent transformations. In CVPR, 2022. 1, 2 +[30] Been Kim, Oluwasanmi Koyejo, and Rajiv Khanna. Examples are not enough, learn to criticize! criticism for interpretability. In NeurIPS, 2016. 7 +[31] Jinkyu Kim and John F. Canny. Interpretable learning for self-driving cars by visualizing causal attention. In ICCV, 2017. 2 +[32] Oran Lang, Yossi Gandelsman, Michal Yarom, Yoav Wald, Gal Elidan, Avinatan Hassidim, William T. Freeman, Phillip Isola, Amir Globerson, Michal Irani, and Inbar Mosseri. Explaining in style: Training a GAN to explain a classifier in stylespace. In ICCV, 2021. 2 +[33] Zhiheng Li and Chenliang Xu. Discover the Unknown Biased Attribute of an Image Classifier. In ICCV, 2021. 2 +[34] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In ICCV, 2015. 2 +[35] Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, and Pascal Frossard. Deepfool: A simple and accurate method to fool deep neural networks. In CVPR, 2016. 2 +[36] Giang Nguyen, Daeyoung Kim, and Anh Nguyen. The effectiveness of feature attribution methods and its correlation with automatic evaluation scores. In NeurIPS, 2021. 7 + +[37] Michael Niemeyer and Andreas Geiger. GIRAFFE: representing scenes as compositional generative neural feature fields. In CVPR, 2021. 4 +[38] Martin Pawelczyk, Shalmali Joshi, Chirag Agarwal, Sohini Upadhyay, and Himabindu Lakkaraju. On the connections between counterfactual explanations and adversarial examples. CoRR, abs/2106.09992, 2021. 2 +[39] Sylvestre-Alvise Rebuffi, Ruth Fong, Xu Ji, and Andrea Vedaldi. There and back again: Revisiting backpropagation saliency methods. In CVPR, 2020. 2 +[40] Marco Túlio Ribeiro, Sameer Singh, and Carlos Guestrin. "why should I trust you?": Explaining the predictions of any classifier. In SIGKDD, 2016. 2 +[41] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: A stylegan encoder for image-to-image translation. In CVPR, 2021. 4 +[42] Pau Rodríguez, Massimo Caccia, Alexandre Lacoste, Lee Zamparo, Issam H. Laradji, Laurent Charlin, and David Vázquez. Beyond trivial counterfactual explanations with diverse valuable explanations. In ICCV, 2021. 1, 2, 3, 8 +[43] Axel Sauer, Nikolay Savinov, and Andreas Geiger. Conditional affordance learning for driving in urban environments. In CoRL, 2018. 2 +[44] Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In ICCV, 2017. 2 +[45] Yuan Shen, Shanduojiao Jiang, Yanlin Chen, Eileen Yang, Xilun Jin, Yuliang Fan, and Katie Driggs Campbell. To explain or not to explain: A study on the necessity of explanations for autonomous vehicles. CoRR, 2020. 2 +[46] Sumedha Singla, Brian Pollack, Junxiang Chen, and Kayhan Batmanghelich. Explanation by progressive exaggeration. In ICLR, 2020. 1, 2, 8 +[47] Mukund Sundararajan, Ankur Taly, and Qiqi Yan. Axiomatic attribution for deep networks. In ICML, 2017. 2 +[48] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian J. Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In *ICLR*, 2014. 2 +[49] Erico Tjoa and Cuntai Guan. A survey on explainable artificial intelligence (XAI): towards medical XAI. CoRR, abs/1907.07374, 2019. 2 +[50] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for stylegan image manipulation. ACM Transactions on Graphics (TOG), 2021. 4 +[51] Simon Vandenhende, Dhruv Kumar Mahajan, Filip Radenovic, and Deepti Ghadiyaram. Making heads or tails: Towards semantically consistent visual counterfactuals. In ECCV, 2022. 2 +[52] Sahil Verma, John P. Dickerson, and Keegan Hines. Counterfactual explanations for machine learning: A review. CoRR, abs/2010.10596, 2020. 1 + +[53] Tom Vermeire, Dieter Brughmans, Sofie Goethals, Raphael Mazzine Barbossa de Oliveira, and David Martens. Explainable image classification with evidence counterfactual. Pattern Anal. Appl., 2022. 2 +[54] Kira Vinogradova, Alexandr Dibrov, and Gene Myers. Towards interpretable semantic segmentation via gradient-weighted class activation mapping (student abstract). In AAAI, 2020. 7 +[55] Sandra Wachter, Brent Mittelstadt, and Chris Russell. Counterfactual explanations without opening the black box: Automated decisions and the gdpr. Harvard Journal of Law & Technology, 2017. 1, 2 +[56] Jörg Wagner, Jan Mathias Kohler, Tobias Gindele, Leon Hetzel, Jakob Thaddäus Wiedemer, and Sven Behnke. Interpretable and fine-grained visual explanations for convolutional neural networks. In CVPR, 2019. 2 +[57] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 3 +[58] Pei Wang and Nuno Vasconcelos. SCOUT: self-aware discriminant counterfactual explanations. In CVPR, 2020. 2 +[59] Yiran Xu, Xiaoyin Yang, Lihang Gong, Hsuan-Chu Lin, TzYing Wu, Yunsheng Li, and Nuno Vasconcelos. Explainable object-induced action decision for autonomous vehicles. In CVPR, 2020. 2, 5, 8 +[60] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darryll. BDD100K: A diverse driving dataset for heterogeneous multitask learning. In CVPR, 2020. 2, 5 +[61] Éloi Zablocki, Hédi Ben-Younes, Patrick Pérez, and Matthieu Cord. Explainability of deep vision-based autonomous driving systems: Review and challenges. Int. J. Comput. Vis., 2022. 2 +[62] Matthew D. Zeiler and Rob Fergus. Visualizing and understanding convolutional networks. In ECCV, 2014. 2 +[63] Quanshi Zhang, Ying Nian Wu, and Song-Chun Zhu. Interpretable convolutional neural networks. In CVPR, 2018. 2 +[64] Qiaoning Zhang, X. Jessie Yang, and Lionel Peter Robert. Expectations and trust in automated vehicles. In CHI, 2020. 2 +[65] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 5 +[66] Bolei Zhou, Aditya Khosla, Ågata Lapedriza, Aude Oliva, and Antonio Torralba. Object detectors emerge in deep scene cnns. In ICLR, 2015. 2 +[67] Jiapeng Zhu, Yujun Shen, Deli Zhao, and Bolei Zhou. Indomain gan inversion for real image editing. In ECCV, 2020. 4 +[68] Peihao Zhu, Rameen Abdal, Yipeng Qin, and Peter Wonka. SEAN: image synthesis with semantic region-adaptive normalization. In CVPR, 2020. 3 \ No newline at end of file diff --git a/2023/OCTET_ Object-Aware Counterfactual Explanations/images.zip b/2023/OCTET_ Object-Aware Counterfactual Explanations/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d21dac7d1e7e5ae6e8c9508c20fb7ab223621132 --- /dev/null +++ b/2023/OCTET_ Object-Aware Counterfactual Explanations/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3e305be2555ad35c4d9385f7eaac7c23a97c8e69e0d32a4a16ea8aa260b59279 +size 577151 diff --git a/2023/OCTET_ Object-Aware Counterfactual Explanations/layout.json b/2023/OCTET_ Object-Aware Counterfactual Explanations/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..27926a1fe2d20733e72056b2ac94de2d56ed14ac --- /dev/null +++ b/2023/OCTET_ Object-Aware Counterfactual Explanations/layout.json @@ -0,0 +1,8219 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 134, + 103, + 460, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 103, + 460, + 121 + ], + "spans": [ + { + "bbox": [ + 134, + 103, + 460, + 121 + ], + "type": "text", + "content": "OCTET: Object-aware Counterfactual Explanations" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 142, + 550, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 142, + 550, + 174 + ], + "spans": [ + { + "bbox": [ + 47, + 142, + 550, + 174 + ], + "type": "text", + "content": "Mehdi Zemni1, Mickaël Chen1, Éloi Zablocki1, Hédi Ben-Younes1, Patrick Pérez1, Matthieu Cord1,2 \n1 Valeo.ai, Paris, France \n2 Sorbonne Université, Paris, France" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 201, + 192, + 215 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 201, + 192, + 215 + ], + "spans": [ + { + "bbox": [ + 143, + 201, + 192, + 215 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 45, + 227, + 290, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 227, + 290, + 563 + ], + "spans": [ + { + "bbox": [ + 45, + 227, + 290, + 563 + ], + "type": "text", + "content": "Nowadays, deep vision models are being widely deployed in safety-critical applications, e.g., autonomous driving, and explainability of such models is becoming a pressing concern. Among explanation methods, counterfactual explanations aim to find minimal and interpretable changes to the input image that would also change the output of the model to be explained. Such explanations point end-users at the main factors that impact the decision of the model. However, previous methods struggle to explain decision models trained on images with many objects, e.g., urban scenes, which are more difficult to work with but also arguably more critical to explain. In this work, we propose to tackle this issue with an object-centric framework for counterfactual explanation generation. Our method, inspired by recent generative modeling works, encodes the query image into a latent space that is structured in a way to ease object-level manipulations. Doing so, it provides the end-user with control over which search directions (e.g., spatial displacement of objects, style modification, etc.) are to be explored during the counterfactual generation. We conduct a set of experiments on counterfactual explanation benchmarks for driving scenes, and we show that our method can be adapted beyond classification, e.g., to explain semantic segmentation models. To complete our analysis, we design and run a user study that measures the usefulness of counterfactual explanations in understanding a decision model. Code is available at https://github.com/valeoaai/OCTET." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 585, + 128, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 585, + 128, + 597 + ], + "spans": [ + { + "bbox": [ + 47, + 585, + 128, + 597 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "Deep learning models are now being widely deployed, notably in safety-critical applications such as autonomous driving. In such contexts, their black-box nature is a major concern, and explainability methods have been developed to improve their trustworthiness. Among them, counterfactual explanations have recently emerged to provide insights into a model's decision [7, 52, 55]. Given a decision model and an input query, a counterfactual explanation is a data point that differs minimally but meaningfully from the query in" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 308, + 201, + 425, + 271 + ], + "blocks": [ + { + "bbox": [ + 308, + 201, + 425, + 271 + ], + "lines": [ + { + "bbox": [ + 308, + 201, + 425, + 271 + ], + "spans": [ + { + "bbox": [ + 308, + 201, + 425, + 271 + ], + "type": "image", + "image_path": "d33776a7a391bf2261a43e245a6f91dce3948c346fb8e8cc627ebdcddf3c8507.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 426, + 202, + 543, + 271 + ], + "blocks": [ + { + "bbox": [ + 426, + 202, + 543, + 271 + ], + "lines": [ + { + "bbox": [ + 426, + 202, + 543, + 271 + ], + "spans": [ + { + "bbox": [ + 426, + 202, + 543, + 271 + ], + "type": "image", + "image_path": "e47858e4c9cc444c0f953f84b0949b80c1da045a5b34ded8728f5484e20ac5d0.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 308, + 272, + 425, + 334 + ], + "blocks": [ + { + "bbox": [ + 308, + 272, + 425, + 334 + ], + "lines": [ + { + "bbox": [ + 308, + 272, + 425, + 334 + ], + "spans": [ + { + "bbox": [ + 308, + 272, + 425, + 334 + ], + "type": "image", + "image_path": "d73cc6e01394921ba2748a6a2fcc1da5b8ef572cb6da23bf7142c222f05c3999.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 343, + 547, + 432 + ], + "lines": [ + { + "bbox": [ + 304, + 343, + 547, + 432 + ], + "spans": [ + { + "bbox": [ + 304, + 343, + 547, + 432 + ], + "type": "text", + "content": "Figure 1. Counterfactual explanations generated by OCTET. Given a classifier that predicts whether or not it is possible to go left, and a query image (top left), OCTET produces a counterfactual explanation where the most influential features that led to the decision are changed (top right). On the bottom row, we show that OCTET can also operate under different settings that result in different focused explanations. We report the prediction made by the decision model at the top left of each image." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 427, + 272, + 543, + 334 + ], + "blocks": [ + { + "bbox": [ + 427, + 272, + 543, + 334 + ], + "lines": [ + { + "bbox": [ + 427, + 272, + 543, + 334 + ], + "spans": [ + { + "bbox": [ + 427, + 272, + 543, + 334 + ], + "type": "image", + "image_path": "9b87e53b767850201cdde52254fad8d6e04c50f1d066d5192fd8d202da78a202.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 459, + 547, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 459, + 547, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 459, + 547, + 639 + ], + "type": "text", + "content": "a way that changes the output decision of the model. By looking at the differences between the query and the explanation, a user is able to infer — by contrast — which elements were essential for the model to come to its decision. However, most counterfactual methods have only shown results for explaining classifiers trained on single-object images such as face portraits [4, 28, 29, 42, 46]. Aside from the technical difficulties of scaling up the resolution of the images, explaining decision models trained on scenes composed of many objects also present the challenge that those decisions are often multi-factorial. In autonomous driving, for example, most decisions have to take into account the position of all other road users, as well as the layout of the road and its markings, the traffic light and signs, the overall visibility, and many other factors." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "text", + "content": "In this paper, we present a new framework, dubbed OCTET for Object-aware Counterfactual ExplanaTions, to generate counterfactual examples for autonomous driving. We leverage recent advances in unsupervised compositional generative modeling [14] to provide a flexible explanation method. Exploiting such a model as our backbone, we can" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 495, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 495, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 495, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "15062" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "type": "text", + "content": "assess the contribution of each object of the scene independently and look for explanations in relation to their positions, their styles, or combinations of both." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 108, + 286, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 108, + 286, + 262 + ], + "spans": [ + { + "bbox": [ + 46, + 108, + 286, + 262 + ], + "type": "text", + "content": "To validate our claims, extensive experiments are conducted for self-driving action decision models trained with the BDD100k dataset [60] and its BDD-OIA extension [59]. Fig. 1 shows counterfactual explanations found by OCTET: given a query image and a visual decision system trained to assess if the ego vehicle is allowed to go left or not, OCTET proposes a counterfactual example with cars parked on the left side that are moved closer. When inspecting specific items (bottom row of the figure), OCTET finds that moving the yellow car to the left, or adding a double line marking on the road, are ways to change the model's decision. These explanations highlight that the absence of such elements on the left side of the road heavily influenced the model." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 264, + 228, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 264, + 228, + 275 + ], + "spans": [ + { + "bbox": [ + 47, + 264, + 228, + 275 + ], + "type": "text", + "content": "To sum up, our contributions are as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 276, + 286, + 502 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 53, + 276, + 286, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 276, + 286, + 323 + ], + "spans": [ + { + "bbox": [ + 53, + 276, + 286, + 323 + ], + "type": "text", + "content": "- We tackle the problem of building counterfactual explanations for visual decision models operating on complex compositional scenes. We specifically target autonomous driving visual scenarios." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 324, + 286, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 324, + 286, + 358 + ], + "spans": [ + { + "bbox": [ + 53, + 324, + 286, + 358 + ], + "type": "text", + "content": "- Our method is also a tool to investigate the role of specific objects in a model's decision, empowering the user with control over which type of explanation to look for." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 360, + 286, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 360, + 286, + 430 + ], + "spans": [ + { + "bbox": [ + 53, + 360, + 286, + 430 + ], + "type": "text", + "content": "- We thoroughly evaluate the realism of our counterfactual images, the minimality and meaningfulness of changes, and compare against previous reference strategies. Beyond explaining classifiers, we also demonstrate the versatility of our method by addressing explanations for a segmentation network." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 431, + 286, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 431, + 286, + 502 + ], + "spans": [ + { + "bbox": [ + 53, + 431, + 286, + 502 + ], + "type": "text", + "content": "- Finally, we conduct a user-centered study to assess the usefulness of our explanations in a practical case. As standard evaluation benchmarks for counterfactual explanations are lacking a concrete way to measure the interpretability of the explanations, our user-centered study is a key element to validate the presented pipeline." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 514, + 222, + 527 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 514, + 222, + 527 + ], + "spans": [ + { + "bbox": [ + 47, + 514, + 222, + 527 + ], + "type": "text", + "content": "2. Background and Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 287, + 713 + ], + "type": "text", + "content": "Local explanations. The overwhelming majority of deep learning based models are designed without explainability in mind, with only a few notable exceptions [8, 63]. This fact has prompted an interest in post-hoc explanations of already trained models that can be useful to analyze corner cases, understand failures, and find biases [15, 40]. In safety-critical applications, e.g., autonomous driving or medical imaging, explanations are especially needed for liability purposes and to foster end-user trust [45, 49, 61, 64]. Post-hoc explanations are global if they provide a holistic view of the main decision factors driving the model [18, 21, 32, 33], or they are local if they target the understanding of the model behavior on a specific input [40, 44]. Historically, the vast majority of local explanation methods are attribution-based: they generate saliency maps high-" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 545, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 251 + ], + "type": "text", + "content": "lighting pixels or regions influencing the most the model's decision [5, 16, 39, 44, 47, 53, 56, 62, 66]. In the case of urban scenes, a saliency method would for instance point at the traffic light, or the presence of a pedestrian to explain why a driving model stops [31, 43]. However, saliency explanations may be misleading as they can be independent of the model at hand and merely act as edge detectors [3]. Besides, saliency methods are by nature restricted to show where the important regions are and they cannot indicate what in these regions was deemed decisive. For example, beyond highlighting things, it would be useful to know if the color of the light (red or green?) or the orientation of the pedestrian's body (towards or away from the car?) are taken into account in the model's decision. Counterfactual explanations are a step in that direction." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 257, + 545, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 257, + 545, + 412 + ], + "spans": [ + { + "bbox": [ + 304, + 257, + 545, + 412 + ], + "type": "text", + "content": "Counterfactual explanations. For a given decision model and query input, a counterfactual explanation is defined as a minimally but meaningfully modified version of the query that changes the decision of the model [55]. The notions of 'minimal' and 'meaningful' are crucial here. Adversarial attacks [19, 35, 48] for instance are minimal but not meaningful perturbations: they are imperceptible by design and thus not informative for the end-user [6, 17, 38]. Conversely, if not minimal, the explanation might exhibit unnecessary changes that would hinder its interpretability. While originally developed for models operating on tabular data [55], counterfactuals have been recently considered to explain deep computer vision models [20, 27, 42, 46]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 413, + 545, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 413, + 545, + 509 + ], + "spans": [ + { + "bbox": [ + 304, + 413, + 545, + 509 + ], + "type": "text", + "content": "Seminal work [23] explored the use of natural language to produce descriptions of elements that, in case of presence, change the image classifications. Later, retrieval-based methods aimed to explain decisions by mining the dataset for counterfactual explanations [20, 23, 51, 58]. While obtained explanations can be meaningful, the difference to the query can only be as minimal as the dataset granularity allows, limiting overall interpretability." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": "More recently, generator-based counterfactual explanations have been proposed [4, 27-29, 42]. They exploit the power of deep generative models that map a latent space to the data distribution. By optimizing for small manipulations in the latent space, these methods ensure for both the minimality and the meaningfulness of changes. The properties of the generative backbone and the way the latent space is structured have a crucial impact on the resulting counterfactual explanations. DIVE [42], for instance, exploits the disentangled latent space of a " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": "-TCVAE [10] to generate diverse counterfactual explanations. However, the low visual quality of its counterfactuals hinders interpretability. DIME [28] and DVCE [4] instead build on recently popularized denoising diffusion probabilistic models [12], achieving better-looking results but losing the more curated control on the diversity. These works can explain classifiers working on 'simple' images, e.g., face portraits [34] or" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15063" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 70, + 545, + 270 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 545, + 270 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 545, + 270 + ], + "type": "image", + "image_path": "3fe7bee604b8e640f8ff7e16ac9095d4843e05e9bd9339eaefa744bb09677681.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 281, + 547, + 327 + ], + "lines": [ + { + "bbox": [ + 46, + 281, + 547, + 327 + ], + "spans": [ + { + "bbox": [ + 46, + 281, + 547, + 327 + ], + "type": "text", + "content": "Figure 2. Overview of OCTET. From the query image (top left), a trained encoder " + }, + { + "bbox": [ + 46, + 281, + 547, + 327 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 46, + 281, + 547, + 327 + ], + "type": "text", + "content": " provides initial blob parameters. The blobs are refined in optimization stage ① using Eq. 3 to better encode the query. The refined blob parameters are again modified in optimization stage ② using Eq. 1 to obtain a counterfactual example. The parameters of the encoder " + }, + { + "bbox": [ + 46, + 281, + 547, + 327 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 46, + 281, + 547, + 327 + ], + "type": "text", + "content": " and generator " + }, + { + "bbox": [ + 46, + 281, + 547, + 327 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 46, + 281, + 547, + 327 + ], + "type": "text", + "content": " are frozen during the optimization, as is " + }, + { + "bbox": [ + 46, + 281, + 547, + 327 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 281, + 547, + 327 + ], + "type": "text", + "content": " the decision model to be explained. Losses are illustrated with dashed arrows whose directions represent the gradient flow." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 347, + 235, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 347, + 235, + 359 + ], + "spans": [ + { + "bbox": [ + 47, + 347, + 235, + 359 + ], + "type": "text", + "content": "featuring a single and centered object [11, 57]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 359, + 290, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 359, + 290, + 587 + ], + "spans": [ + { + "bbox": [ + 46, + 359, + 290, + 587 + ], + "type": "text", + "content": "Scaling up to classifiers operating over more complex images such as crowded urban scenes remains a major challenge. To date, only STEEX [27] proposes a method able to explain decision models dealing with scenes composed of many objects. It employs a segmentation-to-image generation backbone [68], which helps for generating high-quality complex images but comes with two important limitations. Firstly, it requires annotated semantic segmentation masks of images to work, which is not always available or is costly for the user to produce. Secondly, it imposes that the explanation keeps the same spatial and semantic layout as the query, thereby severely restricting the search space. It is, for instance, not able to provide insights about the importance of the position of objects. By contrast, OCTET does not impose a fixed spatial layout. Indeed, the structural layout of the scene is an intermediate representation of the generator, inferred from images only. Therefore, we do not use additional annotations, and the spatial layout can be optimized by back-propagation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 597, + 285, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 597, + 285, + 612 + ], + "spans": [ + { + "bbox": [ + 47, + 597, + 285, + 612 + ], + "type": "text", + "content": "3. Object-aware Counterfactual Explanations" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "type": "text", + "content": "We present here OCTET, a method to produce counterfactual examples for compositional scenes, based on a generative architecture. We first formally describe our approach and objective in Sec. 3.1. Then, we instantiate the generative backbone in Sec. 3.2 and the counterfactual objective in Sec. 3.3. Lastly, we explain how we invert the generator to recover the latent code for the query image in Sec. 3.4. The pipeline of OCTET is shown in Fig. 2." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 346, + 419, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 346, + 419, + 358 + ], + "spans": [ + { + "bbox": [ + 306, + 346, + 419, + 358 + ], + "type": "text", + "content": "3.1. Goal and Notations" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 365, + 547, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 365, + 547, + 498 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 547, + 498 + ], + "type": "text", + "content": "Given a decision model " + }, + { + "bbox": [ + 304, + 365, + 547, + 498 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 365, + 547, + 498 + ], + "type": "text", + "content": ", a query image " + }, + { + "bbox": [ + 304, + 365, + 547, + 498 + ], + "type": "inline_equation", + "content": "x^{q}" + }, + { + "bbox": [ + 304, + 365, + 547, + 498 + ], + "type": "text", + "content": ", and a target decision " + }, + { + "bbox": [ + 304, + 365, + 547, + 498 + ], + "type": "inline_equation", + "content": "y \\neq M(x^{q})" + }, + { + "bbox": [ + 304, + 365, + 547, + 498 + ], + "type": "text", + "content": ", a counterfactual explanation is an image " + }, + { + "bbox": [ + 304, + 365, + 547, + 498 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 365, + 547, + 498 + ], + "type": "text", + "content": " that is close to the query image " + }, + { + "bbox": [ + 304, + 365, + 547, + 498 + ], + "type": "inline_equation", + "content": "x^{q}" + }, + { + "bbox": [ + 304, + 365, + 547, + 498 + ], + "type": "text", + "content": " but modified so that the decision " + }, + { + "bbox": [ + 304, + 365, + 547, + 498 + ], + "type": "inline_equation", + "content": "M(x)" + }, + { + "bbox": [ + 304, + 365, + 547, + 498 + ], + "type": "text", + "content": " of the model changes to " + }, + { + "bbox": [ + 304, + 365, + 547, + 498 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 365, + 547, + 498 + ], + "type": "text", + "content": ". Generative counterfactual methods, such as ours, employ a trained generator " + }, + { + "bbox": [ + 304, + 365, + 547, + 498 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 304, + 365, + 547, + 498 + ], + "type": "text", + "content": " [27, 28, 42]. We denote its latent space " + }, + { + "bbox": [ + 304, + 365, + 547, + 498 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}" + }, + { + "bbox": [ + 304, + 365, + 547, + 498 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 304, + 365, + 547, + 498 + ], + "type": "inline_equation", + "content": "z^{q} \\in \\mathcal{Z}" + }, + { + "bbox": [ + 304, + 365, + 547, + 498 + ], + "type": "text", + "content": " the latent code that produces the best approximation of " + }, + { + "bbox": [ + 304, + 365, + 547, + 498 + ], + "type": "inline_equation", + "content": "x^{q}" + }, + { + "bbox": [ + 304, + 365, + 547, + 498 + ], + "type": "text", + "content": " when passed through " + }, + { + "bbox": [ + 304, + 365, + 547, + 498 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 304, + 365, + 547, + 498 + ], + "type": "text", + "content": ", i.e., so that " + }, + { + "bbox": [ + 304, + 365, + 547, + 498 + ], + "type": "inline_equation", + "content": "G(z^{q}) \\approx x^{q}" + }, + { + "bbox": [ + 304, + 365, + 547, + 498 + ], + "type": "text", + "content": ". To explain differentiable deep vision models, the definition of a counterfactual explanation stated here-above can be translated into the following optimization problem:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 319, + 506, + 545, + 524 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 506, + 545, + 524 + ], + "spans": [ + { + "bbox": [ + 319, + 506, + 545, + 524 + ], + "type": "interline_equation", + "content": "\\underset {z \\in \\mathcal {Z}} {\\operatorname {a r g m i n}} L _ {\\text {d e c i s i o n}} \\left(M (G (z)), y\\right) + \\lambda_ {\\text {d i s t}} L _ {\\text {d i s t}} \\left(z, z ^ {q}\\right). \\tag {1}", + "image_path": "240f5d3a4dc3b7af6e841973341b42bd71dd62f97aed6122916bd7712d7632ce.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 532, + 547, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 532, + 547, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 532, + 547, + 665 + ], + "type": "text", + "content": "The role of the first term " + }, + { + "bbox": [ + 304, + 532, + 547, + 665 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{decision}}" + }, + { + "bbox": [ + 304, + 532, + 547, + 665 + ], + "type": "text", + "content": " is to push the decision of the model " + }, + { + "bbox": [ + 304, + 532, + 547, + 665 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 532, + 547, + 665 + ], + "type": "text", + "content": " for the generated image " + }, + { + "bbox": [ + 304, + 532, + 547, + 665 + ], + "type": "inline_equation", + "content": "G(z)" + }, + { + "bbox": [ + 304, + 532, + 547, + 665 + ], + "type": "text", + "content": " to the target class " + }, + { + "bbox": [ + 304, + 532, + 547, + 665 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 532, + 547, + 665 + ], + "type": "text", + "content": "; The second term " + }, + { + "bbox": [ + 304, + 532, + 547, + 665 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{dist}}" + }, + { + "bbox": [ + 304, + 532, + 547, + 665 + ], + "type": "text", + "content": " is a distance term in the latent space of " + }, + { + "bbox": [ + 304, + 532, + 547, + 665 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 304, + 532, + 547, + 665 + ], + "type": "text", + "content": ", which ensures that the query " + }, + { + "bbox": [ + 304, + 532, + 547, + 665 + ], + "type": "inline_equation", + "content": "x^q" + }, + { + "bbox": [ + 304, + 532, + 547, + 665 + ], + "type": "text", + "content": " and the explanation " + }, + { + "bbox": [ + 304, + 532, + 547, + 665 + ], + "type": "inline_equation", + "content": "G(z)" + }, + { + "bbox": [ + 304, + 532, + 547, + 665 + ], + "type": "text", + "content": " remain similar; " + }, + { + "bbox": [ + 304, + 532, + 547, + 665 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{dist}}" + }, + { + "bbox": [ + 304, + 532, + 547, + 665 + ], + "type": "text", + "content": " is an hyper-parameter that controls the relative weight between the two terms. With optimal " + }, + { + "bbox": [ + 304, + 532, + 547, + 665 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 304, + 532, + 547, + 665 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 532, + 547, + 665 + ], + "type": "inline_equation", + "content": "G(z)" + }, + { + "bbox": [ + 304, + 532, + 547, + 665 + ], + "type": "text", + "content": " then recovers a counterfactual image for the query " + }, + { + "bbox": [ + 304, + 532, + 547, + 665 + ], + "type": "inline_equation", + "content": "x^q" + }, + { + "bbox": [ + 304, + 532, + 547, + 665 + ], + "type": "text", + "content": " and the decision model " + }, + { + "bbox": [ + 304, + 532, + 547, + 665 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 532, + 547, + 665 + ], + "type": "text", + "content": ". Since the choice of " + }, + { + "bbox": [ + 304, + 532, + 547, + 665 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{decision}}" + }, + { + "bbox": [ + 304, + 532, + 547, + 665 + ], + "type": "text", + "content": " is constrained by the nature of the model to explain, e.g., a classification loss if " + }, + { + "bbox": [ + 304, + 532, + 547, + 665 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 532, + 547, + 665 + ], + "type": "text", + "content": " is a classifier, we only have to specify " + }, + { + "bbox": [ + 304, + 532, + 547, + 665 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 304, + 532, + 547, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 532, + 547, + 665 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{dist}}" + }, + { + "bbox": [ + 304, + 532, + 547, + 665 + ], + "type": "text", + "content": " to define our counterfactual method." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 671, + 496, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 671, + 496, + 685 + ], + "spans": [ + { + "bbox": [ + 306, + 671, + 496, + 685 + ], + "type": "text", + "content": "3.2. Compositional generative backbone" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "type": "text", + "content": "In this work, we focus on explaining decision models that operate on compositional visual scenes and we look for" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "15064" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 167 + ], + "type": "text", + "content": "a generative model capable of modeling such scenes. However, most generative models are designed for images containing centered, isolated objects, and they do not learn a latent space that can easily be disentangled into per-object representations. Instead, we propose that our generative backbone should be object-based by design [14, 37]. Such methods are fully differentiable and trained without supervision to generate images in a compositional fashion." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 168, + 288, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 168, + 288, + 420 + ], + "spans": [ + { + "bbox": [ + 46, + 168, + 288, + 420 + ], + "type": "text", + "content": "We choose to build on BlobGAN [14], a compositional generative model that shows good performances and has the advantage of allowing not only edition but also insertion and removal of objects in a differentiable way. In BlobGAN, images are generated in two steps. First, a layout network maps a random Gaussian noise vector to an intermediary representation consisting of an ordered set of " + }, + { + "bbox": [ + 46, + 168, + 288, + 420 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 168, + 288, + 420 + ], + "type": "text", + "content": " 'blobs'. Each blob " + }, + { + "bbox": [ + 46, + 168, + 288, + 420 + ], + "type": "inline_equation", + "content": "k \\in \\{1 \\dots K\\}" + }, + { + "bbox": [ + 46, + 168, + 288, + 420 + ], + "type": "text", + "content": " is defined by spatial parameters " + }, + { + "bbox": [ + 46, + 168, + 288, + 420 + ], + "type": "inline_equation", + "content": "\\phi_k" + }, + { + "bbox": [ + 46, + 168, + 288, + 420 + ], + "type": "text", + "content": " and a style feature vector " + }, + { + "bbox": [ + 46, + 168, + 288, + 420 + ], + "type": "inline_equation", + "content": "\\psi_k" + }, + { + "bbox": [ + 46, + 168, + 288, + 420 + ], + "type": "text", + "content": ". The spatial parameters " + }, + { + "bbox": [ + 46, + 168, + 288, + 420 + ], + "type": "inline_equation", + "content": "\\phi_k" + }, + { + "bbox": [ + 46, + 168, + 288, + 420 + ], + "type": "text", + "content": " consist of five values that are its center coordinates, scale, aspect ratio, and rotation angle. In practice, the layout network is a neural network that outputs vectors " + }, + { + "bbox": [ + 46, + 168, + 288, + 420 + ], + "type": "inline_equation", + "content": "\\phi_k" + }, + { + "bbox": [ + 46, + 168, + 288, + 420 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 168, + 288, + 420 + ], + "type": "inline_equation", + "content": "\\psi_k" + }, + { + "bbox": [ + 46, + 168, + 288, + 420 + ], + "type": "text", + "content": " for each blob. Then, in the second stage, the generator network is tasked to transform the blob parameters into an image. A rendering layer draws the blobs on a canvas as ellipses whose shapes and positions are given directly by " + }, + { + "bbox": [ + 46, + 168, + 288, + 420 + ], + "type": "inline_equation", + "content": "\\phi_k" + }, + { + "bbox": [ + 46, + 168, + 288, + 420 + ], + "type": "text", + "content": ". The output of the rendering layer is a map that associates to each 2D spatial location a feature vector computed from the style vectors of the blobs present at that position. This feature map is transformed by convolutional layers into an image that roughly follows the spatial organization of the canvas." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 421, + 287, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 421, + 287, + 515 + ], + "spans": [ + { + "bbox": [ + 46, + 421, + 287, + 515 + ], + "type": "text", + "content": "The layout network and the generator network are trained end-to-end against a discriminator with standard GAN losses only. In particular, we do not use any annotations for the number, size, class, or location of objects. They are discovered by the generator during training. Note also that scale parameters can take non-positive values. In that case, the object is considered absent and the corresponding blob is not rendered on the canvas in the generator network." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 525, + 239, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 525, + 239, + 539 + ], + "spans": [ + { + "bbox": [ + 47, + 525, + 239, + 539 + ], + "type": "text", + "content": "3.3. Setting " + }, + { + "bbox": [ + 47, + 525, + 239, + 539 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{dist}}" + }, + { + "bbox": [ + 47, + 525, + 239, + 539 + ], + "type": "text", + "content": " for compositional scenes" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 545, + 287, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 287, + 700 + ], + "type": "text", + "content": "Using a compositional generative backbone allows us to design a meaningful, disentangled, per-object distance for images. Indeed, to be interpretable with respect to the query image, a counterfactual explanation should display sparse changes in terms of the number of moved, added, removed, or modified objects. Therefore, we design " + }, + { + "bbox": [ + 46, + 545, + 287, + 700 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{dist}}" + }, + { + "bbox": [ + 46, + 545, + 287, + 700 + ], + "type": "text", + "content": " such that it is split into per-object distances where each object can be addressed independently. Moreover, we propose that modifications of the spatial position of an object should be orthogonal to those made to its shape or colors. Since in a well-trained BlobGAN each blob is associated with a distinct element of the scene (see Appendix), such a distance can be naturally defined in the space of blob parameters." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "text", + "content": "Formally, let us denote " + }, + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "text", + "content": " the concatenation across" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "text", + "content": "all blobs " + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "k \\in \\{1 \\dots K\\}" + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "text", + "content": " of spatial parameters " + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "\\phi_k" + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "text", + "content": " and style vectors " + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "\\psi_k" + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "z = (\\phi, \\psi)" + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "text", + "content": " the complete latent representation that the generator network " + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "text", + "content": " can decode into an image. Accordingly, the latent code for the query image " + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "x^q" + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "text", + "content": " reads " + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "z^q = (\\phi^q, \\psi^q)" + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "text", + "content": ". Then, OCTET implements " + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{dist}}" + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 330, + 143, + 545, + 175 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 143, + 545, + 175 + ], + "spans": [ + { + "bbox": [ + 330, + 143, + 545, + 175 + ], + "type": "interline_equation", + "content": "L _ {\\text {d i s t}} (z, z ^ {q}) = \\sum_ {k = 1} ^ {K} \\| \\phi_ {k} - \\phi_ {k} ^ {q} \\| _ {1} + \\| \\psi_ {k} - \\psi_ {k} ^ {q} \\| _ {1}. \\tag {2}", + "image_path": "38081872ddf8741c1df2387edd4d7332cfe59c1ceb5eba9fcd76252d6a045432.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 186, + 545, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 186, + 545, + 258 + ], + "spans": [ + { + "bbox": [ + 304, + 186, + 545, + 258 + ], + "type": "text", + "content": "We use the " + }, + { + "bbox": [ + 304, + 186, + 545, + 258 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 304, + 186, + 545, + 258 + ], + "type": "text", + "content": " distance here to promote sparsity in the number of modified blobs and the type of edits applied to each blob. If the user wishes to inspect the role of the style of objects, their spatial positions, or both, with a counterfactual explanation, she can instead optimize Eq. 2 only on the corresponding features of the blobs." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 258, + 545, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 258, + 545, + 306 + ], + "spans": [ + { + "bbox": [ + 304, + 258, + 545, + 306 + ], + "type": "text", + "content": "In order to specify Eq. 2, and thus the final counterfactual objective (Eq. 1), it remains to be solved how to obtain the blob-wise latent codes " + }, + { + "bbox": [ + 304, + 258, + 545, + 306 + ], + "type": "inline_equation", + "content": "\\phi_k^q" + }, + { + "bbox": [ + 304, + 258, + 545, + 306 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 258, + 545, + 306 + ], + "type": "inline_equation", + "content": "\\psi_k^q" + }, + { + "bbox": [ + 304, + 258, + 545, + 306 + ], + "type": "text", + "content": " associated with the query image " + }, + { + "bbox": [ + 304, + 258, + 545, + 306 + ], + "type": "inline_equation", + "content": "x^q" + }, + { + "bbox": [ + 304, + 258, + 545, + 306 + ], + "type": "text", + "content": ". We discuss this matter in the next section." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 316, + 450, + 329 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 316, + 450, + 329 + ], + "spans": [ + { + "bbox": [ + 306, + 316, + 450, + 329 + ], + "type": "text", + "content": "3.4. Inverting the query image" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 335, + 545, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 335, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 304, + 335, + 545, + 431 + ], + "type": "text", + "content": "Obtaining the latent code corresponding to the query image, a process called inversion, is not straightforward when using a GAN-based generator. To produce a counterfactual example, such inversion needs to find latent codes " + }, + { + "bbox": [ + 304, + 335, + 545, + 431 + ], + "type": "inline_equation", + "content": "(\\phi^q,\\psi^q)" + }, + { + "bbox": [ + 304, + 335, + 545, + 431 + ], + "type": "text", + "content": " so that " + }, + { + "bbox": [ + 304, + 335, + 545, + 431 + ], + "type": "inline_equation", + "content": "G(\\phi^q,\\psi^q)" + }, + { + "bbox": [ + 304, + 335, + 545, + 431 + ], + "type": "text", + "content": " recovers " + }, + { + "bbox": [ + 304, + 335, + 545, + 431 + ], + "type": "inline_equation", + "content": "x^{q}" + }, + { + "bbox": [ + 304, + 335, + 545, + 431 + ], + "type": "text", + "content": ". For the downstream counterfactual optimization to make sense, the decision of the model for the query image should be preserved for the reconstructed image, i.e., " + }, + { + "bbox": [ + 304, + 335, + 545, + 431 + ], + "type": "inline_equation", + "content": "M(G(\\phi^q,\\psi^q)) = M(x^q)" + }, + { + "bbox": [ + 304, + 335, + 545, + 431 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 432, + 545, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 432, + 545, + 515 + ], + "spans": [ + { + "bbox": [ + 304, + 432, + 545, + 515 + ], + "type": "text", + "content": "To achieve those objectives, we follow best practices for image inversion [1,2,41,50,67]. We first learn an encoder " + }, + { + "bbox": [ + 304, + 432, + 545, + 515 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 304, + 432, + 545, + 515 + ], + "type": "text", + "content": " to predict blob parameters from input images. Generating an image " + }, + { + "bbox": [ + 304, + 432, + 545, + 515 + ], + "type": "inline_equation", + "content": "G(E(x^q))" + }, + { + "bbox": [ + 304, + 432, + 545, + 515 + ], + "type": "text", + "content": " with parameters predicted this way yields a coarse reconstruction of " + }, + { + "bbox": [ + 304, + 432, + 545, + 515 + ], + "type": "inline_equation", + "content": "x^q" + }, + { + "bbox": [ + 304, + 432, + 545, + 515 + ], + "type": "text", + "content": ". To improve the inversion, we start from predicted blob parameters " + }, + { + "bbox": [ + 304, + 432, + 545, + 515 + ], + "type": "inline_equation", + "content": "E(x^q)" + }, + { + "bbox": [ + 304, + 432, + 545, + 515 + ], + "type": "text", + "content": " and further optimize them with the following objective:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 526, + 545, + 571 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 526, + 545, + 571 + ], + "spans": [ + { + "bbox": [ + 306, + 526, + 545, + 571 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\phi^ {q}, \\psi^ {q} = \\underset {\\phi , \\psi} {\\arg \\min } L _ {\\mathrm {L P I P S}} (G (\\phi , \\psi), x ^ {q})) + L _ {2} (G (\\phi , \\psi), x ^ {q}) \\\\ + L _ {2} \\left(f _ {M} \\left(x ^ {q}\\right), f _ {M} (G (\\phi , \\psi))\\right) \\\\ + L _ {2} \\left(\\left(\\phi , \\psi\\right), E \\left(x ^ {q}\\right)\\right). \\tag {3} \\\\ \\end{array}", + "image_path": "ccbfb512a1d9e044e9d39172532827a3c4603c773b247aa12fb287ad6444cf1b.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": "The first two terms ensure that the image " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "G(\\phi, \\psi)" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " generated from the blob parameters reconstructs the query " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "x^q" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " using a perceptual " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{LPIPS}}" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " loss and the " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "L_2" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " loss. The third term aims to preserve the features that are important to the decision model " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " by adding a " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "L_2" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " loss on intermediate features " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "f_M(x)" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": ". Finally, the last term encourages the parameters to stay close to the reasonably good solution given by the encoder. This inversion method concludes the specification of the optimization problem of Eq. 1 in OCTET and of the pipeline depicted in Fig. 2. More details on the inversion and the training of the encoder are in the appendix." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15065" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 70, + 173, + 328 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 173, + 328 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 173, + 328 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 173, + 328 + ], + "type": "image", + "image_path": "f31e535168e4f9d9249280d5abad0182018c73380052c848ce4fd719bd7a5e83.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 332, + 546, + 389 + ], + "lines": [ + { + "bbox": [ + 46, + 332, + 546, + 389 + ], + "spans": [ + { + "bbox": [ + 46, + 332, + 546, + 389 + ], + "type": "text", + "content": "Figure 3. Counterfactual explanations on driving scenes. Explanations are generated for binary classifiers predicting the capacity of the ego-car to respectively Move forward, Stop, turn Left, or Right. We indicate at the top left of each image the output decision of the model for that image. For each row, counterfactuals target the opposite decision to that obtained for the query image. OCTET finds interpretable, sparse, and meaningful semantic and spatial modifications to the query image. OCTET style-only only optimizes on the style features, it is analog to STEEX [27] that can only search for changes in object appearance to flip the decision." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 173, + 70, + 301, + 327 + ], + "blocks": [ + { + "bbox": [ + 173, + 70, + 301, + 327 + ], + "lines": [ + { + "bbox": [ + 173, + 70, + 301, + 327 + ], + "spans": [ + { + "bbox": [ + 173, + 70, + 301, + 327 + ], + "type": "image", + "image_path": "662c96748894516838b570468917ecaf4e572d0a0221a528bbdad3e1b5ac202e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 301, + 70, + 422, + 327 + ], + "blocks": [ + { + "bbox": [ + 301, + 70, + 422, + 327 + ], + "lines": [ + { + "bbox": [ + 301, + 70, + 422, + 327 + ], + "spans": [ + { + "bbox": [ + 301, + 70, + 422, + 327 + ], + "type": "image", + "image_path": "513ce641086b18691d286a66acf1d2e978ae0552214c7927bacd25144a220cc5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 422, + 71, + 544, + 327 + ], + "blocks": [ + { + "bbox": [ + 422, + 71, + 544, + 327 + ], + "lines": [ + { + "bbox": [ + 422, + 71, + 544, + 327 + ], + "spans": [ + { + "bbox": [ + 422, + 71, + 544, + 327 + ], + "type": "image", + "image_path": "868051f78f2c7cfebb90c4a0a9c6249426f0c8d5c0395886a94835af47fdab24.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 407, + 128, + 422 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 407, + 128, + 422 + ], + "spans": [ + { + "bbox": [ + 47, + 407, + 128, + 422 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 427, + 269, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 427, + 269, + 441 + ], + "spans": [ + { + "bbox": [ + 47, + 427, + 269, + 441 + ], + "type": "text", + "content": "4.1. Experimental protocol and training details" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 446, + 289, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 446, + 289, + 638 + ], + "spans": [ + { + "bbox": [ + 46, + 446, + 289, + 638 + ], + "type": "text", + "content": "Data and decision model. Our method is designed to build counterfactual examples to explain decision models trained on compositional images. To evaluate our method, we follow previous work [27] and use the BDD100k dataset [60], which contains 100k images of diverse driving scenes that we resize to a resolution of " + }, + { + "bbox": [ + 46, + 446, + 289, + 638 + ], + "type": "inline_equation", + "content": "512 \\times 256" + }, + { + "bbox": [ + 46, + 446, + 289, + 638 + ], + "type": "text", + "content": ". In addition, we also use BDD-OIA [59], a 20K scene extension of BDD100k that is annotated with 4 binary attributes labeled Move Forward, Stop, Turn Left, and Turn Right. Each label is defined as the capacity in the given situation of the ego-vehicle to perform the corresponding action. Note that by this definition, being allowed to Stop and to Move Forward are not mutually exclusive. The decision models being explained in the experiments are multi-label binary DenseNet121 [26] classifiers trained on BDD-OIA with the four aforementioned labels." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 287, + 714 + ], + "type": "text", + "content": "Baseline. We compare our method against STEEX [27], the only counterfactual explanation method that performs well on compositional images like driving scenes. Note that in STEEX, the conservation of the semantic layout is hardcoded in the architecture: it is less flexible as an explanation method than ours. Moreover, the authors use semantic seg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 409, + 545, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 409, + 545, + 445 + ], + "spans": [ + { + "bbox": [ + 304, + 409, + 545, + 445 + ], + "type": "text", + "content": "mentation annotations of BDD100k to train their generative backbone, while we do not need any additional annotation. We use the pre-trained model from the official release." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 448, + 546, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 448, + 546, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 448, + 546, + 616 + ], + "type": "text", + "content": "Technical setup. For our method, we train the generative backbone (BlobGAN) on the training set of BDD100k, with a number " + }, + { + "bbox": [ + 304, + 448, + 546, + 616 + ], + "type": "inline_equation", + "content": "K = 40" + }, + { + "bbox": [ + 304, + 448, + 546, + 616 + ], + "type": "text", + "content": " of blobs. We adapt the architecture of BlobGAN to generate rectangular " + }, + { + "bbox": [ + 304, + 448, + 546, + 616 + ], + "type": "inline_equation", + "content": "512 \\times 256" + }, + { + "bbox": [ + 304, + 448, + 546, + 616 + ], + "type": "text", + "content": " images, and we decrease the size of the feature vectors of the BlobGAN to compensate for the increased number of blobs. Throughout our experiments, we evaluate two main variants of our method: OCTET where both spatial and style features are optimized altogether and OCTET style-only where only the style vector " + }, + { + "bbox": [ + 304, + 448, + 546, + 616 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 304, + 448, + 546, + 616 + ], + "type": "text", + "content": " is optimized while the spatial vector " + }, + { + "bbox": [ + 304, + 448, + 546, + 616 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 304, + 448, + 546, + 616 + ], + "type": "text", + "content": " is left equal to " + }, + { + "bbox": [ + 304, + 448, + 546, + 616 + ], + "type": "inline_equation", + "content": "\\phi^q" + }, + { + "bbox": [ + 304, + 448, + 546, + 616 + ], + "type": "text", + "content": ". This setting only assesses the contribution of style features, and is most similar to the STEEX baseline [27] that finds explanations while conserving the semantic layout of the query. More details in the appendix." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 623, + 455, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 623, + 455, + 636 + ], + "spans": [ + { + "bbox": [ + 306, + 623, + 455, + 636 + ], + "type": "text", + "content": "4.2. Quality of the explanations" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 641, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 546, + 715 + ], + "type": "text", + "content": "Counterfactual explanations are defined as images that are similar to the query image but effectively change the decision of the target model. To verify that our explanations fit the criteria, we evaluate them with the following metrics: 1) the FID [24] which measures if they are realistic images, and 2) a perceptual distance LPIPS [65] that mea" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15066" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 68, + 548, + 217 + ], + "blocks": [ + { + "bbox": [ + 47, + 68, + 548, + 217 + ], + "lines": [ + { + "bbox": [ + 47, + 68, + 548, + 217 + ], + "spans": [ + { + "bbox": [ + 47, + 68, + 548, + 217 + ], + "type": "image", + "image_path": "06746fddb885cfc21d5ade7ca54e393a7c56d79008c40646fcfdcb216c87ec72.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 220, + 547, + 253 + ], + "lines": [ + { + "bbox": [ + 46, + 220, + 547, + 253 + ], + "spans": [ + { + "bbox": [ + 46, + 220, + 547, + 253 + ], + "type": "text", + "content": "Figure 4. FID vs. LPIPS, labeled with Success rate for OCTET and STEEX [27] explaining a decision model trained on the 4 labels of BDD-OIA. Data points are obtained with different values of " + }, + { + "bbox": [ + 46, + 220, + 547, + 253 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{dist}}" + }, + { + "bbox": [ + 46, + 220, + 547, + 253 + ], + "type": "text", + "content": " yielding different trade-offs. STEEX is directly comparable to OCTET style-only as they operate in similar conditions. See Sec. 4.2." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 269, + 289, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 269, + 289, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 269, + 289, + 604 + ], + "type": "text", + "content": "sures the change magnitude between them and the query image. We also compute the success rate corresponding to the percentage of counterfactual examples that are indeed classified in the target class by the decision model. We compute these metrics for different values of " + }, + { + "bbox": [ + 46, + 269, + 289, + 604 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{dist}}" + }, + { + "bbox": [ + 46, + 269, + 289, + 604 + ], + "type": "text", + "content": " that lead to different trade-offs and plot the scores in Fig. 4. As STEEX explanations are constrained to style features, we compare it to our OCTET style-only setting. Firstly, in terms of FID, we can observe that both versions of OCTET obtain lower FID than STEEX: OCTET produces explanations of better visual quality, even though it does not use any annotations as opposed to STEEX, which needs semantic maps, and handles spatial explanations that are more challenging to generate. Secondly, in terms of LPIPS, for style-based explanations OCTET style-only stays closer to the query than STEEX. This can be explained as STEEX latent representation is semantic-based, while ours is instance-based. Our method is therefore able to seek finer-grained modifications, while STEEX tends to change the style for an entire semantic class even when only one instance is important to the decision model. As for OCTET, because LPIPS disproportionately weights structural modifications, the wide range of attainable values indicates that many degrees of changes can be obtained depending on the choice of " + }, + { + "bbox": [ + 46, + 269, + 289, + 604 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{dist}}" + }, + { + "bbox": [ + 46, + 269, + 289, + 604 + ], + "type": "text", + "content": ". Some settings tend to move objects a lot, while other settings attain values similar to 'style-only' edits. Finally, the very high achievable Success Rates displayed by all three models confirm that they will propose explanations in almost any situation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": "We display examples of OCTET's counterfactual explanations in Fig. 3. OCTET finds sparse spatial and/or semantic modifications to objects that strongly influence the output decision. For example, we can see with OCTET that the distance with the car in front plays a role when deciding for 'Forward' (1st row) or 'Stop' (2nd row). On the other hand, OCTET style-only shows that for those same decisions, the brake lights also are taken into account. For going 'Left' class (3rd row), the absence of the continuous line seems" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 308, + 266, + 430, + 327 + ], + "blocks": [ + { + "bbox": [ + 308, + 266, + 430, + 327 + ], + "lines": [ + { + "bbox": [ + 308, + 266, + 430, + 327 + ], + "spans": [ + { + "bbox": [ + 308, + 266, + 430, + 327 + ], + "type": "image", + "image_path": "02ecf044eeeb09d18676375f0f91775f46ec942e42550eda0fa0ac116ee09fc0.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 431, + 267, + 544, + 326 + ], + "blocks": [ + { + "bbox": [ + 431, + 267, + 544, + 326 + ], + "lines": [ + { + "bbox": [ + 431, + 267, + 544, + 326 + ], + "spans": [ + { + "bbox": [ + 431, + 267, + 544, + 326 + ], + "type": "image", + "image_path": "cf2a1f9c734e9004d3040b14da5d2538c94a2698debb9edb5b6896c23da7e16d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 308, + 327, + 430, + 399 + ], + "blocks": [ + { + "bbox": [ + 308, + 327, + 430, + 399 + ], + "lines": [ + { + "bbox": [ + 308, + 327, + 430, + 399 + ], + "spans": [ + { + "bbox": [ + 308, + 327, + 430, + 399 + ], + "type": "image", + "image_path": "0ba562000555d27052b79a88c49fcfa6b01570a01e619c233aa7dac408cc5c17.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 431, + 327, + 544, + 399 + ], + "blocks": [ + { + "bbox": [ + 431, + 327, + 544, + 399 + ], + "lines": [ + { + "bbox": [ + 431, + 327, + 544, + 399 + ], + "spans": [ + { + "bbox": [ + 431, + 327, + 544, + 399 + ], + "type": "image", + "image_path": "4cc55b71d51124381d24eff6d9917f9818338f0f7b6c629b64984ab2af6408de.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 316, + 384, + 430, + 443 + ], + "blocks": [ + { + "bbox": [ + 316, + 384, + 430, + 443 + ], + "lines": [ + { + "bbox": [ + 316, + 384, + 430, + 443 + ], + "spans": [ + { + "bbox": [ + 316, + 384, + 430, + 443 + ], + "type": "image", + "image_path": "e0695404d0381ce301bcafbc942476d2f8f1bc6bb2920e79b4047bef45f7313c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 455, + 545, + 499 + ], + "lines": [ + { + "bbox": [ + 306, + 455, + 545, + 499 + ], + "spans": [ + { + "bbox": [ + 306, + 455, + 545, + 499 + ], + "type": "text", + "content": "Figure 5. Targeted counterfactual explanations. For each query image (top row), we target different blobs (2nd and 3rd row) to inspect the role of corresponding objects. The output of the decision model is reported on the top right corner for each image." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 431, + 384, + 544, + 443 + ], + "blocks": [ + { + "bbox": [ + 431, + 384, + 544, + 443 + ], + "lines": [ + { + "bbox": [ + 431, + 384, + 544, + 443 + ], + "spans": [ + { + "bbox": [ + 431, + 384, + 544, + 443 + ], + "type": "image", + "image_path": "382776078b54f5ced820578e6c6683388632bf6d795c6dd4e5da08983f492584.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 520, + 547, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 520, + 547, + 568 + ], + "spans": [ + { + "bbox": [ + 304, + 520, + 547, + 568 + ], + "type": "text", + "content": "crucial. On the 3rd and 4th row, we can see that OCTET mixes both style (brake lights) and spatial (proximity) explanations. On the same tasks, STEEX changes the style of the whole class, and not just the necessary instance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 575, + 446, + 587 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 575, + 446, + 587 + ], + "spans": [ + { + "bbox": [ + 306, + 575, + 446, + 587 + ], + "type": "text", + "content": "4.3. Targeted counterfactuals" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": "Models that operate on compositional scenes can have their decisions influenced by many different factors, and we have shown that OCTET can expose those factors in its counterfactual examples. However, the user may want to test some particular hypothesis dealing with specific objects or areas of the scene. Accordingly, the user can decide to optimize Eq. 2 only on selected blobs parameters. Finding the blob corresponding to a specific object in the image is discussed in Appendix. For instance, in the situation shown at the top left of Fig. 5, the car at the front seems to have" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15067" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 547, + 254 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 547, + 254 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 547, + 254 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 547, + 254 + ], + "type": "image", + "image_path": "26199a254b14a14d894dd2752db6136ce1d86e516264b0c238cdd47b2550a2b5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 266, + 546, + 312 + ], + "lines": [ + { + "bbox": [ + 46, + 266, + 546, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 546, + 312 + ], + "type": "text", + "content": "Figure 6. Counterfactual explanations for a segmentation model. The segmentation model predicts the driveable area with primary and secondary lanes shown in red and blue respectively. The first row shows query images and segmentations by the model. In the second row, we add or remove parts of the original masks to build target masks (changes are shown with white dotted lines). The third row displays counterfactual explanations for the target masks and predictions on the new images. Green ellipses are manually added to highlight details." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 331, + 289, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 331, + 289, + 464 + ], + "spans": [ + { + "bbox": [ + 46, + 331, + 289, + 464 + ], + "type": "text", + "content": "its brake lights on, but the decision is still not to 'Stop'. Targeting the front car (2nd row left image) reveals that the decision would have been 'Stop' if the car was closer, and the brake light more clearly contrasted. OCTET can also target blobs that are currently inactive, corresponding to no visible object in the image. In this example, doing so for car blobs shows that if another vehicle also had brake lights on, the model would also decide to 'Stop' (3rd row left image). This finding is unexpected, and could warrant further investigation of the decision model. We show in Fig. 5 and Fig. 1 some other results we obtain by targeting specific blobs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 475, + 270, + 489 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 475, + 270, + 489 + ], + "spans": [ + { + "bbox": [ + 47, + 475, + 270, + 489 + ], + "type": "text", + "content": "4.4. Explaining a semantic segmentation model" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 496, + 287, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 496, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 496, + 287, + 616 + ], + "type": "text", + "content": "For complex tasks, like autonomous driving, decision models usually rely on multiple neural networks that address different subtasks (e.g., traffic sign recognition, depth estimation, motion forecasting). In this context, it is important to explain models beyond classifiers [25,54], a question not addressed in previous counterfactual methods. Here, we show the versatility of our method as we extend it to explain a semantic segmentation network. To do so, we use the same optimization problem described in Eq. 1 and Sec. 3 but we change " + }, + { + "bbox": [ + 46, + 496, + 287, + 616 + ], + "type": "inline_equation", + "content": "L_{\\text{decision}}" + }, + { + "bbox": [ + 46, + 496, + 287, + 616 + ], + "type": "text", + "content": " to an image segmentation loss." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "type": "text", + "content": "In Fig. 6, we show counterfactual explanations for a driveable area segmentation model. More precisely, OCTET explains outputs of a DeepLabv3 [9] trained to segment images from BDD100k into classes of driveable area: the primary lane, the secondary side lanes, and the background. We define the target masks by manually modifying predictions made on the query images as illustrated in the second row. We see in the obtained counterfactuals, in the" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 331, + 547, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 331, + 547, + 464 + ], + "spans": [ + { + "bbox": [ + 304, + 331, + 547, + 464 + ], + "type": "text", + "content": "third row, changes in road markings and vehicles' positions. They confirm that the segmentation model will consider as driveable a lane behind broken white lines (col. 3) but not behind a double yellow line (col. 2), especially if, in addition, the lane is hatched (col. 3). If there is a car on a drivable lane, the model can detect that the occupied portion is non-driveable while considering that the rest of the lane remains accessible. With these examples, OCTET provides explanations for both large semantic regions, e.g., when we add or remove an entire region in the target mask, or small subparts, e.g., when we slightly crop or extend a region." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 475, + 484, + 489 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 475, + 484, + 489 + ], + "spans": [ + { + "bbox": [ + 306, + 475, + 484, + 489 + ], + "type": "text", + "content": "4.5. User-study to assess explanations" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 496, + 545, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 496, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 496, + 545, + 640 + ], + "type": "text", + "content": "Metrics presented in Sec. 4.2 measure how well the produced examples match the definition of counterfactual explanations. We want here to assess how useful the explanations are for a user to better understand a decision model. However, designing such an evaluation is challenging. Indeed, it has to avoid the many biases that appear [13, 36]. Moreover, we usually do not have access to the inner workings of a model, and therefore we cannot compare users' insights on its behavior against a ground truth. To overcome these challenges, we take inspiration from recent work on the usefulness of saliency maps as explanations [15] and we adapt its evaluation protocol to counterfactual explanations." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 641, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 714 + ], + "type": "text", + "content": "The goal is to measure if accessing counterfactual explanations on some examples helps users to better understand the model. Grounded within previous literature [16,22,30], we instantiate understand as 'being able to predict the model's output for new instances', a concept known as simulatability. In particular, our user-study has two phases:" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15068" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 70, + 286, + 115 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 286, + 115 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 286, + 115 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 286, + 115 + ], + "type": "table", + "html": "
Cohort sizeReplicationBias Detection
Group w/ OCTET2070%65%
Group control2052.5%0%
", + "image_path": "d274907b67a0c6707ceab17b4b4752903e4a17788ba6607b774777ed6c7d34d3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 123, + 287, + 189 + ], + "lines": [ + { + "bbox": [ + 46, + 123, + 287, + 189 + ], + "spans": [ + { + "bbox": [ + 46, + 123, + 287, + 189 + ], + "type": "text", + "content": "Table 1. Results of the user-study. The performance of users who had access to explanations ('w/ OCTET') in the observation phase is compared to that of users who didn't ('control'). 'Replication' measures the accuracy of the users in predicting the model's output on new samples and 'bias detection' is the proportion of users who found the spurious correlation in the model." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 49, + 212, + 286, + 307 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 50, + 212, + 286, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 50, + 212, + 286, + 258 + ], + "spans": [ + { + "bbox": [ + 50, + 212, + 286, + 258 + ], + "type": "text", + "content": "1. Observation phase: participants are shown a series of triplets, each being composed of (1) a query image, (2) the model's decision on that image, and (3), a counterfactual explanation for the decision." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 260, + 286, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 260, + 286, + 307 + ], + "spans": [ + { + "bbox": [ + 49, + 260, + 286, + 307 + ], + "type": "text", + "content": "2. Questionnaire phase: participants are presented with new images and are asked to guess the output of the model on these images. We stress that during the questionnaire, no counterfactual images are shown." + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 46, + 309, + 286, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 309, + 286, + 475 + ], + "spans": [ + { + "bbox": [ + 46, + 309, + 286, + 475 + ], + "type": "text", + "content": "To measure the impact that counterfactual explanations of the first phase have on the users' ability to predict the model output on new samples (second phase), we use a control group. This control group, composed of different participants, undertakes the exact same tasks except that counterfactual explanations are not shown during the first phase. The decision model used in this study is a classifier trained on BDD-OIA [59] to predict the binary 'Turn Right' label. Besides, the decision model is trained, by design, with a flaw: the presence of a car on any side of the road can impact the decision, not only obstacles on the right side. As discussed further below, this bias is introduced to study if users are able to identify model defects thanks to counterfactuals. More details are provided in the appendix." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 481, + 286, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 481, + 286, + 647 + ], + "spans": [ + { + "bbox": [ + 46, + 481, + 286, + 647 + ], + "type": "text", + "content": "Replication score. We first measure the 'replication score': the proportion of correct match between the user's answer and the prediction of the decision model, averaged over participants of the group. We report these results in Tab. 1 and observe that participants in the group accessing counterfactuals during the observation phase can successfully guess the decision model's prediction " + }, + { + "bbox": [ + 46, + 481, + 286, + 647 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 46, + 481, + 286, + 647 + ], + "type": "text", + "content": " of the time while the control group's performance is " + }, + { + "bbox": [ + 46, + 481, + 286, + 647 + ], + "type": "inline_equation", + "content": "52.5\\%" + }, + { + "bbox": [ + 46, + 481, + 286, + 647 + ], + "type": "text", + "content": ", barely above random. A statistical unpaired t-test, with as null hypothesis that the counterfactual explanations have no effect, confirms the significance of our result with a p-value of 0.0028. This demonstrates the usefulness of OCTET in understanding the decision model as participants with explanations could better anticipate the model's behavior on new instances." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 653, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 286, + 712 + ], + "type": "text", + "content": "Bias detection. Another promise of counterfactual explanations is to enable bias detection. We then want to assess whether, by looking at explanations, users can tell if and how a decision model is biased. In that direction, previous works dealing with human face datasets train biased models" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 419 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 419 + ], + "type": "text", + "content": "by artificially confounding attributes, e.g., smile and age, in the training set. Then, they develop bias detection benchmarks that measure, by means of an oracle neural network, how well attributes remain confounded in the counterfactual explanations [28, 42, 46]. However, a serious limitation of that benchmark is that it keeps end-users out of the evaluation loop and it cannot tell whether or not a user will be able to uncover the spurious correlations by looking at the counterfactual explanation. Moreover, detecting the issue with the model can be especially difficult when the bias is not obvious or its presence not suspected in the first place. Instead of an automated evaluation, at the end of the questionnaire, we asked the participants to describe in free text what they had understood from the model. The goal is to prompt them to mention the issue we embedded in the decision model, i.e., that the presence of a car on the left-hand side of the image factors in the 'Turn Right' decision. The 'bias detection' score simply counts, for each group, the proportion of participants that used the word 'left' in their answers. In Tab. 1, we see that " + }, + { + "bbox": [ + 304, + 72, + 545, + 419 + ], + "type": "inline_equation", + "content": "65\\%" + }, + { + "bbox": [ + 304, + 72, + 545, + 419 + ], + "type": "text", + "content": " of the participants that had access to the explanations from OCTET identify the spurious correlation learned by the model, while none of the users that did not have the explanations (control group) mentions it. This clearly shows that our method is useful to pinpoint issues with the decision model. Also, the fact that not every participant detected it confirms our suspicions against automated bias detection benchmarks: the presence of the information in the counterfactual examples is not sufficient to ensure that the issue will be found in practice." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 432, + 378, + 444 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 432, + 378, + 444 + ], + "spans": [ + { + "bbox": [ + 306, + 432, + 378, + 444 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 453, + 545, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 453, + 545, + 608 + ], + "spans": [ + { + "bbox": [ + 304, + 453, + 545, + 608 + ], + "type": "text", + "content": "In this work, we presented OCTET, a generative method to produce counterfactual explanations for deep vision models working on complex scenes. Thanks to an object-centric representation of the scene, OCTET is able to process complex compositional scenes and to assess the impact of both the spatial positions and the style features of each object. Our tool allows in-depth exploration of the different explaining factors by letting the user adjust on the fly the weights of the different search directions independently. We evaluated our method thoroughly on real driving scenes from BDD100k, with an improved evaluation benchmark compared to previous works, including a human-centered study that confirmed the usefulness of the approach." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 621, + 408, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 621, + 408, + 635 + ], + "spans": [ + { + "bbox": [ + 306, + 621, + 408, + 635 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 641, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 712 + ], + "type": "text", + "content": "This work was supported in part by the ANR grants VISA DEEP (ANR-20-CHIA-0022) and MultiTrans (ANR-21-CE23-0032). Authors would like to thank the voluntary participants of the user-study, as well as Remi Cadene, Julien Colin, Thomas Fel, and Guillaume Jeanneret for helpful discussions." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15069" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "text", + "content": "[1] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan: How to embed images into the stylegan latent space? In ICCV, 2019. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 124, + 287, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 124, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 53, + 124, + 287, + 156 + ], + "type": "text", + "content": "[2] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan++: How to edit the embedded images? In CVPR, 2020. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 158, + 288, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 158, + 288, + 191 + ], + "spans": [ + { + "bbox": [ + 53, + 158, + 288, + 191 + ], + "type": "text", + "content": "[3] Julius Adebayo, Justin Gilmer, Michael Muelly, Ian J. Goodfellow, Moritz Hardt, and Been Kim. Sanity checks for saliency maps. In NeurIPS, 2018. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 192, + 287, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 192, + 287, + 223 + ], + "spans": [ + { + "bbox": [ + 53, + 192, + 287, + 223 + ], + "type": "text", + "content": "[4] Maximilian Augustin, Valentyn Boreiko, Francesco Croce, and Matthias Hein. Diffusion visual counterfactual explanations. In NeurIPS, 2022. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 224, + 288, + 269 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 224, + 288, + 269 + ], + "spans": [ + { + "bbox": [ + 53, + 224, + 288, + 269 + ], + "type": "text", + "content": "[5] Mariusz Bojarski, Anna Choromanska, Krzysztof Choromanski, Bernhard Firner, Larry J Ackel, Urs Muller, Phil Yeres, and Karol Zieba. Visualbackprop: Efficient visualization of cnns for autonomous driving. In ICRA, 2018. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 270, + 287, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 270, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 53, + 270, + 287, + 312 + ], + "type": "text", + "content": "[6] Kieran Browne and Ben Swift. Semantics and explanation: why counterfactual explanations produce adversarial examples in deep neural networks. CoRR, abs/2012.10076, 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 313, + 288, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 313, + 288, + 346 + ], + "spans": [ + { + "bbox": [ + 53, + 313, + 288, + 346 + ], + "type": "text", + "content": "[7] Chun-Hao Chang, Elliot Creager, Anna Goldenberg, and David Duvenaud. Explaining image classifiers by counterfactual generation. In ICLR, 2019. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 346, + 288, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 346, + 288, + 379 + ], + "spans": [ + { + "bbox": [ + 53, + 346, + 288, + 379 + ], + "type": "text", + "content": "[8] Chaofan Chen, Oscar Li, Daniel Tao, Alina Barnett, Cynthia Rudin, and Jonathan Su. This looks like that: Deep learning for interpretable image recognition. In NeurIPS, 2019. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 380, + 288, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 288, + 413 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 288, + 413 + ], + "type": "text", + "content": "[9] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. CoRR, abs/1706.05587, 2017. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 414, + 287, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 414, + 287, + 445 + ], + "spans": [ + { + "bbox": [ + 48, + 414, + 287, + 445 + ], + "type": "text", + "content": "[10] Ricky T. Q. Chen, Xuechen Li, Roger Grosse, and David Duvenaud. Isolating sources of disentanglement in variational autoencoders. In NeurIPS, 2018. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 447, + 287, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 447, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 48, + 447, + 287, + 479 + ], + "type": "text", + "content": "[11] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 480, + 287, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 480, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 48, + 480, + 287, + 502 + ], + "type": "text", + "content": "[12] Prafulla Dhariwal and Alexander Quinn Nichol. Diffusion models beat gans on image synthesis. In NeurIPS, 2021. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 502, + 287, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 502, + 287, + 534 + ], + "spans": [ + { + "bbox": [ + 48, + 502, + 287, + 534 + ], + "type": "text", + "content": "[13] Finale Doshi-Velez and Been Kim. Towards a rigorous science of interpretable machine learning. arXiv preprint arXiv:1702.08608, 2017. 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 536, + 287, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 287, + 568 + ], + "type": "text", + "content": "[14] Dave Epstein, Taesung Park, Richard Zhang, Eli Shechtman, and Alexei A. Efros. Blobgan: Spatially disentangled scene representations. In ECCV, 2022. 1, 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 570, + 287, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 570, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 570, + 287, + 612 + ], + "type": "text", + "content": "[15] Thomas Fel, Julien Colin, Rémi Cadène, and Thomas Serre. What I cannot predict, I do not understand: A human-centered evaluation framework for explainability methods. In NeurIPS, 2022. 2, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 614, + 287, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 287, + 645 + ], + "type": "text", + "content": "[16] Ruth C. Fong and Andrea Vedaldi. Interpretable explanations of black boxes by meaningful perturbation. In ICCV, 2017. 2, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 647, + 287, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 680 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 680 + ], + "type": "text", + "content": "[17] Timo Freiesleben. Counterfactual explanations & adversarial examples - common grounds, essential differences, and potential transfers. CoRR, abs/2009.05487, 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "text", + "content": "[18] Nicholas Frosst and Geoffrey E. Hinton. Distilling a neural network into a soft decision tree. In Workshop on Comprehensibility and Explanation in AI and ML @AI*IA, 2017. 2" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 38, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[19] Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. In ICLR, 2015. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 108, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 545, + 140 + ], + "type": "text", + "content": "[20] Yash Goyal, Ziyan Wu, Jan Ernst, Dhruv Batra, Devi Parikh, and Stefan Lee. Counterfactual visual explanations. In ICML, 2019. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 141, + 545, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 175 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 175 + ], + "type": "text", + "content": "[21] Michael Harradon, Jeff Druce, and Brian E. Ruttenberg. Causal learning and explanation of deep neural networks via autoencoded activations. CoRR, 2018. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 176, + 545, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 176, + 545, + 209 + ], + "spans": [ + { + "bbox": [ + 307, + 176, + 545, + 209 + ], + "type": "text", + "content": "[22] Peter Hase and Mohit Bansal. Evaluating explainable AI: which algorithmic explanations help users predict model behavior? In ACL, 2020. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 210, + 545, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 210, + 545, + 243 + ], + "spans": [ + { + "bbox": [ + 307, + 210, + 545, + 243 + ], + "type": "text", + "content": "[23] Lisa Anne Hendricks, Ronghang Hu, Trevor Darrell, and Zeynep Akata. Grounding visual explanations. In ECCV, 2018. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 245, + 545, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 245, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 307, + 245, + 545, + 289 + ], + "type": "text", + "content": "[24] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In NeurIPS, 2017. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 291, + 545, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 291, + 545, + 323 + ], + "spans": [ + { + "bbox": [ + 307, + 291, + 545, + 323 + ], + "type": "text", + "content": "[25] Lukas Hoyer, Maurizio Munoz, Prateek Katiyar, Anna Khoreva, and Volker Fischer. Grid saliency for context explanations of semantic segmentation. In NeurIPS, 2019. 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 325, + 545, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 325, + 545, + 357 + ], + "spans": [ + { + "bbox": [ + 307, + 325, + 545, + 357 + ], + "type": "text", + "content": "[26] Gao Huang, Zhuang Liu, Laurens van der Maaten, and Kilian Q. Weinberger. Densely connected convolutional networks. In CVPR, 2017. 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 358, + 545, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 358, + 545, + 402 + ], + "spans": [ + { + "bbox": [ + 307, + 358, + 545, + 402 + ], + "type": "text", + "content": "[27] Paul Jacob, Éloi Zablocki, Hedi Ben-Younes, Mickaël Chen, Patrick Pérez, and Matthieu Cord. STEEX: steering counterfactual explanations with semantics. In ECCV, 2022. 2, 3, 5, 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 404, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 404, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 307, + 404, + 545, + 437 + ], + "type": "text", + "content": "[28] Guillaume Jeanneret, Loic Simon, and Frédéric Jurie. Diffusion models for counterfactual explanations. CoRR, abs/2203.15636, 2022. 1, 2, 3, 8" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 439, + 545, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 439, + 545, + 461 + ], + "spans": [ + { + "bbox": [ + 307, + 439, + 545, + 461 + ], + "type": "text", + "content": "[29] Saeed Khorram and Fuxin Li. Cycle-consistent counterfactuals by latent transformations. In CVPR, 2022. 1, 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 463, + 545, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 463, + 545, + 495 + ], + "spans": [ + { + "bbox": [ + 307, + 463, + 545, + 495 + ], + "type": "text", + "content": "[30] Been Kim, Oluwasanmi Koyejo, and Rajiv Khanna. Examples are not enough, learn to criticize! criticism for interpretability. In NeurIPS, 2016. 7" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 497, + 545, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 497, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 307, + 497, + 545, + 529 + ], + "type": "text", + "content": "[31] Jinkyu Kim and John F. Canny. Interpretable learning for self-driving cars by visualizing causal attention. In ICCV, 2017. 2" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 531, + 545, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 531, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 307, + 531, + 545, + 586 + ], + "type": "text", + "content": "[32] Oran Lang, Yossi Gandelsman, Michal Yarom, Yoav Wald, Gal Elidan, Avinatan Hassidim, William T. Freeman, Phillip Isola, Amir Globerson, Michal Irani, and Inbar Mosseri. Explaining in style: Training a GAN to explain a classifier in stylespace. In ICCV, 2021. 2" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 307, + 588, + 545, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 588, + 545, + 610 + ], + "spans": [ + { + "bbox": [ + 307, + 588, + 545, + 610 + ], + "type": "text", + "content": "[33] Zhiheng Li and Chenliang Xu. Discover the Unknown Biased Attribute of an Image Classifier. In ICCV, 2021. 2" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 307, + 612, + 545, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 612, + 545, + 643 + ], + "spans": [ + { + "bbox": [ + 307, + 612, + 545, + 643 + ], + "type": "text", + "content": "[34] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In ICCV, 2015. 2" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 307, + 646, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 646, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 646, + 545, + 678 + ], + "type": "text", + "content": "[35] Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, and Pascal Frossard. Deepfool: A simple and accurate method to fool deep neural networks. In CVPR, 2016. 2" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "text", + "content": "[36] Giang Nguyen, Daeyoung Kim, and Anh Nguyen. The effectiveness of feature attribution methods and its correlation with automatic evaluation scores. In NeurIPS, 2021. 7" + } + ] + } + ], + "index": 37 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15070" + } + ] + } + ], + "index": 39 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 675 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "type": "text", + "content": "[37] Michael Niemeyer and Andreas Geiger. GIRAFFE: representing scenes as compositional generative neural feature fields. In CVPR, 2021. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 287, + 150 + ], + "type": "text", + "content": "[38] Martin Pawelczyk, Shalmali Joshi, Chirag Agarwal, Sohini Upadhyay, and Himabindu Lakkaraju. On the connections between counterfactual explanations and adversarial examples. CoRR, abs/2106.09992, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 151, + 287, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 151, + 287, + 184 + ], + "spans": [ + { + "bbox": [ + 48, + 151, + 287, + 184 + ], + "type": "text", + "content": "[39] Sylvestre-Alvise Rebuffi, Ruth Fong, Xu Ji, and Andrea Vedaldi. There and back again: Revisiting backpropagation saliency methods. In CVPR, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 185, + 287, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 185, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 185, + 287, + 217 + ], + "type": "text", + "content": "[40] Marco Túlio Ribeiro, Sameer Singh, and Carlos Guestrin. \"why should I trust you?\": Explaining the predictions of any classifier. In SIGKDD, 2016. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 218, + 287, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 218, + 287, + 261 + ], + "spans": [ + { + "bbox": [ + 48, + 218, + 287, + 261 + ], + "type": "text", + "content": "[41] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: A stylegan encoder for image-to-image translation. In CVPR, 2021. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 262, + 287, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 262, + 287, + 306 + ], + "spans": [ + { + "bbox": [ + 48, + 262, + 287, + 306 + ], + "type": "text", + "content": "[42] Pau Rodríguez, Massimo Caccia, Alexandre Lacoste, Lee Zamparo, Issam H. Laradji, Laurent Charlin, and David Vázquez. Beyond trivial counterfactual explanations with diverse valuable explanations. In ICCV, 2021. 1, 2, 3, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 307, + 287, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 307, + 287, + 339 + ], + "spans": [ + { + "bbox": [ + 48, + 307, + 287, + 339 + ], + "type": "text", + "content": "[43] Axel Sauer, Nikolay Savinov, and Andreas Geiger. Conditional affordance learning for driving in urban environments. In CoRL, 2018. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 341, + 287, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 341, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 48, + 341, + 287, + 384 + ], + "type": "text", + "content": "[44] Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In ICCV, 2017. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 386, + 287, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 386, + 287, + 429 + ], + "spans": [ + { + "bbox": [ + 48, + 386, + 287, + 429 + ], + "type": "text", + "content": "[45] Yuan Shen, Shanduojiao Jiang, Yanlin Chen, Eileen Yang, Xilun Jin, Yuliang Fan, and Katie Driggs Campbell. To explain or not to explain: A study on the necessity of explanations for autonomous vehicles. CoRR, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 430, + 287, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 430, + 287, + 462 + ], + "spans": [ + { + "bbox": [ + 48, + 430, + 287, + 462 + ], + "type": "text", + "content": "[46] Sumedha Singla, Brian Pollack, Junxiang Chen, and Kayhan Batmanghelich. Explanation by progressive exaggeration. In ICLR, 2020. 1, 2, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 463, + 287, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 463, + 287, + 485 + ], + "spans": [ + { + "bbox": [ + 48, + 463, + 287, + 485 + ], + "type": "text", + "content": "[47] Mukund Sundararajan, Ankur Taly, and Qiqi Yan. Axiomatic attribution for deep networks. In ICML, 2017. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 486, + 287, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 486, + 287, + 518 + ], + "spans": [ + { + "bbox": [ + 48, + 486, + 287, + 518 + ], + "type": "text", + "content": "[48] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian J. Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In *ICLR*, 2014. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 520, + 287, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 520, + 287, + 552 + ], + "spans": [ + { + "bbox": [ + 48, + 520, + 287, + 552 + ], + "type": "text", + "content": "[49] Erico Tjoa and Cuntai Guan. A survey on explainable artificial intelligence (XAI): towards medical XAI. CoRR, abs/1907.07374, 2019. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 553, + 287, + 596 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 553, + 287, + 596 + ], + "spans": [ + { + "bbox": [ + 48, + 553, + 287, + 596 + ], + "type": "text", + "content": "[50] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for stylegan image manipulation. ACM Transactions on Graphics (TOG), 2021. 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 597, + 287, + 641 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 597, + 287, + 641 + ], + "spans": [ + { + "bbox": [ + 48, + 597, + 287, + 641 + ], + "type": "text", + "content": "[51] Simon Vandenhende, Dhruv Kumar Mahajan, Filip Radenovic, and Deepti Ghadiyaram. Making heads or tails: Towards semantically consistent visual counterfactuals. In ECCV, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 642, + 287, + 675 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 642, + 287, + 675 + ], + "spans": [ + { + "bbox": [ + 48, + 642, + 287, + 675 + ], + "type": "text", + "content": "[52] Sahil Verma, John P. Dickerson, and Keegan Hines. Counterfactual explanations for machine learning: A review. CoRR, abs/2010.10596, 2020. 1" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 545, + 675 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 307, + 72, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 545, + 116 + ], + "type": "text", + "content": "[53] Tom Vermeire, Dieter Brughmans, Sofie Goethals, Raphael Mazzine Barbossa de Oliveira, and David Martens. Explainable image classification with evidence counterfactual. Pattern Anal. Appl., 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "type": "text", + "content": "[54] Kira Vinogradova, Alexandr Dibrov, and Gene Myers. Towards interpretable semantic segmentation via gradient-weighted class activation mapping (student abstract). In AAAI, 2020. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 163, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 163, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 308, + 163, + 545, + 206 + ], + "type": "text", + "content": "[55] Sandra Wachter, Brent Mittelstadt, and Chris Russell. Counterfactual explanations without opening the black box: Automated decisions and the gdpr. Harvard Journal of Law & Technology, 2017. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 208, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 208, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 308, + 208, + 545, + 251 + ], + "type": "text", + "content": "[56] Jörg Wagner, Jan Mathias Kohler, Tobias Gindele, Leon Hetzel, Jakob Thaddäus Wiedemer, and Sven Behnke. Interpretable and fine-grained visual explanations for convolutional neural networks. In CVPR, 2019. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 254, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 254, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 308, + 254, + 545, + 285 + ], + "type": "text", + "content": "[57] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 288, + 545, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 288, + 545, + 310 + ], + "spans": [ + { + "bbox": [ + 308, + 288, + 545, + 310 + ], + "type": "text", + "content": "[58] Pei Wang and Nuno Vasconcelos. SCOUT: self-aware discriminant counterfactual explanations. In CVPR, 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 312, + 545, + 355 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 312, + 545, + 355 + ], + "spans": [ + { + "bbox": [ + 308, + 312, + 545, + 355 + ], + "type": "text", + "content": "[59] Yiran Xu, Xiaoyin Yang, Lihang Gong, Hsuan-Chu Lin, TzYing Wu, Yunsheng Li, and Nuno Vasconcelos. Explainable object-induced action decision for autonomous vehicles. In CVPR, 2020. 2, 5, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 357, + 545, + 400 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 357, + 545, + 400 + ], + "spans": [ + { + "bbox": [ + 308, + 357, + 545, + 400 + ], + "type": "text", + "content": "[60] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darryll. BDD100K: A diverse driving dataset for heterogeneous multitask learning. In CVPR, 2020. 2, 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 402, + 545, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 402, + 545, + 445 + ], + "spans": [ + { + "bbox": [ + 308, + 402, + 545, + 445 + ], + "type": "text", + "content": "[61] Éloi Zablocki, Hédi Ben-Younes, Patrick Pérez, and Matthieu Cord. Explainability of deep vision-based autonomous driving systems: Review and challenges. Int. J. Comput. Vis., 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 447, + 545, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 447, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 308, + 447, + 545, + 469 + ], + "type": "text", + "content": "[62] Matthew D. Zeiler and Rob Fergus. Visualizing and understanding convolutional networks. In ECCV, 2014. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 471, + 545, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 471, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 308, + 471, + 545, + 502 + ], + "type": "text", + "content": "[63] Quanshi Zhang, Ying Nian Wu, and Song-Chun Zhu. Interpretable convolutional neural networks. In CVPR, 2018. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 505, + 545, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 505, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 308, + 505, + 545, + 536 + ], + "type": "text", + "content": "[64] Qiaoning Zhang, X. Jessie Yang, and Lionel Peter Robert. Expectations and trust in automated vehicles. In CHI, 2020. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 539, + 545, + 572 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 539, + 545, + 572 + ], + "spans": [ + { + "bbox": [ + 308, + 539, + 545, + 572 + ], + "type": "text", + "content": "[65] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 5" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 574, + 545, + 605 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 574, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 308, + 574, + 545, + 605 + ], + "type": "text", + "content": "[66] Bolei Zhou, Aditya Khosla, Ågata Lapedriza, Aude Oliva, and Antonio Torralba. Object detectors emerge in deep scene cnns. In ICLR, 2015. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 608, + 545, + 639 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 608, + 545, + 639 + ], + "spans": [ + { + "bbox": [ + 308, + 608, + 545, + 639 + ], + "type": "text", + "content": "[67] Jiapeng Zhu, Yujun Shen, Deli Zhao, and Bolei Zhou. Indomain gan inversion for real image editing. In ECCV, 2020. 4" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 642, + 545, + 675 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 642, + 545, + 675 + ], + "spans": [ + { + "bbox": [ + 308, + 642, + 545, + 675 + ], + "type": "text", + "content": "[68] Peihao Zhu, Rameen Abdal, Yipeng Qin, and Peter Wonka. SEAN: image synthesis with semantic region-adaptive normalization. In CVPR, 2020. 3" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15071" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/OPE-SR_ Orthogonal Position Encoding for Designing a Parameter-Free Upsampling Module in Arbitrary-Scale Image Super-Resolution/538d0672-86b3-4cf7-ab0d-b9344b19df37_content_list.json b/2023/OPE-SR_ Orthogonal Position Encoding for Designing a Parameter-Free Upsampling Module in Arbitrary-Scale Image Super-Resolution/538d0672-86b3-4cf7-ab0d-b9344b19df37_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6ef8601bca31c11ee8a6025184f7890fe19396d7 --- /dev/null +++ b/2023/OPE-SR_ Orthogonal Position Encoding for Designing a Parameter-Free Upsampling Module in Arbitrary-Scale Image Super-Resolution/538d0672-86b3-4cf7-ab0d-b9344b19df37_content_list.json @@ -0,0 +1,1743 @@ +[ + { + "type": "text", + "text": "OPE-SR: Orthogonal Position Encoding for Designing a Parameter-free Upsampling Module in Arbitrary-scale Image Super-Resolution", + "text_level": 1, + "bbox": [ + 124, + 128, + 848, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Gaochao Song $^{1}$ Qian Sun $^{2*}$ Luo Zhang $^{3}$ Ran Su $^{1}$ Jianfeng Shi $^{2}$ Ying He $^{3}$", + "bbox": [ + 150, + 202, + 818, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Tianjin University $^{2}$ Nanjing University of Information Science and Technology $^{3}$ Nanyang Technological University", + "bbox": [ + 158, + 220, + 810, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "gaochaosong_21@tju.edu.cn sunqian@nuist.edu.cn", + "bbox": [ + 271, + 258, + 691, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 308, + 313, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Arbitrary-scale image super-resolution (SR) is often tackled using the implicit neural representation (INR) approach, which relies on a position encoding scheme to improve its representation ability. In this paper, we introduce orthogonal position encoding (OPE), an extension of position encoding, and an OPE-Upscale module to replace the INR-based upsampling module for arbitrary-scale image super-resolution. Our OPE-Upscale module takes 2D coordinates and latent code as inputs, just like INR, but does not require any training parameters. This parameter-free feature allows the OPE-Upscale module to directly perform linear combination operations, resulting in continuous image reconstruction and achieving arbitrary-scale image reconstruction. As a concise SR framework, our method is computationally efficient and consumes less memory than state-of-the-art methods, as confirmed by extensive experiments and evaluations. In addition, our method achieves comparable results with state-of-the-art methods in arbitrary-scale image super-resolution. Lastly, we show that OPE corresponds to a set of orthogonal basis, validating our design principle.", + "bbox": [ + 75, + 340, + 472, + 657 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 686, + 209, + 704 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Photographs are composed of discrete pixels of varying precision due to the limitations of sampling frequency, which breaks the continuous visual world into discrete parts. The single image super-resolution (SISR) task aims to restore the original continuous world in the image as much as possible. In an arbitrary-scale SR task, one often reconstructs the continuous representation of a low-resolution image and then adjusts the resolution of the target image as needed. The recent rise of implicit neural representation (INR) in 3D vision has enabled the representation", + "bbox": [ + 75, + 713, + 468, + 864 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "of complex 3D objects and scenes in a continuous manner [14, 19, 41, 42, 44, 45, 47, 49, 57, 58], which also opens up possibilities for continuous image and arbitrary-scale image super-resolution [5, 18, 32, 72].", + "bbox": [ + 500, + 310, + 890, + 369 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing methods for arbitrary-scale SR typically use a post-upsampling framework [70]. In this approach, low-resolution (LR) images first pass through a deep CNN network (encoder) without improving the resolution, and then pass through an INR-based upsampling module (decoder) with a specified target resolution to reconstruct high-resolution (HR) images. The decoder establishes a mapping from feature maps (the output of encoder) to target image pixels using a pre-assigned grid partitioning and achieves arbitrary-scale with the density of the grid in Cartesian coordinate system. However, the INR approach has a defect of learning low-frequency information, also known as spectral bias [50]. To address this issue, sinusoidal positional encoding is introduced to embed input coordinates to higher dimensions and enable the network to learn high-frequency details. This inspired recent works on arbitrary-scale SR to further improve the representation ability [32, 72].", + "bbox": [ + 496, + 371, + 892, + 627 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite its effectiveness in arbitrary-scale SR, the INR-based upsampling module increases the complexity of the entire SR framework as two different networks are jointly trained. Additionally, as a black-box model, it represents a continuous image with a strong dependency on both the feature map and the decoder (e.g., MLP). However, its representation ability decreases after flipping the feature map, a phenomenon known as flipping consistency decline. As shown in Fig. 1, flipping the feature map horizontally before the upsampling module of LIIF results in a blurred target image that does not have the expected flip transformation. This decline could be due to limitations of the MLP in learning the symmetry feature of the image.", + "bbox": [ + 496, + 628, + 893, + 824 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "MLP is a universal function approximator [17], which tries to fit a mapping function from feature map to the continuous image, therefore, it is reasonable to assume that such process could be solved by an analytical solution. In this paper, we re-examine position encoding from the per", + "bbox": [ + 496, + 824, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 94, + 875, + 222, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1Project page: https://github.com/gaochao-s/ope-sr", + "bbox": [ + 96, + 887, + 452, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "10009", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "spective of orthogonal basis and propose orthogonal position encoding (OPE) for continuous image representation. The linear combination of 1D latent code and OPE can directly reconstruct continuous image patch without using implicit neural function [5]. To prove OPE's rationality, we analyse it both from functional analysis and 2D-Fourier transform. We further embed it into a parameter-free upsampling module, called OPE-Upscale Module, to replace INR-based upsampling module in deep SR framework, then currently deep SR framework can be greatly simplified.", + "bbox": [ + 75, + 90, + 472, + 241 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Unlike the state-of-the-art method by Lee et al. [32], which enhances MLP with position encoding, we explore the possibility of extending position encoding without MLP. By providing a more concise SR framework, our method achieves high computing efficiency and consumes less memory than the state-of-the-art, while also achieving comparable image performance in arbitrary-scale SR tasks.", + "bbox": [ + 75, + 242, + 472, + 348 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are as follows:", + "bbox": [ + 96, + 348, + 315, + 361 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel position encoding, called orthogonal position encoding (OPE), which takes the form of a 2D-Fourier series and corresponds to a set of orthogonal basis. Building on OPE, we introduce the OPE-Upscale Module, a parameter-free upsampling module for arbitrary-scale image super-resolution.", + "- Our method significantly reduces the consumption of computing resources, resulting in high computing efficiency for arbitrary-scale SR tasks.", + "- The OPE-Upscale Module is interpretable, parameter-free and does not require training, resulting in a concise SR framework that elegantly solves the flipping consistency problem.", + "- Extensive experiments demonstrate that our method achieves comparable results with the state-of-the-art. Furthermore, our method enables super-resolution up to a large scale of $\\times 30$ ." + ], + "bbox": [ + 94, + 369, + 470, + 652 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 671, + 218, + 686 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Sinusoidal Positional Encoding", + "text_level": 1, + "bbox": [ + 76, + 696, + 352, + 713 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Sinusoidal positional encoding is widely used to counteract the negative effects of token order and sequence length in sequence models [65], or to guide image generation as the spatial inductive bias in CNNs [7, 23, 36]. In implicit neural representations, it plays a critical role in solving spectral bias [50]. By embedding input coordinates into a higher dimensional space, position encoding greatly improves high-frequency representation of implicit 3D scenes [42, 56] and subsequent works take it as the default operation to improve representation quality [37, 43, 53, 78]. Inspired by these works, positional encoding has been preliminarily explored in representing images in arbitrary-scale SR [32, 72].", + "bbox": [ + 75, + 719, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/431e70fe5e15799c2fd2cfcf81a85fecf1bfe8720acc6cede96f95242475aa4f.jpg", + "image_caption": [ + "LIIF:32.35", + "LIIF-flip: 28.47" + ], + "image_footnote": [], + "bbox": [ + 514, + 88, + 694, + 157 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2c5698403b7ce46cbf5dbe6eb0580985fcc06df8f1c84856cb42c264a35e73da.jpg", + "image_caption": [ + "LIIF:25.43", + "LIIF-flip: 23.11" + ], + "image_footnote": [], + "bbox": [ + 702, + 88, + 875, + 159 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/9a15d2c0a4f1a1b7a11386054ae16ec18bd0c0b1de4f758e9301086228311523.jpg", + "image_caption": [ + "OPE: 32.37", + "OPE-flip: 32.37", + "Figure 1. Flipping consistency decline (PSNR (dB)). LIIF-flip: Flipping the input of the LIIF [5] decoder yields blurred results in the symmetric outputs. OPE-flip: Our method does not exhibit such artifacts. Additional results can be found in the supplementary material." + ], + "image_footnote": [], + "bbox": [ + 514, + 170, + 694, + 239 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/5386d6e74192c053b22db82f332fd7d30e29ed161cc485d8585e5143e92c53d1.jpg", + "image_caption": [ + "OPE:25.50", + "OPE-flip: 25.50" + ], + "image_footnote": [], + "bbox": [ + 702, + 170, + 875, + 239 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Orthogonal Basis Decomposition", + "text_level": 1, + "bbox": [ + 500, + 354, + 787, + 369 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In functional analysis, an orthogonal basis decomposition is a way to represent a vector or function as a linear combination of orthogonal basis functions. Wavelet transform [2, 15, 20, 38] and 2D-Fourier transform [6, 12, 13, 22, 74] are commonly used decomposition techniques for images and videos. In DSGAN [12], the input image is explicitly decomposed into low and high frequencies using high-pass and low-pass filters. Other methods use frequency domain losses to decompose images, either in a supervised manner [22] or in an unsupervised manner [13]. To address with resolution discrepancy of reconstructed images and input images, Rippel et al. [52] employ spectral pooling to decrease resolution by truncating in the Fourier domain, while Zhou et al. [81] explore an up-sampling method in the Fourier domain to increase the resolution. Image moments, which decompose images into two-dimensional orthogonal polynomials [27, 82], are widely used in invariant pattern recognition [30, 77]. Image sparse representation inherits this decomposition idea and performs well in traditional computer vision tasks [39, 71, 73]. In 3D domains, spherical harmonics serve as an orthogonal basis in space to represent view dependence [3, 51, 59] and have recently been proposed as a replacement for MLPs [11] for representing neural radiance fields [42].", + "bbox": [ + 496, + 378, + 893, + 741 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.3. Deep Learning-based SR", + "text_level": 1, + "bbox": [ + 500, + 753, + 728, + 771 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Based on the upsampling operations and their location in the model, deep learning-based SR frameworks can be classified into four categories (see [70] for a comprehensive survey): pre-upsampling [8,24,25,55,61,62], post-upsampling [9,31,35,64,79], progressive-upsampling [28,29,68], and iterative up-and-down sampling [16, 33, 69]. With preupsampling, the LR image is first upsampled by traditional interpolation and then fed into a deep CNN to reconstruct", + "bbox": [ + 496, + 779, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "10010", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "high-quality details. While it was one of the most popular frameworks for arbitrary-scale factors, it has side effects like enlarged noise by interpolation and high time complexity and space consumption. The progressive-upsampling and iterative up-and-down sampling frameworks pose challenges in terms of complicated model designing and unclear design criteria, as noted in [70]. For post-upsampling, the LR image is directly fed as input to a deep CNN, and then a trainable upsampling module (e.g., deconvolution [9], subpixel [54], and interpolation convolution [10]) increases the resolution at the end. Since feature extraction process, which is computationally intensive, only occurs in low-dimensional space, it has become one of the mainstream frameworks [32, 34, 67].", + "bbox": [ + 75, + 90, + 472, + 303 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.4. Arbitrary-scale SR", + "text_level": 1, + "bbox": [ + 76, + 310, + 261, + 325 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the field of arbitrary-scale SR, most existing works are based on the post-upsampling framework and replace the traditional upsampling module with an INR-based one, such as a coordinate-based MLP. Meta-SR [18] was the first arbitrary-scale SR method based on CNN. ArbSR [66] adopts a general plug-in module to solve the scaling problem of different horizontal and vertical scales. SRWarp [60] transforms LR images into HR images with arbitrary shapes via a differential adaptive warping layer. SphereSR [75] explores arbitrary-scale on $360^{\\circ}$ images. LIIF [5] uses coordinates and conditional latent code into an MLP to directly predict target pixel color with an intuitive network structure. LIIF-related follow-up works focus on predicting high-frequency information with position encoding [32,72].", + "bbox": [ + 75, + 333, + 472, + 547 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 556, + 168, + 571 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. OPE-based Image Representation", + "text_level": 1, + "bbox": [ + 76, + 580, + 374, + 598 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given an LR image $I_{LR}$ with resolution of $h \\times w$ , we decide its 2D domain into $h \\times w$ grids, where each grid represents a pixel in the LR image and corresponds to a patch of the high resolution image of size $r_h \\times r_w$ . The output image $I_{SR}$ has a resolution of $H \\times W$ , where $H = r_h \\cdot h$ and $W = r_w \\cdot w$ . Denote by $I \\in \\mathbb{R}^{r_h \\times r_w \\times 1}$ the high resolution image patch with size $r_h \\times r_w$ for a specific color channel. We view each pixel in $I$ as a sample of a continuous bivariate function $f(x,y): [-1,1] \\times [-1,1] \\to \\mathbb{R}$ . We embed the coordinates $x$ and $y$ using sinusoidal positional encoding", + "bbox": [ + 75, + 604, + 468, + 758 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} { X } { = \\gamma ( x ) = \\sqrt { 2 } \\cdot [ \\frac { 1 } { \\sqrt { 2 } } , \\cos ( \\pi x ) , \\sin ( \\pi x ) , \\cos ( 2 \\pi x ) , } \\\\ \\sin (2 \\pi x), \\dots , \\cos (n \\pi x), \\sin (n \\pi x) ] \\tag {1} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 762, + 468, + 811 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} { Y } { = \\gamma ( y ) = \\sqrt { 2 } \\cdot \\left[ \\frac { 1 } { \\sqrt { 2 } } , \\cos ( \\pi y ) , \\sin ( \\pi y ) , \\cos ( 2 \\pi y ) , \\right. } \\\\ \\left. \\sin (2 \\pi y), \\dots , \\cos (n \\pi y), \\sin (n \\pi y) \\right], \\tag {2} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 98, + 813, + 468, + 864 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\gamma (\\cdot):\\mathbb{R}\\to \\mathbb{R}^{1\\times (2n + 1)}$ is a univariate function for position encoding with a predefined maximum frequency", + "bbox": [ + 76, + 869, + 468, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$n\\in \\mathbb{N}$ . We flatten the matrix $X^T Y$ into a row vector $P\\in$ $\\mathbb{R}^{1\\times (2n + 1)^2}$ as", + "bbox": [ + 498, + 90, + 890, + 119 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nP = \\operatorname {f l a t} \\left(X ^ {T} Y\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 637, + 119, + 890, + 137 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where flat: $\\mathbb{R}^{(2n + 1)\\times (2n + 1)}\\to \\mathbb{R}^{1\\times (2n + 1)^2}$ is the flattening operation. Denote by $e_{i,j}$ the element on the $i$ -th row and the $j$ -th column of matrix $X^T Y$ . For example. $e_{4,5} = 2\\cos (2\\pi x)\\sin (2\\pi y)$ . It is easy to verify that", + "bbox": [ + 496, + 141, + 893, + 205 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\langle e _ {i _ {1}, j _ {1}}, e _ {i _ {2}, j _ {2}} \\right\\rangle = \\left\\{ \\begin{array}{l l} 0, & (i _ {1}, j _ {1}) \\neq (i _ {2}, j _ {2}) \\\\ 1, & (i _ {1}, j _ {1}) = (i _ {2}, j _ {2}) \\end{array} \\right. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 214, + 890, + 253 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\langle ,\\rangle$ is the inner product in function space, i.e.,", + "bbox": [ + 500, + 262, + 846, + 279 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\langle g, h \\rangle = \\frac {1}{4} \\int_ {- 1} ^ {1} \\int_ {- 1} ^ {1} g (x, y) h (x, y) \\mathrm {d} x \\mathrm {d} y.\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 286, + 833, + 321 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Therefore the elements $\\{e_{i,j}\\}$ form a set of orthogonal basis, allowing us to approximate $f$ as a linear combination. Consider a pixel $(x,y)$ in the upsampled image patch $I$ , which corresponds to a region $[x - 1 / r_h, x + 1 / r_h] \\times [y - 1 / r_w, y + 1 / r_w]$ . We use $f(x,y)$ as a representative for the entire region and compute $I_{(x,y)}$ as", + "bbox": [ + 498, + 330, + 890, + 422 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nI _ {(x, y)} \\triangleq f (x, y) \\approx Z P ^ {T}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 609, + 430, + 890, + 450 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $Z \\in \\mathbb{R}^{1 \\times (2n + 1)^2}$ represents the projection. Due to the orthogonal property, we call the resulting vector $P$ as orthogonal position encoding (OPE). Fig. 3 illustrates the concept of OPE-based patch representation.", + "bbox": [ + 496, + 459, + 890, + 521 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Remark 1. Our OPE basis can be seen as the real form version of the 2D-Fourier basis, which eliminates the complex exponential term based on conjugate symmetry when representing real signals. See the supplementary material for details.", + "bbox": [ + 496, + 521, + 890, + 595 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Remark 2. OPE differs from the commonly used positional encoding formulation, such as that in [42], in that it includes a constant term and takes the product of each coordinate embedding as a new term. These seemingly minor changes indeed have a deep impact. Conventional positional encoding processes the coordinates separately, thereby encoding frequencies in the horizontal and vertical directions only. In contrast, OPE includes frequencies covering in all directions of the plane due to the product terms, resulting in a better expression capability.", + "bbox": [ + 496, + 597, + 890, + 747 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Remark 3. Unlike LIIF [5] that uses an MLP to approximate the image function $f$ , our method completely eliminates the need of MLP.", + "bbox": [ + 496, + 748, + 890, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. OPE-Upscale Module", + "text_level": 1, + "bbox": [ + 500, + 801, + 702, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We project the latent code onto the OPE basis. OPE with a sufficiently long latent code could represent an image directly in a continuous manner. However, it suffers from long embedding time and is unstable for representing high-frequency details locally, similar to the limitations", + "bbox": [ + 496, + 824, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "10011", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/315202109cbd5ab2b2ae09d337dd9a1ba4711b5da23cf976bd21202fadfd71de.jpg", + "image_caption": [ + "Figure 2. OPE-Upscale Module for arbitrary-scale SR. The encoder $E_{\\theta}$ is the only trainable part. With a pre-defined maximum frequency $n$ of OPE, the OPE-Upscale Module (shaded in grey) takes the feature map from $E_{\\theta}$ and the target resolution $H$ , $W$ as input, and renders the pixels of the target SR image in parallel. $\\odot$ is the matmul function that returns the product of $z^{*} \\in \\mathbb{R}^{1 \\times 3(2n + 1)^{2}}$ and OPEs $\\in \\mathbb{R}^{3(2n + 1)^{2} \\times 1}$ per color channel." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 880, + 244 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d9549fbd46fbe99071b6b32e939af388d83fef11d090526cca8a42697b81241b.jpg", + "image_caption": [ + "Figure 3. OPE-based patch representation. An image patch can be represented as a linear combination of basic plane waves in a continuous manner. The representation is a 2D extension to [46, Chap. 3.3.1]. Refer to the supplementary material for more details." + ], + "image_footnote": [], + "bbox": [ + 89, + 335, + 455, + 478 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "of Fourier transform to describe local information. To address this issue, we represent the input image as the seamless stitching of local patches, whose latent codes are extracted from a feature map over the channel dimension2. As shown in Fig. 2, the OPE-upscale module takes both the target resolution $H = r_h \\cdot h$ , $W = r_w \\cdot w$ and the feature map $\\in \\mathbb{R}^{h \\times w \\times 3(2n + 1)^2}$ from the deep encoder $E_{\\theta}$ as inputs and computes target pixels in parallel.", + "bbox": [ + 75, + 563, + 468, + 686 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Feature map rendering. As shown in Fig. 4, to render a target image $I_{SR}$ with size $H \\times W$ from a LR image $I_{LR}$ with size $h \\times w$ , OPE-Upscale Module firstly divide a 2D domain $[-1,1] \\times [-1,1]$ into $H \\times W$ regions with equal size, so that every pixel in $I_{SR}$ will be associated with an absolute central coordinates $(x_q, y_q)$ in corresponding region. Secondly, the latent codes in the feature map (same dimension with $I_{LR}$ ) also possess corresponding central coordinates $(x_c, y_c) \\in [-1,1] \\times [-1,1]$ by dividing same 2D domain into $h \\times w$ regions, therefore, given a target image pixel with $(x_q, y_q)$ , a specific latent code $z^* \\in \\mathbb{R}^{1 \\times 3(2n + 1)^2}$ with", + "bbox": [ + 75, + 686, + 470, + 854 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6f0579f2f542a32044fcfddc79d9a66ba99c8e42ec8c5e9feb5df76dd793ddb2.jpg", + "image_caption": [ + "Figure 4. Illustration of the mapping from the input LR image to the output SR image. LR image, feature map and SR image are all divided into the same domain $[-1,1] \\times [-1,1]$ . Each LR pixel $I_{LR}(x_{LR},y_{LR})$ corresponds to a latent code with coordinates $(x_c,y_c)$ , while each SR pixel $I_{SR}(x_{SR},y_{SR})$ corresponds to a latent code with coordinates $(x_q,y_q)$ ." + ], + "image_footnote": [], + "bbox": [ + 521, + 335, + 874, + 470 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "coordinates $(x_{c},y_{c})$ , which has the smallest distance from $(x_{q},y_{q})$ could be found. As shown in Eq.(6) and Eq.(7), a render function $\\mathcal{R}$ takes two parts of inputs: $z^{*}$ and $(x_q',y_q')$ , to calculate final target pixel value as following:", + "bbox": [ + 496, + 587, + 893, + 648 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nI _ {S R} \\left(x _ {q}, y _ {q}\\right) = \\mathcal {R} \\left(z ^ {*}, \\left(x _ {q} ^ {\\prime}, y _ {q} ^ {\\prime}\\right)\\right) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 648, + 890, + 669 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nx _ {q} ^ {\\prime} = \\left(x _ {q} - x _ {c}\\right) \\cdot h, \\quad y _ {q} ^ {\\prime} = \\left(y _ {q} - y _ {c}\\right) \\cdot w \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 681, + 890, + 700 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $z^{*}$ is the nearest latent code we found, and $(x_{q}^{\\prime}, y_{q}^{\\prime})$ are relative coordinates, which are calculated based on Eq.(7) to rescale the absolute coordinates (in domain $[-1, 1] \\times [-1, 1]$ ) by times $h$ and $w$ , which is taken as input by function $\\mathcal{R}$ to render target pixel. $\\mathcal{R}$ has the similar calculation as Eq.(5) while the difference is it repeats OPE three times to adapt $z^{*}$ and calculate the linear combination per color channel. In this way, our OPE-upscale module successfully deals with arbitrary size $I_{SR}$ by processing each pixel by $\\mathcal{R}$ , in which feature map rendering process is parameter-free with high computing efficiency and less memory consumption (which has been confirmed by the experiments in Sec. 4.3).", + "bbox": [ + 496, + 704, + 892, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "2To ensure compatibility with color images, we adjust the output channel of the encoder to $3(2n + 1)^{2}$ , where $n$ is the pre-defined maximum frequency $n$ of OPE.", + "bbox": [ + 75, + 862, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "10012", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Patch ensemble. As shown in Fig. 4, when moving $(x_{q},y_{q})$ from location $a$ to $b$ , the pixel value in the target image $I_{SR}$ may change abruptly, as well as the nearest latent code $z^{*}$ . To address this discontinuity issue, we propose a patch ensemble, which is a local ensemble styled interpolation using relative coordinates. Instead of finding a single nearest latent code for $(x_{q},y_{q})$ , we select the nearest four neighbouring latent codes $z_{t}^{*}$ with corresponding central coordinates $(x_{t},y_{t})$ , where $t\\in \\{00,01,10,11\\}$ . Then we calculate the relative coordinates $x_{q}^{\\prime}$ and $y_{q}^{\\prime}$ as", + "bbox": [ + 73, + 90, + 472, + 244 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nx _ {q} ^ {\\prime} = \\frac {\\left(x _ {q} - x _ {t}\\right) \\cdot h}{2}, \\quad y _ {q} ^ {\\prime} = \\frac {\\left(y _ {q} - y _ {t}\\right) \\cdot w}{2}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 127, + 248, + 468, + 279 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "which can guarantee $x_{q}^{\\prime}, y_{q}^{\\prime} \\in [-1, 1]$ . As the inset shows, for the four adjacent pixels from the low resolution image $I_{LR}$ (i.e. related to latent codes $z_{00}^{*}, z_{01}^{*}, z_{10}^{*}$ and $z_{11}^{*}$ ),", + "bbox": [ + 76, + 281, + 468, + 327 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "their corresponding patches in super resolution image $I_{SR}$ are colored in red, green, yellow and blue, respectively. The pixel color in $I_{SR}$ is not solely dependent on the nearest latent code but considers the four neighboring latent codes. Specifically, using the rendering function $\\mathcal{R}$ and the di", + "bbox": [ + 76, + 327, + 300, + 462 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ed462df17db4a6c8995f53e1017ef18600251ebc53569198c6eeee7aa186f906.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 305, + 329, + 467, + 458 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "aagonal rectangle areas $s_t$ as weights, we compute the pixel value as a weighted sum", + "bbox": [ + 76, + 463, + 468, + 492 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nI _ {S R} \\left(x _ {q}, y _ {q}\\right) = \\sum_ {t \\in \\{0 0, 0 1, 1 0, 1 1 \\}} \\frac {s _ {t}}{S} \\cdot \\mathcal {R} \\left(z _ {t} ^ {*}, \\left(x _ {q} ^ {\\prime}, y _ {q} ^ {\\prime}\\right)\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 496, + 468, + 532 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $S = \\sum_{t} s_{t}$ is the sum of areas. Considering the contribution of each latent code allows us to integrate the adjacent latent codes with different significance, thereby providing a seamless stitching of adjacent patches. We call Eq.(9) local ensemble styled interpolation since it takes a similar form of the local ensemble in LIIF [5].", + "bbox": [ + 76, + 536, + 468, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Maximum Frequency $n$", + "text_level": 1, + "bbox": [ + 76, + 633, + 294, + 651 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Selecting a proper maximum frequency $n$ plays an important role in designing and implementing the OPE-upsampling module since it directly determines the network architecture and also has effects on the performance of different SR scales. Given a high resolution image $I_{HR}$ with size $H \\times W$ , $n$ and $r$ , we aim to obtain a feature map with size $\\frac{H}{r} \\times \\frac{W}{r}$ , then we re-render the obtained feature map with the selected $n$ . By the comparison of the rendered $I_{SR}$ , we present the performance of $n \\in \\{1,2,\\dots,8\\}$ under different $r$ values (SR scale), as show in Tab. 1, and select the $n$ with the best performance (the details would be discussed Sec. 4.1). To be specific, we use Eq.(10) as the basic theory and use Eq.(11) to infer the feature map. First, similar to calculate the projection of a normal vector on orthogonal basis, we can calculate projections (or so-called latent code) $Z \\in \\mathbb{R}^{1 \\times (2n + 1)^2}$ of $f(x,y)$ in Eq.(5) as follows:", + "bbox": [ + 76, + 657, + 472, + 902 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nZ [ i ] = \\frac {1}{4} \\int_ {- 1} ^ {1} \\int_ {- 1} ^ {1} f (x, y) P [ i ] (x, y) \\mathrm {d} x \\mathrm {d} y \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 102, + 890, + 138 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $P[i](x,y)$ is a bivariate function taken from the $i$ -th position of OPE and $Z[i]$ is the corresponding projection. Based on Eq.(10) and taking both the discreteness of an image and the design of OPE-Upscale Module into consideration, we calculate the feature map of an image $I_{HR}$ with down-sampling scale $r$ as follows:", + "bbox": [ + 496, + 143, + 893, + 234 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nz ^ {*} [ i ] = \\frac {1}{4} \\sum_ {x ^ {\\prime}} ^ {2 r} \\sum_ {y ^ {\\prime}} ^ {2 r} I _ {H R} \\left(x ^ {\\prime}, y ^ {\\prime}\\right) P [ i ] \\left(x ^ {\\prime}, y ^ {\\prime}\\right) \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 244, + 890, + 287 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "It can be considered as the inverse operation of Eq.(9).", + "bbox": [ + 498, + 294, + 879, + 310 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Take the right inset as an example. We choose the HR image as the ground truth (e.g. HR in Fig. 5), when $r = 4$ , every latent code $z^{*}$ corresponds to a $8 \\times 8$ patch of HR (gray points) in relative coordinate domain (blue region). To calculate the $i$ -th position of $z^{*}$ , we multiply every HR pixel value", + "bbox": [ + 498, + 310, + 723, + 445 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c3e0f18e5283550c5f3e197ca6920b72751ad2be79e0e5b1afd13585496f518a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 728, + 311, + 890, + 441 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$I_{HR}(x',y')$ and basis value $P[i](x',y')$ together and finally sum them. After getting the feature map, we render it to the same size of $I_{HR}$ via OPE-Upscale Module and calculate their Peak Signal-to-Noise Ratio (PSNR).", + "bbox": [ + 498, + 445, + 890, + 506 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 518, + 633, + 537 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Parameter Setting", + "text_level": 1, + "bbox": [ + 500, + 544, + 676, + 561 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluated the performance of different maximum frequencies $n$ on 50 images from the DIV2K validation set [1] under different scales $r$ . Since our method is a local representation, we do not use a large $n$ . As shown in Tab. 1, the optimal sampling frequency for a given $r_i$ is always $r_i - 1$ . This observation can also be explained by the Nyquist-Shannon sampling theorem. For example, when $r_i = 4$ , there are $8 \\times 8$ sampling points for every latent code to \"fit\", so the maximum frequency that can be recovered from these sampling points should be less than 4. We also tested larger frequency with $r_i \\leq n \\leq 2 \\times r_i$ , reaching the upper limit that equals the number of sampling points. We further visualize the reconstructed images for different $n$ . As shown in Fig. 5, with scale factor $\\times 4$ , the larger frequency ( $n > 3$ ) brings redundant high-frequency information that sharpens the resulting images.", + "bbox": [ + 496, + 568, + 890, + 809 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Based on the above analysis, we decide to choose $n = 3$ as the maximum frequency for our OPE-Upscale Module. This is because existing arbitrary-scale SR methods, such as LIIF [5] and LTE [32], are trained with random scale factors up to 4, and the frequency $n = 4 - 1 = 3$ is sufficient to fully capture the ground truth information. Al-", + "bbox": [ + 496, + 810, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "10013", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/39c9f50803cba958c9d706c10d3701d5a21ea936fd7647cc5a97fdfd6276e8f5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
n×2×3×4×5×6×7×8
131.195128.608326.442425.048524.111423.342322.7898
230.747233.658631.209128.802227.370126.191325.3838
322.187133.658535.198332.401130.613528.896427.8159
412.123028.608334.963134.629434.097931.446230.2865
5-22.846529.951234.629337.370433.728532.8190
6-22.846524.312232.401137.150635.304635.8250
7--19.159328.802233.303935.304639.1160
8--12.086325.048529.096633.728638.9863
", + "bbox": [ + 83, + 88, + 460, + 204 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Representation performance (PSNR (dB)). The best value for each upsampling factor is bolded.", + "bbox": [ + 76, + 208, + 468, + 236 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1ed6aacf9d28f0b5fd6221f6ccc37ab98dd42e8b9689c7eff37379a229040c83.jpg", + "image_caption": [ + "Figure 5. Qualitative comparison of different OPE frequency n under scale factor $\\times 4$ (PSNR (dB)). 1-th row: residuals from $n = 0$ in image time domain. 2-th row: fourier frequency domain of HR and rendered image with $n$ . 3-th row: HR image and rendered image with $n$ ." + ], + "image_footnote": [], + "bbox": [ + 83, + 247, + 467, + 422 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "though a larger $n$ could represent more detailed patches of the target SR image, it would also introduce redundant high-frequency information and potentially increase computation time and memory consumption during training. Therefore, we choose $n = 3$ as the maximum frequency for our OPE-Upscale Module to balance performance and efficiency.", + "bbox": [ + 75, + 525, + 468, + 617 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Training", + "text_level": 1, + "bbox": [ + 76, + 631, + 181, + 647 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. Similar to [5, 32], we use the DIV2K dataset [1] of the NTIRE 2017 Challenge [63] for training. For testing, we use the DIV2K validation set [1] with 100 images and four benchmark datasets: Set5 [4], Set14 [76], B100 [40], and Urban100 [21]. We use PSNR as the quality measure.", + "bbox": [ + 75, + 656, + 467, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details. We mainly follow the prior implementation [5, 32] for arbitrary-scale SR training after replacing their upsampling module with OPE-Upscale. We use EDSR-baseline [35] and RDN [80] without their upsampling modules as the encoder, which is the only trainable part of our network. We use $48 \\times 48$ patches cropped from training set as inputs, L1 loss and Adam [26] optimizer for optimization. The network was trained for 1000 epochs with batch size 16, while the initial learning rate is 1e-4 and decayed by factor 0.5 every 200 epochs. More implementation details are presented in supplementary material.", + "bbox": [ + 75, + 734, + 467, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Evaluation", + "text_level": 1, + "bbox": [ + 500, + 90, + 620, + 104 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative results. Tab. 2 and Tab. 3 report quantitative results of OPE and the SOTA arbitrary-scale SR methods on the DIV2K validation set and the benchmark datasets. It is worth noting that we focus on finding an alternative of MLP with position encoding, rather than enhancing it like LTE [32]. We observe that our method achieves comparable results (less than 0.1dB on DIV2K and less than 0.15dB on benchmark), which indicates that our method is a feasible analytical solution with good performance and efficient parameter-free module. As shown in Tab. 2, EDSR [35] and RDN [80] are our selected encoders, and we achieve the highest efficiency (i.e. the shortest inference time in red number) comparing to all the other baselines with both encoders. The higher the scale factor, the better result we achieve. Specifically, in out-scale SR ( $\\times 6$ to $\\times 30$ ), our method outperforms most baselines and just has a small gap with LTE (less than 0.1dB). Such results demonstrate that our method has rich representation capability. We also compared with the benchmark dataset, as shown in Tab. 3, we keep comparable results to baselines (the gap is less than 0.15dB). However, as a nonlinear representation method, MLP still has advantages over our linear representation with low scale factors. See Sec. 5 for discussion on this issue.", + "bbox": [ + 496, + 114, + 890, + 460 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative results. Fig. 6 provides qualitative results with SOTA methods by using different scale factor. We show competitive visual quality against others, more results are provided in supplementary material. From the local perspective, LIIF and LTE only generate smooth patches, while our OPE with max frequency 3 is enough to achieve similar visual quality. We also notice LIIF [5] has artifact (vertical stripes) in the 1st row, this is a common drawback for implicit neural representation and is hard to be explained. However, with our image representation, there is no artifacts. In the 2nd row, we could observe a sprout (in red rectangle) in the GT, the same region of LIIF is vanished, and the boundary of our sprout is more obvious than LTE.", + "bbox": [ + 496, + 462, + 890, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Computing efficiency. We measure computing efficiency with MACs (multiply-accumulate operations), FLOPs (floating point operations) and actual running time. In Tab. 4 column 2-3, judged by the time complexity measured by the number of operations, we save 2 orders of magnitude. In our upsampling module, there is only one matrix operation and essential position encoding between input and output. In Tab. 2 we show shortest inference time benefiting from our compact SR framework. To further demonstrate our time advantage on large size images, we take $256 \\times 256$ as LR input of encoder and calculate time consumption of upsampling module with scale factor $\\times 4 \\times 30$ on NVIDIA RTX 3090. As shown in Tab. 5, our upsampling module shows $26\\% - 57\\%$ time advantage, this advantage keeps growing with larger scale factor. Notice We do not take advantage of GPU acceleration to design the upsampling", + "bbox": [ + 496, + 659, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "10014", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/aac223839dec03f1e84574ef5e7b063b56f045530bb6f881d9a6206fcde2bbd2.jpg", + "image_caption": [ + "Figure 6. Qualitative comparison with SOTA methods for arbitrary-scale SR. RDN [80] is used as encoder for all methods." + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 890, + 292 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2220f6947053ac6a8124e35c4df6a4bec0be982a1da05f288c3f132b3b19f8fc.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodIn-scaleOut-scale
×2×3×4×6×12×18×24×30
Bicubic [35]31.0128.2226.6624.8222.2721.0020.1919.59
EDSR-baseline [35]34.5530.9028.94-----
EDSR-baseline-MetaSR# [5, 18]34.6430.9328.9226.6123.5522.0321.0620.37
EDSR-baseline-LIIF [5]34.67 / 170230.96 / 127729.00 / 114426.75 / 104623.71 / 96522.17 / 95321.18 / 95120.48 / 947
EDSR-baseline-LTE [32]34.72 / 115831.02 / 107929.04 / 104526.81 / 102323.78 / 100722.23 / 100521.24 / 100320.53 / 1000
EDSR-baseline-OPE (ours)34.34 / 47630.94 / 39529.02 / 36426.77 / 34823.74 / 32222.21 / 31821.21 / 31420.52 / 311
RDN-MetaSR# [5, 18]35.0031.2729.2526.8823.7322.1821.1720.47
RDN-LIIF [5]34.99 / 310731.26 / 207329.27 / 151326.99 / 124823.89 / 102522.34 / 99421.31 / 99120.59 / 972
RDN-LTE [32]35.04 / 254931.32 / 183929.33 / 142027.04 / 118423.95 / 104922.40 / 102721.36 / 102520.64 / 1014
RDN-OPE (ours)34.52 / 227731.17 / 149729.26 / 103926.98 / 81323.91 / 66322.36 / 62321.34 / 59620.63 / 590
", + "bbox": [ + 89, + 324, + 875, + 465 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Quantitative comparison with the SOTA methods for arbitrary-scale SR on the DIV2K validation set (PSNR (dB) / running time (ms per image)). $\\sharp$ indicates implementation in LIIF [5]. With a parameter-free upsampling module, we narrow the gap between SOTA and ours in most results less than 0.1dB (blue) and obtain the shortest inference time (red).", + "bbox": [ + 75, + 467, + 892, + 508 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "module carefully, with hardware optimization, we believe our time advantage could be much larger thanks to fewer number of operations required.", + "bbox": [ + 75, + 536, + 468, + 583 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Memory consumption. In Tab. 4 column 4-5 we compare GPU memory consumption of OPE-Upscale Module with LIIF [5] and LTE [32] under training mode and testing mode of Pytorch [48]. For training mode, we use a $48 \\times 48$ patch as input and sample 2304 pixels as output following the default training strategy in arbitrary-scale SR works. For testing mode, we use $512 \\times 512$ image as input with scale factor 4. As a interpretable image representation without network parameters, OPE-Upscale Module saves memory of intermediate data (e.g. gradients, hidden layer outputs), and this advantage is fully reflected in training mode.", + "bbox": [ + 75, + 599, + 468, + 767 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Flipping consistency. As described in Sec. 1, the INR-based upsampling module like [5] is sensitive for the flipping of feature map. However, our method solves this problem completely and elegantly. The orthogonal basis of OPE is based on symmetric sinusoidal function, which leads to advantage of our method for keeping the flipping consistency. Also, more samples are provided in supplementary material for verifying other more flipping transforms.", + "bbox": [ + 75, + 784, + 470, + 906 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f8bf4b45e4adbc2597c95a129188961df8edd51595c6d51117cc75e7a18d9586.jpg", + "image_caption": [ + "Figure 7. Visual results of the ablation study on patch ensemble. See the text for details." + ], + "image_footnote": [], + "bbox": [ + 503, + 532, + 887, + 619 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 500, + 678, + 663, + 693 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In order to examine the effects of local ensemble styled interpolation (I)-Eq.(9) and extension of relative coordinate domain (E)-Eq.(8) in patch ensemble, we conducted experiments using four different settings with EDSR-baseline as the encoder. The four settings were: 1) OPE: OPE-Upscale module with I and E; 2) OPE-E: OPE-Upscale module without E but with I; 3) OPE-I: OPE-Upscale module without I but with E; and 4) OPE-IE: OPE-Upscale module with neither I nor E (that is, without patch ensemble).", + "bbox": [ + 496, + 702, + 890, + 838 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fig. 7 and Tab. 6 present the comparison results of different settings. In the OPE-E setting, only the nearest latent code to the query point can be rescaled to $[-1,1] \\times [-1,1]$ , while the relative coordinates of the other three latent codes", + "bbox": [ + 496, + 839, + 892, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "10015", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/06fa2f2c00a39af62efcf9d0f86c02a1b30212e86aa9548e26dc837ae6ce897a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodSet5Set14B100Urban100
In-scaleOut-scaleIn-scaleOut-scaleOut-scaleIn-scaleOut-scale
×2×3×4×6×8×2×3×4×6×8×2×3×4×6×8×2×3
RDN [80]38.2434.7132.47--34.0130.5728.81--32.3429.2627.72--32.8928.80
RDN-MetaSR# [5, 18]38.2234.6332.3829.0426.9633.9830.5428.7826.5124.9732.3329.2627.7125.9024.8332.9228.82
RDN-LIIF [5]38.1734.6832.5029.1527.1433.9730.5328.8026.6425.1532.3229.2627.7425.9824.9132.8728.82
RDN-LTE [32]38.2334.7232.6129.3227.2634.0930.5828.8826.7125.1632.3629.3027.7726.0124.9533.0428.97
RDN-OPE (ours)37.6034.5932.4729.1727.2233.3930.4928.8026.6525.1732.0529.1927.7225.9624.9131.7828.63
", + "bbox": [ + 78, + 88, + 895, + 188 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/73e639231a12cca83c681a35f039653ff4dfb231ce3dbd155c019f6a39cd30e0.jpg", + "table_caption": [ + "Table 3. Quantitative comparison with SOTA methods for arbitrary-scale image SR on benchmark datasets (PSNR (dB)). $\\sharp$ indicates implementation in LIIF [5]. We narrow the gap between SOTA and ours in most results less than $0.15\\mathrm{dB}$ (blue number). For large scale factor, we keep comparable results to MetaSR [18] and LIIF [5]. The defect in low scale factor will be analysed in Sec. 5." + ], + "table_footnote": [], + "table_body": "
MethodParamsMACsFLOPsMem (training)Mem (Test)
LIIF0.35 M429 K6.2 G85.1 + 1.9 M32 + 96 M
LTE0.26 M526 K7.5 G97.8 + 1.9 M64 + 96 M
OPE (ours)0 M6 K85 M0 + 1.9 M0 + 96 M
", + "bbox": [ + 84, + 244, + 460, + 297 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/8f9e5a50fcdd64c2c19a57192d0b779d5c492534a061883c36fb1dbff447fa1a.jpg", + "table_caption": [ + "Table 4. Parameter number, time complexity and memory consumption. MACs: multiply-accumulate operations, FLOPs: floating point operations, Mem: intermediate data + essential output for GPU memory consumption. We use $n = 3$ as maximum frequency of OPE and test in training mode and test mode on Pytorch with tool: torch.cuda.memory_allocated(). Training mode: $48^2$ to 2304 pixels, test mode: $512^2$ to $2048^2$ ." + ], + "table_footnote": [], + "table_body": "
Method×4×8×12×16×20×24×30
LIIF382152135306004102741835027866
LTE376149033405922102681834027838
OPE2771125249537195673836612012
Percentage28%26%30%39%45%55%57%
", + "bbox": [ + 89, + 405, + 457, + 470 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/729e7c550f780c9a069b700214aacbf9349d7306396b716c192533f8d37216db.jpg", + "table_caption": [ + "Table 5. Rendering time of upsampling module (ms per image) with an input resolution of $256 \\times 256$ . The last row shows the time saving percentage achieved by our method. We use $n = 3$ as the maximum frequency of OPE. Our method provides a time advantage, which increases as the rendering resolution increases. On average, we achieve a $40\\%$ reduction in rendering time." + ], + "table_footnote": [], + "table_body": "
In-scaleOut-scale
×2×3×4×6×8
OPE33.2930.2928.6526.4624.98
OPE-E33.2730.2328.5626.3424.82
OPE-I33.2830.2628.6326.4424.97
OPE-IE33.2030.0928.4426.2524.70
", + "bbox": [ + 109, + 571, + 434, + 647 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Ablation studies on Set14. EDSR-baseline [35] is used as encoder.", + "bbox": [ + 75, + 650, + 468, + 678 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "cannot be rescaled, resulting in periodic stripes in the resulting SR image. On the other hand, the OPE-I result shows no obvious discontinuity between patches since the extension plays a positive role, but this means that only a small region of the patch is presented in the target image. Lastly, the OPE-IE setting shows an obvious thick boundary between patches, indicating that both extension and interpolation are necessary for the best performance.", + "bbox": [ + 75, + 707, + 468, + 829 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Discussions", + "text_level": 1, + "bbox": [ + 76, + 844, + 197, + 859 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We observed that our quantitative results decreases at low scale factors, especially when the input size is small,", + "bbox": [ + 75, + 869, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "such as with benchmark datasets. See Tab. 3 and Tab. 2. This is due to the fact that a smaller target size $(W\\times H)$ leads to a larger grid in the 2D domain $\\left([-1,1]\\times [-1,1]\\right)$ , where utilizing only one central point value to represent the entire larger grid would result in a loss of detailed information compared to a smaller grid. Since the 2D domain we use is continuous, the higher the resolution of the target image, the stronger the representation ability we can achieve. MLP-based representations, such as [5] and [32], can overcome this issue through nonlinear operations. In our method, this defect can be ignored for high SR scale factors where pixels are dense, but for low scale such as $\\times 2$ or $\\times 3$ , our performance may slightly degrade. A possible way to address this issue is to sample more points for every grid region and calculate their mean value, with careful consideration of the time consumption trade-off.", + "bbox": [ + 496, + 250, + 893, + 491 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 506, + 617, + 522 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we proposed an interpretable method for continuous image representation without implicit neural networks. Our method leverages a novel position encoding technique called orthogonal position encoding, which takes the form a 2D-Fourier series and corresponds to 2D image coordinates. As a set of orthogonal basis in inner product space, OPE is both interpretable and rich in representation. Building on OPE, we introduced the OPE-Upscale Module, a parameter-free approach for arbitrary-scale image super-resolution that simplifies the existing deep SR framework, leading to high computing efficiency and less memory consumption. Our OPE-Upscale Module can be easily integrated into existing image super-resolution pipelines, and extensive experiments demonstrate that our method achieves competitive results with the state-of-the-art.", + "bbox": [ + 496, + 531, + 890, + 772 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "It is worth noting that the overall efficiency of SR framework depends on both the encoder and decoder. Since our work focuses on decoder design and efficiency, we leave the development of high-efficient encoder as future work.", + "bbox": [ + 496, + 773, + 890, + 834 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement This project was partially supported by NSFC Grants (61702363, 62222311, 62201274) and the Ministry of Education, Singapore, under its Academic Research Fund Grants (MOE-T2EP20220-0005 & RG20/20).", + "bbox": [ + 496, + 844, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "10016", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Eirikur Agustsson and Radu Timofte. Ntire 2017 challenge on single image super-resolution: Dataset and study. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 126-135, 2017. 5, 6", + "[2] Woong Bae, Jaejun Yoo, and Jong Chul Ye. Beyond deep residual learning for image restoration: Persistent homology-guided manifold simplification. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 145-153, 2017. 2", + "[3] Ronen Basri and David W Jacobs. Lambertian reflectance and linear subspaces. IEEE transactions on pattern analysis and machine intelligence, 25(2):218-233, 2003. 2", + "[4] Marco Bevilacqua, Aline Roumy, Christine Guillemot, and Marie Line Alberi-Morel. Low-complexity single-image super-resolution based on nonnegative neighbor embedding. 2012. 6", + "[5] Yinbo Chen, Sifei Liu, and Xiaolong Wang. Learning continuous image representation with local implicit image function. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8628-8638, 2021. 1, 2, 3, 5, 6, 7, 8", + "[6] Lu Chi, Borui Jiang, and Yadong Mu. Fast fourier convolution. Advances in Neural Information Processing Systems, 33:4479-4488, 2020. 2", + "[7] Jooyoung Choi, Jungbeom Lee, Yonghyun Jeong, and Sungroh Yoon. Toward spatially unbiased generative models. arXiv preprint arXiv:2108.01285, 2021. 2", + "[8] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Learning a deep convolutional network for image super-resolution. In European conference on computer vision, pages 184-199. Springer, 2014. 2", + "[9] Chao Dong, Chen Change Loy, and Xiaou Tang. Accelerating the super-resolution convolutional neural network. In European conference on computer vision, pages 391-407. Springer, 2016. 2, 3", + "[10] Vincent Dumoulin, Jonathon Shlens, and Manjunath Kudur. A learned representation for artistic style. arXiv preprint arXiv:1610.07629, 2016. 3", + "[11] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5501–5510, 2022. 2", + "[12] Manuel Fritsche, Shuhang Gu, and Radu Timofte. Frequency separation for real-world super-resolution. In 2019 IEEE/CVF International Conference on Computer Vision Workshop (ICCVW), pages 3599-3608. IEEE, 2019. 2", + "[13] Dario Fuoli, Luc Van Gool, and Radu Timofte. Fourier space losses for efficient perceptual image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2360-2369, 2021. 2", + "[14] Kyle Genova, Forrester Cole, Avneesh Sud, Aaron Sarna, and Thomas Funkhouser. Local deep implicit functions for 3d shape. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4857-4866, 2020. 1" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Tiantong Guo, Hojjat Seyed Mousavi, Tiep Huu Vu, and Vishal Monga. Deep wavelet prediction for image superresolution. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 104-113, 2017. 2", + "[16] Muhammad Haris, Gregory Shakhnarovich, and Norimichi Ukita. Deep back-projection networks for super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1664-1673, 2018. 2", + "[17] Kurt Hornik, Maxwell Stinchcombe, and Halbert White. Multilayer feedforward networks are universal approximators. Neural networks, 2(5):359-366, 1989. 1", + "[18] Xuecai Hu, Haoyuan Mu, Xiangyu Zhang, Zilei Wang, Tieniu Tan, and Jian Sun. Meta-sr: A magnification-arbitrary network for super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1575-1584, 2019. 1, 3, 7, 8", + "[19] Binbin Huang, Xinhao Yan, Anpei Chen, Shenghua Gao, and Jingyi Yu. Pref: Phasorial embedding fields for compact neural representations. arXiv preprint arXiv:2205.13524, 2022. 1", + "[20] Huaibo Huang, Ran He, Zhenan Sun, and Tieniu Tan. Wavelet-srnet: A wavelet-based cnn for multi-scale face super resolution. In Proceedings of the IEEE International Conference on Computer Vision, pages 1689-1697, 2017. 2", + "[21] Jia-Bin Huang, Abhishek Singh, and Narendra Ahuja. Single image super-resolution from transformed self-exemplars. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5197-5206, 2015. 6", + "[22] Liming Jiang, Bo Dai, Wayne Wu, and Chen Change Loy. Focal frequency loss for image reconstruction and synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13919-13929, 2021. 2", + "[23] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. Advances in Neural Information Processing Systems, 34:852-863, 2021. 2", + "[24] Jiwon Kim, Jung Kwon Lee, and Kyoung Mu Lee. Accurate image super-resolution using very deep convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1646-1654, 2016. 2", + "[25] Jiwon Kim, Jung Kwon Lee, and Kyoung Mu Lee. Deeplyrecursive convolutional network for image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1637-1645, 2016. 2", + "[26] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6", + "[27] Tom Koornwinder. Two-variable analogues of the classical orthogonal polynomials. In Theory and application of special functions, pages 435-495. Elsevier, 1975. 2", + "[28] Wei-Sheng Lai, Jia-Bin Huang, Narendra Ahuja, and Ming-Hsuan Yang. Deep laplacian pyramid networks for fast and accurate super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 624-632, 2017. 2", + "[29] Wei-Sheng Lai, Jia-Bin Huang, Narendra Ahuja, and Ming-Hsuan Yang. Fast and accurate image super-resolution with" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "10017", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "deep laplacian pyramid networks. IEEE transactions on pattern analysis and machine intelligence, 41(11):2599-2613, 2018. 2", + "[30] Seyed Mehdi Lajevardi and Zahir M Hussain. Higher order orthogonal moments for invariant facial expression recognition. Digital Signal Processing, 20(6):1771-1779, 2010. 2", + "[31] Christian Ledig, Lucas Theis, Ferenc Huszár, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photorealistic single image super-resolution using a generative adversarial network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4681-4690, 2017. 2", + "[32] Jaewon Lee and Kyong Hwan Jin. Local texture estimator for implicit representation function. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1929-1938, 2022. 1, 2, 3, 5, 6, 7, 8", + "[33] Zhen Li, Jinglei Yang, Zheng Liu, Xiaomin Yang, Gwang-gil Jeon, and Wei Wu. Feedback network for image super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3867-3876, 2019. 2", + "[34] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1833-1844, 2021. 3", + "[35] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 136-144, 2017. 2, 6, 7, 8", + "[36] Chieh Hubert Lin, Hsin-Ying Lee, Yen-Chi Cheng, Sergey Tulyakov, and Ming-Hsuan Yang. Infinitygan: Towards infinite-pixel image synthesis. arXiv preprint arXiv:2104.03963, 2021. 2", + "[37] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. Advances in Neural Information Processing Systems, 33:15651-15663, 2020. 2", + "[38] Pengju Liu, Hongzhi Zhang, Kai Zhang, Liang Lin, and Wangmeng Zuo. Multi-level wavelet-cnn for image restoration. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 773-782, 2018. 2", + "[39] Julien Mairal, Michael Elad, and Guillermo Sapiro. Sparse representation for color image restoration. IEEE Transactions on image processing, 17(1):53-69, 2007. 2", + "[40] David Martin, Charless Fowlkes, Doron Tal, and Jitendra Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In Proceedings Eighth IEEE International Conference on Computer Vision. ICCV 2001, volume 2, pages 416-423. IEEE, 2001. 6", + "[41] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "of the IEEE/CVF conference on computer vision and pattern recognition, pages 4460-4470, 2019. 1", + "[42] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 2, 3", + "[43] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11453-11464, 2021. 2", + "[44] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3504-3515, 2020. 1", + "[45] Michael Oechsle, Lars Mescheder, Michael Niemeyer, Thilo Strauss, and Andreas Geiger. Texture fields: Learning texture representations in function space. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4531-4540, 2019. 1", + "[46] Alan V. Oppenheim and Alan S. Willsky. Signals and Systems. Prentice Hall, Upper Saddle River, NJ, 2 edition, 1997. 4", + "[47] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 165-174, 2019. 1", + "[48] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019.", + "[49] Songyou Peng, Michael Niemeyer, Lars Mescheder, Marc Pollefeys, and Andreas Geiger. Convolutional occupancy networks. In European Conference on Computer Vision, pages 523-540. Springer, 2020. 1", + "[50] Nasim Rahaman, Aristide Baratin, Devansh Arpit, Felix Draxler, Min Lin, Fred Hamprecht, Yoshua Bengio, and Aaron Courville. On the spectral bias of neural networks. In International Conference on Machine Learning, pages 5301-5310. PMLR, 2019. 1, 2", + "[51] Ravi Ramamoorthi and Pat Hanrahan. On the relationship between radiance and irradiance: determining the illumination from images of a convex lambertian object. JOSA A, 18(10):2448-2459, 2001. 2", + "[52] Oren Rippel, Jasper Snoek, and Ryan P Adams. Spectral representations for convolutional neural networks. Advances in neural information processing systems, 28, 2015. 2", + "[53] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. Graf: Generative radiance fields for 3d-aware image synthesis. Advances in Neural Information Processing Systems, 33:20154-20166, 2020. 2", + "[54] Wenzhe Shi, Jose Caballero, Ferenc Huszár, Johannes Totz, Andrew P Aitken, Rob Bishop, Daniel Rueckert, and Zehan" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10018", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Wang. Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1874-1883, 2016. 3", + "[55] Assaf Shocher, Nadav Cohen, and Michal Irani. “zero-shot” super-resolution using deep internal learning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3118–3126, 2018. 2", + "[56] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. Advances in Neural Information Processing Systems, 33:7462-7473, 2020. 2", + "[57] Vincent Sitzmann, Justus Thies, Felix Heide, Matthias Nießner, Gordon Wetzstein, and Michael Zollhofer. Deepvoxels: Learning persistent 3d feature embeddings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2437-2446, 2019. 1", + "[58] Vincent Sitzmann, Michael Zollhöfer, and Gordon Wetzstein. Scene representation networks: Continuous 3d-structure-aware neural scene representations. Advances in Neural Information Processing Systems, 32, 2019. 1", + "[59] Peter-Pike Sloan, Jan Kautz, and John Snyder. Precomputed radiance transfer for real-time rendering in dynamic, low-frequency lighting environments. In Proceedings of the 29th annual conference on Computer graphics and interactive techniques, pages 527-536, 2002. 2", + "[60] Sanghyun Son and Kyoung Mu Lee. Srwarp: Generalized image super-resolution under arbitrary transformation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7782-7791, 2021. 3", + "[61] Ying Tai, Jian Yang, and Xiaoming Liu. Image superresolution via deep recursive residual network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3147-3155, 2017. 2", + "[62] Ying Tai, Jian Yang, Xiaoming Liu, and Chunyan Xu. Memnet: A persistent memory network for image restoration. In Proceedings of the IEEE international conference on computer vision, pages 4539-4547, 2017. 2", + "[63] Radu Timofte, Eirikur Agustsson, Luc Van Gool, Ming-Hsuan Yang, and Lei Zhang. Ntire 2017 challenge on single image super-resolution: Methods and results. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 114-125, 2017. 6", + "[64] Tong Tong, Gen Li, Xiejie Liu, and Qinquan Gao. Image super-resolution using dense skip connections. In Proceedings of the IEEE international conference on computer vision, pages 4799-4807, 2017. 2", + "[65] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2", + "[66] Longguang Wang, Yingqian Wang, Zaiping Lin, Jungang Yang, Wei An, and Yulan Guo. Learning a single network for scale-arbitrary super-resolution. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4801-4810, 2021. 3", + "[67] Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real-esrgan: Training real-world blind super-resolution with" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "pure synthetic data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1905-1914, 2021. 3", + "[68] Yifan Wang, Federico Perazzi, Brian McWilliams, Alexander Sorkine-Hornung, Olga Sorkine-Hornung, and Christopher Schroers. A fully progressive approach to single-image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 864-873, 2018. 2", + "[69] Yifan Wang, Federico Perazzi, Brian McWilliams, Alexander Sorkine-Hornung, Olga Sorkine-Hornung, and Christopher Schroers. A fully progressive approach to single-image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 864-873, 2018. 2", + "[70] Zhihao Wang, Jian Chen, and Steven CH Hoi. Deep learning for image super-resolution: A survey. IEEE transactions on pattern analysis and machine intelligence, 43(10):3365-3387, 2020. 1, 2, 3", + "[71] John Wright, Allen Y Yang, Arvind Ganesh, S Shankar Sastry, and Yi Ma. Robust face recognition via sparse representation. IEEE transactions on pattern analysis and machine intelligence, 31(2):210-227, 2008. 2", + "[72] Xingqian Xu, Zhangyang Wang, and Humphrey Shi. Ultrasr: Spatial encoding is a missing key for implicit image function-based arbitrary-scale super-resolution. arXiv preprint arXiv:2103.12716, 2021. 1, 2, 3", + "[73] Jianchao Yang, John Wright, Thomas Huang, and Yi Ma. Image super-resolution as sparse representation of raw image patches. In 2008 IEEE conference on computer vision and pattern recognition, pages 1-8. IEEE, 2008. 2", + "[74] Yanchao Yang and Stefano Soatto. Fda: Fourier domain adaptation for semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4085-4095, 2020. 2", + "[75] Youngho Yoon, Inchul Chung, Lin Wang, and Kuk-Jin Yoon. Spheresr: 360deg image super-resolution with arbitrary projection via continuous spherical image representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5677-5686, 2022. 3", + "[76] Roman Zeyde, Michael Elad, and Matan Protter. On single image scale-up using sparse-representations. In International conference on curves and surfaces, pages 711-730. Springer, 2010. 6", + "[77] Hui Zhang, Huazhong Shu, Guoniu N Han, Gouenou Coa-trieux, Limin Luo, and Jean Louis Coa-trieux. Blurred image recognition by legendre moment invariants. IEEE Transactions on Image Processing, 19(3):596-611, 2009. 2", + "[78] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492, 2020. 2", + "[79] Yulun Zhang, Kunpeng Li, Kai Li, Lichen Wang, Bineng Zhong, and Yun Fu. Image super-resolution using very deep residual channel attention networks. In Proceedings of the European conference on computer vision (ECCV), pages 286-301, 2018. 2", + "[80] Yulun Zhang, Yapeng Tian, Yu Kong, Bineng Zhong, and Yun Fu. Residual dense network for image super-resolution." + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "10019", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2472-2481, 2018. 6, 7, 8", + "[81] Man Zhou, Hu Yu, Jie Huang, Feng Zhao, Jinwei Gu, Chen Change Loy, Deyu Meng, and Chongyi Li. Deep fourier up-sampling. arXiv preprint arXiv:2210.05171, 2022. 2", + "[82] Hongqing Zhu. Image representation using separable two-dimensional continuous and discrete orthogonal moments. Pattern Recognition, 45(4):1540-1558, 2012. 2" + ], + "bbox": [ + 78, + 90, + 468, + 219 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "10020", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2023/OPE-SR_ Orthogonal Position Encoding for Designing a Parameter-Free Upsampling Module in Arbitrary-Scale Image Super-Resolution/538d0672-86b3-4cf7-ab0d-b9344b19df37_model.json b/2023/OPE-SR_ Orthogonal Position Encoding for Designing a Parameter-Free Upsampling Module in Arbitrary-Scale Image Super-Resolution/538d0672-86b3-4cf7-ab0d-b9344b19df37_model.json new file mode 100644 index 0000000000000000000000000000000000000000..353ee57691834507ffc74f344799e7960fe41f6f --- /dev/null +++ b/2023/OPE-SR_ Orthogonal Position Encoding for Designing a Parameter-Free Upsampling Module in Arbitrary-Scale Image Super-Resolution/538d0672-86b3-4cf7-ab0d-b9344b19df37_model.json @@ -0,0 +1,2765 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.125, + 0.13, + 0.849, + 0.177 + ], + "angle": 0, + "content": "OPE-SR: Orthogonal Position Encoding for Designing a Parameter-free Upsampling Module in Arbitrary-scale Image Super-Resolution" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.203, + 0.819, + 0.221 + ], + "angle": 0, + "content": "Gaochao Song\\(^{1}\\) Qian Sun\\(^{2*}\\) Luo Zhang\\(^{3}\\) Ran Su\\(^{1}\\) Jianfeng Shi\\(^{2}\\) Ying He\\(^{3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.159, + 0.222, + 0.812, + 0.258 + ], + "angle": 0, + "content": "\\(^{1}\\)Tianjin University \\(^{2}\\)Nanjing University of Information Science and Technology \\(^{3}\\)Nanyang Technological University" + }, + { + "type": "text", + "bbox": [ + 0.272, + 0.26, + 0.692, + 0.275 + ], + "angle": 0, + "content": "gaochaosong_21@tju.edu.cn sunqian@nuist.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.309, + 0.314, + 0.327 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.341, + 0.473, + 0.659 + ], + "angle": 0, + "content": "Arbitrary-scale image super-resolution (SR) is often tackled using the implicit neural representation (INR) approach, which relies on a position encoding scheme to improve its representation ability. In this paper, we introduce orthogonal position encoding (OPE), an extension of position encoding, and an OPE-Upscale module to replace the INR-based upsampling module for arbitrary-scale image super-resolution. Our OPE-Upscale module takes 2D coordinates and latent code as inputs, just like INR, but does not require any training parameters. This parameter-free feature allows the OPE-Upscale module to directly perform linear combination operations, resulting in continuous image reconstruction and achieving arbitrary-scale image reconstruction. As a concise SR framework, our method is computationally efficient and consumes less memory than state-of-the-art methods, as confirmed by extensive experiments and evaluations. In addition, our method achieves comparable results with state-of-the-art methods in arbitrary-scale image super-resolution. Lastly, we show that OPE corresponds to a set of orthogonal basis, validating our design principle." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.688, + 0.21, + 0.705 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.714, + 0.47, + 0.866 + ], + "angle": 0, + "content": "Photographs are composed of discrete pixels of varying precision due to the limitations of sampling frequency, which breaks the continuous visual world into discrete parts. The single image super-resolution (SISR) task aims to restore the original continuous world in the image as much as possible. In an arbitrary-scale SR task, one often reconstructs the continuous representation of a low-resolution image and then adjusts the resolution of the target image as needed. The recent rise of implicit neural representation (INR) in 3D vision has enabled the representation" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.311, + 0.892, + 0.371 + ], + "angle": 0, + "content": "of complex 3D objects and scenes in a continuous manner [14, 19, 41, 42, 44, 45, 47, 49, 57, 58], which also opens up possibilities for continuous image and arbitrary-scale image super-resolution [5, 18, 32, 72]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.372, + 0.893, + 0.628 + ], + "angle": 0, + "content": "Existing methods for arbitrary-scale SR typically use a post-upsampling framework [70]. In this approach, low-resolution (LR) images first pass through a deep CNN network (encoder) without improving the resolution, and then pass through an INR-based upsampling module (decoder) with a specified target resolution to reconstruct high-resolution (HR) images. The decoder establishes a mapping from feature maps (the output of encoder) to target image pixels using a pre-assigned grid partitioning and achieves arbitrary-scale with the density of the grid in Cartesian coordinate system. However, the INR approach has a defect of learning low-frequency information, also known as spectral bias [50]. To address this issue, sinusoidal positional encoding is introduced to embed input coordinates to higher dimensions and enable the network to learn high-frequency details. This inspired recent works on arbitrary-scale SR to further improve the representation ability [32, 72]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.895, + 0.825 + ], + "angle": 0, + "content": "Despite its effectiveness in arbitrary-scale SR, the INR-based upsampling module increases the complexity of the entire SR framework as two different networks are jointly trained. Additionally, as a black-box model, it represents a continuous image with a strong dependency on both the feature map and the decoder (e.g., MLP). However, its representation ability decreases after flipping the feature map, a phenomenon known as flipping consistency decline. As shown in Fig. 1, flipping the feature map horizontally before the upsampling module of LIIF results in a blurred target image that does not have the expected flip transformation. This decline could be due to limitations of the MLP in learning the symmetry feature of the image." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.895, + 0.903 + ], + "angle": 0, + "content": "MLP is a universal function approximator [17], which tries to fit a mapping function from feature map to the continuous image, therefore, it is reasonable to assume that such process could be solved by an analytical solution. In this paper, we re-examine position encoding from the per" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.875, + 0.223, + 0.888 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_footnote", + "bbox": [ + 0.097, + 0.888, + 0.453, + 0.901 + ], + "angle": 0, + "content": "1Project page: https://github.com/gaochao-s/ope-sr" + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.875, + 0.453, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "10009" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.242 + ], + "angle": 0, + "content": "spective of orthogonal basis and propose orthogonal position encoding (OPE) for continuous image representation. The linear combination of 1D latent code and OPE can directly reconstruct continuous image patch without using implicit neural function [5]. To prove OPE's rationality, we analyse it both from functional analysis and 2D-Fourier transform. We further embed it into a parameter-free upsampling module, called OPE-Upscale Module, to replace INR-based upsampling module in deep SR framework, then currently deep SR framework can be greatly simplified." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.243, + 0.473, + 0.349 + ], + "angle": 0, + "content": "Unlike the state-of-the-art method by Lee et al. [32], which enhances MLP with position encoding, we explore the possibility of extending position encoding without MLP. By providing a more concise SR framework, our method achieves high computing efficiency and consumes less memory than the state-of-the-art, while also achieving comparable image performance in arbitrary-scale SR tasks." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.349, + 0.316, + 0.362 + ], + "angle": 0, + "content": "Our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.371, + 0.47, + 0.462 + ], + "angle": 0, + "content": "- We propose a novel position encoding, called orthogonal position encoding (OPE), which takes the form of a 2D-Fourier series and corresponds to a set of orthogonal basis. Building on OPE, we introduce the OPE-Upscale Module, a parameter-free upsampling module for arbitrary-scale image super-resolution." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.47, + 0.472, + 0.515 + ], + "angle": 0, + "content": "- Our method significantly reduces the consumption of computing resources, resulting in high computing efficiency for arbitrary-scale SR tasks." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.523, + 0.47, + 0.585 + ], + "angle": 0, + "content": "- The OPE-Upscale Module is interpretable, parameter-free and does not require training, resulting in a concise SR framework that elegantly solves the flipping consistency problem." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.592, + 0.47, + 0.653 + ], + "angle": 0, + "content": "- Extensive experiments demonstrate that our method achieves comparable results with the state-of-the-art. Furthermore, our method enables super-resolution up to a large scale of \\(\\times 30\\)." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.371, + 0.472, + 0.653 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.672, + 0.22, + 0.687 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.697, + 0.353, + 0.714 + ], + "angle": 0, + "content": "2.1. Sinusoidal Positional Encoding" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Sinusoidal positional encoding is widely used to counteract the negative effects of token order and sequence length in sequence models [65], or to guide image generation as the spatial inductive bias in CNNs [7, 23, 36]. In implicit neural representations, it plays a critical role in solving spectral bias [50]. By embedding input coordinates into a higher dimensional space, position encoding greatly improves high-frequency representation of implicit 3D scenes [42, 56] and subsequent works take it as the default operation to improve representation quality [37, 43, 53, 78]. Inspired by these works, positional encoding has been preliminarily explored in representing images in arbitrary-scale SR [32, 72]." + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.089, + 0.696, + 0.159 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.529, + 0.159, + 0.591, + 0.17 + ], + "angle": 0, + "content": "LIIF:32.35" + }, + { + "type": "image_caption", + "bbox": [ + 0.609, + 0.159, + 0.693, + 0.17 + ], + "angle": 0, + "content": "LIIF-flip: 28.47" + }, + { + "type": "image", + "bbox": [ + 0.703, + 0.089, + 0.877, + 0.16 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.713, + 0.16, + 0.774, + 0.17 + ], + "angle": 0, + "content": "LIIF:25.43" + }, + { + "type": "image_caption", + "bbox": [ + 0.791, + 0.16, + 0.877, + 0.17 + ], + "angle": 0, + "content": "LIIF-flip: 23.11" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.171, + 0.696, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.529, + 0.24, + 0.592, + 0.251 + ], + "angle": 0, + "content": "OPE: 32.37" + }, + { + "type": "image_caption", + "bbox": [ + 0.609, + 0.24, + 0.693, + 0.252 + ], + "angle": 0, + "content": "OPE-flip: 32.37" + }, + { + "type": "image", + "bbox": [ + 0.704, + 0.171, + 0.877, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.712, + 0.24, + 0.775, + 0.251 + ], + "angle": 0, + "content": "OPE:25.50" + }, + { + "type": "image_caption", + "bbox": [ + 0.791, + 0.24, + 0.877, + 0.251 + ], + "angle": 0, + "content": "OPE-flip: 25.50" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.256, + 0.894, + 0.326 + ], + "angle": 0, + "content": "Figure 1. Flipping consistency decline (PSNR (dB)). LIIF-flip: Flipping the input of the LIIF [5] decoder yields blurred results in the symmetric outputs. OPE-flip: Our method does not exhibit such artifacts. Additional results can be found in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.355, + 0.788, + 0.371 + ], + "angle": 0, + "content": "2.2. Orthogonal Basis Decomposition" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.379, + 0.895, + 0.742 + ], + "angle": 0, + "content": "In functional analysis, an orthogonal basis decomposition is a way to represent a vector or function as a linear combination of orthogonal basis functions. Wavelet transform [2, 15, 20, 38] and 2D-Fourier transform [6, 12, 13, 22, 74] are commonly used decomposition techniques for images and videos. In DSGAN [12], the input image is explicitly decomposed into low and high frequencies using high-pass and low-pass filters. Other methods use frequency domain losses to decompose images, either in a supervised manner [22] or in an unsupervised manner [13]. To address with resolution discrepancy of reconstructed images and input images, Rippel et al. [52] employ spectral pooling to decrease resolution by truncating in the Fourier domain, while Zhou et al. [81] explore an up-sampling method in the Fourier domain to increase the resolution. Image moments, which decompose images into two-dimensional orthogonal polynomials [27, 82], are widely used in invariant pattern recognition [30, 77]. Image sparse representation inherits this decomposition idea and performs well in traditional computer vision tasks [39, 71, 73]. In 3D domains, spherical harmonics serve as an orthogonal basis in space to represent view dependence [3, 51, 59] and have recently been proposed as a replacement for MLPs [11] for representing neural radiance fields [42]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.755, + 0.729, + 0.772 + ], + "angle": 0, + "content": "2.3. Deep Learning-based SR" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Based on the upsampling operations and their location in the model, deep learning-based SR frameworks can be classified into four categories (see [70] for a comprehensive survey): pre-upsampling [8,24,25,55,61,62], post-upsampling [9,31,35,64,79], progressive-upsampling [28,29,68], and iterative up-and-down sampling [16, 33, 69]. With preupsampling, the LR image is first upsampled by traditional interpolation and then fed into a deep CNN to reconstruct" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10010" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.304 + ], + "angle": 0, + "content": "high-quality details. While it was one of the most popular frameworks for arbitrary-scale factors, it has side effects like enlarged noise by interpolation and high time complexity and space consumption. The progressive-upsampling and iterative up-and-down sampling frameworks pose challenges in terms of complicated model designing and unclear design criteria, as noted in [70]. For post-upsampling, the LR image is directly fed as input to a deep CNN, and then a trainable upsampling module (e.g., deconvolution [9], subpixel [54], and interpolation convolution [10]) increases the resolution at the end. Since feature extraction process, which is computationally intensive, only occurs in low-dimensional space, it has become one of the mainstream frameworks [32, 34, 67]." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.311, + 0.262, + 0.327 + ], + "angle": 0, + "content": "2.4. Arbitrary-scale SR" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.334, + 0.473, + 0.548 + ], + "angle": 0, + "content": "In the field of arbitrary-scale SR, most existing works are based on the post-upsampling framework and replace the traditional upsampling module with an INR-based one, such as a coordinate-based MLP. Meta-SR [18] was the first arbitrary-scale SR method based on CNN. ArbSR [66] adopts a general plug-in module to solve the scaling problem of different horizontal and vertical scales. SRWarp [60] transforms LR images into HR images with arbitrary shapes via a differential adaptive warping layer. SphereSR [75] explores arbitrary-scale on \\(360^{\\circ}\\) images. LIIF [5] uses coordinates and conditional latent code into an MLP to directly predict target pixel color with an intuitive network structure. LIIF-related follow-up works focus on predicting high-frequency information with position encoding [32,72]." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.558, + 0.169, + 0.572 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.582, + 0.375, + 0.599 + ], + "angle": 0, + "content": "3.1. OPE-based Image Representation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.605, + 0.47, + 0.759 + ], + "angle": 0, + "content": "Given an LR image \\( I_{LR} \\) with resolution of \\( h \\times w \\), we decide its 2D domain into \\( h \\times w \\) grids, where each grid represents a pixel in the LR image and corresponds to a patch of the high resolution image of size \\( r_h \\times r_w \\). The output image \\( I_{SR} \\) has a resolution of \\( H \\times W \\), where \\( H = r_h \\cdot h \\) and \\( W = r_w \\cdot w \\). Denote by \\( I \\in \\mathbb{R}^{r_h \\times r_w \\times 1} \\) the high resolution image patch with size \\( r_h \\times r_w \\) for a specific color channel. We view each pixel in \\( I \\) as a sample of a continuous bivariate function \\( f(x,y): [-1,1] \\times [-1,1] \\to \\mathbb{R} \\). We embed the coordinates \\( x \\) and \\( y \\) using sinusoidal positional encoding" + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.763, + 0.469, + 0.813 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} { X } { = \\gamma ( x ) = \\sqrt { 2 } \\cdot [ \\frac { 1 } { \\sqrt { 2 } } , \\cos ( \\pi x ) , \\sin ( \\pi x ) , \\cos ( 2 \\pi x ) , } \\\\ \\sin (2 \\pi x), \\dots , \\cos (n \\pi x), \\sin (n \\pi x) ] \\tag {1} \\\\ \\end{array}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.099, + 0.814, + 0.469, + 0.865 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} { Y } { = \\gamma ( y ) = \\sqrt { 2 } \\cdot \\left[ \\frac { 1 } { \\sqrt { 2 } } , \\cos ( \\pi y ) , \\sin ( \\pi y ) , \\cos ( 2 \\pi y ) , \\right. } \\\\ \\left. \\sin (2 \\pi y), \\dots , \\cos (n \\pi y), \\sin (n \\pi y) \\right], \\tag {2} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.87, + 0.47, + 0.902 + ], + "angle": 0, + "content": "where \\(\\gamma (\\cdot):\\mathbb{R}\\to \\mathbb{R}^{1\\times (2n + 1)}\\) is a univariate function for position encoding with a predefined maximum frequency" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.091, + 0.892, + 0.12 + ], + "angle": 0, + "content": "\\(n\\in \\mathbb{N}\\) . We flatten the matrix \\(X^T Y\\) into a row vector \\(P\\in\\) \\(\\mathbb{R}^{1\\times (2n + 1)^2}\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.638, + 0.121, + 0.892, + 0.138 + ], + "angle": 0, + "content": "\\[\nP = \\operatorname {f l a t} \\left(X ^ {T} Y\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.142, + 0.894, + 0.207 + ], + "angle": 0, + "content": "where flat: \\(\\mathbb{R}^{(2n + 1)\\times (2n + 1)}\\to \\mathbb{R}^{1\\times (2n + 1)^2}\\) is the flattening operation. Denote by \\(e_{i,j}\\) the element on the \\(i\\) -th row and the \\(j\\) -th column of matrix \\(X^T Y\\) . For example. \\(e_{4,5} = 2\\cos (2\\pi x)\\sin (2\\pi y)\\) . It is easy to verify that" + }, + { + "type": "equation", + "bbox": [ + 0.549, + 0.215, + 0.892, + 0.255 + ], + "angle": 0, + "content": "\\[\n\\left\\langle e _ {i _ {1}, j _ {1}}, e _ {i _ {2}, j _ {2}} \\right\\rangle = \\left\\{ \\begin{array}{l l} 0, & (i _ {1}, j _ {1}) \\neq (i _ {2}, j _ {2}) \\\\ 1, & (i _ {1}, j _ {1}) = (i _ {2}, j _ {2}) \\end{array} \\right. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.263, + 0.847, + 0.28 + ], + "angle": 0, + "content": "where \\(\\langle ,\\rangle\\) is the inner product in function space, i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.561, + 0.287, + 0.834, + 0.323 + ], + "angle": 0, + "content": "\\[\n\\langle g, h \\rangle = \\frac {1}{4} \\int_ {- 1} ^ {1} \\int_ {- 1} ^ {1} g (x, y) h (x, y) \\mathrm {d} x \\mathrm {d} y.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.331, + 0.892, + 0.424 + ], + "angle": 0, + "content": "Therefore the elements \\(\\{e_{i,j}\\}\\) form a set of orthogonal basis, allowing us to approximate \\(f\\) as a linear combination. Consider a pixel \\((x,y)\\) in the upsampled image patch \\(I\\), which corresponds to a region \\([x - 1 / r_h, x + 1 / r_h] \\times [y - 1 / r_w, y + 1 / r_w]\\). We use \\(f(x,y)\\) as a representative for the entire region and compute \\(I_{(x,y)}\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.61, + 0.431, + 0.892, + 0.451 + ], + "angle": 0, + "content": "\\[\nI _ {(x, y)} \\triangleq f (x, y) \\approx Z P ^ {T}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.46, + 0.892, + 0.522 + ], + "angle": 0, + "content": "where \\( Z \\in \\mathbb{R}^{1 \\times (2n + 1)^2} \\) represents the projection. Due to the orthogonal property, we call the resulting vector \\( P \\) as orthogonal position encoding (OPE). Fig. 3 illustrates the concept of OPE-based patch representation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.522, + 0.892, + 0.596 + ], + "angle": 0, + "content": "Remark 1. Our OPE basis can be seen as the real form version of the 2D-Fourier basis, which eliminates the complex exponential term based on conjugate symmetry when representing real signals. See the supplementary material for details." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.598, + 0.892, + 0.748 + ], + "angle": 0, + "content": "Remark 2. OPE differs from the commonly used positional encoding formulation, such as that in [42], in that it includes a constant term and takes the product of each coordinate embedding as a new term. These seemingly minor changes indeed have a deep impact. Conventional positional encoding processes the coordinates separately, thereby encoding frequencies in the horizontal and vertical directions only. In contrast, OPE includes frequencies covering in all directions of the plane due to the product terms, resulting in a better expression capability." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.749, + 0.892, + 0.793 + ], + "angle": 0, + "content": "Remark 3. Unlike LIIF [5] that uses an MLP to approximate the image function \\( f \\), our method completely eliminates the need of MLP." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.802, + 0.704, + 0.818 + ], + "angle": 0, + "content": "3.2. OPE-Upscale Module" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We project the latent code onto the OPE basis. OPE with a sufficiently long latent code could represent an image directly in a continuous manner. However, it suffers from long embedding time and is unstable for representing high-frequency details locally, similar to the limitations" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "10011" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.881, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.255, + 0.895, + 0.316 + ], + "angle": 0, + "content": "Figure 2. OPE-Upscale Module for arbitrary-scale SR. The encoder \\( E_{\\theta} \\) is the only trainable part. With a pre-defined maximum frequency \\( n \\) of OPE, the OPE-Upscale Module (shaded in grey) takes the feature map from \\( E_{\\theta} \\) and the target resolution \\( H \\), \\( W \\) as input, and renders the pixels of the target SR image in parallel. \\( \\odot \\) is the matmul function that returns the product of \\( z^{*} \\in \\mathbb{R}^{1 \\times 3(2n + 1)^{2}} \\) and OPEs \\( \\in \\mathbb{R}^{3(2n + 1)^{2} \\times 1} \\) per color channel." + }, + { + "type": "image", + "bbox": [ + 0.09, + 0.337, + 0.456, + 0.479 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.483, + 0.471, + 0.541 + ], + "angle": 0, + "content": "Figure 3. OPE-based patch representation. An image patch can be represented as a linear combination of basic plane waves in a continuous manner. The representation is a 2D extension to [46, Chap. 3.3.1]. Refer to the supplementary material for more details." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.564, + 0.47, + 0.687 + ], + "angle": 0, + "content": "of Fourier transform to describe local information. To address this issue, we represent the input image as the seamless stitching of local patches, whose latent codes are extracted from a feature map over the channel dimension2. As shown in Fig. 2, the OPE-upscale module takes both the target resolution \\( H = r_h \\cdot h \\), \\( W = r_w \\cdot w \\) and the feature map \\( \\in \\mathbb{R}^{h \\times w \\times 3(2n + 1)^2} \\) from the deep encoder \\( E_{\\theta} \\) as inputs and computes target pixels in parallel." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.687, + 0.471, + 0.856 + ], + "angle": 0, + "content": "Feature map rendering. As shown in Fig. 4, to render a target image \\( I_{SR} \\) with size \\( H \\times W \\) from a LR image \\( I_{LR} \\) with size \\( h \\times w \\), OPE-Upscale Module firstly divide a 2D domain \\( [-1,1] \\times [-1,1] \\) into \\( H \\times W \\) regions with equal size, so that every pixel in \\( I_{SR} \\) will be associated with an absolute central coordinates \\( (x_q, y_q) \\) in corresponding region. Secondly, the latent codes in the feature map (same dimension with \\( I_{LR} \\)) also possess corresponding central coordinates \\( (x_c, y_c) \\in [-1,1] \\times [-1,1] \\) by dividing same 2D domain into \\( h \\times w \\) regions, therefore, given a target image pixel with \\( (x_q, y_q) \\), a specific latent code \\( z^* \\in \\mathbb{R}^{1 \\times 3(2n + 1)^2} \\) with" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.337, + 0.875, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.48, + 0.895, + 0.565 + ], + "angle": 0, + "content": "Figure 4. Illustration of the mapping from the input LR image to the output SR image. LR image, feature map and SR image are all divided into the same domain \\([-1,1] \\times [-1,1]\\). Each LR pixel \\(I_{LR}(x_{LR},y_{LR})\\) corresponds to a latent code with coordinates \\((x_c,y_c)\\), while each SR pixel \\(I_{SR}(x_{SR},y_{SR})\\) corresponds to a latent code with coordinates \\((x_q,y_q)\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.588, + 0.895, + 0.649 + ], + "angle": 0, + "content": "coordinates \\((x_{c},y_{c})\\), which has the smallest distance from \\((x_{q},y_{q})\\) could be found. As shown in Eq.(6) and Eq.(7), a render function \\(\\mathcal{R}\\) takes two parts of inputs: \\(z^{*}\\) and \\((x_q',y_q')\\), to calculate final target pixel value as following:" + }, + { + "type": "equation", + "bbox": [ + 0.592, + 0.65, + 0.892, + 0.67 + ], + "angle": 0, + "content": "\\[\nI _ {S R} \\left(x _ {q}, y _ {q}\\right) = \\mathcal {R} \\left(z ^ {*}, \\left(x _ {q} ^ {\\prime}, y _ {q} ^ {\\prime}\\right)\\right) \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.557, + 0.683, + 0.892, + 0.702 + ], + "angle": 0, + "content": "\\[\nx _ {q} ^ {\\prime} = \\left(x _ {q} - x _ {c}\\right) \\cdot h, \\quad y _ {q} ^ {\\prime} = \\left(y _ {q} - y _ {c}\\right) \\cdot w \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.893, + 0.902 + ], + "angle": 0, + "content": "where \\( z^{*} \\) is the nearest latent code we found, and \\( (x_{q}^{\\prime}, y_{q}^{\\prime}) \\) are relative coordinates, which are calculated based on Eq.(7) to rescale the absolute coordinates (in domain \\( [-1, 1] \\times [-1, 1] \\)) by times \\( h \\) and \\( w \\), which is taken as input by function \\( \\mathcal{R} \\) to render target pixel. \\( \\mathcal{R} \\) has the similar calculation as Eq.(5) while the difference is it repeats OPE three times to adapt \\( z^{*} \\) and calculate the linear combination per color channel. In this way, our OPE-upscale module successfully deals with arbitrary size \\( I_{SR} \\) by processing each pixel by \\( \\mathcal{R} \\), in which feature map rendering process is parameter-free with high computing efficiency and less memory consumption (which has been confirmed by the experiments in Sec. 4.3)." + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.863, + 0.471, + 0.902 + ], + "angle": 0, + "content": "2To ensure compatibility with color images, we adjust the output channel of the encoder to \\(3(2n + 1)^{2}\\), where \\(n\\) is the pre-defined maximum frequency \\(n\\) of OPE." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "10012" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.075, + 0.092, + 0.473, + 0.246 + ], + "angle": 0, + "content": "Patch ensemble. As shown in Fig. 4, when moving \\((x_{q},y_{q})\\) from location \\(a\\) to \\(b\\), the pixel value in the target image \\(I_{SR}\\) may change abruptly, as well as the nearest latent code \\(z^{*}\\). To address this discontinuity issue, we propose a patch ensemble, which is a local ensemble styled interpolation using relative coordinates. Instead of finding a single nearest latent code for \\((x_{q},y_{q})\\), we select the nearest four neighbouring latent codes \\(z_{t}^{*}\\) with corresponding central coordinates \\((x_{t},y_{t})\\), where \\(t\\in \\{00,01,10,11\\}\\). Then we calculate the relative coordinates \\(x_{q}^{\\prime}\\) and \\(y_{q}^{\\prime}\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.129, + 0.249, + 0.47, + 0.28 + ], + "angle": 0, + "content": "\\[\nx _ {q} ^ {\\prime} = \\frac {\\left(x _ {q} - x _ {t}\\right) \\cdot h}{2}, \\quad y _ {q} ^ {\\prime} = \\frac {\\left(y _ {q} - y _ {t}\\right) \\cdot w}{2}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.282, + 0.469, + 0.328 + ], + "angle": 0, + "content": "which can guarantee \\( x_{q}^{\\prime}, y_{q}^{\\prime} \\in [-1, 1] \\). As the inset shows, for the four adjacent pixels from the low resolution image \\( I_{LR} \\) (i.e. related to latent codes \\( z_{00}^{*}, z_{01}^{*}, z_{10}^{*} \\) and \\( z_{11}^{*} \\))," + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.328, + 0.302, + 0.463 + ], + "angle": 0, + "content": "their corresponding patches in super resolution image \\(I_{SR}\\) are colored in red, green, yellow and blue, respectively. The pixel color in \\(I_{SR}\\) is not solely dependent on the nearest latent code but considers the four neighboring latent codes. Specifically, using the rendering function \\(\\mathcal{R}\\) and the di" + }, + { + "type": "image", + "bbox": [ + 0.306, + 0.33, + 0.468, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.464, + 0.469, + 0.493 + ], + "angle": 0, + "content": "aagonal rectangle areas \\( s_t \\) as weights, we compute the pixel value as a weighted sum" + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.497, + 0.47, + 0.533 + ], + "angle": 0, + "content": "\\[\nI _ {S R} \\left(x _ {q}, y _ {q}\\right) = \\sum_ {t \\in \\{0 0, 0 1, 1 0, 1 1 \\}} \\frac {s _ {t}}{S} \\cdot \\mathcal {R} \\left(z _ {t} ^ {*}, \\left(x _ {q} ^ {\\prime}, y _ {q} ^ {\\prime}\\right)\\right), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.537, + 0.47, + 0.629 + ], + "angle": 0, + "content": "where \\( S = \\sum_{t} s_{t} \\) is the sum of areas. Considering the contribution of each latent code allows us to integrate the adjacent latent codes with different significance, thereby providing a seamless stitching of adjacent patches. We call Eq.(9) local ensemble styled interpolation since it takes a similar form of the local ensemble in LIIF [5]." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.635, + 0.295, + 0.652 + ], + "angle": 0, + "content": "3.3. Maximum Frequency \\( n \\)" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.659, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Selecting a proper maximum frequency \\( n \\) plays an important role in designing and implementing the OPE-upsampling module since it directly determines the network architecture and also has effects on the performance of different SR scales. Given a high resolution image \\( I_{HR} \\) with size \\( H \\times W \\), \\( n \\) and \\( r \\), we aim to obtain a feature map with size \\( \\frac{H}{r} \\times \\frac{W}{r} \\), then we re-render the obtained feature map with the selected \\( n \\). By the comparison of the rendered \\( I_{SR} \\), we present the performance of \\( n \\in \\{1,2,\\dots,8\\} \\) under different \\( r \\) values (SR scale), as show in Tab. 1, and select the \\( n \\) with the best performance (the details would be discussed Sec. 4.1). To be specific, we use Eq.(10) as the basic theory and use Eq.(11) to infer the feature map. First, similar to calculate the projection of a normal vector on orthogonal basis, we can calculate projections (or so-called latent code) \\( Z \\in \\mathbb{R}^{1 \\times (2n + 1)^2} \\) of \\( f(x,y) \\) in Eq.(5) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.103, + 0.892, + 0.139 + ], + "angle": 0, + "content": "\\[\nZ [ i ] = \\frac {1}{4} \\int_ {- 1} ^ {1} \\int_ {- 1} ^ {1} f (x, y) P [ i ] (x, y) \\mathrm {d} x \\mathrm {d} y \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.144, + 0.894, + 0.236 + ], + "angle": 0, + "content": "where \\(P[i](x,y)\\) is a bivariate function taken from the \\(i\\)-th position of OPE and \\(Z[i]\\) is the corresponding projection. Based on Eq.(10) and taking both the discreteness of an image and the design of OPE-Upscale Module into consideration, we calculate the feature map of an image \\(I_{HR}\\) with down-sampling scale \\(r\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.246, + 0.892, + 0.289 + ], + "angle": 0, + "content": "\\[\nz ^ {*} [ i ] = \\frac {1}{4} \\sum_ {x ^ {\\prime}} ^ {2 r} \\sum_ {y ^ {\\prime}} ^ {2 r} I _ {H R} \\left(x ^ {\\prime}, y ^ {\\prime}\\right) P [ i ] \\left(x ^ {\\prime}, y ^ {\\prime}\\right) \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.295, + 0.88, + 0.311 + ], + "angle": 0, + "content": "It can be considered as the inverse operation of Eq.(9)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.311, + 0.724, + 0.446 + ], + "angle": 0, + "content": "Take the right inset as an example. We choose the HR image as the ground truth (e.g. HR in Fig. 5), when \\( r = 4 \\), every latent code \\( z^{*} \\) corresponds to a \\( 8 \\times 8 \\) patch of HR (gray points) in relative coordinate domain (blue region). To calculate the \\( i \\)-th position of \\( z^{*} \\), we multiply every HR pixel value" + }, + { + "type": "image", + "bbox": [ + 0.729, + 0.312, + 0.892, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.446, + 0.892, + 0.507 + ], + "angle": 0, + "content": "\\(I_{HR}(x',y')\\) and basis value \\(P[i](x',y')\\) together and finally sum them. After getting the feature map, we render it to the same size of \\(I_{HR}\\) via OPE-Upscale Module and calculate their Peak Signal-to-Noise Ratio (PSNR)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.52, + 0.634, + 0.538 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.545, + 0.677, + 0.562 + ], + "angle": 0, + "content": "4.1. Parameter Setting" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.569, + 0.892, + 0.81 + ], + "angle": 0, + "content": "We evaluated the performance of different maximum frequencies \\( n \\) on 50 images from the DIV2K validation set [1] under different scales \\( r \\). Since our method is a local representation, we do not use a large \\( n \\). As shown in Tab. 1, the optimal sampling frequency for a given \\( r_i \\) is always \\( r_i - 1 \\). This observation can also be explained by the Nyquist-Shannon sampling theorem. For example, when \\( r_i = 4 \\), there are \\( 8 \\times 8 \\) sampling points for every latent code to \"fit\", so the maximum frequency that can be recovered from these sampling points should be less than 4. We also tested larger frequency with \\( r_i \\leq n \\leq 2 \\times r_i \\), reaching the upper limit that equals the number of sampling points. We further visualize the reconstructed images for different \\( n \\). As shown in Fig. 5, with scale factor \\( \\times 4 \\), the larger frequency (\\( n > 3 \\)) brings redundant high-frequency information that sharpens the resulting images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Based on the above analysis, we decide to choose \\( n = 3 \\) as the maximum frequency for our OPE-Upscale Module. This is because existing arbitrary-scale SR methods, such as LIIF [5] and LTE [32], are trained with random scale factors up to 4, and the frequency \\( n = 4 - 1 = 3 \\) is sufficient to fully capture the ground truth information. Al-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10013" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.084, + 0.089, + 0.462, + 0.205 + ], + "angle": 0, + "content": "
n×2×3×4×5×6×7×8
131.195128.608326.442425.048524.111423.342322.7898
230.747233.658631.209128.802227.370126.191325.3838
322.187133.658535.198332.401130.613528.896427.8159
412.123028.608334.963134.629434.097931.446230.2865
5-22.846529.951234.629337.370433.728532.8190
6-22.846524.312232.401137.150635.304635.8250
7--19.159328.802233.303935.304639.1160
8--12.086325.048529.096633.728638.9863
" + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.209, + 0.47, + 0.237 + ], + "angle": 0, + "content": "Table 1. Representation performance (PSNR (dB)). The best value for each upsampling factor is bolded." + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.248, + 0.468, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.426, + 0.47, + 0.496 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparison of different OPE frequency n under scale factor \\(\\times 4\\) (PSNR (dB)). 1-th row: residuals from \\(n = 0\\) in image time domain. 2-th row: fourier frequency domain of HR and rendered image with \\(n\\). 3-th row: HR image and rendered image with \\(n\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.526, + 0.47, + 0.618 + ], + "angle": 0, + "content": "though a larger \\( n \\) could represent more detailed patches of the target SR image, it would also introduce redundant high-frequency information and potentially increase computation time and memory consumption during training. Therefore, we choose \\( n = 3 \\) as the maximum frequency for our OPE-Upscale Module to balance performance and efficiency." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.632, + 0.182, + 0.648 + ], + "angle": 0, + "content": "4.2. Training" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.657, + 0.468, + 0.734 + ], + "angle": 0, + "content": "Datasets. Similar to [5, 32], we use the DIV2K dataset [1] of the NTIRE 2017 Challenge [63] for training. For testing, we use the DIV2K validation set [1] with 100 images and four benchmark datasets: Set5 [4], Set14 [76], B100 [40], and Urban100 [21]. We use PSNR as the quality measure." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Implementation details. We mainly follow the prior implementation [5, 32] for arbitrary-scale SR training after replacing their upsampling module with OPE-Upscale. We use EDSR-baseline [35] and RDN [80] without their upsampling modules as the encoder, which is the only trainable part of our network. We use \\(48 \\times 48\\) patches cropped from training set as inputs, L1 loss and Adam [26] optimizer for optimization. The network was trained for 1000 epochs with batch size 16, while the initial learning rate is 1e-4 and decayed by factor 0.5 every 200 epochs. More implementation details are presented in supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.091, + 0.621, + 0.106 + ], + "angle": 0, + "content": "4.3. Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.115, + 0.892, + 0.461 + ], + "angle": 0, + "content": "Quantitative results. Tab. 2 and Tab. 3 report quantitative results of OPE and the SOTA arbitrary-scale SR methods on the DIV2K validation set and the benchmark datasets. It is worth noting that we focus on finding an alternative of MLP with position encoding, rather than enhancing it like LTE [32]. We observe that our method achieves comparable results (less than 0.1dB on DIV2K and less than 0.15dB on benchmark), which indicates that our method is a feasible analytical solution with good performance and efficient parameter-free module. As shown in Tab. 2, EDSR [35] and RDN [80] are our selected encoders, and we achieve the highest efficiency (i.e. the shortest inference time in red number) comparing to all the other baselines with both encoders. The higher the scale factor, the better result we achieve. Specifically, in out-scale SR (\\( \\times 6 \\) to \\( \\times 30 \\)), our method outperforms most baselines and just has a small gap with LTE (less than 0.1dB). Such results demonstrate that our method has rich representation capability. We also compared with the benchmark dataset, as shown in Tab. 3, we keep comparable results to baselines (the gap is less than 0.15dB). However, as a nonlinear representation method, MLP still has advantages over our linear representation with low scale factors. See Sec. 5 for discussion on this issue." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.463, + 0.892, + 0.658 + ], + "angle": 0, + "content": "Qualitative results. Fig. 6 provides qualitative results with SOTA methods by using different scale factor. We show competitive visual quality against others, more results are provided in supplementary material. From the local perspective, LIIF and LTE only generate smooth patches, while our OPE with max frequency 3 is enough to achieve similar visual quality. We also notice LIIF [5] has artifact (vertical stripes) in the 1st row, this is a common drawback for implicit neural representation and is hard to be explained. However, with our image representation, there is no artifacts. In the 2nd row, we could observe a sprout (in red rectangle) in the GT, the same region of LIIF is vanished, and the boundary of our sprout is more obvious than LTE." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Computing efficiency. We measure computing efficiency with MACs (multiply-accumulate operations), FLOPs (floating point operations) and actual running time. In Tab. 4 column 2-3, judged by the time complexity measured by the number of operations, we save 2 orders of magnitude. In our upsampling module, there is only one matrix operation and essential position encoding between input and output. In Tab. 2 we show shortest inference time benefiting from our compact SR framework. To further demonstrate our time advantage on large size images, we take \\(256 \\times 256\\) as LR input of encoder and calculate time consumption of upsampling module with scale factor \\(\\times 4 \\times 30\\) on NVIDIA RTX 3090. As shown in Tab. 5, our upsampling module shows \\(26\\% - 57\\%\\) time advantage, this advantage keeps growing with larger scale factor. Notice We do not take advantage of GPU acceleration to design the upsampling" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10014" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.083, + 0.089, + 0.891, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.116, + 0.298, + 0.851, + 0.311 + ], + "angle": 0, + "content": "Figure 6. Qualitative comparison with SOTA methods for arbitrary-scale SR. RDN [80] is used as encoder for all methods." + }, + { + "type": "table", + "bbox": [ + 0.09, + 0.325, + 0.876, + 0.466 + ], + "angle": 0, + "content": "
MethodIn-scaleOut-scale
×2×3×4×6×12×18×24×30
Bicubic [35]31.0128.2226.6624.8222.2721.0020.1919.59
EDSR-baseline [35]34.5530.9028.94-----
EDSR-baseline-MetaSR# [5, 18]34.6430.9328.9226.6123.5522.0321.0620.37
EDSR-baseline-LIIF [5]34.67 / 170230.96 / 127729.00 / 114426.75 / 104623.71 / 96522.17 / 95321.18 / 95120.48 / 947
EDSR-baseline-LTE [32]34.72 / 115831.02 / 107929.04 / 104526.81 / 102323.78 / 100722.23 / 100521.24 / 100320.53 / 1000
EDSR-baseline-OPE (ours)34.34 / 47630.94 / 39529.02 / 36426.77 / 34823.74 / 32222.21 / 31821.21 / 31420.52 / 311
RDN-MetaSR# [5, 18]35.0031.2729.2526.8823.7322.1821.1720.47
RDN-LIIF [5]34.99 / 310731.26 / 207329.27 / 151326.99 / 124823.89 / 102522.34 / 99421.31 / 99120.59 / 972
RDN-LTE [32]35.04 / 254931.32 / 183929.33 / 142027.04 / 118423.95 / 104922.40 / 102721.36 / 102520.64 / 1014
RDN-OPE (ours)34.52 / 227731.17 / 149729.26 / 103926.98 / 81323.91 / 66322.36 / 62321.34 / 59620.63 / 590
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.468, + 0.893, + 0.51 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison with the SOTA methods for arbitrary-scale SR on the DIV2K validation set (PSNR (dB) / running time (ms per image)). \\(\\sharp\\) indicates implementation in LIIF [5]. With a parameter-free upsampling module, we narrow the gap between SOTA and ours in most results less than 0.1dB (blue) and obtain the shortest inference time (red)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.537, + 0.47, + 0.584 + ], + "angle": 0, + "content": "module carefully, with hardware optimization, we believe our time advantage could be much larger thanks to fewer number of operations required." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.601, + 0.47, + 0.768 + ], + "angle": 0, + "content": "Memory consumption. In Tab. 4 column 4-5 we compare GPU memory consumption of OPE-Upscale Module with LIIF [5] and LTE [32] under training mode and testing mode of Pytorch [48]. For training mode, we use a \\(48 \\times 48\\) patch as input and sample 2304 pixels as output following the default training strategy in arbitrary-scale SR works. For testing mode, we use \\(512 \\times 512\\) image as input with scale factor 4. As a interpretable image representation without network parameters, OPE-Upscale Module saves memory of intermediate data (e.g. gradients, hidden layer outputs), and this advantage is fully reflected in training mode." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.785, + 0.471, + 0.907 + ], + "angle": 0, + "content": "Flipping consistency. As described in Sec. 1, the INR-based upsampling module like [5] is sensitive for the flipping of feature map. However, our method solves this problem completely and elegantly. The orthogonal basis of OPE is based on symmetric sinusoidal function, which leads to advantage of our method for keeping the flipping consistency. Also, more samples are provided in supplementary material for verifying other more flipping transforms." + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.534, + 0.888, + 0.621 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.623, + 0.892, + 0.649 + ], + "angle": 0, + "content": "Figure 7. Visual results of the ablation study on patch ensemble. See the text for details." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.679, + 0.665, + 0.694 + ], + "angle": 0, + "content": "4.4. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.703, + 0.892, + 0.839 + ], + "angle": 0, + "content": "In order to examine the effects of local ensemble styled interpolation (I)-Eq.(9) and extension of relative coordinate domain (E)-Eq.(8) in patch ensemble, we conducted experiments using four different settings with EDSR-baseline as the encoder. The four settings were: 1) OPE: OPE-Upscale module with I and E; 2) OPE-E: OPE-Upscale module without E but with I; 3) OPE-I: OPE-Upscale module without I but with E; and 4) OPE-IE: OPE-Upscale module with neither I nor E (that is, without patch ensemble)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.893, + 0.9 + ], + "angle": 0, + "content": "Fig. 7 and Tab. 6 present the comparison results of different settings. In the OPE-E setting, only the nearest latent code to the query point can be rescaled to \\([-1,1] \\times [-1,1]\\), while the relative coordinates of the other three latent codes" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10015" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.079, + 0.089, + 0.896, + 0.189 + ], + "angle": 0, + "content": "
MethodSet5Set14B100Urban100
In-scaleOut-scaleIn-scaleOut-scaleOut-scaleIn-scaleOut-scale
×2×3×4×6×8×2×3×4×6×8×2×3×4×6×8×2×3
RDN [80]38.2434.7132.47--34.0130.5728.81--32.3429.2627.72--32.8928.80
RDN-MetaSR# [5, 18]38.2234.6332.3829.0426.9633.9830.5428.7826.5124.9732.3329.2627.7125.9024.8332.9228.82
RDN-LIIF [5]38.1734.6832.5029.1527.1433.9730.5328.8026.6425.1532.3229.2627.7425.9824.9132.8728.82
RDN-LTE [32]38.2334.7232.6129.3227.2634.0930.5828.8826.7125.1632.3629.3027.7726.0124.9533.0428.97
RDN-OPE (ours)37.6034.5932.4729.1727.2233.3930.4928.8026.6525.1732.0529.1927.7225.9624.9131.7828.63
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.193, + 0.895, + 0.237 + ], + "angle": 0, + "content": "Table 3. Quantitative comparison with SOTA methods for arbitrary-scale image SR on benchmark datasets (PSNR (dB)). \\(\\sharp\\) indicates implementation in LIIF [5]. We narrow the gap between SOTA and ours in most results less than \\(0.15\\mathrm{dB}\\) (blue number). For large scale factor, we keep comparable results to MetaSR [18] and LIIF [5]. The defect in low scale factor will be analysed in Sec. 5." + }, + { + "type": "table", + "bbox": [ + 0.086, + 0.246, + 0.462, + 0.299 + ], + "angle": 0, + "content": "
MethodParamsMACsFLOPsMem (training)Mem (Test)
LIIF0.35 M429 K6.2 G85.1 + 1.9 M32 + 96 M
LTE0.26 M526 K7.5 G97.8 + 1.9 M64 + 96 M
OPE (ours)0 M6 K85 M0 + 1.9 M0 + 96 M
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.301, + 0.47, + 0.399 + ], + "angle": 0, + "content": "Table 4. Parameter number, time complexity and memory consumption. MACs: multiply-accumulate operations, FLOPs: floating point operations, Mem: intermediate data + essential output for GPU memory consumption. We use \\( n = 3 \\) as maximum frequency of OPE and test in training mode and test mode on Pytorch with tool: torch.cuda.memory_allocated(). Training mode: \\( 48^2 \\) to 2304 pixels, test mode: \\( 512^2 \\) to \\( 2048^2 \\)." + }, + { + "type": "table", + "bbox": [ + 0.091, + 0.406, + 0.458, + 0.471 + ], + "angle": 0, + "content": "
Method×4×8×12×16×20×24×30
LIIF382152135306004102741835027866
LTE376149033405922102681834027838
OPE2771125249537195673836612012
Percentage28%26%30%39%45%55%57%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.481, + 0.47, + 0.566 + ], + "angle": 0, + "content": "Table 5. Rendering time of upsampling module (ms per image) with an input resolution of \\(256 \\times 256\\). The last row shows the time saving percentage achieved by our method. We use \\(n = 3\\) as the maximum frequency of OPE. Our method provides a time advantage, which increases as the rendering resolution increases. On average, we achieve a \\(40\\%\\) reduction in rendering time." + }, + { + "type": "table", + "bbox": [ + 0.11, + 0.573, + 0.435, + 0.648 + ], + "angle": 0, + "content": "
In-scaleOut-scale
×2×3×4×6×8
OPE33.2930.2928.6526.4624.98
OPE-E33.2730.2328.5626.3424.82
OPE-I33.2830.2628.6326.4424.97
OPE-IE33.2030.0928.4426.2524.70
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.651, + 0.47, + 0.679 + ], + "angle": 0, + "content": "Table 6. Ablation studies on Set14. EDSR-baseline [35] is used as encoder." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.708, + 0.47, + 0.83 + ], + "angle": 0, + "content": "cannot be rescaled, resulting in periodic stripes in the resulting SR image. On the other hand, the OPE-I result shows no obvious discontinuity between patches since the extension plays a positive role, but this means that only a small region of the patch is presented in the target image. Lastly, the OPE-IE setting shows an obvious thick boundary between patches, indicating that both extension and interpolation are necessary for the best performance." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.845, + 0.199, + 0.86 + ], + "angle": 0, + "content": "5. Discussions" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.47, + 0.901 + ], + "angle": 0, + "content": "We observed that our quantitative results decreases at low scale factors, especially when the input size is small," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.25, + 0.895, + 0.492 + ], + "angle": 0, + "content": "such as with benchmark datasets. See Tab. 3 and Tab. 2. This is due to the fact that a smaller target size \\((W\\times H)\\) leads to a larger grid in the 2D domain \\(\\left([-1,1]\\times [-1,1]\\right)\\), where utilizing only one central point value to represent the entire larger grid would result in a loss of detailed information compared to a smaller grid. Since the 2D domain we use is continuous, the higher the resolution of the target image, the stronger the representation ability we can achieve. MLP-based representations, such as [5] and [32], can overcome this issue through nonlinear operations. In our method, this defect can be ignored for high SR scale factors where pixels are dense, but for low scale such as \\(\\times 2\\) or \\(\\times 3\\), our performance may slightly degrade. A possible way to address this issue is to sample more points for every grid region and calculate their mean value, with careful consideration of the time consumption trade-off." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.507, + 0.619, + 0.523 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.532, + 0.892, + 0.773 + ], + "angle": 0, + "content": "In this paper, we proposed an interpretable method for continuous image representation without implicit neural networks. Our method leverages a novel position encoding technique called orthogonal position encoding, which takes the form a 2D-Fourier series and corresponds to 2D image coordinates. As a set of orthogonal basis in inner product space, OPE is both interpretable and rich in representation. Building on OPE, we introduced the OPE-Upscale Module, a parameter-free approach for arbitrary-scale image super-resolution that simplifies the existing deep SR framework, leading to high computing efficiency and less memory consumption. Our OPE-Upscale Module can be easily integrated into existing image super-resolution pipelines, and extensive experiments demonstrate that our method achieves competitive results with the state-of-the-art." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.775, + 0.892, + 0.835 + ], + "angle": 0, + "content": "It is worth noting that the overall efficiency of SR framework depends on both the encoder and decoder. Since our work focuses on decoder design and efficiency, we leave the development of high-efficient encoder as future work." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.845, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Acknowledgement This project was partially supported by NSFC Grants (61702363, 62222311, 62201274) and the Ministry of Education, Singapore, under its Academic Research Fund Grants (MOE-T2EP20220-0005 & RG20/20)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10016" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Eirikur Agustsson and Radu Timofte. Ntire 2017 challenge on single image super-resolution: Dataset and study. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 126-135, 2017. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.172, + 0.472, + 0.241 + ], + "angle": 0, + "content": "[2] Woong Bae, Jaejun Yoo, and Jong Chul Ye. Beyond deep residual learning for image restoration: Persistent homology-guided manifold simplification. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 145-153, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.242, + 0.471, + 0.283 + ], + "angle": 0, + "content": "[3] Ronen Basri and David W Jacobs. Lambertian reflectance and linear subspaces. IEEE transactions on pattern analysis and machine intelligence, 25(2):218-233, 2003. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.284, + 0.471, + 0.338 + ], + "angle": 0, + "content": "[4] Marco Bevilacqua, Aline Roumy, Christine Guillemot, and Marie Line Alberi-Morel. Low-complexity single-image super-resolution based on nonnegative neighbor embedding. 2012. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.34, + 0.471, + 0.408 + ], + "angle": 0, + "content": "[5] Yinbo Chen, Sifei Liu, and Xiaolong Wang. Learning continuous image representation with local implicit image function. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8628-8638, 2021. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.41, + 0.471, + 0.451 + ], + "angle": 0, + "content": "[6] Lu Chi, Borui Jiang, and Yadong Mu. Fast fourier convolution. Advances in Neural Information Processing Systems, 33:4479-4488, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.453, + 0.471, + 0.493 + ], + "angle": 0, + "content": "[7] Jooyoung Choi, Jungbeom Lee, Yonghyun Jeong, and Sungroh Yoon. Toward spatially unbiased generative models. arXiv preprint arXiv:2108.01285, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.495, + 0.471, + 0.55 + ], + "angle": 0, + "content": "[8] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Learning a deep convolutional network for image super-resolution. In European conference on computer vision, pages 184-199. Springer, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.551, + 0.471, + 0.606 + ], + "angle": 0, + "content": "[9] Chao Dong, Chen Change Loy, and Xiaou Tang. Accelerating the super-resolution convolutional neural network. In European conference on computer vision, pages 391-407. Springer, 2016. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.607, + 0.471, + 0.648 + ], + "angle": 0, + "content": "[10] Vincent Dumoulin, Jonathon Shlens, and Manjunath Kudur. A learned representation for artistic style. arXiv preprint arXiv:1610.07629, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.649, + 0.471, + 0.718 + ], + "angle": 0, + "content": "[11] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5501–5510, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.719, + 0.471, + 0.774 + ], + "angle": 0, + "content": "[12] Manuel Fritsche, Shuhang Gu, and Radu Timofte. Frequency separation for real-world super-resolution. In 2019 IEEE/CVF International Conference on Computer Vision Workshop (ICCVW), pages 3599-3608. IEEE, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.776, + 0.471, + 0.831 + ], + "angle": 0, + "content": "[13] Dario Fuoli, Luc Van Gool, and Radu Timofte. Fourier space losses for efficient perceptual image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2360-2369, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.832, + 0.471, + 0.9 + ], + "angle": 0, + "content": "[14] Kyle Genova, Forrester Cole, Avneesh Sud, Aaron Sarna, and Thomas Funkhouser. Local deep implicit functions for 3d shape. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4857-4866, 2020. 1" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.161 + ], + "angle": 0, + "content": "[15] Tiantong Guo, Hojjat Seyed Mousavi, Tiep Huu Vu, and Vishal Monga. Deep wavelet prediction for image superresolution. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 104-113, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.894, + 0.218 + ], + "angle": 0, + "content": "[16] Muhammad Haris, Gregory Shakhnarovich, and Norimichi Ukita. Deep back-projection networks for super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1664-1673, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.219, + 0.894, + 0.259 + ], + "angle": 0, + "content": "[17] Kurt Hornik, Maxwell Stinchcombe, and Halbert White. Multilayer feedforward networks are universal approximators. Neural networks, 2(5):359-366, 1989. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.261, + 0.895, + 0.328 + ], + "angle": 0, + "content": "[18] Xuecai Hu, Haoyuan Mu, Xiangyu Zhang, Zilei Wang, Tieniu Tan, and Jian Sun. Meta-sr: A magnification-arbitrary network for super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1575-1584, 2019. 1, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.33, + 0.894, + 0.383 + ], + "angle": 0, + "content": "[19] Binbin Huang, Xinhao Yan, Anpei Chen, Shenghua Gao, and Jingyi Yu. Pref: Phasorial embedding fields for compact neural representations. arXiv preprint arXiv:2205.13524, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.385, + 0.894, + 0.441 + ], + "angle": 0, + "content": "[20] Huaibo Huang, Ran He, Zhenan Sun, and Tieniu Tan. Wavelet-srnet: A wavelet-based cnn for multi-scale face super resolution. In Proceedings of the IEEE International Conference on Computer Vision, pages 1689-1697, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.442, + 0.894, + 0.496 + ], + "angle": 0, + "content": "[21] Jia-Bin Huang, Abhishek Singh, and Narendra Ahuja. Single image super-resolution from transformed self-exemplars. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5197-5206, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.497, + 0.894, + 0.552 + ], + "angle": 0, + "content": "[22] Liming Jiang, Bo Dai, Wayne Wu, and Chen Change Loy. Focal frequency loss for image reconstruction and synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13919-13929, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.553, + 0.894, + 0.607 + ], + "angle": 0, + "content": "[23] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. Advances in Neural Information Processing Systems, 34:852-863, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.608, + 0.894, + 0.663 + ], + "angle": 0, + "content": "[24] Jiwon Kim, Jung Kwon Lee, and Kyoung Mu Lee. Accurate image super-resolution using very deep convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1646-1654, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.664, + 0.894, + 0.719 + ], + "angle": 0, + "content": "[25] Jiwon Kim, Jung Kwon Lee, and Kyoung Mu Lee. Deeplyrecursive convolutional network for image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1637-1645, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.72, + 0.894, + 0.761 + ], + "angle": 0, + "content": "[26] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.762, + 0.894, + 0.803 + ], + "angle": 0, + "content": "[27] Tom Koornwinder. Two-variable analogues of the classical orthogonal polynomials. In Theory and application of special functions, pages 435-495. Elsevier, 1975. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.804, + 0.894, + 0.871 + ], + "angle": 0, + "content": "[28] Wei-Sheng Lai, Jia-Bin Huang, Narendra Ahuja, and Ming-Hsuan Yang. Deep laplacian pyramid networks for fast and accurate super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 624-632, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[29] Wei-Sheng Lai, Jia-Bin Huang, Narendra Ahuja, and Ming-Hsuan Yang. Fast and accurate image super-resolution with" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.895, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10017" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.133 + ], + "angle": 0, + "content": "deep laplacian pyramid networks. IEEE transactions on pattern analysis and machine intelligence, 41(11):2599-2613, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.177 + ], + "angle": 0, + "content": "[30] Seyed Mehdi Lajevardi and Zahir M Hussain. Higher order orthogonal moments for invariant facial expression recognition. Digital Signal Processing, 20(6):1771-1779, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.469, + 0.274 + ], + "angle": 0, + "content": "[31] Christian Ledig, Lucas Theis, Ferenc Huszár, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photorealistic single image super-resolution using a generative adversarial network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4681-4690, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.278, + 0.471, + 0.332 + ], + "angle": 0, + "content": "[32] Jaewon Lee and Kyong Hwan Jin. Local texture estimator for implicit representation function. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1929-1938, 2022. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.334, + 0.469, + 0.402 + ], + "angle": 0, + "content": "[33] Zhen Li, Jinglei Yang, Zheng Liu, Xiaomin Yang, Gwang-gil Jeon, and Wei Wu. Feedback network for image super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3867-3876, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.405, + 0.469, + 0.473 + ], + "angle": 0, + "content": "[34] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1833-1844, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.476, + 0.469, + 0.545 + ], + "angle": 0, + "content": "[35] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 136-144, 2017. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.548, + 0.469, + 0.601 + ], + "angle": 0, + "content": "[36] Chieh Hubert Lin, Hsin-Ying Lee, Yen-Chi Cheng, Sergey Tulyakov, and Ming-Hsuan Yang. Infinitygan: Towards infinite-pixel image synthesis. arXiv preprint arXiv:2104.03963, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.604, + 0.469, + 0.657 + ], + "angle": 0, + "content": "[37] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. Advances in Neural Information Processing Systems, 33:15651-15663, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.661, + 0.469, + 0.728 + ], + "angle": 0, + "content": "[38] Pengju Liu, Hongzhi Zhang, Kai Zhang, Liang Lin, and Wangmeng Zuo. Multi-level wavelet-cnn for image restoration. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 773-782, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.732, + 0.469, + 0.773 + ], + "angle": 0, + "content": "[39] Julien Mairal, Michael Elad, and Guillermo Sapiro. Sparse representation for color image restoration. IEEE Transactions on image processing, 17(1):53-69, 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.469, + 0.857 + ], + "angle": 0, + "content": "[40] David Martin, Charless Fowlkes, Doron Tal, and Jitendra Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In Proceedings Eighth IEEE International Conference on Computer Vision. ICCV 2001, volume 2, pages 416-423. IEEE, 2001. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[41] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "of the IEEE/CVF conference on computer vision and pattern recognition, pages 4460-4470, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.121, + 0.892, + 0.188 + ], + "angle": 0, + "content": "[42] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.19, + 0.892, + 0.258 + ], + "angle": 0, + "content": "[43] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11453-11464, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.261, + 0.892, + 0.329 + ], + "angle": 0, + "content": "[44] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3504-3515, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.33, + 0.892, + 0.398 + ], + "angle": 0, + "content": "[45] Michael Oechsle, Lars Mescheder, Michael Niemeyer, Thilo Strauss, and Andreas Geiger. Texture fields: Learning texture representations in function space. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4531-4540, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.4, + 0.892, + 0.438 + ], + "angle": 0, + "content": "[46] Alan V. Oppenheim and Alan S. Willsky. Signals and Systems. Prentice Hall, Upper Saddle River, NJ, 2 edition, 1997. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.441, + 0.892, + 0.509 + ], + "angle": 0, + "content": "[47] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 165-174, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.511, + 0.892, + 0.591 + ], + "angle": 0, + "content": "[48] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.595, + 0.892, + 0.649 + ], + "angle": 0, + "content": "[49] Songyou Peng, Michael Niemeyer, Lars Mescheder, Marc Pollefeys, and Andreas Geiger. Convolutional occupancy networks. In European Conference on Computer Vision, pages 523-540. Springer, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.651, + 0.892, + 0.718 + ], + "angle": 0, + "content": "[50] Nasim Rahaman, Aristide Baratin, Devansh Arpit, Felix Draxler, Min Lin, Fred Hamprecht, Yoshua Bengio, and Aaron Courville. On the spectral bias of neural networks. In International Conference on Machine Learning, pages 5301-5310. PMLR, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.72, + 0.892, + 0.773 + ], + "angle": 0, + "content": "[51] Ravi Ramamoorthi and Pat Hanrahan. On the relationship between radiance and irradiance: determining the illumination from images of a convex lambertian object. JOSA A, 18(10):2448-2459, 2001. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.776, + 0.892, + 0.816 + ], + "angle": 0, + "content": "[52] Oren Rippel, Jasper Snoek, and Ryan P Adams. Spectral representations for convolutional neural networks. Advances in neural information processing systems, 28, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.818, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[53] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. Graf: Generative radiance fields for 3d-aware image synthesis. Advances in Neural Information Processing Systems, 33:20154-20166, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[54] Wenzhe Shi, Jose Caballero, Ferenc Huszár, Johannes Totz, Andrew P Aitken, Rob Bishop, Daniel Rueckert, and Zehan" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10018" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "Wang. Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1874-1883, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.47, + 0.204 + ], + "angle": 0, + "content": "[55] Assaf Shocher, Nadav Cohen, and Michal Irani. “zero-shot” super-resolution using deep internal learning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3118–3126, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.205, + 0.471, + 0.259 + ], + "angle": 0, + "content": "[56] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. Advances in Neural Information Processing Systems, 33:7462-7473, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.261, + 0.47, + 0.328 + ], + "angle": 0, + "content": "[57] Vincent Sitzmann, Justus Thies, Felix Heide, Matthias Nießner, Gordon Wetzstein, and Michael Zollhofer. Deepvoxels: Learning persistent 3d feature embeddings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2437-2446, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.33, + 0.47, + 0.385 + ], + "angle": 0, + "content": "[58] Vincent Sitzmann, Michael Zollhöfer, and Gordon Wetzstein. Scene representation networks: Continuous 3d-structure-aware neural scene representations. Advances in Neural Information Processing Systems, 32, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.386, + 0.47, + 0.455 + ], + "angle": 0, + "content": "[59] Peter-Pike Sloan, Jan Kautz, and John Snyder. Precomputed radiance transfer for real-time rendering in dynamic, low-frequency lighting environments. In Proceedings of the 29th annual conference on Computer graphics and interactive techniques, pages 527-536, 2002. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.456, + 0.47, + 0.51 + ], + "angle": 0, + "content": "[60] Sanghyun Son and Kyoung Mu Lee. Srwarp: Generalized image super-resolution under arbitrary transformation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7782-7791, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.511, + 0.47, + 0.566 + ], + "angle": 0, + "content": "[61] Ying Tai, Jian Yang, and Xiaoming Liu. Image superresolution via deep recursive residual network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3147-3155, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.567, + 0.47, + 0.622 + ], + "angle": 0, + "content": "[62] Ying Tai, Jian Yang, Xiaoming Liu, and Chunyan Xu. Memnet: A persistent memory network for image restoration. In Proceedings of the IEEE international conference on computer vision, pages 4539-4547, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.623, + 0.47, + 0.691 + ], + "angle": 0, + "content": "[63] Radu Timofte, Eirikur Agustsson, Luc Van Gool, Ming-Hsuan Yang, and Lei Zhang. Ntire 2017 challenge on single image super-resolution: Methods and results. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 114-125, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.692, + 0.47, + 0.747 + ], + "angle": 0, + "content": "[64] Tong Tong, Gen Li, Xiejie Liu, and Qinquan Gao. Image super-resolution using dense skip connections. In Proceedings of the IEEE international conference on computer vision, pages 4799-4807, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.748, + 0.47, + 0.803 + ], + "angle": 0, + "content": "[65] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.471, + 0.872 + ], + "angle": 0, + "content": "[66] Longguang Wang, Yingqian Wang, Zaiping Lin, Jungang Yang, Wei An, and Yulan Guo. Learning a single network for scale-arbitrary super-resolution. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4801-4810, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[67] Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real-esrgan: Training real-world blind super-resolution with" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.133 + ], + "angle": 0, + "content": "pure synthetic data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1905-1914, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.135, + 0.892, + 0.217 + ], + "angle": 0, + "content": "[68] Yifan Wang, Federico Perazzi, Brian McWilliams, Alexander Sorkine-Hornung, Olga Sorkine-Hornung, and Christopher Schroers. A fully progressive approach to single-image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 864-873, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.218, + 0.892, + 0.3 + ], + "angle": 0, + "content": "[69] Yifan Wang, Federico Perazzi, Brian McWilliams, Alexander Sorkine-Hornung, Olga Sorkine-Hornung, and Christopher Schroers. A fully progressive approach to single-image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 864-873, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.301, + 0.892, + 0.356 + ], + "angle": 0, + "content": "[70] Zhihao Wang, Jian Chen, and Steven CH Hoi. Deep learning for image super-resolution: A survey. IEEE transactions on pattern analysis and machine intelligence, 43(10):3365-3387, 2020. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.357, + 0.892, + 0.413 + ], + "angle": 0, + "content": "[71] John Wright, Allen Y Yang, Arvind Ganesh, S Shankar Sastry, and Yi Ma. Robust face recognition via sparse representation. IEEE transactions on pattern analysis and machine intelligence, 31(2):210-227, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.414, + 0.892, + 0.469 + ], + "angle": 0, + "content": "[72] Xingqian Xu, Zhangyang Wang, and Humphrey Shi. Ultrasr: Spatial encoding is a missing key for implicit image function-based arbitrary-scale super-resolution. arXiv preprint arXiv:2103.12716, 2021. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.469, + 0.892, + 0.525 + ], + "angle": 0, + "content": "[73] Jianchao Yang, John Wright, Thomas Huang, and Yi Ma. Image super-resolution as sparse representation of raw image patches. In 2008 IEEE conference on computer vision and pattern recognition, pages 1-8. IEEE, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.525, + 0.892, + 0.58 + ], + "angle": 0, + "content": "[74] Yanchao Yang and Stefano Soatto. Fda: Fourier domain adaptation for semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4085-4095, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.58, + 0.892, + 0.65 + ], + "angle": 0, + "content": "[75] Youngho Yoon, Inchul Chung, Lin Wang, and Kuk-Jin Yoon. Spheresr: 360deg image super-resolution with arbitrary projection via continuous spherical image representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5677-5686, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.65, + 0.892, + 0.705 + ], + "angle": 0, + "content": "[76] Roman Zeyde, Michael Elad, and Matan Protter. On single image scale-up using sparse-representations. In International conference on curves and surfaces, pages 711-730. Springer, 2010. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.706, + 0.892, + 0.762 + ], + "angle": 0, + "content": "[77] Hui Zhang, Huazhong Shu, Guoniu N Han, Gouenou Coa-trieux, Limin Luo, and Jean Louis Coa-trieux. Blurred image recognition by legendre moment invariants. IEEE Transactions on Image Processing, 19(3):596-611, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.762, + 0.892, + 0.803 + ], + "angle": 0, + "content": "[78] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.804, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[79] Yulun Zhang, Kunpeng Li, Kai Li, Lichen Wang, Bineng Zhong, and Yun Fu. Image super-resolution using very deep residual channel attention networks. In Proceedings of the European conference on computer vision (ECCV), pages 286-301, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.872, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[80] Yulun Zhang, Yapeng Tian, Yu Kong, Bineng Zhong, and Yun Fu. Residual dense network for image super-resolution." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10019" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.469, + 0.12 + ], + "angle": 0, + "content": "In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2472-2481, 2018. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.469, + 0.176 + ], + "angle": 0, + "content": "[81] Man Zhou, Hu Yu, Jie Huang, Feng Zhao, Jinwei Gu, Chen Change Loy, Deyu Meng, and Chongyi Li. Deep fourier up-sampling. arXiv preprint arXiv:2210.05171, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.468, + 0.22 + ], + "angle": 0, + "content": "[82] Hongqing Zhu. Image representation using separable two-dimensional continuous and discrete orthogonal moments. Pattern Recognition, 45(4):1540-1558, 2012. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "10020" + } + ] +] \ No newline at end of file diff --git a/2023/OPE-SR_ Orthogonal Position Encoding for Designing a Parameter-Free Upsampling Module in Arbitrary-Scale Image Super-Resolution/538d0672-86b3-4cf7-ab0d-b9344b19df37_origin.pdf b/2023/OPE-SR_ Orthogonal Position Encoding for Designing a Parameter-Free Upsampling Module in Arbitrary-Scale Image Super-Resolution/538d0672-86b3-4cf7-ab0d-b9344b19df37_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e1106f928f0e88cd3226c466be7804d95870a200 --- /dev/null +++ b/2023/OPE-SR_ Orthogonal Position Encoding for Designing a Parameter-Free Upsampling Module in Arbitrary-Scale Image Super-Resolution/538d0672-86b3-4cf7-ab0d-b9344b19df37_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0f63a320cb3a77e4025bb93217e3f82ea862c5b88c1cee9eccac054fa6e837d3 +size 9832211 diff --git a/2023/OPE-SR_ Orthogonal Position Encoding for Designing a Parameter-Free Upsampling Module in Arbitrary-Scale Image Super-Resolution/full.md b/2023/OPE-SR_ Orthogonal Position Encoding for Designing a Parameter-Free Upsampling Module in Arbitrary-Scale Image Super-Resolution/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e71a619e75d7b96819f35a455b3a48e0e2034a5e --- /dev/null +++ b/2023/OPE-SR_ Orthogonal Position Encoding for Designing a Parameter-Free Upsampling Module in Arbitrary-Scale Image Super-Resolution/full.md @@ -0,0 +1,371 @@ +# OPE-SR: Orthogonal Position Encoding for Designing a Parameter-free Upsampling Module in Arbitrary-scale Image Super-Resolution + +Gaochao Song $^{1}$ Qian Sun $^{2*}$ Luo Zhang $^{3}$ Ran Su $^{1}$ Jianfeng Shi $^{2}$ Ying He $^{3}$ + +$^{1}$ Tianjin University $^{2}$ Nanjing University of Information Science and Technology $^{3}$ Nanyang Technological University + +gaochaosong_21@tju.edu.cn sunqian@nuist.edu.cn + +# Abstract + +Arbitrary-scale image super-resolution (SR) is often tackled using the implicit neural representation (INR) approach, which relies on a position encoding scheme to improve its representation ability. In this paper, we introduce orthogonal position encoding (OPE), an extension of position encoding, and an OPE-Upscale module to replace the INR-based upsampling module for arbitrary-scale image super-resolution. Our OPE-Upscale module takes 2D coordinates and latent code as inputs, just like INR, but does not require any training parameters. This parameter-free feature allows the OPE-Upscale module to directly perform linear combination operations, resulting in continuous image reconstruction and achieving arbitrary-scale image reconstruction. As a concise SR framework, our method is computationally efficient and consumes less memory than state-of-the-art methods, as confirmed by extensive experiments and evaluations. In addition, our method achieves comparable results with state-of-the-art methods in arbitrary-scale image super-resolution. Lastly, we show that OPE corresponds to a set of orthogonal basis, validating our design principle. + +# 1. Introduction + +Photographs are composed of discrete pixels of varying precision due to the limitations of sampling frequency, which breaks the continuous visual world into discrete parts. The single image super-resolution (SISR) task aims to restore the original continuous world in the image as much as possible. In an arbitrary-scale SR task, one often reconstructs the continuous representation of a low-resolution image and then adjusts the resolution of the target image as needed. The recent rise of implicit neural representation (INR) in 3D vision has enabled the representation + +of complex 3D objects and scenes in a continuous manner [14, 19, 41, 42, 44, 45, 47, 49, 57, 58], which also opens up possibilities for continuous image and arbitrary-scale image super-resolution [5, 18, 32, 72]. + +Existing methods for arbitrary-scale SR typically use a post-upsampling framework [70]. In this approach, low-resolution (LR) images first pass through a deep CNN network (encoder) without improving the resolution, and then pass through an INR-based upsampling module (decoder) with a specified target resolution to reconstruct high-resolution (HR) images. The decoder establishes a mapping from feature maps (the output of encoder) to target image pixels using a pre-assigned grid partitioning and achieves arbitrary-scale with the density of the grid in Cartesian coordinate system. However, the INR approach has a defect of learning low-frequency information, also known as spectral bias [50]. To address this issue, sinusoidal positional encoding is introduced to embed input coordinates to higher dimensions and enable the network to learn high-frequency details. This inspired recent works on arbitrary-scale SR to further improve the representation ability [32, 72]. + +Despite its effectiveness in arbitrary-scale SR, the INR-based upsampling module increases the complexity of the entire SR framework as two different networks are jointly trained. Additionally, as a black-box model, it represents a continuous image with a strong dependency on both the feature map and the decoder (e.g., MLP). However, its representation ability decreases after flipping the feature map, a phenomenon known as flipping consistency decline. As shown in Fig. 1, flipping the feature map horizontally before the upsampling module of LIIF results in a blurred target image that does not have the expected flip transformation. This decline could be due to limitations of the MLP in learning the symmetry feature of the image. + +MLP is a universal function approximator [17], which tries to fit a mapping function from feature map to the continuous image, therefore, it is reasonable to assume that such process could be solved by an analytical solution. In this paper, we re-examine position encoding from the per + +spective of orthogonal basis and propose orthogonal position encoding (OPE) for continuous image representation. The linear combination of 1D latent code and OPE can directly reconstruct continuous image patch without using implicit neural function [5]. To prove OPE's rationality, we analyse it both from functional analysis and 2D-Fourier transform. We further embed it into a parameter-free upsampling module, called OPE-Upscale Module, to replace INR-based upsampling module in deep SR framework, then currently deep SR framework can be greatly simplified. + +Unlike the state-of-the-art method by Lee et al. [32], which enhances MLP with position encoding, we explore the possibility of extending position encoding without MLP. By providing a more concise SR framework, our method achieves high computing efficiency and consumes less memory than the state-of-the-art, while also achieving comparable image performance in arbitrary-scale SR tasks. + +Our contributions are as follows: + +- We propose a novel position encoding, called orthogonal position encoding (OPE), which takes the form of a 2D-Fourier series and corresponds to a set of orthogonal basis. Building on OPE, we introduce the OPE-Upscale Module, a parameter-free upsampling module for arbitrary-scale image super-resolution. +- Our method significantly reduces the consumption of computing resources, resulting in high computing efficiency for arbitrary-scale SR tasks. +- The OPE-Upscale Module is interpretable, parameter-free and does not require training, resulting in a concise SR framework that elegantly solves the flipping consistency problem. +- Extensive experiments demonstrate that our method achieves comparable results with the state-of-the-art. Furthermore, our method enables super-resolution up to a large scale of $\times 30$ . + +# 2. Related Work + +# 2.1. Sinusoidal Positional Encoding + +Sinusoidal positional encoding is widely used to counteract the negative effects of token order and sequence length in sequence models [65], or to guide image generation as the spatial inductive bias in CNNs [7, 23, 36]. In implicit neural representations, it plays a critical role in solving spectral bias [50]. By embedding input coordinates into a higher dimensional space, position encoding greatly improves high-frequency representation of implicit 3D scenes [42, 56] and subsequent works take it as the default operation to improve representation quality [37, 43, 53, 78]. Inspired by these works, positional encoding has been preliminarily explored in representing images in arbitrary-scale SR [32, 72]. + +![](images/431e70fe5e15799c2fd2cfcf81a85fecf1bfe8720acc6cede96f95242475aa4f.jpg) +LIIF:32.35 +LIIF-flip: 28.47 + +![](images/2c5698403b7ce46cbf5dbe6eb0580985fcc06df8f1c84856cb42c264a35e73da.jpg) +LIIF:25.43 +LIIF-flip: 23.11 + +![](images/9a15d2c0a4f1a1b7a11386054ae16ec18bd0c0b1de4f758e9301086228311523.jpg) +OPE: 32.37 +OPE-flip: 32.37 +Figure 1. Flipping consistency decline (PSNR (dB)). LIIF-flip: Flipping the input of the LIIF [5] decoder yields blurred results in the symmetric outputs. OPE-flip: Our method does not exhibit such artifacts. Additional results can be found in the supplementary material. + +![](images/5386d6e74192c053b22db82f332fd7d30e29ed161cc485d8585e5143e92c53d1.jpg) +OPE:25.50 +OPE-flip: 25.50 + +# 2.2. Orthogonal Basis Decomposition + +In functional analysis, an orthogonal basis decomposition is a way to represent a vector or function as a linear combination of orthogonal basis functions. Wavelet transform [2, 15, 20, 38] and 2D-Fourier transform [6, 12, 13, 22, 74] are commonly used decomposition techniques for images and videos. In DSGAN [12], the input image is explicitly decomposed into low and high frequencies using high-pass and low-pass filters. Other methods use frequency domain losses to decompose images, either in a supervised manner [22] or in an unsupervised manner [13]. To address with resolution discrepancy of reconstructed images and input images, Rippel et al. [52] employ spectral pooling to decrease resolution by truncating in the Fourier domain, while Zhou et al. [81] explore an up-sampling method in the Fourier domain to increase the resolution. Image moments, which decompose images into two-dimensional orthogonal polynomials [27, 82], are widely used in invariant pattern recognition [30, 77]. Image sparse representation inherits this decomposition idea and performs well in traditional computer vision tasks [39, 71, 73]. In 3D domains, spherical harmonics serve as an orthogonal basis in space to represent view dependence [3, 51, 59] and have recently been proposed as a replacement for MLPs [11] for representing neural radiance fields [42]. + +# 2.3. Deep Learning-based SR + +Based on the upsampling operations and their location in the model, deep learning-based SR frameworks can be classified into four categories (see [70] for a comprehensive survey): pre-upsampling [8,24,25,55,61,62], post-upsampling [9,31,35,64,79], progressive-upsampling [28,29,68], and iterative up-and-down sampling [16, 33, 69]. With preupsampling, the LR image is first upsampled by traditional interpolation and then fed into a deep CNN to reconstruct + +high-quality details. While it was one of the most popular frameworks for arbitrary-scale factors, it has side effects like enlarged noise by interpolation and high time complexity and space consumption. The progressive-upsampling and iterative up-and-down sampling frameworks pose challenges in terms of complicated model designing and unclear design criteria, as noted in [70]. For post-upsampling, the LR image is directly fed as input to a deep CNN, and then a trainable upsampling module (e.g., deconvolution [9], subpixel [54], and interpolation convolution [10]) increases the resolution at the end. Since feature extraction process, which is computationally intensive, only occurs in low-dimensional space, it has become one of the mainstream frameworks [32, 34, 67]. + +# 2.4. Arbitrary-scale SR + +In the field of arbitrary-scale SR, most existing works are based on the post-upsampling framework and replace the traditional upsampling module with an INR-based one, such as a coordinate-based MLP. Meta-SR [18] was the first arbitrary-scale SR method based on CNN. ArbSR [66] adopts a general plug-in module to solve the scaling problem of different horizontal and vertical scales. SRWarp [60] transforms LR images into HR images with arbitrary shapes via a differential adaptive warping layer. SphereSR [75] explores arbitrary-scale on $360^{\circ}$ images. LIIF [5] uses coordinates and conditional latent code into an MLP to directly predict target pixel color with an intuitive network structure. LIIF-related follow-up works focus on predicting high-frequency information with position encoding [32,72]. + +# 3. Method + +# 3.1. OPE-based Image Representation + +Given an LR image $I_{LR}$ with resolution of $h \times w$ , we decide its 2D domain into $h \times w$ grids, where each grid represents a pixel in the LR image and corresponds to a patch of the high resolution image of size $r_h \times r_w$ . The output image $I_{SR}$ has a resolution of $H \times W$ , where $H = r_h \cdot h$ and $W = r_w \cdot w$ . Denote by $I \in \mathbb{R}^{r_h \times r_w \times 1}$ the high resolution image patch with size $r_h \times r_w$ for a specific color channel. We view each pixel in $I$ as a sample of a continuous bivariate function $f(x,y): [-1,1] \times [-1,1] \to \mathbb{R}$ . We embed the coordinates $x$ and $y$ using sinusoidal positional encoding + +$$ +\begin{array}{l} { X } { = \gamma ( x ) = \sqrt { 2 } \cdot [ \frac { 1 } { \sqrt { 2 } } , \cos ( \pi x ) , \sin ( \pi x ) , \cos ( 2 \pi x ) , } \\ \sin (2 \pi x), \dots , \cos (n \pi x), \sin (n \pi x) ] \tag {1} \\ \end{array} +$$ + +$$ +\begin{array}{l} { Y } { = \gamma ( y ) = \sqrt { 2 } \cdot \left[ \frac { 1 } { \sqrt { 2 } } , \cos ( \pi y ) , \sin ( \pi y ) , \cos ( 2 \pi y ) , \right. } \\ \left. \sin (2 \pi y), \dots , \cos (n \pi y), \sin (n \pi y) \right], \tag {2} \\ \end{array} +$$ + +where $\gamma (\cdot):\mathbb{R}\to \mathbb{R}^{1\times (2n + 1)}$ is a univariate function for position encoding with a predefined maximum frequency + +$n\in \mathbb{N}$ . We flatten the matrix $X^T Y$ into a row vector $P\in$ $\mathbb{R}^{1\times (2n + 1)^2}$ as + +$$ +P = \operatorname {f l a t} \left(X ^ {T} Y\right), \tag {3} +$$ + +where flat: $\mathbb{R}^{(2n + 1)\times (2n + 1)}\to \mathbb{R}^{1\times (2n + 1)^2}$ is the flattening operation. Denote by $e_{i,j}$ the element on the $i$ -th row and the $j$ -th column of matrix $X^T Y$ . For example. $e_{4,5} = 2\cos (2\pi x)\sin (2\pi y)$ . It is easy to verify that + +$$ +\left\langle e _ {i _ {1}, j _ {1}}, e _ {i _ {2}, j _ {2}} \right\rangle = \left\{ \begin{array}{l l} 0, & (i _ {1}, j _ {1}) \neq (i _ {2}, j _ {2}) \\ 1, & (i _ {1}, j _ {1}) = (i _ {2}, j _ {2}) \end{array} \right. \tag {4} +$$ + +where $\langle ,\rangle$ is the inner product in function space, i.e., + +$$ +\langle g, h \rangle = \frac {1}{4} \int_ {- 1} ^ {1} \int_ {- 1} ^ {1} g (x, y) h (x, y) \mathrm {d} x \mathrm {d} y. +$$ + +Therefore the elements $\{e_{i,j}\}$ form a set of orthogonal basis, allowing us to approximate $f$ as a linear combination. Consider a pixel $(x,y)$ in the upsampled image patch $I$ , which corresponds to a region $[x - 1 / r_h, x + 1 / r_h] \times [y - 1 / r_w, y + 1 / r_w]$ . We use $f(x,y)$ as a representative for the entire region and compute $I_{(x,y)}$ as + +$$ +I _ {(x, y)} \triangleq f (x, y) \approx Z P ^ {T}, \tag {5} +$$ + +where $Z \in \mathbb{R}^{1 \times (2n + 1)^2}$ represents the projection. Due to the orthogonal property, we call the resulting vector $P$ as orthogonal position encoding (OPE). Fig. 3 illustrates the concept of OPE-based patch representation. + +Remark 1. Our OPE basis can be seen as the real form version of the 2D-Fourier basis, which eliminates the complex exponential term based on conjugate symmetry when representing real signals. See the supplementary material for details. + +Remark 2. OPE differs from the commonly used positional encoding formulation, such as that in [42], in that it includes a constant term and takes the product of each coordinate embedding as a new term. These seemingly minor changes indeed have a deep impact. Conventional positional encoding processes the coordinates separately, thereby encoding frequencies in the horizontal and vertical directions only. In contrast, OPE includes frequencies covering in all directions of the plane due to the product terms, resulting in a better expression capability. + +Remark 3. Unlike LIIF [5] that uses an MLP to approximate the image function $f$ , our method completely eliminates the need of MLP. + +# 3.2. OPE-Upscale Module + +We project the latent code onto the OPE basis. OPE with a sufficiently long latent code could represent an image directly in a continuous manner. However, it suffers from long embedding time and is unstable for representing high-frequency details locally, similar to the limitations + +![](images/315202109cbd5ab2b2ae09d337dd9a1ba4711b5da23cf976bd21202fadfd71de.jpg) +Figure 2. OPE-Upscale Module for arbitrary-scale SR. The encoder $E_{\theta}$ is the only trainable part. With a pre-defined maximum frequency $n$ of OPE, the OPE-Upscale Module (shaded in grey) takes the feature map from $E_{\theta}$ and the target resolution $H$ , $W$ as input, and renders the pixels of the target SR image in parallel. $\odot$ is the matmul function that returns the product of $z^{*} \in \mathbb{R}^{1 \times 3(2n + 1)^{2}}$ and OPEs $\in \mathbb{R}^{3(2n + 1)^{2} \times 1}$ per color channel. + +![](images/d9549fbd46fbe99071b6b32e939af388d83fef11d090526cca8a42697b81241b.jpg) +Figure 3. OPE-based patch representation. An image patch can be represented as a linear combination of basic plane waves in a continuous manner. The representation is a 2D extension to [46, Chap. 3.3.1]. Refer to the supplementary material for more details. + +of Fourier transform to describe local information. To address this issue, we represent the input image as the seamless stitching of local patches, whose latent codes are extracted from a feature map over the channel dimension2. As shown in Fig. 2, the OPE-upscale module takes both the target resolution $H = r_h \cdot h$ , $W = r_w \cdot w$ and the feature map $\in \mathbb{R}^{h \times w \times 3(2n + 1)^2}$ from the deep encoder $E_{\theta}$ as inputs and computes target pixels in parallel. + +Feature map rendering. As shown in Fig. 4, to render a target image $I_{SR}$ with size $H \times W$ from a LR image $I_{LR}$ with size $h \times w$ , OPE-Upscale Module firstly divide a 2D domain $[-1,1] \times [-1,1]$ into $H \times W$ regions with equal size, so that every pixel in $I_{SR}$ will be associated with an absolute central coordinates $(x_q, y_q)$ in corresponding region. Secondly, the latent codes in the feature map (same dimension with $I_{LR}$ ) also possess corresponding central coordinates $(x_c, y_c) \in [-1,1] \times [-1,1]$ by dividing same 2D domain into $h \times w$ regions, therefore, given a target image pixel with $(x_q, y_q)$ , a specific latent code $z^* \in \mathbb{R}^{1 \times 3(2n + 1)^2}$ with + +![](images/6f0579f2f542a32044fcfddc79d9a66ba99c8e42ec8c5e9feb5df76dd793ddb2.jpg) +Figure 4. Illustration of the mapping from the input LR image to the output SR image. LR image, feature map and SR image are all divided into the same domain $[-1,1] \times [-1,1]$ . Each LR pixel $I_{LR}(x_{LR},y_{LR})$ corresponds to a latent code with coordinates $(x_c,y_c)$ , while each SR pixel $I_{SR}(x_{SR},y_{SR})$ corresponds to a latent code with coordinates $(x_q,y_q)$ . + +coordinates $(x_{c},y_{c})$ , which has the smallest distance from $(x_{q},y_{q})$ could be found. As shown in Eq.(6) and Eq.(7), a render function $\mathcal{R}$ takes two parts of inputs: $z^{*}$ and $(x_q',y_q')$ , to calculate final target pixel value as following: + +$$ +I _ {S R} \left(x _ {q}, y _ {q}\right) = \mathcal {R} \left(z ^ {*}, \left(x _ {q} ^ {\prime}, y _ {q} ^ {\prime}\right)\right) \tag {6} +$$ + +$$ +x _ {q} ^ {\prime} = \left(x _ {q} - x _ {c}\right) \cdot h, \quad y _ {q} ^ {\prime} = \left(y _ {q} - y _ {c}\right) \cdot w \tag {7} +$$ + +where $z^{*}$ is the nearest latent code we found, and $(x_{q}^{\prime}, y_{q}^{\prime})$ are relative coordinates, which are calculated based on Eq.(7) to rescale the absolute coordinates (in domain $[-1, 1] \times [-1, 1]$ ) by times $h$ and $w$ , which is taken as input by function $\mathcal{R}$ to render target pixel. $\mathcal{R}$ has the similar calculation as Eq.(5) while the difference is it repeats OPE three times to adapt $z^{*}$ and calculate the linear combination per color channel. In this way, our OPE-upscale module successfully deals with arbitrary size $I_{SR}$ by processing each pixel by $\mathcal{R}$ , in which feature map rendering process is parameter-free with high computing efficiency and less memory consumption (which has been confirmed by the experiments in Sec. 4.3). + +Patch ensemble. As shown in Fig. 4, when moving $(x_{q},y_{q})$ from location $a$ to $b$ , the pixel value in the target image $I_{SR}$ may change abruptly, as well as the nearest latent code $z^{*}$ . To address this discontinuity issue, we propose a patch ensemble, which is a local ensemble styled interpolation using relative coordinates. Instead of finding a single nearest latent code for $(x_{q},y_{q})$ , we select the nearest four neighbouring latent codes $z_{t}^{*}$ with corresponding central coordinates $(x_{t},y_{t})$ , where $t\in \{00,01,10,11\}$ . Then we calculate the relative coordinates $x_{q}^{\prime}$ and $y_{q}^{\prime}$ as + +$$ +x _ {q} ^ {\prime} = \frac {\left(x _ {q} - x _ {t}\right) \cdot h}{2}, \quad y _ {q} ^ {\prime} = \frac {\left(y _ {q} - y _ {t}\right) \cdot w}{2}, \tag {8} +$$ + +which can guarantee $x_{q}^{\prime}, y_{q}^{\prime} \in [-1, 1]$ . As the inset shows, for the four adjacent pixels from the low resolution image $I_{LR}$ (i.e. related to latent codes $z_{00}^{*}, z_{01}^{*}, z_{10}^{*}$ and $z_{11}^{*}$ ), + +their corresponding patches in super resolution image $I_{SR}$ are colored in red, green, yellow and blue, respectively. The pixel color in $I_{SR}$ is not solely dependent on the nearest latent code but considers the four neighboring latent codes. Specifically, using the rendering function $\mathcal{R}$ and the di + +![](images/ed462df17db4a6c8995f53e1017ef18600251ebc53569198c6eeee7aa186f906.jpg) + +aagonal rectangle areas $s_t$ as weights, we compute the pixel value as a weighted sum + +$$ +I _ {S R} \left(x _ {q}, y _ {q}\right) = \sum_ {t \in \{0 0, 0 1, 1 0, 1 1 \}} \frac {s _ {t}}{S} \cdot \mathcal {R} \left(z _ {t} ^ {*}, \left(x _ {q} ^ {\prime}, y _ {q} ^ {\prime}\right)\right), \tag {9} +$$ + +where $S = \sum_{t} s_{t}$ is the sum of areas. Considering the contribution of each latent code allows us to integrate the adjacent latent codes with different significance, thereby providing a seamless stitching of adjacent patches. We call Eq.(9) local ensemble styled interpolation since it takes a similar form of the local ensemble in LIIF [5]. + +# 3.3. Maximum Frequency $n$ + +Selecting a proper maximum frequency $n$ plays an important role in designing and implementing the OPE-upsampling module since it directly determines the network architecture and also has effects on the performance of different SR scales. Given a high resolution image $I_{HR}$ with size $H \times W$ , $n$ and $r$ , we aim to obtain a feature map with size $\frac{H}{r} \times \frac{W}{r}$ , then we re-render the obtained feature map with the selected $n$ . By the comparison of the rendered $I_{SR}$ , we present the performance of $n \in \{1,2,\dots,8\}$ under different $r$ values (SR scale), as show in Tab. 1, and select the $n$ with the best performance (the details would be discussed Sec. 4.1). To be specific, we use Eq.(10) as the basic theory and use Eq.(11) to infer the feature map. First, similar to calculate the projection of a normal vector on orthogonal basis, we can calculate projections (or so-called latent code) $Z \in \mathbb{R}^{1 \times (2n + 1)^2}$ of $f(x,y)$ in Eq.(5) as follows: + +$$ +Z [ i ] = \frac {1}{4} \int_ {- 1} ^ {1} \int_ {- 1} ^ {1} f (x, y) P [ i ] (x, y) \mathrm {d} x \mathrm {d} y \tag {10} +$$ + +where $P[i](x,y)$ is a bivariate function taken from the $i$ -th position of OPE and $Z[i]$ is the corresponding projection. Based on Eq.(10) and taking both the discreteness of an image and the design of OPE-Upscale Module into consideration, we calculate the feature map of an image $I_{HR}$ with down-sampling scale $r$ as follows: + +$$ +z ^ {*} [ i ] = \frac {1}{4} \sum_ {x ^ {\prime}} ^ {2 r} \sum_ {y ^ {\prime}} ^ {2 r} I _ {H R} \left(x ^ {\prime}, y ^ {\prime}\right) P [ i ] \left(x ^ {\prime}, y ^ {\prime}\right) \tag {11} +$$ + +It can be considered as the inverse operation of Eq.(9). + +Take the right inset as an example. We choose the HR image as the ground truth (e.g. HR in Fig. 5), when $r = 4$ , every latent code $z^{*}$ corresponds to a $8 \times 8$ patch of HR (gray points) in relative coordinate domain (blue region). To calculate the $i$ -th position of $z^{*}$ , we multiply every HR pixel value + +![](images/c3e0f18e5283550c5f3e197ca6920b72751ad2be79e0e5b1afd13585496f518a.jpg) + +$I_{HR}(x',y')$ and basis value $P[i](x',y')$ together and finally sum them. After getting the feature map, we render it to the same size of $I_{HR}$ via OPE-Upscale Module and calculate their Peak Signal-to-Noise Ratio (PSNR). + +# 4. Experiments + +# 4.1. Parameter Setting + +We evaluated the performance of different maximum frequencies $n$ on 50 images from the DIV2K validation set [1] under different scales $r$ . Since our method is a local representation, we do not use a large $n$ . As shown in Tab. 1, the optimal sampling frequency for a given $r_i$ is always $r_i - 1$ . This observation can also be explained by the Nyquist-Shannon sampling theorem. For example, when $r_i = 4$ , there are $8 \times 8$ sampling points for every latent code to "fit", so the maximum frequency that can be recovered from these sampling points should be less than 4. We also tested larger frequency with $r_i \leq n \leq 2 \times r_i$ , reaching the upper limit that equals the number of sampling points. We further visualize the reconstructed images for different $n$ . As shown in Fig. 5, with scale factor $\times 4$ , the larger frequency ( $n > 3$ ) brings redundant high-frequency information that sharpens the resulting images. + +Based on the above analysis, we decide to choose $n = 3$ as the maximum frequency for our OPE-Upscale Module. This is because existing arbitrary-scale SR methods, such as LIIF [5] and LTE [32], are trained with random scale factors up to 4, and the frequency $n = 4 - 1 = 3$ is sufficient to fully capture the ground truth information. Al- + +
n×2×3×4×5×6×7×8
131.195128.608326.442425.048524.111423.342322.7898
230.747233.658631.209128.802227.370126.191325.3838
322.187133.658535.198332.401130.613528.896427.8159
412.123028.608334.963134.629434.097931.446230.2865
5-22.846529.951234.629337.370433.728532.8190
6-22.846524.312232.401137.150635.304635.8250
7--19.159328.802233.303935.304639.1160
8--12.086325.048529.096633.728638.9863
+ +Table 1. Representation performance (PSNR (dB)). The best value for each upsampling factor is bolded. + +![](images/1ed6aacf9d28f0b5fd6221f6ccc37ab98dd42e8b9689c7eff37379a229040c83.jpg) +Figure 5. Qualitative comparison of different OPE frequency n under scale factor $\times 4$ (PSNR (dB)). 1-th row: residuals from $n = 0$ in image time domain. 2-th row: fourier frequency domain of HR and rendered image with $n$ . 3-th row: HR image and rendered image with $n$ . + +though a larger $n$ could represent more detailed patches of the target SR image, it would also introduce redundant high-frequency information and potentially increase computation time and memory consumption during training. Therefore, we choose $n = 3$ as the maximum frequency for our OPE-Upscale Module to balance performance and efficiency. + +# 4.2. Training + +Datasets. Similar to [5, 32], we use the DIV2K dataset [1] of the NTIRE 2017 Challenge [63] for training. For testing, we use the DIV2K validation set [1] with 100 images and four benchmark datasets: Set5 [4], Set14 [76], B100 [40], and Urban100 [21]. We use PSNR as the quality measure. + +Implementation details. We mainly follow the prior implementation [5, 32] for arbitrary-scale SR training after replacing their upsampling module with OPE-Upscale. We use EDSR-baseline [35] and RDN [80] without their upsampling modules as the encoder, which is the only trainable part of our network. We use $48 \times 48$ patches cropped from training set as inputs, L1 loss and Adam [26] optimizer for optimization. The network was trained for 1000 epochs with batch size 16, while the initial learning rate is 1e-4 and decayed by factor 0.5 every 200 epochs. More implementation details are presented in supplementary material. + +# 4.3. Evaluation + +Quantitative results. Tab. 2 and Tab. 3 report quantitative results of OPE and the SOTA arbitrary-scale SR methods on the DIV2K validation set and the benchmark datasets. It is worth noting that we focus on finding an alternative of MLP with position encoding, rather than enhancing it like LTE [32]. We observe that our method achieves comparable results (less than 0.1dB on DIV2K and less than 0.15dB on benchmark), which indicates that our method is a feasible analytical solution with good performance and efficient parameter-free module. As shown in Tab. 2, EDSR [35] and RDN [80] are our selected encoders, and we achieve the highest efficiency (i.e. the shortest inference time in red number) comparing to all the other baselines with both encoders. The higher the scale factor, the better result we achieve. Specifically, in out-scale SR ( $\times 6$ to $\times 30$ ), our method outperforms most baselines and just has a small gap with LTE (less than 0.1dB). Such results demonstrate that our method has rich representation capability. We also compared with the benchmark dataset, as shown in Tab. 3, we keep comparable results to baselines (the gap is less than 0.15dB). However, as a nonlinear representation method, MLP still has advantages over our linear representation with low scale factors. See Sec. 5 for discussion on this issue. + +Qualitative results. Fig. 6 provides qualitative results with SOTA methods by using different scale factor. We show competitive visual quality against others, more results are provided in supplementary material. From the local perspective, LIIF and LTE only generate smooth patches, while our OPE with max frequency 3 is enough to achieve similar visual quality. We also notice LIIF [5] has artifact (vertical stripes) in the 1st row, this is a common drawback for implicit neural representation and is hard to be explained. However, with our image representation, there is no artifacts. In the 2nd row, we could observe a sprout (in red rectangle) in the GT, the same region of LIIF is vanished, and the boundary of our sprout is more obvious than LTE. + +Computing efficiency. We measure computing efficiency with MACs (multiply-accumulate operations), FLOPs (floating point operations) and actual running time. In Tab. 4 column 2-3, judged by the time complexity measured by the number of operations, we save 2 orders of magnitude. In our upsampling module, there is only one matrix operation and essential position encoding between input and output. In Tab. 2 we show shortest inference time benefiting from our compact SR framework. To further demonstrate our time advantage on large size images, we take $256 \times 256$ as LR input of encoder and calculate time consumption of upsampling module with scale factor $\times 4 \times 30$ on NVIDIA RTX 3090. As shown in Tab. 5, our upsampling module shows $26\% - 57\%$ time advantage, this advantage keeps growing with larger scale factor. Notice We do not take advantage of GPU acceleration to design the upsampling + +![](images/aac223839dec03f1e84574ef5e7b063b56f045530bb6f881d9a6206fcde2bbd2.jpg) +Figure 6. Qualitative comparison with SOTA methods for arbitrary-scale SR. RDN [80] is used as encoder for all methods. + +
MethodIn-scaleOut-scale
×2×3×4×6×12×18×24×30
Bicubic [35]31.0128.2226.6624.8222.2721.0020.1919.59
EDSR-baseline [35]34.5530.9028.94-----
EDSR-baseline-MetaSR# [5, 18]34.6430.9328.9226.6123.5522.0321.0620.37
EDSR-baseline-LIIF [5]34.67 / 170230.96 / 127729.00 / 114426.75 / 104623.71 / 96522.17 / 95321.18 / 95120.48 / 947
EDSR-baseline-LTE [32]34.72 / 115831.02 / 107929.04 / 104526.81 / 102323.78 / 100722.23 / 100521.24 / 100320.53 / 1000
EDSR-baseline-OPE (ours)34.34 / 47630.94 / 39529.02 / 36426.77 / 34823.74 / 32222.21 / 31821.21 / 31420.52 / 311
RDN-MetaSR# [5, 18]35.0031.2729.2526.8823.7322.1821.1720.47
RDN-LIIF [5]34.99 / 310731.26 / 207329.27 / 151326.99 / 124823.89 / 102522.34 / 99421.31 / 99120.59 / 972
RDN-LTE [32]35.04 / 254931.32 / 183929.33 / 142027.04 / 118423.95 / 104922.40 / 102721.36 / 102520.64 / 1014
RDN-OPE (ours)34.52 / 227731.17 / 149729.26 / 103926.98 / 81323.91 / 66322.36 / 62321.34 / 59620.63 / 590
+ +Table 2. Quantitative comparison with the SOTA methods for arbitrary-scale SR on the DIV2K validation set (PSNR (dB) / running time (ms per image)). $\sharp$ indicates implementation in LIIF [5]. With a parameter-free upsampling module, we narrow the gap between SOTA and ours in most results less than 0.1dB (blue) and obtain the shortest inference time (red). + +module carefully, with hardware optimization, we believe our time advantage could be much larger thanks to fewer number of operations required. + +Memory consumption. In Tab. 4 column 4-5 we compare GPU memory consumption of OPE-Upscale Module with LIIF [5] and LTE [32] under training mode and testing mode of Pytorch [48]. For training mode, we use a $48 \times 48$ patch as input and sample 2304 pixels as output following the default training strategy in arbitrary-scale SR works. For testing mode, we use $512 \times 512$ image as input with scale factor 4. As a interpretable image representation without network parameters, OPE-Upscale Module saves memory of intermediate data (e.g. gradients, hidden layer outputs), and this advantage is fully reflected in training mode. + +Flipping consistency. As described in Sec. 1, the INR-based upsampling module like [5] is sensitive for the flipping of feature map. However, our method solves this problem completely and elegantly. The orthogonal basis of OPE is based on symmetric sinusoidal function, which leads to advantage of our method for keeping the flipping consistency. Also, more samples are provided in supplementary material for verifying other more flipping transforms. + +![](images/f8bf4b45e4adbc2597c95a129188961df8edd51595c6d51117cc75e7a18d9586.jpg) +Figure 7. Visual results of the ablation study on patch ensemble. See the text for details. + +# 4.4. Ablation Studies + +In order to examine the effects of local ensemble styled interpolation (I)-Eq.(9) and extension of relative coordinate domain (E)-Eq.(8) in patch ensemble, we conducted experiments using four different settings with EDSR-baseline as the encoder. The four settings were: 1) OPE: OPE-Upscale module with I and E; 2) OPE-E: OPE-Upscale module without E but with I; 3) OPE-I: OPE-Upscale module without I but with E; and 4) OPE-IE: OPE-Upscale module with neither I nor E (that is, without patch ensemble). + +Fig. 7 and Tab. 6 present the comparison results of different settings. In the OPE-E setting, only the nearest latent code to the query point can be rescaled to $[-1,1] \times [-1,1]$ , while the relative coordinates of the other three latent codes + +
MethodSet5Set14B100Urban100
In-scaleOut-scaleIn-scaleOut-scaleOut-scaleIn-scaleOut-scale
×2×3×4×6×8×2×3×4×6×8×2×3×4×6×8×2×3
RDN [80]38.2434.7132.47--34.0130.5728.81--32.3429.2627.72--32.8928.80
RDN-MetaSR# [5, 18]38.2234.6332.3829.0426.9633.9830.5428.7826.5124.9732.3329.2627.7125.9024.8332.9228.82
RDN-LIIF [5]38.1734.6832.5029.1527.1433.9730.5328.8026.6425.1532.3229.2627.7425.9824.9132.8728.82
RDN-LTE [32]38.2334.7232.6129.3227.2634.0930.5828.8826.7125.1632.3629.3027.7726.0124.9533.0428.97
RDN-OPE (ours)37.6034.5932.4729.1727.2233.3930.4928.8026.6525.1732.0529.1927.7225.9624.9131.7828.63
+ +Table 3. Quantitative comparison with SOTA methods for arbitrary-scale image SR on benchmark datasets (PSNR (dB)). $\sharp$ indicates implementation in LIIF [5]. We narrow the gap between SOTA and ours in most results less than $0.15\mathrm{dB}$ (blue number). For large scale factor, we keep comparable results to MetaSR [18] and LIIF [5]. The defect in low scale factor will be analysed in Sec. 5. + +
MethodParamsMACsFLOPsMem (training)Mem (Test)
LIIF0.35 M429 K6.2 G85.1 + 1.9 M32 + 96 M
LTE0.26 M526 K7.5 G97.8 + 1.9 M64 + 96 M
OPE (ours)0 M6 K85 M0 + 1.9 M0 + 96 M
+ +Table 4. Parameter number, time complexity and memory consumption. MACs: multiply-accumulate operations, FLOPs: floating point operations, Mem: intermediate data + essential output for GPU memory consumption. We use $n = 3$ as maximum frequency of OPE and test in training mode and test mode on Pytorch with tool: torch.cuda.memory_allocated(). Training mode: $48^2$ to 2304 pixels, test mode: $512^2$ to $2048^2$ . + +
Method×4×8×12×16×20×24×30
LIIF382152135306004102741835027866
LTE376149033405922102681834027838
OPE2771125249537195673836612012
Percentage28%26%30%39%45%55%57%
+ +Table 5. Rendering time of upsampling module (ms per image) with an input resolution of $256 \times 256$ . The last row shows the time saving percentage achieved by our method. We use $n = 3$ as the maximum frequency of OPE. Our method provides a time advantage, which increases as the rendering resolution increases. On average, we achieve a $40\%$ reduction in rendering time. + +
In-scaleOut-scale
×2×3×4×6×8
OPE33.2930.2928.6526.4624.98
OPE-E33.2730.2328.5626.3424.82
OPE-I33.2830.2628.6326.4424.97
OPE-IE33.2030.0928.4426.2524.70
+ +Table 6. Ablation studies on Set14. EDSR-baseline [35] is used as encoder. + +cannot be rescaled, resulting in periodic stripes in the resulting SR image. On the other hand, the OPE-I result shows no obvious discontinuity between patches since the extension plays a positive role, but this means that only a small region of the patch is presented in the target image. Lastly, the OPE-IE setting shows an obvious thick boundary between patches, indicating that both extension and interpolation are necessary for the best performance. + +# 5. Discussions + +We observed that our quantitative results decreases at low scale factors, especially when the input size is small, + +such as with benchmark datasets. See Tab. 3 and Tab. 2. This is due to the fact that a smaller target size $(W\times H)$ leads to a larger grid in the 2D domain $\left([-1,1]\times [-1,1]\right)$ , where utilizing only one central point value to represent the entire larger grid would result in a loss of detailed information compared to a smaller grid. Since the 2D domain we use is continuous, the higher the resolution of the target image, the stronger the representation ability we can achieve. MLP-based representations, such as [5] and [32], can overcome this issue through nonlinear operations. In our method, this defect can be ignored for high SR scale factors where pixels are dense, but for low scale such as $\times 2$ or $\times 3$ , our performance may slightly degrade. A possible way to address this issue is to sample more points for every grid region and calculate their mean value, with careful consideration of the time consumption trade-off. + +# 6. Conclusion + +In this paper, we proposed an interpretable method for continuous image representation without implicit neural networks. Our method leverages a novel position encoding technique called orthogonal position encoding, which takes the form a 2D-Fourier series and corresponds to 2D image coordinates. As a set of orthogonal basis in inner product space, OPE is both interpretable and rich in representation. Building on OPE, we introduced the OPE-Upscale Module, a parameter-free approach for arbitrary-scale image super-resolution that simplifies the existing deep SR framework, leading to high computing efficiency and less memory consumption. Our OPE-Upscale Module can be easily integrated into existing image super-resolution pipelines, and extensive experiments demonstrate that our method achieves competitive results with the state-of-the-art. + +It is worth noting that the overall efficiency of SR framework depends on both the encoder and decoder. Since our work focuses on decoder design and efficiency, we leave the development of high-efficient encoder as future work. + +Acknowledgement This project was partially supported by NSFC Grants (61702363, 62222311, 62201274) and the Ministry of Education, Singapore, under its Academic Research Fund Grants (MOE-T2EP20220-0005 & RG20/20). + +# References + +[1] Eirikur Agustsson and Radu Timofte. Ntire 2017 challenge on single image super-resolution: Dataset and study. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 126-135, 2017. 5, 6 +[2] Woong Bae, Jaejun Yoo, and Jong Chul Ye. Beyond deep residual learning for image restoration: Persistent homology-guided manifold simplification. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 145-153, 2017. 2 +[3] Ronen Basri and David W Jacobs. Lambertian reflectance and linear subspaces. IEEE transactions on pattern analysis and machine intelligence, 25(2):218-233, 2003. 2 +[4] Marco Bevilacqua, Aline Roumy, Christine Guillemot, and Marie Line Alberi-Morel. Low-complexity single-image super-resolution based on nonnegative neighbor embedding. 2012. 6 +[5] Yinbo Chen, Sifei Liu, and Xiaolong Wang. Learning continuous image representation with local implicit image function. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8628-8638, 2021. 1, 2, 3, 5, 6, 7, 8 +[6] Lu Chi, Borui Jiang, and Yadong Mu. Fast fourier convolution. Advances in Neural Information Processing Systems, 33:4479-4488, 2020. 2 +[7] Jooyoung Choi, Jungbeom Lee, Yonghyun Jeong, and Sungroh Yoon. Toward spatially unbiased generative models. arXiv preprint arXiv:2108.01285, 2021. 2 +[8] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Learning a deep convolutional network for image super-resolution. In European conference on computer vision, pages 184-199. Springer, 2014. 2 +[9] Chao Dong, Chen Change Loy, and Xiaou Tang. Accelerating the super-resolution convolutional neural network. In European conference on computer vision, pages 391-407. Springer, 2016. 2, 3 +[10] Vincent Dumoulin, Jonathon Shlens, and Manjunath Kudur. A learned representation for artistic style. arXiv preprint arXiv:1610.07629, 2016. 3 +[11] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5501–5510, 2022. 2 +[12] Manuel Fritsche, Shuhang Gu, and Radu Timofte. Frequency separation for real-world super-resolution. In 2019 IEEE/CVF International Conference on Computer Vision Workshop (ICCVW), pages 3599-3608. IEEE, 2019. 2 +[13] Dario Fuoli, Luc Van Gool, and Radu Timofte. Fourier space losses for efficient perceptual image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2360-2369, 2021. 2 +[14] Kyle Genova, Forrester Cole, Avneesh Sud, Aaron Sarna, and Thomas Funkhouser. Local deep implicit functions for 3d shape. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4857-4866, 2020. 1 + +[15] Tiantong Guo, Hojjat Seyed Mousavi, Tiep Huu Vu, and Vishal Monga. Deep wavelet prediction for image superresolution. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 104-113, 2017. 2 +[16] Muhammad Haris, Gregory Shakhnarovich, and Norimichi Ukita. Deep back-projection networks for super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1664-1673, 2018. 2 +[17] Kurt Hornik, Maxwell Stinchcombe, and Halbert White. Multilayer feedforward networks are universal approximators. Neural networks, 2(5):359-366, 1989. 1 +[18] Xuecai Hu, Haoyuan Mu, Xiangyu Zhang, Zilei Wang, Tieniu Tan, and Jian Sun. Meta-sr: A magnification-arbitrary network for super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1575-1584, 2019. 1, 3, 7, 8 +[19] Binbin Huang, Xinhao Yan, Anpei Chen, Shenghua Gao, and Jingyi Yu. Pref: Phasorial embedding fields for compact neural representations. arXiv preprint arXiv:2205.13524, 2022. 1 +[20] Huaibo Huang, Ran He, Zhenan Sun, and Tieniu Tan. Wavelet-srnet: A wavelet-based cnn for multi-scale face super resolution. In Proceedings of the IEEE International Conference on Computer Vision, pages 1689-1697, 2017. 2 +[21] Jia-Bin Huang, Abhishek Singh, and Narendra Ahuja. Single image super-resolution from transformed self-exemplars. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5197-5206, 2015. 6 +[22] Liming Jiang, Bo Dai, Wayne Wu, and Chen Change Loy. Focal frequency loss for image reconstruction and synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13919-13929, 2021. 2 +[23] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. Advances in Neural Information Processing Systems, 34:852-863, 2021. 2 +[24] Jiwon Kim, Jung Kwon Lee, and Kyoung Mu Lee. Accurate image super-resolution using very deep convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1646-1654, 2016. 2 +[25] Jiwon Kim, Jung Kwon Lee, and Kyoung Mu Lee. Deeplyrecursive convolutional network for image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1637-1645, 2016. 2 +[26] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6 +[27] Tom Koornwinder. Two-variable analogues of the classical orthogonal polynomials. In Theory and application of special functions, pages 435-495. Elsevier, 1975. 2 +[28] Wei-Sheng Lai, Jia-Bin Huang, Narendra Ahuja, and Ming-Hsuan Yang. Deep laplacian pyramid networks for fast and accurate super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 624-632, 2017. 2 +[29] Wei-Sheng Lai, Jia-Bin Huang, Narendra Ahuja, and Ming-Hsuan Yang. Fast and accurate image super-resolution with + +deep laplacian pyramid networks. IEEE transactions on pattern analysis and machine intelligence, 41(11):2599-2613, 2018. 2 +[30] Seyed Mehdi Lajevardi and Zahir M Hussain. Higher order orthogonal moments for invariant facial expression recognition. Digital Signal Processing, 20(6):1771-1779, 2010. 2 +[31] Christian Ledig, Lucas Theis, Ferenc Huszár, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photorealistic single image super-resolution using a generative adversarial network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4681-4690, 2017. 2 +[32] Jaewon Lee and Kyong Hwan Jin. Local texture estimator for implicit representation function. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1929-1938, 2022. 1, 2, 3, 5, 6, 7, 8 +[33] Zhen Li, Jinglei Yang, Zheng Liu, Xiaomin Yang, Gwang-gil Jeon, and Wei Wu. Feedback network for image super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3867-3876, 2019. 2 +[34] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1833-1844, 2021. 3 +[35] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 136-144, 2017. 2, 6, 7, 8 +[36] Chieh Hubert Lin, Hsin-Ying Lee, Yen-Chi Cheng, Sergey Tulyakov, and Ming-Hsuan Yang. Infinitygan: Towards infinite-pixel image synthesis. arXiv preprint arXiv:2104.03963, 2021. 2 +[37] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. Advances in Neural Information Processing Systems, 33:15651-15663, 2020. 2 +[38] Pengju Liu, Hongzhi Zhang, Kai Zhang, Liang Lin, and Wangmeng Zuo. Multi-level wavelet-cnn for image restoration. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 773-782, 2018. 2 +[39] Julien Mairal, Michael Elad, and Guillermo Sapiro. Sparse representation for color image restoration. IEEE Transactions on image processing, 17(1):53-69, 2007. 2 +[40] David Martin, Charless Fowlkes, Doron Tal, and Jitendra Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In Proceedings Eighth IEEE International Conference on Computer Vision. ICCV 2001, volume 2, pages 416-423. IEEE, 2001. 6 +[41] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings + +of the IEEE/CVF conference on computer vision and pattern recognition, pages 4460-4470, 2019. 1 +[42] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 2, 3 +[43] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11453-11464, 2021. 2 +[44] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3504-3515, 2020. 1 +[45] Michael Oechsle, Lars Mescheder, Michael Niemeyer, Thilo Strauss, and Andreas Geiger. Texture fields: Learning texture representations in function space. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4531-4540, 2019. 1 +[46] Alan V. Oppenheim and Alan S. Willsky. Signals and Systems. Prentice Hall, Upper Saddle River, NJ, 2 edition, 1997. 4 +[47] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 165-174, 2019. 1 +[48] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. +[49] Songyou Peng, Michael Niemeyer, Lars Mescheder, Marc Pollefeys, and Andreas Geiger. Convolutional occupancy networks. In European Conference on Computer Vision, pages 523-540. Springer, 2020. 1 +[50] Nasim Rahaman, Aristide Baratin, Devansh Arpit, Felix Draxler, Min Lin, Fred Hamprecht, Yoshua Bengio, and Aaron Courville. On the spectral bias of neural networks. In International Conference on Machine Learning, pages 5301-5310. PMLR, 2019. 1, 2 +[51] Ravi Ramamoorthi and Pat Hanrahan. On the relationship between radiance and irradiance: determining the illumination from images of a convex lambertian object. JOSA A, 18(10):2448-2459, 2001. 2 +[52] Oren Rippel, Jasper Snoek, and Ryan P Adams. Spectral representations for convolutional neural networks. Advances in neural information processing systems, 28, 2015. 2 +[53] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. Graf: Generative radiance fields for 3d-aware image synthesis. Advances in Neural Information Processing Systems, 33:20154-20166, 2020. 2 +[54] Wenzhe Shi, Jose Caballero, Ferenc Huszár, Johannes Totz, Andrew P Aitken, Rob Bishop, Daniel Rueckert, and Zehan + +Wang. Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1874-1883, 2016. 3 +[55] Assaf Shocher, Nadav Cohen, and Michal Irani. “zero-shot” super-resolution using deep internal learning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3118–3126, 2018. 2 +[56] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. Advances in Neural Information Processing Systems, 33:7462-7473, 2020. 2 +[57] Vincent Sitzmann, Justus Thies, Felix Heide, Matthias Nießner, Gordon Wetzstein, and Michael Zollhofer. Deepvoxels: Learning persistent 3d feature embeddings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2437-2446, 2019. 1 +[58] Vincent Sitzmann, Michael Zollhöfer, and Gordon Wetzstein. Scene representation networks: Continuous 3d-structure-aware neural scene representations. Advances in Neural Information Processing Systems, 32, 2019. 1 +[59] Peter-Pike Sloan, Jan Kautz, and John Snyder. Precomputed radiance transfer for real-time rendering in dynamic, low-frequency lighting environments. In Proceedings of the 29th annual conference on Computer graphics and interactive techniques, pages 527-536, 2002. 2 +[60] Sanghyun Son and Kyoung Mu Lee. Srwarp: Generalized image super-resolution under arbitrary transformation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7782-7791, 2021. 3 +[61] Ying Tai, Jian Yang, and Xiaoming Liu. Image superresolution via deep recursive residual network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3147-3155, 2017. 2 +[62] Ying Tai, Jian Yang, Xiaoming Liu, and Chunyan Xu. Memnet: A persistent memory network for image restoration. In Proceedings of the IEEE international conference on computer vision, pages 4539-4547, 2017. 2 +[63] Radu Timofte, Eirikur Agustsson, Luc Van Gool, Ming-Hsuan Yang, and Lei Zhang. Ntire 2017 challenge on single image super-resolution: Methods and results. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 114-125, 2017. 6 +[64] Tong Tong, Gen Li, Xiejie Liu, and Qinquan Gao. Image super-resolution using dense skip connections. In Proceedings of the IEEE international conference on computer vision, pages 4799-4807, 2017. 2 +[65] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2 +[66] Longguang Wang, Yingqian Wang, Zaiping Lin, Jungang Yang, Wei An, and Yulan Guo. Learning a single network for scale-arbitrary super-resolution. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4801-4810, 2021. 3 +[67] Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real-esrgan: Training real-world blind super-resolution with + +pure synthetic data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1905-1914, 2021. 3 +[68] Yifan Wang, Federico Perazzi, Brian McWilliams, Alexander Sorkine-Hornung, Olga Sorkine-Hornung, and Christopher Schroers. A fully progressive approach to single-image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 864-873, 2018. 2 +[69] Yifan Wang, Federico Perazzi, Brian McWilliams, Alexander Sorkine-Hornung, Olga Sorkine-Hornung, and Christopher Schroers. A fully progressive approach to single-image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 864-873, 2018. 2 +[70] Zhihao Wang, Jian Chen, and Steven CH Hoi. Deep learning for image super-resolution: A survey. IEEE transactions on pattern analysis and machine intelligence, 43(10):3365-3387, 2020. 1, 2, 3 +[71] John Wright, Allen Y Yang, Arvind Ganesh, S Shankar Sastry, and Yi Ma. Robust face recognition via sparse representation. IEEE transactions on pattern analysis and machine intelligence, 31(2):210-227, 2008. 2 +[72] Xingqian Xu, Zhangyang Wang, and Humphrey Shi. Ultrasr: Spatial encoding is a missing key for implicit image function-based arbitrary-scale super-resolution. arXiv preprint arXiv:2103.12716, 2021. 1, 2, 3 +[73] Jianchao Yang, John Wright, Thomas Huang, and Yi Ma. Image super-resolution as sparse representation of raw image patches. In 2008 IEEE conference on computer vision and pattern recognition, pages 1-8. IEEE, 2008. 2 +[74] Yanchao Yang and Stefano Soatto. Fda: Fourier domain adaptation for semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4085-4095, 2020. 2 +[75] Youngho Yoon, Inchul Chung, Lin Wang, and Kuk-Jin Yoon. Spheresr: 360deg image super-resolution with arbitrary projection via continuous spherical image representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5677-5686, 2022. 3 +[76] Roman Zeyde, Michael Elad, and Matan Protter. On single image scale-up using sparse-representations. In International conference on curves and surfaces, pages 711-730. Springer, 2010. 6 +[77] Hui Zhang, Huazhong Shu, Guoniu N Han, Gouenou Coa-trieux, Limin Luo, and Jean Louis Coa-trieux. Blurred image recognition by legendre moment invariants. IEEE Transactions on Image Processing, 19(3):596-611, 2009. 2 +[78] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492, 2020. 2 +[79] Yulun Zhang, Kunpeng Li, Kai Li, Lichen Wang, Bineng Zhong, and Yun Fu. Image super-resolution using very deep residual channel attention networks. In Proceedings of the European conference on computer vision (ECCV), pages 286-301, 2018. 2 +[80] Yulun Zhang, Yapeng Tian, Yu Kong, Bineng Zhong, and Yun Fu. Residual dense network for image super-resolution. + +In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2472-2481, 2018. 6, 7, 8 +[81] Man Zhou, Hu Yu, Jie Huang, Feng Zhao, Jinwei Gu, Chen Change Loy, Deyu Meng, and Chongyi Li. Deep fourier up-sampling. arXiv preprint arXiv:2210.05171, 2022. 2 +[82] Hongqing Zhu. Image representation using separable two-dimensional continuous and discrete orthogonal moments. Pattern Recognition, 45(4):1540-1558, 2012. 2 \ No newline at end of file diff --git a/2023/OPE-SR_ Orthogonal Position Encoding for Designing a Parameter-Free Upsampling Module in Arbitrary-Scale Image Super-Resolution/images.zip b/2023/OPE-SR_ Orthogonal Position Encoding for Designing a Parameter-Free Upsampling Module in Arbitrary-Scale Image Super-Resolution/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..257678d938cca411ba5f91aa2723ab227b36700c --- /dev/null +++ b/2023/OPE-SR_ Orthogonal Position Encoding for Designing a Parameter-Free Upsampling Module in Arbitrary-Scale Image Super-Resolution/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:78420c5eb4ea2233f951de689a0d44a479a265107b1ace15ad41e4c266f12b36 +size 616693 diff --git a/2023/OPE-SR_ Orthogonal Position Encoding for Designing a Parameter-Free Upsampling Module in Arbitrary-Scale Image Super-Resolution/layout.json b/2023/OPE-SR_ Orthogonal Position Encoding for Designing a Parameter-Free Upsampling Module in Arbitrary-Scale Image Super-Resolution/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f12b94b835197669273e1182008f0de948d04052 --- /dev/null +++ b/2023/OPE-SR_ Orthogonal Position Encoding for Designing a Parameter-Free Upsampling Module in Arbitrary-Scale Image Super-Resolution/layout.json @@ -0,0 +1,12024 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 76, + 102, + 519, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 102, + 519, + 140 + ], + "spans": [ + { + "bbox": [ + 76, + 102, + 519, + 140 + ], + "type": "text", + "content": "OPE-SR: Orthogonal Position Encoding for Designing a Parameter-free Upsampling Module in Arbitrary-scale Image Super-Resolution" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 92, + 160, + 501, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 160, + 501, + 175 + ], + "spans": [ + { + "bbox": [ + 92, + 160, + 501, + 175 + ], + "type": "text", + "content": "Gaochao Song" + }, + { + "bbox": [ + 92, + 160, + 501, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 92, + 160, + 501, + 175 + ], + "type": "text", + "content": " Qian Sun" + }, + { + "bbox": [ + 92, + 160, + 501, + 175 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 92, + 160, + 501, + 175 + ], + "type": "text", + "content": " Luo Zhang" + }, + { + "bbox": [ + 92, + 160, + 501, + 175 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 92, + 160, + 501, + 175 + ], + "type": "text", + "content": " Ran Su" + }, + { + "bbox": [ + 92, + 160, + 501, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 92, + 160, + 501, + 175 + ], + "type": "text", + "content": " Jianfeng Shi" + }, + { + "bbox": [ + 92, + 160, + 501, + 175 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 92, + 160, + 501, + 175 + ], + "type": "text", + "content": " Ying He" + }, + { + "bbox": [ + 92, + 160, + 501, + 175 + ], + "type": "inline_equation", + "content": "^{3}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 97, + 175, + 496, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 175, + 496, + 204 + ], + "spans": [ + { + "bbox": [ + 97, + 175, + 496, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 97, + 175, + 496, + 204 + ], + "type": "text", + "content": "Tianjin University " + }, + { + "bbox": [ + 97, + 175, + 496, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 97, + 175, + 496, + 204 + ], + "type": "text", + "content": "Nanjing University of Information Science and Technology " + }, + { + "bbox": [ + 97, + 175, + 496, + 204 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 97, + 175, + 496, + 204 + ], + "type": "text", + "content": "Nanyang Technological University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 166, + 205, + 423, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 205, + 423, + 217 + ], + "spans": [ + { + "bbox": [ + 166, + 205, + 423, + 217 + ], + "type": "text", + "content": "gaochaosong_21@tju.edu.cn sunqian@nuist.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "spans": [ + { + "bbox": [ + 143, + 244, + 192, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 270, + 289, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 270, + 289, + 521 + ], + "spans": [ + { + "bbox": [ + 46, + 270, + 289, + 521 + ], + "type": "text", + "content": "Arbitrary-scale image super-resolution (SR) is often tackled using the implicit neural representation (INR) approach, which relies on a position encoding scheme to improve its representation ability. In this paper, we introduce orthogonal position encoding (OPE), an extension of position encoding, and an OPE-Upscale module to replace the INR-based upsampling module for arbitrary-scale image super-resolution. Our OPE-Upscale module takes 2D coordinates and latent code as inputs, just like INR, but does not require any training parameters. This parameter-free feature allows the OPE-Upscale module to directly perform linear combination operations, resulting in continuous image reconstruction and achieving arbitrary-scale image reconstruction. As a concise SR framework, our method is computationally efficient and consumes less memory than state-of-the-art methods, as confirmed by extensive experiments and evaluations. In addition, our method achieves comparable results with state-of-the-art methods in arbitrary-scale image super-resolution. Lastly, we show that OPE corresponds to a set of orthogonal basis, validating our design principle." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 544, + 128, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 544, + 128, + 558 + ], + "spans": [ + { + "bbox": [ + 47, + 544, + 128, + 558 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 565, + 287, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 565, + 287, + 685 + ], + "spans": [ + { + "bbox": [ + 46, + 565, + 287, + 685 + ], + "type": "text", + "content": "Photographs are composed of discrete pixels of varying precision due to the limitations of sampling frequency, which breaks the continuous visual world into discrete parts. The single image super-resolution (SISR) task aims to restore the original continuous world in the image as much as possible. In an arbitrary-scale SR task, one often reconstructs the continuous representation of a low-resolution image and then adjusts the resolution of the target image as needed. The recent rise of implicit neural representation (INR) in 3D vision has enabled the representation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 246, + 545, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 246, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 306, + 246, + 545, + 293 + ], + "type": "text", + "content": "of complex 3D objects and scenes in a continuous manner [14, 19, 41, 42, 44, 45, 47, 49, 57, 58], which also opens up possibilities for continuous image and arbitrary-scale image super-resolution [5, 18, 32, 72]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 294, + 546, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 294, + 546, + 497 + ], + "spans": [ + { + "bbox": [ + 304, + 294, + 546, + 497 + ], + "type": "text", + "content": "Existing methods for arbitrary-scale SR typically use a post-upsampling framework [70]. In this approach, low-resolution (LR) images first pass through a deep CNN network (encoder) without improving the resolution, and then pass through an INR-based upsampling module (decoder) with a specified target resolution to reconstruct high-resolution (HR) images. The decoder establishes a mapping from feature maps (the output of encoder) to target image pixels using a pre-assigned grid partitioning and achieves arbitrary-scale with the density of the grid in Cartesian coordinate system. However, the INR approach has a defect of learning low-frequency information, also known as spectral bias [50]. To address this issue, sinusoidal positional encoding is introduced to embed input coordinates to higher dimensions and enable the network to learn high-frequency details. This inspired recent works on arbitrary-scale SR to further improve the representation ability [32, 72]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 498, + 547, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 547, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 547, + 653 + ], + "type": "text", + "content": "Despite its effectiveness in arbitrary-scale SR, the INR-based upsampling module increases the complexity of the entire SR framework as two different networks are jointly trained. Additionally, as a black-box model, it represents a continuous image with a strong dependency on both the feature map and the decoder (e.g., MLP). However, its representation ability decreases after flipping the feature map, a phenomenon known as flipping consistency decline. As shown in Fig. 1, flipping the feature map horizontally before the upsampling module of LIIF results in a blurred target image that does not have the expected flip transformation. This decline could be due to limitations of the MLP in learning the symmetry feature of the image." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 715 + ], + "type": "text", + "content": "MLP is a universal function approximator [17], which tries to fit a mapping function from feature map to the continuous image, therefore, it is reasonable to assume that such process could be solved by an analytical solution. In this paper, we re-examine position encoding from the per" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 693, + 136, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 693, + 136, + 703 + ], + "spans": [ + { + "bbox": [ + 58, + 693, + 136, + 703 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 59, + 703, + 277, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 703, + 277, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 703, + 277, + 713 + ], + "type": "text", + "content": "1Project page: https://github.com/gaochao-s/ope-sr" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "10009" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": "spective of orthogonal basis and propose orthogonal position encoding (OPE) for continuous image representation. The linear combination of 1D latent code and OPE can directly reconstruct continuous image patch without using implicit neural function [5]. To prove OPE's rationality, we analyse it both from functional analysis and 2D-Fourier transform. We further embed it into a parameter-free upsampling module, called OPE-Upscale Module, to replace INR-based upsampling module in deep SR framework, then currently deep SR framework can be greatly simplified." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 192, + 289, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 192, + 289, + 276 + ], + "spans": [ + { + "bbox": [ + 46, + 192, + 289, + 276 + ], + "type": "text", + "content": "Unlike the state-of-the-art method by Lee et al. [32], which enhances MLP with position encoding, we explore the possibility of extending position encoding without MLP. By providing a more concise SR framework, our method achieves high computing efficiency and consumes less memory than the state-of-the-art, while also achieving comparable image performance in arbitrary-scale SR tasks." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 59, + 276, + 193, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 276, + 193, + 286 + ], + "spans": [ + { + "bbox": [ + 59, + 276, + 193, + 286 + ], + "type": "text", + "content": "Our contributions are as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 293, + 288, + 517 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 58, + 293, + 287, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 293, + 287, + 365 + ], + "spans": [ + { + "bbox": [ + 58, + 293, + 287, + 365 + ], + "type": "text", + "content": "- We propose a novel position encoding, called orthogonal position encoding (OPE), which takes the form of a 2D-Fourier series and corresponds to a set of orthogonal basis. Building on OPE, we introduce the OPE-Upscale Module, a parameter-free upsampling module for arbitrary-scale image super-resolution." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 372, + 288, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 372, + 288, + 407 + ], + "spans": [ + { + "bbox": [ + 58, + 372, + 288, + 407 + ], + "type": "text", + "content": "- Our method significantly reduces the consumption of computing resources, resulting in high computing efficiency for arbitrary-scale SR tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 414, + 287, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 414, + 287, + 463 + ], + "spans": [ + { + "bbox": [ + 58, + 414, + 287, + 463 + ], + "type": "text", + "content": "- The OPE-Upscale Module is interpretable, parameter-free and does not require training, resulting in a concise SR framework that elegantly solves the flipping consistency problem." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 58, + 468, + 287, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 468, + 287, + 517 + ], + "spans": [ + { + "bbox": [ + 58, + 468, + 287, + 517 + ], + "type": "text", + "content": "- Extensive experiments demonstrate that our method achieves comparable results with the state-of-the-art. Furthermore, our method enables super-resolution up to a large scale of " + }, + { + "bbox": [ + 58, + 468, + 287, + 517 + ], + "type": "inline_equation", + "content": "\\times 30" + }, + { + "bbox": [ + 58, + 468, + 287, + 517 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 532, + 134, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 532, + 134, + 544 + ], + "spans": [ + { + "bbox": [ + 47, + 532, + 134, + 544 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 552, + 216, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 552, + 216, + 565 + ], + "spans": [ + { + "bbox": [ + 47, + 552, + 216, + 565 + ], + "type": "text", + "content": "2.1. Sinusoidal Positional Encoding" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "text", + "content": "Sinusoidal positional encoding is widely used to counteract the negative effects of token order and sequence length in sequence models [65], or to guide image generation as the spatial inductive bias in CNNs [7, 23, 36]. In implicit neural representations, it plays a critical role in solving spectral bias [50]. By embedding input coordinates into a higher dimensional space, position encoding greatly improves high-frequency representation of implicit 3D scenes [42, 56] and subsequent works take it as the default operation to improve representation quality [37, 43, 53, 78]. Inspired by these works, positional encoding has been preliminarily explored in representing images in arbitrary-scale SR [32, 72]." + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 315, + 70, + 425, + 125 + ], + "blocks": [ + { + "bbox": [ + 315, + 70, + 425, + 125 + ], + "lines": [ + { + "bbox": [ + 315, + 70, + 425, + 125 + ], + "spans": [ + { + "bbox": [ + 315, + 70, + 425, + 125 + ], + "type": "image", + "image_path": "431e70fe5e15799c2fd2cfcf81a85fecf1bfe8720acc6cede96f95242475aa4f.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 323, + 125, + 361, + 134 + ], + "lines": [ + { + "bbox": [ + 323, + 125, + 361, + 134 + ], + "spans": [ + { + "bbox": [ + 323, + 125, + 361, + 134 + ], + "type": "text", + "content": "LIIF:32.35" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 372, + 125, + 424, + 134 + ], + "lines": [ + { + "bbox": [ + 372, + 125, + 424, + 134 + ], + "spans": [ + { + "bbox": [ + 372, + 125, + 424, + 134 + ], + "type": "text", + "content": "LIIF-flip: 28.47" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 430, + 70, + 536, + 126 + ], + "blocks": [ + { + "bbox": [ + 430, + 70, + 536, + 126 + ], + "lines": [ + { + "bbox": [ + 430, + 70, + 536, + 126 + ], + "spans": [ + { + "bbox": [ + 430, + 70, + 536, + 126 + ], + "type": "image", + "image_path": "2c5698403b7ce46cbf5dbe6eb0580985fcc06df8f1c84856cb42c264a35e73da.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 436, + 126, + 473, + 134 + ], + "lines": [ + { + "bbox": [ + 436, + 126, + 473, + 134 + ], + "spans": [ + { + "bbox": [ + 436, + 126, + 473, + 134 + ], + "type": "text", + "content": "LIIF:25.43" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 484, + 126, + 536, + 134 + ], + "lines": [ + { + "bbox": [ + 484, + 126, + 536, + 134 + ], + "spans": [ + { + "bbox": [ + 484, + 126, + 536, + 134 + ], + "type": "text", + "content": "LIIF-flip: 23.11" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 315, + 135, + 425, + 190 + ], + "blocks": [ + { + "bbox": [ + 315, + 135, + 425, + 190 + ], + "lines": [ + { + "bbox": [ + 315, + 135, + 425, + 190 + ], + "spans": [ + { + "bbox": [ + 315, + 135, + 425, + 190 + ], + "type": "image", + "image_path": "9a15d2c0a4f1a1b7a11386054ae16ec18bd0c0b1de4f758e9301086228311523.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 323, + 190, + 362, + 198 + ], + "lines": [ + { + "bbox": [ + 323, + 190, + 362, + 198 + ], + "spans": [ + { + "bbox": [ + 323, + 190, + 362, + 198 + ], + "type": "text", + "content": "OPE: 32.37" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 372, + 190, + 424, + 199 + ], + "lines": [ + { + "bbox": [ + 372, + 190, + 424, + 199 + ], + "spans": [ + { + "bbox": [ + 372, + 190, + 424, + 199 + ], + "type": "text", + "content": "OPE-flip: 32.37" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 202, + 547, + 258 + ], + "lines": [ + { + "bbox": [ + 304, + 202, + 547, + 258 + ], + "spans": [ + { + "bbox": [ + 304, + 202, + 547, + 258 + ], + "type": "text", + "content": "Figure 1. Flipping consistency decline (PSNR (dB)). LIIF-flip: Flipping the input of the LIIF [5] decoder yields blurred results in the symmetric outputs. OPE-flip: Our method does not exhibit such artifacts. Additional results can be found in the supplementary material." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 430, + 135, + 536, + 190 + ], + "blocks": [ + { + "bbox": [ + 430, + 135, + 536, + 190 + ], + "lines": [ + { + "bbox": [ + 430, + 135, + 536, + 190 + ], + "spans": [ + { + "bbox": [ + 430, + 135, + 536, + 190 + ], + "type": "image", + "image_path": "5386d6e74192c053b22db82f332fd7d30e29ed161cc485d8585e5143e92c53d1.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 435, + 190, + 474, + 198 + ], + "lines": [ + { + "bbox": [ + 435, + 190, + 474, + 198 + ], + "spans": [ + { + "bbox": [ + 435, + 190, + 474, + 198 + ], + "type": "text", + "content": "OPE:25.50" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 484, + 190, + 536, + 198 + ], + "lines": [ + { + "bbox": [ + 484, + 190, + 536, + 198 + ], + "spans": [ + { + "bbox": [ + 484, + 190, + 536, + 198 + ], + "type": "text", + "content": "OPE-flip: 25.50" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 281, + 482, + 293 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 281, + 482, + 293 + ], + "spans": [ + { + "bbox": [ + 306, + 281, + 482, + 293 + ], + "type": "text", + "content": "2.2. Orthogonal Basis Decomposition" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 300, + 547, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 300, + 547, + 587 + ], + "spans": [ + { + "bbox": [ + 304, + 300, + 547, + 587 + ], + "type": "text", + "content": "In functional analysis, an orthogonal basis decomposition is a way to represent a vector or function as a linear combination of orthogonal basis functions. Wavelet transform [2, 15, 20, 38] and 2D-Fourier transform [6, 12, 13, 22, 74] are commonly used decomposition techniques for images and videos. In DSGAN [12], the input image is explicitly decomposed into low and high frequencies using high-pass and low-pass filters. Other methods use frequency domain losses to decompose images, either in a supervised manner [22] or in an unsupervised manner [13]. To address with resolution discrepancy of reconstructed images and input images, Rippel et al. [52] employ spectral pooling to decrease resolution by truncating in the Fourier domain, while Zhou et al. [81] explore an up-sampling method in the Fourier domain to increase the resolution. Image moments, which decompose images into two-dimensional orthogonal polynomials [27, 82], are widely used in invariant pattern recognition [30, 77]. Image sparse representation inherits this decomposition idea and performs well in traditional computer vision tasks [39, 71, 73]. In 3D domains, spherical harmonics serve as an orthogonal basis in space to represent view dependence [3, 51, 59] and have recently been proposed as a replacement for MLPs [11] for representing neural radiance fields [42]." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 306, + 597, + 446, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 597, + 446, + 611 + ], + "spans": [ + { + "bbox": [ + 306, + 597, + 446, + 611 + ], + "type": "text", + "content": "2.3. Deep Learning-based SR" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": "Based on the upsampling operations and their location in the model, deep learning-based SR frameworks can be classified into four categories (see [70] for a comprehensive survey): pre-upsampling [8,24,25,55,61,62], post-upsampling [9,31,35,64,79], progressive-upsampling [28,29,68], and iterative up-and-down sampling [16, 33, 69]. With preupsampling, the LR image is first upsampled by traditional interpolation and then fed into a deep CNN to reconstruct" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10010" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 240 + ], + "type": "text", + "content": "high-quality details. While it was one of the most popular frameworks for arbitrary-scale factors, it has side effects like enlarged noise by interpolation and high time complexity and space consumption. The progressive-upsampling and iterative up-and-down sampling frameworks pose challenges in terms of complicated model designing and unclear design criteria, as noted in [70]. For post-upsampling, the LR image is directly fed as input to a deep CNN, and then a trainable upsampling module (e.g., deconvolution [9], subpixel [54], and interpolation convolution [10]) increases the resolution at the end. Since feature extraction process, which is computationally intensive, only occurs in low-dimensional space, it has become one of the mainstream frameworks [32, 34, 67]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 246, + 160, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 246, + 160, + 258 + ], + "spans": [ + { + "bbox": [ + 47, + 246, + 160, + 258 + ], + "type": "text", + "content": "2.4. Arbitrary-scale SR" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 264, + 289, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 264, + 289, + 434 + ], + "spans": [ + { + "bbox": [ + 46, + 264, + 289, + 434 + ], + "type": "text", + "content": "In the field of arbitrary-scale SR, most existing works are based on the post-upsampling framework and replace the traditional upsampling module with an INR-based one, such as a coordinate-based MLP. Meta-SR [18] was the first arbitrary-scale SR method based on CNN. ArbSR [66] adopts a general plug-in module to solve the scaling problem of different horizontal and vertical scales. SRWarp [60] transforms LR images into HR images with arbitrary shapes via a differential adaptive warping layer. SphereSR [75] explores arbitrary-scale on " + }, + { + "bbox": [ + 46, + 264, + 289, + 434 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 46, + 264, + 289, + 434 + ], + "type": "text", + "content": " images. LIIF [5] uses coordinates and conditional latent code into an MLP to directly predict target pixel color with an intuitive network structure. LIIF-related follow-up works focus on predicting high-frequency information with position encoding [32,72]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 441, + 103, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 441, + 103, + 453 + ], + "spans": [ + { + "bbox": [ + 47, + 441, + 103, + 453 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 460, + 229, + 474 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 460, + 229, + 474 + ], + "spans": [ + { + "bbox": [ + 47, + 460, + 229, + 474 + ], + "type": "text", + "content": "3.1. OPE-based Image Representation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 479, + 287, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 479, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 46, + 479, + 287, + 601 + ], + "type": "text", + "content": "Given an LR image " + }, + { + "bbox": [ + 46, + 479, + 287, + 601 + ], + "type": "inline_equation", + "content": "I_{LR}" + }, + { + "bbox": [ + 46, + 479, + 287, + 601 + ], + "type": "text", + "content": " with resolution of " + }, + { + "bbox": [ + 46, + 479, + 287, + 601 + ], + "type": "inline_equation", + "content": "h \\times w" + }, + { + "bbox": [ + 46, + 479, + 287, + 601 + ], + "type": "text", + "content": ", we decide its 2D domain into " + }, + { + "bbox": [ + 46, + 479, + 287, + 601 + ], + "type": "inline_equation", + "content": "h \\times w" + }, + { + "bbox": [ + 46, + 479, + 287, + 601 + ], + "type": "text", + "content": " grids, where each grid represents a pixel in the LR image and corresponds to a patch of the high resolution image of size " + }, + { + "bbox": [ + 46, + 479, + 287, + 601 + ], + "type": "inline_equation", + "content": "r_h \\times r_w" + }, + { + "bbox": [ + 46, + 479, + 287, + 601 + ], + "type": "text", + "content": ". The output image " + }, + { + "bbox": [ + 46, + 479, + 287, + 601 + ], + "type": "inline_equation", + "content": "I_{SR}" + }, + { + "bbox": [ + 46, + 479, + 287, + 601 + ], + "type": "text", + "content": " has a resolution of " + }, + { + "bbox": [ + 46, + 479, + 287, + 601 + ], + "type": "inline_equation", + "content": "H \\times W" + }, + { + "bbox": [ + 46, + 479, + 287, + 601 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 479, + 287, + 601 + ], + "type": "inline_equation", + "content": "H = r_h \\cdot h" + }, + { + "bbox": [ + 46, + 479, + 287, + 601 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 479, + 287, + 601 + ], + "type": "inline_equation", + "content": "W = r_w \\cdot w" + }, + { + "bbox": [ + 46, + 479, + 287, + 601 + ], + "type": "text", + "content": ". Denote by " + }, + { + "bbox": [ + 46, + 479, + 287, + 601 + ], + "type": "inline_equation", + "content": "I \\in \\mathbb{R}^{r_h \\times r_w \\times 1}" + }, + { + "bbox": [ + 46, + 479, + 287, + 601 + ], + "type": "text", + "content": " the high resolution image patch with size " + }, + { + "bbox": [ + 46, + 479, + 287, + 601 + ], + "type": "inline_equation", + "content": "r_h \\times r_w" + }, + { + "bbox": [ + 46, + 479, + 287, + 601 + ], + "type": "text", + "content": " for a specific color channel. We view each pixel in " + }, + { + "bbox": [ + 46, + 479, + 287, + 601 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 46, + 479, + 287, + 601 + ], + "type": "text", + "content": " as a sample of a continuous bivariate function " + }, + { + "bbox": [ + 46, + 479, + 287, + 601 + ], + "type": "inline_equation", + "content": "f(x,y): [-1,1] \\times [-1,1] \\to \\mathbb{R}" + }, + { + "bbox": [ + 46, + 479, + 287, + 601 + ], + "type": "text", + "content": ". We embed the coordinates " + }, + { + "bbox": [ + 46, + 479, + 287, + 601 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 479, + 287, + 601 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 479, + 287, + 601 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 479, + 287, + 601 + ], + "type": "text", + "content": " using sinusoidal positional encoding" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 59, + 604, + 287, + 643 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 604, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 59, + 604, + 287, + 643 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} { X } { = \\gamma ( x ) = \\sqrt { 2 } \\cdot [ \\frac { 1 } { \\sqrt { 2 } } , \\cos ( \\pi x ) , \\sin ( \\pi x ) , \\cos ( 2 \\pi x ) , } \\\\ \\sin (2 \\pi x), \\dots , \\cos (n \\pi x), \\sin (n \\pi x) ] \\tag {1} \\\\ \\end{array}", + "image_path": "64c0db7324cc735b3790712597c9de9bcd0e350e9b1988a4512705874fc3b3e9.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 60, + 644, + 287, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 644, + 287, + 685 + ], + "spans": [ + { + "bbox": [ + 60, + 644, + 287, + 685 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} { Y } { = \\gamma ( y ) = \\sqrt { 2 } \\cdot \\left[ \\frac { 1 } { \\sqrt { 2 } } , \\cos ( \\pi y ) , \\sin ( \\pi y ) , \\cos ( 2 \\pi y ) , \\right. } \\\\ \\left. \\sin (2 \\pi y), \\dots , \\cos (n \\pi y), \\sin (n \\pi y) \\right], \\tag {2} \\\\ \\end{array}", + "image_path": "d8c0bf26f067ede94beb8fb26a08cdba67f790c971a37584479b517afe7c3265.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\gamma (\\cdot):\\mathbb{R}\\to \\mathbb{R}^{1\\times (2n + 1)}" + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": " is a univariate function for position encoding with a predefined maximum frequency" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "inline_equation", + "content": "n\\in \\mathbb{N}" + }, + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "content": " . We flatten the matrix " + }, + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "inline_equation", + "content": "X^T Y" + }, + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "content": " into a row vector " + }, + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "inline_equation", + "content": "P\\in" + }, + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{1\\times (2n + 1)^2}" + }, + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 390, + 95, + 545, + 109 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 390, + 95, + 545, + 109 + ], + "spans": [ + { + "bbox": [ + 390, + 95, + 545, + 109 + ], + "type": "interline_equation", + "content": "P = \\operatorname {f l a t} \\left(X ^ {T} Y\\right), \\tag {3}", + "image_path": "21929a9a92a7c1af3dff7d04c96423421f453aa65eb423ca39aba90d2f7c47ea.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 112, + 547, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 112, + 547, + 163 + ], + "spans": [ + { + "bbox": [ + 304, + 112, + 547, + 163 + ], + "type": "text", + "content": "where flat: " + }, + { + "bbox": [ + 304, + 112, + 547, + 163 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{(2n + 1)\\times (2n + 1)}\\to \\mathbb{R}^{1\\times (2n + 1)^2}" + }, + { + "bbox": [ + 304, + 112, + 547, + 163 + ], + "type": "text", + "content": " is the flattening operation. Denote by " + }, + { + "bbox": [ + 304, + 112, + 547, + 163 + ], + "type": "inline_equation", + "content": "e_{i,j}" + }, + { + "bbox": [ + 304, + 112, + 547, + 163 + ], + "type": "text", + "content": " the element on the " + }, + { + "bbox": [ + 304, + 112, + 547, + 163 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 112, + 547, + 163 + ], + "type": "text", + "content": " -th row and the " + }, + { + "bbox": [ + 304, + 112, + 547, + 163 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 112, + 547, + 163 + ], + "type": "text", + "content": " -th column of matrix " + }, + { + "bbox": [ + 304, + 112, + 547, + 163 + ], + "type": "inline_equation", + "content": "X^T Y" + }, + { + "bbox": [ + 304, + 112, + 547, + 163 + ], + "type": "text", + "content": " . For example. " + }, + { + "bbox": [ + 304, + 112, + 547, + 163 + ], + "type": "inline_equation", + "content": "e_{4,5} = 2\\cos (2\\pi x)\\sin (2\\pi y)" + }, + { + "bbox": [ + 304, + 112, + 547, + 163 + ], + "type": "text", + "content": " . It is easy to verify that" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 335, + 170, + 545, + 201 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 170, + 545, + 201 + ], + "spans": [ + { + "bbox": [ + 335, + 170, + 545, + 201 + ], + "type": "interline_equation", + "content": "\\left\\langle e _ {i _ {1}, j _ {1}}, e _ {i _ {2}, j _ {2}} \\right\\rangle = \\left\\{ \\begin{array}{l l} 0, & (i _ {1}, j _ {1}) \\neq (i _ {2}, j _ {2}) \\\\ 1, & (i _ {1}, j _ {1}) = (i _ {2}, j _ {2}) \\end{array} \\right. \\tag {4}", + "image_path": "ee96091276328b37df2a11b067e550f5b4fdd074e24a21ba958ef12490d88fb9.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 208, + 518, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 208, + 518, + 221 + ], + "spans": [ + { + "bbox": [ + 306, + 208, + 518, + 221 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 208, + 518, + 221 + ], + "type": "inline_equation", + "content": "\\langle ,\\rangle" + }, + { + "bbox": [ + 306, + 208, + 518, + 221 + ], + "type": "text", + "content": " is the inner product in function space, i.e.," + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 343, + 227, + 510, + 255 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 227, + 510, + 255 + ], + "spans": [ + { + "bbox": [ + 343, + 227, + 510, + 255 + ], + "type": "interline_equation", + "content": "\\langle g, h \\rangle = \\frac {1}{4} \\int_ {- 1} ^ {1} \\int_ {- 1} ^ {1} g (x, y) h (x, y) \\mathrm {d} x \\mathrm {d} y.", + "image_path": "6b8a1015007858026df58d7154369b9d93fd38724ce9fa0d81cf32189c5c8b1f.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 262, + 545, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 262, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 305, + 262, + 545, + 335 + ], + "type": "text", + "content": "Therefore the elements " + }, + { + "bbox": [ + 305, + 262, + 545, + 335 + ], + "type": "inline_equation", + "content": "\\{e_{i,j}\\}" + }, + { + "bbox": [ + 305, + 262, + 545, + 335 + ], + "type": "text", + "content": " form a set of orthogonal basis, allowing us to approximate " + }, + { + "bbox": [ + 305, + 262, + 545, + 335 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 305, + 262, + 545, + 335 + ], + "type": "text", + "content": " as a linear combination. Consider a pixel " + }, + { + "bbox": [ + 305, + 262, + 545, + 335 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 305, + 262, + 545, + 335 + ], + "type": "text", + "content": " in the upsampled image patch " + }, + { + "bbox": [ + 305, + 262, + 545, + 335 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 305, + 262, + 545, + 335 + ], + "type": "text", + "content": ", which corresponds to a region " + }, + { + "bbox": [ + 305, + 262, + 545, + 335 + ], + "type": "inline_equation", + "content": "[x - 1 / r_h, x + 1 / r_h] \\times [y - 1 / r_w, y + 1 / r_w]" + }, + { + "bbox": [ + 305, + 262, + 545, + 335 + ], + "type": "text", + "content": ". We use " + }, + { + "bbox": [ + 305, + 262, + 545, + 335 + ], + "type": "inline_equation", + "content": "f(x,y)" + }, + { + "bbox": [ + 305, + 262, + 545, + 335 + ], + "type": "text", + "content": " as a representative for the entire region and compute " + }, + { + "bbox": [ + 305, + 262, + 545, + 335 + ], + "type": "inline_equation", + "content": "I_{(x,y)}" + }, + { + "bbox": [ + 305, + 262, + 545, + 335 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 373, + 341, + 545, + 357 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 341, + 545, + 357 + ], + "spans": [ + { + "bbox": [ + 373, + 341, + 545, + 357 + ], + "type": "interline_equation", + "content": "I _ {(x, y)} \\triangleq f (x, y) \\approx Z P ^ {T}, \\tag {5}", + "image_path": "56e0e7140979744591dcf75eea96c61d3040b77460458930c41d7305fadc0221.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 364, + 545, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 364, + 545, + 413 + ], + "spans": [ + { + "bbox": [ + 304, + 364, + 545, + 413 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 364, + 545, + 413 + ], + "type": "inline_equation", + "content": "Z \\in \\mathbb{R}^{1 \\times (2n + 1)^2}" + }, + { + "bbox": [ + 304, + 364, + 545, + 413 + ], + "type": "text", + "content": " represents the projection. Due to the orthogonal property, we call the resulting vector " + }, + { + "bbox": [ + 304, + 364, + 545, + 413 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 364, + 545, + 413 + ], + "type": "text", + "content": " as orthogonal position encoding (OPE). Fig. 3 illustrates the concept of OPE-based patch representation." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 413, + 545, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 413, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 304, + 413, + 545, + 472 + ], + "type": "text", + "content": "Remark 1. Our OPE basis can be seen as the real form version of the 2D-Fourier basis, which eliminates the complex exponential term based on conjugate symmetry when representing real signals. See the supplementary material for details." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 473, + 545, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 473, + 545, + 592 + ], + "spans": [ + { + "bbox": [ + 304, + 473, + 545, + 592 + ], + "type": "text", + "content": "Remark 2. OPE differs from the commonly used positional encoding formulation, such as that in [42], in that it includes a constant term and takes the product of each coordinate embedding as a new term. These seemingly minor changes indeed have a deep impact. Conventional positional encoding processes the coordinates separately, thereby encoding frequencies in the horizontal and vertical directions only. In contrast, OPE includes frequencies covering in all directions of the plane due to the product terms, resulting in a better expression capability." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 593, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 593, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 593, + 545, + 628 + ], + "type": "text", + "content": "Remark 3. Unlike LIIF [5] that uses an MLP to approximate the image function " + }, + { + "bbox": [ + 304, + 593, + 545, + 628 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 593, + 545, + 628 + ], + "type": "text", + "content": ", our method completely eliminates the need of MLP." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 635, + 430, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 635, + 430, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 635, + 430, + 647 + ], + "type": "text", + "content": "3.2. OPE-Upscale Module" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": "We project the latent code onto the OPE basis. OPE with a sufficiently long latent code could represent an image directly in a continuous manner. However, it suffers from long embedding time and is unstable for representing high-frequency details locally, similar to the limitations" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10011" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 539, + 194 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 539, + 194 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 539, + 194 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 539, + 194 + ], + "type": "image", + "image_path": "315202109cbd5ab2b2ae09d337dd9a1ba4711b5da23cf976bd21202fadfd71de.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 201, + 547, + 250 + ], + "lines": [ + { + "bbox": [ + 46, + 201, + 547, + 250 + ], + "spans": [ + { + "bbox": [ + 46, + 201, + 547, + 250 + ], + "type": "text", + "content": "Figure 2. OPE-Upscale Module for arbitrary-scale SR. The encoder " + }, + { + "bbox": [ + 46, + 201, + 547, + 250 + ], + "type": "inline_equation", + "content": "E_{\\theta}" + }, + { + "bbox": [ + 46, + 201, + 547, + 250 + ], + "type": "text", + "content": " is the only trainable part. With a pre-defined maximum frequency " + }, + { + "bbox": [ + 46, + 201, + 547, + 250 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 201, + 547, + 250 + ], + "type": "text", + "content": " of OPE, the OPE-Upscale Module (shaded in grey) takes the feature map from " + }, + { + "bbox": [ + 46, + 201, + 547, + 250 + ], + "type": "inline_equation", + "content": "E_{\\theta}" + }, + { + "bbox": [ + 46, + 201, + 547, + 250 + ], + "type": "text", + "content": " and the target resolution " + }, + { + "bbox": [ + 46, + 201, + 547, + 250 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 46, + 201, + 547, + 250 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 201, + 547, + 250 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 46, + 201, + 547, + 250 + ], + "type": "text", + "content": " as input, and renders the pixels of the target SR image in parallel. " + }, + { + "bbox": [ + 46, + 201, + 547, + 250 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 46, + 201, + 547, + 250 + ], + "type": "text", + "content": " is the matmul function that returns the product of " + }, + { + "bbox": [ + 46, + 201, + 547, + 250 + ], + "type": "inline_equation", + "content": "z^{*} \\in \\mathbb{R}^{1 \\times 3(2n + 1)^{2}}" + }, + { + "bbox": [ + 46, + 201, + 547, + 250 + ], + "type": "text", + "content": " and OPEs " + }, + { + "bbox": [ + 46, + 201, + 547, + 250 + ], + "type": "inline_equation", + "content": "\\in \\mathbb{R}^{3(2n + 1)^{2} \\times 1}" + }, + { + "bbox": [ + 46, + 201, + 547, + 250 + ], + "type": "text", + "content": " per color channel." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 55, + 266, + 279, + 379 + ], + "blocks": [ + { + "bbox": [ + 55, + 266, + 279, + 379 + ], + "lines": [ + { + "bbox": [ + 55, + 266, + 279, + 379 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 279, + 379 + ], + "type": "image", + "image_path": "d9549fbd46fbe99071b6b32e939af388d83fef11d090526cca8a42697b81241b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 382, + 288, + 428 + ], + "lines": [ + { + "bbox": [ + 46, + 382, + 288, + 428 + ], + "spans": [ + { + "bbox": [ + 46, + 382, + 288, + 428 + ], + "type": "text", + "content": "Figure 3. OPE-based patch representation. An image patch can be represented as a linear combination of basic plane waves in a continuous manner. The representation is a 2D extension to [46, Chap. 3.3.1]. Refer to the supplementary material for more details." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 446, + 287, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 446, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 446, + 287, + 544 + ], + "type": "text", + "content": "of Fourier transform to describe local information. To address this issue, we represent the input image as the seamless stitching of local patches, whose latent codes are extracted from a feature map over the channel dimension2. As shown in Fig. 2, the OPE-upscale module takes both the target resolution " + }, + { + "bbox": [ + 46, + 446, + 287, + 544 + ], + "type": "inline_equation", + "content": "H = r_h \\cdot h" + }, + { + "bbox": [ + 46, + 446, + 287, + 544 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 446, + 287, + 544 + ], + "type": "inline_equation", + "content": "W = r_w \\cdot w" + }, + { + "bbox": [ + 46, + 446, + 287, + 544 + ], + "type": "text", + "content": " and the feature map " + }, + { + "bbox": [ + 46, + 446, + 287, + 544 + ], + "type": "inline_equation", + "content": "\\in \\mathbb{R}^{h \\times w \\times 3(2n + 1)^2}" + }, + { + "bbox": [ + 46, + 446, + 287, + 544 + ], + "type": "text", + "content": " from the deep encoder " + }, + { + "bbox": [ + 46, + 446, + 287, + 544 + ], + "type": "inline_equation", + "content": "E_{\\theta}" + }, + { + "bbox": [ + 46, + 446, + 287, + 544 + ], + "type": "text", + "content": " as inputs and computes target pixels in parallel." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 544, + 288, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 544, + 288, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 544, + 288, + 677 + ], + "type": "text", + "content": "Feature map rendering. As shown in Fig. 4, to render a target image " + }, + { + "bbox": [ + 46, + 544, + 288, + 677 + ], + "type": "inline_equation", + "content": "I_{SR}" + }, + { + "bbox": [ + 46, + 544, + 288, + 677 + ], + "type": "text", + "content": " with size " + }, + { + "bbox": [ + 46, + 544, + 288, + 677 + ], + "type": "inline_equation", + "content": "H \\times W" + }, + { + "bbox": [ + 46, + 544, + 288, + 677 + ], + "type": "text", + "content": " from a LR image " + }, + { + "bbox": [ + 46, + 544, + 288, + 677 + ], + "type": "inline_equation", + "content": "I_{LR}" + }, + { + "bbox": [ + 46, + 544, + 288, + 677 + ], + "type": "text", + "content": " with size " + }, + { + "bbox": [ + 46, + 544, + 288, + 677 + ], + "type": "inline_equation", + "content": "h \\times w" + }, + { + "bbox": [ + 46, + 544, + 288, + 677 + ], + "type": "text", + "content": ", OPE-Upscale Module firstly divide a 2D domain " + }, + { + "bbox": [ + 46, + 544, + 288, + 677 + ], + "type": "inline_equation", + "content": "[-1,1] \\times [-1,1]" + }, + { + "bbox": [ + 46, + 544, + 288, + 677 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 46, + 544, + 288, + 677 + ], + "type": "inline_equation", + "content": "H \\times W" + }, + { + "bbox": [ + 46, + 544, + 288, + 677 + ], + "type": "text", + "content": " regions with equal size, so that every pixel in " + }, + { + "bbox": [ + 46, + 544, + 288, + 677 + ], + "type": "inline_equation", + "content": "I_{SR}" + }, + { + "bbox": [ + 46, + 544, + 288, + 677 + ], + "type": "text", + "content": " will be associated with an absolute central coordinates " + }, + { + "bbox": [ + 46, + 544, + 288, + 677 + ], + "type": "inline_equation", + "content": "(x_q, y_q)" + }, + { + "bbox": [ + 46, + 544, + 288, + 677 + ], + "type": "text", + "content": " in corresponding region. Secondly, the latent codes in the feature map (same dimension with " + }, + { + "bbox": [ + 46, + 544, + 288, + 677 + ], + "type": "inline_equation", + "content": "I_{LR}" + }, + { + "bbox": [ + 46, + 544, + 288, + 677 + ], + "type": "text", + "content": ") also possess corresponding central coordinates " + }, + { + "bbox": [ + 46, + 544, + 288, + 677 + ], + "type": "inline_equation", + "content": "(x_c, y_c) \\in [-1,1] \\times [-1,1]" + }, + { + "bbox": [ + 46, + 544, + 288, + 677 + ], + "type": "text", + "content": " by dividing same 2D domain into " + }, + { + "bbox": [ + 46, + 544, + 288, + 677 + ], + "type": "inline_equation", + "content": "h \\times w" + }, + { + "bbox": [ + 46, + 544, + 288, + 677 + ], + "type": "text", + "content": " regions, therefore, given a target image pixel with " + }, + { + "bbox": [ + 46, + 544, + 288, + 677 + ], + "type": "inline_equation", + "content": "(x_q, y_q)" + }, + { + "bbox": [ + 46, + 544, + 288, + 677 + ], + "type": "text", + "content": ", a specific latent code " + }, + { + "bbox": [ + 46, + 544, + 288, + 677 + ], + "type": "inline_equation", + "content": "z^* \\in \\mathbb{R}^{1 \\times 3(2n + 1)^2}" + }, + { + "bbox": [ + 46, + 544, + 288, + 677 + ], + "type": "text", + "content": " with" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 319, + 266, + 535, + 373 + ], + "blocks": [ + { + "bbox": [ + 319, + 266, + 535, + 373 + ], + "lines": [ + { + "bbox": [ + 319, + 266, + 535, + 373 + ], + "spans": [ + { + "bbox": [ + 319, + 266, + 535, + 373 + ], + "type": "image", + "image_path": "6f0579f2f542a32044fcfddc79d9a66ba99c8e42ec8c5e9feb5df76dd793ddb2.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 380, + 547, + 447 + ], + "lines": [ + { + "bbox": [ + 304, + 380, + 547, + 447 + ], + "spans": [ + { + "bbox": [ + 304, + 380, + 547, + 447 + ], + "type": "text", + "content": "Figure 4. Illustration of the mapping from the input LR image to the output SR image. LR image, feature map and SR image are all divided into the same domain " + }, + { + "bbox": [ + 304, + 380, + 547, + 447 + ], + "type": "inline_equation", + "content": "[-1,1] \\times [-1,1]" + }, + { + "bbox": [ + 304, + 380, + 547, + 447 + ], + "type": "text", + "content": ". Each LR pixel " + }, + { + "bbox": [ + 304, + 380, + 547, + 447 + ], + "type": "inline_equation", + "content": "I_{LR}(x_{LR},y_{LR})" + }, + { + "bbox": [ + 304, + 380, + 547, + 447 + ], + "type": "text", + "content": " corresponds to a latent code with coordinates " + }, + { + "bbox": [ + 304, + 380, + 547, + 447 + ], + "type": "inline_equation", + "content": "(x_c,y_c)" + }, + { + "bbox": [ + 304, + 380, + 547, + 447 + ], + "type": "text", + "content": ", while each SR pixel " + }, + { + "bbox": [ + 304, + 380, + 547, + 447 + ], + "type": "inline_equation", + "content": "I_{SR}(x_{SR},y_{SR})" + }, + { + "bbox": [ + 304, + 380, + 547, + 447 + ], + "type": "text", + "content": " corresponds to a latent code with coordinates " + }, + { + "bbox": [ + 304, + 380, + 547, + 447 + ], + "type": "inline_equation", + "content": "(x_q,y_q)" + }, + { + "bbox": [ + 304, + 380, + 547, + 447 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 465, + 547, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 465, + 547, + 514 + ], + "spans": [ + { + "bbox": [ + 304, + 465, + 547, + 514 + ], + "type": "text", + "content": "coordinates " + }, + { + "bbox": [ + 304, + 465, + 547, + 514 + ], + "type": "inline_equation", + "content": "(x_{c},y_{c})" + }, + { + "bbox": [ + 304, + 465, + 547, + 514 + ], + "type": "text", + "content": ", which has the smallest distance from " + }, + { + "bbox": [ + 304, + 465, + 547, + 514 + ], + "type": "inline_equation", + "content": "(x_{q},y_{q})" + }, + { + "bbox": [ + 304, + 465, + 547, + 514 + ], + "type": "text", + "content": " could be found. As shown in Eq.(6) and Eq.(7), a render function " + }, + { + "bbox": [ + 304, + 465, + 547, + 514 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 304, + 465, + 547, + 514 + ], + "type": "text", + "content": " takes two parts of inputs: " + }, + { + "bbox": [ + 304, + 465, + 547, + 514 + ], + "type": "inline_equation", + "content": "z^{*}" + }, + { + "bbox": [ + 304, + 465, + 547, + 514 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 465, + 547, + 514 + ], + "type": "inline_equation", + "content": "(x_q',y_q')" + }, + { + "bbox": [ + 304, + 465, + 547, + 514 + ], + "type": "text", + "content": ", to calculate final target pixel value as following:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 362, + 514, + 545, + 530 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 514, + 545, + 530 + ], + "spans": [ + { + "bbox": [ + 362, + 514, + 545, + 530 + ], + "type": "interline_equation", + "content": "I _ {S R} \\left(x _ {q}, y _ {q}\\right) = \\mathcal {R} \\left(z ^ {*}, \\left(x _ {q} ^ {\\prime}, y _ {q} ^ {\\prime}\\right)\\right) \\tag {6}", + "image_path": "f5b560c76d9cf7a07f6f76f98f1847838897302a06e90e96d13cf6edb9b2825d.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 340, + 540, + 545, + 555 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 540, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 340, + 540, + 545, + 555 + ], + "type": "interline_equation", + "content": "x _ {q} ^ {\\prime} = \\left(x _ {q} - x _ {c}\\right) \\cdot h, \\quad y _ {q} ^ {\\prime} = \\left(y _ {q} - y _ {c}\\right) \\cdot w \\tag {7}", + "image_path": "a007082a960d3755987657d3a0769c9169a9d75ec033ba68f713835e1ea30cf5.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 558, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 546, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 558, + 546, + 714 + ], + "type": "inline_equation", + "content": "z^{*}" + }, + { + "bbox": [ + 304, + 558, + 546, + 714 + ], + "type": "text", + "content": " is the nearest latent code we found, and " + }, + { + "bbox": [ + 304, + 558, + 546, + 714 + ], + "type": "inline_equation", + "content": "(x_{q}^{\\prime}, y_{q}^{\\prime})" + }, + { + "bbox": [ + 304, + 558, + 546, + 714 + ], + "type": "text", + "content": " are relative coordinates, which are calculated based on Eq.(7) to rescale the absolute coordinates (in domain " + }, + { + "bbox": [ + 304, + 558, + 546, + 714 + ], + "type": "inline_equation", + "content": "[-1, 1] \\times [-1, 1]" + }, + { + "bbox": [ + 304, + 558, + 546, + 714 + ], + "type": "text", + "content": ") by times " + }, + { + "bbox": [ + 304, + 558, + 546, + 714 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 304, + 558, + 546, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 558, + 546, + 714 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 304, + 558, + 546, + 714 + ], + "type": "text", + "content": ", which is taken as input by function " + }, + { + "bbox": [ + 304, + 558, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 304, + 558, + 546, + 714 + ], + "type": "text", + "content": " to render target pixel. " + }, + { + "bbox": [ + 304, + 558, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 304, + 558, + 546, + 714 + ], + "type": "text", + "content": " has the similar calculation as Eq.(5) while the difference is it repeats OPE three times to adapt " + }, + { + "bbox": [ + 304, + 558, + 546, + 714 + ], + "type": "inline_equation", + "content": "z^{*}" + }, + { + "bbox": [ + 304, + 558, + 546, + 714 + ], + "type": "text", + "content": " and calculate the linear combination per color channel. In this way, our OPE-upscale module successfully deals with arbitrary size " + }, + { + "bbox": [ + 304, + 558, + 546, + 714 + ], + "type": "inline_equation", + "content": "I_{SR}" + }, + { + "bbox": [ + 304, + 558, + 546, + 714 + ], + "type": "text", + "content": " by processing each pixel by " + }, + { + "bbox": [ + 304, + 558, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 304, + 558, + 546, + 714 + ], + "type": "text", + "content": ", in which feature map rendering process is parameter-free with high computing efficiency and less memory consumption (which has been confirmed by the experiments in Sec. 4.3)." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 46, + 683, + 288, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 683, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 683, + 288, + 714 + ], + "type": "text", + "content": "2To ensure compatibility with color images, we adjust the output channel of the encoder to " + }, + { + "bbox": [ + 46, + 683, + 288, + 714 + ], + "type": "inline_equation", + "content": "3(2n + 1)^{2}" + }, + { + "bbox": [ + 46, + 683, + 288, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 683, + 288, + 714 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 683, + 288, + 714 + ], + "type": "text", + "content": " is the pre-defined maximum frequency " + }, + { + "bbox": [ + 46, + 683, + 288, + 714 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 683, + 288, + 714 + ], + "type": "text", + "content": " of OPE." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "10012" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 45, + 72, + 289, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 72, + 289, + 194 + ], + "spans": [ + { + "bbox": [ + 45, + 72, + 289, + 194 + ], + "type": "text", + "content": "Patch ensemble. As shown in Fig. 4, when moving " + }, + { + "bbox": [ + 45, + 72, + 289, + 194 + ], + "type": "inline_equation", + "content": "(x_{q},y_{q})" + }, + { + "bbox": [ + 45, + 72, + 289, + 194 + ], + "type": "text", + "content": " from location " + }, + { + "bbox": [ + 45, + 72, + 289, + 194 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 45, + 72, + 289, + 194 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 45, + 72, + 289, + 194 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 45, + 72, + 289, + 194 + ], + "type": "text", + "content": ", the pixel value in the target image " + }, + { + "bbox": [ + 45, + 72, + 289, + 194 + ], + "type": "inline_equation", + "content": "I_{SR}" + }, + { + "bbox": [ + 45, + 72, + 289, + 194 + ], + "type": "text", + "content": " may change abruptly, as well as the nearest latent code " + }, + { + "bbox": [ + 45, + 72, + 289, + 194 + ], + "type": "inline_equation", + "content": "z^{*}" + }, + { + "bbox": [ + 45, + 72, + 289, + 194 + ], + "type": "text", + "content": ". To address this discontinuity issue, we propose a patch ensemble, which is a local ensemble styled interpolation using relative coordinates. Instead of finding a single nearest latent code for " + }, + { + "bbox": [ + 45, + 72, + 289, + 194 + ], + "type": "inline_equation", + "content": "(x_{q},y_{q})" + }, + { + "bbox": [ + 45, + 72, + 289, + 194 + ], + "type": "text", + "content": ", we select the nearest four neighbouring latent codes " + }, + { + "bbox": [ + 45, + 72, + 289, + 194 + ], + "type": "inline_equation", + "content": "z_{t}^{*}" + }, + { + "bbox": [ + 45, + 72, + 289, + 194 + ], + "type": "text", + "content": " with corresponding central coordinates " + }, + { + "bbox": [ + 45, + 72, + 289, + 194 + ], + "type": "inline_equation", + "content": "(x_{t},y_{t})" + }, + { + "bbox": [ + 45, + 72, + 289, + 194 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 45, + 72, + 289, + 194 + ], + "type": "inline_equation", + "content": "t\\in \\{00,01,10,11\\}" + }, + { + "bbox": [ + 45, + 72, + 289, + 194 + ], + "type": "text", + "content": ". Then we calculate the relative coordinates " + }, + { + "bbox": [ + 45, + 72, + 289, + 194 + ], + "type": "inline_equation", + "content": "x_{q}^{\\prime}" + }, + { + "bbox": [ + 45, + 72, + 289, + 194 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 45, + 72, + 289, + 194 + ], + "type": "inline_equation", + "content": "y_{q}^{\\prime}" + }, + { + "bbox": [ + 45, + 72, + 289, + 194 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 78, + 197, + 287, + 221 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 197, + 287, + 221 + ], + "spans": [ + { + "bbox": [ + 78, + 197, + 287, + 221 + ], + "type": "interline_equation", + "content": "x _ {q} ^ {\\prime} = \\frac {\\left(x _ {q} - x _ {t}\\right) \\cdot h}{2}, \\quad y _ {q} ^ {\\prime} = \\frac {\\left(y _ {q} - y _ {t}\\right) \\cdot w}{2}, \\tag {8}", + "image_path": "6b3c9d605621528c9958317a3e570ac3ffa04f91ec81cb4eabb133a159bc8c64.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 223, + 287, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 223, + 287, + 259 + ], + "spans": [ + { + "bbox": [ + 47, + 223, + 287, + 259 + ], + "type": "text", + "content": "which can guarantee " + }, + { + "bbox": [ + 47, + 223, + 287, + 259 + ], + "type": "inline_equation", + "content": "x_{q}^{\\prime}, y_{q}^{\\prime} \\in [-1, 1]" + }, + { + "bbox": [ + 47, + 223, + 287, + 259 + ], + "type": "text", + "content": ". As the inset shows, for the four adjacent pixels from the low resolution image " + }, + { + "bbox": [ + 47, + 223, + 287, + 259 + ], + "type": "inline_equation", + "content": "I_{LR}" + }, + { + "bbox": [ + 47, + 223, + 287, + 259 + ], + "type": "text", + "content": " (i.e. related to latent codes " + }, + { + "bbox": [ + 47, + 223, + 287, + 259 + ], + "type": "inline_equation", + "content": "z_{00}^{*}, z_{01}^{*}, z_{10}^{*}" + }, + { + "bbox": [ + 47, + 223, + 287, + 259 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 223, + 287, + 259 + ], + "type": "inline_equation", + "content": "z_{11}^{*}" + }, + { + "bbox": [ + 47, + 223, + 287, + 259 + ], + "type": "text", + "content": ")," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 259, + 184, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 259, + 184, + 366 + ], + "spans": [ + { + "bbox": [ + 47, + 259, + 184, + 366 + ], + "type": "text", + "content": "their corresponding patches in super resolution image " + }, + { + "bbox": [ + 47, + 259, + 184, + 366 + ], + "type": "inline_equation", + "content": "I_{SR}" + }, + { + "bbox": [ + 47, + 259, + 184, + 366 + ], + "type": "text", + "content": " are colored in red, green, yellow and blue, respectively. The pixel color in " + }, + { + "bbox": [ + 47, + 259, + 184, + 366 + ], + "type": "inline_equation", + "content": "I_{SR}" + }, + { + "bbox": [ + 47, + 259, + 184, + 366 + ], + "type": "text", + "content": " is not solely dependent on the nearest latent code but considers the four neighboring latent codes. Specifically, using the rendering function " + }, + { + "bbox": [ + 47, + 259, + 184, + 366 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 47, + 259, + 184, + 366 + ], + "type": "text", + "content": " and the di" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 187, + 261, + 286, + 363 + ], + "blocks": [ + { + "bbox": [ + 187, + 261, + 286, + 363 + ], + "lines": [ + { + "bbox": [ + 187, + 261, + 286, + 363 + ], + "spans": [ + { + "bbox": [ + 187, + 261, + 286, + 363 + ], + "type": "image", + "image_path": "ed462df17db4a6c8995f53e1017ef18600251ebc53569198c6eeee7aa186f906.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 367, + 287, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 367, + 287, + 390 + ], + "spans": [ + { + "bbox": [ + 47, + 367, + 287, + 390 + ], + "type": "text", + "content": "aagonal rectangle areas " + }, + { + "bbox": [ + 47, + 367, + 287, + 390 + ], + "type": "inline_equation", + "content": "s_t" + }, + { + "bbox": [ + 47, + 367, + 287, + 390 + ], + "type": "text", + "content": " as weights, we compute the pixel value as a weighted sum" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 59, + 393, + 287, + 422 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 393, + 287, + 422 + ], + "spans": [ + { + "bbox": [ + 59, + 393, + 287, + 422 + ], + "type": "interline_equation", + "content": "I _ {S R} \\left(x _ {q}, y _ {q}\\right) = \\sum_ {t \\in \\{0 0, 0 1, 1 0, 1 1 \\}} \\frac {s _ {t}}{S} \\cdot \\mathcal {R} \\left(z _ {t} ^ {*}, \\left(x _ {q} ^ {\\prime}, y _ {q} ^ {\\prime}\\right)\\right), \\tag {9}", + "image_path": "49e2a686a14ce8b8b5340aa4688a91cb8a2df518a6cb0e294c79d831139d0310.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 425, + 287, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 425, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 47, + 425, + 287, + 498 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 425, + 287, + 498 + ], + "type": "inline_equation", + "content": "S = \\sum_{t} s_{t}" + }, + { + "bbox": [ + 47, + 425, + 287, + 498 + ], + "type": "text", + "content": " is the sum of areas. Considering the contribution of each latent code allows us to integrate the adjacent latent codes with different significance, thereby providing a seamless stitching of adjacent patches. We call Eq.(9) local ensemble styled interpolation since it takes a similar form of the local ensemble in LIIF [5]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 502, + 180, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 502, + 180, + 516 + ], + "spans": [ + { + "bbox": [ + 47, + 502, + 180, + 516 + ], + "type": "text", + "content": "3.3. Maximum Frequency " + }, + { + "bbox": [ + 47, + 502, + 180, + 516 + ], + "type": "inline_equation", + "content": "n" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 521, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 521, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 521, + 289, + 715 + ], + "type": "text", + "content": "Selecting a proper maximum frequency " + }, + { + "bbox": [ + 47, + 521, + 289, + 715 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 47, + 521, + 289, + 715 + ], + "type": "text", + "content": " plays an important role in designing and implementing the OPE-upsampling module since it directly determines the network architecture and also has effects on the performance of different SR scales. Given a high resolution image " + }, + { + "bbox": [ + 47, + 521, + 289, + 715 + ], + "type": "inline_equation", + "content": "I_{HR}" + }, + { + "bbox": [ + 47, + 521, + 289, + 715 + ], + "type": "text", + "content": " with size " + }, + { + "bbox": [ + 47, + 521, + 289, + 715 + ], + "type": "inline_equation", + "content": "H \\times W" + }, + { + "bbox": [ + 47, + 521, + 289, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 521, + 289, + 715 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 47, + 521, + 289, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 521, + 289, + 715 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 47, + 521, + 289, + 715 + ], + "type": "text", + "content": ", we aim to obtain a feature map with size " + }, + { + "bbox": [ + 47, + 521, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\frac{H}{r} \\times \\frac{W}{r}" + }, + { + "bbox": [ + 47, + 521, + 289, + 715 + ], + "type": "text", + "content": ", then we re-render the obtained feature map with the selected " + }, + { + "bbox": [ + 47, + 521, + 289, + 715 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 47, + 521, + 289, + 715 + ], + "type": "text", + "content": ". By the comparison of the rendered " + }, + { + "bbox": [ + 47, + 521, + 289, + 715 + ], + "type": "inline_equation", + "content": "I_{SR}" + }, + { + "bbox": [ + 47, + 521, + 289, + 715 + ], + "type": "text", + "content": ", we present the performance of " + }, + { + "bbox": [ + 47, + 521, + 289, + 715 + ], + "type": "inline_equation", + "content": "n \\in \\{1,2,\\dots,8\\}" + }, + { + "bbox": [ + 47, + 521, + 289, + 715 + ], + "type": "text", + "content": " under different " + }, + { + "bbox": [ + 47, + 521, + 289, + 715 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 47, + 521, + 289, + 715 + ], + "type": "text", + "content": " values (SR scale), as show in Tab. 1, and select the " + }, + { + "bbox": [ + 47, + 521, + 289, + 715 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 47, + 521, + 289, + 715 + ], + "type": "text", + "content": " with the best performance (the details would be discussed Sec. 4.1). To be specific, we use Eq.(10) as the basic theory and use Eq.(11) to infer the feature map. First, similar to calculate the projection of a normal vector on orthogonal basis, we can calculate projections (or so-called latent code) " + }, + { + "bbox": [ + 47, + 521, + 289, + 715 + ], + "type": "inline_equation", + "content": "Z \\in \\mathbb{R}^{1 \\times (2n + 1)^2}" + }, + { + "bbox": [ + 47, + 521, + 289, + 715 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 47, + 521, + 289, + 715 + ], + "type": "inline_equation", + "content": "f(x,y)" + }, + { + "bbox": [ + 47, + 521, + 289, + 715 + ], + "type": "text", + "content": " in Eq.(5) as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 341, + 81, + 545, + 110 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 81, + 545, + 110 + ], + "spans": [ + { + "bbox": [ + 341, + 81, + 545, + 110 + ], + "type": "interline_equation", + "content": "Z [ i ] = \\frac {1}{4} \\int_ {- 1} ^ {1} \\int_ {- 1} ^ {1} f (x, y) P [ i ] (x, y) \\mathrm {d} x \\mathrm {d} y \\tag {10}", + "image_path": "7b20f9a1220b924e975a48df96997b3771290e1a86f86018826c41477d6f78ea.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 114, + 547, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 114, + 547, + 186 + ], + "spans": [ + { + "bbox": [ + 304, + 114, + 547, + 186 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 114, + 547, + 186 + ], + "type": "inline_equation", + "content": "P[i](x,y)" + }, + { + "bbox": [ + 304, + 114, + 547, + 186 + ], + "type": "text", + "content": " is a bivariate function taken from the " + }, + { + "bbox": [ + 304, + 114, + 547, + 186 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 114, + 547, + 186 + ], + "type": "text", + "content": "-th position of OPE and " + }, + { + "bbox": [ + 304, + 114, + 547, + 186 + ], + "type": "inline_equation", + "content": "Z[i]" + }, + { + "bbox": [ + 304, + 114, + 547, + 186 + ], + "type": "text", + "content": " is the corresponding projection. Based on Eq.(10) and taking both the discreteness of an image and the design of OPE-Upscale Module into consideration, we calculate the feature map of an image " + }, + { + "bbox": [ + 304, + 114, + 547, + 186 + ], + "type": "inline_equation", + "content": "I_{HR}" + }, + { + "bbox": [ + 304, + 114, + 547, + 186 + ], + "type": "text", + "content": " with down-sampling scale " + }, + { + "bbox": [ + 304, + 114, + 547, + 186 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 304, + 114, + 547, + 186 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 341, + 194, + 545, + 228 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 194, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 341, + 194, + 545, + 228 + ], + "type": "interline_equation", + "content": "z ^ {*} [ i ] = \\frac {1}{4} \\sum_ {x ^ {\\prime}} ^ {2 r} \\sum_ {y ^ {\\prime}} ^ {2 r} I _ {H R} \\left(x ^ {\\prime}, y ^ {\\prime}\\right) P [ i ] \\left(x ^ {\\prime}, y ^ {\\prime}\\right) \\tag {11}", + "image_path": "7ecc6177bb610de5fb049edef1c97a2221e23f5949f3d561ea9b560ceb9f4f59.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 233, + 538, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 233, + 538, + 246 + ], + "spans": [ + { + "bbox": [ + 305, + 233, + 538, + 246 + ], + "type": "text", + "content": "It can be considered as the inverse operation of Eq.(9)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 246, + 443, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 246, + 443, + 353 + ], + "spans": [ + { + "bbox": [ + 305, + 246, + 443, + 353 + ], + "type": "text", + "content": "Take the right inset as an example. We choose the HR image as the ground truth (e.g. HR in Fig. 5), when " + }, + { + "bbox": [ + 305, + 246, + 443, + 353 + ], + "type": "inline_equation", + "content": "r = 4" + }, + { + "bbox": [ + 305, + 246, + 443, + 353 + ], + "type": "text", + "content": ", every latent code " + }, + { + "bbox": [ + 305, + 246, + 443, + 353 + ], + "type": "inline_equation", + "content": "z^{*}" + }, + { + "bbox": [ + 305, + 246, + 443, + 353 + ], + "type": "text", + "content": " corresponds to a " + }, + { + "bbox": [ + 305, + 246, + 443, + 353 + ], + "type": "inline_equation", + "content": "8 \\times 8" + }, + { + "bbox": [ + 305, + 246, + 443, + 353 + ], + "type": "text", + "content": " patch of HR (gray points) in relative coordinate domain (blue region). To calculate the " + }, + { + "bbox": [ + 305, + 246, + 443, + 353 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 305, + 246, + 443, + 353 + ], + "type": "text", + "content": "-th position of " + }, + { + "bbox": [ + 305, + 246, + 443, + 353 + ], + "type": "inline_equation", + "content": "z^{*}" + }, + { + "bbox": [ + 305, + 246, + 443, + 353 + ], + "type": "text", + "content": ", we multiply every HR pixel value" + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 446, + 247, + 545, + 350 + ], + "blocks": [ + { + "bbox": [ + 446, + 247, + 545, + 350 + ], + "lines": [ + { + "bbox": [ + 446, + 247, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 446, + 247, + 545, + 350 + ], + "type": "image", + "image_path": "c3e0f18e5283550c5f3e197ca6920b72751ad2be79e0e5b1afd13585496f518a.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 353, + 545, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 353, + 545, + 401 + ], + "spans": [ + { + "bbox": [ + 305, + 353, + 545, + 401 + ], + "type": "inline_equation", + "content": "I_{HR}(x',y')" + }, + { + "bbox": [ + 305, + 353, + 545, + 401 + ], + "type": "text", + "content": " and basis value " + }, + { + "bbox": [ + 305, + 353, + 545, + 401 + ], + "type": "inline_equation", + "content": "P[i](x',y')" + }, + { + "bbox": [ + 305, + 353, + 545, + 401 + ], + "type": "text", + "content": " together and finally sum them. After getting the feature map, we render it to the same size of " + }, + { + "bbox": [ + 305, + 353, + 545, + 401 + ], + "type": "inline_equation", + "content": "I_{HR}" + }, + { + "bbox": [ + 305, + 353, + 545, + 401 + ], + "type": "text", + "content": " via OPE-Upscale Module and calculate their Peak Signal-to-Noise Ratio (PSNR)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 411, + 388, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 411, + 388, + 426 + ], + "spans": [ + { + "bbox": [ + 306, + 411, + 388, + 426 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 431, + 414, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 431, + 414, + 445 + ], + "spans": [ + { + "bbox": [ + 306, + 431, + 414, + 445 + ], + "type": "text", + "content": "4.1. Parameter Setting" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 450, + 545, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 450, + 545, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 545, + 641 + ], + "type": "text", + "content": "We evaluated the performance of different maximum frequencies " + }, + { + "bbox": [ + 304, + 450, + 545, + 641 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 450, + 545, + 641 + ], + "type": "text", + "content": " on 50 images from the DIV2K validation set [1] under different scales " + }, + { + "bbox": [ + 304, + 450, + 545, + 641 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 304, + 450, + 545, + 641 + ], + "type": "text", + "content": ". Since our method is a local representation, we do not use a large " + }, + { + "bbox": [ + 304, + 450, + 545, + 641 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 450, + 545, + 641 + ], + "type": "text", + "content": ". As shown in Tab. 1, the optimal sampling frequency for a given " + }, + { + "bbox": [ + 304, + 450, + 545, + 641 + ], + "type": "inline_equation", + "content": "r_i" + }, + { + "bbox": [ + 304, + 450, + 545, + 641 + ], + "type": "text", + "content": " is always " + }, + { + "bbox": [ + 304, + 450, + 545, + 641 + ], + "type": "inline_equation", + "content": "r_i - 1" + }, + { + "bbox": [ + 304, + 450, + 545, + 641 + ], + "type": "text", + "content": ". This observation can also be explained by the Nyquist-Shannon sampling theorem. For example, when " + }, + { + "bbox": [ + 304, + 450, + 545, + 641 + ], + "type": "inline_equation", + "content": "r_i = 4" + }, + { + "bbox": [ + 304, + 450, + 545, + 641 + ], + "type": "text", + "content": ", there are " + }, + { + "bbox": [ + 304, + 450, + 545, + 641 + ], + "type": "inline_equation", + "content": "8 \\times 8" + }, + { + "bbox": [ + 304, + 450, + 545, + 641 + ], + "type": "text", + "content": " sampling points for every latent code to \"fit\", so the maximum frequency that can be recovered from these sampling points should be less than 4. We also tested larger frequency with " + }, + { + "bbox": [ + 304, + 450, + 545, + 641 + ], + "type": "inline_equation", + "content": "r_i \\leq n \\leq 2 \\times r_i" + }, + { + "bbox": [ + 304, + 450, + 545, + 641 + ], + "type": "text", + "content": ", reaching the upper limit that equals the number of sampling points. We further visualize the reconstructed images for different " + }, + { + "bbox": [ + 304, + 450, + 545, + 641 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 450, + 545, + 641 + ], + "type": "text", + "content": ". As shown in Fig. 5, with scale factor " + }, + { + "bbox": [ + 304, + 450, + 545, + 641 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 304, + 450, + 545, + 641 + ], + "type": "text", + "content": ", the larger frequency (" + }, + { + "bbox": [ + 304, + 450, + 545, + 641 + ], + "type": "inline_equation", + "content": "n > 3" + }, + { + "bbox": [ + 304, + 450, + 545, + 641 + ], + "type": "text", + "content": ") brings redundant high-frequency information that sharpens the resulting images." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": "Based on the above analysis, we decide to choose " + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "inline_equation", + "content": "n = 3" + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": " as the maximum frequency for our OPE-Upscale Module. This is because existing arbitrary-scale SR methods, such as LIIF [5] and LTE [32], are trained with random scale factors up to 4, and the frequency " + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "inline_equation", + "content": "n = 4 - 1 = 3" + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": " is sufficient to fully capture the ground truth information. Al-" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10013" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 51, + 70, + 282, + 162 + ], + "blocks": [ + { + "bbox": [ + 51, + 70, + 282, + 162 + ], + "lines": [ + { + "bbox": [ + 51, + 70, + 282, + 162 + ], + "spans": [ + { + "bbox": [ + 51, + 70, + 282, + 162 + ], + "type": "table", + "html": "
n×2×3×4×5×6×7×8
131.195128.608326.442425.048524.111423.342322.7898
230.747233.658631.209128.802227.370126.191325.3838
322.187133.658535.198332.401130.613528.896427.8159
412.123028.608334.963134.629434.097931.446230.2865
5-22.846529.951234.629337.370433.728532.8190
6-22.846524.312232.401137.150635.304635.8250
7--19.159328.802233.303935.304639.1160
8--12.086325.048529.096633.728638.9863
", + "image_path": "39c9f50803cba958c9d706c10d3701d5a21ea936fd7647cc5a97fdfd6276e8f5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 165, + 287, + 187 + ], + "lines": [ + { + "bbox": [ + 47, + 165, + 287, + 187 + ], + "spans": [ + { + "bbox": [ + 47, + 165, + 287, + 187 + ], + "type": "text", + "content": "Table 1. Representation performance (PSNR (dB)). The best value for each upsampling factor is bolded." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 51, + 196, + 286, + 335 + ], + "blocks": [ + { + "bbox": [ + 51, + 196, + 286, + 335 + ], + "lines": [ + { + "bbox": [ + 51, + 196, + 286, + 335 + ], + "spans": [ + { + "bbox": [ + 51, + 196, + 286, + 335 + ], + "type": "image", + "image_path": "1ed6aacf9d28f0b5fd6221f6ccc37ab98dd42e8b9689c7eff37379a229040c83.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 337, + 287, + 392 + ], + "lines": [ + { + "bbox": [ + 46, + 337, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 46, + 337, + 287, + 392 + ], + "type": "text", + "content": "Figure 5. Qualitative comparison of different OPE frequency n under scale factor " + }, + { + "bbox": [ + 46, + 337, + 287, + 392 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 46, + 337, + 287, + 392 + ], + "type": "text", + "content": " (PSNR (dB)). 1-th row: residuals from " + }, + { + "bbox": [ + 46, + 337, + 287, + 392 + ], + "type": "inline_equation", + "content": "n = 0" + }, + { + "bbox": [ + 46, + 337, + 287, + 392 + ], + "type": "text", + "content": " in image time domain. 2-th row: fourier frequency domain of HR and rendered image with " + }, + { + "bbox": [ + 46, + 337, + 287, + 392 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 337, + 287, + 392 + ], + "type": "text", + "content": ". 3-th row: HR image and rendered image with " + }, + { + "bbox": [ + 46, + 337, + 287, + 392 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 337, + 287, + 392 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 416, + 287, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 416, + 287, + 489 + ], + "spans": [ + { + "bbox": [ + 46, + 416, + 287, + 489 + ], + "type": "text", + "content": "though a larger " + }, + { + "bbox": [ + 46, + 416, + 287, + 489 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 416, + 287, + 489 + ], + "type": "text", + "content": " could represent more detailed patches of the target SR image, it would also introduce redundant high-frequency information and potentially increase computation time and memory consumption during training. Therefore, we choose " + }, + { + "bbox": [ + 46, + 416, + 287, + 489 + ], + "type": "inline_equation", + "content": "n = 3" + }, + { + "bbox": [ + 46, + 416, + 287, + 489 + ], + "type": "text", + "content": " as the maximum frequency for our OPE-Upscale Module to balance performance and efficiency." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 500, + 111, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 500, + 111, + 513 + ], + "spans": [ + { + "bbox": [ + 47, + 500, + 111, + 513 + ], + "type": "text", + "content": "4.2. Training" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 520, + 286, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 520, + 286, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 520, + 286, + 581 + ], + "type": "text", + "content": "Datasets. Similar to [5, 32], we use the DIV2K dataset [1] of the NTIRE 2017 Challenge [63] for training. For testing, we use the DIV2K validation set [1] with 100 images and four benchmark datasets: Set5 [4], Set14 [76], B100 [40], and Urban100 [21]. We use PSNR as the quality measure." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 582, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 286, + 713 + ], + "type": "text", + "content": "Implementation details. We mainly follow the prior implementation [5, 32] for arbitrary-scale SR training after replacing their upsampling module with OPE-Upscale. We use EDSR-baseline [35] and RDN [80] without their upsampling modules as the encoder, which is the only trainable part of our network. We use " + }, + { + "bbox": [ + 46, + 582, + 286, + 713 + ], + "type": "inline_equation", + "content": "48 \\times 48" + }, + { + "bbox": [ + 46, + 582, + 286, + 713 + ], + "type": "text", + "content": " patches cropped from training set as inputs, L1 loss and Adam [26] optimizer for optimization. The network was trained for 1000 epochs with batch size 16, while the initial learning rate is 1e-4 and decayed by factor 0.5 every 200 epochs. More implementation details are presented in supplementary material." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 72, + 380, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 380, + 83 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 380, + 83 + ], + "type": "text", + "content": "4.3. Evaluation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 91, + 545, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 545, + 365 + ], + "type": "text", + "content": "Quantitative results. Tab. 2 and Tab. 3 report quantitative results of OPE and the SOTA arbitrary-scale SR methods on the DIV2K validation set and the benchmark datasets. It is worth noting that we focus on finding an alternative of MLP with position encoding, rather than enhancing it like LTE [32]. We observe that our method achieves comparable results (less than 0.1dB on DIV2K and less than 0.15dB on benchmark), which indicates that our method is a feasible analytical solution with good performance and efficient parameter-free module. As shown in Tab. 2, EDSR [35] and RDN [80] are our selected encoders, and we achieve the highest efficiency (i.e. the shortest inference time in red number) comparing to all the other baselines with both encoders. The higher the scale factor, the better result we achieve. Specifically, in out-scale SR (" + }, + { + "bbox": [ + 304, + 91, + 545, + 365 + ], + "type": "inline_equation", + "content": "\\times 6" + }, + { + "bbox": [ + 304, + 91, + 545, + 365 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 91, + 545, + 365 + ], + "type": "inline_equation", + "content": "\\times 30" + }, + { + "bbox": [ + 304, + 91, + 545, + 365 + ], + "type": "text", + "content": "), our method outperforms most baselines and just has a small gap with LTE (less than 0.1dB). Such results demonstrate that our method has rich representation capability. We also compared with the benchmark dataset, as shown in Tab. 3, we keep comparable results to baselines (the gap is less than 0.15dB). However, as a nonlinear representation method, MLP still has advantages over our linear representation with low scale factors. See Sec. 5 for discussion on this issue." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 366, + 545, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 366, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 366, + 545, + 521 + ], + "type": "text", + "content": "Qualitative results. Fig. 6 provides qualitative results with SOTA methods by using different scale factor. We show competitive visual quality against others, more results are provided in supplementary material. From the local perspective, LIIF and LTE only generate smooth patches, while our OPE with max frequency 3 is enough to achieve similar visual quality. We also notice LIIF [5] has artifact (vertical stripes) in the 1st row, this is a common drawback for implicit neural representation and is hard to be explained. However, with our image representation, there is no artifacts. In the 2nd row, we could observe a sprout (in red rectangle) in the GT, the same region of LIIF is vanished, and the boundary of our sprout is more obvious than LTE." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "type": "text", + "content": "Computing efficiency. We measure computing efficiency with MACs (multiply-accumulate operations), FLOPs (floating point operations) and actual running time. In Tab. 4 column 2-3, judged by the time complexity measured by the number of operations, we save 2 orders of magnitude. In our upsampling module, there is only one matrix operation and essential position encoding between input and output. In Tab. 2 we show shortest inference time benefiting from our compact SR framework. To further demonstrate our time advantage on large size images, we take " + }, + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "type": "text", + "content": " as LR input of encoder and calculate time consumption of upsampling module with scale factor " + }, + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\times 4 \\times 30" + }, + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "type": "text", + "content": " on NVIDIA RTX 3090. As shown in Tab. 5, our upsampling module shows " + }, + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "type": "inline_equation", + "content": "26\\% - 57\\%" + }, + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "type": "text", + "content": " time advantage, this advantage keeps growing with larger scale factor. Notice We do not take advantage of GPU acceleration to design the upsampling" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10014" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 545, + 232 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 545, + 232 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 545, + 232 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 545, + 232 + ], + "type": "image", + "image_path": "aac223839dec03f1e84574ef5e7b063b56f045530bb6f881d9a6206fcde2bbd2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 70, + 236, + 520, + 246 + ], + "lines": [ + { + "bbox": [ + 70, + 236, + 520, + 246 + ], + "spans": [ + { + "bbox": [ + 70, + 236, + 520, + 246 + ], + "type": "text", + "content": "Figure 6. Qualitative comparison with SOTA methods for arbitrary-scale SR. RDN [80] is used as encoder for all methods." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 55, + 257, + 536, + 369 + ], + "blocks": [ + { + "bbox": [ + 55, + 257, + 536, + 369 + ], + "lines": [ + { + "bbox": [ + 55, + 257, + 536, + 369 + ], + "spans": [ + { + "bbox": [ + 55, + 257, + 536, + 369 + ], + "type": "table", + "html": "
MethodIn-scaleOut-scale
×2×3×4×6×12×18×24×30
Bicubic [35]31.0128.2226.6624.8222.2721.0020.1919.59
EDSR-baseline [35]34.5530.9028.94-----
EDSR-baseline-MetaSR# [5, 18]34.6430.9328.9226.6123.5522.0321.0620.37
EDSR-baseline-LIIF [5]34.67 / 170230.96 / 127729.00 / 114426.75 / 104623.71 / 96522.17 / 95321.18 / 95120.48 / 947
EDSR-baseline-LTE [32]34.72 / 115831.02 / 107929.04 / 104526.81 / 102323.78 / 100722.23 / 100521.24 / 100320.53 / 1000
EDSR-baseline-OPE (ours)34.34 / 47630.94 / 39529.02 / 36426.77 / 34823.74 / 32222.21 / 31821.21 / 31420.52 / 311
RDN-MetaSR# [5, 18]35.0031.2729.2526.8823.7322.1821.1720.47
RDN-LIIF [5]34.99 / 310731.26 / 207329.27 / 151326.99 / 124823.89 / 102522.34 / 99421.31 / 99120.59 / 972
RDN-LTE [32]35.04 / 254931.32 / 183929.33 / 142027.04 / 118423.95 / 104922.40 / 102721.36 / 102520.64 / 1014
RDN-OPE (ours)34.52 / 227731.17 / 149729.26 / 103926.98 / 81323.91 / 66322.36 / 62321.34 / 59620.63 / 590
", + "image_path": "2220f6947053ac6a8124e35c4df6a4bec0be982a1da05f288c3f132b3b19f8fc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 370, + 546, + 403 + ], + "lines": [ + { + "bbox": [ + 46, + 370, + 546, + 403 + ], + "spans": [ + { + "bbox": [ + 46, + 370, + 546, + 403 + ], + "type": "text", + "content": "Table 2. Quantitative comparison with the SOTA methods for arbitrary-scale SR on the DIV2K validation set (PSNR (dB) / running time (ms per image)). " + }, + { + "bbox": [ + 46, + 370, + 546, + 403 + ], + "type": "inline_equation", + "content": "\\sharp" + }, + { + "bbox": [ + 46, + 370, + 546, + 403 + ], + "type": "text", + "content": " indicates implementation in LIIF [5]. With a parameter-free upsampling module, we narrow the gap between SOTA and ours in most results less than 0.1dB (blue) and obtain the shortest inference time (red)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 425, + 287, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 425, + 287, + 462 + ], + "spans": [ + { + "bbox": [ + 46, + 425, + 287, + 462 + ], + "type": "text", + "content": "module carefully, with hardware optimization, we believe our time advantage could be much larger thanks to fewer number of operations required." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 475, + 287, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 475, + 287, + 608 + ], + "spans": [ + { + "bbox": [ + 46, + 475, + 287, + 608 + ], + "type": "text", + "content": "Memory consumption. In Tab. 4 column 4-5 we compare GPU memory consumption of OPE-Upscale Module with LIIF [5] and LTE [32] under training mode and testing mode of Pytorch [48]. For training mode, we use a " + }, + { + "bbox": [ + 46, + 475, + 287, + 608 + ], + "type": "inline_equation", + "content": "48 \\times 48" + }, + { + "bbox": [ + 46, + 475, + 287, + 608 + ], + "type": "text", + "content": " patch as input and sample 2304 pixels as output following the default training strategy in arbitrary-scale SR works. For testing mode, we use " + }, + { + "bbox": [ + 46, + 475, + 287, + 608 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 46, + 475, + 287, + 608 + ], + "type": "text", + "content": " image as input with scale factor 4. As a interpretable image representation without network parameters, OPE-Upscale Module saves memory of intermediate data (e.g. gradients, hidden layer outputs), and this advantage is fully reflected in training mode." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 621, + 288, + 718 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 621, + 288, + 718 + ], + "spans": [ + { + "bbox": [ + 46, + 621, + 288, + 718 + ], + "type": "text", + "content": "Flipping consistency. As described in Sec. 1, the INR-based upsampling module like [5] is sensitive for the flipping of feature map. However, our method solves this problem completely and elegantly. The orthogonal basis of OPE is based on symmetric sinusoidal function, which leads to advantage of our method for keeping the flipping consistency. Also, more samples are provided in supplementary material for verifying other more flipping transforms." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 308, + 422, + 543, + 491 + ], + "blocks": [ + { + "bbox": [ + 308, + 422, + 543, + 491 + ], + "lines": [ + { + "bbox": [ + 308, + 422, + 543, + 491 + ], + "spans": [ + { + "bbox": [ + 308, + 422, + 543, + 491 + ], + "type": "image", + "image_path": "f8bf4b45e4adbc2597c95a129188961df8edd51595c6d51117cc75e7a18d9586.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 493, + 545, + 514 + ], + "lines": [ + { + "bbox": [ + 305, + 493, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 305, + 493, + 545, + 514 + ], + "type": "text", + "content": "Figure 7. Visual results of the ablation study on patch ensemble. See the text for details." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 537, + 406, + 549 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 537, + 406, + 549 + ], + "spans": [ + { + "bbox": [ + 306, + 537, + 406, + 549 + ], + "type": "text", + "content": "4.4. Ablation Studies" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 556, + 545, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 556, + 545, + 664 + ], + "spans": [ + { + "bbox": [ + 304, + 556, + 545, + 664 + ], + "type": "text", + "content": "In order to examine the effects of local ensemble styled interpolation (I)-Eq.(9) and extension of relative coordinate domain (E)-Eq.(8) in patch ensemble, we conducted experiments using four different settings with EDSR-baseline as the encoder. The four settings were: 1) OPE: OPE-Upscale module with I and E; 2) OPE-E: OPE-Upscale module without E but with I; 3) OPE-I: OPE-Upscale module without I but with E; and 4) OPE-IE: OPE-Upscale module with neither I nor E (that is, without patch ensemble)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 665, + 546, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 546, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 546, + 712 + ], + "type": "text", + "content": "Fig. 7 and Tab. 6 present the comparison results of different settings. In the OPE-E setting, only the nearest latent code to the query point can be rescaled to " + }, + { + "bbox": [ + 304, + 665, + 546, + 712 + ], + "type": "inline_equation", + "content": "[-1,1] \\times [-1,1]" + }, + { + "bbox": [ + 304, + 665, + 546, + 712 + ], + "type": "text", + "content": ", while the relative coordinates of the other three latent codes" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10015" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 70, + 548, + 149 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 548, + 149 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 548, + 149 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 548, + 149 + ], + "type": "table", + "html": "
MethodSet5Set14B100Urban100
In-scaleOut-scaleIn-scaleOut-scaleOut-scaleIn-scaleOut-scale
×2×3×4×6×8×2×3×4×6×8×2×3×4×6×8×2×3
RDN [80]38.2434.7132.47--34.0130.5728.81--32.3429.2627.72--32.8928.80
RDN-MetaSR# [5, 18]38.2234.6332.3829.0426.9633.9830.5428.7826.5124.9732.3329.2627.7125.9024.8332.9228.82
RDN-LIIF [5]38.1734.6832.5029.1527.1433.9730.5328.8026.6425.1532.3229.2627.7425.9824.9132.8728.82
RDN-LTE [32]38.2334.7232.6129.3227.2634.0930.5828.8826.7125.1632.3629.3027.7726.0124.9533.0428.97
RDN-OPE (ours)37.6034.5932.4729.1727.2233.3930.4928.8026.6525.1732.0529.1927.7225.9624.9131.7828.63
", + "image_path": "06fa2f2c00a39af62efcf9d0f86c02a1b30212e86aa9548e26dc837ae6ce897a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 52, + 194, + 282, + 236 + ], + "blocks": [ + { + "bbox": [ + 46, + 152, + 547, + 187 + ], + "lines": [ + { + "bbox": [ + 46, + 152, + 547, + 187 + ], + "spans": [ + { + "bbox": [ + 46, + 152, + 547, + 187 + ], + "type": "text", + "content": "Table 3. Quantitative comparison with SOTA methods for arbitrary-scale image SR on benchmark datasets (PSNR (dB)). " + }, + { + "bbox": [ + 46, + 152, + 547, + 187 + ], + "type": "inline_equation", + "content": "\\sharp" + }, + { + "bbox": [ + 46, + 152, + 547, + 187 + ], + "type": "text", + "content": " indicates implementation in LIIF [5]. We narrow the gap between SOTA and ours in most results less than " + }, + { + "bbox": [ + 46, + 152, + 547, + 187 + ], + "type": "inline_equation", + "content": "0.15\\mathrm{dB}" + }, + { + "bbox": [ + 46, + 152, + 547, + 187 + ], + "type": "text", + "content": " (blue number). For large scale factor, we keep comparable results to MetaSR [18] and LIIF [5]. The defect in low scale factor will be analysed in Sec. 5." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 52, + 194, + 282, + 236 + ], + "lines": [ + { + "bbox": [ + 52, + 194, + 282, + 236 + ], + "spans": [ + { + "bbox": [ + 52, + 194, + 282, + 236 + ], + "type": "table", + "html": "
MethodParamsMACsFLOPsMem (training)Mem (Test)
LIIF0.35 M429 K6.2 G85.1 + 1.9 M32 + 96 M
LTE0.26 M526 K7.5 G97.8 + 1.9 M64 + 96 M
OPE (ours)0 M6 K85 M0 + 1.9 M0 + 96 M
", + "image_path": "73e639231a12cca83c681a35f039653ff4dfb231ce3dbd155c019f6a39cd30e0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 55, + 321, + 280, + 373 + ], + "blocks": [ + { + "bbox": [ + 46, + 238, + 287, + 316 + ], + "lines": [ + { + "bbox": [ + 46, + 238, + 287, + 316 + ], + "spans": [ + { + "bbox": [ + 46, + 238, + 287, + 316 + ], + "type": "text", + "content": "Table 4. Parameter number, time complexity and memory consumption. MACs: multiply-accumulate operations, FLOPs: floating point operations, Mem: intermediate data + essential output for GPU memory consumption. We use " + }, + { + "bbox": [ + 46, + 238, + 287, + 316 + ], + "type": "inline_equation", + "content": "n = 3" + }, + { + "bbox": [ + 46, + 238, + 287, + 316 + ], + "type": "text", + "content": " as maximum frequency of OPE and test in training mode and test mode on Pytorch with tool: torch.cuda.memory_allocated(). Training mode: " + }, + { + "bbox": [ + 46, + 238, + 287, + 316 + ], + "type": "inline_equation", + "content": "48^2" + }, + { + "bbox": [ + 46, + 238, + 287, + 316 + ], + "type": "text", + "content": " to 2304 pixels, test mode: " + }, + { + "bbox": [ + 46, + 238, + 287, + 316 + ], + "type": "inline_equation", + "content": "512^2" + }, + { + "bbox": [ + 46, + 238, + 287, + 316 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 238, + 287, + 316 + ], + "type": "inline_equation", + "content": "2048^2" + }, + { + "bbox": [ + 46, + 238, + 287, + 316 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 55, + 321, + 280, + 373 + ], + "lines": [ + { + "bbox": [ + 55, + 321, + 280, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 321, + 280, + 373 + ], + "type": "table", + "html": "
Method×4×8×12×16×20×24×30
LIIF382152135306004102741835027866
LTE376149033405922102681834027838
OPE2771125249537195673836612012
Percentage28%26%30%39%45%55%57%
", + "image_path": "8f9e5a50fcdd64c2c19a57192d0b779d5c492534a061883c36fb1dbff447fa1a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 67, + 453, + 266, + 513 + ], + "blocks": [ + { + "bbox": [ + 46, + 380, + 287, + 448 + ], + "lines": [ + { + "bbox": [ + 46, + 380, + 287, + 448 + ], + "spans": [ + { + "bbox": [ + 46, + 380, + 287, + 448 + ], + "type": "text", + "content": "Table 5. Rendering time of upsampling module (ms per image) with an input resolution of " + }, + { + "bbox": [ + 46, + 380, + 287, + 448 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 46, + 380, + 287, + 448 + ], + "type": "text", + "content": ". The last row shows the time saving percentage achieved by our method. We use " + }, + { + "bbox": [ + 46, + 380, + 287, + 448 + ], + "type": "inline_equation", + "content": "n = 3" + }, + { + "bbox": [ + 46, + 380, + 287, + 448 + ], + "type": "text", + "content": " as the maximum frequency of OPE. Our method provides a time advantage, which increases as the rendering resolution increases. On average, we achieve a " + }, + { + "bbox": [ + 46, + 380, + 287, + 448 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 46, + 380, + 287, + 448 + ], + "type": "text", + "content": " reduction in rendering time." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 67, + 453, + 266, + 513 + ], + "lines": [ + { + "bbox": [ + 67, + 453, + 266, + 513 + ], + "spans": [ + { + "bbox": [ + 67, + 453, + 266, + 513 + ], + "type": "table", + "html": "
In-scaleOut-scale
×2×3×4×6×8
OPE33.2930.2928.6526.4624.98
OPE-E33.2730.2328.5626.3424.82
OPE-I33.2830.2628.6326.4424.97
OPE-IE33.2030.0928.4426.2524.70
", + "image_path": "729e7c550f780c9a069b700214aacbf9349d7306396b716c192533f8d37216db.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 515, + 287, + 537 + ], + "lines": [ + { + "bbox": [ + 46, + 515, + 287, + 537 + ], + "spans": [ + { + "bbox": [ + 46, + 515, + 287, + 537 + ], + "type": "text", + "content": "Table 6. Ablation studies on Set14. EDSR-baseline [35] is used as encoder." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 560, + 287, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 560, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 46, + 560, + 287, + 657 + ], + "type": "text", + "content": "cannot be rescaled, resulting in periodic stripes in the resulting SR image. On the other hand, the OPE-I result shows no obvious discontinuity between patches since the extension plays a positive role, but this means that only a small region of the patch is presented in the target image. Lastly, the OPE-IE setting shows an obvious thick boundary between patches, indicating that both extension and interpolation are necessary for the best performance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 669, + 121, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 669, + 121, + 681 + ], + "spans": [ + { + "bbox": [ + 47, + 669, + 121, + 681 + ], + "type": "text", + "content": "5. Discussions" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": "We observed that our quantitative results decreases at low scale factors, especially when the input size is small," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 198, + 547, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 198, + 547, + 389 + ], + "spans": [ + { + "bbox": [ + 304, + 198, + 547, + 389 + ], + "type": "text", + "content": "such as with benchmark datasets. See Tab. 3 and Tab. 2. This is due to the fact that a smaller target size " + }, + { + "bbox": [ + 304, + 198, + 547, + 389 + ], + "type": "inline_equation", + "content": "(W\\times H)" + }, + { + "bbox": [ + 304, + 198, + 547, + 389 + ], + "type": "text", + "content": " leads to a larger grid in the 2D domain " + }, + { + "bbox": [ + 304, + 198, + 547, + 389 + ], + "type": "inline_equation", + "content": "\\left([-1,1]\\times [-1,1]\\right)" + }, + { + "bbox": [ + 304, + 198, + 547, + 389 + ], + "type": "text", + "content": ", where utilizing only one central point value to represent the entire larger grid would result in a loss of detailed information compared to a smaller grid. Since the 2D domain we use is continuous, the higher the resolution of the target image, the stronger the representation ability we can achieve. MLP-based representations, such as [5] and [32], can overcome this issue through nonlinear operations. In our method, this defect can be ignored for high SR scale factors where pixels are dense, but for low scale such as " + }, + { + "bbox": [ + 304, + 198, + 547, + 389 + ], + "type": "inline_equation", + "content": "\\times 2" + }, + { + "bbox": [ + 304, + 198, + 547, + 389 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 304, + 198, + 547, + 389 + ], + "type": "inline_equation", + "content": "\\times 3" + }, + { + "bbox": [ + 304, + 198, + 547, + 389 + ], + "type": "text", + "content": ", our performance may slightly degrade. A possible way to address this issue is to sample more points for every grid region and calculate their mean value, with careful consideration of the time consumption trade-off." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 401, + 378, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 401, + 378, + 414 + ], + "spans": [ + { + "bbox": [ + 306, + 401, + 378, + 414 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 421, + 545, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 421, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 304, + 421, + 545, + 612 + ], + "type": "text", + "content": "In this paper, we proposed an interpretable method for continuous image representation without implicit neural networks. Our method leverages a novel position encoding technique called orthogonal position encoding, which takes the form a 2D-Fourier series and corresponds to 2D image coordinates. As a set of orthogonal basis in inner product space, OPE is both interpretable and rich in representation. Building on OPE, we introduced the OPE-Upscale Module, a parameter-free approach for arbitrary-scale image super-resolution that simplifies the existing deep SR framework, leading to high computing efficiency and less memory consumption. Our OPE-Upscale Module can be easily integrated into existing image super-resolution pipelines, and extensive experiments demonstrate that our method achieves competitive results with the state-of-the-art." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 613, + 545, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 613, + 545, + 661 + ], + "spans": [ + { + "bbox": [ + 304, + 613, + 545, + 661 + ], + "type": "text", + "content": "It is worth noting that the overall efficiency of SR framework depends on both the encoder and decoder. Since our work focuses on decoder design and efficiency, we leave the development of high-efficient encoder as future work." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 669, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 669, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 669, + 547, + 713 + ], + "type": "text", + "content": "Acknowledgement This project was partially supported by NSFC Grants (61702363, 62222311, 62201274) and the Ministry of Education, Singapore, under its Academic Research Fund Grants (MOE-T2EP20220-0005 & RG20/20)." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10016" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Eirikur Agustsson and Radu Timofte. Ntire 2017 challenge on single image super-resolution: Dataset and study. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 126-135, 2017. 5, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 136, + 288, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 136, + 288, + 190 + ], + "spans": [ + { + "bbox": [ + 53, + 136, + 288, + 190 + ], + "type": "text", + "content": "[2] Woong Bae, Jaejun Yoo, and Jong Chul Ye. Beyond deep residual learning for image restoration: Persistent homology-guided manifold simplification. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 145-153, 2017. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 191, + 288, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 191, + 288, + 224 + ], + "spans": [ + { + "bbox": [ + 53, + 191, + 288, + 224 + ], + "type": "text", + "content": "[3] Ronen Basri and David W Jacobs. Lambertian reflectance and linear subspaces. IEEE transactions on pattern analysis and machine intelligence, 25(2):218-233, 2003. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 224, + 288, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 224, + 288, + 267 + ], + "spans": [ + { + "bbox": [ + 53, + 224, + 288, + 267 + ], + "type": "text", + "content": "[4] Marco Bevilacqua, Aline Roumy, Christine Guillemot, and Marie Line Alberi-Morel. Low-complexity single-image super-resolution based on nonnegative neighbor embedding. 2012. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 269, + 288, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 269, + 288, + 323 + ], + "spans": [ + { + "bbox": [ + 53, + 269, + 288, + 323 + ], + "type": "text", + "content": "[5] Yinbo Chen, Sifei Liu, and Xiaolong Wang. Learning continuous image representation with local implicit image function. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8628-8638, 2021. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 324, + 288, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 324, + 288, + 357 + ], + "spans": [ + { + "bbox": [ + 53, + 324, + 288, + 357 + ], + "type": "text", + "content": "[6] Lu Chi, Borui Jiang, and Yadong Mu. Fast fourier convolution. Advances in Neural Information Processing Systems, 33:4479-4488, 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 358, + 288, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 358, + 288, + 390 + ], + "spans": [ + { + "bbox": [ + 53, + 358, + 288, + 390 + ], + "type": "text", + "content": "[7] Jooyoung Choi, Jungbeom Lee, Yonghyun Jeong, and Sungroh Yoon. Toward spatially unbiased generative models. arXiv preprint arXiv:2108.01285, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 392, + 288, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 392, + 288, + 435 + ], + "spans": [ + { + "bbox": [ + 53, + 392, + 288, + 435 + ], + "type": "text", + "content": "[8] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Learning a deep convolutional network for image super-resolution. In European conference on computer vision, pages 184-199. Springer, 2014. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 436, + 288, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 436, + 288, + 479 + ], + "spans": [ + { + "bbox": [ + 53, + 436, + 288, + 479 + ], + "type": "text", + "content": "[9] Chao Dong, Chen Change Loy, and Xiaou Tang. Accelerating the super-resolution convolutional neural network. In European conference on computer vision, pages 391-407. Springer, 2016. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 480, + 288, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 480, + 288, + 513 + ], + "spans": [ + { + "bbox": [ + 48, + 480, + 288, + 513 + ], + "type": "text", + "content": "[10] Vincent Dumoulin, Jonathon Shlens, and Manjunath Kudur. A learned representation for artistic style. arXiv preprint arXiv:1610.07629, 2016. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 514, + 288, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 288, + 568 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 288, + 568 + ], + "type": "text", + "content": "[11] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5501–5510, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 569, + 288, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 569, + 288, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 569, + 288, + 613 + ], + "type": "text", + "content": "[12] Manuel Fritsche, Shuhang Gu, and Radu Timofte. Frequency separation for real-world super-resolution. In 2019 IEEE/CVF International Conference on Computer Vision Workshop (ICCVW), pages 3599-3608. IEEE, 2019. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 614, + 288, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 288, + 658 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 288, + 658 + ], + "type": "text", + "content": "[13] Dario Fuoli, Luc Van Gool, and Radu Timofte. Fourier space losses for efficient perceptual image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2360-2369, 2021. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 288, + 712 + ], + "type": "text", + "content": "[14] Kyle Genova, Forrester Cole, Avneesh Sud, Aaron Sarna, and Thomas Funkhouser. Local deep implicit functions for 3d shape. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4857-4866, 2020. 1" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 127 + ], + "type": "text", + "content": "[15] Tiantong Guo, Hojjat Seyed Mousavi, Tiep Huu Vu, and Vishal Monga. Deep wavelet prediction for image superresolution. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 104-113, 2017. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 129, + 547, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 547, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 547, + 172 + ], + "type": "text", + "content": "[16] Muhammad Haris, Gregory Shakhnarovich, and Norimichi Ukita. Deep back-projection networks for super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1664-1673, 2018. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 173, + 547, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 173, + 547, + 205 + ], + "spans": [ + { + "bbox": [ + 307, + 173, + 547, + 205 + ], + "type": "text", + "content": "[17] Kurt Hornik, Maxwell Stinchcombe, and Halbert White. Multilayer feedforward networks are universal approximators. Neural networks, 2(5):359-366, 1989. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 206, + 547, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 206, + 547, + 259 + ], + "spans": [ + { + "bbox": [ + 307, + 206, + 547, + 259 + ], + "type": "text", + "content": "[18] Xuecai Hu, Haoyuan Mu, Xiangyu Zhang, Zilei Wang, Tieniu Tan, and Jian Sun. Meta-sr: A magnification-arbitrary network for super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1575-1584, 2019. 1, 3, 7, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 261, + 547, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 261, + 547, + 303 + ], + "spans": [ + { + "bbox": [ + 307, + 261, + 547, + 303 + ], + "type": "text", + "content": "[19] Binbin Huang, Xinhao Yan, Anpei Chen, Shenghua Gao, and Jingyi Yu. Pref: Phasorial embedding fields for compact neural representations. arXiv preprint arXiv:2205.13524, 2022. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 304, + 547, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 304, + 547, + 349 + ], + "spans": [ + { + "bbox": [ + 307, + 304, + 547, + 349 + ], + "type": "text", + "content": "[20] Huaibo Huang, Ran He, Zhenan Sun, and Tieniu Tan. Wavelet-srnet: A wavelet-based cnn for multi-scale face super resolution. In Proceedings of the IEEE International Conference on Computer Vision, pages 1689-1697, 2017. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 350, + 547, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 350, + 547, + 392 + ], + "spans": [ + { + "bbox": [ + 307, + 350, + 547, + 392 + ], + "type": "text", + "content": "[21] Jia-Bin Huang, Abhishek Singh, and Narendra Ahuja. Single image super-resolution from transformed self-exemplars. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5197-5206, 2015. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 393, + 547, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 393, + 547, + 437 + ], + "spans": [ + { + "bbox": [ + 307, + 393, + 547, + 437 + ], + "type": "text", + "content": "[22] Liming Jiang, Bo Dai, Wayne Wu, and Chen Change Loy. Focal frequency loss for image reconstruction and synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13919-13929, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 437, + 547, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 437, + 547, + 480 + ], + "spans": [ + { + "bbox": [ + 307, + 437, + 547, + 480 + ], + "type": "text", + "content": "[23] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. Advances in Neural Information Processing Systems, 34:852-863, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 481, + 547, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 481, + 547, + 525 + ], + "spans": [ + { + "bbox": [ + 307, + 481, + 547, + 525 + ], + "type": "text", + "content": "[24] Jiwon Kim, Jung Kwon Lee, and Kyoung Mu Lee. Accurate image super-resolution using very deep convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1646-1654, 2016. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 525, + 547, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 525, + 547, + 569 + ], + "spans": [ + { + "bbox": [ + 307, + 525, + 547, + 569 + ], + "type": "text", + "content": "[25] Jiwon Kim, Jung Kwon Lee, and Kyoung Mu Lee. Deeplyrecursive convolutional network for image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1637-1645, 2016. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 570, + 547, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 570, + 547, + 602 + ], + "spans": [ + { + "bbox": [ + 307, + 570, + 547, + 602 + ], + "type": "text", + "content": "[26] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 603, + 547, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 603, + 547, + 635 + ], + "spans": [ + { + "bbox": [ + 307, + 603, + 547, + 635 + ], + "type": "text", + "content": "[27] Tom Koornwinder. Two-variable analogues of the classical orthogonal polynomials. In Theory and application of special functions, pages 435-495. Elsevier, 1975. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 636, + 547, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 636, + 547, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 636, + 547, + 689 + ], + "type": "text", + "content": "[28] Wei-Sheng Lai, Jia-Bin Huang, Narendra Ahuja, and Ming-Hsuan Yang. Deep laplacian pyramid networks for fast and accurate super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 624-632, 2017. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 691, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 547, + 713 + ], + "type": "text", + "content": "[29] Wei-Sheng Lai, Jia-Bin Huang, Narendra Ahuja, and Ming-Hsuan Yang. Fast and accurate image super-resolution with" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10017" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 105 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 105 + ], + "type": "text", + "content": "deep laplacian pyramid networks. IEEE transactions on pattern analysis and machine intelligence, 41(11):2599-2613, 2018. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 140 + ], + "type": "text", + "content": "[30] Seyed Mehdi Lajevardi and Zahir M Hussain. Higher order orthogonal moments for invariant facial expression recognition. Digital Signal Processing, 20(6):1771-1779, 2010. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 287, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 287, + 217 + ], + "type": "text", + "content": "[31] Christian Ledig, Lucas Theis, Ferenc Huszár, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photorealistic single image super-resolution using a generative adversarial network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4681-4690, 2017. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 220, + 288, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 220, + 288, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 220, + 288, + 262 + ], + "type": "text", + "content": "[32] Jaewon Lee and Kyong Hwan Jin. Local texture estimator for implicit representation function. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1929-1938, 2022. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 264, + 287, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 264, + 287, + 318 + ], + "spans": [ + { + "bbox": [ + 48, + 264, + 287, + 318 + ], + "type": "text", + "content": "[33] Zhen Li, Jinglei Yang, Zheng Liu, Xiaomin Yang, Gwang-gil Jeon, and Wei Wu. Feedback network for image super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3867-3876, 2019. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 320, + 287, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 320, + 287, + 374 + ], + "spans": [ + { + "bbox": [ + 48, + 320, + 287, + 374 + ], + "type": "text", + "content": "[34] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1833-1844, 2021. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 376, + 287, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 376, + 287, + 431 + ], + "spans": [ + { + "bbox": [ + 48, + 376, + 287, + 431 + ], + "type": "text", + "content": "[35] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 136-144, 2017. 2, 6, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 434, + 287, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 434, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 48, + 434, + 287, + 475 + ], + "type": "text", + "content": "[36] Chieh Hubert Lin, Hsin-Ying Lee, Yen-Chi Cheng, Sergey Tulyakov, and Ming-Hsuan Yang. Infinitygan: Towards infinite-pixel image synthesis. arXiv preprint arXiv:2104.03963, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 478, + 287, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 478, + 287, + 520 + ], + "spans": [ + { + "bbox": [ + 48, + 478, + 287, + 520 + ], + "type": "text", + "content": "[37] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. Advances in Neural Information Processing Systems, 33:15651-15663, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 523, + 287, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 523, + 287, + 576 + ], + "spans": [ + { + "bbox": [ + 48, + 523, + 287, + 576 + ], + "type": "text", + "content": "[38] Pengju Liu, Hongzhi Zhang, Kai Zhang, Liang Lin, and Wangmeng Zuo. Multi-level wavelet-cnn for image restoration. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 773-782, 2018. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 579, + 287, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 579, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 579, + 287, + 612 + ], + "type": "text", + "content": "[39] Julien Mairal, Michael Elad, and Guillermo Sapiro. Sparse representation for color image restoration. IEEE Transactions on image processing, 17(1):53-69, 2007. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 613, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 678 + ], + "type": "text", + "content": "[40] David Martin, Charless Fowlkes, Doron Tal, and Jitendra Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In Proceedings Eighth IEEE International Conference on Computer Vision. ICCV 2001, volume 2, pages 416-423. IEEE, 2001. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[41] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "of the IEEE/CVF conference on computer vision and pattern recognition, pages 4460-4470, 2019. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 95, + 545, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 95, + 545, + 148 + ], + "spans": [ + { + "bbox": [ + 308, + 95, + 545, + 148 + ], + "type": "text", + "content": "[42] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 150, + 545, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 150, + 545, + 204 + ], + "spans": [ + { + "bbox": [ + 308, + 150, + 545, + 204 + ], + "type": "text", + "content": "[43] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11453-11464, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 206, + 545, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 206, + 545, + 260 + ], + "spans": [ + { + "bbox": [ + 308, + 206, + 545, + 260 + ], + "type": "text", + "content": "[44] Michael Niemeyer, Lars Mescheder, Michael Oechsle, and Andreas Geiger. Differentiable volumetric rendering: Learning implicit 3d representations without 3d supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3504-3515, 2020. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 261, + 545, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 261, + 545, + 315 + ], + "spans": [ + { + "bbox": [ + 308, + 261, + 545, + 315 + ], + "type": "text", + "content": "[45] Michael Oechsle, Lars Mescheder, Michael Niemeyer, Thilo Strauss, and Andreas Geiger. Texture fields: Learning texture representations in function space. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4531-4540, 2019. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 316, + 545, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 316, + 545, + 346 + ], + "spans": [ + { + "bbox": [ + 308, + 316, + 545, + 346 + ], + "type": "text", + "content": "[46] Alan V. Oppenheim and Alan S. Willsky. Signals and Systems. Prentice Hall, Upper Saddle River, NJ, 2 edition, 1997. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 349, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 349, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 308, + 349, + 545, + 403 + ], + "type": "text", + "content": "[47] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. Deepsdf: Learning continuous signed distance functions for shape representation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 165-174, 2019. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 404, + 545, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 404, + 545, + 468 + ], + "spans": [ + { + "bbox": [ + 308, + 404, + 545, + 468 + ], + "type": "text", + "content": "[48] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 471, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 471, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 308, + 471, + 545, + 514 + ], + "type": "text", + "content": "[49] Songyou Peng, Michael Niemeyer, Lars Mescheder, Marc Pollefeys, and Andreas Geiger. Convolutional occupancy networks. In European Conference on Computer Vision, pages 523-540. Springer, 2020. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 515, + 545, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 515, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 308, + 515, + 545, + 568 + ], + "type": "text", + "content": "[50] Nasim Rahaman, Aristide Baratin, Devansh Arpit, Felix Draxler, Min Lin, Fred Hamprecht, Yoshua Bengio, and Aaron Courville. On the spectral bias of neural networks. In International Conference on Machine Learning, pages 5301-5310. PMLR, 2019. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 570, + 545, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 570, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 308, + 570, + 545, + 612 + ], + "type": "text", + "content": "[51] Ravi Ramamoorthi and Pat Hanrahan. On the relationship between radiance and irradiance: determining the illumination from images of a convex lambertian object. JOSA A, 18(10):2448-2459, 2001. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 614, + 545, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 614, + 545, + 646 + ], + "spans": [ + { + "bbox": [ + 308, + 614, + 545, + 646 + ], + "type": "text", + "content": "[52] Oren Rippel, Jasper Snoek, and Ryan P Adams. Spectral representations for convolutional neural networks. Advances in neural information processing systems, 28, 2015. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "type": "text", + "content": "[53] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. Graf: Generative radiance fields for 3d-aware image synthesis. Advances in Neural Information Processing Systems, 33:20154-20166, 2020. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "text", + "content": "[54] Wenzhe Shi, Jose Caballero, Ferenc Huszár, Johannes Totz, Andrew P Aitken, Rob Bishop, Daniel Rueckert, and Zehan" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10018" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 116 + ], + "type": "text", + "content": "Wang. Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1874-1883, 2016. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 161 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 161 + ], + "type": "text", + "content": "[55] Assaf Shocher, Nadav Cohen, and Michal Irani. “zero-shot” super-resolution using deep internal learning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3118–3126, 2018. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 162, + 288, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 162, + 288, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 162, + 288, + 205 + ], + "type": "text", + "content": "[56] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. Advances in Neural Information Processing Systems, 33:7462-7473, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 206, + 287, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 206, + 287, + 259 + ], + "spans": [ + { + "bbox": [ + 48, + 206, + 287, + 259 + ], + "type": "text", + "content": "[57] Vincent Sitzmann, Justus Thies, Felix Heide, Matthias Nießner, Gordon Wetzstein, and Michael Zollhofer. Deepvoxels: Learning persistent 3d feature embeddings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2437-2446, 2019. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 261, + 287, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 261, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 48, + 261, + 287, + 304 + ], + "type": "text", + "content": "[58] Vincent Sitzmann, Michael Zollhöfer, and Gordon Wetzstein. Scene representation networks: Continuous 3d-structure-aware neural scene representations. Advances in Neural Information Processing Systems, 32, 2019. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 305, + 287, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 305, + 287, + 360 + ], + "spans": [ + { + "bbox": [ + 48, + 305, + 287, + 360 + ], + "type": "text", + "content": "[59] Peter-Pike Sloan, Jan Kautz, and John Snyder. Precomputed radiance transfer for real-time rendering in dynamic, low-frequency lighting environments. In Proceedings of the 29th annual conference on Computer graphics and interactive techniques, pages 527-536, 2002. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 361, + 287, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 361, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 48, + 361, + 287, + 403 + ], + "type": "text", + "content": "[60] Sanghyun Son and Kyoung Mu Lee. Srwarp: Generalized image super-resolution under arbitrary transformation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7782-7791, 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 404, + 287, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 404, + 287, + 448 + ], + "spans": [ + { + "bbox": [ + 48, + 404, + 287, + 448 + ], + "type": "text", + "content": "[61] Ying Tai, Jian Yang, and Xiaoming Liu. Image superresolution via deep recursive residual network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3147-3155, 2017. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 449, + 287, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 449, + 287, + 492 + ], + "spans": [ + { + "bbox": [ + 48, + 449, + 287, + 492 + ], + "type": "text", + "content": "[62] Ying Tai, Jian Yang, Xiaoming Liu, and Chunyan Xu. Memnet: A persistent memory network for image restoration. In Proceedings of the IEEE international conference on computer vision, pages 4539-4547, 2017. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 493, + 287, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 493, + 287, + 547 + ], + "spans": [ + { + "bbox": [ + 48, + 493, + 287, + 547 + ], + "type": "text", + "content": "[63] Radu Timofte, Eirikur Agustsson, Luc Van Gool, Ming-Hsuan Yang, and Lei Zhang. Ntire 2017 challenge on single image super-resolution: Methods and results. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 114-125, 2017. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 548, + 287, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 287, + 591 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 287, + 591 + ], + "type": "text", + "content": "[64] Tong Tong, Gen Li, Xiejie Liu, and Qinquan Gao. Image super-resolution using dense skip connections. In Proceedings of the IEEE international conference on computer vision, pages 4799-4807, 2017. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 592, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 287, + 635 + ], + "type": "text", + "content": "[65] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 636, + 288, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 288, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 288, + 690 + ], + "type": "text", + "content": "[66] Longguang Wang, Yingqian Wang, Zaiping Lin, Jungang Yang, Wei An, and Yulan Guo. Learning a single network for scale-arbitrary super-resolution. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4801-4810, 2021. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[67] Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real-esrgan: Training real-world blind super-resolution with" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 105 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 105 + ], + "type": "text", + "content": "pure synthetic data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1905-1914, 2021. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 106, + 545, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 106, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 307, + 106, + 545, + 171 + ], + "type": "text", + "content": "[68] Yifan Wang, Federico Perazzi, Brian McWilliams, Alexander Sorkine-Hornung, Olga Sorkine-Hornung, and Christopher Schroers. A fully progressive approach to single-image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 864-873, 2018. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 172, + 545, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 172, + 545, + 237 + ], + "spans": [ + { + "bbox": [ + 307, + 172, + 545, + 237 + ], + "type": "text", + "content": "[69] Yifan Wang, Federico Perazzi, Brian McWilliams, Alexander Sorkine-Hornung, Olga Sorkine-Hornung, and Christopher Schroers. A fully progressive approach to single-image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 864-873, 2018. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 238, + 545, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 238, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 307, + 238, + 545, + 281 + ], + "type": "text", + "content": "[70] Zhihao Wang, Jian Chen, and Steven CH Hoi. Deep learning for image super-resolution: A survey. IEEE transactions on pattern analysis and machine intelligence, 43(10):3365-3387, 2020. 1, 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 282, + 545, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 282, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 307, + 282, + 545, + 327 + ], + "type": "text", + "content": "[71] John Wright, Allen Y Yang, Arvind Ganesh, S Shankar Sastry, and Yi Ma. Robust face recognition via sparse representation. IEEE transactions on pattern analysis and machine intelligence, 31(2):210-227, 2008. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 327, + 545, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 327, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 307, + 327, + 545, + 371 + ], + "type": "text", + "content": "[72] Xingqian Xu, Zhangyang Wang, and Humphrey Shi. Ultrasr: Spatial encoding is a missing key for implicit image function-based arbitrary-scale super-resolution. arXiv preprint arXiv:2103.12716, 2021. 1, 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 371, + 545, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 371, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 307, + 371, + 545, + 415 + ], + "type": "text", + "content": "[73] Jianchao Yang, John Wright, Thomas Huang, and Yi Ma. Image super-resolution as sparse representation of raw image patches. In 2008 IEEE conference on computer vision and pattern recognition, pages 1-8. IEEE, 2008. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 415, + 545, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 415, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 307, + 415, + 545, + 459 + ], + "type": "text", + "content": "[74] Yanchao Yang and Stefano Soatto. Fda: Fourier domain adaptation for semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4085-4095, 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 459, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 459, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 307, + 459, + 545, + 514 + ], + "type": "text", + "content": "[75] Youngho Yoon, Inchul Chung, Lin Wang, and Kuk-Jin Yoon. Spheresr: 360deg image super-resolution with arbitrary projection via continuous spherical image representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5677-5686, 2022. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 514, + 545, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 514, + 545, + 558 + ], + "spans": [ + { + "bbox": [ + 307, + 514, + 545, + 558 + ], + "type": "text", + "content": "[76] Roman Zeyde, Michael Elad, and Matan Protter. On single image scale-up using sparse-representations. In International conference on curves and surfaces, pages 711-730. Springer, 2010. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 559, + 545, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 559, + 545, + 603 + ], + "spans": [ + { + "bbox": [ + 307, + 559, + 545, + 603 + ], + "type": "text", + "content": "[77] Hui Zhang, Huazhong Shu, Guoniu N Han, Gouenou Coa-trieux, Limin Luo, and Jean Louis Coa-trieux. Blurred image recognition by legendre moment invariants. IEEE Transactions on Image Processing, 19(3):596-611, 2009. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 603, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 603, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 307, + 603, + 545, + 635 + ], + "type": "text", + "content": "[78] Kai Zhang, Gernot Riegler, Noah Snavely, and Vladlen Koltun. Nerf++: Analyzing and improving neural radiance fields. arXiv preprint arXiv:2010.07492, 2020. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 636, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 636, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 636, + 545, + 689 + ], + "type": "text", + "content": "[79] Yulun Zhang, Kunpeng Li, Kai Li, Lichen Wang, Bineng Zhong, and Yun Fu. Image super-resolution using very deep residual channel attention networks. In Proceedings of the European conference on computer vision (ECCV), pages 286-301, 2018. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 690, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 690, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 690, + 545, + 713 + ], + "type": "text", + "content": "[80] Yulun Zhang, Yapeng Tian, Yu Kong, Bineng Zhong, and Yun Fu. Residual dense network for image super-resolution." + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10019" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 174 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "text", + "content": "In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2472-2481, 2018. 6, 7, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 139 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 139 + ], + "type": "text", + "content": "[81] Man Zhou, Hu Yu, Jie Huang, Feng Zhao, Jinwei Gu, Chen Change Loy, Deyu Meng, and Chongyi Li. Deep fourier up-sampling. arXiv preprint arXiv:2210.05171, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 286, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 286, + 174 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 286, + 174 + ], + "type": "text", + "content": "[82] Hongqing Zhu. Image representation using separable two-dimensional continuous and discrete orthogonal moments. Pattern Recognition, 45(4):1540-1558, 2012. 2" + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "10020" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/ORCa_ Glossy Objects As Radiance-Field Cameras/da15cd79-f9cc-4796-854b-9e2af61a06e8_content_list.json b/2023/ORCa_ Glossy Objects As Radiance-Field Cameras/da15cd79-f9cc-4796-854b-9e2af61a06e8_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..162967d57af31a106fcef0048a0a7f0cad8a2dbc --- /dev/null +++ b/2023/ORCa_ Glossy Objects As Radiance-Field Cameras/da15cd79-f9cc-4796-854b-9e2af61a06e8_content_list.json @@ -0,0 +1,1866 @@ +[ + { + "type": "text", + "text": "ORCa: Glossy Objects as Radiance-Field Cameras", + "text_level": 1, + "bbox": [ + 227, + 130, + 743, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kushagra Tiwary $^{1*}$ , Akshit Dave $^{2*}$ , Nikhil Behari $^{1}$ , Tzofi Klinghoffer $^{1}$ , Ashok Veeraraghavan $^{2}$ , Ramesh Raskar $^{1}$", + "bbox": [ + 202, + 179, + 764, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Massachusetts Institute of Technology, 2Rice University", + "bbox": [ + 259, + 215, + 709, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{ktiwary,behari,tzofi,raskar}@mit.edu，{akshat,vashok}@rice.edu", + "bbox": [ + 205, + 236, + 759, + 252 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2df5ce9a07841845a9d542726a640bb85b9e6d9291e58e9478efb66eb984ffe0.jpg", + "image_caption": [ + "(a) Scene: Glossy Elephant in a room" + ], + "image_footnote": [], + "bbox": [ + 117, + 311, + 316, + 498 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/dc19608bbd6e867f5468407562bad93e13addd859ab7ff6236d1e71d79f9430b.jpg", + "image_caption": [ + "(b) Input: images with distorted reflections" + ], + "image_footnote": [], + "bbox": [ + 318, + 311, + 449, + 406 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/31e4fd0b5ff2c092526cf8c945975aff577b819c566fc32bc84692aab0160134.jpg", + "image_caption": [ + "Figure 1. Objects as radiance-field cameras. We convert everyday objects with unknown geometry (a) into radiance-field cameras by modeling multi-view reflections (b) as projections of the 5D radiance field of the environment. We convert the object surface into a virtual sensor to capture this radiance field (c), which enables depth and radiance estimation of the surrounding environment. We can then query this radiance field to perform beyond field-of-view novel view synthesis of the environment (d)." + ], + "image_footnote": [], + "bbox": [ + 318, + 406, + 447, + 498 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/319140cdbea5bb5445735e9ea44a9cd49aae383dd620fbddcb858941bfcc8e1d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 455, + 393, + 514, + 415 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4e21bd1d0ee90f2978ca334273eca7b705f28da5a05562d24c5f88cc947fccc9.jpg", + "image_caption": [ + "(c) Using objects to image beyond camera's field-of-view" + ], + "image_footnote": [], + "bbox": [ + 519, + 311, + 718, + 498 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a2b083707994440a1e0da6eb06ee8d40808b8bdec0bcb43e6e6c39ed27b00c62.jpg", + "image_caption": [ + "(d) Recovered and Undistorted RGBs" + ], + "image_footnote": [], + "bbox": [ + 720, + 311, + 849, + 405 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7fae400df005cfb7ceba6e51ff03cb387e2efef519e097a20a61fd25e9dea927.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 720, + 406, + 849, + 498 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 577, + 313, + 593 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Reflections on glossy objects contain valuable and hidden information about the surrounding environment. By converting these objects into cameras, we can unlock exciting applications, including imaging beyond the camera's field-of-view and from seemingly impossible vantage points, e.g. from reflections on the human eye. However, this task is challenging because reflections depend jointly on object geometry, material properties, the 3D environment, and the observer's viewing direction. Our approach converts glossy objects with unknown geometry into radiance-field cameras to image the world from the object's perspective. Our key insight is to convert the object surface into a virtual sensor that captures cast reflections as a 2D projection of the 5D environment radiance field visible to and surrounding the object. We show that recovering the environment radiance fields enables depth and radiance estimation from the object to its surroundings in addition to beyond field-of-view", + "bbox": [ + 73, + 611, + 473, + 869 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "novel-view synthesis, i.e. rendering of novel views that are only directly visible to the glossy object present in the scene, but not the observer. Moreover, using the radiance field we can image around occluders caused by close-by objects in the scene. Our method is trained end-to-end on multi-view images of the object and jointly estimates object geometry, diffuse radiance, and the 5D environment radiance field. For more information, visit our website.", + "bbox": [ + 500, + 579, + 890, + 700 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 709, + 632, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Imagine that you're driving down a city street that is packed with lines of parked cars on both sides. Inspection of the cars' glass windshields, glossy paint, and plastic reveal sharp, but faint and distorted views of the surroundings that might be otherwise hidden from you. Humans can infer depth and semantic cues about the occluded areas in the environment by processing reflections visible on reflective objects, internally decomposing the object geometry and radiance from the specular radiance being reflected onto it. Our aim is to decompose the object from its reflections to \"see\" the world from the object's perspective, effectively turning", + "bbox": [ + 496, + 734, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution", + "bbox": [ + 96, + 886, + 202, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "20773", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/336fad8acd27ed663001f77ba70be48d9fce30a1b40b4f2c16201dee72c312f1.jpg", + "image_caption": [ + "(a) Learning Neural Implicit Surfaces [Sec. 3.2]" + ], + "image_footnote": [ + "Real camera origin" + ], + "bbox": [ + 89, + 119, + 315, + 276 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/31c4ee92e85bdedbf20674e064bab88310bde6e2ece202491963e1b2013d9540.jpg", + "image_caption": [ + "Figure 2. ORCa Overview. We jointly estimate the object's geometry and diffuse along with the environment radiance field estimation through a three-step approach. First, we model the object as a neural implicit surface (a). We model the reflections as probing the environment on virtual viewpoints (b) estimated analytically from surface properties. We model the environment as a radiance field queried on these viewpoints (c). Both neural implicit surface and environment radiance field are trained jointly on multi-view images of the object using a photometric loss." + ], + "image_footnote": [ + "Real pixel cone", + "- Surface intersection", + "Virtual pixel cone", + "Virtual camera origin" + ], + "bbox": [ + 364, + 169, + 591, + 277 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/351c007cb16eae15b70ea2b301d7329906fe4afd514b299eee6a2d2089679df2.jpg", + "image_caption": [ + "(b) Objects Surface as Virtual Sensor [Sec. 3.3]" + ], + "image_footnote": [], + "bbox": [ + 620, + 125, + 709, + 195 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f957743e124559cdadc315e2b75882f4c4ed9a059097d2b671298f01d58da310.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 620, + 196, + 704, + 271 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/79e2e5a652c3fc556155146662f212966fa57dd0870b932f09a0ce16f5fea170.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 717, + 178, + 790, + 258 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/779145651c7b8316abd2c83fd2648e6f16c1d6536a4200c71278aab7c12e34d3.jpg", + "image_caption": [ + "(c) Learning Environment Radiance Fields [Sec. 3.4]" + ], + "image_footnote": [], + "bbox": [ + 802, + 132, + 875, + 190 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/05cf6acc5514021a0e70d0ad07726823cb7affe5b4c7e89d84572cc68281700e.jpg", + "image_caption": [ + "Virtual View", + "Virtual Depth" + ], + "image_footnote": [], + "bbox": [ + 802, + 203, + 888, + 258 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the object into a camera that images its environment. However, reflections pose a long-standing challenge in computer vision as the reflections are a 2D projection of an unknown 3D environment that is distorted based on the shape of the reflector.", + "bbox": [ + 75, + 391, + 470, + 465 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To capture the 3D world from the object's perspective, we model the object's surface as a virtual sensor that captures the 2D projection of a 5D environment radiance field surrounding the object. This environment radiance field consists largely of areas only visible to the observer through the object's reflections. Our use of environment radiance fields not only enables depth and radiance estimation from the object to its surroundings but also enables beyond field-of-view novel-view synthesis, i.e. rendering of novel views that are only directly visible to the glossy object present in the scene but not the observer. Unlike conventional approaches that model the environment as a 2D map, our approach models it as a 5D field without assuming the scene is infinitely far away. Moreover, by sampling the 5D radiance field, instead of a 2D map, we can capture depth and images around occluders, such as close-by objects in the scene, as shown in Fig. 3. These applications cannot be done from a 2D environment map.", + "bbox": [ + 75, + 472, + 468, + 743 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We aim to decompose reflections on the object's surface, from its surface and exploit those reflections to construct a radiance field surrounding the object, therefore capturing the 3D world in the process. This is a challenging task because the reflections are extremely sensitive to local object geometry, viewing direction and inter-reflections due to the object's surface. To capture this radiance field, we convert glossy objects with unknown geometry and texture into radiance-field cameras. Specifically, we exploit neural rendering to estimate the local surface of the object viewed", + "bbox": [ + 75, + 750, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/aeb4110451eb9505f87a500f1fe33aaa68a383a804374a2cba6aee5e98fd4d1e.jpg", + "image_caption": [ + "(a) Sample Input Views" + ], + "image_footnote": [], + "bbox": [ + 514, + 402, + 612, + 516 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/fbafe69d37013587b525762efddc55cca4a5bba27ead0abc2b7e3edc12c30e06.jpg", + "image_caption": [ + "(b) Virtual Views with Translation", + "(c) Virtual View Depth" + ], + "image_footnote": [], + "bbox": [ + 616, + 401, + 784, + 517 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/7f89207a5e51e11d24714ceb23ad8cae15edef672cca9e9b3768842a48d12ba6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 785, + 401, + 874, + 516 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/22bcbfb95f7bdc1a496880cc6aee049956973e1d3b15df3cab447d26c0ae733e.jpg", + "image_caption": [ + "(d) Parallax effect in 5D Env. Radiance Field" + ], + "image_footnote": [], + "bbox": [ + 513, + 531, + 635, + 607 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b20d6b50a427e566c9f341272c5fec2348027903e9276308a271ec6b2683132c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 643, + 531, + 753, + 606 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/7f6a08274bc3335ececbd90b34197b8e3717f76675954522b74248b0ef7f7f0f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 763, + 531, + 872, + 606 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/5ed18484080d02f54b9c078dca844f100f84775f461a893f9e4ababb07ee2d4a.jpg", + "image_caption": [ + "x 0", + "Figure 3. Advantages of 5D environment radiance field. Modeling reflections on object surfaces (a) as a 5D env. radiance field enables beyond field-of-view novel-view synthesis, including rendering of the environment from translated virtual camera views (b). Depth (c) and environment radiance of translated and parallax views can further enable imaging behind occluders, for example revealing the tails behind the primary Pokemon occluders (d)." + ], + "image_footnote": [], + "bbox": [ + 513, + 608, + 635, + 683 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/bb1091cb84235be474a23582eacd3ddda3c203b1817093900e6a0022c52c262d.jpg", + "image_caption": [ + "+0.25" + ], + "image_footnote": [], + "bbox": [ + 643, + 608, + 753, + 681 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/67c7c9425f8b662d3be565cdae4237350d080244862d3ecf337dc718ab2cd1e6.jpg", + "image_caption": [ + "+0.50" + ], + "image_footnote": [], + "bbox": [ + 763, + 608, + 872, + 681 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "from each pixel of the real camera. We then convert this local surface into a virtual pixel that captures radiance from the environment. This virtual pixel captures the environment radiance as shown in Fig 5. We estimate the outgo", + "bbox": [ + 496, + 839, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "20774", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ing frustum from the virtual pixel as a cone that samples the scene. By sampling the scene from many virtual pixels on the object surface, we construct an environment radiance field that can be queried independently of the object surface, enabling beyond field-of-view novel-view synthesis from previously unsampled viewpoints.", + "bbox": [ + 75, + 90, + 470, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our approach jointly estimates object geometry, diffuse radiance, and the environment radiance field from multiview images of glossy objects with unknown geometry and diffuse texture in three steps. First, we use neural signed distance functions (SDF) and an MLP to model the glossy object's geometry as a neural implicit surface and diffuse radiance, respectively, similar to PANDORA [10]. Then, for every pixel on the observer's camera, we estimate the virtual pixels on the object's surface based on the estimated local geometry from the neural SDF. We analytically compute the parameters of the virtual cone through the virtual pixel. Lastly, we use the cone formulation in MipNeRF [5] to cast virtual cones from the virtual camera to recover the environment radiance.", + "bbox": [ + 75, + 181, + 470, + 391 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To summarize, we make the following contributions:", + "bbox": [ + 96, + 393, + 444, + 407 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We present a method to convert implicit surfaces into virtual sensors that can image their surroundings using virtual cones. (Sec. 3.3)", + "- We jointly estimate object geometry, diffuse radiance, and estimate the 5D environment radiance field surrounding the object. (Fig. 7 & 9)", + "- We show that the environment radiance field can be queried to perform beyond-field-of-view novel viewpoint synthesis, i.e. render views only visible to the object in the scene (Section 3.4)" + ], + "bbox": [ + 96, + 412, + 467, + 579 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Scope. We only model glossy objects with low roughness as such specular reflections tend to have a high signal-to-noise ratio, therefore, are a sharper estimate of the environment radiance field. However, we note that the virtual cone computation can be extended to model the cone radius as a function of surface roughness. Deblurring approaches can further improve the resolution of estimated environment. In addition, we approximate the local curvature using mean curvature, which fails for objects with varying radius of curvature along the tangent space. We explain how our virtual cone curvature estimation can be extended to handle general shape operators in the supplementary material. Lastly, similar to other multi-view approaches, our approach relies on a sufficient virtual baseline between virtual viewpoints to recover the environment radiance field.", + "bbox": [ + 75, + 584, + 468, + 811 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 821, + 218, + 835 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1. Modeling reflections", + "text_level": 1, + "bbox": [ + 76, + 847, + 269, + 863 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Catadioptric imaging systems incorporate reflections on curved reflective mirrors to expand the field of view of con", + "bbox": [ + 75, + 869, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ventional cameras [2, 26], to increase the baseline of light field cameras [11, 40] and to perform novel view synthesis from a single capture [44]. These works assume the geometry of the reflecting surface is known or calibrated while we create a catadioptric imaging system from everyday glossy objects of unknown geometry. Grossberg et al. [15] propose a generalized model for light transport through imaging systems, including catadioptric systems. Our work focuses on light transport reflecting off a general object in the scene.", + "bbox": [ + 496, + 90, + 890, + 226 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Light field imaging is shown to be effective for reflection removal [21, 27], reconstructing specular surfaces [18], intrinsic decomposition [1], and neural rendering [35]. These works typically consider planar reflections or require training on synthetic datasets, while our approach models reflections on complex geometry and is unsupervised. Prior works have also utilized additional light properties such as polarization [9, 10, 20] and time-of-flight [17, 33] for the separation of the reflected component and specular surface reconstruction. While the input of our approach is RGB images, there is scope for improving reconstruction quality by supplementing the algorithm with these additional cues.", + "bbox": [ + 496, + 228, + 890, + 409 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recent progress in neural radiance fields has enabled impressive novel view rendering and geometry reconstruction from multi-view images [24]. MipNeRF [5] demonstrates better novel view synthesis by modeling outgoing rays as cones to enable anti-aliasing. RefNeRF [41] proposes Integrated Directional Embeddings for improved novel view synthesis of reflections. NeRFRen [16] separates diffuse and specular radiance by using separate neural networks. Neural Catacaustics [19] propose a neural warping method to model reflections by learning the caustics of the surface. While these works focus on novel-view synthesis of the scene from the primary camera, we perform view synthesis that is beyond the line-of-sight of the primary camera, i.e., rendering views only visible to the objects present in the scene, while jointly estimating object geometry and separating diffuse and specular radiance.", + "bbox": [ + 496, + 410, + 890, + 652 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Environment Estimation", + "text_level": 1, + "bbox": [ + 500, + 665, + 723, + 679 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recovering underlying scene properties from multiple images is inherently ill-posed [32], but can be regularized using the natural statistics of scene properties as a prior [4, 34]. Recent works exploit this prior through deep neural networks and demonstrate inverse rendering of indoor scenes from a single image [12, 22, 43, 48]. However, these techniques typically recover only coarse representations of lighting and cannot reconstruct fine details of the environment. Lombardi et al. [23] recover environment and reflectance, assuming the scene is composed of known geometry and uniform material. Georgolis et al. [13] recover the environment map behind the camera from a single image of a glossy object, assuming the object is composed of textureless materials and using ground truth seg", + "bbox": [ + 496, + 689, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "20775", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/47a18be627fa2f9c56ccaf7a56370b076e9d90c9a3ed8e66f6c506b9a0667948.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ApproachInputScene GeometryAlgorithm TypeEnvironment Dimension
Neural IllumSingle RGBNormalsSupervised2D
LighthouseStereo RGB pairMulti-plane ImageSupervised5D
NeRFactorMulti-view RGBVolumetric AlbedoSemi-supervised2D
RefNeRFMulti-view RGBVolumetric AlbedoSelf-supervised2D
PANDORAMulti-view RGB+Pol.Neural SDFSelf-supervised2D
ORCa (Ours)Multi-view RGBNeural SDFSelf-supervised5D
", + "bbox": [ + 78, + 88, + 470, + 237 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "mentation masks. Song et al. [36] estimate plausible environment maps by mapping reflections in the image and inpainting unmapped regions. Srinivasan et al. [37] capture stereo image pairs and estimate plausible spatially-coherent environment maps. NeRD [7], NeRFactor [47] and NeuralPIL [8] employ data-driven priors for lighting and BRDF in a NeRF-based approach for radiance decomposition. Park et al. [31] use RGB-D videos to estimate environment map. Swedish et al. [39] recover high-frequency illumination map from the shadows of an object with known geometry. PhySG [46] and Munkberg et al. [25] perform inverse rendering from multi-view images by modeling the surface as signed distance functions. PANDORA [10] performs radiance decomposition from polarized RGB images.", + "bbox": [ + 75, + 398, + 472, + 612 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. Learning environment radiance fields from multi-view reflections", + "text_level": 1, + "bbox": [ + 76, + 638, + 468, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 76, + 681, + 187, + 695 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Reflections on glossy objects offer a glimpse into the surrounding environment beyond the camera's field of view. From multi-view images of a glossy object with unknown geometry and albedo, we aim to recover the 5D radiance field of the surrounding environment. The mapping from images captured by the observer to the surrounding environment depends on the glossy object's surface properties, in particular, the surface normals and curvature. We first cast a cone from the observer camera's center-of-projection through each pixel viewing the scene. When the cone intersects the object's surface, it reflects, causing the cone to be transformed (Fig. 5). The transformed cone, referred to as a virtual cone, samples the environment and is primar-", + "bbox": [ + 75, + 704, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/779df1b504cb234de996732bbc934595f060771d044a45fbd83b5e114802095c.jpg", + "image_caption": [ + "Figure 4. Comparison of environment estimation approaches. Approaches such Neural Illum [36], Lighthouse [37], NeRFactor [47] train on datasets of natural illumination maps to regularize ill-posedness of environment estimation. PANDORA [10] and RefNeRF [41] exploit multi-view reflections on objects but assume the surrounding environment is infinitely far away and is modeled with a flat 2D map. From multi-view reflections, ORCa converts the object surface into a virtual sensor and extracts the 5D radiance field of the environment.", + "Figure 5. Virtual Sensor. We image the world through the object by modeling each pixel's specular radiance as a projection of the 5D radiance field of the environment onto the object's surface. We capture the radiance field by treating the surface area on the object that the pixel views, $\\mathbf{dS}_t$ , as a single-pixel virtual camera with its center-of-projection at $\\mathbf{v}_o$ . We cast virtual cones through the virtual sensor to capture the 5D radiance field of the environment." + ], + "image_footnote": [], + "bbox": [ + 532, + 88, + 864, + 280 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ily responsible for the specular radiance observed on the glossy object. Our key insight is that the reflections captured by the observer's camera can be modeled as a projection of the environment radiance field onto the object's surface. By modeling the reflected rays as a cone and computing the parameters of the cone, we can more accurately estimate the projected environment radiance field onto the object surface, as shown in Fig. 10.", + "bbox": [ + 496, + 412, + 890, + 532 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ORCa is composed of three steps: modeling the object's geometry as a neural implicit surface (Sec. 3.2), converting the object's surface into a virtual sensor (Sec. 3.3), and modeling the environment radiance field as a projection along these virtual cones (Sec. 3.4). The learned environment radiance field can then be queried on novel viewpoints to show occluded areas in the scene. Fig. 2 depicts our output for each component on a scene rendered with a complex glossy object and 3D environment. Fig. 6 shows our system architecture. Next, we describe each step in detail.", + "bbox": [ + 496, + 534, + 892, + 685 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Learning Neural implicit Surfaces", + "text_level": 1, + "bbox": [ + 500, + 694, + 799, + 710 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Neural Signed Distance Function We model the object geometry as a neural signed distance function (SDF). $f: \\mathbb{R}^3 \\to \\mathbb{R}$ . SDFs provide a helpful inductive bias for learning smooth surface geometry [30, 42, 45] that assists downstream tasks in our pipeline. Moreover, the surface properties crucial for our framework, surface normals and curvature, can be conveniently computed from SDFs in a differentiable manner. Consider the 3D spatial coordinates, $\\mathbf{x}$ , in the scene. The glossy object surface, $S$ is then represented by the zero-level set of the SDF", + "bbox": [ + 496, + 717, + 893, + 869 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {S} = \\left\\{f _ {\\mathcal {S}} (\\mathbf {x}) = 0 | \\mathbf {x} \\in \\mathbb {R} ^ {3} \\right\\} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 878, + 890, + 897 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "20776", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b38a38c708192bc41253a0ce6fa511bf54b2beedb6b028f437d397b52b2721d8.jpg", + "image_caption": [ + "Figure 6. Overview of our proposed architecture." + ], + "image_footnote": [], + "bbox": [ + 81, + 89, + 460, + 162 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Similar to Yariv et al. [45], we model the SDF $f_{S}$ as a coordinate-based MLP.", + "bbox": [ + 76, + 212, + 468, + 239 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Surface Normals Gradients of the SDF at the zero level set point $S$ towards the surface normals $S$ ,", + "bbox": [ + 76, + 243, + 468, + 272 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {n} (\\mathbf {x}) = \\frac {\\nabla_ {\\mathbf {x}} f _ {\\mathcal {S}} (\\mathbf {x})}{\\| \\nabla_ {\\mathbf {x}} f _ {\\mathcal {S}} (\\mathbf {x}) \\|} \\quad \\mathbf {x} \\in \\mathcal {S} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 173, + 281, + 468, + 314 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Surface Curvature We employ differential geometry techniques developed by Novello et al. [29] to estimate curvature for neural implicit surfaces. In particular, we estimate the mean curvature $K(\\mathbf{x})$ for the implicit surface from the divergence, $\\nabla$ of the surface normals", + "bbox": [ + 76, + 320, + 468, + 396 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nK (\\mathbf {x}) = \\frac {\\nabla \\cdot \\mathbf {n} (\\mathbf {x})}{2} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 207, + 404, + 468, + 434 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Mean curvature approximates the surface with an osculating sphere. Our approach also works for more generalized notions of curvature through the shape operator, at the cost of higher computational complexity. We refer our readers to the supplement for the general case.", + "bbox": [ + 76, + 441, + 468, + 516 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Diffuse Radiance We separate the captured radiance at the observer camera with diffuse radiance, which depends on the glossy object's albedo, and specular radiance which depends on the environment radiance. The diffuse radiance does not have any view dependence and only depends on surface point $\\mathbf{x}$ . We denote the diffuse radiance as $f_{d}$ and model it using a coordinate-based MLP (Fig. 6).", + "bbox": [ + 76, + 517, + 468, + 621 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Volume Rendering As proposed in [45], we perform volumetric rendering on the SDF. We define the volume density $\\sigma(\\mathbf{x})$ as the cumulative distribution function (CDF), denoted as $\\Psi(s)$ , applied to $f_{S}$ :", + "bbox": [ + 76, + 622, + 468, + 681 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\sigma (\\mathbf {x}) = \\alpha \\Psi_ {\\beta} (f _ {\\mathcal {S}}) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 691, + 468, + 709 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In contrast to [45], however, we only aim to recover the diffuse radiance of the object along a particular ray. We define a function $f_{d}$ that estimates the diffuse radiance at each point, $\\mathbf{x}$ , along the ray. To get the final diffuse radiance along a given primary ray, $\\mathbf{r}_p(t)$ , we perform volumetric rendering:", + "bbox": [ + 76, + 717, + 468, + 806 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {c}} _ {d} (\\mathbf {r}) = \\int_ {0} ^ {\\infty} f _ {d} (\\mathbf {r} (t), f _ {S} ^ {k} (\\mathbf {r} (t)) \\tau (t) d t \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 815, + 468, + 848 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Note that there is no view dependence in Eq. 5 and intermediate features, $f_{\\mathcal{S}}^{k}$ , are used as input. $\\tau(t)$ is the accumulated transmittance along the ray.", + "bbox": [ + 76, + 854, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0529fab554ead1c2b7ef537769a18371272aa8a29144141cda2b7c3683399f99.jpg", + "image_caption": [ + "Figure 7. Qualitative comparisons of diffuse-specular separation and geometry estimation on rendered dataset. The environment contains nearby objects with complex occlusions when seen through reflections on the glossy object. RefNeRF fails to perform accurate diffuse-specular separation and PANDORA blurs the nearby objects in the specular map. ORCA can model the complex specular reflections through environment radiance field." + ], + "image_footnote": [], + "bbox": [ + 501, + 88, + 890, + 383 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Objects Surface as Virtual Sensor", + "text_level": 1, + "bbox": [ + 500, + 512, + 794, + 527 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Each pixel, $\\mathbf{p}$ , with a finite surface area, $\\mathbf{dA}_p$ , on the real-camera sensor views the surface of the object through a frustum originating at that pixel. The object then samples the environment radiance field through this finite surface converting the finite surface into a virtual pixel with surface area, $\\mathbf{dS}$ . Through this model, we can interpret the object surface as a virtual sensor consisting of many virtual pixels that sample radiance from the environment field based on the geometry of the object and observer viewing direction. We now formulate a virtual pixel based on real camera pose and implicit surface geometry. Please refer to Fig. 5 for a visualization of the virtual sensor.", + "bbox": [ + 496, + 536, + 890, + 715 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Consider a real camera origin as $\\mathbf{o}$ and a pixel on the real sensor $\\mathbf{p}_{i,j}$ that corresponds to ray direction $\\mathbf{d}$ . The primary ray for pixel $\\mathbf{p}_{i,j}$ is parameterized with ray length $t$ as $\\mathbf{r_p}(t) = \\mathbf{o} + t\\mathbf{d}$", + "bbox": [ + 496, + 717, + 890, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Casting Real Cones We can approximate the outgoing conical frustum from pixel $\\mathbf{p}_{i,j}$ as a cone originating at $\\mathbf{o}$ with axis-of-direction $\\mathbf{d}$ and radius $\\dot{r}$ , equivalent to half the distance of the pixel in the x and y directions. We represent the real cone as a parametric volume,", + "bbox": [ + 496, + 777, + 890, + 853 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {r} _ {\\text {c o n e}} (\\dot {r}, s, \\theta) = \\dot {r} s \\cos (\\theta) \\hat {\\mathbf {e}} _ {u} + \\dot {r} s \\sin (\\theta) \\hat {\\mathbf {e}} _ {v} + \\dot {r} s \\mathbf {d}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 861, + 890, + 877 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\hat{\\mathbf{e}}_u$ and $\\hat{\\mathbf{e}}_v$ are basis vectors in the plane perpendicular", + "bbox": [ + 500, + 885, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "20777", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "to $\\mathbf{d}$ , $\\theta \\in [0, \\pi]$ and $s \\in [0, t_{\\max}]$", + "bbox": [ + 76, + 90, + 297, + 106 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Virtual Pixel. Virtual pixels are characterized by the intersection of the real cone with the object surface. In Sec. 3.2, we model local surface properties using mean curvature which enables efficient analytical computations for the virtual pixel parameters even though our approach works with general shape operators. For a sampled point $t_i$ along the ray, we have the surface normals $\\mathbf{n}(t_i)$ from Eq. 2 and estimated mean curvature $K(t_i)$ from Eq. 3. The local object surface at $t_i$ , can be approximated with an osculating sphere, $\\mathcal{O}(t_i)$ , centered at $\\hat{\\mathbf{o}}_{\\mathbf{S}}(t_i)$ with radius, $R(t_i)$ as", + "bbox": [ + 75, + 106, + 468, + 258 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nR (t _ {i}) = \\frac {2}{K _ {t _ {i}}} \\hat {\\mathbf {o}} _ {\\mathbf {S}} (t _ {i}) = \\mathbf {r} _ {p} (t _ {i}) + R (t _ {i}) \\cdot \\hat {\\mathbf {n}} (t _ {i})\n$$\n", + "text_format": "latex", + "bbox": [ + 122, + 279, + 423, + 311 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For concave surfaces $K_{t_i} < 0$ . So, $\\hat{\\mathbf{o}}_{\\mathbf{S}}$ will lie outside the object. For $K_{t_i} > 0$ , $\\hat{\\mathbf{o}}_{\\mathbf{S}}$ will lie inside the object.", + "bbox": [ + 75, + 321, + 468, + 349 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The edges of the virtual pixel for $\\mathbf{r_p}(t_i)$ would lie at the intersection of the osculating sphere $\\mathcal{O}(t_i)$ and the primary cone given by $\\mathbf{r}_{\\mathrm{cone}}$ . Computing exact cone-sphere intersections are computationally expensive so we approximate the cone-sphere intersection using rays bound cone-sphere intersectional surface dS. We consider four rays that bound the cone and sample them at $\\theta_{j}\\in \\{0,\\pi /2,\\pi ,3\\pi /2\\}$ with Eq. 6. We perform intersections of the corresponding bounding rays with the osculating sphere $\\mathcal{O}(t_i)$ to get corners of the virtual pixel $\\mathbf{ds}_j$ . These ray sphere intersections can be computed analytically in an efficient manner.", + "bbox": [ + 75, + 351, + 468, + 517 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Virtual Cone Origin. From the virtual pixel surface area, we can now compute the virtual cone that samples the environment. We first compute normal vectors at virtual pixel corners $\\mathbf{ds_j}$ from the center of osculating sphere $\\hat{\\mathbf{o}}_{\\mathrm{S}}$", + "bbox": [ + 76, + 517, + 468, + 578 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {n}} _ {\\mathbf {j}} = \\frac {\\mathbf {d} \\mathbf {s} _ {\\mathbf {j}} - \\hat {\\mathbf {o}} _ {\\mathbf {S}}}{\\left| \\left| \\mathbf {d} \\mathbf {s} _ {\\mathbf {j}} - \\hat {\\mathbf {o}} _ {\\mathbf {S}} \\right| \\right|} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 207, + 589, + 468, + 622 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "At each virtual pixel corner, we compute the reflected ray directions, $\\omega_{j}^{r}$ , by computing the dot product between the incoming ray directions, $\\omega_{k}^{i}$ , and the normals, $\\hat{\\mathbf{n}}_k$ , where $\\omega_0^r$ is the primary ray's reflected vector.", + "bbox": [ + 75, + 631, + 468, + 694 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\omega_ {0} ^ {r} = \\mathbf {d} - \\left(\\mathbf {d} \\cdot \\hat {\\mathbf {n}} \\left(t _ {i}\\right)\\right) \\hat {\\mathbf {n}} \\left(t _ {i}\\right) \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 719, + 468, + 736 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\omega_ {j} ^ {r} = \\mathbf {d} _ {j} - \\left(\\mathbf {d} _ {j} \\cdot \\hat {\\mathbf {n}} _ {j} (k)\\right) \\hat {\\mathbf {n}} _ {j} (k) \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 739, + 468, + 757 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$\\mathbf{d}_j$ are the incident directions to the virtual pixel corners $\\mathbf{ds}_j$ . The virtual cone origin is the intersection of these reflected rays at the pixel corners and pixel center. However, these rays might not intersect at a single point so we approximate a virtual origin to be the point that minimizes the sum of distances to the reflected rays $\\omega_j$ .", + "bbox": [ + 75, + 768, + 468, + 859 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {v} _ {\\mathbf {o}} = \\operatorname {a r g m i n} _ {\\mathbf {v}} \\sum_ {j} | (\\mathbf {v} - \\mathbf {d s} _ {\\mathbf {j}}) \\times \\omega_ {j} ^ {r} | \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 869, + 468, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We pose this as a linear least squares problem and estimate the virtual cone origin efficiently through pseudo-inverse.", + "bbox": [ + 498, + 90, + 890, + 121 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Virtual Cones Direction. The reflected ray at the center of the virtual pixel reflects the object surface along the direction $\\omega_0^r$ from Eq. 8. We consider this as the direction-of-axis of the virtual cone.", + "bbox": [ + 498, + 122, + 890, + 181 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {v}} _ {d} = \\omega_ {0} ^ {r} \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 663, + 196, + 890, + 213 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Virtual Cone Radius. We compute the radius of the cone by treating the reflection vectors of the bounding rays as the neighboring \"pixel\" directions. Similar to [5], we can compute the distance between $\\{\\omega_{k_{\\theta}}^{r}\\}_{\\theta = 0}^{2\\pi}$ and the primary reflected ray $\\omega_0^r$ in the $(x,y)$ components (omitted below for clarity).", + "bbox": [ + 498, + 227, + 890, + 318 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {v}} _ {\\dot {r}} = \\left\\| \\left\\{\\omega_ {k _ {\\theta}} ^ {r} \\right\\} _ {\\theta = 0} ^ {2 \\pi} - \\omega_ {0} ^ {r} \\right\\| \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 347, + 890, + 364 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Finally, for each sampled point $t_i$ , we can characterize our single-pixel virtual sensor located at the object surface $\\mathbf{dS}$ as a virtual cone with $\\hat{\\mathbf{v}}_o$ as its apex, $\\hat{\\mathbf{v}}_d$ as axis-direction, $\\hat{\\mathbf{v}}_{\\dot{r}}$ as the radius.", + "bbox": [ + 498, + 378, + 890, + 438 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Connections to caustics. Our work takes inspiration from catadioptric imaging systems. To convert objects into cameras, we compute the surface and find a corresponding center-of-projection for this surface-as-sensor. However, unlike conventional perspective cameras, objects don't have a fixed center-of-projection, other than in a few special configurations [3], but a locus of viewpoints that vary with object geometry and viewing direction. These viewpoints lie on the \"caustic surface\" of the object. While typical works in catadioptric imaging analytically compute the caustic surface by assuming known geometry [14, 38], or making assumptions about placement of the observer [40], our formulation approximates the caustic surface of unknown geometry through the intersection of reflected rays on virtual pixels. We empirically show in the supplement that as the surface area of the virtual pixel goes to $0$ , $d\\mathbf{S} \\rightarrow 0$ , our method estimates the true caustic of the object without assuming geometry. Our method also has applications in estimating the caustic surface of the unknown geometry.", + "bbox": [ + 496, + 440, + 890, + 728 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4. Environment Radiance Fields", + "text_level": 1, + "bbox": [ + 500, + 739, + 763, + 753 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our goal is to capture a 5D environment radiance field of the scene by imaging the world through these single-pixel virtual sensors located at the object's surface. We use our formulation of virtual cones to recover 5D environment radiance fields. We define an environment radiance field as $f_{\\mathcal{E}}: (\\hat{\\mathbf{v}}_o, \\hat{\\mathbf{v}}_d) \\to (\\sigma^{Env}, c_s)$ ,", + "bbox": [ + 498, + 763, + 890, + 854 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $f_{\\varepsilon}$ outputs opacity and radiance along sampled virtual cones. We note that this view-dependent radiance is equivalent to the specular radiance at point $t_i$ sampled", + "bbox": [ + 498, + 854, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "20778", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/413e0df2124e259ffdb6edf7e6eeb29af2b2fa53113aac8e2ddfce0478c9883e.jpg", + "image_caption": [ + "Figure 8. Evaluation of environment depth from reflections. For the scene in Fig. 3, ORCa can faithfully estimate the depth of the environment, especially for nearby surroundings." + ], + "image_footnote": [], + "bbox": [ + 81, + 87, + 467, + 169 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/36134989eff7efd70451770de2133c601015ac8f38333e939ea7ee9bed43dc08.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ApproachDiffuse RadianceSpecular RadianceMixed RadianceNormals
PSNR ↑ (dB)SSIM ↑PSNR ↑ (dB)SSIM ↑PSNR ↑ (dB)SSIM ↑MAE ↓ (°)
Ref-NeRF18.800.730416.990.663320.190.789043.690
PANDORA18.250.726016.250.648318.900.72847.606
ORCA19.840.789320.740.753522.000.79472.339
", + "bbox": [ + 78, + 237, + 465, + 301 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "along the primary-camera ray $\\mathbf{r}_{\\mathbf{p}}(t)$ . We can render the final specular radiance at pixel $\\mathbf{p}_{i,j}$ as follows:", + "bbox": [ + 75, + 422, + 468, + 455 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {c}} _ {s} (\\mathbf {r}) = \\int_ {0} ^ {\\infty} f _ {\\mathcal {E}} (\\hat {\\mathbf {v}} _ {o}, \\hat {\\mathbf {v}} _ {d}) \\tau (t) d t\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 465, + 377, + 500 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {c}} = \\hat {\\mathbf {c _ {d}}} + \\hat {\\mathbf {c _ {s}}}\n$$\n", + "text_format": "latex", + "bbox": [ + 228, + 503, + 316, + 517 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Intuitively, $f_{\\mathcal{E}}$ learns the 5D radiance field by sampling single-pixel virtual sensors from the object surface and must learn geometry and environment radiance that is consistent with the multi-view reflections. Moreover, we can query $f_{\\mathcal{E}}$ to render novel viewpoints and associated depths that are beyond the field-of-view of the real camera. We volume render each virtual cone by dividing them into conical frustums using Integrated-Positional Encoding as proposed in MipNeRF [5]. Our formulation of virtual cones works well with Mip-Nerf's rays-as-cones method.", + "bbox": [ + 75, + 532, + 470, + 684 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 700, + 209, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 76, + 724, + 294, + 742 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As in PANDORA, we parameterize $f_{S}$ with an 8-layer MLP to estimate the surface, and, as in MipNeRF, $f_{d}$ with 4-layer MLP with input geometric features of size 512 from $f_{S}$ . We follow the SDF-to-opacity conversion and the iterative sampling of the ray proposed in [45]. To aid the network to learn geometry quickly, we also train $f_{S}$ with a mask-net as proposed in [10]. We use five losses in our architecture: photometric loss, mask loss [10], normal loss [41], eikonal loss [45], and distortion loss [6]. Additional training details are discussed in the supplement.", + "bbox": [ + 75, + 750, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9dfddaa4d2b31821db93af595b01f37bdd75890edf9dd3a0f7798ce08951ab91.jpg", + "table_caption": [ + "Table 1. Average evaluation metrics on rendered scenes. We compare ORCa to other neural rendering techniques that model reflections, including Ref-NeRF and PANDORA, on six simulated scenes. ORCa provides consistent improvements in geometry estimation, diffuse-specular separation and novel view synthesis. Please refer to the supplement for additional metrics." + ], + "table_footnote": [], + "table_body": "
Scene: Ball-cup in hallway
ApproachMixedRadiance
PSNRSSIM
↑(dB)↑
Ref-NeRF32.750.9617
PANDORA28.830.9758
ORCA30.860.9799
", + "bbox": [ + 519, + 88, + 692, + 172 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/68eacca3ce83ef7e3a267f0d5ead74177f52e370e2e57a2958b3e9834cce8f7f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Scene: Owl in hallway
ApproachMixed Radiance
PSNRSSIM
↑ (dB)↑
Ref-NeRF26.650.8890
PANDORA27.240.9343
ORCA26.840.9299
", + "bbox": [ + 694, + 88, + 870, + 172 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Quantitative metrics for real-world, captured scenes. ORCa demonstrates comparable performance in novel view synthesis on scenes from the PANDORA real dataset.", + "bbox": [ + 498, + 183, + 890, + 224 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Datasets", + "text_level": 1, + "bbox": [ + 500, + 251, + 601, + 266 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct experiments on both simulated and real-world datasets. Simulated datasets are rendered in Mitsu2 [28]. Simulated datasets contain a range of increasingly complex object geometries (elephant, Pokeball, and orca) and scenes (living room and Pokemon). We train with 200 views for simulated datasets. We also show results for a real-world dataset [10] capturing a glossy cup with a black vase sitting atop it using 35 views. All datasets will be publicly released upon publication.", + "bbox": [ + 496, + 275, + 890, + 411 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Advantages of Environment Radiance Fields", + "text_level": 1, + "bbox": [ + 498, + 421, + 877, + 436 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Other neural rendering techniques that handle reflections, Ref-NeRF and PANDORA, estimate the environment as a 2D map, while ORCa recovers a 5D environment radiance field. Fig. 3 shows the advantages of recovering a 5D environment radiance field. Close-by surrounding objects often cause occlusions that cannot be modeled by 2D environment maps. By estimating the radiance field, we can image behind occluders through sampling novel viewpoints such as the translated viewpoints shown in Fig. 3. Moreover, we can also show the depth to the surroundings from these virtual viewpoints. In Fig. 8, we quantitatively evaluate the estimated depth of the surroundings for the synthesized virtual views in Fig. 3 and show that ORCa provides reliable depth estimates especially for nearby objects. We provide additional examples of depth estimation and beyond field-of-view novel-view synthesis in the supplement.", + "bbox": [ + 496, + 444, + 890, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Comparisons with Baselines", + "text_level": 1, + "bbox": [ + 500, + 696, + 751, + 712 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compare ORCa with Ref-NeRF and PANDORA on synthetic and real datasets. As shown in Fig. 7 and 9, ORCa estimates more accurate surface normals than other methods. While the total radiance predicted by Ref-NeRF and PANDORA are visually similar, the surface normals are less smooth than ORCa. We also observe that ORCa is able to achieve better diffuse and specular radiance separation than PANDORA, which is evident in the Pokeball surface normals Fig. 7. In these examples, PANDORA recovers blurry specular radiance. We see that ORCa's predicted depth is highly interpretable and matches the underlying geometry of the environment, as shown in Fig. 3. Even on cylindri", + "bbox": [ + 496, + 719, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "20779", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b8c49e4f653135a5bfcfea828e7f63b0338865f2c803608894e4402fc01b0a06.jpg", + "image_caption": [ + "Figure 9. Comparisons on real dataset. Using a real-world dataset with only 35 views, ORCa can model the sharp specularities on the ball arriving from regions of the nearby scene, such as the table, and far away scene regions, such as the hallway, to learn an environment radiance field. We query the radiance field for the depth of the far hallway (blue) and the nearby objects, such as the table (red). We also render novel viewpoints that are beyond the field of view of the observer camera and show that ORCa interpolates well between those views." + ], + "image_footnote": [], + "bbox": [ + 80, + 88, + 321, + 361 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cb2188ee0525ae84b3267adaa2b3c2983a96e4c7147e764a619a70c63dbd4a81.jpg", + "image_caption": [ + "Figure 10. Ablations on virtual cone computation Accurate estimation of the virtual cones (Sec. 3.3) is crucial for ORCa. If the virtual cone origin is assumed to be at the object surface (left column) or if the surface is assumed to locally have no curvature (right column), the surface normal and specular radiance outputs suffer from artifacts (red boxes)." + ], + "image_footnote": [], + "bbox": [ + 326, + 88, + 470, + 358 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "cal real-world datasets, such as the black vase in Fig. 9, the nearby hallway is visible in both the virtual view and depth, despite never being in the field of view of the primary camera. Unlike Ref-NeRF, our primary objective is not to perform novel-view synthesis, but instead to capture the environment radiance field from the object surface. As shown in Table 1, ORCa improves over Ref-NeRF and PANDORA in estimating diffuse radiance, specular radiance, mixed radiance, and normals. Here we report the image quality metrics, PSNR and SSIM, and surface normal metric, MAE, averaged over six synthetic scenes with varying environments, camera view distribution, and object complexity. Please refer to the supplement for individual metrics on each scene. ORCa also shows competitive performance for novel-view synthesis metrics on real-world scenes from PANDORA dataset, as shown in Table 2.", + "bbox": [ + 75, + 521, + 468, + 762 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Impact of Virtual Cone Computation", + "text_level": 1, + "bbox": [ + 76, + 771, + 398, + 787 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We base our method on a physically accurate formulation by modeling ray-cone intersections and using the surface as a virtual sensor, as described in Sec. 3.3. We demonstrate the importance of this step by setting up two ablation experiments. In the first experiment, which we term Naive Virtual Cone, we place the origin of the cone at the object surface instead of computing the virtual cone origin based", + "bbox": [ + 75, + 795, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a17c9af40a96d28275bd72eb531568ca8e35ab246ece9f22045da15905766aba.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 90, + 893, + 223 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "on the curvature. In the second experiment, which we term No Curvature, we assume that locally the surface is like a flat mirror and has no curvature. In Fig. 10, we show that for both of these ablation experiments, we see worse performance as demonstrated by the artifacts in estimated surface normals and specular components (shown as red boxes).", + "bbox": [ + 498, + 353, + 893, + 444 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 472, + 617, + 487 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present a method to convert glossy objects with unknown geometry and texture into radiance-field cameras that capture the environment's radiance field around them. Our method recovers object geometry and diffuse radiance, in addition to capturing the depth and radiance of the object's surroundings from its perspective. Our modeling of the environment as a radiance field is effective in recovering close-by objects (Fig. 7) and is occlusion aware (Fig. 3). From the recovered environment radiance field, we can perform beyond field-of-view novel-view synthesis. Our work can unleash applications in virtual object insertion and 3D perception, e.g. inferring information beyond the line-of-sight of the camera using predicted virtual views and depth. Our formulation goes beyond the conventional direct-line-of-sight radiance fields and can enable further areas of research to extract more information from multi-view images directly from the environment and the objects present in it.", + "bbox": [ + 496, + 502, + 893, + 760 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 500, + 787, + 658, + 804 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This research was supported by the SMART Contract IARPA Grant 2021-20111000004. We would also like to thank Systems Technology Research (STR). AD and AV are supported by NSF Expeditions-Award IIS-1730574 and NSF CAREER-Award IIS-1652633.", + "bbox": [ + 498, + 816, + 893, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "20780", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Anna Alperovich, Ole Johannsen, and Bastian Goldluecke. Intrinsic light field decomposition and disparity estimation with deep encoder-decoder network. In 2018 26th European Signal Processing Conference (EUSIPCO), pages 2165-2169. IEEE, 2018. 3", + "[2] S. Baker and S.K. Nayar. A theory of catadioptric image formation. In Sixth International Conference on Computer Vision (IEEE Cat. No.98CH36271), pages 35-42, 1998. 3", + "[3] Simon Baker and Shree K. Nayar. A theory of single-viewpoint catadioptric image formation. International Journal of Computer Vision, 35:175-196, 2004. 6", + "[4] Jonathan T Barron and Jitendra Malik. Shape, illumination, and reflectance from shading. IEEE transactions on pattern analysis and machine intelligence, 37(8):1670-1687, 2014. 3", + "[5] Jonathan T. Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P. Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. ICCV, 2021. 3, 6, 7", + "[6] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5470–5479, 2022. 7", + "[7] Mark Boss, Raphael Braun, Varun Jampani, Jonathan T Barron, Ce Liu, and Hendrik Lensch. Nerd: Neural reflectance decomposition from image collections. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12684-12694, 2021. 4", + "[8] Mark Boss, Varun Jampani, Raphael Braun, Ce Liu, Jonathan Barron, and Hendrik Lensch. Neural-pil: Neural pre-integrated lighting for reflectance decomposition. Advances in Neural Information Processing Systems, 34:10691-10704, 2021. 4", + "[9] Akshit Dave, Yannick Hold-Geoffroy, Milos Hasan, Kalyan Sunkavalli, and Ashok Veeraraghavan. Snapshot polarimetric diffuse-specular separation. Optics Express, 30(19):34239-34255, 2022. 3", + "[10] Akshit Dave, Yongyi Zhao, and Ashok Veeraraghavan. Pandora: Polarization-aided neural decomposition of radiance. arXiv preprint arXiv:2203.13458, 2022. 3, 4, 7", + "[11] Michael De Zeeuw and Aswin C Sankaranarayanan. Wide-baseline light fields using ellipsoidal mirrors. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 3", + "[12] Mathieu Garon, Kalyan Sunkavalli, Sunil Hadap, Nathan Carr, and Jean-François Lalonde. Fast spatially-varying indoor lighting estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6908-6917, 2019. 3", + "[13] Stamatios Georgoulis, Konstantinos Rematas, Tobias Ritschel, Mario Fritz, Tinne Tuytelaars, and Luc Van Gool. What is around the camera? In Proceedings of the IEEE International Conference on Computer Vision, pages 5170-5178, 2017. 3" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] J. Gluckman and S.K. Nayar. Planar catadioptric stereo: geometry and calibration. In Proceedings. 1999 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (Cat. No PR00149), volume 1, pages 22-28 Vol. 1, 1999. 6", + "[15] Michael D Grossberg and Shree K Nayar. The raxel imaging model and ray-based calibration. International Journal of Computer Vision, 61(2):119, 2005. 3", + "[16] Yuan-Chen Guo, Di Kang, Linchao Bao, Yu He, and Song-Hai Zhang. Nerfren: Neural radiance fields with reflections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18409-18418, June 2022. 3", + "[17] Connor Henley, Siddharth Somasundaram, Joseph Hollmann, and Ramesh Raskar. Detection and mapping of specular surfaces using multibounce lidar returns. Optics Express, 31(4):6370-6388, 2023. 3", + "[18] Jan Jachnik, Richard A Newcombe, and Andrew J Davison. Real-time surface light-field capture for augmentation of planar specular surfaces. In 2012 IEEE International Symposium on Mixed and Augmented Reality (ISMAR), pages 91-97. IEEE, 2012. 3", + "[19] Georgios Kopanas, Thomas Leimkuhler, Gilles Rainer, Clément Jambon, and George Drettakis. Neural point cata-caustics for novel-view synthesis of reflections. ACM Transactions on Graphics, 41(6):Article-201, 2022. 3", + "[20] Rui Li, Simeng Qiu, Guangming Zang, and Wolfgang Heidrich. Reflection separation via multi-bounce polarization state tracing. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIII 16, pages 781-796. Springer, 2020. 3", + "[21] Tingtian Li, Daniel PK Lun, Yuk-Hee Chan, et al. Robust reflection removal based on light field imaging. IEEE Transactions on Image Processing, 28(4):1798–1812, 2018. 3", + "[22] Zhengqin Li, Mohammad Shafiei, Ravi Ramamoorthi, Kalyan Sunkavalli, and Manmohan Chandraker. Inverse rendering for complex indoor scenes: Shape, spatially-varying lighting and svbrdf from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2475-2484, 2020. 3", + "[23] Stephen Lombardi and Ko Nishino. Reflectance and natural illumination from a single image. In European Conference on Computer Vision, pages 582-595. Springer, 2012. 3", + "[24] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 3", + "[25] Jacob Munkberg, Jon Hasselgren, Tianchang Shen, Jun Gao, Wenzheng Chen, Alex Evans, Thomas Müller, and Sanja Fidler. Extracting triangular 3d models, materials, and lighting from images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8280-8290, 2022. 4", + "[26] Shree K Nayar and Simon Baker. Catadioptric image formation. In Proceedings of the 1997 DARPA Image Understanding Workshop, pages 1431-1437, 1997. 3" + ], + "bbox": [ + 501, + 92, + 893, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "20781", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Yun Ni, Jie Chen, and Lap-Pui Chau. Reflection removal on single light field capture using focus manipulation. IEEE Transactions on Computational Imaging, 4(4):562-572, 2018. 3", + "[28] Merlin Nimier-David, Delio Vicini, Tizian Zeltner, and Wenzel Jakob. Mitsuba 2: A retargetable forward and inverse renderer. ACM Transactions on Graphics (TOG), 38(6):1-17, 2019. 7", + "[29] Tiago Novello, Guilherme Schardong, Luiz Schirmer, Vinicius da Silva, Helio Lopes, and Luiz Velho. Exploring differential geometry in neural implicits, 2022. 5", + "[30] Michael Oechsle, Songyou Peng, and Andreas Geiger. Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5589-5599, 2021. 4", + "[31] Jeong Joon Park, Aleksander Holynski, and Steven M Seitz. Seeing the world in a bag of chips. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1417-1427, 2020. 4", + "[32] Ravi Ramamoorthi and Pat Hanrahan. A signal-processing framework for inverse rendering. In Proceedings of the 28th annual conference on Computer graphics and interactive techniques, pages 117-128, 2001. 3", + "[33] Raskar Ramesh and James Davis. 5d time-light transport matrix: What can we reason about scene properties? Technical report, 2008. 3", + "[34] Fabiano Romeiro and Todd Zickler. Blind reflectometry. In European conference on computer vision, pages 45-58. Springer, 2010. 3", + "[35] Vincent Sitzmann, Semon Rezchikov, Bill Freeman, Josh Tenenbaum, and Fredo Durand. Light field networks: Neural scene representations with single-evaluation rendering. Advances in Neural Information Processing Systems, 34:19313-19325, 2021. 3", + "[36] Shuran Song and Thomas Funkhouser. Neural illumination: Lighting prediction for indoor environments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6918-6926, 2019. 4", + "[37] Pratul P Srinivasan, Ben Mildenhall, Matthew Tancik, Jonathan T Barron, Richard Tucker, and Noah Snavely. Lighthouse: Predicting lighting volumes for spatially-coherent illumination. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8080-8089, 2020. 4", + "[38] R. Swaminathan, M.D. Grossberg, and S.K. Nayar. Cautistics of catadioptric cameras. In Proceedings Eighth IEEE International Conference on Computer Vision. ICCV 2001, volume 2, pages 2-9 vol.2, 2001. 6", + "[39] Tristan Swedish, Connor Henley, and Ramesh Raskar. Objects as cameras: Estimating high-frequency illumination from shadows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2593-2602, 2021. 4", + "[40] Yuichi Taguchi, Amit Agrawal, Ashok Veeraraghavan, Srikumar Ramalingam, and Ramesh Raskar. Axial-cones: Modeling spherical catadioptric cameras for wide-angle light field rendering. ACM Trans. Graph., 29(6), dec 2010. 3, 6" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T. Barron, and Pratul P. Srinivasan. Ref-NeRF: Structured view-dependent appearance for neural radiance fields. CVPR, 2022. 3, 4, 7", + "[42] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021. 4", + "[43] Zian Wang, Jonah Philion, Sanja Fidler, and Jan Kautz. Learning indoor inverse rendering with 3d spatially-varying lighting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12538-12547, 2021. 3", + "[44] Ziyu Wang, Liao Wang, Fuqiang Zhao, Minye Wu, Lan Xu, and Jingyi Yu. Mirrornerf: One-shot neural portrait radiance field from multi-mirror catadioptric imaging. In 2021 IEEE International Conference on Computational Photography (ICCP), pages 1-12. IEEE, 2021. 3", + "[45] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. In Thirty-Fifth Conference on Neural Information Processing Systems, 2021. 4, 5, 7", + "[46] Kai Zhang, Fujun Luan, Qianqian Wang, Kavita Bala, and Noah Snively. Physg: Inverse rendering with spherical gaussians for physics-based material editing and relighting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5453-5462, 2021. 4", + "[47] Xiuming Zhang, Pratul P Srinivasan, Boyang Deng, Paul Debevec, William T Freeman, and Jonathan T Barron. Nerfactor: Neural factorization of shape and reflectance under an unknown illumination. ACM Transactions on Graphics (TOG), 40(6):1-18, 2021. 4", + "[48] Rui Zhu, Zhengqin Li, Janarbek Matai, Fatih Porikli, and Manmohan Chandraker. Irisformer: Dense vision transformers for single-image inverse rendering in indoor scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2822-2831, 2022. 3" + ], + "bbox": [ + 501, + 92, + 890, + 613 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "20782", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/ORCa_ Glossy Objects As Radiance-Field Cameras/da15cd79-f9cc-4796-854b-9e2af61a06e8_model.json b/2023/ORCa_ Glossy Objects As Radiance-Field Cameras/da15cd79-f9cc-4796-854b-9e2af61a06e8_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3106f54050b5200ba72af305271499817f0b68b7 --- /dev/null +++ b/2023/ORCa_ Glossy Objects As Radiance-Field Cameras/da15cd79-f9cc-4796-854b-9e2af61a06e8_model.json @@ -0,0 +1,2585 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.228, + 0.131, + 0.745, + 0.154 + ], + "angle": 0, + "content": "ORCa: Glossy Objects as Radiance-Field Cameras" + }, + { + "type": "text", + "bbox": [ + 0.204, + 0.18, + 0.766, + 0.216 + ], + "angle": 0, + "content": "Kushagra Tiwary\\(^{1*}\\), Akshit Dave\\(^{2*}\\), Nikhil Behari\\(^{1}\\), Tzofi Klinghoffer\\(^{1}\\), Ashok Veeraraghavan\\(^{2}\\), Ramesh Raskar\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.26, + 0.217, + 0.71, + 0.235 + ], + "angle": 0, + "content": "1Massachusetts Institute of Technology, 2Rice University" + }, + { + "type": "text", + "bbox": [ + 0.207, + 0.237, + 0.76, + 0.253 + ], + "angle": 0, + "content": "{ktiwary,behari,tzofi,raskar}@mit.edu，{akshat,vashok}@rice.edu" + }, + { + "type": "image_caption", + "bbox": [ + 0.167, + 0.288, + 0.266, + 0.311 + ], + "angle": 0, + "content": "(a) Scene: Glossy Elephant in a room" + }, + { + "type": "image", + "bbox": [ + 0.119, + 0.313, + 0.318, + 0.499 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.324, + 0.288, + 0.444, + 0.311 + ], + "angle": 0, + "content": "(b) Input: images with distorted reflections" + }, + { + "type": "image", + "bbox": [ + 0.319, + 0.313, + 0.45, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.319, + 0.407, + 0.449, + 0.499 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.456, + 0.394, + 0.516, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.538, + 0.288, + 0.702, + 0.311 + ], + "angle": 0, + "content": "(c) Using objects to image beyond camera's field-of-view" + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.313, + 0.719, + 0.499 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.735, + 0.288, + 0.835, + 0.311 + ], + "angle": 0, + "content": "(d) Recovered and Undistorted RGBs" + }, + { + "type": "image", + "bbox": [ + 0.721, + 0.313, + 0.851, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.721, + 0.407, + 0.851, + 0.499 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.509, + 0.895, + 0.566 + ], + "angle": 0, + "content": "Figure 1. Objects as radiance-field cameras. We convert everyday objects with unknown geometry (a) into radiance-field cameras by modeling multi-view reflections (b) as projections of the 5D radiance field of the environment. We convert the object surface into a virtual sensor to capture this radiance field (c), which enables depth and radiance estimation of the surrounding environment. We can then query this radiance field to perform beyond field-of-view novel view synthesis of the environment (d)." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.578, + 0.314, + 0.594 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.612, + 0.474, + 0.87 + ], + "angle": 0, + "content": "Reflections on glossy objects contain valuable and hidden information about the surrounding environment. By converting these objects into cameras, we can unlock exciting applications, including imaging beyond the camera's field-of-view and from seemingly impossible vantage points, e.g. from reflections on the human eye. However, this task is challenging because reflections depend jointly on object geometry, material properties, the 3D environment, and the observer's viewing direction. Our approach converts glossy objects with unknown geometry into radiance-field cameras to image the world from the object's perspective. Our key insight is to convert the object surface into a virtual sensor that captures cast reflections as a 2D projection of the 5D environment radiance field visible to and surrounding the object. We show that recovering the environment radiance fields enables depth and radiance estimation from the object to its surroundings in addition to beyond field-of-view" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.58, + 0.892, + 0.701 + ], + "angle": 0, + "content": "novel-view synthesis, i.e. rendering of novel views that are only directly visible to the glossy object present in the scene, but not the observer. Moreover, using the radiance field we can image around occluders caused by close-by objects in the scene. Our method is trained end-to-end on multi-view images of the object and jointly estimates object geometry, diffuse radiance, and the 5D environment radiance field. For more information, visit our website." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.71, + 0.633, + 0.725 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Imagine that you're driving down a city street that is packed with lines of parked cars on both sides. Inspection of the cars' glass windshields, glossy paint, and plastic reveal sharp, but faint and distorted views of the surroundings that might be otherwise hidden from you. Humans can infer depth and semantic cues about the occluded areas in the environment by processing reflections visible on reflective objects, internally decomposing the object geometry and radiance from the specular radiance being reflected onto it. Our aim is to decompose the object from its reflections to \"see\" the world from the object's perspective, effectively turning" + }, + { + "type": "page_footnote", + "bbox": [ + 0.097, + 0.887, + 0.204, + 0.9 + ], + "angle": 0, + "content": "* Equal contribution" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "20773" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.102, + 0.093, + 0.33, + 0.106 + ], + "angle": 0, + "content": "(a) Learning Neural Implicit Surfaces [Sec. 3.2]" + }, + { + "type": "image_caption", + "bbox": [ + 0.382, + 0.093, + 0.606, + 0.106 + ], + "angle": 0, + "content": "(b) Objects Surface as Virtual Sensor [Sec. 3.3]" + }, + { + "type": "image_caption", + "bbox": [ + 0.629, + 0.093, + 0.88, + 0.106 + ], + "angle": 0, + "content": "(c) Learning Environment Radiance Fields [Sec. 3.4]" + }, + { + "type": "image", + "bbox": [ + 0.09, + 0.12, + 0.316, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.368, + 0.117, + 0.46, + 0.126 + ], + "angle": 0, + "content": "Real camera origin" + }, + { + "type": "image_footnote", + "bbox": [ + 0.372, + 0.128, + 0.447, + 0.138 + ], + "angle": 0, + "content": "Real pixel cone" + }, + { + "type": "image_footnote", + "bbox": [ + 0.373, + 0.141, + 0.462, + 0.15 + ], + "angle": 0, + "content": "- Surface intersection" + }, + { + "type": "image_footnote", + "bbox": [ + 0.373, + 0.154, + 0.454, + 0.163 + ], + "angle": 0, + "content": "Virtual pixel cone" + }, + { + "type": "image_footnote", + "bbox": [ + 0.373, + 0.165, + 0.468, + 0.175 + ], + "angle": 0, + "content": "Virtual camera origin" + }, + { + "type": "list", + "bbox": [ + 0.368, + 0.117, + 0.468, + 0.175 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.17, + 0.593, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.622, + 0.125, + 0.71, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.621, + 0.198, + 0.705, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.718, + 0.179, + 0.791, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.803, + 0.133, + 0.876, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.815, + 0.193, + 0.865, + 0.202 + ], + "angle": 0, + "content": "Virtual View" + }, + { + "type": "image", + "bbox": [ + 0.803, + 0.204, + 0.89, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.812, + 0.262, + 0.867, + 0.273 + ], + "angle": 0, + "content": "Virtual Depth" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.295, + 0.895, + 0.366 + ], + "angle": 0, + "content": "Figure 2. ORCa Overview. We jointly estimate the object's geometry and diffuse along with the environment radiance field estimation through a three-step approach. First, we model the object as a neural implicit surface (a). We model the reflections as probing the environment on virtual viewpoints (b) estimated analytically from surface properties. We model the environment as a radiance field queried on these viewpoints (c). Both neural implicit surface and environment radiance field are trained jointly on multi-view images of the object using a photometric loss." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.392, + 0.471, + 0.466 + ], + "angle": 0, + "content": "the object into a camera that images its environment. However, reflections pose a long-standing challenge in computer vision as the reflections are a 2D projection of an unknown 3D environment that is distorted based on the shape of the reflector." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.473, + 0.47, + 0.744 + ], + "angle": 0, + "content": "To capture the 3D world from the object's perspective, we model the object's surface as a virtual sensor that captures the 2D projection of a 5D environment radiance field surrounding the object. This environment radiance field consists largely of areas only visible to the observer through the object's reflections. Our use of environment radiance fields not only enables depth and radiance estimation from the object to its surroundings but also enables beyond field-of-view novel-view synthesis, i.e. rendering of novel views that are only directly visible to the glossy object present in the scene but not the observer. Unlike conventional approaches that model the environment as a 2D map, our approach models it as a 5D field without assuming the scene is infinitely far away. Moreover, by sampling the 5D radiance field, instead of a 2D map, we can capture depth and images around occluders, such as close-by objects in the scene, as shown in Fig. 3. These applications cannot be done from a 2D environment map." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.902 + ], + "angle": 0, + "content": "We aim to decompose reflections on the object's surface, from its surface and exploit those reflections to construct a radiance field surrounding the object, therefore capturing the 3D world in the process. This is a challenging task because the reflections are extremely sensitive to local object geometry, viewing direction and inter-reflections due to the object's surface. To capture this radiance field, we convert glossy objects with unknown geometry and texture into radiance-field cameras. Specifically, we exploit neural rendering to estimate the local surface of the object viewed" + }, + { + "type": "image_caption", + "bbox": [ + 0.51, + 0.39, + 0.614, + 0.402 + ], + "angle": 0, + "content": "(a) Sample Input Views" + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.404, + 0.613, + 0.517 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.626, + 0.39, + 0.776, + 0.401 + ], + "angle": 0, + "content": "(b) Virtual Views with Translation" + }, + { + "type": "image", + "bbox": [ + 0.617, + 0.402, + 0.785, + 0.518 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.784, + 0.391, + 0.882, + 0.401 + ], + "angle": 0, + "content": "(c) Virtual View Depth" + }, + { + "type": "image", + "bbox": [ + 0.787, + 0.402, + 0.875, + 0.517 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.603, + 0.522, + 0.798, + 0.531 + ], + "angle": 0, + "content": "(d) Parallax effect in 5D Env. Radiance Field" + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.532, + 0.636, + 0.608 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.644, + 0.532, + 0.754, + 0.607 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.764, + 0.532, + 0.874, + 0.607 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.609, + 0.636, + 0.684 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.644, + 0.609, + 0.754, + 0.683 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.764, + 0.609, + 0.874, + 0.683 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.517, + 0.685, + 0.585, + 0.694 + ], + "angle": 0, + "content": "x 0" + }, + { + "type": "image_caption", + "bbox": [ + 0.687, + 0.685, + 0.712, + 0.694 + ], + "angle": 0, + "content": "+0.25" + }, + { + "type": "image_caption", + "bbox": [ + 0.806, + 0.685, + 0.833, + 0.694 + ], + "angle": 0, + "content": "+0.50" + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.709, + 0.892, + 0.807 + ], + "angle": 0, + "content": "Figure 3. Advantages of 5D environment radiance field. Modeling reflections on object surfaces (a) as a 5D env. radiance field enables beyond field-of-view novel-view synthesis, including rendering of the environment from translated virtual camera views (b). Depth (c) and environment radiance of translated and parallax views can further enable imaging behind occluders, for example revealing the tails behind the primary Pokemon occluders (d)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.903 + ], + "angle": 0, + "content": "from each pixel of the real camera. We then convert this local surface into a virtual pixel that captures radiance from the environment. This virtual pixel captures the environment radiance as shown in Fig 5. We estimate the outgo" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20774" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.182 + ], + "angle": 0, + "content": "ing frustum from the virtual pixel as a cone that samples the scene. By sampling the scene from many virtual pixels on the object surface, we construct an environment radiance field that can be queried independently of the object surface, enabling beyond field-of-view novel-view synthesis from previously unsampled viewpoints." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.183, + 0.471, + 0.392 + ], + "angle": 0, + "content": "Our approach jointly estimates object geometry, diffuse radiance, and the environment radiance field from multiview images of glossy objects with unknown geometry and diffuse texture in three steps. First, we use neural signed distance functions (SDF) and an MLP to model the glossy object's geometry as a neural implicit surface and diffuse radiance, respectively, similar to PANDORA [10]. Then, for every pixel on the observer's camera, we estimate the virtual pixels on the object's surface based on the estimated local geometry from the neural SDF. We analytically compute the parameters of the virtual cone through the virtual pixel. Lastly, we use the cone formulation in MipNeRF [5] to cast virtual cones from the virtual camera to recover the environment radiance." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.394, + 0.445, + 0.409 + ], + "angle": 0, + "content": "To summarize, we make the following contributions:" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.414, + 0.468, + 0.459 + ], + "angle": 0, + "content": "- We present a method to convert implicit surfaces into virtual sensors that can image their surroundings using virtual cones. (Sec. 3.3)" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.467, + 0.468, + 0.512 + ], + "angle": 0, + "content": "- We jointly estimate object geometry, diffuse radiance, and estimate the 5D environment radiance field surrounding the object. (Fig. 7 & 9)" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.52, + 0.468, + 0.58 + ], + "angle": 0, + "content": "- We show that the environment radiance field can be queried to perform beyond-field-of-view novel viewpoint synthesis, i.e. render views only visible to the object in the scene (Section 3.4)" + }, + { + "type": "list", + "bbox": [ + 0.097, + 0.414, + 0.468, + 0.58 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.585, + 0.47, + 0.812 + ], + "angle": 0, + "content": "Scope. We only model glossy objects with low roughness as such specular reflections tend to have a high signal-to-noise ratio, therefore, are a sharper estimate of the environment radiance field. However, we note that the virtual cone computation can be extended to model the cone radius as a function of surface roughness. Deblurring approaches can further improve the resolution of estimated environment. In addition, we approximate the local curvature using mean curvature, which fails for objects with varying radius of curvature along the tangent space. We explain how our virtual cone curvature estimation can be extended to handle general shape operators in the supplementary material. Lastly, similar to other multi-view approaches, our approach relies on a sufficient virtual baseline between virtual viewpoints to recover the environment radiance field." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.822, + 0.22, + 0.837 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.848, + 0.27, + 0.864 + ], + "angle": 0, + "content": "2.1. Modeling reflections" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Catadioptric imaging systems incorporate reflections on curved reflective mirrors to expand the field of view of con" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.227 + ], + "angle": 0, + "content": "ventional cameras [2, 26], to increase the baseline of light field cameras [11, 40] and to perform novel view synthesis from a single capture [44]. These works assume the geometry of the reflecting surface is known or calibrated while we create a catadioptric imaging system from everyday glossy objects of unknown geometry. Grossberg et al. [15] propose a generalized model for light transport through imaging systems, including catadioptric systems. Our work focuses on light transport reflecting off a general object in the scene." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.229, + 0.892, + 0.41 + ], + "angle": 0, + "content": "Light field imaging is shown to be effective for reflection removal [21, 27], reconstructing specular surfaces [18], intrinsic decomposition [1], and neural rendering [35]. These works typically consider planar reflections or require training on synthetic datasets, while our approach models reflections on complex geometry and is unsupervised. Prior works have also utilized additional light properties such as polarization [9, 10, 20] and time-of-flight [17, 33] for the separation of the reflected component and specular surface reconstruction. While the input of our approach is RGB images, there is scope for improving reconstruction quality by supplementing the algorithm with these additional cues." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.411, + 0.892, + 0.653 + ], + "angle": 0, + "content": "Recent progress in neural radiance fields has enabled impressive novel view rendering and geometry reconstruction from multi-view images [24]. MipNeRF [5] demonstrates better novel view synthesis by modeling outgoing rays as cones to enable anti-aliasing. RefNeRF [41] proposes Integrated Directional Embeddings for improved novel view synthesis of reflections. NeRFRen [16] separates diffuse and specular radiance by using separate neural networks. Neural Catacaustics [19] propose a neural warping method to model reflections by learning the caustics of the surface. While these works focus on novel-view synthesis of the scene from the primary camera, we perform view synthesis that is beyond the line-of-sight of the primary camera, i.e., rendering views only visible to the objects present in the scene, while jointly estimating object geometry and separating diffuse and specular radiance." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.666, + 0.725, + 0.68 + ], + "angle": 0, + "content": "2.2. Environment Estimation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Recovering underlying scene properties from multiple images is inherently ill-posed [32], but can be regularized using the natural statistics of scene properties as a prior [4, 34]. Recent works exploit this prior through deep neural networks and demonstrate inverse rendering of indoor scenes from a single image [12, 22, 43, 48]. However, these techniques typically recover only coarse representations of lighting and cannot reconstruct fine details of the environment. Lombardi et al. [23] recover environment and reflectance, assuming the scene is composed of known geometry and uniform material. Georgolis et al. [13] recover the environment map behind the camera from a single image of a glossy object, assuming the object is composed of textureless materials and using ground truth seg" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20775" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.08, + 0.089, + 0.472, + 0.238 + ], + "angle": 0, + "content": "
ApproachInputScene GeometryAlgorithm TypeEnvironment Dimension
Neural IllumSingle RGBNormalsSupervised2D
LighthouseStereo RGB pairMulti-plane ImageSupervised5D
NeRFactorMulti-view RGBVolumetric AlbedoSemi-supervised2D
RefNeRFMulti-view RGBVolumetric AlbedoSelf-supervised2D
PANDORAMulti-view RGB+Pol.Neural SDFSelf-supervised2D
ORCa (Ours)Multi-view RGBNeural SDFSelf-supervised5D
" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.249, + 0.47, + 0.372 + ], + "angle": 0, + "content": "Figure 4. Comparison of environment estimation approaches. Approaches such Neural Illum [36], Lighthouse [37], NeRFactor [47] train on datasets of natural illumination maps to regularize ill-posedness of environment estimation. PANDORA [10] and RefNeRF [41] exploit multi-view reflections on objects but assume the surrounding environment is infinitely far away and is modeled with a flat 2D map. From multi-view reflections, ORCa converts the object surface into a virtual sensor and extracts the 5D radiance field of the environment." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.399, + 0.473, + 0.613 + ], + "angle": 0, + "content": "mentation masks. Song et al. [36] estimate plausible environment maps by mapping reflections in the image and inpainting unmapped regions. Srinivasan et al. [37] capture stereo image pairs and estimate plausible spatially-coherent environment maps. NeRD [7], NeRFactor [47] and NeuralPIL [8] employ data-driven priors for lighting and BRDF in a NeRF-based approach for radiance decomposition. Park et al. [31] use RGB-D videos to estimate environment map. Swedish et al. [39] recover high-frequency illumination map from the shadows of an object with known geometry. PhySG [46] and Munkberg et al. [25] perform inverse rendering from multi-view images by modeling the surface as signed distance functions. PANDORA [10] performs radiance decomposition from polarized RGB images." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.639, + 0.47, + 0.673 + ], + "angle": 0, + "content": "3. Learning environment radiance fields from multi-view reflections" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.682, + 0.188, + 0.696 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Reflections on glossy objects offer a glimpse into the surrounding environment beyond the camera's field of view. From multi-view images of a glossy object with unknown geometry and albedo, we aim to recover the 5D radiance field of the surrounding environment. The mapping from images captured by the observer to the surrounding environment depends on the glossy object's surface properties, in particular, the surface normals and curvature. We first cast a cone from the observer camera's center-of-projection through each pixel viewing the scene. When the cone intersects the object's surface, it reflects, causing the cone to be transformed (Fig. 5). The transformed cone, referred to as a virtual cone, samples the environment and is primar-" + }, + { + "type": "image", + "bbox": [ + 0.533, + 0.089, + 0.865, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.289, + 0.895, + 0.389 + ], + "angle": 0, + "content": "Figure 5. Virtual Sensor. We image the world through the object by modeling each pixel's specular radiance as a projection of the 5D radiance field of the environment onto the object's surface. We capture the radiance field by treating the surface area on the object that the pixel views, \\(\\mathbf{dS}_t\\), as a single-pixel virtual camera with its center-of-projection at \\(\\mathbf{v}_o\\). We cast virtual cones through the virtual sensor to capture the 5D radiance field of the environment." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.414, + 0.892, + 0.534 + ], + "angle": 0, + "content": "ily responsible for the specular radiance observed on the glossy object. Our key insight is that the reflections captured by the observer's camera can be modeled as a projection of the environment radiance field onto the object's surface. By modeling the reflected rays as a cone and computing the parameters of the cone, we can more accurately estimate the projected environment radiance field onto the object surface, as shown in Fig. 10." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.535, + 0.893, + 0.686 + ], + "angle": 0, + "content": "ORCa is composed of three steps: modeling the object's geometry as a neural implicit surface (Sec. 3.2), converting the object's surface into a virtual sensor (Sec. 3.3), and modeling the environment radiance field as a projection along these virtual cones (Sec. 3.4). The learned environment radiance field can then be queried on novel viewpoints to show occluded areas in the scene. Fig. 2 depicts our output for each component on a scene rendered with a complex glossy object and 3D environment. Fig. 6 shows our system architecture. Next, we describe each step in detail." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.695, + 0.8, + 0.712 + ], + "angle": 0, + "content": "3.2. Learning Neural implicit Surfaces" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.718, + 0.895, + 0.87 + ], + "angle": 0, + "content": "Neural Signed Distance Function We model the object geometry as a neural signed distance function (SDF). \\( f: \\mathbb{R}^3 \\to \\mathbb{R} \\). SDFs provide a helpful inductive bias for learning smooth surface geometry [30, 42, 45] that assists downstream tasks in our pipeline. Moreover, the surface properties crucial for our framework, surface normals and curvature, can be conveniently computed from SDFs in a differentiable manner. Consider the 3D spatial coordinates, \\( \\mathbf{x} \\), in the scene. The glossy object surface, \\( S \\) is then represented by the zero-level set of the SDF" + }, + { + "type": "equation", + "bbox": [ + 0.606, + 0.88, + 0.892, + 0.898 + ], + "angle": 0, + "content": "\\[\n\\mathcal {S} = \\left\\{f _ {\\mathcal {S}} (\\mathbf {x}) = 0 | \\mathbf {x} \\in \\mathbb {R} ^ {3} \\right\\} \\tag {1}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20776" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.09, + 0.462, + 0.164 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.119, + 0.174, + 0.427, + 0.188 + ], + "angle": 0, + "content": "Figure 6. Overview of our proposed architecture." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.213, + 0.47, + 0.241 + ], + "angle": 0, + "content": "Similar to Yariv et al. [45], we model the SDF \\( f_{S} \\) as a coordinate-based MLP." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.244, + 0.469, + 0.273 + ], + "angle": 0, + "content": "Surface Normals Gradients of the SDF at the zero level set point \\(S\\) towards the surface normals \\(S\\)," + }, + { + "type": "equation", + "bbox": [ + 0.174, + 0.282, + 0.469, + 0.315 + ], + "angle": 0, + "content": "\\[\n\\mathbf {n} (\\mathbf {x}) = \\frac {\\nabla_ {\\mathbf {x}} f _ {\\mathcal {S}} (\\mathbf {x})}{\\| \\nabla_ {\\mathbf {x}} f _ {\\mathcal {S}} (\\mathbf {x}) \\|} \\quad \\mathbf {x} \\in \\mathcal {S} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.321, + 0.469, + 0.397 + ], + "angle": 0, + "content": "Surface Curvature We employ differential geometry techniques developed by Novello et al. [29] to estimate curvature for neural implicit surfaces. In particular, we estimate the mean curvature \\( K(\\mathbf{x}) \\) for the implicit surface from the divergence, \\( \\nabla \\) of the surface normals" + }, + { + "type": "equation", + "bbox": [ + 0.209, + 0.405, + 0.469, + 0.435 + ], + "angle": 0, + "content": "\\[\nK (\\mathbf {x}) = \\frac {\\nabla \\cdot \\mathbf {n} (\\mathbf {x})}{2} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.442, + 0.469, + 0.517 + ], + "angle": 0, + "content": "Mean curvature approximates the surface with an osculating sphere. Our approach also works for more generalized notions of curvature through the shape operator, at the cost of higher computational complexity. We refer our readers to the supplement for the general case." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.518, + 0.469, + 0.622 + ], + "angle": 0, + "content": "Diffuse Radiance We separate the captured radiance at the observer camera with diffuse radiance, which depends on the glossy object's albedo, and specular radiance which depends on the environment radiance. The diffuse radiance does not have any view dependence and only depends on surface point \\(\\mathbf{x}\\). We denote the diffuse radiance as \\(f_{d}\\) and model it using a coordinate-based MLP (Fig. 6)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.623, + 0.469, + 0.683 + ], + "angle": 0, + "content": "Volume Rendering As proposed in [45], we perform volumetric rendering on the SDF. We define the volume density \\(\\sigma(\\mathbf{x})\\) as the cumulative distribution function (CDF), denoted as \\(\\Psi(s)\\), applied to \\(f_{S}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.212, + 0.693, + 0.469, + 0.71 + ], + "angle": 0, + "content": "\\[\n\\sigma (\\mathbf {x}) = \\alpha \\Psi_ {\\beta} (f _ {\\mathcal {S}}) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.718, + 0.469, + 0.807 + ], + "angle": 0, + "content": "In contrast to [45], however, we only aim to recover the diffuse radiance of the object along a particular ray. We define a function \\( f_{d} \\) that estimates the diffuse radiance at each point, \\( \\mathbf{x} \\), along the ray. To get the final diffuse radiance along a given primary ray, \\( \\mathbf{r}_p(t) \\), we perform volumetric rendering:" + }, + { + "type": "equation", + "bbox": [ + 0.147, + 0.816, + 0.469, + 0.849 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {c}} _ {d} (\\mathbf {r}) = \\int_ {0} ^ {\\infty} f _ {d} (\\mathbf {r} (t), f _ {S} ^ {k} (\\mathbf {r} (t)) \\tau (t) d t \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Note that there is no view dependence in Eq. 5 and intermediate features, \\( f_{\\mathcal{S}}^{k} \\), are used as input. \\( \\tau(t) \\) is the accumulated transmittance along the ray." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.089, + 0.892, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.393, + 0.892, + 0.49 + ], + "angle": 0, + "content": "Figure 7. Qualitative comparisons of diffuse-specular separation and geometry estimation on rendered dataset. The environment contains nearby objects with complex occlusions when seen through reflections on the glossy object. RefNeRF fails to perform accurate diffuse-specular separation and PANDORA blurs the nearby objects in the specular map. ORCA can model the complex specular reflections through environment radiance field." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.513, + 0.795, + 0.529 + ], + "angle": 0, + "content": "3.3. Objects Surface as Virtual Sensor" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.537, + 0.892, + 0.717 + ], + "angle": 0, + "content": "Each pixel, \\(\\mathbf{p}\\), with a finite surface area, \\(\\mathbf{dA}_p\\), on the real-camera sensor views the surface of the object through a frustum originating at that pixel. The object then samples the environment radiance field through this finite surface converting the finite surface into a virtual pixel with surface area, \\(\\mathbf{dS}\\). Through this model, we can interpret the object surface as a virtual sensor consisting of many virtual pixels that sample radiance from the environment field based on the geometry of the object and observer viewing direction. We now formulate a virtual pixel based on real camera pose and implicit surface geometry. Please refer to Fig. 5 for a visualization of the virtual sensor." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.718, + 0.892, + 0.779 + ], + "angle": 0, + "content": "Consider a real camera origin as \\(\\mathbf{o}\\) and a pixel on the real sensor \\(\\mathbf{p}_{i,j}\\) that corresponds to ray direction \\(\\mathbf{d}\\). The primary ray for pixel \\(\\mathbf{p}_{i,j}\\) is parameterized with ray length \\(t\\) as \\(\\mathbf{r_p}(t) = \\mathbf{o} + t\\mathbf{d}\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.779, + 0.892, + 0.854 + ], + "angle": 0, + "content": "Casting Real Cones We can approximate the outgoing conical frustum from pixel \\(\\mathbf{p}_{i,j}\\) as a cone originating at \\(\\mathbf{o}\\) with axis-of-direction \\(\\mathbf{d}\\) and radius \\(\\dot{r}\\), equivalent to half the distance of the pixel in the x and y directions. We represent the real cone as a parametric volume," + }, + { + "type": "equation", + "bbox": [ + 0.518, + 0.862, + 0.891, + 0.878 + ], + "angle": 0, + "content": "\\[\n\\mathbf {r} _ {\\text {c o n e}} (\\dot {r}, s, \\theta) = \\dot {r} s \\cos (\\theta) \\hat {\\mathbf {e}} _ {u} + \\dot {r} s \\sin (\\theta) \\hat {\\mathbf {e}} _ {v} + \\dot {r} s \\mathbf {d}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.886, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where \\(\\hat{\\mathbf{e}}_u\\) and \\(\\hat{\\mathbf{e}}_v\\) are basis vectors in the plane perpendicular" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20777" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.298, + 0.107 + ], + "angle": 0, + "content": "to \\(\\mathbf{d}\\), \\(\\theta \\in [0, \\pi]\\) and \\(s \\in [0, t_{\\max}]\\)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.107, + 0.47, + 0.259 + ], + "angle": 0, + "content": "Virtual Pixel. Virtual pixels are characterized by the intersection of the real cone with the object surface. In Sec. 3.2, we model local surface properties using mean curvature which enables efficient analytical computations for the virtual pixel parameters even though our approach works with general shape operators. For a sampled point \\( t_i \\) along the ray, we have the surface normals \\( \\mathbf{n}(t_i) \\) from Eq. 2 and estimated mean curvature \\( K(t_i) \\) from Eq. 3. The local object surface at \\( t_i \\), can be approximated with an osculating sphere, \\( \\mathcal{O}(t_i) \\), centered at \\( \\hat{\\mathbf{o}}_{\\mathbf{S}}(t_i) \\) with radius, \\( R(t_i) \\) as" + }, + { + "type": "equation", + "bbox": [ + 0.124, + 0.28, + 0.424, + 0.313 + ], + "angle": 0, + "content": "\\[\nR (t _ {i}) = \\frac {2}{K _ {t _ {i}}} \\hat {\\mathbf {o}} _ {\\mathbf {S}} (t _ {i}) = \\mathbf {r} _ {p} (t _ {i}) + R (t _ {i}) \\cdot \\hat {\\mathbf {n}} (t _ {i})\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.322, + 0.469, + 0.351 + ], + "angle": 0, + "content": "For concave surfaces \\( K_{t_i} < 0 \\). So, \\( \\hat{\\mathbf{o}}_{\\mathbf{S}} \\) will lie outside the object. For \\( K_{t_i} > 0 \\), \\( \\hat{\\mathbf{o}}_{\\mathbf{S}} \\) will lie inside the object." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.352, + 0.469, + 0.518 + ], + "angle": 0, + "content": "The edges of the virtual pixel for \\(\\mathbf{r_p}(t_i)\\) would lie at the intersection of the osculating sphere \\(\\mathcal{O}(t_i)\\) and the primary cone given by \\(\\mathbf{r}_{\\mathrm{cone}}\\). Computing exact cone-sphere intersections are computationally expensive so we approximate the cone-sphere intersection using rays bound cone-sphere intersectional surface dS. We consider four rays that bound the cone and sample them at \\(\\theta_{j}\\in \\{0,\\pi /2,\\pi ,3\\pi /2\\}\\) with Eq. 6. We perform intersections of the corresponding bounding rays with the osculating sphere \\(\\mathcal{O}(t_i)\\) to get corners of the virtual pixel \\(\\mathbf{ds}_j\\). These ray sphere intersections can be computed analytically in an efficient manner." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.518, + 0.469, + 0.579 + ], + "angle": 0, + "content": "Virtual Cone Origin. From the virtual pixel surface area, we can now compute the virtual cone that samples the environment. We first compute normal vectors at virtual pixel corners \\(\\mathbf{ds_j}\\) from the center of osculating sphere \\(\\hat{\\mathbf{o}}_{\\mathrm{S}}\\)" + }, + { + "type": "equation", + "bbox": [ + 0.209, + 0.59, + 0.469, + 0.623 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {n}} _ {\\mathbf {j}} = \\frac {\\mathbf {d} \\mathbf {s} _ {\\mathbf {j}} - \\hat {\\mathbf {o}} _ {\\mathbf {S}}}{\\left| \\left| \\mathbf {d} \\mathbf {s} _ {\\mathbf {j}} - \\hat {\\mathbf {o}} _ {\\mathbf {S}} \\right| \\right|} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.632, + 0.469, + 0.695 + ], + "angle": 0, + "content": "At each virtual pixel corner, we compute the reflected ray directions, \\(\\omega_{j}^{r}\\), by computing the dot product between the incoming ray directions, \\(\\omega_{k}^{i}\\), and the normals, \\(\\hat{\\mathbf{n}}_k\\), where \\(\\omega_0^r\\) is the primary ray's reflected vector." + }, + { + "type": "equation", + "bbox": [ + 0.195, + 0.72, + 0.469, + 0.737 + ], + "angle": 0, + "content": "\\[\n\\omega_ {0} ^ {r} = \\mathbf {d} - \\left(\\mathbf {d} \\cdot \\hat {\\mathbf {n}} \\left(t _ {i}\\right)\\right) \\hat {\\mathbf {n}} \\left(t _ {i}\\right) \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.172, + 0.741, + 0.469, + 0.758 + ], + "angle": 0, + "content": "\\[\n\\omega_ {j} ^ {r} = \\mathbf {d} _ {j} - \\left(\\mathbf {d} _ {j} \\cdot \\hat {\\mathbf {n}} _ {j} (k)\\right) \\hat {\\mathbf {n}} _ {j} (k) \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.769, + 0.469, + 0.86 + ], + "angle": 0, + "content": "\\(\\mathbf{d}_j\\) are the incident directions to the virtual pixel corners \\(\\mathbf{ds}_j\\). The virtual cone origin is the intersection of these reflected rays at the pixel corners and pixel center. However, these rays might not intersect at a single point so we approximate a virtual origin to be the point that minimizes the sum of distances to the reflected rays \\(\\omega_j\\)." + }, + { + "type": "equation", + "bbox": [ + 0.152, + 0.871, + 0.469, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\mathbf {v} _ {\\mathbf {o}} = \\operatorname {a r g m i n} _ {\\mathbf {v}} \\sum_ {j} | (\\mathbf {v} - \\mathbf {d s} _ {\\mathbf {j}}) \\times \\omega_ {j} ^ {r} | \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "We pose this as a linear least squares problem and estimate the virtual cone origin efficiently through pseudo-inverse." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.123, + 0.892, + 0.182 + ], + "angle": 0, + "content": "Virtual Cones Direction. The reflected ray at the center of the virtual pixel reflects the object surface along the direction \\(\\omega_0^r\\) from Eq. 8. We consider this as the direction-of-axis of the virtual cone." + }, + { + "type": "equation", + "bbox": [ + 0.665, + 0.197, + 0.892, + 0.214 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {v}} _ {d} = \\omega_ {0} ^ {r} \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.228, + 0.892, + 0.319 + ], + "angle": 0, + "content": "Virtual Cone Radius. We compute the radius of the cone by treating the reflection vectors of the bounding rays as the neighboring \"pixel\" directions. Similar to [5], we can compute the distance between \\(\\{\\omega_{k_{\\theta}}^{r}\\}_{\\theta = 0}^{2\\pi}\\) and the primary reflected ray \\(\\omega_0^r\\) in the \\((x,y)\\) components (omitted below for clarity)." + }, + { + "type": "equation", + "bbox": [ + 0.614, + 0.348, + 0.892, + 0.366 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {v}} _ {\\dot {r}} = \\left\\| \\left\\{\\omega_ {k _ {\\theta}} ^ {r} \\right\\} _ {\\theta = 0} ^ {2 \\pi} - \\omega_ {0} ^ {r} \\right\\| \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.38, + 0.892, + 0.439 + ], + "angle": 0, + "content": "Finally, for each sampled point \\( t_i \\), we can characterize our single-pixel virtual sensor located at the object surface \\( \\mathbf{dS} \\) as a virtual cone with \\( \\hat{\\mathbf{v}}_o \\) as its apex, \\( \\hat{\\mathbf{v}}_d \\) as axis-direction, \\( \\hat{\\mathbf{v}}_{\\dot{r}} \\) as the radius." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.441, + 0.892, + 0.729 + ], + "angle": 0, + "content": "Connections to caustics. Our work takes inspiration from catadioptric imaging systems. To convert objects into cameras, we compute the surface and find a corresponding center-of-projection for this surface-as-sensor. However, unlike conventional perspective cameras, objects don't have a fixed center-of-projection, other than in a few special configurations [3], but a locus of viewpoints that vary with object geometry and viewing direction. These viewpoints lie on the \"caustic surface\" of the object. While typical works in catadioptric imaging analytically compute the caustic surface by assuming known geometry [14, 38], or making assumptions about placement of the observer [40], our formulation approximates the caustic surface of unknown geometry through the intersection of reflected rays on virtual pixels. We empirically show in the supplement that as the surface area of the virtual pixel goes to \\(0\\), \\(d\\mathbf{S} \\rightarrow 0\\), our method estimates the true caustic of the object without assuming geometry. Our method also has applications in estimating the caustic surface of the unknown geometry." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.74, + 0.764, + 0.755 + ], + "angle": 0, + "content": "3.4. Environment Radiance Fields" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.764, + 0.892, + 0.855 + ], + "angle": 0, + "content": "Our goal is to capture a 5D environment radiance field of the scene by imaging the world through these single-pixel virtual sensors located at the object's surface. We use our formulation of virtual cones to recover 5D environment radiance fields. We define an environment radiance field as \\( f_{\\mathcal{E}}: (\\hat{\\mathbf{v}}_o, \\hat{\\mathbf{v}}_d) \\to (\\sigma^{Env}, c_s) \\)," + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where \\( f_{\\varepsilon} \\) outputs opacity and radiance along sampled virtual cones. We note that this view-dependent radiance is equivalent to the specular radiance at point \\( t_i \\) sampled" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20778" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.083, + 0.088, + 0.468, + 0.17 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.18, + 0.47, + 0.222 + ], + "angle": 0, + "content": "Figure 8. Evaluation of environment depth from reflections. For the scene in Fig. 3, ORCa can faithfully estimate the depth of the environment, especially for nearby surroundings." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.238, + 0.466, + 0.303 + ], + "angle": 0, + "content": "
ApproachDiffuse RadianceSpecular RadianceMixed RadianceNormals
PSNR ↑ (dB)SSIM ↑PSNR ↑ (dB)SSIM ↑PSNR ↑ (dB)SSIM ↑MAE ↓ (°)
Ref-NeRF18.800.730416.990.663320.190.789043.690
PANDORA18.250.726016.250.648318.900.72847.606
ORCA19.840.789320.740.753522.000.79472.339
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.313, + 0.47, + 0.397 + ], + "angle": 0, + "content": "Table 1. Average evaluation metrics on rendered scenes. We compare ORCa to other neural rendering techniques that model reflections, including Ref-NeRF and PANDORA, on six simulated scenes. ORCa provides consistent improvements in geometry estimation, diffuse-specular separation and novel view synthesis. Please refer to the supplement for additional metrics." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.424, + 0.47, + 0.456 + ], + "angle": 0, + "content": "along the primary-camera ray \\(\\mathbf{r}_{\\mathbf{p}}(t)\\). We can render the final specular radiance at pixel \\(\\mathbf{p}_{i,j}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.467, + 0.378, + 0.5 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {c}} _ {s} (\\mathbf {r}) = \\int_ {0} ^ {\\infty} f _ {\\mathcal {E}} (\\hat {\\mathbf {v}} _ {o}, \\hat {\\mathbf {v}} _ {d}) \\tau (t) d t\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.23, + 0.504, + 0.317, + 0.518 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {c}} = \\hat {\\mathbf {c _ {d}}} + \\hat {\\mathbf {c _ {s}}}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.534, + 0.471, + 0.685 + ], + "angle": 0, + "content": "Intuitively, \\( f_{\\mathcal{E}} \\) learns the 5D radiance field by sampling single-pixel virtual sensors from the object surface and must learn geometry and environment radiance that is consistent with the multi-view reflections. Moreover, we can query \\( f_{\\mathcal{E}} \\) to render novel viewpoints and associated depths that are beyond the field-of-view of the real camera. We volume render each virtual cone by dividing them into conical frustums using Integrated-Positional Encoding as proposed in MipNeRF [5]. Our formulation of virtual cones works well with Mip-Nerf's rays-as-cones method." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.701, + 0.21, + 0.717 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.726, + 0.295, + 0.743 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.903 + ], + "angle": 0, + "content": "As in PANDORA, we parameterize \\( f_{S} \\) with an 8-layer MLP to estimate the surface, and, as in MipNeRF, \\( f_{d} \\) with 4-layer MLP with input geometric features of size 512 from \\( f_{S} \\). We follow the SDF-to-opacity conversion and the iterative sampling of the ray proposed in [45]. To aid the network to learn geometry quickly, we also train \\( f_{S} \\) with a mask-net as proposed in [10]. We use five losses in our architecture: photometric loss, mask loss [10], normal loss [41], eikonal loss [45], and distortion loss [6]. Additional training details are discussed in the supplement." + }, + { + "type": "table", + "bbox": [ + 0.52, + 0.089, + 0.694, + 0.173 + ], + "angle": 0, + "content": "
Scene: Ball-cup in hallway
ApproachMixedRadiance
PSNRSSIM
↑(dB)↑
Ref-NeRF32.750.9617
PANDORA28.830.9758
ORCA30.860.9799
" + }, + { + "type": "table", + "bbox": [ + 0.696, + 0.089, + 0.872, + 0.173 + ], + "angle": 0, + "content": "
Scene: Owl in hallway
ApproachMixed Radiance
PSNRSSIM
↑ (dB)↑
Ref-NeRF26.650.8890
PANDORA27.240.9343
ORCA26.840.9299
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.184, + 0.892, + 0.225 + ], + "angle": 0, + "content": "Table 2. Quantitative metrics for real-world, captured scenes. ORCa demonstrates comparable performance in novel view synthesis on scenes from the PANDORA real dataset." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.252, + 0.602, + 0.267 + ], + "angle": 0, + "content": "4.2. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.276, + 0.892, + 0.412 + ], + "angle": 0, + "content": "We conduct experiments on both simulated and real-world datasets. Simulated datasets are rendered in Mitsu2 [28]. Simulated datasets contain a range of increasingly complex object geometries (elephant, Pokeball, and orca) and scenes (living room and Pokemon). We train with 200 views for simulated datasets. We also show results for a real-world dataset [10] capturing a glossy cup with a black vase sitting atop it using 35 views. All datasets will be publicly released upon publication." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.422, + 0.878, + 0.437 + ], + "angle": 0, + "content": "4.3. Advantages of Environment Radiance Fields" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.445, + 0.892, + 0.688 + ], + "angle": 0, + "content": "Other neural rendering techniques that handle reflections, Ref-NeRF and PANDORA, estimate the environment as a 2D map, while ORCa recovers a 5D environment radiance field. Fig. 3 shows the advantages of recovering a 5D environment radiance field. Close-by surrounding objects often cause occlusions that cannot be modeled by 2D environment maps. By estimating the radiance field, we can image behind occluders through sampling novel viewpoints such as the translated viewpoints shown in Fig. 3. Moreover, we can also show the depth to the surroundings from these virtual viewpoints. In Fig. 8, we quantitatively evaluate the estimated depth of the surroundings for the synthesized virtual views in Fig. 3 and show that ORCa provides reliable depth estimates especially for nearby objects. We provide additional examples of depth estimation and beyond field-of-view novel-view synthesis in the supplement." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.697, + 0.753, + 0.713 + ], + "angle": 0, + "content": "4.4. Comparisons with Baselines" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We compare ORCa with Ref-NeRF and PANDORA on synthetic and real datasets. As shown in Fig. 7 and 9, ORCa estimates more accurate surface normals than other methods. While the total radiance predicted by Ref-NeRF and PANDORA are visually similar, the surface normals are less smooth than ORCa. We also observe that ORCa is able to achieve better diffuse and specular radiance separation than PANDORA, which is evident in the Pokeball surface normals Fig. 7. In these examples, PANDORA recovers blurry specular radiance. We see that ORCa's predicted depth is highly interpretable and matches the underlying geometry of the environment, as shown in Fig. 3. Even on cylindri" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20779" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.322, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.327, + 0.089, + 0.471, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.371, + 0.472, + 0.496 + ], + "angle": 0, + "content": "Figure 9. Comparisons on real dataset. Using a real-world dataset with only 35 views, ORCa can model the sharp specularities on the ball arriving from regions of the nearby scene, such as the table, and far away scene regions, such as the hallway, to learn an environment radiance field. We query the radiance field for the depth of the far hallway (blue) and the nearby objects, such as the table (red). We also render novel viewpoints that are beyond the field of view of the observer camera and show that ORCa interpolates well between those views." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.522, + 0.47, + 0.763 + ], + "angle": 0, + "content": "cal real-world datasets, such as the black vase in Fig. 9, the nearby hallway is visible in both the virtual view and depth, despite never being in the field of view of the primary camera. Unlike Ref-NeRF, our primary objective is not to perform novel-view synthesis, but instead to capture the environment radiance field from the object surface. As shown in Table 1, ORCa improves over Ref-NeRF and PANDORA in estimating diffuse radiance, specular radiance, mixed radiance, and normals. Here we report the image quality metrics, PSNR and SSIM, and surface normal metric, MAE, averaged over six synthetic scenes with varying environments, camera view distribution, and object complexity. Please refer to the supplement for individual metrics on each scene. ORCa also shows competitive performance for novel-view synthesis metrics on real-world scenes from PANDORA dataset, as shown in Table 2." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.772, + 0.399, + 0.789 + ], + "angle": 0, + "content": "4.5. Impact of Virtual Cone Computation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.471, + 0.903 + ], + "angle": 0, + "content": "We base our method on a physically accurate formulation by modeling ray-cone intersections and using the surface as a virtual sensor, as described in Sec. 3.3. We demonstrate the importance of this step by setting up two ablation experiments. In the first experiment, which we term Naive Virtual Cone, we place the origin of the cone at the object surface instead of computing the virtual cone origin based" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.091, + 0.895, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.234, + 0.895, + 0.318 + ], + "angle": 0, + "content": "Figure 10. Ablations on virtual cone computation Accurate estimation of the virtual cones (Sec. 3.3) is crucial for ORCa. If the virtual cone origin is assumed to be at the object surface (left column) or if the surface is assumed to locally have no curvature (right column), the surface normal and specular radiance outputs suffer from artifacts (red boxes)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.354, + 0.895, + 0.445 + ], + "angle": 0, + "content": "on the curvature. In the second experiment, which we term No Curvature, we assume that locally the surface is like a flat mirror and has no curvature. In Fig. 10, we show that for both of these ablation experiments, we see worse performance as demonstrated by the artifacts in estimated surface normals and specular components (shown as red boxes)." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.473, + 0.619, + 0.488 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.503, + 0.895, + 0.761 + ], + "angle": 0, + "content": "We present a method to convert glossy objects with unknown geometry and texture into radiance-field cameras that capture the environment's radiance field around them. Our method recovers object geometry and diffuse radiance, in addition to capturing the depth and radiance of the object's surroundings from its perspective. Our modeling of the environment as a radiance field is effective in recovering close-by objects (Fig. 7) and is occlusion aware (Fig. 3). From the recovered environment radiance field, we can perform beyond field-of-view novel-view synthesis. Our work can unleash applications in virtual object insertion and 3D perception, e.g. inferring information beyond the line-of-sight of the camera using predicted virtual views and depth. Our formulation goes beyond the conventional direct-line-of-sight radiance fields and can enable further areas of research to extract more information from multi-view images directly from the environment and the objects present in it." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.788, + 0.66, + 0.805 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.817, + 0.894, + 0.9 + ], + "angle": 0, + "content": "This research was supported by the SMART Contract IARPA Grant 2021-20111000004. We would also like to thank Systems Technology Research (STR). AD and AV are supported by NSF Expeditions-Award IIS-1730574 and NSF CAREER-Award IIS-1652633." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20780" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.185 + ], + "angle": 0, + "content": "[1] Anna Alperovich, Ole Johannsen, and Bastian Goldluecke. Intrinsic light field decomposition and disparity estimation with deep encoder-decoder network. In 2018 26th European Signal Processing Conference (EUSIPCO), pages 2165-2169. IEEE, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.187, + 0.472, + 0.228 + ], + "angle": 0, + "content": "[2] S. Baker and S.K. Nayar. A theory of catadioptric image formation. In Sixth International Conference on Computer Vision (IEEE Cat. No.98CH36271), pages 35-42, 1998. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.231, + 0.471, + 0.273 + ], + "angle": 0, + "content": "[3] Simon Baker and Shree K. Nayar. A theory of single-viewpoint catadioptric image formation. International Journal of Computer Vision, 35:175-196, 2004. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.274, + 0.47, + 0.328 + ], + "angle": 0, + "content": "[4] Jonathan T Barron and Jitendra Malik. Shape, illumination, and reflectance from shading. IEEE transactions on pattern analysis and machine intelligence, 37(8):1670-1687, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.331, + 0.47, + 0.387 + ], + "angle": 0, + "content": "[5] Jonathan T. Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P. Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. ICCV, 2021. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.389, + 0.471, + 0.459 + ], + "angle": 0, + "content": "[6] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5470–5479, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.46, + 0.471, + 0.529 + ], + "angle": 0, + "content": "[7] Mark Boss, Raphael Braun, Varun Jampani, Jonathan T Barron, Ce Liu, and Hendrik Lensch. Nerd: Neural reflectance decomposition from image collections. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12684-12694, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.531, + 0.47, + 0.6 + ], + "angle": 0, + "content": "[8] Mark Boss, Varun Jampani, Raphael Braun, Ce Liu, Jonathan Barron, and Hendrik Lensch. Neural-pil: Neural pre-integrated lighting for reflectance decomposition. Advances in Neural Information Processing Systems, 34:10691-10704, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.602, + 0.47, + 0.657 + ], + "angle": 0, + "content": "[9] Akshit Dave, Yannick Hold-Geoffroy, Milos Hasan, Kalyan Sunkavalli, and Ashok Veeraraghavan. Snapshot polarimetric diffuse-specular separation. Optics Express, 30(19):34239-34255, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.66, + 0.469, + 0.702 + ], + "angle": 0, + "content": "[10] Akshit Dave, Yongyi Zhao, and Ashok Veeraraghavan. Pandora: Polarization-aided neural decomposition of radiance. arXiv preprint arXiv:2203.13458, 2022. 3, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.703, + 0.469, + 0.758 + ], + "angle": 0, + "content": "[11] Michael De Zeeuw and Aswin C Sankaranarayanan. Wide-baseline light fields using ellipsoidal mirrors. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.469, + 0.829 + ], + "angle": 0, + "content": "[12] Mathieu Garon, Kalyan Sunkavalli, Sunil Hadap, Nathan Carr, and Jean-François Lalonde. Fast spatially-varying indoor lighting estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6908-6917, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[13] Stamatios Georgoulis, Konstantinos Rematas, Tobias Ritschel, Mario Fritz, Tinne Tuytelaars, and Luc Van Gool. What is around the camera? In Proceedings of the IEEE International Conference on Computer Vision, pages 5170-5178, 2017. 3" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.116, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.161 + ], + "angle": 0, + "content": "[14] J. Gluckman and S.K. Nayar. Planar catadioptric stereo: geometry and calibration. In Proceedings. 1999 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (Cat. No PR00149), volume 1, pages 22-28 Vol. 1, 1999. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.894, + 0.205 + ], + "angle": 0, + "content": "[15] Michael D Grossberg and Shree K Nayar. The raxel imaging model and ray-based calibration. International Journal of Computer Vision, 61(2):119, 2005. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.207, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[16] Yuan-Chen Guo, Di Kang, Linchao Bao, Yu He, and Song-Hai Zhang. Nerfren: Neural radiance fields with reflections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18409-18418, June 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.278, + 0.892, + 0.332 + ], + "angle": 0, + "content": "[17] Connor Henley, Siddharth Somasundaram, Joseph Hollmann, and Ramesh Raskar. Detection and mapping of specular surfaces using multibounce lidar returns. Optics Express, 31(4):6370-6388, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.334, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[18] Jan Jachnik, Richard A Newcombe, and Andrew J Davison. Real-time surface light-field capture for augmentation of planar specular surfaces. In 2012 IEEE International Symposium on Mixed and Augmented Reality (ISMAR), pages 91-97. IEEE, 2012. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.406, + 0.892, + 0.46 + ], + "angle": 0, + "content": "[19] Georgios Kopanas, Thomas Leimkuhler, Gilles Rainer, Clément Jambon, and George Drettakis. Neural point cata-caustics for novel-view synthesis of reflections. ACM Transactions on Graphics, 41(6):Article-201, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.462, + 0.892, + 0.532 + ], + "angle": 0, + "content": "[20] Rui Li, Simeng Qiu, Guangming Zang, and Wolfgang Heidrich. Reflection separation via multi-bounce polarization state tracing. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIII 16, pages 781-796. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.533, + 0.892, + 0.575 + ], + "angle": 0, + "content": "[21] Tingtian Li, Daniel PK Lun, Yuk-Hee Chan, et al. Robust reflection removal based on light field imaging. IEEE Transactions on Image Processing, 28(4):1798–1812, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.577, + 0.892, + 0.659 + ], + "angle": 0, + "content": "[22] Zhengqin Li, Mohammad Shafiei, Ravi Ramamoorthi, Kalyan Sunkavalli, and Manmohan Chandraker. Inverse rendering for complex indoor scenes: Shape, spatially-varying lighting and svbrdf from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2475-2484, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.661, + 0.892, + 0.703 + ], + "angle": 0, + "content": "[23] Stephen Lombardi and Ko Nishino. Reflectance and natural illumination from a single image. In European Conference on Computer Vision, pages 582-595. Springer, 2012. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.704, + 0.892, + 0.771 + ], + "angle": 0, + "content": "[24] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.775, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[25] Jacob Munkberg, Jon Hasselgren, Tianchang Shen, Jun Gao, Wenzheng Chen, Alex Evans, Thomas Müller, and Sanja Fidler. Extracting triangular 3d models, materials, and lighting from images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8280-8290, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.859, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[26] Shree K Nayar and Simon Baker. Catadioptric image formation. In Proceedings of the 1997 DARPA Image Understanding Workshop, pages 1431-1437, 1997. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20781" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[27] Yun Ni, Jie Chen, and Lap-Pui Chau. Reflection removal on single light field capture using focus manipulation. IEEE Transactions on Computational Imaging, 4(4):562-572, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.47, + 0.203 + ], + "angle": 0, + "content": "[28] Merlin Nimier-David, Delio Vicini, Tizian Zeltner, and Wenzel Jakob. Mitsuba 2: A retargetable forward and inverse renderer. ACM Transactions on Graphics (TOG), 38(6):1-17, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.205, + 0.469, + 0.247 + ], + "angle": 0, + "content": "[29] Tiago Novello, Guilherme Schardong, Luiz Schirmer, Vinicius da Silva, Helio Lopes, and Luiz Velho. Exploring differential geometry in neural implicits, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.248, + 0.469, + 0.315 + ], + "angle": 0, + "content": "[30] Michael Oechsle, Songyou Peng, and Andreas Geiger. Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5589-5599, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.317, + 0.47, + 0.372 + ], + "angle": 0, + "content": "[31] Jeong Joon Park, Aleksander Holynski, and Steven M Seitz. Seeing the world in a bag of chips. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1417-1427, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.373, + 0.469, + 0.427 + ], + "angle": 0, + "content": "[32] Ravi Ramamoorthi and Pat Hanrahan. A signal-processing framework for inverse rendering. In Proceedings of the 28th annual conference on Computer graphics and interactive techniques, pages 117-128, 2001. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.428, + 0.469, + 0.469 + ], + "angle": 0, + "content": "[33] Raskar Ramesh and James Davis. 5d time-light transport matrix: What can we reason about scene properties? Technical report, 2008. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.47, + 0.468, + 0.511 + ], + "angle": 0, + "content": "[34] Fabiano Romeiro and Todd Zickler. Blind reflectometry. In European conference on computer vision, pages 45-58. Springer, 2010. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.512, + 0.468, + 0.58 + ], + "angle": 0, + "content": "[35] Vincent Sitzmann, Semon Rezchikov, Bill Freeman, Josh Tenenbaum, and Fredo Durand. Light field networks: Neural scene representations with single-evaluation rendering. Advances in Neural Information Processing Systems, 34:19313-19325, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.582, + 0.468, + 0.636 + ], + "angle": 0, + "content": "[36] Shuran Song and Thomas Funkhouser. Neural illumination: Lighting prediction for indoor environments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6918-6926, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.637, + 0.468, + 0.718 + ], + "angle": 0, + "content": "[37] Pratul P Srinivasan, Ben Mildenhall, Matthew Tancik, Jonathan T Barron, Richard Tucker, and Noah Snavely. Lighthouse: Predicting lighting volumes for spatially-coherent illumination. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8080-8089, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.721, + 0.468, + 0.775 + ], + "angle": 0, + "content": "[38] R. Swaminathan, M.D. Grossberg, and S.K. Nayar. Cautistics of catadioptric cameras. In Proceedings Eighth IEEE International Conference on Computer Vision. ICCV 2001, volume 2, pages 2-9 vol.2, 2001. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.777, + 0.468, + 0.843 + ], + "angle": 0, + "content": "[39] Tristan Swedish, Connor Henley, and Ramesh Raskar. Objects as cameras: Estimating high-frequency illumination from shadows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2593-2602, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.846, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[40] Yuichi Taguchi, Amit Agrawal, Ashok Veeraraghavan, Srikumar Ramalingam, and Ramesh Raskar. Axial-cones: Modeling spherical catadioptric cameras for wide-angle light field rendering. ACM Trans. Graph., 29(6), dec 2010. 3, 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[41] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T. Barron, and Pratul P. Srinivasan. Ref-NeRF: Structured view-dependent appearance for neural radiance fields. CVPR, 2022. 3, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.149, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[42] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.207, + 0.892, + 0.274 + ], + "angle": 0, + "content": "[43] Zian Wang, Jonah Philion, Sanja Fidler, and Jan Kautz. Learning indoor inverse rendering with 3d spatially-varying lighting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12538-12547, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.277, + 0.892, + 0.346 + ], + "angle": 0, + "content": "[44] Ziyu Wang, Liao Wang, Fuqiang Zhao, Minye Wu, Lan Xu, and Jingyi Yu. Mirrornerf: One-shot neural portrait radiance field from multi-mirror catadioptric imaging. In 2021 IEEE International Conference on Computational Photography (ICCP), pages 1-12. IEEE, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.348, + 0.892, + 0.401 + ], + "angle": 0, + "content": "[45] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. In Thirty-Fifth Conference on Neural Information Processing Systems, 2021. 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.404, + 0.892, + 0.472 + ], + "angle": 0, + "content": "[46] Kai Zhang, Fujun Luan, Qianqian Wang, Kavita Bala, and Noah Snively. Physg: Inverse rendering with spherical gaussians for physics-based material editing and relighting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5453-5462, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.475, + 0.892, + 0.543 + ], + "angle": 0, + "content": "[47] Xiuming Zhang, Pratul P Srinivasan, Boyang Deng, Paul Debevec, William T Freeman, and Jonathan T Barron. Nerfactor: Neural factorization of shape and reflectance under an unknown illumination. ACM Transactions on Graphics (TOG), 40(6):1-18, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.545, + 0.892, + 0.614 + ], + "angle": 0, + "content": "[48] Rui Zhu, Zhengqin Li, Janarbek Matai, Fatih Porikli, and Manmohan Chandraker. Irisformer: Dense vision transformers for single-image inverse rendering in indoor scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2822-2831, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.614 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20782" + } + ] +] \ No newline at end of file diff --git a/2023/ORCa_ Glossy Objects As Radiance-Field Cameras/da15cd79-f9cc-4796-854b-9e2af61a06e8_origin.pdf b/2023/ORCa_ Glossy Objects As Radiance-Field Cameras/da15cd79-f9cc-4796-854b-9e2af61a06e8_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..204df6db931b04231a24d7a447b245318a3a8957 --- /dev/null +++ b/2023/ORCa_ Glossy Objects As Radiance-Field Cameras/da15cd79-f9cc-4796-854b-9e2af61a06e8_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f80d42acc7d794a29a11d75ef8b2c9231d7e7c9ae133dbab7af492b6bb28596b +size 9549980 diff --git a/2023/ORCa_ Glossy Objects As Radiance-Field Cameras/full.md b/2023/ORCa_ Glossy Objects As Radiance-Field Cameras/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7278dd6d398685712f50e0b5bc8513497040ab9c --- /dev/null +++ b/2023/ORCa_ Glossy Objects As Radiance-Field Cameras/full.md @@ -0,0 +1,380 @@ +# ORCa: Glossy Objects as Radiance-Field Cameras + +Kushagra Tiwary $^{1*}$ , Akshit Dave $^{2*}$ , Nikhil Behari $^{1}$ , Tzofi Klinghoffer $^{1}$ , Ashok Veeraraghavan $^{2}$ , Ramesh Raskar $^{1}$ + +1Massachusetts Institute of Technology, 2Rice University + +{ktiwary,behari,tzofi,raskar}@mit.edu，{akshat,vashok}@rice.edu + +![](images/2df5ce9a07841845a9d542726a640bb85b9e6d9291e58e9478efb66eb984ffe0.jpg) +(a) Scene: Glossy Elephant in a room + +![](images/dc19608bbd6e867f5468407562bad93e13addd859ab7ff6236d1e71d79f9430b.jpg) +(b) Input: images with distorted reflections + +![](images/31e4fd0b5ff2c092526cf8c945975aff577b819c566fc32bc84692aab0160134.jpg) +Figure 1. Objects as radiance-field cameras. We convert everyday objects with unknown geometry (a) into radiance-field cameras by modeling multi-view reflections (b) as projections of the 5D radiance field of the environment. We convert the object surface into a virtual sensor to capture this radiance field (c), which enables depth and radiance estimation of the surrounding environment. We can then query this radiance field to perform beyond field-of-view novel view synthesis of the environment (d). + +![](images/319140cdbea5bb5445735e9ea44a9cd49aae383dd620fbddcb858941bfcc8e1d.jpg) + +![](images/4e21bd1d0ee90f2978ca334273eca7b705f28da5a05562d24c5f88cc947fccc9.jpg) +(c) Using objects to image beyond camera's field-of-view + +![](images/a2b083707994440a1e0da6eb06ee8d40808b8bdec0bcb43e6e6c39ed27b00c62.jpg) +(d) Recovered and Undistorted RGBs + +![](images/7fae400df005cfb7ceba6e51ff03cb387e2efef519e097a20a61fd25e9dea927.jpg) + +# Abstract + +Reflections on glossy objects contain valuable and hidden information about the surrounding environment. By converting these objects into cameras, we can unlock exciting applications, including imaging beyond the camera's field-of-view and from seemingly impossible vantage points, e.g. from reflections on the human eye. However, this task is challenging because reflections depend jointly on object geometry, material properties, the 3D environment, and the observer's viewing direction. Our approach converts glossy objects with unknown geometry into radiance-field cameras to image the world from the object's perspective. Our key insight is to convert the object surface into a virtual sensor that captures cast reflections as a 2D projection of the 5D environment radiance field visible to and surrounding the object. We show that recovering the environment radiance fields enables depth and radiance estimation from the object to its surroundings in addition to beyond field-of-view + +novel-view synthesis, i.e. rendering of novel views that are only directly visible to the glossy object present in the scene, but not the observer. Moreover, using the radiance field we can image around occluders caused by close-by objects in the scene. Our method is trained end-to-end on multi-view images of the object and jointly estimates object geometry, diffuse radiance, and the 5D environment radiance field. For more information, visit our website. + +# 1. Introduction + +Imagine that you're driving down a city street that is packed with lines of parked cars on both sides. Inspection of the cars' glass windshields, glossy paint, and plastic reveal sharp, but faint and distorted views of the surroundings that might be otherwise hidden from you. Humans can infer depth and semantic cues about the occluded areas in the environment by processing reflections visible on reflective objects, internally decomposing the object geometry and radiance from the specular radiance being reflected onto it. Our aim is to decompose the object from its reflections to "see" the world from the object's perspective, effectively turning + +(a) Learning Neural Implicit Surfaces [Sec. 3.2] +![](images/336fad8acd27ed663001f77ba70be48d9fce30a1b40b4f2c16201dee72c312f1.jpg) +Real camera origin + +Figure 2. ORCa Overview. We jointly estimate the object's geometry and diffuse along with the environment radiance field estimation through a three-step approach. First, we model the object as a neural implicit surface (a). We model the reflections as probing the environment on virtual viewpoints (b) estimated analytically from surface properties. We model the environment as a radiance field queried on these viewpoints (c). Both neural implicit surface and environment radiance field are trained jointly on multi-view images of the object using a photometric loss. +![](images/31c4ee92e85bdedbf20674e064bab88310bde6e2ece202491963e1b2013d9540.jpg) +Real pixel cone +- Surface intersection +Virtual pixel cone +Virtual camera origin + +![](images/351c007cb16eae15b70ea2b301d7329906fe4afd514b299eee6a2d2089679df2.jpg) +(b) Objects Surface as Virtual Sensor [Sec. 3.3] + +![](images/f957743e124559cdadc315e2b75882f4c4ed9a059097d2b671298f01d58da310.jpg) + +![](images/79e2e5a652c3fc556155146662f212966fa57dd0870b932f09a0ce16f5fea170.jpg) + +![](images/779145651c7b8316abd2c83fd2648e6f16c1d6536a4200c71278aab7c12e34d3.jpg) +(c) Learning Environment Radiance Fields [Sec. 3.4] + +![](images/05cf6acc5514021a0e70d0ad07726823cb7affe5b4c7e89d84572cc68281700e.jpg) +Virtual View +Virtual Depth + +the object into a camera that images its environment. However, reflections pose a long-standing challenge in computer vision as the reflections are a 2D projection of an unknown 3D environment that is distorted based on the shape of the reflector. + +To capture the 3D world from the object's perspective, we model the object's surface as a virtual sensor that captures the 2D projection of a 5D environment radiance field surrounding the object. This environment radiance field consists largely of areas only visible to the observer through the object's reflections. Our use of environment radiance fields not only enables depth and radiance estimation from the object to its surroundings but also enables beyond field-of-view novel-view synthesis, i.e. rendering of novel views that are only directly visible to the glossy object present in the scene but not the observer. Unlike conventional approaches that model the environment as a 2D map, our approach models it as a 5D field without assuming the scene is infinitely far away. Moreover, by sampling the 5D radiance field, instead of a 2D map, we can capture depth and images around occluders, such as close-by objects in the scene, as shown in Fig. 3. These applications cannot be done from a 2D environment map. + +We aim to decompose reflections on the object's surface, from its surface and exploit those reflections to construct a radiance field surrounding the object, therefore capturing the 3D world in the process. This is a challenging task because the reflections are extremely sensitive to local object geometry, viewing direction and inter-reflections due to the object's surface. To capture this radiance field, we convert glossy objects with unknown geometry and texture into radiance-field cameras. Specifically, we exploit neural rendering to estimate the local surface of the object viewed + +![](images/aeb4110451eb9505f87a500f1fe33aaa68a383a804374a2cba6aee5e98fd4d1e.jpg) +(a) Sample Input Views + +![](images/fbafe69d37013587b525762efddc55cca4a5bba27ead0abc2b7e3edc12c30e06.jpg) +(b) Virtual Views with Translation +(c) Virtual View Depth + +![](images/7f89207a5e51e11d24714ceb23ad8cae15edef672cca9e9b3768842a48d12ba6.jpg) + +![](images/22bcbfb95f7bdc1a496880cc6aee049956973e1d3b15df3cab447d26c0ae733e.jpg) +(d) Parallax effect in 5D Env. Radiance Field + +![](images/b20d6b50a427e566c9f341272c5fec2348027903e9276308a271ec6b2683132c.jpg) + +![](images/7f6a08274bc3335ececbd90b34197b8e3717f76675954522b74248b0ef7f7f0f.jpg) + +![](images/5ed18484080d02f54b9c078dca844f100f84775f461a893f9e4ababb07ee2d4a.jpg) +x 0 +Figure 3. Advantages of 5D environment radiance field. Modeling reflections on object surfaces (a) as a 5D env. radiance field enables beyond field-of-view novel-view synthesis, including rendering of the environment from translated virtual camera views (b). Depth (c) and environment radiance of translated and parallax views can further enable imaging behind occluders, for example revealing the tails behind the primary Pokemon occluders (d). + +![](images/bb1091cb84235be474a23582eacd3ddda3c203b1817093900e6a0022c52c262d.jpg) ++0.25 + +![](images/67c7c9425f8b662d3be565cdae4237350d080244862d3ecf337dc718ab2cd1e6.jpg) ++0.50 + +from each pixel of the real camera. We then convert this local surface into a virtual pixel that captures radiance from the environment. This virtual pixel captures the environment radiance as shown in Fig 5. We estimate the outgo + +ing frustum from the virtual pixel as a cone that samples the scene. By sampling the scene from many virtual pixels on the object surface, we construct an environment radiance field that can be queried independently of the object surface, enabling beyond field-of-view novel-view synthesis from previously unsampled viewpoints. + +Our approach jointly estimates object geometry, diffuse radiance, and the environment radiance field from multiview images of glossy objects with unknown geometry and diffuse texture in three steps. First, we use neural signed distance functions (SDF) and an MLP to model the glossy object's geometry as a neural implicit surface and diffuse radiance, respectively, similar to PANDORA [10]. Then, for every pixel on the observer's camera, we estimate the virtual pixels on the object's surface based on the estimated local geometry from the neural SDF. We analytically compute the parameters of the virtual cone through the virtual pixel. Lastly, we use the cone formulation in MipNeRF [5] to cast virtual cones from the virtual camera to recover the environment radiance. + +To summarize, we make the following contributions: + +- We present a method to convert implicit surfaces into virtual sensors that can image their surroundings using virtual cones. (Sec. 3.3) +- We jointly estimate object geometry, diffuse radiance, and estimate the 5D environment radiance field surrounding the object. (Fig. 7 & 9) +- We show that the environment radiance field can be queried to perform beyond-field-of-view novel viewpoint synthesis, i.e. render views only visible to the object in the scene (Section 3.4) + +Scope. We only model glossy objects with low roughness as such specular reflections tend to have a high signal-to-noise ratio, therefore, are a sharper estimate of the environment radiance field. However, we note that the virtual cone computation can be extended to model the cone radius as a function of surface roughness. Deblurring approaches can further improve the resolution of estimated environment. In addition, we approximate the local curvature using mean curvature, which fails for objects with varying radius of curvature along the tangent space. We explain how our virtual cone curvature estimation can be extended to handle general shape operators in the supplementary material. Lastly, similar to other multi-view approaches, our approach relies on a sufficient virtual baseline between virtual viewpoints to recover the environment radiance field. + +# 2. Related Work + +# 2.1. Modeling reflections + +Catadioptric imaging systems incorporate reflections on curved reflective mirrors to expand the field of view of con + +ventional cameras [2, 26], to increase the baseline of light field cameras [11, 40] and to perform novel view synthesis from a single capture [44]. These works assume the geometry of the reflecting surface is known or calibrated while we create a catadioptric imaging system from everyday glossy objects of unknown geometry. Grossberg et al. [15] propose a generalized model for light transport through imaging systems, including catadioptric systems. Our work focuses on light transport reflecting off a general object in the scene. + +Light field imaging is shown to be effective for reflection removal [21, 27], reconstructing specular surfaces [18], intrinsic decomposition [1], and neural rendering [35]. These works typically consider planar reflections or require training on synthetic datasets, while our approach models reflections on complex geometry and is unsupervised. Prior works have also utilized additional light properties such as polarization [9, 10, 20] and time-of-flight [17, 33] for the separation of the reflected component and specular surface reconstruction. While the input of our approach is RGB images, there is scope for improving reconstruction quality by supplementing the algorithm with these additional cues. + +Recent progress in neural radiance fields has enabled impressive novel view rendering and geometry reconstruction from multi-view images [24]. MipNeRF [5] demonstrates better novel view synthesis by modeling outgoing rays as cones to enable anti-aliasing. RefNeRF [41] proposes Integrated Directional Embeddings for improved novel view synthesis of reflections. NeRFRen [16] separates diffuse and specular radiance by using separate neural networks. Neural Catacaustics [19] propose a neural warping method to model reflections by learning the caustics of the surface. While these works focus on novel-view synthesis of the scene from the primary camera, we perform view synthesis that is beyond the line-of-sight of the primary camera, i.e., rendering views only visible to the objects present in the scene, while jointly estimating object geometry and separating diffuse and specular radiance. + +# 2.2. Environment Estimation + +Recovering underlying scene properties from multiple images is inherently ill-posed [32], but can be regularized using the natural statistics of scene properties as a prior [4, 34]. Recent works exploit this prior through deep neural networks and demonstrate inverse rendering of indoor scenes from a single image [12, 22, 43, 48]. However, these techniques typically recover only coarse representations of lighting and cannot reconstruct fine details of the environment. Lombardi et al. [23] recover environment and reflectance, assuming the scene is composed of known geometry and uniform material. Georgolis et al. [13] recover the environment map behind the camera from a single image of a glossy object, assuming the object is composed of textureless materials and using ground truth seg + +
ApproachInputScene GeometryAlgorithm TypeEnvironment Dimension
Neural IllumSingle RGBNormalsSupervised2D
LighthouseStereo RGB pairMulti-plane ImageSupervised5D
NeRFactorMulti-view RGBVolumetric AlbedoSemi-supervised2D
RefNeRFMulti-view RGBVolumetric AlbedoSelf-supervised2D
PANDORAMulti-view RGB+Pol.Neural SDFSelf-supervised2D
ORCa (Ours)Multi-view RGBNeural SDFSelf-supervised5D
+ +mentation masks. Song et al. [36] estimate plausible environment maps by mapping reflections in the image and inpainting unmapped regions. Srinivasan et al. [37] capture stereo image pairs and estimate plausible spatially-coherent environment maps. NeRD [7], NeRFactor [47] and NeuralPIL [8] employ data-driven priors for lighting and BRDF in a NeRF-based approach for radiance decomposition. Park et al. [31] use RGB-D videos to estimate environment map. Swedish et al. [39] recover high-frequency illumination map from the shadows of an object with known geometry. PhySG [46] and Munkberg et al. [25] perform inverse rendering from multi-view images by modeling the surface as signed distance functions. PANDORA [10] performs radiance decomposition from polarized RGB images. + +# 3. Learning environment radiance fields from multi-view reflections + +# 3.1. Overview + +Reflections on glossy objects offer a glimpse into the surrounding environment beyond the camera's field of view. From multi-view images of a glossy object with unknown geometry and albedo, we aim to recover the 5D radiance field of the surrounding environment. The mapping from images captured by the observer to the surrounding environment depends on the glossy object's surface properties, in particular, the surface normals and curvature. We first cast a cone from the observer camera's center-of-projection through each pixel viewing the scene. When the cone intersects the object's surface, it reflects, causing the cone to be transformed (Fig. 5). The transformed cone, referred to as a virtual cone, samples the environment and is primar- + +![](images/779df1b504cb234de996732bbc934595f060771d044a45fbd83b5e114802095c.jpg) +Figure 4. Comparison of environment estimation approaches. Approaches such Neural Illum [36], Lighthouse [37], NeRFactor [47] train on datasets of natural illumination maps to regularize ill-posedness of environment estimation. PANDORA [10] and RefNeRF [41] exploit multi-view reflections on objects but assume the surrounding environment is infinitely far away and is modeled with a flat 2D map. From multi-view reflections, ORCa converts the object surface into a virtual sensor and extracts the 5D radiance field of the environment. +Figure 5. Virtual Sensor. We image the world through the object by modeling each pixel's specular radiance as a projection of the 5D radiance field of the environment onto the object's surface. We capture the radiance field by treating the surface area on the object that the pixel views, $\mathbf{dS}_t$ , as a single-pixel virtual camera with its center-of-projection at $\mathbf{v}_o$ . We cast virtual cones through the virtual sensor to capture the 5D radiance field of the environment. + +ily responsible for the specular radiance observed on the glossy object. Our key insight is that the reflections captured by the observer's camera can be modeled as a projection of the environment radiance field onto the object's surface. By modeling the reflected rays as a cone and computing the parameters of the cone, we can more accurately estimate the projected environment radiance field onto the object surface, as shown in Fig. 10. + +ORCa is composed of three steps: modeling the object's geometry as a neural implicit surface (Sec. 3.2), converting the object's surface into a virtual sensor (Sec. 3.3), and modeling the environment radiance field as a projection along these virtual cones (Sec. 3.4). The learned environment radiance field can then be queried on novel viewpoints to show occluded areas in the scene. Fig. 2 depicts our output for each component on a scene rendered with a complex glossy object and 3D environment. Fig. 6 shows our system architecture. Next, we describe each step in detail. + +# 3.2. Learning Neural implicit Surfaces + +Neural Signed Distance Function We model the object geometry as a neural signed distance function (SDF). $f: \mathbb{R}^3 \to \mathbb{R}$ . SDFs provide a helpful inductive bias for learning smooth surface geometry [30, 42, 45] that assists downstream tasks in our pipeline. Moreover, the surface properties crucial for our framework, surface normals and curvature, can be conveniently computed from SDFs in a differentiable manner. Consider the 3D spatial coordinates, $\mathbf{x}$ , in the scene. The glossy object surface, $S$ is then represented by the zero-level set of the SDF + +$$ +\mathcal {S} = \left\{f _ {\mathcal {S}} (\mathbf {x}) = 0 | \mathbf {x} \in \mathbb {R} ^ {3} \right\} \tag {1} +$$ + +![](images/b38a38c708192bc41253a0ce6fa511bf54b2beedb6b028f437d397b52b2721d8.jpg) +Figure 6. Overview of our proposed architecture. + +Similar to Yariv et al. [45], we model the SDF $f_{S}$ as a coordinate-based MLP. + +Surface Normals Gradients of the SDF at the zero level set point $S$ towards the surface normals $S$ , + +$$ +\mathbf {n} (\mathbf {x}) = \frac {\nabla_ {\mathbf {x}} f _ {\mathcal {S}} (\mathbf {x})}{\| \nabla_ {\mathbf {x}} f _ {\mathcal {S}} (\mathbf {x}) \|} \quad \mathbf {x} \in \mathcal {S} \tag {2} +$$ + +Surface Curvature We employ differential geometry techniques developed by Novello et al. [29] to estimate curvature for neural implicit surfaces. In particular, we estimate the mean curvature $K(\mathbf{x})$ for the implicit surface from the divergence, $\nabla$ of the surface normals + +$$ +K (\mathbf {x}) = \frac {\nabla \cdot \mathbf {n} (\mathbf {x})}{2} \tag {3} +$$ + +Mean curvature approximates the surface with an osculating sphere. Our approach also works for more generalized notions of curvature through the shape operator, at the cost of higher computational complexity. We refer our readers to the supplement for the general case. + +Diffuse Radiance We separate the captured radiance at the observer camera with diffuse radiance, which depends on the glossy object's albedo, and specular radiance which depends on the environment radiance. The diffuse radiance does not have any view dependence and only depends on surface point $\mathbf{x}$ . We denote the diffuse radiance as $f_{d}$ and model it using a coordinate-based MLP (Fig. 6). + +Volume Rendering As proposed in [45], we perform volumetric rendering on the SDF. We define the volume density $\sigma(\mathbf{x})$ as the cumulative distribution function (CDF), denoted as $\Psi(s)$ , applied to $f_{S}$ : + +$$ +\sigma (\mathbf {x}) = \alpha \Psi_ {\beta} (f _ {\mathcal {S}}) \tag {4} +$$ + +In contrast to [45], however, we only aim to recover the diffuse radiance of the object along a particular ray. We define a function $f_{d}$ that estimates the diffuse radiance at each point, $\mathbf{x}$ , along the ray. To get the final diffuse radiance along a given primary ray, $\mathbf{r}_p(t)$ , we perform volumetric rendering: + +$$ +\hat {\mathbf {c}} _ {d} (\mathbf {r}) = \int_ {0} ^ {\infty} f _ {d} (\mathbf {r} (t), f _ {S} ^ {k} (\mathbf {r} (t)) \tau (t) d t \tag {5} +$$ + +Note that there is no view dependence in Eq. 5 and intermediate features, $f_{\mathcal{S}}^{k}$ , are used as input. $\tau(t)$ is the accumulated transmittance along the ray. + +![](images/0529fab554ead1c2b7ef537769a18371272aa8a29144141cda2b7c3683399f99.jpg) +Figure 7. Qualitative comparisons of diffuse-specular separation and geometry estimation on rendered dataset. The environment contains nearby objects with complex occlusions when seen through reflections on the glossy object. RefNeRF fails to perform accurate diffuse-specular separation and PANDORA blurs the nearby objects in the specular map. ORCA can model the complex specular reflections through environment radiance field. + +# 3.3. Objects Surface as Virtual Sensor + +Each pixel, $\mathbf{p}$ , with a finite surface area, $\mathbf{dA}_p$ , on the real-camera sensor views the surface of the object through a frustum originating at that pixel. The object then samples the environment radiance field through this finite surface converting the finite surface into a virtual pixel with surface area, $\mathbf{dS}$ . Through this model, we can interpret the object surface as a virtual sensor consisting of many virtual pixels that sample radiance from the environment field based on the geometry of the object and observer viewing direction. We now formulate a virtual pixel based on real camera pose and implicit surface geometry. Please refer to Fig. 5 for a visualization of the virtual sensor. + +Consider a real camera origin as $\mathbf{o}$ and a pixel on the real sensor $\mathbf{p}_{i,j}$ that corresponds to ray direction $\mathbf{d}$ . The primary ray for pixel $\mathbf{p}_{i,j}$ is parameterized with ray length $t$ as $\mathbf{r_p}(t) = \mathbf{o} + t\mathbf{d}$ + +Casting Real Cones We can approximate the outgoing conical frustum from pixel $\mathbf{p}_{i,j}$ as a cone originating at $\mathbf{o}$ with axis-of-direction $\mathbf{d}$ and radius $\dot{r}$ , equivalent to half the distance of the pixel in the x and y directions. We represent the real cone as a parametric volume, + +$$ +\mathbf {r} _ {\text {c o n e}} (\dot {r}, s, \theta) = \dot {r} s \cos (\theta) \hat {\mathbf {e}} _ {u} + \dot {r} s \sin (\theta) \hat {\mathbf {e}} _ {v} + \dot {r} s \mathbf {d}, \tag {6} +$$ + +where $\hat{\mathbf{e}}_u$ and $\hat{\mathbf{e}}_v$ are basis vectors in the plane perpendicular + +to $\mathbf{d}$ , $\theta \in [0, \pi]$ and $s \in [0, t_{\max}]$ + +Virtual Pixel. Virtual pixels are characterized by the intersection of the real cone with the object surface. In Sec. 3.2, we model local surface properties using mean curvature which enables efficient analytical computations for the virtual pixel parameters even though our approach works with general shape operators. For a sampled point $t_i$ along the ray, we have the surface normals $\mathbf{n}(t_i)$ from Eq. 2 and estimated mean curvature $K(t_i)$ from Eq. 3. The local object surface at $t_i$ , can be approximated with an osculating sphere, $\mathcal{O}(t_i)$ , centered at $\hat{\mathbf{o}}_{\mathbf{S}}(t_i)$ with radius, $R(t_i)$ as + +$$ +R (t _ {i}) = \frac {2}{K _ {t _ {i}}} \hat {\mathbf {o}} _ {\mathbf {S}} (t _ {i}) = \mathbf {r} _ {p} (t _ {i}) + R (t _ {i}) \cdot \hat {\mathbf {n}} (t _ {i}) +$$ + +For concave surfaces $K_{t_i} < 0$ . So, $\hat{\mathbf{o}}_{\mathbf{S}}$ will lie outside the object. For $K_{t_i} > 0$ , $\hat{\mathbf{o}}_{\mathbf{S}}$ will lie inside the object. + +The edges of the virtual pixel for $\mathbf{r_p}(t_i)$ would lie at the intersection of the osculating sphere $\mathcal{O}(t_i)$ and the primary cone given by $\mathbf{r}_{\mathrm{cone}}$ . Computing exact cone-sphere intersections are computationally expensive so we approximate the cone-sphere intersection using rays bound cone-sphere intersectional surface dS. We consider four rays that bound the cone and sample them at $\theta_{j}\in \{0,\pi /2,\pi ,3\pi /2\}$ with Eq. 6. We perform intersections of the corresponding bounding rays with the osculating sphere $\mathcal{O}(t_i)$ to get corners of the virtual pixel $\mathbf{ds}_j$ . These ray sphere intersections can be computed analytically in an efficient manner. + +Virtual Cone Origin. From the virtual pixel surface area, we can now compute the virtual cone that samples the environment. We first compute normal vectors at virtual pixel corners $\mathbf{ds_j}$ from the center of osculating sphere $\hat{\mathbf{o}}_{\mathrm{S}}$ + +$$ +\hat {\mathbf {n}} _ {\mathbf {j}} = \frac {\mathbf {d} \mathbf {s} _ {\mathbf {j}} - \hat {\mathbf {o}} _ {\mathbf {S}}}{\left| \left| \mathbf {d} \mathbf {s} _ {\mathbf {j}} - \hat {\mathbf {o}} _ {\mathbf {S}} \right| \right|} \tag {7} +$$ + +At each virtual pixel corner, we compute the reflected ray directions, $\omega_{j}^{r}$ , by computing the dot product between the incoming ray directions, $\omega_{k}^{i}$ , and the normals, $\hat{\mathbf{n}}_k$ , where $\omega_0^r$ is the primary ray's reflected vector. + +$$ +\omega_ {0} ^ {r} = \mathbf {d} - \left(\mathbf {d} \cdot \hat {\mathbf {n}} \left(t _ {i}\right)\right) \hat {\mathbf {n}} \left(t _ {i}\right) \tag {8} +$$ + +$$ +\omega_ {j} ^ {r} = \mathbf {d} _ {j} - \left(\mathbf {d} _ {j} \cdot \hat {\mathbf {n}} _ {j} (k)\right) \hat {\mathbf {n}} _ {j} (k) \tag {9} +$$ + +$\mathbf{d}_j$ are the incident directions to the virtual pixel corners $\mathbf{ds}_j$ . The virtual cone origin is the intersection of these reflected rays at the pixel corners and pixel center. However, these rays might not intersect at a single point so we approximate a virtual origin to be the point that minimizes the sum of distances to the reflected rays $\omega_j$ . + +$$ +\mathbf {v} _ {\mathbf {o}} = \operatorname {a r g m i n} _ {\mathbf {v}} \sum_ {j} | (\mathbf {v} - \mathbf {d s} _ {\mathbf {j}}) \times \omega_ {j} ^ {r} | \tag {10} +$$ + +We pose this as a linear least squares problem and estimate the virtual cone origin efficiently through pseudo-inverse. + +Virtual Cones Direction. The reflected ray at the center of the virtual pixel reflects the object surface along the direction $\omega_0^r$ from Eq. 8. We consider this as the direction-of-axis of the virtual cone. + +$$ +\hat {\mathbf {v}} _ {d} = \omega_ {0} ^ {r} \tag {11} +$$ + +Virtual Cone Radius. We compute the radius of the cone by treating the reflection vectors of the bounding rays as the neighboring "pixel" directions. Similar to [5], we can compute the distance between $\{\omega_{k_{\theta}}^{r}\}_{\theta = 0}^{2\pi}$ and the primary reflected ray $\omega_0^r$ in the $(x,y)$ components (omitted below for clarity). + +$$ +\hat {\mathbf {v}} _ {\dot {r}} = \left\| \left\{\omega_ {k _ {\theta}} ^ {r} \right\} _ {\theta = 0} ^ {2 \pi} - \omega_ {0} ^ {r} \right\| \tag {12} +$$ + +Finally, for each sampled point $t_i$ , we can characterize our single-pixel virtual sensor located at the object surface $\mathbf{dS}$ as a virtual cone with $\hat{\mathbf{v}}_o$ as its apex, $\hat{\mathbf{v}}_d$ as axis-direction, $\hat{\mathbf{v}}_{\dot{r}}$ as the radius. + +Connections to caustics. Our work takes inspiration from catadioptric imaging systems. To convert objects into cameras, we compute the surface and find a corresponding center-of-projection for this surface-as-sensor. However, unlike conventional perspective cameras, objects don't have a fixed center-of-projection, other than in a few special configurations [3], but a locus of viewpoints that vary with object geometry and viewing direction. These viewpoints lie on the "caustic surface" of the object. While typical works in catadioptric imaging analytically compute the caustic surface by assuming known geometry [14, 38], or making assumptions about placement of the observer [40], our formulation approximates the caustic surface of unknown geometry through the intersection of reflected rays on virtual pixels. We empirically show in the supplement that as the surface area of the virtual pixel goes to $0$ , $d\mathbf{S} \rightarrow 0$ , our method estimates the true caustic of the object without assuming geometry. Our method also has applications in estimating the caustic surface of the unknown geometry. + +# 3.4. Environment Radiance Fields + +Our goal is to capture a 5D environment radiance field of the scene by imaging the world through these single-pixel virtual sensors located at the object's surface. We use our formulation of virtual cones to recover 5D environment radiance fields. We define an environment radiance field as $f_{\mathcal{E}}: (\hat{\mathbf{v}}_o, \hat{\mathbf{v}}_d) \to (\sigma^{Env}, c_s)$ , + +where $f_{\varepsilon}$ outputs opacity and radiance along sampled virtual cones. We note that this view-dependent radiance is equivalent to the specular radiance at point $t_i$ sampled + +![](images/413e0df2124e259ffdb6edf7e6eeb29af2b2fa53113aac8e2ddfce0478c9883e.jpg) +Figure 8. Evaluation of environment depth from reflections. For the scene in Fig. 3, ORCa can faithfully estimate the depth of the environment, especially for nearby surroundings. + +
ApproachDiffuse RadianceSpecular RadianceMixed RadianceNormals
PSNR ↑ (dB)SSIM ↑PSNR ↑ (dB)SSIM ↑PSNR ↑ (dB)SSIM ↑MAE ↓ (°)
Ref-NeRF18.800.730416.990.663320.190.789043.690
PANDORA18.250.726016.250.648318.900.72847.606
ORCA19.840.789320.740.753522.000.79472.339
+ +along the primary-camera ray $\mathbf{r}_{\mathbf{p}}(t)$ . We can render the final specular radiance at pixel $\mathbf{p}_{i,j}$ as follows: + +$$ +\hat {\mathbf {c}} _ {s} (\mathbf {r}) = \int_ {0} ^ {\infty} f _ {\mathcal {E}} (\hat {\mathbf {v}} _ {o}, \hat {\mathbf {v}} _ {d}) \tau (t) d t +$$ + +$$ +\hat {\mathbf {c}} = \hat {\mathbf {c _ {d}}} + \hat {\mathbf {c _ {s}}} +$$ + +Intuitively, $f_{\mathcal{E}}$ learns the 5D radiance field by sampling single-pixel virtual sensors from the object surface and must learn geometry and environment radiance that is consistent with the multi-view reflections. Moreover, we can query $f_{\mathcal{E}}$ to render novel viewpoints and associated depths that are beyond the field-of-view of the real camera. We volume render each virtual cone by dividing them into conical frustums using Integrated-Positional Encoding as proposed in MipNeRF [5]. Our formulation of virtual cones works well with Mip-Nerf's rays-as-cones method. + +# 4. Experiments + +# 4.1. Implementation Details + +As in PANDORA, we parameterize $f_{S}$ with an 8-layer MLP to estimate the surface, and, as in MipNeRF, $f_{d}$ with 4-layer MLP with input geometric features of size 512 from $f_{S}$ . We follow the SDF-to-opacity conversion and the iterative sampling of the ray proposed in [45]. To aid the network to learn geometry quickly, we also train $f_{S}$ with a mask-net as proposed in [10]. We use five losses in our architecture: photometric loss, mask loss [10], normal loss [41], eikonal loss [45], and distortion loss [6]. Additional training details are discussed in the supplement. + +Table 1. Average evaluation metrics on rendered scenes. We compare ORCa to other neural rendering techniques that model reflections, including Ref-NeRF and PANDORA, on six simulated scenes. ORCa provides consistent improvements in geometry estimation, diffuse-specular separation and novel view synthesis. Please refer to the supplement for additional metrics. + +
Scene: Ball-cup in hallway
ApproachMixedRadiance
PSNRSSIM
↑(dB)↑
Ref-NeRF32.750.9617
PANDORA28.830.9758
ORCA30.860.9799
+ +
Scene: Owl in hallway
ApproachMixed Radiance
PSNRSSIM
↑ (dB)↑
Ref-NeRF26.650.8890
PANDORA27.240.9343
ORCA26.840.9299
+ +Table 2. Quantitative metrics for real-world, captured scenes. ORCa demonstrates comparable performance in novel view synthesis on scenes from the PANDORA real dataset. + +# 4.2. Datasets + +We conduct experiments on both simulated and real-world datasets. Simulated datasets are rendered in Mitsu2 [28]. Simulated datasets contain a range of increasingly complex object geometries (elephant, Pokeball, and orca) and scenes (living room and Pokemon). We train with 200 views for simulated datasets. We also show results for a real-world dataset [10] capturing a glossy cup with a black vase sitting atop it using 35 views. All datasets will be publicly released upon publication. + +# 4.3. Advantages of Environment Radiance Fields + +Other neural rendering techniques that handle reflections, Ref-NeRF and PANDORA, estimate the environment as a 2D map, while ORCa recovers a 5D environment radiance field. Fig. 3 shows the advantages of recovering a 5D environment radiance field. Close-by surrounding objects often cause occlusions that cannot be modeled by 2D environment maps. By estimating the radiance field, we can image behind occluders through sampling novel viewpoints such as the translated viewpoints shown in Fig. 3. Moreover, we can also show the depth to the surroundings from these virtual viewpoints. In Fig. 8, we quantitatively evaluate the estimated depth of the surroundings for the synthesized virtual views in Fig. 3 and show that ORCa provides reliable depth estimates especially for nearby objects. We provide additional examples of depth estimation and beyond field-of-view novel-view synthesis in the supplement. + +# 4.4. Comparisons with Baselines + +We compare ORCa with Ref-NeRF and PANDORA on synthetic and real datasets. As shown in Fig. 7 and 9, ORCa estimates more accurate surface normals than other methods. While the total radiance predicted by Ref-NeRF and PANDORA are visually similar, the surface normals are less smooth than ORCa. We also observe that ORCa is able to achieve better diffuse and specular radiance separation than PANDORA, which is evident in the Pokeball surface normals Fig. 7. In these examples, PANDORA recovers blurry specular radiance. We see that ORCa's predicted depth is highly interpretable and matches the underlying geometry of the environment, as shown in Fig. 3. Even on cylindri + +![](images/b8c49e4f653135a5bfcfea828e7f63b0338865f2c803608894e4402fc01b0a06.jpg) +Figure 9. Comparisons on real dataset. Using a real-world dataset with only 35 views, ORCa can model the sharp specularities on the ball arriving from regions of the nearby scene, such as the table, and far away scene regions, such as the hallway, to learn an environment radiance field. We query the radiance field for the depth of the far hallway (blue) and the nearby objects, such as the table (red). We also render novel viewpoints that are beyond the field of view of the observer camera and show that ORCa interpolates well between those views. + +![](images/cb2188ee0525ae84b3267adaa2b3c2983a96e4c7147e764a619a70c63dbd4a81.jpg) +Figure 10. Ablations on virtual cone computation Accurate estimation of the virtual cones (Sec. 3.3) is crucial for ORCa. If the virtual cone origin is assumed to be at the object surface (left column) or if the surface is assumed to locally have no curvature (right column), the surface normal and specular radiance outputs suffer from artifacts (red boxes). + +cal real-world datasets, such as the black vase in Fig. 9, the nearby hallway is visible in both the virtual view and depth, despite never being in the field of view of the primary camera. Unlike Ref-NeRF, our primary objective is not to perform novel-view synthesis, but instead to capture the environment radiance field from the object surface. As shown in Table 1, ORCa improves over Ref-NeRF and PANDORA in estimating diffuse radiance, specular radiance, mixed radiance, and normals. Here we report the image quality metrics, PSNR and SSIM, and surface normal metric, MAE, averaged over six synthetic scenes with varying environments, camera view distribution, and object complexity. Please refer to the supplement for individual metrics on each scene. ORCa also shows competitive performance for novel-view synthesis metrics on real-world scenes from PANDORA dataset, as shown in Table 2. + +# 4.5. Impact of Virtual Cone Computation + +We base our method on a physically accurate formulation by modeling ray-cone intersections and using the surface as a virtual sensor, as described in Sec. 3.3. We demonstrate the importance of this step by setting up two ablation experiments. In the first experiment, which we term Naive Virtual Cone, we place the origin of the cone at the object surface instead of computing the virtual cone origin based + +![](images/a17c9af40a96d28275bd72eb531568ca8e35ab246ece9f22045da15905766aba.jpg) + +on the curvature. In the second experiment, which we term No Curvature, we assume that locally the surface is like a flat mirror and has no curvature. In Fig. 10, we show that for both of these ablation experiments, we see worse performance as demonstrated by the artifacts in estimated surface normals and specular components (shown as red boxes). + +# 5. Conclusion + +We present a method to convert glossy objects with unknown geometry and texture into radiance-field cameras that capture the environment's radiance field around them. Our method recovers object geometry and diffuse radiance, in addition to capturing the depth and radiance of the object's surroundings from its perspective. Our modeling of the environment as a radiance field is effective in recovering close-by objects (Fig. 7) and is occlusion aware (Fig. 3). From the recovered environment radiance field, we can perform beyond field-of-view novel-view synthesis. Our work can unleash applications in virtual object insertion and 3D perception, e.g. inferring information beyond the line-of-sight of the camera using predicted virtual views and depth. Our formulation goes beyond the conventional direct-line-of-sight radiance fields and can enable further areas of research to extract more information from multi-view images directly from the environment and the objects present in it. + +# Acknowledgments + +This research was supported by the SMART Contract IARPA Grant 2021-20111000004. We would also like to thank Systems Technology Research (STR). AD and AV are supported by NSF Expeditions-Award IIS-1730574 and NSF CAREER-Award IIS-1652633. + +# References + +[1] Anna Alperovich, Ole Johannsen, and Bastian Goldluecke. Intrinsic light field decomposition and disparity estimation with deep encoder-decoder network. In 2018 26th European Signal Processing Conference (EUSIPCO), pages 2165-2169. IEEE, 2018. 3 +[2] S. Baker and S.K. Nayar. A theory of catadioptric image formation. In Sixth International Conference on Computer Vision (IEEE Cat. No.98CH36271), pages 35-42, 1998. 3 +[3] Simon Baker and Shree K. Nayar. A theory of single-viewpoint catadioptric image formation. International Journal of Computer Vision, 35:175-196, 2004. 6 +[4] Jonathan T Barron and Jitendra Malik. Shape, illumination, and reflectance from shading. IEEE transactions on pattern analysis and machine intelligence, 37(8):1670-1687, 2014. 3 +[5] Jonathan T. Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P. Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. ICCV, 2021. 3, 6, 7 +[6] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5470–5479, 2022. 7 +[7] Mark Boss, Raphael Braun, Varun Jampani, Jonathan T Barron, Ce Liu, and Hendrik Lensch. Nerd: Neural reflectance decomposition from image collections. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12684-12694, 2021. 4 +[8] Mark Boss, Varun Jampani, Raphael Braun, Ce Liu, Jonathan Barron, and Hendrik Lensch. Neural-pil: Neural pre-integrated lighting for reflectance decomposition. Advances in Neural Information Processing Systems, 34:10691-10704, 2021. 4 +[9] Akshit Dave, Yannick Hold-Geoffroy, Milos Hasan, Kalyan Sunkavalli, and Ashok Veeraraghavan. Snapshot polarimetric diffuse-specular separation. Optics Express, 30(19):34239-34255, 2022. 3 +[10] Akshit Dave, Yongyi Zhao, and Ashok Veeraraghavan. Pandora: Polarization-aided neural decomposition of radiance. arXiv preprint arXiv:2203.13458, 2022. 3, 4, 7 +[11] Michael De Zeeuw and Aswin C Sankaranarayanan. Wide-baseline light fields using ellipsoidal mirrors. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 3 +[12] Mathieu Garon, Kalyan Sunkavalli, Sunil Hadap, Nathan Carr, and Jean-François Lalonde. Fast spatially-varying indoor lighting estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6908-6917, 2019. 3 +[13] Stamatios Georgoulis, Konstantinos Rematas, Tobias Ritschel, Mario Fritz, Tinne Tuytelaars, and Luc Van Gool. What is around the camera? In Proceedings of the IEEE International Conference on Computer Vision, pages 5170-5178, 2017. 3 + +[14] J. Gluckman and S.K. Nayar. Planar catadioptric stereo: geometry and calibration. In Proceedings. 1999 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (Cat. No PR00149), volume 1, pages 22-28 Vol. 1, 1999. 6 +[15] Michael D Grossberg and Shree K Nayar. The raxel imaging model and ray-based calibration. International Journal of Computer Vision, 61(2):119, 2005. 3 +[16] Yuan-Chen Guo, Di Kang, Linchao Bao, Yu He, and Song-Hai Zhang. Nerfren: Neural radiance fields with reflections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18409-18418, June 2022. 3 +[17] Connor Henley, Siddharth Somasundaram, Joseph Hollmann, and Ramesh Raskar. Detection and mapping of specular surfaces using multibounce lidar returns. Optics Express, 31(4):6370-6388, 2023. 3 +[18] Jan Jachnik, Richard A Newcombe, and Andrew J Davison. Real-time surface light-field capture for augmentation of planar specular surfaces. In 2012 IEEE International Symposium on Mixed and Augmented Reality (ISMAR), pages 91-97. IEEE, 2012. 3 +[19] Georgios Kopanas, Thomas Leimkuhler, Gilles Rainer, Clément Jambon, and George Drettakis. Neural point cata-caustics for novel-view synthesis of reflections. ACM Transactions on Graphics, 41(6):Article-201, 2022. 3 +[20] Rui Li, Simeng Qiu, Guangming Zang, and Wolfgang Heidrich. Reflection separation via multi-bounce polarization state tracing. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIII 16, pages 781-796. Springer, 2020. 3 +[21] Tingtian Li, Daniel PK Lun, Yuk-Hee Chan, et al. Robust reflection removal based on light field imaging. IEEE Transactions on Image Processing, 28(4):1798–1812, 2018. 3 +[22] Zhengqin Li, Mohammad Shafiei, Ravi Ramamoorthi, Kalyan Sunkavalli, and Manmohan Chandraker. Inverse rendering for complex indoor scenes: Shape, spatially-varying lighting and svbrdf from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2475-2484, 2020. 3 +[23] Stephen Lombardi and Ko Nishino. Reflectance and natural illumination from a single image. In European Conference on Computer Vision, pages 582-595. Springer, 2012. 3 +[24] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 3 +[25] Jacob Munkberg, Jon Hasselgren, Tianchang Shen, Jun Gao, Wenzheng Chen, Alex Evans, Thomas Müller, and Sanja Fidler. Extracting triangular 3d models, materials, and lighting from images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8280-8290, 2022. 4 +[26] Shree K Nayar and Simon Baker. Catadioptric image formation. In Proceedings of the 1997 DARPA Image Understanding Workshop, pages 1431-1437, 1997. 3 + +[27] Yun Ni, Jie Chen, and Lap-Pui Chau. Reflection removal on single light field capture using focus manipulation. IEEE Transactions on Computational Imaging, 4(4):562-572, 2018. 3 +[28] Merlin Nimier-David, Delio Vicini, Tizian Zeltner, and Wenzel Jakob. Mitsuba 2: A retargetable forward and inverse renderer. ACM Transactions on Graphics (TOG), 38(6):1-17, 2019. 7 +[29] Tiago Novello, Guilherme Schardong, Luiz Schirmer, Vinicius da Silva, Helio Lopes, and Luiz Velho. Exploring differential geometry in neural implicits, 2022. 5 +[30] Michael Oechsle, Songyou Peng, and Andreas Geiger. Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5589-5599, 2021. 4 +[31] Jeong Joon Park, Aleksander Holynski, and Steven M Seitz. Seeing the world in a bag of chips. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1417-1427, 2020. 4 +[32] Ravi Ramamoorthi and Pat Hanrahan. A signal-processing framework for inverse rendering. In Proceedings of the 28th annual conference on Computer graphics and interactive techniques, pages 117-128, 2001. 3 +[33] Raskar Ramesh and James Davis. 5d time-light transport matrix: What can we reason about scene properties? Technical report, 2008. 3 +[34] Fabiano Romeiro and Todd Zickler. Blind reflectometry. In European conference on computer vision, pages 45-58. Springer, 2010. 3 +[35] Vincent Sitzmann, Semon Rezchikov, Bill Freeman, Josh Tenenbaum, and Fredo Durand. Light field networks: Neural scene representations with single-evaluation rendering. Advances in Neural Information Processing Systems, 34:19313-19325, 2021. 3 +[36] Shuran Song and Thomas Funkhouser. Neural illumination: Lighting prediction for indoor environments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6918-6926, 2019. 4 +[37] Pratul P Srinivasan, Ben Mildenhall, Matthew Tancik, Jonathan T Barron, Richard Tucker, and Noah Snavely. Lighthouse: Predicting lighting volumes for spatially-coherent illumination. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8080-8089, 2020. 4 +[38] R. Swaminathan, M.D. Grossberg, and S.K. Nayar. Cautistics of catadioptric cameras. In Proceedings Eighth IEEE International Conference on Computer Vision. ICCV 2001, volume 2, pages 2-9 vol.2, 2001. 6 +[39] Tristan Swedish, Connor Henley, and Ramesh Raskar. Objects as cameras: Estimating high-frequency illumination from shadows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2593-2602, 2021. 4 +[40] Yuichi Taguchi, Amit Agrawal, Ashok Veeraraghavan, Srikumar Ramalingam, and Ramesh Raskar. Axial-cones: Modeling spherical catadioptric cameras for wide-angle light field rendering. ACM Trans. Graph., 29(6), dec 2010. 3, 6 + +[41] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T. Barron, and Pratul P. Srinivasan. Ref-NeRF: Structured view-dependent appearance for neural radiance fields. CVPR, 2022. 3, 4, 7 +[42] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021. 4 +[43] Zian Wang, Jonah Philion, Sanja Fidler, and Jan Kautz. Learning indoor inverse rendering with 3d spatially-varying lighting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12538-12547, 2021. 3 +[44] Ziyu Wang, Liao Wang, Fuqiang Zhao, Minye Wu, Lan Xu, and Jingyi Yu. Mirrornerf: One-shot neural portrait radiance field from multi-mirror catadioptric imaging. In 2021 IEEE International Conference on Computational Photography (ICCP), pages 1-12. IEEE, 2021. 3 +[45] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. In Thirty-Fifth Conference on Neural Information Processing Systems, 2021. 4, 5, 7 +[46] Kai Zhang, Fujun Luan, Qianqian Wang, Kavita Bala, and Noah Snively. Physg: Inverse rendering with spherical gaussians for physics-based material editing and relighting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5453-5462, 2021. 4 +[47] Xiuming Zhang, Pratul P Srinivasan, Boyang Deng, Paul Debevec, William T Freeman, and Jonathan T Barron. Nerfactor: Neural factorization of shape and reflectance under an unknown illumination. ACM Transactions on Graphics (TOG), 40(6):1-18, 2021. 4 +[48] Rui Zhu, Zhengqin Li, Janarbek Matai, Fatih Porikli, and Manmohan Chandraker. Irisformer: Dense vision transformers for single-image inverse rendering in indoor scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2822-2831, 2022. 3 \ No newline at end of file diff --git a/2023/ORCa_ Glossy Objects As Radiance-Field Cameras/images.zip b/2023/ORCa_ Glossy Objects As Radiance-Field Cameras/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..3d26c9d79a456c15c85dddf6a80df806c373db57 --- /dev/null +++ b/2023/ORCa_ Glossy Objects As Radiance-Field Cameras/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8ae951d249604528d4368570d3278439162659a2c5f2fb666201d1c4e49776f1 +size 558201 diff --git a/2023/ORCa_ Glossy Objects As Radiance-Field Cameras/layout.json b/2023/ORCa_ Glossy Objects As Radiance-Field Cameras/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ccba937b95fc531f7fe8ed256ddec6d646dd6df5 --- /dev/null +++ b/2023/ORCa_ Glossy Objects As Radiance-Field Cameras/layout.json @@ -0,0 +1,9894 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 139, + 103, + 455, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 103, + 455, + 121 + ], + "spans": [ + { + "bbox": [ + 139, + 103, + 455, + 121 + ], + "type": "text", + "content": "ORCa: Glossy Objects as Radiance-Field Cameras" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 124, + 142, + 468, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 142, + 468, + 171 + ], + "spans": [ + { + "bbox": [ + 124, + 142, + 468, + 171 + ], + "type": "text", + "content": "Kushagra Tiwary" + }, + { + "bbox": [ + 124, + 142, + 468, + 171 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 124, + 142, + 468, + 171 + ], + "type": "text", + "content": ", Akshit Dave" + }, + { + "bbox": [ + 124, + 142, + 468, + 171 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 124, + 142, + 468, + 171 + ], + "type": "text", + "content": ", Nikhil Behari" + }, + { + "bbox": [ + 124, + 142, + 468, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 124, + 142, + 468, + 171 + ], + "type": "text", + "content": ", Tzofi Klinghoffer" + }, + { + "bbox": [ + 124, + 142, + 468, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 124, + 142, + 468, + 171 + ], + "type": "text", + "content": ", Ashok Veeraraghavan" + }, + { + "bbox": [ + 124, + 142, + 468, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 124, + 142, + 468, + 171 + ], + "type": "text", + "content": ", Ramesh Raskar" + }, + { + "bbox": [ + 124, + 142, + 468, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 159, + 171, + 434, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 171, + 434, + 186 + ], + "spans": [ + { + "bbox": [ + 159, + 171, + 434, + 186 + ], + "type": "text", + "content": "1Massachusetts Institute of Technology, 2Rice University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 126, + 187, + 465, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 187, + 465, + 200 + ], + "spans": [ + { + "bbox": [ + 126, + 187, + 465, + 200 + ], + "type": "text", + "content": "{ktiwary,behari,tzofi,raskar}@mit.edu，{akshat,vashok}@rice.edu" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 72, + 247, + 194, + 395 + ], + "blocks": [ + { + "bbox": [ + 102, + 228, + 162, + 246 + ], + "lines": [ + { + "bbox": [ + 102, + 228, + 162, + 246 + ], + "spans": [ + { + "bbox": [ + 102, + 228, + 162, + 246 + ], + "type": "text", + "content": "(a) Scene: Glossy Elephant in a room" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 72, + 247, + 194, + 395 + ], + "lines": [ + { + "bbox": [ + 72, + 247, + 194, + 395 + ], + "spans": [ + { + "bbox": [ + 72, + 247, + 194, + 395 + ], + "type": "image", + "image_path": "2df5ce9a07841845a9d542726a640bb85b9e6d9291e58e9478efb66eb984ffe0.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 195, + 247, + 275, + 322 + ], + "blocks": [ + { + "bbox": [ + 198, + 228, + 271, + 246 + ], + "lines": [ + { + "bbox": [ + 198, + 228, + 271, + 246 + ], + "spans": [ + { + "bbox": [ + 198, + 228, + 271, + 246 + ], + "type": "text", + "content": "(b) Input: images with distorted reflections" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 195, + 247, + 275, + 322 + ], + "lines": [ + { + "bbox": [ + 195, + 247, + 275, + 322 + ], + "spans": [ + { + "bbox": [ + 195, + 247, + 275, + 322 + ], + "type": "image", + "image_path": "dc19608bbd6e867f5468407562bad93e13addd859ab7ff6236d1e71d79f9430b.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 195, + 322, + 274, + 395 + ], + "blocks": [ + { + "bbox": [ + 195, + 322, + 274, + 395 + ], + "lines": [ + { + "bbox": [ + 195, + 322, + 274, + 395 + ], + "spans": [ + { + "bbox": [ + 195, + 322, + 274, + 395 + ], + "type": "image", + "image_path": "31e4fd0b5ff2c092526cf8c945975aff577b819c566fc32bc84692aab0160134.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 403, + 547, + 448 + ], + "lines": [ + { + "bbox": [ + 46, + 403, + 547, + 448 + ], + "spans": [ + { + "bbox": [ + 46, + 403, + 547, + 448 + ], + "type": "text", + "content": "Figure 1. Objects as radiance-field cameras. We convert everyday objects with unknown geometry (a) into radiance-field cameras by modeling multi-view reflections (b) as projections of the 5D radiance field of the environment. We convert the object surface into a virtual sensor to capture this radiance field (c), which enables depth and radiance estimation of the surrounding environment. We can then query this radiance field to perform beyond field-of-view novel view synthesis of the environment (d)." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 279, + 312, + 315, + 329 + ], + "blocks": [ + { + "bbox": [ + 279, + 312, + 315, + 329 + ], + "lines": [ + { + "bbox": [ + 279, + 312, + 315, + 329 + ], + "spans": [ + { + "bbox": [ + 279, + 312, + 315, + 329 + ], + "type": "image", + "image_path": "319140cdbea5bb5445735e9ea44a9cd49aae383dd620fbddcb858941bfcc8e1d.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 318, + 247, + 440, + 395 + ], + "blocks": [ + { + "bbox": [ + 329, + 228, + 429, + 246 + ], + "lines": [ + { + "bbox": [ + 329, + 228, + 429, + 246 + ], + "spans": [ + { + "bbox": [ + 329, + 228, + 429, + 246 + ], + "type": "text", + "content": "(c) Using objects to image beyond camera's field-of-view" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 318, + 247, + 440, + 395 + ], + "lines": [ + { + "bbox": [ + 318, + 247, + 440, + 395 + ], + "spans": [ + { + "bbox": [ + 318, + 247, + 440, + 395 + ], + "type": "image", + "image_path": "4e21bd1d0ee90f2978ca334273eca7b705f28da5a05562d24c5f88cc947fccc9.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 441, + 247, + 520, + 321 + ], + "blocks": [ + { + "bbox": [ + 449, + 228, + 511, + 246 + ], + "lines": [ + { + "bbox": [ + 449, + 228, + 511, + 246 + ], + "spans": [ + { + "bbox": [ + 449, + 228, + 511, + 246 + ], + "type": "text", + "content": "(d) Recovered and Undistorted RGBs" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 441, + 247, + 520, + 321 + ], + "lines": [ + { + "bbox": [ + 441, + 247, + 520, + 321 + ], + "spans": [ + { + "bbox": [ + 441, + 247, + 520, + 321 + ], + "type": "image", + "image_path": "a2b083707994440a1e0da6eb06ee8d40808b8bdec0bcb43e6e6c39ed27b00c62.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 441, + 322, + 520, + 395 + ], + "blocks": [ + { + "bbox": [ + 441, + 322, + 520, + 395 + ], + "lines": [ + { + "bbox": [ + 441, + 322, + 520, + 395 + ], + "spans": [ + { + "bbox": [ + 441, + 322, + 520, + 395 + ], + "type": "image", + "image_path": "7fae400df005cfb7ceba6e51ff03cb387e2efef519e097a20a61fd25e9dea927.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 143, + 457, + 192, + 470 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 457, + 192, + 470 + ], + "spans": [ + { + "bbox": [ + 143, + 457, + 192, + 470 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 45, + 484, + 290, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 484, + 290, + 689 + ], + "spans": [ + { + "bbox": [ + 45, + 484, + 290, + 689 + ], + "type": "text", + "content": "Reflections on glossy objects contain valuable and hidden information about the surrounding environment. By converting these objects into cameras, we can unlock exciting applications, including imaging beyond the camera's field-of-view and from seemingly impossible vantage points, e.g. from reflections on the human eye. However, this task is challenging because reflections depend jointly on object geometry, material properties, the 3D environment, and the observer's viewing direction. Our approach converts glossy objects with unknown geometry into radiance-field cameras to image the world from the object's perspective. Our key insight is to convert the object surface into a virtual sensor that captures cast reflections as a 2D projection of the 5D environment radiance field visible to and surrounding the object. We show that recovering the environment radiance fields enables depth and radiance estimation from the object to its surroundings in addition to beyond field-of-view" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 459, + 545, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 459, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 306, + 459, + 545, + 555 + ], + "type": "text", + "content": "novel-view synthesis, i.e. rendering of novel views that are only directly visible to the glossy object present in the scene, but not the observer. Moreover, using the radiance field we can image around occluders caused by close-by objects in the scene. Our method is trained end-to-end on multi-view images of the object and jointly estimates object geometry, diffuse radiance, and the 5D environment radiance field. For more information, visit our website." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 562, + 387, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 562, + 387, + 574 + ], + "spans": [ + { + "bbox": [ + 306, + 562, + 387, + 574 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": "Imagine that you're driving down a city street that is packed with lines of parked cars on both sides. Inspection of the cars' glass windshields, glossy paint, and plastic reveal sharp, but faint and distorted views of the surroundings that might be otherwise hidden from you. Humans can infer depth and semantic cues about the occluded areas in the environment by processing reflections visible on reflective objects, internally decomposing the object geometry and radiance from the specular radiance being reflected onto it. Our aim is to decompose the object from its reflections to \"see\" the world from the object's perspective, effectively turning" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 59, + 702, + 124, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 702, + 124, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 702, + 124, + 712 + ], + "type": "text", + "content": "* Equal contribution" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20773" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 95, + 193, + 219 + ], + "blocks": [ + { + "bbox": [ + 62, + 73, + 201, + 83 + ], + "lines": [ + { + "bbox": [ + 62, + 73, + 201, + 83 + ], + "spans": [ + { + "bbox": [ + 62, + 73, + 201, + 83 + ], + "type": "text", + "content": "(a) Learning Neural Implicit Surfaces [Sec. 3.2]" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 95, + 193, + 219 + ], + "lines": [ + { + "bbox": [ + 55, + 95, + 193, + 219 + ], + "spans": [ + { + "bbox": [ + 55, + 95, + 193, + 219 + ], + "type": "image", + "image_path": "336fad8acd27ed663001f77ba70be48d9fce30a1b40b4f2c16201dee72c312f1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 225, + 92, + 281, + 99 + ], + "lines": [ + { + "bbox": [ + 225, + 92, + 281, + 99 + ], + "spans": [ + { + "bbox": [ + 225, + 92, + 281, + 99 + ], + "type": "text", + "content": "Real camera origin" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 223, + 134, + 362, + 220 + ], + "blocks": [ + { + "bbox": [ + 227, + 101, + 273, + 109 + ], + "lines": [ + { + "bbox": [ + 227, + 101, + 273, + 109 + ], + "spans": [ + { + "bbox": [ + 227, + 101, + 273, + 109 + ], + "type": "text", + "content": "Real pixel cone" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 228, + 111, + 282, + 118 + ], + "lines": [ + { + "bbox": [ + 228, + 111, + 282, + 118 + ], + "spans": [ + { + "bbox": [ + 228, + 111, + 282, + 118 + ], + "type": "text", + "content": "- Surface intersection" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 228, + 121, + 277, + 129 + ], + "lines": [ + { + "bbox": [ + 228, + 121, + 277, + 129 + ], + "spans": [ + { + "bbox": [ + 228, + 121, + 277, + 129 + ], + "type": "text", + "content": "Virtual pixel cone" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 228, + 130, + 286, + 138 + ], + "lines": [ + { + "bbox": [ + 228, + 130, + 286, + 138 + ], + "spans": [ + { + "bbox": [ + 228, + 130, + 286, + 138 + ], + "type": "text", + "content": "Virtual camera origin" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 223, + 134, + 362, + 220 + ], + "lines": [ + { + "bbox": [ + 223, + 134, + 362, + 220 + ], + "spans": [ + { + "bbox": [ + 223, + 134, + 362, + 220 + ], + "type": "image", + "image_path": "31c4ee92e85bdedbf20674e064bab88310bde6e2ece202491963e1b2013d9540.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 233, + 547, + 289 + ], + "lines": [ + { + "bbox": [ + 46, + 233, + 547, + 289 + ], + "spans": [ + { + "bbox": [ + 46, + 233, + 547, + 289 + ], + "type": "text", + "content": "Figure 2. ORCa Overview. We jointly estimate the object's geometry and diffuse along with the environment radiance field estimation through a three-step approach. First, we model the object as a neural implicit surface (a). We model the reflections as probing the environment on virtual viewpoints (b) estimated analytically from surface properties. We model the environment as a radiance field queried on these viewpoints (c). Both neural implicit surface and environment radiance field are trained jointly on multi-view images of the object using a photometric loss." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 380, + 99, + 434, + 155 + ], + "blocks": [ + { + "bbox": [ + 233, + 73, + 370, + 83 + ], + "lines": [ + { + "bbox": [ + 233, + 73, + 370, + 83 + ], + "spans": [ + { + "bbox": [ + 233, + 73, + 370, + 83 + ], + "type": "text", + "content": "(b) Objects Surface as Virtual Sensor [Sec. 3.3]" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 380, + 99, + 434, + 155 + ], + "lines": [ + { + "bbox": [ + 380, + 99, + 434, + 155 + ], + "spans": [ + { + "bbox": [ + 380, + 99, + 434, + 155 + ], + "type": "image", + "image_path": "351c007cb16eae15b70ea2b301d7329906fe4afd514b299eee6a2d2089679df2.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 380, + 156, + 431, + 215 + ], + "blocks": [ + { + "bbox": [ + 380, + 156, + 431, + 215 + ], + "lines": [ + { + "bbox": [ + 380, + 156, + 431, + 215 + ], + "spans": [ + { + "bbox": [ + 380, + 156, + 431, + 215 + ], + "type": "image", + "image_path": "f957743e124559cdadc315e2b75882f4c4ed9a059097d2b671298f01d58da310.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 439, + 141, + 484, + 205 + ], + "blocks": [ + { + "bbox": [ + 439, + 141, + 484, + 205 + ], + "lines": [ + { + "bbox": [ + 439, + 141, + 484, + 205 + ], + "spans": [ + { + "bbox": [ + 439, + 141, + 484, + 205 + ], + "type": "image", + "image_path": "79e2e5a652c3fc556155146662f212966fa57dd0870b932f09a0ce16f5fea170.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 491, + 105, + 536, + 151 + ], + "blocks": [ + { + "bbox": [ + 384, + 73, + 538, + 83 + ], + "lines": [ + { + "bbox": [ + 384, + 73, + 538, + 83 + ], + "spans": [ + { + "bbox": [ + 384, + 73, + 538, + 83 + ], + "type": "text", + "content": "(c) Learning Environment Radiance Fields [Sec. 3.4]" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 491, + 105, + 536, + 151 + ], + "lines": [ + { + "bbox": [ + 491, + 105, + 536, + 151 + ], + "spans": [ + { + "bbox": [ + 491, + 105, + 536, + 151 + ], + "type": "image", + "image_path": "779145651c7b8316abd2c83fd2648e6f16c1d6536a4200c71278aab7c12e34d3.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 491, + 161, + 544, + 205 + ], + "blocks": [ + { + "bbox": [ + 498, + 152, + 529, + 159 + ], + "lines": [ + { + "bbox": [ + 498, + 152, + 529, + 159 + ], + "spans": [ + { + "bbox": [ + 498, + 152, + 529, + 159 + ], + "type": "text", + "content": "Virtual View" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 491, + 161, + 544, + 205 + ], + "lines": [ + { + "bbox": [ + 491, + 161, + 544, + 205 + ], + "spans": [ + { + "bbox": [ + 491, + 161, + 544, + 205 + ], + "type": "image", + "image_path": "05cf6acc5514021a0e70d0ad07726823cb7affe5b4c7e89d84572cc68281700e.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 496, + 207, + 530, + 216 + ], + "lines": [ + { + "bbox": [ + 496, + 207, + 530, + 216 + ], + "spans": [ + { + "bbox": [ + 496, + 207, + 530, + 216 + ], + "type": "text", + "content": "Virtual Depth" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 310, + 288, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 310, + 288, + 369 + ], + "spans": [ + { + "bbox": [ + 46, + 310, + 288, + 369 + ], + "type": "text", + "content": "the object into a camera that images its environment. However, reflections pose a long-standing challenge in computer vision as the reflections are a 2D projection of an unknown 3D environment that is distorted based on the shape of the reflector." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 46, + 374, + 287, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 374, + 287, + 589 + ], + "spans": [ + { + "bbox": [ + 46, + 374, + 287, + 589 + ], + "type": "text", + "content": "To capture the 3D world from the object's perspective, we model the object's surface as a virtual sensor that captures the 2D projection of a 5D environment radiance field surrounding the object. This environment radiance field consists largely of areas only visible to the observer through the object's reflections. Our use of environment radiance fields not only enables depth and radiance estimation from the object to its surroundings but also enables beyond field-of-view novel-view synthesis, i.e. rendering of novel views that are only directly visible to the glossy object present in the scene but not the observer. Unlike conventional approaches that model the environment as a 2D map, our approach models it as a 5D field without assuming the scene is infinitely far away. Moreover, by sampling the 5D radiance field, instead of a 2D map, we can capture depth and images around occluders, such as close-by objects in the scene, as shown in Fig. 3. These applications cannot be done from a 2D environment map." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": "We aim to decompose reflections on the object's surface, from its surface and exploit those reflections to construct a radiance field surrounding the object, therefore capturing the 3D world in the process. This is a challenging task because the reflections are extremely sensitive to local object geometry, viewing direction and inter-reflections due to the object's surface. To capture this radiance field, we convert glossy objects with unknown geometry and texture into radiance-field cameras. Specifically, we exploit neural rendering to estimate the local surface of the object viewed" + } + ] + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 315, + 319, + 375, + 409 + ], + "blocks": [ + { + "bbox": [ + 312, + 308, + 375, + 318 + ], + "lines": [ + { + "bbox": [ + 312, + 308, + 375, + 318 + ], + "spans": [ + { + "bbox": [ + 312, + 308, + 375, + 318 + ], + "type": "text", + "content": "(a) Sample Input Views" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 315, + 319, + 375, + 409 + ], + "lines": [ + { + "bbox": [ + 315, + 319, + 375, + 409 + ], + "spans": [ + { + "bbox": [ + 315, + 319, + 375, + 409 + ], + "type": "image", + "image_path": "aeb4110451eb9505f87a500f1fe33aaa68a383a804374a2cba6aee5e98fd4d1e.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 377, + 318, + 480, + 410 + ], + "blocks": [ + { + "bbox": [ + 383, + 308, + 474, + 317 + ], + "lines": [ + { + "bbox": [ + 383, + 308, + 474, + 317 + ], + "spans": [ + { + "bbox": [ + 383, + 308, + 474, + 317 + ], + "type": "text", + "content": "(b) Virtual Views with Translation" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 377, + 318, + 480, + 410 + ], + "lines": [ + { + "bbox": [ + 377, + 318, + 480, + 410 + ], + "spans": [ + { + "bbox": [ + 377, + 318, + 480, + 410 + ], + "type": "image", + "image_path": "fbafe69d37013587b525762efddc55cca4a5bba27ead0abc2b7e3edc12c30e06.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 479, + 309, + 539, + 317 + ], + "lines": [ + { + "bbox": [ + 479, + 309, + 539, + 317 + ], + "spans": [ + { + "bbox": [ + 479, + 309, + 539, + 317 + ], + "type": "text", + "content": "(c) Virtual View Depth" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 481, + 318, + 535, + 409 + ], + "blocks": [ + { + "bbox": [ + 481, + 318, + 535, + 409 + ], + "lines": [ + { + "bbox": [ + 481, + 318, + 535, + 409 + ], + "spans": [ + { + "bbox": [ + 481, + 318, + 535, + 409 + ], + "type": "image", + "image_path": "7f89207a5e51e11d24714ceb23ad8cae15edef672cca9e9b3768842a48d12ba6.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 314, + 421, + 389, + 481 + ], + "blocks": [ + { + "bbox": [ + 369, + 413, + 488, + 420 + ], + "lines": [ + { + "bbox": [ + 369, + 413, + 488, + 420 + ], + "spans": [ + { + "bbox": [ + 369, + 413, + 488, + 420 + ], + "type": "text", + "content": "(d) Parallax effect in 5D Env. Radiance Field" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 314, + 421, + 389, + 481 + ], + "lines": [ + { + "bbox": [ + 314, + 421, + 389, + 481 + ], + "spans": [ + { + "bbox": [ + 314, + 421, + 389, + 481 + ], + "type": "image", + "image_path": "22bcbfb95f7bdc1a496880cc6aee049956973e1d3b15df3cab447d26c0ae733e.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 394, + 421, + 461, + 480 + ], + "blocks": [ + { + "bbox": [ + 394, + 421, + 461, + 480 + ], + "lines": [ + { + "bbox": [ + 394, + 421, + 461, + 480 + ], + "spans": [ + { + "bbox": [ + 394, + 421, + 461, + 480 + ], + "type": "image", + "image_path": "b20d6b50a427e566c9f341272c5fec2348027903e9276308a271ec6b2683132c.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 467, + 421, + 534, + 480 + ], + "blocks": [ + { + "bbox": [ + 467, + 421, + 534, + 480 + ], + "lines": [ + { + "bbox": [ + 467, + 421, + 534, + 480 + ], + "spans": [ + { + "bbox": [ + 467, + 421, + 534, + 480 + ], + "type": "image", + "image_path": "7f6a08274bc3335ececbd90b34197b8e3717f76675954522b74248b0ef7f7f0f.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 314, + 482, + 389, + 541 + ], + "blocks": [ + { + "bbox": [ + 314, + 482, + 389, + 541 + ], + "lines": [ + { + "bbox": [ + 314, + 482, + 389, + 541 + ], + "spans": [ + { + "bbox": [ + 314, + 482, + 389, + 541 + ], + "type": "image", + "image_path": "5ed18484080d02f54b9c078dca844f100f84775f461a893f9e4ababb07ee2d4a.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 316, + 542, + 358, + 549 + ], + "lines": [ + { + "bbox": [ + 316, + 542, + 358, + 549 + ], + "spans": [ + { + "bbox": [ + 316, + 542, + 358, + 549 + ], + "type": "text", + "content": "x 0" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 561, + 545, + 639 + ], + "lines": [ + { + "bbox": [ + 306, + 561, + 545, + 639 + ], + "spans": [ + { + "bbox": [ + 306, + 561, + 545, + 639 + ], + "type": "text", + "content": "Figure 3. Advantages of 5D environment radiance field. Modeling reflections on object surfaces (a) as a 5D env. radiance field enables beyond field-of-view novel-view synthesis, including rendering of the environment from translated virtual camera views (b). Depth (c) and environment radiance of translated and parallax views can further enable imaging behind occluders, for example revealing the tails behind the primary Pokemon occluders (d)." + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 394, + 482, + 461, + 540 + ], + "blocks": [ + { + "bbox": [ + 394, + 482, + 461, + 540 + ], + "lines": [ + { + "bbox": [ + 394, + 482, + 461, + 540 + ], + "spans": [ + { + "bbox": [ + 394, + 482, + 461, + 540 + ], + "type": "image", + "image_path": "bb1091cb84235be474a23582eacd3ddda3c203b1817093900e6a0022c52c262d.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 420, + 542, + 435, + 549 + ], + "lines": [ + { + "bbox": [ + 420, + 542, + 435, + 549 + ], + "spans": [ + { + "bbox": [ + 420, + 542, + 435, + 549 + ], + "type": "text", + "content": "+0.25" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 467, + 482, + 534, + 540 + ], + "blocks": [ + { + "bbox": [ + 467, + 482, + 534, + 540 + ], + "lines": [ + { + "bbox": [ + 467, + 482, + 534, + 540 + ], + "spans": [ + { + "bbox": [ + 467, + 482, + 534, + 540 + ], + "type": "image", + "image_path": "67c7c9425f8b662d3be565cdae4237350d080244862d3ecf337dc718ab2cd1e6.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 493, + 542, + 509, + 549 + ], + "lines": [ + { + "bbox": [ + 493, + 542, + 509, + 549 + ], + "spans": [ + { + "bbox": [ + 493, + 542, + 509, + 549 + ], + "type": "text", + "content": "+0.50" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + } + ], + "index": 34 + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": "from each pixel of the real camera. We then convert this local surface into a virtual pixel that captures radiance from the environment. This virtual pixel captures the environment radiance as shown in Fig 5. We estimate the outgo" + } + ] + } + ], + "index": 39 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20774" + } + ] + } + ], + "index": 40 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "text", + "content": "ing frustum from the virtual pixel as a cone that samples the scene. By sampling the scene from many virtual pixels on the object surface, we construct an environment radiance field that can be queried independently of the object surface, enabling beyond field-of-view novel-view synthesis from previously unsampled viewpoints." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 144, + 288, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 144, + 288, + 310 + ], + "spans": [ + { + "bbox": [ + 46, + 144, + 288, + 310 + ], + "type": "text", + "content": "Our approach jointly estimates object geometry, diffuse radiance, and the environment radiance field from multiview images of glossy objects with unknown geometry and diffuse texture in three steps. First, we use neural signed distance functions (SDF) and an MLP to model the glossy object's geometry as a neural implicit surface and diffuse radiance, respectively, similar to PANDORA [10]. Then, for every pixel on the observer's camera, we estimate the virtual pixels on the object's surface based on the estimated local geometry from the neural SDF. We analytically compute the parameters of the virtual cone through the virtual pixel. Lastly, we use the cone formulation in MipNeRF [5] to cast virtual cones from the virtual camera to recover the environment radiance." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 59, + 312, + 272, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 312, + 272, + 323 + ], + "spans": [ + { + "bbox": [ + 59, + 312, + 272, + 323 + ], + "type": "text", + "content": "To summarize, we make the following contributions:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 327, + 286, + 459 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 59, + 327, + 286, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 327, + 286, + 363 + ], + "spans": [ + { + "bbox": [ + 59, + 327, + 286, + 363 + ], + "type": "text", + "content": "- We present a method to convert implicit surfaces into virtual sensors that can image their surroundings using virtual cones. (Sec. 3.3)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 369, + 286, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 369, + 286, + 405 + ], + "spans": [ + { + "bbox": [ + 59, + 369, + 286, + 405 + ], + "type": "text", + "content": "- We jointly estimate object geometry, diffuse radiance, and estimate the 5D environment radiance field surrounding the object. (Fig. 7 & 9)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 411, + 286, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 411, + 286, + 459 + ], + "spans": [ + { + "bbox": [ + 59, + 411, + 286, + 459 + ], + "type": "text", + "content": "- We show that the environment radiance field can be queried to perform beyond-field-of-view novel viewpoint synthesis, i.e. render views only visible to the object in the scene (Section 3.4)" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 46, + 463, + 287, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 463, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 46, + 463, + 287, + 643 + ], + "type": "text", + "content": "Scope. We only model glossy objects with low roughness as such specular reflections tend to have a high signal-to-noise ratio, therefore, are a sharper estimate of the environment radiance field. However, we note that the virtual cone computation can be extended to model the cone radius as a function of surface roughness. Deblurring approaches can further improve the resolution of estimated environment. In addition, we approximate the local curvature using mean curvature, which fails for objects with varying radius of curvature along the tangent space. We explain how our virtual cone curvature estimation can be extended to handle general shape operators in the supplementary material. Lastly, similar to other multi-view approaches, our approach relies on a sufficient virtual baseline between virtual viewpoints to recover the environment radiance field." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 651, + 134, + 662 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 651, + 134, + 662 + ], + "spans": [ + { + "bbox": [ + 47, + 651, + 134, + 662 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 671, + 165, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 165, + 684 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 165, + 684 + ], + "type": "text", + "content": "2.1. Modeling reflections" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 713 + ], + "type": "text", + "content": "Catadioptric imaging systems incorporate reflections on curved reflective mirrors to expand the field of view of con" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 179 + ], + "type": "text", + "content": "ventional cameras [2, 26], to increase the baseline of light field cameras [11, 40] and to perform novel view synthesis from a single capture [44]. These works assume the geometry of the reflecting surface is known or calibrated while we create a catadioptric imaging system from everyday glossy objects of unknown geometry. Grossberg et al. [15] propose a generalized model for light transport through imaging systems, including catadioptric systems. Our work focuses on light transport reflecting off a general object in the scene." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 181, + 545, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 181, + 545, + 324 + ], + "spans": [ + { + "bbox": [ + 304, + 181, + 545, + 324 + ], + "type": "text", + "content": "Light field imaging is shown to be effective for reflection removal [21, 27], reconstructing specular surfaces [18], intrinsic decomposition [1], and neural rendering [35]. These works typically consider planar reflections or require training on synthetic datasets, while our approach models reflections on complex geometry and is unsupervised. Prior works have also utilized additional light properties such as polarization [9, 10, 20] and time-of-flight [17, 33] for the separation of the reflected component and specular surface reconstruction. While the input of our approach is RGB images, there is scope for improving reconstruction quality by supplementing the algorithm with these additional cues." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 325, + 545, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 325, + 545, + 517 + ], + "spans": [ + { + "bbox": [ + 304, + 325, + 545, + 517 + ], + "type": "text", + "content": "Recent progress in neural radiance fields has enabled impressive novel view rendering and geometry reconstruction from multi-view images [24]. MipNeRF [5] demonstrates better novel view synthesis by modeling outgoing rays as cones to enable anti-aliasing. RefNeRF [41] proposes Integrated Directional Embeddings for improved novel view synthesis of reflections. NeRFRen [16] separates diffuse and specular radiance by using separate neural networks. Neural Catacaustics [19] propose a neural warping method to model reflections by learning the caustics of the surface. While these works focus on novel-view synthesis of the scene from the primary camera, we perform view synthesis that is beyond the line-of-sight of the primary camera, i.e., rendering views only visible to the objects present in the scene, while jointly estimating object geometry and separating diffuse and specular radiance." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 527, + 443, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 527, + 443, + 538 + ], + "spans": [ + { + "bbox": [ + 306, + 527, + 443, + 538 + ], + "type": "text", + "content": "2.2. Environment Estimation" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 545, + 713 + ], + "type": "text", + "content": "Recovering underlying scene properties from multiple images is inherently ill-posed [32], but can be regularized using the natural statistics of scene properties as a prior [4, 34]. Recent works exploit this prior through deep neural networks and demonstrate inverse rendering of indoor scenes from a single image [12, 22, 43, 48]. However, these techniques typically recover only coarse representations of lighting and cannot reconstruct fine details of the environment. Lombardi et al. [23] recover environment and reflectance, assuming the scene is composed of known geometry and uniform material. Georgolis et al. [13] recover the environment map behind the camera from a single image of a glossy object, assuming the object is composed of textureless materials and using ground truth seg" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20775" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 70, + 288, + 188 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 288, + 188 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 288, + 188 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 288, + 188 + ], + "type": "table", + "html": "
ApproachInputScene GeometryAlgorithm TypeEnvironment Dimension
Neural IllumSingle RGBNormalsSupervised2D
LighthouseStereo RGB pairMulti-plane ImageSupervised5D
NeRFactorMulti-view RGBVolumetric AlbedoSemi-supervised2D
RefNeRFMulti-view RGBVolumetric AlbedoSelf-supervised2D
PANDORAMulti-view RGB+Pol.Neural SDFSelf-supervised2D
ORCa (Ours)Multi-view RGBNeural SDFSelf-supervised5D
", + "image_path": "47a18be627fa2f9c56ccaf7a56370b076e9d90c9a3ed8e66f6c506b9a0667948.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 316, + 289, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 316, + 289, + 485 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 289, + 485 + ], + "type": "text", + "content": "mentation masks. Song et al. [36] estimate plausible environment maps by mapping reflections in the image and inpainting unmapped regions. Srinivasan et al. [37] capture stereo image pairs and estimate plausible spatially-coherent environment maps. NeRD [7], NeRFactor [47] and NeuralPIL [8] employ data-driven priors for lighting and BRDF in a NeRF-based approach for radiance decomposition. Park et al. [31] use RGB-D videos to estimate environment map. Swedish et al. [39] recover high-frequency illumination map from the shadows of an object with known geometry. PhySG [46] and Munkberg et al. [25] perform inverse rendering from multi-view images by modeling the surface as signed distance functions. PANDORA [10] performs radiance decomposition from polarized RGB images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 506, + 287, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 506, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 47, + 506, + 287, + 533 + ], + "type": "text", + "content": "3. Learning environment radiance fields from multi-view reflections" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 540, + 115, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 540, + 115, + 551 + ], + "spans": [ + { + "bbox": [ + 47, + 540, + 115, + 551 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "content": "Reflections on glossy objects offer a glimpse into the surrounding environment beyond the camera's field of view. From multi-view images of a glossy object with unknown geometry and albedo, we aim to recover the 5D radiance field of the surrounding environment. The mapping from images captured by the observer to the surrounding environment depends on the glossy object's surface properties, in particular, the surface normals and curvature. We first cast a cone from the observer camera's center-of-projection through each pixel viewing the scene. When the cone intersects the object's surface, it reflects, causing the cone to be transformed (Fig. 5). The transformed cone, referred to as a virtual cone, samples the environment and is primar-" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 326, + 70, + 529, + 222 + ], + "blocks": [ + { + "bbox": [ + 46, + 197, + 287, + 294 + ], + "lines": [ + { + "bbox": [ + 46, + 197, + 287, + 294 + ], + "spans": [ + { + "bbox": [ + 46, + 197, + 287, + 294 + ], + "type": "text", + "content": "Figure 4. Comparison of environment estimation approaches. Approaches such Neural Illum [36], Lighthouse [37], NeRFactor [47] train on datasets of natural illumination maps to regularize ill-posedness of environment estimation. PANDORA [10] and RefNeRF [41] exploit multi-view reflections on objects but assume the surrounding environment is infinitely far away and is modeled with a flat 2D map. From multi-view reflections, ORCa converts the object surface into a virtual sensor and extracts the 5D radiance field of the environment." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 326, + 70, + 529, + 222 + ], + "lines": [ + { + "bbox": [ + 326, + 70, + 529, + 222 + ], + "spans": [ + { + "bbox": [ + 326, + 70, + 529, + 222 + ], + "type": "image", + "image_path": "779df1b504cb234de996732bbc934595f060771d044a45fbd83b5e114802095c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 228, + 547, + 308 + ], + "lines": [ + { + "bbox": [ + 304, + 228, + 547, + 308 + ], + "spans": [ + { + "bbox": [ + 304, + 228, + 547, + 308 + ], + "type": "text", + "content": "Figure 5. Virtual Sensor. We image the world through the object by modeling each pixel's specular radiance as a projection of the 5D radiance field of the environment onto the object's surface. We capture the radiance field by treating the surface area on the object that the pixel views, " + }, + { + "bbox": [ + 304, + 228, + 547, + 308 + ], + "type": "inline_equation", + "content": "\\mathbf{dS}_t" + }, + { + "bbox": [ + 304, + 228, + 547, + 308 + ], + "type": "text", + "content": ", as a single-pixel virtual camera with its center-of-projection at " + }, + { + "bbox": [ + 304, + 228, + 547, + 308 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_o" + }, + { + "bbox": [ + 304, + 228, + 547, + 308 + ], + "type": "text", + "content": ". We cast virtual cones through the virtual sensor to capture the 5D radiance field of the environment." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 327, + 545, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 327, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 304, + 327, + 545, + 422 + ], + "type": "text", + "content": "ily responsible for the specular radiance observed on the glossy object. Our key insight is that the reflections captured by the observer's camera can be modeled as a projection of the environment radiance field onto the object's surface. By modeling the reflected rays as a cone and computing the parameters of the cone, we can more accurately estimate the projected environment radiance field onto the object surface, as shown in Fig. 10." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 423, + 546, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 423, + 546, + 543 + ], + "spans": [ + { + "bbox": [ + 304, + 423, + 546, + 543 + ], + "type": "text", + "content": "ORCa is composed of three steps: modeling the object's geometry as a neural implicit surface (Sec. 3.2), converting the object's surface into a virtual sensor (Sec. 3.3), and modeling the environment radiance field as a projection along these virtual cones (Sec. 3.4). The learned environment radiance field can then be queried on novel viewpoints to show occluded areas in the scene. Fig. 2 depicts our output for each component on a scene rendered with a complex glossy object and 3D environment. Fig. 6 shows our system architecture. Next, we describe each step in detail." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 550, + 489, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 550, + 489, + 563 + ], + "spans": [ + { + "bbox": [ + 306, + 550, + 489, + 563 + ], + "type": "text", + "content": "3.2. Learning Neural implicit Surfaces" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 568, + 547, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 568, + 547, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 568, + 547, + 689 + ], + "type": "text", + "content": "Neural Signed Distance Function We model the object geometry as a neural signed distance function (SDF). " + }, + { + "bbox": [ + 304, + 568, + 547, + 689 + ], + "type": "inline_equation", + "content": "f: \\mathbb{R}^3 \\to \\mathbb{R}" + }, + { + "bbox": [ + 304, + 568, + 547, + 689 + ], + "type": "text", + "content": ". SDFs provide a helpful inductive bias for learning smooth surface geometry [30, 42, 45] that assists downstream tasks in our pipeline. Moreover, the surface properties crucial for our framework, surface normals and curvature, can be conveniently computed from SDFs in a differentiable manner. Consider the 3D spatial coordinates, " + }, + { + "bbox": [ + 304, + 568, + 547, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 568, + 547, + 689 + ], + "type": "text", + "content": ", in the scene. The glossy object surface, " + }, + { + "bbox": [ + 304, + 568, + 547, + 689 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 568, + 547, + 689 + ], + "type": "text", + "content": " is then represented by the zero-level set of the SDF" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 370, + 696, + 545, + 711 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 696, + 545, + 711 + ], + "spans": [ + { + "bbox": [ + 370, + 696, + 545, + 711 + ], + "type": "interline_equation", + "content": "\\mathcal {S} = \\left\\{f _ {\\mathcal {S}} (\\mathbf {x}) = 0 | \\mathbf {x} \\in \\mathbb {R} ^ {3} \\right\\} \\tag {1}", + "image_path": "94fb121b997c504be0898264972e7e772b6f372b3c804de624e987223322a271.jpg" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20776" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 282, + 129 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 282, + 129 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 282, + 129 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 282, + 129 + ], + "type": "image", + "image_path": "b38a38c708192bc41253a0ce6fa511bf54b2beedb6b028f437d397b52b2721d8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 72, + 137, + 261, + 148 + ], + "lines": [ + { + "bbox": [ + 72, + 137, + 261, + 148 + ], + "spans": [ + { + "bbox": [ + 72, + 137, + 261, + 148 + ], + "type": "text", + "content": "Figure 6. Overview of our proposed architecture." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 168, + 287, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 168, + 287, + 190 + ], + "spans": [ + { + "bbox": [ + 47, + 168, + 287, + 190 + ], + "type": "text", + "content": "Similar to Yariv et al. [45], we model the SDF " + }, + { + "bbox": [ + 47, + 168, + 287, + 190 + ], + "type": "inline_equation", + "content": "f_{S}" + }, + { + "bbox": [ + 47, + 168, + 287, + 190 + ], + "type": "text", + "content": " as a coordinate-based MLP." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 193, + 287, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 193, + 287, + 216 + ], + "spans": [ + { + "bbox": [ + 47, + 193, + 287, + 216 + ], + "type": "text", + "content": "Surface Normals Gradients of the SDF at the zero level set point " + }, + { + "bbox": [ + 47, + 193, + 287, + 216 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 47, + 193, + 287, + 216 + ], + "type": "text", + "content": " towards the surface normals " + }, + { + "bbox": [ + 47, + 193, + 287, + 216 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 47, + 193, + 287, + 216 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 223, + 287, + 249 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 223, + 287, + 249 + ], + "spans": [ + { + "bbox": [ + 106, + 223, + 287, + 249 + ], + "type": "interline_equation", + "content": "\\mathbf {n} (\\mathbf {x}) = \\frac {\\nabla_ {\\mathbf {x}} f _ {\\mathcal {S}} (\\mathbf {x})}{\\| \\nabla_ {\\mathbf {x}} f _ {\\mathcal {S}} (\\mathbf {x}) \\|} \\quad \\mathbf {x} \\in \\mathcal {S} \\tag {2}", + "image_path": "901094170f7b5a040d003b6311b55153a46f73c6459bec815f51732c980508a2.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 254, + 287, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 254, + 287, + 314 + ], + "spans": [ + { + "bbox": [ + 47, + 254, + 287, + 314 + ], + "type": "text", + "content": "Surface Curvature We employ differential geometry techniques developed by Novello et al. [29] to estimate curvature for neural implicit surfaces. In particular, we estimate the mean curvature " + }, + { + "bbox": [ + 47, + 254, + 287, + 314 + ], + "type": "inline_equation", + "content": "K(\\mathbf{x})" + }, + { + "bbox": [ + 47, + 254, + 287, + 314 + ], + "type": "text", + "content": " for the implicit surface from the divergence, " + }, + { + "bbox": [ + 47, + 254, + 287, + 314 + ], + "type": "inline_equation", + "content": "\\nabla" + }, + { + "bbox": [ + 47, + 254, + 287, + 314 + ], + "type": "text", + "content": " of the surface normals" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 127, + 320, + 287, + 344 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 320, + 287, + 344 + ], + "spans": [ + { + "bbox": [ + 127, + 320, + 287, + 344 + ], + "type": "interline_equation", + "content": "K (\\mathbf {x}) = \\frac {\\nabla \\cdot \\mathbf {n} (\\mathbf {x})}{2} \\tag {3}", + "image_path": "60a35236cd5a88b4aaa9cbe26879ce741771a1a99de2217b7df6889e17965c8a.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 350, + 287, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 350, + 287, + 409 + ], + "spans": [ + { + "bbox": [ + 47, + 350, + 287, + 409 + ], + "type": "text", + "content": "Mean curvature approximates the surface with an osculating sphere. Our approach also works for more generalized notions of curvature through the shape operator, at the cost of higher computational complexity. We refer our readers to the supplement for the general case." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 410, + 287, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 410, + 287, + 492 + ], + "spans": [ + { + "bbox": [ + 47, + 410, + 287, + 492 + ], + "type": "text", + "content": "Diffuse Radiance We separate the captured radiance at the observer camera with diffuse radiance, which depends on the glossy object's albedo, and specular radiance which depends on the environment radiance. The diffuse radiance does not have any view dependence and only depends on surface point " + }, + { + "bbox": [ + 47, + 410, + 287, + 492 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 47, + 410, + 287, + 492 + ], + "type": "text", + "content": ". We denote the diffuse radiance as " + }, + { + "bbox": [ + 47, + 410, + 287, + 492 + ], + "type": "inline_equation", + "content": "f_{d}" + }, + { + "bbox": [ + 47, + 410, + 287, + 492 + ], + "type": "text", + "content": " and model it using a coordinate-based MLP (Fig. 6)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 493, + 287, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 493, + 287, + 540 + ], + "spans": [ + { + "bbox": [ + 47, + 493, + 287, + 540 + ], + "type": "text", + "content": "Volume Rendering As proposed in [45], we perform volumetric rendering on the SDF. We define the volume density " + }, + { + "bbox": [ + 47, + 493, + 287, + 540 + ], + "type": "inline_equation", + "content": "\\sigma(\\mathbf{x})" + }, + { + "bbox": [ + 47, + 493, + 287, + 540 + ], + "type": "text", + "content": " as the cumulative distribution function (CDF), denoted as " + }, + { + "bbox": [ + 47, + 493, + 287, + 540 + ], + "type": "inline_equation", + "content": "\\Psi(s)" + }, + { + "bbox": [ + 47, + 493, + 287, + 540 + ], + "type": "text", + "content": ", applied to " + }, + { + "bbox": [ + 47, + 493, + 287, + 540 + ], + "type": "inline_equation", + "content": "f_{S}" + }, + { + "bbox": [ + 47, + 493, + 287, + 540 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 129, + 548, + 287, + 562 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 548, + 287, + 562 + ], + "spans": [ + { + "bbox": [ + 129, + 548, + 287, + 562 + ], + "type": "interline_equation", + "content": "\\sigma (\\mathbf {x}) = \\alpha \\Psi_ {\\beta} (f _ {\\mathcal {S}}) \\tag {4}", + "image_path": "000e9c138540f8c6baab57404d6f88ac8f637219ff5facda1c2721f9a40bdd8f.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 568, + 287, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 568, + 287, + 639 + ], + "spans": [ + { + "bbox": [ + 47, + 568, + 287, + 639 + ], + "type": "text", + "content": "In contrast to [45], however, we only aim to recover the diffuse radiance of the object along a particular ray. We define a function " + }, + { + "bbox": [ + 47, + 568, + 287, + 639 + ], + "type": "inline_equation", + "content": "f_{d}" + }, + { + "bbox": [ + 47, + 568, + 287, + 639 + ], + "type": "text", + "content": " that estimates the diffuse radiance at each point, " + }, + { + "bbox": [ + 47, + 568, + 287, + 639 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 47, + 568, + 287, + 639 + ], + "type": "text", + "content": ", along the ray. To get the final diffuse radiance along a given primary ray, " + }, + { + "bbox": [ + 47, + 568, + 287, + 639 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_p(t)" + }, + { + "bbox": [ + 47, + 568, + 287, + 639 + ], + "type": "text", + "content": ", we perform volumetric rendering:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 89, + 646, + 287, + 672 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 646, + 287, + 672 + ], + "spans": [ + { + "bbox": [ + 89, + 646, + 287, + 672 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {c}} _ {d} (\\mathbf {r}) = \\int_ {0} ^ {\\infty} f _ {d} (\\mathbf {r} (t), f _ {S} ^ {k} (\\mathbf {r} (t)) \\tau (t) d t \\tag {5}", + "image_path": "17ef3e0a0c0424de8366e86eb492bffb08443e02a39feda1951f8caaeeca2ab3.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "content": "Note that there is no view dependence in Eq. 5 and intermediate features, " + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "f_{\\mathcal{S}}^{k}" + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "content": ", are used as input. " + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\tau(t)" + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "content": " is the accumulated transmittance along the ray." + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 307, + 70, + 545, + 304 + ], + "blocks": [ + { + "bbox": [ + 307, + 70, + 545, + 304 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 545, + 304 + ], + "type": "image", + "image_path": "0529fab554ead1c2b7ef537769a18371272aa8a29144141cda2b7c3683399f99.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 311, + 545, + 388 + ], + "lines": [ + { + "bbox": [ + 305, + 311, + 545, + 388 + ], + "spans": [ + { + "bbox": [ + 305, + 311, + 545, + 388 + ], + "type": "text", + "content": "Figure 7. Qualitative comparisons of diffuse-specular separation and geometry estimation on rendered dataset. The environment contains nearby objects with complex occlusions when seen through reflections on the glossy object. RefNeRF fails to perform accurate diffuse-specular separation and PANDORA blurs the nearby objects in the specular map. ORCA can model the complex specular reflections through environment radiance field." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 406, + 486, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 406, + 486, + 418 + ], + "spans": [ + { + "bbox": [ + 306, + 406, + 486, + 418 + ], + "type": "text", + "content": "3.3. Objects Surface as Virtual Sensor" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 425, + 545, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 425, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 425, + 545, + 567 + ], + "type": "text", + "content": "Each pixel, " + }, + { + "bbox": [ + 304, + 425, + 545, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 304, + 425, + 545, + 567 + ], + "type": "text", + "content": ", with a finite surface area, " + }, + { + "bbox": [ + 304, + 425, + 545, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{dA}_p" + }, + { + "bbox": [ + 304, + 425, + 545, + 567 + ], + "type": "text", + "content": ", on the real-camera sensor views the surface of the object through a frustum originating at that pixel. The object then samples the environment radiance field through this finite surface converting the finite surface into a virtual pixel with surface area, " + }, + { + "bbox": [ + 304, + 425, + 545, + 567 + ], + "type": "inline_equation", + "content": "\\mathbf{dS}" + }, + { + "bbox": [ + 304, + 425, + 545, + 567 + ], + "type": "text", + "content": ". Through this model, we can interpret the object surface as a virtual sensor consisting of many virtual pixels that sample radiance from the environment field based on the geometry of the object and observer viewing direction. We now formulate a virtual pixel based on real camera pose and implicit surface geometry. Please refer to Fig. 5 for a visualization of the virtual sensor." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 568, + 545, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 568, + 545, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 568, + 545, + 616 + ], + "type": "text", + "content": "Consider a real camera origin as " + }, + { + "bbox": [ + 304, + 568, + 545, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{o}" + }, + { + "bbox": [ + 304, + 568, + 545, + 616 + ], + "type": "text", + "content": " and a pixel on the real sensor " + }, + { + "bbox": [ + 304, + 568, + 545, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_{i,j}" + }, + { + "bbox": [ + 304, + 568, + 545, + 616 + ], + "type": "text", + "content": " that corresponds to ray direction " + }, + { + "bbox": [ + 304, + 568, + 545, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{d}" + }, + { + "bbox": [ + 304, + 568, + 545, + 616 + ], + "type": "text", + "content": ". The primary ray for pixel " + }, + { + "bbox": [ + 304, + 568, + 545, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_{i,j}" + }, + { + "bbox": [ + 304, + 568, + 545, + 616 + ], + "type": "text", + "content": " is parameterized with ray length " + }, + { + "bbox": [ + 304, + 568, + 545, + 616 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 568, + 545, + 616 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 304, + 568, + 545, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{r_p}(t) = \\mathbf{o} + t\\mathbf{d}" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 616, + 545, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 616, + 545, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 616, + 545, + 676 + ], + "type": "text", + "content": "Casting Real Cones We can approximate the outgoing conical frustum from pixel " + }, + { + "bbox": [ + 304, + 616, + 545, + 676 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_{i,j}" + }, + { + "bbox": [ + 304, + 616, + 545, + 676 + ], + "type": "text", + "content": " as a cone originating at " + }, + { + "bbox": [ + 304, + 616, + 545, + 676 + ], + "type": "inline_equation", + "content": "\\mathbf{o}" + }, + { + "bbox": [ + 304, + 616, + 545, + 676 + ], + "type": "text", + "content": " with axis-of-direction " + }, + { + "bbox": [ + 304, + 616, + 545, + 676 + ], + "type": "inline_equation", + "content": "\\mathbf{d}" + }, + { + "bbox": [ + 304, + 616, + 545, + 676 + ], + "type": "text", + "content": " and radius " + }, + { + "bbox": [ + 304, + 616, + 545, + 676 + ], + "type": "inline_equation", + "content": "\\dot{r}" + }, + { + "bbox": [ + 304, + 616, + 545, + 676 + ], + "type": "text", + "content": ", equivalent to half the distance of the pixel in the x and y directions. We represent the real cone as a parametric volume," + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 682, + 545, + 695 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 682, + 545, + 695 + ], + "spans": [ + { + "bbox": [ + 317, + 682, + 545, + 695 + ], + "type": "interline_equation", + "content": "\\mathbf {r} _ {\\text {c o n e}} (\\dot {r}, s, \\theta) = \\dot {r} s \\cos (\\theta) \\hat {\\mathbf {e}} _ {u} + \\dot {r} s \\sin (\\theta) \\hat {\\mathbf {e}} _ {v} + \\dot {r} s \\mathbf {d}, \\tag {6}", + "image_path": "f8539a4055fd270c1f8c00a869cf7c5656b71b86471b27cefc9781f908fbe039.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{e}}_u" + }, + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{e}}_v" + }, + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "type": "text", + "content": " are basis vectors in the plane perpendicular" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20777" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 182, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 182, + 84 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 182, + 84 + ], + "type": "text", + "content": "to " + }, + { + "bbox": [ + 47, + 72, + 182, + 84 + ], + "type": "inline_equation", + "content": "\\mathbf{d}" + }, + { + "bbox": [ + 47, + 72, + 182, + 84 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 72, + 182, + 84 + ], + "type": "inline_equation", + "content": "\\theta \\in [0, \\pi]" + }, + { + "bbox": [ + 47, + 72, + 182, + 84 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 72, + 182, + 84 + ], + "type": "inline_equation", + "content": "s \\in [0, t_{\\max}]" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 84, + 287, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 84, + 287, + 205 + ], + "spans": [ + { + "bbox": [ + 46, + 84, + 287, + 205 + ], + "type": "text", + "content": "Virtual Pixel. Virtual pixels are characterized by the intersection of the real cone with the object surface. In Sec. 3.2, we model local surface properties using mean curvature which enables efficient analytical computations for the virtual pixel parameters even though our approach works with general shape operators. For a sampled point " + }, + { + "bbox": [ + 46, + 84, + 287, + 205 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 46, + 84, + 287, + 205 + ], + "type": "text", + "content": " along the ray, we have the surface normals " + }, + { + "bbox": [ + 46, + 84, + 287, + 205 + ], + "type": "inline_equation", + "content": "\\mathbf{n}(t_i)" + }, + { + "bbox": [ + 46, + 84, + 287, + 205 + ], + "type": "text", + "content": " from Eq. 2 and estimated mean curvature " + }, + { + "bbox": [ + 46, + 84, + 287, + 205 + ], + "type": "inline_equation", + "content": "K(t_i)" + }, + { + "bbox": [ + 46, + 84, + 287, + 205 + ], + "type": "text", + "content": " from Eq. 3. The local object surface at " + }, + { + "bbox": [ + 46, + 84, + 287, + 205 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 46, + 84, + 287, + 205 + ], + "type": "text", + "content": ", can be approximated with an osculating sphere, " + }, + { + "bbox": [ + 46, + 84, + 287, + 205 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(t_i)" + }, + { + "bbox": [ + 46, + 84, + 287, + 205 + ], + "type": "text", + "content": ", centered at " + }, + { + "bbox": [ + 46, + 84, + 287, + 205 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{o}}_{\\mathbf{S}}(t_i)" + }, + { + "bbox": [ + 46, + 84, + 287, + 205 + ], + "type": "text", + "content": " with radius, " + }, + { + "bbox": [ + 46, + 84, + 287, + 205 + ], + "type": "inline_equation", + "content": "R(t_i)" + }, + { + "bbox": [ + 46, + 84, + 287, + 205 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 75, + 221, + 259, + 247 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 221, + 259, + 247 + ], + "spans": [ + { + "bbox": [ + 75, + 221, + 259, + 247 + ], + "type": "interline_equation", + "content": "R (t _ {i}) = \\frac {2}{K _ {t _ {i}}} \\hat {\\mathbf {o}} _ {\\mathbf {S}} (t _ {i}) = \\mathbf {r} _ {p} (t _ {i}) + R (t _ {i}) \\cdot \\hat {\\mathbf {n}} (t _ {i})", + "image_path": "ffa331cadb420ceebf3f9e40486ddd7535dd5e35682178193624980d0ed2eb0f.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 255, + 287, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 255, + 287, + 277 + ], + "spans": [ + { + "bbox": [ + 46, + 255, + 287, + 277 + ], + "type": "text", + "content": "For concave surfaces " + }, + { + "bbox": [ + 46, + 255, + 287, + 277 + ], + "type": "inline_equation", + "content": "K_{t_i} < 0" + }, + { + "bbox": [ + 46, + 255, + 287, + 277 + ], + "type": "text", + "content": ". So, " + }, + { + "bbox": [ + 46, + 255, + 287, + 277 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{o}}_{\\mathbf{S}}" + }, + { + "bbox": [ + 46, + 255, + 287, + 277 + ], + "type": "text", + "content": " will lie outside the object. For " + }, + { + "bbox": [ + 46, + 255, + 287, + 277 + ], + "type": "inline_equation", + "content": "K_{t_i} > 0" + }, + { + "bbox": [ + 46, + 255, + 287, + 277 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 255, + 287, + 277 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{o}}_{\\mathbf{S}}" + }, + { + "bbox": [ + 46, + 255, + 287, + 277 + ], + "type": "text", + "content": " will lie inside the object." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 278, + 287, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 278, + 287, + 410 + ], + "spans": [ + { + "bbox": [ + 46, + 278, + 287, + 410 + ], + "type": "text", + "content": "The edges of the virtual pixel for " + }, + { + "bbox": [ + 46, + 278, + 287, + 410 + ], + "type": "inline_equation", + "content": "\\mathbf{r_p}(t_i)" + }, + { + "bbox": [ + 46, + 278, + 287, + 410 + ], + "type": "text", + "content": " would lie at the intersection of the osculating sphere " + }, + { + "bbox": [ + 46, + 278, + 287, + 410 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(t_i)" + }, + { + "bbox": [ + 46, + 278, + 287, + 410 + ], + "type": "text", + "content": " and the primary cone given by " + }, + { + "bbox": [ + 46, + 278, + 287, + 410 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_{\\mathrm{cone}}" + }, + { + "bbox": [ + 46, + 278, + 287, + 410 + ], + "type": "text", + "content": ". Computing exact cone-sphere intersections are computationally expensive so we approximate the cone-sphere intersection using rays bound cone-sphere intersectional surface dS. We consider four rays that bound the cone and sample them at " + }, + { + "bbox": [ + 46, + 278, + 287, + 410 + ], + "type": "inline_equation", + "content": "\\theta_{j}\\in \\{0,\\pi /2,\\pi ,3\\pi /2\\}" + }, + { + "bbox": [ + 46, + 278, + 287, + 410 + ], + "type": "text", + "content": " with Eq. 6. We perform intersections of the corresponding bounding rays with the osculating sphere " + }, + { + "bbox": [ + 46, + 278, + 287, + 410 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(t_i)" + }, + { + "bbox": [ + 46, + 278, + 287, + 410 + ], + "type": "text", + "content": " to get corners of the virtual pixel " + }, + { + "bbox": [ + 46, + 278, + 287, + 410 + ], + "type": "inline_equation", + "content": "\\mathbf{ds}_j" + }, + { + "bbox": [ + 46, + 278, + 287, + 410 + ], + "type": "text", + "content": ". These ray sphere intersections can be computed analytically in an efficient manner." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 410, + 287, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 410, + 287, + 458 + ], + "spans": [ + { + "bbox": [ + 47, + 410, + 287, + 458 + ], + "type": "text", + "content": "Virtual Cone Origin. From the virtual pixel surface area, we can now compute the virtual cone that samples the environment. We first compute normal vectors at virtual pixel corners " + }, + { + "bbox": [ + 47, + 410, + 287, + 458 + ], + "type": "inline_equation", + "content": "\\mathbf{ds_j}" + }, + { + "bbox": [ + 47, + 410, + 287, + 458 + ], + "type": "text", + "content": " from the center of osculating sphere " + }, + { + "bbox": [ + 47, + 410, + 287, + 458 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{o}}_{\\mathrm{S}}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 127, + 467, + 287, + 493 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 467, + 287, + 493 + ], + "spans": [ + { + "bbox": [ + 127, + 467, + 287, + 493 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {n}} _ {\\mathbf {j}} = \\frac {\\mathbf {d} \\mathbf {s} _ {\\mathbf {j}} - \\hat {\\mathbf {o}} _ {\\mathbf {S}}}{\\left| \\left| \\mathbf {d} \\mathbf {s} _ {\\mathbf {j}} - \\hat {\\mathbf {o}} _ {\\mathbf {S}} \\right| \\right|} \\tag {7}", + "image_path": "f955eb87b6fd7f7b53beeecc16acc65bd51c47f7dfb20935f05539cc1e54350c.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 500, + 287, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 500, + 287, + 550 + ], + "spans": [ + { + "bbox": [ + 46, + 500, + 287, + 550 + ], + "type": "text", + "content": "At each virtual pixel corner, we compute the reflected ray directions, " + }, + { + "bbox": [ + 46, + 500, + 287, + 550 + ], + "type": "inline_equation", + "content": "\\omega_{j}^{r}" + }, + { + "bbox": [ + 46, + 500, + 287, + 550 + ], + "type": "text", + "content": ", by computing the dot product between the incoming ray directions, " + }, + { + "bbox": [ + 46, + 500, + 287, + 550 + ], + "type": "inline_equation", + "content": "\\omega_{k}^{i}" + }, + { + "bbox": [ + 46, + 500, + 287, + 550 + ], + "type": "text", + "content": ", and the normals, " + }, + { + "bbox": [ + 46, + 500, + 287, + 550 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{n}}_k" + }, + { + "bbox": [ + 46, + 500, + 287, + 550 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 500, + 287, + 550 + ], + "type": "inline_equation", + "content": "\\omega_0^r" + }, + { + "bbox": [ + 46, + 500, + 287, + 550 + ], + "type": "text", + "content": " is the primary ray's reflected vector." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 119, + 570, + 287, + 583 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 570, + 287, + 583 + ], + "spans": [ + { + "bbox": [ + 119, + 570, + 287, + 583 + ], + "type": "interline_equation", + "content": "\\omega_ {0} ^ {r} = \\mathbf {d} - \\left(\\mathbf {d} \\cdot \\hat {\\mathbf {n}} \\left(t _ {i}\\right)\\right) \\hat {\\mathbf {n}} \\left(t _ {i}\\right) \\tag {8}", + "image_path": "37f797b0e8380188535382bfd064e7a41eb17a2d0b94a4316ba1685181f050ee.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 586, + 287, + 600 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 586, + 287, + 600 + ], + "spans": [ + { + "bbox": [ + 105, + 586, + 287, + 600 + ], + "type": "interline_equation", + "content": "\\omega_ {j} ^ {r} = \\mathbf {d} _ {j} - \\left(\\mathbf {d} _ {j} \\cdot \\hat {\\mathbf {n}} _ {j} (k)\\right) \\hat {\\mathbf {n}} _ {j} (k) \\tag {9}", + "image_path": "499086cae09ade9b3c1e835a3a1c30a1fe912df93b623edaa0b39195f5b6865a.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 609, + 287, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 609, + 287, + 681 + ], + "spans": [ + { + "bbox": [ + 46, + 609, + 287, + 681 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_j" + }, + { + "bbox": [ + 46, + 609, + 287, + 681 + ], + "type": "text", + "content": " are the incident directions to the virtual pixel corners " + }, + { + "bbox": [ + 46, + 609, + 287, + 681 + ], + "type": "inline_equation", + "content": "\\mathbf{ds}_j" + }, + { + "bbox": [ + 46, + 609, + 287, + 681 + ], + "type": "text", + "content": ". The virtual cone origin is the intersection of these reflected rays at the pixel corners and pixel center. However, these rays might not intersect at a single point so we approximate a virtual origin to be the point that minimizes the sum of distances to the reflected rays " + }, + { + "bbox": [ + 46, + 609, + 287, + 681 + ], + "type": "inline_equation", + "content": "\\omega_j" + }, + { + "bbox": [ + 46, + 609, + 287, + 681 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 93, + 689, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 93, + 689, + 287, + 715 + ], + "type": "interline_equation", + "content": "\\mathbf {v} _ {\\mathbf {o}} = \\operatorname {a r g m i n} _ {\\mathbf {v}} \\sum_ {j} | (\\mathbf {v} - \\mathbf {d s} _ {\\mathbf {j}}) \\times \\omega_ {j} ^ {r} | \\tag {10}", + "image_path": "5429af46c0acc395307477c2e2b71f6e0475e91ca1654894f5ff9140bffcdcbf.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": "We pose this as a linear least squares problem and estimate the virtual cone origin efficiently through pseudo-inverse." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 97, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 97, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 305, + 97, + 545, + 144 + ], + "type": "text", + "content": "Virtual Cones Direction. The reflected ray at the center of the virtual pixel reflects the object surface along the direction " + }, + { + "bbox": [ + 305, + 97, + 545, + 144 + ], + "type": "inline_equation", + "content": "\\omega_0^r" + }, + { + "bbox": [ + 305, + 97, + 545, + 144 + ], + "type": "text", + "content": " from Eq. 8. We consider this as the direction-of-axis of the virtual cone." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 406, + 156, + 545, + 169 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 406, + 156, + 545, + 169 + ], + "spans": [ + { + "bbox": [ + 406, + 156, + 545, + 169 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {v}} _ {d} = \\omega_ {0} ^ {r} \\tag {11}", + "image_path": "e904f06699a6193d6d32c3e877af96a1efec4adb68d8d1f32b612e6452ad31c5.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 180, + 545, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 180, + 545, + 252 + ], + "spans": [ + { + "bbox": [ + 305, + 180, + 545, + 252 + ], + "type": "text", + "content": "Virtual Cone Radius. We compute the radius of the cone by treating the reflection vectors of the bounding rays as the neighboring \"pixel\" directions. Similar to [5], we can compute the distance between " + }, + { + "bbox": [ + 305, + 180, + 545, + 252 + ], + "type": "inline_equation", + "content": "\\{\\omega_{k_{\\theta}}^{r}\\}_{\\theta = 0}^{2\\pi}" + }, + { + "bbox": [ + 305, + 180, + 545, + 252 + ], + "type": "text", + "content": " and the primary reflected ray " + }, + { + "bbox": [ + 305, + 180, + 545, + 252 + ], + "type": "inline_equation", + "content": "\\omega_0^r" + }, + { + "bbox": [ + 305, + 180, + 545, + 252 + ], + "type": "text", + "content": " in the " + }, + { + "bbox": [ + 305, + 180, + 545, + 252 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 305, + 180, + 545, + 252 + ], + "type": "text", + "content": " components (omitted below for clarity)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 375, + 275, + 545, + 289 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 275, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 375, + 275, + 545, + 289 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {v}} _ {\\dot {r}} = \\left\\| \\left\\{\\omega_ {k _ {\\theta}} ^ {r} \\right\\} _ {\\theta = 0} ^ {2 \\pi} - \\omega_ {0} ^ {r} \\right\\| \\tag {12}", + "image_path": "edbd66947887b40c428a10a6fea8d9a36a7aa8cb9e067622ff885d9ad9b73848.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 300, + 545, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 300, + 545, + 347 + ], + "spans": [ + { + "bbox": [ + 305, + 300, + 545, + 347 + ], + "type": "text", + "content": "Finally, for each sampled point " + }, + { + "bbox": [ + 305, + 300, + 545, + 347 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 305, + 300, + 545, + 347 + ], + "type": "text", + "content": ", we can characterize our single-pixel virtual sensor located at the object surface " + }, + { + "bbox": [ + 305, + 300, + 545, + 347 + ], + "type": "inline_equation", + "content": "\\mathbf{dS}" + }, + { + "bbox": [ + 305, + 300, + 545, + 347 + ], + "type": "text", + "content": " as a virtual cone with " + }, + { + "bbox": [ + 305, + 300, + 545, + 347 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{v}}_o" + }, + { + "bbox": [ + 305, + 300, + 545, + 347 + ], + "type": "text", + "content": " as its apex, " + }, + { + "bbox": [ + 305, + 300, + 545, + 347 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{v}}_d" + }, + { + "bbox": [ + 305, + 300, + 545, + 347 + ], + "type": "text", + "content": " as axis-direction, " + }, + { + "bbox": [ + 305, + 300, + 545, + 347 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{v}}_{\\dot{r}}" + }, + { + "bbox": [ + 305, + 300, + 545, + 347 + ], + "type": "text", + "content": " as the radius." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 349, + 545, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 349, + 545, + 577 + ], + "spans": [ + { + "bbox": [ + 304, + 349, + 545, + 577 + ], + "type": "text", + "content": "Connections to caustics. Our work takes inspiration from catadioptric imaging systems. To convert objects into cameras, we compute the surface and find a corresponding center-of-projection for this surface-as-sensor. However, unlike conventional perspective cameras, objects don't have a fixed center-of-projection, other than in a few special configurations [3], but a locus of viewpoints that vary with object geometry and viewing direction. These viewpoints lie on the \"caustic surface\" of the object. While typical works in catadioptric imaging analytically compute the caustic surface by assuming known geometry [14, 38], or making assumptions about placement of the observer [40], our formulation approximates the caustic surface of unknown geometry through the intersection of reflected rays on virtual pixels. We empirically show in the supplement that as the surface area of the virtual pixel goes to " + }, + { + "bbox": [ + 304, + 349, + 545, + 577 + ], + "type": "inline_equation", + "content": "0" + }, + { + "bbox": [ + 304, + 349, + 545, + 577 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 349, + 545, + 577 + ], + "type": "inline_equation", + "content": "d\\mathbf{S} \\rightarrow 0" + }, + { + "bbox": [ + 304, + 349, + 545, + 577 + ], + "type": "text", + "content": ", our method estimates the true caustic of the object without assuming geometry. Our method also has applications in estimating the caustic surface of the unknown geometry." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 586, + 467, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 586, + 467, + 597 + ], + "spans": [ + { + "bbox": [ + 306, + 586, + 467, + 597 + ], + "type": "text", + "content": "3.4. Environment Radiance Fields" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 605, + 545, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 605, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 305, + 605, + 545, + 677 + ], + "type": "text", + "content": "Our goal is to capture a 5D environment radiance field of the scene by imaging the world through these single-pixel virtual sensors located at the object's surface. We use our formulation of virtual cones to recover 5D environment radiance fields. We define an environment radiance field as " + }, + { + "bbox": [ + 305, + 605, + 545, + 677 + ], + "type": "inline_equation", + "content": "f_{\\mathcal{E}}: (\\hat{\\mathbf{v}}_o, \\hat{\\mathbf{v}}_d) \\to (\\sigma^{Env}, c_s)" + }, + { + "bbox": [ + 305, + 605, + 545, + 677 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "f_{\\varepsilon}" + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": " outputs opacity and radiance along sampled virtual cones. We note that this view-dependent radiance is equivalent to the specular radiance at point " + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": " sampled" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20778" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 69, + 286, + 134 + ], + "blocks": [ + { + "bbox": [ + 50, + 69, + 286, + 134 + ], + "lines": [ + { + "bbox": [ + 50, + 69, + 286, + 134 + ], + "spans": [ + { + "bbox": [ + 50, + 69, + 286, + 134 + ], + "type": "image", + "image_path": "413e0df2124e259ffdb6edf7e6eeb29af2b2fa53113aac8e2ddfce0478c9883e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 142, + 287, + 175 + ], + "lines": [ + { + "bbox": [ + 46, + 142, + 287, + 175 + ], + "spans": [ + { + "bbox": [ + 46, + 142, + 287, + 175 + ], + "type": "text", + "content": "Figure 8. Evaluation of environment depth from reflections. For the scene in Fig. 3, ORCa can faithfully estimate the depth of the environment, especially for nearby surroundings." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 48, + 188, + 285, + 239 + ], + "blocks": [ + { + "bbox": [ + 48, + 188, + 285, + 239 + ], + "lines": [ + { + "bbox": [ + 48, + 188, + 285, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 188, + 285, + 239 + ], + "type": "table", + "html": "
ApproachDiffuse RadianceSpecular RadianceMixed RadianceNormals
PSNR ↑ (dB)SSIM ↑PSNR ↑ (dB)SSIM ↑PSNR ↑ (dB)SSIM ↑MAE ↓ (°)
Ref-NeRF18.800.730416.990.663320.190.789043.690
PANDORA18.250.726016.250.648318.900.72847.606
ORCA19.840.789320.740.753522.000.79472.339
", + "image_path": "36134989eff7efd70451770de2133c601015ac8f38333e939ea7ee9bed43dc08.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 335, + 287, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 335, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 335, + 287, + 361 + ], + "type": "text", + "content": "along the primary-camera ray " + }, + { + "bbox": [ + 46, + 335, + 287, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_{\\mathbf{p}}(t)" + }, + { + "bbox": [ + 46, + 335, + 287, + 361 + ], + "type": "text", + "content": ". We can render the final specular radiance at pixel " + }, + { + "bbox": [ + 46, + 335, + 287, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_{i,j}" + }, + { + "bbox": [ + 46, + 335, + 287, + 361 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 102, + 369, + 231, + 396 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 369, + 231, + 396 + ], + "spans": [ + { + "bbox": [ + 102, + 369, + 231, + 396 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {c}} _ {s} (\\mathbf {r}) = \\int_ {0} ^ {\\infty} f _ {\\mathcal {E}} (\\hat {\\mathbf {v}} _ {o}, \\hat {\\mathbf {v}} _ {d}) \\tau (t) d t", + "image_path": "c8634c241b663e83862f41e0b05ba89b3b6c5a3dba04f3dac471e4d204e86006.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 140, + 399, + 194, + 410 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 399, + 194, + 410 + ], + "spans": [ + { + "bbox": [ + 140, + 399, + 194, + 410 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {c}} = \\hat {\\mathbf {c _ {d}}} + \\hat {\\mathbf {c _ {s}}}", + "image_path": "917e4f68c8a7f739d4e3425abc21e46e0f20423455daa4846bae6c2aeea87680.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 422, + 288, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 422, + 288, + 542 + ], + "spans": [ + { + "bbox": [ + 46, + 422, + 288, + 542 + ], + "type": "text", + "content": "Intuitively, " + }, + { + "bbox": [ + 46, + 422, + 288, + 542 + ], + "type": "inline_equation", + "content": "f_{\\mathcal{E}}" + }, + { + "bbox": [ + 46, + 422, + 288, + 542 + ], + "type": "text", + "content": " learns the 5D radiance field by sampling single-pixel virtual sensors from the object surface and must learn geometry and environment radiance that is consistent with the multi-view reflections. Moreover, we can query " + }, + { + "bbox": [ + 46, + 422, + 288, + 542 + ], + "type": "inline_equation", + "content": "f_{\\mathcal{E}}" + }, + { + "bbox": [ + 46, + 422, + 288, + 542 + ], + "type": "text", + "content": " to render novel viewpoints and associated depths that are beyond the field-of-view of the real camera. We volume render each virtual cone by dividing them into conical frustums using Integrated-Positional Encoding as proposed in MipNeRF [5]. Our formulation of virtual cones works well with Mip-Nerf's rays-as-cones method." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 555, + 128, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 555, + 128, + 567 + ], + "spans": [ + { + "bbox": [ + 47, + 555, + 128, + 567 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 574, + 180, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 574, + 180, + 588 + ], + "spans": [ + { + "bbox": [ + 47, + 574, + 180, + 588 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "content": "As in PANDORA, we parameterize " + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "inline_equation", + "content": "f_{S}" + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "content": " with an 8-layer MLP to estimate the surface, and, as in MipNeRF, " + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "inline_equation", + "content": "f_{d}" + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "content": " with 4-layer MLP with input geometric features of size 512 from " + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "inline_equation", + "content": "f_{S}" + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "content": ". We follow the SDF-to-opacity conversion and the iterative sampling of the ray proposed in [45]. To aid the network to learn geometry quickly, we also train " + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "inline_equation", + "content": "f_{S}" + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "content": " with a mask-net as proposed in [10]. We use five losses in our architecture: photometric loss, mask loss [10], normal loss [41], eikonal loss [45], and distortion loss [6]. Additional training details are discussed in the supplement." + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 318, + 70, + 424, + 137 + ], + "blocks": [ + { + "bbox": [ + 46, + 247, + 287, + 314 + ], + "lines": [ + { + "bbox": [ + 46, + 247, + 287, + 314 + ], + "spans": [ + { + "bbox": [ + 46, + 247, + 287, + 314 + ], + "type": "text", + "content": "Table 1. Average evaluation metrics on rendered scenes. We compare ORCa to other neural rendering techniques that model reflections, including Ref-NeRF and PANDORA, on six simulated scenes. ORCa provides consistent improvements in geometry estimation, diffuse-specular separation and novel view synthesis. Please refer to the supplement for additional metrics." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 70, + 424, + 137 + ], + "lines": [ + { + "bbox": [ + 318, + 70, + 424, + 137 + ], + "spans": [ + { + "bbox": [ + 318, + 70, + 424, + 137 + ], + "type": "table", + "html": "
Scene: Ball-cup in hallway
ApproachMixedRadiance
PSNRSSIM
↑(dB)↑
Ref-NeRF32.750.9617
PANDORA28.830.9758
ORCA30.860.9799
", + "image_path": "9dfddaa4d2b31821db93af595b01f37bdd75890edf9dd3a0f7798ce08951ab91.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 425, + 70, + 533, + 137 + ], + "blocks": [ + { + "bbox": [ + 425, + 70, + 533, + 137 + ], + "lines": [ + { + "bbox": [ + 425, + 70, + 533, + 137 + ], + "spans": [ + { + "bbox": [ + 425, + 70, + 533, + 137 + ], + "type": "table", + "html": "
Scene: Owl in hallway
ApproachMixed Radiance
PSNRSSIM
↑ (dB)↑
Ref-NeRF26.650.8890
PANDORA27.240.9343
ORCA26.840.9299
", + "image_path": "68eacca3ce83ef7e3a267f0d5ead74177f52e370e2e57a2958b3e9834cce8f7f.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 145, + 545, + 178 + ], + "lines": [ + { + "bbox": [ + 305, + 145, + 545, + 178 + ], + "spans": [ + { + "bbox": [ + 305, + 145, + 545, + 178 + ], + "type": "text", + "content": "Table 2. Quantitative metrics for real-world, captured scenes. ORCa demonstrates comparable performance in novel view synthesis on scenes from the PANDORA real dataset." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 306, + 199, + 368, + 211 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 199, + 368, + 211 + ], + "spans": [ + { + "bbox": [ + 306, + 199, + 368, + 211 + ], + "type": "text", + "content": "4.2. Datasets" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 218, + 545, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 218, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 304, + 218, + 545, + 326 + ], + "type": "text", + "content": "We conduct experiments on both simulated and real-world datasets. Simulated datasets are rendered in Mitsu2 [28]. Simulated datasets contain a range of increasingly complex object geometries (elephant, Pokeball, and orca) and scenes (living room and Pokemon). We train with 200 views for simulated datasets. We also show results for a real-world dataset [10] capturing a glossy cup with a black vase sitting atop it using 35 views. All datasets will be publicly released upon publication." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 334, + 537, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 334, + 537, + 346 + ], + "spans": [ + { + "bbox": [ + 305, + 334, + 537, + 346 + ], + "type": "text", + "content": "4.3. Advantages of Environment Radiance Fields" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 352, + 545, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 352, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 352, + 545, + 544 + ], + "type": "text", + "content": "Other neural rendering techniques that handle reflections, Ref-NeRF and PANDORA, estimate the environment as a 2D map, while ORCa recovers a 5D environment radiance field. Fig. 3 shows the advantages of recovering a 5D environment radiance field. Close-by surrounding objects often cause occlusions that cannot be modeled by 2D environment maps. By estimating the radiance field, we can image behind occluders through sampling novel viewpoints such as the translated viewpoints shown in Fig. 3. Moreover, we can also show the depth to the surroundings from these virtual viewpoints. In Fig. 8, we quantitatively evaluate the estimated depth of the surroundings for the synthesized virtual views in Fig. 3 and show that ORCa provides reliable depth estimates especially for nearby objects. We provide additional examples of depth estimation and beyond field-of-view novel-view synthesis in the supplement." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 552, + 460, + 564 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 552, + 460, + 564 + ], + "spans": [ + { + "bbox": [ + 306, + 552, + 460, + 564 + ], + "type": "text", + "content": "4.4. Comparisons with Baselines" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": "We compare ORCa with Ref-NeRF and PANDORA on synthetic and real datasets. As shown in Fig. 7 and 9, ORCa estimates more accurate surface normals than other methods. While the total radiance predicted by Ref-NeRF and PANDORA are visually similar, the surface normals are less smooth than ORCa. We also observe that ORCa is able to achieve better diffuse and specular radiance separation than PANDORA, which is evident in the Pokeball surface normals Fig. 7. In these examples, PANDORA recovers blurry specular radiance. We see that ORCa's predicted depth is highly interpretable and matches the underlying geometry of the environment, as shown in Fig. 3. Even on cylindri" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20779" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 197, + 286 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 197, + 286 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 197, + 286 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 197, + 286 + ], + "type": "image", + "image_path": "b8c49e4f653135a5bfcfea828e7f63b0338865f2c803608894e4402fc01b0a06.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 293, + 288, + 392 + ], + "lines": [ + { + "bbox": [ + 46, + 293, + 288, + 392 + ], + "spans": [ + { + "bbox": [ + 46, + 293, + 288, + 392 + ], + "type": "text", + "content": "Figure 9. Comparisons on real dataset. Using a real-world dataset with only 35 views, ORCa can model the sharp specularities on the ball arriving from regions of the nearby scene, such as the table, and far away scene regions, such as the hallway, to learn an environment radiance field. We query the radiance field for the depth of the far hallway (blue) and the nearby objects, such as the table (red). We also render novel viewpoints that are beyond the field of view of the observer camera and show that ORCa interpolates well between those views." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 200, + 70, + 288, + 284 + ], + "blocks": [ + { + "bbox": [ + 200, + 70, + 288, + 284 + ], + "lines": [ + { + "bbox": [ + 200, + 70, + 288, + 284 + ], + "spans": [ + { + "bbox": [ + 200, + 70, + 288, + 284 + ], + "type": "image", + "image_path": "cb2188ee0525ae84b3267adaa2b3c2983a96e4c7147e764a619a70c63dbd4a81.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 185, + 547, + 251 + ], + "lines": [ + { + "bbox": [ + 305, + 185, + 547, + 251 + ], + "spans": [ + { + "bbox": [ + 305, + 185, + 547, + 251 + ], + "type": "text", + "content": "Figure 10. Ablations on virtual cone computation Accurate estimation of the virtual cones (Sec. 3.3) is crucial for ORCa. If the virtual cone origin is assumed to be at the object surface (left column) or if the surface is assumed to locally have no curvature (right column), the surface normal and specular radiance outputs suffer from artifacts (red boxes)." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 413, + 287, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 413, + 287, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 413, + 287, + 604 + ], + "type": "text", + "content": "cal real-world datasets, such as the black vase in Fig. 9, the nearby hallway is visible in both the virtual view and depth, despite never being in the field of view of the primary camera. Unlike Ref-NeRF, our primary objective is not to perform novel-view synthesis, but instead to capture the environment radiance field from the object surface. As shown in Table 1, ORCa improves over Ref-NeRF and PANDORA in estimating diffuse radiance, specular radiance, mixed radiance, and normals. Here we report the image quality metrics, PSNR and SSIM, and surface normal metric, MAE, averaged over six synthetic scenes with varying environments, camera view distribution, and object complexity. Please refer to the supplement for individual metrics on each scene. ORCa also shows competitive performance for novel-view synthesis metrics on real-world scenes from PANDORA dataset, as shown in Table 2." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 611, + 244, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 611, + 244, + 624 + ], + "spans": [ + { + "bbox": [ + 47, + 611, + 244, + 624 + ], + "type": "text", + "content": "4.5. Impact of Virtual Cone Computation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "content": "We base our method on a physically accurate formulation by modeling ray-cone intersections and using the surface as a virtual sensor, as described in Sec. 3.3. We demonstrate the importance of this step by setting up two ablation experiments. In the first experiment, which we term Naive Virtual Cone, we place the origin of the cone at the object surface instead of computing the virtual cone origin based" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 307, + 72, + 547, + 177 + ], + "blocks": [ + { + "bbox": [ + 307, + 72, + 547, + 177 + ], + "lines": [ + { + "bbox": [ + 307, + 72, + 547, + 177 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 547, + 177 + ], + "type": "image", + "image_path": "a17c9af40a96d28275bd72eb531568ca8e35ab246ece9f22045da15905766aba.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 280, + 547, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 280, + 547, + 352 + ], + "spans": [ + { + "bbox": [ + 305, + 280, + 547, + 352 + ], + "type": "text", + "content": "on the curvature. In the second experiment, which we term No Curvature, we assume that locally the surface is like a flat mirror and has no curvature. In Fig. 10, we show that for both of these ablation experiments, we see worse performance as demonstrated by the artifacts in estimated surface normals and specular components (shown as red boxes)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 374, + 378, + 386 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 374, + 378, + 386 + ], + "spans": [ + { + "bbox": [ + 306, + 374, + 378, + 386 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 398, + 547, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 398, + 547, + 602 + ], + "spans": [ + { + "bbox": [ + 304, + 398, + 547, + 602 + ], + "type": "text", + "content": "We present a method to convert glossy objects with unknown geometry and texture into radiance-field cameras that capture the environment's radiance field around them. Our method recovers object geometry and diffuse radiance, in addition to capturing the depth and radiance of the object's surroundings from its perspective. Our modeling of the environment as a radiance field is effective in recovering close-by objects (Fig. 7) and is occlusion aware (Fig. 3). From the recovered environment radiance field, we can perform beyond field-of-view novel-view synthesis. Our work can unleash applications in virtual object insertion and 3D perception, e.g. inferring information beyond the line-of-sight of the camera using predicted virtual views and depth. Our formulation goes beyond the conventional direct-line-of-sight radiance fields and can enable further areas of research to extract more information from multi-view images directly from the environment and the objects present in it." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 624, + 403, + 637 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 624, + 403, + 637 + ], + "spans": [ + { + "bbox": [ + 306, + 624, + 403, + 637 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 647, + 547, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 647, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 305, + 647, + 547, + 712 + ], + "type": "text", + "content": "This research was supported by the SMART Contract IARPA Grant 2021-20111000004. We would also like to thank Systems Technology Research (STR). AD and AV are supported by NSF Expeditions-Award IIS-1730574 and NSF CAREER-Award IIS-1652633." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20780" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 146 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 146 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 146 + ], + "type": "text", + "content": "[1] Anna Alperovich, Ole Johannsen, and Bastian Goldluecke. Intrinsic light field decomposition and disparity estimation with deep encoder-decoder network. In 2018 26th European Signal Processing Conference (EUSIPCO), pages 2165-2169. IEEE, 2018. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 148, + 288, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 288, + 180 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 288, + 180 + ], + "type": "text", + "content": "[2] S. Baker and S.K. Nayar. A theory of catadioptric image formation. In Sixth International Conference on Computer Vision (IEEE Cat. No.98CH36271), pages 35-42, 1998. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 182, + 288, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 182, + 288, + 216 + ], + "spans": [ + { + "bbox": [ + 53, + 182, + 288, + 216 + ], + "type": "text", + "content": "[3] Simon Baker and Shree K. Nayar. A theory of single-viewpoint catadioptric image formation. International Journal of Computer Vision, 35:175-196, 2004. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 217, + 287, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 217, + 287, + 259 + ], + "spans": [ + { + "bbox": [ + 53, + 217, + 287, + 259 + ], + "type": "text", + "content": "[4] Jonathan T Barron and Jitendra Malik. Shape, illumination, and reflectance from shading. IEEE transactions on pattern analysis and machine intelligence, 37(8):1670-1687, 2014. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 262, + 287, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 262, + 287, + 306 + ], + "spans": [ + { + "bbox": [ + 53, + 262, + 287, + 306 + ], + "type": "text", + "content": "[5] Jonathan T. Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P. Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. ICCV, 2021. 3, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 308, + 288, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 308, + 288, + 363 + ], + "spans": [ + { + "bbox": [ + 53, + 308, + 288, + 363 + ], + "type": "text", + "content": "[6] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5470–5479, 2022. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 364, + 288, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 364, + 288, + 418 + ], + "spans": [ + { + "bbox": [ + 53, + 364, + 288, + 418 + ], + "type": "text", + "content": "[7] Mark Boss, Raphael Braun, Varun Jampani, Jonathan T Barron, Ce Liu, and Hendrik Lensch. Nerd: Neural reflectance decomposition from image collections. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12684-12694, 2021. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 420, + 287, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 420, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 53, + 420, + 287, + 475 + ], + "type": "text", + "content": "[8] Mark Boss, Varun Jampani, Raphael Braun, Ce Liu, Jonathan Barron, and Hendrik Lensch. Neural-pil: Neural pre-integrated lighting for reflectance decomposition. Advances in Neural Information Processing Systems, 34:10691-10704, 2021. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 476, + 287, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 476, + 287, + 520 + ], + "spans": [ + { + "bbox": [ + 53, + 476, + 287, + 520 + ], + "type": "text", + "content": "[9] Akshit Dave, Yannick Hold-Geoffroy, Milos Hasan, Kalyan Sunkavalli, and Ashok Veeraraghavan. Snapshot polarimetric diffuse-specular separation. Optics Express, 30(19):34239-34255, 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 522, + 287, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 522, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 522, + 287, + 555 + ], + "type": "text", + "content": "[10] Akshit Dave, Yongyi Zhao, and Ashok Veeraraghavan. Pandora: Polarization-aided neural decomposition of radiance. arXiv preprint arXiv:2203.13458, 2022. 3, 4, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 556, + 287, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 556, + 287, + 600 + ], + "spans": [ + { + "bbox": [ + 48, + 556, + 287, + 600 + ], + "type": "text", + "content": "[11] Michael De Zeeuw and Aswin C Sankaranarayanan. Wide-baseline light fields using ellipsoidal mirrors. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 602, + 287, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 287, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 287, + 656 + ], + "type": "text", + "content": "[12] Mathieu Garon, Kalyan Sunkavalli, Sunil Hadap, Nathan Carr, and Jean-François Lalonde. Fast spatially-varying indoor lighting estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6908-6917, 2019. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[13] Stamatios Georgoulis, Konstantinos Rematas, Tobias Ritschel, Mario Fritz, Tinne Tuytelaars, and Luc Van Gool. What is around the camera? In Proceedings of the IEEE International Conference on Computer Vision, pages 5170-5178, 2017. 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "text", + "content": "[14] J. Gluckman and S.K. Nayar. Planar catadioptric stereo: geometry and calibration. In Proceedings. 1999 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (Cat. No PR00149), volume 1, pages 22-28 Vol. 1, 1999. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 129, + 547, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 547, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 547, + 162 + ], + "type": "text", + "content": "[15] Michael D Grossberg and Shree K Nayar. The raxel imaging model and ray-based calibration. International Journal of Computer Vision, 61(2):119, 2005. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 163, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 163, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 308, + 163, + 545, + 217 + ], + "type": "text", + "content": "[16] Yuan-Chen Guo, Di Kang, Linchao Bao, Yu He, and Song-Hai Zhang. Nerfren: Neural radiance fields with reflections. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18409-18418, June 2022. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 220, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 220, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 308, + 220, + 545, + 262 + ], + "type": "text", + "content": "[17] Connor Henley, Siddharth Somasundaram, Joseph Hollmann, and Ramesh Raskar. Detection and mapping of specular surfaces using multibounce lidar returns. Optics Express, 31(4):6370-6388, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 264, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 264, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 308, + 264, + 545, + 319 + ], + "type": "text", + "content": "[18] Jan Jachnik, Richard A Newcombe, and Andrew J Davison. Real-time surface light-field capture for augmentation of planar specular surfaces. In 2012 IEEE International Symposium on Mixed and Augmented Reality (ISMAR), pages 91-97. IEEE, 2012. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 321, + 545, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 321, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 308, + 321, + 545, + 364 + ], + "type": "text", + "content": "[19] Georgios Kopanas, Thomas Leimkuhler, Gilles Rainer, Clément Jambon, and George Drettakis. Neural point cata-caustics for novel-view synthesis of reflections. ACM Transactions on Graphics, 41(6):Article-201, 2022. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 365, + 545, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 365, + 545, + 421 + ], + "spans": [ + { + "bbox": [ + 308, + 365, + 545, + 421 + ], + "type": "text", + "content": "[20] Rui Li, Simeng Qiu, Guangming Zang, and Wolfgang Heidrich. Reflection separation via multi-bounce polarization state tracing. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIII 16, pages 781-796. Springer, 2020. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 422, + 545, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 422, + 545, + 455 + ], + "spans": [ + { + "bbox": [ + 308, + 422, + 545, + 455 + ], + "type": "text", + "content": "[21] Tingtian Li, Daniel PK Lun, Yuk-Hee Chan, et al. Robust reflection removal based on light field imaging. IEEE Transactions on Image Processing, 28(4):1798–1812, 2018. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 456, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 456, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 308, + 456, + 545, + 521 + ], + "type": "text", + "content": "[22] Zhengqin Li, Mohammad Shafiei, Ravi Ramamoorthi, Kalyan Sunkavalli, and Manmohan Chandraker. Inverse rendering for complex indoor scenes: Shape, spatially-varying lighting and svbrdf from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2475-2484, 2020. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 523, + 545, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 523, + 545, + 556 + ], + "spans": [ + { + "bbox": [ + 308, + 523, + 545, + 556 + ], + "type": "text", + "content": "[23] Stephen Lombardi and Ko Nishino. Reflectance and natural illumination from a single image. In European Conference on Computer Vision, pages 582-595. Springer, 2012. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 557, + 545, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 557, + 545, + 610 + ], + "spans": [ + { + "bbox": [ + 308, + 557, + 545, + 610 + ], + "type": "text", + "content": "[24] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 613, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 613, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 308, + 613, + 545, + 678 + ], + "type": "text", + "content": "[25] Jacob Munkberg, Jon Hasselgren, Tianchang Shen, Jun Gao, Wenzheng Chen, Alex Evans, Thomas Müller, and Sanja Fidler. Extracting triangular 3d models, materials, and lighting from images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8280-8290, 2022. 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 680, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 680, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 308, + 680, + 545, + 714 + ], + "type": "text", + "content": "[26] Shree K Nayar and Simon Baker. Catadioptric image formation. In Proceedings of the 1997 DARPA Image Understanding Workshop, pages 1431-1437, 1997. 3" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20781" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[27] Yun Ni, Jie Chen, and Lap-Pui Chau. Reflection removal on single light field capture using focus manipulation. IEEE Transactions on Computational Imaging, 4(4):562-572, 2018. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 287, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 287, + 160 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 287, + 160 + ], + "type": "text", + "content": "[28] Merlin Nimier-David, Delio Vicini, Tizian Zeltner, and Wenzel Jakob. Mitsuba 2: A retargetable forward and inverse renderer. ACM Transactions on Graphics (TOG), 38(6):1-17, 2019. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 162, + 287, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 162, + 287, + 195 + ], + "spans": [ + { + "bbox": [ + 49, + 162, + 287, + 195 + ], + "type": "text", + "content": "[29] Tiago Novello, Guilherme Schardong, Luiz Schirmer, Vinicius da Silva, Helio Lopes, and Luiz Velho. Exploring differential geometry in neural implicits, 2022. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 196, + 287, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 196, + 287, + 249 + ], + "spans": [ + { + "bbox": [ + 49, + 196, + 287, + 249 + ], + "type": "text", + "content": "[30] Michael Oechsle, Songyou Peng, and Andreas Geiger. Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5589-5599, 2021. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 251, + 287, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 251, + 287, + 294 + ], + "spans": [ + { + "bbox": [ + 49, + 251, + 287, + 294 + ], + "type": "text", + "content": "[31] Jeong Joon Park, Aleksander Holynski, and Steven M Seitz. Seeing the world in a bag of chips. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1417-1427, 2020. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 295, + 287, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 295, + 287, + 338 + ], + "spans": [ + { + "bbox": [ + 49, + 295, + 287, + 338 + ], + "type": "text", + "content": "[32] Ravi Ramamoorthi and Pat Hanrahan. A signal-processing framework for inverse rendering. In Proceedings of the 28th annual conference on Computer graphics and interactive techniques, pages 117-128, 2001. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 338, + 287, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 338, + 287, + 371 + ], + "spans": [ + { + "bbox": [ + 49, + 338, + 287, + 371 + ], + "type": "text", + "content": "[33] Raskar Ramesh and James Davis. 5d time-light transport matrix: What can we reason about scene properties? Technical report, 2008. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 372, + 286, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 372, + 286, + 404 + ], + "spans": [ + { + "bbox": [ + 49, + 372, + 286, + 404 + ], + "type": "text", + "content": "[34] Fabiano Romeiro and Todd Zickler. Blind reflectometry. In European conference on computer vision, pages 45-58. Springer, 2010. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 405, + 286, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 405, + 286, + 459 + ], + "spans": [ + { + "bbox": [ + 49, + 405, + 286, + 459 + ], + "type": "text", + "content": "[35] Vincent Sitzmann, Semon Rezchikov, Bill Freeman, Josh Tenenbaum, and Fredo Durand. Light field networks: Neural scene representations with single-evaluation rendering. Advances in Neural Information Processing Systems, 34:19313-19325, 2021. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 460, + 286, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 460, + 286, + 503 + ], + "spans": [ + { + "bbox": [ + 49, + 460, + 286, + 503 + ], + "type": "text", + "content": "[36] Shuran Song and Thomas Funkhouser. Neural illumination: Lighting prediction for indoor environments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6918-6926, 2019. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 504, + 286, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 504, + 286, + 568 + ], + "spans": [ + { + "bbox": [ + 49, + 504, + 286, + 568 + ], + "type": "text", + "content": "[37] Pratul P Srinivasan, Ben Mildenhall, Matthew Tancik, Jonathan T Barron, Richard Tucker, and Noah Snavely. Lighthouse: Predicting lighting volumes for spatially-coherent illumination. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8080-8089, 2020. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 571, + 286, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 571, + 286, + 613 + ], + "spans": [ + { + "bbox": [ + 49, + 571, + 286, + 613 + ], + "type": "text", + "content": "[38] R. Swaminathan, M.D. Grossberg, and S.K. Nayar. Cautistics of catadioptric cameras. In Proceedings Eighth IEEE International Conference on Computer Vision. ICCV 2001, volume 2, pages 2-9 vol.2, 2001. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 615, + 286, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 615, + 286, + 667 + ], + "spans": [ + { + "bbox": [ + 49, + 615, + 286, + 667 + ], + "type": "text", + "content": "[39] Tristan Swedish, Connor Henley, and Ramesh Raskar. Objects as cameras: Estimating high-frequency illumination from shadows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2593-2602, 2021. 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 670, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 670, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 49, + 670, + 286, + 713 + ], + "type": "text", + "content": "[40] Yuichi Taguchi, Amit Agrawal, Ashok Veeraraghavan, Srikumar Ramalingam, and Ramesh Raskar. Axial-cones: Modeling spherical catadioptric cameras for wide-angle light field rendering. ACM Trans. Graph., 29(6), dec 2010. 3, 6" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 486 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[41] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T. Barron, and Pratul P. Srinivasan. Ref-NeRF: Structured view-dependent appearance for neural radiance fields. CVPR, 2022. 3, 4, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 118, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 118, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 308, + 118, + 545, + 162 + ], + "type": "text", + "content": "[42] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 163, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 163, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 308, + 163, + 545, + 217 + ], + "type": "text", + "content": "[43] Zian Wang, Jonah Philion, Sanja Fidler, and Jan Kautz. Learning indoor inverse rendering with 3d spatially-varying lighting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12538-12547, 2021. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 219, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 219, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 308, + 219, + 545, + 274 + ], + "type": "text", + "content": "[44] Ziyu Wang, Liao Wang, Fuqiang Zhao, Minye Wu, Lan Xu, and Jingyi Yu. Mirrornerf: One-shot neural portrait radiance field from multi-mirror catadioptric imaging. In 2021 IEEE International Conference on Computational Photography (ICCP), pages 1-12. IEEE, 2021. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 275, + 545, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 275, + 545, + 317 + ], + "spans": [ + { + "bbox": [ + 308, + 275, + 545, + 317 + ], + "type": "text", + "content": "[45] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. In Thirty-Fifth Conference on Neural Information Processing Systems, 2021. 4, 5, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 319, + 545, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 319, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 308, + 319, + 545, + 373 + ], + "type": "text", + "content": "[46] Kai Zhang, Fujun Luan, Qianqian Wang, Kavita Bala, and Noah Snively. Physg: Inverse rendering with spherical gaussians for physics-based material editing and relighting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5453-5462, 2021. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 376, + 545, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 376, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 308, + 376, + 545, + 430 + ], + "type": "text", + "content": "[47] Xiuming Zhang, Pratul P Srinivasan, Boyang Deng, Paul Debevec, William T Freeman, and Jonathan T Barron. Nerfactor: Neural factorization of shape and reflectance under an unknown illumination. ACM Transactions on Graphics (TOG), 40(6):1-18, 2021. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 431, + 545, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 431, + 545, + 486 + ], + "spans": [ + { + "bbox": [ + 308, + 431, + 545, + 486 + ], + "type": "text", + "content": "[48] Rui Zhu, Zhengqin Li, Janarbek Matai, Fatih Porikli, and Manmohan Chandraker. Irisformer: Dense vision transformers for single-image inverse rendering in indoor scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2822-2831, 2022. 3" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20782" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/OReX_ Object Reconstruction From Planar Cross-Sections Using Neural Fields/748d73f6-6f8e-4794-a6a8-f014dea611bf_content_list.json b/2023/OReX_ Object Reconstruction From Planar Cross-Sections Using Neural Fields/748d73f6-6f8e-4794-a6a8-f014dea611bf_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ead6440d77966486f7b254c0a3b58fafe6976814 --- /dev/null +++ b/2023/OReX_ Object Reconstruction From Planar Cross-Sections Using Neural Fields/748d73f6-6f8e-4794-a6a8-f014dea611bf_content_list.json @@ -0,0 +1,1123 @@ +[ + { + "type": "text", + "text": "OReX: Object Reconstruction from Planar Cross-sections Using Neural Fields", + "text_level": 1, + "bbox": [ + 89, + 130, + 879, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Haim Sawdayee \nThe Blavatnik School of Computer Science \nTel Aviv University \nhaimsawdayee@mail.tau.ac.il", + "bbox": [ + 173, + 181, + 517, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Amir Vaxman \nSchool of Informatics \nThe University of Edinburgh \navaxman@inf.ed.ac.uk", + "bbox": [ + 566, + 181, + 795, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Amit H. Bermano \nThe Blavatnik School of Computer Science \nTel Aviv University", + "bbox": [ + 313, + 258, + 656, + 311 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "amberman@tauex.tau.ac.il", + "bbox": [ + 374, + 315, + 589, + 327 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e488693ee20eac151f657d31b60dfa8f6abe8109e87c4f4218c30181db8be148.jpg", + "image_caption": [ + "Figure 1. OREx reconstructs smooth 3D shapes (right) from input planar cross-sections (left). Our neural field-based prior allows smooth interpolation between slices while respecting high-frequency features and self-similarities." + ], + "image_footnote": [], + "bbox": [ + 101, + 342, + 869, + 515 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 564, + 313, + 580 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Reconstructing 3D shapes from planar cross-sections is a challenge inspired by downstream applications like medical imaging and geographic informatics. The input is an in/out indicator function fully defined on a sparse collection of planes in space, and the output is an interpolation of the indicator function to the entire volume. Previous works addressing this sparse and ill-posed problem either produce low quality results, or rely on additional priors such as target topology, appearance information, or input normal directions. In this paper, we present OREx, a method for 3D shape reconstruction from slices alone, featuring a Neural Field as the interpolation prior. A modest neural network is trained on the input planes to return an inside/outside estimate for a given 3D coordinate, yielding a powerful prior that induces smoothness and self-similarities. The main challenge for this approach is high-frequency details, as the neural prior is overly smoothing. To alleviate this, we offer an iterative estimation architecture and a hierarchical input sampling scheme that encourage coarse-to-fine training, allowing the training process to focus on high frequencies at later stages.", + "bbox": [ + 75, + 598, + 473, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In addition, we identify and analyze a ripple-like effect stemming from the mesh extraction step. We mitigate it by regularizing the spatial gradients of the indicator function around input in/out boundaries during network training, tackling the problem at the root. Through extensive qualitative and quantitative experimentation, we demonstrate our method is robust, accurate, and scales well with the size of the input. We report state-of-the-art results compared to previous approaches and recent potential solutions, and demonstrate the benefit of our individual contributions through analysis and ablation studies.", + "bbox": [ + 496, + 566, + 895, + 731 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 762, + 632, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Reconstructing a 3D object from its cross-sections is a long-standing task. It persists in fields including medical imaging, topography mapping, and manufacturing. The typical setting is where a sparse set of arbitrary planes is given, upon which the \"inside\" and \"outside\" regions of the de", + "bbox": [ + 496, + 787, + 895, + 864 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1Code and data available at https://github.com/haimsaw/ ORex", + "bbox": [ + 500, + 875, + 892, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "20854", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "picted domain are labeled, and the entire shape in 3D is to be estimated (see Fig. 1). This is a challenging and ill-posed problem, especially due to the sparse and irregular nature of the data. Classical approaches first localize the problem by constructing an arrangement of the input planes, and then introduce a local regularizer that governs the interpolation of the input to within each cell. While sound, these approaches typically involve simplistic regularization functions, that only interpolate the volume within a cell bounded by the relevant cross-sections; as a consequence, they introduce over-smoothed solutions that do not respect features. In addition, finding a cellular arrangement of planes is a computationally-heavy procedure, adding considerable complexity to the problem and rendering it quickly infeasible for large inputs (see Sec. 4). As we demonstrate (Sec. 4), recent approaches that reconstruct a mesh from an input point cloud are not well suited to our setting, as they assume a rather dense sampling of the entire shape. In addition, these methods do not consider the information of an entire cross-sectional plane, but rather only on the shape boundary.", + "bbox": [ + 76, + 90, + 472, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we introduce $OReX$ a reconstruction approach based on neural networks that estimates an entire shape from its cross-sectional slices. Similar to recent approaches, the neural network constitutes the prior that extrapolates the input to the entire volume. Neural networks in general have already been shown to inherently induce smoothness [17], and self-similarities [12], allowing natural recurrence of patterns. Specifically, we argue that Neural Fields [25] are a promising choice for the task at hand. Neural Fields represent a 3D scene by estimating its density and other local geometric properties for every given 3D coordinate. They are typically trained on 2D planar images, and are required to complete the entire 3D scene according to multiview renderings or photographs. This neural representation is differentiable by construction, and hence allows native geometric optimization of the scene, realized via training. We pose the reconstruction problem as a classification of space into \"in\" and \"out\" regions, which are known for the entire slice planes, and thus generate the indicator function which its decision boundary is the output shape.", + "bbox": [ + 76, + 397, + 472, + 700 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The main challenge with applying neural fields to this problem is high-frequency details. Directly applying established training schemes [19] shows strong spectral bias, yielding overly smoothed results and other artifacts (Fig. 12). Spectral bias is a well-known effect, indicating that higher frequency is effectively learned slower [22]. To facilitate effective high-frequency learning, avoiding the shadow cast by the low frequency, we introduce two alterations. First, we sample the planar data mostly around the inside/outside transition regions, where the frequencies are higher. This sampling is further ordered from low to high-frequency regions (according to the distance from the inside/outside boundary), to encourage a low-to-high-frequency training progression.", + "bbox": [ + 76, + 704, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In addition, we follow recent literature and allow the network to iteratively infer the result, where later iterations are responsible for finer, higher-frequency corrections [1, 23].", + "bbox": [ + 496, + 90, + 893, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Finally, we consider another high-frequency artifact, also found in other neural-field-based works [9]. The desired density (or indicator) function dictates a sharp drop in value at the shape boundary. This is contradictory to the induced neural prior, causing sampling-related artifacts in the downstream task of mesh extraction (Sec. 3.4). To alleviate this, we penalize strong spatial gradients around the boundary contours. This enforces a smoother transition between the in and out regions, allowing higher-quality mesh extraction.", + "bbox": [ + 496, + 136, + 893, + 271 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As we demonstrate (see Fig. 1), our method yields state-of-the-art reconstructions from planar data, both for woman-made and organic shapes. The careful loss and training schemes are validated and analyzed through quantitative and qualitative experimentation. Our method is arrangement-free, and thus both interpolates all data globally, avoiding local artifacts, and scales well to a large number of slices (see Sec. 4).", + "bbox": [ + 496, + 272, + 893, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 500, + 404, + 637, + 420 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Reconstruction from cross-sections Reconstruction from planar cross-section has been a long-standing challenge in geometry processing and computational geometry. The problem was mostly studied in the parallel planes setting (e.g., [2,3,5]), where the 3D object was reconstructed per two such \"keyframes\". Later work offered general solutions for any orientation or distribution of planes [4, 16]. The general approach was based on tessellating the planes into convex cells by the planes of the arrangement and reconstructing the object in each cell, by defining some interpolant inside it. These methods suffer from several issues: the construction of the arrangement is computationally expensive (with at least cubic asymptotic times), and the local reconstruction introduces continuity artifacts (see Sec. 4). Furthermore, the interpolants were usually designed for smoothness or proximity, and failed to capture more global trends in the geometry of the slices, such as twists or extrusions (see Fig. 6). Some works provide topological guarantees [14, 26] or limited the solution to given templates [13], but do not provide a general solution without these priors. Our work is arrangement-free and provides a flexible high-parametric model, using neural networks, to capture the details of the reconstructed object.", + "bbox": [ + 496, + 429, + 893, + 779 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Neural reconstruction Most recent reconstruction works employing neural networks address reconstruction from dense point clouds. Our problem could be cast as a point cloud reconstruction one, after sampling the planes. In this setting, the high-frequency problem is pronounced as in our case, and even more so since more information regarding the target shape is available. Almost all recent ap", + "bbox": [ + 496, + 795, + 895, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "20855", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e369288c32fccc3835d8c30e2514d62bdfa40b82052a65401407687ac1f0cbed.jpg", + "image_caption": [ + "Figure 2. OREx method overview. The input (left) is a set of planar cross-sections with inside/outside information on them. The planes are sampled into points in a hierarchical scheme, and are fed to our iterative Neural Field for training. After training, the final shape is estimated using an off-the-shelf mesh extractor." + ], + "image_footnote": [], + "bbox": [ + 84, + 95, + 890, + 287 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ba8977a5aa5189aa268bec55621ca9fa5bfccf0affd88d5fff97c20cf28f6b66.jpg", + "image_caption": [ + "Figure 3. We train OReX on three types of sampling distributions of the input: A scaled-up version of the 3D convex hull bounds the reconstruction volume (left); Uniform sampling within each plane helps stabilize the learning (middle); most of the samples are concentrated where accuracy matters most—on and around the boundary contours on the input planes (right)." + ], + "image_footnote": [], + "bbox": [ + 89, + 363, + 459, + 486 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "proaches employ an implicit representation of the shape. Some works populate 3D (sometimes adaptive) grids to handle fine-details [7,8,18,24], which is challenging to scale. Other works [6, 21] perform local operations, which rely on a rather dense neighborhood. To alleviate this, a mesh can be directly optimized to match the input [12]. Other neural fields have been employed as well in this context [17], without additional information such as normals; see a recent survey for more approaches [25]. As we demonstrate (Sec. 4), this general line of works employs weaker priors, which do not fit the sparse nature of our problem well.", + "bbox": [ + 75, + 611, + 470, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Perhaps a more similar setting is the one of reconstruction from projections or photographs. Here the reconstructions are from 2D data as well, which is also of sparser nature, albeit using appearance information that we do not have. Literature in this field is deep and wide, with several surveys [10, 11], including the usage of neural fields. We argue our work is orthogonal to that listed here, as it can be plugged in to replace our basic neural-field baseline. We leave for", + "bbox": [ + 75, + 779, + 470, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "future work to inspect the most performing architecture. Addressing the most similar problem setting, concurrent work by Ostonov et al. [20] employs reinforcement learning and orthographic projections to ensure proper reconstruction. Except for requiring additional information during inference, they also require a training phase and are restricted to the domain trained on. Our approach, in contrast, requires no additional information besides the planes.", + "bbox": [ + 496, + 367, + 893, + 489 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 508, + 591, + 523 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We next lay out the details for effective high-quality shape reconstruction from slices. Our approach is based on a neural field, OReXNet. OReXNet outputs $Y(x)$ , an extrapolation of the inside/outside indicator function for a query point $x \\in \\mathbb{R}^3$ (Sec. 3.1). Given a single set of input cross-sections, the network undergoes training to approximate the target function on the input. After training is complete, $Y$ is sampled on the entire volume, and a resulting shape is extracted using the Dual Contouring (DC) approach [15]. This pipeline is depicted in Fig. 2.", + "bbox": [ + 496, + 536, + 893, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As motioned in Sec. 1, the main challenge in reconstruction quality is high-frequency details. The straightforward approach to our problem would be to train a neural field for the desired indicator function by uniformly sampling all planes, and subsequently training a network using a Binary Cross Entropy (BCE) loss on all sampled points. Typical approaches also represent the input coordinates using Positional Encoding (PE) [19]. As it turns out, this approach yields overly smooth results failing even to interpolate the input (Fig. 12). To improve reconstruction fidelity, and allow higher frequency details in the resulting shape, we introduce two alterations to the aforementioned training scheme. We present our hierarchical input sampling scheme in Sec. 3.2, and our iterative-refinement architecture in Sec. 3.3.", + "bbox": [ + 496, + 689, + 895, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "20856", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Finally, we also address a common artifact in implicit mesh extraction [9]. Repeating patterns can be seen that correlate to the sampling pattern of the mesh extraction phase, as demonstrated in Fig. 8. We describe how we design the loss function to mitigate this artifact in Sec. 3.4. See the Supplementary material and code for implementation details and hyper-parameter values.", + "bbox": [ + 75, + 90, + 472, + 196 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Problem Setting", + "text_level": 1, + "bbox": [ + 76, + 207, + 238, + 223 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We consider a set $\\mathcal{P} = \\{P_1,\\dots ,P_k\\}$ of 2D planes embedded in $\\mathbb{R}^3$ , with arbitrary offsets and orientations. Each plane $P_{i}$ contains an arbitrary set of (softly) non-intersecting oriented contours $C_i = \\{c_{i,1},\\dots ,c_{i,l_i}\\}$ that consistently partition the plane into regions of \"inside\" and \"outside\" of an unknown domain $\\Omega \\subset \\mathbb{R}^3$ with boundary $\\partial \\Omega$ (Fig. 2). The target output of our method is an indicator function $Y:\\mathbb{R}^3\\to \\mathbb{R}$ defining $\\Omega$ as:", + "bbox": [ + 75, + 229, + 473, + 351 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nY (x) = \\left\\{ \\begin{array}{l l} 1 & x \\notin \\Omega \\\\ 0 & x \\in \\Omega \\\\ 0. 5 & x \\in \\partial \\Omega \\end{array} \\right. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 363, + 470, + 419 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In practice, we approximate $Y$ using a function $f$ , such that $Y(x) \\approx \\sigma(f(x))$ , where $\\sigma(z) = \\frac{1}{1 + e^{-z}}$ is the sigmoid function.", + "bbox": [ + 76, + 430, + 470, + 474 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Input Sampling", + "text_level": 1, + "bbox": [ + 76, + 486, + 235, + 503 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Regularly or randomly sampling the input planes for training yields inaccurate and overly smooth results (see Fig. 12). Instead, we define three types of relevant point distributions to sample from, as depicted in Fig. 3:", + "bbox": [ + 75, + 508, + 472, + 571 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We bound the reconstructed volume by the 3D convex hull of all the input contours, and scale it up by $5\\%$ . We consequently sample it uniformly, where the sampled points are all outside.", + "2. We compute the bounding box (aligned to the principal axes) of the contours in each plane, and sample it with uniform distribution.", + "3. Most samples are taken around the boundaries of the contours, as these are the regions where accuracy matters most. We then sample every contour edge evenly, and further add off-surface points of varying distances for each edge sample. Off-surface locations are found by moving away from the contour on the plane (i.e., in the direction of the plane normal). We further sample off-surface points in a circle around each vertex." + ], + "bbox": [ + 88, + 580, + 470, + 829 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Each sampling point $x_{i}$ is matched with a label $Y_{i}$ according to its position on the slice, and the pairs $\\{x_{i}, Y_{i}\\}$ constitute the input to our training. See supplementary material for the relevant hyperparameters.", + "bbox": [ + 75, + 839, + 472, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Frequency-Oriented Sampling. In order to encourage better high-frequency learning, we sample the $3^{rd}$ type of points around a set of varying distances, from $0.1\\%$ of the bounding box away from the contours to three orders of magnitude closer. In every epoch, we only use points sampled around three consecutive distance ranges. For early training iterations, we use the three largest distances, since further away points translate to lower frequency information about the shape. As the training progresses, and the lower frequencies are assumed to stabilize, we train with points that grow closer and closer to the actual contour, thus focusing the learning process on higher and higher frequencies (see Fig. 11). See supplementary material for more details and exact scheduling.", + "bbox": [ + 496, + 90, + 893, + 303 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Architecture", + "text_level": 1, + "bbox": [ + 500, + 313, + 633, + 328 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ORexNet is simply an MLP that takes a 3D coordinate as input, represented using positional encoding, and produces the function $Y$ (Sec. 3.1) at any query point. In order to encourage high-frequency details, we introduce an iterative refinement mechanism. Inspired by recent work [1], we allow the network to refine its own results by running them through the network for a small number of iterations. This process was previously shown to produce a coarse-to-fine evolution in the realm of 2D images [1], and we argue it applies to our case as well. As demonstrated in Fig. 2, ORexNet is hence a residual MLP that is fed the result of the previous iteration, and a small hidden state code, along with the query point. These iterations (10 in our experiments) are performed both during training and inference; at inference time, only the last result is taken, and during training, the loss is applied to the results of all steps. The first iteration starts from a learned constant $C$ . We show that this process indeed sharpens features and allows the incorporation of higher-frequency details in Fig. 13.", + "bbox": [ + 496, + 335, + 893, + 625 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Loss Function and Inference", + "text_level": 1, + "bbox": [ + 500, + 635, + 756, + 650 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In order to present our loss function, we must first attend to an issue in the final stage of our pipeline. In this stage, we extract an explicit mesh for the boundary of the indicator function using Dual Contouring [15]. This step uses a discretely sampled version of $Y(x)$ and $\\nabla Y(x)$ on a regular 3D grid to extract the mesh. As can be seen in Fig. 8 (and as witnessed other works [9]), this creates a ripples-like artifact, that correlates to the 3D grid resolution. This results from an aliasing of the sampling in these regions where the gradient magnitudes exhibit high variance. This effect could be mitigated via higher resolution grid sampling, which is expensive, or a post-process operation that may compromise the geometry. Instead, we propose revisiting the loss function of the training process, and incorporating regularization to reduce this sharp drop in value at the shape boundary. First, we omit the activation—in other words, we use the", + "bbox": [ + 496, + 659, + 893, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "20857", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/aaa56d05df9f225105016f491599929176101356d798e60e998657efd31547a8.jpg", + "image_caption": [ + "Figure 4. Results gallery. A qualitative demonstration of a collection of man-made and medical inputs. Zoomed-in viewing is recommended." + ], + "image_footnote": [], + "bbox": [ + 101, + 94, + 869, + 424 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2d05ff135b53b6e06f211fb095439b6119caac82a2573d5bb4c1f8c288ae8250.jpg", + "image_caption": [ + "Figure 5. Additional qualitative results, where transparency reveals the reconstruction of internal cavities and tunnels. Zoomed-in viewing is recommended." + ], + "image_footnote": [], + "bbox": [ + 81, + 472, + 486, + 652 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "function $f(x)$ for contouring, while using $Y(x)$ only for training (Sec 3.1). This de-radicalizes the function values, already offering a softer transition between inside and outside regions. In addition, we explicitly penalize strong transitions by using a hinge loss (Eq. 2). With this, we reduce the gradient magnitude variance considerably by limiting them close to the input shape boundary. Our loss function for a single", + "bbox": [ + 75, + 794, + 470, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "point $x$ is then", + "bbox": [ + 500, + 477, + 599, + 491 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} (x, \\theta) = \\sum_ {i = 0} ^ {N - 1} \\operatorname {B C E} \\left(Y _ {i} (x)\\right) + \\lambda \\max (0, | | \\nabla f _ {N - 1} (x) | | - \\alpha) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 503, + 893, + 556 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Where $\\sigma$ is the sigmoid function, $\\theta$ are the parameters of the network, $BCE()$ is the binary cross-entropy loss applied to all outputs of our iterative scheme, $max(0,x)$ is a hinge loss applied only to the last OReXnet iteration, $N$ the number of iterations, and $\\alpha$ and $\\lambda$ are hyperparameters. Fig. 8 demonstrates the correction effect of the hinge loss.", + "bbox": [ + 496, + 556, + 893, + 648 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 660, + 632, + 678 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To evaluate our method, we developed a prototype, and ran it on a NVIDIA GeForce GTX 1080 Ti. See the Supplementary Material for exact timing and memory consumption statistics. Generally speaking, all training processes took less than four hours on the single GPU.", + "bbox": [ + 496, + 686, + 893, + 762 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Results and Comparisons", + "text_level": 1, + "bbox": [ + 500, + 771, + 732, + 787 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We first qualitatively demonstrate the result of our method on a variety of slice inputs, from both the medical and graphics worlds (Fig. 4). This demonstrates the versatility and generality of our method. We further show our our algorithm correctly reproduces the internal cavities and details of reconstructed objects in Fig. 5. Note how intricate details are learned along side a smooth interpolation between slices,", + "bbox": [ + 496, + 794, + 893, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "20858", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5864da259dc40994c5510a921b02c5ff5ea6b4e1ade1d8d8f76918ba0b663a8e.jpg", + "image_caption": [ + "Figure 6. Qualitative comparisons to other approaches. Our result is smoother, more aligned to features, and with less artifacts. Note how the reconstructed shape silhouettes does not suffer imprints from the input planes. Check Table 1 for quantitative comparison." + ], + "image_footnote": [], + "bbox": [ + 76, + 85, + 836, + 441 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "leaving no slice transition artifacts on the resulting shape.", + "bbox": [ + 75, + 508, + 455, + 523 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In terms of other methods, we qualitatively (Fig. 6) and quantitatively (Table 1) compare our method to Bermano et al. [4], which is a reconstruction from cross-section method with the same input-output as ours. We further compare to general state-of-the-art reconstruction methods that target point clouds [12]. Hanocka et al. [12] expect point normals as input as well. Hence we compare our result to the latter work with Ground Truth (GT) normals at input, and with ones coming from the input plane normals for a fairer comparison. We use the following metrics: 1) intersection-over-union volume of inside regions in reconstructed volumes and 2) GT planes, and 3) symmetric Hausdorff distance. These metrics demonstrate our superiority over the compared methods. See Supplementary Material for more comparisons.", + "bbox": [ + 75, + 523, + 472, + 736 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "**Scalability** We show that our method does not suffer the computational cost that arrangement-based methods [4, 14, 16] must bear, since we do not tessellate the space, and hence scale well with the input. In addition, we show (Fig. 7) how our method converges with the addition of slices.", + "bbox": [ + 75, + 758, + 470, + 835 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Ablations", + "text_level": 1, + "bbox": [ + 76, + 845, + 189, + 861 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We perform ablation testing to evaluate important aspects of our method and our design choices.", + "bbox": [ + 75, + 869, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8b367136416dde2717c7abc95eec0e40af5edfac6287f63649a5b034fecac0f7.jpg", + "image_caption": [ + "Figure 7. Increasing number of slices. As can be expected, introducing more slices converges consistently converges to the solution." + ], + "image_footnote": [], + "bbox": [ + 503, + 510, + 908, + 628 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Gradient magnitude regularization (Fig. 8) We show the effect of the hinge-loss regularizer on gradient magnitudes, with increasing $\\lambda$ values. It is evident the hinge loss effectively filters out the high variance, and smooths the ripple artifacts. As can be seen (bottom row), without our hinge loss, geometry vertices tend to snap to discrete locations during the mesh extraction phase. This is because a large gradient induces a large bias in the magnitude of the indicator function, encouraging vertices to remain close to the grid points sampled during the extraction step. Regularizing the gradients around the boundary regions (top row) decreases the gap between the positive and negative sides of the function, allowing a more uniform, and hence continuous, distribution", + "bbox": [ + 496, + 704, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "20859", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "of the mesh vertices. With this bias removed, the resulting geometry is better sampled and hence the ripple artifact is mitigated (middle).", + "bbox": [ + 76, + 90, + 470, + 136 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/30a8e65595c84cd5c6cc6264b27abefaca22af12dc1ead2e5adfecffdb9659fb.jpg", + "image_caption": [ + "Figure 8. Left to right: our Hinge loss effect for mitigating ripple-like aliasing artifacts. Top: output values of $f(x)$ on a single slice (arrows show gradients). Middle: The reconstructed mesh with a closer view on the top of the skull. Bottom: distribution of vertices along grid edges during mesh extraction. It is evident the hinge loss effectively controls gradient magnitudes at the zero set, curbing the aliasing artifacts. The hinge regularization further allows to place more vertices in the middle of the sampled grid cells, and not only on their edges." + ], + "image_footnote": [], + "bbox": [ + 89, + 154, + 457, + 393 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Iterative architecture (Fig. 9) This ablations examines the effect of the number of refinement iterations on the accuracy of the mesh produced. As can be expected, the results show that as the number of iterations increases, high-frequency details are better represented.", + "bbox": [ + 76, + 570, + 468, + 647 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3852e3adc0901289721be3a1cea9dad6a53f356651fbe67c30786541190b859f.jpg", + "image_caption": [ + "Figure 9. Iterative refinement experiment. Training and inferring with more OReXNet refinement iterations allows the network to perform smaller scale corrections and reduces spectral bias." + ], + "image_footnote": [], + "bbox": [ + 81, + 660, + 465, + 815 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Sampling method To evaluate our importance sampling scheme, we first replace it with regular grid sampling in two resolutions (Fig. 10). As can be seen, using each of the resolutions alone yields overly smoothed results due to spectral bias. Training with a gradually increasing resolution allows better handling of high-frequency details. In addition, we have demonstrate the effects on quality when using our importance scheme in a non-hierarchical manner, i.e. using all points uniformly without consideration of distance from the boundary (Fig. 11). As can be seen, our scheme improves in detail quality upon both experiments, demonstrating that our sampling scheme produces more accurate results with less spectral bias while using fewer samples.", + "bbox": [ + 496, + 90, + 893, + 287 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dd91b74219f224ad277f141b755395c9ce44fc9fb7bd03f183df50409134bf64.jpg", + "image_caption": [ + "Figure 10. Sampling scheme Ablations. Left: ground truth. Middle: grid sampling with different resolutions, and a gradually increasing sampling resolution over the training process. Right: our method." + ], + "image_footnote": [], + "bbox": [ + 506, + 300, + 890, + 465 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/257734fcec337945ae835bb473ac3c885d312e63ab418626c197090aa3c9dc1c.jpg", + "image_caption": [ + "Figure 11. Left: non-hierarchical sampling, using points from all off-surface distances uniformly. Right: using our hierarchical sampling scheme." + ], + "image_footnote": [], + "bbox": [ + 531, + 565, + 893, + 718 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Choice of architecture and sampling (Fig. 12) We justify our design choices by training a baseline architecture with uniform grid sampling and no iteration refinement. It is evident such a model does not preserve details nor interpolate well, even for a relatively simple case.", + "bbox": [ + 496, + 824, + 892, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "20860", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ad33d7f6008402f492e8d468ace478a94ef461f9ccdd9406eddbe3f8a0a51036.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
InputHausdorff distanceIoU in 3dIoU in 2d
ORExBermano et. alpoint2mesh plane normalspoint2mesh GT normalsORExBermano et. alpoint2mesh plane normalspoint2mesh GT normalsORExBermano et. alpoint2mesh plane normalspoint2mesh GT normals
Eight (S)0.0180.0650.2190.0460.9840.8650.8420.9610.9880.9840.7950.980
Eight (M)0.0060.0330.0450.0140.9870.8930.9150.9740.9860.9760.9610.971
Elephant0.0560.0810.1000.0860.9660.9080.8850.9350.9750.9690.8500.908
Balloon dog0.0490.1940.0780.0860.9570.8680.8970.9280.9880.9770.9260.956
Hand OK0.0630.1770.1950.1350.9550.8600.9210.9310.9870.9680.9080.882
Armadillo0.0500.1210.0570.0590.9390.8910.8910.9210.9640.7760.8500.868
", + "bbox": [ + 78, + 89, + 893, + 186 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 1. Quantitative comparisons. We measure performance using Hausdorff distance, IoU of the inner volume compared to the GT shape, and IoU of the inner surface on the input cross-sections. We compare our result to a dedicated cross-section based reconstruction method [4], and to two flavors of a pointcloud reconstruction method [12]. See more comparisons in the Supplementary Material.", + "bbox": [ + 75, + 196, + 893, + 241 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0867482e84d4e25861d89138a15b1e6b5756b1775009965359e10ee8626c7648.jpg", + "image_caption": [ + "Figure 12. Left and middle: the baseline experiment with regular grid samples of 64 and 128 resolution over the slices. Right: our result for the same amount for samples (our variable number of samples is provided at its mean)." + ], + "image_footnote": [], + "bbox": [ + 78, + 284, + 486, + 426 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/73edd476b5527a81cf34571a53dd7d7feabb4a2897acd525b7241e2605c4554a.jpg", + "image_caption": [ + "Figure 13. Refinement progression throughout the iteration process. The maximum change in the output value decreases along the iterations. Right: points that were added to the shape (Red) and that were removed from it (blue) in the last refinement iteration. These are concentrated on the level set, whereas most other points (green, covering all internal regions) have already converged." + ], + "image_footnote": [], + "bbox": [ + 86, + 530, + 455, + 683 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusions", + "text_level": 1, + "bbox": [ + 76, + 813, + 204, + 829 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have presented OREx, a state-of-the-art method for the long-standing problem of shape reconstruction from planar cross-sectional indicator data. Free of datasets and training requirements, OREx is simple and intuitive to use. Our", + "bbox": [ + 75, + 839, + 472, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "work balances the smoothness of a neural prior with high-frequency features. We show our approach successfully produces smooth interpolation between contours while respecting high-frequency features and repeating patterns. In addition, we believe some of the analysis and insights presented here can be applied to neural fields in similar tasks.", + "bbox": [ + 496, + 265, + 893, + 354 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The advantage of our method is also its disadvantage: using only binary data, much of the information is lost in the process. For example, for medical imaging an interesting future direction would be to directly use the raw grayscale density input of a planar probe such as that comes from an ultrasound. Another interesting direction for research is extending this work for multi-labeled volumes (e.g., the reconstruction of several organs simultaneously from a scan), and using partial or noisy data.", + "bbox": [ + 496, + 357, + 893, + 492 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 501, + 506, + 596, + 522 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Yuval Alaluf, Or Patashnik, and Daniel Cohen-Or. Restyle: A residual-based stylegan encoder via iterative refinement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6711–6720, 2021. 2, 4", + "[2] Chandrajit L Bajaj, Edward J Coyle, and Kwun-Nan Lin. Arbitrary topology shape reconstruction from planar cross sections. Graphical models and image processing, 58(6):524-543, 1996. 2", + "[3] Gill Barequet and Micha Sharir. Piecewise-linear interpolation between polygonal slices. In Proceedings of the tenth annual symposium on Computational geometry, pages 93-102, 1994. 2", + "[4] Amit Bermano, Amir Vaxman, and Craig Gotsman. Online reconstruction of 3d objects from arbitrary cross-sections. ACM Transactions on Graphics (TOG), 30(5):1-11, 2011. 2, 6, 8", + "[5] Jean-Daniel Boissonnat and Pooran Memari. Shape reconstruction from unorganized cross-sections. In Symposium on geometry processing, pages 89-98, 2007. 2", + "[6] Alexandre Boulch and Renaud Marlet. Poco: Point convolution for surface reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6302-6314, 2022. 3", + "[7] Zhang Chen, Yinda Zhang, Kyle Genova, Sean Fanello, Sofien Bouaziz, Christian Hane, Ruofei Du, Cem Keskin, Thomas Funkhouser, and Danhang Tang. Multiresolution" + ], + "bbox": [ + 506, + 531, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "20861", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "deep implicit functions for 3d shape representation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13087-13096, 2021. 3", + "[8] Julian Chibane, Thiemo Alldieck, and Gerard Pons-Moll. Implicit functions in feature space for 3d shape reconstruction and completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6970-6981, 2020. 3", + "[9] Philipp Erler, Paul Guerrero, Stefan Ohrhallinger, Niloy J Mitra, and Michael Wimmer. Points2surf learning implicit surfaces from point clouds. In European Conference on Computer Vision, pages 108-124. Springer, 2020. 2, 4", + "[10] Kyle Gao, Yina Gao, Hongjie He, Denning Lu, Linlin Xu, and Jonathan Li. Nerf: Neural radiance field in 3d vision, a comprehensive review. arXiv preprint arXiv:2210.00379, 2022. 3", + "[11] Henry Ham, Julian Wesley, and Hendra Hendra. Computer vision based 3d reconstruction: A review. International Journal of Electrical and Computer Engineering, 9(4):2394, 2019. 3", + "[12] Rana Hanocka, Gal Metzer, Raja Giryes, and Daniel Cohen-Or. Point2mesh: A self-prior for deformable meshes. arXiv preprint arXiv:2005.11084, 2020. 2, 3, 6, 8", + "[13] Michelle Holloway, Cindy Grimm, and Tao Ju. Template-based surface reconstruction from cross-sections. Computers & Graphics, 58:84-91, 2016. Shape Modeling International 2016. 2", + "[14] Zhiyang Huang, Ming Zou, Nathan Carr, and Tao Ju. Topology-controlled reconstruction of multi-labelled domains from cross-sections. ACM Transactions on Graphics (TOG), 36(4):1-12, 2017. 2, 6", + "[15] Tao Ju, Frank Losasso, Scott Schaefer, and Joe Warren. Dual contouring of hermite data. In Proceedings of the 29th annual conference on Computer graphics and interactive techniques, pages 339-346, 2002. 3, 4", + "[16] Lu Liu, Chandrajit Bajaj, Joseph O Deasy, Daniel A Low, and Tao Ju. Surface reconstruction from non-parallel curve networks. In Computer Graphics Forum, volume 27, pages 155–163. Wiley Online Library, 2008. 2, 6", + "[17] Baorui Ma, Zhizhong Han, Yu-Shen Liu, and Matthias Zwicker. Neural-pull: Learning signed distance function from point clouds by learning to pull space onto surface. In International Conference on Machine Learning, pages 7246-7257. PMLR, 2021. 2, 3", + "[18] Julien NP Martel, David B Lindell, Connor Z Lin, Eric R Chan, Marco Monteiro, and Gordon Wetzstein. Acorn: Adaptive coordinate networks for neural scene representation. arXiv preprint arXiv:2105.02788, 2021. 3", + "[19] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2, 3", + "[20] Azimkhon Ostonov. Cut-and-approximate: 3d shape reconstruction from planar cross-sections with deep reinforcement learning. arXiv preprint arXiv:2210.12509, 2022. 3" + ], + "bbox": [ + 78, + 90, + 472, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[21] Songyou Peng, Michael Niemeyer, Lars Mescheder, Marc Pollefeys, and Andreas Geiger. Convolutional occupancy networks. In European Conference on Computer Vision, pages 523-540. Springer, 2020. 3", + "[22] Nasim Rahaman, Aristide Baratin, Devansh Arpit, Felix Draxler, Min Lin, Fred Hamprecht, Yoshua Bengio, and Aaron Courville. On the spectral bias of neural networks. In International Conference on Machine Learning, pages 5301-5310. PMLR, 2019. 2", + "[23] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 2", + "[24] Towaki Takikawa, Joey Litalien, Kangxue Yin, Karsten Kreis, Charles Loop, Derek Nowrouzezahrai, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Neural geometric level of detail: Real-time rendering with implicit 3d shapes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11358-11367, 2021. 3", + "[25] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond. In Computer Graphics Forum, volume 41, pages 641-676. Wiley Online Library, 2022. 2, 3", + "[26] Ming Zou, Michelle Holloway, Nathan Carr, and Tao Ju. Topology-constrained surface reconstruction from cross-sections. ACM Transactions on Graphics (TOG), 34(4):1-10, 2015. 2" + ], + "bbox": [ + 501, + 92, + 893, + 470 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "20862", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + } +] \ No newline at end of file diff --git a/2023/OReX_ Object Reconstruction From Planar Cross-Sections Using Neural Fields/748d73f6-6f8e-4794-a6a8-f014dea611bf_model.json b/2023/OReX_ Object Reconstruction From Planar Cross-Sections Using Neural Fields/748d73f6-6f8e-4794-a6a8-f014dea611bf_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7827079e935078ef3850991d855880c8f1b954d4 --- /dev/null +++ b/2023/OReX_ Object Reconstruction From Planar Cross-Sections Using Neural Fields/748d73f6-6f8e-4794-a6a8-f014dea611bf_model.json @@ -0,0 +1,1505 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.131, + 0.88, + 0.153 + ], + "angle": 0, + "content": "OReX: Object Reconstruction from Planar Cross-sections Using Neural Fields" + }, + { + "type": "text", + "bbox": [ + 0.174, + 0.182, + 0.518, + 0.251 + ], + "angle": 0, + "content": "Haim Sawdayee \nThe Blavatnik School of Computer Science \nTel Aviv University \nhaimsawdayee@mail.tau.ac.il" + }, + { + "type": "text", + "bbox": [ + 0.567, + 0.182, + 0.796, + 0.251 + ], + "angle": 0, + "content": "Amir Vaxman \nSchool of Informatics \nThe University of Edinburgh \navaxman@inf.ed.ac.uk" + }, + { + "type": "text", + "bbox": [ + 0.314, + 0.26, + 0.658, + 0.313 + ], + "angle": 0, + "content": "Amit H. Bermano \nThe Blavatnik School of Computer Science \nTel Aviv University" + }, + { + "type": "text", + "bbox": [ + 0.375, + 0.316, + 0.591, + 0.328 + ], + "angle": 0, + "content": "amberman@tauex.tau.ac.il" + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.343, + 0.87, + 0.516 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.522, + 0.893, + 0.551 + ], + "angle": 0, + "content": "Figure 1. OREx reconstructs smooth 3D shapes (right) from input planar cross-sections (left). Our neural field-based prior allows smooth interpolation between slices while respecting high-frequency features and self-similarities." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.565, + 0.314, + 0.581 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.474, + 0.903 + ], + "angle": 0, + "content": "Reconstructing 3D shapes from planar cross-sections is a challenge inspired by downstream applications like medical imaging and geographic informatics. The input is an in/out indicator function fully defined on a sparse collection of planes in space, and the output is an interpolation of the indicator function to the entire volume. Previous works addressing this sparse and ill-posed problem either produce low quality results, or rely on additional priors such as target topology, appearance information, or input normal directions. In this paper, we present OREx, a method for 3D shape reconstruction from slices alone, featuring a Neural Field as the interpolation prior. A modest neural network is trained on the input planes to return an inside/outside estimate for a given 3D coordinate, yielding a powerful prior that induces smoothness and self-similarities. The main challenge for this approach is high-frequency details, as the neural prior is overly smoothing. To alleviate this, we offer an iterative estimation architecture and a hierarchical input sampling scheme that encourage coarse-to-fine training, allowing the training process to focus on high frequencies at later stages." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.567, + 0.896, + 0.732 + ], + "angle": 0, + "content": "In addition, we identify and analyze a ripple-like effect stemming from the mesh extraction step. We mitigate it by regularizing the spatial gradients of the indicator function around input in/out boundaries during network training, tackling the problem at the root. Through extensive qualitative and quantitative experimentation, we demonstrate our method is robust, accurate, and scales well with the size of the input. We report state-of-the-art results compared to previous approaches and recent potential solutions, and demonstrate the benefit of our individual contributions through analysis and ablation studies." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.763, + 0.633, + 0.779 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.789, + 0.896, + 0.866 + ], + "angle": 0, + "content": "Reconstructing a 3D object from its cross-sections is a long-standing task. It persists in fields including medical imaging, topography mapping, and manufacturing. The typical setting is where a sparse set of arbitrary planes is given, upon which the \"inside\" and \"outside\" regions of the de" + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.875, + 0.893, + 0.9 + ], + "angle": 0, + "content": "1Code and data available at https://github.com/haimsaw/ ORex" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20854" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.091, + 0.473, + 0.395 + ], + "angle": 0, + "content": "picted domain are labeled, and the entire shape in 3D is to be estimated (see Fig. 1). This is a challenging and ill-posed problem, especially due to the sparse and irregular nature of the data. Classical approaches first localize the problem by constructing an arrangement of the input planes, and then introduce a local regularizer that governs the interpolation of the input to within each cell. While sound, these approaches typically involve simplistic regularization functions, that only interpolate the volume within a cell bounded by the relevant cross-sections; as a consequence, they introduce over-smoothed solutions that do not respect features. In addition, finding a cellular arrangement of planes is a computationally-heavy procedure, adding considerable complexity to the problem and rendering it quickly infeasible for large inputs (see Sec. 4). As we demonstrate (Sec. 4), recent approaches that reconstruct a mesh from an input point cloud are not well suited to our setting, as they assume a rather dense sampling of the entire shape. In addition, these methods do not consider the information of an entire cross-sectional plane, but rather only on the shape boundary." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.398, + 0.473, + 0.701 + ], + "angle": 0, + "content": "In this paper, we introduce \\( OReX \\) a reconstruction approach based on neural networks that estimates an entire shape from its cross-sectional slices. Similar to recent approaches, the neural network constitutes the prior that extrapolates the input to the entire volume. Neural networks in general have already been shown to inherently induce smoothness [17], and self-similarities [12], allowing natural recurrence of patterns. Specifically, we argue that Neural Fields [25] are a promising choice for the task at hand. Neural Fields represent a 3D scene by estimating its density and other local geometric properties for every given 3D coordinate. They are typically trained on 2D planar images, and are required to complete the entire 3D scene according to multiview renderings or photographs. This neural representation is differentiable by construction, and hence allows native geometric optimization of the scene, realized via training. We pose the reconstruction problem as a classification of space into \"in\" and \"out\" regions, which are known for the entire slice planes, and thus generate the indicator function which its decision boundary is the output shape." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.705, + 0.473, + 0.903 + ], + "angle": 0, + "content": "The main challenge with applying neural fields to this problem is high-frequency details. Directly applying established training schemes [19] shows strong spectral bias, yielding overly smoothed results and other artifacts (Fig. 12). Spectral bias is a well-known effect, indicating that higher frequency is effectively learned slower [22]. To facilitate effective high-frequency learning, avoiding the shadow cast by the low frequency, we introduce two alterations. First, we sample the planar data mostly around the inside/outside transition regions, where the frequencies are higher. This sampling is further ordered from low to high-frequency regions (according to the distance from the inside/outside boundary), to encourage a low-to-high-frequency training progression." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.894, + 0.137 + ], + "angle": 0, + "content": "In addition, we follow recent literature and allow the network to iteratively infer the result, where later iterations are responsible for finer, higher-frequency corrections [1, 23]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.137, + 0.895, + 0.272 + ], + "angle": 0, + "content": "Finally, we consider another high-frequency artifact, also found in other neural-field-based works [9]. The desired density (or indicator) function dictates a sharp drop in value at the shape boundary. This is contradictory to the induced neural prior, causing sampling-related artifacts in the downstream task of mesh extraction (Sec. 3.4). To alleviate this, we penalize strong spatial gradients around the boundary contours. This enforces a smoother transition between the in and out regions, allowing higher-quality mesh extraction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.273, + 0.895, + 0.394 + ], + "angle": 0, + "content": "As we demonstrate (see Fig. 1), our method yields state-of-the-art reconstructions from planar data, both for woman-made and organic shapes. The careful loss and training schemes are validated and analyzed through quantitative and qualitative experimentation. Our method is arrangement-free, and thus both interpolates all data globally, avoiding local artifacts, and scales well to a large number of slices (see Sec. 4)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.405, + 0.638, + 0.421 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.43, + 0.895, + 0.78 + ], + "angle": 0, + "content": "Reconstruction from cross-sections Reconstruction from planar cross-section has been a long-standing challenge in geometry processing and computational geometry. The problem was mostly studied in the parallel planes setting (e.g., [2,3,5]), where the 3D object was reconstructed per two such \"keyframes\". Later work offered general solutions for any orientation or distribution of planes [4, 16]. The general approach was based on tessellating the planes into convex cells by the planes of the arrangement and reconstructing the object in each cell, by defining some interpolant inside it. These methods suffer from several issues: the construction of the arrangement is computationally expensive (with at least cubic asymptotic times), and the local reconstruction introduces continuity artifacts (see Sec. 4). Furthermore, the interpolants were usually designed for smoothness or proximity, and failed to capture more global trends in the geometry of the slices, such as twists or extrusions (see Fig. 6). Some works provide topological guarantees [14, 26] or limited the solution to given templates [13], but do not provide a general solution without these priors. Our work is arrangement-free and provides a flexible high-parametric model, using neural networks, to capture the details of the reconstructed object." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.897, + 0.903 + ], + "angle": 0, + "content": "Neural reconstruction Most recent reconstruction works employing neural networks address reconstruction from dense point clouds. Our problem could be cast as a point cloud reconstruction one, after sampling the planes. In this setting, the high-frequency problem is pronounced as in our case, and even more so since more information regarding the target shape is available. Almost all recent ap" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20855" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.096, + 0.891, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.299, + 0.893, + 0.342 + ], + "angle": 0, + "content": "Figure 2. OREx method overview. The input (left) is a set of planar cross-sections with inside/outside information on them. The planes are sampled into points in a hierarchical scheme, and are fed to our iterative Neural Field for training. After training, the final shape is estimated using an off-the-shelf mesh extractor." + }, + { + "type": "image", + "bbox": [ + 0.091, + 0.364, + 0.46, + 0.487 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.496, + 0.471, + 0.581 + ], + "angle": 0, + "content": "Figure 3. We train OReX on three types of sampling distributions of the input: A scaled-up version of the 3D convex hull bounds the reconstruction volume (left); Uniform sampling within each plane helps stabilize the learning (middle); most of the samples are concentrated where accuracy matters most—on and around the boundary contours on the input planes (right)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.612, + 0.472, + 0.779 + ], + "angle": 0, + "content": "proaches employ an implicit representation of the shape. Some works populate 3D (sometimes adaptive) grids to handle fine-details [7,8,18,24], which is challenging to scale. Other works [6, 21] perform local operations, which rely on a rather dense neighborhood. To alleviate this, a mesh can be directly optimized to match the input [12]. Other neural fields have been employed as well in this context [17], without additional information such as normals; see a recent survey for more approaches [25]. As we demonstrate (Sec. 4), this general line of works employs weaker priors, which do not fit the sparse nature of our problem well." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.472, + 0.901 + ], + "angle": 0, + "content": "Perhaps a more similar setting is the one of reconstruction from projections or photographs. Here the reconstructions are from 2D data as well, which is also of sparser nature, albeit using appearance information that we do not have. Literature in this field is deep and wide, with several surveys [10, 11], including the usage of neural fields. We argue our work is orthogonal to that listed here, as it can be plugged in to replace our basic neural-field baseline. We leave for" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.368, + 0.895, + 0.49 + ], + "angle": 0, + "content": "future work to inspect the most performing architecture. Addressing the most similar problem setting, concurrent work by Ostonov et al. [20] employs reinforcement learning and orthographic projections to ensure proper reconstruction. Except for requiring additional information during inference, they also require a training phase and are restricted to the domain trained on. Our approach, in contrast, requires no additional information besides the planes." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.509, + 0.593, + 0.525 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.537, + 0.895, + 0.688 + ], + "angle": 0, + "content": "We next lay out the details for effective high-quality shape reconstruction from slices. Our approach is based on a neural field, OReXNet. OReXNet outputs \\( Y(x) \\), an extrapolation of the inside/outside indicator function for a query point \\( x \\in \\mathbb{R}^3 \\) (Sec. 3.1). Given a single set of input cross-sections, the network undergoes training to approximate the target function on the input. After training is complete, \\( Y \\) is sampled on the entire volume, and a resulting shape is extracted using the Dual Contouring (DC) approach [15]. This pipeline is depicted in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.896, + 0.901 + ], + "angle": 0, + "content": "As motioned in Sec. 1, the main challenge in reconstruction quality is high-frequency details. The straightforward approach to our problem would be to train a neural field for the desired indicator function by uniformly sampling all planes, and subsequently training a network using a Binary Cross Entropy (BCE) loss on all sampled points. Typical approaches also represent the input coordinates using Positional Encoding (PE) [19]. As it turns out, this approach yields overly smooth results failing even to interpolate the input (Fig. 12). To improve reconstruction fidelity, and allow higher frequency details in the resulting shape, we introduce two alterations to the aforementioned training scheme. We present our hierarchical input sampling scheme in Sec. 3.2, and our iterative-refinement architecture in Sec. 3.3." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20856" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.198 + ], + "angle": 0, + "content": "Finally, we also address a common artifact in implicit mesh extraction [9]. Repeating patterns can be seen that correlate to the sampling pattern of the mesh extraction phase, as demonstrated in Fig. 8. We describe how we design the loss function to mitigate this artifact in Sec. 3.4. See the Supplementary material and code for implementation details and hyper-parameter values." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.208, + 0.24, + 0.224 + ], + "angle": 0, + "content": "3.1. Problem Setting" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.231, + 0.474, + 0.352 + ], + "angle": 0, + "content": "We consider a set \\(\\mathcal{P} = \\{P_1,\\dots ,P_k\\}\\) of 2D planes embedded in \\(\\mathbb{R}^3\\) , with arbitrary offsets and orientations. Each plane \\(P_{i}\\) contains an arbitrary set of (softly) non-intersecting oriented contours \\(C_i = \\{c_{i,1},\\dots ,c_{i,l_i}\\}\\) that consistently partition the plane into regions of \"inside\" and \"outside\" of an unknown domain \\(\\Omega \\subset \\mathbb{R}^3\\) with boundary \\(\\partial \\Omega\\) (Fig. 2). The target output of our method is an indicator function \\(Y:\\mathbb{R}^3\\to \\mathbb{R}\\) defining \\(\\Omega\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.364, + 0.471, + 0.42 + ], + "angle": 0, + "content": "\\[\nY (x) = \\left\\{ \\begin{array}{l l} 1 & x \\notin \\Omega \\\\ 0 & x \\in \\Omega \\\\ 0. 5 & x \\in \\partial \\Omega \\end{array} \\right. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.431, + 0.471, + 0.476 + ], + "angle": 0, + "content": "In practice, we approximate \\( Y \\) using a function \\( f \\), such that \\( Y(x) \\approx \\sigma(f(x)) \\), where \\( \\sigma(z) = \\frac{1}{1 + e^{-z}} \\) is the sigmoid function." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.487, + 0.236, + 0.504 + ], + "angle": 0, + "content": "3.2. Input Sampling" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.51, + 0.473, + 0.572 + ], + "angle": 0, + "content": "Regularly or randomly sampling the input planes for training yields inaccurate and overly smooth results (see Fig. 12). Instead, we define three types of relevant point distributions to sample from, as depicted in Fig. 3:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.581, + 0.47, + 0.642 + ], + "angle": 0, + "content": "1. We bound the reconstructed volume by the 3D convex hull of all the input contours, and scale it up by \\(5\\%\\). We consequently sample it uniformly, where the sampled points are all outside." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.653, + 0.47, + 0.697 + ], + "angle": 0, + "content": "2. We compute the bounding box (aligned to the principal axes) of the contours in each plane, and sample it with uniform distribution." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.709, + 0.471, + 0.83 + ], + "angle": 0, + "content": "3. Most samples are taken around the boundaries of the contours, as these are the regions where accuracy matters most. We then sample every contour edge evenly, and further add off-surface points of varying distances for each edge sample. Off-surface locations are found by moving away from the contour on the plane (i.e., in the direction of the plane normal). We further sample off-surface points in a circle around each vertex." + }, + { + "type": "list", + "bbox": [ + 0.089, + 0.581, + 0.471, + 0.83 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Each sampling point \\( x_{i} \\) is matched with a label \\( Y_{i} \\) according to its position on the slice, and the pairs \\( \\{x_{i}, Y_{i}\\} \\) constitute the input to our training. See supplementary material for the relevant hyperparameters." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.304 + ], + "angle": 0, + "content": "Frequency-Oriented Sampling. In order to encourage better high-frequency learning, we sample the \\(3^{rd}\\) type of points around a set of varying distances, from \\(0.1\\%\\) of the bounding box away from the contours to three orders of magnitude closer. In every epoch, we only use points sampled around three consecutive distance ranges. For early training iterations, we use the three largest distances, since further away points translate to lower frequency information about the shape. As the training progresses, and the lower frequencies are assumed to stabilize, we train with points that grow closer and closer to the actual contour, thus focusing the learning process on higher and higher frequencies (see Fig. 11). See supplementary material for more details and exact scheduling." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.314, + 0.634, + 0.329 + ], + "angle": 0, + "content": "3.3. Architecture" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.337, + 0.895, + 0.626 + ], + "angle": 0, + "content": "ORexNet is simply an MLP that takes a 3D coordinate as input, represented using positional encoding, and produces the function \\( Y \\) (Sec. 3.1) at any query point. In order to encourage high-frequency details, we introduce an iterative refinement mechanism. Inspired by recent work [1], we allow the network to refine its own results by running them through the network for a small number of iterations. This process was previously shown to produce a coarse-to-fine evolution in the realm of 2D images [1], and we argue it applies to our case as well. As demonstrated in Fig. 2, ORexNet is hence a residual MLP that is fed the result of the previous iteration, and a small hidden state code, along with the query point. These iterations (10 in our experiments) are performed both during training and inference; at inference time, only the last result is taken, and during training, the loss is applied to the results of all steps. The first iteration starts from a learned constant \\( C \\). We show that this process indeed sharpens features and allows the incorporation of higher-frequency details in Fig. 13." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.636, + 0.757, + 0.651 + ], + "angle": 0, + "content": "3.4. Loss Function and Inference" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.895, + 0.901 + ], + "angle": 0, + "content": "In order to present our loss function, we must first attend to an issue in the final stage of our pipeline. In this stage, we extract an explicit mesh for the boundary of the indicator function using Dual Contouring [15]. This step uses a discretely sampled version of \\( Y(x) \\) and \\( \\nabla Y(x) \\) on a regular 3D grid to extract the mesh. As can be seen in Fig. 8 (and as witnessed other works [9]), this creates a ripples-like artifact, that correlates to the 3D grid resolution. This results from an aliasing of the sampling in these regions where the gradient magnitudes exhibit high variance. This effect could be mitigated via higher resolution grid sampling, which is expensive, or a post-process operation that may compromise the geometry. Instead, we propose revisiting the loss function of the training process, and incorporating regularization to reduce this sharp drop in value at the shape boundary. First, we omit the activation—in other words, we use the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20857" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.095, + 0.87, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.436, + 0.892, + 0.451 + ], + "angle": 0, + "content": "Figure 4. Results gallery. A qualitative demonstration of a collection of man-made and medical inputs. Zoomed-in viewing is recommended." + }, + { + "type": "image", + "bbox": [ + 0.083, + 0.473, + 0.487, + 0.653 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.664, + 0.47, + 0.706 + ], + "angle": 0, + "content": "Figure 5. Additional qualitative results, where transparency reveals the reconstruction of internal cavities and tunnels. Zoomed-in viewing is recommended." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.471, + 0.903 + ], + "angle": 0, + "content": "function \\( f(x) \\) for contouring, while using \\( Y(x) \\) only for training (Sec 3.1). This de-radicalizes the function values, already offering a softer transition between inside and outside regions. In addition, we explicitly penalize strong transitions by using a hinge loss (Eq. 2). With this, we reduce the gradient magnitude variance considerably by limiting them close to the input shape boundary. Our loss function for a single" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.478, + 0.6, + 0.492 + ], + "angle": 0, + "content": "point \\(x\\) is then" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.504, + 0.894, + 0.557 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} (x, \\theta) = \\sum_ {i = 0} ^ {N - 1} \\operatorname {B C E} \\left(Y _ {i} (x)\\right) + \\lambda \\max (0, | | \\nabla f _ {N - 1} (x) | | - \\alpha) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.558, + 0.894, + 0.649 + ], + "angle": 0, + "content": "Where \\(\\sigma\\) is the sigmoid function, \\(\\theta\\) are the parameters of the network, \\(BCE()\\) is the binary cross-entropy loss applied to all outputs of our iterative scheme, \\(max(0,x)\\) is a hinge loss applied only to the last OReXnet iteration, \\(N\\) the number of iterations, and \\(\\alpha\\) and \\(\\lambda\\) are hyperparameters. Fig. 8 demonstrates the correction effect of the hinge loss." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.661, + 0.633, + 0.679 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.687, + 0.894, + 0.763 + ], + "angle": 0, + "content": "To evaluate our method, we developed a prototype, and ran it on a NVIDIA GeForce GTX 1080 Ti. See the Supplementary Material for exact timing and memory consumption statistics. Generally speaking, all training processes took less than four hours on the single GPU." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.772, + 0.733, + 0.789 + ], + "angle": 0, + "content": "4.1. Results and Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.895, + 0.901 + ], + "angle": 0, + "content": "We first qualitatively demonstrate the result of our method on a variety of slice inputs, from both the medical and graphics worlds (Fig. 4). This demonstrates the versatility and generality of our method. We further show our our algorithm correctly reproduces the internal cavities and details of reconstructed objects in Fig. 5. Note how intricate details are learned along side a smooth interpolation between slices," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20858" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.086, + 0.837, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.453, + 0.895, + 0.484 + ], + "angle": 0, + "content": "Figure 6. Qualitative comparisons to other approaches. Our result is smoother, more aligned to features, and with less artifacts. Note how the reconstructed shape silhouettes does not suffer imprints from the input planes. Check Table 1 for quantitative comparison." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.509, + 0.457, + 0.524 + ], + "angle": 0, + "content": "leaving no slice transition artifacts on the resulting shape." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.525, + 0.473, + 0.737 + ], + "angle": 0, + "content": "In terms of other methods, we qualitatively (Fig. 6) and quantitatively (Table 1) compare our method to Bermano et al. [4], which is a reconstruction from cross-section method with the same input-output as ours. We further compare to general state-of-the-art reconstruction methods that target point clouds [12]. Hanocka et al. [12] expect point normals as input as well. Hence we compare our result to the latter work with Ground Truth (GT) normals at input, and with ones coming from the input plane normals for a fairer comparison. We use the following metrics: 1) intersection-over-union volume of inside regions in reconstructed volumes and 2) GT planes, and 3) symmetric Hausdorff distance. These metrics demonstrate our superiority over the compared methods. See Supplementary Material for more comparisons." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.759, + 0.472, + 0.837 + ], + "angle": 0, + "content": "**Scalability** We show that our method does not suffer the computational cost that arrangement-based methods [4, 14, 16] must bear, since we do not tessellate the space, and hence scale well with the input. In addition, we show (Fig. 7) how our method converges with the addition of slices." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.847, + 0.19, + 0.862 + ], + "angle": 0, + "content": "4.2. Ablations" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.471, + 0.903 + ], + "angle": 0, + "content": "We perform ablation testing to evaluate important aspects of our method and our design choices." + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.511, + 0.91, + 0.629 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.643, + 0.895, + 0.673 + ], + "angle": 0, + "content": "Figure 7. Increasing number of slices. As can be expected, introducing more slices converges consistently converges to the solution." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Gradient magnitude regularization (Fig. 8) We show the effect of the hinge-loss regularizer on gradient magnitudes, with increasing \\(\\lambda\\) values. It is evident the hinge loss effectively filters out the high variance, and smooths the ripple artifacts. As can be seen (bottom row), without our hinge loss, geometry vertices tend to snap to discrete locations during the mesh extraction phase. This is because a large gradient induces a large bias in the magnitude of the indicator function, encouraging vertices to remain close to the grid points sampled during the extraction step. Regularizing the gradients around the boundary regions (top row) decreases the gap between the positive and negative sides of the function, allowing a more uniform, and hence continuous, distribution" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20859" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.471, + 0.137 + ], + "angle": 0, + "content": "of the mesh vertices. With this bias removed, the resulting geometry is better sampled and hence the ripple artifact is mitigated (middle)." + }, + { + "type": "image", + "bbox": [ + 0.091, + 0.155, + 0.458, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.408, + 0.471, + 0.533 + ], + "angle": 0, + "content": "Figure 8. Left to right: our Hinge loss effect for mitigating ripple-like aliasing artifacts. Top: output values of \\( f(x) \\) on a single slice (arrows show gradients). Middle: The reconstructed mesh with a closer view on the top of the skull. Bottom: distribution of vertices along grid edges during mesh extraction. It is evident the hinge loss effectively controls gradient magnitudes at the zero set, curbing the aliasing artifacts. The hinge regularization further allows to place more vertices in the middle of the sampled grid cells, and not only on their edges." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.571, + 0.47, + 0.648 + ], + "angle": 0, + "content": "Iterative architecture (Fig. 9) This ablations examines the effect of the number of refinement iterations on the accuracy of the mesh produced. As can be expected, the results show that as the number of iterations increases, high-frequency details are better represented." + }, + { + "type": "image", + "bbox": [ + 0.083, + 0.661, + 0.466, + 0.816 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.839, + 0.47, + 0.882 + ], + "angle": 0, + "content": "Figure 9. Iterative refinement experiment. Training and inferring with more OReXNet refinement iterations allows the network to perform smaller scale corrections and reduces spectral bias." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.289 + ], + "angle": 0, + "content": "Sampling method To evaluate our importance sampling scheme, we first replace it with regular grid sampling in two resolutions (Fig. 10). As can be seen, using each of the resolutions alone yields overly smoothed results due to spectral bias. Training with a gradually increasing resolution allows better handling of high-frequency details. In addition, we have demonstrate the effects on quality when using our importance scheme in a non-hierarchical manner, i.e. using all points uniformly without consideration of distance from the boundary (Fig. 11). As can be seen, our scheme improves in detail quality upon both experiments, demonstrating that our sampling scheme produces more accurate results with less spectral bias while using fewer samples." + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.301, + 0.891, + 0.467 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.481, + 0.894, + 0.523 + ], + "angle": 0, + "content": "Figure 10. Sampling scheme Ablations. Left: ground truth. Middle: grid sampling with different resolutions, and a gradually increasing sampling resolution over the training process. Right: our method." + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.566, + 0.895, + 0.719 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.743, + 0.893, + 0.786 + ], + "angle": 0, + "content": "Figure 11. Left: non-hierarchical sampling, using points from all off-surface distances uniformly. Right: using our hierarchical sampling scheme." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Choice of architecture and sampling (Fig. 12) We justify our design choices by training a baseline architecture with uniform grid sampling and no iteration refinement. It is evident such a model does not preserve details nor interpolate well, even for a relatively simple case." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20860" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.079, + 0.09, + 0.894, + 0.188 + ], + "angle": 0, + "content": "
InputHausdorff distanceIoU in 3dIoU in 2d
ORExBermano et. alpoint2mesh plane normalspoint2mesh GT normalsORExBermano et. alpoint2mesh plane normalspoint2mesh GT normalsORExBermano et. alpoint2mesh plane normalspoint2mesh GT normals
Eight (S)0.0180.0650.2190.0460.9840.8650.8420.9610.9880.9840.7950.980
Eight (M)0.0060.0330.0450.0140.9870.8930.9150.9740.9860.9760.9610.971
Elephant0.0560.0810.1000.0860.9660.9080.8850.9350.9750.9690.8500.908
Balloon dog0.0490.1940.0780.0860.9570.8680.8970.9280.9880.9770.9260.956
Hand OK0.0630.1770.1950.1350.9550.8600.9210.9310.9870.9680.9080.882
Armadillo0.0500.1210.0570.0590.9390.8910.8910.9210.9640.7760.8500.868
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.198, + 0.895, + 0.242 + ], + "angle": 0, + "content": "Table 1. Quantitative comparisons. We measure performance using Hausdorff distance, IoU of the inner volume compared to the GT shape, and IoU of the inner surface on the input cross-sections. We compare our result to a dedicated cross-section based reconstruction method [4], and to two flavors of a pointcloud reconstruction method [12]. See more comparisons in the Supplementary Material." + }, + { + "type": "image", + "bbox": [ + 0.079, + 0.285, + 0.487, + 0.427 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.452, + 0.472, + 0.51 + ], + "angle": 0, + "content": "Figure 12. Left and middle: the baseline experiment with regular grid samples of 64 and 128 resolution over the slices. Right: our result for the same amount for samples (our variable number of samples is provided at its mean)." + }, + { + "type": "image", + "bbox": [ + 0.087, + 0.531, + 0.456, + 0.684 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.703, + 0.472, + 0.788 + ], + "angle": 0, + "content": "Figure 13. Refinement progression throughout the iteration process. The maximum change in the output value decreases along the iterations. Right: points that were added to the shape (Red) and that were removed from it (blue) in the last refinement iteration. These are concentrated on the level set, whereas most other points (green, covering all internal regions) have already converged." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.814, + 0.205, + 0.83 + ], + "angle": 0, + "content": "5. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.473, + 0.903 + ], + "angle": 0, + "content": "We have presented OREx, a state-of-the-art method for the long-standing problem of shape reconstruction from planar cross-sectional indicator data. Free of datasets and training requirements, OREx is simple and intuitive to use. Our" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.266, + 0.895, + 0.356 + ], + "angle": 0, + "content": "work balances the smoothness of a neural prior with high-frequency features. We show our approach successfully produces smooth interpolation between contours while respecting high-frequency features and repeating patterns. In addition, we believe some of the analysis and insights presented here can be applied to neural fields in similar tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.358, + 0.895, + 0.493 + ], + "angle": 0, + "content": "The advantage of our method is also its disadvantage: using only binary data, much of the information is lost in the process. For example, for medical imaging an interesting future direction would be to directly use the raw grayscale density input of a planar probe such as that comes from an ultrasound. Another interesting direction for research is extending this work for multi-labeled volumes (e.g., the reconstruction of several organs simultaneously from a scan), and using partial or noisy data." + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.507, + 0.597, + 0.523 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.508, + 0.532, + 0.895, + 0.588 + ], + "angle": 0, + "content": "[1] Yuval Alaluf, Or Patashnik, and Daniel Cohen-Or. Restyle: A residual-based stylegan encoder via iterative refinement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6711–6720, 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.589, + 0.895, + 0.644 + ], + "angle": 0, + "content": "[2] Chandrajit L Bajaj, Edward J Coyle, and Kwun-Nan Lin. Arbitrary topology shape reconstruction from planar cross sections. Graphical models and image processing, 58(6):524-543, 1996. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.646, + 0.895, + 0.701 + ], + "angle": 0, + "content": "[3] Gill Barequet and Micha Sharir. Piecewise-linear interpolation between polygonal slices. In Proceedings of the tenth annual symposium on Computational geometry, pages 93-102, 1994. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.703, + 0.895, + 0.757 + ], + "angle": 0, + "content": "[4] Amit Bermano, Amir Vaxman, and Craig Gotsman. Online reconstruction of 3d objects from arbitrary cross-sections. ACM Transactions on Graphics (TOG), 30(5):1-11, 2011. 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.76, + 0.894, + 0.801 + ], + "angle": 0, + "content": "[5] Jean-Daniel Boissonnat and Pooran Memari. Shape reconstruction from unorganized cross-sections. In Symposium on geometry processing, pages 89-98, 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.802, + 0.895, + 0.858 + ], + "angle": 0, + "content": "[6] Alexandre Boulch and Renaud Marlet. Poco: Point convolution for surface reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6302-6314, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.509, + 0.86, + 0.895, + 0.901 + ], + "angle": 0, + "content": "[7] Zhang Chen, Yinda Zhang, Kyle Genova, Sean Fanello, Sofien Bouaziz, Christian Hane, Ruofei Du, Cem Keskin, Thomas Funkhouser, and Danhang Tang. Multiresolution" + }, + { + "type": "list", + "bbox": [ + 0.508, + 0.532, + 0.895, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20861" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.472, + 0.135 + ], + "angle": 0, + "content": "deep implicit functions for 3d shape representation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13087-13096, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.137, + 0.473, + 0.207 + ], + "angle": 0, + "content": "[8] Julian Chibane, Thiemo Alldieck, and Gerard Pons-Moll. Implicit functions in feature space for 3d shape reconstruction and completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6970-6981, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.208, + 0.473, + 0.264 + ], + "angle": 0, + "content": "[9] Philipp Erler, Paul Guerrero, Stefan Ohrhallinger, Niloy J Mitra, and Michael Wimmer. Points2surf learning implicit surfaces from point clouds. In European Conference on Computer Vision, pages 108-124. Springer, 2020. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.266, + 0.472, + 0.322 + ], + "angle": 0, + "content": "[10] Kyle Gao, Yina Gao, Hongjie He, Denning Lu, Linlin Xu, and Jonathan Li. Nerf: Neural radiance field in 3d vision, a comprehensive review. arXiv preprint arXiv:2210.00379, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.324, + 0.472, + 0.379 + ], + "angle": 0, + "content": "[11] Henry Ham, Julian Wesley, and Hendra Hendra. Computer vision based 3d reconstruction: A review. International Journal of Electrical and Computer Engineering, 9(4):2394, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.382, + 0.472, + 0.424 + ], + "angle": 0, + "content": "[12] Rana Hanocka, Gal Metzer, Raja Giryes, and Daniel Cohen-Or. Point2mesh: A self-prior for deformable meshes. arXiv preprint arXiv:2005.11084, 2020. 2, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.426, + 0.472, + 0.481 + ], + "angle": 0, + "content": "[13] Michelle Holloway, Cindy Grimm, and Tao Ju. Template-based surface reconstruction from cross-sections. Computers & Graphics, 58:84-91, 2016. Shape Modeling International 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.484, + 0.472, + 0.539 + ], + "angle": 0, + "content": "[14] Zhiyang Huang, Ming Zou, Nathan Carr, and Tao Ju. Topology-controlled reconstruction of multi-labelled domains from cross-sections. ACM Transactions on Graphics (TOG), 36(4):1-12, 2017. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.542, + 0.472, + 0.598 + ], + "angle": 0, + "content": "[15] Tao Ju, Frank Losasso, Scott Schaefer, and Joe Warren. Dual contouring of hermite data. In Proceedings of the 29th annual conference on Computer graphics and interactive techniques, pages 339-346, 2002. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.601, + 0.472, + 0.656 + ], + "angle": 0, + "content": "[16] Lu Liu, Chandrajit Bajaj, Joseph O Deasy, Daniel A Low, and Tao Ju. Surface reconstruction from non-parallel curve networks. In Computer Graphics Forum, volume 27, pages 155–163. Wiley Online Library, 2008. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.657, + 0.472, + 0.727 + ], + "angle": 0, + "content": "[17] Baorui Ma, Zhizhong Han, Yu-Shen Liu, and Matthias Zwicker. Neural-pull: Learning signed distance function from point clouds by learning to pull space onto surface. In International Conference on Machine Learning, pages 7246-7257. PMLR, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.729, + 0.472, + 0.785 + ], + "angle": 0, + "content": "[18] Julien NP Martel, David B Lindell, Connor Z Lin, Eric R Chan, Marco Monteiro, and Gordon Wetzstein. Acorn: Adaptive coordinate networks for neural scene representation. arXiv preprint arXiv:2105.02788, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.788, + 0.472, + 0.855 + ], + "angle": 0, + "content": "[19] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.472, + 0.902 + ], + "angle": 0, + "content": "[20] Azimkhon Ostonov. Cut-and-approximate: 3d shape reconstruction from planar cross-sections with deep reinforcement learning. arXiv preprint arXiv:2210.12509, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.092, + 0.473, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.147 + ], + "angle": 0, + "content": "[21] Songyou Peng, Michael Niemeyer, Lars Mescheder, Marc Pollefeys, and Andreas Geiger. Convolutional occupancy networks. In European Conference on Computer Vision, pages 523-540. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.894, + 0.218 + ], + "angle": 0, + "content": "[22] Nasim Rahaman, Aristide Baratin, Devansh Arpit, Felix Draxler, Min Lin, Fred Hamprecht, Yoshua Bengio, and Aaron Courville. On the spectral bias of neural networks. In International Conference on Machine Learning, pages 5301-5310. PMLR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.22, + 0.894, + 0.26 + ], + "angle": 0, + "content": "[23] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.262, + 0.894, + 0.346 + ], + "angle": 0, + "content": "[24] Towaki Takikawa, Joey Litalien, Kangxue Yin, Karsten Kreis, Charles Loop, Derek Nowrouzezahrai, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Neural geometric level of detail: Real-time rendering with implicit 3d shapes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11358-11367, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.347, + 0.894, + 0.416 + ], + "angle": 0, + "content": "[25] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond. In Computer Graphics Forum, volume 41, pages 641-676. Wiley Online Library, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.418, + 0.894, + 0.472 + ], + "angle": 0, + "content": "[26] Ming Zou, Michelle Holloway, Nathan Carr, and Tao Ju. Topology-constrained surface reconstruction from cross-sections. ACM Transactions on Graphics (TOG), 34(4):1-10, 2015. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20862" + } + ] +] \ No newline at end of file diff --git a/2023/OReX_ Object Reconstruction From Planar Cross-Sections Using Neural Fields/748d73f6-6f8e-4794-a6a8-f014dea611bf_origin.pdf b/2023/OReX_ Object Reconstruction From Planar Cross-Sections Using Neural Fields/748d73f6-6f8e-4794-a6a8-f014dea611bf_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..bcddfed0c3a1ce960289359a856dd4c46c78bacc --- /dev/null +++ b/2023/OReX_ Object Reconstruction From Planar Cross-Sections Using Neural Fields/748d73f6-6f8e-4794-a6a8-f014dea611bf_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf5feba0cd80f66e27d9080d559c59d8e5828c6e721687c3875aea6af3936b90 +size 7315561 diff --git a/2023/OReX_ Object Reconstruction From Planar Cross-Sections Using Neural Fields/full.md b/2023/OReX_ Object Reconstruction From Planar Cross-Sections Using Neural Fields/full.md new file mode 100644 index 0000000000000000000000000000000000000000..37b429bb11d4c2dca9ce56c1b34618a7abe6f37d --- /dev/null +++ b/2023/OReX_ Object Reconstruction From Planar Cross-Sections Using Neural Fields/full.md @@ -0,0 +1,210 @@ +# OReX: Object Reconstruction from Planar Cross-sections Using Neural Fields + +Haim Sawdayee +The Blavatnik School of Computer Science +Tel Aviv University +haimsawdayee@mail.tau.ac.il + +Amir Vaxman +School of Informatics +The University of Edinburgh +avaxman@inf.ed.ac.uk + +Amit H. Bermano +The Blavatnik School of Computer Science +Tel Aviv University + +amberman@tauex.tau.ac.il + +![](images/e488693ee20eac151f657d31b60dfa8f6abe8109e87c4f4218c30181db8be148.jpg) +Figure 1. OREx reconstructs smooth 3D shapes (right) from input planar cross-sections (left). Our neural field-based prior allows smooth interpolation between slices while respecting high-frequency features and self-similarities. + +# Abstract + +Reconstructing 3D shapes from planar cross-sections is a challenge inspired by downstream applications like medical imaging and geographic informatics. The input is an in/out indicator function fully defined on a sparse collection of planes in space, and the output is an interpolation of the indicator function to the entire volume. Previous works addressing this sparse and ill-posed problem either produce low quality results, or rely on additional priors such as target topology, appearance information, or input normal directions. In this paper, we present OREx, a method for 3D shape reconstruction from slices alone, featuring a Neural Field as the interpolation prior. A modest neural network is trained on the input planes to return an inside/outside estimate for a given 3D coordinate, yielding a powerful prior that induces smoothness and self-similarities. The main challenge for this approach is high-frequency details, as the neural prior is overly smoothing. To alleviate this, we offer an iterative estimation architecture and a hierarchical input sampling scheme that encourage coarse-to-fine training, allowing the training process to focus on high frequencies at later stages. + +In addition, we identify and analyze a ripple-like effect stemming from the mesh extraction step. We mitigate it by regularizing the spatial gradients of the indicator function around input in/out boundaries during network training, tackling the problem at the root. Through extensive qualitative and quantitative experimentation, we demonstrate our method is robust, accurate, and scales well with the size of the input. We report state-of-the-art results compared to previous approaches and recent potential solutions, and demonstrate the benefit of our individual contributions through analysis and ablation studies. + +# 1. Introduction + +Reconstructing a 3D object from its cross-sections is a long-standing task. It persists in fields including medical imaging, topography mapping, and manufacturing. The typical setting is where a sparse set of arbitrary planes is given, upon which the "inside" and "outside" regions of the de + +picted domain are labeled, and the entire shape in 3D is to be estimated (see Fig. 1). This is a challenging and ill-posed problem, especially due to the sparse and irregular nature of the data. Classical approaches first localize the problem by constructing an arrangement of the input planes, and then introduce a local regularizer that governs the interpolation of the input to within each cell. While sound, these approaches typically involve simplistic regularization functions, that only interpolate the volume within a cell bounded by the relevant cross-sections; as a consequence, they introduce over-smoothed solutions that do not respect features. In addition, finding a cellular arrangement of planes is a computationally-heavy procedure, adding considerable complexity to the problem and rendering it quickly infeasible for large inputs (see Sec. 4). As we demonstrate (Sec. 4), recent approaches that reconstruct a mesh from an input point cloud are not well suited to our setting, as they assume a rather dense sampling of the entire shape. In addition, these methods do not consider the information of an entire cross-sectional plane, but rather only on the shape boundary. + +In this paper, we introduce $OReX$ a reconstruction approach based on neural networks that estimates an entire shape from its cross-sectional slices. Similar to recent approaches, the neural network constitutes the prior that extrapolates the input to the entire volume. Neural networks in general have already been shown to inherently induce smoothness [17], and self-similarities [12], allowing natural recurrence of patterns. Specifically, we argue that Neural Fields [25] are a promising choice for the task at hand. Neural Fields represent a 3D scene by estimating its density and other local geometric properties for every given 3D coordinate. They are typically trained on 2D planar images, and are required to complete the entire 3D scene according to multiview renderings or photographs. This neural representation is differentiable by construction, and hence allows native geometric optimization of the scene, realized via training. We pose the reconstruction problem as a classification of space into "in" and "out" regions, which are known for the entire slice planes, and thus generate the indicator function which its decision boundary is the output shape. + +The main challenge with applying neural fields to this problem is high-frequency details. Directly applying established training schemes [19] shows strong spectral bias, yielding overly smoothed results and other artifacts (Fig. 12). Spectral bias is a well-known effect, indicating that higher frequency is effectively learned slower [22]. To facilitate effective high-frequency learning, avoiding the shadow cast by the low frequency, we introduce two alterations. First, we sample the planar data mostly around the inside/outside transition regions, where the frequencies are higher. This sampling is further ordered from low to high-frequency regions (according to the distance from the inside/outside boundary), to encourage a low-to-high-frequency training progression. + +In addition, we follow recent literature and allow the network to iteratively infer the result, where later iterations are responsible for finer, higher-frequency corrections [1, 23]. + +Finally, we consider another high-frequency artifact, also found in other neural-field-based works [9]. The desired density (or indicator) function dictates a sharp drop in value at the shape boundary. This is contradictory to the induced neural prior, causing sampling-related artifacts in the downstream task of mesh extraction (Sec. 3.4). To alleviate this, we penalize strong spatial gradients around the boundary contours. This enforces a smoother transition between the in and out regions, allowing higher-quality mesh extraction. + +As we demonstrate (see Fig. 1), our method yields state-of-the-art reconstructions from planar data, both for woman-made and organic shapes. The careful loss and training schemes are validated and analyzed through quantitative and qualitative experimentation. Our method is arrangement-free, and thus both interpolates all data globally, avoiding local artifacts, and scales well to a large number of slices (see Sec. 4). + +# 2. Related work + +Reconstruction from cross-sections Reconstruction from planar cross-section has been a long-standing challenge in geometry processing and computational geometry. The problem was mostly studied in the parallel planes setting (e.g., [2,3,5]), where the 3D object was reconstructed per two such "keyframes". Later work offered general solutions for any orientation or distribution of planes [4, 16]. The general approach was based on tessellating the planes into convex cells by the planes of the arrangement and reconstructing the object in each cell, by defining some interpolant inside it. These methods suffer from several issues: the construction of the arrangement is computationally expensive (with at least cubic asymptotic times), and the local reconstruction introduces continuity artifacts (see Sec. 4). Furthermore, the interpolants were usually designed for smoothness or proximity, and failed to capture more global trends in the geometry of the slices, such as twists or extrusions (see Fig. 6). Some works provide topological guarantees [14, 26] or limited the solution to given templates [13], but do not provide a general solution without these priors. Our work is arrangement-free and provides a flexible high-parametric model, using neural networks, to capture the details of the reconstructed object. + +Neural reconstruction Most recent reconstruction works employing neural networks address reconstruction from dense point clouds. Our problem could be cast as a point cloud reconstruction one, after sampling the planes. In this setting, the high-frequency problem is pronounced as in our case, and even more so since more information regarding the target shape is available. Almost all recent ap + +![](images/e369288c32fccc3835d8c30e2514d62bdfa40b82052a65401407687ac1f0cbed.jpg) +Figure 2. OREx method overview. The input (left) is a set of planar cross-sections with inside/outside information on them. The planes are sampled into points in a hierarchical scheme, and are fed to our iterative Neural Field for training. After training, the final shape is estimated using an off-the-shelf mesh extractor. + +![](images/ba8977a5aa5189aa268bec55621ca9fa5bfccf0affd88d5fff97c20cf28f6b66.jpg) +Figure 3. We train OReX on three types of sampling distributions of the input: A scaled-up version of the 3D convex hull bounds the reconstruction volume (left); Uniform sampling within each plane helps stabilize the learning (middle); most of the samples are concentrated where accuracy matters most—on and around the boundary contours on the input planes (right). + +proaches employ an implicit representation of the shape. Some works populate 3D (sometimes adaptive) grids to handle fine-details [7,8,18,24], which is challenging to scale. Other works [6, 21] perform local operations, which rely on a rather dense neighborhood. To alleviate this, a mesh can be directly optimized to match the input [12]. Other neural fields have been employed as well in this context [17], without additional information such as normals; see a recent survey for more approaches [25]. As we demonstrate (Sec. 4), this general line of works employs weaker priors, which do not fit the sparse nature of our problem well. + +Perhaps a more similar setting is the one of reconstruction from projections or photographs. Here the reconstructions are from 2D data as well, which is also of sparser nature, albeit using appearance information that we do not have. Literature in this field is deep and wide, with several surveys [10, 11], including the usage of neural fields. We argue our work is orthogonal to that listed here, as it can be plugged in to replace our basic neural-field baseline. We leave for + +future work to inspect the most performing architecture. Addressing the most similar problem setting, concurrent work by Ostonov et al. [20] employs reinforcement learning and orthographic projections to ensure proper reconstruction. Except for requiring additional information during inference, they also require a training phase and are restricted to the domain trained on. Our approach, in contrast, requires no additional information besides the planes. + +# 3. Method + +We next lay out the details for effective high-quality shape reconstruction from slices. Our approach is based on a neural field, OReXNet. OReXNet outputs $Y(x)$ , an extrapolation of the inside/outside indicator function for a query point $x \in \mathbb{R}^3$ (Sec. 3.1). Given a single set of input cross-sections, the network undergoes training to approximate the target function on the input. After training is complete, $Y$ is sampled on the entire volume, and a resulting shape is extracted using the Dual Contouring (DC) approach [15]. This pipeline is depicted in Fig. 2. + +As motioned in Sec. 1, the main challenge in reconstruction quality is high-frequency details. The straightforward approach to our problem would be to train a neural field for the desired indicator function by uniformly sampling all planes, and subsequently training a network using a Binary Cross Entropy (BCE) loss on all sampled points. Typical approaches also represent the input coordinates using Positional Encoding (PE) [19]. As it turns out, this approach yields overly smooth results failing even to interpolate the input (Fig. 12). To improve reconstruction fidelity, and allow higher frequency details in the resulting shape, we introduce two alterations to the aforementioned training scheme. We present our hierarchical input sampling scheme in Sec. 3.2, and our iterative-refinement architecture in Sec. 3.3. + +Finally, we also address a common artifact in implicit mesh extraction [9]. Repeating patterns can be seen that correlate to the sampling pattern of the mesh extraction phase, as demonstrated in Fig. 8. We describe how we design the loss function to mitigate this artifact in Sec. 3.4. See the Supplementary material and code for implementation details and hyper-parameter values. + +# 3.1. Problem Setting + +We consider a set $\mathcal{P} = \{P_1,\dots ,P_k\}$ of 2D planes embedded in $\mathbb{R}^3$ , with arbitrary offsets and orientations. Each plane $P_{i}$ contains an arbitrary set of (softly) non-intersecting oriented contours $C_i = \{c_{i,1},\dots ,c_{i,l_i}\}$ that consistently partition the plane into regions of "inside" and "outside" of an unknown domain $\Omega \subset \mathbb{R}^3$ with boundary $\partial \Omega$ (Fig. 2). The target output of our method is an indicator function $Y:\mathbb{R}^3\to \mathbb{R}$ defining $\Omega$ as: + +$$ +Y (x) = \left\{ \begin{array}{l l} 1 & x \notin \Omega \\ 0 & x \in \Omega \\ 0. 5 & x \in \partial \Omega \end{array} \right. \tag {1} +$$ + +In practice, we approximate $Y$ using a function $f$ , such that $Y(x) \approx \sigma(f(x))$ , where $\sigma(z) = \frac{1}{1 + e^{-z}}$ is the sigmoid function. + +# 3.2. Input Sampling + +Regularly or randomly sampling the input planes for training yields inaccurate and overly smooth results (see Fig. 12). Instead, we define three types of relevant point distributions to sample from, as depicted in Fig. 3: + +1. We bound the reconstructed volume by the 3D convex hull of all the input contours, and scale it up by $5\%$ . We consequently sample it uniformly, where the sampled points are all outside. +2. We compute the bounding box (aligned to the principal axes) of the contours in each plane, and sample it with uniform distribution. +3. Most samples are taken around the boundaries of the contours, as these are the regions where accuracy matters most. We then sample every contour edge evenly, and further add off-surface points of varying distances for each edge sample. Off-surface locations are found by moving away from the contour on the plane (i.e., in the direction of the plane normal). We further sample off-surface points in a circle around each vertex. + +Each sampling point $x_{i}$ is matched with a label $Y_{i}$ according to its position on the slice, and the pairs $\{x_{i}, Y_{i}\}$ constitute the input to our training. See supplementary material for the relevant hyperparameters. + +Frequency-Oriented Sampling. In order to encourage better high-frequency learning, we sample the $3^{rd}$ type of points around a set of varying distances, from $0.1\%$ of the bounding box away from the contours to three orders of magnitude closer. In every epoch, we only use points sampled around three consecutive distance ranges. For early training iterations, we use the three largest distances, since further away points translate to lower frequency information about the shape. As the training progresses, and the lower frequencies are assumed to stabilize, we train with points that grow closer and closer to the actual contour, thus focusing the learning process on higher and higher frequencies (see Fig. 11). See supplementary material for more details and exact scheduling. + +# 3.3. Architecture + +ORexNet is simply an MLP that takes a 3D coordinate as input, represented using positional encoding, and produces the function $Y$ (Sec. 3.1) at any query point. In order to encourage high-frequency details, we introduce an iterative refinement mechanism. Inspired by recent work [1], we allow the network to refine its own results by running them through the network for a small number of iterations. This process was previously shown to produce a coarse-to-fine evolution in the realm of 2D images [1], and we argue it applies to our case as well. As demonstrated in Fig. 2, ORexNet is hence a residual MLP that is fed the result of the previous iteration, and a small hidden state code, along with the query point. These iterations (10 in our experiments) are performed both during training and inference; at inference time, only the last result is taken, and during training, the loss is applied to the results of all steps. The first iteration starts from a learned constant $C$ . We show that this process indeed sharpens features and allows the incorporation of higher-frequency details in Fig. 13. + +# 3.4. Loss Function and Inference + +In order to present our loss function, we must first attend to an issue in the final stage of our pipeline. In this stage, we extract an explicit mesh for the boundary of the indicator function using Dual Contouring [15]. This step uses a discretely sampled version of $Y(x)$ and $\nabla Y(x)$ on a regular 3D grid to extract the mesh. As can be seen in Fig. 8 (and as witnessed other works [9]), this creates a ripples-like artifact, that correlates to the 3D grid resolution. This results from an aliasing of the sampling in these regions where the gradient magnitudes exhibit high variance. This effect could be mitigated via higher resolution grid sampling, which is expensive, or a post-process operation that may compromise the geometry. Instead, we propose revisiting the loss function of the training process, and incorporating regularization to reduce this sharp drop in value at the shape boundary. First, we omit the activation—in other words, we use the + +![](images/aaa56d05df9f225105016f491599929176101356d798e60e998657efd31547a8.jpg) +Figure 4. Results gallery. A qualitative demonstration of a collection of man-made and medical inputs. Zoomed-in viewing is recommended. + +![](images/2d05ff135b53b6e06f211fb095439b6119caac82a2573d5bb4c1f8c288ae8250.jpg) +Figure 5. Additional qualitative results, where transparency reveals the reconstruction of internal cavities and tunnels. Zoomed-in viewing is recommended. + +function $f(x)$ for contouring, while using $Y(x)$ only for training (Sec 3.1). This de-radicalizes the function values, already offering a softer transition between inside and outside regions. In addition, we explicitly penalize strong transitions by using a hinge loss (Eq. 2). With this, we reduce the gradient magnitude variance considerably by limiting them close to the input shape boundary. Our loss function for a single + +point $x$ is then + +$$ +\mathcal {L} (x, \theta) = \sum_ {i = 0} ^ {N - 1} \operatorname {B C E} \left(Y _ {i} (x)\right) + \lambda \max (0, | | \nabla f _ {N - 1} (x) | | - \alpha) \tag {2} +$$ + +Where $\sigma$ is the sigmoid function, $\theta$ are the parameters of the network, $BCE()$ is the binary cross-entropy loss applied to all outputs of our iterative scheme, $max(0,x)$ is a hinge loss applied only to the last OReXnet iteration, $N$ the number of iterations, and $\alpha$ and $\lambda$ are hyperparameters. Fig. 8 demonstrates the correction effect of the hinge loss. + +# 4. Experiments + +To evaluate our method, we developed a prototype, and ran it on a NVIDIA GeForce GTX 1080 Ti. See the Supplementary Material for exact timing and memory consumption statistics. Generally speaking, all training processes took less than four hours on the single GPU. + +# 4.1. Results and Comparisons + +We first qualitatively demonstrate the result of our method on a variety of slice inputs, from both the medical and graphics worlds (Fig. 4). This demonstrates the versatility and generality of our method. We further show our our algorithm correctly reproduces the internal cavities and details of reconstructed objects in Fig. 5. Note how intricate details are learned along side a smooth interpolation between slices, + +![](images/5864da259dc40994c5510a921b02c5ff5ea6b4e1ade1d8d8f76918ba0b663a8e.jpg) +Figure 6. Qualitative comparisons to other approaches. Our result is smoother, more aligned to features, and with less artifacts. Note how the reconstructed shape silhouettes does not suffer imprints from the input planes. Check Table 1 for quantitative comparison. + +leaving no slice transition artifacts on the resulting shape. + +In terms of other methods, we qualitatively (Fig. 6) and quantitatively (Table 1) compare our method to Bermano et al. [4], which is a reconstruction from cross-section method with the same input-output as ours. We further compare to general state-of-the-art reconstruction methods that target point clouds [12]. Hanocka et al. [12] expect point normals as input as well. Hence we compare our result to the latter work with Ground Truth (GT) normals at input, and with ones coming from the input plane normals for a fairer comparison. We use the following metrics: 1) intersection-over-union volume of inside regions in reconstructed volumes and 2) GT planes, and 3) symmetric Hausdorff distance. These metrics demonstrate our superiority over the compared methods. See Supplementary Material for more comparisons. + +**Scalability** We show that our method does not suffer the computational cost that arrangement-based methods [4, 14, 16] must bear, since we do not tessellate the space, and hence scale well with the input. In addition, we show (Fig. 7) how our method converges with the addition of slices. + +# 4.2. Ablations + +We perform ablation testing to evaluate important aspects of our method and our design choices. + +![](images/8b367136416dde2717c7abc95eec0e40af5edfac6287f63649a5b034fecac0f7.jpg) +Figure 7. Increasing number of slices. As can be expected, introducing more slices converges consistently converges to the solution. + +Gradient magnitude regularization (Fig. 8) We show the effect of the hinge-loss regularizer on gradient magnitudes, with increasing $\lambda$ values. It is evident the hinge loss effectively filters out the high variance, and smooths the ripple artifacts. As can be seen (bottom row), without our hinge loss, geometry vertices tend to snap to discrete locations during the mesh extraction phase. This is because a large gradient induces a large bias in the magnitude of the indicator function, encouraging vertices to remain close to the grid points sampled during the extraction step. Regularizing the gradients around the boundary regions (top row) decreases the gap between the positive and negative sides of the function, allowing a more uniform, and hence continuous, distribution + +of the mesh vertices. With this bias removed, the resulting geometry is better sampled and hence the ripple artifact is mitigated (middle). + +![](images/30a8e65595c84cd5c6cc6264b27abefaca22af12dc1ead2e5adfecffdb9659fb.jpg) +Figure 8. Left to right: our Hinge loss effect for mitigating ripple-like aliasing artifacts. Top: output values of $f(x)$ on a single slice (arrows show gradients). Middle: The reconstructed mesh with a closer view on the top of the skull. Bottom: distribution of vertices along grid edges during mesh extraction. It is evident the hinge loss effectively controls gradient magnitudes at the zero set, curbing the aliasing artifacts. The hinge regularization further allows to place more vertices in the middle of the sampled grid cells, and not only on their edges. + +Iterative architecture (Fig. 9) This ablations examines the effect of the number of refinement iterations on the accuracy of the mesh produced. As can be expected, the results show that as the number of iterations increases, high-frequency details are better represented. + +![](images/3852e3adc0901289721be3a1cea9dad6a53f356651fbe67c30786541190b859f.jpg) +Figure 9. Iterative refinement experiment. Training and inferring with more OReXNet refinement iterations allows the network to perform smaller scale corrections and reduces spectral bias. + +Sampling method To evaluate our importance sampling scheme, we first replace it with regular grid sampling in two resolutions (Fig. 10). As can be seen, using each of the resolutions alone yields overly smoothed results due to spectral bias. Training with a gradually increasing resolution allows better handling of high-frequency details. In addition, we have demonstrate the effects on quality when using our importance scheme in a non-hierarchical manner, i.e. using all points uniformly without consideration of distance from the boundary (Fig. 11). As can be seen, our scheme improves in detail quality upon both experiments, demonstrating that our sampling scheme produces more accurate results with less spectral bias while using fewer samples. + +![](images/dd91b74219f224ad277f141b755395c9ce44fc9fb7bd03f183df50409134bf64.jpg) +Figure 10. Sampling scheme Ablations. Left: ground truth. Middle: grid sampling with different resolutions, and a gradually increasing sampling resolution over the training process. Right: our method. + +![](images/257734fcec337945ae835bb473ac3c885d312e63ab418626c197090aa3c9dc1c.jpg) +Figure 11. Left: non-hierarchical sampling, using points from all off-surface distances uniformly. Right: using our hierarchical sampling scheme. + +Choice of architecture and sampling (Fig. 12) We justify our design choices by training a baseline architecture with uniform grid sampling and no iteration refinement. It is evident such a model does not preserve details nor interpolate well, even for a relatively simple case. + +
InputHausdorff distanceIoU in 3dIoU in 2d
ORExBermano et. alpoint2mesh plane normalspoint2mesh GT normalsORExBermano et. alpoint2mesh plane normalspoint2mesh GT normalsORExBermano et. alpoint2mesh plane normalspoint2mesh GT normals
Eight (S)0.0180.0650.2190.0460.9840.8650.8420.9610.9880.9840.7950.980
Eight (M)0.0060.0330.0450.0140.9870.8930.9150.9740.9860.9760.9610.971
Elephant0.0560.0810.1000.0860.9660.9080.8850.9350.9750.9690.8500.908
Balloon dog0.0490.1940.0780.0860.9570.8680.8970.9280.9880.9770.9260.956
Hand OK0.0630.1770.1950.1350.9550.8600.9210.9310.9870.9680.9080.882
Armadillo0.0500.1210.0570.0590.9390.8910.8910.9210.9640.7760.8500.868
+ +Table 1. Quantitative comparisons. We measure performance using Hausdorff distance, IoU of the inner volume compared to the GT shape, and IoU of the inner surface on the input cross-sections. We compare our result to a dedicated cross-section based reconstruction method [4], and to two flavors of a pointcloud reconstruction method [12]. See more comparisons in the Supplementary Material. + +![](images/0867482e84d4e25861d89138a15b1e6b5756b1775009965359e10ee8626c7648.jpg) +Figure 12. Left and middle: the baseline experiment with regular grid samples of 64 and 128 resolution over the slices. Right: our result for the same amount for samples (our variable number of samples is provided at its mean). + +![](images/73edd476b5527a81cf34571a53dd7d7feabb4a2897acd525b7241e2605c4554a.jpg) +Figure 13. Refinement progression throughout the iteration process. The maximum change in the output value decreases along the iterations. Right: points that were added to the shape (Red) and that were removed from it (blue) in the last refinement iteration. These are concentrated on the level set, whereas most other points (green, covering all internal regions) have already converged. + +# 5. Conclusions + +We have presented OREx, a state-of-the-art method for the long-standing problem of shape reconstruction from planar cross-sectional indicator data. Free of datasets and training requirements, OREx is simple and intuitive to use. Our + +work balances the smoothness of a neural prior with high-frequency features. We show our approach successfully produces smooth interpolation between contours while respecting high-frequency features and repeating patterns. In addition, we believe some of the analysis and insights presented here can be applied to neural fields in similar tasks. + +The advantage of our method is also its disadvantage: using only binary data, much of the information is lost in the process. For example, for medical imaging an interesting future direction would be to directly use the raw grayscale density input of a planar probe such as that comes from an ultrasound. Another interesting direction for research is extending this work for multi-labeled volumes (e.g., the reconstruction of several organs simultaneously from a scan), and using partial or noisy data. + +# References + +[1] Yuval Alaluf, Or Patashnik, and Daniel Cohen-Or. Restyle: A residual-based stylegan encoder via iterative refinement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6711–6720, 2021. 2, 4 +[2] Chandrajit L Bajaj, Edward J Coyle, and Kwun-Nan Lin. Arbitrary topology shape reconstruction from planar cross sections. Graphical models and image processing, 58(6):524-543, 1996. 2 +[3] Gill Barequet and Micha Sharir. Piecewise-linear interpolation between polygonal slices. In Proceedings of the tenth annual symposium on Computational geometry, pages 93-102, 1994. 2 +[4] Amit Bermano, Amir Vaxman, and Craig Gotsman. Online reconstruction of 3d objects from arbitrary cross-sections. ACM Transactions on Graphics (TOG), 30(5):1-11, 2011. 2, 6, 8 +[5] Jean-Daniel Boissonnat and Pooran Memari. Shape reconstruction from unorganized cross-sections. In Symposium on geometry processing, pages 89-98, 2007. 2 +[6] Alexandre Boulch and Renaud Marlet. Poco: Point convolution for surface reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6302-6314, 2022. 3 +[7] Zhang Chen, Yinda Zhang, Kyle Genova, Sean Fanello, Sofien Bouaziz, Christian Hane, Ruofei Du, Cem Keskin, Thomas Funkhouser, and Danhang Tang. Multiresolution + +deep implicit functions for 3d shape representation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13087-13096, 2021. 3 +[8] Julian Chibane, Thiemo Alldieck, and Gerard Pons-Moll. Implicit functions in feature space for 3d shape reconstruction and completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6970-6981, 2020. 3 +[9] Philipp Erler, Paul Guerrero, Stefan Ohrhallinger, Niloy J Mitra, and Michael Wimmer. Points2surf learning implicit surfaces from point clouds. In European Conference on Computer Vision, pages 108-124. Springer, 2020. 2, 4 +[10] Kyle Gao, Yina Gao, Hongjie He, Denning Lu, Linlin Xu, and Jonathan Li. Nerf: Neural radiance field in 3d vision, a comprehensive review. arXiv preprint arXiv:2210.00379, 2022. 3 +[11] Henry Ham, Julian Wesley, and Hendra Hendra. Computer vision based 3d reconstruction: A review. International Journal of Electrical and Computer Engineering, 9(4):2394, 2019. 3 +[12] Rana Hanocka, Gal Metzer, Raja Giryes, and Daniel Cohen-Or. Point2mesh: A self-prior for deformable meshes. arXiv preprint arXiv:2005.11084, 2020. 2, 3, 6, 8 +[13] Michelle Holloway, Cindy Grimm, and Tao Ju. Template-based surface reconstruction from cross-sections. Computers & Graphics, 58:84-91, 2016. Shape Modeling International 2016. 2 +[14] Zhiyang Huang, Ming Zou, Nathan Carr, and Tao Ju. Topology-controlled reconstruction of multi-labelled domains from cross-sections. ACM Transactions on Graphics (TOG), 36(4):1-12, 2017. 2, 6 +[15] Tao Ju, Frank Losasso, Scott Schaefer, and Joe Warren. Dual contouring of hermite data. In Proceedings of the 29th annual conference on Computer graphics and interactive techniques, pages 339-346, 2002. 3, 4 +[16] Lu Liu, Chandrajit Bajaj, Joseph O Deasy, Daniel A Low, and Tao Ju. Surface reconstruction from non-parallel curve networks. In Computer Graphics Forum, volume 27, pages 155–163. Wiley Online Library, 2008. 2, 6 +[17] Baorui Ma, Zhizhong Han, Yu-Shen Liu, and Matthias Zwicker. Neural-pull: Learning signed distance function from point clouds by learning to pull space onto surface. In International Conference on Machine Learning, pages 7246-7257. PMLR, 2021. 2, 3 +[18] Julien NP Martel, David B Lindell, Connor Z Lin, Eric R Chan, Marco Monteiro, and Gordon Wetzstein. Acorn: Adaptive coordinate networks for neural scene representation. arXiv preprint arXiv:2105.02788, 2021. 3 +[19] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2, 3 +[20] Azimkhon Ostonov. Cut-and-approximate: 3d shape reconstruction from planar cross-sections with deep reinforcement learning. arXiv preprint arXiv:2210.12509, 2022. 3 + +[21] Songyou Peng, Michael Niemeyer, Lars Mescheder, Marc Pollefeys, and Andreas Geiger. Convolutional occupancy networks. In European Conference on Computer Vision, pages 523-540. Springer, 2020. 3 +[22] Nasim Rahaman, Aristide Baratin, Devansh Arpit, Felix Draxler, Min Lin, Fred Hamprecht, Yoshua Bengio, and Aaron Courville. On the spectral bias of neural networks. In International Conference on Machine Learning, pages 5301-5310. PMLR, 2019. 2 +[23] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 2 +[24] Towaki Takikawa, Joey Litalien, Kangxue Yin, Karsten Kreis, Charles Loop, Derek Nowrouzezahrai, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Neural geometric level of detail: Real-time rendering with implicit 3d shapes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11358-11367, 2021. 3 +[25] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond. In Computer Graphics Forum, volume 41, pages 641-676. Wiley Online Library, 2022. 2, 3 +[26] Ming Zou, Michelle Holloway, Nathan Carr, and Tao Ju. Topology-constrained surface reconstruction from cross-sections. ACM Transactions on Graphics (TOG), 34(4):1-10, 2015. 2 \ No newline at end of file diff --git a/2023/OReX_ Object Reconstruction From Planar Cross-Sections Using Neural Fields/images.zip b/2023/OReX_ Object Reconstruction From Planar Cross-Sections Using Neural Fields/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..46cc85749f41fefd7fb531aa37ac51180b145629 --- /dev/null +++ b/2023/OReX_ Object Reconstruction From Planar Cross-Sections Using Neural Fields/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5a722ab661f245c65247eb3955a2a5d3870bfdda40ea1ebeee44f4cf8eba5238 +size 802858 diff --git a/2023/OReX_ Object Reconstruction From Planar Cross-Sections Using Neural Fields/layout.json b/2023/OReX_ Object Reconstruction From Planar Cross-Sections Using Neural Fields/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..cf7976efe2fdc89b41cc272e5bbfb3915ade2a0b --- /dev/null +++ b/2023/OReX_ Object Reconstruction From Planar Cross-Sections Using Neural Fields/layout.json @@ -0,0 +1,5415 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 55, + 103, + 538, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 103, + 538, + 121 + ], + "spans": [ + { + "bbox": [ + 55, + 103, + 538, + 121 + ], + "type": "text", + "content": "OReX: Object Reconstruction from Planar Cross-sections Using Neural Fields" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 144, + 317, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 144, + 317, + 198 + ], + "spans": [ + { + "bbox": [ + 106, + 144, + 317, + 198 + ], + "type": "text", + "content": "Haim Sawdayee \nThe Blavatnik School of Computer Science \nTel Aviv University \nhaimsawdayee@mail.tau.ac.il" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 347, + 144, + 487, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 144, + 487, + 198 + ], + "spans": [ + { + "bbox": [ + 347, + 144, + 487, + 198 + ], + "type": "text", + "content": "Amir Vaxman \nSchool of Informatics \nThe University of Edinburgh \navaxman@inf.ed.ac.uk" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 192, + 205, + 402, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 205, + 402, + 247 + ], + "spans": [ + { + "bbox": [ + 192, + 205, + 402, + 247 + ], + "type": "text", + "content": "Amit H. Bermano \nThe Blavatnik School of Computer Science \nTel Aviv University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 229, + 250, + 361, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 229, + 250, + 361, + 259 + ], + "spans": [ + { + "bbox": [ + 229, + 250, + 361, + 259 + ], + "type": "text", + "content": "amberman@tauex.tau.ac.il" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 62, + 271, + 532, + 408 + ], + "blocks": [ + { + "bbox": [ + 62, + 271, + 532, + 408 + ], + "lines": [ + { + "bbox": [ + 62, + 271, + 532, + 408 + ], + "spans": [ + { + "bbox": [ + 62, + 271, + 532, + 408 + ], + "type": "image", + "image_path": "e488693ee20eac151f657d31b60dfa8f6abe8109e87c4f4218c30181db8be148.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 413, + 546, + 436 + ], + "lines": [ + { + "bbox": [ + 46, + 413, + 546, + 436 + ], + "spans": [ + { + "bbox": [ + 46, + 413, + 546, + 436 + ], + "type": "text", + "content": "Figure 1. OREx reconstructs smooth 3D shapes (right) from input planar cross-sections (left). Our neural field-based prior allows smooth interpolation between slices while respecting high-frequency features and self-similarities." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 143, + 447, + 192, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 447, + 192, + 460 + ], + "spans": [ + { + "bbox": [ + 143, + 447, + 192, + 460 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 474, + 290, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 290, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 290, + 715 + ], + "type": "text", + "content": "Reconstructing 3D shapes from planar cross-sections is a challenge inspired by downstream applications like medical imaging and geographic informatics. The input is an in/out indicator function fully defined on a sparse collection of planes in space, and the output is an interpolation of the indicator function to the entire volume. Previous works addressing this sparse and ill-posed problem either produce low quality results, or rely on additional priors such as target topology, appearance information, or input normal directions. In this paper, we present OREx, a method for 3D shape reconstruction from slices alone, featuring a Neural Field as the interpolation prior. A modest neural network is trained on the input planes to return an inside/outside estimate for a given 3D coordinate, yielding a powerful prior that induces smoothness and self-similarities. The main challenge for this approach is high-frequency details, as the neural prior is overly smoothing. To alleviate this, we offer an iterative estimation architecture and a hierarchical input sampling scheme that encourage coarse-to-fine training, allowing the training process to focus on high frequencies at later stages." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 449, + 548, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 548, + 579 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 548, + 579 + ], + "type": "text", + "content": "In addition, we identify and analyze a ripple-like effect stemming from the mesh extraction step. We mitigate it by regularizing the spatial gradients of the indicator function around input in/out boundaries during network training, tackling the problem at the root. Through extensive qualitative and quantitative experimentation, we demonstrate our method is robust, accurate, and scales well with the size of the input. We report state-of-the-art results compared to previous approaches and recent potential solutions, and demonstrate the benefit of our individual contributions through analysis and ablation studies." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 604, + 387, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 604, + 387, + 616 + ], + "spans": [ + { + "bbox": [ + 306, + 604, + 387, + 616 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 624, + 548, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 624, + 548, + 685 + ], + "spans": [ + { + "bbox": [ + 304, + 624, + 548, + 685 + ], + "type": "text", + "content": "Reconstructing a 3D object from its cross-sections is a long-standing task. It persists in fields including medical imaging, topography mapping, and manufacturing. The typical setting is where a sparse set of arbitrary planes is given, upon which the \"inside\" and \"outside\" regions of the de" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 306, + 693, + 546, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 693, + 546, + 712 + ], + "spans": [ + { + "bbox": [ + 306, + 693, + 546, + 712 + ], + "type": "text", + "content": "1Code and data available at https://github.com/haimsaw/ ORex" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20854" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 312 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 312 + ], + "type": "text", + "content": "picted domain are labeled, and the entire shape in 3D is to be estimated (see Fig. 1). This is a challenging and ill-posed problem, especially due to the sparse and irregular nature of the data. Classical approaches first localize the problem by constructing an arrangement of the input planes, and then introduce a local regularizer that governs the interpolation of the input to within each cell. While sound, these approaches typically involve simplistic regularization functions, that only interpolate the volume within a cell bounded by the relevant cross-sections; as a consequence, they introduce over-smoothed solutions that do not respect features. In addition, finding a cellular arrangement of planes is a computationally-heavy procedure, adding considerable complexity to the problem and rendering it quickly infeasible for large inputs (see Sec. 4). As we demonstrate (Sec. 4), recent approaches that reconstruct a mesh from an input point cloud are not well suited to our setting, as they assume a rather dense sampling of the entire shape. In addition, these methods do not consider the information of an entire cross-sectional plane, but rather only on the shape boundary." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 315, + 289, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 315, + 289, + 555 + ], + "spans": [ + { + "bbox": [ + 47, + 315, + 289, + 555 + ], + "type": "text", + "content": "In this paper, we introduce " + }, + { + "bbox": [ + 47, + 315, + 289, + 555 + ], + "type": "inline_equation", + "content": "OReX" + }, + { + "bbox": [ + 47, + 315, + 289, + 555 + ], + "type": "text", + "content": " a reconstruction approach based on neural networks that estimates an entire shape from its cross-sectional slices. Similar to recent approaches, the neural network constitutes the prior that extrapolates the input to the entire volume. Neural networks in general have already been shown to inherently induce smoothness [17], and self-similarities [12], allowing natural recurrence of patterns. Specifically, we argue that Neural Fields [25] are a promising choice for the task at hand. Neural Fields represent a 3D scene by estimating its density and other local geometric properties for every given 3D coordinate. They are typically trained on 2D planar images, and are required to complete the entire 3D scene according to multiview renderings or photographs. This neural representation is differentiable by construction, and hence allows native geometric optimization of the scene, realized via training. We pose the reconstruction problem as a classification of space into \"in\" and \"out\" regions, which are known for the entire slice planes, and thus generate the indicator function which its decision boundary is the output shape." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 558, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 558, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 558, + 289, + 715 + ], + "type": "text", + "content": "The main challenge with applying neural fields to this problem is high-frequency details. Directly applying established training schemes [19] shows strong spectral bias, yielding overly smoothed results and other artifacts (Fig. 12). Spectral bias is a well-known effect, indicating that higher frequency is effectively learned slower [22]. To facilitate effective high-frequency learning, avoiding the shadow cast by the low frequency, we introduce two alterations. First, we sample the planar data mostly around the inside/outside transition regions, where the frequencies are higher. This sampling is further ordered from low to high-frequency regions (according to the distance from the inside/outside boundary), to encourage a low-to-high-frequency training progression." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 304, + 72, + 547, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 108 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 108 + ], + "type": "text", + "content": "In addition, we follow recent literature and allow the network to iteratively infer the result, where later iterations are responsible for finer, higher-frequency corrections [1, 23]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 108, + 547, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 108, + 547, + 215 + ], + "spans": [ + { + "bbox": [ + 304, + 108, + 547, + 215 + ], + "type": "text", + "content": "Finally, we consider another high-frequency artifact, also found in other neural-field-based works [9]. The desired density (or indicator) function dictates a sharp drop in value at the shape boundary. This is contradictory to the induced neural prior, causing sampling-related artifacts in the downstream task of mesh extraction (Sec. 3.4). To alleviate this, we penalize strong spatial gradients around the boundary contours. This enforces a smoother transition between the in and out regions, allowing higher-quality mesh extraction." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 216, + 547, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 216, + 547, + 312 + ], + "spans": [ + { + "bbox": [ + 304, + 216, + 547, + 312 + ], + "type": "text", + "content": "As we demonstrate (see Fig. 1), our method yields state-of-the-art reconstructions from planar data, both for woman-made and organic shapes. The careful loss and training schemes are validated and analyzed through quantitative and qualitative experimentation. Our method is arrangement-free, and thus both interpolates all data globally, avoiding local artifacts, and scales well to a large number of slices (see Sec. 4)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 320, + 390, + 333 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 320, + 390, + 333 + ], + "spans": [ + { + "bbox": [ + 306, + 320, + 390, + 333 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 340, + 547, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 340, + 547, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 340, + 547, + 617 + ], + "type": "text", + "content": "Reconstruction from cross-sections Reconstruction from planar cross-section has been a long-standing challenge in geometry processing and computational geometry. The problem was mostly studied in the parallel planes setting (e.g., [2,3,5]), where the 3D object was reconstructed per two such \"keyframes\". Later work offered general solutions for any orientation or distribution of planes [4, 16]. The general approach was based on tessellating the planes into convex cells by the planes of the arrangement and reconstructing the object in each cell, by defining some interpolant inside it. These methods suffer from several issues: the construction of the arrangement is computationally expensive (with at least cubic asymptotic times), and the local reconstruction introduces continuity artifacts (see Sec. 4). Furthermore, the interpolants were usually designed for smoothness or proximity, and failed to capture more global trends in the geometry of the slices, such as twists or extrusions (see Fig. 6). Some works provide topological guarantees [14, 26] or limited the solution to given templates [13], but do not provide a general solution without these priors. Our work is arrangement-free and provides a flexible high-parametric model, using neural networks, to capture the details of the reconstructed object." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 630, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 548, + 715 + ], + "type": "text", + "content": "Neural reconstruction Most recent reconstruction works employing neural networks address reconstruction from dense point clouds. Our problem could be cast as a point cloud reconstruction one, after sampling the planes. In this setting, the high-frequency problem is pronounced as in our case, and even more so since more information regarding the target shape is available. Almost all recent ap" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20855" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 76, + 545, + 228 + ], + "blocks": [ + { + "bbox": [ + 52, + 76, + 545, + 228 + ], + "lines": [ + { + "bbox": [ + 52, + 76, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 52, + 76, + 545, + 228 + ], + "type": "image", + "image_path": "e369288c32fccc3835d8c30e2514d62bdfa40b82052a65401407687ac1f0cbed.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 236, + 546, + 270 + ], + "lines": [ + { + "bbox": [ + 46, + 236, + 546, + 270 + ], + "spans": [ + { + "bbox": [ + 46, + 236, + 546, + 270 + ], + "type": "text", + "content": "Figure 2. OREx method overview. The input (left) is a set of planar cross-sections with inside/outside information on them. The planes are sampled into points in a hierarchical scheme, and are fed to our iterative Neural Field for training. After training, the final shape is estimated using an off-the-shelf mesh extractor." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 55, + 288, + 281, + 385 + ], + "blocks": [ + { + "bbox": [ + 55, + 288, + 281, + 385 + ], + "lines": [ + { + "bbox": [ + 55, + 288, + 281, + 385 + ], + "spans": [ + { + "bbox": [ + 55, + 288, + 281, + 385 + ], + "type": "image", + "image_path": "ba8977a5aa5189aa268bec55621ca9fa5bfccf0affd88d5fff97c20cf28f6b66.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 392, + 288, + 460 + ], + "lines": [ + { + "bbox": [ + 46, + 392, + 288, + 460 + ], + "spans": [ + { + "bbox": [ + 46, + 392, + 288, + 460 + ], + "type": "text", + "content": "Figure 3. We train OReX on three types of sampling distributions of the input: A scaled-up version of the 3D convex hull bounds the reconstruction volume (left); Uniform sampling within each plane helps stabilize the learning (middle); most of the samples are concentrated where accuracy matters most—on and around the boundary contours on the input planes (right)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 484, + 288, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 484, + 288, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 484, + 288, + 616 + ], + "type": "text", + "content": "proaches employ an implicit representation of the shape. Some works populate 3D (sometimes adaptive) grids to handle fine-details [7,8,18,24], which is challenging to scale. Other works [6, 21] perform local operations, which rely on a rather dense neighborhood. To alleviate this, a mesh can be directly optimized to match the input [12]. Other neural fields have been employed as well in this context [17], without additional information such as normals; see a recent survey for more approaches [25]. As we demonstrate (Sec. 4), this general line of works employs weaker priors, which do not fit the sparse nature of our problem well." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 713 + ], + "type": "text", + "content": "Perhaps a more similar setting is the one of reconstruction from projections or photographs. Here the reconstructions are from 2D data as well, which is also of sparser nature, albeit using appearance information that we do not have. Literature in this field is deep and wide, with several surveys [10, 11], including the usage of neural fields. We argue our work is orthogonal to that listed here, as it can be plugged in to replace our basic neural-field baseline. We leave for" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 291, + 547, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 291, + 547, + 388 + ], + "spans": [ + { + "bbox": [ + 304, + 291, + 547, + 388 + ], + "type": "text", + "content": "future work to inspect the most performing architecture. Addressing the most similar problem setting, concurrent work by Ostonov et al. [20] employs reinforcement learning and orthographic projections to ensure proper reconstruction. Except for requiring additional information during inference, they also require a training phase and are restricted to the domain trained on. Our approach, in contrast, requires no additional information besides the planes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 403, + 362, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 403, + 362, + 415 + ], + "spans": [ + { + "bbox": [ + 306, + 403, + 362, + 415 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 425, + 547, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 425, + 547, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 425, + 547, + 544 + ], + "type": "text", + "content": "We next lay out the details for effective high-quality shape reconstruction from slices. Our approach is based on a neural field, OReXNet. OReXNet outputs " + }, + { + "bbox": [ + 304, + 425, + 547, + 544 + ], + "type": "inline_equation", + "content": "Y(x)" + }, + { + "bbox": [ + 304, + 425, + 547, + 544 + ], + "type": "text", + "content": ", an extrapolation of the inside/outside indicator function for a query point " + }, + { + "bbox": [ + 304, + 425, + 547, + 544 + ], + "type": "inline_equation", + "content": "x \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 425, + 547, + 544 + ], + "type": "text", + "content": " (Sec. 3.1). Given a single set of input cross-sections, the network undergoes training to approximate the target function on the input. After training is complete, " + }, + { + "bbox": [ + 304, + 425, + 547, + 544 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 304, + 425, + 547, + 544 + ], + "type": "text", + "content": " is sampled on the entire volume, and a resulting shape is extracted using the Dual Contouring (DC) approach [15]. This pipeline is depicted in Fig. 2." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 546, + 548, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 548, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 548, + 713 + ], + "type": "text", + "content": "As motioned in Sec. 1, the main challenge in reconstruction quality is high-frequency details. The straightforward approach to our problem would be to train a neural field for the desired indicator function by uniformly sampling all planes, and subsequently training a network using a Binary Cross Entropy (BCE) loss on all sampled points. Typical approaches also represent the input coordinates using Positional Encoding (PE) [19]. As it turns out, this approach yields overly smooth results failing even to interpolate the input (Fig. 12). To improve reconstruction fidelity, and allow higher frequency details in the resulting shape, we introduce two alterations to the aforementioned training scheme. We present our hierarchical input sampling scheme in Sec. 3.2, and our iterative-refinement architecture in Sec. 3.3." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20856" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": "Finally, we also address a common artifact in implicit mesh extraction [9]. Repeating patterns can be seen that correlate to the sampling pattern of the mesh extraction phase, as demonstrated in Fig. 8. We describe how we design the loss function to mitigate this artifact in Sec. 3.4. See the Supplementary material and code for implementation details and hyper-parameter values." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 164, + 146, + 177 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 164, + 146, + 177 + ], + "spans": [ + { + "bbox": [ + 47, + 164, + 146, + 177 + ], + "type": "text", + "content": "3.1. Problem Setting" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 182, + 290, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 182, + 290, + 278 + ], + "spans": [ + { + "bbox": [ + 46, + 182, + 290, + 278 + ], + "type": "text", + "content": "We consider a set " + }, + { + "bbox": [ + 46, + 182, + 290, + 278 + ], + "type": "inline_equation", + "content": "\\mathcal{P} = \\{P_1,\\dots ,P_k\\}" + }, + { + "bbox": [ + 46, + 182, + 290, + 278 + ], + "type": "text", + "content": " of 2D planes embedded in " + }, + { + "bbox": [ + 46, + 182, + 290, + 278 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^3" + }, + { + "bbox": [ + 46, + 182, + 290, + 278 + ], + "type": "text", + "content": " , with arbitrary offsets and orientations. Each plane " + }, + { + "bbox": [ + 46, + 182, + 290, + 278 + ], + "type": "inline_equation", + "content": "P_{i}" + }, + { + "bbox": [ + 46, + 182, + 290, + 278 + ], + "type": "text", + "content": " contains an arbitrary set of (softly) non-intersecting oriented contours " + }, + { + "bbox": [ + 46, + 182, + 290, + 278 + ], + "type": "inline_equation", + "content": "C_i = \\{c_{i,1},\\dots ,c_{i,l_i}\\}" + }, + { + "bbox": [ + 46, + 182, + 290, + 278 + ], + "type": "text", + "content": " that consistently partition the plane into regions of \"inside\" and \"outside\" of an unknown domain " + }, + { + "bbox": [ + 46, + 182, + 290, + 278 + ], + "type": "inline_equation", + "content": "\\Omega \\subset \\mathbb{R}^3" + }, + { + "bbox": [ + 46, + 182, + 290, + 278 + ], + "type": "text", + "content": " with boundary " + }, + { + "bbox": [ + 46, + 182, + 290, + 278 + ], + "type": "inline_equation", + "content": "\\partial \\Omega" + }, + { + "bbox": [ + 46, + 182, + 290, + 278 + ], + "type": "text", + "content": " (Fig. 2). The target output of our method is an indicator function " + }, + { + "bbox": [ + 46, + 182, + 290, + 278 + ], + "type": "inline_equation", + "content": "Y:\\mathbb{R}^3\\to \\mathbb{R}" + }, + { + "bbox": [ + 46, + 182, + 290, + 278 + ], + "type": "text", + "content": " defining " + }, + { + "bbox": [ + 46, + 182, + 290, + 278 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 46, + 182, + 290, + 278 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 116, + 288, + 288, + 332 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 288, + 288, + 332 + ], + "spans": [ + { + "bbox": [ + 116, + 288, + 288, + 332 + ], + "type": "interline_equation", + "content": "Y (x) = \\left\\{ \\begin{array}{l l} 1 & x \\notin \\Omega \\\\ 0 & x \\in \\Omega \\\\ 0. 5 & x \\in \\partial \\Omega \\end{array} \\right. \\tag {1}", + "image_path": "3ef4a9979c5f9b05a5a2a3a865a57d12cbbf25156f82fb17f5e494c7da5bbea4.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 341, + 288, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 341, + 288, + 376 + ], + "spans": [ + { + "bbox": [ + 47, + 341, + 288, + 376 + ], + "type": "text", + "content": "In practice, we approximate " + }, + { + "bbox": [ + 47, + 341, + 288, + 376 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 47, + 341, + 288, + 376 + ], + "type": "text", + "content": " using a function " + }, + { + "bbox": [ + 47, + 341, + 288, + 376 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 47, + 341, + 288, + 376 + ], + "type": "text", + "content": ", such that " + }, + { + "bbox": [ + 47, + 341, + 288, + 376 + ], + "type": "inline_equation", + "content": "Y(x) \\approx \\sigma(f(x))" + }, + { + "bbox": [ + 47, + 341, + 288, + 376 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 341, + 288, + 376 + ], + "type": "inline_equation", + "content": "\\sigma(z) = \\frac{1}{1 + e^{-z}}" + }, + { + "bbox": [ + 47, + 341, + 288, + 376 + ], + "type": "text", + "content": " is the sigmoid function." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 385, + 144, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 385, + 144, + 399 + ], + "spans": [ + { + "bbox": [ + 47, + 385, + 144, + 399 + ], + "type": "text", + "content": "3.2. Input Sampling" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 403, + 289, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 403, + 289, + 453 + ], + "spans": [ + { + "bbox": [ + 46, + 403, + 289, + 453 + ], + "type": "text", + "content": "Regularly or randomly sampling the input planes for training yields inaccurate and overly smooth results (see Fig. 12). Instead, we define three types of relevant point distributions to sample from, as depicted in Fig. 3:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 460, + 288, + 657 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 55, + 460, + 287, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 287, + 508 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 287, + 508 + ], + "type": "text", + "content": "1. We bound the reconstructed volume by the 3D convex hull of all the input contours, and scale it up by " + }, + { + "bbox": [ + 55, + 460, + 287, + 508 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 55, + 460, + 287, + 508 + ], + "type": "text", + "content": ". We consequently sample it uniformly, where the sampled points are all outside." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 517, + 287, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 517, + 287, + 552 + ], + "spans": [ + { + "bbox": [ + 54, + 517, + 287, + 552 + ], + "type": "text", + "content": "2. We compute the bounding box (aligned to the principal axes) of the contours in each plane, and sample it with uniform distribution." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 561, + 288, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 561, + 288, + 657 + ], + "spans": [ + { + "bbox": [ + 54, + 561, + 288, + 657 + ], + "type": "text", + "content": "3. Most samples are taken around the boundaries of the contours, as these are the regions where accuracy matters most. We then sample every contour edge evenly, and further add off-surface points of varying distances for each edge sample. Off-surface locations are found by moving away from the contour on the plane (i.e., in the direction of the plane normal). We further sample off-surface points in a circle around each vertex." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 46, + 665, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 289, + 714 + ], + "type": "text", + "content": "Each sampling point " + }, + { + "bbox": [ + 46, + 665, + 289, + 714 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 46, + 665, + 289, + 714 + ], + "type": "text", + "content": " is matched with a label " + }, + { + "bbox": [ + 46, + 665, + 289, + 714 + ], + "type": "inline_equation", + "content": "Y_{i}" + }, + { + "bbox": [ + 46, + 665, + 289, + 714 + ], + "type": "text", + "content": " according to its position on the slice, and the pairs " + }, + { + "bbox": [ + 46, + 665, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\{x_{i}, Y_{i}\\}" + }, + { + "bbox": [ + 46, + 665, + 289, + 714 + ], + "type": "text", + "content": " constitute the input to our training. See supplementary material for the relevant hyperparameters." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "type": "text", + "content": "Frequency-Oriented Sampling. In order to encourage better high-frequency learning, we sample the " + }, + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "type": "inline_equation", + "content": "3^{rd}" + }, + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "type": "text", + "content": " type of points around a set of varying distances, from " + }, + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "type": "inline_equation", + "content": "0.1\\%" + }, + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "type": "text", + "content": " of the bounding box away from the contours to three orders of magnitude closer. In every epoch, we only use points sampled around three consecutive distance ranges. For early training iterations, we use the three largest distances, since further away points translate to lower frequency information about the shape. As the training progresses, and the lower frequencies are assumed to stabilize, we train with points that grow closer and closer to the actual contour, thus focusing the learning process on higher and higher frequencies (see Fig. 11). See supplementary material for more details and exact scheduling." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 248, + 388, + 260 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 248, + 388, + 260 + ], + "spans": [ + { + "bbox": [ + 306, + 248, + 388, + 260 + ], + "type": "text", + "content": "3.3. Architecture" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 266, + 547, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 266, + 547, + 495 + ], + "spans": [ + { + "bbox": [ + 304, + 266, + 547, + 495 + ], + "type": "text", + "content": "ORexNet is simply an MLP that takes a 3D coordinate as input, represented using positional encoding, and produces the function " + }, + { + "bbox": [ + 304, + 266, + 547, + 495 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 304, + 266, + 547, + 495 + ], + "type": "text", + "content": " (Sec. 3.1) at any query point. In order to encourage high-frequency details, we introduce an iterative refinement mechanism. Inspired by recent work [1], we allow the network to refine its own results by running them through the network for a small number of iterations. This process was previously shown to produce a coarse-to-fine evolution in the realm of 2D images [1], and we argue it applies to our case as well. As demonstrated in Fig. 2, ORexNet is hence a residual MLP that is fed the result of the previous iteration, and a small hidden state code, along with the query point. These iterations (10 in our experiments) are performed both during training and inference; at inference time, only the last result is taken, and during training, the loss is applied to the results of all steps. The first iteration starts from a learned constant " + }, + { + "bbox": [ + 304, + 266, + 547, + 495 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 266, + 547, + 495 + ], + "type": "text", + "content": ". We show that this process indeed sharpens features and allows the incorporation of higher-frequency details in Fig. 13." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 503, + 463, + 515 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 503, + 463, + 515 + ], + "spans": [ + { + "bbox": [ + 306, + 503, + 463, + 515 + ], + "type": "text", + "content": "3.4. Loss Function and Inference" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 522, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 547, + 713 + ], + "type": "text", + "content": "In order to present our loss function, we must first attend to an issue in the final stage of our pipeline. In this stage, we extract an explicit mesh for the boundary of the indicator function using Dual Contouring [15]. This step uses a discretely sampled version of " + }, + { + "bbox": [ + 304, + 522, + 547, + 713 + ], + "type": "inline_equation", + "content": "Y(x)" + }, + { + "bbox": [ + 304, + 522, + 547, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 522, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\nabla Y(x)" + }, + { + "bbox": [ + 304, + 522, + 547, + 713 + ], + "type": "text", + "content": " on a regular 3D grid to extract the mesh. As can be seen in Fig. 8 (and as witnessed other works [9]), this creates a ripples-like artifact, that correlates to the 3D grid resolution. This results from an aliasing of the sampling in these regions where the gradient magnitudes exhibit high variance. This effect could be mitigated via higher resolution grid sampling, which is expensive, or a post-process operation that may compromise the geometry. Instead, we propose revisiting the loss function of the training process, and incorporating regularization to reduce this sharp drop in value at the shape boundary. First, we omit the activation—in other words, we use the" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20857" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 75, + 532, + 336 + ], + "blocks": [ + { + "bbox": [ + 62, + 75, + 532, + 336 + ], + "lines": [ + { + "bbox": [ + 62, + 75, + 532, + 336 + ], + "spans": [ + { + "bbox": [ + 62, + 75, + 532, + 336 + ], + "type": "image", + "image_path": "aaa56d05df9f225105016f491599929176101356d798e60e998657efd31547a8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 345, + 545, + 357 + ], + "lines": [ + { + "bbox": [ + 47, + 345, + 545, + 357 + ], + "spans": [ + { + "bbox": [ + 47, + 345, + 545, + 357 + ], + "type": "text", + "content": "Figure 4. Results gallery. A qualitative demonstration of a collection of man-made and medical inputs. Zoomed-in viewing is recommended." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 50, + 374, + 298, + 517 + ], + "blocks": [ + { + "bbox": [ + 50, + 374, + 298, + 517 + ], + "lines": [ + { + "bbox": [ + 50, + 374, + 298, + 517 + ], + "spans": [ + { + "bbox": [ + 50, + 374, + 298, + 517 + ], + "type": "image", + "image_path": "2d05ff135b53b6e06f211fb095439b6119caac82a2573d5bb4c1f8c288ae8250.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 525, + 287, + 559 + ], + "lines": [ + { + "bbox": [ + 46, + 525, + 287, + 559 + ], + "spans": [ + { + "bbox": [ + 46, + 525, + 287, + 559 + ], + "type": "text", + "content": "Figure 5. Additional qualitative results, where transparency reveals the reconstruction of internal cavities and tunnels. Zoomed-in viewing is recommended." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": "function " + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "inline_equation", + "content": "f(x)" + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": " for contouring, while using " + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "inline_equation", + "content": "Y(x)" + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": " only for training (Sec 3.1). This de-radicalizes the function values, already offering a softer transition between inside and outside regions. In addition, we explicitly penalize strong transitions by using a hinge loss (Eq. 2). With this, we reduce the gradient magnitude variance considerably by limiting them close to the input shape boundary. Our loss function for a single" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 378, + 367, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 378, + 367, + 389 + ], + "spans": [ + { + "bbox": [ + 306, + 378, + 367, + 389 + ], + "type": "text", + "content": "point " + }, + { + "bbox": [ + 306, + 378, + 367, + 389 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 306, + 378, + 367, + 389 + ], + "type": "text", + "content": " is then" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 399, + 547, + 441 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 399, + 547, + 441 + ], + "spans": [ + { + "bbox": [ + 306, + 399, + 547, + 441 + ], + "type": "interline_equation", + "content": "\\mathcal {L} (x, \\theta) = \\sum_ {i = 0} ^ {N - 1} \\operatorname {B C E} \\left(Y _ {i} (x)\\right) + \\lambda \\max (0, | | \\nabla f _ {N - 1} (x) | | - \\alpha) \\tag {2}", + "image_path": "544dea31bfbbeaa1d9d2ae8024c93cd42df2f1e88d4636db422211515cf7a720.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 441, + 547, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 441, + 547, + 514 + ], + "spans": [ + { + "bbox": [ + 304, + 441, + 547, + 514 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 304, + 441, + 547, + 514 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 304, + 441, + 547, + 514 + ], + "type": "text", + "content": " is the sigmoid function, " + }, + { + "bbox": [ + 304, + 441, + 547, + 514 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 441, + 547, + 514 + ], + "type": "text", + "content": " are the parameters of the network, " + }, + { + "bbox": [ + 304, + 441, + 547, + 514 + ], + "type": "inline_equation", + "content": "BCE()" + }, + { + "bbox": [ + 304, + 441, + 547, + 514 + ], + "type": "text", + "content": " is the binary cross-entropy loss applied to all outputs of our iterative scheme, " + }, + { + "bbox": [ + 304, + 441, + 547, + 514 + ], + "type": "inline_equation", + "content": "max(0,x)" + }, + { + "bbox": [ + 304, + 441, + 547, + 514 + ], + "type": "text", + "content": " is a hinge loss applied only to the last OReXnet iteration, " + }, + { + "bbox": [ + 304, + 441, + 547, + 514 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 441, + 547, + 514 + ], + "type": "text", + "content": " the number of iterations, and " + }, + { + "bbox": [ + 304, + 441, + 547, + 514 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 441, + 547, + 514 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 441, + 547, + 514 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 441, + 547, + 514 + ], + "type": "text", + "content": " are hyperparameters. Fig. 8 demonstrates the correction effect of the hinge loss." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 523, + 387, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 523, + 387, + 537 + ], + "spans": [ + { + "bbox": [ + 306, + 523, + 387, + 537 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 544, + 547, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 544, + 547, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 544, + 547, + 604 + ], + "type": "text", + "content": "To evaluate our method, we developed a prototype, and ran it on a NVIDIA GeForce GTX 1080 Ti. See the Supplementary Material for exact timing and memory consumption statistics. Generally speaking, all training processes took less than four hours on the single GPU." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 611, + 448, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 611, + 448, + 624 + ], + "spans": [ + { + "bbox": [ + 306, + 611, + 448, + 624 + ], + "type": "text", + "content": "4.1. Results and Comparisons" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "type": "text", + "content": "We first qualitatively demonstrate the result of our method on a variety of slice inputs, from both the medical and graphics worlds (Fig. 4). This demonstrates the versatility and generality of our method. We further show our our algorithm correctly reproduces the internal cavities and details of reconstructed objects in Fig. 5. Note how intricate details are learned along side a smooth interpolation between slices," + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20858" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 68, + 512, + 350 + ], + "blocks": [ + { + "bbox": [ + 47, + 68, + 512, + 350 + ], + "lines": [ + { + "bbox": [ + 47, + 68, + 512, + 350 + ], + "spans": [ + { + "bbox": [ + 47, + 68, + 512, + 350 + ], + "type": "image", + "image_path": "5864da259dc40994c5510a921b02c5ff5ea6b4e1ade1d8d8f76918ba0b663a8e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 358, + 547, + 383 + ], + "lines": [ + { + "bbox": [ + 46, + 358, + 547, + 383 + ], + "spans": [ + { + "bbox": [ + 46, + 358, + 547, + 383 + ], + "type": "text", + "content": "Figure 6. Qualitative comparisons to other approaches. Our result is smoother, more aligned to features, and with less artifacts. Note how the reconstructed shape silhouettes does not suffer imprints from the input planes. Check Table 1 for quantitative comparison." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 403, + 279, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 403, + 279, + 415 + ], + "spans": [ + { + "bbox": [ + 46, + 403, + 279, + 415 + ], + "type": "text", + "content": "leaving no slice transition artifacts on the resulting shape." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 415, + 289, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 415, + 289, + 583 + ], + "spans": [ + { + "bbox": [ + 46, + 415, + 289, + 583 + ], + "type": "text", + "content": "In terms of other methods, we qualitatively (Fig. 6) and quantitatively (Table 1) compare our method to Bermano et al. [4], which is a reconstruction from cross-section method with the same input-output as ours. We further compare to general state-of-the-art reconstruction methods that target point clouds [12]. Hanocka et al. [12] expect point normals as input as well. Hence we compare our result to the latter work with Ground Truth (GT) normals at input, and with ones coming from the input plane normals for a fairer comparison. We use the following metrics: 1) intersection-over-union volume of inside regions in reconstructed volumes and 2) GT planes, and 3) symmetric Hausdorff distance. These metrics demonstrate our superiority over the compared methods. See Supplementary Material for more comparisons." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 601, + 288, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 601, + 288, + 662 + ], + "spans": [ + { + "bbox": [ + 46, + 601, + 288, + 662 + ], + "type": "text", + "content": "**Scalability** We show that our method does not suffer the computational cost that arrangement-based methods [4, 14, 16] must bear, since we do not tessellate the space, and hence scale well with the input. In addition, we show (Fig. 7) how our method converges with the addition of slices." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 670, + 116, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 116, + 682 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 116, + 682 + ], + "type": "text", + "content": "4.2. Ablations" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "text", + "content": "We perform ablation testing to evaluate important aspects of our method and our design choices." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 308, + 404, + 556, + 498 + ], + "blocks": [ + { + "bbox": [ + 308, + 404, + 556, + 498 + ], + "lines": [ + { + "bbox": [ + 308, + 404, + 556, + 498 + ], + "spans": [ + { + "bbox": [ + 308, + 404, + 556, + 498 + ], + "type": "image", + "image_path": "8b367136416dde2717c7abc95eec0e40af5edfac6287f63649a5b034fecac0f7.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 509, + 547, + 533 + ], + "lines": [ + { + "bbox": [ + 305, + 509, + 547, + 533 + ], + "spans": [ + { + "bbox": [ + 305, + 509, + 547, + 533 + ], + "type": "text", + "content": "Figure 7. Increasing number of slices. As can be expected, introducing more slices converges consistently converges to the solution." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 558, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 547, + 713 + ], + "type": "text", + "content": "Gradient magnitude regularization (Fig. 8) We show the effect of the hinge-loss regularizer on gradient magnitudes, with increasing " + }, + { + "bbox": [ + 304, + 558, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 558, + 547, + 713 + ], + "type": "text", + "content": " values. It is evident the hinge loss effectively filters out the high variance, and smooths the ripple artifacts. As can be seen (bottom row), without our hinge loss, geometry vertices tend to snap to discrete locations during the mesh extraction phase. This is because a large gradient induces a large bias in the magnitude of the indicator function, encouraging vertices to remain close to the grid points sampled during the extraction step. Regularizing the gradients around the boundary regions (top row) decreases the gap between the positive and negative sides of the function, allowing a more uniform, and hence continuous, distribution" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20859" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 288, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 288, + 108 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 288, + 108 + ], + "type": "text", + "content": "of the mesh vertices. With this bias removed, the resulting geometry is better sampled and hence the ripple artifact is mitigated (middle)." + } + ] + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 55, + 122, + 280, + 312 + ], + "blocks": [ + { + "bbox": [ + 55, + 122, + 280, + 312 + ], + "lines": [ + { + "bbox": [ + 55, + 122, + 280, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 122, + 280, + 312 + ], + "type": "image", + "image_path": "30a8e65595c84cd5c6cc6264b27abefaca22af12dc1ead2e5adfecffdb9659fb.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 323, + 288, + 422 + ], + "lines": [ + { + "bbox": [ + 46, + 323, + 288, + 422 + ], + "spans": [ + { + "bbox": [ + 46, + 323, + 288, + 422 + ], + "type": "text", + "content": "Figure 8. Left to right: our Hinge loss effect for mitigating ripple-like aliasing artifacts. Top: output values of " + }, + { + "bbox": [ + 46, + 323, + 288, + 422 + ], + "type": "inline_equation", + "content": "f(x)" + }, + { + "bbox": [ + 46, + 323, + 288, + 422 + ], + "type": "text", + "content": " on a single slice (arrows show gradients). Middle: The reconstructed mesh with a closer view on the top of the skull. Bottom: distribution of vertices along grid edges during mesh extraction. It is evident the hinge loss effectively controls gradient magnitudes at the zero set, curbing the aliasing artifacts. The hinge regularization further allows to place more vertices in the middle of the sampled grid cells, and not only on their edges." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 452, + 287, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 452, + 287, + 513 + ], + "spans": [ + { + "bbox": [ + 47, + 452, + 287, + 513 + ], + "type": "text", + "content": "Iterative architecture (Fig. 9) This ablations examines the effect of the number of refinement iterations on the accuracy of the mesh produced. As can be expected, the results show that as the number of iterations increases, high-frequency details are better represented." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 50, + 523, + 285, + 646 + ], + "blocks": [ + { + "bbox": [ + 50, + 523, + 285, + 646 + ], + "lines": [ + { + "bbox": [ + 50, + 523, + 285, + 646 + ], + "spans": [ + { + "bbox": [ + 50, + 523, + 285, + 646 + ], + "type": "image", + "image_path": "3852e3adc0901289721be3a1cea9dad6a53f356651fbe67c30786541190b859f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 664, + 287, + 698 + ], + "lines": [ + { + "bbox": [ + 47, + 664, + 287, + 698 + ], + "spans": [ + { + "bbox": [ + 47, + 664, + 287, + 698 + ], + "type": "text", + "content": "Figure 9. Iterative refinement experiment. Training and inferring with more OReXNet refinement iterations allows the network to perform smaller scale corrections and reduces spectral bias." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 228 + ], + "type": "text", + "content": "Sampling method To evaluate our importance sampling scheme, we first replace it with regular grid sampling in two resolutions (Fig. 10). As can be seen, using each of the resolutions alone yields overly smoothed results due to spectral bias. Training with a gradually increasing resolution allows better handling of high-frequency details. In addition, we have demonstrate the effects on quality when using our importance scheme in a non-hierarchical manner, i.e. using all points uniformly without consideration of distance from the boundary (Fig. 11). As can be seen, our scheme improves in detail quality upon both experiments, demonstrating that our sampling scheme produces more accurate results with less spectral bias while using fewer samples." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 310, + 238, + 545, + 369 + ], + "blocks": [ + { + "bbox": [ + 310, + 238, + 545, + 369 + ], + "lines": [ + { + "bbox": [ + 310, + 238, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 310, + 238, + 545, + 369 + ], + "type": "image", + "image_path": "dd91b74219f224ad277f141b755395c9ce44fc9fb7bd03f183df50409134bf64.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 380, + 547, + 414 + ], + "lines": [ + { + "bbox": [ + 305, + 380, + 547, + 414 + ], + "spans": [ + { + "bbox": [ + 305, + 380, + 547, + 414 + ], + "type": "text", + "content": "Figure 10. Sampling scheme Ablations. Left: ground truth. Middle: grid sampling with different resolutions, and a gradually increasing sampling resolution over the training process. Right: our method." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 325, + 448, + 547, + 569 + ], + "blocks": [ + { + "bbox": [ + 325, + 448, + 547, + 569 + ], + "lines": [ + { + "bbox": [ + 325, + 448, + 547, + 569 + ], + "spans": [ + { + "bbox": [ + 325, + 448, + 547, + 569 + ], + "type": "image", + "image_path": "257734fcec337945ae835bb473ac3c885d312e63ab418626c197090aa3c9dc1c.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 588, + 546, + 622 + ], + "lines": [ + { + "bbox": [ + 305, + 588, + 546, + 622 + ], + "spans": [ + { + "bbox": [ + 305, + 588, + 546, + 622 + ], + "type": "text", + "content": "Figure 11. Left: non-hierarchical sampling, using points from all off-surface distances uniformly. Right: using our hierarchical sampling scheme." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 653, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 546, + 713 + ], + "type": "text", + "content": "Choice of architecture and sampling (Fig. 12) We justify our design choices by training a baseline architecture with uniform grid sampling and no iteration refinement. It is evident such a model does not preserve details nor interpolate well, even for a relatively simple case." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20860" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 71, + 547, + 148 + ], + "blocks": [ + { + "bbox": [ + 48, + 71, + 547, + 148 + ], + "lines": [ + { + "bbox": [ + 48, + 71, + 547, + 148 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 547, + 148 + ], + "type": "table", + "html": "
InputHausdorff distanceIoU in 3dIoU in 2d
ORExBermano et. alpoint2mesh plane normalspoint2mesh GT normalsORExBermano et. alpoint2mesh plane normalspoint2mesh GT normalsORExBermano et. alpoint2mesh plane normalspoint2mesh GT normals
Eight (S)0.0180.0650.2190.0460.9840.8650.8420.9610.9880.9840.7950.980
Eight (M)0.0060.0330.0450.0140.9870.8930.9150.9740.9860.9760.9610.971
Elephant0.0560.0810.1000.0860.9660.9080.8850.9350.9750.9690.8500.908
Balloon dog0.0490.1940.0780.0860.9570.8680.8970.9280.9880.9770.9260.956
Hand OK0.0630.1770.1950.1350.9550.8600.9210.9310.9870.9680.9080.882
Armadillo0.0500.1210.0570.0590.9390.8910.8910.9210.9640.7760.8500.868
", + "image_path": "ad33d7f6008402f492e8d468ace478a94ef461f9ccdd9406eddbe3f8a0a51036.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 156, + 547, + 191 + ], + "lines": [ + { + "bbox": [ + 46, + 156, + 547, + 191 + ], + "spans": [ + { + "bbox": [ + 46, + 156, + 547, + 191 + ], + "type": "text", + "content": "Table 1. Quantitative comparisons. We measure performance using Hausdorff distance, IoU of the inner volume compared to the GT shape, and IoU of the inner surface on the input cross-sections. We compare our result to a dedicated cross-section based reconstruction method [4], and to two flavors of a pointcloud reconstruction method [12]. See more comparisons in the Supplementary Material." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 48, + 225, + 298, + 338 + ], + "blocks": [ + { + "bbox": [ + 48, + 225, + 298, + 338 + ], + "lines": [ + { + "bbox": [ + 48, + 225, + 298, + 338 + ], + "spans": [ + { + "bbox": [ + 48, + 225, + 298, + 338 + ], + "type": "image", + "image_path": "0867482e84d4e25861d89138a15b1e6b5756b1775009965359e10ee8626c7648.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 357, + 288, + 403 + ], + "lines": [ + { + "bbox": [ + 46, + 357, + 288, + 403 + ], + "spans": [ + { + "bbox": [ + 46, + 357, + 288, + 403 + ], + "type": "text", + "content": "Figure 12. Left and middle: the baseline experiment with regular grid samples of 64 and 128 resolution over the slices. Right: our result for the same amount for samples (our variable number of samples is provided at its mean)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 53, + 420, + 279, + 541 + ], + "blocks": [ + { + "bbox": [ + 53, + 420, + 279, + 541 + ], + "lines": [ + { + "bbox": [ + 53, + 420, + 279, + 541 + ], + "spans": [ + { + "bbox": [ + 53, + 420, + 279, + 541 + ], + "type": "image", + "image_path": "73edd476b5527a81cf34571a53dd7d7feabb4a2897acd525b7241e2605c4554a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 556, + 288, + 624 + ], + "lines": [ + { + "bbox": [ + 46, + 556, + 288, + 624 + ], + "spans": [ + { + "bbox": [ + 46, + 556, + 288, + 624 + ], + "type": "text", + "content": "Figure 13. Refinement progression throughout the iteration process. The maximum change in the output value decreases along the iterations. Right: points that were added to the shape (Red) and that were removed from it (blue) in the last refinement iteration. These are concentrated on the level set, whereas most other points (green, covering all internal regions) have already converged." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 644, + 125, + 657 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 644, + 125, + 657 + ], + "spans": [ + { + "bbox": [ + 47, + 644, + 125, + 657 + ], + "type": "text", + "content": "5. Conclusions" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 665, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 289, + 715 + ], + "type": "text", + "content": "We have presented OREx, a state-of-the-art method for the long-standing problem of shape reconstruction from planar cross-sectional indicator data. Free of datasets and training requirements, OREx is simple and intuitive to use. Our" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 210, + 547, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 210, + 547, + 281 + ], + "spans": [ + { + "bbox": [ + 304, + 210, + 547, + 281 + ], + "type": "text", + "content": "work balances the smoothness of a neural prior with high-frequency features. We show our approach successfully produces smooth interpolation between contours while respecting high-frequency features and repeating patterns. In addition, we believe some of the analysis and insights presented here can be applied to neural fields in similar tasks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 283, + 547, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 283, + 547, + 390 + ], + "spans": [ + { + "bbox": [ + 304, + 283, + 547, + 390 + ], + "type": "text", + "content": "The advantage of our method is also its disadvantage: using only binary data, much of the information is lost in the process. For example, for medical imaging an interesting future direction would be to directly use the raw grayscale density input of a planar probe such as that comes from an ultrasound. Another interesting direction for research is extending this work for multi-labeled volumes (e.g., the reconstruction of several organs simultaneously from a scan), and using partial or noisy data." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 307, + 401, + 365, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 401, + 365, + 414 + ], + "spans": [ + { + "bbox": [ + 307, + 401, + 365, + 414 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 310, + 421, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 310, + 421, + 547, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 421, + 547, + 465 + ], + "spans": [ + { + "bbox": [ + 310, + 421, + 547, + 465 + ], + "type": "text", + "content": "[1] Yuval Alaluf, Or Patashnik, and Daniel Cohen-Or. Restyle: A residual-based stylegan encoder via iterative refinement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6711–6720, 2021. 2, 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 311, + 466, + 547, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 466, + 547, + 510 + ], + "spans": [ + { + "bbox": [ + 311, + 466, + 547, + 510 + ], + "type": "text", + "content": "[2] Chandrajit L Bajaj, Edward J Coyle, and Kwun-Nan Lin. Arbitrary topology shape reconstruction from planar cross sections. Graphical models and image processing, 58(6):524-543, 1996. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 311, + 511, + 547, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 511, + 547, + 555 + ], + "spans": [ + { + "bbox": [ + 311, + 511, + 547, + 555 + ], + "type": "text", + "content": "[3] Gill Barequet and Micha Sharir. Piecewise-linear interpolation between polygonal slices. In Proceedings of the tenth annual symposium on Computational geometry, pages 93-102, 1994. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 311, + 556, + 547, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 556, + 547, + 599 + ], + "spans": [ + { + "bbox": [ + 311, + 556, + 547, + 599 + ], + "type": "text", + "content": "[4] Amit Bermano, Amir Vaxman, and Craig Gotsman. Online reconstruction of 3d objects from arbitrary cross-sections. ACM Transactions on Graphics (TOG), 30(5):1-11, 2011. 2, 6, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 311, + 601, + 547, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 601, + 547, + 634 + ], + "spans": [ + { + "bbox": [ + 311, + 601, + 547, + 634 + ], + "type": "text", + "content": "[5] Jean-Daniel Boissonnat and Pooran Memari. Shape reconstruction from unorganized cross-sections. In Symposium on geometry processing, pages 89-98, 2007. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 311, + 635, + 547, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 635, + 547, + 679 + ], + "spans": [ + { + "bbox": [ + 311, + 635, + 547, + 679 + ], + "type": "text", + "content": "[6] Alexandre Boulch and Renaud Marlet. Poco: Point convolution for surface reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6302-6314, 2022. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 311, + 681, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 681, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 311, + 681, + 547, + 713 + ], + "type": "text", + "content": "[7] Zhang Chen, Yinda Zhang, Kyle Genova, Sean Fanello, Sofien Bouaziz, Christian Hane, Ruofei Du, Cem Keskin, Thomas Funkhouser, and Danhang Tang. Multiresolution" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20861" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 289, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 66, + 72, + 288, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 288, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 288, + 106 + ], + "type": "text", + "content": "deep implicit functions for 3d shape representation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13087-13096, 2021. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 108, + 289, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 108, + 289, + 163 + ], + "spans": [ + { + "bbox": [ + 53, + 108, + 289, + 163 + ], + "type": "text", + "content": "[8] Julian Chibane, Thiemo Alldieck, and Gerard Pons-Moll. Implicit functions in feature space for 3d shape reconstruction and completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6970-6981, 2020. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 164, + 289, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 164, + 289, + 209 + ], + "spans": [ + { + "bbox": [ + 53, + 164, + 289, + 209 + ], + "type": "text", + "content": "[9] Philipp Erler, Paul Guerrero, Stefan Ohrhallinger, Niloy J Mitra, and Michael Wimmer. Points2surf learning implicit surfaces from point clouds. In European Conference on Computer Vision, pages 108-124. Springer, 2020. 2, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 210, + 288, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 210, + 288, + 255 + ], + "spans": [ + { + "bbox": [ + 48, + 210, + 288, + 255 + ], + "type": "text", + "content": "[10] Kyle Gao, Yina Gao, Hongjie He, Denning Lu, Linlin Xu, and Jonathan Li. Nerf: Neural radiance field in 3d vision, a comprehensive review. arXiv preprint arXiv:2210.00379, 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 256, + 288, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 256, + 288, + 300 + ], + "spans": [ + { + "bbox": [ + 48, + 256, + 288, + 300 + ], + "type": "text", + "content": "[11] Henry Ham, Julian Wesley, and Hendra Hendra. Computer vision based 3d reconstruction: A review. International Journal of Electrical and Computer Engineering, 9(4):2394, 2019. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 302, + 288, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 302, + 288, + 335 + ], + "spans": [ + { + "bbox": [ + 48, + 302, + 288, + 335 + ], + "type": "text", + "content": "[12] Rana Hanocka, Gal Metzer, Raja Giryes, and Daniel Cohen-Or. Point2mesh: A self-prior for deformable meshes. arXiv preprint arXiv:2005.11084, 2020. 2, 3, 6, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 337, + 288, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 337, + 288, + 380 + ], + "spans": [ + { + "bbox": [ + 48, + 337, + 288, + 380 + ], + "type": "text", + "content": "[13] Michelle Holloway, Cindy Grimm, and Tao Ju. Template-based surface reconstruction from cross-sections. Computers & Graphics, 58:84-91, 2016. Shape Modeling International 2016. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 383, + 288, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 383, + 288, + 426 + ], + "spans": [ + { + "bbox": [ + 48, + 383, + 288, + 426 + ], + "type": "text", + "content": "[14] Zhiyang Huang, Ming Zou, Nathan Carr, and Tao Ju. Topology-controlled reconstruction of multi-labelled domains from cross-sections. ACM Transactions on Graphics (TOG), 36(4):1-12, 2017. 2, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 429, + 288, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 429, + 288, + 473 + ], + "spans": [ + { + "bbox": [ + 48, + 429, + 288, + 473 + ], + "type": "text", + "content": "[15] Tao Ju, Frank Losasso, Scott Schaefer, and Joe Warren. Dual contouring of hermite data. In Proceedings of the 29th annual conference on Computer graphics and interactive techniques, pages 339-346, 2002. 3, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 475, + 288, + 519 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 475, + 288, + 519 + ], + "spans": [ + { + "bbox": [ + 48, + 475, + 288, + 519 + ], + "type": "text", + "content": "[16] Lu Liu, Chandrajit Bajaj, Joseph O Deasy, Daniel A Low, and Tao Ju. Surface reconstruction from non-parallel curve networks. In Computer Graphics Forum, volume 27, pages 155–163. Wiley Online Library, 2008. 2, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 520, + 288, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 520, + 288, + 575 + ], + "spans": [ + { + "bbox": [ + 48, + 520, + 288, + 575 + ], + "type": "text", + "content": "[17] Baorui Ma, Zhizhong Han, Yu-Shen Liu, and Matthias Zwicker. Neural-pull: Learning signed distance function from point clouds by learning to pull space onto surface. In International Conference on Machine Learning, pages 7246-7257. PMLR, 2021. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 577, + 288, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 577, + 288, + 621 + ], + "spans": [ + { + "bbox": [ + 48, + 577, + 288, + 621 + ], + "type": "text", + "content": "[18] Julien NP Martel, David B Lindell, Connor Z Lin, Eric R Chan, Marco Monteiro, and Gordon Wetzstein. Acorn: Adaptive coordinate networks for neural scene representation. arXiv preprint arXiv:2105.02788, 2021. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 624, + 288, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 288, + 677 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 288, + 677 + ], + "type": "text", + "content": "[19] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 680, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 288, + 714 + ], + "type": "text", + "content": "[20] Azimkhon Ostonov. Cut-and-approximate: 3d shape reconstruction from planar cross-sections with deep reinforcement learning. arXiv preprint arXiv:2210.12509, 2022. 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 373 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 116 + ], + "type": "text", + "content": "[21] Songyou Peng, Michael Niemeyer, Lars Mescheder, Marc Pollefeys, and Andreas Geiger. Convolutional occupancy networks. In European Conference on Computer Vision, pages 523-540. Springer, 2020. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 547, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 547, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 547, + 172 + ], + "type": "text", + "content": "[22] Nasim Rahaman, Aristide Baratin, Devansh Arpit, Felix Draxler, Min Lin, Fred Hamprecht, Yoshua Bengio, and Aaron Courville. On the spectral bias of neural networks. In International Conference on Machine Learning, pages 5301-5310. PMLR, 2019. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 174, + 547, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 174, + 547, + 205 + ], + "spans": [ + { + "bbox": [ + 307, + 174, + 547, + 205 + ], + "type": "text", + "content": "[23] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 207, + 547, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 207, + 547, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 207, + 547, + 274 + ], + "type": "text", + "content": "[24] Towaki Takikawa, Joey Litalien, Kangxue Yin, Karsten Kreis, Charles Loop, Derek Nowrouzezahrai, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Neural geometric level of detail: Real-time rendering with implicit 3d shapes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11358-11367, 2021. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 274, + 547, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 274, + 547, + 329 + ], + "spans": [ + { + "bbox": [ + 307, + 274, + 547, + 329 + ], + "type": "text", + "content": "[25] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond. In Computer Graphics Forum, volume 41, pages 641-676. Wiley Online Library, 2022. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 331, + 547, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 331, + 547, + 373 + ], + "spans": [ + { + "bbox": [ + 307, + 331, + 547, + 373 + ], + "type": "text", + "content": "[26] Ming Zou, Michelle Holloway, Nathan Carr, and Tao Ju. Topology-constrained surface reconstruction from cross-sections. ACM Transactions on Graphics (TOG), 34(4):1-10, 2015. 2" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20862" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/OSAN_ A One-Stage Alignment Network To Unify Multimodal Alignment and Unsupervised Domain Adaptation/4429e1d5-6fb8-40a5-9804-367d3c0ae901_content_list.json b/2023/OSAN_ A One-Stage Alignment Network To Unify Multimodal Alignment and Unsupervised Domain Adaptation/4429e1d5-6fb8-40a5-9804-367d3c0ae901_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8659c525dd411dc81116b7c42007b115a7aabeed --- /dev/null +++ b/2023/OSAN_ A One-Stage Alignment Network To Unify Multimodal Alignment and Unsupervised Domain Adaptation/4429e1d5-6fb8-40a5-9804-367d3c0ae901_content_list.json @@ -0,0 +1,1497 @@ +[ + { + "type": "text", + "text": "OSAN: A One-Stage Alignment Network to Unify Multimodal Alignment and Unsupervised Domain Adaptation", + "text_level": 1, + "bbox": [ + 94, + 128, + 875, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ye Liu, Lingfeng Qiao, Changchong Lu, Di Yin, Chen Lin, Haoyuan Peng, Bo Ren \nTencent Youtu Lab", + "bbox": [ + 155, + 203, + 815, + 237 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{rafelliu, leafqiao, chonglu, endymecyyin, iwlliamlin, haoyuanpeng, timren}@tencent.com", + "bbox": [ + 153, + 241, + 805, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 291, + 313, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Extending from unimodal to multimodal is a critical challenge for unsupervised domain adaptation (UDA). Two major problems emerge in unsupervised multimodal domain adaptation: domain adaptation and modality alignment. An intuitive way to handle these two problems is to fulfill these tasks in two separate stages: aligning modalities followed by domain adaptation, or vice versa. However, domains and modalities are not associated in most existing two-stage studies, and the relationship between them is not leveraged which can provide complementary information to each other. In this paper, we unify these two stages into one to align domains and modalities simultaneously. In our model, a tensor-based alignment module (TAL) is presented to explore the relationship between domains and modalities. By this means, domains and modalities can interact sufficiently and guide them to utilize complementary information for better results. Furthermore, to establish a bridge between domains, a dynamic domain generator (DDG) module is proposed to build transitional samples by mixing the shared information of two domains in a self-supervised manner, which helps our model learn a domain-invariant common representation space. Extensive experiments prove that our method can achieve superior performance in two real-world applications. The code will be publicly available.", + "bbox": [ + 75, + 325, + 473, + 690 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 722, + 209, + 738 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With explosively emerging multimedia data on the Internet, the field of multimodal analysis achieves more and more attention [10, 13, 18, 19, 43]. Compared to extensive unimodal models in NLP and CV, learning adequate knowledge from multimodal signals is still preliminary but very important. Abundant data plays a key role in different scenarios of multimodal analysis, such as pre-training or downstream multimedia tasks. However, it is prohibitively expensive and time-consuming to obtain large amounts of labeled data. To eliminate this issue, domain adaptation (DA)", + "bbox": [ + 75, + 750, + 470, + 902 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/884fcf7f5421700cb5403e4a222adf854fe52088a0f25c74bd6263691899488a.jpg", + "image_caption": [ + "Figure 1. Conception of our one-stage model." + ], + "image_footnote": [], + "bbox": [ + 513, + 292, + 890, + 417 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "is raised to learn a model from a labeled dataset (source domain) that can be generalized to other related tasks without sufficient labeled data (target domain) [3]. Classical domain adaptation can be classified into different categories: unsupervised domain adaptation (UDA), fully supervised domain adaptation, and semi-supervised domain adaptation [31]. In this paper, we focus on UDA where no samples in target domain are annotated. With this technique, it is not necessary to prepare a customized training dataset for a specific task, but it can perform the task effectively and efficiently.", + "bbox": [ + 496, + 460, + 892, + 628 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "There are two challenges when applying domain adaptation to multimodal scenarios [17]: (1) how to align the source and target domains and remit domain discrepancy, and (2) how to align multiple modalities and leverage multimodal information. Most existing works address these two problems in two consecutive stages: multimodal alignment followed by domain adaptation [34, 41], or vice versa [14, 44]. However, they solve these two issues separately without considering their relationship: domain and modality can be treated as two views to portray the intrinsic characteristic of multimodal data [8], and the hidden underlying relationship in these two views can provide complementary information to each other. Unimodal domain adaptation methods can not work well in multimodal tasks due to the inability to preserve the relations between modalities at the same time. Through our experiments and analysis, we observe that the two-stage model could not achieve ideal performance. Fig.2 shows the learning curve of two-stage", + "bbox": [ + 496, + 628, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "3551", + "bbox": [ + 482, + 944, + 513, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "model during training phase by 800 iterations for the task of multimodal sentiment analysis. It can be found that the learning curve of two-stage model is oscillating and converges slowly, which indicates that two-stage model is probably not a superior solution. To handle these challenges, the objective of multimodal domain adaptation can be defined as: (1) Exploring the relationship between domains and modalities; (2) Finding a common domain-invariant cross-modal representation space to align domains and modalities simultaneously.", + "bbox": [ + 75, + 90, + 472, + 241 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Therefore, in this paper, we design a One-Stage Alignment Network (OSAN) to unify multimodal alignment and domain adaptation in one stage. Fig.1 shows the conception of our one-stage model. Our method benefits from: (1) The modality and domain are associated and interacted to capture the relationship between domains and modalities, which can provide rich complementary information to each other. (2) Multimodal alignment and domain adaptation are unified in one stage, which allows our model to perform domain adaptation and leverage multimodal information at the same time. In Fig.2, we observe that the learning curve of our method is relatively stable and converges better, which indicates that exploring the relation between modality and domain contributes to our task.", + "bbox": [ + 75, + 242, + 472, + 454 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions are as follows:", + "bbox": [ + 96, + 454, + 398, + 468 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(1) To capture the relationship between domain and modality, we propose a one-stage alignment network, called OSAN, to associate domain and modality. In this way, a joint domain-invariant and cross-modal representation space is learned in one stage.", + "(2) We design a TAL module to bring sufficient interactions between domains and modalities and guide them to utilize complementary information for each other.", + "(3) To effectively bridge distinct domains, a DDG module is developed to dynamically construct multiple new domains by combining knowledge of source and target domains and exploring intrinsic structure of data distribution.", + "(4) Extensive experiments on two totally different tasks demonstrate the effectiveness of our method compared to the supervised and strongly UDA methods." + ], + "bbox": [ + 75, + 468, + 468, + 695 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 709, + 218, + 724 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Unimodal Domain Adaption The main approaches for domain adaptation or domain alignment can be categorized as discrepancy-based and adversarial-based methods [1, 23, 30]. Discrepancy-based methods design statistics, such as maximum mean discrepancy (MMD) [16], correlation alignment (CORAL) [45] and etc., to measure the difference between two domain distributions. For adversarial-based methods, Ganin et al. [4] used the concept of generative adversarial networks (GAN) to obtain domain-invariant characteristics. Wu et al. [32] proposed a one-stage adaptation framework for nighttime semantic segmentation to per", + "bbox": [ + 75, + 734, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/eddb63d74bc5cbe37d8b01f66f52ee8d5afa958fd9aa4be022d75dfe9406e75e.jpg", + "image_caption": [ + "Figure 2. The learning curves of our one-stage method and traditional two-stage method in multimodal sentiment analysis task." + ], + "image_footnote": [], + "bbox": [ + 513, + 93, + 882, + 280 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "form an end-to-end learning where a separate common preprocessing step was integrated, which is different from our motivation. However, these studies only focus on the elimination of domain shift in unimodal scenarios, which may lose effect when applied in multimodal area.", + "bbox": [ + 496, + 345, + 893, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multimodal Domain Adaption Qi et al. [17] first analyzed the multimodal domain adaptation and raised two problems, i.e., domain adaptation and modality alignment. To alleviate them, hybrid domain constraints and attention-based modality fusion module were introduced to learn domain-invariant fused features. More recently, some studies [14, 34, 41, 42, 44] related to this area have tried to address the multimodal problem by extending the unimodal domain adaptation methods. Xu et al. [34] proposed a supervised multimodal domain adaptation method for VQA to learn joint feature embeddings across different domains and modalities. Munro et al. [14] exploited the correspondence of modalities as a self-supervised alignment approach for UDA in addition to adversarial alignment. Furthermore, Zhou et al. [44] proposed MDMN for early rumor detection, which can combine textual and visual information with two heterogeneous feature extractors. However, all of these studies solved the two problems separately in two stages. Moreover, the relationship between modality and domain is missing, thus they cannot achieve ideal performance.", + "bbox": [ + 496, + 438, + 893, + 742 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Proposed Method", + "text_level": 1, + "bbox": [ + 500, + 753, + 671, + 771 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Fig.3 presents an overview of our method that contains four parts: multimodal feature extraction, tensor-based alignment, dynamic domain generator and task-specific heads. First, encoders, as multimodal feature extractor, map source and target multimodal data to different latent spaces. Then, by means of an effective and efficient way of tensor representation, we propose a TAL module to acquire relation information through consecutive interactions between", + "bbox": [ + 496, + 779, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "3552", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/58c94c1fa496aeb949ff7c4f381a7d8ad55528349d45db9521c07bd93c96f56c.jpg", + "image_caption": [ + "Figure 3. Overview of our method OSAN. We use circle, square and triangle to represent three modalities of text, vision and audio. In the figure, $\\times$ means mode- $d$ product of a tensor and a matrix, $\\odot$ represents element-wise product, and $\\ast$ means tensor factorization." + ], + "image_footnote": [], + "bbox": [ + 89, + 95, + 877, + 431 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "modalities and domains. Thus, the modality and domain are aligned simultaneously in one stage. After that, we develop a DDG module to mix information from different domains to create new transitional domains. Finally, our network is split into three branches: category classification, domain adversarial learning and discrepancy elimination.", + "bbox": [ + 75, + 496, + 470, + 587 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Tensor-based Alignment", + "text_level": 1, + "bbox": [ + 76, + 602, + 302, + 618 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We develop a TAL module to simultaneously address the challenges of multimodal alignment and domain adaptation. We first resort to tensor representation, an effective and natural representation of multimodal data, to develop models capturing inter-modality dynamics [11, 37]. Then, we utilize statistical measurements to establish an interaction between domains and modalities to explore the relation information. In this way, we associate domains and modalities so that they can communicate and align together. Specifically, we perform tensor factorization to fuse multimodal information and maximize statistical measurements to simultaneously eliminate domain gaps.", + "bbox": [ + 75, + 627, + 468, + 808 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Specifically, we use $\\underline{\\mathbf{X}}^{I_1\\times \\dots \\times I_N\\times M}$ and $\\underline{\\mathbf{Y}}^{J_1\\times \\dots \\times J_N\\times M}$ to represent source and target dataset with $N$ modalities and $M$ samples. For example, a video can be represented by a third-order tensor using a differentiable outer product between visual, textual, and acoustic representations [37]. For simplicity, we first make the training samples zero mean.", + "bbox": [ + 75, + 806, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To perform multimodal alignment, TAL aims to find pairs of linear transformations $\\mathbf{U}^{(n)}|_{n = 1}^{N}$ and $\\mathbf{V}^{(n)}|_{n = 1}^{N}$ for each modality of source and target domains to project samples of two sets into low dimensional subspaces. During this process, we establish an interaction between domain and modality by maximizing a statistical measurement of covariance given a normalized standard deviation. As reported in [45], we choose covariance instead of the first-order mean statistics used in MK-MMD [5, 21], as it is more powerful and achieves better performance. By this means, domains and modalities can be aligned simultaneously in one stage. Thus the objective function is defined as follows:", + "bbox": [ + 496, + 496, + 892, + 679 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\left\\{\\mathbf {U} ^ {(n)} | _ {n = 1} ^ {N}, \\mathbf {V} ^ {(n)} | _ {n = 1} ^ {N} \\right\\} \\\\ = \\max \\llbracket (\\underline {{\\mathbf {X}}} \\prod_ {n = 1} ^ {N} \\times_ {n} \\mathbf {U} ^ {(n) ^ {T}}) \\otimes (\\underline {{\\mathbf {Y}}} \\prod_ {n = 1} ^ {N} \\times_ {n} \\mathbf {V} ^ {(n) ^ {T}}); \\rrbracket \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 686, + 826, + 750 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} s. t. \\left(\\underline {{\\mathbf {X}}} \\prod_ {n = 1} ^ {N} \\times_ {n} \\mathbf {U} ^ {(n) ^ {T}}\\right) _ {(N + 1)} ^ {T} \\left(\\underline {{\\mathbf {X}}} \\prod_ {n = 1} ^ {N} \\times_ {n} \\mathbf {U} ^ {(n) ^ {T}}\\right) _ {(N + 1)} = \\mathbf {I} \\tag {1} \\\\ \\left(\\underline {{\\mathbf {Y}}} \\prod_ {n = 1} ^ {N} \\times_ {n} \\mathbf {V} ^ {(n) ^ {T}}\\right) _ {(N + 1)} ^ {T} \\left(\\underline {{\\mathbf {Y}}} \\prod_ {n = 1} ^ {N} \\times_ {n} \\mathbf {V} ^ {(n) ^ {T}}\\right) _ {(N + 1)} = \\mathbf {I} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 752, + 890, + 830 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The objective function above does not have a closed-form solution, so instead we derive a suboptimal solution following the principle of the alternating projection method [24], where the complicated optimization problem is re", + "bbox": [ + 498, + 839, + 892, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3553", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "duced into smaller conditional subproblems that can be solved by simple established methods. Therefore, the objective function is decomposed into $N$ different subproblems:", + "bbox": [ + 75, + 90, + 468, + 148 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{\\mathbf {U} ^ {(n)}, \\mathbf {V} ^ {(n)} \\right\\} = \\max t r \\left\\{\\mathbf {U} ^ {(n) ^ {T}} \\mathbf {C} _ {x y} ^ {(n)} \\mathbf {V} ^ {(n)} \\right\\} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 125, + 159, + 468, + 186 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ns. t. \\mathbf {U} ^ {(n) ^ {T}} \\mathbf {C} _ {x x} ^ {(n)} \\mathbf {U} ^ {(n)} = \\mathbf {I}, \\mathbf {V} ^ {(n) ^ {T}} \\mathbf {C} _ {y y} ^ {(n)} \\mathbf {V} ^ {(n)} = \\mathbf {I}\n$$\n", + "text_format": "latex", + "bbox": [ + 124, + 186, + 424, + 204 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where, for simplification, we define $\\mathbf{C}_{xy}^{(n)}$ , $\\mathbf{C}_{xx}^{(n)}$ as", + "bbox": [ + 76, + 217, + 410, + 236 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {C} _ {x y} ^ {(n)} = \\left(\\llbracket (\\underline {{\\mathbf {X}}} ^ {(\\bar {n})}) \\otimes (\\underline {{\\mathbf {Y}}} ^ {(\\bar {n})}); (\\bar {n}) (\\bar {n}) \\rrbracket\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 246, + 468, + 277 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {C} _ {x x} ^ {(n)} = \\left(\\llbracket (\\underline {{\\mathbf {X}}} ^ {(\\bar {n})}) \\otimes (\\underline {{\\mathbf {X}}} ^ {(\\bar {n})}); (\\bar {n}) (\\bar {n}) \\rrbracket\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 273, + 388, + 295 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $([\\cdot \\otimes ;(\\bar{n})(\\bar{n})])$ means tensor contraction on all indices except the $n$ -th index on the tensor product [25]. Similarly, $\\mathbf{C}_{yy}^{(n)}$ and $\\mathbf{C}_{yx}^{(n)}$ is same as $\\mathbf{C}_{xx}^{(n)}$ and $\\mathbf{C}_{xy}^{(n)}$ in form. By Eq.2, an interaction between source and target domains for the $n$ -th modality is established.", + "bbox": [ + 76, + 306, + 468, + 383 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Dynamic Domain Generator", + "text_level": 1, + "bbox": [ + 76, + 393, + 331, + 409 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In domain adaptation, adversarial learning is often used to align domains by developing a domain discriminator to judge samples from the source or target domains [27, 33]. By this means, it brings a two-class classification task with hard assignments of $\\mathbf{1}$ or $\\mathbf{0}$ . They try to confuse the domain discriminator to learn a shared common representation space between the source and target domains. However, it usually achieves worse performance due to a huge variance between the two domains.", + "bbox": [ + 75, + 417, + 468, + 551 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To overcome this problem, several works have been proposed to produce some transitional domains based on Mixup [35, 40]. They mixed images from the source and target domains to generate more soft and mixed samples [26, 33]. Suppose there are one sample $s$ from source domain and another sample $t$ from target domain, $s$ and $t$ are linearly interpolated [40] to generate a new sample. This strategy can fully utilize the inter-domain information and improve the generalization ability of models. However, it has an obvious shortcoming: the generated sample fuses full information of two domains, which contains redundant information and speciality of domains that may induce undesirable oscillations on training the domain discriminator.", + "bbox": [ + 75, + 553, + 468, + 748 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To solve this issue, we propose a DDG module to establish a bridge between domains. Specifically, DDG explicitly captures commonality and abandons the specialty of domains. By highlighting this commonality, we make the domain discriminator focus on commonality rather than full information, which helps our model learn a domain-invariant common representation space.", + "bbox": [ + 75, + 750, + 468, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given one sample $s$ from source domain and another sample $t$ from target domain, DDG is proposed to dynamically generate soft samples $s^{\\prime}$ and $t^\\prime$ to smoothly bridge the", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "domain gap. Specifically, $s'$ represents a variety of $s$ which maintains information of $s$ and mix source-relevant commonality from $t'$ . Similarly, $t'$ is a variety of $t$ . As shown in Fig.3, we capture the commonality between $s$ and $t$ with the help of the popular and effective self-attention mechanism used in Transformers [29]. We construct a sequence in which the length is 2 by stacking $s$ and $t$ . Denote query $Q_{s}$ and key $K_{s}$ is projected from the source embedding $s$ respectively, while query $Q_{t}$ and key $K_{t}$ are from the target embedding $t$ , respectively. With dot-product between queries and keys, the source-target attention matrix $M_{st}$ is obtained, which represents the relations between source and target. Afterwards, we design a gating to highlight the commonality between $s$ and $t$ , and fuse the commonality to generate $s'$ and $t'$ respectively. In detail, to generate $s'$ , we calculate the feature-relevant distribution $p$ by summing $M_{st}$ along the dimension of the query. Larger value in $p$ means related element in $t$ is more-relevant to $s$ , which should be chosen for commonality. As a result, the commonality $c$ is obtained by $c = t \\odot \\text{Norm}(p)$ , where $\\text{Norm}(\\cdot)$ means the normalized operator and $\\odot$ is the element-wise multiplication. Finally, $s'$ is obtained by fusing commonality $c$ and information from $s$ .", + "bbox": [ + 496, + 90, + 890, + 436 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ns ^ {\\prime} = \\lambda s + (1 - \\lambda) c \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 627, + 445, + 890, + 462 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, we use multi-head attention to generate various new domains to enrich the sample space. Samples from new domains and raw source and target domains are fed to domain discriminator, by which the domain discriminator is guided by the hard label information and well-designed soft domains. Each sample from these soft domains explores the intrinsic structure of data distribution from raw domains and enriches feature patterns by the interaction of two domains.", + "bbox": [ + 496, + 469, + 890, + 589 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Tasks", + "text_level": 1, + "bbox": [ + 500, + 597, + 578, + 612 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our model consists of three tasks: category classification, domain adversarial learning, and domain discrepancy elimination. First, we perform $K$ -way object classification on the source domain. Second, a domain discriminator is introduced to judge samples from source, target or the created soft domains. Finally, we try to learn transferable features by minimizing domain discrepancy. In summary, we have:", + "bbox": [ + 496, + 619, + 890, + 727 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL = L _ {c l s} + \\alpha L _ {a d v} + \\beta L _ {d i s} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 734, + 890, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $L_{cls}$ is the category classification loss, $L_{adv}$ is the adversarial loss, and $L_{dis}$ is the discrepancy loss. $\\alpha$ and $\\beta$ are weights to balance these three losses. Cross entropy loss is used as classification and adversarial loss, while MK-MMD [5] is employed as discrepancy loss.", + "bbox": [ + 496, + 757, + 890, + 834 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 844, + 632, + 861 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We conduct experiments on two totally different tasks to demonstrate the efficacy of our model. One task is popu", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "3554", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/e9cb6c510b2a1e1bfce9d9241e0a8719a3ec39ab3d401fbf0c570ad2881a594a.jpg", + "table_caption": [ + "Table 1. Multimodal sentiment analysis results on CMU-MOSI and CMU-MOSEI. †: results come from [7]; ‡: results come from [36]; ◇: results come from [6]; ↓: the lower the better." + ], + "table_footnote": [], + "table_body": "
MethodsCMU-MOSEI → CMU-MOSICMU-MOSI → CMU-MOSEI
MAE ↓CorrAcc-7Acc-2F1MAE ↓CorrAcc-7Acc-2F1
Direct Transfer0.7940.76439.579.7/81.579.5/81.40.6210.68551.379.54/82.1480.84/81.33
Supervised
TFN [37] †0.9010.69834.9-/80.8-/80.70.5930.70050.2-/82.5-/82.1
ICCN [22] †0.8620.71439.0-/83.0-/83.00.5650.71351.6-/84.2-/84.2
MISA [7] ‡0.8040.764-80.79/82.1080.77/82.030.5680.724-82.59/84.2382.67/83.97
MAG-BERT [20] ‡0.7310.789-82.50/84.3082.60/84.300.5390.753-83.80/85.2083.70/85.10
Self-MM [36] ‡0.7130.798-84.00/85.9884.42/85.950.5300.765-82.81/85.1782.53/85.30
MMIM [6] ◇0.7000.80046.6584.14/86.0684.00/85.980.5260.77254.2482.24/85.9782.66/85.94
UDA
DAN [12]0.7770.77439.7980.03/81.7179.74/81.490.6140.69351.680.24/81.3281.36/82.47
ADDA [27]0.7840.77340.1480.12/82.2680.13/82.320.6360.70751.480.47/81.5981.53/82.76
MM-SADA [15]0.7870.76940.5280.9/82.7780.68/82.630.6670.68452.180.32/81.4481.26/81.95
MDMN [44]0.7780.77439.6581.92/82.0181.97/82.110.6020.71252.882.24/82.3882.95/83.26
OSAN(TAL + Mixup)0.7530.78242.6482.44/83.3282.14/83.210.5420.75753.1482.76/82.8883.13/83.96
OSAN(TAL + DDG)0.7130.80146.3883.12/84.5883.02/84.510.5320.76853.8483.41/84.3683.31/84.47
", + "bbox": [ + 80, + 128, + 888, + 400 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "lar multimodal sentiment analysis, the other is a new task called video text classification.", + "bbox": [ + 75, + 419, + 468, + 450 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Multimodal Sentiment Analysis", + "text_level": 1, + "bbox": [ + 76, + 460, + 356, + 477 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Multimodal sentiment analysis is a new dimension of traditional text-based sentiment analysis, which goes beyond the analysis of texts, and includes other modalities data. This task processes data from multiple sources, such as acoustic, visual, and textual information to understand various human emotions.", + "bbox": [ + 75, + 483, + 470, + 575 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1.1 Experimental Setup", + "text_level": 1, + "bbox": [ + 76, + 594, + 271, + 611 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets and Model Structure We use two classical datasets, called CMU-MOSEI [39] and CMU-MOSI [38], for the multimodal sentiment analysis task.", + "bbox": [ + 75, + 618, + 468, + 664 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Metrics We use the metrics that have been presented in [6]: mean absolute error (MAE), which is the average mean absolute difference value between predicted values and truth values. Pearson correlation (Corr) that measures the degree of prediction skew, seven-class classification accuracy (Acc-7) indicating the proportion of predictions that correctly fall into the same interval of seven intervals between -3 and +3 as the corresponding truths, binary classification accuracy (Acc-2) and F1 score for positive/negative and non-negative/negative classification results.", + "bbox": [ + 75, + 683, + 468, + 835 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines To inspect the relative performance of our model, we compare our model with various baselines in three aspects: Direct Transfer, UDA and Supervised", + "bbox": [ + 75, + 854, + 470, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "methods. Direct transfer means that we train a model using source dataset and then predict test samples in target dataset. For UDA, we use model structures of discrepancy-based model DAN [12], adversarial-based model ADDA [27], and some typical studies focused on multimodal area such as MM-SADA [14] and MDMN [44] where multimodal alignment and domain adaptation are performed in two separate stages. Besides direct transfer and UDA methods, we also bring supervised methods for evaluation. For supervised methods, we consider the pure learning-based model TFN [37], feature space manipulation like ICCN [22], MISA [7], more recent and competitive baselines including MAG-BERT [20], Self-MM [36] and MMIM [6].", + "bbox": [ + 496, + 419, + 893, + 618 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1.2 Experimental Results", + "text_level": 1, + "bbox": [ + 498, + 638, + 705, + 656 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Overall The overall results are shown in Table 1. In detail, we find that OSAN achieves the best results among all UDA methods on all metrics with large margin. Surprisingly, OSAN even achieves better performance than many supervised methods, such as TFN, MISA and MAG-BERT.", + "bbox": [ + 496, + 664, + 893, + 742 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Ablation Study To verify the contributions of each component of OSAN, we design a series of ablation experiments. First, we evaluate the effectiveness of the two proposed modules TAL and DDG by eliminating one or both of them from our model, as illustrated in Table 2. We notice a manifest performance degradation after removing TAL or DDG, and the results are even worse when both modules are removed. Furthermore, we conduct additional experiments to explore the contributions of the three losses. In Table 3,", + "bbox": [ + 496, + 763, + 893, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "3555", + "bbox": [ + 482, + 944, + 514, + 957 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "we find that only using classification loss $L_{cls}$ , our model can achieve competitive performance compared to traditional multimodal DA methods such as MDMN, demonstrating that both modality and domain are well aligned using TAL and DDG. Finally, we adjust some important hyperparameters of DDG and TAL to evaluate their importance. In terms of DDG, we adjust the number of created soft domains by changing multi-head $H$ . Furthermore, we adjust the fusion coefficient $\\lambda$ to control the fusion ratio between the source and target samples. In Table 4, we observe that the performance improves significantly as the number of domains increases. We also find that the fusion coefficient set to 0.85 can achieve the best result. In terms of TAL, we reduce multimodal features to different combinations of dimensions. We find that the combination of $384 * 2 * 2$ achieves the best performance, as shown in Table 5.", + "bbox": [ + 75, + 90, + 472, + 333 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/3545f081dee40d265105b137e1c7d7ace9feea4524a54a03854d63cacfc9fed9.jpg", + "table_caption": [ + "Table 2. Ablation study of the proposed modules on CMU-MOSI." + ], + "table_footnote": [], + "table_body": "
Modulew/o DDGw/o TALw/o Both
MAE ↓+0.039+0.051+0.074
Corr-0.023-0.021-0.032
Acc-7-3.55-4.65-5.86
Acc-2-1.03/-1.30-1.89/-1.72-2.22/-1.81
F1-1.15/-1.52-1.85/-1.61-2.34/-1.88
", + "bbox": [ + 78, + 371, + 460, + 470 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/46d0554b44a06aa4a64292332db292441ec0dd9dd0f5727fb9c75c65f08f8f31.jpg", + "table_caption": [ + "Table 3. Ablation study of three losses on CMU-MOSI." + ], + "table_footnote": [], + "table_body": "
LossLclsLcls + LadvLcls + Ldis
MAE ↓+0.018+0.009+0.016
Corr-0.016-0.012-0.014
Acc-7-1.07-0.82-0.67
Acc-2-0.65/-0.95-0.32/-0.28-0.67/-0.74
F1-0.48/-0.69-0.7/-0.49-0.31/-0.58
", + "bbox": [ + 78, + 518, + 464, + 619 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/4a1c4d7a031d0f50679247b48ca2c4b0221e5e332cd34c1eedc12aff50b8a23c.jpg", + "table_caption": [ + "Table 4. Ablation study of DDG hyperparameters on CMU-MOSI." + ], + "table_footnote": [], + "table_body": "
DDGλ = 0.85\nH = 64λ = 0.9\nH = 64λ = 0.85\nH = 32λ = 0.9\nH = 32
MAE ↓0.00+0.0040.001+0.006
Corr0.00-0.01-0.003-0.016
Acc-70.00-0.590-0.27-0.75
Acc-20.00/0.00-0.27/-1.14-0.16/-1.07-0.39/-1.29
F10.00/0.00-0.259/-0.15-0.14/-0.04-0.39/-0.30
", + "bbox": [ + 78, + 667, + 460, + 782 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Video Text Classification", + "text_level": 1, + "bbox": [ + 76, + 801, + 303, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Video text classification is a real industrial task for video understanding [9]. With massive videos generated everyday, video textual information extraction is an essential work in many applications. However, numerous useless texts, such as rolling texts and blurred background texts,", + "bbox": [ + 75, + 824, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/3d57b9497d9e1d85c6c0516786e6e5339a4a2025c77528b2bef30aff8c4185e1.jpg", + "table_caption": [ + "Table 5. Ablation study of TAL hyperparameters on CMU-MOSI." + ], + "table_footnote": [], + "table_body": "
TALt:384t:192t:192
v:2,a:2v:4,a:2v:2,a:4
MAE↓0.000+0.007+0.006
Corr0.00-0.009-0.006
Acc-70.00-1.17-0.91
Acc-20.00/0.00-0.77/-1.16-0.30/-0.91
F10.00/0.00-0.59/-2.02-0.13/-0.79
", + "bbox": [ + 501, + 114, + 887, + 229 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "have no good effect or even side effects on downstream tasks. To control and filter video content, we propose a new video text classification task to classify video texts into valuable categories.", + "bbox": [ + 496, + 250, + 893, + 310 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.1 Experimental Setup", + "text_level": 1, + "bbox": [ + 498, + 337, + 692, + 353 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets and Model Structure We construct a well-defined industrial-grade dataset, called Text-news, which is dedicated to promote video text extraction research for news applications. The extraction of text in videos usually involves two steps: (1) text recognition and (2) text classification. We prepare various videos from news programs for annotation and use a general Optical Character Recognition (OCR) engine to locate and recognize all the texts in the videos. Then, an annotation system is designed for the video text classification task to tag video texts. Three representative categories are defined which are Caption, Subtitle and Others. To verify the effectiveness and generality of our method, we use Text-news as source domain dataset and construct a dataset named Text-show with a few annotated samples as the target domain dataset. The difference between the source and target datasets is that the video program is totally different. The source dataset is constructed from news videos, while the target dataset is constructed from videos of variety shows. We use the standard Precision, Recall, and F1 score for evaluation.", + "bbox": [ + 496, + 364, + 890, + 667 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines To inspect the relative performance of our model OSAN, we compare it with various baselines in two aspects: Direct Transfer and UDA methods. Unlike multimodal sentiment analysis, we do not use Supervised methods for comparison because the target dataset does not have annotated training data.", + "bbox": [ + 496, + 694, + 890, + 787 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.2 Experimental Results", + "text_level": 1, + "bbox": [ + 498, + 813, + 705, + 828 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Overall The results are shown in Table 6. All the metrics of OSAN achieve the best results when compared to other UDA methods and direct transfer, leading to a splendid performance of video text classification task.", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "3556", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/929645d2f94acbcb8a3154a1132894c1c6feef1959d0fa7f848a13a337e93462.jpg", + "table_caption": [ + "Table 6. Video text classification results on Text-show." + ], + "table_footnote": [], + "table_body": "
MethodsText-news → Text-show
PrecisionRecallF1
Direct Transfer80.277.9879.08
UDA
DAN [12]87.4480.5883.87
ADDA [27]91.6683.6687.48
MM-SADA [15]94.0783.4988.46
MDMN [44]93.5483.6988.34
OSAN(TAL + Mixup)94.4284.7989.35
OSAN(TAL + DDG)95.0386.4490.53
", + "bbox": [ + 84, + 114, + 454, + 281 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/85b98dbc349d4579515fa84be50bde415d5431e8c5beecde31e4ad5a03a6ddf4.jpg", + "table_caption": [ + "Table 7. Ablation study of the proposed modules on Text-show." + ], + "table_footnote": [], + "table_body": "
Modulew/o DDGw/o TALw/o Both
Precision-0.90-0.51-0.96
Recall-1.24-2.37-2.95
F1-1.09-1.55-2.07
", + "bbox": [ + 86, + 315, + 455, + 383 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6324a2056683b0ab713b88431a8f6ad1619d2eea3a909e012f782bbeb55b7bc6.jpg", + "table_caption": [ + "Table 8. Ablation study of three losses on Text-show." + ], + "table_footnote": [], + "table_body": "
LossLclsLcls + LadvLcls + Ldis
Precision-0.35-0.17-0.45
Recall-0.42-0.24-0.14
F1-0.39-0.21-0.28
", + "bbox": [ + 86, + 425, + 457, + 494 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/249e7bef64259e425b618b0b48e73145d23afe445ad759fcd1f1bc6baff9363e.jpg", + "table_caption": [ + "Table 9. Ablation study of DDG hyperparameters on Text-show." + ], + "table_footnote": [], + "table_body": "
DDGλ = 0.9\nH = 128λ = 0.95\nH = 128λ = 0.9\nH = 64λ = 0.95\nH = 64
Precision0.00-0.16-0.75-0.67
Recall0.00-0.32-1.21-1.33
F10.00-0.25-1.01-1.03
", + "bbox": [ + 86, + 535, + 455, + 619 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a3ea129e6d16fb6ad69f3e7cd599baa941b39204441b1ee23f58b2bf7a0b5036.jpg", + "table_caption": [ + "Table 10. Ablation study of TAL hyperparameters on Text-show." + ], + "table_footnote": [], + "table_body": "
TALt:56,v:14t:14,v:56t:28,v:28
Precision0.000-0.96+0.24
Recall0.000+0.05-0.59
F10.000-0.41-0.22
", + "bbox": [ + 86, + 650, + 455, + 718 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation Study Similar to multimodal sentiment analysis, in order to show the benefits of TAL and DDG modules, we perform a series of ablation experiments. First, we eliminate one or both modules from our model. In Table 7, the overall F1 score decreases after removing single module, and the results are even worse when removing both modules, which demonstrates the efficacy of these two modules. Furthermore, we study the impact of three losses of category classification loss, adversarial loss, and discrepancy loss, as shown in Table 8. The results show that all three losses contribute to performance. Moreover, by removing", + "bbox": [ + 75, + 734, + 472, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "DA related losses such as adversarial loss and discrepancy loss, it performs slightly worse than keeping all three losses, but still better than traditional multimodal DA methods such as MM-SADA and MDMN. This observation indicates that different domains are well aligned by TAL and DDG. Finally, we tuned some hyperparameters of TAL and DDG to explore their impact. For the DDG hyperparameters, in Table 9, we observe that a smaller fusion coefficient $\\lambda$ and a larger multi-head $H$ can achieve better results. For the TAL hyperparameters, we reduce multimodal features to different dimensions. In Table 10, the combination $56*14$ for textual and visual features achieves the best result.", + "bbox": [ + 496, + 90, + 893, + 272 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/840d248501d4438e62a3cdfe44e12b849405d601bbd5f71333a9677d3dfa676e.jpg", + "image_caption": [ + "(a) Directly transfer" + ], + "image_footnote": [], + "bbox": [ + 521, + 285, + 625, + 364 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/18bcbe9e4f1318cf0419f661e21a0db4a5140ef1749f31a0910d4c1eaa256dc5.jpg", + "image_caption": [ + "(b) Our method w/o TAL&DDG" + ], + "image_footnote": [], + "bbox": [ + 648, + 285, + 751, + 364 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b97a35b19be5df092e1814c4f6975a49a3770ab886d41c439fb39c42280df987.jpg", + "image_caption": [ + "(c) Our method" + ], + "image_footnote": [], + "bbox": [ + 772, + 285, + 875, + 364 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c009a84b8e253a616ccc50dcd967bb72587c0390e6ce73e71bb50392b10bc3eb.jpg", + "image_caption": [ + "Figure 4. Visualization of the feature distribution of the target domain for multimodal sentiment analysis.", + "(a) Directly transfer", + "Figure 5. Visualization of the feature distribution of the target domain for video text classification." + ], + "image_footnote": [], + "bbox": [ + 516, + 434, + 620, + 515 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/212c66b99d36fc9d5a8926204770b18d8b453565ec8481f2c22ac38cd988c136.jpg", + "image_caption": [ + "(b) Our method w/o TAL&DDG" + ], + "image_footnote": [], + "bbox": [ + 643, + 434, + 751, + 515 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e8bdbbb9e2bbeb10ddf20ad13275b340e6af69e92d63c3c6d6e5f16a17b817b2.jpg", + "image_caption": [ + "(c) Our method" + ], + "image_footnote": [], + "bbox": [ + 769, + 434, + 875, + 515 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Discussion", + "text_level": 1, + "bbox": [ + 500, + 579, + 612, + 594 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Visual Analysis of Feature Distribution", + "text_level": 1, + "bbox": [ + 500, + 604, + 834, + 619 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We use t-SNE [28] to visualize the feature distribution of test data in target domain. Fig.5 illustrates the t-SNE visualization of the feature distribution in the video text classification task. We find that the features of different classes are separated more clearly, and the features in the same class are more centralized. A similar phenomenon is observed in Fig.4 of the feature distribution for the multimodal sentiment analysis task. This makes sense because TAL can explore the relationship between domains and modalities and guide them to utilize complementary information. Moreover, DDG can construct new domains that guarantee domain-invariance in a more continuous latent space.", + "bbox": [ + 496, + 627, + 890, + 809 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Distribution Discrepancy", + "text_level": 1, + "bbox": [ + 500, + 816, + 728, + 833 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To quantify the distribution discrepancy of source and target domains, we use the classical metric of symmetric Jensen-Shannon divergence (JSD) [2]. Fig.7 shows the feature distribution discrepancy based on two baselines and", + "bbox": [ + 496, + 839, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "3557", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7e10fa6a31a82bcf3938021c99413f1534ee29bed1161c81f69b2291b4701958.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DomainImageTextMDMN ResultImageTextOSAN(ours) Result
SourceBeyond that, I actually excelled academically from some of my decisions um i say go see it coz i i enjoyed it nowgt: 0 prediction: --gt: 2.0 prediction:-0.37Thank you for that Godelieve, we all enjoyed your comments! um i say go see it coz i i enjoyed it nowgt: 2 prediction: --gt: 2.0 prediction:1.85
Target
Sourcehey were all randomly chosen ad we're going to have my lovely assistant here choose from among these three people and um i liked you know catherine hardwicke is like you know kind of ridiculousgt: 0 prediction: --gt: -0.6 prediction:0.42This story demonstrates how ridiculous and petty fighting, or war... and um i liked you know catherine hardwicke is like you know kind of ridiculousgt: -2 prediction: --gt: -0.6 prediction: -0.56
Target
Sourcewell if they had kept their mouth shut, that movie would have been fabulous. and it was kind of interestinggt: -1 prediction: --gt: 1.2 prediction:-1.94Learn a language if you can, because that will make your life more interesting and it was kind of interestinggt: 2.0 prediction: --gt: 1.2 prediction:1.53
Target
", + "bbox": [ + 102, + 95, + 866, + 330 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "our model on two tasks. Two baselines are Direct Transfer and our model without the two proposed modules TAL and DDG. We observe that JSD of our model is smaller than JSD of both baselines. This phenomenon implies that our model can reduce the domain gap more effectively than the baselines.", + "bbox": [ + 75, + 405, + 470, + 494 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1bdd0ff247178dfbcc7e838ba9559a1f5ec8830e1e8aad738f256c93fe6aac28.jpg", + "image_caption": [ + "Figure 6. Some samples analysis of one baseline MDMN and our model OSAN for UDA task CMU-MOSEI $\\rightarrow$ CMU-MOSI. Columns show information of one representative frame extracted from video clip, text message and the prediction result. Rows give information of target samples and closest source samples.", + "Figure 7. Distribution discrepancy of different methods on two different tasks CMU-MOSEI $\\rightarrow$ CMU-MOSI and Text-news $\\rightarrow$ Text-show." + ], + "image_footnote": [], + "bbox": [ + 78, + 511, + 468, + 672 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Case Studies", + "text_level": 1, + "bbox": [ + 76, + 741, + 210, + 755 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To deeply explore the advantages of our model over other baselines, we give some predictions of a two-stage baseline MDMN and our model OSAN for the UDA task CMU-MOSEI $\\rightarrow$ CMU-MOSI, as shown in Fig.6. We select some typical samples from the target domain to explore how our model works. For each target sample, we find the closest sample from source domain by calculating cosine similarity between target sample and source sample. We observe that the closest source samples picked by our model OSAN are", + "bbox": [ + 75, + 763, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "more relevant to the target samples. Specifically, take the target sample and source sample in the first row for example, with our model, speakers in both the source frame and target frame are very happy. Moreover, there is one common keyword in what the speakers said: enjoyed. Combining these two modalities of image and text, our model correctly classifies this target sample as positive sentiment with a score of 1.85, which is very close to the ground truth score of 2. In contrast, source sample found by the MDMN is not relevant to the target sample. There is no similar semantic information in what they said. The expressions in the frames are dissimilar. The phenomenon in these cases implies that our model can learn a common latent space in which the source domain and target domain are well aligned.", + "bbox": [ + 496, + 405, + 893, + 632 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 645, + 617, + 660 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To address the problems of domain adaptation and modality alignment in multimodal domain adaptation, we propose a novel one-stage alignment network to unify these two problems in one stage. In our model, the TAL module is proposed to explore the relationship between domains and modalities, which can provide complementary information to each other. Moreover, the DDG module is designed to generate new soft samples that mix information from both the source and target domains, guiding our model in a more continuous space. Extensive experiments on two multimodal tasks with different extents of domain shift demonstrate the excellent performance of our model.", + "bbox": [ + 496, + 670, + 893, + 852 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "3558", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Shai Ben-David, John Blitzer, Koby Crammer, Alex Kulesza, Fernando Pereira, and Jennifer Wortman Vaughan. A theory of learning from different domains. Machine learning, 79(1):151-175, 2010. 2", + "[2] Bent Fuglede and Flemming Topsoe. Jensen-shannon divergence and hilbert space embedding. In International Symposium on Information Theory, 2004. ISIT 2004. Proceedings., page 31. IEEE, 2004. 7", + "[3] Yaroslav Ganin and Victor Lempitsky. Unsupervised domain adaptation by backpropagation. In International conference on machine learning, pages 1180-1189. PMLR, 2015. 1", + "[4] Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario Marchand, and Victor Lempitsky. Domain-adversarial training of neural networks. The journal of machine learning research, 17(1):2096-2030, 2016. 2", + "[5] Arthur Gretton, Dino Sejdinovic, Heiko Strathmann, Sivaraman Balakrishnan, Massimiliano Pontil, Kenji Fukumizu, and Bharath K Striperumbudur. Optimal kernel choice for large-scale two-sample tests. In Advances in neural information processing systems, pages 1205-1213. Citeseer, 2012. 3, 4", + "[6] Wei Han, Hui Chen, and Soujanya Poria. Improving multimodal fusion with hierarchical mutual information maximization for multimodal sentiment analysis, 2021. 5", + "[7] Devamanyu Hazarika, Roger Zimmermann, and Soujanya Poria. Misa: Modality-invariant and-specific representations for multimodal sentiment analysis. In Proceedings of the 28th ACM International Conference on Multimedia, pages 1122-1131, 2020. 5", + "[8] Donghyun Kim, Yi-Hsuan Tsai, Bingbing Zhuang, Xiang Yu, Stan Sclaroff, Kate Saenko, and Manmohan Chandraker. Learning cross-modal contrastive features for video domain adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13618-13627, 2021. 1", + "[9] Ye Liu, Changchong Lu, Chen Lin, Di Yin, and Bo Ren. Contrastive graph multimodal model for text classification in videos. arXiv preprint arXiv:2206.02343, 2022. 6", + "[10] Ye Liu, Lingfeng Qiao, Di Yin, Zhuoxuan Jiang, Xinghua Jiang, Deqiang Jiang, and Bo Ren. Os-msl: One stage multimodal sequential link framework for scene segmentation and classification. In Proceedings of the 30th ACM International Conference on Multimedia, pages 6269-6277, 2022. 1", + "[11] Zhun Liu, Ying Shen, Varun Bharadhwaj Lakshminarasimhan, Paul Pu Liang, Amir Zadeh, and Louis-Philippe Morency. Efficient low-rank multimodal fusion with modality-specific factors. arXiv preprint arXiv:1806.00064, 2018.3", + "[12] Mingsheng Long, Yue Cao, Jianmin Wang, and Michael Jordan. Learning transferable features with deep adaptation networks. In International conference on machine learning, pages 97-105. PMLR, 2015. 5, 7", + "[13] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representational" + ], + "bbox": [ + 78, + 114, + 470, + 902 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tations for vision-and-language tasks. arXiv preprint arXiv:1908.02265, 2019. 1", + "[14] Jonathan Munro and Dima Damen. Multi-modal domain adaptation for fine-grained action recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 122-132, 2020. 1, 2, 5", + "[15] Jonathan Munro and Dima Damen. Multi-modal domain adaptation for fine-grained action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 5, 7", + "[16] Sinno Jialin Pan, Ivor W Tsang, James T Kwok, and Qiang Yang. Domain adaptation via transfer component analysis. IEEE transactions on neural networks, 22(2):199-210, 2010. 2", + "[17] Fan Qi, Xiaoshan Yang, and Changsheng Xu. A unified framework for multimodal domain adaptation. In Proceedings of the 26th ACM International Conference on Multimedia, MM '18, page 429-437, New York, NY, USA, 2018. Association for Computing Machinery. 1, 2", + "[18] Lingfeng Qiao, Chen Wu, Ye Liu, Haoyuan Peng, Di Yin, and Bo Ren. Grafting pre-trained models for multimodal headline generation. arXiv preprint arXiv:2211.07210, 2022. 1", + "[19] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 1", + "[20] Wasifur Rahman, Md Kamrul Hasan, Sangwu Lee, Amir Zadeh, Chengfeng Mao, Louis-Philippe Morency, and Ehsan Hoque. Integrating multimodal information in large pretrained transformers. In Proceedings of the conference. Association for Computational Linguistics. Meeting, volume 2020, page 2359. NIH Public Access, 2020. 5", + "[21] Baochen Sun and Kate Saenko. Deep coral: Correlation alignment for deep domain adaptation. In European conference on computer vision, pages 443-450. Springer, 2016. 3", + "[22] Zhongkai Sun, Prathusha Sarma, William Sethares, and Yingyu Liang. Learning relationships between text, audio, and video via deep canonical correlation for multimodal language analysis. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 8992-8999, 2020. 5", + "[23] Hui Tang and Kui Jia. Discriminative adversarial domain adaptation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 5940-5947, 2020. 2", + "[24] Dacheng Tao, Xuelong Li, Xindong Wu, and Steve Maybank. Tensor rank one discriminant analysis-a convergent method for discriminative multilinear subspace selection. Neurocomputing, 71(10):1866-1882, 2008. 3", + "[25] Dacheng Tao, Xuelong Li, Xindong Wu, and Stephen J Maybank. General tensor discriminant analysis and gabor features for gait recognition. IEEE transactions on pattern analysis and machine intelligence, 29(10):1700-1715, 2007. 4", + "[26] Wilhelm Tranheden, Viktor Olsson, Juliano Pinto, and Lennart Svensson. Dacs: Domain adaptation via cross-domain mixed sampling. In Proceedings of the IEEE/CVF" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "3559", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Winter Conference on Applications of Computer Vision, pages 1379-1389, 2021. 4", + "[27] Eric Tzeng, Judy Hoffman, Kate Saenko, and Trevor Darrell. Adversarial discriminative domain adaptation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7167-7176, 2017. 4, 5, 7", + "[28] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9(11), 2008. 7", + "[29] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Ilia Polosukhin. Attention is all you need. arXiv preprint arXiv:1706.03762, 2017. 4", + "[30] Wei Wang, Haojie Li, Zhengming Ding, Feiping Nie, Junyang Chen, Xiao Dong, and Zhihui Wang. Rethinking maximum mean discrepancy for visual domain adaptation. IEEE Transactions on Neural Networks and Learning Systems, 2021. 2", + "[31] Garrett Wilson and Diane J Cook. A survey of unsupervised deep domain adaptation. ACM Transactions on Intelligent Systems and Technology (TIST), 11(5):1-46, 2020. 1", + "[32] Xinyi Wu, Zhenyao Wu, Hao Guo, Lili Ju, and Song Wang. Darnet: A one-stage domain adaptation network for unsupervised nighttime semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15769-15778, 2021. 2", + "[33] Minghao Xu, Jian Zhang, Bingbing Ni, Teng Li, Chengjie Wang, Qi Tian, and Wenjun Zhang. Adversarial domain adaptation with domain mixup. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 6502-6509, 2020. 4", + "[34] Yiming Xu, Lin Chen, Zhongwei Cheng, Lixin Duan, and Jiebo Luo. Open-ended visual question answering by multimodal domain adaptation. arXiv preprint arXiv:1911.04058, 2019. 1, 2", + "[35] Shen Yan, Huan Song, Nanxiang Li, Lincan Zou, and Liu Ren. Improve unsupervised domain adaptation with mixup training. arXiv preprint arXiv:2001.00677, 2020. 4", + "[36] Wenmeng Yu, Hua Xu, Ziqi Yuan, and Jiele Wu. Learning modality-specific representations with self-supervised multi-task learning for multimodal sentiment analysis. arXiv preprint arXiv:2102.04830, 2021. 5", + "[37] Amir Zadeh, Minghai Chen, Soujanya Poria, Erik Cambria, and Louis-Philippe Morency. Tensor fusion network for multimodal sentiment analysis. arXiv preprint arXiv:1707.07250, 2017. 3, 5", + "[38] Amir Zadeh, Rowan Zellers, Eli Pincus, and Louis-Philippe Morency. Mosi: multimodal corpus of sentiment intensity and subjectivity analysis in online opinion videos. arXiv preprint arXiv:1606.06259, 2016. 5", + "[39] AmirAli Bagher Zadeh, Paul Pu Liang, Soujanya Poria, Erik Cambria, and Louis-Philippe Morency. Multimodal language analysis in the wild: Cmu-mosei dataset and interpretable dynamic fusion graph. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2236–2246, 2018. 5" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412, 2017. 4", + "[41] Tong Zhang, Di Wang, Huanhuan Chen, Zhiwei Zeng, Wei Guo, Chunyan Miao, and Lizhen Cui. Bdann: Bert-based domain adaptation neural network for multi-modal fake news detection. In 2020 International Joint Conference on Neural Networks (IJCNN), pages 1-8, 2020. 1, 2", + "[42] Weichen Zhang, Dong Xu, Jing Zhang, and Wanli Ouyang. Progressive modality cooperation for multi-modality domain adaptation. IEEE Transactions on Image Processing, 30:3293-3306, 2021. 2", + "[43] Yiyi Zhang, Li Niu, Ziqi Pan, Meichao Luo, Jianfu Zhang, Dawei Cheng, and Liqing Zhang. Exploiting motion information from unlabeled videos for static image action recognition. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 12918-12925, 2020. 1", + "[44] Honghao Zhou, Tinghuai Ma, Huan Rong, Yurong Qian, Yuan Tian, and Najla Al-Nabhan. Mdmn: Multi-task and domain adaptation based multi-modal network for early rumor detection. Expert Systems with Applications, 195:116517, 2022. 1, 2, 5, 7", + "[45] Junbao Zhuo, Shuhui Wang, Weigang Zhang, and Qingming Huang. Deep unsupervised convolutional domain adaptation. In Proceedings of the 25th ACM international conference on Multimedia, pages 261-269, 2017. 2, 3" + ], + "bbox": [ + 501, + 92, + 890, + 458 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "3560", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/OSAN_ A One-Stage Alignment Network To Unify Multimodal Alignment and Unsupervised Domain Adaptation/4429e1d5-6fb8-40a5-9804-367d3c0ae901_model.json b/2023/OSAN_ A One-Stage Alignment Network To Unify Multimodal Alignment and Unsupervised Domain Adaptation/4429e1d5-6fb8-40a5-9804-367d3c0ae901_model.json new file mode 100644 index 0000000000000000000000000000000000000000..dbdff6f62e32cc449fd6ddb956538ef6bfdfdd43 --- /dev/null +++ b/2023/OSAN_ A One-Stage Alignment Network To Unify Multimodal Alignment and Unsupervised Domain Adaptation/4429e1d5-6fb8-40a5-9804-367d3c0ae901_model.json @@ -0,0 +1,2145 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.095, + 0.13, + 0.877, + 0.177 + ], + "angle": 0, + "content": "OSAN: A One-Stage Alignment Network to Unify Multimodal Alignment and Unsupervised Domain Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.156, + 0.204, + 0.816, + 0.238 + ], + "angle": 0, + "content": "Ye Liu, Lingfeng Qiao, Changchong Lu, Di Yin, Chen Lin, Haoyuan Peng, Bo Ren \nTencent Youtu Lab" + }, + { + "type": "text", + "bbox": [ + 0.154, + 0.242, + 0.807, + 0.258 + ], + "angle": 0, + "content": "{rafelliu, leafqiao, chonglu, endymecyyin, iwlliamlin, haoyuanpeng, timren}@tencent.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.292, + 0.314, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.326, + 0.474, + 0.691 + ], + "angle": 0, + "content": "Extending from unimodal to multimodal is a critical challenge for unsupervised domain adaptation (UDA). Two major problems emerge in unsupervised multimodal domain adaptation: domain adaptation and modality alignment. An intuitive way to handle these two problems is to fulfill these tasks in two separate stages: aligning modalities followed by domain adaptation, or vice versa. However, domains and modalities are not associated in most existing two-stage studies, and the relationship between them is not leveraged which can provide complementary information to each other. In this paper, we unify these two stages into one to align domains and modalities simultaneously. In our model, a tensor-based alignment module (TAL) is presented to explore the relationship between domains and modalities. By this means, domains and modalities can interact sufficiently and guide them to utilize complementary information for better results. Furthermore, to establish a bridge between domains, a dynamic domain generator (DDG) module is proposed to build transitional samples by mixing the shared information of two domains in a self-supervised manner, which helps our model learn a domain-invariant common representation space. Extensive experiments prove that our method can achieve superior performance in two real-world applications. The code will be publicly available." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.723, + 0.21, + 0.739 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.903 + ], + "angle": 0, + "content": "With explosively emerging multimedia data on the Internet, the field of multimodal analysis achieves more and more attention [10, 13, 18, 19, 43]. Compared to extensive unimodal models in NLP and CV, learning adequate knowledge from multimodal signals is still preliminary but very important. Abundant data plays a key role in different scenarios of multimodal analysis, such as pre-training or downstream multimedia tasks. However, it is prohibitively expensive and time-consuming to obtain large amounts of labeled data. To eliminate this issue, domain adaptation (DA)" + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.294, + 0.891, + 0.419 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.558, + 0.43, + 0.833, + 0.445 + ], + "angle": 0, + "content": "Figure 1. Conception of our one-stage model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.462, + 0.893, + 0.629 + ], + "angle": 0, + "content": "is raised to learn a model from a labeled dataset (source domain) that can be generalized to other related tasks without sufficient labeled data (target domain) [3]. Classical domain adaptation can be classified into different categories: unsupervised domain adaptation (UDA), fully supervised domain adaptation, and semi-supervised domain adaptation [31]. In this paper, we focus on UDA where no samples in target domain are annotated. With this technique, it is not necessary to prepare a customized training dataset for a specific task, but it can perform the task effectively and efficiently." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.63, + 0.895, + 0.903 + ], + "angle": 0, + "content": "There are two challenges when applying domain adaptation to multimodal scenarios [17]: (1) how to align the source and target domains and remit domain discrepancy, and (2) how to align multiple modalities and leverage multimodal information. Most existing works address these two problems in two consecutive stages: multimodal alignment followed by domain adaptation [34, 41], or vice versa [14, 44]. However, they solve these two issues separately without considering their relationship: domain and modality can be treated as two views to portray the intrinsic characteristic of multimodal data [8], and the hidden underlying relationship in these two views can provide complementary information to each other. Unimodal domain adaptation methods can not work well in multimodal tasks due to the inability to preserve the relations between modalities at the same time. Through our experiments and analysis, we observe that the two-stage model could not achieve ideal performance. Fig.2 shows the learning curve of two-stage" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.958 + ], + "angle": 0, + "content": "3551" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.242 + ], + "angle": 0, + "content": "model during training phase by 800 iterations for the task of multimodal sentiment analysis. It can be found that the learning curve of two-stage model is oscillating and converges slowly, which indicates that two-stage model is probably not a superior solution. To handle these challenges, the objective of multimodal domain adaptation can be defined as: (1) Exploring the relationship between domains and modalities; (2) Finding a common domain-invariant cross-modal representation space to align domains and modalities simultaneously." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.243, + 0.473, + 0.455 + ], + "angle": 0, + "content": "Therefore, in this paper, we design a One-Stage Alignment Network (OSAN) to unify multimodal alignment and domain adaptation in one stage. Fig.1 shows the conception of our one-stage model. Our method benefits from: (1) The modality and domain are associated and interacted to capture the relationship between domains and modalities, which can provide rich complementary information to each other. (2) Multimodal alignment and domain adaptation are unified in one stage, which allows our model to perform domain adaptation and leverage multimodal information at the same time. In Fig.2, we observe that the learning curve of our method is relatively stable and converges better, which indicates that exploring the relation between modality and domain contributes to our task." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.455, + 0.399, + 0.469 + ], + "angle": 0, + "content": "In summary, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.469, + 0.47, + 0.545 + ], + "angle": 0, + "content": "(1) To capture the relationship between domain and modality, we propose a one-stage alignment network, called OSAN, to associate domain and modality. In this way, a joint domain-invariant and cross-modal representation space is learned in one stage." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.545, + 0.47, + 0.589 + ], + "angle": 0, + "content": "(2) We design a TAL module to bring sufficient interactions between domains and modalities and guide them to utilize complementary information for each other." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.59, + 0.47, + 0.651 + ], + "angle": 0, + "content": "(3) To effectively bridge distinct domains, a DDG module is developed to dynamically construct multiple new domains by combining knowledge of source and target domains and exploring intrinsic structure of data distribution." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.651, + 0.47, + 0.696 + ], + "angle": 0, + "content": "(4) Extensive experiments on two totally different tasks demonstrate the effectiveness of our method compared to the supervised and strongly UDA methods." + }, + { + "type": "list", + "bbox": [ + 0.076, + 0.469, + 0.47, + 0.696 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.71, + 0.22, + 0.725 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Unimodal Domain Adaption The main approaches for domain adaptation or domain alignment can be categorized as discrepancy-based and adversarial-based methods [1, 23, 30]. Discrepancy-based methods design statistics, such as maximum mean discrepancy (MMD) [16], correlation alignment (CORAL) [45] and etc., to measure the difference between two domain distributions. For adversarial-based methods, Ganin et al. [4] used the concept of generative adversarial networks (GAN) to obtain domain-invariant characteristics. Wu et al. [32] proposed a one-stage adaptation framework for nighttime semantic segmentation to per" + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.094, + 0.883, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.292, + 0.892, + 0.322 + ], + "angle": 0, + "content": "Figure 2. The learning curves of our one-stage method and traditional two-stage method in multimodal sentiment analysis task." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.346, + 0.894, + 0.424 + ], + "angle": 0, + "content": "form an end-to-end learning where a separate common preprocessing step was integrated, which is different from our motivation. However, these studies only focus on the elimination of domain shift in unimodal scenarios, which may lose effect when applied in multimodal area." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.439, + 0.895, + 0.743 + ], + "angle": 0, + "content": "Multimodal Domain Adaption Qi et al. [17] first analyzed the multimodal domain adaptation and raised two problems, i.e., domain adaptation and modality alignment. To alleviate them, hybrid domain constraints and attention-based modality fusion module were introduced to learn domain-invariant fused features. More recently, some studies [14, 34, 41, 42, 44] related to this area have tried to address the multimodal problem by extending the unimodal domain adaptation methods. Xu et al. [34] proposed a supervised multimodal domain adaptation method for VQA to learn joint feature embeddings across different domains and modalities. Munro et al. [14] exploited the correspondence of modalities as a self-supervised alignment approach for UDA in addition to adversarial alignment. Furthermore, Zhou et al. [44] proposed MDMN for early rumor detection, which can combine textual and visual information with two heterogeneous feature extractors. However, all of these studies solved the two problems separately in two stages. Moreover, the relationship between modality and domain is missing, thus they cannot achieve ideal performance." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.755, + 0.673, + 0.772 + ], + "angle": 0, + "content": "3. Proposed Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Fig.3 presents an overview of our method that contains four parts: multimodal feature extraction, tensor-based alignment, dynamic domain generator and task-specific heads. First, encoders, as multimodal feature extractor, map source and target multimodal data to different latent spaces. Then, by means of an effective and efficient way of tensor representation, we propose a TAL module to acquire relation information through consecutive interactions between" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3552" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.097, + 0.878, + 0.433 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.454, + 0.893, + 0.484 + ], + "angle": 0, + "content": "Figure 3. Overview of our method OSAN. We use circle, square and triangle to represent three modalities of text, vision and audio. In the figure, \\(\\times\\) means mode- \\(d\\) product of a tensor and a matrix, \\(\\odot\\) represents element-wise product, and \\(\\ast\\) means tensor factorization." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.497, + 0.471, + 0.588 + ], + "angle": 0, + "content": "modalities and domains. Thus, the modality and domain are aligned simultaneously in one stage. After that, we develop a DDG module to mix information from different domains to create new transitional domains. Finally, our network is split into three branches: category classification, domain adversarial learning and discrepancy elimination." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.603, + 0.303, + 0.619 + ], + "angle": 0, + "content": "3.1. Tensor-based Alignment" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.628, + 0.47, + 0.809 + ], + "angle": 0, + "content": "We develop a TAL module to simultaneously address the challenges of multimodal alignment and domain adaptation. We first resort to tensor representation, an effective and natural representation of multimodal data, to develop models capturing inter-modality dynamics [11, 37]. Then, we utilize statistical measurements to establish an interaction between domains and modalities to explore the relation information. In this way, we associate domains and modalities so that they can communicate and align together. Specifically, we perform tensor factorization to fuse multimodal information and maximize statistical measurements to simultaneously eliminate domain gaps." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.808, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Specifically, we use \\(\\underline{\\mathbf{X}}^{I_1\\times \\dots \\times I_N\\times M}\\) and \\(\\underline{\\mathbf{Y}}^{J_1\\times \\dots \\times J_N\\times M}\\) to represent source and target dataset with \\(N\\) modalities and \\(M\\) samples. For example, a video can be represented by a third-order tensor using a differentiable outer product between visual, textual, and acoustic representations [37]. For simplicity, we first make the training samples zero mean." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.497, + 0.893, + 0.68 + ], + "angle": 0, + "content": "To perform multimodal alignment, TAL aims to find pairs of linear transformations \\(\\mathbf{U}^{(n)}|_{n = 1}^{N}\\) and \\(\\mathbf{V}^{(n)}|_{n = 1}^{N}\\) for each modality of source and target domains to project samples of two sets into low dimensional subspaces. During this process, we establish an interaction between domain and modality by maximizing a statistical measurement of covariance given a normalized standard deviation. As reported in [45], we choose covariance instead of the first-order mean statistics used in MK-MMD [5, 21], as it is more powerful and achieves better performance. By this means, domains and modalities can be aligned simultaneously in one stage. Thus the objective function is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.512, + 0.688, + 0.828, + 0.751 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\left\\{\\mathbf {U} ^ {(n)} | _ {n = 1} ^ {N}, \\mathbf {V} ^ {(n)} | _ {n = 1} ^ {N} \\right\\} \\\\ = \\max \\llbracket (\\underline {{\\mathbf {X}}} \\prod_ {n = 1} ^ {N} \\times_ {n} \\mathbf {U} ^ {(n) ^ {T}}) \\otimes (\\underline {{\\mathbf {Y}}} \\prod_ {n = 1} ^ {N} \\times_ {n} \\mathbf {V} ^ {(n) ^ {T}}); \\rrbracket \\\\ \\end{array}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.51, + 0.753, + 0.892, + 0.832 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} s. t. \\left(\\underline {{\\mathbf {X}}} \\prod_ {n = 1} ^ {N} \\times_ {n} \\mathbf {U} ^ {(n) ^ {T}}\\right) _ {(N + 1)} ^ {T} \\left(\\underline {{\\mathbf {X}}} \\prod_ {n = 1} ^ {N} \\times_ {n} \\mathbf {U} ^ {(n) ^ {T}}\\right) _ {(N + 1)} = \\mathbf {I} \\tag {1} \\\\ \\left(\\underline {{\\mathbf {Y}}} \\prod_ {n = 1} ^ {N} \\times_ {n} \\mathbf {V} ^ {(n) ^ {T}}\\right) _ {(N + 1)} ^ {T} \\left(\\underline {{\\mathbf {Y}}} \\prod_ {n = 1} ^ {N} \\times_ {n} \\mathbf {V} ^ {(n) ^ {T}}\\right) _ {(N + 1)} = \\mathbf {I} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.84, + 0.893, + 0.902 + ], + "angle": 0, + "content": "The objective function above does not have a closed-form solution, so instead we derive a suboptimal solution following the principle of the alternating projection method [24], where the complicated optimization problem is re" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3553" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.15 + ], + "angle": 0, + "content": "duced into smaller conditional subproblems that can be solved by simple established methods. Therefore, the objective function is decomposed into \\(N\\) different subproblems:" + }, + { + "type": "equation", + "bbox": [ + 0.126, + 0.16, + 0.47, + 0.188 + ], + "angle": 0, + "content": "\\[\n\\left\\{\\mathbf {U} ^ {(n)}, \\mathbf {V} ^ {(n)} \\right\\} = \\max t r \\left\\{\\mathbf {U} ^ {(n) ^ {T}} \\mathbf {C} _ {x y} ^ {(n)} \\mathbf {V} ^ {(n)} \\right\\} \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.125, + 0.188, + 0.425, + 0.205 + ], + "angle": 0, + "content": "\\[\ns. t. \\mathbf {U} ^ {(n) ^ {T}} \\mathbf {C} _ {x x} ^ {(n)} \\mathbf {U} ^ {(n)} = \\mathbf {I}, \\mathbf {V} ^ {(n) ^ {T}} \\mathbf {C} _ {y y} ^ {(n)} \\mathbf {V} ^ {(n)} = \\mathbf {I}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.218, + 0.411, + 0.237 + ], + "angle": 0, + "content": "where, for simplification, we define \\(\\mathbf{C}_{xy}^{(n)}\\), \\(\\mathbf{C}_{xx}^{(n)}\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.247, + 0.47, + 0.278 + ], + "angle": 0, + "content": "\\[\n\\mathbf {C} _ {x y} ^ {(n)} = \\left(\\llbracket (\\underline {{\\mathbf {X}}} ^ {(\\bar {n})}) \\otimes (\\underline {{\\mathbf {Y}}} ^ {(\\bar {n})}); (\\bar {n}) (\\bar {n}) \\rrbracket\\right) \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.155, + 0.274, + 0.39, + 0.296 + ], + "angle": 0, + "content": "\\[\n\\mathbf {C} _ {x x} ^ {(n)} = \\left(\\llbracket (\\underline {{\\mathbf {X}}} ^ {(\\bar {n})}) \\otimes (\\underline {{\\mathbf {X}}} ^ {(\\bar {n})}); (\\bar {n}) (\\bar {n}) \\rrbracket\\right)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.307, + 0.469, + 0.384 + ], + "angle": 0, + "content": "where \\(([\\cdot \\otimes ;(\\bar{n})(\\bar{n})])\\) means tensor contraction on all indices except the \\(n\\)-th index on the tensor product [25]. Similarly, \\(\\mathbf{C}_{yy}^{(n)}\\) and \\(\\mathbf{C}_{yx}^{(n)}\\) is same as \\(\\mathbf{C}_{xx}^{(n)}\\) and \\(\\mathbf{C}_{xy}^{(n)}\\) in form. By Eq.2, an interaction between source and target domains for the \\(n\\)-th modality is established." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.395, + 0.332, + 0.41 + ], + "angle": 0, + "content": "3.2. Dynamic Domain Generator" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.418, + 0.469, + 0.552 + ], + "angle": 0, + "content": "In domain adaptation, adversarial learning is often used to align domains by developing a domain discriminator to judge samples from the source or target domains [27, 33]. By this means, it brings a two-class classification task with hard assignments of \\( \\mathbf{1} \\) or \\( \\mathbf{0} \\). They try to confuse the domain discriminator to learn a shared common representation space between the source and target domains. However, it usually achieves worse performance due to a huge variance between the two domains." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.554, + 0.469, + 0.749 + ], + "angle": 0, + "content": "To overcome this problem, several works have been proposed to produce some transitional domains based on Mixup [35, 40]. They mixed images from the source and target domains to generate more soft and mixed samples [26, 33]. Suppose there are one sample \\( s \\) from source domain and another sample \\( t \\) from target domain, \\( s \\) and \\( t \\) are linearly interpolated [40] to generate a new sample. This strategy can fully utilize the inter-domain information and improve the generalization ability of models. However, it has an obvious shortcoming: the generated sample fuses full information of two domains, which contains redundant information and speciality of domains that may induce undesirable oscillations on training the domain discriminator." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.469, + 0.856 + ], + "angle": 0, + "content": "To solve this issue, we propose a DDG module to establish a bridge between domains. Specifically, DDG explicitly captures commonality and abandons the specialty of domains. By highlighting this commonality, we make the domain discriminator focus on commonality rather than full information, which helps our model learn a domain-invariant common representation space." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Given one sample \\(s\\) from source domain and another sample \\(t\\) from target domain, DDG is proposed to dynamically generate soft samples \\(s^{\\prime}\\) and \\(t^\\prime\\) to smoothly bridge the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.438 + ], + "angle": 0, + "content": "domain gap. Specifically, \\( s' \\) represents a variety of \\( s \\) which maintains information of \\( s \\) and mix source-relevant commonality from \\( t' \\). Similarly, \\( t' \\) is a variety of \\( t \\). As shown in Fig.3, we capture the commonality between \\( s \\) and \\( t \\) with the help of the popular and effective self-attention mechanism used in Transformers [29]. We construct a sequence in which the length is 2 by stacking \\( s \\) and \\( t \\). Denote query \\( Q_{s} \\) and key \\( K_{s} \\) is projected from the source embedding \\( s \\) respectively, while query \\( Q_{t} \\) and key \\( K_{t} \\) are from the target embedding \\( t \\), respectively. With dot-product between queries and keys, the source-target attention matrix \\( M_{st} \\) is obtained, which represents the relations between source and target. Afterwards, we design a gating to highlight the commonality between \\( s \\) and \\( t \\), and fuse the commonality to generate \\( s' \\) and \\( t' \\) respectively. In detail, to generate \\( s' \\), we calculate the feature-relevant distribution \\( p \\) by summing \\( M_{st} \\) along the dimension of the query. Larger value in \\( p \\) means related element in \\( t \\) is more-relevant to \\( s \\), which should be chosen for commonality. As a result, the commonality \\( c \\) is obtained by \\( c = t \\odot \\text{Norm}(p) \\), where \\( \\text{Norm}(\\cdot) \\) means the normalized operator and \\( \\odot \\) is the element-wise multiplication. Finally, \\( s' \\) is obtained by fusing commonality \\( c \\) and information from \\( s \\)." + }, + { + "type": "equation", + "bbox": [ + 0.629, + 0.446, + 0.891, + 0.463 + ], + "angle": 0, + "content": "\\[\ns ^ {\\prime} = \\lambda s + (1 - \\lambda) c \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.47, + 0.892, + 0.59 + ], + "angle": 0, + "content": "Finally, we use multi-head attention to generate various new domains to enrich the sample space. Samples from new domains and raw source and target domains are fed to domain discriminator, by which the domain discriminator is guided by the hard label information and well-designed soft domains. Each sample from these soft domains explores the intrinsic structure of data distribution from raw domains and enriches feature patterns by the interaction of two domains." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.598, + 0.58, + 0.613 + ], + "angle": 0, + "content": "3.3. Tasks" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.621, + 0.892, + 0.728 + ], + "angle": 0, + "content": "Our model consists of three tasks: category classification, domain adversarial learning, and domain discrepancy elimination. First, we perform \\( K \\)-way object classification on the source domain. Second, a domain discriminator is introduced to judge samples from source, target or the created soft domains. Finally, we try to learn transferable features by minimizing domain discrepancy. In summary, we have:" + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.735, + 0.891, + 0.751 + ], + "angle": 0, + "content": "\\[\nL = L _ {c l s} + \\alpha L _ {a d v} + \\beta L _ {d i s} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.758, + 0.892, + 0.835 + ], + "angle": 0, + "content": "where \\( L_{cls} \\) is the category classification loss, \\( L_{adv} \\) is the adversarial loss, and \\( L_{dis} \\) is the discrepancy loss. \\( \\alpha \\) and \\( \\beta \\) are weights to balance these three losses. Cross entropy loss is used as classification and adversarial loss, while MK-MMD [5] is employed as discrepancy loss." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.845, + 0.633, + 0.862 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We conduct experiments on two totally different tasks to demonstrate the efficacy of our model. One task is popu" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3554" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.892, + 0.118 + ], + "angle": 0, + "content": "Table 1. Multimodal sentiment analysis results on CMU-MOSI and CMU-MOSEI. †: results come from [7]; ‡: results come from [36]; ◇: results come from [6]; ↓: the lower the better." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.13, + 0.89, + 0.401 + ], + "angle": 0, + "content": "
MethodsCMU-MOSEI → CMU-MOSICMU-MOSI → CMU-MOSEI
MAE ↓CorrAcc-7Acc-2F1MAE ↓CorrAcc-7Acc-2F1
Direct Transfer0.7940.76439.579.7/81.579.5/81.40.6210.68551.379.54/82.1480.84/81.33
Supervised
TFN [37] †0.9010.69834.9-/80.8-/80.70.5930.70050.2-/82.5-/82.1
ICCN [22] †0.8620.71439.0-/83.0-/83.00.5650.71351.6-/84.2-/84.2
MISA [7] ‡0.8040.764-80.79/82.1080.77/82.030.5680.724-82.59/84.2382.67/83.97
MAG-BERT [20] ‡0.7310.789-82.50/84.3082.60/84.300.5390.753-83.80/85.2083.70/85.10
Self-MM [36] ‡0.7130.798-84.00/85.9884.42/85.950.5300.765-82.81/85.1782.53/85.30
MMIM [6] ◇0.7000.80046.6584.14/86.0684.00/85.980.5260.77254.2482.24/85.9782.66/85.94
UDA
DAN [12]0.7770.77439.7980.03/81.7179.74/81.490.6140.69351.680.24/81.3281.36/82.47
ADDA [27]0.7840.77340.1480.12/82.2680.13/82.320.6360.70751.480.47/81.5981.53/82.76
MM-SADA [15]0.7870.76940.5280.9/82.7780.68/82.630.6670.68452.180.32/81.4481.26/81.95
MDMN [44]0.7780.77439.6581.92/82.0181.97/82.110.6020.71252.882.24/82.3882.95/83.26
OSAN(TAL + Mixup)0.7530.78242.6482.44/83.3282.14/83.210.5420.75753.1482.76/82.8883.13/83.96
OSAN(TAL + DDG)0.7130.80146.3883.12/84.5883.02/84.510.5320.76853.8483.41/84.3683.31/84.47
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.42, + 0.47, + 0.451 + ], + "angle": 0, + "content": "lar multimodal sentiment analysis, the other is a new task called video text classification." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.461, + 0.357, + 0.478 + ], + "angle": 0, + "content": "4.1. Multimodal Sentiment Analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.484, + 0.471, + 0.576 + ], + "angle": 0, + "content": "Multimodal sentiment analysis is a new dimension of traditional text-based sentiment analysis, which goes beyond the analysis of texts, and includes other modalities data. This task processes data from multiple sources, such as acoustic, visual, and textual information to understand various human emotions." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.595, + 0.272, + 0.612 + ], + "angle": 0, + "content": "4.1.1 Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.619, + 0.47, + 0.665 + ], + "angle": 0, + "content": "Datasets and Model Structure We use two classical datasets, called CMU-MOSEI [39] and CMU-MOSI [38], for the multimodal sentiment analysis task." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.684, + 0.47, + 0.837 + ], + "angle": 0, + "content": "Metrics We use the metrics that have been presented in [6]: mean absolute error (MAE), which is the average mean absolute difference value between predicted values and truth values. Pearson correlation (Corr) that measures the degree of prediction skew, seven-class classification accuracy (Acc-7) indicating the proportion of predictions that correctly fall into the same interval of seven intervals between -3 and +3 as the corresponding truths, binary classification accuracy (Acc-2) and F1 score for positive/negative and non-negative/negative classification results." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Baselines To inspect the relative performance of our model, we compare our model with various baselines in three aspects: Direct Transfer, UDA and Supervised" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.42, + 0.895, + 0.619 + ], + "angle": 0, + "content": "methods. Direct transfer means that we train a model using source dataset and then predict test samples in target dataset. For UDA, we use model structures of discrepancy-based model DAN [12], adversarial-based model ADDA [27], and some typical studies focused on multimodal area such as MM-SADA [14] and MDMN [44] where multimodal alignment and domain adaptation are performed in two separate stages. Besides direct transfer and UDA methods, we also bring supervised methods for evaluation. For supervised methods, we consider the pure learning-based model TFN [37], feature space manipulation like ICCN [22], MISA [7], more recent and competitive baselines including MAG-BERT [20], Self-MM [36] and MMIM [6]." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.64, + 0.707, + 0.657 + ], + "angle": 0, + "content": "4.1.2 Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.665, + 0.894, + 0.743 + ], + "angle": 0, + "content": "Overall The overall results are shown in Table 1. In detail, we find that OSAN achieves the best results among all UDA methods on all metrics with large margin. Surprisingly, OSAN even achieves better performance than many supervised methods, such as TFN, MISA and MAG-BERT." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Ablation Study To verify the contributions of each component of OSAN, we design a series of ablation experiments. First, we evaluate the effectiveness of the two proposed modules TAL and DDG by eliminating one or both of them from our model, as illustrated in Table 2. We notice a manifest performance degradation after removing TAL or DDG, and the results are even worse when both modules are removed. Furthermore, we conduct additional experiments to explore the contributions of the three losses. In Table 3," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.958 + ], + "angle": 0, + "content": "3555" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.334 + ], + "angle": 0, + "content": "we find that only using classification loss \\( L_{cls} \\), our model can achieve competitive performance compared to traditional multimodal DA methods such as MDMN, demonstrating that both modality and domain are well aligned using TAL and DDG. Finally, we adjust some important hyperparameters of DDG and TAL to evaluate their importance. In terms of DDG, we adjust the number of created soft domains by changing multi-head \\( H \\). Furthermore, we adjust the fusion coefficient \\( \\lambda \\) to control the fusion ratio between the source and target samples. In Table 4, we observe that the performance improves significantly as the number of domains increases. We also find that the fusion coefficient set to 0.85 can achieve the best result. In terms of TAL, we reduce multimodal features to different combinations of dimensions. We find that the combination of \\( 384 * 2 * 2 \\) achieves the best performance, as shown in Table 5." + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.346, + 0.468, + 0.361 + ], + "angle": 0, + "content": "Table 2. Ablation study of the proposed modules on CMU-MOSI." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.372, + 0.462, + 0.472 + ], + "angle": 0, + "content": "
Modulew/o DDGw/o TALw/o Both
MAE ↓+0.039+0.051+0.074
Corr-0.023-0.021-0.032
Acc-7-3.55-4.65-5.86
Acc-2-1.03/-1.30-1.89/-1.72-2.22/-1.81
F1-1.15/-1.52-1.85/-1.61-2.34/-1.88
" + }, + { + "type": "table_caption", + "bbox": [ + 0.107, + 0.494, + 0.438, + 0.508 + ], + "angle": 0, + "content": "Table 3. Ablation study of three losses on CMU-MOSI." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.52, + 0.465, + 0.62 + ], + "angle": 0, + "content": "
LossLclsLcls + LadvLcls + Ldis
MAE ↓+0.018+0.009+0.016
Corr-0.016-0.012-0.014
Acc-7-1.07-0.82-0.67
Acc-2-0.65/-0.95-0.32/-0.28-0.67/-0.74
F1-0.48/-0.69-0.7/-0.49-0.31/-0.58
" + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.643, + 0.468, + 0.658 + ], + "angle": 0, + "content": "Table 4. Ablation study of DDG hyperparameters on CMU-MOSI." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.669, + 0.462, + 0.784 + ], + "angle": 0, + "content": "
DDGλ = 0.85\nH = 64λ = 0.9\nH = 64λ = 0.85\nH = 32λ = 0.9\nH = 32
MAE ↓0.00+0.0040.001+0.006
Corr0.00-0.01-0.003-0.016
Acc-70.00-0.590-0.27-0.75
Acc-20.00/0.00-0.27/-1.14-0.16/-1.07-0.39/-1.29
F10.00/0.00-0.259/-0.15-0.14/-0.04-0.39/-0.30
" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.802, + 0.305, + 0.817 + ], + "angle": 0, + "content": "4.2. Video Text Classification" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Video text classification is a real industrial task for video understanding [9]. With massive videos generated everyday, video textual information extraction is an essential work in many applications. However, numerous useless texts, such as rolling texts and blurred background texts," + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.089, + 0.892, + 0.104 + ], + "angle": 0, + "content": "Table 5. Ablation study of TAL hyperparameters on CMU-MOSI." + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.116, + 0.888, + 0.23 + ], + "angle": 0, + "content": "
TALt:384t:192t:192
v:2,a:2v:4,a:2v:2,a:4
MAE↓0.000+0.007+0.006
Corr0.00-0.009-0.006
Acc-70.00-1.17-0.91
Acc-20.00/0.00-0.77/-1.16-0.30/-0.91
F10.00/0.00-0.59/-2.02-0.13/-0.79
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.25, + 0.894, + 0.311 + ], + "angle": 0, + "content": "have no good effect or even side effects on downstream tasks. To control and filter video content, we propose a new video text classification task to classify video texts into valuable categories." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.338, + 0.694, + 0.354 + ], + "angle": 0, + "content": "4.2.1 Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.365, + 0.892, + 0.668 + ], + "angle": 0, + "content": "Datasets and Model Structure We construct a well-defined industrial-grade dataset, called Text-news, which is dedicated to promote video text extraction research for news applications. The extraction of text in videos usually involves two steps: (1) text recognition and (2) text classification. We prepare various videos from news programs for annotation and use a general Optical Character Recognition (OCR) engine to locate and recognize all the texts in the videos. Then, an annotation system is designed for the video text classification task to tag video texts. Three representative categories are defined which are Caption, Subtitle and Others. To verify the effectiveness and generality of our method, we use Text-news as source domain dataset and construct a dataset named Text-show with a few annotated samples as the target domain dataset. The difference between the source and target datasets is that the video program is totally different. The source dataset is constructed from news videos, while the target dataset is constructed from videos of variety shows. We use the standard Precision, Recall, and F1 score for evaluation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.695, + 0.892, + 0.788 + ], + "angle": 0, + "content": "Baselines To inspect the relative performance of our model OSAN, we compare it with various baselines in two aspects: Direct Transfer and UDA methods. Unlike multimodal sentiment analysis, we do not use Supervised methods for comparison because the target dataset does not have annotated training data." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.814, + 0.706, + 0.829 + ], + "angle": 0, + "content": "4.2.2 Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Overall The results are shown in Table 6. All the metrics of OSAN achieve the best results when compared to other UDA methods and direct transfer, leading to a splendid performance of video text classification task." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3556" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.111, + 0.09, + 0.437, + 0.103 + ], + "angle": 0, + "content": "Table 6. Video text classification results on Text-show." + }, + { + "type": "table", + "bbox": [ + 0.086, + 0.116, + 0.455, + 0.282 + ], + "angle": 0, + "content": "
MethodsText-news → Text-show
PrecisionRecallF1
Direct Transfer80.277.9879.08
UDA
DAN [12]87.4480.5883.87
ADDA [27]91.6683.6687.48
MM-SADA [15]94.0783.4988.46
MDMN [44]93.5483.6988.34
OSAN(TAL + Mixup)94.4284.7989.35
OSAN(TAL + DDG)95.0386.4490.53
" + }, + { + "type": "table_caption", + "bbox": [ + 0.085, + 0.29, + 0.461, + 0.304 + ], + "angle": 0, + "content": "Table 7. Ablation study of the proposed modules on Text-show." + }, + { + "type": "table", + "bbox": [ + 0.088, + 0.316, + 0.456, + 0.385 + ], + "angle": 0, + "content": "
Modulew/o DDGw/o TALw/o Both
Precision-0.90-0.51-0.96
Recall-1.24-2.37-2.95
F1-1.09-1.55-2.07
" + }, + { + "type": "table_caption", + "bbox": [ + 0.115, + 0.4, + 0.432, + 0.414 + ], + "angle": 0, + "content": "Table 8. Ablation study of three losses on Text-show." + }, + { + "type": "table", + "bbox": [ + 0.088, + 0.426, + 0.458, + 0.495 + ], + "angle": 0, + "content": "
LossLclsLcls + LadvLcls + Ldis
Precision-0.35-0.17-0.45
Recall-0.42-0.24-0.14
F1-0.39-0.21-0.28
" + }, + { + "type": "table_caption", + "bbox": [ + 0.084, + 0.51, + 0.463, + 0.525 + ], + "angle": 0, + "content": "Table 9. Ablation study of DDG hyperparameters on Text-show." + }, + { + "type": "table", + "bbox": [ + 0.088, + 0.536, + 0.456, + 0.62 + ], + "angle": 0, + "content": "
DDGλ = 0.9\nH = 128λ = 0.95\nH = 128λ = 0.9\nH = 64λ = 0.95\nH = 64
Precision0.00-0.16-0.75-0.67
Recall0.00-0.32-1.21-1.33
F10.00-0.25-1.01-1.03
" + }, + { + "type": "table_caption", + "bbox": [ + 0.082, + 0.624, + 0.465, + 0.639 + ], + "angle": 0, + "content": "Table 10. Ablation study of TAL hyperparameters on Text-show." + }, + { + "type": "table", + "bbox": [ + 0.088, + 0.651, + 0.456, + 0.719 + ], + "angle": 0, + "content": "
TALt:56,v:14t:14,v:56t:28,v:28
Precision0.000-0.96+0.24
Recall0.000+0.05-0.59
F10.000-0.41-0.22
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Ablation Study Similar to multimodal sentiment analysis, in order to show the benefits of TAL and DDG modules, we perform a series of ablation experiments. First, we eliminate one or both modules from our model. In Table 7, the overall F1 score decreases after removing single module, and the results are even worse when removing both modules, which demonstrates the efficacy of these two modules. Furthermore, we study the impact of three losses of category classification loss, adversarial loss, and discrepancy loss, as shown in Table 8. The results show that all three losses contribute to performance. Moreover, by removing" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.273 + ], + "angle": 0, + "content": "DA related losses such as adversarial loss and discrepancy loss, it performs slightly worse than keeping all three losses, but still better than traditional multimodal DA methods such as MM-SADA and MDMN. This observation indicates that different domains are well aligned by TAL and DDG. Finally, we tuned some hyperparameters of TAL and DDG to explore their impact. For the DDG hyperparameters, in Table 9, we observe that a smaller fusion coefficient \\(\\lambda\\) and a larger multi-head \\(H\\) can achieve better results. For the TAL hyperparameters, we reduce multimodal features to different dimensions. In Table 10, the combination \\(56*14\\) for textual and visual features achieves the best result." + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.286, + 0.627, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.545, + 0.368, + 0.6, + 0.376 + ], + "angle": 0, + "content": "(a) Directly transfer" + }, + { + "type": "image", + "bbox": [ + 0.649, + 0.286, + 0.753, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.655, + 0.367, + 0.744, + 0.374 + ], + "angle": 0, + "content": "(b) Our method w/o TAL&DDG" + }, + { + "type": "image", + "bbox": [ + 0.773, + 0.286, + 0.876, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.799, + 0.368, + 0.844, + 0.375 + ], + "angle": 0, + "content": "(c) Our method" + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.387, + 0.892, + 0.415 + ], + "angle": 0, + "content": "Figure 4. Visualization of the feature distribution of the target domain for multimodal sentiment analysis." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.435, + 0.622, + 0.516 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.539, + 0.519, + 0.595, + 0.526 + ], + "angle": 0, + "content": "(a) Directly transfer" + }, + { + "type": "image", + "bbox": [ + 0.645, + 0.435, + 0.752, + 0.516 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.648, + 0.519, + 0.739, + 0.526 + ], + "angle": 0, + "content": "(b) Our method w/o TAL&DDG" + }, + { + "type": "image", + "bbox": [ + 0.771, + 0.435, + 0.877, + 0.516 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.799, + 0.519, + 0.844, + 0.526 + ], + "angle": 0, + "content": "(c) Our method" + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.54, + 0.892, + 0.567 + ], + "angle": 0, + "content": "Figure 5. Visualization of the feature distribution of the target domain for video text classification." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.58, + 0.614, + 0.595 + ], + "angle": 0, + "content": "5. Discussion" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.605, + 0.836, + 0.621 + ], + "angle": 0, + "content": "5.1. Visual Analysis of Feature Distribution" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.628, + 0.892, + 0.81 + ], + "angle": 0, + "content": "We use t-SNE [28] to visualize the feature distribution of test data in target domain. Fig.5 illustrates the t-SNE visualization of the feature distribution in the video text classification task. We find that the features of different classes are separated more clearly, and the features in the same class are more centralized. A similar phenomenon is observed in Fig.4 of the feature distribution for the multimodal sentiment analysis task. This makes sense because TAL can explore the relationship between domains and modalities and guide them to utilize complementary information. Moreover, DDG can construct new domains that guarantee domain-invariance in a more continuous latent space." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.818, + 0.729, + 0.834 + ], + "angle": 0, + "content": "5.2. Distribution Discrepancy" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.901 + ], + "angle": 0, + "content": "To quantify the distribution discrepancy of source and target domains, we use the classical metric of symmetric Jensen-Shannon divergence (JSD) [2]. Fig.7 shows the feature distribution discrepancy based on two baselines and" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3557" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.104, + 0.096, + 0.867, + 0.332 + ], + "angle": 0, + "content": "
DomainImageTextMDMN ResultImageTextOSAN(ours) Result
SourceBeyond that, I actually excelled academically from some of my decisions um i say go see it coz i i enjoyed it nowgt: 0 prediction: --gt: 2.0 prediction:-0.37Thank you for that Godelieve, we all enjoyed your comments! um i say go see it coz i i enjoyed it nowgt: 2 prediction: --gt: 2.0 prediction:1.85
Target
Sourcehey were all randomly chosen ad we're going to have my lovely assistant here choose from among these three people and um i liked you know catherine hardwicke is like you know kind of ridiculousgt: 0 prediction: --gt: -0.6 prediction:0.42This story demonstrates how ridiculous and petty fighting, or war... and um i liked you know catherine hardwicke is like you know kind of ridiculousgt: -2 prediction: --gt: -0.6 prediction: -0.56
Target
Sourcewell if they had kept their mouth shut, that movie would have been fabulous. and it was kind of interestinggt: -1 prediction: --gt: 1.2 prediction:-1.94Learn a language if you can, because that will make your life more interesting and it was kind of interestinggt: 2.0 prediction: --gt: 1.2 prediction:1.53
Target
" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.348, + 0.895, + 0.391 + ], + "angle": 0, + "content": "Figure 6. Some samples analysis of one baseline MDMN and our model OSAN for UDA task CMU-MOSEI \\(\\rightarrow\\) CMU-MOSI. Columns show information of one representative frame extracted from video clip, text message and the prediction result. Rows give information of target samples and closest source samples." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.406, + 0.471, + 0.496 + ], + "angle": 0, + "content": "our model on two tasks. Two baselines are Direct Transfer and our model without the two proposed modules TAL and DDG. We observe that JSD of our model is smaller than JSD of both baselines. This phenomenon implies that our model can reduce the domain gap more effectively than the baselines." + }, + { + "type": "image", + "bbox": [ + 0.079, + 0.512, + 0.47, + 0.673 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.685, + 0.471, + 0.726 + ], + "angle": 0, + "content": "Figure 7. Distribution discrepancy of different methods on two different tasks CMU-MOSEI \\(\\rightarrow\\) CMU-MOSI and Text-news \\(\\rightarrow\\) Text-show." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.742, + 0.212, + 0.756 + ], + "angle": 0, + "content": "5.3. Case Studies" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.903 + ], + "angle": 0, + "content": "To deeply explore the advantages of our model over other baselines, we give some predictions of a two-stage baseline MDMN and our model OSAN for the UDA task CMU-MOSEI \\(\\rightarrow\\) CMU-MOSI, as shown in Fig.6. We select some typical samples from the target domain to explore how our model works. For each target sample, we find the closest sample from source domain by calculating cosine similarity between target sample and source sample. We observe that the closest source samples picked by our model OSAN are" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.406, + 0.895, + 0.633 + ], + "angle": 0, + "content": "more relevant to the target samples. Specifically, take the target sample and source sample in the first row for example, with our model, speakers in both the source frame and target frame are very happy. Moreover, there is one common keyword in what the speakers said: enjoyed. Combining these two modalities of image and text, our model correctly classifies this target sample as positive sentiment with a score of 1.85, which is very close to the ground truth score of 2. In contrast, source sample found by the MDMN is not relevant to the target sample. There is no similar semantic information in what they said. The expressions in the frames are dissimilar. The phenomenon in these cases implies that our model can learn a common latent space in which the source domain and target domain are well aligned." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.646, + 0.619, + 0.661 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.671, + 0.895, + 0.853 + ], + "angle": 0, + "content": "To address the problems of domain adaptation and modality alignment in multimodal domain adaptation, we propose a novel one-stage alignment network to unify these two problems in one stage. In our model, the TAL module is proposed to explore the relationship between domains and modalities, which can provide complementary information to each other. Moreover, the DDG module is designed to generate new soft samples that mix information from both the source and target domains, guiding our model in a more continuous space. Extensive experiments on two multimodal tasks with different extents of domain shift demonstrate the excellent performance of our model." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3558" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Shai Ben-David, John Blitzer, Koby Crammer, Alex Kulesza, Fernando Pereira, and Jennifer Wortman Vaughan. A theory of learning from different domains. Machine learning, 79(1):151-175, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.471, + 0.229 + ], + "angle": 0, + "content": "[2] Bent Fuglede and Flemming Topsoe. Jensen-shannon divergence and hilbert space embedding. In International Symposium on Information Theory, 2004. ISIT 2004. Proceedings., page 31. IEEE, 2004. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.231, + 0.471, + 0.273 + ], + "angle": 0, + "content": "[3] Yaroslav Ganin and Victor Lempitsky. Unsupervised domain adaptation by backpropagation. In International conference on machine learning, pages 1180-1189. PMLR, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.274, + 0.471, + 0.343 + ], + "angle": 0, + "content": "[4] Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario Marchand, and Victor Lempitsky. Domain-adversarial training of neural networks. The journal of machine learning research, 17(1):2096-2030, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.345, + 0.471, + 0.427 + ], + "angle": 0, + "content": "[5] Arthur Gretton, Dino Sejdinovic, Heiko Strathmann, Sivaraman Balakrishnan, Massimiliano Pontil, Kenji Fukumizu, and Bharath K Striperumbudur. Optimal kernel choice for large-scale two-sample tests. In Advances in neural information processing systems, pages 1205-1213. Citeseer, 2012. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.43, + 0.471, + 0.472 + ], + "angle": 0, + "content": "[6] Wei Han, Hui Chen, and Soujanya Poria. Improving multimodal fusion with hierarchical mutual information maximization for multimodal sentiment analysis, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.474, + 0.471, + 0.543 + ], + "angle": 0, + "content": "[7] Devamanyu Hazarika, Roger Zimmermann, and Soujanya Poria. Misa: Modality-invariant and-specific representations for multimodal sentiment analysis. In Proceedings of the 28th ACM International Conference on Multimedia, pages 1122-1131, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.545, + 0.471, + 0.627 + ], + "angle": 0, + "content": "[8] Donghyun Kim, Yi-Hsuan Tsai, Bingbing Zhuang, Xiang Yu, Stan Sclaroff, Kate Saenko, and Manmohan Chandraker. Learning cross-modal contrastive features for video domain adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13618-13627, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.63, + 0.471, + 0.672 + ], + "angle": 0, + "content": "[9] Ye Liu, Changchong Lu, Chen Lin, Di Yin, and Bo Ren. Contrastive graph multimodal model for text classification in videos. arXiv preprint arXiv:2206.02343, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.673, + 0.471, + 0.743 + ], + "angle": 0, + "content": "[10] Ye Liu, Lingfeng Qiao, Di Yin, Zhuoxuan Jiang, Xinghua Jiang, Deqiang Jiang, and Bo Ren. Os-msl: One stage multimodal sequential link framework for scene segmentation and classification. In Proceedings of the 30th ACM International Conference on Multimedia, pages 6269-6277, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.745, + 0.471, + 0.813 + ], + "angle": 0, + "content": "[11] Zhun Liu, Ying Shen, Varun Bharadhwaj Lakshminarasimhan, Paul Pu Liang, Amir Zadeh, and Louis-Philippe Morency. Efficient low-rank multimodal fusion with modality-specific factors. arXiv preprint arXiv:1806.00064, 2018.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.816, + 0.471, + 0.871 + ], + "angle": 0, + "content": "[12] Mingsheng Long, Yue Cao, Jianmin Wang, and Michael Jordan. Learning transferable features with deep adaptation networks. In International conference on machine learning, pages 97-105. PMLR, 2015. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.873, + 0.471, + 0.903 + ], + "angle": 0, + "content": "[13] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representational" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.116, + 0.471, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "tations for vision-and-language tasks. arXiv preprint arXiv:1908.02265, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.121, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[14] Jonathan Munro and Dima Damen. Multi-modal domain adaptation for fine-grained action recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 122-132, 2020. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.177, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[15] Jonathan Munro and Dima Damen. Multi-modal domain adaptation for fine-grained action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.233, + 0.892, + 0.287 + ], + "angle": 0, + "content": "[16] Sinno Jialin Pan, Ivor W Tsang, James T Kwok, and Qiang Yang. Domain adaptation via transfer component analysis. IEEE transactions on neural networks, 22(2):199-210, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.288, + 0.892, + 0.357 + ], + "angle": 0, + "content": "[17] Fan Qi, Xiaoshan Yang, and Changsheng Xu. A unified framework for multimodal domain adaptation. In Proceedings of the 26th ACM International Conference on Multimedia, MM '18, page 429-437, New York, NY, USA, 2018. Association for Computing Machinery. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.358, + 0.892, + 0.412 + ], + "angle": 0, + "content": "[18] Lingfeng Qiao, Chen Wu, Ye Liu, Haoyuan Peng, Di Yin, and Bo Ren. Grafting pre-trained models for multimodal headline generation. arXiv preprint arXiv:2211.07210, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.413, + 0.892, + 0.496 + ], + "angle": 0, + "content": "[19] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.497, + 0.892, + 0.58 + ], + "angle": 0, + "content": "[20] Wasifur Rahman, Md Kamrul Hasan, Sangwu Lee, Amir Zadeh, Chengfeng Mao, Louis-Philippe Morency, and Ehsan Hoque. Integrating multimodal information in large pretrained transformers. In Proceedings of the conference. Association for Computational Linguistics. Meeting, volume 2020, page 2359. NIH Public Access, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.58, + 0.892, + 0.634 + ], + "angle": 0, + "content": "[21] Baochen Sun and Kate Saenko. Deep coral: Correlation alignment for deep domain adaptation. In European conference on computer vision, pages 443-450. Springer, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.636, + 0.892, + 0.705 + ], + "angle": 0, + "content": "[22] Zhongkai Sun, Prathusha Sarma, William Sethares, and Yingyu Liang. Learning relationships between text, audio, and video via deep canonical correlation for multimodal language analysis. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 8992-8999, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.706, + 0.892, + 0.748 + ], + "angle": 0, + "content": "[23] Hui Tang and Kui Jia. Discriminative adversarial domain adaptation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 5940-5947, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.748, + 0.892, + 0.803 + ], + "angle": 0, + "content": "[24] Dacheng Tao, Xuelong Li, Xindong Wu, and Steve Maybank. Tensor rank one discriminant analysis-a convergent method for discriminative multilinear subspace selection. Neurocomputing, 71(10):1866-1882, 2008. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.804, + 0.892, + 0.859 + ], + "angle": 0, + "content": "[25] Dacheng Tao, Xuelong Li, Xindong Wu, and Stephen J Maybank. General tensor discriminant analysis and gabor features for gait recognition. IEEE transactions on pattern analysis and machine intelligence, 29(10):1700-1715, 2007. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[26] Wilhelm Tranheden, Viktor Olsson, Juliano Pinto, and Lennart Svensson. Dacs: Domain adaptation via cross-domain mixed sampling. In Proceedings of the IEEE/CVF" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3559" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.468, + 0.12 + ], + "angle": 0, + "content": "Winter Conference on Applications of Computer Vision, pages 1379-1389, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.469, + 0.177 + ], + "angle": 0, + "content": "[27] Eric Tzeng, Judy Hoffman, Kate Saenko, and Trevor Darrell. Adversarial discriminative domain adaptation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7167-7176, 2017. 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.469, + 0.22 + ], + "angle": 0, + "content": "[28] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9(11), 2008. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.222, + 0.469, + 0.276 + ], + "angle": 0, + "content": "[29] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Ilia Polosukhin. Attention is all you need. arXiv preprint arXiv:1706.03762, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.278, + 0.469, + 0.346 + ], + "angle": 0, + "content": "[30] Wei Wang, Haojie Li, Zhengming Ding, Feiping Nie, Junyang Chen, Xiao Dong, and Zhihui Wang. Rethinking maximum mean discrepancy for visual domain adaptation. IEEE Transactions on Neural Networks and Learning Systems, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.349, + 0.469, + 0.391 + ], + "angle": 0, + "content": "[31] Garrett Wilson and Diane J Cook. A survey of unsupervised deep domain adaptation. ACM Transactions on Intelligent Systems and Technology (TIST), 11(5):1-46, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.393, + 0.469, + 0.461 + ], + "angle": 0, + "content": "[32] Xinyi Wu, Zhenyao Wu, Hao Guo, Lili Ju, and Song Wang. Darnet: A one-stage domain adaptation network for unsupervised nighttime semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15769-15778, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.463, + 0.469, + 0.532 + ], + "angle": 0, + "content": "[33] Minghao Xu, Jian Zhang, Bingbing Ni, Teng Li, Chengjie Wang, Qi Tian, and Wenjun Zhang. Adversarial domain adaptation with domain mixup. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 6502-6509, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.534, + 0.469, + 0.588 + ], + "angle": 0, + "content": "[34] Yiming Xu, Lin Chen, Zhongwei Cheng, Lixin Duan, and Jiebo Luo. Open-ended visual question answering by multimodal domain adaptation. arXiv preprint arXiv:1911.04058, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.591, + 0.469, + 0.632 + ], + "angle": 0, + "content": "[35] Shen Yan, Huan Song, Nanxiang Li, Lincan Zou, and Liu Ren. Improve unsupervised domain adaptation with mixup training. arXiv preprint arXiv:2001.00677, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.634, + 0.469, + 0.689 + ], + "angle": 0, + "content": "[36] Wenmeng Yu, Hua Xu, Ziqi Yuan, and Jiele Wu. Learning modality-specific representations with self-supervised multi-task learning for multimodal sentiment analysis. arXiv preprint arXiv:2102.04830, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.69, + 0.469, + 0.745 + ], + "angle": 0, + "content": "[37] Amir Zadeh, Minghai Chen, Soujanya Poria, Erik Cambria, and Louis-Philippe Morency. Tensor fusion network for multimodal sentiment analysis. arXiv preprint arXiv:1707.07250, 2017. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.747, + 0.469, + 0.802 + ], + "angle": 0, + "content": "[38] Amir Zadeh, Rowan Zellers, Eli Pincus, and Louis-Philippe Morency. Mosi: multimodal corpus of sentiment intensity and subjectivity analysis in online opinion videos. arXiv preprint arXiv:1606.06259, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.469, + 0.899 + ], + "angle": 0, + "content": "[39] AmirAli Bagher Zadeh, Paul Pu Liang, Soujanya Poria, Erik Cambria, and Louis-Philippe Morency. Multimodal language analysis in the wild: Cmu-mosei dataset and interpretable dynamic fusion graph. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2236–2246, 2018. 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[40] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[41] Tong Zhang, Di Wang, Huanhuan Chen, Zhiwei Zeng, Wei Guo, Chunyan Miao, and Lizhen Cui. Bdann: Bert-based domain adaptation neural network for multi-modal fake news detection. In 2020 International Joint Conference on Neural Networks (IJCNN), pages 1-8, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.892, + 0.26 + ], + "angle": 0, + "content": "[42] Weichen Zhang, Dong Xu, Jing Zhang, and Wanli Ouyang. Progressive modality cooperation for multi-modality domain adaptation. IEEE Transactions on Image Processing, 30:3293-3306, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.262, + 0.892, + 0.331 + ], + "angle": 0, + "content": "[43] Yiyi Zhang, Li Niu, Ziqi Pan, Meichao Luo, Jianfu Zhang, Dawei Cheng, and Liqing Zhang. Exploiting motion information from unlabeled videos for static image action recognition. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 12918-12925, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.333, + 0.892, + 0.401 + ], + "angle": 0, + "content": "[44] Honghao Zhou, Tinghuai Ma, Huan Rong, Yurong Qian, Yuan Tian, and Najla Al-Nabhan. Mdmn: Multi-task and domain adaptation based multi-modal network for early rumor detection. Expert Systems with Applications, 195:116517, 2022. 1, 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.403, + 0.892, + 0.459 + ], + "angle": 0, + "content": "[45] Junbao Zhuo, Shuhui Wang, Weigang Zhang, and Qingming Huang. Deep unsupervised convolutional domain adaptation. In Proceedings of the 25th ACM international conference on Multimedia, pages 261-269, 2017. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3560" + } + ] +] \ No newline at end of file diff --git a/2023/OSAN_ A One-Stage Alignment Network To Unify Multimodal Alignment and Unsupervised Domain Adaptation/4429e1d5-6fb8-40a5-9804-367d3c0ae901_origin.pdf b/2023/OSAN_ A One-Stage Alignment Network To Unify Multimodal Alignment and Unsupervised Domain Adaptation/4429e1d5-6fb8-40a5-9804-367d3c0ae901_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2fa19392656c33854daf80a9ea54e40582f27425 --- /dev/null +++ b/2023/OSAN_ A One-Stage Alignment Network To Unify Multimodal Alignment and Unsupervised Domain Adaptation/4429e1d5-6fb8-40a5-9804-367d3c0ae901_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4dbae156120c8b1ac213124206e705c6f59e13dff3de30758a8797d971e8926d +size 2543291 diff --git a/2023/OSAN_ A One-Stage Alignment Network To Unify Multimodal Alignment and Unsupervised Domain Adaptation/full.md b/2023/OSAN_ A One-Stage Alignment Network To Unify Multimodal Alignment and Unsupervised Domain Adaptation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..61ca06ecb563e7e2968aab0335f2e9dfb99ba2f4 --- /dev/null +++ b/2023/OSAN_ A One-Stage Alignment Network To Unify Multimodal Alignment and Unsupervised Domain Adaptation/full.md @@ -0,0 +1,317 @@ +# OSAN: A One-Stage Alignment Network to Unify Multimodal Alignment and Unsupervised Domain Adaptation + +Ye Liu, Lingfeng Qiao, Changchong Lu, Di Yin, Chen Lin, Haoyuan Peng, Bo Ren +Tencent Youtu Lab + +{rafelliu, leafqiao, chonglu, endymecyyin, iwlliamlin, haoyuanpeng, timren}@tencent.com + +# Abstract + +Extending from unimodal to multimodal is a critical challenge for unsupervised domain adaptation (UDA). Two major problems emerge in unsupervised multimodal domain adaptation: domain adaptation and modality alignment. An intuitive way to handle these two problems is to fulfill these tasks in two separate stages: aligning modalities followed by domain adaptation, or vice versa. However, domains and modalities are not associated in most existing two-stage studies, and the relationship between them is not leveraged which can provide complementary information to each other. In this paper, we unify these two stages into one to align domains and modalities simultaneously. In our model, a tensor-based alignment module (TAL) is presented to explore the relationship between domains and modalities. By this means, domains and modalities can interact sufficiently and guide them to utilize complementary information for better results. Furthermore, to establish a bridge between domains, a dynamic domain generator (DDG) module is proposed to build transitional samples by mixing the shared information of two domains in a self-supervised manner, which helps our model learn a domain-invariant common representation space. Extensive experiments prove that our method can achieve superior performance in two real-world applications. The code will be publicly available. + +# 1. Introduction + +With explosively emerging multimedia data on the Internet, the field of multimodal analysis achieves more and more attention [10, 13, 18, 19, 43]. Compared to extensive unimodal models in NLP and CV, learning adequate knowledge from multimodal signals is still preliminary but very important. Abundant data plays a key role in different scenarios of multimodal analysis, such as pre-training or downstream multimedia tasks. However, it is prohibitively expensive and time-consuming to obtain large amounts of labeled data. To eliminate this issue, domain adaptation (DA) + +![](images/884fcf7f5421700cb5403e4a222adf854fe52088a0f25c74bd6263691899488a.jpg) +Figure 1. Conception of our one-stage model. + +is raised to learn a model from a labeled dataset (source domain) that can be generalized to other related tasks without sufficient labeled data (target domain) [3]. Classical domain adaptation can be classified into different categories: unsupervised domain adaptation (UDA), fully supervised domain adaptation, and semi-supervised domain adaptation [31]. In this paper, we focus on UDA where no samples in target domain are annotated. With this technique, it is not necessary to prepare a customized training dataset for a specific task, but it can perform the task effectively and efficiently. + +There are two challenges when applying domain adaptation to multimodal scenarios [17]: (1) how to align the source and target domains and remit domain discrepancy, and (2) how to align multiple modalities and leverage multimodal information. Most existing works address these two problems in two consecutive stages: multimodal alignment followed by domain adaptation [34, 41], or vice versa [14, 44]. However, they solve these two issues separately without considering their relationship: domain and modality can be treated as two views to portray the intrinsic characteristic of multimodal data [8], and the hidden underlying relationship in these two views can provide complementary information to each other. Unimodal domain adaptation methods can not work well in multimodal tasks due to the inability to preserve the relations between modalities at the same time. Through our experiments and analysis, we observe that the two-stage model could not achieve ideal performance. Fig.2 shows the learning curve of two-stage + +model during training phase by 800 iterations for the task of multimodal sentiment analysis. It can be found that the learning curve of two-stage model is oscillating and converges slowly, which indicates that two-stage model is probably not a superior solution. To handle these challenges, the objective of multimodal domain adaptation can be defined as: (1) Exploring the relationship between domains and modalities; (2) Finding a common domain-invariant cross-modal representation space to align domains and modalities simultaneously. + +Therefore, in this paper, we design a One-Stage Alignment Network (OSAN) to unify multimodal alignment and domain adaptation in one stage. Fig.1 shows the conception of our one-stage model. Our method benefits from: (1) The modality and domain are associated and interacted to capture the relationship between domains and modalities, which can provide rich complementary information to each other. (2) Multimodal alignment and domain adaptation are unified in one stage, which allows our model to perform domain adaptation and leverage multimodal information at the same time. In Fig.2, we observe that the learning curve of our method is relatively stable and converges better, which indicates that exploring the relation between modality and domain contributes to our task. + +In summary, our contributions are as follows: + +(1) To capture the relationship between domain and modality, we propose a one-stage alignment network, called OSAN, to associate domain and modality. In this way, a joint domain-invariant and cross-modal representation space is learned in one stage. +(2) We design a TAL module to bring sufficient interactions between domains and modalities and guide them to utilize complementary information for each other. +(3) To effectively bridge distinct domains, a DDG module is developed to dynamically construct multiple new domains by combining knowledge of source and target domains and exploring intrinsic structure of data distribution. +(4) Extensive experiments on two totally different tasks demonstrate the effectiveness of our method compared to the supervised and strongly UDA methods. + +# 2. Related Work + +Unimodal Domain Adaption The main approaches for domain adaptation or domain alignment can be categorized as discrepancy-based and adversarial-based methods [1, 23, 30]. Discrepancy-based methods design statistics, such as maximum mean discrepancy (MMD) [16], correlation alignment (CORAL) [45] and etc., to measure the difference between two domain distributions. For adversarial-based methods, Ganin et al. [4] used the concept of generative adversarial networks (GAN) to obtain domain-invariant characteristics. Wu et al. [32] proposed a one-stage adaptation framework for nighttime semantic segmentation to per + +![](images/eddb63d74bc5cbe37d8b01f66f52ee8d5afa958fd9aa4be022d75dfe9406e75e.jpg) +Figure 2. The learning curves of our one-stage method and traditional two-stage method in multimodal sentiment analysis task. + +form an end-to-end learning where a separate common preprocessing step was integrated, which is different from our motivation. However, these studies only focus on the elimination of domain shift in unimodal scenarios, which may lose effect when applied in multimodal area. + +Multimodal Domain Adaption Qi et al. [17] first analyzed the multimodal domain adaptation and raised two problems, i.e., domain adaptation and modality alignment. To alleviate them, hybrid domain constraints and attention-based modality fusion module were introduced to learn domain-invariant fused features. More recently, some studies [14, 34, 41, 42, 44] related to this area have tried to address the multimodal problem by extending the unimodal domain adaptation methods. Xu et al. [34] proposed a supervised multimodal domain adaptation method for VQA to learn joint feature embeddings across different domains and modalities. Munro et al. [14] exploited the correspondence of modalities as a self-supervised alignment approach for UDA in addition to adversarial alignment. Furthermore, Zhou et al. [44] proposed MDMN for early rumor detection, which can combine textual and visual information with two heterogeneous feature extractors. However, all of these studies solved the two problems separately in two stages. Moreover, the relationship between modality and domain is missing, thus they cannot achieve ideal performance. + +# 3. Proposed Method + +Fig.3 presents an overview of our method that contains four parts: multimodal feature extraction, tensor-based alignment, dynamic domain generator and task-specific heads. First, encoders, as multimodal feature extractor, map source and target multimodal data to different latent spaces. Then, by means of an effective and efficient way of tensor representation, we propose a TAL module to acquire relation information through consecutive interactions between + +![](images/58c94c1fa496aeb949ff7c4f381a7d8ad55528349d45db9521c07bd93c96f56c.jpg) +Figure 3. Overview of our method OSAN. We use circle, square and triangle to represent three modalities of text, vision and audio. In the figure, $\times$ means mode- $d$ product of a tensor and a matrix, $\odot$ represents element-wise product, and $\ast$ means tensor factorization. + +modalities and domains. Thus, the modality and domain are aligned simultaneously in one stage. After that, we develop a DDG module to mix information from different domains to create new transitional domains. Finally, our network is split into three branches: category classification, domain adversarial learning and discrepancy elimination. + +# 3.1. Tensor-based Alignment + +We develop a TAL module to simultaneously address the challenges of multimodal alignment and domain adaptation. We first resort to tensor representation, an effective and natural representation of multimodal data, to develop models capturing inter-modality dynamics [11, 37]. Then, we utilize statistical measurements to establish an interaction between domains and modalities to explore the relation information. In this way, we associate domains and modalities so that they can communicate and align together. Specifically, we perform tensor factorization to fuse multimodal information and maximize statistical measurements to simultaneously eliminate domain gaps. + +Specifically, we use $\underline{\mathbf{X}}^{I_1\times \dots \times I_N\times M}$ and $\underline{\mathbf{Y}}^{J_1\times \dots \times J_N\times M}$ to represent source and target dataset with $N$ modalities and $M$ samples. For example, a video can be represented by a third-order tensor using a differentiable outer product between visual, textual, and acoustic representations [37]. For simplicity, we first make the training samples zero mean. + +To perform multimodal alignment, TAL aims to find pairs of linear transformations $\mathbf{U}^{(n)}|_{n = 1}^{N}$ and $\mathbf{V}^{(n)}|_{n = 1}^{N}$ for each modality of source and target domains to project samples of two sets into low dimensional subspaces. During this process, we establish an interaction between domain and modality by maximizing a statistical measurement of covariance given a normalized standard deviation. As reported in [45], we choose covariance instead of the first-order mean statistics used in MK-MMD [5, 21], as it is more powerful and achieves better performance. By this means, domains and modalities can be aligned simultaneously in one stage. Thus the objective function is defined as follows: + +$$ +\begin{array}{l} \left\{\mathbf {U} ^ {(n)} | _ {n = 1} ^ {N}, \mathbf {V} ^ {(n)} | _ {n = 1} ^ {N} \right\} \\ = \max \llbracket (\underline {{\mathbf {X}}} \prod_ {n = 1} ^ {N} \times_ {n} \mathbf {U} ^ {(n) ^ {T}}) \otimes (\underline {{\mathbf {Y}}} \prod_ {n = 1} ^ {N} \times_ {n} \mathbf {V} ^ {(n) ^ {T}}); \rrbracket \\ \end{array} +$$ + +$$ +\begin{array}{l} s. t. \left(\underline {{\mathbf {X}}} \prod_ {n = 1} ^ {N} \times_ {n} \mathbf {U} ^ {(n) ^ {T}}\right) _ {(N + 1)} ^ {T} \left(\underline {{\mathbf {X}}} \prod_ {n = 1} ^ {N} \times_ {n} \mathbf {U} ^ {(n) ^ {T}}\right) _ {(N + 1)} = \mathbf {I} \tag {1} \\ \left(\underline {{\mathbf {Y}}} \prod_ {n = 1} ^ {N} \times_ {n} \mathbf {V} ^ {(n) ^ {T}}\right) _ {(N + 1)} ^ {T} \left(\underline {{\mathbf {Y}}} \prod_ {n = 1} ^ {N} \times_ {n} \mathbf {V} ^ {(n) ^ {T}}\right) _ {(N + 1)} = \mathbf {I} \\ \end{array} +$$ + +The objective function above does not have a closed-form solution, so instead we derive a suboptimal solution following the principle of the alternating projection method [24], where the complicated optimization problem is re + +duced into smaller conditional subproblems that can be solved by simple established methods. Therefore, the objective function is decomposed into $N$ different subproblems: + +$$ +\left\{\mathbf {U} ^ {(n)}, \mathbf {V} ^ {(n)} \right\} = \max t r \left\{\mathbf {U} ^ {(n) ^ {T}} \mathbf {C} _ {x y} ^ {(n)} \mathbf {V} ^ {(n)} \right\} \tag {2} +$$ + +$$ +s. t. \mathbf {U} ^ {(n) ^ {T}} \mathbf {C} _ {x x} ^ {(n)} \mathbf {U} ^ {(n)} = \mathbf {I}, \mathbf {V} ^ {(n) ^ {T}} \mathbf {C} _ {y y} ^ {(n)} \mathbf {V} ^ {(n)} = \mathbf {I} +$$ + +where, for simplification, we define $\mathbf{C}_{xy}^{(n)}$ , $\mathbf{C}_{xx}^{(n)}$ as + +$$ +\mathbf {C} _ {x y} ^ {(n)} = \left(\llbracket (\underline {{\mathbf {X}}} ^ {(\bar {n})}) \otimes (\underline {{\mathbf {Y}}} ^ {(\bar {n})}); (\bar {n}) (\bar {n}) \rrbracket\right) \tag {3} +$$ + +$$ +\mathbf {C} _ {x x} ^ {(n)} = \left(\llbracket (\underline {{\mathbf {X}}} ^ {(\bar {n})}) \otimes (\underline {{\mathbf {X}}} ^ {(\bar {n})}); (\bar {n}) (\bar {n}) \rrbracket\right) +$$ + +where $([\cdot \otimes ;(\bar{n})(\bar{n})])$ means tensor contraction on all indices except the $n$ -th index on the tensor product [25]. Similarly, $\mathbf{C}_{yy}^{(n)}$ and $\mathbf{C}_{yx}^{(n)}$ is same as $\mathbf{C}_{xx}^{(n)}$ and $\mathbf{C}_{xy}^{(n)}$ in form. By Eq.2, an interaction between source and target domains for the $n$ -th modality is established. + +# 3.2. Dynamic Domain Generator + +In domain adaptation, adversarial learning is often used to align domains by developing a domain discriminator to judge samples from the source or target domains [27, 33]. By this means, it brings a two-class classification task with hard assignments of $\mathbf{1}$ or $\mathbf{0}$ . They try to confuse the domain discriminator to learn a shared common representation space between the source and target domains. However, it usually achieves worse performance due to a huge variance between the two domains. + +To overcome this problem, several works have been proposed to produce some transitional domains based on Mixup [35, 40]. They mixed images from the source and target domains to generate more soft and mixed samples [26, 33]. Suppose there are one sample $s$ from source domain and another sample $t$ from target domain, $s$ and $t$ are linearly interpolated [40] to generate a new sample. This strategy can fully utilize the inter-domain information and improve the generalization ability of models. However, it has an obvious shortcoming: the generated sample fuses full information of two domains, which contains redundant information and speciality of domains that may induce undesirable oscillations on training the domain discriminator. + +To solve this issue, we propose a DDG module to establish a bridge between domains. Specifically, DDG explicitly captures commonality and abandons the specialty of domains. By highlighting this commonality, we make the domain discriminator focus on commonality rather than full information, which helps our model learn a domain-invariant common representation space. + +Given one sample $s$ from source domain and another sample $t$ from target domain, DDG is proposed to dynamically generate soft samples $s^{\prime}$ and $t^\prime$ to smoothly bridge the + +domain gap. Specifically, $s'$ represents a variety of $s$ which maintains information of $s$ and mix source-relevant commonality from $t'$ . Similarly, $t'$ is a variety of $t$ . As shown in Fig.3, we capture the commonality between $s$ and $t$ with the help of the popular and effective self-attention mechanism used in Transformers [29]. We construct a sequence in which the length is 2 by stacking $s$ and $t$ . Denote query $Q_{s}$ and key $K_{s}$ is projected from the source embedding $s$ respectively, while query $Q_{t}$ and key $K_{t}$ are from the target embedding $t$ , respectively. With dot-product between queries and keys, the source-target attention matrix $M_{st}$ is obtained, which represents the relations between source and target. Afterwards, we design a gating to highlight the commonality between $s$ and $t$ , and fuse the commonality to generate $s'$ and $t'$ respectively. In detail, to generate $s'$ , we calculate the feature-relevant distribution $p$ by summing $M_{st}$ along the dimension of the query. Larger value in $p$ means related element in $t$ is more-relevant to $s$ , which should be chosen for commonality. As a result, the commonality $c$ is obtained by $c = t \odot \text{Norm}(p)$ , where $\text{Norm}(\cdot)$ means the normalized operator and $\odot$ is the element-wise multiplication. Finally, $s'$ is obtained by fusing commonality $c$ and information from $s$ . + +$$ +s ^ {\prime} = \lambda s + (1 - \lambda) c \tag {4} +$$ + +Finally, we use multi-head attention to generate various new domains to enrich the sample space. Samples from new domains and raw source and target domains are fed to domain discriminator, by which the domain discriminator is guided by the hard label information and well-designed soft domains. Each sample from these soft domains explores the intrinsic structure of data distribution from raw domains and enriches feature patterns by the interaction of two domains. + +# 3.3. Tasks + +Our model consists of three tasks: category classification, domain adversarial learning, and domain discrepancy elimination. First, we perform $K$ -way object classification on the source domain. Second, a domain discriminator is introduced to judge samples from source, target or the created soft domains. Finally, we try to learn transferable features by minimizing domain discrepancy. In summary, we have: + +$$ +L = L _ {c l s} + \alpha L _ {a d v} + \beta L _ {d i s} \tag {5} +$$ + +where $L_{cls}$ is the category classification loss, $L_{adv}$ is the adversarial loss, and $L_{dis}$ is the discrepancy loss. $\alpha$ and $\beta$ are weights to balance these three losses. Cross entropy loss is used as classification and adversarial loss, while MK-MMD [5] is employed as discrepancy loss. + +# 4. Experiments + +We conduct experiments on two totally different tasks to demonstrate the efficacy of our model. One task is popu + +Table 1. Multimodal sentiment analysis results on CMU-MOSI and CMU-MOSEI. †: results come from [7]; ‡: results come from [36]; ◇: results come from [6]; ↓: the lower the better. + +
MethodsCMU-MOSEI → CMU-MOSICMU-MOSI → CMU-MOSEI
MAE ↓CorrAcc-7Acc-2F1MAE ↓CorrAcc-7Acc-2F1
Direct Transfer0.7940.76439.579.7/81.579.5/81.40.6210.68551.379.54/82.1480.84/81.33
Supervised
TFN [37] †0.9010.69834.9-/80.8-/80.70.5930.70050.2-/82.5-/82.1
ICCN [22] †0.8620.71439.0-/83.0-/83.00.5650.71351.6-/84.2-/84.2
MISA [7] ‡0.8040.764-80.79/82.1080.77/82.030.5680.724-82.59/84.2382.67/83.97
MAG-BERT [20] ‡0.7310.789-82.50/84.3082.60/84.300.5390.753-83.80/85.2083.70/85.10
Self-MM [36] ‡0.7130.798-84.00/85.9884.42/85.950.5300.765-82.81/85.1782.53/85.30
MMIM [6] ◇0.7000.80046.6584.14/86.0684.00/85.980.5260.77254.2482.24/85.9782.66/85.94
UDA
DAN [12]0.7770.77439.7980.03/81.7179.74/81.490.6140.69351.680.24/81.3281.36/82.47
ADDA [27]0.7840.77340.1480.12/82.2680.13/82.320.6360.70751.480.47/81.5981.53/82.76
MM-SADA [15]0.7870.76940.5280.9/82.7780.68/82.630.6670.68452.180.32/81.4481.26/81.95
MDMN [44]0.7780.77439.6581.92/82.0181.97/82.110.6020.71252.882.24/82.3882.95/83.26
OSAN(TAL + Mixup)0.7530.78242.6482.44/83.3282.14/83.210.5420.75753.1482.76/82.8883.13/83.96
OSAN(TAL + DDG)0.7130.80146.3883.12/84.5883.02/84.510.5320.76853.8483.41/84.3683.31/84.47
+ +lar multimodal sentiment analysis, the other is a new task called video text classification. + +# 4.1. Multimodal Sentiment Analysis + +Multimodal sentiment analysis is a new dimension of traditional text-based sentiment analysis, which goes beyond the analysis of texts, and includes other modalities data. This task processes data from multiple sources, such as acoustic, visual, and textual information to understand various human emotions. + +# 4.1.1 Experimental Setup + +Datasets and Model Structure We use two classical datasets, called CMU-MOSEI [39] and CMU-MOSI [38], for the multimodal sentiment analysis task. + +Metrics We use the metrics that have been presented in [6]: mean absolute error (MAE), which is the average mean absolute difference value between predicted values and truth values. Pearson correlation (Corr) that measures the degree of prediction skew, seven-class classification accuracy (Acc-7) indicating the proportion of predictions that correctly fall into the same interval of seven intervals between -3 and +3 as the corresponding truths, binary classification accuracy (Acc-2) and F1 score for positive/negative and non-negative/negative classification results. + +Baselines To inspect the relative performance of our model, we compare our model with various baselines in three aspects: Direct Transfer, UDA and Supervised + +methods. Direct transfer means that we train a model using source dataset and then predict test samples in target dataset. For UDA, we use model structures of discrepancy-based model DAN [12], adversarial-based model ADDA [27], and some typical studies focused on multimodal area such as MM-SADA [14] and MDMN [44] where multimodal alignment and domain adaptation are performed in two separate stages. Besides direct transfer and UDA methods, we also bring supervised methods for evaluation. For supervised methods, we consider the pure learning-based model TFN [37], feature space manipulation like ICCN [22], MISA [7], more recent and competitive baselines including MAG-BERT [20], Self-MM [36] and MMIM [6]. + +# 4.1.2 Experimental Results + +Overall The overall results are shown in Table 1. In detail, we find that OSAN achieves the best results among all UDA methods on all metrics with large margin. Surprisingly, OSAN even achieves better performance than many supervised methods, such as TFN, MISA and MAG-BERT. + +Ablation Study To verify the contributions of each component of OSAN, we design a series of ablation experiments. First, we evaluate the effectiveness of the two proposed modules TAL and DDG by eliminating one or both of them from our model, as illustrated in Table 2. We notice a manifest performance degradation after removing TAL or DDG, and the results are even worse when both modules are removed. Furthermore, we conduct additional experiments to explore the contributions of the three losses. In Table 3, + +we find that only using classification loss $L_{cls}$ , our model can achieve competitive performance compared to traditional multimodal DA methods such as MDMN, demonstrating that both modality and domain are well aligned using TAL and DDG. Finally, we adjust some important hyperparameters of DDG and TAL to evaluate their importance. In terms of DDG, we adjust the number of created soft domains by changing multi-head $H$ . Furthermore, we adjust the fusion coefficient $\lambda$ to control the fusion ratio between the source and target samples. In Table 4, we observe that the performance improves significantly as the number of domains increases. We also find that the fusion coefficient set to 0.85 can achieve the best result. In terms of TAL, we reduce multimodal features to different combinations of dimensions. We find that the combination of $384 * 2 * 2$ achieves the best performance, as shown in Table 5. + +Table 2. Ablation study of the proposed modules on CMU-MOSI. + +
Modulew/o DDGw/o TALw/o Both
MAE ↓+0.039+0.051+0.074
Corr-0.023-0.021-0.032
Acc-7-3.55-4.65-5.86
Acc-2-1.03/-1.30-1.89/-1.72-2.22/-1.81
F1-1.15/-1.52-1.85/-1.61-2.34/-1.88
+ +Table 3. Ablation study of three losses on CMU-MOSI. + +
LossLclsLcls + LadvLcls + Ldis
MAE ↓+0.018+0.009+0.016
Corr-0.016-0.012-0.014
Acc-7-1.07-0.82-0.67
Acc-2-0.65/-0.95-0.32/-0.28-0.67/-0.74
F1-0.48/-0.69-0.7/-0.49-0.31/-0.58
+ +Table 4. Ablation study of DDG hyperparameters on CMU-MOSI. + +
DDGλ = 0.85 +H = 64λ = 0.9 +H = 64λ = 0.85 +H = 32λ = 0.9 +H = 32
MAE ↓0.00+0.0040.001+0.006
Corr0.00-0.01-0.003-0.016
Acc-70.00-0.590-0.27-0.75
Acc-20.00/0.00-0.27/-1.14-0.16/-1.07-0.39/-1.29
F10.00/0.00-0.259/-0.15-0.14/-0.04-0.39/-0.30
+ +# 4.2. Video Text Classification + +Video text classification is a real industrial task for video understanding [9]. With massive videos generated everyday, video textual information extraction is an essential work in many applications. However, numerous useless texts, such as rolling texts and blurred background texts, + +Table 5. Ablation study of TAL hyperparameters on CMU-MOSI. + +
TALt:384t:192t:192
v:2,a:2v:4,a:2v:2,a:4
MAE↓0.000+0.007+0.006
Corr0.00-0.009-0.006
Acc-70.00-1.17-0.91
Acc-20.00/0.00-0.77/-1.16-0.30/-0.91
F10.00/0.00-0.59/-2.02-0.13/-0.79
+ +have no good effect or even side effects on downstream tasks. To control and filter video content, we propose a new video text classification task to classify video texts into valuable categories. + +# 4.2.1 Experimental Setup + +Datasets and Model Structure We construct a well-defined industrial-grade dataset, called Text-news, which is dedicated to promote video text extraction research for news applications. The extraction of text in videos usually involves two steps: (1) text recognition and (2) text classification. We prepare various videos from news programs for annotation and use a general Optical Character Recognition (OCR) engine to locate and recognize all the texts in the videos. Then, an annotation system is designed for the video text classification task to tag video texts. Three representative categories are defined which are Caption, Subtitle and Others. To verify the effectiveness and generality of our method, we use Text-news as source domain dataset and construct a dataset named Text-show with a few annotated samples as the target domain dataset. The difference between the source and target datasets is that the video program is totally different. The source dataset is constructed from news videos, while the target dataset is constructed from videos of variety shows. We use the standard Precision, Recall, and F1 score for evaluation. + +Baselines To inspect the relative performance of our model OSAN, we compare it with various baselines in two aspects: Direct Transfer and UDA methods. Unlike multimodal sentiment analysis, we do not use Supervised methods for comparison because the target dataset does not have annotated training data. + +# 4.2.2 Experimental Results + +Overall The results are shown in Table 6. All the metrics of OSAN achieve the best results when compared to other UDA methods and direct transfer, leading to a splendid performance of video text classification task. + +Table 6. Video text classification results on Text-show. + +
MethodsText-news → Text-show
PrecisionRecallF1
Direct Transfer80.277.9879.08
UDA
DAN [12]87.4480.5883.87
ADDA [27]91.6683.6687.48
MM-SADA [15]94.0783.4988.46
MDMN [44]93.5483.6988.34
OSAN(TAL + Mixup)94.4284.7989.35
OSAN(TAL + DDG)95.0386.4490.53
+ +Table 7. Ablation study of the proposed modules on Text-show. + +
Modulew/o DDGw/o TALw/o Both
Precision-0.90-0.51-0.96
Recall-1.24-2.37-2.95
F1-1.09-1.55-2.07
+ +Table 8. Ablation study of three losses on Text-show. + +
LossLclsLcls + LadvLcls + Ldis
Precision-0.35-0.17-0.45
Recall-0.42-0.24-0.14
F1-0.39-0.21-0.28
+ +Table 9. Ablation study of DDG hyperparameters on Text-show. + +
DDGλ = 0.9 +H = 128λ = 0.95 +H = 128λ = 0.9 +H = 64λ = 0.95 +H = 64
Precision0.00-0.16-0.75-0.67
Recall0.00-0.32-1.21-1.33
F10.00-0.25-1.01-1.03
+ +Table 10. Ablation study of TAL hyperparameters on Text-show. + +
TALt:56,v:14t:14,v:56t:28,v:28
Precision0.000-0.96+0.24
Recall0.000+0.05-0.59
F10.000-0.41-0.22
+ +Ablation Study Similar to multimodal sentiment analysis, in order to show the benefits of TAL and DDG modules, we perform a series of ablation experiments. First, we eliminate one or both modules from our model. In Table 7, the overall F1 score decreases after removing single module, and the results are even worse when removing both modules, which demonstrates the efficacy of these two modules. Furthermore, we study the impact of three losses of category classification loss, adversarial loss, and discrepancy loss, as shown in Table 8. The results show that all three losses contribute to performance. Moreover, by removing + +DA related losses such as adversarial loss and discrepancy loss, it performs slightly worse than keeping all three losses, but still better than traditional multimodal DA methods such as MM-SADA and MDMN. This observation indicates that different domains are well aligned by TAL and DDG. Finally, we tuned some hyperparameters of TAL and DDG to explore their impact. For the DDG hyperparameters, in Table 9, we observe that a smaller fusion coefficient $\lambda$ and a larger multi-head $H$ can achieve better results. For the TAL hyperparameters, we reduce multimodal features to different dimensions. In Table 10, the combination $56*14$ for textual and visual features achieves the best result. + +![](images/840d248501d4438e62a3cdfe44e12b849405d601bbd5f71333a9677d3dfa676e.jpg) +(a) Directly transfer + +![](images/18bcbe9e4f1318cf0419f661e21a0db4a5140ef1749f31a0910d4c1eaa256dc5.jpg) +(b) Our method w/o TAL&DDG + +![](images/b97a35b19be5df092e1814c4f6975a49a3770ab886d41c439fb39c42280df987.jpg) +(c) Our method + +![](images/c009a84b8e253a616ccc50dcd967bb72587c0390e6ce73e71bb50392b10bc3eb.jpg) +Figure 4. Visualization of the feature distribution of the target domain for multimodal sentiment analysis. +(a) Directly transfer +Figure 5. Visualization of the feature distribution of the target domain for video text classification. + +![](images/212c66b99d36fc9d5a8926204770b18d8b453565ec8481f2c22ac38cd988c136.jpg) +(b) Our method w/o TAL&DDG + +![](images/e8bdbbb9e2bbeb10ddf20ad13275b340e6af69e92d63c3c6d6e5f16a17b817b2.jpg) +(c) Our method + +# 5. Discussion + +# 5.1. Visual Analysis of Feature Distribution + +We use t-SNE [28] to visualize the feature distribution of test data in target domain. Fig.5 illustrates the t-SNE visualization of the feature distribution in the video text classification task. We find that the features of different classes are separated more clearly, and the features in the same class are more centralized. A similar phenomenon is observed in Fig.4 of the feature distribution for the multimodal sentiment analysis task. This makes sense because TAL can explore the relationship between domains and modalities and guide them to utilize complementary information. Moreover, DDG can construct new domains that guarantee domain-invariance in a more continuous latent space. + +# 5.2. Distribution Discrepancy + +To quantify the distribution discrepancy of source and target domains, we use the classical metric of symmetric Jensen-Shannon divergence (JSD) [2]. Fig.7 shows the feature distribution discrepancy based on two baselines and + +
DomainImageTextMDMN ResultImageTextOSAN(ours) Result
SourceBeyond that, I actually excelled academically from some of my decisions um i say go see it coz i i enjoyed it nowgt: 0 prediction: --gt: 2.0 prediction:-0.37Thank you for that Godelieve, we all enjoyed your comments! um i say go see it coz i i enjoyed it nowgt: 2 prediction: --gt: 2.0 prediction:1.85
Target
Sourcehey were all randomly chosen ad we're going to have my lovely assistant here choose from among these three people and um i liked you know catherine hardwicke is like you know kind of ridiculousgt: 0 prediction: --gt: -0.6 prediction:0.42This story demonstrates how ridiculous and petty fighting, or war... and um i liked you know catherine hardwicke is like you know kind of ridiculousgt: -2 prediction: --gt: -0.6 prediction: -0.56
Target
Sourcewell if they had kept their mouth shut, that movie would have been fabulous. and it was kind of interestinggt: -1 prediction: --gt: 1.2 prediction:-1.94Learn a language if you can, because that will make your life more interesting and it was kind of interestinggt: 2.0 prediction: --gt: 1.2 prediction:1.53
Target
+ +our model on two tasks. Two baselines are Direct Transfer and our model without the two proposed modules TAL and DDG. We observe that JSD of our model is smaller than JSD of both baselines. This phenomenon implies that our model can reduce the domain gap more effectively than the baselines. + +![](images/1bdd0ff247178dfbcc7e838ba9559a1f5ec8830e1e8aad738f256c93fe6aac28.jpg) +Figure 6. Some samples analysis of one baseline MDMN and our model OSAN for UDA task CMU-MOSEI $\rightarrow$ CMU-MOSI. Columns show information of one representative frame extracted from video clip, text message and the prediction result. Rows give information of target samples and closest source samples. +Figure 7. Distribution discrepancy of different methods on two different tasks CMU-MOSEI $\rightarrow$ CMU-MOSI and Text-news $\rightarrow$ Text-show. + +# 5.3. Case Studies + +To deeply explore the advantages of our model over other baselines, we give some predictions of a two-stage baseline MDMN and our model OSAN for the UDA task CMU-MOSEI $\rightarrow$ CMU-MOSI, as shown in Fig.6. We select some typical samples from the target domain to explore how our model works. For each target sample, we find the closest sample from source domain by calculating cosine similarity between target sample and source sample. We observe that the closest source samples picked by our model OSAN are + +more relevant to the target samples. Specifically, take the target sample and source sample in the first row for example, with our model, speakers in both the source frame and target frame are very happy. Moreover, there is one common keyword in what the speakers said: enjoyed. Combining these two modalities of image and text, our model correctly classifies this target sample as positive sentiment with a score of 1.85, which is very close to the ground truth score of 2. In contrast, source sample found by the MDMN is not relevant to the target sample. There is no similar semantic information in what they said. The expressions in the frames are dissimilar. The phenomenon in these cases implies that our model can learn a common latent space in which the source domain and target domain are well aligned. + +# 6. Conclusion + +To address the problems of domain adaptation and modality alignment in multimodal domain adaptation, we propose a novel one-stage alignment network to unify these two problems in one stage. In our model, the TAL module is proposed to explore the relationship between domains and modalities, which can provide complementary information to each other. Moreover, the DDG module is designed to generate new soft samples that mix information from both the source and target domains, guiding our model in a more continuous space. Extensive experiments on two multimodal tasks with different extents of domain shift demonstrate the excellent performance of our model. + +# References + +[1] Shai Ben-David, John Blitzer, Koby Crammer, Alex Kulesza, Fernando Pereira, and Jennifer Wortman Vaughan. A theory of learning from different domains. Machine learning, 79(1):151-175, 2010. 2 +[2] Bent Fuglede and Flemming Topsoe. Jensen-shannon divergence and hilbert space embedding. In International Symposium on Information Theory, 2004. ISIT 2004. Proceedings., page 31. IEEE, 2004. 7 +[3] Yaroslav Ganin and Victor Lempitsky. Unsupervised domain adaptation by backpropagation. In International conference on machine learning, pages 1180-1189. PMLR, 2015. 1 +[4] Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario Marchand, and Victor Lempitsky. Domain-adversarial training of neural networks. The journal of machine learning research, 17(1):2096-2030, 2016. 2 +[5] Arthur Gretton, Dino Sejdinovic, Heiko Strathmann, Sivaraman Balakrishnan, Massimiliano Pontil, Kenji Fukumizu, and Bharath K Striperumbudur. Optimal kernel choice for large-scale two-sample tests. In Advances in neural information processing systems, pages 1205-1213. Citeseer, 2012. 3, 4 +[6] Wei Han, Hui Chen, and Soujanya Poria. Improving multimodal fusion with hierarchical mutual information maximization for multimodal sentiment analysis, 2021. 5 +[7] Devamanyu Hazarika, Roger Zimmermann, and Soujanya Poria. Misa: Modality-invariant and-specific representations for multimodal sentiment analysis. In Proceedings of the 28th ACM International Conference on Multimedia, pages 1122-1131, 2020. 5 +[8] Donghyun Kim, Yi-Hsuan Tsai, Bingbing Zhuang, Xiang Yu, Stan Sclaroff, Kate Saenko, and Manmohan Chandraker. Learning cross-modal contrastive features for video domain adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13618-13627, 2021. 1 +[9] Ye Liu, Changchong Lu, Chen Lin, Di Yin, and Bo Ren. Contrastive graph multimodal model for text classification in videos. arXiv preprint arXiv:2206.02343, 2022. 6 +[10] Ye Liu, Lingfeng Qiao, Di Yin, Zhuoxuan Jiang, Xinghua Jiang, Deqiang Jiang, and Bo Ren. Os-msl: One stage multimodal sequential link framework for scene segmentation and classification. In Proceedings of the 30th ACM International Conference on Multimedia, pages 6269-6277, 2022. 1 +[11] Zhun Liu, Ying Shen, Varun Bharadhwaj Lakshminarasimhan, Paul Pu Liang, Amir Zadeh, and Louis-Philippe Morency. Efficient low-rank multimodal fusion with modality-specific factors. arXiv preprint arXiv:1806.00064, 2018.3 +[12] Mingsheng Long, Yue Cao, Jianmin Wang, and Michael Jordan. Learning transferable features with deep adaptation networks. In International conference on machine learning, pages 97-105. PMLR, 2015. 5, 7 +[13] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representational + +tations for vision-and-language tasks. arXiv preprint arXiv:1908.02265, 2019. 1 +[14] Jonathan Munro and Dima Damen. Multi-modal domain adaptation for fine-grained action recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 122-132, 2020. 1, 2, 5 +[15] Jonathan Munro and Dima Damen. Multi-modal domain adaptation for fine-grained action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 5, 7 +[16] Sinno Jialin Pan, Ivor W Tsang, James T Kwok, and Qiang Yang. Domain adaptation via transfer component analysis. IEEE transactions on neural networks, 22(2):199-210, 2010. 2 +[17] Fan Qi, Xiaoshan Yang, and Changsheng Xu. A unified framework for multimodal domain adaptation. In Proceedings of the 26th ACM International Conference on Multimedia, MM '18, page 429-437, New York, NY, USA, 2018. Association for Computing Machinery. 1, 2 +[18] Lingfeng Qiao, Chen Wu, Ye Liu, Haoyuan Peng, Di Yin, and Bo Ren. Grafting pre-trained models for multimodal headline generation. arXiv preprint arXiv:2211.07210, 2022. 1 +[19] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 1 +[20] Wasifur Rahman, Md Kamrul Hasan, Sangwu Lee, Amir Zadeh, Chengfeng Mao, Louis-Philippe Morency, and Ehsan Hoque. Integrating multimodal information in large pretrained transformers. In Proceedings of the conference. Association for Computational Linguistics. Meeting, volume 2020, page 2359. NIH Public Access, 2020. 5 +[21] Baochen Sun and Kate Saenko. Deep coral: Correlation alignment for deep domain adaptation. In European conference on computer vision, pages 443-450. Springer, 2016. 3 +[22] Zhongkai Sun, Prathusha Sarma, William Sethares, and Yingyu Liang. Learning relationships between text, audio, and video via deep canonical correlation for multimodal language analysis. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 8992-8999, 2020. 5 +[23] Hui Tang and Kui Jia. Discriminative adversarial domain adaptation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 5940-5947, 2020. 2 +[24] Dacheng Tao, Xuelong Li, Xindong Wu, and Steve Maybank. Tensor rank one discriminant analysis-a convergent method for discriminative multilinear subspace selection. Neurocomputing, 71(10):1866-1882, 2008. 3 +[25] Dacheng Tao, Xuelong Li, Xindong Wu, and Stephen J Maybank. General tensor discriminant analysis and gabor features for gait recognition. IEEE transactions on pattern analysis and machine intelligence, 29(10):1700-1715, 2007. 4 +[26] Wilhelm Tranheden, Viktor Olsson, Juliano Pinto, and Lennart Svensson. Dacs: Domain adaptation via cross-domain mixed sampling. In Proceedings of the IEEE/CVF + +Winter Conference on Applications of Computer Vision, pages 1379-1389, 2021. 4 +[27] Eric Tzeng, Judy Hoffman, Kate Saenko, and Trevor Darrell. Adversarial discriminative domain adaptation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7167-7176, 2017. 4, 5, 7 +[28] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9(11), 2008. 7 +[29] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Ilia Polosukhin. Attention is all you need. arXiv preprint arXiv:1706.03762, 2017. 4 +[30] Wei Wang, Haojie Li, Zhengming Ding, Feiping Nie, Junyang Chen, Xiao Dong, and Zhihui Wang. Rethinking maximum mean discrepancy for visual domain adaptation. IEEE Transactions on Neural Networks and Learning Systems, 2021. 2 +[31] Garrett Wilson and Diane J Cook. A survey of unsupervised deep domain adaptation. ACM Transactions on Intelligent Systems and Technology (TIST), 11(5):1-46, 2020. 1 +[32] Xinyi Wu, Zhenyao Wu, Hao Guo, Lili Ju, and Song Wang. Darnet: A one-stage domain adaptation network for unsupervised nighttime semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15769-15778, 2021. 2 +[33] Minghao Xu, Jian Zhang, Bingbing Ni, Teng Li, Chengjie Wang, Qi Tian, and Wenjun Zhang. Adversarial domain adaptation with domain mixup. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 6502-6509, 2020. 4 +[34] Yiming Xu, Lin Chen, Zhongwei Cheng, Lixin Duan, and Jiebo Luo. Open-ended visual question answering by multimodal domain adaptation. arXiv preprint arXiv:1911.04058, 2019. 1, 2 +[35] Shen Yan, Huan Song, Nanxiang Li, Lincan Zou, and Liu Ren. Improve unsupervised domain adaptation with mixup training. arXiv preprint arXiv:2001.00677, 2020. 4 +[36] Wenmeng Yu, Hua Xu, Ziqi Yuan, and Jiele Wu. Learning modality-specific representations with self-supervised multi-task learning for multimodal sentiment analysis. arXiv preprint arXiv:2102.04830, 2021. 5 +[37] Amir Zadeh, Minghai Chen, Soujanya Poria, Erik Cambria, and Louis-Philippe Morency. Tensor fusion network for multimodal sentiment analysis. arXiv preprint arXiv:1707.07250, 2017. 3, 5 +[38] Amir Zadeh, Rowan Zellers, Eli Pincus, and Louis-Philippe Morency. Mosi: multimodal corpus of sentiment intensity and subjectivity analysis in online opinion videos. arXiv preprint arXiv:1606.06259, 2016. 5 +[39] AmirAli Bagher Zadeh, Paul Pu Liang, Soujanya Poria, Erik Cambria, and Louis-Philippe Morency. Multimodal language analysis in the wild: Cmu-mosei dataset and interpretable dynamic fusion graph. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2236–2246, 2018. 5 + +[40] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412, 2017. 4 +[41] Tong Zhang, Di Wang, Huanhuan Chen, Zhiwei Zeng, Wei Guo, Chunyan Miao, and Lizhen Cui. Bdann: Bert-based domain adaptation neural network for multi-modal fake news detection. In 2020 International Joint Conference on Neural Networks (IJCNN), pages 1-8, 2020. 1, 2 +[42] Weichen Zhang, Dong Xu, Jing Zhang, and Wanli Ouyang. Progressive modality cooperation for multi-modality domain adaptation. IEEE Transactions on Image Processing, 30:3293-3306, 2021. 2 +[43] Yiyi Zhang, Li Niu, Ziqi Pan, Meichao Luo, Jianfu Zhang, Dawei Cheng, and Liqing Zhang. Exploiting motion information from unlabeled videos for static image action recognition. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 12918-12925, 2020. 1 +[44] Honghao Zhou, Tinghuai Ma, Huan Rong, Yurong Qian, Yuan Tian, and Najla Al-Nabhan. Mdmn: Multi-task and domain adaptation based multi-modal network for early rumor detection. Expert Systems with Applications, 195:116517, 2022. 1, 2, 5, 7 +[45] Junbao Zhuo, Shuhui Wang, Weigang Zhang, and Qingming Huang. Deep unsupervised convolutional domain adaptation. In Proceedings of the 25th ACM international conference on Multimedia, pages 261-269, 2017. 2, 3 \ No newline at end of file diff --git a/2023/OSAN_ A One-Stage Alignment Network To Unify Multimodal Alignment and Unsupervised Domain Adaptation/images.zip b/2023/OSAN_ A One-Stage Alignment Network To Unify Multimodal Alignment and Unsupervised Domain Adaptation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..21a3ce6c8c0705928f5eff6364148828eee31261 --- /dev/null +++ b/2023/OSAN_ A One-Stage Alignment Network To Unify Multimodal Alignment and Unsupervised Domain Adaptation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e02c1f7f714547ebb40348fe5ea0f3b4b1b8f1f43eeed6f868a45caf01a2392e +size 779014 diff --git a/2023/OSAN_ A One-Stage Alignment Network To Unify Multimodal Alignment and Unsupervised Domain Adaptation/layout.json b/2023/OSAN_ A One-Stage Alignment Network To Unify Multimodal Alignment and Unsupervised Domain Adaptation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4ba19125647fb7c92c1fc2bf0692146064803107 --- /dev/null +++ b/2023/OSAN_ A One-Stage Alignment Network To Unify Multimodal Alignment and Unsupervised Domain Adaptation/layout.json @@ -0,0 +1,8281 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 58, + 102, + 536, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 102, + 536, + 140 + ], + "spans": [ + { + "bbox": [ + 58, + 102, + 536, + 140 + ], + "type": "text", + "content": "OSAN: A One-Stage Alignment Network to Unify Multimodal Alignment and Unsupervised Domain Adaptation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 95, + 161, + 499, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 161, + 499, + 188 + ], + "spans": [ + { + "bbox": [ + 95, + 161, + 499, + 188 + ], + "type": "text", + "content": "Ye Liu, Lingfeng Qiao, Changchong Lu, Di Yin, Chen Lin, Haoyuan Peng, Bo Ren \nTencent Youtu Lab" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 94, + 191, + 493, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 191, + 493, + 204 + ], + "spans": [ + { + "bbox": [ + 94, + 191, + 493, + 204 + ], + "type": "text", + "content": "{rafelliu, leafqiao, chonglu, endymecyyin, iwlliamlin, haoyuanpeng, timren}@tencent.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "spans": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 258, + 290, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 258, + 290, + 547 + ], + "spans": [ + { + "bbox": [ + 46, + 258, + 290, + 547 + ], + "type": "text", + "content": "Extending from unimodal to multimodal is a critical challenge for unsupervised domain adaptation (UDA). Two major problems emerge in unsupervised multimodal domain adaptation: domain adaptation and modality alignment. An intuitive way to handle these two problems is to fulfill these tasks in two separate stages: aligning modalities followed by domain adaptation, or vice versa. However, domains and modalities are not associated in most existing two-stage studies, and the relationship between them is not leveraged which can provide complementary information to each other. In this paper, we unify these two stages into one to align domains and modalities simultaneously. In our model, a tensor-based alignment module (TAL) is presented to explore the relationship between domains and modalities. By this means, domains and modalities can interact sufficiently and guide them to utilize complementary information for better results. Furthermore, to establish a bridge between domains, a dynamic domain generator (DDG) module is proposed to build transitional samples by mixing the shared information of two domains in a self-supervised manner, which helps our model learn a domain-invariant common representation space. Extensive experiments prove that our method can achieve superior performance in two real-world applications. The code will be publicly available." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 572, + 128, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 572, + 128, + 585 + ], + "spans": [ + { + "bbox": [ + 47, + 572, + 128, + 585 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "content": "With explosively emerging multimedia data on the Internet, the field of multimodal analysis achieves more and more attention [10, 13, 18, 19, 43]. Compared to extensive unimodal models in NLP and CV, learning adequate knowledge from multimodal signals is still preliminary but very important. Abundant data plays a key role in different scenarios of multimodal analysis, such as pre-training or downstream multimedia tasks. However, it is prohibitively expensive and time-consuming to obtain large amounts of labeled data. To eliminate this issue, domain adaptation (DA)" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 314, + 232, + 545, + 331 + ], + "blocks": [ + { + "bbox": [ + 314, + 232, + 545, + 331 + ], + "lines": [ + { + "bbox": [ + 314, + 232, + 545, + 331 + ], + "spans": [ + { + "bbox": [ + 314, + 232, + 545, + 331 + ], + "type": "image", + "image_path": "884fcf7f5421700cb5403e4a222adf854fe52088a0f25c74bd6263691899488a.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 341, + 340, + 509, + 352 + ], + "lines": [ + { + "bbox": [ + 341, + 340, + 509, + 352 + ], + "spans": [ + { + "bbox": [ + 341, + 340, + 509, + 352 + ], + "type": "text", + "content": "Figure 1. Conception of our one-stage model." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 365, + 546, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 365, + 546, + 498 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 546, + 498 + ], + "type": "text", + "content": "is raised to learn a model from a labeled dataset (source domain) that can be generalized to other related tasks without sufficient labeled data (target domain) [3]. Classical domain adaptation can be classified into different categories: unsupervised domain adaptation (UDA), fully supervised domain adaptation, and semi-supervised domain adaptation [31]. In this paper, we focus on UDA where no samples in target domain are annotated. With this technique, it is not necessary to prepare a customized training dataset for a specific task, but it can perform the task effectively and efficiently." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 498, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 547, + 715 + ], + "type": "text", + "content": "There are two challenges when applying domain adaptation to multimodal scenarios [17]: (1) how to align the source and target domains and remit domain discrepancy, and (2) how to align multiple modalities and leverage multimodal information. Most existing works address these two problems in two consecutive stages: multimodal alignment followed by domain adaptation [34, 41], or vice versa [14, 44]. However, they solve these two issues separately without considering their relationship: domain and modality can be treated as two views to portray the intrinsic characteristic of multimodal data [8], and the hidden underlying relationship in these two views can provide complementary information to each other. Unimodal domain adaptation methods can not work well in multimodal tasks due to the inability to preserve the relations between modalities at the same time. Through our experiments and analysis, we observe that the two-stage model could not achieve ideal performance. Fig.2 shows the learning curve of two-stage" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 314, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 758 + ], + "type": "text", + "content": "3551" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": "model during training phase by 800 iterations for the task of multimodal sentiment analysis. It can be found that the learning curve of two-stage model is oscillating and converges slowly, which indicates that two-stage model is probably not a superior solution. To handle these challenges, the objective of multimodal domain adaptation can be defined as: (1) Exploring the relationship between domains and modalities; (2) Finding a common domain-invariant cross-modal representation space to align domains and modalities simultaneously." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 192, + 289, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 192, + 289, + 360 + ], + "spans": [ + { + "bbox": [ + 46, + 192, + 289, + 360 + ], + "type": "text", + "content": "Therefore, in this paper, we design a One-Stage Alignment Network (OSAN) to unify multimodal alignment and domain adaptation in one stage. Fig.1 shows the conception of our one-stage model. Our method benefits from: (1) The modality and domain are associated and interacted to capture the relationship between domains and modalities, which can provide rich complementary information to each other. (2) Multimodal alignment and domain adaptation are unified in one stage, which allows our model to perform domain adaptation and leverage multimodal information at the same time. In Fig.2, we observe that the learning curve of our method is relatively stable and converges better, which indicates that exploring the relation between modality and domain contributes to our task." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 59, + 360, + 244, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 360, + 244, + 371 + ], + "spans": [ + { + "bbox": [ + 59, + 360, + 244, + 371 + ], + "type": "text", + "content": "In summary, our contributions are as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 371, + 287, + 551 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 46, + 371, + 287, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 371, + 287, + 431 + ], + "spans": [ + { + "bbox": [ + 46, + 371, + 287, + 431 + ], + "type": "text", + "content": "(1) To capture the relationship between domain and modality, we propose a one-stage alignment network, called OSAN, to associate domain and modality. In this way, a joint domain-invariant and cross-modal representation space is learned in one stage." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 431, + 287, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 431, + 287, + 466 + ], + "spans": [ + { + "bbox": [ + 46, + 431, + 287, + 466 + ], + "type": "text", + "content": "(2) We design a TAL module to bring sufficient interactions between domains and modalities and guide them to utilize complementary information for each other." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 467, + 287, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 467, + 287, + 515 + ], + "spans": [ + { + "bbox": [ + 46, + 467, + 287, + 515 + ], + "type": "text", + "content": "(3) To effectively bridge distinct domains, a DDG module is developed to dynamically construct multiple new domains by combining knowledge of source and target domains and exploring intrinsic structure of data distribution." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 515, + 287, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 515, + 287, + 551 + ], + "spans": [ + { + "bbox": [ + 46, + 515, + 287, + 551 + ], + "type": "text", + "content": "(4) Extensive experiments on two totally different tasks demonstrate the effectiveness of our method compared to the supervised and strongly UDA methods." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 562, + 134, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 562, + 134, + 574 + ], + "spans": [ + { + "bbox": [ + 47, + 562, + 134, + 574 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": "Unimodal Domain Adaption The main approaches for domain adaptation or domain alignment can be categorized as discrepancy-based and adversarial-based methods [1, 23, 30]. Discrepancy-based methods design statistics, such as maximum mean discrepancy (MMD) [16], correlation alignment (CORAL) [45] and etc., to measure the difference between two domain distributions. For adversarial-based methods, Ganin et al. [4] used the concept of generative adversarial networks (GAN) to obtain domain-invariant characteristics. Wu et al. [32] proposed a one-stage adaptation framework for nighttime semantic segmentation to per" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 314, + 74, + 540, + 222 + ], + "blocks": [ + { + "bbox": [ + 314, + 74, + 540, + 222 + ], + "lines": [ + { + "bbox": [ + 314, + 74, + 540, + 222 + ], + "spans": [ + { + "bbox": [ + 314, + 74, + 540, + 222 + ], + "type": "image", + "image_path": "eddb63d74bc5cbe37d8b01f66f52ee8d5afa958fd9aa4be022d75dfe9406e75e.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 231, + 545, + 255 + ], + "lines": [ + { + "bbox": [ + 305, + 231, + 545, + 255 + ], + "spans": [ + { + "bbox": [ + 305, + 231, + 545, + 255 + ], + "type": "text", + "content": "Figure 2. The learning curves of our one-stage method and traditional two-stage method in multimodal sentiment analysis task." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 274, + 547, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 274, + 547, + 335 + ], + "spans": [ + { + "bbox": [ + 304, + 274, + 547, + 335 + ], + "type": "text", + "content": "form an end-to-end learning where a separate common preprocessing step was integrated, which is different from our motivation. However, these studies only focus on the elimination of domain shift in unimodal scenarios, which may lose effect when applied in multimodal area." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 347, + 547, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 347, + 547, + 588 + ], + "spans": [ + { + "bbox": [ + 304, + 347, + 547, + 588 + ], + "type": "text", + "content": "Multimodal Domain Adaption Qi et al. [17] first analyzed the multimodal domain adaptation and raised two problems, i.e., domain adaptation and modality alignment. To alleviate them, hybrid domain constraints and attention-based modality fusion module were introduced to learn domain-invariant fused features. More recently, some studies [14, 34, 41, 42, 44] related to this area have tried to address the multimodal problem by extending the unimodal domain adaptation methods. Xu et al. [34] proposed a supervised multimodal domain adaptation method for VQA to learn joint feature embeddings across different domains and modalities. Munro et al. [14] exploited the correspondence of modalities as a self-supervised alignment approach for UDA in addition to adversarial alignment. Furthermore, Zhou et al. [44] proposed MDMN for early rumor detection, which can combine textual and visual information with two heterogeneous feature extractors. However, all of these studies solved the two problems separately in two stages. Moreover, the relationship between modality and domain is missing, thus they cannot achieve ideal performance." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 597, + 411, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 597, + 411, + 611 + ], + "spans": [ + { + "bbox": [ + 306, + 597, + 411, + 611 + ], + "type": "text", + "content": "3. Proposed Method" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 547, + 715 + ], + "type": "text", + "content": "Fig.3 presents an overview of our method that contains four parts: multimodal feature extraction, tensor-based alignment, dynamic domain generator and task-specific heads. First, encoders, as multimodal feature extractor, map source and target multimodal data to different latent spaces. Then, by means of an effective and efficient way of tensor representation, we propose a TAL module to acquire relation information through consecutive interactions between" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3552" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 76, + 537, + 342 + ], + "blocks": [ + { + "bbox": [ + 55, + 76, + 537, + 342 + ], + "lines": [ + { + "bbox": [ + 55, + 76, + 537, + 342 + ], + "spans": [ + { + "bbox": [ + 55, + 76, + 537, + 342 + ], + "type": "image", + "image_path": "58c94c1fa496aeb949ff7c4f381a7d8ad55528349d45db9521c07bd93c96f56c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 359, + 546, + 383 + ], + "lines": [ + { + "bbox": [ + 46, + 359, + 546, + 383 + ], + "spans": [ + { + "bbox": [ + 46, + 359, + 546, + 383 + ], + "type": "text", + "content": "Figure 3. Overview of our method OSAN. We use circle, square and triangle to represent three modalities of text, vision and audio. In the figure, " + }, + { + "bbox": [ + 46, + 359, + 546, + 383 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 46, + 359, + 546, + 383 + ], + "type": "text", + "content": " means mode- " + }, + { + "bbox": [ + 46, + 359, + 546, + 383 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 359, + 546, + 383 + ], + "type": "text", + "content": " product of a tensor and a matrix, " + }, + { + "bbox": [ + 46, + 359, + 546, + 383 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 46, + 359, + 546, + 383 + ], + "type": "text", + "content": " represents element-wise product, and " + }, + { + "bbox": [ + 46, + 359, + 546, + 383 + ], + "type": "inline_equation", + "content": "\\ast" + }, + { + "bbox": [ + 46, + 359, + 546, + 383 + ], + "type": "text", + "content": " means tensor factorization." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 393, + 288, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 393, + 288, + 465 + ], + "spans": [ + { + "bbox": [ + 46, + 393, + 288, + 465 + ], + "type": "text", + "content": "modalities and domains. Thus, the modality and domain are aligned simultaneously in one stage. After that, we develop a DDG module to mix information from different domains to create new transitional domains. Finally, our network is split into three branches: category classification, domain adversarial learning and discrepancy elimination." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 477, + 185, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 477, + 185, + 490 + ], + "spans": [ + { + "bbox": [ + 47, + 477, + 185, + 490 + ], + "type": "text", + "content": "3.1. Tensor-based Alignment" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 497, + 287, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 497, + 287, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 497, + 287, + 640 + ], + "type": "text", + "content": "We develop a TAL module to simultaneously address the challenges of multimodal alignment and domain adaptation. We first resort to tensor representation, an effective and natural representation of multimodal data, to develop models capturing inter-modality dynamics [11, 37]. Then, we utilize statistical measurements to establish an interaction between domains and modalities to explore the relation information. In this way, we associate domains and modalities so that they can communicate and align together. Specifically, we perform tensor factorization to fuse multimodal information and maximize statistical measurements to simultaneously eliminate domain gaps." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 639, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 639, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 639, + 288, + 714 + ], + "type": "text", + "content": "Specifically, we use " + }, + { + "bbox": [ + 46, + 639, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\underline{\\mathbf{X}}^{I_1\\times \\dots \\times I_N\\times M}" + }, + { + "bbox": [ + 46, + 639, + 288, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 639, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\underline{\\mathbf{Y}}^{J_1\\times \\dots \\times J_N\\times M}" + }, + { + "bbox": [ + 46, + 639, + 288, + 714 + ], + "type": "text", + "content": " to represent source and target dataset with " + }, + { + "bbox": [ + 46, + 639, + 288, + 714 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 639, + 288, + 714 + ], + "type": "text", + "content": " modalities and " + }, + { + "bbox": [ + 46, + 639, + 288, + 714 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 639, + 288, + 714 + ], + "type": "text", + "content": " samples. For example, a video can be represented by a third-order tensor using a differentiable outer product between visual, textual, and acoustic representations [37]. For simplicity, we first make the training samples zero mean." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 393, + 546, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 393, + 546, + 538 + ], + "spans": [ + { + "bbox": [ + 304, + 393, + 546, + 538 + ], + "type": "text", + "content": "To perform multimodal alignment, TAL aims to find pairs of linear transformations " + }, + { + "bbox": [ + 304, + 393, + 546, + 538 + ], + "type": "inline_equation", + "content": "\\mathbf{U}^{(n)}|_{n = 1}^{N}" + }, + { + "bbox": [ + 304, + 393, + 546, + 538 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 393, + 546, + 538 + ], + "type": "inline_equation", + "content": "\\mathbf{V}^{(n)}|_{n = 1}^{N}" + }, + { + "bbox": [ + 304, + 393, + 546, + 538 + ], + "type": "text", + "content": " for each modality of source and target domains to project samples of two sets into low dimensional subspaces. During this process, we establish an interaction between domain and modality by maximizing a statistical measurement of covariance given a normalized standard deviation. As reported in [45], we choose covariance instead of the first-order mean statistics used in MK-MMD [5, 21], as it is more powerful and achieves better performance. By this means, domains and modalities can be aligned simultaneously in one stage. Thus the objective function is defined as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 544, + 506, + 594 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 506, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 506, + 594 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\left\\{\\mathbf {U} ^ {(n)} | _ {n = 1} ^ {N}, \\mathbf {V} ^ {(n)} | _ {n = 1} ^ {N} \\right\\} \\\\ = \\max \\llbracket (\\underline {{\\mathbf {X}}} \\prod_ {n = 1} ^ {N} \\times_ {n} \\mathbf {U} ^ {(n) ^ {T}}) \\otimes (\\underline {{\\mathbf {Y}}} \\prod_ {n = 1} ^ {N} \\times_ {n} \\mathbf {V} ^ {(n) ^ {T}}); \\rrbracket \\\\ \\end{array}", + "image_path": "ed73b5d8e3a45dcefd076573846dc98bcb1c4277480db585b651473deafa62cb.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 312, + 596, + 545, + 658 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 596, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 312, + 596, + 545, + 658 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} s. t. \\left(\\underline {{\\mathbf {X}}} \\prod_ {n = 1} ^ {N} \\times_ {n} \\mathbf {U} ^ {(n) ^ {T}}\\right) _ {(N + 1)} ^ {T} \\left(\\underline {{\\mathbf {X}}} \\prod_ {n = 1} ^ {N} \\times_ {n} \\mathbf {U} ^ {(n) ^ {T}}\\right) _ {(N + 1)} = \\mathbf {I} \\tag {1} \\\\ \\left(\\underline {{\\mathbf {Y}}} \\prod_ {n = 1} ^ {N} \\times_ {n} \\mathbf {V} ^ {(n) ^ {T}}\\right) _ {(N + 1)} ^ {T} \\left(\\underline {{\\mathbf {Y}}} \\prod_ {n = 1} ^ {N} \\times_ {n} \\mathbf {V} ^ {(n) ^ {T}}\\right) _ {(N + 1)} = \\mathbf {I} \\\\ \\end{array}", + "image_path": "d48c8924c7b8177059b87b308203cc7fe460000e20d0da9e38734472f33d7437.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 665, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 665, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 665, + 546, + 714 + ], + "type": "text", + "content": "The objective function above does not have a closed-form solution, so instead we derive a suboptimal solution following the principle of the alternating projection method [24], where the complicated optimization problem is re" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3553" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 118 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 118 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 118 + ], + "type": "text", + "content": "duced into smaller conditional subproblems that can be solved by simple established methods. Therefore, the objective function is decomposed into " + }, + { + "bbox": [ + 46, + 72, + 287, + 118 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 72, + 287, + 118 + ], + "type": "text", + "content": " different subproblems:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 77, + 126, + 287, + 148 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 126, + 287, + 148 + ], + "spans": [ + { + "bbox": [ + 77, + 126, + 287, + 148 + ], + "type": "interline_equation", + "content": "\\left\\{\\mathbf {U} ^ {(n)}, \\mathbf {V} ^ {(n)} \\right\\} = \\max t r \\left\\{\\mathbf {U} ^ {(n) ^ {T}} \\mathbf {C} _ {x y} ^ {(n)} \\mathbf {V} ^ {(n)} \\right\\} \\tag {2}", + "image_path": "322666e3fb358099e5b9459372340e623bfc11e1302e3d2f3ce290ed1a2a57d1.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 76, + 148, + 260, + 162 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 148, + 260, + 162 + ], + "spans": [ + { + "bbox": [ + 76, + 148, + 260, + 162 + ], + "type": "interline_equation", + "content": "s. t. \\mathbf {U} ^ {(n) ^ {T}} \\mathbf {C} _ {x x} ^ {(n)} \\mathbf {U} ^ {(n)} = \\mathbf {I}, \\mathbf {V} ^ {(n) ^ {T}} \\mathbf {C} _ {y y} ^ {(n)} \\mathbf {V} ^ {(n)} = \\mathbf {I}", + "image_path": "70a51a3bf247bde0283f7362149e27276ed1da3e7dbf25f6da28e4da13151f42.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 172, + 251, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 172, + 251, + 187 + ], + "spans": [ + { + "bbox": [ + 47, + 172, + 251, + 187 + ], + "type": "text", + "content": "where, for simplification, we define " + }, + { + "bbox": [ + 47, + 172, + 251, + 187 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_{xy}^{(n)}" + }, + { + "bbox": [ + 47, + 172, + 251, + 187 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 172, + 251, + 187 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_{xx}^{(n)}" + }, + { + "bbox": [ + 47, + 172, + 251, + 187 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 94, + 195, + 287, + 220 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 195, + 287, + 220 + ], + "spans": [ + { + "bbox": [ + 94, + 195, + 287, + 220 + ], + "type": "interline_equation", + "content": "\\mathbf {C} _ {x y} ^ {(n)} = \\left(\\llbracket (\\underline {{\\mathbf {X}}} ^ {(\\bar {n})}) \\otimes (\\underline {{\\mathbf {Y}}} ^ {(\\bar {n})}); (\\bar {n}) (\\bar {n}) \\rrbracket\\right) \\tag {3}", + "image_path": "047c81f259e1ec9c72d0458e6ee16e596ba44b9dbb9d42544e152c1f0ed70bee.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 94, + 217, + 238, + 234 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 217, + 238, + 234 + ], + "spans": [ + { + "bbox": [ + 94, + 217, + 238, + 234 + ], + "type": "interline_equation", + "content": "\\mathbf {C} _ {x x} ^ {(n)} = \\left(\\llbracket (\\underline {{\\mathbf {X}}} ^ {(\\bar {n})}) \\otimes (\\underline {{\\mathbf {X}}} ^ {(\\bar {n})}); (\\bar {n}) (\\bar {n}) \\rrbracket\\right)", + "image_path": "df682a2d9d8e36fedcb1ea69535d8d4b600499c3901c61a8db4b949aac3902cb.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 243, + 287, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 243, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 47, + 243, + 287, + 304 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 243, + 287, + 304 + ], + "type": "inline_equation", + "content": "([\\cdot \\otimes ;(\\bar{n})(\\bar{n})])" + }, + { + "bbox": [ + 47, + 243, + 287, + 304 + ], + "type": "text", + "content": " means tensor contraction on all indices except the " + }, + { + "bbox": [ + 47, + 243, + 287, + 304 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 47, + 243, + 287, + 304 + ], + "type": "text", + "content": "-th index on the tensor product [25]. Similarly, " + }, + { + "bbox": [ + 47, + 243, + 287, + 304 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_{yy}^{(n)}" + }, + { + "bbox": [ + 47, + 243, + 287, + 304 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 243, + 287, + 304 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_{yx}^{(n)}" + }, + { + "bbox": [ + 47, + 243, + 287, + 304 + ], + "type": "text", + "content": " is same as " + }, + { + "bbox": [ + 47, + 243, + 287, + 304 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_{xx}^{(n)}" + }, + { + "bbox": [ + 47, + 243, + 287, + 304 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 243, + 287, + 304 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_{xy}^{(n)}" + }, + { + "bbox": [ + 47, + 243, + 287, + 304 + ], + "type": "text", + "content": " in form. By Eq.2, an interaction between source and target domains for the " + }, + { + "bbox": [ + 47, + 243, + 287, + 304 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 47, + 243, + 287, + 304 + ], + "type": "text", + "content": "-th modality is established." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 312, + 203, + 324 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 312, + 203, + 324 + ], + "spans": [ + { + "bbox": [ + 47, + 312, + 203, + 324 + ], + "type": "text", + "content": "3.2. Dynamic Domain Generator" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 331, + 287, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 331, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 46, + 331, + 287, + 437 + ], + "type": "text", + "content": "In domain adaptation, adversarial learning is often used to align domains by developing a domain discriminator to judge samples from the source or target domains [27, 33]. By this means, it brings a two-class classification task with hard assignments of " + }, + { + "bbox": [ + 46, + 331, + 287, + 437 + ], + "type": "inline_equation", + "content": "\\mathbf{1}" + }, + { + "bbox": [ + 46, + 331, + 287, + 437 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 46, + 331, + 287, + 437 + ], + "type": "inline_equation", + "content": "\\mathbf{0}" + }, + { + "bbox": [ + 46, + 331, + 287, + 437 + ], + "type": "text", + "content": ". They try to confuse the domain discriminator to learn a shared common representation space between the source and target domains. However, it usually achieves worse performance due to a huge variance between the two domains." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 438, + 287, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 438, + 287, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 438, + 287, + 593 + ], + "type": "text", + "content": "To overcome this problem, several works have been proposed to produce some transitional domains based on Mixup [35, 40]. They mixed images from the source and target domains to generate more soft and mixed samples [26, 33]. Suppose there are one sample " + }, + { + "bbox": [ + 46, + 438, + 287, + 593 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 438, + 287, + 593 + ], + "type": "text", + "content": " from source domain and another sample " + }, + { + "bbox": [ + 46, + 438, + 287, + 593 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 438, + 287, + 593 + ], + "type": "text", + "content": " from target domain, " + }, + { + "bbox": [ + 46, + 438, + 287, + 593 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 438, + 287, + 593 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 438, + 287, + 593 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 438, + 287, + 593 + ], + "type": "text", + "content": " are linearly interpolated [40] to generate a new sample. This strategy can fully utilize the inter-domain information and improve the generalization ability of models. However, it has an obvious shortcoming: the generated sample fuses full information of two domains, which contains redundant information and speciality of domains that may induce undesirable oscillations on training the domain discriminator." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 594, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 677 + ], + "type": "text", + "content": "To solve this issue, we propose a DDG module to establish a bridge between domains. Specifically, DDG explicitly captures commonality and abandons the specialty of domains. By highlighting this commonality, we make the domain discriminator focus on commonality rather than full information, which helps our model learn a domain-invariant common representation space." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "Given one sample " + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": " from source domain and another sample " + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": " from target domain, DDG is proposed to dynamically generate soft samples " + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "s^{\\prime}" + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "inline_equation", + "content": "t^\\prime" + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": " to smoothly bridge the" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 72, + 545, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 346 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 346 + ], + "type": "text", + "content": "domain gap. Specifically, " + }, + { + "bbox": [ + 304, + 72, + 545, + 346 + ], + "type": "inline_equation", + "content": "s'" + }, + { + "bbox": [ + 304, + 72, + 545, + 346 + ], + "type": "text", + "content": " represents a variety of " + }, + { + "bbox": [ + 304, + 72, + 545, + 346 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 72, + 545, + 346 + ], + "type": "text", + "content": " which maintains information of " + }, + { + "bbox": [ + 304, + 72, + 545, + 346 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 72, + 545, + 346 + ], + "type": "text", + "content": " and mix source-relevant commonality from " + }, + { + "bbox": [ + 304, + 72, + 545, + 346 + ], + "type": "inline_equation", + "content": "t'" + }, + { + "bbox": [ + 304, + 72, + 545, + 346 + ], + "type": "text", + "content": ". Similarly, " + }, + { + "bbox": [ + 304, + 72, + 545, + 346 + ], + "type": "inline_equation", + "content": "t'" + }, + { + "bbox": [ + 304, + 72, + 545, + 346 + ], + "type": "text", + "content": " is a variety of " + }, + { + "bbox": [ + 304, + 72, + 545, + 346 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 72, + 545, + 346 + ], + "type": "text", + "content": ". As shown in Fig.3, we capture the commonality between " + }, + { + "bbox": [ + 304, + 72, + 545, + 346 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 72, + 545, + 346 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 346 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 72, + 545, + 346 + ], + "type": "text", + "content": " with the help of the popular and effective self-attention mechanism used in Transformers [29]. We construct a sequence in which the length is 2 by stacking " + }, + { + "bbox": [ + 304, + 72, + 545, + 346 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 72, + 545, + 346 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 346 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 72, + 545, + 346 + ], + "type": "text", + "content": ". Denote query " + }, + { + "bbox": [ + 304, + 72, + 545, + 346 + ], + "type": "inline_equation", + "content": "Q_{s}" + }, + { + "bbox": [ + 304, + 72, + 545, + 346 + ], + "type": "text", + "content": " and key " + }, + { + "bbox": [ + 304, + 72, + 545, + 346 + ], + "type": "inline_equation", + "content": "K_{s}" + }, + { + "bbox": [ + 304, + 72, + 545, + 346 + ], + "type": "text", + "content": " is projected from the source embedding " + }, + { + "bbox": [ + 304, + 72, + 545, + 346 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 72, + 545, + 346 + ], + "type": "text", + "content": " respectively, while query " + }, + { + "bbox": [ + 304, + 72, + 545, + 346 + ], + "type": "inline_equation", + "content": "Q_{t}" + }, + { + "bbox": [ + 304, + 72, + 545, + 346 + ], + "type": "text", + "content": " and key " + }, + { + "bbox": [ + 304, + 72, + 545, + 346 + ], + "type": "inline_equation", + "content": "K_{t}" + }, + { + "bbox": [ + 304, + 72, + 545, + 346 + ], + "type": "text", + "content": " are from the target embedding " + }, + { + "bbox": [ + 304, + 72, + 545, + 346 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 72, + 545, + 346 + ], + "type": "text", + "content": ", respectively. With dot-product between queries and keys, the source-target attention matrix " + }, + { + "bbox": [ + 304, + 72, + 545, + 346 + ], + "type": "inline_equation", + "content": "M_{st}" + }, + { + "bbox": [ + 304, + 72, + 545, + 346 + ], + "type": "text", + "content": " is obtained, which represents the relations between source and target. Afterwards, we design a gating to highlight the commonality between " + }, + { + "bbox": [ + 304, + 72, + 545, + 346 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 72, + 545, + 346 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 346 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 72, + 545, + 346 + ], + "type": "text", + "content": ", and fuse the commonality to generate " + }, + { + "bbox": [ + 304, + 72, + 545, + 346 + ], + "type": "inline_equation", + "content": "s'" + }, + { + "bbox": [ + 304, + 72, + 545, + 346 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 72, + 545, + 346 + ], + "type": "inline_equation", + "content": "t'" + }, + { + "bbox": [ + 304, + 72, + 545, + 346 + ], + "type": "text", + "content": " respectively. In detail, to generate " + }, + { + "bbox": [ + 304, + 72, + 545, + 346 + ], + "type": "inline_equation", + "content": "s'" + }, + { + "bbox": [ + 304, + 72, + 545, + 346 + ], + "type": "text", + "content": ", we calculate the feature-relevant distribution " + }, + { + "bbox": [ + 304, + 72, + 545, + 346 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 72, + 545, + 346 + ], + "type": "text", + "content": " by summing " + }, + { + "bbox": [ + 304, + 72, + 545, + 346 + ], + "type": "inline_equation", + "content": "M_{st}" + }, + { + "bbox": [ + 304, + 72, + 545, + 346 + ], + "type": "text", + "content": " along the dimension of the query. Larger value in " + }, + { + "bbox": [ + 304, + 72, + 545, + 346 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 72, + 545, + 346 + ], + "type": "text", + "content": " means related element in " + }, + { + "bbox": [ + 304, + 72, + 545, + 346 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 72, + 545, + 346 + ], + "type": "text", + "content": " is more-relevant to " + }, + { + "bbox": [ + 304, + 72, + 545, + 346 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 72, + 545, + 346 + ], + "type": "text", + "content": ", which should be chosen for commonality. As a result, the commonality " + }, + { + "bbox": [ + 304, + 72, + 545, + 346 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 72, + 545, + 346 + ], + "type": "text", + "content": " is obtained by " + }, + { + "bbox": [ + 304, + 72, + 545, + 346 + ], + "type": "inline_equation", + "content": "c = t \\odot \\text{Norm}(p)" + }, + { + "bbox": [ + 304, + 72, + 545, + 346 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 72, + 545, + 346 + ], + "type": "inline_equation", + "content": "\\text{Norm}(\\cdot)" + }, + { + "bbox": [ + 304, + 72, + 545, + 346 + ], + "type": "text", + "content": " means the normalized operator and " + }, + { + "bbox": [ + 304, + 72, + 545, + 346 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 304, + 72, + 545, + 346 + ], + "type": "text", + "content": " is the element-wise multiplication. Finally, " + }, + { + "bbox": [ + 304, + 72, + 545, + 346 + ], + "type": "inline_equation", + "content": "s'" + }, + { + "bbox": [ + 304, + 72, + 545, + 346 + ], + "type": "text", + "content": " is obtained by fusing commonality " + }, + { + "bbox": [ + 304, + 72, + 545, + 346 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 72, + 545, + 346 + ], + "type": "text", + "content": " and information from " + }, + { + "bbox": [ + 304, + 72, + 545, + 346 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 72, + 545, + 346 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 384, + 353, + 545, + 366 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 353, + 545, + 366 + ], + "spans": [ + { + "bbox": [ + 384, + 353, + 545, + 366 + ], + "type": "interline_equation", + "content": "s ^ {\\prime} = \\lambda s + (1 - \\lambda) c \\tag {4}", + "image_path": "18341c72e43093d7458b8dbfa6d77923c74f97fe8bb66571bdbfe805a9415fb1.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 372, + 545, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 372, + 545, + 467 + ], + "spans": [ + { + "bbox": [ + 304, + 372, + 545, + 467 + ], + "type": "text", + "content": "Finally, we use multi-head attention to generate various new domains to enrich the sample space. Samples from new domains and raw source and target domains are fed to domain discriminator, by which the domain discriminator is guided by the hard label information and well-designed soft domains. Each sample from these soft domains explores the intrinsic structure of data distribution from raw domains and enriches feature patterns by the interaction of two domains." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 473, + 354, + 485 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 473, + 354, + 485 + ], + "spans": [ + { + "bbox": [ + 306, + 473, + 354, + 485 + ], + "type": "text", + "content": "3.3. Tasks" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 491, + 545, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 491, + 545, + 576 + ], + "spans": [ + { + "bbox": [ + 304, + 491, + 545, + 576 + ], + "type": "text", + "content": "Our model consists of three tasks: category classification, domain adversarial learning, and domain discrepancy elimination. First, we perform " + }, + { + "bbox": [ + 304, + 491, + 545, + 576 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 491, + 545, + 576 + ], + "type": "text", + "content": "-way object classification on the source domain. Second, a domain discriminator is introduced to judge samples from source, target or the created soft domains. Finally, we try to learn transferable features by minimizing domain discrepancy. In summary, we have:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 367, + 582, + 545, + 594 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 582, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 367, + 582, + 545, + 594 + ], + "type": "interline_equation", + "content": "L = L _ {c l s} + \\alpha L _ {a d v} + \\beta L _ {d i s} \\tag {5}", + "image_path": "d25cb9030ae1979ae122d21c556337153d2362a4c2666bd48bcdab47f5942972.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 600, + 545, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 600, + 545, + 661 + ], + "spans": [ + { + "bbox": [ + 304, + 600, + 545, + 661 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 600, + 545, + 661 + ], + "type": "inline_equation", + "content": "L_{cls}" + }, + { + "bbox": [ + 304, + 600, + 545, + 661 + ], + "type": "text", + "content": " is the category classification loss, " + }, + { + "bbox": [ + 304, + 600, + 545, + 661 + ], + "type": "inline_equation", + "content": "L_{adv}" + }, + { + "bbox": [ + 304, + 600, + 545, + 661 + ], + "type": "text", + "content": " is the adversarial loss, and " + }, + { + "bbox": [ + 304, + 600, + 545, + 661 + ], + "type": "inline_equation", + "content": "L_{dis}" + }, + { + "bbox": [ + 304, + 600, + 545, + 661 + ], + "type": "text", + "content": " is the discrepancy loss. " + }, + { + "bbox": [ + 304, + 600, + 545, + 661 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 600, + 545, + 661 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 600, + 545, + 661 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 600, + 545, + 661 + ], + "type": "text", + "content": " are weights to balance these three losses. Cross entropy loss is used as classification and adversarial loss, while MK-MMD [5] is employed as discrepancy loss." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 669, + 387, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 669, + 387, + 682 + ], + "spans": [ + { + "bbox": [ + 306, + 669, + 387, + 682 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "We conduct experiments on two totally different tasks to demonstrate the efficacy of our model. One task is popu" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3554" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 49, + 102, + 544, + 317 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 545, + 93 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 545, + 93 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 545, + 93 + ], + "type": "text", + "content": "Table 1. Multimodal sentiment analysis results on CMU-MOSI and CMU-MOSEI. †: results come from [7]; ‡: results come from [36]; ◇: results come from [6]; ↓: the lower the better." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 102, + 544, + 317 + ], + "lines": [ + { + "bbox": [ + 49, + 102, + 544, + 317 + ], + "spans": [ + { + "bbox": [ + 49, + 102, + 544, + 317 + ], + "type": "table", + "html": "
MethodsCMU-MOSEI → CMU-MOSICMU-MOSI → CMU-MOSEI
MAE ↓CorrAcc-7Acc-2F1MAE ↓CorrAcc-7Acc-2F1
Direct Transfer0.7940.76439.579.7/81.579.5/81.40.6210.68551.379.54/82.1480.84/81.33
Supervised
TFN [37] †0.9010.69834.9-/80.8-/80.70.5930.70050.2-/82.5-/82.1
ICCN [22] †0.8620.71439.0-/83.0-/83.00.5650.71351.6-/84.2-/84.2
MISA [7] ‡0.8040.764-80.79/82.1080.77/82.030.5680.724-82.59/84.2382.67/83.97
MAG-BERT [20] ‡0.7310.789-82.50/84.3082.60/84.300.5390.753-83.80/85.2083.70/85.10
Self-MM [36] ‡0.7130.798-84.00/85.9884.42/85.950.5300.765-82.81/85.1782.53/85.30
MMIM [6] ◇0.7000.80046.6584.14/86.0684.00/85.980.5260.77254.2482.24/85.9782.66/85.94
UDA
DAN [12]0.7770.77439.7980.03/81.7179.74/81.490.6140.69351.680.24/81.3281.36/82.47
ADDA [27]0.7840.77340.1480.12/82.2680.13/82.320.6360.70751.480.47/81.5981.53/82.76
MM-SADA [15]0.7870.76940.5280.9/82.7780.68/82.630.6670.68452.180.32/81.4481.26/81.95
MDMN [44]0.7780.77439.6581.92/82.0181.97/82.110.6020.71252.882.24/82.3882.95/83.26
OSAN(TAL + Mixup)0.7530.78242.6482.44/83.3282.14/83.210.5420.75753.1482.76/82.8883.13/83.96
OSAN(TAL + DDG)0.7130.80146.3883.12/84.5883.02/84.510.5320.76853.8483.41/84.3683.31/84.47
", + "image_path": "e9cb6c510b2a1e1bfce9d9241e0a8719a3ec39ab3d401fbf0c570ad2881a594a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 332, + 287, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 332, + 287, + 357 + ], + "spans": [ + { + "bbox": [ + 46, + 332, + 287, + 357 + ], + "type": "text", + "content": "lar multimodal sentiment analysis, the other is a new task called video text classification." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 365, + 218, + 378 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 365, + 218, + 378 + ], + "spans": [ + { + "bbox": [ + 47, + 365, + 218, + 378 + ], + "type": "text", + "content": "4.1. Multimodal Sentiment Analysis" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 383, + 288, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 383, + 288, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 383, + 288, + 456 + ], + "type": "text", + "content": "Multimodal sentiment analysis is a new dimension of traditional text-based sentiment analysis, which goes beyond the analysis of texts, and includes other modalities data. This task processes data from multiple sources, such as acoustic, visual, and textual information to understand various human emotions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 471, + 166, + 484 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 471, + 166, + 484 + ], + "spans": [ + { + "bbox": [ + 47, + 471, + 166, + 484 + ], + "type": "text", + "content": "4.1.1 Experimental Setup" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 490, + 287, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 490, + 287, + 526 + ], + "spans": [ + { + "bbox": [ + 46, + 490, + 287, + 526 + ], + "type": "text", + "content": "Datasets and Model Structure We use two classical datasets, called CMU-MOSEI [39] and CMU-MOSI [38], for the multimodal sentiment analysis task." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 541, + 287, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 541, + 287, + 662 + ], + "spans": [ + { + "bbox": [ + 46, + 541, + 287, + 662 + ], + "type": "text", + "content": "Metrics We use the metrics that have been presented in [6]: mean absolute error (MAE), which is the average mean absolute difference value between predicted values and truth values. Pearson correlation (Corr) that measures the degree of prediction skew, seven-class classification accuracy (Acc-7) indicating the proportion of predictions that correctly fall into the same interval of seven intervals between -3 and +3 as the corresponding truths, binary classification accuracy (Acc-2) and F1 score for positive/negative and non-negative/negative classification results." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": "Baselines To inspect the relative performance of our model, we compare our model with various baselines in three aspects: Direct Transfer, UDA and Supervised" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 332, + 547, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 332, + 547, + 490 + ], + "spans": [ + { + "bbox": [ + 304, + 332, + 547, + 490 + ], + "type": "text", + "content": "methods. Direct transfer means that we train a model using source dataset and then predict test samples in target dataset. For UDA, we use model structures of discrepancy-based model DAN [12], adversarial-based model ADDA [27], and some typical studies focused on multimodal area such as MM-SADA [14] and MDMN [44] where multimodal alignment and domain adaptation are performed in two separate stages. Besides direct transfer and UDA methods, we also bring supervised methods for evaluation. For supervised methods, we consider the pure learning-based model TFN [37], feature space manipulation like ICCN [22], MISA [7], more recent and competitive baselines including MAG-BERT [20], Self-MM [36] and MMIM [6]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 506, + 432, + 520 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 506, + 432, + 520 + ], + "spans": [ + { + "bbox": [ + 305, + 506, + 432, + 520 + ], + "type": "text", + "content": "4.1.2 Experimental Results" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 526, + 547, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 526, + 547, + 588 + ], + "spans": [ + { + "bbox": [ + 304, + 526, + 547, + 588 + ], + "type": "text", + "content": "Overall The overall results are shown in Table 1. In detail, we find that OSAN achieves the best results among all UDA methods on all metrics with large margin. Surprisingly, OSAN even achieves better performance than many supervised methods, such as TFN, MISA and MAG-BERT." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": "Ablation Study To verify the contributions of each component of OSAN, we design a series of ablation experiments. First, we evaluate the effectiveness of the two proposed modules TAL and DDG by eliminating one or both of them from our model, as illustrated in Table 2. We notice a manifest performance degradation after removing TAL or DDG, and the results are even worse when both modules are removed. Furthermore, we conduct additional experiments to explore the contributions of the three losses. In Table 3," + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "text", + "content": "3555" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 264 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 264 + ], + "type": "text", + "content": "we find that only using classification loss " + }, + { + "bbox": [ + 46, + 72, + 289, + 264 + ], + "type": "inline_equation", + "content": "L_{cls}" + }, + { + "bbox": [ + 46, + 72, + 289, + 264 + ], + "type": "text", + "content": ", our model can achieve competitive performance compared to traditional multimodal DA methods such as MDMN, demonstrating that both modality and domain are well aligned using TAL and DDG. Finally, we adjust some important hyperparameters of DDG and TAL to evaluate their importance. In terms of DDG, we adjust the number of created soft domains by changing multi-head " + }, + { + "bbox": [ + 46, + 72, + 289, + 264 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 46, + 72, + 289, + 264 + ], + "type": "text", + "content": ". Furthermore, we adjust the fusion coefficient " + }, + { + "bbox": [ + 46, + 72, + 289, + 264 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 72, + 289, + 264 + ], + "type": "text", + "content": " to control the fusion ratio between the source and target samples. In Table 4, we observe that the performance improves significantly as the number of domains increases. We also find that the fusion coefficient set to 0.85 can achieve the best result. In terms of TAL, we reduce multimodal features to different combinations of dimensions. We find that the combination of " + }, + { + "bbox": [ + 46, + 72, + 289, + 264 + ], + "type": "inline_equation", + "content": "384 * 2 * 2" + }, + { + "bbox": [ + 46, + 72, + 289, + 264 + ], + "type": "text", + "content": " achieves the best performance, as shown in Table 5." + } + ] + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 48, + 294, + 282, + 373 + ], + "blocks": [ + { + "bbox": [ + 47, + 274, + 286, + 285 + ], + "lines": [ + { + "bbox": [ + 47, + 274, + 286, + 285 + ], + "spans": [ + { + "bbox": [ + 47, + 274, + 286, + 285 + ], + "type": "text", + "content": "Table 2. Ablation study of the proposed modules on CMU-MOSI." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 294, + 282, + 373 + ], + "lines": [ + { + "bbox": [ + 48, + 294, + 282, + 373 + ], + "spans": [ + { + "bbox": [ + 48, + 294, + 282, + 373 + ], + "type": "table", + "html": "
Modulew/o DDGw/o TALw/o Both
MAE ↓+0.039+0.051+0.074
Corr-0.023-0.021-0.032
Acc-7-3.55-4.65-5.86
Acc-2-1.03/-1.30-1.89/-1.72-2.22/-1.81
F1-1.15/-1.52-1.85/-1.61-2.34/-1.88
", + "image_path": "3545f081dee40d265105b137e1c7d7ace9feea4524a54a03854d63cacfc9fed9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 48, + 411, + 284, + 491 + ], + "blocks": [ + { + "bbox": [ + 65, + 391, + 268, + 402 + ], + "lines": [ + { + "bbox": [ + 65, + 391, + 268, + 402 + ], + "spans": [ + { + "bbox": [ + 65, + 391, + 268, + 402 + ], + "type": "text", + "content": "Table 3. Ablation study of three losses on CMU-MOSI." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 411, + 284, + 491 + ], + "lines": [ + { + "bbox": [ + 48, + 411, + 284, + 491 + ], + "spans": [ + { + "bbox": [ + 48, + 411, + 284, + 491 + ], + "type": "table", + "html": "
LossLclsLcls + LadvLcls + Ldis
MAE ↓+0.018+0.009+0.016
Corr-0.016-0.012-0.014
Acc-7-1.07-0.82-0.67
Acc-2-0.65/-0.95-0.32/-0.28-0.67/-0.74
F1-0.48/-0.69-0.7/-0.49-0.31/-0.58
", + "image_path": "46d0554b44a06aa4a64292332db292441ec0dd9dd0f5727fb9c75c65f08f8f31.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 48, + 529, + 282, + 620 + ], + "blocks": [ + { + "bbox": [ + 47, + 509, + 286, + 521 + ], + "lines": [ + { + "bbox": [ + 47, + 509, + 286, + 521 + ], + "spans": [ + { + "bbox": [ + 47, + 509, + 286, + 521 + ], + "type": "text", + "content": "Table 4. Ablation study of DDG hyperparameters on CMU-MOSI." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 529, + 282, + 620 + ], + "lines": [ + { + "bbox": [ + 48, + 529, + 282, + 620 + ], + "spans": [ + { + "bbox": [ + 48, + 529, + 282, + 620 + ], + "type": "table", + "html": "
DDGλ = 0.85\nH = 64λ = 0.9\nH = 64λ = 0.85\nH = 32λ = 0.9\nH = 32
MAE ↓0.00+0.0040.001+0.006
Corr0.00-0.01-0.003-0.016
Acc-70.00-0.590-0.27-0.75
Acc-20.00/0.00-0.27/-1.14-0.16/-1.07-0.39/-1.29
F10.00/0.00-0.259/-0.15-0.14/-0.04-0.39/-0.30
", + "image_path": "4a1c4d7a031d0f50679247b48ca2c4b0221e5e332cd34c1eedc12aff50b8a23c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 635, + 186, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 186, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 186, + 647 + ], + "type": "text", + "content": "4.2. Video Text Classification" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": "Video text classification is a real industrial task for video understanding [9]. With massive videos generated everyday, video textual information extraction is an essential work in many applications. However, numerous useless texts, such as rolling texts and blurred background texts," + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 307, + 91, + 543, + 182 + ], + "blocks": [ + { + "bbox": [ + 306, + 70, + 545, + 82 + ], + "lines": [ + { + "bbox": [ + 306, + 70, + 545, + 82 + ], + "spans": [ + { + "bbox": [ + 306, + 70, + 545, + 82 + ], + "type": "text", + "content": "Table 5. Ablation study of TAL hyperparameters on CMU-MOSI." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 91, + 543, + 182 + ], + "lines": [ + { + "bbox": [ + 307, + 91, + 543, + 182 + ], + "spans": [ + { + "bbox": [ + 307, + 91, + 543, + 182 + ], + "type": "table", + "html": "
TALt:384t:192t:192
v:2,a:2v:4,a:2v:2,a:4
MAE↓0.000+0.007+0.006
Corr0.00-0.009-0.006
Acc-70.00-1.17-0.91
Acc-20.00/0.00-0.77/-1.16-0.30/-0.91
F10.00/0.00-0.59/-2.02-0.13/-0.79
", + "image_path": "3d57b9497d9e1d85c6c0516786e6e5339a4a2025c77528b2bef30aff8c4185e1.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 198, + 547, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 198, + 547, + 246 + ], + "spans": [ + { + "bbox": [ + 304, + 198, + 547, + 246 + ], + "type": "text", + "content": "have no good effect or even side effects on downstream tasks. To control and filter video content, we propose a new video text classification task to classify video texts into valuable categories." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 267, + 424, + 280 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 267, + 424, + 280 + ], + "spans": [ + { + "bbox": [ + 305, + 267, + 424, + 280 + ], + "type": "text", + "content": "4.2.1 Experimental Setup" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 289, + 545, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 289, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 304, + 289, + 545, + 529 + ], + "type": "text", + "content": "Datasets and Model Structure We construct a well-defined industrial-grade dataset, called Text-news, which is dedicated to promote video text extraction research for news applications. The extraction of text in videos usually involves two steps: (1) text recognition and (2) text classification. We prepare various videos from news programs for annotation and use a general Optical Character Recognition (OCR) engine to locate and recognize all the texts in the videos. Then, an annotation system is designed for the video text classification task to tag video texts. Three representative categories are defined which are Caption, Subtitle and Others. To verify the effectiveness and generality of our method, we use Text-news as source domain dataset and construct a dataset named Text-show with a few annotated samples as the target domain dataset. The difference between the source and target datasets is that the video program is totally different. The source dataset is constructed from news videos, while the target dataset is constructed from videos of variety shows. We use the standard Precision, Recall, and F1 score for evaluation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 550, + 545, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 550, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 304, + 550, + 545, + 624 + ], + "type": "text", + "content": "Baselines To inspect the relative performance of our model OSAN, we compare it with various baselines in two aspects: Direct Transfer and UDA methods. Unlike multimodal sentiment analysis, we do not use Supervised methods for comparison because the target dataset does not have annotated training data." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 644, + 432, + 656 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 644, + 432, + 656 + ], + "spans": [ + { + "bbox": [ + 305, + 644, + 432, + 656 + ], + "type": "text", + "content": "4.2.2 Experimental Results" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "Overall The results are shown in Table 6. All the metrics of OSAN achieve the best results when compared to other UDA methods and direct transfer, leading to a splendid performance of video text classification task." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3556" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 52, + 91, + 278, + 223 + ], + "blocks": [ + { + "bbox": [ + 67, + 71, + 267, + 81 + ], + "lines": [ + { + "bbox": [ + 67, + 71, + 267, + 81 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 267, + 81 + ], + "type": "text", + "content": "Table 6. Video text classification results on Text-show." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 52, + 91, + 278, + 223 + ], + "lines": [ + { + "bbox": [ + 52, + 91, + 278, + 223 + ], + "spans": [ + { + "bbox": [ + 52, + 91, + 278, + 223 + ], + "type": "table", + "html": "
MethodsText-news → Text-show
PrecisionRecallF1
Direct Transfer80.277.9879.08
UDA
DAN [12]87.4480.5883.87
ADDA [27]91.6683.6687.48
MM-SADA [15]94.0783.4988.46
MDMN [44]93.5483.6988.34
OSAN(TAL + Mixup)94.4284.7989.35
OSAN(TAL + DDG)95.0386.4490.53
", + "image_path": "929645d2f94acbcb8a3154a1132894c1c6feef1959d0fa7f848a13a337e93462.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 53, + 250, + 279, + 304 + ], + "blocks": [ + { + "bbox": [ + 52, + 229, + 282, + 240 + ], + "lines": [ + { + "bbox": [ + 52, + 229, + 282, + 240 + ], + "spans": [ + { + "bbox": [ + 52, + 229, + 282, + 240 + ], + "type": "text", + "content": "Table 7. Ablation study of the proposed modules on Text-show." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 53, + 250, + 279, + 304 + ], + "lines": [ + { + "bbox": [ + 53, + 250, + 279, + 304 + ], + "spans": [ + { + "bbox": [ + 53, + 250, + 279, + 304 + ], + "type": "table", + "html": "
Modulew/o DDGw/o TALw/o Both
Precision-0.90-0.51-0.96
Recall-1.24-2.37-2.95
F1-1.09-1.55-2.07
", + "image_path": "85b98dbc349d4579515fa84be50bde415d5431e8c5beecde31e4ad5a03a6ddf4.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 53, + 337, + 280, + 392 + ], + "blocks": [ + { + "bbox": [ + 70, + 316, + 264, + 327 + ], + "lines": [ + { + "bbox": [ + 70, + 316, + 264, + 327 + ], + "spans": [ + { + "bbox": [ + 70, + 316, + 264, + 327 + ], + "type": "text", + "content": "Table 8. Ablation study of three losses on Text-show." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 53, + 337, + 280, + 392 + ], + "lines": [ + { + "bbox": [ + 53, + 337, + 280, + 392 + ], + "spans": [ + { + "bbox": [ + 53, + 337, + 280, + 392 + ], + "type": "table", + "html": "
LossLclsLcls + LadvLcls + Ldis
Precision-0.35-0.17-0.45
Recall-0.42-0.24-0.14
F1-0.39-0.21-0.28
", + "image_path": "6324a2056683b0ab713b88431a8f6ad1619d2eea3a909e012f782bbeb55b7bc6.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 53, + 424, + 279, + 491 + ], + "blocks": [ + { + "bbox": [ + 51, + 403, + 283, + 415 + ], + "lines": [ + { + "bbox": [ + 51, + 403, + 283, + 415 + ], + "spans": [ + { + "bbox": [ + 51, + 403, + 283, + 415 + ], + "type": "text", + "content": "Table 9. Ablation study of DDG hyperparameters on Text-show." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 53, + 424, + 279, + 491 + ], + "lines": [ + { + "bbox": [ + 53, + 424, + 279, + 491 + ], + "spans": [ + { + "bbox": [ + 53, + 424, + 279, + 491 + ], + "type": "table", + "html": "
DDGλ = 0.9\nH = 128λ = 0.95\nH = 128λ = 0.9\nH = 64λ = 0.95\nH = 64
Precision0.00-0.16-0.75-0.67
Recall0.00-0.32-1.21-1.33
F10.00-0.25-1.01-1.03
", + "image_path": "249e7bef64259e425b618b0b48e73145d23afe445ad759fcd1f1bc6baff9363e.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 53, + 515, + 279, + 569 + ], + "blocks": [ + { + "bbox": [ + 50, + 494, + 284, + 506 + ], + "lines": [ + { + "bbox": [ + 50, + 494, + 284, + 506 + ], + "spans": [ + { + "bbox": [ + 50, + 494, + 284, + 506 + ], + "type": "text", + "content": "Table 10. Ablation study of TAL hyperparameters on Text-show." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 53, + 515, + 279, + 569 + ], + "lines": [ + { + "bbox": [ + 53, + 515, + 279, + 569 + ], + "spans": [ + { + "bbox": [ + 53, + 515, + 279, + 569 + ], + "type": "table", + "html": "
TALt:56,v:14t:14,v:56t:28,v:28
Precision0.000-0.96+0.24
Recall0.000+0.05-0.59
F10.000-0.41-0.22
", + "image_path": "a3ea129e6d16fb6ad69f3e7cd599baa941b39204441b1ee23f58b2bf7a0b5036.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": "Ablation Study Similar to multimodal sentiment analysis, in order to show the benefits of TAL and DDG modules, we perform a series of ablation experiments. First, we eliminate one or both modules from our model. In Table 7, the overall F1 score decreases after removing single module, and the results are even worse when removing both modules, which demonstrates the efficacy of these two modules. Furthermore, we study the impact of three losses of category classification loss, adversarial loss, and discrepancy loss, as shown in Table 8. The results show that all three losses contribute to performance. Moreover, by removing" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 72, + 547, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 216 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 216 + ], + "type": "text", + "content": "DA related losses such as adversarial loss and discrepancy loss, it performs slightly worse than keeping all three losses, but still better than traditional multimodal DA methods such as MM-SADA and MDMN. This observation indicates that different domains are well aligned by TAL and DDG. Finally, we tuned some hyperparameters of TAL and DDG to explore their impact. For the DDG hyperparameters, in Table 9, we observe that a smaller fusion coefficient " + }, + { + "bbox": [ + 304, + 72, + 547, + 216 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 72, + 547, + 216 + ], + "type": "text", + "content": " and a larger multi-head " + }, + { + "bbox": [ + 304, + 72, + 547, + 216 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 304, + 72, + 547, + 216 + ], + "type": "text", + "content": " can achieve better results. For the TAL hyperparameters, we reduce multimodal features to different dimensions. In Table 10, the combination " + }, + { + "bbox": [ + 304, + 72, + 547, + 216 + ], + "type": "inline_equation", + "content": "56*14" + }, + { + "bbox": [ + 304, + 72, + 547, + 216 + ], + "type": "text", + "content": " for textual and visual features achieves the best result." + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 319, + 226, + 383, + 289 + ], + "blocks": [ + { + "bbox": [ + 319, + 226, + 383, + 289 + ], + "lines": [ + { + "bbox": [ + 319, + 226, + 383, + 289 + ], + "spans": [ + { + "bbox": [ + 319, + 226, + 383, + 289 + ], + "type": "image", + "image_path": "840d248501d4438e62a3cdfe44e12b849405d601bbd5f71333a9677d3dfa676e.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 333, + 291, + 367, + 297 + ], + "lines": [ + { + "bbox": [ + 333, + 291, + 367, + 297 + ], + "spans": [ + { + "bbox": [ + 333, + 291, + 367, + 297 + ], + "type": "text", + "content": "(a) Directly transfer" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 397, + 226, + 460, + 289 + ], + "blocks": [ + { + "bbox": [ + 397, + 226, + 460, + 289 + ], + "lines": [ + { + "bbox": [ + 397, + 226, + 460, + 289 + ], + "spans": [ + { + "bbox": [ + 397, + 226, + 460, + 289 + ], + "type": "image", + "image_path": "18bcbe9e4f1318cf0419f661e21a0db4a5140ef1749f31a0910d4c1eaa256dc5.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 400, + 290, + 455, + 296 + ], + "lines": [ + { + "bbox": [ + 400, + 290, + 455, + 296 + ], + "spans": [ + { + "bbox": [ + 400, + 290, + 455, + 296 + ], + "type": "text", + "content": "(b) Our method w/o TAL&DDG" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 473, + 226, + 536, + 289 + ], + "blocks": [ + { + "bbox": [ + 473, + 226, + 536, + 289 + ], + "lines": [ + { + "bbox": [ + 473, + 226, + 536, + 289 + ], + "spans": [ + { + "bbox": [ + 473, + 226, + 536, + 289 + ], + "type": "image", + "image_path": "b97a35b19be5df092e1814c4f6975a49a3770ab886d41c439fb39c42280df987.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 488, + 291, + 516, + 297 + ], + "lines": [ + { + "bbox": [ + 488, + 291, + 516, + 297 + ], + "spans": [ + { + "bbox": [ + 488, + 291, + 516, + 297 + ], + "type": "text", + "content": "(c) Our method" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 316, + 344, + 380, + 408 + ], + "blocks": [ + { + "bbox": [ + 306, + 306, + 545, + 328 + ], + "lines": [ + { + "bbox": [ + 306, + 306, + 545, + 328 + ], + "spans": [ + { + "bbox": [ + 306, + 306, + 545, + 328 + ], + "type": "text", + "content": "Figure 4. Visualization of the feature distribution of the target domain for multimodal sentiment analysis." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 316, + 344, + 380, + 408 + ], + "lines": [ + { + "bbox": [ + 316, + 344, + 380, + 408 + ], + "spans": [ + { + "bbox": [ + 316, + 344, + 380, + 408 + ], + "type": "image", + "image_path": "c009a84b8e253a616ccc50dcd967bb72587c0390e6ce73e71bb50392b10bc3eb.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 329, + 411, + 364, + 416 + ], + "lines": [ + { + "bbox": [ + 329, + 411, + 364, + 416 + ], + "spans": [ + { + "bbox": [ + 329, + 411, + 364, + 416 + ], + "type": "text", + "content": "(a) Directly transfer" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 427, + 545, + 449 + ], + "lines": [ + { + "bbox": [ + 306, + 427, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 306, + 427, + 545, + 449 + ], + "type": "text", + "content": "Figure 5. Visualization of the feature distribution of the target domain for video text classification." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 394, + 344, + 460, + 408 + ], + "blocks": [ + { + "bbox": [ + 394, + 344, + 460, + 408 + ], + "lines": [ + { + "bbox": [ + 394, + 344, + 460, + 408 + ], + "spans": [ + { + "bbox": [ + 394, + 344, + 460, + 408 + ], + "type": "image", + "image_path": "212c66b99d36fc9d5a8926204770b18d8b453565ec8481f2c22ac38cd988c136.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 396, + 411, + 452, + 416 + ], + "lines": [ + { + "bbox": [ + 396, + 411, + 452, + 416 + ], + "spans": [ + { + "bbox": [ + 396, + 411, + 452, + 416 + ], + "type": "text", + "content": "(b) Our method w/o TAL&DDG" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 471, + 344, + 536, + 408 + ], + "blocks": [ + { + "bbox": [ + 471, + 344, + 536, + 408 + ], + "lines": [ + { + "bbox": [ + 471, + 344, + 536, + 408 + ], + "spans": [ + { + "bbox": [ + 471, + 344, + 536, + 408 + ], + "type": "image", + "image_path": "e8bdbbb9e2bbeb10ddf20ad13275b340e6af69e92d63c3c6d6e5f16a17b817b2.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 488, + 411, + 516, + 416 + ], + "lines": [ + { + "bbox": [ + 488, + 411, + 516, + 416 + ], + "spans": [ + { + "bbox": [ + 488, + 411, + 516, + 416 + ], + "type": "text", + "content": "(c) Our method" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 459, + 375, + 471 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 459, + 375, + 471 + ], + "spans": [ + { + "bbox": [ + 306, + 459, + 375, + 471 + ], + "type": "text", + "content": "5. Discussion" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 306, + 479, + 511, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 479, + 511, + 491 + ], + "spans": [ + { + "bbox": [ + 306, + 479, + 511, + 491 + ], + "type": "text", + "content": "5.1. Visual Analysis of Feature Distribution" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 497, + 545, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 497, + 545, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 497, + 545, + 641 + ], + "type": "text", + "content": "We use t-SNE [28] to visualize the feature distribution of test data in target domain. Fig.5 illustrates the t-SNE visualization of the feature distribution in the video text classification task. We find that the features of different classes are separated more clearly, and the features in the same class are more centralized. A similar phenomenon is observed in Fig.4 of the feature distribution for the multimodal sentiment analysis task. This makes sense because TAL can explore the relationship between domains and modalities and guide them to utilize complementary information. Moreover, DDG can construct new domains that guarantee domain-invariance in a more continuous latent space." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 306, + 647, + 446, + 660 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 647, + 446, + 660 + ], + "spans": [ + { + "bbox": [ + 306, + 647, + 446, + 660 + ], + "type": "text", + "content": "5.2. Distribution Discrepancy" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": "To quantify the distribution discrepancy of source and target domains, we use the classical metric of symmetric Jensen-Shannon divergence (JSD) [2]. Fig.7 shows the feature distribution discrepancy based on two baselines and" + } + ] + } + ], + "index": 30 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3557" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 63, + 76, + 530, + 262 + ], + "blocks": [ + { + "bbox": [ + 63, + 76, + 530, + 262 + ], + "lines": [ + { + "bbox": [ + 63, + 76, + 530, + 262 + ], + "spans": [ + { + "bbox": [ + 63, + 76, + 530, + 262 + ], + "type": "table", + "html": "
DomainImageTextMDMN ResultImageTextOSAN(ours) Result
SourceBeyond that, I actually excelled academically from some of my decisions um i say go see it coz i i enjoyed it nowgt: 0 prediction: --gt: 2.0 prediction:-0.37Thank you for that Godelieve, we all enjoyed your comments! um i say go see it coz i i enjoyed it nowgt: 2 prediction: --gt: 2.0 prediction:1.85
Target
Sourcehey were all randomly chosen ad we're going to have my lovely assistant here choose from among these three people and um i liked you know catherine hardwicke is like you know kind of ridiculousgt: 0 prediction: --gt: -0.6 prediction:0.42This story demonstrates how ridiculous and petty fighting, or war... and um i liked you know catherine hardwicke is like you know kind of ridiculousgt: -2 prediction: --gt: -0.6 prediction: -0.56
Target
Sourcewell if they had kept their mouth shut, that movie would have been fabulous. and it was kind of interestinggt: -1 prediction: --gt: 1.2 prediction:-1.94Learn a language if you can, because that will make your life more interesting and it was kind of interestinggt: 2.0 prediction: --gt: 1.2 prediction:1.53
Target
", + "image_path": "7e10fa6a31a82bcf3938021c99413f1534ee29bed1161c81f69b2291b4701958.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 321, + 288, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 321, + 288, + 392 + ], + "spans": [ + { + "bbox": [ + 46, + 321, + 288, + 392 + ], + "type": "text", + "content": "our model on two tasks. Two baselines are Direct Transfer and our model without the two proposed modules TAL and DDG. We observe that JSD of our model is smaller than JSD of both baselines. This phenomenon implies that our model can reduce the domain gap more effectively than the baselines." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 48, + 405, + 287, + 533 + ], + "blocks": [ + { + "bbox": [ + 46, + 275, + 547, + 309 + ], + "lines": [ + { + "bbox": [ + 46, + 275, + 547, + 309 + ], + "spans": [ + { + "bbox": [ + 46, + 275, + 547, + 309 + ], + "type": "text", + "content": "Figure 6. Some samples analysis of one baseline MDMN and our model OSAN for UDA task CMU-MOSEI " + }, + { + "bbox": [ + 46, + 275, + 547, + 309 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 275, + 547, + 309 + ], + "type": "text", + "content": " CMU-MOSI. Columns show information of one representative frame extracted from video clip, text message and the prediction result. Rows give information of target samples and closest source samples." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 48, + 405, + 287, + 533 + ], + "lines": [ + { + "bbox": [ + 48, + 405, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 405, + 287, + 533 + ], + "type": "image", + "image_path": "1bdd0ff247178dfbcc7e838ba9559a1f5ec8830e1e8aad738f256c93fe6aac28.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 542, + 288, + 574 + ], + "lines": [ + { + "bbox": [ + 46, + 542, + 288, + 574 + ], + "spans": [ + { + "bbox": [ + 46, + 542, + 288, + 574 + ], + "type": "text", + "content": "Figure 7. Distribution discrepancy of different methods on two different tasks CMU-MOSEI " + }, + { + "bbox": [ + 46, + 542, + 288, + 574 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 542, + 288, + 574 + ], + "type": "text", + "content": " CMU-MOSI and Text-news " + }, + { + "bbox": [ + 46, + 542, + 288, + 574 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 542, + 288, + 574 + ], + "type": "text", + "content": " Text-show." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 587, + 129, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 587, + 129, + 598 + ], + "spans": [ + { + "bbox": [ + 47, + 587, + 129, + 598 + ], + "type": "text", + "content": "5.3. Case Studies" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": "To deeply explore the advantages of our model over other baselines, we give some predictions of a two-stage baseline MDMN and our model OSAN for the UDA task CMU-MOSEI " + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 605, + 288, + 715 + ], + "type": "text", + "content": " CMU-MOSI, as shown in Fig.6. We select some typical samples from the target domain to explore how our model works. For each target sample, we find the closest sample from source domain by calculating cosine similarity between target sample and source sample. We observe that the closest source samples picked by our model OSAN are" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 321, + 547, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 321, + 547, + 501 + ], + "spans": [ + { + "bbox": [ + 304, + 321, + 547, + 501 + ], + "type": "text", + "content": "more relevant to the target samples. Specifically, take the target sample and source sample in the first row for example, with our model, speakers in both the source frame and target frame are very happy. Moreover, there is one common keyword in what the speakers said: enjoyed. Combining these two modalities of image and text, our model correctly classifies this target sample as positive sentiment with a score of 1.85, which is very close to the ground truth score of 2. In contrast, source sample found by the MDMN is not relevant to the target sample. There is no similar semantic information in what they said. The expressions in the frames are dissimilar. The phenomenon in these cases implies that our model can learn a common latent space in which the source domain and target domain are well aligned." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 511, + 378, + 523 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 511, + 378, + 523 + ], + "spans": [ + { + "bbox": [ + 306, + 511, + 378, + 523 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 531, + 547, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 531, + 547, + 675 + ], + "spans": [ + { + "bbox": [ + 304, + 531, + 547, + 675 + ], + "type": "text", + "content": "To address the problems of domain adaptation and modality alignment in multimodal domain adaptation, we propose a novel one-stage alignment network to unify these two problems in one stage. In our model, the TAL module is proposed to explore the relationship between domains and modalities, which can provide complementary information to each other. Moreover, the DDG module is designed to generate new soft samples that mix information from both the source and target domains, guiding our model in a more continuous space. Extensive experiments on two multimodal tasks with different extents of domain shift demonstrate the excellent performance of our model." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3558" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 715 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Shai Ben-David, John Blitzer, Koby Crammer, Alex Kulesza, Fernando Pereira, and Jennifer Wortman Vaughan. A theory of learning from different domains. Machine learning, 79(1):151-175, 2010. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 288, + 181 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 288, + 181 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 288, + 181 + ], + "type": "text", + "content": "[2] Bent Fuglede and Flemming Topsoe. Jensen-shannon divergence and hilbert space embedding. In International Symposium on Information Theory, 2004. ISIT 2004. Proceedings., page 31. IEEE, 2004. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 182, + 288, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 182, + 288, + 216 + ], + "spans": [ + { + "bbox": [ + 53, + 182, + 288, + 216 + ], + "type": "text", + "content": "[3] Yaroslav Ganin and Victor Lempitsky. Unsupervised domain adaptation by backpropagation. In International conference on machine learning, pages 1180-1189. PMLR, 2015. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 217, + 288, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 217, + 288, + 271 + ], + "spans": [ + { + "bbox": [ + 53, + 217, + 288, + 271 + ], + "type": "text", + "content": "[4] Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, François Laviolette, Mario Marchand, and Victor Lempitsky. Domain-adversarial training of neural networks. The journal of machine learning research, 17(1):2096-2030, 2016. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 273, + 288, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 273, + 288, + 338 + ], + "spans": [ + { + "bbox": [ + 53, + 273, + 288, + 338 + ], + "type": "text", + "content": "[5] Arthur Gretton, Dino Sejdinovic, Heiko Strathmann, Sivaraman Balakrishnan, Massimiliano Pontil, Kenji Fukumizu, and Bharath K Striperumbudur. Optimal kernel choice for large-scale two-sample tests. In Advances in neural information processing systems, pages 1205-1213. Citeseer, 2012. 3, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 340, + 288, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 340, + 288, + 373 + ], + "spans": [ + { + "bbox": [ + 53, + 340, + 288, + 373 + ], + "type": "text", + "content": "[6] Wei Han, Hui Chen, and Soujanya Poria. Improving multimodal fusion with hierarchical mutual information maximization for multimodal sentiment analysis, 2021. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 375, + 288, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 375, + 288, + 430 + ], + "spans": [ + { + "bbox": [ + 53, + 375, + 288, + 430 + ], + "type": "text", + "content": "[7] Devamanyu Hazarika, Roger Zimmermann, and Soujanya Poria. Misa: Modality-invariant and-specific representations for multimodal sentiment analysis. In Proceedings of the 28th ACM International Conference on Multimedia, pages 1122-1131, 2020. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 431, + 288, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 431, + 288, + 496 + ], + "spans": [ + { + "bbox": [ + 53, + 431, + 288, + 496 + ], + "type": "text", + "content": "[8] Donghyun Kim, Yi-Hsuan Tsai, Bingbing Zhuang, Xiang Yu, Stan Sclaroff, Kate Saenko, and Manmohan Chandraker. Learning cross-modal contrastive features for video domain adaptation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13618-13627, 2021. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 498, + 288, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 498, + 288, + 532 + ], + "spans": [ + { + "bbox": [ + 53, + 498, + 288, + 532 + ], + "type": "text", + "content": "[9] Ye Liu, Changchong Lu, Chen Lin, Di Yin, and Bo Ren. Contrastive graph multimodal model for text classification in videos. arXiv preprint arXiv:2206.02343, 2022. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 533, + 288, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 533, + 288, + 588 + ], + "spans": [ + { + "bbox": [ + 48, + 533, + 288, + 588 + ], + "type": "text", + "content": "[10] Ye Liu, Lingfeng Qiao, Di Yin, Zhuoxuan Jiang, Xinghua Jiang, Deqiang Jiang, and Bo Ren. Os-msl: One stage multimodal sequential link framework for scene segmentation and classification. In Proceedings of the 30th ACM International Conference on Multimedia, pages 6269-6277, 2022. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 590, + 288, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 288, + 643 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 288, + 643 + ], + "type": "text", + "content": "[11] Zhun Liu, Ying Shen, Varun Bharadhwaj Lakshminarasimhan, Paul Pu Liang, Amir Zadeh, and Louis-Philippe Morency. Efficient low-rank multimodal fusion with modality-specific factors. arXiv preprint arXiv:1806.00064, 2018.3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 646, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 646, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 646, + 288, + 689 + ], + "type": "text", + "content": "[12] Mingsheng Long, Yue Cao, Jianmin Wang, and Michael Jordan. Learning transferable features with deep adaptation networks. In International conference on machine learning, pages 97-105. PMLR, 2015. 5, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 715 + ], + "type": "text", + "content": "[13] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representational" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "tations for vision-and-language tasks. arXiv preprint arXiv:1908.02265, 2019. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 95, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 95, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 308, + 95, + 545, + 139 + ], + "type": "text", + "content": "[14] Jonathan Munro and Dima Damen. Multi-modal domain adaptation for fine-grained action recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 122-132, 2020. 1, 2, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 140, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 140, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 308, + 140, + 545, + 183 + ], + "type": "text", + "content": "[15] Jonathan Munro and Dima Damen. Multi-modal domain adaptation for fine-grained action recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020. 5, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 184, + 545, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 184, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 308, + 184, + 545, + 227 + ], + "type": "text", + "content": "[16] Sinno Jialin Pan, Ivor W Tsang, James T Kwok, and Qiang Yang. Domain adaptation via transfer component analysis. IEEE transactions on neural networks, 22(2):199-210, 2010. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 228, + 545, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 228, + 545, + 282 + ], + "spans": [ + { + "bbox": [ + 308, + 228, + 545, + 282 + ], + "type": "text", + "content": "[17] Fan Qi, Xiaoshan Yang, and Changsheng Xu. A unified framework for multimodal domain adaptation. In Proceedings of the 26th ACM International Conference on Multimedia, MM '18, page 429-437, New York, NY, USA, 2018. Association for Computing Machinery. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 283, + 545, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 283, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 308, + 283, + 545, + 326 + ], + "type": "text", + "content": "[18] Lingfeng Qiao, Chen Wu, Ye Liu, Haoyuan Peng, Di Yin, and Bo Ren. Grafting pre-trained models for multimodal headline generation. arXiv preprint arXiv:2211.07210, 2022. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 327, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 327, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 308, + 327, + 545, + 392 + ], + "type": "text", + "content": "[19] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 393, + 545, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 393, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 308, + 393, + 545, + 459 + ], + "type": "text", + "content": "[20] Wasifur Rahman, Md Kamrul Hasan, Sangwu Lee, Amir Zadeh, Chengfeng Mao, Louis-Philippe Morency, and Ehsan Hoque. Integrating multimodal information in large pretrained transformers. In Proceedings of the conference. Association for Computational Linguistics. Meeting, volume 2020, page 2359. NIH Public Access, 2020. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 459, + 545, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 459, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 308, + 459, + 545, + 502 + ], + "type": "text", + "content": "[21] Baochen Sun and Kate Saenko. Deep coral: Correlation alignment for deep domain adaptation. In European conference on computer vision, pages 443-450. Springer, 2016. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 503, + 545, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 503, + 545, + 558 + ], + "spans": [ + { + "bbox": [ + 308, + 503, + 545, + 558 + ], + "type": "text", + "content": "[22] Zhongkai Sun, Prathusha Sarma, William Sethares, and Yingyu Liang. Learning relationships between text, audio, and video via deep canonical correlation for multimodal language analysis. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 8992-8999, 2020. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 559, + 545, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 559, + 545, + 592 + ], + "spans": [ + { + "bbox": [ + 308, + 559, + 545, + 592 + ], + "type": "text", + "content": "[23] Hui Tang and Kui Jia. Discriminative adversarial domain adaptation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 5940-5947, 2020. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 592, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 592, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 308, + 592, + 545, + 635 + ], + "type": "text", + "content": "[24] Dacheng Tao, Xuelong Li, Xindong Wu, and Steve Maybank. Tensor rank one discriminant analysis-a convergent method for discriminative multilinear subspace selection. Neurocomputing, 71(10):1866-1882, 2008. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 636, + 545, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 636, + 545, + 680 + ], + "spans": [ + { + "bbox": [ + 308, + 636, + 545, + 680 + ], + "type": "text", + "content": "[25] Dacheng Tao, Xuelong Li, Xindong Wu, and Stephen J Maybank. General tensor discriminant analysis and gabor features for gait recognition. IEEE transactions on pattern analysis and machine intelligence, 29(10):1700-1715, 2007. 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "text", + "content": "[26] Wilhelm Tranheden, Viktor Olsson, Juliano Pinto, and Lennart Svensson. Dacs: Domain adaptation via cross-domain mixed sampling. In Proceedings of the IEEE/CVF" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3559" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 67, + 72, + 286, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 286, + 95 + ], + "type": "text", + "content": "Winter Conference on Applications of Computer Vision, pages 1379-1389, 2021. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 140 + ], + "type": "text", + "content": "[27] Eric Tzeng, Judy Hoffman, Kate Saenko, and Trevor Darrell. Adversarial discriminative domain adaptation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7167-7176, 2017. 4, 5, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 287, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 287, + 174 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 287, + 174 + ], + "type": "text", + "content": "[28] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of machine learning research, 9(11), 2008. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 175, + 287, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 287, + 218 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 287, + 218 + ], + "type": "text", + "content": "[29] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Ilia Polosukhin. Attention is all you need. arXiv preprint arXiv:1706.03762, 2017. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 220, + 287, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 220, + 287, + 274 + ], + "spans": [ + { + "bbox": [ + 48, + 220, + 287, + 274 + ], + "type": "text", + "content": "[30] Wei Wang, Haojie Li, Zhengming Ding, Feiping Nie, Junyang Chen, Xiao Dong, and Zhihui Wang. Rethinking maximum mean discrepancy for visual domain adaptation. IEEE Transactions on Neural Networks and Learning Systems, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 276, + 287, + 309 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 276, + 287, + 309 + ], + "spans": [ + { + "bbox": [ + 48, + 276, + 287, + 309 + ], + "type": "text", + "content": "[31] Garrett Wilson and Diane J Cook. A survey of unsupervised deep domain adaptation. ACM Transactions on Intelligent Systems and Technology (TIST), 11(5):1-46, 2020. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 311, + 287, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 311, + 287, + 365 + ], + "spans": [ + { + "bbox": [ + 48, + 311, + 287, + 365 + ], + "type": "text", + "content": "[32] Xinyi Wu, Zhenyao Wu, Hao Guo, Lili Ju, and Song Wang. Darnet: A one-stage domain adaptation network for unsupervised nighttime semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15769-15778, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 366, + 287, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 366, + 287, + 421 + ], + "spans": [ + { + "bbox": [ + 48, + 366, + 287, + 421 + ], + "type": "text", + "content": "[33] Minghao Xu, Jian Zhang, Bingbing Ni, Teng Li, Chengjie Wang, Qi Tian, and Wenjun Zhang. Adversarial domain adaptation with domain mixup. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 6502-6509, 2020. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 422, + 287, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 422, + 287, + 465 + ], + "spans": [ + { + "bbox": [ + 48, + 422, + 287, + 465 + ], + "type": "text", + "content": "[34] Yiming Xu, Lin Chen, Zhongwei Cheng, Lixin Duan, and Jiebo Luo. Open-ended visual question answering by multimodal domain adaptation. arXiv preprint arXiv:1911.04058, 2019. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 468, + 287, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 468, + 287, + 500 + ], + "spans": [ + { + "bbox": [ + 48, + 468, + 287, + 500 + ], + "type": "text", + "content": "[35] Shen Yan, Huan Song, Nanxiang Li, Lincan Zou, and Liu Ren. Improve unsupervised domain adaptation with mixup training. arXiv preprint arXiv:2001.00677, 2020. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 502, + 287, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 502, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 48, + 502, + 287, + 545 + ], + "type": "text", + "content": "[36] Wenmeng Yu, Hua Xu, Ziqi Yuan, and Jiele Wu. Learning modality-specific representations with self-supervised multi-task learning for multimodal sentiment analysis. arXiv preprint arXiv:2102.04830, 2021. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 546, + 287, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 546, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 546, + 287, + 590 + ], + "type": "text", + "content": "[37] Amir Zadeh, Minghai Chen, Soujanya Poria, Erik Cambria, and Louis-Philippe Morency. Tensor fusion network for multimodal sentiment analysis. arXiv preprint arXiv:1707.07250, 2017. 3, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 591, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 591, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 591, + 287, + 635 + ], + "type": "text", + "content": "[38] Amir Zadeh, Rowan Zellers, Eli Pincus, and Louis-Philippe Morency. Mosi: multimodal corpus of sentiment intensity and subjectivity analysis in online opinion videos. arXiv preprint arXiv:1606.06259, 2016. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 636, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 287, + 712 + ], + "type": "text", + "content": "[39] AmirAli Bagher Zadeh, Paul Pu Liang, Soujanya Poria, Erik Cambria, and Louis-Philippe Morency. Multimodal language analysis in the wild: Cmu-mosei dataset and interpretable dynamic fusion graph. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2236–2246, 2018. 5" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 363 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[40] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412, 2017. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 107, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 162 + ], + "type": "text", + "content": "[41] Tong Zhang, Di Wang, Huanhuan Chen, Zhiwei Zeng, Wei Guo, Chunyan Miao, and Lizhen Cui. Bdann: Bert-based domain adaptation neural network for multi-modal fake news detection. In 2020 International Joint Conference on Neural Networks (IJCNN), pages 1-8, 2020. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 163, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 205 + ], + "type": "text", + "content": "[42] Weichen Zhang, Dong Xu, Jing Zhang, and Wanli Ouyang. Progressive modality cooperation for multi-modality domain adaptation. IEEE Transactions on Image Processing, 30:3293-3306, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 207, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 207, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 207, + 545, + 262 + ], + "type": "text", + "content": "[43] Yiyi Zhang, Li Niu, Ziqi Pan, Meichao Luo, Jianfu Zhang, Dawei Cheng, and Liqing Zhang. Exploiting motion information from unlabeled videos for static image action recognition. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 12918-12925, 2020. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 263, + 545, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 263, + 545, + 317 + ], + "spans": [ + { + "bbox": [ + 307, + 263, + 545, + 317 + ], + "type": "text", + "content": "[44] Honghao Zhou, Tinghuai Ma, Huan Rong, Yurong Qian, Yuan Tian, and Najla Al-Nabhan. Mdmn: Multi-task and domain adaptation based multi-modal network for early rumor detection. Expert Systems with Applications, 195:116517, 2022. 1, 2, 5, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 319, + 545, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 319, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 307, + 319, + 545, + 363 + ], + "type": "text", + "content": "[45] Junbao Zhuo, Shuhui Wang, Weigang Zhang, and Qingming Huang. Deep unsupervised convolutional domain adaptation. In Proceedings of the 25th ACM international conference on Multimedia, pages 261-269, 2017. 2, 3" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3560" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/OSRT_ Omnidirectional Image Super-Resolution With Distortion-Aware Transformer/b6aea18d-97cd-464e-a4ad-626c19fdc060_content_list.json b/2023/OSRT_ Omnidirectional Image Super-Resolution With Distortion-Aware Transformer/b6aea18d-97cd-464e-a4ad-626c19fdc060_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..38b7ddc015ce08cde9d731bdc47f9599727ab5f6 --- /dev/null +++ b/2023/OSRT_ Omnidirectional Image Super-Resolution With Distortion-Aware Transformer/b6aea18d-97cd-464e-a4ad-626c19fdc060_content_list.json @@ -0,0 +1,2170 @@ +[ + { + "type": "text", + "text": "OSRT: Omnidirectional Image Super-Resolution with Distortion-aware Transformer", + "text_level": 1, + "bbox": [ + 203, + 138, + 763, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Fanghua Yu $^{1*}$ Xintao Wang $^{2*}$ Mingdeng Cao $^{2,3}$ Gen Li $^{4}$ Ying Shan $^{2}$ Chao Dong $^{1,5\\dagger}$ $^{1}$ ShenZhen Key Lab of Computer Vision and Pattern Recognition \nShenzhen Institute of Advanced Technology, Chinese Academy of Sciences \n $^{2}$ ARC, Tencent PCG ${}^{3}$ The University of Tokyo", + "bbox": [ + 102, + 179, + 870, + 245 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4 Platform Technologies, Tencent Online Video Shanghai Artificial Intelligence Laboratory", + "bbox": [ + 104, + 247, + 863, + 264 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "fanghuayu96@gmail.com, xintaowang@tencent.com, cmd@g.ecc.u-tokyo.ac.jp", + "bbox": [ + 164, + 266, + 801, + 279 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{genli, yingsshan}@tencent.com, chao.dong@siat.ac.cn", + "bbox": [ + 247, + 282, + 719, + 296 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 226, + 319, + 307, + 333 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Omnidirectional images (ODIs) have obtained lots of research interest for immersive experiences. Although ODIs require extremely high resolution to capture details of the entire scene, the resolutions of most ODIs are insufficient. Previous methods attempt to solve this issue by image super-resolution (SR) on equirectangular projection (ERP) images. However, they omit geometric properties of ERP in the degradation process, and their models can hardly generalize to real ERP images. In this paper, we propose Fisheye downsampling, which mimics the real-world imaging process and synthesizes more realistic low-resolution samples. Then we design a distortion-aware Transformer (OSRT) to modulate ERP distortions continuously and self-adaptively. Without a cumbersome process, OSRT outperforms previous methods by about 0.2dB on PSNR. Moreover, we propose a convenient data augmentation strategy, which synthesizes pseudo ERP images from plain images. This simple strategy can alleviate the over-fitting problem of large networks and significantly boost the performance of ODISR. Extensive experiments have demonstrated the state-of-the-art performance of our OSRT.", + "bbox": [ + 63, + 342, + 473, + 642 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 65, + 661, + 200, + 676 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In pursuit of the realistic visual experience, omnidirectional images (ODIs), also known as $360^{\\circ}$ images or panoramic images, have obtained lots of research interest in the computer vision community. In reality, we usually view ODIs with a narrow field-of-view (FOV), e.g., viewing in a headset. To capture details of the entire scene, ODIs require extremely high resolution, e.g., $4\\mathrm{K} \\times 8\\mathrm{K}$ [1]. However, due to the high industrial cost of camera sensors with high precision, the resolutions of most ODIs are insufficient.", + "bbox": [ + 63, + 686, + 468, + 813 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, some attempts have been made to solve this problem by image super-resolution (SR) [12, 15, 28, 39, 40].", + "bbox": [ + 63, + 814, + 468, + 843 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5aadb5dd34e8dc3d24b12e025b3c1637a487761ffde252181d26af69da3856a5.jpg", + "image_caption": [ + "Unseen LR" + ], + "image_footnote": [], + "bbox": [ + 510, + 311, + 697, + 410 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b9145473a0eb5e8a56509988face13f4e65bd2a193ad7d1b3a0a1f8980421b57.jpg", + "image_caption": [ + "LAU-Net [12] w/o Fisheye" + ], + "image_footnote": [], + "bbox": [ + 712, + 311, + 897, + 410 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ae3130b9f69db040cc2b88220b207a6a8af801674510da05fa5bff35b377ed7e.jpg", + "image_caption": [ + "OSRT w/o Fisheye" + ], + "image_footnote": [], + "bbox": [ + 510, + 427, + 697, + 527 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b255ba70f6bcacee07d49fa892047a8c4a2f0b7753dbac13d64c8af1dac46e81.jpg", + "image_caption": [ + "OSRT w/ Fisheye", + "Figure 1. Visual comparisons of $\\times 8$ SR results on LR images1 with unknown degradations. Fisheye denotes that the downsampling process in training stages is under Fisheye images." + ], + "image_footnote": [], + "bbox": [ + 710, + 426, + 897, + 527 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As most of the ODIs are stored and transmitted in the equirectangular projection (ERP) type, the SR process is usually performed on the ERP images. To generate high- /low-resolution training pairs, existing ODISR methods [12, 15, 28, 39, 40] directly apply uniform bicubic down-sampling on the original ERP images (called ERP down-sampling), which is identical to general image SR settings [24, 43]. While omitting geometric properties of ERP in the degradation process, their models can hardly generalize to real ERP images. We can observe missing structures and blur textures in Fig. 1. Therefore, we need a more appropriate degradation model before studying SR algorithms. In practice, ODIs are acquired by the fisheye lens and stored in ERP. Given that the low-resolution issue in real-world scenarios is caused by insufficient sensor precision and density, the downsampling process should be applied to original-formatted images before converting into other storage types. Thus, to be conformed with real-world imaging processes, we propose to apply uniform bicubic", + "bbox": [ + 497, + 581, + 904, + 853 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 119, + 36, + 189, + 73 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 245, + 35, + 800, + 76 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution", + "bbox": [ + 82, + 853, + 194, + 863 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author (e-mail: chao.dong@siat.ac.cn)", + "bbox": [ + 82, + 864, + 381, + 876 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Photoed by Peter Leth on Flickr, with CC license.", + "bbox": [ + 517, + 863, + 794, + 876 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "13283", + "bbox": [ + 465, + 891, + 504, + 903 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "downsampling on Fisheye images, which are the original format of ODIs. The new downsampling process (called Fisheye downsampling) applies uniform bicubic downsampling on Fisheye images before converting them to ERP images. Our Fisheye downsampling is more conducive to exploring the geometric property of ODIs.", + "bbox": [ + 63, + 114, + 468, + 201 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The key issue of ODISR algorithm design is to utilize the geometric properties of ERP images, which is also the focus of previous methods. For example, Nishiyama et al. [28] add a distortion-related condition as an additional input. LAU-Net [12] splits the whole ERP image into patches by latitude band and learns upscaling processes separately. However, the separated learning process will lead to information disconnection between adjacent patches. SphereSR [40] learns different upscaling functions on various projection types, but will inevitably introduce multiple-time computation costs. To push the performance upper bound, we propose the first Transformer for Omnidirectional image Super-Resolution (OSRT), and incorporate geometric properties in a distortion-aware manner. Specifically, to modulate distorted feature maps, we implement feature-level warping, in which offsets are learned from latitude conditions. In OSRT, we introduce two dedicated blocks to adapt latitude-related distortion: distortion-aware attention block (DAAB), and distortion-aware convolution block (DACB). DAAB and DACB are designed to perform distortion modulation in arbitrary Transformers and ConvNets. These two blocks can directly replace the multi-head self-attention block and convolution layer, respectively. The benefit of DAAB and DACB can be further improved when being inserted into the same backbone network. OSRT outperforms previous methods by about 0.2dB on PSNR (Tab. 2).", + "bbox": [ + 63, + 203, + 470, + 572 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, the increase of network capacity will also enlarge the overfitting problem of ODISR, which is rarely mentioned before. The largest ODIs dataset [12] contains only 1K images, which cannot provide enough diversity for training Transformers. Given that acquiring ODIs requires expensive equipment and tedious work, we propose to generate distorted ERP samples from plain images for data augmentation. In practice, we regard a plain image as a sampled perspective, and project it back to the ERP format. Then we can introduce 146K additional training patches, 6 times of the previous dataset. This simple strategy can significantly boost the performance of ODISR (Tab. 4) and alleviate the over-fitting problem of large networks (Fig. 9). A similar data augmentation method is also applied in Nishiyama et al. [28], but shows marginal improvement on small models under ERP downsampling settings.", + "bbox": [ + 63, + 574, + 468, + 804 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are threefold. 1) For problem formulation: To generate more realistic ERP low-resolution images, we propose Fisheye downsampling, which mimics the real-world imaging process. 2) For method: Combined with the geometric properties of ERP, we design a distortion-aware", + "bbox": [ + 63, + 805, + 468, + 878 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Transformer, which modulates distortions continuously and self-adaptively without cumbersome process. 3) For data: To reduce overfitting, we propose a convenient data augmentation strategy, which synthesizes pseudo ERP images from plain images. Extensive experiments have demonstrated the state-of-the-art performance of our OSRT2.", + "bbox": [ + 497, + 114, + 904, + 203 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 499, + 216, + 645, + 231 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Single Image Super-Resolution (SISR). Deep learning for single image SR (SISR) is first introduced in [13]. Further works boost SR performance by CNNs [11, 14, 22, 24, 26, 29, 43], Vision Transformers (ViTs) [7, 8, 21, 23] and generative adversarial networks (GANs) [20, 35, 36, 42]. For instance, EDSR [24] removes Batch Normalization layers and applies a more complicated residual block. RCAN [43] introduces channel-wise attention mechanisms to a deeper network. SwinIR [23] proposes an image restoration Transformer based on [25]. To improve perceptual quality, adversarial training are performed as a tuning process to generate more realistic results [35, 36]. Moreover, various flexible degradation models are proposed in [35, 41] to synthesize more practical degradations.", + "bbox": [ + 497, + 241, + 904, + 440 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Omnidirectional Image Super-Resolution (ODISR). Initially, ODISR models focus on the spherical assembling of LR ODIs under various projection types [2-4, 17, 27]. Recent ODISR models are performed on plane images and are fine-tuned from existing SISR models with L1 loss [15] or GAN loss [30, 44]. The improvements are limited, for they only concern the distribution gap between ODIs and plain images. Since LAU-Net [12] found pixel density in ERP ODIs is non-uniform, many studies attempt to design specific backbone networks to overcome this issue. LAU-Net [12] manually splits the whole ERP image into latitude-related patches and learns ERP distortion over different latitude ranges separately. While LAU-Net learns latitude-related ERP distortion somewhat, its non-overlapped patches lead to disconnection in whole ERP images. Nishiyama et al. [28] treats area stretching ratio as additional input. However, these conditions are tough to be utilized with an unmodified SISR backbone network. SphereSR [40] learns upsampling processes on various projection types to mitigate the influence of non-uniformity in specific projection types. Although SphereSR improves information consistency between various ODI projection types by LIIF [9], they apply multiple networks to learn the upscaling process of each projection type. Given that all other projection types in SphereSR are converted from ERP, patterns under various types are reusable when distortions are properly rectified. Moreover, the complex and unstructured image data in polyhedron projection hinders further research of ODISR.", + "bbox": [ + 497, + 442, + 904, + 853 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "$^{2}$ https://github.com/Fanghua-Yu/OSRT", + "bbox": [ + 517, + 863, + 732, + 876 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "13284", + "bbox": [ + 465, + 891, + 505, + 903 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Deformable Mechanism. Dai et al. [30] first propose deformable convolutions to obtain information out of its regular neighborhood. Xia et al. [38] further verified that Vision Transformers also benefit from applying deformable mechanisms on self-attention blocks. In Video SR tasks, the deformable mechanism can be adapted to align features between adjacent frames [5,6,34].", + "bbox": [ + 63, + 114, + 468, + 214 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 65, + 226, + 159, + 242 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first analyze the cause of ERP and Fisheye distortions, as well as the relationship between these two distortions (Sec. 3.1). Then, we discuss the designs of Fisheye downsampling (Sec. 3.2), distortion-aware Transformer (OSRT) (Sec. 3.3), and the convenient data augmentation strategy (Sec. 3.4).", + "bbox": [ + 63, + 250, + 468, + 336 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Revisiting Distortions in ODIs", + "text_level": 1, + "bbox": [ + 65, + 344, + 341, + 359 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/fb8333e243e50452470bb4f9fe79baddd5ab63b25639d7b4b09b9ed9068e956d.jpg", + "image_caption": [ + "Figure 2. Geometric explanation of transforming between ERP, Fisheye, and the ideal spherical surface. To simplify, we discuss the horizontal spliced Fisheye with an aperture degree of $\\pi$ ." + ], + "image_footnote": [], + "bbox": [ + 72, + 370, + 468, + 559 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As ODIs under each projection type are constrained by different transforming equations, the distortion caused by each type is inconsistent, indicating that applying matrix operations under one projection type can introduce unexpected changes when being converted to other types. Specifically, applying uniformed bicubic downsampling on ERP images will affect the distribution of pixel density on Fisheye images, which are the original-formatted image type of imaging process in real-world scenarios. To analyze the specific effect of ERP downsampling on the Fisheye image, we revisit the cause of distortions in ERP and Fisheye.", + "bbox": [ + 63, + 606, + 468, + 762 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As we assume that viewing directions are uniformly distributed, the data points in an ideal ODI should be uniformly distributed on a spherical surface. In practice, there is a trade-off between the uniformity of the spherical surface and the structural degree. ERP is the most convenient projection type for storage or transmission, but it is also the projection type that suffers the heaviest distortion. To better explain the causation of distortions, we follow the definition", + "bbox": [ + 63, + 763, + 468, + 876 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "of stretching ratio $(\\mathbf{K})$ in [31], which represents distortion degree at different locations from the target projection type to the ideal spherical surface. $\\mathbf{K}$ is determined by area variation from one projection type to another. When the target type is uniforming spherical surface, $\\mathbf{K}$ is defined as:", + "bbox": [ + 497, + 114, + 904, + 187 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {K} (x, y) = \\frac {\\delta S (\\theta , \\varphi)}{\\delta P (x , y)} = \\frac {\\cos (\\varphi) | d \\theta d \\varphi |}{| d x d y |} = \\frac {\\cos (\\varphi)}{| J (\\theta , \\varphi) |}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 512, + 194, + 902, + 226 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\delta S(\\cdot ,\\cdot)$ and $\\delta P(\\cdot ,\\cdot)$ represent the area on the spherical surface and the projection plane, respectively. $|didj|$ represents plane microunit. $|J(\\theta ,\\varphi)|$ is the Jacobian determinant from spherical coordinate to projection coordinate.", + "bbox": [ + 497, + 235, + 902, + 292 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ERP distortion. The coordinate in ERP is defined as $x = \\theta$ and $y = \\varphi$ . ERP stretching ratio can be derived as:", + "bbox": [ + 497, + 292, + 902, + 321 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {K} _ {\\mathrm {E R P}} (x, y) = \\cos (\\varphi) = \\cos (y), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 331, + 902, + 347 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $x\\in (-\\pi ,\\pi)$ $y\\in (-\\frac{\\pi}{2},\\frac{\\pi}{2})$", + "bbox": [ + 497, + 356, + 732, + 372 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "From Eq. (2), we conclude that ERP distortion is only determined by its latitude degree. $\\mathbf{K}_{\\mathrm{ERP}}$ is reduced to zero when the absolute value of latitude degree increases to $\\pi /2$ , which represents that pixel density on the polar areas of ERP images is closer to zero. As shown in Fig. 2 (c), with the increasing of the absolutely value of latitude degree $(|\\varphi_{s}|)$ , the corresponding area on the spherical surface of an ERP microunit is gradually decreased to zero. In conclusion, ERP distortion is caused by variable stretching ratios $\\mathbf{K}_{\\mathrm{ERP}}$ , and is the heaviest in the polar areas.", + "bbox": [ + 497, + 371, + 902, + 513 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Fisheye distortion. The coordinate in Fisheye can be derived from $\\theta = \\arctan \\left(\\frac{y}{x}\\right)$ and $\\varphi = (1 - \\sqrt{x^2 + y^2})\\times \\frac{\\pi}{2}$ . The stretching ratios of Fisheye can be derived as:", + "bbox": [ + 497, + 513, + 902, + 555 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {K} _ {\\text {F i s h e y e}} (x, y) = \\frac {\\frac {2}{\\pi} \\sin \\left(\\frac {\\pi}{2} \\sqrt {x ^ {2} + y ^ {2}}\\right)}{\\sqrt {x ^ {2} + y ^ {2}}}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 564, + 902, + 602 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\sqrt{x^2 + y^2} \\in (0,1)$ .", + "bbox": [ + 497, + 611, + 682, + 627 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$\\mathbf{K}_{\\mathrm{Fisheye}}$ is determined by distance from the fisheye center. As $(\\mathbf{K}_{\\mathrm{Fisheye}})^{-1}$ is bounded, fisheye projection is closer to uniform distribution than ERP. Moreover, it introduces much slighter distortion at the polar.", + "bbox": [ + 497, + 627, + 902, + 684 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Relationship between ERP and Fisheye distortions. To simplify, here we only discuss a typical Fisheye with an aperture degree of $\\pi$ and a horizontal slicing plane4. In this case, the ERP coordinates and Fisheye's polar coordinates correspond linearly. We can quantize the relationship by:", + "bbox": [ + 497, + 684, + 902, + 755 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {K} _ {\\mathrm {E R P} \\mid \\text {F i s h e y e}} (\\theta , \\varphi) = \\frac {\\mathbf {K} _ {\\mathrm {E R P}} \\left(x _ {E} , y _ {E}\\right)}{\\mathbf {K} _ {\\text {F i s h e y e}} \\left(x _ {F} , y _ {F}\\right)} = \\frac {\\pi}{2} - | \\varphi |, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 507, + 764, + 902, + 796 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\theta, \\varphi$ are spherical coordinates on the sphere, $x_{E}, y_{E}$ ( $x_{F}, y_{F}$ ) denotes the plain coordinate under ERP (Fisheye).", + "bbox": [ + 497, + 804, + 902, + 833 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "3Detailed derivative processes can be found in the supplementary file.", + "bbox": [ + 515, + 840, + 895, + 853 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "4The influence of Fisheye formats with arbitrary splicing plane is discussed in the supplementary file.", + "bbox": [ + 500, + 853, + 900, + 876 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "13285", + "bbox": [ + 465, + 891, + 504, + 902 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/7b63ac2ee9f66d49a42ce1663d2e22e5d7a14e6a3c7436c95582879c443377da.jpg", + "image_caption": [ + "Figure 3. Downsampling process of ODIs (left) and imaging process in real world (right). * denotes that LR images synthesized from different downsampling processes are inconsistent." + ], + "image_footnote": [], + "bbox": [ + 68, + 114, + 460, + 264 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "From Eqs. (2) to (4), we conclude that when uniformed downsampling is performed on ERP, the kernel size of equivalent Fisheye downsampling is non-uniformed. Especially when fisheye projection is spliced horizontally, the kernel size is proportional with $\\pi / 2 - |\\varphi|$ .", + "bbox": [ + 63, + 319, + 468, + 390 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Learning with More Realistic Degradation", + "text_level": 1, + "bbox": [ + 65, + 399, + 436, + 414 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As depicted in Fig. 3, the original-formatted projection type in ODI acquiring process is fisheye projection. Given that real-world low-resolution issues are caused by insufficient precision and density of sensors, we consider that the degeneration process should be directly applied to original-formatted images before the type conversion.", + "bbox": [ + 63, + 421, + 467, + 505 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Ideally, as camera sensors are arranged in uniform arrays, pixel density on original-formatted images is consistent everywhere. Thus, for a realistic ODI, the pixel density on Fisheye should be a constant. As discussed in Sec. 3.1, applying uniformed downsampling on ERP means applying downsampling of variable kernel size on Fisheye. The variable kernel size leads to variable Fisheye pixel density, which results in unrealistic LR images. In conclusion, the ERP downsampling in previous methods influences the intrinsic distribution of pixel density in original-formatted images, which leads to unrealistic ODIs. When the downsampling process happens on Fisheye, the Fisheye pixel density is unchanged, which fits the real-world imaging process and synthesizes more realistic LR pairs.", + "bbox": [ + 63, + 507, + 467, + 705 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Process of Fisheye downsampling. To generate more realistic LR ODIs, we mimic the real-world imaging process and apply bicubic downsampling on Fisheye images. One single Fisheye image can only store information about a hemisphere. Hence, ERP images are converted to dual Fisheye images. Before downsampling, Fisheye images are padded by a FOV larger than $180^{\\circ}$ to avoid edge disconnections. This padding operation will not influence the geometric transforming relation between ERP and Fisheye. As Fisheye data is unstructured and Fisheye distortion is more complicated than ERP distortion, we still learn the upscaling process under ERP. Thus we reconvert LR images to the", + "bbox": [ + 63, + 706, + 467, + 876 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ERP format. The overall process of Fisheye downsampling are described in Fig. 3.", + "bbox": [ + 499, + 115, + 900, + 143 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.OSRT:Modulate Distortion in ODIs", + "text_level": 1, + "bbox": [ + 499, + 152, + 823, + 165 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Overall. As discussed in Sec. 3.1, ERP images suffer a distortion caused by a non-consistency area stretching ratio from an ideal spherical surface. Referred from Eq. (2), for an LR input $X_{i}\\in \\mathbb{R}^{C\\times M\\times N}$ , the distortion map $C_d\\in \\mathbb{R}^{1\\times M\\times N}$ is derived by:", + "bbox": [ + 499, + 174, + 900, + 245 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nC _ {d} = \\cos \\left(\\frac {m + 0 . 5 - M / 2}{M} \\pi\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 254, + 900, + 285 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $m$ is the current height of LR input.", + "bbox": [ + 499, + 294, + 786, + 308 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Previous methods tend to treat $C_d$ as an additional input of $X_i$ [28], or re-weighting parameters by $C_d$ [18]. Although these solutions can benefit from building awareness of distortion, continuous and amorphous distortions cannot be adequately fitted by scattering and structured convolution operations. While previous methods cannot fully explore the advantage of $C_d$ , we intend to design a novel block for learning distorted patterns continuously. In VSR tasks, the deformable mechanism is proposed to align features between adjacent frames [32, 34]. Unlike standard DCN [10], which calculates offsets from the input feature map, offsets are calculated from bi-directional optical flow in VSR pipelines. Inspired by feature-level flow warping in VSR, we find that the deformable mechanism is a feasible solution for continuous mappings. Consequently, we modulate ERP distortion by feature-level warping operations. As shown in Fig. 4, $C_d$ is only utilized to calculate the deformable offsets $\\theta$ . To keep compatibility with arbitrary ConvNets and Transformers, we propose two blocks to modulate ERP distortion, which can directly replace the multi-head self-attention blocks in Transformers and the standard convolution layers in ConvNets, respectively.", + "bbox": [ + 497, + 308, + 902, + 619 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Distortion-aware attention block (DAAB). As depicted in Fig. 4 (a), a distortion condition guided deformable self-attention is proposed to learn correlations between the distorted input $X_{i}$ and its corresponding modulated feature map $\\tilde{X}_{i}$ . DAAB is formulated as:", + "bbox": [ + 497, + 621, + 902, + 691 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {i} = H _ {\\text {o f f s e t} _ {i}} \\left(C _ {d}, C _ {w}\\right), \\tilde {X} _ {i} = \\phi \\left(X _ {i}; p _ {0} + \\theta_ {i}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 544, + 701, + 900, + 717 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nX _ {o} = H _ {\\mathrm {S A}} \\left(X _ {i} W _ {q _ {i}}, \\tilde {X} _ {i} W _ {k _ {i}}, \\tilde {X} _ {i} W _ {v _ {i}}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 726, + 900, + 743 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $H_{\\mathrm{offset}_i}(\\cdot)$ denotes the $i$ -th convolution block to calculate offset maps $\\theta_i \\in \\mathbb{R}^{2 \\times H \\times W}$ , and $H_{\\mathrm{SA}}$ denotes standard self-attention formula. $H_{\\mathrm{offset}}(\\cdot)$ consists of $1 \\times 1$ convolution block with two hidden layers. The input of $H_{\\mathrm{offset}}(\\cdot)$ is concatenated by the latitude-related distortion condition $C_d \\in \\mathbb{R}^{1 \\times H \\times W}$ and the window condition $C_w \\in \\mathbb{R}^{2 \\times H \\times W}$ . $C_w$ is a linear position encoding within a self-attention kernel. $\\phi(\\cdot, \\cdot)$ denotes a bilinear interpolation, and $W_{q_i}, W_{k_i}, W_{v_i}$ denote $i$ -th weight matrix of query, key, and", + "bbox": [ + 497, + 748, + 902, + 876 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "13286", + "bbox": [ + 465, + 891, + 504, + 902 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9a56a726695fd45479635425fdb85149a3f247dcad2883417c18b0ba871511bf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 82, + 104, + 890, + 223 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/72c5be446d0de73ba147472880a1c4c2d735a796168cf0b389d1d4d73f1f7b1e.jpg", + "image_caption": [ + "Figure 4. Overall illustration of OSRT. From SwinIR [23], we replace the standard multi-head self-attention block with DAAB and insert DACB behind the end of the RSTB. Channel dimensions of $\\theta_{\\mathrm{offsets}}$ in DAAB and DACB are 2 and 18, respectively." + ], + "image_footnote": [], + "bbox": [ + 75, + 225, + 507, + 378 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a884f0fbde197c90940bbf89f5d5f09dd47afaa086cf5514ba23ee5e36df7a20.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 512, + 225, + 727, + 380 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/154b285d4c54c969b4bd95e5610f05814e7fd4b54cb468d4d98ec941b8f4e740.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 731, + 225, + 889, + 380 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "value, respectively. For multi-head self-attention blocks, $H_{\\mathrm{offset}_i}(\\cdot)$ is identical in calculations of parallel heads.", + "bbox": [ + 65, + 418, + 465, + 446 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Distortion-aware convolution block (DACB). As shown in Fig. 4 (b), we apply a standard deformable convolution layer with a substituted input for offset calculation. Modulated output $X_{o}$ is extracted as:", + "bbox": [ + 63, + 447, + 467, + 503 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {i} = H _ {\\text {o f f s e t} _ {i}} \\left(C _ {d}\\right), X _ {o} = H _ {\\mathrm {D C N} _ {i}} \\left(X _ {i}, \\theta_ {i}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 122, + 511, + 467, + 527 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $H_{\\mathrm{DCN}}(X,\\theta)$ denotes standard deformable convolution layer in [46]. The architecture of $H_{\\mathrm{offset}_i}(\\cdot)$ is identical to that in DAAB. As the kernel size of DCN is $3\\times 3$ in DACB, the output channel dimension of offsets maps is 18.", + "bbox": [ + 63, + 534, + 467, + 591 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "OSRT. In practice, we propose an Omnidirectional image Super-Resolution Transformer, named OSRT. SwinIR [23] is selected as the basic architecture for its strong reconstruction ability in the SISR task. To learn distortion rectified representations, we stack a DACB after the last convolution layer of each residual swin Transformer block and replace all self-attention blocks as DAAB. The feature dimension of OSRT is reduced from 180 to 156 to maintain identical parameters with SwinIR.", + "bbox": [ + 63, + 591, + 467, + 719 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Boosting ODISR Performance by Plain Images", + "text_level": 1, + "bbox": [ + 65, + 726, + 467, + 742 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As the capacity of OSRT is relatively large, it suffers overfitting for large upscaling factors (Fig. 9). Given that acquiring ODIs are expensive, we propose to generate pseudo ERP images from 2D plain images to tackle this issue. After being sampled by sliding windows, the patch of plain images is treated as a plain perspective. By converting from Perspective to ERP, plain images are distorted in the same way as ERP. Considering that distortion of a Perspective is enlarged by its FOV degree, a relatively small FOV", + "bbox": [ + 63, + 749, + 467, + 876 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "degree of $90^{\\circ}$ is applied. For a given pseudo Perspective, $\\theta_{p}$ is fixed at 0 and $\\varphi_{p}$ is derived by:", + "bbox": [ + 497, + 418, + 900, + 448 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Phi_ {p} = \\varphi_ {h} + z _ {0}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 647, + 461, + 900, + 478 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\varphi_h$ is determined by patch locations and $z_0$ is orderly sampled from $\\{-15^\\circ, 0^\\circ, 15^\\circ\\}$ .", + "bbox": [ + 497, + 491, + 900, + 520 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To maximize the approximate data distribution of ODIs, we horizontally split a plain image into three sub-images and define $\\varphi_h$ as $-30^\\circ, 0^\\circ, 30^\\circ$ respectively. Pseudo ERP images are cropped to remove the black border. As shown in Fig. 5, we get a new ERP dataset (called DF2K-ERP) by implementing the augmentation pipeline on widely-used plain image dataset DF2K [24, 33]. The DF2K-ERP dataset consists of 146K high-quality ERP image patches with a patch size larger than 256.", + "bbox": [ + 497, + 521, + 902, + 650 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c1511d336fd6d67d2883beed07de61b73a9cf216ef514f4c0641c079d11f14f5.jpg", + "image_caption": [ + "Figure 5. Synthetic process of DF2K-ERP." + ], + "image_footnote": [], + "bbox": [ + 502, + 661, + 895, + 852 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "13287", + "bbox": [ + 465, + 891, + 504, + 902 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/2dba8417b1c2eef86537e9b579bebb779cb3a310fdab95e94001c7905e89784c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodScaleODI-SRPSNRSUN 360 Panorama
PSNRSSIMWS-PSNRWS-SSIMSSIMWS-PSNRWS-SSIM
Bicubic28.210.821527.610.815628.140.811828.010.8321
RCAN [43]30.080.872329.490.871430.560.871231.180.8969
SRResNet [36]30.160.871729.590.869730.650.871431.200.8953
EDSR [24]×230.320.877029.680.872730.890.878431.420.8995
SwinIR [23]30.520.881929.870.877231.210.885231.780.9051
SwinIR†[23]30.640.882130.000.877731.330.885531.980.9059
OSRT†30.770.884630.110.879531.520.888832.140.9081
Bicubic25.590.711824.950.692325.290.699324.900.7083
RCAN [43]26.850.762126.150.748527.100.766026.990.7856
SRResNet [36]26.910.759726.240.745727.100.761826.990.7812
EDSR [24]×426.970.758926.300.745827.190.763327.100.7827
SwinIR [23]27.120.766326.440.752327.390.770727.300.7901
SwinIR†[23]27.310.773526.610.758927.710.780427.640.7996
OSRT†27.410.776226.700.760927.840.783527.770.8020
", + "bbox": [ + 134, + 112, + 831, + 289 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. SR results under Fisheye downsampling. † denotes applying DF2K-ERP as augmented dataset. Best results are shown in Bold.", + "bbox": [ + 72, + 292, + 895, + 307 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c932d1fb890ac706ad04dc32f9289fcb0ce311fa8652d4bd663864d44ac46f38.jpg", + "image_caption": [ + "SUN360 $(\\times 4)$ : 034" + ], + "image_footnote": [], + "bbox": [ + 78, + 313, + 299, + 469 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b7df64b1a03e7f8a36ff41abe586ecaa72ca27ccbbbf0899d86d9fe1d3913914.jpg", + "image_caption": [ + "HR" + ], + "image_footnote": [], + "bbox": [ + 314, + 312, + 442, + 381 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f9475fdab31f0999131642a7793161d6eab43d3d887afc1e31b6d85db4da8f12.jpg", + "image_caption": [ + "PSNR/SSIM", + "EDSR [24]" + ], + "image_footnote": [], + "bbox": [ + 314, + 400, + 442, + 467 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/aab48792041743755ca62115a4c4922972a25ffec0420064182fff9ccf5cc5c4.jpg", + "image_caption": [ + "Bicubic" + ], + "image_footnote": [], + "bbox": [ + 457, + 312, + 584, + 381 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7015d250894c2c694b918b664e9cbb5a0fdb672a2a087fce115626fc1c18f96e.jpg", + "image_caption": [ + "24.38dB/0.6872", + "SwinJR [23]" + ], + "image_footnote": [], + "bbox": [ + 457, + 401, + 583, + 467 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0acdf7d88bfcb6b4afc4eab6d1201a356f5846e79a01346926f7b16b25ac6c5c.jpg", + "image_caption": [ + "RCAN [43]" + ], + "image_footnote": [], + "bbox": [ + 600, + 312, + 727, + 381 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/481744445b0f343e23c06bcce4ed9331562c5e7d171db802b6c79cd1ca07a558.jpg", + "image_caption": [ + "26.40dB/0.8137", + "SwinIR† [23]" + ], + "image_footnote": [], + "bbox": [ + 600, + 401, + 727, + 467 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8477afb1d91308b13e0c72f6ce1f6ca8850546571d439b61c17ab62270f44d79.jpg", + "image_caption": [ + "SRResNet [36]" + ], + "image_footnote": [], + "bbox": [ + 742, + 312, + 868, + 381 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4815ee82c3c4c7985323b1e9e0aabe47ead9697211ab2aa4f21f51d45cb9ff57.jpg", + "image_caption": [ + "26.21dB/0.7999", + "OSRT†" + ], + "image_footnote": [], + "bbox": [ + 742, + 401, + 868, + 467 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3bffb7b07707ac3d5c03dc456cbf1a86803f201e902c9c43ec23d3ff454e4205.jpg", + "image_caption": [ + "SUN360 $(\\times 4)$ : 095", + "Figure 6. Visual comparisons of $\\times 4$ SR results under Fisheye downsampling." + ], + "image_footnote": [], + "bbox": [ + 78, + 495, + 299, + 650 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d4f1a7a11ae19a88ec5eefd9b6e5bb209f898b01ae13657c1b6232d8ab83d34b.jpg", + "image_caption": [ + "26.38dB/0.8072", + "HR" + ], + "image_footnote": [], + "bbox": [ + 315, + 495, + 442, + 562 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1a1d5efe555b1f5035ecc94306435532e175fa4e6656e29c696c66d460bff762.jpg", + "image_caption": [ + "PSNR/SSIM", + "EDSR [24]", + "33.64dB/0.9074" + ], + "image_footnote": [], + "bbox": [ + 314, + 584, + 442, + 650 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/df19881ea35c55f6ef96db6fcb7f3ccacb62078b221a2cfdaab5f58e5a22304a.jpg", + "image_caption": [ + "26.77dB/0.8234", + "Bicubic" + ], + "image_footnote": [], + "bbox": [ + 457, + 495, + 584, + 562 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c54e91b8db836f67d26c9e9067e41436ab9c905563cd77742ff61876f914e714.jpg", + "image_caption": [ + "30.20dB/0.8506", + "SwinIR [23]", + "34.05dB/0.9119" + ], + "image_footnote": [], + "bbox": [ + 457, + 584, + 584, + 650 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1e3682281a265e04e93b2dba4f4cec276c1167bb85a93b03f68b256eea66dd1d.jpg", + "image_caption": [ + "27.34dB/0.8462", + "RCAN [43]" + ], + "image_footnote": [], + "bbox": [ + 598, + 495, + 727, + 562 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/bdfb984ea97a83e001e31d4a1f23bc3afb561016c77dcf1ec9ce44a13b1e5345.jpg", + "image_caption": [ + "33.59dB/0.9088", + "SwinIR [23]", + "34.41dB/0.9158" + ], + "image_footnote": [], + "bbox": [ + 598, + 583, + 727, + 650 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1448538722ba337494966ef75c9fbf0d4bc040e329acc5140d8213ccd5bff992.jpg", + "image_caption": [ + "27.68dB/0.8561", + "SRResNet [36]" + ], + "image_footnote": [], + "bbox": [ + 742, + 495, + 868, + 562 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b350bdce5d2de18e8b6e05b88ebcdc8e43d8ae5cadca5f2dfd3f96e718b7dd4c.jpg", + "image_caption": [ + "33.43dB/0.9043", + "OSRT†", + "34.77dB/0.9187" + ], + "image_footnote": [], + "bbox": [ + 742, + 583, + 868, + 650 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 65, + 698, + 201, + 713 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 65, + 724, + 258, + 739 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ODI-SR dataset [12] and SUN360 Panorama dataset [39] are used in our experiment. In the training phase, we follow the data split setting in [12] and train on the ODI-SR training set. The resolution of the ERP HR is $1024 \\times 2048$ , and the upscaling factors are $\\times 2$ and $\\times 4$ . Fisheye downsampling is applied as our pre-defined downsampling kernel. Loss is calculated by L1 distance and optimized by Adam [19], with an initial learning rate of $2 \\times 10^{-4}$ , a total batch size of 32, and an input patch size of 64. We", + "bbox": [ + 63, + 748, + 468, + 876 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "train OSRT for $500k$ iterations and halve the learning rate at $250k$ , $400k$ , $450k$ and $475k$ . In evaluation, we test on the ODI-SR testing set and SUN360 dataset. PSNR [16], SSIM [37], and their distortion re-weighted versions (WS-PSNR [31], WS-SSIM [45]) are used as evaluation metrics.", + "bbox": [ + 497, + 699, + 902, + 771 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Evaluation under Fisheye Downsampling", + "text_level": 1, + "bbox": [ + 499, + 782, + 862, + 798 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "When the downsampling process is performed on Fisheye images, we train SRResNet [36], EDSR [24], RCAN [43], and SwinIR [23] for comparison.", + "bbox": [ + 497, + 805, + 902, + 847 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative results. As shown in Tab. 1, with the help of additional DF2K-ERP training patches, OSRT out", + "bbox": [ + 497, + 847, + 902, + 876 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "13288", + "bbox": [ + 465, + 891, + 504, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/9a33e0913f340d6848a3d7a1d728d0fe6e853a0283c0a1d970c86bc8b557e6d5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Scale×8×16
MethodODI-SRSUN 360 PanoramaODI-SRSUN 360 Panorama
WS-PSNRWS-SSIMWS-PSNRWS-SSIMWS-PSNRWS-SSIMWS-PSNRWS-SSIM
Bicubic19.640.590819.720.540317.120.433217.560.4638
SRCNN [13]20.080.611219.460.570118.080.450117.950.4684
EDSR [24]23.970.641722.460.634121.120.569821.060.5645
RCAN [43]24.260.662823.880.654221.940.582421.740.5742
360-SS [30]21.650.641721.480.635219.650.543119.620.5308
LAU-Net [12]24.360.680124.020.670822.070.590121.820.5824
SphereSR [40]24.370.677724.170.682022.510.637021.950.6342
OSRT24.530.678024.380.707222.690.626122.130.6388
", + "bbox": [ + 154, + 112, + 811, + 238 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "performs previous methods by 0.3dB on PSNR. Although directly applying SwinIR on the ODISR task has already reached SOTA performance, OSRT surpasses SwinIR over 0.1dB on two datasets for both $\\times 2$ and $\\times 4$ SR tasks, which demonstrates the effectiveness of its distortion modulation ability. The performance of RCAN degrades under Fisheye downsampling, which is caused by the incompatibility between channel attention and Fisheye downsampling5.", + "bbox": [ + 63, + 263, + 468, + 377 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative comparison. Fig. 6 shows the visualization results of $\\times 4$ ODISR task. While other methods struggle to understand the geometric transformation process in distorted images, OSRT can reconstruct sharp and accurate boundaries with the advantages of distortion modulation. It is observed that OSRT is skilled at reconstructing rigid texture. Moreover, benefiting from the distortion modulation ability, OSRT can preserve the original structure as most when being projected to other projection types (Fig. 7).", + "bbox": [ + 63, + 377, + 468, + 505 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Evaluation under ERP Downsampling", + "text_level": 1, + "bbox": [ + 63, + 515, + 405, + 530 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To compare with previous ODISR methods [12, 30, 40], we train OSRT under the previous ERP setting. Regardless of over-fitting issues, we only train on the dataset provided by [12] for fairness. As shown in Tab. 2, OSRT still outperforms LAU-Net [12] and SphereSR [40] under large upscaling factor and ERP downsampling. Without a complicated training pipeline and discrete inference process, OSRT yields the best PSNR values and surpasses all previous methods on most SSIM-related metrics (three of four).", + "bbox": [ + 63, + 536, + 468, + 665 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Study and Discussion", + "text_level": 1, + "bbox": [ + 63, + 675, + 344, + 690 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we prove the effectiveness of Fisheye downsampling, OSRT components, and augmented DF2K-ERP. We then explain the distortion modulation ability of OSRT by visualizing offsets in deformable blocks.", + "bbox": [ + 63, + 697, + 468, + 754 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fisheye downsampling. As shown in Fig. 1, the SR model trained under ERP downsampling is more likely to generate blur details and missing structures in real-world scenarios. These artifacts cannot be removed by a superior backbone network, but can be eliminated by a more realistic imaging process. More importantly, ERP downsampling directly covers the geometric property of ERP images", + "bbox": [ + 63, + 754, + 468, + 853 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3cf94c84f54a3a4b5ff65cc77204388439d79b1a4cdf4eeb5d2eec389e6fffff.jpg", + "table_caption": [ + "Table 2. SR results under ERP downsampling." + ], + "table_footnote": [], + "table_body": "
feature dimDACBDAABODI-SRSUN360Params. (M)
PSNRSSIMPSNRSSIM
60××30.270.873930.780.87420.91
60✓×30.410.877531.000.87931.16
60×w/o Cw30.310.874630.830.87551.00
60×w/ Cw30.320.874630.840.87531.01
60✓w/ Cw30.440.878031.040.88001.26
72××30.320.874830.850.87551.29
", + "bbox": [ + 505, + 261, + 890, + 351 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Ablation study on OSRT components. All models are trained on $\\times 2$ SR task under Fisheye downsampling.", + "bbox": [ + 497, + 355, + 900, + 382 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "and makes the ODISR task identical to the standard plain image super-resolution task. The evidence is that a standard SISR model (SwinIR) trained on a plain image dataset (DF2K) can outperform previous SOTA in the ODISR task, which yields WS-PSNR results of 24.63dB/24.49dB (22.68dB/22.13dB) on $\\times 8$ ( $\\times 16$ ) ODI-SR/SUN360 testing set, respectively. In conclusion, when the intrinsic property of ODIs is broken by ERP downsampling, the ODISR task degenerates into a plain image super-resolution task with a particular data distribution.", + "bbox": [ + 497, + 384, + 902, + 526 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d263c9dc37bd87f334f11c9005cdf5c152dd27cca0279982cf36f6cb5dda0676.jpg", + "image_caption": [ + "ODI-SR $(\\times 4)$ :003 Fisheye (Vertical, Right)" + ], + "image_footnote": [], + "bbox": [ + 502, + 528, + 690, + 661 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/347399a281d81a2d0d22053c52bed700fd8cfce33057400626ea2f8cc673852b.jpg", + "image_caption": [ + "EDSR [24]" + ], + "image_footnote": [], + "bbox": [ + 697, + 528, + 794, + 597 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8aa834e957f53eb38f677568c4bf692307999de4ef356c7c58422c3c26310c02.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 798, + 529, + 894, + 596 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9963fa8d1acfcc5a6d149a19bfb179db69dd6d77b260bc2ca6ba3bdc5faf8c48.jpg", + "image_caption": [ + "SwinIR† [23]" + ], + "image_footnote": [], + "bbox": [ + 697, + 606, + 794, + 673 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ec3d96c5c427240db388b5eacae55786ddb7b31ba10c483e2efd6dd8118ce7c0.jpg", + "image_caption": [ + "SwinIR [23]", + "OSRT†" + ], + "image_footnote": [], + "bbox": [ + 798, + 606, + 894, + 673 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/044705c461a57792277caae6ab00859e0349e3d4c68851aa8489c1f61b722fd8.jpg", + "image_caption": [ + "SUN360 $(\\times 2)$ : 062 Perspective $(\\varphi :30^{\\circ}$ ;FOV: $90^{\\circ}$" + ], + "image_footnote": [], + "bbox": [ + 504, + 687, + 690, + 819 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ab76630d7f4eb043277a7f12d79ac163f08a3ef4c81ee0291e596541b8c53988.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 687, + 793, + 754 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7765713b1be23316de722e73c0ce74cede6ff58d5d4247c9cc4212714f3e92b9.jpg", + "image_caption": [ + "EDSR [24]", + "SwinIR† [23]" + ], + "image_footnote": [], + "bbox": [ + 697, + 764, + 793, + 831 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5caee10a8ce95acf83a71b25eeb960c6896770a0a72d1c8dacffb3c246e1d105.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 798, + 687, + 894, + 754 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/159410d8b784f09b44c67d733c2fc6325a05f36296e01d3b0f733d950853a79b.jpg", + "image_caption": [ + "SwinIR [23]", + "OSRT†", + "Figure 7. Visual comparisons for SR of Fisheye and Perspective images. $\\dagger$ denotes applying DF2K-ERP as augmented dataset." + ], + "image_footnote": [], + "bbox": [ + 798, + 764, + 894, + 832 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "5The cause is discussed in the supplementary file.", + "bbox": [ + 82, + 863, + 354, + 876 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "13289", + "bbox": [ + 465, + 891, + 504, + 902 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c3cf00c8f3d468d3e2c7a3efbf4bd7a8c653e5e8e31469311d32541521f07351.jpg", + "image_caption": [ + "DAAB (29-th)" + ], + "image_footnote": [], + "bbox": [ + 112, + 115, + 226, + 195 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/aaf70f510f46675cf63b9135c08d88c8212849ebe829cdaca705a4dadd8456ea.jpg", + "image_caption": [ + "$\\varphi = 45^{\\circ}$" + ], + "image_footnote": [], + "bbox": [ + 231, + 116, + 344, + 195 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/17e28f2558a2bba12c2cedbe02d22d7006226271a10f0413a0f789e57c345074.jpg", + "image_caption": [ + "$\\varphi = 90^{\\circ}$" + ], + "image_footnote": [], + "bbox": [ + 351, + 116, + 463, + 195 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/45e10ebd222dcf2a24dc360e989bc54ed34567825a2779a634b800c19babe609.jpg", + "image_caption": [ + "$\\omega = 0^{\\circ}$", + "DAAB (33-th)" + ], + "image_footnote": [], + "bbox": [ + 112, + 206, + 226, + 286 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1d82e6dc3dc63cc5f2405848b899de9664c0e7d2b89e8b2cdc5fa89a40e3ec2e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 231, + 206, + 344, + 286 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8b703d5dee9b0a31f011e5c6c872d5cbce0c5fba1282cf61e05e732ca49178a0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 351, + 206, + 463, + 286 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/17ce58425cfcda28af96cdc102bcb82b07e11f6257b1859af081fb33bdad73f4.jpg", + "image_caption": [ + "$\\varphi = 0^{\\circ}$", + "DACB", + "$\\varphi = 0^{\\circ}$", + "Figure 8. Visualizations of offset maps in OSRT. Reference and deformed points are depicted in green and red, respectively. The deformable kernel is sparse in the polar area." + ], + "image_footnote": [], + "bbox": [ + 112, + 298, + 196, + 357 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b197eba21c2d9107a864f5589706d7516219b43891386c3ca6eb8d3e277022d3.jpg", + "image_caption": [ + "$\\varphi = 45^{\\circ}$", + "$\\varphi = 30^{\\circ}$" + ], + "image_footnote": [], + "bbox": [ + 201, + 298, + 285, + 357 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4db4315bd083d2575e395724e124f4e6134db2cb434b7d98c625c52210a3e8e1.jpg", + "image_caption": [ + "$\\varphi = 60^{\\circ}$" + ], + "image_footnote": [], + "bbox": [ + 290, + 298, + 374, + 357 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7ebc60cf1b6d62fb9d134d64abd88daba08c92ed823a05a7d574caf9b86f4609.jpg", + "image_caption": [ + "$\\varphi = 90^{\\circ}$", + "$\\varphi = 90^{\\circ}$" + ], + "image_footnote": [], + "bbox": [ + 381, + 298, + 463, + 357 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "OSRT components. To study the effectiveness of each component in OSRT, we propose a light version of OSRT (OSRT-light) for ablation study, which corresponds with the official SwinIR-light [23]. As proofed in Tab. 3, all components in OSRT are beneficial for modulating ERP distortion. The advantages of DACB and DAAB can be stacked when being applied in the same network. Compared with simply expanding the feature dimension of SwinIR to match the network complexity, the overall improvements of OSRT is more significant $(+0.05\\mathrm{dB}$ vs. $+0.2\\mathrm{dB})$", + "bbox": [ + 63, + 423, + 467, + 565 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Offsets in OSRT. Offsets map in a well-trained OSRT are visualized in Fig. 8. Deformable kernels in both DAAB and DACB tend to gather at the equator and scatter at the polar, which conforms to the geometric distribution of pixel density in ERP images.", + "bbox": [ + 63, + 568, + 468, + 638 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/86d33b68e48f5512cb07c309a6291a07500b6c66226197bbb47b164d05215df9.jpg", + "image_caption": [ + "Figure 9. Training process of Transformers on $\\times 4$ ODISR task. The overfitting issue is tackled by our augmentation scheme." + ], + "image_footnote": [], + "bbox": [ + 67, + 650, + 468, + 843 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/177e1bd4b84fbb0d117f1efcfb20da93a91db993aaf1a7694baf1b25a2e70697.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Backbone networkDatasetsTraining schemeScaleSUN360
PSNRSSIM
SwinIRODI-SRN/A31.210.8852
SwinIRDF2K/ODI-SRone-stage31.260.8841
SwinIRDF2K-ERP/ODI-SRone-stage×231.330.8855
SwinIRDF2K-ERP/ODI-SRtwo-stage31.170.8818
OSRTDF2K-ERP/ODI-SRone-stage31.520.8888
SwinIRODI-SRN/A27.390.7707
SwinIRDF2K/ODI-SRone-stage27.590.7768
SwinIRDF2K-ERP/ODI-SRone-stage×427.710.7804
SwinIRDF2K-ERP/ODI-SRtwo-stage27.740.7795
OSRTDF2K-ERP/ODI-SRone-stage27.840.7835
", + "bbox": [ + 515, + 112, + 884, + 245 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Ablation study on data augmentation. The results of ODISR (In the supplementary file) are in the same trend as SUN360.", + "bbox": [ + 499, + 247, + 900, + 274 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Pseudo ERP patches. In Sec. 3.4, we propose a distorted dataset DF2K-ERP to tackle over-fitting issues. We train a standard SwinIR on diverse datasets and training schemes to study the influence of data augmentation separately. As shown in Tab. 4, while training on ODI-SR and DF2K, distortion operations in DF2K lead to better performance. Compared with fine-tuning on DF2K-ERP pretrained models (two-stage), training on two datasets jointly (one-stage) shows better results. We infer that there is a domain gap between ODI-SR and DF2K-ERP, which is caused by omitted Perspective distortion. Moreover, the advantage of distortion modulation mechanisms in OSRT is enlarged when additional training patches are applied. Fig. 9 proves that our data augmentation scheme overcomes the over-fitting issue and improves the reconstruction ability.", + "bbox": [ + 497, + 283, + 904, + 496 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 510, + 621, + 524 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we find that the previous downsampling process in the ODISR task harms the intrinsic distribution of pixel density in ODIs, which leads to poor generalization ability in real-world scenarios. To tackle this issue, we propose Fisheye downsampling, which mimics the real-world imaging process to preserve the realistic density distribution. After refining the downsampling process, we design a distortion-aware Transformer (OSRT) to modulate distortions continuously and self-adaptively. OSRT learns offsets from the distortion-related condition and rectifies distortion by feature-level warping. Moreover, to alleviate the overfitting problem of large networks, we propose to synthesize additional ERP training data from the plain images. Extensive experiments have demonstrated the state-of-the-art performance of our OSRT.", + "bbox": [ + 497, + 534, + 902, + 747 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitation. The process of sampling ERP images into viewing types also requires careful design.", + "bbox": [ + 499, + 748, + 900, + 776 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement.", + "text_level": 1, + "bbox": [ + 519, + 777, + 660, + 790 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported in part by the National Key R&D Program of China (NO.2022ZD0160505), in part by the National Natural Science Foundation of China under Grant (62276251), the Joint Lab of CAS-HK, and in part by the Youth Innovation Promotion Association of Chinese Academy of Sciences (No.2020356).", + "bbox": [ + 497, + 790, + 900, + 876 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "13290", + "bbox": [ + 465, + 891, + 505, + 903 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 65, + 112, + 164, + 128 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Hao Ai, Zidong Cao, Jinjing Zhu, Haotian Bai, Yucheng Chen, and Ling Wang. Deep learning for omnidirectional vision: A survey and new perspectives. arXiv preprint arXiv:2205.10468, 2022. 1", + "[2] Zafer Arican and Pascal Frossard. L1 regularized superresolution from unregistered omnidirectional images. In 2009 IEEE International Conference on Acoustics, Speech and Signal Processing, pages 829-832. IEEE, 2009. 2", + "[3] Zafer Arican and Pascal Frossard. Joint registration and super-resolution with omnidirectional images. IEEE Transactions on Image Processing, 20(11):3151-3162, 2011. 2", + "[4] Luigi Bagnato, Yannick Boursier, Pascal Frossard, and Pierre Vandergheynst. Plenoptic based super-resolution for omnidirectional image sequences. In 2010 IEEE International Conference on Image Processing, pages 2829-2832. IEEE, 2010. 2", + "[5] Kelvin CK Chan, Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Basicvsr: The search for essential components in video super-resolution and beyond. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4947-4956, 2021. 3", + "[6] Kelvin CK Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. Basicvsr++: Improving video superresolution with enhanced propagation and alignment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5972-5981, 2022. 3", + "[7] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12299-12310, 2021. 2", + "[8] Xiangyu Chen, Xintao Wang, Jiantao Zhou, and Chao Dong. Activating more pixels in image super-resolution transformer. arXiv preprint arXiv:2205.04437, 2022. 2", + "[9] Yinbo Chen, Sifei Liu, and Xiaolong Wang. Learning continuous image representation with local implicit image function. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8628-8638, 2021. 2", + "[10] Jifeng Dai, Haozhi Qi, Yuwen Xiong, Yi Li, Guodong Zhang, Han Hu, and Yichen Wei. Deformable convolutional networks. In Proceedings of the IEEE international conference on computer vision, pages 764-773, 2017. 4", + "[11] Tao Dai, Jianrui Cai, Yongbing Zhang, Shu-Tao Xia, and Lei Zhang. Second-order attention network for single image super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11065-11074, 2019. 2", + "[12] Xin Deng, Hao Wang, Mai Xu, Yichen Guo, Yuhang Song, and Li Yang. Lau-net: Latitude adaptive upscaling network for omnidirectional image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9189-9198, 2021. 1, 2, 6, 7", + "[13] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Image super-resolution using deep convolutional net" + ], + "bbox": [ + 67, + 137, + 468, + 876 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "works. IEEE transactions on pattern analysis and machine intelligence, 38(2):295-307, 2015. 2, 7", + "[14] Chao Dong, Chen Change Loy, and Xiaou Tang. Accelerating the super-resolution convolutional neural network. In European conference on computer vision, pages 391-407. Springer, 2016. 2", + "[15] Vida Fakour-Sevom, Esin Guldogan, and Joni-Kristian Kämäräinen. 360 panorama super-resolution using deep convolutional networks. In Int. Conf. on Computer Vision Theory and Applications (VISAPP), volume 1, 2018. 1, 2", + "[16] Quan Huynh-Thu and Mohammed Ghanbari. Scope of validity of psnr in image/video quality assessment. *Electronics letters*, 44(13):800-801, 2008. 6", + "[17] Hiroshi Kawasaki, Katsushi Ikeuchi, and Masao Sakauchi. Super-resolution omnidirectional camera images using spatio-temporal analysis. *Electronics and Communications in Japan* (Part III: Fundamental Electronic Science), 89(6):47-59, 2006. 2", + "[18] Renata Khasanova and Pascal Frossard. Geometry aware convolutional filters for omnidirectional images representation. In International Conference on Machine Learning, pages 3351-3359. PMLR, 2019. 4", + "[19] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6", + "[20] Christian Ledig, Lucas Theis, Ferenc Huszar, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photorealistic single image super-resolution using a generative adversarial network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4681-4690, 2017. 2", + "[21] Wenbo Li, Xin Lu, Jiangbo Lu, Xiangyu Zhang, and Jiaya Jia. On efficient transformer and image pre-training for low-level vision. arXiv preprint arXiv:2112.10175, 2021. 2", + "[22] Zheyuan Li, Yingqi Liu, Xiangyu Chen, Haoming Cai, Jinjin Gu, Yu Qiao, and Chao Dong. Blueprint separable residual network for efficient image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 833-843, 2022. 2", + "[23] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1833-1844, 2021. 2, 5, 6, 7, 8", + "[24] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 136-144, 2017. 1, 2, 5, 6, 7", + "[25] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 2", + "[26] Yiqun Mei, Yuchen Fan, and Yuqian Zhou. Image superresolution with non-local sparse attention. In Proceedings of" + ], + "bbox": [ + 502, + 116, + 900, + 876 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "13291", + "bbox": [ + 465, + 891, + 502, + 902 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3517-3526, 2021. 2", + "[27] Hajime Nagahara, Yasushi Yagi, and Masahiko Yachida. Super-resolution from an omnidirectional image sequence. In 2000 26th Annual Conference of the IEEE Industrial Electronics Society. IECON 2000. 2000 IEEE International Conference on Industrial Electronics, Control and Instrumentation. 21st Century Technologies, volume 4, pages 2559-2564. IEEE, 2000. 2", + "[28] Akito Nishiyama, Satoshi Ikehata, and Kiyoharu Aizawa. 360 single image super resolution via distortion-aware network and distorted perspective images. In 2021 IEEE International Conference on Image Processing (ICIP), pages 1829-1833. IEEE, 2021. 1, 2, 4", + "[29] Ben Niu, Weilei Wen, Wenqi Ren, Xiangde Zhang, Lianping Yang, Shuzhen Wang, Kaihao Zhang, Xiaochun Cao, and Haifeng Shen. Single image super-resolution via a holistic attention network. In European conference on computer vision, pages 191-207. Springer, 2020. 2", + "[30] Cagri Ozcinar, Aakanksha Rana, and Aljosa Smolic. Superresolution of omnidirectional images using adversarial learning. In 2019 IEEE 21st International Workshop on Multimedia Signal Processing (MMSP), pages 1-6. IEEE, 2019. 2, 3, 7", + "[31] Yule Sun, Ang Lu, and Lu Yu. Weighted-to-spherically-uniform quality evaluation for omnidirectional video. IEEE signal processing letters, 24(9):1408-1412, 2017. 3, 6", + "[32] Yapeng Tian, Yulun Zhang, Yun Fu, and Chenliang Xu. Tdan: Temporally-deformable alignment network for video super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3360-3369, 2020. 4", + "[33] Radu Timofte, Eirikur Agustsson, Luc Van Gool, Ming-Hsuan Yang, and Lei Zhang. Ntire 2017 challenge on single image super-resolution: Methods and results. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 114-125, 2017. 5", + "[34] Xintao Wang, Kelvin CK Chan, Ke Yu, Chao Dong, and Chen Change Loy. Edvr: Video restoration with enhanced deformable convolutional networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 0–0, 2019. 3, 4", + "[35] Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1905-1914, 2021. 2", + "[36] Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Yu Qiao, and Chen Change Loy. Esrgan: Enhanced super-resolution generative adversarial networks. In Proceedings of the European conference on computer vision (ECCV) workshops, pages 0–0, 2018. 2, 6", + "[37] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6", + "[38] Zhuofan Xia, Xuran Pan, Shiji Song, Li Erran Li, and Gao Huang. Vision transformer with deformable attention. In" + ], + "bbox": [ + 68, + 116, + 467, + 876 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4794-4803, 2022. 3", + "[39] Jianxiong Xiao, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Recognizing scene viewpoint using panoramic place representation. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, pages 2695-2702. IEEE, 2012. 1, 6", + "[40] Youngho Yoon, Inchul Chung, Lin Wang, and Kuk-Jin Yoon. Spheresr: 360deg image super-resolution with arbitrary projection via continuous spherical image representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5677-5686, 2022. 1, 2, 7", + "[41] Kai Zhang, Jingyun Liang, Luc Van Gool, and Radu Timofte. Designing a practical degradation model for deep blind image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4791-4800, 2021. 2", + "[42] Wenlong Zhang, Yihao Liu, Chao Dong, and Yu Qiao. Ranksrgan: Generative adversarial networks with ranker for image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3096-3105, 2019. 2", + "[43] Yulun Zhang, Kunpeng Li, Kai Li, Lichen Wang, Bineng Zhong, and Yun Fu. Image super-resolution using very deep residual channel attention networks. In Proceedings of the European conference on computer vision (ECCV), pages 286-301, 2018. 1, 2, 6, 7", + "[44] Yupeng Zhang, Hengzhi Zhang, Daojing Li, Liyan Liu, Hong Yi, Wei Wang, Hiroshi Suitoh, and Makoto Odamaki. Toward real-world panoramic image enhancement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 628-629, 2020. 2", + "[45] Yufeng Zhou, Mei Yu, Hualin Ma, Hua Shao, and Gangyi Jiang. Weighted-to-spherically-uniform sSIM objective quality evaluation for panoramic video. In 2018 14th IEEE International Conference on Signal Processing (ICSP), pages 54-57. IEEE, 2018. 6", + "[46] Xizhou Zhu, Han Hu, Stephen Lin, and Jifeng Dai. Deformable convnets v2: More deformable, better results. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9308-9316, 2019. 5" + ], + "bbox": [ + 502, + 116, + 902, + 686 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "13292", + "bbox": [ + 465, + 891, + 504, + 902 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/OSRT_ Omnidirectional Image Super-Resolution With Distortion-Aware Transformer/b6aea18d-97cd-464e-a4ad-626c19fdc060_model.json b/2023/OSRT_ Omnidirectional Image Super-Resolution With Distortion-Aware Transformer/b6aea18d-97cd-464e-a4ad-626c19fdc060_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6cb30274bc74c3693cf37693762321f6ad424099 --- /dev/null +++ b/2023/OSRT_ Omnidirectional Image Super-Resolution With Distortion-Aware Transformer/b6aea18d-97cd-464e-a4ad-626c19fdc060_model.json @@ -0,0 +1,3223 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.12, + 0.038, + 0.191, + 0.074 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.246, + 0.036, + 0.8, + 0.077 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.205, + 0.139, + 0.764, + 0.178 + ], + "angle": 0, + "content": "OSRT: Omnidirectional Image Super-Resolution with Distortion-aware Transformer" + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.18, + 0.871, + 0.247 + ], + "angle": 0, + "content": "Fanghua Yu\\(^{1*}\\) Xintao Wang\\(^{2*}\\) Mingdeng Cao\\(^{2,3}\\) Gen Li\\(^{4}\\) Ying Shan\\(^{2}\\) Chao Dong\\(^{1,5\\dagger}\\) \n\\(^{1}\\)ShenZhen Key Lab of Computer Vision and Pattern Recognition \nShenzhen Institute of Advanced Technology, Chinese Academy of Sciences \n\\(^{2}\\)ARC, Tencent PCG \\({}^{3}\\)The University of Tokyo" + }, + { + "type": "text", + "bbox": [ + 0.105, + 0.248, + 0.864, + 0.265 + ], + "angle": 0, + "content": "4 Platform Technologies, Tencent Online Video Shanghai Artificial Intelligence Laboratory" + }, + { + "type": "text", + "bbox": [ + 0.166, + 0.267, + 0.803, + 0.28 + ], + "angle": 0, + "content": "fanghuayu96@gmail.com, xintaowang@tencent.com, cmd@g.ecc.u-tokyo.ac.jp" + }, + { + "type": "text", + "bbox": [ + 0.248, + 0.283, + 0.72, + 0.297 + ], + "angle": 0, + "content": "{genli, yingsshan}@tencent.com, chao.dong@siat.ac.cn" + }, + { + "type": "title", + "bbox": [ + 0.228, + 0.32, + 0.308, + 0.334 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.065, + 0.343, + 0.474, + 0.643 + ], + "angle": 0, + "content": "Omnidirectional images (ODIs) have obtained lots of research interest for immersive experiences. Although ODIs require extremely high resolution to capture details of the entire scene, the resolutions of most ODIs are insufficient. Previous methods attempt to solve this issue by image super-resolution (SR) on equirectangular projection (ERP) images. However, they omit geometric properties of ERP in the degradation process, and their models can hardly generalize to real ERP images. In this paper, we propose Fisheye downsampling, which mimics the real-world imaging process and synthesizes more realistic low-resolution samples. Then we design a distortion-aware Transformer (OSRT) to modulate ERP distortions continuously and self-adaptively. Without a cumbersome process, OSRT outperforms previous methods by about 0.2dB on PSNR. Moreover, we propose a convenient data augmentation strategy, which synthesizes pseudo ERP images from plain images. This simple strategy can alleviate the over-fitting problem of large networks and significantly boost the performance of ODISR. Extensive experiments have demonstrated the state-of-the-art performance of our OSRT." + }, + { + "type": "title", + "bbox": [ + 0.067, + 0.662, + 0.201, + 0.677 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.064, + 0.687, + 0.469, + 0.814 + ], + "angle": 0, + "content": "In pursuit of the realistic visual experience, omnidirectional images (ODIs), also known as \\(360^{\\circ}\\) images or panoramic images, have obtained lots of research interest in the computer vision community. In reality, we usually view ODIs with a narrow field-of-view (FOV), e.g., viewing in a headset. To capture details of the entire scene, ODIs require extremely high resolution, e.g., \\(4\\mathrm{K} \\times 8\\mathrm{K}\\) [1]. However, due to the high industrial cost of camera sensors with high precision, the resolutions of most ODIs are insufficient." + }, + { + "type": "text", + "bbox": [ + 0.065, + 0.815, + 0.469, + 0.844 + ], + "angle": 0, + "content": "Recently, some attempts have been made to solve this problem by image super-resolution (SR) [12, 15, 28, 39, 40]." + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.312, + 0.699, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.572, + 0.413, + 0.637, + 0.423 + ], + "angle": 0, + "content": "Unseen LR" + }, + { + "type": "image", + "bbox": [ + 0.713, + 0.312, + 0.899, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.729, + 0.413, + 0.878, + 0.424 + ], + "angle": 0, + "content": "LAU-Net [12] w/o Fisheye" + }, + { + "type": "image", + "bbox": [ + 0.512, + 0.428, + 0.699, + 0.528 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.551, + 0.53, + 0.657, + 0.541 + ], + "angle": 0, + "content": "OSRT w/o Fisheye" + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.427, + 0.899, + 0.528 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.756, + 0.53, + 0.855, + 0.541 + ], + "angle": 0, + "content": "OSRT w/ Fisheye" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.544, + 0.904, + 0.583 + ], + "angle": 0, + "content": "Figure 1. Visual comparisons of \\(\\times 8\\) SR results on LR images1 with unknown degradations. Fisheye denotes that the downsampling process in training stages is under Fisheye images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.583, + 0.905, + 0.854 + ], + "angle": 0, + "content": "As most of the ODIs are stored and transmitted in the equirectangular projection (ERP) type, the SR process is usually performed on the ERP images. To generate high- /low-resolution training pairs, existing ODISR methods [12, 15, 28, 39, 40] directly apply uniform bicubic down-sampling on the original ERP images (called ERP down-sampling), which is identical to general image SR settings [24, 43]. While omitting geometric properties of ERP in the degradation process, their models can hardly generalize to real ERP images. We can observe missing structures and blur textures in Fig. 1. Therefore, we need a more appropriate degradation model before studying SR algorithms. In practice, ODIs are acquired by the fisheye lens and stored in ERP. Given that the low-resolution issue in real-world scenarios is caused by insufficient sensor precision and density, the downsampling process should be applied to original-formatted images before converting into other storage types. Thus, to be conformed with real-world imaging processes, we propose to apply uniform bicubic" + }, + { + "type": "page_footnote", + "bbox": [ + 0.083, + 0.854, + 0.195, + 0.864 + ], + "angle": 0, + "content": "*Equal contribution" + }, + { + "type": "page_footnote", + "bbox": [ + 0.083, + 0.865, + 0.383, + 0.877 + ], + "angle": 0, + "content": "† Corresponding author (e-mail: chao.dong@siat.ac.cn)" + }, + { + "type": "list", + "bbox": [ + 0.083, + 0.854, + 0.383, + 0.877 + ], + "angle": 0, + "content": null + }, + { + "type": "page_footnote", + "bbox": [ + 0.518, + 0.864, + 0.796, + 0.877 + ], + "angle": 0, + "content": "Photoed by Peter Leth on Flickr, with CC license." + }, + { + "type": "page_number", + "bbox": [ + 0.466, + 0.892, + 0.505, + 0.904 + ], + "angle": 0, + "content": "13283" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.065, + 0.115, + 0.469, + 0.203 + ], + "angle": 0, + "content": "downsampling on Fisheye images, which are the original format of ODIs. The new downsampling process (called Fisheye downsampling) applies uniform bicubic downsampling on Fisheye images before converting them to ERP images. Our Fisheye downsampling is more conducive to exploring the geometric property of ODIs." + }, + { + "type": "text", + "bbox": [ + 0.065, + 0.204, + 0.471, + 0.573 + ], + "angle": 0, + "content": "The key issue of ODISR algorithm design is to utilize the geometric properties of ERP images, which is also the focus of previous methods. For example, Nishiyama et al. [28] add a distortion-related condition as an additional input. LAU-Net [12] splits the whole ERP image into patches by latitude band and learns upscaling processes separately. However, the separated learning process will lead to information disconnection between adjacent patches. SphereSR [40] learns different upscaling functions on various projection types, but will inevitably introduce multiple-time computation costs. To push the performance upper bound, we propose the first Transformer for Omnidirectional image Super-Resolution (OSRT), and incorporate geometric properties in a distortion-aware manner. Specifically, to modulate distorted feature maps, we implement feature-level warping, in which offsets are learned from latitude conditions. In OSRT, we introduce two dedicated blocks to adapt latitude-related distortion: distortion-aware attention block (DAAB), and distortion-aware convolution block (DACB). DAAB and DACB are designed to perform distortion modulation in arbitrary Transformers and ConvNets. These two blocks can directly replace the multi-head self-attention block and convolution layer, respectively. The benefit of DAAB and DACB can be further improved when being inserted into the same backbone network. OSRT outperforms previous methods by about 0.2dB on PSNR (Tab. 2)." + }, + { + "type": "text", + "bbox": [ + 0.065, + 0.576, + 0.47, + 0.805 + ], + "angle": 0, + "content": "However, the increase of network capacity will also enlarge the overfitting problem of ODISR, which is rarely mentioned before. The largest ODIs dataset [12] contains only 1K images, which cannot provide enough diversity for training Transformers. Given that acquiring ODIs requires expensive equipment and tedious work, we propose to generate distorted ERP samples from plain images for data augmentation. In practice, we regard a plain image as a sampled perspective, and project it back to the ERP format. Then we can introduce 146K additional training patches, 6 times of the previous dataset. This simple strategy can significantly boost the performance of ODISR (Tab. 4) and alleviate the over-fitting problem of large networks (Fig. 9). A similar data augmentation method is also applied in Nishiyama et al. [28], but shows marginal improvement on small models under ERP downsampling settings." + }, + { + "type": "text", + "bbox": [ + 0.065, + 0.806, + 0.47, + 0.879 + ], + "angle": 0, + "content": "Our contributions are threefold. 1) For problem formulation: To generate more realistic ERP low-resolution images, we propose Fisheye downsampling, which mimics the real-world imaging process. 2) For method: Combined with the geometric properties of ERP, we design a distortion-aware" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.115, + 0.905, + 0.204 + ], + "angle": 0, + "content": "Transformer, which modulates distortions continuously and self-adaptively without cumbersome process. 3) For data: To reduce overfitting, we propose a convenient data augmentation strategy, which synthesizes pseudo ERP images from plain images. Extensive experiments have demonstrated the state-of-the-art performance of our OSRT2." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.217, + 0.646, + 0.232 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.242, + 0.905, + 0.441 + ], + "angle": 0, + "content": "Single Image Super-Resolution (SISR). Deep learning for single image SR (SISR) is first introduced in [13]. Further works boost SR performance by CNNs [11, 14, 22, 24, 26, 29, 43], Vision Transformers (ViTs) [7, 8, 21, 23] and generative adversarial networks (GANs) [20, 35, 36, 42]. For instance, EDSR [24] removes Batch Normalization layers and applies a more complicated residual block. RCAN [43] introduces channel-wise attention mechanisms to a deeper network. SwinIR [23] proposes an image restoration Transformer based on [25]. To improve perceptual quality, adversarial training are performed as a tuning process to generate more realistic results [35, 36]. Moreover, various flexible degradation models are proposed in [35, 41] to synthesize more practical degradations." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.443, + 0.905, + 0.854 + ], + "angle": 0, + "content": "Omnidirectional Image Super-Resolution (ODISR). Initially, ODISR models focus on the spherical assembling of LR ODIs under various projection types [2-4, 17, 27]. Recent ODISR models are performed on plane images and are fine-tuned from existing SISR models with L1 loss [15] or GAN loss [30, 44]. The improvements are limited, for they only concern the distribution gap between ODIs and plain images. Since LAU-Net [12] found pixel density in ERP ODIs is non-uniform, many studies attempt to design specific backbone networks to overcome this issue. LAU-Net [12] manually splits the whole ERP image into latitude-related patches and learns ERP distortion over different latitude ranges separately. While LAU-Net learns latitude-related ERP distortion somewhat, its non-overlapped patches lead to disconnection in whole ERP images. Nishiyama et al. [28] treats area stretching ratio as additional input. However, these conditions are tough to be utilized with an unmodified SISR backbone network. SphereSR [40] learns upsampling processes on various projection types to mitigate the influence of non-uniformity in specific projection types. Although SphereSR improves information consistency between various ODI projection types by LIIF [9], they apply multiple networks to learn the upscaling process of each projection type. Given that all other projection types in SphereSR are converted from ERP, patterns under various types are reusable when distortions are properly rectified. Moreover, the complex and unstructured image data in polyhedron projection hinders further research of ODISR." + }, + { + "type": "page_footnote", + "bbox": [ + 0.518, + 0.864, + 0.733, + 0.877 + ], + "angle": 0, + "content": "\\(^{2}\\)https://github.com/Fanghua-Yu/OSRT" + }, + { + "type": "page_number", + "bbox": [ + 0.466, + 0.892, + 0.506, + 0.904 + ], + "angle": 0, + "content": "13284" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.065, + 0.115, + 0.47, + 0.215 + ], + "angle": 0, + "content": "Deformable Mechanism. Dai et al. [30] first propose deformable convolutions to obtain information out of its regular neighborhood. Xia et al. [38] further verified that Vision Transformers also benefit from applying deformable mechanisms on self-attention blocks. In Video SR tasks, the deformable mechanism can be adapted to align features between adjacent frames [5,6,34]." + }, + { + "type": "title", + "bbox": [ + 0.066, + 0.227, + 0.16, + 0.243 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.065, + 0.251, + 0.469, + 0.337 + ], + "angle": 0, + "content": "In this section, we first analyze the cause of ERP and Fisheye distortions, as well as the relationship between these two distortions (Sec. 3.1). Then, we discuss the designs of Fisheye downsampling (Sec. 3.2), distortion-aware Transformer (OSRT) (Sec. 3.3), and the convenient data augmentation strategy (Sec. 3.4)." + }, + { + "type": "title", + "bbox": [ + 0.066, + 0.345, + 0.342, + 0.361 + ], + "angle": 0, + "content": "3.1. Revisiting Distortions in ODIs" + }, + { + "type": "image", + "bbox": [ + 0.073, + 0.371, + 0.469, + 0.56 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.065, + 0.564, + 0.469, + 0.603 + ], + "angle": 0, + "content": "Figure 2. Geometric explanation of transforming between ERP, Fisheye, and the ideal spherical surface. To simplify, we discuss the horizontal spliced Fisheye with an aperture degree of \\(\\pi\\)." + }, + { + "type": "text", + "bbox": [ + 0.064, + 0.607, + 0.469, + 0.763 + ], + "angle": 0, + "content": "As ODIs under each projection type are constrained by different transforming equations, the distortion caused by each type is inconsistent, indicating that applying matrix operations under one projection type can introduce unexpected changes when being converted to other types. Specifically, applying uniformed bicubic downsampling on ERP images will affect the distribution of pixel density on Fisheye images, which are the original-formatted image type of imaging process in real-world scenarios. To analyze the specific effect of ERP downsampling on the Fisheye image, we revisit the cause of distortions in ERP and Fisheye." + }, + { + "type": "text", + "bbox": [ + 0.064, + 0.764, + 0.469, + 0.877 + ], + "angle": 0, + "content": "As we assume that viewing directions are uniformly distributed, the data points in an ideal ODI should be uniformly distributed on a spherical surface. In practice, there is a trade-off between the uniformity of the spherical surface and the structural degree. ERP is the most convenient projection type for storage or transmission, but it is also the projection type that suffers the heaviest distortion. To better explain the causation of distortions, we follow the definition" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.115, + 0.905, + 0.188 + ], + "angle": 0, + "content": "of stretching ratio \\((\\mathbf{K})\\) in [31], which represents distortion degree at different locations from the target projection type to the ideal spherical surface. \\(\\mathbf{K}\\) is determined by area variation from one projection type to another. When the target type is uniforming spherical surface, \\(\\mathbf{K}\\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.195, + 0.903, + 0.228 + ], + "angle": 0, + "content": "\\[\n\\mathbf {K} (x, y) = \\frac {\\delta S (\\theta , \\varphi)}{\\delta P (x , y)} = \\frac {\\cos (\\varphi) | d \\theta d \\varphi |}{| d x d y |} = \\frac {\\cos (\\varphi)}{| J (\\theta , \\varphi) |}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.236, + 0.903, + 0.293 + ], + "angle": 0, + "content": "where \\(\\delta S(\\cdot ,\\cdot)\\) and \\(\\delta P(\\cdot ,\\cdot)\\) represent the area on the spherical surface and the projection plane, respectively. \\(|didj|\\) represents plane microunit. \\(|J(\\theta ,\\varphi)|\\) is the Jacobian determinant from spherical coordinate to projection coordinate." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.293, + 0.904, + 0.322 + ], + "angle": 0, + "content": "ERP distortion. The coordinate in ERP is defined as \\( x = \\theta \\) and \\( y = \\varphi \\). ERP stretching ratio can be derived as:" + }, + { + "type": "equation", + "bbox": [ + 0.587, + 0.332, + 0.903, + 0.349 + ], + "angle": 0, + "content": "\\[\n\\mathbf {K} _ {\\mathrm {E R P}} (x, y) = \\cos (\\varphi) = \\cos (y), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.357, + 0.734, + 0.373 + ], + "angle": 0, + "content": "where \\(x\\in (-\\pi ,\\pi)\\) \\(y\\in (-\\frac{\\pi}{2},\\frac{\\pi}{2})\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.372, + 0.904, + 0.514 + ], + "angle": 0, + "content": "From Eq. (2), we conclude that ERP distortion is only determined by its latitude degree. \\(\\mathbf{K}_{\\mathrm{ERP}}\\) is reduced to zero when the absolute value of latitude degree increases to \\(\\pi /2\\), which represents that pixel density on the polar areas of ERP images is closer to zero. As shown in Fig. 2 (c), with the increasing of the absolutely value of latitude degree \\((|\\varphi_{s}|)\\), the corresponding area on the spherical surface of an ERP microunit is gradually decreased to zero. In conclusion, ERP distortion is caused by variable stretching ratios \\(\\mathbf{K}_{\\mathrm{ERP}}\\), and is the heaviest in the polar areas." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.514, + 0.904, + 0.557 + ], + "angle": 0, + "content": "Fisheye distortion. The coordinate in Fisheye can be derived from \\(\\theta = \\arctan \\left(\\frac{y}{x}\\right)\\) and \\(\\varphi = (1 - \\sqrt{x^2 + y^2})\\times \\frac{\\pi}{2}\\). The stretching ratios of Fisheye can be derived as:" + }, + { + "type": "equation", + "bbox": [ + 0.567, + 0.565, + 0.903, + 0.603 + ], + "angle": 0, + "content": "\\[\n\\mathbf {K} _ {\\text {F i s h e y e}} (x, y) = \\frac {\\frac {2}{\\pi} \\sin \\left(\\frac {\\pi}{2} \\sqrt {x ^ {2} + y ^ {2}}\\right)}{\\sqrt {x ^ {2} + y ^ {2}}}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.612, + 0.684, + 0.628 + ], + "angle": 0, + "content": "where \\(\\sqrt{x^2 + y^2} \\in (0,1)\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.628, + 0.903, + 0.685 + ], + "angle": 0, + "content": "\\(\\mathbf{K}_{\\mathrm{Fisheye}}\\) is determined by distance from the fisheye center. As \\((\\mathbf{K}_{\\mathrm{Fisheye}})^{-1}\\) is bounded, fisheye projection is closer to uniform distribution than ERP. Moreover, it introduces much slighter distortion at the polar." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.685, + 0.903, + 0.756 + ], + "angle": 0, + "content": "Relationship between ERP and Fisheye distortions. To simplify, here we only discuss a typical Fisheye with an aperture degree of \\(\\pi\\) and a horizontal slicing plane4. In this case, the ERP coordinates and Fisheye's polar coordinates correspond linearly. We can quantize the relationship by:" + }, + { + "type": "equation", + "bbox": [ + 0.509, + 0.765, + 0.903, + 0.797 + ], + "angle": 0, + "content": "\\[\n\\mathbf {K} _ {\\mathrm {E R P} \\mid \\text {F i s h e y e}} (\\theta , \\varphi) = \\frac {\\mathbf {K} _ {\\mathrm {E R P}} \\left(x _ {E} , y _ {E}\\right)}{\\mathbf {K} _ {\\text {F i s h e y e}} \\left(x _ {F} , y _ {F}\\right)} = \\frac {\\pi}{2} - | \\varphi |, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.805, + 0.903, + 0.834 + ], + "angle": 0, + "content": "where \\(\\theta, \\varphi\\) are spherical coordinates on the sphere, \\(x_{E}, y_{E}\\) (\\(x_{F}, y_{F}\\)) denotes the plain coordinate under ERP (Fisheye)." + }, + { + "type": "page_footnote", + "bbox": [ + 0.516, + 0.841, + 0.897, + 0.854 + ], + "angle": 0, + "content": "3Detailed derivative processes can be found in the supplementary file." + }, + { + "type": "page_footnote", + "bbox": [ + 0.502, + 0.854, + 0.902, + 0.877 + ], + "angle": 0, + "content": "4The influence of Fisheye formats with arbitrary splicing plane is discussed in the supplementary file." + }, + { + "type": "list", + "bbox": [ + 0.502, + 0.841, + 0.902, + 0.877 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.466, + 0.892, + 0.505, + 0.903 + ], + "angle": 0, + "content": "13285" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.07, + 0.115, + 0.462, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.066, + 0.271, + 0.468, + 0.311 + ], + "angle": 0, + "content": "Figure 3. Downsampling process of ODIs (left) and imaging process in real world (right). * denotes that LR images synthesized from different downsampling processes are inconsistent." + }, + { + "type": "text", + "bbox": [ + 0.065, + 0.32, + 0.469, + 0.391 + ], + "angle": 0, + "content": "From Eqs. (2) to (4), we conclude that when uniformed downsampling is performed on ERP, the kernel size of equivalent Fisheye downsampling is non-uniformed. Especially when fisheye projection is spliced horizontally, the kernel size is proportional with \\(\\pi / 2 - |\\varphi|\\)." + }, + { + "type": "title", + "bbox": [ + 0.066, + 0.4, + 0.437, + 0.415 + ], + "angle": 0, + "content": "3.2. Learning with More Realistic Degradation" + }, + { + "type": "text", + "bbox": [ + 0.065, + 0.422, + 0.468, + 0.507 + ], + "angle": 0, + "content": "As depicted in Fig. 3, the original-formatted projection type in ODI acquiring process is fisheye projection. Given that real-world low-resolution issues are caused by insufficient precision and density of sensors, we consider that the degeneration process should be directly applied to original-formatted images before the type conversion." + }, + { + "type": "text", + "bbox": [ + 0.065, + 0.508, + 0.468, + 0.706 + ], + "angle": 0, + "content": "Ideally, as camera sensors are arranged in uniform arrays, pixel density on original-formatted images is consistent everywhere. Thus, for a realistic ODI, the pixel density on Fisheye should be a constant. As discussed in Sec. 3.1, applying uniformed downsampling on ERP means applying downsampling of variable kernel size on Fisheye. The variable kernel size leads to variable Fisheye pixel density, which results in unrealistic LR images. In conclusion, the ERP downsampling in previous methods influences the intrinsic distribution of pixel density in original-formatted images, which leads to unrealistic ODIs. When the downsampling process happens on Fisheye, the Fisheye pixel density is unchanged, which fits the real-world imaging process and synthesizes more realistic LR pairs." + }, + { + "type": "text", + "bbox": [ + 0.065, + 0.707, + 0.468, + 0.877 + ], + "angle": 0, + "content": "Process of Fisheye downsampling. To generate more realistic LR ODIs, we mimic the real-world imaging process and apply bicubic downsampling on Fisheye images. One single Fisheye image can only store information about a hemisphere. Hence, ERP images are converted to dual Fisheye images. Before downsampling, Fisheye images are padded by a FOV larger than \\(180^{\\circ}\\) to avoid edge disconnections. This padding operation will not influence the geometric transforming relation between ERP and Fisheye. As Fisheye data is unstructured and Fisheye distortion is more complicated than ERP distortion, we still learn the upscaling process under ERP. Thus we reconvert LR images to the" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.116, + 0.902, + 0.144 + ], + "angle": 0, + "content": "ERP format. The overall process of Fisheye downsampling are described in Fig. 3." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.153, + 0.824, + 0.166 + ], + "angle": 0, + "content": "3.3.OSRT:Modulate Distortion in ODIs" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.175, + 0.902, + 0.246 + ], + "angle": 0, + "content": "Overall. As discussed in Sec. 3.1, ERP images suffer a distortion caused by a non-consistency area stretching ratio from an ideal spherical surface. Referred from Eq. (2), for an LR input \\(X_{i}\\in \\mathbb{R}^{C\\times M\\times N}\\) , the distortion map \\(C_d\\in \\mathbb{R}^{1\\times M\\times N}\\) is derived by:" + }, + { + "type": "equation", + "bbox": [ + 0.586, + 0.255, + 0.902, + 0.286 + ], + "angle": 0, + "content": "\\[\nC _ {d} = \\cos \\left(\\frac {m + 0 . 5 - M / 2}{M} \\pi\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.295, + 0.788, + 0.309 + ], + "angle": 0, + "content": "where \\(m\\) is the current height of LR input." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.309, + 0.903, + 0.621 + ], + "angle": 0, + "content": "Previous methods tend to treat \\( C_d \\) as an additional input of \\( X_i \\) [28], or re-weighting parameters by \\( C_d \\) [18]. Although these solutions can benefit from building awareness of distortion, continuous and amorphous distortions cannot be adequately fitted by scattering and structured convolution operations. While previous methods cannot fully explore the advantage of \\( C_d \\), we intend to design a novel block for learning distorted patterns continuously. In VSR tasks, the deformable mechanism is proposed to align features between adjacent frames [32, 34]. Unlike standard DCN [10], which calculates offsets from the input feature map, offsets are calculated from bi-directional optical flow in VSR pipelines. Inspired by feature-level flow warping in VSR, we find that the deformable mechanism is a feasible solution for continuous mappings. Consequently, we modulate ERP distortion by feature-level warping operations. As shown in Fig. 4, \\( C_d \\) is only utilized to calculate the deformable offsets \\( \\theta \\). To keep compatibility with arbitrary ConvNets and Transformers, we propose two blocks to modulate ERP distortion, which can directly replace the multi-head self-attention blocks in Transformers and the standard convolution layers in ConvNets, respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.622, + 0.903, + 0.692 + ], + "angle": 0, + "content": "Distortion-aware attention block (DAAB). As depicted in Fig. 4 (a), a distortion condition guided deformable self-attention is proposed to learn correlations between the distorted input \\( X_{i} \\) and its corresponding modulated feature map \\( \\tilde{X}_{i} \\). DAAB is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.546, + 0.702, + 0.902, + 0.718 + ], + "angle": 0, + "content": "\\[\n\\theta_ {i} = H _ {\\text {o f f s e t} _ {i}} \\left(C _ {d}, C _ {w}\\right), \\tilde {X} _ {i} = \\phi \\left(X _ {i}; p _ {0} + \\theta_ {i}\\right), \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.574, + 0.727, + 0.902, + 0.744 + ], + "angle": 0, + "content": "\\[\nX _ {o} = H _ {\\mathrm {S A}} \\left(X _ {i} W _ {q _ {i}}, \\tilde {X} _ {i} W _ {k _ {i}}, \\tilde {X} _ {i} W _ {v _ {i}}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.749, + 0.903, + 0.877 + ], + "angle": 0, + "content": "where \\(H_{\\mathrm{offset}_i}(\\cdot)\\) denotes the \\(i\\)-th convolution block to calculate offset maps \\(\\theta_i \\in \\mathbb{R}^{2 \\times H \\times W}\\), and \\(H_{\\mathrm{SA}}\\) denotes standard self-attention formula. \\(H_{\\mathrm{offset}}(\\cdot)\\) consists of \\(1 \\times 1\\) convolution block with two hidden layers. The input of \\(H_{\\mathrm{offset}}(\\cdot)\\) is concatenated by the latitude-related distortion condition \\(C_d \\in \\mathbb{R}^{1 \\times H \\times W}\\) and the window condition \\(C_w \\in \\mathbb{R}^{2 \\times H \\times W}\\). \\(C_w\\) is a linear position encoding within a self-attention kernel. \\(\\phi(\\cdot, \\cdot)\\) denotes a bilinear interpolation, and \\(W_{q_i}, W_{k_i}, W_{v_i}\\) denote \\(i\\)-th weight matrix of query, key, and" + }, + { + "type": "page_number", + "bbox": [ + 0.466, + 0.892, + 0.505, + 0.903 + ], + "angle": 0, + "content": "13286" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.105, + 0.892, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.077, + 0.226, + 0.508, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.226, + 0.728, + 0.381 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.732, + 0.226, + 0.89, + 0.381 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.065, + 0.388, + 0.904, + 0.414 + ], + "angle": 0, + "content": "Figure 4. Overall illustration of OSRT. From SwinIR [23], we replace the standard multi-head self-attention block with DAAB and insert DACB behind the end of the RSTB. Channel dimensions of \\(\\theta_{\\mathrm{offsets}}\\) in DAAB and DACB are 2 and 18, respectively." + }, + { + "type": "text", + "bbox": [ + 0.066, + 0.419, + 0.467, + 0.447 + ], + "angle": 0, + "content": "value, respectively. For multi-head self-attention blocks, \\(H_{\\mathrm{offset}_i}(\\cdot)\\) is identical in calculations of parallel heads." + }, + { + "type": "text", + "bbox": [ + 0.065, + 0.448, + 0.468, + 0.504 + ], + "angle": 0, + "content": "Distortion-aware convolution block (DACB). As shown in Fig. 4 (b), we apply a standard deformable convolution layer with a substituted input for offset calculation. Modulated output \\( X_{o} \\) is extracted as:" + }, + { + "type": "equation", + "bbox": [ + 0.124, + 0.512, + 0.468, + 0.528 + ], + "angle": 0, + "content": "\\[\n\\theta_ {i} = H _ {\\text {o f f s e t} _ {i}} \\left(C _ {d}\\right), X _ {o} = H _ {\\mathrm {D C N} _ {i}} \\left(X _ {i}, \\theta_ {i}\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.065, + 0.535, + 0.468, + 0.592 + ], + "angle": 0, + "content": "where \\( H_{\\mathrm{DCN}}(X,\\theta) \\) denotes standard deformable convolution layer in [46]. The architecture of \\( H_{\\mathrm{offset}_i}(\\cdot) \\) is identical to that in DAAB. As the kernel size of DCN is \\( 3\\times 3 \\) in DACB, the output channel dimension of offsets maps is 18." + }, + { + "type": "text", + "bbox": [ + 0.065, + 0.592, + 0.468, + 0.72 + ], + "angle": 0, + "content": "OSRT. In practice, we propose an Omnidirectional image Super-Resolution Transformer, named OSRT. SwinIR [23] is selected as the basic architecture for its strong reconstruction ability in the SISR task. To learn distortion rectified representations, we stack a DACB after the last convolution layer of each residual swin Transformer block and replace all self-attention blocks as DAAB. The feature dimension of OSRT is reduced from 180 to 156 to maintain identical parameters with SwinIR." + }, + { + "type": "title", + "bbox": [ + 0.066, + 0.727, + 0.468, + 0.743 + ], + "angle": 0, + "content": "3.4. Boosting ODISR Performance by Plain Images" + }, + { + "type": "text", + "bbox": [ + 0.065, + 0.75, + 0.468, + 0.877 + ], + "angle": 0, + "content": "As the capacity of OSRT is relatively large, it suffers overfitting for large upscaling factors (Fig. 9). Given that acquiring ODIs are expensive, we propose to generate pseudo ERP images from 2D plain images to tackle this issue. After being sampled by sliding windows, the patch of plain images is treated as a plain perspective. By converting from Perspective to ERP, plain images are distorted in the same way as ERP. Considering that distortion of a Perspective is enlarged by its FOV degree, a relatively small FOV" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.419, + 0.902, + 0.449 + ], + "angle": 0, + "content": "degree of \\(90^{\\circ}\\) is applied. For a given pseudo Perspective, \\(\\theta_{p}\\) is fixed at 0 and \\(\\varphi_{p}\\) is derived by:" + }, + { + "type": "equation", + "bbox": [ + 0.648, + 0.463, + 0.902, + 0.479 + ], + "angle": 0, + "content": "\\[\n\\Phi_ {p} = \\varphi_ {h} + z _ {0}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.492, + 0.902, + 0.521 + ], + "angle": 0, + "content": "where \\(\\varphi_h\\) is determined by patch locations and \\(z_0\\) is orderly sampled from \\(\\{-15^\\circ, 0^\\circ, 15^\\circ\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.522, + 0.903, + 0.651 + ], + "angle": 0, + "content": "To maximize the approximate data distribution of ODIs, we horizontally split a plain image into three sub-images and define \\(\\varphi_h\\) as \\(-30^\\circ, 0^\\circ, 30^\\circ\\) respectively. Pseudo ERP images are cropped to remove the black border. As shown in Fig. 5, we get a new ERP dataset (called DF2K-ERP) by implementing the augmentation pipeline on widely-used plain image dataset DF2K [24, 33]. The DF2K-ERP dataset consists of 146K high-quality ERP image patches with a patch size larger than 256." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.662, + 0.897, + 0.853 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.57, + 0.86, + 0.833, + 0.873 + ], + "angle": 0, + "content": "Figure 5. Synthetic process of DF2K-ERP." + }, + { + "type": "page_number", + "bbox": [ + 0.466, + 0.892, + 0.505, + 0.903 + ], + "angle": 0, + "content": "13287" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.135, + 0.113, + 0.833, + 0.29 + ], + "angle": 0, + "content": "
MethodScaleODI-SRPSNRSUN 360 Panorama
PSNRSSIMWS-PSNRWS-SSIMSSIMWS-PSNRWS-SSIM
Bicubic28.210.821527.610.815628.140.811828.010.8321
RCAN [43]30.080.872329.490.871430.560.871231.180.8969
SRResNet [36]30.160.871729.590.869730.650.871431.200.8953
EDSR [24]×230.320.877029.680.872730.890.878431.420.8995
SwinIR [23]30.520.881929.870.877231.210.885231.780.9051
SwinIR†[23]30.640.882130.000.877731.330.885531.980.9059
OSRT†30.770.884630.110.879531.520.888832.140.9081
Bicubic25.590.711824.950.692325.290.699324.900.7083
RCAN [43]26.850.762126.150.748527.100.766026.990.7856
SRResNet [36]26.910.759726.240.745727.100.761826.990.7812
EDSR [24]×426.970.758926.300.745827.190.763327.100.7827
SwinIR [23]27.120.766326.440.752327.390.770727.300.7901
SwinIR†[23]27.310.773526.610.758927.710.780427.640.7996
OSRT†27.410.776226.700.760927.840.783527.770.8020
" + }, + { + "type": "table_caption", + "bbox": [ + 0.073, + 0.293, + 0.896, + 0.308 + ], + "angle": 0, + "content": "Table 1. SR results under Fisheye downsampling. † denotes applying DF2K-ERP as augmented dataset. Best results are shown in Bold." + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.314, + 0.3, + 0.47 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.143, + 0.474, + 0.239, + 0.485 + ], + "angle": 0, + "content": "SUN360 \\((\\times 4)\\) : 034" + }, + { + "type": "image", + "bbox": [ + 0.315, + 0.313, + 0.443, + 0.382 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.371, + 0.384, + 0.387, + 0.392 + ], + "angle": 0, + "content": "HR" + }, + { + "type": "image_caption", + "bbox": [ + 0.348, + 0.393, + 0.409, + 0.401 + ], + "angle": 0, + "content": "PSNR/SSIM" + }, + { + "type": "image", + "bbox": [ + 0.315, + 0.401, + 0.443, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.351, + 0.471, + 0.407, + 0.479 + ], + "angle": 0, + "content": "EDSR [24]" + }, + { + "type": "image_caption", + "bbox": [ + 0.341, + 0.481, + 0.418, + 0.489 + ], + "angle": 0, + "content": "26.38dB/0.8072" + }, + { + "type": "image", + "bbox": [ + 0.458, + 0.313, + 0.585, + 0.382 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.502, + 0.384, + 0.54, + 0.392 + ], + "angle": 0, + "content": "Bicubic" + }, + { + "type": "image_caption", + "bbox": [ + 0.483, + 0.393, + 0.56, + 0.401 + ], + "angle": 0, + "content": "24.38dB/0.6872" + }, + { + "type": "image", + "bbox": [ + 0.458, + 0.402, + 0.584, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.491, + 0.471, + 0.551, + 0.479 + ], + "angle": 0, + "content": "SwinJR [23]" + }, + { + "type": "image_caption", + "bbox": [ + 0.48, + 0.481, + 0.562, + 0.489 + ], + "angle": 0, + "content": "26.77dB/0.8234" + }, + { + "type": "image", + "bbox": [ + 0.6, + 0.313, + 0.728, + 0.382 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.636, + 0.383, + 0.692, + 0.391 + ], + "angle": 0, + "content": "RCAN [43]" + }, + { + "type": "image_caption", + "bbox": [ + 0.625, + 0.393, + 0.702, + 0.401 + ], + "angle": 0, + "content": "26.40dB/0.8137" + }, + { + "type": "image", + "bbox": [ + 0.6, + 0.402, + 0.728, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.628, + 0.47, + 0.698, + 0.479 + ], + "angle": 0, + "content": "SwinIR† [23]" + }, + { + "type": "image_caption", + "bbox": [ + 0.625, + 0.48, + 0.702, + 0.489 + ], + "angle": 0, + "content": "27.34dB/0.8462" + }, + { + "type": "image", + "bbox": [ + 0.743, + 0.313, + 0.87, + 0.382 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.768, + 0.383, + 0.842, + 0.391 + ], + "angle": 0, + "content": "SRResNet [36]" + }, + { + "type": "image_caption", + "bbox": [ + 0.767, + 0.393, + 0.845, + 0.401 + ], + "angle": 0, + "content": "26.21dB/0.7999" + }, + { + "type": "image", + "bbox": [ + 0.743, + 0.402, + 0.87, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.787, + 0.471, + 0.824, + 0.48 + ], + "angle": 0, + "content": "OSRT†" + }, + { + "type": "image_caption", + "bbox": [ + 0.767, + 0.48, + 0.844, + 0.489 + ], + "angle": 0, + "content": "27.68dB/0.8561" + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.496, + 0.3, + 0.652 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.143, + 0.658, + 0.239, + 0.667 + ], + "angle": 0, + "content": "SUN360 \\((\\times 4)\\) : 095" + }, + { + "type": "image", + "bbox": [ + 0.316, + 0.496, + 0.443, + 0.564 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.371, + 0.566, + 0.387, + 0.574 + ], + "angle": 0, + "content": "HR" + }, + { + "type": "image_caption", + "bbox": [ + 0.348, + 0.575, + 0.41, + 0.583 + ], + "angle": 0, + "content": "PSNR/SSIM" + }, + { + "type": "image", + "bbox": [ + 0.315, + 0.585, + 0.443, + 0.652 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.352, + 0.654, + 0.406, + 0.662 + ], + "angle": 0, + "content": "EDSR [24]" + }, + { + "type": "image_caption", + "bbox": [ + 0.34, + 0.664, + 0.418, + 0.673 + ], + "angle": 0, + "content": "33.64dB/0.9074" + }, + { + "type": "image", + "bbox": [ + 0.458, + 0.496, + 0.585, + 0.564 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.502, + 0.566, + 0.54, + 0.574 + ], + "angle": 0, + "content": "Bicubic" + }, + { + "type": "image_caption", + "bbox": [ + 0.483, + 0.575, + 0.56, + 0.583 + ], + "angle": 0, + "content": "30.20dB/0.8506" + }, + { + "type": "image", + "bbox": [ + 0.458, + 0.585, + 0.585, + 0.652 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.491, + 0.654, + 0.551, + 0.662 + ], + "angle": 0, + "content": "SwinIR [23]" + }, + { + "type": "image_caption", + "bbox": [ + 0.482, + 0.664, + 0.56, + 0.673 + ], + "angle": 0, + "content": "34.05dB/0.9119" + }, + { + "type": "image", + "bbox": [ + 0.599, + 0.496, + 0.728, + 0.564 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.634, + 0.566, + 0.692, + 0.574 + ], + "angle": 0, + "content": "RCAN [43]" + }, + { + "type": "image_caption", + "bbox": [ + 0.626, + 0.575, + 0.702, + 0.583 + ], + "angle": 0, + "content": "33.59dB/0.9088" + }, + { + "type": "image", + "bbox": [ + 0.599, + 0.584, + 0.728, + 0.652 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.629, + 0.654, + 0.697, + 0.662 + ], + "angle": 0, + "content": "SwinIR [23]" + }, + { + "type": "image_caption", + "bbox": [ + 0.625, + 0.664, + 0.702, + 0.672 + ], + "angle": 0, + "content": "34.41dB/0.9158" + }, + { + "type": "image", + "bbox": [ + 0.743, + 0.496, + 0.87, + 0.564 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.768, + 0.566, + 0.842, + 0.574 + ], + "angle": 0, + "content": "SRResNet [36]" + }, + { + "type": "image_caption", + "bbox": [ + 0.767, + 0.575, + 0.844, + 0.583 + ], + "angle": 0, + "content": "33.43dB/0.9043" + }, + { + "type": "image", + "bbox": [ + 0.743, + 0.584, + 0.87, + 0.652 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.786, + 0.654, + 0.824, + 0.662 + ], + "angle": 0, + "content": "OSRT†" + }, + { + "type": "image_caption", + "bbox": [ + 0.767, + 0.664, + 0.845, + 0.672 + ], + "angle": 0, + "content": "34.77dB/0.9187" + }, + { + "type": "image_caption", + "bbox": [ + 0.248, + 0.675, + 0.72, + 0.689 + ], + "angle": 0, + "content": "Figure 6. Visual comparisons of \\(\\times 4\\) SR results under Fisheye downsampling." + }, + { + "type": "title", + "bbox": [ + 0.066, + 0.699, + 0.202, + 0.714 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.066, + 0.725, + 0.26, + 0.74 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.065, + 0.749, + 0.469, + 0.877 + ], + "angle": 0, + "content": "ODI-SR dataset [12] and SUN360 Panorama dataset [39] are used in our experiment. In the training phase, we follow the data split setting in [12] and train on the ODI-SR training set. The resolution of the ERP HR is \\(1024 \\times 2048\\), and the upscaling factors are \\(\\times 2\\) and \\(\\times 4\\). Fisheye downsampling is applied as our pre-defined downsampling kernel. Loss is calculated by L1 distance and optimized by Adam [19], with an initial learning rate of \\(2 \\times 10^{-4}\\), a total batch size of 32, and an input patch size of 64. We" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.7, + 0.904, + 0.772 + ], + "angle": 0, + "content": "train OSRT for \\(500k\\) iterations and halve the learning rate at \\(250k\\), \\(400k\\), \\(450k\\) and \\(475k\\). In evaluation, we test on the ODI-SR testing set and SUN360 dataset. PSNR [16], SSIM [37], and their distortion re-weighted versions (WS-PSNR [31], WS-SSIM [45]) are used as evaluation metrics." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.783, + 0.863, + 0.799 + ], + "angle": 0, + "content": "4.2. Evaluation under Fisheye Downsampling" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.806, + 0.903, + 0.848 + ], + "angle": 0, + "content": "When the downsampling process is performed on Fisheye images, we train SRResNet [36], EDSR [24], RCAN [43], and SwinIR [23] for comparison." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.849, + 0.903, + 0.877 + ], + "angle": 0, + "content": "Quantitative results. As shown in Tab. 1, with the help of additional DF2K-ERP training patches, OSRT out" + }, + { + "type": "page_number", + "bbox": [ + 0.466, + 0.892, + 0.505, + 0.903 + ], + "angle": 0, + "content": "13288" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.156, + 0.113, + 0.813, + 0.239 + ], + "angle": 0, + "content": "
Scale×8×16
MethodODI-SRSUN 360 PanoramaODI-SRSUN 360 Panorama
WS-PSNRWS-SSIMWS-PSNRWS-SSIMWS-PSNRWS-SSIMWS-PSNRWS-SSIM
Bicubic19.640.590819.720.540317.120.433217.560.4638
SRCNN [13]20.080.611219.460.570118.080.450117.950.4684
EDSR [24]23.970.641722.460.634121.120.569821.060.5645
RCAN [43]24.260.662823.880.654221.940.582421.740.5742
360-SS [30]21.650.641721.480.635219.650.543119.620.5308
LAU-Net [12]24.360.680124.020.670822.070.590121.820.5824
SphereSR [40]24.370.677724.170.682022.510.637021.950.6342
OSRT24.530.678024.380.707222.690.626122.130.6388
" + }, + { + "type": "table_caption", + "bbox": [ + 0.342, + 0.242, + 0.625, + 0.256 + ], + "angle": 0, + "content": "Table 2. SR results under ERP downsampling." + }, + { + "type": "text", + "bbox": [ + 0.064, + 0.264, + 0.469, + 0.378 + ], + "angle": 0, + "content": "performs previous methods by 0.3dB on PSNR. Although directly applying SwinIR on the ODISR task has already reached SOTA performance, OSRT surpasses SwinIR over 0.1dB on two datasets for both \\(\\times 2\\) and \\(\\times 4\\) SR tasks, which demonstrates the effectiveness of its distortion modulation ability. The performance of RCAN degrades under Fisheye downsampling, which is caused by the incompatibility between channel attention and Fisheye downsampling5." + }, + { + "type": "text", + "bbox": [ + 0.064, + 0.378, + 0.469, + 0.507 + ], + "angle": 0, + "content": "Qualitative comparison. Fig. 6 shows the visualization results of \\(\\times 4\\) ODISR task. While other methods struggle to understand the geometric transformation process in distorted images, OSRT can reconstruct sharp and accurate boundaries with the advantages of distortion modulation. It is observed that OSRT is skilled at reconstructing rigid texture. Moreover, benefiting from the distortion modulation ability, OSRT can preserve the original structure as most when being projected to other projection types (Fig. 7)." + }, + { + "type": "title", + "bbox": [ + 0.065, + 0.516, + 0.406, + 0.532 + ], + "angle": 0, + "content": "4.3. Evaluation under ERP Downsampling" + }, + { + "type": "text", + "bbox": [ + 0.064, + 0.538, + 0.469, + 0.666 + ], + "angle": 0, + "content": "To compare with previous ODISR methods [12, 30, 40], we train OSRT under the previous ERP setting. Regardless of over-fitting issues, we only train on the dataset provided by [12] for fairness. As shown in Tab. 2, OSRT still outperforms LAU-Net [12] and SphereSR [40] under large upscaling factor and ERP downsampling. Without a complicated training pipeline and discrete inference process, OSRT yields the best PSNR values and surpasses all previous methods on most SSIM-related metrics (three of four)." + }, + { + "type": "title", + "bbox": [ + 0.065, + 0.676, + 0.345, + 0.691 + ], + "angle": 0, + "content": "4.4. Ablation Study and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.064, + 0.698, + 0.469, + 0.755 + ], + "angle": 0, + "content": "In this section, we prove the effectiveness of Fisheye downsampling, OSRT components, and augmented DF2K-ERP. We then explain the distortion modulation ability of OSRT by visualizing offsets in deformable blocks." + }, + { + "type": "text", + "bbox": [ + 0.064, + 0.755, + 0.469, + 0.854 + ], + "angle": 0, + "content": "Fisheye downsampling. As shown in Fig. 1, the SR model trained under ERP downsampling is more likely to generate blur details and missing structures in real-world scenarios. These artifacts cannot be removed by a superior backbone network, but can be eliminated by a more realistic imaging process. More importantly, ERP downsampling directly covers the geometric property of ERP images" + }, + { + "type": "table", + "bbox": [ + 0.507, + 0.262, + 0.892, + 0.352 + ], + "angle": 0, + "content": "
feature dimDACBDAABODI-SRSUN360Params. (M)
PSNRSSIMPSNRSSIM
60××30.270.873930.780.87420.91
60✓×30.410.877531.000.87931.16
60×w/o Cw30.310.874630.830.87551.00
60×w/ Cw30.320.874630.840.87531.01
60✓w/ Cw30.440.878031.040.88001.26
72××30.320.874830.850.87551.29
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.356, + 0.902, + 0.383 + ], + "angle": 0, + "content": "Table 3. Ablation study on OSRT components. All models are trained on \\(\\times 2\\) SR task under Fisheye downsampling." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.385, + 0.904, + 0.527 + ], + "angle": 0, + "content": "and makes the ODISR task identical to the standard plain image super-resolution task. The evidence is that a standard SISR model (SwinIR) trained on a plain image dataset (DF2K) can outperform previous SOTA in the ODISR task, which yields WS-PSNR results of 24.63dB/24.49dB (22.68dB/22.13dB) on \\(\\times 8\\) (\\(\\times 16\\)) ODI-SR/SUN360 testing set, respectively. In conclusion, when the intrinsic property of ODIs is broken by ERP downsampling, the ODISR task degenerates into a plain image super-resolution task with a particular data distribution." + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.529, + 0.692, + 0.662 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.538, + 0.663, + 0.658, + 0.683 + ], + "angle": 0, + "content": "ODI-SR \\((\\times 4)\\) :003 Fisheye (Vertical, Right)" + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.529, + 0.795, + 0.598 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.721, + 0.598, + 0.774, + 0.608 + ], + "angle": 0, + "content": "EDSR [24]" + }, + { + "type": "image", + "bbox": [ + 0.799, + 0.53, + 0.895, + 0.597 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.819, + 0.598, + 0.877, + 0.608 + ], + "angle": 0, + "content": "SwinIR [23]" + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.608, + 0.795, + 0.674 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.713, + 0.675, + 0.78, + 0.686 + ], + "angle": 0, + "content": "SwinIR† [23]" + }, + { + "type": "image", + "bbox": [ + 0.799, + 0.608, + 0.895, + 0.674 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.828, + 0.675, + 0.864, + 0.686 + ], + "angle": 0, + "content": "OSRT†" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.688, + 0.691, + 0.82 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.52, + 0.821, + 0.677, + 0.841 + ], + "angle": 0, + "content": "SUN360 \\((\\times 2)\\) : 062 Perspective \\((\\varphi :30^{\\circ}\\) ;FOV: \\(90^{\\circ}\\)" + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.688, + 0.794, + 0.755 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.721, + 0.756, + 0.774, + 0.766 + ], + "angle": 0, + "content": "EDSR [24]" + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.766, + 0.794, + 0.832 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.713, + 0.834, + 0.781, + 0.844 + ], + "angle": 0, + "content": "SwinIR† [23]" + }, + { + "type": "image", + "bbox": [ + 0.799, + 0.688, + 0.895, + 0.755 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.818, + 0.756, + 0.877, + 0.766 + ], + "angle": 0, + "content": "SwinIR [23]" + }, + { + "type": "image", + "bbox": [ + 0.799, + 0.766, + 0.895, + 0.833 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.828, + 0.834, + 0.864, + 0.844 + ], + "angle": 0, + "content": "OSRT†" + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.847, + 0.902, + 0.873 + ], + "angle": 0, + "content": "Figure 7. Visual comparisons for SR of Fisheye and Perspective images. \\(\\dagger\\) denotes applying DF2K-ERP as augmented dataset." + }, + { + "type": "page_footnote", + "bbox": [ + 0.084, + 0.864, + 0.355, + 0.877 + ], + "angle": 0, + "content": "5The cause is discussed in the supplementary file." + }, + { + "type": "page_number", + "bbox": [ + 0.466, + 0.892, + 0.505, + 0.903 + ], + "angle": 0, + "content": "13289" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.073, + 0.147, + 0.11, + 0.168 + ], + "angle": 0, + "content": "DAAB (29-th)" + }, + { + "type": "image", + "bbox": [ + 0.114, + 0.116, + 0.227, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.151, + 0.198, + 0.184, + 0.205 + ], + "angle": 0, + "content": "\\(\\omega = 0^{\\circ}\\)" + }, + { + "type": "image", + "bbox": [ + 0.232, + 0.117, + 0.346, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.266, + 0.198, + 0.307, + 0.205 + ], + "angle": 0, + "content": "\\(\\varphi = 45^{\\circ}\\)" + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.117, + 0.465, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.385, + 0.198, + 0.427, + 0.206 + ], + "angle": 0, + "content": "\\(\\varphi = 90^{\\circ}\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.073, + 0.241, + 0.11, + 0.26 + ], + "angle": 0, + "content": "DAAB (33-th)" + }, + { + "type": "image", + "bbox": [ + 0.113, + 0.207, + 0.227, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.151, + 0.289, + 0.185, + 0.297 + ], + "angle": 0, + "content": "\\(\\varphi = 0^{\\circ}\\)" + }, + { + "type": "image", + "bbox": [ + 0.232, + 0.207, + 0.346, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.266, + 0.289, + 0.307, + 0.297 + ], + "angle": 0, + "content": "\\(\\varphi = 45^{\\circ}\\)" + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.207, + 0.465, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.385, + 0.289, + 0.427, + 0.297 + ], + "angle": 0, + "content": "\\(\\varphi = 90^{\\circ}\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.074, + 0.324, + 0.109, + 0.333 + ], + "angle": 0, + "content": "DACB" + }, + { + "type": "image", + "bbox": [ + 0.113, + 0.299, + 0.198, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.135, + 0.36, + 0.169, + 0.368 + ], + "angle": 0, + "content": "\\(\\varphi = 0^{\\circ}\\)" + }, + { + "type": "image", + "bbox": [ + 0.203, + 0.299, + 0.287, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.222, + 0.36, + 0.264, + 0.368 + ], + "angle": 0, + "content": "\\(\\varphi = 30^{\\circ}\\)" + }, + { + "type": "image", + "bbox": [ + 0.292, + 0.299, + 0.376, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.311, + 0.36, + 0.352, + 0.368 + ], + "angle": 0, + "content": "\\(\\varphi = 60^{\\circ}\\)" + }, + { + "type": "image", + "bbox": [ + 0.382, + 0.299, + 0.465, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.4, + 0.36, + 0.442, + 0.368 + ], + "angle": 0, + "content": "\\(\\varphi = 90^{\\circ}\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.065, + 0.376, + 0.469, + 0.416 + ], + "angle": 0, + "content": "Figure 8. Visualizations of offset maps in OSRT. Reference and deformed points are depicted in green and red, respectively. The deformable kernel is sparse in the polar area." + }, + { + "type": "text", + "bbox": [ + 0.064, + 0.424, + 0.468, + 0.566 + ], + "angle": 0, + "content": "OSRT components. To study the effectiveness of each component in OSRT, we propose a light version of OSRT (OSRT-light) for ablation study, which corresponds with the official SwinIR-light [23]. As proofed in Tab. 3, all components in OSRT are beneficial for modulating ERP distortion. The advantages of DACB and DAAB can be stacked when being applied in the same network. Compared with simply expanding the feature dimension of SwinIR to match the network complexity, the overall improvements of OSRT is more significant \\((+0.05\\mathrm{dB}\\) vs. \\(+0.2\\mathrm{dB})\\)" + }, + { + "type": "text", + "bbox": [ + 0.064, + 0.569, + 0.469, + 0.64 + ], + "angle": 0, + "content": "Offsets in OSRT. Offsets map in a well-trained OSRT are visualized in Fig. 8. Deformable kernels in both DAAB and DACB tend to gather at the equator and scatter at the polar, which conforms to the geometric distribution of pixel density in ERP images." + }, + { + "type": "image", + "bbox": [ + 0.068, + 0.651, + 0.469, + 0.844 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.065, + 0.847, + 0.468, + 0.873 + ], + "angle": 0, + "content": "Figure 9. Training process of Transformers on \\(\\times 4\\) ODISR task. The overfitting issue is tackled by our augmentation scheme." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.114, + 0.885, + 0.246 + ], + "angle": 0, + "content": "
Backbone networkDatasetsTraining schemeScaleSUN360
PSNRSSIM
SwinIRODI-SRN/A31.210.8852
SwinIRDF2K/ODI-SRone-stage31.260.8841
SwinIRDF2K-ERP/ODI-SRone-stage×231.330.8855
SwinIRDF2K-ERP/ODI-SRtwo-stage31.170.8818
OSRTDF2K-ERP/ODI-SRone-stage31.520.8888
SwinIRODI-SRN/A27.390.7707
SwinIRDF2K/ODI-SRone-stage27.590.7768
SwinIRDF2K-ERP/ODI-SRone-stage×427.710.7804
SwinIRDF2K-ERP/ODI-SRtwo-stage27.740.7795
OSRTDF2K-ERP/ODI-SRone-stage27.840.7835
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.248, + 0.902, + 0.275 + ], + "angle": 0, + "content": "Table 4. Ablation study on data augmentation. The results of ODISR (In the supplementary file) are in the same trend as SUN360." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.284, + 0.905, + 0.497 + ], + "angle": 0, + "content": "Pseudo ERP patches. In Sec. 3.4, we propose a distorted dataset DF2K-ERP to tackle over-fitting issues. We train a standard SwinIR on diverse datasets and training schemes to study the influence of data augmentation separately. As shown in Tab. 4, while training on ODI-SR and DF2K, distortion operations in DF2K lead to better performance. Compared with fine-tuning on DF2K-ERP pretrained models (two-stage), training on two datasets jointly (one-stage) shows better results. We infer that there is a domain gap between ODI-SR and DF2K-ERP, which is caused by omitted Perspective distortion. Moreover, the advantage of distortion modulation mechanisms in OSRT is enlarged when additional training patches are applied. Fig. 9 proves that our data augmentation scheme overcomes the over-fitting issue and improves the reconstruction ability." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.511, + 0.623, + 0.525 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.535, + 0.904, + 0.748 + ], + "angle": 0, + "content": "In this paper, we find that the previous downsampling process in the ODISR task harms the intrinsic distribution of pixel density in ODIs, which leads to poor generalization ability in real-world scenarios. To tackle this issue, we propose Fisheye downsampling, which mimics the real-world imaging process to preserve the realistic density distribution. After refining the downsampling process, we design a distortion-aware Transformer (OSRT) to modulate distortions continuously and self-adaptively. OSRT learns offsets from the distortion-related condition and rectifies distortion by feature-level warping. Moreover, to alleviate the overfitting problem of large networks, we propose to synthesize additional ERP training data from the plain images. Extensive experiments have demonstrated the state-of-the-art performance of our OSRT." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.749, + 0.902, + 0.777 + ], + "angle": 0, + "content": "Limitation. The process of sampling ERP images into viewing types also requires careful design." + }, + { + "type": "title", + "bbox": [ + 0.52, + 0.778, + 0.662, + 0.791 + ], + "angle": 0, + "content": "Acknowledgement." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.792, + 0.902, + 0.877 + ], + "angle": 0, + "content": "This work was supported in part by the National Key R&D Program of China (NO.2022ZD0160505), in part by the National Natural Science Foundation of China under Grant (62276251), the Joint Lab of CAS-HK, and in part by the Youth Innovation Promotion Association of Chinese Academy of Sciences (No.2020356)." + }, + { + "type": "page_number", + "bbox": [ + 0.466, + 0.892, + 0.506, + 0.904 + ], + "angle": 0, + "content": "13290" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.067, + 0.114, + 0.165, + 0.129 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.075, + 0.138, + 0.47, + 0.191 + ], + "angle": 0, + "content": "[1] Hao Ai, Zidong Cao, Jinjing Zhu, Haotian Bai, Yucheng Chen, and Ling Wang. Deep learning for omnidirectional vision: A survey and new perspectives. arXiv preprint arXiv:2205.10468, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.075, + 0.192, + 0.469, + 0.245 + ], + "angle": 0, + "content": "[2] Zafer Arican and Pascal Frossard. L1 regularized superresolution from unregistered omnidirectional images. In 2009 IEEE International Conference on Acoustics, Speech and Signal Processing, pages 829-832. IEEE, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.076, + 0.247, + 0.469, + 0.286 + ], + "angle": 0, + "content": "[3] Zafer Arican and Pascal Frossard. Joint registration and super-resolution with omnidirectional images. IEEE Transactions on Image Processing, 20(11):3151-3162, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.076, + 0.288, + 0.469, + 0.352 + ], + "angle": 0, + "content": "[4] Luigi Bagnato, Yannick Boursier, Pascal Frossard, and Pierre Vandergheynst. Plenoptic based super-resolution for omnidirectional image sequences. In 2010 IEEE International Conference on Image Processing, pages 2829-2832. IEEE, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.076, + 0.354, + 0.469, + 0.42 + ], + "angle": 0, + "content": "[5] Kelvin CK Chan, Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Basicvsr: The search for essential components in video super-resolution and beyond. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4947-4956, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.076, + 0.421, + 0.468, + 0.487 + ], + "angle": 0, + "content": "[6] Kelvin CK Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. Basicvsr++: Improving video superresolution with enhanced propagation and alignment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5972-5981, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.076, + 0.488, + 0.469, + 0.554 + ], + "angle": 0, + "content": "[7] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12299-12310, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.076, + 0.555, + 0.468, + 0.594 + ], + "angle": 0, + "content": "[8] Xiangyu Chen, Xintao Wang, Jiantao Zhou, and Chao Dong. Activating more pixels in image super-resolution transformer. arXiv preprint arXiv:2205.04437, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.076, + 0.596, + 0.468, + 0.66 + ], + "angle": 0, + "content": "[9] Yinbo Chen, Sifei Liu, and Xiaolong Wang. Learning continuous image representation with local implicit image function. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8628-8638, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.068, + 0.662, + 0.468, + 0.716 + ], + "angle": 0, + "content": "[10] Jifeng Dai, Haozhi Qi, Yuwen Xiong, Yi Li, Guodong Zhang, Han Hu, and Yichen Wei. Deformable convolutional networks. In Proceedings of the IEEE international conference on computer vision, pages 764-773, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.068, + 0.717, + 0.468, + 0.782 + ], + "angle": 0, + "content": "[11] Tao Dai, Jianrui Cai, Yongbing Zhang, Shu-Tao Xia, and Lei Zhang. Second-order attention network for single image super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11065-11074, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.068, + 0.784, + 0.468, + 0.849 + ], + "angle": 0, + "content": "[12] Xin Deng, Hao Wang, Mai Xu, Yichen Guo, Yuhang Song, and Li Yang. Lau-net: Latitude adaptive upscaling network for omnidirectional image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9189-9198, 2021. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.068, + 0.851, + 0.468, + 0.877 + ], + "angle": 0, + "content": "[13] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Image super-resolution using deep convolutional net" + }, + { + "type": "list", + "bbox": [ + 0.068, + 0.138, + 0.47, + 0.877 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.536, + 0.117, + 0.902, + 0.143 + ], + "angle": 0, + "content": "works. IEEE transactions on pattern analysis and machine intelligence, 38(2):295-307, 2015. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.145, + 0.902, + 0.197 + ], + "angle": 0, + "content": "[14] Chao Dong, Chen Change Loy, and Xiaou Tang. Accelerating the super-resolution convolutional neural network. In European conference on computer vision, pages 391-407. Springer, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.198, + 0.902, + 0.25 + ], + "angle": 0, + "content": "[15] Vida Fakour-Sevom, Esin Guldogan, and Joni-Kristian Kämäräinen. 360 panorama super-resolution using deep convolutional networks. In Int. Conf. on Computer Vision Theory and Applications (VISAPP), volume 1, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.251, + 0.902, + 0.29 + ], + "angle": 0, + "content": "[16] Quan Huynh-Thu and Mohammed Ghanbari. Scope of validity of psnr in image/video quality assessment. *Electronics letters*, 44(13):800-801, 2008. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.292, + 0.902, + 0.356 + ], + "angle": 0, + "content": "[17] Hiroshi Kawasaki, Katsushi Ikeuchi, and Masao Sakauchi. Super-resolution omnidirectional camera images using spatio-temporal analysis. *Electronics and Communications in Japan* (Part III: Fundamental Electronic Science), 89(6):47-59, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.358, + 0.902, + 0.41 + ], + "angle": 0, + "content": "[18] Renata Khasanova and Pascal Frossard. Geometry aware convolutional filters for omnidirectional images representation. In International Conference on Machine Learning, pages 3351-3359. PMLR, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.412, + 0.902, + 0.45 + ], + "angle": 0, + "content": "[19] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.452, + 0.902, + 0.543 + ], + "angle": 0, + "content": "[20] Christian Ledig, Lucas Theis, Ferenc Huszar, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photorealistic single image super-resolution using a generative adversarial network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4681-4690, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.545, + 0.902, + 0.584 + ], + "angle": 0, + "content": "[21] Wenbo Li, Xin Lu, Jiangbo Lu, Xiangyu Zhang, and Jiaya Jia. On efficient transformer and image pre-training for low-level vision. arXiv preprint arXiv:2112.10175, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.585, + 0.902, + 0.65 + ], + "angle": 0, + "content": "[22] Zheyuan Li, Yingqi Liu, Xiangyu Chen, Haoming Cai, Jinjin Gu, Yu Qiao, and Chao Dong. Blueprint separable residual network for efficient image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 833-843, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.652, + 0.902, + 0.716 + ], + "angle": 0, + "content": "[23] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1833-1844, 2021. 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.718, + 0.902, + 0.783 + ], + "angle": 0, + "content": "[24] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 136-144, 2017. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.785, + 0.902, + 0.849 + ], + "angle": 0, + "content": "[25] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.851, + 0.902, + 0.877 + ], + "angle": 0, + "content": "[26] Yiqun Mei, Yuchen Fan, and Yuqian Zhou. Image superresolution with non-local sparse attention. In Proceedings of" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.117, + 0.902, + 0.877 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.467, + 0.892, + 0.503, + 0.903 + ], + "angle": 0, + "content": "13291" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.117, + 0.468, + 0.142 + ], + "angle": 0, + "content": "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3517-3526, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.069, + 0.144, + 0.468, + 0.234 + ], + "angle": 0, + "content": "[27] Hajime Nagahara, Yasushi Yagi, and Masahiko Yachida. Super-resolution from an omnidirectional image sequence. In 2000 26th Annual Conference of the IEEE Industrial Electronics Society. IECON 2000. 2000 IEEE International Conference on Industrial Electronics, Control and Instrumentation. 21st Century Technologies, volume 4, pages 2559-2564. IEEE, 2000. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.069, + 0.236, + 0.468, + 0.299 + ], + "angle": 0, + "content": "[28] Akito Nishiyama, Satoshi Ikehata, and Kiyoharu Aizawa. 360 single image super resolution via distortion-aware network and distorted perspective images. In 2021 IEEE International Conference on Image Processing (ICIP), pages 1829-1833. IEEE, 2021. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.069, + 0.3, + 0.468, + 0.365 + ], + "angle": 0, + "content": "[29] Ben Niu, Weilei Wen, Wenqi Ren, Xiangde Zhang, Lianping Yang, Shuzhen Wang, Kaihao Zhang, Xiaochun Cao, and Haifeng Shen. Single image super-resolution via a holistic attention network. In European conference on computer vision, pages 191-207. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.069, + 0.367, + 0.468, + 0.43 + ], + "angle": 0, + "content": "[30] Cagri Ozcinar, Aakanksha Rana, and Aljosa Smolic. Superresolution of omnidirectional images using adversarial learning. In 2019 IEEE 21st International Workshop on Multimedia Signal Processing (MMSP), pages 1-6. IEEE, 2019. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.069, + 0.432, + 0.468, + 0.47 + ], + "angle": 0, + "content": "[31] Yule Sun, Ang Lu, and Lu Yu. Weighted-to-spherically-uniform quality evaluation for omnidirectional video. IEEE signal processing letters, 24(9):1408-1412, 2017. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.069, + 0.471, + 0.468, + 0.535 + ], + "angle": 0, + "content": "[32] Yapeng Tian, Yulun Zhang, Yun Fu, and Chenliang Xu. Tdan: Temporally-deformable alignment network for video super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3360-3369, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.069, + 0.537, + 0.468, + 0.601 + ], + "angle": 0, + "content": "[33] Radu Timofte, Eirikur Agustsson, Luc Van Gool, Ming-Hsuan Yang, and Lei Zhang. Ntire 2017 challenge on single image super-resolution: Methods and results. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 114-125, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.069, + 0.602, + 0.468, + 0.667 + ], + "angle": 0, + "content": "[34] Xintao Wang, Kelvin CK Chan, Ke Yu, Chao Dong, and Chen Change Loy. Edvr: Video restoration with enhanced deformable convolutional networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 0–0, 2019. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.069, + 0.668, + 0.468, + 0.731 + ], + "angle": 0, + "content": "[35] Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1905-1914, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.069, + 0.733, + 0.468, + 0.798 + ], + "angle": 0, + "content": "[36] Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Yu Qiao, and Chen Change Loy. Esrgan: Enhanced super-resolution generative adversarial networks. In Proceedings of the European conference on computer vision (ECCV) workshops, pages 0–0, 2018. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.069, + 0.799, + 0.468, + 0.85 + ], + "angle": 0, + "content": "[37] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.069, + 0.851, + 0.468, + 0.877 + ], + "angle": 0, + "content": "[38] Zhuofan Xia, Xuran Pan, Shiji Song, Li Erran Li, and Gao Huang. Vision transformer with deformable attention. In" + }, + { + "type": "list", + "bbox": [ + 0.069, + 0.117, + 0.468, + 0.877 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.117, + 0.901, + 0.143 + ], + "angle": 0, + "content": "Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4794-4803, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.145, + 0.903, + 0.208 + ], + "angle": 0, + "content": "[39] Jianxiong Xiao, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Recognizing scene viewpoint using panoramic place representation. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, pages 2695-2702. IEEE, 2012. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.211, + 0.903, + 0.286 + ], + "angle": 0, + "content": "[40] Youngho Yoon, Inchul Chung, Lin Wang, and Kuk-Jin Yoon. Spheresr: 360deg image super-resolution with arbitrary projection via continuous spherical image representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5677-5686, 2022. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.29, + 0.903, + 0.353 + ], + "angle": 0, + "content": "[41] Kai Zhang, Jingyun Liang, Luc Van Gool, and Radu Timofte. Designing a practical degradation model for deep blind image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4791-4800, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.356, + 0.903, + 0.42 + ], + "angle": 0, + "content": "[42] Wenlong Zhang, Yihao Liu, Chao Dong, and Yu Qiao. Ranksrgan: Generative adversarial networks with ranker for image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3096-3105, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.422, + 0.903, + 0.487 + ], + "angle": 0, + "content": "[43] Yulun Zhang, Kunpeng Li, Kai Li, Lichen Wang, Bineng Zhong, and Yun Fu. Image super-resolution using very deep residual channel attention networks. In Proceedings of the European conference on computer vision (ECCV), pages 286-301, 2018. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.489, + 0.903, + 0.564 + ], + "angle": 0, + "content": "[44] Yupeng Zhang, Hengzhi Zhang, Daojing Li, Liyan Liu, Hong Yi, Wei Wang, Hiroshi Suitoh, and Makoto Odamaki. Toward real-world panoramic image enhancement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 628-629, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.568, + 0.903, + 0.632 + ], + "angle": 0, + "content": "[45] Yufeng Zhou, Mei Yu, Hualin Ma, Hua Shao, and Gangyi Jiang. Weighted-to-spherically-uniform sSIM objective quality evaluation for panoramic video. In 2018 14th IEEE International Conference on Signal Processing (ICSP), pages 54-57. IEEE, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.634, + 0.903, + 0.687 + ], + "angle": 0, + "content": "[46] Xizhou Zhu, Han Hu, Stephen Lin, and Jifeng Dai. Deformable convnets v2: More deformable, better results. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9308-9316, 2019. 5" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.117, + 0.903, + 0.687 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.467, + 0.892, + 0.505, + 0.903 + ], + "angle": 0, + "content": "13292" + } + ] +] \ No newline at end of file diff --git a/2023/OSRT_ Omnidirectional Image Super-Resolution With Distortion-Aware Transformer/b6aea18d-97cd-464e-a4ad-626c19fdc060_origin.pdf b/2023/OSRT_ Omnidirectional Image Super-Resolution With Distortion-Aware Transformer/b6aea18d-97cd-464e-a4ad-626c19fdc060_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..acf8704b7f389d23214a92f3e88d76f903525a15 --- /dev/null +++ b/2023/OSRT_ Omnidirectional Image Super-Resolution With Distortion-Aware Transformer/b6aea18d-97cd-464e-a4ad-626c19fdc060_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ec5eb8bb4a429b7a144e8aa1b4bd0a773fc3d3b924389dba6ca7eb96cf6eb793 +size 2212002 diff --git a/2023/OSRT_ Omnidirectional Image Super-Resolution With Distortion-Aware Transformer/full.md b/2023/OSRT_ Omnidirectional Image Super-Resolution With Distortion-Aware Transformer/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d003439494867dbae2d0fbca6bcbbf4ccd7cd77a --- /dev/null +++ b/2023/OSRT_ Omnidirectional Image Super-Resolution With Distortion-Aware Transformer/full.md @@ -0,0 +1,436 @@ +# OSRT: Omnidirectional Image Super-Resolution with Distortion-aware Transformer + +Fanghua Yu $^{1*}$ Xintao Wang $^{2*}$ Mingdeng Cao $^{2,3}$ Gen Li $^{4}$ Ying Shan $^{2}$ Chao Dong $^{1,5\dagger}$ $^{1}$ ShenZhen Key Lab of Computer Vision and Pattern Recognition +Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences + $^{2}$ ARC, Tencent PCG ${}^{3}$ The University of Tokyo + +4 Platform Technologies, Tencent Online Video Shanghai Artificial Intelligence Laboratory + +fanghuayu96@gmail.com, xintaowang@tencent.com, cmd@g.ecc.u-tokyo.ac.jp + +{genli, yingsshan}@tencent.com, chao.dong@siat.ac.cn + +# Abstract + +Omnidirectional images (ODIs) have obtained lots of research interest for immersive experiences. Although ODIs require extremely high resolution to capture details of the entire scene, the resolutions of most ODIs are insufficient. Previous methods attempt to solve this issue by image super-resolution (SR) on equirectangular projection (ERP) images. However, they omit geometric properties of ERP in the degradation process, and their models can hardly generalize to real ERP images. In this paper, we propose Fisheye downsampling, which mimics the real-world imaging process and synthesizes more realistic low-resolution samples. Then we design a distortion-aware Transformer (OSRT) to modulate ERP distortions continuously and self-adaptively. Without a cumbersome process, OSRT outperforms previous methods by about 0.2dB on PSNR. Moreover, we propose a convenient data augmentation strategy, which synthesizes pseudo ERP images from plain images. This simple strategy can alleviate the over-fitting problem of large networks and significantly boost the performance of ODISR. Extensive experiments have demonstrated the state-of-the-art performance of our OSRT. + +# 1. Introduction + +In pursuit of the realistic visual experience, omnidirectional images (ODIs), also known as $360^{\circ}$ images or panoramic images, have obtained lots of research interest in the computer vision community. In reality, we usually view ODIs with a narrow field-of-view (FOV), e.g., viewing in a headset. To capture details of the entire scene, ODIs require extremely high resolution, e.g., $4\mathrm{K} \times 8\mathrm{K}$ [1]. However, due to the high industrial cost of camera sensors with high precision, the resolutions of most ODIs are insufficient. + +Recently, some attempts have been made to solve this problem by image super-resolution (SR) [12, 15, 28, 39, 40]. + +![](images/5aadb5dd34e8dc3d24b12e025b3c1637a487761ffde252181d26af69da3856a5.jpg) +Unseen LR + +![](images/b9145473a0eb5e8a56509988face13f4e65bd2a193ad7d1b3a0a1f8980421b57.jpg) +LAU-Net [12] w/o Fisheye + +![](images/ae3130b9f69db040cc2b88220b207a6a8af801674510da05fa5bff35b377ed7e.jpg) +OSRT w/o Fisheye + +![](images/b255ba70f6bcacee07d49fa892047a8c4a2f0b7753dbac13d64c8af1dac46e81.jpg) +OSRT w/ Fisheye +Figure 1. Visual comparisons of $\times 8$ SR results on LR images1 with unknown degradations. Fisheye denotes that the downsampling process in training stages is under Fisheye images. + +As most of the ODIs are stored and transmitted in the equirectangular projection (ERP) type, the SR process is usually performed on the ERP images. To generate high- /low-resolution training pairs, existing ODISR methods [12, 15, 28, 39, 40] directly apply uniform bicubic down-sampling on the original ERP images (called ERP down-sampling), which is identical to general image SR settings [24, 43]. While omitting geometric properties of ERP in the degradation process, their models can hardly generalize to real ERP images. We can observe missing structures and blur textures in Fig. 1. Therefore, we need a more appropriate degradation model before studying SR algorithms. In practice, ODIs are acquired by the fisheye lens and stored in ERP. Given that the low-resolution issue in real-world scenarios is caused by insufficient sensor precision and density, the downsampling process should be applied to original-formatted images before converting into other storage types. Thus, to be conformed with real-world imaging processes, we propose to apply uniform bicubic + +downsampling on Fisheye images, which are the original format of ODIs. The new downsampling process (called Fisheye downsampling) applies uniform bicubic downsampling on Fisheye images before converting them to ERP images. Our Fisheye downsampling is more conducive to exploring the geometric property of ODIs. + +The key issue of ODISR algorithm design is to utilize the geometric properties of ERP images, which is also the focus of previous methods. For example, Nishiyama et al. [28] add a distortion-related condition as an additional input. LAU-Net [12] splits the whole ERP image into patches by latitude band and learns upscaling processes separately. However, the separated learning process will lead to information disconnection between adjacent patches. SphereSR [40] learns different upscaling functions on various projection types, but will inevitably introduce multiple-time computation costs. To push the performance upper bound, we propose the first Transformer for Omnidirectional image Super-Resolution (OSRT), and incorporate geometric properties in a distortion-aware manner. Specifically, to modulate distorted feature maps, we implement feature-level warping, in which offsets are learned from latitude conditions. In OSRT, we introduce two dedicated blocks to adapt latitude-related distortion: distortion-aware attention block (DAAB), and distortion-aware convolution block (DACB). DAAB and DACB are designed to perform distortion modulation in arbitrary Transformers and ConvNets. These two blocks can directly replace the multi-head self-attention block and convolution layer, respectively. The benefit of DAAB and DACB can be further improved when being inserted into the same backbone network. OSRT outperforms previous methods by about 0.2dB on PSNR (Tab. 2). + +However, the increase of network capacity will also enlarge the overfitting problem of ODISR, which is rarely mentioned before. The largest ODIs dataset [12] contains only 1K images, which cannot provide enough diversity for training Transformers. Given that acquiring ODIs requires expensive equipment and tedious work, we propose to generate distorted ERP samples from plain images for data augmentation. In practice, we regard a plain image as a sampled perspective, and project it back to the ERP format. Then we can introduce 146K additional training patches, 6 times of the previous dataset. This simple strategy can significantly boost the performance of ODISR (Tab. 4) and alleviate the over-fitting problem of large networks (Fig. 9). A similar data augmentation method is also applied in Nishiyama et al. [28], but shows marginal improvement on small models under ERP downsampling settings. + +Our contributions are threefold. 1) For problem formulation: To generate more realistic ERP low-resolution images, we propose Fisheye downsampling, which mimics the real-world imaging process. 2) For method: Combined with the geometric properties of ERP, we design a distortion-aware + +Transformer, which modulates distortions continuously and self-adaptively without cumbersome process. 3) For data: To reduce overfitting, we propose a convenient data augmentation strategy, which synthesizes pseudo ERP images from plain images. Extensive experiments have demonstrated the state-of-the-art performance of our OSRT2. + +# 2. Related Work + +Single Image Super-Resolution (SISR). Deep learning for single image SR (SISR) is first introduced in [13]. Further works boost SR performance by CNNs [11, 14, 22, 24, 26, 29, 43], Vision Transformers (ViTs) [7, 8, 21, 23] and generative adversarial networks (GANs) [20, 35, 36, 42]. For instance, EDSR [24] removes Batch Normalization layers and applies a more complicated residual block. RCAN [43] introduces channel-wise attention mechanisms to a deeper network. SwinIR [23] proposes an image restoration Transformer based on [25]. To improve perceptual quality, adversarial training are performed as a tuning process to generate more realistic results [35, 36]. Moreover, various flexible degradation models are proposed in [35, 41] to synthesize more practical degradations. + +Omnidirectional Image Super-Resolution (ODISR). Initially, ODISR models focus on the spherical assembling of LR ODIs under various projection types [2-4, 17, 27]. Recent ODISR models are performed on plane images and are fine-tuned from existing SISR models with L1 loss [15] or GAN loss [30, 44]. The improvements are limited, for they only concern the distribution gap between ODIs and plain images. Since LAU-Net [12] found pixel density in ERP ODIs is non-uniform, many studies attempt to design specific backbone networks to overcome this issue. LAU-Net [12] manually splits the whole ERP image into latitude-related patches and learns ERP distortion over different latitude ranges separately. While LAU-Net learns latitude-related ERP distortion somewhat, its non-overlapped patches lead to disconnection in whole ERP images. Nishiyama et al. [28] treats area stretching ratio as additional input. However, these conditions are tough to be utilized with an unmodified SISR backbone network. SphereSR [40] learns upsampling processes on various projection types to mitigate the influence of non-uniformity in specific projection types. Although SphereSR improves information consistency between various ODI projection types by LIIF [9], they apply multiple networks to learn the upscaling process of each projection type. Given that all other projection types in SphereSR are converted from ERP, patterns under various types are reusable when distortions are properly rectified. Moreover, the complex and unstructured image data in polyhedron projection hinders further research of ODISR. + +Deformable Mechanism. Dai et al. [30] first propose deformable convolutions to obtain information out of its regular neighborhood. Xia et al. [38] further verified that Vision Transformers also benefit from applying deformable mechanisms on self-attention blocks. In Video SR tasks, the deformable mechanism can be adapted to align features between adjacent frames [5,6,34]. + +# 3. Method + +In this section, we first analyze the cause of ERP and Fisheye distortions, as well as the relationship between these two distortions (Sec. 3.1). Then, we discuss the designs of Fisheye downsampling (Sec. 3.2), distortion-aware Transformer (OSRT) (Sec. 3.3), and the convenient data augmentation strategy (Sec. 3.4). + +# 3.1. Revisiting Distortions in ODIs + +![](images/fb8333e243e50452470bb4f9fe79baddd5ab63b25639d7b4b09b9ed9068e956d.jpg) +Figure 2. Geometric explanation of transforming between ERP, Fisheye, and the ideal spherical surface. To simplify, we discuss the horizontal spliced Fisheye with an aperture degree of $\pi$ . + +As ODIs under each projection type are constrained by different transforming equations, the distortion caused by each type is inconsistent, indicating that applying matrix operations under one projection type can introduce unexpected changes when being converted to other types. Specifically, applying uniformed bicubic downsampling on ERP images will affect the distribution of pixel density on Fisheye images, which are the original-formatted image type of imaging process in real-world scenarios. To analyze the specific effect of ERP downsampling on the Fisheye image, we revisit the cause of distortions in ERP and Fisheye. + +As we assume that viewing directions are uniformly distributed, the data points in an ideal ODI should be uniformly distributed on a spherical surface. In practice, there is a trade-off between the uniformity of the spherical surface and the structural degree. ERP is the most convenient projection type for storage or transmission, but it is also the projection type that suffers the heaviest distortion. To better explain the causation of distortions, we follow the definition + +of stretching ratio $(\mathbf{K})$ in [31], which represents distortion degree at different locations from the target projection type to the ideal spherical surface. $\mathbf{K}$ is determined by area variation from one projection type to another. When the target type is uniforming spherical surface, $\mathbf{K}$ is defined as: + +$$ +\mathbf {K} (x, y) = \frac {\delta S (\theta , \varphi)}{\delta P (x , y)} = \frac {\cos (\varphi) | d \theta d \varphi |}{| d x d y |} = \frac {\cos (\varphi)}{| J (\theta , \varphi) |}, \tag {1} +$$ + +where $\delta S(\cdot ,\cdot)$ and $\delta P(\cdot ,\cdot)$ represent the area on the spherical surface and the projection plane, respectively. $|didj|$ represents plane microunit. $|J(\theta ,\varphi)|$ is the Jacobian determinant from spherical coordinate to projection coordinate. + +ERP distortion. The coordinate in ERP is defined as $x = \theta$ and $y = \varphi$ . ERP stretching ratio can be derived as: + +$$ +\mathbf {K} _ {\mathrm {E R P}} (x, y) = \cos (\varphi) = \cos (y), \tag {2} +$$ + +where $x\in (-\pi ,\pi)$ $y\in (-\frac{\pi}{2},\frac{\pi}{2})$ + +From Eq. (2), we conclude that ERP distortion is only determined by its latitude degree. $\mathbf{K}_{\mathrm{ERP}}$ is reduced to zero when the absolute value of latitude degree increases to $\pi /2$ , which represents that pixel density on the polar areas of ERP images is closer to zero. As shown in Fig. 2 (c), with the increasing of the absolutely value of latitude degree $(|\varphi_{s}|)$ , the corresponding area on the spherical surface of an ERP microunit is gradually decreased to zero. In conclusion, ERP distortion is caused by variable stretching ratios $\mathbf{K}_{\mathrm{ERP}}$ , and is the heaviest in the polar areas. + +Fisheye distortion. The coordinate in Fisheye can be derived from $\theta = \arctan \left(\frac{y}{x}\right)$ and $\varphi = (1 - \sqrt{x^2 + y^2})\times \frac{\pi}{2}$ . The stretching ratios of Fisheye can be derived as: + +$$ +\mathbf {K} _ {\text {F i s h e y e}} (x, y) = \frac {\frac {2}{\pi} \sin \left(\frac {\pi}{2} \sqrt {x ^ {2} + y ^ {2}}\right)}{\sqrt {x ^ {2} + y ^ {2}}}, \tag {3} +$$ + +where $\sqrt{x^2 + y^2} \in (0,1)$ . + +$\mathbf{K}_{\mathrm{Fisheye}}$ is determined by distance from the fisheye center. As $(\mathbf{K}_{\mathrm{Fisheye}})^{-1}$ is bounded, fisheye projection is closer to uniform distribution than ERP. Moreover, it introduces much slighter distortion at the polar. + +Relationship between ERP and Fisheye distortions. To simplify, here we only discuss a typical Fisheye with an aperture degree of $\pi$ and a horizontal slicing plane4. In this case, the ERP coordinates and Fisheye's polar coordinates correspond linearly. We can quantize the relationship by: + +$$ +\mathbf {K} _ {\mathrm {E R P} \mid \text {F i s h e y e}} (\theta , \varphi) = \frac {\mathbf {K} _ {\mathrm {E R P}} \left(x _ {E} , y _ {E}\right)}{\mathbf {K} _ {\text {F i s h e y e}} \left(x _ {F} , y _ {F}\right)} = \frac {\pi}{2} - | \varphi |, \tag {4} +$$ + +where $\theta, \varphi$ are spherical coordinates on the sphere, $x_{E}, y_{E}$ ( $x_{F}, y_{F}$ ) denotes the plain coordinate under ERP (Fisheye). + +![](images/7b63ac2ee9f66d49a42ce1663d2e22e5d7a14e6a3c7436c95582879c443377da.jpg) +Figure 3. Downsampling process of ODIs (left) and imaging process in real world (right). * denotes that LR images synthesized from different downsampling processes are inconsistent. + +From Eqs. (2) to (4), we conclude that when uniformed downsampling is performed on ERP, the kernel size of equivalent Fisheye downsampling is non-uniformed. Especially when fisheye projection is spliced horizontally, the kernel size is proportional with $\pi / 2 - |\varphi|$ . + +# 3.2. Learning with More Realistic Degradation + +As depicted in Fig. 3, the original-formatted projection type in ODI acquiring process is fisheye projection. Given that real-world low-resolution issues are caused by insufficient precision and density of sensors, we consider that the degeneration process should be directly applied to original-formatted images before the type conversion. + +Ideally, as camera sensors are arranged in uniform arrays, pixel density on original-formatted images is consistent everywhere. Thus, for a realistic ODI, the pixel density on Fisheye should be a constant. As discussed in Sec. 3.1, applying uniformed downsampling on ERP means applying downsampling of variable kernel size on Fisheye. The variable kernel size leads to variable Fisheye pixel density, which results in unrealistic LR images. In conclusion, the ERP downsampling in previous methods influences the intrinsic distribution of pixel density in original-formatted images, which leads to unrealistic ODIs. When the downsampling process happens on Fisheye, the Fisheye pixel density is unchanged, which fits the real-world imaging process and synthesizes more realistic LR pairs. + +Process of Fisheye downsampling. To generate more realistic LR ODIs, we mimic the real-world imaging process and apply bicubic downsampling on Fisheye images. One single Fisheye image can only store information about a hemisphere. Hence, ERP images are converted to dual Fisheye images. Before downsampling, Fisheye images are padded by a FOV larger than $180^{\circ}$ to avoid edge disconnections. This padding operation will not influence the geometric transforming relation between ERP and Fisheye. As Fisheye data is unstructured and Fisheye distortion is more complicated than ERP distortion, we still learn the upscaling process under ERP. Thus we reconvert LR images to the + +ERP format. The overall process of Fisheye downsampling are described in Fig. 3. + +# 3.3.OSRT:Modulate Distortion in ODIs + +Overall. As discussed in Sec. 3.1, ERP images suffer a distortion caused by a non-consistency area stretching ratio from an ideal spherical surface. Referred from Eq. (2), for an LR input $X_{i}\in \mathbb{R}^{C\times M\times N}$ , the distortion map $C_d\in \mathbb{R}^{1\times M\times N}$ is derived by: + +$$ +C _ {d} = \cos \left(\frac {m + 0 . 5 - M / 2}{M} \pi\right), \tag {5} +$$ + +where $m$ is the current height of LR input. + +Previous methods tend to treat $C_d$ as an additional input of $X_i$ [28], or re-weighting parameters by $C_d$ [18]. Although these solutions can benefit from building awareness of distortion, continuous and amorphous distortions cannot be adequately fitted by scattering and structured convolution operations. While previous methods cannot fully explore the advantage of $C_d$ , we intend to design a novel block for learning distorted patterns continuously. In VSR tasks, the deformable mechanism is proposed to align features between adjacent frames [32, 34]. Unlike standard DCN [10], which calculates offsets from the input feature map, offsets are calculated from bi-directional optical flow in VSR pipelines. Inspired by feature-level flow warping in VSR, we find that the deformable mechanism is a feasible solution for continuous mappings. Consequently, we modulate ERP distortion by feature-level warping operations. As shown in Fig. 4, $C_d$ is only utilized to calculate the deformable offsets $\theta$ . To keep compatibility with arbitrary ConvNets and Transformers, we propose two blocks to modulate ERP distortion, which can directly replace the multi-head self-attention blocks in Transformers and the standard convolution layers in ConvNets, respectively. + +Distortion-aware attention block (DAAB). As depicted in Fig. 4 (a), a distortion condition guided deformable self-attention is proposed to learn correlations between the distorted input $X_{i}$ and its corresponding modulated feature map $\tilde{X}_{i}$ . DAAB is formulated as: + +$$ +\theta_ {i} = H _ {\text {o f f s e t} _ {i}} \left(C _ {d}, C _ {w}\right), \tilde {X} _ {i} = \phi \left(X _ {i}; p _ {0} + \theta_ {i}\right), \tag {6} +$$ + +$$ +X _ {o} = H _ {\mathrm {S A}} \left(X _ {i} W _ {q _ {i}}, \tilde {X} _ {i} W _ {k _ {i}}, \tilde {X} _ {i} W _ {v _ {i}}\right), \tag {7} +$$ + +where $H_{\mathrm{offset}_i}(\cdot)$ denotes the $i$ -th convolution block to calculate offset maps $\theta_i \in \mathbb{R}^{2 \times H \times W}$ , and $H_{\mathrm{SA}}$ denotes standard self-attention formula. $H_{\mathrm{offset}}(\cdot)$ consists of $1 \times 1$ convolution block with two hidden layers. The input of $H_{\mathrm{offset}}(\cdot)$ is concatenated by the latitude-related distortion condition $C_d \in \mathbb{R}^{1 \times H \times W}$ and the window condition $C_w \in \mathbb{R}^{2 \times H \times W}$ . $C_w$ is a linear position encoding within a self-attention kernel. $\phi(\cdot, \cdot)$ denotes a bilinear interpolation, and $W_{q_i}, W_{k_i}, W_{v_i}$ denote $i$ -th weight matrix of query, key, and + +![](images/9a56a726695fd45479635425fdb85149a3f247dcad2883417c18b0ba871511bf.jpg) + +![](images/72c5be446d0de73ba147472880a1c4c2d735a796168cf0b389d1d4d73f1f7b1e.jpg) +Figure 4. Overall illustration of OSRT. From SwinIR [23], we replace the standard multi-head self-attention block with DAAB and insert DACB behind the end of the RSTB. Channel dimensions of $\theta_{\mathrm{offsets}}$ in DAAB and DACB are 2 and 18, respectively. + +![](images/a884f0fbde197c90940bbf89f5d5f09dd47afaa086cf5514ba23ee5e36df7a20.jpg) + +![](images/154b285d4c54c969b4bd95e5610f05814e7fd4b54cb468d4d98ec941b8f4e740.jpg) + +value, respectively. For multi-head self-attention blocks, $H_{\mathrm{offset}_i}(\cdot)$ is identical in calculations of parallel heads. + +Distortion-aware convolution block (DACB). As shown in Fig. 4 (b), we apply a standard deformable convolution layer with a substituted input for offset calculation. Modulated output $X_{o}$ is extracted as: + +$$ +\theta_ {i} = H _ {\text {o f f s e t} _ {i}} \left(C _ {d}\right), X _ {o} = H _ {\mathrm {D C N} _ {i}} \left(X _ {i}, \theta_ {i}\right), \tag {8} +$$ + +where $H_{\mathrm{DCN}}(X,\theta)$ denotes standard deformable convolution layer in [46]. The architecture of $H_{\mathrm{offset}_i}(\cdot)$ is identical to that in DAAB. As the kernel size of DCN is $3\times 3$ in DACB, the output channel dimension of offsets maps is 18. + +OSRT. In practice, we propose an Omnidirectional image Super-Resolution Transformer, named OSRT. SwinIR [23] is selected as the basic architecture for its strong reconstruction ability in the SISR task. To learn distortion rectified representations, we stack a DACB after the last convolution layer of each residual swin Transformer block and replace all self-attention blocks as DAAB. The feature dimension of OSRT is reduced from 180 to 156 to maintain identical parameters with SwinIR. + +# 3.4. Boosting ODISR Performance by Plain Images + +As the capacity of OSRT is relatively large, it suffers overfitting for large upscaling factors (Fig. 9). Given that acquiring ODIs are expensive, we propose to generate pseudo ERP images from 2D plain images to tackle this issue. After being sampled by sliding windows, the patch of plain images is treated as a plain perspective. By converting from Perspective to ERP, plain images are distorted in the same way as ERP. Considering that distortion of a Perspective is enlarged by its FOV degree, a relatively small FOV + +degree of $90^{\circ}$ is applied. For a given pseudo Perspective, $\theta_{p}$ is fixed at 0 and $\varphi_{p}$ is derived by: + +$$ +\Phi_ {p} = \varphi_ {h} + z _ {0}, \tag {9} +$$ + +where $\varphi_h$ is determined by patch locations and $z_0$ is orderly sampled from $\{-15^\circ, 0^\circ, 15^\circ\}$ . + +To maximize the approximate data distribution of ODIs, we horizontally split a plain image into three sub-images and define $\varphi_h$ as $-30^\circ, 0^\circ, 30^\circ$ respectively. Pseudo ERP images are cropped to remove the black border. As shown in Fig. 5, we get a new ERP dataset (called DF2K-ERP) by implementing the augmentation pipeline on widely-used plain image dataset DF2K [24, 33]. The DF2K-ERP dataset consists of 146K high-quality ERP image patches with a patch size larger than 256. + +![](images/c1511d336fd6d67d2883beed07de61b73a9cf216ef514f4c0641c079d11f14f5.jpg) +Figure 5. Synthetic process of DF2K-ERP. + +
MethodScaleODI-SRPSNRSUN 360 Panorama
PSNRSSIMWS-PSNRWS-SSIMSSIMWS-PSNRWS-SSIM
Bicubic28.210.821527.610.815628.140.811828.010.8321
RCAN [43]30.080.872329.490.871430.560.871231.180.8969
SRResNet [36]30.160.871729.590.869730.650.871431.200.8953
EDSR [24]×230.320.877029.680.872730.890.878431.420.8995
SwinIR [23]30.520.881929.870.877231.210.885231.780.9051
SwinIR†[23]30.640.882130.000.877731.330.885531.980.9059
OSRT†30.770.884630.110.879531.520.888832.140.9081
Bicubic25.590.711824.950.692325.290.699324.900.7083
RCAN [43]26.850.762126.150.748527.100.766026.990.7856
SRResNet [36]26.910.759726.240.745727.100.761826.990.7812
EDSR [24]×426.970.758926.300.745827.190.763327.100.7827
SwinIR [23]27.120.766326.440.752327.390.770727.300.7901
SwinIR†[23]27.310.773526.610.758927.710.780427.640.7996
OSRT†27.410.776226.700.760927.840.783527.770.8020
+ +Table 1. SR results under Fisheye downsampling. † denotes applying DF2K-ERP as augmented dataset. Best results are shown in Bold. + +![](images/c932d1fb890ac706ad04dc32f9289fcb0ce311fa8652d4bd663864d44ac46f38.jpg) +SUN360 $(\times 4)$ : 034 + +![](images/b7df64b1a03e7f8a36ff41abe586ecaa72ca27ccbbbf0899d86d9fe1d3913914.jpg) +HR + +![](images/f9475fdab31f0999131642a7793161d6eab43d3d887afc1e31b6d85db4da8f12.jpg) +PSNR/SSIM +EDSR [24] + +![](images/aab48792041743755ca62115a4c4922972a25ffec0420064182fff9ccf5cc5c4.jpg) +Bicubic + +![](images/7015d250894c2c694b918b664e9cbb5a0fdb672a2a087fce115626fc1c18f96e.jpg) +24.38dB/0.6872 +SwinJR [23] + +![](images/0acdf7d88bfcb6b4afc4eab6d1201a356f5846e79a01346926f7b16b25ac6c5c.jpg) +RCAN [43] + +![](images/481744445b0f343e23c06bcce4ed9331562c5e7d171db802b6c79cd1ca07a558.jpg) +26.40dB/0.8137 +SwinIR† [23] + +![](images/8477afb1d91308b13e0c72f6ce1f6ca8850546571d439b61c17ab62270f44d79.jpg) +SRResNet [36] + +![](images/4815ee82c3c4c7985323b1e9e0aabe47ead9697211ab2aa4f21f51d45cb9ff57.jpg) +26.21dB/0.7999 +OSRT† + +![](images/3bffb7b07707ac3d5c03dc456cbf1a86803f201e902c9c43ec23d3ff454e4205.jpg) +SUN360 $(\times 4)$ : 095 +Figure 6. Visual comparisons of $\times 4$ SR results under Fisheye downsampling. + +![](images/d4f1a7a11ae19a88ec5eefd9b6e5bb209f898b01ae13657c1b6232d8ab83d34b.jpg) +26.38dB/0.8072 +HR + +![](images/1a1d5efe555b1f5035ecc94306435532e175fa4e6656e29c696c66d460bff762.jpg) +PSNR/SSIM +EDSR [24] +33.64dB/0.9074 + +![](images/df19881ea35c55f6ef96db6fcb7f3ccacb62078b221a2cfdaab5f58e5a22304a.jpg) +26.77dB/0.8234 +Bicubic + +![](images/c54e91b8db836f67d26c9e9067e41436ab9c905563cd77742ff61876f914e714.jpg) +30.20dB/0.8506 +SwinIR [23] +34.05dB/0.9119 + +![](images/1e3682281a265e04e93b2dba4f4cec276c1167bb85a93b03f68b256eea66dd1d.jpg) +27.34dB/0.8462 +RCAN [43] + +![](images/bdfb984ea97a83e001e31d4a1f23bc3afb561016c77dcf1ec9ce44a13b1e5345.jpg) +33.59dB/0.9088 +SwinIR [23] +34.41dB/0.9158 + +![](images/1448538722ba337494966ef75c9fbf0d4bc040e329acc5140d8213ccd5bff992.jpg) +27.68dB/0.8561 +SRResNet [36] + +![](images/b350bdce5d2de18e8b6e05b88ebcdc8e43d8ae5cadca5f2dfd3f96e718b7dd4c.jpg) +33.43dB/0.9043 +OSRT† +34.77dB/0.9187 + +# 4. Experiments + +# 4.1. Experimental Setup + +ODI-SR dataset [12] and SUN360 Panorama dataset [39] are used in our experiment. In the training phase, we follow the data split setting in [12] and train on the ODI-SR training set. The resolution of the ERP HR is $1024 \times 2048$ , and the upscaling factors are $\times 2$ and $\times 4$ . Fisheye downsampling is applied as our pre-defined downsampling kernel. Loss is calculated by L1 distance and optimized by Adam [19], with an initial learning rate of $2 \times 10^{-4}$ , a total batch size of 32, and an input patch size of 64. We + +train OSRT for $500k$ iterations and halve the learning rate at $250k$ , $400k$ , $450k$ and $475k$ . In evaluation, we test on the ODI-SR testing set and SUN360 dataset. PSNR [16], SSIM [37], and their distortion re-weighted versions (WS-PSNR [31], WS-SSIM [45]) are used as evaluation metrics. + +# 4.2. Evaluation under Fisheye Downsampling + +When the downsampling process is performed on Fisheye images, we train SRResNet [36], EDSR [24], RCAN [43], and SwinIR [23] for comparison. + +Quantitative results. As shown in Tab. 1, with the help of additional DF2K-ERP training patches, OSRT out + +
Scale×8×16
MethodODI-SRSUN 360 PanoramaODI-SRSUN 360 Panorama
WS-PSNRWS-SSIMWS-PSNRWS-SSIMWS-PSNRWS-SSIMWS-PSNRWS-SSIM
Bicubic19.640.590819.720.540317.120.433217.560.4638
SRCNN [13]20.080.611219.460.570118.080.450117.950.4684
EDSR [24]23.970.641722.460.634121.120.569821.060.5645
RCAN [43]24.260.662823.880.654221.940.582421.740.5742
360-SS [30]21.650.641721.480.635219.650.543119.620.5308
LAU-Net [12]24.360.680124.020.670822.070.590121.820.5824
SphereSR [40]24.370.677724.170.682022.510.637021.950.6342
OSRT24.530.678024.380.707222.690.626122.130.6388
+ +performs previous methods by 0.3dB on PSNR. Although directly applying SwinIR on the ODISR task has already reached SOTA performance, OSRT surpasses SwinIR over 0.1dB on two datasets for both $\times 2$ and $\times 4$ SR tasks, which demonstrates the effectiveness of its distortion modulation ability. The performance of RCAN degrades under Fisheye downsampling, which is caused by the incompatibility between channel attention and Fisheye downsampling5. + +Qualitative comparison. Fig. 6 shows the visualization results of $\times 4$ ODISR task. While other methods struggle to understand the geometric transformation process in distorted images, OSRT can reconstruct sharp and accurate boundaries with the advantages of distortion modulation. It is observed that OSRT is skilled at reconstructing rigid texture. Moreover, benefiting from the distortion modulation ability, OSRT can preserve the original structure as most when being projected to other projection types (Fig. 7). + +# 4.3. Evaluation under ERP Downsampling + +To compare with previous ODISR methods [12, 30, 40], we train OSRT under the previous ERP setting. Regardless of over-fitting issues, we only train on the dataset provided by [12] for fairness. As shown in Tab. 2, OSRT still outperforms LAU-Net [12] and SphereSR [40] under large upscaling factor and ERP downsampling. Without a complicated training pipeline and discrete inference process, OSRT yields the best PSNR values and surpasses all previous methods on most SSIM-related metrics (three of four). + +# 4.4. Ablation Study and Discussion + +In this section, we prove the effectiveness of Fisheye downsampling, OSRT components, and augmented DF2K-ERP. We then explain the distortion modulation ability of OSRT by visualizing offsets in deformable blocks. + +Fisheye downsampling. As shown in Fig. 1, the SR model trained under ERP downsampling is more likely to generate blur details and missing structures in real-world scenarios. These artifacts cannot be removed by a superior backbone network, but can be eliminated by a more realistic imaging process. More importantly, ERP downsampling directly covers the geometric property of ERP images + +Table 2. SR results under ERP downsampling. + +
feature dimDACBDAABODI-SRSUN360Params. (M)
PSNRSSIMPSNRSSIM
60××30.270.873930.780.87420.91
60✓×30.410.877531.000.87931.16
60×w/o Cw30.310.874630.830.87551.00
60×w/ Cw30.320.874630.840.87531.01
60✓w/ Cw30.440.878031.040.88001.26
72××30.320.874830.850.87551.29
+ +Table 3. Ablation study on OSRT components. All models are trained on $\times 2$ SR task under Fisheye downsampling. + +and makes the ODISR task identical to the standard plain image super-resolution task. The evidence is that a standard SISR model (SwinIR) trained on a plain image dataset (DF2K) can outperform previous SOTA in the ODISR task, which yields WS-PSNR results of 24.63dB/24.49dB (22.68dB/22.13dB) on $\times 8$ ( $\times 16$ ) ODI-SR/SUN360 testing set, respectively. In conclusion, when the intrinsic property of ODIs is broken by ERP downsampling, the ODISR task degenerates into a plain image super-resolution task with a particular data distribution. + +![](images/d263c9dc37bd87f334f11c9005cdf5c152dd27cca0279982cf36f6cb5dda0676.jpg) +ODI-SR $(\times 4)$ :003 Fisheye (Vertical, Right) + +![](images/347399a281d81a2d0d22053c52bed700fd8cfce33057400626ea2f8cc673852b.jpg) +EDSR [24] + +![](images/8aa834e957f53eb38f677568c4bf692307999de4ef356c7c58422c3c26310c02.jpg) + +![](images/9963fa8d1acfcc5a6d149a19bfb179db69dd6d77b260bc2ca6ba3bdc5faf8c48.jpg) +SwinIR† [23] + +![](images/ec3d96c5c427240db388b5eacae55786ddb7b31ba10c483e2efd6dd8118ce7c0.jpg) +SwinIR [23] +OSRT† + +![](images/044705c461a57792277caae6ab00859e0349e3d4c68851aa8489c1f61b722fd8.jpg) +SUN360 $(\times 2)$ : 062 Perspective $(\varphi :30^{\circ}$ ;FOV: $90^{\circ}$ + +![](images/ab76630d7f4eb043277a7f12d79ac163f08a3ef4c81ee0291e596541b8c53988.jpg) + +![](images/7765713b1be23316de722e73c0ce74cede6ff58d5d4247c9cc4212714f3e92b9.jpg) +EDSR [24] +SwinIR† [23] + +![](images/5caee10a8ce95acf83a71b25eeb960c6896770a0a72d1c8dacffb3c246e1d105.jpg) + +![](images/159410d8b784f09b44c67d733c2fc6325a05f36296e01d3b0f733d950853a79b.jpg) +SwinIR [23] +OSRT† +Figure 7. Visual comparisons for SR of Fisheye and Perspective images. $\dagger$ denotes applying DF2K-ERP as augmented dataset. + +![](images/c3cf00c8f3d468d3e2c7a3efbf4bd7a8c653e5e8e31469311d32541521f07351.jpg) +DAAB (29-th) + +![](images/aaf70f510f46675cf63b9135c08d88c8212849ebe829cdaca705a4dadd8456ea.jpg) +$\varphi = 45^{\circ}$ + +![](images/17e28f2558a2bba12c2cedbe02d22d7006226271a10f0413a0f789e57c345074.jpg) +$\varphi = 90^{\circ}$ + +![](images/45e10ebd222dcf2a24dc360e989bc54ed34567825a2779a634b800c19babe609.jpg) +$\omega = 0^{\circ}$ +DAAB (33-th) + +![](images/1d82e6dc3dc63cc5f2405848b899de9664c0e7d2b89e8b2cdc5fa89a40e3ec2e.jpg) + +![](images/8b703d5dee9b0a31f011e5c6c872d5cbce0c5fba1282cf61e05e732ca49178a0.jpg) + +![](images/17ce58425cfcda28af96cdc102bcb82b07e11f6257b1859af081fb33bdad73f4.jpg) +$\varphi = 0^{\circ}$ +DACB +$\varphi = 0^{\circ}$ +Figure 8. Visualizations of offset maps in OSRT. Reference and deformed points are depicted in green and red, respectively. The deformable kernel is sparse in the polar area. + +![](images/b197eba21c2d9107a864f5589706d7516219b43891386c3ca6eb8d3e277022d3.jpg) +$\varphi = 45^{\circ}$ +$\varphi = 30^{\circ}$ + +![](images/4db4315bd083d2575e395724e124f4e6134db2cb434b7d98c625c52210a3e8e1.jpg) +$\varphi = 60^{\circ}$ + +![](images/7ebc60cf1b6d62fb9d134d64abd88daba08c92ed823a05a7d574caf9b86f4609.jpg) +$\varphi = 90^{\circ}$ +$\varphi = 90^{\circ}$ + +OSRT components. To study the effectiveness of each component in OSRT, we propose a light version of OSRT (OSRT-light) for ablation study, which corresponds with the official SwinIR-light [23]. As proofed in Tab. 3, all components in OSRT are beneficial for modulating ERP distortion. The advantages of DACB and DAAB can be stacked when being applied in the same network. Compared with simply expanding the feature dimension of SwinIR to match the network complexity, the overall improvements of OSRT is more significant $(+0.05\mathrm{dB}$ vs. $+0.2\mathrm{dB})$ + +Offsets in OSRT. Offsets map in a well-trained OSRT are visualized in Fig. 8. Deformable kernels in both DAAB and DACB tend to gather at the equator and scatter at the polar, which conforms to the geometric distribution of pixel density in ERP images. + +![](images/86d33b68e48f5512cb07c309a6291a07500b6c66226197bbb47b164d05215df9.jpg) +Figure 9. Training process of Transformers on $\times 4$ ODISR task. The overfitting issue is tackled by our augmentation scheme. + +
Backbone networkDatasetsTraining schemeScaleSUN360
PSNRSSIM
SwinIRODI-SRN/A31.210.8852
SwinIRDF2K/ODI-SRone-stage31.260.8841
SwinIRDF2K-ERP/ODI-SRone-stage×231.330.8855
SwinIRDF2K-ERP/ODI-SRtwo-stage31.170.8818
OSRTDF2K-ERP/ODI-SRone-stage31.520.8888
SwinIRODI-SRN/A27.390.7707
SwinIRDF2K/ODI-SRone-stage27.590.7768
SwinIRDF2K-ERP/ODI-SRone-stage×427.710.7804
SwinIRDF2K-ERP/ODI-SRtwo-stage27.740.7795
OSRTDF2K-ERP/ODI-SRone-stage27.840.7835
+ +Table 4. Ablation study on data augmentation. The results of ODISR (In the supplementary file) are in the same trend as SUN360. + +Pseudo ERP patches. In Sec. 3.4, we propose a distorted dataset DF2K-ERP to tackle over-fitting issues. We train a standard SwinIR on diverse datasets and training schemes to study the influence of data augmentation separately. As shown in Tab. 4, while training on ODI-SR and DF2K, distortion operations in DF2K lead to better performance. Compared with fine-tuning on DF2K-ERP pretrained models (two-stage), training on two datasets jointly (one-stage) shows better results. We infer that there is a domain gap between ODI-SR and DF2K-ERP, which is caused by omitted Perspective distortion. Moreover, the advantage of distortion modulation mechanisms in OSRT is enlarged when additional training patches are applied. Fig. 9 proves that our data augmentation scheme overcomes the over-fitting issue and improves the reconstruction ability. + +# 5. Conclusion + +In this paper, we find that the previous downsampling process in the ODISR task harms the intrinsic distribution of pixel density in ODIs, which leads to poor generalization ability in real-world scenarios. To tackle this issue, we propose Fisheye downsampling, which mimics the real-world imaging process to preserve the realistic density distribution. After refining the downsampling process, we design a distortion-aware Transformer (OSRT) to modulate distortions continuously and self-adaptively. OSRT learns offsets from the distortion-related condition and rectifies distortion by feature-level warping. Moreover, to alleviate the overfitting problem of large networks, we propose to synthesize additional ERP training data from the plain images. Extensive experiments have demonstrated the state-of-the-art performance of our OSRT. + +Limitation. The process of sampling ERP images into viewing types also requires careful design. + +# Acknowledgement. + +This work was supported in part by the National Key R&D Program of China (NO.2022ZD0160505), in part by the National Natural Science Foundation of China under Grant (62276251), the Joint Lab of CAS-HK, and in part by the Youth Innovation Promotion Association of Chinese Academy of Sciences (No.2020356). + +# References + +[1] Hao Ai, Zidong Cao, Jinjing Zhu, Haotian Bai, Yucheng Chen, and Ling Wang. Deep learning for omnidirectional vision: A survey and new perspectives. arXiv preprint arXiv:2205.10468, 2022. 1 +[2] Zafer Arican and Pascal Frossard. L1 regularized superresolution from unregistered omnidirectional images. In 2009 IEEE International Conference on Acoustics, Speech and Signal Processing, pages 829-832. IEEE, 2009. 2 +[3] Zafer Arican and Pascal Frossard. Joint registration and super-resolution with omnidirectional images. IEEE Transactions on Image Processing, 20(11):3151-3162, 2011. 2 +[4] Luigi Bagnato, Yannick Boursier, Pascal Frossard, and Pierre Vandergheynst. Plenoptic based super-resolution for omnidirectional image sequences. In 2010 IEEE International Conference on Image Processing, pages 2829-2832. IEEE, 2010. 2 +[5] Kelvin CK Chan, Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Basicvsr: The search for essential components in video super-resolution and beyond. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4947-4956, 2021. 3 +[6] Kelvin CK Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. Basicvsr++: Improving video superresolution with enhanced propagation and alignment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5972-5981, 2022. 3 +[7] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12299-12310, 2021. 2 +[8] Xiangyu Chen, Xintao Wang, Jiantao Zhou, and Chao Dong. Activating more pixels in image super-resolution transformer. arXiv preprint arXiv:2205.04437, 2022. 2 +[9] Yinbo Chen, Sifei Liu, and Xiaolong Wang. Learning continuous image representation with local implicit image function. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8628-8638, 2021. 2 +[10] Jifeng Dai, Haozhi Qi, Yuwen Xiong, Yi Li, Guodong Zhang, Han Hu, and Yichen Wei. Deformable convolutional networks. In Proceedings of the IEEE international conference on computer vision, pages 764-773, 2017. 4 +[11] Tao Dai, Jianrui Cai, Yongbing Zhang, Shu-Tao Xia, and Lei Zhang. Second-order attention network for single image super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11065-11074, 2019. 2 +[12] Xin Deng, Hao Wang, Mai Xu, Yichen Guo, Yuhang Song, and Li Yang. Lau-net: Latitude adaptive upscaling network for omnidirectional image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9189-9198, 2021. 1, 2, 6, 7 +[13] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Image super-resolution using deep convolutional net + +works. IEEE transactions on pattern analysis and machine intelligence, 38(2):295-307, 2015. 2, 7 +[14] Chao Dong, Chen Change Loy, and Xiaou Tang. Accelerating the super-resolution convolutional neural network. In European conference on computer vision, pages 391-407. Springer, 2016. 2 +[15] Vida Fakour-Sevom, Esin Guldogan, and Joni-Kristian Kämäräinen. 360 panorama super-resolution using deep convolutional networks. In Int. Conf. on Computer Vision Theory and Applications (VISAPP), volume 1, 2018. 1, 2 +[16] Quan Huynh-Thu and Mohammed Ghanbari. Scope of validity of psnr in image/video quality assessment. *Electronics letters*, 44(13):800-801, 2008. 6 +[17] Hiroshi Kawasaki, Katsushi Ikeuchi, and Masao Sakauchi. Super-resolution omnidirectional camera images using spatio-temporal analysis. *Electronics and Communications in Japan* (Part III: Fundamental Electronic Science), 89(6):47-59, 2006. 2 +[18] Renata Khasanova and Pascal Frossard. Geometry aware convolutional filters for omnidirectional images representation. In International Conference on Machine Learning, pages 3351-3359. PMLR, 2019. 4 +[19] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6 +[20] Christian Ledig, Lucas Theis, Ferenc Huszar, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photorealistic single image super-resolution using a generative adversarial network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4681-4690, 2017. 2 +[21] Wenbo Li, Xin Lu, Jiangbo Lu, Xiangyu Zhang, and Jiaya Jia. On efficient transformer and image pre-training for low-level vision. arXiv preprint arXiv:2112.10175, 2021. 2 +[22] Zheyuan Li, Yingqi Liu, Xiangyu Chen, Haoming Cai, Jinjin Gu, Yu Qiao, and Chao Dong. Blueprint separable residual network for efficient image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 833-843, 2022. 2 +[23] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1833-1844, 2021. 2, 5, 6, 7, 8 +[24] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 136-144, 2017. 1, 2, 5, 6, 7 +[25] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 2 +[26] Yiqun Mei, Yuchen Fan, and Yuqian Zhou. Image superresolution with non-local sparse attention. In Proceedings of + +the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3517-3526, 2021. 2 +[27] Hajime Nagahara, Yasushi Yagi, and Masahiko Yachida. Super-resolution from an omnidirectional image sequence. In 2000 26th Annual Conference of the IEEE Industrial Electronics Society. IECON 2000. 2000 IEEE International Conference on Industrial Electronics, Control and Instrumentation. 21st Century Technologies, volume 4, pages 2559-2564. IEEE, 2000. 2 +[28] Akito Nishiyama, Satoshi Ikehata, and Kiyoharu Aizawa. 360 single image super resolution via distortion-aware network and distorted perspective images. In 2021 IEEE International Conference on Image Processing (ICIP), pages 1829-1833. IEEE, 2021. 1, 2, 4 +[29] Ben Niu, Weilei Wen, Wenqi Ren, Xiangde Zhang, Lianping Yang, Shuzhen Wang, Kaihao Zhang, Xiaochun Cao, and Haifeng Shen. Single image super-resolution via a holistic attention network. In European conference on computer vision, pages 191-207. Springer, 2020. 2 +[30] Cagri Ozcinar, Aakanksha Rana, and Aljosa Smolic. Superresolution of omnidirectional images using adversarial learning. In 2019 IEEE 21st International Workshop on Multimedia Signal Processing (MMSP), pages 1-6. IEEE, 2019. 2, 3, 7 +[31] Yule Sun, Ang Lu, and Lu Yu. Weighted-to-spherically-uniform quality evaluation for omnidirectional video. IEEE signal processing letters, 24(9):1408-1412, 2017. 3, 6 +[32] Yapeng Tian, Yulun Zhang, Yun Fu, and Chenliang Xu. Tdan: Temporally-deformable alignment network for video super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3360-3369, 2020. 4 +[33] Radu Timofte, Eirikur Agustsson, Luc Van Gool, Ming-Hsuan Yang, and Lei Zhang. Ntire 2017 challenge on single image super-resolution: Methods and results. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 114-125, 2017. 5 +[34] Xintao Wang, Kelvin CK Chan, Ke Yu, Chao Dong, and Chen Change Loy. Edvr: Video restoration with enhanced deformable convolutional networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 0–0, 2019. 3, 4 +[35] Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1905-1914, 2021. 2 +[36] Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Yu Qiao, and Chen Change Loy. Esrgan: Enhanced super-resolution generative adversarial networks. In Proceedings of the European conference on computer vision (ECCV) workshops, pages 0–0, 2018. 2, 6 +[37] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6 +[38] Zhuofan Xia, Xuran Pan, Shiji Song, Li Erran Li, and Gao Huang. Vision transformer with deformable attention. In + +Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4794-4803, 2022. 3 +[39] Jianxiong Xiao, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Recognizing scene viewpoint using panoramic place representation. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, pages 2695-2702. IEEE, 2012. 1, 6 +[40] Youngho Yoon, Inchul Chung, Lin Wang, and Kuk-Jin Yoon. Spheresr: 360deg image super-resolution with arbitrary projection via continuous spherical image representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5677-5686, 2022. 1, 2, 7 +[41] Kai Zhang, Jingyun Liang, Luc Van Gool, and Radu Timofte. Designing a practical degradation model for deep blind image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4791-4800, 2021. 2 +[42] Wenlong Zhang, Yihao Liu, Chao Dong, and Yu Qiao. Ranksrgan: Generative adversarial networks with ranker for image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3096-3105, 2019. 2 +[43] Yulun Zhang, Kunpeng Li, Kai Li, Lichen Wang, Bineng Zhong, and Yun Fu. Image super-resolution using very deep residual channel attention networks. In Proceedings of the European conference on computer vision (ECCV), pages 286-301, 2018. 1, 2, 6, 7 +[44] Yupeng Zhang, Hengzhi Zhang, Daojing Li, Liyan Liu, Hong Yi, Wei Wang, Hiroshi Suitoh, and Makoto Odamaki. Toward real-world panoramic image enhancement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 628-629, 2020. 2 +[45] Yufeng Zhou, Mei Yu, Hualin Ma, Hua Shao, and Gangyi Jiang. Weighted-to-spherically-uniform sSIM objective quality evaluation for panoramic video. In 2018 14th IEEE International Conference on Signal Processing (ICSP), pages 54-57. IEEE, 2018. 6 +[46] Xizhou Zhu, Han Hu, Stephen Lin, and Jifeng Dai. Deformable convnets v2: More deformable, better results. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9308-9316, 2019. 5 \ No newline at end of file diff --git a/2023/OSRT_ Omnidirectional Image Super-Resolution With Distortion-Aware Transformer/images.zip b/2023/OSRT_ Omnidirectional Image Super-Resolution With Distortion-Aware Transformer/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..4b125aae0060673079222dfd940a9fbaee3d9689 --- /dev/null +++ b/2023/OSRT_ Omnidirectional Image Super-Resolution With Distortion-Aware Transformer/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a13b938960205634bfad513ea41273c4d3c940096768220b7f71adf13d82d624 +size 812744 diff --git a/2023/OSRT_ Omnidirectional Image Super-Resolution With Distortion-Aware Transformer/layout.json b/2023/OSRT_ Omnidirectional Image Super-Resolution With Distortion-Aware Transformer/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..73709780d7819b81305908be68ea5a87795599e4 --- /dev/null +++ b/2023/OSRT_ Omnidirectional Image Super-Resolution With Distortion-Aware Transformer/layout.json @@ -0,0 +1,12248 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 121, + 117, + 454, + 149 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 117, + 454, + 149 + ], + "spans": [ + { + "bbox": [ + 121, + 117, + 454, + 149 + ], + "type": "text", + "content": "OSRT: Omnidirectional Image Super-Resolution with Distortion-aware Transformer" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 151, + 518, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 151, + 518, + 207 + ], + "spans": [ + { + "bbox": [ + 61, + 151, + 518, + 207 + ], + "type": "text", + "content": "Fanghua Yu" + }, + { + "bbox": [ + 61, + 151, + 518, + 207 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 61, + 151, + 518, + 207 + ], + "type": "text", + "content": " Xintao Wang" + }, + { + "bbox": [ + 61, + 151, + 518, + 207 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 61, + 151, + 518, + 207 + ], + "type": "text", + "content": " Mingdeng Cao" + }, + { + "bbox": [ + 61, + 151, + 518, + 207 + ], + "type": "inline_equation", + "content": "^{2,3}" + }, + { + "bbox": [ + 61, + 151, + 518, + 207 + ], + "type": "text", + "content": " Gen Li" + }, + { + "bbox": [ + 61, + 151, + 518, + 207 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 61, + 151, + 518, + 207 + ], + "type": "text", + "content": " Ying Shan" + }, + { + "bbox": [ + 61, + 151, + 518, + 207 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 151, + 518, + 207 + ], + "type": "text", + "content": " Chao Dong" + }, + { + "bbox": [ + 61, + 151, + 518, + 207 + ], + "type": "inline_equation", + "content": "^{1,5\\dagger}" + }, + { + "bbox": [ + 61, + 151, + 518, + 207 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 151, + 518, + 207 + ], + "type": "text", + "content": "ShenZhen Key Lab of Computer Vision and Pattern Recognition \nShenzhen Institute of Advanced Technology, Chinese Academy of Sciences \n" + }, + { + "bbox": [ + 61, + 151, + 518, + 207 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 151, + 518, + 207 + ], + "type": "text", + "content": "ARC, Tencent PCG " + }, + { + "bbox": [ + 61, + 151, + 518, + 207 + ], + "type": "inline_equation", + "content": "{}^{3}" + }, + { + "bbox": [ + 61, + 151, + 518, + 207 + ], + "type": "text", + "content": "The University of Tokyo" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 208, + 514, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 208, + 514, + 223 + ], + "spans": [ + { + "bbox": [ + 62, + 208, + 514, + 223 + ], + "type": "text", + "content": "4 Platform Technologies, Tencent Online Video Shanghai Artificial Intelligence Laboratory" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 98, + 224, + 477, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 98, + 224, + 477, + 235 + ], + "spans": [ + { + "bbox": [ + 98, + 224, + 477, + 235 + ], + "type": "text", + "content": "fanghuayu96@gmail.com, xintaowang@tencent.com, cmd@g.ecc.u-tokyo.ac.jp" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 147, + 238, + 428, + 250 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 147, + 238, + 428, + 250 + ], + "spans": [ + { + "bbox": [ + 147, + 238, + 428, + 250 + ], + "type": "text", + "content": "{genli, yingsshan}@tencent.com, chao.dong@siat.ac.cn" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 135, + 269, + 183, + 281 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 269, + 183, + 281 + ], + "spans": [ + { + "bbox": [ + 135, + 269, + 183, + 281 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 38, + 288, + 282, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 288, + 282, + 541 + ], + "spans": [ + { + "bbox": [ + 38, + 288, + 282, + 541 + ], + "type": "text", + "content": "Omnidirectional images (ODIs) have obtained lots of research interest for immersive experiences. Although ODIs require extremely high resolution to capture details of the entire scene, the resolutions of most ODIs are insufficient. Previous methods attempt to solve this issue by image super-resolution (SR) on equirectangular projection (ERP) images. However, they omit geometric properties of ERP in the degradation process, and their models can hardly generalize to real ERP images. In this paper, we propose Fisheye downsampling, which mimics the real-world imaging process and synthesizes more realistic low-resolution samples. Then we design a distortion-aware Transformer (OSRT) to modulate ERP distortions continuously and self-adaptively. Without a cumbersome process, OSRT outperforms previous methods by about 0.2dB on PSNR. Moreover, we propose a convenient data augmentation strategy, which synthesizes pseudo ERP images from plain images. This simple strategy can alleviate the over-fitting problem of large networks and significantly boost the performance of ODISR. Extensive experiments have demonstrated the state-of-the-art performance of our OSRT." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 39, + 557, + 119, + 570 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 39, + 557, + 119, + 570 + ], + "spans": [ + { + "bbox": [ + 39, + 557, + 119, + 570 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 38, + 578, + 279, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 578, + 279, + 685 + ], + "spans": [ + { + "bbox": [ + 38, + 578, + 279, + 685 + ], + "type": "text", + "content": "In pursuit of the realistic visual experience, omnidirectional images (ODIs), also known as " + }, + { + "bbox": [ + 38, + 578, + 279, + 685 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 38, + 578, + 279, + 685 + ], + "type": "text", + "content": " images or panoramic images, have obtained lots of research interest in the computer vision community. In reality, we usually view ODIs with a narrow field-of-view (FOV), e.g., viewing in a headset. To capture details of the entire scene, ODIs require extremely high resolution, e.g., " + }, + { + "bbox": [ + 38, + 578, + 279, + 685 + ], + "type": "inline_equation", + "content": "4\\mathrm{K} \\times 8\\mathrm{K}" + }, + { + "bbox": [ + 38, + 578, + 279, + 685 + ], + "type": "text", + "content": " [1]. However, due to the high industrial cost of camera sensors with high precision, the resolutions of most ODIs are insufficient." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 38, + 686, + 279, + 710 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 686, + 279, + 710 + ], + "spans": [ + { + "bbox": [ + 38, + 686, + 279, + 710 + ], + "type": "text", + "content": "Recently, some attempts have been made to solve this problem by image super-resolution (SR) [12, 15, 28, 39, 40]." + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 304, + 262, + 415, + 346 + ], + "blocks": [ + { + "bbox": [ + 304, + 262, + 415, + 346 + ], + "lines": [ + { + "bbox": [ + 304, + 262, + 415, + 346 + ], + "spans": [ + { + "bbox": [ + 304, + 262, + 415, + 346 + ], + "type": "image", + "image_path": "5aadb5dd34e8dc3d24b12e025b3c1637a487761ffde252181d26af69da3856a5.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 340, + 347, + 379, + 356 + ], + "lines": [ + { + "bbox": [ + 340, + 347, + 379, + 356 + ], + "spans": [ + { + "bbox": [ + 340, + 347, + 379, + 356 + ], + "type": "text", + "content": "Unseen LR" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 424, + 262, + 534, + 346 + ], + "blocks": [ + { + "bbox": [ + 424, + 262, + 534, + 346 + ], + "lines": [ + { + "bbox": [ + 424, + 262, + 534, + 346 + ], + "spans": [ + { + "bbox": [ + 424, + 262, + 534, + 346 + ], + "type": "image", + "image_path": "b9145473a0eb5e8a56509988face13f4e65bd2a193ad7d1b3a0a1f8980421b57.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 433, + 347, + 522, + 357 + ], + "lines": [ + { + "bbox": [ + 433, + 347, + 522, + 357 + ], + "spans": [ + { + "bbox": [ + 433, + 347, + 522, + 357 + ], + "type": "text", + "content": "LAU-Net [12] w/o Fisheye" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 304, + 360, + 415, + 444 + ], + "blocks": [ + { + "bbox": [ + 304, + 360, + 415, + 444 + ], + "lines": [ + { + "bbox": [ + 304, + 360, + 415, + 444 + ], + "spans": [ + { + "bbox": [ + 304, + 360, + 415, + 444 + ], + "type": "image", + "image_path": "ae3130b9f69db040cc2b88220b207a6a8af801674510da05fa5bff35b377ed7e.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 327, + 446, + 390, + 455 + ], + "lines": [ + { + "bbox": [ + 327, + 446, + 390, + 455 + ], + "spans": [ + { + "bbox": [ + 327, + 446, + 390, + 455 + ], + "type": "text", + "content": "OSRT w/o Fisheye" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 423, + 359, + 534, + 444 + ], + "blocks": [ + { + "bbox": [ + 423, + 359, + 534, + 444 + ], + "lines": [ + { + "bbox": [ + 423, + 359, + 534, + 444 + ], + "spans": [ + { + "bbox": [ + 423, + 359, + 534, + 444 + ], + "type": "image", + "image_path": "b255ba70f6bcacee07d49fa892047a8c4a2f0b7753dbac13d64c8af1dac46e81.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 449, + 446, + 508, + 455 + ], + "lines": [ + { + "bbox": [ + 449, + 446, + 508, + 455 + ], + "spans": [ + { + "bbox": [ + 449, + 446, + 508, + 455 + ], + "type": "text", + "content": "OSRT w/ Fisheye" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 296, + 458, + 537, + 490 + ], + "lines": [ + { + "bbox": [ + 296, + 458, + 537, + 490 + ], + "spans": [ + { + "bbox": [ + 296, + 458, + 537, + 490 + ], + "type": "text", + "content": "Figure 1. Visual comparisons of " + }, + { + "bbox": [ + 296, + 458, + 537, + 490 + ], + "type": "inline_equation", + "content": "\\times 8" + }, + { + "bbox": [ + 296, + 458, + 537, + 490 + ], + "type": "text", + "content": " SR results on LR images1 with unknown degradations. Fisheye denotes that the downsampling process in training stages is under Fisheye images." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 296, + 490, + 538, + 719 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 490, + 538, + 719 + ], + "spans": [ + { + "bbox": [ + 296, + 490, + 538, + 719 + ], + "type": "text", + "content": "As most of the ODIs are stored and transmitted in the equirectangular projection (ERP) type, the SR process is usually performed on the ERP images. To generate high- /low-resolution training pairs, existing ODISR methods [12, 15, 28, 39, 40] directly apply uniform bicubic down-sampling on the original ERP images (called ERP down-sampling), which is identical to general image SR settings [24, 43]. While omitting geometric properties of ERP in the degradation process, their models can hardly generalize to real ERP images. We can observe missing structures and blur textures in Fig. 1. Therefore, we need a more appropriate degradation model before studying SR algorithms. In practice, ODIs are acquired by the fisheye lens and stored in ERP. Given that the low-resolution issue in real-world scenarios is caused by insufficient sensor precision and density, the downsampling process should be applied to original-formatted images before converting into other storage types. Thus, to be conformed with real-world imaging processes, we propose to apply uniform bicubic" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 71, + 31, + 113, + 62 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 31, + 113, + 62 + ], + "spans": [ + { + "bbox": [ + 71, + 31, + 113, + 62 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 30, + 476, + 64 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 30, + 476, + 64 + ], + "spans": [ + { + "bbox": [ + 146, + 30, + 476, + 64 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 719, + 116, + 727 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 719, + 116, + 727 + ], + "spans": [ + { + "bbox": [ + 49, + 719, + 116, + 727 + ], + "type": "text", + "content": "*Equal contribution" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 49, + 728, + 227, + 738 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 728, + 227, + 738 + ], + "spans": [ + { + "bbox": [ + 49, + 728, + 227, + 738 + ], + "type": "text", + "content": "† Corresponding author (e-mail: chao.dong@siat.ac.cn)" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 727, + 473, + 738 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 727, + 473, + 738 + ], + "spans": [ + { + "bbox": [ + 308, + 727, + 473, + 738 + ], + "type": "text", + "content": "Photoed by Peter Leth on Flickr, with CC license." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 277, + 751, + 300, + 761 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 751, + 300, + 761 + ], + "spans": [ + { + "bbox": [ + 277, + 751, + 300, + 761 + ], + "type": "text", + "content": "13283" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 38, + 96, + 279, + 170 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 96, + 279, + 170 + ], + "spans": [ + { + "bbox": [ + 38, + 96, + 279, + 170 + ], + "type": "text", + "content": "downsampling on Fisheye images, which are the original format of ODIs. The new downsampling process (called Fisheye downsampling) applies uniform bicubic downsampling on Fisheye images before converting them to ERP images. Our Fisheye downsampling is more conducive to exploring the geometric property of ODIs." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 38, + 171, + 280, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 171, + 280, + 482 + ], + "spans": [ + { + "bbox": [ + 38, + 171, + 280, + 482 + ], + "type": "text", + "content": "The key issue of ODISR algorithm design is to utilize the geometric properties of ERP images, which is also the focus of previous methods. For example, Nishiyama et al. [28] add a distortion-related condition as an additional input. LAU-Net [12] splits the whole ERP image into patches by latitude band and learns upscaling processes separately. However, the separated learning process will lead to information disconnection between adjacent patches. SphereSR [40] learns different upscaling functions on various projection types, but will inevitably introduce multiple-time computation costs. To push the performance upper bound, we propose the first Transformer for Omnidirectional image Super-Resolution (OSRT), and incorporate geometric properties in a distortion-aware manner. Specifically, to modulate distorted feature maps, we implement feature-level warping, in which offsets are learned from latitude conditions. In OSRT, we introduce two dedicated blocks to adapt latitude-related distortion: distortion-aware attention block (DAAB), and distortion-aware convolution block (DACB). DAAB and DACB are designed to perform distortion modulation in arbitrary Transformers and ConvNets. These two blocks can directly replace the multi-head self-attention block and convolution layer, respectively. The benefit of DAAB and DACB can be further improved when being inserted into the same backbone network. OSRT outperforms previous methods by about 0.2dB on PSNR (Tab. 2)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 38, + 484, + 279, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 484, + 279, + 677 + ], + "spans": [ + { + "bbox": [ + 38, + 484, + 279, + 677 + ], + "type": "text", + "content": "However, the increase of network capacity will also enlarge the overfitting problem of ODISR, which is rarely mentioned before. The largest ODIs dataset [12] contains only 1K images, which cannot provide enough diversity for training Transformers. Given that acquiring ODIs requires expensive equipment and tedious work, we propose to generate distorted ERP samples from plain images for data augmentation. In practice, we regard a plain image as a sampled perspective, and project it back to the ERP format. Then we can introduce 146K additional training patches, 6 times of the previous dataset. This simple strategy can significantly boost the performance of ODISR (Tab. 4) and alleviate the over-fitting problem of large networks (Fig. 9). A similar data augmentation method is also applied in Nishiyama et al. [28], but shows marginal improvement on small models under ERP downsampling settings." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 38, + 678, + 279, + 740 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 678, + 279, + 740 + ], + "spans": [ + { + "bbox": [ + 38, + 678, + 279, + 740 + ], + "type": "text", + "content": "Our contributions are threefold. 1) For problem formulation: To generate more realistic ERP low-resolution images, we propose Fisheye downsampling, which mimics the real-world imaging process. 2) For method: Combined with the geometric properties of ERP, we design a distortion-aware" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 296, + 96, + 538, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 96, + 538, + 171 + ], + "spans": [ + { + "bbox": [ + 296, + 96, + 538, + 171 + ], + "type": "text", + "content": "Transformer, which modulates distortions continuously and self-adaptively without cumbersome process. 3) For data: To reduce overfitting, we propose a convenient data augmentation strategy, which synthesizes pseudo ERP images from plain images. Extensive experiments have demonstrated the state-of-the-art performance of our OSRT2." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 297, + 182, + 384, + 195 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 182, + 384, + 195 + ], + "spans": [ + { + "bbox": [ + 297, + 182, + 384, + 195 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 296, + 203, + 538, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 203, + 538, + 371 + ], + "spans": [ + { + "bbox": [ + 296, + 203, + 538, + 371 + ], + "type": "text", + "content": "Single Image Super-Resolution (SISR). Deep learning for single image SR (SISR) is first introduced in [13]. Further works boost SR performance by CNNs [11, 14, 22, 24, 26, 29, 43], Vision Transformers (ViTs) [7, 8, 21, 23] and generative adversarial networks (GANs) [20, 35, 36, 42]. For instance, EDSR [24] removes Batch Normalization layers and applies a more complicated residual block. RCAN [43] introduces channel-wise attention mechanisms to a deeper network. SwinIR [23] proposes an image restoration Transformer based on [25]. To improve perceptual quality, adversarial training are performed as a tuning process to generate more realistic results [35, 36]. Moreover, various flexible degradation models are proposed in [35, 41] to synthesize more practical degradations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 296, + 373, + 538, + 719 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 373, + 538, + 719 + ], + "spans": [ + { + "bbox": [ + 296, + 373, + 538, + 719 + ], + "type": "text", + "content": "Omnidirectional Image Super-Resolution (ODISR). Initially, ODISR models focus on the spherical assembling of LR ODIs under various projection types [2-4, 17, 27]. Recent ODISR models are performed on plane images and are fine-tuned from existing SISR models with L1 loss [15] or GAN loss [30, 44]. The improvements are limited, for they only concern the distribution gap between ODIs and plain images. Since LAU-Net [12] found pixel density in ERP ODIs is non-uniform, many studies attempt to design specific backbone networks to overcome this issue. LAU-Net [12] manually splits the whole ERP image into latitude-related patches and learns ERP distortion over different latitude ranges separately. While LAU-Net learns latitude-related ERP distortion somewhat, its non-overlapped patches lead to disconnection in whole ERP images. Nishiyama et al. [28] treats area stretching ratio as additional input. However, these conditions are tough to be utilized with an unmodified SISR backbone network. SphereSR [40] learns upsampling processes on various projection types to mitigate the influence of non-uniformity in specific projection types. Although SphereSR improves information consistency between various ODI projection types by LIIF [9], they apply multiple networks to learn the upscaling process of each projection type. Given that all other projection types in SphereSR are converted from ERP, patterns under various types are reusable when distortions are properly rectified. Moreover, the complex and unstructured image data in polyhedron projection hinders further research of ODISR." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 308, + 727, + 436, + 738 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 727, + 436, + 738 + ], + "spans": [ + { + "bbox": [ + 308, + 727, + 436, + 738 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 308, + 727, + 436, + 738 + ], + "type": "text", + "content": "https://github.com/Fanghua-Yu/OSRT" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 277, + 751, + 301, + 761 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 751, + 301, + 761 + ], + "spans": [ + { + "bbox": [ + 277, + 751, + 301, + 761 + ], + "type": "text", + "content": "13284" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 38, + 96, + 279, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 96, + 279, + 181 + ], + "spans": [ + { + "bbox": [ + 38, + 96, + 279, + 181 + ], + "type": "text", + "content": "Deformable Mechanism. Dai et al. [30] first propose deformable convolutions to obtain information out of its regular neighborhood. Xia et al. [38] further verified that Vision Transformers also benefit from applying deformable mechanisms on self-attention blocks. In Video SR tasks, the deformable mechanism can be adapted to align features between adjacent frames [5,6,34]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 39, + 191, + 95, + 204 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 39, + 191, + 95, + 204 + ], + "spans": [ + { + "bbox": [ + 39, + 191, + 95, + 204 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 38, + 211, + 279, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 211, + 279, + 283 + ], + "spans": [ + { + "bbox": [ + 38, + 211, + 279, + 283 + ], + "type": "text", + "content": "In this section, we first analyze the cause of ERP and Fisheye distortions, as well as the relationship between these two distortions (Sec. 3.1). Then, we discuss the designs of Fisheye downsampling (Sec. 3.2), distortion-aware Transformer (OSRT) (Sec. 3.3), and the convenient data augmentation strategy (Sec. 3.4)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 39, + 290, + 203, + 303 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 39, + 290, + 203, + 303 + ], + "spans": [ + { + "bbox": [ + 39, + 290, + 203, + 303 + ], + "type": "text", + "content": "3.1. Revisiting Distortions in ODIs" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 43, + 312, + 279, + 471 + ], + "blocks": [ + { + "bbox": [ + 43, + 312, + 279, + 471 + ], + "lines": [ + { + "bbox": [ + 43, + 312, + 279, + 471 + ], + "spans": [ + { + "bbox": [ + 43, + 312, + 279, + 471 + ], + "type": "image", + "image_path": "fb8333e243e50452470bb4f9fe79baddd5ab63b25639d7b4b09b9ed9068e956d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 38, + 474, + 279, + 507 + ], + "lines": [ + { + "bbox": [ + 38, + 474, + 279, + 507 + ], + "spans": [ + { + "bbox": [ + 38, + 474, + 279, + 507 + ], + "type": "text", + "content": "Figure 2. Geometric explanation of transforming between ERP, Fisheye, and the ideal spherical surface. To simplify, we discuss the horizontal spliced Fisheye with an aperture degree of " + }, + { + "bbox": [ + 38, + 474, + 279, + 507 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 38, + 474, + 279, + 507 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 38, + 511, + 279, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 511, + 279, + 642 + ], + "spans": [ + { + "bbox": [ + 38, + 511, + 279, + 642 + ], + "type": "text", + "content": "As ODIs under each projection type are constrained by different transforming equations, the distortion caused by each type is inconsistent, indicating that applying matrix operations under one projection type can introduce unexpected changes when being converted to other types. Specifically, applying uniformed bicubic downsampling on ERP images will affect the distribution of pixel density on Fisheye images, which are the original-formatted image type of imaging process in real-world scenarios. To analyze the specific effect of ERP downsampling on the Fisheye image, we revisit the cause of distortions in ERP and Fisheye." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 38, + 643, + 279, + 738 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 643, + 279, + 738 + ], + "spans": [ + { + "bbox": [ + 38, + 643, + 279, + 738 + ], + "type": "text", + "content": "As we assume that viewing directions are uniformly distributed, the data points in an ideal ODI should be uniformly distributed on a spherical surface. In practice, there is a trade-off between the uniformity of the spherical surface and the structural degree. ERP is the most convenient projection type for storage or transmission, but it is also the projection type that suffers the heaviest distortion. To better explain the causation of distortions, we follow the definition" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 296, + 96, + 538, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 96, + 538, + 158 + ], + "spans": [ + { + "bbox": [ + 296, + 96, + 538, + 158 + ], + "type": "text", + "content": "of stretching ratio " + }, + { + "bbox": [ + 296, + 96, + 538, + 158 + ], + "type": "inline_equation", + "content": "(\\mathbf{K})" + }, + { + "bbox": [ + 296, + 96, + 538, + 158 + ], + "type": "text", + "content": " in [31], which represents distortion degree at different locations from the target projection type to the ideal spherical surface. " + }, + { + "bbox": [ + 296, + 96, + 538, + 158 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 296, + 96, + 538, + 158 + ], + "type": "text", + "content": " is determined by area variation from one projection type to another. When the target type is uniforming spherical surface, " + }, + { + "bbox": [ + 296, + 96, + 538, + 158 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 296, + 96, + 538, + 158 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 164, + 537, + 191 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 164, + 537, + 191 + ], + "spans": [ + { + "bbox": [ + 305, + 164, + 537, + 191 + ], + "type": "interline_equation", + "content": "\\mathbf {K} (x, y) = \\frac {\\delta S (\\theta , \\varphi)}{\\delta P (x , y)} = \\frac {\\cos (\\varphi) | d \\theta d \\varphi |}{| d x d y |} = \\frac {\\cos (\\varphi)}{| J (\\theta , \\varphi) |}, \\tag {1}", + "image_path": "d59f194dfdf4d53a77cafb085d533510a40b2bc7d065a3cd561fdde9f2f7187e.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 296, + 198, + 537, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 198, + 537, + 246 + ], + "spans": [ + { + "bbox": [ + 296, + 198, + 537, + 246 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 296, + 198, + 537, + 246 + ], + "type": "inline_equation", + "content": "\\delta S(\\cdot ,\\cdot)" + }, + { + "bbox": [ + 296, + 198, + 537, + 246 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 296, + 198, + 537, + 246 + ], + "type": "inline_equation", + "content": "\\delta P(\\cdot ,\\cdot)" + }, + { + "bbox": [ + 296, + 198, + 537, + 246 + ], + "type": "text", + "content": " represent the area on the spherical surface and the projection plane, respectively. " + }, + { + "bbox": [ + 296, + 198, + 537, + 246 + ], + "type": "inline_equation", + "content": "|didj|" + }, + { + "bbox": [ + 296, + 198, + 537, + 246 + ], + "type": "text", + "content": " represents plane microunit. " + }, + { + "bbox": [ + 296, + 198, + 537, + 246 + ], + "type": "inline_equation", + "content": "|J(\\theta ,\\varphi)|" + }, + { + "bbox": [ + 296, + 198, + 537, + 246 + ], + "type": "text", + "content": " is the Jacobian determinant from spherical coordinate to projection coordinate." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 296, + 246, + 537, + 271 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 246, + 537, + 271 + ], + "spans": [ + { + "bbox": [ + 296, + 246, + 537, + 271 + ], + "type": "text", + "content": "ERP distortion. The coordinate in ERP is defined as " + }, + { + "bbox": [ + 296, + 246, + 537, + 271 + ], + "type": "inline_equation", + "content": "x = \\theta" + }, + { + "bbox": [ + 296, + 246, + 537, + 271 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 296, + 246, + 537, + 271 + ], + "type": "inline_equation", + "content": "y = \\varphi" + }, + { + "bbox": [ + 296, + 246, + 537, + 271 + ], + "type": "text", + "content": ". ERP stretching ratio can be derived as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 349, + 279, + 537, + 293 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 279, + 537, + 293 + ], + "spans": [ + { + "bbox": [ + 349, + 279, + 537, + 293 + ], + "type": "interline_equation", + "content": "\\mathbf {K} _ {\\mathrm {E R P}} (x, y) = \\cos (\\varphi) = \\cos (y), \\tag {2}", + "image_path": "d03c56fa4d9cdc4753b5c5d97573c66655f4b8453da8d49dadcfc192333f4df2.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 296, + 300, + 436, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 300, + 436, + 314 + ], + "spans": [ + { + "bbox": [ + 296, + 300, + 436, + 314 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 296, + 300, + 436, + 314 + ], + "type": "inline_equation", + "content": "x\\in (-\\pi ,\\pi)" + }, + { + "bbox": [ + 296, + 300, + 436, + 314 + ], + "type": "inline_equation", + "content": "y\\in (-\\frac{\\pi}{2},\\frac{\\pi}{2})" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 296, + 313, + 537, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 313, + 537, + 432 + ], + "spans": [ + { + "bbox": [ + 296, + 313, + 537, + 432 + ], + "type": "text", + "content": "From Eq. (2), we conclude that ERP distortion is only determined by its latitude degree. " + }, + { + "bbox": [ + 296, + 313, + 537, + 432 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_{\\mathrm{ERP}}" + }, + { + "bbox": [ + 296, + 313, + 537, + 432 + ], + "type": "text", + "content": " is reduced to zero when the absolute value of latitude degree increases to " + }, + { + "bbox": [ + 296, + 313, + 537, + 432 + ], + "type": "inline_equation", + "content": "\\pi /2" + }, + { + "bbox": [ + 296, + 313, + 537, + 432 + ], + "type": "text", + "content": ", which represents that pixel density on the polar areas of ERP images is closer to zero. As shown in Fig. 2 (c), with the increasing of the absolutely value of latitude degree " + }, + { + "bbox": [ + 296, + 313, + 537, + 432 + ], + "type": "inline_equation", + "content": "(|\\varphi_{s}|)" + }, + { + "bbox": [ + 296, + 313, + 537, + 432 + ], + "type": "text", + "content": ", the corresponding area on the spherical surface of an ERP microunit is gradually decreased to zero. In conclusion, ERP distortion is caused by variable stretching ratios " + }, + { + "bbox": [ + 296, + 313, + 537, + 432 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_{\\mathrm{ERP}}" + }, + { + "bbox": [ + 296, + 313, + 537, + 432 + ], + "type": "text", + "content": ", and is the heaviest in the polar areas." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 296, + 432, + 537, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 432, + 537, + 468 + ], + "spans": [ + { + "bbox": [ + 296, + 432, + 537, + 468 + ], + "type": "text", + "content": "Fisheye distortion. The coordinate in Fisheye can be derived from " + }, + { + "bbox": [ + 296, + 432, + 537, + 468 + ], + "type": "inline_equation", + "content": "\\theta = \\arctan \\left(\\frac{y}{x}\\right)" + }, + { + "bbox": [ + 296, + 432, + 537, + 468 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 296, + 432, + 537, + 468 + ], + "type": "inline_equation", + "content": "\\varphi = (1 - \\sqrt{x^2 + y^2})\\times \\frac{\\pi}{2}" + }, + { + "bbox": [ + 296, + 432, + 537, + 468 + ], + "type": "text", + "content": ". The stretching ratios of Fisheye can be derived as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 337, + 475, + 537, + 507 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 475, + 537, + 507 + ], + "spans": [ + { + "bbox": [ + 337, + 475, + 537, + 507 + ], + "type": "interline_equation", + "content": "\\mathbf {K} _ {\\text {F i s h e y e}} (x, y) = \\frac {\\frac {2}{\\pi} \\sin \\left(\\frac {\\pi}{2} \\sqrt {x ^ {2} + y ^ {2}}\\right)}{\\sqrt {x ^ {2} + y ^ {2}}}, \\tag {3}", + "image_path": "a640c1ce00ee31ec50364f60e5a0e29f1f6ee9a6e8e5c3a4837b3ddcb1598145.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 296, + 515, + 406, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 515, + 406, + 528 + ], + "spans": [ + { + "bbox": [ + 296, + 515, + 406, + 528 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 296, + 515, + 406, + 528 + ], + "type": "inline_equation", + "content": "\\sqrt{x^2 + y^2} \\in (0,1)" + }, + { + "bbox": [ + 296, + 515, + 406, + 528 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 296, + 528, + 537, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 528, + 537, + 576 + ], + "spans": [ + { + "bbox": [ + 296, + 528, + 537, + 576 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_{\\mathrm{Fisheye}}" + }, + { + "bbox": [ + 296, + 528, + 537, + 576 + ], + "type": "text", + "content": " is determined by distance from the fisheye center. As " + }, + { + "bbox": [ + 296, + 528, + 537, + 576 + ], + "type": "inline_equation", + "content": "(\\mathbf{K}_{\\mathrm{Fisheye}})^{-1}" + }, + { + "bbox": [ + 296, + 528, + 537, + 576 + ], + "type": "text", + "content": " is bounded, fisheye projection is closer to uniform distribution than ERP. Moreover, it introduces much slighter distortion at the polar." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 296, + 576, + 537, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 576, + 537, + 636 + ], + "spans": [ + { + "bbox": [ + 296, + 576, + 537, + 636 + ], + "type": "text", + "content": "Relationship between ERP and Fisheye distortions. To simplify, here we only discuss a typical Fisheye with an aperture degree of " + }, + { + "bbox": [ + 296, + 576, + 537, + 636 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 296, + 576, + 537, + 636 + ], + "type": "text", + "content": " and a horizontal slicing plane4. In this case, the ERP coordinates and Fisheye's polar coordinates correspond linearly. We can quantize the relationship by:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 302, + 644, + 537, + 671 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 644, + 537, + 671 + ], + "spans": [ + { + "bbox": [ + 302, + 644, + 537, + 671 + ], + "type": "interline_equation", + "content": "\\mathbf {K} _ {\\mathrm {E R P} \\mid \\text {F i s h e y e}} (\\theta , \\varphi) = \\frac {\\mathbf {K} _ {\\mathrm {E R P}} \\left(x _ {E} , y _ {E}\\right)}{\\mathbf {K} _ {\\text {F i s h e y e}} \\left(x _ {F} , y _ {F}\\right)} = \\frac {\\pi}{2} - | \\varphi |, \\tag {4}", + "image_path": "e33aa38ab9b29fd0ec3295af541153f9648c9a950f84dcca01b0bcc38a7ee5da.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 296, + 677, + 537, + 702 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 677, + 537, + 702 + ], + "spans": [ + { + "bbox": [ + 296, + 677, + 537, + 702 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 296, + 677, + 537, + 702 + ], + "type": "inline_equation", + "content": "\\theta, \\varphi" + }, + { + "bbox": [ + 296, + 677, + 537, + 702 + ], + "type": "text", + "content": " are spherical coordinates on the sphere, " + }, + { + "bbox": [ + 296, + 677, + 537, + 702 + ], + "type": "inline_equation", + "content": "x_{E}, y_{E}" + }, + { + "bbox": [ + 296, + 677, + 537, + 702 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 296, + 677, + 537, + 702 + ], + "type": "inline_equation", + "content": "x_{F}, y_{F}" + }, + { + "bbox": [ + 296, + 677, + 537, + 702 + ], + "type": "text", + "content": ") denotes the plain coordinate under ERP (Fisheye)." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 307, + 708, + 533, + 719 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 708, + 533, + 719 + ], + "spans": [ + { + "bbox": [ + 307, + 708, + 533, + 719 + ], + "type": "text", + "content": "3Detailed derivative processes can be found in the supplementary file." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 298, + 719, + 536, + 738 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 719, + 536, + 738 + ], + "spans": [ + { + "bbox": [ + 298, + 719, + 536, + 738 + ], + "type": "text", + "content": "4The influence of Fisheye formats with arbitrary splicing plane is discussed in the supplementary file." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 277, + 751, + 300, + 760 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 751, + 300, + 760 + ], + "spans": [ + { + "bbox": [ + 277, + 751, + 300, + 760 + ], + "type": "text", + "content": "13285" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 41, + 96, + 274, + 223 + ], + "blocks": [ + { + "bbox": [ + 41, + 96, + 274, + 223 + ], + "lines": [ + { + "bbox": [ + 41, + 96, + 274, + 223 + ], + "spans": [ + { + "bbox": [ + 41, + 96, + 274, + 223 + ], + "type": "image", + "image_path": "7b63ac2ee9f66d49a42ce1663d2e22e5d7a14e6a3c7436c95582879c443377da.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 39, + 228, + 278, + 261 + ], + "lines": [ + { + "bbox": [ + 39, + 228, + 278, + 261 + ], + "spans": [ + { + "bbox": [ + 39, + 228, + 278, + 261 + ], + "type": "text", + "content": "Figure 3. Downsampling process of ODIs (left) and imaging process in real world (right). * denotes that LR images synthesized from different downsampling processes are inconsistent." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 38, + 269, + 279, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 269, + 279, + 329 + ], + "spans": [ + { + "bbox": [ + 38, + 269, + 279, + 329 + ], + "type": "text", + "content": "From Eqs. (2) to (4), we conclude that when uniformed downsampling is performed on ERP, the kernel size of equivalent Fisheye downsampling is non-uniformed. Especially when fisheye projection is spliced horizontally, the kernel size is proportional with " + }, + { + "bbox": [ + 38, + 269, + 279, + 329 + ], + "type": "inline_equation", + "content": "\\pi / 2 - |\\varphi|" + }, + { + "bbox": [ + 38, + 269, + 279, + 329 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 39, + 336, + 260, + 349 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 39, + 336, + 260, + 349 + ], + "spans": [ + { + "bbox": [ + 39, + 336, + 260, + 349 + ], + "type": "text", + "content": "3.2. Learning with More Realistic Degradation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 38, + 355, + 278, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 355, + 278, + 426 + ], + "spans": [ + { + "bbox": [ + 38, + 355, + 278, + 426 + ], + "type": "text", + "content": "As depicted in Fig. 3, the original-formatted projection type in ODI acquiring process is fisheye projection. Given that real-world low-resolution issues are caused by insufficient precision and density of sensors, we consider that the degeneration process should be directly applied to original-formatted images before the type conversion." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 38, + 427, + 278, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 427, + 278, + 594 + ], + "spans": [ + { + "bbox": [ + 38, + 427, + 278, + 594 + ], + "type": "text", + "content": "Ideally, as camera sensors are arranged in uniform arrays, pixel density on original-formatted images is consistent everywhere. Thus, for a realistic ODI, the pixel density on Fisheye should be a constant. As discussed in Sec. 3.1, applying uniformed downsampling on ERP means applying downsampling of variable kernel size on Fisheye. The variable kernel size leads to variable Fisheye pixel density, which results in unrealistic LR images. In conclusion, the ERP downsampling in previous methods influences the intrinsic distribution of pixel density in original-formatted images, which leads to unrealistic ODIs. When the downsampling process happens on Fisheye, the Fisheye pixel density is unchanged, which fits the real-world imaging process and synthesizes more realistic LR pairs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 38, + 595, + 278, + 738 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 595, + 278, + 738 + ], + "spans": [ + { + "bbox": [ + 38, + 595, + 278, + 738 + ], + "type": "text", + "content": "Process of Fisheye downsampling. To generate more realistic LR ODIs, we mimic the real-world imaging process and apply bicubic downsampling on Fisheye images. One single Fisheye image can only store information about a hemisphere. Hence, ERP images are converted to dual Fisheye images. Before downsampling, Fisheye images are padded by a FOV larger than " + }, + { + "bbox": [ + 38, + 595, + 278, + 738 + ], + "type": "inline_equation", + "content": "180^{\\circ}" + }, + { + "bbox": [ + 38, + 595, + 278, + 738 + ], + "type": "text", + "content": " to avoid edge disconnections. This padding operation will not influence the geometric transforming relation between ERP and Fisheye. As Fisheye data is unstructured and Fisheye distortion is more complicated than ERP distortion, we still learn the upscaling process under ERP. Thus we reconvert LR images to the" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 297, + 97, + 536, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 97, + 536, + 121 + ], + "spans": [ + { + "bbox": [ + 297, + 97, + 536, + 121 + ], + "type": "text", + "content": "ERP format. The overall process of Fisheye downsampling are described in Fig. 3." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 297, + 128, + 490, + 139 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 128, + 490, + 139 + ], + "spans": [ + { + "bbox": [ + 297, + 128, + 490, + 139 + ], + "type": "text", + "content": "3.3.OSRT:Modulate Distortion in ODIs" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 297, + 147, + 536, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 147, + 536, + 207 + ], + "spans": [ + { + "bbox": [ + 297, + 147, + 536, + 207 + ], + "type": "text", + "content": "Overall. As discussed in Sec. 3.1, ERP images suffer a distortion caused by a non-consistency area stretching ratio from an ideal spherical surface. Referred from Eq. (2), for an LR input " + }, + { + "bbox": [ + 297, + 147, + 536, + 207 + ], + "type": "inline_equation", + "content": "X_{i}\\in \\mathbb{R}^{C\\times M\\times N}" + }, + { + "bbox": [ + 297, + 147, + 536, + 207 + ], + "type": "text", + "content": " , the distortion map " + }, + { + "bbox": [ + 297, + 147, + 536, + 207 + ], + "type": "inline_equation", + "content": "C_d\\in \\mathbb{R}^{1\\times M\\times N}" + }, + { + "bbox": [ + 297, + 147, + 536, + 207 + ], + "type": "text", + "content": " is derived by:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 348, + 214, + 536, + 240 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 214, + 536, + 240 + ], + "spans": [ + { + "bbox": [ + 348, + 214, + 536, + 240 + ], + "type": "interline_equation", + "content": "C _ {d} = \\cos \\left(\\frac {m + 0 . 5 - M / 2}{M} \\pi\\right), \\tag {5}", + "image_path": "fff3f6a372657bb1a90ddc0b54c975c2028ea844c91cb2734d7609c0111228bb.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 297, + 248, + 468, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 248, + 468, + 260 + ], + "spans": [ + { + "bbox": [ + 297, + 248, + 468, + 260 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 297, + 248, + 468, + 260 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 297, + 248, + 468, + 260 + ], + "type": "text", + "content": " is the current height of LR input." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 296, + 260, + 537, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 260, + 537, + 522 + ], + "spans": [ + { + "bbox": [ + 296, + 260, + 537, + 522 + ], + "type": "text", + "content": "Previous methods tend to treat " + }, + { + "bbox": [ + 296, + 260, + 537, + 522 + ], + "type": "inline_equation", + "content": "C_d" + }, + { + "bbox": [ + 296, + 260, + 537, + 522 + ], + "type": "text", + "content": " as an additional input of " + }, + { + "bbox": [ + 296, + 260, + 537, + 522 + ], + "type": "inline_equation", + "content": "X_i" + }, + { + "bbox": [ + 296, + 260, + 537, + 522 + ], + "type": "text", + "content": " [28], or re-weighting parameters by " + }, + { + "bbox": [ + 296, + 260, + 537, + 522 + ], + "type": "inline_equation", + "content": "C_d" + }, + { + "bbox": [ + 296, + 260, + 537, + 522 + ], + "type": "text", + "content": " [18]. Although these solutions can benefit from building awareness of distortion, continuous and amorphous distortions cannot be adequately fitted by scattering and structured convolution operations. While previous methods cannot fully explore the advantage of " + }, + { + "bbox": [ + 296, + 260, + 537, + 522 + ], + "type": "inline_equation", + "content": "C_d" + }, + { + "bbox": [ + 296, + 260, + 537, + 522 + ], + "type": "text", + "content": ", we intend to design a novel block for learning distorted patterns continuously. In VSR tasks, the deformable mechanism is proposed to align features between adjacent frames [32, 34]. Unlike standard DCN [10], which calculates offsets from the input feature map, offsets are calculated from bi-directional optical flow in VSR pipelines. Inspired by feature-level flow warping in VSR, we find that the deformable mechanism is a feasible solution for continuous mappings. Consequently, we modulate ERP distortion by feature-level warping operations. As shown in Fig. 4, " + }, + { + "bbox": [ + 296, + 260, + 537, + 522 + ], + "type": "inline_equation", + "content": "C_d" + }, + { + "bbox": [ + 296, + 260, + 537, + 522 + ], + "type": "text", + "content": " is only utilized to calculate the deformable offsets " + }, + { + "bbox": [ + 296, + 260, + 537, + 522 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 296, + 260, + 537, + 522 + ], + "type": "text", + "content": ". To keep compatibility with arbitrary ConvNets and Transformers, we propose two blocks to modulate ERP distortion, which can directly replace the multi-head self-attention blocks in Transformers and the standard convolution layers in ConvNets, respectively." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 296, + 523, + 537, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 523, + 537, + 582 + ], + "spans": [ + { + "bbox": [ + 296, + 523, + 537, + 582 + ], + "type": "text", + "content": "Distortion-aware attention block (DAAB). As depicted in Fig. 4 (a), a distortion condition guided deformable self-attention is proposed to learn correlations between the distorted input " + }, + { + "bbox": [ + 296, + 523, + 537, + 582 + ], + "type": "inline_equation", + "content": "X_{i}" + }, + { + "bbox": [ + 296, + 523, + 537, + 582 + ], + "type": "text", + "content": " and its corresponding modulated feature map " + }, + { + "bbox": [ + 296, + 523, + 537, + 582 + ], + "type": "inline_equation", + "content": "\\tilde{X}_{i}" + }, + { + "bbox": [ + 296, + 523, + 537, + 582 + ], + "type": "text", + "content": ". DAAB is formulated as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 324, + 591, + 536, + 604 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 591, + 536, + 604 + ], + "spans": [ + { + "bbox": [ + 324, + 591, + 536, + 604 + ], + "type": "interline_equation", + "content": "\\theta_ {i} = H _ {\\text {o f f s e t} _ {i}} \\left(C _ {d}, C _ {w}\\right), \\tilde {X} _ {i} = \\phi \\left(X _ {i}; p _ {0} + \\theta_ {i}\\right), \\tag {6}", + "image_path": "c30a186d2dae748f58dde8c44a8d826e53d4c58cb6f49bb37aa32fe5a4216600.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 341, + 612, + 536, + 626 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 612, + 536, + 626 + ], + "spans": [ + { + "bbox": [ + 341, + 612, + 536, + 626 + ], + "type": "interline_equation", + "content": "X _ {o} = H _ {\\mathrm {S A}} \\left(X _ {i} W _ {q _ {i}}, \\tilde {X} _ {i} W _ {k _ {i}}, \\tilde {X} _ {i} W _ {v _ {i}}\\right), \\tag {7}", + "image_path": "4b7c58c10bba443165f64ac229f59a59d79e7950b4e37d9547dcd56bd1b80aed.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 296, + 630, + 537, + 738 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 630, + 537, + 738 + ], + "spans": [ + { + "bbox": [ + 296, + 630, + 537, + 738 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 296, + 630, + 537, + 738 + ], + "type": "inline_equation", + "content": "H_{\\mathrm{offset}_i}(\\cdot)" + }, + { + "bbox": [ + 296, + 630, + 537, + 738 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 296, + 630, + 537, + 738 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 296, + 630, + 537, + 738 + ], + "type": "text", + "content": "-th convolution block to calculate offset maps " + }, + { + "bbox": [ + 296, + 630, + 537, + 738 + ], + "type": "inline_equation", + "content": "\\theta_i \\in \\mathbb{R}^{2 \\times H \\times W}" + }, + { + "bbox": [ + 296, + 630, + 537, + 738 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 296, + 630, + 537, + 738 + ], + "type": "inline_equation", + "content": "H_{\\mathrm{SA}}" + }, + { + "bbox": [ + 296, + 630, + 537, + 738 + ], + "type": "text", + "content": " denotes standard self-attention formula. " + }, + { + "bbox": [ + 296, + 630, + 537, + 738 + ], + "type": "inline_equation", + "content": "H_{\\mathrm{offset}}(\\cdot)" + }, + { + "bbox": [ + 296, + 630, + 537, + 738 + ], + "type": "text", + "content": " consists of " + }, + { + "bbox": [ + 296, + 630, + 537, + 738 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 296, + 630, + 537, + 738 + ], + "type": "text", + "content": " convolution block with two hidden layers. The input of " + }, + { + "bbox": [ + 296, + 630, + 537, + 738 + ], + "type": "inline_equation", + "content": "H_{\\mathrm{offset}}(\\cdot)" + }, + { + "bbox": [ + 296, + 630, + 537, + 738 + ], + "type": "text", + "content": " is concatenated by the latitude-related distortion condition " + }, + { + "bbox": [ + 296, + 630, + 537, + 738 + ], + "type": "inline_equation", + "content": "C_d \\in \\mathbb{R}^{1 \\times H \\times W}" + }, + { + "bbox": [ + 296, + 630, + 537, + 738 + ], + "type": "text", + "content": " and the window condition " + }, + { + "bbox": [ + 296, + 630, + 537, + 738 + ], + "type": "inline_equation", + "content": "C_w \\in \\mathbb{R}^{2 \\times H \\times W}" + }, + { + "bbox": [ + 296, + 630, + 537, + 738 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 296, + 630, + 537, + 738 + ], + "type": "inline_equation", + "content": "C_w" + }, + { + "bbox": [ + 296, + 630, + 537, + 738 + ], + "type": "text", + "content": " is a linear position encoding within a self-attention kernel. " + }, + { + "bbox": [ + 296, + 630, + 537, + 738 + ], + "type": "inline_equation", + "content": "\\phi(\\cdot, \\cdot)" + }, + { + "bbox": [ + 296, + 630, + 537, + 738 + ], + "type": "text", + "content": " denotes a bilinear interpolation, and " + }, + { + "bbox": [ + 296, + 630, + 537, + 738 + ], + "type": "inline_equation", + "content": "W_{q_i}, W_{k_i}, W_{v_i}" + }, + { + "bbox": [ + 296, + 630, + 537, + 738 + ], + "type": "text", + "content": " denote " + }, + { + "bbox": [ + 296, + 630, + 537, + 738 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 296, + 630, + 537, + 738 + ], + "type": "text", + "content": "-th weight matrix of query, key, and" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 277, + 751, + 300, + 760 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 751, + 300, + 760 + ], + "spans": [ + { + "bbox": [ + 277, + 751, + 300, + 760 + ], + "type": "text", + "content": "13286" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 88, + 530, + 188 + ], + "blocks": [ + { + "bbox": [ + 49, + 88, + 530, + 188 + ], + "lines": [ + { + "bbox": [ + 49, + 88, + 530, + 188 + ], + "spans": [ + { + "bbox": [ + 49, + 88, + 530, + 188 + ], + "type": "image", + "image_path": "9a56a726695fd45479635425fdb85149a3f247dcad2883417c18b0ba871511bf.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 45, + 190, + 302, + 319 + ], + "blocks": [ + { + "bbox": [ + 45, + 190, + 302, + 319 + ], + "lines": [ + { + "bbox": [ + 45, + 190, + 302, + 319 + ], + "spans": [ + { + "bbox": [ + 45, + 190, + 302, + 319 + ], + "type": "image", + "image_path": "72c5be446d0de73ba147472880a1c4c2d735a796168cf0b389d1d4d73f1f7b1e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 38, + 326, + 537, + 348 + ], + "lines": [ + { + "bbox": [ + 38, + 326, + 537, + 348 + ], + "spans": [ + { + "bbox": [ + 38, + 326, + 537, + 348 + ], + "type": "text", + "content": "Figure 4. Overall illustration of OSRT. From SwinIR [23], we replace the standard multi-head self-attention block with DAAB and insert DACB behind the end of the RSTB. Channel dimensions of " + }, + { + "bbox": [ + 38, + 326, + 537, + 348 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathrm{offsets}}" + }, + { + "bbox": [ + 38, + 326, + 537, + 348 + ], + "type": "text", + "content": " in DAAB and DACB are 2 and 18, respectively." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 305, + 190, + 433, + 320 + ], + "blocks": [ + { + "bbox": [ + 305, + 190, + 433, + 320 + ], + "lines": [ + { + "bbox": [ + 305, + 190, + 433, + 320 + ], + "spans": [ + { + "bbox": [ + 305, + 190, + 433, + 320 + ], + "type": "image", + "image_path": "a884f0fbde197c90940bbf89f5d5f09dd47afaa086cf5514ba23ee5e36df7a20.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 435, + 190, + 529, + 320 + ], + "blocks": [ + { + "bbox": [ + 435, + 190, + 529, + 320 + ], + "lines": [ + { + "bbox": [ + 435, + 190, + 529, + 320 + ], + "spans": [ + { + "bbox": [ + 435, + 190, + 529, + 320 + ], + "type": "image", + "image_path": "154b285d4c54c969b4bd95e5610f05814e7fd4b54cb468d4d98ec941b8f4e740.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 39, + 352, + 277, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 39, + 352, + 277, + 376 + ], + "spans": [ + { + "bbox": [ + 39, + 352, + 277, + 376 + ], + "type": "text", + "content": "value, respectively. For multi-head self-attention blocks, " + }, + { + "bbox": [ + 39, + 352, + 277, + 376 + ], + "type": "inline_equation", + "content": "H_{\\mathrm{offset}_i}(\\cdot)" + }, + { + "bbox": [ + 39, + 352, + 277, + 376 + ], + "type": "text", + "content": " is identical in calculations of parallel heads." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 38, + 377, + 278, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 377, + 278, + 424 + ], + "spans": [ + { + "bbox": [ + 38, + 377, + 278, + 424 + ], + "type": "text", + "content": "Distortion-aware convolution block (DACB). As shown in Fig. 4 (b), we apply a standard deformable convolution layer with a substituted input for offset calculation. Modulated output " + }, + { + "bbox": [ + 38, + 377, + 278, + 424 + ], + "type": "inline_equation", + "content": "X_{o}" + }, + { + "bbox": [ + 38, + 377, + 278, + 424 + ], + "type": "text", + "content": " is extracted as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 73, + 431, + 278, + 444 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 431, + 278, + 444 + ], + "spans": [ + { + "bbox": [ + 73, + 431, + 278, + 444 + ], + "type": "interline_equation", + "content": "\\theta_ {i} = H _ {\\text {o f f s e t} _ {i}} \\left(C _ {d}\\right), X _ {o} = H _ {\\mathrm {D C N} _ {i}} \\left(X _ {i}, \\theta_ {i}\\right), \\tag {8}", + "image_path": "e058c68acffa795dd19b10c4ab3effced124cba8c311610109d2d00a0a390bf4.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 38, + 450, + 278, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 450, + 278, + 498 + ], + "spans": [ + { + "bbox": [ + 38, + 450, + 278, + 498 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 38, + 450, + 278, + 498 + ], + "type": "inline_equation", + "content": "H_{\\mathrm{DCN}}(X,\\theta)" + }, + { + "bbox": [ + 38, + 450, + 278, + 498 + ], + "type": "text", + "content": " denotes standard deformable convolution layer in [46]. The architecture of " + }, + { + "bbox": [ + 38, + 450, + 278, + 498 + ], + "type": "inline_equation", + "content": "H_{\\mathrm{offset}_i}(\\cdot)" + }, + { + "bbox": [ + 38, + 450, + 278, + 498 + ], + "type": "text", + "content": " is identical to that in DAAB. As the kernel size of DCN is " + }, + { + "bbox": [ + 38, + 450, + 278, + 498 + ], + "type": "inline_equation", + "content": "3\\times 3" + }, + { + "bbox": [ + 38, + 450, + 278, + 498 + ], + "type": "text", + "content": " in DACB, the output channel dimension of offsets maps is 18." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 38, + 498, + 278, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 498, + 278, + 606 + ], + "spans": [ + { + "bbox": [ + 38, + 498, + 278, + 606 + ], + "type": "text", + "content": "OSRT. In practice, we propose an Omnidirectional image Super-Resolution Transformer, named OSRT. SwinIR [23] is selected as the basic architecture for its strong reconstruction ability in the SISR task. To learn distortion rectified representations, we stack a DACB after the last convolution layer of each residual swin Transformer block and replace all self-attention blocks as DAAB. The feature dimension of OSRT is reduced from 180 to 156 to maintain identical parameters with SwinIR." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 39, + 612, + 278, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 39, + 612, + 278, + 625 + ], + "spans": [ + { + "bbox": [ + 39, + 612, + 278, + 625 + ], + "type": "text", + "content": "3.4. Boosting ODISR Performance by Plain Images" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 38, + 631, + 278, + 738 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 631, + 278, + 738 + ], + "spans": [ + { + "bbox": [ + 38, + 631, + 278, + 738 + ], + "type": "text", + "content": "As the capacity of OSRT is relatively large, it suffers overfitting for large upscaling factors (Fig. 9). Given that acquiring ODIs are expensive, we propose to generate pseudo ERP images from 2D plain images to tackle this issue. After being sampled by sliding windows, the patch of plain images is treated as a plain perspective. By converting from Perspective to ERP, plain images are distorted in the same way as ERP. Considering that distortion of a Perspective is enlarged by its FOV degree, a relatively small FOV" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 296, + 352, + 536, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 352, + 536, + 378 + ], + "spans": [ + { + "bbox": [ + 296, + 352, + 536, + 378 + ], + "type": "text", + "content": "degree of " + }, + { + "bbox": [ + 296, + 352, + 536, + 378 + ], + "type": "inline_equation", + "content": "90^{\\circ}" + }, + { + "bbox": [ + 296, + 352, + 536, + 378 + ], + "type": "text", + "content": " is applied. For a given pseudo Perspective, " + }, + { + "bbox": [ + 296, + 352, + 536, + 378 + ], + "type": "inline_equation", + "content": "\\theta_{p}" + }, + { + "bbox": [ + 296, + 352, + 536, + 378 + ], + "type": "text", + "content": " is fixed at 0 and " + }, + { + "bbox": [ + 296, + 352, + 536, + 378 + ], + "type": "inline_equation", + "content": "\\varphi_{p}" + }, + { + "bbox": [ + 296, + 352, + 536, + 378 + ], + "type": "text", + "content": " is derived by:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 385, + 389, + 536, + 403 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 385, + 389, + 536, + 403 + ], + "spans": [ + { + "bbox": [ + 385, + 389, + 536, + 403 + ], + "type": "interline_equation", + "content": "\\Phi_ {p} = \\varphi_ {h} + z _ {0}, \\tag {9}", + "image_path": "350d4ebe225347aa72537a285006f0654ee3b071c69bef9e37b4777ffeb879fd.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 296, + 414, + 536, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 414, + 536, + 438 + ], + "spans": [ + { + "bbox": [ + 296, + 414, + 536, + 438 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 296, + 414, + 536, + 438 + ], + "type": "inline_equation", + "content": "\\varphi_h" + }, + { + "bbox": [ + 296, + 414, + 536, + 438 + ], + "type": "text", + "content": " is determined by patch locations and " + }, + { + "bbox": [ + 296, + 414, + 536, + 438 + ], + "type": "inline_equation", + "content": "z_0" + }, + { + "bbox": [ + 296, + 414, + 536, + 438 + ], + "type": "text", + "content": " is orderly sampled from " + }, + { + "bbox": [ + 296, + 414, + 536, + 438 + ], + "type": "inline_equation", + "content": "\\{-15^\\circ, 0^\\circ, 15^\\circ\\}" + }, + { + "bbox": [ + 296, + 414, + 536, + 438 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 296, + 439, + 537, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 439, + 537, + 548 + ], + "spans": [ + { + "bbox": [ + 296, + 439, + 537, + 548 + ], + "type": "text", + "content": "To maximize the approximate data distribution of ODIs, we horizontally split a plain image into three sub-images and define " + }, + { + "bbox": [ + 296, + 439, + 537, + 548 + ], + "type": "inline_equation", + "content": "\\varphi_h" + }, + { + "bbox": [ + 296, + 439, + 537, + 548 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 296, + 439, + 537, + 548 + ], + "type": "inline_equation", + "content": "-30^\\circ, 0^\\circ, 30^\\circ" + }, + { + "bbox": [ + 296, + 439, + 537, + 548 + ], + "type": "text", + "content": " respectively. Pseudo ERP images are cropped to remove the black border. As shown in Fig. 5, we get a new ERP dataset (called DF2K-ERP) by implementing the augmentation pipeline on widely-used plain image dataset DF2K [24, 33]. The DF2K-ERP dataset consists of 146K high-quality ERP image patches with a patch size larger than 256." + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 299, + 557, + 533, + 718 + ], + "blocks": [ + { + "bbox": [ + 299, + 557, + 533, + 718 + ], + "lines": [ + { + "bbox": [ + 299, + 557, + 533, + 718 + ], + "spans": [ + { + "bbox": [ + 299, + 557, + 533, + 718 + ], + "type": "image", + "image_path": "c1511d336fd6d67d2883beed07de61b73a9cf216ef514f4c0641c079d11f14f5.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 339, + 724, + 495, + 735 + ], + "lines": [ + { + "bbox": [ + 339, + 724, + 495, + 735 + ], + "spans": [ + { + "bbox": [ + 339, + 724, + 495, + 735 + ], + "type": "text", + "content": "Figure 5. Synthetic process of DF2K-ERP." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 277, + 751, + 300, + 760 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 751, + 300, + 760 + ], + "spans": [ + { + "bbox": [ + 277, + 751, + 300, + 760 + ], + "type": "text", + "content": "13287" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 80, + 95, + 495, + 244 + ], + "blocks": [ + { + "bbox": [ + 80, + 95, + 495, + 244 + ], + "lines": [ + { + "bbox": [ + 80, + 95, + 495, + 244 + ], + "spans": [ + { + "bbox": [ + 80, + 95, + 495, + 244 + ], + "type": "table", + "html": "
MethodScaleODI-SRPSNRSUN 360 Panorama
PSNRSSIMWS-PSNRWS-SSIMSSIMWS-PSNRWS-SSIM
Bicubic28.210.821527.610.815628.140.811828.010.8321
RCAN [43]30.080.872329.490.871430.560.871231.180.8969
SRResNet [36]30.160.871729.590.869730.650.871431.200.8953
EDSR [24]×230.320.877029.680.872730.890.878431.420.8995
SwinIR [23]30.520.881929.870.877231.210.885231.780.9051
SwinIR†[23]30.640.882130.000.877731.330.885531.980.9059
OSRT†30.770.884630.110.879531.520.888832.140.9081
Bicubic25.590.711824.950.692325.290.699324.900.7083
RCAN [43]26.850.762126.150.748527.100.766026.990.7856
SRResNet [36]26.910.759726.240.745727.100.761826.990.7812
EDSR [24]×426.970.758926.300.745827.190.763327.100.7827
SwinIR [23]27.120.766326.440.752327.390.770727.300.7901
SwinIR†[23]27.310.773526.610.758927.710.780427.640.7996
OSRT†27.410.776226.700.760927.840.783527.770.8020
", + "image_path": "2dba8417b1c2eef86537e9b579bebb779cb3a310fdab95e94001c7905e89784c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 43, + 246, + 533, + 259 + ], + "lines": [ + { + "bbox": [ + 43, + 246, + 533, + 259 + ], + "spans": [ + { + "bbox": [ + 43, + 246, + 533, + 259 + ], + "type": "text", + "content": "Table 1. SR results under Fisheye downsampling. † denotes applying DF2K-ERP as augmented dataset. Best results are shown in Bold." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 47, + 264, + 178, + 395 + ], + "blocks": [ + { + "bbox": [ + 47, + 264, + 178, + 395 + ], + "lines": [ + { + "bbox": [ + 47, + 264, + 178, + 395 + ], + "spans": [ + { + "bbox": [ + 47, + 264, + 178, + 395 + ], + "type": "image", + "image_path": "c932d1fb890ac706ad04dc32f9289fcb0ce311fa8652d4bd663864d44ac46f38.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 85, + 399, + 142, + 408 + ], + "lines": [ + { + "bbox": [ + 85, + 399, + 142, + 408 + ], + "spans": [ + { + "bbox": [ + 85, + 399, + 142, + 408 + ], + "type": "text", + "content": "SUN360 " + }, + { + "bbox": [ + 85, + 399, + 142, + 408 + ], + "type": "inline_equation", + "content": "(\\times 4)" + }, + { + "bbox": [ + 85, + 399, + 142, + 408 + ], + "type": "text", + "content": " : 034" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 187, + 263, + 263, + 321 + ], + "blocks": [ + { + "bbox": [ + 187, + 263, + 263, + 321 + ], + "lines": [ + { + "bbox": [ + 187, + 263, + 263, + 321 + ], + "spans": [ + { + "bbox": [ + 187, + 263, + 263, + 321 + ], + "type": "image", + "image_path": "b7df64b1a03e7f8a36ff41abe586ecaa72ca27ccbbbf0899d86d9fe1d3913914.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 220, + 323, + 230, + 330 + ], + "lines": [ + { + "bbox": [ + 220, + 323, + 230, + 330 + ], + "spans": [ + { + "bbox": [ + 220, + 323, + 230, + 330 + ], + "type": "text", + "content": "HR" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 187, + 337, + 263, + 394 + ], + "blocks": [ + { + "bbox": [ + 207, + 330, + 243, + 337 + ], + "lines": [ + { + "bbox": [ + 207, + 330, + 243, + 337 + ], + "spans": [ + { + "bbox": [ + 207, + 330, + 243, + 337 + ], + "type": "text", + "content": "PSNR/SSIM" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 187, + 337, + 263, + 394 + ], + "lines": [ + { + "bbox": [ + 187, + 337, + 263, + 394 + ], + "spans": [ + { + "bbox": [ + 187, + 337, + 263, + 394 + ], + "type": "image", + "image_path": "f9475fdab31f0999131642a7793161d6eab43d3d887afc1e31b6d85db4da8f12.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 208, + 396, + 242, + 403 + ], + "lines": [ + { + "bbox": [ + 208, + 396, + 242, + 403 + ], + "spans": [ + { + "bbox": [ + 208, + 396, + 242, + 403 + ], + "type": "text", + "content": "EDSR [24]" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 272, + 263, + 348, + 321 + ], + "blocks": [ + { + "bbox": [ + 272, + 263, + 348, + 321 + ], + "lines": [ + { + "bbox": [ + 272, + 263, + 348, + 321 + ], + "spans": [ + { + "bbox": [ + 272, + 263, + 348, + 321 + ], + "type": "image", + "image_path": "aab48792041743755ca62115a4c4922972a25ffec0420064182fff9ccf5cc5c4.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 298, + 323, + 321, + 330 + ], + "lines": [ + { + "bbox": [ + 298, + 323, + 321, + 330 + ], + "spans": [ + { + "bbox": [ + 298, + 323, + 321, + 330 + ], + "type": "text", + "content": "Bicubic" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 272, + 338, + 347, + 394 + ], + "blocks": [ + { + "bbox": [ + 287, + 330, + 333, + 337 + ], + "lines": [ + { + "bbox": [ + 287, + 330, + 333, + 337 + ], + "spans": [ + { + "bbox": [ + 287, + 330, + 333, + 337 + ], + "type": "text", + "content": "24.38dB/0.6872" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 272, + 338, + 347, + 394 + ], + "lines": [ + { + "bbox": [ + 272, + 338, + 347, + 394 + ], + "spans": [ + { + "bbox": [ + 272, + 338, + 347, + 394 + ], + "type": "image", + "image_path": "7015d250894c2c694b918b664e9cbb5a0fdb672a2a087fce115626fc1c18f96e.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 292, + 396, + 327, + 403 + ], + "lines": [ + { + "bbox": [ + 292, + 396, + 327, + 403 + ], + "spans": [ + { + "bbox": [ + 292, + 396, + 327, + 403 + ], + "type": "text", + "content": "SwinJR [23]" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 357, + 263, + 433, + 321 + ], + "blocks": [ + { + "bbox": [ + 357, + 263, + 433, + 321 + ], + "lines": [ + { + "bbox": [ + 357, + 263, + 433, + 321 + ], + "spans": [ + { + "bbox": [ + 357, + 263, + 433, + 321 + ], + "type": "image", + "image_path": "0acdf7d88bfcb6b4afc4eab6d1201a356f5846e79a01346926f7b16b25ac6c5c.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 378, + 322, + 411, + 329 + ], + "lines": [ + { + "bbox": [ + 378, + 322, + 411, + 329 + ], + "spans": [ + { + "bbox": [ + 378, + 322, + 411, + 329 + ], + "type": "text", + "content": "RCAN [43]" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 357, + 338, + 433, + 394 + ], + "blocks": [ + { + "bbox": [ + 371, + 330, + 417, + 337 + ], + "lines": [ + { + "bbox": [ + 371, + 330, + 417, + 337 + ], + "spans": [ + { + "bbox": [ + 371, + 330, + 417, + 337 + ], + "type": "text", + "content": "26.40dB/0.8137" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 357, + 338, + 433, + 394 + ], + "lines": [ + { + "bbox": [ + 357, + 338, + 433, + 394 + ], + "spans": [ + { + "bbox": [ + 357, + 338, + 433, + 394 + ], + "type": "image", + "image_path": "481744445b0f343e23c06bcce4ed9331562c5e7d171db802b6c79cd1ca07a558.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 373, + 395, + 415, + 403 + ], + "lines": [ + { + "bbox": [ + 373, + 395, + 415, + 403 + ], + "spans": [ + { + "bbox": [ + 373, + 395, + 415, + 403 + ], + "type": "text", + "content": "SwinIR† [23]" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 442, + 263, + 517, + 321 + ], + "blocks": [ + { + "bbox": [ + 442, + 263, + 517, + 321 + ], + "lines": [ + { + "bbox": [ + 442, + 263, + 517, + 321 + ], + "spans": [ + { + "bbox": [ + 442, + 263, + 517, + 321 + ], + "type": "image", + "image_path": "8477afb1d91308b13e0c72f6ce1f6ca8850546571d439b61c17ab62270f44d79.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 456, + 322, + 500, + 329 + ], + "lines": [ + { + "bbox": [ + 456, + 322, + 500, + 329 + ], + "spans": [ + { + "bbox": [ + 456, + 322, + 500, + 329 + ], + "type": "text", + "content": "SRResNet [36]" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 442, + 338, + 517, + 394 + ], + "blocks": [ + { + "bbox": [ + 456, + 330, + 502, + 337 + ], + "lines": [ + { + "bbox": [ + 456, + 330, + 502, + 337 + ], + "spans": [ + { + "bbox": [ + 456, + 330, + 502, + 337 + ], + "type": "text", + "content": "26.21dB/0.7999" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 442, + 338, + 517, + 394 + ], + "lines": [ + { + "bbox": [ + 442, + 338, + 517, + 394 + ], + "spans": [ + { + "bbox": [ + 442, + 338, + 517, + 394 + ], + "type": "image", + "image_path": "4815ee82c3c4c7985323b1e9e0aabe47ead9697211ab2aa4f21f51d45cb9ff57.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 468, + 396, + 490, + 404 + ], + "lines": [ + { + "bbox": [ + 468, + 396, + 490, + 404 + ], + "spans": [ + { + "bbox": [ + 468, + 396, + 490, + 404 + ], + "type": "text", + "content": "OSRT†" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 47, + 417, + 178, + 548 + ], + "blocks": [ + { + "bbox": [ + 47, + 417, + 178, + 548 + ], + "lines": [ + { + "bbox": [ + 47, + 417, + 178, + 548 + ], + "spans": [ + { + "bbox": [ + 47, + 417, + 178, + 548 + ], + "type": "image", + "image_path": "3bffb7b07707ac3d5c03dc456cbf1a86803f201e902c9c43ec23d3ff454e4205.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 85, + 554, + 142, + 561 + ], + "lines": [ + { + "bbox": [ + 85, + 554, + 142, + 561 + ], + "spans": [ + { + "bbox": [ + 85, + 554, + 142, + 561 + ], + "type": "text", + "content": "SUN360 " + }, + { + "bbox": [ + 85, + 554, + 142, + 561 + ], + "type": "inline_equation", + "content": "(\\times 4)" + }, + { + "bbox": [ + 85, + 554, + 142, + 561 + ], + "type": "text", + "content": " : 095" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 147, + 568, + 428, + 580 + ], + "lines": [ + { + "bbox": [ + 147, + 568, + 428, + 580 + ], + "spans": [ + { + "bbox": [ + 147, + 568, + 428, + 580 + ], + "type": "text", + "content": "Figure 6. Visual comparisons of " + }, + { + "bbox": [ + 147, + 568, + 428, + 580 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 147, + 568, + 428, + 580 + ], + "type": "text", + "content": " SR results under Fisheye downsampling." + } + ] + } + ], + "index": 54, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 188, + 417, + 263, + 474 + ], + "blocks": [ + { + "bbox": [ + 202, + 405, + 248, + 411 + ], + "lines": [ + { + "bbox": [ + 202, + 405, + 248, + 411 + ], + "spans": [ + { + "bbox": [ + 202, + 405, + 248, + 411 + ], + "type": "text", + "content": "26.38dB/0.8072" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 188, + 417, + 263, + 474 + ], + "lines": [ + { + "bbox": [ + 188, + 417, + 263, + 474 + ], + "spans": [ + { + "bbox": [ + 188, + 417, + 263, + 474 + ], + "type": "image", + "image_path": "d4f1a7a11ae19a88ec5eefd9b6e5bb209f898b01ae13657c1b6232d8ab83d34b.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 220, + 476, + 230, + 483 + ], + "lines": [ + { + "bbox": [ + 220, + 476, + 230, + 483 + ], + "spans": [ + { + "bbox": [ + 220, + 476, + 230, + 483 + ], + "type": "text", + "content": "HR" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 187, + 492, + 263, + 548 + ], + "blocks": [ + { + "bbox": [ + 207, + 484, + 243, + 490 + ], + "lines": [ + { + "bbox": [ + 207, + 484, + 243, + 490 + ], + "spans": [ + { + "bbox": [ + 207, + 484, + 243, + 490 + ], + "type": "text", + "content": "PSNR/SSIM" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 187, + 492, + 263, + 548 + ], + "lines": [ + { + "bbox": [ + 187, + 492, + 263, + 548 + ], + "spans": [ + { + "bbox": [ + 187, + 492, + 263, + 548 + ], + "type": "image", + "image_path": "1a1d5efe555b1f5035ecc94306435532e175fa4e6656e29c696c66d460bff762.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 209, + 550, + 241, + 557 + ], + "lines": [ + { + "bbox": [ + 209, + 550, + 241, + 557 + ], + "spans": [ + { + "bbox": [ + 209, + 550, + 241, + 557 + ], + "type": "text", + "content": "EDSR [24]" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 202, + 559, + 248, + 566 + ], + "lines": [ + { + "bbox": [ + 202, + 559, + 248, + 566 + ], + "spans": [ + { + "bbox": [ + 202, + 559, + 248, + 566 + ], + "type": "text", + "content": "33.64dB/0.9074" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 272, + 417, + 348, + 474 + ], + "blocks": [ + { + "bbox": [ + 285, + 405, + 334, + 411 + ], + "lines": [ + { + "bbox": [ + 285, + 405, + 334, + 411 + ], + "spans": [ + { + "bbox": [ + 285, + 405, + 334, + 411 + ], + "type": "text", + "content": "26.77dB/0.8234" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 272, + 417, + 348, + 474 + ], + "lines": [ + { + "bbox": [ + 272, + 417, + 348, + 474 + ], + "spans": [ + { + "bbox": [ + 272, + 417, + 348, + 474 + ], + "type": "image", + "image_path": "df19881ea35c55f6ef96db6fcb7f3ccacb62078b221a2cfdaab5f58e5a22304a.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 298, + 476, + 321, + 483 + ], + "lines": [ + { + "bbox": [ + 298, + 476, + 321, + 483 + ], + "spans": [ + { + "bbox": [ + 298, + 476, + 321, + 483 + ], + "type": "text", + "content": "Bicubic" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 272, + 492, + 348, + 548 + ], + "blocks": [ + { + "bbox": [ + 287, + 484, + 333, + 490 + ], + "lines": [ + { + "bbox": [ + 287, + 484, + 333, + 490 + ], + "spans": [ + { + "bbox": [ + 287, + 484, + 333, + 490 + ], + "type": "text", + "content": "30.20dB/0.8506" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 272, + 492, + 348, + 548 + ], + "lines": [ + { + "bbox": [ + 272, + 492, + 348, + 548 + ], + "spans": [ + { + "bbox": [ + 272, + 492, + 348, + 548 + ], + "type": "image", + "image_path": "c54e91b8db836f67d26c9e9067e41436ab9c905563cd77742ff61876f914e714.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 292, + 550, + 327, + 557 + ], + "lines": [ + { + "bbox": [ + 292, + 550, + 327, + 557 + ], + "spans": [ + { + "bbox": [ + 292, + 550, + 327, + 557 + ], + "type": "text", + "content": "SwinIR [23]" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 286, + 559, + 333, + 566 + ], + "lines": [ + { + "bbox": [ + 286, + 559, + 333, + 566 + ], + "spans": [ + { + "bbox": [ + 286, + 559, + 333, + 566 + ], + "type": "text", + "content": "34.05dB/0.9119" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_caption" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 356, + 417, + 433, + 474 + ], + "blocks": [ + { + "bbox": [ + 371, + 404, + 417, + 411 + ], + "lines": [ + { + "bbox": [ + 371, + 404, + 417, + 411 + ], + "spans": [ + { + "bbox": [ + 371, + 404, + 417, + 411 + ], + "type": "text", + "content": "27.34dB/0.8462" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 356, + 417, + 433, + 474 + ], + "lines": [ + { + "bbox": [ + 356, + 417, + 433, + 474 + ], + "spans": [ + { + "bbox": [ + 356, + 417, + 433, + 474 + ], + "type": "image", + "image_path": "1e3682281a265e04e93b2dba4f4cec276c1167bb85a93b03f68b256eea66dd1d.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 377, + 476, + 411, + 483 + ], + "lines": [ + { + "bbox": [ + 377, + 476, + 411, + 483 + ], + "spans": [ + { + "bbox": [ + 377, + 476, + 411, + 483 + ], + "type": "text", + "content": "RCAN [43]" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_caption" + } + ], + "index": 42 + }, + { + "type": "image", + "bbox": [ + 356, + 491, + 433, + 548 + ], + "blocks": [ + { + "bbox": [ + 372, + 484, + 417, + 490 + ], + "lines": [ + { + "bbox": [ + 372, + 484, + 417, + 490 + ], + "spans": [ + { + "bbox": [ + 372, + 484, + 417, + 490 + ], + "type": "text", + "content": "33.59dB/0.9088" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 356, + 491, + 433, + 548 + ], + "lines": [ + { + "bbox": [ + 356, + 491, + 433, + 548 + ], + "spans": [ + { + "bbox": [ + 356, + 491, + 433, + 548 + ], + "type": "image", + "image_path": "bdfb984ea97a83e001e31d4a1f23bc3afb561016c77dcf1ec9ce44a13b1e5345.jpg" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 374, + 550, + 414, + 557 + ], + "lines": [ + { + "bbox": [ + 374, + 550, + 414, + 557 + ], + "spans": [ + { + "bbox": [ + 374, + 550, + 414, + 557 + ], + "type": "text", + "content": "SwinIR [23]" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 371, + 559, + 417, + 565 + ], + "lines": [ + { + "bbox": [ + 371, + 559, + 417, + 565 + ], + "spans": [ + { + "bbox": [ + 371, + 559, + 417, + 565 + ], + "type": "text", + "content": "34.41dB/0.9158" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_caption" + } + ], + "index": 45 + }, + { + "type": "image", + "bbox": [ + 442, + 417, + 517, + 474 + ], + "blocks": [ + { + "bbox": [ + 456, + 404, + 502, + 411 + ], + "lines": [ + { + "bbox": [ + 456, + 404, + 502, + 411 + ], + "spans": [ + { + "bbox": [ + 456, + 404, + 502, + 411 + ], + "type": "text", + "content": "27.68dB/0.8561" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 442, + 417, + 517, + 474 + ], + "lines": [ + { + "bbox": [ + 442, + 417, + 517, + 474 + ], + "spans": [ + { + "bbox": [ + 442, + 417, + 517, + 474 + ], + "type": "image", + "image_path": "1448538722ba337494966ef75c9fbf0d4bc040e329acc5140d8213ccd5bff992.jpg" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 456, + 476, + 500, + 483 + ], + "lines": [ + { + "bbox": [ + 456, + 476, + 500, + 483 + ], + "spans": [ + { + "bbox": [ + 456, + 476, + 500, + 483 + ], + "type": "text", + "content": "SRResNet [36]" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_caption" + } + ], + "index": 48 + }, + { + "type": "image", + "bbox": [ + 442, + 491, + 517, + 548 + ], + "blocks": [ + { + "bbox": [ + 456, + 484, + 502, + 490 + ], + "lines": [ + { + "bbox": [ + 456, + 484, + 502, + 490 + ], + "spans": [ + { + "bbox": [ + 456, + 484, + 502, + 490 + ], + "type": "text", + "content": "33.43dB/0.9043" + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 442, + 491, + 517, + 548 + ], + "lines": [ + { + "bbox": [ + 442, + 491, + 517, + 548 + ], + "spans": [ + { + "bbox": [ + 442, + 491, + 517, + 548 + ], + "type": "image", + "image_path": "b350bdce5d2de18e8b6e05b88ebcdc8e43d8ae5cadca5f2dfd3f96e718b7dd4c.jpg" + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 467, + 550, + 490, + 557 + ], + "lines": [ + { + "bbox": [ + 467, + 550, + 490, + 557 + ], + "spans": [ + { + "bbox": [ + 467, + 550, + 490, + 557 + ], + "type": "text", + "content": "OSRT†" + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 456, + 559, + 502, + 565 + ], + "lines": [ + { + "bbox": [ + 456, + 559, + 502, + 565 + ], + "spans": [ + { + "bbox": [ + 456, + 559, + 502, + 565 + ], + "type": "text", + "content": "34.77dB/0.9187" + } + ] + } + ], + "index": 53, + "angle": 0, + "type": "image_caption" + } + ], + "index": 51 + }, + { + "bbox": [ + 39, + 588, + 120, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 39, + 588, + 120, + 601 + ], + "spans": [ + { + "bbox": [ + 39, + 588, + 120, + 601 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 55 + }, + { + "bbox": [ + 39, + 610, + 154, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 39, + 610, + 154, + 623 + ], + "spans": [ + { + "bbox": [ + 39, + 610, + 154, + 623 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 56 + }, + { + "bbox": [ + 38, + 630, + 279, + 738 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 630, + 279, + 738 + ], + "spans": [ + { + "bbox": [ + 38, + 630, + 279, + 738 + ], + "type": "text", + "content": "ODI-SR dataset [12] and SUN360 Panorama dataset [39] are used in our experiment. In the training phase, we follow the data split setting in [12] and train on the ODI-SR training set. The resolution of the ERP HR is " + }, + { + "bbox": [ + 38, + 630, + 279, + 738 + ], + "type": "inline_equation", + "content": "1024 \\times 2048" + }, + { + "bbox": [ + 38, + 630, + 279, + 738 + ], + "type": "text", + "content": ", and the upscaling factors are " + }, + { + "bbox": [ + 38, + 630, + 279, + 738 + ], + "type": "inline_equation", + "content": "\\times 2" + }, + { + "bbox": [ + 38, + 630, + 279, + 738 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 38, + 630, + 279, + 738 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 38, + 630, + 279, + 738 + ], + "type": "text", + "content": ". Fisheye downsampling is applied as our pre-defined downsampling kernel. Loss is calculated by L1 distance and optimized by Adam [19], with an initial learning rate of " + }, + { + "bbox": [ + 38, + 630, + 279, + 738 + ], + "type": "inline_equation", + "content": "2 \\times 10^{-4}" + }, + { + "bbox": [ + 38, + 630, + 279, + 738 + ], + "type": "text", + "content": ", a total batch size of 32, and an input patch size of 64. We" + } + ] + } + ], + "index": 57 + }, + { + "bbox": [ + 296, + 589, + 537, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 589, + 537, + 650 + ], + "spans": [ + { + "bbox": [ + 296, + 589, + 537, + 650 + ], + "type": "text", + "content": "train OSRT for " + }, + { + "bbox": [ + 296, + 589, + 537, + 650 + ], + "type": "inline_equation", + "content": "500k" + }, + { + "bbox": [ + 296, + 589, + 537, + 650 + ], + "type": "text", + "content": " iterations and halve the learning rate at " + }, + { + "bbox": [ + 296, + 589, + 537, + 650 + ], + "type": "inline_equation", + "content": "250k" + }, + { + "bbox": [ + 296, + 589, + 537, + 650 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 296, + 589, + 537, + 650 + ], + "type": "inline_equation", + "content": "400k" + }, + { + "bbox": [ + 296, + 589, + 537, + 650 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 296, + 589, + 537, + 650 + ], + "type": "inline_equation", + "content": "450k" + }, + { + "bbox": [ + 296, + 589, + 537, + 650 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 296, + 589, + 537, + 650 + ], + "type": "inline_equation", + "content": "475k" + }, + { + "bbox": [ + 296, + 589, + 537, + 650 + ], + "type": "text", + "content": ". In evaluation, we test on the ODI-SR testing set and SUN360 dataset. PSNR [16], SSIM [37], and their distortion re-weighted versions (WS-PSNR [31], WS-SSIM [45]) are used as evaluation metrics." + } + ] + } + ], + "index": 58 + }, + { + "bbox": [ + 297, + 659, + 513, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 659, + 513, + 672 + ], + "spans": [ + { + "bbox": [ + 297, + 659, + 513, + 672 + ], + "type": "text", + "content": "4.2. Evaluation under Fisheye Downsampling" + } + ] + } + ], + "index": 59 + }, + { + "bbox": [ + 296, + 678, + 537, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 678, + 537, + 714 + ], + "spans": [ + { + "bbox": [ + 296, + 678, + 537, + 714 + ], + "type": "text", + "content": "When the downsampling process is performed on Fisheye images, we train SRResNet [36], EDSR [24], RCAN [43], and SwinIR [23] for comparison." + } + ] + } + ], + "index": 60 + }, + { + "bbox": [ + 296, + 714, + 537, + 738 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 714, + 537, + 738 + ], + "spans": [ + { + "bbox": [ + 296, + 714, + 537, + 738 + ], + "type": "text", + "content": "Quantitative results. As shown in Tab. 1, with the help of additional DF2K-ERP training patches, OSRT out" + } + ] + } + ], + "index": 61 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 277, + 751, + 300, + 760 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 751, + 300, + 760 + ], + "spans": [ + { + "bbox": [ + 277, + 751, + 300, + 760 + ], + "type": "text", + "content": "13288" + } + ] + } + ], + "index": 62 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 92, + 95, + 483, + 201 + ], + "blocks": [ + { + "bbox": [ + 92, + 95, + 483, + 201 + ], + "lines": [ + { + "bbox": [ + 92, + 95, + 483, + 201 + ], + "spans": [ + { + "bbox": [ + 92, + 95, + 483, + 201 + ], + "type": "table", + "html": "
Scale×8×16
MethodODI-SRSUN 360 PanoramaODI-SRSUN 360 Panorama
WS-PSNRWS-SSIMWS-PSNRWS-SSIMWS-PSNRWS-SSIMWS-PSNRWS-SSIM
Bicubic19.640.590819.720.540317.120.433217.560.4638
SRCNN [13]20.080.611219.460.570118.080.450117.950.4684
EDSR [24]23.970.641722.460.634121.120.569821.060.5645
RCAN [43]24.260.662823.880.654221.940.582421.740.5742
360-SS [30]21.650.641721.480.635219.650.543119.620.5308
LAU-Net [12]24.360.680124.020.670822.070.590121.820.5824
SphereSR [40]24.370.677724.170.682022.510.637021.950.6342
OSRT24.530.678024.380.707222.690.626122.130.6388
", + "image_path": "9a33e0913f340d6848a3d7a1d728d0fe6e853a0283c0a1d970c86bc8b557e6d5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 38, + 222, + 279, + 318 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 222, + 279, + 318 + ], + "spans": [ + { + "bbox": [ + 38, + 222, + 279, + 318 + ], + "type": "text", + "content": "performs previous methods by 0.3dB on PSNR. Although directly applying SwinIR on the ODISR task has already reached SOTA performance, OSRT surpasses SwinIR over 0.1dB on two datasets for both " + }, + { + "bbox": [ + 38, + 222, + 279, + 318 + ], + "type": "inline_equation", + "content": "\\times 2" + }, + { + "bbox": [ + 38, + 222, + 279, + 318 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 38, + 222, + 279, + 318 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 38, + 222, + 279, + 318 + ], + "type": "text", + "content": " SR tasks, which demonstrates the effectiveness of its distortion modulation ability. The performance of RCAN degrades under Fisheye downsampling, which is caused by the incompatibility between channel attention and Fisheye downsampling5." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 38, + 318, + 279, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 318, + 279, + 426 + ], + "spans": [ + { + "bbox": [ + 38, + 318, + 279, + 426 + ], + "type": "text", + "content": "Qualitative comparison. Fig. 6 shows the visualization results of " + }, + { + "bbox": [ + 38, + 318, + 279, + 426 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 38, + 318, + 279, + 426 + ], + "type": "text", + "content": " ODISR task. While other methods struggle to understand the geometric transformation process in distorted images, OSRT can reconstruct sharp and accurate boundaries with the advantages of distortion modulation. It is observed that OSRT is skilled at reconstructing rigid texture. Moreover, benefiting from the distortion modulation ability, OSRT can preserve the original structure as most when being projected to other projection types (Fig. 7)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 38, + 434, + 241, + 447 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 434, + 241, + 447 + ], + "spans": [ + { + "bbox": [ + 38, + 434, + 241, + 447 + ], + "type": "text", + "content": "4.3. Evaluation under ERP Downsampling" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 38, + 452, + 279, + 560 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 452, + 279, + 560 + ], + "spans": [ + { + "bbox": [ + 38, + 452, + 279, + 560 + ], + "type": "text", + "content": "To compare with previous ODISR methods [12, 30, 40], we train OSRT under the previous ERP setting. Regardless of over-fitting issues, we only train on the dataset provided by [12] for fairness. As shown in Tab. 2, OSRT still outperforms LAU-Net [12] and SphereSR [40] under large upscaling factor and ERP downsampling. Without a complicated training pipeline and discrete inference process, OSRT yields the best PSNR values and surpasses all previous methods on most SSIM-related metrics (three of four)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 38, + 569, + 205, + 581 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 569, + 205, + 581 + ], + "spans": [ + { + "bbox": [ + 38, + 569, + 205, + 581 + ], + "type": "text", + "content": "4.4. Ablation Study and Discussion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 38, + 587, + 279, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 587, + 279, + 635 + ], + "spans": [ + { + "bbox": [ + 38, + 587, + 279, + 635 + ], + "type": "text", + "content": "In this section, we prove the effectiveness of Fisheye downsampling, OSRT components, and augmented DF2K-ERP. We then explain the distortion modulation ability of OSRT by visualizing offsets in deformable blocks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 38, + 635, + 279, + 719 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 635, + 279, + 719 + ], + "spans": [ + { + "bbox": [ + 38, + 635, + 279, + 719 + ], + "type": "text", + "content": "Fisheye downsampling. As shown in Fig. 1, the SR model trained under ERP downsampling is more likely to generate blur details and missing structures in real-world scenarios. These artifacts cannot be removed by a superior backbone network, but can be eliminated by a more realistic imaging process. More importantly, ERP downsampling directly covers the geometric property of ERP images" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 301, + 220, + 530, + 296 + ], + "blocks": [ + { + "bbox": [ + 203, + 203, + 371, + 215 + ], + "lines": [ + { + "bbox": [ + 203, + 203, + 371, + 215 + ], + "spans": [ + { + "bbox": [ + 203, + 203, + 371, + 215 + ], + "type": "text", + "content": "Table 2. SR results under ERP downsampling." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 301, + 220, + 530, + 296 + ], + "lines": [ + { + "bbox": [ + 301, + 220, + 530, + 296 + ], + "spans": [ + { + "bbox": [ + 301, + 220, + 530, + 296 + ], + "type": "table", + "html": "
feature dimDACBDAABODI-SRSUN360Params. (M)
PSNRSSIMPSNRSSIM
60××30.270.873930.780.87420.91
60✓×30.410.877531.000.87931.16
60×w/o Cw30.310.874630.830.87551.00
60×w/ Cw30.320.874630.840.87531.01
60✓w/ Cw30.440.878031.040.88001.26
72××30.320.874830.850.87551.29
", + "image_path": "3cf94c84f54a3a4b5ff65cc77204388439d79b1a4cdf4eeb5d2eec389e6fffff.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 296, + 299, + 536, + 322 + ], + "lines": [ + { + "bbox": [ + 296, + 299, + 536, + 322 + ], + "spans": [ + { + "bbox": [ + 296, + 299, + 536, + 322 + ], + "type": "text", + "content": "Table 3. Ablation study on OSRT components. All models are trained on " + }, + { + "bbox": [ + 296, + 299, + 536, + 322 + ], + "type": "inline_equation", + "content": "\\times 2" + }, + { + "bbox": [ + 296, + 299, + 536, + 322 + ], + "type": "text", + "content": " SR task under Fisheye downsampling." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 296, + 324, + 537, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 324, + 537, + 443 + ], + "spans": [ + { + "bbox": [ + 296, + 324, + 537, + 443 + ], + "type": "text", + "content": "and makes the ODISR task identical to the standard plain image super-resolution task. The evidence is that a standard SISR model (SwinIR) trained on a plain image dataset (DF2K) can outperform previous SOTA in the ODISR task, which yields WS-PSNR results of 24.63dB/24.49dB (22.68dB/22.13dB) on " + }, + { + "bbox": [ + 296, + 324, + 537, + 443 + ], + "type": "inline_equation", + "content": "\\times 8" + }, + { + "bbox": [ + 296, + 324, + 537, + 443 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 296, + 324, + 537, + 443 + ], + "type": "inline_equation", + "content": "\\times 16" + }, + { + "bbox": [ + 296, + 324, + 537, + 443 + ], + "type": "text", + "content": ") ODI-SR/SUN360 testing set, respectively. In conclusion, when the intrinsic property of ODIs is broken by ERP downsampling, the ODISR task degenerates into a plain image super-resolution task with a particular data distribution." + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 299, + 445, + 411, + 557 + ], + "blocks": [ + { + "bbox": [ + 299, + 445, + 411, + 557 + ], + "lines": [ + { + "bbox": [ + 299, + 445, + 411, + 557 + ], + "spans": [ + { + "bbox": [ + 299, + 445, + 411, + 557 + ], + "type": "image", + "image_path": "d263c9dc37bd87f334f11c9005cdf5c152dd27cca0279982cf36f6cb5dda0676.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 320, + 558, + 391, + 575 + ], + "lines": [ + { + "bbox": [ + 320, + 558, + 391, + 575 + ], + "spans": [ + { + "bbox": [ + 320, + 558, + 391, + 575 + ], + "type": "text", + "content": "ODI-SR " + }, + { + "bbox": [ + 320, + 558, + 391, + 575 + ], + "type": "inline_equation", + "content": "(\\times 4)" + }, + { + "bbox": [ + 320, + 558, + 391, + 575 + ], + "type": "text", + "content": " :003 Fisheye (Vertical, Right)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 415, + 445, + 473, + 503 + ], + "blocks": [ + { + "bbox": [ + 415, + 445, + 473, + 503 + ], + "lines": [ + { + "bbox": [ + 415, + 445, + 473, + 503 + ], + "spans": [ + { + "bbox": [ + 415, + 445, + 473, + 503 + ], + "type": "image", + "image_path": "347399a281d81a2d0d22053c52bed700fd8cfce33057400626ea2f8cc673852b.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 428, + 503, + 460, + 511 + ], + "lines": [ + { + "bbox": [ + 428, + 503, + 460, + 511 + ], + "spans": [ + { + "bbox": [ + 428, + 503, + 460, + 511 + ], + "type": "text", + "content": "EDSR [24]" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 475, + 446, + 532, + 502 + ], + "blocks": [ + { + "bbox": [ + 475, + 446, + 532, + 502 + ], + "lines": [ + { + "bbox": [ + 475, + 446, + 532, + 502 + ], + "spans": [ + { + "bbox": [ + 475, + 446, + 532, + 502 + ], + "type": "image", + "image_path": "8aa834e957f53eb38f677568c4bf692307999de4ef356c7c58422c3c26310c02.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 415, + 511, + 473, + 567 + ], + "blocks": [ + { + "bbox": [ + 415, + 511, + 473, + 567 + ], + "lines": [ + { + "bbox": [ + 415, + 511, + 473, + 567 + ], + "spans": [ + { + "bbox": [ + 415, + 511, + 473, + 567 + ], + "type": "image", + "image_path": "9963fa8d1acfcc5a6d149a19bfb179db69dd6d77b260bc2ca6ba3bdc5faf8c48.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 424, + 568, + 464, + 577 + ], + "lines": [ + { + "bbox": [ + 424, + 568, + 464, + 577 + ], + "spans": [ + { + "bbox": [ + 424, + 568, + 464, + 577 + ], + "type": "text", + "content": "SwinIR† [23]" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 475, + 511, + 532, + 567 + ], + "blocks": [ + { + "bbox": [ + 487, + 503, + 521, + 511 + ], + "lines": [ + { + "bbox": [ + 487, + 503, + 521, + 511 + ], + "spans": [ + { + "bbox": [ + 487, + 503, + 521, + 511 + ], + "type": "text", + "content": "SwinIR [23]" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 475, + 511, + 532, + 567 + ], + "lines": [ + { + "bbox": [ + 475, + 511, + 532, + 567 + ], + "spans": [ + { + "bbox": [ + 475, + 511, + 532, + 567 + ], + "type": "image", + "image_path": "ec3d96c5c427240db388b5eacae55786ddb7b31ba10c483e2efd6dd8118ce7c0.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 492, + 568, + 514, + 577 + ], + "lines": [ + { + "bbox": [ + 492, + 568, + 514, + 577 + ], + "spans": [ + { + "bbox": [ + 492, + 568, + 514, + 577 + ], + "type": "text", + "content": "OSRT†" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 300, + 579, + 411, + 690 + ], + "blocks": [ + { + "bbox": [ + 300, + 579, + 411, + 690 + ], + "lines": [ + { + "bbox": [ + 300, + 579, + 411, + 690 + ], + "spans": [ + { + "bbox": [ + 300, + 579, + 411, + 690 + ], + "type": "image", + "image_path": "044705c461a57792277caae6ab00859e0349e3d4c68851aa8489c1f61b722fd8.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 309, + 691, + 402, + 708 + ], + "lines": [ + { + "bbox": [ + 309, + 691, + 402, + 708 + ], + "spans": [ + { + "bbox": [ + 309, + 691, + 402, + 708 + ], + "type": "text", + "content": "SUN360 " + }, + { + "bbox": [ + 309, + 691, + 402, + 708 + ], + "type": "inline_equation", + "content": "(\\times 2)" + }, + { + "bbox": [ + 309, + 691, + 402, + 708 + ], + "type": "text", + "content": " : 062 Perspective " + }, + { + "bbox": [ + 309, + 691, + 402, + 708 + ], + "type": "inline_equation", + "content": "(\\varphi :30^{\\circ}" + }, + { + "bbox": [ + 309, + 691, + 402, + 708 + ], + "type": "text", + "content": " ;FOV: " + }, + { + "bbox": [ + 309, + 691, + 402, + 708 + ], + "type": "inline_equation", + "content": "90^{\\circ}" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 415, + 579, + 472, + 635 + ], + "blocks": [ + { + "bbox": [ + 415, + 579, + 472, + 635 + ], + "lines": [ + { + "bbox": [ + 415, + 579, + 472, + 635 + ], + "spans": [ + { + "bbox": [ + 415, + 579, + 472, + 635 + ], + "type": "image", + "image_path": "ab76630d7f4eb043277a7f12d79ac163f08a3ef4c81ee0291e596541b8c53988.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 415, + 644, + 472, + 700 + ], + "blocks": [ + { + "bbox": [ + 428, + 636, + 460, + 644 + ], + "lines": [ + { + "bbox": [ + 428, + 636, + 460, + 644 + ], + "spans": [ + { + "bbox": [ + 428, + 636, + 460, + 644 + ], + "type": "text", + "content": "EDSR [24]" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 415, + 644, + 472, + 700 + ], + "lines": [ + { + "bbox": [ + 415, + 644, + 472, + 700 + ], + "spans": [ + { + "bbox": [ + 415, + 644, + 472, + 700 + ], + "type": "image", + "image_path": "7765713b1be23316de722e73c0ce74cede6ff58d5d4247c9cc4212714f3e92b9.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 424, + 702, + 464, + 710 + ], + "lines": [ + { + "bbox": [ + 424, + 702, + 464, + 710 + ], + "spans": [ + { + "bbox": [ + 424, + 702, + 464, + 710 + ], + "type": "text", + "content": "SwinIR† [23]" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 475, + 579, + 532, + 635 + ], + "blocks": [ + { + "bbox": [ + 475, + 579, + 532, + 635 + ], + "lines": [ + { + "bbox": [ + 475, + 579, + 532, + 635 + ], + "spans": [ + { + "bbox": [ + 475, + 579, + 532, + 635 + ], + "type": "image", + "image_path": "5caee10a8ce95acf83a71b25eeb960c6896770a0a72d1c8dacffb3c246e1d105.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 475, + 644, + 532, + 701 + ], + "blocks": [ + { + "bbox": [ + 486, + 636, + 521, + 644 + ], + "lines": [ + { + "bbox": [ + 486, + 636, + 521, + 644 + ], + "spans": [ + { + "bbox": [ + 486, + 636, + 521, + 644 + ], + "type": "text", + "content": "SwinIR [23]" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 475, + 644, + 532, + 701 + ], + "lines": [ + { + "bbox": [ + 475, + 644, + 532, + 701 + ], + "spans": [ + { + "bbox": [ + 475, + 644, + 532, + 701 + ], + "type": "image", + "image_path": "159410d8b784f09b44c67d733c2fc6325a05f36296e01d3b0f733d950853a79b.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 492, + 702, + 514, + 710 + ], + "lines": [ + { + "bbox": [ + 492, + 702, + 514, + 710 + ], + "spans": [ + { + "bbox": [ + 492, + 702, + 514, + 710 + ], + "type": "text", + "content": "OSRT†" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 297, + 713, + 536, + 735 + ], + "lines": [ + { + "bbox": [ + 297, + 713, + 536, + 735 + ], + "spans": [ + { + "bbox": [ + 297, + 713, + 536, + 735 + ], + "type": "text", + "content": "Figure 7. Visual comparisons for SR of Fisheye and Perspective images. " + }, + { + "bbox": [ + 297, + 713, + 536, + 735 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 297, + 713, + 536, + 735 + ], + "type": "text", + "content": " denotes applying DF2K-ERP as augmented dataset." + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 49, + 727, + 211, + 738 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 727, + 211, + 738 + ], + "spans": [ + { + "bbox": [ + 49, + 727, + 211, + 738 + ], + "type": "text", + "content": "5The cause is discussed in the supplementary file." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 277, + 751, + 300, + 760 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 751, + 300, + 760 + ], + "spans": [ + { + "bbox": [ + 277, + 751, + 300, + 760 + ], + "type": "text", + "content": "13289" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 97, + 135, + 165 + ], + "blocks": [ + { + "bbox": [ + 43, + 123, + 65, + 141 + ], + "lines": [ + { + "bbox": [ + 43, + 123, + 65, + 141 + ], + "spans": [ + { + "bbox": [ + 43, + 123, + 65, + 141 + ], + "type": "text", + "content": "DAAB (29-th)" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 67, + 97, + 135, + 165 + ], + "lines": [ + { + "bbox": [ + 67, + 97, + 135, + 165 + ], + "spans": [ + { + "bbox": [ + 67, + 97, + 135, + 165 + ], + "type": "image", + "image_path": "c3cf00c8f3d468d3e2c7a3efbf4bd7a8c653e5e8e31469311d32541521f07351.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 138, + 98, + 205, + 165 + ], + "blocks": [ + { + "bbox": [ + 138, + 98, + 205, + 165 + ], + "lines": [ + { + "bbox": [ + 138, + 98, + 205, + 165 + ], + "spans": [ + { + "bbox": [ + 138, + 98, + 205, + 165 + ], + "type": "image", + "image_path": "aaf70f510f46675cf63b9135c08d88c8212849ebe829cdaca705a4dadd8456ea.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 158, + 166, + 182, + 172 + ], + "lines": [ + { + "bbox": [ + 158, + 166, + 182, + 172 + ], + "spans": [ + { + "bbox": [ + 158, + 166, + 182, + 172 + ], + "type": "inline_equation", + "content": "\\varphi = 45^{\\circ}" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 209, + 98, + 276, + 165 + ], + "blocks": [ + { + "bbox": [ + 209, + 98, + 276, + 165 + ], + "lines": [ + { + "bbox": [ + 209, + 98, + 276, + 165 + ], + "spans": [ + { + "bbox": [ + 209, + 98, + 276, + 165 + ], + "type": "image", + "image_path": "17e28f2558a2bba12c2cedbe02d22d7006226271a10f0413a0f789e57c345074.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 229, + 166, + 254, + 173 + ], + "lines": [ + { + "bbox": [ + 229, + 166, + 254, + 173 + ], + "spans": [ + { + "bbox": [ + 229, + 166, + 254, + 173 + ], + "type": "inline_equation", + "content": "\\varphi = 90^{\\circ}" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 67, + 174, + 135, + 241 + ], + "blocks": [ + { + "bbox": [ + 89, + 166, + 109, + 172 + ], + "lines": [ + { + "bbox": [ + 89, + 166, + 109, + 172 + ], + "spans": [ + { + "bbox": [ + 89, + 166, + 109, + 172 + ], + "type": "inline_equation", + "content": "\\omega = 0^{\\circ}" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 43, + 202, + 65, + 218 + ], + "lines": [ + { + "bbox": [ + 43, + 202, + 65, + 218 + ], + "spans": [ + { + "bbox": [ + 43, + 202, + 65, + 218 + ], + "type": "text", + "content": "DAAB (33-th)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 67, + 174, + 135, + 241 + ], + "lines": [ + { + "bbox": [ + 67, + 174, + 135, + 241 + ], + "spans": [ + { + "bbox": [ + 67, + 174, + 135, + 241 + ], + "type": "image", + "image_path": "45e10ebd222dcf2a24dc360e989bc54ed34567825a2779a634b800c19babe609.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 138, + 174, + 205, + 241 + ], + "blocks": [ + { + "bbox": [ + 138, + 174, + 205, + 241 + ], + "lines": [ + { + "bbox": [ + 138, + 174, + 205, + 241 + ], + "spans": [ + { + "bbox": [ + 138, + 174, + 205, + 241 + ], + "type": "image", + "image_path": "1d82e6dc3dc63cc5f2405848b899de9664c0e7d2b89e8b2cdc5fa89a40e3ec2e.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 209, + 174, + 276, + 241 + ], + "blocks": [ + { + "bbox": [ + 209, + 174, + 276, + 241 + ], + "lines": [ + { + "bbox": [ + 209, + 174, + 276, + 241 + ], + "spans": [ + { + "bbox": [ + 209, + 174, + 276, + 241 + ], + "type": "image", + "image_path": "8b703d5dee9b0a31f011e5c6c872d5cbce0c5fba1282cf61e05e732ca49178a0.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 67, + 251, + 117, + 301 + ], + "blocks": [ + { + "bbox": [ + 89, + 243, + 110, + 250 + ], + "lines": [ + { + "bbox": [ + 89, + 243, + 110, + 250 + ], + "spans": [ + { + "bbox": [ + 89, + 243, + 110, + 250 + ], + "type": "inline_equation", + "content": "\\varphi = 0^{\\circ}" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 44, + 272, + 64, + 280 + ], + "lines": [ + { + "bbox": [ + 44, + 272, + 64, + 280 + ], + "spans": [ + { + "bbox": [ + 44, + 272, + 64, + 280 + ], + "type": "text", + "content": "DACB" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 67, + 251, + 117, + 301 + ], + "lines": [ + { + "bbox": [ + 67, + 251, + 117, + 301 + ], + "spans": [ + { + "bbox": [ + 67, + 251, + 117, + 301 + ], + "type": "image", + "image_path": "17ce58425cfcda28af96cdc102bcb82b07e11f6257b1859af081fb33bdad73f4.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 80, + 303, + 100, + 309 + ], + "lines": [ + { + "bbox": [ + 80, + 303, + 100, + 309 + ], + "spans": [ + { + "bbox": [ + 80, + 303, + 100, + 309 + ], + "type": "inline_equation", + "content": "\\varphi = 0^{\\circ}" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 38, + 316, + 279, + 350 + ], + "lines": [ + { + "bbox": [ + 38, + 316, + 279, + 350 + ], + "spans": [ + { + "bbox": [ + 38, + 316, + 279, + 350 + ], + "type": "text", + "content": "Figure 8. Visualizations of offset maps in OSRT. Reference and deformed points are depicted in green and red, respectively. The deformable kernel is sparse in the polar area." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 120, + 251, + 170, + 301 + ], + "blocks": [ + { + "bbox": [ + 158, + 243, + 182, + 250 + ], + "lines": [ + { + "bbox": [ + 158, + 243, + 182, + 250 + ], + "spans": [ + { + "bbox": [ + 158, + 243, + 182, + 250 + ], + "type": "inline_equation", + "content": "\\varphi = 45^{\\circ}" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 120, + 251, + 170, + 301 + ], + "lines": [ + { + "bbox": [ + 120, + 251, + 170, + 301 + ], + "spans": [ + { + "bbox": [ + 120, + 251, + 170, + 301 + ], + "type": "image", + "image_path": "b197eba21c2d9107a864f5589706d7516219b43891386c3ca6eb8d3e277022d3.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 303, + 157, + 309 + ], + "lines": [ + { + "bbox": [ + 132, + 303, + 157, + 309 + ], + "spans": [ + { + "bbox": [ + 132, + 303, + 157, + 309 + ], + "type": "inline_equation", + "content": "\\varphi = 30^{\\circ}" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 173, + 251, + 223, + 301 + ], + "blocks": [ + { + "bbox": [ + 173, + 251, + 223, + 301 + ], + "lines": [ + { + "bbox": [ + 173, + 251, + 223, + 301 + ], + "spans": [ + { + "bbox": [ + 173, + 251, + 223, + 301 + ], + "type": "image", + "image_path": "4db4315bd083d2575e395724e124f4e6134db2cb434b7d98c625c52210a3e8e1.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 185, + 303, + 209, + 309 + ], + "lines": [ + { + "bbox": [ + 185, + 303, + 209, + 309 + ], + "spans": [ + { + "bbox": [ + 185, + 303, + 209, + 309 + ], + "type": "inline_equation", + "content": "\\varphi = 60^{\\circ}" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 227, + 251, + 276, + 301 + ], + "blocks": [ + { + "bbox": [ + 229, + 243, + 254, + 250 + ], + "lines": [ + { + "bbox": [ + 229, + 243, + 254, + 250 + ], + "spans": [ + { + "bbox": [ + 229, + 243, + 254, + 250 + ], + "type": "inline_equation", + "content": "\\varphi = 90^{\\circ}" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 227, + 251, + 276, + 301 + ], + "lines": [ + { + "bbox": [ + 227, + 251, + 276, + 301 + ], + "spans": [ + { + "bbox": [ + 227, + 251, + 276, + 301 + ], + "type": "image", + "image_path": "7ebc60cf1b6d62fb9d134d64abd88daba08c92ed823a05a7d574caf9b86f4609.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 238, + 303, + 262, + 309 + ], + "lines": [ + { + "bbox": [ + 238, + 303, + 262, + 309 + ], + "spans": [ + { + "bbox": [ + 238, + 303, + 262, + 309 + ], + "type": "inline_equation", + "content": "\\varphi = 90^{\\circ}" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "bbox": [ + 38, + 357, + 278, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 357, + 278, + 476 + ], + "spans": [ + { + "bbox": [ + 38, + 357, + 278, + 476 + ], + "type": "text", + "content": "OSRT components. To study the effectiveness of each component in OSRT, we propose a light version of OSRT (OSRT-light) for ablation study, which corresponds with the official SwinIR-light [23]. As proofed in Tab. 3, all components in OSRT are beneficial for modulating ERP distortion. The advantages of DACB and DAAB can be stacked when being applied in the same network. Compared with simply expanding the feature dimension of SwinIR to match the network complexity, the overall improvements of OSRT is more significant " + }, + { + "bbox": [ + 38, + 357, + 278, + 476 + ], + "type": "inline_equation", + "content": "(+0.05\\mathrm{dB}" + }, + { + "bbox": [ + 38, + 357, + 278, + 476 + ], + "type": "text", + "content": " vs. " + }, + { + "bbox": [ + 38, + 357, + 278, + 476 + ], + "type": "inline_equation", + "content": "+0.2\\mathrm{dB})" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 38, + 479, + 279, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 38, + 479, + 279, + 538 + ], + "spans": [ + { + "bbox": [ + 38, + 479, + 279, + 538 + ], + "type": "text", + "content": "Offsets in OSRT. Offsets map in a well-trained OSRT are visualized in Fig. 8. Deformable kernels in both DAAB and DACB tend to gather at the equator and scatter at the polar, which conforms to the geometric distribution of pixel density in ERP images." + } + ] + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 40, + 548, + 279, + 710 + ], + "blocks": [ + { + "bbox": [ + 40, + 548, + 279, + 710 + ], + "lines": [ + { + "bbox": [ + 40, + 548, + 279, + 710 + ], + "spans": [ + { + "bbox": [ + 40, + 548, + 279, + 710 + ], + "type": "image", + "image_path": "86d33b68e48f5512cb07c309a6291a07500b6c66226197bbb47b164d05215df9.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 38, + 713, + 278, + 735 + ], + "lines": [ + { + "bbox": [ + 38, + 713, + 278, + 735 + ], + "spans": [ + { + "bbox": [ + 38, + 713, + 278, + 735 + ], + "type": "text", + "content": "Figure 9. Training process of Transformers on " + }, + { + "bbox": [ + 38, + 713, + 278, + 735 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 38, + 713, + 278, + 735 + ], + "type": "text", + "content": " ODISR task. The overfitting issue is tackled by our augmentation scheme." + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "table", + "bbox": [ + 307, + 95, + 526, + 207 + ], + "blocks": [ + { + "bbox": [ + 307, + 95, + 526, + 207 + ], + "lines": [ + { + "bbox": [ + 307, + 95, + 526, + 207 + ], + "spans": [ + { + "bbox": [ + 307, + 95, + 526, + 207 + ], + "type": "table", + "html": "
Backbone networkDatasetsTraining schemeScaleSUN360
PSNRSSIM
SwinIRODI-SRN/A31.210.8852
SwinIRDF2K/ODI-SRone-stage31.260.8841
SwinIRDF2K-ERP/ODI-SRone-stage×231.330.8855
SwinIRDF2K-ERP/ODI-SRtwo-stage31.170.8818
OSRTDF2K-ERP/ODI-SRone-stage31.520.8888
SwinIRODI-SRN/A27.390.7707
SwinIRDF2K/ODI-SRone-stage27.590.7768
SwinIRDF2K-ERP/ODI-SRone-stage×427.710.7804
SwinIRDF2K-ERP/ODI-SRtwo-stage27.740.7795
OSRTDF2K-ERP/ODI-SRone-stage27.840.7835
", + "image_path": "177e1bd4b84fbb0d117f1efcfb20da93a91db993aaf1a7694baf1b25a2e70697.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "table_body" + } + ], + "index": 28 + }, + { + "bbox": [ + 297, + 208, + 536, + 231 + ], + "lines": [ + { + "bbox": [ + 297, + 208, + 536, + 231 + ], + "spans": [ + { + "bbox": [ + 297, + 208, + 536, + 231 + ], + "type": "text", + "content": "Table 4. Ablation study on data augmentation. The results of ODISR (In the supplementary file) are in the same trend as SUN360." + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 296, + 239, + 538, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 239, + 538, + 418 + ], + "spans": [ + { + "bbox": [ + 296, + 239, + 538, + 418 + ], + "type": "text", + "content": "Pseudo ERP patches. In Sec. 3.4, we propose a distorted dataset DF2K-ERP to tackle over-fitting issues. We train a standard SwinIR on diverse datasets and training schemes to study the influence of data augmentation separately. As shown in Tab. 4, while training on ODI-SR and DF2K, distortion operations in DF2K lead to better performance. Compared with fine-tuning on DF2K-ERP pretrained models (two-stage), training on two datasets jointly (one-stage) shows better results. We infer that there is a domain gap between ODI-SR and DF2K-ERP, which is caused by omitted Perspective distortion. Moreover, the advantage of distortion modulation mechanisms in OSRT is enlarged when additional training patches are applied. Fig. 9 proves that our data augmentation scheme overcomes the over-fitting issue and improves the reconstruction ability." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 298, + 430, + 370, + 442 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 430, + 370, + 442 + ], + "spans": [ + { + "bbox": [ + 298, + 430, + 370, + 442 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 296, + 450, + 537, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 450, + 537, + 629 + ], + "spans": [ + { + "bbox": [ + 296, + 450, + 537, + 629 + ], + "type": "text", + "content": "In this paper, we find that the previous downsampling process in the ODISR task harms the intrinsic distribution of pixel density in ODIs, which leads to poor generalization ability in real-world scenarios. To tackle this issue, we propose Fisheye downsampling, which mimics the real-world imaging process to preserve the realistic density distribution. After refining the downsampling process, we design a distortion-aware Transformer (OSRT) to modulate distortions continuously and self-adaptively. OSRT learns offsets from the distortion-related condition and rectifies distortion by feature-level warping. Moreover, to alleviate the overfitting problem of large networks, we propose to synthesize additional ERP training data from the plain images. Extensive experiments have demonstrated the state-of-the-art performance of our OSRT." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 297, + 630, + 536, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 630, + 536, + 654 + ], + "spans": [ + { + "bbox": [ + 297, + 630, + 536, + 654 + ], + "type": "text", + "content": "Limitation. The process of sampling ERP images into viewing types also requires careful design." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 309, + 655, + 393, + 666 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 655, + 393, + 666 + ], + "spans": [ + { + "bbox": [ + 309, + 655, + 393, + 666 + ], + "type": "text", + "content": "Acknowledgement." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 296, + 666, + 536, + 738 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 666, + 536, + 738 + ], + "spans": [ + { + "bbox": [ + 296, + 666, + 536, + 738 + ], + "type": "text", + "content": "This work was supported in part by the National Key R&D Program of China (NO.2022ZD0160505), in part by the National Natural Science Foundation of China under Grant (62276251), the Joint Lab of CAS-HK, and in part by the Youth Innovation Promotion Association of Chinese Academy of Sciences (No.2020356)." + } + ] + } + ], + "index": 35 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 277, + 751, + 301, + 761 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 751, + 301, + 761 + ], + "spans": [ + { + "bbox": [ + 277, + 751, + 301, + 761 + ], + "type": "text", + "content": "13290" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 39, + 95, + 98, + 108 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 39, + 95, + 98, + 108 + ], + "spans": [ + { + "bbox": [ + 39, + 95, + 98, + 108 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 40, + 116, + 279, + 738 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 44, + 116, + 279, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 44, + 116, + 279, + 160 + ], + "spans": [ + { + "bbox": [ + 44, + 116, + 279, + 160 + ], + "type": "text", + "content": "[1] Hao Ai, Zidong Cao, Jinjing Zhu, Haotian Bai, Yucheng Chen, and Ling Wang. Deep learning for omnidirectional vision: A survey and new perspectives. arXiv preprint arXiv:2205.10468, 2022. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 44, + 161, + 279, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 44, + 161, + 279, + 206 + ], + "spans": [ + { + "bbox": [ + 44, + 161, + 279, + 206 + ], + "type": "text", + "content": "[2] Zafer Arican and Pascal Frossard. L1 regularized superresolution from unregistered omnidirectional images. In 2009 IEEE International Conference on Acoustics, Speech and Signal Processing, pages 829-832. IEEE, 2009. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 207, + 279, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 207, + 279, + 240 + ], + "spans": [ + { + "bbox": [ + 45, + 207, + 279, + 240 + ], + "type": "text", + "content": "[3] Zafer Arican and Pascal Frossard. Joint registration and super-resolution with omnidirectional images. IEEE Transactions on Image Processing, 20(11):3151-3162, 2011. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 45, + 242, + 279, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 242, + 279, + 296 + ], + "spans": [ + { + "bbox": [ + 45, + 242, + 279, + 296 + ], + "type": "text", + "content": "[4] Luigi Bagnato, Yannick Boursier, Pascal Frossard, and Pierre Vandergheynst. Plenoptic based super-resolution for omnidirectional image sequences. In 2010 IEEE International Conference on Image Processing, pages 2829-2832. IEEE, 2010. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 45, + 298, + 279, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 298, + 279, + 353 + ], + "spans": [ + { + "bbox": [ + 45, + 298, + 279, + 353 + ], + "type": "text", + "content": "[5] Kelvin CK Chan, Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Basicvsr: The search for essential components in video super-resolution and beyond. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4947-4956, 2021. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 354, + 278, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 354, + 278, + 410 + ], + "spans": [ + { + "bbox": [ + 45, + 354, + 278, + 410 + ], + "type": "text", + "content": "[6] Kelvin CK Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. Basicvsr++: Improving video superresolution with enhanced propagation and alignment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5972-5981, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 45, + 410, + 279, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 410, + 279, + 466 + ], + "spans": [ + { + "bbox": [ + 45, + 410, + 279, + 466 + ], + "type": "text", + "content": "[7] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12299-12310, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 45, + 467, + 278, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 467, + 278, + 500 + ], + "spans": [ + { + "bbox": [ + 45, + 467, + 278, + 500 + ], + "type": "text", + "content": "[8] Xiangyu Chen, Xintao Wang, Jiantao Zhou, and Chao Dong. Activating more pixels in image super-resolution transformer. arXiv preprint arXiv:2205.04437, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 45, + 501, + 278, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 501, + 278, + 555 + ], + "spans": [ + { + "bbox": [ + 45, + 501, + 278, + 555 + ], + "type": "text", + "content": "[9] Yinbo Chen, Sifei Liu, and Xiaolong Wang. Learning continuous image representation with local implicit image function. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8628-8638, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 40, + 557, + 278, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 40, + 557, + 278, + 602 + ], + "spans": [ + { + "bbox": [ + 40, + 557, + 278, + 602 + ], + "type": "text", + "content": "[10] Jifeng Dai, Haozhi Qi, Yuwen Xiong, Yi Li, Guodong Zhang, Han Hu, and Yichen Wei. Deformable convolutional networks. In Proceedings of the IEEE international conference on computer vision, pages 764-773, 2017. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 40, + 603, + 278, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 40, + 603, + 278, + 658 + ], + "spans": [ + { + "bbox": [ + 40, + 603, + 278, + 658 + ], + "type": "text", + "content": "[11] Tao Dai, Jianrui Cai, Yongbing Zhang, Shu-Tao Xia, and Lei Zhang. Second-order attention network for single image super-resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11065-11074, 2019. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 40, + 660, + 278, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 40, + 660, + 278, + 714 + ], + "spans": [ + { + "bbox": [ + 40, + 660, + 278, + 714 + ], + "type": "text", + "content": "[12] Xin Deng, Hao Wang, Mai Xu, Yichen Guo, Yuhang Song, and Li Yang. Lau-net: Latitude adaptive upscaling network for omnidirectional image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9189-9198, 2021. 1, 2, 6, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 40, + 716, + 278, + 738 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 40, + 716, + 278, + 738 + ], + "spans": [ + { + "bbox": [ + 40, + 716, + 278, + 738 + ], + "type": "text", + "content": "[13] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Image super-resolution using deep convolutional net" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 299, + 98, + 536, + 738 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 318, + 98, + 536, + 120 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 98, + 536, + 120 + ], + "spans": [ + { + "bbox": [ + 318, + 98, + 536, + 120 + ], + "type": "text", + "content": "works. IEEE transactions on pattern analysis and machine intelligence, 38(2):295-307, 2015. 2, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 299, + 122, + 536, + 165 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 122, + 536, + 165 + ], + "spans": [ + { + "bbox": [ + 299, + 122, + 536, + 165 + ], + "type": "text", + "content": "[14] Chao Dong, Chen Change Loy, and Xiaou Tang. Accelerating the super-resolution convolutional neural network. In European conference on computer vision, pages 391-407. Springer, 2016. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 299, + 166, + 536, + 210 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 166, + 536, + 210 + ], + "spans": [ + { + "bbox": [ + 299, + 166, + 536, + 210 + ], + "type": "text", + "content": "[15] Vida Fakour-Sevom, Esin Guldogan, and Joni-Kristian Kämäräinen. 360 panorama super-resolution using deep convolutional networks. In Int. Conf. on Computer Vision Theory and Applications (VISAPP), volume 1, 2018. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 299, + 211, + 536, + 244 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 211, + 536, + 244 + ], + "spans": [ + { + "bbox": [ + 299, + 211, + 536, + 244 + ], + "type": "text", + "content": "[16] Quan Huynh-Thu and Mohammed Ghanbari. Scope of validity of psnr in image/video quality assessment. *Electronics letters*, 44(13):800-801, 2008. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 299, + 245, + 536, + 299 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 245, + 536, + 299 + ], + "spans": [ + { + "bbox": [ + 299, + 245, + 536, + 299 + ], + "type": "text", + "content": "[17] Hiroshi Kawasaki, Katsushi Ikeuchi, and Masao Sakauchi. Super-resolution omnidirectional camera images using spatio-temporal analysis. *Electronics and Communications in Japan* (Part III: Fundamental Electronic Science), 89(6):47-59, 2006. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 299, + 301, + 536, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 301, + 536, + 345 + ], + "spans": [ + { + "bbox": [ + 299, + 301, + 536, + 345 + ], + "type": "text", + "content": "[18] Renata Khasanova and Pascal Frossard. Geometry aware convolutional filters for omnidirectional images representation. In International Conference on Machine Learning, pages 3351-3359. PMLR, 2019. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 299, + 346, + 536, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 346, + 536, + 378 + ], + "spans": [ + { + "bbox": [ + 299, + 346, + 536, + 378 + ], + "type": "text", + "content": "[19] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 299, + 380, + 536, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 380, + 536, + 457 + ], + "spans": [ + { + "bbox": [ + 299, + 380, + 536, + 457 + ], + "type": "text", + "content": "[20] Christian Ledig, Lucas Theis, Ferenc Huszar, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, et al. Photorealistic single image super-resolution using a generative adversarial network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4681-4690, 2017. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 299, + 458, + 536, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 458, + 536, + 491 + ], + "spans": [ + { + "bbox": [ + 299, + 458, + 536, + 491 + ], + "type": "text", + "content": "[21] Wenbo Li, Xin Lu, Jiangbo Lu, Xiangyu Zhang, and Jiaya Jia. On efficient transformer and image pre-training for low-level vision. arXiv preprint arXiv:2112.10175, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 299, + 492, + 536, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 492, + 536, + 547 + ], + "spans": [ + { + "bbox": [ + 299, + 492, + 536, + 547 + ], + "type": "text", + "content": "[22] Zheyuan Li, Yingqi Liu, Xiangyu Chen, Haoming Cai, Jinjin Gu, Yu Qiao, and Chao Dong. Blueprint separable residual network for efficient image super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 833-843, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 299, + 548, + 536, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 548, + 536, + 602 + ], + "spans": [ + { + "bbox": [ + 299, + 548, + 536, + 602 + ], + "type": "text", + "content": "[23] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1833-1844, 2021. 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 299, + 604, + 536, + 659 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 604, + 536, + 659 + ], + "spans": [ + { + "bbox": [ + 299, + 604, + 536, + 659 + ], + "type": "text", + "content": "[24] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 136-144, 2017. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 299, + 660, + 536, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 660, + 536, + 714 + ], + "spans": [ + { + "bbox": [ + 299, + 660, + 536, + 714 + ], + "type": "text", + "content": "[25] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 299, + 716, + 536, + 738 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 716, + 536, + 738 + ], + "spans": [ + { + "bbox": [ + 299, + 716, + 536, + 738 + ], + "type": "text", + "content": "[26] Yiqun Mei, Yuchen Fan, and Yuqian Zhou. Image superresolution with non-local sparse attention. In Proceedings of" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 277, + 751, + 299, + 760 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 751, + 299, + 760 + ], + "spans": [ + { + "bbox": [ + 277, + 751, + 299, + 760 + ], + "type": "text", + "content": "13291" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 41, + 98, + 278, + 738 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 60, + 98, + 278, + 119 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 98, + 278, + 119 + ], + "spans": [ + { + "bbox": [ + 60, + 98, + 278, + 119 + ], + "type": "text", + "content": "the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3517-3526, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 41, + 121, + 278, + 197 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 41, + 121, + 278, + 197 + ], + "spans": [ + { + "bbox": [ + 41, + 121, + 278, + 197 + ], + "type": "text", + "content": "[27] Hajime Nagahara, Yasushi Yagi, and Masahiko Yachida. Super-resolution from an omnidirectional image sequence. In 2000 26th Annual Conference of the IEEE Industrial Electronics Society. IECON 2000. 2000 IEEE International Conference on Industrial Electronics, Control and Instrumentation. 21st Century Technologies, volume 4, pages 2559-2564. IEEE, 2000. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 41, + 198, + 278, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 41, + 198, + 278, + 251 + ], + "spans": [ + { + "bbox": [ + 41, + 198, + 278, + 251 + ], + "type": "text", + "content": "[28] Akito Nishiyama, Satoshi Ikehata, and Kiyoharu Aizawa. 360 single image super resolution via distortion-aware network and distorted perspective images. In 2021 IEEE International Conference on Image Processing (ICIP), pages 1829-1833. IEEE, 2021. 1, 2, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 41, + 252, + 278, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 41, + 252, + 278, + 307 + ], + "spans": [ + { + "bbox": [ + 41, + 252, + 278, + 307 + ], + "type": "text", + "content": "[29] Ben Niu, Weilei Wen, Wenqi Ren, Xiangde Zhang, Lianping Yang, Shuzhen Wang, Kaihao Zhang, Xiaochun Cao, and Haifeng Shen. Single image super-resolution via a holistic attention network. In European conference on computer vision, pages 191-207. Springer, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 41, + 309, + 278, + 362 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 41, + 309, + 278, + 362 + ], + "spans": [ + { + "bbox": [ + 41, + 309, + 278, + 362 + ], + "type": "text", + "content": "[30] Cagri Ozcinar, Aakanksha Rana, and Aljosa Smolic. Superresolution of omnidirectional images using adversarial learning. In 2019 IEEE 21st International Workshop on Multimedia Signal Processing (MMSP), pages 1-6. IEEE, 2019. 2, 3, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 41, + 363, + 278, + 395 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 41, + 363, + 278, + 395 + ], + "spans": [ + { + "bbox": [ + 41, + 363, + 278, + 395 + ], + "type": "text", + "content": "[31] Yule Sun, Ang Lu, and Lu Yu. Weighted-to-spherically-uniform quality evaluation for omnidirectional video. IEEE signal processing letters, 24(9):1408-1412, 2017. 3, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 41, + 396, + 278, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 41, + 396, + 278, + 450 + ], + "spans": [ + { + "bbox": [ + 41, + 396, + 278, + 450 + ], + "type": "text", + "content": "[32] Yapeng Tian, Yulun Zhang, Yun Fu, and Chenliang Xu. Tdan: Temporally-deformable alignment network for video super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3360-3369, 2020. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 41, + 452, + 278, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 41, + 452, + 278, + 506 + ], + "spans": [ + { + "bbox": [ + 41, + 452, + 278, + 506 + ], + "type": "text", + "content": "[33] Radu Timofte, Eirikur Agustsson, Luc Van Gool, Ming-Hsuan Yang, and Lei Zhang. Ntire 2017 challenge on single image super-resolution: Methods and results. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 114-125, 2017. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 41, + 506, + 278, + 561 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 41, + 506, + 278, + 561 + ], + "spans": [ + { + "bbox": [ + 41, + 506, + 278, + 561 + ], + "type": "text", + "content": "[34] Xintao Wang, Kelvin CK Chan, Ke Yu, Chao Dong, and Chen Change Loy. Edvr: Video restoration with enhanced deformable convolutional networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 0–0, 2019. 3, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 41, + 562, + 278, + 615 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 41, + 562, + 278, + 615 + ], + "spans": [ + { + "bbox": [ + 41, + 562, + 278, + 615 + ], + "type": "text", + "content": "[35] Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1905-1914, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 41, + 617, + 278, + 671 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 41, + 617, + 278, + 671 + ], + "spans": [ + { + "bbox": [ + 41, + 617, + 278, + 671 + ], + "type": "text", + "content": "[36] Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Yu Qiao, and Chen Change Loy. Esrgan: Enhanced super-resolution generative adversarial networks. In Proceedings of the European conference on computer vision (ECCV) workshops, pages 0–0, 2018. 2, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 41, + 672, + 278, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 41, + 672, + 278, + 715 + ], + "spans": [ + { + "bbox": [ + 41, + 672, + 278, + 715 + ], + "type": "text", + "content": "[37] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 41, + 716, + 278, + 738 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 41, + 716, + 278, + 738 + ], + "spans": [ + { + "bbox": [ + 41, + 716, + 278, + 738 + ], + "type": "text", + "content": "[38] Zhuofan Xia, Xuran Pan, Shiji Song, Li Erran Li, and Gao Huang. Vision transformer with deformable attention. In" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 299, + 98, + 537, + 578 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 318, + 98, + 536, + 120 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 98, + 536, + 120 + ], + "spans": [ + { + "bbox": [ + 318, + 98, + 536, + 120 + ], + "type": "text", + "content": "Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4794-4803, 2022. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 299, + 122, + 537, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 122, + 537, + 175 + ], + "spans": [ + { + "bbox": [ + 299, + 122, + 537, + 175 + ], + "type": "text", + "content": "[39] Jianxiong Xiao, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Recognizing scene viewpoint using panoramic place representation. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, pages 2695-2702. IEEE, 2012. 1, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 299, + 177, + 537, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 177, + 537, + 240 + ], + "spans": [ + { + "bbox": [ + 299, + 177, + 537, + 240 + ], + "type": "text", + "content": "[40] Youngho Yoon, Inchul Chung, Lin Wang, and Kuk-Jin Yoon. Spheresr: 360deg image super-resolution with arbitrary projection via continuous spherical image representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5677-5686, 2022. 1, 2, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 299, + 244, + 537, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 244, + 537, + 297 + ], + "spans": [ + { + "bbox": [ + 299, + 244, + 537, + 297 + ], + "type": "text", + "content": "[41] Kai Zhang, Jingyun Liang, Luc Van Gool, and Radu Timofte. Designing a practical degradation model for deep blind image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4791-4800, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 299, + 299, + 537, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 299, + 537, + 353 + ], + "spans": [ + { + "bbox": [ + 299, + 299, + 537, + 353 + ], + "type": "text", + "content": "[42] Wenlong Zhang, Yihao Liu, Chao Dong, and Yu Qiao. Ranksrgan: Generative adversarial networks with ranker for image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3096-3105, 2019. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 299, + 355, + 537, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 355, + 537, + 410 + ], + "spans": [ + { + "bbox": [ + 299, + 355, + 537, + 410 + ], + "type": "text", + "content": "[43] Yulun Zhang, Kunpeng Li, Kai Li, Lichen Wang, Bineng Zhong, and Yun Fu. Image super-resolution using very deep residual channel attention networks. In Proceedings of the European conference on computer vision (ECCV), pages 286-301, 2018. 1, 2, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 299, + 411, + 537, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 411, + 537, + 474 + ], + "spans": [ + { + "bbox": [ + 299, + 411, + 537, + 474 + ], + "type": "text", + "content": "[44] Yupeng Zhang, Hengzhi Zhang, Daojing Li, Liyan Liu, Hong Yi, Wei Wang, Hiroshi Suitoh, and Makoto Odamaki. Toward real-world panoramic image enhancement. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 628-629, 2020. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 299, + 478, + 537, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 478, + 537, + 532 + ], + "spans": [ + { + "bbox": [ + 299, + 478, + 537, + 532 + ], + "type": "text", + "content": "[45] Yufeng Zhou, Mei Yu, Hualin Ma, Hua Shao, and Gangyi Jiang. Weighted-to-spherically-uniform sSIM objective quality evaluation for panoramic video. In 2018 14th IEEE International Conference on Signal Processing (ICSP), pages 54-57. IEEE, 2018. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 299, + 533, + 537, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 533, + 537, + 578 + ], + "spans": [ + { + "bbox": [ + 299, + 533, + 537, + 578 + ], + "type": "text", + "content": "[46] Xizhou Zhu, Han Hu, Stephen Lin, and Jifeng Dai. Deformable convnets v2: More deformable, better results. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9308-9316, 2019. 5" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 277, + 751, + 300, + 760 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 751, + 300, + 760 + ], + "spans": [ + { + "bbox": [ + 277, + 751, + 300, + 760 + ], + "type": "text", + "content": "13292" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 595, + 842 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/OT-Filter_ An Optimal Transport Filter for Learning With Noisy Labels/69b93138-511e-410f-9d28-563936aa9d2f_content_list.json b/2023/OT-Filter_ An Optimal Transport Filter for Learning With Noisy Labels/69b93138-511e-410f-9d28-563936aa9d2f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e1d17b550a321e4d31922effe5ca234f14de8aaa --- /dev/null +++ b/2023/OT-Filter_ An Optimal Transport Filter for Learning With Noisy Labels/69b93138-511e-410f-9d28-563936aa9d2f_content_list.json @@ -0,0 +1,1759 @@ +[ + { + "type": "text", + "text": "OT-Filter: An Optimal Transport Filter for Learning with Noisy Labels", + "text_level": 1, + "bbox": [ + 124, + 130, + 844, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chuanwen Feng† Yilong Ren† Xike Xie \nUniversity of Science and Technology of China \nData Darkness Lab, MIRACLE Center, Suzhou Institute for Advanced Research, USTC \n{chuanwen,ylren}@mail.ustc.edu.cn, {xkxie}@ustc.edu.cn", + "bbox": [ + 137, + 179, + 831, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 286, + 313, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The success of deep learning is largely attributed to the training over clean data. However, data is often coupled with noisy labels in practice. Learning with noisy labels is challenging because the performance of the deep neural networks (DNN) drastically degenerates, due to confirmation bias caused by the network memorization over noisy labels. To alleviate that, a recent prominent direction is on sample selection, which retrieves clean data samples from noisy samples, so as to enhance the model's robustness and tolerance to noisy labels. In this paper, we revamp the sample selection from the perspective of optimal transport theory and propose a novel method, called the OT-Filter. The OT-Filter provides geometrically meaningful distances and preserves distribution patterns to measure the data discrepancy, thus alleviating the confirmation bias. Extensive experiments on benchmarks, such as Clothing1M and ANIMAL-10N, show that the performance of the OT-Filter outperforms its counterparts. Meanwhile, results on benchmarks with synthetic labels, such as CIFAR-10/100, show the superiority of the OT-Filter in handling data labels of high noise.", + "bbox": [ + 75, + 318, + 472, + 636 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 667, + 209, + 681 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep learning has achieved great success on a flurry of emerging applications, such as [28, 29, 32, 49]. It is believed that the phenomenal achievement of deep learning is largely attributed to accurate labels. However, the inaccuracy or imprecision of labels is inherent in real-world datasets, the so-called noisy label challenge. One way to alleviate that is to collect labels from internet queries over data-level tags, but the performance of deep neural networks (DNN) suffers drastically from the inaccuracy of such labels [27, 32]. A higher quality of data labels can be obtained by employing human workers, but the seemingly \"ground truth annotations\" inevitably involve human biases or mis", + "bbox": [ + 75, + 693, + 470, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "takes [45,66]. More, the human annotation is expensive and time-consuming, especially for large-scale datasets.", + "bbox": [ + 500, + 287, + 890, + 318 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "There have been many works on handling noisy labels, such as regularization [23] and transition matrix [13, 43]. The regularization approach leverages the regularization bias to overcome the label noise issue. But the regularization bias is permanent [62], thus overfitting models to noisy labels. The transition matrix approach assumes that the transition probabilities between clean and noisy labels are fixed and independent of data samples. However, a quality label transition matrix is hard to be estimated, especially when the number of classes is big, making it fall short in handling noisy real-world datasets, such as [60] and [36]. A recent prominent direction is on adopting sample selection [27, 38, 58] for enhancing the label quality, by selecting clean samples from the noisy training dataset. In general, existing literatures in sample selection can be grouped to two categories, co-training networking [27, 62] and criterion-based filtering [31, 34, 57, 58].", + "bbox": [ + 496, + 320, + 892, + 578 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The former utilizes the memorization of DNNs and multiple networks (e.g., co-teaching [27] and its variants [38, 62]) to filter label noise with small loss trick, so that a small set of clean samples are used as training examples. Letting alone the high training overhead of multiple networks, the disadvantages are two-fold: 1) it may require the a-priori knowledge of noisy rates to select the specified proportion of small loss samples as clean samples; 2) the small-loss trick is not tolerant to the error accumulation of network training once a clean sample is falsely recognized, the so-called confirmation bias, especially for labels with high noise where clean and noisy samples largely overlap.", + "bbox": [ + 496, + 580, + 893, + 762 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The latter alleviates the problem by setting a specific criterion. Mostly, existing works [31] [58] adopt Euclidean distances for measuring the similarity between data samples. Distance-based filtering iteratively explores the neighborhood of the feature representation and infers/cleans sample labels by aggregating the information from their neighborhoods. Despite the simplicity, distance-based filtering is insufficient to address noisy labels, especially when the label noise is high, e.g., overlapped label classes.", + "bbox": [ + 496, + 763, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\dagger$ Equal Contribution.", + "bbox": [ + 94, + 886, + 209, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "16164", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we revamp the sample selection from the perspective of optimal transport [56] and propose a novel filtering method, called the OT-Filter. In light of the optimal transport, we construct discrete probability measures over the sample features, which lifts the Euclidean space of feature vectors to a probability space. It thus enables a geometric way of measuring the discrepancy between probability measures, representing the corresponding sample features. In addition to the distance-based metric in the Euclidean space [31, 58], the OT-Filter also captures the distribution information in the probability space so as to alleviate the confirmation bias. Accordingly, a clean representation can be obtained for each class in the probability space. By optimizing the transport plan from a sample to a clean representation, one can better determine if a sample is clean or noisy, thus improving the quality of sample selection.", + "bbox": [ + 80, + 90, + 467, + 332 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In general, the merits of the OT-Filter can be summarized as follows. First, it does not require any a-priori knowledge about the noisy rate of a dataset. Second, it utilizes the optimal transport which provides geometrically meaningful distances to exploit the sample discrepancy yet preserving the distribution patterns in corresponding probability space, making the sample selection of high quality and theoretical support. Third, it can be plugged to existing robust training paradigms, e.g., supervised and semi-supervised robust training.", + "bbox": [ + 80, + 333, + 467, + 484 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We conduct extensive experiments with a series of synthetic and real datasets to gain insights into our proposals. The result shows that the OT-Filter follows state-of-the-art (SOTA) [38] when the noise rate is low, and dominates SOTA when the noise rate is high. For instance, the OT-Filter achieves about $14\\%$ and $12\\%$ higher accuracy than SOTA, in the presence of $90\\%$ noise rate, on synthetic datasets CIFAR-10 and CIFAR-100, respectively. Moreover, the OT-Filter outperforms the competitors on real datasets Clothing1M and ANIMAL-10N.", + "bbox": [ + 80, + 486, + 467, + 635 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The rest of the paper is organized as follows. We review the existing literature in Section 2. In Section 3, we present preliminaries of optimal transport. We investigate our proposed OT-Filter in Section 4. Furthermore, we conduct extensive empirical studies in Section 5 and conclude the paper in Section 6.", + "bbox": [ + 80, + 638, + 467, + 728 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 80, + 744, + 215, + 760 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Learning with Noisy Label", + "text_level": 1, + "bbox": [ + 80, + 770, + 316, + 785 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A flurry of research methods [5, 13, 14, 27, 40, 54] was proposed for learning with noisy labels, which can be roughly divided into two categories, robust model [13, 22, 39, 43, 54, 61, 64] and clean sample selection [27, 32, 62]. The clean sample selection aims to select a clean subset from a noisy dataset [14, 27, 38]. One way is to train multiple networks to filter noisy samples. For instance, [32]", + "bbox": [ + 80, + 794, + 467, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "proposed to pre-train an extra network to select clean samples for the main network. The inferiority of this method is on error accumulation. To relieve that, Han et al. [27] proposed Co-teaching that maintains two networks simultaneously, where one network selects clean samples for the other network by small loss trick. However, as the increase of training epochs, Co-teaching would converge to consensus gradually. Then, $\\mathrm{Co - teaching}^{+}$ [62] was proposed to keep two networks diverged. To avoid confirmation bias, DivMix [38] proposed a hybrid framework with semi-supervised training, achieving the state-of-the-art performance. Based on DivMix, [41, 65] modified the training schema, such as self-supervised training [11, 33], to boost the performance.", + "bbox": [ + 503, + 90, + 888, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The other research line of sample selection is to filter noisy samples based on specific criteria [31, 34, 42, 59]. For example, Xia et al. [59] proposed to reduce the uncertainty of loss using specific strategies. Iscen et al. [31] proposed to leverage similarities between training examples in the feature space, encouraging the prediction of each example to be close to its nearest neighbors. Also, in Euclidean space, Wu et al. [58] proposed the TopoFilter that filters noisy samples using k-nearest neighborhood distance between prelogits. Fine [34] proposed to use the principal components of latent representations to select clean samples.", + "bbox": [ + 503, + 303, + 888, + 468 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Optimal Transport", + "text_level": 1, + "bbox": [ + 503, + 482, + 679, + 498 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The optimal transport problem [55, 56] aims to move mass from a probability measure to another probability measure at a minimum cost. It defines the Wasserstein distance [24] and provides a geometric way to measure the discrepancy between probability measures. One factor that limits the wide application of optimal transport is the high computational cost. To improve its scalability, Cuturi [17] proposed an entropic regularization for the transport plan, which yields an efficient algorithm with the matrix scaling method of Sinkhorn-Knopp [35]. Recently, computational optimal transport [44, 48] has found many applications in various areas, e.g., generative models [4, 25], domain adaptation [15, 16], and semi-supervised learning [52, 53]. In particular, [1] studied the barycenter in Wasserstein space. [18] explored the efficient computing of Wasserstein barycenter. [19] proposed to use entropic optimal transport loss, based on joint distribution optimal transport [15], to build a robust training.", + "bbox": [ + 503, + 506, + 888, + 777 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we study the problem of clean sample selection from the perspective of optimal transport to combat label noise. The proposed method can potentially capture geometric information from the probability space. Unlike the mechanism of multiple network training, it only trains a single network without a-priori information on noise rates. Moreover, it can be easily plugged to multiple robust training paradigms.", + "bbox": [ + 503, + 780, + 888, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "16165", + "bbox": [ + 482, + 946, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Preliminary", + "text_level": 1, + "bbox": [ + 76, + 89, + 202, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we present preliminaries on optimal transport, making a basis for subsequent proposed techniques. The optimal transport [56] is to seek an optimal transport plan between two measures at a minimal cost (e.g. the Wasserstein distance), which provides a geometric way of matching probability measures. Here we first introduce the general definition of Wasserstein distance induced by optimal transport problem. More details about optimal transport can be found in [44, 48, 55].", + "bbox": [ + 75, + 114, + 468, + 250 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Wassertein Distance. Let $\\mathcal{S}$ be a locally complete and separable metric space, the $\\mathcal{P}(S)$ be the Borel probability measures set on $\\mathcal{S}$ . For any $\\mathcal{X}, \\mathcal{Y} \\subset \\mathcal{S}$ , given $\\mu \\in \\mathcal{P}(\\mathcal{X})$ and $\\nu \\in \\mathcal{P}(\\mathcal{Y})$ , the optimal transport defines a Wasserstein distance between two probability measures, denoted as", + "bbox": [ + 75, + 250, + 468, + 327 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nW _ {p} (\\mu , \\nu) := \\left(\\inf _ {\\pi (\\mu , \\nu)} \\int_ {\\mathcal {X} \\times \\mathcal {Y}} | | x - y | | ^ {p} d \\pi (\\mu , \\nu)\\right) ^ {\\frac {1}{p}}\n$$\n", + "text_format": "latex", + "bbox": [ + 106, + 332, + 437, + 369 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$p \\geq 1$ , where the $\\pi(\\mu, \\nu)$ is the set of joint probability measures with marginal $\\mu$ and $\\nu$ .", + "bbox": [ + 75, + 375, + 468, + 405 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Discrete Optimal Transport", + "text_level": 1, + "bbox": [ + 76, + 412, + 326, + 429 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let $\\Delta_{n} = \\{\\mathbf{a}\\in \\mathbb{R}_{+}^{n}|\\sum_{i = 1}^{n}a_{i} = 1,\\forall a_{i}\\geq 0\\}$ be a probability simplex in dimension $n$ . Consider two empirical probability measures $\\mu = \\sum_{i = 1}^{n}a_{i}\\delta_{x_{i}}$ and $\\nu = \\sum_{j = 1}^{m}b_{j}\\delta_{y_{j}}$ , defined on metric space $\\mathcal{X}$ with support $\\{x_i\\}_{i = 1}^n$ and $\\mathcal{V}$ with support $\\{y_j\\}_{j = 1}^m$ respectively. Here the weight vector $\\mathbf{a} = (a_1,a_2,\\dots,a_n)$ and $\\mathbf{b} = (b_{1},b_{2},\\dots,b_{m})$ live in $\\Delta_{n}$ and $\\Delta_{m}$ , respectively. The $\\delta$ stands for the Dirac unit mass function. Given a transport cost $c:\\mathcal{X}\\times \\mathcal{Y}\\to \\mathbb{R}_+$ the discrete optimal transport between probability measures $\\mu$ and $\\nu$ can be formulated as", + "bbox": [ + 75, + 435, + 468, + 585 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} W _ {p} ^ {p} (\\mu , \\nu) := \\min _ {M \\in \\Pi (\\mu , \\nu)} \\langle C, M \\rangle_ {F} \\tag {1} \\\\ s. t. M \\mathbf {1} _ {m} = \\mu , M ^ {T} \\mathbf {1} _ {n} = \\nu \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 592, + 468, + 636 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $C \\in \\mathbb{R}_+^{n \\times m}$ is the transport cost matrix, and $c_{ij}$ represents a unit transport cost from $x_i$ to $y_j$ . The $M \\in \\mathbb{R}_+^{n \\times m}$ is transport plan in which $m_{ij}$ denotes the amount of mass transported from $x_i$ to $y_j$ . All feasible transport plans constitute transport polytope $\\Pi(\\mu, \\nu)$ . The $\\langle C, M \\rangle_F$ is the Frobenius inner product of matrices and equals to $tr(C^T M)$ .", + "bbox": [ + 75, + 643, + 468, + 751 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Regularized Optimal Transport", + "text_level": 1, + "bbox": [ + 76, + 758, + 356, + 773 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The discrete optimal transport formulation, in essence, is a convex optimization problem. More precisely, it's a linear programming problem. Unfortunately, this linear programming problem has a cubic computing complexity. A way to relieve this is to leverage entropic regularization [17] formulated as", + "bbox": [ + 75, + 781, + 468, + 869 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {M \\in \\Pi (\\mu , \\nu)} \\langle C, M \\rangle_ {F} - \\epsilon H (M) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 173, + 878, + 468, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\epsilon > 0$ is the regularization coefficient, and $H(M)$ is the entropic regularization term, which promotes an efficient computation for transportation via the matrix scaling algorithm [17], given by", + "bbox": [ + 496, + 90, + 890, + 151 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nH (M) := - \\sum_ {i j} M _ {i j} \\left(\\log \\left(M _ {i j}\\right) - 1\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 162, + 820, + 195 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Wasserstein Barycenter", + "text_level": 1, + "bbox": [ + 500, + 205, + 718, + 222 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given $N$ probability measures $\\{\\nu_{1},\\nu_{2},\\dots,\\nu_{N}\\}$ , $\\nu_{i} \\in \\mathcal{P}(S)$ , each of which has finite supports and second moments. A Wasserstein barycenter of these measures is a probability measure $\\mu$ , satisfying:", + "bbox": [ + 498, + 229, + 890, + 290 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mu := \\inf \\sum_ {i = 1} ^ {N} \\lambda_ {i} W _ {2} ^ {2} (\\nu_ {i}, \\mu), \\quad s. t. \\sum_ {i = 1} ^ {N} \\lambda_ {i} = 1, \\forall \\lambda_ {i} \\geq 0\n$$\n", + "text_format": "latex", + "bbox": [ + 508, + 300, + 880, + 340 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The notion was first proposed by [1], where some elegant properties of Wasserstein barycenter were presented. Then [2] discussed it on the discrete case and showed that the problem of finding Wasserstein barycenter over the space $\\mathcal{P}(\\mathcal{S})$ can be reduced to a simpler space $\\mathcal{O}_r(S)$ , where $r = \\sum_{i=1}^{N} e_i - N + 1$ , and $e_i$ is the number of components of $\\nu_i, i \\in [1, N]$ . To find Wasserstein barycenter over space $\\mathcal{O}_r(S)$ , a set of efficient algorithms [7, 18, 26] were proposed.", + "bbox": [ + 498, + 351, + 890, + 489 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Methods", + "text_level": 1, + "bbox": [ + 500, + 503, + 596, + 517 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Overview", + "text_level": 1, + "bbox": [ + 500, + 527, + 609, + 541 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we study the OT-Filter, the proposed method based on optimal transport for learning with noisy labels in a semi-supervised paradigm. The key idea of our proposed method is to transport samples with noisy labels to samples with clean labels at a minimum cost, which makes sample selection of high quality and theoretical support.", + "bbox": [ + 496, + 550, + 890, + 641 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The mechanism of the OT-Filter, consists of two phases, representation phase (Section 4.2) and transportation phase (Section 4.3). Then, the OT-Filter can be plugged to off-the-shelf SSLs for robust training, e.g., MixMatch [8], by treating clean samples as labeled data and noisy samples as unlabeled samples (Section 4.4).", + "bbox": [ + 496, + 641, + 890, + 731 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.2. Representation Phase: Finding Clean Representations via Wasserstein Barycenters", + "text_level": 1, + "bbox": [ + 498, + 741, + 890, + 772 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the representation phase, we first use a pre-trained network, e.g., ResNet [29], to extract a set of feature representations of the labeled samples. Then, we define a discrete probability measure over the feature representations of each class to find the clean representations (or prototypes) via Wasserstein barycenters. Since there are mistrabeled samples, the barycenters we obtained may be noisy. Therefore, we iteratively optimize the barycenters with the expectation", + "bbox": [ + 496, + 780, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "16166", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "maximization (EM) algorithm [20] to converge to clean representations. The details about the representation phase are covered in Algorithm 1.", + "bbox": [ + 75, + 90, + 468, + 135 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Feature Extraction. The purpose of the feature extraction is to transform the input noisy label dataset into a feature-label dataset via a network. A noisy label dataset of $N$ samples can be represented by $\\hat{\\mathcal{D}} = \\{(\\mathbf{x}_i,\\hat{y}_i)\\}_{i = 1}^N$ , where $\\mathbf{x}_i\\in \\mathbb{R}^d$ denotes the $i$ -th sample and $\\dot{y}_i\\in \\{0,1\\} ^K$ denotes the corresponding noisy label over $K$ classes. Then, a feature extraction network $f(\\omega ,\\mathbf{x}_i):\\mathbb{R}^d\\to \\mathbb{R}^m$ with parameters $\\omega$ maps a sample $\\mathbf{x}_i$ to a $m$ -dimensional feature representation, denoted as $\\tilde{\\mathbf{x}}_i\\in \\mathbb{R}^m$ . Therefore, all output feature representations and their corresponding labels constitute a feature-label set, denoted as $\\tilde{D} = \\{(\\tilde{\\mathbf{x}}_i,\\hat{y}_i)\\}_{i = 1}^N$ .", + "bbox": [ + 75, + 136, + 467, + 303 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Probability Measure Modeling. The purpose of probability measure modeling is to define a probability measure over the feature-label set. This definition will potentially lift a Euclidean space $\\mathcal{D}$ , in which the feature representations originally live, to a probability space, abbreviated as $\\mathcal{P}(\\mathcal{D})$ . Assume that the labels are of $K$ classes. First, we split the feature-label set $\\tilde{D}$ into $K$ subsets, i.e., $\\{s_1,s_2,\\dots,s_K\\}$ according to the label class. For one feature representation $\\tilde{\\mathbf{x}}_i$ in class $k\\in [1,K]$ , we represent it as $\\tilde{\\mathbf{x}}_i^k,i\\in [1,|s_k|]$ . Then, all feature representations in subset $s_k$ construct a probability measure, denoted as", + "bbox": [ + 75, + 301, + 467, + 468 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {Q} = \\frac {1}{| s _ {k} |} \\sum_ {i = 1} ^ {| s _ {k} |} \\delta_ {\\tilde {\\mathbf {x}} _ {i} ^ {k}}\n$$\n", + "text_format": "latex", + "bbox": [ + 212, + 479, + 331, + 522 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the $\\delta_{\\widetilde{\\mathbf{x}}_i^k}$ is a Dirac unit mass on $\\widetilde{\\mathbf{x}}_i^k$ , and for simplicity, we use uniform weights.", + "bbox": [ + 76, + 535, + 467, + 565 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Clean Representation Retrieval. The problem of finding the representation of any subset $s_k$ is equivalent to finding the corresponding Wasserstein barycenter $\\mathcal{B}_k$ with finite supports, formalized as follows:", + "bbox": [ + 75, + 566, + 467, + 626 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {B} _ {k} := \\inf _ {\\mathcal {B} \\in \\mathcal {O} _ {r} (\\tilde {\\mathcal {D}})} W _ {2} ^ {2} (\\mathcal {B}, \\mathcal {Q}) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 637, + 467, + 664 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the $W$ is the Wasserstein distance, and $r = |s_k|$ . As mentioned in Section 3, from the perspective of linear programming, the above optimization admits the dual:", + "bbox": [ + 76, + 676, + 467, + 722 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nD (\\alpha , \\beta) = \\max _ {(\\alpha , \\beta) \\in R (C)} \\alpha^ {T} \\mathcal {B} + \\beta^ {T} \\mathcal {Q}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 732, + 393, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the polyhedron of dual variables is:", + "bbox": [ + 76, + 768, + 354, + 784 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nR (C) = \\left\\{\\left(\\alpha , \\beta\\right) \\in \\mathbb {R} ^ {r} \\times \\mathbb {R} ^ {r} \\mid \\alpha_ {i} + \\beta_ {j} \\leq C _ {i j} \\wedge i, j \\in [ r ] \\right\\}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 796, + 459, + 813 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "By means of duality theory [6,9,10], we can find the relation of solution between the primal and dual problem. Here, the optimal solution of $D(\\alpha, \\beta)$ is a subgradient of $W_2^2(\\mathcal{Q}, \\mathcal{B})$ with respect to $\\mathcal{B}$ . Thus, we employ a simple projected subgradient [18] to optimize Equation 3. over $\\mathcal{B}$ .", + "bbox": [ + 75, + 824, + 467, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Despite we have found a set of Wasserstein barycenters $\\{\\mathcal{B}_i\\}_{i=1}^K$ , the barycenters may be noisy since there are mislabeled data. Moreover, the clean barycenters are unknown. Therefore, for each class, we consider the ideal and clean barycenter $\\mathcal{B}_c$ as a hidden variable, the obtained noisy barycenter $\\mathcal{B}$ as an observed variable. Then, the problem of finding clean barycenter $\\mathcal{B}_c$ is to maximize the following log-likelihood:", + "bbox": [ + 496, + 90, + 890, + 210 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\theta} \\leftarrow \\arg \\max _ {\\theta} \\log \\sum_ {\\mathcal {B} _ {c}} p (\\mathcal {B} _ {c}, \\mathcal {B} | \\theta)\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 223, + 805, + 255 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the $\\theta$ is the parameters of probabilistic model. By this way, we can optimize the noisy barycenters iteratively with the EM algorithm to converge to clean barycenters. In E-step, we construct optimal transport between all samples and clean barycenters, thereby we can infer the labels of all samples. In M-step, we refine the barycenters with inferred label obtained from E-step. The overall computing details see Algorithm 1.", + "bbox": [ + 496, + 267, + 890, + 387 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 Find Clean Representation via Wasserstein Barycenter" + ], + "code_body": "Require: $\\mathcal{Q} = \\{\\tilde{\\mathbf{x}}_i\\}_{i=1}^{|s_k|} \\in \\mathbb{R}^{m \\times r}$ , regularization coefficient $\\epsilon$ , $\\mathcal{B} \\in \\mathbb{R}^{m \\times e}$ \n1: Initialize: $\\mathcal{B}$ , $\\nu \\gets \\mathcal{B}$ , $\\mu \\gets \\mathcal{Q}$ , $\\eta = 1/2$ , $t = 2$ \n2: while $\\mathcal{B}$ not converged do \n3: set $\\nu = \\mathbf{1}_e / e$ , $\\mu = \\mathbf{1}_r / r$ , $\\hat{\\nu} = \\tilde{\\nu} = \\mathbf{1} / e$ \n4: while $\\nu$ not converged do \n5: $\\beta \\gets (t + 1) / 2$ , $\\nu \\gets (1 - \\beta^{-1})\\hat{\\nu} + \\beta^{-1}\\tilde{\\nu}$ \n6: $\\alpha \\gets$ optimizer of $\\mathbf{d}(\\nu, \\mu, C)$ \n7: $\\tilde{\\nu} \\gets \\tilde{\\nu} \\circ \\exp(-\\beta\\alpha)$ , $\\tilde{\\nu} \\gets \\tilde{\\nu} / \\tilde{\\nu}^T\\mathbf{1}$ \n8: $\\hat{\\nu} \\gets (1 - \\beta^{-1})\\hat{\\nu} + \\beta^{-1}\\tilde{\\nu}$ \n9: $t \\gets t + 1$ \n10: end while \n11: $\\nu \\gets \\hat{\\nu}$ \n12: $\\mathcal{M} \\gets$ optimizer of $\\mathbf{p}(\\nu, \\mu, C)$ \n13: $\\mathcal{B} \\gets (1 - \\eta)\\mathcal{B} + \\eta (\\mathcal{QM})diag(\\nu^{-1})$ \n14: end while \n15: while $\\mathcal{B}_c$ not converged do \n16: E-Step: \n17: $\\mathbb{L}_{\\hat{\\mathcal{D}}} \\gets \\mathrm{OT}(\\mu, \\nu, C)$ \n18: M-Step: \n19: $\\mathcal{B}_c \\gets$ refine $\\mathcal{B}$ with $\\mathbb{L}_{\\hat{\\mathcal{D}}}$ \n20: end while \n21: Return $\\mathcal{B}_c$", + "bbox": [ + 500, + 435, + 890, + 790 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.3. Transportation Phase: Transporting Noisy Labels via Regularized Optimal Transport", + "text_level": 1, + "bbox": [ + 500, + 816, + 890, + 848 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the transportation phase, noisy feature representations are transported to clean representations. In particular, we align all feature representations, including those of noisy", + "bbox": [ + 500, + 854, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "16167", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "samples, to the clean representations (obtained in the representation phase) via sparsity regularized optimal transport. This operation potentially detects the corrupted labels based on the result of the optimal transport. Here, the mechanism of our sample selection via optimal transport is described as follows. First, we align the clean barycenters to the all rest samples via optimal transport. Based on the result of optimal transport, we can infer the labels of the all rest samples. We consider the route with the max mass transport that forms a coupling and then assign the label of a clean barycenter to the samples on this route. Second, we select the sample whose inferred label is identical to the original label as a clean sample.", + "bbox": [ + 75, + 90, + 472, + 286 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Sparsity Regularization. After the representation phase, clean representations are obtained, which are essentially the barycenters of label classes. The clean representation of a class provides the guidance of identifying whether samples are clean or noisy w.r.t. the given class.", + "bbox": [ + 75, + 287, + 472, + 362 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We consider the barycenters of all the $K$ classes as a discrete probability measure, denoted as $\\mu = \\frac{1}{K}\\sum_{k = 1}^{K}\\delta_{u_k}$ where $\\delta_{\\mu_k}$ is a Dirac unit mass on $\\mu_{k}$ . Accordingly, the feature representations of the entire dataset construct a discrete probability measure $\\nu = \\frac{1}{N}\\sum_{i = 1}^{N}\\delta_{\\tilde{\\mathbf{x}}_i}$ . Then, we transport feature representations $\\nu$ of all samples to the clean barycenters $\\mu$ , in order to detect if a sample is clean or mislabeled. Inspired by [16,46], we view the clean barycenters as the source domain and the feature representations of samples as the target domain. Then, we recast the alignments as an optimal transport problem with a sparsity regularization [16], written as", + "bbox": [ + 75, + 363, + 472, + 544 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {M \\in \\Pi (\\mu , \\nu)} \\langle C, M \\rangle_ {F} - \\epsilon H (M) + \\gamma \\sum_ {j, c} | | M (i _ {c}, j) | | _ {1} ^ {\\frac {1}{2}} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 551, + 468, + 587 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The second term is an entropic regularization, as described in Equation 2. The third term is a sparsity regularization with coefficient $\\gamma$ , where $i_c$ index the line if its element in class $c$ , and $M(i_c, j)$ is a vector consisting of the $j$ -th column of $M$ in class $c$ . The $\\left\\| \\cdot \\right\\|_1^{1/2}$ denotes an $L_1$ norm with the power of $\\frac{1}{2}$ . The entropic regularization enables discrete optimal transport to be efficient. However, it disperses the transport route, which negatively affects sample selection. Therefore, to improve the quality of sample selection, we employ a sparsity regularization. The sparsity regularization promotes a feature representation that would be aligned to one of the clean barycenters and penalizes transportation matrix $M$ that aligns together feature representations with different labels.", + "bbox": [ + 75, + 595, + 472, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Optimization. Despite the consistent sparsity achieved by the regularization, the objective function of Equation 4 is non-convex. Moreover, the $L_{1}$ norm with the power in label regularization term is a concave function. A common optimization strategy [37] is to construct a convex upper bound, represented by convex functions, over the non-convex prob", + "bbox": [ + 75, + 810, + 472, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "lem. By doing so, we can use the optimal solution of convex upper bound to approximate the optimal solution of the original non-convex problem. If applying the simplest linear approximation, we have", + "bbox": [ + 496, + 90, + 890, + 151 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\sum_ {j, c} \\left| \\left| M \\left(i _ {c}, j\\right) \\right| \\right| _ {1} ^ {\\frac {1}{2}} \\leq \\langle G, M \\rangle_ {F} + \\mathcal {L}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 161, + 815, + 196 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where matrix $G$ is $\\frac{1}{2} (||\\hat{M}(i_c, j) + \\delta||^{-\\frac{1}{2}})$ , and $\\hat{M}$ is a given start point. The steps of optimization process are depicted as Algorithm 2.", + "bbox": [ + 496, + 208, + 893, + 256 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Algorithm 2 Transportation: Transport Noisy Label via Regularized Optimal Transport", + "bbox": [ + 500, + 270, + 890, + 301 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Require: optimal cost $C_{\\text{min}}$ from Equation 1,", + "bbox": [ + 500, + 304, + 808, + 320 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: Initialize: $G = 0$", + "2: while G not converged do", + "3: $C_{min} \\gets C_{min} + G$", + "4: $M\\gets$ optimizer of Equation 2 with $C_{min}$", + "5: $G\\gets$ update $G$ using $M$", + "6: end while" + ], + "bbox": [ + 509, + 320, + 826, + 409 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Robust Training in SSL", + "text_level": 1, + "bbox": [ + 500, + 438, + 718, + 454 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After all feature representations are transported to the clean barycenters. We filter noisy samples and obtain two data subsets, namely the clean sample set and noisy sample set. As aforementioned, our sample selection method is flexible enough for supporting various robust training paradigms, e.g., the robust supervised training [64] and the robust semi-supervised learning [50]. Following [38], we select MixMatch [8] with data augmentation strategy [41] as the practice of robust semi-supervised training but with fewer operations. We admit the labels of clean samples and neglect the labels of noisy samples, and then we use labeled data and unlabeled data for training in a semi-supervised learning fashion. The overall training process is described as Algorithm 3.", + "bbox": [ + 496, + 460, + 890, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 500, + 686, + 633, + 704 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Datasets", + "text_level": 1, + "bbox": [ + 500, + 710, + 601, + 726 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We perform extensive experiments on four benchmark datasets: CIFAR-10 [36] and CIFAR-100 [36] with synthetic label noise, and Clothing1M [60] and ANIMAL-10N [51] with real label noise.", + "bbox": [ + 496, + 734, + 890, + 794 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "CIFAR10/100. We evaluate the OT-Filter on CIFAR10 and CIFAR-100 [36] with synthetic noise. Both datasets contain 60,000 $32 \\times 32$ color images, in 10 and 100 classes, respectively. From both datasets, 50K images are used for training, and 10K images are used for testing. Since noise characteristics can hardly be determined in advance, synthetic noise is commonly taken for controlling the noise rate", + "bbox": [ + 496, + 794, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "16168", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 3 Robust Training in SSL" + ], + "code_body": "Require: network parameter $\\omega$ , Beta distribution parameter $\\alpha$ , weight of unlabeled loss $\\lambda_{\\mathcal{U}}$ , batch size $B$ , number of augmentations $K$ , sharpening temperature $T$ , labeled data $\\mathcal{X}$ , unlabeled data $\\mathcal{U}$ \n1: while $b$ in $B$ do \n2: $\\hat{x}_b \\gets \\text{augment}(x_b)$ \n3: for $k$ in $K$ do \n4: $\\hat{u}_{b,k} \\gets \\text{augment}(u_b)$ \n5: end for \n6: $\\hat{q}_b \\gets \\frac{1}{K} \\sum_k p_{model}(\\hat{u}_{b,k}; \\omega)$ \n7: $q_b \\gets \\text{sharpen}(\\hat{q}_b, T)$ \n8: end while \n9: $\\hat{\\mathcal{X}} = \\{(\\hat{x}_b, p_b)\\}_{b=1}^B, \\hat{\\mathcal{U}} = \\{(\\hat{u}_{b,k}, q_b)\\}_{b,k=1}^{B,K}$ \n10: $\\mathcal{W} = \\text{shuffle}(\\text{concat}(\\hat{\\mathcal{X}}, \\hat{\\mathcal{U}}))$ \n11: $\\hat{\\mathcal{X}}' = \\text{Mixup}(\\hat{\\mathcal{X}}_i, \\mathcal{W}_i), i \\in [1, |\\hat{\\mathcal{X}}|]$ \n12: $\\hat{\\mathcal{U}}' = \\text{Mixup}(\\hat{\\mathcal{U}}_i, \\mathcal{W}_{i+ \\hat{\\mathcal{U}}}), i \\in [1, |\\hat{\\mathcal{U}}|]$ \n13: $\\mathcal{L}_{\\mathcal{X}} \\gets \\text{CE}(\\mathcal{X}')$ , $\\mathcal{L}_{\\mathcal{U}} \\gets \\text{MAE}(\\mathcal{U}')$ \n14: $\\mathcal{L} \\gets \\mathcal{L}_{\\mathcal{X}} + \\lambda_{\\mathcal{U}} \\mathcal{L}_{\\mathcal{U}}$", + "bbox": [ + 78, + 109, + 468, + 391 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "to deliberately evaluate learning algorithms with noisy labels. Thus, following previous works [27, 38], we consider two types of label noise, i.e., symmetric and asymmetric label noise. The symmetric label noise is generated by randomly flipping labels of a portion of samples from one class to all other possible classes. The asymmetric label noise is designed to follow the structure of real-world label noise, where labels are flipped to similar classes within the superclasses.", + "bbox": [ + 75, + 421, + 468, + 556 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Clothing1M. Clothing1M is a large-scale dataset with noisy labels [60], containing over 1M images obtained from online shopping websites. The labels are from 14 classes generated based on surrounding texts provided by the sellers, and the noise rate is estimated around $38.5\\%$ [39]. Also, the dataset provides 3 clean datasets for training, validation, and testing, containing 50K, 14K, and 10K images, respectively.", + "bbox": [ + 75, + 558, + 468, + 678 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ANIMAL-10N. ANIMAL-10N is a real-world noisy dataset released by [51], which is crawled from several online search engines using predefined labels as searching keywords. There are in total 55K images, of which 50K images are for training, and 5K images are for testing. The noise rate was estimated to be around $8\\%$ .", + "bbox": [ + 75, + 680, + 468, + 770 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Implementation Details", + "text_level": 1, + "bbox": [ + 76, + 785, + 292, + 800 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For CIFAR-10 and CIFAR-100, we use the PreAct ResNet18 [29] as the backbone and train it using the SGD optimizer with the following settings: a momentum of 0.9, a weight decay of 5e-4, and a batch size of 128. The learning rate was initialized as 0.02 and reduced by a factor of 10, after 150 epochs. The network was trained for 300", + "bbox": [ + 75, + 809, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "epochs. The warmup period is 10 epochs for CIFAR-10, and 30 epochs for CIFAR-100.", + "bbox": [ + 496, + 90, + 890, + 119 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For Clothing1M, we use the ResNet50 [28] pre-trained on ImageNet [21] as the backbone and train it using the SGD optimizer with the following settings: a momentum of 0.9, a weight decay of 1e-3, and a batch size of 32. The learning rate was initialized as 0.002 and reduced by a factor of 10 after 50 epochs. The network was trained for 120 epochs. For each epoch, we sample 1000 mini-batches from the training data.", + "bbox": [ + 496, + 122, + 890, + 241 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For ANIMAL-10N, we use VGG19 [49] with batch normalization [30] as the backbone and train it using the SGD optimizer. We train the network for 100 epochs. The initial learning rate is set as 0.01 and reduced by a factor of 5 after 50 and 75 epochs.", + "bbox": [ + 496, + 243, + 890, + 318 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In addition, we set the entropic regularization coefficient $\\epsilon$ and sparsity regularization coefficient $\\gamma$ as 10 and 1, respectively, which are consistent for all training implementations.", + "bbox": [ + 496, + 319, + 890, + 378 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Experimental Results", + "text_level": 1, + "bbox": [ + 500, + 391, + 700, + 407 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we present experimental results of the OT-Filter on benchmark datasets with both synthetic and real label noises. For CIFAR-10 and CIFAR-100, we consider the noise rate of $20\\%$ , $50\\%$ , $80\\%$ , and $90\\%$ for symmetric noise, and noise rate of $40\\%$ for asymmetric noise, respectively. The results regarding precision and recall were drawn from the 100-th epoch of robust training.", + "bbox": [ + 496, + 415, + 890, + 520 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Sample Selection Performance. We first present the empirical study on the quality of sample selection. The dataset CIFAR-10 under noise rate of $90\\%$ (symmetric) and noise rate (asymmetric) of $40\\%$ was employed to compare our method with state-of-the-art sample selection method DivMix [38]. Here, we consider two metrics, precision and recall. As shown in Figure 1, with the increase of training epochs, our method outperforms the DivMix [38] in terms of both precision and recall.", + "bbox": [ + 496, + 521, + 890, + 656 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For example, on $90\\%$ symmetric noise, when the training epoch is above 100, the precision of the OT-Filter is steadily above $90\\%$ , while the precision of DivMix is less than $73\\%$ . In terms of the recall, the OT-Filter is above $90\\%$ , while the DivMix is below $82\\%$ , when the training epochs are higher than 100. It also shows that our method converges steadily to the optimal solution, which DivMix yields fluctuations in the convergence process. Similar trends can be observed on $40\\%$ asymmetric noise.", + "bbox": [ + 496, + 657, + 890, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "CIFAR10 and CIFAR100 Datasets. Table 1 shows the test accuracy on CIFAR-10 for symmetric and asymmetric noise, respectively. In all testings, the OT-Filter demonstrates good performance in test accuracy. In particular, in the presence of high noise, the OT-Filter significantly outperforms other competitors. For $90\\%$ noise rate, the test accuracy of our OT-Filter is $90.5\\%$ , which is about $14\\%$", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "16169", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5d9c3fad1e1a6576ba873a292b6cbb4de291f5ab4c0b3ad4590e1966a76763f0.jpg", + "image_caption": [ + "Figure 1. Sample selection performance (Precision & Recall) w.r.t. Training Epochs on CIFAR-10 under synthetic label noise." + ], + "image_footnote": [], + "bbox": [ + 81, + 99, + 277, + 205 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/747b89a5ea3e898f42eab527e153797e3482b5a7014a3263d8bf02bfe7344e86.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 279, + 99, + 473, + 205 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/aff6f3d61ce1e94f8fe84f1f0c57d7c1ba55d4122aa154909e24d109f62034d3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 478, + 99, + 673, + 205 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7535333730364ceb8b133640270ce0e96f7200df21589f3c04ae1028fd48e47a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 676, + 99, + 870, + 205 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/bd5c2aa2080134747a9bd57817140c99b2c2ed422bc0898b38300bca76316bf9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodCIFAR-10Asym
20%50%80%90%40%
CE86.879.462.942.777.3
Co-teaching+ [62]89.585.767.447.971.3
Co-learning [54]92.284.561.2-81.4
TopoFilter [58]90.2-45.7-87.9
CRUST [40]91.186.358.3-88.8
Fine [34]91.087.369.4-89.5
M-correction [3]93.691.875.874.793.3
CTRR [61]93.1-83.781.789.0
DivMix [38]96.194.693.276.093.4
Fine+DivMix [34]96.194.993.590.593.8
OT-Filter96.095.394.090.595.1
", + "bbox": [ + 101, + 271, + 446, + 458 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e82289c0246a3842f2b79edaf512c2658e5ad7bc613bc39d7b4aeeb9740680a5.jpg", + "table_caption": [ + "Table 1. Test accuracies(%) obtained from state-of-the-art sample selection methods. The best results are in bold. The data are copied from respective papers. The - denotes the lack of respective data." + ], + "table_footnote": [], + "table_body": "
MethodCIFAR-100Asym
20%50%80%90%40%
CE62.046.719.910.144.5
Co-teaching+ [62]65.651.827.913.7-
Co-learning [54]66.654.535.5-47.6
CRUST [40]65.256.4--53.0
M-correction [3]73.966.141.624.347.4
CTRR [61]70.1-43.7-54.5
TopoFilter [58]65.6-20.7--
Fine [34]70.364.225.6-61.7
DivMix [38]77.374.660.231.555.1
Fine+DivMix [34]79.174.661.034.3-
OT-Filter76.773.861.842.876.5
", + "bbox": [ + 101, + 541, + 446, + 729 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Test accuracies(%) obtained from state-of-the-art sample selection methods. The best results are in bold. The data are copied from respective papers. The - denotes the lack of respective data.", + "bbox": [ + 75, + 739, + 468, + 795 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "better than DivMix. Although [38] performs slightly better than the OT-Filter at the $20\\%$ noise rate, our method is also competitive. Similar trends can be observed on CIFAR100 (Table 2), where our method mostly follows the SOTA method and significantly outperforms its competitors when", + "bbox": [ + 75, + 824, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7decc5d7358a203eb2da4b4b96a51733ed5cb6cb1a2ba66eb4d03d9ff37a9868.jpg", + "table_caption": [ + "the noise rate is high." + ], + "table_footnote": [], + "table_body": "
MethodBackboneTest Accuracy
CEResNet-5069.2
M-correction [1]ResNet-5071.0
Co-teaching [27]ResNet-5071.7
CTRR [61]ResNet-5072.7
Fine [34]ResNet-5072.9
CRUST [40]ResNet-5073.5
TopoFilter [58]ResNet-5074.1
DivMix* [38]ResNet-5074.3
Fine+DivMix [34]ResNet-5074.4
OT-FilterResNet-5074.5
", + "bbox": [ + 542, + 301, + 849, + 468 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Clothing1M Dataset. Table 3 shows the result of test accuracy on Clothing1M by comparing different competitors. The result validates the effectiveness of the OT-Filter on the large-scale real-world dataset. Moreover, It can be observed that the OT-Filter follows the state-of-the-art performance regarding test accuracy.", + "bbox": [ + 498, + 537, + 893, + 630 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/35a9871031d3b044a6e1af243ed6f33ad4627d766b83967a021b10874128eb2a.jpg", + "table_caption": [ + "Table 3. Test accuracy(%) on Clothing1M. The * denotes we have run the algorithm based on the official implementation. The data are copied from respective papers. The best result is in bold." + ], + "table_footnote": [], + "table_body": "
MethodTest Accuracy
CE79.4
Nested [47]81.3
SELFIE [51]81.8
PLC [63]83.4
Co-teaching+Nested [12]84.1
GJS [22]84.2
OT-Filter85.5
", + "bbox": [ + 562, + 641, + 831, + 770 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Test accuracy(%) on ANIMAL-10N. The data are copied from respective papers. The best result is in bold.", + "bbox": [ + 500, + 779, + 890, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ANIMAL-10N Dataset. Table 4 shows the experimental results on ANIMAL-10N, a real-world dataset with moderate label noise. We compared our method with multiple state-of-the-art baseline methods that are not limited to sample selection. For fair comparison, we use the same VGG-", + "bbox": [ + 498, + 824, + 892, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "16170", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/782face7b230ec611903c1836eb79df154e24bfdc0274d3dc37e91f3c6ba9182.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetCIFAR-100
Noise Rate80%90%80%90%
MethodBestLastBestLastPrecisionRecallPrecisionRecall
OT-Filter w/o Spar-Reg51.050.723.823.790.447.162.119.7
OT-Filter61.861.542.842.193.761.872.340.0
", + "bbox": [ + 81, + 89, + 465, + 154 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f0e34641445936ad2f39c604e9eb6256f5c2dc7c84bb1bb5cee9e416a4382229.jpg", + "image_caption": [ + "Figure 2. The ablation study for the quality of sample selection w/o sparsity regularization on CIFAR-100 under noise rate of $90\\%$ . The result shows that sparsity regularization brings a relatively uniform sample selection." + ], + "image_footnote": [], + "bbox": [ + 83, + 222, + 467, + 371 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "19 network architecture. The results show we outperform all competitors, and have $2.1\\%$ performance improvement over PLC [63], and $3.7\\%$ performance improvement over SELFIE [51].", + "bbox": [ + 75, + 468, + 467, + 529 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Analysis", + "text_level": 1, + "bbox": [ + 76, + 542, + 179, + 558 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conduct a series of ablation studies to understand the effectiveness of the key components of the OT-Filter, including the sparsity regularization of the optimal transport, the EM algorithm, and the entropic regularization coefficient.", + "bbox": [ + 75, + 566, + 467, + 641 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of Sparsity Regularization. The sparsity regularization discussed in Section 4.3 is one of the key components in our OT-Filter, which is designed to improve the accuracy of the optimal transport by controlling the sparsity of transport. We therefore study the effect of the sparsity regularization over sample selection and performance under high noise on CIFAR-100 in Table 5. It can be observed, with the sparsity regularization, both the quality of sample selection and test accuracy are much better than the counterpart without it. In particular, when the noise rate is $90\\%$ , it shows that the technique of sparsity regularization brings in over $10\\%$ improvement in precision, about $21\\%$ improvement in recall, and $19\\%$ improvement in performance. Moreover, as shown in Figure 2, with the sparsity regularization, the number of samples selected for each class is more uniform. It is mainly attributed to discrete optimal transport allowing mass split.", + "bbox": [ + 75, + 643, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/cf45aa418ad277fff815903c981a2726dd1413a6371a8944a3dc1c7518db33d5.jpg", + "table_caption": [ + "Table 5. Ablation study for sparsity regularization on CIFAR-100 dataset. The left part is the performance comparison and the right part is the quality of sample selection." + ], + "table_footnote": [], + "table_body": "
DatasetCIFAR-10
Noise Rate20%50%80%90%
MethodBestLastBestLastBestLastBestLast
OT-Filter w/o EM95.995.795.094.893.693.490.089.8
OT-Filter96.095.895.395.193.793.590.590.2
", + "bbox": [ + 504, + 89, + 887, + 169 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/ebd05a991058b5c6e8e076fac2100472583f2e4affb6800b0512244bd9c76793.jpg", + "table_caption": [ + "Table 6. Ablation study for EM algorithm on CIFAR-10 dataset." + ], + "table_footnote": [], + "table_body": "
DatasetCIFAR-10
Noise Rate20%50%80%90%
MethodPrecisionRecallPrecisionRecallPrecisionRecallPrecisionRecall
ε = 199.793.999.093.994.690.688.684.2
ε = 599.793.899.093.894.690.688.684.2
ε = 1099.793.899.093.894.590.688.684.2
", + "bbox": [ + 503, + 208, + 888, + 279 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 7. Ablation study for entropic regularization coefficient on CIFAR-10 dataset.", + "bbox": [ + 500, + 289, + 890, + 316 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of Expectation Maximization. To filter noisy labels, we align all samples to clean barycenters via optimal transport. Therefore, the quality of clean barycenters affects the quality of sample selection and therefore the performance of the robust training. Table 6 indicates the impact of EM on CIFAR-10 dataset. It shows that the EM helps in improving the performance under different noise rates, and the significance is higher for high noise rates.", + "bbox": [ + 496, + 344, + 890, + 465 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of Entropic Regularization Coefficient $\\epsilon$ . The motivation for equipping the optimal transport with entropic regularization is to speed up its computation. However, the entropic regularization could potentially degrade the transport sparsity and therefore the quality of sample selection. To alleviate that, we propose sparsity regularization, whose effectiveness is examined in Table 5. Then, a by-product of sparsity regularization is the parameter robustness of the regularization coefficient $\\epsilon$ . As shown in Table 7, the performance of sample selection stays quite stable when varying $\\epsilon$ from 1 to 10.", + "bbox": [ + 496, + 465, + 890, + 630 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 645, + 617, + 660 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we study the problem of sample selection for learning with noisy labels. We propose the OT-Filter, a novel technique for retrieving clean samples, which can also be combined with existing semi-supervised learning techniques for handling noisy labels. Unlike previous works relying on a-priori knowledge or conforming to confirmation bias, the OT-Filter enhances the measurement of data discrepancy, by lifting data from Euclidean space to the probability space, and thus improves the quality of sample selection. Extensive experiments on synthetic and real-world noisy datasets are conducted to evaluate our proposals. The results show its effectiveness and superiority.", + "bbox": [ + 496, + 670, + 890, + 851 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. This work is supported by NSFC (No.61772492, 62072428) and the CAS Pioneer Hundred Talents Program. Xike Xie is the corresponding author.", + "bbox": [ + 496, + 851, + 890, + 897 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "16171", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Martial Agueh and Guillaume Carlier. Barycenters in the wasserstein space. SIAM Journal on Mathematical Analysis, 43(2):904-924, 2011. 2, 3, 7", + "[2] Ethan Anderes, Steffen Borgwardt, and Jacob Miller. Discrete Wasserstein barycenters: Optimal transport for discrete data. Mathematical Methods of Operations Research, 84(2):389-409, 2016. 3", + "[3] Eric Arazo, Diego Ortego, Paul Albert, Noel O'Connor, and Kevin McGuinness. Unsupervised label noise modeling and loss correction. In International conference on machine learning, pages 312-321. PMLR, 2019. 7", + "[4] Martin Arjovsky, Soumith Chintala, and Léon Bottou. Wasserstein generative adversarial networks. In International conference on machine learning, pages 214-223. PMLR, 2017. 2", + "[5] Yingbin Bai, Erkun Yang, Bo Han, Yanhua Yang, Jiatong Li, Yinian Mao, Gang Niu, and Tongliang Liu. Understanding and improving early stopping for learning with noisy labels. Advances in Neural Information Processing Systems, 34:24392-24403, 2021. 2", + "[6] Mokhtar S Bazaraa, Hanif D Sherali, and Chitharanjan M Shetty. Nonlinear programming: theory and algorithms. John Wiley & Sons, 2013. 4", + "[7] Jean-David Benamou, Guillaume Carlier, Marco Cuturi, Luca Nenna, and Gabriel Peyre. Iterative bregman projections for regularized transportation problems. SIAM Journal on Scientific Computing, 37(2):A1111-A1138, 2015. 3", + "[8] David Berthelot, Nicholas Carlini, Ian Goodfellow, Nicolas Papernot, Avital Oliver, and Colin A Raffel. Mixmatch: A holistic approach to semi-supervised learning. Advances in neural information processing systems, 32, 2019. 3, 5", + "[9] Dimitris Bertsimas and John N Tsitsiklis. Introduction to linear optimization, volume 6. 4", + "[10] Stephen Boyd, Stephen P Boyd, and Lieven Vandenberghe. Convex optimization. Cambridge university press, 2004. 4", + "[11] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2", + "[12] Yingyi Chen, Xi Shen, Shell Xu Hu, and Johan AK Suykens. Boosting co-teaching with compression regularization for label noise. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2688-2692, 2021. 7", + "[13] De Cheng, Tongliang Liu, Yixiong Ning, Nannan Wang, Bo Han, Gang Niu, Xinbo Gao, and Masashi Sugiyama. Instance-dependent label-noise learning with manifold-regularized transition matrix estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16630–16639, 2022. 1, 2", + "[14] Hao Cheng, Zhaowei Zhu, Xingyu Li, Yifei Gong, Xing Sun, and Yang Liu. Learning with instance-dependent label noise: A sample sieve approach. In International Conference on Learning Representations, 2021. 2", + "[15] Nicolas Courty, Rémi Flamary, Amaury Habrard, and Alain Rakotomamonjy. Joint distribution optimal transportation" + ], + "bbox": [ + 78, + 114, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "for domain adaptation. Advances in Neural Information Processing Systems, 30, 2017. 2", + "[16] Nicolas Courty, Rémi Flamary, and Devis Tuia. Domain adaptation with regularized optimal transport. In Joint European Conference on Machine Learning and Knowledge Discovery in Databases, pages 274-289. Springer, 2014. 2, 5", + "[17] Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. Advances in neural information processing systems, 26, 2013. 2, 3", + "[18] Marco Cuturi and Arnaud Doucet. Fast computation of wasserstein barycenters. In International conference on machine learning, pages 685-693. PMLR, 2014. 2, 3, 4", + "[19] Bharath Bhushan Damodaran, Rémi Flamary, Vivien Seguy, and Nicolas Courty. An entropic optimal transport loss for learning deep neural networks under label noise in remote sensing images. Comput. Vis. Image Underst., 191:102863, 2018. 2", + "[20] Arthur P Dempster, Nan M Laird, and Donald B Rubin. Maximum likelihood from incomplete data via the em algorithm. Journal of the Royal Statistical Society: Series B (Methodological), 39(1):1-22, 1977. 4", + "[21] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6", + "[22] Erik Englesson and Hossein Azizpour. Generalized jensen-shannon divergence loss for learning with noisy labels. Advances in Neural Information Processing Systems, 34:30284-30297, 2021. 2, 7", + "[23] Kilian Fatras, Bharath Bhushan Damodaran, Sylvain Lobry, Remi Flamary, Devis Tuia, and Nicolas Courty. Wasserstein adversarial regularization for learning with label noise. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 1", + "[24] Alessio Figalli and Federico Glaudo. An Invitation to Optimal Transport, Wasserstein Distances, and Gradient Flows. 2021. 2", + "[25] Aude Geneva, Gabriel Peyré, and Marco Cuturi. Learning generative models with sinkhorn divergences. In International Conference on Artificial Intelligence and Statistics, pages 1608-1617. PMLR, 2018. 2", + "[26] Wenshuo Guo, Nhat Ho, and Michael Jordan. Fast algorithms for computational optimal transport and wasserstein barycenter. In International Conference on Artificial Intelligence and Statistics, pages 2088-2097. PMLR, 2020. 3", + "[27] Bo Han, Quanming Yao, Xingrui Yu, Gang Niu, Miao Xu, Weihua Hu, Ivor Tsang, and Masashi Sugiyama. Coteaching: Robust training of deep neural networks with extremely noisy labels. Advances in neural information processing systems, 31, 2018. 1, 2, 6, 7", + "[28] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1, 6", + "[29] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Identity mappings in deep residual networks. In European conference on computer vision, pages 630-645. Springer, 2016. 1, 3, 6" + ], + "bbox": [ + 503, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "16172", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In International conference on machine learning, pages 448-456. pmlr, 2015. 6", + "[31] Ahmet Iscen, Jack Valmadre, Anurag Arnab, and Cordelia Schmid. Learning with neighbor consistency for noisy labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4672-4681, 2022. 1, 2", + "[32] Lu Jiang, Zhengyuan Zhou, Thomas Leung, Li-Jia Li, and Li Fei-Fei. Mentornet: Learning data-driven curriculum for very deep neural networks on corrupted labels. In International conference on machine learning, pages 2304–2313. PMLR, 2018. 1, 2", + "[33] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. Advances in neural information processing systems, 33:18661-18673, 2020. 2", + "[34] Taehyeon Kim, Jongwoo Ko, JinHwan Choi, Se-Young Yun, et al. Fine samples for learning with noisy labels. Advances in Neural Information Processing Systems, 34:24137–24149, 2021. 1, 2, 7", + "[35] Philip A Knight. The sinkhorn-knopp algorithm: convergence and applications. SIAM Journal on Matrix Analysis and Applications, 30(1):261-275, 2008. 2", + "[36] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 1, 5", + "[37] Kenneth Lange. MM optimization algorithms. SIAM, 2016. 5", + "[38] Junnan Li, Richard Socher, and Steven CH Hoi. Dividemix: Learning with noisy labels as semi-supervised learning. arXiv preprint arXiv:2002.07394, 2020. 1, 2, 5, 6, 7", + "[39] Sheng Liu, Jonathan Niles-Weed, Narges Razavian, and Carlos Fernandez-Granda. Early-learning regularization prevents memorization of noisy labels. Advances in Neural Information Processing Systems, 33, 2020. 2, 6", + "[40] Baharan Mirzasoleiman, Kaidi Cao, and Jure Leskovec. Coresets for robust training of deep neural networks against noisy labels. Advances in Neural Information Processing Systems, 33:11465-11477, 2020. 2, 7", + "[41] Kento Nishi, Yi Ding, Alex Rich, and Tobias Hollerer. Augmentation strategies for learning with noisy labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8022-8031, 2021. 2, 5", + "[42] Diego Ortego, Eric Arazo, Paul Albert, Noel E O'Connor, and Kevin McGuinness. Multi-objective interpolation training for robustness to label noise. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6606-6615, 2021. 2", + "[43] Giorgio Patrini, Alessandro Rozza, Aditya Krishna Menon, Richard Nock, and Lizhen Qu. Making deep neural networks robust to label noise: A loss correction approach. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1944-1952, 2017. 1, 2", + "[44] Gabriel Peyré, Marco Cuturi, et al. Computational optimal transport: With applications to data science. Foundations" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "and Trends® in Machine Learning, 11(5-6):355-607, 2019. 2, 3", + "[45] Vikas C Raykar, Shipeng Yu, Linda H Zhao, Gerardo Hermosillo Valadez, Charles Florin, Luca Bogoni, and Linda Moy. Learning from crowds. Journal of machine learning research, 11(4), 2010. 1", + "[46] Ievgen Redko, Nicolas Courty, Rémi Flamary, and Devis Tuia. Optimal transport for multi-source domain adaptation under target shift. In The 22nd International Conference on Artificial Intelligence and Statistics, pages 849-858. PMLR, 2019. 5", + "[47] Oren Rippel, Michael Gelbart, and Ryan Adams. Learning ordered representations with nested dropout. In International Conference on Machine Learning, pages 1746-1754. PMLR, 2014. 7", + "[48] Filippo Santambrogio. Optimal transport for applied mathematicians. Birkhäuser, NY, 55(58-63):94, 2015. 2, 3", + "[49] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 1, 6", + "[50] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. Advances in neural information processing systems, 33:596-608, 2020. 5", + "[51] Hwanjun Song, Minseok Kim, and Jae-Gil Lee. Selfie: Refurbishing unclean samples for robust deep learning. In International Conference on Machine Learning, pages 5907-5915. PMLR, 2019. 5, 6, 7, 8", + "[52] Fariborz Taherkhani, Ali Dabouei, Sobhan Soleymani, Jeremy Dawson, and Nasser M Nasrabadi. Transporting labels via hierarchical optimal transport for semi-supervised learning. In European Conference on Computer Vision, pages 509-526. Springer, 2020. 2", + "[53] Fariborz Taherkhani, Ali Dabouei, Sobhan Soleymani, Jeremy Dawson, and Nasser M Nasrabadi. Self-supervised Wasserstein pseudo-labeling for semi-supervised image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12267-12277, 2021. 2", + "[54] Cheng Tan, Jun Xia, Lirong Wu, and Stan Z Li. Co-learning: Learning from noisy labels with self-supervision. In Proceedings of the 29th ACM International Conference on Multimedia, pages 1405-1413, 2021. 2, 7", + "[55] Cédric Villani. Optimal transport: old and new, volume 338. Springer, 2009. 2, 3", + "[56] Cédric Villani. Topics in optimal transportation, volume 58. American Mathematical Soc., 2021. 2, 3", + "[57] Yikai Wang, Xinwei Sun, and Yanwei Fu. Scalable penalized regression for noise detection in learning with noisy labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 346-355, 2022. 1", + "[58] Pengxiang Wu, Songzhu Zheng, Mayank Goswami, Dimitris Metaxas, and Chao Chen. A topological filter for learning with label noise. Advances in neural information processing systems, 33:21382-21393, 2020. 1, 2, 7" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "16173", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[59] Xiaobo Xia, Tongliang Liu, Bo Han, Mingming Gong, Jun Yu, Gang Niu, and Masashi Sugiyama. Sample selection with uncertainty of losses for learning with noisy labels. In International Conference on Learning Representations, 2022. 2", + "[60] Tong Xiao, Tian Xia, Yi Yang, Chang Huang, and Xiaogang Wang. Learning from massive noisy labeled data for image classification. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2691-2699, 2015. 1, 5, 6", + "[61] Li Yi, Sheng Liu, Qi She, A Ian McLeod, and Boyu Wang. On learning contrastive representations for learning with noisy labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16682-16691, 2022. 2, 7", + "[62] Xingrui Yu, Bo Han, Jiangchao Yao, Gang Niu, Ivor Tsang, and Masashi Sugiyama. How does disagreement help generalization against label corruption? In International Conference on Machine Learning, pages 7164-7173. PMLR, 2019. 1, 2, 7", + "[63] Yikai Zhang, Songzhu Zheng, Pengxiang Wu, Mayank Goswami, and Chao Chen. Learning with feature-dependent label noise: A progressive approach. In Ninth International Conference on Learning Representations, volume 9, 2021. 7, 8", + "[64] Zhilu Zhang and Mert Sabuncu. Generalized cross entropy loss for training deep neural networks with noisy labels. Advances in neural information processing systems, 31, 2018. 2, 5", + "[65] Evgenii Zheltonozhskii, Chaim Baskin, Avi Mendelson, Alex M Bronstein, and Or Litany. Contrast to divide: Self-supervised pre-training for learning with noisy labels. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1657-1667, 2022. 2", + "[66] Yudian Zheng, Guoliang Li, Yuanbing Li, Caihua Shan, and Reynold Cheng. Truth inference in crowdsourcing: Is the problem solved? Proceedings of the VLDB Endowment, 10(5):541-552, 2017. 1" + ], + "bbox": [ + 78, + 90, + 470, + 626 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "16174", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/OT-Filter_ An Optimal Transport Filter for Learning With Noisy Labels/69b93138-511e-410f-9d28-563936aa9d2f_model.json b/2023/OT-Filter_ An Optimal Transport Filter for Learning With Noisy Labels/69b93138-511e-410f-9d28-563936aa9d2f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2cc45811f3b1937f567876369efab605dc2f096f --- /dev/null +++ b/2023/OT-Filter_ An Optimal Transport Filter for Learning With Noisy Labels/69b93138-511e-410f-9d28-563936aa9d2f_model.json @@ -0,0 +1,2521 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.125, + 0.131, + 0.846, + 0.153 + ], + "angle": 0, + "content": "OT-Filter: An Optimal Transport Filter for Learning with Noisy Labels" + }, + { + "type": "text", + "bbox": [ + 0.138, + 0.18, + 0.832, + 0.252 + ], + "angle": 0, + "content": "Chuanwen Feng† Yilong Ren† Xike Xie \nUniversity of Science and Technology of China \nData Darkness Lab, MIRACLE Center, Suzhou Institute for Advanced Research, USTC \n{chuanwen,ylren}@mail.ustc.edu.cn, {xkxie}@ustc.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.287, + 0.314, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.319, + 0.473, + 0.637 + ], + "angle": 0, + "content": "The success of deep learning is largely attributed to the training over clean data. However, data is often coupled with noisy labels in practice. Learning with noisy labels is challenging because the performance of the deep neural networks (DNN) drastically degenerates, due to confirmation bias caused by the network memorization over noisy labels. To alleviate that, a recent prominent direction is on sample selection, which retrieves clean data samples from noisy samples, so as to enhance the model's robustness and tolerance to noisy labels. In this paper, we revamp the sample selection from the perspective of optimal transport theory and propose a novel method, called the OT-Filter. The OT-Filter provides geometrically meaningful distances and preserves distribution patterns to measure the data discrepancy, thus alleviating the confirmation bias. Extensive experiments on benchmarks, such as Clothing1M and ANIMAL-10N, show that the performance of the OT-Filter outperforms its counterparts. Meanwhile, results on benchmarks with synthetic labels, such as CIFAR-10/100, show the superiority of the OT-Filter in handling data labels of high noise." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.668, + 0.21, + 0.683 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.694, + 0.471, + 0.875 + ], + "angle": 0, + "content": "Deep learning has achieved great success on a flurry of emerging applications, such as [28, 29, 32, 49]. It is believed that the phenomenal achievement of deep learning is largely attributed to accurate labels. However, the inaccuracy or imprecision of labels is inherent in real-world datasets, the so-called noisy label challenge. One way to alleviate that is to collect labels from internet queries over data-level tags, but the performance of deep neural networks (DNN) suffers drastically from the inaccuracy of such labels [27, 32]. A higher quality of data labels can be obtained by employing human workers, but the seemingly \"ground truth annotations\" inevitably involve human biases or mis" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.288, + 0.892, + 0.319 + ], + "angle": 0, + "content": "takes [45,66]. More, the human annotation is expensive and time-consuming, especially for large-scale datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.321, + 0.893, + 0.579 + ], + "angle": 0, + "content": "There have been many works on handling noisy labels, such as regularization [23] and transition matrix [13, 43]. The regularization approach leverages the regularization bias to overcome the label noise issue. But the regularization bias is permanent [62], thus overfitting models to noisy labels. The transition matrix approach assumes that the transition probabilities between clean and noisy labels are fixed and independent of data samples. However, a quality label transition matrix is hard to be estimated, especially when the number of classes is big, making it fall short in handling noisy real-world datasets, such as [60] and [36]. A recent prominent direction is on adopting sample selection [27, 38, 58] for enhancing the label quality, by selecting clean samples from the noisy training dataset. In general, existing literatures in sample selection can be grouped to two categories, co-training networking [27, 62] and criterion-based filtering [31, 34, 57, 58]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.581, + 0.895, + 0.763 + ], + "angle": 0, + "content": "The former utilizes the memorization of DNNs and multiple networks (e.g., co-teaching [27] and its variants [38, 62]) to filter label noise with small loss trick, so that a small set of clean samples are used as training examples. Letting alone the high training overhead of multiple networks, the disadvantages are two-fold: 1) it may require the a-priori knowledge of noisy rates to select the specified proportion of small loss samples as clean samples; 2) the small-loss trick is not tolerant to the error accumulation of network training once a clean sample is falsely recognized, the so-called confirmation bias, especially for labels with high noise where clean and noisy samples largely overlap." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.895, + 0.903 + ], + "angle": 0, + "content": "The latter alleviates the problem by setting a specific criterion. Mostly, existing works [31] [58] adopt Euclidean distances for measuring the similarity between data samples. Distance-based filtering iteratively explores the neighborhood of the feature representation and infers/cleans sample labels by aggregating the information from their neighborhoods. Despite the simplicity, distance-based filtering is insufficient to address noisy labels, especially when the label noise is high, e.g., overlapped label classes." + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.21, + 0.901 + ], + "angle": 0, + "content": "\\(\\dagger\\) Equal Contribution." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "16164" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.081, + 0.092, + 0.468, + 0.333 + ], + "angle": 0, + "content": "In this paper, we revamp the sample selection from the perspective of optimal transport [56] and propose a novel filtering method, called the OT-Filter. In light of the optimal transport, we construct discrete probability measures over the sample features, which lifts the Euclidean space of feature vectors to a probability space. It thus enables a geometric way of measuring the discrepancy between probability measures, representing the corresponding sample features. In addition to the distance-based metric in the Euclidean space [31, 58], the OT-Filter also captures the distribution information in the probability space so as to alleviate the confirmation bias. Accordingly, a clean representation can be obtained for each class in the probability space. By optimizing the transport plan from a sample to a clean representation, one can better determine if a sample is clean or noisy, thus improving the quality of sample selection." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.334, + 0.468, + 0.485 + ], + "angle": 0, + "content": "In general, the merits of the OT-Filter can be summarized as follows. First, it does not require any a-priori knowledge about the noisy rate of a dataset. Second, it utilizes the optimal transport which provides geometrically meaningful distances to exploit the sample discrepancy yet preserving the distribution patterns in corresponding probability space, making the sample selection of high quality and theoretical support. Third, it can be plugged to existing robust training paradigms, e.g., supervised and semi-supervised robust training." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.487, + 0.468, + 0.636 + ], + "angle": 0, + "content": "We conduct extensive experiments with a series of synthetic and real datasets to gain insights into our proposals. The result shows that the OT-Filter follows state-of-the-art (SOTA) [38] when the noise rate is low, and dominates SOTA when the noise rate is high. For instance, the OT-Filter achieves about \\(14\\%\\) and \\(12\\%\\) higher accuracy than SOTA, in the presence of \\(90\\%\\) noise rate, on synthetic datasets CIFAR-10 and CIFAR-100, respectively. Moreover, the OT-Filter outperforms the competitors on real datasets Clothing1M and ANIMAL-10N." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.639, + 0.468, + 0.729 + ], + "angle": 0, + "content": "The rest of the paper is organized as follows. We review the existing literature in Section 2. In Section 3, we present preliminaries of optimal transport. We investigate our proposed OT-Filter in Section 4. Furthermore, we conduct extensive empirical studies in Section 5 and conclude the paper in Section 6." + }, + { + "type": "title", + "bbox": [ + 0.081, + 0.746, + 0.217, + 0.761 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.081, + 0.771, + 0.317, + 0.786 + ], + "angle": 0, + "content": "2.1. Learning with Noisy Label" + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.795, + 0.468, + 0.9 + ], + "angle": 0, + "content": "A flurry of research methods [5, 13, 14, 27, 40, 54] was proposed for learning with noisy labels, which can be roughly divided into two categories, robust model [13, 22, 39, 43, 54, 61, 64] and clean sample selection [27, 32, 62]. The clean sample selection aims to select a clean subset from a noisy dataset [14, 27, 38]. One way is to train multiple networks to filter noisy samples. For instance, [32]" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.092, + 0.89, + 0.302 + ], + "angle": 0, + "content": "proposed to pre-train an extra network to select clean samples for the main network. The inferiority of this method is on error accumulation. To relieve that, Han et al. [27] proposed Co-teaching that maintains two networks simultaneously, where one network selects clean samples for the other network by small loss trick. However, as the increase of training epochs, Co-teaching would converge to consensus gradually. Then, \\(\\mathrm{Co - teaching}^{+}\\) [62] was proposed to keep two networks diverged. To avoid confirmation bias, DivMix [38] proposed a hybrid framework with semi-supervised training, achieving the state-of-the-art performance. Based on DivMix, [41, 65] modified the training schema, such as self-supervised training [11, 33], to boost the performance." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.304, + 0.89, + 0.469 + ], + "angle": 0, + "content": "The other research line of sample selection is to filter noisy samples based on specific criteria [31, 34, 42, 59]. For example, Xia et al. [59] proposed to reduce the uncertainty of loss using specific strategies. Iscen et al. [31] proposed to leverage similarities between training examples in the feature space, encouraging the prediction of each example to be close to its nearest neighbors. Also, in Euclidean space, Wu et al. [58] proposed the TopoFilter that filters noisy samples using k-nearest neighborhood distance between prelogits. Fine [34] proposed to use the principal components of latent representations to select clean samples." + }, + { + "type": "title", + "bbox": [ + 0.504, + 0.483, + 0.681, + 0.499 + ], + "angle": 0, + "content": "2.2. Optimal Transport" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.507, + 0.89, + 0.779 + ], + "angle": 0, + "content": "The optimal transport problem [55, 56] aims to move mass from a probability measure to another probability measure at a minimum cost. It defines the Wasserstein distance [24] and provides a geometric way to measure the discrepancy between probability measures. One factor that limits the wide application of optimal transport is the high computational cost. To improve its scalability, Cuturi [17] proposed an entropic regularization for the transport plan, which yields an efficient algorithm with the matrix scaling method of Sinkhorn-Knopp [35]. Recently, computational optimal transport [44, 48] has found many applications in various areas, e.g., generative models [4, 25], domain adaptation [15, 16], and semi-supervised learning [52, 53]. In particular, [1] studied the barycenter in Wasserstein space. [18] explored the efficient computing of Wasserstein barycenter. [19] proposed to use entropic optimal transport loss, based on joint distribution optimal transport [15], to build a robust training." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.781, + 0.89, + 0.9 + ], + "angle": 0, + "content": "In this paper, we study the problem of clean sample selection from the perspective of optimal transport to combat label noise. The proposed method can potentially capture geometric information from the probability space. Unlike the mechanism of multiple network training, it only trains a single network without a-priori information on noise rates. Moreover, it can be easily plugged to multiple robust training paradigms." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.947, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16165" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.09, + 0.203, + 0.108 + ], + "angle": 0, + "content": "3. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.115, + 0.469, + 0.251 + ], + "angle": 0, + "content": "In this section, we present preliminaries on optimal transport, making a basis for subsequent proposed techniques. The optimal transport [56] is to seek an optimal transport plan between two measures at a minimal cost (e.g. the Wasserstein distance), which provides a geometric way of matching probability measures. Here we first introduce the general definition of Wasserstein distance induced by optimal transport problem. More details about optimal transport can be found in [44, 48, 55]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.251, + 0.469, + 0.328 + ], + "angle": 0, + "content": "Wassertein Distance. Let \\( \\mathcal{S} \\) be a locally complete and separable metric space, the \\( \\mathcal{P}(S) \\) be the Borel probability measures set on \\( \\mathcal{S} \\). For any \\( \\mathcal{X}, \\mathcal{Y} \\subset \\mathcal{S} \\), given \\( \\mu \\in \\mathcal{P}(\\mathcal{X}) \\) and \\( \\nu \\in \\mathcal{P}(\\mathcal{Y}) \\), the optimal transport defines a Wasserstein distance between two probability measures, denoted as" + }, + { + "type": "equation", + "bbox": [ + 0.107, + 0.333, + 0.438, + 0.371 + ], + "angle": 0, + "content": "\\[\nW _ {p} (\\mu , \\nu) := \\left(\\inf _ {\\pi (\\mu , \\nu)} \\int_ {\\mathcal {X} \\times \\mathcal {Y}} | | x - y | | ^ {p} d \\pi (\\mu , \\nu)\\right) ^ {\\frac {1}{p}}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.376, + 0.469, + 0.406 + ], + "angle": 0, + "content": "\\(p \\geq 1\\), where the \\(\\pi(\\mu, \\nu)\\) is the set of joint probability measures with marginal \\(\\mu\\) and \\(\\nu\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.413, + 0.327, + 0.43 + ], + "angle": 0, + "content": "3.1. Discrete Optimal Transport" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.436, + 0.469, + 0.587 + ], + "angle": 0, + "content": "Let \\(\\Delta_{n} = \\{\\mathbf{a}\\in \\mathbb{R}_{+}^{n}|\\sum_{i = 1}^{n}a_{i} = 1,\\forall a_{i}\\geq 0\\}\\) be a probability simplex in dimension \\(n\\) . Consider two empirical probability measures \\(\\mu = \\sum_{i = 1}^{n}a_{i}\\delta_{x_{i}}\\) and \\(\\nu = \\sum_{j = 1}^{m}b_{j}\\delta_{y_{j}}\\) , defined on metric space \\(\\mathcal{X}\\) with support \\(\\{x_i\\}_{i = 1}^n\\) and \\(\\mathcal{V}\\) with support \\(\\{y_j\\}_{j = 1}^m\\) respectively. Here the weight vector \\(\\mathbf{a} = (a_1,a_2,\\dots,a_n)\\) and \\(\\mathbf{b} = (b_{1},b_{2},\\dots,b_{m})\\) live in \\(\\Delta_{n}\\) and \\(\\Delta_{m}\\) , respectively. The \\(\\delta\\) stands for the Dirac unit mass function. Given a transport cost \\(c:\\mathcal{X}\\times \\mathcal{Y}\\to \\mathbb{R}_+\\) the discrete optimal transport between probability measures \\(\\mu\\) and \\(\\nu\\) can be formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.164, + 0.593, + 0.469, + 0.637 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} W _ {p} ^ {p} (\\mu , \\nu) := \\min _ {M \\in \\Pi (\\mu , \\nu)} \\langle C, M \\rangle_ {F} \\tag {1} \\\\ s. t. M \\mathbf {1} _ {m} = \\mu , M ^ {T} \\mathbf {1} _ {n} = \\nu \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.645, + 0.469, + 0.752 + ], + "angle": 0, + "content": "where \\(C \\in \\mathbb{R}_+^{n \\times m}\\) is the transport cost matrix, and \\(c_{ij}\\) represents a unit transport cost from \\(x_i\\) to \\(y_j\\). The \\(M \\in \\mathbb{R}_+^{n \\times m}\\) is transport plan in which \\(m_{ij}\\) denotes the amount of mass transported from \\(x_i\\) to \\(y_j\\). All feasible transport plans constitute transport polytope \\(\\Pi(\\mu, \\nu)\\). The \\(\\langle C, M \\rangle_F\\) is the Frobenius inner product of matrices and equals to \\(tr(C^T M)\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.759, + 0.357, + 0.775 + ], + "angle": 0, + "content": "3.2. Regularized Optimal Transport" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.782, + 0.469, + 0.871 + ], + "angle": 0, + "content": "The discrete optimal transport formulation, in essence, is a convex optimization problem. More precisely, it's a linear programming problem. Unfortunately, this linear programming problem has a cubic computing complexity. A way to relieve this is to leverage entropic regularization [17] formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.174, + 0.88, + 0.469, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\min _ {M \\in \\Pi (\\mu , \\nu)} \\langle C, M \\rangle_ {F} - \\epsilon H (M) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "where \\(\\epsilon > 0\\) is the regularization coefficient, and \\(H(M)\\) is the entropic regularization term, which promotes an efficient computation for transportation via the matrix scaling algorithm [17], given by" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.163, + 0.821, + 0.196 + ], + "angle": 0, + "content": "\\[\nH (M) := - \\sum_ {i j} M _ {i j} \\left(\\log \\left(M _ {i j}\\right) - 1\\right)\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.207, + 0.719, + 0.223 + ], + "angle": 0, + "content": "3.3. Wasserstein Barycenter" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.23, + 0.892, + 0.291 + ], + "angle": 0, + "content": "Given \\(N\\) probability measures \\(\\{\\nu_{1},\\nu_{2},\\dots,\\nu_{N}\\}\\), \\(\\nu_{i} \\in \\mathcal{P}(S)\\), each of which has finite supports and second moments. A Wasserstein barycenter of these measures is a probability measure \\(\\mu\\), satisfying:" + }, + { + "type": "equation", + "bbox": [ + 0.509, + 0.301, + 0.881, + 0.342 + ], + "angle": 0, + "content": "\\[\n\\mu := \\inf \\sum_ {i = 1} ^ {N} \\lambda_ {i} W _ {2} ^ {2} (\\nu_ {i}, \\mu), \\quad s. t. \\sum_ {i = 1} ^ {N} \\lambda_ {i} = 1, \\forall \\lambda_ {i} \\geq 0\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.352, + 0.892, + 0.49 + ], + "angle": 0, + "content": "The notion was first proposed by [1], where some elegant properties of Wasserstein barycenter were presented. Then [2] discussed it on the discrete case and showed that the problem of finding Wasserstein barycenter over the space \\(\\mathcal{P}(\\mathcal{S})\\) can be reduced to a simpler space \\(\\mathcal{O}_r(S)\\), where \\(r = \\sum_{i=1}^{N} e_i - N + 1\\), and \\(e_i\\) is the number of components of \\(\\nu_i, i \\in [1, N]\\). To find Wasserstein barycenter over space \\(\\mathcal{O}_r(S)\\), a set of efficient algorithms [7, 18, 26] were proposed." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.504, + 0.598, + 0.518 + ], + "angle": 0, + "content": "4. Methods" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.528, + 0.61, + 0.542 + ], + "angle": 0, + "content": "4.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.551, + 0.892, + 0.642 + ], + "angle": 0, + "content": "In this section, we study the OT-Filter, the proposed method based on optimal transport for learning with noisy labels in a semi-supervised paradigm. The key idea of our proposed method is to transport samples with noisy labels to samples with clean labels at a minimum cost, which makes sample selection of high quality and theoretical support." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.642, + 0.892, + 0.732 + ], + "angle": 0, + "content": "The mechanism of the OT-Filter, consists of two phases, representation phase (Section 4.2) and transportation phase (Section 4.3). Then, the OT-Filter can be plugged to off-the-shelf SSLs for robust training, e.g., MixMatch [8], by treating clean samples as labeled data and noisy samples as unlabeled samples (Section 4.4)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.742, + 0.891, + 0.773 + ], + "angle": 0, + "content": "4.2. Representation Phase: Finding Clean Representations via Wasserstein Barycenters" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In the representation phase, we first use a pre-trained network, e.g., ResNet [29], to extract a set of feature representations of the labeled samples. Then, we define a discrete probability measure over the feature representations of each class to find the clean representations (or prototypes) via Wasserstein barycenters. Since there are mistrabeled samples, the barycenters we obtained may be noisy. Therefore, we iteratively optimize the barycenters with the expectation" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16166" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.136 + ], + "angle": 0, + "content": "maximization (EM) algorithm [20] to converge to clean representations. The details about the representation phase are covered in Algorithm 1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.137, + 0.468, + 0.304 + ], + "angle": 0, + "content": "Feature Extraction. The purpose of the feature extraction is to transform the input noisy label dataset into a feature-label dataset via a network. A noisy label dataset of \\(N\\) samples can be represented by \\(\\hat{\\mathcal{D}} = \\{(\\mathbf{x}_i,\\hat{y}_i)\\}_{i = 1}^N\\), where \\(\\mathbf{x}_i\\in \\mathbb{R}^d\\) denotes the \\(i\\)-th sample and \\(\\dot{y}_i\\in \\{0,1\\} ^K\\) denotes the corresponding noisy label over \\(K\\) classes. Then, a feature extraction network \\(f(\\omega ,\\mathbf{x}_i):\\mathbb{R}^d\\to \\mathbb{R}^m\\) with parameters \\(\\omega\\) maps a sample \\(\\mathbf{x}_i\\) to a \\(m\\)-dimensional feature representation, denoted as \\(\\tilde{\\mathbf{x}}_i\\in \\mathbb{R}^m\\). Therefore, all output feature representations and their corresponding labels constitute a feature-label set, denoted as \\(\\tilde{D} = \\{(\\tilde{\\mathbf{x}}_i,\\hat{y}_i)\\}_{i = 1}^N\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.303, + 0.468, + 0.469 + ], + "angle": 0, + "content": "Probability Measure Modeling. The purpose of probability measure modeling is to define a probability measure over the feature-label set. This definition will potentially lift a Euclidean space \\(\\mathcal{D}\\), in which the feature representations originally live, to a probability space, abbreviated as \\(\\mathcal{P}(\\mathcal{D})\\). Assume that the labels are of \\(K\\) classes. First, we split the feature-label set \\(\\tilde{D}\\) into \\(K\\) subsets, i.e., \\(\\{s_1,s_2,\\dots,s_K\\}\\) according to the label class. For one feature representation \\(\\tilde{\\mathbf{x}}_i\\) in class \\(k\\in [1,K]\\), we represent it as \\(\\tilde{\\mathbf{x}}_i^k,i\\in [1,|s_k|]\\). Then, all feature representations in subset \\(s_k\\) construct a probability measure, denoted as" + }, + { + "type": "equation", + "bbox": [ + 0.214, + 0.481, + 0.332, + 0.523 + ], + "angle": 0, + "content": "\\[\n\\mathcal {Q} = \\frac {1}{| s _ {k} |} \\sum_ {i = 1} ^ {| s _ {k} |} \\delta_ {\\tilde {\\mathbf {x}} _ {i} ^ {k}}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.536, + 0.468, + 0.566 + ], + "angle": 0, + "content": "where the \\(\\delta_{\\widetilde{\\mathbf{x}}_i^k}\\) is a Dirac unit mass on \\(\\widetilde{\\mathbf{x}}_i^k\\), and for simplicity, we use uniform weights." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.567, + 0.468, + 0.627 + ], + "angle": 0, + "content": "Clean Representation Retrieval. The problem of finding the representation of any subset \\( s_k \\) is equivalent to finding the corresponding Wasserstein barycenter \\( \\mathcal{B}_k \\) with finite supports, formalized as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.638, + 0.468, + 0.665 + ], + "angle": 0, + "content": "\\[\n\\mathcal {B} _ {k} := \\inf _ {\\mathcal {B} \\in \\mathcal {O} _ {r} (\\tilde {\\mathcal {D}})} W _ {2} ^ {2} (\\mathcal {B}, \\mathcal {Q}) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.677, + 0.468, + 0.723 + ], + "angle": 0, + "content": "where the \\(W\\) is the Wasserstein distance, and \\(r = |s_k|\\). As mentioned in Section 3, from the perspective of linear programming, the above optimization admits the dual:" + }, + { + "type": "equation", + "bbox": [ + 0.152, + 0.733, + 0.395, + 0.759 + ], + "angle": 0, + "content": "\\[\nD (\\alpha , \\beta) = \\max _ {(\\alpha , \\beta) \\in R (C)} \\alpha^ {T} \\mathcal {B} + \\beta^ {T} \\mathcal {Q}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.77, + 0.356, + 0.785 + ], + "angle": 0, + "content": "where the polyhedron of dual variables is:" + }, + { + "type": "equation", + "bbox": [ + 0.085, + 0.797, + 0.46, + 0.814 + ], + "angle": 0, + "content": "\\[\nR (C) = \\left\\{\\left(\\alpha , \\beta\\right) \\in \\mathbb {R} ^ {r} \\times \\mathbb {R} ^ {r} \\mid \\alpha_ {i} + \\beta_ {j} \\leq C _ {i j} \\wedge i, j \\in [ r ] \\right\\}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.468, + 0.901 + ], + "angle": 0, + "content": "By means of duality theory [6,9,10], we can find the relation of solution between the primal and dual problem. Here, the optimal solution of \\( D(\\alpha, \\beta) \\) is a subgradient of \\( W_2^2(\\mathcal{Q}, \\mathcal{B}) \\) with respect to \\( \\mathcal{B} \\). Thus, we employ a simple projected subgradient [18] to optimize Equation 3. over \\( \\mathcal{B} \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.212 + ], + "angle": 0, + "content": "Despite we have found a set of Wasserstein barycenters \\(\\{\\mathcal{B}_i\\}_{i=1}^K\\), the barycenters may be noisy since there are mislabeled data. Moreover, the clean barycenters are unknown. Therefore, for each class, we consider the ideal and clean barycenter \\(\\mathcal{B}_c\\) as a hidden variable, the obtained noisy barycenter \\(\\mathcal{B}\\) as an observed variable. Then, the problem of finding clean barycenter \\(\\mathcal{B}_c\\) is to maximize the following log-likelihood:" + }, + { + "type": "equation", + "bbox": [ + 0.585, + 0.224, + 0.806, + 0.256 + ], + "angle": 0, + "content": "\\[\n\\hat {\\theta} \\leftarrow \\arg \\max _ {\\theta} \\log \\sum_ {\\mathcal {B} _ {c}} p (\\mathcal {B} _ {c}, \\mathcal {B} | \\theta)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.268, + 0.892, + 0.388 + ], + "angle": 0, + "content": "where the \\(\\theta\\) is the parameters of probabilistic model. By this way, we can optimize the noisy barycenters iteratively with the EM algorithm to converge to clean barycenters. In E-step, we construct optimal transport between all samples and clean barycenters, thereby we can infer the labels of all samples. In M-step, we refine the barycenters with inferred label obtained from E-step. The overall computing details see Algorithm 1." + }, + { + "type": "code_caption", + "bbox": [ + 0.5, + 0.404, + 0.892, + 0.433 + ], + "angle": 0, + "content": "Algorithm 1 Find Clean Representation via Wasserstein Barycenter" + }, + { + "type": "algorithm", + "bbox": [ + 0.501, + 0.436, + 0.892, + 0.791 + ], + "angle": 0, + "content": "Require: \\(\\mathcal{Q} = \\{\\tilde{\\mathbf{x}}_i\\}_{i=1}^{|s_k|} \\in \\mathbb{R}^{m \\times r}\\), regularization coefficient \\(\\epsilon\\), \\(\\mathcal{B} \\in \\mathbb{R}^{m \\times e}\\) \n1: Initialize: \\(\\mathcal{B}\\), \\(\\nu \\gets \\mathcal{B}\\), \\(\\mu \\gets \\mathcal{Q}\\), \\(\\eta = 1/2\\), \\(t = 2\\) \n2: while \\(\\mathcal{B}\\) not converged do \n3: set \\(\\nu = \\mathbf{1}_e / e\\), \\(\\mu = \\mathbf{1}_r / r\\), \\(\\hat{\\nu} = \\tilde{\\nu} = \\mathbf{1} / e\\) \n4: while \\(\\nu\\) not converged do \n5: \\(\\beta \\gets (t + 1) / 2\\), \\(\\nu \\gets (1 - \\beta^{-1})\\hat{\\nu} + \\beta^{-1}\\tilde{\\nu}\\) \n6: \\(\\alpha \\gets\\) optimizer of \\(\\mathbf{d}(\\nu, \\mu, C)\\) \n7: \\(\\tilde{\\nu} \\gets \\tilde{\\nu} \\circ \\exp(-\\beta\\alpha)\\), \\(\\tilde{\\nu} \\gets \\tilde{\\nu} / \\tilde{\\nu}^T\\mathbf{1}\\) \n8: \\(\\hat{\\nu} \\gets (1 - \\beta^{-1})\\hat{\\nu} + \\beta^{-1}\\tilde{\\nu}\\) \n9: \\(t \\gets t + 1\\) \n10: end while \n11: \\(\\nu \\gets \\hat{\\nu}\\) \n12: \\(\\mathcal{M} \\gets\\) optimizer of \\(\\mathbf{p}(\\nu, \\mu, C)\\) \n13: \\(\\mathcal{B} \\gets (1 - \\eta)\\mathcal{B} + \\eta (\\mathcal{QM})diag(\\nu^{-1})\\) \n14: end while \n15: while \\(\\mathcal{B}_c\\) not converged do \n16: E-Step: \n17: \\(\\mathbb{L}_{\\hat{\\mathcal{D}}} \\gets \\mathrm{OT}(\\mu, \\nu, C)\\) \n18: M-Step: \n19: \\(\\mathcal{B}_c \\gets\\) refine \\(\\mathcal{B}\\) with \\(\\mathbb{L}_{\\hat{\\mathcal{D}}}\\) \n20: end while \n21: Return \\(\\mathcal{B}_c\\)" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.817, + 0.891, + 0.849 + ], + "angle": 0, + "content": "4.3. Transportation Phase: Transporting Noisy Labels via Regularized Optimal Transport" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In the transportation phase, noisy feature representations are transported to clean representations. In particular, we align all feature representations, including those of noisy" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16167" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.287 + ], + "angle": 0, + "content": "samples, to the clean representations (obtained in the representation phase) via sparsity regularized optimal transport. This operation potentially detects the corrupted labels based on the result of the optimal transport. Here, the mechanism of our sample selection via optimal transport is described as follows. First, we align the clean barycenters to the all rest samples via optimal transport. Based on the result of optimal transport, we can infer the labels of the all rest samples. We consider the route with the max mass transport that forms a coupling and then assign the label of a clean barycenter to the samples on this route. Second, we select the sample whose inferred label is identical to the original label as a clean sample." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.288, + 0.473, + 0.363 + ], + "angle": 0, + "content": "Sparsity Regularization. After the representation phase, clean representations are obtained, which are essentially the barycenters of label classes. The clean representation of a class provides the guidance of identifying whether samples are clean or noisy w.r.t. the given class." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.364, + 0.473, + 0.545 + ], + "angle": 0, + "content": "We consider the barycenters of all the \\(K\\) classes as a discrete probability measure, denoted as \\(\\mu = \\frac{1}{K}\\sum_{k = 1}^{K}\\delta_{u_k}\\) where \\(\\delta_{\\mu_k}\\) is a Dirac unit mass on \\(\\mu_{k}\\). Accordingly, the feature representations of the entire dataset construct a discrete probability measure \\(\\nu = \\frac{1}{N}\\sum_{i = 1}^{N}\\delta_{\\tilde{\\mathbf{x}}_i}\\). Then, we transport feature representations \\(\\nu\\) of all samples to the clean barycenters \\(\\mu\\), in order to detect if a sample is clean or mislabeled. Inspired by [16,46], we view the clean barycenters as the source domain and the feature representations of samples as the target domain. Then, we recast the alignments as an optimal transport problem with a sparsity regularization [16], written as" + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.553, + 0.47, + 0.588 + ], + "angle": 0, + "content": "\\[\n\\min _ {M \\in \\Pi (\\mu , \\nu)} \\langle C, M \\rangle_ {F} - \\epsilon H (M) + \\gamma \\sum_ {j, c} | | M (i _ {c}, j) | | _ {1} ^ {\\frac {1}{2}} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.597, + 0.473, + 0.809 + ], + "angle": 0, + "content": "The second term is an entropic regularization, as described in Equation 2. The third term is a sparsity regularization with coefficient \\(\\gamma\\), where \\(i_c\\) index the line if its element in class \\(c\\), and \\(M(i_c, j)\\) is a vector consisting of the \\(j\\)-th column of \\(M\\) in class \\(c\\). The \\(\\left\\| \\cdot \\right\\|_1^{1/2}\\) denotes an \\(L_1\\) norm with the power of \\(\\frac{1}{2}\\). The entropic regularization enables discrete optimal transport to be efficient. However, it disperses the transport route, which negatively affects sample selection. Therefore, to improve the quality of sample selection, we employ a sparsity regularization. The sparsity regularization promotes a feature representation that would be aligned to one of the clean barycenters and penalizes transportation matrix \\(M\\) that aligns together feature representations with different labels." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Optimization. Despite the consistent sparsity achieved by the regularization, the objective function of Equation 4 is non-convex. Moreover, the \\(L_{1}\\) norm with the power in label regularization term is a concave function. A common optimization strategy [37] is to construct a convex upper bound, represented by convex functions, over the non-convex prob" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "lem. By doing so, we can use the optimal solution of convex upper bound to approximate the optimal solution of the original non-convex problem. If applying the simplest linear approximation, we have" + }, + { + "type": "equation", + "bbox": [ + 0.579, + 0.162, + 0.816, + 0.198 + ], + "angle": 0, + "content": "\\[\n\\sum_ {j, c} \\left| \\left| M \\left(i _ {c}, j\\right) \\right| \\right| _ {1} ^ {\\frac {1}{2}} \\leq \\langle G, M \\rangle_ {F} + \\mathcal {L}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.209, + 0.894, + 0.257 + ], + "angle": 0, + "content": "where matrix \\(G\\) is \\(\\frac{1}{2} (||\\hat{M}(i_c, j) + \\delta||^{-\\frac{1}{2}})\\), and \\(\\hat{M}\\) is a given start point. The steps of optimization process are depicted as Algorithm 2." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.271, + 0.892, + 0.302 + ], + "angle": 0, + "content": "Algorithm 2 Transportation: Transport Noisy Label via Regularized Optimal Transport" + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.305, + 0.81, + 0.321 + ], + "angle": 0, + "content": "Require: optimal cost \\( C_{\\text{min}} \\) from Equation 1," + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.321, + 0.645, + 0.335 + ], + "angle": 0, + "content": "1: Initialize: \\( G = 0 \\)" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.336, + 0.704, + 0.35 + ], + "angle": 0, + "content": "2: while G not converged do" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.351, + 0.69, + 0.366 + ], + "angle": 0, + "content": "3: \\( C_{min} \\gets C_{min} + G \\)" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.367, + 0.828, + 0.381 + ], + "angle": 0, + "content": "4: \\(M\\gets\\) optimizer of Equation 2 with \\(C_{min}\\)" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.382, + 0.716, + 0.396 + ], + "angle": 0, + "content": "5: \\(G\\gets\\) update \\(G\\) using \\(M\\)" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.397, + 0.6, + 0.41 + ], + "angle": 0, + "content": "6: end while" + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.321, + 0.828, + 0.41 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.439, + 0.719, + 0.455 + ], + "angle": 0, + "content": "4.4. Robust Training in SSL" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.462, + 0.892, + 0.674 + ], + "angle": 0, + "content": "After all feature representations are transported to the clean barycenters. We filter noisy samples and obtain two data subsets, namely the clean sample set and noisy sample set. As aforementioned, our sample selection method is flexible enough for supporting various robust training paradigms, e.g., the robust supervised training [64] and the robust semi-supervised learning [50]. Following [38], we select MixMatch [8] with data augmentation strategy [41] as the practice of robust semi-supervised training but with fewer operations. We admit the labels of clean samples and neglect the labels of noisy samples, and then we use labeled data and unlabeled data for training in a semi-supervised learning fashion. The overall training process is described as Algorithm 3." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.687, + 0.634, + 0.705 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.712, + 0.602, + 0.727 + ], + "angle": 0, + "content": "5.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.795 + ], + "angle": 0, + "content": "We perform extensive experiments on four benchmark datasets: CIFAR-10 [36] and CIFAR-100 [36] with synthetic label noise, and Clothing1M [60] and ANIMAL-10N [51] with real label noise." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "CIFAR10/100. We evaluate the OT-Filter on CIFAR10 and CIFAR-100 [36] with synthetic noise. Both datasets contain 60,000 \\(32 \\times 32\\) color images, in 10 and 100 classes, respectively. From both datasets, 50K images are used for training, and 10K images are used for testing. Since noise characteristics can hardly be determined in advance, synthetic noise is commonly taken for controlling the noise rate" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16168" + } + ], + [ + { + "type": "code_caption", + "bbox": [ + 0.08, + 0.092, + 0.327, + 0.106 + ], + "angle": 0, + "content": "Algorithm 3 Robust Training in SSL" + }, + { + "type": "algorithm", + "bbox": [ + 0.08, + 0.111, + 0.47, + 0.392 + ], + "angle": 0, + "content": "Require: network parameter \\(\\omega\\), Beta distribution parameter \\(\\alpha\\), weight of unlabeled loss \\(\\lambda_{\\mathcal{U}}\\), batch size \\(B\\), number of augmentations \\(K\\), sharpening temperature \\(T\\), labeled data \\(\\mathcal{X}\\), unlabeled data \\(\\mathcal{U}\\) \n1: while \\(b\\) in \\(B\\) do \n2: \\(\\hat{x}_b \\gets \\text{augment}(x_b)\\) \n3: for \\(k\\) in \\(K\\) do \n4: \\(\\hat{u}_{b,k} \\gets \\text{augment}(u_b)\\) \n5: end for \n6: \\(\\hat{q}_b \\gets \\frac{1}{K} \\sum_k p_{model}(\\hat{u}_{b,k}; \\omega)\\) \n7: \\(q_b \\gets \\text{sharpen}(\\hat{q}_b, T)\\) \n8: end while \n9: \\(\\hat{\\mathcal{X}} = \\{(\\hat{x}_b, p_b)\\}_{b=1}^B, \\hat{\\mathcal{U}} = \\{(\\hat{u}_{b,k}, q_b)\\}_{b,k=1}^{B,K}\\) \n10: \\(\\mathcal{W} = \\text{shuffle}(\\text{concat}(\\hat{\\mathcal{X}}, \\hat{\\mathcal{U}}))\\) \n11: \\(\\hat{\\mathcal{X}}' = \\text{Mixup}(\\hat{\\mathcal{X}}_i, \\mathcal{W}_i), i \\in [1, |\\hat{\\mathcal{X}}|]\\) \n12: \\(\\hat{\\mathcal{U}}' = \\text{Mixup}(\\hat{\\mathcal{U}}_i, \\mathcal{W}_{i+ \\hat{\\mathcal{U}}}), i \\in [1, |\\hat{\\mathcal{U}}|]\\) \n13: \\(\\mathcal{L}_{\\mathcal{X}} \\gets \\text{CE}(\\mathcal{X}')\\), \\(\\mathcal{L}_{\\mathcal{U}} \\gets \\text{MAE}(\\mathcal{U}')\\) \n14: \\(\\mathcal{L} \\gets \\mathcal{L}_{\\mathcal{X}} + \\lambda_{\\mathcal{U}} \\mathcal{L}_{\\mathcal{U}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.422, + 0.469, + 0.557 + ], + "angle": 0, + "content": "to deliberately evaluate learning algorithms with noisy labels. Thus, following previous works [27, 38], we consider two types of label noise, i.e., symmetric and asymmetric label noise. The symmetric label noise is generated by randomly flipping labels of a portion of samples from one class to all other possible classes. The asymmetric label noise is designed to follow the structure of real-world label noise, where labels are flipped to similar classes within the superclasses." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.559, + 0.469, + 0.679 + ], + "angle": 0, + "content": "Clothing1M. Clothing1M is a large-scale dataset with noisy labels [60], containing over 1M images obtained from online shopping websites. The labels are from 14 classes generated based on surrounding texts provided by the sellers, and the noise rate is estimated around \\(38.5\\%\\) [39]. Also, the dataset provides 3 clean datasets for training, validation, and testing, containing 50K, 14K, and 10K images, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.681, + 0.469, + 0.771 + ], + "angle": 0, + "content": "ANIMAL-10N. ANIMAL-10N is a real-world noisy dataset released by [51], which is crawled from several online search engines using predefined labels as searching keywords. There are in total 55K images, of which 50K images are for training, and 5K images are for testing. The noise rate was estimated to be around \\(8\\%\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.786, + 0.294, + 0.801 + ], + "angle": 0, + "content": "5.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.469, + 0.901 + ], + "angle": 0, + "content": "For CIFAR-10 and CIFAR-100, we use the PreAct ResNet18 [29] as the backbone and train it using the SGD optimizer with the following settings: a momentum of 0.9, a weight decay of 5e-4, and a batch size of 128. The learning rate was initialized as 0.02 and reduced by a factor of 10, after 150 epochs. The network was trained for 300" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.891, + 0.121 + ], + "angle": 0, + "content": "epochs. The warmup period is 10 epochs for CIFAR-10, and 30 epochs for CIFAR-100." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.123, + 0.892, + 0.242 + ], + "angle": 0, + "content": "For Clothing1M, we use the ResNet50 [28] pre-trained on ImageNet [21] as the backbone and train it using the SGD optimizer with the following settings: a momentum of 0.9, a weight decay of 1e-3, and a batch size of 32. The learning rate was initialized as 0.002 and reduced by a factor of 10 after 50 epochs. The network was trained for 120 epochs. For each epoch, we sample 1000 mini-batches from the training data." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.244, + 0.892, + 0.319 + ], + "angle": 0, + "content": "For ANIMAL-10N, we use VGG19 [49] with batch normalization [30] as the backbone and train it using the SGD optimizer. We train the network for 100 epochs. The initial learning rate is set as 0.01 and reduced by a factor of 5 after 50 and 75 epochs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.32, + 0.892, + 0.379 + ], + "angle": 0, + "content": "In addition, we set the entropic regularization coefficient \\(\\epsilon\\) and sparsity regularization coefficient \\(\\gamma\\) as 10 and 1, respectively, which are consistent for all training implementations." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.392, + 0.702, + 0.408 + ], + "angle": 0, + "content": "5.3. Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.416, + 0.892, + 0.521 + ], + "angle": 0, + "content": "In this section, we present experimental results of the OT-Filter on benchmark datasets with both synthetic and real label noises. For CIFAR-10 and CIFAR-100, we consider the noise rate of \\(20\\%\\), \\(50\\%\\), \\(80\\%\\), and \\(90\\%\\) for symmetric noise, and noise rate of \\(40\\%\\) for asymmetric noise, respectively. The results regarding precision and recall were drawn from the 100-th epoch of robust training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.522, + 0.892, + 0.657 + ], + "angle": 0, + "content": "Sample Selection Performance. We first present the empirical study on the quality of sample selection. The dataset CIFAR-10 under noise rate of \\(90\\%\\) (symmetric) and noise rate (asymmetric) of \\(40\\%\\) was employed to compare our method with state-of-the-art sample selection method DivMix [38]. Here, we consider two metrics, precision and recall. As shown in Figure 1, with the increase of training epochs, our method outperforms the DivMix [38] in terms of both precision and recall." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.892, + 0.794 + ], + "angle": 0, + "content": "For example, on \\(90\\%\\) symmetric noise, when the training epoch is above 100, the precision of the OT-Filter is steadily above \\(90\\%\\), while the precision of DivMix is less than \\(73\\%\\). In terms of the recall, the OT-Filter is above \\(90\\%\\), while the DivMix is below \\(82\\%\\), when the training epochs are higher than 100. It also shows that our method converges steadily to the optimal solution, which DivMix yields fluctuations in the convergence process. Similar trends can be observed on \\(40\\%\\) asymmetric noise." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "CIFAR10 and CIFAR100 Datasets. Table 1 shows the test accuracy on CIFAR-10 for symmetric and asymmetric noise, respectively. In all testings, the OT-Filter demonstrates good performance in test accuracy. In particular, in the presence of high noise, the OT-Filter significantly outperforms other competitors. For \\(90\\%\\) noise rate, the test accuracy of our OT-Filter is \\(90.5\\%\\), which is about \\(14\\%\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16169" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.101, + 0.278, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.28, + 0.101, + 0.475, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.48, + 0.101, + 0.674, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.677, + 0.101, + 0.871, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.11, + 0.232, + 0.856, + 0.248 + ], + "angle": 0, + "content": "Figure 1. Sample selection performance (Precision & Recall) w.r.t. Training Epochs on CIFAR-10 under synthetic label noise." + }, + { + "type": "table", + "bbox": [ + 0.102, + 0.272, + 0.447, + 0.459 + ], + "angle": 0, + "content": "
MethodCIFAR-10Asym
20%50%80%90%40%
CE86.879.462.942.777.3
Co-teaching+ [62]89.585.767.447.971.3
Co-learning [54]92.284.561.2-81.4
TopoFilter [58]90.2-45.7-87.9
CRUST [40]91.186.358.3-88.8
Fine [34]91.087.369.4-89.5
M-correction [3]93.691.875.874.793.3
CTRR [61]93.1-83.781.789.0
DivMix [38]96.194.693.276.093.4
Fine+DivMix [34]96.194.993.590.593.8
OT-Filter96.095.394.090.595.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.469, + 0.471, + 0.525 + ], + "angle": 0, + "content": "Table 1. Test accuracies(%) obtained from state-of-the-art sample selection methods. The best results are in bold. The data are copied from respective papers. The - denotes the lack of respective data." + }, + { + "type": "table", + "bbox": [ + 0.102, + 0.542, + 0.447, + 0.73 + ], + "angle": 0, + "content": "
MethodCIFAR-100Asym
20%50%80%90%40%
CE62.046.719.910.144.5
Co-teaching+ [62]65.651.827.913.7-
Co-learning [54]66.654.535.5-47.6
CRUST [40]65.256.4--53.0
M-correction [3]73.966.141.624.347.4
CTRR [61]70.1-43.7-54.5
TopoFilter [58]65.6-20.7--
Fine [34]70.364.225.6-61.7
DivMix [38]77.374.660.231.555.1
Fine+DivMix [34]79.174.661.034.3-
OT-Filter76.773.861.842.876.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.74, + 0.47, + 0.796 + ], + "angle": 0, + "content": "Table 2. Test accuracies(%) obtained from state-of-the-art sample selection methods. The best results are in bold. The data are copied from respective papers. The - denotes the lack of respective data." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.471, + 0.903 + ], + "angle": 0, + "content": "better than DivMix. Although [38] performs slightly better than the OT-Filter at the \\(20\\%\\) noise rate, our method is also competitive. Similar trends can be observed on CIFAR100 (Table 2), where our method mostly follows the SOTA method and significantly outperforms its competitors when" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.275, + 0.644, + 0.29 + ], + "angle": 0, + "content": "the noise rate is high." + }, + { + "type": "table", + "bbox": [ + 0.543, + 0.302, + 0.851, + 0.469 + ], + "angle": 0, + "content": "
MethodBackboneTest Accuracy
CEResNet-5069.2
M-correction [1]ResNet-5071.0
Co-teaching [27]ResNet-5071.7
CTRR [61]ResNet-5072.7
Fine [34]ResNet-5072.9
CRUST [40]ResNet-5073.5
TopoFilter [58]ResNet-5074.1
DivMix* [38]ResNet-5074.3
Fine+DivMix [34]ResNet-5074.4
OT-FilterResNet-5074.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.479, + 0.893, + 0.523 + ], + "angle": 0, + "content": "Table 3. Test accuracy(%) on Clothing1M. The * denotes we have run the algorithm based on the official implementation. The data are copied from respective papers. The best result is in bold." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.539, + 0.894, + 0.631 + ], + "angle": 0, + "content": "Clothing1M Dataset. Table 3 shows the result of test accuracy on Clothing1M by comparing different competitors. The result validates the effectiveness of the OT-Filter on the large-scale real-world dataset. Moreover, It can be observed that the OT-Filter follows the state-of-the-art performance regarding test accuracy." + }, + { + "type": "table", + "bbox": [ + 0.563, + 0.642, + 0.832, + 0.771 + ], + "angle": 0, + "content": "
MethodTest Accuracy
CE79.4
Nested [47]81.3
SELFIE [51]81.8
PLC [63]83.4
Co-teaching+Nested [12]84.1
GJS [22]84.2
OT-Filter85.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.78, + 0.892, + 0.809 + ], + "angle": 0, + "content": "Table 4. Test accuracy(%) on ANIMAL-10N. The data are copied from respective papers. The best result is in bold." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.825, + 0.893, + 0.903 + ], + "angle": 0, + "content": "ANIMAL-10N Dataset. Table 4 shows the experimental results on ANIMAL-10N, a real-world dataset with moderate label noise. We compared our method with multiple state-of-the-art baseline methods that are not limited to sample selection. For fair comparison, we use the same VGG-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16170" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.09, + 0.467, + 0.155 + ], + "angle": 0, + "content": "
DatasetCIFAR-100
Noise Rate80%90%80%90%
MethodBestLastBestLastPrecisionRecallPrecisionRecall
OT-Filter w/o Spar-Reg51.050.723.823.790.447.162.119.7
OT-Filter61.861.542.842.193.761.872.340.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.166, + 0.47, + 0.207 + ], + "angle": 0, + "content": "Table 5. Ablation study for sparsity regularization on CIFAR-100 dataset. The left part is the performance comparison and the right part is the quality of sample selection." + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.223, + 0.468, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.384, + 0.468, + 0.44 + ], + "angle": 0, + "content": "Figure 2. The ablation study for the quality of sample selection w/o sparsity regularization on CIFAR-100 under noise rate of \\(90\\%\\). The result shows that sparsity regularization brings a relatively uniform sample selection." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.469, + 0.468, + 0.53 + ], + "angle": 0, + "content": "19 network architecture. The results show we outperform all competitors, and have \\(2.1\\%\\) performance improvement over PLC [63], and \\(3.7\\%\\) performance improvement over SELFIE [51]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.543, + 0.18, + 0.559 + ], + "angle": 0, + "content": "5.4. Analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.568, + 0.468, + 0.642 + ], + "angle": 0, + "content": "We conduct a series of ablation studies to understand the effectiveness of the key components of the OT-Filter, including the sparsity regularization of the optimal transport, the EM algorithm, and the entropic regularization coefficient." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Effect of Sparsity Regularization. The sparsity regularization discussed in Section 4.3 is one of the key components in our OT-Filter, which is designed to improve the accuracy of the optimal transport by controlling the sparsity of transport. We therefore study the effect of the sparsity regularization over sample selection and performance under high noise on CIFAR-100 in Table 5. It can be observed, with the sparsity regularization, both the quality of sample selection and test accuracy are much better than the counterpart without it. In particular, when the noise rate is \\(90\\%\\), it shows that the technique of sparsity regularization brings in over \\(10\\%\\) improvement in precision, about \\(21\\%\\) improvement in recall, and \\(19\\%\\) improvement in performance. Moreover, as shown in Figure 2, with the sparsity regularization, the number of samples selected for each class is more uniform. It is mainly attributed to discrete optimal transport allowing mass split." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.09, + 0.888, + 0.17 + ], + "angle": 0, + "content": "
DatasetCIFAR-10
Noise Rate20%50%80%90%
MethodBestLastBestLastBestLastBestLast
OT-Filter w/o EM95.995.795.094.893.693.490.089.8
OT-Filter96.095.895.395.193.793.590.590.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.505, + 0.181, + 0.885, + 0.194 + ], + "angle": 0, + "content": "Table 6. Ablation study for EM algorithm on CIFAR-10 dataset." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.209, + 0.889, + 0.28 + ], + "angle": 0, + "content": "
DatasetCIFAR-10
Noise Rate20%50%80%90%
MethodPrecisionRecallPrecisionRecallPrecisionRecallPrecisionRecall
ε = 199.793.999.093.994.690.688.684.2
ε = 599.793.899.093.894.690.688.684.2
ε = 1099.793.899.093.894.590.688.684.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.29, + 0.891, + 0.318 + ], + "angle": 0, + "content": "Table 7. Ablation study for entropic regularization coefficient on CIFAR-10 dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.345, + 0.892, + 0.466 + ], + "angle": 0, + "content": "Effect of Expectation Maximization. To filter noisy labels, we align all samples to clean barycenters via optimal transport. Therefore, the quality of clean barycenters affects the quality of sample selection and therefore the performance of the robust training. Table 6 indicates the impact of EM on CIFAR-10 dataset. It shows that the EM helps in improving the performance under different noise rates, and the significance is higher for high noise rates." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.466, + 0.892, + 0.631 + ], + "angle": 0, + "content": "Effect of Entropic Regularization Coefficient \\(\\epsilon\\). The motivation for equipping the optimal transport with entropic regularization is to speed up its computation. However, the entropic regularization could potentially degrade the transport sparsity and therefore the quality of sample selection. To alleviate that, we propose sparsity regularization, whose effectiveness is examined in Table 5. Then, a by-product of sparsity regularization is the parameter robustness of the regularization coefficient \\(\\epsilon\\). As shown in Table 7, the performance of sample selection stays quite stable when varying \\(\\epsilon\\) from 1 to 10." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.646, + 0.618, + 0.661 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.671, + 0.892, + 0.852 + ], + "angle": 0, + "content": "In this work, we study the problem of sample selection for learning with noisy labels. We propose the OT-Filter, a novel technique for retrieving clean samples, which can also be combined with existing semi-supervised learning techniques for handling noisy labels. Unlike previous works relying on a-priori knowledge or conforming to confirmation bias, the OT-Filter enhances the measurement of data discrepancy, by lifting data from Euclidean space to the probability space, and thus improves the quality of sample selection. Extensive experiments on synthetic and real-world noisy datasets are conducted to evaluate our proposals. The results show its effectiveness and superiority." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.852, + 0.892, + 0.898 + ], + "angle": 0, + "content": "Acknowledgement. This work is supported by NSFC (No.61772492, 62072428) and the CAS Pioneer Hundred Talents Program. Xike Xie is the corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16171" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.469, + 0.156 + ], + "angle": 0, + "content": "[1] Martial Agueh and Guillaume Carlier. Barycenters in the wasserstein space. SIAM Journal on Mathematical Analysis, 43(2):904-924, 2011. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.157, + 0.469, + 0.212 + ], + "angle": 0, + "content": "[2] Ethan Anderes, Steffen Borgwardt, and Jacob Miller. Discrete Wasserstein barycenters: Optimal transport for discrete data. Mathematical Methods of Operations Research, 84(2):389-409, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.213, + 0.469, + 0.269 + ], + "angle": 0, + "content": "[3] Eric Arazo, Diego Ortego, Paul Albert, Noel O'Connor, and Kevin McGuinness. Unsupervised label noise modeling and loss correction. In International conference on machine learning, pages 312-321. PMLR, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.27, + 0.468, + 0.324 + ], + "angle": 0, + "content": "[4] Martin Arjovsky, Soumith Chintala, and Léon Bottou. Wasserstein generative adversarial networks. In International conference on machine learning, pages 214-223. PMLR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.326, + 0.469, + 0.394 + ], + "angle": 0, + "content": "[5] Yingbin Bai, Erkun Yang, Bo Han, Yanhua Yang, Jiatong Li, Yinian Mao, Gang Niu, and Tongliang Liu. Understanding and improving early stopping for learning with noisy labels. Advances in Neural Information Processing Systems, 34:24392-24403, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.395, + 0.469, + 0.437 + ], + "angle": 0, + "content": "[6] Mokhtar S Bazaraa, Hanif D Sherali, and Chitharanjan M Shetty. Nonlinear programming: theory and algorithms. John Wiley & Sons, 2013. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.438, + 0.469, + 0.493 + ], + "angle": 0, + "content": "[7] Jean-David Benamou, Guillaume Carlier, Marco Cuturi, Luca Nenna, and Gabriel Peyre. Iterative bregman projections for regularized transportation problems. SIAM Journal on Scientific Computing, 37(2):A1111-A1138, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.494, + 0.469, + 0.549 + ], + "angle": 0, + "content": "[8] David Berthelot, Nicholas Carlini, Ian Goodfellow, Nicolas Papernot, Avital Oliver, and Colin A Raffel. Mixmatch: A holistic approach to semi-supervised learning. Advances in neural information processing systems, 32, 2019. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.55, + 0.469, + 0.578 + ], + "angle": 0, + "content": "[9] Dimitris Bertsimas and John N Tsitsiklis. Introduction to linear optimization, volume 6. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.579, + 0.468, + 0.606 + ], + "angle": 0, + "content": "[10] Stephen Boyd, Stephen P Boyd, and Lieven Vandenberghe. Convex optimization. Cambridge university press, 2004. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.607, + 0.469, + 0.662 + ], + "angle": 0, + "content": "[11] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.663, + 0.469, + 0.732 + ], + "angle": 0, + "content": "[12] Yingyi Chen, Xi Shen, Shell Xu Hu, and Johan AK Suykens. Boosting co-teaching with compression regularization for label noise. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2688-2692, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.733, + 0.469, + 0.816 + ], + "angle": 0, + "content": "[13] De Cheng, Tongliang Liu, Yixiong Ning, Nannan Wang, Bo Han, Gang Niu, Xinbo Gao, and Masashi Sugiyama. Instance-dependent label-noise learning with manifold-regularized transition matrix estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16630–16639, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.817, + 0.469, + 0.872 + ], + "angle": 0, + "content": "[14] Hao Cheng, Zhaowei Zhu, Xingyu Li, Yifei Gong, Xing Sun, and Yang Liu. Learning with instance-dependent label noise: A sample sieve approach. In International Conference on Learning Representations, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[15] Nicolas Courty, Rémi Flamary, Amaury Habrard, and Alain Rakotomamonjy. Joint distribution optimal transportation" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "for domain adaptation. Advances in Neural Information Processing Systems, 30, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.121, + 0.892, + 0.177 + ], + "angle": 0, + "content": "[16] Nicolas Courty, Rémi Flamary, and Devis Tuia. Domain adaptation with regularized optimal transport. In Joint European Conference on Machine Learning and Knowledge Discovery in Databases, pages 274-289. Springer, 2014. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.178, + 0.892, + 0.218 + ], + "angle": 0, + "content": "[17] Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. Advances in neural information processing systems, 26, 2013. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.219, + 0.892, + 0.26 + ], + "angle": 0, + "content": "[18] Marco Cuturi and Arnaud Doucet. Fast computation of wasserstein barycenters. In International conference on machine learning, pages 685-693. PMLR, 2014. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.26, + 0.892, + 0.327 + ], + "angle": 0, + "content": "[19] Bharath Bhushan Damodaran, Rémi Flamary, Vivien Seguy, and Nicolas Courty. An entropic optimal transport loss for learning deep neural networks under label noise in remote sensing images. Comput. Vis. Image Underst., 191:102863, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.329, + 0.892, + 0.385 + ], + "angle": 0, + "content": "[20] Arthur P Dempster, Nan M Laird, and Donald B Rubin. Maximum likelihood from incomplete data via the em algorithm. Journal of the Royal Statistical Society: Series B (Methodological), 39(1):1-22, 1977. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.386, + 0.892, + 0.441 + ], + "angle": 0, + "content": "[21] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.442, + 0.892, + 0.496 + ], + "angle": 0, + "content": "[22] Erik Englesson and Hossein Azizpour. Generalized jensen-shannon divergence loss for learning with noisy labels. Advances in Neural Information Processing Systems, 34:30284-30297, 2021. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.497, + 0.892, + 0.565 + ], + "angle": 0, + "content": "[23] Kilian Fatras, Bharath Bhushan Damodaran, Sylvain Lobry, Remi Flamary, Devis Tuia, and Nicolas Courty. Wasserstein adversarial regularization for learning with label noise. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.567, + 0.892, + 0.607 + ], + "angle": 0, + "content": "[24] Alessio Figalli and Federico Glaudo. An Invitation to Optimal Transport, Wasserstein Distances, and Gradient Flows. 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.609, + 0.892, + 0.663 + ], + "angle": 0, + "content": "[25] Aude Geneva, Gabriel Peyré, and Marco Cuturi. Learning generative models with sinkhorn divergences. In International Conference on Artificial Intelligence and Statistics, pages 1608-1617. PMLR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.664, + 0.892, + 0.72 + ], + "angle": 0, + "content": "[26] Wenshuo Guo, Nhat Ho, and Michael Jordan. Fast algorithms for computational optimal transport and wasserstein barycenter. In International Conference on Artificial Intelligence and Statistics, pages 2088-2097. PMLR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.72, + 0.892, + 0.789 + ], + "angle": 0, + "content": "[27] Bo Han, Quanming Yao, Xingrui Yu, Gang Niu, Miao Xu, Weihua Hu, Ivor Tsang, and Masashi Sugiyama. Coteaching: Robust training of deep neural networks with extremely noisy labels. Advances in neural information processing systems, 31, 2018. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.79, + 0.892, + 0.845 + ], + "angle": 0, + "content": "[28] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[29] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Identity mappings in deep residual networks. In European conference on computer vision, pages 630-645. Springer, 2016. 1, 3, 6" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16172" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[30] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In International conference on machine learning, pages 448-456. pmlr, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.47, + 0.218 + ], + "angle": 0, + "content": "[31] Ahmet Iscen, Jack Valmadre, Anurag Arnab, and Cordelia Schmid. Learning with neighbor consistency for noisy labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4672-4681, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.22, + 0.47, + 0.288 + ], + "angle": 0, + "content": "[32] Lu Jiang, Zhengyuan Zhou, Thomas Leung, Li-Jia Li, and Li Fei-Fei. Mentornet: Learning data-driven curriculum for very deep neural networks on corrupted labels. In International conference on machine learning, pages 2304–2313. PMLR, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.291, + 0.47, + 0.358 + ], + "angle": 0, + "content": "[33] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. Advances in neural information processing systems, 33:18661-18673, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.361, + 0.469, + 0.416 + ], + "angle": 0, + "content": "[34] Taehyeon Kim, Jongwoo Ko, JinHwan Choi, Se-Young Yun, et al. Fine samples for learning with noisy labels. Advances in Neural Information Processing Systems, 34:24137–24149, 2021. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.418, + 0.469, + 0.459 + ], + "angle": 0, + "content": "[35] Philip A Knight. The sinkhorn-knopp algorithm: convergence and applications. SIAM Journal on Matrix Analysis and Applications, 30(1):261-275, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.461, + 0.469, + 0.489 + ], + "angle": 0, + "content": "[36] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.49, + 0.469, + 0.517 + ], + "angle": 0, + "content": "[37] Kenneth Lange. MM optimization algorithms. SIAM, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.519, + 0.469, + 0.56 + ], + "angle": 0, + "content": "[38] Junnan Li, Richard Socher, and Steven CH Hoi. Dividemix: Learning with noisy labels as semi-supervised learning. arXiv preprint arXiv:2002.07394, 2020. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.562, + 0.469, + 0.617 + ], + "angle": 0, + "content": "[39] Sheng Liu, Jonathan Niles-Weed, Narges Razavian, and Carlos Fernandez-Granda. Early-learning regularization prevents memorization of noisy labels. Advances in Neural Information Processing Systems, 33, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.619, + 0.469, + 0.673 + ], + "angle": 0, + "content": "[40] Baharan Mirzasoleiman, Kaidi Cao, and Jure Leskovec. Coresets for robust training of deep neural networks against noisy labels. Advances in Neural Information Processing Systems, 33:11465-11477, 2020. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.676, + 0.469, + 0.731 + ], + "angle": 0, + "content": "[41] Kento Nishi, Yi Ding, Alex Rich, and Tobias Hollerer. Augmentation strategies for learning with noisy labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8022-8031, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.733, + 0.469, + 0.801 + ], + "angle": 0, + "content": "[42] Diego Ortego, Eric Arazo, Paul Albert, Noel E O'Connor, and Kevin McGuinness. Multi-objective interpolation training for robustness to label noise. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6606-6615, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[43] Giorgio Patrini, Alessandro Rozza, Aditya Krishna Menon, Richard Nock, and Lizhen Qu. Making deep neural networks robust to label noise: A loss correction approach. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1944-1952, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[44] Gabriel Peyré, Marco Cuturi, et al. Computational optimal transport: With applications to data science. Foundations" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "and Trends® in Machine Learning, 11(5-6):355-607, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.177 + ], + "angle": 0, + "content": "[45] Vikas C Raykar, Shipeng Yu, Linda H Zhao, Gerardo Hermosillo Valadez, Charles Florin, Luca Bogoni, and Linda Moy. Learning from crowds. Journal of machine learning research, 11(4), 2010. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.179, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[46] Ievgen Redko, Nicolas Courty, Rémi Flamary, and Devis Tuia. Optimal transport for multi-source domain adaptation under target shift. In The 22nd International Conference on Artificial Intelligence and Statistics, pages 849-858. PMLR, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.249, + 0.892, + 0.303 + ], + "angle": 0, + "content": "[47] Oren Rippel, Michael Gelbart, and Ryan Adams. Learning ordered representations with nested dropout. In International Conference on Machine Learning, pages 1746-1754. PMLR, 2014. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.305, + 0.892, + 0.333 + ], + "angle": 0, + "content": "[48] Filippo Santambrogio. Optimal transport for applied mathematicians. Birkhäuser, NY, 55(58-63):94, 2015. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.335, + 0.892, + 0.376 + ], + "angle": 0, + "content": "[49] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.378, + 0.892, + 0.459 + ], + "angle": 0, + "content": "[50] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. Advances in neural information processing systems, 33:596-608, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.462, + 0.892, + 0.517 + ], + "angle": 0, + "content": "[51] Hwanjun Song, Minseok Kim, and Jae-Gil Lee. Selfie: Refurbishing unclean samples for robust deep learning. In International Conference on Machine Learning, pages 5907-5915. PMLR, 2019. 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.519, + 0.892, + 0.588 + ], + "angle": 0, + "content": "[52] Fariborz Taherkhani, Ali Dabouei, Sobhan Soleymani, Jeremy Dawson, and Nasser M Nasrabadi. Transporting labels via hierarchical optimal transport for semi-supervised learning. In European Conference on Computer Vision, pages 509-526. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.59, + 0.892, + 0.671 + ], + "angle": 0, + "content": "[53] Fariborz Taherkhani, Ali Dabouei, Sobhan Soleymani, Jeremy Dawson, and Nasser M Nasrabadi. Self-supervised Wasserstein pseudo-labeling for semi-supervised image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12267-12277, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.674, + 0.892, + 0.73 + ], + "angle": 0, + "content": "[54] Cheng Tan, Jun Xia, Lirong Wu, and Stan Z Li. Co-learning: Learning from noisy labels with self-supervision. In Proceedings of the 29th ACM International Conference on Multimedia, pages 1405-1413, 2021. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.731, + 0.892, + 0.758 + ], + "angle": 0, + "content": "[55] Cédric Villani. Optimal transport: old and new, volume 338. Springer, 2009. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.76, + 0.892, + 0.787 + ], + "angle": 0, + "content": "[56] Cédric Villani. Topics in optimal transportation, volume 58. American Mathematical Soc., 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.789, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[57] Yikai Wang, Xinwei Sun, and Yanwei Fu. Scalable penalized regression for noise detection in learning with noisy labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 346-355, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[58] Pengxiang Wu, Songzhu Zheng, Mayank Goswami, Dimitris Metaxas, and Chao Chen. A topological filter for learning with label noise. Advances in neural information processing systems, 33:21382-21393, 2020. 1, 2, 7" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16173" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.16 + ], + "angle": 0, + "content": "[59] Xiaobo Xia, Tongliang Liu, Bo Han, Mingming Gong, Jun Yu, Gang Niu, and Masashi Sugiyama. Sample selection with uncertainty of losses for learning with noisy labels. In International Conference on Learning Representations, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.47, + 0.232 + ], + "angle": 0, + "content": "[60] Tong Xiao, Tian Xia, Yi Yang, Chang Huang, and Xiaogang Wang. Learning from massive noisy labeled data for image classification. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2691-2699, 2015. 1, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.234, + 0.471, + 0.302 + ], + "angle": 0, + "content": "[61] Li Yi, Sheng Liu, Qi She, A Ian McLeod, and Boyu Wang. On learning contrastive representations for learning with noisy labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16682-16691, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.304, + 0.47, + 0.372 + ], + "angle": 0, + "content": "[62] Xingrui Yu, Bo Han, Jiangchao Yao, Gang Niu, Ivor Tsang, and Masashi Sugiyama. How does disagreement help generalization against label corruption? In International Conference on Machine Learning, pages 7164-7173. PMLR, 2019. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.374, + 0.47, + 0.442 + ], + "angle": 0, + "content": "[63] Yikai Zhang, Songzhu Zheng, Pengxiang Wu, Mayank Goswami, and Chao Chen. Learning with feature-dependent label noise: A progressive approach. In Ninth International Conference on Learning Representations, volume 9, 2021. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.445, + 0.469, + 0.499 + ], + "angle": 0, + "content": "[64] Zhilu Zhang and Mert Sabuncu. Generalized cross entropy loss for training deep neural networks with noisy labels. Advances in neural information processing systems, 31, 2018. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.501, + 0.469, + 0.571 + ], + "angle": 0, + "content": "[65] Evgenii Zheltonozhskii, Chaim Baskin, Avi Mendelson, Alex M Bronstein, and Or Litany. Contrast to divide: Self-supervised pre-training for learning with noisy labels. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1657-1667, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.572, + 0.469, + 0.627 + ], + "angle": 0, + "content": "[66] Yudian Zheng, Guoliang Li, Yuanbing Li, Caihua Shan, and Reynold Cheng. Truth inference in crowdsourcing: Is the problem solved? Proceedings of the VLDB Endowment, 10(5):541-552, 2017. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.627 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16174" + } + ] +] \ No newline at end of file diff --git a/2023/OT-Filter_ An Optimal Transport Filter for Learning With Noisy Labels/69b93138-511e-410f-9d28-563936aa9d2f_origin.pdf b/2023/OT-Filter_ An Optimal Transport Filter for Learning With Noisy Labels/69b93138-511e-410f-9d28-563936aa9d2f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..99f02104a85558b81a0fbf7b04f5816661d7c7ce --- /dev/null +++ b/2023/OT-Filter_ An Optimal Transport Filter for Learning With Noisy Labels/69b93138-511e-410f-9d28-563936aa9d2f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:42c42dc45fc1361d2bfe8b13aa6db367d54246747672a823df48f5b339490f68 +size 2400534 diff --git a/2023/OT-Filter_ An Optimal Transport Filter for Learning With Noisy Labels/full.md b/2023/OT-Filter_ An Optimal Transport Filter for Learning With Noisy Labels/full.md new file mode 100644 index 0000000000000000000000000000000000000000..dbca2327067795799a6867fca7700693d0414c77 --- /dev/null +++ b/2023/OT-Filter_ An Optimal Transport Filter for Learning With Noisy Labels/full.md @@ -0,0 +1,403 @@ +# OT-Filter: An Optimal Transport Filter for Learning with Noisy Labels + +Chuanwen Feng† Yilong Ren† Xike Xie +University of Science and Technology of China +Data Darkness Lab, MIRACLE Center, Suzhou Institute for Advanced Research, USTC +{chuanwen,ylren}@mail.ustc.edu.cn, {xkxie}@ustc.edu.cn + +# Abstract + +The success of deep learning is largely attributed to the training over clean data. However, data is often coupled with noisy labels in practice. Learning with noisy labels is challenging because the performance of the deep neural networks (DNN) drastically degenerates, due to confirmation bias caused by the network memorization over noisy labels. To alleviate that, a recent prominent direction is on sample selection, which retrieves clean data samples from noisy samples, so as to enhance the model's robustness and tolerance to noisy labels. In this paper, we revamp the sample selection from the perspective of optimal transport theory and propose a novel method, called the OT-Filter. The OT-Filter provides geometrically meaningful distances and preserves distribution patterns to measure the data discrepancy, thus alleviating the confirmation bias. Extensive experiments on benchmarks, such as Clothing1M and ANIMAL-10N, show that the performance of the OT-Filter outperforms its counterparts. Meanwhile, results on benchmarks with synthetic labels, such as CIFAR-10/100, show the superiority of the OT-Filter in handling data labels of high noise. + +# 1. Introduction + +Deep learning has achieved great success on a flurry of emerging applications, such as [28, 29, 32, 49]. It is believed that the phenomenal achievement of deep learning is largely attributed to accurate labels. However, the inaccuracy or imprecision of labels is inherent in real-world datasets, the so-called noisy label challenge. One way to alleviate that is to collect labels from internet queries over data-level tags, but the performance of deep neural networks (DNN) suffers drastically from the inaccuracy of such labels [27, 32]. A higher quality of data labels can be obtained by employing human workers, but the seemingly "ground truth annotations" inevitably involve human biases or mis + +takes [45,66]. More, the human annotation is expensive and time-consuming, especially for large-scale datasets. + +There have been many works on handling noisy labels, such as regularization [23] and transition matrix [13, 43]. The regularization approach leverages the regularization bias to overcome the label noise issue. But the regularization bias is permanent [62], thus overfitting models to noisy labels. The transition matrix approach assumes that the transition probabilities between clean and noisy labels are fixed and independent of data samples. However, a quality label transition matrix is hard to be estimated, especially when the number of classes is big, making it fall short in handling noisy real-world datasets, such as [60] and [36]. A recent prominent direction is on adopting sample selection [27, 38, 58] for enhancing the label quality, by selecting clean samples from the noisy training dataset. In general, existing literatures in sample selection can be grouped to two categories, co-training networking [27, 62] and criterion-based filtering [31, 34, 57, 58]. + +The former utilizes the memorization of DNNs and multiple networks (e.g., co-teaching [27] and its variants [38, 62]) to filter label noise with small loss trick, so that a small set of clean samples are used as training examples. Letting alone the high training overhead of multiple networks, the disadvantages are two-fold: 1) it may require the a-priori knowledge of noisy rates to select the specified proportion of small loss samples as clean samples; 2) the small-loss trick is not tolerant to the error accumulation of network training once a clean sample is falsely recognized, the so-called confirmation bias, especially for labels with high noise where clean and noisy samples largely overlap. + +The latter alleviates the problem by setting a specific criterion. Mostly, existing works [31] [58] adopt Euclidean distances for measuring the similarity between data samples. Distance-based filtering iteratively explores the neighborhood of the feature representation and infers/cleans sample labels by aggregating the information from their neighborhoods. Despite the simplicity, distance-based filtering is insufficient to address noisy labels, especially when the label noise is high, e.g., overlapped label classes. + +In this paper, we revamp the sample selection from the perspective of optimal transport [56] and propose a novel filtering method, called the OT-Filter. In light of the optimal transport, we construct discrete probability measures over the sample features, which lifts the Euclidean space of feature vectors to a probability space. It thus enables a geometric way of measuring the discrepancy between probability measures, representing the corresponding sample features. In addition to the distance-based metric in the Euclidean space [31, 58], the OT-Filter also captures the distribution information in the probability space so as to alleviate the confirmation bias. Accordingly, a clean representation can be obtained for each class in the probability space. By optimizing the transport plan from a sample to a clean representation, one can better determine if a sample is clean or noisy, thus improving the quality of sample selection. + +In general, the merits of the OT-Filter can be summarized as follows. First, it does not require any a-priori knowledge about the noisy rate of a dataset. Second, it utilizes the optimal transport which provides geometrically meaningful distances to exploit the sample discrepancy yet preserving the distribution patterns in corresponding probability space, making the sample selection of high quality and theoretical support. Third, it can be plugged to existing robust training paradigms, e.g., supervised and semi-supervised robust training. + +We conduct extensive experiments with a series of synthetic and real datasets to gain insights into our proposals. The result shows that the OT-Filter follows state-of-the-art (SOTA) [38] when the noise rate is low, and dominates SOTA when the noise rate is high. For instance, the OT-Filter achieves about $14\%$ and $12\%$ higher accuracy than SOTA, in the presence of $90\%$ noise rate, on synthetic datasets CIFAR-10 and CIFAR-100, respectively. Moreover, the OT-Filter outperforms the competitors on real datasets Clothing1M and ANIMAL-10N. + +The rest of the paper is organized as follows. We review the existing literature in Section 2. In Section 3, we present preliminaries of optimal transport. We investigate our proposed OT-Filter in Section 4. Furthermore, we conduct extensive empirical studies in Section 5 and conclude the paper in Section 6. + +# 2. Related Work + +# 2.1. Learning with Noisy Label + +A flurry of research methods [5, 13, 14, 27, 40, 54] was proposed for learning with noisy labels, which can be roughly divided into two categories, robust model [13, 22, 39, 43, 54, 61, 64] and clean sample selection [27, 32, 62]. The clean sample selection aims to select a clean subset from a noisy dataset [14, 27, 38]. One way is to train multiple networks to filter noisy samples. For instance, [32] + +proposed to pre-train an extra network to select clean samples for the main network. The inferiority of this method is on error accumulation. To relieve that, Han et al. [27] proposed Co-teaching that maintains two networks simultaneously, where one network selects clean samples for the other network by small loss trick. However, as the increase of training epochs, Co-teaching would converge to consensus gradually. Then, $\mathrm{Co - teaching}^{+}$ [62] was proposed to keep two networks diverged. To avoid confirmation bias, DivMix [38] proposed a hybrid framework with semi-supervised training, achieving the state-of-the-art performance. Based on DivMix, [41, 65] modified the training schema, such as self-supervised training [11, 33], to boost the performance. + +The other research line of sample selection is to filter noisy samples based on specific criteria [31, 34, 42, 59]. For example, Xia et al. [59] proposed to reduce the uncertainty of loss using specific strategies. Iscen et al. [31] proposed to leverage similarities between training examples in the feature space, encouraging the prediction of each example to be close to its nearest neighbors. Also, in Euclidean space, Wu et al. [58] proposed the TopoFilter that filters noisy samples using k-nearest neighborhood distance between prelogits. Fine [34] proposed to use the principal components of latent representations to select clean samples. + +# 2.2. Optimal Transport + +The optimal transport problem [55, 56] aims to move mass from a probability measure to another probability measure at a minimum cost. It defines the Wasserstein distance [24] and provides a geometric way to measure the discrepancy between probability measures. One factor that limits the wide application of optimal transport is the high computational cost. To improve its scalability, Cuturi [17] proposed an entropic regularization for the transport plan, which yields an efficient algorithm with the matrix scaling method of Sinkhorn-Knopp [35]. Recently, computational optimal transport [44, 48] has found many applications in various areas, e.g., generative models [4, 25], domain adaptation [15, 16], and semi-supervised learning [52, 53]. In particular, [1] studied the barycenter in Wasserstein space. [18] explored the efficient computing of Wasserstein barycenter. [19] proposed to use entropic optimal transport loss, based on joint distribution optimal transport [15], to build a robust training. + +In this paper, we study the problem of clean sample selection from the perspective of optimal transport to combat label noise. The proposed method can potentially capture geometric information from the probability space. Unlike the mechanism of multiple network training, it only trains a single network without a-priori information on noise rates. Moreover, it can be easily plugged to multiple robust training paradigms. + +# 3. Preliminary + +In this section, we present preliminaries on optimal transport, making a basis for subsequent proposed techniques. The optimal transport [56] is to seek an optimal transport plan between two measures at a minimal cost (e.g. the Wasserstein distance), which provides a geometric way of matching probability measures. Here we first introduce the general definition of Wasserstein distance induced by optimal transport problem. More details about optimal transport can be found in [44, 48, 55]. + +Wassertein Distance. Let $\mathcal{S}$ be a locally complete and separable metric space, the $\mathcal{P}(S)$ be the Borel probability measures set on $\mathcal{S}$ . For any $\mathcal{X}, \mathcal{Y} \subset \mathcal{S}$ , given $\mu \in \mathcal{P}(\mathcal{X})$ and $\nu \in \mathcal{P}(\mathcal{Y})$ , the optimal transport defines a Wasserstein distance between two probability measures, denoted as + +$$ +W _ {p} (\mu , \nu) := \left(\inf _ {\pi (\mu , \nu)} \int_ {\mathcal {X} \times \mathcal {Y}} | | x - y | | ^ {p} d \pi (\mu , \nu)\right) ^ {\frac {1}{p}} +$$ + +$p \geq 1$ , where the $\pi(\mu, \nu)$ is the set of joint probability measures with marginal $\mu$ and $\nu$ . + +# 3.1. Discrete Optimal Transport + +Let $\Delta_{n} = \{\mathbf{a}\in \mathbb{R}_{+}^{n}|\sum_{i = 1}^{n}a_{i} = 1,\forall a_{i}\geq 0\}$ be a probability simplex in dimension $n$ . Consider two empirical probability measures $\mu = \sum_{i = 1}^{n}a_{i}\delta_{x_{i}}$ and $\nu = \sum_{j = 1}^{m}b_{j}\delta_{y_{j}}$ , defined on metric space $\mathcal{X}$ with support $\{x_i\}_{i = 1}^n$ and $\mathcal{V}$ with support $\{y_j\}_{j = 1}^m$ respectively. Here the weight vector $\mathbf{a} = (a_1,a_2,\dots,a_n)$ and $\mathbf{b} = (b_{1},b_{2},\dots,b_{m})$ live in $\Delta_{n}$ and $\Delta_{m}$ , respectively. The $\delta$ stands for the Dirac unit mass function. Given a transport cost $c:\mathcal{X}\times \mathcal{Y}\to \mathbb{R}_+$ the discrete optimal transport between probability measures $\mu$ and $\nu$ can be formulated as + +$$ +\begin{array}{l} W _ {p} ^ {p} (\mu , \nu) := \min _ {M \in \Pi (\mu , \nu)} \langle C, M \rangle_ {F} \tag {1} \\ s. t. M \mathbf {1} _ {m} = \mu , M ^ {T} \mathbf {1} _ {n} = \nu \\ \end{array} +$$ + +where $C \in \mathbb{R}_+^{n \times m}$ is the transport cost matrix, and $c_{ij}$ represents a unit transport cost from $x_i$ to $y_j$ . The $M \in \mathbb{R}_+^{n \times m}$ is transport plan in which $m_{ij}$ denotes the amount of mass transported from $x_i$ to $y_j$ . All feasible transport plans constitute transport polytope $\Pi(\mu, \nu)$ . The $\langle C, M \rangle_F$ is the Frobenius inner product of matrices and equals to $tr(C^T M)$ . + +# 3.2. Regularized Optimal Transport + +The discrete optimal transport formulation, in essence, is a convex optimization problem. More precisely, it's a linear programming problem. Unfortunately, this linear programming problem has a cubic computing complexity. A way to relieve this is to leverage entropic regularization [17] formulated as + +$$ +\min _ {M \in \Pi (\mu , \nu)} \langle C, M \rangle_ {F} - \epsilon H (M) \tag {2} +$$ + +where $\epsilon > 0$ is the regularization coefficient, and $H(M)$ is the entropic regularization term, which promotes an efficient computation for transportation via the matrix scaling algorithm [17], given by + +$$ +H (M) := - \sum_ {i j} M _ {i j} \left(\log \left(M _ {i j}\right) - 1\right) +$$ + +# 3.3. Wasserstein Barycenter + +Given $N$ probability measures $\{\nu_{1},\nu_{2},\dots,\nu_{N}\}$ , $\nu_{i} \in \mathcal{P}(S)$ , each of which has finite supports and second moments. A Wasserstein barycenter of these measures is a probability measure $\mu$ , satisfying: + +$$ +\mu := \inf \sum_ {i = 1} ^ {N} \lambda_ {i} W _ {2} ^ {2} (\nu_ {i}, \mu), \quad s. t. \sum_ {i = 1} ^ {N} \lambda_ {i} = 1, \forall \lambda_ {i} \geq 0 +$$ + +The notion was first proposed by [1], where some elegant properties of Wasserstein barycenter were presented. Then [2] discussed it on the discrete case and showed that the problem of finding Wasserstein barycenter over the space $\mathcal{P}(\mathcal{S})$ can be reduced to a simpler space $\mathcal{O}_r(S)$ , where $r = \sum_{i=1}^{N} e_i - N + 1$ , and $e_i$ is the number of components of $\nu_i, i \in [1, N]$ . To find Wasserstein barycenter over space $\mathcal{O}_r(S)$ , a set of efficient algorithms [7, 18, 26] were proposed. + +# 4. Methods + +# 4.1. Overview + +In this section, we study the OT-Filter, the proposed method based on optimal transport for learning with noisy labels in a semi-supervised paradigm. The key idea of our proposed method is to transport samples with noisy labels to samples with clean labels at a minimum cost, which makes sample selection of high quality and theoretical support. + +The mechanism of the OT-Filter, consists of two phases, representation phase (Section 4.2) and transportation phase (Section 4.3). Then, the OT-Filter can be plugged to off-the-shelf SSLs for robust training, e.g., MixMatch [8], by treating clean samples as labeled data and noisy samples as unlabeled samples (Section 4.4). + +# 4.2. Representation Phase: Finding Clean Representations via Wasserstein Barycenters + +In the representation phase, we first use a pre-trained network, e.g., ResNet [29], to extract a set of feature representations of the labeled samples. Then, we define a discrete probability measure over the feature representations of each class to find the clean representations (or prototypes) via Wasserstein barycenters. Since there are mistrabeled samples, the barycenters we obtained may be noisy. Therefore, we iteratively optimize the barycenters with the expectation + +maximization (EM) algorithm [20] to converge to clean representations. The details about the representation phase are covered in Algorithm 1. + +Feature Extraction. The purpose of the feature extraction is to transform the input noisy label dataset into a feature-label dataset via a network. A noisy label dataset of $N$ samples can be represented by $\hat{\mathcal{D}} = \{(\mathbf{x}_i,\hat{y}_i)\}_{i = 1}^N$ , where $\mathbf{x}_i\in \mathbb{R}^d$ denotes the $i$ -th sample and $\dot{y}_i\in \{0,1\} ^K$ denotes the corresponding noisy label over $K$ classes. Then, a feature extraction network $f(\omega ,\mathbf{x}_i):\mathbb{R}^d\to \mathbb{R}^m$ with parameters $\omega$ maps a sample $\mathbf{x}_i$ to a $m$ -dimensional feature representation, denoted as $\tilde{\mathbf{x}}_i\in \mathbb{R}^m$ . Therefore, all output feature representations and their corresponding labels constitute a feature-label set, denoted as $\tilde{D} = \{(\tilde{\mathbf{x}}_i,\hat{y}_i)\}_{i = 1}^N$ . + +Probability Measure Modeling. The purpose of probability measure modeling is to define a probability measure over the feature-label set. This definition will potentially lift a Euclidean space $\mathcal{D}$ , in which the feature representations originally live, to a probability space, abbreviated as $\mathcal{P}(\mathcal{D})$ . Assume that the labels are of $K$ classes. First, we split the feature-label set $\tilde{D}$ into $K$ subsets, i.e., $\{s_1,s_2,\dots,s_K\}$ according to the label class. For one feature representation $\tilde{\mathbf{x}}_i$ in class $k\in [1,K]$ , we represent it as $\tilde{\mathbf{x}}_i^k,i\in [1,|s_k|]$ . Then, all feature representations in subset $s_k$ construct a probability measure, denoted as + +$$ +\mathcal {Q} = \frac {1}{| s _ {k} |} \sum_ {i = 1} ^ {| s _ {k} |} \delta_ {\tilde {\mathbf {x}} _ {i} ^ {k}} +$$ + +where the $\delta_{\widetilde{\mathbf{x}}_i^k}$ is a Dirac unit mass on $\widetilde{\mathbf{x}}_i^k$ , and for simplicity, we use uniform weights. + +Clean Representation Retrieval. The problem of finding the representation of any subset $s_k$ is equivalent to finding the corresponding Wasserstein barycenter $\mathcal{B}_k$ with finite supports, formalized as follows: + +$$ +\mathcal {B} _ {k} := \inf _ {\mathcal {B} \in \mathcal {O} _ {r} (\tilde {\mathcal {D}})} W _ {2} ^ {2} (\mathcal {B}, \mathcal {Q}) \tag {3} +$$ + +where the $W$ is the Wasserstein distance, and $r = |s_k|$ . As mentioned in Section 3, from the perspective of linear programming, the above optimization admits the dual: + +$$ +D (\alpha , \beta) = \max _ {(\alpha , \beta) \in R (C)} \alpha^ {T} \mathcal {B} + \beta^ {T} \mathcal {Q} +$$ + +where the polyhedron of dual variables is: + +$$ +R (C) = \left\{\left(\alpha , \beta\right) \in \mathbb {R} ^ {r} \times \mathbb {R} ^ {r} \mid \alpha_ {i} + \beta_ {j} \leq C _ {i j} \wedge i, j \in [ r ] \right\} +$$ + +By means of duality theory [6,9,10], we can find the relation of solution between the primal and dual problem. Here, the optimal solution of $D(\alpha, \beta)$ is a subgradient of $W_2^2(\mathcal{Q}, \mathcal{B})$ with respect to $\mathcal{B}$ . Thus, we employ a simple projected subgradient [18] to optimize Equation 3. over $\mathcal{B}$ . + +Despite we have found a set of Wasserstein barycenters $\{\mathcal{B}_i\}_{i=1}^K$ , the barycenters may be noisy since there are mislabeled data. Moreover, the clean barycenters are unknown. Therefore, for each class, we consider the ideal and clean barycenter $\mathcal{B}_c$ as a hidden variable, the obtained noisy barycenter $\mathcal{B}$ as an observed variable. Then, the problem of finding clean barycenter $\mathcal{B}_c$ is to maximize the following log-likelihood: + +$$ +\hat {\theta} \leftarrow \arg \max _ {\theta} \log \sum_ {\mathcal {B} _ {c}} p (\mathcal {B} _ {c}, \mathcal {B} | \theta) +$$ + +where the $\theta$ is the parameters of probabilistic model. By this way, we can optimize the noisy barycenters iteratively with the EM algorithm to converge to clean barycenters. In E-step, we construct optimal transport between all samples and clean barycenters, thereby we can infer the labels of all samples. In M-step, we refine the barycenters with inferred label obtained from E-step. The overall computing details see Algorithm 1. + +Algorithm 1 Find Clean Representation via Wasserstein Barycenter +Require: $\mathcal{Q} = \{\tilde{\mathbf{x}}_i\}_{i=1}^{|s_k|} \in \mathbb{R}^{m \times r}$ , regularization coefficient $\epsilon$ , $\mathcal{B} \in \mathbb{R}^{m \times e}$ +1: Initialize: $\mathcal{B}$ , $\nu \gets \mathcal{B}$ , $\mu \gets \mathcal{Q}$ , $\eta = 1/2$ , $t = 2$ +2: while $\mathcal{B}$ not converged do +3: set $\nu = \mathbf{1}_e / e$ , $\mu = \mathbf{1}_r / r$ , $\hat{\nu} = \tilde{\nu} = \mathbf{1} / e$ +4: while $\nu$ not converged do +5: $\beta \gets (t + 1) / 2$ , $\nu \gets (1 - \beta^{-1})\hat{\nu} + \beta^{-1}\tilde{\nu}$ +6: $\alpha \gets$ optimizer of $\mathbf{d}(\nu, \mu, C)$ +7: $\tilde{\nu} \gets \tilde{\nu} \circ \exp(-\beta\alpha)$ , $\tilde{\nu} \gets \tilde{\nu} / \tilde{\nu}^T\mathbf{1}$ +8: $\hat{\nu} \gets (1 - \beta^{-1})\hat{\nu} + \beta^{-1}\tilde{\nu}$ +9: $t \gets t + 1$ +10: end while +11: $\nu \gets \hat{\nu}$ +12: $\mathcal{M} \gets$ optimizer of $\mathbf{p}(\nu, \mu, C)$ +13: $\mathcal{B} \gets (1 - \eta)\mathcal{B} + \eta (\mathcal{QM})diag(\nu^{-1})$ +14: end while +15: while $\mathcal{B}_c$ not converged do +16: E-Step: +17: $\mathbb{L}_{\hat{\mathcal{D}}} \gets \mathrm{OT}(\mu, \nu, C)$ +18: M-Step: +19: $\mathcal{B}_c \gets$ refine $\mathcal{B}$ with $\mathbb{L}_{\hat{\mathcal{D}}}$ +20: end while +21: Return $\mathcal{B}_c$ + +# 4.3. Transportation Phase: Transporting Noisy Labels via Regularized Optimal Transport + +In the transportation phase, noisy feature representations are transported to clean representations. In particular, we align all feature representations, including those of noisy + +samples, to the clean representations (obtained in the representation phase) via sparsity regularized optimal transport. This operation potentially detects the corrupted labels based on the result of the optimal transport. Here, the mechanism of our sample selection via optimal transport is described as follows. First, we align the clean barycenters to the all rest samples via optimal transport. Based on the result of optimal transport, we can infer the labels of the all rest samples. We consider the route with the max mass transport that forms a coupling and then assign the label of a clean barycenter to the samples on this route. Second, we select the sample whose inferred label is identical to the original label as a clean sample. + +Sparsity Regularization. After the representation phase, clean representations are obtained, which are essentially the barycenters of label classes. The clean representation of a class provides the guidance of identifying whether samples are clean or noisy w.r.t. the given class. + +We consider the barycenters of all the $K$ classes as a discrete probability measure, denoted as $\mu = \frac{1}{K}\sum_{k = 1}^{K}\delta_{u_k}$ where $\delta_{\mu_k}$ is a Dirac unit mass on $\mu_{k}$ . Accordingly, the feature representations of the entire dataset construct a discrete probability measure $\nu = \frac{1}{N}\sum_{i = 1}^{N}\delta_{\tilde{\mathbf{x}}_i}$ . Then, we transport feature representations $\nu$ of all samples to the clean barycenters $\mu$ , in order to detect if a sample is clean or mislabeled. Inspired by [16,46], we view the clean barycenters as the source domain and the feature representations of samples as the target domain. Then, we recast the alignments as an optimal transport problem with a sparsity regularization [16], written as + +$$ +\min _ {M \in \Pi (\mu , \nu)} \langle C, M \rangle_ {F} - \epsilon H (M) + \gamma \sum_ {j, c} | | M (i _ {c}, j) | | _ {1} ^ {\frac {1}{2}} \tag {4} +$$ + +The second term is an entropic regularization, as described in Equation 2. The third term is a sparsity regularization with coefficient $\gamma$ , where $i_c$ index the line if its element in class $c$ , and $M(i_c, j)$ is a vector consisting of the $j$ -th column of $M$ in class $c$ . The $\left\| \cdot \right\|_1^{1/2}$ denotes an $L_1$ norm with the power of $\frac{1}{2}$ . The entropic regularization enables discrete optimal transport to be efficient. However, it disperses the transport route, which negatively affects sample selection. Therefore, to improve the quality of sample selection, we employ a sparsity regularization. The sparsity regularization promotes a feature representation that would be aligned to one of the clean barycenters and penalizes transportation matrix $M$ that aligns together feature representations with different labels. + +Optimization. Despite the consistent sparsity achieved by the regularization, the objective function of Equation 4 is non-convex. Moreover, the $L_{1}$ norm with the power in label regularization term is a concave function. A common optimization strategy [37] is to construct a convex upper bound, represented by convex functions, over the non-convex prob + +lem. By doing so, we can use the optimal solution of convex upper bound to approximate the optimal solution of the original non-convex problem. If applying the simplest linear approximation, we have + +$$ +\sum_ {j, c} \left| \left| M \left(i _ {c}, j\right) \right| \right| _ {1} ^ {\frac {1}{2}} \leq \langle G, M \rangle_ {F} + \mathcal {L} +$$ + +where matrix $G$ is $\frac{1}{2} (||\hat{M}(i_c, j) + \delta||^{-\frac{1}{2}})$ , and $\hat{M}$ is a given start point. The steps of optimization process are depicted as Algorithm 2. + +Algorithm 2 Transportation: Transport Noisy Label via Regularized Optimal Transport + +Require: optimal cost $C_{\text{min}}$ from Equation 1, + +1: Initialize: $G = 0$ +2: while G not converged do +3: $C_{min} \gets C_{min} + G$ +4: $M\gets$ optimizer of Equation 2 with $C_{min}$ +5: $G\gets$ update $G$ using $M$ +6: end while + +# 4.4. Robust Training in SSL + +After all feature representations are transported to the clean barycenters. We filter noisy samples and obtain two data subsets, namely the clean sample set and noisy sample set. As aforementioned, our sample selection method is flexible enough for supporting various robust training paradigms, e.g., the robust supervised training [64] and the robust semi-supervised learning [50]. Following [38], we select MixMatch [8] with data augmentation strategy [41] as the practice of robust semi-supervised training but with fewer operations. We admit the labels of clean samples and neglect the labels of noisy samples, and then we use labeled data and unlabeled data for training in a semi-supervised learning fashion. The overall training process is described as Algorithm 3. + +# 5. Experiments + +# 5.1. Datasets + +We perform extensive experiments on four benchmark datasets: CIFAR-10 [36] and CIFAR-100 [36] with synthetic label noise, and Clothing1M [60] and ANIMAL-10N [51] with real label noise. + +CIFAR10/100. We evaluate the OT-Filter on CIFAR10 and CIFAR-100 [36] with synthetic noise. Both datasets contain 60,000 $32 \times 32$ color images, in 10 and 100 classes, respectively. From both datasets, 50K images are used for training, and 10K images are used for testing. Since noise characteristics can hardly be determined in advance, synthetic noise is commonly taken for controlling the noise rate + +Algorithm 3 Robust Training in SSL +Require: network parameter $\omega$ , Beta distribution parameter $\alpha$ , weight of unlabeled loss $\lambda_{\mathcal{U}}$ , batch size $B$ , number of augmentations $K$ , sharpening temperature $T$ , labeled data $\mathcal{X}$ , unlabeled data $\mathcal{U}$ +1: while $b$ in $B$ do +2: $\hat{x}_b \gets \text{augment}(x_b)$ +3: for $k$ in $K$ do +4: $\hat{u}_{b,k} \gets \text{augment}(u_b)$ +5: end for +6: $\hat{q}_b \gets \frac{1}{K} \sum_k p_{model}(\hat{u}_{b,k}; \omega)$ +7: $q_b \gets \text{sharpen}(\hat{q}_b, T)$ +8: end while +9: $\hat{\mathcal{X}} = \{(\hat{x}_b, p_b)\}_{b=1}^B, \hat{\mathcal{U}} = \{(\hat{u}_{b,k}, q_b)\}_{b,k=1}^{B,K}$ +10: $\mathcal{W} = \text{shuffle}(\text{concat}(\hat{\mathcal{X}}, \hat{\mathcal{U}}))$ +11: $\hat{\mathcal{X}}' = \text{Mixup}(\hat{\mathcal{X}}_i, \mathcal{W}_i), i \in [1, |\hat{\mathcal{X}}|]$ +12: $\hat{\mathcal{U}}' = \text{Mixup}(\hat{\mathcal{U}}_i, \mathcal{W}_{i+ \hat{\mathcal{U}}}), i \in [1, |\hat{\mathcal{U}}|]$ +13: $\mathcal{L}_{\mathcal{X}} \gets \text{CE}(\mathcal{X}')$ , $\mathcal{L}_{\mathcal{U}} \gets \text{MAE}(\mathcal{U}')$ +14: $\mathcal{L} \gets \mathcal{L}_{\mathcal{X}} + \lambda_{\mathcal{U}} \mathcal{L}_{\mathcal{U}}$ + +to deliberately evaluate learning algorithms with noisy labels. Thus, following previous works [27, 38], we consider two types of label noise, i.e., symmetric and asymmetric label noise. The symmetric label noise is generated by randomly flipping labels of a portion of samples from one class to all other possible classes. The asymmetric label noise is designed to follow the structure of real-world label noise, where labels are flipped to similar classes within the superclasses. + +Clothing1M. Clothing1M is a large-scale dataset with noisy labels [60], containing over 1M images obtained from online shopping websites. The labels are from 14 classes generated based on surrounding texts provided by the sellers, and the noise rate is estimated around $38.5\%$ [39]. Also, the dataset provides 3 clean datasets for training, validation, and testing, containing 50K, 14K, and 10K images, respectively. + +ANIMAL-10N. ANIMAL-10N is a real-world noisy dataset released by [51], which is crawled from several online search engines using predefined labels as searching keywords. There are in total 55K images, of which 50K images are for training, and 5K images are for testing. The noise rate was estimated to be around $8\%$ . + +# 5.2. Implementation Details + +For CIFAR-10 and CIFAR-100, we use the PreAct ResNet18 [29] as the backbone and train it using the SGD optimizer with the following settings: a momentum of 0.9, a weight decay of 5e-4, and a batch size of 128. The learning rate was initialized as 0.02 and reduced by a factor of 10, after 150 epochs. The network was trained for 300 + +epochs. The warmup period is 10 epochs for CIFAR-10, and 30 epochs for CIFAR-100. + +For Clothing1M, we use the ResNet50 [28] pre-trained on ImageNet [21] as the backbone and train it using the SGD optimizer with the following settings: a momentum of 0.9, a weight decay of 1e-3, and a batch size of 32. The learning rate was initialized as 0.002 and reduced by a factor of 10 after 50 epochs. The network was trained for 120 epochs. For each epoch, we sample 1000 mini-batches from the training data. + +For ANIMAL-10N, we use VGG19 [49] with batch normalization [30] as the backbone and train it using the SGD optimizer. We train the network for 100 epochs. The initial learning rate is set as 0.01 and reduced by a factor of 5 after 50 and 75 epochs. + +In addition, we set the entropic regularization coefficient $\epsilon$ and sparsity regularization coefficient $\gamma$ as 10 and 1, respectively, which are consistent for all training implementations. + +# 5.3. Experimental Results + +In this section, we present experimental results of the OT-Filter on benchmark datasets with both synthetic and real label noises. For CIFAR-10 and CIFAR-100, we consider the noise rate of $20\%$ , $50\%$ , $80\%$ , and $90\%$ for symmetric noise, and noise rate of $40\%$ for asymmetric noise, respectively. The results regarding precision and recall were drawn from the 100-th epoch of robust training. + +Sample Selection Performance. We first present the empirical study on the quality of sample selection. The dataset CIFAR-10 under noise rate of $90\%$ (symmetric) and noise rate (asymmetric) of $40\%$ was employed to compare our method with state-of-the-art sample selection method DivMix [38]. Here, we consider two metrics, precision and recall. As shown in Figure 1, with the increase of training epochs, our method outperforms the DivMix [38] in terms of both precision and recall. + +For example, on $90\%$ symmetric noise, when the training epoch is above 100, the precision of the OT-Filter is steadily above $90\%$ , while the precision of DivMix is less than $73\%$ . In terms of the recall, the OT-Filter is above $90\%$ , while the DivMix is below $82\%$ , when the training epochs are higher than 100. It also shows that our method converges steadily to the optimal solution, which DivMix yields fluctuations in the convergence process. Similar trends can be observed on $40\%$ asymmetric noise. + +CIFAR10 and CIFAR100 Datasets. Table 1 shows the test accuracy on CIFAR-10 for symmetric and asymmetric noise, respectively. In all testings, the OT-Filter demonstrates good performance in test accuracy. In particular, in the presence of high noise, the OT-Filter significantly outperforms other competitors. For $90\%$ noise rate, the test accuracy of our OT-Filter is $90.5\%$ , which is about $14\%$ + +![](images/5d9c3fad1e1a6576ba873a292b6cbb4de291f5ab4c0b3ad4590e1966a76763f0.jpg) +Figure 1. Sample selection performance (Precision & Recall) w.r.t. Training Epochs on CIFAR-10 under synthetic label noise. + +![](images/747b89a5ea3e898f42eab527e153797e3482b5a7014a3263d8bf02bfe7344e86.jpg) + +![](images/aff6f3d61ce1e94f8fe84f1f0c57d7c1ba55d4122aa154909e24d109f62034d3.jpg) + +![](images/7535333730364ceb8b133640270ce0e96f7200df21589f3c04ae1028fd48e47a.jpg) + +
MethodCIFAR-10Asym
20%50%80%90%40%
CE86.879.462.942.777.3
Co-teaching+ [62]89.585.767.447.971.3
Co-learning [54]92.284.561.2-81.4
TopoFilter [58]90.2-45.7-87.9
CRUST [40]91.186.358.3-88.8
Fine [34]91.087.369.4-89.5
M-correction [3]93.691.875.874.793.3
CTRR [61]93.1-83.781.789.0
DivMix [38]96.194.693.276.093.4
Fine+DivMix [34]96.194.993.590.593.8
OT-Filter96.095.394.090.595.1
+ +Table 1. Test accuracies(%) obtained from state-of-the-art sample selection methods. The best results are in bold. The data are copied from respective papers. The - denotes the lack of respective data. + +
MethodCIFAR-100Asym
20%50%80%90%40%
CE62.046.719.910.144.5
Co-teaching+ [62]65.651.827.913.7-
Co-learning [54]66.654.535.5-47.6
CRUST [40]65.256.4--53.0
M-correction [3]73.966.141.624.347.4
CTRR [61]70.1-43.7-54.5
TopoFilter [58]65.6-20.7--
Fine [34]70.364.225.6-61.7
DivMix [38]77.374.660.231.555.1
Fine+DivMix [34]79.174.661.034.3-
OT-Filter76.773.861.842.876.5
+ +Table 2. Test accuracies(%) obtained from state-of-the-art sample selection methods. The best results are in bold. The data are copied from respective papers. The - denotes the lack of respective data. + +better than DivMix. Although [38] performs slightly better than the OT-Filter at the $20\%$ noise rate, our method is also competitive. Similar trends can be observed on CIFAR100 (Table 2), where our method mostly follows the SOTA method and significantly outperforms its competitors when + +the noise rate is high. + +
MethodBackboneTest Accuracy
CEResNet-5069.2
M-correction [1]ResNet-5071.0
Co-teaching [27]ResNet-5071.7
CTRR [61]ResNet-5072.7
Fine [34]ResNet-5072.9
CRUST [40]ResNet-5073.5
TopoFilter [58]ResNet-5074.1
DivMix* [38]ResNet-5074.3
Fine+DivMix [34]ResNet-5074.4
OT-FilterResNet-5074.5
+ +Clothing1M Dataset. Table 3 shows the result of test accuracy on Clothing1M by comparing different competitors. The result validates the effectiveness of the OT-Filter on the large-scale real-world dataset. Moreover, It can be observed that the OT-Filter follows the state-of-the-art performance regarding test accuracy. + +Table 3. Test accuracy(%) on Clothing1M. The * denotes we have run the algorithm based on the official implementation. The data are copied from respective papers. The best result is in bold. + +
MethodTest Accuracy
CE79.4
Nested [47]81.3
SELFIE [51]81.8
PLC [63]83.4
Co-teaching+Nested [12]84.1
GJS [22]84.2
OT-Filter85.5
+ +Table 4. Test accuracy(%) on ANIMAL-10N. The data are copied from respective papers. The best result is in bold. + +ANIMAL-10N Dataset. Table 4 shows the experimental results on ANIMAL-10N, a real-world dataset with moderate label noise. We compared our method with multiple state-of-the-art baseline methods that are not limited to sample selection. For fair comparison, we use the same VGG- + +
DatasetCIFAR-100
Noise Rate80%90%80%90%
MethodBestLastBestLastPrecisionRecallPrecisionRecall
OT-Filter w/o Spar-Reg51.050.723.823.790.447.162.119.7
OT-Filter61.861.542.842.193.761.872.340.0
+ +![](images/f0e34641445936ad2f39c604e9eb6256f5c2dc7c84bb1bb5cee9e416a4382229.jpg) +Figure 2. The ablation study for the quality of sample selection w/o sparsity regularization on CIFAR-100 under noise rate of $90\%$ . The result shows that sparsity regularization brings a relatively uniform sample selection. + +19 network architecture. The results show we outperform all competitors, and have $2.1\%$ performance improvement over PLC [63], and $3.7\%$ performance improvement over SELFIE [51]. + +# 5.4. Analysis + +We conduct a series of ablation studies to understand the effectiveness of the key components of the OT-Filter, including the sparsity regularization of the optimal transport, the EM algorithm, and the entropic regularization coefficient. + +Effect of Sparsity Regularization. The sparsity regularization discussed in Section 4.3 is one of the key components in our OT-Filter, which is designed to improve the accuracy of the optimal transport by controlling the sparsity of transport. We therefore study the effect of the sparsity regularization over sample selection and performance under high noise on CIFAR-100 in Table 5. It can be observed, with the sparsity regularization, both the quality of sample selection and test accuracy are much better than the counterpart without it. In particular, when the noise rate is $90\%$ , it shows that the technique of sparsity regularization brings in over $10\%$ improvement in precision, about $21\%$ improvement in recall, and $19\%$ improvement in performance. Moreover, as shown in Figure 2, with the sparsity regularization, the number of samples selected for each class is more uniform. It is mainly attributed to discrete optimal transport allowing mass split. + +Table 5. Ablation study for sparsity regularization on CIFAR-100 dataset. The left part is the performance comparison and the right part is the quality of sample selection. + +
DatasetCIFAR-10
Noise Rate20%50%80%90%
MethodBestLastBestLastBestLastBestLast
OT-Filter w/o EM95.995.795.094.893.693.490.089.8
OT-Filter96.095.895.395.193.793.590.590.2
+ +Table 6. Ablation study for EM algorithm on CIFAR-10 dataset. + +
DatasetCIFAR-10
Noise Rate20%50%80%90%
MethodPrecisionRecallPrecisionRecallPrecisionRecallPrecisionRecall
ε = 199.793.999.093.994.690.688.684.2
ε = 599.793.899.093.894.690.688.684.2
ε = 1099.793.899.093.894.590.688.684.2
+ +Table 7. Ablation study for entropic regularization coefficient on CIFAR-10 dataset. + +Effect of Expectation Maximization. To filter noisy labels, we align all samples to clean barycenters via optimal transport. Therefore, the quality of clean barycenters affects the quality of sample selection and therefore the performance of the robust training. Table 6 indicates the impact of EM on CIFAR-10 dataset. It shows that the EM helps in improving the performance under different noise rates, and the significance is higher for high noise rates. + +Effect of Entropic Regularization Coefficient $\epsilon$ . The motivation for equipping the optimal transport with entropic regularization is to speed up its computation. However, the entropic regularization could potentially degrade the transport sparsity and therefore the quality of sample selection. To alleviate that, we propose sparsity regularization, whose effectiveness is examined in Table 5. Then, a by-product of sparsity regularization is the parameter robustness of the regularization coefficient $\epsilon$ . As shown in Table 7, the performance of sample selection stays quite stable when varying $\epsilon$ from 1 to 10. + +# 6. Conclusion + +In this work, we study the problem of sample selection for learning with noisy labels. We propose the OT-Filter, a novel technique for retrieving clean samples, which can also be combined with existing semi-supervised learning techniques for handling noisy labels. Unlike previous works relying on a-priori knowledge or conforming to confirmation bias, the OT-Filter enhances the measurement of data discrepancy, by lifting data from Euclidean space to the probability space, and thus improves the quality of sample selection. Extensive experiments on synthetic and real-world noisy datasets are conducted to evaluate our proposals. The results show its effectiveness and superiority. + +Acknowledgement. This work is supported by NSFC (No.61772492, 62072428) and the CAS Pioneer Hundred Talents Program. Xike Xie is the corresponding author. + +# References + +[1] Martial Agueh and Guillaume Carlier. Barycenters in the wasserstein space. SIAM Journal on Mathematical Analysis, 43(2):904-924, 2011. 2, 3, 7 +[2] Ethan Anderes, Steffen Borgwardt, and Jacob Miller. Discrete Wasserstein barycenters: Optimal transport for discrete data. Mathematical Methods of Operations Research, 84(2):389-409, 2016. 3 +[3] Eric Arazo, Diego Ortego, Paul Albert, Noel O'Connor, and Kevin McGuinness. Unsupervised label noise modeling and loss correction. In International conference on machine learning, pages 312-321. PMLR, 2019. 7 +[4] Martin Arjovsky, Soumith Chintala, and Léon Bottou. Wasserstein generative adversarial networks. In International conference on machine learning, pages 214-223. PMLR, 2017. 2 +[5] Yingbin Bai, Erkun Yang, Bo Han, Yanhua Yang, Jiatong Li, Yinian Mao, Gang Niu, and Tongliang Liu. Understanding and improving early stopping for learning with noisy labels. Advances in Neural Information Processing Systems, 34:24392-24403, 2021. 2 +[6] Mokhtar S Bazaraa, Hanif D Sherali, and Chitharanjan M Shetty. Nonlinear programming: theory and algorithms. John Wiley & Sons, 2013. 4 +[7] Jean-David Benamou, Guillaume Carlier, Marco Cuturi, Luca Nenna, and Gabriel Peyre. Iterative bregman projections for regularized transportation problems. SIAM Journal on Scientific Computing, 37(2):A1111-A1138, 2015. 3 +[8] David Berthelot, Nicholas Carlini, Ian Goodfellow, Nicolas Papernot, Avital Oliver, and Colin A Raffel. Mixmatch: A holistic approach to semi-supervised learning. Advances in neural information processing systems, 32, 2019. 3, 5 +[9] Dimitris Bertsimas and John N Tsitsiklis. Introduction to linear optimization, volume 6. 4 +[10] Stephen Boyd, Stephen P Boyd, and Lieven Vandenberghe. Convex optimization. Cambridge university press, 2004. 4 +[11] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2 +[12] Yingyi Chen, Xi Shen, Shell Xu Hu, and Johan AK Suykens. Boosting co-teaching with compression regularization for label noise. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2688-2692, 2021. 7 +[13] De Cheng, Tongliang Liu, Yixiong Ning, Nannan Wang, Bo Han, Gang Niu, Xinbo Gao, and Masashi Sugiyama. Instance-dependent label-noise learning with manifold-regularized transition matrix estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16630–16639, 2022. 1, 2 +[14] Hao Cheng, Zhaowei Zhu, Xingyu Li, Yifei Gong, Xing Sun, and Yang Liu. Learning with instance-dependent label noise: A sample sieve approach. In International Conference on Learning Representations, 2021. 2 +[15] Nicolas Courty, Rémi Flamary, Amaury Habrard, and Alain Rakotomamonjy. Joint distribution optimal transportation + +for domain adaptation. Advances in Neural Information Processing Systems, 30, 2017. 2 +[16] Nicolas Courty, Rémi Flamary, and Devis Tuia. Domain adaptation with regularized optimal transport. In Joint European Conference on Machine Learning and Knowledge Discovery in Databases, pages 274-289. Springer, 2014. 2, 5 +[17] Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. Advances in neural information processing systems, 26, 2013. 2, 3 +[18] Marco Cuturi and Arnaud Doucet. Fast computation of wasserstein barycenters. In International conference on machine learning, pages 685-693. PMLR, 2014. 2, 3, 4 +[19] Bharath Bhushan Damodaran, Rémi Flamary, Vivien Seguy, and Nicolas Courty. An entropic optimal transport loss for learning deep neural networks under label noise in remote sensing images. Comput. Vis. Image Underst., 191:102863, 2018. 2 +[20] Arthur P Dempster, Nan M Laird, and Donald B Rubin. Maximum likelihood from incomplete data via the em algorithm. Journal of the Royal Statistical Society: Series B (Methodological), 39(1):1-22, 1977. 4 +[21] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6 +[22] Erik Englesson and Hossein Azizpour. Generalized jensen-shannon divergence loss for learning with noisy labels. Advances in Neural Information Processing Systems, 34:30284-30297, 2021. 2, 7 +[23] Kilian Fatras, Bharath Bhushan Damodaran, Sylvain Lobry, Remi Flamary, Devis Tuia, and Nicolas Courty. Wasserstein adversarial regularization for learning with label noise. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 1 +[24] Alessio Figalli and Federico Glaudo. An Invitation to Optimal Transport, Wasserstein Distances, and Gradient Flows. 2021. 2 +[25] Aude Geneva, Gabriel Peyré, and Marco Cuturi. Learning generative models with sinkhorn divergences. In International Conference on Artificial Intelligence and Statistics, pages 1608-1617. PMLR, 2018. 2 +[26] Wenshuo Guo, Nhat Ho, and Michael Jordan. Fast algorithms for computational optimal transport and wasserstein barycenter. In International Conference on Artificial Intelligence and Statistics, pages 2088-2097. PMLR, 2020. 3 +[27] Bo Han, Quanming Yao, Xingrui Yu, Gang Niu, Miao Xu, Weihua Hu, Ivor Tsang, and Masashi Sugiyama. Coteaching: Robust training of deep neural networks with extremely noisy labels. Advances in neural information processing systems, 31, 2018. 1, 2, 6, 7 +[28] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1, 6 +[29] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Identity mappings in deep residual networks. In European conference on computer vision, pages 630-645. Springer, 2016. 1, 3, 6 + +[30] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In International conference on machine learning, pages 448-456. pmlr, 2015. 6 +[31] Ahmet Iscen, Jack Valmadre, Anurag Arnab, and Cordelia Schmid. Learning with neighbor consistency for noisy labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4672-4681, 2022. 1, 2 +[32] Lu Jiang, Zhengyuan Zhou, Thomas Leung, Li-Jia Li, and Li Fei-Fei. Mentornet: Learning data-driven curriculum for very deep neural networks on corrupted labels. In International conference on machine learning, pages 2304–2313. PMLR, 2018. 1, 2 +[33] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. Advances in neural information processing systems, 33:18661-18673, 2020. 2 +[34] Taehyeon Kim, Jongwoo Ko, JinHwan Choi, Se-Young Yun, et al. Fine samples for learning with noisy labels. Advances in Neural Information Processing Systems, 34:24137–24149, 2021. 1, 2, 7 +[35] Philip A Knight. The sinkhorn-knopp algorithm: convergence and applications. SIAM Journal on Matrix Analysis and Applications, 30(1):261-275, 2008. 2 +[36] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 1, 5 +[37] Kenneth Lange. MM optimization algorithms. SIAM, 2016. 5 +[38] Junnan Li, Richard Socher, and Steven CH Hoi. Dividemix: Learning with noisy labels as semi-supervised learning. arXiv preprint arXiv:2002.07394, 2020. 1, 2, 5, 6, 7 +[39] Sheng Liu, Jonathan Niles-Weed, Narges Razavian, and Carlos Fernandez-Granda. Early-learning regularization prevents memorization of noisy labels. Advances in Neural Information Processing Systems, 33, 2020. 2, 6 +[40] Baharan Mirzasoleiman, Kaidi Cao, and Jure Leskovec. Coresets for robust training of deep neural networks against noisy labels. Advances in Neural Information Processing Systems, 33:11465-11477, 2020. 2, 7 +[41] Kento Nishi, Yi Ding, Alex Rich, and Tobias Hollerer. Augmentation strategies for learning with noisy labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8022-8031, 2021. 2, 5 +[42] Diego Ortego, Eric Arazo, Paul Albert, Noel E O'Connor, and Kevin McGuinness. Multi-objective interpolation training for robustness to label noise. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6606-6615, 2021. 2 +[43] Giorgio Patrini, Alessandro Rozza, Aditya Krishna Menon, Richard Nock, and Lizhen Qu. Making deep neural networks robust to label noise: A loss correction approach. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1944-1952, 2017. 1, 2 +[44] Gabriel Peyré, Marco Cuturi, et al. Computational optimal transport: With applications to data science. Foundations + +and Trends® in Machine Learning, 11(5-6):355-607, 2019. 2, 3 +[45] Vikas C Raykar, Shipeng Yu, Linda H Zhao, Gerardo Hermosillo Valadez, Charles Florin, Luca Bogoni, and Linda Moy. Learning from crowds. Journal of machine learning research, 11(4), 2010. 1 +[46] Ievgen Redko, Nicolas Courty, Rémi Flamary, and Devis Tuia. Optimal transport for multi-source domain adaptation under target shift. In The 22nd International Conference on Artificial Intelligence and Statistics, pages 849-858. PMLR, 2019. 5 +[47] Oren Rippel, Michael Gelbart, and Ryan Adams. Learning ordered representations with nested dropout. In International Conference on Machine Learning, pages 1746-1754. PMLR, 2014. 7 +[48] Filippo Santambrogio. Optimal transport for applied mathematicians. Birkhäuser, NY, 55(58-63):94, 2015. 2, 3 +[49] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 1, 6 +[50] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. Advances in neural information processing systems, 33:596-608, 2020. 5 +[51] Hwanjun Song, Minseok Kim, and Jae-Gil Lee. Selfie: Refurbishing unclean samples for robust deep learning. In International Conference on Machine Learning, pages 5907-5915. PMLR, 2019. 5, 6, 7, 8 +[52] Fariborz Taherkhani, Ali Dabouei, Sobhan Soleymani, Jeremy Dawson, and Nasser M Nasrabadi. Transporting labels via hierarchical optimal transport for semi-supervised learning. In European Conference on Computer Vision, pages 509-526. Springer, 2020. 2 +[53] Fariborz Taherkhani, Ali Dabouei, Sobhan Soleymani, Jeremy Dawson, and Nasser M Nasrabadi. Self-supervised Wasserstein pseudo-labeling for semi-supervised image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12267-12277, 2021. 2 +[54] Cheng Tan, Jun Xia, Lirong Wu, and Stan Z Li. Co-learning: Learning from noisy labels with self-supervision. In Proceedings of the 29th ACM International Conference on Multimedia, pages 1405-1413, 2021. 2, 7 +[55] Cédric Villani. Optimal transport: old and new, volume 338. Springer, 2009. 2, 3 +[56] Cédric Villani. Topics in optimal transportation, volume 58. American Mathematical Soc., 2021. 2, 3 +[57] Yikai Wang, Xinwei Sun, and Yanwei Fu. Scalable penalized regression for noise detection in learning with noisy labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 346-355, 2022. 1 +[58] Pengxiang Wu, Songzhu Zheng, Mayank Goswami, Dimitris Metaxas, and Chao Chen. A topological filter for learning with label noise. Advances in neural information processing systems, 33:21382-21393, 2020. 1, 2, 7 + +[59] Xiaobo Xia, Tongliang Liu, Bo Han, Mingming Gong, Jun Yu, Gang Niu, and Masashi Sugiyama. Sample selection with uncertainty of losses for learning with noisy labels. In International Conference on Learning Representations, 2022. 2 +[60] Tong Xiao, Tian Xia, Yi Yang, Chang Huang, and Xiaogang Wang. Learning from massive noisy labeled data for image classification. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2691-2699, 2015. 1, 5, 6 +[61] Li Yi, Sheng Liu, Qi She, A Ian McLeod, and Boyu Wang. On learning contrastive representations for learning with noisy labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16682-16691, 2022. 2, 7 +[62] Xingrui Yu, Bo Han, Jiangchao Yao, Gang Niu, Ivor Tsang, and Masashi Sugiyama. How does disagreement help generalization against label corruption? In International Conference on Machine Learning, pages 7164-7173. PMLR, 2019. 1, 2, 7 +[63] Yikai Zhang, Songzhu Zheng, Pengxiang Wu, Mayank Goswami, and Chao Chen. Learning with feature-dependent label noise: A progressive approach. In Ninth International Conference on Learning Representations, volume 9, 2021. 7, 8 +[64] Zhilu Zhang and Mert Sabuncu. Generalized cross entropy loss for training deep neural networks with noisy labels. Advances in neural information processing systems, 31, 2018. 2, 5 +[65] Evgenii Zheltonozhskii, Chaim Baskin, Avi Mendelson, Alex M Bronstein, and Or Litany. Contrast to divide: Self-supervised pre-training for learning with noisy labels. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1657-1667, 2022. 2 +[66] Yudian Zheng, Guoliang Li, Yuanbing Li, Caihua Shan, and Reynold Cheng. Truth inference in crowdsourcing: Is the problem solved? Proceedings of the VLDB Endowment, 10(5):541-552, 2017. 1 \ No newline at end of file diff --git a/2023/OT-Filter_ An Optimal Transport Filter for Learning With Noisy Labels/images.zip b/2023/OT-Filter_ An Optimal Transport Filter for Learning With Noisy Labels/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..fb41ed98645c9f92cfc4c0dd335ea26a2ed7b276 --- /dev/null +++ b/2023/OT-Filter_ An Optimal Transport Filter for Learning With Noisy Labels/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1d3debc7c5b2222c540b500f072dfd76d2e25affa022b0b020c5d4bc86f6e5dd +size 319497 diff --git a/2023/OT-Filter_ An Optimal Transport Filter for Learning With Noisy Labels/layout.json b/2023/OT-Filter_ An Optimal Transport Filter for Learning With Noisy Labels/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..fec7a7604aa521d7bec7e48cd32658148aa5f783 --- /dev/null +++ b/2023/OT-Filter_ An Optimal Transport Filter for Learning With Noisy Labels/layout.json @@ -0,0 +1,11743 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 76, + 103, + 517, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 103, + 517, + 121 + ], + "spans": [ + { + "bbox": [ + 76, + 103, + 517, + 121 + ], + "type": "text", + "content": "OT-Filter: An Optimal Transport Filter for Learning with Noisy Labels" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 84, + 142, + 509, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 142, + 509, + 199 + ], + "spans": [ + { + "bbox": [ + 84, + 142, + 509, + 199 + ], + "type": "text", + "content": "Chuanwen Feng† Yilong Ren† Xike Xie \nUniversity of Science and Technology of China \nData Darkness Lab, MIRACLE Center, Suzhou Institute for Advanced Research, USTC \n{chuanwen,ylren}@mail.ustc.edu.cn, {xkxie}@ustc.edu.cn" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "spans": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 252, + 289, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 252, + 289, + 504 + ], + "spans": [ + { + "bbox": [ + 46, + 252, + 289, + 504 + ], + "type": "text", + "content": "The success of deep learning is largely attributed to the training over clean data. However, data is often coupled with noisy labels in practice. Learning with noisy labels is challenging because the performance of the deep neural networks (DNN) drastically degenerates, due to confirmation bias caused by the network memorization over noisy labels. To alleviate that, a recent prominent direction is on sample selection, which retrieves clean data samples from noisy samples, so as to enhance the model's robustness and tolerance to noisy labels. In this paper, we revamp the sample selection from the perspective of optimal transport theory and propose a novel method, called the OT-Filter. The OT-Filter provides geometrically meaningful distances and preserves distribution patterns to measure the data discrepancy, thus alleviating the confirmation bias. Extensive experiments on benchmarks, such as Clothing1M and ANIMAL-10N, show that the performance of the OT-Filter outperforms its counterparts. Meanwhile, results on benchmarks with synthetic labels, such as CIFAR-10/100, show the superiority of the OT-Filter in handling data labels of high noise." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 529, + 128, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 529, + 128, + 540 + ], + "spans": [ + { + "bbox": [ + 47, + 529, + 128, + 540 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 549, + 288, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 549, + 288, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 549, + 288, + 693 + ], + "type": "text", + "content": "Deep learning has achieved great success on a flurry of emerging applications, such as [28, 29, 32, 49]. It is believed that the phenomenal achievement of deep learning is largely attributed to accurate labels. However, the inaccuracy or imprecision of labels is inherent in real-world datasets, the so-called noisy label challenge. One way to alleviate that is to collect labels from internet queries over data-level tags, but the performance of deep neural networks (DNN) suffers drastically from the inaccuracy of such labels [27, 32]. A higher quality of data labels can be obtained by employing human workers, but the seemingly \"ground truth annotations\" inevitably involve human biases or mis" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 228, + 545, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 228, + 545, + 252 + ], + "spans": [ + { + "bbox": [ + 306, + 228, + 545, + 252 + ], + "type": "text", + "content": "takes [45,66]. More, the human annotation is expensive and time-consuming, especially for large-scale datasets." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 254, + 546, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 254, + 546, + 458 + ], + "spans": [ + { + "bbox": [ + 304, + 254, + 546, + 458 + ], + "type": "text", + "content": "There have been many works on handling noisy labels, such as regularization [23] and transition matrix [13, 43]. The regularization approach leverages the regularization bias to overcome the label noise issue. But the regularization bias is permanent [62], thus overfitting models to noisy labels. The transition matrix approach assumes that the transition probabilities between clean and noisy labels are fixed and independent of data samples. However, a quality label transition matrix is hard to be estimated, especially when the number of classes is big, making it fall short in handling noisy real-world datasets, such as [60] and [36]. A recent prominent direction is on adopting sample selection [27, 38, 58] for enhancing the label quality, by selecting clean samples from the noisy training dataset. In general, existing literatures in sample selection can be grouped to two categories, co-training networking [27, 62] and criterion-based filtering [31, 34, 57, 58]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 460, + 547, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 547, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 547, + 604 + ], + "type": "text", + "content": "The former utilizes the memorization of DNNs and multiple networks (e.g., co-teaching [27] and its variants [38, 62]) to filter label noise with small loss trick, so that a small set of clean samples are used as training examples. Letting alone the high training overhead of multiple networks, the disadvantages are two-fold: 1) it may require the a-priori knowledge of noisy rates to select the specified proportion of small loss samples as clean samples; 2) the small-loss trick is not tolerant to the error accumulation of network training once a clean sample is falsely recognized, the so-called confirmation bias, especially for labels with high noise where clean and noisy samples largely overlap." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": "The latter alleviates the problem by setting a specific criterion. Mostly, existing works [31] [58] adopt Euclidean distances for measuring the similarity between data samples. Distance-based filtering iteratively explores the neighborhood of the feature representation and infers/cleans sample labels by aggregating the information from their neighborhoods. Despite the simplicity, distance-based filtering is insufficient to address noisy labels, especially when the label noise is high, e.g., overlapped label classes." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 128, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 128, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 128, + 713 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 58, + 702, + 128, + 713 + ], + "type": "text", + "content": " Equal Contribution." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "16164" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 49, + 72, + 286, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 72, + 286, + 263 + ], + "spans": [ + { + "bbox": [ + 49, + 72, + 286, + 263 + ], + "type": "text", + "content": "In this paper, we revamp the sample selection from the perspective of optimal transport [56] and propose a novel filtering method, called the OT-Filter. In light of the optimal transport, we construct discrete probability measures over the sample features, which lifts the Euclidean space of feature vectors to a probability space. It thus enables a geometric way of measuring the discrepancy between probability measures, representing the corresponding sample features. In addition to the distance-based metric in the Euclidean space [31, 58], the OT-Filter also captures the distribution information in the probability space so as to alleviate the confirmation bias. Accordingly, a clean representation can be obtained for each class in the probability space. By optimizing the transport plan from a sample to a clean representation, one can better determine if a sample is clean or noisy, thus improving the quality of sample selection." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 264, + 286, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 264, + 286, + 384 + ], + "spans": [ + { + "bbox": [ + 49, + 264, + 286, + 384 + ], + "type": "text", + "content": "In general, the merits of the OT-Filter can be summarized as follows. First, it does not require any a-priori knowledge about the noisy rate of a dataset. Second, it utilizes the optimal transport which provides geometrically meaningful distances to exploit the sample discrepancy yet preserving the distribution patterns in corresponding probability space, making the sample selection of high quality and theoretical support. Third, it can be plugged to existing robust training paradigms, e.g., supervised and semi-supervised robust training." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 385, + 286, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 385, + 286, + 503 + ], + "spans": [ + { + "bbox": [ + 49, + 385, + 286, + 503 + ], + "type": "text", + "content": "We conduct extensive experiments with a series of synthetic and real datasets to gain insights into our proposals. The result shows that the OT-Filter follows state-of-the-art (SOTA) [38] when the noise rate is low, and dominates SOTA when the noise rate is high. For instance, the OT-Filter achieves about " + }, + { + "bbox": [ + 49, + 385, + 286, + 503 + ], + "type": "inline_equation", + "content": "14\\%" + }, + { + "bbox": [ + 49, + 385, + 286, + 503 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 49, + 385, + 286, + 503 + ], + "type": "inline_equation", + "content": "12\\%" + }, + { + "bbox": [ + 49, + 385, + 286, + 503 + ], + "type": "text", + "content": " higher accuracy than SOTA, in the presence of " + }, + { + "bbox": [ + 49, + 385, + 286, + 503 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 49, + 385, + 286, + 503 + ], + "type": "text", + "content": " noise rate, on synthetic datasets CIFAR-10 and CIFAR-100, respectively. Moreover, the OT-Filter outperforms the competitors on real datasets Clothing1M and ANIMAL-10N." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 506, + 286, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 506, + 286, + 577 + ], + "spans": [ + { + "bbox": [ + 49, + 506, + 286, + 577 + ], + "type": "text", + "content": "The rest of the paper is organized as follows. We review the existing literature in Section 2. In Section 3, we present preliminaries of optimal transport. We investigate our proposed OT-Filter in Section 4. Furthermore, we conduct extensive empirical studies in Section 5 and conclude the paper in Section 6." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 590, + 132, + 602 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 590, + 132, + 602 + ], + "spans": [ + { + "bbox": [ + 49, + 590, + 132, + 602 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 610, + 194, + 622 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 610, + 194, + 622 + ], + "spans": [ + { + "bbox": [ + 49, + 610, + 194, + 622 + ], + "type": "text", + "content": "2.1. Learning with Noisy Label" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 629, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 629, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 629, + 286, + 712 + ], + "type": "text", + "content": "A flurry of research methods [5, 13, 14, 27, 40, 54] was proposed for learning with noisy labels, which can be roughly divided into two categories, robust model [13, 22, 39, 43, 54, 61, 64] and clean sample selection [27, 32, 62]. The clean sample selection aims to select a clean subset from a noisy dataset [14, 27, 38]. One way is to train multiple networks to filter noisy samples. For instance, [32]" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 308, + 72, + 544, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 72, + 544, + 239 + ], + "spans": [ + { + "bbox": [ + 308, + 72, + 544, + 239 + ], + "type": "text", + "content": "proposed to pre-train an extra network to select clean samples for the main network. The inferiority of this method is on error accumulation. To relieve that, Han et al. [27] proposed Co-teaching that maintains two networks simultaneously, where one network selects clean samples for the other network by small loss trick. However, as the increase of training epochs, Co-teaching would converge to consensus gradually. Then, " + }, + { + "bbox": [ + 308, + 72, + 544, + 239 + ], + "type": "inline_equation", + "content": "\\mathrm{Co - teaching}^{+}" + }, + { + "bbox": [ + 308, + 72, + 544, + 239 + ], + "type": "text", + "content": " [62] was proposed to keep two networks diverged. To avoid confirmation bias, DivMix [38] proposed a hybrid framework with semi-supervised training, achieving the state-of-the-art performance. Based on DivMix, [41, 65] modified the training schema, such as self-supervised training [11, 33], to boost the performance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 308, + 240, + 544, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 240, + 544, + 371 + ], + "spans": [ + { + "bbox": [ + 308, + 240, + 544, + 371 + ], + "type": "text", + "content": "The other research line of sample selection is to filter noisy samples based on specific criteria [31, 34, 42, 59]. For example, Xia et al. [59] proposed to reduce the uncertainty of loss using specific strategies. Iscen et al. [31] proposed to leverage similarities between training examples in the feature space, encouraging the prediction of each example to be close to its nearest neighbors. Also, in Euclidean space, Wu et al. [58] proposed the TopoFilter that filters noisy samples using k-nearest neighborhood distance between prelogits. Fine [34] proposed to use the principal components of latent representations to select clean samples." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 308, + 382, + 416, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 382, + 416, + 395 + ], + "spans": [ + { + "bbox": [ + 308, + 382, + 416, + 395 + ], + "type": "text", + "content": "2.2. Optimal Transport" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 308, + 401, + 544, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 401, + 544, + 616 + ], + "spans": [ + { + "bbox": [ + 308, + 401, + 544, + 616 + ], + "type": "text", + "content": "The optimal transport problem [55, 56] aims to move mass from a probability measure to another probability measure at a minimum cost. It defines the Wasserstein distance [24] and provides a geometric way to measure the discrepancy between probability measures. One factor that limits the wide application of optimal transport is the high computational cost. To improve its scalability, Cuturi [17] proposed an entropic regularization for the transport plan, which yields an efficient algorithm with the matrix scaling method of Sinkhorn-Knopp [35]. Recently, computational optimal transport [44, 48] has found many applications in various areas, e.g., generative models [4, 25], domain adaptation [15, 16], and semi-supervised learning [52, 53]. In particular, [1] studied the barycenter in Wasserstein space. [18] explored the efficient computing of Wasserstein barycenter. [19] proposed to use entropic optimal transport loss, based on joint distribution optimal transport [15], to build a robust training." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 308, + 618, + 544, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 618, + 544, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 618, + 544, + 712 + ], + "type": "text", + "content": "In this paper, we study the problem of clean sample selection from the perspective of optimal transport to combat label noise. The proposed method can potentially capture geometric information from the probability space. Unlike the mechanism of multiple network training, it only trains a single network without a-priori information on noise rates. Moreover, it can be easily plugged to multiple robust training paradigms." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 750, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 750, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 750, + 317, + 757 + ], + "type": "text", + "content": "16165" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 71, + 124, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 71, + 124, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 124, + 85 + ], + "type": "text", + "content": "3. Preliminary" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 91, + 287, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 91, + 287, + 198 + ], + "spans": [ + { + "bbox": [ + 46, + 91, + 287, + 198 + ], + "type": "text", + "content": "In this section, we present preliminaries on optimal transport, making a basis for subsequent proposed techniques. The optimal transport [56] is to seek an optimal transport plan between two measures at a minimal cost (e.g. the Wasserstein distance), which provides a geometric way of matching probability measures. Here we first introduce the general definition of Wasserstein distance induced by optimal transport problem. More details about optimal transport can be found in [44, 48, 55]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 198, + 287, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 198, + 287, + 259 + ], + "spans": [ + { + "bbox": [ + 46, + 198, + 287, + 259 + ], + "type": "text", + "content": "Wassertein Distance. Let " + }, + { + "bbox": [ + 46, + 198, + 287, + 259 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 46, + 198, + 287, + 259 + ], + "type": "text", + "content": " be a locally complete and separable metric space, the " + }, + { + "bbox": [ + 46, + 198, + 287, + 259 + ], + "type": "inline_equation", + "content": "\\mathcal{P}(S)" + }, + { + "bbox": [ + 46, + 198, + 287, + 259 + ], + "type": "text", + "content": " be the Borel probability measures set on " + }, + { + "bbox": [ + 46, + 198, + 287, + 259 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 46, + 198, + 287, + 259 + ], + "type": "text", + "content": ". For any " + }, + { + "bbox": [ + 46, + 198, + 287, + 259 + ], + "type": "inline_equation", + "content": "\\mathcal{X}, \\mathcal{Y} \\subset \\mathcal{S}" + }, + { + "bbox": [ + 46, + 198, + 287, + 259 + ], + "type": "text", + "content": ", given " + }, + { + "bbox": [ + 46, + 198, + 287, + 259 + ], + "type": "inline_equation", + "content": "\\mu \\in \\mathcal{P}(\\mathcal{X})" + }, + { + "bbox": [ + 46, + 198, + 287, + 259 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 198, + 287, + 259 + ], + "type": "inline_equation", + "content": "\\nu \\in \\mathcal{P}(\\mathcal{Y})" + }, + { + "bbox": [ + 46, + 198, + 287, + 259 + ], + "type": "text", + "content": ", the optimal transport defines a Wasserstein distance between two probability measures, denoted as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 65, + 263, + 268, + 293 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 263, + 268, + 293 + ], + "spans": [ + { + "bbox": [ + 65, + 263, + 268, + 293 + ], + "type": "interline_equation", + "content": "W _ {p} (\\mu , \\nu) := \\left(\\inf _ {\\pi (\\mu , \\nu)} \\int_ {\\mathcal {X} \\times \\mathcal {Y}} | | x - y | | ^ {p} d \\pi (\\mu , \\nu)\\right) ^ {\\frac {1}{p}}", + "image_path": "bae6f368836ae28e62c6ac339754c8f0a6808518166f1d94a36c71243233f500.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 297, + 287, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 297, + 287, + 321 + ], + "spans": [ + { + "bbox": [ + 46, + 297, + 287, + 321 + ], + "type": "inline_equation", + "content": "p \\geq 1" + }, + { + "bbox": [ + 46, + 297, + 287, + 321 + ], + "type": "text", + "content": ", where the " + }, + { + "bbox": [ + 46, + 297, + 287, + 321 + ], + "type": "inline_equation", + "content": "\\pi(\\mu, \\nu)" + }, + { + "bbox": [ + 46, + 297, + 287, + 321 + ], + "type": "text", + "content": " is the set of joint probability measures with marginal " + }, + { + "bbox": [ + 46, + 297, + 287, + 321 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 46, + 297, + 287, + 321 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 297, + 287, + 321 + ], + "type": "inline_equation", + "content": "\\nu" + }, + { + "bbox": [ + 46, + 297, + 287, + 321 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 327, + 200, + 340 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 327, + 200, + 340 + ], + "spans": [ + { + "bbox": [ + 47, + 327, + 200, + 340 + ], + "type": "text", + "content": "3.1. Discrete Optimal Transport" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 345, + 287, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 345, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 46, + 345, + 287, + 464 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 46, + 345, + 287, + 464 + ], + "type": "inline_equation", + "content": "\\Delta_{n} = \\{\\mathbf{a}\\in \\mathbb{R}_{+}^{n}|\\sum_{i = 1}^{n}a_{i} = 1,\\forall a_{i}\\geq 0\\}" + }, + { + "bbox": [ + 46, + 345, + 287, + 464 + ], + "type": "text", + "content": " be a probability simplex in dimension " + }, + { + "bbox": [ + 46, + 345, + 287, + 464 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 345, + 287, + 464 + ], + "type": "text", + "content": " . Consider two empirical probability measures " + }, + { + "bbox": [ + 46, + 345, + 287, + 464 + ], + "type": "inline_equation", + "content": "\\mu = \\sum_{i = 1}^{n}a_{i}\\delta_{x_{i}}" + }, + { + "bbox": [ + 46, + 345, + 287, + 464 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 345, + 287, + 464 + ], + "type": "inline_equation", + "content": "\\nu = \\sum_{j = 1}^{m}b_{j}\\delta_{y_{j}}" + }, + { + "bbox": [ + 46, + 345, + 287, + 464 + ], + "type": "text", + "content": " , defined on metric space " + }, + { + "bbox": [ + 46, + 345, + 287, + 464 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 46, + 345, + 287, + 464 + ], + "type": "text", + "content": " with support " + }, + { + "bbox": [ + 46, + 345, + 287, + 464 + ], + "type": "inline_equation", + "content": "\\{x_i\\}_{i = 1}^n" + }, + { + "bbox": [ + 46, + 345, + 287, + 464 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 345, + 287, + 464 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 46, + 345, + 287, + 464 + ], + "type": "text", + "content": " with support " + }, + { + "bbox": [ + 46, + 345, + 287, + 464 + ], + "type": "inline_equation", + "content": "\\{y_j\\}_{j = 1}^m" + }, + { + "bbox": [ + 46, + 345, + 287, + 464 + ], + "type": "text", + "content": " respectively. Here the weight vector " + }, + { + "bbox": [ + 46, + 345, + 287, + 464 + ], + "type": "inline_equation", + "content": "\\mathbf{a} = (a_1,a_2,\\dots,a_n)" + }, + { + "bbox": [ + 46, + 345, + 287, + 464 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 345, + 287, + 464 + ], + "type": "inline_equation", + "content": "\\mathbf{b} = (b_{1},b_{2},\\dots,b_{m})" + }, + { + "bbox": [ + 46, + 345, + 287, + 464 + ], + "type": "text", + "content": " live in " + }, + { + "bbox": [ + 46, + 345, + 287, + 464 + ], + "type": "inline_equation", + "content": "\\Delta_{n}" + }, + { + "bbox": [ + 46, + 345, + 287, + 464 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 345, + 287, + 464 + ], + "type": "inline_equation", + "content": "\\Delta_{m}" + }, + { + "bbox": [ + 46, + 345, + 287, + 464 + ], + "type": "text", + "content": " , respectively. The " + }, + { + "bbox": [ + 46, + 345, + 287, + 464 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 46, + 345, + 287, + 464 + ], + "type": "text", + "content": " stands for the Dirac unit mass function. Given a transport cost " + }, + { + "bbox": [ + 46, + 345, + 287, + 464 + ], + "type": "inline_equation", + "content": "c:\\mathcal{X}\\times \\mathcal{Y}\\to \\mathbb{R}_+" + }, + { + "bbox": [ + 46, + 345, + 287, + 464 + ], + "type": "text", + "content": " the discrete optimal transport between probability measures " + }, + { + "bbox": [ + 46, + 345, + 287, + 464 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 46, + 345, + 287, + 464 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 345, + 287, + 464 + ], + "type": "inline_equation", + "content": "\\nu" + }, + { + "bbox": [ + 46, + 345, + 287, + 464 + ], + "type": "text", + "content": " can be formulated as" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 100, + 469, + 287, + 504 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 469, + 287, + 504 + ], + "spans": [ + { + "bbox": [ + 100, + 469, + 287, + 504 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} W _ {p} ^ {p} (\\mu , \\nu) := \\min _ {M \\in \\Pi (\\mu , \\nu)} \\langle C, M \\rangle_ {F} \\tag {1} \\\\ s. t. M \\mathbf {1} _ {m} = \\mu , M ^ {T} \\mathbf {1} _ {n} = \\nu \\\\ \\end{array}", + "image_path": "b35dbc2ff158be82e2777ba64c83842db7014e50226fe120cbf2b8a0bb0e44d6.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 510, + 287, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 287, + 595 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 287, + 595 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 510, + 287, + 595 + ], + "type": "inline_equation", + "content": "C \\in \\mathbb{R}_+^{n \\times m}" + }, + { + "bbox": [ + 46, + 510, + 287, + 595 + ], + "type": "text", + "content": " is the transport cost matrix, and " + }, + { + "bbox": [ + 46, + 510, + 287, + 595 + ], + "type": "inline_equation", + "content": "c_{ij}" + }, + { + "bbox": [ + 46, + 510, + 287, + 595 + ], + "type": "text", + "content": " represents a unit transport cost from " + }, + { + "bbox": [ + 46, + 510, + 287, + 595 + ], + "type": "inline_equation", + "content": "x_i" + }, + { + "bbox": [ + 46, + 510, + 287, + 595 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 510, + 287, + 595 + ], + "type": "inline_equation", + "content": "y_j" + }, + { + "bbox": [ + 46, + 510, + 287, + 595 + ], + "type": "text", + "content": ". The " + }, + { + "bbox": [ + 46, + 510, + 287, + 595 + ], + "type": "inline_equation", + "content": "M \\in \\mathbb{R}_+^{n \\times m}" + }, + { + "bbox": [ + 46, + 510, + 287, + 595 + ], + "type": "text", + "content": " is transport plan in which " + }, + { + "bbox": [ + 46, + 510, + 287, + 595 + ], + "type": "inline_equation", + "content": "m_{ij}" + }, + { + "bbox": [ + 46, + 510, + 287, + 595 + ], + "type": "text", + "content": " denotes the amount of mass transported from " + }, + { + "bbox": [ + 46, + 510, + 287, + 595 + ], + "type": "inline_equation", + "content": "x_i" + }, + { + "bbox": [ + 46, + 510, + 287, + 595 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 510, + 287, + 595 + ], + "type": "inline_equation", + "content": "y_j" + }, + { + "bbox": [ + 46, + 510, + 287, + 595 + ], + "type": "text", + "content": ". All feasible transport plans constitute transport polytope " + }, + { + "bbox": [ + 46, + 510, + 287, + 595 + ], + "type": "inline_equation", + "content": "\\Pi(\\mu, \\nu)" + }, + { + "bbox": [ + 46, + 510, + 287, + 595 + ], + "type": "text", + "content": ". The " + }, + { + "bbox": [ + 46, + 510, + 287, + 595 + ], + "type": "inline_equation", + "content": "\\langle C, M \\rangle_F" + }, + { + "bbox": [ + 46, + 510, + 287, + 595 + ], + "type": "text", + "content": " is the Frobenius inner product of matrices and equals to " + }, + { + "bbox": [ + 46, + 510, + 287, + 595 + ], + "type": "inline_equation", + "content": "tr(C^T M)" + }, + { + "bbox": [ + 46, + 510, + 287, + 595 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 601, + 218, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 601, + 218, + 613 + ], + "spans": [ + { + "bbox": [ + 47, + 601, + 218, + 613 + ], + "type": "text", + "content": "3.2. Regularized Optimal Transport" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 619, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 619, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 619, + 287, + 689 + ], + "type": "text", + "content": "The discrete optimal transport formulation, in essence, is a convex optimization problem. More precisely, it's a linear programming problem. Unfortunately, this linear programming problem has a cubic computing complexity. A way to relieve this is to leverage entropic regularization [17] formulated as" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 106, + 696, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 696, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 106, + 696, + 287, + 715 + ], + "type": "interline_equation", + "content": "\\min _ {M \\in \\Pi (\\mu , \\nu)} \\langle C, M \\rangle_ {F} - \\epsilon H (M) \\tag {2}", + "image_path": "c673f913941fc9fa82af952fa5fb6825115075b29a6a6684674999896ad3841f.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "\\epsilon > 0" + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": " is the regularization coefficient, and " + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "H(M)" + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": " is the entropic regularization term, which promotes an efficient computation for transportation via the matrix scaling algorithm [17], given by" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 349, + 129, + 502, + 155 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 129, + 502, + 155 + ], + "spans": [ + { + "bbox": [ + 349, + 129, + 502, + 155 + ], + "type": "interline_equation", + "content": "H (M) := - \\sum_ {i j} M _ {i j} \\left(\\log \\left(M _ {i j}\\right) - 1\\right)", + "image_path": "dbfe8d1afc9b28c5e6b6cb42cb86e945a1ce0ef26e06916c1c5d30b29dd45d08.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 163, + 440, + 176 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 163, + 440, + 176 + ], + "spans": [ + { + "bbox": [ + 306, + 163, + 440, + 176 + ], + "type": "text", + "content": "3.3. Wasserstein Barycenter" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 182, + 545, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 182, + 545, + 230 + ], + "spans": [ + { + "bbox": [ + 305, + 182, + 545, + 230 + ], + "type": "text", + "content": "Given " + }, + { + "bbox": [ + 305, + 182, + 545, + 230 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 305, + 182, + 545, + 230 + ], + "type": "text", + "content": " probability measures " + }, + { + "bbox": [ + 305, + 182, + 545, + 230 + ], + "type": "inline_equation", + "content": "\\{\\nu_{1},\\nu_{2},\\dots,\\nu_{N}\\}" + }, + { + "bbox": [ + 305, + 182, + 545, + 230 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 182, + 545, + 230 + ], + "type": "inline_equation", + "content": "\\nu_{i} \\in \\mathcal{P}(S)" + }, + { + "bbox": [ + 305, + 182, + 545, + 230 + ], + "type": "text", + "content": ", each of which has finite supports and second moments. A Wasserstein barycenter of these measures is a probability measure " + }, + { + "bbox": [ + 305, + 182, + 545, + 230 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 305, + 182, + 545, + 230 + ], + "type": "text", + "content": ", satisfying:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 311, + 238, + 539, + 270 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 238, + 539, + 270 + ], + "spans": [ + { + "bbox": [ + 311, + 238, + 539, + 270 + ], + "type": "interline_equation", + "content": "\\mu := \\inf \\sum_ {i = 1} ^ {N} \\lambda_ {i} W _ {2} ^ {2} (\\nu_ {i}, \\mu), \\quad s. t. \\sum_ {i = 1} ^ {N} \\lambda_ {i} = 1, \\forall \\lambda_ {i} \\geq 0", + "image_path": "9c16baa692146853fee53f247a120ad6624a20f96b5504d2f258322379f420da.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 278, + 545, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 278, + 545, + 388 + ], + "spans": [ + { + "bbox": [ + 305, + 278, + 545, + 388 + ], + "type": "text", + "content": "The notion was first proposed by [1], where some elegant properties of Wasserstein barycenter were presented. Then [2] discussed it on the discrete case and showed that the problem of finding Wasserstein barycenter over the space " + }, + { + "bbox": [ + 305, + 278, + 545, + 388 + ], + "type": "inline_equation", + "content": "\\mathcal{P}(\\mathcal{S})" + }, + { + "bbox": [ + 305, + 278, + 545, + 388 + ], + "type": "text", + "content": " can be reduced to a simpler space " + }, + { + "bbox": [ + 305, + 278, + 545, + 388 + ], + "type": "inline_equation", + "content": "\\mathcal{O}_r(S)" + }, + { + "bbox": [ + 305, + 278, + 545, + 388 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 305, + 278, + 545, + 388 + ], + "type": "inline_equation", + "content": "r = \\sum_{i=1}^{N} e_i - N + 1" + }, + { + "bbox": [ + 305, + 278, + 545, + 388 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 305, + 278, + 545, + 388 + ], + "type": "inline_equation", + "content": "e_i" + }, + { + "bbox": [ + 305, + 278, + 545, + 388 + ], + "type": "text", + "content": " is the number of components of " + }, + { + "bbox": [ + 305, + 278, + 545, + 388 + ], + "type": "inline_equation", + "content": "\\nu_i, i \\in [1, N]" + }, + { + "bbox": [ + 305, + 278, + 545, + 388 + ], + "type": "text", + "content": ". To find Wasserstein barycenter over space " + }, + { + "bbox": [ + 305, + 278, + 545, + 388 + ], + "type": "inline_equation", + "content": "\\mathcal{O}_r(S)" + }, + { + "bbox": [ + 305, + 278, + 545, + 388 + ], + "type": "text", + "content": ", a set of efficient algorithms [7, 18, 26] were proposed." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 399, + 365, + 410 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 399, + 365, + 410 + ], + "spans": [ + { + "bbox": [ + 306, + 399, + 365, + 410 + ], + "type": "text", + "content": "4. Methods" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 418, + 373, + 429 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 418, + 373, + 429 + ], + "spans": [ + { + "bbox": [ + 306, + 418, + 373, + 429 + ], + "type": "text", + "content": "4.1. Overview" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 436, + 545, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 436, + 545, + 508 + ], + "spans": [ + { + "bbox": [ + 304, + 436, + 545, + 508 + ], + "type": "text", + "content": "In this section, we study the OT-Filter, the proposed method based on optimal transport for learning with noisy labels in a semi-supervised paradigm. The key idea of our proposed method is to transport samples with noisy labels to samples with clean labels at a minimum cost, which makes sample selection of high quality and theoretical support." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 508, + 545, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 508, + 545, + 579 + ], + "spans": [ + { + "bbox": [ + 304, + 508, + 545, + 579 + ], + "type": "text", + "content": "The mechanism of the OT-Filter, consists of two phases, representation phase (Section 4.2) and transportation phase (Section 4.3). Then, the OT-Filter can be plugged to off-the-shelf SSLs for robust training, e.g., MixMatch [8], by treating clean samples as labeled data and noisy samples as unlabeled samples (Section 4.4)." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 305, + 587, + 545, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 587, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 305, + 587, + 545, + 612 + ], + "type": "text", + "content": "4.2. Representation Phase: Finding Clean Representations via Wasserstein Barycenters" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 545, + 713 + ], + "type": "text", + "content": "In the representation phase, we first use a pre-trained network, e.g., ResNet [29], to extract a set of feature representations of the labeled samples. Then, we define a discrete probability measure over the feature representations of each class to find the clean representations (or prototypes) via Wasserstein barycenters. Since there are mistrabeled samples, the barycenters we obtained may be noisy. Therefore, we iteratively optimize the barycenters with the expectation" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "16166" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 107 + ], + "type": "text", + "content": "maximization (EM) algorithm [20] to converge to clean representations. The details about the representation phase are covered in Algorithm 1." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 108, + 286, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 108, + 286, + 240 + ], + "spans": [ + { + "bbox": [ + 46, + 108, + 286, + 240 + ], + "type": "text", + "content": "Feature Extraction. The purpose of the feature extraction is to transform the input noisy label dataset into a feature-label dataset via a network. A noisy label dataset of " + }, + { + "bbox": [ + 46, + 108, + 286, + 240 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 108, + 286, + 240 + ], + "type": "text", + "content": " samples can be represented by " + }, + { + "bbox": [ + 46, + 108, + 286, + 240 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{D}} = \\{(\\mathbf{x}_i,\\hat{y}_i)\\}_{i = 1}^N" + }, + { + "bbox": [ + 46, + 108, + 286, + 240 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 108, + 286, + 240 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i\\in \\mathbb{R}^d" + }, + { + "bbox": [ + 46, + 108, + 286, + 240 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 46, + 108, + 286, + 240 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 108, + 286, + 240 + ], + "type": "text", + "content": "-th sample and " + }, + { + "bbox": [ + 46, + 108, + 286, + 240 + ], + "type": "inline_equation", + "content": "\\dot{y}_i\\in \\{0,1\\} ^K" + }, + { + "bbox": [ + 46, + 108, + 286, + 240 + ], + "type": "text", + "content": " denotes the corresponding noisy label over " + }, + { + "bbox": [ + 46, + 108, + 286, + 240 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 108, + 286, + 240 + ], + "type": "text", + "content": " classes. Then, a feature extraction network " + }, + { + "bbox": [ + 46, + 108, + 286, + 240 + ], + "type": "inline_equation", + "content": "f(\\omega ,\\mathbf{x}_i):\\mathbb{R}^d\\to \\mathbb{R}^m" + }, + { + "bbox": [ + 46, + 108, + 286, + 240 + ], + "type": "text", + "content": " with parameters " + }, + { + "bbox": [ + 46, + 108, + 286, + 240 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 46, + 108, + 286, + 240 + ], + "type": "text", + "content": " maps a sample " + }, + { + "bbox": [ + 46, + 108, + 286, + 240 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 46, + 108, + 286, + 240 + ], + "type": "text", + "content": " to a " + }, + { + "bbox": [ + 46, + 108, + 286, + 240 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 108, + 286, + 240 + ], + "type": "text", + "content": "-dimensional feature representation, denoted as " + }, + { + "bbox": [ + 46, + 108, + 286, + 240 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}_i\\in \\mathbb{R}^m" + }, + { + "bbox": [ + 46, + 108, + 286, + 240 + ], + "type": "text", + "content": ". Therefore, all output feature representations and their corresponding labels constitute a feature-label set, denoted as " + }, + { + "bbox": [ + 46, + 108, + 286, + 240 + ], + "type": "inline_equation", + "content": "\\tilde{D} = \\{(\\tilde{\\mathbf{x}}_i,\\hat{y}_i)\\}_{i = 1}^N" + }, + { + "bbox": [ + 46, + 108, + 286, + 240 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 239, + 286, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 239, + 286, + 371 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 286, + 371 + ], + "type": "text", + "content": "Probability Measure Modeling. The purpose of probability measure modeling is to define a probability measure over the feature-label set. This definition will potentially lift a Euclidean space " + }, + { + "bbox": [ + 46, + 239, + 286, + 371 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 46, + 239, + 286, + 371 + ], + "type": "text", + "content": ", in which the feature representations originally live, to a probability space, abbreviated as " + }, + { + "bbox": [ + 46, + 239, + 286, + 371 + ], + "type": "inline_equation", + "content": "\\mathcal{P}(\\mathcal{D})" + }, + { + "bbox": [ + 46, + 239, + 286, + 371 + ], + "type": "text", + "content": ". Assume that the labels are of " + }, + { + "bbox": [ + 46, + 239, + 286, + 371 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 239, + 286, + 371 + ], + "type": "text", + "content": " classes. First, we split the feature-label set " + }, + { + "bbox": [ + 46, + 239, + 286, + 371 + ], + "type": "inline_equation", + "content": "\\tilde{D}" + }, + { + "bbox": [ + 46, + 239, + 286, + 371 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 46, + 239, + 286, + 371 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 239, + 286, + 371 + ], + "type": "text", + "content": " subsets, i.e., " + }, + { + "bbox": [ + 46, + 239, + 286, + 371 + ], + "type": "inline_equation", + "content": "\\{s_1,s_2,\\dots,s_K\\}" + }, + { + "bbox": [ + 46, + 239, + 286, + 371 + ], + "type": "text", + "content": " according to the label class. For one feature representation " + }, + { + "bbox": [ + 46, + 239, + 286, + 371 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}_i" + }, + { + "bbox": [ + 46, + 239, + 286, + 371 + ], + "type": "text", + "content": " in class " + }, + { + "bbox": [ + 46, + 239, + 286, + 371 + ], + "type": "inline_equation", + "content": "k\\in [1,K]" + }, + { + "bbox": [ + 46, + 239, + 286, + 371 + ], + "type": "text", + "content": ", we represent it as " + }, + { + "bbox": [ + 46, + 239, + 286, + 371 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}_i^k,i\\in [1,|s_k|]" + }, + { + "bbox": [ + 46, + 239, + 286, + 371 + ], + "type": "text", + "content": ". Then, all feature representations in subset " + }, + { + "bbox": [ + 46, + 239, + 286, + 371 + ], + "type": "inline_equation", + "content": "s_k" + }, + { + "bbox": [ + 46, + 239, + 286, + 371 + ], + "type": "text", + "content": " construct a probability measure, denoted as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 380, + 203, + 414 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 380, + 203, + 414 + ], + "spans": [ + { + "bbox": [ + 130, + 380, + 203, + 414 + ], + "type": "interline_equation", + "content": "\\mathcal {Q} = \\frac {1}{| s _ {k} |} \\sum_ {i = 1} ^ {| s _ {k} |} \\delta_ {\\tilde {\\mathbf {x}} _ {i} ^ {k}}", + "image_path": "f395722378f7164f464c518a50d2811bc5481fb70639bb4ceaf266d84e63dc41.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 424, + 286, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 424, + 286, + 448 + ], + "spans": [ + { + "bbox": [ + 47, + 424, + 286, + 448 + ], + "type": "text", + "content": "where the " + }, + { + "bbox": [ + 47, + 424, + 286, + 448 + ], + "type": "inline_equation", + "content": "\\delta_{\\widetilde{\\mathbf{x}}_i^k}" + }, + { + "bbox": [ + 47, + 424, + 286, + 448 + ], + "type": "text", + "content": " is a Dirac unit mass on " + }, + { + "bbox": [ + 47, + 424, + 286, + 448 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathbf{x}}_i^k" + }, + { + "bbox": [ + 47, + 424, + 286, + 448 + ], + "type": "text", + "content": ", and for simplicity, we use uniform weights." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 449, + 286, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 286, + 496 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 286, + 496 + ], + "type": "text", + "content": "Clean Representation Retrieval. The problem of finding the representation of any subset " + }, + { + "bbox": [ + 46, + 449, + 286, + 496 + ], + "type": "inline_equation", + "content": "s_k" + }, + { + "bbox": [ + 46, + 449, + 286, + 496 + ], + "type": "text", + "content": " is equivalent to finding the corresponding Wasserstein barycenter " + }, + { + "bbox": [ + 46, + 449, + 286, + 496 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_k" + }, + { + "bbox": [ + 46, + 449, + 286, + 496 + ], + "type": "text", + "content": " with finite supports, formalized as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 111, + 505, + 286, + 526 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 505, + 286, + 526 + ], + "spans": [ + { + "bbox": [ + 111, + 505, + 286, + 526 + ], + "type": "interline_equation", + "content": "\\mathcal {B} _ {k} := \\inf _ {\\mathcal {B} \\in \\mathcal {O} _ {r} (\\tilde {\\mathcal {D}})} W _ {2} ^ {2} (\\mathcal {B}, \\mathcal {Q}) \\tag {3}", + "image_path": "76b8c488c4e4618667737afcfb2a07bb8d4e63d8977a5b9ce8356159c37a2a47.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 536, + 286, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 536, + 286, + 572 + ], + "spans": [ + { + "bbox": [ + 47, + 536, + 286, + 572 + ], + "type": "text", + "content": "where the " + }, + { + "bbox": [ + 47, + 536, + 286, + 572 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 47, + 536, + 286, + 572 + ], + "type": "text", + "content": " is the Wasserstein distance, and " + }, + { + "bbox": [ + 47, + 536, + 286, + 572 + ], + "type": "inline_equation", + "content": "r = |s_k|" + }, + { + "bbox": [ + 47, + 536, + 286, + 572 + ], + "type": "text", + "content": ". As mentioned in Section 3, from the perspective of linear programming, the above optimization admits the dual:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 93, + 580, + 241, + 601 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 580, + 241, + 601 + ], + "spans": [ + { + "bbox": [ + 93, + 580, + 241, + 601 + ], + "type": "interline_equation", + "content": "D (\\alpha , \\beta) = \\max _ {(\\alpha , \\beta) \\in R (C)} \\alpha^ {T} \\mathcal {B} + \\beta^ {T} \\mathcal {Q}", + "image_path": "5c6c32a2322294fcf360e35ef2ed5c92c9e963f1a99f64b4d13343b1c962b544.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 609, + 217, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 609, + 217, + 621 + ], + "spans": [ + { + "bbox": [ + 47, + 609, + 217, + 621 + ], + "type": "text", + "content": "where the polyhedron of dual variables is:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 52, + 631, + 281, + 644 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 631, + 281, + 644 + ], + "spans": [ + { + "bbox": [ + 52, + 631, + 281, + 644 + ], + "type": "interline_equation", + "content": "R (C) = \\left\\{\\left(\\alpha , \\beta\\right) \\in \\mathbb {R} ^ {r} \\times \\mathbb {R} ^ {r} \\mid \\alpha_ {i} + \\beta_ {j} \\leq C _ {i j} \\wedge i, j \\in [ r ] \\right\\}", + "image_path": "3d5822bfbdc1eb7aae4951a03a50abd665f792d9699bc14b92843fef7465f4d0.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 653, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 286, + 713 + ], + "type": "text", + "content": "By means of duality theory [6,9,10], we can find the relation of solution between the primal and dual problem. Here, the optimal solution of " + }, + { + "bbox": [ + 46, + 653, + 286, + 713 + ], + "type": "inline_equation", + "content": "D(\\alpha, \\beta)" + }, + { + "bbox": [ + 46, + 653, + 286, + 713 + ], + "type": "text", + "content": " is a subgradient of " + }, + { + "bbox": [ + 46, + 653, + 286, + 713 + ], + "type": "inline_equation", + "content": "W_2^2(\\mathcal{Q}, \\mathcal{B})" + }, + { + "bbox": [ + 46, + 653, + 286, + 713 + ], + "type": "text", + "content": " with respect to " + }, + { + "bbox": [ + 46, + 653, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 46, + 653, + 286, + 713 + ], + "type": "text", + "content": ". Thus, we employ a simple projected subgradient [18] to optimize Equation 3. over " + }, + { + "bbox": [ + 46, + 653, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 46, + 653, + 286, + 713 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "text", + "content": "Despite we have found a set of Wasserstein barycenters " + }, + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{B}_i\\}_{i=1}^K" + }, + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "text", + "content": ", the barycenters may be noisy since there are mislabeled data. Moreover, the clean barycenters are unknown. Therefore, for each class, we consider the ideal and clean barycenter " + }, + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_c" + }, + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "text", + "content": " as a hidden variable, the obtained noisy barycenter " + }, + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "text", + "content": " as an observed variable. Then, the problem of finding clean barycenter " + }, + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_c" + }, + { + "bbox": [ + 304, + 72, + 545, + 167 + ], + "type": "text", + "content": " is to maximize the following log-likelihood:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 358, + 177, + 493, + 202 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 177, + 493, + 202 + ], + "spans": [ + { + "bbox": [ + 358, + 177, + 493, + 202 + ], + "type": "interline_equation", + "content": "\\hat {\\theta} \\leftarrow \\arg \\max _ {\\theta} \\log \\sum_ {\\mathcal {B} _ {c}} p (\\mathcal {B} _ {c}, \\mathcal {B} | \\theta)", + "image_path": "e03eded9d01d42593da79918d887cb9e35daeb8b5391cb8a17db6defd7b6cc13.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 212, + 545, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 212, + 545, + 307 + ], + "spans": [ + { + "bbox": [ + 304, + 212, + 545, + 307 + ], + "type": "text", + "content": "where the " + }, + { + "bbox": [ + 304, + 212, + 545, + 307 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 212, + 545, + 307 + ], + "type": "text", + "content": " is the parameters of probabilistic model. By this way, we can optimize the noisy barycenters iteratively with the EM algorithm to converge to clean barycenters. In E-step, we construct optimal transport between all samples and clean barycenters, thereby we can infer the labels of all samples. In M-step, we refine the barycenters with inferred label obtained from E-step. The overall computing details see Algorithm 1." + } + ] + } + ], + "index": 14 + }, + { + "type": "code", + "bbox": [ + 306, + 345, + 545, + 626 + ], + "blocks": [ + { + "bbox": [ + 306, + 319, + 545, + 342 + ], + "lines": [ + { + "bbox": [ + 306, + 319, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 306, + 319, + 545, + 342 + ], + "type": "text", + "content": "Algorithm 1 Find Clean Representation via Wasserstein Barycenter" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 306, + 345, + 545, + 626 + ], + "lines": [ + { + "bbox": [ + 306, + 345, + 545, + 626 + ], + "spans": [ + { + "bbox": [ + 306, + 345, + 545, + 626 + ], + "type": "text", + "content": "Require: " + }, + { + "bbox": [ + 306, + 345, + 545, + 626 + ], + "type": "inline_equation", + "content": "\\mathcal{Q} = \\{\\tilde{\\mathbf{x}}_i\\}_{i=1}^{|s_k|} \\in \\mathbb{R}^{m \\times r}" + }, + { + "bbox": [ + 306, + 345, + 545, + 626 + ], + "type": "text", + "content": ", regularization coefficient " + }, + { + "bbox": [ + 306, + 345, + 545, + 626 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 306, + 345, + 545, + 626 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 306, + 345, + 545, + 626 + ], + "type": "inline_equation", + "content": "\\mathcal{B} \\in \\mathbb{R}^{m \\times e}" + }, + { + "bbox": [ + 306, + 345, + 545, + 626 + ], + "type": "text", + "content": " \n1: Initialize: " + }, + { + "bbox": [ + 306, + 345, + 545, + 626 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 306, + 345, + 545, + 626 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 306, + 345, + 545, + 626 + ], + "type": "inline_equation", + "content": "\\nu \\gets \\mathcal{B}" + }, + { + "bbox": [ + 306, + 345, + 545, + 626 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 306, + 345, + 545, + 626 + ], + "type": "inline_equation", + "content": "\\mu \\gets \\mathcal{Q}" + }, + { + "bbox": [ + 306, + 345, + 545, + 626 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 306, + 345, + 545, + 626 + ], + "type": "inline_equation", + "content": "\\eta = 1/2" + }, + { + "bbox": [ + 306, + 345, + 545, + 626 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 306, + 345, + 545, + 626 + ], + "type": "inline_equation", + "content": "t = 2" + }, + { + "bbox": [ + 306, + 345, + 545, + 626 + ], + "type": "text", + "content": " \n2: while " + }, + { + "bbox": [ + 306, + 345, + 545, + 626 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 306, + 345, + 545, + 626 + ], + "type": "text", + "content": " not converged do \n3: set " + }, + { + "bbox": [ + 306, + 345, + 545, + 626 + ], + "type": "inline_equation", + "content": "\\nu = \\mathbf{1}_e / e" + }, + { + "bbox": [ + 306, + 345, + 545, + 626 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 306, + 345, + 545, + 626 + ], + "type": "inline_equation", + "content": "\\mu = \\mathbf{1}_r / r" + }, + { + "bbox": [ + 306, + 345, + 545, + 626 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 306, + 345, + 545, + 626 + ], + "type": "inline_equation", + "content": "\\hat{\\nu} = \\tilde{\\nu} = \\mathbf{1} / e" + }, + { + "bbox": [ + 306, + 345, + 545, + 626 + ], + "type": "text", + "content": " \n4: while " + }, + { + "bbox": [ + 306, + 345, + 545, + 626 + ], + "type": "inline_equation", + "content": "\\nu" + }, + { + "bbox": [ + 306, + 345, + 545, + 626 + ], + "type": "text", + "content": " not converged do \n5: " + }, + { + "bbox": [ + 306, + 345, + 545, + 626 + ], + "type": "inline_equation", + "content": "\\beta \\gets (t + 1) / 2" + }, + { + "bbox": [ + 306, + 345, + 545, + 626 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 306, + 345, + 545, + 626 + ], + "type": "inline_equation", + "content": "\\nu \\gets (1 - \\beta^{-1})\\hat{\\nu} + \\beta^{-1}\\tilde{\\nu}" + }, + { + "bbox": [ + 306, + 345, + 545, + 626 + ], + "type": "text", + "content": " \n6: " + }, + { + "bbox": [ + 306, + 345, + 545, + 626 + ], + "type": "inline_equation", + "content": "\\alpha \\gets" + }, + { + "bbox": [ + 306, + 345, + 545, + 626 + ], + "type": "text", + "content": " optimizer of " + }, + { + "bbox": [ + 306, + 345, + 545, + 626 + ], + "type": "inline_equation", + "content": "\\mathbf{d}(\\nu, \\mu, C)" + }, + { + "bbox": [ + 306, + 345, + 545, + 626 + ], + "type": "text", + "content": " \n7: " + }, + { + "bbox": [ + 306, + 345, + 545, + 626 + ], + "type": "inline_equation", + "content": "\\tilde{\\nu} \\gets \\tilde{\\nu} \\circ \\exp(-\\beta\\alpha)" + }, + { + "bbox": [ + 306, + 345, + 545, + 626 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 306, + 345, + 545, + 626 + ], + "type": "inline_equation", + "content": "\\tilde{\\nu} \\gets \\tilde{\\nu} / \\tilde{\\nu}^T\\mathbf{1}" + }, + { + "bbox": [ + 306, + 345, + 545, + 626 + ], + "type": "text", + "content": " \n8: " + }, + { + "bbox": [ + 306, + 345, + 545, + 626 + ], + "type": "inline_equation", + "content": "\\hat{\\nu} \\gets (1 - \\beta^{-1})\\hat{\\nu} + \\beta^{-1}\\tilde{\\nu}" + }, + { + "bbox": [ + 306, + 345, + 545, + 626 + ], + "type": "text", + "content": " \n9: " + }, + { + "bbox": [ + 306, + 345, + 545, + 626 + ], + "type": "inline_equation", + "content": "t \\gets t + 1" + }, + { + "bbox": [ + 306, + 345, + 545, + 626 + ], + "type": "text", + "content": " \n10: end while \n11: " + }, + { + "bbox": [ + 306, + 345, + 545, + 626 + ], + "type": "inline_equation", + "content": "\\nu \\gets \\hat{\\nu}" + }, + { + "bbox": [ + 306, + 345, + 545, + 626 + ], + "type": "text", + "content": " \n12: " + }, + { + "bbox": [ + 306, + 345, + 545, + 626 + ], + "type": "inline_equation", + "content": "\\mathcal{M} \\gets" + }, + { + "bbox": [ + 306, + 345, + 545, + 626 + ], + "type": "text", + "content": " optimizer of " + }, + { + "bbox": [ + 306, + 345, + 545, + 626 + ], + "type": "inline_equation", + "content": "\\mathbf{p}(\\nu, \\mu, C)" + }, + { + "bbox": [ + 306, + 345, + 545, + 626 + ], + "type": "text", + "content": " \n13: " + }, + { + "bbox": [ + 306, + 345, + 545, + 626 + ], + "type": "inline_equation", + "content": "\\mathcal{B} \\gets (1 - \\eta)\\mathcal{B} + \\eta (\\mathcal{QM})diag(\\nu^{-1})" + }, + { + "bbox": [ + 306, + 345, + 545, + 626 + ], + "type": "text", + "content": " \n14: end while \n15: while " + }, + { + "bbox": [ + 306, + 345, + 545, + 626 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_c" + }, + { + "bbox": [ + 306, + 345, + 545, + 626 + ], + "type": "text", + "content": " not converged do \n16: E-Step: \n17: " + }, + { + "bbox": [ + 306, + 345, + 545, + 626 + ], + "type": "inline_equation", + "content": "\\mathbb{L}_{\\hat{\\mathcal{D}}} \\gets \\mathrm{OT}(\\mu, \\nu, C)" + }, + { + "bbox": [ + 306, + 345, + 545, + 626 + ], + "type": "text", + "content": " \n18: M-Step: \n19: " + }, + { + "bbox": [ + 306, + 345, + 545, + 626 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_c \\gets" + }, + { + "bbox": [ + 306, + 345, + 545, + 626 + ], + "type": "text", + "content": " refine " + }, + { + "bbox": [ + 306, + 345, + 545, + 626 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 306, + 345, + 545, + 626 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 306, + 345, + 545, + 626 + ], + "type": "inline_equation", + "content": "\\mathbb{L}_{\\hat{\\mathcal{D}}}" + }, + { + "bbox": [ + 306, + 345, + 545, + 626 + ], + "type": "text", + "content": " \n20: end while \n21: Return " + }, + { + "bbox": [ + 306, + 345, + 545, + 626 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_c" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "code_body" + } + ], + "index": 16, + "sub_type": "algorithm" + }, + { + "bbox": [ + 306, + 647, + 545, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 647, + 545, + 672 + ], + "spans": [ + { + "bbox": [ + 306, + 647, + 545, + 672 + ], + "type": "text", + "content": "4.3. Transportation Phase: Transporting Noisy Labels via Regularized Optimal Transport" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 677, + 545, + 713 + ], + "type": "text", + "content": "In the transportation phase, noisy feature representations are transported to clean representations. In particular, we align all feature representations, including those of noisy" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16167" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 227 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 227 + ], + "type": "text", + "content": "samples, to the clean representations (obtained in the representation phase) via sparsity regularized optimal transport. This operation potentially detects the corrupted labels based on the result of the optimal transport. Here, the mechanism of our sample selection via optimal transport is described as follows. First, we align the clean barycenters to the all rest samples via optimal transport. Based on the result of optimal transport, we can infer the labels of the all rest samples. We consider the route with the max mass transport that forms a coupling and then assign the label of a clean barycenter to the samples on this route. Second, we select the sample whose inferred label is identical to the original label as a clean sample." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 228, + 289, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 228, + 289, + 287 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 289, + 287 + ], + "type": "text", + "content": "Sparsity Regularization. After the representation phase, clean representations are obtained, which are essentially the barycenters of label classes. The clean representation of a class provides the guidance of identifying whether samples are clean or noisy w.r.t. the given class." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 288, + 289, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 288, + 289, + 431 + ], + "spans": [ + { + "bbox": [ + 46, + 288, + 289, + 431 + ], + "type": "text", + "content": "We consider the barycenters of all the " + }, + { + "bbox": [ + 46, + 288, + 289, + 431 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 288, + 289, + 431 + ], + "type": "text", + "content": " classes as a discrete probability measure, denoted as " + }, + { + "bbox": [ + 46, + 288, + 289, + 431 + ], + "type": "inline_equation", + "content": "\\mu = \\frac{1}{K}\\sum_{k = 1}^{K}\\delta_{u_k}" + }, + { + "bbox": [ + 46, + 288, + 289, + 431 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 46, + 288, + 289, + 431 + ], + "type": "inline_equation", + "content": "\\delta_{\\mu_k}" + }, + { + "bbox": [ + 46, + 288, + 289, + 431 + ], + "type": "text", + "content": " is a Dirac unit mass on " + }, + { + "bbox": [ + 46, + 288, + 289, + 431 + ], + "type": "inline_equation", + "content": "\\mu_{k}" + }, + { + "bbox": [ + 46, + 288, + 289, + 431 + ], + "type": "text", + "content": ". Accordingly, the feature representations of the entire dataset construct a discrete probability measure " + }, + { + "bbox": [ + 46, + 288, + 289, + 431 + ], + "type": "inline_equation", + "content": "\\nu = \\frac{1}{N}\\sum_{i = 1}^{N}\\delta_{\\tilde{\\mathbf{x}}_i}" + }, + { + "bbox": [ + 46, + 288, + 289, + 431 + ], + "type": "text", + "content": ". Then, we transport feature representations " + }, + { + "bbox": [ + 46, + 288, + 289, + 431 + ], + "type": "inline_equation", + "content": "\\nu" + }, + { + "bbox": [ + 46, + 288, + 289, + 431 + ], + "type": "text", + "content": " of all samples to the clean barycenters " + }, + { + "bbox": [ + 46, + 288, + 289, + 431 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 46, + 288, + 289, + 431 + ], + "type": "text", + "content": ", in order to detect if a sample is clean or mislabeled. Inspired by [16,46], we view the clean barycenters as the source domain and the feature representations of samples as the target domain. Then, we recast the alignments as an optimal transport problem with a sparsity regularization [16], written as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 437, + 287, + 465 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 287, + 465 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 287, + 465 + ], + "type": "interline_equation", + "content": "\\min _ {M \\in \\Pi (\\mu , \\nu)} \\langle C, M \\rangle_ {F} - \\epsilon H (M) + \\gamma \\sum_ {j, c} | | M (i _ {c}, j) | | _ {1} ^ {\\frac {1}{2}} \\tag {4}", + "image_path": "c0fd24d5f7622cbaa3aa30efbd41591c9eba5b3ee976c4a90656b90b549f3715.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 472, + 289, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 472, + 289, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 289, + 640 + ], + "type": "text", + "content": "The second term is an entropic regularization, as described in Equation 2. The third term is a sparsity regularization with coefficient " + }, + { + "bbox": [ + 46, + 472, + 289, + 640 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 472, + 289, + 640 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 472, + 289, + 640 + ], + "type": "inline_equation", + "content": "i_c" + }, + { + "bbox": [ + 46, + 472, + 289, + 640 + ], + "type": "text", + "content": " index the line if its element in class " + }, + { + "bbox": [ + 46, + 472, + 289, + 640 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 472, + 289, + 640 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 472, + 289, + 640 + ], + "type": "inline_equation", + "content": "M(i_c, j)" + }, + { + "bbox": [ + 46, + 472, + 289, + 640 + ], + "type": "text", + "content": " is a vector consisting of the " + }, + { + "bbox": [ + 46, + 472, + 289, + 640 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 472, + 289, + 640 + ], + "type": "text", + "content": "-th column of " + }, + { + "bbox": [ + 46, + 472, + 289, + 640 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 472, + 289, + 640 + ], + "type": "text", + "content": " in class " + }, + { + "bbox": [ + 46, + 472, + 289, + 640 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 472, + 289, + 640 + ], + "type": "text", + "content": ". The " + }, + { + "bbox": [ + 46, + 472, + 289, + 640 + ], + "type": "inline_equation", + "content": "\\left\\| \\cdot \\right\\|_1^{1/2}" + }, + { + "bbox": [ + 46, + 472, + 289, + 640 + ], + "type": "text", + "content": " denotes an " + }, + { + "bbox": [ + 46, + 472, + 289, + 640 + ], + "type": "inline_equation", + "content": "L_1" + }, + { + "bbox": [ + 46, + 472, + 289, + 640 + ], + "type": "text", + "content": " norm with the power of " + }, + { + "bbox": [ + 46, + 472, + 289, + 640 + ], + "type": "inline_equation", + "content": "\\frac{1}{2}" + }, + { + "bbox": [ + 46, + 472, + 289, + 640 + ], + "type": "text", + "content": ". The entropic regularization enables discrete optimal transport to be efficient. However, it disperses the transport route, which negatively affects sample selection. Therefore, to improve the quality of sample selection, we employ a sparsity regularization. The sparsity regularization promotes a feature representation that would be aligned to one of the clean barycenters and penalizes transportation matrix " + }, + { + "bbox": [ + 46, + 472, + 289, + 640 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 472, + 289, + 640 + ], + "type": "text", + "content": " that aligns together feature representations with different labels." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 642, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 289, + 715 + ], + "type": "text", + "content": "Optimization. Despite the consistent sparsity achieved by the regularization, the objective function of Equation 4 is non-convex. Moreover, the " + }, + { + "bbox": [ + 46, + 642, + 289, + 715 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 46, + 642, + 289, + 715 + ], + "type": "text", + "content": " norm with the power in label regularization term is a concave function. A common optimization strategy [37] is to construct a convex upper bound, represented by convex functions, over the non-convex prob" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": "lem. By doing so, we can use the optimal solution of convex upper bound to approximate the optimal solution of the original non-convex problem. If applying the simplest linear approximation, we have" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 354, + 128, + 499, + 156 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 128, + 499, + 156 + ], + "spans": [ + { + "bbox": [ + 354, + 128, + 499, + 156 + ], + "type": "interline_equation", + "content": "\\sum_ {j, c} \\left| \\left| M \\left(i _ {c}, j\\right) \\right| \\right| _ {1} ^ {\\frac {1}{2}} \\leq \\langle G, M \\rangle_ {F} + \\mathcal {L}", + "image_path": "a836af3417f2b3c3db0be20a3bc3c1998f6ba7a6a193a8347a721c2824871724.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 165, + 547, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 165, + 547, + 203 + ], + "spans": [ + { + "bbox": [ + 304, + 165, + 547, + 203 + ], + "type": "text", + "content": "where matrix " + }, + { + "bbox": [ + 304, + 165, + 547, + 203 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 304, + 165, + 547, + 203 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 304, + 165, + 547, + 203 + ], + "type": "inline_equation", + "content": "\\frac{1}{2} (||\\hat{M}(i_c, j) + \\delta||^{-\\frac{1}{2}})" + }, + { + "bbox": [ + 304, + 165, + 547, + 203 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 165, + 547, + 203 + ], + "type": "inline_equation", + "content": "\\hat{M}" + }, + { + "bbox": [ + 304, + 165, + 547, + 203 + ], + "type": "text", + "content": " is a given start point. The steps of optimization process are depicted as Algorithm 2." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 214, + 545, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 214, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 306, + 214, + 545, + 239 + ], + "type": "text", + "content": "Algorithm 2 Transportation: Transport Noisy Label via Regularized Optimal Transport" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 241, + 495, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 241, + 495, + 254 + ], + "spans": [ + { + "bbox": [ + 306, + 241, + 495, + 254 + ], + "type": "text", + "content": "Require: optimal cost " + }, + { + "bbox": [ + 306, + 241, + 495, + 254 + ], + "type": "inline_equation", + "content": "C_{\\text{min}}" + }, + { + "bbox": [ + 306, + 241, + 495, + 254 + ], + "type": "text", + "content": " from Equation 1," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 254, + 506, + 324 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 312, + 254, + 394, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 254, + 394, + 265 + ], + "spans": [ + { + "bbox": [ + 312, + 254, + 394, + 265 + ], + "type": "text", + "content": "1: Initialize: " + }, + { + "bbox": [ + 312, + 254, + 394, + 265 + ], + "type": "inline_equation", + "content": "G = 0" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 266, + 430, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 266, + 430, + 277 + ], + "spans": [ + { + "bbox": [ + 312, + 266, + 430, + 277 + ], + "type": "text", + "content": "2: while G not converged do" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 277, + 422, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 277, + 422, + 289 + ], + "spans": [ + { + "bbox": [ + 312, + 277, + 422, + 289 + ], + "type": "text", + "content": "3: " + }, + { + "bbox": [ + 312, + 277, + 422, + 289 + ], + "type": "inline_equation", + "content": "C_{min} \\gets C_{min} + G" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 290, + 506, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 290, + 506, + 301 + ], + "spans": [ + { + "bbox": [ + 312, + 290, + 506, + 301 + ], + "type": "text", + "content": "4: " + }, + { + "bbox": [ + 312, + 290, + 506, + 301 + ], + "type": "inline_equation", + "content": "M\\gets" + }, + { + "bbox": [ + 312, + 290, + 506, + 301 + ], + "type": "text", + "content": " optimizer of Equation 2 with " + }, + { + "bbox": [ + 312, + 290, + 506, + 301 + ], + "type": "inline_equation", + "content": "C_{min}" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 302, + 438, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 302, + 438, + 313 + ], + "spans": [ + { + "bbox": [ + 312, + 302, + 438, + 313 + ], + "type": "text", + "content": "5: " + }, + { + "bbox": [ + 312, + 302, + 438, + 313 + ], + "type": "inline_equation", + "content": "G\\gets" + }, + { + "bbox": [ + 312, + 302, + 438, + 313 + ], + "type": "text", + "content": " update " + }, + { + "bbox": [ + 312, + 302, + 438, + 313 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 312, + 302, + 438, + 313 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 312, + 302, + 438, + 313 + ], + "type": "inline_equation", + "content": "M" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 314, + 367, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 314, + 367, + 324 + ], + "spans": [ + { + "bbox": [ + 312, + 314, + 367, + 324 + ], + "type": "text", + "content": "6: end while" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 347, + 440, + 360 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 347, + 440, + 360 + ], + "spans": [ + { + "bbox": [ + 306, + 347, + 440, + 360 + ], + "type": "text", + "content": "4.4. Robust Training in SSL" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 365, + 545, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 365, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 545, + 533 + ], + "type": "text", + "content": "After all feature representations are transported to the clean barycenters. We filter noisy samples and obtain two data subsets, namely the clean sample set and noisy sample set. As aforementioned, our sample selection method is flexible enough for supporting various robust training paradigms, e.g., the robust supervised training [64] and the robust semi-supervised learning [50]. Following [38], we select MixMatch [8] with data augmentation strategy [41] as the practice of robust semi-supervised training but with fewer operations. We admit the labels of clean samples and neglect the labels of noisy samples, and then we use labeled data and unlabeled data for training in a semi-supervised learning fashion. The overall training process is described as Algorithm 3." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 544, + 388, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 544, + 388, + 558 + ], + "spans": [ + { + "bbox": [ + 306, + 544, + 388, + 558 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 563, + 368, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 563, + 368, + 575 + ], + "spans": [ + { + "bbox": [ + 306, + 563, + 368, + 575 + ], + "type": "text", + "content": "5.1. Datasets" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 582, + 545, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 629 + ], + "type": "text", + "content": "We perform extensive experiments on four benchmark datasets: CIFAR-10 [36] and CIFAR-100 [36] with synthetic label noise, and Clothing1M [60] and ANIMAL-10N [51] with real label noise." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": "CIFAR10/100. We evaluate the OT-Filter on CIFAR10 and CIFAR-100 [36] with synthetic noise. Both datasets contain 60,000 " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": " color images, in 10 and 100 classes, respectively. From both datasets, 50K images are used for training, and 10K images are used for testing. Since noise characteristics can hardly be determined in advance, synthetic noise is commonly taken for controlling the noise rate" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16168" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 48, + 87, + 287, + 310 + ], + "blocks": [ + { + "bbox": [ + 48, + 72, + 200, + 83 + ], + "lines": [ + { + "bbox": [ + 48, + 72, + 200, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 200, + 83 + ], + "type": "text", + "content": "Algorithm 3 Robust Training in SSL" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 48, + 87, + 287, + 310 + ], + "lines": [ + { + "bbox": [ + 48, + 87, + 287, + 310 + ], + "spans": [ + { + "bbox": [ + 48, + 87, + 287, + 310 + ], + "type": "text", + "content": "Require: network parameter " + }, + { + "bbox": [ + 48, + 87, + 287, + 310 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 48, + 87, + 287, + 310 + ], + "type": "text", + "content": ", Beta distribution parameter " + }, + { + "bbox": [ + 48, + 87, + 287, + 310 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 48, + 87, + 287, + 310 + ], + "type": "text", + "content": ", weight of unlabeled loss " + }, + { + "bbox": [ + 48, + 87, + 287, + 310 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathcal{U}}" + }, + { + "bbox": [ + 48, + 87, + 287, + 310 + ], + "type": "text", + "content": ", batch size " + }, + { + "bbox": [ + 48, + 87, + 287, + 310 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 48, + 87, + 287, + 310 + ], + "type": "text", + "content": ", number of augmentations " + }, + { + "bbox": [ + 48, + 87, + 287, + 310 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 48, + 87, + 287, + 310 + ], + "type": "text", + "content": ", sharpening temperature " + }, + { + "bbox": [ + 48, + 87, + 287, + 310 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 48, + 87, + 287, + 310 + ], + "type": "text", + "content": ", labeled data " + }, + { + "bbox": [ + 48, + 87, + 287, + 310 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 48, + 87, + 287, + 310 + ], + "type": "text", + "content": ", unlabeled data " + }, + { + "bbox": [ + 48, + 87, + 287, + 310 + ], + "type": "inline_equation", + "content": "\\mathcal{U}" + }, + { + "bbox": [ + 48, + 87, + 287, + 310 + ], + "type": "text", + "content": " \n1: while " + }, + { + "bbox": [ + 48, + 87, + 287, + 310 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 48, + 87, + 287, + 310 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 48, + 87, + 287, + 310 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 48, + 87, + 287, + 310 + ], + "type": "text", + "content": " do \n2: " + }, + { + "bbox": [ + 48, + 87, + 287, + 310 + ], + "type": "inline_equation", + "content": "\\hat{x}_b \\gets \\text{augment}(x_b)" + }, + { + "bbox": [ + 48, + 87, + 287, + 310 + ], + "type": "text", + "content": " \n3: for " + }, + { + "bbox": [ + 48, + 87, + 287, + 310 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 48, + 87, + 287, + 310 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 48, + 87, + 287, + 310 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 48, + 87, + 287, + 310 + ], + "type": "text", + "content": " do \n4: " + }, + { + "bbox": [ + 48, + 87, + 287, + 310 + ], + "type": "inline_equation", + "content": "\\hat{u}_{b,k} \\gets \\text{augment}(u_b)" + }, + { + "bbox": [ + 48, + 87, + 287, + 310 + ], + "type": "text", + "content": " \n5: end for \n6: " + }, + { + "bbox": [ + 48, + 87, + 287, + 310 + ], + "type": "inline_equation", + "content": "\\hat{q}_b \\gets \\frac{1}{K} \\sum_k p_{model}(\\hat{u}_{b,k}; \\omega)" + }, + { + "bbox": [ + 48, + 87, + 287, + 310 + ], + "type": "text", + "content": " \n7: " + }, + { + "bbox": [ + 48, + 87, + 287, + 310 + ], + "type": "inline_equation", + "content": "q_b \\gets \\text{sharpen}(\\hat{q}_b, T)" + }, + { + "bbox": [ + 48, + 87, + 287, + 310 + ], + "type": "text", + "content": " \n8: end while \n9: " + }, + { + "bbox": [ + 48, + 87, + 287, + 310 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{X}} = \\{(\\hat{x}_b, p_b)\\}_{b=1}^B, \\hat{\\mathcal{U}} = \\{(\\hat{u}_{b,k}, q_b)\\}_{b,k=1}^{B,K}" + }, + { + "bbox": [ + 48, + 87, + 287, + 310 + ], + "type": "text", + "content": " \n10: " + }, + { + "bbox": [ + 48, + 87, + 287, + 310 + ], + "type": "inline_equation", + "content": "\\mathcal{W} = \\text{shuffle}(\\text{concat}(\\hat{\\mathcal{X}}, \\hat{\\mathcal{U}}))" + }, + { + "bbox": [ + 48, + 87, + 287, + 310 + ], + "type": "text", + "content": " \n11: " + }, + { + "bbox": [ + 48, + 87, + 287, + 310 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{X}}' = \\text{Mixup}(\\hat{\\mathcal{X}}_i, \\mathcal{W}_i), i \\in [1, |\\hat{\\mathcal{X}}|]" + }, + { + "bbox": [ + 48, + 87, + 287, + 310 + ], + "type": "text", + "content": " \n12: " + }, + { + "bbox": [ + 48, + 87, + 287, + 310 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{U}}' = \\text{Mixup}(\\hat{\\mathcal{U}}_i, \\mathcal{W}_{i+ \\hat{\\mathcal{U}}}), i \\in [1, |\\hat{\\mathcal{U}}|]" + }, + { + "bbox": [ + 48, + 87, + 287, + 310 + ], + "type": "text", + "content": " \n13: " + }, + { + "bbox": [ + 48, + 87, + 287, + 310 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathcal{X}} \\gets \\text{CE}(\\mathcal{X}')" + }, + { + "bbox": [ + 48, + 87, + 287, + 310 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 48, + 87, + 287, + 310 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathcal{U}} \\gets \\text{MAE}(\\mathcal{U}')" + }, + { + "bbox": [ + 48, + 87, + 287, + 310 + ], + "type": "text", + "content": " \n14: " + }, + { + "bbox": [ + 48, + 87, + 287, + 310 + ], + "type": "inline_equation", + "content": "\\mathcal{L} \\gets \\mathcal{L}_{\\mathcal{X}} + \\lambda_{\\mathcal{U}} \\mathcal{L}_{\\mathcal{U}}" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "code_body" + } + ], + "index": 1, + "sub_type": "algorithm" + }, + { + "bbox": [ + 46, + 334, + 287, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 334, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 46, + 334, + 287, + 441 + ], + "type": "text", + "content": "to deliberately evaluate learning algorithms with noisy labels. Thus, following previous works [27, 38], we consider two types of label noise, i.e., symmetric and asymmetric label noise. The symmetric label noise is generated by randomly flipping labels of a portion of samples from one class to all other possible classes. The asymmetric label noise is designed to follow the structure of real-world label noise, where labels are flipped to similar classes within the superclasses." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 442, + 287, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 442, + 287, + 537 + ], + "spans": [ + { + "bbox": [ + 46, + 442, + 287, + 537 + ], + "type": "text", + "content": "Clothing1M. Clothing1M is a large-scale dataset with noisy labels [60], containing over 1M images obtained from online shopping websites. The labels are from 14 classes generated based on surrounding texts provided by the sellers, and the noise rate is estimated around " + }, + { + "bbox": [ + 46, + 442, + 287, + 537 + ], + "type": "inline_equation", + "content": "38.5\\%" + }, + { + "bbox": [ + 46, + 442, + 287, + 537 + ], + "type": "text", + "content": " [39]. Also, the dataset provides 3 clean datasets for training, validation, and testing, containing 50K, 14K, and 10K images, respectively." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 539, + 287, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 539, + 287, + 610 + ], + "spans": [ + { + "bbox": [ + 46, + 539, + 287, + 610 + ], + "type": "text", + "content": "ANIMAL-10N. ANIMAL-10N is a real-world noisy dataset released by [51], which is crawled from several online search engines using predefined labels as searching keywords. There are in total 55K images, of which 50K images are for training, and 5K images are for testing. The noise rate was estimated to be around " + }, + { + "bbox": [ + 46, + 539, + 287, + 610 + ], + "type": "inline_equation", + "content": "8\\%" + }, + { + "bbox": [ + 46, + 539, + 287, + 610 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 622, + 179, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 622, + 179, + 634 + ], + "spans": [ + { + "bbox": [ + 47, + 622, + 179, + 634 + ], + "type": "text", + "content": "5.2. Implementation Details" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 641, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 287, + 713 + ], + "type": "text", + "content": "For CIFAR-10 and CIFAR-100, we use the PreAct ResNet18 [29] as the backbone and train it using the SGD optimizer with the following settings: a momentum of 0.9, a weight decay of 5e-4, and a batch size of 128. The learning rate was initialized as 0.02 and reduced by a factor of 10, after 150 epochs. The network was trained for 300" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 95 + ], + "type": "text", + "content": "epochs. The warmup period is 10 epochs for CIFAR-10, and 30 epochs for CIFAR-100." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 97, + 545, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 97, + 545, + 191 + ], + "spans": [ + { + "bbox": [ + 304, + 97, + 545, + 191 + ], + "type": "text", + "content": "For Clothing1M, we use the ResNet50 [28] pre-trained on ImageNet [21] as the backbone and train it using the SGD optimizer with the following settings: a momentum of 0.9, a weight decay of 1e-3, and a batch size of 32. The learning rate was initialized as 0.002 and reduced by a factor of 10 after 50 epochs. The network was trained for 120 epochs. For each epoch, we sample 1000 mini-batches from the training data." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 193, + 545, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 193, + 545, + 252 + ], + "spans": [ + { + "bbox": [ + 304, + 193, + 545, + 252 + ], + "type": "text", + "content": "For ANIMAL-10N, we use VGG19 [49] with batch normalization [30] as the backbone and train it using the SGD optimizer. We train the network for 100 epochs. The initial learning rate is set as 0.01 and reduced by a factor of 5 after 50 and 75 epochs." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 253, + 545, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 253, + 545, + 300 + ], + "spans": [ + { + "bbox": [ + 304, + 253, + 545, + 300 + ], + "type": "text", + "content": "In addition, we set the entropic regularization coefficient " + }, + { + "bbox": [ + 304, + 253, + 545, + 300 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 304, + 253, + 545, + 300 + ], + "type": "text", + "content": " and sparsity regularization coefficient " + }, + { + "bbox": [ + 304, + 253, + 545, + 300 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 253, + 545, + 300 + ], + "type": "text", + "content": " as 10 and 1, respectively, which are consistent for all training implementations." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 310, + 429, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 310, + 429, + 323 + ], + "spans": [ + { + "bbox": [ + 306, + 310, + 429, + 323 + ], + "type": "text", + "content": "5.3. Experimental Results" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 329, + 545, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 329, + 545, + 412 + ], + "spans": [ + { + "bbox": [ + 304, + 329, + 545, + 412 + ], + "type": "text", + "content": "In this section, we present experimental results of the OT-Filter on benchmark datasets with both synthetic and real label noises. For CIFAR-10 and CIFAR-100, we consider the noise rate of " + }, + { + "bbox": [ + 304, + 329, + 545, + 412 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 304, + 329, + 545, + 412 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 329, + 545, + 412 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 304, + 329, + 545, + 412 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 329, + 545, + 412 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 304, + 329, + 545, + 412 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 329, + 545, + 412 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 304, + 329, + 545, + 412 + ], + "type": "text", + "content": " for symmetric noise, and noise rate of " + }, + { + "bbox": [ + 304, + 329, + 545, + 412 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 304, + 329, + 545, + 412 + ], + "type": "text", + "content": " for asymmetric noise, respectively. The results regarding precision and recall were drawn from the 100-th epoch of robust training." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 413, + 545, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 413, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 304, + 413, + 545, + 520 + ], + "type": "text", + "content": "Sample Selection Performance. We first present the empirical study on the quality of sample selection. The dataset CIFAR-10 under noise rate of " + }, + { + "bbox": [ + 304, + 413, + 545, + 520 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 304, + 413, + 545, + 520 + ], + "type": "text", + "content": " (symmetric) and noise rate (asymmetric) of " + }, + { + "bbox": [ + 304, + 413, + 545, + 520 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 304, + 413, + 545, + 520 + ], + "type": "text", + "content": " was employed to compare our method with state-of-the-art sample selection method DivMix [38]. Here, we consider two metrics, precision and recall. As shown in Figure 1, with the increase of training epochs, our method outperforms the DivMix [38] in terms of both precision and recall." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 521, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 628 + ], + "type": "text", + "content": "For example, on " + }, + { + "bbox": [ + 304, + 521, + 545, + 628 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 304, + 521, + 545, + 628 + ], + "type": "text", + "content": " symmetric noise, when the training epoch is above 100, the precision of the OT-Filter is steadily above " + }, + { + "bbox": [ + 304, + 521, + 545, + 628 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 304, + 521, + 545, + 628 + ], + "type": "text", + "content": ", while the precision of DivMix is less than " + }, + { + "bbox": [ + 304, + 521, + 545, + 628 + ], + "type": "inline_equation", + "content": "73\\%" + }, + { + "bbox": [ + 304, + 521, + 545, + 628 + ], + "type": "text", + "content": ". In terms of the recall, the OT-Filter is above " + }, + { + "bbox": [ + 304, + 521, + 545, + 628 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 304, + 521, + 545, + 628 + ], + "type": "text", + "content": ", while the DivMix is below " + }, + { + "bbox": [ + 304, + 521, + 545, + 628 + ], + "type": "inline_equation", + "content": "82\\%" + }, + { + "bbox": [ + 304, + 521, + 545, + 628 + ], + "type": "text", + "content": ", when the training epochs are higher than 100. It also shows that our method converges steadily to the optimal solution, which DivMix yields fluctuations in the convergence process. Similar trends can be observed on " + }, + { + "bbox": [ + 304, + 521, + 545, + 628 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 304, + 521, + 545, + 628 + ], + "type": "text", + "content": " asymmetric noise." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "CIFAR10 and CIFAR100 Datasets. Table 1 shows the test accuracy on CIFAR-10 for symmetric and asymmetric noise, respectively. In all testings, the OT-Filter demonstrates good performance in test accuracy. In particular, in the presence of high noise, the OT-Filter significantly outperforms other competitors. For " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": " noise rate, the test accuracy of our OT-Filter is " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "90.5\\%" + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": ", which is about " + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "inline_equation", + "content": "14\\%" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16169" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 79, + 170, + 163 + ], + "blocks": [ + { + "bbox": [ + 50, + 79, + 170, + 163 + ], + "lines": [ + { + "bbox": [ + 50, + 79, + 170, + 163 + ], + "spans": [ + { + "bbox": [ + 50, + 79, + 170, + 163 + ], + "type": "image", + "image_path": "5d9c3fad1e1a6576ba873a292b6cbb4de291f5ab4c0b3ad4590e1966a76763f0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 183, + 523, + 196 + ], + "lines": [ + { + "bbox": [ + 67, + 183, + 523, + 196 + ], + "spans": [ + { + "bbox": [ + 67, + 183, + 523, + 196 + ], + "type": "text", + "content": "Figure 1. Sample selection performance (Precision & Recall) w.r.t. Training Epochs on CIFAR-10 under synthetic label noise." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 171, + 79, + 290, + 163 + ], + "blocks": [ + { + "bbox": [ + 171, + 79, + 290, + 163 + ], + "lines": [ + { + "bbox": [ + 171, + 79, + 290, + 163 + ], + "spans": [ + { + "bbox": [ + 171, + 79, + 290, + 163 + ], + "type": "image", + "image_path": "747b89a5ea3e898f42eab527e153797e3482b5a7014a3263d8bf02bfe7344e86.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 293, + 79, + 412, + 163 + ], + "blocks": [ + { + "bbox": [ + 293, + 79, + 412, + 163 + ], + "lines": [ + { + "bbox": [ + 293, + 79, + 412, + 163 + ], + "spans": [ + { + "bbox": [ + 293, + 79, + 412, + 163 + ], + "type": "image", + "image_path": "aff6f3d61ce1e94f8fe84f1f0c57d7c1ba55d4122aa154909e24d109f62034d3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 414, + 79, + 533, + 163 + ], + "blocks": [ + { + "bbox": [ + 414, + 79, + 533, + 163 + ], + "lines": [ + { + "bbox": [ + 414, + 79, + 533, + 163 + ], + "spans": [ + { + "bbox": [ + 414, + 79, + 533, + 163 + ], + "type": "image", + "image_path": "7535333730364ceb8b133640270ce0e96f7200df21589f3c04ae1028fd48e47a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 62, + 215, + 273, + 363 + ], + "blocks": [ + { + "bbox": [ + 62, + 215, + 273, + 363 + ], + "lines": [ + { + "bbox": [ + 62, + 215, + 273, + 363 + ], + "spans": [ + { + "bbox": [ + 62, + 215, + 273, + 363 + ], + "type": "table", + "html": "
MethodCIFAR-10Asym
20%50%80%90%40%
CE86.879.462.942.777.3
Co-teaching+ [62]89.585.767.447.971.3
Co-learning [54]92.284.561.2-81.4
TopoFilter [58]90.2-45.7-87.9
CRUST [40]91.186.358.3-88.8
Fine [34]91.087.369.4-89.5
M-correction [3]93.691.875.874.793.3
CTRR [61]93.1-83.781.789.0
DivMix [38]96.194.693.276.093.4
Fine+DivMix [34]96.194.993.590.593.8
OT-Filter96.095.394.090.595.1
", + "image_path": "bd5c2aa2080134747a9bd57817140c99b2c2ed422bc0898b38300bca76316bf9.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 62, + 429, + 273, + 578 + ], + "blocks": [ + { + "bbox": [ + 46, + 371, + 288, + 415 + ], + "lines": [ + { + "bbox": [ + 46, + 371, + 288, + 415 + ], + "spans": [ + { + "bbox": [ + 46, + 371, + 288, + 415 + ], + "type": "text", + "content": "Table 1. Test accuracies(%) obtained from state-of-the-art sample selection methods. The best results are in bold. The data are copied from respective papers. The - denotes the lack of respective data." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 62, + 429, + 273, + 578 + ], + "lines": [ + { + "bbox": [ + 62, + 429, + 273, + 578 + ], + "spans": [ + { + "bbox": [ + 62, + 429, + 273, + 578 + ], + "type": "table", + "html": "
MethodCIFAR-100Asym
20%50%80%90%40%
CE62.046.719.910.144.5
Co-teaching+ [62]65.651.827.913.7-
Co-learning [54]66.654.535.5-47.6
CRUST [40]65.256.4--53.0
M-correction [3]73.966.141.624.347.4
CTRR [61]70.1-43.7-54.5
TopoFilter [58]65.6-20.7--
Fine [34]70.364.225.6-61.7
DivMix [38]77.374.660.231.555.1
Fine+DivMix [34]79.174.661.034.3-
OT-Filter76.773.861.842.876.5
", + "image_path": "e82289c0246a3842f2b79edaf512c2658e5ad7bc613bc39d7b4aeeb9740680a5.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 586, + 287, + 630 + ], + "lines": [ + { + "bbox": [ + 46, + 586, + 287, + 630 + ], + "spans": [ + { + "bbox": [ + 46, + 586, + 287, + 630 + ], + "type": "text", + "content": "Table 2. Test accuracies(%) obtained from state-of-the-art sample selection methods. The best results are in bold. The data are copied from respective papers. The - denotes the lack of respective data." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": "better than DivMix. Although [38] performs slightly better than the OT-Filter at the " + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": " noise rate, our method is also competitive. Similar trends can be observed on CIFAR100 (Table 2), where our method mostly follows the SOTA method and significantly outperforms its competitors when" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 332, + 239, + 520, + 371 + ], + "blocks": [ + { + "bbox": [ + 306, + 217, + 394, + 229 + ], + "lines": [ + { + "bbox": [ + 306, + 217, + 394, + 229 + ], + "spans": [ + { + "bbox": [ + 306, + 217, + 394, + 229 + ], + "type": "text", + "content": "the noise rate is high." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 332, + 239, + 520, + 371 + ], + "lines": [ + { + "bbox": [ + 332, + 239, + 520, + 371 + ], + "spans": [ + { + "bbox": [ + 332, + 239, + 520, + 371 + ], + "type": "table", + "html": "
MethodBackboneTest Accuracy
CEResNet-5069.2
M-correction [1]ResNet-5071.0
Co-teaching [27]ResNet-5071.7
CTRR [61]ResNet-5072.7
Fine [34]ResNet-5072.9
CRUST [40]ResNet-5073.5
TopoFilter [58]ResNet-5074.1
DivMix* [38]ResNet-5074.3
Fine+DivMix [34]ResNet-5074.4
OT-FilterResNet-5074.5
", + "image_path": "7decc5d7358a203eb2da4b4b96a51733ed5cb6cb1a2ba66eb4d03d9ff37a9868.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 426, + 547, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 426, + 547, + 499 + ], + "spans": [ + { + "bbox": [ + 305, + 426, + 547, + 499 + ], + "type": "text", + "content": "Clothing1M Dataset. Table 3 shows the result of test accuracy on Clothing1M by comparing different competitors. The result validates the effectiveness of the OT-Filter on the large-scale real-world dataset. Moreover, It can be observed that the OT-Filter follows the state-of-the-art performance regarding test accuracy." + } + ] + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 344, + 508, + 509, + 610 + ], + "blocks": [ + { + "bbox": [ + 305, + 379, + 546, + 414 + ], + "lines": [ + { + "bbox": [ + 305, + 379, + 546, + 414 + ], + "spans": [ + { + "bbox": [ + 305, + 379, + 546, + 414 + ], + "type": "text", + "content": "Table 3. Test accuracy(%) on Clothing1M. The * denotes we have run the algorithm based on the official implementation. The data are copied from respective papers. The best result is in bold." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 344, + 508, + 509, + 610 + ], + "lines": [ + { + "bbox": [ + 344, + 508, + 509, + 610 + ], + "spans": [ + { + "bbox": [ + 344, + 508, + 509, + 610 + ], + "type": "table", + "html": "
MethodTest Accuracy
CE79.4
Nested [47]81.3
SELFIE [51]81.8
PLC [63]83.4
Co-teaching+Nested [12]84.1
GJS [22]84.2
OT-Filter85.5
", + "image_path": "35a9871031d3b044a6e1af243ed6f33ad4627d766b83967a021b10874128eb2a.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 617, + 545, + 640 + ], + "lines": [ + { + "bbox": [ + 306, + 617, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 306, + 617, + 545, + 640 + ], + "type": "text", + "content": "Table 4. Test accuracy(%) on ANIMAL-10N. The data are copied from respective papers. The best result is in bold." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 653, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 653, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 653, + 546, + 715 + ], + "type": "text", + "content": "ANIMAL-10N Dataset. Table 4 shows the experimental results on ANIMAL-10N, a real-world dataset with moderate label noise. We compared our method with multiple state-of-the-art baseline methods that are not limited to sample selection. For fair comparison, we use the same VGG-" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16170" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 71, + 285, + 122 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 285, + 122 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 285, + 122 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 285, + 122 + ], + "type": "table", + "html": "
DatasetCIFAR-100
Noise Rate80%90%80%90%
MethodBestLastBestLastPrecisionRecallPrecisionRecall
OT-Filter w/o Spar-Reg51.050.723.823.790.447.162.119.7
OT-Filter61.861.542.842.193.761.872.340.0
", + "image_path": "782face7b230ec611903c1836eb79df154e24bfdc0274d3dc37e91f3c6ba9182.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 51, + 176, + 286, + 294 + ], + "blocks": [ + { + "bbox": [ + 51, + 176, + 286, + 294 + ], + "lines": [ + { + "bbox": [ + 51, + 176, + 286, + 294 + ], + "spans": [ + { + "bbox": [ + 51, + 176, + 286, + 294 + ], + "type": "image", + "image_path": "f0e34641445936ad2f39c604e9eb6256f5c2dc7c84bb1bb5cee9e416a4382229.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 304, + 286, + 348 + ], + "lines": [ + { + "bbox": [ + 46, + 304, + 286, + 348 + ], + "spans": [ + { + "bbox": [ + 46, + 304, + 286, + 348 + ], + "type": "text", + "content": "Figure 2. The ablation study for the quality of sample selection w/o sparsity regularization on CIFAR-100 under noise rate of " + }, + { + "bbox": [ + 46, + 304, + 286, + 348 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 46, + 304, + 286, + 348 + ], + "type": "text", + "content": ". The result shows that sparsity regularization brings a relatively uniform sample selection." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 371, + 286, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 371, + 286, + 419 + ], + "spans": [ + { + "bbox": [ + 46, + 371, + 286, + 419 + ], + "type": "text", + "content": "19 network architecture. The results show we outperform all competitors, and have " + }, + { + "bbox": [ + 46, + 371, + 286, + 419 + ], + "type": "inline_equation", + "content": "2.1\\%" + }, + { + "bbox": [ + 46, + 371, + 286, + 419 + ], + "type": "text", + "content": " performance improvement over PLC [63], and " + }, + { + "bbox": [ + 46, + 371, + 286, + 419 + ], + "type": "inline_equation", + "content": "3.7\\%" + }, + { + "bbox": [ + 46, + 371, + 286, + 419 + ], + "type": "text", + "content": " performance improvement over SELFIE [51]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 430, + 110, + 442 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 430, + 110, + 442 + ], + "spans": [ + { + "bbox": [ + 47, + 430, + 110, + 442 + ], + "type": "text", + "content": "5.4. Analysis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 449, + 286, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 286, + 508 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 286, + 508 + ], + "type": "text", + "content": "We conduct a series of ablation studies to understand the effectiveness of the key components of the OT-Filter, including the sparsity regularization of the optimal transport, the EM algorithm, and the entropic regularization coefficient." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 510, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 286, + 713 + ], + "type": "text", + "content": "Effect of Sparsity Regularization. The sparsity regularization discussed in Section 4.3 is one of the key components in our OT-Filter, which is designed to improve the accuracy of the optimal transport by controlling the sparsity of transport. We therefore study the effect of the sparsity regularization over sample selection and performance under high noise on CIFAR-100 in Table 5. It can be observed, with the sparsity regularization, both the quality of sample selection and test accuracy are much better than the counterpart without it. In particular, when the noise rate is " + }, + { + "bbox": [ + 46, + 510, + 286, + 713 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 46, + 510, + 286, + 713 + ], + "type": "text", + "content": ", it shows that the technique of sparsity regularization brings in over " + }, + { + "bbox": [ + 46, + 510, + 286, + 713 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 510, + 286, + 713 + ], + "type": "text", + "content": " improvement in precision, about " + }, + { + "bbox": [ + 46, + 510, + 286, + 713 + ], + "type": "inline_equation", + "content": "21\\%" + }, + { + "bbox": [ + 46, + 510, + 286, + 713 + ], + "type": "text", + "content": " improvement in recall, and " + }, + { + "bbox": [ + 46, + 510, + 286, + 713 + ], + "type": "inline_equation", + "content": "19\\%" + }, + { + "bbox": [ + 46, + 510, + 286, + 713 + ], + "type": "text", + "content": " improvement in performance. Moreover, as shown in Figure 2, with the sparsity regularization, the number of samples selected for each class is more uniform. It is mainly attributed to discrete optimal transport allowing mass split." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 309, + 71, + 543, + 134 + ], + "blocks": [ + { + "bbox": [ + 46, + 131, + 287, + 163 + ], + "lines": [ + { + "bbox": [ + 46, + 131, + 287, + 163 + ], + "spans": [ + { + "bbox": [ + 46, + 131, + 287, + 163 + ], + "type": "text", + "content": "Table 5. Ablation study for sparsity regularization on CIFAR-100 dataset. The left part is the performance comparison and the right part is the quality of sample selection." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 71, + 543, + 134 + ], + "lines": [ + { + "bbox": [ + 309, + 71, + 543, + 134 + ], + "spans": [ + { + "bbox": [ + 309, + 71, + 543, + 134 + ], + "type": "table", + "html": "
DatasetCIFAR-10
Noise Rate20%50%80%90%
MethodBestLastBestLastBestLastBestLast
OT-Filter w/o EM95.995.795.094.893.693.490.089.8
OT-Filter96.095.895.395.193.793.590.590.2
", + "image_path": "cf45aa418ad277fff815903c981a2726dd1413a6371a8944a3dc1c7518db33d5.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 308, + 165, + 544, + 221 + ], + "blocks": [ + { + "bbox": [ + 309, + 143, + 541, + 153 + ], + "lines": [ + { + "bbox": [ + 309, + 143, + 541, + 153 + ], + "spans": [ + { + "bbox": [ + 309, + 143, + 541, + 153 + ], + "type": "text", + "content": "Table 6. Ablation study for EM algorithm on CIFAR-10 dataset." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 165, + 544, + 221 + ], + "lines": [ + { + "bbox": [ + 308, + 165, + 544, + 221 + ], + "spans": [ + { + "bbox": [ + 308, + 165, + 544, + 221 + ], + "type": "table", + "html": "
DatasetCIFAR-10
Noise Rate20%50%80%90%
MethodPrecisionRecallPrecisionRecallPrecisionRecallPrecisionRecall
ε = 199.793.999.093.994.690.688.684.2
ε = 599.793.899.093.894.690.688.684.2
ε = 1099.793.899.093.894.590.688.684.2
", + "image_path": "ebd05a991058b5c6e8e076fac2100472583f2e4affb6800b0512244bd9c76793.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 229, + 545, + 251 + ], + "lines": [ + { + "bbox": [ + 306, + 229, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 306, + 229, + 545, + 251 + ], + "type": "text", + "content": "Table 7. Ablation study for entropic regularization coefficient on CIFAR-10 dataset." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 273, + 545, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 273, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 304, + 273, + 545, + 369 + ], + "type": "text", + "content": "Effect of Expectation Maximization. To filter noisy labels, we align all samples to clean barycenters via optimal transport. Therefore, the quality of clean barycenters affects the quality of sample selection and therefore the performance of the robust training. Table 6 indicates the impact of EM on CIFAR-10 dataset. It shows that the EM helps in improving the performance under different noise rates, and the significance is higher for high noise rates." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 369, + 545, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 369, + 545, + 499 + ], + "spans": [ + { + "bbox": [ + 304, + 369, + 545, + 499 + ], + "type": "text", + "content": "Effect of Entropic Regularization Coefficient " + }, + { + "bbox": [ + 304, + 369, + 545, + 499 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 304, + 369, + 545, + 499 + ], + "type": "text", + "content": ". The motivation for equipping the optimal transport with entropic regularization is to speed up its computation. However, the entropic regularization could potentially degrade the transport sparsity and therefore the quality of sample selection. To alleviate that, we propose sparsity regularization, whose effectiveness is examined in Table 5. Then, a by-product of sparsity regularization is the parameter robustness of the regularization coefficient " + }, + { + "bbox": [ + 304, + 369, + 545, + 499 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 304, + 369, + 545, + 499 + ], + "type": "text", + "content": ". As shown in Table 7, the performance of sample selection stays quite stable when varying " + }, + { + "bbox": [ + 304, + 369, + 545, + 499 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 304, + 369, + 545, + 499 + ], + "type": "text", + "content": " from 1 to 10." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 511, + 378, + 523 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 511, + 378, + 523 + ], + "spans": [ + { + "bbox": [ + 306, + 511, + 378, + 523 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 531, + 545, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 531, + 545, + 674 + ], + "spans": [ + { + "bbox": [ + 304, + 531, + 545, + 674 + ], + "type": "text", + "content": "In this work, we study the problem of sample selection for learning with noisy labels. We propose the OT-Filter, a novel technique for retrieving clean samples, which can also be combined with existing semi-supervised learning techniques for handling noisy labels. Unlike previous works relying on a-priori knowledge or conforming to confirmation bias, the OT-Filter enhances the measurement of data discrepancy, by lifting data from Euclidean space to the probability space, and thus improves the quality of sample selection. Extensive experiments on synthetic and real-world noisy datasets are conducted to evaluate our proposals. The results show its effectiveness and superiority." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 674, + 545, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 674, + 545, + 711 + ], + "spans": [ + { + "bbox": [ + 304, + 674, + 545, + 711 + ], + "type": "text", + "content": "Acknowledgement. This work is supported by NSFC (No.61772492, 62072428) and the CAS Pioneer Hundred Talents Program. Xike Xie is the corresponding author." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16171" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "text", + "content": "[1] Martial Agueh and Guillaume Carlier. Barycenters in the wasserstein space. SIAM Journal on Mathematical Analysis, 43(2):904-924, 2011. 2, 3, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 124, + 287, + 167 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 124, + 287, + 167 + ], + "spans": [ + { + "bbox": [ + 53, + 124, + 287, + 167 + ], + "type": "text", + "content": "[2] Ethan Anderes, Steffen Borgwardt, and Jacob Miller. Discrete Wasserstein barycenters: Optimal transport for discrete data. Mathematical Methods of Operations Research, 84(2):389-409, 2016. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 168, + 287, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 168, + 287, + 213 + ], + "spans": [ + { + "bbox": [ + 53, + 168, + 287, + 213 + ], + "type": "text", + "content": "[3] Eric Arazo, Diego Ortego, Paul Albert, Noel O'Connor, and Kevin McGuinness. Unsupervised label noise modeling and loss correction. In International conference on machine learning, pages 312-321. PMLR, 2019. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 213, + 286, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 213, + 286, + 256 + ], + "spans": [ + { + "bbox": [ + 53, + 213, + 286, + 256 + ], + "type": "text", + "content": "[4] Martin Arjovsky, Soumith Chintala, and Léon Bottou. Wasserstein generative adversarial networks. In International conference on machine learning, pages 214-223. PMLR, 2017. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 258, + 287, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 258, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 53, + 258, + 287, + 312 + ], + "type": "text", + "content": "[5] Yingbin Bai, Erkun Yang, Bo Han, Yanhua Yang, Jiatong Li, Yinian Mao, Gang Niu, and Tongliang Liu. Understanding and improving early stopping for learning with noisy labels. Advances in Neural Information Processing Systems, 34:24392-24403, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 312, + 287, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 312, + 287, + 346 + ], + "spans": [ + { + "bbox": [ + 53, + 312, + 287, + 346 + ], + "type": "text", + "content": "[6] Mokhtar S Bazaraa, Hanif D Sherali, and Chitharanjan M Shetty. Nonlinear programming: theory and algorithms. John Wiley & Sons, 2013. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 346, + 287, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 346, + 287, + 390 + ], + "spans": [ + { + "bbox": [ + 53, + 346, + 287, + 390 + ], + "type": "text", + "content": "[7] Jean-David Benamou, Guillaume Carlier, Marco Cuturi, Luca Nenna, and Gabriel Peyre. Iterative bregman projections for regularized transportation problems. SIAM Journal on Scientific Computing, 37(2):A1111-A1138, 2015. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 391, + 287, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 391, + 287, + 434 + ], + "spans": [ + { + "bbox": [ + 53, + 391, + 287, + 434 + ], + "type": "text", + "content": "[8] David Berthelot, Nicholas Carlini, Ian Goodfellow, Nicolas Papernot, Avital Oliver, and Colin A Raffel. Mixmatch: A holistic approach to semi-supervised learning. Advances in neural information processing systems, 32, 2019. 3, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 435, + 287, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 435, + 287, + 457 + ], + "spans": [ + { + "bbox": [ + 53, + 435, + 287, + 457 + ], + "type": "text", + "content": "[9] Dimitris Bertsimas and John N Tsitsiklis. Introduction to linear optimization, volume 6. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 458, + 286, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 458, + 286, + 479 + ], + "spans": [ + { + "bbox": [ + 48, + 458, + 286, + 479 + ], + "type": "text", + "content": "[10] Stephen Boyd, Stephen P Boyd, and Lieven Vandenberghe. Convex optimization. Cambridge university press, 2004. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 480, + 287, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 480, + 287, + 524 + ], + "spans": [ + { + "bbox": [ + 48, + 480, + 287, + 524 + ], + "type": "text", + "content": "[11] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 525, + 287, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 287, + 579 + ], + "type": "text", + "content": "[12] Yingyi Chen, Xi Shen, Shell Xu Hu, and Johan AK Suykens. Boosting co-teaching with compression regularization for label noise. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2688-2692, 2021. 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 580, + 287, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 580, + 287, + 646 + ], + "spans": [ + { + "bbox": [ + 48, + 580, + 287, + 646 + ], + "type": "text", + "content": "[13] De Cheng, Tongliang Liu, Yixiong Ning, Nannan Wang, Bo Han, Gang Niu, Xinbo Gao, and Masashi Sugiyama. Instance-dependent label-noise learning with manifold-regularized transition matrix estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16630–16639, 2022. 1, 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 647, + 287, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 690 + ], + "type": "text", + "content": "[14] Hao Cheng, Zhaowei Zhu, Xingyu Li, Yifei Gong, Xing Sun, and Yang Liu. Learning with instance-dependent label noise: A sample sieve approach. In International Conference on Learning Representations, 2021. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[15] Nicolas Courty, Rémi Flamary, Amaury Habrard, and Alain Rakotomamonjy. Joint distribution optimal transportation" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "for domain adaptation. Advances in Neural Information Processing Systems, 30, 2017. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 95, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 95, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 308, + 95, + 545, + 140 + ], + "type": "text", + "content": "[16] Nicolas Courty, Rémi Flamary, and Devis Tuia. Domain adaptation with regularized optimal transport. In Joint European Conference on Machine Learning and Knowledge Discovery in Databases, pages 274-289. Springer, 2014. 2, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 140, + 545, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 140, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 308, + 140, + 545, + 172 + ], + "type": "text", + "content": "[17] Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. Advances in neural information processing systems, 26, 2013. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 173, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 173, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 308, + 173, + 545, + 205 + ], + "type": "text", + "content": "[18] Marco Cuturi and Arnaud Doucet. Fast computation of wasserstein barycenters. In International conference on machine learning, pages 685-693. PMLR, 2014. 2, 3, 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 205, + 545, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 205, + 545, + 258 + ], + "spans": [ + { + "bbox": [ + 308, + 205, + 545, + 258 + ], + "type": "text", + "content": "[19] Bharath Bhushan Damodaran, Rémi Flamary, Vivien Seguy, and Nicolas Courty. An entropic optimal transport loss for learning deep neural networks under label noise in remote sensing images. Comput. Vis. Image Underst., 191:102863, 2018. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 260, + 545, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 260, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 308, + 260, + 545, + 304 + ], + "type": "text", + "content": "[20] Arthur P Dempster, Nan M Laird, and Donald B Rubin. Maximum likelihood from incomplete data via the em algorithm. Journal of the Royal Statistical Society: Series B (Methodological), 39(1):1-22, 1977. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 305, + 545, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 305, + 545, + 349 + ], + "spans": [ + { + "bbox": [ + 308, + 305, + 545, + 349 + ], + "type": "text", + "content": "[21] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 350, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 350, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 308, + 350, + 545, + 392 + ], + "type": "text", + "content": "[22] Erik Englesson and Hossein Azizpour. Generalized jensen-shannon divergence loss for learning with noisy labels. Advances in Neural Information Processing Systems, 34:30284-30297, 2021. 2, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 393, + 545, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 393, + 545, + 447 + ], + "spans": [ + { + "bbox": [ + 308, + 393, + 545, + 447 + ], + "type": "text", + "content": "[23] Kilian Fatras, Bharath Bhushan Damodaran, Sylvain Lobry, Remi Flamary, Devis Tuia, and Nicolas Courty. Wasserstein adversarial regularization for learning with label noise. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 449, + 545, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 449, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 308, + 449, + 545, + 480 + ], + "type": "text", + "content": "[24] Alessio Figalli and Federico Glaudo. An Invitation to Optimal Transport, Wasserstein Distances, and Gradient Flows. 2021. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 482, + 545, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 482, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 308, + 482, + 545, + 525 + ], + "type": "text", + "content": "[25] Aude Geneva, Gabriel Peyré, and Marco Cuturi. Learning generative models with sinkhorn divergences. In International Conference on Artificial Intelligence and Statistics, pages 1608-1617. PMLR, 2018. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 525, + 545, + 570 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 525, + 545, + 570 + ], + "spans": [ + { + "bbox": [ + 308, + 525, + 545, + 570 + ], + "type": "text", + "content": "[26] Wenshuo Guo, Nhat Ho, and Michael Jordan. Fast algorithms for computational optimal transport and wasserstein barycenter. In International Conference on Artificial Intelligence and Statistics, pages 2088-2097. PMLR, 2020. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 570, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 570, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 570, + 545, + 624 + ], + "type": "text", + "content": "[27] Bo Han, Quanming Yao, Xingrui Yu, Gang Niu, Miao Xu, Weihua Hu, Ivor Tsang, and Masashi Sugiyama. Coteaching: Robust training of deep neural networks with extremely noisy labels. Advances in neural information processing systems, 31, 2018. 1, 2, 6, 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 625, + 545, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 625, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 308, + 625, + 545, + 669 + ], + "type": "text", + "content": "[28] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1, 6" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 712 + ], + "type": "text", + "content": "[29] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Identity mappings in deep residual networks. In European conference on computer vision, pages 630-645. Springer, 2016. 1, 3, 6" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16172" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[30] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In International conference on machine learning, pages 448-456. pmlr, 2015. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 172 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 172 + ], + "type": "text", + "content": "[31] Ahmet Iscen, Jack Valmadre, Anurag Arnab, and Cordelia Schmid. Learning with neighbor consistency for noisy labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4672-4681, 2022. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 174, + 287, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 174, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 48, + 174, + 287, + 228 + ], + "type": "text", + "content": "[32] Lu Jiang, Zhengyuan Zhou, Thomas Leung, Li-Jia Li, and Li Fei-Fei. Mentornet: Learning data-driven curriculum for very deep neural networks on corrupted labels. In International conference on machine learning, pages 2304–2313. PMLR, 2018. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 230, + 287, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 230, + 287, + 283 + ], + "spans": [ + { + "bbox": [ + 48, + 230, + 287, + 283 + ], + "type": "text", + "content": "[33] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. Advances in neural information processing systems, 33:18661-18673, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 285, + 287, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 285, + 287, + 329 + ], + "spans": [ + { + "bbox": [ + 48, + 285, + 287, + 329 + ], + "type": "text", + "content": "[34] Taehyeon Kim, Jongwoo Ko, JinHwan Choi, Se-Young Yun, et al. Fine samples for learning with noisy labels. Advances in Neural Information Processing Systems, 34:24137–24149, 2021. 1, 2, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 331, + 287, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 331, + 287, + 363 + ], + "spans": [ + { + "bbox": [ + 48, + 331, + 287, + 363 + ], + "type": "text", + "content": "[35] Philip A Knight. The sinkhorn-knopp algorithm: convergence and applications. SIAM Journal on Matrix Analysis and Applications, 30(1):261-275, 2008. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 365, + 287, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 365, + 287, + 387 + ], + "spans": [ + { + "bbox": [ + 48, + 365, + 287, + 387 + ], + "type": "text", + "content": "[36] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 1, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 388, + 287, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 287, + 409 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 287, + 409 + ], + "type": "text", + "content": "[37] Kenneth Lange. MM optimization algorithms. SIAM, 2016. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 411, + 287, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 411, + 287, + 443 + ], + "spans": [ + { + "bbox": [ + 48, + 411, + 287, + 443 + ], + "type": "text", + "content": "[38] Junnan Li, Richard Socher, and Steven CH Hoi. Dividemix: Learning with noisy labels as semi-supervised learning. arXiv preprint arXiv:2002.07394, 2020. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 445, + 287, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 445, + 287, + 488 + ], + "spans": [ + { + "bbox": [ + 48, + 445, + 287, + 488 + ], + "type": "text", + "content": "[39] Sheng Liu, Jonathan Niles-Weed, Narges Razavian, and Carlos Fernandez-Granda. Early-learning regularization prevents memorization of noisy labels. Advances in Neural Information Processing Systems, 33, 2020. 2, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 490, + 287, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 490, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 490, + 287, + 533 + ], + "type": "text", + "content": "[40] Baharan Mirzasoleiman, Kaidi Cao, and Jure Leskovec. Coresets for robust training of deep neural networks against noisy labels. Advances in Neural Information Processing Systems, 33:11465-11477, 2020. 2, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 535, + 287, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 535, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 535, + 287, + 578 + ], + "type": "text", + "content": "[41] Kento Nishi, Yi Ding, Alex Rich, and Tobias Hollerer. Augmentation strategies for learning with noisy labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8022-8031, 2021. 2, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 580, + 287, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 580, + 287, + 634 + ], + "spans": [ + { + "bbox": [ + 48, + 580, + 287, + 634 + ], + "type": "text", + "content": "[42] Diego Ortego, Eric Arazo, Paul Albert, Noel E O'Connor, and Kevin McGuinness. Multi-objective interpolation training for robustness to label noise. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6606-6615, 2021. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 689 + ], + "type": "text", + "content": "[43] Giorgio Patrini, Alessandro Rozza, Aditya Krishna Menon, Richard Nock, and Lizhen Qu. Making deep neural networks robust to label noise: A loss correction approach. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1944-1952, 2017. 1, 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[44] Gabriel Peyré, Marco Cuturi, et al. Computational optimal transport: With applications to data science. Foundations" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "and Trends® in Machine Learning, 11(5-6):355-607, 2019. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 96, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 140 + ], + "type": "text", + "content": "[45] Vikas C Raykar, Shipeng Yu, Linda H Zhao, Gerardo Hermosillo Valadez, Charles Florin, Luca Bogoni, and Linda Moy. Learning from crowds. Journal of machine learning research, 11(4), 2010. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 141, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 141, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 308, + 141, + 545, + 194 + ], + "type": "text", + "content": "[46] Ievgen Redko, Nicolas Courty, Rémi Flamary, and Devis Tuia. Optimal transport for multi-source domain adaptation under target shift. In The 22nd International Conference on Artificial Intelligence and Statistics, pages 849-858. PMLR, 2019. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 197, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 197, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 308, + 197, + 545, + 239 + ], + "type": "text", + "content": "[47] Oren Rippel, Michael Gelbart, and Ryan Adams. Learning ordered representations with nested dropout. In International Conference on Machine Learning, pages 1746-1754. PMLR, 2014. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 241, + 545, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 241, + 545, + 263 + ], + "spans": [ + { + "bbox": [ + 308, + 241, + 545, + 263 + ], + "type": "text", + "content": "[48] Filippo Santambrogio. Optimal transport for applied mathematicians. Birkhäuser, NY, 55(58-63):94, 2015. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 265, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 265, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 308, + 265, + 545, + 297 + ], + "type": "text", + "content": "[49] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 1, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 299, + 545, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 299, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 308, + 299, + 545, + 363 + ], + "type": "text", + "content": "[50] Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. Advances in neural information processing systems, 33:596-608, 2020. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 365, + 545, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 365, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 308, + 365, + 545, + 409 + ], + "type": "text", + "content": "[51] Hwanjun Song, Minseok Kim, and Jae-Gil Lee. Selfie: Refurbishing unclean samples for robust deep learning. In International Conference on Machine Learning, pages 5907-5915. PMLR, 2019. 5, 6, 7, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 411, + 545, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 411, + 545, + 465 + ], + "spans": [ + { + "bbox": [ + 308, + 411, + 545, + 465 + ], + "type": "text", + "content": "[52] Fariborz Taherkhani, Ali Dabouei, Sobhan Soleymani, Jeremy Dawson, and Nasser M Nasrabadi. Transporting labels via hierarchical optimal transport for semi-supervised learning. In European Conference on Computer Vision, pages 509-526. Springer, 2020. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 467, + 545, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 467, + 545, + 531 + ], + "spans": [ + { + "bbox": [ + 308, + 467, + 545, + 531 + ], + "type": "text", + "content": "[53] Fariborz Taherkhani, Ali Dabouei, Sobhan Soleymani, Jeremy Dawson, and Nasser M Nasrabadi. Self-supervised Wasserstein pseudo-labeling for semi-supervised image classification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12267-12277, 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 533, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 533, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 308, + 533, + 545, + 578 + ], + "type": "text", + "content": "[54] Cheng Tan, Jun Xia, Lirong Wu, and Stan Z Li. Co-learning: Learning from noisy labels with self-supervision. In Proceedings of the 29th ACM International Conference on Multimedia, pages 1405-1413, 2021. 2, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 578, + 545, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 578, + 545, + 600 + ], + "spans": [ + { + "bbox": [ + 308, + 578, + 545, + 600 + ], + "type": "text", + "content": "[55] Cédric Villani. Optimal transport: old and new, volume 338. Springer, 2009. 2, 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 601, + 545, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 601, + 545, + 623 + ], + "spans": [ + { + "bbox": [ + 308, + 601, + 545, + 623 + ], + "type": "text", + "content": "[56] Cédric Villani. Topics in optimal transportation, volume 58. American Mathematical Soc., 2021. 2, 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 624, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 624, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 308, + 624, + 545, + 668 + ], + "type": "text", + "content": "[57] Yikai Wang, Xinwei Sun, and Yanwei Fu. Scalable penalized regression for noise detection in learning with noisy labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 346-355, 2022. 1" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "text", + "content": "[58] Pengxiang Wu, Songzhu Zheng, Mayank Goswami, Dimitris Metaxas, and Chao Chen. A topological filter for learning with label noise. Advances in neural information processing systems, 33:21382-21393, 2020. 1, 2, 7" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16173" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 496 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "type": "text", + "content": "[59] Xiaobo Xia, Tongliang Liu, Bo Han, Mingming Gong, Jun Yu, Gang Niu, and Masashi Sugiyama. Sample selection with uncertainty of losses for learning with noisy labels. In International Conference on Learning Representations, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 287, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 183 + ], + "type": "text", + "content": "[60] Tong Xiao, Tian Xia, Yi Yang, Chang Huang, and Xiaogang Wang. Learning from massive noisy labeled data for image classification. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2691-2699, 2015. 1, 5, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 185, + 288, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 185, + 288, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 185, + 288, + 239 + ], + "type": "text", + "content": "[61] Li Yi, Sheng Liu, Qi She, A Ian McLeod, and Boyu Wang. On learning contrastive representations for learning with noisy labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16682-16691, 2022. 2, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 240, + 287, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 240, + 287, + 294 + ], + "spans": [ + { + "bbox": [ + 48, + 240, + 287, + 294 + ], + "type": "text", + "content": "[62] Xingrui Yu, Bo Han, Jiangchao Yao, Gang Niu, Ivor Tsang, and Masashi Sugiyama. How does disagreement help generalization against label corruption? In International Conference on Machine Learning, pages 7164-7173. PMLR, 2019. 1, 2, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 296, + 287, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 296, + 287, + 350 + ], + "spans": [ + { + "bbox": [ + 48, + 296, + 287, + 350 + ], + "type": "text", + "content": "[63] Yikai Zhang, Songzhu Zheng, Pengxiang Wu, Mayank Goswami, and Chao Chen. Learning with feature-dependent label noise: A progressive approach. In Ninth International Conference on Learning Representations, volume 9, 2021. 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 352, + 287, + 395 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 352, + 287, + 395 + ], + "spans": [ + { + "bbox": [ + 48, + 352, + 287, + 395 + ], + "type": "text", + "content": "[64] Zhilu Zhang and Mert Sabuncu. Generalized cross entropy loss for training deep neural networks with noisy labels. Advances in neural information processing systems, 31, 2018. 2, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 396, + 287, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 396, + 287, + 452 + ], + "spans": [ + { + "bbox": [ + 48, + 396, + 287, + 452 + ], + "type": "text", + "content": "[65] Evgenii Zheltonozhskii, Chaim Baskin, Avi Mendelson, Alex M Bronstein, and Or Litany. Contrast to divide: Self-supervised pre-training for learning with noisy labels. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1657-1667, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 453, + 287, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 453, + 287, + 496 + ], + "spans": [ + { + "bbox": [ + 48, + 453, + 287, + 496 + ], + "type": "text", + "content": "[66] Yudian Zheng, Guoliang Li, Yuanbing Li, Caihua Shan, and Reynold Cheng. Truth inference in crowdsourcing: Is the problem solved? Proceedings of the VLDB Endowment, 10(5):541-552, 2017. 1" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16174" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/OTAvatar_ One-Shot Talking Face Avatar With Controllable Tri-Plane Rendering/aea7643c-93a9-4960-bd4a-4385733b4c38_content_list.json b/2023/OTAvatar_ One-Shot Talking Face Avatar With Controllable Tri-Plane Rendering/aea7643c-93a9-4960-bd4a-4385733b4c38_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6a5b8755943c5ca970e9641dd682b330e241e8b4 --- /dev/null +++ b/2023/OTAvatar_ One-Shot Talking Face Avatar With Controllable Tri-Plane Rendering/aea7643c-93a9-4960-bd4a-4385733b4c38_content_list.json @@ -0,0 +1,1597 @@ +[ + { + "type": "text", + "text": "OTAvatar: One-shot Talking Face Avatar with Controllable Tri-plane Rendering", + "text_level": 1, + "bbox": [ + 78, + 130, + 890, + 154 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhiyuan Ma $^{1,2*}$", + "bbox": [ + 138, + 178, + 267, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xiangyu Zhu $^{3*}$", + "bbox": [ + 300, + 179, + 421, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Guojun Qi", + "bbox": [ + 450, + 180, + 545, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhen Lei $^{1,2,3\\dagger}$", + "bbox": [ + 584, + 180, + 696, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Lei Zhang", + "bbox": [ + 728, + 180, + 821, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1The Hong Kong Polytechnic University", + "bbox": [ + 325, + 199, + 645, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Center for Artificial Intelligence and Robotics, HKISI CAS", + "bbox": [ + 246, + 215, + 722, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ State Key Laboratory of Multimodal Artificial Intelligence Systems, CASIA", + "bbox": [ + 181, + 234, + 787, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{4}$ OPPO Research", + "bbox": [ + 416, + 252, + 553, + 267 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "zm2354.ma@connect.polyu.hk, xiangyu.zhu@nlpr.ia.ac.cn", + "bbox": [ + 253, + 272, + 725, + 286 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "guojunq@gmail.com, zlei@nlpr.ia.ac.cn, cslzhang@comp.polyu.edu.hk", + "bbox": [ + 192, + 290, + 767, + 304 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 339, + 313, + 354 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Controllability, generalizability and efficiency are the major objectives of constructing face avatars represented by neural implicit field. However, existing methods have not managed to accommodate the three requirements simultaneously. They either focus on static portraits, restricting the representation ability to a specific subject, or suffer from substantial computational cost, limiting their flexibility. In this paper, we propose One-shot Talking face Avatar (OTAvatar), which constructs face avatars by a generalized controllable tri-plane rendering solution so that each personalized avatar can be constructed from only one portrait as the reference. Specifically, OTAvatar first inverts a portrait image to a motion-free identity code. Second, the identity code and a motion code are utilized to modulate an efficient CNN to generate a tri-plane formulated volume, which encodes the subject in the desired motion. Finally, volume rendering is employed to generate an image in any view. The core of our solution is a novel decoupling-by-inverting strategy that disentangles identity and motion in the latent code via optimization-based inversion. Benefiting from the efficient tri-plane representation, we achieve controllable rendering of generalized face avatar at 35 FPS on A100. Experiments show promising performance of cross-identity reenactment on subjects out of the training set and better 3D consistency. The code is available at https://github.com/theEricMa/OTAvatar.", + "bbox": [ + 75, + 371, + 473, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 792, + 209, + 808 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Neural rendering has achieved remarkable progress and promising results in 3D reconstruction. Thanks to the differentiability in neuron computation, the neural rendering", + "bbox": [ + 76, + 818, + 468, + 864 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/374d618906637c283a736c98874ae17ab80e20c09c7399c9b680ad80e2949c8a.jpg", + "image_caption": [ + "Figure 1. OTAvatar animation results. The source subjects in HDTF [43] dataset are animated by OTAvatar using a single portrait as the reference. We use the pose and expression coefficients of 3DMM to represent motion and drive the avatar. Note that these subjects are not included in the training data of OTAvatar." + ], + "image_footnote": [], + "bbox": [ + 509, + 343, + 890, + 597 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "methods bypass the expensive cost of high-fidelity digital 3D modeling, thus attracting the attention of many researchers from academia and industry. In this work, we aim to generate a talking face avatar via neural rendering techniques, which is controllable by driving signals like video and audio segments. Such animation is expected to inherit the identity from reference portraits, while the expression and pose can keep in sync with the driving signals.", + "bbox": [ + 496, + 714, + 893, + 835 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The early works about talking face focus on expression animation consistent with driving signals in constrained frontal face images [5, 7, 27]. It is then extended to in-the-wild scenes with pose variations [18, 35, 46, 47]. Many", + "bbox": [ + 498, + 839, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 94, + 875, + 205, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 96, + 887, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "16901", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "previous works are 2D methods, where the image warping adopted to stimulate the motion in 3D space is learned in 2D space [29, 31, 32, 36, 40]. These methods tend to collapse under large pose variation. In contrast, there are 3D methods that can address the pose problem, since the head pose can be treated as a novel view. Both explicit [11, 20] and implicit 3D representations [23] are introduced to face rendering [15, 17, 21, 22, 25, 28, 45]. However, these methods either overfit to a single identity or fail to produce high-quality animation for different identities.", + "bbox": [ + 75, + 90, + 472, + 241 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we propose a one-shot talking face avatar (OTAvatar), which can generate mimic expressions with good 3D consistency and be generalized to different identities with only one portrait reference. Some avatar animation examples are shown in Fig. 1. Given a single reference image, OTAvatar can drive the subject with motion signals to generate corresponding face images. We realize it under the framework of volume rendering. Current methods usually render static subjects [22, 28]. Although there are works [12, 15, 25, 26, 45] proposed to implement dynamic rendering for face avatars, they need one model per subject. Therefore, the generalization is poor. HeadNeRF [17] is a similar work to us. However, its reconstruction results are unnatural in the talking face case.", + "bbox": [ + 75, + 242, + 472, + 450 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our method is built on a 3D generative model pre-trained on a large-scale face database to guarantee identity generalization ability. Besides, we employ a motion controller to decouple the motion and identity in latent space when performing the optimization-based GAN inversion, to make the motion controllable and transferable to different identities. The network architecture is compact to ensure inference efficiency. In the utilization of our OTAvatar, given a single reference image, we fix the motion code predicted by the controller and only optimize the identity code so that a new avatar of the reference image can be constructed. Such a disentanglement enables the rendering of any desired motion by simply alternating the motion-related latent code to be that of the specific motion representation.", + "bbox": [ + 75, + 453, + 472, + 664 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The major contributions of this work can be summarized as follows.", + "bbox": [ + 76, + 675, + 468, + 705 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We make the first attempt for one-shot 3D face reconstruction and motion-controllable rendering by taming a pre-trained 3D generative model for motion control.", + "- We propose to decouple motion-related and motion-free latent code in inversion optimization by prompting the motion fraction of latent code ahead of the optimization using a decoupling-by-inverting strategy.", + "- Our method can photo-realistically render any identity with the desired expression and pose at 35FPS. The experiment shows promising results of natural motion and 3D consistency on both 2D and 3D datasets." + ], + "bbox": [ + 94, + 710, + 467, + 891 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 500, + 89, + 648, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Talking Face", + "text_level": 1, + "bbox": [ + 500, + 114, + 635, + 131 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Regarding the driving signal, the talking face methods can be roughly divided into three categories: audio-driven, image-driven and coefficients-driven. Our approach is most related to coefficients-driven methods, which use either facial landmarks or 3DMM coefficients to represent motion. Facial landmarks involve identity and expression information and is challenging to transfer motion across different subjects. 3DMM coefficients disentangle the identity, expression and pose, and thus they are good driven signals to drive different faces.", + "bbox": [ + 496, + 138, + 890, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Thies et al. [34] used 3D rendering to maintain the shape and illumination attributes when transferring expression and refining mouth details with a mouth retrieval algorithm. Geng et al. [14] used 3DMM to render a given subject with different expressions, followed by a neural network to refine the detail and harmonize with the background. These methods cannot change the face pose. Ren et al. [29] mapped 3DMM expression and pose to a high-dimensional motion representation, then predicted the dense image flow to reenact faces via image warping. Doukas et al. [10] used additional 3DMM mesh fitting results to assist the dense image flow prediction. Yin et al. [40] integrated pre-trained StyleGAN [6] to generate high-resolution talking face prediction by warping low-resolution feature maps. These warping-based methods follow the two-stage workflow by first warping images and then refining facial detail, demonstrating superior performance for same identity reenactment. But in cross-identity reenactment, especially when there is a large pose variation against the source portrait, the animation tends to collapse, since image warping is merely trained to stimulate 3D motion in the 2D space. Our work animates face avatars using 3D rendering and explicitly ensures 3D consistency to handle large pose variations.", + "bbox": [ + 496, + 290, + 892, + 638 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Volume Rendering", + "text_level": 1, + "bbox": [ + 500, + 650, + 681, + 666 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Volume rendering is a 3D rendering strategy and has demonstrated its success in novel view synthesis. Existing methods either use volume [11, 21] to explicitly represent 3D space or implicitly store the 3D scene in the MLP network [23]. These methods are introduced to animate 3D-consistent portraits. Yu et al. [41] and Raj et al. [28] retrieved the spatial information from sparse support views to render portraits. On the other side, The 3D face generative model can reconstruct a given identity using GAN Inversion. These models vary from pure MLP architecture [4], to low-resolution neural rendering followed by an up-sampling strategy [24, 30], and to using more efficient 3D presentation [3]. All these methods only support rendering static portraits. To animate the avatar with controllable motion, AD-NeRF [15] and NeRFace [13] introduce", + "bbox": [ + 496, + 674, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "16902", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/bfcd6020e08ff84474dce29dfa2a4659e7f33557fd48d2e046a45ebe554931d4.jpg", + "image_caption": [ + "Figure 2. Overview of OTAvatar. OTAvatar contains a 3D face animatior network and a latent code decoupling strategy, namely decoupling-by-inverting. The 3D face animatior is composed of a pre-trained 3D face generator [3] and a motion controller module. The decoupling-by-inverting algorithm is an optimization-based image inversion that can decouple the latent code $\\mathbf{w}$ into identity code $\\mathbf{w}_{id}$ and motion code $\\mathbf{w}_x$ . When the model is well-trained, the motion-free identity code can be inferred from a single reference image, and an avatar of an unseen subject can be constructed. The identity code can be integrated with any other motion code predicted by the controller to animate the identity with desired motion." + ], + "image_footnote": [], + "bbox": [ + 78, + 88, + 890, + 395 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "either the audio feature or 3DMM expression coefficients to the NeRF model to render dynamic heads. Nerfies [25] employs a deformation field conditioned on spatial points to represent the face motion of different frames. HyperNeRF [26] adds an ambient slicing network to tackle the topological drawback of Nerfies. Though these methods have achieved motion control and view consistency for face avatars, they all overfit each NeRF model to one subject, thus lack generalizability. HeadNeRF [17] is trained on both multi-view and frontal face datasets and can be generalized to multiple identities and support motion control using 3DMM coefficients, but its rendering quality is not satisfactory compared with the given portrait, and the motion control is unnatural with random jitters. In contrast, our method can reconstruct photo-realistic face avatars in one shot and animate natural talking motion.", + "bbox": [ + 75, + 518, + 472, + 761 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Generative Prior", + "text_level": 1, + "bbox": [ + 76, + 770, + 243, + 786 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Many methods rely on pre-trained generative models. Employing a generative model as the prior has the following advantages. First, it provides rich and diverse facial priors, such as texture, color, shape and pose, which can help to restore realistic and faithful facial details. For instance, Yang et al. [39] leveraged a pre-trained GAN model as a prior decoder for blind face restoration. Second, it also enables one-", + "bbox": [ + 75, + 794, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "shot image manipulation both inside and outside of the domain. Well-designed schemes for disentanglement are necessary to provide acceptable performance on fine-grained out-of-domain image manipulation. Zhang et al. [44] suggested separating the attributes into global style—like texture and color—and structure—like shape and pose, and performing domain-adaptive generation by transferring the decoupled style attributes. The image manipulation also includes synthesizing human talking videos, Ivan et al. [33] modeled the temporal dynamics of video frames through the latent code disentanglement. Unlike other talking head generation methods [31, 40], it only synthesizes center-aligned faces. The recently presented 3D GANs have been popularized for their ability to generate 3D-consistent photorealistic human faces with explicit pose control. Katja et.al [30] introduced 3D scene volume rendering without resorting to computationally demanding voxel-based representation, and the model is trained from unposed 2D photos in an adversarial manner. Chan et al. [3] proposed to generate high-quality geometry and multi-view consistent images from 2D photos through a hybrid explicit-implicit network architecture based on StyleGAN2 [19]. Despite the fact that 3D GANs' explicit pose control and 3D-consistent generation are appealing for talking face synthesis, no previous work has studied their application to talking face avatars.", + "bbox": [ + 496, + 518, + 893, + 897 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "16903", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 89, + 169, + 104 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Talking face avatar aims to synthesize face images with controllable expression and pose. In this paper, we investigate the possibility of endowing volume rendering with the ability to 1) build a faithful identity representation in one shot, 2) enable natural motion control, and 3) achieve real-time inference speed. Our framework takes an identity code and a motion signal as input and generates a tri-plane-based [3] volume through a CNN architecture. By performing volume rendering on the tri-plane representation, where each point is projected on the three feature planes and the sampled features are fused to occupancy and color, a face image at a given camera view can be generated as:", + "bbox": [ + 75, + 114, + 472, + 297 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {I} (\\mathbf {p}) = \\mathcal {R} (G (\\mathbf {w} _ {i d}, \\mathbf {x}; \\Theta), \\mathbf {p}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 314, + 468, + 330 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{w}_{id}$ is the identity code, $\\mathbf{x}$ is the motion signal, $G(\\cdot ,\\cdot ;\\Theta)$ is a face volume generation network with $\\Theta$ as its network weights to reconstruct an encoded portrait in three orthogonal feature planes, $\\mathcal{R}(\\mathbf{V},\\mathbf{p})$ is the volume rendering with $\\mathbf{V}$ as the volume and $\\mathbf{p}$ as the camera view, and $\\mathbf{I}(\\mathbf{p})\\in \\mathbb{R}^{3\\times H\\times W}$ is the rendered result. In our implementation, the identity code $\\mathbf{w}_{id}$ can be extracted by optimizing on a single reference image using our proposed decoupling-by-inverting strategy (Sec. 3.3), achieving one-shot avatar reconstruction. The motion signal $\\mathbf{x}$ is 3DMM coefficients for flexible control. We call $G(\\cdot ,\\cdot ;\\Theta)$ as 3D Animator because of its ability to animate an avatar with desired motion. Compared with 2D methods, we construct a neural 3D space to promise multi-view consistency. Compared with NeRF, we build a 3D space in the propagation of a CNN, and directly sample features on the tri-plane representation, saving the MLP computation on each sampled point.", + "bbox": [ + 75, + 339, + 472, + 597 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. 3D Animator Network Structure", + "text_level": 1, + "bbox": [ + 76, + 607, + 366, + 622 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Tri-plane volume representation. The output of the 3D anemotor network $G(\\cdot, \\cdot; \\Theta)$ in Eqn. 1 is a tri-plane volume representation, which is composed of three feature planes:", + "bbox": [ + 75, + 631, + 468, + 676 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {V} _ {t r i} = (\\mathbf {F} _ {x y}, \\mathbf {F} _ {x z}, \\mathbf {F} _ {y z}) = G (\\mathbf {w} _ {i d}, \\mathbf {x}; \\boldsymbol {\\Theta}), \\qquad (2)\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 689, + 468, + 708 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{F}_{xy}$ , $\\mathbf{F}_{xz}$ , and $\\mathbf{F}_{yz}$ are three axis-aligned orthogonal feature maps in the 3D space, implicitly forming the volume $\\mathbf{V}_{tri}$ . When performing volume rendering, for each queried point $(x,y,z)$ , we project it onto each of three feature maps and retrieve the corresponding features $(\\mathbf{F}_{xy}(x,y),\\mathbf{F}_{xz}(x,z),\\mathbf{F}_{yz}(y,z))$ . The summation of features is sent to a lightweight MLP to decode color and opacity. Compared to fully implicit MLP architectures [4,15,45] or volume representation [21], we can efficiently regress the three 2D feature maps through a CNN architecture. In our implementation, $G$ is a deconvolutional network [6] that outputs three $256\\times 256\\times 32$ feature maps.", + "bbox": [ + 75, + 719, + 472, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Animator structure. Towards animating the face avatar of a given portrait, an intuitive solution is representing $\\mathbf{w}_{id}$ with a pre-defined feature and training $G(\\cdot)$ from scratch. However, we find that the model does not work well due to the sub-optimal $\\mathbf{w}_{id}$ . For example, when employing 3DMM shape coefficients [9] and face recognition features [8], the identity information is not preserved in the rendering results because $\\mathbf{w}_{id}$ does not encode face appearances well. These models are trained on cropped facial images which contain no torso and hairstyle. Therefore, $\\mathbf{w}_{id}$ losses such information that is necessary when animating portraits. The network even collapses if we make $\\mathbf{w}_{id}$ learnable.", + "bbox": [ + 496, + 90, + 890, + 271 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this work, we employ a two-phase strategy to achieve one-shot avatar reconstruction: 1) building a 3D face generator, and 2) making the generator controllable. We build our 3D face generator on a pre-trained EG3D [3] network, which incorporates tri-plane representation for efficient 3D face generation:", + "bbox": [ + 496, + 272, + 890, + 362 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {V} _ {t r i} = G _ {e g} (\\mathbf {w}; \\Theta_ {e g}), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 372, + 890, + 388 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{w}$ is an uninterpretable latent code drawn from the latent space of the generator $G_{eg}$ . The latent space is determined by the generator weight $\\Theta_{eg}$ . To control the latent code $\\mathbf{w}$ by the given identity represented by $\\mathbf{w}_{id}$ and motion signal $\\mathbf{x}$ , we propose a motion controller module $C$ parameterized by $\\Theta_c$ , which maps motion signal $\\mathbf{x}$ to the motion code $\\mathbf{w}_x$ , such that:", + "bbox": [ + 496, + 397, + 890, + 503 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {w} = \\mathbf {w} _ {i d} + \\mathbf {w} _ {x} = \\mathbf {w} _ {i d} + C (\\mathbf {x}; \\Theta_ {c}). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 575, + 513, + 890, + 529 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Injecting Eqn. 4 to Eqn. 3, we have the structure of the 3D animatior:", + "bbox": [ + 496, + 537, + 890, + 566 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{c} G (\\mathbf {w} _ {i d}, \\mathbf {x}; \\Theta) = G _ {e g} (\\mathbf {w} _ {i d} + C (\\mathbf {x}; \\Theta_ {c}); \\Theta_ {e g}), \\\\ \\Theta = \\Theta_ {e g} \\cup \\Theta_ {c}. \\end{array} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 545, + 574, + 890, + 611 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In our implementation, the motion signal $\\mathbf{x}$ is constructed by 3DMM pose and expression coefficients [9], which are more flexible compared with image-based driving signals [31, 32, 36], especially in cross-identity driving.", + "bbox": [ + 496, + 619, + 890, + 679 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Controller structure. We use the concatenation of 3DMM pose and expression coefficients to represent motion $\\mathbf{x}$ . Practically, we use the adjacent coefficients of pose and expression within a particular radius to represent the motion signal of the current frame, firstly the weighted summation of the coefficients across the temporal dimension is conducted to avoid noises, which is achieved by three 1D convolution layers [29, 40]; secondly, a five-layer MLP is employed to transform the weighted summation into a motion feature; finally, a codebook with learnable orthogonal bases [36] is built, on which each motion feature is projected to get the final motion code $\\mathbf{w}_x$ . The aforementioned operation is summarized as follows:", + "bbox": [ + 496, + 679, + 890, + 875 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {w} _ {x} = C (\\mathbf {x}; \\Theta_ {c}) = \\mathbf {D} * F _ {M} \\left(F _ {T} (\\mathbf {x}; \\Theta_ {T}); \\Theta_ {M}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 883, + 890, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "16904", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $F_{T}$ parameterized by $\\Theta_T$ is the temporal smoothing network of 1D convolution layers, $F_{M}$ parameterized by $\\Theta_M$ is the five-layer MLP which projects motion to the magnitudes of bases contained in the codebook $\\mathbf{D}$ . It can be seen that the magnitude projection parameters, temporal smoothing weights, and the codebook are the parameters of the controller $C$ to be trained, that is:", + "bbox": [ + 75, + 90, + 472, + 196 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Theta_ {c} = \\mathbf {D} \\cup \\Theta_ {T} \\cup \\Theta_ {M}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 196, + 207, + 468, + 224 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. Controller Training", + "text_level": 1, + "bbox": [ + 76, + 232, + 266, + 248 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given a couple of source frame $\\mathbf{I}_s$ and driving frame $\\mathbf{I}_d$ , along with their motion signals $\\mathbf{x}_s, \\mathbf{x}_d$ and camera views $\\mathbf{p}_s, \\mathbf{p}_d$ , we perform the dual-objective optimization:", + "bbox": [ + 76, + 255, + 468, + 301 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf{w}_{id}^{*} = \\operatorname *{arg min}_{\\mathbf{w}_{id}}(\\mathcal{L}_{s} + \\mathcal{L}_{d}),\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 310, + 359, + 335 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s} = \\mathcal {L} \\left(I _ {s}, \\hat {I} _ {s}\\right) = \\mathcal {L} \\left(I _ {s}, \\mathcal {R} \\left(G \\left(\\mathbf {w} _ {i d}, \\mathbf {x} _ {s}; \\Theta\\right), \\mathbf {p} _ {s}\\right), \\right. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 338, + 468, + 356 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {d} = \\mathcal {L} (I _ {d}, \\hat {I} _ {d}) = \\mathcal {L} (I _ {d}, \\mathcal {R} (G (\\mathbf {w} _ {i d}, \\mathbf {x} _ {d}; \\Theta), \\mathbf {p} _ {d}).\n$$\n", + "text_format": "latex", + "bbox": [ + 106, + 361, + 421, + 378 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "It is implemented to estimate the identity code $\\mathbf{w}_{id}$ shared between $\\mathbf{I}_s$ and $\\mathbf{I}_d$ . $G$ , $\\mathcal{R}$ , and $\\mathcal{L}$ indicate 3D face generator, volume rendering, and loss function, respectively. The 3D face generator $G$ is parameterized by $\\Theta = \\Theta_{eg} \\cup \\Theta_c$ . $\\Theta_{eg}$ includes the parameters of the pre-trained 3D face generator, and $\\Theta_c$ includes the parameters of the controller which is the focus of our training scheme.", + "bbox": [ + 75, + 387, + 468, + 492 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training scheme. The purpose of controller training is to decouple the identity code and motion code from the latent code of the generator, making identity and motion replaceable in our generalized avatars. To this end, we propose the decoupling-by-inverting training strategy, which is generally an alternation of identity optimization and controller training at different steps. Specifically, we first freeze $\\Theta_c$ and perform $N_{id}$ steps of back-propagation on $\\mathbf{w}_{id}$ . Then we freeze $\\mathbf{w}_{id}$ and perform $N_{mo}$ steps of backpropagation on $\\Theta_c$ so that the controller learns to add the motion code on top of the identity code. In our implementation, $N_{id} = 90$ and $N_{mo} = 10$ . After $N = N_{id} + N_{mo}$ steps, we slightly finetune $\\Theta_{eg}$ with a learning rate of $10^{-4}$ to make the rendering fit the driving signal better. A brief pseudo-code of the whole scheme is provided in the supplementary material.", + "bbox": [ + 75, + 492, + 468, + 734 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. One-shot Avatar Construction", + "text_level": 1, + "bbox": [ + 76, + 743, + 344, + 757 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the inference stage, given a single portrait image $\\mathbf{I}_r$ , its 3DMM coefficients $\\mathbf{x}_r$ , and camera view $\\mathbf{p}_r$ as the reference of avatar construction, we can estimate the identity code $\\mathbf{w}_r$ of the reference portrait via:", + "bbox": [ + 75, + 766, + 468, + 827 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {w} _ {r} = \\underset {\\mathbf {w} _ {i d}} {\\arg \\min } \\mathcal {L} (\\mathbf {I} _ {r}, \\mathcal {R} (G (\\mathbf {w} _ {i d}, \\mathbf {x} _ {r}), \\mathbf {p} _ {r}). \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 137, + 837, + 468, + 862 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Afterwards, we can animate the identity with given driving motion $\\mathbf{x}_d$ and camera view $\\mathbf{p}_d$ , either from the same or", + "bbox": [ + 76, + 869, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "different subject, through the following operation:", + "bbox": [ + 500, + 90, + 830, + 106 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {I} \\left(\\mathbf {w} _ {r}, \\mathbf {x} _ {d}, \\mathbf {p} _ {d}\\right) = \\mathcal {R} \\left(G \\left(\\mathbf {w} _ {r}, \\mathbf {x} _ {d}\\right), \\mathbf {p} _ {d}\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 116, + 890, + 133 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{I}(\\mathbf{w}_r,\\mathbf{c}_d,\\mathbf{p}_d)$ has the identity estimated as $\\mathbf{I}_r$ , and is animated with the motion $\\mathbf{c}_d$ and camera view $\\mathbf{p}_d$ . We name Eqn. 9 as the decoupling-by-inverting inference strategy since it can estimate decoupled identity code from latent code in the iterative optimization-based GAN Inversion, and can be used to render different motions.", + "bbox": [ + 496, + 142, + 890, + 233 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Loss Terms", + "text_level": 1, + "bbox": [ + 500, + 242, + 624, + 257 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The loss $\\mathcal{L}(\\mathbf{I},\\hat{\\mathbf{I}})$ in Eqn. 8 and Eqn. 9 measures the error between the animated result $\\hat{\\mathbf{I}}$ and the ground truth $\\mathbf{I}$ . First, a pre-trained VGG-19 network is implemented to calculate the distance of multi-scale activation maps $\\phi_{i}(\\mathbf{I})$ via:", + "bbox": [ + 496, + 265, + 890, + 328 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c} = \\sum_ {i} \\left\\| \\phi_ {i} (\\mathbf {I}) - \\phi_ {i} (\\hat {\\mathbf {I}}) \\right\\| _ {1}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 598, + 335, + 890, + 369 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\phi_{i}$ is the $i$ -th layer of VGG. Second, the distance between the Gram matrices $G_{j}^{\\phi}$ constructed from the $j$ -th activation maps $\\phi_{j}$ [29] of $\\mathbf{I}$ and $\\hat{\\mathbf{I}}$ is measured:", + "bbox": [ + 496, + 378, + 890, + 428 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s} = \\sum_ {j} \\left\\| G _ {j} ^ {\\phi} (\\mathbf {I}) - G _ {j} ^ {\\phi} (\\hat {\\mathbf {I}}) \\right\\| _ {1}. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 436, + 890, + 472 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Third, we utilize the L1 distance performed on the eyes and mouth regions to supervise the expression detail:", + "bbox": [ + 496, + 479, + 890, + 511 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {m} = \\sum_ {n} \\left\\| R _ {n} (\\mathbf {I}) - R _ {n} (\\hat {\\mathbf {I}}) \\right\\| _ {1}, \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 518, + 890, + 554 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $R_{n}$ is either the eyes or mouth region extracted using RoI-align on the bounding boxes calculated using landmarks. Finally, the ID Loss is performed to preserve the identity consistency:", + "bbox": [ + 496, + 561, + 890, + 622 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {i d} = 1 - \\cos (E (\\mathbf {I}), E (\\tilde {\\mathbf {I}})), \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 631, + 890, + 648 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $E$ is the Arcface face recognition model [8].", + "bbox": [ + 496, + 657, + 836, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The animation loss $\\mathcal{L}(\\mathbf{I},\\hat{\\mathbf{I}})$ is a weighted summation of the above loss terms and is implemented in both training and test-time one-shot avatar construction. Additionally, in the training, we maintain the stability of tri-plane representation with monotonic and TV loss [3], and $\\mathcal{L}_1$ loss on $\\Theta_c$ to avoid the latent code being out of the latent distribution of $G_{eg}$ .", + "bbox": [ + 496, + 672, + 890, + 765 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 777, + 633, + 794 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experiment Setting", + "text_level": 1, + "bbox": [ + 500, + 801, + 686, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we first describe the implementation detail of our proposed method, the dataset used, and the baselines of our work. Then we compare our proposed method with previous 2D and 3D methods on motion controllability and multi-view consistency.", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "16905", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation details. We use the off-the-shelf 3D face reconstruction model to extract the expression and pose coefficients of 3DMM [9]. The motion of any timestamp is represented by the window of adjacent 27 frames of expression and pose coefficients. In practice, we find that the extracted head poses contain too much noises to serve as the camera pose. Therefore, we extract the camera pose using the method in [15]. Our model is implemented in Pytorch using four A100 GPUs. The total batch size is 24, with six images per GPU. We use Adam optimizer with a 0.0001 learning rate to train the motion controller $C$ and finetune the generator $G$ . Exponential Moving Average (EMA) is employed to update $\\Theta_{c}$ since it can stabilize the training. The model is trained for 2000 iterations, and each iteration contains $N = N_{id} + N_{mo} = 100$ steps to optimize the identity codes and train the motion controller. For each minibatch of data, a new batch of latent codes are initialized and optimized using the Adam optimizer with a 0.01 learning rate. Both the $64 \\times 64$ resolution volume rendering results and $512 \\times 512$ super-resolution results are implemented to calculate the loss. During training, the identity code $\\mathbf{w}_{id}$ is optimized in $\\mathcal{W}$ space. After being extended to the $\\mathcal{W}^{+}$ space by channel-wisely repeated 14 times, it is combined with the motion code $\\mathbf{w}_x$ and then fed into the face generator $G$ . It takes less than two days to train the network entirely. During inference, we first optimize the identity code in $\\mathcal{W}$ space, then use another $N = 100$ steps to finetune it in $\\mathcal{W}^{+}$ space [1, 2].", + "bbox": [ + 76, + 90, + 472, + 513 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset. Our model is trained on the HDTF [43] dataset, which contains frontal talking faces from 362 videos and over 300 subjects. Following [40], we use the pre-processing step in [31] and resize images to $512 \\times 512$ resolution. The test dataset contains 20 videos of subjects out of the training set. To evaluate the novel subject generalizability and controllability in the 3D-consistent fashion, we sample subjects from the Multiface [38] dataset, which is particularly proposed for 3D-consistent face avatars.", + "bbox": [ + 76, + 516, + 468, + 652 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baseline methods. We mainly compare our method with previous works that incorporate 3DMM coefficients to reenact face avatars in either 2D/3D fashion and are generalizable to unseen identities. For 2D methods, PIRenderer [29] and StyleHEAT [40] achieve SOTA performance on face reenactment. For 3D methods, HeadNeRF [17] is the most advanced volume rendering method that incorporates controllability and generalizability. Beyond coefficients-based methods, the SOTA image-driven method FOMM [31] is also taken into comparison.", + "bbox": [ + 76, + 656, + 468, + 806 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation metrics. We adopt peak-signal-to-noise ratio (PSNR), structural similarity index measure (SSIM) [37], and learned perceptual image patch similarity (LPIPS) [42] to evaluate the visual quality. To measure the realism of the synthesized results and the identity preservation, we use frechet inception distance (FID) [16] and", + "bbox": [ + 76, + 810, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/18b56b69d9f6f066ca3869b1992f733f12eaf712519100cc1a856d8cbc38bb90.jpg", + "image_caption": [ + "Figure 3. Qualitative result for cross-identity reenactment. Examples are sampled from the HDTF dataset [43]. Both source and driving subjects are not included in the training set." + ], + "image_footnote": [], + "bbox": [ + 501, + 90, + 890, + 361 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "cosine similarity of identity embedding (CSIM) [8] based on ArcFace model, respectively. Besides, the average expression distance (AED), average pose distance (APD), and average keypoint distance (AKD) are employed to evaluate the facial expression and pose.", + "bbox": [ + 498, + 440, + 890, + 516 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Evaluation Result", + "text_level": 1, + "bbox": [ + 500, + 527, + 671, + 542 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate OTAvatar on animating photo-realistic talking head videos and compare them with state-of-the-art models that support identity-generalized animation methods. In the HDTF [43] dataset, we examine the identity-motion disentanglement by transferring the motion of one subject to drive the other subjects, namely cross-identity reenactment. The employed motion is extracted from a video with large motion variation to evaluate the result in extreme conditions. In the Multiface [38] dataset, we evaluate the consistency of the animation in different views, namely multi-view reenactment. Note that none of the data in this dataset has been used in training our method and baselines. For each talking corpus, we choose the first frame of the frontal-view camera recording as the reference and enforce the network to animate the following frames in frontal view and other views.", + "bbox": [ + 496, + 551, + 890, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative comparison. Fig. 3 shows the results of different methods on cross-identity generation. Compared to FOMM [31], PIRenderer [29] and StyleHEAT [40], which use warping fields for reenacting faces, our method can handle extreme pose variation and maintain identity consistency. Compared to the 3D method of HeadNeRF, our model fully reconstructs the identity-specific detail and syn", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "16906", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2d9797dc277adee4d2c846da9fe23815799d108bc8d8dbeb9857ce3cff89e9b3.jpg", + "image_caption": [ + "Figure 4. Qualitative result for multi-view reenactment. Examples are sampled from the multi-view dataset [38]. All methods use the first frame of the frontal-view portrait to extract the identity feature, and take the expressions of sequential frames and poses of different camera views to generate the talking face. Note that this subject is not included in the training set of any methods." + ], + "image_footnote": [], + "bbox": [ + 83, + 92, + 893, + 469 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f88b995063f61e75164c5e51dd6032615075cc4680279fc54582e4ea9b5e74bd.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Multi-View ReenactmentCross-Identity Reenactment
PSNR ↑SSIM ↑CSIM ↑AED ↓APD ↓AKD ↓LIPIPS ↓FID ↓CSIM ↑AED ↓APD ↓AKD ↓FID ↓
FOMM20.750.6390.5052.0040.5455.0520.308101.60.6723.1960.5004.198113.2
PIRenderer20.040.5860.4932.2030.6806.5660.299100.60.6323.0180.4984.977103.7
StyleHEAT20.030.6320.3872.1790.4725.5220.284123.80.6142.8600.4713.592239.1
HeadNeRF17.600.5460.2392.0860.7764.1660.367212.30.2822.8730.5673.465233.0
Ours21.190.6570.5741.8740.4283.7310.288137.30.6942.8500.4054.307101.8
", + "bbox": [ + 76, + 542, + 895, + 654 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. Quantitative comparisons on multi-View reenactment and cross-identity reenactment.", + "bbox": [ + 187, + 662, + 779, + 676 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "thesizes more natural expression. The multi-view consistency results are shown in Fig. 4. First, the 2D warping method suffers facial malformation, which becomes more serious in larger poses. Second, though HeadNeRF renders accurate head poses, the generated results have noticeable deterioration compared to the ground truth. Finally, the face avatar constructed by our method preserves identity details and multi-view consistency.", + "bbox": [ + 75, + 704, + 468, + 824 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative comparison. The quantitative comparison among competing methods is shown in Table 1. We can see that the proposed method achieves the best performance in terms of most of the criteria.", + "bbox": [ + 75, + 825, + 468, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Inference speed. Table 2 lists the inference speed com", + "bbox": [ + 96, + 885, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "parison among state-of-the-art 3D face avatar methods. By utilizing the pre-trained 3D face generator and the compact motion controller (0.8M parameters), our method achieves the highest inference speed, demonstrating its effectiveness.", + "bbox": [ + 496, + 704, + 890, + 765 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 771, + 653, + 787 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Decoupling-by-inverting. The controller is trained by the decoupling-by-inverting strategy described in Sec. 3.2, where the training alternates between $N_{id}$ steps of identity code optimization and $N_{mo}$ steps of controller parameters training. To validate the effectiveness of the alternative manner, we discard it and jointly train the controller parameters and the identity code. The qualitative comparison is", + "bbox": [ + 496, + 794, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "16907", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5f76f54b7f3a5ce6c202ee1f30fa63b1513283ca567d0913b97a55e7c3cac869.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsFrames Per Second ↑
IMAvatar [45]0.03
ADNeRF [15]0.13
HeadNeRF [17]25
Ours35
", + "bbox": [ + 84, + 88, + 460, + 179 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/53172a8c4df956366c8f40b2f19a5b8e53dce57e021ffbeff17931d0bdcb8572.jpg", + "image_caption": [ + "Figure 5. Qualitative comparison of decoupling-by-inverting training and joint training. Joint training cannot preserve identity information in one-shot avatar construction." + ], + "image_footnote": [], + "bbox": [ + 135, + 239, + 410, + 510 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "in Fig. 5. We see that the jointly trained model cannot preserve identity during the animation because the controller manages to overfit the identity information in the training set, and the identity information is not fully encoded in the identity code. The performance degradation caused by joint training is also shown in Table 4. We also analyze the setting of identity optimization step $N_{id}$ and controller training step $N_{mo}$ , and list the results in Table 3. We observe that $N_{id} = 90$ , $N_{mo} = 10$ achieves a balance of expression reconstruction (AED) and identity preserving (CSIM).", + "bbox": [ + 75, + 594, + 467, + 744 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Losses. We also perform an ablation study on the loss functions, including the landmark region losses $\\mathcal{L}_m$ and ID losses $\\mathcal{L}_{id}$ , as shown in Table 4. We can see that the absence of landmark region losses $\\mathcal{L}_m$ causes a deterioration of expression reconstruction, measured by AED, while the absence of ID loss causes a performance drop in both expression and identity consistency.", + "bbox": [ + 75, + 747, + 467, + 853 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3D generator finetuning. When the controller training and identity code optimization are finished, we further slightly finetune the pre-trained EG3D [3] face generator", + "bbox": [ + 76, + 854, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/aaf24cafba30073bc774d69dc20ee838e1cd7ef917b4353ee571f26b03b9a279.jpg", + "table_caption": [ + "Table 2. Quantitative comparison on the inference speed. The comparison is conducted on an A100 GPU." + ], + "table_footnote": [], + "table_body": "
NmoNidAED ↓CSIM ↑
1993.2850.7200
5953.1780.716
10902.8500.694
20803.2310.592
", + "bbox": [ + 527, + 88, + 864, + 179 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e1afd0adc05fec1c4a39978e4242d20d9b763787c645a1c8a4f51366c1ed917d.jpg", + "table_caption": [ + "Table 3. Ablation study on the decoupling-by-inverting hyperparameters. Experiments are conducted under the different settings of $N_{mo}$ and $N_{id}$ with $N = N_{mo} + N_{id} = 100$ ." + ], + "table_footnote": [], + "table_body": "
AED ↓CSIM ↑
ours2.8500.694
joint3.0090.689
w/o Lm3.1010.687
w/o Lid3.1640.592
w/o finetune3.1450.687
", + "bbox": [ + 535, + 247, + 857, + 354 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Ablation study on the joint training, loss terms, and finetuning.", + "bbox": [ + 498, + 366, + 890, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "with a learning rate of 0.0001. In Table 4, we show the results without finetuning. One can see that slightly finetuning the face generator in the final stage improves identity and expression reconstruction performance.", + "bbox": [ + 498, + 422, + 890, + 484 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 500, + 617, + 515 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We proposed a novel framework of one-shot 3D-consistent talking face avatar, namely OTAvatar, using volume rendering. It jointly addressed the three challenges of face avatars, i.e., generalizability, controllability, and efficiency. For identity-generalized rendering, we implemented a pre-trained 3D face generator to reconstruct faces faithfully, given a single portrait reference. For motion control, we proposed a motion controller module, which predicts the motion code conditioned on 3DMM coefficients. For efficiency, benefiting from the compact architecture of both the 3D generator and controller, our model can animate avatars at a high speed. Besides, we proposed a decoupling-by-inverting approach, which is both a training scheme and a test-time disentangle strategy that decouples the latent code into identity and motion codes via GAN inversion, so that one can animate the avatar with different motions using the motion codes predicted by the controller. Comprehensive experiments were conducted to prove the advantage of our proposed framework.", + "bbox": [ + 496, + 525, + 890, + 813 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 829, + 660, + 845 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work is supported in part by the Chinese National Natural Science Foundation Projects #62176256, #62276254, and the InnoHK program.", + "bbox": [ + 500, + 854, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "16908", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan: How to embed images into the stylegan latent space? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4432-4441, 2019.", + "[2] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan++: How to edit the embedded images? In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8296-8305, 2020.", + "[3] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In CVPR, 2022.", + "[4] Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5799-5809, 2021.", + "[5] Lele Chen, Ross K Maddox, Zhiyao Duan, and Chenliang Xu. Hierarchical cross-modal talking face generation with dynamic pixel-wise loss. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7832-7841, 2019.", + "[6] Min Jin Chong, Hsin-Ying Lee, and David Forsyth. Stylegan of all trades: Image manipulation with only pretrained stylegan. arXiv preprint arXiv:2111.01619, 2021.", + "[7] Joon Son Chung and Andrew Zisserman. Lip reading in the wild. In Asian conference on computer vision, pages 87-103. Springer, 2016.", + "[8] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019.", + "[9] Yu Deng, Jiaolong Yang, Sicheng Xu, Dong Chen, Yunde Jia, and Xin Tong. Accurate 3d face reconstruction with weakly-supervised learning: From single image to image set. In IEEE Computer Vision and Pattern Recognition Workshops, 2019.", + "[10] Michail Christos Doukas, Stefanos Zafeiriou, and Viktoriia Sharmanska. Headgan: One-shot neural head synthesis and editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14398-14407, 2021.", + "[11] Randima Fernando et al. GPU gems: programming techniques, tips, and tricks for real-time graphics, volume 590. Addison-Wesley Reading, 2004.", + "[12] Guy Gafni, Justus Thies, Michael Zollhöfer, and Matthias Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8649-8658, June 2021.", + "[13] Guy Gafni, Justus Thies, Michael Zollhofer, and Matthias Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8649-8658, 2021." + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Zhenglin Geng, Chen Cao, and Sergey Tulyakov. 3d guided fine-grained face manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9821-9830, 2019.", + "[15] Yudong Guo, Keyu Chen, Sen Liang, Yong-Jin Liu, Hujun Bao, and Juyong Zhang. Ad-nerf: Audio driven neural radiance fields for talking head synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5784–5794, 2021.", + "[16] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017.", + "[17] Yang Hong, Bo Peng, Haiyao Xiao, Ligang Liu, and Juyong Zhang. Headnerf: A real-time nerf-based parametric head model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20374-20384, 2022.", + "[18] Xinya Ji, Hang Zhou, Kaisiyuan Wang, Wayne Wu, Chen Change Loy, Xun Cao, and Feng Xu. Audio-driven emotional video portraits. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14080-14089, 2021.", + "[19] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020.", + "[20] Marc Levoy. Display of surfaces from volume data. IEEE Computer graphics and Applications, 8(3):29-37, 1988.", + "[21] Stephen Lombardi, Tomas Simon, Jason Saragih, Gabriel Schwartz, Andreas Lehrmann, and Yaser Sheikh. Neural volumes: Learning dynamic renderable volumes from images. ACM Trans. Graph., 38(4):65:1-65:14, July 2019.", + "[22] Marko Mihajlovic, Aayush Bansal, Michael Zollhoefer, Siyu Tang, and Shunsuke Saito. KeypointNeRF: Generalizing image-based volumetric avatars using relative spatial encoding of keypoints. In European conference on computer vision, 2022.", + "[23] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021.", + "[24] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11453-11464, 2021.", + "[25] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865-5874, 2021.", + "[26] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Ricardo Martin-Brualla, and Steven M. Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. ACM Trans. Graph., 40(6), dec 2021." + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "16909", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] KR Prajwal, Rudrabha Mukhopadhyay, Vinay P Namboodiri, and CV Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In Proceedings of the 28th ACM International Conference on Multimedia, pages 484-492, 2020.", + "[28] Amit Raj, Michael Zollhoefer, Tomas Simon, Jason Saragih, Shunsuke Saito, James Hays, and Stephen Lombardi. Pva: Pixel-aligned volumetric avatars. arXiv preprint arXiv:2101.02697, 2021.", + "[29] Yurui Ren, Ge Li, Yuanqi Chen, Thomas H Li, and Shan Liu. Pirenderer: Controllable portrait image generation via semantic neural rendering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13759-13768, 2021.", + "[30] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. Graf: Generative radiance fields for 3d-aware image synthesis. Advances in Neural Information Processing Systems, 33:20154-20166, 2020.", + "[31] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. First order motion model for image animation. Advances in Neural Information Processing Systems, 32, 2019.", + "[32] Aliaksandr Siarohin, Oliver J Woodford, Jian Ren, Menglei Chai, and Sergey Tulyakov. Motion representations for articulated animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13653-13662, 2021.", + "[33] Ivan Skorokhodov, Sergey Tulyakov, and Mohamed Elhoseiny. Stylegan-v: A continuous video generator with the price, image quality and perks of stylegan2. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3626-3636, 2022.", + "[34] Justus Thies, Michael Zollhofer, Marc Stamminger, Christian Theobalt, and Matthias Nießner. Face2face: Real-time face capture and reenactment of rgb videos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2387-2395, 2016.", + "[35] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Guilin Liu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. Video-to-video synthesis. In Advances in Neural Information Processing Systems (NeurIPS), 2018.", + "[36] Yaohui Wang, Di Yang, Francois Bremond, and Antitza Dantcheva. Latent image animator: Learning to animate images via latent space navigation. arXiv preprint arXiv:2203.09043, 2022.", + "[37] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004.", + "[38] Cheng-hsin Wu, Ningyuan Zheng, Scott Ardisson, Rohan Bali, Danielle Belko, Eric Brockmeyer, Lucas Evans, Timothy Godisart, Hyowon Ha, Alexander Hypes, Taylor Koska, Steven Krenn, Stephen Lombardi, Xiaomin Luo, Kevyn McPhail, Laura Millerschoen, Michal Perdoch, Mark Pitts, Alexander Richard, Jason Saragih, Junko Saragih, Takaaki Shiratori, Tomas Simon, Matt Stewart, Autumn Trimble, Xinshuo Weng, David Whitewolf, Chenglei Wu, Shouu-I Yu," + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "and Yaser Sheikh. Multiface: A dataset for neural face rendering. In arXiv, 2022.", + "[39] Tao Yang, Peiran Ren, Xuansong Xie, and Lei Zhang. Gan prior embedded network for blind face restoration in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 672-681, June 2021.", + "[40] Fei Yin, Yong Zhang, Xiaodong Cun, Mingdeng Cao, Yanbo Fan, Xuan Wang, Qingyan Bai, Baoyuan Wu, Jue Wang, and Yujiu Yang. Styleheat: One-shot high-resolution editable talking face generation via pretrained stylegan. arXiv preprint arXiv:2203.04036, 2022.", + "[41] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4578-4587, 2021.", + "[42] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018.", + "[43] Zhimeng Zhang, Lincheng Li, Yu Ding, and Changjie Fan. Flow-guided one-shot talking face generation with a high-resolution audio-visual dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3661–3670, 2021.", + "[44] Zicheng Zhang, Yinglu Liu, Congying Han, Tiande Guo, Ting Yao, and Tao Mei. Generalized one-shot domain adaption of generative adversarial networks. arXiv preprint arXiv:2209.03665, 2022.", + "[45] Yufeng Zheng, Victoria Fernández Abrevaya, Marcel C. Bühler, Xu Chen, Michael J. Black, and Otmar Hilliges. I M Avatar: Implicit morphable head avatars from videos. In Computer Vision and Pattern Recognition (CVPR), 2022.", + "[46] Hang Zhou, Yu Liu, Ziwei Liu, Ping Luo, and Xiaogang Wang. Talking face generation by adversarially disentangled audio-visual representation. In AAAI Conference on Artificial Intelligence (AAAI), 2019.", + "[47] Hang Zhou, Yasheng Sun, Wayne Wu, Chen Change Loy, Xiaogang Wang, and Ziwei Liu. Pose-controllable talking face generation by implicitly modularized audio-visual representation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021." + ], + "bbox": [ + 501, + 92, + 890, + 698 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "16910", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/OTAvatar_ One-Shot Talking Face Avatar With Controllable Tri-Plane Rendering/aea7643c-93a9-4960-bd4a-4385733b4c38_model.json b/2023/OTAvatar_ One-Shot Talking Face Avatar With Controllable Tri-Plane Rendering/aea7643c-93a9-4960-bd4a-4385733b4c38_model.json new file mode 100644 index 0000000000000000000000000000000000000000..caa2d694cd82f4ba21f48206685d5e3a730271d2 --- /dev/null +++ b/2023/OTAvatar_ One-Shot Talking Face Avatar With Controllable Tri-Plane Rendering/aea7643c-93a9-4960-bd4a-4385733b4c38_model.json @@ -0,0 +1,2134 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.079, + 0.131, + 0.892, + 0.155 + ], + "angle": 0, + "content": "OTAvatar: One-shot Talking Face Avatar with Controllable Tri-plane Rendering" + }, + { + "type": "text", + "bbox": [ + 0.14, + 0.179, + 0.269, + 0.199 + ], + "angle": 0, + "content": "Zhiyuan Ma\\(^{1,2*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.301, + 0.18, + 0.422, + 0.199 + ], + "angle": 0, + "content": "Xiangyu Zhu\\(^{3*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.452, + 0.181, + 0.546, + 0.199 + ], + "angle": 0, + "content": "Guojun Qi" + }, + { + "type": "text", + "bbox": [ + 0.586, + 0.181, + 0.697, + 0.199 + ], + "angle": 0, + "content": "Zhen Lei\\(^{1,2,3\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.73, + 0.181, + 0.822, + 0.199 + ], + "angle": 0, + "content": "Lei Zhang" + }, + { + "type": "text", + "bbox": [ + 0.326, + 0.2, + 0.646, + 0.217 + ], + "angle": 0, + "content": "1The Hong Kong Polytechnic University" + }, + { + "type": "text", + "bbox": [ + 0.248, + 0.217, + 0.723, + 0.234 + ], + "angle": 0, + "content": "\\(^{2}\\)Center for Artificial Intelligence and Robotics, HKISI CAS" + }, + { + "type": "text", + "bbox": [ + 0.182, + 0.235, + 0.789, + 0.252 + ], + "angle": 0, + "content": "\\(^{3}\\)State Key Laboratory of Multimodal Artificial Intelligence Systems, CASIA" + }, + { + "type": "text", + "bbox": [ + 0.418, + 0.253, + 0.555, + 0.268 + ], + "angle": 0, + "content": "\\(^{4}\\)OPPO Research" + }, + { + "type": "text", + "bbox": [ + 0.254, + 0.273, + 0.726, + 0.287 + ], + "angle": 0, + "content": "zm2354.ma@connect.polyu.hk, xiangyu.zhu@nlpr.ia.ac.cn" + }, + { + "type": "text", + "bbox": [ + 0.193, + 0.291, + 0.769, + 0.305 + ], + "angle": 0, + "content": "guojunq@gmail.com, zlei@nlpr.ia.ac.cn, cslzhang@comp.polyu.edu.hk" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.34, + 0.314, + 0.356 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.372, + 0.474, + 0.765 + ], + "angle": 0, + "content": "Controllability, generalizability and efficiency are the major objectives of constructing face avatars represented by neural implicit field. However, existing methods have not managed to accommodate the three requirements simultaneously. They either focus on static portraits, restricting the representation ability to a specific subject, or suffer from substantial computational cost, limiting their flexibility. In this paper, we propose One-shot Talking face Avatar (OTAvatar), which constructs face avatars by a generalized controllable tri-plane rendering solution so that each personalized avatar can be constructed from only one portrait as the reference. Specifically, OTAvatar first inverts a portrait image to a motion-free identity code. Second, the identity code and a motion code are utilized to modulate an efficient CNN to generate a tri-plane formulated volume, which encodes the subject in the desired motion. Finally, volume rendering is employed to generate an image in any view. The core of our solution is a novel decoupling-by-inverting strategy that disentangles identity and motion in the latent code via optimization-based inversion. Benefiting from the efficient tri-plane representation, we achieve controllable rendering of generalized face avatar at 35 FPS on A100. Experiments show promising performance of cross-identity reenactment on subjects out of the training set and better 3D consistency. The code is available at https://github.com/theEricMa/OTAvatar." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.794, + 0.21, + 0.809 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.819, + 0.47, + 0.866 + ], + "angle": 0, + "content": "Neural rendering has achieved remarkable progress and promising results in 3D reconstruction. Thanks to the differentiability in neuron computation, the neural rendering" + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.344, + 0.891, + 0.598 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.611, + 0.894, + 0.682 + ], + "angle": 0, + "content": "Figure 1. OTAvatar animation results. The source subjects in HDTF [43] dataset are animated by OTAvatar using a single portrait as the reference. We use the pose and expression coefficients of 3DMM to represent motion and drive the avatar. Note that these subjects are not included in the training data of OTAvatar." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.715, + 0.894, + 0.837 + ], + "angle": 0, + "content": "methods bypass the expensive cost of high-fidelity digital 3D modeling, thus attracting the attention of many researchers from academia and industry. In this work, we aim to generate a talking face avatar via neural rendering techniques, which is controllable by driving signals like video and audio segments. Such animation is expected to inherit the identity from reference portraits, while the expression and pose can keep in sync with the driving signals." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.84, + 0.894, + 0.902 + ], + "angle": 0, + "content": "The early works about talking face focus on expression animation consistent with driving signals in constrained frontal face images [5, 7, 27]. It is then extended to in-the-wild scenes with pose variations [18, 35, 46, 47]. Many" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.875, + 0.207, + 0.888 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.098, + 0.888, + 0.222, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.875, + 0.222, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16901" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.242 + ], + "angle": 0, + "content": "previous works are 2D methods, where the image warping adopted to stimulate the motion in 3D space is learned in 2D space [29, 31, 32, 36, 40]. These methods tend to collapse under large pose variation. In contrast, there are 3D methods that can address the pose problem, since the head pose can be treated as a novel view. Both explicit [11, 20] and implicit 3D representations [23] are introduced to face rendering [15, 17, 21, 22, 25, 28, 45]. However, these methods either overfit to a single identity or fail to produce high-quality animation for different identities." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.243, + 0.473, + 0.452 + ], + "angle": 0, + "content": "In this work, we propose a one-shot talking face avatar (OTAvatar), which can generate mimic expressions with good 3D consistency and be generalized to different identities with only one portrait reference. Some avatar animation examples are shown in Fig. 1. Given a single reference image, OTAvatar can drive the subject with motion signals to generate corresponding face images. We realize it under the framework of volume rendering. Current methods usually render static subjects [22, 28]. Although there are works [12, 15, 25, 26, 45] proposed to implement dynamic rendering for face avatars, they need one model per subject. Therefore, the generalization is poor. HeadNeRF [17] is a similar work to us. However, its reconstruction results are unnatural in the talking face case." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.454, + 0.473, + 0.665 + ], + "angle": 0, + "content": "Our method is built on a 3D generative model pre-trained on a large-scale face database to guarantee identity generalization ability. Besides, we employ a motion controller to decouple the motion and identity in latent space when performing the optimization-based GAN inversion, to make the motion controllable and transferable to different identities. The network architecture is compact to ensure inference efficiency. In the utilization of our OTAvatar, given a single reference image, we fix the motion code predicted by the controller and only optimize the identity code so that a new avatar of the reference image can be constructed. Such a disentanglement enables the rendering of any desired motion by simply alternating the motion-related latent code to be that of the specific motion representation." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.676, + 0.47, + 0.707 + ], + "angle": 0, + "content": "The major contributions of this work can be summarized as follows." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.711, + 0.468, + 0.756 + ], + "angle": 0, + "content": "- We make the first attempt for one-shot 3D face reconstruction and motion-controllable rendering by taming a pre-trained 3D generative model for motion control." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.765, + 0.468, + 0.825 + ], + "angle": 0, + "content": "- We propose to decouple motion-related and motion-free latent code in inversion optimization by prompting the motion fraction of latent code ahead of the optimization using a decoupling-by-inverting strategy." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.833, + 0.468, + 0.892 + ], + "angle": 0, + "content": "- Our method can photo-realistically render any identity with the desired expression and pose at 35FPS. The experiment shows promising results of natural motion and 3D consistency on both 2D and 3D datasets." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.711, + 0.468, + 0.892 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.09, + 0.65, + 0.107 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.115, + 0.637, + 0.132 + ], + "angle": 0, + "content": "2.1. Talking Face" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.14, + 0.892, + 0.289 + ], + "angle": 0, + "content": "Regarding the driving signal, the talking face methods can be roughly divided into three categories: audio-driven, image-driven and coefficients-driven. Our approach is most related to coefficients-driven methods, which use either facial landmarks or 3DMM coefficients to represent motion. Facial landmarks involve identity and expression information and is challenging to transfer motion across different subjects. 3DMM coefficients disentangle the identity, expression and pose, and thus they are good driven signals to drive different faces." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.291, + 0.893, + 0.639 + ], + "angle": 0, + "content": "Thies et al. [34] used 3D rendering to maintain the shape and illumination attributes when transferring expression and refining mouth details with a mouth retrieval algorithm. Geng et al. [14] used 3DMM to render a given subject with different expressions, followed by a neural network to refine the detail and harmonize with the background. These methods cannot change the face pose. Ren et al. [29] mapped 3DMM expression and pose to a high-dimensional motion representation, then predicted the dense image flow to reenact faces via image warping. Doukas et al. [10] used additional 3DMM mesh fitting results to assist the dense image flow prediction. Yin et al. [40] integrated pre-trained StyleGAN [6] to generate high-resolution talking face prediction by warping low-resolution feature maps. These warping-based methods follow the two-stage workflow by first warping images and then refining facial detail, demonstrating superior performance for same identity reenactment. But in cross-identity reenactment, especially when there is a large pose variation against the source portrait, the animation tends to collapse, since image warping is merely trained to stimulate 3D motion in the 2D space. Our work animates face avatars using 3D rendering and explicitly ensures 3D consistency to handle large pose variations." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.651, + 0.682, + 0.667 + ], + "angle": 0, + "content": "2.2. Volume Rendering" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Volume rendering is a 3D rendering strategy and has demonstrated its success in novel view synthesis. Existing methods either use volume [11, 21] to explicitly represent 3D space or implicitly store the 3D scene in the MLP network [23]. These methods are introduced to animate 3D-consistent portraits. Yu et al. [41] and Raj et al. [28] retrieved the spatial information from sparse support views to render portraits. On the other side, The 3D face generative model can reconstruct a given identity using GAN Inversion. These models vary from pure MLP architecture [4], to low-resolution neural rendering followed by an up-sampling strategy [24, 30], and to using more efficient 3D presentation [3]. All these methods only support rendering static portraits. To animate the avatar with controllable motion, AD-NeRF [15] and NeRFace [13] introduce" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16902" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.089, + 0.891, + 0.396 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.408, + 0.896, + 0.493 + ], + "angle": 0, + "content": "Figure 2. Overview of OTAvatar. OTAvatar contains a 3D face animatior network and a latent code decoupling strategy, namely decoupling-by-inverting. The 3D face animatior is composed of a pre-trained 3D face generator [3] and a motion controller module. The decoupling-by-inverting algorithm is an optimization-based image inversion that can decouple the latent code \\(\\mathbf{w}\\) into identity code \\(\\mathbf{w}_{id}\\) and motion code \\(\\mathbf{w}_x\\). When the model is well-trained, the motion-free identity code can be inferred from a single reference image, and an avatar of an unseen subject can be constructed. The identity code can be integrated with any other motion code predicted by the controller to animate the identity with desired motion." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.519, + 0.473, + 0.762 + ], + "angle": 0, + "content": "either the audio feature or 3DMM expression coefficients to the NeRF model to render dynamic heads. Nerfies [25] employs a deformation field conditioned on spatial points to represent the face motion of different frames. HyperNeRF [26] adds an ambient slicing network to tackle the topological drawback of Nerfies. Though these methods have achieved motion control and view consistency for face avatars, they all overfit each NeRF model to one subject, thus lack generalizability. HeadNeRF [17] is trained on both multi-view and frontal face datasets and can be generalized to multiple identities and support motion control using 3DMM coefficients, but its rendering quality is not satisfactory compared with the given portrait, and the motion control is unnatural with random jitters. In contrast, our method can reconstruct photo-realistic face avatars in one shot and animate natural talking motion." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.771, + 0.245, + 0.787 + ], + "angle": 0, + "content": "2.3. Generative Prior" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Many methods rely on pre-trained generative models. Employing a generative model as the prior has the following advantages. First, it provides rich and diverse facial priors, such as texture, color, shape and pose, which can help to restore realistic and faithful facial details. For instance, Yang et al. [39] leveraged a pre-trained GAN model as a prior decoder for blind face restoration. Second, it also enables one-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.519, + 0.895, + 0.898 + ], + "angle": 0, + "content": "shot image manipulation both inside and outside of the domain. Well-designed schemes for disentanglement are necessary to provide acceptable performance on fine-grained out-of-domain image manipulation. Zhang et al. [44] suggested separating the attributes into global style—like texture and color—and structure—like shape and pose, and performing domain-adaptive generation by transferring the decoupled style attributes. The image manipulation also includes synthesizing human talking videos, Ivan et al. [33] modeled the temporal dynamics of video frames through the latent code disentanglement. Unlike other talking head generation methods [31, 40], it only synthesizes center-aligned faces. The recently presented 3D GANs have been popularized for their ability to generate 3D-consistent photorealistic human faces with explicit pose control. Katja et.al [30] introduced 3D scene volume rendering without resorting to computationally demanding voxel-based representation, and the model is trained from unposed 2D photos in an adversarial manner. Chan et al. [3] proposed to generate high-quality geometry and multi-view consistent images from 2D photos through a hybrid explicit-implicit network architecture based on StyleGAN2 [19]. Despite the fact that 3D GANs' explicit pose control and 3D-consistent generation are appealing for talking face synthesis, no previous work has studied their application to talking face avatars." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16903" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.09, + 0.17, + 0.105 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.116, + 0.473, + 0.299 + ], + "angle": 0, + "content": "Talking face avatar aims to synthesize face images with controllable expression and pose. In this paper, we investigate the possibility of endowing volume rendering with the ability to 1) build a faithful identity representation in one shot, 2) enable natural motion control, and 3) achieve real-time inference speed. Our framework takes an identity code and a motion signal as input and generates a tri-plane-based [3] volume through a CNN architecture. By performing volume rendering on the tri-plane representation, where each point is projected on the three feature planes and the sampled features are fused to occupancy and color, a face image at a given camera view can be generated as:" + }, + { + "type": "equation", + "bbox": [ + 0.176, + 0.315, + 0.47, + 0.332 + ], + "angle": 0, + "content": "\\[\n\\mathbf {I} (\\mathbf {p}) = \\mathcal {R} (G (\\mathbf {w} _ {i d}, \\mathbf {x}; \\Theta), \\mathbf {p}), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.34, + 0.473, + 0.598 + ], + "angle": 0, + "content": "where \\(\\mathbf{w}_{id}\\) is the identity code, \\(\\mathbf{x}\\) is the motion signal, \\(G(\\cdot ,\\cdot ;\\Theta)\\) is a face volume generation network with \\(\\Theta\\) as its network weights to reconstruct an encoded portrait in three orthogonal feature planes, \\(\\mathcal{R}(\\mathbf{V},\\mathbf{p})\\) is the volume rendering with \\(\\mathbf{V}\\) as the volume and \\(\\mathbf{p}\\) as the camera view, and \\(\\mathbf{I}(\\mathbf{p})\\in \\mathbb{R}^{3\\times H\\times W}\\) is the rendered result. In our implementation, the identity code \\(\\mathbf{w}_{id}\\) can be extracted by optimizing on a single reference image using our proposed decoupling-by-inverting strategy (Sec. 3.3), achieving one-shot avatar reconstruction. The motion signal \\(\\mathbf{x}\\) is 3DMM coefficients for flexible control. We call \\(G(\\cdot ,\\cdot ;\\Theta)\\) as 3D Animator because of its ability to animate an avatar with desired motion. Compared with 2D methods, we construct a neural 3D space to promise multi-view consistency. Compared with NeRF, we build a 3D space in the propagation of a CNN, and directly sample features on the tri-plane representation, saving the MLP computation on each sampled point." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.608, + 0.367, + 0.623 + ], + "angle": 0, + "content": "3.1. 3D Animator Network Structure" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.632, + 0.47, + 0.678 + ], + "angle": 0, + "content": "Tri-plane volume representation. The output of the 3D anemotor network \\( G(\\cdot, \\cdot; \\Theta) \\) in Eqn. 1 is a tri-plane volume representation, which is composed of three feature planes:" + }, + { + "type": "equation", + "bbox": [ + 0.134, + 0.69, + 0.47, + 0.709 + ], + "angle": 0, + "content": "\\[\n\\mathbf {V} _ {t r i} = (\\mathbf {F} _ {x y}, \\mathbf {F} _ {x z}, \\mathbf {F} _ {y z}) = G (\\mathbf {w} _ {i d}, \\mathbf {x}; \\boldsymbol {\\Theta}), \\qquad (2)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.473, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathbf{F}_{xy}\\), \\(\\mathbf{F}_{xz}\\), and \\(\\mathbf{F}_{yz}\\) are three axis-aligned orthogonal feature maps in the 3D space, implicitly forming the volume \\(\\mathbf{V}_{tri}\\). When performing volume rendering, for each queried point \\((x,y,z)\\), we project it onto each of three feature maps and retrieve the corresponding features \\((\\mathbf{F}_{xy}(x,y),\\mathbf{F}_{xz}(x,z),\\mathbf{F}_{yz}(y,z))\\). The summation of features is sent to a lightweight MLP to decode color and opacity. Compared to fully implicit MLP architectures [4,15,45] or volume representation [21], we can efficiently regress the three 2D feature maps through a CNN architecture. In our implementation, \\(G\\) is a deconvolutional network [6] that outputs three \\(256\\times 256\\times 32\\) feature maps." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.272 + ], + "angle": 0, + "content": "Animator structure. Towards animating the face avatar of a given portrait, an intuitive solution is representing \\(\\mathbf{w}_{id}\\) with a pre-defined feature and training \\(G(\\cdot)\\) from scratch. However, we find that the model does not work well due to the sub-optimal \\(\\mathbf{w}_{id}\\). For example, when employing 3DMM shape coefficients [9] and face recognition features [8], the identity information is not preserved in the rendering results because \\(\\mathbf{w}_{id}\\) does not encode face appearances well. These models are trained on cropped facial images which contain no torso and hairstyle. Therefore, \\(\\mathbf{w}_{id}\\) losses such information that is necessary when animating portraits. The network even collapses if we make \\(\\mathbf{w}_{id}\\) learnable." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.273, + 0.892, + 0.363 + ], + "angle": 0, + "content": "In this work, we employ a two-phase strategy to achieve one-shot avatar reconstruction: 1) building a 3D face generator, and 2) making the generator controllable. We build our 3D face generator on a pre-trained EG3D [3] network, which incorporates tri-plane representation for efficient 3D face generation:" + }, + { + "type": "equation", + "bbox": [ + 0.622, + 0.373, + 0.892, + 0.39 + ], + "angle": 0, + "content": "\\[\n\\mathbf {V} _ {t r i} = G _ {e g} (\\mathbf {w}; \\Theta_ {e g}), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.398, + 0.892, + 0.504 + ], + "angle": 0, + "content": "where \\(\\mathbf{w}\\) is an uninterpretable latent code drawn from the latent space of the generator \\(G_{eg}\\). The latent space is determined by the generator weight \\(\\Theta_{eg}\\). To control the latent code \\(\\mathbf{w}\\) by the given identity represented by \\(\\mathbf{w}_{id}\\) and motion signal \\(\\mathbf{x}\\), we propose a motion controller module \\(C\\) parameterized by \\(\\Theta_c\\), which maps motion signal \\(\\mathbf{x}\\) to the motion code \\(\\mathbf{w}_x\\), such that:" + }, + { + "type": "equation", + "bbox": [ + 0.576, + 0.514, + 0.892, + 0.53 + ], + "angle": 0, + "content": "\\[\n\\mathbf {w} = \\mathbf {w} _ {i d} + \\mathbf {w} _ {x} = \\mathbf {w} _ {i d} + C (\\mathbf {x}; \\Theta_ {c}). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.539, + 0.892, + 0.568 + ], + "angle": 0, + "content": "Injecting Eqn. 4 to Eqn. 3, we have the structure of the 3D animatior:" + }, + { + "type": "equation", + "bbox": [ + 0.547, + 0.575, + 0.892, + 0.612 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{c} G (\\mathbf {w} _ {i d}, \\mathbf {x}; \\Theta) = G _ {e g} (\\mathbf {w} _ {i d} + C (\\mathbf {x}; \\Theta_ {c}); \\Theta_ {e g}), \\\\ \\Theta = \\Theta_ {e g} \\cup \\Theta_ {c}. \\end{array} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.62, + 0.892, + 0.68 + ], + "angle": 0, + "content": "In our implementation, the motion signal \\(\\mathbf{x}\\) is constructed by 3DMM pose and expression coefficients [9], which are more flexible compared with image-based driving signals [31, 32, 36], especially in cross-identity driving." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.68, + 0.892, + 0.876 + ], + "angle": 0, + "content": "Controller structure. We use the concatenation of 3DMM pose and expression coefficients to represent motion \\(\\mathbf{x}\\). Practically, we use the adjacent coefficients of pose and expression within a particular radius to represent the motion signal of the current frame, firstly the weighted summation of the coefficients across the temporal dimension is conducted to avoid noises, which is achieved by three 1D convolution layers [29, 40]; secondly, a five-layer MLP is employed to transform the weighted summation into a motion feature; finally, a codebook with learnable orthogonal bases [36] is built, on which each motion feature is projected to get the final motion code \\(\\mathbf{w}_x\\). The aforementioned operation is summarized as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.885, + 0.892, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\mathbf {w} _ {x} = C (\\mathbf {x}; \\Theta_ {c}) = \\mathbf {D} * F _ {M} \\left(F _ {T} (\\mathbf {x}; \\Theta_ {T}); \\Theta_ {M}\\right), \\tag {6}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16904" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.198 + ], + "angle": 0, + "content": "where \\(F_{T}\\) parameterized by \\(\\Theta_T\\) is the temporal smoothing network of 1D convolution layers, \\(F_{M}\\) parameterized by \\(\\Theta_M\\) is the five-layer MLP which projects motion to the magnitudes of bases contained in the codebook \\(\\mathbf{D}\\). It can be seen that the magnitude projection parameters, temporal smoothing weights, and the codebook are the parameters of the controller \\(C\\) to be trained, that is:" + }, + { + "type": "equation", + "bbox": [ + 0.197, + 0.208, + 0.47, + 0.226 + ], + "angle": 0, + "content": "\\[\n\\Theta_ {c} = \\mathbf {D} \\cup \\Theta_ {T} \\cup \\Theta_ {M}. \\tag {7}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.233, + 0.267, + 0.249 + ], + "angle": 0, + "content": "3.2. Controller Training" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.256, + 0.47, + 0.303 + ], + "angle": 0, + "content": "Given a couple of source frame \\(\\mathbf{I}_s\\) and driving frame \\(\\mathbf{I}_d\\), along with their motion signals \\(\\mathbf{x}_s, \\mathbf{x}_d\\) and camera views \\(\\mathbf{p}_s, \\mathbf{p}_d\\), we perform the dual-objective optimization:" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.311, + 0.36, + 0.336 + ], + "angle": 0, + "content": "\\[\n\\mathbf{w}_{id}^{*} = \\operatorname *{arg min}_{\\mathbf{w}_{id}}(\\mathcal{L}_{s} + \\mathcal{L}_{d}),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.108, + 0.339, + 0.469, + 0.357 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s} = \\mathcal {L} \\left(I _ {s}, \\hat {I} _ {s}\\right) = \\mathcal {L} \\left(I _ {s}, \\mathcal {R} \\left(G \\left(\\mathbf {w} _ {i d}, \\mathbf {x} _ {s}; \\Theta\\right), \\mathbf {p} _ {s}\\right), \\right. \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.107, + 0.362, + 0.423, + 0.38 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {d} = \\mathcal {L} (I _ {d}, \\hat {I} _ {d}) = \\mathcal {L} (I _ {d}, \\mathcal {R} (G (\\mathbf {w} _ {i d}, \\mathbf {x} _ {d}; \\Theta), \\mathbf {p} _ {d}).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.388, + 0.469, + 0.493 + ], + "angle": 0, + "content": "It is implemented to estimate the identity code \\(\\mathbf{w}_{id}\\) shared between \\(\\mathbf{I}_s\\) and \\(\\mathbf{I}_d\\). \\(G\\), \\(\\mathcal{R}\\), and \\(\\mathcal{L}\\) indicate 3D face generator, volume rendering, and loss function, respectively. The 3D face generator \\(G\\) is parameterized by \\(\\Theta = \\Theta_{eg} \\cup \\Theta_c\\). \\(\\Theta_{eg}\\) includes the parameters of the pre-trained 3D face generator, and \\(\\Theta_c\\) includes the parameters of the controller which is the focus of our training scheme." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.493, + 0.47, + 0.735 + ], + "angle": 0, + "content": "Training scheme. The purpose of controller training is to decouple the identity code and motion code from the latent code of the generator, making identity and motion replaceable in our generalized avatars. To this end, we propose the decoupling-by-inverting training strategy, which is generally an alternation of identity optimization and controller training at different steps. Specifically, we first freeze \\(\\Theta_c\\) and perform \\(N_{id}\\) steps of back-propagation on \\(\\mathbf{w}_{id}\\). Then we freeze \\(\\mathbf{w}_{id}\\) and perform \\(N_{mo}\\) steps of backpropagation on \\(\\Theta_c\\) so that the controller learns to add the motion code on top of the identity code. In our implementation, \\(N_{id} = 90\\) and \\(N_{mo} = 10\\). After \\(N = N_{id} + N_{mo}\\) steps, we slightly finetune \\(\\Theta_{eg}\\) with a learning rate of \\(10^{-4}\\) to make the rendering fit the driving signal better. A brief pseudo-code of the whole scheme is provided in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.744, + 0.345, + 0.758 + ], + "angle": 0, + "content": "3.3. One-shot Avatar Construction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.767, + 0.469, + 0.828 + ], + "angle": 0, + "content": "In the inference stage, given a single portrait image \\(\\mathbf{I}_r\\), its 3DMM coefficients \\(\\mathbf{x}_r\\), and camera view \\(\\mathbf{p}_r\\) as the reference of avatar construction, we can estimate the identity code \\(\\mathbf{w}_r\\) of the reference portrait via:" + }, + { + "type": "equation", + "bbox": [ + 0.138, + 0.838, + 0.469, + 0.863 + ], + "angle": 0, + "content": "\\[\n\\mathbf {w} _ {r} = \\underset {\\mathbf {w} _ {i d}} {\\arg \\min } \\mathcal {L} (\\mathbf {I} _ {r}, \\mathcal {R} (G (\\mathbf {w} _ {i d}, \\mathbf {x} _ {r}), \\mathbf {p} _ {r}). \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Afterwards, we can animate the identity with given driving motion \\(\\mathbf{x}_d\\) and camera view \\(\\mathbf{p}_d\\), either from the same or" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.831, + 0.107 + ], + "angle": 0, + "content": "different subject, through the following operation:" + }, + { + "type": "equation", + "bbox": [ + 0.573, + 0.117, + 0.892, + 0.134 + ], + "angle": 0, + "content": "\\[\n\\mathbf {I} \\left(\\mathbf {w} _ {r}, \\mathbf {x} _ {d}, \\mathbf {p} _ {d}\\right) = \\mathcal {R} \\left(G \\left(\\mathbf {w} _ {r}, \\mathbf {x} _ {d}\\right), \\mathbf {p} _ {d}\\right), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.143, + 0.892, + 0.234 + ], + "angle": 0, + "content": "where \\(\\mathbf{I}(\\mathbf{w}_r,\\mathbf{c}_d,\\mathbf{p}_d)\\) has the identity estimated as \\(\\mathbf{I}_r\\), and is animated with the motion \\(\\mathbf{c}_d\\) and camera view \\(\\mathbf{p}_d\\). We name Eqn. 9 as the decoupling-by-inverting inference strategy since it can estimate decoupled identity code from latent code in the iterative optimization-based GAN Inversion, and can be used to render different motions." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.243, + 0.625, + 0.258 + ], + "angle": 0, + "content": "3.4. Loss Terms" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.266, + 0.892, + 0.329 + ], + "angle": 0, + "content": "The loss \\(\\mathcal{L}(\\mathbf{I},\\hat{\\mathbf{I}})\\) in Eqn. 8 and Eqn. 9 measures the error between the animated result \\(\\hat{\\mathbf{I}}\\) and the ground truth \\(\\mathbf{I}\\). First, a pre-trained VGG-19 network is implemented to calculate the distance of multi-scale activation maps \\(\\phi_{i}(\\mathbf{I})\\) via:" + }, + { + "type": "equation", + "bbox": [ + 0.599, + 0.337, + 0.892, + 0.37 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c} = \\sum_ {i} \\left\\| \\phi_ {i} (\\mathbf {I}) - \\phi_ {i} (\\hat {\\mathbf {I}}) \\right\\| _ {1}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.379, + 0.892, + 0.429 + ], + "angle": 0, + "content": "where \\(\\phi_{i}\\) is the \\(i\\)-th layer of VGG. Second, the distance between the Gram matrices \\(G_{j}^{\\phi}\\) constructed from the \\(j\\)-th activation maps \\(\\phi_{j}\\) [29] of \\(\\mathbf{I}\\) and \\(\\hat{\\mathbf{I}}\\) is measured:" + }, + { + "type": "equation", + "bbox": [ + 0.593, + 0.437, + 0.892, + 0.473 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s} = \\sum_ {j} \\left\\| G _ {j} ^ {\\phi} (\\mathbf {I}) - G _ {j} ^ {\\phi} (\\hat {\\mathbf {I}}) \\right\\| _ {1}. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.481, + 0.892, + 0.512 + ], + "angle": 0, + "content": "Third, we utilize the L1 distance performed on the eyes and mouth regions to supervise the expression detail:" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.52, + 0.892, + 0.555 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {m} = \\sum_ {n} \\left\\| R _ {n} (\\mathbf {I}) - R _ {n} (\\hat {\\mathbf {I}}) \\right\\| _ {1}, \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.562, + 0.892, + 0.623 + ], + "angle": 0, + "content": "where \\( R_{n} \\) is either the eyes or mouth region extracted using RoI-align on the bounding boxes calculated using landmarks. Finally, the ID Loss is performed to preserve the identity consistency:" + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.632, + 0.892, + 0.65 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {i d} = 1 - \\cos (E (\\mathbf {I}), E (\\tilde {\\mathbf {I}})), \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.838, + 0.674 + ], + "angle": 0, + "content": "where \\(E\\) is the Arcface face recognition model [8]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.674, + 0.892, + 0.766 + ], + "angle": 0, + "content": "The animation loss \\(\\mathcal{L}(\\mathbf{I},\\hat{\\mathbf{I}})\\) is a weighted summation of the above loss terms and is implemented in both training and test-time one-shot avatar construction. Additionally, in the training, we maintain the stability of tri-plane representation with monotonic and TV loss [3], and \\(\\mathcal{L}_1\\) loss on \\(\\Theta_c\\) to avoid the latent code being out of the latent distribution of \\(G_{eg}\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.778, + 0.634, + 0.795 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.802, + 0.687, + 0.818 + ], + "angle": 0, + "content": "4.1. Experiment Setting" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In this section, we first describe the implementation detail of our proposed method, the dataset used, and the baselines of our work. Then we compare our proposed method with previous 2D and 3D methods on motion controllability and multi-view consistency." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16905" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.514 + ], + "angle": 0, + "content": "Implementation details. We use the off-the-shelf 3D face reconstruction model to extract the expression and pose coefficients of 3DMM [9]. The motion of any timestamp is represented by the window of adjacent 27 frames of expression and pose coefficients. In practice, we find that the extracted head poses contain too much noises to serve as the camera pose. Therefore, we extract the camera pose using the method in [15]. Our model is implemented in Pytorch using four A100 GPUs. The total batch size is 24, with six images per GPU. We use Adam optimizer with a 0.0001 learning rate to train the motion controller \\( C \\) and finetune the generator \\( G \\). Exponential Moving Average (EMA) is employed to update \\( \\Theta_{c} \\) since it can stabilize the training. The model is trained for 2000 iterations, and each iteration contains \\( N = N_{id} + N_{mo} = 100 \\) steps to optimize the identity codes and train the motion controller. For each minibatch of data, a new batch of latent codes are initialized and optimized using the Adam optimizer with a 0.01 learning rate. Both the \\( 64 \\times 64 \\) resolution volume rendering results and \\( 512 \\times 512 \\) super-resolution results are implemented to calculate the loss. During training, the identity code \\( \\mathbf{w}_{id} \\) is optimized in \\( \\mathcal{W} \\) space. After being extended to the \\( \\mathcal{W}^{+} \\)space by channel-wisely repeated 14 times, it is combined with the motion code \\( \\mathbf{w}_x \\) and then fed into the face generator \\( G \\). It takes less than two days to train the network entirely. During inference, we first optimize the identity code in \\( \\mathcal{W} \\) space, then use another \\( N = 100 \\) steps to finetune it in \\( \\mathcal{W}^{+} \\)space [1, 2]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.517, + 0.47, + 0.653 + ], + "angle": 0, + "content": "Dataset. Our model is trained on the HDTF [43] dataset, which contains frontal talking faces from 362 videos and over 300 subjects. Following [40], we use the pre-processing step in [31] and resize images to \\(512 \\times 512\\) resolution. The test dataset contains 20 videos of subjects out of the training set. To evaluate the novel subject generalizability and controllability in the 3D-consistent fashion, we sample subjects from the Multiface [38] dataset, which is particularly proposed for 3D-consistent face avatars." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.657, + 0.47, + 0.807 + ], + "angle": 0, + "content": "Baseline methods. We mainly compare our method with previous works that incorporate 3DMM coefficients to reenact face avatars in either 2D/3D fashion and are generalizable to unseen identities. For 2D methods, PIRenderer [29] and StyleHEAT [40] achieve SOTA performance on face reenactment. For 3D methods, HeadNeRF [17] is the most advanced volume rendering method that incorporates controllability and generalizability. Beyond coefficients-based methods, the SOTA image-driven method FOMM [31] is also taken into comparison." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.811, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Evaluation metrics. We adopt peak-signal-to-noise ratio (PSNR), structural similarity index measure (SSIM) [37], and learned perceptual image patch similarity (LPIPS) [42] to evaluate the visual quality. To measure the realism of the synthesized results and the identity preservation, we use frechet inception distance (FID) [16] and" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.371, + 0.892, + 0.413 + ], + "angle": 0, + "content": "Figure 3. Qualitative result for cross-identity reenactment. Examples are sampled from the HDTF dataset [43]. Both source and driving subjects are not included in the training set." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.441, + 0.892, + 0.517 + ], + "angle": 0, + "content": "cosine similarity of identity embedding (CSIM) [8] based on ArcFace model, respectively. Besides, the average expression distance (AED), average pose distance (APD), and average keypoint distance (AKD) are employed to evaluate the facial expression and pose." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.529, + 0.673, + 0.543 + ], + "angle": 0, + "content": "4.2. Evaluation Result" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.553, + 0.892, + 0.793 + ], + "angle": 0, + "content": "We evaluate OTAvatar on animating photo-realistic talking head videos and compare them with state-of-the-art models that support identity-generalized animation methods. In the HDTF [43] dataset, we examine the identity-motion disentanglement by transferring the motion of one subject to drive the other subjects, namely cross-identity reenactment. The employed motion is extracted from a video with large motion variation to evaluate the result in extreme conditions. In the Multiface [38] dataset, we evaluate the consistency of the animation in different views, namely multi-view reenactment. Note that none of the data in this dataset has been used in training our method and baselines. For each talking corpus, we choose the first frame of the frontal-view camera recording as the reference and enforce the network to animate the following frames in frontal view and other views." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Qualitative comparison. Fig. 3 shows the results of different methods on cross-identity generation. Compared to FOMM [31], PIRenderer [29] and StyleHEAT [40], which use warping fields for reenacting faces, our method can handle extreme pose variation and maintain identity consistency. Compared to the 3D method of HeadNeRF, our model fully reconstructs the identity-specific detail and syn" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16906" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.093, + 0.895, + 0.47 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.48, + 0.895, + 0.522 + ], + "angle": 0, + "content": "Figure 4. Qualitative result for multi-view reenactment. Examples are sampled from the multi-view dataset [38]. All methods use the first frame of the frontal-view portrait to extract the identity feature, and take the expressions of sequential frames and poses of different camera views to generate the talking face. Note that this subject is not included in the training set of any methods." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.543, + 0.897, + 0.655 + ], + "angle": 0, + "content": "
Multi-View ReenactmentCross-Identity Reenactment
PSNR ↑SSIM ↑CSIM ↑AED ↓APD ↓AKD ↓LIPIPS ↓FID ↓CSIM ↑AED ↓APD ↓AKD ↓FID ↓
FOMM20.750.6390.5052.0040.5455.0520.308101.60.6723.1960.5004.198113.2
PIRenderer20.040.5860.4932.2030.6806.5660.299100.60.6323.0180.4984.977103.7
StyleHEAT20.030.6320.3872.1790.4725.5220.284123.80.6142.8600.4713.592239.1
HeadNeRF17.600.5460.2392.0860.7764.1660.367212.30.2822.8730.5673.465233.0
Ours21.190.6570.5741.8740.4283.7310.288137.30.6942.8500.4054.307101.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.188, + 0.664, + 0.781, + 0.678 + ], + "angle": 0, + "content": "Table 1. Quantitative comparisons on multi-View reenactment and cross-identity reenactment." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.47, + 0.825 + ], + "angle": 0, + "content": "thesizes more natural expression. The multi-view consistency results are shown in Fig. 4. First, the 2D warping method suffers facial malformation, which becomes more serious in larger poses. Second, though HeadNeRF renders accurate head poses, the generated results have noticeable deterioration compared to the ground truth. Finally, the face avatar constructed by our method preserves identity details and multi-view consistency." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.469, + 0.884 + ], + "angle": 0, + "content": "Quantitative comparison. The quantitative comparison among competing methods is shown in Table 1. We can see that the proposed method achieves the best performance in terms of most of the criteria." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Inference speed. Table 2 lists the inference speed com" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.766 + ], + "angle": 0, + "content": "parison among state-of-the-art 3D face avatar methods. By utilizing the pre-trained 3D face generator and the compact motion controller (0.8M parameters), our method achieves the highest inference speed, demonstrating its effectiveness." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.772, + 0.655, + 0.789 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Decoupling-by-inverting. The controller is trained by the decoupling-by-inverting strategy described in Sec. 3.2, where the training alternates between \\( N_{id} \\) steps of identity code optimization and \\( N_{mo} \\) steps of controller parameters training. To validate the effectiveness of the alternative manner, we discard it and jointly train the controller parameters and the identity code. The qualitative comparison is" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16907" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.086, + 0.089, + 0.462, + 0.18 + ], + "angle": 0, + "content": "
MethodsFrames Per Second ↑
IMAvatar [45]0.03
ADNeRF [15]0.13
HeadNeRF [17]25
Ours35
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.192, + 0.47, + 0.22 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison on the inference speed. The comparison is conducted on an A100 GPU." + }, + { + "type": "image", + "bbox": [ + 0.137, + 0.241, + 0.411, + 0.511 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.522, + 0.47, + 0.565 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparison of decoupling-by-inverting training and joint training. Joint training cannot preserve identity information in one-shot avatar construction." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.595, + 0.468, + 0.746 + ], + "angle": 0, + "content": "in Fig. 5. We see that the jointly trained model cannot preserve identity during the animation because the controller manages to overfit the identity information in the training set, and the identity information is not fully encoded in the identity code. The performance degradation caused by joint training is also shown in Table 4. We also analyze the setting of identity optimization step \\( N_{id} \\) and controller training step \\( N_{mo} \\), and list the results in Table 3. We observe that \\( N_{id} = 90 \\), \\( N_{mo} = 10 \\) achieves a balance of expression reconstruction (AED) and identity preserving (CSIM)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.748, + 0.468, + 0.854 + ], + "angle": 0, + "content": "Losses. We also perform an ablation study on the loss functions, including the landmark region losses \\(\\mathcal{L}_m\\) and ID losses \\(\\mathcal{L}_{id}\\), as shown in Table 4. We can see that the absence of landmark region losses \\(\\mathcal{L}_m\\) causes a deterioration of expression reconstruction, measured by AED, while the absence of ID loss causes a performance drop in both expression and identity consistency." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "3D generator finetuning. When the controller training and identity code optimization are finished, we further slightly finetune the pre-trained EG3D [3] face generator" + }, + { + "type": "table", + "bbox": [ + 0.529, + 0.089, + 0.865, + 0.18 + ], + "angle": 0, + "content": "
NmoNidAED ↓CSIM ↑
1993.2850.7200
5953.1780.716
10902.8500.694
20803.2310.592
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.192, + 0.892, + 0.234 + ], + "angle": 0, + "content": "Table 3. Ablation study on the decoupling-by-inverting hyperparameters. Experiments are conducted under the different settings of \\( N_{mo} \\) and \\( N_{id} \\) with \\( N = N_{mo} + N_{id} = 100 \\)." + }, + { + "type": "table", + "bbox": [ + 0.536, + 0.248, + 0.858, + 0.356 + ], + "angle": 0, + "content": "
AED ↓CSIM ↑
ours2.8500.694
joint3.0090.689
w/o Lm3.1010.687
w/o Lid3.1640.592
w/o finetune3.1450.687
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.367, + 0.892, + 0.395 + ], + "angle": 0, + "content": "Table 4. Ablation study on the joint training, loss terms, and finetuning." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.424, + 0.892, + 0.485 + ], + "angle": 0, + "content": "with a learning rate of 0.0001. In Table 4, we show the results without finetuning. One can see that slightly finetuning the face generator in the final stage improves identity and expression reconstruction performance." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.5, + 0.619, + 0.516 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.526, + 0.892, + 0.814 + ], + "angle": 0, + "content": "We proposed a novel framework of one-shot 3D-consistent talking face avatar, namely OTAvatar, using volume rendering. It jointly addressed the three challenges of face avatars, i.e., generalizability, controllability, and efficiency. For identity-generalized rendering, we implemented a pre-trained 3D face generator to reconstruct faces faithfully, given a single portrait reference. For motion control, we proposed a motion controller module, which predicts the motion code conditioned on 3DMM coefficients. For efficiency, benefiting from the compact architecture of both the 3D generator and controller, our model can animate avatars at a high speed. Besides, we proposed a decoupling-by-inverting approach, which is both a training scheme and a test-time disentangle strategy that decouples the latent code into identity and motion codes via GAN inversion, so that one can animate the avatar with different motions using the motion codes predicted by the controller. Comprehensive experiments were conducted to prove the advantage of our proposed framework." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.83, + 0.661, + 0.846 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This work is supported in part by the Chinese National Natural Science Foundation Projects #62176256, #62276254, and the InnoHK program." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16908" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan: How to embed images into the stylegan latent space? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4432-4441, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.172, + 0.472, + 0.227 + ], + "angle": 0, + "content": "[2] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan++: How to edit the embedded images? In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8296-8305, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.228, + 0.47, + 0.297 + ], + "angle": 0, + "content": "[3] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In CVPR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.298, + 0.47, + 0.368 + ], + "angle": 0, + "content": "[4] Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5799-5809, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.368, + 0.47, + 0.436 + ], + "angle": 0, + "content": "[5] Lele Chen, Ross K Maddox, Zhiyao Duan, and Chenliang Xu. Hierarchical cross-modal talking face generation with dynamic pixel-wise loss. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7832-7841, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.438, + 0.47, + 0.48 + ], + "angle": 0, + "content": "[6] Min Jin Chong, Hsin-Ying Lee, and David Forsyth. Stylegan of all trades: Image manipulation with only pretrained stylegan. arXiv preprint arXiv:2111.01619, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.481, + 0.47, + 0.522 + ], + "angle": 0, + "content": "[7] Joon Son Chung and Andrew Zisserman. Lip reading in the wild. In Asian conference on computer vision, pages 87-103. Springer, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.524, + 0.47, + 0.592 + ], + "angle": 0, + "content": "[8] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.593, + 0.47, + 0.662 + ], + "angle": 0, + "content": "[9] Yu Deng, Jiaolong Yang, Sicheng Xu, Dong Chen, Yunde Jia, and Xin Tong. Accurate 3d face reconstruction with weakly-supervised learning: From single image to image set. In IEEE Computer Vision and Pattern Recognition Workshops, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.663, + 0.47, + 0.718 + ], + "angle": 0, + "content": "[10] Michail Christos Doukas, Stefanos Zafeiriou, and Viktoriia Sharmanska. Headgan: One-shot neural head synthesis and editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14398-14407, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.72, + 0.47, + 0.761 + ], + "angle": 0, + "content": "[11] Randima Fernando et al. GPU gems: programming techniques, tips, and tricks for real-time graphics, volume 590. Addison-Wesley Reading, 2004." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.762, + 0.47, + 0.831 + ], + "angle": 0, + "content": "[12] Guy Gafni, Justus Thies, Michael Zollhöfer, and Matthias Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8649-8658, June 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[13] Guy Gafni, Justus Thies, Michael Zollhofer, and Matthias Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8649-8658, 2021." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.147 + ], + "angle": 0, + "content": "[14] Zhenglin Geng, Chen Cao, and Sergey Tulyakov. 3d guided fine-grained face manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9821-9830, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.894, + 0.218 + ], + "angle": 0, + "content": "[15] Yudong Guo, Keyu Chen, Sen Liang, Yong-Jin Liu, Hujun Bao, and Juyong Zhang. Ad-nerf: Audio driven neural radiance fields for talking head synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5784–5794, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.219, + 0.893, + 0.286 + ], + "angle": 0, + "content": "[16] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.288, + 0.893, + 0.356 + ], + "angle": 0, + "content": "[17] Yang Hong, Bo Peng, Haiyao Xiao, Ligang Liu, and Juyong Zhang. Headnerf: A real-time nerf-based parametric head model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20374-20384, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.358, + 0.893, + 0.427 + ], + "angle": 0, + "content": "[18] Xinya Ji, Hang Zhou, Kaisiyuan Wang, Wayne Wu, Chen Change Loy, Xun Cao, and Feng Xu. Audio-driven emotional video portraits. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14080-14089, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.428, + 0.894, + 0.496 + ], + "angle": 0, + "content": "[19] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.498, + 0.893, + 0.525 + ], + "angle": 0, + "content": "[20] Marc Levoy. Display of surfaces from volume data. IEEE Computer graphics and Applications, 8(3):29-37, 1988." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.526, + 0.893, + 0.58 + ], + "angle": 0, + "content": "[21] Stephen Lombardi, Tomas Simon, Jason Saragih, Gabriel Schwartz, Andreas Lehrmann, and Yaser Sheikh. Neural volumes: Learning dynamic renderable volumes from images. ACM Trans. Graph., 38(4):65:1-65:14, July 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.581, + 0.893, + 0.649 + ], + "angle": 0, + "content": "[22] Marko Mihajlovic, Aayush Bansal, Michael Zollhoefer, Siyu Tang, and Shunsuke Saito. KeypointNeRF: Generalizing image-based volumetric avatars using relative spatial encoding of keypoints. In European conference on computer vision, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.65, + 0.893, + 0.705 + ], + "angle": 0, + "content": "[23] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.707, + 0.893, + 0.762 + ], + "angle": 0, + "content": "[24] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11453-11464, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.762, + 0.893, + 0.831 + ], + "angle": 0, + "content": "[25] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865-5874, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.832, + 0.893, + 0.901 + ], + "angle": 0, + "content": "[26] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Ricardo Martin-Brualla, and Steven M. Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. ACM Trans. Graph., 40(6), dec 2021." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16909" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[27] KR Prajwal, Rudrabha Mukhopadhyay, Vinay P Namboodiri, and CV Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In Proceedings of the 28th ACM International Conference on Multimedia, pages 484-492, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.164, + 0.47, + 0.219 + ], + "angle": 0, + "content": "[28] Amit Raj, Michael Zollhoefer, Tomas Simon, Jason Saragih, Shunsuke Saito, James Hays, and Stephen Lombardi. Pva: Pixel-aligned volumetric avatars. arXiv preprint arXiv:2101.02697, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.221, + 0.47, + 0.289 + ], + "angle": 0, + "content": "[29] Yurui Ren, Ge Li, Yuanqi Chen, Thomas H Li, and Shan Liu. Pirenderer: Controllable portrait image generation via semantic neural rendering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13759-13768, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.291, + 0.47, + 0.347 + ], + "angle": 0, + "content": "[30] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. Graf: Generative radiance fields for 3d-aware image synthesis. Advances in Neural Information Processing Systems, 33:20154-20166, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.349, + 0.469, + 0.404 + ], + "angle": 0, + "content": "[31] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. First order motion model for image animation. Advances in Neural Information Processing Systems, 32, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.406, + 0.469, + 0.475 + ], + "angle": 0, + "content": "[32] Aliaksandr Siarohin, Oliver J Woodford, Jian Ren, Menglei Chai, and Sergey Tulyakov. Motion representations for articulated animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13653-13662, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.477, + 0.469, + 0.547 + ], + "angle": 0, + "content": "[33] Ivan Skorokhodov, Sergey Tulyakov, and Mohamed Elhoseiny. Stylegan-v: A continuous video generator with the price, image quality and perks of stylegan2. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3626-3636, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.548, + 0.469, + 0.617 + ], + "angle": 0, + "content": "[34] Justus Thies, Michael Zollhofer, Marc Stamminger, Christian Theobalt, and Matthias Nießner. Face2face: Real-time face capture and reenactment of rgb videos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2387-2395, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.619, + 0.469, + 0.674 + ], + "angle": 0, + "content": "[35] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Guilin Liu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. Video-to-video synthesis. In Advances in Neural Information Processing Systems (NeurIPS), 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.676, + 0.469, + 0.731 + ], + "angle": 0, + "content": "[36] Yaohui Wang, Di Yang, Francois Bremond, and Antitza Dantcheva. Latent image animator: Learning to animate images via latent space navigation. arXiv preprint arXiv:2203.09043, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.733, + 0.469, + 0.788 + ], + "angle": 0, + "content": "[37] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[38] Cheng-hsin Wu, Ningyuan Zheng, Scott Ardisson, Rohan Bali, Danielle Belko, Eric Brockmeyer, Lucas Evans, Timothy Godisart, Hyowon Ha, Alexander Hypes, Taylor Koska, Steven Krenn, Stephen Lombardi, Xiaomin Luo, Kevyn McPhail, Laura Millerschoen, Michal Perdoch, Mark Pitts, Alexander Richard, Jason Saragih, Junko Saragih, Takaaki Shiratori, Tomas Simon, Matt Stewart, Autumn Trimble, Xinshuo Weng, David Whitewolf, Chenglei Wu, Shouu-I Yu," + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "and Yaser Sheikh. Multiface: A dataset for neural face rendering. In arXiv, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[39] Tao Yang, Peiran Ren, Xuansong Xie, and Lei Zhang. Gan prior embedded network for blind face restoration in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 672-681, June 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[40] Fei Yin, Yong Zhang, Xiaodong Cun, Mingdeng Cao, Yanbo Fan, Xuan Wang, Qingyan Bai, Baoyuan Wu, Jue Wang, and Yujiu Yang. Styleheat: One-shot high-resolution editable talking face generation via pretrained stylegan. arXiv preprint arXiv:2203.04036, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.263, + 0.892, + 0.319 + ], + "angle": 0, + "content": "[41] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4578-4587, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.32, + 0.892, + 0.388 + ], + "angle": 0, + "content": "[42] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.39, + 0.892, + 0.459 + ], + "angle": 0, + "content": "[43] Zhimeng Zhang, Lincheng Li, Yu Ding, and Changjie Fan. Flow-guided one-shot talking face generation with a high-resolution audio-visual dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3661–3670, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.46, + 0.892, + 0.514 + ], + "angle": 0, + "content": "[44] Zicheng Zhang, Yinglu Liu, Congying Han, Tiande Guo, Ting Yao, and Tao Mei. Generalized one-shot domain adaption of generative adversarial networks. arXiv preprint arXiv:2209.03665, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.517, + 0.892, + 0.572 + ], + "angle": 0, + "content": "[45] Yufeng Zheng, Victoria Fernández Abrevaya, Marcel C. Bühler, Xu Chen, Michael J. Black, and Otmar Hilliges. I M Avatar: Implicit morphable head avatars from videos. In Computer Vision and Pattern Recognition (CVPR), 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.574, + 0.892, + 0.628 + ], + "angle": 0, + "content": "[46] Hang Zhou, Yu Liu, Ziwei Liu, Ping Luo, and Xiaogang Wang. Talking face generation by adversarially disentangled audio-visual representation. In AAAI Conference on Artificial Intelligence (AAAI), 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.63, + 0.892, + 0.699 + ], + "angle": 0, + "content": "[47] Hang Zhou, Yasheng Sun, Wayne Wu, Chen Change Loy, Xiaogang Wang, and Ziwei Liu. Pose-controllable talking face generation by implicitly modularized audio-visual representation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.699 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16910" + } + ] +] \ No newline at end of file diff --git a/2023/OTAvatar_ One-Shot Talking Face Avatar With Controllable Tri-Plane Rendering/aea7643c-93a9-4960-bd4a-4385733b4c38_origin.pdf b/2023/OTAvatar_ One-Shot Talking Face Avatar With Controllable Tri-Plane Rendering/aea7643c-93a9-4960-bd4a-4385733b4c38_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5519dc8793e5764589caeb314d3723282cd3339b --- /dev/null +++ b/2023/OTAvatar_ One-Shot Talking Face Avatar With Controllable Tri-Plane Rendering/aea7643c-93a9-4960-bd4a-4385733b4c38_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6347bbc8f1b6022cf3502a8345104a95a44adfc1cc5c8abfd142a5a720e4ba13 +size 1468903 diff --git a/2023/OTAvatar_ One-Shot Talking Face Avatar With Controllable Tri-Plane Rendering/full.md b/2023/OTAvatar_ One-Shot Talking Face Avatar With Controllable Tri-Plane Rendering/full.md new file mode 100644 index 0000000000000000000000000000000000000000..60f659cef90efc64fc5397dc3612a8911bd6f8e1 --- /dev/null +++ b/2023/OTAvatar_ One-Shot Talking Face Avatar With Controllable Tri-Plane Rendering/full.md @@ -0,0 +1,324 @@ +# OTAvatar: One-shot Talking Face Avatar with Controllable Tri-plane Rendering + +Zhiyuan Ma $^{1,2*}$ + +Xiangyu Zhu $^{3*}$ + +Guojun Qi + +Zhen Lei $^{1,2,3\dagger}$ + +Lei Zhang + +1The Hong Kong Polytechnic University + +$^{2}$ Center for Artificial Intelligence and Robotics, HKISI CAS + +$^{3}$ State Key Laboratory of Multimodal Artificial Intelligence Systems, CASIA + +$^{4}$ OPPO Research + +zm2354.ma@connect.polyu.hk, xiangyu.zhu@nlpr.ia.ac.cn + +guojunq@gmail.com, zlei@nlpr.ia.ac.cn, cslzhang@comp.polyu.edu.hk + +# Abstract + +Controllability, generalizability and efficiency are the major objectives of constructing face avatars represented by neural implicit field. However, existing methods have not managed to accommodate the three requirements simultaneously. They either focus on static portraits, restricting the representation ability to a specific subject, or suffer from substantial computational cost, limiting their flexibility. In this paper, we propose One-shot Talking face Avatar (OTAvatar), which constructs face avatars by a generalized controllable tri-plane rendering solution so that each personalized avatar can be constructed from only one portrait as the reference. Specifically, OTAvatar first inverts a portrait image to a motion-free identity code. Second, the identity code and a motion code are utilized to modulate an efficient CNN to generate a tri-plane formulated volume, which encodes the subject in the desired motion. Finally, volume rendering is employed to generate an image in any view. The core of our solution is a novel decoupling-by-inverting strategy that disentangles identity and motion in the latent code via optimization-based inversion. Benefiting from the efficient tri-plane representation, we achieve controllable rendering of generalized face avatar at 35 FPS on A100. Experiments show promising performance of cross-identity reenactment on subjects out of the training set and better 3D consistency. The code is available at https://github.com/theEricMa/OTAvatar. + +# 1. Introduction + +Neural rendering has achieved remarkable progress and promising results in 3D reconstruction. Thanks to the differentiability in neuron computation, the neural rendering + +![](images/374d618906637c283a736c98874ae17ab80e20c09c7399c9b680ad80e2949c8a.jpg) +Figure 1. OTAvatar animation results. The source subjects in HDTF [43] dataset are animated by OTAvatar using a single portrait as the reference. We use the pose and expression coefficients of 3DMM to represent motion and drive the avatar. Note that these subjects are not included in the training data of OTAvatar. + +methods bypass the expensive cost of high-fidelity digital 3D modeling, thus attracting the attention of many researchers from academia and industry. In this work, we aim to generate a talking face avatar via neural rendering techniques, which is controllable by driving signals like video and audio segments. Such animation is expected to inherit the identity from reference portraits, while the expression and pose can keep in sync with the driving signals. + +The early works about talking face focus on expression animation consistent with driving signals in constrained frontal face images [5, 7, 27]. It is then extended to in-the-wild scenes with pose variations [18, 35, 46, 47]. Many + +previous works are 2D methods, where the image warping adopted to stimulate the motion in 3D space is learned in 2D space [29, 31, 32, 36, 40]. These methods tend to collapse under large pose variation. In contrast, there are 3D methods that can address the pose problem, since the head pose can be treated as a novel view. Both explicit [11, 20] and implicit 3D representations [23] are introduced to face rendering [15, 17, 21, 22, 25, 28, 45]. However, these methods either overfit to a single identity or fail to produce high-quality animation for different identities. + +In this work, we propose a one-shot talking face avatar (OTAvatar), which can generate mimic expressions with good 3D consistency and be generalized to different identities with only one portrait reference. Some avatar animation examples are shown in Fig. 1. Given a single reference image, OTAvatar can drive the subject with motion signals to generate corresponding face images. We realize it under the framework of volume rendering. Current methods usually render static subjects [22, 28]. Although there are works [12, 15, 25, 26, 45] proposed to implement dynamic rendering for face avatars, they need one model per subject. Therefore, the generalization is poor. HeadNeRF [17] is a similar work to us. However, its reconstruction results are unnatural in the talking face case. + +Our method is built on a 3D generative model pre-trained on a large-scale face database to guarantee identity generalization ability. Besides, we employ a motion controller to decouple the motion and identity in latent space when performing the optimization-based GAN inversion, to make the motion controllable and transferable to different identities. The network architecture is compact to ensure inference efficiency. In the utilization of our OTAvatar, given a single reference image, we fix the motion code predicted by the controller and only optimize the identity code so that a new avatar of the reference image can be constructed. Such a disentanglement enables the rendering of any desired motion by simply alternating the motion-related latent code to be that of the specific motion representation. + +The major contributions of this work can be summarized as follows. + +- We make the first attempt for one-shot 3D face reconstruction and motion-controllable rendering by taming a pre-trained 3D generative model for motion control. +- We propose to decouple motion-related and motion-free latent code in inversion optimization by prompting the motion fraction of latent code ahead of the optimization using a decoupling-by-inverting strategy. +- Our method can photo-realistically render any identity with the desired expression and pose at 35FPS. The experiment shows promising results of natural motion and 3D consistency on both 2D and 3D datasets. + +# 2. Related Works + +# 2.1. Talking Face + +Regarding the driving signal, the talking face methods can be roughly divided into three categories: audio-driven, image-driven and coefficients-driven. Our approach is most related to coefficients-driven methods, which use either facial landmarks or 3DMM coefficients to represent motion. Facial landmarks involve identity and expression information and is challenging to transfer motion across different subjects. 3DMM coefficients disentangle the identity, expression and pose, and thus they are good driven signals to drive different faces. + +Thies et al. [34] used 3D rendering to maintain the shape and illumination attributes when transferring expression and refining mouth details with a mouth retrieval algorithm. Geng et al. [14] used 3DMM to render a given subject with different expressions, followed by a neural network to refine the detail and harmonize with the background. These methods cannot change the face pose. Ren et al. [29] mapped 3DMM expression and pose to a high-dimensional motion representation, then predicted the dense image flow to reenact faces via image warping. Doukas et al. [10] used additional 3DMM mesh fitting results to assist the dense image flow prediction. Yin et al. [40] integrated pre-trained StyleGAN [6] to generate high-resolution talking face prediction by warping low-resolution feature maps. These warping-based methods follow the two-stage workflow by first warping images and then refining facial detail, demonstrating superior performance for same identity reenactment. But in cross-identity reenactment, especially when there is a large pose variation against the source portrait, the animation tends to collapse, since image warping is merely trained to stimulate 3D motion in the 2D space. Our work animates face avatars using 3D rendering and explicitly ensures 3D consistency to handle large pose variations. + +# 2.2. Volume Rendering + +Volume rendering is a 3D rendering strategy and has demonstrated its success in novel view synthesis. Existing methods either use volume [11, 21] to explicitly represent 3D space or implicitly store the 3D scene in the MLP network [23]. These methods are introduced to animate 3D-consistent portraits. Yu et al. [41] and Raj et al. [28] retrieved the spatial information from sparse support views to render portraits. On the other side, The 3D face generative model can reconstruct a given identity using GAN Inversion. These models vary from pure MLP architecture [4], to low-resolution neural rendering followed by an up-sampling strategy [24, 30], and to using more efficient 3D presentation [3]. All these methods only support rendering static portraits. To animate the avatar with controllable motion, AD-NeRF [15] and NeRFace [13] introduce + +![](images/bfcd6020e08ff84474dce29dfa2a4659e7f33557fd48d2e046a45ebe554931d4.jpg) +Figure 2. Overview of OTAvatar. OTAvatar contains a 3D face animatior network and a latent code decoupling strategy, namely decoupling-by-inverting. The 3D face animatior is composed of a pre-trained 3D face generator [3] and a motion controller module. The decoupling-by-inverting algorithm is an optimization-based image inversion that can decouple the latent code $\mathbf{w}$ into identity code $\mathbf{w}_{id}$ and motion code $\mathbf{w}_x$ . When the model is well-trained, the motion-free identity code can be inferred from a single reference image, and an avatar of an unseen subject can be constructed. The identity code can be integrated with any other motion code predicted by the controller to animate the identity with desired motion. + +either the audio feature or 3DMM expression coefficients to the NeRF model to render dynamic heads. Nerfies [25] employs a deformation field conditioned on spatial points to represent the face motion of different frames. HyperNeRF [26] adds an ambient slicing network to tackle the topological drawback of Nerfies. Though these methods have achieved motion control and view consistency for face avatars, they all overfit each NeRF model to one subject, thus lack generalizability. HeadNeRF [17] is trained on both multi-view and frontal face datasets and can be generalized to multiple identities and support motion control using 3DMM coefficients, but its rendering quality is not satisfactory compared with the given portrait, and the motion control is unnatural with random jitters. In contrast, our method can reconstruct photo-realistic face avatars in one shot and animate natural talking motion. + +# 2.3. Generative Prior + +Many methods rely on pre-trained generative models. Employing a generative model as the prior has the following advantages. First, it provides rich and diverse facial priors, such as texture, color, shape and pose, which can help to restore realistic and faithful facial details. For instance, Yang et al. [39] leveraged a pre-trained GAN model as a prior decoder for blind face restoration. Second, it also enables one- + +shot image manipulation both inside and outside of the domain. Well-designed schemes for disentanglement are necessary to provide acceptable performance on fine-grained out-of-domain image manipulation. Zhang et al. [44] suggested separating the attributes into global style—like texture and color—and structure—like shape and pose, and performing domain-adaptive generation by transferring the decoupled style attributes. The image manipulation also includes synthesizing human talking videos, Ivan et al. [33] modeled the temporal dynamics of video frames through the latent code disentanglement. Unlike other talking head generation methods [31, 40], it only synthesizes center-aligned faces. The recently presented 3D GANs have been popularized for their ability to generate 3D-consistent photorealistic human faces with explicit pose control. Katja et.al [30] introduced 3D scene volume rendering without resorting to computationally demanding voxel-based representation, and the model is trained from unposed 2D photos in an adversarial manner. Chan et al. [3] proposed to generate high-quality geometry and multi-view consistent images from 2D photos through a hybrid explicit-implicit network architecture based on StyleGAN2 [19]. Despite the fact that 3D GANs' explicit pose control and 3D-consistent generation are appealing for talking face synthesis, no previous work has studied their application to talking face avatars. + +# 3. Method + +Talking face avatar aims to synthesize face images with controllable expression and pose. In this paper, we investigate the possibility of endowing volume rendering with the ability to 1) build a faithful identity representation in one shot, 2) enable natural motion control, and 3) achieve real-time inference speed. Our framework takes an identity code and a motion signal as input and generates a tri-plane-based [3] volume through a CNN architecture. By performing volume rendering on the tri-plane representation, where each point is projected on the three feature planes and the sampled features are fused to occupancy and color, a face image at a given camera view can be generated as: + +$$ +\mathbf {I} (\mathbf {p}) = \mathcal {R} (G (\mathbf {w} _ {i d}, \mathbf {x}; \Theta), \mathbf {p}), \tag {1} +$$ + +where $\mathbf{w}_{id}$ is the identity code, $\mathbf{x}$ is the motion signal, $G(\cdot ,\cdot ;\Theta)$ is a face volume generation network with $\Theta$ as its network weights to reconstruct an encoded portrait in three orthogonal feature planes, $\mathcal{R}(\mathbf{V},\mathbf{p})$ is the volume rendering with $\mathbf{V}$ as the volume and $\mathbf{p}$ as the camera view, and $\mathbf{I}(\mathbf{p})\in \mathbb{R}^{3\times H\times W}$ is the rendered result. In our implementation, the identity code $\mathbf{w}_{id}$ can be extracted by optimizing on a single reference image using our proposed decoupling-by-inverting strategy (Sec. 3.3), achieving one-shot avatar reconstruction. The motion signal $\mathbf{x}$ is 3DMM coefficients for flexible control. We call $G(\cdot ,\cdot ;\Theta)$ as 3D Animator because of its ability to animate an avatar with desired motion. Compared with 2D methods, we construct a neural 3D space to promise multi-view consistency. Compared with NeRF, we build a 3D space in the propagation of a CNN, and directly sample features on the tri-plane representation, saving the MLP computation on each sampled point. + +# 3.1. 3D Animator Network Structure + +Tri-plane volume representation. The output of the 3D anemotor network $G(\cdot, \cdot; \Theta)$ in Eqn. 1 is a tri-plane volume representation, which is composed of three feature planes: + +$$ +\mathbf {V} _ {t r i} = (\mathbf {F} _ {x y}, \mathbf {F} _ {x z}, \mathbf {F} _ {y z}) = G (\mathbf {w} _ {i d}, \mathbf {x}; \boldsymbol {\Theta}), \qquad (2) +$$ + +where $\mathbf{F}_{xy}$ , $\mathbf{F}_{xz}$ , and $\mathbf{F}_{yz}$ are three axis-aligned orthogonal feature maps in the 3D space, implicitly forming the volume $\mathbf{V}_{tri}$ . When performing volume rendering, for each queried point $(x,y,z)$ , we project it onto each of three feature maps and retrieve the corresponding features $(\mathbf{F}_{xy}(x,y),\mathbf{F}_{xz}(x,z),\mathbf{F}_{yz}(y,z))$ . The summation of features is sent to a lightweight MLP to decode color and opacity. Compared to fully implicit MLP architectures [4,15,45] or volume representation [21], we can efficiently regress the three 2D feature maps through a CNN architecture. In our implementation, $G$ is a deconvolutional network [6] that outputs three $256\times 256\times 32$ feature maps. + +Animator structure. Towards animating the face avatar of a given portrait, an intuitive solution is representing $\mathbf{w}_{id}$ with a pre-defined feature and training $G(\cdot)$ from scratch. However, we find that the model does not work well due to the sub-optimal $\mathbf{w}_{id}$ . For example, when employing 3DMM shape coefficients [9] and face recognition features [8], the identity information is not preserved in the rendering results because $\mathbf{w}_{id}$ does not encode face appearances well. These models are trained on cropped facial images which contain no torso and hairstyle. Therefore, $\mathbf{w}_{id}$ losses such information that is necessary when animating portraits. The network even collapses if we make $\mathbf{w}_{id}$ learnable. + +In this work, we employ a two-phase strategy to achieve one-shot avatar reconstruction: 1) building a 3D face generator, and 2) making the generator controllable. We build our 3D face generator on a pre-trained EG3D [3] network, which incorporates tri-plane representation for efficient 3D face generation: + +$$ +\mathbf {V} _ {t r i} = G _ {e g} (\mathbf {w}; \Theta_ {e g}), \tag {3} +$$ + +where $\mathbf{w}$ is an uninterpretable latent code drawn from the latent space of the generator $G_{eg}$ . The latent space is determined by the generator weight $\Theta_{eg}$ . To control the latent code $\mathbf{w}$ by the given identity represented by $\mathbf{w}_{id}$ and motion signal $\mathbf{x}$ , we propose a motion controller module $C$ parameterized by $\Theta_c$ , which maps motion signal $\mathbf{x}$ to the motion code $\mathbf{w}_x$ , such that: + +$$ +\mathbf {w} = \mathbf {w} _ {i d} + \mathbf {w} _ {x} = \mathbf {w} _ {i d} + C (\mathbf {x}; \Theta_ {c}). \tag {4} +$$ + +Injecting Eqn. 4 to Eqn. 3, we have the structure of the 3D animatior: + +$$ +\begin{array}{c} G (\mathbf {w} _ {i d}, \mathbf {x}; \Theta) = G _ {e g} (\mathbf {w} _ {i d} + C (\mathbf {x}; \Theta_ {c}); \Theta_ {e g}), \\ \Theta = \Theta_ {e g} \cup \Theta_ {c}. \end{array} \tag {5} +$$ + +In our implementation, the motion signal $\mathbf{x}$ is constructed by 3DMM pose and expression coefficients [9], which are more flexible compared with image-based driving signals [31, 32, 36], especially in cross-identity driving. + +Controller structure. We use the concatenation of 3DMM pose and expression coefficients to represent motion $\mathbf{x}$ . Practically, we use the adjacent coefficients of pose and expression within a particular radius to represent the motion signal of the current frame, firstly the weighted summation of the coefficients across the temporal dimension is conducted to avoid noises, which is achieved by three 1D convolution layers [29, 40]; secondly, a five-layer MLP is employed to transform the weighted summation into a motion feature; finally, a codebook with learnable orthogonal bases [36] is built, on which each motion feature is projected to get the final motion code $\mathbf{w}_x$ . The aforementioned operation is summarized as follows: + +$$ +\mathbf {w} _ {x} = C (\mathbf {x}; \Theta_ {c}) = \mathbf {D} * F _ {M} \left(F _ {T} (\mathbf {x}; \Theta_ {T}); \Theta_ {M}\right), \tag {6} +$$ + +where $F_{T}$ parameterized by $\Theta_T$ is the temporal smoothing network of 1D convolution layers, $F_{M}$ parameterized by $\Theta_M$ is the five-layer MLP which projects motion to the magnitudes of bases contained in the codebook $\mathbf{D}$ . It can be seen that the magnitude projection parameters, temporal smoothing weights, and the codebook are the parameters of the controller $C$ to be trained, that is: + +$$ +\Theta_ {c} = \mathbf {D} \cup \Theta_ {T} \cup \Theta_ {M}. \tag {7} +$$ + +# 3.2. Controller Training + +Given a couple of source frame $\mathbf{I}_s$ and driving frame $\mathbf{I}_d$ , along with their motion signals $\mathbf{x}_s, \mathbf{x}_d$ and camera views $\mathbf{p}_s, \mathbf{p}_d$ , we perform the dual-objective optimization: + +$$ +\mathbf{w}_{id}^{*} = \operatorname *{arg min}_{\mathbf{w}_{id}}(\mathcal{L}_{s} + \mathcal{L}_{d}), +$$ + +$$ +\mathcal {L} _ {s} = \mathcal {L} \left(I _ {s}, \hat {I} _ {s}\right) = \mathcal {L} \left(I _ {s}, \mathcal {R} \left(G \left(\mathbf {w} _ {i d}, \mathbf {x} _ {s}; \Theta\right), \mathbf {p} _ {s}\right), \right. \tag {8} +$$ + +$$ +\mathcal {L} _ {d} = \mathcal {L} (I _ {d}, \hat {I} _ {d}) = \mathcal {L} (I _ {d}, \mathcal {R} (G (\mathbf {w} _ {i d}, \mathbf {x} _ {d}; \Theta), \mathbf {p} _ {d}). +$$ + +It is implemented to estimate the identity code $\mathbf{w}_{id}$ shared between $\mathbf{I}_s$ and $\mathbf{I}_d$ . $G$ , $\mathcal{R}$ , and $\mathcal{L}$ indicate 3D face generator, volume rendering, and loss function, respectively. The 3D face generator $G$ is parameterized by $\Theta = \Theta_{eg} \cup \Theta_c$ . $\Theta_{eg}$ includes the parameters of the pre-trained 3D face generator, and $\Theta_c$ includes the parameters of the controller which is the focus of our training scheme. + +Training scheme. The purpose of controller training is to decouple the identity code and motion code from the latent code of the generator, making identity and motion replaceable in our generalized avatars. To this end, we propose the decoupling-by-inverting training strategy, which is generally an alternation of identity optimization and controller training at different steps. Specifically, we first freeze $\Theta_c$ and perform $N_{id}$ steps of back-propagation on $\mathbf{w}_{id}$ . Then we freeze $\mathbf{w}_{id}$ and perform $N_{mo}$ steps of backpropagation on $\Theta_c$ so that the controller learns to add the motion code on top of the identity code. In our implementation, $N_{id} = 90$ and $N_{mo} = 10$ . After $N = N_{id} + N_{mo}$ steps, we slightly finetune $\Theta_{eg}$ with a learning rate of $10^{-4}$ to make the rendering fit the driving signal better. A brief pseudo-code of the whole scheme is provided in the supplementary material. + +# 3.3. One-shot Avatar Construction + +In the inference stage, given a single portrait image $\mathbf{I}_r$ , its 3DMM coefficients $\mathbf{x}_r$ , and camera view $\mathbf{p}_r$ as the reference of avatar construction, we can estimate the identity code $\mathbf{w}_r$ of the reference portrait via: + +$$ +\mathbf {w} _ {r} = \underset {\mathbf {w} _ {i d}} {\arg \min } \mathcal {L} (\mathbf {I} _ {r}, \mathcal {R} (G (\mathbf {w} _ {i d}, \mathbf {x} _ {r}), \mathbf {p} _ {r}). \tag {9} +$$ + +Afterwards, we can animate the identity with given driving motion $\mathbf{x}_d$ and camera view $\mathbf{p}_d$ , either from the same or + +different subject, through the following operation: + +$$ +\mathbf {I} \left(\mathbf {w} _ {r}, \mathbf {x} _ {d}, \mathbf {p} _ {d}\right) = \mathcal {R} \left(G \left(\mathbf {w} _ {r}, \mathbf {x} _ {d}\right), \mathbf {p} _ {d}\right), \tag {10} +$$ + +where $\mathbf{I}(\mathbf{w}_r,\mathbf{c}_d,\mathbf{p}_d)$ has the identity estimated as $\mathbf{I}_r$ , and is animated with the motion $\mathbf{c}_d$ and camera view $\mathbf{p}_d$ . We name Eqn. 9 as the decoupling-by-inverting inference strategy since it can estimate decoupled identity code from latent code in the iterative optimization-based GAN Inversion, and can be used to render different motions. + +# 3.4. Loss Terms + +The loss $\mathcal{L}(\mathbf{I},\hat{\mathbf{I}})$ in Eqn. 8 and Eqn. 9 measures the error between the animated result $\hat{\mathbf{I}}$ and the ground truth $\mathbf{I}$ . First, a pre-trained VGG-19 network is implemented to calculate the distance of multi-scale activation maps $\phi_{i}(\mathbf{I})$ via: + +$$ +\mathcal {L} _ {c} = \sum_ {i} \left\| \phi_ {i} (\mathbf {I}) - \phi_ {i} (\hat {\mathbf {I}}) \right\| _ {1}, \tag {11} +$$ + +where $\phi_{i}$ is the $i$ -th layer of VGG. Second, the distance between the Gram matrices $G_{j}^{\phi}$ constructed from the $j$ -th activation maps $\phi_{j}$ [29] of $\mathbf{I}$ and $\hat{\mathbf{I}}$ is measured: + +$$ +\mathcal {L} _ {s} = \sum_ {j} \left\| G _ {j} ^ {\phi} (\mathbf {I}) - G _ {j} ^ {\phi} (\hat {\mathbf {I}}) \right\| _ {1}. \tag {12} +$$ + +Third, we utilize the L1 distance performed on the eyes and mouth regions to supervise the expression detail: + +$$ +\mathcal {L} _ {m} = \sum_ {n} \left\| R _ {n} (\mathbf {I}) - R _ {n} (\hat {\mathbf {I}}) \right\| _ {1}, \tag {13} +$$ + +where $R_{n}$ is either the eyes or mouth region extracted using RoI-align on the bounding boxes calculated using landmarks. Finally, the ID Loss is performed to preserve the identity consistency: + +$$ +\mathcal {L} _ {i d} = 1 - \cos (E (\mathbf {I}), E (\tilde {\mathbf {I}})), \tag {14} +$$ + +where $E$ is the Arcface face recognition model [8]. + +The animation loss $\mathcal{L}(\mathbf{I},\hat{\mathbf{I}})$ is a weighted summation of the above loss terms and is implemented in both training and test-time one-shot avatar construction. Additionally, in the training, we maintain the stability of tri-plane representation with monotonic and TV loss [3], and $\mathcal{L}_1$ loss on $\Theta_c$ to avoid the latent code being out of the latent distribution of $G_{eg}$ . + +# 4. Experiments + +# 4.1. Experiment Setting + +In this section, we first describe the implementation detail of our proposed method, the dataset used, and the baselines of our work. Then we compare our proposed method with previous 2D and 3D methods on motion controllability and multi-view consistency. + +Implementation details. We use the off-the-shelf 3D face reconstruction model to extract the expression and pose coefficients of 3DMM [9]. The motion of any timestamp is represented by the window of adjacent 27 frames of expression and pose coefficients. In practice, we find that the extracted head poses contain too much noises to serve as the camera pose. Therefore, we extract the camera pose using the method in [15]. Our model is implemented in Pytorch using four A100 GPUs. The total batch size is 24, with six images per GPU. We use Adam optimizer with a 0.0001 learning rate to train the motion controller $C$ and finetune the generator $G$ . Exponential Moving Average (EMA) is employed to update $\Theta_{c}$ since it can stabilize the training. The model is trained for 2000 iterations, and each iteration contains $N = N_{id} + N_{mo} = 100$ steps to optimize the identity codes and train the motion controller. For each minibatch of data, a new batch of latent codes are initialized and optimized using the Adam optimizer with a 0.01 learning rate. Both the $64 \times 64$ resolution volume rendering results and $512 \times 512$ super-resolution results are implemented to calculate the loss. During training, the identity code $\mathbf{w}_{id}$ is optimized in $\mathcal{W}$ space. After being extended to the $\mathcal{W}^{+}$ space by channel-wisely repeated 14 times, it is combined with the motion code $\mathbf{w}_x$ and then fed into the face generator $G$ . It takes less than two days to train the network entirely. During inference, we first optimize the identity code in $\mathcal{W}$ space, then use another $N = 100$ steps to finetune it in $\mathcal{W}^{+}$ space [1, 2]. + +Dataset. Our model is trained on the HDTF [43] dataset, which contains frontal talking faces from 362 videos and over 300 subjects. Following [40], we use the pre-processing step in [31] and resize images to $512 \times 512$ resolution. The test dataset contains 20 videos of subjects out of the training set. To evaluate the novel subject generalizability and controllability in the 3D-consistent fashion, we sample subjects from the Multiface [38] dataset, which is particularly proposed for 3D-consistent face avatars. + +Baseline methods. We mainly compare our method with previous works that incorporate 3DMM coefficients to reenact face avatars in either 2D/3D fashion and are generalizable to unseen identities. For 2D methods, PIRenderer [29] and StyleHEAT [40] achieve SOTA performance on face reenactment. For 3D methods, HeadNeRF [17] is the most advanced volume rendering method that incorporates controllability and generalizability. Beyond coefficients-based methods, the SOTA image-driven method FOMM [31] is also taken into comparison. + +Evaluation metrics. We adopt peak-signal-to-noise ratio (PSNR), structural similarity index measure (SSIM) [37], and learned perceptual image patch similarity (LPIPS) [42] to evaluate the visual quality. To measure the realism of the synthesized results and the identity preservation, we use frechet inception distance (FID) [16] and + +![](images/18b56b69d9f6f066ca3869b1992f733f12eaf712519100cc1a856d8cbc38bb90.jpg) +Figure 3. Qualitative result for cross-identity reenactment. Examples are sampled from the HDTF dataset [43]. Both source and driving subjects are not included in the training set. + +cosine similarity of identity embedding (CSIM) [8] based on ArcFace model, respectively. Besides, the average expression distance (AED), average pose distance (APD), and average keypoint distance (AKD) are employed to evaluate the facial expression and pose. + +# 4.2. Evaluation Result + +We evaluate OTAvatar on animating photo-realistic talking head videos and compare them with state-of-the-art models that support identity-generalized animation methods. In the HDTF [43] dataset, we examine the identity-motion disentanglement by transferring the motion of one subject to drive the other subjects, namely cross-identity reenactment. The employed motion is extracted from a video with large motion variation to evaluate the result in extreme conditions. In the Multiface [38] dataset, we evaluate the consistency of the animation in different views, namely multi-view reenactment. Note that none of the data in this dataset has been used in training our method and baselines. For each talking corpus, we choose the first frame of the frontal-view camera recording as the reference and enforce the network to animate the following frames in frontal view and other views. + +Qualitative comparison. Fig. 3 shows the results of different methods on cross-identity generation. Compared to FOMM [31], PIRenderer [29] and StyleHEAT [40], which use warping fields for reenacting faces, our method can handle extreme pose variation and maintain identity consistency. Compared to the 3D method of HeadNeRF, our model fully reconstructs the identity-specific detail and syn + +![](images/2d9797dc277adee4d2c846da9fe23815799d108bc8d8dbeb9857ce3cff89e9b3.jpg) +Figure 4. Qualitative result for multi-view reenactment. Examples are sampled from the multi-view dataset [38]. All methods use the first frame of the frontal-view portrait to extract the identity feature, and take the expressions of sequential frames and poses of different camera views to generate the talking face. Note that this subject is not included in the training set of any methods. + +
Multi-View ReenactmentCross-Identity Reenactment
PSNR ↑SSIM ↑CSIM ↑AED ↓APD ↓AKD ↓LIPIPS ↓FID ↓CSIM ↑AED ↓APD ↓AKD ↓FID ↓
FOMM20.750.6390.5052.0040.5455.0520.308101.60.6723.1960.5004.198113.2
PIRenderer20.040.5860.4932.2030.6806.5660.299100.60.6323.0180.4984.977103.7
StyleHEAT20.030.6320.3872.1790.4725.5220.284123.80.6142.8600.4713.592239.1
HeadNeRF17.600.5460.2392.0860.7764.1660.367212.30.2822.8730.5673.465233.0
Ours21.190.6570.5741.8740.4283.7310.288137.30.6942.8500.4054.307101.8
+ +Table 1. Quantitative comparisons on multi-View reenactment and cross-identity reenactment. + +thesizes more natural expression. The multi-view consistency results are shown in Fig. 4. First, the 2D warping method suffers facial malformation, which becomes more serious in larger poses. Second, though HeadNeRF renders accurate head poses, the generated results have noticeable deterioration compared to the ground truth. Finally, the face avatar constructed by our method preserves identity details and multi-view consistency. + +Quantitative comparison. The quantitative comparison among competing methods is shown in Table 1. We can see that the proposed method achieves the best performance in terms of most of the criteria. + +Inference speed. Table 2 lists the inference speed com + +parison among state-of-the-art 3D face avatar methods. By utilizing the pre-trained 3D face generator and the compact motion controller (0.8M parameters), our method achieves the highest inference speed, demonstrating its effectiveness. + +# 4.3. Ablation Study + +Decoupling-by-inverting. The controller is trained by the decoupling-by-inverting strategy described in Sec. 3.2, where the training alternates between $N_{id}$ steps of identity code optimization and $N_{mo}$ steps of controller parameters training. To validate the effectiveness of the alternative manner, we discard it and jointly train the controller parameters and the identity code. The qualitative comparison is + +
MethodsFrames Per Second ↑
IMAvatar [45]0.03
ADNeRF [15]0.13
HeadNeRF [17]25
Ours35
+ +![](images/53172a8c4df956366c8f40b2f19a5b8e53dce57e021ffbeff17931d0bdcb8572.jpg) +Figure 5. Qualitative comparison of decoupling-by-inverting training and joint training. Joint training cannot preserve identity information in one-shot avatar construction. + +in Fig. 5. We see that the jointly trained model cannot preserve identity during the animation because the controller manages to overfit the identity information in the training set, and the identity information is not fully encoded in the identity code. The performance degradation caused by joint training is also shown in Table 4. We also analyze the setting of identity optimization step $N_{id}$ and controller training step $N_{mo}$ , and list the results in Table 3. We observe that $N_{id} = 90$ , $N_{mo} = 10$ achieves a balance of expression reconstruction (AED) and identity preserving (CSIM). + +Losses. We also perform an ablation study on the loss functions, including the landmark region losses $\mathcal{L}_m$ and ID losses $\mathcal{L}_{id}$ , as shown in Table 4. We can see that the absence of landmark region losses $\mathcal{L}_m$ causes a deterioration of expression reconstruction, measured by AED, while the absence of ID loss causes a performance drop in both expression and identity consistency. + +3D generator finetuning. When the controller training and identity code optimization are finished, we further slightly finetune the pre-trained EG3D [3] face generator + +Table 2. Quantitative comparison on the inference speed. The comparison is conducted on an A100 GPU. + +
NmoNidAED ↓CSIM ↑
1993.2850.7200
5953.1780.716
10902.8500.694
20803.2310.592
+ +Table 3. Ablation study on the decoupling-by-inverting hyperparameters. Experiments are conducted under the different settings of $N_{mo}$ and $N_{id}$ with $N = N_{mo} + N_{id} = 100$ . + +
AED ↓CSIM ↑
ours2.8500.694
joint3.0090.689
w/o Lm3.1010.687
w/o Lid3.1640.592
w/o finetune3.1450.687
+ +Table 4. Ablation study on the joint training, loss terms, and finetuning. + +with a learning rate of 0.0001. In Table 4, we show the results without finetuning. One can see that slightly finetuning the face generator in the final stage improves identity and expression reconstruction performance. + +# 5. Conclusion + +We proposed a novel framework of one-shot 3D-consistent talking face avatar, namely OTAvatar, using volume rendering. It jointly addressed the three challenges of face avatars, i.e., generalizability, controllability, and efficiency. For identity-generalized rendering, we implemented a pre-trained 3D face generator to reconstruct faces faithfully, given a single portrait reference. For motion control, we proposed a motion controller module, which predicts the motion code conditioned on 3DMM coefficients. For efficiency, benefiting from the compact architecture of both the 3D generator and controller, our model can animate avatars at a high speed. Besides, we proposed a decoupling-by-inverting approach, which is both a training scheme and a test-time disentangle strategy that decouples the latent code into identity and motion codes via GAN inversion, so that one can animate the avatar with different motions using the motion codes predicted by the controller. Comprehensive experiments were conducted to prove the advantage of our proposed framework. + +# Acknowledgement + +This work is supported in part by the Chinese National Natural Science Foundation Projects #62176256, #62276254, and the InnoHK program. + +# References + +[1] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan: How to embed images into the stylegan latent space? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4432-4441, 2019. +[2] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan++: How to edit the embedded images? In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8296-8305, 2020. +[3] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In CVPR, 2022. +[4] Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5799-5809, 2021. +[5] Lele Chen, Ross K Maddox, Zhiyao Duan, and Chenliang Xu. Hierarchical cross-modal talking face generation with dynamic pixel-wise loss. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7832-7841, 2019. +[6] Min Jin Chong, Hsin-Ying Lee, and David Forsyth. Stylegan of all trades: Image manipulation with only pretrained stylegan. arXiv preprint arXiv:2111.01619, 2021. +[7] Joon Son Chung and Andrew Zisserman. Lip reading in the wild. In Asian conference on computer vision, pages 87-103. Springer, 2016. +[8] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019. +[9] Yu Deng, Jiaolong Yang, Sicheng Xu, Dong Chen, Yunde Jia, and Xin Tong. Accurate 3d face reconstruction with weakly-supervised learning: From single image to image set. In IEEE Computer Vision and Pattern Recognition Workshops, 2019. +[10] Michail Christos Doukas, Stefanos Zafeiriou, and Viktoriia Sharmanska. Headgan: One-shot neural head synthesis and editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14398-14407, 2021. +[11] Randima Fernando et al. GPU gems: programming techniques, tips, and tricks for real-time graphics, volume 590. Addison-Wesley Reading, 2004. +[12] Guy Gafni, Justus Thies, Michael Zollhöfer, and Matthias Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8649-8658, June 2021. +[13] Guy Gafni, Justus Thies, Michael Zollhofer, and Matthias Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8649-8658, 2021. + +[14] Zhenglin Geng, Chen Cao, and Sergey Tulyakov. 3d guided fine-grained face manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9821-9830, 2019. +[15] Yudong Guo, Keyu Chen, Sen Liang, Yong-Jin Liu, Hujun Bao, and Juyong Zhang. Ad-nerf: Audio driven neural radiance fields for talking head synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5784–5794, 2021. +[16] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. +[17] Yang Hong, Bo Peng, Haiyao Xiao, Ligang Liu, and Juyong Zhang. Headnerf: A real-time nerf-based parametric head model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20374-20384, 2022. +[18] Xinya Ji, Hang Zhou, Kaisiyuan Wang, Wayne Wu, Chen Change Loy, Xun Cao, and Feng Xu. Audio-driven emotional video portraits. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14080-14089, 2021. +[19] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. +[20] Marc Levoy. Display of surfaces from volume data. IEEE Computer graphics and Applications, 8(3):29-37, 1988. +[21] Stephen Lombardi, Tomas Simon, Jason Saragih, Gabriel Schwartz, Andreas Lehrmann, and Yaser Sheikh. Neural volumes: Learning dynamic renderable volumes from images. ACM Trans. Graph., 38(4):65:1-65:14, July 2019. +[22] Marko Mihajlovic, Aayush Bansal, Michael Zollhoefer, Siyu Tang, and Shunsuke Saito. KeypointNeRF: Generalizing image-based volumetric avatars using relative spatial encoding of keypoints. In European conference on computer vision, 2022. +[23] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. +[24] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11453-11464, 2021. +[25] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865-5874, 2021. +[26] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Ricardo Martin-Brualla, and Steven M. Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. ACM Trans. Graph., 40(6), dec 2021. + +[27] KR Prajwal, Rudrabha Mukhopadhyay, Vinay P Namboodiri, and CV Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In Proceedings of the 28th ACM International Conference on Multimedia, pages 484-492, 2020. +[28] Amit Raj, Michael Zollhoefer, Tomas Simon, Jason Saragih, Shunsuke Saito, James Hays, and Stephen Lombardi. Pva: Pixel-aligned volumetric avatars. arXiv preprint arXiv:2101.02697, 2021. +[29] Yurui Ren, Ge Li, Yuanqi Chen, Thomas H Li, and Shan Liu. Pirenderer: Controllable portrait image generation via semantic neural rendering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13759-13768, 2021. +[30] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. Graf: Generative radiance fields for 3d-aware image synthesis. Advances in Neural Information Processing Systems, 33:20154-20166, 2020. +[31] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. First order motion model for image animation. Advances in Neural Information Processing Systems, 32, 2019. +[32] Aliaksandr Siarohin, Oliver J Woodford, Jian Ren, Menglei Chai, and Sergey Tulyakov. Motion representations for articulated animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13653-13662, 2021. +[33] Ivan Skorokhodov, Sergey Tulyakov, and Mohamed Elhoseiny. Stylegan-v: A continuous video generator with the price, image quality and perks of stylegan2. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3626-3636, 2022. +[34] Justus Thies, Michael Zollhofer, Marc Stamminger, Christian Theobalt, and Matthias Nießner. Face2face: Real-time face capture and reenactment of rgb videos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2387-2395, 2016. +[35] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Guilin Liu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. Video-to-video synthesis. In Advances in Neural Information Processing Systems (NeurIPS), 2018. +[36] Yaohui Wang, Di Yang, Francois Bremond, and Antitza Dantcheva. Latent image animator: Learning to animate images via latent space navigation. arXiv preprint arXiv:2203.09043, 2022. +[37] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. +[38] Cheng-hsin Wu, Ningyuan Zheng, Scott Ardisson, Rohan Bali, Danielle Belko, Eric Brockmeyer, Lucas Evans, Timothy Godisart, Hyowon Ha, Alexander Hypes, Taylor Koska, Steven Krenn, Stephen Lombardi, Xiaomin Luo, Kevyn McPhail, Laura Millerschoen, Michal Perdoch, Mark Pitts, Alexander Richard, Jason Saragih, Junko Saragih, Takaaki Shiratori, Tomas Simon, Matt Stewart, Autumn Trimble, Xinshuo Weng, David Whitewolf, Chenglei Wu, Shouu-I Yu, + +and Yaser Sheikh. Multiface: A dataset for neural face rendering. In arXiv, 2022. +[39] Tao Yang, Peiran Ren, Xuansong Xie, and Lei Zhang. Gan prior embedded network for blind face restoration in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 672-681, June 2021. +[40] Fei Yin, Yong Zhang, Xiaodong Cun, Mingdeng Cao, Yanbo Fan, Xuan Wang, Qingyan Bai, Baoyuan Wu, Jue Wang, and Yujiu Yang. Styleheat: One-shot high-resolution editable talking face generation via pretrained stylegan. arXiv preprint arXiv:2203.04036, 2022. +[41] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4578-4587, 2021. +[42] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. +[43] Zhimeng Zhang, Lincheng Li, Yu Ding, and Changjie Fan. Flow-guided one-shot talking face generation with a high-resolution audio-visual dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3661–3670, 2021. +[44] Zicheng Zhang, Yinglu Liu, Congying Han, Tiande Guo, Ting Yao, and Tao Mei. Generalized one-shot domain adaption of generative adversarial networks. arXiv preprint arXiv:2209.03665, 2022. +[45] Yufeng Zheng, Victoria Fernández Abrevaya, Marcel C. Bühler, Xu Chen, Michael J. Black, and Otmar Hilliges. I M Avatar: Implicit morphable head avatars from videos. In Computer Vision and Pattern Recognition (CVPR), 2022. +[46] Hang Zhou, Yu Liu, Ziwei Liu, Ping Luo, and Xiaogang Wang. Talking face generation by adversarially disentangled audio-visual representation. In AAAI Conference on Artificial Intelligence (AAAI), 2019. +[47] Hang Zhou, Yasheng Sun, Wayne Wu, Chen Change Loy, Xiaogang Wang, and Ziwei Liu. Pose-controllable talking face generation by implicitly modularized audio-visual representation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. \ No newline at end of file diff --git a/2023/OTAvatar_ One-Shot Talking Face Avatar With Controllable Tri-Plane Rendering/images.zip b/2023/OTAvatar_ One-Shot Talking Face Avatar With Controllable Tri-Plane Rendering/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f773bde18f496a6d9d77b3a23c45361b8aea9e15 --- /dev/null +++ b/2023/OTAvatar_ One-Shot Talking Face Avatar With Controllable Tri-Plane Rendering/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a48923c1371046c7f4d7d805331a776ed7666aab14b74b8cce9476939f010ea0 +size 553159 diff --git a/2023/OTAvatar_ One-Shot Talking Face Avatar With Controllable Tri-Plane Rendering/layout.json b/2023/OTAvatar_ One-Shot Talking Face Avatar With Controllable Tri-Plane Rendering/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..fe5f7a9b64a90519a427f76d3433cb28cb789f84 --- /dev/null +++ b/2023/OTAvatar_ One-Shot Talking Face Avatar With Controllable Tri-Plane Rendering/layout.json @@ -0,0 +1,9052 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 48, + 103, + 545, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 103, + 545, + 122 + ], + "spans": [ + { + "bbox": [ + 48, + 103, + 545, + 122 + ], + "type": "text", + "content": "OTAvatar: One-shot Talking Face Avatar with Controllable Tri-plane Rendering" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 85, + 141, + 164, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 141, + 164, + 157 + ], + "spans": [ + { + "bbox": [ + 85, + 141, + 164, + 157 + ], + "type": "text", + "content": "Zhiyuan Ma" + }, + { + "bbox": [ + 85, + 141, + 164, + 157 + ], + "type": "inline_equation", + "content": "^{1,2*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 184, + 142, + 258, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 142, + 258, + 157 + ], + "spans": [ + { + "bbox": [ + 184, + 142, + 258, + 157 + ], + "type": "text", + "content": "Xiangyu Zhu" + }, + { + "bbox": [ + 184, + 142, + 258, + 157 + ], + "type": "inline_equation", + "content": "^{3*}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 276, + 143, + 334, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 276, + 143, + 334, + 157 + ], + "spans": [ + { + "bbox": [ + 276, + 143, + 334, + 157 + ], + "type": "text", + "content": "Guojun Qi" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 358, + 143, + 426, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 143, + 426, + 157 + ], + "spans": [ + { + "bbox": [ + 358, + 143, + 426, + 157 + ], + "type": "text", + "content": "Zhen Lei" + }, + { + "bbox": [ + 358, + 143, + 426, + 157 + ], + "type": "inline_equation", + "content": "^{1,2,3\\dagger}" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 446, + 143, + 503, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 446, + 143, + 503, + 157 + ], + "spans": [ + { + "bbox": [ + 446, + 143, + 503, + 157 + ], + "type": "text", + "content": "Lei Zhang" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 199, + 158, + 395, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 199, + 158, + 395, + 171 + ], + "spans": [ + { + "bbox": [ + 199, + 158, + 395, + 171 + ], + "type": "text", + "content": "1The Hong Kong Polytechnic University" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 151, + 171, + 442, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 171, + 442, + 185 + ], + "spans": [ + { + "bbox": [ + 151, + 171, + 442, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 151, + 171, + 442, + 185 + ], + "type": "text", + "content": "Center for Artificial Intelligence and Robotics, HKISI CAS" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 111, + 186, + 482, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 186, + 482, + 199 + ], + "spans": [ + { + "bbox": [ + 111, + 186, + 482, + 199 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 111, + 186, + 482, + 199 + ], + "type": "text", + "content": "State Key Laboratory of Multimodal Artificial Intelligence Systems, CASIA" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 255, + 200, + 339, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 200, + 339, + 212 + ], + "spans": [ + { + "bbox": [ + 255, + 200, + 339, + 212 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 255, + 200, + 339, + 212 + ], + "type": "text", + "content": "OPPO Research" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 155, + 216, + 444, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 216, + 444, + 227 + ], + "spans": [ + { + "bbox": [ + 155, + 216, + 444, + 227 + ], + "type": "text", + "content": "zm2354.ma@connect.polyu.hk, xiangyu.zhu@nlpr.ia.ac.cn" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 118, + 230, + 470, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 230, + 470, + 241 + ], + "spans": [ + { + "bbox": [ + 118, + 230, + 470, + 241 + ], + "type": "text", + "content": "guojunq@gmail.com, zlei@nlpr.ia.ac.cn, cslzhang@comp.polyu.edu.hk" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 143, + 269, + 192, + 281 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 269, + 192, + 281 + ], + "spans": [ + { + "bbox": [ + 143, + 269, + 192, + 281 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 294, + 290, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 294, + 290, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 294, + 290, + 605 + ], + "type": "text", + "content": "Controllability, generalizability and efficiency are the major objectives of constructing face avatars represented by neural implicit field. However, existing methods have not managed to accommodate the three requirements simultaneously. They either focus on static portraits, restricting the representation ability to a specific subject, or suffer from substantial computational cost, limiting their flexibility. In this paper, we propose One-shot Talking face Avatar (OTAvatar), which constructs face avatars by a generalized controllable tri-plane rendering solution so that each personalized avatar can be constructed from only one portrait as the reference. Specifically, OTAvatar first inverts a portrait image to a motion-free identity code. Second, the identity code and a motion code are utilized to modulate an efficient CNN to generate a tri-plane formulated volume, which encodes the subject in the desired motion. Finally, volume rendering is employed to generate an image in any view. The core of our solution is a novel decoupling-by-inverting strategy that disentangles identity and motion in the latent code via optimization-based inversion. Benefiting from the efficient tri-plane representation, we achieve controllable rendering of generalized face avatar at 35 FPS on A100. Experiments show promising performance of cross-identity reenactment on subjects out of the training set and better 3D consistency. The code is available at https://github.com/theEricMa/OTAvatar." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 47, + 628, + 128, + 640 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 628, + 128, + 640 + ], + "spans": [ + { + "bbox": [ + 47, + 628, + 128, + 640 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 47, + 648, + 287, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 648, + 287, + 685 + ], + "spans": [ + { + "bbox": [ + 47, + 648, + 287, + 685 + ], + "type": "text", + "content": "Neural rendering has achieved remarkable progress and promising results in 3D reconstruction. Thanks to the differentiability in neuron computation, the neural rendering" + } + ] + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 312, + 272, + 545, + 473 + ], + "blocks": [ + { + "bbox": [ + 312, + 272, + 545, + 473 + ], + "lines": [ + { + "bbox": [ + 312, + 272, + 545, + 473 + ], + "spans": [ + { + "bbox": [ + 312, + 272, + 545, + 473 + ], + "type": "image", + "image_path": "374d618906637c283a736c98874ae17ab80e20c09c7399c9b680ad80e2949c8a.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 483, + 547, + 540 + ], + "lines": [ + { + "bbox": [ + 305, + 483, + 547, + 540 + ], + "spans": [ + { + "bbox": [ + 305, + 483, + 547, + 540 + ], + "type": "text", + "content": "Figure 1. OTAvatar animation results. The source subjects in HDTF [43] dataset are animated by OTAvatar using a single portrait as the reference. We use the pose and expression coefficients of 3DMM to represent motion and drive the avatar. Note that these subjects are not included in the training data of OTAvatar." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 566, + 547, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 566, + 547, + 662 + ], + "spans": [ + { + "bbox": [ + 304, + 566, + 547, + 662 + ], + "type": "text", + "content": "methods bypass the expensive cost of high-fidelity digital 3D modeling, thus attracting the attention of many researchers from academia and industry. In this work, we aim to generate a talking face avatar via neural rendering techniques, which is controllable by driving signals like video and audio segments. Such animation is expected to inherit the identity from reference portraits, while the expression and pose can keep in sync with the driving signals." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 665, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 665, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 665, + 547, + 714 + ], + "type": "text", + "content": "The early works about talking face focus on expression animation consistent with driving signals in constrained frontal face images [5, 7, 27]. It is then extended to in-the-wild scenes with pose variations [18, 35, 46, 47]. Many" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 693, + 126, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 693, + 126, + 703 + ], + "spans": [ + { + "bbox": [ + 58, + 693, + 126, + 703 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 703, + 135, + 712 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16901" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": "previous works are 2D methods, where the image warping adopted to stimulate the motion in 3D space is learned in 2D space [29, 31, 32, 36, 40]. These methods tend to collapse under large pose variation. In contrast, there are 3D methods that can address the pose problem, since the head pose can be treated as a novel view. Both explicit [11, 20] and implicit 3D representations [23] are introduced to face rendering [15, 17, 21, 22, 25, 28, 45]. However, these methods either overfit to a single identity or fail to produce high-quality animation for different identities." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 192, + 289, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 192, + 289, + 357 + ], + "spans": [ + { + "bbox": [ + 46, + 192, + 289, + 357 + ], + "type": "text", + "content": "In this work, we propose a one-shot talking face avatar (OTAvatar), which can generate mimic expressions with good 3D consistency and be generalized to different identities with only one portrait reference. Some avatar animation examples are shown in Fig. 1. Given a single reference image, OTAvatar can drive the subject with motion signals to generate corresponding face images. We realize it under the framework of volume rendering. Current methods usually render static subjects [22, 28]. Although there are works [12, 15, 25, 26, 45] proposed to implement dynamic rendering for face avatars, they need one model per subject. Therefore, the generalization is poor. HeadNeRF [17] is a similar work to us. However, its reconstruction results are unnatural in the talking face case." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 359, + 289, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 359, + 289, + 526 + ], + "spans": [ + { + "bbox": [ + 46, + 359, + 289, + 526 + ], + "type": "text", + "content": "Our method is built on a 3D generative model pre-trained on a large-scale face database to guarantee identity generalization ability. Besides, we employ a motion controller to decouple the motion and identity in latent space when performing the optimization-based GAN inversion, to make the motion controllable and transferable to different identities. The network architecture is compact to ensure inference efficiency. In the utilization of our OTAvatar, given a single reference image, we fix the motion code predicted by the controller and only optimize the identity code so that a new avatar of the reference image can be constructed. Such a disentanglement enables the rendering of any desired motion by simply alternating the motion-related latent code to be that of the specific motion representation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 535, + 287, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 535, + 287, + 559 + ], + "spans": [ + { + "bbox": [ + 47, + 535, + 287, + 559 + ], + "type": "text", + "content": "The major contributions of this work can be summarized as follows." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 563, + 286, + 706 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 58, + 563, + 286, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 563, + 286, + 598 + ], + "spans": [ + { + "bbox": [ + 58, + 563, + 286, + 598 + ], + "type": "text", + "content": "- We make the first attempt for one-shot 3D face reconstruction and motion-controllable rendering by taming a pre-trained 3D generative model for motion control." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 605, + 286, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 605, + 286, + 653 + ], + "spans": [ + { + "bbox": [ + 58, + 605, + 286, + 653 + ], + "type": "text", + "content": "- We propose to decouple motion-related and motion-free latent code in inversion optimization by prompting the motion fraction of latent code ahead of the optimization using a decoupling-by-inverting strategy." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 58, + 659, + 286, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 659, + 286, + 706 + ], + "spans": [ + { + "bbox": [ + 58, + 659, + 286, + 706 + ], + "type": "text", + "content": "- Our method can photo-realistically render any identity with the desired expression and pose at 35FPS. The experiment shows promising results of natural motion and 3D consistency on both 2D and 3D datasets." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 71, + 397, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 397, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 397, + 84 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 91, + 389, + 104 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 91, + 389, + 104 + ], + "spans": [ + { + "bbox": [ + 306, + 91, + 389, + 104 + ], + "type": "text", + "content": "2.1. Talking Face" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 110, + 545, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 110, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 304, + 110, + 545, + 228 + ], + "type": "text", + "content": "Regarding the driving signal, the talking face methods can be roughly divided into three categories: audio-driven, image-driven and coefficients-driven. Our approach is most related to coefficients-driven methods, which use either facial landmarks or 3DMM coefficients to represent motion. Facial landmarks involve identity and expression information and is challenging to transfer motion across different subjects. 3DMM coefficients disentangle the identity, expression and pose, and thus they are good driven signals to drive different faces." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 230, + 546, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 230, + 546, + 506 + ], + "spans": [ + { + "bbox": [ + 304, + 230, + 546, + 506 + ], + "type": "text", + "content": "Thies et al. [34] used 3D rendering to maintain the shape and illumination attributes when transferring expression and refining mouth details with a mouth retrieval algorithm. Geng et al. [14] used 3DMM to render a given subject with different expressions, followed by a neural network to refine the detail and harmonize with the background. These methods cannot change the face pose. Ren et al. [29] mapped 3DMM expression and pose to a high-dimensional motion representation, then predicted the dense image flow to reenact faces via image warping. Doukas et al. [10] used additional 3DMM mesh fitting results to assist the dense image flow prediction. Yin et al. [40] integrated pre-trained StyleGAN [6] to generate high-resolution talking face prediction by warping low-resolution feature maps. These warping-based methods follow the two-stage workflow by first warping images and then refining facial detail, demonstrating superior performance for same identity reenactment. But in cross-identity reenactment, especially when there is a large pose variation against the source portrait, the animation tends to collapse, since image warping is merely trained to stimulate 3D motion in the 2D space. Our work animates face avatars using 3D rendering and explicitly ensures 3D consistency to handle large pose variations." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 515, + 417, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 515, + 417, + 528 + ], + "spans": [ + { + "bbox": [ + 306, + 515, + 417, + 528 + ], + "type": "text", + "content": "2.2. Volume Rendering" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 545, + 713 + ], + "type": "text", + "content": "Volume rendering is a 3D rendering strategy and has demonstrated its success in novel view synthesis. Existing methods either use volume [11, 21] to explicitly represent 3D space or implicitly store the 3D scene in the MLP network [23]. These methods are introduced to animate 3D-consistent portraits. Yu et al. [41] and Raj et al. [28] retrieved the spatial information from sparse support views to render portraits. On the other side, The 3D face generative model can reconstruct a given identity using GAN Inversion. These models vary from pure MLP architecture [4], to low-resolution neural rendering followed by an up-sampling strategy [24, 30], and to using more efficient 3D presentation [3]. All these methods only support rendering static portraits. To animate the avatar with controllable motion, AD-NeRF [15] and NeRFace [13] introduce" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16902" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 70, + 545, + 313 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 545, + 313 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 545, + 313 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 545, + 313 + ], + "type": "image", + "image_path": "bfcd6020e08ff84474dce29dfa2a4659e7f33557fd48d2e046a45ebe554931d4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 323, + 548, + 390 + ], + "lines": [ + { + "bbox": [ + 46, + 323, + 548, + 390 + ], + "spans": [ + { + "bbox": [ + 46, + 323, + 548, + 390 + ], + "type": "text", + "content": "Figure 2. Overview of OTAvatar. OTAvatar contains a 3D face animatior network and a latent code decoupling strategy, namely decoupling-by-inverting. The 3D face animatior is composed of a pre-trained 3D face generator [3] and a motion controller module. The decoupling-by-inverting algorithm is an optimization-based image inversion that can decouple the latent code " + }, + { + "bbox": [ + 46, + 323, + 548, + 390 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 46, + 323, + 548, + 390 + ], + "type": "text", + "content": " into identity code " + }, + { + "bbox": [ + 46, + 323, + 548, + 390 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_{id}" + }, + { + "bbox": [ + 46, + 323, + 548, + 390 + ], + "type": "text", + "content": " and motion code " + }, + { + "bbox": [ + 46, + 323, + 548, + 390 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_x" + }, + { + "bbox": [ + 46, + 323, + 548, + 390 + ], + "type": "text", + "content": ". When the model is well-trained, the motion-free identity code can be inferred from a single reference image, and an avatar of an unseen subject can be constructed. The identity code can be integrated with any other motion code predicted by the controller to animate the identity with desired motion." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 411, + 289, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 411, + 289, + 603 + ], + "spans": [ + { + "bbox": [ + 46, + 411, + 289, + 603 + ], + "type": "text", + "content": "either the audio feature or 3DMM expression coefficients to the NeRF model to render dynamic heads. Nerfies [25] employs a deformation field conditioned on spatial points to represent the face motion of different frames. HyperNeRF [26] adds an ambient slicing network to tackle the topological drawback of Nerfies. Though these methods have achieved motion control and view consistency for face avatars, they all overfit each NeRF model to one subject, thus lack generalizability. HeadNeRF [17] is trained on both multi-view and frontal face datasets and can be generalized to multiple identities and support motion control using 3DMM coefficients, but its rendering quality is not satisfactory compared with the given portrait, and the motion control is unnatural with random jitters. In contrast, our method can reconstruct photo-realistic face avatars in one shot and animate natural talking motion." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 610, + 149, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 610, + 149, + 623 + ], + "spans": [ + { + "bbox": [ + 47, + 610, + 149, + 623 + ], + "type": "text", + "content": "2.3. Generative Prior" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "content": "Many methods rely on pre-trained generative models. Employing a generative model as the prior has the following advantages. First, it provides rich and diverse facial priors, such as texture, color, shape and pose, which can help to restore realistic and faithful facial details. For instance, Yang et al. [39] leveraged a pre-trained GAN model as a prior decoder for blind face restoration. Second, it also enables one-" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 411, + 547, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 411, + 547, + 711 + ], + "spans": [ + { + "bbox": [ + 304, + 411, + 547, + 711 + ], + "type": "text", + "content": "shot image manipulation both inside and outside of the domain. Well-designed schemes for disentanglement are necessary to provide acceptable performance on fine-grained out-of-domain image manipulation. Zhang et al. [44] suggested separating the attributes into global style—like texture and color—and structure—like shape and pose, and performing domain-adaptive generation by transferring the decoupled style attributes. The image manipulation also includes synthesizing human talking videos, Ivan et al. [33] modeled the temporal dynamics of video frames through the latent code disentanglement. Unlike other talking head generation methods [31, 40], it only synthesizes center-aligned faces. The recently presented 3D GANs have been popularized for their ability to generate 3D-consistent photorealistic human faces with explicit pose control. Katja et.al [30] introduced 3D scene volume rendering without resorting to computationally demanding voxel-based representation, and the model is trained from unposed 2D photos in an adversarial manner. Chan et al. [3] proposed to generate high-quality geometry and multi-view consistent images from 2D photos through a hybrid explicit-implicit network architecture based on StyleGAN2 [19]. Despite the fact that 3D GANs' explicit pose control and 3D-consistent generation are appealing for talking face synthesis, no previous work has studied their application to talking face avatars." + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16903" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 71, + 104, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 71, + 104, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 104, + 83 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 91, + 289, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 91, + 289, + 236 + ], + "spans": [ + { + "bbox": [ + 46, + 91, + 289, + 236 + ], + "type": "text", + "content": "Talking face avatar aims to synthesize face images with controllable expression and pose. In this paper, we investigate the possibility of endowing volume rendering with the ability to 1) build a faithful identity representation in one shot, 2) enable natural motion control, and 3) achieve real-time inference speed. Our framework takes an identity code and a motion signal as input and generates a tri-plane-based [3] volume through a CNN architecture. By performing volume rendering on the tri-plane representation, where each point is projected on the three feature planes and the sampled features are fused to occupancy and color, a face image at a given camera view can be generated as:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 107, + 249, + 287, + 262 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 249, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 107, + 249, + 287, + 262 + ], + "type": "interline_equation", + "content": "\\mathbf {I} (\\mathbf {p}) = \\mathcal {R} (G (\\mathbf {w} _ {i d}, \\mathbf {x}; \\Theta), \\mathbf {p}), \\tag {1}", + "image_path": "92ba2a56cccfc2bbcf67d607162e4fc1913ae9adcfc587ce2a282d9847691992.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 269, + 289, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 269, + 289, + 473 + ], + "spans": [ + { + "bbox": [ + 46, + 269, + 289, + 473 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 269, + 289, + 473 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_{id}" + }, + { + "bbox": [ + 46, + 269, + 289, + 473 + ], + "type": "text", + "content": " is the identity code, " + }, + { + "bbox": [ + 46, + 269, + 289, + 473 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 46, + 269, + 289, + 473 + ], + "type": "text", + "content": " is the motion signal, " + }, + { + "bbox": [ + 46, + 269, + 289, + 473 + ], + "type": "inline_equation", + "content": "G(\\cdot ,\\cdot ;\\Theta)" + }, + { + "bbox": [ + 46, + 269, + 289, + 473 + ], + "type": "text", + "content": " is a face volume generation network with " + }, + { + "bbox": [ + 46, + 269, + 289, + 473 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 46, + 269, + 289, + 473 + ], + "type": "text", + "content": " as its network weights to reconstruct an encoded portrait in three orthogonal feature planes, " + }, + { + "bbox": [ + 46, + 269, + 289, + 473 + ], + "type": "inline_equation", + "content": "\\mathcal{R}(\\mathbf{V},\\mathbf{p})" + }, + { + "bbox": [ + 46, + 269, + 289, + 473 + ], + "type": "text", + "content": " is the volume rendering with " + }, + { + "bbox": [ + 46, + 269, + 289, + 473 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 46, + 269, + 289, + 473 + ], + "type": "text", + "content": " as the volume and " + }, + { + "bbox": [ + 46, + 269, + 289, + 473 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 46, + 269, + 289, + 473 + ], + "type": "text", + "content": " as the camera view, and " + }, + { + "bbox": [ + 46, + 269, + 289, + 473 + ], + "type": "inline_equation", + "content": "\\mathbf{I}(\\mathbf{p})\\in \\mathbb{R}^{3\\times H\\times W}" + }, + { + "bbox": [ + 46, + 269, + 289, + 473 + ], + "type": "text", + "content": " is the rendered result. In our implementation, the identity code " + }, + { + "bbox": [ + 46, + 269, + 289, + 473 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_{id}" + }, + { + "bbox": [ + 46, + 269, + 289, + 473 + ], + "type": "text", + "content": " can be extracted by optimizing on a single reference image using our proposed decoupling-by-inverting strategy (Sec. 3.3), achieving one-shot avatar reconstruction. The motion signal " + }, + { + "bbox": [ + 46, + 269, + 289, + 473 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 46, + 269, + 289, + 473 + ], + "type": "text", + "content": " is 3DMM coefficients for flexible control. We call " + }, + { + "bbox": [ + 46, + 269, + 289, + 473 + ], + "type": "inline_equation", + "content": "G(\\cdot ,\\cdot ;\\Theta)" + }, + { + "bbox": [ + 46, + 269, + 289, + 473 + ], + "type": "text", + "content": " as 3D Animator because of its ability to animate an avatar with desired motion. Compared with 2D methods, we construct a neural 3D space to promise multi-view consistency. Compared with NeRF, we build a 3D space in the propagation of a CNN, and directly sample features on the tri-plane representation, saving the MLP computation on each sampled point." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 481, + 224, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 481, + 224, + 493 + ], + "spans": [ + { + "bbox": [ + 47, + 481, + 224, + 493 + ], + "type": "text", + "content": "3.1. 3D Animator Network Structure" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 500, + 287, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 500, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 46, + 500, + 287, + 536 + ], + "type": "text", + "content": "Tri-plane volume representation. The output of the 3D anemotor network " + }, + { + "bbox": [ + 46, + 500, + 287, + 536 + ], + "type": "inline_equation", + "content": "G(\\cdot, \\cdot; \\Theta)" + }, + { + "bbox": [ + 46, + 500, + 287, + 536 + ], + "type": "text", + "content": " in Eqn. 1 is a tri-plane volume representation, which is composed of three feature planes:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 82, + 546, + 287, + 561 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 546, + 287, + 561 + ], + "spans": [ + { + "bbox": [ + 82, + 546, + 287, + 561 + ], + "type": "interline_equation", + "content": "\\mathbf {V} _ {t r i} = (\\mathbf {F} _ {x y}, \\mathbf {F} _ {x z}, \\mathbf {F} _ {y z}) = G (\\mathbf {w} _ {i d}, \\mathbf {x}; \\boldsymbol {\\Theta}), \\qquad (2)", + "image_path": "549e71135f3797ac5d11d1c3128086af27da2243b8a459f36a0bb117c8a03242.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{xy}" + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{xz}" + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{yz}" + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "content": " are three axis-aligned orthogonal feature maps in the 3D space, implicitly forming the volume " + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_{tri}" + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "content": ". When performing volume rendering, for each queried point " + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "inline_equation", + "content": "(x,y,z)" + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "content": ", we project it onto each of three feature maps and retrieve the corresponding features " + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "inline_equation", + "content": "(\\mathbf{F}_{xy}(x,y),\\mathbf{F}_{xz}(x,z),\\mathbf{F}_{yz}(y,z))" + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "content": ". The summation of features is sent to a lightweight MLP to decode color and opacity. Compared to fully implicit MLP architectures [4,15,45] or volume representation [21], we can efficiently regress the three 2D feature maps through a CNN architecture. In our implementation, " + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "content": " is a deconvolutional network [6] that outputs three " + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "inline_equation", + "content": "256\\times 256\\times 32" + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "content": " feature maps." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 215 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 215 + ], + "type": "text", + "content": "Animator structure. Towards animating the face avatar of a given portrait, an intuitive solution is representing " + }, + { + "bbox": [ + 304, + 72, + 545, + 215 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_{id}" + }, + { + "bbox": [ + 304, + 72, + 545, + 215 + ], + "type": "text", + "content": " with a pre-defined feature and training " + }, + { + "bbox": [ + 304, + 72, + 545, + 215 + ], + "type": "inline_equation", + "content": "G(\\cdot)" + }, + { + "bbox": [ + 304, + 72, + 545, + 215 + ], + "type": "text", + "content": " from scratch. However, we find that the model does not work well due to the sub-optimal " + }, + { + "bbox": [ + 304, + 72, + 545, + 215 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_{id}" + }, + { + "bbox": [ + 304, + 72, + 545, + 215 + ], + "type": "text", + "content": ". For example, when employing 3DMM shape coefficients [9] and face recognition features [8], the identity information is not preserved in the rendering results because " + }, + { + "bbox": [ + 304, + 72, + 545, + 215 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_{id}" + }, + { + "bbox": [ + 304, + 72, + 545, + 215 + ], + "type": "text", + "content": " does not encode face appearances well. These models are trained on cropped facial images which contain no torso and hairstyle. Therefore, " + }, + { + "bbox": [ + 304, + 72, + 545, + 215 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_{id}" + }, + { + "bbox": [ + 304, + 72, + 545, + 215 + ], + "type": "text", + "content": " losses such information that is necessary when animating portraits. The network even collapses if we make " + }, + { + "bbox": [ + 304, + 72, + 545, + 215 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_{id}" + }, + { + "bbox": [ + 304, + 72, + 545, + 215 + ], + "type": "text", + "content": " learnable." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 216, + 545, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 216, + 545, + 287 + ], + "spans": [ + { + "bbox": [ + 304, + 216, + 545, + 287 + ], + "type": "text", + "content": "In this work, we employ a two-phase strategy to achieve one-shot avatar reconstruction: 1) building a 3D face generator, and 2) making the generator controllable. We build our 3D face generator on a pre-trained EG3D [3] network, which incorporates tri-plane representation for efficient 3D face generation:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 380, + 295, + 545, + 308 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 295, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 380, + 295, + 545, + 308 + ], + "type": "interline_equation", + "content": "\\mathbf {V} _ {t r i} = G _ {e g} (\\mathbf {w}; \\Theta_ {e g}), \\tag {3}", + "image_path": "be35c58b1094daeae268563cd989e080ff4a9c7b5dbaae43854a488928a2fb7e.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 315, + 545, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 315, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 304, + 315, + 545, + 399 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 315, + 545, + 399 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 304, + 315, + 545, + 399 + ], + "type": "text", + "content": " is an uninterpretable latent code drawn from the latent space of the generator " + }, + { + "bbox": [ + 304, + 315, + 545, + 399 + ], + "type": "inline_equation", + "content": "G_{eg}" + }, + { + "bbox": [ + 304, + 315, + 545, + 399 + ], + "type": "text", + "content": ". The latent space is determined by the generator weight " + }, + { + "bbox": [ + 304, + 315, + 545, + 399 + ], + "type": "inline_equation", + "content": "\\Theta_{eg}" + }, + { + "bbox": [ + 304, + 315, + 545, + 399 + ], + "type": "text", + "content": ". To control the latent code " + }, + { + "bbox": [ + 304, + 315, + 545, + 399 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 304, + 315, + 545, + 399 + ], + "type": "text", + "content": " by the given identity represented by " + }, + { + "bbox": [ + 304, + 315, + 545, + 399 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_{id}" + }, + { + "bbox": [ + 304, + 315, + 545, + 399 + ], + "type": "text", + "content": " and motion signal " + }, + { + "bbox": [ + 304, + 315, + 545, + 399 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 315, + 545, + 399 + ], + "type": "text", + "content": ", we propose a motion controller module " + }, + { + "bbox": [ + 304, + 315, + 545, + 399 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 315, + 545, + 399 + ], + "type": "text", + "content": " parameterized by " + }, + { + "bbox": [ + 304, + 315, + 545, + 399 + ], + "type": "inline_equation", + "content": "\\Theta_c" + }, + { + "bbox": [ + 304, + 315, + 545, + 399 + ], + "type": "text", + "content": ", which maps motion signal " + }, + { + "bbox": [ + 304, + 315, + 545, + 399 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 315, + 545, + 399 + ], + "type": "text", + "content": " to the motion code " + }, + { + "bbox": [ + 304, + 315, + 545, + 399 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_x" + }, + { + "bbox": [ + 304, + 315, + 545, + 399 + ], + "type": "text", + "content": ", such that:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 352, + 407, + 545, + 419 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 352, + 407, + 545, + 419 + ], + "spans": [ + { + "bbox": [ + 352, + 407, + 545, + 419 + ], + "type": "interline_equation", + "content": "\\mathbf {w} = \\mathbf {w} _ {i d} + \\mathbf {w} _ {x} = \\mathbf {w} _ {i d} + C (\\mathbf {x}; \\Theta_ {c}). \\tag {4}", + "image_path": "d8a80e2fa7de4722b1b3e35a4dd22fcd298fdf298ef1a5159a849eaecce7cc5b.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 426, + 545, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 545, + 449 + ], + "type": "text", + "content": "Injecting Eqn. 4 to Eqn. 3, we have the structure of the 3D animatior:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 334, + 455, + 545, + 484 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 455, + 545, + 484 + ], + "spans": [ + { + "bbox": [ + 334, + 455, + 545, + 484 + ], + "type": "interline_equation", + "content": "\\begin{array}{c} G (\\mathbf {w} _ {i d}, \\mathbf {x}; \\Theta) = G _ {e g} (\\mathbf {w} _ {i d} + C (\\mathbf {x}; \\Theta_ {c}); \\Theta_ {e g}), \\\\ \\Theta = \\Theta_ {e g} \\cup \\Theta_ {c}. \\end{array} \\tag {5}", + "image_path": "78d7eea58a69f63f147569875e1a93934ca60f7a0aa896684975873bc03a7702.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 491, + 545, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 491, + 545, + 538 + ], + "spans": [ + { + "bbox": [ + 304, + 491, + 545, + 538 + ], + "type": "text", + "content": "In our implementation, the motion signal " + }, + { + "bbox": [ + 304, + 491, + 545, + 538 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 491, + 545, + 538 + ], + "type": "text", + "content": " is constructed by 3DMM pose and expression coefficients [9], which are more flexible compared with image-based driving signals [31, 32, 36], especially in cross-identity driving." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 538, + 545, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 538, + 545, + 693 + ], + "spans": [ + { + "bbox": [ + 304, + 538, + 545, + 693 + ], + "type": "text", + "content": "Controller structure. We use the concatenation of 3DMM pose and expression coefficients to represent motion " + }, + { + "bbox": [ + 304, + 538, + 545, + 693 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 538, + 545, + 693 + ], + "type": "text", + "content": ". Practically, we use the adjacent coefficients of pose and expression within a particular radius to represent the motion signal of the current frame, firstly the weighted summation of the coefficients across the temporal dimension is conducted to avoid noises, which is achieved by three 1D convolution layers [29, 40]; secondly, a five-layer MLP is employed to transform the weighted summation into a motion feature; finally, a codebook with learnable orthogonal bases [36] is built, on which each motion feature is projected to get the final motion code " + }, + { + "bbox": [ + 304, + 538, + 545, + 693 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_x" + }, + { + "bbox": [ + 304, + 538, + 545, + 693 + ], + "type": "text", + "content": ". The aforementioned operation is summarized as follows:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 318, + 700, + 545, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 700, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 318, + 700, + 545, + 715 + ], + "type": "interline_equation", + "content": "\\mathbf {w} _ {x} = C (\\mathbf {x}; \\Theta_ {c}) = \\mathbf {D} * F _ {M} \\left(F _ {T} (\\mathbf {x}; \\Theta_ {T}); \\Theta_ {M}\\right), \\tag {6}", + "image_path": "29514a4491190649bb8e2d8c145c36e123d57dcc374e76b0b4d12e02f78dc47a.jpg" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16904" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "F_{T}" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": " parameterized by " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "\\Theta_T" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": " is the temporal smoothing network of 1D convolution layers, " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "F_{M}" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": " parameterized by " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "\\Theta_M" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": " is the five-layer MLP which projects motion to the magnitudes of bases contained in the codebook " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "\\mathbf{D}" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": ". It can be seen that the magnitude projection parameters, temporal smoothing weights, and the codebook are the parameters of the controller " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": " to be trained, that is:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 120, + 164, + 287, + 178 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 164, + 287, + 178 + ], + "spans": [ + { + "bbox": [ + 120, + 164, + 287, + 178 + ], + "type": "interline_equation", + "content": "\\Theta_ {c} = \\mathbf {D} \\cup \\Theta_ {T} \\cup \\Theta_ {M}. \\tag {7}", + "image_path": "5573d895d755639788e25e8fe37c251056cbf4ce49948d152a204444e74e7833.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 184, + 163, + 197 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 184, + 163, + 197 + ], + "spans": [ + { + "bbox": [ + 47, + 184, + 163, + 197 + ], + "type": "text", + "content": "3.2. Controller Training" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 202, + 287, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 202, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 47, + 202, + 287, + 239 + ], + "type": "text", + "content": "Given a couple of source frame " + }, + { + "bbox": [ + 47, + 202, + 287, + 239 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_s" + }, + { + "bbox": [ + 47, + 202, + 287, + 239 + ], + "type": "text", + "content": " and driving frame " + }, + { + "bbox": [ + 47, + 202, + 287, + 239 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_d" + }, + { + "bbox": [ + 47, + 202, + 287, + 239 + ], + "type": "text", + "content": ", along with their motion signals " + }, + { + "bbox": [ + 47, + 202, + 287, + 239 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_s, \\mathbf{x}_d" + }, + { + "bbox": [ + 47, + 202, + 287, + 239 + ], + "type": "text", + "content": " and camera views " + }, + { + "bbox": [ + 47, + 202, + 287, + 239 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_s, \\mathbf{p}_d" + }, + { + "bbox": [ + 47, + 202, + 287, + 239 + ], + "type": "text", + "content": ", we perform the dual-objective optimization:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 110, + 246, + 220, + 266 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 246, + 220, + 266 + ], + "spans": [ + { + "bbox": [ + 110, + 246, + 220, + 266 + ], + "type": "interline_equation", + "content": "\\mathbf{w}_{id}^{*} = \\operatorname *{arg min}_{\\mathbf{w}_{id}}(\\mathcal{L}_{s} + \\mathcal{L}_{d}),", + "image_path": "fd36da4811d3135c77d1c8aff8bd3ea993afa861f1385c2b2b54b42049ff98bc.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 66, + 268, + 287, + 282 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 268, + 287, + 282 + ], + "spans": [ + { + "bbox": [ + 66, + 268, + 287, + 282 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s} = \\mathcal {L} \\left(I _ {s}, \\hat {I} _ {s}\\right) = \\mathcal {L} \\left(I _ {s}, \\mathcal {R} \\left(G \\left(\\mathbf {w} _ {i d}, \\mathbf {x} _ {s}; \\Theta\\right), \\mathbf {p} _ {s}\\right), \\right. \\tag {8}", + "image_path": "f5bbef868d4b0e1fb9bbd6dd8f43b39f020c0b2a8a705e2c446ae83a68d6e818.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 65, + 286, + 258, + 300 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 286, + 258, + 300 + ], + "spans": [ + { + "bbox": [ + 65, + 286, + 258, + 300 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {d} = \\mathcal {L} (I _ {d}, \\hat {I} _ {d}) = \\mathcal {L} (I _ {d}, \\mathcal {R} (G (\\mathbf {w} _ {i d}, \\mathbf {x} _ {d}; \\Theta), \\mathbf {p} _ {d}).", + "image_path": "9882ae8b94bf7cb8d19851ed7bde63b7dd6defbd51487d4201948b2b6716598a.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 307, + 287, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 307, + 287, + 390 + ], + "spans": [ + { + "bbox": [ + 46, + 307, + 287, + 390 + ], + "type": "text", + "content": "It is implemented to estimate the identity code " + }, + { + "bbox": [ + 46, + 307, + 287, + 390 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_{id}" + }, + { + "bbox": [ + 46, + 307, + 287, + 390 + ], + "type": "text", + "content": " shared between " + }, + { + "bbox": [ + 46, + 307, + 287, + 390 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_s" + }, + { + "bbox": [ + 46, + 307, + 287, + 390 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 307, + 287, + 390 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_d" + }, + { + "bbox": [ + 46, + 307, + 287, + 390 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 307, + 287, + 390 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 46, + 307, + 287, + 390 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 307, + 287, + 390 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 46, + 307, + 287, + 390 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 307, + 287, + 390 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 46, + 307, + 287, + 390 + ], + "type": "text", + "content": " indicate 3D face generator, volume rendering, and loss function, respectively. The 3D face generator " + }, + { + "bbox": [ + 46, + 307, + 287, + 390 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 46, + 307, + 287, + 390 + ], + "type": "text", + "content": " is parameterized by " + }, + { + "bbox": [ + 46, + 307, + 287, + 390 + ], + "type": "inline_equation", + "content": "\\Theta = \\Theta_{eg} \\cup \\Theta_c" + }, + { + "bbox": [ + 46, + 307, + 287, + 390 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 307, + 287, + 390 + ], + "type": "inline_equation", + "content": "\\Theta_{eg}" + }, + { + "bbox": [ + 46, + 307, + 287, + 390 + ], + "type": "text", + "content": " includes the parameters of the pre-trained 3D face generator, and " + }, + { + "bbox": [ + 46, + 307, + 287, + 390 + ], + "type": "inline_equation", + "content": "\\Theta_c" + }, + { + "bbox": [ + 46, + 307, + 287, + 390 + ], + "type": "text", + "content": " includes the parameters of the controller which is the focus of our training scheme." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 390, + 287, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 390, + 287, + 582 + ], + "spans": [ + { + "bbox": [ + 46, + 390, + 287, + 582 + ], + "type": "text", + "content": "Training scheme. The purpose of controller training is to decouple the identity code and motion code from the latent code of the generator, making identity and motion replaceable in our generalized avatars. To this end, we propose the decoupling-by-inverting training strategy, which is generally an alternation of identity optimization and controller training at different steps. Specifically, we first freeze " + }, + { + "bbox": [ + 46, + 390, + 287, + 582 + ], + "type": "inline_equation", + "content": "\\Theta_c" + }, + { + "bbox": [ + 46, + 390, + 287, + 582 + ], + "type": "text", + "content": " and perform " + }, + { + "bbox": [ + 46, + 390, + 287, + 582 + ], + "type": "inline_equation", + "content": "N_{id}" + }, + { + "bbox": [ + 46, + 390, + 287, + 582 + ], + "type": "text", + "content": " steps of back-propagation on " + }, + { + "bbox": [ + 46, + 390, + 287, + 582 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_{id}" + }, + { + "bbox": [ + 46, + 390, + 287, + 582 + ], + "type": "text", + "content": ". Then we freeze " + }, + { + "bbox": [ + 46, + 390, + 287, + 582 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_{id}" + }, + { + "bbox": [ + 46, + 390, + 287, + 582 + ], + "type": "text", + "content": " and perform " + }, + { + "bbox": [ + 46, + 390, + 287, + 582 + ], + "type": "inline_equation", + "content": "N_{mo}" + }, + { + "bbox": [ + 46, + 390, + 287, + 582 + ], + "type": "text", + "content": " steps of backpropagation on " + }, + { + "bbox": [ + 46, + 390, + 287, + 582 + ], + "type": "inline_equation", + "content": "\\Theta_c" + }, + { + "bbox": [ + 46, + 390, + 287, + 582 + ], + "type": "text", + "content": " so that the controller learns to add the motion code on top of the identity code. In our implementation, " + }, + { + "bbox": [ + 46, + 390, + 287, + 582 + ], + "type": "inline_equation", + "content": "N_{id} = 90" + }, + { + "bbox": [ + 46, + 390, + 287, + 582 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 390, + 287, + 582 + ], + "type": "inline_equation", + "content": "N_{mo} = 10" + }, + { + "bbox": [ + 46, + 390, + 287, + 582 + ], + "type": "text", + "content": ". After " + }, + { + "bbox": [ + 46, + 390, + 287, + 582 + ], + "type": "inline_equation", + "content": "N = N_{id} + N_{mo}" + }, + { + "bbox": [ + 46, + 390, + 287, + 582 + ], + "type": "text", + "content": " steps, we slightly finetune " + }, + { + "bbox": [ + 46, + 390, + 287, + 582 + ], + "type": "inline_equation", + "content": "\\Theta_{eg}" + }, + { + "bbox": [ + 46, + 390, + 287, + 582 + ], + "type": "text", + "content": " with a learning rate of " + }, + { + "bbox": [ + 46, + 390, + 287, + 582 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 46, + 390, + 287, + 582 + ], + "type": "text", + "content": " to make the rendering fit the driving signal better. A brief pseudo-code of the whole scheme is provided in the supplementary material." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 589, + 211, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 589, + 211, + 600 + ], + "spans": [ + { + "bbox": [ + 47, + 589, + 211, + 600 + ], + "type": "text", + "content": "3.3. One-shot Avatar Construction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 607, + 287, + 655 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 607, + 287, + 655 + ], + "spans": [ + { + "bbox": [ + 46, + 607, + 287, + 655 + ], + "type": "text", + "content": "In the inference stage, given a single portrait image " + }, + { + "bbox": [ + 46, + 607, + 287, + 655 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_r" + }, + { + "bbox": [ + 46, + 607, + 287, + 655 + ], + "type": "text", + "content": ", its 3DMM coefficients " + }, + { + "bbox": [ + 46, + 607, + 287, + 655 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_r" + }, + { + "bbox": [ + 46, + 607, + 287, + 655 + ], + "type": "text", + "content": ", and camera view " + }, + { + "bbox": [ + 46, + 607, + 287, + 655 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_r" + }, + { + "bbox": [ + 46, + 607, + 287, + 655 + ], + "type": "text", + "content": " as the reference of avatar construction, we can estimate the identity code " + }, + { + "bbox": [ + 46, + 607, + 287, + 655 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_r" + }, + { + "bbox": [ + 46, + 607, + 287, + 655 + ], + "type": "text", + "content": " of the reference portrait via:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 84, + 663, + 287, + 683 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 663, + 287, + 683 + ], + "spans": [ + { + "bbox": [ + 84, + 663, + 287, + 683 + ], + "type": "interline_equation", + "content": "\\mathbf {w} _ {r} = \\underset {\\mathbf {w} _ {i d}} {\\arg \\min } \\mathcal {L} (\\mathbf {I} _ {r}, \\mathcal {R} (G (\\mathbf {w} _ {i d}, \\mathbf {x} _ {r}), \\mathbf {p} _ {r}). \\tag {9}", + "image_path": "d8eb46b9e10e706c6bbb7ef9a309088bbc5c826382c1d99be83cf2ba04546776.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "Afterwards, we can animate the identity with given driving motion " + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_d" + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": " and camera view " + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_d" + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": ", either from the same or" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 72, + 508, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 508, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 508, + 84 + ], + "type": "text", + "content": "different subject, through the following operation:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 350, + 92, + 545, + 106 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 92, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 350, + 92, + 545, + 106 + ], + "type": "interline_equation", + "content": "\\mathbf {I} \\left(\\mathbf {w} _ {r}, \\mathbf {x} _ {d}, \\mathbf {p} _ {d}\\right) = \\mathcal {R} \\left(G \\left(\\mathbf {w} _ {r}, \\mathbf {x} _ {d}\\right), \\mathbf {p} _ {d}\\right), \\tag {10}", + "image_path": "7eecd3ba238b6a8449be47528bd65c1533af4214d8552ff7724213f643cc4631.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 113, + 545, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 113, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 304, + 113, + 545, + 185 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 113, + 545, + 185 + ], + "type": "inline_equation", + "content": "\\mathbf{I}(\\mathbf{w}_r,\\mathbf{c}_d,\\mathbf{p}_d)" + }, + { + "bbox": [ + 304, + 113, + 545, + 185 + ], + "type": "text", + "content": " has the identity estimated as " + }, + { + "bbox": [ + 304, + 113, + 545, + 185 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_r" + }, + { + "bbox": [ + 304, + 113, + 545, + 185 + ], + "type": "text", + "content": ", and is animated with the motion " + }, + { + "bbox": [ + 304, + 113, + 545, + 185 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_d" + }, + { + "bbox": [ + 304, + 113, + 545, + 185 + ], + "type": "text", + "content": " and camera view " + }, + { + "bbox": [ + 304, + 113, + 545, + 185 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_d" + }, + { + "bbox": [ + 304, + 113, + 545, + 185 + ], + "type": "text", + "content": ". We name Eqn. 9 as the decoupling-by-inverting inference strategy since it can estimate decoupled identity code from latent code in the iterative optimization-based GAN Inversion, and can be used to render different motions." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 192, + 382, + 204 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 192, + 382, + 204 + ], + "spans": [ + { + "bbox": [ + 306, + 192, + 382, + 204 + ], + "type": "text", + "content": "3.4. Loss Terms" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 210, + 545, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 210, + 545, + 260 + ], + "spans": [ + { + "bbox": [ + 304, + 210, + 545, + 260 + ], + "type": "text", + "content": "The loss " + }, + { + "bbox": [ + 304, + 210, + 545, + 260 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(\\mathbf{I},\\hat{\\mathbf{I}})" + }, + { + "bbox": [ + 304, + 210, + 545, + 260 + ], + "type": "text", + "content": " in Eqn. 8 and Eqn. 9 measures the error between the animated result " + }, + { + "bbox": [ + 304, + 210, + 545, + 260 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{I}}" + }, + { + "bbox": [ + 304, + 210, + 545, + 260 + ], + "type": "text", + "content": " and the ground truth " + }, + { + "bbox": [ + 304, + 210, + 545, + 260 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 304, + 210, + 545, + 260 + ], + "type": "text", + "content": ". First, a pre-trained VGG-19 network is implemented to calculate the distance of multi-scale activation maps " + }, + { + "bbox": [ + 304, + 210, + 545, + 260 + ], + "type": "inline_equation", + "content": "\\phi_{i}(\\mathbf{I})" + }, + { + "bbox": [ + 304, + 210, + 545, + 260 + ], + "type": "text", + "content": " via:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 366, + 266, + 545, + 293 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 366, + 266, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 366, + 266, + 545, + 293 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c} = \\sum_ {i} \\left\\| \\phi_ {i} (\\mathbf {I}) - \\phi_ {i} (\\hat {\\mathbf {I}}) \\right\\| _ {1}, \\tag {11}", + "image_path": "e4d49b425b8e070d45da7267bfdf296db36a9716eed2aca00e4eaf6e246ef348.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 300, + 545, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 300, + 545, + 339 + ], + "spans": [ + { + "bbox": [ + 304, + 300, + 545, + 339 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 300, + 545, + 339 + ], + "type": "inline_equation", + "content": "\\phi_{i}" + }, + { + "bbox": [ + 304, + 300, + 545, + 339 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 304, + 300, + 545, + 339 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 300, + 545, + 339 + ], + "type": "text", + "content": "-th layer of VGG. Second, the distance between the Gram matrices " + }, + { + "bbox": [ + 304, + 300, + 545, + 339 + ], + "type": "inline_equation", + "content": "G_{j}^{\\phi}" + }, + { + "bbox": [ + 304, + 300, + 545, + 339 + ], + "type": "text", + "content": " constructed from the " + }, + { + "bbox": [ + 304, + 300, + 545, + 339 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 300, + 545, + 339 + ], + "type": "text", + "content": "-th activation maps " + }, + { + "bbox": [ + 304, + 300, + 545, + 339 + ], + "type": "inline_equation", + "content": "\\phi_{j}" + }, + { + "bbox": [ + 304, + 300, + 545, + 339 + ], + "type": "text", + "content": " [29] of " + }, + { + "bbox": [ + 304, + 300, + 545, + 339 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 304, + 300, + 545, + 339 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 300, + 545, + 339 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{I}}" + }, + { + "bbox": [ + 304, + 300, + 545, + 339 + ], + "type": "text", + "content": " is measured:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 362, + 346, + 545, + 374 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 346, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 362, + 346, + 545, + 374 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s} = \\sum_ {j} \\left\\| G _ {j} ^ {\\phi} (\\mathbf {I}) - G _ {j} ^ {\\phi} (\\hat {\\mathbf {I}}) \\right\\| _ {1}. \\tag {12}", + "image_path": "12f21610cf7ec1e355a76b9c44782e1cb22818e7bd0d497077fdcb5b93fddb44.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 380, + 545, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 380, + 545, + 405 + ], + "spans": [ + { + "bbox": [ + 304, + 380, + 545, + 405 + ], + "type": "text", + "content": "Third, we utilize the L1 distance performed on the eyes and mouth regions to supervise the expression detail:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 361, + 411, + 545, + 439 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 411, + 545, + 439 + ], + "spans": [ + { + "bbox": [ + 361, + 411, + 545, + 439 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {m} = \\sum_ {n} \\left\\| R _ {n} (\\mathbf {I}) - R _ {n} (\\hat {\\mathbf {I}}) \\right\\| _ {1}, \\tag {13}", + "image_path": "62edef548b888dc81c78b4a7ca55e5e178167512bdc4cef726b4e6a807e3d726.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 445, + 545, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 445, + 545, + 493 + ], + "spans": [ + { + "bbox": [ + 304, + 445, + 545, + 493 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 445, + 545, + 493 + ], + "type": "inline_equation", + "content": "R_{n}" + }, + { + "bbox": [ + 304, + 445, + 545, + 493 + ], + "type": "text", + "content": " is either the eyes or mouth region extracted using RoI-align on the bounding boxes calculated using landmarks. Finally, the ID Loss is performed to preserve the identity consistency:" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 367, + 500, + 545, + 514 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 500, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 367, + 500, + 545, + 514 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {i d} = 1 - \\cos (E (\\mathbf {I}), E (\\tilde {\\mathbf {I}})), \\tag {14}", + "image_path": "b2956a43792e55bb513a890f99e470f4b6b4148ee0b77e906dbc34294391083b.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 521, + 512, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 512, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 512, + 533 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 521, + 512, + 533 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 304, + 521, + 512, + 533 + ], + "type": "text", + "content": " is the Arcface face recognition model [8]." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 533, + 545, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 545, + 606 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 545, + 606 + ], + "type": "text", + "content": "The animation loss " + }, + { + "bbox": [ + 304, + 533, + 545, + 606 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(\\mathbf{I},\\hat{\\mathbf{I}})" + }, + { + "bbox": [ + 304, + 533, + 545, + 606 + ], + "type": "text", + "content": " is a weighted summation of the above loss terms and is implemented in both training and test-time one-shot avatar construction. Additionally, in the training, we maintain the stability of tri-plane representation with monotonic and TV loss [3], and " + }, + { + "bbox": [ + 304, + 533, + 545, + 606 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_1" + }, + { + "bbox": [ + 304, + 533, + 545, + 606 + ], + "type": "text", + "content": " loss on " + }, + { + "bbox": [ + 304, + 533, + 545, + 606 + ], + "type": "inline_equation", + "content": "\\Theta_c" + }, + { + "bbox": [ + 304, + 533, + 545, + 606 + ], + "type": "text", + "content": " to avoid the latent code being out of the latent distribution of " + }, + { + "bbox": [ + 304, + 533, + 545, + 606 + ], + "type": "inline_equation", + "content": "G_{eg}" + }, + { + "bbox": [ + 304, + 533, + 545, + 606 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 306, + 616, + 388, + 629 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 616, + 388, + 629 + ], + "spans": [ + { + "bbox": [ + 306, + 616, + 388, + 629 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 306, + 635, + 420, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 635, + 420, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 635, + 420, + 647 + ], + "type": "text", + "content": "4.1. Experiment Setting" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "In this section, we first describe the implementation detail of our proposed method, the dataset used, and the baselines of our work. Then we compare our proposed method with previous 2D and 3D methods on motion controllability and multi-view consistency." + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16905" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 407 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 407 + ], + "type": "text", + "content": "Implementation details. We use the off-the-shelf 3D face reconstruction model to extract the expression and pose coefficients of 3DMM [9]. The motion of any timestamp is represented by the window of adjacent 27 frames of expression and pose coefficients. In practice, we find that the extracted head poses contain too much noises to serve as the camera pose. Therefore, we extract the camera pose using the method in [15]. Our model is implemented in Pytorch using four A100 GPUs. The total batch size is 24, with six images per GPU. We use Adam optimizer with a 0.0001 learning rate to train the motion controller " + }, + { + "bbox": [ + 47, + 72, + 289, + 407 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 47, + 72, + 289, + 407 + ], + "type": "text", + "content": " and finetune the generator " + }, + { + "bbox": [ + 47, + 72, + 289, + 407 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 47, + 72, + 289, + 407 + ], + "type": "text", + "content": ". Exponential Moving Average (EMA) is employed to update " + }, + { + "bbox": [ + 47, + 72, + 289, + 407 + ], + "type": "inline_equation", + "content": "\\Theta_{c}" + }, + { + "bbox": [ + 47, + 72, + 289, + 407 + ], + "type": "text", + "content": " since it can stabilize the training. The model is trained for 2000 iterations, and each iteration contains " + }, + { + "bbox": [ + 47, + 72, + 289, + 407 + ], + "type": "inline_equation", + "content": "N = N_{id} + N_{mo} = 100" + }, + { + "bbox": [ + 47, + 72, + 289, + 407 + ], + "type": "text", + "content": " steps to optimize the identity codes and train the motion controller. For each minibatch of data, a new batch of latent codes are initialized and optimized using the Adam optimizer with a 0.01 learning rate. Both the " + }, + { + "bbox": [ + 47, + 72, + 289, + 407 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 47, + 72, + 289, + 407 + ], + "type": "text", + "content": " resolution volume rendering results and " + }, + { + "bbox": [ + 47, + 72, + 289, + 407 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 47, + 72, + 289, + 407 + ], + "type": "text", + "content": " super-resolution results are implemented to calculate the loss. During training, the identity code " + }, + { + "bbox": [ + 47, + 72, + 289, + 407 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_{id}" + }, + { + "bbox": [ + 47, + 72, + 289, + 407 + ], + "type": "text", + "content": " is optimized in " + }, + { + "bbox": [ + 47, + 72, + 289, + 407 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 47, + 72, + 289, + 407 + ], + "type": "text", + "content": " space. After being extended to the " + }, + { + "bbox": [ + 47, + 72, + 289, + 407 + ], + "type": "inline_equation", + "content": "\\mathcal{W}^{+}" + }, + { + "bbox": [ + 47, + 72, + 289, + 407 + ], + "type": "text", + "content": "space by channel-wisely repeated 14 times, it is combined with the motion code " + }, + { + "bbox": [ + 47, + 72, + 289, + 407 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_x" + }, + { + "bbox": [ + 47, + 72, + 289, + 407 + ], + "type": "text", + "content": " and then fed into the face generator " + }, + { + "bbox": [ + 47, + 72, + 289, + 407 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 47, + 72, + 289, + 407 + ], + "type": "text", + "content": ". It takes less than two days to train the network entirely. During inference, we first optimize the identity code in " + }, + { + "bbox": [ + 47, + 72, + 289, + 407 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 47, + 72, + 289, + 407 + ], + "type": "text", + "content": " space, then use another " + }, + { + "bbox": [ + 47, + 72, + 289, + 407 + ], + "type": "inline_equation", + "content": "N = 100" + }, + { + "bbox": [ + 47, + 72, + 289, + 407 + ], + "type": "text", + "content": " steps to finetune it in " + }, + { + "bbox": [ + 47, + 72, + 289, + 407 + ], + "type": "inline_equation", + "content": "\\mathcal{W}^{+}" + }, + { + "bbox": [ + 47, + 72, + 289, + 407 + ], + "type": "text", + "content": "space [1, 2]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 409, + 287, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 409, + 287, + 517 + ], + "spans": [ + { + "bbox": [ + 47, + 409, + 287, + 517 + ], + "type": "text", + "content": "Dataset. Our model is trained on the HDTF [43] dataset, which contains frontal talking faces from 362 videos and over 300 subjects. Following [40], we use the pre-processing step in [31] and resize images to " + }, + { + "bbox": [ + 47, + 409, + 287, + 517 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 47, + 409, + 287, + 517 + ], + "type": "text", + "content": " resolution. The test dataset contains 20 videos of subjects out of the training set. To evaluate the novel subject generalizability and controllability in the 3D-consistent fashion, we sample subjects from the Multiface [38] dataset, which is particularly proposed for 3D-consistent face avatars." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 520, + 287, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 520, + 287, + 639 + ], + "spans": [ + { + "bbox": [ + 47, + 520, + 287, + 639 + ], + "type": "text", + "content": "Baseline methods. We mainly compare our method with previous works that incorporate 3DMM coefficients to reenact face avatars in either 2D/3D fashion and are generalizable to unseen identities. For 2D methods, PIRenderer [29] and StyleHEAT [40] achieve SOTA performance on face reenactment. For 3D methods, HeadNeRF [17] is the most advanced volume rendering method that incorporates controllability and generalizability. Beyond coefficients-based methods, the SOTA image-driven method FOMM [31] is also taken into comparison." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 642, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 642, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 642, + 287, + 713 + ], + "type": "text", + "content": "Evaluation metrics. We adopt peak-signal-to-noise ratio (PSNR), structural similarity index measure (SSIM) [37], and learned perceptual image patch similarity (LPIPS) [42] to evaluate the visual quality. To measure the realism of the synthesized results and the identity preservation, we use frechet inception distance (FID) [16] and" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 307, + 72, + 545, + 286 + ], + "blocks": [ + { + "bbox": [ + 307, + 72, + 545, + 286 + ], + "lines": [ + { + "bbox": [ + 307, + 72, + 545, + 286 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 545, + 286 + ], + "type": "image", + "image_path": "18b56b69d9f6f066ca3869b1992f733f12eaf712519100cc1a856d8cbc38bb90.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 293, + 545, + 327 + ], + "lines": [ + { + "bbox": [ + 305, + 293, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 305, + 293, + 545, + 327 + ], + "type": "text", + "content": "Figure 3. Qualitative result for cross-identity reenactment. Examples are sampled from the HDTF dataset [43]. Both source and driving subjects are not included in the training set." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 349, + 545, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 349, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 305, + 349, + 545, + 409 + ], + "type": "text", + "content": "cosine similarity of identity embedding (CSIM) [8] based on ArcFace model, respectively. Besides, the average expression distance (AED), average pose distance (APD), and average keypoint distance (AKD) are employed to evaluate the facial expression and pose." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 418, + 411, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 418, + 411, + 430 + ], + "spans": [ + { + "bbox": [ + 306, + 418, + 411, + 430 + ], + "type": "text", + "content": "4.2. Evaluation Result" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 437, + 545, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 437, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 437, + 545, + 628 + ], + "type": "text", + "content": "We evaluate OTAvatar on animating photo-realistic talking head videos and compare them with state-of-the-art models that support identity-generalized animation methods. In the HDTF [43] dataset, we examine the identity-motion disentanglement by transferring the motion of one subject to drive the other subjects, namely cross-identity reenactment. The employed motion is extracted from a video with large motion variation to evaluate the result in extreme conditions. In the Multiface [38] dataset, we evaluate the consistency of the animation in different views, namely multi-view reenactment. Note that none of the data in this dataset has been used in training our method and baselines. For each talking corpus, we choose the first frame of the frontal-view camera recording as the reference and enforce the network to animate the following frames in frontal view and other views." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "Qualitative comparison. Fig. 3 shows the results of different methods on cross-identity generation. Compared to FOMM [31], PIRenderer [29] and StyleHEAT [40], which use warping fields for reenacting faces, our method can handle extreme pose variation and maintain identity consistency. Compared to the 3D method of HeadNeRF, our model fully reconstructs the identity-specific detail and syn" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "16906" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 73, + 547, + 372 + ], + "blocks": [ + { + "bbox": [ + 51, + 73, + 547, + 372 + ], + "lines": [ + { + "bbox": [ + 51, + 73, + 547, + 372 + ], + "spans": [ + { + "bbox": [ + 51, + 73, + 547, + 372 + ], + "type": "image", + "image_path": "2d9797dc277adee4d2c846da9fe23815799d108bc8d8dbeb9857ce3cff89e9b3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 380, + 547, + 413 + ], + "lines": [ + { + "bbox": [ + 46, + 380, + 547, + 413 + ], + "spans": [ + { + "bbox": [ + 46, + 380, + 547, + 413 + ], + "type": "text", + "content": "Figure 4. Qualitative result for multi-view reenactment. Examples are sampled from the multi-view dataset [38]. All methods use the first frame of the frontal-view portrait to extract the identity feature, and take the expressions of sequential frames and poses of different camera views to generate the talking face. Note that this subject is not included in the training set of any methods." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 47, + 430, + 548, + 518 + ], + "blocks": [ + { + "bbox": [ + 47, + 430, + 548, + 518 + ], + "lines": [ + { + "bbox": [ + 47, + 430, + 548, + 518 + ], + "spans": [ + { + "bbox": [ + 47, + 430, + 548, + 518 + ], + "type": "table", + "html": "
Multi-View ReenactmentCross-Identity Reenactment
PSNR ↑SSIM ↑CSIM ↑AED ↓APD ↓AKD ↓LIPIPS ↓FID ↓CSIM ↑AED ↓APD ↓AKD ↓FID ↓
FOMM20.750.6390.5052.0040.5455.0520.308101.60.6723.1960.5004.198113.2
PIRenderer20.040.5860.4932.2030.6806.5660.299100.60.6323.0180.4984.977103.7
StyleHEAT20.030.6320.3872.1790.4725.5220.284123.80.6142.8600.4713.592239.1
HeadNeRF17.600.5460.2392.0860.7764.1660.367212.30.2822.8730.5673.465233.0
Ours21.190.6570.5741.8740.4283.7310.288137.30.6942.8500.4054.307101.8
", + "image_path": "f88b995063f61e75164c5e51dd6032615075cc4680279fc54582e4ea9b5e74bd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 525, + 477, + 536 + ], + "lines": [ + { + "bbox": [ + 115, + 525, + 477, + 536 + ], + "spans": [ + { + "bbox": [ + 115, + 525, + 477, + 536 + ], + "type": "text", + "content": "Table 1. Quantitative comparisons on multi-View reenactment and cross-identity reenactment." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 558, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 287, + 653 + ], + "type": "text", + "content": "thesizes more natural expression. The multi-view consistency results are shown in Fig. 4. First, the 2D warping method suffers facial malformation, which becomes more serious in larger poses. Second, though HeadNeRF renders accurate head poses, the generated results have noticeable deterioration compared to the ground truth. Finally, the face avatar constructed by our method preserves identity details and multi-view consistency." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 654, + 287, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 287, + 700 + ], + "type": "text", + "content": "Quantitative comparison. The quantitative comparison among competing methods is shown in Table 1. We can see that the proposed method achieves the best performance in terms of most of the criteria." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "text", + "content": "Inference speed. Table 2 lists the inference speed com" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 558, + 545, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 606 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 606 + ], + "type": "text", + "content": "parison among state-of-the-art 3D face avatar methods. By utilizing the pre-trained 3D face generator and the compact motion controller (0.8M parameters), our method achieves the highest inference speed, demonstrating its effectiveness." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 611, + 400, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 611, + 400, + 624 + ], + "spans": [ + { + "bbox": [ + 306, + 611, + 400, + 624 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": "Decoupling-by-inverting. The controller is trained by the decoupling-by-inverting strategy described in Sec. 3.2, where the training alternates between " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "N_{id}" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": " steps of identity code optimization and " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "N_{mo}" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": " steps of controller parameters training. To validate the effectiveness of the alternative manner, we discard it and jointly train the controller parameters and the identity code. The qualitative comparison is" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16907" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 52, + 70, + 282, + 142 + ], + "blocks": [ + { + "bbox": [ + 52, + 70, + 282, + 142 + ], + "lines": [ + { + "bbox": [ + 52, + 70, + 282, + 142 + ], + "spans": [ + { + "bbox": [ + 52, + 70, + 282, + 142 + ], + "type": "table", + "html": "
MethodsFrames Per Second ↑
IMAvatar [45]0.03
ADNeRF [15]0.13
HeadNeRF [17]25
Ours35
", + "image_path": "5f76f54b7f3a5ce6c202ee1f30fa63b1513283ca567d0913b97a55e7c3cac869.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 83, + 190, + 251, + 404 + ], + "blocks": [ + { + "bbox": [ + 83, + 190, + 251, + 404 + ], + "lines": [ + { + "bbox": [ + 83, + 190, + 251, + 404 + ], + "spans": [ + { + "bbox": [ + 83, + 190, + 251, + 404 + ], + "type": "image", + "image_path": "53172a8c4df956366c8f40b2f19a5b8e53dce57e021ffbeff17931d0bdcb8572.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 413, + 287, + 447 + ], + "lines": [ + { + "bbox": [ + 46, + 413, + 287, + 447 + ], + "spans": [ + { + "bbox": [ + 46, + 413, + 287, + 447 + ], + "type": "text", + "content": "Figure 5. Qualitative comparison of decoupling-by-inverting training and joint training. Joint training cannot preserve identity information in one-shot avatar construction." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 471, + 286, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 471, + 286, + 590 + ], + "spans": [ + { + "bbox": [ + 46, + 471, + 286, + 590 + ], + "type": "text", + "content": "in Fig. 5. We see that the jointly trained model cannot preserve identity during the animation because the controller manages to overfit the identity information in the training set, and the identity information is not fully encoded in the identity code. The performance degradation caused by joint training is also shown in Table 4. We also analyze the setting of identity optimization step " + }, + { + "bbox": [ + 46, + 471, + 286, + 590 + ], + "type": "inline_equation", + "content": "N_{id}" + }, + { + "bbox": [ + 46, + 471, + 286, + 590 + ], + "type": "text", + "content": " and controller training step " + }, + { + "bbox": [ + 46, + 471, + 286, + 590 + ], + "type": "inline_equation", + "content": "N_{mo}" + }, + { + "bbox": [ + 46, + 471, + 286, + 590 + ], + "type": "text", + "content": ", and list the results in Table 3. We observe that " + }, + { + "bbox": [ + 46, + 471, + 286, + 590 + ], + "type": "inline_equation", + "content": "N_{id} = 90" + }, + { + "bbox": [ + 46, + 471, + 286, + 590 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 471, + 286, + 590 + ], + "type": "inline_equation", + "content": "N_{mo} = 10" + }, + { + "bbox": [ + 46, + 471, + 286, + 590 + ], + "type": "text", + "content": " achieves a balance of expression reconstruction (AED) and identity preserving (CSIM)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 592, + 286, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 592, + 286, + 676 + ], + "spans": [ + { + "bbox": [ + 46, + 592, + 286, + 676 + ], + "type": "text", + "content": "Losses. We also perform an ablation study on the loss functions, including the landmark region losses " + }, + { + "bbox": [ + 46, + 592, + 286, + 676 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_m" + }, + { + "bbox": [ + 46, + 592, + 286, + 676 + ], + "type": "text", + "content": " and ID losses " + }, + { + "bbox": [ + 46, + 592, + 286, + 676 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{id}" + }, + { + "bbox": [ + 46, + 592, + 286, + 676 + ], + "type": "text", + "content": ", as shown in Table 4. We can see that the absence of landmark region losses " + }, + { + "bbox": [ + 46, + 592, + 286, + 676 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_m" + }, + { + "bbox": [ + 46, + 592, + 286, + 676 + ], + "type": "text", + "content": " causes a deterioration of expression reconstruction, measured by AED, while the absence of ID loss causes a performance drop in both expression and identity consistency." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "content": "3D generator finetuning. When the controller training and identity code optimization are finished, we further slightly finetune the pre-trained EG3D [3] face generator" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 323, + 70, + 529, + 142 + ], + "blocks": [ + { + "bbox": [ + 47, + 152, + 287, + 174 + ], + "lines": [ + { + "bbox": [ + 47, + 152, + 287, + 174 + ], + "spans": [ + { + "bbox": [ + 47, + 152, + 287, + 174 + ], + "type": "text", + "content": "Table 2. Quantitative comparison on the inference speed. The comparison is conducted on an A100 GPU." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 323, + 70, + 529, + 142 + ], + "lines": [ + { + "bbox": [ + 323, + 70, + 529, + 142 + ], + "spans": [ + { + "bbox": [ + 323, + 70, + 529, + 142 + ], + "type": "table", + "html": "
NmoNidAED ↓CSIM ↑
1993.2850.7200
5953.1780.716
10902.8500.694
20803.2310.592
", + "image_path": "aaf24cafba30073bc774d69dc20ee838e1cd7ef917b4353ee571f26b03b9a279.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 328, + 196, + 525, + 281 + ], + "blocks": [ + { + "bbox": [ + 305, + 152, + 545, + 185 + ], + "lines": [ + { + "bbox": [ + 305, + 152, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 305, + 152, + 545, + 185 + ], + "type": "text", + "content": "Table 3. Ablation study on the decoupling-by-inverting hyperparameters. Experiments are conducted under the different settings of " + }, + { + "bbox": [ + 305, + 152, + 545, + 185 + ], + "type": "inline_equation", + "content": "N_{mo}" + }, + { + "bbox": [ + 305, + 152, + 545, + 185 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 152, + 545, + 185 + ], + "type": "inline_equation", + "content": "N_{id}" + }, + { + "bbox": [ + 305, + 152, + 545, + 185 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 305, + 152, + 545, + 185 + ], + "type": "inline_equation", + "content": "N = N_{mo} + N_{id} = 100" + }, + { + "bbox": [ + 305, + 152, + 545, + 185 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 328, + 196, + 525, + 281 + ], + "lines": [ + { + "bbox": [ + 328, + 196, + 525, + 281 + ], + "spans": [ + { + "bbox": [ + 328, + 196, + 525, + 281 + ], + "type": "table", + "html": "
AED ↓CSIM ↑
ours2.8500.694
joint3.0090.689
w/o Lm3.1010.687
w/o Lid3.1640.592
w/o finetune3.1450.687
", + "image_path": "e1afd0adc05fec1c4a39978e4242d20d9b763787c645a1c8a4f51366c1ed917d.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 290, + 545, + 312 + ], + "lines": [ + { + "bbox": [ + 305, + 290, + 545, + 312 + ], + "spans": [ + { + "bbox": [ + 305, + 290, + 545, + 312 + ], + "type": "text", + "content": "Table 4. Ablation study on the joint training, loss terms, and finetuning." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 335, + 545, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 335, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 305, + 335, + 545, + 384 + ], + "type": "text", + "content": "with a learning rate of 0.0001. In Table 4, we show the results without finetuning. One can see that slightly finetuning the face generator in the final stage improves identity and expression reconstruction performance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 396, + 378, + 408 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 396, + 378, + 408 + ], + "spans": [ + { + "bbox": [ + 306, + 396, + 378, + 408 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 416, + 545, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 416, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 304, + 416, + 545, + 644 + ], + "type": "text", + "content": "We proposed a novel framework of one-shot 3D-consistent talking face avatar, namely OTAvatar, using volume rendering. It jointly addressed the three challenges of face avatars, i.e., generalizability, controllability, and efficiency. For identity-generalized rendering, we implemented a pre-trained 3D face generator to reconstruct faces faithfully, given a single portrait reference. For motion control, we proposed a motion controller module, which predicts the motion code conditioned on 3DMM coefficients. For efficiency, benefiting from the compact architecture of both the 3D generator and controller, our model can animate avatars at a high speed. Besides, we proposed a decoupling-by-inverting approach, which is both a training scheme and a test-time disentangle strategy that decouples the latent code into identity and motion codes via GAN inversion, so that one can animate the avatar with different motions using the motion codes predicted by the controller. Comprehensive experiments were conducted to prove the advantage of our proposed framework." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 657, + 404, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 657, + 404, + 670 + ], + "spans": [ + { + "bbox": [ + 306, + 657, + 404, + 670 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 677, + 545, + 713 + ], + "type": "text", + "content": "This work is supported in part by the Chinese National Natural Science Foundation Projects #62176256, #62276254, and the InnoHK program." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16908" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan: How to embed images into the stylegan latent space? In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4432-4441, 2019." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 136, + 288, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 136, + 288, + 179 + ], + "spans": [ + { + "bbox": [ + 53, + 136, + 288, + 179 + ], + "type": "text", + "content": "[2] Rameen Abdul, Yipeng Qin, and Peter Wonka. Image2stylegan++: How to edit the embedded images? In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8296-8305, 2020." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 180, + 287, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 180, + 287, + 235 + ], + "spans": [ + { + "bbox": [ + 53, + 180, + 287, + 235 + ], + "type": "text", + "content": "[3] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In CVPR, 2022." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 236, + 287, + 291 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 236, + 287, + 291 + ], + "spans": [ + { + "bbox": [ + 53, + 236, + 287, + 291 + ], + "type": "text", + "content": "[4] Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5799-5809, 2021." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 291, + 287, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 291, + 287, + 345 + ], + "spans": [ + { + "bbox": [ + 53, + 291, + 287, + 345 + ], + "type": "text", + "content": "[5] Lele Chen, Ross K Maddox, Zhiyao Duan, and Chenliang Xu. Hierarchical cross-modal talking face generation with dynamic pixel-wise loss. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7832-7841, 2019." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 346, + 287, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 346, + 287, + 380 + ], + "spans": [ + { + "bbox": [ + 53, + 346, + 287, + 380 + ], + "type": "text", + "content": "[6] Min Jin Chong, Hsin-Ying Lee, and David Forsyth. Stylegan of all trades: Image manipulation with only pretrained stylegan. arXiv preprint arXiv:2111.01619, 2021." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 380, + 287, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 287, + 413 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 287, + 413 + ], + "type": "text", + "content": "[7] Joon Son Chung and Andrew Zisserman. Lip reading in the wild. In Asian conference on computer vision, pages 87-103. Springer, 2016." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 415, + 287, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 415, + 287, + 468 + ], + "spans": [ + { + "bbox": [ + 53, + 415, + 287, + 468 + ], + "type": "text", + "content": "[8] Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2019." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 469, + 287, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 469, + 287, + 524 + ], + "spans": [ + { + "bbox": [ + 53, + 469, + 287, + 524 + ], + "type": "text", + "content": "[9] Yu Deng, Jiaolong Yang, Sicheng Xu, Dong Chen, Yunde Jia, and Xin Tong. Accurate 3d face reconstruction with weakly-supervised learning: From single image to image set. In IEEE Computer Vision and Pattern Recognition Workshops, 2019." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 525, + 287, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 287, + 568 + ], + "type": "text", + "content": "[10] Michail Christos Doukas, Stefanos Zafeiriou, and Viktoriia Sharmanska. Headgan: One-shot neural head synthesis and editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14398-14407, 2021." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 570, + 287, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 570, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 48, + 570, + 287, + 602 + ], + "type": "text", + "content": "[11] Randima Fernando et al. GPU gems: programming techniques, tips, and tricks for real-time graphics, volume 590. Addison-Wesley Reading, 2004." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 603, + 287, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 603, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 48, + 603, + 287, + 658 + ], + "type": "text", + "content": "[12] Guy Gafni, Justus Thies, Michael Zollhöfer, and Matthias Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8649-8658, June 2021." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 713 + ], + "type": "text", + "content": "[13] Guy Gafni, Justus Thies, Michael Zollhofer, and Matthias Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8649-8658, 2021." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 116 + ], + "type": "text", + "content": "[14] Zhenglin Geng, Chen Cao, and Sergey Tulyakov. 3d guided fine-grained face manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9821-9830, 2019." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 547, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 547, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 547, + 172 + ], + "type": "text", + "content": "[15] Yudong Guo, Keyu Chen, Sen Liang, Yong-Jin Liu, Hujun Bao, and Juyong Zhang. Ad-nerf: Audio driven neural radiance fields for talking head synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5784–5794, 2021." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 173, + 546, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 173, + 546, + 226 + ], + "spans": [ + { + "bbox": [ + 307, + 173, + 546, + 226 + ], + "type": "text", + "content": "[16] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 228, + 546, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 228, + 546, + 281 + ], + "spans": [ + { + "bbox": [ + 307, + 228, + 546, + 281 + ], + "type": "text", + "content": "[17] Yang Hong, Bo Peng, Haiyao Xiao, Ligang Liu, and Juyong Zhang. Headnerf: A real-time nerf-based parametric head model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20374-20384, 2022." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 283, + 546, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 283, + 546, + 338 + ], + "spans": [ + { + "bbox": [ + 307, + 283, + 546, + 338 + ], + "type": "text", + "content": "[18] Xinya Ji, Hang Zhou, Kaisiyuan Wang, Wayne Wu, Chen Change Loy, Xun Cao, and Feng Xu. Audio-driven emotional video portraits. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14080-14089, 2021." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 338, + 547, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 338, + 547, + 392 + ], + "spans": [ + { + "bbox": [ + 307, + 338, + 547, + 392 + ], + "type": "text", + "content": "[19] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 394, + 546, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 394, + 546, + 415 + ], + "spans": [ + { + "bbox": [ + 307, + 394, + 546, + 415 + ], + "type": "text", + "content": "[20] Marc Levoy. Display of surfaces from volume data. IEEE Computer graphics and Applications, 8(3):29-37, 1988." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 416, + 546, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 416, + 546, + 459 + ], + "spans": [ + { + "bbox": [ + 307, + 416, + 546, + 459 + ], + "type": "text", + "content": "[21] Stephen Lombardi, Tomas Simon, Jason Saragih, Gabriel Schwartz, Andreas Lehrmann, and Yaser Sheikh. Neural volumes: Learning dynamic renderable volumes from images. ACM Trans. Graph., 38(4):65:1-65:14, July 2019." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 460, + 546, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 460, + 546, + 514 + ], + "spans": [ + { + "bbox": [ + 307, + 460, + 546, + 514 + ], + "type": "text", + "content": "[22] Marko Mihajlovic, Aayush Bansal, Michael Zollhoefer, Siyu Tang, and Shunsuke Saito. KeypointNeRF: Generalizing image-based volumetric avatars using relative spatial encoding of keypoints. In European conference on computer vision, 2022." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 514, + 546, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 514, + 546, + 558 + ], + "spans": [ + { + "bbox": [ + 307, + 514, + 546, + 558 + ], + "type": "text", + "content": "[23] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 559, + 546, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 559, + 546, + 603 + ], + "spans": [ + { + "bbox": [ + 307, + 559, + 546, + 603 + ], + "type": "text", + "content": "[24] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11453-11464, 2021." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 603, + 546, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 603, + 546, + 658 + ], + "spans": [ + { + "bbox": [ + 307, + 603, + 546, + 658 + ], + "type": "text", + "content": "[25] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865-5874, 2021." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 658, + 546, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 546, + 713 + ], + "type": "text", + "content": "[26] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Ricardo Martin-Brualla, and Steven M. Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. ACM Trans. Graph., 40(6), dec 2021." + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "16909" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[27] KR Prajwal, Rudrabha Mukhopadhyay, Vinay P Namboodiri, and CV Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In Proceedings of the 28th ACM International Conference on Multimedia, pages 484-492, 2020." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 287, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 173 + ], + "type": "text", + "content": "[28] Amit Raj, Michael Zollhoefer, Tomas Simon, Jason Saragih, Shunsuke Saito, James Hays, and Stephen Lombardi. Pva: Pixel-aligned volumetric avatars. arXiv preprint arXiv:2101.02697, 2021." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 175, + 287, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 287, + 228 + ], + "type": "text", + "content": "[29] Yurui Ren, Ge Li, Yuanqi Chen, Thomas H Li, and Shan Liu. Pirenderer: Controllable portrait image generation via semantic neural rendering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13759-13768, 2021." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 230, + 287, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 230, + 287, + 274 + ], + "spans": [ + { + "bbox": [ + 48, + 230, + 287, + 274 + ], + "type": "text", + "content": "[30] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. Graf: Generative radiance fields for 3d-aware image synthesis. Advances in Neural Information Processing Systems, 33:20154-20166, 2020." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 276, + 287, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 276, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 276, + 287, + 319 + ], + "type": "text", + "content": "[31] Aliaksandr Siarohin, Stéphane Lathuilière, Sergey Tulyakov, Elisa Ricci, and Nicu Sebe. First order motion model for image animation. Advances in Neural Information Processing Systems, 32, 2019." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 321, + 287, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 321, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 48, + 321, + 287, + 376 + ], + "type": "text", + "content": "[32] Aliaksandr Siarohin, Oliver J Woodford, Jian Ren, Menglei Chai, and Sergey Tulyakov. Motion representations for articulated animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13653-13662, 2021." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 377, + 287, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 377, + 287, + 433 + ], + "spans": [ + { + "bbox": [ + 48, + 377, + 287, + 433 + ], + "type": "text", + "content": "[33] Ivan Skorokhodov, Sergey Tulyakov, and Mohamed Elhoseiny. Stylegan-v: A continuous video generator with the price, image quality and perks of stylegan2. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3626-3636, 2022." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 434, + 287, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 434, + 287, + 488 + ], + "spans": [ + { + "bbox": [ + 48, + 434, + 287, + 488 + ], + "type": "text", + "content": "[34] Justus Thies, Michael Zollhofer, Marc Stamminger, Christian Theobalt, and Matthias Nießner. Face2face: Real-time face capture and reenactment of rgb videos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2387-2395, 2016." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 490, + 287, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 490, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 490, + 287, + 533 + ], + "type": "text", + "content": "[35] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Guilin Liu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. Video-to-video synthesis. In Advances in Neural Information Processing Systems (NeurIPS), 2018." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 535, + 287, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 535, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 535, + 287, + 578 + ], + "type": "text", + "content": "[36] Yaohui Wang, Di Yang, Francois Bremond, and Antitza Dantcheva. Latent image animator: Learning to animate images via latent space navigation. arXiv preprint arXiv:2203.09043, 2022." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 580, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 580, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 580, + 287, + 624 + ], + "type": "text", + "content": "[37] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 625, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 287, + 713 + ], + "type": "text", + "content": "[38] Cheng-hsin Wu, Ningyuan Zheng, Scott Ardisson, Rohan Bali, Danielle Belko, Eric Brockmeyer, Lucas Evans, Timothy Godisart, Hyowon Ha, Alexander Hypes, Taylor Koska, Steven Krenn, Stephen Lombardi, Xiaomin Luo, Kevyn McPhail, Laura Millerschoen, Michal Perdoch, Mark Pitts, Alexander Richard, Jason Saragih, Junko Saragih, Takaaki Shiratori, Tomas Simon, Matt Stewart, Autumn Trimble, Xinshuo Weng, David Whitewolf, Chenglei Wu, Shouu-I Yu," + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 553 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "and Yaser Sheikh. Multiface: A dataset for neural face rendering. In arXiv, 2022." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 96, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 150 + ], + "type": "text", + "content": "[39] Tao Yang, Peiran Ren, Xuansong Xie, and Lei Zhang. Gan prior embedded network for blind face restoration in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 672-681, June 2021." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 152, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 206 + ], + "type": "text", + "content": "[40] Fei Yin, Yong Zhang, Xiaodong Cun, Mingdeng Cao, Yanbo Fan, Xuan Wang, Qingyan Bai, Baoyuan Wu, Jue Wang, and Yujiu Yang. Styleheat: One-shot high-resolution editable talking face generation via pretrained stylegan. arXiv preprint arXiv:2203.04036, 2022." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 208, + 545, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 208, + 545, + 252 + ], + "spans": [ + { + "bbox": [ + 307, + 208, + 545, + 252 + ], + "type": "text", + "content": "[41] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4578-4587, 2021." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 253, + 545, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 253, + 545, + 307 + ], + "spans": [ + { + "bbox": [ + 307, + 253, + 545, + 307 + ], + "type": "text", + "content": "[42] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 308, + 545, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 308, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 307, + 308, + 545, + 363 + ], + "type": "text", + "content": "[43] Zhimeng Zhang, Lincheng Li, Yu Ding, and Changjie Fan. Flow-guided one-shot talking face generation with a high-resolution audio-visual dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3661–3670, 2021." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 364, + 545, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 364, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 307, + 364, + 545, + 407 + ], + "type": "text", + "content": "[44] Zicheng Zhang, Yinglu Liu, Congying Han, Tiande Guo, Ting Yao, and Tao Mei. Generalized one-shot domain adaption of generative adversarial networks. arXiv preprint arXiv:2209.03665, 2022." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 409, + 545, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 409, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 307, + 409, + 545, + 453 + ], + "type": "text", + "content": "[45] Yufeng Zheng, Victoria Fernández Abrevaya, Marcel C. Bühler, Xu Chen, Michael J. Black, and Otmar Hilliges. I M Avatar: Implicit morphable head avatars from videos. In Computer Vision and Pattern Recognition (CVPR), 2022." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 454, + 545, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 454, + 545, + 497 + ], + "spans": [ + { + "bbox": [ + 307, + 454, + 545, + 497 + ], + "type": "text", + "content": "[46] Hang Zhou, Yu Liu, Ziwei Liu, Ping Luo, and Xiaogang Wang. Talking face generation by adversarially disentangled audio-visual representation. In AAAI Conference on Artificial Intelligence (AAAI), 2019." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 498, + 545, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 498, + 545, + 553 + ], + "spans": [ + { + "bbox": [ + 307, + 498, + 545, + 553 + ], + "type": "text", + "content": "[47] Hang Zhou, Yasheng Sun, Wayne Wu, Chen Change Loy, Xiaogang Wang, and Ziwei Liu. Pose-controllable talking face generation by implicitly modularized audio-visual representation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021." + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "16910" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/OVTrack_ Open-Vocabulary Multiple Object Tracking/092f1f8a-50aa-41bc-8e59-a89dde3c1cc2_content_list.json b/2023/OVTrack_ Open-Vocabulary Multiple Object Tracking/092f1f8a-50aa-41bc-8e59-a89dde3c1cc2_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ea3f96f31ccf5da1e8364e84f263414c1ea5256f --- /dev/null +++ b/2023/OVTrack_ Open-Vocabulary Multiple Object Tracking/092f1f8a-50aa-41bc-8e59-a89dde3c1cc2_content_list.json @@ -0,0 +1,1389 @@ +[ + { + "type": "text", + "text": "OVTrack: Open-Vocabulary Multiple Object Tracking", + "text_level": 1, + "bbox": [ + 207, + 130, + 763, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Siyuan Li* Tobias Fischer* Lei Ke Henghui Ding Martin Danelljan Fisher Yu Computer Vision Lab, ETH Zürich https://www.vis.xyz/pub/ovtrack/", + "bbox": [ + 235, + 180, + 730, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 286, + 313, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The ability to recognize, localize and track dynamic objects in a scene is fundamental to many real-world applications, such as self-driving and robotic systems. Yet, traditional multiple object tracking (MOT) benchmarks rely only on a few object categories that hardly represent the multitude of possible objects that are encountered in the real world. This leaves contemporary MOT methods limited to a small set of pre-defined object categories. In this paper, we address this limitation by tackling a novel task, open-vocabulary MOT, that aims to evaluate tracking beyond predefined training categories. We further develop OVTrack, an open-vocabulary tracker that is capable of tracking arbitrary object classes. Its design is based on two key ingredients: First, leveraging vision-language models for both classification and association via knowledge distillation; second, a data hallucination strategy for robust appearance feature learning from denoising diffusion probabilistic models. The result is an extremely data-efficient open-vocabulary tracker that sets a new state-of-the-art on the large-scale, large-vocabulary TAO benchmark, while being trained solely on static images.", + "bbox": [ + 75, + 318, + 473, + 636 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 667, + 209, + 681 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Multiple Object Tracking (MOT) aims to recognize, localize and track objects in a given video sequence. It is a cornerstone of dynamic scene analysis and vital for many real-world applications such as autonomous driving, augmented reality, and video surveillance. Traditionally, MOT benchmarks [9, 11, 19, 64, 71] define a set of semantic categories that constitute the objects to be tracked in the training and testing data distributions. The potential of traditional MOT methods [3, 4, 33, 43] is therefore limited by the taxonomies of those benchmarks. As a consequence, contemporary MOT methods struggle with unseen events, leading to a gap between evaluation performance and real-world", + "bbox": [ + 75, + 693, + 470, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/394eae8e3a8d336e2657acfe267ee8f7a5eb0040cfba683c429a42f56bc6dfde.jpg", + "image_caption": [ + "Figure 1. OVTrack. We approach the task of open-vocabulary multiple object tracking. During training, we leverage vision-language (VL) models both for generating samples and knowledge distillation. During testing, we track both base and novel classes unseen during training by querying a vision-language model." + ], + "image_footnote": [], + "bbox": [ + 504, + 285, + 890, + 503 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "deployment.", + "text_level": 1, + "bbox": [ + 500, + 613, + 586, + 627 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To bridge this gap, previous works have tackled MOT in an open-world context. In particular, Osep et al. [46, 48] approach generic object tracking by first segmenting the scene and performing tracking before classification. Other works have used class agnostic localizers [10, 49] to perform MOT on arbitrary objects. Recently, Liu et al. [37] defined open-world tracking, a task that focuses on the evaluation of previously unseen objects. In particular, it requires any-object tracking as a stage that precedes object classification. This setup comes with two inherent difficulties. First, in an open-world context, densely annotating all objects is prohibitively expensive. Second, without a pre-defined taxonomy of categories, the notion of what is an object is ambiguous. As a consequence, Liu et al. resort to recall-based evaluation, which is limited in two ways. Penalizing false positives (FP) becomes impossible, i.e. we cannot measure the tracker precision. Moreover, by evaluating tracking in a class-agnostic manner, we lose the ability to evaluate how", + "bbox": [ + 496, + 628, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 94, + 886, + 205, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "5567", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "well a tracker can infer the semantic category of an object.", + "bbox": [ + 76, + 90, + 464, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose open-vocabulary MOT as an effective solution to these problems. Similar to open-world MOT, open-vocabulary MOT aims to track multiple objects beyond the pre-defined training categories. However, instead of dismissing the classification problem and resorting to recall-based evaluation, we assume that at test time we are given the classes of objects we are interested in. This allows us to apply existing closed-set tracking metrics [34, 70] that capture both precision and recall, while still evaluating the tracker's ability to track arbitrary objects during inference.", + "bbox": [ + 75, + 106, + 467, + 256 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We further present the first Open-Vocabulary Tracker, OVTrack (see Fig. 1). To this end, we identify and address two fundamental challenges to the design of an open-vocabulary multi-object tracker. The first is that closed-set MOT methods are simply not capable of extending their pre-defined taxonomies. The second is data availability, i.e. scaling video data annotation to a large vocabulary of classes is extremely costly. Inspired by existing works in open-vocabulary detection [1, 13, 20, 76], we replace our classifier with an embedding head, which allows us to measure similarities of localized objects to an open vocabulary of semantic categories. In particular, we distill knowledge from CLIP [52] into our model by aligning the image feature representations of object proposals with the corresponding CLIP image and text embeddings.", + "bbox": [ + 75, + 257, + 468, + 483 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Beyond detection, association is the core of modern MOT methods. It is driven by two affinity cues: motion and appearance. In an open-vocabulary context, motion cues are brittle since arbitrary scenery contains complex and diverse camera and object motion patterns. In contrast, diverse objects usually exhibit heterogeneous appearance. However, relying on appearance cues requires robust representations that generalize to novel object categories. We find that CLIP feature distillation helps in learning better appearance representations for improved association. This is especially intriguing since object classification and appearance modeling are usually distinct in the MOT pipeline [3, 16, 65].", + "bbox": [ + 75, + 484, + 468, + 665 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Learning robust appearance features also requires strong supervision that captures object appearance changes in different viewpoints, background, and lighting. To approach the data availability problem, we utilize the recent success of denoising diffusion probabilistic models (DDPMs) in image synthesis [54, 59] and propose an effective data hallucination strategy tailored to appearance modeling. In particular, from a static image, we generate both simulated positive and negative instances along with random background perturbations.", + "bbox": [ + 75, + 666, + 470, + 800 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The main contributions are summarized as follows:", + "bbox": [ + 96, + 801, + 434, + 815 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We define the task of open-vocabulary MOT and provide a suitable benchmark setting on the large-scale, large-vocabulary MOT benchmark TAO [9].", + "2. We develop OVTrack, the first open-vocabulary multi-object tracker. It leverages vision-language models to" + ], + "bbox": [ + 89, + 819, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/dc380e7f66d95f62446cb991af7fa3aa0f494037ef2d7b4804959ed3bc5610ad.jpg", + "image_caption": [ + "Figure 2. OVTrack qualitative results. We condition our tracker on text prompts unseen during training, namely 'heron', 'hippo' and 'drone', and successfully track the corresponding objects in the videos. The box color depicts object identity." + ], + "image_footnote": [], + "bbox": [ + 506, + 90, + 890, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "improve both classification and association compared to closed-set trackers.", + "bbox": [ + 531, + 371, + 890, + 400 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. We propose an effective data hallucination strategy that allows us to address the data availability problem in open-vocabulary settings by leveraging DDPMs.", + "bbox": [ + 511, + 407, + 890, + 454 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Owing to its thoughtful design, OVTrack sets a new state-of-the-art on the challenging TAO benchmark [9], outperforming existing trackers by a significant margin while being trained on static images only. In addition, OVTrack is capable of tracking arbitrary object classes (see Fig. 2), overcoming the limitation of closed-set trackers.", + "bbox": [ + 496, + 465, + 893, + 555 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 500, + 571, + 635, + 588 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multiple object tracking. The dominant paradigm in MOT literature is tracking-by-detection [53], where objects are first detected in each frame, and subsequently associated across time. Thus, many works have focused on data association, aiming to exploit similarity cues such as visual appearance [3, 16, 32, 42, 50, 60, 65, 69], 2D object motion [4,5,15,26,68] or 3D object motion [25,40,44,47,48,62] most effectively. Recently, researchers have focused on learning data association with graph neural networks [6, 61] or transformers [41,63,73,78]. However, those works dismiss a more profound problem in the tracking-by-detection pipeline that precedes data association: Contemporary object detectors [23, 36, 55-57] are designed for closed-set scenarios where all objects appear frequently in the training and testing data distributions. Hence, Dave et al. [9] proposed a new benchmark, TAO, that focuses on studying MOT in the long-tail of the object category distribution. On this benchmark, GTR [78], AOA [12], QDTrack [16] and TET [34] achieve impressive performance. However, those works are still limited to pre-defined object categories and thus do not scale", + "bbox": [ + 496, + 598, + 893, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "5568", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "to the diversity of real-world settings. Our work enables tracking of unseen classes from an open vocabulary.", + "bbox": [ + 75, + 90, + 468, + 121 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Open-world detection and tracking. Open-world detection methods aim to detect any salient object in a given input image irrespective of its category and beyond the training data distribution in particular. However, object classification under such a setting is ill-posed since novel classes will be unknown by definition [2, 29]. As such, open-world detection methods utilize class agnostic localizers [10] and treat classification as a clustering problem [29], estimating a similarity between novel instances and grouping them into novel classes via incremental learning.", + "bbox": [ + 75, + 123, + 470, + 273 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Instead, open-vocabulary object detection methods aim to detect arbitrary, but given classes of objects at test time [72]. For this, Bansal et al. [1] connect an object detector with word representations [51]. Recently, models like CLIP [52] learn visual representations from natural language supervision. Their main advantage over word representations is better alignment of visual concepts and language description. Consequently, many works have focused on leveraging image-text representations for open-vocabulary and few-shot object detection [20, 76]. ViLD [20] distills CLIP image features, while Detic [76] leverages classification data for joint training. Other works have focused on learning good language prompts for open-vocabulary object detection [13].", + "bbox": [ + 75, + 276, + 470, + 474 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Fewer works have tackled the open-world problem in the MOT domain. Existing works perform scene segmentation and class agnostic tracking before classification [44, 46, 48] or utilize class-agnostic proposal generation [10, 49], similar to open-world detection methods. Liu et al. [37] propose an open-world tracking benchmark, TAO-OW, that evaluates class-agnostic tracking as a task that precedes classification. However, this comes with the limitation that the evaluation only captures tracker recall and no classification accuracy. Instead of dismissing classification, we pose the problem in a different way, i.e. at test time we know the novel classes we are interested in. This allows us to capture both the precision and recall of novel classes in our evaluation, while our method maintains the ability to track any object.", + "bbox": [ + 75, + 474, + 470, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Learning tracking from static images. Since labelled video data is expensive to acquire at scale, recent methods have proposed to use static images to supervise MOT methods [16, 66, 74, 77]. CenterTrack [77] proposes to learn motion offsets from static images by random translation of the input, while FairMOT [74] treats objects in a dataset of static images as unique classes to distinguish. Inspired by recent progress in self-supervised representation learning [7, 22, 45], Fischer et al. [16] propose to utilize data augmentation in combination with a contrastive learning objective to learn appearance-based tracking from static images. We go beyond classic data augmentation used in existing works by generating positive and negative examples of objects along with background perturbations via generative", + "bbox": [ + 75, + 688, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "models, offering a more targeted approach to guiding appearance similarity learning from static images.", + "bbox": [ + 498, + 90, + 893, + 121 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Data generation for tracking. While aforementioned methods alleviate the data availability problem in MOT, there is still room for improvement when it comes to data generation in video tasks. Therefore, a large body of research has focused on data generation strategies that can benefit tracking methods [8, 14, 17, 28, 30, 31, 58]. While early works focused on obtaining synthetic data from computer graphics engines [14, 17, 28, 58], newer approaches combine 3D assets with generative models for improved realism [8, 31]. Fewer works have tackled data generation with generative models only [30]. Recently, DDPMs [27, 54, 59] showed impressive results in image synthesis. We leverage their data generation fidelity to address the data availability problem that is particularly pronounced in open-vocabulary MOT with a novel data hallucination strategy tailored to appearance modeling.", + "bbox": [ + 496, + 121, + 893, + 348 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Open-Vocabulary MOT", + "text_level": 1, + "bbox": [ + 498, + 361, + 723, + 378 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In real-world scenarios, object categories follow a long-tailed distribution with a rich vocabulary. The remarkable diversity of the open world cannot be covered by a monolithic dataset. However, existing MOT benchmarks focus on closed-set evaluation, with often only a handful of object classes being evaluated. Furthermore, the task setup requires trackers to only track objects within a small set of training categories. To bridge the gap between existing MOT benchmarks and algorithms and real-world settings, we propose the task of open-vocabulary MOT and define its training and evaluation setup as follows.", + "bbox": [ + 496, + 387, + 893, + 551 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "At training time, we train a tracker $M$ on the training data distribution $\\mathcal{D}^{\\mathrm{train}} = \\{\\mathbf{X}^{\\mathrm{train}},\\mathcal{A}^{\\mathrm{train}}\\}$ that contains video sequences $\\mathbf{X}^{\\mathrm{train}}$ and their respective annotations $\\mathcal{A}^{\\mathrm{train}}$ of objects with semantic categories $\\mathcal{C}^{\\mathrm{base}}\\subset \\mathbb{N}$ . Each annotation $\\alpha \\in \\mathcal{A}^{\\mathrm{train}}$ consists of a set of states $\\{\\alpha_{t}\\}_{t\\in T}$ for each frame $t\\in T$ that the object is visible in. A state $\\alpha_{t} = (\\mathbf{b}_{t},c_{t})$ comprises the object class $c\\in \\mathbb{N}$ and the 2D bounding box $\\mathbf{b} = [x,y,w,h]$ , where $(x,y)$ is the center location in pixel coordinates and $(w,h)$ are width and height, respectively. At test time, we are given video sequences $\\mathbf{X}^{\\mathrm{test}}$ and a set of object classes $\\mathcal{C}^{\\mathrm{novel}}\\subset \\mathbb{N}\\setminus \\mathcal{C}^{\\mathrm{base}}$ that we are interested in. We aim to find all tracks $\\mathcal{T}$ of objects in $\\mathbf{X}^{\\mathrm{test}}$ belonging to classes $\\mathcal{C}^{\\mathrm{base}}\\cup \\mathcal{C}^{\\mathrm{novel}}$ . Each track state $\\tau_{t} = (\\mathbf{b}_{t},p_{t},c_{t})\\in \\mathcal{T}$ contains predicted object confidence $p\\in [0,1]$ , class $c\\in \\mathbb{N}$ and 2D bounding box $\\mathbf{b} = [x,y,w,h]$ . The important distinctions to closed-set tracking are two-fold: 1) We evaluate the tracker $M$ not only on $\\mathcal{C}^{\\mathrm{base}}$ but also on $\\mathcal{C}^{\\mathrm{novel}}$ with $\\mathcal{C}^{\\mathrm{novel}}\\cap \\mathcal{C}^{\\mathrm{base}} = \\emptyset$ , and 2) While $\\mathcal{C}^{\\mathrm{novel}}$ is known at test time, our setup requires the tracker $M$ to track arbitrary object classes $c\\in \\mathbb{N}$ since $\\mathcal{C}^{\\mathrm{novel}}$ remains unknown at training time. In particular, the evaluation of $\\mathcal{C}^{\\mathrm{novel}}$ illustrates the ability of tracker $M$ to track any unknown class $c\\in \\mathbb{N}\\setminus \\mathcal{C}^{\\mathrm{base}}$ , while the classes in $\\mathcal{C}^{\\mathrm{novel}}$ serve as proxy.", + "bbox": [ + 496, + 553, + 895, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "5569", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/cdb9b574ef41e6b3717b075a9a0509d1fc1a49d112ced4c0a08dd85b4963cbd3.jpg", + "image_caption": [ + "Figure 3. OVTrack training. From a single static $I_{\\mathrm{key}}$ , we generate $I_{\\mathrm{ref}}$ with our data hallucination strategy. We extract RoIs via a RPN [57] and perform knowledge distillation from CLIP [52] via the embeddings of the text and image heads. Note that we train classification only on $C^{\\mathrm{base}}$ . Further, we obtain appearance embeddings from the tracking head and apply our instance similarity loss on the image pair." + ], + "image_footnote": [], + "bbox": [ + 81, + 80, + 893, + 296 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Benchmark", + "text_level": 1, + "bbox": [ + 76, + 362, + 205, + 376 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We utilize the large-scale, large-vocabulary MOT dataset TAO [9] to establish a suitable benchmark for open-vocabulary MOT. TAO mostly follows the taxonomy of LVIS [21], which divides classes according to their occurrence into frequent, common and rare classes. To obtain our held-out set $\\mathcal{C}^{\\mathrm{novel}}$ , we follow open-vocabulary detection literature [20] and use the rare classes as defined by LVIS. The intuition behind this is that the occurrence of rare classes is correlated with uncommon scenarios and events that we are particularly interested in evaluating.", + "bbox": [ + 75, + 387, + 470, + 537 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With respect to evaluation, the advantage of defining $C^{\\mathrm{novel}}$ is that we can apply closed-set tracking metrics in a straightforward manner, while open-world MOT [37] needs to resort to recall-based evaluation. Further, previous works [34, 37] have shown that the official evaluation metric in TAO, Track mAP [70], is sub-optimal in terms of handling FPs in presence of missing annotations. On the contrary, the recently proposed TETA metric [34] handles this shortcoming via local cluster evaluation. Also, TETA disentangles classification from localization and association performance. Thus, we choose TETA as the evaluation metric for our setup to provide a comprehensive insight into the localization, association, and open-vocabulary classification performance of tracker $M$ .", + "bbox": [ + 75, + 539, + 473, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. OVTrack", + "text_level": 1, + "bbox": [ + 76, + 767, + 181, + 782 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We present our Open-Vocabulary Tracker, OVTrack. We address two perspectives of its design: 1) Model perspective: We show how to handle the open-vocabulary setting in the localization, classification, and association modules of the tracker in Section 4.1; 2) Data perspective: Collecting and annotating the necessary amount of training videos is impractical for open-vocabulary MOT. Therefore, we contribute a", + "bbox": [ + 75, + 794, + 470, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "novel training approach for learning object tracking without video data in Section 4.2.", + "bbox": [ + 498, + 363, + 892, + 393 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Model design", + "text_level": 1, + "bbox": [ + 500, + 405, + 638, + 421 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We decompose OVTrack's functionality into localization, classification and association and discuss our open-vocabulary design philosophy in tackling difficulties for each of those parts. The model design is illustrated in Fig. 3.", + "bbox": [ + 496, + 429, + 893, + 489 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1) Localization: To localize objects of arbitrary and possibly unknown classes $c \\in \\mathbb{N}$ in a video, we train Faster R-CNN [57] in a class-agnostic manner, i.e. we use only the RPN and regression losses defined in [57]. We find that this localization procedure can generalize well to object classes that are unknown at training time, as also validated by previous works [10, 20, 76]. During training, we use RPN proposals as object candidates $P$ for greater diversity, while during inference, we use the refined RCNN outputs as object candidates. Each candidate $r \\in P$ is defined by confidence $p_r$ and bounding box $\\mathbf{b}_r$ .", + "2) Classification: Existing closed-set trackers [3,4,16,77] can only track objects of categories in $\\mathcal{C}^{\\mathrm{base}}$ , i.e. objects present and annotated in the training data distribution $\\mathcal{D}^{\\mathrm{train}}$ . To enable open-vocabulary classification, we need to be able to configure the classes we are interested in without re-training. Inspired by open-vocabulary detection literature [1], we connect our Faster R-CNN with the vision-language model CLIP [52] that has been pre-trained on over 400 million image-text pairs for contrastive learning." + ], + "bbox": [ + 496, + 489, + 893, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After extracting the RoI feature embeddings $\\mathbf{f}_r = \\mathcal{R}(\\phi(I), \\mathbf{b}_r), \\forall r \\in P$ from the backbone $\\phi$ , we replace the original classifier in Faster R-CNN with a text head and add an image head generating the embeddings $\\hat{\\mathbf{t}}_r$ and $\\hat{\\mathbf{i}}_r$ , for each $\\mathbf{f}_r$ . We use the CLIP text and image encoders to supervise the heads following [13, 20]. In particular, we use the class names to generate text prompts $\\mathcal{P}(c) = \\{\\mathbf{v}_1^c, \\dots, \\mathbf{v}_L^c, \\mathbf{w}_c\\}$", + "bbox": [ + 496, + 795, + 893, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "5570", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0f6b8be833e232ee13125ff5a83561aa957cd3a3d12ff7efa1a326edf3612bd3.jpg", + "image_caption": [ + "Figure 4. OVTrack inference. Given an input video stream, we track objects of arbitrary classes, e.g. $\\mathcal{C}^{\\mathrm{base}}\\cup \\mathcal{C}^{\\mathrm{novel}}$ . We first localize objects agnostic of their class, then assign a semantic class label via the text embedding head, and finally associate them to existing tracks by comparing their appearance embedding $\\mathbf{q}$ obtained from the tracking head to the appearance embeddings in the track memory." + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 893, + 200 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "that consist of $L$ context vectors $\\mathbf{v}^c$ and a class name embedding $\\mathbf{w}_c$ . We feed the prompts into the CLIP text encoder $\\mathcal{E}$ , generating text embeddings $\\mathbf{t}_c = \\mathcal{E}(\\mathcal{P}(c)), \\forall c \\in \\mathcal{C}^{\\mathrm{base}}$ . We compute the affinity between the predicted embeddings $\\hat{\\mathbf{t}}_r$ and their CLIP counterpart $\\mathbf{t}_c$ .", + "bbox": [ + 75, + 268, + 472, + 345 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {z} (r) = \\left[ \\cos \\left(\\hat {\\mathbf {t}} _ {r}, \\mathbf {t} _ {b g}\\right), \\cos \\left(\\hat {\\mathbf {t}} _ {r}, \\mathbf {t} _ {1}\\right), \\dots , \\cos \\left(\\hat {\\mathbf {t}} _ {r}, \\mathbf {t} _ {| \\mathcal {C} ^ {\\text {b a s e}} |}\\right) \\right] \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 352, + 470, + 371 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t e x t}} = \\frac {1}{| P |} \\sum_ {r \\in P} \\mathcal {L} _ {\\mathrm {C E}} (\\operatorname {s o f t m a x} (\\mathbf {z} (r) / \\lambda), c _ {r}), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 373, + 470, + 409 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\cos (\\mathbf{v},\\mathbf{k}) = \\frac{\\mathbf{v}\\cdot\\mathbf{k}}{||\\mathbf{v}||||\\mathbf{k}||}$ , $\\mathbf{t}_{bg}$ a learned background prompt, $\\lambda$ a temperature parameter, $\\mathcal{L}_{\\mathrm{CE}}$ the cross-entropy loss and $c_{r}$ is the class label of $r$ . Furthermore, we align each $\\hat{\\mathbf{i}}_r$ with the CLIP image encoder $\\mathcal{I}$ . For each $r$ , we crop the input image to $\\mathbf{b}_r$ , and resize it to the required input size to obtain the image embedding $\\mathbf{i}_r = \\mathcal{I}(\\mathcal{R}(I,\\mathbf{b}_r))$ . We minimize the distance between the corresponding $\\hat{\\mathbf{i}}_r$ and $\\mathbf{i}_r$ .", + "bbox": [ + 75, + 417, + 472, + 527 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {i m a g e}} = \\frac {1}{| P |} \\sum_ {r \\in P} \\left\\| \\hat {\\mathbf {i}} _ {r} - \\mathbf {i} _ {r} \\right\\| _ {1}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 173, + 534, + 468, + 570 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3) Association: An open-vocabulary tracker should handle diverse scenarios that comprise complex camera motion and heterogeneous object motion patterns. However, those patterns are difficult to model especially when there are not enough video annotations available [16, 37]. Therefore, we rely on appearance cues to robustly track objects in an open-vocabulary context. Specifically, we employ a contrastive learning approach inspired by [16, 34]. Given an image pair $(I_{\\mathrm{key}}, I_{\\mathrm{ref}})$ we extract RoIs from both images and match the RoIs to the annotations using intersection-over-union (IoU). For each matched RoI in $I_{\\mathrm{key}}$ with appearance embedding $\\mathbf{q} \\in Q$ , we cluster objects $Q^{+}$ with the same identity and divide objects $Q^{-}$ with different identity in $I_{\\mathrm{ref}}$ .", + "bbox": [ + 75, + 577, + 472, + 773 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {P o s D} (\\mathbf {q}) = \\frac {1}{| Q ^ {+} (\\mathbf {q})) |} \\sum_ {\\mathbf {q} ^ {+} \\in Q ^ {+}} \\exp \\left(\\mathbf {q} \\cdot \\mathbf {q} ^ {+}\\right) / \\tau), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 780, + 468, + 818 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {S i m} (\\mathbf {q}) = \\frac {\\exp \\left(\\mathbf {q} \\cdot \\mathbf {q} ^ {+} / \\tau\\right)}{\\operatorname {P o s D} (\\mathbf {q}) + \\sum_ {\\mathbf {q} ^ {-} \\in Q ^ {-}} \\exp \\left(\\mathbf {q} \\cdot \\mathbf {q} ^ {-} / \\tau\\right)}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 821, + 468, + 857 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t r a c k}} = - \\sum_ {\\mathbf {q} \\in Q} \\frac {1}{| Q ^ {+} (\\mathbf {q}) |} \\sum_ {\\mathbf {q} ^ {+} \\in Q ^ {+} (\\mathbf {q})} \\log \\left(\\operatorname {S i m} \\left(\\mathbf {q} ^ {+}\\right)\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 859, + 468, + 897 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We further apply an auxiliary loss $\\mathcal{L}_{\\mathrm{aux}}$ to constrain the magnitude of the logits following [16].", + "bbox": [ + 498, + 268, + 893, + 299 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During inference, we use straightforward appearance feature similarity for associating existing tracks $\\mathcal{T}$ with objects in $P$ . In particular, for each track $\\tau \\in \\mathcal{T}$ and its corresponding appearance embedding $\\mathbf{q}_{\\tau}$ , we compare its similarity with all candidate objects $r \\in P$ using appearance embedding $\\mathbf{q}_r$ . We measure the similarity $\\mathbf{s}(\\tau, r)$ of existing tracks with the candidate objects using both bi-directional softmax [16] and cosine similarity. We assign $r$ to the track $\\tau$ with its maximum similarity, if $\\mathbf{s}(\\tau, r) > \\beta$ . If $r$ does not have a matching track, it starts a new track if its confidence $p_r > \\gamma$ and is discarded otherwise. The inference pipeline is illustrated in Fig. 4.", + "bbox": [ + 496, + 301, + 893, + 482 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Learning to track without video data", + "text_level": 1, + "bbox": [ + 500, + 496, + 818, + 512 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we focus on how to train OVTrack in an open-vocabulary context. In particular, open-vocabulary MOT is challenging from a data perspective since we need to localize, classify, and associate possibly unknown objects of extremely diverse appearance with fixed method components. Thus, it is of fundamental importance to align our training data distribution $\\mathcal{D}^{\\mathrm{train}}$ with the conditions found in evaluation. However, existing video datasets lack the diversity of contemporary image datasets. Hence, it is essential for open-vocabulary trackers to leverage not only video but more importantly static image data during training.", + "bbox": [ + 496, + 521, + 893, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use the large-scale, diverse image dataset LVIS [21] to train OVTrack. In particular, for each image $I_{\\mathrm{key}}$ we generate a reference image $I_{\\mathrm{ref}}$ . Referring to our instance similarity loss in Eq. 6, the appearance similarity learning is constituted by contrasting positive and negative examples $\\mathbf{q}^{+}$ and $\\mathbf{q}^{-}$ . The corollary of this is that learning will be optimal if $\\mathbf{q}^{+}$ consists of examples with distortions commonly encountered in video data, such as change in object scale, viewpoint or lighting, while $\\mathbf{q}^{-}$ contains examples with appearance changes associated with object identity, such as different material. While distortions like translation, scaling, and rotation can be simulated via classic data augmentation strategies [16, 74, 77], there are certain phenomena like changes in object viewpoint, lighting or context that cannot", + "bbox": [ + 496, + 688, + 895, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5571", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f0d1096e6bcba4323d732a4a65eda141184a4087260d160b51dd1be8db4a2778.jpg", + "image_caption": [ + "Figure 5. Data hallucination strategy. Given an input image, its annotations and its caption, we generate $x_0 \\sim n$ and input it to the diffusion model [59], which progressively denoises $x$ from $\\delta_0 = 0.75$ to $\\eta$ . At the same time, the foreground regions are kept fixed at each iteration. Specifically, we compose the generated images $x_k \\sim m_\\theta$ with foreground regions $\\hat{x}_k \\sim n$ at the current noise level, to obtain $x_k$ as input for the next iteration. Finally, we arrive at $I_{\\mathrm{ref}}$ which we use for instance similarity learning." + ], + "image_footnote": [], + "bbox": [ + 80, + 92, + 888, + 277 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "be simulated by these.", + "bbox": [ + 76, + 361, + 225, + 375 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Therefore, to simulate all desired properties of our instance embedding space in $I_{\\mathrm{ref}}$ , we combine classic data augmentations with a DDPM-based data hallucination strategy. We use the data generation fidelity of stable diffusion [59] to simulate $\\mathbf{q}^{+}$ and $\\mathbf{q}^{-}$ via a specialized denoising process. Generally, the denoising process of a DPPM can be viewed as the inversion of a forward process that maps an input $x$ to Gaussian white noise $\\mathcal{N}(0,\\mathbf{I})$ . In the forward direction, Gaussian noise is added to the input image according to a variance schedule $\\delta_{k}$ in $K$ steps.", + "bbox": [ + 75, + 376, + 473, + 527 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nn \\left(x _ {k} \\mid x _ {k - 1}\\right) = \\mathcal {N} \\left(x _ {k}; \\sqrt {1 - \\delta_ {k}} \\cdot x _ {k - 1}, \\delta_ {k} \\mathbf {I}\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 122, + 540, + 468, + 559 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the backward direction, a neural network with parameters $\\theta$ predicts the parameters $\\mu$ and $\\Sigma$ of a Gaussian distribution that reverses a forward step.", + "bbox": [ + 76, + 571, + 468, + 617 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nm _ {\\theta} \\left(x _ {k - 1} \\mid x _ {k}\\right) = \\mathcal {N} \\left(x _ {k - 1}; \\mu_ {\\theta} \\left(x _ {k}, k\\right), \\Sigma_ {\\theta} \\left(x _ {k}, k\\right)\\right). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 628, + 468, + 646 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our specialized denoising process is illustrated in Fig. 5. We initialize $I_{\\mathrm{ref}}$ as $I_{\\mathrm{key}}$ and apply a random geometric transformation to $I_{\\mathrm{ref}}$ . Next, we use the instance mask annotations in LVIS to define the set of positive examples $A^{+}$ . We divide each iteration of the denoising process of image $I_{\\mathrm{ref}}$ using the union of all object masks in $A^{+}$ into two branches following [39]. In addition, we use the conditioning mechanism in [59] to guide the backward process with the corresponding image caption. To initialize the backward process, we set $x_0$ to $I_{\\mathrm{ref}}$ at $\\delta_0 = 0.75$ via the forward process $n(x_k|x_{k-1})$ . Note that $x$ corresponds to a latent representation of $I_{\\mathrm{ref}}$ obtained via the encoder of stable diffusion. In each iteration, we apply $m_\\theta(x_{k-1}|x_k)$ to obtain a new sample $x_k \\sim m_\\theta$ . At the same time, we use the forward process $n(x_k|x_{k-1})$ on the areas of $A^{+}$ to generate $\\hat{x}_k \\sim n$ . At the end of each reverse iteration, we compose the two versions via", + "bbox": [ + 75, + 657, + 472, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$x_{k} = A^{+}\\cdot \\hat{x}_{k} + (1 - A^{+})\\cdot x_{k}$ . We iterate until $\\delta_k\\leq \\eta$ Finally, we apply $m_{\\theta}(x_{k - 1}|x_k)$ to the whole image, without branching, as a homogenization step between $A^{+}x_{k}$ and $(1 - A^{+})x_{k}$ , while $\\eta >\\delta_k > 0$", + "bbox": [ + 496, + 359, + 893, + 422 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "By this process, we achieve three goals. First, we generate random perturbations of the background. Second, we keep the areas of $A^{+}$ close to its original content in each denoising step so that positive instances are integrated well into the new background. Third, we generate distractor objects by caption guided hallucination.", + "bbox": [ + 496, + 422, + 892, + 513 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 500, + 532, + 633, + 549 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Evaluation metrics", + "text_level": 1, + "bbox": [ + 500, + 558, + 681, + 571 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "TETA. The tracking-everything accuracy (TETA) [34] is calculated from three independent scores. First, the localization accuracy (LocA) is calculated by matching all annotated boxes $\\alpha$ to the predicted boxes of $\\mathcal{T}$ without taking classification into account: $\\mathrm{LocA} = \\frac{|\\mathrm{TPL}|}{|\\mathrm{TPL}| + |\\mathrm{FPL}| + |\\mathrm{FNL}|}$ . Next, classification accuracy (ClsA) is computed based on all well-localized TPL, comparing the predicted semantic classes to the matched ground-truths: $\\mathrm{ClsA} = \\frac{|\\mathrm{TPC}|}{|\\mathrm{TPC}| + |\\mathrm{FPC}| + |\\mathrm{FNC}|}$ . Finally, association accuracy (AssocA) is computed in a similar fashion, comparing the identity of associated ground truths with well-localized predictions: AssocA = $\\frac{1}{|\\mathrm{TPL}|}\\sum_{b\\in \\mathrm{TPL}}\\frac{|\\mathrm{TPA}(b)|}{|\\mathrm{TPA}(b)| + |\\mathrm{FPA}(b)| + |\\mathrm{FNA}(b)|}$ . The TETA score is computed as the arithmetic mean of the three scores.", + "bbox": [ + 496, + 582, + 893, + 787 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Track mAP. The Track mAP [70] is calculated using the 3D IoU between the bounding boxes of a predicted track $\\tau$ and an annotated track $\\alpha$ by $\\mathrm{IoU}_{\\mathrm{3D}}(\\tau, \\alpha) = \\frac{\\sum_{t \\in T} \\tau_t \\cap \\alpha_t}{\\sum_{t \\in T} \\tau_t \\cup \\alpha_t}$ . It is used analogous to 2D bounding box IoU to calculate the popular average precision metric per class as in [21]. The Track mAP is the average of the per-class scores across a set of IoU3D thresholds.", + "bbox": [ + 496, + 789, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "5572", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/36c604f62529a7ae52773518b5bef973c64818e93e56b1c266d71a92448fd3bc.jpg", + "table_caption": [ + "Table 1. Open-vocabulary MOT comparison. We compare our method with existing closed-set trackers and off-the-shelf open-vocabulary baselines on base and novel classes on the validation and test sets of TAO [9]. We indicate the classes and data the methods trained on. Note that methods using TAO data utilize videos for training. All methods use ResNet50 [18] as backbone." + ], + "table_footnote": [], + "table_body": "
MethodClassesDataBaseNovel
Validation setBaseNovelCC3MLVISTAOTETALocAAssocAClsATETALocAAssocAClsA
QDTrack [16]✓✓-✓✓27.145.624.711.022.542.724.40.4
TETer [34]✓✓-✓✓30.347.431.612.125.745.931.10.2
DeepSORT (ViLD) [65]✓--✓✓26.947.115.817.721.146.414.72.3
Tracktor++ (ViLD) [3]✓--✓✓28.347.420.517.022.746.719.32.2
OVTrack✓--✓-35.549.336.920.227.848.833.61.5
RegionCLIP [75]
+ DeepSORT [65]✓-✓✓✓28.452.515.617.024.549.215.39.0
+ Tracktor++ [3]✓-✓✓✓29.652.419.616.925.750.118.98.1
+ OVTrack✓-✓✓-36.353.936.318.732.051.433.211.4
Test setBaseNovelCC3MLVISTAOTETALocAAssocAClsATETALocAAssocAClsA
QDTrack [16]✓✓-✓✓25.843.223.510.620.239.720.90.2
TETer [34]✓✓-✓✓29.244.030.410.721.739.125.90.0
DeepSORT (ViLD) [65]✓--✓✓24.543.814.615.217.238.411.61.7
Tracktor++ (ViLD) [3]✓--✓✓26.044.119.014.818.039.013.41.7
OVTrack✓--✓-32.645.635.416.924.141.828.71.8
RegionCLIP [75]
+ DeepSORT [65]✓-✓✓✓27.049.815.116.118.741.89.15.2
+ Tracktor++ [3]✓-✓✓✓28.049.418.815.720.042.412.05.7
+ OVTrack✓-✓✓-34.851.136.117.325.744.826.26.1
", + "bbox": [ + 158, + 132, + 810, + 386 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/bab4e8d00462206cd00f4a5e657ae82b2e728619bfadca0c9258cfecf1afd66d.jpg", + "table_caption": [ + "Table 2. Closed-set MOT Track mAP comparison. We compare to existing trackers on TAO [9] validation. Competing methods use ResNet101 [24], we use ResNet50 as backbone. All methods use Faster R-CNN [57]. We include results with stronger detectors and additional data in gray. $\\dagger$ does not use videos for training." + ], + "table_footnote": [], + "table_body": "
MethodTrack mAP50Track mAP75Track mAP
SORT-TAO [9]13.2--
QDTrack [16]15.95.010.6
GTR† [78]20.4--
TAC [67]17.75.807.30
BIV [66]19.67.3013.6
OVTrack†21.210.615.9
GTR + CenterNet2† [78]22.5--
AOA [12]25.8--
", + "bbox": [ + 88, + 474, + 457, + 595 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Implementation details", + "text_level": 1, + "bbox": [ + 76, + 616, + 290, + 633 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We use ResNet50 [24] with FPN [35]. We filter object candidates $P$ by non-maximum suppression (NMS) with an IoU threshold of 0.7 and randomly select $|P| = 256$ candidates per image. We set $\\lambda = 0.07$ in $\\mathcal{L}_{\\mathrm{text}}$ . We use a two-stage training process, first training the detection components following [13, 20], second fine-tuning the model for tracking with loss weights 0.25 for $\\mathcal{L}_{\\mathrm{track}}$ and 1.0 for $\\mathcal{L}_{\\mathrm{aux}}$ following [16]. We train on the LVIS dataset, with one hallucinated counterpart per image in the dataset. While we use the full dataset for the state-of-the-art comparison, we use a subset of 10,000 images for the ablation studies due to resource constraints. Unless otherwise noted, we use the following data augmentations in training: resizing, random horizontal flipping, color jittering, random affine transformation, and mosaic composition with varying parameters between $I_{\\mathrm{key}}$ and $I_{\\mathrm{ref}}$ . We use $\\eta = 0.02$ for data generation. For inference, we select object candidates $P$ by NMS with", + "bbox": [ + 75, + 643, + 473, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9f016f8382f1c61f4495848df41ecbb5a1a651061a8e5600ad45123799ad25b4.jpg", + "table_caption": [ + "Table 3. Closed-set MOT TETA comparison. We compare to existing trackers on the TAO [9] validation. Benchmark results are taken from [34]. All competing methods use ResNet101 [24] except AOA [12], we use ResNet50 as backbone. All methods use Faster R-CNN [57]. $\\dagger$ does not use videos for training." + ], + "table_footnote": [], + "table_body": "
MethodTETALocAAssocAClsA
SORT-TAO [9]24.848.114.312.1
Tracktor [3]24.247.413.012.1
DeepSORT [65]26.048.417.512.1
AOA [12]25.323.430.621.9
Tracktor++ [9]28.049.022.812.1
QDTrack [16]30.050.527.412.1
TETer [34]33.351.635.013.2
OVTrack†34.749.336.718.1
", + "bbox": [ + 555, + 474, + 836, + 592 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "an IoU threshold of 0.5. We keep a track memory of 10 frames to re-identify objects after occlusion and set $\\beta = 0.5$ and $\\gamma = 0.0001$ (see Sec. 4.1).", + "bbox": [ + 496, + 606, + 893, + 654 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Comparison to state-of-the-art", + "text_level": 1, + "bbox": [ + 498, + 664, + 769, + 680 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Open-vocabulary MOT. In Tab. 1, we show the open-vocabulary MOT evaluation on the TAO validation and test sets, divided into base classes $\\mathcal{C}^{\\mathrm{base}}$ and novel classes $\\mathcal{C}^{\\mathrm{novel}}$ . For details on the setup, please refer to the supplemental material. The baselines we establish are composed of both closed-set and open-vocabulary trackers. We choose the two state-of-the-art closed-set trackers, TETer [34] and QDTrack [16], trained on $\\mathcal{C}^{\\mathrm{base}} \\cup \\mathcal{C}^{\\mathrm{novel}}$ . In addition, we combine off-the-shelf trackers DeepSORT [65] and Tracktor++ [3] with the open-vocabulary detector ViLD [20] as baseline open-vocabulary trackers. These are, like OVTrack, trained on $\\mathcal{C}^{\\mathrm{base}}$ only. Note that all baselines use video data for training, while we use only static images.", + "bbox": [ + 496, + 688, + 893, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our approach substantially outperforms all closed-set and", + "bbox": [ + 517, + 885, + 890, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "5573", + "bbox": [ + 482, + 944, + 514, + 957 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "open-vocabulary baselines. We achieve consistent improvement across LocA, AssocA, and ClsA on both base and novel classes. The baselines trained on $\\mathcal{C}^{\\mathrm{base}}\\cup \\mathcal{C}^{\\mathrm{novel}}$ can, in some cases, correctly classify objects in $\\mathcal{C}^{\\mathrm{novel}}$ but achieve poor results. On the contrary, both the open-vocabulary baselines and our tracker achieve significantly higher ClsA on novel classes. However, we note that classification on the TAO dataset remains a very challenging task. The absolute ClsA scores on novel classes are low. This is partially due to the nature of the ClsA metric, which only considers top-1 classification accuracy, while classes on the TAO dataset are diverse and fine-grained.", + "bbox": [ + 75, + 90, + 472, + 272 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Therefore, we investigate the use of stronger, recently proposed open-vocabulary detectors. We combine RegionCLIP [75] with our off-the-shelf baselines and OVTrack. We replace the localization and classification parts of OVTrack with RegionCLIP while keeping the association fixed. We observe that ClsA increases substantially for all trackers on novel classes. Our method achieves the best performance by a wide margin and achieves the best ClsA scores with 11.4 and 6.1 on the validation and test sets, respectively. Note however that RegionCLIP makes use of additional data.", + "bbox": [ + 75, + 273, + 472, + 422 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Closed-set MOT. In Tab. 2 and Tab. 3 we compare to existing works on the validation split of TAO using Track mAP and TETA metrics, respectively. Note that our method neither uses video data for training, nor is it trained on rare classes as defined in Sec. 3.1, while all of the compared closed-set trackers train on video data and use the held-out rare classes for training as they are part of the closed-set evaluation in TAO. We outperform all previous works by a sizable margin on both metrics. By examining the TETA scores in Tab. 3, we observe that our tracker obtains 2.3 points less in LocA compared to TETer [34]. However, our approach beats TETer in terms of AssocA by 1.7 points and greatly improves in ClsA by 4.9 points, illustrating the positive effect of CLIP distillation on both classification and associated compared to closed-set trackers. This validates our design in Sec. 4.1. Note that while AOA [12] has a better ClsA, it ensembles multiple few-shot detection and re-identification models trained on additional datasets as reported by previous works [34, 78]. Overall, our approach surpasses the previous state-of-the-art by 1.4 points in TETA and 2.3 points in Track mAP while using a weaker backbone and the same detector.", + "bbox": [ + 75, + 425, + 472, + 757 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Ablation studies", + "text_level": 1, + "bbox": [ + 76, + 768, + 238, + 784 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "CLIP knowledge distillation. In Tab. 4 we analyze the effect of the knowledge distillation described in Sec. 4.1. In particular, we observe that using both $\\mathcal{L}_{\\mathrm{text}}$ and $\\mathcal{L}_{\\mathrm{image}}$ is more effective for classification than using only $\\mathcal{L}_{\\mathrm{text}}$ , improving ClsA significantly from 15.6 to 18.1, while LocA and AssocA stay at the same level of performance.", + "bbox": [ + 75, + 792, + 470, + 883 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Data hallucination strategy. We validate the effective", + "bbox": [ + 76, + 885, + 472, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e8c787f701f246c3b9d3724eb14d5bc3cff096815622b6c91bbac73a30e67a99.jpg", + "table_caption": [ + "Table 4. Ablation study on CLIP knowledge distillation. We show that using both $\\mathcal{L}_{\\mathrm{text}}$ and $\\mathcal{L}_{\\mathrm{image}}$ is important to classification performance when doing CLIP knowledge distillation (Sec. 4.1)." + ], + "table_footnote": [], + "table_body": "
\\( \\mathcal{L}_{\\text{text}} \\)\\( \\mathcal{L}_{\\text{text}} \\mid \\text{image} \\)TETALocAAssocAClsA
✓-34.050.535.715.6
✓✓34.349.335.418.1
", + "bbox": [ + 547, + 135, + 841, + 185 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/8a8a8292bb1812b9f7f06dad2bc6898f66af85be677b106078372f8ba294debd.jpg", + "table_caption": [ + "Table 5. Ablation study on data hallucination strategy. We show that our data hallucination strategy ('DDPM', Sec. 4.2) improves the association of a closed-set tracker [34] and our OVTrack on TAO [9] validation. We ensemble it with data augmentations, where 'Standard' refers to random resize and horizontal flip, 'Heavy' to color jitter, random affine transformation and mosaic." + ], + "table_footnote": [], + "table_body": "
StandardDDPMHeavyTETALocAAssocAClsA
TETer-SwinT
✓--32.350.730.615.5
✓✓-33.251.233.015.4
✓✓✓34.351.435.515.8
OVTrack
✓--32.548.931.117.6
✓✓-33.348.932.918.0
✓✓✓34.449.135.718.3
", + "bbox": [ + 503, + 282, + 893, + 406 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ness of our data hallucination strategy described in Sec. 4.2 by training both the closed-set tracker TETer [34] and our OVTrack with it. Note that we choose to use SwinT [38] with TETer and ResNet50 [24] with our OVTrack to achieve similar performance, in order to fairly compare the performance difference on both trackers. Tab. 5 shows the TETA results on the TAO validation set. We observe that our data hallucination strategy improves the AssocA significantly for both trackers, while LocA and ClsA are comparable. In particular, we improve 2.4 and 1.8 points in AssocA for TETer and OVTrack, respectively. Further, ensembling our data generation strategy with heavy data augmentations yields another 2.5 and 1.8 points improvement. Overall, we show that our data generation strategy improves instance similarity learning across both closed-set and open-vocabulary trackers while being complementary to classic data augmentation.", + "bbox": [ + 496, + 421, + 893, + 664 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 678, + 619, + 694 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work introduced open-vocabulary MOT as an effective solution to evaluating multi-object trackers beyond pre-defined training categories. We defined a suitable benchmark setting and presented OVTrack, a data-efficient open-vocabulary tracker. By using knowledge distillation from vision-language models, we improve tracking while going beyond limited dataset taxonomies. In addition, we put forth a data hallucination strategy tailored to instance similarity learning that addresses the data availability problem in open-vocabulary MOT. As a result, OVTrack learns tracking from static images and is able to track arbitrary objects in videos while outperforming existing trackers by a sizable margin on the large-scale, large-vocabulary TAO [9] benchmark.", + "bbox": [ + 496, + 703, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "5574", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Ankan Bansal, Karan Sikka, Gaurav Sharma, Rama Chellappa, and Ajay Divakaran. Zero-shot object detection. In ECCV, 2018. 2, 3, 4", + "[2] Abhijit Bendale and Terrance Boult. Towards open world recognition. In CVPR, 2015. 3", + "[3] Philipp Bergmann, Tim Meinhardt, and Laura Leal-Taixe. Tracking without bells and whistles. In ICCV, 2019. 1, 2, 4, 7", + "[4] Alex Bewley, Zongyuan Ge, Lionel Ott, Fabio Ramos, and Ben Upcroft. Simple online and realtime tracking. In ICIP, 2016. 1, 2, 4", + "[5] Erik Bochinski, Volker Eiselein, and Thomas Sikora. High-speed tracking-by-detection without using image information. In AVSS, 2017. 2", + "[6] Guillem Brasó and Laura Leal-Taixe. Learning a neural solver for multiple object tracking. In CVPR, 2020. 2", + "[7] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 3", + "[8] Yun Chen, Frieda Rong, Shivam Duggal, Shenlong Wang, Xinchen Yan, Sivabalan Manivasagam, Shangjie Xue, Ersin Yumer, and Raquel Urtasun. Geosim: Realistic video simulation via geometry-aware composition for self-driving. In CVPR, 2021. 3", + "[9] Achal Dave, Tarasha Khurana, Pavel Tokmakov, Cordelia Schmid, and Deva Ramanan. Tao: A large-scale benchmark for tracking any object. In ECCV, 2020. 1, 2, 4, 7, 8", + "[10] Achal Dave, Pavel Tokmakov, and Deva Ramanan. Towards segmenting anything that moves. In CVPRW, 2019. 1, 3, 4", + "[11] Patrick Dendorfer, Aljosa Osep, Anton Milan, Konrad Schindler, Daniel Cremers, Ian Reid, Stefan Roth, and Laura Leal-Taixe. Motchallenge: A benchmark for single-camera multiple target tracking. IJCV, 129(4):845-881, 2021. 1", + "[12] Fei Du, Bo Xu, Jiasheng Tang, Yuqi Zhang, Fan Wang, and Hao Li. 1st place solution to eccv-tao-2020: Detect and represent any object for tracking. arXiv preprint arXiv:2101.08040, 2021. 2, 7, 8", + "[13] Yu Du, Fangyun Wei, Zihe Zhang, Miaojing Shi, Yue Gao, and Guoqi Li. Learning to prompt for open-vocabulary object detection with vision-language model. In CVPR, 2022. 2, 3, 4, 7", + "[14] Matteo Fabbri, Guillem Brasó, Gianluca Maugeri, Orcun Cetintas, Riccardo Gasparini, Aljoša Osep, Simone Calderara, Laura Leal-Taixe, and Rita Cucchiara. Motsynth: How can synthetic data help pedestrian detection and tracking? In CVPR, 2021. 3", + "[15] Christoph Feichtenhofer, Axel Pinz, and Andrew Zisserman. Detect to track and track to detect. In ICCV, 2017. 2", + "[16] Tobias Fischer, Jiangmiao Pang, Thomas E Huang, Linlu Qiu, Haofeng Chen, Trevor Darrell, and Fisher Yu. Qdtrack: Quasi-dense similarity learning for appearance-only multiple object tracking. arXiv preprint arXiv:2210.06984, 2022. 2, 3, 4, 5, 7", + "[17] Adrien Gaidon, Qiao Wang, Yohann Cabon, and Eleonora Vig. Virtual worlds as proxy for multi-object tracking analysis. In CVPR, 2016. 3" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[18] Shang-Hua Gao, Ming-Ming Cheng, Kai Zhao, Xin-Yu Zhang, Ming-Hsuan Yang, and Philip Torr. Res2net: A new multi-scale backbone architecture. IEEE TPAMI, 43(2):652-662, 2019. 7", + "[19] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In CVPR, 2012. 1", + "[20] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. arXiv preprint arXiv:2104.13921, 2021. 2, 3, 4, 7", + "[21] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In CVPR, 2019. 4, 5, 6", + "[22] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, 2020. 3", + "[23] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, 2017. 2", + "[24] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 7, 8", + "[25] David Held, Jesse Levinson, and Sebastian Thrun. Precision tracking with sparse 3d and dense color 2d data. In ICRA, 2013. 2", + "[26] David Held, Sebastian Thrun, and Silvio Savarese. Learning to track at 100 FPS with deep regression networks. In ECCV, 2016. 2", + "[27] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 2020. 3", + "[28] Hou-Ning Hu, Yung-Hsu Yang, Tobias Fischer, Trevor Darrell, Fisher Yu, and Min Sun. Monocular quasi-dense 3d object tracking. IEEE TPAMI, 2022. 3", + "[29] KJ Joseph, Salman Khan, Fahad Shahbaz Khan, and Vineeth N Balasubramanian. Towards open world object detection. In CVPR, 2021. 3", + "[30] Anna Khoreva, Rodrigo Benenson, Eddy Ilg, Thomas Brox, and Bernt Schiele. Lucid data dreaming for video object segmentation. *IJCV*, 127(9):1175–1197, 2019. 3", + "[31] Seung Wook Kim, Jonah Philion, Antonio Torralba, and Sanja Fidler. Drivegan: Towards a controllable high-quality neural simulation. In CVPR, 2021. 3", + "[32] Laura Leal-Taixe, Cristian Canton-Ferrer, and Konrad Schindler. Learning by tracking: Siamese CNN for robust target association. In CVPRW, 2016. 2", + "[33] Bastian Leibe, Ales Leonardis, and Bernt Schiele. Robust object detection with interleaved categorization and segmentation. IJCV, 77(1):259-289, 2008. 1", + "[34] Siyuan Li, Martin Danelljan, Henghui Ding, Thomas E Huang, and Fisher Yu. Tracking every thing in the wild. In ECCV, 2022. 2, 4, 5, 6, 7, 8", + "[35] Tsung-Yi Lin, Piotr Dollar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In CVPR, 2017. 7", + "[36] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C Berg. Ssd: Single shot multibox detector. In ECCV, 2016. 2" + ], + "bbox": [ + 501, + 92, + 893, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "5575", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] Yang Liu, Idil Esen Zulfikar, Jonathon Luiten, Achal Dave, Deva Ramanan, Bastian Leibe, Aljosa Osep, and Laura Leal-Taixe. Opening up open world tracking. In CVPR, 2022. 1, 3, 4, 5", + "[38] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 8", + "[39] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In CVPR, 2022. 6", + "[40] Jonathon Luiten, Tobias Fischer, and Bastian Leibe. Track to reconstruct and reconstruct to track. RA-L, 5(2):1803-1810, 2020. 2", + "[41] Tim Meinhardt, Alexander Kirillov, Laura Leal-Taixe, and Christoph Feichtenhofer. Trackformer: Multi-object tracking with transformers. In CVPR, 2022. 2", + "[42] Anton Milan, Seyed Hamid Rezatofighi, Anthony R. Dick, Ian D. Reid, and Konrad Schindler. Online multi-target tracking using recurrent neural networks. In AAAI, 2017. 2", + "[43] Anton Milan, Stefan Roth, and Konrad Schindler. Continuous energy minimization for multitarget tracking. IEEE TPAMI, 36(1):58-72, 2013. 1", + "[44] Dennis Mitzel and Bastian Leibe. Taking mobile multi-object tracking to the next level: People, unknown objects, and carried items. In ECCV, 2012. 2, 3", + "[45] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 3", + "[46] Aljoša Osep, Alexander Hermans, Francis Engelmann, Dirk Klostermann, Markus Mathias, and Bastian Leibe. Multiscale object candidates for generic object tracking in street scenes. In ICRA, 2016. 1, 3", + "[47] Aljoša Osep, Wolfgang Mehner, Markus Mathias, and Bastian Leibe. Combined image-and world-space tracking in traffic scenes. In ICRA, 2017. 2", + "[48] Aljoša Osep, Wolfgang Mehner, Paul Voigtlaender, and Bastian Leibe. Track, then decide: Category-agnostic vision-based multi-object tracking. In ICRA, 2018. 1, 2, 3", + "[49] Aljoša Osep, Paul Voigtlaender, Mark Weber, Jonathon Luiten, and Bastian Leibe. 4d generic video object proposals. In ICRA, 2020. 1, 3", + "[50] Jiangmiao Pang, Linlu Qiu, Xia Li, Haofeng Chen, Qi Li, Trevor Darrell, and Fisher Yu. Quasi-dense similarity learning for multiple object tracking. In CVPR, 2021. 2", + "[51] Jeffrey Pennington, Richard Socher, and Christopher D Manning. Glove: Global vectors for word representation. In EMNLP, 2014. 3", + "[52] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 2, 3, 4", + "[53] Deva Ramanan and David A Forsyth. Finding and tracking people from the bottom up. In CVPR, 2003. 2" + ], + "bbox": [ + 78, + 90, + 470, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[54] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, 2021. 2, 3", + "[55] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In CVPR, 2016. 2", + "[56] Joseph Redmon and Ali Farhadi. Yolo9000: better, faster, stronger. In CVPR, 2017. 2", + "[57] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In NeurIPS, 2015. 2, 4, 7", + "[58] Stephan R Richter, Vibhav Vineet, Stefan Roth, and Vladlen Koltun. Playing for data: Ground truth from computer games. In ECCV, 2016. 3", + "[59] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 3, 6", + "[60] Amir Sadeghian, Alexandre Alahi, and Silvio Savarese. Tracking the untrackable: Learning to track multiple cues with long-term dependencies. In ICCV, 2017. 2", + "[61] Samuel Schulter, Paul Vernaza, Wongun Choi, and Manmohan Chandraker. Deep network flow for multi-object tracking. In CVPR, 2017. 2", + "[62] Sarthak Sharma, Junaid Ahmed Ansari, J. Krishna Murthy, and K. Madhava Krishna. Beyond pixels: Leveraging geometry and shape cues for online multi-object tracking. In ICRA, 2018. 2", + "[63] Peize Sun, Jinkun Cao, Yi Jiang, Rufeng Zhang, Enze Xie, Zehuan Yuan, Changhu Wang, and Ping Luo. Transtrack: Multiple object tracking with transformer. arXiv preprint arXiv:2012.15460, 2020. 2", + "[64] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In CVPR, 2020. 1", + "[65] Nicolai Wojke, Alex Bewley, and Dietrich Paulus. Simple online and realtime tracking with a deep association metric. In ICIP, 2017. 2, 7", + "[66] Sanghyun Woo, Kwanyong Park, Seoung Wug Oh, In So Kweon, and Joon-Young Lee. Bridging images and videos: A simple learning framework for large vocabulary video object detection. In ECCV, 2022. 3, 7", + "[67] Sanghyun Woo, Kwanyong Park, Seoung Wug Oh, In So Kweon, and Joon-Young Lee. Tracking by associating clips. In ECCV, 2022. 7", + "[68] Bin Xiao, Haiping Wu, and Yichen Wei. Simple baselines for human pose estimation and tracking. In ECCV, 2018. 2", + "[69] Bo Yang and Ram Nevatia. An online learned CRF model for multi-target tracking. In CVPR, 2012. 2", + "[70] Linjie Yang, Yuchen Fan, and Ning Xu. Video instance segmentation. In ICCV, 2019. 2, 4, 6", + "[71] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darryll. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In CVPR, 2020. 1" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "5576", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[72] Alireza Zareian, Kevin Dela Rosa, Derek Hao Hu, and Shih-Fu Chang. Open-vocabulary object detection using captions. In CVPR, 2021. 3", + "[73] Fangao Zeng, Bin Dong, Tiancai Wang, Xiangyu Zhang, and Yichen Wei. Motr: End-to-end multiple-object tracking with transformer. arXiv preprint arXiv:2105.03247, 2021. 2", + "[74] Yifu Zhang, Chunyu Wang, Xinggang Wang, Wenjun Zeng, and Wenyu Liu. Fairmot: On the fairness of detection and re-identification in multiple object tracking. IJCV, 129(11):3069-3087, 2021. 3, 5", + "[75] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Liunian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, et al. Regionclip: Region-based language-image pretraining. In CVPR, 2022. 7, 8", + "[76] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Phillip Krahenbuhl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision. In ECCV, 2022. 2, 3, 4", + "[77] Xingyi Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Tracking objects as points. In ECCV, 2020. 3, 4, 5", + "[78] Xingyi Zhou, Tianwei Yin, Vladlen Koltun, and Philipp Krahenbihl. Global tracking transformers. In CVPR, 2022. 2, 7, 8" + ], + "bbox": [ + 78, + 90, + 470, + 416 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "5577", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/OVTrack_ Open-Vocabulary Multiple Object Tracking/092f1f8a-50aa-41bc-8e59-a89dde3c1cc2_model.json b/2023/OVTrack_ Open-Vocabulary Multiple Object Tracking/092f1f8a-50aa-41bc-8e59-a89dde3c1cc2_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2e939f8071780d40b9c8a8d1ed71ab7621694cc1 --- /dev/null +++ b/2023/OVTrack_ Open-Vocabulary Multiple Object Tracking/092f1f8a-50aa-41bc-8e59-a89dde3c1cc2_model.json @@ -0,0 +1,2257 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.209, + 0.131, + 0.764, + 0.154 + ], + "angle": 0, + "content": "OVTrack: Open-Vocabulary Multiple Object Tracking" + }, + { + "type": "text", + "bbox": [ + 0.236, + 0.181, + 0.731, + 0.253 + ], + "angle": 0, + "content": "Siyuan Li* Tobias Fischer* Lei Ke Henghui Ding Martin Danelljan Fisher Yu Computer Vision Lab, ETH Zürich https://www.vis.xyz/pub/ovtrack/" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.287, + 0.314, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.319, + 0.474, + 0.637 + ], + "angle": 0, + "content": "The ability to recognize, localize and track dynamic objects in a scene is fundamental to many real-world applications, such as self-driving and robotic systems. Yet, traditional multiple object tracking (MOT) benchmarks rely only on a few object categories that hardly represent the multitude of possible objects that are encountered in the real world. This leaves contemporary MOT methods limited to a small set of pre-defined object categories. In this paper, we address this limitation by tackling a novel task, open-vocabulary MOT, that aims to evaluate tracking beyond predefined training categories. We further develop OVTrack, an open-vocabulary tracker that is capable of tracking arbitrary object classes. Its design is based on two key ingredients: First, leveraging vision-language models for both classification and association via knowledge distillation; second, a data hallucination strategy for robust appearance feature learning from denoising diffusion probabilistic models. The result is an extremely data-efficient open-vocabulary tracker that sets a new state-of-the-art on the large-scale, large-vocabulary TAO benchmark, while being trained solely on static images." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.668, + 0.21, + 0.683 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.694, + 0.472, + 0.875 + ], + "angle": 0, + "content": "Multiple Object Tracking (MOT) aims to recognize, localize and track objects in a given video sequence. It is a cornerstone of dynamic scene analysis and vital for many real-world applications such as autonomous driving, augmented reality, and video surveillance. Traditionally, MOT benchmarks [9, 11, 19, 64, 71] define a set of semantic categories that constitute the objects to be tracked in the training and testing data distributions. The potential of traditional MOT methods [3, 4, 33, 43] is therefore limited by the taxonomies of those benchmarks. As a consequence, contemporary MOT methods struggle with unseen events, leading to a gap between evaluation performance and real-world" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.286, + 0.891, + 0.504 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.523, + 0.896, + 0.594 + ], + "angle": 0, + "content": "Figure 1. OVTrack. We approach the task of open-vocabulary multiple object tracking. During training, we leverage vision-language (VL) models both for generating samples and knowledge distillation. During testing, we track both base and novel classes unseen during training by querying a vision-language model." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.614, + 0.587, + 0.628 + ], + "angle": 0, + "content": "deployment." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.895, + 0.901 + ], + "angle": 0, + "content": "To bridge this gap, previous works have tackled MOT in an open-world context. In particular, Osep et al. [46, 48] approach generic object tracking by first segmenting the scene and performing tracking before classification. Other works have used class agnostic localizers [10, 49] to perform MOT on arbitrary objects. Recently, Liu et al. [37] defined open-world tracking, a task that focuses on the evaluation of previously unseen objects. In particular, it requires any-object tracking as a stage that precedes object classification. This setup comes with two inherent difficulties. First, in an open-world context, densely annotating all objects is prohibitively expensive. Second, without a pre-defined taxonomy of categories, the notion of what is an object is ambiguous. As a consequence, Liu et al. resort to recall-based evaluation, which is limited in two ways. Penalizing false positives (FP) becomes impossible, i.e. we cannot measure the tracker precision. Moreover, by evaluating tracking in a class-agnostic manner, we lose the ability to evaluate how" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.206, + 0.9 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5567" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.465, + 0.106 + ], + "angle": 0, + "content": "well a tracker can infer the semantic category of an object." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.107, + 0.468, + 0.257 + ], + "angle": 0, + "content": "In this paper, we propose open-vocabulary MOT as an effective solution to these problems. Similar to open-world MOT, open-vocabulary MOT aims to track multiple objects beyond the pre-defined training categories. However, instead of dismissing the classification problem and resorting to recall-based evaluation, we assume that at test time we are given the classes of objects we are interested in. This allows us to apply existing closed-set tracking metrics [34, 70] that capture both precision and recall, while still evaluating the tracker's ability to track arbitrary objects during inference." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.258, + 0.47, + 0.484 + ], + "angle": 0, + "content": "We further present the first Open-Vocabulary Tracker, OVTrack (see Fig. 1). To this end, we identify and address two fundamental challenges to the design of an open-vocabulary multi-object tracker. The first is that closed-set MOT methods are simply not capable of extending their pre-defined taxonomies. The second is data availability, i.e. scaling video data annotation to a large vocabulary of classes is extremely costly. Inspired by existing works in open-vocabulary detection [1, 13, 20, 76], we replace our classifier with an embedding head, which allows us to measure similarities of localized objects to an open vocabulary of semantic categories. In particular, we distill knowledge from CLIP [52] into our model by aligning the image feature representations of object proposals with the corresponding CLIP image and text embeddings." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.485, + 0.47, + 0.666 + ], + "angle": 0, + "content": "Beyond detection, association is the core of modern MOT methods. It is driven by two affinity cues: motion and appearance. In an open-vocabulary context, motion cues are brittle since arbitrary scenery contains complex and diverse camera and object motion patterns. In contrast, diverse objects usually exhibit heterogeneous appearance. However, relying on appearance cues requires robust representations that generalize to novel object categories. We find that CLIP feature distillation helps in learning better appearance representations for improved association. This is especially intriguing since object classification and appearance modeling are usually distinct in the MOT pipeline [3, 16, 65]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.667, + 0.471, + 0.801 + ], + "angle": 0, + "content": "Learning robust appearance features also requires strong supervision that captures object appearance changes in different viewpoints, background, and lighting. To approach the data availability problem, we utilize the recent success of denoising diffusion probabilistic models (DDPMs) in image synthesis [54, 59] and propose an effective data hallucination strategy tailored to appearance modeling. In particular, from a static image, we generate both simulated positive and negative instances along with random background perturbations." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.802, + 0.436, + 0.816 + ], + "angle": 0, + "content": "The main contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.82, + 0.47, + 0.865 + ], + "angle": 0, + "content": "1. We define the task of open-vocabulary MOT and provide a suitable benchmark setting on the large-scale, large-vocabulary MOT benchmark TAO [9]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.471, + 0.902 + ], + "angle": 0, + "content": "2. We develop OVTrack, the first open-vocabulary multi-object tracker. It leverages vision-language models to" + }, + { + "type": "list", + "bbox": [ + 0.09, + 0.82, + 0.471, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.092, + 0.892, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.295, + 0.894, + 0.351 + ], + "angle": 0, + "content": "Figure 2. OVTrack qualitative results. We condition our tracker on text prompts unseen during training, namely 'heron', 'hippo' and 'drone', and successfully track the corresponding objects in the videos. The box color depicts object identity." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.372, + 0.892, + 0.401 + ], + "angle": 0, + "content": "improve both classification and association compared to closed-set trackers." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.408, + 0.892, + 0.455 + ], + "angle": 0, + "content": "3. We propose an effective data hallucination strategy that allows us to address the data availability problem in open-vocabulary settings by leveraging DDPMs." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.466, + 0.895, + 0.556 + ], + "angle": 0, + "content": "Owing to its thoughtful design, OVTrack sets a new state-of-the-art on the challenging TAO benchmark [9], outperforming existing trackers by a significant margin while being trained on static images only. In addition, OVTrack is capable of tracking arbitrary object classes (see Fig. 2), overcoming the limitation of closed-set trackers." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.573, + 0.637, + 0.589 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Multiple object tracking. The dominant paradigm in MOT literature is tracking-by-detection [53], where objects are first detected in each frame, and subsequently associated across time. Thus, many works have focused on data association, aiming to exploit similarity cues such as visual appearance [3, 16, 32, 42, 50, 60, 65, 69], 2D object motion [4,5,15,26,68] or 3D object motion [25,40,44,47,48,62] most effectively. Recently, researchers have focused on learning data association with graph neural networks [6, 61] or transformers [41,63,73,78]. However, those works dismiss a more profound problem in the tracking-by-detection pipeline that precedes data association: Contemporary object detectors [23, 36, 55-57] are designed for closed-set scenarios where all objects appear frequently in the training and testing data distributions. Hence, Dave et al. [9] proposed a new benchmark, TAO, that focuses on studying MOT in the long-tail of the object category distribution. On this benchmark, GTR [78], AOA [12], QDTrack [16] and TET [34] achieve impressive performance. However, those works are still limited to pre-defined object categories and thus do not scale" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5568" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.122 + ], + "angle": 0, + "content": "to the diversity of real-world settings. Our work enables tracking of unseen classes from an open vocabulary." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.124, + 0.471, + 0.275 + ], + "angle": 0, + "content": "Open-world detection and tracking. Open-world detection methods aim to detect any salient object in a given input image irrespective of its category and beyond the training data distribution in particular. However, object classification under such a setting is ill-posed since novel classes will be unknown by definition [2, 29]. As such, open-world detection methods utilize class agnostic localizers [10] and treat classification as a clustering problem [29], estimating a similarity between novel instances and grouping them into novel classes via incremental learning." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.277, + 0.471, + 0.475 + ], + "angle": 0, + "content": "Instead, open-vocabulary object detection methods aim to detect arbitrary, but given classes of objects at test time [72]. For this, Bansal et al. [1] connect an object detector with word representations [51]. Recently, models like CLIP [52] learn visual representations from natural language supervision. Their main advantage over word representations is better alignment of visual concepts and language description. Consequently, many works have focused on leveraging image-text representations for open-vocabulary and few-shot object detection [20, 76]. ViLD [20] distills CLIP image features, while Detic [76] leverages classification data for joint training. Other works have focused on learning good language prompts for open-vocabulary object detection [13]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.476, + 0.471, + 0.688 + ], + "angle": 0, + "content": "Fewer works have tackled the open-world problem in the MOT domain. Existing works perform scene segmentation and class agnostic tracking before classification [44, 46, 48] or utilize class-agnostic proposal generation [10, 49], similar to open-world detection methods. Liu et al. [37] propose an open-world tracking benchmark, TAO-OW, that evaluates class-agnostic tracking as a task that precedes classification. However, this comes with the limitation that the evaluation only captures tracker recall and no classification accuracy. Instead of dismissing classification, we pose the problem in a different way, i.e. at test time we know the novel classes we are interested in. This allows us to capture both the precision and recall of novel classes in our evaluation, while our method maintains the ability to track any object." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Learning tracking from static images. Since labelled video data is expensive to acquire at scale, recent methods have proposed to use static images to supervise MOT methods [16, 66, 74, 77]. CenterTrack [77] proposes to learn motion offsets from static images by random translation of the input, while FairMOT [74] treats objects in a dataset of static images as unique classes to distinguish. Inspired by recent progress in self-supervised representation learning [7, 22, 45], Fischer et al. [16] propose to utilize data augmentation in combination with a contrastive learning objective to learn appearance-based tracking from static images. We go beyond classic data augmentation used in existing works by generating positive and negative examples of objects along with background perturbations via generative" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.894, + 0.122 + ], + "angle": 0, + "content": "models, offering a more targeted approach to guiding appearance similarity learning from static images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.122, + 0.895, + 0.349 + ], + "angle": 0, + "content": "Data generation for tracking. While aforementioned methods alleviate the data availability problem in MOT, there is still room for improvement when it comes to data generation in video tasks. Therefore, a large body of research has focused on data generation strategies that can benefit tracking methods [8, 14, 17, 28, 30, 31, 58]. While early works focused on obtaining synthetic data from computer graphics engines [14, 17, 28, 58], newer approaches combine 3D assets with generative models for improved realism [8, 31]. Fewer works have tackled data generation with generative models only [30]. Recently, DDPMs [27, 54, 59] showed impressive results in image synthesis. We leverage their data generation fidelity to address the data availability problem that is particularly pronounced in open-vocabulary MOT with a novel data hallucination strategy tailored to appearance modeling." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.362, + 0.724, + 0.38 + ], + "angle": 0, + "content": "3. Open-Vocabulary MOT" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.388, + 0.895, + 0.553 + ], + "angle": 0, + "content": "In real-world scenarios, object categories follow a long-tailed distribution with a rich vocabulary. The remarkable diversity of the open world cannot be covered by a monolithic dataset. However, existing MOT benchmarks focus on closed-set evaluation, with often only a handful of object classes being evaluated. Furthermore, the task setup requires trackers to only track objects within a small set of training categories. To bridge the gap between existing MOT benchmarks and algorithms and real-world settings, we propose the task of open-vocabulary MOT and define its training and evaluation setup as follows." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.554, + 0.897, + 0.903 + ], + "angle": 0, + "content": "At training time, we train a tracker \\(M\\) on the training data distribution \\(\\mathcal{D}^{\\mathrm{train}} = \\{\\mathbf{X}^{\\mathrm{train}},\\mathcal{A}^{\\mathrm{train}}\\}\\) that contains video sequences \\(\\mathbf{X}^{\\mathrm{train}}\\) and their respective annotations \\(\\mathcal{A}^{\\mathrm{train}}\\) of objects with semantic categories \\(\\mathcal{C}^{\\mathrm{base}}\\subset \\mathbb{N}\\). Each annotation \\(\\alpha \\in \\mathcal{A}^{\\mathrm{train}}\\) consists of a set of states \\(\\{\\alpha_{t}\\}_{t\\in T}\\) for each frame \\(t\\in T\\) that the object is visible in. A state \\(\\alpha_{t} = (\\mathbf{b}_{t},c_{t})\\) comprises the object class \\(c\\in \\mathbb{N}\\) and the 2D bounding box \\(\\mathbf{b} = [x,y,w,h]\\), where \\((x,y)\\) is the center location in pixel coordinates and \\((w,h)\\) are width and height, respectively. At test time, we are given video sequences \\(\\mathbf{X}^{\\mathrm{test}}\\) and a set of object classes \\(\\mathcal{C}^{\\mathrm{novel}}\\subset \\mathbb{N}\\setminus \\mathcal{C}^{\\mathrm{base}}\\) that we are interested in. We aim to find all tracks \\(\\mathcal{T}\\) of objects in \\(\\mathbf{X}^{\\mathrm{test}}\\) belonging to classes \\(\\mathcal{C}^{\\mathrm{base}}\\cup \\mathcal{C}^{\\mathrm{novel}}\\). Each track state \\(\\tau_{t} = (\\mathbf{b}_{t},p_{t},c_{t})\\in \\mathcal{T}\\) contains predicted object confidence \\(p\\in [0,1]\\), class \\(c\\in \\mathbb{N}\\) and 2D bounding box \\(\\mathbf{b} = [x,y,w,h]\\). The important distinctions to closed-set tracking are two-fold: 1) We evaluate the tracker \\(M\\) not only on \\(\\mathcal{C}^{\\mathrm{base}}\\) but also on \\(\\mathcal{C}^{\\mathrm{novel}}\\) with \\(\\mathcal{C}^{\\mathrm{novel}}\\cap \\mathcal{C}^{\\mathrm{base}} = \\emptyset\\), and 2) While \\(\\mathcal{C}^{\\mathrm{novel}}\\) is known at test time, our setup requires the tracker \\(M\\) to track arbitrary object classes \\(c\\in \\mathbb{N}\\) since \\(\\mathcal{C}^{\\mathrm{novel}}\\) remains unknown at training time. In particular, the evaluation of \\(\\mathcal{C}^{\\mathrm{novel}}\\) illustrates the ability of tracker \\(M\\) to track any unknown class \\(c\\in \\mathbb{N}\\setminus \\mathcal{C}^{\\mathrm{base}}\\), while the classes in \\(\\mathcal{C}^{\\mathrm{novel}}\\) serve as proxy." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5569" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.082, + 0.895, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.302, + 0.895, + 0.348 + ], + "angle": 0, + "content": "Figure 3. OVTrack training. From a single static \\( I_{\\mathrm{key}} \\), we generate \\( I_{\\mathrm{ref}} \\) with our data hallucination strategy. We extract RoIs via a RPN [57] and perform knowledge distillation from CLIP [52] via the embeddings of the text and image heads. Note that we train classification only on \\( C^{\\mathrm{base}} \\). Further, we obtain appearance embeddings from the tracking head and apply our instance similarity loss on the image pair." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.363, + 0.206, + 0.377 + ], + "angle": 0, + "content": "3.1. Benchmark" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.388, + 0.471, + 0.538 + ], + "angle": 0, + "content": "We utilize the large-scale, large-vocabulary MOT dataset TAO [9] to establish a suitable benchmark for open-vocabulary MOT. TAO mostly follows the taxonomy of LVIS [21], which divides classes according to their occurrence into frequent, common and rare classes. To obtain our held-out set \\(\\mathcal{C}^{\\mathrm{novel}}\\), we follow open-vocabulary detection literature [20] and use the rare classes as defined by LVIS. The intuition behind this is that the occurrence of rare classes is correlated with uncommon scenarios and events that we are particularly interested in evaluating." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.54, + 0.474, + 0.751 + ], + "angle": 0, + "content": "With respect to evaluation, the advantage of defining \\( C^{\\mathrm{novel}} \\) is that we can apply closed-set tracking metrics in a straightforward manner, while open-world MOT [37] needs to resort to recall-based evaluation. Further, previous works [34, 37] have shown that the official evaluation metric in TAO, Track mAP [70], is sub-optimal in terms of handling FPs in presence of missing annotations. On the contrary, the recently proposed TETA metric [34] handles this shortcoming via local cluster evaluation. Also, TETA disentangles classification from localization and association performance. Thus, we choose TETA as the evaluation metric for our setup to provide a comprehensive insight into the localization, association, and open-vocabulary classification performance of tracker \\( M \\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.768, + 0.182, + 0.784 + ], + "angle": 0, + "content": "4. OVTrack" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.472, + 0.901 + ], + "angle": 0, + "content": "We present our Open-Vocabulary Tracker, OVTrack. We address two perspectives of its design: 1) Model perspective: We show how to handle the open-vocabulary setting in the localization, classification, and association modules of the tracker in Section 4.1; 2) Data perspective: Collecting and annotating the necessary amount of training videos is impractical for open-vocabulary MOT. Therefore, we contribute a" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.364, + 0.893, + 0.394 + ], + "angle": 0, + "content": "novel training approach for learning object tracking without video data in Section 4.2." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.406, + 0.64, + 0.422 + ], + "angle": 0, + "content": "4.1. Model design" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.43, + 0.895, + 0.491 + ], + "angle": 0, + "content": "We decompose OVTrack's functionality into localization, classification and association and discuss our open-vocabulary design philosophy in tackling difficulties for each of those parts. The model design is illustrated in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.491, + 0.895, + 0.657 + ], + "angle": 0, + "content": "1) Localization: To localize objects of arbitrary and possibly unknown classes \\( c \\in \\mathbb{N} \\) in a video, we train Faster R-CNN [57] in a class-agnostic manner, i.e. we use only the RPN and regression losses defined in [57]. We find that this localization procedure can generalize well to object classes that are unknown at training time, as also validated by previous works [10, 20, 76]. During training, we use RPN proposals as object candidates \\( P \\) for greater diversity, while during inference, we use the refined RCNN outputs as object candidates. Each candidate \\( r \\in P \\) is defined by confidence \\( p_r \\) and bounding box \\( \\mathbf{b}_r \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.895, + 0.794 + ], + "angle": 0, + "content": "2) Classification: Existing closed-set trackers [3,4,16,77] can only track objects of categories in \\( \\mathcal{C}^{\\mathrm{base}} \\), i.e. objects present and annotated in the training data distribution \\( \\mathcal{D}^{\\mathrm{train}} \\). To enable open-vocabulary classification, we need to be able to configure the classes we are interested in without re-training. Inspired by open-vocabulary detection literature [1], we connect our Faster R-CNN with the vision-language model CLIP [52] that has been pre-trained on over 400 million image-text pairs for contrastive learning." + }, + { + "type": "list", + "bbox": [ + 0.498, + 0.491, + 0.895, + 0.794 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.894, + 0.903 + ], + "angle": 0, + "content": "After extracting the RoI feature embeddings \\(\\mathbf{f}_r = \\mathcal{R}(\\phi(I), \\mathbf{b}_r), \\forall r \\in P\\) from the backbone \\(\\phi\\), we replace the original classifier in Faster R-CNN with a text head and add an image head generating the embeddings \\(\\hat{\\mathbf{t}}_r\\) and \\(\\hat{\\mathbf{i}}_r\\), for each \\(\\mathbf{f}_r\\). We use the CLIP text and image encoders to supervise the heads following [13, 20]. In particular, we use the class names to generate text prompts \\(\\mathcal{P}(c) = \\{\\mathbf{v}_1^c, \\dots, \\mathbf{v}_L^c, \\mathbf{w}_c\\}\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5570" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.895, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.205, + 0.895, + 0.248 + ], + "angle": 0, + "content": "Figure 4. OVTrack inference. Given an input video stream, we track objects of arbitrary classes, e.g. \\(\\mathcal{C}^{\\mathrm{base}}\\cup \\mathcal{C}^{\\mathrm{novel}}\\). We first localize objects agnostic of their class, then assign a semantic class label via the text embedding head, and finally associate them to existing tracks by comparing their appearance embedding \\(\\mathbf{q}\\) obtained from the tracking head to the appearance embeddings in the track memory." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.27, + 0.473, + 0.346 + ], + "angle": 0, + "content": "that consist of \\(L\\) context vectors \\(\\mathbf{v}^c\\) and a class name embedding \\(\\mathbf{w}_c\\). We feed the prompts into the CLIP text encoder \\(\\mathcal{E}\\), generating text embeddings \\(\\mathbf{t}_c = \\mathcal{E}(\\mathcal{P}(c)), \\forall c \\in \\mathcal{C}^{\\mathrm{base}}\\). We compute the affinity between the predicted embeddings \\(\\hat{\\mathbf{t}}_r\\) and their CLIP counterpart \\(\\mathbf{t}_c\\)." + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.353, + 0.471, + 0.372 + ], + "angle": 0, + "content": "\\[\n\\mathbf {z} (r) = \\left[ \\cos \\left(\\hat {\\mathbf {t}} _ {r}, \\mathbf {t} _ {b g}\\right), \\cos \\left(\\hat {\\mathbf {t}} _ {r}, \\mathbf {t} _ {1}\\right), \\dots , \\cos \\left(\\hat {\\mathbf {t}} _ {r}, \\mathbf {t} _ {| \\mathcal {C} ^ {\\text {b a s e}} |}\\right) \\right] \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.374, + 0.471, + 0.41 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t e x t}} = \\frac {1}{| P |} \\sum_ {r \\in P} \\mathcal {L} _ {\\mathrm {C E}} (\\operatorname {s o f t m a x} (\\mathbf {z} (r) / \\lambda), c _ {r}), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.418, + 0.473, + 0.528 + ], + "angle": 0, + "content": "where \\(\\cos (\\mathbf{v},\\mathbf{k}) = \\frac{\\mathbf{v}\\cdot\\mathbf{k}}{||\\mathbf{v}||||\\mathbf{k}||}\\), \\(\\mathbf{t}_{bg}\\) a learned background prompt, \\(\\lambda\\) a temperature parameter, \\(\\mathcal{L}_{\\mathrm{CE}}\\) the cross-entropy loss and \\(c_{r}\\) is the class label of \\(r\\). Furthermore, we align each \\(\\hat{\\mathbf{i}}_r\\) with the CLIP image encoder \\(\\mathcal{I}\\). For each \\(r\\), we crop the input image to \\(\\mathbf{b}_r\\), and resize it to the required input size to obtain the image embedding \\(\\mathbf{i}_r = \\mathcal{I}(\\mathcal{R}(I,\\mathbf{b}_r))\\). We minimize the distance between the corresponding \\(\\hat{\\mathbf{i}}_r\\) and \\(\\mathbf{i}_r\\)." + }, + { + "type": "equation", + "bbox": [ + 0.174, + 0.535, + 0.47, + 0.571 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {i m a g e}} = \\frac {1}{| P |} \\sum_ {r \\in P} \\left\\| \\hat {\\mathbf {i}} _ {r} - \\mathbf {i} _ {r} \\right\\| _ {1}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.578, + 0.473, + 0.775 + ], + "angle": 0, + "content": "3) Association: An open-vocabulary tracker should handle diverse scenarios that comprise complex camera motion and heterogeneous object motion patterns. However, those patterns are difficult to model especially when there are not enough video annotations available [16, 37]. Therefore, we rely on appearance cues to robustly track objects in an open-vocabulary context. Specifically, we employ a contrastive learning approach inspired by [16, 34]. Given an image pair \\((I_{\\mathrm{key}}, I_{\\mathrm{ref}})\\) we extract RoIs from both images and match the RoIs to the annotations using intersection-over-union (IoU). For each matched RoI in \\(I_{\\mathrm{key}}\\) with appearance embedding \\(\\mathbf{q} \\in Q\\), we cluster objects \\(Q^{+}\\) with the same identity and divide objects \\(Q^{-}\\) with different identity in \\(I_{\\mathrm{ref}}\\)." + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.781, + 0.47, + 0.819 + ], + "angle": 0, + "content": "\\[\n\\operatorname {P o s D} (\\mathbf {q}) = \\frac {1}{| Q ^ {+} (\\mathbf {q})) |} \\sum_ {\\mathbf {q} ^ {+} \\in Q ^ {+}} \\exp \\left(\\mathbf {q} \\cdot \\mathbf {q} ^ {+}\\right) / \\tau), \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.822, + 0.47, + 0.858 + ], + "angle": 0, + "content": "\\[\n\\operatorname {S i m} (\\mathbf {q}) = \\frac {\\exp \\left(\\mathbf {q} \\cdot \\mathbf {q} ^ {+} / \\tau\\right)}{\\operatorname {P o s D} (\\mathbf {q}) + \\sum_ {\\mathbf {q} ^ {-} \\in Q ^ {-}} \\exp \\left(\\mathbf {q} \\cdot \\mathbf {q} ^ {-} / \\tau\\right)}, \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.86, + 0.47, + 0.898 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t r a c k}} = - \\sum_ {\\mathbf {q} \\in Q} \\frac {1}{| Q ^ {+} (\\mathbf {q}) |} \\sum_ {\\mathbf {q} ^ {+} \\in Q ^ {+} (\\mathbf {q})} \\log \\left(\\operatorname {S i m} \\left(\\mathbf {q} ^ {+}\\right)\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.27, + 0.895, + 0.3 + ], + "angle": 0, + "content": "We further apply an auxiliary loss \\(\\mathcal{L}_{\\mathrm{aux}}\\) to constrain the magnitude of the logits following [16]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.302, + 0.895, + 0.483 + ], + "angle": 0, + "content": "During inference, we use straightforward appearance feature similarity for associating existing tracks \\(\\mathcal{T}\\) with objects in \\(P\\). In particular, for each track \\(\\tau \\in \\mathcal{T}\\) and its corresponding appearance embedding \\(\\mathbf{q}_{\\tau}\\), we compare its similarity with all candidate objects \\(r \\in P\\) using appearance embedding \\(\\mathbf{q}_r\\). We measure the similarity \\(\\mathbf{s}(\\tau, r)\\) of existing tracks with the candidate objects using both bi-directional softmax [16] and cosine similarity. We assign \\(r\\) to the track \\(\\tau\\) with its maximum similarity, if \\(\\mathbf{s}(\\tau, r) > \\beta\\). If \\(r\\) does not have a matching track, it starts a new track if its confidence \\(p_r > \\gamma\\) and is discarded otherwise. The inference pipeline is illustrated in Fig. 4." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.497, + 0.82, + 0.513 + ], + "angle": 0, + "content": "4.2. Learning to track without video data" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.522, + 0.895, + 0.688 + ], + "angle": 0, + "content": "In this section, we focus on how to train OVTrack in an open-vocabulary context. In particular, open-vocabulary MOT is challenging from a data perspective since we need to localize, classify, and associate possibly unknown objects of extremely diverse appearance with fixed method components. Thus, it is of fundamental importance to align our training data distribution \\(\\mathcal{D}^{\\mathrm{train}}\\) with the conditions found in evaluation. However, existing video datasets lack the diversity of contemporary image datasets. Hence, it is essential for open-vocabulary trackers to leverage not only video but more importantly static image data during training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.897, + 0.901 + ], + "angle": 0, + "content": "We use the large-scale, diverse image dataset LVIS [21] to train OVTrack. In particular, for each image \\( I_{\\mathrm{key}} \\) we generate a reference image \\( I_{\\mathrm{ref}} \\). Referring to our instance similarity loss in Eq. 6, the appearance similarity learning is constituted by contrasting positive and negative examples \\( \\mathbf{q}^{+} \\) and \\( \\mathbf{q}^{-} \\). The corollary of this is that learning will be optimal if \\( \\mathbf{q}^{+} \\) consists of examples with distortions commonly encountered in video data, such as change in object scale, viewpoint or lighting, while \\( \\mathbf{q}^{-} \\) contains examples with appearance changes associated with object identity, such as different material. While distortions like translation, scaling, and rotation can be simulated via classic data augmentation strategies [16, 74, 77], there are certain phenomena like changes in object viewpoint, lighting or context that cannot" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "5571" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.093, + 0.89, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.286, + 0.893, + 0.343 + ], + "angle": 0, + "content": "Figure 5. Data hallucination strategy. Given an input image, its annotations and its caption, we generate \\( x_0 \\sim n \\) and input it to the diffusion model [59], which progressively denoises \\( x \\) from \\( \\delta_0 = 0.75 \\) to \\( \\eta \\). At the same time, the foreground regions are kept fixed at each iteration. Specifically, we compose the generated images \\( x_k \\sim m_\\theta \\) with foreground regions \\( \\hat{x}_k \\sim n \\) at the current noise level, to obtain \\( x_k \\) as input for the next iteration. Finally, we arrive at \\( I_{\\mathrm{ref}} \\) which we use for instance similarity learning." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.362, + 0.226, + 0.376 + ], + "angle": 0, + "content": "be simulated by these." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.377, + 0.474, + 0.528 + ], + "angle": 0, + "content": "Therefore, to simulate all desired properties of our instance embedding space in \\( I_{\\mathrm{ref}} \\), we combine classic data augmentations with a DDPM-based data hallucination strategy. We use the data generation fidelity of stable diffusion [59] to simulate \\( \\mathbf{q}^{+} \\) and \\( \\mathbf{q}^{-} \\) via a specialized denoising process. Generally, the denoising process of a DPPM can be viewed as the inversion of a forward process that maps an input \\( x \\) to Gaussian white noise \\( \\mathcal{N}(0,\\mathbf{I}) \\). In the forward direction, Gaussian noise is added to the input image according to a variance schedule \\( \\delta_{k} \\) in \\( K \\) steps." + }, + { + "type": "equation", + "bbox": [ + 0.124, + 0.541, + 0.47, + 0.56 + ], + "angle": 0, + "content": "\\[\nn \\left(x _ {k} \\mid x _ {k - 1}\\right) = \\mathcal {N} \\left(x _ {k}; \\sqrt {1 - \\delta_ {k}} \\cdot x _ {k - 1}, \\delta_ {k} \\mathbf {I}\\right). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.572, + 0.47, + 0.618 + ], + "angle": 0, + "content": "In the backward direction, a neural network with parameters \\(\\theta\\) predicts the parameters \\(\\mu\\) and \\(\\Sigma\\) of a Gaussian distribution that reverses a forward step." + }, + { + "type": "equation", + "bbox": [ + 0.101, + 0.63, + 0.47, + 0.647 + ], + "angle": 0, + "content": "\\[\nm _ {\\theta} \\left(x _ {k - 1} \\mid x _ {k}\\right) = \\mathcal {N} \\left(x _ {k - 1}; \\mu_ {\\theta} \\left(x _ {k}, k\\right), \\Sigma_ {\\theta} \\left(x _ {k}, k\\right)\\right). \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.473, + 0.901 + ], + "angle": 0, + "content": "Our specialized denoising process is illustrated in Fig. 5. We initialize \\( I_{\\mathrm{ref}} \\) as \\( I_{\\mathrm{key}} \\) and apply a random geometric transformation to \\( I_{\\mathrm{ref}} \\). Next, we use the instance mask annotations in LVIS to define the set of positive examples \\( A^{+} \\). We divide each iteration of the denoising process of image \\( I_{\\mathrm{ref}} \\) using the union of all object masks in \\( A^{+} \\) into two branches following [39]. In addition, we use the conditioning mechanism in [59] to guide the backward process with the corresponding image caption. To initialize the backward process, we set \\( x_0 \\) to \\( I_{\\mathrm{ref}} \\) at \\( \\delta_0 = 0.75 \\) via the forward process \\( n(x_k|x_{k-1}) \\). Note that \\( x \\) corresponds to a latent representation of \\( I_{\\mathrm{ref}} \\) obtained via the encoder of stable diffusion. In each iteration, we apply \\( m_\\theta(x_{k-1}|x_k) \\) to obtain a new sample \\( x_k \\sim m_\\theta \\). At the same time, we use the forward process \\( n(x_k|x_{k-1}) \\) on the areas of \\( A^{+} \\) to generate \\( \\hat{x}_k \\sim n \\). At the end of each reverse iteration, we compose the two versions via" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.361, + 0.895, + 0.423 + ], + "angle": 0, + "content": "\\(x_{k} = A^{+}\\cdot \\hat{x}_{k} + (1 - A^{+})\\cdot x_{k}\\) . We iterate until \\(\\delta_k\\leq \\eta\\) Finally, we apply \\(m_{\\theta}(x_{k - 1}|x_k)\\) to the whole image, without branching, as a homogenization step between \\(A^{+}x_{k}\\) and \\((1 - A^{+})x_{k}\\) , while \\(\\eta >\\delta_k > 0\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.424, + 0.893, + 0.514 + ], + "angle": 0, + "content": "By this process, we achieve three goals. First, we generate random perturbations of the background. Second, we keep the areas of \\( A^{+} \\) close to its original content in each denoising step so that positive instances are integrated well into the new background. Third, we generate distractor objects by caption guided hallucination." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.533, + 0.634, + 0.55 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.559, + 0.682, + 0.573 + ], + "angle": 0, + "content": "5.1. Evaluation metrics" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.583, + 0.895, + 0.788 + ], + "angle": 0, + "content": "TETA. The tracking-everything accuracy (TETA) [34] is calculated from three independent scores. First, the localization accuracy (LocA) is calculated by matching all annotated boxes \\(\\alpha\\) to the predicted boxes of \\(\\mathcal{T}\\) without taking classification into account: \\(\\mathrm{LocA} = \\frac{|\\mathrm{TPL}|}{|\\mathrm{TPL}| + |\\mathrm{FPL}| + |\\mathrm{FNL}|}\\). Next, classification accuracy (ClsA) is computed based on all well-localized TPL, comparing the predicted semantic classes to the matched ground-truths: \\(\\mathrm{ClsA} = \\frac{|\\mathrm{TPC}|}{|\\mathrm{TPC}| + |\\mathrm{FPC}| + |\\mathrm{FNC}|}\\). Finally, association accuracy (AssocA) is computed in a similar fashion, comparing the identity of associated ground truths with well-localized predictions: AssocA = \\(\\frac{1}{|\\mathrm{TPL}|}\\sum_{b\\in \\mathrm{TPL}}\\frac{|\\mathrm{TPA}(b)|}{|\\mathrm{TPA}(b)| + |\\mathrm{FPA}(b)| + |\\mathrm{FNA}(b)|}\\). The TETA score is computed as the arithmetic mean of the three scores." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.79, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Track mAP. The Track mAP [70] is calculated using the 3D IoU between the bounding boxes of a predicted track \\(\\tau\\) and an annotated track \\(\\alpha\\) by \\(\\mathrm{IoU}_{\\mathrm{3D}}(\\tau, \\alpha) = \\frac{\\sum_{t \\in T} \\tau_t \\cap \\alpha_t}{\\sum_{t \\in T} \\tau_t \\cup \\alpha_t}\\). It is used analogous to 2D bounding box IoU to calculate the popular average precision metric per class as in [21]. The Track mAP is the average of the per-class scores across a set of IoU3D thresholds." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5572" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.895, + 0.131 + ], + "angle": 0, + "content": "Table 1. Open-vocabulary MOT comparison. We compare our method with existing closed-set trackers and off-the-shelf open-vocabulary baselines on base and novel classes on the validation and test sets of TAO [9]. We indicate the classes and data the methods trained on. Note that methods using TAO data utilize videos for training. All methods use ResNet50 [18] as backbone." + }, + { + "type": "table", + "bbox": [ + 0.159, + 0.133, + 0.812, + 0.387 + ], + "angle": 0, + "content": "
MethodClassesDataBaseNovel
Validation setBaseNovelCC3MLVISTAOTETALocAAssocAClsATETALocAAssocAClsA
QDTrack [16]✓✓-✓✓27.145.624.711.022.542.724.40.4
TETer [34]✓✓-✓✓30.347.431.612.125.745.931.10.2
DeepSORT (ViLD) [65]✓--✓✓26.947.115.817.721.146.414.72.3
Tracktor++ (ViLD) [3]✓--✓✓28.347.420.517.022.746.719.32.2
OVTrack✓--✓-35.549.336.920.227.848.833.61.5
RegionCLIP [75]
+ DeepSORT [65]✓-✓✓✓28.452.515.617.024.549.215.39.0
+ Tracktor++ [3]✓-✓✓✓29.652.419.616.925.750.118.98.1
+ OVTrack✓-✓✓-36.353.936.318.732.051.433.211.4
Test setBaseNovelCC3MLVISTAOTETALocAAssocAClsATETALocAAssocAClsA
QDTrack [16]✓✓-✓✓25.843.223.510.620.239.720.90.2
TETer [34]✓✓-✓✓29.244.030.410.721.739.125.90.0
DeepSORT (ViLD) [65]✓--✓✓24.543.814.615.217.238.411.61.7
Tracktor++ (ViLD) [3]✓--✓✓26.044.119.014.818.039.013.41.7
OVTrack✓--✓-32.645.635.416.924.141.828.71.8
RegionCLIP [75]
+ DeepSORT [65]✓-✓✓✓27.049.815.116.118.741.89.15.2
+ Tracktor++ [3]✓-✓✓✓28.049.418.815.720.042.412.05.7
+ OVTrack✓-✓✓-34.851.136.117.325.744.826.26.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.402, + 0.472, + 0.473 + ], + "angle": 0, + "content": "Table 2. Closed-set MOT Track mAP comparison. We compare to existing trackers on TAO [9] validation. Competing methods use ResNet101 [24], we use ResNet50 as backbone. All methods use Faster R-CNN [57]. We include results with stronger detectors and additional data in gray. \\(\\dagger\\) does not use videos for training." + }, + { + "type": "table", + "bbox": [ + 0.089, + 0.476, + 0.459, + 0.597 + ], + "angle": 0, + "content": "
MethodTrack mAP50Track mAP75Track mAP
SORT-TAO [9]13.2--
QDTrack [16]15.95.010.6
GTR† [78]20.4--
TAC [67]17.75.807.30
BIV [66]19.67.3013.6
OVTrack†21.210.615.9
GTR + CenterNet2† [78]22.5--
AOA [12]25.8--
" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.617, + 0.292, + 0.634 + ], + "angle": 0, + "content": "5.2. Implementation details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.474, + 0.903 + ], + "angle": 0, + "content": "We use ResNet50 [24] with FPN [35]. We filter object candidates \\(P\\) by non-maximum suppression (NMS) with an IoU threshold of 0.7 and randomly select \\(|P| = 256\\) candidates per image. We set \\(\\lambda = 0.07\\) in \\(\\mathcal{L}_{\\mathrm{text}}\\). We use a two-stage training process, first training the detection components following [13, 20], second fine-tuning the model for tracking with loss weights 0.25 for \\(\\mathcal{L}_{\\mathrm{track}}\\) and 1.0 for \\(\\mathcal{L}_{\\mathrm{aux}}\\) following [16]. We train on the LVIS dataset, with one hallucinated counterpart per image in the dataset. While we use the full dataset for the state-of-the-art comparison, we use a subset of 10,000 images for the ablation studies due to resource constraints. Unless otherwise noted, we use the following data augmentations in training: resizing, random horizontal flipping, color jittering, random affine transformation, and mosaic composition with varying parameters between \\(I_{\\mathrm{key}}\\) and \\(I_{\\mathrm{ref}}\\). We use \\(\\eta = 0.02\\) for data generation. For inference, we select object candidates \\(P\\) by NMS with" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.402, + 0.895, + 0.473 + ], + "angle": 0, + "content": "Table 3. Closed-set MOT TETA comparison. We compare to existing trackers on the TAO [9] validation. Benchmark results are taken from [34]. All competing methods use ResNet101 [24] except AOA [12], we use ResNet50 as backbone. All methods use Faster R-CNN [57]. \\(\\dagger\\) does not use videos for training." + }, + { + "type": "table", + "bbox": [ + 0.557, + 0.476, + 0.837, + 0.593 + ], + "angle": 0, + "content": "
MethodTETALocAAssocAClsA
SORT-TAO [9]24.848.114.312.1
Tracktor [3]24.247.413.012.1
DeepSORT [65]26.048.417.512.1
AOA [12]25.323.430.621.9
Tracktor++ [9]28.049.022.812.1
QDTrack [16]30.050.527.412.1
TETer [34]33.351.635.013.2
OVTrack†34.749.336.718.1
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.607, + 0.895, + 0.655 + ], + "angle": 0, + "content": "an IoU threshold of 0.5. We keep a track memory of 10 frames to re-identify objects after occlusion and set \\(\\beta = 0.5\\) and \\(\\gamma = 0.0001\\) (see Sec. 4.1)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.665, + 0.771, + 0.681 + ], + "angle": 0, + "content": "5.3. Comparison to state-of-the-art" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.895, + 0.885 + ], + "angle": 0, + "content": "Open-vocabulary MOT. In Tab. 1, we show the open-vocabulary MOT evaluation on the TAO validation and test sets, divided into base classes \\(\\mathcal{C}^{\\mathrm{base}}\\) and novel classes \\(\\mathcal{C}^{\\mathrm{novel}}\\). For details on the setup, please refer to the supplemental material. The baselines we establish are composed of both closed-set and open-vocabulary trackers. We choose the two state-of-the-art closed-set trackers, TETer [34] and QDTrack [16], trained on \\(\\mathcal{C}^{\\mathrm{base}} \\cup \\mathcal{C}^{\\mathrm{novel}}\\). In addition, we combine off-the-shelf trackers DeepSORT [65] and Tracktor++ [3] with the open-vocabulary detector ViLD [20] as baseline open-vocabulary trackers. These are, like OVTrack, trained on \\(\\mathcal{C}^{\\mathrm{base}}\\) only. Note that all baselines use video data for training, while we use only static images." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Our approach substantially outperforms all closed-set and" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.958 + ], + "angle": 0, + "content": "5573" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.273 + ], + "angle": 0, + "content": "open-vocabulary baselines. We achieve consistent improvement across LocA, AssocA, and ClsA on both base and novel classes. The baselines trained on \\(\\mathcal{C}^{\\mathrm{base}}\\cup \\mathcal{C}^{\\mathrm{novel}}\\) can, in some cases, correctly classify objects in \\(\\mathcal{C}^{\\mathrm{novel}}\\) but achieve poor results. On the contrary, both the open-vocabulary baselines and our tracker achieve significantly higher ClsA on novel classes. However, we note that classification on the TAO dataset remains a very challenging task. The absolute ClsA scores on novel classes are low. This is partially due to the nature of the ClsA metric, which only considers top-1 classification accuracy, while classes on the TAO dataset are diverse and fine-grained." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.274, + 0.473, + 0.424 + ], + "angle": 0, + "content": "Therefore, we investigate the use of stronger, recently proposed open-vocabulary detectors. We combine RegionCLIP [75] with our off-the-shelf baselines and OVTrack. We replace the localization and classification parts of OVTrack with RegionCLIP while keeping the association fixed. We observe that ClsA increases substantially for all trackers on novel classes. Our method achieves the best performance by a wide margin and achieves the best ClsA scores with 11.4 and 6.1 on the validation and test sets, respectively. Note however that RegionCLIP makes use of additional data." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.426, + 0.473, + 0.758 + ], + "angle": 0, + "content": "Closed-set MOT. In Tab. 2 and Tab. 3 we compare to existing works on the validation split of TAO using Track mAP and TETA metrics, respectively. Note that our method neither uses video data for training, nor is it trained on rare classes as defined in Sec. 3.1, while all of the compared closed-set trackers train on video data and use the held-out rare classes for training as they are part of the closed-set evaluation in TAO. We outperform all previous works by a sizable margin on both metrics. By examining the TETA scores in Tab. 3, we observe that our tracker obtains 2.3 points less in LocA compared to TETer [34]. However, our approach beats TETer in terms of AssocA by 1.7 points and greatly improves in ClsA by 4.9 points, illustrating the positive effect of CLIP distillation on both classification and associated compared to closed-set trackers. This validates our design in Sec. 4.1. Note that while AOA [12] has a better ClsA, it ensembles multiple few-shot detection and re-identification models trained on additional datasets as reported by previous works [34, 78]. Overall, our approach surpasses the previous state-of-the-art by 1.4 points in TETA and 2.3 points in Track mAP while using a weaker backbone and the same detector." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.77, + 0.24, + 0.785 + ], + "angle": 0, + "content": "5.4. Ablation studies" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.794, + 0.472, + 0.884 + ], + "angle": 0, + "content": "CLIP knowledge distillation. In Tab. 4 we analyze the effect of the knowledge distillation described in Sec. 4.1. In particular, we observe that using both \\(\\mathcal{L}_{\\mathrm{text}}\\) and \\(\\mathcal{L}_{\\mathrm{image}}\\) is more effective for classification than using only \\(\\mathcal{L}_{\\mathrm{text}}\\), improving ClsA significantly from 15.6 to 18.1, while LocA and AssocA stay at the same level of performance." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.886, + 0.473, + 0.901 + ], + "angle": 0, + "content": "Data hallucination strategy. We validate the effective" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.089, + 0.892, + 0.132 + ], + "angle": 0, + "content": "Table 4. Ablation study on CLIP knowledge distillation. We show that using both \\(\\mathcal{L}_{\\mathrm{text}}\\) and \\(\\mathcal{L}_{\\mathrm{image}}\\) is important to classification performance when doing CLIP knowledge distillation (Sec. 4.1)." + }, + { + "type": "table", + "bbox": [ + 0.549, + 0.136, + 0.842, + 0.186 + ], + "angle": 0, + "content": "
\\( \\mathcal{L}_{\\text{text}} \\)\\( \\mathcal{L}_{\\text{text}} \\mid \\text{image} \\)TETALocAAssocAClsA
✓-34.050.535.715.6
✓✓34.349.335.418.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.196, + 0.894, + 0.28 + ], + "angle": 0, + "content": "Table 5. Ablation study on data hallucination strategy. We show that our data hallucination strategy ('DDPM', Sec. 4.2) improves the association of a closed-set tracker [34] and our OVTrack on TAO [9] validation. We ensemble it with data augmentations, where 'Standard' refers to random resize and horizontal flip, 'Heavy' to color jitter, random affine transformation and mosaic." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.283, + 0.895, + 0.407 + ], + "angle": 0, + "content": "
StandardDDPMHeavyTETALocAAssocAClsA
TETer-SwinT
✓--32.350.730.615.5
✓✓-33.251.233.015.4
✓✓✓34.351.435.515.8
OVTrack
✓--32.548.931.117.6
✓✓-33.348.932.918.0
✓✓✓34.449.135.718.3
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.422, + 0.895, + 0.665 + ], + "angle": 0, + "content": "ness of our data hallucination strategy described in Sec. 4.2 by training both the closed-set tracker TETer [34] and our OVTrack with it. Note that we choose to use SwinT [38] with TETer and ResNet50 [24] with our OVTrack to achieve similar performance, in order to fairly compare the performance difference on both trackers. Tab. 5 shows the TETA results on the TAO validation set. We observe that our data hallucination strategy improves the AssocA significantly for both trackers, while LocA and ClsA are comparable. In particular, we improve 2.4 and 1.8 points in AssocA for TETer and OVTrack, respectively. Further, ensembling our data generation strategy with heavy data augmentations yields another 2.5 and 1.8 points improvement. Overall, we show that our data generation strategy improves instance similarity learning across both closed-set and open-vocabulary trackers while being complementary to classic data augmentation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.679, + 0.62, + 0.695 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.895, + 0.901 + ], + "angle": 0, + "content": "This work introduced open-vocabulary MOT as an effective solution to evaluating multi-object trackers beyond pre-defined training categories. We defined a suitable benchmark setting and presented OVTrack, a data-efficient open-vocabulary tracker. By using knowledge distillation from vision-language models, we improve tracking while going beyond limited dataset taxonomies. In addition, we put forth a data hallucination strategy tailored to instance similarity learning that addresses the data availability problem in open-vocabulary MOT. As a result, OVTrack learns tracking from static images and is able to track arbitrary objects in videos while outperforming existing trackers by a sizable margin on the large-scale, large-vocabulary TAO [9] benchmark." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5574" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.471, + 0.156 + ], + "angle": 0, + "content": "[1] Ankan Bansal, Karan Sikka, Gaurav Sharma, Rama Chellappa, and Ajay Divakaran. Zero-shot object detection. In ECCV, 2018. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.159, + 0.471, + 0.186 + ], + "angle": 0, + "content": "[2] Abhijit Bendale and Terrance Boult. Towards open world recognition. In CVPR, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.188, + 0.472, + 0.215 + ], + "angle": 0, + "content": "[3] Philipp Bergmann, Tim Meinhardt, and Laura Leal-Taixe. Tracking without bells and whistles. In ICCV, 2019. 1, 2, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.217, + 0.471, + 0.257 + ], + "angle": 0, + "content": "[4] Alex Bewley, Zongyuan Ge, Lionel Ott, Fabio Ramos, and Ben Upcroft. Simple online and realtime tracking. In ICIP, 2016. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.26, + 0.471, + 0.3 + ], + "angle": 0, + "content": "[5] Erik Bochinski, Volker Eiselein, and Thomas Sikora. High-speed tracking-by-detection without using image information. In AVSS, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.303, + 0.471, + 0.33 + ], + "angle": 0, + "content": "[6] Guillem Brasó and Laura Leal-Taixe. Learning a neural solver for multiple object tracking. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.333, + 0.471, + 0.373 + ], + "angle": 0, + "content": "[7] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.376, + 0.471, + 0.443 + ], + "angle": 0, + "content": "[8] Yun Chen, Frieda Rong, Shivam Duggal, Shenlong Wang, Xinchen Yan, Sivabalan Manivasagam, Shangjie Xue, Ersin Yumer, and Raquel Urtasun. Geosim: Realistic video simulation via geometry-aware composition for self-driving. In CVPR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.446, + 0.471, + 0.488 + ], + "angle": 0, + "content": "[9] Achal Dave, Tarasha Khurana, Pavel Tokmakov, Cordelia Schmid, and Deva Ramanan. Tao: A large-scale benchmark for tracking any object. In ECCV, 2020. 1, 2, 4, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.49, + 0.469, + 0.517 + ], + "angle": 0, + "content": "[10] Achal Dave, Pavel Tokmakov, and Deva Ramanan. Towards segmenting anything that moves. In CVPRW, 2019. 1, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.519, + 0.469, + 0.573 + ], + "angle": 0, + "content": "[11] Patrick Dendorfer, Aljosa Osep, Anton Milan, Konrad Schindler, Daniel Cremers, Ian Reid, Stefan Roth, and Laura Leal-Taixe. Motchallenge: A benchmark for single-camera multiple target tracking. IJCV, 129(4):845-881, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.576, + 0.471, + 0.629 + ], + "angle": 0, + "content": "[12] Fei Du, Bo Xu, Jiasheng Tang, Yuqi Zhang, Fan Wang, and Hao Li. 1st place solution to eccv-tao-2020: Detect and represent any object for tracking. arXiv preprint arXiv:2101.08040, 2021. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.633, + 0.471, + 0.686 + ], + "angle": 0, + "content": "[13] Yu Du, Fangyun Wei, Zihe Zhang, Miaojing Shi, Yue Gao, and Guoqi Li. Learning to prompt for open-vocabulary object detection with vision-language model. In CVPR, 2022. 2, 3, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.689, + 0.471, + 0.757 + ], + "angle": 0, + "content": "[14] Matteo Fabbri, Guillem Brasó, Gianluca Maugeri, Orcun Cetintas, Riccardo Gasparini, Aljoša Osep, Simone Calderara, Laura Leal-Taixe, and Rita Cucchiara. Motsynth: How can synthetic data help pedestrian detection and tracking? In CVPR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.76, + 0.471, + 0.786 + ], + "angle": 0, + "content": "[15] Christoph Feichtenhofer, Axel Pinz, and Andrew Zisserman. Detect to track and track to detect. In ICCV, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.471, + 0.856 + ], + "angle": 0, + "content": "[16] Tobias Fischer, Jiangmiao Pang, Thomas E Huang, Linlu Qiu, Haofeng Chen, Trevor Darrell, and Fisher Yu. Qdtrack: Quasi-dense similarity learning for appearance-only multiple object tracking. arXiv preprint arXiv:2210.06984, 2022. 2, 3, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.471, + 0.9 + ], + "angle": 0, + "content": "[17] Adrien Gaidon, Qiao Wang, Yohann Cabon, and Eleonora Vig. Virtual worlds as proxy for multi-object tracking analysis. In CVPR, 2016. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.147 + ], + "angle": 0, + "content": "[18] Shang-Hua Gao, Ming-Ming Cheng, Kai Zhao, Xin-Yu Zhang, Ming-Hsuan Yang, and Philip Torr. Res2net: A new multi-scale backbone architecture. IEEE TPAMI, 43(2):652-662, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.893, + 0.19 + ], + "angle": 0, + "content": "[19] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In CVPR, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.894, + 0.246 + ], + "angle": 0, + "content": "[20] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. arXiv preprint arXiv:2104.13921, 2021. 2, 3, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.894, + 0.289 + ], + "angle": 0, + "content": "[21] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In CVPR, 2019. 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.291, + 0.894, + 0.332 + ], + "angle": 0, + "content": "[22] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.334, + 0.894, + 0.36 + ], + "angle": 0, + "content": "[23] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.362, + 0.894, + 0.402 + ], + "angle": 0, + "content": "[24] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.405, + 0.893, + 0.445 + ], + "angle": 0, + "content": "[25] David Held, Jesse Levinson, and Sebastian Thrun. Precision tracking with sparse 3d and dense color 2d data. In ICRA, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.448, + 0.894, + 0.488 + ], + "angle": 0, + "content": "[26] David Held, Sebastian Thrun, and Silvio Savarese. Learning to track at 100 FPS with deep regression networks. In ECCV, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.49, + 0.893, + 0.517 + ], + "angle": 0, + "content": "[27] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.519, + 0.893, + 0.56 + ], + "angle": 0, + "content": "[28] Hou-Ning Hu, Yung-Hsu Yang, Tobias Fischer, Trevor Darrell, Fisher Yu, and Min Sun. Monocular quasi-dense 3d object tracking. IEEE TPAMI, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.562, + 0.894, + 0.602 + ], + "angle": 0, + "content": "[29] KJ Joseph, Salman Khan, Fahad Shahbaz Khan, and Vineeth N Balasubramanian. Towards open world object detection. In CVPR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.604, + 0.893, + 0.645 + ], + "angle": 0, + "content": "[30] Anna Khoreva, Rodrigo Benenson, Eddy Ilg, Thomas Brox, and Bernt Schiele. Lucid data dreaming for video object segmentation. *IJCV*, 127(9):1175–1197, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.647, + 0.893, + 0.687 + ], + "angle": 0, + "content": "[31] Seung Wook Kim, Jonah Philion, Antonio Torralba, and Sanja Fidler. Drivegan: Towards a controllable high-quality neural simulation. In CVPR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.689, + 0.893, + 0.73 + ], + "angle": 0, + "content": "[32] Laura Leal-Taixe, Cristian Canton-Ferrer, and Konrad Schindler. Learning by tracking: Siamese CNN for robust target association. In CVPRW, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.732, + 0.893, + 0.772 + ], + "angle": 0, + "content": "[33] Bastian Leibe, Ales Leonardis, and Bernt Schiele. Robust object detection with interleaved categorization and segmentation. IJCV, 77(1):259-289, 2008. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.775, + 0.893, + 0.815 + ], + "angle": 0, + "content": "[34] Siyuan Li, Martin Danelljan, Henghui Ding, Thomas E Huang, and Fisher Yu. Tracking every thing in the wild. In ECCV, 2022. 2, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.817, + 0.893, + 0.858 + ], + "angle": 0, + "content": "[35] Tsung-Yi Lin, Piotr Dollar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In CVPR, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.86, + 0.893, + 0.9 + ], + "angle": 0, + "content": "[36] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C Berg. Ssd: Single shot multibox detector. In ECCV, 2016. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5575" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.146 + ], + "angle": 0, + "content": "[37] Yang Liu, Idil Esen Zulfikar, Jonathon Luiten, Achal Dave, Deva Ramanan, Bastian Leibe, Aljosa Osep, and Laura Leal-Taixe. Opening up open world tracking. In CVPR, 2022. 1, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.15, + 0.472, + 0.205 + ], + "angle": 0, + "content": "[38] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.208, + 0.472, + 0.262 + ], + "angle": 0, + "content": "[39] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In CVPR, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.265, + 0.472, + 0.305 + ], + "angle": 0, + "content": "[40] Jonathon Luiten, Tobias Fischer, and Bastian Leibe. Track to reconstruct and reconstruct to track. RA-L, 5(2):1803-1810, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.309, + 0.472, + 0.35 + ], + "angle": 0, + "content": "[41] Tim Meinhardt, Alexander Kirillov, Laura Leal-Taixe, and Christoph Feichtenhofer. Trackformer: Multi-object tracking with transformers. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.352, + 0.472, + 0.394 + ], + "angle": 0, + "content": "[42] Anton Milan, Seyed Hamid Rezatofighi, Anthony R. Dick, Ian D. Reid, and Konrad Schindler. Online multi-target tracking using recurrent neural networks. In AAAI, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.396, + 0.472, + 0.437 + ], + "angle": 0, + "content": "[43] Anton Milan, Stefan Roth, and Konrad Schindler. Continuous energy minimization for multitarget tracking. IEEE TPAMI, 36(1):58-72, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.44, + 0.472, + 0.48 + ], + "angle": 0, + "content": "[44] Dennis Mitzel and Bastian Leibe. Taking mobile multi-object tracking to the next level: People, unknown objects, and carried items. In ECCV, 2012. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.483, + 0.472, + 0.525 + ], + "angle": 0, + "content": "[45] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.527, + 0.472, + 0.58 + ], + "angle": 0, + "content": "[46] Aljoša Osep, Alexander Hermans, Francis Engelmann, Dirk Klostermann, Markus Mathias, and Bastian Leibe. Multiscale object candidates for generic object tracking in street scenes. In ICRA, 2016. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.584, + 0.472, + 0.625 + ], + "angle": 0, + "content": "[47] Aljoša Osep, Wolfgang Mehner, Markus Mathias, and Bastian Leibe. Combined image-and world-space tracking in traffic scenes. In ICRA, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.628, + 0.472, + 0.669 + ], + "angle": 0, + "content": "[48] Aljoša Osep, Wolfgang Mehner, Paul Voigtlaender, and Bastian Leibe. Track, then decide: Category-agnostic vision-based multi-object tracking. In ICRA, 2018. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.672, + 0.472, + 0.712 + ], + "angle": 0, + "content": "[49] Aljoša Osep, Paul Voigtlaender, Mark Weber, Jonathon Luiten, and Bastian Leibe. 4d generic video object proposals. In ICRA, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.715, + 0.472, + 0.757 + ], + "angle": 0, + "content": "[50] Jiangmiao Pang, Linlu Qiu, Xia Li, Haofeng Chen, Qi Li, Trevor Darrell, and Fisher Yu. Quasi-dense similarity learning for multiple object tracking. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.759, + 0.472, + 0.799 + ], + "angle": 0, + "content": "[51] Jeffrey Pennington, Richard Socher, and Christopher D Manning. Glove: Global vectors for word representation. In EMNLP, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.472, + 0.87 + ], + "angle": 0, + "content": "[52] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.472, + 0.902 + ], + "angle": 0, + "content": "[53] Deva Ramanan and David A Forsyth. Finding and tracking people from the bottom up. In CVPR, 2003. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.134 + ], + "angle": 0, + "content": "[54] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.894, + 0.175 + ], + "angle": 0, + "content": "[55] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In CVPR, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.894, + 0.205 + ], + "angle": 0, + "content": "[56] Joseph Redmon and Ali Farhadi. Yolo9000: better, faster, stronger. In CVPR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.208, + 0.894, + 0.248 + ], + "angle": 0, + "content": "[57] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In NeurIPS, 2015. 2, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.251, + 0.894, + 0.289 + ], + "angle": 0, + "content": "[58] Stephan R Richter, Vibhav Vineet, Stefan Roth, and Vladlen Koltun. Playing for data: Ground truth from computer games. In ECCV, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.293, + 0.894, + 0.345 + ], + "angle": 0, + "content": "[59] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.349, + 0.894, + 0.39 + ], + "angle": 0, + "content": "[60] Amir Sadeghian, Alexandre Alahi, and Silvio Savarese. Tracking the untrackable: Learning to track multiple cues with long-term dependencies. In ICCV, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.392, + 0.894, + 0.432 + ], + "angle": 0, + "content": "[61] Samuel Schulter, Paul Vernaza, Wongun Choi, and Manmohan Chandraker. Deep network flow for multi-object tracking. In CVPR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.434, + 0.894, + 0.488 + ], + "angle": 0, + "content": "[62] Sarthak Sharma, Junaid Ahmed Ansari, J. Krishna Murthy, and K. Madhava Krishna. Beyond pixels: Leveraging geometry and shape cues for online multi-object tracking. In ICRA, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.491, + 0.894, + 0.544 + ], + "angle": 0, + "content": "[63] Peize Sun, Jinkun Cao, Yi Jiang, Rufeng Zhang, Enze Xie, Zehuan Yuan, Changhu Wang, and Ping Luo. Transtrack: Multiple object tracking with transformer. arXiv preprint arXiv:2012.15460, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.547, + 0.894, + 0.615 + ], + "angle": 0, + "content": "[64] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In CVPR, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.618, + 0.894, + 0.657 + ], + "angle": 0, + "content": "[65] Nicolai Wojke, Alex Bewley, and Dietrich Paulus. Simple online and realtime tracking with a deep association metric. In ICIP, 2017. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.66, + 0.894, + 0.715 + ], + "angle": 0, + "content": "[66] Sanghyun Woo, Kwanyong Park, Seoung Wug Oh, In So Kweon, and Joon-Young Lee. Bridging images and videos: A simple learning framework for large vocabulary video object detection. In ECCV, 2022. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.717, + 0.894, + 0.757 + ], + "angle": 0, + "content": "[67] Sanghyun Woo, Kwanyong Park, Seoung Wug Oh, In So Kweon, and Joon-Young Lee. Tracking by associating clips. In ECCV, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.759, + 0.894, + 0.787 + ], + "angle": 0, + "content": "[68] Bin Xiao, Haiping Wu, and Yichen Wei. Simple baselines for human pose estimation and tracking. In ECCV, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.789, + 0.894, + 0.815 + ], + "angle": 0, + "content": "[69] Bo Yang and Ram Nevatia. An online learned CRF model for multi-target tracking. In CVPR, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.817, + 0.894, + 0.844 + ], + "angle": 0, + "content": "[70] Linjie Yang, Yuchen Fan, and Ning Xu. Video instance segmentation. In ICCV, 2019. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[71] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darryll. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In CVPR, 2020. 1" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5576" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.133 + ], + "angle": 0, + "content": "[72] Alireza Zareian, Kevin Dela Rosa, Derek Hao Hu, and Shih-Fu Chang. Open-vocabulary object detection using captions. In CVPR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.136, + 0.47, + 0.177 + ], + "angle": 0, + "content": "[73] Fangao Zeng, Bin Dong, Tiancai Wang, Xiangyu Zhang, and Yichen Wei. Motr: End-to-end multiple-object tracking with transformer. arXiv preprint arXiv:2105.03247, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.179, + 0.472, + 0.233 + ], + "angle": 0, + "content": "[74] Yifu Zhang, Chunyu Wang, Xinggang Wang, Wenjun Zeng, and Wenyu Liu. Fairmot: On the fairness of detection and re-identification in multiple object tracking. IJCV, 129(11):3069-3087, 2021. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.235, + 0.472, + 0.29 + ], + "angle": 0, + "content": "[75] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Liunian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, et al. Regionclip: Region-based language-image pretraining. In CVPR, 2022. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.292, + 0.471, + 0.345 + ], + "angle": 0, + "content": "[76] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Phillip Krahenbuhl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision. In ECCV, 2022. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.348, + 0.472, + 0.375 + ], + "angle": 0, + "content": "[77] Xingyi Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Tracking objects as points. In ECCV, 2020. 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.378, + 0.471, + 0.417 + ], + "angle": 0, + "content": "[78] Xingyi Zhou, Tianwei Yin, Vladlen Koltun, and Philipp Krahenbihl. Global tracking transformers. In CVPR, 2022. 2, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.417 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5577" + } + ] +] \ No newline at end of file diff --git a/2023/OVTrack_ Open-Vocabulary Multiple Object Tracking/092f1f8a-50aa-41bc-8e59-a89dde3c1cc2_origin.pdf b/2023/OVTrack_ Open-Vocabulary Multiple Object Tracking/092f1f8a-50aa-41bc-8e59-a89dde3c1cc2_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2ae0bfc3c139e8215dd144539f332e8f3e977222 --- /dev/null +++ b/2023/OVTrack_ Open-Vocabulary Multiple Object Tracking/092f1f8a-50aa-41bc-8e59-a89dde3c1cc2_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c106ed198c34b813018080e45988a77d1b1aa19b657eab713facf0f52f58da76 +size 3457633 diff --git a/2023/OVTrack_ Open-Vocabulary Multiple Object Tracking/full.md b/2023/OVTrack_ Open-Vocabulary Multiple Object Tracking/full.md new file mode 100644 index 0000000000000000000000000000000000000000..28b45d7e82b536528abd96c431fee4609ca2b76b --- /dev/null +++ b/2023/OVTrack_ Open-Vocabulary Multiple Object Tracking/full.md @@ -0,0 +1,299 @@ +# OVTrack: Open-Vocabulary Multiple Object Tracking + +Siyuan Li* Tobias Fischer* Lei Ke Henghui Ding Martin Danelljan Fisher Yu Computer Vision Lab, ETH Zürich https://www.vis.xyz/pub/ovtrack/ + +# Abstract + +The ability to recognize, localize and track dynamic objects in a scene is fundamental to many real-world applications, such as self-driving and robotic systems. Yet, traditional multiple object tracking (MOT) benchmarks rely only on a few object categories that hardly represent the multitude of possible objects that are encountered in the real world. This leaves contemporary MOT methods limited to a small set of pre-defined object categories. In this paper, we address this limitation by tackling a novel task, open-vocabulary MOT, that aims to evaluate tracking beyond predefined training categories. We further develop OVTrack, an open-vocabulary tracker that is capable of tracking arbitrary object classes. Its design is based on two key ingredients: First, leveraging vision-language models for both classification and association via knowledge distillation; second, a data hallucination strategy for robust appearance feature learning from denoising diffusion probabilistic models. The result is an extremely data-efficient open-vocabulary tracker that sets a new state-of-the-art on the large-scale, large-vocabulary TAO benchmark, while being trained solely on static images. + +# 1. Introduction + +Multiple Object Tracking (MOT) aims to recognize, localize and track objects in a given video sequence. It is a cornerstone of dynamic scene analysis and vital for many real-world applications such as autonomous driving, augmented reality, and video surveillance. Traditionally, MOT benchmarks [9, 11, 19, 64, 71] define a set of semantic categories that constitute the objects to be tracked in the training and testing data distributions. The potential of traditional MOT methods [3, 4, 33, 43] is therefore limited by the taxonomies of those benchmarks. As a consequence, contemporary MOT methods struggle with unseen events, leading to a gap between evaluation performance and real-world + +![](images/394eae8e3a8d336e2657acfe267ee8f7a5eb0040cfba683c429a42f56bc6dfde.jpg) +Figure 1. OVTrack. We approach the task of open-vocabulary multiple object tracking. During training, we leverage vision-language (VL) models both for generating samples and knowledge distillation. During testing, we track both base and novel classes unseen during training by querying a vision-language model. + +# deployment. + +To bridge this gap, previous works have tackled MOT in an open-world context. In particular, Osep et al. [46, 48] approach generic object tracking by first segmenting the scene and performing tracking before classification. Other works have used class agnostic localizers [10, 49] to perform MOT on arbitrary objects. Recently, Liu et al. [37] defined open-world tracking, a task that focuses on the evaluation of previously unseen objects. In particular, it requires any-object tracking as a stage that precedes object classification. This setup comes with two inherent difficulties. First, in an open-world context, densely annotating all objects is prohibitively expensive. Second, without a pre-defined taxonomy of categories, the notion of what is an object is ambiguous. As a consequence, Liu et al. resort to recall-based evaluation, which is limited in two ways. Penalizing false positives (FP) becomes impossible, i.e. we cannot measure the tracker precision. Moreover, by evaluating tracking in a class-agnostic manner, we lose the ability to evaluate how + +well a tracker can infer the semantic category of an object. + +In this paper, we propose open-vocabulary MOT as an effective solution to these problems. Similar to open-world MOT, open-vocabulary MOT aims to track multiple objects beyond the pre-defined training categories. However, instead of dismissing the classification problem and resorting to recall-based evaluation, we assume that at test time we are given the classes of objects we are interested in. This allows us to apply existing closed-set tracking metrics [34, 70] that capture both precision and recall, while still evaluating the tracker's ability to track arbitrary objects during inference. + +We further present the first Open-Vocabulary Tracker, OVTrack (see Fig. 1). To this end, we identify and address two fundamental challenges to the design of an open-vocabulary multi-object tracker. The first is that closed-set MOT methods are simply not capable of extending their pre-defined taxonomies. The second is data availability, i.e. scaling video data annotation to a large vocabulary of classes is extremely costly. Inspired by existing works in open-vocabulary detection [1, 13, 20, 76], we replace our classifier with an embedding head, which allows us to measure similarities of localized objects to an open vocabulary of semantic categories. In particular, we distill knowledge from CLIP [52] into our model by aligning the image feature representations of object proposals with the corresponding CLIP image and text embeddings. + +Beyond detection, association is the core of modern MOT methods. It is driven by two affinity cues: motion and appearance. In an open-vocabulary context, motion cues are brittle since arbitrary scenery contains complex and diverse camera and object motion patterns. In contrast, diverse objects usually exhibit heterogeneous appearance. However, relying on appearance cues requires robust representations that generalize to novel object categories. We find that CLIP feature distillation helps in learning better appearance representations for improved association. This is especially intriguing since object classification and appearance modeling are usually distinct in the MOT pipeline [3, 16, 65]. + +Learning robust appearance features also requires strong supervision that captures object appearance changes in different viewpoints, background, and lighting. To approach the data availability problem, we utilize the recent success of denoising diffusion probabilistic models (DDPMs) in image synthesis [54, 59] and propose an effective data hallucination strategy tailored to appearance modeling. In particular, from a static image, we generate both simulated positive and negative instances along with random background perturbations. + +The main contributions are summarized as follows: + +1. We define the task of open-vocabulary MOT and provide a suitable benchmark setting on the large-scale, large-vocabulary MOT benchmark TAO [9]. +2. We develop OVTrack, the first open-vocabulary multi-object tracker. It leverages vision-language models to + +![](images/dc380e7f66d95f62446cb991af7fa3aa0f494037ef2d7b4804959ed3bc5610ad.jpg) +Figure 2. OVTrack qualitative results. We condition our tracker on text prompts unseen during training, namely 'heron', 'hippo' and 'drone', and successfully track the corresponding objects in the videos. The box color depicts object identity. + +improve both classification and association compared to closed-set trackers. + +3. We propose an effective data hallucination strategy that allows us to address the data availability problem in open-vocabulary settings by leveraging DDPMs. + +Owing to its thoughtful design, OVTrack sets a new state-of-the-art on the challenging TAO benchmark [9], outperforming existing trackers by a significant margin while being trained on static images only. In addition, OVTrack is capable of tracking arbitrary object classes (see Fig. 2), overcoming the limitation of closed-set trackers. + +# 2. Related work + +Multiple object tracking. The dominant paradigm in MOT literature is tracking-by-detection [53], where objects are first detected in each frame, and subsequently associated across time. Thus, many works have focused on data association, aiming to exploit similarity cues such as visual appearance [3, 16, 32, 42, 50, 60, 65, 69], 2D object motion [4,5,15,26,68] or 3D object motion [25,40,44,47,48,62] most effectively. Recently, researchers have focused on learning data association with graph neural networks [6, 61] or transformers [41,63,73,78]. However, those works dismiss a more profound problem in the tracking-by-detection pipeline that precedes data association: Contemporary object detectors [23, 36, 55-57] are designed for closed-set scenarios where all objects appear frequently in the training and testing data distributions. Hence, Dave et al. [9] proposed a new benchmark, TAO, that focuses on studying MOT in the long-tail of the object category distribution. On this benchmark, GTR [78], AOA [12], QDTrack [16] and TET [34] achieve impressive performance. However, those works are still limited to pre-defined object categories and thus do not scale + +to the diversity of real-world settings. Our work enables tracking of unseen classes from an open vocabulary. + +Open-world detection and tracking. Open-world detection methods aim to detect any salient object in a given input image irrespective of its category and beyond the training data distribution in particular. However, object classification under such a setting is ill-posed since novel classes will be unknown by definition [2, 29]. As such, open-world detection methods utilize class agnostic localizers [10] and treat classification as a clustering problem [29], estimating a similarity between novel instances and grouping them into novel classes via incremental learning. + +Instead, open-vocabulary object detection methods aim to detect arbitrary, but given classes of objects at test time [72]. For this, Bansal et al. [1] connect an object detector with word representations [51]. Recently, models like CLIP [52] learn visual representations from natural language supervision. Their main advantage over word representations is better alignment of visual concepts and language description. Consequently, many works have focused on leveraging image-text representations for open-vocabulary and few-shot object detection [20, 76]. ViLD [20] distills CLIP image features, while Detic [76] leverages classification data for joint training. Other works have focused on learning good language prompts for open-vocabulary object detection [13]. + +Fewer works have tackled the open-world problem in the MOT domain. Existing works perform scene segmentation and class agnostic tracking before classification [44, 46, 48] or utilize class-agnostic proposal generation [10, 49], similar to open-world detection methods. Liu et al. [37] propose an open-world tracking benchmark, TAO-OW, that evaluates class-agnostic tracking as a task that precedes classification. However, this comes with the limitation that the evaluation only captures tracker recall and no classification accuracy. Instead of dismissing classification, we pose the problem in a different way, i.e. at test time we know the novel classes we are interested in. This allows us to capture both the precision and recall of novel classes in our evaluation, while our method maintains the ability to track any object. + +Learning tracking from static images. Since labelled video data is expensive to acquire at scale, recent methods have proposed to use static images to supervise MOT methods [16, 66, 74, 77]. CenterTrack [77] proposes to learn motion offsets from static images by random translation of the input, while FairMOT [74] treats objects in a dataset of static images as unique classes to distinguish. Inspired by recent progress in self-supervised representation learning [7, 22, 45], Fischer et al. [16] propose to utilize data augmentation in combination with a contrastive learning objective to learn appearance-based tracking from static images. We go beyond classic data augmentation used in existing works by generating positive and negative examples of objects along with background perturbations via generative + +models, offering a more targeted approach to guiding appearance similarity learning from static images. + +Data generation for tracking. While aforementioned methods alleviate the data availability problem in MOT, there is still room for improvement when it comes to data generation in video tasks. Therefore, a large body of research has focused on data generation strategies that can benefit tracking methods [8, 14, 17, 28, 30, 31, 58]. While early works focused on obtaining synthetic data from computer graphics engines [14, 17, 28, 58], newer approaches combine 3D assets with generative models for improved realism [8, 31]. Fewer works have tackled data generation with generative models only [30]. Recently, DDPMs [27, 54, 59] showed impressive results in image synthesis. We leverage their data generation fidelity to address the data availability problem that is particularly pronounced in open-vocabulary MOT with a novel data hallucination strategy tailored to appearance modeling. + +# 3. Open-Vocabulary MOT + +In real-world scenarios, object categories follow a long-tailed distribution with a rich vocabulary. The remarkable diversity of the open world cannot be covered by a monolithic dataset. However, existing MOT benchmarks focus on closed-set evaluation, with often only a handful of object classes being evaluated. Furthermore, the task setup requires trackers to only track objects within a small set of training categories. To bridge the gap between existing MOT benchmarks and algorithms and real-world settings, we propose the task of open-vocabulary MOT and define its training and evaluation setup as follows. + +At training time, we train a tracker $M$ on the training data distribution $\mathcal{D}^{\mathrm{train}} = \{\mathbf{X}^{\mathrm{train}},\mathcal{A}^{\mathrm{train}}\}$ that contains video sequences $\mathbf{X}^{\mathrm{train}}$ and their respective annotations $\mathcal{A}^{\mathrm{train}}$ of objects with semantic categories $\mathcal{C}^{\mathrm{base}}\subset \mathbb{N}$ . Each annotation $\alpha \in \mathcal{A}^{\mathrm{train}}$ consists of a set of states $\{\alpha_{t}\}_{t\in T}$ for each frame $t\in T$ that the object is visible in. A state $\alpha_{t} = (\mathbf{b}_{t},c_{t})$ comprises the object class $c\in \mathbb{N}$ and the 2D bounding box $\mathbf{b} = [x,y,w,h]$ , where $(x,y)$ is the center location in pixel coordinates and $(w,h)$ are width and height, respectively. At test time, we are given video sequences $\mathbf{X}^{\mathrm{test}}$ and a set of object classes $\mathcal{C}^{\mathrm{novel}}\subset \mathbb{N}\setminus \mathcal{C}^{\mathrm{base}}$ that we are interested in. We aim to find all tracks $\mathcal{T}$ of objects in $\mathbf{X}^{\mathrm{test}}$ belonging to classes $\mathcal{C}^{\mathrm{base}}\cup \mathcal{C}^{\mathrm{novel}}$ . Each track state $\tau_{t} = (\mathbf{b}_{t},p_{t},c_{t})\in \mathcal{T}$ contains predicted object confidence $p\in [0,1]$ , class $c\in \mathbb{N}$ and 2D bounding box $\mathbf{b} = [x,y,w,h]$ . The important distinctions to closed-set tracking are two-fold: 1) We evaluate the tracker $M$ not only on $\mathcal{C}^{\mathrm{base}}$ but also on $\mathcal{C}^{\mathrm{novel}}$ with $\mathcal{C}^{\mathrm{novel}}\cap \mathcal{C}^{\mathrm{base}} = \emptyset$ , and 2) While $\mathcal{C}^{\mathrm{novel}}$ is known at test time, our setup requires the tracker $M$ to track arbitrary object classes $c\in \mathbb{N}$ since $\mathcal{C}^{\mathrm{novel}}$ remains unknown at training time. In particular, the evaluation of $\mathcal{C}^{\mathrm{novel}}$ illustrates the ability of tracker $M$ to track any unknown class $c\in \mathbb{N}\setminus \mathcal{C}^{\mathrm{base}}$ , while the classes in $\mathcal{C}^{\mathrm{novel}}$ serve as proxy. + +![](images/cdb9b574ef41e6b3717b075a9a0509d1fc1a49d112ced4c0a08dd85b4963cbd3.jpg) +Figure 3. OVTrack training. From a single static $I_{\mathrm{key}}$ , we generate $I_{\mathrm{ref}}$ with our data hallucination strategy. We extract RoIs via a RPN [57] and perform knowledge distillation from CLIP [52] via the embeddings of the text and image heads. Note that we train classification only on $C^{\mathrm{base}}$ . Further, we obtain appearance embeddings from the tracking head and apply our instance similarity loss on the image pair. + +# 3.1. Benchmark + +We utilize the large-scale, large-vocabulary MOT dataset TAO [9] to establish a suitable benchmark for open-vocabulary MOT. TAO mostly follows the taxonomy of LVIS [21], which divides classes according to their occurrence into frequent, common and rare classes. To obtain our held-out set $\mathcal{C}^{\mathrm{novel}}$ , we follow open-vocabulary detection literature [20] and use the rare classes as defined by LVIS. The intuition behind this is that the occurrence of rare classes is correlated with uncommon scenarios and events that we are particularly interested in evaluating. + +With respect to evaluation, the advantage of defining $C^{\mathrm{novel}}$ is that we can apply closed-set tracking metrics in a straightforward manner, while open-world MOT [37] needs to resort to recall-based evaluation. Further, previous works [34, 37] have shown that the official evaluation metric in TAO, Track mAP [70], is sub-optimal in terms of handling FPs in presence of missing annotations. On the contrary, the recently proposed TETA metric [34] handles this shortcoming via local cluster evaluation. Also, TETA disentangles classification from localization and association performance. Thus, we choose TETA as the evaluation metric for our setup to provide a comprehensive insight into the localization, association, and open-vocabulary classification performance of tracker $M$ . + +# 4. OVTrack + +We present our Open-Vocabulary Tracker, OVTrack. We address two perspectives of its design: 1) Model perspective: We show how to handle the open-vocabulary setting in the localization, classification, and association modules of the tracker in Section 4.1; 2) Data perspective: Collecting and annotating the necessary amount of training videos is impractical for open-vocabulary MOT. Therefore, we contribute a + +novel training approach for learning object tracking without video data in Section 4.2. + +# 4.1. Model design + +We decompose OVTrack's functionality into localization, classification and association and discuss our open-vocabulary design philosophy in tackling difficulties for each of those parts. The model design is illustrated in Fig. 3. + +1) Localization: To localize objects of arbitrary and possibly unknown classes $c \in \mathbb{N}$ in a video, we train Faster R-CNN [57] in a class-agnostic manner, i.e. we use only the RPN and regression losses defined in [57]. We find that this localization procedure can generalize well to object classes that are unknown at training time, as also validated by previous works [10, 20, 76]. During training, we use RPN proposals as object candidates $P$ for greater diversity, while during inference, we use the refined RCNN outputs as object candidates. Each candidate $r \in P$ is defined by confidence $p_r$ and bounding box $\mathbf{b}_r$ . +2) Classification: Existing closed-set trackers [3,4,16,77] can only track objects of categories in $\mathcal{C}^{\mathrm{base}}$ , i.e. objects present and annotated in the training data distribution $\mathcal{D}^{\mathrm{train}}$ . To enable open-vocabulary classification, we need to be able to configure the classes we are interested in without re-training. Inspired by open-vocabulary detection literature [1], we connect our Faster R-CNN with the vision-language model CLIP [52] that has been pre-trained on over 400 million image-text pairs for contrastive learning. + +After extracting the RoI feature embeddings $\mathbf{f}_r = \mathcal{R}(\phi(I), \mathbf{b}_r), \forall r \in P$ from the backbone $\phi$ , we replace the original classifier in Faster R-CNN with a text head and add an image head generating the embeddings $\hat{\mathbf{t}}_r$ and $\hat{\mathbf{i}}_r$ , for each $\mathbf{f}_r$ . We use the CLIP text and image encoders to supervise the heads following [13, 20]. In particular, we use the class names to generate text prompts $\mathcal{P}(c) = \{\mathbf{v}_1^c, \dots, \mathbf{v}_L^c, \mathbf{w}_c\}$ + +![](images/0f6b8be833e232ee13125ff5a83561aa957cd3a3d12ff7efa1a326edf3612bd3.jpg) +Figure 4. OVTrack inference. Given an input video stream, we track objects of arbitrary classes, e.g. $\mathcal{C}^{\mathrm{base}}\cup \mathcal{C}^{\mathrm{novel}}$ . We first localize objects agnostic of their class, then assign a semantic class label via the text embedding head, and finally associate them to existing tracks by comparing their appearance embedding $\mathbf{q}$ obtained from the tracking head to the appearance embeddings in the track memory. + +that consist of $L$ context vectors $\mathbf{v}^c$ and a class name embedding $\mathbf{w}_c$ . We feed the prompts into the CLIP text encoder $\mathcal{E}$ , generating text embeddings $\mathbf{t}_c = \mathcal{E}(\mathcal{P}(c)), \forall c \in \mathcal{C}^{\mathrm{base}}$ . We compute the affinity between the predicted embeddings $\hat{\mathbf{t}}_r$ and their CLIP counterpart $\mathbf{t}_c$ . + +$$ +\mathbf {z} (r) = \left[ \cos \left(\hat {\mathbf {t}} _ {r}, \mathbf {t} _ {b g}\right), \cos \left(\hat {\mathbf {t}} _ {r}, \mathbf {t} _ {1}\right), \dots , \cos \left(\hat {\mathbf {t}} _ {r}, \mathbf {t} _ {| \mathcal {C} ^ {\text {b a s e}} |}\right) \right] \tag {1} +$$ + +$$ +\mathcal {L} _ {\text {t e x t}} = \frac {1}{| P |} \sum_ {r \in P} \mathcal {L} _ {\mathrm {C E}} (\operatorname {s o f t m a x} (\mathbf {z} (r) / \lambda), c _ {r}), \tag {2} +$$ + +where $\cos (\mathbf{v},\mathbf{k}) = \frac{\mathbf{v}\cdot\mathbf{k}}{||\mathbf{v}||||\mathbf{k}||}$ , $\mathbf{t}_{bg}$ a learned background prompt, $\lambda$ a temperature parameter, $\mathcal{L}_{\mathrm{CE}}$ the cross-entropy loss and $c_{r}$ is the class label of $r$ . Furthermore, we align each $\hat{\mathbf{i}}_r$ with the CLIP image encoder $\mathcal{I}$ . For each $r$ , we crop the input image to $\mathbf{b}_r$ , and resize it to the required input size to obtain the image embedding $\mathbf{i}_r = \mathcal{I}(\mathcal{R}(I,\mathbf{b}_r))$ . We minimize the distance between the corresponding $\hat{\mathbf{i}}_r$ and $\mathbf{i}_r$ . + +$$ +\mathcal {L} _ {\text {i m a g e}} = \frac {1}{| P |} \sum_ {r \in P} \left\| \hat {\mathbf {i}} _ {r} - \mathbf {i} _ {r} \right\| _ {1}. \tag {3} +$$ + +3) Association: An open-vocabulary tracker should handle diverse scenarios that comprise complex camera motion and heterogeneous object motion patterns. However, those patterns are difficult to model especially when there are not enough video annotations available [16, 37]. Therefore, we rely on appearance cues to robustly track objects in an open-vocabulary context. Specifically, we employ a contrastive learning approach inspired by [16, 34]. Given an image pair $(I_{\mathrm{key}}, I_{\mathrm{ref}})$ we extract RoIs from both images and match the RoIs to the annotations using intersection-over-union (IoU). For each matched RoI in $I_{\mathrm{key}}$ with appearance embedding $\mathbf{q} \in Q$ , we cluster objects $Q^{+}$ with the same identity and divide objects $Q^{-}$ with different identity in $I_{\mathrm{ref}}$ . + +$$ +\operatorname {P o s D} (\mathbf {q}) = \frac {1}{| Q ^ {+} (\mathbf {q})) |} \sum_ {\mathbf {q} ^ {+} \in Q ^ {+}} \exp \left(\mathbf {q} \cdot \mathbf {q} ^ {+}\right) / \tau), \tag {4} +$$ + +$$ +\operatorname {S i m} (\mathbf {q}) = \frac {\exp \left(\mathbf {q} \cdot \mathbf {q} ^ {+} / \tau\right)}{\operatorname {P o s D} (\mathbf {q}) + \sum_ {\mathbf {q} ^ {-} \in Q ^ {-}} \exp \left(\mathbf {q} \cdot \mathbf {q} ^ {-} / \tau\right)}, \tag {5} +$$ + +$$ +\mathcal {L} _ {\text {t r a c k}} = - \sum_ {\mathbf {q} \in Q} \frac {1}{| Q ^ {+} (\mathbf {q}) |} \sum_ {\mathbf {q} ^ {+} \in Q ^ {+} (\mathbf {q})} \log \left(\operatorname {S i m} \left(\mathbf {q} ^ {+}\right)\right). \tag {6} +$$ + +We further apply an auxiliary loss $\mathcal{L}_{\mathrm{aux}}$ to constrain the magnitude of the logits following [16]. + +During inference, we use straightforward appearance feature similarity for associating existing tracks $\mathcal{T}$ with objects in $P$ . In particular, for each track $\tau \in \mathcal{T}$ and its corresponding appearance embedding $\mathbf{q}_{\tau}$ , we compare its similarity with all candidate objects $r \in P$ using appearance embedding $\mathbf{q}_r$ . We measure the similarity $\mathbf{s}(\tau, r)$ of existing tracks with the candidate objects using both bi-directional softmax [16] and cosine similarity. We assign $r$ to the track $\tau$ with its maximum similarity, if $\mathbf{s}(\tau, r) > \beta$ . If $r$ does not have a matching track, it starts a new track if its confidence $p_r > \gamma$ and is discarded otherwise. The inference pipeline is illustrated in Fig. 4. + +# 4.2. Learning to track without video data + +In this section, we focus on how to train OVTrack in an open-vocabulary context. In particular, open-vocabulary MOT is challenging from a data perspective since we need to localize, classify, and associate possibly unknown objects of extremely diverse appearance with fixed method components. Thus, it is of fundamental importance to align our training data distribution $\mathcal{D}^{\mathrm{train}}$ with the conditions found in evaluation. However, existing video datasets lack the diversity of contemporary image datasets. Hence, it is essential for open-vocabulary trackers to leverage not only video but more importantly static image data during training. + +We use the large-scale, diverse image dataset LVIS [21] to train OVTrack. In particular, for each image $I_{\mathrm{key}}$ we generate a reference image $I_{\mathrm{ref}}$ . Referring to our instance similarity loss in Eq. 6, the appearance similarity learning is constituted by contrasting positive and negative examples $\mathbf{q}^{+}$ and $\mathbf{q}^{-}$ . The corollary of this is that learning will be optimal if $\mathbf{q}^{+}$ consists of examples with distortions commonly encountered in video data, such as change in object scale, viewpoint or lighting, while $\mathbf{q}^{-}$ contains examples with appearance changes associated with object identity, such as different material. While distortions like translation, scaling, and rotation can be simulated via classic data augmentation strategies [16, 74, 77], there are certain phenomena like changes in object viewpoint, lighting or context that cannot + +![](images/f0d1096e6bcba4323d732a4a65eda141184a4087260d160b51dd1be8db4a2778.jpg) +Figure 5. Data hallucination strategy. Given an input image, its annotations and its caption, we generate $x_0 \sim n$ and input it to the diffusion model [59], which progressively denoises $x$ from $\delta_0 = 0.75$ to $\eta$ . At the same time, the foreground regions are kept fixed at each iteration. Specifically, we compose the generated images $x_k \sim m_\theta$ with foreground regions $\hat{x}_k \sim n$ at the current noise level, to obtain $x_k$ as input for the next iteration. Finally, we arrive at $I_{\mathrm{ref}}$ which we use for instance similarity learning. + +be simulated by these. + +Therefore, to simulate all desired properties of our instance embedding space in $I_{\mathrm{ref}}$ , we combine classic data augmentations with a DDPM-based data hallucination strategy. We use the data generation fidelity of stable diffusion [59] to simulate $\mathbf{q}^{+}$ and $\mathbf{q}^{-}$ via a specialized denoising process. Generally, the denoising process of a DPPM can be viewed as the inversion of a forward process that maps an input $x$ to Gaussian white noise $\mathcal{N}(0,\mathbf{I})$ . In the forward direction, Gaussian noise is added to the input image according to a variance schedule $\delta_{k}$ in $K$ steps. + +$$ +n \left(x _ {k} \mid x _ {k - 1}\right) = \mathcal {N} \left(x _ {k}; \sqrt {1 - \delta_ {k}} \cdot x _ {k - 1}, \delta_ {k} \mathbf {I}\right). \tag {7} +$$ + +In the backward direction, a neural network with parameters $\theta$ predicts the parameters $\mu$ and $\Sigma$ of a Gaussian distribution that reverses a forward step. + +$$ +m _ {\theta} \left(x _ {k - 1} \mid x _ {k}\right) = \mathcal {N} \left(x _ {k - 1}; \mu_ {\theta} \left(x _ {k}, k\right), \Sigma_ {\theta} \left(x _ {k}, k\right)\right). \tag {8} +$$ + +Our specialized denoising process is illustrated in Fig. 5. We initialize $I_{\mathrm{ref}}$ as $I_{\mathrm{key}}$ and apply a random geometric transformation to $I_{\mathrm{ref}}$ . Next, we use the instance mask annotations in LVIS to define the set of positive examples $A^{+}$ . We divide each iteration of the denoising process of image $I_{\mathrm{ref}}$ using the union of all object masks in $A^{+}$ into two branches following [39]. In addition, we use the conditioning mechanism in [59] to guide the backward process with the corresponding image caption. To initialize the backward process, we set $x_0$ to $I_{\mathrm{ref}}$ at $\delta_0 = 0.75$ via the forward process $n(x_k|x_{k-1})$ . Note that $x$ corresponds to a latent representation of $I_{\mathrm{ref}}$ obtained via the encoder of stable diffusion. In each iteration, we apply $m_\theta(x_{k-1}|x_k)$ to obtain a new sample $x_k \sim m_\theta$ . At the same time, we use the forward process $n(x_k|x_{k-1})$ on the areas of $A^{+}$ to generate $\hat{x}_k \sim n$ . At the end of each reverse iteration, we compose the two versions via + +$x_{k} = A^{+}\cdot \hat{x}_{k} + (1 - A^{+})\cdot x_{k}$ . We iterate until $\delta_k\leq \eta$ Finally, we apply $m_{\theta}(x_{k - 1}|x_k)$ to the whole image, without branching, as a homogenization step between $A^{+}x_{k}$ and $(1 - A^{+})x_{k}$ , while $\eta >\delta_k > 0$ + +By this process, we achieve three goals. First, we generate random perturbations of the background. Second, we keep the areas of $A^{+}$ close to its original content in each denoising step so that positive instances are integrated well into the new background. Third, we generate distractor objects by caption guided hallucination. + +# 5. Experiments + +# 5.1. Evaluation metrics + +TETA. The tracking-everything accuracy (TETA) [34] is calculated from three independent scores. First, the localization accuracy (LocA) is calculated by matching all annotated boxes $\alpha$ to the predicted boxes of $\mathcal{T}$ without taking classification into account: $\mathrm{LocA} = \frac{|\mathrm{TPL}|}{|\mathrm{TPL}| + |\mathrm{FPL}| + |\mathrm{FNL}|}$ . Next, classification accuracy (ClsA) is computed based on all well-localized TPL, comparing the predicted semantic classes to the matched ground-truths: $\mathrm{ClsA} = \frac{|\mathrm{TPC}|}{|\mathrm{TPC}| + |\mathrm{FPC}| + |\mathrm{FNC}|}$ . Finally, association accuracy (AssocA) is computed in a similar fashion, comparing the identity of associated ground truths with well-localized predictions: AssocA = $\frac{1}{|\mathrm{TPL}|}\sum_{b\in \mathrm{TPL}}\frac{|\mathrm{TPA}(b)|}{|\mathrm{TPA}(b)| + |\mathrm{FPA}(b)| + |\mathrm{FNA}(b)|}$ . The TETA score is computed as the arithmetic mean of the three scores. + +Track mAP. The Track mAP [70] is calculated using the 3D IoU between the bounding boxes of a predicted track $\tau$ and an annotated track $\alpha$ by $\mathrm{IoU}_{\mathrm{3D}}(\tau, \alpha) = \frac{\sum_{t \in T} \tau_t \cap \alpha_t}{\sum_{t \in T} \tau_t \cup \alpha_t}$ . It is used analogous to 2D bounding box IoU to calculate the popular average precision metric per class as in [21]. The Track mAP is the average of the per-class scores across a set of IoU3D thresholds. + +Table 1. Open-vocabulary MOT comparison. We compare our method with existing closed-set trackers and off-the-shelf open-vocabulary baselines on base and novel classes on the validation and test sets of TAO [9]. We indicate the classes and data the methods trained on. Note that methods using TAO data utilize videos for training. All methods use ResNet50 [18] as backbone. + +
MethodClassesDataBaseNovel
Validation setBaseNovelCC3MLVISTAOTETALocAAssocAClsATETALocAAssocAClsA
QDTrack [16]✓✓-✓✓27.145.624.711.022.542.724.40.4
TETer [34]✓✓-✓✓30.347.431.612.125.745.931.10.2
DeepSORT (ViLD) [65]✓--✓✓26.947.115.817.721.146.414.72.3
Tracktor++ (ViLD) [3]✓--✓✓28.347.420.517.022.746.719.32.2
OVTrack✓--✓-35.549.336.920.227.848.833.61.5
RegionCLIP [75]
+ DeepSORT [65]✓-✓✓✓28.452.515.617.024.549.215.39.0
+ Tracktor++ [3]✓-✓✓✓29.652.419.616.925.750.118.98.1
+ OVTrack✓-✓✓-36.353.936.318.732.051.433.211.4
Test setBaseNovelCC3MLVISTAOTETALocAAssocAClsATETALocAAssocAClsA
QDTrack [16]✓✓-✓✓25.843.223.510.620.239.720.90.2
TETer [34]✓✓-✓✓29.244.030.410.721.739.125.90.0
DeepSORT (ViLD) [65]✓--✓✓24.543.814.615.217.238.411.61.7
Tracktor++ (ViLD) [3]✓--✓✓26.044.119.014.818.039.013.41.7
OVTrack✓--✓-32.645.635.416.924.141.828.71.8
RegionCLIP [75]
+ DeepSORT [65]✓-✓✓✓27.049.815.116.118.741.89.15.2
+ Tracktor++ [3]✓-✓✓✓28.049.418.815.720.042.412.05.7
+ OVTrack✓-✓✓-34.851.136.117.325.744.826.26.1
+ +Table 2. Closed-set MOT Track mAP comparison. We compare to existing trackers on TAO [9] validation. Competing methods use ResNet101 [24], we use ResNet50 as backbone. All methods use Faster R-CNN [57]. We include results with stronger detectors and additional data in gray. $\dagger$ does not use videos for training. + +
MethodTrack mAP50Track mAP75Track mAP
SORT-TAO [9]13.2--
QDTrack [16]15.95.010.6
GTR† [78]20.4--
TAC [67]17.75.807.30
BIV [66]19.67.3013.6
OVTrack†21.210.615.9
GTR + CenterNet2† [78]22.5--
AOA [12]25.8--
+ +# 5.2. Implementation details + +We use ResNet50 [24] with FPN [35]. We filter object candidates $P$ by non-maximum suppression (NMS) with an IoU threshold of 0.7 and randomly select $|P| = 256$ candidates per image. We set $\lambda = 0.07$ in $\mathcal{L}_{\mathrm{text}}$ . We use a two-stage training process, first training the detection components following [13, 20], second fine-tuning the model for tracking with loss weights 0.25 for $\mathcal{L}_{\mathrm{track}}$ and 1.0 for $\mathcal{L}_{\mathrm{aux}}$ following [16]. We train on the LVIS dataset, with one hallucinated counterpart per image in the dataset. While we use the full dataset for the state-of-the-art comparison, we use a subset of 10,000 images for the ablation studies due to resource constraints. Unless otherwise noted, we use the following data augmentations in training: resizing, random horizontal flipping, color jittering, random affine transformation, and mosaic composition with varying parameters between $I_{\mathrm{key}}$ and $I_{\mathrm{ref}}$ . We use $\eta = 0.02$ for data generation. For inference, we select object candidates $P$ by NMS with + +Table 3. Closed-set MOT TETA comparison. We compare to existing trackers on the TAO [9] validation. Benchmark results are taken from [34]. All competing methods use ResNet101 [24] except AOA [12], we use ResNet50 as backbone. All methods use Faster R-CNN [57]. $\dagger$ does not use videos for training. + +
MethodTETALocAAssocAClsA
SORT-TAO [9]24.848.114.312.1
Tracktor [3]24.247.413.012.1
DeepSORT [65]26.048.417.512.1
AOA [12]25.323.430.621.9
Tracktor++ [9]28.049.022.812.1
QDTrack [16]30.050.527.412.1
TETer [34]33.351.635.013.2
OVTrack†34.749.336.718.1
+ +an IoU threshold of 0.5. We keep a track memory of 10 frames to re-identify objects after occlusion and set $\beta = 0.5$ and $\gamma = 0.0001$ (see Sec. 4.1). + +# 5.3. Comparison to state-of-the-art + +Open-vocabulary MOT. In Tab. 1, we show the open-vocabulary MOT evaluation on the TAO validation and test sets, divided into base classes $\mathcal{C}^{\mathrm{base}}$ and novel classes $\mathcal{C}^{\mathrm{novel}}$ . For details on the setup, please refer to the supplemental material. The baselines we establish are composed of both closed-set and open-vocabulary trackers. We choose the two state-of-the-art closed-set trackers, TETer [34] and QDTrack [16], trained on $\mathcal{C}^{\mathrm{base}} \cup \mathcal{C}^{\mathrm{novel}}$ . In addition, we combine off-the-shelf trackers DeepSORT [65] and Tracktor++ [3] with the open-vocabulary detector ViLD [20] as baseline open-vocabulary trackers. These are, like OVTrack, trained on $\mathcal{C}^{\mathrm{base}}$ only. Note that all baselines use video data for training, while we use only static images. + +Our approach substantially outperforms all closed-set and + +open-vocabulary baselines. We achieve consistent improvement across LocA, AssocA, and ClsA on both base and novel classes. The baselines trained on $\mathcal{C}^{\mathrm{base}}\cup \mathcal{C}^{\mathrm{novel}}$ can, in some cases, correctly classify objects in $\mathcal{C}^{\mathrm{novel}}$ but achieve poor results. On the contrary, both the open-vocabulary baselines and our tracker achieve significantly higher ClsA on novel classes. However, we note that classification on the TAO dataset remains a very challenging task. The absolute ClsA scores on novel classes are low. This is partially due to the nature of the ClsA metric, which only considers top-1 classification accuracy, while classes on the TAO dataset are diverse and fine-grained. + +Therefore, we investigate the use of stronger, recently proposed open-vocabulary detectors. We combine RegionCLIP [75] with our off-the-shelf baselines and OVTrack. We replace the localization and classification parts of OVTrack with RegionCLIP while keeping the association fixed. We observe that ClsA increases substantially for all trackers on novel classes. Our method achieves the best performance by a wide margin and achieves the best ClsA scores with 11.4 and 6.1 on the validation and test sets, respectively. Note however that RegionCLIP makes use of additional data. + +Closed-set MOT. In Tab. 2 and Tab. 3 we compare to existing works on the validation split of TAO using Track mAP and TETA metrics, respectively. Note that our method neither uses video data for training, nor is it trained on rare classes as defined in Sec. 3.1, while all of the compared closed-set trackers train on video data and use the held-out rare classes for training as they are part of the closed-set evaluation in TAO. We outperform all previous works by a sizable margin on both metrics. By examining the TETA scores in Tab. 3, we observe that our tracker obtains 2.3 points less in LocA compared to TETer [34]. However, our approach beats TETer in terms of AssocA by 1.7 points and greatly improves in ClsA by 4.9 points, illustrating the positive effect of CLIP distillation on both classification and associated compared to closed-set trackers. This validates our design in Sec. 4.1. Note that while AOA [12] has a better ClsA, it ensembles multiple few-shot detection and re-identification models trained on additional datasets as reported by previous works [34, 78]. Overall, our approach surpasses the previous state-of-the-art by 1.4 points in TETA and 2.3 points in Track mAP while using a weaker backbone and the same detector. + +# 5.4. Ablation studies + +CLIP knowledge distillation. In Tab. 4 we analyze the effect of the knowledge distillation described in Sec. 4.1. In particular, we observe that using both $\mathcal{L}_{\mathrm{text}}$ and $\mathcal{L}_{\mathrm{image}}$ is more effective for classification than using only $\mathcal{L}_{\mathrm{text}}$ , improving ClsA significantly from 15.6 to 18.1, while LocA and AssocA stay at the same level of performance. + +Data hallucination strategy. We validate the effective + +Table 4. Ablation study on CLIP knowledge distillation. We show that using both $\mathcal{L}_{\mathrm{text}}$ and $\mathcal{L}_{\mathrm{image}}$ is important to classification performance when doing CLIP knowledge distillation (Sec. 4.1). + +
\( \mathcal{L}_{\text{text}} \)\( \mathcal{L}_{\text{text}} \mid \text{image} \)TETALocAAssocAClsA
✓-34.050.535.715.6
✓✓34.349.335.418.1
+ +Table 5. Ablation study on data hallucination strategy. We show that our data hallucination strategy ('DDPM', Sec. 4.2) improves the association of a closed-set tracker [34] and our OVTrack on TAO [9] validation. We ensemble it with data augmentations, where 'Standard' refers to random resize and horizontal flip, 'Heavy' to color jitter, random affine transformation and mosaic. + +
StandardDDPMHeavyTETALocAAssocAClsA
TETer-SwinT
✓--32.350.730.615.5
✓✓-33.251.233.015.4
✓✓✓34.351.435.515.8
OVTrack
✓--32.548.931.117.6
✓✓-33.348.932.918.0
✓✓✓34.449.135.718.3
+ +ness of our data hallucination strategy described in Sec. 4.2 by training both the closed-set tracker TETer [34] and our OVTrack with it. Note that we choose to use SwinT [38] with TETer and ResNet50 [24] with our OVTrack to achieve similar performance, in order to fairly compare the performance difference on both trackers. Tab. 5 shows the TETA results on the TAO validation set. We observe that our data hallucination strategy improves the AssocA significantly for both trackers, while LocA and ClsA are comparable. In particular, we improve 2.4 and 1.8 points in AssocA for TETer and OVTrack, respectively. Further, ensembling our data generation strategy with heavy data augmentations yields another 2.5 and 1.8 points improvement. Overall, we show that our data generation strategy improves instance similarity learning across both closed-set and open-vocabulary trackers while being complementary to classic data augmentation. + +# 6. Conclusion + +This work introduced open-vocabulary MOT as an effective solution to evaluating multi-object trackers beyond pre-defined training categories. We defined a suitable benchmark setting and presented OVTrack, a data-efficient open-vocabulary tracker. By using knowledge distillation from vision-language models, we improve tracking while going beyond limited dataset taxonomies. In addition, we put forth a data hallucination strategy tailored to instance similarity learning that addresses the data availability problem in open-vocabulary MOT. As a result, OVTrack learns tracking from static images and is able to track arbitrary objects in videos while outperforming existing trackers by a sizable margin on the large-scale, large-vocabulary TAO [9] benchmark. + +# References + +[1] Ankan Bansal, Karan Sikka, Gaurav Sharma, Rama Chellappa, and Ajay Divakaran. Zero-shot object detection. In ECCV, 2018. 2, 3, 4 +[2] Abhijit Bendale and Terrance Boult. Towards open world recognition. In CVPR, 2015. 3 +[3] Philipp Bergmann, Tim Meinhardt, and Laura Leal-Taixe. Tracking without bells and whistles. In ICCV, 2019. 1, 2, 4, 7 +[4] Alex Bewley, Zongyuan Ge, Lionel Ott, Fabio Ramos, and Ben Upcroft. Simple online and realtime tracking. In ICIP, 2016. 1, 2, 4 +[5] Erik Bochinski, Volker Eiselein, and Thomas Sikora. High-speed tracking-by-detection without using image information. In AVSS, 2017. 2 +[6] Guillem Brasó and Laura Leal-Taixe. Learning a neural solver for multiple object tracking. In CVPR, 2020. 2 +[7] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 3 +[8] Yun Chen, Frieda Rong, Shivam Duggal, Shenlong Wang, Xinchen Yan, Sivabalan Manivasagam, Shangjie Xue, Ersin Yumer, and Raquel Urtasun. Geosim: Realistic video simulation via geometry-aware composition for self-driving. In CVPR, 2021. 3 +[9] Achal Dave, Tarasha Khurana, Pavel Tokmakov, Cordelia Schmid, and Deva Ramanan. Tao: A large-scale benchmark for tracking any object. In ECCV, 2020. 1, 2, 4, 7, 8 +[10] Achal Dave, Pavel Tokmakov, and Deva Ramanan. Towards segmenting anything that moves. In CVPRW, 2019. 1, 3, 4 +[11] Patrick Dendorfer, Aljosa Osep, Anton Milan, Konrad Schindler, Daniel Cremers, Ian Reid, Stefan Roth, and Laura Leal-Taixe. Motchallenge: A benchmark for single-camera multiple target tracking. IJCV, 129(4):845-881, 2021. 1 +[12] Fei Du, Bo Xu, Jiasheng Tang, Yuqi Zhang, Fan Wang, and Hao Li. 1st place solution to eccv-tao-2020: Detect and represent any object for tracking. arXiv preprint arXiv:2101.08040, 2021. 2, 7, 8 +[13] Yu Du, Fangyun Wei, Zihe Zhang, Miaojing Shi, Yue Gao, and Guoqi Li. Learning to prompt for open-vocabulary object detection with vision-language model. In CVPR, 2022. 2, 3, 4, 7 +[14] Matteo Fabbri, Guillem Brasó, Gianluca Maugeri, Orcun Cetintas, Riccardo Gasparini, Aljoša Osep, Simone Calderara, Laura Leal-Taixe, and Rita Cucchiara. Motsynth: How can synthetic data help pedestrian detection and tracking? In CVPR, 2021. 3 +[15] Christoph Feichtenhofer, Axel Pinz, and Andrew Zisserman. Detect to track and track to detect. In ICCV, 2017. 2 +[16] Tobias Fischer, Jiangmiao Pang, Thomas E Huang, Linlu Qiu, Haofeng Chen, Trevor Darrell, and Fisher Yu. Qdtrack: Quasi-dense similarity learning for appearance-only multiple object tracking. arXiv preprint arXiv:2210.06984, 2022. 2, 3, 4, 5, 7 +[17] Adrien Gaidon, Qiao Wang, Yohann Cabon, and Eleonora Vig. Virtual worlds as proxy for multi-object tracking analysis. In CVPR, 2016. 3 + +[18] Shang-Hua Gao, Ming-Ming Cheng, Kai Zhao, Xin-Yu Zhang, Ming-Hsuan Yang, and Philip Torr. Res2net: A new multi-scale backbone architecture. IEEE TPAMI, 43(2):652-662, 2019. 7 +[19] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In CVPR, 2012. 1 +[20] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. arXiv preprint arXiv:2104.13921, 2021. 2, 3, 4, 7 +[21] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In CVPR, 2019. 4, 5, 6 +[22] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, 2020. 3 +[23] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, 2017. 2 +[24] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 7, 8 +[25] David Held, Jesse Levinson, and Sebastian Thrun. Precision tracking with sparse 3d and dense color 2d data. In ICRA, 2013. 2 +[26] David Held, Sebastian Thrun, and Silvio Savarese. Learning to track at 100 FPS with deep regression networks. In ECCV, 2016. 2 +[27] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 2020. 3 +[28] Hou-Ning Hu, Yung-Hsu Yang, Tobias Fischer, Trevor Darrell, Fisher Yu, and Min Sun. Monocular quasi-dense 3d object tracking. IEEE TPAMI, 2022. 3 +[29] KJ Joseph, Salman Khan, Fahad Shahbaz Khan, and Vineeth N Balasubramanian. Towards open world object detection. In CVPR, 2021. 3 +[30] Anna Khoreva, Rodrigo Benenson, Eddy Ilg, Thomas Brox, and Bernt Schiele. Lucid data dreaming for video object segmentation. *IJCV*, 127(9):1175–1197, 2019. 3 +[31] Seung Wook Kim, Jonah Philion, Antonio Torralba, and Sanja Fidler. Drivegan: Towards a controllable high-quality neural simulation. In CVPR, 2021. 3 +[32] Laura Leal-Taixe, Cristian Canton-Ferrer, and Konrad Schindler. Learning by tracking: Siamese CNN for robust target association. In CVPRW, 2016. 2 +[33] Bastian Leibe, Ales Leonardis, and Bernt Schiele. Robust object detection with interleaved categorization and segmentation. IJCV, 77(1):259-289, 2008. 1 +[34] Siyuan Li, Martin Danelljan, Henghui Ding, Thomas E Huang, and Fisher Yu. Tracking every thing in the wild. In ECCV, 2022. 2, 4, 5, 6, 7, 8 +[35] Tsung-Yi Lin, Piotr Dollar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In CVPR, 2017. 7 +[36] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C Berg. Ssd: Single shot multibox detector. In ECCV, 2016. 2 + +[37] Yang Liu, Idil Esen Zulfikar, Jonathon Luiten, Achal Dave, Deva Ramanan, Bastian Leibe, Aljosa Osep, and Laura Leal-Taixe. Opening up open world tracking. In CVPR, 2022. 1, 3, 4, 5 +[38] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 8 +[39] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In CVPR, 2022. 6 +[40] Jonathon Luiten, Tobias Fischer, and Bastian Leibe. Track to reconstruct and reconstruct to track. RA-L, 5(2):1803-1810, 2020. 2 +[41] Tim Meinhardt, Alexander Kirillov, Laura Leal-Taixe, and Christoph Feichtenhofer. Trackformer: Multi-object tracking with transformers. In CVPR, 2022. 2 +[42] Anton Milan, Seyed Hamid Rezatofighi, Anthony R. Dick, Ian D. Reid, and Konrad Schindler. Online multi-target tracking using recurrent neural networks. In AAAI, 2017. 2 +[43] Anton Milan, Stefan Roth, and Konrad Schindler. Continuous energy minimization for multitarget tracking. IEEE TPAMI, 36(1):58-72, 2013. 1 +[44] Dennis Mitzel and Bastian Leibe. Taking mobile multi-object tracking to the next level: People, unknown objects, and carried items. In ECCV, 2012. 2, 3 +[45] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 3 +[46] Aljoša Osep, Alexander Hermans, Francis Engelmann, Dirk Klostermann, Markus Mathias, and Bastian Leibe. Multiscale object candidates for generic object tracking in street scenes. In ICRA, 2016. 1, 3 +[47] Aljoša Osep, Wolfgang Mehner, Markus Mathias, and Bastian Leibe. Combined image-and world-space tracking in traffic scenes. In ICRA, 2017. 2 +[48] Aljoša Osep, Wolfgang Mehner, Paul Voigtlaender, and Bastian Leibe. Track, then decide: Category-agnostic vision-based multi-object tracking. In ICRA, 2018. 1, 2, 3 +[49] Aljoša Osep, Paul Voigtlaender, Mark Weber, Jonathon Luiten, and Bastian Leibe. 4d generic video object proposals. In ICRA, 2020. 1, 3 +[50] Jiangmiao Pang, Linlu Qiu, Xia Li, Haofeng Chen, Qi Li, Trevor Darrell, and Fisher Yu. Quasi-dense similarity learning for multiple object tracking. In CVPR, 2021. 2 +[51] Jeffrey Pennington, Richard Socher, and Christopher D Manning. Glove: Global vectors for word representation. In EMNLP, 2014. 3 +[52] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 2, 3, 4 +[53] Deva Ramanan and David A Forsyth. Finding and tracking people from the bottom up. In CVPR, 2003. 2 + +[54] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, 2021. 2, 3 +[55] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In CVPR, 2016. 2 +[56] Joseph Redmon and Ali Farhadi. Yolo9000: better, faster, stronger. In CVPR, 2017. 2 +[57] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In NeurIPS, 2015. 2, 4, 7 +[58] Stephan R Richter, Vibhav Vineet, Stefan Roth, and Vladlen Koltun. Playing for data: Ground truth from computer games. In ECCV, 2016. 3 +[59] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 3, 6 +[60] Amir Sadeghian, Alexandre Alahi, and Silvio Savarese. Tracking the untrackable: Learning to track multiple cues with long-term dependencies. In ICCV, 2017. 2 +[61] Samuel Schulter, Paul Vernaza, Wongun Choi, and Manmohan Chandraker. Deep network flow for multi-object tracking. In CVPR, 2017. 2 +[62] Sarthak Sharma, Junaid Ahmed Ansari, J. Krishna Murthy, and K. Madhava Krishna. Beyond pixels: Leveraging geometry and shape cues for online multi-object tracking. In ICRA, 2018. 2 +[63] Peize Sun, Jinkun Cao, Yi Jiang, Rufeng Zhang, Enze Xie, Zehuan Yuan, Changhu Wang, and Ping Luo. Transtrack: Multiple object tracking with transformer. arXiv preprint arXiv:2012.15460, 2020. 2 +[64] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In CVPR, 2020. 1 +[65] Nicolai Wojke, Alex Bewley, and Dietrich Paulus. Simple online and realtime tracking with a deep association metric. In ICIP, 2017. 2, 7 +[66] Sanghyun Woo, Kwanyong Park, Seoung Wug Oh, In So Kweon, and Joon-Young Lee. Bridging images and videos: A simple learning framework for large vocabulary video object detection. In ECCV, 2022. 3, 7 +[67] Sanghyun Woo, Kwanyong Park, Seoung Wug Oh, In So Kweon, and Joon-Young Lee. Tracking by associating clips. In ECCV, 2022. 7 +[68] Bin Xiao, Haiping Wu, and Yichen Wei. Simple baselines for human pose estimation and tracking. In ECCV, 2018. 2 +[69] Bo Yang and Ram Nevatia. An online learned CRF model for multi-target tracking. In CVPR, 2012. 2 +[70] Linjie Yang, Yuchen Fan, and Ning Xu. Video instance segmentation. In ICCV, 2019. 2, 4, 6 +[71] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darryll. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In CVPR, 2020. 1 + +[72] Alireza Zareian, Kevin Dela Rosa, Derek Hao Hu, and Shih-Fu Chang. Open-vocabulary object detection using captions. In CVPR, 2021. 3 +[73] Fangao Zeng, Bin Dong, Tiancai Wang, Xiangyu Zhang, and Yichen Wei. Motr: End-to-end multiple-object tracking with transformer. arXiv preprint arXiv:2105.03247, 2021. 2 +[74] Yifu Zhang, Chunyu Wang, Xinggang Wang, Wenjun Zeng, and Wenyu Liu. Fairmot: On the fairness of detection and re-identification in multiple object tracking. IJCV, 129(11):3069-3087, 2021. 3, 5 +[75] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Liunian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, et al. Regionclip: Region-based language-image pretraining. In CVPR, 2022. 7, 8 +[76] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Phillip Krahenbuhl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision. In ECCV, 2022. 2, 3, 4 +[77] Xingyi Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Tracking objects as points. In ECCV, 2020. 3, 4, 5 +[78] Xingyi Zhou, Tianwei Yin, Vladlen Koltun, and Philipp Krahenbihl. Global tracking transformers. In CVPR, 2022. 2, 7, 8 \ No newline at end of file diff --git a/2023/OVTrack_ Open-Vocabulary Multiple Object Tracking/images.zip b/2023/OVTrack_ Open-Vocabulary Multiple Object Tracking/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7e6ef4c9c7b73d31cbb9d29f391bf9a8f23bf379 --- /dev/null +++ b/2023/OVTrack_ Open-Vocabulary Multiple Object Tracking/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c2ec111f744d54d481370750144df6901def7f360ccdf6f8faf8d169a0ce42d9 +size 583246 diff --git a/2023/OVTrack_ Open-Vocabulary Multiple Object Tracking/layout.json b/2023/OVTrack_ Open-Vocabulary Multiple Object Tracking/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f11dacaadaa2694bc1a8660fdd0f302739cbf3b8 --- /dev/null +++ b/2023/OVTrack_ Open-Vocabulary Multiple Object Tracking/layout.json @@ -0,0 +1,10316 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 127, + 103, + 467, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 103, + 467, + 121 + ], + "spans": [ + { + "bbox": [ + 127, + 103, + 467, + 121 + ], + "type": "text", + "content": "OVTrack: Open-Vocabulary Multiple Object Tracking" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 144, + 143, + 447, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 143, + 447, + 200 + ], + "spans": [ + { + "bbox": [ + 144, + 143, + 447, + 200 + ], + "type": "text", + "content": "Siyuan Li* Tobias Fischer* Lei Ke Henghui Ding Martin Danelljan Fisher Yu Computer Vision Lab, ETH Zürich https://www.vis.xyz/pub/ovtrack/" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "spans": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 252, + 290, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 252, + 290, + 504 + ], + "spans": [ + { + "bbox": [ + 46, + 252, + 290, + 504 + ], + "type": "text", + "content": "The ability to recognize, localize and track dynamic objects in a scene is fundamental to many real-world applications, such as self-driving and robotic systems. Yet, traditional multiple object tracking (MOT) benchmarks rely only on a few object categories that hardly represent the multitude of possible objects that are encountered in the real world. This leaves contemporary MOT methods limited to a small set of pre-defined object categories. In this paper, we address this limitation by tackling a novel task, open-vocabulary MOT, that aims to evaluate tracking beyond predefined training categories. We further develop OVTrack, an open-vocabulary tracker that is capable of tracking arbitrary object classes. Its design is based on two key ingredients: First, leveraging vision-language models for both classification and association via knowledge distillation; second, a data hallucination strategy for robust appearance feature learning from denoising diffusion probabilistic models. The result is an extremely data-efficient open-vocabulary tracker that sets a new state-of-the-art on the large-scale, large-vocabulary TAO benchmark, while being trained solely on static images." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 529, + 128, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 529, + 128, + 540 + ], + "spans": [ + { + "bbox": [ + 47, + 529, + 128, + 540 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 549, + 288, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 549, + 288, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 549, + 288, + 693 + ], + "type": "text", + "content": "Multiple Object Tracking (MOT) aims to recognize, localize and track objects in a given video sequence. It is a cornerstone of dynamic scene analysis and vital for many real-world applications such as autonomous driving, augmented reality, and video surveillance. Traditionally, MOT benchmarks [9, 11, 19, 64, 71] define a set of semantic categories that constitute the objects to be tracked in the training and testing data distributions. The potential of traditional MOT methods [3, 4, 33, 43] is therefore limited by the taxonomies of those benchmarks. As a consequence, contemporary MOT methods struggle with unseen events, leading to a gap between evaluation performance and real-world" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 309, + 226, + 545, + 399 + ], + "blocks": [ + { + "bbox": [ + 309, + 226, + 545, + 399 + ], + "lines": [ + { + "bbox": [ + 309, + 226, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 309, + 226, + 545, + 399 + ], + "type": "image", + "image_path": "394eae8e3a8d336e2657acfe267ee8f7a5eb0040cfba683c429a42f56bc6dfde.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 414, + 548, + 470 + ], + "lines": [ + { + "bbox": [ + 304, + 414, + 548, + 470 + ], + "spans": [ + { + "bbox": [ + 304, + 414, + 548, + 470 + ], + "type": "text", + "content": "Figure 1. OVTrack. We approach the task of open-vocabulary multiple object tracking. During training, we leverage vision-language (VL) models both for generating samples and knowledge distillation. During testing, we track both base and novel classes unseen during training by querying a vision-language model." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 486, + 359, + 497 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 486, + 359, + 497 + ], + "spans": [ + { + "bbox": [ + 306, + 486, + 359, + 497 + ], + "type": "text", + "content": "deployment." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 498, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 547, + 713 + ], + "type": "text", + "content": "To bridge this gap, previous works have tackled MOT in an open-world context. In particular, Osep et al. [46, 48] approach generic object tracking by first segmenting the scene and performing tracking before classification. Other works have used class agnostic localizers [10, 49] to perform MOT on arbitrary objects. Recently, Liu et al. [37] defined open-world tracking, a task that focuses on the evaluation of previously unseen objects. In particular, it requires any-object tracking as a stage that precedes object classification. This setup comes with two inherent difficulties. First, in an open-world context, densely annotating all objects is prohibitively expensive. Second, without a pre-defined taxonomy of categories, the notion of what is an object is ambiguous. As a consequence, Liu et al. resort to recall-based evaluation, which is limited in two ways. Penalizing false positives (FP) becomes impossible, i.e. we cannot measure the tracker precision. Moreover, by evaluating tracking in a class-agnostic manner, we lose the ability to evaluate how" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 126, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 126, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 126, + 712 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5567" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 284, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 284, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 284, + 83 + ], + "type": "text", + "content": "well a tracker can infer the semantic category of an object." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 84, + 286, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 84, + 286, + 203 + ], + "spans": [ + { + "bbox": [ + 46, + 84, + 286, + 203 + ], + "type": "text", + "content": "In this paper, we propose open-vocabulary MOT as an effective solution to these problems. Similar to open-world MOT, open-vocabulary MOT aims to track multiple objects beyond the pre-defined training categories. However, instead of dismissing the classification problem and resorting to recall-based evaluation, we assume that at test time we are given the classes of objects we are interested in. This allows us to apply existing closed-set tracking metrics [34, 70] that capture both precision and recall, while still evaluating the tracker's ability to track arbitrary objects during inference." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 204, + 287, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 204, + 287, + 383 + ], + "spans": [ + { + "bbox": [ + 46, + 204, + 287, + 383 + ], + "type": "text", + "content": "We further present the first Open-Vocabulary Tracker, OVTrack (see Fig. 1). To this end, we identify and address two fundamental challenges to the design of an open-vocabulary multi-object tracker. The first is that closed-set MOT methods are simply not capable of extending their pre-defined taxonomies. The second is data availability, i.e. scaling video data annotation to a large vocabulary of classes is extremely costly. Inspired by existing works in open-vocabulary detection [1, 13, 20, 76], we replace our classifier with an embedding head, which allows us to measure similarities of localized objects to an open vocabulary of semantic categories. In particular, we distill knowledge from CLIP [52] into our model by aligning the image feature representations of object proposals with the corresponding CLIP image and text embeddings." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 384, + 287, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 384, + 287, + 527 + ], + "spans": [ + { + "bbox": [ + 46, + 384, + 287, + 527 + ], + "type": "text", + "content": "Beyond detection, association is the core of modern MOT methods. It is driven by two affinity cues: motion and appearance. In an open-vocabulary context, motion cues are brittle since arbitrary scenery contains complex and diverse camera and object motion patterns. In contrast, diverse objects usually exhibit heterogeneous appearance. However, relying on appearance cues requires robust representations that generalize to novel object categories. We find that CLIP feature distillation helps in learning better appearance representations for improved association. This is especially intriguing since object classification and appearance modeling are usually distinct in the MOT pipeline [3, 16, 65]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 528, + 288, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 528, + 288, + 634 + ], + "spans": [ + { + "bbox": [ + 46, + 528, + 288, + 634 + ], + "type": "text", + "content": "Learning robust appearance features also requires strong supervision that captures object appearance changes in different viewpoints, background, and lighting. To approach the data availability problem, we utilize the recent success of denoising diffusion probabilistic models (DDPMs) in image synthesis [54, 59] and propose an effective data hallucination strategy tailored to appearance modeling. In particular, from a static image, we generate both simulated positive and negative instances along with random background perturbations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 635, + 266, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 635, + 266, + 646 + ], + "spans": [ + { + "bbox": [ + 59, + 635, + 266, + 646 + ], + "type": "text", + "content": "The main contributions are summarized as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 649, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 55, + 649, + 287, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 649, + 287, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 649, + 287, + 685 + ], + "type": "text", + "content": "1. We define the task of open-vocabulary MOT and provide a suitable benchmark setting on the large-scale, large-vocabulary MOT benchmark TAO [9]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 288, + 714 + ], + "type": "text", + "content": "2. We develop OVTrack, the first open-vocabulary multi-object tracker. It leverages vision-language models to" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "type": "image", + "bbox": [ + 310, + 72, + 545, + 228 + ], + "blocks": [ + { + "bbox": [ + 310, + 72, + 545, + 228 + ], + "lines": [ + { + "bbox": [ + 310, + 72, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 310, + 72, + 545, + 228 + ], + "type": "image", + "image_path": "dc380e7f66d95f62446cb991af7fa3aa0f494037ef2d7b4804959ed3bc5610ad.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 233, + 547, + 277 + ], + "lines": [ + { + "bbox": [ + 305, + 233, + 547, + 277 + ], + "spans": [ + { + "bbox": [ + 305, + 233, + 547, + 277 + ], + "type": "text", + "content": "Figure 2. OVTrack qualitative results. We condition our tracker on text prompts unseen during training, namely 'heron', 'hippo' and 'drone', and successfully track the corresponding objects in the videos. The box color depicts object identity." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 325, + 294, + 545, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 294, + 545, + 317 + ], + "spans": [ + { + "bbox": [ + 325, + 294, + 545, + 317 + ], + "type": "text", + "content": "improve both classification and association compared to closed-set trackers." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 323, + 545, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 323, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 313, + 323, + 545, + 360 + ], + "type": "text", + "content": "3. We propose an effective data hallucination strategy that allows us to address the data availability problem in open-vocabulary settings by leveraging DDPMs." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 369, + 547, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 369, + 547, + 440 + ], + "spans": [ + { + "bbox": [ + 304, + 369, + 547, + 440 + ], + "type": "text", + "content": "Owing to its thoughtful design, OVTrack sets a new state-of-the-art on the challenging TAO benchmark [9], outperforming existing trackers by a significant margin while being trained on static images only. In addition, OVTrack is capable of tracking arbitrary object classes (see Fig. 2), overcoming the limitation of closed-set trackers." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 453, + 389, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 453, + 389, + 466 + ], + "spans": [ + { + "bbox": [ + 306, + 453, + 389, + 466 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 474, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 547, + 713 + ], + "type": "text", + "content": "Multiple object tracking. The dominant paradigm in MOT literature is tracking-by-detection [53], where objects are first detected in each frame, and subsequently associated across time. Thus, many works have focused on data association, aiming to exploit similarity cues such as visual appearance [3, 16, 32, 42, 50, 60, 65, 69], 2D object motion [4,5,15,26,68] or 3D object motion [25,40,44,47,48,62] most effectively. Recently, researchers have focused on learning data association with graph neural networks [6, 61] or transformers [41,63,73,78]. However, those works dismiss a more profound problem in the tracking-by-detection pipeline that precedes data association: Contemporary object detectors [23, 36, 55-57] are designed for closed-set scenarios where all objects appear frequently in the training and testing data distributions. Hence, Dave et al. [9] proposed a new benchmark, TAO, that focuses on studying MOT in the long-tail of the object category distribution. On this benchmark, GTR [78], AOA [12], QDTrack [16] and TET [34] achieve impressive performance. However, those works are still limited to pre-defined object categories and thus do not scale" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5568" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 96 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 96 + ], + "type": "text", + "content": "to the diversity of real-world settings. Our work enables tracking of unseen classes from an open vocabulary." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 98, + 288, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 98, + 288, + 217 + ], + "spans": [ + { + "bbox": [ + 46, + 98, + 288, + 217 + ], + "type": "text", + "content": "Open-world detection and tracking. Open-world detection methods aim to detect any salient object in a given input image irrespective of its category and beyond the training data distribution in particular. However, object classification under such a setting is ill-posed since novel classes will be unknown by definition [2, 29]. As such, open-world detection methods utilize class agnostic localizers [10] and treat classification as a clustering problem [29], estimating a similarity between novel instances and grouping them into novel classes via incremental learning." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 219, + 288, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 219, + 288, + 376 + ], + "spans": [ + { + "bbox": [ + 46, + 219, + 288, + 376 + ], + "type": "text", + "content": "Instead, open-vocabulary object detection methods aim to detect arbitrary, but given classes of objects at test time [72]. For this, Bansal et al. [1] connect an object detector with word representations [51]. Recently, models like CLIP [52] learn visual representations from natural language supervision. Their main advantage over word representations is better alignment of visual concepts and language description. Consequently, many works have focused on leveraging image-text representations for open-vocabulary and few-shot object detection [20, 76]. ViLD [20] distills CLIP image features, while Detic [76] leverages classification data for joint training. Other works have focused on learning good language prompts for open-vocabulary object detection [13]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 376, + 288, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 376, + 288, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 288, + 544 + ], + "type": "text", + "content": "Fewer works have tackled the open-world problem in the MOT domain. Existing works perform scene segmentation and class agnostic tracking before classification [44, 46, 48] or utilize class-agnostic proposal generation [10, 49], similar to open-world detection methods. Liu et al. [37] propose an open-world tracking benchmark, TAO-OW, that evaluates class-agnostic tracking as a task that precedes classification. However, this comes with the limitation that the evaluation only captures tracker recall and no classification accuracy. Instead of dismissing classification, we pose the problem in a different way, i.e. at test time we know the novel classes we are interested in. This allows us to capture both the precision and recall of novel classes in our evaluation, while our method maintains the ability to track any object." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 545, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 288, + 714 + ], + "type": "text", + "content": "Learning tracking from static images. Since labelled video data is expensive to acquire at scale, recent methods have proposed to use static images to supervise MOT methods [16, 66, 74, 77]. CenterTrack [77] proposes to learn motion offsets from static images by random translation of the input, while FairMOT [74] treats objects in a dataset of static images as unique classes to distinguish. Inspired by recent progress in self-supervised representation learning [7, 22, 45], Fischer et al. [16] propose to utilize data augmentation in combination with a contrastive learning objective to learn appearance-based tracking from static images. We go beyond classic data augmentation used in existing works by generating positive and negative examples of objects along with background perturbations via generative" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 72, + 547, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 547, + 96 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 547, + 96 + ], + "type": "text", + "content": "models, offering a more targeted approach to guiding appearance similarity learning from static images." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 96, + 547, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 96, + 547, + 276 + ], + "spans": [ + { + "bbox": [ + 304, + 96, + 547, + 276 + ], + "type": "text", + "content": "Data generation for tracking. While aforementioned methods alleviate the data availability problem in MOT, there is still room for improvement when it comes to data generation in video tasks. Therefore, a large body of research has focused on data generation strategies that can benefit tracking methods [8, 14, 17, 28, 30, 31, 58]. While early works focused on obtaining synthetic data from computer graphics engines [14, 17, 28, 58], newer approaches combine 3D assets with generative models for improved realism [8, 31]. Fewer works have tackled data generation with generative models only [30]. Recently, DDPMs [27, 54, 59] showed impressive results in image synthesis. We leverage their data generation fidelity to address the data availability problem that is particularly pronounced in open-vocabulary MOT with a novel data hallucination strategy tailored to appearance modeling." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 286, + 443, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 286, + 443, + 300 + ], + "spans": [ + { + "bbox": [ + 305, + 286, + 443, + 300 + ], + "type": "text", + "content": "3. Open-Vocabulary MOT" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 307, + 547, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 307, + 547, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 307, + 547, + 437 + ], + "type": "text", + "content": "In real-world scenarios, object categories follow a long-tailed distribution with a rich vocabulary. The remarkable diversity of the open world cannot be covered by a monolithic dataset. However, existing MOT benchmarks focus on closed-set evaluation, with often only a handful of object classes being evaluated. Furthermore, the task setup requires trackers to only track objects within a small set of training categories. To bridge the gap between existing MOT benchmarks and algorithms and real-world settings, we propose the task of open-vocabulary MOT and define its training and evaluation setup as follows." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 438, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 438, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 438, + 548, + 715 + ], + "type": "text", + "content": "At training time, we train a tracker " + }, + { + "bbox": [ + 304, + 438, + 548, + 715 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 438, + 548, + 715 + ], + "type": "text", + "content": " on the training data distribution " + }, + { + "bbox": [ + 304, + 438, + 548, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{\\mathrm{train}} = \\{\\mathbf{X}^{\\mathrm{train}},\\mathcal{A}^{\\mathrm{train}}\\}" + }, + { + "bbox": [ + 304, + 438, + 548, + 715 + ], + "type": "text", + "content": " that contains video sequences " + }, + { + "bbox": [ + 304, + 438, + 548, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^{\\mathrm{train}}" + }, + { + "bbox": [ + 304, + 438, + 548, + 715 + ], + "type": "text", + "content": " and their respective annotations " + }, + { + "bbox": [ + 304, + 438, + 548, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{A}^{\\mathrm{train}}" + }, + { + "bbox": [ + 304, + 438, + 548, + 715 + ], + "type": "text", + "content": " of objects with semantic categories " + }, + { + "bbox": [ + 304, + 438, + 548, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^{\\mathrm{base}}\\subset \\mathbb{N}" + }, + { + "bbox": [ + 304, + 438, + 548, + 715 + ], + "type": "text", + "content": ". Each annotation " + }, + { + "bbox": [ + 304, + 438, + 548, + 715 + ], + "type": "inline_equation", + "content": "\\alpha \\in \\mathcal{A}^{\\mathrm{train}}" + }, + { + "bbox": [ + 304, + 438, + 548, + 715 + ], + "type": "text", + "content": " consists of a set of states " + }, + { + "bbox": [ + 304, + 438, + 548, + 715 + ], + "type": "inline_equation", + "content": "\\{\\alpha_{t}\\}_{t\\in T}" + }, + { + "bbox": [ + 304, + 438, + 548, + 715 + ], + "type": "text", + "content": " for each frame " + }, + { + "bbox": [ + 304, + 438, + 548, + 715 + ], + "type": "inline_equation", + "content": "t\\in T" + }, + { + "bbox": [ + 304, + 438, + 548, + 715 + ], + "type": "text", + "content": " that the object is visible in. A state " + }, + { + "bbox": [ + 304, + 438, + 548, + 715 + ], + "type": "inline_equation", + "content": "\\alpha_{t} = (\\mathbf{b}_{t},c_{t})" + }, + { + "bbox": [ + 304, + 438, + 548, + 715 + ], + "type": "text", + "content": " comprises the object class " + }, + { + "bbox": [ + 304, + 438, + 548, + 715 + ], + "type": "inline_equation", + "content": "c\\in \\mathbb{N}" + }, + { + "bbox": [ + 304, + 438, + 548, + 715 + ], + "type": "text", + "content": " and the 2D bounding box " + }, + { + "bbox": [ + 304, + 438, + 548, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{b} = [x,y,w,h]" + }, + { + "bbox": [ + 304, + 438, + 548, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 438, + 548, + 715 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 304, + 438, + 548, + 715 + ], + "type": "text", + "content": " is the center location in pixel coordinates and " + }, + { + "bbox": [ + 304, + 438, + 548, + 715 + ], + "type": "inline_equation", + "content": "(w,h)" + }, + { + "bbox": [ + 304, + 438, + 548, + 715 + ], + "type": "text", + "content": " are width and height, respectively. At test time, we are given video sequences " + }, + { + "bbox": [ + 304, + 438, + 548, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^{\\mathrm{test}}" + }, + { + "bbox": [ + 304, + 438, + 548, + 715 + ], + "type": "text", + "content": " and a set of object classes " + }, + { + "bbox": [ + 304, + 438, + 548, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^{\\mathrm{novel}}\\subset \\mathbb{N}\\setminus \\mathcal{C}^{\\mathrm{base}}" + }, + { + "bbox": [ + 304, + 438, + 548, + 715 + ], + "type": "text", + "content": " that we are interested in. We aim to find all tracks " + }, + { + "bbox": [ + 304, + 438, + 548, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 304, + 438, + 548, + 715 + ], + "type": "text", + "content": " of objects in " + }, + { + "bbox": [ + 304, + 438, + 548, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^{\\mathrm{test}}" + }, + { + "bbox": [ + 304, + 438, + 548, + 715 + ], + "type": "text", + "content": " belonging to classes " + }, + { + "bbox": [ + 304, + 438, + 548, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^{\\mathrm{base}}\\cup \\mathcal{C}^{\\mathrm{novel}}" + }, + { + "bbox": [ + 304, + 438, + 548, + 715 + ], + "type": "text", + "content": ". Each track state " + }, + { + "bbox": [ + 304, + 438, + 548, + 715 + ], + "type": "inline_equation", + "content": "\\tau_{t} = (\\mathbf{b}_{t},p_{t},c_{t})\\in \\mathcal{T}" + }, + { + "bbox": [ + 304, + 438, + 548, + 715 + ], + "type": "text", + "content": " contains predicted object confidence " + }, + { + "bbox": [ + 304, + 438, + 548, + 715 + ], + "type": "inline_equation", + "content": "p\\in [0,1]" + }, + { + "bbox": [ + 304, + 438, + 548, + 715 + ], + "type": "text", + "content": ", class " + }, + { + "bbox": [ + 304, + 438, + 548, + 715 + ], + "type": "inline_equation", + "content": "c\\in \\mathbb{N}" + }, + { + "bbox": [ + 304, + 438, + 548, + 715 + ], + "type": "text", + "content": " and 2D bounding box " + }, + { + "bbox": [ + 304, + 438, + 548, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{b} = [x,y,w,h]" + }, + { + "bbox": [ + 304, + 438, + 548, + 715 + ], + "type": "text", + "content": ". The important distinctions to closed-set tracking are two-fold: 1) We evaluate the tracker " + }, + { + "bbox": [ + 304, + 438, + 548, + 715 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 438, + 548, + 715 + ], + "type": "text", + "content": " not only on " + }, + { + "bbox": [ + 304, + 438, + 548, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^{\\mathrm{base}}" + }, + { + "bbox": [ + 304, + 438, + 548, + 715 + ], + "type": "text", + "content": " but also on " + }, + { + "bbox": [ + 304, + 438, + 548, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^{\\mathrm{novel}}" + }, + { + "bbox": [ + 304, + 438, + 548, + 715 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 304, + 438, + 548, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^{\\mathrm{novel}}\\cap \\mathcal{C}^{\\mathrm{base}} = \\emptyset" + }, + { + "bbox": [ + 304, + 438, + 548, + 715 + ], + "type": "text", + "content": ", and 2) While " + }, + { + "bbox": [ + 304, + 438, + 548, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^{\\mathrm{novel}}" + }, + { + "bbox": [ + 304, + 438, + 548, + 715 + ], + "type": "text", + "content": " is known at test time, our setup requires the tracker " + }, + { + "bbox": [ + 304, + 438, + 548, + 715 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 438, + 548, + 715 + ], + "type": "text", + "content": " to track arbitrary object classes " + }, + { + "bbox": [ + 304, + 438, + 548, + 715 + ], + "type": "inline_equation", + "content": "c\\in \\mathbb{N}" + }, + { + "bbox": [ + 304, + 438, + 548, + 715 + ], + "type": "text", + "content": " since " + }, + { + "bbox": [ + 304, + 438, + 548, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^{\\mathrm{novel}}" + }, + { + "bbox": [ + 304, + 438, + 548, + 715 + ], + "type": "text", + "content": " remains unknown at training time. In particular, the evaluation of " + }, + { + "bbox": [ + 304, + 438, + 548, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^{\\mathrm{novel}}" + }, + { + "bbox": [ + 304, + 438, + 548, + 715 + ], + "type": "text", + "content": " illustrates the ability of tracker " + }, + { + "bbox": [ + 304, + 438, + 548, + 715 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 438, + 548, + 715 + ], + "type": "text", + "content": " to track any unknown class " + }, + { + "bbox": [ + 304, + 438, + 548, + 715 + ], + "type": "inline_equation", + "content": "c\\in \\mathbb{N}\\setminus \\mathcal{C}^{\\mathrm{base}}" + }, + { + "bbox": [ + 304, + 438, + 548, + 715 + ], + "type": "text", + "content": ", while the classes in " + }, + { + "bbox": [ + 304, + 438, + 548, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^{\\mathrm{novel}}" + }, + { + "bbox": [ + 304, + 438, + 548, + 715 + ], + "type": "text", + "content": " serve as proxy." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5569" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 64, + 547, + 235 + ], + "blocks": [ + { + "bbox": [ + 50, + 64, + 547, + 235 + ], + "lines": [ + { + "bbox": [ + 50, + 64, + 547, + 235 + ], + "spans": [ + { + "bbox": [ + 50, + 64, + 547, + 235 + ], + "type": "image", + "image_path": "cdb9b574ef41e6b3717b075a9a0509d1fc1a49d112ced4c0a08dd85b4963cbd3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 239, + 547, + 275 + ], + "lines": [ + { + "bbox": [ + 46, + 239, + 547, + 275 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 547, + 275 + ], + "type": "text", + "content": "Figure 3. OVTrack training. From a single static " + }, + { + "bbox": [ + 46, + 239, + 547, + 275 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{key}}" + }, + { + "bbox": [ + 46, + 239, + 547, + 275 + ], + "type": "text", + "content": ", we generate " + }, + { + "bbox": [ + 46, + 239, + 547, + 275 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{ref}}" + }, + { + "bbox": [ + 46, + 239, + 547, + 275 + ], + "type": "text", + "content": " with our data hallucination strategy. We extract RoIs via a RPN [57] and perform knowledge distillation from CLIP [52] via the embeddings of the text and image heads. Note that we train classification only on " + }, + { + "bbox": [ + 46, + 239, + 547, + 275 + ], + "type": "inline_equation", + "content": "C^{\\mathrm{base}}" + }, + { + "bbox": [ + 46, + 239, + 547, + 275 + ], + "type": "text", + "content": ". Further, we obtain appearance embeddings from the tracking head and apply our instance similarity loss on the image pair." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 287, + 126, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 287, + 126, + 298 + ], + "spans": [ + { + "bbox": [ + 47, + 287, + 126, + 298 + ], + "type": "text", + "content": "3.1. Benchmark" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 307, + 288, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 307, + 288, + 426 + ], + "spans": [ + { + "bbox": [ + 46, + 307, + 288, + 426 + ], + "type": "text", + "content": "We utilize the large-scale, large-vocabulary MOT dataset TAO [9] to establish a suitable benchmark for open-vocabulary MOT. TAO mostly follows the taxonomy of LVIS [21], which divides classes according to their occurrence into frequent, common and rare classes. To obtain our held-out set " + }, + { + "bbox": [ + 46, + 307, + 288, + 426 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^{\\mathrm{novel}}" + }, + { + "bbox": [ + 46, + 307, + 288, + 426 + ], + "type": "text", + "content": ", we follow open-vocabulary detection literature [20] and use the rare classes as defined by LVIS. The intuition behind this is that the occurrence of rare classes is correlated with uncommon scenarios and events that we are particularly interested in evaluating." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 427, + 290, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 427, + 290, + 594 + ], + "spans": [ + { + "bbox": [ + 46, + 427, + 290, + 594 + ], + "type": "text", + "content": "With respect to evaluation, the advantage of defining " + }, + { + "bbox": [ + 46, + 427, + 290, + 594 + ], + "type": "inline_equation", + "content": "C^{\\mathrm{novel}}" + }, + { + "bbox": [ + 46, + 427, + 290, + 594 + ], + "type": "text", + "content": " is that we can apply closed-set tracking metrics in a straightforward manner, while open-world MOT [37] needs to resort to recall-based evaluation. Further, previous works [34, 37] have shown that the official evaluation metric in TAO, Track mAP [70], is sub-optimal in terms of handling FPs in presence of missing annotations. On the contrary, the recently proposed TETA metric [34] handles this shortcoming via local cluster evaluation. Also, TETA disentangles classification from localization and association performance. Thus, we choose TETA as the evaluation metric for our setup to provide a comprehensive insight into the localization, association, and open-vocabulary classification performance of tracker " + }, + { + "bbox": [ + 46, + 427, + 290, + 594 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 427, + 290, + 594 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 608, + 111, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 608, + 111, + 620 + ], + "spans": [ + { + "bbox": [ + 47, + 608, + 111, + 620 + ], + "type": "text", + "content": "4. OVTrack" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "type": "text", + "content": "We present our Open-Vocabulary Tracker, OVTrack. We address two perspectives of its design: 1) Model perspective: We show how to handle the open-vocabulary setting in the localization, classification, and association modules of the tracker in Section 4.1; 2) Data perspective: Collecting and annotating the necessary amount of training videos is impractical for open-vocabulary MOT. Therefore, we contribute a" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 288, + 546, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 288, + 546, + 312 + ], + "spans": [ + { + "bbox": [ + 305, + 288, + 546, + 312 + ], + "type": "text", + "content": "novel training approach for learning object tracking without video data in Section 4.2." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 321, + 391, + 334 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 321, + 391, + 334 + ], + "spans": [ + { + "bbox": [ + 306, + 321, + 391, + 334 + ], + "type": "text", + "content": "4.1. Model design" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 340, + 547, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 340, + 547, + 388 + ], + "spans": [ + { + "bbox": [ + 304, + 340, + 547, + 388 + ], + "type": "text", + "content": "We decompose OVTrack's functionality into localization, classification and association and discuss our open-vocabulary design philosophy in tackling difficulties for each of those parts. The model design is illustrated in Fig. 3." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 388, + 547, + 628 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 304, + 388, + 547, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 388, + 547, + 520 + ], + "spans": [ + { + "bbox": [ + 304, + 388, + 547, + 520 + ], + "type": "text", + "content": "1) Localization: To localize objects of arbitrary and possibly unknown classes " + }, + { + "bbox": [ + 304, + 388, + 547, + 520 + ], + "type": "inline_equation", + "content": "c \\in \\mathbb{N}" + }, + { + "bbox": [ + 304, + 388, + 547, + 520 + ], + "type": "text", + "content": " in a video, we train Faster R-CNN [57] in a class-agnostic manner, i.e. we use only the RPN and regression losses defined in [57]. We find that this localization procedure can generalize well to object classes that are unknown at training time, as also validated by previous works [10, 20, 76]. During training, we use RPN proposals as object candidates " + }, + { + "bbox": [ + 304, + 388, + 547, + 520 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 388, + 547, + 520 + ], + "type": "text", + "content": " for greater diversity, while during inference, we use the refined RCNN outputs as object candidates. Each candidate " + }, + { + "bbox": [ + 304, + 388, + 547, + 520 + ], + "type": "inline_equation", + "content": "r \\in P" + }, + { + "bbox": [ + 304, + 388, + 547, + 520 + ], + "type": "text", + "content": " is defined by confidence " + }, + { + "bbox": [ + 304, + 388, + 547, + 520 + ], + "type": "inline_equation", + "content": "p_r" + }, + { + "bbox": [ + 304, + 388, + 547, + 520 + ], + "type": "text", + "content": " and bounding box " + }, + { + "bbox": [ + 304, + 388, + 547, + 520 + ], + "type": "inline_equation", + "content": "\\mathbf{b}_r" + }, + { + "bbox": [ + 304, + 388, + 547, + 520 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 521, + 547, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 547, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 547, + 628 + ], + "type": "text", + "content": "2) Classification: Existing closed-set trackers [3,4,16,77] can only track objects of categories in " + }, + { + "bbox": [ + 304, + 521, + 547, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^{\\mathrm{base}}" + }, + { + "bbox": [ + 304, + 521, + 547, + 628 + ], + "type": "text", + "content": ", i.e. objects present and annotated in the training data distribution " + }, + { + "bbox": [ + 304, + 521, + 547, + 628 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{\\mathrm{train}}" + }, + { + "bbox": [ + 304, + 521, + 547, + 628 + ], + "type": "text", + "content": ". To enable open-vocabulary classification, we need to be able to configure the classes we are interested in without re-training. Inspired by open-vocabulary detection literature [1], we connect our Faster R-CNN with the vision-language model CLIP [52] that has been pre-trained on over 400 million image-text pairs for contrastive learning." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "content": "After extracting the RoI feature embeddings " + }, + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_r = \\mathcal{R}(\\phi(I), \\mathbf{b}_r), \\forall r \\in P" + }, + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "content": " from the backbone " + }, + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "content": ", we replace the original classifier in Faster R-CNN with a text head and add an image head generating the embeddings " + }, + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{t}}_r" + }, + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{i}}_r" + }, + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "content": ", for each " + }, + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_r" + }, + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "content": ". We use the CLIP text and image encoders to supervise the heads following [13, 20]. In particular, we use the class names to generate text prompts " + }, + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{P}(c) = \\{\\mathbf{v}_1^c, \\dots, \\mathbf{v}_L^c, \\mathbf{w}_c\\}" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5570" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 547, + 159 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 547, + 159 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 547, + 159 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 547, + 159 + ], + "type": "image", + "image_path": "0f6b8be833e232ee13125ff5a83561aa957cd3a3d12ff7efa1a326edf3612bd3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 162, + 547, + 196 + ], + "lines": [ + { + "bbox": [ + 46, + 162, + 547, + 196 + ], + "spans": [ + { + "bbox": [ + 46, + 162, + 547, + 196 + ], + "type": "text", + "content": "Figure 4. OVTrack inference. Given an input video stream, we track objects of arbitrary classes, e.g. " + }, + { + "bbox": [ + 46, + 162, + 547, + 196 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^{\\mathrm{base}}\\cup \\mathcal{C}^{\\mathrm{novel}}" + }, + { + "bbox": [ + 46, + 162, + 547, + 196 + ], + "type": "text", + "content": ". We first localize objects agnostic of their class, then assign a semantic class label via the text embedding head, and finally associate them to existing tracks by comparing their appearance embedding " + }, + { + "bbox": [ + 46, + 162, + 547, + 196 + ], + "type": "inline_equation", + "content": "\\mathbf{q}" + }, + { + "bbox": [ + 46, + 162, + 547, + 196 + ], + "type": "text", + "content": " obtained from the tracking head to the appearance embeddings in the track memory." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 213, + 289, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 213, + 289, + 274 + ], + "spans": [ + { + "bbox": [ + 46, + 213, + 289, + 274 + ], + "type": "text", + "content": "that consist of " + }, + { + "bbox": [ + 46, + 213, + 289, + 274 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 46, + 213, + 289, + 274 + ], + "type": "text", + "content": " context vectors " + }, + { + "bbox": [ + 46, + 213, + 289, + 274 + ], + "type": "inline_equation", + "content": "\\mathbf{v}^c" + }, + { + "bbox": [ + 46, + 213, + 289, + 274 + ], + "type": "text", + "content": " and a class name embedding " + }, + { + "bbox": [ + 46, + 213, + 289, + 274 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_c" + }, + { + "bbox": [ + 46, + 213, + 289, + 274 + ], + "type": "text", + "content": ". We feed the prompts into the CLIP text encoder " + }, + { + "bbox": [ + 46, + 213, + 289, + 274 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 46, + 213, + 289, + 274 + ], + "type": "text", + "content": ", generating text embeddings " + }, + { + "bbox": [ + 46, + 213, + 289, + 274 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_c = \\mathcal{E}(\\mathcal{P}(c)), \\forall c \\in \\mathcal{C}^{\\mathrm{base}}" + }, + { + "bbox": [ + 46, + 213, + 289, + 274 + ], + "type": "text", + "content": ". We compute the affinity between the predicted embeddings " + }, + { + "bbox": [ + 46, + 213, + 289, + 274 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{t}}_r" + }, + { + "bbox": [ + 46, + 213, + 289, + 274 + ], + "type": "text", + "content": " and their CLIP counterpart " + }, + { + "bbox": [ + 46, + 213, + 289, + 274 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_c" + }, + { + "bbox": [ + 46, + 213, + 289, + 274 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 279, + 288, + 294 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 279, + 288, + 294 + ], + "spans": [ + { + "bbox": [ + 55, + 279, + 288, + 294 + ], + "type": "interline_equation", + "content": "\\mathbf {z} (r) = \\left[ \\cos \\left(\\hat {\\mathbf {t}} _ {r}, \\mathbf {t} _ {b g}\\right), \\cos \\left(\\hat {\\mathbf {t}} _ {r}, \\mathbf {t} _ {1}\\right), \\dots , \\cos \\left(\\hat {\\mathbf {t}} _ {r}, \\mathbf {t} _ {| \\mathcal {C} ^ {\\text {b a s e}} |}\\right) \\right] \\tag {1}", + "image_path": "9168b904c023707c5ed930a44ac9a928a3ede9a9e793a2c83b197b2d6663ed7a.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 296, + 288, + 324 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 296, + 288, + 324 + ], + "spans": [ + { + "bbox": [ + 55, + 296, + 288, + 324 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t e x t}} = \\frac {1}{| P |} \\sum_ {r \\in P} \\mathcal {L} _ {\\mathrm {C E}} (\\operatorname {s o f t m a x} (\\mathbf {z} (r) / \\lambda), c _ {r}), \\tag {2}", + "image_path": "a626d841478107333f9d4555bb604ec8c84e548fe9f07622b4448950453c2a0c.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 331, + 289, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 331, + 289, + 418 + ], + "spans": [ + { + "bbox": [ + 46, + 331, + 289, + 418 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 331, + 289, + 418 + ], + "type": "inline_equation", + "content": "\\cos (\\mathbf{v},\\mathbf{k}) = \\frac{\\mathbf{v}\\cdot\\mathbf{k}}{||\\mathbf{v}||||\\mathbf{k}||}" + }, + { + "bbox": [ + 46, + 331, + 289, + 418 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 331, + 289, + 418 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_{bg}" + }, + { + "bbox": [ + 46, + 331, + 289, + 418 + ], + "type": "text", + "content": " a learned background prompt, " + }, + { + "bbox": [ + 46, + 331, + 289, + 418 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 331, + 289, + 418 + ], + "type": "text", + "content": " a temperature parameter, " + }, + { + "bbox": [ + 46, + 331, + 289, + 418 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{CE}}" + }, + { + "bbox": [ + 46, + 331, + 289, + 418 + ], + "type": "text", + "content": " the cross-entropy loss and " + }, + { + "bbox": [ + 46, + 331, + 289, + 418 + ], + "type": "inline_equation", + "content": "c_{r}" + }, + { + "bbox": [ + 46, + 331, + 289, + 418 + ], + "type": "text", + "content": " is the class label of " + }, + { + "bbox": [ + 46, + 331, + 289, + 418 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 46, + 331, + 289, + 418 + ], + "type": "text", + "content": ". Furthermore, we align each " + }, + { + "bbox": [ + 46, + 331, + 289, + 418 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{i}}_r" + }, + { + "bbox": [ + 46, + 331, + 289, + 418 + ], + "type": "text", + "content": " with the CLIP image encoder " + }, + { + "bbox": [ + 46, + 331, + 289, + 418 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 46, + 331, + 289, + 418 + ], + "type": "text", + "content": ". For each " + }, + { + "bbox": [ + 46, + 331, + 289, + 418 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 46, + 331, + 289, + 418 + ], + "type": "text", + "content": ", we crop the input image to " + }, + { + "bbox": [ + 46, + 331, + 289, + 418 + ], + "type": "inline_equation", + "content": "\\mathbf{b}_r" + }, + { + "bbox": [ + 46, + 331, + 289, + 418 + ], + "type": "text", + "content": ", and resize it to the required input size to obtain the image embedding " + }, + { + "bbox": [ + 46, + 331, + 289, + 418 + ], + "type": "inline_equation", + "content": "\\mathbf{i}_r = \\mathcal{I}(\\mathcal{R}(I,\\mathbf{b}_r))" + }, + { + "bbox": [ + 46, + 331, + 289, + 418 + ], + "type": "text", + "content": ". We minimize the distance between the corresponding " + }, + { + "bbox": [ + 46, + 331, + 289, + 418 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{i}}_r" + }, + { + "bbox": [ + 46, + 331, + 289, + 418 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 331, + 289, + 418 + ], + "type": "inline_equation", + "content": "\\mathbf{i}_r" + }, + { + "bbox": [ + 46, + 331, + 289, + 418 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 423, + 287, + 452 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 423, + 287, + 452 + ], + "spans": [ + { + "bbox": [ + 106, + 423, + 287, + 452 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {i m a g e}} = \\frac {1}{| P |} \\sum_ {r \\in P} \\left\\| \\hat {\\mathbf {i}} _ {r} - \\mathbf {i} _ {r} \\right\\| _ {1}. \\tag {3}", + "image_path": "65232a919d34f65868e2590738c76fe7d0d6c9535b27a02e8ad445cf433a3ae1.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 457, + 289, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 457, + 289, + 613 + ], + "spans": [ + { + "bbox": [ + 46, + 457, + 289, + 613 + ], + "type": "text", + "content": "3) Association: An open-vocabulary tracker should handle diverse scenarios that comprise complex camera motion and heterogeneous object motion patterns. However, those patterns are difficult to model especially when there are not enough video annotations available [16, 37]. Therefore, we rely on appearance cues to robustly track objects in an open-vocabulary context. Specifically, we employ a contrastive learning approach inspired by [16, 34]. Given an image pair " + }, + { + "bbox": [ + 46, + 457, + 289, + 613 + ], + "type": "inline_equation", + "content": "(I_{\\mathrm{key}}, I_{\\mathrm{ref}})" + }, + { + "bbox": [ + 46, + 457, + 289, + 613 + ], + "type": "text", + "content": " we extract RoIs from both images and match the RoIs to the annotations using intersection-over-union (IoU). For each matched RoI in " + }, + { + "bbox": [ + 46, + 457, + 289, + 613 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{key}}" + }, + { + "bbox": [ + 46, + 457, + 289, + 613 + ], + "type": "text", + "content": " with appearance embedding " + }, + { + "bbox": [ + 46, + 457, + 289, + 613 + ], + "type": "inline_equation", + "content": "\\mathbf{q} \\in Q" + }, + { + "bbox": [ + 46, + 457, + 289, + 613 + ], + "type": "text", + "content": ", we cluster objects " + }, + { + "bbox": [ + 46, + 457, + 289, + 613 + ], + "type": "inline_equation", + "content": "Q^{+}" + }, + { + "bbox": [ + 46, + 457, + 289, + 613 + ], + "type": "text", + "content": " with the same identity and divide objects " + }, + { + "bbox": [ + 46, + 457, + 289, + 613 + ], + "type": "inline_equation", + "content": "Q^{-}" + }, + { + "bbox": [ + 46, + 457, + 289, + 613 + ], + "type": "text", + "content": " with different identity in " + }, + { + "bbox": [ + 46, + 457, + 289, + 613 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{ref}}" + }, + { + "bbox": [ + 46, + 457, + 289, + 613 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 59, + 618, + 287, + 648 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 618, + 287, + 648 + ], + "spans": [ + { + "bbox": [ + 59, + 618, + 287, + 648 + ], + "type": "interline_equation", + "content": "\\operatorname {P o s D} (\\mathbf {q}) = \\frac {1}{| Q ^ {+} (\\mathbf {q})) |} \\sum_ {\\mathbf {q} ^ {+} \\in Q ^ {+}} \\exp \\left(\\mathbf {q} \\cdot \\mathbf {q} ^ {+}\\right) / \\tau), \\tag {4}", + "image_path": "78627641bc4ba6b040975b8a0cd906a408bd84f37c79e072fe81ffca549b0391.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 59, + 651, + 287, + 679 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 651, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 59, + 651, + 287, + 679 + ], + "type": "interline_equation", + "content": "\\operatorname {S i m} (\\mathbf {q}) = \\frac {\\exp \\left(\\mathbf {q} \\cdot \\mathbf {q} ^ {+} / \\tau\\right)}{\\operatorname {P o s D} (\\mathbf {q}) + \\sum_ {\\mathbf {q} ^ {-} \\in Q ^ {-}} \\exp \\left(\\mathbf {q} \\cdot \\mathbf {q} ^ {-} / \\tau\\right)}, \\tag {5}", + "image_path": "880844be19d37f064c5585b758517544f70befc1827f54d92356f09ff8a30e72.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 59, + 681, + 287, + 711 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 681, + 287, + 711 + ], + "spans": [ + { + "bbox": [ + 59, + 681, + 287, + 711 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t r a c k}} = - \\sum_ {\\mathbf {q} \\in Q} \\frac {1}{| Q ^ {+} (\\mathbf {q}) |} \\sum_ {\\mathbf {q} ^ {+} \\in Q ^ {+} (\\mathbf {q})} \\log \\left(\\operatorname {S i m} \\left(\\mathbf {q} ^ {+}\\right)\\right). \\tag {6}", + "image_path": "22da398c1564580ebeccd2600d9c12e2041c537e3395df5d9aa261540dc6ade8.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 213, + 547, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 213, + 547, + 237 + ], + "spans": [ + { + "bbox": [ + 305, + 213, + 547, + 237 + ], + "type": "text", + "content": "We further apply an auxiliary loss " + }, + { + "bbox": [ + 305, + 213, + 547, + 237 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{aux}}" + }, + { + "bbox": [ + 305, + 213, + 547, + 237 + ], + "type": "text", + "content": " to constrain the magnitude of the logits following [16]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 239, + 547, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 239, + 547, + 382 + ], + "spans": [ + { + "bbox": [ + 304, + 239, + 547, + 382 + ], + "type": "text", + "content": "During inference, we use straightforward appearance feature similarity for associating existing tracks " + }, + { + "bbox": [ + 304, + 239, + 547, + 382 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 304, + 239, + 547, + 382 + ], + "type": "text", + "content": " with objects in " + }, + { + "bbox": [ + 304, + 239, + 547, + 382 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 239, + 547, + 382 + ], + "type": "text", + "content": ". In particular, for each track " + }, + { + "bbox": [ + 304, + 239, + 547, + 382 + ], + "type": "inline_equation", + "content": "\\tau \\in \\mathcal{T}" + }, + { + "bbox": [ + 304, + 239, + 547, + 382 + ], + "type": "text", + "content": " and its corresponding appearance embedding " + }, + { + "bbox": [ + 304, + 239, + 547, + 382 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_{\\tau}" + }, + { + "bbox": [ + 304, + 239, + 547, + 382 + ], + "type": "text", + "content": ", we compare its similarity with all candidate objects " + }, + { + "bbox": [ + 304, + 239, + 547, + 382 + ], + "type": "inline_equation", + "content": "r \\in P" + }, + { + "bbox": [ + 304, + 239, + 547, + 382 + ], + "type": "text", + "content": " using appearance embedding " + }, + { + "bbox": [ + 304, + 239, + 547, + 382 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_r" + }, + { + "bbox": [ + 304, + 239, + 547, + 382 + ], + "type": "text", + "content": ". We measure the similarity " + }, + { + "bbox": [ + 304, + 239, + 547, + 382 + ], + "type": "inline_equation", + "content": "\\mathbf{s}(\\tau, r)" + }, + { + "bbox": [ + 304, + 239, + 547, + 382 + ], + "type": "text", + "content": " of existing tracks with the candidate objects using both bi-directional softmax [16] and cosine similarity. We assign " + }, + { + "bbox": [ + 304, + 239, + 547, + 382 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 304, + 239, + 547, + 382 + ], + "type": "text", + "content": " to the track " + }, + { + "bbox": [ + 304, + 239, + 547, + 382 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 239, + 547, + 382 + ], + "type": "text", + "content": " with its maximum similarity, if " + }, + { + "bbox": [ + 304, + 239, + 547, + 382 + ], + "type": "inline_equation", + "content": "\\mathbf{s}(\\tau, r) > \\beta" + }, + { + "bbox": [ + 304, + 239, + 547, + 382 + ], + "type": "text", + "content": ". If " + }, + { + "bbox": [ + 304, + 239, + 547, + 382 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 304, + 239, + 547, + 382 + ], + "type": "text", + "content": " does not have a matching track, it starts a new track if its confidence " + }, + { + "bbox": [ + 304, + 239, + 547, + 382 + ], + "type": "inline_equation", + "content": "p_r > \\gamma" + }, + { + "bbox": [ + 304, + 239, + 547, + 382 + ], + "type": "text", + "content": " and is discarded otherwise. The inference pipeline is illustrated in Fig. 4." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 393, + 501, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 393, + 501, + 406 + ], + "spans": [ + { + "bbox": [ + 306, + 393, + 501, + 406 + ], + "type": "text", + "content": "4.2. Learning to track without video data" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 413, + 547, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 413, + 547, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 413, + 547, + 544 + ], + "type": "text", + "content": "In this section, we focus on how to train OVTrack in an open-vocabulary context. In particular, open-vocabulary MOT is challenging from a data perspective since we need to localize, classify, and associate possibly unknown objects of extremely diverse appearance with fixed method components. Thus, it is of fundamental importance to align our training data distribution " + }, + { + "bbox": [ + 304, + 413, + 547, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{\\mathrm{train}}" + }, + { + "bbox": [ + 304, + 413, + 547, + 544 + ], + "type": "text", + "content": " with the conditions found in evaluation. However, existing video datasets lack the diversity of contemporary image datasets. Hence, it is essential for open-vocabulary trackers to leverage not only video but more importantly static image data during training." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 545, + 548, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 548, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 548, + 713 + ], + "type": "text", + "content": "We use the large-scale, diverse image dataset LVIS [21] to train OVTrack. In particular, for each image " + }, + { + "bbox": [ + 304, + 545, + 548, + 713 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{key}}" + }, + { + "bbox": [ + 304, + 545, + 548, + 713 + ], + "type": "text", + "content": " we generate a reference image " + }, + { + "bbox": [ + 304, + 545, + 548, + 713 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{ref}}" + }, + { + "bbox": [ + 304, + 545, + 548, + 713 + ], + "type": "text", + "content": ". Referring to our instance similarity loss in Eq. 6, the appearance similarity learning is constituted by contrasting positive and negative examples " + }, + { + "bbox": [ + 304, + 545, + 548, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{q}^{+}" + }, + { + "bbox": [ + 304, + 545, + 548, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 545, + 548, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{q}^{-}" + }, + { + "bbox": [ + 304, + 545, + 548, + 713 + ], + "type": "text", + "content": ". The corollary of this is that learning will be optimal if " + }, + { + "bbox": [ + 304, + 545, + 548, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{q}^{+}" + }, + { + "bbox": [ + 304, + 545, + 548, + 713 + ], + "type": "text", + "content": " consists of examples with distortions commonly encountered in video data, such as change in object scale, viewpoint or lighting, while " + }, + { + "bbox": [ + 304, + 545, + 548, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{q}^{-}" + }, + { + "bbox": [ + 304, + 545, + 548, + 713 + ], + "type": "text", + "content": " contains examples with appearance changes associated with object identity, such as different material. While distortions like translation, scaling, and rotation can be simulated via classic data augmentation strategies [16, 74, 77], there are certain phenomena like changes in object viewpoint, lighting or context that cannot" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "5571" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 73, + 544, + 220 + ], + "blocks": [ + { + "bbox": [ + 49, + 73, + 544, + 220 + ], + "lines": [ + { + "bbox": [ + 49, + 73, + 544, + 220 + ], + "spans": [ + { + "bbox": [ + 49, + 73, + 544, + 220 + ], + "type": "image", + "image_path": "f0d1096e6bcba4323d732a4a65eda141184a4087260d160b51dd1be8db4a2778.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 226, + 546, + 271 + ], + "lines": [ + { + "bbox": [ + 46, + 226, + 546, + 271 + ], + "spans": [ + { + "bbox": [ + 46, + 226, + 546, + 271 + ], + "type": "text", + "content": "Figure 5. Data hallucination strategy. Given an input image, its annotations and its caption, we generate " + }, + { + "bbox": [ + 46, + 226, + 546, + 271 + ], + "type": "inline_equation", + "content": "x_0 \\sim n" + }, + { + "bbox": [ + 46, + 226, + 546, + 271 + ], + "type": "text", + "content": " and input it to the diffusion model [59], which progressively denoises " + }, + { + "bbox": [ + 46, + 226, + 546, + 271 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 226, + 546, + 271 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 46, + 226, + 546, + 271 + ], + "type": "inline_equation", + "content": "\\delta_0 = 0.75" + }, + { + "bbox": [ + 46, + 226, + 546, + 271 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 226, + 546, + 271 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 46, + 226, + 546, + 271 + ], + "type": "text", + "content": ". At the same time, the foreground regions are kept fixed at each iteration. Specifically, we compose the generated images " + }, + { + "bbox": [ + 46, + 226, + 546, + 271 + ], + "type": "inline_equation", + "content": "x_k \\sim m_\\theta" + }, + { + "bbox": [ + 46, + 226, + 546, + 271 + ], + "type": "text", + "content": " with foreground regions " + }, + { + "bbox": [ + 46, + 226, + 546, + 271 + ], + "type": "inline_equation", + "content": "\\hat{x}_k \\sim n" + }, + { + "bbox": [ + 46, + 226, + 546, + 271 + ], + "type": "text", + "content": " at the current noise level, to obtain " + }, + { + "bbox": [ + 46, + 226, + 546, + 271 + ], + "type": "inline_equation", + "content": "x_k" + }, + { + "bbox": [ + 46, + 226, + 546, + 271 + ], + "type": "text", + "content": " as input for the next iteration. Finally, we arrive at " + }, + { + "bbox": [ + 46, + 226, + 546, + 271 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{ref}}" + }, + { + "bbox": [ + 46, + 226, + 546, + 271 + ], + "type": "text", + "content": " which we use for instance similarity learning." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 286, + 138, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 286, + 138, + 297 + ], + "spans": [ + { + "bbox": [ + 47, + 286, + 138, + 297 + ], + "type": "text", + "content": "be simulated by these." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 298, + 290, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 298, + 290, + 418 + ], + "spans": [ + { + "bbox": [ + 46, + 298, + 290, + 418 + ], + "type": "text", + "content": "Therefore, to simulate all desired properties of our instance embedding space in " + }, + { + "bbox": [ + 46, + 298, + 290, + 418 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{ref}}" + }, + { + "bbox": [ + 46, + 298, + 290, + 418 + ], + "type": "text", + "content": ", we combine classic data augmentations with a DDPM-based data hallucination strategy. We use the data generation fidelity of stable diffusion [59] to simulate " + }, + { + "bbox": [ + 46, + 298, + 290, + 418 + ], + "type": "inline_equation", + "content": "\\mathbf{q}^{+}" + }, + { + "bbox": [ + 46, + 298, + 290, + 418 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 298, + 290, + 418 + ], + "type": "inline_equation", + "content": "\\mathbf{q}^{-}" + }, + { + "bbox": [ + 46, + 298, + 290, + 418 + ], + "type": "text", + "content": " via a specialized denoising process. Generally, the denoising process of a DPPM can be viewed as the inversion of a forward process that maps an input " + }, + { + "bbox": [ + 46, + 298, + 290, + 418 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 298, + 290, + 418 + ], + "type": "text", + "content": " to Gaussian white noise " + }, + { + "bbox": [ + 46, + 298, + 290, + 418 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(0,\\mathbf{I})" + }, + { + "bbox": [ + 46, + 298, + 290, + 418 + ], + "type": "text", + "content": ". In the forward direction, Gaussian noise is added to the input image according to a variance schedule " + }, + { + "bbox": [ + 46, + 298, + 290, + 418 + ], + "type": "inline_equation", + "content": "\\delta_{k}" + }, + { + "bbox": [ + 46, + 298, + 290, + 418 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 46, + 298, + 290, + 418 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 298, + 290, + 418 + ], + "type": "text", + "content": " steps." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 75, + 428, + 287, + 443 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 428, + 287, + 443 + ], + "spans": [ + { + "bbox": [ + 75, + 428, + 287, + 443 + ], + "type": "interline_equation", + "content": "n \\left(x _ {k} \\mid x _ {k - 1}\\right) = \\mathcal {N} \\left(x _ {k}; \\sqrt {1 - \\delta_ {k}} \\cdot x _ {k - 1}, \\delta_ {k} \\mathbf {I}\\right). \\tag {7}", + "image_path": "91fc0b51414df4e8f47f19f726daf1a1fcae54584f9959c0a4d252f4b07f22f2.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 453, + 287, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 453, + 287, + 489 + ], + "spans": [ + { + "bbox": [ + 47, + 453, + 287, + 489 + ], + "type": "text", + "content": "In the backward direction, a neural network with parameters " + }, + { + "bbox": [ + 47, + 453, + 287, + 489 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 47, + 453, + 287, + 489 + ], + "type": "text", + "content": " predicts the parameters " + }, + { + "bbox": [ + 47, + 453, + 287, + 489 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 47, + 453, + 287, + 489 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 453, + 287, + 489 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 47, + 453, + 287, + 489 + ], + "type": "text", + "content": " of a Gaussian distribution that reverses a forward step." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 498, + 287, + 512 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 498, + 287, + 512 + ], + "spans": [ + { + "bbox": [ + 61, + 498, + 287, + 512 + ], + "type": "interline_equation", + "content": "m _ {\\theta} \\left(x _ {k - 1} \\mid x _ {k}\\right) = \\mathcal {N} \\left(x _ {k - 1}; \\mu_ {\\theta} \\left(x _ {k}, k\\right), \\Sigma_ {\\theta} \\left(x _ {k}, k\\right)\\right). \\tag {8}", + "image_path": "97cb0438cdcf5ba4344b47257405b3da1d624df3b3503d69580d59c01cb6e925.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 521, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 289, + 713 + ], + "type": "text", + "content": "Our specialized denoising process is illustrated in Fig. 5. We initialize " + }, + { + "bbox": [ + 46, + 521, + 289, + 713 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{ref}}" + }, + { + "bbox": [ + 46, + 521, + 289, + 713 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 46, + 521, + 289, + 713 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{key}}" + }, + { + "bbox": [ + 46, + 521, + 289, + 713 + ], + "type": "text", + "content": " and apply a random geometric transformation to " + }, + { + "bbox": [ + 46, + 521, + 289, + 713 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{ref}}" + }, + { + "bbox": [ + 46, + 521, + 289, + 713 + ], + "type": "text", + "content": ". Next, we use the instance mask annotations in LVIS to define the set of positive examples " + }, + { + "bbox": [ + 46, + 521, + 289, + 713 + ], + "type": "inline_equation", + "content": "A^{+}" + }, + { + "bbox": [ + 46, + 521, + 289, + 713 + ], + "type": "text", + "content": ". We divide each iteration of the denoising process of image " + }, + { + "bbox": [ + 46, + 521, + 289, + 713 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{ref}}" + }, + { + "bbox": [ + 46, + 521, + 289, + 713 + ], + "type": "text", + "content": " using the union of all object masks in " + }, + { + "bbox": [ + 46, + 521, + 289, + 713 + ], + "type": "inline_equation", + "content": "A^{+}" + }, + { + "bbox": [ + 46, + 521, + 289, + 713 + ], + "type": "text", + "content": " into two branches following [39]. In addition, we use the conditioning mechanism in [59] to guide the backward process with the corresponding image caption. To initialize the backward process, we set " + }, + { + "bbox": [ + 46, + 521, + 289, + 713 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 46, + 521, + 289, + 713 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 521, + 289, + 713 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{ref}}" + }, + { + "bbox": [ + 46, + 521, + 289, + 713 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 46, + 521, + 289, + 713 + ], + "type": "inline_equation", + "content": "\\delta_0 = 0.75" + }, + { + "bbox": [ + 46, + 521, + 289, + 713 + ], + "type": "text", + "content": " via the forward process " + }, + { + "bbox": [ + 46, + 521, + 289, + 713 + ], + "type": "inline_equation", + "content": "n(x_k|x_{k-1})" + }, + { + "bbox": [ + 46, + 521, + 289, + 713 + ], + "type": "text", + "content": ". Note that " + }, + { + "bbox": [ + 46, + 521, + 289, + 713 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 46, + 521, + 289, + 713 + ], + "type": "text", + "content": " corresponds to a latent representation of " + }, + { + "bbox": [ + 46, + 521, + 289, + 713 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{ref}}" + }, + { + "bbox": [ + 46, + 521, + 289, + 713 + ], + "type": "text", + "content": " obtained via the encoder of stable diffusion. In each iteration, we apply " + }, + { + "bbox": [ + 46, + 521, + 289, + 713 + ], + "type": "inline_equation", + "content": "m_\\theta(x_{k-1}|x_k)" + }, + { + "bbox": [ + 46, + 521, + 289, + 713 + ], + "type": "text", + "content": " to obtain a new sample " + }, + { + "bbox": [ + 46, + 521, + 289, + 713 + ], + "type": "inline_equation", + "content": "x_k \\sim m_\\theta" + }, + { + "bbox": [ + 46, + 521, + 289, + 713 + ], + "type": "text", + "content": ". At the same time, we use the forward process " + }, + { + "bbox": [ + 46, + 521, + 289, + 713 + ], + "type": "inline_equation", + "content": "n(x_k|x_{k-1})" + }, + { + "bbox": [ + 46, + 521, + 289, + 713 + ], + "type": "text", + "content": " on the areas of " + }, + { + "bbox": [ + 46, + 521, + 289, + 713 + ], + "type": "inline_equation", + "content": "A^{+}" + }, + { + "bbox": [ + 46, + 521, + 289, + 713 + ], + "type": "text", + "content": " to generate " + }, + { + "bbox": [ + 46, + 521, + 289, + 713 + ], + "type": "inline_equation", + "content": "\\hat{x}_k \\sim n" + }, + { + "bbox": [ + 46, + 521, + 289, + 713 + ], + "type": "text", + "content": ". At the end of each reverse iteration, we compose the two versions via" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 285, + 547, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 285, + 547, + 335 + ], + "spans": [ + { + "bbox": [ + 304, + 285, + 547, + 335 + ], + "type": "inline_equation", + "content": "x_{k} = A^{+}\\cdot \\hat{x}_{k} + (1 - A^{+})\\cdot x_{k}" + }, + { + "bbox": [ + 304, + 285, + 547, + 335 + ], + "type": "text", + "content": " . We iterate until " + }, + { + "bbox": [ + 304, + 285, + 547, + 335 + ], + "type": "inline_equation", + "content": "\\delta_k\\leq \\eta" + }, + { + "bbox": [ + 304, + 285, + 547, + 335 + ], + "type": "text", + "content": " Finally, we apply " + }, + { + "bbox": [ + 304, + 285, + 547, + 335 + ], + "type": "inline_equation", + "content": "m_{\\theta}(x_{k - 1}|x_k)" + }, + { + "bbox": [ + 304, + 285, + 547, + 335 + ], + "type": "text", + "content": " to the whole image, without branching, as a homogenization step between " + }, + { + "bbox": [ + 304, + 285, + 547, + 335 + ], + "type": "inline_equation", + "content": "A^{+}x_{k}" + }, + { + "bbox": [ + 304, + 285, + 547, + 335 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 285, + 547, + 335 + ], + "type": "inline_equation", + "content": "(1 - A^{+})x_{k}" + }, + { + "bbox": [ + 304, + 285, + 547, + 335 + ], + "type": "text", + "content": " , while " + }, + { + "bbox": [ + 304, + 285, + 547, + 335 + ], + "type": "inline_equation", + "content": "\\eta >\\delta_k > 0" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 335, + 546, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 335, + 546, + 407 + ], + "spans": [ + { + "bbox": [ + 304, + 335, + 546, + 407 + ], + "type": "text", + "content": "By this process, we achieve three goals. First, we generate random perturbations of the background. Second, we keep the areas of " + }, + { + "bbox": [ + 304, + 335, + 546, + 407 + ], + "type": "inline_equation", + "content": "A^{+}" + }, + { + "bbox": [ + 304, + 335, + 546, + 407 + ], + "type": "text", + "content": " close to its original content in each denoising step so that positive instances are integrated well into the new background. Third, we generate distractor objects by caption guided hallucination." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 422, + 388, + 435 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 422, + 388, + 435 + ], + "spans": [ + { + "bbox": [ + 306, + 422, + 388, + 435 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 442, + 417, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 442, + 417, + 453 + ], + "spans": [ + { + "bbox": [ + 306, + 442, + 417, + 453 + ], + "type": "text", + "content": "5.1. Evaluation metrics" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 461, + 547, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 461, + 547, + 624 + ], + "spans": [ + { + "bbox": [ + 304, + 461, + 547, + 624 + ], + "type": "text", + "content": "TETA. The tracking-everything accuracy (TETA) [34] is calculated from three independent scores. First, the localization accuracy (LocA) is calculated by matching all annotated boxes " + }, + { + "bbox": [ + 304, + 461, + 547, + 624 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 461, + 547, + 624 + ], + "type": "text", + "content": " to the predicted boxes of " + }, + { + "bbox": [ + 304, + 461, + 547, + 624 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 304, + 461, + 547, + 624 + ], + "type": "text", + "content": " without taking classification into account: " + }, + { + "bbox": [ + 304, + 461, + 547, + 624 + ], + "type": "inline_equation", + "content": "\\mathrm{LocA} = \\frac{|\\mathrm{TPL}|}{|\\mathrm{TPL}| + |\\mathrm{FPL}| + |\\mathrm{FNL}|}" + }, + { + "bbox": [ + 304, + 461, + 547, + 624 + ], + "type": "text", + "content": ". Next, classification accuracy (ClsA) is computed based on all well-localized TPL, comparing the predicted semantic classes to the matched ground-truths: " + }, + { + "bbox": [ + 304, + 461, + 547, + 624 + ], + "type": "inline_equation", + "content": "\\mathrm{ClsA} = \\frac{|\\mathrm{TPC}|}{|\\mathrm{TPC}| + |\\mathrm{FPC}| + |\\mathrm{FNC}|}" + }, + { + "bbox": [ + 304, + 461, + 547, + 624 + ], + "type": "text", + "content": ". Finally, association accuracy (AssocA) is computed in a similar fashion, comparing the identity of associated ground truths with well-localized predictions: AssocA = " + }, + { + "bbox": [ + 304, + 461, + 547, + 624 + ], + "type": "inline_equation", + "content": "\\frac{1}{|\\mathrm{TPL}|}\\sum_{b\\in \\mathrm{TPL}}\\frac{|\\mathrm{TPA}(b)|}{|\\mathrm{TPA}(b)| + |\\mathrm{FPA}(b)| + |\\mathrm{FNA}(b)|}" + }, + { + "bbox": [ + 304, + 461, + 547, + 624 + ], + "type": "text", + "content": ". The TETA score is computed as the arithmetic mean of the three scores." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 625, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 625, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 625, + 547, + 713 + ], + "type": "text", + "content": "Track mAP. The Track mAP [70] is calculated using the 3D IoU between the bounding boxes of a predicted track " + }, + { + "bbox": [ + 304, + 625, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 625, + 547, + 713 + ], + "type": "text", + "content": " and an annotated track " + }, + { + "bbox": [ + 304, + 625, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 625, + 547, + 713 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 304, + 625, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{IoU}_{\\mathrm{3D}}(\\tau, \\alpha) = \\frac{\\sum_{t \\in T} \\tau_t \\cap \\alpha_t}{\\sum_{t \\in T} \\tau_t \\cup \\alpha_t}" + }, + { + "bbox": [ + 304, + 625, + 547, + 713 + ], + "type": "text", + "content": ". It is used analogous to 2D bounding box IoU to calculate the popular average precision metric per class as in [21]. The Track mAP is the average of the per-class scores across a set of IoU3D thresholds." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5572" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 97, + 105, + 496, + 306 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 547, + 103 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 547, + 103 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 547, + 103 + ], + "type": "text", + "content": "Table 1. Open-vocabulary MOT comparison. We compare our method with existing closed-set trackers and off-the-shelf open-vocabulary baselines on base and novel classes on the validation and test sets of TAO [9]. We indicate the classes and data the methods trained on. Note that methods using TAO data utilize videos for training. All methods use ResNet50 [18] as backbone." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 97, + 105, + 496, + 306 + ], + "lines": [ + { + "bbox": [ + 97, + 105, + 496, + 306 + ], + "spans": [ + { + "bbox": [ + 97, + 105, + 496, + 306 + ], + "type": "table", + "html": "
MethodClassesDataBaseNovel
Validation setBaseNovelCC3MLVISTAOTETALocAAssocAClsATETALocAAssocAClsA
QDTrack [16]✓✓-✓✓27.145.624.711.022.542.724.40.4
TETer [34]✓✓-✓✓30.347.431.612.125.745.931.10.2
DeepSORT (ViLD) [65]✓--✓✓26.947.115.817.721.146.414.72.3
Tracktor++ (ViLD) [3]✓--✓✓28.347.420.517.022.746.719.32.2
OVTrack✓--✓-35.549.336.920.227.848.833.61.5
RegionCLIP [75]
+ DeepSORT [65]✓-✓✓✓28.452.515.617.024.549.215.39.0
+ Tracktor++ [3]✓-✓✓✓29.652.419.616.925.750.118.98.1
+ OVTrack✓-✓✓-36.353.936.318.732.051.433.211.4
Test setBaseNovelCC3MLVISTAOTETALocAAssocAClsATETALocAAssocAClsA
QDTrack [16]✓✓-✓✓25.843.223.510.620.239.720.90.2
TETer [34]✓✓-✓✓29.244.030.410.721.739.125.90.0
DeepSORT (ViLD) [65]✓--✓✓24.543.814.615.217.238.411.61.7
Tracktor++ (ViLD) [3]✓--✓✓26.044.119.014.818.039.013.41.7
OVTrack✓--✓-32.645.635.416.924.141.828.71.8
RegionCLIP [75]
+ DeepSORT [65]✓-✓✓✓27.049.815.116.118.741.89.15.2
+ Tracktor++ [3]✓-✓✓✓28.049.418.815.720.042.412.05.7
+ OVTrack✓-✓✓-34.851.136.117.325.744.826.26.1
", + "image_path": "36c604f62529a7ae52773518b5bef973c64818e93e56b1c266d71a92448fd3bc.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 54, + 376, + 280, + 472 + ], + "blocks": [ + { + "bbox": [ + 46, + 318, + 288, + 374 + ], + "lines": [ + { + "bbox": [ + 46, + 318, + 288, + 374 + ], + "spans": [ + { + "bbox": [ + 46, + 318, + 288, + 374 + ], + "type": "text", + "content": "Table 2. Closed-set MOT Track mAP comparison. We compare to existing trackers on TAO [9] validation. Competing methods use ResNet101 [24], we use ResNet50 as backbone. All methods use Faster R-CNN [57]. We include results with stronger detectors and additional data in gray. " + }, + { + "bbox": [ + 46, + 318, + 288, + 374 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 46, + 318, + 288, + 374 + ], + "type": "text", + "content": " does not use videos for training." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 54, + 376, + 280, + 472 + ], + "lines": [ + { + "bbox": [ + 54, + 376, + 280, + 472 + ], + "spans": [ + { + "bbox": [ + 54, + 376, + 280, + 472 + ], + "type": "table", + "html": "
MethodTrack mAP50Track mAP75Track mAP
SORT-TAO [9]13.2--
QDTrack [16]15.95.010.6
GTR† [78]20.4--
TAC [67]17.75.807.30
BIV [66]19.67.3013.6
OVTrack†21.210.615.9
GTR + CenterNet2† [78]22.5--
AOA [12]25.8--
", + "image_path": "bab4e8d00462206cd00f4a5e657ae82b2e728619bfadca0c9258cfecf1afd66d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 488, + 178, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 488, + 178, + 502 + ], + "spans": [ + { + "bbox": [ + 47, + 488, + 178, + 502 + ], + "type": "text", + "content": "5.2. Implementation details" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 510, + 290, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 290, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 290, + 715 + ], + "type": "text", + "content": "We use ResNet50 [24] with FPN [35]. We filter object candidates " + }, + { + "bbox": [ + 46, + 510, + 290, + 715 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 510, + 290, + 715 + ], + "type": "text", + "content": " by non-maximum suppression (NMS) with an IoU threshold of 0.7 and randomly select " + }, + { + "bbox": [ + 46, + 510, + 290, + 715 + ], + "type": "inline_equation", + "content": "|P| = 256" + }, + { + "bbox": [ + 46, + 510, + 290, + 715 + ], + "type": "text", + "content": " candidates per image. We set " + }, + { + "bbox": [ + 46, + 510, + 290, + 715 + ], + "type": "inline_equation", + "content": "\\lambda = 0.07" + }, + { + "bbox": [ + 46, + 510, + 290, + 715 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 46, + 510, + 290, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{text}}" + }, + { + "bbox": [ + 46, + 510, + 290, + 715 + ], + "type": "text", + "content": ". We use a two-stage training process, first training the detection components following [13, 20], second fine-tuning the model for tracking with loss weights 0.25 for " + }, + { + "bbox": [ + 46, + 510, + 290, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{track}}" + }, + { + "bbox": [ + 46, + 510, + 290, + 715 + ], + "type": "text", + "content": " and 1.0 for " + }, + { + "bbox": [ + 46, + 510, + 290, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{aux}}" + }, + { + "bbox": [ + 46, + 510, + 290, + 715 + ], + "type": "text", + "content": " following [16]. We train on the LVIS dataset, with one hallucinated counterpart per image in the dataset. While we use the full dataset for the state-of-the-art comparison, we use a subset of 10,000 images for the ablation studies due to resource constraints. Unless otherwise noted, we use the following data augmentations in training: resizing, random horizontal flipping, color jittering, random affine transformation, and mosaic composition with varying parameters between " + }, + { + "bbox": [ + 46, + 510, + 290, + 715 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{key}}" + }, + { + "bbox": [ + 46, + 510, + 290, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 510, + 290, + 715 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{ref}}" + }, + { + "bbox": [ + 46, + 510, + 290, + 715 + ], + "type": "text", + "content": ". We use " + }, + { + "bbox": [ + 46, + 510, + 290, + 715 + ], + "type": "inline_equation", + "content": "\\eta = 0.02" + }, + { + "bbox": [ + 46, + 510, + 290, + 715 + ], + "type": "text", + "content": " for data generation. For inference, we select object candidates " + }, + { + "bbox": [ + 46, + 510, + 290, + 715 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 46, + 510, + 290, + 715 + ], + "type": "text", + "content": " by NMS with" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 340, + 376, + 512, + 469 + ], + "blocks": [ + { + "bbox": [ + 304, + 318, + 547, + 374 + ], + "lines": [ + { + "bbox": [ + 304, + 318, + 547, + 374 + ], + "spans": [ + { + "bbox": [ + 304, + 318, + 547, + 374 + ], + "type": "text", + "content": "Table 3. Closed-set MOT TETA comparison. We compare to existing trackers on the TAO [9] validation. Benchmark results are taken from [34]. All competing methods use ResNet101 [24] except AOA [12], we use ResNet50 as backbone. All methods use Faster R-CNN [57]. " + }, + { + "bbox": [ + 304, + 318, + 547, + 374 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 304, + 318, + 547, + 374 + ], + "type": "text", + "content": " does not use videos for training." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 340, + 376, + 512, + 469 + ], + "lines": [ + { + "bbox": [ + 340, + 376, + 512, + 469 + ], + "spans": [ + { + "bbox": [ + 340, + 376, + 512, + 469 + ], + "type": "table", + "html": "
MethodTETALocAAssocAClsA
SORT-TAO [9]24.848.114.312.1
Tracktor [3]24.247.413.012.1
DeepSORT [65]26.048.417.512.1
AOA [12]25.323.430.621.9
Tracktor++ [9]28.049.022.812.1
QDTrack [16]30.050.527.412.1
TETer [34]33.351.635.013.2
OVTrack†34.749.336.718.1
", + "image_path": "9f016f8382f1c61f4495848df41ecbb5a1a651061a8e5600ad45123799ad25b4.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 480, + 547, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 480, + 547, + 518 + ], + "spans": [ + { + "bbox": [ + 304, + 480, + 547, + 518 + ], + "type": "text", + "content": "an IoU threshold of 0.5. We keep a track memory of 10 frames to re-identify objects after occlusion and set " + }, + { + "bbox": [ + 304, + 480, + 547, + 518 + ], + "type": "inline_equation", + "content": "\\beta = 0.5" + }, + { + "bbox": [ + 304, + 480, + 547, + 518 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 480, + 547, + 518 + ], + "type": "inline_equation", + "content": "\\gamma = 0.0001" + }, + { + "bbox": [ + 304, + 480, + 547, + 518 + ], + "type": "text", + "content": " (see Sec. 4.1)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 526, + 471, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 526, + 471, + 539 + ], + "spans": [ + { + "bbox": [ + 305, + 526, + 471, + 539 + ], + "type": "text", + "content": "5.3. Comparison to state-of-the-art" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 545, + 547, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 547, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 547, + 700 + ], + "type": "text", + "content": "Open-vocabulary MOT. In Tab. 1, we show the open-vocabulary MOT evaluation on the TAO validation and test sets, divided into base classes " + }, + { + "bbox": [ + 304, + 545, + 547, + 700 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^{\\mathrm{base}}" + }, + { + "bbox": [ + 304, + 545, + 547, + 700 + ], + "type": "text", + "content": " and novel classes " + }, + { + "bbox": [ + 304, + 545, + 547, + 700 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^{\\mathrm{novel}}" + }, + { + "bbox": [ + 304, + 545, + 547, + 700 + ], + "type": "text", + "content": ". For details on the setup, please refer to the supplemental material. The baselines we establish are composed of both closed-set and open-vocabulary trackers. We choose the two state-of-the-art closed-set trackers, TETer [34] and QDTrack [16], trained on " + }, + { + "bbox": [ + 304, + 545, + 547, + 700 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^{\\mathrm{base}} \\cup \\mathcal{C}^{\\mathrm{novel}}" + }, + { + "bbox": [ + 304, + 545, + 547, + 700 + ], + "type": "text", + "content": ". In addition, we combine off-the-shelf trackers DeepSORT [65] and Tracktor++ [3] with the open-vocabulary detector ViLD [20] as baseline open-vocabulary trackers. These are, like OVTrack, trained on " + }, + { + "bbox": [ + 304, + 545, + 547, + 700 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^{\\mathrm{base}}" + }, + { + "bbox": [ + 304, + 545, + 547, + 700 + ], + "type": "text", + "content": " only. Note that all baselines use video data for training, while we use only static images." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "type": "text", + "content": "Our approach substantially outperforms all closed-set and" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "text", + "content": "5573" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": "open-vocabulary baselines. We achieve consistent improvement across LocA, AssocA, and ClsA on both base and novel classes. The baselines trained on " + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^{\\mathrm{base}}\\cup \\mathcal{C}^{\\mathrm{novel}}" + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": " can, in some cases, correctly classify objects in " + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^{\\mathrm{novel}}" + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": " but achieve poor results. On the contrary, both the open-vocabulary baselines and our tracker achieve significantly higher ClsA on novel classes. However, we note that classification on the TAO dataset remains a very challenging task. The absolute ClsA scores on novel classes are low. This is partially due to the nature of the ClsA metric, which only considers top-1 classification accuracy, while classes on the TAO dataset are diverse and fine-grained." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 217, + 289, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 217, + 289, + 335 + ], + "spans": [ + { + "bbox": [ + 46, + 217, + 289, + 335 + ], + "type": "text", + "content": "Therefore, we investigate the use of stronger, recently proposed open-vocabulary detectors. We combine RegionCLIP [75] with our off-the-shelf baselines and OVTrack. We replace the localization and classification parts of OVTrack with RegionCLIP while keeping the association fixed. We observe that ClsA increases substantially for all trackers on novel classes. Our method achieves the best performance by a wide margin and achieves the best ClsA scores with 11.4 and 6.1 on the validation and test sets, respectively. Note however that RegionCLIP makes use of additional data." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 337, + 289, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 337, + 289, + 600 + ], + "spans": [ + { + "bbox": [ + 46, + 337, + 289, + 600 + ], + "type": "text", + "content": "Closed-set MOT. In Tab. 2 and Tab. 3 we compare to existing works on the validation split of TAO using Track mAP and TETA metrics, respectively. Note that our method neither uses video data for training, nor is it trained on rare classes as defined in Sec. 3.1, while all of the compared closed-set trackers train on video data and use the held-out rare classes for training as they are part of the closed-set evaluation in TAO. We outperform all previous works by a sizable margin on both metrics. By examining the TETA scores in Tab. 3, we observe that our tracker obtains 2.3 points less in LocA compared to TETer [34]. However, our approach beats TETer in terms of AssocA by 1.7 points and greatly improves in ClsA by 4.9 points, illustrating the positive effect of CLIP distillation on both classification and associated compared to closed-set trackers. This validates our design in Sec. 4.1. Note that while AOA [12] has a better ClsA, it ensembles multiple few-shot detection and re-identification models trained on additional datasets as reported by previous works [34, 78]. Overall, our approach surpasses the previous state-of-the-art by 1.4 points in TETA and 2.3 points in Track mAP while using a weaker backbone and the same detector." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 609, + 146, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 609, + 146, + 621 + ], + "spans": [ + { + "bbox": [ + 47, + 609, + 146, + 621 + ], + "type": "text", + "content": "5.4. Ablation studies" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 628, + 288, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 628, + 288, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 628, + 288, + 700 + ], + "type": "text", + "content": "CLIP knowledge distillation. In Tab. 4 we analyze the effect of the knowledge distillation described in Sec. 4.1. In particular, we observe that using both " + }, + { + "bbox": [ + 46, + 628, + 288, + 700 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{text}}" + }, + { + "bbox": [ + 46, + 628, + 288, + 700 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 628, + 288, + 700 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{image}}" + }, + { + "bbox": [ + 46, + 628, + 288, + 700 + ], + "type": "text", + "content": " is more effective for classification than using only " + }, + { + "bbox": [ + 46, + 628, + 288, + 700 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{text}}" + }, + { + "bbox": [ + 46, + 628, + 288, + 700 + ], + "type": "text", + "content": ", improving ClsA significantly from 15.6 to 18.1, while LocA and AssocA stay at the same level of performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 701, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 289, + 713 + ], + "type": "text", + "content": "Data hallucination strategy. We validate the effective" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 335, + 107, + 515, + 147 + ], + "blocks": [ + { + "bbox": [ + 304, + 70, + 545, + 104 + ], + "lines": [ + { + "bbox": [ + 304, + 70, + 545, + 104 + ], + "spans": [ + { + "bbox": [ + 304, + 70, + 545, + 104 + ], + "type": "text", + "content": "Table 4. Ablation study on CLIP knowledge distillation. We show that using both " + }, + { + "bbox": [ + 304, + 70, + 545, + 104 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{text}}" + }, + { + "bbox": [ + 304, + 70, + 545, + 104 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 70, + 545, + 104 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{image}}" + }, + { + "bbox": [ + 304, + 70, + 545, + 104 + ], + "type": "text", + "content": " is important to classification performance when doing CLIP knowledge distillation (Sec. 4.1)." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 335, + 107, + 515, + 147 + ], + "lines": [ + { + "bbox": [ + 335, + 107, + 515, + 147 + ], + "spans": [ + { + "bbox": [ + 335, + 107, + 515, + 147 + ], + "type": "table", + "html": "
\\( \\mathcal{L}_{\\text{text}} \\)\\( \\mathcal{L}_{\\text{text}} \\mid \\text{image} \\)TETALocAAssocAClsA
✓-34.050.535.715.6
✓✓34.349.335.418.1
", + "image_path": "e8c787f701f246c3b9d3724eb14d5bc3cff096815622b6c91bbac73a30e67a99.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 308, + 224, + 547, + 322 + ], + "blocks": [ + { + "bbox": [ + 304, + 155, + 547, + 221 + ], + "lines": [ + { + "bbox": [ + 304, + 155, + 547, + 221 + ], + "spans": [ + { + "bbox": [ + 304, + 155, + 547, + 221 + ], + "type": "text", + "content": "Table 5. Ablation study on data hallucination strategy. We show that our data hallucination strategy ('DDPM', Sec. 4.2) improves the association of a closed-set tracker [34] and our OVTrack on TAO [9] validation. We ensemble it with data augmentations, where 'Standard' refers to random resize and horizontal flip, 'Heavy' to color jitter, random affine transformation and mosaic." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 224, + 547, + 322 + ], + "lines": [ + { + "bbox": [ + 308, + 224, + 547, + 322 + ], + "spans": [ + { + "bbox": [ + 308, + 224, + 547, + 322 + ], + "type": "table", + "html": "
StandardDDPMHeavyTETALocAAssocAClsA
TETer-SwinT
✓--32.350.730.615.5
✓✓-33.251.233.015.4
✓✓✓34.351.435.515.8
OVTrack
✓--32.548.931.117.6
✓✓-33.348.932.918.0
✓✓✓34.449.135.718.3
", + "image_path": "8a8a8292bb1812b9f7f06dad2bc6898f66af85be677b106078372f8ba294debd.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 334, + 547, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 334, + 547, + 526 + ], + "spans": [ + { + "bbox": [ + 304, + 334, + 547, + 526 + ], + "type": "text", + "content": "ness of our data hallucination strategy described in Sec. 4.2 by training both the closed-set tracker TETer [34] and our OVTrack with it. Note that we choose to use SwinT [38] with TETer and ResNet50 [24] with our OVTrack to achieve similar performance, in order to fairly compare the performance difference on both trackers. Tab. 5 shows the TETA results on the TAO validation set. We observe that our data hallucination strategy improves the AssocA significantly for both trackers, while LocA and ClsA are comparable. In particular, we improve 2.4 and 1.8 points in AssocA for TETer and OVTrack, respectively. Further, ensembling our data generation strategy with heavy data augmentations yields another 2.5 and 1.8 points improvement. Overall, we show that our data generation strategy improves instance similarity learning across both closed-set and open-vocabulary trackers while being complementary to classic data augmentation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 537, + 379, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 537, + 379, + 550 + ], + "spans": [ + { + "bbox": [ + 306, + 537, + 379, + 550 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "text", + "content": "This work introduced open-vocabulary MOT as an effective solution to evaluating multi-object trackers beyond pre-defined training categories. We defined a suitable benchmark setting and presented OVTrack, a data-efficient open-vocabulary tracker. By using knowledge distillation from vision-language models, we improve tracking while going beyond limited dataset taxonomies. In addition, we put forth a data hallucination strategy tailored to instance similarity learning that addresses the data availability problem in open-vocabulary MOT. As a result, OVTrack learns tracking from static images and is able to track arbitrary objects in videos while outperforming existing trackers by a sizable margin on the large-scale, large-vocabulary TAO [9] benchmark." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5574" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 53, + 91, + 288, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 288, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 288, + 123 + ], + "type": "text", + "content": "[1] Ankan Bansal, Karan Sikka, Gaurav Sharma, Rama Chellappa, and Ajay Divakaran. Zero-shot object detection. In ECCV, 2018. 2, 3, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 125, + 288, + 147 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 288, + 147 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 288, + 147 + ], + "type": "text", + "content": "[2] Abhijit Bendale and Terrance Boult. Towards open world recognition. In CVPR, 2015. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 148, + 288, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 288, + 170 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 288, + 170 + ], + "type": "text", + "content": "[3] Philipp Bergmann, Tim Meinhardt, and Laura Leal-Taixe. Tracking without bells and whistles. In ICCV, 2019. 1, 2, 4, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 171, + 288, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 171, + 288, + 203 + ], + "spans": [ + { + "bbox": [ + 53, + 171, + 288, + 203 + ], + "type": "text", + "content": "[4] Alex Bewley, Zongyuan Ge, Lionel Ott, Fabio Ramos, and Ben Upcroft. Simple online and realtime tracking. In ICIP, 2016. 1, 2, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 205, + 288, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 205, + 288, + 237 + ], + "spans": [ + { + "bbox": [ + 53, + 205, + 288, + 237 + ], + "type": "text", + "content": "[5] Erik Bochinski, Volker Eiselein, and Thomas Sikora. High-speed tracking-by-detection without using image information. In AVSS, 2017. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 239, + 288, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 239, + 288, + 261 + ], + "spans": [ + { + "bbox": [ + 53, + 239, + 288, + 261 + ], + "type": "text", + "content": "[6] Guillem Brasó and Laura Leal-Taixe. Learning a neural solver for multiple object tracking. In CVPR, 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 263, + 288, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 263, + 288, + 295 + ], + "spans": [ + { + "bbox": [ + 53, + 263, + 288, + 295 + ], + "type": "text", + "content": "[7] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 297, + 288, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 297, + 288, + 350 + ], + "spans": [ + { + "bbox": [ + 53, + 297, + 288, + 350 + ], + "type": "text", + "content": "[8] Yun Chen, Frieda Rong, Shivam Duggal, Shenlong Wang, Xinchen Yan, Sivabalan Manivasagam, Shangjie Xue, Ersin Yumer, and Raquel Urtasun. Geosim: Realistic video simulation via geometry-aware composition for self-driving. In CVPR, 2021. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 353, + 288, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 353, + 288, + 386 + ], + "spans": [ + { + "bbox": [ + 53, + 353, + 288, + 386 + ], + "type": "text", + "content": "[9] Achal Dave, Tarasha Khurana, Pavel Tokmakov, Cordelia Schmid, and Deva Ramanan. Tao: A large-scale benchmark for tracking any object. In ECCV, 2020. 1, 2, 4, 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 388, + 287, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 287, + 409 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 287, + 409 + ], + "type": "text", + "content": "[10] Achal Dave, Pavel Tokmakov, and Deva Ramanan. Towards segmenting anything that moves. In CVPRW, 2019. 1, 3, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 411, + 287, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 411, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 48, + 411, + 287, + 453 + ], + "type": "text", + "content": "[11] Patrick Dendorfer, Aljosa Osep, Anton Milan, Konrad Schindler, Daniel Cremers, Ian Reid, Stefan Roth, and Laura Leal-Taixe. Motchallenge: A benchmark for single-camera multiple target tracking. IJCV, 129(4):845-881, 2021. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 456, + 288, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 288, + 498 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 288, + 498 + ], + "type": "text", + "content": "[12] Fei Du, Bo Xu, Jiasheng Tang, Yuqi Zhang, Fan Wang, and Hao Li. 1st place solution to eccv-tao-2020: Detect and represent any object for tracking. arXiv preprint arXiv:2101.08040, 2021. 2, 7, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 501, + 288, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 501, + 288, + 543 + ], + "spans": [ + { + "bbox": [ + 48, + 501, + 288, + 543 + ], + "type": "text", + "content": "[13] Yu Du, Fangyun Wei, Zihe Zhang, Miaojing Shi, Yue Gao, and Guoqi Li. Learning to prompt for open-vocabulary object detection with vision-language model. In CVPR, 2022. 2, 3, 4, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 545, + 288, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 545, + 288, + 599 + ], + "spans": [ + { + "bbox": [ + 48, + 545, + 288, + 599 + ], + "type": "text", + "content": "[14] Matteo Fabbri, Guillem Brasó, Gianluca Maugeri, Orcun Cetintas, Riccardo Gasparini, Aljoša Osep, Simone Calderara, Laura Leal-Taixe, and Rita Cucchiara. Motsynth: How can synthetic data help pedestrian detection and tracking? In CVPR, 2021. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 601, + 288, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 288, + 622 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 288, + 622 + ], + "type": "text", + "content": "[15] Christoph Feichtenhofer, Axel Pinz, and Andrew Zisserman. Detect to track and track to detect. In ICCV, 2017. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 624, + 288, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 288, + 677 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 288, + 677 + ], + "type": "text", + "content": "[16] Tobias Fischer, Jiangmiao Pang, Thomas E Huang, Linlu Qiu, Haofeng Chen, Trevor Darrell, and Fisher Yu. Qdtrack: Quasi-dense similarity learning for appearance-only multiple object tracking. arXiv preprint arXiv:2210.06984, 2022. 2, 3, 4, 5, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 680, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 288, + 712 + ], + "type": "text", + "content": "[17] Adrien Gaidon, Qiao Wang, Yohann Cabon, and Eleonora Vig. Virtual worlds as proxy for multi-object tracking analysis. In CVPR, 2016. 3" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 38, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 116 + ], + "type": "text", + "content": "[18] Shang-Hua Gao, Ming-Ming Cheng, Kai Zhao, Xin-Yu Zhang, Ming-Hsuan Yang, and Philip Torr. Res2net: A new multi-scale backbone architecture. IEEE TPAMI, 43(2):652-662, 2019. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 118, + 546, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 546, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 546, + 150 + ], + "type": "text", + "content": "[19] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In CVPR, 2012. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 152, + 547, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 547, + 194 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 547, + 194 + ], + "type": "text", + "content": "[20] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. arXiv preprint arXiv:2104.13921, 2021. 2, 3, 4, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 197, + 547, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 547, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 547, + 228 + ], + "type": "text", + "content": "[21] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In CVPR, 2019. 4, 5, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 230, + 547, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 230, + 547, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 230, + 547, + 262 + ], + "type": "text", + "content": "[22] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, 2020. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 264, + 547, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 264, + 547, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 264, + 547, + 285 + ], + "type": "text", + "content": "[23] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, 2017. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 286, + 547, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 286, + 547, + 318 + ], + "spans": [ + { + "bbox": [ + 307, + 286, + 547, + 318 + ], + "type": "text", + "content": "[24] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 7, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 320, + 546, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 320, + 546, + 352 + ], + "spans": [ + { + "bbox": [ + 307, + 320, + 546, + 352 + ], + "type": "text", + "content": "[25] David Held, Jesse Levinson, and Sebastian Thrun. Precision tracking with sparse 3d and dense color 2d data. In ICRA, 2013. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 354, + 547, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 354, + 547, + 386 + ], + "spans": [ + { + "bbox": [ + 307, + 354, + 547, + 386 + ], + "type": "text", + "content": "[26] David Held, Sebastian Thrun, and Silvio Savarese. Learning to track at 100 FPS with deep regression networks. In ECCV, 2016. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 388, + 546, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 546, + 409 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 546, + 409 + ], + "type": "text", + "content": "[27] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 2020. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 411, + 546, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 411, + 546, + 443 + ], + "spans": [ + { + "bbox": [ + 307, + 411, + 546, + 443 + ], + "type": "text", + "content": "[28] Hou-Ning Hu, Yung-Hsu Yang, Tobias Fischer, Trevor Darrell, Fisher Yu, and Min Sun. Monocular quasi-dense 3d object tracking. IEEE TPAMI, 2022. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 445, + 547, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 445, + 547, + 476 + ], + "spans": [ + { + "bbox": [ + 307, + 445, + 547, + 476 + ], + "type": "text", + "content": "[29] KJ Joseph, Salman Khan, Fahad Shahbaz Khan, and Vineeth N Balasubramanian. Towards open world object detection. In CVPR, 2021. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 478, + 546, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 478, + 546, + 510 + ], + "spans": [ + { + "bbox": [ + 307, + 478, + 546, + 510 + ], + "type": "text", + "content": "[30] Anna Khoreva, Rodrigo Benenson, Eddy Ilg, Thomas Brox, and Bernt Schiele. Lucid data dreaming for video object segmentation. *IJCV*, 127(9):1175–1197, 2019. 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 512, + 546, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 512, + 546, + 544 + ], + "spans": [ + { + "bbox": [ + 307, + 512, + 546, + 544 + ], + "type": "text", + "content": "[31] Seung Wook Kim, Jonah Philion, Antonio Torralba, and Sanja Fidler. Drivegan: Towards a controllable high-quality neural simulation. In CVPR, 2021. 3" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 545, + 546, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 545, + 546, + 578 + ], + "spans": [ + { + "bbox": [ + 307, + 545, + 546, + 578 + ], + "type": "text", + "content": "[32] Laura Leal-Taixe, Cristian Canton-Ferrer, and Konrad Schindler. Learning by tracking: Siamese CNN for robust target association. In CVPRW, 2016. 2" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 307, + 579, + 546, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 579, + 546, + 611 + ], + "spans": [ + { + "bbox": [ + 307, + 579, + 546, + 611 + ], + "type": "text", + "content": "[33] Bastian Leibe, Ales Leonardis, and Bernt Schiele. Robust object detection with interleaved categorization and segmentation. IJCV, 77(1):259-289, 2008. 1" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 307, + 613, + 546, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 546, + 645 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 546, + 645 + ], + "type": "text", + "content": "[34] Siyuan Li, Martin Danelljan, Henghui Ding, Thomas E Huang, and Fisher Yu. Tracking every thing in the wild. In ECCV, 2022. 2, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 307, + 647, + 546, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 546, + 679 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 546, + 679 + ], + "type": "text", + "content": "[35] Tsung-Yi Lin, Piotr Dollar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In CVPR, 2017. 7" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 307, + 681, + 546, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 681, + 546, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 681, + 546, + 712 + ], + "type": "text", + "content": "[36] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C Berg. Ssd: Single shot multibox detector. In ECCV, 2016. 2" + } + ] + } + ], + "index": 37 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "5575" + } + ] + } + ], + "index": 39 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 115 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 115 + ], + "type": "text", + "content": "[37] Yang Liu, Idil Esen Zulfikar, Jonathon Luiten, Achal Dave, Deva Ramanan, Bastian Leibe, Aljosa Osep, and Laura Leal-Taixe. Opening up open world tracking. In CVPR, 2022. 1, 3, 4, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 288, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 288, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 288, + 162 + ], + "type": "text", + "content": "[38] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 164, + 288, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 164, + 288, + 207 + ], + "spans": [ + { + "bbox": [ + 48, + 164, + 288, + 207 + ], + "type": "text", + "content": "[39] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In CVPR, 2022. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 209, + 288, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 288, + 241 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 288, + 241 + ], + "type": "text", + "content": "[40] Jonathon Luiten, Tobias Fischer, and Bastian Leibe. Track to reconstruct and reconstruct to track. RA-L, 5(2):1803-1810, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 244, + 288, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 244, + 288, + 277 + ], + "spans": [ + { + "bbox": [ + 48, + 244, + 288, + 277 + ], + "type": "text", + "content": "[41] Tim Meinhardt, Alexander Kirillov, Laura Leal-Taixe, and Christoph Feichtenhofer. Trackformer: Multi-object tracking with transformers. In CVPR, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 278, + 288, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 278, + 288, + 312 + ], + "spans": [ + { + "bbox": [ + 48, + 278, + 288, + 312 + ], + "type": "text", + "content": "[42] Anton Milan, Seyed Hamid Rezatofighi, Anthony R. Dick, Ian D. Reid, and Konrad Schindler. Online multi-target tracking using recurrent neural networks. In AAAI, 2017. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 313, + 288, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 313, + 288, + 346 + ], + "spans": [ + { + "bbox": [ + 48, + 313, + 288, + 346 + ], + "type": "text", + "content": "[43] Anton Milan, Stefan Roth, and Konrad Schindler. Continuous energy minimization for multitarget tracking. IEEE TPAMI, 36(1):58-72, 2013. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 348, + 288, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 348, + 288, + 380 + ], + "spans": [ + { + "bbox": [ + 48, + 348, + 288, + 380 + ], + "type": "text", + "content": "[44] Dennis Mitzel and Bastian Leibe. Taking mobile multi-object tracking to the next level: People, unknown objects, and carried items. In ECCV, 2012. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 382, + 288, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 382, + 288, + 415 + ], + "spans": [ + { + "bbox": [ + 48, + 382, + 288, + 415 + ], + "type": "text", + "content": "[45] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 417, + 288, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 417, + 288, + 459 + ], + "spans": [ + { + "bbox": [ + 48, + 417, + 288, + 459 + ], + "type": "text", + "content": "[46] Aljoša Osep, Alexander Hermans, Francis Engelmann, Dirk Klostermann, Markus Mathias, and Bastian Leibe. Multiscale object candidates for generic object tracking in street scenes. In ICRA, 2016. 1, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 462, + 288, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 462, + 288, + 495 + ], + "spans": [ + { + "bbox": [ + 48, + 462, + 288, + 495 + ], + "type": "text", + "content": "[47] Aljoša Osep, Wolfgang Mehner, Markus Mathias, and Bastian Leibe. Combined image-and world-space tracking in traffic scenes. In ICRA, 2017. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 497, + 288, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 497, + 288, + 529 + ], + "spans": [ + { + "bbox": [ + 48, + 497, + 288, + 529 + ], + "type": "text", + "content": "[48] Aljoša Osep, Wolfgang Mehner, Paul Voigtlaender, and Bastian Leibe. Track, then decide: Category-agnostic vision-based multi-object tracking. In ICRA, 2018. 1, 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 532, + 288, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 532, + 288, + 563 + ], + "spans": [ + { + "bbox": [ + 48, + 532, + 288, + 563 + ], + "type": "text", + "content": "[49] Aljoša Osep, Paul Voigtlaender, Mark Weber, Jonathon Luiten, and Bastian Leibe. 4d generic video object proposals. In ICRA, 2020. 1, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 566, + 288, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 566, + 288, + 599 + ], + "spans": [ + { + "bbox": [ + 48, + 566, + 288, + 599 + ], + "type": "text", + "content": "[50] Jiangmiao Pang, Linlu Qiu, Xia Li, Haofeng Chen, Qi Li, Trevor Darrell, and Fisher Yu. Quasi-dense similarity learning for multiple object tracking. In CVPR, 2021. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 601, + 288, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 288, + 632 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 288, + 632 + ], + "type": "text", + "content": "[51] Jeffrey Pennington, Richard Socher, and Christopher D Manning. Glove: Global vectors for word representation. In EMNLP, 2014. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 635, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 288, + 689 + ], + "type": "text", + "content": "[52] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 2, 3, 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 691, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 714 + ], + "type": "text", + "content": "[53] Deva Ramanan and David A Forsyth. Finding and tracking people from the bottom up. In CVPR, 2003. 2" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 36, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 106 + ], + "type": "text", + "content": "[54] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, 2021. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 107, + 547, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 547, + 138 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 547, + 138 + ], + "type": "text", + "content": "[55] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In CVPR, 2016. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 141, + 547, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 547, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 547, + 162 + ], + "type": "text", + "content": "[56] Joseph Redmon and Ali Farhadi. Yolo9000: better, faster, stronger. In CVPR, 2017. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 164, + 547, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 164, + 547, + 196 + ], + "spans": [ + { + "bbox": [ + 307, + 164, + 547, + 196 + ], + "type": "text", + "content": "[57] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In NeurIPS, 2015. 2, 4, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 198, + 547, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 198, + 547, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 198, + 547, + 228 + ], + "type": "text", + "content": "[58] Stephan R Richter, Vibhav Vineet, Stefan Roth, and Vladlen Koltun. Playing for data: Ground truth from computer games. In ECCV, 2016. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 232, + 547, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 232, + 547, + 273 + ], + "spans": [ + { + "bbox": [ + 307, + 232, + 547, + 273 + ], + "type": "text", + "content": "[59] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 3, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 276, + 547, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 276, + 547, + 308 + ], + "spans": [ + { + "bbox": [ + 307, + 276, + 547, + 308 + ], + "type": "text", + "content": "[60] Amir Sadeghian, Alexandre Alahi, and Silvio Savarese. Tracking the untrackable: Learning to track multiple cues with long-term dependencies. In ICCV, 2017. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 310, + 547, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 310, + 547, + 342 + ], + "spans": [ + { + "bbox": [ + 307, + 310, + 547, + 342 + ], + "type": "text", + "content": "[61] Samuel Schulter, Paul Vernaza, Wongun Choi, and Manmohan Chandraker. Deep network flow for multi-object tracking. In CVPR, 2017. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 343, + 547, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 343, + 547, + 386 + ], + "spans": [ + { + "bbox": [ + 307, + 343, + 547, + 386 + ], + "type": "text", + "content": "[62] Sarthak Sharma, Junaid Ahmed Ansari, J. Krishna Murthy, and K. Madhava Krishna. Beyond pixels: Leveraging geometry and shape cues for online multi-object tracking. In ICRA, 2018. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 388, + 547, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 547, + 430 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 547, + 430 + ], + "type": "text", + "content": "[63] Peize Sun, Jinkun Cao, Yi Jiang, Rufeng Zhang, Enze Xie, Zehuan Yuan, Changhu Wang, and Ping Luo. Transtrack: Multiple object tracking with transformer. arXiv preprint arXiv:2012.15460, 2020. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 433, + 547, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 433, + 547, + 487 + ], + "spans": [ + { + "bbox": [ + 307, + 433, + 547, + 487 + ], + "type": "text", + "content": "[64] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In CVPR, 2020. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 489, + 547, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 489, + 547, + 520 + ], + "spans": [ + { + "bbox": [ + 307, + 489, + 547, + 520 + ], + "type": "text", + "content": "[65] Nicolai Wojke, Alex Bewley, and Dietrich Paulus. Simple online and realtime tracking with a deep association metric. In ICIP, 2017. 2, 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 522, + 547, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 522, + 547, + 566 + ], + "spans": [ + { + "bbox": [ + 307, + 522, + 547, + 566 + ], + "type": "text", + "content": "[66] Sanghyun Woo, Kwanyong Park, Seoung Wug Oh, In So Kweon, and Joon-Young Lee. Bridging images and videos: A simple learning framework for large vocabulary video object detection. In ECCV, 2022. 3, 7" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 567, + 547, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 567, + 547, + 599 + ], + "spans": [ + { + "bbox": [ + 307, + 567, + 547, + 599 + ], + "type": "text", + "content": "[67] Sanghyun Woo, Kwanyong Park, Seoung Wug Oh, In So Kweon, and Joon-Young Lee. Tracking by associating clips. In ECCV, 2022. 7" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 601, + 547, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 601, + 547, + 623 + ], + "spans": [ + { + "bbox": [ + 307, + 601, + 547, + 623 + ], + "type": "text", + "content": "[68] Bin Xiao, Haiping Wu, and Yichen Wei. Simple baselines for human pose estimation and tracking. In ECCV, 2018. 2" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 624, + 547, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 547, + 645 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 547, + 645 + ], + "type": "text", + "content": "[69] Bo Yang and Ram Nevatia. An online learned CRF model for multi-target tracking. In CVPR, 2012. 2" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 307, + 647, + 547, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 547, + 668 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 547, + 668 + ], + "type": "text", + "content": "[70] Linjie Yang, Yuchen Fan, and Ning Xu. Video instance segmentation. In ICCV, 2019. 2, 4, 6" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 307, + 670, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 547, + 713 + ], + "type": "text", + "content": "[71] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darryll. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In CVPR, 2020. 1" + } + ] + } + ], + "index": 35 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "5576" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 330 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 105 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 105 + ], + "type": "text", + "content": "[72] Alireza Zareian, Kevin Dela Rosa, Derek Hao Hu, and Shih-Fu Chang. Open-vocabulary object detection using captions. In CVPR, 2021. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 107, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 107, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 49, + 107, + 287, + 140 + ], + "type": "text", + "content": "[73] Fangao Zeng, Bin Dong, Tiancai Wang, Xiangyu Zhang, and Yichen Wei. Motr: End-to-end multiple-object tracking with transformer. arXiv preprint arXiv:2105.03247, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 141, + 288, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 141, + 288, + 184 + ], + "spans": [ + { + "bbox": [ + 49, + 141, + 288, + 184 + ], + "type": "text", + "content": "[74] Yifu Zhang, Chunyu Wang, Xinggang Wang, Wenjun Zeng, and Wenyu Liu. Fairmot: On the fairness of detection and re-identification in multiple object tracking. IJCV, 129(11):3069-3087, 2021. 3, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 186, + 288, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 186, + 288, + 229 + ], + "spans": [ + { + "bbox": [ + 49, + 186, + 288, + 229 + ], + "type": "text", + "content": "[75] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Liunian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, et al. Regionclip: Region-based language-image pretraining. In CVPR, 2022. 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 231, + 288, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 231, + 288, + 273 + ], + "spans": [ + { + "bbox": [ + 49, + 231, + 288, + 273 + ], + "type": "text", + "content": "[76] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Phillip Krahenbuhl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision. In ECCV, 2022. 2, 3, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 275, + 288, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 275, + 288, + 297 + ], + "spans": [ + { + "bbox": [ + 49, + 275, + 288, + 297 + ], + "type": "text", + "content": "[77] Xingyi Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Tracking objects as points. In ECCV, 2020. 3, 4, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 299, + 288, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 299, + 288, + 330 + ], + "spans": [ + { + "bbox": [ + 49, + 299, + 288, + 330 + ], + "type": "text", + "content": "[78] Xingyi Zhou, Tianwei Yin, Vladlen Koltun, and Philipp Krahenbihl. Global tracking transformers. In CVPR, 2022. 2, 7, 8" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "5577" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Objaverse_ A Universe of Annotated 3D Objects/ea3a0563-acc5-4afc-aad6-362a9401875b_content_list.json b/2023/Objaverse_ A Universe of Annotated 3D Objects/ea3a0563-acc5-4afc-aad6-362a9401875b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..069795769335ef5979d71f8b1e5a040553a1c4bd --- /dev/null +++ b/2023/Objaverse_ A Universe of Annotated 3D Objects/ea3a0563-acc5-4afc-aad6-362a9401875b_content_list.json @@ -0,0 +1,1528 @@ +[ + { + "type": "text", + "text": "Objaverse: A Universe of Annotated 3D Objects", + "text_level": 1, + "bbox": [ + 241, + 130, + 728, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Matt Deitke $^{\\dagger \\psi}$ , Dustin Schwenk $^{\\dagger}$ , Jordi Salvador $^{\\dagger}$ , Luca Weihs $^{\\dagger}$ , Oscar Michel $^{\\dagger}$ , Eli VanderBilt $^{\\dagger}$ , Ludwig Schmidt $^{\\psi}$ , Kiana Ehsani $^{\\dagger}$ , Aniruddha Kembhavi $^{\\dagger \\psi}$ , Ali Farhadi $^{\\psi}$ , PRIOR @ Allen Institute for AI, $^{\\psi}$ University of Washington, Seattle objaverse.allenai.org", + "bbox": [ + 109, + 169, + 856, + 244 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d27174bf5aa9b8c6b03aef22ab2204e1fece6cc396a02880cc669df3e5ab6385.jpg", + "image_caption": [ + "Figure 1. Example instances from our large-scale 3D asset dataset OBJVERSE. OBJVERSE 3D assets are semantically diverse, high-quality, and paired with natural-language descriptions." + ], + "image_footnote": [], + "bbox": [ + 81, + 255, + 890, + 542 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 603, + 313, + 619 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Massive data corpora like WebText, Wikipedia, Conceptual Captions, WebImageText, and LAION have propelled recent dramatic progress in AI. Large neural models trained on such datasets produce impressive results and top many of today's benchmarks. A notable omission1ion within this family of large-scale datasets is 3D data. Despite considerable interest and potential applications in 3D vision, datasets of high-fidelity 3D models continue to be mid-sized with limited diversity of object categories. Addressing this gap, we present Objaverse 1.0, a large dataset of objects with $800K+$ (and growing) 3D models with descriptive captions, tags, and animations. Objaverse improves upon present day 3D repositories in terms of scale, number of categories, and in the visual diversity of instances within a category. We demonstrate the large potential of Objaverse via four diverse applications: training generative 3D models, im", + "bbox": [ + 75, + 626, + 473, + 868 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "proving tail category segmentation on the LVIS benchmark, training open-vocabulary object-navigation models for Embodied AI, and creating a new benchmark for robustness analysis of vision models. Objaverse can open new directions for research and enable new applications across the field of AI.", + "bbox": [ + 500, + 599, + 890, + 688 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 501, + 694, + 632, + 709 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Massive datasets have enabled and driven rapid progress in AI. Language corpora on the web led to large language models like GPT-3 [4]; paired image and text datasets like Conceptual Captions [63] led to vision-and-language pretrained models like VilBERT [42]; YouTube video datasets led to video capable models like Merlot-Reserve [82]; and massive multimodal datasets like WebImageText [65] and LAION [61, 62] led to models like CLIP [55] and StableDiffusion [59]. These leaps in dataset scale and diversity were triggered by moving from manually curated datasets to harnessing the power of the web and its creative content.", + "bbox": [ + 496, + 719, + 890, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In contrast to the datasets described above, the size of", + "bbox": [ + 517, + 885, + 892, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Correspondence to .", + "bbox": [ + 99, + 887, + 328, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "13142", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "the datasets we are feeding to our data-hungry deep learning models in many other areas of research is simply not comparable. For instance, the number of 3D assets used in training generative 3D models is, maximally, on the order of thousands [22] and the simulators used to train embodied AI models typically have only between a few dozen to a thousand unique scenes [36, 39, 58, 67]. The startling advances brought about by developing large-scale datasets for images, videos, and natural language, demand that an equivalent dataset be built for 3D assets.", + "bbox": [ + 75, + 90, + 472, + 241 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We present OBJAYERSE 1.0, a large scale corpus of high-quality, richly annotated, 3D objects; see Fig. 1. Objects in our dataset are free to use1 and sourced from Sketchfab, a leading online platform for managing, viewing, and distributing 3D models. In total, OBJAYERSE contains over 800K 3D assets designed by over 150K artists which makes this data large and diversely sourced. Assets not only belong to varied categories like animals, humans, and vehicles, but also include interiors and exteriors of large spaces that can be used, e.g., to train embodied agents. OBJAYERSE is a universe of rich 3D data with detailed metadata that can support many different annotations to enable new applications. With this remarkable increase in scale, we see an incredible opportunity for OBJAYERSE to impact research progress across domains. In this work, we provide promising results to answer three questions.", + "bbox": [ + 75, + 242, + 472, + 484 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Can 3D vision benefit from a large-scale dataset? First, as a 3D asset resource, OBJVERSE can support the exciting field of 3D generative modeling. We use data extracted from OBJVERSE to train generative models for single and multiple categories using GET3D [22] and find that we are able to generate high-quality objects. Moreover, we find that our generated objects are found by human annotators to be more diverse than those generated by a model trained on ShapeNet objects in $91\\%$ of cases.", + "bbox": [ + 75, + 484, + 472, + 619 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Can the diversity of 3D models help improve classical 2D vision task performance? To answer this question, we use the diversity of OBJVERSE to improve the performance of long tail instance segmentation models. Instance segmentation data can be expensive to obtain owing to the cost of annotating contours around objects. The recent LVIS dataset contains segmentation annotations for 1,230 categories but the task remains very challenging for present day models, particularly on tail categories that have few examples. We show that increasing the volume of data by leveraging a simple Copy+Paste augmentation method with OBJVERSE assets can improve the performance of state-of-the-art segmentation methods.", + "bbox": [ + 75, + 619, + 472, + 816 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We also use OBJVERSE to build a benchmark for evaluating the robustness of state-of-the-art visual classification models to perspective shifts. We render objects in OBJVERSE from random orientations, which is how one", + "bbox": [ + 76, + 816, + 470, + 878 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "might expect to see them in the real world, and test the ability of CLIP-style visual backbones to correctly classify these images. Our experiments show that current state-of-the-art models' performance degrades dramatically in this setting when viewing objects from arbitrary views. OBJVERSE allows us to build benchmarks to test (and potentially train) for orientation robustness for a long tail distribution of asset categories. Building such benchmarks is made uniquely possible by the scale and diversity of 3D assets in OBJVERSE. This would simply not be feasible to create in the real world nor can they be generated from existing 2D images.", + "bbox": [ + 500, + 90, + 893, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Can a large-scale 3D dataset help us train perform- \nmant embodied agents? We use assets in ObjAVERSE to \npopulate procedurally generated simulated environments in \nProcTHOR [16] that are used to train Embodied AI agents. \nThis results in an orders of magnitude increase in the num \nber of unique assets available for use in ProcTHOR scenes \n(Previously limited to AI2-THOR's [36] asset library of a \nfew thousand unique instances each assigned to one of 108 \nobject categories). Using ObjAVERSE populated scenes en \nables open vocabulary object navigation from any text de \nscription. In this paper, we provide quantitative results for \nnavigating to 1.1K semantic object categories, roughly a \n50x increase.", + "bbox": [ + 496, + 273, + 893, + 468 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "These findings represent just a small fraction of what can be accomplished using OBJVERSE. We are excited to see how the research community will leverage OBJVERSE to enable fast and exciting progress in 2D and 3D computer vision applications and beyond.", + "bbox": [ + 496, + 469, + 893, + 546 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 556, + 640, + 571 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Large scale datasets. Scaling the size and scope of training datasets has widely been demonstrated to be an effective avenue of improvement for model performance. In computer vision, the adoption of early large scale datasets such as Imagenet [17,60] and MS-COCO [41] has dramatically accelerated progress on a variety of tasks including classification, object detection, captioning, and more. Ever since, the diversity and scale of datasets have continued to grow. YFCC100M is a dataset of 99.2M images and 800K videos [72]. OpenImages [37] is a large scale dataset of 9M images that contains labeled subsets bounding boxes, visual relationships, segmentation masks, localized narratives, and categorical annotations. Massive web-scraped datasets containing image-text pairs such as Conceptual Captions [63], WIT [65], and LAION [61,62] have seen increased popularity recently as they have been used to train impressive models for vision-language representation learning [27,30,55], text-to-image generation [30,56,57,59], and vision-language multitasking [8,9,68,74].", + "bbox": [ + 496, + 582, + 893, + 869 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D datasets. Current large-scale 2D image datasets offer three crucial components that benefit learning: scale,", + "bbox": [ + 500, + 869, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "$^{1}$ Creative Commons license", + "bbox": [ + 94, + 886, + 246, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "13143", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/5c950e32123e8db1b6a14ea9fd110fc402b44b9c0b32f85745854c1ce63842dc.jpg", + "image_caption": [ + "Objverse" + ], + "image_footnote": [], + "bbox": [ + 78, + 113, + 367, + 276 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b75316214db6200a6ee1919761e93615b32977476914291829d738de2c7ab97d.jpg", + "image_caption": [ + "ShapeNet", + "Figure 2. Comparison between OBJVERSE and existing 3D object datasets. (Left:) Visual comparison of instances from OBJVERSE and ShapeNet for the categories of CAR, BED, VASE, and BAG. OBJVERSE instances are substantially more diverse since objects can come from many 3D content creation platforms, whereas ShapeNet models look more similar and all come from SketchUp, a 3D modeling platform built for simple architectural modeling. (Right:) Scale comparison table between existing 3D object datasets." + ], + "image_footnote": [], + "bbox": [ + 383, + 113, + 663, + 276 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/df511b3d9e9e2a89178ecd629235c88f23d63f6c0045d6b4d3ecbd7b2f946b99.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Dataset# Objects# Classes
YCB [5]775
BigBIRD [64]125-
KIT [33]219145
IKEA [40]21911
Pix3D [66]3959
GSO [18]1K17
EGAD [48]2K-
PhotoShape [50]5K1
ABO [11]8K63
3D-Future [20]10K34
ShapeNet [6]51K55
Objaverse818K21K
", + "bbox": [ + 684, + 104, + 888, + 286 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "diversity, and realism. Ideally, models that reason about 3D objects should have access to datasets that meet these same criteria. However, of the numerous 3D object datasets that currently exist, none are able to excel in all three categories to the same degree as their 2D counterparts. Datasets such as KIT [33], YCB [5], BigBIRD [64], IKEA [40], and Pix3D [66] provide image-calibrated models over a diverse set of household objects, but severely lack in scale with only a few hundred objects at most. EGAD [48] procedurally generates 2K objects for grasping, but produces objects that are not that realistic or diverse. Slightly larger datasets of photo-realistic objects include GSO [18], PhotoShape [50], ABO [11] and 3D-Future [20], and ShapeNet [6] with object counts in the tens of thousands, see Fig. 2 for comparisons between OBJVERSE and these datasets. Datasets for CAD models, such as ModelNet [78] and DeepCAD [77], and ABC [35] do not include textures or materials, which limits their ability to represent objects that could plausibly be found in the real world. Datasets of scanned 3D objects and environments are valuable for real-world understanding [10, 13, 14, 38], but are quite small and limited. In addition to containing numerous artist designed objects, OBJVERSE contains many scanned assets, making it a useful source of data for learning from real-world distributions.", + "bbox": [ + 75, + 367, + 472, + 731 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While rapid progress has been made in developing datasets that combine image and text, in contrast, only a few datasets that pair language and 3D data exist. Text2Shape [7] released a dataset of 15,038 chairs and tables from ShapeNet each with around 5 text captions, giving 75,344 total text-shape pairs. ShapeGlot [1] released the CiC (Chairs in Context) dataset which contains 4,511 chairs from ShapeNet along with 78,789 descriptive utterances generated from a referential game. Due to the small scale and limited diversity of these datasets, current SoTA text-to-3D models [29,44,53] forgo the use of 3D datasets entirely", + "bbox": [ + 75, + 734, + 468, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and instead rely on 2D image-text supervision.", + "bbox": [ + 498, + 366, + 808, + 382 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Objverse", + "text_level": 1, + "bbox": [ + 500, + 388, + 609, + 407 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "OBJAVERSE is a massive annotated 3D dataset that can be used to enable research in a wide range of areas across computer vision. The objects are sourced from Sketchfab, an online 3D marketplace where users can upload and share models for both free and commercial use. Objects selected for OBJVERSE have a distributable Creative Commons license and were obtained using Sketchfab's public API. Aside from licensing consideration, models marked as restricted due to objectionable or adult thematic content were excluded from the dataset.", + "bbox": [ + 496, + 415, + 892, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Model metadata. OBJVERSE objects inherit a set of foundational annotations supplied by their creator when uploaded to Sketchfab. Figure 4 shows an example of the metadata available for each model. The metadata includes a name, assignments to a set of fixed categories, a set of unrestricted tags, and a natural language description.", + "bbox": [ + 496, + 566, + 893, + 657 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "OBJAVERSE-LVIS. While OBJAVERSE metadata contains a great deal of information about objects, Sketchfab's existing categorization scheme covers only 18 categories, too coarse for most applications. Object names, categories, and tags provide multiple potential categorizations at varying levels of specificity and with some inherent noise. However, for many existing computer vision tasks, it is useful to assign objects to a single category drawn from a predetermined set of the right size and level of semantic granularity.", + "bbox": [ + 496, + 657, + 893, + 794 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We choose the categories from the LVIS dataset [24] for categorizing a long-tail subset of objects in OBJVERSE. We construct a 47K LVIS categorized object subset, called OBJVERSE-LVIS, comprised of objects uniquely assigned to one of 1156 LVIS categories. We perform these assignments by first selecting 500 candidate objects per category using a combination of predictions from a CLIP classifi-", + "bbox": [ + 496, + 795, + 893, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "13144", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/efa256944f5ac1bfacdc0fd3b9c2eec90f9b3b63f6cc5f8aa15a004e42d6e850.jpg", + "image_caption": [ + "(a) Breakdown of high-level OBJVERSE categories" + ], + "image_footnote": [], + "bbox": [ + 76, + 90, + 472, + 241 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/130ef3e48f4de85c2ef3d74081848d75aba43bbce8e716255dfc9f15a22e3bbe.jpg", + "image_caption": [ + "(b) Word cloud of OBJVERSE metadata tags." + ], + "image_footnote": [], + "bbox": [ + 498, + 87, + 890, + 241 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6cd2455fa6ddc613c8418b6c2bb78ba3dbe2f5546204618ce46ab2d428931a2d.jpg", + "image_caption": [ + "(c) OBJAYERSE-LVIS category occurrence distribution." + ], + "image_footnote": [], + "bbox": [ + 76, + 268, + 472, + 444 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c4778b2d2da25378220938528fd057d369eb877a05541899a2ad194e445ce874.jpg", + "image_caption": [ + "(d) OBJAYERSE tag occurrence distribution.", + "Figure 3. ObjAverse statistics. (a) All 18 high-level categories present in ObjAverse's metadata with their corresponding number of occurrences. The relative share of most popular categories are evenly split, with a small number of less frequently categories. (b) A sample of several thousand popular object tags found in ObjAverse log-scaled by their frequency. (c) A histogram of fine-grained ObjAverse-LVIS categories with representative members from several bins highlighted. (d) A histogram of ObjAverse tags with representative members from several bins highlighted (note y-axis log scale). Tags from the low-occurrence side of the distribution correspond to unique objects that, taken individually, are rarely seen in the world. Frequently used tags like \"furniture\" and \"car\" reflect their real-world normalcy, but the high frequency of assets like \"sword\" diverge from their real-world counterparts." + ], + "image_footnote": [], + "bbox": [ + 486, + 268, + 890, + 445 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "cation model and candidates suggested by terms in their metadata. This combined pool contains objects visually resembling the target category (from the CLIP features of their thumbnail images) that might have missing metadata, as well as visually unusual instances of a category that are accurately named or tagged. These 250k candidate objects were then manually filtered and their assigned categories verified by crowdworkers. Since we only presented 500 object candidates per class, many popular categories, such as chair or car, have substantially more objects that could be included in OBJAYERSE-LVIS with future annotations.", + "bbox": [ + 75, + 582, + 468, + 747 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Animated objects and rigged characters. OBJVERSE includes 44K animated objects and over 63K objects self-categorized as characters. Examples of animations include fridge doors opening, animals running, and the hands on a clock moving. Rigged characters can be set up for animation and rendering, and may often come annotated with bone mappings. The vast scale of animations available in OBJVERSE can support a wide range of research in temporal 3D learning, such as building text-based animation generative models [71], representing object changes over time", + "bbox": [ + 75, + 750, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "with NERFs [51, 54], and temporal self-supervised learning via. future frame prediction [28, 82].", + "bbox": [ + 496, + 582, + 890, + 612 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Articulated objects. Decomposing 3D objects into parts has led to a flurry of research in the past few years, including work in learning robotic grasping policies [79, 81], 3D semantic segmentation [47], and shape generation [46]. Since many objects in OBJAYERSE were uploaded by artists, the objects often come separated into parts. Figure 5 shows an example, where a chair is separated by its backrest, wheels, and legs, among many smaller parts.", + "bbox": [ + 496, + 613, + 890, + 733 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Exteriors. Photogrammetry and NERF advances have enabled the commercialization of capturing high-quality 3D objects of large exteriors by taking pictures [70, 80]. In OBJVERSE, there are a large number of scanned buildings, cities, and stadiums. Figure 5 shows an example of a 3D object of NYC's skyline captured through a scan.", + "bbox": [ + 496, + 734, + 890, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "OBJAVERSE-Interiors. There are $16\\mathrm{K}+$ interior scenes in OBJAVERSE, including houses, classrooms, and offices. The scenes often have multiple floors, many types of rooms, and are densely populated with objects from human input. Objects in the scenes are separable into parts, which allows", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "13145", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2ae6e510eb49b547d316e5c34919efdc85c3ce5acd8eb4f3806f824b58cb61e8.jpg", + "image_caption": [ + "Figure 4. An example of metadata available for each object in OBJVERSE. Each uploaded object has a 3D model, user-selected rendered thumbnail image, name, description, tags, category, and stats, among additional metadata." + ], + "image_footnote": [], + "bbox": [ + 83, + 89, + 890, + 210 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "them to be usable for interactive robotics, embodied AI, and scene synthesis. To put the scale of OBJVERSE-Interiors in perspective, the number of scenes in OBJVERSE-Interiors is significantly larger than the 400 or so existing hand-built interactive embodied AI scenes [21,36,39,67].", + "bbox": [ + 75, + 261, + 470, + 338 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Visual styles. Objects in the world can be constructed in many styles and often differ in style based on the time-period, geographic location, and artist's style. OBJVERSE objects cover a vast set of visual styles, including 3D scans, 3D modeled objects from virtually any platform, point clouds, and photo-realism via physically based rendering (PBR) [52]. Moreover, instances of objects often appear with many styles, which is critical for training and evaluating robust computer vision models [55]. Figure 5 shows examples of chairs in OBJVERSE in many different styles, including Gothic, modern, Victorian, cartoon, and abstract.", + "bbox": [ + 75, + 343, + 468, + 508 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Statistics. OBJVERSE 1.0 includes 818K 3D objects, designed by 160K artists. There are $>2.35\\mathrm{M}$ tags on the objects, with $>170\\mathrm{K}$ of them being unique. We estimate that the objects have coverage for nearly 21K WordNet entities [45] (see appendix for details). Objects were uploaded between 2012 and 2022, with over 200K objects uploaded uploaded just in 2021. Figure 3 visualizes several statistics of the dataset, including the breakdown of objects into their", + "bbox": [ + 75, + 515, + 470, + 636 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "self-assigned Sketchfab categories, a word cloud over the tags, a frequency plot of the tags, and the number of objects in OBJAYERSE-LVIS categories.", + "bbox": [ + 498, + 261, + 890, + 308 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Applications", + "text_level": 1, + "bbox": [ + 500, + 321, + 630, + 339 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we present 4 initial distinct applications of OBJVERSE, including 3D generative modeling, instance segmentation with CP3D, open-vocabulary ObjectNav, and analyze robustness in computer vision models.", + "bbox": [ + 498, + 345, + 890, + 407 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. 3D Generative Modeling", + "text_level": 1, + "bbox": [ + 500, + 416, + 725, + 431 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3D generative modeling has shown much improvement recently with models such as GET3D [22] delivering impressive high quality results with rich geometric details. GET3D is trained to generate 3D textured meshes for a category and produces impressive 3D objects for categories like Car, Chair, and Motorcycle using data from ShapeNet [6]. OBJVERSE contains 3D models for many diverse categories including tail categories which are not represented in other datasets. It also contains diverse and realistic object instances per category. This scale and diversity can be used to train large vocabulary and high quality 3D generative models. In this work, we showcase the potential of this data as follows. We choose three categories of ob-", + "bbox": [ + 496, + 439, + 892, + 636 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/44c8b5c5e5a1e4701007798a436663cb732d5c01ad604d78470a7168ad1b4dcc.jpg", + "image_caption": [ + "Figure 5. Highlights of the visual diversity of objects that appear in OBJVERSE, including animated objects, rigged (body-part annotated) characters, models separatable into parts, exterior environments, interior environments, and a wide range visual styles." + ], + "image_footnote": [], + "bbox": [ + 78, + 656, + 893, + 869 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "13146", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1897e9edb06e5ebeb723e88e3d50a6f5e9c5c4a01c2229ebd026eaed4048aa37.jpg", + "image_caption": [ + "Objaverse Bag Generations" + ], + "image_footnote": [], + "bbox": [ + 86, + 104, + 452, + 183 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/76725f113584d350d32777b8cd96b334d519f4be3129402e97db4ef290d081d0.jpg", + "image_caption": [ + "ShapeNet Bag Generations" + ], + "image_footnote": [], + "bbox": [ + 96, + 210, + 449, + 284 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7721da45e4abe5aad684934b546f37b24ac8a6160b7947af563dabef504b58ed.jpg", + "image_caption": [ + "(a) Comparison of Bags generated with OBJVERSE and ShapeNet.", + "Shoe Generations" + ], + "image_footnote": [], + "bbox": [ + 96, + 343, + 449, + 376 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e55973c2cfbc54edc9b1cb6eb6fb9683ffbd125d85de9f822d47a873b599e22e.jpg", + "image_caption": [ + "Fruit&Veg Generations", + "(b) Shoe and Fruit&Veg. generations." + ], + "image_footnote": [], + "bbox": [ + 94, + 411, + 449, + 448 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4206ebb7e776c70eb499cdbf2192ca5ffcf3fc6d15acdee0b7ed329ba07ff600.jpg", + "image_caption": [ + "Fruit&Veg Interpolation", + "(c) Fruit&Veg. interpolation.", + "Figure 6. (a) Example GET3D Bag object generations using OBJVERSE and ShapeNet models for training. (b) Additional Shoe and Fruit&Veg generations from OBJVERSE models. (c) models generated when interpolating between two, randomly sampled, latent encodings with our trained Fruit&Veg. model; what appears to be a pumpkin smoothly transforms into a mushroom." + ], + "image_footnote": [], + "bbox": [ + 91, + 503, + 449, + 532 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "jects, Shoe, Bag, and Fruit&Veg, and subsample objects from OBJVERSE to create three separate datasets containing, respectively, 143 shoes, 816 bags, and 571 fruits & vegetables (116 apples, 112 gourds, 92 mushrooms, 68 bananas, 52 oranges, 52 pears, 31 potatoes 24 lemons, and 24 pineapples). For comparison, we also train a GET3D model on the set of 83 bags from the ShapeNet dataset. Fig. 6 shows a collection of 3D objects generated by our trained GET3D models. Qualitatively, the 3D-meshes generated by the OBJVERSE-trained models are high-quality and diverse, especially when compared to the generations from the ShapeNet-trained model. To quantify this observation, we asked crowdworkers to rate the diversity of Bag generations produced by the OBJVERSE and ShapeNet trained models. When shown collections of nine randomly sampled", + "bbox": [ + 75, + 674, + 472, + 902 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2d568ef167cdfef8cbd5757de0bf5a3138b984dd8bd4e39ada1779261896e044.jpg", + "image_caption": [ + "Figure 7. An illustration of CP3D (copy-paste 3D) for segmentation augmentation. We render 3D objects from multiple views and paste them over LVIS training images." + ], + "image_footnote": [], + "bbox": [ + 509, + 79, + 887, + 266 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "generations from both models, workers rated the collection generated from the OBJVERSE trained model as more diverse in appearance $91\\%$ of the time.", + "bbox": [ + 498, + 330, + 890, + 375 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our fruits and vegetables, composed of 9 varieties produces perhaps the highest quality output, a promising signal that can inspire future work in text-to-3D generation.", + "bbox": [ + 498, + 376, + 890, + 422 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Instance Segmentation with CP3D", + "text_level": 1, + "bbox": [ + 500, + 431, + 799, + 449 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "A key advantage of using simulated data for computer vision is that it is much cheaper to obtain expert annotations. Annotated OBJVERSE objects can be rendered into images, allowing them to serve as a rich source of additional data that can be used to enhance model performance on 2D computer vision tasks. As a proof-of-concept demonstrating the effectiveness of this approach, we use segmented data from OBJVERSE objects as auxiliary labels for training models on the LVIS dataset for Large Scale Instance Segmentation [24]. The LVIS dataset contains instance segmentation masks for 1200 object categories that occur throughout a set of $164\\mathrm{k}$ images. Recognition is especially challenging in this task due to the long tail of the object category distribution in this dataset. LVIS categories only con", + "bbox": [ + 496, + 455, + 890, + 669 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2db9e376a2d7d27686fb04d8283faae22010d6d4e8ff38cbb71263fc052b24e2.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodAPAPrAPcAPf
RFS [24]23.713.323.029.0
EQLv2 [69]25.517.724.330.2
LOCE [19]26.618.526.230.7
NorCal with RFS [49]25.219.324.229.0
Seesaw [73]26.419.526.129.7
GOL [3]27.721.427.730.4
GOL + CP3D28.321.828.331.1
", + "bbox": [ + 516, + 691, + 877, + 828 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Comparison of our approach (GOL+CP3D) against SoTA Mask-RCNN ResNet-50 models on LVIS. We report results for APr, APc, and APf which measure AP for categories that are rare (appear in 1-10 images), common (appear in 11-100 images), and frequent (appear in $>100$ images), respectively", + "bbox": [ + 498, + 839, + 890, + 910 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "13147", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4fa6782d8fa095cdf09a093835f3cede63e70411b507d47c3530b45458562905.jpg", + "image_caption": [ + "Figure 8. An existing ProcTHOR scene (left) and a semantically similar ProcTHOR generatable scene with OBJVERSE objects (right)." + ], + "image_footnote": [], + "bbox": [ + 156, + 88, + 478, + 244 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/096eec98402182a571df44cf631a5a93dbb7f7defbf4a80fb36a5536d8c1582b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 483, + 88, + 810, + 244 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tain an average 9 instances across the dataset, so training on simulated data is a promising approach for overcoming the challenges of learning in this low-sample regime.", + "bbox": [ + 75, + 266, + 468, + 311 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Using the LVIS-annotated subset of ObjAVERSE, we introduce CP3D: an enhancement to the simple, but effective, copy-and-paste technique of [23]. Figure 7 shows an example of the setup for CP3D. Here, we render different views of 3D objects and paste them on-top of existing LVIS images. We render 5 distinct views of each object and cache them for use throughout training. During training, an image is selected for the copy-paste augmentation with 0.5 probability, and once selected, 1-3 images of randomly chosen LVIS-annotated ObjAVERSE objects are pasted onto the selected training image. The segmentation masks of the selected objects are added to the training image's annotation as well. Object images and masks are randomly scaled and translated before being pasted. We use this strategy to finetune the pretrained ResNet-50 Mask-RCNN [25, 26] of [3]. As shown in Tab.1, simply finetuning this model for 24 epochs yields performance gains across several metrics.", + "bbox": [ + 75, + 311, + 472, + 570 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Open-Vocabulary ObjectNav", + "text_level": 1, + "bbox": [ + 76, + 577, + 336, + 593 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we introduce open-vocabulary ObjectNav, a new task propelled by the vast diversity of objects that appear in OBJAYERSE. Here, an agent is placed at a random starting location inside of a home and tasked to navigate to a target object provided from a text description (e.g. \"Raspberry Pi Pico\"). To facilitate this task, we procedurally generate 10K new homes in ProcTHOR [16] fully populated with objects from OBJAYERSE-LVIS. Until now, ObjectNav tasks have focused on training agents to navigate to 20 or so target objects provided their category label [15, 16, 58], and existing interactive embodied AI simulations, including ProcTHOR, only include around 2K total objects across around 100 object types [16, 39, 67]. In this work, we take a large step to massively scale the number of target objects used in ObjectNav ( $20 \\to \\text{OpenVocab}$ ), the number of objects available in simulation ( $2\\mathbf{K} \\to 36\\mathbf{K}$ ), and the number of object types of the objects ( $100 \\to 1.1\\mathbf{K}$ ).", + "bbox": [ + 75, + 597, + 468, + 853 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Object placement. To make the placement of objects in the houses more natural, we use the OBJVERSE-LVIS subset and annotate placement constraints for each object", + "bbox": [ + 75, + 854, + 470, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "category. Specifically, we annotate if objects of a given category typically appears on the floor, on-top of a surface, or on a wall. If instances of the object category may appear on the floor, we also annotate whether it may appear in the middle of the scene (e.g. a clutter object like a basketball) or on the edge of the scene (e.g. a toilet or a fridge). For objects placed on the floor, we also to automatically detect flat regions on top of the object's mesh to place surface object types. The annotations are used by ProcTHOR for sampling objects to place in a scene. We also filter out OBJAYERSE-LVIS objects that do not appear inside of homes, such as a jet plane. Structural objects, like doors and windows, are inherited from ProcTHOR as they would require additional cleanup.", + "bbox": [ + 496, + 266, + 890, + 477 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Object size correction. Objects in Sketchfab may be uploaded at unnatural scales (e.g. a plant being as large as a tower). We therefore scale the objects to be of a reasonable size for them to look natural in a house. Here, for each object category, we annotate the maximum bounding box dimension length that every instance of the object category should be scaled to. For example, we annotate the maximum bounding box dimension for bookcase to be 2 meters and fork to be 0.18 meters. If a 3D modeled bookcase then has a bounding box of $20\\mathrm{m}\\times 6\\mathrm{m}\\times 3\\mathrm{m}$ , we shrink each side by a factor of $\\max (20,6,3) / 2 = 5$ .", + "bbox": [ + 496, + 478, + 892, + 643 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Preprocessing for AI2-THOR. We add support to AI2-THOR for loading objects on the fly at runtime. Previously, all objects had to be stored in a Unity build, but such an approach is impractical when working with orders of magnitude more object data. For each object, we compress it with Blender [12] by joining all of its meshes together, decimate the joined mesh such that it has at most 5K vertices, and bake all the UV texture maps into a single texture map. We then generate colliders using V-HACD [43] to support rigid-body interactions.", + "bbox": [ + 496, + 643, + 890, + 794 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Approach. Given procedural houses populated with OBJVERSE-LVIS, the task is to navigate to the proximity of a chosen target object and invoke a task-completion action when the target object is in sight, given an open-vocabulary description formed with the template \"a {name} {category}\" The name is the object name given by its creator, which is often descriptive. We filter", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "13148", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ce79a2e0bd4d79922e4ca0ff9aa7070a8f2cc0e617560e82d4ca0acf156c29d9.jpg", + "image_caption": [ + "Television 99.6%" + ], + "image_footnote": [], + "bbox": [ + 80, + 88, + 171, + 160 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/82cf596b8e34decf4cfac9cc4bdf62e0bf7756ca0da27d19d6853a58c9194d2a.jpg", + "image_caption": [ + "Television 96.1%" + ], + "image_footnote": [], + "bbox": [ + 179, + 88, + 271, + 160 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bc6c4c147c226e984ff05e3fc2dcf17cc7a5cffc713823bd2ce8508a2a253844.jpg", + "image_caption": [ + "Vase 39.0%" + ], + "image_footnote": [], + "bbox": [ + 277, + 88, + 370, + 160 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3823427c5bc8dae555d3db806c11590db8b0e0a463eefddaca8f23e9ac685ba6.jpg", + "image_caption": [ + "Flowerpot 34.9%" + ], + "image_footnote": [], + "bbox": [ + 374, + 88, + 468, + 160 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0822e0addd82a3b04016ec6b6ac45f00366446491301187cbb71ea99c4e8b1f0.jpg", + "image_caption": [ + "Clock $99.8\\%$" + ], + "image_footnote": [], + "bbox": [ + 80, + 176, + 171, + 248 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/32954a33ab0101bec606b19c6b1c7c2e0aa942305dabb95bb800f2f3ecba188f.jpg", + "image_caption": [ + "Clock 98.2%" + ], + "image_footnote": [], + "bbox": [ + 179, + 176, + 271, + 248 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/24c7d257474f5c15ac3b633adeaefbfdf6ecbf4d829c9b8706e6a24236f71bbb.jpg", + "image_caption": [ + "Flowerpot 32.4%" + ], + "image_footnote": [], + "bbox": [ + 277, + 176, + 370, + 248 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7ae741bf261e216b98d703f1aaaf74e83979a66e44b237c86e4755ee6eec8595.jpg", + "image_caption": [ + "Vase 33.2%", + "Figure 9. Examples of objects rendered from random orientations and their 0-shot classification categories with the CLIP ViT-B/32." + ], + "image_footnote": [], + "bbox": [ + 374, + 176, + 468, + 248 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "each by whether it is detected as being written in English by a language detector [31,32], and fall back to a class-only description for non-English name. Examples of the possible expressions include “a victorian-monobike motorcycle”, “a unicorn pony”, or “a dino ghost lizard”. The agent, similar to the ones in [34], observes an RGB egocentric view of the environment, pre-processed by the visual branch of a frozen ResNet-50 CLIP model [55] – the target description is pre-processed by the corresponding text branch. We train the agent with DD-PPO [76] and evaluate on houses with floor plans, objects, and descriptions unseen in training. We use the AllenAct [75] framework to train our agent. Our trained agent achieves a success rate of $19.9\\%$ , for a random policy success of $5.1\\%$ . For more details about the experiment refer to the appendix.", + "bbox": [ + 75, + 316, + 468, + 542 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Analyzing Robustness", + "text_level": 1, + "bbox": [ + 76, + 549, + 282, + 565 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A persistent bias present in many image datasets, e.g. ImageNet [60], is that the subjects of interest are generally photographed from a forward-facing, canonical, orientation. When, for example, taking a photograph of a television, few would choose to take this photograph crouched on the floor behind the television. This impact of this bias was studied by Alcorn et al. [2] who find that modern computer vision systems are highly susceptible to deviations from canonical poses. This is more than a theoretical problem: computer vision systems deployed in the real world will frequently encounter objects in non-canonical orientations and in many applications, e.g. autonomous driving, it will be safety critical that they behave well.", + "bbox": [ + 75, + 568, + 468, + 763 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Given the above, we adopt the experimental design of Alcorn et al. and design, using OBJVERSE assets, a benchmark for evaluating the robustness of state-of-the-art computer vision classification models to orientation shifts. In particular, for each object in our OBJVERSE-LVIS subset, we render 12 images of the object from random orientations rendered upon a background with RGB values equalling the mean RGB values from ImageNet; see Fig. 9 for examples. This ability to, at scale, render objects from random view", + "bbox": [ + 75, + 763, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/57d8679f1191c16454ebcc916eb23bf0d71c95bd62220f1230d05fe2e031ecb1.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelRandom RotationAny RotationΔTop-1
Top-1Top-5Top-1Top-5
OpenAI-400M [55]
RN5021.4%45.0%43.9%70.8%22.5%
ViT-L/1429.1%54.5%52.3%77.2%23.2%
LAION-400M [62]
ViT-B/3224.1%48.5%46.9%74.2%22.8%
ViT-L/1430.6%56.8%50.5%77.0%19.9%
LAION-2B [61]
ViT-B/3227.0%51.8%50.3%76.1%23.3%
ViT-L/1432.9%59.2%52.1%78.0%19.2%
ViT-H/1432.3%58.8%50.1%77.3%17.8%
", + "bbox": [ + 504, + 88, + 888, + 271 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 2. Evaluating 0-shot CLIP classification models on our rotational robustness benchmark. $\\Delta$ Top-1 denotes the difference between Top-1 Any Rotation and Top-1 Random Rotation. Models are strongly overfit to standard views of objects.", + "bbox": [ + 498, + 282, + 890, + 338 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "points is a practical impossibility in the real world but is made trivial when using 3D assets. We then evaluate several modern open-domain image-classification networks (constrained to the $\\approx 1,200$ LVIS categories) on these images and report 4 metrics for each model. These metrics include:", + "bbox": [ + 496, + 353, + 890, + 428 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Top-1 Random Rotation - the frequency with which a model correctly classifies an image as belonging to the respective LVIS category.", + "- Top-1 Any Rotation - the frequency with which a model classifies an image correctly from at least one of the 12 random orientations." + ], + "bbox": [ + 500, + 430, + 890, + 518 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This second metric is diagnostic and serves to represent a model's performance when shown an object from a canonical pose. We also have Top-5 variants of the above metric where the correct category need only be in the top 5 predictions from the model. We report our results in Tab. 2 in which we evaluate a variety of performant pretrained models. Comparing the gap in performance between the Top-k Random Rotation and Top-k Any Rotation metrics we find that model performance dramatically degrades when viewing objects from unusual orientations.", + "bbox": [ + 496, + 521, + 890, + 672 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 683, + 617, + 698 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present OBJAYERSE, a next-generation 3D asset library containing 818K high-quality, diverse, 3D models with paired text descriptions, titles, and tags. As a small glimpse of the potential uses of OBJAYERSE, we present four experimental studies showing how OBJAYERSE can be used to power (1) generative 3D models with clear future applications to text-to-3D generation, (2) improvements to classical computer vision tasks such as instance segmentation, (3) the creation of novel embodied AI tasks like Open Vocabulary Object Navigation, and (4) quantifying the rotational robustness of vision models on renderings of objects. We hope to see OBJAYERSE enable a new universe of new applications for computer vision.", + "bbox": [ + 496, + 704, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "13149", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Panos Achlioptas, Judy Fan, Robert Hawkins, Noah Goodman, and Leonidas J Guibas. Shapeplot: Learning language for shape differentiation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8938-8947, 2019. 3", + "[2] Michael A. Alcorn, Qi Li, Zhitao Gong, Chengfei Wang, Long Mai, Wei-Shinn Ku, and Anh Nguyen. Strike (with) a pose: Neural networks are easily fooled by strange poses of familiar objects. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, pages 4845-4854. Computer Vision Foundation / IEEE, 2019. 8", + "[3] Konstantinos Panagiotis Alexandridis, Jiankang Deng, Anh Nguyen, and Shan Luo. Long-tailed instance segmentation using gumbel optimized loss. arXiv preprint arXiv:2207.10936, 2022. 6, 7", + "[4] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 1", + "[5] Berk Calli, Aaron Walsman, Arjun Singh, Siddhartha Srinivasa, Pieter Abbeel, and Aaron M Dollar. Benchmarking in manipulation research: The ycb object and model set and benchmarking protocols. arXiv preprint arXiv:1502.03143, 2015. 3", + "[6] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 3, 5", + "[7] Kevin Chen, Christopher B Choy, Manolis Savva, Angel X Chang, Thomas Funkhouser, and Silvio Savarese. Text2shape: Generating shapes from natural language by learning joint embeddings. In Asian conference on computer vision, pages 100-116. Springer, 2018. 3", + "[8] Xi Chen, Xiao Wang, Soravit Changpinyo, AJ Piergiovanni, Piotr Padlewski, Daniel Salz, Sebastian Goodman, Adam Grycner, Basil Mustafa, Lucas Beyer, et al. Pali: A jointly-scaled multilingual language-image model. arXiv preprint arXiv:2209.06794, 2022. 2", + "[9] Jaemin Cho, Jie Lei, Hao Tan, and Mohit Bansal. Unifying vision-and-language tasks via text generation. In International Conference on Machine Learning, pages 1931-1942. PMLR, 2021. 2", + "[10] Sungjoon Choi, Qian-Yi Zhou, Stephen Miller, and Vladlen Koltun. A large dataset of object scans. arXiv preprint arXiv:1602.02481, 2016. 3", + "[11] Jasmine Collins, Shubham Goel, Kenan Deng, Achleshwar Luthra, Leon Xu, Erhan Gundogdu, Xi Zhang, Tomas F Yago Vicente, Thomas Dideriksien, Himanshu Arora, et al. Abo: Dataset and benchmarks for real-world 3d object understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21126-21136, 2022. 3" + ], + "bbox": [ + 78, + 116, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Blender Online Community. Blender - a 3d modelling and rendering package. http://www.blender.org, 2018.7", + "[13] Camille Couprie, Clément Farabet, Laurent Najman, and Yann LeCun. Indoor semantic segmentation using depth information. arXiv preprint arXiv:1301.3572, 2013. 3", + "[14] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 3", + "[15] Matt Deitke, Dhruv Batra, Yonatan Bisk, Tommaso Campari, Angel X Chang, Devendra Singh Chaplot, Changan Chen, Claudia Pérez D'Arpino, Kiana Ehsani, Ali Farhadi, et al. Retrospectives on the embodied ai workshop. arXiv preprint arXiv:2210.06849, 2022. 7", + "[16] Matt Deitke, Eli VanderBilt, Alvaro Herrasti, Luca Weihs, Jordi Salvador, Kiana Ehsani, Winson Han, Eric Kolve, Ali Farhadi, Aniruddha Kembhavi, et al. Procthor: Large-scale embodied ai using procedural generation. Conference on Neural Information Processing Systems, 2022. 2, 7", + "[17] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 2", + "[18] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas B McHugh, and Vincent Vanhoucke. Google scanned objects: A high-quality dataset of 3d scanned household items. arXiv preprint arXiv:2204.11918, 2022. 3", + "[19] Chengjian Feng, Yujie Zhong, and Weilin Huang. Exploring classification equilibrium in long-tailed object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3417-3426, 2021. 6", + "[20] Huan Fu, Rongfei Jia, Lin Gao, Mingming Gong, Binqiang Zhao, Steve Maybank, and Dacheng Tao. 3d-future: 3d furniture shape with texture. International Journal of Computer Vision, 129(12):3313-3337, 2021. 3", + "[21] Chuang Gan, Jeremy Schwartz, Seth Alter, Martin Schrimpf, James Traer, Julian De Freitas, Jonas Kubilius, Abhishek Bhandwaldar, Nick Haber, Megumi Sano, et al. Threadworld: A platform for interactive multi-modal physical simulation. arXiv preprint arXiv:2007.04954, 2020. 5", + "[22] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. In Advances In Neural Information Processing Systems, 2022. 2, 5", + "[23] Golnaz Ghiasi, Yin Cui, Aravind Srinivas, Rui Qian, Tsung-Yi Lin, Ekin D Cubuk, Quoc V Le, and Barret Zoph. Simple copy-paste is a strong data augmentation method for instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2918-2928, 2021. 7", + "[24] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5356-5364, 2019. 3, 6" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "13150", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Kaiming He, Georgia Gkioxari, Piotr Dollár, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 7", + "[26] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 7", + "[27] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. OpenCLIP, July 2021. 2", + "[28] Allan Jabri, Andrew Owens, and Alexei Efros. Space-time correspondence as a contrastive random walk. Advances in neural information processing systems, 33:19545-19560, 2020. 4", + "[29] Ajay Jain, Ben Mildenhall, Jonathan T Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 867-876, 2022. 3", + "[30] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916. PMLR, 2021. 2", + "[31] Armand Joulin, Edouard Grave, Piotr Bojanowski, Matthijs Douze, Hérve Jégou, and Tomas Mikolov. Fasttext.zip: Compressing text classification models. arXiv preprint arXiv:1612.03651, 2016. 8", + "[32] Armand Joulin, Edouard Grave, Piotr Bojanowski, and Tomas Mikolov. Bag of tricks for efficient text classification. arXiv preprint arXiv:1607.01759, 2016. 8", + "[33] Alexander Kasper, Zhixing Xue, and Rüdiger Dillmann. The kit object models database: An object model database for object recognition, localization and manipulation in service robotics. The International Journal of Robotics Research, 31(8):927-934, 2012. 3", + "[34] Apoorv Khandelwal, Luca Weihs, Roozbeh Mottaghi, and Aniruddha Kembhavi. Simple but effective: Clip embeddings for embodied ai. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14809-14818, 2022. 8", + "[35] Sebastian Koch, Albert Matveev, Zhongshi Jiang, Francis Williams, Alexey Artemov, Evgeny Burnaev, Marc Alexa, Denis Zorin, and Daniele Panozzo. Abc: A big cad model dataset for geometric deep learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9601-9611, 2019. 3", + "[36] Eric Kolve, Roozbeh Mottaghi, Winson Han, Eli VanderBilt, Luca Weihs, Alvaro Herrasti, Matt Deitke, Kiana Ehsani, Daniel Gordon, Yuke Zhu, Aniruddha Kembhavi, Abhinav Kumar Gupta, and Ali Farhadi. Ai2-thor: An interactive 3d environment for visual ai. arXiv e-prints, pages arXiv-1712, 2017. 2, 5", + "[37] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan" + ], + "bbox": [ + 78, + 90, + 468, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4. International Journal of Computer Vision, 128(7):1956-1981, 2020. 2", + "[38] Marc Levoy, Kari Pulli, Brian Curless, Szymon Rusinkiewicz, David Koller, Lucas Pereira, Matt Ginzton, Sean Anderson, James Davis, Jeremy Ginsberg, et al. The digital michelangelo project: 3d scanning of large statues. In Proceedings of the 27th annual conference on Computer graphics and interactive techniques, pages 131-144, 2000. 3", + "[39] Chengshu Li, Fei Xia, Roberto Martín-Martín, Michael Lingelbach, Sanjana Srivastava, Bokui Shen, Kent Vainio, Cem Gokmen, Gokul Dharan, Tanish Jain, et al. igibson 2.0: Object-centric simulation for robot learning of everyday household tasks. arXiv preprint arXiv:2108.03272, 2021. 2, 5, 7", + "[40] Joseph J Lim, Hamed Pirsiavash, and Antonio Torralba. Parsing IKEA objects: Fine pose estimation. In Proceedings of the IEEE international conference on computer vision, pages 2992-2999, 2013. 3", + "[41] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 2", + "[42] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. Advances in neural information processing systems, 32, 2019. 1", + "[43] Khaled Mamou, E Lengyel, and A Peters. Volumetric hierarchical approximate convex decomposition. In Game Engine Gems 3, pages 141-158. AK Peters, 2016. 7", + "[44] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13492-13502, 2022. 3", + "[45] George A Miller. Wordnet: a lexical database for english. Communications of the ACM, 38(11):39-41, 1995. 5", + "[46] Kaichun Mo, Paul Guerrero, Li Yi, Hao Su, Peter Wonka, Niloy Mitra, and Leonidas J Guibas. Structurenet: Hierarchical graph networks for 3d shape generation. arXiv preprint arXiv:1908.00575, 2019. 4", + "[47] Kaichun Mo, Shilin Zhu, Angel X Chang, Li Yi, Subarna Tripathi, Leonidas J Guibas, and Hao Su. Partnet: A largescale benchmark for fine-grained and hierarchical part-level 3d object understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 909-918, 2019. 4", + "[48] D. Morrison, P. Corke, and J. Leitner. Egad! an evolved grasping analysis dataset for diversity and reproducibility in robotic manipulation. IEEE Robotics and Automation Letters, 5(3):4368-4375, 2020. 3", + "[49] Tai-Yu Pan, Cheng Zhang, Yandong Li, Hexiang Hu, Dong Xuan, Soravit Changpinyo, Boqing Gong, and Wei-Lun Chao. On model calibration for long-tailed object detection and instance segmentation. Advances in Neural Information Processing Systems, 34:2529-2542, 2021. 6" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "13151", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[50] Keunhong Park, Konstantinos Rematas, Ali Farhadi, and Steven M Seitz. Photoshape: Photorealistic materials for large-scale shape collections. arXiv preprint arXiv:1809.09761, 2018.3", + "[51] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865-5874, 2021. 4", + "[52] Matt Pharr, Wenzel Jakob, and Greg Humphreys. Physically based rendering: From theory to implementation. Morgan Kaufmann, 2016. 5", + "[53] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 3", + "[54] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10318–10327, 2021. 4", + "[55] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 1, 2, 5, 8", + "[56] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2", + "[57] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 2", + "[58] Ram Ramrakhya, Eric Undersander, Dhruv Batra, and Abhishek Das. Habitat-web: Learning embodied object-search strategies from human demonstrations at scale. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5173-5183, 2022. 2, 7", + "[59] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 1, 2", + "[60] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115(3):211-252, 2015. 2, 8", + "[61] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. arXiv preprint arXiv:2210.08402, 2022. 1, 2, 8", + "[62] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo" + ], + "bbox": [ + 78, + 90, + 470, + 901 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 1, 2, 8", + "[63] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2556-2565, 2018. 1, 2", + "[64] Arjun Singh, James Sha, Karthik S. Narayan, Tudor Achim, and P. Abbeel. Bigbird: A large-scale 3d database of object instances. 2014 IEEE International Conference on Robotics and Automation (ICRA), pages 509-516, 2014. 3", + "[65] Krishna Srinivasan, Karthik Raman, Jiecao Chen, Michael Bendersky, and Marc Najork. Wit: Wikipedia-based image text dataset for multimodal multilingual machine learning. In Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 2443-2449, 2021. 1, 2", + "[66] Xingyuan Sun, Jiajun Wu, Xiuming Zhang, Zhoutong Zhang, Chengkai Zhang, Tianfan Xue, Joshua B Tenenbaum, and William T Freeman. Pix3d: Dataset and methods for single-image 3d shape modeling. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2974-2983, 2018. 3", + "[67] Andrew Szot, Alexander Clegg, Eric Undersander, Erik Wijmans, Yili Zhao, John Turner, Noah Maestre, Mustafa Mukadam, Devendra Singh Chaplot, Oleksandr Maksymets, et al. Habitat 2.0: Training home assistants to rearrange their habitat. Advances in Neural Information Processing Systems, 34:251-266, 2021. 2, 5, 7", + "[68] Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. arXiv preprint arXiv:1908.07490, 2019. 2", + "[69] Jingru Tan, Xin Lu, Gang Zhang, Changqing Yin, and Quanquan Li. Equalization loss v2: A new gradient balance approach for long-tailed object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1685-1694, 2021. 6", + "[70] Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P Srinivasan, Jonathan T Barron, and Henrik Kretzschmar. Block-nerf: Scalable large scene neural view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8248–8258, 2022. 4", + "[71] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Amit H Bermano, and Daniel Cohen-Or. Human motion diffusion model. arXiv preprint arXiv:2209.14916, 2022. 4", + "[72] Bart Thomee, David A Shamma, Gerald Friedland, Benjamin Elizalde, Karl Ni, Douglas Poland, Damian Borth, and Li-Jia Li. Yfcc100m: The new data in multimedia research. Communications of the ACM, 59(2):64-73, 2016. 2", + "[73] Jiaqi Wang, Wenwei Zhang, Yuhang Zang, Yuhang Cao, Jiangmiao Pang, Tao Gong, Kai Chen, Ziwei Liu, Chen Change Loy, and Dahua Lin. Seesaw loss for long-tailed instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9695-9704, 2021. 6" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "13152", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[74] Wenhui Wang, Hangbo Bao, Li Dong, Johan Bjorck, Zhiliang Peng, Qiang Liu, Kriti Aggarwal, Owais Khan Mohammed, Saksham Singhal, Subhojit Som, et al. Image as a foreign language: Beit pretraining for all vision and vision-language tasks. arXiv preprint arXiv:2208.10442, 2022. 2", + "[75] Luca Weihs, Jordi Salvador, Klemen Kotar, Unnat Jain, Kuo-Hao Zeng, Roozbeh Mottaghi, and Aniruddha Kembhavi. Allenact: A framework for embodied ai research. arXiv preprint arXiv:2008.12760, 2020. 8", + "[76] Erik Wijmans, Abhishek Kadian, Ari S. Morcos, Stefan Lee, Irfan Essa, Devi Parikh, Manolis Savva, and Dhruv Batra. Dd-ppo: Learning near-perfect pointgoal navigators from 2.5 billion frames. In ICLR, 2020. 8", + "[77] Rundi Wu, Chang Xiao, and Changxi Zheng. Deepcad: A deep generative network for computer-aided design models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6772–6782, 2021. 3", + "[78] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguuang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1912-1920, 2015. 3", + "[79] Fanbo Xiang, Yuzhe Qin, Kaichun Mo, Yikuan Xia, Hao Zhu, Fangchen Liu, Minghua Liu, Hanxiao Jiang, Yifu Yuan, He Wang, et al. Sapien: A simulated part-based interactive environment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11097-11107, 2020. 4", + "[80] Yuanbo Xiangli, Linning Xu, Xingang Pan, Nanxuan Zhao, Anyi Rao, Christian Theobalt, Bo Dai, and Dahua Lin. Bungeenerf: Progressive neural radiance field for extreme multi-scale scene rendering. In The European Conference on Computer Vision (ECCV), 2022. 4", + "[81] Zhenjia Xu, Beichun Qi, Shubham Agrawal, and Shuran Song. Adagrasp: Learning an adaptive gripper-aware grasping policy. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 4620-4626. IEEE, 2021. 4", + "[82] Rowan Zellers, Jiasen Lu, Ximing Lu, Youngjae Yu, Yanpeng Zhao, Mohammadreza Salehi, Aditya Kusupati, Jack Hessel, Ali Farhadi, and Yejin Choi. Merlot reserve: Neural script knowledge through vision and language and sound. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16375-16387, 2022. 1, 4" + ], + "bbox": [ + 78, + 90, + 468, + 723 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "13153", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2023/Objaverse_ A Universe of Annotated 3D Objects/ea3a0563-acc5-4afc-aad6-362a9401875b_model.json b/2023/Objaverse_ A Universe of Annotated 3D Objects/ea3a0563-acc5-4afc-aad6-362a9401875b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..eacf47ee98caf0b16e12c55e906c9ddf10bd0580 --- /dev/null +++ b/2023/Objaverse_ A Universe of Annotated 3D Objects/ea3a0563-acc5-4afc-aad6-362a9401875b_model.json @@ -0,0 +1,2611 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.242, + 0.131, + 0.73, + 0.154 + ], + "angle": 0, + "content": "Objaverse: A Universe of Annotated 3D Objects" + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.17, + 0.857, + 0.245 + ], + "angle": 0, + "content": "Matt Deitke\\(^{\\dagger \\psi}\\), Dustin Schwenk\\(^{\\dagger}\\), Jordi Salvador\\(^{\\dagger}\\), Luca Weihs\\(^{\\dagger}\\), Oscar Michel\\(^{\\dagger}\\), Eli VanderBilt\\(^{\\dagger}\\), Ludwig Schmidt\\(^{\\psi}\\), Kiana Ehsani\\(^{\\dagger}\\), Aniruddha Kembhavi\\(^{\\dagger \\psi}\\), Ali Farhadi\\(^{\\psi}\\), PRIOR @ Allen Institute for AI, \\(^{\\psi}\\)University of Washington, Seattle objaverse.allenai.org" + }, + { + "type": "image", + "bbox": [ + 0.083, + 0.256, + 0.891, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.554, + 0.893, + 0.584 + ], + "angle": 0, + "content": "Figure 1. Example instances from our large-scale 3D asset dataset OBJVERSE. OBJVERSE 3D assets are semantically diverse, high-quality, and paired with natural-language descriptions." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.604, + 0.314, + 0.62 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.627, + 0.474, + 0.869 + ], + "angle": 0, + "content": "Massive data corpora like WebText, Wikipedia, Conceptual Captions, WebImageText, and LAION have propelled recent dramatic progress in AI. Large neural models trained on such datasets produce impressive results and top many of today's benchmarks. A notable omission1ion within this family of large-scale datasets is 3D data. Despite considerable interest and potential applications in 3D vision, datasets of high-fidelity 3D models continue to be mid-sized with limited diversity of object categories. Addressing this gap, we present Objaverse 1.0, a large dataset of objects with \\(800K+\\) (and growing) 3D models with descriptive captions, tags, and animations. Objaverse improves upon present day 3D repositories in terms of scale, number of categories, and in the visual diversity of instances within a category. We demonstrate the large potential of Objaverse via four diverse applications: training generative 3D models, im" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.6, + 0.892, + 0.689 + ], + "angle": 0, + "content": "proving tail category segmentation on the LVIS benchmark, training open-vocabulary object-navigation models for Embodied AI, and creating a new benchmark for robustness analysis of vision models. Objaverse can open new directions for research and enable new applications across the field of AI." + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.695, + 0.633, + 0.71 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.885 + ], + "angle": 0, + "content": "Massive datasets have enabled and driven rapid progress in AI. Language corpora on the web led to large language models like GPT-3 [4]; paired image and text datasets like Conceptual Captions [63] led to vision-and-language pretrained models like VilBERT [42]; YouTube video datasets led to video capable models like Merlot-Reserve [82]; and massive multimodal datasets like WebImageText [65] and LAION [61, 62] led to models like CLIP [55] and StableDiffusion [59]. These leaps in dataset scale and diversity were triggered by moving from manually curated datasets to harnessing the power of the web and its creative content." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.886, + 0.893, + 0.901 + ], + "angle": 0, + "content": "In contrast to the datasets described above, the size of" + }, + { + "type": "page_footnote", + "bbox": [ + 0.101, + 0.888, + 0.329, + 0.901 + ], + "angle": 0, + "content": "Correspondence to ." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "13142" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.242 + ], + "angle": 0, + "content": "the datasets we are feeding to our data-hungry deep learning models in many other areas of research is simply not comparable. For instance, the number of 3D assets used in training generative 3D models is, maximally, on the order of thousands [22] and the simulators used to train embodied AI models typically have only between a few dozen to a thousand unique scenes [36, 39, 58, 67]. The startling advances brought about by developing large-scale datasets for images, videos, and natural language, demand that an equivalent dataset be built for 3D assets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.243, + 0.473, + 0.485 + ], + "angle": 0, + "content": "We present OBJAYERSE 1.0, a large scale corpus of high-quality, richly annotated, 3D objects; see Fig. 1. Objects in our dataset are free to use1 and sourced from Sketchfab, a leading online platform for managing, viewing, and distributing 3D models. In total, OBJAYERSE contains over 800K 3D assets designed by over 150K artists which makes this data large and diversely sourced. Assets not only belong to varied categories like animals, humans, and vehicles, but also include interiors and exteriors of large spaces that can be used, e.g., to train embodied agents. OBJAYERSE is a universe of rich 3D data with detailed metadata that can support many different annotations to enable new applications. With this remarkable increase in scale, we see an incredible opportunity for OBJAYERSE to impact research progress across domains. In this work, we provide promising results to answer three questions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.485, + 0.473, + 0.621 + ], + "angle": 0, + "content": "Can 3D vision benefit from a large-scale dataset? First, as a 3D asset resource, OBJVERSE can support the exciting field of 3D generative modeling. We use data extracted from OBJVERSE to train generative models for single and multiple categories using GET3D [22] and find that we are able to generate high-quality objects. Moreover, we find that our generated objects are found by human annotators to be more diverse than those generated by a model trained on ShapeNet objects in \\(91\\%\\) of cases." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.621, + 0.473, + 0.817 + ], + "angle": 0, + "content": "Can the diversity of 3D models help improve classical 2D vision task performance? To answer this question, we use the diversity of OBJVERSE to improve the performance of long tail instance segmentation models. Instance segmentation data can be expensive to obtain owing to the cost of annotating contours around objects. The recent LVIS dataset contains segmentation annotations for 1,230 categories but the task remains very challenging for present day models, particularly on tail categories that have few examples. We show that increasing the volume of data by leveraging a simple Copy+Paste augmentation method with OBJVERSE assets can improve the performance of state-of-the-art segmentation methods." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.818, + 0.472, + 0.879 + ], + "angle": 0, + "content": "We also use OBJVERSE to build a benchmark for evaluating the robustness of state-of-the-art visual classification models to perspective shifts. We render objects in OBJVERSE from random orientations, which is how one" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.895, + 0.273 + ], + "angle": 0, + "content": "might expect to see them in the real world, and test the ability of CLIP-style visual backbones to correctly classify these images. Our experiments show that current state-of-the-art models' performance degrades dramatically in this setting when viewing objects from arbitrary views. OBJVERSE allows us to build benchmarks to test (and potentially train) for orientation robustness for a long tail distribution of asset categories. Building such benchmarks is made uniquely possible by the scale and diversity of 3D assets in OBJVERSE. This would simply not be feasible to create in the real world nor can they be generated from existing 2D images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.274, + 0.895, + 0.469 + ], + "angle": 0, + "content": "Can a large-scale 3D dataset help us train perform- \nmant embodied agents? We use assets in ObjAVERSE to \npopulate procedurally generated simulated environments in \nProcTHOR [16] that are used to train Embodied AI agents. \nThis results in an orders of magnitude increase in the num \nber of unique assets available for use in ProcTHOR scenes \n(Previously limited to AI2-THOR's [36] asset library of a \nfew thousand unique instances each assigned to one of 108 \nobject categories). Using ObjAVERSE populated scenes en \nables open vocabulary object navigation from any text de \nscription. In this paper, we provide quantitative results for \nnavigating to 1.1K semantic object categories, roughly a \n50x increase." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.47, + 0.895, + 0.547 + ], + "angle": 0, + "content": "These findings represent just a small fraction of what can be accomplished using OBJVERSE. We are excited to see how the research community will leverage OBJVERSE to enable fast and exciting progress in 2D and 3D computer vision applications and beyond." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.558, + 0.642, + 0.573 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.583, + 0.895, + 0.87 + ], + "angle": 0, + "content": "Large scale datasets. Scaling the size and scope of training datasets has widely been demonstrated to be an effective avenue of improvement for model performance. In computer vision, the adoption of early large scale datasets such as Imagenet [17,60] and MS-COCO [41] has dramatically accelerated progress on a variety of tasks including classification, object detection, captioning, and more. Ever since, the diversity and scale of datasets have continued to grow. YFCC100M is a dataset of 99.2M images and 800K videos [72]. OpenImages [37] is a large scale dataset of 9M images that contains labeled subsets bounding boxes, visual relationships, segmentation masks, localized narratives, and categorical annotations. Massive web-scraped datasets containing image-text pairs such as Conceptual Captions [63], WIT [65], and LAION [61,62] have seen increased popularity recently as they have been used to train impressive models for vision-language representation learning [27,30,55], text-to-image generation [30,56,57,59], and vision-language multitasking [8,9,68,74]." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.895, + 0.903 + ], + "angle": 0, + "content": "3D datasets. Current large-scale 2D image datasets offer three crucial components that benefit learning: scale," + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.248, + 0.9 + ], + "angle": 0, + "content": "\\(^{1}\\)Creative Commons license" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13143" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.084, + 0.09, + 0.147, + 0.104 + ], + "angle": 0, + "content": "Objverse" + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.114, + 0.368, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.387, + 0.091, + 0.447, + 0.104 + ], + "angle": 0, + "content": "ShapeNet" + }, + { + "type": "image", + "bbox": [ + 0.384, + 0.114, + 0.664, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "table", + "bbox": [ + 0.686, + 0.106, + 0.89, + 0.287 + ], + "angle": 0, + "content": "
Dataset# Objects# Classes
YCB [5]775
BigBIRD [64]125-
KIT [33]219145
IKEA [40]21911
Pix3D [66]3959
GSO [18]1K17
EGAD [48]2K-
PhotoShape [50]5K1
ABO [11]8K63
3D-Future [20]10K34
ShapeNet [6]51K55
Objaverse818K21K
" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.298, + 0.893, + 0.355 + ], + "angle": 0, + "content": "Figure 2. Comparison between OBJVERSE and existing 3D object datasets. (Left:) Visual comparison of instances from OBJVERSE and ShapeNet for the categories of CAR, BED, VASE, and BAG. OBJVERSE instances are substantially more diverse since objects can come from many 3D content creation platforms, whereas ShapeNet models look more similar and all come from SketchUp, a 3D modeling platform built for simple architectural modeling. (Right:) Scale comparison table between existing 3D object datasets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.368, + 0.473, + 0.732 + ], + "angle": 0, + "content": "diversity, and realism. Ideally, models that reason about 3D objects should have access to datasets that meet these same criteria. However, of the numerous 3D object datasets that currently exist, none are able to excel in all three categories to the same degree as their 2D counterparts. Datasets such as KIT [33], YCB [5], BigBIRD [64], IKEA [40], and Pix3D [66] provide image-calibrated models over a diverse set of household objects, but severely lack in scale with only a few hundred objects at most. EGAD [48] procedurally generates 2K objects for grasping, but produces objects that are not that realistic or diverse. Slightly larger datasets of photo-realistic objects include GSO [18], PhotoShape [50], ABO [11] and 3D-Future [20], and ShapeNet [6] with object counts in the tens of thousands, see Fig. 2 for comparisons between OBJVERSE and these datasets. Datasets for CAD models, such as ModelNet [78] and DeepCAD [77], and ABC [35] do not include textures or materials, which limits their ability to represent objects that could plausibly be found in the real world. Datasets of scanned 3D objects and environments are valuable for real-world understanding [10, 13, 14, 38], but are quite small and limited. In addition to containing numerous artist designed objects, OBJVERSE contains many scanned assets, making it a useful source of data for learning from real-world distributions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.47, + 0.903 + ], + "angle": 0, + "content": "While rapid progress has been made in developing datasets that combine image and text, in contrast, only a few datasets that pair language and 3D data exist. Text2Shape [7] released a dataset of 15,038 chairs and tables from ShapeNet each with around 5 text captions, giving 75,344 total text-shape pairs. ShapeGlot [1] released the CiC (Chairs in Context) dataset which contains 4,511 chairs from ShapeNet along with 78,789 descriptive utterances generated from a referential game. Due to the small scale and limited diversity of these datasets, current SoTA text-to-3D models [29,44,53] forgo the use of 3D datasets entirely" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.367, + 0.81, + 0.383 + ], + "angle": 0, + "content": "and instead rely on 2D image-text supervision." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.39, + 0.611, + 0.408 + ], + "angle": 0, + "content": "3. Objverse" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.416, + 0.893, + 0.567 + ], + "angle": 0, + "content": "OBJAVERSE is a massive annotated 3D dataset that can be used to enable research in a wide range of areas across computer vision. The objects are sourced from Sketchfab, an online 3D marketplace where users can upload and share models for both free and commercial use. Objects selected for OBJVERSE have a distributable Creative Commons license and were obtained using Sketchfab's public API. Aside from licensing consideration, models marked as restricted due to objectionable or adult thematic content were excluded from the dataset." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.568, + 0.894, + 0.659 + ], + "angle": 0, + "content": "Model metadata. OBJVERSE objects inherit a set of foundational annotations supplied by their creator when uploaded to Sketchfab. Figure 4 shows an example of the metadata available for each model. The metadata includes a name, assignments to a set of fixed categories, a set of unrestricted tags, and a natural language description." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.894, + 0.795 + ], + "angle": 0, + "content": "OBJAVERSE-LVIS. While OBJAVERSE metadata contains a great deal of information about objects, Sketchfab's existing categorization scheme covers only 18 categories, too coarse for most applications. Object names, categories, and tags provide multiple potential categorizations at varying levels of specificity and with some inherent noise. However, for many existing computer vision tasks, it is useful to assign objects to a single category drawn from a predetermined set of the right size and level of semantic granularity." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.894, + 0.901 + ], + "angle": 0, + "content": "We choose the categories from the LVIS dataset [24] for categorizing a long-tail subset of objects in OBJVERSE. We construct a 47K LVIS categorized object subset, called OBJVERSE-LVIS, comprised of objects uniquely assigned to one of 1156 LVIS categories. We perform these assignments by first selecting 500 candidate objects per category using a combination of predictions from a CLIP classifi-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13144" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.092, + 0.473, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.138, + 0.244, + 0.413, + 0.256 + ], + "angle": 0, + "content": "(a) Breakdown of high-level OBJVERSE categories" + }, + { + "type": "image", + "bbox": [ + 0.499, + 0.088, + 0.891, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.575, + 0.244, + 0.816, + 0.256 + ], + "angle": 0, + "content": "(b) Word cloud of OBJVERSE metadata tags." + }, + { + "type": "image", + "bbox": [ + 0.078, + 0.269, + 0.473, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.128, + 0.447, + 0.422, + 0.46 + ], + "angle": 0, + "content": "(c) OBJAYERSE-LVIS category occurrence distribution." + }, + { + "type": "image", + "bbox": [ + 0.487, + 0.269, + 0.891, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.574, + 0.447, + 0.805, + 0.46 + ], + "angle": 0, + "content": "(d) OBJAYERSE tag occurrence distribution." + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.471, + 0.893, + 0.569 + ], + "angle": 0, + "content": "Figure 3. ObjAverse statistics. (a) All 18 high-level categories present in ObjAverse's metadata with their corresponding number of occurrences. The relative share of most popular categories are evenly split, with a small number of less frequently categories. (b) A sample of several thousand popular object tags found in ObjAverse log-scaled by their frequency. (c) A histogram of fine-grained ObjAverse-LVIS categories with representative members from several bins highlighted. (d) A histogram of ObjAverse tags with representative members from several bins highlighted (note y-axis log scale). Tags from the low-occurrence side of the distribution correspond to unique objects that, taken individually, are rarely seen in the world. Frequently used tags like \"furniture\" and \"car\" reflect their real-world normalcy, but the high frequency of assets like \"sword\" diverge from their real-world counterparts." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.583, + 0.47, + 0.748 + ], + "angle": 0, + "content": "cation model and candidates suggested by terms in their metadata. This combined pool contains objects visually resembling the target category (from the CLIP features of their thumbnail images) that might have missing metadata, as well as visually unusual instances of a category that are accurately named or tagged. These 250k candidate objects were then manually filtered and their assigned categories verified by crowdworkers. Since we only presented 500 object candidates per class, many popular categories, such as chair or car, have substantially more objects that could be included in OBJAYERSE-LVIS with future annotations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Animated objects and rigged characters. OBJVERSE includes 44K animated objects and over 63K objects self-categorized as characters. Examples of animations include fridge doors opening, animals running, and the hands on a clock moving. Rigged characters can be set up for animation and rendering, and may often come annotated with bone mappings. The vast scale of animations available in OBJVERSE can support a wide range of research in temporal 3D learning, such as building text-based animation generative models [71], representing object changes over time" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.583, + 0.891, + 0.613 + ], + "angle": 0, + "content": "with NERFs [51, 54], and temporal self-supervised learning via. future frame prediction [28, 82]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.892, + 0.734 + ], + "angle": 0, + "content": "Articulated objects. Decomposing 3D objects into parts has led to a flurry of research in the past few years, including work in learning robotic grasping policies [79, 81], 3D semantic segmentation [47], and shape generation [46]. Since many objects in OBJAYERSE were uploaded by artists, the objects often come separated into parts. Figure 5 shows an example, where a chair is separated by its backrest, wheels, and legs, among many smaller parts." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.825 + ], + "angle": 0, + "content": "Exteriors. Photogrammetry and NERF advances have enabled the commercialization of capturing high-quality 3D objects of large exteriors by taking pictures [70, 80]. In OBJVERSE, there are a large number of scanned buildings, cities, and stadiums. Figure 5 shows an example of a 3D object of NYC's skyline captured through a scan." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "OBJAVERSE-Interiors. There are \\(16\\mathrm{K}+\\) interior scenes in OBJAVERSE, including houses, classrooms, and offices. The scenes often have multiple floors, many types of rooms, and are densely populated with objects from human input. Objects in the scenes are separable into parts, which allows" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13145" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.09, + 0.891, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.22, + 0.895, + 0.25 + ], + "angle": 0, + "content": "Figure 4. An example of metadata available for each object in OBJVERSE. Each uploaded object has a 3D model, user-selected rendered thumbnail image, name, description, tags, category, and stats, among additional metadata." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.262, + 0.471, + 0.339 + ], + "angle": 0, + "content": "them to be usable for interactive robotics, embodied AI, and scene synthesis. To put the scale of OBJVERSE-Interiors in perspective, the number of scenes in OBJVERSE-Interiors is significantly larger than the 400 or so existing hand-built interactive embodied AI scenes [21,36,39,67]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.344, + 0.47, + 0.51 + ], + "angle": 0, + "content": "Visual styles. Objects in the world can be constructed in many styles and often differ in style based on the time-period, geographic location, and artist's style. OBJVERSE objects cover a vast set of visual styles, including 3D scans, 3D modeled objects from virtually any platform, point clouds, and photo-realism via physically based rendering (PBR) [52]. Moreover, instances of objects often appear with many styles, which is critical for training and evaluating robust computer vision models [55]. Figure 5 shows examples of chairs in OBJVERSE in many different styles, including Gothic, modern, Victorian, cartoon, and abstract." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.516, + 0.471, + 0.637 + ], + "angle": 0, + "content": "Statistics. OBJVERSE 1.0 includes 818K 3D objects, designed by 160K artists. There are \\(>2.35\\mathrm{M}\\) tags on the objects, with \\(>170\\mathrm{K}\\) of them being unique. We estimate that the objects have coverage for nearly 21K WordNet entities [45] (see appendix for details). Objects were uploaded between 2012 and 2022, with over 200K objects uploaded uploaded just in 2021. Figure 3 visualizes several statistics of the dataset, including the breakdown of objects into their" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.262, + 0.892, + 0.309 + ], + "angle": 0, + "content": "self-assigned Sketchfab categories, a word cloud over the tags, a frequency plot of the tags, and the number of objects in OBJAYERSE-LVIS categories." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.322, + 0.631, + 0.34 + ], + "angle": 0, + "content": "4. Applications" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.347, + 0.892, + 0.408 + ], + "angle": 0, + "content": "In this section, we present 4 initial distinct applications of OBJVERSE, including 3D generative modeling, instance segmentation with CP3D, open-vocabulary ObjectNav, and analyze robustness in computer vision models." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.417, + 0.726, + 0.433 + ], + "angle": 0, + "content": "4.1. 3D Generative Modeling" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.44, + 0.893, + 0.637 + ], + "angle": 0, + "content": "3D generative modeling has shown much improvement recently with models such as GET3D [22] delivering impressive high quality results with rich geometric details. GET3D is trained to generate 3D textured meshes for a category and produces impressive 3D objects for categories like Car, Chair, and Motorcycle using data from ShapeNet [6]. OBJVERSE contains 3D models for many diverse categories including tail categories which are not represented in other datasets. It also contains diverse and realistic object instances per category. This scale and diversity can be used to train large vocabulary and high quality 3D generative models. In this work, we showcase the potential of this data as follows. We choose three categories of ob-" + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.657, + 0.895, + 0.871 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.881, + 0.892, + 0.91 + ], + "angle": 0, + "content": "Figure 5. Highlights of the visual diversity of objects that appear in OBJVERSE, including animated objects, rigged (body-part annotated) characters, models separatable into parts, exterior environments, interior environments, and a wide range visual styles." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13146" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.092, + 0.086, + 0.235, + 0.098 + ], + "angle": 0, + "content": "Objaverse Bag Generations" + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.105, + 0.454, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.092, + 0.193, + 0.234, + 0.205 + ], + "angle": 0, + "content": "ShapeNet Bag Generations" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.212, + 0.45, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.096, + 0.297, + 0.451, + 0.31 + ], + "angle": 0, + "content": "(a) Comparison of Bags generated with OBJVERSE and ShapeNet." + }, + { + "type": "image_caption", + "bbox": [ + 0.093, + 0.323, + 0.187, + 0.334 + ], + "angle": 0, + "content": "Shoe Generations" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.344, + 0.45, + 0.377 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.093, + 0.394, + 0.211, + 0.406 + ], + "angle": 0, + "content": "Fruit&Veg Generations" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.412, + 0.45, + 0.449 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.176, + 0.459, + 0.372, + 0.471 + ], + "angle": 0, + "content": "(b) Shoe and Fruit&Veg. generations." + }, + { + "type": "image_caption", + "bbox": [ + 0.093, + 0.484, + 0.212, + 0.496 + ], + "angle": 0, + "content": "Fruit&Veg Interpolation" + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.505, + 0.45, + 0.533 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.198, + 0.55, + 0.349, + 0.561 + ], + "angle": 0, + "content": "(c) Fruit&Veg. interpolation." + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.574, + 0.47, + 0.658 + ], + "angle": 0, + "content": "Figure 6. (a) Example GET3D Bag object generations using OBJVERSE and ShapeNet models for training. (b) Additional Shoe and Fruit&Veg generations from OBJVERSE models. (c) models generated when interpolating between two, randomly sampled, latent encodings with our trained Fruit&Veg. model; what appears to be a pumpkin smoothly transforms into a mushroom." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.473, + 0.903 + ], + "angle": 0, + "content": "jects, Shoe, Bag, and Fruit&Veg, and subsample objects from OBJVERSE to create three separate datasets containing, respectively, 143 shoes, 816 bags, and 571 fruits & vegetables (116 apples, 112 gourds, 92 mushrooms, 68 bananas, 52 oranges, 52 pears, 31 potatoes 24 lemons, and 24 pineapples). For comparison, we also train a GET3D model on the set of 83 bags from the ShapeNet dataset. Fig. 6 shows a collection of 3D objects generated by our trained GET3D models. Qualitatively, the 3D-meshes generated by the OBJVERSE-trained models are high-quality and diverse, especially when compared to the generations from the ShapeNet-trained model. To quantify this observation, we asked crowdworkers to rate the diversity of Bag generations produced by the OBJVERSE and ShapeNet trained models. When shown collections of nine randomly sampled" + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.08, + 0.888, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.275, + 0.892, + 0.317 + ], + "angle": 0, + "content": "Figure 7. An illustration of CP3D (copy-paste 3D) for segmentation augmentation. We render 3D objects from multiple views and paste them over LVIS training images." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.332, + 0.892, + 0.376 + ], + "angle": 0, + "content": "generations from both models, workers rated the collection generated from the OBJVERSE trained model as more diverse in appearance \\(91\\%\\) of the time." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.377, + 0.892, + 0.423 + ], + "angle": 0, + "content": "Our fruits and vegetables, composed of 9 varieties produces perhaps the highest quality output, a promising signal that can inspire future work in text-to-3D generation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.433, + 0.8, + 0.45 + ], + "angle": 0, + "content": "4.2. Instance Segmentation with CP3D" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.457, + 0.892, + 0.67 + ], + "angle": 0, + "content": "A key advantage of using simulated data for computer vision is that it is much cheaper to obtain expert annotations. Annotated OBJVERSE objects can be rendered into images, allowing them to serve as a rich source of additional data that can be used to enhance model performance on 2D computer vision tasks. As a proof-of-concept demonstrating the effectiveness of this approach, we use segmented data from OBJVERSE objects as auxiliary labels for training models on the LVIS dataset for Large Scale Instance Segmentation [24]. The LVIS dataset contains instance segmentation masks for 1200 object categories that occur throughout a set of \\(164\\mathrm{k}\\) images. Recognition is especially challenging in this task due to the long tail of the object category distribution in this dataset. LVIS categories only con" + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.692, + 0.878, + 0.829 + ], + "angle": 0, + "content": "
MethodAPAPrAPcAPf
RFS [24]23.713.323.029.0
EQLv2 [69]25.517.724.330.2
LOCE [19]26.618.526.230.7
NorCal with RFS [49]25.219.324.229.0
Seesaw [73]26.419.526.129.7
GOL [3]27.721.427.730.4
GOL + CP3D28.321.828.331.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.84, + 0.892, + 0.911 + ], + "angle": 0, + "content": "Table 1. Comparison of our approach (GOL+CP3D) against SoTA Mask-RCNN ResNet-50 models on LVIS. We report results for APr, APc, and APf which measure AP for categories that are rare (appear in 1-10 images), common (appear in 11-100 images), and frequent (appear in \\(>100\\) images), respectively" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13147" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.157, + 0.089, + 0.48, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.484, + 0.089, + 0.811, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.082, + 0.248, + 0.885, + 0.263 + ], + "angle": 0, + "content": "Figure 8. An existing ProcTHOR scene (left) and a semantically similar ProcTHOR generatable scene with OBJVERSE objects (right)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.267, + 0.47, + 0.312 + ], + "angle": 0, + "content": "tain an average 9 instances across the dataset, so training on simulated data is a promising approach for overcoming the challenges of learning in this low-sample regime." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.313, + 0.473, + 0.571 + ], + "angle": 0, + "content": "Using the LVIS-annotated subset of ObjAVERSE, we introduce CP3D: an enhancement to the simple, but effective, copy-and-paste technique of [23]. Figure 7 shows an example of the setup for CP3D. Here, we render different views of 3D objects and paste them on-top of existing LVIS images. We render 5 distinct views of each object and cache them for use throughout training. During training, an image is selected for the copy-paste augmentation with 0.5 probability, and once selected, 1-3 images of randomly chosen LVIS-annotated ObjAVERSE objects are pasted onto the selected training image. The segmentation masks of the selected objects are added to the training image's annotation as well. Object images and masks are randomly scaled and translated before being pasted. We use this strategy to finetune the pretrained ResNet-50 Mask-RCNN [25, 26] of [3]. As shown in Tab.1, simply finetuning this model for 24 epochs yields performance gains across several metrics." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.578, + 0.337, + 0.594 + ], + "angle": 0, + "content": "4.3. Open-Vocabulary ObjectNav" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.598, + 0.47, + 0.854 + ], + "angle": 0, + "content": "In this section, we introduce open-vocabulary ObjectNav, a new task propelled by the vast diversity of objects that appear in OBJAYERSE. Here, an agent is placed at a random starting location inside of a home and tasked to navigate to a target object provided from a text description (e.g. \"Raspberry Pi Pico\"). To facilitate this task, we procedurally generate 10K new homes in ProcTHOR [16] fully populated with objects from OBJAYERSE-LVIS. Until now, ObjectNav tasks have focused on training agents to navigate to 20 or so target objects provided their category label [15, 16, 58], and existing interactive embodied AI simulations, including ProcTHOR, only include around 2K total objects across around 100 object types [16, 39, 67]. In this work, we take a large step to massively scale the number of target objects used in ObjectNav (\\(20 \\to \\text{OpenVocab}\\)), the number of objects available in simulation (\\(2\\mathbf{K} \\to 36\\mathbf{K}\\)), and the number of object types of the objects (\\(100 \\to 1.1\\mathbf{K}\\))." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Object placement. To make the placement of objects in the houses more natural, we use the OBJVERSE-LVIS subset and annotate placement constraints for each object" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.267, + 0.892, + 0.478 + ], + "angle": 0, + "content": "category. Specifically, we annotate if objects of a given category typically appears on the floor, on-top of a surface, or on a wall. If instances of the object category may appear on the floor, we also annotate whether it may appear in the middle of the scene (e.g. a clutter object like a basketball) or on the edge of the scene (e.g. a toilet or a fridge). For objects placed on the floor, we also to automatically detect flat regions on top of the object's mesh to place surface object types. The annotations are used by ProcTHOR for sampling objects to place in a scene. We also filter out OBJAYERSE-LVIS objects that do not appear inside of homes, such as a jet plane. Structural objects, like doors and windows, are inherited from ProcTHOR as they would require additional cleanup." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.479, + 0.893, + 0.644 + ], + "angle": 0, + "content": "Object size correction. Objects in Sketchfab may be uploaded at unnatural scales (e.g. a plant being as large as a tower). We therefore scale the objects to be of a reasonable size for them to look natural in a house. Here, for each object category, we annotate the maximum bounding box dimension length that every instance of the object category should be scaled to. For example, we annotate the maximum bounding box dimension for bookcase to be 2 meters and fork to be 0.18 meters. If a 3D modeled bookcase then has a bounding box of \\(20\\mathrm{m}\\times 6\\mathrm{m}\\times 3\\mathrm{m}\\), we shrink each side by a factor of \\(\\max (20,6,3) / 2 = 5\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.892, + 0.795 + ], + "angle": 0, + "content": "Preprocessing for AI2-THOR. We add support to AI2-THOR for loading objects on the fly at runtime. Previously, all objects had to be stored in a Unity build, but such an approach is impractical when working with orders of magnitude more object data. For each object, we compress it with Blender [12] by joining all of its meshes together, decimate the joined mesh such that it has at most 5K vertices, and bake all the UV texture maps into a single texture map. We then generate colliders using V-HACD [43] to support rigid-body interactions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Approach. Given procedural houses populated with OBJVERSE-LVIS, the task is to navigate to the proximity of a chosen target object and invoke a task-completion action when the target object is in sight, given an open-vocabulary description formed with the template \"a {name} {category}\" The name is the object name given by its creator, which is often descriptive. We filter" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13148" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.172, + 0.161 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.084, + 0.163, + 0.168, + 0.171 + ], + "angle": 0, + "content": "Television 99.6%" + }, + { + "type": "image", + "bbox": [ + 0.18, + 0.089, + 0.272, + 0.161 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.184, + 0.163, + 0.267, + 0.171 + ], + "angle": 0, + "content": "Television 96.1%" + }, + { + "type": "image", + "bbox": [ + 0.278, + 0.089, + 0.371, + 0.161 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.291, + 0.163, + 0.354, + 0.171 + ], + "angle": 0, + "content": "Vase 39.0%" + }, + { + "type": "image", + "bbox": [ + 0.375, + 0.089, + 0.47, + 0.161 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.382, + 0.163, + 0.465, + 0.171 + ], + "angle": 0, + "content": "Flowerpot 34.9%" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.177, + 0.172, + 0.249 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.094, + 0.251, + 0.159, + 0.26 + ], + "angle": 0, + "content": "Clock \\(99.8\\%\\)" + }, + { + "type": "image", + "bbox": [ + 0.18, + 0.177, + 0.272, + 0.249 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.193, + 0.252, + 0.258, + 0.259 + ], + "angle": 0, + "content": "Clock 98.2%" + }, + { + "type": "image", + "bbox": [ + 0.278, + 0.177, + 0.371, + 0.249 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.282, + 0.251, + 0.366, + 0.259 + ], + "angle": 0, + "content": "Flowerpot 32.4%" + }, + { + "type": "image", + "bbox": [ + 0.375, + 0.177, + 0.47, + 0.249 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.392, + 0.251, + 0.454, + 0.259 + ], + "angle": 0, + "content": "Vase 33.2%" + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.274, + 0.468, + 0.302 + ], + "angle": 0, + "content": "Figure 9. Examples of objects rendered from random orientations and their 0-shot classification categories with the CLIP ViT-B/32." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.317, + 0.47, + 0.543 + ], + "angle": 0, + "content": "each by whether it is detected as being written in English by a language detector [31,32], and fall back to a class-only description for non-English name. Examples of the possible expressions include “a victorian-monobike motorcycle”, “a unicorn pony”, or “a dino ghost lizard”. The agent, similar to the ones in [34], observes an RGB egocentric view of the environment, pre-processed by the visual branch of a frozen ResNet-50 CLIP model [55] – the target description is pre-processed by the corresponding text branch. We train the agent with DD-PPO [76] and evaluate on houses with floor plans, objects, and descriptions unseen in training. We use the AllenAct [75] framework to train our agent. Our trained agent achieves a success rate of \\(19.9\\%\\), for a random policy success of \\(5.1\\%\\). For more details about the experiment refer to the appendix." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.55, + 0.283, + 0.566 + ], + "angle": 0, + "content": "4.4. Analyzing Robustness" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.569, + 0.469, + 0.764 + ], + "angle": 0, + "content": "A persistent bias present in many image datasets, e.g. ImageNet [60], is that the subjects of interest are generally photographed from a forward-facing, canonical, orientation. When, for example, taking a photograph of a television, few would choose to take this photograph crouched on the floor behind the television. This impact of this bias was studied by Alcorn et al. [2] who find that modern computer vision systems are highly susceptible to deviations from canonical poses. This is more than a theoretical problem: computer vision systems deployed in the real world will frequently encounter objects in non-canonical orientations and in many applications, e.g. autonomous driving, it will be safety critical that they behave well." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Given the above, we adopt the experimental design of Alcorn et al. and design, using OBJVERSE assets, a benchmark for evaluating the robustness of state-of-the-art computer vision classification models to orientation shifts. In particular, for each object in our OBJVERSE-LVIS subset, we render 12 images of the object from random orientations rendered upon a background with RGB values equalling the mean RGB values from ImageNet; see Fig. 9 for examples. This ability to, at scale, render objects from random view" + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.089, + 0.89, + 0.272 + ], + "angle": 0, + "content": "
ModelRandom RotationAny RotationΔTop-1
Top-1Top-5Top-1Top-5
OpenAI-400M [55]
RN5021.4%45.0%43.9%70.8%22.5%
ViT-L/1429.1%54.5%52.3%77.2%23.2%
LAION-400M [62]
ViT-B/3224.1%48.5%46.9%74.2%22.8%
ViT-L/1430.6%56.8%50.5%77.0%19.9%
LAION-2B [61]
ViT-B/3227.0%51.8%50.3%76.1%23.3%
ViT-L/1432.9%59.2%52.1%78.0%19.2%
ViT-H/1432.3%58.8%50.1%77.3%17.8%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.283, + 0.892, + 0.339 + ], + "angle": 0, + "content": "Table 2. Evaluating 0-shot CLIP classification models on our rotational robustness benchmark. \\(\\Delta\\)Top-1 denotes the difference between Top-1 Any Rotation and Top-1 Random Rotation. Models are strongly overfit to standard views of objects." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.354, + 0.892, + 0.429 + ], + "angle": 0, + "content": "points is a practical impossibility in the real world but is made trivial when using 3D assets. We then evaluate several modern open-domain image-classification networks (constrained to the \\(\\approx 1,200\\) LVIS categories) on these images and report 4 metrics for each model. These metrics include:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.431, + 0.892, + 0.476 + ], + "angle": 0, + "content": "- Top-1 Random Rotation - the frequency with which a model correctly classifies an image as belonging to the respective LVIS category." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.476, + 0.892, + 0.52 + ], + "angle": 0, + "content": "- Top-1 Any Rotation - the frequency with which a model classifies an image correctly from at least one of the 12 random orientations." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.431, + 0.892, + 0.52 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.522, + 0.892, + 0.673 + ], + "angle": 0, + "content": "This second metric is diagnostic and serves to represent a model's performance when shown an object from a canonical pose. We also have Top-5 variants of the above metric where the correct category need only be in the top 5 predictions from the model. We report our results in Tab. 2 in which we evaluate a variety of performant pretrained models. Comparing the gap in performance between the Top-k Random Rotation and Top-k Any Rotation metrics we find that model performance dramatically degrades when viewing objects from unusual orientations." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.684, + 0.619, + 0.699 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We present OBJAYERSE, a next-generation 3D asset library containing 818K high-quality, diverse, 3D models with paired text descriptions, titles, and tags. As a small glimpse of the potential uses of OBJAYERSE, we present four experimental studies showing how OBJAYERSE can be used to power (1) generative 3D models with clear future applications to text-to-3D generation, (2) improvements to classical computer vision tasks such as instance segmentation, (3) the creation of novel embodied AI tasks like Open Vocabulary Object Navigation, and (4) quantifying the rotational robustness of vision models on renderings of objects. We hope to see OBJAYERSE enable a new universe of new applications for computer vision." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13149" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.117, + 0.47, + 0.185 + ], + "angle": 0, + "content": "[1] Panos Achlioptas, Judy Fan, Robert Hawkins, Noah Goodman, and Leonidas J Guibas. Shapeplot: Learning language for shape differentiation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8938-8947, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.188, + 0.471, + 0.284 + ], + "angle": 0, + "content": "[2] Michael A. Alcorn, Qi Li, Zhitao Gong, Chengfei Wang, Long Mai, Wei-Shinn Ku, and Anh Nguyen. Strike (with) a pose: Neural networks are easily fooled by strange poses of familiar objects. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, pages 4845-4854. Computer Vision Foundation / IEEE, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.287, + 0.471, + 0.342 + ], + "angle": 0, + "content": "[3] Konstantinos Panagiotis Alexandridis, Jiankang Deng, Anh Nguyen, and Shan Luo. Long-tailed instance segmentation using gumbel optimized loss. arXiv preprint arXiv:2207.10936, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.345, + 0.47, + 0.413 + ], + "angle": 0, + "content": "[4] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.416, + 0.47, + 0.484 + ], + "angle": 0, + "content": "[5] Berk Calli, Aaron Walsman, Arjun Singh, Siddhartha Srinivasa, Pieter Abbeel, and Aaron M Dollar. Benchmarking in manipulation research: The ycb object and model set and benchmarking protocols. arXiv preprint arXiv:1502.03143, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.487, + 0.47, + 0.557 + ], + "angle": 0, + "content": "[6] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.559, + 0.47, + 0.629 + ], + "angle": 0, + "content": "[7] Kevin Chen, Christopher B Choy, Manolis Savva, Angel X Chang, Thomas Funkhouser, and Silvio Savarese. Text2shape: Generating shapes from natural language by learning joint embeddings. In Asian conference on computer vision, pages 100-116. Springer, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.631, + 0.47, + 0.699 + ], + "angle": 0, + "content": "[8] Xi Chen, Xiao Wang, Soravit Changpinyo, AJ Piergiovanni, Piotr Padlewski, Daniel Salz, Sebastian Goodman, Adam Grycner, Basil Mustafa, Lucas Beyer, et al. Pali: A jointly-scaled multilingual language-image model. arXiv preprint arXiv:2209.06794, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.702, + 0.47, + 0.757 + ], + "angle": 0, + "content": "[9] Jaemin Cho, Jie Lei, Hao Tan, and Mohit Bansal. Unifying vision-and-language tasks via text generation. In International Conference on Machine Learning, pages 1931-1942. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.76, + 0.47, + 0.801 + ], + "angle": 0, + "content": "[10] Sungjoon Choi, Qian-Yi Zhou, Stephen Miller, and Vladlen Koltun. A large dataset of object scans. arXiv preprint arXiv:1602.02481, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.47, + 0.899 + ], + "angle": 0, + "content": "[11] Jasmine Collins, Shubham Goel, Kenan Deng, Achleshwar Luthra, Leon Xu, Erhan Gundogdu, Xi Zhang, Tomas F Yago Vicente, Thomas Dideriksien, Himanshu Arora, et al. Abo: Dataset and benchmarks for real-world 3d object understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21126-21136, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.117, + 0.471, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.133 + ], + "angle": 0, + "content": "[12] Blender Online Community. Blender - a 3d modelling and rendering package. http://www.blender.org, 2018.7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.134, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[13] Camille Couprie, Clément Farabet, Laurent Najman, and Yann LeCun. Indoor semantic segmentation using depth information. arXiv preprint arXiv:1301.3572, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.178, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[14] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.247, + 0.892, + 0.314 + ], + "angle": 0, + "content": "[15] Matt Deitke, Dhruv Batra, Yonatan Bisk, Tommaso Campari, Angel X Chang, Devendra Singh Chaplot, Changan Chen, Claudia Pérez D'Arpino, Kiana Ehsani, Ali Farhadi, et al. Retrospectives on the embodied ai workshop. arXiv preprint arXiv:2210.06849, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.316, + 0.892, + 0.385 + ], + "angle": 0, + "content": "[16] Matt Deitke, Eli VanderBilt, Alvaro Herrasti, Luca Weihs, Jordi Salvador, Kiana Ehsani, Winson Han, Eric Kolve, Ali Farhadi, Aniruddha Kembhavi, et al. Procthor: Large-scale embodied ai using procedural generation. Conference on Neural Information Processing Systems, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.385, + 0.892, + 0.441 + ], + "angle": 0, + "content": "[17] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.442, + 0.892, + 0.51 + ], + "angle": 0, + "content": "[18] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas B McHugh, and Vincent Vanhoucke. Google scanned objects: A high-quality dataset of 3d scanned household items. arXiv preprint arXiv:2204.11918, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.511, + 0.892, + 0.566 + ], + "angle": 0, + "content": "[19] Chengjian Feng, Yujie Zhong, and Weilin Huang. Exploring classification equilibrium in long-tailed object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3417-3426, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.567, + 0.892, + 0.621 + ], + "angle": 0, + "content": "[20] Huan Fu, Rongfei Jia, Lin Gao, Mingming Gong, Binqiang Zhao, Steve Maybank, and Dacheng Tao. 3d-future: 3d furniture shape with texture. International Journal of Computer Vision, 129(12):3313-3337, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.622, + 0.892, + 0.691 + ], + "angle": 0, + "content": "[21] Chuang Gan, Jeremy Schwartz, Seth Alter, Martin Schrimpf, James Traer, Julian De Freitas, Jonas Kubilius, Abhishek Bhandwaldar, Nick Haber, Megumi Sano, et al. Threadworld: A platform for interactive multi-modal physical simulation. arXiv preprint arXiv:2007.04954, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.692, + 0.892, + 0.761 + ], + "angle": 0, + "content": "[22] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. In Advances In Neural Information Processing Systems, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.762, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[23] Golnaz Ghiasi, Yin Cui, Aravind Srinivas, Rui Qian, Tsung-Yi Lin, Ekin D Cubuk, Quoc V Le, and Barret Zoph. Simple copy-paste is a strong data augmentation method for instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2918-2928, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.845, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[24] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5356-5364, 2019. 3, 6" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13150" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[25] Kaiming He, Georgia Gkioxari, Piotr Dollár, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.134, + 0.47, + 0.19 + ], + "angle": 0, + "content": "[26] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.19, + 0.47, + 0.259 + ], + "angle": 0, + "content": "[27] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. OpenCLIP, July 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.26, + 0.47, + 0.314 + ], + "angle": 0, + "content": "[28] Allan Jabri, Andrew Owens, and Alexei Efros. Space-time correspondence as a contrastive random walk. Advances in neural information processing systems, 33:19545-19560, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.316, + 0.47, + 0.384 + ], + "angle": 0, + "content": "[29] Ajay Jain, Ben Mildenhall, Jonathan T Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 867-876, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.385, + 0.47, + 0.467 + ], + "angle": 0, + "content": "[30] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.469, + 0.47, + 0.523 + ], + "angle": 0, + "content": "[31] Armand Joulin, Edouard Grave, Piotr Bojanowski, Matthijs Douze, Hérve Jégou, and Tomas Mikolov. Fasttext.zip: Compressing text classification models. arXiv preprint arXiv:1612.03651, 2016. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.525, + 0.47, + 0.566 + ], + "angle": 0, + "content": "[32] Armand Joulin, Edouard Grave, Piotr Bojanowski, and Tomas Mikolov. Bag of tricks for efficient text classification. arXiv preprint arXiv:1607.01759, 2016. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.567, + 0.47, + 0.635 + ], + "angle": 0, + "content": "[33] Alexander Kasper, Zhixing Xue, and Rüdiger Dillmann. The kit object models database: An object model database for object recognition, localization and manipulation in service robotics. The International Journal of Robotics Research, 31(8):927-934, 2012. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.636, + 0.47, + 0.704 + ], + "angle": 0, + "content": "[34] Apoorv Khandelwal, Luca Weihs, Roozbeh Mottaghi, and Aniruddha Kembhavi. Simple but effective: Clip embeddings for embodied ai. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14809-14818, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.706, + 0.47, + 0.789 + ], + "angle": 0, + "content": "[35] Sebastian Koch, Albert Matveev, Zhongshi Jiang, Francis Williams, Alexey Artemov, Evgeny Burnaev, Marc Alexa, Denis Zorin, and Daniele Panozzo. Abc: A big cad model dataset for geometric deep learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9601-9611, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.47, + 0.871 + ], + "angle": 0, + "content": "[36] Eric Kolve, Roozbeh Mottaghi, Winson Han, Eli VanderBilt, Luca Weihs, Alvaro Herrasti, Matt Deitke, Kiana Ehsani, Daniel Gordon, Yuke Zhu, Aniruddha Kembhavi, Abhinav Kumar Gupta, and Ali Farhadi. Ai2-thor: An interactive 3d environment for visual ai. arXiv e-prints, pages arXiv-1712, 2017. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.872, + 0.47, + 0.902 + ], + "angle": 0, + "content": "[37] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4. International Journal of Computer Vision, 128(7):1956-1981, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.135, + 0.892, + 0.231 + ], + "angle": 0, + "content": "[38] Marc Levoy, Kari Pulli, Brian Curless, Szymon Rusinkiewicz, David Koller, Lucas Pereira, Matt Ginzton, Sean Anderson, James Davis, Jeremy Ginsberg, et al. The digital michelangelo project: 3d scanning of large statues. In Proceedings of the 27th annual conference on Computer graphics and interactive techniques, pages 131-144, 2000. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.232, + 0.892, + 0.314 + ], + "angle": 0, + "content": "[39] Chengshu Li, Fei Xia, Roberto Martín-Martín, Michael Lingelbach, Sanjana Srivastava, Bokui Shen, Kent Vainio, Cem Gokmen, Gokul Dharan, Tanish Jain, et al. igibson 2.0: Object-centric simulation for robot learning of everyday household tasks. arXiv preprint arXiv:2108.03272, 2021. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.315, + 0.892, + 0.37 + ], + "angle": 0, + "content": "[40] Joseph J Lim, Hamed Pirsiavash, and Antonio Torralba. Parsing IKEA objects: Fine pose estimation. In Proceedings of the IEEE international conference on computer vision, pages 2992-2999, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.371, + 0.892, + 0.44 + ], + "angle": 0, + "content": "[41] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.441, + 0.892, + 0.496 + ], + "angle": 0, + "content": "[42] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. Advances in neural information processing systems, 32, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.497, + 0.892, + 0.538 + ], + "angle": 0, + "content": "[43] Khaled Mamou, E Lengyel, and A Peters. Volumetric hierarchical approximate convex decomposition. In Game Engine Gems 3, pages 141-158. AK Peters, 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.539, + 0.892, + 0.607 + ], + "angle": 0, + "content": "[44] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13492-13502, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.608, + 0.892, + 0.636 + ], + "angle": 0, + "content": "[45] George A Miller. Wordnet: a lexical database for english. Communications of the ACM, 38(11):39-41, 1995. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.636, + 0.892, + 0.691 + ], + "angle": 0, + "content": "[46] Kaichun Mo, Paul Guerrero, Li Yi, Hao Su, Peter Wonka, Niloy Mitra, and Leonidas J Guibas. Structurenet: Hierarchical graph networks for 3d shape generation. arXiv preprint arXiv:1908.00575, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.692, + 0.892, + 0.775 + ], + "angle": 0, + "content": "[47] Kaichun Mo, Shilin Zhu, Angel X Chang, Li Yi, Subarna Tripathi, Leonidas J Guibas, and Hao Su. Partnet: A largescale benchmark for fine-grained and hierarchical part-level 3d object understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 909-918, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.776, + 0.892, + 0.831 + ], + "angle": 0, + "content": "[48] D. Morrison, P. Corke, and J. Leitner. Egad! an evolved grasping analysis dataset for diversity and reproducibility in robotic manipulation. IEEE Robotics and Automation Letters, 5(3):4368-4375, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[49] Tai-Yu Pan, Cheng Zhang, Yandong Li, Hexiang Hu, Dong Xuan, Soravit Changpinyo, Boqing Gong, and Wei-Lun Chao. On model calibration for long-tailed object detection and instance segmentation. Advances in Neural Information Processing Systems, 34:2529-2542, 2021. 6" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "13151" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[50] Keunhong Park, Konstantinos Rematas, Ali Farhadi, and Steven M Seitz. Photoshape: Photorealistic materials for large-scale shape collections. arXiv preprint arXiv:1809.09761, 2018.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.148, + 0.471, + 0.218 + ], + "angle": 0, + "content": "[51] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865-5874, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.218, + 0.471, + 0.259 + ], + "angle": 0, + "content": "[52] Matt Pharr, Wenzel Jakob, and Greg Humphreys. Physically based rendering: From theory to implementation. Morgan Kaufmann, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.26, + 0.471, + 0.302 + ], + "angle": 0, + "content": "[53] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.302, + 0.471, + 0.37 + ], + "angle": 0, + "content": "[54] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10318–10327, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.372, + 0.471, + 0.455 + ], + "angle": 0, + "content": "[55] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 1, 2, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.455, + 0.471, + 0.509 + ], + "angle": 0, + "content": "[56] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.511, + 0.471, + 0.579 + ], + "angle": 0, + "content": "[57] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.581, + 0.471, + 0.65 + ], + "angle": 0, + "content": "[58] Ram Ramrakhya, Eric Undersander, Dhruv Batra, and Abhishek Das. Habitat-web: Learning embodied object-search strategies from human demonstrations at scale. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5173-5183, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.65, + 0.471, + 0.719 + ], + "angle": 0, + "content": "[59] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.72, + 0.471, + 0.789 + ], + "angle": 0, + "content": "[60] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115(3):211-252, 2015. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.471, + 0.871 + ], + "angle": 0, + "content": "[61] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. arXiv preprint arXiv:2210.08402, 2022. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.872, + 0.471, + 0.902 + ], + "angle": 0, + "content": "[62] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.135, + 0.894, + 0.217 + ], + "angle": 0, + "content": "[63] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2556-2565, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.218, + 0.894, + 0.273 + ], + "angle": 0, + "content": "[64] Arjun Singh, James Sha, Karthik S. Narayan, Tudor Achim, and P. Abbeel. Bigbird: A large-scale 3d database of object instances. 2014 IEEE International Conference on Robotics and Automation (ICRA), pages 509-516, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.274, + 0.894, + 0.357 + ], + "angle": 0, + "content": "[65] Krishna Srinivasan, Karthik Raman, Jiecao Chen, Michael Bendersky, and Marc Najork. Wit: Wikipedia-based image text dataset for multimodal multilingual machine learning. In Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 2443-2449, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.358, + 0.894, + 0.44 + ], + "angle": 0, + "content": "[66] Xingyuan Sun, Jiajun Wu, Xiuming Zhang, Zhoutong Zhang, Chengkai Zhang, Tianfan Xue, Joshua B Tenenbaum, and William T Freeman. Pix3d: Dataset and methods for single-image 3d shape modeling. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2974-2983, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.441, + 0.894, + 0.523 + ], + "angle": 0, + "content": "[67] Andrew Szot, Alexander Clegg, Eric Undersander, Erik Wijmans, Yili Zhao, John Turner, Noah Maestre, Mustafa Mukadam, Devendra Singh Chaplot, Oleksandr Maksymets, et al. Habitat 2.0: Training home assistants to rearrange their habitat. Advances in Neural Information Processing Systems, 34:251-266, 2021. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.525, + 0.894, + 0.566 + ], + "angle": 0, + "content": "[68] Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. arXiv preprint arXiv:1908.07490, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.567, + 0.894, + 0.636 + ], + "angle": 0, + "content": "[69] Jingru Tan, Xin Lu, Gang Zhang, Changqing Yin, and Quanquan Li. Equalization loss v2: A new gradient balance approach for long-tailed object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1685-1694, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.636, + 0.894, + 0.718 + ], + "angle": 0, + "content": "[70] Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P Srinivasan, Jonathan T Barron, and Henrik Kretzschmar. Block-nerf: Scalable large scene neural view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8248–8258, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.719, + 0.894, + 0.761 + ], + "angle": 0, + "content": "[71] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Amit H Bermano, and Daniel Cohen-Or. Human motion diffusion model. arXiv preprint arXiv:2209.14916, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.761, + 0.894, + 0.816 + ], + "angle": 0, + "content": "[72] Bart Thomee, David A Shamma, Gerald Friedland, Benjamin Elizalde, Karl Ni, Douglas Poland, Damian Borth, and Li-Jia Li. Yfcc100m: The new data in multimedia research. Communications of the ACM, 59(2):64-73, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.818, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[73] Jiaqi Wang, Wenwei Zhang, Yuhang Zang, Yuhang Cao, Jiangmiao Pang, Tao Gong, Kai Chen, Ziwei Liu, Chen Change Loy, and Dahua Lin. Seesaw loss for long-tailed instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9695-9704, 2021. 6" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13152" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[74] Wenhui Wang, Hangbo Bao, Li Dong, Johan Bjorck, Zhiliang Peng, Qiang Liu, Kriti Aggarwal, Owais Khan Mohammed, Saksham Singhal, Subhojit Som, et al. Image as a foreign language: Beit pretraining for all vision and vision-language tasks. arXiv preprint arXiv:2208.10442, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.47, + 0.219 + ], + "angle": 0, + "content": "[75] Luca Weihs, Jordi Salvador, Klemen Kotar, Unnat Jain, Kuo-Hao Zeng, Roozbeh Mottaghi, and Aniruddha Kembhavi. Allenact: A framework for embodied ai research. arXiv preprint arXiv:2008.12760, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.22, + 0.469, + 0.275 + ], + "angle": 0, + "content": "[76] Erik Wijmans, Abhishek Kadian, Ari S. Morcos, Stefan Lee, Irfan Essa, Devi Parikh, Manolis Savva, and Dhruv Batra. Dd-ppo: Learning near-perfect pointgoal navigators from 2.5 billion frames. In ICLR, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.277, + 0.469, + 0.332 + ], + "angle": 0, + "content": "[77] Rundi Wu, Chang Xiao, and Changxi Zheng. Deepcad: A deep generative network for computer-aided design models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6772–6782, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.334, + 0.469, + 0.403 + ], + "angle": 0, + "content": "[78] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguuang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1912-1920, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.404, + 0.469, + 0.485 + ], + "angle": 0, + "content": "[79] Fanbo Xiang, Yuzhe Qin, Kaichun Mo, Yikuan Xia, Hao Zhu, Fangchen Liu, Minghua Liu, Hanxiao Jiang, Yifu Yuan, He Wang, et al. Sapien: A simulated part-based interactive environment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11097-11107, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.488, + 0.469, + 0.557 + ], + "angle": 0, + "content": "[80] Yuanbo Xiangli, Linning Xu, Xingang Pan, Nanxuan Zhao, Anyi Rao, Christian Theobalt, Bo Dai, and Dahua Lin. Bungeenerf: Progressive neural radiance field for extreme multi-scale scene rendering. In The European Conference on Computer Vision (ECCV), 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.559, + 0.469, + 0.626 + ], + "angle": 0, + "content": "[81] Zhenjia Xu, Beichun Qi, Shubham Agrawal, and Shuran Song. Adagrasp: Learning an adaptive gripper-aware grasping policy. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 4620-4626. IEEE, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.629, + 0.469, + 0.724 + ], + "angle": 0, + "content": "[82] Rowan Zellers, Jiasen Lu, Ximing Lu, Youngjae Yu, Yanpeng Zhao, Mohammadreza Salehi, Aditya Kusupati, Jack Hessel, Ali Farhadi, and Yejin Choi. Merlot reserve: Neural script knowledge through vision and language and sound. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16375-16387, 2022. 1, 4" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.724 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13153" + } + ] +] \ No newline at end of file diff --git a/2023/Objaverse_ A Universe of Annotated 3D Objects/ea3a0563-acc5-4afc-aad6-362a9401875b_origin.pdf b/2023/Objaverse_ A Universe of Annotated 3D Objects/ea3a0563-acc5-4afc-aad6-362a9401875b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..456f79b509b6d81fd281a054038929c173bb1672 --- /dev/null +++ b/2023/Objaverse_ A Universe of Annotated 3D Objects/ea3a0563-acc5-4afc-aad6-362a9401875b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a1a65a90ab972c0e64a1ceadb92d37b70bc252ea974ba2f6b27076125cf87d79 +size 684353 diff --git a/2023/Objaverse_ A Universe of Annotated 3D Objects/full.md b/2023/Objaverse_ A Universe of Annotated 3D Objects/full.md new file mode 100644 index 0000000000000000000000000000000000000000..790317642715fa358d6e18050e0efef987059bca --- /dev/null +++ b/2023/Objaverse_ A Universe of Annotated 3D Objects/full.md @@ -0,0 +1,324 @@ +# Objaverse: A Universe of Annotated 3D Objects + +Matt Deitke $^{\dagger \psi}$ , Dustin Schwenk $^{\dagger}$ , Jordi Salvador $^{\dagger}$ , Luca Weihs $^{\dagger}$ , Oscar Michel $^{\dagger}$ , Eli VanderBilt $^{\dagger}$ , Ludwig Schmidt $^{\psi}$ , Kiana Ehsani $^{\dagger}$ , Aniruddha Kembhavi $^{\dagger \psi}$ , Ali Farhadi $^{\psi}$ , PRIOR @ Allen Institute for AI, $^{\psi}$ University of Washington, Seattle objaverse.allenai.org + +![](images/d27174bf5aa9b8c6b03aef22ab2204e1fece6cc396a02880cc669df3e5ab6385.jpg) +Figure 1. Example instances from our large-scale 3D asset dataset OBJVERSE. OBJVERSE 3D assets are semantically diverse, high-quality, and paired with natural-language descriptions. + +# Abstract + +Massive data corpora like WebText, Wikipedia, Conceptual Captions, WebImageText, and LAION have propelled recent dramatic progress in AI. Large neural models trained on such datasets produce impressive results and top many of today's benchmarks. A notable omission1ion within this family of large-scale datasets is 3D data. Despite considerable interest and potential applications in 3D vision, datasets of high-fidelity 3D models continue to be mid-sized with limited diversity of object categories. Addressing this gap, we present Objaverse 1.0, a large dataset of objects with $800K+$ (and growing) 3D models with descriptive captions, tags, and animations. Objaverse improves upon present day 3D repositories in terms of scale, number of categories, and in the visual diversity of instances within a category. We demonstrate the large potential of Objaverse via four diverse applications: training generative 3D models, im + +proving tail category segmentation on the LVIS benchmark, training open-vocabulary object-navigation models for Embodied AI, and creating a new benchmark for robustness analysis of vision models. Objaverse can open new directions for research and enable new applications across the field of AI. + +# 1. Introduction + +Massive datasets have enabled and driven rapid progress in AI. Language corpora on the web led to large language models like GPT-3 [4]; paired image and text datasets like Conceptual Captions [63] led to vision-and-language pretrained models like VilBERT [42]; YouTube video datasets led to video capable models like Merlot-Reserve [82]; and massive multimodal datasets like WebImageText [65] and LAION [61, 62] led to models like CLIP [55] and StableDiffusion [59]. These leaps in dataset scale and diversity were triggered by moving from manually curated datasets to harnessing the power of the web and its creative content. + +In contrast to the datasets described above, the size of + +the datasets we are feeding to our data-hungry deep learning models in many other areas of research is simply not comparable. For instance, the number of 3D assets used in training generative 3D models is, maximally, on the order of thousands [22] and the simulators used to train embodied AI models typically have only between a few dozen to a thousand unique scenes [36, 39, 58, 67]. The startling advances brought about by developing large-scale datasets for images, videos, and natural language, demand that an equivalent dataset be built for 3D assets. + +We present OBJAYERSE 1.0, a large scale corpus of high-quality, richly annotated, 3D objects; see Fig. 1. Objects in our dataset are free to use1 and sourced from Sketchfab, a leading online platform for managing, viewing, and distributing 3D models. In total, OBJAYERSE contains over 800K 3D assets designed by over 150K artists which makes this data large and diversely sourced. Assets not only belong to varied categories like animals, humans, and vehicles, but also include interiors and exteriors of large spaces that can be used, e.g., to train embodied agents. OBJAYERSE is a universe of rich 3D data with detailed metadata that can support many different annotations to enable new applications. With this remarkable increase in scale, we see an incredible opportunity for OBJAYERSE to impact research progress across domains. In this work, we provide promising results to answer three questions. + +Can 3D vision benefit from a large-scale dataset? First, as a 3D asset resource, OBJVERSE can support the exciting field of 3D generative modeling. We use data extracted from OBJVERSE to train generative models for single and multiple categories using GET3D [22] and find that we are able to generate high-quality objects. Moreover, we find that our generated objects are found by human annotators to be more diverse than those generated by a model trained on ShapeNet objects in $91\%$ of cases. + +Can the diversity of 3D models help improve classical 2D vision task performance? To answer this question, we use the diversity of OBJVERSE to improve the performance of long tail instance segmentation models. Instance segmentation data can be expensive to obtain owing to the cost of annotating contours around objects. The recent LVIS dataset contains segmentation annotations for 1,230 categories but the task remains very challenging for present day models, particularly on tail categories that have few examples. We show that increasing the volume of data by leveraging a simple Copy+Paste augmentation method with OBJVERSE assets can improve the performance of state-of-the-art segmentation methods. + +We also use OBJVERSE to build a benchmark for evaluating the robustness of state-of-the-art visual classification models to perspective shifts. We render objects in OBJVERSE from random orientations, which is how one + +might expect to see them in the real world, and test the ability of CLIP-style visual backbones to correctly classify these images. Our experiments show that current state-of-the-art models' performance degrades dramatically in this setting when viewing objects from arbitrary views. OBJVERSE allows us to build benchmarks to test (and potentially train) for orientation robustness for a long tail distribution of asset categories. Building such benchmarks is made uniquely possible by the scale and diversity of 3D assets in OBJVERSE. This would simply not be feasible to create in the real world nor can they be generated from existing 2D images. + +Can a large-scale 3D dataset help us train perform- +mant embodied agents? We use assets in ObjAVERSE to +populate procedurally generated simulated environments in +ProcTHOR [16] that are used to train Embodied AI agents. +This results in an orders of magnitude increase in the num +ber of unique assets available for use in ProcTHOR scenes +(Previously limited to AI2-THOR's [36] asset library of a +few thousand unique instances each assigned to one of 108 +object categories). Using ObjAVERSE populated scenes en +ables open vocabulary object navigation from any text de +scription. In this paper, we provide quantitative results for +navigating to 1.1K semantic object categories, roughly a +50x increase. + +These findings represent just a small fraction of what can be accomplished using OBJVERSE. We are excited to see how the research community will leverage OBJVERSE to enable fast and exciting progress in 2D and 3D computer vision applications and beyond. + +# 2. Related Work + +Large scale datasets. Scaling the size and scope of training datasets has widely been demonstrated to be an effective avenue of improvement for model performance. In computer vision, the adoption of early large scale datasets such as Imagenet [17,60] and MS-COCO [41] has dramatically accelerated progress on a variety of tasks including classification, object detection, captioning, and more. Ever since, the diversity and scale of datasets have continued to grow. YFCC100M is a dataset of 99.2M images and 800K videos [72]. OpenImages [37] is a large scale dataset of 9M images that contains labeled subsets bounding boxes, visual relationships, segmentation masks, localized narratives, and categorical annotations. Massive web-scraped datasets containing image-text pairs such as Conceptual Captions [63], WIT [65], and LAION [61,62] have seen increased popularity recently as they have been used to train impressive models for vision-language representation learning [27,30,55], text-to-image generation [30,56,57,59], and vision-language multitasking [8,9,68,74]. + +3D datasets. Current large-scale 2D image datasets offer three crucial components that benefit learning: scale, + +![](images/5c950e32123e8db1b6a14ea9fd110fc402b44b9c0b32f85745854c1ce63842dc.jpg) +Objverse + +![](images/b75316214db6200a6ee1919761e93615b32977476914291829d738de2c7ab97d.jpg) +ShapeNet +Figure 2. Comparison between OBJVERSE and existing 3D object datasets. (Left:) Visual comparison of instances from OBJVERSE and ShapeNet for the categories of CAR, BED, VASE, and BAG. OBJVERSE instances are substantially more diverse since objects can come from many 3D content creation platforms, whereas ShapeNet models look more similar and all come from SketchUp, a 3D modeling platform built for simple architectural modeling. (Right:) Scale comparison table between existing 3D object datasets. + +
Dataset# Objects# Classes
YCB [5]775
BigBIRD [64]125-
KIT [33]219145
IKEA [40]21911
Pix3D [66]3959
GSO [18]1K17
EGAD [48]2K-
PhotoShape [50]5K1
ABO [11]8K63
3D-Future [20]10K34
ShapeNet [6]51K55
Objaverse818K21K
+ +diversity, and realism. Ideally, models that reason about 3D objects should have access to datasets that meet these same criteria. However, of the numerous 3D object datasets that currently exist, none are able to excel in all three categories to the same degree as their 2D counterparts. Datasets such as KIT [33], YCB [5], BigBIRD [64], IKEA [40], and Pix3D [66] provide image-calibrated models over a diverse set of household objects, but severely lack in scale with only a few hundred objects at most. EGAD [48] procedurally generates 2K objects for grasping, but produces objects that are not that realistic or diverse. Slightly larger datasets of photo-realistic objects include GSO [18], PhotoShape [50], ABO [11] and 3D-Future [20], and ShapeNet [6] with object counts in the tens of thousands, see Fig. 2 for comparisons between OBJVERSE and these datasets. Datasets for CAD models, such as ModelNet [78] and DeepCAD [77], and ABC [35] do not include textures or materials, which limits their ability to represent objects that could plausibly be found in the real world. Datasets of scanned 3D objects and environments are valuable for real-world understanding [10, 13, 14, 38], but are quite small and limited. In addition to containing numerous artist designed objects, OBJVERSE contains many scanned assets, making it a useful source of data for learning from real-world distributions. + +While rapid progress has been made in developing datasets that combine image and text, in contrast, only a few datasets that pair language and 3D data exist. Text2Shape [7] released a dataset of 15,038 chairs and tables from ShapeNet each with around 5 text captions, giving 75,344 total text-shape pairs. ShapeGlot [1] released the CiC (Chairs in Context) dataset which contains 4,511 chairs from ShapeNet along with 78,789 descriptive utterances generated from a referential game. Due to the small scale and limited diversity of these datasets, current SoTA text-to-3D models [29,44,53] forgo the use of 3D datasets entirely + +and instead rely on 2D image-text supervision. + +# 3. Objverse + +OBJAVERSE is a massive annotated 3D dataset that can be used to enable research in a wide range of areas across computer vision. The objects are sourced from Sketchfab, an online 3D marketplace where users can upload and share models for both free and commercial use. Objects selected for OBJVERSE have a distributable Creative Commons license and were obtained using Sketchfab's public API. Aside from licensing consideration, models marked as restricted due to objectionable or adult thematic content were excluded from the dataset. + +Model metadata. OBJVERSE objects inherit a set of foundational annotations supplied by their creator when uploaded to Sketchfab. Figure 4 shows an example of the metadata available for each model. The metadata includes a name, assignments to a set of fixed categories, a set of unrestricted tags, and a natural language description. + +OBJAVERSE-LVIS. While OBJAVERSE metadata contains a great deal of information about objects, Sketchfab's existing categorization scheme covers only 18 categories, too coarse for most applications. Object names, categories, and tags provide multiple potential categorizations at varying levels of specificity and with some inherent noise. However, for many existing computer vision tasks, it is useful to assign objects to a single category drawn from a predetermined set of the right size and level of semantic granularity. + +We choose the categories from the LVIS dataset [24] for categorizing a long-tail subset of objects in OBJVERSE. We construct a 47K LVIS categorized object subset, called OBJVERSE-LVIS, comprised of objects uniquely assigned to one of 1156 LVIS categories. We perform these assignments by first selecting 500 candidate objects per category using a combination of predictions from a CLIP classifi- + +![](images/efa256944f5ac1bfacdc0fd3b9c2eec90f9b3b63f6cc5f8aa15a004e42d6e850.jpg) +(a) Breakdown of high-level OBJVERSE categories + +![](images/130ef3e48f4de85c2ef3d74081848d75aba43bbce8e716255dfc9f15a22e3bbe.jpg) +(b) Word cloud of OBJVERSE metadata tags. + +![](images/6cd2455fa6ddc613c8418b6c2bb78ba3dbe2f5546204618ce46ab2d428931a2d.jpg) +(c) OBJAYERSE-LVIS category occurrence distribution. + +![](images/c4778b2d2da25378220938528fd057d369eb877a05541899a2ad194e445ce874.jpg) +(d) OBJAYERSE tag occurrence distribution. +Figure 3. ObjAverse statistics. (a) All 18 high-level categories present in ObjAverse's metadata with their corresponding number of occurrences. The relative share of most popular categories are evenly split, with a small number of less frequently categories. (b) A sample of several thousand popular object tags found in ObjAverse log-scaled by their frequency. (c) A histogram of fine-grained ObjAverse-LVIS categories with representative members from several bins highlighted. (d) A histogram of ObjAverse tags with representative members from several bins highlighted (note y-axis log scale). Tags from the low-occurrence side of the distribution correspond to unique objects that, taken individually, are rarely seen in the world. Frequently used tags like "furniture" and "car" reflect their real-world normalcy, but the high frequency of assets like "sword" diverge from their real-world counterparts. + +cation model and candidates suggested by terms in their metadata. This combined pool contains objects visually resembling the target category (from the CLIP features of their thumbnail images) that might have missing metadata, as well as visually unusual instances of a category that are accurately named or tagged. These 250k candidate objects were then manually filtered and their assigned categories verified by crowdworkers. Since we only presented 500 object candidates per class, many popular categories, such as chair or car, have substantially more objects that could be included in OBJAYERSE-LVIS with future annotations. + +Animated objects and rigged characters. OBJVERSE includes 44K animated objects and over 63K objects self-categorized as characters. Examples of animations include fridge doors opening, animals running, and the hands on a clock moving. Rigged characters can be set up for animation and rendering, and may often come annotated with bone mappings. The vast scale of animations available in OBJVERSE can support a wide range of research in temporal 3D learning, such as building text-based animation generative models [71], representing object changes over time + +with NERFs [51, 54], and temporal self-supervised learning via. future frame prediction [28, 82]. + +Articulated objects. Decomposing 3D objects into parts has led to a flurry of research in the past few years, including work in learning robotic grasping policies [79, 81], 3D semantic segmentation [47], and shape generation [46]. Since many objects in OBJAYERSE were uploaded by artists, the objects often come separated into parts. Figure 5 shows an example, where a chair is separated by its backrest, wheels, and legs, among many smaller parts. + +Exteriors. Photogrammetry and NERF advances have enabled the commercialization of capturing high-quality 3D objects of large exteriors by taking pictures [70, 80]. In OBJVERSE, there are a large number of scanned buildings, cities, and stadiums. Figure 5 shows an example of a 3D object of NYC's skyline captured through a scan. + +OBJAVERSE-Interiors. There are $16\mathrm{K}+$ interior scenes in OBJAVERSE, including houses, classrooms, and offices. The scenes often have multiple floors, many types of rooms, and are densely populated with objects from human input. Objects in the scenes are separable into parts, which allows + +![](images/2ae6e510eb49b547d316e5c34919efdc85c3ce5acd8eb4f3806f824b58cb61e8.jpg) +Figure 4. An example of metadata available for each object in OBJVERSE. Each uploaded object has a 3D model, user-selected rendered thumbnail image, name, description, tags, category, and stats, among additional metadata. + +them to be usable for interactive robotics, embodied AI, and scene synthesis. To put the scale of OBJVERSE-Interiors in perspective, the number of scenes in OBJVERSE-Interiors is significantly larger than the 400 or so existing hand-built interactive embodied AI scenes [21,36,39,67]. + +Visual styles. Objects in the world can be constructed in many styles and often differ in style based on the time-period, geographic location, and artist's style. OBJVERSE objects cover a vast set of visual styles, including 3D scans, 3D modeled objects from virtually any platform, point clouds, and photo-realism via physically based rendering (PBR) [52]. Moreover, instances of objects often appear with many styles, which is critical for training and evaluating robust computer vision models [55]. Figure 5 shows examples of chairs in OBJVERSE in many different styles, including Gothic, modern, Victorian, cartoon, and abstract. + +Statistics. OBJVERSE 1.0 includes 818K 3D objects, designed by 160K artists. There are $>2.35\mathrm{M}$ tags on the objects, with $>170\mathrm{K}$ of them being unique. We estimate that the objects have coverage for nearly 21K WordNet entities [45] (see appendix for details). Objects were uploaded between 2012 and 2022, with over 200K objects uploaded uploaded just in 2021. Figure 3 visualizes several statistics of the dataset, including the breakdown of objects into their + +self-assigned Sketchfab categories, a word cloud over the tags, a frequency plot of the tags, and the number of objects in OBJAYERSE-LVIS categories. + +# 4. Applications + +In this section, we present 4 initial distinct applications of OBJVERSE, including 3D generative modeling, instance segmentation with CP3D, open-vocabulary ObjectNav, and analyze robustness in computer vision models. + +# 4.1. 3D Generative Modeling + +3D generative modeling has shown much improvement recently with models such as GET3D [22] delivering impressive high quality results with rich geometric details. GET3D is trained to generate 3D textured meshes for a category and produces impressive 3D objects for categories like Car, Chair, and Motorcycle using data from ShapeNet [6]. OBJVERSE contains 3D models for many diverse categories including tail categories which are not represented in other datasets. It also contains diverse and realistic object instances per category. This scale and diversity can be used to train large vocabulary and high quality 3D generative models. In this work, we showcase the potential of this data as follows. We choose three categories of ob- + +![](images/44c8b5c5e5a1e4701007798a436663cb732d5c01ad604d78470a7168ad1b4dcc.jpg) +Figure 5. Highlights of the visual diversity of objects that appear in OBJVERSE, including animated objects, rigged (body-part annotated) characters, models separatable into parts, exterior environments, interior environments, and a wide range visual styles. + +![](images/1897e9edb06e5ebeb723e88e3d50a6f5e9c5c4a01c2229ebd026eaed4048aa37.jpg) +Objaverse Bag Generations + +![](images/76725f113584d350d32777b8cd96b334d519f4be3129402e97db4ef290d081d0.jpg) +ShapeNet Bag Generations + +![](images/7721da45e4abe5aad684934b546f37b24ac8a6160b7947af563dabef504b58ed.jpg) +(a) Comparison of Bags generated with OBJVERSE and ShapeNet. +Shoe Generations + +![](images/e55973c2cfbc54edc9b1cb6eb6fb9683ffbd125d85de9f822d47a873b599e22e.jpg) +Fruit&Veg Generations +(b) Shoe and Fruit&Veg. generations. + +![](images/4206ebb7e776c70eb499cdbf2192ca5ffcf3fc6d15acdee0b7ed329ba07ff600.jpg) +Fruit&Veg Interpolation +(c) Fruit&Veg. interpolation. +Figure 6. (a) Example GET3D Bag object generations using OBJVERSE and ShapeNet models for training. (b) Additional Shoe and Fruit&Veg generations from OBJVERSE models. (c) models generated when interpolating between two, randomly sampled, latent encodings with our trained Fruit&Veg. model; what appears to be a pumpkin smoothly transforms into a mushroom. + +jects, Shoe, Bag, and Fruit&Veg, and subsample objects from OBJVERSE to create three separate datasets containing, respectively, 143 shoes, 816 bags, and 571 fruits & vegetables (116 apples, 112 gourds, 92 mushrooms, 68 bananas, 52 oranges, 52 pears, 31 potatoes 24 lemons, and 24 pineapples). For comparison, we also train a GET3D model on the set of 83 bags from the ShapeNet dataset. Fig. 6 shows a collection of 3D objects generated by our trained GET3D models. Qualitatively, the 3D-meshes generated by the OBJVERSE-trained models are high-quality and diverse, especially when compared to the generations from the ShapeNet-trained model. To quantify this observation, we asked crowdworkers to rate the diversity of Bag generations produced by the OBJVERSE and ShapeNet trained models. When shown collections of nine randomly sampled + +![](images/2d568ef167cdfef8cbd5757de0bf5a3138b984dd8bd4e39ada1779261896e044.jpg) +Figure 7. An illustration of CP3D (copy-paste 3D) for segmentation augmentation. We render 3D objects from multiple views and paste them over LVIS training images. + +generations from both models, workers rated the collection generated from the OBJVERSE trained model as more diverse in appearance $91\%$ of the time. + +Our fruits and vegetables, composed of 9 varieties produces perhaps the highest quality output, a promising signal that can inspire future work in text-to-3D generation. + +# 4.2. Instance Segmentation with CP3D + +A key advantage of using simulated data for computer vision is that it is much cheaper to obtain expert annotations. Annotated OBJVERSE objects can be rendered into images, allowing them to serve as a rich source of additional data that can be used to enhance model performance on 2D computer vision tasks. As a proof-of-concept demonstrating the effectiveness of this approach, we use segmented data from OBJVERSE objects as auxiliary labels for training models on the LVIS dataset for Large Scale Instance Segmentation [24]. The LVIS dataset contains instance segmentation masks for 1200 object categories that occur throughout a set of $164\mathrm{k}$ images. Recognition is especially challenging in this task due to the long tail of the object category distribution in this dataset. LVIS categories only con + +
MethodAPAPrAPcAPf
RFS [24]23.713.323.029.0
EQLv2 [69]25.517.724.330.2
LOCE [19]26.618.526.230.7
NorCal with RFS [49]25.219.324.229.0
Seesaw [73]26.419.526.129.7
GOL [3]27.721.427.730.4
GOL + CP3D28.321.828.331.1
+ +Table 1. Comparison of our approach (GOL+CP3D) against SoTA Mask-RCNN ResNet-50 models on LVIS. We report results for APr, APc, and APf which measure AP for categories that are rare (appear in 1-10 images), common (appear in 11-100 images), and frequent (appear in $>100$ images), respectively + +![](images/4fa6782d8fa095cdf09a093835f3cede63e70411b507d47c3530b45458562905.jpg) +Figure 8. An existing ProcTHOR scene (left) and a semantically similar ProcTHOR generatable scene with OBJVERSE objects (right). + +![](images/096eec98402182a571df44cf631a5a93dbb7f7defbf4a80fb36a5536d8c1582b.jpg) + +tain an average 9 instances across the dataset, so training on simulated data is a promising approach for overcoming the challenges of learning in this low-sample regime. + +Using the LVIS-annotated subset of ObjAVERSE, we introduce CP3D: an enhancement to the simple, but effective, copy-and-paste technique of [23]. Figure 7 shows an example of the setup for CP3D. Here, we render different views of 3D objects and paste them on-top of existing LVIS images. We render 5 distinct views of each object and cache them for use throughout training. During training, an image is selected for the copy-paste augmentation with 0.5 probability, and once selected, 1-3 images of randomly chosen LVIS-annotated ObjAVERSE objects are pasted onto the selected training image. The segmentation masks of the selected objects are added to the training image's annotation as well. Object images and masks are randomly scaled and translated before being pasted. We use this strategy to finetune the pretrained ResNet-50 Mask-RCNN [25, 26] of [3]. As shown in Tab.1, simply finetuning this model for 24 epochs yields performance gains across several metrics. + +# 4.3. Open-Vocabulary ObjectNav + +In this section, we introduce open-vocabulary ObjectNav, a new task propelled by the vast diversity of objects that appear in OBJAYERSE. Here, an agent is placed at a random starting location inside of a home and tasked to navigate to a target object provided from a text description (e.g. "Raspberry Pi Pico"). To facilitate this task, we procedurally generate 10K new homes in ProcTHOR [16] fully populated with objects from OBJAYERSE-LVIS. Until now, ObjectNav tasks have focused on training agents to navigate to 20 or so target objects provided their category label [15, 16, 58], and existing interactive embodied AI simulations, including ProcTHOR, only include around 2K total objects across around 100 object types [16, 39, 67]. In this work, we take a large step to massively scale the number of target objects used in ObjectNav ( $20 \to \text{OpenVocab}$ ), the number of objects available in simulation ( $2\mathbf{K} \to 36\mathbf{K}$ ), and the number of object types of the objects ( $100 \to 1.1\mathbf{K}$ ). + +Object placement. To make the placement of objects in the houses more natural, we use the OBJVERSE-LVIS subset and annotate placement constraints for each object + +category. Specifically, we annotate if objects of a given category typically appears on the floor, on-top of a surface, or on a wall. If instances of the object category may appear on the floor, we also annotate whether it may appear in the middle of the scene (e.g. a clutter object like a basketball) or on the edge of the scene (e.g. a toilet or a fridge). For objects placed on the floor, we also to automatically detect flat regions on top of the object's mesh to place surface object types. The annotations are used by ProcTHOR for sampling objects to place in a scene. We also filter out OBJAYERSE-LVIS objects that do not appear inside of homes, such as a jet plane. Structural objects, like doors and windows, are inherited from ProcTHOR as they would require additional cleanup. + +Object size correction. Objects in Sketchfab may be uploaded at unnatural scales (e.g. a plant being as large as a tower). We therefore scale the objects to be of a reasonable size for them to look natural in a house. Here, for each object category, we annotate the maximum bounding box dimension length that every instance of the object category should be scaled to. For example, we annotate the maximum bounding box dimension for bookcase to be 2 meters and fork to be 0.18 meters. If a 3D modeled bookcase then has a bounding box of $20\mathrm{m}\times 6\mathrm{m}\times 3\mathrm{m}$ , we shrink each side by a factor of $\max (20,6,3) / 2 = 5$ . + +Preprocessing for AI2-THOR. We add support to AI2-THOR for loading objects on the fly at runtime. Previously, all objects had to be stored in a Unity build, but such an approach is impractical when working with orders of magnitude more object data. For each object, we compress it with Blender [12] by joining all of its meshes together, decimate the joined mesh such that it has at most 5K vertices, and bake all the UV texture maps into a single texture map. We then generate colliders using V-HACD [43] to support rigid-body interactions. + +Approach. Given procedural houses populated with OBJVERSE-LVIS, the task is to navigate to the proximity of a chosen target object and invoke a task-completion action when the target object is in sight, given an open-vocabulary description formed with the template "a {name} {category}" The name is the object name given by its creator, which is often descriptive. We filter + +![](images/ce79a2e0bd4d79922e4ca0ff9aa7070a8f2cc0e617560e82d4ca0acf156c29d9.jpg) +Television 99.6% + +![](images/82cf596b8e34decf4cfac9cc4bdf62e0bf7756ca0da27d19d6853a58c9194d2a.jpg) +Television 96.1% + +![](images/bc6c4c147c226e984ff05e3fc2dcf17cc7a5cffc713823bd2ce8508a2a253844.jpg) +Vase 39.0% + +![](images/3823427c5bc8dae555d3db806c11590db8b0e0a463eefddaca8f23e9ac685ba6.jpg) +Flowerpot 34.9% + +![](images/0822e0addd82a3b04016ec6b6ac45f00366446491301187cbb71ea99c4e8b1f0.jpg) +Clock $99.8\%$ + +![](images/32954a33ab0101bec606b19c6b1c7c2e0aa942305dabb95bb800f2f3ecba188f.jpg) +Clock 98.2% + +![](images/24c7d257474f5c15ac3b633adeaefbfdf6ecbf4d829c9b8706e6a24236f71bbb.jpg) +Flowerpot 32.4% + +![](images/7ae741bf261e216b98d703f1aaaf74e83979a66e44b237c86e4755ee6eec8595.jpg) +Vase 33.2% +Figure 9. Examples of objects rendered from random orientations and their 0-shot classification categories with the CLIP ViT-B/32. + +each by whether it is detected as being written in English by a language detector [31,32], and fall back to a class-only description for non-English name. Examples of the possible expressions include “a victorian-monobike motorcycle”, “a unicorn pony”, or “a dino ghost lizard”. The agent, similar to the ones in [34], observes an RGB egocentric view of the environment, pre-processed by the visual branch of a frozen ResNet-50 CLIP model [55] – the target description is pre-processed by the corresponding text branch. We train the agent with DD-PPO [76] and evaluate on houses with floor plans, objects, and descriptions unseen in training. We use the AllenAct [75] framework to train our agent. Our trained agent achieves a success rate of $19.9\%$ , for a random policy success of $5.1\%$ . For more details about the experiment refer to the appendix. + +# 4.4. Analyzing Robustness + +A persistent bias present in many image datasets, e.g. ImageNet [60], is that the subjects of interest are generally photographed from a forward-facing, canonical, orientation. When, for example, taking a photograph of a television, few would choose to take this photograph crouched on the floor behind the television. This impact of this bias was studied by Alcorn et al. [2] who find that modern computer vision systems are highly susceptible to deviations from canonical poses. This is more than a theoretical problem: computer vision systems deployed in the real world will frequently encounter objects in non-canonical orientations and in many applications, e.g. autonomous driving, it will be safety critical that they behave well. + +Given the above, we adopt the experimental design of Alcorn et al. and design, using OBJVERSE assets, a benchmark for evaluating the robustness of state-of-the-art computer vision classification models to orientation shifts. In particular, for each object in our OBJVERSE-LVIS subset, we render 12 images of the object from random orientations rendered upon a background with RGB values equalling the mean RGB values from ImageNet; see Fig. 9 for examples. This ability to, at scale, render objects from random view + +
ModelRandom RotationAny RotationΔTop-1
Top-1Top-5Top-1Top-5
OpenAI-400M [55]
RN5021.4%45.0%43.9%70.8%22.5%
ViT-L/1429.1%54.5%52.3%77.2%23.2%
LAION-400M [62]
ViT-B/3224.1%48.5%46.9%74.2%22.8%
ViT-L/1430.6%56.8%50.5%77.0%19.9%
LAION-2B [61]
ViT-B/3227.0%51.8%50.3%76.1%23.3%
ViT-L/1432.9%59.2%52.1%78.0%19.2%
ViT-H/1432.3%58.8%50.1%77.3%17.8%
+ +Table 2. Evaluating 0-shot CLIP classification models on our rotational robustness benchmark. $\Delta$ Top-1 denotes the difference between Top-1 Any Rotation and Top-1 Random Rotation. Models are strongly overfit to standard views of objects. + +points is a practical impossibility in the real world but is made trivial when using 3D assets. We then evaluate several modern open-domain image-classification networks (constrained to the $\approx 1,200$ LVIS categories) on these images and report 4 metrics for each model. These metrics include: + +- Top-1 Random Rotation - the frequency with which a model correctly classifies an image as belonging to the respective LVIS category. +- Top-1 Any Rotation - the frequency with which a model classifies an image correctly from at least one of the 12 random orientations. + +This second metric is diagnostic and serves to represent a model's performance when shown an object from a canonical pose. We also have Top-5 variants of the above metric where the correct category need only be in the top 5 predictions from the model. We report our results in Tab. 2 in which we evaluate a variety of performant pretrained models. Comparing the gap in performance between the Top-k Random Rotation and Top-k Any Rotation metrics we find that model performance dramatically degrades when viewing objects from unusual orientations. + +# 5. Conclusion + +We present OBJAYERSE, a next-generation 3D asset library containing 818K high-quality, diverse, 3D models with paired text descriptions, titles, and tags. As a small glimpse of the potential uses of OBJAYERSE, we present four experimental studies showing how OBJAYERSE can be used to power (1) generative 3D models with clear future applications to text-to-3D generation, (2) improvements to classical computer vision tasks such as instance segmentation, (3) the creation of novel embodied AI tasks like Open Vocabulary Object Navigation, and (4) quantifying the rotational robustness of vision models on renderings of objects. We hope to see OBJAYERSE enable a new universe of new applications for computer vision. + +# References + +[1] Panos Achlioptas, Judy Fan, Robert Hawkins, Noah Goodman, and Leonidas J Guibas. Shapeplot: Learning language for shape differentiation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8938-8947, 2019. 3 +[2] Michael A. Alcorn, Qi Li, Zhitao Gong, Chengfei Wang, Long Mai, Wei-Shinn Ku, and Anh Nguyen. Strike (with) a pose: Neural networks are easily fooled by strange poses of familiar objects. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, pages 4845-4854. Computer Vision Foundation / IEEE, 2019. 8 +[3] Konstantinos Panagiotis Alexandridis, Jiankang Deng, Anh Nguyen, and Shan Luo. Long-tailed instance segmentation using gumbel optimized loss. arXiv preprint arXiv:2207.10936, 2022. 6, 7 +[4] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 1 +[5] Berk Calli, Aaron Walsman, Arjun Singh, Siddhartha Srinivasa, Pieter Abbeel, and Aaron M Dollar. Benchmarking in manipulation research: The ycb object and model set and benchmarking protocols. arXiv preprint arXiv:1502.03143, 2015. 3 +[6] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 3, 5 +[7] Kevin Chen, Christopher B Choy, Manolis Savva, Angel X Chang, Thomas Funkhouser, and Silvio Savarese. Text2shape: Generating shapes from natural language by learning joint embeddings. In Asian conference on computer vision, pages 100-116. Springer, 2018. 3 +[8] Xi Chen, Xiao Wang, Soravit Changpinyo, AJ Piergiovanni, Piotr Padlewski, Daniel Salz, Sebastian Goodman, Adam Grycner, Basil Mustafa, Lucas Beyer, et al. Pali: A jointly-scaled multilingual language-image model. arXiv preprint arXiv:2209.06794, 2022. 2 +[9] Jaemin Cho, Jie Lei, Hao Tan, and Mohit Bansal. Unifying vision-and-language tasks via text generation. In International Conference on Machine Learning, pages 1931-1942. PMLR, 2021. 2 +[10] Sungjoon Choi, Qian-Yi Zhou, Stephen Miller, and Vladlen Koltun. A large dataset of object scans. arXiv preprint arXiv:1602.02481, 2016. 3 +[11] Jasmine Collins, Shubham Goel, Kenan Deng, Achleshwar Luthra, Leon Xu, Erhan Gundogdu, Xi Zhang, Tomas F Yago Vicente, Thomas Dideriksien, Himanshu Arora, et al. Abo: Dataset and benchmarks for real-world 3d object understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21126-21136, 2022. 3 + +[12] Blender Online Community. Blender - a 3d modelling and rendering package. http://www.blender.org, 2018.7 +[13] Camille Couprie, Clément Farabet, Laurent Najman, and Yann LeCun. Indoor semantic segmentation using depth information. arXiv preprint arXiv:1301.3572, 2013. 3 +[14] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 3 +[15] Matt Deitke, Dhruv Batra, Yonatan Bisk, Tommaso Campari, Angel X Chang, Devendra Singh Chaplot, Changan Chen, Claudia Pérez D'Arpino, Kiana Ehsani, Ali Farhadi, et al. Retrospectives on the embodied ai workshop. arXiv preprint arXiv:2210.06849, 2022. 7 +[16] Matt Deitke, Eli VanderBilt, Alvaro Herrasti, Luca Weihs, Jordi Salvador, Kiana Ehsani, Winson Han, Eric Kolve, Ali Farhadi, Aniruddha Kembhavi, et al. Procthor: Large-scale embodied ai using procedural generation. Conference on Neural Information Processing Systems, 2022. 2, 7 +[17] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 2 +[18] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas B McHugh, and Vincent Vanhoucke. Google scanned objects: A high-quality dataset of 3d scanned household items. arXiv preprint arXiv:2204.11918, 2022. 3 +[19] Chengjian Feng, Yujie Zhong, and Weilin Huang. Exploring classification equilibrium in long-tailed object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3417-3426, 2021. 6 +[20] Huan Fu, Rongfei Jia, Lin Gao, Mingming Gong, Binqiang Zhao, Steve Maybank, and Dacheng Tao. 3d-future: 3d furniture shape with texture. International Journal of Computer Vision, 129(12):3313-3337, 2021. 3 +[21] Chuang Gan, Jeremy Schwartz, Seth Alter, Martin Schrimpf, James Traer, Julian De Freitas, Jonas Kubilius, Abhishek Bhandwaldar, Nick Haber, Megumi Sano, et al. Threadworld: A platform for interactive multi-modal physical simulation. arXiv preprint arXiv:2007.04954, 2020. 5 +[22] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. In Advances In Neural Information Processing Systems, 2022. 2, 5 +[23] Golnaz Ghiasi, Yin Cui, Aravind Srinivas, Rui Qian, Tsung-Yi Lin, Ekin D Cubuk, Quoc V Le, and Barret Zoph. Simple copy-paste is a strong data augmentation method for instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2918-2928, 2021. 7 +[24] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5356-5364, 2019. 3, 6 + +[25] Kaiming He, Georgia Gkioxari, Piotr Dollár, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 7 +[26] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 7 +[27] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. OpenCLIP, July 2021. 2 +[28] Allan Jabri, Andrew Owens, and Alexei Efros. Space-time correspondence as a contrastive random walk. Advances in neural information processing systems, 33:19545-19560, 2020. 4 +[29] Ajay Jain, Ben Mildenhall, Jonathan T Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 867-876, 2022. 3 +[30] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916. PMLR, 2021. 2 +[31] Armand Joulin, Edouard Grave, Piotr Bojanowski, Matthijs Douze, Hérve Jégou, and Tomas Mikolov. Fasttext.zip: Compressing text classification models. arXiv preprint arXiv:1612.03651, 2016. 8 +[32] Armand Joulin, Edouard Grave, Piotr Bojanowski, and Tomas Mikolov. Bag of tricks for efficient text classification. arXiv preprint arXiv:1607.01759, 2016. 8 +[33] Alexander Kasper, Zhixing Xue, and Rüdiger Dillmann. The kit object models database: An object model database for object recognition, localization and manipulation in service robotics. The International Journal of Robotics Research, 31(8):927-934, 2012. 3 +[34] Apoorv Khandelwal, Luca Weihs, Roozbeh Mottaghi, and Aniruddha Kembhavi. Simple but effective: Clip embeddings for embodied ai. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14809-14818, 2022. 8 +[35] Sebastian Koch, Albert Matveev, Zhongshi Jiang, Francis Williams, Alexey Artemov, Evgeny Burnaev, Marc Alexa, Denis Zorin, and Daniele Panozzo. Abc: A big cad model dataset for geometric deep learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9601-9611, 2019. 3 +[36] Eric Kolve, Roozbeh Mottaghi, Winson Han, Eli VanderBilt, Luca Weihs, Alvaro Herrasti, Matt Deitke, Kiana Ehsani, Daniel Gordon, Yuke Zhu, Aniruddha Kembhavi, Abhinav Kumar Gupta, and Ali Farhadi. Ai2-thor: An interactive 3d environment for visual ai. arXiv e-prints, pages arXiv-1712, 2017. 2, 5 +[37] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan + +Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4. International Journal of Computer Vision, 128(7):1956-1981, 2020. 2 +[38] Marc Levoy, Kari Pulli, Brian Curless, Szymon Rusinkiewicz, David Koller, Lucas Pereira, Matt Ginzton, Sean Anderson, James Davis, Jeremy Ginsberg, et al. The digital michelangelo project: 3d scanning of large statues. In Proceedings of the 27th annual conference on Computer graphics and interactive techniques, pages 131-144, 2000. 3 +[39] Chengshu Li, Fei Xia, Roberto Martín-Martín, Michael Lingelbach, Sanjana Srivastava, Bokui Shen, Kent Vainio, Cem Gokmen, Gokul Dharan, Tanish Jain, et al. igibson 2.0: Object-centric simulation for robot learning of everyday household tasks. arXiv preprint arXiv:2108.03272, 2021. 2, 5, 7 +[40] Joseph J Lim, Hamed Pirsiavash, and Antonio Torralba. Parsing IKEA objects: Fine pose estimation. In Proceedings of the IEEE international conference on computer vision, pages 2992-2999, 2013. 3 +[41] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 2 +[42] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. Advances in neural information processing systems, 32, 2019. 1 +[43] Khaled Mamou, E Lengyel, and A Peters. Volumetric hierarchical approximate convex decomposition. In Game Engine Gems 3, pages 141-158. AK Peters, 2016. 7 +[44] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13492-13502, 2022. 3 +[45] George A Miller. Wordnet: a lexical database for english. Communications of the ACM, 38(11):39-41, 1995. 5 +[46] Kaichun Mo, Paul Guerrero, Li Yi, Hao Su, Peter Wonka, Niloy Mitra, and Leonidas J Guibas. Structurenet: Hierarchical graph networks for 3d shape generation. arXiv preprint arXiv:1908.00575, 2019. 4 +[47] Kaichun Mo, Shilin Zhu, Angel X Chang, Li Yi, Subarna Tripathi, Leonidas J Guibas, and Hao Su. Partnet: A largescale benchmark for fine-grained and hierarchical part-level 3d object understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 909-918, 2019. 4 +[48] D. Morrison, P. Corke, and J. Leitner. Egad! an evolved grasping analysis dataset for diversity and reproducibility in robotic manipulation. IEEE Robotics and Automation Letters, 5(3):4368-4375, 2020. 3 +[49] Tai-Yu Pan, Cheng Zhang, Yandong Li, Hexiang Hu, Dong Xuan, Soravit Changpinyo, Boqing Gong, and Wei-Lun Chao. On model calibration for long-tailed object detection and instance segmentation. Advances in Neural Information Processing Systems, 34:2529-2542, 2021. 6 + +[50] Keunhong Park, Konstantinos Rematas, Ali Farhadi, and Steven M Seitz. Photoshape: Photorealistic materials for large-scale shape collections. arXiv preprint arXiv:1809.09761, 2018.3 +[51] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865-5874, 2021. 4 +[52] Matt Pharr, Wenzel Jakob, and Greg Humphreys. Physically based rendering: From theory to implementation. Morgan Kaufmann, 2016. 5 +[53] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 3 +[54] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10318–10327, 2021. 4 +[55] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 1, 2, 5, 8 +[56] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2 +[57] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 2 +[58] Ram Ramrakhya, Eric Undersander, Dhruv Batra, and Abhishek Das. Habitat-web: Learning embodied object-search strategies from human demonstrations at scale. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5173-5183, 2022. 2, 7 +[59] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 1, 2 +[60] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115(3):211-252, 2015. 2, 8 +[61] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. arXiv preprint arXiv:2210.08402, 2022. 1, 2, 8 +[62] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo + +Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 1, 2, 8 +[63] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2556-2565, 2018. 1, 2 +[64] Arjun Singh, James Sha, Karthik S. Narayan, Tudor Achim, and P. Abbeel. Bigbird: A large-scale 3d database of object instances. 2014 IEEE International Conference on Robotics and Automation (ICRA), pages 509-516, 2014. 3 +[65] Krishna Srinivasan, Karthik Raman, Jiecao Chen, Michael Bendersky, and Marc Najork. Wit: Wikipedia-based image text dataset for multimodal multilingual machine learning. In Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 2443-2449, 2021. 1, 2 +[66] Xingyuan Sun, Jiajun Wu, Xiuming Zhang, Zhoutong Zhang, Chengkai Zhang, Tianfan Xue, Joshua B Tenenbaum, and William T Freeman. Pix3d: Dataset and methods for single-image 3d shape modeling. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2974-2983, 2018. 3 +[67] Andrew Szot, Alexander Clegg, Eric Undersander, Erik Wijmans, Yili Zhao, John Turner, Noah Maestre, Mustafa Mukadam, Devendra Singh Chaplot, Oleksandr Maksymets, et al. Habitat 2.0: Training home assistants to rearrange their habitat. Advances in Neural Information Processing Systems, 34:251-266, 2021. 2, 5, 7 +[68] Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. arXiv preprint arXiv:1908.07490, 2019. 2 +[69] Jingru Tan, Xin Lu, Gang Zhang, Changqing Yin, and Quanquan Li. Equalization loss v2: A new gradient balance approach for long-tailed object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1685-1694, 2021. 6 +[70] Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P Srinivasan, Jonathan T Barron, and Henrik Kretzschmar. Block-nerf: Scalable large scene neural view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8248–8258, 2022. 4 +[71] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Amit H Bermano, and Daniel Cohen-Or. Human motion diffusion model. arXiv preprint arXiv:2209.14916, 2022. 4 +[72] Bart Thomee, David A Shamma, Gerald Friedland, Benjamin Elizalde, Karl Ni, Douglas Poland, Damian Borth, and Li-Jia Li. Yfcc100m: The new data in multimedia research. Communications of the ACM, 59(2):64-73, 2016. 2 +[73] Jiaqi Wang, Wenwei Zhang, Yuhang Zang, Yuhang Cao, Jiangmiao Pang, Tao Gong, Kai Chen, Ziwei Liu, Chen Change Loy, and Dahua Lin. Seesaw loss for long-tailed instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9695-9704, 2021. 6 + +[74] Wenhui Wang, Hangbo Bao, Li Dong, Johan Bjorck, Zhiliang Peng, Qiang Liu, Kriti Aggarwal, Owais Khan Mohammed, Saksham Singhal, Subhojit Som, et al. Image as a foreign language: Beit pretraining for all vision and vision-language tasks. arXiv preprint arXiv:2208.10442, 2022. 2 +[75] Luca Weihs, Jordi Salvador, Klemen Kotar, Unnat Jain, Kuo-Hao Zeng, Roozbeh Mottaghi, and Aniruddha Kembhavi. Allenact: A framework for embodied ai research. arXiv preprint arXiv:2008.12760, 2020. 8 +[76] Erik Wijmans, Abhishek Kadian, Ari S. Morcos, Stefan Lee, Irfan Essa, Devi Parikh, Manolis Savva, and Dhruv Batra. Dd-ppo: Learning near-perfect pointgoal navigators from 2.5 billion frames. In ICLR, 2020. 8 +[77] Rundi Wu, Chang Xiao, and Changxi Zheng. Deepcad: A deep generative network for computer-aided design models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6772–6782, 2021. 3 +[78] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguuang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1912-1920, 2015. 3 +[79] Fanbo Xiang, Yuzhe Qin, Kaichun Mo, Yikuan Xia, Hao Zhu, Fangchen Liu, Minghua Liu, Hanxiao Jiang, Yifu Yuan, He Wang, et al. Sapien: A simulated part-based interactive environment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11097-11107, 2020. 4 +[80] Yuanbo Xiangli, Linning Xu, Xingang Pan, Nanxuan Zhao, Anyi Rao, Christian Theobalt, Bo Dai, and Dahua Lin. Bungeenerf: Progressive neural radiance field for extreme multi-scale scene rendering. In The European Conference on Computer Vision (ECCV), 2022. 4 +[81] Zhenjia Xu, Beichun Qi, Shubham Agrawal, and Shuran Song. Adagrasp: Learning an adaptive gripper-aware grasping policy. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 4620-4626. IEEE, 2021. 4 +[82] Rowan Zellers, Jiasen Lu, Ximing Lu, Youngjae Yu, Yanpeng Zhao, Mohammadreza Salehi, Aditya Kusupati, Jack Hessel, Ali Farhadi, and Yejin Choi. Merlot reserve: Neural script knowledge through vision and language and sound. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16375-16387, 2022. 1, 4 \ No newline at end of file diff --git a/2023/Objaverse_ A Universe of Annotated 3D Objects/images.zip b/2023/Objaverse_ A Universe of Annotated 3D Objects/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d85f7614b3a64ef128c11615bc2338bf3ed977ed --- /dev/null +++ b/2023/Objaverse_ A Universe of Annotated 3D Objects/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:15fdfb0ebf635f64a8e354d9eacfa5d1c56c9fe507629d979f4c1efe37a1df83 +size 700020 diff --git a/2023/Objaverse_ A Universe of Annotated 3D Objects/layout.json b/2023/Objaverse_ A Universe of Annotated 3D Objects/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..01c4b075927a48cffda734850bb5f6af3959810e --- /dev/null +++ b/2023/Objaverse_ A Universe of Annotated 3D Objects/layout.json @@ -0,0 +1,8660 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 148, + 103, + 446, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 103, + 446, + 121 + ], + "spans": [ + { + "bbox": [ + 148, + 103, + 446, + 121 + ], + "type": "text", + "content": "Objaverse: A Universe of Annotated 3D Objects" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 134, + 524, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 134, + 524, + 194 + ], + "spans": [ + { + "bbox": [ + 67, + 134, + 524, + 194 + ], + "type": "text", + "content": "Matt Deitke" + }, + { + "bbox": [ + 67, + 134, + 524, + 194 + ], + "type": "inline_equation", + "content": "^{\\dagger \\psi}" + }, + { + "bbox": [ + 67, + 134, + 524, + 194 + ], + "type": "text", + "content": ", Dustin Schwenk" + }, + { + "bbox": [ + 67, + 134, + 524, + 194 + ], + "type": "inline_equation", + "content": "^{\\dagger}" + }, + { + "bbox": [ + 67, + 134, + 524, + 194 + ], + "type": "text", + "content": ", Jordi Salvador" + }, + { + "bbox": [ + 67, + 134, + 524, + 194 + ], + "type": "inline_equation", + "content": "^{\\dagger}" + }, + { + "bbox": [ + 67, + 134, + 524, + 194 + ], + "type": "text", + "content": ", Luca Weihs" + }, + { + "bbox": [ + 67, + 134, + 524, + 194 + ], + "type": "inline_equation", + "content": "^{\\dagger}" + }, + { + "bbox": [ + 67, + 134, + 524, + 194 + ], + "type": "text", + "content": ", Oscar Michel" + }, + { + "bbox": [ + 67, + 134, + 524, + 194 + ], + "type": "inline_equation", + "content": "^{\\dagger}" + }, + { + "bbox": [ + 67, + 134, + 524, + 194 + ], + "type": "text", + "content": ", Eli VanderBilt" + }, + { + "bbox": [ + 67, + 134, + 524, + 194 + ], + "type": "inline_equation", + "content": "^{\\dagger}" + }, + { + "bbox": [ + 67, + 134, + 524, + 194 + ], + "type": "text", + "content": ", Ludwig Schmidt" + }, + { + "bbox": [ + 67, + 134, + 524, + 194 + ], + "type": "inline_equation", + "content": "^{\\psi}" + }, + { + "bbox": [ + 67, + 134, + 524, + 194 + ], + "type": "text", + "content": ", Kiana Ehsani" + }, + { + "bbox": [ + 67, + 134, + 524, + 194 + ], + "type": "inline_equation", + "content": "^{\\dagger}" + }, + { + "bbox": [ + 67, + 134, + 524, + 194 + ], + "type": "text", + "content": ", Aniruddha Kembhavi" + }, + { + "bbox": [ + 67, + 134, + 524, + 194 + ], + "type": "inline_equation", + "content": "^{\\dagger \\psi}" + }, + { + "bbox": [ + 67, + 134, + 524, + 194 + ], + "type": "text", + "content": ", Ali Farhadi" + }, + { + "bbox": [ + 67, + 134, + 524, + 194 + ], + "type": "inline_equation", + "content": "^{\\psi}" + }, + { + "bbox": [ + 67, + 134, + 524, + 194 + ], + "type": "text", + "content": ", PRIOR @ Allen Institute for AI, " + }, + { + "bbox": [ + 67, + 134, + 524, + 194 + ], + "type": "inline_equation", + "content": "^{\\psi}" + }, + { + "bbox": [ + 67, + 134, + 524, + 194 + ], + "type": "text", + "content": "University of Washington, Seattle objaverse.allenai.org" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 50, + 202, + 545, + 430 + ], + "blocks": [ + { + "bbox": [ + 50, + 202, + 545, + 430 + ], + "lines": [ + { + "bbox": [ + 50, + 202, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 50, + 202, + 545, + 430 + ], + "type": "image", + "image_path": "d27174bf5aa9b8c6b03aef22ab2204e1fece6cc396a02880cc669df3e5ab6385.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 438, + 546, + 462 + ], + "lines": [ + { + "bbox": [ + 46, + 438, + 546, + 462 + ], + "spans": [ + { + "bbox": [ + 46, + 438, + 546, + 462 + ], + "type": "text", + "content": "Figure 1. Example instances from our large-scale 3D asset dataset OBJVERSE. OBJVERSE 3D assets are semantically diverse, high-quality, and paired with natural-language descriptions." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 478, + 192, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 478, + 192, + 491 + ], + "spans": [ + { + "bbox": [ + 143, + 478, + 192, + 491 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 496, + 290, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 496, + 290, + 688 + ], + "spans": [ + { + "bbox": [ + 46, + 496, + 290, + 688 + ], + "type": "text", + "content": "Massive data corpora like WebText, Wikipedia, Conceptual Captions, WebImageText, and LAION have propelled recent dramatic progress in AI. Large neural models trained on such datasets produce impressive results and top many of today's benchmarks. A notable omission1ion within this family of large-scale datasets is 3D data. Despite considerable interest and potential applications in 3D vision, datasets of high-fidelity 3D models continue to be mid-sized with limited diversity of object categories. Addressing this gap, we present Objaverse 1.0, a large dataset of objects with " + }, + { + "bbox": [ + 46, + 496, + 290, + 688 + ], + "type": "inline_equation", + "content": "800K+" + }, + { + "bbox": [ + 46, + 496, + 290, + 688 + ], + "type": "text", + "content": " (and growing) 3D models with descriptive captions, tags, and animations. Objaverse improves upon present day 3D repositories in terms of scale, number of categories, and in the visual diversity of instances within a category. We demonstrate the large potential of Objaverse via four diverse applications: training generative 3D models, im" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 475, + 545, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 475, + 545, + 545 + ], + "spans": [ + { + "bbox": [ + 306, + 475, + 545, + 545 + ], + "type": "text", + "content": "proving tail category segmentation on the LVIS benchmark, training open-vocabulary object-navigation models for Embodied AI, and creating a new benchmark for robustness analysis of vision models. Objaverse can open new directions for research and enable new applications across the field of AI." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 307, + 550, + 387, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 550, + 387, + 562 + ], + "spans": [ + { + "bbox": [ + 307, + 550, + 387, + 562 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 570, + 545, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 700 + ], + "type": "text", + "content": "Massive datasets have enabled and driven rapid progress in AI. Language corpora on the web led to large language models like GPT-3 [4]; paired image and text datasets like Conceptual Captions [63] led to vision-and-language pretrained models like VilBERT [42]; YouTube video datasets led to video capable models like Merlot-Reserve [82]; and massive multimodal datasets like WebImageText [65] and LAION [61, 62] led to models like CLIP [55] and StableDiffusion [59]. These leaps in dataset scale and diversity were triggered by moving from manually curated datasets to harnessing the power of the web and its creative content." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 317, + 701, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 701, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 701, + 546, + 713 + ], + "type": "text", + "content": "In contrast to the datasets described above, the size of" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 703, + 201, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 703, + 201, + 713 + ], + "spans": [ + { + "bbox": [ + 61, + 703, + 201, + 713 + ], + "type": "text", + "content": "Correspondence to ." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "13142" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": "the datasets we are feeding to our data-hungry deep learning models in many other areas of research is simply not comparable. For instance, the number of 3D assets used in training generative 3D models is, maximally, on the order of thousands [22] and the simulators used to train embodied AI models typically have only between a few dozen to a thousand unique scenes [36, 39, 58, 67]. The startling advances brought about by developing large-scale datasets for images, videos, and natural language, demand that an equivalent dataset be built for 3D assets." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 192, + 289, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 192, + 289, + 384 + ], + "spans": [ + { + "bbox": [ + 46, + 192, + 289, + 384 + ], + "type": "text", + "content": "We present OBJAYERSE 1.0, a large scale corpus of high-quality, richly annotated, 3D objects; see Fig. 1. Objects in our dataset are free to use1 and sourced from Sketchfab, a leading online platform for managing, viewing, and distributing 3D models. In total, OBJAYERSE contains over 800K 3D assets designed by over 150K artists which makes this data large and diversely sourced. Assets not only belong to varied categories like animals, humans, and vehicles, but also include interiors and exteriors of large spaces that can be used, e.g., to train embodied agents. OBJAYERSE is a universe of rich 3D data with detailed metadata that can support many different annotations to enable new applications. With this remarkable increase in scale, we see an incredible opportunity for OBJAYERSE to impact research progress across domains. In this work, we provide promising results to answer three questions." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 384, + 289, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 384, + 289, + 491 + ], + "spans": [ + { + "bbox": [ + 46, + 384, + 289, + 491 + ], + "type": "text", + "content": "Can 3D vision benefit from a large-scale dataset? First, as a 3D asset resource, OBJVERSE can support the exciting field of 3D generative modeling. We use data extracted from OBJVERSE to train generative models for single and multiple categories using GET3D [22] and find that we are able to generate high-quality objects. Moreover, we find that our generated objects are found by human annotators to be more diverse than those generated by a model trained on ShapeNet objects in " + }, + { + "bbox": [ + 46, + 384, + 289, + 491 + ], + "type": "inline_equation", + "content": "91\\%" + }, + { + "bbox": [ + 46, + 384, + 289, + 491 + ], + "type": "text", + "content": " of cases." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 491, + 289, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 491, + 289, + 647 + ], + "spans": [ + { + "bbox": [ + 46, + 491, + 289, + 647 + ], + "type": "text", + "content": "Can the diversity of 3D models help improve classical 2D vision task performance? To answer this question, we use the diversity of OBJVERSE to improve the performance of long tail instance segmentation models. Instance segmentation data can be expensive to obtain owing to the cost of annotating contours around objects. The recent LVIS dataset contains segmentation annotations for 1,230 categories but the task remains very challenging for present day models, particularly on tail categories that have few examples. We show that increasing the volume of data by leveraging a simple Copy+Paste augmentation method with OBJVERSE assets can improve the performance of state-of-the-art segmentation methods." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 647, + 288, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 288, + 696 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 288, + 696 + ], + "type": "text", + "content": "We also use OBJVERSE to build a benchmark for evaluating the robustness of state-of-the-art visual classification models to perspective shifts. We render objects in OBJVERSE from random orientations, which is how one" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 72, + 547, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 547, + 216 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 547, + 216 + ], + "type": "text", + "content": "might expect to see them in the real world, and test the ability of CLIP-style visual backbones to correctly classify these images. Our experiments show that current state-of-the-art models' performance degrades dramatically in this setting when viewing objects from arbitrary views. OBJVERSE allows us to build benchmarks to test (and potentially train) for orientation robustness for a long tail distribution of asset categories. Building such benchmarks is made uniquely possible by the scale and diversity of 3D assets in OBJVERSE. This would simply not be feasible to create in the real world nor can they be generated from existing 2D images." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 217, + 547, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 217, + 547, + 371 + ], + "spans": [ + { + "bbox": [ + 304, + 217, + 547, + 371 + ], + "type": "text", + "content": "Can a large-scale 3D dataset help us train perform- \nmant embodied agents? We use assets in ObjAVERSE to \npopulate procedurally generated simulated environments in \nProcTHOR [16] that are used to train Embodied AI agents. \nThis results in an orders of magnitude increase in the num \nber of unique assets available for use in ProcTHOR scenes \n(Previously limited to AI2-THOR's [36] asset library of a \nfew thousand unique instances each assigned to one of 108 \nobject categories). Using ObjAVERSE populated scenes en \nables open vocabulary object navigation from any text de \nscription. In this paper, we provide quantitative results for \nnavigating to 1.1K semantic object categories, roughly a \n50x increase." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 372, + 547, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 372, + 547, + 433 + ], + "spans": [ + { + "bbox": [ + 304, + 372, + 547, + 433 + ], + "type": "text", + "content": "These findings represent just a small fraction of what can be accomplished using OBJVERSE. We are excited to see how the research community will leverage OBJVERSE to enable fast and exciting progress in 2D and 3D computer vision applications and beyond." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 441, + 392, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 441, + 392, + 453 + ], + "spans": [ + { + "bbox": [ + 306, + 441, + 392, + 453 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 461, + 547, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 461, + 547, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 461, + 547, + 689 + ], + "type": "text", + "content": "Large scale datasets. Scaling the size and scope of training datasets has widely been demonstrated to be an effective avenue of improvement for model performance. In computer vision, the adoption of early large scale datasets such as Imagenet [17,60] and MS-COCO [41] has dramatically accelerated progress on a variety of tasks including classification, object detection, captioning, and more. Ever since, the diversity and scale of datasets have continued to grow. YFCC100M is a dataset of 99.2M images and 800K videos [72]. OpenImages [37] is a large scale dataset of 9M images that contains labeled subsets bounding boxes, visual relationships, segmentation masks, localized narratives, and categorical annotations. Massive web-scraped datasets containing image-text pairs such as Conceptual Captions [63], WIT [65], and LAION [61,62] have seen increased popularity recently as they have been used to train impressive models for vision-language representation learning [27,30,55], text-to-image generation [30,56,57,59], and vision-language multitasking [8,9,68,74]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 689, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 547, + 715 + ], + "type": "text", + "content": "3D datasets. Current large-scale 2D image datasets offer three crucial components that benefit learning: scale," + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 58, + 702, + 151, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 151, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 151, + 712 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 58, + 702, + 151, + 712 + ], + "type": "text", + "content": "Creative Commons license" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13143" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 90, + 225, + 219 + ], + "blocks": [ + { + "bbox": [ + 51, + 71, + 89, + 82 + ], + "lines": [ + { + "bbox": [ + 51, + 71, + 89, + 82 + ], + "spans": [ + { + "bbox": [ + 51, + 71, + 89, + 82 + ], + "type": "text", + "content": "Objverse" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 48, + 90, + 225, + 219 + ], + "lines": [ + { + "bbox": [ + 48, + 90, + 225, + 219 + ], + "spans": [ + { + "bbox": [ + 48, + 90, + 225, + 219 + ], + "type": "image", + "image_path": "5c950e32123e8db1b6a14ea9fd110fc402b44b9c0b32f85745854c1ce63842dc.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 235, + 90, + 406, + 219 + ], + "blocks": [ + { + "bbox": [ + 236, + 72, + 273, + 82 + ], + "lines": [ + { + "bbox": [ + 236, + 72, + 273, + 82 + ], + "spans": [ + { + "bbox": [ + 236, + 72, + 273, + 82 + ], + "type": "text", + "content": "ShapeNet" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 235, + 90, + 406, + 219 + ], + "lines": [ + { + "bbox": [ + 235, + 90, + 406, + 219 + ], + "spans": [ + { + "bbox": [ + 235, + 90, + 406, + 219 + ], + "type": "image", + "image_path": "b75316214db6200a6ee1919761e93615b32977476914291829d738de2c7ab97d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 236, + 546, + 281 + ], + "lines": [ + { + "bbox": [ + 46, + 236, + 546, + 281 + ], + "spans": [ + { + "bbox": [ + 46, + 236, + 546, + 281 + ], + "type": "text", + "content": "Figure 2. Comparison between OBJVERSE and existing 3D object datasets. (Left:) Visual comparison of instances from OBJVERSE and ShapeNet for the categories of CAR, BED, VASE, and BAG. OBJVERSE instances are substantially more diverse since objects can come from many 3D content creation platforms, whereas ShapeNet models look more similar and all come from SketchUp, a 3D modeling platform built for simple architectural modeling. (Right:) Scale comparison table between existing 3D object datasets." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 419, + 83, + 544, + 227 + ], + "blocks": [ + { + "bbox": [ + 419, + 83, + 544, + 227 + ], + "lines": [ + { + "bbox": [ + 419, + 83, + 544, + 227 + ], + "spans": [ + { + "bbox": [ + 419, + 83, + 544, + 227 + ], + "type": "table", + "html": "
Dataset# Objects# Classes
YCB [5]775
BigBIRD [64]125-
KIT [33]219145
IKEA [40]21911
Pix3D [66]3959
GSO [18]1K17
EGAD [48]2K-
PhotoShape [50]5K1
ABO [11]8K63
3D-Future [20]10K34
ShapeNet [6]51K55
Objaverse818K21K
", + "image_path": "df511b3d9e9e2a89178ecd629235c88f23d63f6c0045d6b4d3ecbd7b2f946b99.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 291, + 289, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 291, + 289, + 579 + ], + "spans": [ + { + "bbox": [ + 46, + 291, + 289, + 579 + ], + "type": "text", + "content": "diversity, and realism. Ideally, models that reason about 3D objects should have access to datasets that meet these same criteria. However, of the numerous 3D object datasets that currently exist, none are able to excel in all three categories to the same degree as their 2D counterparts. Datasets such as KIT [33], YCB [5], BigBIRD [64], IKEA [40], and Pix3D [66] provide image-calibrated models over a diverse set of household objects, but severely lack in scale with only a few hundred objects at most. EGAD [48] procedurally generates 2K objects for grasping, but produces objects that are not that realistic or diverse. Slightly larger datasets of photo-realistic objects include GSO [18], PhotoShape [50], ABO [11] and 3D-Future [20], and ShapeNet [6] with object counts in the tens of thousands, see Fig. 2 for comparisons between OBJVERSE and these datasets. Datasets for CAD models, such as ModelNet [78] and DeepCAD [77], and ABC [35] do not include textures or materials, which limits their ability to represent objects that could plausibly be found in the real world. Datasets of scanned 3D objects and environments are valuable for real-world understanding [10, 13, 14, 38], but are quite small and limited. In addition to containing numerous artist designed objects, OBJVERSE contains many scanned assets, making it a useful source of data for learning from real-world distributions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 715 + ], + "type": "text", + "content": "While rapid progress has been made in developing datasets that combine image and text, in contrast, only a few datasets that pair language and 3D data exist. Text2Shape [7] released a dataset of 15,038 chairs and tables from ShapeNet each with around 5 text captions, giving 75,344 total text-shape pairs. ShapeGlot [1] released the CiC (Chairs in Context) dataset which contains 4,511 chairs from ShapeNet along with 78,789 descriptive utterances generated from a referential game. Due to the small scale and limited diversity of these datasets, current SoTA text-to-3D models [29,44,53] forgo the use of 3D datasets entirely" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 290, + 495, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 290, + 495, + 303 + ], + "spans": [ + { + "bbox": [ + 305, + 290, + 495, + 303 + ], + "type": "text", + "content": "and instead rely on 2D image-text supervision." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 308, + 373, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 308, + 373, + 323 + ], + "spans": [ + { + "bbox": [ + 306, + 308, + 373, + 323 + ], + "type": "text", + "content": "3. Objverse" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 329, + 546, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 329, + 546, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 329, + 546, + 449 + ], + "type": "text", + "content": "OBJAVERSE is a massive annotated 3D dataset that can be used to enable research in a wide range of areas across computer vision. The objects are sourced from Sketchfab, an online 3D marketplace where users can upload and share models for both free and commercial use. Objects selected for OBJVERSE have a distributable Creative Commons license and were obtained using Sketchfab's public API. Aside from licensing consideration, models marked as restricted due to objectionable or adult thematic content were excluded from the dataset." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 449, + 547, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 547, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 547, + 521 + ], + "type": "text", + "content": "Model metadata. OBJVERSE objects inherit a set of foundational annotations supplied by their creator when uploaded to Sketchfab. Figure 4 shows an example of the metadata available for each model. The metadata includes a name, assignments to a set of fixed categories, a set of unrestricted tags, and a natural language description." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 521, + 547, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 547, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 547, + 629 + ], + "type": "text", + "content": "OBJAVERSE-LVIS. While OBJAVERSE metadata contains a great deal of information about objects, Sketchfab's existing categorization scheme covers only 18 categories, too coarse for most applications. Object names, categories, and tags provide multiple potential categorizations at varying levels of specificity and with some inherent noise. However, for many existing computer vision tasks, it is useful to assign objects to a single category drawn from a predetermined set of the right size and level of semantic granularity." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "type": "text", + "content": "We choose the categories from the LVIS dataset [24] for categorizing a long-tail subset of objects in OBJVERSE. We construct a 47K LVIS categorized object subset, called OBJVERSE-LVIS, comprised of objects uniquely assigned to one of 1156 LVIS categories. We perform these assignments by first selecting 500 candidate objects per category using a combination of predictions from a CLIP classifi-" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13144" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 72, + 289, + 191 + ], + "blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 191 + ], + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 191 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 191 + ], + "type": "image", + "image_path": "efa256944f5ac1bfacdc0fd3b9c2eec90f9b3b63f6cc5f8aa15a004e42d6e850.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 84, + 193, + 252, + 202 + ], + "lines": [ + { + "bbox": [ + 84, + 193, + 252, + 202 + ], + "spans": [ + { + "bbox": [ + 84, + 193, + 252, + 202 + ], + "type": "text", + "content": "(a) Breakdown of high-level OBJVERSE categories" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 305, + 69, + 545, + 191 + ], + "blocks": [ + { + "bbox": [ + 305, + 69, + 545, + 191 + ], + "lines": [ + { + "bbox": [ + 305, + 69, + 545, + 191 + ], + "spans": [ + { + "bbox": [ + 305, + 69, + 545, + 191 + ], + "type": "image", + "image_path": "130ef3e48f4de85c2ef3d74081848d75aba43bbce8e716255dfc9f15a22e3bbe.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 351, + 193, + 499, + 202 + ], + "lines": [ + { + "bbox": [ + 351, + 193, + 499, + 202 + ], + "spans": [ + { + "bbox": [ + 351, + 193, + 499, + 202 + ], + "type": "text", + "content": "(b) Word cloud of OBJVERSE metadata tags." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 47, + 213, + 289, + 352 + ], + "blocks": [ + { + "bbox": [ + 47, + 213, + 289, + 352 + ], + "lines": [ + { + "bbox": [ + 47, + 213, + 289, + 352 + ], + "spans": [ + { + "bbox": [ + 47, + 213, + 289, + 352 + ], + "type": "image", + "image_path": "6cd2455fa6ddc613c8418b6c2bb78ba3dbe2f5546204618ce46ab2d428931a2d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 78, + 354, + 258, + 364 + ], + "lines": [ + { + "bbox": [ + 78, + 354, + 258, + 364 + ], + "spans": [ + { + "bbox": [ + 78, + 354, + 258, + 364 + ], + "type": "text", + "content": "(c) OBJAYERSE-LVIS category occurrence distribution." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 298, + 213, + 545, + 353 + ], + "blocks": [ + { + "bbox": [ + 298, + 213, + 545, + 353 + ], + "lines": [ + { + "bbox": [ + 298, + 213, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 298, + 213, + 545, + 353 + ], + "type": "image", + "image_path": "c4778b2d2da25378220938528fd057d369eb877a05541899a2ad194e445ce874.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 351, + 354, + 492, + 364 + ], + "lines": [ + { + "bbox": [ + 351, + 354, + 492, + 364 + ], + "spans": [ + { + "bbox": [ + 351, + 354, + 492, + 364 + ], + "type": "text", + "content": "(d) OBJAYERSE tag occurrence distribution." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 45, + 373, + 546, + 450 + ], + "lines": [ + { + "bbox": [ + 45, + 373, + 546, + 450 + ], + "spans": [ + { + "bbox": [ + 45, + 373, + 546, + 450 + ], + "type": "text", + "content": "Figure 3. ObjAverse statistics. (a) All 18 high-level categories present in ObjAverse's metadata with their corresponding number of occurrences. The relative share of most popular categories are evenly split, with a small number of less frequently categories. (b) A sample of several thousand popular object tags found in ObjAverse log-scaled by their frequency. (c) A histogram of fine-grained ObjAverse-LVIS categories with representative members from several bins highlighted. (d) A histogram of ObjAverse tags with representative members from several bins highlighted (note y-axis log scale). Tags from the low-occurrence side of the distribution correspond to unique objects that, taken individually, are rarely seen in the world. Frequently used tags like \"furniture\" and \"car\" reflect their real-world normalcy, but the high frequency of assets like \"sword\" diverge from their real-world counterparts." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 461, + 287, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 461, + 287, + 592 + ], + "spans": [ + { + "bbox": [ + 46, + 461, + 287, + 592 + ], + "type": "text", + "content": "cation model and candidates suggested by terms in their metadata. This combined pool contains objects visually resembling the target category (from the CLIP features of their thumbnail images) that might have missing metadata, as well as visually unusual instances of a category that are accurately named or tagged. These 250k candidate objects were then manually filtered and their assigned categories verified by crowdworkers. Since we only presented 500 object candidates per class, many popular categories, such as chair or car, have substantially more objects that could be included in OBJAYERSE-LVIS with future annotations." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": "Animated objects and rigged characters. OBJVERSE includes 44K animated objects and over 63K objects self-categorized as characters. Examples of animations include fridge doors opening, animals running, and the hands on a clock moving. Rigged characters can be set up for animation and rendering, and may often come annotated with bone mappings. The vast scale of animations available in OBJVERSE can support a wide range of research in temporal 3D learning, such as building text-based animation generative models [71], representing object changes over time" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 461, + 545, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 461, + 545, + 485 + ], + "spans": [ + { + "bbox": [ + 304, + 461, + 545, + 485 + ], + "type": "text", + "content": "with NERFs [51, 54], and temporal self-supervised learning via. future frame prediction [28, 82]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 486, + 545, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 545, + 581 + ], + "type": "text", + "content": "Articulated objects. Decomposing 3D objects into parts has led to a flurry of research in the past few years, including work in learning robotic grasping policies [79, 81], 3D semantic segmentation [47], and shape generation [46]. Since many objects in OBJAYERSE were uploaded by artists, the objects often come separated into parts. Figure 5 shows an example, where a chair is separated by its backrest, wheels, and legs, among many smaller parts." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 582, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 653 + ], + "type": "text", + "content": "Exteriors. Photogrammetry and NERF advances have enabled the commercialization of capturing high-quality 3D objects of large exteriors by taking pictures [70, 80]. In OBJVERSE, there are a large number of scanned buildings, cities, and stadiums. Figure 5 shows an example of a 3D object of NYC's skyline captured through a scan." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "OBJAVERSE-Interiors. There are " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "16\\mathrm{K}+" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": " interior scenes in OBJAVERSE, including houses, classrooms, and offices. The scenes often have multiple floors, many types of rooms, and are densely populated with objects from human input. Objects in the scenes are separable into parts, which allows" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13145" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 71, + 545, + 167 + ], + "blocks": [ + { + "bbox": [ + 51, + 71, + 545, + 167 + ], + "lines": [ + { + "bbox": [ + 51, + 71, + 545, + 167 + ], + "spans": [ + { + "bbox": [ + 51, + 71, + 545, + 167 + ], + "type": "image", + "image_path": "2ae6e510eb49b547d316e5c34919efdc85c3ce5acd8eb4f3806f824b58cb61e8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 174, + 547, + 198 + ], + "lines": [ + { + "bbox": [ + 46, + 174, + 547, + 198 + ], + "spans": [ + { + "bbox": [ + 46, + 174, + 547, + 198 + ], + "type": "text", + "content": "Figure 4. An example of metadata available for each object in OBJVERSE. Each uploaded object has a 3D model, user-selected rendered thumbnail image, name, description, tags, category, and stats, among additional metadata." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 207, + 288, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 207, + 288, + 268 + ], + "spans": [ + { + "bbox": [ + 46, + 207, + 288, + 268 + ], + "type": "text", + "content": "them to be usable for interactive robotics, embodied AI, and scene synthesis. To put the scale of OBJVERSE-Interiors in perspective, the number of scenes in OBJVERSE-Interiors is significantly larger than the 400 or so existing hand-built interactive embodied AI scenes [21,36,39,67]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 272, + 287, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 272, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 46, + 272, + 287, + 403 + ], + "type": "text", + "content": "Visual styles. Objects in the world can be constructed in many styles and often differ in style based on the time-period, geographic location, and artist's style. OBJVERSE objects cover a vast set of visual styles, including 3D scans, 3D modeled objects from virtually any platform, point clouds, and photo-realism via physically based rendering (PBR) [52]. Moreover, instances of objects often appear with many styles, which is critical for training and evaluating robust computer vision models [55]. Figure 5 shows examples of chairs in OBJVERSE in many different styles, including Gothic, modern, Victorian, cartoon, and abstract." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 408, + 288, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 408, + 288, + 504 + ], + "spans": [ + { + "bbox": [ + 46, + 408, + 288, + 504 + ], + "type": "text", + "content": "Statistics. OBJVERSE 1.0 includes 818K 3D objects, designed by 160K artists. There are " + }, + { + "bbox": [ + 46, + 408, + 288, + 504 + ], + "type": "inline_equation", + "content": ">2.35\\mathrm{M}" + }, + { + "bbox": [ + 46, + 408, + 288, + 504 + ], + "type": "text", + "content": " tags on the objects, with " + }, + { + "bbox": [ + 46, + 408, + 288, + 504 + ], + "type": "inline_equation", + "content": ">170\\mathrm{K}" + }, + { + "bbox": [ + 46, + 408, + 288, + 504 + ], + "type": "text", + "content": " of them being unique. We estimate that the objects have coverage for nearly 21K WordNet entities [45] (see appendix for details). Objects were uploaded between 2012 and 2022, with over 200K objects uploaded uploaded just in 2021. Figure 3 visualizes several statistics of the dataset, including the breakdown of objects into their" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 207, + 545, + 244 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 207, + 545, + 244 + ], + "spans": [ + { + "bbox": [ + 305, + 207, + 545, + 244 + ], + "type": "text", + "content": "self-assigned Sketchfab categories, a word cloud over the tags, a frequency plot of the tags, and the number of objects in OBJAYERSE-LVIS categories." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 255, + 386, + 269 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 255, + 386, + 269 + ], + "spans": [ + { + "bbox": [ + 306, + 255, + 386, + 269 + ], + "type": "text", + "content": "4. Applications" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 274, + 545, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 274, + 545, + 323 + ], + "spans": [ + { + "bbox": [ + 305, + 274, + 545, + 323 + ], + "type": "text", + "content": "In this section, we present 4 initial distinct applications of OBJVERSE, including 3D generative modeling, instance segmentation with CP3D, open-vocabulary ObjectNav, and analyze robustness in computer vision models." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 330, + 444, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 330, + 444, + 342 + ], + "spans": [ + { + "bbox": [ + 306, + 330, + 444, + 342 + ], + "type": "text", + "content": "4.1. 3D Generative Modeling" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 348, + 546, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 348, + 546, + 504 + ], + "spans": [ + { + "bbox": [ + 304, + 348, + 546, + 504 + ], + "type": "text", + "content": "3D generative modeling has shown much improvement recently with models such as GET3D [22] delivering impressive high quality results with rich geometric details. GET3D is trained to generate 3D textured meshes for a category and produces impressive 3D objects for categories like Car, Chair, and Motorcycle using data from ShapeNet [6]. OBJVERSE contains 3D models for many diverse categories including tail categories which are not represented in other datasets. It also contains diverse and realistic object instances per category. This scale and diversity can be used to train large vocabulary and high quality 3D generative models. In this work, we showcase the potential of this data as follows. We choose three categories of ob-" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 48, + 520, + 547, + 689 + ], + "blocks": [ + { + "bbox": [ + 48, + 520, + 547, + 689 + ], + "lines": [ + { + "bbox": [ + 48, + 520, + 547, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 520, + 547, + 689 + ], + "type": "image", + "image_path": "44c8b5c5e5a1e4701007798a436663cb732d5c01ad604d78470a7168ad1b4dcc.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 697, + 545, + 720 + ], + "lines": [ + { + "bbox": [ + 46, + 697, + 545, + 720 + ], + "spans": [ + { + "bbox": [ + 46, + 697, + 545, + 720 + ], + "type": "text", + "content": "Figure 5. Highlights of the visual diversity of objects that appear in OBJVERSE, including animated objects, rigged (body-part annotated) characters, models separatable into parts, exterior environments, interior environments, and a wide range visual styles." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13146" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 83, + 277, + 145 + ], + "blocks": [ + { + "bbox": [ + 56, + 68, + 143, + 77 + ], + "lines": [ + { + "bbox": [ + 56, + 68, + 143, + 77 + ], + "spans": [ + { + "bbox": [ + 56, + 68, + 143, + 77 + ], + "type": "text", + "content": "Objaverse Bag Generations" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 53, + 83, + 277, + 145 + ], + "lines": [ + { + "bbox": [ + 53, + 83, + 277, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 83, + 277, + 145 + ], + "type": "image", + "image_path": "1897e9edb06e5ebeb723e88e3d50a6f5e9c5c4a01c2229ebd026eaed4048aa37.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 59, + 167, + 275, + 225 + ], + "blocks": [ + { + "bbox": [ + 56, + 152, + 143, + 162 + ], + "lines": [ + { + "bbox": [ + 56, + 152, + 143, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 143, + 162 + ], + "type": "text", + "content": "ShapeNet Bag Generations" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 167, + 275, + 225 + ], + "lines": [ + { + "bbox": [ + 59, + 167, + 275, + 225 + ], + "spans": [ + { + "bbox": [ + 59, + 167, + 275, + 225 + ], + "type": "image", + "image_path": "76725f113584d350d32777b8cd96b334d519f4be3129402e97db4ef290d081d0.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 59, + 272, + 275, + 298 + ], + "blocks": [ + { + "bbox": [ + 58, + 235, + 276, + 245 + ], + "lines": [ + { + "bbox": [ + 58, + 235, + 276, + 245 + ], + "spans": [ + { + "bbox": [ + 58, + 235, + 276, + 245 + ], + "type": "text", + "content": "(a) Comparison of Bags generated with OBJVERSE and ShapeNet." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 56, + 255, + 114, + 264 + ], + "lines": [ + { + "bbox": [ + 56, + 255, + 114, + 264 + ], + "spans": [ + { + "bbox": [ + 56, + 255, + 114, + 264 + ], + "type": "text", + "content": "Shoe Generations" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 272, + 275, + 298 + ], + "lines": [ + { + "bbox": [ + 59, + 272, + 275, + 298 + ], + "spans": [ + { + "bbox": [ + 59, + 272, + 275, + 298 + ], + "type": "image", + "image_path": "7721da45e4abe5aad684934b546f37b24ac8a6160b7947af563dabef504b58ed.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 58, + 326, + 275, + 355 + ], + "blocks": [ + { + "bbox": [ + 56, + 312, + 129, + 321 + ], + "lines": [ + { + "bbox": [ + 56, + 312, + 129, + 321 + ], + "spans": [ + { + "bbox": [ + 56, + 312, + 129, + 321 + ], + "type": "text", + "content": "Fruit&Veg Generations" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 326, + 275, + 355 + ], + "lines": [ + { + "bbox": [ + 58, + 326, + 275, + 355 + ], + "spans": [ + { + "bbox": [ + 58, + 326, + 275, + 355 + ], + "type": "image", + "image_path": "e55973c2cfbc54edc9b1cb6eb6fb9683ffbd125d85de9f822d47a873b599e22e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 107, + 363, + 227, + 373 + ], + "lines": [ + { + "bbox": [ + 107, + 363, + 227, + 373 + ], + "spans": [ + { + "bbox": [ + 107, + 363, + 227, + 373 + ], + "type": "text", + "content": "(b) Shoe and Fruit&Veg. generations." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 56, + 399, + 275, + 422 + ], + "blocks": [ + { + "bbox": [ + 56, + 383, + 129, + 392 + ], + "lines": [ + { + "bbox": [ + 56, + 383, + 129, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 383, + 129, + 392 + ], + "type": "text", + "content": "Fruit&Veg Interpolation" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 56, + 399, + 275, + 422 + ], + "lines": [ + { + "bbox": [ + 56, + 399, + 275, + 422 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 275, + 422 + ], + "type": "image", + "image_path": "4206ebb7e776c70eb499cdbf2192ca5ffcf3fc6d15acdee0b7ed329ba07ff600.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 121, + 435, + 213, + 444 + ], + "lines": [ + { + "bbox": [ + 121, + 435, + 213, + 444 + ], + "spans": [ + { + "bbox": [ + 121, + 435, + 213, + 444 + ], + "type": "text", + "content": "(c) Fruit&Veg. interpolation." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 454, + 287, + 521 + ], + "lines": [ + { + "bbox": [ + 46, + 454, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 46, + 454, + 287, + 521 + ], + "type": "text", + "content": "Figure 6. (a) Example GET3D Bag object generations using OBJVERSE and ShapeNet models for training. (b) Additional Shoe and Fruit&Veg generations from OBJVERSE models. (c) models generated when interpolating between two, randomly sampled, latent encodings with our trained Fruit&Veg. model; what appears to be a pumpkin smoothly transforms into a mushroom." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 289, + 715 + ], + "type": "text", + "content": "jects, Shoe, Bag, and Fruit&Veg, and subsample objects from OBJVERSE to create three separate datasets containing, respectively, 143 shoes, 816 bags, and 571 fruits & vegetables (116 apples, 112 gourds, 92 mushrooms, 68 bananas, 52 oranges, 52 pears, 31 potatoes 24 lemons, and 24 pineapples). For comparison, we also train a GET3D model on the set of 83 bags from the ShapeNet dataset. Fig. 6 shows a collection of 3D objects generated by our trained GET3D models. Qualitatively, the 3D-meshes generated by the OBJVERSE-trained models are high-quality and diverse, especially when compared to the generations from the ShapeNet-trained model. To quantify this observation, we asked crowdworkers to rate the diversity of Bag generations produced by the OBJVERSE and ShapeNet trained models. When shown collections of nine randomly sampled" + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 312, + 63, + 543, + 211 + ], + "blocks": [ + { + "bbox": [ + 312, + 63, + 543, + 211 + ], + "lines": [ + { + "bbox": [ + 312, + 63, + 543, + 211 + ], + "spans": [ + { + "bbox": [ + 312, + 63, + 543, + 211 + ], + "type": "image", + "image_path": "2d568ef167cdfef8cbd5757de0bf5a3138b984dd8bd4e39ada1779261896e044.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 217, + 545, + 251 + ], + "lines": [ + { + "bbox": [ + 305, + 217, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 305, + 217, + 545, + 251 + ], + "type": "text", + "content": "Figure 7. An illustration of CP3D (copy-paste 3D) for segmentation augmentation. We render 3D objects from multiple views and paste them over LVIS training images." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 262, + 545, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 262, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 305, + 262, + 545, + 297 + ], + "type": "text", + "content": "generations from both models, workers rated the collection generated from the OBJVERSE trained model as more diverse in appearance " + }, + { + "bbox": [ + 305, + 262, + 545, + 297 + ], + "type": "inline_equation", + "content": "91\\%" + }, + { + "bbox": [ + 305, + 262, + 545, + 297 + ], + "type": "text", + "content": " of the time." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 298, + 545, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 298, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 305, + 298, + 545, + 335 + ], + "type": "text", + "content": "Our fruits and vegetables, composed of 9 varieties produces perhaps the highest quality output, a promising signal that can inspire future work in text-to-3D generation." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 342, + 489, + 356 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 342, + 489, + 356 + ], + "spans": [ + { + "bbox": [ + 306, + 342, + 489, + 356 + ], + "type": "text", + "content": "4.2. Instance Segmentation with CP3D" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 361, + 545, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 361, + 545, + 530 + ], + "spans": [ + { + "bbox": [ + 304, + 361, + 545, + 530 + ], + "type": "text", + "content": "A key advantage of using simulated data for computer vision is that it is much cheaper to obtain expert annotations. Annotated OBJVERSE objects can be rendered into images, allowing them to serve as a rich source of additional data that can be used to enhance model performance on 2D computer vision tasks. As a proof-of-concept demonstrating the effectiveness of this approach, we use segmented data from OBJVERSE objects as auxiliary labels for training models on the LVIS dataset for Large Scale Instance Segmentation [24]. The LVIS dataset contains instance segmentation masks for 1200 object categories that occur throughout a set of " + }, + { + "bbox": [ + 304, + 361, + 545, + 530 + ], + "type": "inline_equation", + "content": "164\\mathrm{k}" + }, + { + "bbox": [ + 304, + 361, + 545, + 530 + ], + "type": "text", + "content": " images. Recognition is especially challenging in this task due to the long tail of the object category distribution in this dataset. LVIS categories only con" + } + ] + } + ], + "index": 20 + }, + { + "type": "table", + "bbox": [ + 316, + 548, + 537, + 656 + ], + "blocks": [ + { + "bbox": [ + 316, + 548, + 537, + 656 + ], + "lines": [ + { + "bbox": [ + 316, + 548, + 537, + 656 + ], + "spans": [ + { + "bbox": [ + 316, + 548, + 537, + 656 + ], + "type": "table", + "html": "
MethodAPAPrAPcAPf
RFS [24]23.713.323.029.0
EQLv2 [69]25.517.724.330.2
LOCE [19]26.618.526.230.7
NorCal with RFS [49]25.219.324.229.0
Seesaw [73]26.419.526.129.7
GOL [3]27.721.427.730.4
GOL + CP3D28.321.828.331.1
", + "image_path": "2db9e376a2d7d27686fb04d8283faae22010d6d4e8ff38cbb71263fc052b24e2.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "table_body" + } + ], + "index": 21 + }, + { + "bbox": [ + 305, + 665, + 545, + 721 + ], + "lines": [ + { + "bbox": [ + 305, + 665, + 545, + 721 + ], + "spans": [ + { + "bbox": [ + 305, + 665, + 545, + 721 + ], + "type": "text", + "content": "Table 1. Comparison of our approach (GOL+CP3D) against SoTA Mask-RCNN ResNet-50 models on LVIS. We report results for APr, APc, and APf which measure AP for categories that are rare (appear in 1-10 images), common (appear in 11-100 images), and frequent (appear in " + }, + { + "bbox": [ + 305, + 665, + 545, + 721 + ], + "type": "inline_equation", + "content": ">100" + }, + { + "bbox": [ + 305, + 665, + 545, + 721 + ], + "type": "text", + "content": " images), respectively" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13147" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 96, + 70, + 293, + 194 + ], + "blocks": [ + { + "bbox": [ + 96, + 70, + 293, + 194 + ], + "lines": [ + { + "bbox": [ + 96, + 70, + 293, + 194 + ], + "spans": [ + { + "bbox": [ + 96, + 70, + 293, + 194 + ], + "type": "image", + "image_path": "4fa6782d8fa095cdf09a093835f3cede63e70411b507d47c3530b45458562905.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 50, + 196, + 541, + 208 + ], + "lines": [ + { + "bbox": [ + 50, + 196, + 541, + 208 + ], + "spans": [ + { + "bbox": [ + 50, + 196, + 541, + 208 + ], + "type": "text", + "content": "Figure 8. An existing ProcTHOR scene (left) and a semantically similar ProcTHOR generatable scene with OBJVERSE objects (right)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 296, + 70, + 496, + 194 + ], + "blocks": [ + { + "bbox": [ + 296, + 70, + 496, + 194 + ], + "lines": [ + { + "bbox": [ + 296, + 70, + 496, + 194 + ], + "spans": [ + { + "bbox": [ + 296, + 70, + 496, + 194 + ], + "type": "image", + "image_path": "096eec98402182a571df44cf631a5a93dbb7f7defbf4a80fb36a5536d8c1582b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 211, + 287, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 211, + 287, + 247 + ], + "spans": [ + { + "bbox": [ + 46, + 211, + 287, + 247 + ], + "type": "text", + "content": "tain an average 9 instances across the dataset, so training on simulated data is a promising approach for overcoming the challenges of learning in this low-sample regime." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 247, + 289, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 247, + 289, + 452 + ], + "spans": [ + { + "bbox": [ + 46, + 247, + 289, + 452 + ], + "type": "text", + "content": "Using the LVIS-annotated subset of ObjAVERSE, we introduce CP3D: an enhancement to the simple, but effective, copy-and-paste technique of [23]. Figure 7 shows an example of the setup for CP3D. Here, we render different views of 3D objects and paste them on-top of existing LVIS images. We render 5 distinct views of each object and cache them for use throughout training. During training, an image is selected for the copy-paste augmentation with 0.5 probability, and once selected, 1-3 images of randomly chosen LVIS-annotated ObjAVERSE objects are pasted onto the selected training image. The segmentation masks of the selected objects are added to the training image's annotation as well. Object images and masks are randomly scaled and translated before being pasted. We use this strategy to finetune the pretrained ResNet-50 Mask-RCNN [25, 26] of [3]. As shown in Tab.1, simply finetuning this model for 24 epochs yields performance gains across several metrics." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 457, + 206, + 470 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 457, + 206, + 470 + ], + "spans": [ + { + "bbox": [ + 47, + 457, + 206, + 470 + ], + "type": "text", + "content": "4.3. Open-Vocabulary ObjectNav" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 473, + 287, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 473, + 287, + 676 + ], + "spans": [ + { + "bbox": [ + 46, + 473, + 287, + 676 + ], + "type": "text", + "content": "In this section, we introduce open-vocabulary ObjectNav, a new task propelled by the vast diversity of objects that appear in OBJAYERSE. Here, an agent is placed at a random starting location inside of a home and tasked to navigate to a target object provided from a text description (e.g. \"Raspberry Pi Pico\"). To facilitate this task, we procedurally generate 10K new homes in ProcTHOR [16] fully populated with objects from OBJAYERSE-LVIS. Until now, ObjectNav tasks have focused on training agents to navigate to 20 or so target objects provided their category label [15, 16, 58], and existing interactive embodied AI simulations, including ProcTHOR, only include around 2K total objects across around 100 object types [16, 39, 67]. In this work, we take a large step to massively scale the number of target objects used in ObjectNav (" + }, + { + "bbox": [ + 46, + 473, + 287, + 676 + ], + "type": "inline_equation", + "content": "20 \\to \\text{OpenVocab}" + }, + { + "bbox": [ + 46, + 473, + 287, + 676 + ], + "type": "text", + "content": "), the number of objects available in simulation (" + }, + { + "bbox": [ + 46, + 473, + 287, + 676 + ], + "type": "inline_equation", + "content": "2\\mathbf{K} \\to 36\\mathbf{K}" + }, + { + "bbox": [ + 46, + 473, + 287, + 676 + ], + "type": "text", + "content": "), and the number of object types of the objects (" + }, + { + "bbox": [ + 46, + 473, + 287, + 676 + ], + "type": "inline_equation", + "content": "100 \\to 1.1\\mathbf{K}" + }, + { + "bbox": [ + 46, + 473, + 287, + 676 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "content": "Object placement. To make the placement of objects in the houses more natural, we use the OBJVERSE-LVIS subset and annotate placement constraints for each object" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 211, + 545, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 211, + 545, + 378 + ], + "spans": [ + { + "bbox": [ + 304, + 211, + 545, + 378 + ], + "type": "text", + "content": "category. Specifically, we annotate if objects of a given category typically appears on the floor, on-top of a surface, or on a wall. If instances of the object category may appear on the floor, we also annotate whether it may appear in the middle of the scene (e.g. a clutter object like a basketball) or on the edge of the scene (e.g. a toilet or a fridge). For objects placed on the floor, we also to automatically detect flat regions on top of the object's mesh to place surface object types. The annotations are used by ProcTHOR for sampling objects to place in a scene. We also filter out OBJAYERSE-LVIS objects that do not appear inside of homes, such as a jet plane. Structural objects, like doors and windows, are inherited from ProcTHOR as they would require additional cleanup." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 379, + 546, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 379, + 546, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 379, + 546, + 510 + ], + "type": "text", + "content": "Object size correction. Objects in Sketchfab may be uploaded at unnatural scales (e.g. a plant being as large as a tower). We therefore scale the objects to be of a reasonable size for them to look natural in a house. Here, for each object category, we annotate the maximum bounding box dimension length that every instance of the object category should be scaled to. For example, we annotate the maximum bounding box dimension for bookcase to be 2 meters and fork to be 0.18 meters. If a 3D modeled bookcase then has a bounding box of " + }, + { + "bbox": [ + 304, + 379, + 546, + 510 + ], + "type": "inline_equation", + "content": "20\\mathrm{m}\\times 6\\mathrm{m}\\times 3\\mathrm{m}" + }, + { + "bbox": [ + 304, + 379, + 546, + 510 + ], + "type": "text", + "content": ", we shrink each side by a factor of " + }, + { + "bbox": [ + 304, + 379, + 546, + 510 + ], + "type": "inline_equation", + "content": "\\max (20,6,3) / 2 = 5" + }, + { + "bbox": [ + 304, + 379, + 546, + 510 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 510, + 545, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 629 + ], + "type": "text", + "content": "Preprocessing for AI2-THOR. We add support to AI2-THOR for loading objects on the fly at runtime. Previously, all objects had to be stored in a Unity build, but such an approach is impractical when working with orders of magnitude more object data. For each object, we compress it with Blender [12] by joining all of its meshes together, decimate the joined mesh such that it has at most 5K vertices, and bake all the UV texture maps into a single texture map. We then generate colliders using V-HACD [43] to support rigid-body interactions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "Approach. Given procedural houses populated with OBJVERSE-LVIS, the task is to navigate to the proximity of a chosen target object and invoke a task-completion action when the target object is in sight, given an open-vocabulary description formed with the template \"a {name} {category}\" The name is the object name given by its creator, which is often descriptive. We filter" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13148" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 105, + 127 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 105, + 127 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 105, + 127 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 105, + 127 + ], + "type": "image", + "image_path": "ce79a2e0bd4d79922e4ca0ff9aa7070a8f2cc0e617560e82d4ca0acf156c29d9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 51, + 129, + 102, + 135 + ], + "lines": [ + { + "bbox": [ + 51, + 129, + 102, + 135 + ], + "spans": [ + { + "bbox": [ + 51, + 129, + 102, + 135 + ], + "type": "text", + "content": "Television 99.6%" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 110, + 70, + 166, + 127 + ], + "blocks": [ + { + "bbox": [ + 110, + 70, + 166, + 127 + ], + "lines": [ + { + "bbox": [ + 110, + 70, + 166, + 127 + ], + "spans": [ + { + "bbox": [ + 110, + 70, + 166, + 127 + ], + "type": "image", + "image_path": "82cf596b8e34decf4cfac9cc4bdf62e0bf7756ca0da27d19d6853a58c9194d2a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 112, + 129, + 163, + 135 + ], + "lines": [ + { + "bbox": [ + 112, + 129, + 163, + 135 + ], + "spans": [ + { + "bbox": [ + 112, + 129, + 163, + 135 + ], + "type": "text", + "content": "Television 96.1%" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 170, + 70, + 227, + 127 + ], + "blocks": [ + { + "bbox": [ + 170, + 70, + 227, + 127 + ], + "lines": [ + { + "bbox": [ + 170, + 70, + 227, + 127 + ], + "spans": [ + { + "bbox": [ + 170, + 70, + 227, + 127 + ], + "type": "image", + "image_path": "bc6c4c147c226e984ff05e3fc2dcf17cc7a5cffc713823bd2ce8508a2a253844.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 178, + 129, + 216, + 135 + ], + "lines": [ + { + "bbox": [ + 178, + 129, + 216, + 135 + ], + "spans": [ + { + "bbox": [ + 178, + 129, + 216, + 135 + ], + "type": "text", + "content": "Vase 39.0%" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 229, + 70, + 287, + 127 + ], + "blocks": [ + { + "bbox": [ + 229, + 70, + 287, + 127 + ], + "lines": [ + { + "bbox": [ + 229, + 70, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 229, + 70, + 287, + 127 + ], + "type": "image", + "image_path": "3823427c5bc8dae555d3db806c11590db8b0e0a463eefddaca8f23e9ac685ba6.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 233, + 129, + 284, + 135 + ], + "lines": [ + { + "bbox": [ + 233, + 129, + 284, + 135 + ], + "spans": [ + { + "bbox": [ + 233, + 129, + 284, + 135 + ], + "type": "text", + "content": "Flowerpot 34.9%" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 49, + 140, + 105, + 197 + ], + "blocks": [ + { + "bbox": [ + 49, + 140, + 105, + 197 + ], + "lines": [ + { + "bbox": [ + 49, + 140, + 105, + 197 + ], + "spans": [ + { + "bbox": [ + 49, + 140, + 105, + 197 + ], + "type": "image", + "image_path": "0822e0addd82a3b04016ec6b6ac45f00366446491301187cbb71ea99c4e8b1f0.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 57, + 198, + 97, + 205 + ], + "lines": [ + { + "bbox": [ + 57, + 198, + 97, + 205 + ], + "spans": [ + { + "bbox": [ + 57, + 198, + 97, + 205 + ], + "type": "text", + "content": "Clock " + }, + { + "bbox": [ + 57, + 198, + 97, + 205 + ], + "type": "inline_equation", + "content": "99.8\\%" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 110, + 140, + 166, + 197 + ], + "blocks": [ + { + "bbox": [ + 110, + 140, + 166, + 197 + ], + "lines": [ + { + "bbox": [ + 110, + 140, + 166, + 197 + ], + "spans": [ + { + "bbox": [ + 110, + 140, + 166, + 197 + ], + "type": "image", + "image_path": "32954a33ab0101bec606b19c6b1c7c2e0aa942305dabb95bb800f2f3ecba188f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 118, + 199, + 157, + 205 + ], + "lines": [ + { + "bbox": [ + 118, + 199, + 157, + 205 + ], + "spans": [ + { + "bbox": [ + 118, + 199, + 157, + 205 + ], + "type": "text", + "content": "Clock 98.2%" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 170, + 140, + 227, + 197 + ], + "blocks": [ + { + "bbox": [ + 170, + 140, + 227, + 197 + ], + "lines": [ + { + "bbox": [ + 170, + 140, + 227, + 197 + ], + "spans": [ + { + "bbox": [ + 170, + 140, + 227, + 197 + ], + "type": "image", + "image_path": "24c7d257474f5c15ac3b633adeaefbfdf6ecbf4d829c9b8706e6a24236f71bbb.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 172, + 198, + 223, + 205 + ], + "lines": [ + { + "bbox": [ + 172, + 198, + 223, + 205 + ], + "spans": [ + { + "bbox": [ + 172, + 198, + 223, + 205 + ], + "type": "text", + "content": "Flowerpot 32.4%" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 229, + 140, + 287, + 197 + ], + "blocks": [ + { + "bbox": [ + 229, + 140, + 287, + 197 + ], + "lines": [ + { + "bbox": [ + 229, + 140, + 287, + 197 + ], + "spans": [ + { + "bbox": [ + 229, + 140, + 287, + 197 + ], + "type": "image", + "image_path": "7ae741bf261e216b98d703f1aaaf74e83979a66e44b237c86e4755ee6eec8595.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 239, + 198, + 277, + 205 + ], + "lines": [ + { + "bbox": [ + 239, + 198, + 277, + 205 + ], + "spans": [ + { + "bbox": [ + 239, + 198, + 277, + 205 + ], + "type": "text", + "content": "Vase 33.2%" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 47, + 217, + 286, + 239 + ], + "lines": [ + { + "bbox": [ + 47, + 217, + 286, + 239 + ], + "spans": [ + { + "bbox": [ + 47, + 217, + 286, + 239 + ], + "type": "text", + "content": "Figure 9. Examples of objects rendered from random orientations and their 0-shot classification categories with the CLIP ViT-B/32." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 251, + 287, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 251, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 287, + 430 + ], + "type": "text", + "content": "each by whether it is detected as being written in English by a language detector [31,32], and fall back to a class-only description for non-English name. Examples of the possible expressions include “a victorian-monobike motorcycle”, “a unicorn pony”, or “a dino ghost lizard”. The agent, similar to the ones in [34], observes an RGB egocentric view of the environment, pre-processed by the visual branch of a frozen ResNet-50 CLIP model [55] – the target description is pre-processed by the corresponding text branch. We train the agent with DD-PPO [76] and evaluate on houses with floor plans, objects, and descriptions unseen in training. We use the AllenAct [75] framework to train our agent. Our trained agent achieves a success rate of " + }, + { + "bbox": [ + 46, + 251, + 287, + 430 + ], + "type": "inline_equation", + "content": "19.9\\%" + }, + { + "bbox": [ + 46, + 251, + 287, + 430 + ], + "type": "text", + "content": ", for a random policy success of " + }, + { + "bbox": [ + 46, + 251, + 287, + 430 + ], + "type": "inline_equation", + "content": "5.1\\%" + }, + { + "bbox": [ + 46, + 251, + 287, + 430 + ], + "type": "text", + "content": ". For more details about the experiment refer to the appendix." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 47, + 435, + 173, + 448 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 435, + 173, + 448 + ], + "spans": [ + { + "bbox": [ + 47, + 435, + 173, + 448 + ], + "type": "text", + "content": "4.4. Analyzing Robustness" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 46, + 450, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 450, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 450, + 287, + 605 + ], + "type": "text", + "content": "A persistent bias present in many image datasets, e.g. ImageNet [60], is that the subjects of interest are generally photographed from a forward-facing, canonical, orientation. When, for example, taking a photograph of a television, few would choose to take this photograph crouched on the floor behind the television. This impact of this bias was studied by Alcorn et al. [2] who find that modern computer vision systems are highly susceptible to deviations from canonical poses. This is more than a theoretical problem: computer vision systems deployed in the real world will frequently encounter objects in non-canonical orientations and in many applications, e.g. autonomous driving, it will be safety critical that they behave well." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": "Given the above, we adopt the experimental design of Alcorn et al. and design, using OBJVERSE assets, a benchmark for evaluating the robustness of state-of-the-art computer vision classification models to orientation shifts. In particular, for each object in our OBJVERSE-LVIS subset, we render 12 images of the object from random orientations rendered upon a background with RGB values equalling the mean RGB values from ImageNet; see Fig. 9 for examples. This ability to, at scale, render objects from random view" + } + ] + } + ], + "index": 20 + }, + { + "type": "table", + "bbox": [ + 309, + 70, + 544, + 215 + ], + "blocks": [ + { + "bbox": [ + 309, + 70, + 544, + 215 + ], + "lines": [ + { + "bbox": [ + 309, + 70, + 544, + 215 + ], + "spans": [ + { + "bbox": [ + 309, + 70, + 544, + 215 + ], + "type": "table", + "html": "
ModelRandom RotationAny RotationΔTop-1
Top-1Top-5Top-1Top-5
OpenAI-400M [55]
RN5021.4%45.0%43.9%70.8%22.5%
ViT-L/1429.1%54.5%52.3%77.2%23.2%
LAION-400M [62]
ViT-B/3224.1%48.5%46.9%74.2%22.8%
ViT-L/1430.6%56.8%50.5%77.0%19.9%
LAION-2B [61]
ViT-B/3227.0%51.8%50.3%76.1%23.3%
ViT-L/1432.9%59.2%52.1%78.0%19.2%
ViT-H/1432.3%58.8%50.1%77.3%17.8%
", + "image_path": "57d8679f1191c16454ebcc916eb23bf0d71c95bd62220f1230d05fe2e031ecb1.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "table_body" + } + ], + "index": 21 + }, + { + "bbox": [ + 305, + 224, + 545, + 268 + ], + "lines": [ + { + "bbox": [ + 305, + 224, + 545, + 268 + ], + "spans": [ + { + "bbox": [ + 305, + 224, + 545, + 268 + ], + "type": "text", + "content": "Table 2. Evaluating 0-shot CLIP classification models on our rotational robustness benchmark. " + }, + { + "bbox": [ + 305, + 224, + 545, + 268 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 305, + 224, + 545, + 268 + ], + "type": "text", + "content": "Top-1 denotes the difference between Top-1 Any Rotation and Top-1 Random Rotation. Models are strongly overfit to standard views of objects." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 280, + 545, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 280, + 545, + 339 + ], + "spans": [ + { + "bbox": [ + 304, + 280, + 545, + 339 + ], + "type": "text", + "content": "points is a practical impossibility in the real world but is made trivial when using 3D assets. We then evaluate several modern open-domain image-classification networks (constrained to the " + }, + { + "bbox": [ + 304, + 280, + 545, + 339 + ], + "type": "inline_equation", + "content": "\\approx 1,200" + }, + { + "bbox": [ + 304, + 280, + 545, + 339 + ], + "type": "text", + "content": " LVIS categories) on these images and report 4 metrics for each model. These metrics include:" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 341, + 545, + 411 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 306, + 341, + 545, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 341, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 306, + 341, + 545, + 376 + ], + "type": "text", + "content": "- Top-1 Random Rotation - the frequency with which a model correctly classifies an image as belonging to the respective LVIS category." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 306, + 376, + 545, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 376, + 545, + 411 + ], + "spans": [ + { + "bbox": [ + 306, + 376, + 545, + 411 + ], + "type": "text", + "content": "- Top-1 Any Rotation - the frequency with which a model classifies an image correctly from at least one of the 12 random orientations." + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 304, + 413, + 545, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 413, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 413, + 545, + 533 + ], + "type": "text", + "content": "This second metric is diagnostic and serves to represent a model's performance when shown an object from a canonical pose. We also have Top-5 variants of the above metric where the correct category need only be in the top 5 predictions from the model. We report our results in Tab. 2 in which we evaluate a variety of performant pretrained models. Comparing the gap in performance between the Top-k Random Rotation and Top-k Any Rotation metrics we find that model performance dramatically degrades when viewing objects from unusual orientations." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 306, + 541, + 378, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 541, + 378, + 553 + ], + "spans": [ + { + "bbox": [ + 306, + 541, + 378, + 553 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "content": "We present OBJAYERSE, a next-generation 3D asset library containing 818K high-quality, diverse, 3D models with paired text descriptions, titles, and tags. As a small glimpse of the potential uses of OBJAYERSE, we present four experimental studies showing how OBJAYERSE can be used to power (1) generative 3D models with clear future applications to text-to-3D generation, (2) improvements to classical computer vision tasks such as instance segmentation, (3) the creation of novel embodied AI tasks like Open Vocabulary Object Navigation, and (4) quantifying the rotational robustness of vision models on renderings of objects. We hope to see OBJAYERSE enable a new universe of new applications for computer vision." + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13149" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 92, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 53, + 92, + 287, + 146 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 92, + 287, + 146 + ], + "spans": [ + { + "bbox": [ + 53, + 92, + 287, + 146 + ], + "type": "text", + "content": "[1] Panos Achlioptas, Judy Fan, Robert Hawkins, Noah Goodman, and Leonidas J Guibas. Shapeplot: Learning language for shape differentiation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8938-8947, 2019. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 148, + 288, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 288, + 224 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 288, + 224 + ], + "type": "text", + "content": "[2] Michael A. Alcorn, Qi Li, Zhitao Gong, Chengfei Wang, Long Mai, Wei-Shinn Ku, and Anh Nguyen. Strike (with) a pose: Neural networks are easily fooled by strange poses of familiar objects. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, pages 4845-4854. Computer Vision Foundation / IEEE, 2019. 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 227, + 288, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 227, + 288, + 270 + ], + "spans": [ + { + "bbox": [ + 53, + 227, + 288, + 270 + ], + "type": "text", + "content": "[3] Konstantinos Panagiotis Alexandridis, Jiankang Deng, Anh Nguyen, and Shan Luo. Long-tailed instance segmentation using gumbel optimized loss. arXiv preprint arXiv:2207.10936, 2022. 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 273, + 287, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 273, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 53, + 273, + 287, + 327 + ], + "type": "text", + "content": "[4] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 329, + 287, + 383 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 329, + 287, + 383 + ], + "spans": [ + { + "bbox": [ + 53, + 329, + 287, + 383 + ], + "type": "text", + "content": "[5] Berk Calli, Aaron Walsman, Arjun Singh, Siddhartha Srinivasa, Pieter Abbeel, and Aaron M Dollar. Benchmarking in manipulation research: The ycb object and model set and benchmarking protocols. arXiv preprint arXiv:1502.03143, 2015. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 385, + 287, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 385, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 53, + 385, + 287, + 441 + ], + "type": "text", + "content": "[6] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 3, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 442, + 287, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 442, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 53, + 442, + 287, + 498 + ], + "type": "text", + "content": "[7] Kevin Chen, Christopher B Choy, Manolis Savva, Angel X Chang, Thomas Funkhouser, and Silvio Savarese. Text2shape: Generating shapes from natural language by learning joint embeddings. In Asian conference on computer vision, pages 100-116. Springer, 2018. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 499, + 287, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 499, + 287, + 553 + ], + "spans": [ + { + "bbox": [ + 53, + 499, + 287, + 553 + ], + "type": "text", + "content": "[8] Xi Chen, Xiao Wang, Soravit Changpinyo, AJ Piergiovanni, Piotr Padlewski, Daniel Salz, Sebastian Goodman, Adam Grycner, Basil Mustafa, Lucas Beyer, et al. Pali: A jointly-scaled multilingual language-image model. arXiv preprint arXiv:2209.06794, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 555, + 287, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 555, + 287, + 599 + ], + "spans": [ + { + "bbox": [ + 53, + 555, + 287, + 599 + ], + "type": "text", + "content": "[9] Jaemin Cho, Jie Lei, Hao Tan, and Mohit Bansal. Unifying vision-and-language tasks via text generation. In International Conference on Machine Learning, pages 1931-1942. PMLR, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 601, + 287, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 287, + 634 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 287, + 634 + ], + "type": "text", + "content": "[10] Sungjoon Choi, Qian-Yi Zhou, Stephen Miller, and Vladlen Koltun. A large dataset of object scans. arXiv preprint arXiv:1602.02481, 2016. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 636, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 287, + 712 + ], + "type": "text", + "content": "[11] Jasmine Collins, Shubham Goel, Kenan Deng, Achleshwar Luthra, Leon Xu, Erhan Gundogdu, Xi Zhang, Tomas F Yago Vicente, Thomas Dideriksien, Himanshu Arora, et al. Abo: Dataset and benchmarks for real-world 3d object understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21126-21136, 2022. 3" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 105 + ], + "type": "text", + "content": "[12] Blender Online Community. Blender - a 3d modelling and rendering package. http://www.blender.org, 2018.7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 106, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 106, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 106, + 545, + 139 + ], + "type": "text", + "content": "[13] Camille Couprie, Clément Farabet, Laurent Najman, and Yann LeCun. Indoor semantic segmentation using depth information. arXiv preprint arXiv:1301.3572, 2013. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 140, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 140, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 308, + 140, + 545, + 194 + ], + "type": "text", + "content": "[14] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 195, + 545, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 195, + 545, + 248 + ], + "spans": [ + { + "bbox": [ + 308, + 195, + 545, + 248 + ], + "type": "text", + "content": "[15] Matt Deitke, Dhruv Batra, Yonatan Bisk, Tommaso Campari, Angel X Chang, Devendra Singh Chaplot, Changan Chen, Claudia Pérez D'Arpino, Kiana Ehsani, Ali Farhadi, et al. Retrospectives on the embodied ai workshop. arXiv preprint arXiv:2210.06849, 2022. 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 250, + 545, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 250, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 308, + 250, + 545, + 304 + ], + "type": "text", + "content": "[16] Matt Deitke, Eli VanderBilt, Alvaro Herrasti, Luca Weihs, Jordi Salvador, Kiana Ehsani, Winson Han, Eric Kolve, Ali Farhadi, Aniruddha Kembhavi, et al. Procthor: Large-scale embodied ai using procedural generation. Conference on Neural Information Processing Systems, 2022. 2, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 304, + 545, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 304, + 545, + 349 + ], + "spans": [ + { + "bbox": [ + 308, + 304, + 545, + 349 + ], + "type": "text", + "content": "[17] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 350, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 350, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 308, + 350, + 545, + 403 + ], + "type": "text", + "content": "[18] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas B McHugh, and Vincent Vanhoucke. Google scanned objects: A high-quality dataset of 3d scanned household items. arXiv preprint arXiv:2204.11918, 2022. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 404, + 545, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 404, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 308, + 404, + 545, + 448 + ], + "type": "text", + "content": "[19] Chengjian Feng, Yujie Zhong, and Weilin Huang. Exploring classification equilibrium in long-tailed object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3417-3426, 2021. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 449, + 545, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 449, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 308, + 449, + 545, + 491 + ], + "type": "text", + "content": "[20] Huan Fu, Rongfei Jia, Lin Gao, Mingming Gong, Binqiang Zhao, Steve Maybank, and Dacheng Tao. 3d-future: 3d furniture shape with texture. International Journal of Computer Vision, 129(12):3313-3337, 2021. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 492, + 545, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 492, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 308, + 492, + 545, + 547 + ], + "type": "text", + "content": "[21] Chuang Gan, Jeremy Schwartz, Seth Alter, Martin Schrimpf, James Traer, Julian De Freitas, Jonas Kubilius, Abhishek Bhandwaldar, Nick Haber, Megumi Sano, et al. Threadworld: A platform for interactive multi-modal physical simulation. arXiv preprint arXiv:2007.04954, 2020. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 548, + 545, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 548, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 308, + 548, + 545, + 602 + ], + "type": "text", + "content": "[22] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. In Advances In Neural Information Processing Systems, 2022. 2, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 603, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 603, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 308, + 603, + 545, + 668 + ], + "type": "text", + "content": "[23] Golnaz Ghiasi, Yin Cui, Aravind Srinivas, Rui Qian, Tsung-Yi Lin, Ekin D Cubuk, Quoc V Le, and Barret Zoph. Simple copy-paste is a strong data augmentation method for instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2918-2928, 2021. 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 669, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 669, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 669, + 545, + 713 + ], + "type": "text", + "content": "[24] Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5356-5364, 2019. 3, 6" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13150" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[25] Kaiming He, Georgia Gkioxari, Piotr Dollár, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 287, + 150 + ], + "type": "text", + "content": "[26] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 150, + 287, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 150, + 287, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 150, + 287, + 205 + ], + "type": "text", + "content": "[27] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. OpenCLIP, July 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 205, + 287, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 205, + 287, + 248 + ], + "spans": [ + { + "bbox": [ + 48, + 205, + 287, + 248 + ], + "type": "text", + "content": "[28] Allan Jabri, Andrew Owens, and Alexei Efros. Space-time correspondence as a contrastive random walk. Advances in neural information processing systems, 33:19545-19560, 2020. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 250, + 287, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 250, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 48, + 250, + 287, + 304 + ], + "type": "text", + "content": "[29] Ajay Jain, Ben Mildenhall, Jonathan T Barron, Pieter Abbeel, and Ben Poole. Zero-shot text-guided object generation with dream fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 867-876, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 304, + 287, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 304, + 287, + 369 + ], + "spans": [ + { + "bbox": [ + 48, + 304, + 287, + 369 + ], + "type": "text", + "content": "[30] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, pages 4904-4916. PMLR, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 371, + 287, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 371, + 287, + 414 + ], + "spans": [ + { + "bbox": [ + 48, + 371, + 287, + 414 + ], + "type": "text", + "content": "[31] Armand Joulin, Edouard Grave, Piotr Bojanowski, Matthijs Douze, Hérve Jégou, and Tomas Mikolov. Fasttext.zip: Compressing text classification models. arXiv preprint arXiv:1612.03651, 2016. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 415, + 287, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 287, + 448 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 287, + 448 + ], + "type": "text", + "content": "[32] Armand Joulin, Edouard Grave, Piotr Bojanowski, and Tomas Mikolov. Bag of tricks for efficient text classification. arXiv preprint arXiv:1607.01759, 2016. 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 449, + 287, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 449, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 48, + 449, + 287, + 502 + ], + "type": "text", + "content": "[33] Alexander Kasper, Zhixing Xue, and Rüdiger Dillmann. The kit object models database: An object model database for object recognition, localization and manipulation in service robotics. The International Journal of Robotics Research, 31(8):927-934, 2012. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 503, + 287, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 503, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 48, + 503, + 287, + 557 + ], + "type": "text", + "content": "[34] Apoorv Khandelwal, Luca Weihs, Roozbeh Mottaghi, and Aniruddha Kembhavi. Simple but effective: Clip embeddings for embodied ai. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14809-14818, 2022. 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 559, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 287, + 624 + ], + "type": "text", + "content": "[35] Sebastian Koch, Albert Matveev, Zhongshi Jiang, Francis Williams, Alexey Artemov, Evgeny Burnaev, Marc Alexa, Denis Zorin, and Daniele Panozzo. Abc: A big cad model dataset for geometric deep learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9601-9611, 2019. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 625, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 287, + 689 + ], + "type": "text", + "content": "[36] Eric Kolve, Roozbeh Mottaghi, Winson Han, Eli VanderBilt, Luca Weihs, Alvaro Herrasti, Matt Deitke, Kiana Ehsani, Daniel Gordon, Yuke Zhu, Aniruddha Kembhavi, Abhinav Kumar Gupta, and Ali Farhadi. Ai2-thor: An interactive 3d environment for visual ai. arXiv e-prints, pages arXiv-1712, 2017. 2, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 690, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 690, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 690, + 287, + 714 + ], + "type": "text", + "content": "[37] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "text", + "content": "Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4. International Journal of Computer Vision, 128(7):1956-1981, 2020. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 106, + 545, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 106, + 545, + 182 + ], + "spans": [ + { + "bbox": [ + 308, + 106, + 545, + 182 + ], + "type": "text", + "content": "[38] Marc Levoy, Kari Pulli, Brian Curless, Szymon Rusinkiewicz, David Koller, Lucas Pereira, Matt Ginzton, Sean Anderson, James Davis, Jeremy Ginsberg, et al. The digital michelangelo project: 3d scanning of large statues. In Proceedings of the 27th annual conference on Computer graphics and interactive techniques, pages 131-144, 2000. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 183, + 545, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 183, + 545, + 248 + ], + "spans": [ + { + "bbox": [ + 308, + 183, + 545, + 248 + ], + "type": "text", + "content": "[39] Chengshu Li, Fei Xia, Roberto Martín-Martín, Michael Lingelbach, Sanjana Srivastava, Bokui Shen, Kent Vainio, Cem Gokmen, Gokul Dharan, Tanish Jain, et al. igibson 2.0: Object-centric simulation for robot learning of everyday household tasks. arXiv preprint arXiv:2108.03272, 2021. 2, 5, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 249, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 249, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 308, + 249, + 545, + 293 + ], + "type": "text", + "content": "[40] Joseph J Lim, Hamed Pirsiavash, and Antonio Torralba. Parsing IKEA objects: Fine pose estimation. In Proceedings of the IEEE international conference on computer vision, pages 2992-2999, 2013. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 293, + 545, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 293, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 308, + 293, + 545, + 348 + ], + "type": "text", + "content": "[41] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 349, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 349, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 308, + 349, + 545, + 392 + ], + "type": "text", + "content": "[42] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. Advances in neural information processing systems, 32, 2019. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 393, + 545, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 393, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 308, + 393, + 545, + 426 + ], + "type": "text", + "content": "[43] Khaled Mamou, E Lengyel, and A Peters. Volumetric hierarchical approximate convex decomposition. In Game Engine Gems 3, pages 141-158. AK Peters, 2016. 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 426, + 545, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 426, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 308, + 426, + 545, + 480 + ], + "type": "text", + "content": "[44] Oscar Michel, Roi Bar-On, Richard Liu, Sagie Benaim, and Rana Hanocka. Text2mesh: Text-driven neural stylization for meshes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13492-13502, 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 481, + 545, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 481, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 308, + 481, + 545, + 503 + ], + "type": "text", + "content": "[45] George A Miller. Wordnet: a lexical database for english. Communications of the ACM, 38(11):39-41, 1995. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 503, + 545, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 503, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 308, + 503, + 545, + 547 + ], + "type": "text", + "content": "[46] Kaichun Mo, Paul Guerrero, Li Yi, Hao Su, Peter Wonka, Niloy Mitra, and Leonidas J Guibas. Structurenet: Hierarchical graph networks for 3d shape generation. arXiv preprint arXiv:1908.00575, 2019. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 548, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 548, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 308, + 548, + 545, + 613 + ], + "type": "text", + "content": "[47] Kaichun Mo, Shilin Zhu, Angel X Chang, Li Yi, Subarna Tripathi, Leonidas J Guibas, and Hao Su. Partnet: A largescale benchmark for fine-grained and hierarchical part-level 3d object understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 909-918, 2019. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 614, + 545, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 614, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 308, + 614, + 545, + 658 + ], + "type": "text", + "content": "[48] D. Morrison, P. Corke, and J. Leitner. Egad! an evolved grasping analysis dataset for diversity and reproducibility in robotic manipulation. IEEE Robotics and Automation Letters, 5(3):4368-4375, 2020. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "type": "text", + "content": "[49] Tai-Yu Pan, Cheng Zhang, Yandong Li, Hexiang Hu, Dong Xuan, Soravit Changpinyo, Boqing Gong, and Wei-Lun Chao. On model calibration for long-tailed object detection and instance segmentation. Advances in Neural Information Processing Systems, 34:2529-2542, 2021. 6" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13151" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[50] Keunhong Park, Konstantinos Rematas, Ali Farhadi, and Steven M Seitz. Photoshape: Photorealistic materials for large-scale shape collections. arXiv preprint arXiv:1809.09761, 2018.3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 117, + 288, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 117, + 288, + 172 + ], + "spans": [ + { + "bbox": [ + 48, + 117, + 288, + 172 + ], + "type": "text", + "content": "[51] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5865-5874, 2021. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 172, + 288, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 172, + 288, + 205 + ], + "spans": [ + { + "bbox": [ + 48, + 172, + 288, + 205 + ], + "type": "text", + "content": "[52] Matt Pharr, Wenzel Jakob, and Greg Humphreys. Physically based rendering: From theory to implementation. Morgan Kaufmann, 2016. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 205, + 288, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 205, + 288, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 205, + 288, + 239 + ], + "type": "text", + "content": "[53] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 239, + 288, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 239, + 288, + 293 + ], + "spans": [ + { + "bbox": [ + 48, + 239, + 288, + 293 + ], + "type": "text", + "content": "[54] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10318–10327, 2021. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 294, + 288, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 294, + 288, + 360 + ], + "spans": [ + { + "bbox": [ + 48, + 294, + 288, + 360 + ], + "type": "text", + "content": "[55] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 1, 2, 5, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 360, + 288, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 360, + 288, + 403 + ], + "spans": [ + { + "bbox": [ + 48, + 360, + 288, + 403 + ], + "type": "text", + "content": "[56] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 404, + 288, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 404, + 288, + 458 + ], + "spans": [ + { + "bbox": [ + 48, + 404, + 288, + 458 + ], + "type": "text", + "content": "[57] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 460, + 288, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 460, + 288, + 514 + ], + "spans": [ + { + "bbox": [ + 48, + 460, + 288, + 514 + ], + "type": "text", + "content": "[58] Ram Ramrakhya, Eric Undersander, Dhruv Batra, and Abhishek Das. Habitat-web: Learning embodied object-search strategies from human demonstrations at scale. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5173-5183, 2022. 2, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 514, + 288, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 514, + 288, + 569 + ], + "spans": [ + { + "bbox": [ + 48, + 514, + 288, + 569 + ], + "type": "text", + "content": "[59] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 570, + 288, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 570, + 288, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 570, + 288, + 624 + ], + "type": "text", + "content": "[60] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115(3):211-252, 2015. 2, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 625, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 288, + 689 + ], + "type": "text", + "content": "[61] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. arXiv preprint arXiv:2210.08402, 2022. 1, 2, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 690, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 690, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 690, + 288, + 714 + ], + "type": "text", + "content": "[62] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 106 + ], + "type": "text", + "content": "Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 1, 2, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 106, + 547, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 106, + 547, + 171 + ], + "spans": [ + { + "bbox": [ + 307, + 106, + 547, + 171 + ], + "type": "text", + "content": "[63] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 2556-2565, 2018. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 172, + 547, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 172, + 547, + 216 + ], + "spans": [ + { + "bbox": [ + 307, + 172, + 547, + 216 + ], + "type": "text", + "content": "[64] Arjun Singh, James Sha, Karthik S. Narayan, Tudor Achim, and P. Abbeel. Bigbird: A large-scale 3d database of object instances. 2014 IEEE International Conference on Robotics and Automation (ICRA), pages 509-516, 2014. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 217, + 547, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 217, + 547, + 282 + ], + "spans": [ + { + "bbox": [ + 307, + 217, + 547, + 282 + ], + "type": "text", + "content": "[65] Krishna Srinivasan, Karthik Raman, Jiecao Chen, Michael Bendersky, and Marc Najork. Wit: Wikipedia-based image text dataset for multimodal multilingual machine learning. In Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 2443-2449, 2021. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 283, + 547, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 283, + 547, + 348 + ], + "spans": [ + { + "bbox": [ + 307, + 283, + 547, + 348 + ], + "type": "text", + "content": "[66] Xingyuan Sun, Jiajun Wu, Xiuming Zhang, Zhoutong Zhang, Chengkai Zhang, Tianfan Xue, Joshua B Tenenbaum, and William T Freeman. Pix3d: Dataset and methods for single-image 3d shape modeling. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2974-2983, 2018. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 349, + 547, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 349, + 547, + 414 + ], + "spans": [ + { + "bbox": [ + 307, + 349, + 547, + 414 + ], + "type": "text", + "content": "[67] Andrew Szot, Alexander Clegg, Eric Undersander, Erik Wijmans, Yili Zhao, John Turner, Noah Maestre, Mustafa Mukadam, Devendra Singh Chaplot, Oleksandr Maksymets, et al. Habitat 2.0: Training home assistants to rearrange their habitat. Advances in Neural Information Processing Systems, 34:251-266, 2021. 2, 5, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 415, + 547, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 415, + 547, + 448 + ], + "spans": [ + { + "bbox": [ + 307, + 415, + 547, + 448 + ], + "type": "text", + "content": "[68] Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. arXiv preprint arXiv:1908.07490, 2019. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 449, + 547, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 449, + 547, + 503 + ], + "spans": [ + { + "bbox": [ + 307, + 449, + 547, + 503 + ], + "type": "text", + "content": "[69] Jingru Tan, Xin Lu, Gang Zhang, Changqing Yin, and Quanquan Li. Equalization loss v2: A new gradient balance approach for long-tailed object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1685-1694, 2021. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 503, + 547, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 503, + 547, + 568 + ], + "spans": [ + { + "bbox": [ + 307, + 503, + 547, + 568 + ], + "type": "text", + "content": "[70] Matthew Tancik, Vincent Casser, Xinchen Yan, Sabeek Pradhan, Ben Mildenhall, Pratul P Srinivasan, Jonathan T Barron, and Henrik Kretzschmar. Block-nerf: Scalable large scene neural view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8248–8258, 2022. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 569, + 547, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 569, + 547, + 602 + ], + "spans": [ + { + "bbox": [ + 307, + 569, + 547, + 602 + ], + "type": "text", + "content": "[71] Guy Tevet, Sigal Raab, Brian Gordon, Yonatan Shafir, Amit H Bermano, and Daniel Cohen-Or. Human motion diffusion model. arXiv preprint arXiv:2209.14916, 2022. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 602, + 547, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 602, + 547, + 646 + ], + "spans": [ + { + "bbox": [ + 307, + 602, + 547, + 646 + ], + "type": "text", + "content": "[72] Bart Thomee, David A Shamma, Gerald Friedland, Benjamin Elizalde, Karl Ni, Douglas Poland, Damian Borth, and Li-Jia Li. Yfcc100m: The new data in multimedia research. Communications of the ACM, 59(2):64-73, 2016. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 647, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 547, + 713 + ], + "type": "text", + "content": "[73] Jiaqi Wang, Wenwei Zhang, Yuhang Zang, Yuhang Cao, Jiangmiao Pang, Tao Gong, Kai Chen, Ziwei Liu, Chen Change Loy, and Dahua Lin. Seesaw loss for long-tailed instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9695-9704, 2021. 6" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13152" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 573 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[74] Wenhui Wang, Hangbo Bao, Li Dong, Johan Bjorck, Zhiliang Peng, Qiang Liu, Kriti Aggarwal, Owais Khan Mohammed, Saksham Singhal, Subhojit Som, et al. Image as a foreign language: Beit pretraining for all vision and vision-language tasks. arXiv preprint arXiv:2208.10442, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 287, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 173 + ], + "type": "text", + "content": "[75] Luca Weihs, Jordi Salvador, Klemen Kotar, Unnat Jain, Kuo-Hao Zeng, Roozbeh Mottaghi, and Aniruddha Kembhavi. Allenact: A framework for embodied ai research. arXiv preprint arXiv:2008.12760, 2020. 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 174, + 287, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 174, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 174, + 287, + 217 + ], + "type": "text", + "content": "[76] Erik Wijmans, Abhishek Kadian, Ari S. Morcos, Stefan Lee, Irfan Essa, Devi Parikh, Manolis Savva, and Dhruv Batra. Dd-ppo: Learning near-perfect pointgoal navigators from 2.5 billion frames. In ICLR, 2020. 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 219, + 287, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 219, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 219, + 287, + 262 + ], + "type": "text", + "content": "[77] Rundi Wu, Chang Xiao, and Changxi Zheng. Deepcad: A deep generative network for computer-aided design models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6772–6782, 2021. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 264, + 287, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 264, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 264, + 287, + 319 + ], + "type": "text", + "content": "[78] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguuang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1912-1920, 2015. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 319, + 287, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 319, + 287, + 384 + ], + "spans": [ + { + "bbox": [ + 48, + 319, + 287, + 384 + ], + "type": "text", + "content": "[79] Fanbo Xiang, Yuzhe Qin, Kaichun Mo, Yikuan Xia, Hao Zhu, Fangchen Liu, Minghua Liu, Hanxiao Jiang, Yifu Yuan, He Wang, et al. Sapien: A simulated part-based interactive environment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11097-11107, 2020. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 386, + 287, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 386, + 287, + 441 + ], + "spans": [ + { + "bbox": [ + 48, + 386, + 287, + 441 + ], + "type": "text", + "content": "[80] Yuanbo Xiangli, Linning Xu, Xingang Pan, Nanxuan Zhao, Anyi Rao, Christian Theobalt, Bo Dai, and Dahua Lin. Bungeenerf: Progressive neural radiance field for extreme multi-scale scene rendering. In The European Conference on Computer Vision (ECCV), 2022. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 442, + 287, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 442, + 287, + 495 + ], + "spans": [ + { + "bbox": [ + 48, + 442, + 287, + 495 + ], + "type": "text", + "content": "[81] Zhenjia Xu, Beichun Qi, Shubham Agrawal, and Shuran Song. Adagrasp: Learning an adaptive gripper-aware grasping policy. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 4620-4626. IEEE, 2021. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 498, + 287, + 573 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 498, + 287, + 573 + ], + "spans": [ + { + "bbox": [ + 48, + 498, + 287, + 573 + ], + "type": "text", + "content": "[82] Rowan Zellers, Jiasen Lu, Ximing Lu, Youngjae Yu, Yanpeng Zhao, Mohammadreza Salehi, Aditya Kusupati, Jack Hessel, Ali Farhadi, and Yejin Choi. Merlot reserve: Neural script knowledge through vision and language and sound. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16375-16387, 2022. 1, 4" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13153" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Object Detection With Self-Supervised Scene Adaptation/b774b6cf-eacd-4660-845f-f32b5ec4dc83_content_list.json b/2023/Object Detection With Self-Supervised Scene Adaptation/b774b6cf-eacd-4660-845f-f32b5ec4dc83_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1495df30deb31acb49aa477142168fdac5424347 --- /dev/null +++ b/2023/Object Detection With Self-Supervised Scene Adaptation/b774b6cf-eacd-4660-845f-f32b5ec4dc83_content_list.json @@ -0,0 +1,1927 @@ +[ + { + "type": "text", + "text": "Object Detection with Self-Supervised Scene Adaptation", + "text_level": 1, + "bbox": [ + 200, + 130, + 769, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zekun Zhang1 Minh Hoai1,2", + "bbox": [ + 323, + 179, + 643, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Stony Brook University, Stony Brook, NY 11794, USA", + "bbox": [ + 261, + 204, + 707, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ VinAI Research, Hanoi, Vietnam", + "bbox": [ + 349, + 223, + 620, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{zekzhang,minhhooai}@cs.stonyrbook.edu", + "bbox": [ + 318, + 243, + 645, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 292, + 313, + 309 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This paper proposes a novel method to improve the performance of a trained object detector on scenes with fixed camera perspectives based on self-supervised adaptation. Given a specific scene, the trained detector is adapted using pseudo-ground truth labels generated by the detector itself and an object tracker in a cross-teaching manner. When the camera perspective is fixed, our method can utilize the background equivariance by proposing artifact-free object mixup as a means of data augmentation, and utilize accurate background extraction as an additional input modality. We also introduce a large-scale and diverse dataset for the development and evaluation of scene-adaptive object detection. Experiments on this dataset show that our method can improve the average precision of the original detector, outperforming the previous state-of-the-art self-supervised domain adaptive object detection methods by a large margin. Our dataset and code are published at https://github.com/cvlab-stonybrook/scenes100.", + "bbox": [ + 73, + 325, + 473, + 598 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 630, + 209, + 646 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The need to detect objects in video streams from stationary cameras arises in many computer vision applications, including video surveillance and autonomous retail. In general, different applications require the detection of different object categories, and each computer-vision-based product will have its own detector. However, for a specific product, there is typically a single detector that will be used for many cameras/scenes. For example, a typical video surveillance product would use the same detector to detect pedestrians and vehicles for network cameras installed at different locations. Unfortunately, a single detector might not work well for all scenes, leading to trivial and unforgiving mistakes.", + "bbox": [ + 75, + 657, + 468, + 838 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This fundamental problem of many computer vision products stems from the limited generalization power of a single model, due to limited training data, limited model capacity, or both. One can attempt to address this problem by", + "bbox": [ + 75, + 839, + 470, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "using more training data, but it will incur additional cost for data collection and annotation. Furthermore, in many cases, due to the low latency requirement or the limited computing resources for inference, a product is forced to use a very lightweight network, and this network will have limited representation capacity to generalize across many scenes.", + "bbox": [ + 496, + 294, + 890, + 385 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, instead of having a single scene-generic detector, we propose using scene-specific detectors. This yields higher detection performance as each detector is customized for a specific scene, and allows us to use a lightweight model without sacrificing accuracy as each detector is only responsible for one scene.", + "bbox": [ + 496, + 386, + 890, + 474 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Obtaining scene-specific detectors, however, is very challenging. A trivial approach is to train a detector for each scene separately, but this requires an enormous amount of annotated training data. Instead, we propose a self-supervised method to adapt a pre-trained detector to each scene. Our method records the unlabeled video frames in the past, uses the trained detector to detect objects in those frames, and generates augmented training data based on those detections. Although the detections made by the pre-trained model can be noisy, they can still be useful for generating pseudo annotated data. We further extend those pseudo bounding boxes by applying object tracking [2, 57] along the video timeline, aiming to propagate the detections to adjacent frames to recover some of the false negatives not returned by the detector. We also use multiple detectors to obtain the pseudo labels and train the detector in a cross-teaching manner, taking the advantage of the ensemble of models [13, 24].", + "bbox": [ + 496, + 477, + 892, + 748 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Exploiting the stationary nature of the camera, we propose two additional techniques to boost the detection performance: location-aware mixup and background-augmented input. The former is to generate more samples during training through object mixup [76] that contains less artifacts, based on the aforementioned pseudo boxes generated from detection and tracking. The latter involves estimating the background image and fusing it with the detector's input.", + "bbox": [ + 496, + 750, + 890, + 869 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In short, the main contribution of our paper is a novel framework that utilizes self-supervision, location-aware ob", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "21589", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ject mixup, and background modeling to improve the detection performance of a pre-trained object detector on scenes with stationary cameras. We also contribute a large scale and diverse dataset for the development of scene adaptive object detection, which contains sufficient quantity and quality annotations for evaluation.", + "bbox": [ + 75, + 90, + 472, + 183 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 195, + 218, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite much recent improvement in object detection research [3,5,6,8,12,18,19,21,25,28,31,36,40,43,45,47-50, 53-56,64,65,67-69,77,79], trained detectors still encounter problems with domain shift or domain gap. Scene adaptive object detection can be viewed as a special case of domain adaption, in which an object detector trained on a fully-supervised source domain is adapted to a target domain. Most research in this direction focuses on semi-supervised, weakly-supervised, or self-supervised adaptation. In addition, source-free adaptation [26, 33, 35, 71] has been proposed to address the situation where the source domain data is unavailable during adaptation.", + "bbox": [ + 75, + 220, + 468, + 400 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Self-labeling methods [11,27,29,33,37,44,46,57,60,73] uses the teacher-student setup from semi-supervised image classification [30, 62]. A teacher detector trained on the source domain generates pseudo bounding boxes on the target domain images, and a student model is trained with those boxes to improve its performance on the target domain. Techniques such as weak/strong augmentation [37,44,60], knowledge distillation [11], and weight averaging [33, 37] have been used to deal with noisy pseudo labels from the teacher detector due to the domain gap. Our proposed pseudo-labeling method uses two teacher models trained on the source domain to train a student model, and their pseudo boxes are aggregated and refined. We further apply tracking as in [57], but with both forward and backward directions, to extend the pseudo labels. The pseudo boxes also form the basis for later location-aware mixup and dynamic background extraction steps.", + "bbox": [ + 75, + 402, + 470, + 657 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Domain alignment methods [7, 11, 22, 33, 34, 58, 72, 78] aim to reduce the domain gap by enforcing the models to output similarly on the source and target domain at image, proposal, or instance levels. They use domain adversarial learning through gradient reversal [15, 16] or graph matching [34] for domain alignment. This approach involves adding domain alignment losses to existing object detection models and is complementary to our proposed method.", + "bbox": [ + 75, + 659, + 468, + 779 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Another seemingly related research area is continual or incremental learning. Scene adaptive and domain adaptive object detection are in between task-continual learning [38] and data-continual learning [52], for that the object categories of the source and target domains are the same, but the data distribution differs. It is possible to formulate scene adaptive object detection as an online learning problem, but this approach would require additional annotations.", + "bbox": [ + 75, + 780, + 470, + 902 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/57ebeeecc3d09929f50445e8a6752363c0424a4c7ac22dbb1bfdd43d990bb94f.jpg", + "image_caption": [ + "(a) Input image $\\mathcal{I}$ ." + ], + "image_footnote": [], + "bbox": [ + 501, + 85, + 684, + 172 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/360d51fccf503508fa75a943d66deaef16f2f5d493c82dc5ac89ee2b717c23eb.jpg", + "image_caption": [ + "(b) Learned background image $\\mathcal{B}$ ." + ], + "image_footnote": [], + "bbox": [ + 702, + 90, + 885, + 171 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0b7af89dc2853a1079de2d8f2c6ac3e668569254650dddd42448c5a64941b95b.jpg", + "image_caption": [ + "(c) Object mask $\\mathcal{M}^O$", + "Figure 1. Inference flow of the proposed detector that can adapt to a specific scene. The background image and fusion faster-RCNN model are learned during training, discussed in Sec. 3." + ], + "image_footnote": [], + "bbox": [ + 501, + 181, + 683, + 260 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Methods", + "text_level": 1, + "bbox": [ + 500, + 328, + 598, + 344 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For an object detector pre-trained on a source dataset referred as the base model, we aim to improve its performance on video stream from a stationary camera in a given scene, with unlabeled videos captured by the same camera in the past as adaptation training samples. The categories of object of interest in the scene are also present in the source dataset, but the data distribution is shifted. The base model is adapted to each scene independently.", + "bbox": [ + 496, + 354, + 890, + 474 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Assuming the distribution shift is moderate, the base detector can generate partially decent bounding boxes on the unlabeled frames. We use them as pseudo-labels to improve the base model itself. Many objects are ignored by the base detector, as their appearance can be very different from the source dataset. However, videos contain temporal correlation information, making it possible to apply object tracking in both time directions to recover some of those false negatives. The pseudo-labels are then refined to remove low-confidence boxes and duplicates. We refer to this process of obtaining pseudo bounding boxes as pseudo-labeling.", + "bbox": [ + 496, + 476, + 892, + 643 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As data augmentation is beneficial for training on less reliable labels [51], we take advantages of object mixup to generate new sample images. Since the background is stationary, those images contain fewer artifacts, which is beneficial for detectors. We further adopt the idea of ensemble models [13, 24] by using two base detectors to obtain the pseudo labels, for they can be complementary to each other. The pseudo boxes are used to train one base model, which we refer as cross-teaching.", + "bbox": [ + 496, + 643, + 890, + 779 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Since the job of a detector is to separate foreground objects from the background, having information about the background is advantageous. In a stationary camera video with moving objects, different parts of the background are covered by objects in different frames, making it possible to model the complete background. We use the aforementioned pseudo bounding boxes to mark the uncovered parts of the background in each frame. Combining the partial", + "bbox": [ + 496, + 780, + 890, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "21590", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/77df66211452e80f23b46a4eed31c96da2966fc781a8a6759bf5232e0449e947.jpg", + "image_caption": [ + "Figure 2. Training flow of the proposed self-supervised scene adaptation object detection framework with sample images. The mixup training images with pseudo bounding boxes and background image are used in the fusion training. The details are discussed in Sec. 3." + ], + "image_footnote": [], + "bbox": [ + 76, + 84, + 893, + 247 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "background from a sequence of frames can give an accurate and complete background model. We then modify the Faster-RCNN [56] architecture to fuse the background with the input image to improve detection performance.", + "bbox": [ + 75, + 306, + 468, + 366 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The overall structure and data flows of the proposed methods are shown in Fig. 1 and Fig. 2. Our proposed methods are highly flexible and can be configured to trade-off between performance and speed. Object tracking can be bypassed to reduce the time to get the pseudo-bounding boxes. Mixup and object mask fusion can also be disabled for faster training and inference. In Sec. 5.5 we will show ablation study results on how those components affect the performance of the adapted models. We now describe each component in more details. For consistency, the images are all from the first video in our collected dataset, which will be described in Sec. 4.", + "bbox": [ + 75, + 366, + 472, + 547 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Base Detection Models on Source Domain", + "text_level": 1, + "bbox": [ + 76, + 556, + 433, + 571 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We use Faster-RCNN [56] in our experiments for its high precision and flexible modular architecture. Among the techniques proposed for self-supervised scene adaptation, both pseudo-labeling and mixup can be directly applied to any object detection model. The object mask fusion technique requires modification to the network architecture, but it can be easily extended to most types of detectors by adding a parallel branch to the existing network.", + "bbox": [ + 75, + 580, + 468, + 700 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The object labels in MSCOCO training set are remapped to person and vehicle as described in Sec. 4.2. All objects other than person, car, bus, and truck are discarded. We take Faster-RCNNs pre-trained on the original MSCOCO training set, and finetune them on the remapped training set. We choose two models with ResNet-50 and ResNet101 backbones as base models, referred as M1 and M2.", + "bbox": [ + 75, + 700, + 468, + 806 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Pseudo Labels from Detection and Tracking", + "text_level": 1, + "bbox": [ + 76, + 816, + 450, + 832 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We take the frames from the first training portion of each video, and adopt the idea of detection and tracking similar to [57], then apply our proposed bounding box refinement and cross-teaching schemes. Base detectors described in", + "bbox": [ + 75, + 839, + 470, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Sec. 3.1 are applied on the frames, and any bounding box with confidence score higher than $\\lambda_{det}$ is kept as a valid pseudo box.", + "bbox": [ + 496, + 306, + 890, + 349 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Single-object trackers are initialized with detected objects having confidence scores higher than $\\lambda_{\\text{tot}}$ . We use DiMP-50 [2] to track the objects in both forward and backward directions in the video timeline, and the tracked object bounding boxes are used as pseudo boxes. The reason for tracking in both directions is that if an object is moving towards the camera and appears bigger by time, backward tracking is more accurate because the initial bounding box contains more detailed texture. To prevent drifting, the maximum length of each track is capped at two seconds.", + "bbox": [ + 496, + 351, + 892, + 502 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Bounding box refinement is applied to eliminate duplicated boxes of the same object from combining pseudo bounding boxes from detector and tracker. Each candidate box in the same frame is regarded as a graph node. If two boxes have the same object label, and their IoU is above a threshold $\\lambda_{iou}$ , then an edge is added to connect them. We assume that each connected component of the graph represents one single object, and the nodes in the component are duplicated. We only keep the node with the highest degree (i.e., the node with the highest number of connected edges). An example of refinement results is shown in Fig. 2b. Refinement can effectively remove duplicated bounding boxes. However, it can neither remove false positives nor recover false negatives.", + "bbox": [ + 496, + 503, + 893, + 714 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Both M1 and M2 are used in detection and tracking. The pseudo bounding boxes from both base models are then refined jointly, and the refined boxes are used to train M2 in adaptation. We refer this setting as cross-teaching.", + "bbox": [ + 496, + 714, + 893, + 776 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Location-Aware Object Mixup", + "text_level": 1, + "bbox": [ + 500, + 786, + 769, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Mixup [76] is originally proposed as a simple but powerful data augmentation method in for image classification. It is quickly improved by many modifications [1, 23, 32, 66, 75]. Mixup is also used in object detection tasks [4, 75], where it can be applied at image or instance levels. During adaptation training, a frame with its pseudo bounding", + "bbox": [ + 496, + 810, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "21591", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/03bbdeb8e94af07618547e6e68937a16b75f0aa3c9a6fc5ce3fc841c9d9c5d34.jpg", + "image_caption": [ + "(a) Random mixup." + ], + "image_footnote": [], + "bbox": [ + 80, + 88, + 272, + 172 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/74d6a0c14a49681b388cd6772e9f45ba252a2a8a14825519122ba04d6ba94dab.jpg", + "image_caption": [ + "(b) Location-aware mixup.", + "Figure 3. An example of location-aware object mixup. Compared to random mixup, less artifacts are introduced." + ], + "image_footnote": [], + "bbox": [ + 274, + 88, + 468, + 172 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "boxes has the probability $p_{\\text{mixup}}$ of being paired with another randomly select source frame for mixup. A portion $r_{\\text{mixup}}$ of the pseudo bounding boxes in the source frame are cropped out and pasted onto the original frame, along with their locations and labels. If a pasted box covers more than $\\alpha_{\\text{cover}}$ of a pseudo box in the original frame, the pseudo box will be excluded from the pseudo labels. An example of this location-aware object mixup is shown in Fig. 2c and Fig. 3. Since the background is stationary, we propose to use location-aware mixup instead of the original random mixup to reduce the amount of artifacts in the generated mixup images and to keep the location distribution of the objects in the scene. This will be shown to be beneficial in the experiments.", + "bbox": [ + 75, + 229, + 472, + 441 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Dynamic Background Extraction", + "text_level": 1, + "bbox": [ + 76, + 453, + 369, + 469 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For a video frame $\\mathcal{I}$ of dimension $W\\times H$ associated with a set of $K$ pseudo-annotated object bounding boxes $\\{(x_k^1,y_k^1,x_k^2,y_k^2)|k = 1\\dots K\\}$ , a background mask $\\mathcal{M}$ of the same dimension as $\\mathcal{I}$ can be constructed as follows. For each location $(x,y)$ in $\\mathcal{M}$ , we set $M[x,y] = 0$ if $(x,y)$ is inside of any pseudo-annotated bounding box and 1 otherwise. Then for a sequence of frame-mask pairs $\\{(\\mathcal{I}_l,\\mathcal{M}_l)|l = 1\\dots L\\}$ , the background image is determined as", + "bbox": [ + 76, + 477, + 468, + 611 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {B} = \\frac {\\sum_ {l = 1} ^ {L} \\mathcal {I} _ {l} \\otimes \\mathcal {M} _ {l}}{\\sum_ {l = 1} ^ {L} \\mathcal {M} _ {l}}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 199, + 611, + 468, + 650 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\otimes$ is the pixel-wise multiplication operator.", + "bbox": [ + 76, + 657, + 406, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "There might exist a location $(x', y')$ that lies inside an object bounding box in every image, i.e., $\\mathcal{M}_l[x', y'] = 0$ . In this case, the background at this location is never observed and its pixel value cannot be determined. In this case, we \"guess\" the background value by inpainting inside such areas using the inpainting algorithm of [63].", + "bbox": [ + 75, + 672, + 468, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fig. 2d shows an example background, which is reasonably accurate. Although in each video, the camera perspective is fixed, the background can change overtime due to some factors such as illumination change. To incorporate this, the background extraction is operated every $T_{bg}$ seconds, giving a dynamic background modeling.", + "bbox": [ + 75, + 763, + 468, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We construct an object mask as an additional input modality for the detector, to utilize the extracted background to improve detection performance during adapta", + "bbox": [ + 75, + 854, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tion. For a frame $\\mathcal{I}$ and its corresponding background model image $\\mathcal{B}$ , we define its object mask image as", + "bbox": [ + 498, + 90, + 890, + 121 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {M} ^ {O} = (\\mathcal {I} - \\mathcal {B} + 1) \\times 0. 5, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 127, + 890, + 143 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where we assume the pixel values in both $\\mathcal{I}$ and $\\mathcal{B}$ have been normalized to the range [0, 1]. An object mask example is shown in Fig. 1c, in which objects are clearly separated from the background. Note that since $\\mathcal{I},\\mathcal{B}$ , and $\\mathcal{M}^O$ are linearly dependent, no information is lost.", + "bbox": [ + 496, + 151, + 890, + 226 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Before the adaptation training process, the base detector needs to be modified to use object mask input modality and trained on the MSCOCO training set with background images. However, dynamic background extraction on MSCOCO is impossible since it only consists of static images. Therefore, we rely on the object mask annotations in the dataset and apply the same inpainting algorithm [63] to generate the background image for each training image.", + "bbox": [ + 496, + 227, + 890, + 348 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.5. Object Mask Fusion", + "text_level": 1, + "bbox": [ + 500, + 354, + 692, + 371 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The architecture of Faster-RCNN [56] is depicted in Fig. 4a. First, a CNN backbone with a feature pyramid network (FPN) [39] is utilized to extract the Feature Pyramid (FP). The first layer in the CNN that takes the input image is also shown here. Then, an RPN [39] is employed to produce object bounding box proposals. Lastly, an ROI head [19,20] is used to assign object labels and refine the bounding boxes on the pooled feature maps of the proposals. The entire network is trained utilizing two sets of losses: localization loss and objectness loss from RPN, and localization loss and classification loss from the ROI head. The fusion models explained below start with the base model that is trained on the remapped MSCOCO training set.", + "bbox": [ + 496, + 378, + 890, + 574 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Early-fusion. The input to the Faster-RCNN network is the stacked image $[\\mathcal{I};\\mathcal{M}^{\\mathcal{O}}]$ , with the two modalities stacked along the color channels, as shown in Fig. 4b. To enable the network to take a 6-channel image as input instead of the vanilla 3-channel image, the first layer of the backbone CNN is duplicated and stacked. The convolution kernel weights are first copied, then halved. This ensures that the feature map after the first layer conv1|conv2 is the same as in the vanilla model when $[\\mathcal{I};\\mathcal{I}]$ is fed into it, providing a smooth start for adaptation training. This modification introduces only a small number of additional model parameters, and the increase in computational cost is negligible.", + "bbox": [ + 496, + 575, + 890, + 756 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Mid-fusion. Both $\\mathcal{I}$ and $\\mathcal{M}^O$ are fed into the FPN in parallel, which yields two feature pyramid FP1 and FP2, as shown in Fig. 4c. Two sets of RPNs and ROI heads are used in parallel, and their initial weights are copied from the base model. FP1 is used in the left branch with RPN1 and ROI1, which is the same as the vanilla model. In the right branch with RPN2 and ROI2, the input feature pyramid is the fusion of FP1 and FP2:", + "bbox": [ + 496, + 757, + 890, + 876 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {F P} _ {\\text {f u s i o n}} = \\mathrm {F P 1} \\oplus \\mathrm {F P 2}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 885, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "21592", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/94794aeb6e2f4a0443ae56d799a03a8631adf755045aa8e2a605359b16d69a6d.jpg", + "image_caption": [ + "(a) Vanilla Faster-RCNN." + ], + "image_footnote": [], + "bbox": [ + 124, + 88, + 218, + 170 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c3e7a84c7d82014934e52d0acf8860030d9f47dac13140a529bfa5275090b02f.jpg", + "image_caption": [ + "(c) Mid-fusion Faster-RCNN." + ], + "image_footnote": [], + "bbox": [ + 86, + 189, + 272, + 271 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c6b44f02073ad3f51063f656a09dbf4347ff86f0f440124b09146577ffd79648.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 316, + 88, + 436, + 170 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3c26c955f7096337991c93c9b2f70ab6e26b1cdb38ce2160a9b4e60f00f2a0b1.jpg", + "image_caption": [ + "(b) Early-fusion Faster-RCNN.", + "(d) Late-fusion Faster-RCNN." + ], + "image_footnote": [], + "bbox": [ + 282, + 189, + 468, + 271 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The fused feature pyramid has the same dimension as both FP1 and FP2. The fusion operator can be either non-parametric such as average pooling, or parametric such as convolution layers. The whole network is trained with losses from both branches", + "bbox": [ + 75, + 344, + 468, + 417 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {m i d} = \\left(1 - \\alpha_ {m i d}\\right) \\left(\\mathcal {L} _ {r p n} ^ {\\mathcal {I}} + \\mathcal {L} _ {r o i} ^ {\\mathcal {I}}\\right) + \\alpha_ {m i d} \\left(\\mathcal {L} _ {r p n} ^ {f u s i o n} + \\mathcal {L} _ {r o i} ^ {f u s i o n}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 422, + 475, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\alpha_{mid}$ is the weights of different losses.", + "bbox": [ + 76, + 457, + 377, + 470 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Late-fusion is similar to mid-fusion as shown in Fig. 4d. The key difference is that only the ROI head is branched. The region proposals are only generated from FP1, and the fused feature pyramid is only used in ROI head in the right branch. As a result, the total loss is", + "bbox": [ + 76, + 472, + 468, + 547 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {l a t e} = \\mathcal {L} _ {r p n} ^ {\\mathcal {I}} + (1 - \\alpha_ {l a t e}) \\mathcal {L} _ {r o i} ^ {\\mathcal {I}} + \\alpha_ {l a t e} \\mathcal {L} _ {r o i} ^ {f u s i o n}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 551, + 468, + 571 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Both mid-fusion and late-fusion introduce significant additional parameters and computation. $\\mathcal{I}$ and $\\mathcal{M}^O$ need to pass the backbone, requiring more computation even at inference time. The quantitative comparison of inference speed of different fusion models can be found in the supplementary material.", + "bbox": [ + 75, + 592, + 468, + 681 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To train the fusion models in adaptation, we first fine-tune the base models on MSCOCO training set with the inpainted background models described in Sec. 3.4 until convergence. During evaluation, we use the last background image to obtain the object masks, as it is the closest to the evaluation images on the timeline.", + "bbox": [ + 75, + 683, + 468, + 772 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Scenes100 Dataset", + "text_level": 1, + "bbox": [ + 76, + 784, + 254, + 800 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For our proposed scene-adaptive object detection problem, there is no existing dataset with long enough videos of stationary backgrounds for the development and evaluation of self-supervised adaptation techniques. We therefore collected a new dataset called Scenes100, which will be described in this section.", + "bbox": [ + 75, + 810, + 468, + 898 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/fead86c9d20fe3d1cddfaf756c62add0c97143aac5a8b44b51a98ab63949fc69.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 88, + 629, + 143 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/13cf271c9c66c7f42573759012b562a1f7a517e55b89d4f212c4f74c5e16ab16.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 143, + 629, + 200 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/fb1b35766b0f16ffc593fc7a7ece9de10cec27d49e4f0e12865bd4cdb8ea09ff.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 503, + 200, + 629, + 257 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c7750e5360e574180f3330f991604afb14d8562ab07f68c02a6e9ff28cf52711.jpg", + "image_caption": [ + "Figure 4. Illustration of different fusion models for Faster-RCNN. Network modules are indicated by green, and feature maps and outputs are indicated by yellow.", + "Figure 5. Some annotated frames from the evaluation portions. The green mask shows the part of the frame that is excluded for annotation and evaluation. Objects are labeled by bounding boxes. Please see the supplementary material for detailed descriptions." + ], + "image_footnote": [], + "bbox": [ + 503, + 257, + 629, + 311 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/32ab7d85f076faa1a46c3d8d80d7aba865923f82d3dec575e7ff24e010374235.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 632, + 89, + 758, + 143 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4013098c739e593643edbb8079a3dfc399eef4c9440df24b292264126343df3a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 632, + 143, + 758, + 200 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/17fff89abfcb963752a9856b21da43afa732d96f21713b98397d582c9e9d9070.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 632, + 200, + 758, + 257 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/07c1f575910c86ef3307858986c1a1735eede54df05eb00091566aea3bad2ef0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 632, + 257, + 758, + 311 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/aab208c4b08ac9b07b770d63acf81d18fdf26f78afd12259277ad642c0c82fc6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 761, + 89, + 888, + 143 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/fd3fc721cbb5f65ec85d87cbd64fd49639a1b01edf4ee4a5860c3b19635499c6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 761, + 143, + 888, + 200 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5c428924b8bca60fe8bfee2c14e7e1dfc0bd0ad3901bb838146ca04f011277f5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 761, + 200, + 888, + 257 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/21197db50c1e6ad1809beaa6ddee2ad0d2e1d371d1474781c5374f296b0e9609.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 761, + 257, + 888, + 311 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Data Collection", + "text_level": 1, + "bbox": [ + 500, + 380, + 655, + 395 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We used keywords such as \"live view\", \"street camera\", and \"live webcam\" to search for live streams on YouTube. We look for streams that are of decent quality and have fixed camera perspective. Around 200 candidate videos were recorded between September 2020 and February 2022. We picked 100 videos to compile our Scenes100 dataset, ensuring that each video is longer than two hours and has minimum width and height of 720 pixels. We also aimed for a dataset with high diversities. The 100 videos of Scenes100 were recorded in a variety of 16 countries and territories, and during different time periods of day and year. The dataset contain videos from both indoor and outdoor scenes, with different camera perspectives. The density and scale of visual objects also vary significantly from one video to another. Some sample frames is presented in Fig. 5 to show the diversity of our dataset.", + "bbox": [ + 496, + 405, + 890, + 647 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Manual Annotation", + "text_level": 1, + "bbox": [ + 500, + 662, + 687, + 676 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We are interested in detecting objects from two categories: person and vehicle. The person category includes any visible and recognizable people in the frames; this corresponds to the person category in MSCOCO [41] dataset. The vehicle category includes all vehicles with four or more wheels, so motorbikes, bicycles, tricycles are not included. This category corresponds to the car, bus, and truck categories in MSCOCO.", + "bbox": [ + 496, + 686, + 890, + 806 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For each video, we used the first 1.5 hours for self-supervised and weakly-supervised training, and the remaining part for evaluation. Since many videos were taken with wide-angle lenses, some frames included faraway objects that were too small to be considered in evaluation. To address this, we manually drew polygon masks to exclude", + "bbox": [ + 496, + 810, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "21593", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/ab6bea824d15f1a8e4a1c1a9b1c3a631b3c496ecaf39d5795de1ac2768a189d5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
datasetcontain videosaverage lengthframescountriesboundong boxesboxes per video
MSCOCO [41]No---897K-
KITTI [17]No--180K-
BDD100K [74]Yes40s120M11.8M18
CityScapes [9]Yes1.8s150K265K13
Scenes100Yes2h21.6M1684K840
", + "bbox": [ + 78, + 88, + 467, + 193 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Comparison of Scenes100 with other object detection datasets $(\\mathbf{K} = 10^{3},\\mathbf{M} = 10^{6})$", + "bbox": [ + 76, + 196, + 468, + 224 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "these faraway parts from annotation and evaluation. We then estimated the number of visual objects in the unmasked parts by first using a pre-trained detector, and then refining the estimate through human annotation. Frames for annotation were uniformly sampled from the second portion of each video, with the number of frames being inversely proportional to the estimated number of objects in the frames. This ensured that each video had roughly the same number of annotated objects. We contracted a data annotation company to perform bounding box annotation on those sampled frames. Some annotated frame samples for evaluation is shown in Fig. 5.", + "bbox": [ + 75, + 233, + 470, + 416 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Dataset Statistics", + "text_level": 1, + "bbox": [ + 76, + 424, + 245, + 438 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Tab. 1, we compare Scenes100 with some popular object detection datasets. MSCOCO [41] is a fully-supervised general-purpose object detection dataset, which acts as the supervised source domain. KITTI [17] only contains object detection annotation associated to individual images. In BDD100K [74], object detection annotation is given for one key frame in each video. For CityScapes [9], we consider the finely labeled instance-level segmentation masks, which are given for the $20^{\\mathrm{th}}$ image from a 30 frame video snippets lasting 1.8s.", + "bbox": [ + 75, + 446, + 468, + 598 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The scale of Scenes100 is comparable to that of KITTI and CityScapes. However, Scenes100 has three unique features. First, the videos in Scenes100 are much longer than those in any other datasets with scenes, providing sufficient data for a model to adapt to each of them. Second, Scenes100 has greater diversity. Unlike some other datasets that are recorded in only a few cities with similar camera angles, background environments, and object types, Scenes100 covers several different countries, with varying weather conditions, road conditions, view of field, camera perspectives, and vehicle types. Third, each video in Scenes100 has a fixed camera perspective and stationary background scene, making background modeling possible.", + "bbox": [ + 75, + 598, + 468, + 794 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Note that many currently available datasets are not primarily intended for self-supervised adaptive object detection tasks. They may possess other features, such as segmentation masks, multi-modal sensor data, depth maps, or annotations for object tracking. However, despite the absence of these attributes, Scenes100 holds value as the first extensive and varied dataset that features fixed camera per", + "bbox": [ + 75, + 795, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "spectives and lengthy videos. As such, it can complement existing dataset collections and serve as a valuable resource for researchers exploring scene adaptive object detection.", + "bbox": [ + 496, + 90, + 890, + 137 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 500, + 148, + 633, + 166 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Evaluation Protocols", + "text_level": 1, + "bbox": [ + 500, + 174, + 696, + 189 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate the performance of object detectors following the COCO evaluation protocol [41] to calculate the average precision at $IoU = 50\\%$ ( $AP^{50}$ ) and mean average precision of different $IoU$ thresholds from 0.5 to 0.95 ( $AP^m$ ), with some modifications as follows.", + "bbox": [ + 496, + 196, + 890, + 271 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Non-evaluation region: before feeding the ground-truth and detected bounding boxes to the COCO evaluator, the bounding boxes that have at least one corner inside the nonevaluation mask will be removed. So the faraway parts in the frames where objects are deemed too small and blurry will not affect the evaluation results. Please see Sec. 4.2 and Fig. 5 for more details.", + "bbox": [ + 496, + 273, + 890, + 377 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Category weighting: in the standard COCO evaluation protocol, the overall multi-class $AP$ is the simple average of the $AP_{s}$ of the different categories. If a class is sparse in the training set and validation set, the model tends to perform poorly due to the scarcity of the training samples. In this case the average $AP$ does not truly reflect the performance of a detector. Hence we propose an additional multi-class $AP$ metric with weighted average where the weight is determined by the prevalence of ground truth object instances in the evaluation set. This weighted $AP$ can better portray the performance of a detector in the case of long-tailed distribution. We refer the standard classes-mean COCO $AP$ as $AP_{co}$ , and the proposed weighted $AP$ as $AP_{w}$ . We refer the $AP$ averaged over IoU thresholds and IoU = 50% as $AP^{m}$ and $AP^{50}$ .", + "bbox": [ + 496, + 378, + 892, + 603 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Per-video evaluation: we aim to evaluate the performance of a scene adaption method, instead of a specific detection model. For each of the video, a model is adapted to it from the same base model trained on the source dataset. The performance of the base model and adapted model on each video is calculated individually as $\\{(AP_{v,\\text{base}}, AP_{v,\\text{adapt}}) | v = 1 \\ldots 100\\}$ . Then we evaluate the overall effectiveness of the adaptation method using the averaged $AP$ gain as:", + "bbox": [ + 496, + 604, + 890, + 741 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nA P G = \\frac {1}{1 0 0} \\sum_ {v = 1} ^ {1 0 0} \\left(A P _ {v, a d a p t} - A P _ {v, b a s e}\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 751, + 890, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Implementation Details", + "text_level": 1, + "bbox": [ + 500, + 801, + 718, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our implementation is based on Detector2 [70]. The DiMP tracker is directly taken from the official PyTracking [10] implementation without any change. More details and hyper-parameters can be found in the supplementary material. During training, we include the same number of", + "bbox": [ + 496, + 824, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "21594", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "images from MSCOCO training set as the number of unlabeled frames from the videos. The hyper-parameters are the same for all videos. We check the training loss for all experiments, including the baseline methods, to ensure convergence. We also train \"compound\" models where all videos are used together for adaptation. The results and further details can be found in the supplementary material.", + "bbox": [ + 75, + 90, + 472, + 196 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Domain Adaptation Baselines", + "text_level": 1, + "bbox": [ + 76, + 212, + 341, + 229 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compare our methods with several domain adaptive object detection baselines. In their original papers, the methods are evaluated on other domain adaptive object detection datasets. However, those datasets do not contains videos with fixed camera perspectives, so the proposed location-aware mixup and dynamic background extraction cannot be applied. We instead apply the methods in other papers on Scenes100. For a fair comparison, all methods start with the same base model and weights, and then are adapted to and evaluated on each scene individually. We keep the values of the hyper-parameters that are specific to a baseline method used in the original papers. We use the same learning rate, batch size, and number of iterations as our experiments in those baseline methods. More details can be found in the supplementary material.", + "bbox": [ + 75, + 238, + 468, + 464 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Self-Train (ST) [57] is a method that uses detection and tracking to obtain pseudo bounding boxes, similar to ours. But ST applies tracking only in one direction. It assigns different weights to the pseudo bounding boxes during training, but no refinement nor cross-teaching.", + "bbox": [ + 75, + 465, + 468, + 542 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "STAC [60] uses the base model to get pseudo bounding boxes, and trains the model on a strongly augmented version of the target domain image with those boxes.", + "bbox": [ + 75, + 544, + 468, + 589 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Adaptive Teacher (AT) [37] also uses self-training, but the pseudo bounding boxes are detected on the fly. It utilizes exponential moving average to update the model gradually. It also uses weak/strong data augmentation, and a domain classifier to align the features of the two domains.", + "bbox": [ + 75, + 592, + 468, + 666 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "$\\mathbf{H}^2\\mathbf{FA}$ R-CNN [72] applies feature alignment of source and target domains at various levels of Faster-RCNN. It assumes that image-level weak annotation is available for the target domain. So we use the labels from our pseudo-labeling process as image-level class labels.", + "bbox": [ + 75, + 667, + 468, + 744 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "TIA [78] is a method that instantiates feature alignment by using both domain classifier and auxiliary classification and localization heads. The model is trained in a mini-max manner using gradient reverse layers [15].", + "bbox": [ + 75, + 747, + 468, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "LODS [33] uses style enhancement as data augmentation. The pseudo bounding boxes are generated from the original images, and the features from both versions of the same image are aligned. It also utilizes exponential moving average for the weights. It does not use any supervised training data from the source domain.", + "bbox": [ + 75, + 810, + 468, + 898 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c445f87920c29a6edffe5d333512502ed2f46e5fff917f5349ef87c38e777300.jpg", + "table_caption": [], + "table_footnote": [ + "Table 2. Performance of base models on MSCOCO validation set with remapped categories and Scenes100 before any adaptation. For Scenes100, $APs$ are evaluated on each video individually and then averaged. R- stands for ResNet. # indicates the number of parameters $(M = 10^6)$ . See Sec. 5.1 for $AP$ notation details." + ], + "table_body": "
Model# paramsMSCOCOScenes100
APmcoAP50coAPmcoAP50coAPmcoAP50w
M1 (R-50)41.4M50.0576.4840.5262.1241.2864.65
M2 (R-101)60.5M51.2977.4641.1163.1041.9665.74
", + "bbox": [ + 503, + 89, + 888, + 165 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Results", + "text_level": 1, + "bbox": [ + 500, + 244, + 591, + 260 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The performance of the base models is shown in Tab. 2, along with their number of trainable parameters. M2 is a bigger model with more parameters and representation capacity, so it outperforms M1 by a noticeable margin in all the metrics before adaptation. All the $AP$ gains shown in other tables are based on the performance of base model M2. Note that the $AP$ numbers on Scenes100 is already relatively high at only about 10 points below the performance on MSCOCO, so we do not observe very high ( $>5$ points) $AP$ gains in the results shown later.", + "bbox": [ + 496, + 272, + 890, + 422 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compare the performance of different adaptation methods in terms of average $AP$ gain in Tab. 3. Here we show our best combination of methods, which incorporates pseudo-labeling, location-aware mixup, and object mask mid-fusion. Among the baseline methods for domain adaptive object detection, only ST and LODS improves the performance significantly. AT only yields marginal improvement in classes-weighted $AP$ . STAC, $\\mathrm{H}^2\\mathrm{FA}$ , and TIA actually degrade the performance on Scenes100. LODS performs surprisingly well considering it does not use source domain images. However, we found that any training schedule longer than 1000 iterations would lead to severe performance degradation. The proposed method yields much higher, consistent, and stable improvement.", + "bbox": [ + 496, + 428, + 890, + 638 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In general, self-training-based methods (ST, AT, STAC, and the proposed method) perform better than domain alignment-based methods $\\mathrm{(H^2FA}$ and TIA). Domain alignment is usually achieved through adversarial learning, which is known to have unstable objectives and is more difficult to optimize [14, 42, 59, 61]. Therefore, these methods are more sensitive to hyperparameters and training schedules, making it challenging to determine the optimal settings for all datasets. Applying domain alignment methods directly to our scene adaptive dataset without careful tuning can cause problems as our dataset has less in-domain variance than domain adaptation datasets. Furthermore, since the scenes in our dataset are very different, they might require different settings. This implies that Scenes100 has its uniqueness compared to more generally purported domain adaptive object detection datasets. Our proposed method is more robust to this setting and does not rely heavily on", + "bbox": [ + 496, + 643, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "21595", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c842656924f859237c74ca029d3a4fe34d88e85333aa90c1767bfc0ee76f263a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Method\\(APG_{co}^{m}\\)\\(APG_{co}^{50}\\)\\(APG_{w}^{m}\\)\\(APG_{w}^{50}\\)
ST [57]+0.80+0.24+1.39+1.03
STAC [60]-1.26-5.12-1.97-6.64
AT [37]-0.75-1.11+0.06+0.04
\\(\\mathrm{H^{2}FA}\\) [72]-3.10-4.97-3.77-6.01
TIA [78]-0.32-0.37-0.32-0.33
LODS [33]+0.45+1.28+1.02+2.28
Proposed+3.76+4.45+3.78+4.65
", + "bbox": [ + 114, + 88, + 431, + 232 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5f603bdc41fe198eae3473fbb0711ffb455110ba4d0fa7d514a40303b09c33c6.jpg", + "table_caption": [ + "Table 3. Averaged $AP$ gain of different adaptation methods. See Sec. 5.1 for $AP$ notation details." + ], + "table_footnote": [], + "table_body": "
MixupFusion\\(APG_{co}^{m}\\)\\(APG_{co}^{50}\\)\\(APG_{w}^{m}\\)\\(APG_{w}^{50}\\)
XX+0.95+0.54+1.67+1.55
Location-awareX+1.72+1.67+2.25+2.53
RandomX+1.22+1.13+1.76+2.03
Xearly+1.85+2.12+2.22+2.73
Xmid+3.40+3.81+3.67+3.98
Xlate+3.34+3.60+3.38+3.72
Location-awareearly+2.25+2.82+2.59+3.46
Location-awaremid+3.76+4.45+3.78+4.65
Location-awarelate+3.66+4.10+3.73+4.31
", + "bbox": [ + 81, + 279, + 465, + 460 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/4f97855a54a27a12202bc95bfb60a1c4b9023b050798dd977102582cdfca4249.jpg", + "table_caption": [ + "Table 4. Ablation study for the proposed components of our scene-adaptive object detection method. $\\times$ means not being applied. Pseudo-labeling is always applied. This table shows the averaged $AP$ gain; see Sec. 5.1 for $AP$ notation details." + ], + "table_footnote": [], + "table_body": "
MixupFusion\\(APG_{co}^{m}\\)\\(APG_{co}^{50}\\)\\(APG_{w}^{m}\\)\\(APG_{w}^{50}\\)
✘AVG+3.40+3.81+3.67+3.98
✘CNN+3.25+3.52+3.26+3.79
✘ATTN+3.22+3.73+3.23+3.95
Location-awareAVG+3.76+4.45+3.78+4.65
Location-awareCNN+3.60+4.20+3.71+4.50
Location-awareATTN+3.30+3.89+3.45+4.26
", + "bbox": [ + 81, + 536, + 465, + 661 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Ablation study on different feature pyramid fusion methods in term of averaged $AP$ gain. AVG, CNN, and ATTN indicate mid-fusion based on average pooling, CNN, or attention module, respectively. $\\pmb{x}$ means not being applied. Pseudo-labeling is always applied. See Sec. 5.1 for $AP$ notation details.", + "bbox": [ + 75, + 664, + 468, + 736 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "hyperparameter tuning, as we use the same set of hyperparameters across all 100 videos. Additional analysis and success/failure cases can be found in the supplementary material.", + "bbox": [ + 75, + 743, + 468, + 803 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.5. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 815, + 230, + 832 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We explore the effect of the different components in our methods by experimenting various combinations of them. For all experiments, we keep the pseudo-labeling unchanged, as it is the basis of other components. We compare", + "bbox": [ + 75, + 839, + 468, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "different types of object mask fusion (Fusion): early, mid, and late. We also compare the proposed Location-aware object mixup with Random mixup. Tab. 4 shows the results. The experiments on hyper-parameters can be found in the supplementary material.", + "bbox": [ + 496, + 90, + 890, + 167 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The proposed pseudo-labeling, mixup, and object mask fusion all benefit the adaptation performance. Applying only the pseudo-labeling already leads to $AP$ gain higher than ST, which is more complicated involving pseudo box weighting. Location-aware mixup is also more advantageous than Random mixup, validating the assumption in Sec. 3.3 that having fewer artifacts is better for adaptation. Among three fusion options, mid-fusion yields the best result. That is perhaps because in a mid-fusion model, the RPN utilizes the fused feature pyramid, which can contain critical information of object boundaries. The combination of mixup and object mask fusion can further improve the performance.", + "bbox": [ + 496, + 171, + 892, + 369 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We also compare different feature pyramid fusion methods described in Eq. (3). In our main experiments, we simply average the feature pyramids from the original image and the object mask. Now at each level of the pyramid, we use a separate CNN or an attention module to fuse the feature maps. The comparison is shown in Tab. 5. The parametric method performs slightly worse than average-pooling, probably because the newly introduced modules have randomly initialized weights and thus require additional supervised training data.", + "bbox": [ + 496, + 375, + 890, + 526 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Summary", + "text_level": 1, + "bbox": [ + 500, + 556, + 607, + 573 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have presented a self-supervised framework for scene-adaptive object detection. Base detectors and generic trackers are used to generate pseudo object labels as the adaptive training targets in a cross-teaching manner. To fully utilize the background equivariance when camera perspective is fixed, we propose artifacts-free location-aware object mixup to augment the input images, and dynamic background extraction for additional input modality to the detector. We also introduce the first large-scale scene adaptive object detection dataset, Scenes100, with several unique features compared to other domain adaptive object detection datasets. Our method outperforms domain adaptive object detection baselines by a large margin on Scenes100. We also conduct extensive experiments to illustrate the effectiveness of the components in our framework. We hope this work has demonstrated the importance of scene adaptation, and it will spur further research interest in this impactful but understudied area.", + "bbox": [ + 496, + 587, + 890, + 859 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This research was partially supported by NSF grants IIS-1763981 & NSDF DUE-2055406.", + "bbox": [ + 498, + 869, + 890, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "21596", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] David Berthelot, Nicholas Carlini, Ian Goodfellow, Avital Oliver, Nicolas Papernot, and Colin Raffel. Mixmatch: A holistic approach to semi-supervised learning. In Conference on Neural Information Processing Systems (NeurIPS), 2019. 3", + "[2] Goutam Bhat, Martin Danelljan, Luc Van Gool, and Radu Timofte. Learning discriminative model prediction for tracking. In International Conference on Computer Vision (ICCV), 2019. 1, 3", + "[3] Alexey Bochkovskiy, Chien-Yao Wang, and Hong-Yuan Mark Liao. Yolov4: Optimal speed and accuracy of object detection. In ArXiv, 2020. 2", + "[4] Shahine Bouabid and Vincent Delaitre. Mixup regularization for region proposal based object detectors. In ArXiv, 2020. 3", + "[5] Zhaowei Cai and Nuno Vasconcelos. Cascade r-cnn: Delving into high quality object detection. In Computer Vision and Pattern Recognition Conference (CVPR), 2018. 2", + "[6] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European Conference on Computer Vision (ECCV), 2020. 2", + "[7] Yuhua Chen, Wen Li, Christos Sakaridis, Dengxin Dai, and Luc Van Gool. Domain adaptive faster r-cnn for object detection in the wild. In Computer Vision and Pattern Recognition Conference (CVPR), 2018. 2", + "[8] Xiangxiang Chu, Zhi Tian, Yuqing Wang, Bo Zhang, Haibing Ren, Xiaolin Wei, Huaxia Xia, and Chunhua Shen. Twins: Revisiting the design of spatial attention in vision transformers. In Conference on Neural Information Processing Systems (NeurIPS), 2021. 2", + "[9] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Computer Vision and Pattern Recognition Conference (CVPR), 2016. 6", + "[10] Martin Danelljan, Goutam Bhat, Christoph Mayer, and Matthieu Paul. Pytracking. https://github.com/visionml/pytracking, 2019.6", + "[11] Jinhong Deng, Wen Li, Yuhua Chen, and Lixin Duan. Unbiased mean teacher for cross-domain object detection. In Computer Vision and Pattern Recognition Conference (CVPR), 2021. 2", + "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations (ICLR), 2021. 2", + "[13] Yoav Freund and Robert E Schapire. A decision-theoretic generalization of on-line learning and an application to boosting. Journal of Computer and System Sciences, 55(1):119-139, 1997. 1, 2" + ], + "bbox": [ + 78, + 116, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Angus Galloway, Anna Golubeva, Thomas Tanay, Medhat Moussa, and Graham W. Taylor. Batch normalization is a cause of adversarial vulnerability. In *ArXiv*, 2020. 7", + "[15] Yaroslav Ganin and Victor Lempitsky. Unsupervised domain adaptation by backpropagation. In International Conference on Machine Learning (ICML), 2015. 2, 7", + "[16] Yaroslav Ganin, E. Ustinova, Hana Ajakan, Pascal Germain, H. Larochelle, Francois Lavolette, Mario Marchand, and Victor S. Lempitsky. Domain-adversarial training of neural networks. In Journal of Machine Learning Research, 2016. 2", + "[17] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In Computer Vision and Pattern Recognition Conference (CVPR), 2012. 6", + "[18] Ross Girshick. Fast r-cnn. In International Conference on Computer Vision (ICCV), 2015. 2", + "[19] Ross Girshick, Jeff Donahue, Trevor Darrell, and Jitendra Malik. Rich feature hierarchies for accurate object detection and semantic segmentation. In Computer Vision and Pattern Recognition Conference (CVPR), 2014. 2, 4", + "[20] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In International Conference on Computer Vision (ICCV), 2017. 4", + "[21] Kaiming He, X. Zhang, Shaoqing Ren, and Jian Sun. Spatial pyramid pooling in deep convolutional networks for visual recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 37, 2015. 2", + "[22] Zhenwei He and Lei Zhang. Multi-adversarial faster-rcnn for unrestricted object detection. In International Conference on Computer Vision (ICCV), 2019. 2", + "[23] Dan Hendrycks, Norman Mu, Ekin D. Cubuk, Barret Zoph, Justin Gilmer, and Balaji Lakshminarayanan. AugMix: A simple data processing method to improve robustness and uncertainty. In International Conference on Learning Representations (ICLR), 2020. 3", + "[24] Tin Kam Ho. Random decision forests. In International Conference on Document Analysis and Recognition (ICDAR), volume 1, 1995. 1, 2", + "[25] Minh Hoai and Andrew Zisserman. Talking heads: Detecting humans and recognizing their interactions. In Computer Vision and Pattern Recognition Conference (CVPR), 2014. 2", + "[26] Jiaxing Huang, Dayan Guan, Aoran Xiao, and Shijian Lu. Model adaptation: Historical contrastive learning for unsupervised domain adaptation without source data. In Conference on Neural Information Processing Systems (NeurIPS), 2021. 2", + "[27] Naoto Inoue, Ryosuke Furuta, Toshihiko Yamasaki, and Kiyoharu Aizawa. Cross-domain weakly-supervised object detection through progressive domain adaptation. In Computer Vision and Pattern Recognition Conference (CVPR), 2018. 2", + "[28] Glenn Jocher. Yolov5. https://github.com/ultralytics/yolov5, 2020.2", + "[29] Seunghyeon Kim, Jaehoon Choi, Taekyung Kim, and Chang-ick Kim. Self-training and adversarial background regularization for unsupervised domain adaptive one-stage object detection. In International Conference on Computer Vision (ICCV), 2019. 2" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "21597", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] Samuli Laine and Timo Aila. Temporal ensembling for semi-supervised learning. In International Conference on Learning Representations (ICLR), 2017. 2", + "[31] Chuyi Li, Lulu Li, Hongliang Jiang, Kaiheng Weng, Yifei Geng, Liang Li, Zaidan Ke, Qingyuan Li, Meng Cheng, Weiqiang Nie, Yiduo Li, Bo Zhang, Yufei Liang, Linyuan Zhou, Xiaoming Xu, Xiangxiang Chu, Xiaoming Wei, and Xiaolin Wei. Yolov6: A single-stage object detection framework for industrial applications. In ArXiv, 2022. 2", + "[32] Junnan Li, Richard Socher, and Steven C.H. Hoi. Dividemix: Learning with noisy labels as semi-supervised learning. In International Conference on Learning Representations (ICLR), 2020. 3", + "[33] Shuaifeng Li, Mao Ye, Xiatian Zhu, Lihua Zhou, and Lin Xiong. Source-free object detection by learning to overlook domain style. In Computer Vision and Pattern Recognition Conference (CVPR), 2022. 2, 7, 8", + "[34] Wuyang Li, Xinyu Liu, and Yixuan Yuan. Sigma: Semantic-complete graph matching for domain adaptive object. In Computer Vision and Pattern Recognition Conference (CVPR), 2022. 2", + "[35] Xianfeng Li, Weijie Chen, Di Xie, Shicai Yang, Peng Yuan, Shiliang Pu, and Yueting Zhuang. A free lunch for unsupervised domain adaptive object detection without source data. In AAAI Conference on Artificial Intelligence (AAAI), 2021. 2", + "[36] Yanghao Li, Chao-Yuan Wu, Haoqi Fan, Karttikeya Mangalam, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. Mvity2: Improved multiscale vision transformers for classification and detection. In Computer Vision and Pattern Recognition Conference (CVPR), 2022. 2", + "[37] Yu-Jhe Li, Xiaoliang Dai, Chih-Yao Ma, Yen-Cheng Liu, Kan Chen, Bichen Wu, Zijian He, Kris Kitani, and Peter Vajda. Cross-domain adaptive teacher for object detection. In Computer Vision and Pattern Recognition Conference (CVPR), 2022. 2, 7, 8", + "[38] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE Transactions on Pattern Analysis and Machine Intelligence, 40(12):2935-2947, dec 2018. 2", + "[39] Tsung-Yi Lin, Piotr Dollar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Computer Vision and Pattern Recognition Conference (CVPR), 2017. 4", + "[40] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In International Conference on Computer Vision (ICCV), 2017. 2", + "[41] Tsung-Yi Lin, Michael Maire, Serge Belongie, Lubomir Bourdev, Ross Girshick, James Hays, Pietro Perona, Deva Ramanan, C. Lawrence Zitnick, and Piotr Dólar. Microsoft coco: Common objects in context. In ArXiv, 2014. 5, 6", + "[42] Chen Liu, Mathieu Salzmann, Tao Lin, Ryota Tomioka, and Sabine Susstrunk. On the loss landscape of adversarial training: Identifying challenges and how to overcome them. In Conference on Neural Information Processing Systems (NeurIPS), 2020. 7", + "[43] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C." + ], + "bbox": [ + 78, + 90, + 470, + 902 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Berg. Ssd: Single shot multibox detector. In European Conference on Computer Vision (ECCV), 2016. 2", + "[44] Yen-Cheng Liu, Chih-Yao Ma, Zijian He, Chia-Wen Kuo, Kan Chen, Peizhao Zhang, Bichen Wu, Zsolt Kira, and Peter Vajda. Unbiased teacher for semi-supervised object detection. In International Conference on Learning Representations (ICLR), 2021. 2", + "[45] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In International Conference on Computer Vision (ICCV), 2021. 2", + "[46] Muhammad Akhtar Munir, Muhammad Haris Khan, M. Saquib Sarfraz, and Mohsen Ali. Ssal: Synergizing between self-training and adversarial learning for domain adaptive object detection. In Conference on Neural Information Processing Systems (NeurIPS), 2021. 2", + "[47] Supreeth Narasimhaswamy, Trung Nguyen, and Minh Hoai. Detecting hands and recognizing physical contact in the wild. In Conference on Neural Information Processing Systems (NeurIPS), 2020. 2", + "[48] Supreeth Narasimhaswamy, Thanh Nguyen, Mingzhen Huang, and Minh Hoai. Whose hands are these? hand detection and hand-body association in the wild. In Computer Vision and Pattern Recognition Conference (CVPR), 2022. 2", + "[49] Supreeth Narasimhaswamy, Zhengwei Wei, Yang Wang, Justin Zhang, and Minh Hoai. Contextual attention for hand detection in the wild. In International Conference on Computer Vision (ICCV), 2019. 2", + "[50] Thanh Nguyen, Chau Pham, Khoi Nguyen, and Minh Hoai. Few-shot object counting and detection. In European Conference on Computer Vision (ECCV), 2022. 2", + "[51] Kento Nishi, Yi Ding, Alex Rich, and Tobias Höllerer. Augmentation strategies for learning with noisy labels. In Computer Vision and Pattern Recognition Conference (CVPR), 2021. 2", + "[52] German I. Parisi, Ronald Kemker, Jose L. Part, Christopher Kanan, and Stefan Wermter. Continual lifelong learning with neural networks: A review. Neural Networks, 113:54-71, 2019. 2", + "[53] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In ArXiv, 2015. 2", + "[54] Joseph Redmon and Ali Farhadi. YOLO9000: better, faster, stronger. In ArXiv, 2016. 2", + "[55] Joseph Redmon and Ali Farhadi. Yolov3: An incremental improvement. In ArXiv, 2018. 2", + "[56] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In Conference on Neural Information Processing Systems (NeurIPS), 2015. 2, 3, 4", + "[57] Aruni RoyChowdhury, Prithvijit Chakrabarty, Ashish Singh, SouYoung Jin, Huaizu Jiang, Liangliang Cao, and Erik Learned-Miller. Automatic adaptation of object detectors to new domains using self-training. In Computer Vision and Pattern Recognition Conference (CVPR), 2019. 1, 2, 3, 7, 8" + ], + "bbox": [ + 501, + 90, + 890, + 902 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "21598", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[58] Kuniaki Saito, Yoshitaka Ushiku, Tatsuya Harada, and Kate Saenko. Strong-weak distribution alignment for adaptive object detection. In Computer Vision and Pattern Recognition Conference (CVPR), 2019. 2", + "[59] Chawin Sitawarin, Supriyo Chakraborty, and David A. Wagner. Improving adversarial robustness through progressive hardening. In ArXiv, 2020. 7", + "[60] Kihyuk Sohn, Zizhao Zhang, Chun-Liang Li, Han Zhang, Chen-Yu Lee, and Tomas Pfister. A simple semi-supervised learning framework for object detection. In arXiv, 2020. 2, 7, 8", + "[61] Arvind P. Sridhar, Chawin Sitawarin, and David Wagner. Mitigating adversarial training instability with batch normalization. In International Conference on Learning Representations Workshop (ICLRW), 2021. 7", + "[62] Antti Tarvainen and Harri Valpola. Weight-averaged consistency targets improve semi-supervised deep learning results. In Conference on Neural Information Processing Systems (NeurIPS), 03 2017. 2", + "[63] Alexandru Cristian Telea. An image inpainting technique based on the fast marching method. Journal of Graphics Tools, 9:23-34, 2004. 4", + "[64] Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. FCOS: Fully convolutional one-stage object detection. In International Conference on Computer Vision (ICCV), 2019. 2", + "[65] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Conference on Neural Information Processing Systems (NeurIPS), 2017. 2", + "[66] Vikas Verma, Alex Lamb, Christopher Beckham, Amir Najafi, Ioannis Mitliagkas, David Lopez-Paz, and Yoshua Bengio. *Manifold mixup: Better representations by interpolating hidden states.* In International Conference on Machine Learning (ICML), 2019. 3", + "[67] Chien-Yao Wang, Alexey Bochkovskiy, and Hong-Yuan Mark Liao. Yolov7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors. In ArXiv, 2022. 2", + "[68] Chien-Yao Wang, I-Hau Yeh, and Hongpeng Liao. You only learn one representation: Unified network for multiple tasks. In ArXiv, 2021. 2", + "[69] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In International Conference on Computer Vision (ICCV), 2021. 2", + "[70] Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. Detector2. https://github.com/facebookresearch/detectron2, 2019.6", + "[71] Lin Xiong, Mao Ye, Dan Zhang, Yan Gan, Xue Li, and Yingying Zhu. Source data-free domain adaptation of object detector through domain-specific perturbation. International Journal of Intelligent Systems, 36:3746 - 3766, 2021. 2", + "[72] Yunqiu Xu, Yifan Sun, Zongxin Yang, Jiaxu Miao, and Yi Yang. H $^2$ FA R-CNN: Holistic and hierarchical feature alignment for cross-domain weakly supervised object detection. In Computer Vision and Pattern Recognition Conference (CVPR), 2022. 2, 7, 8" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[73] Jihan Yang, Shaoshuai Shi, Zhe Wang, Hongsheng Li, and Xiaojuan Qi. St3d: Self-training for unsupervised domain adaptation on 3d object detection. In Computer Vision and Pattern Recognition Conference (CVPR), 2021. 2", + "[74] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darryll. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In Computer Vision and Pattern Recognition Conference (CVPR), 2020. 6", + "[75] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In International Conference on Computer Vision (ICCV), 2019. 3", + "[76] Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In ArXiv, 2017. 1, 3", + "[77] Pengchuan Zhang, Xiyang Dai, Jianwei Yang, Bin Xiao, Lu Yuan, Lei Zhang, and Jianfeng Gao. Multi-scale vision longformer: A new vision transformer for high-resolution image encoding. In International Conference on Computer Vision (ICCV), 2021. 2", + "[78] Liang Zhao and Limin Wang. Task-specific inconsistency alignment for domain adaptive object detection. In Computer Vision and Pattern Recognition Conference (CVPR), 2022. 2, 7, 8", + "[79] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable DETR: deformable transformers for end-to-end object detection. In ArXiv, 2020. 2" + ], + "bbox": [ + 501, + 92, + 890, + 500 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "21599", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Object Detection With Self-Supervised Scene Adaptation/b774b6cf-eacd-4660-845f-f32b5ec4dc83_model.json b/2023/Object Detection With Self-Supervised Scene Adaptation/b774b6cf-eacd-4660-845f-f32b5ec4dc83_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e7e34cc7ed211c2f59b9aa441cf0d7249358b592 --- /dev/null +++ b/2023/Object Detection With Self-Supervised Scene Adaptation/b774b6cf-eacd-4660-845f-f32b5ec4dc83_model.json @@ -0,0 +1,2807 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.202, + 0.131, + 0.77, + 0.154 + ], + "angle": 0, + "content": "Object Detection with Self-Supervised Scene Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.324, + 0.18, + 0.644, + 0.199 + ], + "angle": 0, + "content": "Zekun Zhang1 Minh Hoai1,2" + }, + { + "type": "text", + "bbox": [ + 0.263, + 0.205, + 0.708, + 0.223 + ], + "angle": 0, + "content": "1Stony Brook University, Stony Brook, NY 11794, USA" + }, + { + "type": "text", + "bbox": [ + 0.35, + 0.224, + 0.621, + 0.24 + ], + "angle": 0, + "content": "\\(^{2}\\)VinAI Research, Hanoi, Vietnam" + }, + { + "type": "text", + "bbox": [ + 0.32, + 0.244, + 0.647, + 0.258 + ], + "angle": 0, + "content": "{zekzhang,minhhooai}@cs.stonyrbook.edu" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.294, + 0.314, + 0.31 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.327, + 0.474, + 0.599 + ], + "angle": 0, + "content": "This paper proposes a novel method to improve the performance of a trained object detector on scenes with fixed camera perspectives based on self-supervised adaptation. Given a specific scene, the trained detector is adapted using pseudo-ground truth labels generated by the detector itself and an object tracker in a cross-teaching manner. When the camera perspective is fixed, our method can utilize the background equivariance by proposing artifact-free object mixup as a means of data augmentation, and utilize accurate background extraction as an additional input modality. We also introduce a large-scale and diverse dataset for the development and evaluation of scene-adaptive object detection. Experiments on this dataset show that our method can improve the average precision of the original detector, outperforming the previous state-of-the-art self-supervised domain adaptive object detection methods by a large margin. Our dataset and code are published at https://github.com/cvlab-stonybrook/scenes100." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.631, + 0.21, + 0.647 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.658, + 0.47, + 0.839 + ], + "angle": 0, + "content": "The need to detect objects in video streams from stationary cameras arises in many computer vision applications, including video surveillance and autonomous retail. In general, different applications require the detection of different object categories, and each computer-vision-based product will have its own detector. However, for a specific product, there is typically a single detector that will be used for many cameras/scenes. For example, a typical video surveillance product would use the same detector to detect pedestrians and vehicles for network cameras installed at different locations. Unfortunately, a single detector might not work well for all scenes, leading to trivial and unforgiving mistakes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.902 + ], + "angle": 0, + "content": "This fundamental problem of many computer vision products stems from the limited generalization power of a single model, due to limited training data, limited model capacity, or both. One can attempt to address this problem by" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.295, + 0.892, + 0.386 + ], + "angle": 0, + "content": "using more training data, but it will incur additional cost for data collection and annotation. Furthermore, in many cases, due to the low latency requirement or the limited computing resources for inference, a product is forced to use a very lightweight network, and this network will have limited representation capacity to generalize across many scenes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.387, + 0.892, + 0.476 + ], + "angle": 0, + "content": "In this paper, instead of having a single scene-generic detector, we propose using scene-specific detectors. This yields higher detection performance as each detector is customized for a specific scene, and allows us to use a lightweight model without sacrificing accuracy as each detector is only responsible for one scene." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.478, + 0.893, + 0.749 + ], + "angle": 0, + "content": "Obtaining scene-specific detectors, however, is very challenging. A trivial approach is to train a detector for each scene separately, but this requires an enormous amount of annotated training data. Instead, we propose a self-supervised method to adapt a pre-trained detector to each scene. Our method records the unlabeled video frames in the past, uses the trained detector to detect objects in those frames, and generates augmented training data based on those detections. Although the detections made by the pre-trained model can be noisy, they can still be useful for generating pseudo annotated data. We further extend those pseudo bounding boxes by applying object tracking [2, 57] along the video timeline, aiming to propagate the detections to adjacent frames to recover some of the false negatives not returned by the detector. We also use multiple detectors to obtain the pseudo labels and train the detector in a cross-teaching manner, taking the advantage of the ensemble of models [13, 24]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.87 + ], + "angle": 0, + "content": "Exploiting the stationary nature of the camera, we propose two additional techniques to boost the detection performance: location-aware mixup and background-augmented input. The former is to generate more samples during training through object mixup [76] that contains less artifacts, based on the aforementioned pseudo boxes generated from detection and tracking. The latter involves estimating the background image and fusing it with the detector's input." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In short, the main contribution of our paper is a novel framework that utilizes self-supervision, location-aware ob" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21589" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.184 + ], + "angle": 0, + "content": "ject mixup, and background modeling to improve the detection performance of a pre-trained object detector on scenes with stationary cameras. We also contribute a large scale and diverse dataset for the development of scene adaptive object detection, which contains sufficient quantity and quality annotations for evaluation." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.196, + 0.22, + 0.212 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.222, + 0.47, + 0.401 + ], + "angle": 0, + "content": "Despite much recent improvement in object detection research [3,5,6,8,12,18,19,21,25,28,31,36,40,43,45,47-50, 53-56,64,65,67-69,77,79], trained detectors still encounter problems with domain shift or domain gap. Scene adaptive object detection can be viewed as a special case of domain adaption, in which an object detector trained on a fully-supervised source domain is adapted to a target domain. Most research in this direction focuses on semi-supervised, weakly-supervised, or self-supervised adaptation. In addition, source-free adaptation [26, 33, 35, 71] has been proposed to address the situation where the source domain data is unavailable during adaptation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.403, + 0.471, + 0.659 + ], + "angle": 0, + "content": "Self-labeling methods [11,27,29,33,37,44,46,57,60,73] uses the teacher-student setup from semi-supervised image classification [30, 62]. A teacher detector trained on the source domain generates pseudo bounding boxes on the target domain images, and a student model is trained with those boxes to improve its performance on the target domain. Techniques such as weak/strong augmentation [37,44,60], knowledge distillation [11], and weight averaging [33, 37] have been used to deal with noisy pseudo labels from the teacher detector due to the domain gap. Our proposed pseudo-labeling method uses two teacher models trained on the source domain to train a student model, and their pseudo boxes are aggregated and refined. We further apply tracking as in [57], but with both forward and backward directions, to extend the pseudo labels. The pseudo boxes also form the basis for later location-aware mixup and dynamic background extraction steps." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.47, + 0.78 + ], + "angle": 0, + "content": "Domain alignment methods [7, 11, 22, 33, 34, 58, 72, 78] aim to reduce the domain gap by enforcing the models to output similarly on the source and target domain at image, proposal, or instance levels. They use domain adversarial learning through gradient reversal [15, 16] or graph matching [34] for domain alignment. This approach involves adding domain alignment losses to existing object detection models and is complementary to our proposed method." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Another seemingly related research area is continual or incremental learning. Scene adaptive and domain adaptive object detection are in between task-continual learning [38] and data-continual learning [52], for that the object categories of the source and target domains are the same, but the data distribution differs. It is possible to formulate scene adaptive object detection as an online learning problem, but this approach would require additional annotations." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.087, + 0.685, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.538, + 0.173, + 0.636, + 0.182 + ], + "angle": 0, + "content": "(a) Input image \\(\\mathcal{I}\\)." + }, + { + "type": "image", + "bbox": [ + 0.704, + 0.092, + 0.886, + 0.172 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.743, + 0.173, + 0.871, + 0.192 + ], + "angle": 0, + "content": "(b) Learned background image \\(\\mathcal{B}\\)." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.183, + 0.684, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.534, + 0.261, + 0.653, + 0.273 + ], + "angle": 0, + "content": "(c) Object mask \\(\\mathcal{M}^O\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.28, + 0.892, + 0.323 + ], + "angle": 0, + "content": "Figure 1. Inference flow of the proposed detector that can adapt to a specific scene. The background image and fusion faster-RCNN model are learned during training, discussed in Sec. 3." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.329, + 0.599, + 0.345 + ], + "angle": 0, + "content": "3. Methods" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.355, + 0.892, + 0.476 + ], + "angle": 0, + "content": "For an object detector pre-trained on a source dataset referred as the base model, we aim to improve its performance on video stream from a stationary camera in a given scene, with unlabeled videos captured by the same camera in the past as adaptation training samples. The categories of object of interest in the scene are also present in the source dataset, but the data distribution is shifted. The base model is adapted to each scene independently." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.477, + 0.893, + 0.644 + ], + "angle": 0, + "content": "Assuming the distribution shift is moderate, the base detector can generate partially decent bounding boxes on the unlabeled frames. We use them as pseudo-labels to improve the base model itself. Many objects are ignored by the base detector, as their appearance can be very different from the source dataset. However, videos contain temporal correlation information, making it possible to apply object tracking in both time directions to recover some of those false negatives. The pseudo-labels are then refined to remove low-confidence boxes and duplicates. We refer to this process of obtaining pseudo bounding boxes as pseudo-labeling." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.892, + 0.78 + ], + "angle": 0, + "content": "As data augmentation is beneficial for training on less reliable labels [51], we take advantages of object mixup to generate new sample images. Since the background is stationary, those images contain fewer artifacts, which is beneficial for detectors. We further adopt the idea of ensemble models [13, 24] by using two base detectors to obtain the pseudo labels, for they can be complementary to each other. The pseudo boxes are used to train one base model, which we refer as cross-teaching." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Since the job of a detector is to separate foreground objects from the background, having information about the background is advantageous. In a stationary camera video with moving objects, different parts of the background are covered by objects in different frames, making it possible to model the complete background. We use the aforementioned pseudo bounding boxes to mark the uncovered parts of the background in each frame. Combining the partial" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21590" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.085, + 0.895, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.252, + 0.895, + 0.281 + ], + "angle": 0, + "content": "Figure 2. Training flow of the proposed self-supervised scene adaptation object detection framework with sample images. The mixup training images with pseudo bounding boxes and background image are used in the fusion training. The details are discussed in Sec. 3." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.307, + 0.47, + 0.367 + ], + "angle": 0, + "content": "background from a sequence of frames can give an accurate and complete background model. We then modify the Faster-RCNN [56] architecture to fuse the background with the input image to improve detection performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.367, + 0.473, + 0.548 + ], + "angle": 0, + "content": "The overall structure and data flows of the proposed methods are shown in Fig. 1 and Fig. 2. Our proposed methods are highly flexible and can be configured to trade-off between performance and speed. Object tracking can be bypassed to reduce the time to get the pseudo-bounding boxes. Mixup and object mask fusion can also be disabled for faster training and inference. In Sec. 5.5 we will show ablation study results on how those components affect the performance of the adapted models. We now describe each component in more details. For consistency, the images are all from the first video in our collected dataset, which will be described in Sec. 4." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.558, + 0.434, + 0.573 + ], + "angle": 0, + "content": "3.1. Base Detection Models on Source Domain" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.581, + 0.47, + 0.701 + ], + "angle": 0, + "content": "We use Faster-RCNN [56] in our experiments for its high precision and flexible modular architecture. Among the techniques proposed for self-supervised scene adaptation, both pseudo-labeling and mixup can be directly applied to any object detection model. The object mask fusion technique requires modification to the network architecture, but it can be easily extended to most types of detectors by adding a parallel branch to the existing network." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.702, + 0.47, + 0.807 + ], + "angle": 0, + "content": "The object labels in MSCOCO training set are remapped to person and vehicle as described in Sec. 4.2. All objects other than person, car, bus, and truck are discarded. We take Faster-RCNNs pre-trained on the original MSCOCO training set, and finetune them on the remapped training set. We choose two models with ResNet-50 and ResNet101 backbones as base models, referred as M1 and M2." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.817, + 0.452, + 0.833 + ], + "angle": 0, + "content": "3.2. Pseudo Labels from Detection and Tracking" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.901 + ], + "angle": 0, + "content": "We take the frames from the first training portion of each video, and adopt the idea of detection and tracking similar to [57], then apply our proposed bounding box refinement and cross-teaching schemes. Base detectors described in" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.307, + 0.892, + 0.351 + ], + "angle": 0, + "content": "Sec. 3.1 are applied on the frames, and any bounding box with confidence score higher than \\(\\lambda_{det}\\) is kept as a valid pseudo box." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.352, + 0.893, + 0.503 + ], + "angle": 0, + "content": "Single-object trackers are initialized with detected objects having confidence scores higher than \\(\\lambda_{\\text{tot}}\\). We use DiMP-50 [2] to track the objects in both forward and backward directions in the video timeline, and the tracked object bounding boxes are used as pseudo boxes. The reason for tracking in both directions is that if an object is moving towards the camera and appears bigger by time, backward tracking is more accurate because the initial bounding box contains more detailed texture. To prevent drifting, the maximum length of each track is capped at two seconds." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.504, + 0.895, + 0.715 + ], + "angle": 0, + "content": "Bounding box refinement is applied to eliminate duplicated boxes of the same object from combining pseudo bounding boxes from detector and tracker. Each candidate box in the same frame is regarded as a graph node. If two boxes have the same object label, and their IoU is above a threshold \\(\\lambda_{iou}\\), then an edge is added to connect them. We assume that each connected component of the graph represents one single object, and the nodes in the component are duplicated. We only keep the node with the highest degree (i.e., the node with the highest number of connected edges). An example of refinement results is shown in Fig. 2b. Refinement can effectively remove duplicated bounding boxes. However, it can neither remove false positives nor recover false negatives." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.715, + 0.894, + 0.777 + ], + "angle": 0, + "content": "Both M1 and M2 are used in detection and tracking. The pseudo bounding boxes from both base models are then refined jointly, and the refined boxes are used to train M2 in adaptation. We refer this setting as cross-teaching." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.787, + 0.771, + 0.803 + ], + "angle": 0, + "content": "3.3. Location-Aware Object Mixup" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Mixup [76] is originally proposed as a simple but powerful data augmentation method in for image classification. It is quickly improved by many modifications [1, 23, 32, 66, 75]. Mixup is also used in object detection tasks [4, 75], where it can be applied at image or instance levels. During adaptation training, a frame with its pseudo bounding" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "21591" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.273, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.125, + 0.177, + 0.228, + 0.189 + ], + "angle": 0, + "content": "(a) Random mixup." + }, + { + "type": "image", + "bbox": [ + 0.275, + 0.089, + 0.47, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.301, + 0.177, + 0.442, + 0.189 + ], + "angle": 0, + "content": "(b) Location-aware mixup." + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.194, + 0.47, + 0.222 + ], + "angle": 0, + "content": "Figure 3. An example of location-aware object mixup. Compared to random mixup, less artifacts are introduced." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.231, + 0.473, + 0.442 + ], + "angle": 0, + "content": "boxes has the probability \\( p_{\\text{mixup}} \\) of being paired with another randomly select source frame for mixup. A portion \\( r_{\\text{mixup}} \\) of the pseudo bounding boxes in the source frame are cropped out and pasted onto the original frame, along with their locations and labels. If a pasted box covers more than \\( \\alpha_{\\text{cover}} \\) of a pseudo box in the original frame, the pseudo box will be excluded from the pseudo labels. An example of this location-aware object mixup is shown in Fig. 2c and Fig. 3. Since the background is stationary, we propose to use location-aware mixup instead of the original random mixup to reduce the amount of artifacts in the generated mixup images and to keep the location distribution of the objects in the scene. This will be shown to be beneficial in the experiments." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.454, + 0.37, + 0.47 + ], + "angle": 0, + "content": "3.4. Dynamic Background Extraction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.478, + 0.47, + 0.612 + ], + "angle": 0, + "content": "For a video frame \\(\\mathcal{I}\\) of dimension \\(W\\times H\\) associated with a set of \\(K\\) pseudo-annotated object bounding boxes \\(\\{(x_k^1,y_k^1,x_k^2,y_k^2)|k = 1\\dots K\\}\\), a background mask \\(\\mathcal{M}\\) of the same dimension as \\(\\mathcal{I}\\) can be constructed as follows. For each location \\((x,y)\\) in \\(\\mathcal{M}\\), we set \\(M[x,y] = 0\\) if \\((x,y)\\) is inside of any pseudo-annotated bounding box and 1 otherwise. Then for a sequence of frame-mask pairs \\(\\{(\\mathcal{I}_l,\\mathcal{M}_l)|l = 1\\dots L\\}\\), the background image is determined as" + }, + { + "type": "equation", + "bbox": [ + 0.2, + 0.612, + 0.47, + 0.651 + ], + "angle": 0, + "content": "\\[\n\\mathcal {B} = \\frac {\\sum_ {l = 1} ^ {L} \\mathcal {I} _ {l} \\otimes \\mathcal {M} _ {l}}{\\sum_ {l = 1} ^ {L} \\mathcal {M} _ {l}}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.658, + 0.408, + 0.673 + ], + "angle": 0, + "content": "where \\(\\otimes\\) is the pixel-wise multiplication operator." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.673, + 0.47, + 0.764 + ], + "angle": 0, + "content": "There might exist a location \\((x', y')\\) that lies inside an object bounding box in every image, i.e., \\(\\mathcal{M}_l[x', y'] = 0\\). In this case, the background at this location is never observed and its pixel value cannot be determined. In this case, we \"guess\" the background value by inpainting inside such areas using the inpainting algorithm of [63]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.469, + 0.855 + ], + "angle": 0, + "content": "Fig. 2d shows an example background, which is reasonably accurate. Although in each video, the camera perspective is fixed, the background can change overtime due to some factors such as illumination change. To incorporate this, the background extraction is operated every \\( T_{bg} \\) seconds, giving a dynamic background modeling." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.902 + ], + "angle": 0, + "content": "We construct an object mask as an additional input modality for the detector, to utilize the extracted background to improve detection performance during adapta" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "tion. For a frame \\(\\mathcal{I}\\) and its corresponding background model image \\(\\mathcal{B}\\), we define its object mask image as" + }, + { + "type": "equation", + "bbox": [ + 0.603, + 0.128, + 0.892, + 0.145 + ], + "angle": 0, + "content": "\\[\n\\mathcal {M} ^ {O} = (\\mathcal {I} - \\mathcal {B} + 1) \\times 0. 5, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.152, + 0.892, + 0.227 + ], + "angle": 0, + "content": "where we assume the pixel values in both \\(\\mathcal{I}\\) and \\(\\mathcal{B}\\) have been normalized to the range [0, 1]. An object mask example is shown in Fig. 1c, in which objects are clearly separated from the background. Note that since \\(\\mathcal{I},\\mathcal{B}\\), and \\(\\mathcal{M}^O\\) are linearly dependent, no information is lost." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.228, + 0.892, + 0.349 + ], + "angle": 0, + "content": "Before the adaptation training process, the base detector needs to be modified to use object mask input modality and trained on the MSCOCO training set with background images. However, dynamic background extraction on MSCOCO is impossible since it only consists of static images. Therefore, we rely on the object mask annotations in the dataset and apply the same inpainting algorithm [63] to generate the background image for each training image." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.356, + 0.693, + 0.372 + ], + "angle": 0, + "content": "3.5. Object Mask Fusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.379, + 0.892, + 0.575 + ], + "angle": 0, + "content": "The architecture of Faster-RCNN [56] is depicted in Fig. 4a. First, a CNN backbone with a feature pyramid network (FPN) [39] is utilized to extract the Feature Pyramid (FP). The first layer in the CNN that takes the input image is also shown here. Then, an RPN [39] is employed to produce object bounding box proposals. Lastly, an ROI head [19,20] is used to assign object labels and refine the bounding boxes on the pooled feature maps of the proposals. The entire network is trained utilizing two sets of losses: localization loss and objectness loss from RPN, and localization loss and classification loss from the ROI head. The fusion models explained below start with the base model that is trained on the remapped MSCOCO training set." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.576, + 0.892, + 0.757 + ], + "angle": 0, + "content": "Early-fusion. The input to the Faster-RCNN network is the stacked image \\([\\mathcal{I};\\mathcal{M}^{\\mathcal{O}}]\\), with the two modalities stacked along the color channels, as shown in Fig. 4b. To enable the network to take a 6-channel image as input instead of the vanilla 3-channel image, the first layer of the backbone CNN is duplicated and stacked. The convolution kernel weights are first copied, then halved. This ensures that the feature map after the first layer conv1|conv2 is the same as in the vanilla model when \\([\\mathcal{I};\\mathcal{I}]\\) is fed into it, providing a smooth start for adaptation training. This modification introduces only a small number of additional model parameters, and the increase in computational cost is negligible." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.758, + 0.892, + 0.877 + ], + "angle": 0, + "content": "Mid-fusion. Both \\(\\mathcal{I}\\) and \\(\\mathcal{M}^O\\) are fed into the FPN in parallel, which yields two feature pyramid FP1 and FP2, as shown in Fig. 4c. Two sets of RPNs and ROI heads are used in parallel, and their initial weights are copied from the base model. FP1 is used in the left branch with RPN1 and ROI1, which is the same as the vanilla model. In the right branch with RPN2 and ROI2, the input feature pyramid is the fusion of FP1 and FP2:" + }, + { + "type": "equation", + "bbox": [ + 0.614, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "\\[\n\\mathrm {F P} _ {\\text {f u s i o n}} = \\mathrm {F P 1} \\oplus \\mathrm {F P 2}. \\tag {3}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21592" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.125, + 0.089, + 0.22, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.108, + 0.176, + 0.243, + 0.188 + ], + "angle": 0, + "content": "(a) Vanilla Faster-RCNN." + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.19, + 0.274, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.098, + 0.277, + 0.255, + 0.289 + ], + "angle": 0, + "content": "(c) Mid-fusion Faster-RCNN." + }, + { + "type": "image", + "bbox": [ + 0.318, + 0.089, + 0.437, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.289, + 0.176, + 0.454, + 0.188 + ], + "angle": 0, + "content": "(b) Early-fusion Faster-RCNN." + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.19, + 0.469, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.293, + 0.277, + 0.451, + 0.289 + ], + "angle": 0, + "content": "(d) Late-fusion Faster-RCNN." + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.294, + 0.47, + 0.336 + ], + "angle": 0, + "content": "Figure 4. Illustration of different fusion models for Faster-RCNN. Network modules are indicated by green, and feature maps and outputs are indicated by yellow." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.345, + 0.47, + 0.419 + ], + "angle": 0, + "content": "The fused feature pyramid has the same dimension as both FP1 and FP2. The fusion operator can be either non-parametric such as average pooling, or parametric such as convolution layers. The whole network is trained with losses from both branches" + }, + { + "type": "equation", + "bbox": [ + 0.078, + 0.424, + 0.476, + 0.457 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {m i d} = \\left(1 - \\alpha_ {m i d}\\right) \\left(\\mathcal {L} _ {r p n} ^ {\\mathcal {I}} + \\mathcal {L} _ {r o i} ^ {\\mathcal {I}}\\right) + \\alpha_ {m i d} \\left(\\mathcal {L} _ {r p n} ^ {f u s i o n} + \\mathcal {L} _ {r o i} ^ {f u s i o n}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.458, + 0.378, + 0.472 + ], + "angle": 0, + "content": "where \\(\\alpha_{mid}\\) is the weights of different losses." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.473, + 0.469, + 0.548 + ], + "angle": 0, + "content": "Late-fusion is similar to mid-fusion as shown in Fig. 4d. The key difference is that only the ROI head is branched. The region proposals are only generated from FP1, and the fused feature pyramid is only used in ROI head in the right branch. As a result, the total loss is" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.553, + 0.469, + 0.573 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {l a t e} = \\mathcal {L} _ {r p n} ^ {\\mathcal {I}} + (1 - \\alpha_ {l a t e}) \\mathcal {L} _ {r o i} ^ {\\mathcal {I}} + \\alpha_ {l a t e} \\mathcal {L} _ {r o i} ^ {f u s i o n}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.593, + 0.469, + 0.682 + ], + "angle": 0, + "content": "Both mid-fusion and late-fusion introduce significant additional parameters and computation. \\(\\mathcal{I}\\) and \\(\\mathcal{M}^O\\) need to pass the backbone, requiring more computation even at inference time. The quantitative comparison of inference speed of different fusion models can be found in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.684, + 0.469, + 0.773 + ], + "angle": 0, + "content": "To train the fusion models in adaptation, we first fine-tune the base models on MSCOCO training set with the inpainted background models described in Sec. 3.4 until convergence. During evaluation, we use the last background image to obtain the object masks, as it is the closest to the evaluation images on the timeline." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.785, + 0.255, + 0.801 + ], + "angle": 0, + "content": "4. Scenes100 Dataset" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.469, + 0.9 + ], + "angle": 0, + "content": "For our proposed scene-adaptive object detection problem, there is no existing dataset with long enough videos of stationary backgrounds for the development and evaluation of self-supervised adaptation techniques. We therefore collected a new dataset called Scenes100, which will be described in this section." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.089, + 0.63, + 0.145 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.145, + 0.63, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.201, + 0.63, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.258, + 0.63, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.633, + 0.09, + 0.759, + 0.145 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.633, + 0.145, + 0.759, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.633, + 0.201, + 0.759, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.633, + 0.258, + 0.759, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.763, + 0.09, + 0.89, + 0.145 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.763, + 0.145, + 0.89, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.763, + 0.201, + 0.89, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.763, + 0.258, + 0.89, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.317, + 0.892, + 0.373 + ], + "angle": 0, + "content": "Figure 5. Some annotated frames from the evaluation portions. The green mask shows the part of the frame that is excluded for annotation and evaluation. Objects are labeled by bounding boxes. Please see the supplementary material for detailed descriptions." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.381, + 0.656, + 0.396 + ], + "angle": 0, + "content": "4.1. Data Collection" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.406, + 0.892, + 0.648 + ], + "angle": 0, + "content": "We used keywords such as \"live view\", \"street camera\", and \"live webcam\" to search for live streams on YouTube. We look for streams that are of decent quality and have fixed camera perspective. Around 200 candidate videos were recorded between September 2020 and February 2022. We picked 100 videos to compile our Scenes100 dataset, ensuring that each video is longer than two hours and has minimum width and height of 720 pixels. We also aimed for a dataset with high diversities. The 100 videos of Scenes100 were recorded in a variety of 16 countries and territories, and during different time periods of day and year. The dataset contain videos from both indoor and outdoor scenes, with different camera perspectives. The density and scale of visual objects also vary significantly from one video to another. Some sample frames is presented in Fig. 5 to show the diversity of our dataset." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.663, + 0.688, + 0.677 + ], + "angle": 0, + "content": "4.2. Manual Annotation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.688, + 0.891, + 0.808 + ], + "angle": 0, + "content": "We are interested in detecting objects from two categories: person and vehicle. The person category includes any visible and recognizable people in the frames; this corresponds to the person category in MSCOCO [41] dataset. The vehicle category includes all vehicles with four or more wheels, so motorbikes, bicycles, tricycles are not included. This category corresponds to the car, bus, and truck categories in MSCOCO." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.901 + ], + "angle": 0, + "content": "For each video, we used the first 1.5 hours for self-supervised and weakly-supervised training, and the remaining part for evaluation. Since many videos were taken with wide-angle lenses, some frames included faraway objects that were too small to be considered in evaluation. To address this, we manually drew polygon masks to exclude" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "21593" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.079, + 0.089, + 0.468, + 0.194 + ], + "angle": 0, + "content": "
datasetcontain videosaverage lengthframescountriesboundong boxesboxes per video
MSCOCO [41]No---897K-
KITTI [17]No--180K-
BDD100K [74]Yes40s120M11.8M18
CityScapes [9]Yes1.8s150K265K13
Scenes100Yes2h21.6M1684K840
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.197, + 0.47, + 0.226 + ], + "angle": 0, + "content": "Table 1. Comparison of Scenes100 with other object detection datasets \\((\\mathbf{K} = 10^{3},\\mathbf{M} = 10^{6})\\)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.234, + 0.471, + 0.417 + ], + "angle": 0, + "content": "these faraway parts from annotation and evaluation. We then estimated the number of visual objects in the unmasked parts by first using a pre-trained detector, and then refining the estimate through human annotation. Frames for annotation were uniformly sampled from the second portion of each video, with the number of frames being inversely proportional to the estimated number of objects in the frames. This ensured that each video had roughly the same number of annotated objects. We contracted a data annotation company to perform bounding box annotation on those sampled frames. Some annotated frame samples for evaluation is shown in Fig. 5." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.425, + 0.246, + 0.439 + ], + "angle": 0, + "content": "4.3. Dataset Statistics" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.448, + 0.47, + 0.599 + ], + "angle": 0, + "content": "In Tab. 1, we compare Scenes100 with some popular object detection datasets. MSCOCO [41] is a fully-supervised general-purpose object detection dataset, which acts as the supervised source domain. KITTI [17] only contains object detection annotation associated to individual images. In BDD100K [74], object detection annotation is given for one key frame in each video. For CityScapes [9], we consider the finely labeled instance-level segmentation masks, which are given for the \\(20^{\\mathrm{th}}\\) image from a 30 frame video snippets lasting 1.8s." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.47, + 0.795 + ], + "angle": 0, + "content": "The scale of Scenes100 is comparable to that of KITTI and CityScapes. However, Scenes100 has three unique features. First, the videos in Scenes100 are much longer than those in any other datasets with scenes, providing sufficient data for a model to adapt to each of them. Second, Scenes100 has greater diversity. Unlike some other datasets that are recorded in only a few cities with similar camera angles, background environments, and object types, Scenes100 covers several different countries, with varying weather conditions, road conditions, view of field, camera perspectives, and vehicle types. Third, each video in Scenes100 has a fixed camera perspective and stationary background scene, making background modeling possible." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Note that many currently available datasets are not primarily intended for self-supervised adaptive object detection tasks. They may possess other features, such as segmentation masks, multi-modal sensor data, depth maps, or annotations for object tracking. However, despite the absence of these attributes, Scenes100 holds value as the first extensive and varied dataset that features fixed camera per" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.138 + ], + "angle": 0, + "content": "spectives and lengthy videos. As such, it can complement existing dataset collections and serve as a valuable resource for researchers exploring scene adaptive object detection." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.15, + 0.634, + 0.167 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.175, + 0.697, + 0.19 + ], + "angle": 0, + "content": "5.1. Evaluation Protocols" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.198, + 0.892, + 0.272 + ], + "angle": 0, + "content": "We evaluate the performance of object detectors following the COCO evaluation protocol [41] to calculate the average precision at \\( IoU = 50\\% \\) (\\( AP^{50} \\)) and mean average precision of different \\( IoU \\) thresholds from 0.5 to 0.95 (\\( AP^m \\)), with some modifications as follows." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.274, + 0.892, + 0.378 + ], + "angle": 0, + "content": "Non-evaluation region: before feeding the ground-truth and detected bounding boxes to the COCO evaluator, the bounding boxes that have at least one corner inside the nonevaluation mask will be removed. So the faraway parts in the frames where objects are deemed too small and blurry will not affect the evaluation results. Please see Sec. 4.2 and Fig. 5 for more details." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.38, + 0.893, + 0.604 + ], + "angle": 0, + "content": "Category weighting: in the standard COCO evaluation protocol, the overall multi-class \\( AP \\) is the simple average of the \\( AP_{s} \\) of the different categories. If a class is sparse in the training set and validation set, the model tends to perform poorly due to the scarcity of the training samples. In this case the average \\( AP \\) does not truly reflect the performance of a detector. Hence we propose an additional multi-class \\( AP \\) metric with weighted average where the weight is determined by the prevalence of ground truth object instances in the evaluation set. This weighted \\( AP \\) can better portray the performance of a detector in the case of long-tailed distribution. We refer the standard classes-mean COCO \\( AP \\) as \\( AP_{co} \\), and the proposed weighted \\( AP \\) as \\( AP_{w} \\). We refer the \\( AP \\) averaged over IoU thresholds and IoU = 50% as \\( AP^{m} \\) and \\( AP^{50} \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.606, + 0.892, + 0.742 + ], + "angle": 0, + "content": "Per-video evaluation: we aim to evaluate the performance of a scene adaption method, instead of a specific detection model. For each of the video, a model is adapted to it from the same base model trained on the source dataset. The performance of the base model and adapted model on each video is calculated individually as \\(\\{(AP_{v,\\text{base}}, AP_{v,\\text{adapt}}) | v = 1 \\ldots 100\\}\\). Then we evaluate the overall effectiveness of the adaptation method using the averaged \\(AP\\) gain as:" + }, + { + "type": "equation", + "bbox": [ + 0.555, + 0.752, + 0.892, + 0.793 + ], + "angle": 0, + "content": "\\[\nA P G = \\frac {1}{1 0 0} \\sum_ {v = 1} ^ {1 0 0} \\left(A P _ {v, a d a p t} - A P _ {v, b a s e}\\right). \\tag {6}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.802, + 0.719, + 0.818 + ], + "angle": 0, + "content": "5.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Our implementation is based on Detector2 [70]. The DiMP tracker is directly taken from the official PyTracking [10] implementation without any change. More details and hyper-parameters can be found in the supplementary material. During training, we include the same number of" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21594" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.198 + ], + "angle": 0, + "content": "images from MSCOCO training set as the number of unlabeled frames from the videos. The hyper-parameters are the same for all videos. We check the training loss for all experiments, including the baseline methods, to ensure convergence. We also train \"compound\" models where all videos are used together for adaptation. The results and further details can be found in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.213, + 0.342, + 0.23 + ], + "angle": 0, + "content": "5.3. Domain Adaptation Baselines" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.239, + 0.47, + 0.465 + ], + "angle": 0, + "content": "We compare our methods with several domain adaptive object detection baselines. In their original papers, the methods are evaluated on other domain adaptive object detection datasets. However, those datasets do not contains videos with fixed camera perspectives, so the proposed location-aware mixup and dynamic background extraction cannot be applied. We instead apply the methods in other papers on Scenes100. For a fair comparison, all methods start with the same base model and weights, and then are adapted to and evaluated on each scene individually. We keep the values of the hyper-parameters that are specific to a baseline method used in the original papers. We use the same learning rate, batch size, and number of iterations as our experiments in those baseline methods. More details can be found in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.467, + 0.47, + 0.543 + ], + "angle": 0, + "content": "Self-Train (ST) [57] is a method that uses detection and tracking to obtain pseudo bounding boxes, similar to ours. But ST applies tracking only in one direction. It assigns different weights to the pseudo bounding boxes during training, but no refinement nor cross-teaching." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.545, + 0.469, + 0.59 + ], + "angle": 0, + "content": "STAC [60] uses the base model to get pseudo bounding boxes, and trains the model on a strongly augmented version of the target domain image with those boxes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.593, + 0.469, + 0.667 + ], + "angle": 0, + "content": "Adaptive Teacher (AT) [37] also uses self-training, but the pseudo bounding boxes are detected on the fly. It utilizes exponential moving average to update the model gradually. It also uses weak/strong data augmentation, and a domain classifier to align the features of the two domains." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.669, + 0.469, + 0.745 + ], + "angle": 0, + "content": "\\(\\mathbf{H}^2\\mathbf{FA}\\) R-CNN [72] applies feature alignment of source and target domains at various levels of Faster-RCNN. It assumes that image-level weak annotation is available for the target domain. So we use the labels from our pseudo-labeling process as image-level class labels." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.748, + 0.469, + 0.809 + ], + "angle": 0, + "content": "TIA [78] is a method that instantiates feature alignment by using both domain classifier and auxiliary classification and localization heads. The model is trained in a mini-max manner using gradient reverse layers [15]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.469, + 0.9 + ], + "angle": 0, + "content": "LODS [33] uses style enhancement as data augmentation. The pseudo bounding boxes are generated from the original images, and the features from both versions of the same image are aligned. It also utilizes exponential moving average for the weights. It does not use any supervised training data from the source domain." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.09, + 0.89, + 0.166 + ], + "angle": 0, + "content": "
Model# paramsMSCOCOScenes100
APmcoAP50coAPmcoAP50coAPmcoAP50w
M1 (R-50)41.4M50.0576.4840.5262.1241.2864.65
M2 (R-101)60.5M51.2977.4641.1163.1041.9665.74
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.499, + 0.168, + 0.892, + 0.238 + ], + "angle": 0, + "content": "Table 2. Performance of base models on MSCOCO validation set with remapped categories and Scenes100 before any adaptation. For Scenes100, \\( APs \\) are evaluated on each video individually and then averaged. R- stands for ResNet. # indicates the number of parameters \\( (M = 10^6) \\). See Sec. 5.1 for \\( AP \\) notation details." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.246, + 0.593, + 0.261 + ], + "angle": 0, + "content": "5.4. Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.273, + 0.892, + 0.424 + ], + "angle": 0, + "content": "The performance of the base models is shown in Tab. 2, along with their number of trainable parameters. M2 is a bigger model with more parameters and representation capacity, so it outperforms M1 by a noticeable margin in all the metrics before adaptation. All the \\(AP\\) gains shown in other tables are based on the performance of base model M2. Note that the \\(AP\\) numbers on Scenes100 is already relatively high at only about 10 points below the performance on MSCOCO, so we do not observe very high (\\(>5\\) points) \\(AP\\) gains in the results shown later." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.429, + 0.892, + 0.64 + ], + "angle": 0, + "content": "We compare the performance of different adaptation methods in terms of average \\(AP\\) gain in Tab. 3. Here we show our best combination of methods, which incorporates pseudo-labeling, location-aware mixup, and object mask mid-fusion. Among the baseline methods for domain adaptive object detection, only ST and LODS improves the performance significantly. AT only yields marginal improvement in classes-weighted \\(AP\\). STAC, \\(\\mathrm{H}^2\\mathrm{FA}\\), and TIA actually degrade the performance on Scenes100. LODS performs surprisingly well considering it does not use source domain images. However, we found that any training schedule longer than 1000 iterations would lead to severe performance degradation. The proposed method yields much higher, consistent, and stable improvement." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In general, self-training-based methods (ST, AT, STAC, and the proposed method) perform better than domain alignment-based methods \\(\\mathrm{(H^2FA}\\) and TIA). Domain alignment is usually achieved through adversarial learning, which is known to have unstable objectives and is more difficult to optimize [14, 42, 59, 61]. Therefore, these methods are more sensitive to hyperparameters and training schedules, making it challenging to determine the optimal settings for all datasets. Applying domain alignment methods directly to our scene adaptive dataset without careful tuning can cause problems as our dataset has less in-domain variance than domain adaptation datasets. Furthermore, since the scenes in our dataset are very different, they might require different settings. This implies that Scenes100 has its uniqueness compared to more generally purported domain adaptive object detection datasets. Our proposed method is more robust to this setting and does not rely heavily on" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21595" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.116, + 0.089, + 0.432, + 0.233 + ], + "angle": 0, + "content": "
Method\\(APG_{co}^{m}\\)\\(APG_{co}^{50}\\)\\(APG_{w}^{m}\\)\\(APG_{w}^{50}\\)
ST [57]+0.80+0.24+1.39+1.03
STAC [60]-1.26-5.12-1.97-6.64
AT [37]-0.75-1.11+0.06+0.04
\\(\\mathrm{H^{2}FA}\\) [72]-3.10-4.97-3.77-6.01
TIA [78]-0.32-0.37-0.32-0.33
LODS [33]+0.45+1.28+1.02+2.28
Proposed+3.76+4.45+3.78+4.65
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.236, + 0.47, + 0.264 + ], + "angle": 0, + "content": "Table 3. Averaged \\(AP\\) gain of different adaptation methods. See Sec. 5.1 for \\(AP\\) notation details." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.28, + 0.466, + 0.461 + ], + "angle": 0, + "content": "
MixupFusion\\(APG_{co}^{m}\\)\\(APG_{co}^{50}\\)\\(APG_{w}^{m}\\)\\(APG_{w}^{50}\\)
XX+0.95+0.54+1.67+1.55
Location-awareX+1.72+1.67+2.25+2.53
RandomX+1.22+1.13+1.76+2.03
Xearly+1.85+2.12+2.22+2.73
Xmid+3.40+3.81+3.67+3.98
Xlate+3.34+3.60+3.38+3.72
Location-awareearly+2.25+2.82+2.59+3.46
Location-awaremid+3.76+4.45+3.78+4.65
Location-awarelate+3.66+4.10+3.73+4.31
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.465, + 0.469, + 0.521 + ], + "angle": 0, + "content": "Table 4. Ablation study for the proposed components of our scene-adaptive object detection method. \\(\\times\\) means not being applied. Pseudo-labeling is always applied. This table shows the averaged \\(AP\\) gain; see Sec. 5.1 for \\(AP\\) notation details." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.537, + 0.466, + 0.662 + ], + "angle": 0, + "content": "
MixupFusion\\(APG_{co}^{m}\\)\\(APG_{co}^{50}\\)\\(APG_{w}^{m}\\)\\(APG_{w}^{50}\\)
✘AVG+3.40+3.81+3.67+3.98
✘CNN+3.25+3.52+3.26+3.79
✘ATTN+3.22+3.73+3.23+3.95
Location-awareAVG+3.76+4.45+3.78+4.65
Location-awareCNN+3.60+4.20+3.71+4.50
Location-awareATTN+3.30+3.89+3.45+4.26
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.665, + 0.469, + 0.737 + ], + "angle": 0, + "content": "Table 5. Ablation study on different feature pyramid fusion methods in term of averaged \\(AP\\) gain. AVG, CNN, and ATTN indicate mid-fusion based on average pooling, CNN, or attention module, respectively. \\(\\pmb{x}\\) means not being applied. Pseudo-labeling is always applied. See Sec. 5.1 for \\(AP\\) notation details." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.744, + 0.469, + 0.804 + ], + "angle": 0, + "content": "hyperparameter tuning, as we use the same set of hyperparameters across all 100 videos. Additional analysis and success/failure cases can be found in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.816, + 0.231, + 0.833 + ], + "angle": 0, + "content": "5.5. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.903 + ], + "angle": 0, + "content": "We explore the effect of the different components in our methods by experimenting various combinations of them. For all experiments, we keep the pseudo-labeling unchanged, as it is the basis of other components. We compare" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "different types of object mask fusion (Fusion): early, mid, and late. We also compare the proposed Location-aware object mixup with Random mixup. Tab. 4 shows the results. The experiments on hyper-parameters can be found in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.172, + 0.893, + 0.37 + ], + "angle": 0, + "content": "The proposed pseudo-labeling, mixup, and object mask fusion all benefit the adaptation performance. Applying only the pseudo-labeling already leads to \\(AP\\) gain higher than ST, which is more complicated involving pseudo box weighting. Location-aware mixup is also more advantageous than Random mixup, validating the assumption in Sec. 3.3 that having fewer artifacts is better for adaptation. Among three fusion options, mid-fusion yields the best result. That is perhaps because in a mid-fusion model, the RPN utilizes the fused feature pyramid, which can contain critical information of object boundaries. The combination of mixup and object mask fusion can further improve the performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.375, + 0.892, + 0.527 + ], + "angle": 0, + "content": "We also compare different feature pyramid fusion methods described in Eq. (3). In our main experiments, we simply average the feature pyramids from the original image and the object mask. Now at each level of the pyramid, we use a separate CNN or an attention module to fuse the feature maps. The comparison is shown in Tab. 5. The parametric method performs slightly worse than average-pooling, probably because the newly introduced modules have randomly initialized weights and thus require additional supervised training data." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.557, + 0.608, + 0.574 + ], + "angle": 0, + "content": "6. Summary" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.588, + 0.892, + 0.86 + ], + "angle": 0, + "content": "We have presented a self-supervised framework for scene-adaptive object detection. Base detectors and generic trackers are used to generate pseudo object labels as the adaptive training targets in a cross-teaching manner. To fully utilize the background equivariance when camera perspective is fixed, we propose artifacts-free location-aware object mixup to augment the input images, and dynamic background extraction for additional input modality to the detector. We also introduce the first large-scale scene adaptive object detection dataset, Scenes100, with several unique features compared to other domain adaptive object detection datasets. Our method outperforms domain adaptive object detection baselines by a large margin on Scenes100. We also conduct extensive experiments to illustrate the effectiveness of the components in our framework. We hope this work has demonstrated the importance of scene adaptation, and it will spur further research interest in this impactful but understudied area." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Acknowledgements. This research was partially supported by NSF grants IIS-1763981 & NSDF DUE-2055406." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21596" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.117, + 0.47, + 0.185 + ], + "angle": 0, + "content": "[1] David Berthelot, Nicholas Carlini, Ian Goodfellow, Avital Oliver, Nicolas Papernot, and Colin Raffel. Mixmatch: A holistic approach to semi-supervised learning. In Conference on Neural Information Processing Systems (NeurIPS), 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.189, + 0.472, + 0.245 + ], + "angle": 0, + "content": "[2] Goutam Bhat, Martin Danelljan, Luc Van Gool, and Radu Timofte. Learning discriminative model prediction for tracking. In International Conference on Computer Vision (ICCV), 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.248, + 0.471, + 0.289 + ], + "angle": 0, + "content": "[3] Alexey Bochkovskiy, Chien-Yao Wang, and Hong-Yuan Mark Liao. Yolov4: Optimal speed and accuracy of object detection. In ArXiv, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.293, + 0.471, + 0.321 + ], + "angle": 0, + "content": "[4] Shahine Bouabid and Vincent Delaitre. Mixup regularization for region proposal based object detectors. In ArXiv, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.323, + 0.471, + 0.365 + ], + "angle": 0, + "content": "[5] Zhaowei Cai and Nuno Vasconcelos. Cascade r-cnn: Delving into high quality object detection. In Computer Vision and Pattern Recognition Conference (CVPR), 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.368, + 0.471, + 0.424 + ], + "angle": 0, + "content": "[6] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European Conference on Computer Vision (ECCV), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.426, + 0.471, + 0.482 + ], + "angle": 0, + "content": "[7] Yuhua Chen, Wen Li, Christos Sakaridis, Dengxin Dai, and Luc Van Gool. Domain adaptive faster r-cnn for object detection in the wild. In Computer Vision and Pattern Recognition Conference (CVPR), 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.485, + 0.471, + 0.554 + ], + "angle": 0, + "content": "[8] Xiangxiang Chu, Zhi Tian, Yuqing Wang, Bo Zhang, Haibing Ren, Xiaolin Wei, Huaxia Xia, and Chunhua Shen. Twins: Revisiting the design of spatial attention in vision transformers. In Conference on Neural Information Processing Systems (NeurIPS), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.557, + 0.471, + 0.639 + ], + "angle": 0, + "content": "[9] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Computer Vision and Pattern Recognition Conference (CVPR), 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.643, + 0.47, + 0.685 + ], + "angle": 0, + "content": "[10] Martin Danelljan, Goutam Bhat, Christoph Mayer, and Matthieu Paul. Pytracking. https://github.com/visionml/pytracking, 2019.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.687, + 0.47, + 0.743 + ], + "angle": 0, + "content": "[11] Jinhong Deng, Wen Li, Yuhua Chen, and Lixin Duan. Unbiased mean teacher for cross-domain object detection. In Computer Vision and Pattern Recognition Conference (CVPR), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.746, + 0.47, + 0.843 + ], + "angle": 0, + "content": "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations (ICLR), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[13] Yoav Freund and Robert E Schapire. A decision-theoretic generalization of on-line learning and an application to boosting. Journal of Computer and System Sciences, 55(1):119-139, 1997. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.117, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "[14] Angus Galloway, Anna Golubeva, Thomas Tanay, Medhat Moussa, and Graham W. Taylor. Batch normalization is a cause of adversarial vulnerability. In *ArXiv*, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[15] Yaroslav Ganin and Victor Lempitsky. Unsupervised domain adaptation by backpropagation. In International Conference on Machine Learning (ICML), 2015. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.178, + 0.892, + 0.244 + ], + "angle": 0, + "content": "[16] Yaroslav Ganin, E. Ustinova, Hana Ajakan, Pascal Germain, H. Larochelle, Francois Lavolette, Mario Marchand, and Victor S. Lempitsky. Domain-adversarial training of neural networks. In Journal of Machine Learning Research, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.247, + 0.892, + 0.3 + ], + "angle": 0, + "content": "[17] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In Computer Vision and Pattern Recognition Conference (CVPR), 2012. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.302, + 0.892, + 0.329 + ], + "angle": 0, + "content": "[18] Ross Girshick. Fast r-cnn. In International Conference on Computer Vision (ICCV), 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.33, + 0.892, + 0.385 + ], + "angle": 0, + "content": "[19] Ross Girshick, Jeff Donahue, Trevor Darrell, and Jitendra Malik. Rich feature hierarchies for accurate object detection and semantic segmentation. In Computer Vision and Pattern Recognition Conference (CVPR), 2014. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.386, + 0.892, + 0.427 + ], + "angle": 0, + "content": "[20] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In International Conference on Computer Vision (ICCV), 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.428, + 0.892, + 0.482 + ], + "angle": 0, + "content": "[21] Kaiming He, X. Zhang, Shaoqing Ren, and Jian Sun. Spatial pyramid pooling in deep convolutional networks for visual recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 37, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.484, + 0.892, + 0.525 + ], + "angle": 0, + "content": "[22] Zhenwei He and Lei Zhang. Multi-adversarial faster-rcnn for unrestricted object detection. In International Conference on Computer Vision (ICCV), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.526, + 0.892, + 0.593 + ], + "angle": 0, + "content": "[23] Dan Hendrycks, Norman Mu, Ekin D. Cubuk, Barret Zoph, Justin Gilmer, and Balaji Lakshminarayanan. AugMix: A simple data processing method to improve robustness and uncertainty. In International Conference on Learning Representations (ICLR), 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.595, + 0.892, + 0.636 + ], + "angle": 0, + "content": "[24] Tin Kam Ho. Random decision forests. In International Conference on Document Analysis and Recognition (ICDAR), volume 1, 1995. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.637, + 0.892, + 0.678 + ], + "angle": 0, + "content": "[25] Minh Hoai and Andrew Zisserman. Talking heads: Detecting humans and recognizing their interactions. In Computer Vision and Pattern Recognition Conference (CVPR), 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.679, + 0.892, + 0.747 + ], + "angle": 0, + "content": "[26] Jiaxing Huang, Dayan Guan, Aoran Xiao, and Shijian Lu. Model adaptation: Historical contrastive learning for unsupervised domain adaptation without source data. In Conference on Neural Information Processing Systems (NeurIPS), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.748, + 0.892, + 0.803 + ], + "angle": 0, + "content": "[27] Naoto Inoue, Ryosuke Furuta, Toshihiko Yamasaki, and Kiyoharu Aizawa. Cross-domain weakly-supervised object detection through progressive domain adaptation. In Computer Vision and Pattern Recognition Conference (CVPR), 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.804, + 0.892, + 0.831 + ], + "angle": 0, + "content": "[28] Glenn Jocher. Yolov5. https://github.com/ultralytics/yolov5, 2020.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[29] Seunghyeon Kim, Jaehoon Choi, Taekyung Kim, and Chang-ick Kim. Self-training and adversarial background regularization for unsupervised domain adaptive one-stage object detection. In International Conference on Computer Vision (ICCV), 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21597" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[30] Samuli Laine and Timo Aila. Temporal ensembling for semi-supervised learning. In International Conference on Learning Representations (ICLR), 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.134, + 0.471, + 0.218 + ], + "angle": 0, + "content": "[31] Chuyi Li, Lulu Li, Hongliang Jiang, Kaiheng Weng, Yifei Geng, Liang Li, Zaidan Ke, Qingyuan Li, Meng Cheng, Weiqiang Nie, Yiduo Li, Bo Zhang, Yufei Liang, Linyuan Zhou, Xiaoming Xu, Xiangxiang Chu, Xiaoming Wei, and Xiaolin Wei. Yolov6: A single-stage object detection framework for industrial applications. In ArXiv, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.218, + 0.471, + 0.273 + ], + "angle": 0, + "content": "[32] Junnan Li, Richard Socher, and Steven C.H. Hoi. Dividemix: Learning with noisy labels as semi-supervised learning. In International Conference on Learning Representations (ICLR), 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.273, + 0.471, + 0.329 + ], + "angle": 0, + "content": "[33] Shuaifeng Li, Mao Ye, Xiatian Zhu, Lihua Zhou, and Lin Xiong. Source-free object detection by learning to overlook domain style. In Computer Vision and Pattern Recognition Conference (CVPR), 2022. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.33, + 0.471, + 0.385 + ], + "angle": 0, + "content": "[34] Wuyang Li, Xinyu Liu, and Yixuan Yuan. Sigma: Semantic-complete graph matching for domain adaptive object. In Computer Vision and Pattern Recognition Conference (CVPR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.385, + 0.471, + 0.453 + ], + "angle": 0, + "content": "[35] Xianfeng Li, Weijie Chen, Di Xie, Shicai Yang, Peng Yuan, Shiliang Pu, and Yueting Zhuang. A free lunch for unsupervised domain adaptive object detection without source data. In AAAI Conference on Artificial Intelligence (AAAI), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.454, + 0.471, + 0.525 + ], + "angle": 0, + "content": "[36] Yanghao Li, Chao-Yuan Wu, Haoqi Fan, Karttikeya Mangalam, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. Mvity2: Improved multiscale vision transformers for classification and detection. In Computer Vision and Pattern Recognition Conference (CVPR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.525, + 0.471, + 0.594 + ], + "angle": 0, + "content": "[37] Yu-Jhe Li, Xiaoliang Dai, Chih-Yao Ma, Yen-Cheng Liu, Kan Chen, Bichen Wu, Zijian He, Kris Kitani, and Peter Vajda. Cross-domain adaptive teacher for object detection. In Computer Vision and Pattern Recognition Conference (CVPR), 2022. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.594, + 0.471, + 0.636 + ], + "angle": 0, + "content": "[38] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE Transactions on Pattern Analysis and Machine Intelligence, 40(12):2935-2947, dec 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.636, + 0.471, + 0.692 + ], + "angle": 0, + "content": "[39] Tsung-Yi Lin, Piotr Dollar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Computer Vision and Pattern Recognition Conference (CVPR), 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.692, + 0.471, + 0.747 + ], + "angle": 0, + "content": "[40] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In International Conference on Computer Vision (ICCV), 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.747, + 0.471, + 0.803 + ], + "angle": 0, + "content": "[41] Tsung-Yi Lin, Michael Maire, Serge Belongie, Lubomir Bourdev, Ross Girshick, James Hays, Pietro Perona, Deva Ramanan, C. Lawrence Zitnick, and Piotr Dólar. Microsoft coco: Common objects in context. In ArXiv, 2014. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.471, + 0.872 + ], + "angle": 0, + "content": "[42] Chen Liu, Mathieu Salzmann, Tao Lin, Ryota Tomioka, and Sabine Susstrunk. On the loss landscape of adversarial training: Identifying challenges and how to overcome them. In Conference on Neural Information Processing Systems (NeurIPS), 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.872, + 0.471, + 0.903 + ], + "angle": 0, + "content": "[43] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.092, + 0.892, + 0.121 + ], + "angle": 0, + "content": "Berg. Ssd: Single shot multibox detector. In European Conference on Computer Vision (ECCV), 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.192 + ], + "angle": 0, + "content": "[44] Yen-Cheng Liu, Chih-Yao Ma, Zijian He, Chia-Wen Kuo, Kan Chen, Peizhao Zhang, Bichen Wu, Zsolt Kira, and Peter Vajda. Unbiased teacher for semi-supervised object detection. In International Conference on Learning Representations (ICLR), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.263 + ], + "angle": 0, + "content": "[45] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In International Conference on Computer Vision (ICCV), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.263, + 0.892, + 0.335 + ], + "angle": 0, + "content": "[46] Muhammad Akhtar Munir, Muhammad Haris Khan, M. Saquib Sarfraz, and Mohsen Ali. Ssal: Synergizing between self-training and adversarial learning for domain adaptive object detection. In Conference on Neural Information Processing Systems (NeurIPS), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.335, + 0.892, + 0.393 + ], + "angle": 0, + "content": "[47] Supreeth Narasimhaswamy, Trung Nguyen, and Minh Hoai. Detecting hands and recognizing physical contact in the wild. In Conference on Neural Information Processing Systems (NeurIPS), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.393, + 0.892, + 0.452 + ], + "angle": 0, + "content": "[48] Supreeth Narasimhaswamy, Thanh Nguyen, Mingzhen Huang, and Minh Hoai. Whose hands are these? hand detection and hand-body association in the wild. In Computer Vision and Pattern Recognition Conference (CVPR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.452, + 0.892, + 0.509 + ], + "angle": 0, + "content": "[49] Supreeth Narasimhaswamy, Zhengwei Wei, Yang Wang, Justin Zhang, and Minh Hoai. Contextual attention for hand detection in the wild. In International Conference on Computer Vision (ICCV), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.509, + 0.892, + 0.553 + ], + "angle": 0, + "content": "[50] Thanh Nguyen, Chau Pham, Khoi Nguyen, and Minh Hoai. Few-shot object counting and detection. In European Conference on Computer Vision (ECCV), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.554, + 0.892, + 0.609 + ], + "angle": 0, + "content": "[51] Kento Nishi, Yi Ding, Alex Rich, and Tobias Höllerer. Augmentation strategies for learning with noisy labels. In Computer Vision and Pattern Recognition Conference (CVPR), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.61, + 0.892, + 0.667 + ], + "angle": 0, + "content": "[52] German I. Parisi, Ronald Kemker, Jose L. Part, Christopher Kanan, and Stefan Wermter. Continual lifelong learning with neural networks: A review. Neural Networks, 113:54-71, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.668, + 0.892, + 0.711 + ], + "angle": 0, + "content": "[53] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In ArXiv, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.712, + 0.892, + 0.742 + ], + "angle": 0, + "content": "[54] Joseph Redmon and Ali Farhadi. YOLO9000: better, faster, stronger. In ArXiv, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.742, + 0.892, + 0.772 + ], + "angle": 0, + "content": "[55] Joseph Redmon and Ali Farhadi. Yolov3: An incremental improvement. In ArXiv, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.772, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[56] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In Conference on Neural Information Processing Systems (NeurIPS), 2015. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.831, + 0.892, + 0.903 + ], + "angle": 0, + "content": "[57] Aruni RoyChowdhury, Prithvijit Chakrabarty, Ashish Singh, SouYoung Jin, Huaizu Jiang, Liangliang Cao, and Erik Learned-Miller. Automatic adaptation of object detectors to new domains using self-training. In Computer Vision and Pattern Recognition Conference (CVPR), 2019. 1, 2, 3, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "21598" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[58] Kuniaki Saito, Yoshitaka Ushiku, Tatsuya Harada, and Kate Saenko. Strong-weak distribution alignment for adaptive object detection. In Computer Vision and Pattern Recognition Conference (CVPR), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.47, + 0.189 + ], + "angle": 0, + "content": "[59] Chawin Sitawarin, Supriyo Chakraborty, and David A. Wagner. Improving adversarial robustness through progressive hardening. In ArXiv, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.191, + 0.469, + 0.245 + ], + "angle": 0, + "content": "[60] Kihyuk Sohn, Zizhao Zhang, Chun-Liang Li, Han Zhang, Chen-Yu Lee, and Tomas Pfister. A simple semi-supervised learning framework for object detection. In arXiv, 2020. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.247, + 0.469, + 0.3 + ], + "angle": 0, + "content": "[61] Arvind P. Sridhar, Chawin Sitawarin, and David Wagner. Mitigating adversarial training instability with batch normalization. In International Conference on Learning Representations Workshop (ICLRW), 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.302, + 0.469, + 0.357 + ], + "angle": 0, + "content": "[62] Antti Tarvainen and Harri Valpola. Weight-averaged consistency targets improve semi-supervised deep learning results. In Conference on Neural Information Processing Systems (NeurIPS), 03 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.359, + 0.469, + 0.398 + ], + "angle": 0, + "content": "[63] Alexandru Cristian Telea. An image inpainting technique based on the fast marching method. Journal of Graphics Tools, 9:23-34, 2004. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.401, + 0.469, + 0.44 + ], + "angle": 0, + "content": "[64] Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. FCOS: Fully convolutional one-stage object detection. In International Conference on Computer Vision (ICCV), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.442, + 0.469, + 0.496 + ], + "angle": 0, + "content": "[65] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Conference on Neural Information Processing Systems (NeurIPS), 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.498, + 0.469, + 0.566 + ], + "angle": 0, + "content": "[66] Vikas Verma, Alex Lamb, Christopher Beckham, Amir Najafi, Ioannis Mitliagkas, David Lopez-Paz, and Yoshua Bengio. *Manifold mixup: Better representations by interpolating hidden states.* In International Conference on Machine Learning (ICML), 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.568, + 0.469, + 0.62 + ], + "angle": 0, + "content": "[67] Chien-Yao Wang, Alexey Bochkovskiy, and Hong-Yuan Mark Liao. Yolov7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors. In ArXiv, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.623, + 0.469, + 0.663 + ], + "angle": 0, + "content": "[68] Chien-Yao Wang, I-Hau Yeh, and Hongpeng Liao. You only learn one representation: Unified network for multiple tasks. In ArXiv, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.665, + 0.469, + 0.733 + ], + "angle": 0, + "content": "[69] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In International Conference on Computer Vision (ICCV), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.735, + 0.469, + 0.774 + ], + "angle": 0, + "content": "[70] Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. Detector2. https://github.com/facebookresearch/detectron2, 2019.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.777, + 0.469, + 0.83 + ], + "angle": 0, + "content": "[71] Lin Xiong, Mao Ye, Dan Zhang, Yan Gan, Xue Li, and Yingying Zhu. Source data-free domain adaptation of object detector through domain-specific perturbation. International Journal of Intelligent Systems, 36:3746 - 3766, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.832, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[72] Yunqiu Xu, Yifan Sun, Zongxin Yang, Jiaxu Miao, and Yi Yang. H\\(^2\\)FA R-CNN: Holistic and hierarchical feature alignment for cross-domain weakly supervised object detection. In Computer Vision and Pattern Recognition Conference (CVPR), 2022. 2, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[73] Jihan Yang, Shaoshuai Shi, Zhe Wang, Hongsheng Li, and Xiaojuan Qi. St3d: Self-training for unsupervised domain adaptation on 3d object detection. In Computer Vision and Pattern Recognition Conference (CVPR), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.218 + ], + "angle": 0, + "content": "[74] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darryll. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In Computer Vision and Pattern Recognition Conference (CVPR), 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.892, + 0.288 + ], + "angle": 0, + "content": "[75] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In International Conference on Computer Vision (ICCV), 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.291, + 0.892, + 0.33 + ], + "angle": 0, + "content": "[76] Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In ArXiv, 2017. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.334, + 0.892, + 0.401 + ], + "angle": 0, + "content": "[77] Pengchuan Zhang, Xiyang Dai, Jianwei Yang, Bin Xiao, Lu Yuan, Lei Zhang, and Jianfeng Gao. Multi-scale vision longformer: A new vision transformer for high-resolution image encoding. In International Conference on Computer Vision (ICCV), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.404, + 0.892, + 0.458 + ], + "angle": 0, + "content": "[78] Liang Zhao and Limin Wang. Task-specific inconsistency alignment for domain adaptive object detection. In Computer Vision and Pattern Recognition Conference (CVPR), 2022. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.461, + 0.892, + 0.501 + ], + "angle": 0, + "content": "[79] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable DETR: deformable transformers for end-to-end object detection. In ArXiv, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.501 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "21599" + } + ] +] \ No newline at end of file diff --git a/2023/Object Detection With Self-Supervised Scene Adaptation/b774b6cf-eacd-4660-845f-f32b5ec4dc83_origin.pdf b/2023/Object Detection With Self-Supervised Scene Adaptation/b774b6cf-eacd-4660-845f-f32b5ec4dc83_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..db3cd224d956c2299f95b7587751a2cabfcf290e --- /dev/null +++ b/2023/Object Detection With Self-Supervised Scene Adaptation/b774b6cf-eacd-4660-845f-f32b5ec4dc83_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:42ac724c09a6501dd1f17441e79cb6c84013d35a422412797d992da3ab45372c +size 7506294 diff --git a/2023/Object Detection With Self-Supervised Scene Adaptation/full.md b/2023/Object Detection With Self-Supervised Scene Adaptation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a302468f1f0e67c34700464bc5b7b7a7ef8fe0f4 --- /dev/null +++ b/2023/Object Detection With Self-Supervised Scene Adaptation/full.md @@ -0,0 +1,391 @@ +# Object Detection with Self-Supervised Scene Adaptation + +Zekun Zhang1 Minh Hoai1,2 + +1Stony Brook University, Stony Brook, NY 11794, USA + +$^{2}$ VinAI Research, Hanoi, Vietnam + +{zekzhang,minhhooai}@cs.stonyrbook.edu + +# Abstract + +This paper proposes a novel method to improve the performance of a trained object detector on scenes with fixed camera perspectives based on self-supervised adaptation. Given a specific scene, the trained detector is adapted using pseudo-ground truth labels generated by the detector itself and an object tracker in a cross-teaching manner. When the camera perspective is fixed, our method can utilize the background equivariance by proposing artifact-free object mixup as a means of data augmentation, and utilize accurate background extraction as an additional input modality. We also introduce a large-scale and diverse dataset for the development and evaluation of scene-adaptive object detection. Experiments on this dataset show that our method can improve the average precision of the original detector, outperforming the previous state-of-the-art self-supervised domain adaptive object detection methods by a large margin. Our dataset and code are published at https://github.com/cvlab-stonybrook/scenes100. + +# 1. Introduction + +The need to detect objects in video streams from stationary cameras arises in many computer vision applications, including video surveillance and autonomous retail. In general, different applications require the detection of different object categories, and each computer-vision-based product will have its own detector. However, for a specific product, there is typically a single detector that will be used for many cameras/scenes. For example, a typical video surveillance product would use the same detector to detect pedestrians and vehicles for network cameras installed at different locations. Unfortunately, a single detector might not work well for all scenes, leading to trivial and unforgiving mistakes. + +This fundamental problem of many computer vision products stems from the limited generalization power of a single model, due to limited training data, limited model capacity, or both. One can attempt to address this problem by + +using more training data, but it will incur additional cost for data collection and annotation. Furthermore, in many cases, due to the low latency requirement or the limited computing resources for inference, a product is forced to use a very lightweight network, and this network will have limited representation capacity to generalize across many scenes. + +In this paper, instead of having a single scene-generic detector, we propose using scene-specific detectors. This yields higher detection performance as each detector is customized for a specific scene, and allows us to use a lightweight model without sacrificing accuracy as each detector is only responsible for one scene. + +Obtaining scene-specific detectors, however, is very challenging. A trivial approach is to train a detector for each scene separately, but this requires an enormous amount of annotated training data. Instead, we propose a self-supervised method to adapt a pre-trained detector to each scene. Our method records the unlabeled video frames in the past, uses the trained detector to detect objects in those frames, and generates augmented training data based on those detections. Although the detections made by the pre-trained model can be noisy, they can still be useful for generating pseudo annotated data. We further extend those pseudo bounding boxes by applying object tracking [2, 57] along the video timeline, aiming to propagate the detections to adjacent frames to recover some of the false negatives not returned by the detector. We also use multiple detectors to obtain the pseudo labels and train the detector in a cross-teaching manner, taking the advantage of the ensemble of models [13, 24]. + +Exploiting the stationary nature of the camera, we propose two additional techniques to boost the detection performance: location-aware mixup and background-augmented input. The former is to generate more samples during training through object mixup [76] that contains less artifacts, based on the aforementioned pseudo boxes generated from detection and tracking. The latter involves estimating the background image and fusing it with the detector's input. + +In short, the main contribution of our paper is a novel framework that utilizes self-supervision, location-aware ob + +ject mixup, and background modeling to improve the detection performance of a pre-trained object detector on scenes with stationary cameras. We also contribute a large scale and diverse dataset for the development of scene adaptive object detection, which contains sufficient quantity and quality annotations for evaluation. + +# 2. Related Work + +Despite much recent improvement in object detection research [3,5,6,8,12,18,19,21,25,28,31,36,40,43,45,47-50, 53-56,64,65,67-69,77,79], trained detectors still encounter problems with domain shift or domain gap. Scene adaptive object detection can be viewed as a special case of domain adaption, in which an object detector trained on a fully-supervised source domain is adapted to a target domain. Most research in this direction focuses on semi-supervised, weakly-supervised, or self-supervised adaptation. In addition, source-free adaptation [26, 33, 35, 71] has been proposed to address the situation where the source domain data is unavailable during adaptation. + +Self-labeling methods [11,27,29,33,37,44,46,57,60,73] uses the teacher-student setup from semi-supervised image classification [30, 62]. A teacher detector trained on the source domain generates pseudo bounding boxes on the target domain images, and a student model is trained with those boxes to improve its performance on the target domain. Techniques such as weak/strong augmentation [37,44,60], knowledge distillation [11], and weight averaging [33, 37] have been used to deal with noisy pseudo labels from the teacher detector due to the domain gap. Our proposed pseudo-labeling method uses two teacher models trained on the source domain to train a student model, and their pseudo boxes are aggregated and refined. We further apply tracking as in [57], but with both forward and backward directions, to extend the pseudo labels. The pseudo boxes also form the basis for later location-aware mixup and dynamic background extraction steps. + +Domain alignment methods [7, 11, 22, 33, 34, 58, 72, 78] aim to reduce the domain gap by enforcing the models to output similarly on the source and target domain at image, proposal, or instance levels. They use domain adversarial learning through gradient reversal [15, 16] or graph matching [34] for domain alignment. This approach involves adding domain alignment losses to existing object detection models and is complementary to our proposed method. + +Another seemingly related research area is continual or incremental learning. Scene adaptive and domain adaptive object detection are in between task-continual learning [38] and data-continual learning [52], for that the object categories of the source and target domains are the same, but the data distribution differs. It is possible to formulate scene adaptive object detection as an online learning problem, but this approach would require additional annotations. + +![](images/57ebeeecc3d09929f50445e8a6752363c0424a4c7ac22dbb1bfdd43d990bb94f.jpg) +(a) Input image $\mathcal{I}$ . + +![](images/360d51fccf503508fa75a943d66deaef16f2f5d493c82dc5ac89ee2b717c23eb.jpg) +(b) Learned background image $\mathcal{B}$ . + +![](images/0b7af89dc2853a1079de2d8f2c6ac3e668569254650dddd42448c5a64941b95b.jpg) +(c) Object mask $\mathcal{M}^O$ +Figure 1. Inference flow of the proposed detector that can adapt to a specific scene. The background image and fusion faster-RCNN model are learned during training, discussed in Sec. 3. + +# 3. Methods + +For an object detector pre-trained on a source dataset referred as the base model, we aim to improve its performance on video stream from a stationary camera in a given scene, with unlabeled videos captured by the same camera in the past as adaptation training samples. The categories of object of interest in the scene are also present in the source dataset, but the data distribution is shifted. The base model is adapted to each scene independently. + +Assuming the distribution shift is moderate, the base detector can generate partially decent bounding boxes on the unlabeled frames. We use them as pseudo-labels to improve the base model itself. Many objects are ignored by the base detector, as their appearance can be very different from the source dataset. However, videos contain temporal correlation information, making it possible to apply object tracking in both time directions to recover some of those false negatives. The pseudo-labels are then refined to remove low-confidence boxes and duplicates. We refer to this process of obtaining pseudo bounding boxes as pseudo-labeling. + +As data augmentation is beneficial for training on less reliable labels [51], we take advantages of object mixup to generate new sample images. Since the background is stationary, those images contain fewer artifacts, which is beneficial for detectors. We further adopt the idea of ensemble models [13, 24] by using two base detectors to obtain the pseudo labels, for they can be complementary to each other. The pseudo boxes are used to train one base model, which we refer as cross-teaching. + +Since the job of a detector is to separate foreground objects from the background, having information about the background is advantageous. In a stationary camera video with moving objects, different parts of the background are covered by objects in different frames, making it possible to model the complete background. We use the aforementioned pseudo bounding boxes to mark the uncovered parts of the background in each frame. Combining the partial + +![](images/77df66211452e80f23b46a4eed31c96da2966fc781a8a6759bf5232e0449e947.jpg) +Figure 2. Training flow of the proposed self-supervised scene adaptation object detection framework with sample images. The mixup training images with pseudo bounding boxes and background image are used in the fusion training. The details are discussed in Sec. 3. + +background from a sequence of frames can give an accurate and complete background model. We then modify the Faster-RCNN [56] architecture to fuse the background with the input image to improve detection performance. + +The overall structure and data flows of the proposed methods are shown in Fig. 1 and Fig. 2. Our proposed methods are highly flexible and can be configured to trade-off between performance and speed. Object tracking can be bypassed to reduce the time to get the pseudo-bounding boxes. Mixup and object mask fusion can also be disabled for faster training and inference. In Sec. 5.5 we will show ablation study results on how those components affect the performance of the adapted models. We now describe each component in more details. For consistency, the images are all from the first video in our collected dataset, which will be described in Sec. 4. + +# 3.1. Base Detection Models on Source Domain + +We use Faster-RCNN [56] in our experiments for its high precision and flexible modular architecture. Among the techniques proposed for self-supervised scene adaptation, both pseudo-labeling and mixup can be directly applied to any object detection model. The object mask fusion technique requires modification to the network architecture, but it can be easily extended to most types of detectors by adding a parallel branch to the existing network. + +The object labels in MSCOCO training set are remapped to person and vehicle as described in Sec. 4.2. All objects other than person, car, bus, and truck are discarded. We take Faster-RCNNs pre-trained on the original MSCOCO training set, and finetune them on the remapped training set. We choose two models with ResNet-50 and ResNet101 backbones as base models, referred as M1 and M2. + +# 3.2. Pseudo Labels from Detection and Tracking + +We take the frames from the first training portion of each video, and adopt the idea of detection and tracking similar to [57], then apply our proposed bounding box refinement and cross-teaching schemes. Base detectors described in + +Sec. 3.1 are applied on the frames, and any bounding box with confidence score higher than $\lambda_{det}$ is kept as a valid pseudo box. + +Single-object trackers are initialized with detected objects having confidence scores higher than $\lambda_{\text{tot}}$ . We use DiMP-50 [2] to track the objects in both forward and backward directions in the video timeline, and the tracked object bounding boxes are used as pseudo boxes. The reason for tracking in both directions is that if an object is moving towards the camera and appears bigger by time, backward tracking is more accurate because the initial bounding box contains more detailed texture. To prevent drifting, the maximum length of each track is capped at two seconds. + +Bounding box refinement is applied to eliminate duplicated boxes of the same object from combining pseudo bounding boxes from detector and tracker. Each candidate box in the same frame is regarded as a graph node. If two boxes have the same object label, and their IoU is above a threshold $\lambda_{iou}$ , then an edge is added to connect them. We assume that each connected component of the graph represents one single object, and the nodes in the component are duplicated. We only keep the node with the highest degree (i.e., the node with the highest number of connected edges). An example of refinement results is shown in Fig. 2b. Refinement can effectively remove duplicated bounding boxes. However, it can neither remove false positives nor recover false negatives. + +Both M1 and M2 are used in detection and tracking. The pseudo bounding boxes from both base models are then refined jointly, and the refined boxes are used to train M2 in adaptation. We refer this setting as cross-teaching. + +# 3.3. Location-Aware Object Mixup + +Mixup [76] is originally proposed as a simple but powerful data augmentation method in for image classification. It is quickly improved by many modifications [1, 23, 32, 66, 75]. Mixup is also used in object detection tasks [4, 75], where it can be applied at image or instance levels. During adaptation training, a frame with its pseudo bounding + +![](images/03bbdeb8e94af07618547e6e68937a16b75f0aa3c9a6fc5ce3fc841c9d9c5d34.jpg) +(a) Random mixup. + +![](images/74d6a0c14a49681b388cd6772e9f45ba252a2a8a14825519122ba04d6ba94dab.jpg) +(b) Location-aware mixup. +Figure 3. An example of location-aware object mixup. Compared to random mixup, less artifacts are introduced. + +boxes has the probability $p_{\text{mixup}}$ of being paired with another randomly select source frame for mixup. A portion $r_{\text{mixup}}$ of the pseudo bounding boxes in the source frame are cropped out and pasted onto the original frame, along with their locations and labels. If a pasted box covers more than $\alpha_{\text{cover}}$ of a pseudo box in the original frame, the pseudo box will be excluded from the pseudo labels. An example of this location-aware object mixup is shown in Fig. 2c and Fig. 3. Since the background is stationary, we propose to use location-aware mixup instead of the original random mixup to reduce the amount of artifacts in the generated mixup images and to keep the location distribution of the objects in the scene. This will be shown to be beneficial in the experiments. + +# 3.4. Dynamic Background Extraction + +For a video frame $\mathcal{I}$ of dimension $W\times H$ associated with a set of $K$ pseudo-annotated object bounding boxes $\{(x_k^1,y_k^1,x_k^2,y_k^2)|k = 1\dots K\}$ , a background mask $\mathcal{M}$ of the same dimension as $\mathcal{I}$ can be constructed as follows. For each location $(x,y)$ in $\mathcal{M}$ , we set $M[x,y] = 0$ if $(x,y)$ is inside of any pseudo-annotated bounding box and 1 otherwise. Then for a sequence of frame-mask pairs $\{(\mathcal{I}_l,\mathcal{M}_l)|l = 1\dots L\}$ , the background image is determined as + +$$ +\mathcal {B} = \frac {\sum_ {l = 1} ^ {L} \mathcal {I} _ {l} \otimes \mathcal {M} _ {l}}{\sum_ {l = 1} ^ {L} \mathcal {M} _ {l}}, \tag {1} +$$ + +where $\otimes$ is the pixel-wise multiplication operator. + +There might exist a location $(x', y')$ that lies inside an object bounding box in every image, i.e., $\mathcal{M}_l[x', y'] = 0$ . In this case, the background at this location is never observed and its pixel value cannot be determined. In this case, we "guess" the background value by inpainting inside such areas using the inpainting algorithm of [63]. + +Fig. 2d shows an example background, which is reasonably accurate. Although in each video, the camera perspective is fixed, the background can change overtime due to some factors such as illumination change. To incorporate this, the background extraction is operated every $T_{bg}$ seconds, giving a dynamic background modeling. + +We construct an object mask as an additional input modality for the detector, to utilize the extracted background to improve detection performance during adapta + +tion. For a frame $\mathcal{I}$ and its corresponding background model image $\mathcal{B}$ , we define its object mask image as + +$$ +\mathcal {M} ^ {O} = (\mathcal {I} - \mathcal {B} + 1) \times 0. 5, \tag {2} +$$ + +where we assume the pixel values in both $\mathcal{I}$ and $\mathcal{B}$ have been normalized to the range [0, 1]. An object mask example is shown in Fig. 1c, in which objects are clearly separated from the background. Note that since $\mathcal{I},\mathcal{B}$ , and $\mathcal{M}^O$ are linearly dependent, no information is lost. + +Before the adaptation training process, the base detector needs to be modified to use object mask input modality and trained on the MSCOCO training set with background images. However, dynamic background extraction on MSCOCO is impossible since it only consists of static images. Therefore, we rely on the object mask annotations in the dataset and apply the same inpainting algorithm [63] to generate the background image for each training image. + +# 3.5. Object Mask Fusion + +The architecture of Faster-RCNN [56] is depicted in Fig. 4a. First, a CNN backbone with a feature pyramid network (FPN) [39] is utilized to extract the Feature Pyramid (FP). The first layer in the CNN that takes the input image is also shown here. Then, an RPN [39] is employed to produce object bounding box proposals. Lastly, an ROI head [19,20] is used to assign object labels and refine the bounding boxes on the pooled feature maps of the proposals. The entire network is trained utilizing two sets of losses: localization loss and objectness loss from RPN, and localization loss and classification loss from the ROI head. The fusion models explained below start with the base model that is trained on the remapped MSCOCO training set. + +Early-fusion. The input to the Faster-RCNN network is the stacked image $[\mathcal{I};\mathcal{M}^{\mathcal{O}}]$ , with the two modalities stacked along the color channels, as shown in Fig. 4b. To enable the network to take a 6-channel image as input instead of the vanilla 3-channel image, the first layer of the backbone CNN is duplicated and stacked. The convolution kernel weights are first copied, then halved. This ensures that the feature map after the first layer conv1|conv2 is the same as in the vanilla model when $[\mathcal{I};\mathcal{I}]$ is fed into it, providing a smooth start for adaptation training. This modification introduces only a small number of additional model parameters, and the increase in computational cost is negligible. + +Mid-fusion. Both $\mathcal{I}$ and $\mathcal{M}^O$ are fed into the FPN in parallel, which yields two feature pyramid FP1 and FP2, as shown in Fig. 4c. Two sets of RPNs and ROI heads are used in parallel, and their initial weights are copied from the base model. FP1 is used in the left branch with RPN1 and ROI1, which is the same as the vanilla model. In the right branch with RPN2 and ROI2, the input feature pyramid is the fusion of FP1 and FP2: + +$$ +\mathrm {F P} _ {\text {f u s i o n}} = \mathrm {F P 1} \oplus \mathrm {F P 2}. \tag {3} +$$ + +![](images/94794aeb6e2f4a0443ae56d799a03a8631adf755045aa8e2a605359b16d69a6d.jpg) +(a) Vanilla Faster-RCNN. + +![](images/c3e7a84c7d82014934e52d0acf8860030d9f47dac13140a529bfa5275090b02f.jpg) +(c) Mid-fusion Faster-RCNN. + +![](images/c6b44f02073ad3f51063f656a09dbf4347ff86f0f440124b09146577ffd79648.jpg) + +![](images/3c26c955f7096337991c93c9b2f70ab6e26b1cdb38ce2160a9b4e60f00f2a0b1.jpg) +(b) Early-fusion Faster-RCNN. +(d) Late-fusion Faster-RCNN. + +The fused feature pyramid has the same dimension as both FP1 and FP2. The fusion operator can be either non-parametric such as average pooling, or parametric such as convolution layers. The whole network is trained with losses from both branches + +$$ +\mathcal {L} ^ {m i d} = \left(1 - \alpha_ {m i d}\right) \left(\mathcal {L} _ {r p n} ^ {\mathcal {I}} + \mathcal {L} _ {r o i} ^ {\mathcal {I}}\right) + \alpha_ {m i d} \left(\mathcal {L} _ {r p n} ^ {f u s i o n} + \mathcal {L} _ {r o i} ^ {f u s i o n}\right), \tag {4} +$$ + +where $\alpha_{mid}$ is the weights of different losses. + +Late-fusion is similar to mid-fusion as shown in Fig. 4d. The key difference is that only the ROI head is branched. The region proposals are only generated from FP1, and the fused feature pyramid is only used in ROI head in the right branch. As a result, the total loss is + +$$ +\mathcal {L} ^ {l a t e} = \mathcal {L} _ {r p n} ^ {\mathcal {I}} + (1 - \alpha_ {l a t e}) \mathcal {L} _ {r o i} ^ {\mathcal {I}} + \alpha_ {l a t e} \mathcal {L} _ {r o i} ^ {f u s i o n}. \tag {5} +$$ + +Both mid-fusion and late-fusion introduce significant additional parameters and computation. $\mathcal{I}$ and $\mathcal{M}^O$ need to pass the backbone, requiring more computation even at inference time. The quantitative comparison of inference speed of different fusion models can be found in the supplementary material. + +To train the fusion models in adaptation, we first fine-tune the base models on MSCOCO training set with the inpainted background models described in Sec. 3.4 until convergence. During evaluation, we use the last background image to obtain the object masks, as it is the closest to the evaluation images on the timeline. + +# 4. Scenes100 Dataset + +For our proposed scene-adaptive object detection problem, there is no existing dataset with long enough videos of stationary backgrounds for the development and evaluation of self-supervised adaptation techniques. We therefore collected a new dataset called Scenes100, which will be described in this section. + +![](images/fead86c9d20fe3d1cddfaf756c62add0c97143aac5a8b44b51a98ab63949fc69.jpg) + +![](images/13cf271c9c66c7f42573759012b562a1f7a517e55b89d4f212c4f74c5e16ab16.jpg) + +![](images/fb1b35766b0f16ffc593fc7a7ece9de10cec27d49e4f0e12865bd4cdb8ea09ff.jpg) + +![](images/c7750e5360e574180f3330f991604afb14d8562ab07f68c02a6e9ff28cf52711.jpg) +Figure 4. Illustration of different fusion models for Faster-RCNN. Network modules are indicated by green, and feature maps and outputs are indicated by yellow. +Figure 5. Some annotated frames from the evaluation portions. The green mask shows the part of the frame that is excluded for annotation and evaluation. Objects are labeled by bounding boxes. Please see the supplementary material for detailed descriptions. + +![](images/32ab7d85f076faa1a46c3d8d80d7aba865923f82d3dec575e7ff24e010374235.jpg) + +![](images/4013098c739e593643edbb8079a3dfc399eef4c9440df24b292264126343df3a.jpg) + +![](images/17fff89abfcb963752a9856b21da43afa732d96f21713b98397d582c9e9d9070.jpg) + +![](images/07c1f575910c86ef3307858986c1a1735eede54df05eb00091566aea3bad2ef0.jpg) + +![](images/aab208c4b08ac9b07b770d63acf81d18fdf26f78afd12259277ad642c0c82fc6.jpg) + +![](images/fd3fc721cbb5f65ec85d87cbd64fd49639a1b01edf4ee4a5860c3b19635499c6.jpg) + +![](images/5c428924b8bca60fe8bfee2c14e7e1dfc0bd0ad3901bb838146ca04f011277f5.jpg) + +![](images/21197db50c1e6ad1809beaa6ddee2ad0d2e1d371d1474781c5374f296b0e9609.jpg) + +# 4.1. Data Collection + +We used keywords such as "live view", "street camera", and "live webcam" to search for live streams on YouTube. We look for streams that are of decent quality and have fixed camera perspective. Around 200 candidate videos were recorded between September 2020 and February 2022. We picked 100 videos to compile our Scenes100 dataset, ensuring that each video is longer than two hours and has minimum width and height of 720 pixels. We also aimed for a dataset with high diversities. The 100 videos of Scenes100 were recorded in a variety of 16 countries and territories, and during different time periods of day and year. The dataset contain videos from both indoor and outdoor scenes, with different camera perspectives. The density and scale of visual objects also vary significantly from one video to another. Some sample frames is presented in Fig. 5 to show the diversity of our dataset. + +# 4.2. Manual Annotation + +We are interested in detecting objects from two categories: person and vehicle. The person category includes any visible and recognizable people in the frames; this corresponds to the person category in MSCOCO [41] dataset. The vehicle category includes all vehicles with four or more wheels, so motorbikes, bicycles, tricycles are not included. This category corresponds to the car, bus, and truck categories in MSCOCO. + +For each video, we used the first 1.5 hours for self-supervised and weakly-supervised training, and the remaining part for evaluation. Since many videos were taken with wide-angle lenses, some frames included faraway objects that were too small to be considered in evaluation. To address this, we manually drew polygon masks to exclude + +
datasetcontain videosaverage lengthframescountriesboundong boxesboxes per video
MSCOCO [41]No---897K-
KITTI [17]No--180K-
BDD100K [74]Yes40s120M11.8M18
CityScapes [9]Yes1.8s150K265K13
Scenes100Yes2h21.6M1684K840
+ +Table 1. Comparison of Scenes100 with other object detection datasets $(\mathbf{K} = 10^{3},\mathbf{M} = 10^{6})$ + +these faraway parts from annotation and evaluation. We then estimated the number of visual objects in the unmasked parts by first using a pre-trained detector, and then refining the estimate through human annotation. Frames for annotation were uniformly sampled from the second portion of each video, with the number of frames being inversely proportional to the estimated number of objects in the frames. This ensured that each video had roughly the same number of annotated objects. We contracted a data annotation company to perform bounding box annotation on those sampled frames. Some annotated frame samples for evaluation is shown in Fig. 5. + +# 4.3. Dataset Statistics + +In Tab. 1, we compare Scenes100 with some popular object detection datasets. MSCOCO [41] is a fully-supervised general-purpose object detection dataset, which acts as the supervised source domain. KITTI [17] only contains object detection annotation associated to individual images. In BDD100K [74], object detection annotation is given for one key frame in each video. For CityScapes [9], we consider the finely labeled instance-level segmentation masks, which are given for the $20^{\mathrm{th}}$ image from a 30 frame video snippets lasting 1.8s. + +The scale of Scenes100 is comparable to that of KITTI and CityScapes. However, Scenes100 has three unique features. First, the videos in Scenes100 are much longer than those in any other datasets with scenes, providing sufficient data for a model to adapt to each of them. Second, Scenes100 has greater diversity. Unlike some other datasets that are recorded in only a few cities with similar camera angles, background environments, and object types, Scenes100 covers several different countries, with varying weather conditions, road conditions, view of field, camera perspectives, and vehicle types. Third, each video in Scenes100 has a fixed camera perspective and stationary background scene, making background modeling possible. + +Note that many currently available datasets are not primarily intended for self-supervised adaptive object detection tasks. They may possess other features, such as segmentation masks, multi-modal sensor data, depth maps, or annotations for object tracking. However, despite the absence of these attributes, Scenes100 holds value as the first extensive and varied dataset that features fixed camera per + +spectives and lengthy videos. As such, it can complement existing dataset collections and serve as a valuable resource for researchers exploring scene adaptive object detection. + +# 5. Experiments + +# 5.1. Evaluation Protocols + +We evaluate the performance of object detectors following the COCO evaluation protocol [41] to calculate the average precision at $IoU = 50\%$ ( $AP^{50}$ ) and mean average precision of different $IoU$ thresholds from 0.5 to 0.95 ( $AP^m$ ), with some modifications as follows. + +Non-evaluation region: before feeding the ground-truth and detected bounding boxes to the COCO evaluator, the bounding boxes that have at least one corner inside the nonevaluation mask will be removed. So the faraway parts in the frames where objects are deemed too small and blurry will not affect the evaluation results. Please see Sec. 4.2 and Fig. 5 for more details. + +Category weighting: in the standard COCO evaluation protocol, the overall multi-class $AP$ is the simple average of the $AP_{s}$ of the different categories. If a class is sparse in the training set and validation set, the model tends to perform poorly due to the scarcity of the training samples. In this case the average $AP$ does not truly reflect the performance of a detector. Hence we propose an additional multi-class $AP$ metric with weighted average where the weight is determined by the prevalence of ground truth object instances in the evaluation set. This weighted $AP$ can better portray the performance of a detector in the case of long-tailed distribution. We refer the standard classes-mean COCO $AP$ as $AP_{co}$ , and the proposed weighted $AP$ as $AP_{w}$ . We refer the $AP$ averaged over IoU thresholds and IoU = 50% as $AP^{m}$ and $AP^{50}$ . + +Per-video evaluation: we aim to evaluate the performance of a scene adaption method, instead of a specific detection model. For each of the video, a model is adapted to it from the same base model trained on the source dataset. The performance of the base model and adapted model on each video is calculated individually as $\{(AP_{v,\text{base}}, AP_{v,\text{adapt}}) | v = 1 \ldots 100\}$ . Then we evaluate the overall effectiveness of the adaptation method using the averaged $AP$ gain as: + +$$ +A P G = \frac {1}{1 0 0} \sum_ {v = 1} ^ {1 0 0} \left(A P _ {v, a d a p t} - A P _ {v, b a s e}\right). \tag {6} +$$ + +# 5.2. Implementation Details + +Our implementation is based on Detector2 [70]. The DiMP tracker is directly taken from the official PyTracking [10] implementation without any change. More details and hyper-parameters can be found in the supplementary material. During training, we include the same number of + +images from MSCOCO training set as the number of unlabeled frames from the videos. The hyper-parameters are the same for all videos. We check the training loss for all experiments, including the baseline methods, to ensure convergence. We also train "compound" models where all videos are used together for adaptation. The results and further details can be found in the supplementary material. + +# 5.3. Domain Adaptation Baselines + +We compare our methods with several domain adaptive object detection baselines. In their original papers, the methods are evaluated on other domain adaptive object detection datasets. However, those datasets do not contains videos with fixed camera perspectives, so the proposed location-aware mixup and dynamic background extraction cannot be applied. We instead apply the methods in other papers on Scenes100. For a fair comparison, all methods start with the same base model and weights, and then are adapted to and evaluated on each scene individually. We keep the values of the hyper-parameters that are specific to a baseline method used in the original papers. We use the same learning rate, batch size, and number of iterations as our experiments in those baseline methods. More details can be found in the supplementary material. + +Self-Train (ST) [57] is a method that uses detection and tracking to obtain pseudo bounding boxes, similar to ours. But ST applies tracking only in one direction. It assigns different weights to the pseudo bounding boxes during training, but no refinement nor cross-teaching. + +STAC [60] uses the base model to get pseudo bounding boxes, and trains the model on a strongly augmented version of the target domain image with those boxes. + +Adaptive Teacher (AT) [37] also uses self-training, but the pseudo bounding boxes are detected on the fly. It utilizes exponential moving average to update the model gradually. It also uses weak/strong data augmentation, and a domain classifier to align the features of the two domains. + +$\mathbf{H}^2\mathbf{FA}$ R-CNN [72] applies feature alignment of source and target domains at various levels of Faster-RCNN. It assumes that image-level weak annotation is available for the target domain. So we use the labels from our pseudo-labeling process as image-level class labels. + +TIA [78] is a method that instantiates feature alignment by using both domain classifier and auxiliary classification and localization heads. The model is trained in a mini-max manner using gradient reverse layers [15]. + +LODS [33] uses style enhancement as data augmentation. The pseudo bounding boxes are generated from the original images, and the features from both versions of the same image are aligned. It also utilizes exponential moving average for the weights. It does not use any supervised training data from the source domain. + +
Model# paramsMSCOCOScenes100
APmcoAP50coAPmcoAP50coAPmcoAP50w
M1 (R-50)41.4M50.0576.4840.5262.1241.2864.65
M2 (R-101)60.5M51.2977.4641.1163.1041.9665.74
+ +Table 2. Performance of base models on MSCOCO validation set with remapped categories and Scenes100 before any adaptation. For Scenes100, $APs$ are evaluated on each video individually and then averaged. R- stands for ResNet. # indicates the number of parameters $(M = 10^6)$ . See Sec. 5.1 for $AP$ notation details. + +# 5.4. Results + +The performance of the base models is shown in Tab. 2, along with their number of trainable parameters. M2 is a bigger model with more parameters and representation capacity, so it outperforms M1 by a noticeable margin in all the metrics before adaptation. All the $AP$ gains shown in other tables are based on the performance of base model M2. Note that the $AP$ numbers on Scenes100 is already relatively high at only about 10 points below the performance on MSCOCO, so we do not observe very high ( $>5$ points) $AP$ gains in the results shown later. + +We compare the performance of different adaptation methods in terms of average $AP$ gain in Tab. 3. Here we show our best combination of methods, which incorporates pseudo-labeling, location-aware mixup, and object mask mid-fusion. Among the baseline methods for domain adaptive object detection, only ST and LODS improves the performance significantly. AT only yields marginal improvement in classes-weighted $AP$ . STAC, $\mathrm{H}^2\mathrm{FA}$ , and TIA actually degrade the performance on Scenes100. LODS performs surprisingly well considering it does not use source domain images. However, we found that any training schedule longer than 1000 iterations would lead to severe performance degradation. The proposed method yields much higher, consistent, and stable improvement. + +In general, self-training-based methods (ST, AT, STAC, and the proposed method) perform better than domain alignment-based methods $\mathrm{(H^2FA}$ and TIA). Domain alignment is usually achieved through adversarial learning, which is known to have unstable objectives and is more difficult to optimize [14, 42, 59, 61]. Therefore, these methods are more sensitive to hyperparameters and training schedules, making it challenging to determine the optimal settings for all datasets. Applying domain alignment methods directly to our scene adaptive dataset without careful tuning can cause problems as our dataset has less in-domain variance than domain adaptation datasets. Furthermore, since the scenes in our dataset are very different, they might require different settings. This implies that Scenes100 has its uniqueness compared to more generally purported domain adaptive object detection datasets. Our proposed method is more robust to this setting and does not rely heavily on + +
Method\(APG_{co}^{m}\)\(APG_{co}^{50}\)\(APG_{w}^{m}\)\(APG_{w}^{50}\)
ST [57]+0.80+0.24+1.39+1.03
STAC [60]-1.26-5.12-1.97-6.64
AT [37]-0.75-1.11+0.06+0.04
\(\mathrm{H^{2}FA}\) [72]-3.10-4.97-3.77-6.01
TIA [78]-0.32-0.37-0.32-0.33
LODS [33]+0.45+1.28+1.02+2.28
Proposed+3.76+4.45+3.78+4.65
+ +Table 3. Averaged $AP$ gain of different adaptation methods. See Sec. 5.1 for $AP$ notation details. + +
MixupFusion\(APG_{co}^{m}\)\(APG_{co}^{50}\)\(APG_{w}^{m}\)\(APG_{w}^{50}\)
XX+0.95+0.54+1.67+1.55
Location-awareX+1.72+1.67+2.25+2.53
RandomX+1.22+1.13+1.76+2.03
Xearly+1.85+2.12+2.22+2.73
Xmid+3.40+3.81+3.67+3.98
Xlate+3.34+3.60+3.38+3.72
Location-awareearly+2.25+2.82+2.59+3.46
Location-awaremid+3.76+4.45+3.78+4.65
Location-awarelate+3.66+4.10+3.73+4.31
+ +Table 4. Ablation study for the proposed components of our scene-adaptive object detection method. $\times$ means not being applied. Pseudo-labeling is always applied. This table shows the averaged $AP$ gain; see Sec. 5.1 for $AP$ notation details. + +
MixupFusion\(APG_{co}^{m}\)\(APG_{co}^{50}\)\(APG_{w}^{m}\)\(APG_{w}^{50}\)
✘AVG+3.40+3.81+3.67+3.98
✘CNN+3.25+3.52+3.26+3.79
✘ATTN+3.22+3.73+3.23+3.95
Location-awareAVG+3.76+4.45+3.78+4.65
Location-awareCNN+3.60+4.20+3.71+4.50
Location-awareATTN+3.30+3.89+3.45+4.26
+ +Table 5. Ablation study on different feature pyramid fusion methods in term of averaged $AP$ gain. AVG, CNN, and ATTN indicate mid-fusion based on average pooling, CNN, or attention module, respectively. $\pmb{x}$ means not being applied. Pseudo-labeling is always applied. See Sec. 5.1 for $AP$ notation details. + +hyperparameter tuning, as we use the same set of hyperparameters across all 100 videos. Additional analysis and success/failure cases can be found in the supplementary material. + +# 5.5. Ablation Study + +We explore the effect of the different components in our methods by experimenting various combinations of them. For all experiments, we keep the pseudo-labeling unchanged, as it is the basis of other components. We compare + +different types of object mask fusion (Fusion): early, mid, and late. We also compare the proposed Location-aware object mixup with Random mixup. Tab. 4 shows the results. The experiments on hyper-parameters can be found in the supplementary material. + +The proposed pseudo-labeling, mixup, and object mask fusion all benefit the adaptation performance. Applying only the pseudo-labeling already leads to $AP$ gain higher than ST, which is more complicated involving pseudo box weighting. Location-aware mixup is also more advantageous than Random mixup, validating the assumption in Sec. 3.3 that having fewer artifacts is better for adaptation. Among three fusion options, mid-fusion yields the best result. That is perhaps because in a mid-fusion model, the RPN utilizes the fused feature pyramid, which can contain critical information of object boundaries. The combination of mixup and object mask fusion can further improve the performance. + +We also compare different feature pyramid fusion methods described in Eq. (3). In our main experiments, we simply average the feature pyramids from the original image and the object mask. Now at each level of the pyramid, we use a separate CNN or an attention module to fuse the feature maps. The comparison is shown in Tab. 5. The parametric method performs slightly worse than average-pooling, probably because the newly introduced modules have randomly initialized weights and thus require additional supervised training data. + +# 6. Summary + +We have presented a self-supervised framework for scene-adaptive object detection. Base detectors and generic trackers are used to generate pseudo object labels as the adaptive training targets in a cross-teaching manner. To fully utilize the background equivariance when camera perspective is fixed, we propose artifacts-free location-aware object mixup to augment the input images, and dynamic background extraction for additional input modality to the detector. We also introduce the first large-scale scene adaptive object detection dataset, Scenes100, with several unique features compared to other domain adaptive object detection datasets. Our method outperforms domain adaptive object detection baselines by a large margin on Scenes100. We also conduct extensive experiments to illustrate the effectiveness of the components in our framework. We hope this work has demonstrated the importance of scene adaptation, and it will spur further research interest in this impactful but understudied area. + +Acknowledgements. This research was partially supported by NSF grants IIS-1763981 & NSDF DUE-2055406. + +# References + +[1] David Berthelot, Nicholas Carlini, Ian Goodfellow, Avital Oliver, Nicolas Papernot, and Colin Raffel. Mixmatch: A holistic approach to semi-supervised learning. In Conference on Neural Information Processing Systems (NeurIPS), 2019. 3 +[2] Goutam Bhat, Martin Danelljan, Luc Van Gool, and Radu Timofte. Learning discriminative model prediction for tracking. In International Conference on Computer Vision (ICCV), 2019. 1, 3 +[3] Alexey Bochkovskiy, Chien-Yao Wang, and Hong-Yuan Mark Liao. Yolov4: Optimal speed and accuracy of object detection. In ArXiv, 2020. 2 +[4] Shahine Bouabid and Vincent Delaitre. Mixup regularization for region proposal based object detectors. In ArXiv, 2020. 3 +[5] Zhaowei Cai and Nuno Vasconcelos. Cascade r-cnn: Delving into high quality object detection. In Computer Vision and Pattern Recognition Conference (CVPR), 2018. 2 +[6] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European Conference on Computer Vision (ECCV), 2020. 2 +[7] Yuhua Chen, Wen Li, Christos Sakaridis, Dengxin Dai, and Luc Van Gool. Domain adaptive faster r-cnn for object detection in the wild. In Computer Vision and Pattern Recognition Conference (CVPR), 2018. 2 +[8] Xiangxiang Chu, Zhi Tian, Yuqing Wang, Bo Zhang, Haibing Ren, Xiaolin Wei, Huaxia Xia, and Chunhua Shen. Twins: Revisiting the design of spatial attention in vision transformers. In Conference on Neural Information Processing Systems (NeurIPS), 2021. 2 +[9] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Computer Vision and Pattern Recognition Conference (CVPR), 2016. 6 +[10] Martin Danelljan, Goutam Bhat, Christoph Mayer, and Matthieu Paul. Pytracking. https://github.com/visionml/pytracking, 2019.6 +[11] Jinhong Deng, Wen Li, Yuhua Chen, and Lixin Duan. Unbiased mean teacher for cross-domain object detection. In Computer Vision and Pattern Recognition Conference (CVPR), 2021. 2 +[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations (ICLR), 2021. 2 +[13] Yoav Freund and Robert E Schapire. A decision-theoretic generalization of on-line learning and an application to boosting. Journal of Computer and System Sciences, 55(1):119-139, 1997. 1, 2 + +[14] Angus Galloway, Anna Golubeva, Thomas Tanay, Medhat Moussa, and Graham W. Taylor. Batch normalization is a cause of adversarial vulnerability. In *ArXiv*, 2020. 7 +[15] Yaroslav Ganin and Victor Lempitsky. Unsupervised domain adaptation by backpropagation. In International Conference on Machine Learning (ICML), 2015. 2, 7 +[16] Yaroslav Ganin, E. Ustinova, Hana Ajakan, Pascal Germain, H. Larochelle, Francois Lavolette, Mario Marchand, and Victor S. Lempitsky. Domain-adversarial training of neural networks. In Journal of Machine Learning Research, 2016. 2 +[17] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In Computer Vision and Pattern Recognition Conference (CVPR), 2012. 6 +[18] Ross Girshick. Fast r-cnn. In International Conference on Computer Vision (ICCV), 2015. 2 +[19] Ross Girshick, Jeff Donahue, Trevor Darrell, and Jitendra Malik. Rich feature hierarchies for accurate object detection and semantic segmentation. In Computer Vision and Pattern Recognition Conference (CVPR), 2014. 2, 4 +[20] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In International Conference on Computer Vision (ICCV), 2017. 4 +[21] Kaiming He, X. Zhang, Shaoqing Ren, and Jian Sun. Spatial pyramid pooling in deep convolutional networks for visual recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 37, 2015. 2 +[22] Zhenwei He and Lei Zhang. Multi-adversarial faster-rcnn for unrestricted object detection. In International Conference on Computer Vision (ICCV), 2019. 2 +[23] Dan Hendrycks, Norman Mu, Ekin D. Cubuk, Barret Zoph, Justin Gilmer, and Balaji Lakshminarayanan. AugMix: A simple data processing method to improve robustness and uncertainty. In International Conference on Learning Representations (ICLR), 2020. 3 +[24] Tin Kam Ho. Random decision forests. In International Conference on Document Analysis and Recognition (ICDAR), volume 1, 1995. 1, 2 +[25] Minh Hoai and Andrew Zisserman. Talking heads: Detecting humans and recognizing their interactions. In Computer Vision and Pattern Recognition Conference (CVPR), 2014. 2 +[26] Jiaxing Huang, Dayan Guan, Aoran Xiao, and Shijian Lu. Model adaptation: Historical contrastive learning for unsupervised domain adaptation without source data. In Conference on Neural Information Processing Systems (NeurIPS), 2021. 2 +[27] Naoto Inoue, Ryosuke Furuta, Toshihiko Yamasaki, and Kiyoharu Aizawa. Cross-domain weakly-supervised object detection through progressive domain adaptation. In Computer Vision and Pattern Recognition Conference (CVPR), 2018. 2 +[28] Glenn Jocher. Yolov5. https://github.com/ultralytics/yolov5, 2020.2 +[29] Seunghyeon Kim, Jaehoon Choi, Taekyung Kim, and Chang-ick Kim. Self-training and adversarial background regularization for unsupervised domain adaptive one-stage object detection. In International Conference on Computer Vision (ICCV), 2019. 2 + +[30] Samuli Laine and Timo Aila. Temporal ensembling for semi-supervised learning. In International Conference on Learning Representations (ICLR), 2017. 2 +[31] Chuyi Li, Lulu Li, Hongliang Jiang, Kaiheng Weng, Yifei Geng, Liang Li, Zaidan Ke, Qingyuan Li, Meng Cheng, Weiqiang Nie, Yiduo Li, Bo Zhang, Yufei Liang, Linyuan Zhou, Xiaoming Xu, Xiangxiang Chu, Xiaoming Wei, and Xiaolin Wei. Yolov6: A single-stage object detection framework for industrial applications. In ArXiv, 2022. 2 +[32] Junnan Li, Richard Socher, and Steven C.H. Hoi. Dividemix: Learning with noisy labels as semi-supervised learning. In International Conference on Learning Representations (ICLR), 2020. 3 +[33] Shuaifeng Li, Mao Ye, Xiatian Zhu, Lihua Zhou, and Lin Xiong. Source-free object detection by learning to overlook domain style. In Computer Vision and Pattern Recognition Conference (CVPR), 2022. 2, 7, 8 +[34] Wuyang Li, Xinyu Liu, and Yixuan Yuan. Sigma: Semantic-complete graph matching for domain adaptive object. In Computer Vision and Pattern Recognition Conference (CVPR), 2022. 2 +[35] Xianfeng Li, Weijie Chen, Di Xie, Shicai Yang, Peng Yuan, Shiliang Pu, and Yueting Zhuang. A free lunch for unsupervised domain adaptive object detection without source data. In AAAI Conference on Artificial Intelligence (AAAI), 2021. 2 +[36] Yanghao Li, Chao-Yuan Wu, Haoqi Fan, Karttikeya Mangalam, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. Mvity2: Improved multiscale vision transformers for classification and detection. In Computer Vision and Pattern Recognition Conference (CVPR), 2022. 2 +[37] Yu-Jhe Li, Xiaoliang Dai, Chih-Yao Ma, Yen-Cheng Liu, Kan Chen, Bichen Wu, Zijian He, Kris Kitani, and Peter Vajda. Cross-domain adaptive teacher for object detection. In Computer Vision and Pattern Recognition Conference (CVPR), 2022. 2, 7, 8 +[38] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE Transactions on Pattern Analysis and Machine Intelligence, 40(12):2935-2947, dec 2018. 2 +[39] Tsung-Yi Lin, Piotr Dollar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Computer Vision and Pattern Recognition Conference (CVPR), 2017. 4 +[40] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In International Conference on Computer Vision (ICCV), 2017. 2 +[41] Tsung-Yi Lin, Michael Maire, Serge Belongie, Lubomir Bourdev, Ross Girshick, James Hays, Pietro Perona, Deva Ramanan, C. Lawrence Zitnick, and Piotr Dólar. Microsoft coco: Common objects in context. In ArXiv, 2014. 5, 6 +[42] Chen Liu, Mathieu Salzmann, Tao Lin, Ryota Tomioka, and Sabine Susstrunk. On the loss landscape of adversarial training: Identifying challenges and how to overcome them. In Conference on Neural Information Processing Systems (NeurIPS), 2020. 7 +[43] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C. + +Berg. Ssd: Single shot multibox detector. In European Conference on Computer Vision (ECCV), 2016. 2 +[44] Yen-Cheng Liu, Chih-Yao Ma, Zijian He, Chia-Wen Kuo, Kan Chen, Peizhao Zhang, Bichen Wu, Zsolt Kira, and Peter Vajda. Unbiased teacher for semi-supervised object detection. In International Conference on Learning Representations (ICLR), 2021. 2 +[45] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In International Conference on Computer Vision (ICCV), 2021. 2 +[46] Muhammad Akhtar Munir, Muhammad Haris Khan, M. Saquib Sarfraz, and Mohsen Ali. Ssal: Synergizing between self-training and adversarial learning for domain adaptive object detection. In Conference on Neural Information Processing Systems (NeurIPS), 2021. 2 +[47] Supreeth Narasimhaswamy, Trung Nguyen, and Minh Hoai. Detecting hands and recognizing physical contact in the wild. In Conference on Neural Information Processing Systems (NeurIPS), 2020. 2 +[48] Supreeth Narasimhaswamy, Thanh Nguyen, Mingzhen Huang, and Minh Hoai. Whose hands are these? hand detection and hand-body association in the wild. In Computer Vision and Pattern Recognition Conference (CVPR), 2022. 2 +[49] Supreeth Narasimhaswamy, Zhengwei Wei, Yang Wang, Justin Zhang, and Minh Hoai. Contextual attention for hand detection in the wild. In International Conference on Computer Vision (ICCV), 2019. 2 +[50] Thanh Nguyen, Chau Pham, Khoi Nguyen, and Minh Hoai. Few-shot object counting and detection. In European Conference on Computer Vision (ECCV), 2022. 2 +[51] Kento Nishi, Yi Ding, Alex Rich, and Tobias Höllerer. Augmentation strategies for learning with noisy labels. In Computer Vision and Pattern Recognition Conference (CVPR), 2021. 2 +[52] German I. Parisi, Ronald Kemker, Jose L. Part, Christopher Kanan, and Stefan Wermter. Continual lifelong learning with neural networks: A review. Neural Networks, 113:54-71, 2019. 2 +[53] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In ArXiv, 2015. 2 +[54] Joseph Redmon and Ali Farhadi. YOLO9000: better, faster, stronger. In ArXiv, 2016. 2 +[55] Joseph Redmon and Ali Farhadi. Yolov3: An incremental improvement. In ArXiv, 2018. 2 +[56] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In Conference on Neural Information Processing Systems (NeurIPS), 2015. 2, 3, 4 +[57] Aruni RoyChowdhury, Prithvijit Chakrabarty, Ashish Singh, SouYoung Jin, Huaizu Jiang, Liangliang Cao, and Erik Learned-Miller. Automatic adaptation of object detectors to new domains using self-training. In Computer Vision and Pattern Recognition Conference (CVPR), 2019. 1, 2, 3, 7, 8 + +[58] Kuniaki Saito, Yoshitaka Ushiku, Tatsuya Harada, and Kate Saenko. Strong-weak distribution alignment for adaptive object detection. In Computer Vision and Pattern Recognition Conference (CVPR), 2019. 2 +[59] Chawin Sitawarin, Supriyo Chakraborty, and David A. Wagner. Improving adversarial robustness through progressive hardening. In ArXiv, 2020. 7 +[60] Kihyuk Sohn, Zizhao Zhang, Chun-Liang Li, Han Zhang, Chen-Yu Lee, and Tomas Pfister. A simple semi-supervised learning framework for object detection. In arXiv, 2020. 2, 7, 8 +[61] Arvind P. Sridhar, Chawin Sitawarin, and David Wagner. Mitigating adversarial training instability with batch normalization. In International Conference on Learning Representations Workshop (ICLRW), 2021. 7 +[62] Antti Tarvainen and Harri Valpola. Weight-averaged consistency targets improve semi-supervised deep learning results. In Conference on Neural Information Processing Systems (NeurIPS), 03 2017. 2 +[63] Alexandru Cristian Telea. An image inpainting technique based on the fast marching method. Journal of Graphics Tools, 9:23-34, 2004. 4 +[64] Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. FCOS: Fully convolutional one-stage object detection. In International Conference on Computer Vision (ICCV), 2019. 2 +[65] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Conference on Neural Information Processing Systems (NeurIPS), 2017. 2 +[66] Vikas Verma, Alex Lamb, Christopher Beckham, Amir Najafi, Ioannis Mitliagkas, David Lopez-Paz, and Yoshua Bengio. *Manifold mixup: Better representations by interpolating hidden states.* In International Conference on Machine Learning (ICML), 2019. 3 +[67] Chien-Yao Wang, Alexey Bochkovskiy, and Hong-Yuan Mark Liao. Yolov7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors. In ArXiv, 2022. 2 +[68] Chien-Yao Wang, I-Hau Yeh, and Hongpeng Liao. You only learn one representation: Unified network for multiple tasks. In ArXiv, 2021. 2 +[69] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In International Conference on Computer Vision (ICCV), 2021. 2 +[70] Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. Detector2. https://github.com/facebookresearch/detectron2, 2019.6 +[71] Lin Xiong, Mao Ye, Dan Zhang, Yan Gan, Xue Li, and Yingying Zhu. Source data-free domain adaptation of object detector through domain-specific perturbation. International Journal of Intelligent Systems, 36:3746 - 3766, 2021. 2 +[72] Yunqiu Xu, Yifan Sun, Zongxin Yang, Jiaxu Miao, and Yi Yang. H $^2$ FA R-CNN: Holistic and hierarchical feature alignment for cross-domain weakly supervised object detection. In Computer Vision and Pattern Recognition Conference (CVPR), 2022. 2, 7, 8 + +[73] Jihan Yang, Shaoshuai Shi, Zhe Wang, Hongsheng Li, and Xiaojuan Qi. St3d: Self-training for unsupervised domain adaptation on 3d object detection. In Computer Vision and Pattern Recognition Conference (CVPR), 2021. 2 +[74] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darryll. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In Computer Vision and Pattern Recognition Conference (CVPR), 2020. 6 +[75] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In International Conference on Computer Vision (ICCV), 2019. 3 +[76] Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In ArXiv, 2017. 1, 3 +[77] Pengchuan Zhang, Xiyang Dai, Jianwei Yang, Bin Xiao, Lu Yuan, Lei Zhang, and Jianfeng Gao. Multi-scale vision longformer: A new vision transformer for high-resolution image encoding. In International Conference on Computer Vision (ICCV), 2021. 2 +[78] Liang Zhao and Limin Wang. Task-specific inconsistency alignment for domain adaptive object detection. In Computer Vision and Pattern Recognition Conference (CVPR), 2022. 2, 7, 8 +[79] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable DETR: deformable transformers for end-to-end object detection. In ArXiv, 2020. 2 \ No newline at end of file diff --git a/2023/Object Detection With Self-Supervised Scene Adaptation/images.zip b/2023/Object Detection With Self-Supervised Scene Adaptation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..578df48050fe5e7f8cacf6a239168cf928b34807 --- /dev/null +++ b/2023/Object Detection With Self-Supervised Scene Adaptation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b10127948ecf7bd479870132858cd2b2e0df82a23a049620a435342bebf4e01b +size 475674 diff --git a/2023/Object Detection With Self-Supervised Scene Adaptation/layout.json b/2023/Object Detection With Self-Supervised Scene Adaptation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a74a2c3f02face0d563c442d66d6706bc298d959 --- /dev/null +++ b/2023/Object Detection With Self-Supervised Scene Adaptation/layout.json @@ -0,0 +1,10257 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 123, + 103, + 471, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 103, + 471, + 121 + ], + "spans": [ + { + "bbox": [ + 123, + 103, + 471, + 121 + ], + "type": "text", + "content": "Object Detection with Self-Supervised Scene Adaptation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 198, + 142, + 394, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 142, + 394, + 157 + ], + "spans": [ + { + "bbox": [ + 198, + 142, + 394, + 157 + ], + "type": "text", + "content": "Zekun Zhang1 Minh Hoai1,2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 160, + 162, + 433, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 162, + 433, + 176 + ], + "spans": [ + { + "bbox": [ + 160, + 162, + 433, + 176 + ], + "type": "text", + "content": "1Stony Brook University, Stony Brook, NY 11794, USA" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 214, + 177, + 380, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 177, + 380, + 190 + ], + "spans": [ + { + "bbox": [ + 214, + 177, + 380, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 214, + 177, + 380, + 190 + ], + "type": "text", + "content": "VinAI Research, Hanoi, Vietnam" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 195, + 193, + 395, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 195, + 193, + 395, + 204 + ], + "spans": [ + { + "bbox": [ + 195, + 193, + 395, + 204 + ], + "type": "text", + "content": "{zekzhang,minhhooai}@cs.stonyrbook.edu" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 232, + 192, + 245 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 232, + 192, + 245 + ], + "spans": [ + { + "bbox": [ + 143, + 232, + 192, + 245 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 45, + 258, + 290, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 258, + 290, + 474 + ], + "spans": [ + { + "bbox": [ + 45, + 258, + 290, + 474 + ], + "type": "text", + "content": "This paper proposes a novel method to improve the performance of a trained object detector on scenes with fixed camera perspectives based on self-supervised adaptation. Given a specific scene, the trained detector is adapted using pseudo-ground truth labels generated by the detector itself and an object tracker in a cross-teaching manner. When the camera perspective is fixed, our method can utilize the background equivariance by proposing artifact-free object mixup as a means of data augmentation, and utilize accurate background extraction as an additional input modality. We also introduce a large-scale and diverse dataset for the development and evaluation of scene-adaptive object detection. Experiments on this dataset show that our method can improve the average precision of the original detector, outperforming the previous state-of-the-art self-supervised domain adaptive object detection methods by a large margin. Our dataset and code are published at https://github.com/cvlab-stonybrook/scenes100." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 499, + 128, + 512 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 499, + 128, + 512 + ], + "spans": [ + { + "bbox": [ + 47, + 499, + 128, + 512 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 521, + 287, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 287, + 664 + ], + "type": "text", + "content": "The need to detect objects in video streams from stationary cameras arises in many computer vision applications, including video surveillance and autonomous retail. In general, different applications require the detection of different object categories, and each computer-vision-based product will have its own detector. However, for a specific product, there is typically a single detector that will be used for many cameras/scenes. For example, a typical video surveillance product would use the same detector to detect pedestrians and vehicles for network cameras installed at different locations. Unfortunately, a single detector might not work well for all scenes, leading to trivial and unforgiving mistakes." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": "This fundamental problem of many computer vision products stems from the limited generalization power of a single model, due to limited training data, limited model capacity, or both. One can attempt to address this problem by" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 233, + 545, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 233, + 545, + 305 + ], + "spans": [ + { + "bbox": [ + 304, + 233, + 545, + 305 + ], + "type": "text", + "content": "using more training data, but it will incur additional cost for data collection and annotation. Furthermore, in many cases, due to the low latency requirement or the limited computing resources for inference, a product is forced to use a very lightweight network, and this network will have limited representation capacity to generalize across many scenes." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 306, + 545, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 306, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 304, + 306, + 545, + 376 + ], + "type": "text", + "content": "In this paper, instead of having a single scene-generic detector, we propose using scene-specific detectors. This yields higher detection performance as each detector is customized for a specific scene, and allows us to use a lightweight model without sacrificing accuracy as each detector is only responsible for one scene." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 378, + 546, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 378, + 546, + 593 + ], + "spans": [ + { + "bbox": [ + 304, + 378, + 546, + 593 + ], + "type": "text", + "content": "Obtaining scene-specific detectors, however, is very challenging. A trivial approach is to train a detector for each scene separately, but this requires an enormous amount of annotated training data. Instead, we propose a self-supervised method to adapt a pre-trained detector to each scene. Our method records the unlabeled video frames in the past, uses the trained detector to detect objects in those frames, and generates augmented training data based on those detections. Although the detections made by the pre-trained model can be noisy, they can still be useful for generating pseudo annotated data. We further extend those pseudo bounding boxes by applying object tracking [2, 57] along the video timeline, aiming to propagate the detections to adjacent frames to recover some of the false negatives not returned by the detector. We also use multiple detectors to obtain the pseudo labels and train the detector in a cross-teaching manner, taking the advantage of the ensemble of models [13, 24]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 594, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 689 + ], + "type": "text", + "content": "Exploiting the stationary nature of the camera, we propose two additional techniques to boost the detection performance: location-aware mixup and background-augmented input. The former is to generate more samples during training through object mixup [76] that contains less artifacts, based on the aforementioned pseudo boxes generated from detection and tracking. The latter involves estimating the background image and fusing it with the detector's input." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "In short, the main contribution of our paper is a novel framework that utilizes self-supervision, location-aware ob" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21589" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "content": "ject mixup, and background modeling to improve the detection performance of a pre-trained object detector on scenes with stationary cameras. We also contribute a large scale and diverse dataset for the development of scene adaptive object detection, which contains sufficient quantity and quality annotations for evaluation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 155, + 134, + 167 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 155, + 134, + 167 + ], + "spans": [ + { + "bbox": [ + 47, + 155, + 134, + 167 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 175, + 287, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 175, + 287, + 317 + ], + "spans": [ + { + "bbox": [ + 46, + 175, + 287, + 317 + ], + "type": "text", + "content": "Despite much recent improvement in object detection research [3,5,6,8,12,18,19,21,25,28,31,36,40,43,45,47-50, 53-56,64,65,67-69,77,79], trained detectors still encounter problems with domain shift or domain gap. Scene adaptive object detection can be viewed as a special case of domain adaption, in which an object detector trained on a fully-supervised source domain is adapted to a target domain. Most research in this direction focuses on semi-supervised, weakly-supervised, or self-supervised adaptation. In addition, source-free adaptation [26, 33, 35, 71] has been proposed to address the situation where the source domain data is unavailable during adaptation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 319, + 288, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 319, + 288, + 521 + ], + "spans": [ + { + "bbox": [ + 46, + 319, + 288, + 521 + ], + "type": "text", + "content": "Self-labeling methods [11,27,29,33,37,44,46,57,60,73] uses the teacher-student setup from semi-supervised image classification [30, 62]. A teacher detector trained on the source domain generates pseudo bounding boxes on the target domain images, and a student model is trained with those boxes to improve its performance on the target domain. Techniques such as weak/strong augmentation [37,44,60], knowledge distillation [11], and weight averaging [33, 37] have been used to deal with noisy pseudo labels from the teacher detector due to the domain gap. Our proposed pseudo-labeling method uses two teacher models trained on the source domain to train a student model, and their pseudo boxes are aggregated and refined. We further apply tracking as in [57], but with both forward and backward directions, to extend the pseudo labels. The pseudo boxes also form the basis for later location-aware mixup and dynamic background extraction steps." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 522, + 287, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 287, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 287, + 617 + ], + "type": "text", + "content": "Domain alignment methods [7, 11, 22, 33, 34, 58, 72, 78] aim to reduce the domain gap by enforcing the models to output similarly on the source and target domain at image, proposal, or instance levels. They use domain adversarial learning through gradient reversal [15, 16] or graph matching [34] for domain alignment. This approach involves adding domain alignment losses to existing object detection models and is complementary to our proposed method." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "type": "text", + "content": "Another seemingly related research area is continual or incremental learning. Scene adaptive and domain adaptive object detection are in between task-continual learning [38] and data-continual learning [52], for that the object categories of the source and target domains are the same, but the data distribution differs. It is possible to formulate scene adaptive object detection as an online learning problem, but this approach would require additional annotations." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 307, + 68, + 419, + 137 + ], + "blocks": [ + { + "bbox": [ + 307, + 68, + 419, + 137 + ], + "lines": [ + { + "bbox": [ + 307, + 68, + 419, + 137 + ], + "spans": [ + { + "bbox": [ + 307, + 68, + 419, + 137 + ], + "type": "image", + "image_path": "57ebeeecc3d09929f50445e8a6752363c0424a4c7ac22dbb1bfdd43d990bb94f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 329, + 137, + 389, + 144 + ], + "lines": [ + { + "bbox": [ + 329, + 137, + 389, + 144 + ], + "spans": [ + { + "bbox": [ + 329, + 137, + 389, + 144 + ], + "type": "text", + "content": "(a) Input image " + }, + { + "bbox": [ + 329, + 137, + 389, + 144 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 329, + 137, + 389, + 144 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 430, + 72, + 542, + 136 + ], + "blocks": [ + { + "bbox": [ + 430, + 72, + 542, + 136 + ], + "lines": [ + { + "bbox": [ + 430, + 72, + 542, + 136 + ], + "spans": [ + { + "bbox": [ + 430, + 72, + 542, + 136 + ], + "type": "image", + "image_path": "360d51fccf503508fa75a943d66deaef16f2f5d493c82dc5ac89ee2b717c23eb.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 454, + 137, + 533, + 152 + ], + "lines": [ + { + "bbox": [ + 454, + 137, + 533, + 152 + ], + "spans": [ + { + "bbox": [ + 454, + 137, + 533, + 152 + ], + "type": "text", + "content": "(b) Learned background image " + }, + { + "bbox": [ + 454, + 137, + 533, + 152 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 454, + 137, + 533, + 152 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 307, + 144, + 418, + 206 + ], + "blocks": [ + { + "bbox": [ + 307, + 144, + 418, + 206 + ], + "lines": [ + { + "bbox": [ + 307, + 144, + 418, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 144, + 418, + 206 + ], + "type": "image", + "image_path": "0b7af89dc2853a1079de2d8f2c6ac3e668569254650dddd42448c5a64941b95b.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 326, + 206, + 399, + 216 + ], + "lines": [ + { + "bbox": [ + 326, + 206, + 399, + 216 + ], + "spans": [ + { + "bbox": [ + 326, + 206, + 399, + 216 + ], + "type": "text", + "content": "(c) Object mask " + }, + { + "bbox": [ + 326, + 206, + 399, + 216 + ], + "type": "inline_equation", + "content": "\\mathcal{M}^O" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 221, + 545, + 255 + ], + "lines": [ + { + "bbox": [ + 304, + 221, + 545, + 255 + ], + "spans": [ + { + "bbox": [ + 304, + 221, + 545, + 255 + ], + "type": "text", + "content": "Figure 1. Inference flow of the proposed detector that can adapt to a specific scene. The background image and fusion faster-RCNN model are learned during training, discussed in Sec. 3." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 260, + 366, + 273 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 260, + 366, + 273 + ], + "spans": [ + { + "bbox": [ + 306, + 260, + 366, + 273 + ], + "type": "text", + "content": "3. Methods" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 281, + 545, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 281, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 304, + 281, + 545, + 376 + ], + "type": "text", + "content": "For an object detector pre-trained on a source dataset referred as the base model, we aim to improve its performance on video stream from a stationary camera in a given scene, with unlabeled videos captured by the same camera in the past as adaptation training samples. The categories of object of interest in the scene are also present in the source dataset, but the data distribution is shifted. The base model is adapted to each scene independently." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 377, + 546, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 377, + 546, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 377, + 546, + 510 + ], + "type": "text", + "content": "Assuming the distribution shift is moderate, the base detector can generate partially decent bounding boxes on the unlabeled frames. We use them as pseudo-labels to improve the base model itself. Many objects are ignored by the base detector, as their appearance can be very different from the source dataset. However, videos contain temporal correlation information, making it possible to apply object tracking in both time directions to recover some of those false negatives. The pseudo-labels are then refined to remove low-confidence boxes and duplicates. We refer to this process of obtaining pseudo bounding boxes as pseudo-labeling." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 510, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 617 + ], + "type": "text", + "content": "As data augmentation is beneficial for training on less reliable labels [51], we take advantages of object mixup to generate new sample images. Since the background is stationary, those images contain fewer artifacts, which is beneficial for detectors. We further adopt the idea of ensemble models [13, 24] by using two base detectors to obtain the pseudo labels, for they can be complementary to each other. The pseudo boxes are used to train one base model, which we refer as cross-teaching." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 618, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 545, + 714 + ], + "type": "text", + "content": "Since the job of a detector is to separate foreground objects from the background, having information about the background is advantageous. In a stationary camera video with moving objects, different parts of the background are covered by objects in different frames, making it possible to model the complete background. We use the aforementioned pseudo bounding boxes to mark the uncovered parts of the background in each frame. Combining the partial" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21590" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 67, + 547, + 196 + ], + "blocks": [ + { + "bbox": [ + 47, + 67, + 547, + 196 + ], + "lines": [ + { + "bbox": [ + 47, + 67, + 547, + 196 + ], + "spans": [ + { + "bbox": [ + 47, + 67, + 547, + 196 + ], + "type": "image", + "image_path": "77df66211452e80f23b46a4eed31c96da2966fc781a8a6759bf5232e0449e947.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 199, + 547, + 222 + ], + "lines": [ + { + "bbox": [ + 46, + 199, + 547, + 222 + ], + "spans": [ + { + "bbox": [ + 46, + 199, + 547, + 222 + ], + "type": "text", + "content": "Figure 2. Training flow of the proposed self-supervised scene adaptation object detection framework with sample images. The mixup training images with pseudo bounding boxes and background image are used in the fusion training. The details are discussed in Sec. 3." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 243, + 287, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 243, + 287, + 290 + ], + "spans": [ + { + "bbox": [ + 46, + 243, + 287, + 290 + ], + "type": "text", + "content": "background from a sequence of frames can give an accurate and complete background model. We then modify the Faster-RCNN [56] architecture to fuse the background with the input image to improve detection performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 290, + 289, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 290, + 289, + 434 + ], + "spans": [ + { + "bbox": [ + 46, + 290, + 289, + 434 + ], + "type": "text", + "content": "The overall structure and data flows of the proposed methods are shown in Fig. 1 and Fig. 2. Our proposed methods are highly flexible and can be configured to trade-off between performance and speed. Object tracking can be bypassed to reduce the time to get the pseudo-bounding boxes. Mixup and object mask fusion can also be disabled for faster training and inference. In Sec. 5.5 we will show ablation study results on how those components affect the performance of the adapted models. We now describe each component in more details. For consistency, the images are all from the first video in our collected dataset, which will be described in Sec. 4." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 441, + 265, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 441, + 265, + 453 + ], + "spans": [ + { + "bbox": [ + 47, + 441, + 265, + 453 + ], + "type": "text", + "content": "3.1. Base Detection Models on Source Domain" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 460, + 287, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 460, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 287, + 555 + ], + "type": "text", + "content": "We use Faster-RCNN [56] in our experiments for its high precision and flexible modular architecture. Among the techniques proposed for self-supervised scene adaptation, both pseudo-labeling and mixup can be directly applied to any object detection model. The object mask fusion technique requires modification to the network architecture, but it can be easily extended to most types of detectors by adding a parallel branch to the existing network." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 555, + 287, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 555, + 287, + 639 + ], + "spans": [ + { + "bbox": [ + 46, + 555, + 287, + 639 + ], + "type": "text", + "content": "The object labels in MSCOCO training set are remapped to person and vehicle as described in Sec. 4.2. All objects other than person, car, bus, and truck are discarded. We take Faster-RCNNs pre-trained on the original MSCOCO training set, and finetune them on the remapped training set. We choose two models with ResNet-50 and ResNet101 backbones as base models, referred as M1 and M2." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 647, + 276, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 647, + 276, + 659 + ], + "spans": [ + { + "bbox": [ + 47, + 647, + 276, + 659 + ], + "type": "text", + "content": "3.2. Pseudo Labels from Detection and Tracking" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 665, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 713 + ], + "type": "text", + "content": "We take the frames from the first training portion of each video, and adopt the idea of detection and tracking similar to [57], then apply our proposed bounding box refinement and cross-teaching schemes. Base detectors described in" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 243, + 545, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 243, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 304, + 243, + 545, + 277 + ], + "type": "text", + "content": "Sec. 3.1 are applied on the frames, and any bounding box with confidence score higher than " + }, + { + "bbox": [ + 304, + 243, + 545, + 277 + ], + "type": "inline_equation", + "content": "\\lambda_{det}" + }, + { + "bbox": [ + 304, + 243, + 545, + 277 + ], + "type": "text", + "content": " is kept as a valid pseudo box." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 278, + 546, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 278, + 546, + 398 + ], + "spans": [ + { + "bbox": [ + 304, + 278, + 546, + 398 + ], + "type": "text", + "content": "Single-object trackers are initialized with detected objects having confidence scores higher than " + }, + { + "bbox": [ + 304, + 278, + 546, + 398 + ], + "type": "inline_equation", + "content": "\\lambda_{\\text{tot}}" + }, + { + "bbox": [ + 304, + 278, + 546, + 398 + ], + "type": "text", + "content": ". We use DiMP-50 [2] to track the objects in both forward and backward directions in the video timeline, and the tracked object bounding boxes are used as pseudo boxes. The reason for tracking in both directions is that if an object is moving towards the camera and appears bigger by time, backward tracking is more accurate because the initial bounding box contains more detailed texture. To prevent drifting, the maximum length of each track is capped at two seconds." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 399, + 547, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 399, + 547, + 566 + ], + "spans": [ + { + "bbox": [ + 304, + 399, + 547, + 566 + ], + "type": "text", + "content": "Bounding box refinement is applied to eliminate duplicated boxes of the same object from combining pseudo bounding boxes from detector and tracker. Each candidate box in the same frame is regarded as a graph node. If two boxes have the same object label, and their IoU is above a threshold " + }, + { + "bbox": [ + 304, + 399, + 547, + 566 + ], + "type": "inline_equation", + "content": "\\lambda_{iou}" + }, + { + "bbox": [ + 304, + 399, + 547, + 566 + ], + "type": "text", + "content": ", then an edge is added to connect them. We assume that each connected component of the graph represents one single object, and the nodes in the component are duplicated. We only keep the node with the highest degree (i.e., the node with the highest number of connected edges). An example of refinement results is shown in Fig. 2b. Refinement can effectively remove duplicated bounding boxes. However, it can neither remove false positives nor recover false negatives." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 566, + 547, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 566, + 547, + 615 + ], + "spans": [ + { + "bbox": [ + 304, + 566, + 547, + 615 + ], + "type": "text", + "content": "Both M1 and M2 are used in detection and tracking. The pseudo bounding boxes from both base models are then refined jointly, and the refined boxes are used to train M2 in adaptation. We refer this setting as cross-teaching." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 623, + 471, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 623, + 471, + 635 + ], + "spans": [ + { + "bbox": [ + 306, + 623, + 471, + 635 + ], + "type": "text", + "content": "3.3. Location-Aware Object Mixup" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 547, + 715 + ], + "type": "text", + "content": "Mixup [76] is originally proposed as a simple but powerful data augmentation method in for image classification. It is quickly improved by many modifications [1, 23, 32, 66, 75]. Mixup is also used in object detection tasks [4, 75], where it can be applied at image or instance levels. During adaptation training, a frame with its pseudo bounding" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21591" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 167, + 137 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 167, + 137 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 167, + 137 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 167, + 137 + ], + "type": "image", + "image_path": "03bbdeb8e94af07618547e6e68937a16b75f0aa3c9a6fc5ce3fc841c9d9c5d34.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 76, + 140, + 139, + 149 + ], + "lines": [ + { + "bbox": [ + 76, + 140, + 139, + 149 + ], + "spans": [ + { + "bbox": [ + 76, + 140, + 139, + 149 + ], + "type": "text", + "content": "(a) Random mixup." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 168, + 70, + 287, + 137 + ], + "blocks": [ + { + "bbox": [ + 168, + 70, + 287, + 137 + ], + "lines": [ + { + "bbox": [ + 168, + 70, + 287, + 137 + ], + "spans": [ + { + "bbox": [ + 168, + 70, + 287, + 137 + ], + "type": "image", + "image_path": "74d6a0c14a49681b388cd6772e9f45ba252a2a8a14825519122ba04d6ba94dab.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 184, + 140, + 270, + 149 + ], + "lines": [ + { + "bbox": [ + 184, + 140, + 270, + 149 + ], + "spans": [ + { + "bbox": [ + 184, + 140, + 270, + 149 + ], + "type": "text", + "content": "(b) Location-aware mixup." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 153, + 287, + 175 + ], + "lines": [ + { + "bbox": [ + 46, + 153, + 287, + 175 + ], + "spans": [ + { + "bbox": [ + 46, + 153, + 287, + 175 + ], + "type": "text", + "content": "Figure 3. An example of location-aware object mixup. Compared to random mixup, less artifacts are introduced." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 182, + 289, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 182, + 289, + 350 + ], + "spans": [ + { + "bbox": [ + 46, + 182, + 289, + 350 + ], + "type": "text", + "content": "boxes has the probability " + }, + { + "bbox": [ + 46, + 182, + 289, + 350 + ], + "type": "inline_equation", + "content": "p_{\\text{mixup}}" + }, + { + "bbox": [ + 46, + 182, + 289, + 350 + ], + "type": "text", + "content": " of being paired with another randomly select source frame for mixup. A portion " + }, + { + "bbox": [ + 46, + 182, + 289, + 350 + ], + "type": "inline_equation", + "content": "r_{\\text{mixup}}" + }, + { + "bbox": [ + 46, + 182, + 289, + 350 + ], + "type": "text", + "content": " of the pseudo bounding boxes in the source frame are cropped out and pasted onto the original frame, along with their locations and labels. If a pasted box covers more than " + }, + { + "bbox": [ + 46, + 182, + 289, + 350 + ], + "type": "inline_equation", + "content": "\\alpha_{\\text{cover}}" + }, + { + "bbox": [ + 46, + 182, + 289, + 350 + ], + "type": "text", + "content": " of a pseudo box in the original frame, the pseudo box will be excluded from the pseudo labels. An example of this location-aware object mixup is shown in Fig. 2c and Fig. 3. Since the background is stationary, we propose to use location-aware mixup instead of the original random mixup to reduce the amount of artifacts in the generated mixup images and to keep the location distribution of the objects in the scene. This will be shown to be beneficial in the experiments." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 359, + 226, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 359, + 226, + 372 + ], + "spans": [ + { + "bbox": [ + 47, + 359, + 226, + 372 + ], + "type": "text", + "content": "3.4. Dynamic Background Extraction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 378, + 287, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 378, + 287, + 484 + ], + "spans": [ + { + "bbox": [ + 47, + 378, + 287, + 484 + ], + "type": "text", + "content": "For a video frame " + }, + { + "bbox": [ + 47, + 378, + 287, + 484 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 47, + 378, + 287, + 484 + ], + "type": "text", + "content": " of dimension " + }, + { + "bbox": [ + 47, + 378, + 287, + 484 + ], + "type": "inline_equation", + "content": "W\\times H" + }, + { + "bbox": [ + 47, + 378, + 287, + 484 + ], + "type": "text", + "content": " associated with a set of " + }, + { + "bbox": [ + 47, + 378, + 287, + 484 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 47, + 378, + 287, + 484 + ], + "type": "text", + "content": " pseudo-annotated object bounding boxes " + }, + { + "bbox": [ + 47, + 378, + 287, + 484 + ], + "type": "inline_equation", + "content": "\\{(x_k^1,y_k^1,x_k^2,y_k^2)|k = 1\\dots K\\}" + }, + { + "bbox": [ + 47, + 378, + 287, + 484 + ], + "type": "text", + "content": ", a background mask " + }, + { + "bbox": [ + 47, + 378, + 287, + 484 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 47, + 378, + 287, + 484 + ], + "type": "text", + "content": " of the same dimension as " + }, + { + "bbox": [ + 47, + 378, + 287, + 484 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 47, + 378, + 287, + 484 + ], + "type": "text", + "content": " can be constructed as follows. For each location " + }, + { + "bbox": [ + 47, + 378, + 287, + 484 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 47, + 378, + 287, + 484 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 47, + 378, + 287, + 484 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 47, + 378, + 287, + 484 + ], + "type": "text", + "content": ", we set " + }, + { + "bbox": [ + 47, + 378, + 287, + 484 + ], + "type": "inline_equation", + "content": "M[x,y] = 0" + }, + { + "bbox": [ + 47, + 378, + 287, + 484 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 47, + 378, + 287, + 484 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 47, + 378, + 287, + 484 + ], + "type": "text", + "content": " is inside of any pseudo-annotated bounding box and 1 otherwise. Then for a sequence of frame-mask pairs " + }, + { + "bbox": [ + 47, + 378, + 287, + 484 + ], + "type": "inline_equation", + "content": "\\{(\\mathcal{I}_l,\\mathcal{M}_l)|l = 1\\dots L\\}" + }, + { + "bbox": [ + 47, + 378, + 287, + 484 + ], + "type": "text", + "content": ", the background image is determined as" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 122, + 484, + 287, + 515 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 122, + 484, + 287, + 515 + ], + "spans": [ + { + "bbox": [ + 122, + 484, + 287, + 515 + ], + "type": "interline_equation", + "content": "\\mathcal {B} = \\frac {\\sum_ {l = 1} ^ {L} \\mathcal {I} _ {l} \\otimes \\mathcal {M} _ {l}}{\\sum_ {l = 1} ^ {L} \\mathcal {M} _ {l}}, \\tag {1}", + "image_path": "fe954c1e262fed2d5fdc067ea2503b78d74851a724c512fe5a7a85c0fc622673.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 521, + 249, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 521, + 249, + 533 + ], + "spans": [ + { + "bbox": [ + 47, + 521, + 249, + 533 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 521, + 249, + 533 + ], + "type": "inline_equation", + "content": "\\otimes" + }, + { + "bbox": [ + 47, + 521, + 249, + 533 + ], + "type": "text", + "content": " is the pixel-wise multiplication operator." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 533, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 287, + 605 + ], + "type": "text", + "content": "There might exist a location " + }, + { + "bbox": [ + 46, + 533, + 287, + 605 + ], + "type": "inline_equation", + "content": "(x', y')" + }, + { + "bbox": [ + 46, + 533, + 287, + 605 + ], + "type": "text", + "content": " that lies inside an object bounding box in every image, i.e., " + }, + { + "bbox": [ + 46, + 533, + 287, + 605 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_l[x', y'] = 0" + }, + { + "bbox": [ + 46, + 533, + 287, + 605 + ], + "type": "text", + "content": ". In this case, the background at this location is never observed and its pixel value cannot be determined. In this case, we \"guess\" the background value by inpainting inside such areas using the inpainting algorithm of [63]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 605, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 677 + ], + "type": "text", + "content": "Fig. 2d shows an example background, which is reasonably accurate. Although in each video, the camera perspective is fixed, the background can change overtime due to some factors such as illumination change. To incorporate this, the background extraction is operated every " + }, + { + "bbox": [ + 46, + 605, + 287, + 677 + ], + "type": "inline_equation", + "content": "T_{bg}" + }, + { + "bbox": [ + 46, + 605, + 287, + 677 + ], + "type": "text", + "content": " seconds, giving a dynamic background modeling." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 714 + ], + "type": "text", + "content": "We construct an object mask as an additional input modality for the detector, to utilize the extracted background to improve detection performance during adapta" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": "tion. For a frame " + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": " and its corresponding background model image " + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": ", we define its object mask image as" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 369, + 101, + 545, + 114 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 101, + 545, + 114 + ], + "spans": [ + { + "bbox": [ + 369, + 101, + 545, + 114 + ], + "type": "interline_equation", + "content": "\\mathcal {M} ^ {O} = (\\mathcal {I} - \\mathcal {B} + 1) \\times 0. 5, \\tag {2}", + "image_path": "3a5b371d2a85696b82f076fe56919a11208576416ed7fd9d94469ea9a223d4df.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 120, + 545, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 120, + 545, + 179 + ], + "spans": [ + { + "bbox": [ + 304, + 120, + 545, + 179 + ], + "type": "text", + "content": "where we assume the pixel values in both " + }, + { + "bbox": [ + 304, + 120, + 545, + 179 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 304, + 120, + 545, + 179 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 120, + 545, + 179 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 304, + 120, + 545, + 179 + ], + "type": "text", + "content": " have been normalized to the range [0, 1]. An object mask example is shown in Fig. 1c, in which objects are clearly separated from the background. Note that since " + }, + { + "bbox": [ + 304, + 120, + 545, + 179 + ], + "type": "inline_equation", + "content": "\\mathcal{I},\\mathcal{B}" + }, + { + "bbox": [ + 304, + 120, + 545, + 179 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 120, + 545, + 179 + ], + "type": "inline_equation", + "content": "\\mathcal{M}^O" + }, + { + "bbox": [ + 304, + 120, + 545, + 179 + ], + "type": "text", + "content": " are linearly dependent, no information is lost." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 180, + 545, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 180, + 545, + 276 + ], + "spans": [ + { + "bbox": [ + 304, + 180, + 545, + 276 + ], + "type": "text", + "content": "Before the adaptation training process, the base detector needs to be modified to use object mask input modality and trained on the MSCOCO training set with background images. However, dynamic background extraction on MSCOCO is impossible since it only consists of static images. Therefore, we rely on the object mask annotations in the dataset and apply the same inpainting algorithm [63] to generate the background image for each training image." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 281, + 424, + 294 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 281, + 424, + 294 + ], + "spans": [ + { + "bbox": [ + 306, + 281, + 424, + 294 + ], + "type": "text", + "content": "3.5. Object Mask Fusion" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 300, + 545, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 300, + 545, + 455 + ], + "spans": [ + { + "bbox": [ + 304, + 300, + 545, + 455 + ], + "type": "text", + "content": "The architecture of Faster-RCNN [56] is depicted in Fig. 4a. First, a CNN backbone with a feature pyramid network (FPN) [39] is utilized to extract the Feature Pyramid (FP). The first layer in the CNN that takes the input image is also shown here. Then, an RPN [39] is employed to produce object bounding box proposals. Lastly, an ROI head [19,20] is used to assign object labels and refine the bounding boxes on the pooled feature maps of the proposals. The entire network is trained utilizing two sets of losses: localization loss and objectness loss from RPN, and localization loss and classification loss from the ROI head. The fusion models explained below start with the base model that is trained on the remapped MSCOCO training set." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 456, + 545, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 456, + 545, + 599 + ], + "spans": [ + { + "bbox": [ + 304, + 456, + 545, + 599 + ], + "type": "text", + "content": "Early-fusion. The input to the Faster-RCNN network is the stacked image " + }, + { + "bbox": [ + 304, + 456, + 545, + 599 + ], + "type": "inline_equation", + "content": "[\\mathcal{I};\\mathcal{M}^{\\mathcal{O}}]" + }, + { + "bbox": [ + 304, + 456, + 545, + 599 + ], + "type": "text", + "content": ", with the two modalities stacked along the color channels, as shown in Fig. 4b. To enable the network to take a 6-channel image as input instead of the vanilla 3-channel image, the first layer of the backbone CNN is duplicated and stacked. The convolution kernel weights are first copied, then halved. This ensures that the feature map after the first layer conv1|conv2 is the same as in the vanilla model when " + }, + { + "bbox": [ + 304, + 456, + 545, + 599 + ], + "type": "inline_equation", + "content": "[\\mathcal{I};\\mathcal{I}]" + }, + { + "bbox": [ + 304, + 456, + 545, + 599 + ], + "type": "text", + "content": " is fed into it, providing a smooth start for adaptation training. This modification introduces only a small number of additional model parameters, and the increase in computational cost is negligible." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 600, + 545, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 600, + 545, + 694 + ], + "spans": [ + { + "bbox": [ + 304, + 600, + 545, + 694 + ], + "type": "text", + "content": "Mid-fusion. Both " + }, + { + "bbox": [ + 304, + 600, + 545, + 694 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 304, + 600, + 545, + 694 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 600, + 545, + 694 + ], + "type": "inline_equation", + "content": "\\mathcal{M}^O" + }, + { + "bbox": [ + 304, + 600, + 545, + 694 + ], + "type": "text", + "content": " are fed into the FPN in parallel, which yields two feature pyramid FP1 and FP2, as shown in Fig. 4c. Two sets of RPNs and ROI heads are used in parallel, and their initial weights are copied from the base model. FP1 is used in the left branch with RPN1 and ROI1, which is the same as the vanilla model. In the right branch with RPN2 and ROI2, the input feature pyramid is the fusion of FP1 and FP2:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 375, + 701, + 545, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 375, + 701, + 545, + 714 + ], + "type": "interline_equation", + "content": "\\mathrm {F P} _ {\\text {f u s i o n}} = \\mathrm {F P 1} \\oplus \\mathrm {F P 2}. \\tag {3}", + "image_path": "786334bdd4df0bc53a6711bef4d6b518b58ec1188f53af0eacbafdd37f382a7b.jpg" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21592" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 76, + 70, + 134, + 135 + ], + "blocks": [ + { + "bbox": [ + 76, + 70, + 134, + 135 + ], + "lines": [ + { + "bbox": [ + 76, + 70, + 134, + 135 + ], + "spans": [ + { + "bbox": [ + 76, + 70, + 134, + 135 + ], + "type": "image", + "image_path": "94794aeb6e2f4a0443ae56d799a03a8631adf755045aa8e2a605359b16d69a6d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 66, + 139, + 148, + 148 + ], + "lines": [ + { + "bbox": [ + 66, + 139, + 148, + 148 + ], + "spans": [ + { + "bbox": [ + 66, + 139, + 148, + 148 + ], + "type": "text", + "content": "(a) Vanilla Faster-RCNN." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 53, + 150, + 167, + 215 + ], + "blocks": [ + { + "bbox": [ + 53, + 150, + 167, + 215 + ], + "lines": [ + { + "bbox": [ + 53, + 150, + 167, + 215 + ], + "spans": [ + { + "bbox": [ + 53, + 150, + 167, + 215 + ], + "type": "image", + "image_path": "c3e7a84c7d82014934e52d0acf8860030d9f47dac13140a529bfa5275090b02f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 59, + 219, + 156, + 228 + ], + "lines": [ + { + "bbox": [ + 59, + 219, + 156, + 228 + ], + "spans": [ + { + "bbox": [ + 59, + 219, + 156, + 228 + ], + "type": "text", + "content": "(c) Mid-fusion Faster-RCNN." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 194, + 70, + 267, + 135 + ], + "blocks": [ + { + "bbox": [ + 194, + 70, + 267, + 135 + ], + "lines": [ + { + "bbox": [ + 194, + 70, + 267, + 135 + ], + "spans": [ + { + "bbox": [ + 194, + 70, + 267, + 135 + ], + "type": "image", + "image_path": "c6b44f02073ad3f51063f656a09dbf4347ff86f0f440124b09146577ffd79648.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 173, + 150, + 287, + 215 + ], + "blocks": [ + { + "bbox": [ + 176, + 139, + 277, + 148 + ], + "lines": [ + { + "bbox": [ + 176, + 139, + 277, + 148 + ], + "spans": [ + { + "bbox": [ + 176, + 139, + 277, + 148 + ], + "type": "text", + "content": "(b) Early-fusion Faster-RCNN." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 173, + 150, + 287, + 215 + ], + "lines": [ + { + "bbox": [ + 173, + 150, + 287, + 215 + ], + "spans": [ + { + "bbox": [ + 173, + 150, + 287, + 215 + ], + "type": "image", + "image_path": "3c26c955f7096337991c93c9b2f70ab6e26b1cdb38ce2160a9b4e60f00f2a0b1.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 179, + 219, + 276, + 228 + ], + "lines": [ + { + "bbox": [ + 179, + 219, + 276, + 228 + ], + "spans": [ + { + "bbox": [ + 179, + 219, + 276, + 228 + ], + "type": "text", + "content": "(d) Late-fusion Faster-RCNN." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 273, + 287, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 273, + 287, + 331 + ], + "spans": [ + { + "bbox": [ + 46, + 273, + 287, + 331 + ], + "type": "text", + "content": "The fused feature pyramid has the same dimension as both FP1 and FP2. The fusion operator can be either non-parametric such as average pooling, or parametric such as convolution layers. The whole network is trained with losses from both branches" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 335, + 291, + 361 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 335, + 291, + 361 + ], + "spans": [ + { + "bbox": [ + 47, + 335, + 291, + 361 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {m i d} = \\left(1 - \\alpha_ {m i d}\\right) \\left(\\mathcal {L} _ {r p n} ^ {\\mathcal {I}} + \\mathcal {L} _ {r o i} ^ {\\mathcal {I}}\\right) + \\alpha_ {m i d} \\left(\\mathcal {L} _ {r p n} ^ {f u s i o n} + \\mathcal {L} _ {r o i} ^ {f u s i o n}\\right), \\tag {4}", + "image_path": "52b73bffbca130073dced3915a4588b1ba3f0f521c047f11c07c74990c7d7bed.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 362, + 231, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 362, + 231, + 373 + ], + "spans": [ + { + "bbox": [ + 47, + 362, + 231, + 373 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 362, + 231, + 373 + ], + "type": "inline_equation", + "content": "\\alpha_{mid}" + }, + { + "bbox": [ + 47, + 362, + 231, + 373 + ], + "type": "text", + "content": " is the weights of different losses." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 374, + 287, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 374, + 287, + 434 + ], + "spans": [ + { + "bbox": [ + 47, + 374, + 287, + 434 + ], + "type": "text", + "content": "Late-fusion is similar to mid-fusion as shown in Fig. 4d. The key difference is that only the ROI head is branched. The region proposals are only generated from FP1, and the fused feature pyramid is only used in ROI head in the right branch. As a result, the total loss is" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 61, + 437, + 287, + 453 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 437, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 61, + 437, + 287, + 453 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {l a t e} = \\mathcal {L} _ {r p n} ^ {\\mathcal {I}} + (1 - \\alpha_ {l a t e}) \\mathcal {L} _ {r o i} ^ {\\mathcal {I}} + \\alpha_ {l a t e} \\mathcal {L} _ {r o i} ^ {f u s i o n}. \\tag {5}", + "image_path": "9ba0f9b042bf393a9c8f0bb6a84ac753a36d2f76e67dd63f6907c6784fa6a93b.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 469, + 287, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 469, + 287, + 540 + ], + "spans": [ + { + "bbox": [ + 46, + 469, + 287, + 540 + ], + "type": "text", + "content": "Both mid-fusion and late-fusion introduce significant additional parameters and computation. " + }, + { + "bbox": [ + 46, + 469, + 287, + 540 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 46, + 469, + 287, + 540 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 469, + 287, + 540 + ], + "type": "inline_equation", + "content": "\\mathcal{M}^O" + }, + { + "bbox": [ + 46, + 469, + 287, + 540 + ], + "type": "text", + "content": " need to pass the backbone, requiring more computation even at inference time. The quantitative comparison of inference speed of different fusion models can be found in the supplementary material." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 541, + 287, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 541, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 46, + 541, + 287, + 612 + ], + "type": "text", + "content": "To train the fusion models in adaptation, we first fine-tune the base models on MSCOCO training set with the inpainted background models described in Sec. 3.4 until convergence. During evaluation, we use the last background image to obtain the object masks, as it is the closest to the evaluation images on the timeline." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 47, + 621, + 156, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 621, + 156, + 634 + ], + "spans": [ + { + "bbox": [ + 47, + 621, + 156, + 634 + ], + "type": "text", + "content": "4. Scenes100 Dataset" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 642, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 712 + ], + "type": "text", + "content": "For our proposed scene-adaptive object detection problem, there is no existing dataset with long enough videos of stationary backgrounds for the development and evaluation of self-supervised adaptation techniques. We therefore collected a new dataset called Scenes100, which will be described in this section." + } + ] + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 307, + 70, + 385, + 114 + ], + "blocks": [ + { + "bbox": [ + 307, + 70, + 385, + 114 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 385, + 114 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 385, + 114 + ], + "type": "image", + "image_path": "fead86c9d20fe3d1cddfaf756c62add0c97143aac5a8b44b51a98ab63949fc69.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 307, + 114, + 385, + 159 + ], + "blocks": [ + { + "bbox": [ + 307, + 114, + 385, + 159 + ], + "lines": [ + { + "bbox": [ + 307, + 114, + 385, + 159 + ], + "spans": [ + { + "bbox": [ + 307, + 114, + 385, + 159 + ], + "type": "image", + "image_path": "13cf271c9c66c7f42573759012b562a1f7a517e55b89d4f212c4f74c5e16ab16.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 308, + 159, + 385, + 204 + ], + "blocks": [ + { + "bbox": [ + 308, + 159, + 385, + 204 + ], + "lines": [ + { + "bbox": [ + 308, + 159, + 385, + 204 + ], + "spans": [ + { + "bbox": [ + 308, + 159, + 385, + 204 + ], + "type": "image", + "image_path": "fb1b35766b0f16ffc593fc7a7ece9de10cec27d49e4f0e12865bd4cdb8ea09ff.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 308, + 204, + 385, + 247 + ], + "blocks": [ + { + "bbox": [ + 46, + 232, + 287, + 266 + ], + "lines": [ + { + "bbox": [ + 46, + 232, + 287, + 266 + ], + "spans": [ + { + "bbox": [ + 46, + 232, + 287, + 266 + ], + "type": "text", + "content": "Figure 4. Illustration of different fusion models for Faster-RCNN. Network modules are indicated by green, and feature maps and outputs are indicated by yellow." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 308, + 204, + 385, + 247 + ], + "lines": [ + { + "bbox": [ + 308, + 204, + 385, + 247 + ], + "spans": [ + { + "bbox": [ + 308, + 204, + 385, + 247 + ], + "type": "image", + "image_path": "c7750e5360e574180f3330f991604afb14d8562ab07f68c02a6e9ff28cf52711.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 251, + 545, + 295 + ], + "lines": [ + { + "bbox": [ + 306, + 251, + 545, + 295 + ], + "spans": [ + { + "bbox": [ + 306, + 251, + 545, + 295 + ], + "type": "text", + "content": "Figure 5. Some annotated frames from the evaluation portions. The green mask shows the part of the frame that is excluded for annotation and evaluation. Objects are labeled by bounding boxes. Please see the supplementary material for detailed descriptions." + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 387, + 71, + 464, + 114 + ], + "blocks": [ + { + "bbox": [ + 387, + 71, + 464, + 114 + ], + "lines": [ + { + "bbox": [ + 387, + 71, + 464, + 114 + ], + "spans": [ + { + "bbox": [ + 387, + 71, + 464, + 114 + ], + "type": "image", + "image_path": "32ab7d85f076faa1a46c3d8d80d7aba865923f82d3dec575e7ff24e010374235.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 387, + 114, + 464, + 159 + ], + "blocks": [ + { + "bbox": [ + 387, + 114, + 464, + 159 + ], + "lines": [ + { + "bbox": [ + 387, + 114, + 464, + 159 + ], + "spans": [ + { + "bbox": [ + 387, + 114, + 464, + 159 + ], + "type": "image", + "image_path": "4013098c739e593643edbb8079a3dfc399eef4c9440df24b292264126343df3a.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 387, + 159, + 464, + 204 + ], + "blocks": [ + { + "bbox": [ + 387, + 159, + 464, + 204 + ], + "lines": [ + { + "bbox": [ + 387, + 159, + 464, + 204 + ], + "spans": [ + { + "bbox": [ + 387, + 159, + 464, + 204 + ], + "type": "image", + "image_path": "17fff89abfcb963752a9856b21da43afa732d96f21713b98397d582c9e9d9070.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 387, + 204, + 464, + 247 + ], + "blocks": [ + { + "bbox": [ + 387, + 204, + 464, + 247 + ], + "lines": [ + { + "bbox": [ + 387, + 204, + 464, + 247 + ], + "spans": [ + { + "bbox": [ + 387, + 204, + 464, + 247 + ], + "type": "image", + "image_path": "07c1f575910c86ef3307858986c1a1735eede54df05eb00091566aea3bad2ef0.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 466, + 71, + 544, + 114 + ], + "blocks": [ + { + "bbox": [ + 466, + 71, + 544, + 114 + ], + "lines": [ + { + "bbox": [ + 466, + 71, + 544, + 114 + ], + "spans": [ + { + "bbox": [ + 466, + 71, + 544, + 114 + ], + "type": "image", + "image_path": "aab208c4b08ac9b07b770d63acf81d18fdf26f78afd12259277ad642c0c82fc6.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 466, + 114, + 544, + 159 + ], + "blocks": [ + { + "bbox": [ + 466, + 114, + 544, + 159 + ], + "lines": [ + { + "bbox": [ + 466, + 114, + 544, + 159 + ], + "spans": [ + { + "bbox": [ + 466, + 114, + 544, + 159 + ], + "type": "image", + "image_path": "fd3fc721cbb5f65ec85d87cbd64fd49639a1b01edf4ee4a5860c3b19635499c6.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 466, + 159, + 544, + 204 + ], + "blocks": [ + { + "bbox": [ + 466, + 159, + 544, + 204 + ], + "lines": [ + { + "bbox": [ + 466, + 159, + 544, + 204 + ], + "spans": [ + { + "bbox": [ + 466, + 159, + 544, + 204 + ], + "type": "image", + "image_path": "5c428924b8bca60fe8bfee2c14e7e1dfc0bd0ad3901bb838146ca04f011277f5.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 466, + 204, + 544, + 247 + ], + "blocks": [ + { + "bbox": [ + 466, + 204, + 544, + 247 + ], + "lines": [ + { + "bbox": [ + 466, + 204, + 544, + 247 + ], + "spans": [ + { + "bbox": [ + 466, + 204, + 544, + 247 + ], + "type": "image", + "image_path": "21197db50c1e6ad1809beaa6ddee2ad0d2e1d371d1474781c5374f296b0e9609.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "bbox": [ + 306, + 301, + 401, + 313 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 301, + 401, + 313 + ], + "spans": [ + { + "bbox": [ + 306, + 301, + 401, + 313 + ], + "type": "text", + "content": "4.1. Data Collection" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 304, + 321, + 545, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 321, + 545, + 513 + ], + "spans": [ + { + "bbox": [ + 304, + 321, + 545, + 513 + ], + "type": "text", + "content": "We used keywords such as \"live view\", \"street camera\", and \"live webcam\" to search for live streams on YouTube. We look for streams that are of decent quality and have fixed camera perspective. Around 200 candidate videos were recorded between September 2020 and February 2022. We picked 100 videos to compile our Scenes100 dataset, ensuring that each video is longer than two hours and has minimum width and height of 720 pixels. We also aimed for a dataset with high diversities. The 100 videos of Scenes100 were recorded in a variety of 16 countries and territories, and during different time periods of day and year. The dataset contain videos from both indoor and outdoor scenes, with different camera perspectives. The density and scale of visual objects also vary significantly from one video to another. Some sample frames is presented in Fig. 5 to show the diversity of our dataset." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 306, + 525, + 421, + 536 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 525, + 421, + 536 + ], + "spans": [ + { + "bbox": [ + 306, + 525, + 421, + 536 + ], + "type": "text", + "content": "4.2. Manual Annotation" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 304, + 544, + 545, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 544, + 545, + 639 + ], + "spans": [ + { + "bbox": [ + 304, + 544, + 545, + 639 + ], + "type": "text", + "content": "We are interested in detecting objects from two categories: person and vehicle. The person category includes any visible and recognizable people in the frames; this corresponds to the person category in MSCOCO [41] dataset. The vehicle category includes all vehicles with four or more wheels, so motorbikes, bicycles, tricycles are not included. This category corresponds to the car, bus, and truck categories in MSCOCO." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": "For each video, we used the first 1.5 hours for self-supervised and weakly-supervised training, and the remaining part for evaluation. Since many videos were taken with wide-angle lenses, some frames included faraway objects that were too small to be considered in evaluation. To address this, we manually drew polygon masks to exclude" + } + ] + } + ], + "index": 35 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21593" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 70, + 286, + 153 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 286, + 153 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 286, + 153 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 286, + 153 + ], + "type": "table", + "html": "
datasetcontain videosaverage lengthframescountriesboundong boxesboxes per video
MSCOCO [41]No---897K-
KITTI [17]No--180K-
BDD100K [74]Yes40s120M11.8M18
CityScapes [9]Yes1.8s150K265K13
Scenes100Yes2h21.6M1684K840
", + "image_path": "ab6bea824d15f1a8e4a1c1a9b1c3a631b3c496ecaf39d5795de1ac2768a189d5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 156, + 287, + 178 + ], + "lines": [ + { + "bbox": [ + 47, + 156, + 287, + 178 + ], + "spans": [ + { + "bbox": [ + 47, + 156, + 287, + 178 + ], + "type": "text", + "content": "Table 1. Comparison of Scenes100 with other object detection datasets " + }, + { + "bbox": [ + 47, + 156, + 287, + 178 + ], + "type": "inline_equation", + "content": "(\\mathbf{K} = 10^{3},\\mathbf{M} = 10^{6})" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 185, + 288, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 185, + 288, + 330 + ], + "spans": [ + { + "bbox": [ + 46, + 185, + 288, + 330 + ], + "type": "text", + "content": "these faraway parts from annotation and evaluation. We then estimated the number of visual objects in the unmasked parts by first using a pre-trained detector, and then refining the estimate through human annotation. Frames for annotation were uniformly sampled from the second portion of each video, with the number of frames being inversely proportional to the estimated number of objects in the frames. This ensured that each video had roughly the same number of annotated objects. We contracted a data annotation company to perform bounding box annotation on those sampled frames. Some annotated frame samples for evaluation is shown in Fig. 5." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 336, + 150, + 347 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 336, + 150, + 347 + ], + "spans": [ + { + "bbox": [ + 47, + 336, + 150, + 347 + ], + "type": "text", + "content": "4.3. Dataset Statistics" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 354, + 287, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 354, + 287, + 474 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 287, + 474 + ], + "type": "text", + "content": "In Tab. 1, we compare Scenes100 with some popular object detection datasets. MSCOCO [41] is a fully-supervised general-purpose object detection dataset, which acts as the supervised source domain. KITTI [17] only contains object detection annotation associated to individual images. In BDD100K [74], object detection annotation is given for one key frame in each video. For CityScapes [9], we consider the finely labeled instance-level segmentation masks, which are given for the " + }, + { + "bbox": [ + 46, + 354, + 287, + 474 + ], + "type": "inline_equation", + "content": "20^{\\mathrm{th}}" + }, + { + "bbox": [ + 46, + 354, + 287, + 474 + ], + "type": "text", + "content": " image from a 30 frame video snippets lasting 1.8s." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 474, + 287, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 287, + 629 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 287, + 629 + ], + "type": "text", + "content": "The scale of Scenes100 is comparable to that of KITTI and CityScapes. However, Scenes100 has three unique features. First, the videos in Scenes100 are much longer than those in any other datasets with scenes, providing sufficient data for a model to adapt to each of them. Second, Scenes100 has greater diversity. Unlike some other datasets that are recorded in only a few cities with similar camera angles, background environments, and object types, Scenes100 covers several different countries, with varying weather conditions, road conditions, view of field, camera perspectives, and vehicle types. Third, each video in Scenes100 has a fixed camera perspective and stationary background scene, making background modeling possible." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "text", + "content": "Note that many currently available datasets are not primarily intended for self-supervised adaptive object detection tasks. They may possess other features, such as segmentation masks, multi-modal sensor data, depth maps, or annotations for object tracking. However, despite the absence of these attributes, Scenes100 holds value as the first extensive and varied dataset that features fixed camera per" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 109 + ], + "type": "text", + "content": "spectives and lengthy videos. As such, it can complement existing dataset collections and serve as a valuable resource for researchers exploring scene adaptive object detection." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 118, + 388, + 132 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 118, + 388, + 132 + ], + "spans": [ + { + "bbox": [ + 306, + 118, + 388, + 132 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 138, + 426, + 150 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 138, + 426, + 150 + ], + "spans": [ + { + "bbox": [ + 306, + 138, + 426, + 150 + ], + "type": "text", + "content": "5.1. Evaluation Protocols" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 156, + 545, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 156, + 545, + 215 + ], + "spans": [ + { + "bbox": [ + 304, + 156, + 545, + 215 + ], + "type": "text", + "content": "We evaluate the performance of object detectors following the COCO evaluation protocol [41] to calculate the average precision at " + }, + { + "bbox": [ + 304, + 156, + 545, + 215 + ], + "type": "inline_equation", + "content": "IoU = 50\\%" + }, + { + "bbox": [ + 304, + 156, + 545, + 215 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 304, + 156, + 545, + 215 + ], + "type": "inline_equation", + "content": "AP^{50}" + }, + { + "bbox": [ + 304, + 156, + 545, + 215 + ], + "type": "text", + "content": ") and mean average precision of different " + }, + { + "bbox": [ + 304, + 156, + 545, + 215 + ], + "type": "inline_equation", + "content": "IoU" + }, + { + "bbox": [ + 304, + 156, + 545, + 215 + ], + "type": "text", + "content": " thresholds from 0.5 to 0.95 (" + }, + { + "bbox": [ + 304, + 156, + 545, + 215 + ], + "type": "inline_equation", + "content": "AP^m" + }, + { + "bbox": [ + 304, + 156, + 545, + 215 + ], + "type": "text", + "content": "), with some modifications as follows." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 217, + 545, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 217, + 545, + 299 + ], + "spans": [ + { + "bbox": [ + 304, + 217, + 545, + 299 + ], + "type": "text", + "content": "Non-evaluation region: before feeding the ground-truth and detected bounding boxes to the COCO evaluator, the bounding boxes that have at least one corner inside the nonevaluation mask will be removed. So the faraway parts in the frames where objects are deemed too small and blurry will not affect the evaluation results. Please see Sec. 4.2 and Fig. 5 for more details." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 300, + 546, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 300, + 546, + 478 + ], + "spans": [ + { + "bbox": [ + 304, + 300, + 546, + 478 + ], + "type": "text", + "content": "Category weighting: in the standard COCO evaluation protocol, the overall multi-class " + }, + { + "bbox": [ + 304, + 300, + 546, + 478 + ], + "type": "inline_equation", + "content": "AP" + }, + { + "bbox": [ + 304, + 300, + 546, + 478 + ], + "type": "text", + "content": " is the simple average of the " + }, + { + "bbox": [ + 304, + 300, + 546, + 478 + ], + "type": "inline_equation", + "content": "AP_{s}" + }, + { + "bbox": [ + 304, + 300, + 546, + 478 + ], + "type": "text", + "content": " of the different categories. If a class is sparse in the training set and validation set, the model tends to perform poorly due to the scarcity of the training samples. In this case the average " + }, + { + "bbox": [ + 304, + 300, + 546, + 478 + ], + "type": "inline_equation", + "content": "AP" + }, + { + "bbox": [ + 304, + 300, + 546, + 478 + ], + "type": "text", + "content": " does not truly reflect the performance of a detector. Hence we propose an additional multi-class " + }, + { + "bbox": [ + 304, + 300, + 546, + 478 + ], + "type": "inline_equation", + "content": "AP" + }, + { + "bbox": [ + 304, + 300, + 546, + 478 + ], + "type": "text", + "content": " metric with weighted average where the weight is determined by the prevalence of ground truth object instances in the evaluation set. This weighted " + }, + { + "bbox": [ + 304, + 300, + 546, + 478 + ], + "type": "inline_equation", + "content": "AP" + }, + { + "bbox": [ + 304, + 300, + 546, + 478 + ], + "type": "text", + "content": " can better portray the performance of a detector in the case of long-tailed distribution. We refer the standard classes-mean COCO " + }, + { + "bbox": [ + 304, + 300, + 546, + 478 + ], + "type": "inline_equation", + "content": "AP" + }, + { + "bbox": [ + 304, + 300, + 546, + 478 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 304, + 300, + 546, + 478 + ], + "type": "inline_equation", + "content": "AP_{co}" + }, + { + "bbox": [ + 304, + 300, + 546, + 478 + ], + "type": "text", + "content": ", and the proposed weighted " + }, + { + "bbox": [ + 304, + 300, + 546, + 478 + ], + "type": "inline_equation", + "content": "AP" + }, + { + "bbox": [ + 304, + 300, + 546, + 478 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 304, + 300, + 546, + 478 + ], + "type": "inline_equation", + "content": "AP_{w}" + }, + { + "bbox": [ + 304, + 300, + 546, + 478 + ], + "type": "text", + "content": ". We refer the " + }, + { + "bbox": [ + 304, + 300, + 546, + 478 + ], + "type": "inline_equation", + "content": "AP" + }, + { + "bbox": [ + 304, + 300, + 546, + 478 + ], + "type": "text", + "content": " averaged over IoU thresholds and IoU = 50% as " + }, + { + "bbox": [ + 304, + 300, + 546, + 478 + ], + "type": "inline_equation", + "content": "AP^{m}" + }, + { + "bbox": [ + 304, + 300, + 546, + 478 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 300, + 546, + 478 + ], + "type": "inline_equation", + "content": "AP^{50}" + }, + { + "bbox": [ + 304, + 300, + 546, + 478 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 479, + 545, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 479, + 545, + 587 + ], + "spans": [ + { + "bbox": [ + 304, + 479, + 545, + 587 + ], + "type": "text", + "content": "Per-video evaluation: we aim to evaluate the performance of a scene adaption method, instead of a specific detection model. For each of the video, a model is adapted to it from the same base model trained on the source dataset. The performance of the base model and adapted model on each video is calculated individually as " + }, + { + "bbox": [ + 304, + 479, + 545, + 587 + ], + "type": "inline_equation", + "content": "\\{(AP_{v,\\text{base}}, AP_{v,\\text{adapt}}) | v = 1 \\ldots 100\\}" + }, + { + "bbox": [ + 304, + 479, + 545, + 587 + ], + "type": "text", + "content": ". Then we evaluate the overall effectiveness of the adaptation method using the averaged " + }, + { + "bbox": [ + 304, + 479, + 545, + 587 + ], + "type": "inline_equation", + "content": "AP" + }, + { + "bbox": [ + 304, + 479, + 545, + 587 + ], + "type": "text", + "content": " gain as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 339, + 595, + 545, + 628 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 595, + 545, + 628 + ], + "spans": [ + { + "bbox": [ + 339, + 595, + 545, + 628 + ], + "type": "interline_equation", + "content": "A P G = \\frac {1}{1 0 0} \\sum_ {v = 1} ^ {1 0 0} \\left(A P _ {v, a d a p t} - A P _ {v, b a s e}\\right). \\tag {6}", + "image_path": "036dd8cc8b5d52a883f32865aabdfe3fff7eeef7336c448efc7db16c7e58554b.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 635, + 440, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 635, + 440, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 635, + 440, + 647 + ], + "type": "text", + "content": "5.2. Implementation Details" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": "Our implementation is based on Detector2 [70]. The DiMP tracker is directly taken from the official PyTracking [10] implementation without any change. More details and hyper-parameters can be found in the supplementary material. During training, we include the same number of" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21594" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": "images from MSCOCO training set as the number of unlabeled frames from the videos. The hyper-parameters are the same for all videos. We check the training loss for all experiments, including the baseline methods, to ensure convergence. We also train \"compound\" models where all videos are used together for adaptation. The results and further details can be found in the supplementary material." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 168, + 209, + 182 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 168, + 209, + 182 + ], + "spans": [ + { + "bbox": [ + 47, + 168, + 209, + 182 + ], + "type": "text", + "content": "5.3. Domain Adaptation Baselines" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 189, + 287, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 189, + 287, + 368 + ], + "spans": [ + { + "bbox": [ + 46, + 189, + 287, + 368 + ], + "type": "text", + "content": "We compare our methods with several domain adaptive object detection baselines. In their original papers, the methods are evaluated on other domain adaptive object detection datasets. However, those datasets do not contains videos with fixed camera perspectives, so the proposed location-aware mixup and dynamic background extraction cannot be applied. We instead apply the methods in other papers on Scenes100. For a fair comparison, all methods start with the same base model and weights, and then are adapted to and evaluated on each scene individually. We keep the values of the hyper-parameters that are specific to a baseline method used in the original papers. We use the same learning rate, batch size, and number of iterations as our experiments in those baseline methods. More details can be found in the supplementary material." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 369, + 287, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 369, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 46, + 369, + 287, + 430 + ], + "type": "text", + "content": "Self-Train (ST) [57] is a method that uses detection and tracking to obtain pseudo bounding boxes, similar to ours. But ST applies tracking only in one direction. It assigns different weights to the pseudo bounding boxes during training, but no refinement nor cross-teaching." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 431, + 287, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 431, + 287, + 467 + ], + "spans": [ + { + "bbox": [ + 46, + 431, + 287, + 467 + ], + "type": "text", + "content": "STAC [60] uses the base model to get pseudo bounding boxes, and trains the model on a strongly augmented version of the target domain image with those boxes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 469, + 287, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 469, + 287, + 528 + ], + "spans": [ + { + "bbox": [ + 46, + 469, + 287, + 528 + ], + "type": "text", + "content": "Adaptive Teacher (AT) [37] also uses self-training, but the pseudo bounding boxes are detected on the fly. It utilizes exponential moving average to update the model gradually. It also uses weak/strong data augmentation, and a domain classifier to align the features of the two domains." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 529, + 287, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 529, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 46, + 529, + 287, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{H}^2\\mathbf{FA}" + }, + { + "bbox": [ + 46, + 529, + 287, + 590 + ], + "type": "text", + "content": " R-CNN [72] applies feature alignment of source and target domains at various levels of Faster-RCNN. It assumes that image-level weak annotation is available for the target domain. So we use the labels from our pseudo-labeling process as image-level class labels." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 592, + 287, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 592, + 287, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 592, + 287, + 640 + ], + "type": "text", + "content": "TIA [78] is a method that instantiates feature alignment by using both domain classifier and auxiliary classification and localization heads. The model is trained in a mini-max manner using gradient reverse layers [15]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 642, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 712 + ], + "type": "text", + "content": "LODS [33] uses style enhancement as data augmentation. The pseudo bounding boxes are generated from the original images, and the features from both versions of the same image are aligned. It also utilizes exponential moving average for the weights. It does not use any supervised training data from the source domain." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 308, + 71, + 544, + 131 + ], + "blocks": [ + { + "bbox": [ + 308, + 71, + 544, + 131 + ], + "lines": [ + { + "bbox": [ + 308, + 71, + 544, + 131 + ], + "spans": [ + { + "bbox": [ + 308, + 71, + 544, + 131 + ], + "type": "table", + "html": "
Model# paramsMSCOCOScenes100
APmcoAP50coAPmcoAP50coAPmcoAP50w
M1 (R-50)41.4M50.0576.4840.5262.1241.2864.65
M2 (R-101)60.5M51.2977.4641.1163.1041.9665.74
", + "image_path": "c445f87920c29a6edffe5d333512502ed2f46e5fff917f5349ef87c38e777300.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 305, + 133, + 545, + 188 + ], + "lines": [ + { + "bbox": [ + 305, + 133, + 545, + 188 + ], + "spans": [ + { + "bbox": [ + 305, + 133, + 545, + 188 + ], + "type": "text", + "content": "Table 2. Performance of base models on MSCOCO validation set with remapped categories and Scenes100 before any adaptation. For Scenes100, " + }, + { + "bbox": [ + 305, + 133, + 545, + 188 + ], + "type": "inline_equation", + "content": "APs" + }, + { + "bbox": [ + 305, + 133, + 545, + 188 + ], + "type": "text", + "content": " are evaluated on each video individually and then averaged. R- stands for ResNet. # indicates the number of parameters " + }, + { + "bbox": [ + 305, + 133, + 545, + 188 + ], + "type": "inline_equation", + "content": "(M = 10^6)" + }, + { + "bbox": [ + 305, + 133, + 545, + 188 + ], + "type": "text", + "content": ". See Sec. 5.1 for " + }, + { + "bbox": [ + 305, + 133, + 545, + 188 + ], + "type": "inline_equation", + "content": "AP" + }, + { + "bbox": [ + 305, + 133, + 545, + 188 + ], + "type": "text", + "content": " notation details." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 194, + 362, + 206 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 194, + 362, + 206 + ], + "spans": [ + { + "bbox": [ + 306, + 194, + 362, + 206 + ], + "type": "text", + "content": "5.4. Results" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 216, + 545, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 216, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 304, + 216, + 545, + 335 + ], + "type": "text", + "content": "The performance of the base models is shown in Tab. 2, along with their number of trainable parameters. M2 is a bigger model with more parameters and representation capacity, so it outperforms M1 by a noticeable margin in all the metrics before adaptation. All the " + }, + { + "bbox": [ + 304, + 216, + 545, + 335 + ], + "type": "inline_equation", + "content": "AP" + }, + { + "bbox": [ + 304, + 216, + 545, + 335 + ], + "type": "text", + "content": " gains shown in other tables are based on the performance of base model M2. Note that the " + }, + { + "bbox": [ + 304, + 216, + 545, + 335 + ], + "type": "inline_equation", + "content": "AP" + }, + { + "bbox": [ + 304, + 216, + 545, + 335 + ], + "type": "text", + "content": " numbers on Scenes100 is already relatively high at only about 10 points below the performance on MSCOCO, so we do not observe very high (" + }, + { + "bbox": [ + 304, + 216, + 545, + 335 + ], + "type": "inline_equation", + "content": ">5" + }, + { + "bbox": [ + 304, + 216, + 545, + 335 + ], + "type": "text", + "content": " points) " + }, + { + "bbox": [ + 304, + 216, + 545, + 335 + ], + "type": "inline_equation", + "content": "AP" + }, + { + "bbox": [ + 304, + 216, + 545, + 335 + ], + "type": "text", + "content": " gains in the results shown later." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 339, + 545, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 339, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 304, + 339, + 545, + 506 + ], + "type": "text", + "content": "We compare the performance of different adaptation methods in terms of average " + }, + { + "bbox": [ + 304, + 339, + 545, + 506 + ], + "type": "inline_equation", + "content": "AP" + }, + { + "bbox": [ + 304, + 339, + 545, + 506 + ], + "type": "text", + "content": " gain in Tab. 3. Here we show our best combination of methods, which incorporates pseudo-labeling, location-aware mixup, and object mask mid-fusion. Among the baseline methods for domain adaptive object detection, only ST and LODS improves the performance significantly. AT only yields marginal improvement in classes-weighted " + }, + { + "bbox": [ + 304, + 339, + 545, + 506 + ], + "type": "inline_equation", + "content": "AP" + }, + { + "bbox": [ + 304, + 339, + 545, + 506 + ], + "type": "text", + "content": ". STAC, " + }, + { + "bbox": [ + 304, + 339, + 545, + 506 + ], + "type": "inline_equation", + "content": "\\mathrm{H}^2\\mathrm{FA}" + }, + { + "bbox": [ + 304, + 339, + 545, + 506 + ], + "type": "text", + "content": ", and TIA actually degrade the performance on Scenes100. LODS performs surprisingly well considering it does not use source domain images. However, we found that any training schedule longer than 1000 iterations would lead to severe performance degradation. The proposed method yields much higher, consistent, and stable improvement." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": "In general, self-training-based methods (ST, AT, STAC, and the proposed method) perform better than domain alignment-based methods " + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{(H^2FA}" + }, + { + "bbox": [ + 304, + 510, + 545, + 713 + ], + "type": "text", + "content": " and TIA). Domain alignment is usually achieved through adversarial learning, which is known to have unstable objectives and is more difficult to optimize [14, 42, 59, 61]. Therefore, these methods are more sensitive to hyperparameters and training schedules, making it challenging to determine the optimal settings for all datasets. Applying domain alignment methods directly to our scene adaptive dataset without careful tuning can cause problems as our dataset has less in-domain variance than domain adaptation datasets. Furthermore, since the scenes in our dataset are very different, they might require different settings. This implies that Scenes100 has its uniqueness compared to more generally purported domain adaptive object detection datasets. Our proposed method is more robust to this setting and does not rely heavily on" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21595" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 70, + 70, + 264, + 184 + ], + "blocks": [ + { + "bbox": [ + 70, + 70, + 264, + 184 + ], + "lines": [ + { + "bbox": [ + 70, + 70, + 264, + 184 + ], + "spans": [ + { + "bbox": [ + 70, + 70, + 264, + 184 + ], + "type": "table", + "html": "
Method\\(APG_{co}^{m}\\)\\(APG_{co}^{50}\\)\\(APG_{w}^{m}\\)\\(APG_{w}^{50}\\)
ST [57]+0.80+0.24+1.39+1.03
STAC [60]-1.26-5.12-1.97-6.64
AT [37]-0.75-1.11+0.06+0.04
\\(\\mathrm{H^{2}FA}\\) [72]-3.10-4.97-3.77-6.01
TIA [78]-0.32-0.37-0.32-0.33
LODS [33]+0.45+1.28+1.02+2.28
Proposed+3.76+4.45+3.78+4.65
", + "image_path": "c842656924f859237c74ca029d3a4fe34d88e85333aa90c1767bfc0ee76f263a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 221, + 285, + 365 + ], + "blocks": [ + { + "bbox": [ + 47, + 186, + 287, + 209 + ], + "lines": [ + { + "bbox": [ + 47, + 186, + 287, + 209 + ], + "spans": [ + { + "bbox": [ + 47, + 186, + 287, + 209 + ], + "type": "text", + "content": "Table 3. Averaged " + }, + { + "bbox": [ + 47, + 186, + 287, + 209 + ], + "type": "inline_equation", + "content": "AP" + }, + { + "bbox": [ + 47, + 186, + 287, + 209 + ], + "type": "text", + "content": " gain of different adaptation methods. See Sec. 5.1 for " + }, + { + "bbox": [ + 47, + 186, + 287, + 209 + ], + "type": "inline_equation", + "content": "AP" + }, + { + "bbox": [ + 47, + 186, + 287, + 209 + ], + "type": "text", + "content": " notation details." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 221, + 285, + 365 + ], + "lines": [ + { + "bbox": [ + 50, + 221, + 285, + 365 + ], + "spans": [ + { + "bbox": [ + 50, + 221, + 285, + 365 + ], + "type": "table", + "html": "
MixupFusion\\(APG_{co}^{m}\\)\\(APG_{co}^{50}\\)\\(APG_{w}^{m}\\)\\(APG_{w}^{50}\\)
XX+0.95+0.54+1.67+1.55
Location-awareX+1.72+1.67+2.25+2.53
RandomX+1.22+1.13+1.76+2.03
Xearly+1.85+2.12+2.22+2.73
Xmid+3.40+3.81+3.67+3.98
Xlate+3.34+3.60+3.38+3.72
Location-awareearly+2.25+2.82+2.59+3.46
Location-awaremid+3.76+4.45+3.78+4.65
Location-awarelate+3.66+4.10+3.73+4.31
", + "image_path": "5f603bdc41fe198eae3473fbb0711ffb455110ba4d0fa7d514a40303b09c33c6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 50, + 425, + 285, + 524 + ], + "blocks": [ + { + "bbox": [ + 47, + 368, + 287, + 412 + ], + "lines": [ + { + "bbox": [ + 47, + 368, + 287, + 412 + ], + "spans": [ + { + "bbox": [ + 47, + 368, + 287, + 412 + ], + "type": "text", + "content": "Table 4. Ablation study for the proposed components of our scene-adaptive object detection method. " + }, + { + "bbox": [ + 47, + 368, + 287, + 412 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 47, + 368, + 287, + 412 + ], + "type": "text", + "content": " means not being applied. Pseudo-labeling is always applied. This table shows the averaged " + }, + { + "bbox": [ + 47, + 368, + 287, + 412 + ], + "type": "inline_equation", + "content": "AP" + }, + { + "bbox": [ + 47, + 368, + 287, + 412 + ], + "type": "text", + "content": " gain; see Sec. 5.1 for " + }, + { + "bbox": [ + 47, + 368, + 287, + 412 + ], + "type": "inline_equation", + "content": "AP" + }, + { + "bbox": [ + 47, + 368, + 287, + 412 + ], + "type": "text", + "content": " notation details." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 425, + 285, + 524 + ], + "lines": [ + { + "bbox": [ + 50, + 425, + 285, + 524 + ], + "spans": [ + { + "bbox": [ + 50, + 425, + 285, + 524 + ], + "type": "table", + "html": "
MixupFusion\\(APG_{co}^{m}\\)\\(APG_{co}^{50}\\)\\(APG_{w}^{m}\\)\\(APG_{w}^{50}\\)
✘AVG+3.40+3.81+3.67+3.98
✘CNN+3.25+3.52+3.26+3.79
✘ATTN+3.22+3.73+3.23+3.95
Location-awareAVG+3.76+4.45+3.78+4.65
Location-awareCNN+3.60+4.20+3.71+4.50
Location-awareATTN+3.30+3.89+3.45+4.26
", + "image_path": "4f97855a54a27a12202bc95bfb60a1c4b9023b050798dd977102582cdfca4249.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 526, + 287, + 583 + ], + "lines": [ + { + "bbox": [ + 46, + 526, + 287, + 583 + ], + "spans": [ + { + "bbox": [ + 46, + 526, + 287, + 583 + ], + "type": "text", + "content": "Table 5. Ablation study on different feature pyramid fusion methods in term of averaged " + }, + { + "bbox": [ + 46, + 526, + 287, + 583 + ], + "type": "inline_equation", + "content": "AP" + }, + { + "bbox": [ + 46, + 526, + 287, + 583 + ], + "type": "text", + "content": " gain. AVG, CNN, and ATTN indicate mid-fusion based on average pooling, CNN, or attention module, respectively. " + }, + { + "bbox": [ + 46, + 526, + 287, + 583 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 46, + 526, + 287, + 583 + ], + "type": "text", + "content": " means not being applied. Pseudo-labeling is always applied. See Sec. 5.1 for " + }, + { + "bbox": [ + 46, + 526, + 287, + 583 + ], + "type": "inline_equation", + "content": "AP" + }, + { + "bbox": [ + 46, + 526, + 287, + 583 + ], + "type": "text", + "content": " notation details." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 589, + 287, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 589, + 287, + 636 + ], + "spans": [ + { + "bbox": [ + 46, + 589, + 287, + 636 + ], + "type": "text", + "content": "hyperparameter tuning, as we use the same set of hyperparameters across all 100 videos. Additional analysis and success/failure cases can be found in the supplementary material." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 646, + 141, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 646, + 141, + 659 + ], + "spans": [ + { + "bbox": [ + 47, + 646, + 141, + 659 + ], + "type": "text", + "content": "5.5. Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "type": "text", + "content": "We explore the effect of the different components in our methods by experimenting various combinations of them. For all experiments, we keep the pseudo-labeling unchanged, as it is the basis of other components. We compare" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": "different types of object mask fusion (Fusion): early, mid, and late. We also compare the proposed Location-aware object mixup with Random mixup. Tab. 4 shows the results. The experiments on hyper-parameters can be found in the supplementary material." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 136, + 546, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 136, + 546, + 293 + ], + "spans": [ + { + "bbox": [ + 304, + 136, + 546, + 293 + ], + "type": "text", + "content": "The proposed pseudo-labeling, mixup, and object mask fusion all benefit the adaptation performance. Applying only the pseudo-labeling already leads to " + }, + { + "bbox": [ + 304, + 136, + 546, + 293 + ], + "type": "inline_equation", + "content": "AP" + }, + { + "bbox": [ + 304, + 136, + 546, + 293 + ], + "type": "text", + "content": " gain higher than ST, which is more complicated involving pseudo box weighting. Location-aware mixup is also more advantageous than Random mixup, validating the assumption in Sec. 3.3 that having fewer artifacts is better for adaptation. Among three fusion options, mid-fusion yields the best result. That is perhaps because in a mid-fusion model, the RPN utilizes the fused feature pyramid, which can contain critical information of object boundaries. The combination of mixup and object mask fusion can further improve the performance." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 297, + 545, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 297, + 545, + 417 + ], + "spans": [ + { + "bbox": [ + 304, + 297, + 545, + 417 + ], + "type": "text", + "content": "We also compare different feature pyramid fusion methods described in Eq. (3). In our main experiments, we simply average the feature pyramids from the original image and the object mask. Now at each level of the pyramid, we use a separate CNN or an attention module to fuse the feature maps. The comparison is shown in Tab. 5. The parametric method performs slightly worse than average-pooling, probably because the newly introduced modules have randomly initialized weights and thus require additional supervised training data." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 441, + 372, + 454 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 441, + 372, + 454 + ], + "spans": [ + { + "bbox": [ + 306, + 441, + 372, + 454 + ], + "type": "text", + "content": "6. Summary" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 465, + 545, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 465, + 545, + 681 + ], + "spans": [ + { + "bbox": [ + 304, + 465, + 545, + 681 + ], + "type": "text", + "content": "We have presented a self-supervised framework for scene-adaptive object detection. Base detectors and generic trackers are used to generate pseudo object labels as the adaptive training targets in a cross-teaching manner. To fully utilize the background equivariance when camera perspective is fixed, we propose artifacts-free location-aware object mixup to augment the input images, and dynamic background extraction for additional input modality to the detector. We also introduce the first large-scale scene adaptive object detection dataset, Scenes100, with several unique features compared to other domain adaptive object detection datasets. Our method outperforms domain adaptive object detection baselines by a large margin on Scenes100. We also conduct extensive experiments to illustrate the effectiveness of the components in our framework. We hope this work has demonstrated the importance of scene adaptation, and it will spur further research interest in this impactful but understudied area." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "content": "Acknowledgements. This research was partially supported by NSF grants IIS-1763981 & NSDF DUE-2055406." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21596" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 92, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 92, + 287, + 146 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 92, + 287, + 146 + ], + "spans": [ + { + "bbox": [ + 53, + 92, + 287, + 146 + ], + "type": "text", + "content": "[1] David Berthelot, Nicholas Carlini, Ian Goodfellow, Avital Oliver, Nicolas Papernot, and Colin Raffel. Mixmatch: A holistic approach to semi-supervised learning. In Conference on Neural Information Processing Systems (NeurIPS), 2019. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 149, + 288, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 149, + 288, + 194 + ], + "spans": [ + { + "bbox": [ + 53, + 149, + 288, + 194 + ], + "type": "text", + "content": "[2] Goutam Bhat, Martin Danelljan, Luc Van Gool, and Radu Timofte. Learning discriminative model prediction for tracking. In International Conference on Computer Vision (ICCV), 2019. 1, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 196, + 288, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 196, + 288, + 228 + ], + "spans": [ + { + "bbox": [ + 53, + 196, + 288, + 228 + ], + "type": "text", + "content": "[3] Alexey Bochkovskiy, Chien-Yao Wang, and Hong-Yuan Mark Liao. Yolov4: Optimal speed and accuracy of object detection. In ArXiv, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 232, + 288, + 254 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 232, + 288, + 254 + ], + "spans": [ + { + "bbox": [ + 53, + 232, + 288, + 254 + ], + "type": "text", + "content": "[4] Shahine Bouabid and Vincent Delaitre. Mixup regularization for region proposal based object detectors. In ArXiv, 2020. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 255, + 288, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 255, + 288, + 289 + ], + "spans": [ + { + "bbox": [ + 53, + 255, + 288, + 289 + ], + "type": "text", + "content": "[5] Zhaowei Cai and Nuno Vasconcelos. Cascade r-cnn: Delving into high quality object detection. In Computer Vision and Pattern Recognition Conference (CVPR), 2018. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 291, + 288, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 291, + 288, + 335 + ], + "spans": [ + { + "bbox": [ + 53, + 291, + 288, + 335 + ], + "type": "text", + "content": "[6] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European Conference on Computer Vision (ECCV), 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 337, + 288, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 337, + 288, + 381 + ], + "spans": [ + { + "bbox": [ + 53, + 337, + 288, + 381 + ], + "type": "text", + "content": "[7] Yuhua Chen, Wen Li, Christos Sakaridis, Dengxin Dai, and Luc Van Gool. Domain adaptive faster r-cnn for object detection in the wild. In Computer Vision and Pattern Recognition Conference (CVPR), 2018. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 384, + 288, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 384, + 288, + 438 + ], + "spans": [ + { + "bbox": [ + 53, + 384, + 288, + 438 + ], + "type": "text", + "content": "[8] Xiangxiang Chu, Zhi Tian, Yuqing Wang, Bo Zhang, Haibing Ren, Xiaolin Wei, Huaxia Xia, and Chunhua Shen. Twins: Revisiting the design of spatial attention in vision transformers. In Conference on Neural Information Processing Systems (NeurIPS), 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 441, + 288, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 441, + 288, + 506 + ], + "spans": [ + { + "bbox": [ + 53, + 441, + 288, + 506 + ], + "type": "text", + "content": "[9] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Computer Vision and Pattern Recognition Conference (CVPR), 2016. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 509, + 287, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 509, + 287, + 542 + ], + "spans": [ + { + "bbox": [ + 48, + 509, + 287, + 542 + ], + "type": "text", + "content": "[10] Martin Danelljan, Goutam Bhat, Christoph Mayer, and Matthieu Paul. Pytracking. https://github.com/visionml/pytracking, 2019.6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 544, + 287, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 287, + 588 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 287, + 588 + ], + "type": "text", + "content": "[11] Jinhong Deng, Wen Li, Yuhua Chen, and Lixin Duan. Unbiased mean teacher for cross-domain object detection. In Computer Vision and Pattern Recognition Conference (CVPR), 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 590, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 667 + ], + "type": "text", + "content": "[12] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations (ICLR), 2021. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "text", + "content": "[13] Yoav Freund and Robert E Schapire. A decision-theoretic generalization of on-line learning and an application to boosting. Journal of Computer and System Sciences, 55(1):119-139, 1997. 1, 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[14] Angus Galloway, Anna Golubeva, Thomas Tanay, Medhat Moussa, and Graham W. Taylor. Batch normalization is a cause of adversarial vulnerability. In *ArXiv*, 2020. 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 107, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 139 + ], + "type": "text", + "content": "[15] Yaroslav Ganin and Victor Lempitsky. Unsupervised domain adaptation by backpropagation. In International Conference on Machine Learning (ICML), 2015. 2, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 140, + 545, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 140, + 545, + 193 + ], + "spans": [ + { + "bbox": [ + 308, + 140, + 545, + 193 + ], + "type": "text", + "content": "[16] Yaroslav Ganin, E. Ustinova, Hana Ajakan, Pascal Germain, H. Larochelle, Francois Lavolette, Mario Marchand, and Victor S. Lempitsky. Domain-adversarial training of neural networks. In Journal of Machine Learning Research, 2016. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 195, + 545, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 195, + 545, + 237 + ], + "spans": [ + { + "bbox": [ + 308, + 195, + 545, + 237 + ], + "type": "text", + "content": "[17] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In Computer Vision and Pattern Recognition Conference (CVPR), 2012. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 239, + 545, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 239, + 545, + 260 + ], + "spans": [ + { + "bbox": [ + 308, + 239, + 545, + 260 + ], + "type": "text", + "content": "[18] Ross Girshick. Fast r-cnn. In International Conference on Computer Vision (ICCV), 2015. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 261, + 545, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 261, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 308, + 261, + 545, + 304 + ], + "type": "text", + "content": "[19] Ross Girshick, Jeff Donahue, Trevor Darrell, and Jitendra Malik. Rich feature hierarchies for accurate object detection and semantic segmentation. In Computer Vision and Pattern Recognition Conference (CVPR), 2014. 2, 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 305, + 545, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 305, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 308, + 305, + 545, + 338 + ], + "type": "text", + "content": "[20] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In International Conference on Computer Vision (ICCV), 2017. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 338, + 545, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 338, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 308, + 338, + 545, + 381 + ], + "type": "text", + "content": "[21] Kaiming He, X. Zhang, Shaoqing Ren, and Jian Sun. Spatial pyramid pooling in deep convolutional networks for visual recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 37, 2015. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 383, + 545, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 383, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 308, + 383, + 545, + 415 + ], + "type": "text", + "content": "[22] Zhenwei He and Lei Zhang. Multi-adversarial faster-rcnn for unrestricted object detection. In International Conference on Computer Vision (ICCV), 2019. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 416, + 545, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 416, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 308, + 416, + 545, + 469 + ], + "type": "text", + "content": "[23] Dan Hendrycks, Norman Mu, Ekin D. Cubuk, Barret Zoph, Justin Gilmer, and Balaji Lakshminarayanan. AugMix: A simple data processing method to improve robustness and uncertainty. In International Conference on Learning Representations (ICLR), 2020. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 471, + 545, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 471, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 308, + 471, + 545, + 503 + ], + "type": "text", + "content": "[24] Tin Kam Ho. Random decision forests. In International Conference on Document Analysis and Recognition (ICDAR), volume 1, 1995. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 504, + 545, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 504, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 308, + 504, + 545, + 536 + ], + "type": "text", + "content": "[25] Minh Hoai and Andrew Zisserman. Talking heads: Detecting humans and recognizing their interactions. In Computer Vision and Pattern Recognition Conference (CVPR), 2014. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 537, + 545, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 537, + 545, + 591 + ], + "spans": [ + { + "bbox": [ + 308, + 537, + 545, + 591 + ], + "type": "text", + "content": "[26] Jiaxing Huang, Dayan Guan, Aoran Xiao, and Shijian Lu. Model adaptation: Historical contrastive learning for unsupervised domain adaptation without source data. In Conference on Neural Information Processing Systems (NeurIPS), 2021. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 592, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 592, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 308, + 592, + 545, + 635 + ], + "type": "text", + "content": "[27] Naoto Inoue, Ryosuke Furuta, Toshihiko Yamasaki, and Kiyoharu Aizawa. Cross-domain weakly-supervised object detection through progressive domain adaptation. In Computer Vision and Pattern Recognition Conference (CVPR), 2018. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 636, + 545, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 636, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 308, + 636, + 545, + 658 + ], + "type": "text", + "content": "[28] Glenn Jocher. Yolov5. https://github.com/ultralytics/yolov5, 2020.2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 713 + ], + "type": "text", + "content": "[29] Seunghyeon Kim, Jaehoon Choi, Taekyung Kim, and Chang-ick Kim. Self-training and adversarial background regularization for unsupervised domain adaptive one-stage object detection. In International Conference on Computer Vision (ICCV), 2019. 2" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21597" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 715 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[30] Samuli Laine and Timo Aila. Temporal ensembling for semi-supervised learning. In International Conference on Learning Representations (ICLR), 2017. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 288, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 288, + 172 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 288, + 172 + ], + "type": "text", + "content": "[31] Chuyi Li, Lulu Li, Hongliang Jiang, Kaiheng Weng, Yifei Geng, Liang Li, Zaidan Ke, Qingyuan Li, Meng Cheng, Weiqiang Nie, Yiduo Li, Bo Zhang, Yufei Liang, Linyuan Zhou, Xiaoming Xu, Xiangxiang Chu, Xiaoming Wei, and Xiaolin Wei. Yolov6: A single-stage object detection framework for industrial applications. In ArXiv, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 172, + 288, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 172, + 288, + 216 + ], + "spans": [ + { + "bbox": [ + 48, + 172, + 288, + 216 + ], + "type": "text", + "content": "[32] Junnan Li, Richard Socher, and Steven C.H. Hoi. Dividemix: Learning with noisy labels as semi-supervised learning. In International Conference on Learning Representations (ICLR), 2020. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 216, + 288, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 216, + 288, + 260 + ], + "spans": [ + { + "bbox": [ + 48, + 216, + 288, + 260 + ], + "type": "text", + "content": "[33] Shuaifeng Li, Mao Ye, Xiatian Zhu, Lihua Zhou, and Lin Xiong. Source-free object detection by learning to overlook domain style. In Computer Vision and Pattern Recognition Conference (CVPR), 2022. 2, 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 261, + 288, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 261, + 288, + 304 + ], + "spans": [ + { + "bbox": [ + 48, + 261, + 288, + 304 + ], + "type": "text", + "content": "[34] Wuyang Li, Xinyu Liu, and Yixuan Yuan. Sigma: Semantic-complete graph matching for domain adaptive object. In Computer Vision and Pattern Recognition Conference (CVPR), 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 304, + 288, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 304, + 288, + 358 + ], + "spans": [ + { + "bbox": [ + 48, + 304, + 288, + 358 + ], + "type": "text", + "content": "[35] Xianfeng Li, Weijie Chen, Di Xie, Shicai Yang, Peng Yuan, Shiliang Pu, and Yueting Zhuang. A free lunch for unsupervised domain adaptive object detection without source data. In AAAI Conference on Artificial Intelligence (AAAI), 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 359, + 288, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 359, + 288, + 415 + ], + "spans": [ + { + "bbox": [ + 48, + 359, + 288, + 415 + ], + "type": "text", + "content": "[36] Yanghao Li, Chao-Yuan Wu, Haoqi Fan, Karttikeya Mangalam, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. Mvity2: Improved multiscale vision transformers for classification and detection. In Computer Vision and Pattern Recognition Conference (CVPR), 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 415, + 288, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 288, + 470 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 288, + 470 + ], + "type": "text", + "content": "[37] Yu-Jhe Li, Xiaoliang Dai, Chih-Yao Ma, Yen-Cheng Liu, Kan Chen, Bichen Wu, Zijian He, Kris Kitani, and Peter Vajda. Cross-domain adaptive teacher for object detection. In Computer Vision and Pattern Recognition Conference (CVPR), 2022. 2, 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 470, + 288, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 470, + 288, + 503 + ], + "spans": [ + { + "bbox": [ + 48, + 470, + 288, + 503 + ], + "type": "text", + "content": "[38] Zhizhong Li and Derek Hoiem. Learning without forgetting. IEEE Transactions on Pattern Analysis and Machine Intelligence, 40(12):2935-2947, dec 2018. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 503, + 288, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 503, + 288, + 548 + ], + "spans": [ + { + "bbox": [ + 48, + 503, + 288, + 548 + ], + "type": "text", + "content": "[39] Tsung-Yi Lin, Piotr Dollar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Computer Vision and Pattern Recognition Conference (CVPR), 2017. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 548, + 288, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 288, + 591 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 288, + 591 + ], + "type": "text", + "content": "[40] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In International Conference on Computer Vision (ICCV), 2017. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 591, + 288, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 591, + 288, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 591, + 288, + 635 + ], + "type": "text", + "content": "[41] Tsung-Yi Lin, Michael Maire, Serge Belongie, Lubomir Bourdev, Ross Girshick, James Hays, Pietro Perona, Deva Ramanan, C. Lawrence Zitnick, and Piotr Dólar. Microsoft coco: Common objects in context. In ArXiv, 2014. 5, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 635, + 288, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 288, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 288, + 690 + ], + "type": "text", + "content": "[42] Chen Liu, Mathieu Salzmann, Tao Lin, Ryota Tomioka, and Sabine Susstrunk. On the loss landscape of adversarial training: Identifying challenges and how to overcome them. In Conference on Neural Information Processing Systems (NeurIPS), 2020. 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 690, + 288, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 690, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 48, + 690, + 288, + 715 + ], + "type": "text", + "content": "[43] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 545, + 715 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 326, + 72, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 72, + 545, + 95 + ], + "type": "text", + "content": "Berg. Ssd: Single shot multibox detector. In European Conference on Computer Vision (ECCV), 2016. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 96, + 545, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 152 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 152 + ], + "type": "text", + "content": "[44] Yen-Cheng Liu, Chih-Yao Ma, Zijian He, Chia-Wen Kuo, Kan Chen, Peizhao Zhang, Bichen Wu, Zsolt Kira, and Peter Vajda. Unbiased teacher for semi-supervised object detection. In International Conference on Learning Representations (ICLR), 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 152, + 545, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 208 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 208 + ], + "type": "text", + "content": "[45] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In International Conference on Computer Vision (ICCV), 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 208, + 545, + 265 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 208, + 545, + 265 + ], + "spans": [ + { + "bbox": [ + 307, + 208, + 545, + 265 + ], + "type": "text", + "content": "[46] Muhammad Akhtar Munir, Muhammad Haris Khan, M. Saquib Sarfraz, and Mohsen Ali. Ssal: Synergizing between self-training and adversarial learning for domain adaptive object detection. In Conference on Neural Information Processing Systems (NeurIPS), 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 265, + 545, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 265, + 545, + 311 + ], + "spans": [ + { + "bbox": [ + 307, + 265, + 545, + 311 + ], + "type": "text", + "content": "[47] Supreeth Narasimhaswamy, Trung Nguyen, and Minh Hoai. Detecting hands and recognizing physical contact in the wild. In Conference on Neural Information Processing Systems (NeurIPS), 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 311, + 545, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 311, + 545, + 357 + ], + "spans": [ + { + "bbox": [ + 307, + 311, + 545, + 357 + ], + "type": "text", + "content": "[48] Supreeth Narasimhaswamy, Thanh Nguyen, Mingzhen Huang, and Minh Hoai. Whose hands are these? hand detection and hand-body association in the wild. In Computer Vision and Pattern Recognition Conference (CVPR), 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 357, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 357, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 307, + 357, + 545, + 403 + ], + "type": "text", + "content": "[49] Supreeth Narasimhaswamy, Zhengwei Wei, Yang Wang, Justin Zhang, and Minh Hoai. Contextual attention for hand detection in the wild. In International Conference on Computer Vision (ICCV), 2019. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 403, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 403, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 307, + 403, + 545, + 437 + ], + "type": "text", + "content": "[50] Thanh Nguyen, Chau Pham, Khoi Nguyen, and Minh Hoai. Few-shot object counting and detection. In European Conference on Computer Vision (ECCV), 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 438, + 545, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 438, + 545, + 482 + ], + "spans": [ + { + "bbox": [ + 307, + 438, + 545, + 482 + ], + "type": "text", + "content": "[51] Kento Nishi, Yi Ding, Alex Rich, and Tobias Höllerer. Augmentation strategies for learning with noisy labels. In Computer Vision and Pattern Recognition Conference (CVPR), 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 483, + 545, + 528 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 483, + 545, + 528 + ], + "spans": [ + { + "bbox": [ + 307, + 483, + 545, + 528 + ], + "type": "text", + "content": "[52] German I. Parisi, Ronald Kemker, Jose L. Part, Christopher Kanan, and Stefan Wermter. Continual lifelong learning with neural networks: A review. Neural Networks, 113:54-71, 2019. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 529, + 545, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 529, + 545, + 563 + ], + "spans": [ + { + "bbox": [ + 307, + 529, + 545, + 563 + ], + "type": "text", + "content": "[53] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In ArXiv, 2015. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 563, + 545, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 563, + 545, + 587 + ], + "spans": [ + { + "bbox": [ + 307, + 563, + 545, + 587 + ], + "type": "text", + "content": "[54] Joseph Redmon and Ali Farhadi. YOLO9000: better, faster, stronger. In ArXiv, 2016. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 587, + 545, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 587, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 307, + 587, + 545, + 611 + ], + "type": "text", + "content": "[55] Joseph Redmon and Ali Farhadi. Yolov3: An incremental improvement. In ArXiv, 2018. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 611, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 611, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 307, + 611, + 545, + 657 + ], + "type": "text", + "content": "[56] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In Conference on Neural Information Processing Systems (NeurIPS), 2015. 2, 3, 4" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 658, + 545, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 545, + 715 + ], + "type": "text", + "content": "[57] Aruni RoyChowdhury, Prithvijit Chakrabarty, Ashish Singh, SouYoung Jin, Huaizu Jiang, Liangliang Cao, and Erik Learned-Miller. Automatic adaptation of object detectors to new domains using self-training. In Computer Vision and Pattern Recognition Conference (CVPR), 2019. 1, 2, 3, 7, 8" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21598" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[58] Kuniaki Saito, Yoshitaka Ushiku, Tatsuya Harada, and Kate Saenko. Strong-weak distribution alignment for adaptive object detection. In Computer Vision and Pattern Recognition Conference (CVPR), 2019. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 287, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 287, + 149 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 287, + 149 + ], + "type": "text", + "content": "[59] Chawin Sitawarin, Supriyo Chakraborty, and David A. Wagner. Improving adversarial robustness through progressive hardening. In ArXiv, 2020. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 151, + 287, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 151, + 287, + 194 + ], + "spans": [ + { + "bbox": [ + 49, + 151, + 287, + 194 + ], + "type": "text", + "content": "[60] Kihyuk Sohn, Zizhao Zhang, Chun-Liang Li, Han Zhang, Chen-Yu Lee, and Tomas Pfister. A simple semi-supervised learning framework for object detection. In arXiv, 2020. 2, 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 195, + 287, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 195, + 287, + 237 + ], + "spans": [ + { + "bbox": [ + 49, + 195, + 287, + 237 + ], + "type": "text", + "content": "[61] Arvind P. Sridhar, Chawin Sitawarin, and David Wagner. Mitigating adversarial training instability with batch normalization. In International Conference on Learning Representations Workshop (ICLRW), 2021. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 239, + 287, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 239, + 287, + 282 + ], + "spans": [ + { + "bbox": [ + 49, + 239, + 287, + 282 + ], + "type": "text", + "content": "[62] Antti Tarvainen and Harri Valpola. Weight-averaged consistency targets improve semi-supervised deep learning results. In Conference on Neural Information Processing Systems (NeurIPS), 03 2017. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 284, + 287, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 284, + 287, + 315 + ], + "spans": [ + { + "bbox": [ + 49, + 284, + 287, + 315 + ], + "type": "text", + "content": "[63] Alexandru Cristian Telea. An image inpainting technique based on the fast marching method. Journal of Graphics Tools, 9:23-34, 2004. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 317, + 287, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 317, + 287, + 348 + ], + "spans": [ + { + "bbox": [ + 49, + 317, + 287, + 348 + ], + "type": "text", + "content": "[64] Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. FCOS: Fully convolutional one-stage object detection. In International Conference on Computer Vision (ICCV), 2019. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 350, + 287, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 350, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 49, + 350, + 287, + 392 + ], + "type": "text", + "content": "[65] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Conference on Neural Information Processing Systems (NeurIPS), 2017. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 394, + 287, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 394, + 287, + 448 + ], + "spans": [ + { + "bbox": [ + 49, + 394, + 287, + 448 + ], + "type": "text", + "content": "[66] Vikas Verma, Alex Lamb, Christopher Beckham, Amir Najafi, Ioannis Mitliagkas, David Lopez-Paz, and Yoshua Bengio. *Manifold mixup: Better representations by interpolating hidden states.* In International Conference on Machine Learning (ICML), 2019. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 449, + 287, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 449, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 49, + 449, + 287, + 491 + ], + "type": "text", + "content": "[67] Chien-Yao Wang, Alexey Bochkovskiy, and Hong-Yuan Mark Liao. Yolov7: Trainable bag-of-freebies sets new state-of-the-art for real-time object detectors. In ArXiv, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 493, + 287, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 493, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 49, + 493, + 287, + 525 + ], + "type": "text", + "content": "[68] Chien-Yao Wang, I-Hau Yeh, and Hongpeng Liao. You only learn one representation: Unified network for multiple tasks. In ArXiv, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 526, + 287, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 526, + 287, + 580 + ], + "spans": [ + { + "bbox": [ + 49, + 526, + 287, + 580 + ], + "type": "text", + "content": "[69] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In International Conference on Computer Vision (ICCV), 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 582, + 287, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 582, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 49, + 582, + 287, + 613 + ], + "type": "text", + "content": "[70] Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. Detector2. https://github.com/facebookresearch/detectron2, 2019.6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 615, + 287, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 615, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 49, + 615, + 287, + 657 + ], + "type": "text", + "content": "[71] Lin Xiong, Mao Ye, Dan Zhang, Yan Gan, Xue Li, and Yingying Zhu. Source data-free domain adaptation of object detector through domain-specific perturbation. International Journal of Intelligent Systems, 36:3746 - 3766, 2021. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 49, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 658, + 287, + 712 + ], + "type": "text", + "content": "[72] Yunqiu Xu, Yifan Sun, Zongxin Yang, Jiaxu Miao, and Yi Yang. H" + }, + { + "bbox": [ + 49, + 658, + 287, + 712 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 49, + 658, + 287, + 712 + ], + "type": "text", + "content": "FA R-CNN: Holistic and hierarchical feature alignment for cross-domain weakly supervised object detection. In Computer Vision and Pattern Recognition Conference (CVPR), 2022. 2, 7, 8" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 396 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[73] Jihan Yang, Shaoshuai Shi, Zhe Wang, Hongsheng Li, and Xiaojuan Qi. St3d: Self-training for unsupervised domain adaptation on 3d object detection. In Computer Vision and Pattern Recognition Conference (CVPR), 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 118, + 545, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 172 + ], + "type": "text", + "content": "[74] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darryll. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In Computer Vision and Pattern Recognition Conference (CVPR), 2020. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 175, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 228 + ], + "type": "text", + "content": "[75] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In International Conference on Computer Vision (ICCV), 2019. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 230, + 545, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 230, + 545, + 261 + ], + "spans": [ + { + "bbox": [ + 308, + 230, + 545, + 261 + ], + "type": "text", + "content": "[76] Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In ArXiv, 2017. 1, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 264, + 545, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 264, + 545, + 317 + ], + "spans": [ + { + "bbox": [ + 308, + 264, + 545, + 317 + ], + "type": "text", + "content": "[77] Pengchuan Zhang, Xiyang Dai, Jianwei Yang, Bin Xiao, Lu Yuan, Lei Zhang, and Jianfeng Gao. Multi-scale vision longformer: A new vision transformer for high-resolution image encoding. In International Conference on Computer Vision (ICCV), 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 319, + 545, + 362 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 319, + 545, + 362 + ], + "spans": [ + { + "bbox": [ + 308, + 319, + 545, + 362 + ], + "type": "text", + "content": "[78] Liang Zhao and Limin Wang. Task-specific inconsistency alignment for domain adaptive object detection. In Computer Vision and Pattern Recognition Conference (CVPR), 2022. 2, 7, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 365, + 545, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 365, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 308, + 365, + 545, + 396 + ], + "type": "text", + "content": "[79] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable DETR: deformable transformers for end-to-end object detection. In ArXiv, 2020. 2" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "21599" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Object Discovery From Motion-Guided Tokens/cce05c92-f0c6-4fb5-8422-8cb2ccb76230_content_list.json b/2023/Object Discovery From Motion-Guided Tokens/cce05c92-f0c6-4fb5-8422-8cb2ccb76230_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d31d64d69510c71feadcea583aeca9bda6063b85 --- /dev/null +++ b/2023/Object Discovery From Motion-Guided Tokens/cce05c92-f0c6-4fb5-8422-8cb2ccb76230_content_list.json @@ -0,0 +1,1771 @@ +[ + { + "type": "text", + "text": "Object Discovery from Motion-Guided Tokens", + "text_level": 1, + "bbox": [ + 250, + 130, + 720, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhipeng Bao\\*,1 Pavel Tokmakov2 Yu-Xiong Wang3 Adrien Gaidon2 Martial Hebert1 CMU 2Toyota Research Institute 3UIUC", + "bbox": [ + 241, + 179, + 725, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{zbao, hebert}@cs.cmu.edu", + "bbox": [ + 125, + 237, + 323, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{pavel).(tokmakov, adrien).(gaidon}@tri.global", + "bbox": [ + 354, + 237, + 683, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "yxw@illinois.edu", + "bbox": [ + 714, + 238, + 841, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 286, + 313, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Object discovery - separating objects from the background without manual labels - is a fundamental open challenge in computer vision. Previous methods struggle to go beyond clustering of low-level cues, whether handcrafted (e.g., color, texture) or learned (e.g., from auto-encoders). In this work, we augment the auto-encoder representation learning framework with two key components: motion-guidance and mid-level feature tokenization. Although both have been separately investigated, we introduce a new transformer decoder showing that their benefits can compound thanks to motion-guided vector quantization. We show that our architecture effectively leverages the synergy between motion and tokenization, improving upon the state of the art on both synthetic and real datasets. Our approach enables the emergence of interpretable object-specific mid-level features, demonstrating the benefits of motion-guidance (no labeling) and quantization (interpretability, memory efficiency).", + "bbox": [ + 75, + 318, + 473, + 575 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 590, + 209, + 607 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Objects are central in human and computer vision. In the former, they are a fundamental primitive used to decompose the complexity of the visual world into an actionable representation. This abstraction in turn enables higher-level cognitive abilities, such as casual reasoning and planning [32, 54]. In computer vision, object detection has achieved remarkable progress [7, 45] and is now an essential component in many applications (e.g., driving, robotics). However, these models require a large amount of manual labels from a fixed vocabulary of categories. Consequently, learning unsupervised, object-centric representations is an important step towards scaling up computer vision to the real world.", + "bbox": [ + 75, + 617, + 470, + 797 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This topic has received renewed attention recently thanks to structured generative networks with iterative inference over a fixed set of variables [6, 16, 22, 38, 39]. These methods cluster pixels in the feature space of an auto-encoder, exhibiting behavior similar to grouping based on low-level cues,", + "bbox": [ + 75, + 799, + 470, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bd50f31a004d67ebf426562ab8a6f0cdc614f4ea730ecdde0bf8fb70e446e90c.jpg", + "image_caption": [ + "Slot Representation" + ], + "image_footnote": [], + "bbox": [ + 504, + 287, + 694, + 361 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/00bb9122d6004180f5f3a421aff54e947c80f7e98bbf1f829bbf45e00202e82f.jpg", + "image_caption": [ + "Token Representation", + "Figure 1. TRI-PD dataset results: (left) top-10 foreground slot segments produced by our approach; (right) corresponding token representations. Compared to raw images, tokens in our framework present a more structured and compact space for reconstruction." + ], + "image_footnote": [], + "bbox": [ + 697, + 287, + 888, + 361 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "such as color or texture. Hence, they are restricted to toy images with colored geometric shapes on a plain background, and fail on more complex realistic scenes [4].", + "bbox": [ + 496, + 446, + 893, + 491 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Two main types of works attempt to address this shortcoming. The first family of methods sets out to simplify the grouping problem by introducing more structure into the output space, e.g., reconstructing optical flow [36] or depth [15]. They, however, require supervision, either in the form of known poses [36] or ground truth bounding boxes [15], veering away from the unsupervised goal of object discovery. In contrast, Bao et al. [4] resolve the object-background ambiguity by explicitly integrating an unsupervised motion segmentation algorithm [11] into the pipeline, showing substantial progress on realistic scenes. The second main direction to improve object discovery focuses on improving the decoder part of auto-encoding architectures [52, 53], replacing convolutional decoders with transformers [43, 60] combined with discrete variational auto-encoders (DVAE) [47] to reduce memory footprint. These more sophisticated architectures improve performance without additional supervision, including on real-world sequences. However, these methods are evaluated with different protocols (metrics, datasets) and therefore no clear architectural principles have emerged yet for unsupervised object discovery.", + "bbox": [ + 496, + 492, + 895, + 809 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we introduce a novel architecture, Motion-guided Tokens (MoTok), based on the combination of two fundamental structural principles: motion and discretization. We define objects as discrete entities that might have an independent motion. As prior works have shown encouraging results thanks to unsupervised motion guidance and better", + "bbox": [ + 496, + 810, + 895, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Work done during an internship at TRI", + "bbox": [ + 94, + 886, + 308, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "22972", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "transformer-based decoders, we propose to leverage motion to guide tokenization, the vector quantization process at the heart of attention mechanisms in transformer architectures (See Figure 1). In addition, to comprehensively evaluate the contributions of prior works, we ablate key design choices proposed in the past, such as the decoder architecture and reconstruction space, in a unified framework.", + "bbox": [ + 75, + 90, + 470, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our key contributions are as follows. (1) We introduce a novel auto-encoder architecture, MoTok, for unsupervised video object discovery with a new transformer decoder leveraging unsupervised motion-guided tokenization. (2) Our results on real and synthetic datasets show that with sufficient capacity in the decoder, motion guidance alleviates the need for labels, optical flow, or depth decoding thanks to tokenization, improving upon the state of the art. (3) We show that our motion-guided tokens map to interpretable mid-level features, going beyond typical clusters of low-level features, thus explaining why our method scales to challenging realistic videos. Our code, models, and synthetic data are made available at https://github.com/zpbao/MoTok/.", + "bbox": [ + 75, + 198, + 473, + 409 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 419, + 218, + 435 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we tackle the object discovery problem in realistic videos with object-centric representation by capitalizing on motion-guided tokens. We review the most relevant works in these areas below.", + "bbox": [ + 75, + 445, + 470, + 505 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Object Discovery tackles the problem of separating objects from background without manual labels [4]. Classic computer vision methods use appearance-based perceptual grouping to parse a scene into object-like regions [2, 18, 37]. Among them, [2] first propose a multiscale fast normalized cuts algorithm and achieved robust object segmentation performance for static images.", + "bbox": [ + 75, + 508, + 470, + 613 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, the topic of object discovery has experienced renewed attention under the name of unsupervised object-centric representation learning. A wide variety of learning-based methods has been introduced [6, 16, 22, 23, 31, 38, 39, 52, 61, 64], usually with an encoder-decoder architecture [35, 46]. These methods aim to learn compositional feature representations, e.g., a set of variables that can bind to objects in an image [6, 16, 22, 23, 39, 52, 61, 64], or a video [4, 15, 31, 33, 36, 38, 48, 51, 53, 62]. Among them, [39] first formulate the SlotAttention framework, which is used to bind a set of variables, called slots, to image locations. The slots are then decoded individually and combined to reconstruct the image.", + "bbox": [ + 75, + 614, + 473, + 810 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, without additional constraints, such methods tend to converge to pixel grouping based on low-level cues, such as color, and do not generalize to realistic images or videos with complex backgrounds. To address this limitation, [36] and [15] extend the slot concept from static images to videos via reconstructing in the optical flow or depth space", + "bbox": [ + 75, + 810, + 470, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "respectively. The intuition behind these methods is that this space provides stronger cues to separate the objects from the background. Separately, Bao et al. [4] use motion cues to guide the slots to find moving objects and then generalize to all the objects that can move.", + "bbox": [ + 496, + 90, + 890, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In another line of work, Singh at al. [53] show that a combination of a more powerful transformer decoder [13] and discrete variational auto-encoder [47] can enhance the object discovery performance. Different from these works, we propose a unified architecture for object discovery that is flexible to different choices of decoders and reconstruction space. We also introduce the vector quantized features as an additional reconstruction space with motion guidance.", + "bbox": [ + 496, + 171, + 893, + 294 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Finally, several recent works also leverage 3D geometry as inductive biases to enforce the learning-based models to focus on object-like regions [10, 14, 26, 55]. Though these methods remain limited to the toy, synthetic environments, the underlying geometric priors are orthogonal to our approach and have a great potential to be combined with our proposed method as a future direction.", + "bbox": [ + 496, + 297, + 893, + 404 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Vector quantization is originally developed for data compression in signal processing [20]. More recently, [58] propose the Vector-Quantized Variational Autoencoder (VQ-VAE), which learns a discrete representation of images, and models their distribution aggressively. This technique has motivated a series of investigations in solving different computer vision tasks including image synthesis [17, 24, 44], video generation [56, 63], and language modeling [24, 42, 43], to name a few. In this work, we adopt the vector quantization technique as a mid-level representation in our object discovery framework. The intuition is that reconstructing in a structured, low-dimensional feature space rather than the high-variance color space should simplify the task of resolving the object and background ambiguity. In comparison, [53] also introduce a mid-level feature space for object discovery in videos. However, they use DVAE, not VQ-VAE, and are motivated primarily by improving training efficiency. Moreover, they did not explore combing vector quantization with motion cues in an end-to-end trainable framework.", + "bbox": [ + 496, + 409, + 895, + 698 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Transformers are originally proposed for sequence-to-sequence modeling in natural language processing [60]. They use a multi-head attention mechanism, instead of the recurrent memory units, to aggregate information from the input. Recently, vision transformer (ViT) [13] and its derivations have achieved state-of-the-art in several visual tasks [7,29,30,41,48,49,67]. Among them, Perceiver IO [29] design a computationally efficient architecture to handle arbitrary outputs in addition to arbitrary inputs, greatly enhancing the generalizability of the transformer architecture. We employ a variant of a Perceiver module as a decoder in our object discovery framework to combine high representational power with computational efficiency.", + "bbox": [ + 496, + 704, + 895, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "22973", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2bf1afd0f9ca6580390c5f700aebeeff549fe9f4019eac6dc97d1852ab074670.jpg", + "image_caption": [ + "Figure 2. Model architecture of the proposed Motion-guided Tokens (MoTok) framework. MoTok is a unified framework for video object discovery that is flexible with different choices of decoders and reconstruction spaces. Our framework effectively leverages the synergy between motion and tokenization, and enables the emergence of interpretable object-specific mid-level features." + ], + "image_footnote": [], + "bbox": [ + 81, + 92, + 890, + 244 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 311, + 168, + 325 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We now explain the proposed Motion-guided Tokens (Mo-Tok) framework in detail. Its architecture is shown in Figure 2. We first introduce a motion-guided slot learning framework [4] in Section 3.1. We then describe the slot decoder units in Section 3.2, which decode the slot features into a reconstruction space. In Section 3.3, we describe the choices of reconstruction space and explain the vector-quantized reconstruction space in detail. Finally, we demonstrate how we optimize the model in Section 3.4.", + "bbox": [ + 75, + 337, + 473, + 472 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary: Motion-Guided Slot Learning", + "text_level": 1, + "bbox": [ + 76, + 483, + 447, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our object-centric representation learning module is derived from [4] (Figure 2, left). Concretely, given a sequence of video frames $\\{I^1, I^2, \\dots, I^T\\}$ , we first process each frame through an encoder CNN to obtain an individual frame representation $H^t = f_{\\mathrm{enc}}(I^t)$ . These individual representations are aggregated by a spatio-temporal Convolutional Gated Recurrent Unit (ConvGRU) [3] to obtain video encoding via $H'^t = \\mathrm{ConvGRU}(R^{t-1}, H^t)$ , where $R^{t-1} \\in \\mathbb{R}^{h' \\times w' \\times d_{\\mathrm{inp}}}$ is the recurrent memory state.", + "bbox": [ + 75, + 507, + 470, + 642 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Next, we perform a single attention operation with $K$ slots to directly compute the slot state $S^t = W^t v(H'^t)$ , where $W^t \\in \\mathbb{R}^{K \\times N}$ is the attention matrix, $N = h' \\times w'$ indicates the flapped shape dimension, and $v(\\cdot)$ is the value embedding function. $W^t$ is computed using the slot state in the previous frame $S^{t-1}$ and the input feature $H'^t$ . For the first frame, we use a learnable initial state $S^0$ . At the same time, for each slot $s_i^t$ , we can also obtain the attention mask $W_{i,i}^t$ .", + "bbox": [ + 75, + 643, + 470, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In [4], a motion cue is also added to guide the slots to find moving objects. A set of sparse, instance-level motion segmentation masks $\\mathcal{M} = \\{M^1,M^2,\\dots,M^T\\}$ is assumed to be provided with every video, with $M^{t} = \\{m_{1},m_{2},\\dots,m_{C^{t}}\\}$ , where $C^t$ is the number of moving objects that were successfully segmented in frame $t$ , and $m_j\\in \\{0,1\\}^{h'\\times w'}$ is a binary mask. The attention mask $W^{t}\\in \\mathbb{R}^{K\\times N}$ , is then supervised with the motion segments. $M^t$ is also considered", + "bbox": [ + 75, + 780, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "as a set of length $K$ padded with $\\emptyset$ (no object), and a bipartite matching is found between them with the lowest cost:", + "bbox": [ + 498, + 311, + 890, + 343 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\sigma} = \\underset {\\sigma} {\\arg \\min } \\sum_ {i = 1} ^ {K} \\mathcal {L} _ {\\operatorname {s e g}} \\left(m _ {i}, W _ {:, \\sigma (i)} ^ {t}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 351, + 892, + 392 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{\\mathrm{seg}}(m_i,W_{:, \\sigma(i)}^t)$ is the segmentation loss between the motion mask $m_i$ and the attention map of the slot with index $\\sigma(i)$ . Once the assignment $\\hat{\\sigma}$ has been computed, the final motion supervision objective is defined as follows:", + "bbox": [ + 498, + 401, + 890, + 463 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {m o t i o n}} = \\sum_ {i = 1} ^ {K} \\mathbb {1} _ {\\{m _ {i} \\neq \\emptyset \\}} \\mathcal {L} _ {\\text {s e g}} \\left(m _ {i}, W _ {:, \\hat {\\sigma} (i)} ^ {t}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 470, + 890, + 511 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbb{1}_{\\{m_i \\neq \\emptyset\\}}$ denotes that the loss is only computed for the matched slots and $\\mathcal{L}_{\\mathrm{seg}}$ is the binary cross entropy.", + "bbox": [ + 498, + 518, + 890, + 550 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Slot Decoders", + "text_level": 1, + "bbox": [ + 500, + 558, + 640, + 571 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The goal of the Slot Decoder is to map the slot representation $(S^t, W^t)$ to a 2D feature map $F^t$ for the reconstruction space. As shown in Figure 2 (middle), we propose four choices for the slot decoder.", + "bbox": [ + 498, + 580, + 893, + 640 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Linear Decoder directly maps the slot features $S^t$ to their corresponding positions based on the attention mask $W^t$ :", + "bbox": [ + 498, + 641, + 890, + 672 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nF _ {\\text {l i n e a r}} ^ {t} (\\mathbf {x}) = \\frac {\\sum_ {i = 1} ^ {K} S _ {i} ^ {t} (\\mathbf {x}) W _ {i , \\mathbf {x}} ^ {t}}{\\sum_ {i = 1} ^ {K} W _ {i , \\mathbf {x}} ^ {t}}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 679, + 890, + 720 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{x}$ is an arbitrary 2D position.", + "bbox": [ + 498, + 729, + 736, + 743 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "CNN Decoder further adds two convolutional layers to the 2D feature map formed by Equation 3:", + "bbox": [ + 498, + 743, + 890, + 773 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nF _ {\\mathrm {C N N}} ^ {t} = \\operatorname {C N N} \\left(\\frac {\\sum_ {i = 1} ^ {K} S _ {i} ^ {t} W _ {i , :} ^ {t}}{\\sum_ {i = 1} ^ {K} W _ {i , :} ^ {t}}\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 583, + 780, + 890, + 823 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Transformer Decoder decodes the feature by querying the slot representation with a 2D positional embedding through a transformer decoder:", + "bbox": [ + 498, + 830, + 890, + 873 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nF _ {\\text {t r a n s f o r m e r}} ^ {t} = \\operatorname {T r a n s f o r m e r} \\left(P, S ^ {t}, S ^ {t}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 883, + 890, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "22974", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1: Perceiver decoder" + ], + "code_body": "Perceiver $(S^t,P)$ .. $S^t\\to \\mathrm{Norm}(S^t)$ $\\hat{S}^t =$ SelfAttention $(S^t) + S^t$ $\\hat{S}^t\\rightarrow \\mathrm{Norm}(\\hat{S}^t)$ $\\tilde{S}^t = \\mathrm{MLP}(\\hat{S}^t) + \\hat{S}^t$ $\\tilde{S}^t\\rightarrow \\mathrm{Norm}(\\tilde{S}^t)\\quad P\\rightarrow \\mathrm{Norm}(P)$ $F_{\\mathrm{Perceiver}}^{t} =$ CrossAttention $(P,\\tilde{S}^{t},\\tilde{S}^{t})$", + "bbox": [ + 78, + 111, + 468, + 223 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the query $P\\in \\mathbb{R}^{N\\times d_p}$ is a learnable positional embedding.", + "bbox": [ + 75, + 253, + 470, + 282 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Compared with the previous two linear decoders, the transformer decoder further considers the global connections between the slot features and the input query, so that it will form a more powerful feature map. However, an obvious limitation is that the transformer decoder applies self-attention to the input positional query, which is 1) redundant since the positional embedding itself is learnable, and 2) not computationally efficient therefore limits the scalability of the whole model. To solve this limitation, we further propose the perceiver decoder.", + "bbox": [ + 75, + 285, + 470, + 435 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Perceiver Decoder is inspired by [29], which designs a computationally efficient architecture to handle arbitrary outputs in addition to arbitrary inputs. Different from the original architecture, we only add a self-attention layer to the slot representations, followed by a cross-attention layer for the output and positional embedding queries. The whole procedure is illustrated in Algorithm 1.", + "bbox": [ + 75, + 436, + 468, + 542 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After forming the 2D feature map $F^t$ , we decode it to a reconstruction space with a CNN-based decoder, except for when vector-quantized space is used. As we will show next, an additional decoder is redundant in this case.", + "bbox": [ + 75, + 542, + 470, + 603 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Reconstruction Space", + "text_level": 1, + "bbox": [ + 76, + 616, + 279, + 633 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "There are four choices of reconstruction space in our framework. The plain RGB space contains the most information but also is the most complex to resolve the object/background ambiguity. The flow and depth spaces, as shown in Figure 2 (right), are more structured. Reconstructing in these spaces makes the grouping problem easier. However, there are still shortcomings, e.g. non-moving objects are not captured in the flow space and depth cannot easily distinguish between-objects that are near each other. Moreover, these two spaces are not as informative as the RGB space. In addition, we introduce the VQ-space, which is end-to-end trainable and is both structured and informative. We describe the VQ-space below.", + "bbox": [ + 75, + 640, + 472, + 835 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Vector-quantized reconstruction space. Different from the other three reconstruction spaces, here we do not predict the reconstruction directly, but rather supervise the feature map $F^{t}$ to match the latent embedding space of VQ-VAE.", + "bbox": [ + 75, + 839, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In particular, following [58], we define a latent embedding space $E$ as the set of $M$ vectors $e_i$ of dimension $d_{vq}$ , $E = \\{e_i\\in \\mathbb{R}^{d_{vq}}|i = 1,2,\\dots ,M\\}$ . Given an input image $I^t$ , VQ-VAE first processes it with an encoder to get the output $z_e^t = \\mathrm{Encoder}_{\\mathrm{VQ}}(I^t)$ , and then the discrete latent variables $z$ are calculated by a nearest neighborhood search among the discrete feature set $E$ :", + "bbox": [ + 496, + 90, + 893, + 196 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nz _ {q} ^ {t} (\\mathbf {x}) = e _ {k}, \\quad \\text {w h e r e} k = \\operatorname {a r g m i n} _ {j} \\| z _ {e} ^ {t} (\\mathbf {x}) - e _ {j} \\| _ {2}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 205, + 893, + 226 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and $\\mathbf{x}$ is an arbitrary 2D position. The reconstructed image $\\hat{I}^t$ is then decoded by $\\hat{I}^t = \\mathrm{Decoder}_{\\mathrm{VQ}}(z_q^t)$ . The objective of VQ-VAE is:", + "bbox": [ + 496, + 234, + 893, + 279 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {V Q V A E}} = \\log P \\left(I ^ {t} \\mid z _ {q} ^ {t}\\right) + \\left\\| s g \\left[ z _ {e} ^ {t} \\right] - z _ {q} ^ {t} \\right\\| _ {2} + \\left\\| s g \\left[ z _ {q} ^ {t} \\right] - z _ {e} ^ {t} \\right\\| _ {2}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 287, + 892, + 305 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $sg[\\cdot]$ is the stop-gradient operation.", + "bbox": [ + 496, + 311, + 777, + 327 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then we use the quantized feature map $z_{q}^{t}$ as the target signal for the slot feature map $F^{t}$ . The final objective of VQ-VAE and the VQ reconstruction is:", + "bbox": [ + 496, + 327, + 893, + 373 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {V Q}} = \\mathcal {L} _ {\\mathrm {V Q V A E}} + \\| s g [ F ^ {t} ] - z _ {q} ^ {t} \\| _ {2} + \\| s g [ z _ {q} ^ {t} ] - F ^ {t} \\| _ {2}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 382, + 892, + 402 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Motion-guided token representation. The last term in Equation 8, $||sg[F^t] - z_q^t||_2$ , enables the motion signal from slot learning to jointly optimize the token space through the output of the slot decoder. Furthermore, the token representation and the motion cues build a connection linked by the slot learning, thus enabling the emergence of interpretable object-specific mid-level features of tokens. In addition, reconstructing in a more compact token space also benefits the model by better utilizing the motion signal to achieve an improved slot representation and temporal consistency.", + "bbox": [ + 496, + 412, + 893, + 565 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Optimization", + "text_level": 1, + "bbox": [ + 500, + 573, + 638, + 589 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Token contrastive constraint. The goal of reconstructing in the VQ-space is that it is more compact and of lower variation compared with the RGB space. To make the VQ-space more structured, we add an optional contrastive constraint below to the vector space, which increases the independence between latent vectors:", + "bbox": [ + 496, + 599, + 893, + 688 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {c o n t r a s t i v e}} = \\left| \\left| \\mathbb {I} - \\operatorname {s o f t m a x} \\left(E \\cdot E ^ {T}\\right) \\right| \\right|, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 699, + 892, + 717 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbb{I}$ is the identity matrix and $E\\in \\mathbb{R}^{N\\times d_{vq}}$ is the matrix of the feature embedding space $S$ .", + "bbox": [ + 496, + 726, + 890, + 756 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The final loss function is a combination of the reconstruction objective, the motion objective, and the optional contrastive constraint:", + "bbox": [ + 496, + 757, + 893, + 801 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\lambda \\mathcal {L} _ {\\text {m o t i o n}} + \\mathcal {L} _ {\\text {r e c o n}} + \\lambda_ {c} \\mathbb {1} _ {\\mathrm {V Q}} \\mathcal {L} _ {\\text {c o n t r a s t i v e}}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 813, + 892, + 830 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\lambda$ and $\\lambda_{c}$ are weighting factors and $\\mathbb{1}_{\\mathrm{VQ}}$ is an indicator function. For the reconstruction loss, we set $\\mathcal{L}_{\\mathrm{recon}} = \\mathcal{L}_{\\mathrm{VQ}}$ when performing reconstruction in the VQ-space. Otherwise, we use an $L_{2}$ loss for reconstruction in the other three spaces.", + "bbox": [ + 496, + 839, + 893, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "22975", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/58fe986f8c7ec3db75011d693a1a648516c53d1132b86d6cd3e8f5323688aa63.jpg", + "image_caption": [ + "Figure 3. Frame samples from the video datasets used in our experiments. MOVi [21] (left) is a multi-object video dataset created by simulating rigid body dynamics. TRI-PD [4] (middle) is a collection of photo-realistic, synthetic driving videos. KITTI [19] (right) is a real-world benchmark with city driving scenes." + ], + "image_footnote": [], + "bbox": [ + 81, + 90, + 202, + 191 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/77f1baf06ed23c0c31053cf62f6c4478f10efd24de4c939d4d0f0fa7443bca4c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 210, + 90, + 328, + 191 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5e8f6aa14e950aa5c277583a1f42bf4622fe94107069b50b2ba639b9c6d2ee04.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 338, + 90, + 464, + 191 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experimental Evaluation", + "text_level": 1, + "bbox": [ + 76, + 281, + 310, + 297 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 76, + 306, + 282, + 324 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Benchmarks. We evaluate our approach on three popular benchmarks with different complexity (Figure 3).", + "bbox": [ + 75, + 334, + 468, + 364 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "MOVi [21] is a synthetic multi-object video dataset, created by simulating rigid body dynamics. We use this benchmark to ablate and analyze our model architecture. Following previous works [15, 53], we use the most complex subset, MOVi-E, for evaluation. MOVi-E contains both moving and static objects (maximum 20 objects) with linear random camera motion. The resolution of this dataset is $128 \\times 128$ and each video contains 24 frames with a sample rate of 12 FPS. We use standard train and test split for MOVi-E.", + "bbox": [ + 75, + 366, + 468, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "TRI-PD [4,25] is a synthetic dataset based on street driving scenarios collected by using a state-of-the-art synthetic data generation service [1]. The dataset comes with a collection of accurate annotations (e.g. for flow or depth), which we use to ablate the impact of additional information and its quality on various models. Each video in TRI-PD is 10 seconds captured at 20 FPS. There are 924 videos in the training set and 51 videos in the test set. We crop and resize each frame to the resolution of $480 \\times 968$ .", + "bbox": [ + 75, + 503, + 468, + 637 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "KITTI [19] is a real-world benchmark with city driving scenes. We use the whole 151 KITTI videos for training and the instance segmentation subsets with 200 single-frame images for evaluation. The frames are resized to $368 \\times 1248$ .", + "bbox": [ + 75, + 640, + 468, + 700 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines. We compare our methods against the most recent learning-based object discovery models. In particular, SAVi [36] and SAVi++ [15] are two direct extensions of the original slot attention [39] using optical flow and depth to facilitate object-centric learning. Alternatively, STEVE [53] uses a more powerful transformer decoder to enhance the object discovery performance. Bao et al. [4], Karazija et al. [33] utilize motion cues to guide object discovery. In addition, in the supplementary, we compare our approach to [51], which is a concurrent work that relies on ImageNet [12] pretraining [9]. For [4, 15, 36, 53], we use either the official code or public implementation (see supplementary for details). Due to the lack of implementation, we reuse reported results", + "bbox": [ + 75, + 704, + 470, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/88da2141d7738868c42fb1955ea8e0fc4db66be8716a02f1b177e7e3c52a3781.jpg", + "image_caption": [ + "Figure 4. Object discovery in different reconstruction spaces on the evaluation set of MOVi-E. Compared with the other reconstruction spaces, the model trained using the VQ-space demonstrates a more accurate performance and better temporal consistency." + ], + "image_footnote": [], + "bbox": [ + 501, + 88, + 890, + 205 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "with the two very recent approaches [33, 51].", + "bbox": [ + 498, + 297, + 795, + 311 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Metrics. Following prior works, we use the Foreground Adjusted Rand Index (FG. ARI) to evaluate the performance of the models, which captures how well the predicted segmentation masks match ground-truth masks in a permutation-invariant fashion. Notice that FG. ARI is measured in terms of the whole video, so that the temporal consistency of the slots is considered with this metric.", + "bbox": [ + 496, + 321, + 890, + 428 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. We use the same ResNet-18 ConvGRU encoder backbone for all the compared methods following [4]. We set the number of slots as 24 for MOVi-E and 45 for PD and KITTI based on the maximal number of objects in these datasets. On TRI-PD and KITTI, we further downsample the images by 4 for SAVi and SAVi++ to fit the GPU memory. Notice that, the other methods also produce downsampled slot masks, leading to the evaluation under the same resolution for all the methods. All the models are trained for 500 epochs using Adam [34]. During training, we train all the models with a randomly sampled video sequence of a fixed length (6, 5, 5 for MOVi-E, TRI-PD, and KITTI respectively). During the inference time, all the models are evaluated frame by frame until the end. For the FG. ARI measurement on TRI-PD, we discard any instance labels covering an area of $0.5\\%$ or less of the first sampled video frame following [15]. We also evaluate FG. ARI at 5 FPS due to memory limitation. We use batch size 64 for MOVi-E, and 8 for TRI-PD and KITTI.", + "bbox": [ + 496, + 438, + 892, + 724 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our method and Bao et al. [4] require motion segmentation signals. We apply [11], a powerful method pre-trained on the toy FlyingThings3D dataset [40], taking the ground-truth flow (MOVi-E) or RAFT [57] flow as the input. For SAVi and $\\mathrm{SAV_i + + }$ , we do not use the first-frame bounding box supervision for a fair comparison. We generate the optical flow and depth annotation for KITTI using two state-of-the-art methods, RAFT [57] and VIDAR [25]. We also use them for ablations on TRI-PD. More details about the hyper-parameters, training scheme, and annotation generation are provided in the supplementary.", + "bbox": [ + 496, + 734, + 893, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "22976", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/68e7b927209bef216bf007d74f96bc38a4cc2ecce589d06941c406cddc3c4387.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MotionSpaceDecoderFG. ARI
XVQLinear14.6
XVQLinear-CNN16.3
XVQTransformer40.1
XVQPerceiver52.4
XRGBPerceiver49.2
XFlowPerceiver36.3
XDepthPerceiver18.4
XFlow + DepthPerceiver38.0
XVQ (flow)Perceiver44.6
", + "bbox": [ + 117, + 88, + 424, + 219 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Object Discovery on MOVi", + "text_level": 1, + "bbox": [ + 76, + 335, + 320, + 353 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We use MOVi-E dataset to ablate our model with different choices of slot decoder, and reconstruction spaces, aiming to elucidate their roles in object discovery. We also compare the performance with state-of-the-art models in this section.", + "bbox": [ + 75, + 361, + 468, + 421 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Set up. We first fix the reconstruction space as the VQ-space and ablate the decoder architectures (first 4 lines in Table 1), and then we fix the decoder architecture with the best component and reconstruct in different reconstruction space (Line 5 to 9 in Table 1). Since SAVi++ [15] argues that reconstructing in the combined space of flow and depth is the key to their success, we add a variant $Flow + Depth$ , which also reconstructs in this combined space. Furthermore, as VQ-VAE can be trained to reconstruct the image in the flow space as well, we also report a $VQ$ (flow) variant for completeness. We do not include motion cues for the architecture analysis. A more comprehensive ablation is reported in the supplementary.", + "bbox": [ + 75, + 425, + 470, + 621 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Decoder analysis. By comparing the results on the first four lines in Table 1, similar to [53], we find that the capacity of the decoder indeed plays a key role in the object discovery performance. A simple linear decoder fails, but adding an additional CNN decoder can bring limited performance gains. Introducing powerful transformer-based decoders, on the other hand, greatly improves the object discovery capabilities of the model. Between the two transformer variants, the more advanced perceiver decoder achieves better results, while being more computationally efficient.", + "bbox": [ + 75, + 625, + 468, + 776 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Reconstruction space analysis. For the four reconstruction spaces, the VQ-space yields the best object discovery performance. We analyze the learned slots and token representations in Figure 4 and make the following discoveries. Firstly, the model trained using the VQ-space better separates the objects from the background and shows stronger temporal consistency. This is due to the more structured, compact, and lower-variance reconstruction space provided", + "bbox": [ + 75, + 780, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/78987071883cdceba9c7ac07a6b1fa9355cdeef9b64ea16d3da18fc038d70020.jpg", + "table_caption": [ + "Table 1. Model architecture analysis with different choices of the slot decoder and reconstruction space on MOVi-E. Perceiver decoder + VQ reconstruction space yields the best performance, indicating that (1) the capacity of the decoder plays a key role in object discovery; (2) learnable, vector-quantized reconstruction space outperforms fixed alternatives like depth or flow." + ], + "table_footnote": [], + "table_body": "
ModelFG. ARI
Bao et al.51.6
STEVE51.2
SAVi28.1
SAVi++31.7
MoTok (no motion)52.4
MoTok63.8
SAVi (G.T. box)53.4
SAVi++ (G.T. box)84.1
", + "bbox": [ + 588, + 88, + 799, + 215 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Comparison to the state-of-the-art object discovery approaches on the validation sets of MOVi-E using FG. ARI. Our approach outperforms all the recent methods with the help of motion cues and vector quantization.", + "bbox": [ + 496, + 220, + 893, + 277 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "by the quantized features, which simplifies the task of grouping, compared to the raw RGB space.", + "bbox": [ + 496, + 290, + 893, + 319 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Secondly, by comparing Flow and $\\text{Flow+Depth}$ variants, we find that the latter indeed carries more information, allowing the model to better group the objects, which is consistent with [15]. However, interestingly, we also find that given a sufficiently strong decoder, reconstructing in depth or flow space does not bring further improvements compared to RGB, and can even decrease the performance. Finally, comparing Flow and $\\text{VQ}(\\text{flow})$ , we find that although adding the quantization improves performance significantly, this variant still lags behind reconstructing in the raw RGB space, reinforcing our previous conclusions.", + "bbox": [ + 496, + 320, + 893, + 484 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison to the state of the art. In Table 2, we compare our approach to the state-of-the-art methods. Our model achieves the best performance among them. Additionally, compared to SAVi and SAVi++, even our variants that reconstructs in the same space outperform them, indicating that the perceiver decoder and single-stage decoding strategy are the optimal choices. Our model also outperforms STEVE, due to the fact that, in contrast to DVAE, the codebook in VQ-VAE can be jointly optimized with slot learning (see Equation 8). Compared with Bao et al., our model shows better performance even without motion cues used by that method, but further incorporating them into our approach allows it to achieve top results.", + "bbox": [ + 496, + 488, + 893, + 684 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Notice that, we compare with the variants of SAVi and SAVi++ that do not rely on ground-truth bounding boxes at test time. Their published numbers are included at the bottom of the table for reference, but they are not comparable to the other unsupervised methods.", + "bbox": [ + 496, + 685, + 893, + 760 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Object Discovery with Realistic Driving Videos", + "text_level": 1, + "bbox": [ + 498, + 768, + 890, + 785 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Set up. Firstly, following [4], we capitalize on the ground-truth annotations available in TRI-PD, and evaluate the importance of the quality of external signals for various approaches. In particular, we separately evaluate using ground-truth (GT) flow, depth, and motion segmentation, as well as estimating those using state-of-the-art algorithms [11,25,57]. In addition, we report a variant STEVE- $m$ for which we add", + "bbox": [ + 496, + 794, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "22977", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1e22679eb8d0f042d634ca737a3968d2d9f3f7d2ca9bcca49bb70b68632f8324.jpg", + "image_caption": [ + "Figure 5. Reconstruction, slot representation, and token masks for [4], STEVE-m [53], and our method. We use ground-truth motion cues for supervision, and visualize top-10 masks excluding background slots. MoTok effectively leverages the synergy between motion and tokenization, enabling the emergence of interpretable object-specific mid-level features, which simplify the problem of object discovery." + ], + "image_footnote": [], + "bbox": [ + 83, + 88, + 890, + 228 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5bd0306cb78f4e2d829440cb2142cf17ecb273bd11b6cc46b5c31a07f2f78dc5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelSignalGPU Mem. (GB)FG. ARI
GTEST
Bao et al.Motion21.350.434.7
SAViFlow17.317.914.2
SAVi++Depth + Flow17.518.415.3
STEVE-66.712.5-
STEVE-mMotion68.245.732.2
MoTokMotion10.960.655.0
", + "bbox": [ + 117, + 279, + 424, + 377 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Object discovery evaluation of different models on TRI-PD dataset. GT denotes the ground-truth guidance given to each model and EST denotes the estimated cues. State-of-the-art models fail to work without motion cues; introducing token representation helps our model better utilize the motion signal.", + "bbox": [ + 75, + 382, + 470, + 452 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the same motion segmentation supervision as to [4] and our approach for a fair comparison.", + "bbox": [ + 75, + 483, + 468, + 513 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "From Table 3, we can observe that our method strongly outperforms all the baselines. Additionally, we find that even with the ground-truth flow or depth, SAVI and SAVI++ do not work well for realistic videos with complex background and crowded scenes. Without motion cues, even equipped with a powerful transformer decoder, STEVE fails to work in this challenging setting as well. Finally, our model achieves better performance compared to Bao et al. with both GT and estimated motion segments, indicating its better robustness.", + "bbox": [ + 75, + 516, + 468, + 651 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of motion-guided tokens: In Figure 5, we visualize the RGB reconstruction, slot representation, and token masks (if possible) for Bao et al., STEVE-m, and our MoTok. There are several key observations. Firstly, compared with STEVE-m, MoTok enables the emergence of more interpretable mid-level features, even though both models are trained with motion cues and tokenization. This is due to the interplay between object and slot discovery in our end-to-end framework, whereas STEVE-m trains DVAE separately from the rest of the model.", + "bbox": [ + 75, + 656, + 470, + 806 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Secondly, our model achieves the best RGB reconstruction result, even for the latter frames. The high quality of the reconstruction indicates that the model can better take advantage of the appearance signal to optimize the slot representation. Finally, our model demonstrates stronger temporal consistency compared with the baselines, thanks to the struc", + "bbox": [ + 75, + 809, + 470, + 898 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/75203d917d95230b77c51b97a02aa668a0e9ff61b7d6c1d31c443472f0b2aa52.jpg", + "image_caption": [ + "Figure 6. Visualization of the learned tokens for vanilla VQ-VAE and MoTok. The proposed MoTok model shows a better alignment with semantic categories." + ], + "image_footnote": [], + "bbox": [ + 503, + 280, + 887, + 339 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tured reconstruction space. To sum up, our architecture effectively leverages the synergy between motion cues and tokenization, enabling the emergence of interpretable mid-level features, which greatly simplifies the task of object discovery.", + "bbox": [ + 496, + 406, + 892, + 481 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Interpretable tokens. To better illustrate the interpretability of motion-guided tokens, we measured their alignment with the ground-truth semantic labels using cluster purity [50]. Vanilla VQ-VAE (no motion guidance) achieves a purity of 40.1, whereas ours reaches 65.7 (chance performance: 38.5). Qualitative comparison in Figure 6 clearly demonstrates our better semantic alignment.", + "bbox": [ + 496, + 484, + 892, + 590 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Scalability to realistic videos: We additionally report the GPU memory usage per batch in Table 3. The per-slot decoding strategy and the use of transformer decoder limit the scalability of the other baselines. In comparison, our model achieves the best performance with the lowest memory consumption, showing a great generalization capability to real-world videos, thanks to the efficient perceiver decoder and the single-shot decoding strategy.", + "bbox": [ + 496, + 595, + 892, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of objectives. Finally, we ablate the objective design of our model with the GT annotations in Table 4. We build three variants of our model, one without the additional token contrastive constraint defined by Equation 10 (no contrastive), the second without the motion cues (no motion), and the last trained with ground-truth instance masks for all objects to indicate the performance upper bound $(^{*})$ . By comparing the performance of these variants, we make the following observations: (1) in realistic synthetic videos even the strongest architectures fail to resolve the object/background ambiguity in the absence of motion cues, demonstrating the difficulty of object discovery tasks in the real world; (2)", + "bbox": [ + 496, + 719, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "22978", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/47e9175bfed9f7976a7e9220e04369752f643bdd669cc0a0d60fc44da9ec8280.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelFG. ARI
MoTok60.6
MoTok (no contrastive)57.4
MoTok (no motion)15.6
MoTok*64.7
", + "bbox": [ + 81, + 88, + 254, + 141 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/6e7fbb38f7d6eef07001296af5be85c46bcd1f51989dbb349b476088a5129ceb.jpg", + "table_caption": [ + "Table 4. Ablation study on different learning signals. MoTok* is our model trained with ground-truth instance segmentation. The motion cue and the contrastive constraint help improve the capability of the proposed model, almost reaching the performance upper bound (MoTok*)." + ], + "table_footnote": [], + "table_body": "
ModelFG. ARI
Bao et al. [4]47.1
SAVi20.0
SAVi++23.9
STEVE11.9
Karazija et al. [33]50.8
Karazija et al. [33] (WL)51.9
MoTok64.4
", + "bbox": [ + 271, + 88, + 467, + 179 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Evaluation of object discovery in the real-world KITTI dataset. Karazija et al. [33] (WL) is a variant trained with their proposed wrapping loss. Our model achieves state-of-the-art in real-world object discovery.", + "bbox": [ + 264, + 184, + 470, + 282 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "contrastive constraint facilitates diversity during vector quantization, increasing the information content of each token, which helps reduce the object/background ambiguity; (3) the performance of our model is close to the upper bound, indicating the effectiveness of the motion-guided tokens.", + "bbox": [ + 75, + 311, + 472, + 388 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Object Discovery in the Real World", + "text_level": 1, + "bbox": [ + 76, + 397, + 385, + 414 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Set up. To train the object discovery model on the real-world KITTI benchmark, we first pre-train all the models on PD using ground-truth flow and depth and then fine-tune them using estimated annotations on KITTI.", + "bbox": [ + 75, + 422, + 468, + 483 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The comparisons on KITTI are shown in Table 5 and Figure 7. We find that, consistent with our observations on TRI-PD, without motion guidance, both more powerful decoders (STEVE) and the simpler reconstruction space (SAVi and $\\mathrm{SAV_i + + }$ ) fail to work. However, our model still captures good dynamic object segmentation results. Compared to Bao et al., we still achieve a better FG. ARI score and a more solid segmentation mask, indicating the benefit of the model design and the motion-guided tokens. Finally, we outperform the very recent approach of Karazija et al. [33] even when they include an additional warping loss objective.", + "bbox": [ + 75, + 484, + 472, + 650 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Limitations and Future Work", + "text_level": 1, + "bbox": [ + 76, + 660, + 341, + 674 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Object/background ambiguity in the real world. Even after successfully parsing the world into object and background slots, separating those from each other in an unsupervised way is an open challenge. Promising directions include temporal contrastive learning [5, 28], unsupervised clustering [8, 59], and utilizing geometric cues [65, 66].", + "bbox": [ + 75, + 685, + 468, + 776 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Slot drift. From Figure 5, we notice that the final RGB reconstruction quality decreases with longer videos, which indicates that the distribution of the slot representation has shifted in the latter frames. This issue leads to degraded temporal consistency and ambiguous object masks. Better training schemes and model improvements, such as adding data augmentation during training [15], replacing GRU units with long-short memory units [27], and enforcing temporal", + "bbox": [ + 75, + 779, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ebf2ce38a48df3be5a2aedea02d2731cdfc99ad2e300a320cf7a8a6dd50a7f60.jpg", + "image_caption": [ + "Figure 7. Top-10 slot visualizations of all the compared methods on the evaluation set of TRI-PD and KITTI datasets. The samples on KITTI are cropped for better visualization. We discard the masks taking more than $20\\%$ of the whole pixels. Both the more powerful decoders (STEVE) and the simpler reconstruction space (SAVi and $\\mathrm{SAV_i + + }$ ) fail to work in realistic driving scenes. Our model produces the most accurate masks and captures more objects." + ], + "image_footnote": [], + "bbox": [ + 504, + 88, + 890, + 335 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "consistency in the token space can help.", + "bbox": [ + 498, + 465, + 764, + 481 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Better measuring object discovery in the real world. Measuring the performance of object discovery in the real world is challenging. As noted in [15], one slot may re-bind to another object after the tracked object moves out of the frame and a new object enters the scene. The FG. ARI score is not suitable for such scenarios but currently, there are no better metrics for object discovery which is able to tackle the re-binding issue. Developing a novel metric for this problem could have a significant impact on the community.", + "bbox": [ + 496, + 486, + 893, + 623 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 642, + 617, + 657 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work proposes MoTok, a unified architecture that leverages motion-guided tokenization for object discovery. By jointly training the slot representation with motion cues and vector quantization, our model enables the emergence of interpretable mid-level features which simplifies the problem of object discovery. Comprehensive evaluation on both synthetic and real-world benchmarks shows that with sufficient capacity of the slot decoder, motion guidance alleviates the need for labels, optical flow, or depth decoding, thanks to tokenization, achieving state-of-the-art results.", + "bbox": [ + 496, + 667, + 893, + 819 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. We thank Dian Chen, Alexei Efros, and Andrew Owens for their valuable comments. This research was supported by Toyota Research Institute. YXW was supported in part by NSF Grant 2106825, NIFA Award 2020-67021-32799, and the NCSA Fellows program.", + "bbox": [ + 496, + 824, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "22979", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Parallel domain. https://paralleldomain.com/, November 2021. 5", + "[2] Pablo Arbeláez, Jordi Pont-Tuset, Jonathan T Barron, Ferran Marques, and Jitendra Malik. Multiscale combinatorial grouping. In CVPR, 2014. 2", + "[3] Nicolas Ballas, Li Yao, Chris Pal, and Aaron Courville. Delving deeper into convolutional networks for learning video representations. In ICLR, 2016. 3", + "[4] Zhipeng Bao, Pavel Tokmakov, Allan Jabri, Yu-Xiong Wang, Adrien Gaidon, and Martial Hebert. Discovering objects that can move. In CVPR, 2022. 1, 2, 3, 5, 6, 7, 8", + "[5] Zhangxing Bian, Allan Jabri, Alexei A Efros, and Andrew Owens. Learning pixel trajectories with multiscale contrastive random walks. In CVPR, 2022. 8", + "[6] Christopher P Burgess, Loic Matthew, Nicholas Watters, Rishabh Kabra, Irina Higgins, Matt Botvinick, and Alexander Lerchner. MONet: Unsupervised scene decomposition and representation. arXiv preprint arXiv:1901.11390, 2019. 1, 2", + "[7] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 1, 2", + "[8] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features. In ECCV, 2018. 8", + "[9] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 5", + "[10] Chang Chen, Fei Deng, and Sungjin Ahn. Roots: Object-centric representation and rendering of 3D scenes. JMLR, 2021. 2", + "[11] Achal Dave, Pavel Tokmakov, and Deva Ramanan. Towards segmenting anything that moves. In ICCV Workshops, 2019. 1, 5, 6", + "[12] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 5", + "[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 2", + "[14] Yilun Du, Kevin Smith, Tomer Ulman, Joshua Tenenbaum, and Jiajun Wu. Unsupervised discovery of 3D physical objects from video. In ICLR, 2021. 2", + "[15] Gamaleldin F Elsayed, Aravindh Mahendran, Sjoerd van Steenkiste, Klaus Greff, Michael C Mozer, and Thomas Kipf. Savi++: Towards end-to-end object-centric learning from real-world videos. In NeurIPS, 2022. 1, 2, 5, 6, 8", + "[16] Martin Engelcke, Adam R Kosiorek, Oiwi Parker Jones, and Ingmar Posner. GENESIS: Generative scene inference and sampling with object-centric latent representations. In ICLR, 2020. 1, 2" + ], + "bbox": [ + 78, + 116, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[17] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In CVPR, 2021. 2", + "[18] Pedro F Felzenszwalb and Daniel P Huttenlocher. Efficient graph-based image segmentation. IJCV, 2004. 2", + "[19] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In CVPR, 2012. 5", + "[20] Robert Gray. Vector quantization. IEEE Assp Magazine, 1984. 2", + "[21] Klaus Greff, Francois Belletti, Lucas Beyer, Carl Doersch, Yilun Du, Daniel Duckworth, David J Fleet, Dan Gnanapragasam, Florian Golemo, Charles Herrmann, et al. Kubric: A scalable dataset generator. In CVPR, 2022. 5", + "[22] Klaus Greff, Raphaël Lopez Kaufman, Rishabh Kabra, Nick Watters, Christopher Burgess, Daniel Zoran, Loic Matthey, Matthew Botvinick, and Alexander Lerchner. Multi-object representation learning with iterative variational inference. In ICML, 2019. 1, 2", + "[23] Klaus Greff, Antti Rasmus, Mathias Berglund, Tele Hao, Harri Valpola, and Jürgen Schmidhuber. Tagger: Deep unsupervised perceptual grouping. In NeurIPS, 2016. 2", + "[24] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In CVPR, 2022. 2", + "[25] Rares Guizilini, Vitor abd Ambrus, Sudeep Pillai, Allan Raventos, and Adrien Gaidon. 3d packing for self-supervised monocular depth estimation. In CVPR, 2020. 5, 6", + "[26] Paul Henderson and Christoph H Lampert. Unsupervised object-centric video generation and decomposition in 3D. In NeurIPS, 2020. 2", + "[27] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation, 1997. 8", + "[28] Allan Jabri, Andrew Owens, and Alexei Efros. Space-time correspondence as a contrastive random walk. NeurIPS, 2020. 8", + "[29] Andrew Jaegle, Sebastian Borgeaud, Jean-Baptiste Alayrac, Carl Doersch, Catalin Ionescu, David Ding, Skanda Koppula, Daniel Zoran, Andrew Brock, Evan Shelhamer, et al. Perceiver io: A general architecture for structured inputs & outputs. In ICLR, 2022. 2, 4", + "[30] Andrew Jaegle, Felix Gimeno, Andy Brock, Oriol Vinyals, Andrew Zisserman, and Joao Carreira. Perceiver: General perception with iterative attention. In ICML, 2021. 2", + "[31] Jindong Jiang, Sepehr Janghorbani, Gerard De Melo, and Sungjin Ahn. SCALOR: Generative world models with scalable object representations. In ICLR, 2020. 2", + "[32] Daniel Kahneman, Anne Treisman, and Brian J Gibbs. The reviewing of object files: Object-specific integration of information. Cognitive psychology, 1992. 1", + "[33] Laurynas Karazija, Subhabrataand Choudhury, Iro Laina, Christian Rupprecht, and Andrea Vedaldi. Unsupervised multi-object segmentation by predicting probable motion patterns. In NeurIPS, 2022. 2, 5, 8", + "[34] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 5" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "22980", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[35] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. In ICLR, 2014. 2", + "[36] Thomas Kipf, Gamaleldin F Elsayed, Aravindh Mahendran, Austin Stone, Sara Sabour, Georg Heigold, Rico Jonschkowski, Alexey Dosovitskiy, and Klaus Greff. Conditional object-centric learning from video. In ICLR, 2022. 1, 2, 5", + "[37] Kurt Koffka. Principles of Gestalt psychology. Routledge, 2013. 2", + "[38] Zhixuan Lin, Yi-Fu Wu, Skand Vishwanath Peri, Weihao Sun, Gautam Singh, Fei Deng, Jindong Jiang, and Sungjin Ahn. Space: Unsupervised object-oriented scene representation via spatial attention and decomposition. In ICLR, 2020. 1, 2", + "[39] Francesco Locatello, Dirk Weissenborn, Thomas Unterthiner, Aravindh Mahendran, Georg Heigold, Jakob Uszkoreit, Alexey Dosovitskiy, and Thomas Kipf. Object-centric learning with slot attention. In NeurIPS, 2020. 1, 2, 5", + "[40] Nikolaus Mayer, Eddy Ilg, Philip Hausser, Philipp Fischer, Daniel Cremers, Alexey Dosovitskiy, and Thomas Brox. A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In CVPR, 2016. 5", + "[41] Tim Meinhardt, Alexander Kirillov, Laura Leal-Taixe, and Christoph Feichtenhofer. Trackformer: Multi-object tracking with transformers. In CVPR, 2022. 2", + "[42] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2", + "[43] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, 2021. 1, 2", + "[44] Ali Razavi, Aaron Van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. NeurIPS, 2019. 2", + "[45] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster R-CNN: Towards real-time object detection with region proposal networks. In NeurIPS, 2015. 1", + "[46] Danilo Jimenez Rezende, Shakir Mohamed, and Daan Wierstra. Stochastic backpropagation and approximate inference in deep generative models. In ICML, 2014. 2", + "[47] Jason Tyler Rolfe. Discrete variational autoencoders. In ICLR, 2017. 1, 2", + "[48] Mehdi SM Sajjadi, Daniel Duckworth, Aravindh Mahendran, Sjoerd van Steenkiste, Filip Pavetic, Mario Lucic, Leonidas J Guibas, Klaus Greff, and Thomas Kipf. Object scene representation transformer. In NeurIPS, 2022. 2", + "[49] Mehdi SM Sajjadi, Henning Meyer, Etienne Pot, Urs Bergmann, Klaus Greff, Noha Radwan, Suhani Vora, Mario Lucic, Daniel Duckworth, Alexey Dosovitskiy, et al. Scene representation transformer: Geometry-free novel view synthesis through set-latent scene representations. In CVPR, 2022. 2", + "[50] Hinrich Schütze, Christopher D Manning, and Prabhakar Raghavan. Introduction to information retrieval. Cambridge University Press Cambridge, 2008. 7", + "[51] Maximilian Seitzer, Max Horn, Andrii Zadaianchuk, Dominik Zietlow, Tianjun Xiao, Carl-Johann Simon-Gabriel, Tong He, Zheng Zhang, Bernhard Scholkopf, Thomas Brox, et al." + ], + "bbox": [ + 78, + 92, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Bridging the gap to real-world object-centric learning. In ICLR, 2023. 2, 5", + "[52] Gautam Singh, Fei Deng, and Sungjin Ahn. Illiterate dall-e learns to compose. In ICLR, 2021. 1, 2", + "[53] Gautam Singh, Yi-Fu Wu, and Sungjin Ahn. Simple unsupervised object-centric learning for complex and naturalistic videos. In NeurlPS, 2022. 1, 2, 5, 6, 7", + "[54] Elizabeth S Spelke and Katherine D Kinzler. Core knowledge. Developmental science, 2007. 1", + "[55] Karl Stelzner, Kristian Kersting, and Adam R Kosiorek. Decomposing 3D scenes into objects via unsupervised volume segmentation. arXiv preprint arXiv:2104.01148, 2021. 2", + "[56] Chen Sun, Austin Myers, Carl Vondrick, Kevin Murphy, and Cordelia Schmid. Videobert: A joint model for video and language representation learning. In CVPR, 2019. 2", + "[57] Zachary Teed and Jia Deng. RAFT: Recurrent all-pairs field transforms for optical flow. In ECCV, 2020. 5, 6", + "[58] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. NeurIPS, 2017. 2, 4", + "[59] Wouter Van Gansbeke, Simon Vandenhende, Stamatios Georgoulis, and Luc Van Gool. Unsupervised semantic segmentation by contrasting object mask proposals. In ICCV, 2021. 8", + "[60] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 2017. 1, 2", + "[61] Rishi Veerapaneni, John D Co-Reyes, Michael Chang, Michael Janner, Chelsea Finn, Jiajun Wu, Joshua Tenenbaum, and Sergey Levine. Entity abstraction in visual model-based reinforcement learning. In CoRL, 2020. 2", + "[62] Ziyi Wu, Nikita Dvornik, Klaus Greff, Thomas Kipf, and Animesh Garg. Slotformer: Unsupervised visual dynamics simulation with object-centric models. In ICLR, 2023. 2", + "[63] Wilson Yan, Yunzhi Zhang, Pieter Abbeel, and Aravind Srinivas. Videogpt: Video generation using vq-vae and transformers. arXiv preprint arXiv:2104.10157, 2021. 2", + "[64] Peiyu Yu, Sirui Xie, Xiaojian Ma, Yixin Zhu, Ying Nian Wu, and Song-Chun Zhu. Unsupervised foreground extraction via deep region competition. In NeurIPS, 2021. 2", + "[65] Jiahui Zhang, Dawei Sun, Zixin Luo, Anbang Yao, Lei Zhou, Tianwei Shen, Yurong Chen, Long Quan, and Honggen Liao. Learning two-view correspondences and geometry using order-aware network. In ICCV, 2019. 8", + "[66] Brady Zhou and Philipp Krahenbuhl. Cross-view transformers for real-time map-view semantic segmentation. In CVPR, 2022. 8", + "[67] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In ICLR, 2021. 2" + ], + "bbox": [ + 503, + 92, + 890, + 789 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "22981", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Object Discovery From Motion-Guided Tokens/cce05c92-f0c6-4fb5-8422-8cb2ccb76230_model.json b/2023/Object Discovery From Motion-Guided Tokens/cce05c92-f0c6-4fb5-8422-8cb2ccb76230_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8e85617a16f458cff8b12450a7f263c1e68c345c --- /dev/null +++ b/2023/Object Discovery From Motion-Guided Tokens/cce05c92-f0c6-4fb5-8422-8cb2ccb76230_model.json @@ -0,0 +1,2508 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.25, + 0.131, + 0.722, + 0.154 + ], + "angle": 0, + "content": "Object Discovery from Motion-Guided Tokens" + }, + { + "type": "text", + "bbox": [ + 0.242, + 0.18, + 0.726, + 0.234 + ], + "angle": 0, + "content": "Zhipeng Bao\\*,1 Pavel Tokmakov2 Yu-Xiong Wang3 Adrien Gaidon2 Martial Hebert1 CMU 2Toyota Research Institute 3UIUC" + }, + { + "type": "text", + "bbox": [ + 0.126, + 0.238, + 0.325, + 0.251 + ], + "angle": 0, + "content": "{zbao, hebert}@cs.cmu.edu" + }, + { + "type": "text", + "bbox": [ + 0.356, + 0.238, + 0.684, + 0.252 + ], + "angle": 0, + "content": "{pavel).(tokmakov, adrien).(gaidon}@tri.global" + }, + { + "type": "text", + "bbox": [ + 0.715, + 0.239, + 0.843, + 0.251 + ], + "angle": 0, + "content": "yxw@illinois.edu" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.287, + 0.314, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.319, + 0.474, + 0.577 + ], + "angle": 0, + "content": "Object discovery - separating objects from the background without manual labels - is a fundamental open challenge in computer vision. Previous methods struggle to go beyond clustering of low-level cues, whether handcrafted (e.g., color, texture) or learned (e.g., from auto-encoders). In this work, we augment the auto-encoder representation learning framework with two key components: motion-guidance and mid-level feature tokenization. Although both have been separately investigated, we introduce a new transformer decoder showing that their benefits can compound thanks to motion-guided vector quantization. We show that our architecture effectively leverages the synergy between motion and tokenization, improving upon the state of the art on both synthetic and real datasets. Our approach enables the emergence of interpretable object-specific mid-level features, demonstrating the benefits of motion-guidance (no labeling) and quantization (interpretability, memory efficiency)." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.592, + 0.21, + 0.608 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.618, + 0.471, + 0.799 + ], + "angle": 0, + "content": "Objects are central in human and computer vision. In the former, they are a fundamental primitive used to decompose the complexity of the visual world into an actionable representation. This abstraction in turn enables higher-level cognitive abilities, such as casual reasoning and planning [32, 54]. In computer vision, object detection has achieved remarkable progress [7, 45] and is now an essential component in many applications (e.g., driving, robotics). However, these models require a large amount of manual labels from a fixed vocabulary of categories. Consequently, learning unsupervised, object-centric representations is an important step towards scaling up computer vision to the real world." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.8, + 0.472, + 0.876 + ], + "angle": 0, + "content": "This topic has received renewed attention recently thanks to structured generative networks with iterative inference over a fixed set of variables [6, 16, 22, 38, 39]. These methods cluster pixels in the feature space of an auto-encoder, exhibiting behavior similar to grouping based on low-level cues," + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.288, + 0.695, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.547, + 0.363, + 0.647, + 0.372 + ], + "angle": 0, + "content": "Slot Representation" + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.288, + 0.889, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.736, + 0.363, + 0.847, + 0.373 + ], + "angle": 0, + "content": "Token Representation" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.377, + 0.893, + 0.432 + ], + "angle": 0, + "content": "Figure 1. TRI-PD dataset results: (left) top-10 foreground slot segments produced by our approach; (right) corresponding token representations. Compared to raw images, tokens in our framework present a more structured and compact space for reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.447, + 0.895, + 0.492 + ], + "angle": 0, + "content": "such as color or texture. Hence, they are restricted to toy images with colored geometric shapes on a plain background, and fail on more complex realistic scenes [4]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.493, + 0.896, + 0.81 + ], + "angle": 0, + "content": "Two main types of works attempt to address this shortcoming. The first family of methods sets out to simplify the grouping problem by introducing more structure into the output space, e.g., reconstructing optical flow [36] or depth [15]. They, however, require supervision, either in the form of known poses [36] or ground truth bounding boxes [15], veering away from the unsupervised goal of object discovery. In contrast, Bao et al. [4] resolve the object-background ambiguity by explicitly integrating an unsupervised motion segmentation algorithm [11] into the pipeline, showing substantial progress on realistic scenes. The second main direction to improve object discovery focuses on improving the decoder part of auto-encoding architectures [52, 53], replacing convolutional decoders with transformers [43, 60] combined with discrete variational auto-encoders (DVAE) [47] to reduce memory footprint. These more sophisticated architectures improve performance without additional supervision, including on real-world sequences. However, these methods are evaluated with different protocols (metrics, datasets) and therefore no clear architectural principles have emerged yet for unsupervised object discovery." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.897, + 0.901 + ], + "angle": 0, + "content": "In this work, we introduce a novel architecture, Motion-guided Tokens (MoTok), based on the combination of two fundamental structural principles: motion and discretization. We define objects as discrete entities that might have an independent motion. As prior works have shown encouraging results thanks to unsupervised motion guidance and better" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.309, + 0.9 + ], + "angle": 0, + "content": "*Work done during an internship at TRI" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22972" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.198 + ], + "angle": 0, + "content": "transformer-based decoders, we propose to leverage motion to guide tokenization, the vector quantization process at the heart of attention mechanisms in transformer architectures (See Figure 1). In addition, to comprehensively evaluate the contributions of prior works, we ablate key design choices proposed in the past, such as the decoder architecture and reconstruction space, in a unified framework." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.199, + 0.474, + 0.41 + ], + "angle": 0, + "content": "Our key contributions are as follows. (1) We introduce a novel auto-encoder architecture, MoTok, for unsupervised video object discovery with a new transformer decoder leveraging unsupervised motion-guided tokenization. (2) Our results on real and synthetic datasets show that with sufficient capacity in the decoder, motion guidance alleviates the need for labels, optical flow, or depth decoding thanks to tokenization, improving upon the state of the art. (3) We show that our motion-guided tokens map to interpretable mid-level features, going beyond typical clusters of low-level features, thus explaining why our method scales to challenging realistic videos. Our code, models, and synthetic data are made available at https://github.com/zpbao/MoTok/." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.42, + 0.22, + 0.436 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.446, + 0.471, + 0.506 + ], + "angle": 0, + "content": "In this work, we tackle the object discovery problem in realistic videos with object-centric representation by capitalizing on motion-guided tokens. We review the most relevant works in these areas below." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.509, + 0.471, + 0.614 + ], + "angle": 0, + "content": "Object Discovery tackles the problem of separating objects from background without manual labels [4]. Classic computer vision methods use appearance-based perceptual grouping to parse a scene into object-like regions [2, 18, 37]. Among them, [2] first propose a multiscale fast normalized cuts algorithm and achieved robust object segmentation performance for static images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.615, + 0.474, + 0.811 + ], + "angle": 0, + "content": "Recently, the topic of object discovery has experienced renewed attention under the name of unsupervised object-centric representation learning. A wide variety of learning-based methods has been introduced [6, 16, 22, 23, 31, 38, 39, 52, 61, 64], usually with an encoder-decoder architecture [35, 46]. These methods aim to learn compositional feature representations, e.g., a set of variables that can bind to objects in an image [6, 16, 22, 23, 39, 52, 61, 64], or a video [4, 15, 31, 33, 36, 38, 48, 51, 53, 62]. Among them, [39] first formulate the SlotAttention framework, which is used to bind a set of variables, called slots, to image locations. The slots are then decoded individually and combined to reconstruct the image." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.471, + 0.903 + ], + "angle": 0, + "content": "However, without additional constraints, such methods tend to converge to pixel grouping based on low-level cues, such as color, and do not generalize to realistic images or videos with complex backgrounds. To address this limitation, [36] and [15] extend the slot concept from static images to videos via reconstructing in the optical flow or depth space" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "respectively. The intuition behind these methods is that this space provides stronger cues to separate the objects from the background. Separately, Bao et al. [4] use motion cues to guide the slots to find moving objects and then generalize to all the objects that can move." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.172, + 0.895, + 0.295 + ], + "angle": 0, + "content": "In another line of work, Singh at al. [53] show that a combination of a more powerful transformer decoder [13] and discrete variational auto-encoder [47] can enhance the object discovery performance. Different from these works, we propose a unified architecture for object discovery that is flexible to different choices of decoders and reconstruction space. We also introduce the vector quantized features as an additional reconstruction space with motion guidance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.298, + 0.895, + 0.405 + ], + "angle": 0, + "content": "Finally, several recent works also leverage 3D geometry as inductive biases to enforce the learning-based models to focus on object-like regions [10, 14, 26, 55]. Though these methods remain limited to the toy, synthetic environments, the underlying geometric priors are orthogonal to our approach and have a great potential to be combined with our proposed method as a future direction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.41, + 0.897, + 0.699 + ], + "angle": 0, + "content": "Vector quantization is originally developed for data compression in signal processing [20]. More recently, [58] propose the Vector-Quantized Variational Autoencoder (VQ-VAE), which learns a discrete representation of images, and models their distribution aggressively. This technique has motivated a series of investigations in solving different computer vision tasks including image synthesis [17, 24, 44], video generation [56, 63], and language modeling [24, 42, 43], to name a few. In this work, we adopt the vector quantization technique as a mid-level representation in our object discovery framework. The intuition is that reconstructing in a structured, low-dimensional feature space rather than the high-variance color space should simplify the task of resolving the object and background ambiguity. In comparison, [53] also introduce a mid-level feature space for object discovery in videos. However, they use DVAE, not VQ-VAE, and are motivated primarily by improving training efficiency. Moreover, they did not explore combing vector quantization with motion cues in an end-to-end trainable framework." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.897, + 0.903 + ], + "angle": 0, + "content": "Transformers are originally proposed for sequence-to-sequence modeling in natural language processing [60]. They use a multi-head attention mechanism, instead of the recurrent memory units, to aggregate information from the input. Recently, vision transformer (ViT) [13] and its derivations have achieved state-of-the-art in several visual tasks [7,29,30,41,48,49,67]. Among them, Perceiver IO [29] design a computationally efficient architecture to handle arbitrary outputs in addition to arbitrary inputs, greatly enhancing the generalizability of the transformer architecture. We employ a variant of a Perceiver module as a decoder in our object discovery framework to combine high representational power with computational efficiency." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22973" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.093, + 0.891, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.255, + 0.894, + 0.298 + ], + "angle": 0, + "content": "Figure 2. Model architecture of the proposed Motion-guided Tokens (MoTok) framework. MoTok is a unified framework for video object discovery that is flexible with different choices of decoders and reconstruction spaces. Our framework effectively leverages the synergy between motion and tokenization, and enables the emergence of interpretable object-specific mid-level features." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.312, + 0.169, + 0.327 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.338, + 0.474, + 0.473 + ], + "angle": 0, + "content": "We now explain the proposed Motion-guided Tokens (Mo-Tok) framework in detail. Its architecture is shown in Figure 2. We first introduce a motion-guided slot learning framework [4] in Section 3.1. We then describe the slot decoder units in Section 3.2, which decode the slot features into a reconstruction space. In Section 3.3, we describe the choices of reconstruction space and explain the vector-quantized reconstruction space in detail. Finally, we demonstrate how we optimize the model in Section 3.4." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.484, + 0.449, + 0.501 + ], + "angle": 0, + "content": "3.1. Preliminary: Motion-Guided Slot Learning" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.508, + 0.471, + 0.643 + ], + "angle": 0, + "content": "Our object-centric representation learning module is derived from [4] (Figure 2, left). Concretely, given a sequence of video frames \\(\\{I^1, I^2, \\dots, I^T\\}\\), we first process each frame through an encoder CNN to obtain an individual frame representation \\(H^t = f_{\\mathrm{enc}}(I^t)\\). These individual representations are aggregated by a spatio-temporal Convolutional Gated Recurrent Unit (ConvGRU) [3] to obtain video encoding via \\(H'^t = \\mathrm{ConvGRU}(R^{t-1}, H^t)\\), where \\(R^{t-1} \\in \\mathbb{R}^{h' \\times w' \\times d_{\\mathrm{inp}}}\\) is the recurrent memory state." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.471, + 0.78 + ], + "angle": 0, + "content": "Next, we perform a single attention operation with \\(K\\) slots to directly compute the slot state \\(S^t = W^t v(H'^t)\\), where \\(W^t \\in \\mathbb{R}^{K \\times N}\\) is the attention matrix, \\(N = h' \\times w'\\) indicates the flapped shape dimension, and \\(v(\\cdot)\\) is the value embedding function. \\(W^t\\) is computed using the slot state in the previous frame \\(S^{t-1}\\) and the input feature \\(H'^t\\). For the first frame, we use a learnable initial state \\(S^0\\). At the same time, for each slot \\(s_i^t\\), we can also obtain the attention mask \\(W_{i,i}^t\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.471, + 0.902 + ], + "angle": 0, + "content": "In [4], a motion cue is also added to guide the slots to find moving objects. A set of sparse, instance-level motion segmentation masks \\(\\mathcal{M} = \\{M^1,M^2,\\dots,M^T\\}\\) is assumed to be provided with every video, with \\(M^{t} = \\{m_{1},m_{2},\\dots,m_{C^{t}}\\}\\), where \\(C^t\\) is the number of moving objects that were successfully segmented in frame \\(t\\), and \\(m_j\\in \\{0,1\\}^{h'\\times w'}\\) is a binary mask. The attention mask \\(W^{t}\\in \\mathbb{R}^{K\\times N}\\), is then supervised with the motion segments. \\(M^t\\) is also considered" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.313, + 0.892, + 0.344 + ], + "angle": 0, + "content": "as a set of length \\(K\\) padded with \\(\\emptyset\\) (no object), and a bipartite matching is found between them with the lowest cost:" + }, + { + "type": "equation", + "bbox": [ + 0.575, + 0.352, + 0.893, + 0.393 + ], + "angle": 0, + "content": "\\[\n\\hat {\\sigma} = \\underset {\\sigma} {\\arg \\min } \\sum_ {i = 1} ^ {K} \\mathcal {L} _ {\\operatorname {s e g}} \\left(m _ {i}, W _ {:, \\sigma (i)} ^ {t}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.402, + 0.892, + 0.464 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{\\mathrm{seg}}(m_i,W_{:, \\sigma(i)}^t)\\) is the segmentation loss between the motion mask \\(m_i\\) and the attention map of the slot with index \\(\\sigma(i)\\). Once the assignment \\(\\hat{\\sigma}\\) has been computed, the final motion supervision objective is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.554, + 0.472, + 0.892, + 0.512 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {m o t i o n}} = \\sum_ {i = 1} ^ {K} \\mathbb {1} _ {\\{m _ {i} \\neq \\emptyset \\}} \\mathcal {L} _ {\\text {s e g}} \\left(m _ {i}, W _ {:, \\hat {\\sigma} (i)} ^ {t}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.52, + 0.892, + 0.551 + ], + "angle": 0, + "content": "where \\(\\mathbb{1}_{\\{m_i \\neq \\emptyset\\}}\\) denotes that the loss is only computed for the matched slots and \\(\\mathcal{L}_{\\mathrm{seg}}\\) is the binary cross entropy." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.559, + 0.642, + 0.573 + ], + "angle": 0, + "content": "3.2. Slot Decoders" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.582, + 0.894, + 0.641 + ], + "angle": 0, + "content": "The goal of the Slot Decoder is to map the slot representation \\((S^t, W^t)\\) to a 2D feature map \\(F^t\\) for the reconstruction space. As shown in Figure 2 (middle), we propose four choices for the slot decoder." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.642, + 0.892, + 0.673 + ], + "angle": 0, + "content": "Linear Decoder directly maps the slot features \\( S^t \\) to their corresponding positions based on the attention mask \\( W^t \\):" + }, + { + "type": "equation", + "bbox": [ + 0.588, + 0.68, + 0.892, + 0.722 + ], + "angle": 0, + "content": "\\[\nF _ {\\text {l i n e a r}} ^ {t} (\\mathbf {x}) = \\frac {\\sum_ {i = 1} ^ {K} S _ {i} ^ {t} (\\mathbf {x}) W _ {i , \\mathbf {x}} ^ {t}}{\\sum_ {i = 1} ^ {K} W _ {i , \\mathbf {x}} ^ {t}}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.73, + 0.738, + 0.744 + ], + "angle": 0, + "content": "where \\(\\mathbf{x}\\) is an arbitrary 2D position." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.744, + 0.892, + 0.774 + ], + "angle": 0, + "content": "CNN Decoder further adds two convolutional layers to the 2D feature map formed by Equation 3:" + }, + { + "type": "equation", + "bbox": [ + 0.584, + 0.781, + 0.892, + 0.824 + ], + "angle": 0, + "content": "\\[\nF _ {\\mathrm {C N N}} ^ {t} = \\operatorname {C N N} \\left(\\frac {\\sum_ {i = 1} ^ {K} S _ {i} ^ {t} W _ {i , :} ^ {t}}{\\sum_ {i = 1} ^ {K} W _ {i , :} ^ {t}}\\right). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.831, + 0.892, + 0.874 + ], + "angle": 0, + "content": "Transformer Decoder decodes the feature by querying the slot representation with a 2D positional embedding through a transformer decoder:" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.885, + 0.892, + 0.903 + ], + "angle": 0, + "content": "\\[\nF _ {\\text {t r a n s f o r m e r}} ^ {t} = \\operatorname {T r a n s f o r m e r} \\left(P, S ^ {t}, S ^ {t}\\right), \\tag {5}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22974" + } + ], + [ + { + "type": "code_caption", + "bbox": [ + 0.088, + 0.096, + 0.303, + 0.109 + ], + "angle": 0, + "content": "Algorithm 1: Perceiver decoder" + }, + { + "type": "algorithm", + "bbox": [ + 0.08, + 0.112, + 0.47, + 0.224 + ], + "angle": 0, + "content": "Perceiver \\((S^t,P)\\) .. \\(S^t\\to \\mathrm{Norm}(S^t)\\) \n\\(\\hat{S}^t =\\) SelfAttention \\((S^t) + S^t\\) \n\\(\\hat{S}^t\\rightarrow \\mathrm{Norm}(\\hat{S}^t)\\) \n\\(\\tilde{S}^t = \\mathrm{MLP}(\\hat{S}^t) + \\hat{S}^t\\) \n\\(\\tilde{S}^t\\rightarrow \\mathrm{Norm}(\\tilde{S}^t)\\quad P\\rightarrow \\mathrm{Norm}(P)\\) \n\\(F_{\\mathrm{Perceiver}}^{t} =\\) CrossAttention \\((P,\\tilde{S}^{t},\\tilde{S}^{t})\\)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.254, + 0.471, + 0.284 + ], + "angle": 0, + "content": "where the query \\(P\\in \\mathbb{R}^{N\\times d_p}\\) is a learnable positional embedding." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.286, + 0.471, + 0.436 + ], + "angle": 0, + "content": "Compared with the previous two linear decoders, the transformer decoder further considers the global connections between the slot features and the input query, so that it will form a more powerful feature map. However, an obvious limitation is that the transformer decoder applies self-attention to the input positional query, which is 1) redundant since the positional embedding itself is learnable, and 2) not computationally efficient therefore limits the scalability of the whole model. To solve this limitation, we further propose the perceiver decoder." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.438, + 0.47, + 0.543 + ], + "angle": 0, + "content": "Perceiver Decoder is inspired by [29], which designs a computationally efficient architecture to handle arbitrary outputs in addition to arbitrary inputs. Different from the original architecture, we only add a self-attention layer to the slot representations, followed by a cross-attention layer for the output and positional embedding queries. The whole procedure is illustrated in Algorithm 1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.544, + 0.471, + 0.604 + ], + "angle": 0, + "content": "After forming the 2D feature map \\( F^t \\), we decode it to a reconstruction space with a CNN-based decoder, except for when vector-quantized space is used. As we will show next, an additional decoder is redundant in this case." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.617, + 0.281, + 0.634 + ], + "angle": 0, + "content": "3.3. Reconstruction Space" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.641, + 0.473, + 0.837 + ], + "angle": 0, + "content": "There are four choices of reconstruction space in our framework. The plain RGB space contains the most information but also is the most complex to resolve the object/background ambiguity. The flow and depth spaces, as shown in Figure 2 (right), are more structured. Reconstructing in these spaces makes the grouping problem easier. However, there are still shortcomings, e.g. non-moving objects are not captured in the flow space and depth cannot easily distinguish between-objects that are near each other. Moreover, these two spaces are not as informative as the RGB space. In addition, we introduce the VQ-space, which is end-to-end trainable and is both structured and informative. We describe the VQ-space below." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Vector-quantized reconstruction space. Different from the other three reconstruction spaces, here we do not predict the reconstruction directly, but rather supervise the feature map \\( F^{t} \\) to match the latent embedding space of VQ-VAE." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.198 + ], + "angle": 0, + "content": "In particular, following [58], we define a latent embedding space \\(E\\) as the set of \\(M\\) vectors \\(e_i\\) of dimension \\(d_{vq}\\), \\(E = \\{e_i\\in \\mathbb{R}^{d_{vq}}|i = 1,2,\\dots ,M\\}\\). Given an input image \\(I^t\\), VQ-VAE first processes it with an encoder to get the output \\(z_e^t = \\mathrm{Encoder}_{\\mathrm{VQ}}(I^t)\\), and then the discrete latent variables \\(z\\) are calculated by a nearest neighborhood search among the discrete feature set \\(E\\):" + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.207, + 0.894, + 0.227 + ], + "angle": 0, + "content": "\\[\nz _ {q} ^ {t} (\\mathbf {x}) = e _ {k}, \\quad \\text {w h e r e} k = \\operatorname {a r g m i n} _ {j} \\| z _ {e} ^ {t} (\\mathbf {x}) - e _ {j} \\| _ {2}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.235, + 0.894, + 0.28 + ], + "angle": 0, + "content": "and \\(\\mathbf{x}\\) is an arbitrary 2D position. The reconstructed image \\(\\hat{I}^t\\) is then decoded by \\(\\hat{I}^t = \\mathrm{Decoder}_{\\mathrm{VQ}}(z_q^t)\\). The objective of VQ-VAE is:" + }, + { + "type": "equation", + "bbox": [ + 0.507, + 0.288, + 0.893, + 0.306 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {V Q V A E}} = \\log P \\left(I ^ {t} \\mid z _ {q} ^ {t}\\right) + \\left\\| s g \\left[ z _ {e} ^ {t} \\right] - z _ {q} ^ {t} \\right\\| _ {2} + \\left\\| s g \\left[ z _ {q} ^ {t} \\right] - z _ {e} ^ {t} \\right\\| _ {2}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.313, + 0.779, + 0.328 + ], + "angle": 0, + "content": "where \\( sg[\\cdot] \\) is the stop-gradient operation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.328, + 0.894, + 0.374 + ], + "angle": 0, + "content": "Then we use the quantized feature map \\( z_{q}^{t} \\) as the target signal for the slot feature map \\( F^{t} \\). The final objective of VQ-VAE and the VQ reconstruction is:" + }, + { + "type": "equation", + "bbox": [ + 0.508, + 0.383, + 0.893, + 0.403 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {V Q}} = \\mathcal {L} _ {\\mathrm {V Q V A E}} + \\| s g [ F ^ {t} ] - z _ {q} ^ {t} \\| _ {2} + \\| s g [ z _ {q} ^ {t} ] - F ^ {t} \\| _ {2}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.414, + 0.895, + 0.566 + ], + "angle": 0, + "content": "Motion-guided token representation. The last term in Equation 8, \\( ||sg[F^t] - z_q^t||_2 \\), enables the motion signal from slot learning to jointly optimize the token space through the output of the slot decoder. Furthermore, the token representation and the motion cues build a connection linked by the slot learning, thus enabling the emergence of interpretable object-specific mid-level features of tokens. In addition, reconstructing in a more compact token space also benefits the model by better utilizing the motion signal to achieve an improved slot representation and temporal consistency." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.574, + 0.64, + 0.59 + ], + "angle": 0, + "content": "3.4. Optimization" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.6, + 0.894, + 0.689 + ], + "angle": 0, + "content": "Token contrastive constraint. The goal of reconstructing in the VQ-space is that it is more compact and of lower variation compared with the RGB space. To make the VQ-space more structured, we add an optional contrastive constraint below to the vector space, which increases the independence between latent vectors:" + }, + { + "type": "equation", + "bbox": [ + 0.564, + 0.7, + 0.893, + 0.718 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {c o n t r a s t i v e}} = \\left| \\left| \\mathbb {I} - \\operatorname {s o f t m a x} \\left(E \\cdot E ^ {T}\\right) \\right| \\right|, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.727, + 0.892, + 0.757 + ], + "angle": 0, + "content": "where \\(\\mathbb{I}\\) is the identity matrix and \\(E\\in \\mathbb{R}^{N\\times d_{vq}}\\) is the matrix of the feature embedding space \\(S\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.758, + 0.894, + 0.802 + ], + "angle": 0, + "content": "The final loss function is a combination of the reconstruction objective, the motion objective, and the optional contrastive constraint:" + }, + { + "type": "equation", + "bbox": [ + 0.53, + 0.814, + 0.893, + 0.831 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\lambda \\mathcal {L} _ {\\text {m o t i o n}} + \\mathcal {L} _ {\\text {r e c o n}} + \\lambda_ {c} \\mathbb {1} _ {\\mathrm {V Q}} \\mathcal {L} _ {\\text {c o n t r a s t i v e}}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.895, + 0.901 + ], + "angle": 0, + "content": "where \\(\\lambda\\) and \\(\\lambda_{c}\\) are weighting factors and \\(\\mathbb{1}_{\\mathrm{VQ}}\\) is an indicator function. For the reconstruction loss, we set \\(\\mathcal{L}_{\\mathrm{recon}} = \\mathcal{L}_{\\mathrm{VQ}}\\) when performing reconstruction in the VQ-space. Otherwise, we use an \\(L_{2}\\) loss for reconstruction in the other three spaces." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22975" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.091, + 0.203, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.212, + 0.092, + 0.33, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.339, + 0.092, + 0.465, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.195, + 0.471, + 0.266 + ], + "angle": 0, + "content": "Figure 3. Frame samples from the video datasets used in our experiments. MOVi [21] (left) is a multi-object video dataset created by simulating rigid body dynamics. TRI-PD [4] (middle) is a collection of photo-realistic, synthetic driving videos. KITTI [19] (right) is a real-world benchmark with city driving scenes." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.282, + 0.312, + 0.299 + ], + "angle": 0, + "content": "4. Experimental Evaluation" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.308, + 0.284, + 0.325 + ], + "angle": 0, + "content": "4.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.335, + 0.47, + 0.365 + ], + "angle": 0, + "content": "Benchmarks. We evaluate our approach on three popular benchmarks with different complexity (Figure 3)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.367, + 0.47, + 0.501 + ], + "angle": 0, + "content": "MOVi [21] is a synthetic multi-object video dataset, created by simulating rigid body dynamics. We use this benchmark to ablate and analyze our model architecture. Following previous works [15, 53], we use the most complex subset, MOVi-E, for evaluation. MOVi-E contains both moving and static objects (maximum 20 objects) with linear random camera motion. The resolution of this dataset is \\(128 \\times 128\\) and each video contains 24 frames with a sample rate of 12 FPS. We use standard train and test split for MOVi-E." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.504, + 0.47, + 0.638 + ], + "angle": 0, + "content": "TRI-PD [4,25] is a synthetic dataset based on street driving scenarios collected by using a state-of-the-art synthetic data generation service [1]. The dataset comes with a collection of accurate annotations (e.g. for flow or depth), which we use to ablate the impact of additional information and its quality on various models. Each video in TRI-PD is 10 seconds captured at 20 FPS. There are 924 videos in the training set and 51 videos in the test set. We crop and resize each frame to the resolution of \\(480 \\times 968\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.641, + 0.47, + 0.701 + ], + "angle": 0, + "content": "KITTI [19] is a real-world benchmark with city driving scenes. We use the whole 151 KITTI videos for training and the instance segmentation subsets with 200 single-frame images for evaluation. The frames are resized to \\(368 \\times 1248\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Baselines. We compare our methods against the most recent learning-based object discovery models. In particular, SAVi [36] and SAVi++ [15] are two direct extensions of the original slot attention [39] using optical flow and depth to facilitate object-centric learning. Alternatively, STEVE [53] uses a more powerful transformer decoder to enhance the object discovery performance. Bao et al. [4], Karazija et al. [33] utilize motion cues to guide object discovery. In addition, in the supplementary, we compare our approach to [51], which is a concurrent work that relies on ImageNet [12] pretraining [9]. For [4, 15, 36, 53], we use either the official code or public implementation (see supplementary for details). Due to the lack of implementation, we reuse reported results" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.089, + 0.892, + 0.206 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.208, + 0.892, + 0.264 + ], + "angle": 0, + "content": "Figure 4. Object discovery in different reconstruction spaces on the evaluation set of MOVi-E. Compared with the other reconstruction spaces, the model trained using the VQ-space demonstrates a more accurate performance and better temporal consistency." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.298, + 0.797, + 0.313 + ], + "angle": 0, + "content": "with the two very recent approaches [33, 51]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.323, + 0.892, + 0.429 + ], + "angle": 0, + "content": "Evaluation Metrics. Following prior works, we use the Foreground Adjusted Rand Index (FG. ARI) to evaluate the performance of the models, which captures how well the predicted segmentation masks match ground-truth masks in a permutation-invariant fashion. Notice that FG. ARI is measured in terms of the whole video, so that the temporal consistency of the slots is considered with this metric." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.439, + 0.893, + 0.726 + ], + "angle": 0, + "content": "Implementation Details. We use the same ResNet-18 ConvGRU encoder backbone for all the compared methods following [4]. We set the number of slots as 24 for MOVi-E and 45 for PD and KITTI based on the maximal number of objects in these datasets. On TRI-PD and KITTI, we further downsample the images by 4 for SAVi and SAVi++ to fit the GPU memory. Notice that, the other methods also produce downsampled slot masks, leading to the evaluation under the same resolution for all the methods. All the models are trained for 500 epochs using Adam [34]. During training, we train all the models with a randomly sampled video sequence of a fixed length (6, 5, 5 for MOVi-E, TRI-PD, and KITTI respectively). During the inference time, all the models are evaluated frame by frame until the end. For the FG. ARI measurement on TRI-PD, we discard any instance labels covering an area of \\(0.5\\%\\) or less of the first sampled video frame following [15]. We also evaluate FG. ARI at 5 FPS due to memory limitation. We use batch size 64 for MOVi-E, and 8 for TRI-PD and KITTI." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Our method and Bao et al. [4] require motion segmentation signals. We apply [11], a powerful method pre-trained on the toy FlyingThings3D dataset [40], taking the ground-truth flow (MOVi-E) or RAFT [57] flow as the input. For SAVi and \\(\\mathrm{SAV_i + + }\\), we do not use the first-frame bounding box supervision for a fair comparison. We generate the optical flow and depth annotation for KITTI using two state-of-the-art methods, RAFT [57] and VIDAR [25]. We also use them for ablations on TRI-PD. More details about the hyper-parameters, training scheme, and annotation generation are provided in the supplementary." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "22976" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.119, + 0.089, + 0.426, + 0.22 + ], + "angle": 0, + "content": "
MotionSpaceDecoderFG. ARI
XVQLinear14.6
XVQLinear-CNN16.3
XVQTransformer40.1
XVQPerceiver52.4
XRGBPerceiver49.2
XFlowPerceiver36.3
XDepthPerceiver18.4
XFlow + DepthPerceiver38.0
XVQ (flow)Perceiver44.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.226, + 0.472, + 0.31 + ], + "angle": 0, + "content": "Table 1. Model architecture analysis with different choices of the slot decoder and reconstruction space on MOVi-E. Perceiver decoder + VQ reconstruction space yields the best performance, indicating that (1) the capacity of the decoder plays a key role in object discovery; (2) learnable, vector-quantized reconstruction space outperforms fixed alternatives like depth or flow." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.337, + 0.321, + 0.354 + ], + "angle": 0, + "content": "4.2. Object Discovery on MOVi" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.362, + 0.47, + 0.422 + ], + "angle": 0, + "content": "We use MOVi-E dataset to ablate our model with different choices of slot decoder, and reconstruction spaces, aiming to elucidate their roles in object discovery. We also compare the performance with state-of-the-art models in this section." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.426, + 0.471, + 0.622 + ], + "angle": 0, + "content": "Set up. We first fix the reconstruction space as the VQ-space and ablate the decoder architectures (first 4 lines in Table 1), and then we fix the decoder architecture with the best component and reconstruct in different reconstruction space (Line 5 to 9 in Table 1). Since SAVi++ [15] argues that reconstructing in the combined space of flow and depth is the key to their success, we add a variant \\( Flow + Depth \\), which also reconstructs in this combined space. Furthermore, as VQ-VAE can be trained to reconstruct the image in the flow space as well, we also report a \\( VQ \\) (flow) variant for completeness. We do not include motion cues for the architecture analysis. A more comprehensive ablation is reported in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.626, + 0.47, + 0.777 + ], + "angle": 0, + "content": "Decoder analysis. By comparing the results on the first four lines in Table 1, similar to [53], we find that the capacity of the decoder indeed plays a key role in the object discovery performance. A simple linear decoder fails, but adding an additional CNN decoder can bring limited performance gains. Introducing powerful transformer-based decoders, on the other hand, greatly improves the object discovery capabilities of the model. Between the two transformer variants, the more advanced perceiver decoder achieves better results, while being more computationally efficient." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Reconstruction space analysis. For the four reconstruction spaces, the VQ-space yields the best object discovery performance. We analyze the learned slots and token representations in Figure 4 and make the following discoveries. Firstly, the model trained using the VQ-space better separates the objects from the background and shows stronger temporal consistency. This is due to the more structured, compact, and lower-variance reconstruction space provided" + }, + { + "type": "table", + "bbox": [ + 0.589, + 0.089, + 0.8, + 0.216 + ], + "angle": 0, + "content": "
ModelFG. ARI
Bao et al.51.6
STEVE51.2
SAVi28.1
SAVi++31.7
MoTok (no motion)52.4
MoTok63.8
SAVi (G.T. box)53.4
SAVi++ (G.T. box)84.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.222, + 0.895, + 0.278 + ], + "angle": 0, + "content": "Table 2. Comparison to the state-of-the-art object discovery approaches on the validation sets of MOVi-E using FG. ARI. Our approach outperforms all the recent methods with the help of motion cues and vector quantization." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.291, + 0.895, + 0.32 + ], + "angle": 0, + "content": "by the quantized features, which simplifies the task of grouping, compared to the raw RGB space." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.321, + 0.895, + 0.486 + ], + "angle": 0, + "content": "Secondly, by comparing Flow and \\( \\text{Flow+Depth} \\) variants, we find that the latter indeed carries more information, allowing the model to better group the objects, which is consistent with [15]. However, interestingly, we also find that given a sufficiently strong decoder, reconstructing in depth or flow space does not bring further improvements compared to RGB, and can even decrease the performance. Finally, comparing Flow and \\( \\text{VQ}(\\text{flow}) \\), we find that although adding the quantization improves performance significantly, this variant still lags behind reconstructing in the raw RGB space, reinforcing our previous conclusions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.489, + 0.895, + 0.685 + ], + "angle": 0, + "content": "Comparison to the state of the art. In Table 2, we compare our approach to the state-of-the-art methods. Our model achieves the best performance among them. Additionally, compared to SAVi and SAVi++, even our variants that reconstructs in the same space outperform them, indicating that the perceiver decoder and single-stage decoding strategy are the optimal choices. Our model also outperforms STEVE, due to the fact that, in contrast to DVAE, the codebook in VQ-VAE can be jointly optimized with slot learning (see Equation 8). Compared with Bao et al., our model shows better performance even without motion cues used by that method, but further incorporating them into our approach allows it to achieve top results." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.686, + 0.894, + 0.761 + ], + "angle": 0, + "content": "Notice that, we compare with the variants of SAVi and SAVi++ that do not rely on ground-truth bounding boxes at test time. Their published numbers are included at the bottom of the table for reference, but they are not comparable to the other unsupervised methods." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.769, + 0.892, + 0.786 + ], + "angle": 0, + "content": "4.3. Object Discovery with Realistic Driving Videos" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Set up. Firstly, following [4], we capitalize on the ground-truth annotations available in TRI-PD, and evaluate the importance of the quality of external signals for various approaches. In particular, we separately evaluate using ground-truth (GT) flow, depth, and motion segmentation, as well as estimating those using state-of-the-art algorithms [11,25,57]. In addition, we report a variant STEVE-\\(m\\) for which we add" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "22977" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.089, + 0.891, + 0.229 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.23, + 0.893, + 0.273 + ], + "angle": 0, + "content": "Figure 5. Reconstruction, slot representation, and token masks for [4], STEVE-m [53], and our method. We use ground-truth motion cues for supervision, and visualize top-10 masks excluding background slots. MoTok effectively leverages the synergy between motion and tokenization, enabling the emergence of interpretable object-specific mid-level features, which simplify the problem of object discovery." + }, + { + "type": "table", + "bbox": [ + 0.119, + 0.28, + 0.426, + 0.378 + ], + "angle": 0, + "content": "
ModelSignalGPU Mem. (GB)FG. ARI
GTEST
Bao et al.Motion21.350.434.7
SAViFlow17.317.914.2
SAVi++Depth + Flow17.518.415.3
STEVE-66.712.5-
STEVE-mMotion68.245.732.2
MoTokMotion10.960.655.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.383, + 0.471, + 0.453 + ], + "angle": 0, + "content": "Table 3. Object discovery evaluation of different models on TRI-PD dataset. GT denotes the ground-truth guidance given to each model and EST denotes the estimated cues. State-of-the-art models fail to work without motion cues; introducing token representation helps our model better utilize the motion signal." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.484, + 0.47, + 0.514 + ], + "angle": 0, + "content": "the same motion segmentation supervision as to [4] and our approach for a fair comparison." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.517, + 0.47, + 0.652 + ], + "angle": 0, + "content": "From Table 3, we can observe that our method strongly outperforms all the baselines. Additionally, we find that even with the ground-truth flow or depth, SAVI and SAVI++ do not work well for realistic videos with complex background and crowded scenes. Without motion cues, even equipped with a powerful transformer decoder, STEVE fails to work in this challenging setting as well. Finally, our model achieves better performance compared to Bao et al. with both GT and estimated motion segments, indicating its better robustness." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.657, + 0.471, + 0.807 + ], + "angle": 0, + "content": "Impact of motion-guided tokens: In Figure 5, we visualize the RGB reconstruction, slot representation, and token masks (if possible) for Bao et al., STEVE-m, and our MoTok. There are several key observations. Firstly, compared with STEVE-m, MoTok enables the emergence of more interpretable mid-level features, even though both models are trained with motion cues and tokenization. This is due to the interplay between object and slot discovery in our end-to-end framework, whereas STEVE-m trains DVAE separately from the rest of the model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.471, + 0.9 + ], + "angle": 0, + "content": "Secondly, our model achieves the best RGB reconstruction result, even for the latter frames. The high quality of the reconstruction indicates that the model can better take advantage of the appearance signal to optimize the slot representation. Finally, our model demonstrates stronger temporal consistency compared with the baselines, thanks to the struc" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.281, + 0.888, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.342, + 0.893, + 0.384 + ], + "angle": 0, + "content": "Figure 6. Visualization of the learned tokens for vanilla VQ-VAE and MoTok. The proposed MoTok model shows a better alignment with semantic categories." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.407, + 0.893, + 0.482 + ], + "angle": 0, + "content": "tured reconstruction space. To sum up, our architecture effectively leverages the synergy between motion cues and tokenization, enabling the emergence of interpretable mid-level features, which greatly simplifies the task of object discovery." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.486, + 0.893, + 0.592 + ], + "angle": 0, + "content": "Interpretable tokens. To better illustrate the interpretability of motion-guided tokens, we measured their alignment with the ground-truth semantic labels using cluster purity [50]. Vanilla VQ-VAE (no motion guidance) achieves a purity of 40.1, whereas ours reaches 65.7 (chance performance: 38.5). Qualitative comparison in Figure 6 clearly demonstrates our better semantic alignment." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.596, + 0.893, + 0.716 + ], + "angle": 0, + "content": "Scalability to realistic videos: We additionally report the GPU memory usage per batch in Table 3. The per-slot decoding strategy and the use of transformer decoder limit the scalability of the other baselines. In comparison, our model achieves the best performance with the lowest memory consumption, showing a great generalization capability to real-world videos, thanks to the efficient perceiver decoder and the single-shot decoding strategy." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Impact of objectives. Finally, we ablate the objective design of our model with the GT annotations in Table 4. We build three variants of our model, one without the additional token contrastive constraint defined by Equation 10 (no contrastive), the second without the motion cues (no motion), and the last trained with ground-truth instance masks for all objects to indicate the performance upper bound \\((^{*})\\). By comparing the performance of these variants, we make the following observations: (1) in realistic synthetic videos even the strongest architectures fail to resolve the object/background ambiguity in the absence of motion cues, demonstrating the difficulty of object discovery tasks in the real world; (2)" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22978" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.255, + 0.142 + ], + "angle": 0, + "content": "
ModelFG. ARI
MoTok60.6
MoTok (no contrastive)57.4
MoTok (no motion)15.6
MoTok*64.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.147, + 0.258, + 0.286 + ], + "angle": 0, + "content": "Table 4. Ablation study on different learning signals. MoTok* is our model trained with ground-truth instance segmentation. The motion cue and the contrastive constraint help improve the capability of the proposed model, almost reaching the performance upper bound (MoTok*)." + }, + { + "type": "table", + "bbox": [ + 0.272, + 0.089, + 0.468, + 0.18 + ], + "angle": 0, + "content": "
ModelFG. ARI
Bao et al. [4]47.1
SAVi20.0
SAVi++23.9
STEVE11.9
Karazija et al. [33]50.8
Karazija et al. [33] (WL)51.9
MoTok64.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.266, + 0.185, + 0.472, + 0.283 + ], + "angle": 0, + "content": "Table 5. Evaluation of object discovery in the real-world KITTI dataset. Karazija et al. [33] (WL) is a variant trained with their proposed wrapping loss. Our model achieves state-of-the-art in real-world object discovery." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.313, + 0.473, + 0.389 + ], + "angle": 0, + "content": "contrastive constraint facilitates diversity during vector quantization, increasing the information content of each token, which helps reduce the object/background ambiguity; (3) the performance of our model is close to the upper bound, indicating the effectiveness of the motion-guided tokens." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.398, + 0.386, + 0.415 + ], + "angle": 0, + "content": "4.4. Object Discovery in the Real World" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.424, + 0.47, + 0.484 + ], + "angle": 0, + "content": "Set up. To train the object discovery model on the real-world KITTI benchmark, we first pre-train all the models on PD using ground-truth flow and depth and then fine-tune them using estimated annotations on KITTI." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.485, + 0.473, + 0.651 + ], + "angle": 0, + "content": "The comparisons on KITTI are shown in Table 5 and Figure 7. We find that, consistent with our observations on TRI-PD, without motion guidance, both more powerful decoders (STEVE) and the simpler reconstruction space (SAVi and \\(\\mathrm{SAV_i + + }\\)) fail to work. However, our model still captures good dynamic object segmentation results. Compared to Bao et al., we still achieve a better FG. ARI score and a more solid segmentation mask, indicating the benefit of the model design and the motion-guided tokens. Finally, we outperform the very recent approach of Karazija et al. [33] even when they include an additional warping loss objective." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.661, + 0.342, + 0.675 + ], + "angle": 0, + "content": "4.5. Limitations and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.686, + 0.47, + 0.777 + ], + "angle": 0, + "content": "Object/background ambiguity in the real world. Even after successfully parsing the world into object and background slots, separating those from each other in an unsupervised way is an open challenge. Promising directions include temporal contrastive learning [5, 28], unsupervised clustering [8, 59], and utilizing geometric cues [65, 66]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Slot drift. From Figure 5, we notice that the final RGB reconstruction quality decreases with longer videos, which indicates that the distribution of the slot representation has shifted in the latter frames. This issue leads to degraded temporal consistency and ambiguous object masks. Better training schemes and model improvements, such as adding data augmentation during training [15], replacing GRU units with long-short memory units [27], and enforcing temporal" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.089, + 0.891, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.349, + 0.895, + 0.448 + ], + "angle": 0, + "content": "Figure 7. Top-10 slot visualizations of all the compared methods on the evaluation set of TRI-PD and KITTI datasets. The samples on KITTI are cropped for better visualization. We discard the masks taking more than \\(20\\%\\) of the whole pixels. Both the more powerful decoders (STEVE) and the simpler reconstruction space (SAVi and \\(\\mathrm{SAV_i + + }\\)) fail to work in realistic driving scenes. Our model produces the most accurate masks and captures more objects." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.467, + 0.766, + 0.482 + ], + "angle": 0, + "content": "consistency in the token space can help." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.487, + 0.895, + 0.624 + ], + "angle": 0, + "content": "Better measuring object discovery in the real world. Measuring the performance of object discovery in the real world is challenging. As noted in [15], one slot may re-bind to another object after the tracked object moves out of the frame and a new object enters the scene. The FG. ARI score is not suitable for such scenarios but currently, there are no better metrics for object discovery which is able to tackle the re-binding issue. Developing a novel metric for this problem could have a significant impact on the community." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.643, + 0.619, + 0.658 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.669, + 0.895, + 0.82 + ], + "angle": 0, + "content": "This work proposes MoTok, a unified architecture that leverages motion-guided tokenization for object discovery. By jointly training the slot representation with motion cues and vector quantization, our model enables the emergence of interpretable mid-level features which simplifies the problem of object discovery. Comprehensive evaluation on both synthetic and real-world benchmarks shows that with sufficient capacity of the slot decoder, motion guidance alleviates the need for labels, optical flow, or depth decoding, thanks to tokenization, achieving state-of-the-art results." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Acknowledgements. We thank Dian Chen, Alexei Efros, and Andrew Owens for their valuable comments. This research was supported by Toyota Research Institute. YXW was supported in part by NSF Grant 2106825, NIFA Award 2020-67021-32799, and the NCSA Fellows program." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22979" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.117, + 0.47, + 0.143 + ], + "angle": 0, + "content": "[1] Parallel domain. https://paralleldomain.com/, November 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.147, + 0.472, + 0.188 + ], + "angle": 0, + "content": "[2] Pablo Arbeláez, Jordi Pont-Tuset, Jonathan T Barron, Ferran Marques, and Jitendra Malik. Multiscale combinatorial grouping. In CVPR, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.19, + 0.472, + 0.233 + ], + "angle": 0, + "content": "[3] Nicolas Ballas, Li Yao, Chris Pal, and Aaron Courville. Delving deeper into convolutional networks for learning video representations. In ICLR, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.235, + 0.471, + 0.276 + ], + "angle": 0, + "content": "[4] Zhipeng Bao, Pavel Tokmakov, Allan Jabri, Yu-Xiong Wang, Adrien Gaidon, and Martial Hebert. Discovering objects that can move. In CVPR, 2022. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.279, + 0.471, + 0.32 + ], + "angle": 0, + "content": "[5] Zhangxing Bian, Allan Jabri, Alexei A Efros, and Andrew Owens. Learning pixel trajectories with multiscale contrastive random walks. In CVPR, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.323, + 0.471, + 0.378 + ], + "angle": 0, + "content": "[6] Christopher P Burgess, Loic Matthew, Nicholas Watters, Rishabh Kabra, Irina Higgins, Matt Botvinick, and Alexander Lerchner. MONet: Unsupervised scene decomposition and representation. arXiv preprint arXiv:1901.11390, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.381, + 0.471, + 0.435 + ], + "angle": 0, + "content": "[7] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.438, + 0.471, + 0.479 + ], + "angle": 0, + "content": "[8] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features. In ECCV, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.482, + 0.471, + 0.536 + ], + "angle": 0, + "content": "[9] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.54, + 0.471, + 0.581 + ], + "angle": 0, + "content": "[10] Chang Chen, Fei Deng, and Sungjin Ahn. Roots: Object-centric representation and rendering of 3D scenes. JMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.584, + 0.471, + 0.625 + ], + "angle": 0, + "content": "[11] Achal Dave, Pavel Tokmakov, and Deva Ramanan. Towards segmenting anything that moves. In ICCV Workshops, 2019. 1, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.628, + 0.471, + 0.668 + ], + "angle": 0, + "content": "[12] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.672, + 0.471, + 0.741 + ], + "angle": 0, + "content": "[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.744, + 0.471, + 0.785 + ], + "angle": 0, + "content": "[14] Yilun Du, Kevin Smith, Tomer Ulman, Joshua Tenenbaum, and Jiajun Wu. Unsupervised discovery of 3D physical objects from video. In ICLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.788, + 0.471, + 0.843 + ], + "angle": 0, + "content": "[15] Gamaleldin F Elsayed, Aravindh Mahendran, Sjoerd van Steenkiste, Klaus Greff, Michael C Mozer, and Thomas Kipf. Savi++: Towards end-to-end object-centric learning from real-world videos. In NeurIPS, 2022. 1, 2, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.471, + 0.9 + ], + "angle": 0, + "content": "[16] Martin Engelcke, Adam R Kosiorek, Oiwi Parker Jones, and Ingmar Posner. GENESIS: Generative scene inference and sampling with object-centric latent representations. In ICLR, 2020. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.117, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.134 + ], + "angle": 0, + "content": "[17] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.893, + 0.164 + ], + "angle": 0, + "content": "[18] Pedro F Felzenszwalb and Daniel P Huttenlocher. Efficient graph-based image segmentation. IJCV, 2004. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.165, + 0.893, + 0.205 + ], + "angle": 0, + "content": "[19] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In CVPR, 2012. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.208, + 0.894, + 0.234 + ], + "angle": 0, + "content": "[20] Robert Gray. Vector quantization. IEEE Assp Magazine, 1984. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.236, + 0.894, + 0.29 + ], + "angle": 0, + "content": "[21] Klaus Greff, Francois Belletti, Lucas Beyer, Carl Doersch, Yilun Du, Daniel Duckworth, David J Fleet, Dan Gnanapragasam, Florian Golemo, Charles Herrmann, et al. Kubric: A scalable dataset generator. In CVPR, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.292, + 0.893, + 0.36 + ], + "angle": 0, + "content": "[22] Klaus Greff, Raphaël Lopez Kaufman, Rishabh Kabra, Nick Watters, Christopher Burgess, Daniel Zoran, Loic Matthey, Matthew Botvinick, and Alexander Lerchner. Multi-object representation learning with iterative variational inference. In ICML, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.363, + 0.893, + 0.404 + ], + "angle": 0, + "content": "[23] Klaus Greff, Antti Rasmus, Mathias Berglund, Tele Hao, Harri Valpola, and Jürgen Schmidhuber. Tagger: Deep unsupervised perceptual grouping. In NeurIPS, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.406, + 0.893, + 0.459 + ], + "angle": 0, + "content": "[24] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.462, + 0.893, + 0.503 + ], + "angle": 0, + "content": "[25] Rares Guizilini, Vitor abd Ambrus, Sudeep Pillai, Allan Raventos, and Adrien Gaidon. 3d packing for self-supervised monocular depth estimation. In CVPR, 2020. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.505, + 0.893, + 0.545 + ], + "angle": 0, + "content": "[26] Paul Henderson and Christoph H Lampert. Unsupervised object-centric video generation and decomposition in 3D. In NeurIPS, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.547, + 0.893, + 0.574 + ], + "angle": 0, + "content": "[27] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation, 1997. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.576, + 0.893, + 0.615 + ], + "angle": 0, + "content": "[28] Allan Jabri, Andrew Owens, and Alexei Efros. Space-time correspondence as a contrastive random walk. NeurIPS, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.618, + 0.893, + 0.687 + ], + "angle": 0, + "content": "[29] Andrew Jaegle, Sebastian Borgeaud, Jean-Baptiste Alayrac, Carl Doersch, Catalin Ionescu, David Ding, Skanda Koppula, Daniel Zoran, Andrew Brock, Evan Shelhamer, et al. Perceiver io: A general architecture for structured inputs & outputs. In ICLR, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.689, + 0.893, + 0.731 + ], + "angle": 0, + "content": "[30] Andrew Jaegle, Felix Gimeno, Andy Brock, Oriol Vinyals, Andrew Zisserman, and Joao Carreira. Perceiver: General perception with iterative attention. In ICML, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.732, + 0.893, + 0.773 + ], + "angle": 0, + "content": "[31] Jindong Jiang, Sepehr Janghorbani, Gerard De Melo, and Sungjin Ahn. SCALOR: Generative world models with scalable object representations. In ICLR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.775, + 0.893, + 0.815 + ], + "angle": 0, + "content": "[32] Daniel Kahneman, Anne Treisman, and Brian J Gibbs. The reviewing of object files: Object-specific integration of information. Cognitive psychology, 1992. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.817, + 0.893, + 0.871 + ], + "angle": 0, + "content": "[33] Laurynas Karazija, Subhabrataand Choudhury, Iro Laina, Christian Rupprecht, and Andrea Vedaldi. Unsupervised multi-object segmentation by predicting probable motion patterns. In NeurIPS, 2022. 2, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.893, + 0.901 + ], + "angle": 0, + "content": "[34] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 5" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "22980" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.093, + 0.47, + 0.119 + ], + "angle": 0, + "content": "[35] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. In ICLR, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.472, + 0.177 + ], + "angle": 0, + "content": "[36] Thomas Kipf, Gamaleldin F Elsayed, Aravindh Mahendran, Austin Stone, Sara Sabour, Georg Heigold, Rico Jonschkowski, Alexey Dosovitskiy, and Klaus Greff. Conditional object-centric learning from video. In ICLR, 2022. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.471, + 0.204 + ], + "angle": 0, + "content": "[37] Kurt Koffka. Principles of Gestalt psychology. Routledge, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.471, + 0.262 + ], + "angle": 0, + "content": "[38] Zhixuan Lin, Yi-Fu Wu, Skand Vishwanath Peri, Weihao Sun, Gautam Singh, Fei Deng, Jindong Jiang, and Sungjin Ahn. Space: Unsupervised object-oriented scene representation via spatial attention and decomposition. In ICLR, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.264, + 0.471, + 0.319 + ], + "angle": 0, + "content": "[39] Francesco Locatello, Dirk Weissenborn, Thomas Unterthiner, Aravindh Mahendran, Georg Heigold, Jakob Uszkoreit, Alexey Dosovitskiy, and Thomas Kipf. Object-centric learning with slot attention. In NeurIPS, 2020. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.321, + 0.471, + 0.375 + ], + "angle": 0, + "content": "[40] Nikolaus Mayer, Eddy Ilg, Philip Hausser, Philipp Fischer, Daniel Cremers, Alexey Dosovitskiy, and Thomas Brox. A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In CVPR, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.377, + 0.469, + 0.417 + ], + "angle": 0, + "content": "[41] Tim Meinhardt, Alexander Kirillov, Laura Leal-Taixe, and Christoph Feichtenhofer. Trackformer: Multi-object tracking with transformers. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.42, + 0.471, + 0.473 + ], + "angle": 0, + "content": "[42] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.476, + 0.471, + 0.517 + ], + "angle": 0, + "content": "[43] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.519, + 0.471, + 0.559 + ], + "angle": 0, + "content": "[44] Ali Razavi, Aaron Van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. NeurIPS, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.562, + 0.471, + 0.603 + ], + "angle": 0, + "content": "[45] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster R-CNN: Towards real-time object detection with region proposal networks. In NeurIPS, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.605, + 0.471, + 0.645 + ], + "angle": 0, + "content": "[46] Danilo Jimenez Rezende, Shakir Mohamed, and Daan Wierstra. Stochastic backpropagation and approximate inference in deep generative models. In ICML, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.647, + 0.471, + 0.673 + ], + "angle": 0, + "content": "[47] Jason Tyler Rolfe. Discrete variational autoencoders. In ICLR, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.676, + 0.471, + 0.731 + ], + "angle": 0, + "content": "[48] Mehdi SM Sajjadi, Daniel Duckworth, Aravindh Mahendran, Sjoerd van Steenkiste, Filip Pavetic, Mario Lucic, Leonidas J Guibas, Klaus Greff, and Thomas Kipf. Object scene representation transformer. In NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.733, + 0.471, + 0.813 + ], + "angle": 0, + "content": "[49] Mehdi SM Sajjadi, Henning Meyer, Etienne Pot, Urs Bergmann, Klaus Greff, Noha Radwan, Suhani Vora, Mario Lucic, Daniel Duckworth, Alexey Dosovitskiy, et al. Scene representation transformer: Geometry-free novel view synthesis through set-latent scene representations. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.471, + 0.858 + ], + "angle": 0, + "content": "[50] Hinrich Schütze, Christopher D Manning, and Prabhakar Raghavan. Introduction to information retrieval. Cambridge University Press Cambridge, 2008. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[51] Maximilian Seitzer, Max Horn, Andrii Zadaianchuk, Dominik Zietlow, Tianjun Xiao, Carl-Johann Simon-Gabriel, Tong He, Zheng Zhang, Bernhard Scholkopf, Thomas Brox, et al." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.119 + ], + "angle": 0, + "content": "Bridging the gap to real-world object-centric learning. In ICLR, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.149 + ], + "angle": 0, + "content": "[52] Gautam Singh, Fei Deng, and Sungjin Ahn. Illiterate dall-e learns to compose. In ICLR, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.151, + 0.892, + 0.191 + ], + "angle": 0, + "content": "[53] Gautam Singh, Yi-Fu Wu, and Sungjin Ahn. Simple unsupervised object-centric learning for complex and naturalistic videos. In NeurlPS, 2022. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.193, + 0.892, + 0.221 + ], + "angle": 0, + "content": "[54] Elizabeth S Spelke and Katherine D Kinzler. Core knowledge. Developmental science, 2007. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.223, + 0.892, + 0.263 + ], + "angle": 0, + "content": "[55] Karl Stelzner, Kristian Kersting, and Adam R Kosiorek. Decomposing 3D scenes into objects via unsupervised volume segmentation. arXiv preprint arXiv:2104.01148, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.265, + 0.892, + 0.306 + ], + "angle": 0, + "content": "[56] Chen Sun, Austin Myers, Carl Vondrick, Kevin Murphy, and Cordelia Schmid. Videobert: A joint model for video and language representation learning. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.308, + 0.892, + 0.335 + ], + "angle": 0, + "content": "[57] Zachary Teed and Jia Deng. RAFT: Recurrent all-pairs field transforms for optical flow. In ECCV, 2020. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.337, + 0.892, + 0.364 + ], + "angle": 0, + "content": "[58] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. NeurIPS, 2017. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.366, + 0.892, + 0.42 + ], + "angle": 0, + "content": "[59] Wouter Van Gansbeke, Simon Vandenhende, Stamatios Georgoulis, and Luc Van Gool. Unsupervised semantic segmentation by contrasting object mask proposals. In ICCV, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.422, + 0.892, + 0.464 + ], + "angle": 0, + "content": "[60] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.466, + 0.892, + 0.52 + ], + "angle": 0, + "content": "[61] Rishi Veerapaneni, John D Co-Reyes, Michael Chang, Michael Janner, Chelsea Finn, Jiajun Wu, Joshua Tenenbaum, and Sergey Levine. Entity abstraction in visual model-based reinforcement learning. In CoRL, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.522, + 0.892, + 0.563 + ], + "angle": 0, + "content": "[62] Ziyi Wu, Nikita Dvornik, Klaus Greff, Thomas Kipf, and Animesh Garg. Slotformer: Unsupervised visual dynamics simulation with object-centric models. In ICLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.565, + 0.892, + 0.605 + ], + "angle": 0, + "content": "[63] Wilson Yan, Yunzhi Zhang, Pieter Abbeel, and Aravind Srinivas. Videogpt: Video generation using vq-vae and transformers. arXiv preprint arXiv:2104.10157, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.607, + 0.892, + 0.648 + ], + "angle": 0, + "content": "[64] Peiyu Yu, Sirui Xie, Xiaojian Ma, Yixin Zhu, Ying Nian Wu, and Song-Chun Zhu. Unsupervised foreground extraction via deep region competition. In NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.65, + 0.892, + 0.705 + ], + "angle": 0, + "content": "[65] Jiahui Zhang, Dawei Sun, Zixin Luo, Anbang Yao, Lei Zhou, Tianwei Shen, Yurong Chen, Long Quan, and Honggen Liao. Learning two-view correspondences and geometry using order-aware network. In ICCV, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.707, + 0.892, + 0.747 + ], + "angle": 0, + "content": "[66] Brady Zhou and Philipp Krahenbuhl. Cross-view transformers for real-time map-view semantic segmentation. In CVPR, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.749, + 0.892, + 0.79 + ], + "angle": 0, + "content": "[67] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In ICLR, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.79 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22981" + } + ] +] \ No newline at end of file diff --git a/2023/Object Discovery From Motion-Guided Tokens/cce05c92-f0c6-4fb5-8422-8cb2ccb76230_origin.pdf b/2023/Object Discovery From Motion-Guided Tokens/cce05c92-f0c6-4fb5-8422-8cb2ccb76230_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f58555aae2e3c8987f217fd2c6b26ad977e45e81 --- /dev/null +++ b/2023/Object Discovery From Motion-Guided Tokens/cce05c92-f0c6-4fb5-8422-8cb2ccb76230_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:52e6e1120753d636b0a5484e1c997b3e2560a2aaf158ddb91419c102c05cc2c6 +size 5232036 diff --git a/2023/Object Discovery From Motion-Guided Tokens/full.md b/2023/Object Discovery From Motion-Guided Tokens/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5215d0bb19fc0f65d02618b89b38658722867744 --- /dev/null +++ b/2023/Object Discovery From Motion-Guided Tokens/full.md @@ -0,0 +1,363 @@ +# Object Discovery from Motion-Guided Tokens + +Zhipeng Bao\*,1 Pavel Tokmakov2 Yu-Xiong Wang3 Adrien Gaidon2 Martial Hebert1 CMU 2Toyota Research Institute 3UIUC + +{zbao, hebert}@cs.cmu.edu + +{pavel).(tokmakov, adrien).(gaidon}@tri.global + +yxw@illinois.edu + +# Abstract + +Object discovery - separating objects from the background without manual labels - is a fundamental open challenge in computer vision. Previous methods struggle to go beyond clustering of low-level cues, whether handcrafted (e.g., color, texture) or learned (e.g., from auto-encoders). In this work, we augment the auto-encoder representation learning framework with two key components: motion-guidance and mid-level feature tokenization. Although both have been separately investigated, we introduce a new transformer decoder showing that their benefits can compound thanks to motion-guided vector quantization. We show that our architecture effectively leverages the synergy between motion and tokenization, improving upon the state of the art on both synthetic and real datasets. Our approach enables the emergence of interpretable object-specific mid-level features, demonstrating the benefits of motion-guidance (no labeling) and quantization (interpretability, memory efficiency). + +# 1. Introduction + +Objects are central in human and computer vision. In the former, they are a fundamental primitive used to decompose the complexity of the visual world into an actionable representation. This abstraction in turn enables higher-level cognitive abilities, such as casual reasoning and planning [32, 54]. In computer vision, object detection has achieved remarkable progress [7, 45] and is now an essential component in many applications (e.g., driving, robotics). However, these models require a large amount of manual labels from a fixed vocabulary of categories. Consequently, learning unsupervised, object-centric representations is an important step towards scaling up computer vision to the real world. + +This topic has received renewed attention recently thanks to structured generative networks with iterative inference over a fixed set of variables [6, 16, 22, 38, 39]. These methods cluster pixels in the feature space of an auto-encoder, exhibiting behavior similar to grouping based on low-level cues, + +![](images/bd50f31a004d67ebf426562ab8a6f0cdc614f4ea730ecdde0bf8fb70e446e90c.jpg) +Slot Representation + +![](images/00bb9122d6004180f5f3a421aff54e947c80f7e98bbf1f829bbf45e00202e82f.jpg) +Token Representation +Figure 1. TRI-PD dataset results: (left) top-10 foreground slot segments produced by our approach; (right) corresponding token representations. Compared to raw images, tokens in our framework present a more structured and compact space for reconstruction. + +such as color or texture. Hence, they are restricted to toy images with colored geometric shapes on a plain background, and fail on more complex realistic scenes [4]. + +Two main types of works attempt to address this shortcoming. The first family of methods sets out to simplify the grouping problem by introducing more structure into the output space, e.g., reconstructing optical flow [36] or depth [15]. They, however, require supervision, either in the form of known poses [36] or ground truth bounding boxes [15], veering away from the unsupervised goal of object discovery. In contrast, Bao et al. [4] resolve the object-background ambiguity by explicitly integrating an unsupervised motion segmentation algorithm [11] into the pipeline, showing substantial progress on realistic scenes. The second main direction to improve object discovery focuses on improving the decoder part of auto-encoding architectures [52, 53], replacing convolutional decoders with transformers [43, 60] combined with discrete variational auto-encoders (DVAE) [47] to reduce memory footprint. These more sophisticated architectures improve performance without additional supervision, including on real-world sequences. However, these methods are evaluated with different protocols (metrics, datasets) and therefore no clear architectural principles have emerged yet for unsupervised object discovery. + +In this work, we introduce a novel architecture, Motion-guided Tokens (MoTok), based on the combination of two fundamental structural principles: motion and discretization. We define objects as discrete entities that might have an independent motion. As prior works have shown encouraging results thanks to unsupervised motion guidance and better + +transformer-based decoders, we propose to leverage motion to guide tokenization, the vector quantization process at the heart of attention mechanisms in transformer architectures (See Figure 1). In addition, to comprehensively evaluate the contributions of prior works, we ablate key design choices proposed in the past, such as the decoder architecture and reconstruction space, in a unified framework. + +Our key contributions are as follows. (1) We introduce a novel auto-encoder architecture, MoTok, for unsupervised video object discovery with a new transformer decoder leveraging unsupervised motion-guided tokenization. (2) Our results on real and synthetic datasets show that with sufficient capacity in the decoder, motion guidance alleviates the need for labels, optical flow, or depth decoding thanks to tokenization, improving upon the state of the art. (3) We show that our motion-guided tokens map to interpretable mid-level features, going beyond typical clusters of low-level features, thus explaining why our method scales to challenging realistic videos. Our code, models, and synthetic data are made available at https://github.com/zpbao/MoTok/. + +# 2. Related Work + +In this work, we tackle the object discovery problem in realistic videos with object-centric representation by capitalizing on motion-guided tokens. We review the most relevant works in these areas below. + +Object Discovery tackles the problem of separating objects from background without manual labels [4]. Classic computer vision methods use appearance-based perceptual grouping to parse a scene into object-like regions [2, 18, 37]. Among them, [2] first propose a multiscale fast normalized cuts algorithm and achieved robust object segmentation performance for static images. + +Recently, the topic of object discovery has experienced renewed attention under the name of unsupervised object-centric representation learning. A wide variety of learning-based methods has been introduced [6, 16, 22, 23, 31, 38, 39, 52, 61, 64], usually with an encoder-decoder architecture [35, 46]. These methods aim to learn compositional feature representations, e.g., a set of variables that can bind to objects in an image [6, 16, 22, 23, 39, 52, 61, 64], or a video [4, 15, 31, 33, 36, 38, 48, 51, 53, 62]. Among them, [39] first formulate the SlotAttention framework, which is used to bind a set of variables, called slots, to image locations. The slots are then decoded individually and combined to reconstruct the image. + +However, without additional constraints, such methods tend to converge to pixel grouping based on low-level cues, such as color, and do not generalize to realistic images or videos with complex backgrounds. To address this limitation, [36] and [15] extend the slot concept from static images to videos via reconstructing in the optical flow or depth space + +respectively. The intuition behind these methods is that this space provides stronger cues to separate the objects from the background. Separately, Bao et al. [4] use motion cues to guide the slots to find moving objects and then generalize to all the objects that can move. + +In another line of work, Singh at al. [53] show that a combination of a more powerful transformer decoder [13] and discrete variational auto-encoder [47] can enhance the object discovery performance. Different from these works, we propose a unified architecture for object discovery that is flexible to different choices of decoders and reconstruction space. We also introduce the vector quantized features as an additional reconstruction space with motion guidance. + +Finally, several recent works also leverage 3D geometry as inductive biases to enforce the learning-based models to focus on object-like regions [10, 14, 26, 55]. Though these methods remain limited to the toy, synthetic environments, the underlying geometric priors are orthogonal to our approach and have a great potential to be combined with our proposed method as a future direction. + +Vector quantization is originally developed for data compression in signal processing [20]. More recently, [58] propose the Vector-Quantized Variational Autoencoder (VQ-VAE), which learns a discrete representation of images, and models their distribution aggressively. This technique has motivated a series of investigations in solving different computer vision tasks including image synthesis [17, 24, 44], video generation [56, 63], and language modeling [24, 42, 43], to name a few. In this work, we adopt the vector quantization technique as a mid-level representation in our object discovery framework. The intuition is that reconstructing in a structured, low-dimensional feature space rather than the high-variance color space should simplify the task of resolving the object and background ambiguity. In comparison, [53] also introduce a mid-level feature space for object discovery in videos. However, they use DVAE, not VQ-VAE, and are motivated primarily by improving training efficiency. Moreover, they did not explore combing vector quantization with motion cues in an end-to-end trainable framework. + +Transformers are originally proposed for sequence-to-sequence modeling in natural language processing [60]. They use a multi-head attention mechanism, instead of the recurrent memory units, to aggregate information from the input. Recently, vision transformer (ViT) [13] and its derivations have achieved state-of-the-art in several visual tasks [7,29,30,41,48,49,67]. Among them, Perceiver IO [29] design a computationally efficient architecture to handle arbitrary outputs in addition to arbitrary inputs, greatly enhancing the generalizability of the transformer architecture. We employ a variant of a Perceiver module as a decoder in our object discovery framework to combine high representational power with computational efficiency. + +![](images/2bf1afd0f9ca6580390c5f700aebeeff549fe9f4019eac6dc97d1852ab074670.jpg) +Figure 2. Model architecture of the proposed Motion-guided Tokens (MoTok) framework. MoTok is a unified framework for video object discovery that is flexible with different choices of decoders and reconstruction spaces. Our framework effectively leverages the synergy between motion and tokenization, and enables the emergence of interpretable object-specific mid-level features. + +# 3. Method + +We now explain the proposed Motion-guided Tokens (Mo-Tok) framework in detail. Its architecture is shown in Figure 2. We first introduce a motion-guided slot learning framework [4] in Section 3.1. We then describe the slot decoder units in Section 3.2, which decode the slot features into a reconstruction space. In Section 3.3, we describe the choices of reconstruction space and explain the vector-quantized reconstruction space in detail. Finally, we demonstrate how we optimize the model in Section 3.4. + +# 3.1. Preliminary: Motion-Guided Slot Learning + +Our object-centric representation learning module is derived from [4] (Figure 2, left). Concretely, given a sequence of video frames $\{I^1, I^2, \dots, I^T\}$ , we first process each frame through an encoder CNN to obtain an individual frame representation $H^t = f_{\mathrm{enc}}(I^t)$ . These individual representations are aggregated by a spatio-temporal Convolutional Gated Recurrent Unit (ConvGRU) [3] to obtain video encoding via $H'^t = \mathrm{ConvGRU}(R^{t-1}, H^t)$ , where $R^{t-1} \in \mathbb{R}^{h' \times w' \times d_{\mathrm{inp}}}$ is the recurrent memory state. + +Next, we perform a single attention operation with $K$ slots to directly compute the slot state $S^t = W^t v(H'^t)$ , where $W^t \in \mathbb{R}^{K \times N}$ is the attention matrix, $N = h' \times w'$ indicates the flapped shape dimension, and $v(\cdot)$ is the value embedding function. $W^t$ is computed using the slot state in the previous frame $S^{t-1}$ and the input feature $H'^t$ . For the first frame, we use a learnable initial state $S^0$ . At the same time, for each slot $s_i^t$ , we can also obtain the attention mask $W_{i,i}^t$ . + +In [4], a motion cue is also added to guide the slots to find moving objects. A set of sparse, instance-level motion segmentation masks $\mathcal{M} = \{M^1,M^2,\dots,M^T\}$ is assumed to be provided with every video, with $M^{t} = \{m_{1},m_{2},\dots,m_{C^{t}}\}$ , where $C^t$ is the number of moving objects that were successfully segmented in frame $t$ , and $m_j\in \{0,1\}^{h'\times w'}$ is a binary mask. The attention mask $W^{t}\in \mathbb{R}^{K\times N}$ , is then supervised with the motion segments. $M^t$ is also considered + +as a set of length $K$ padded with $\emptyset$ (no object), and a bipartite matching is found between them with the lowest cost: + +$$ +\hat {\sigma} = \underset {\sigma} {\arg \min } \sum_ {i = 1} ^ {K} \mathcal {L} _ {\operatorname {s e g}} \left(m _ {i}, W _ {:, \sigma (i)} ^ {t}\right), \tag {1} +$$ + +where $\mathcal{L}_{\mathrm{seg}}(m_i,W_{:, \sigma(i)}^t)$ is the segmentation loss between the motion mask $m_i$ and the attention map of the slot with index $\sigma(i)$ . Once the assignment $\hat{\sigma}$ has been computed, the final motion supervision objective is defined as follows: + +$$ +\mathcal {L} _ {\text {m o t i o n}} = \sum_ {i = 1} ^ {K} \mathbb {1} _ {\{m _ {i} \neq \emptyset \}} \mathcal {L} _ {\text {s e g}} \left(m _ {i}, W _ {:, \hat {\sigma} (i)} ^ {t}\right), \tag {2} +$$ + +where $\mathbb{1}_{\{m_i \neq \emptyset\}}$ denotes that the loss is only computed for the matched slots and $\mathcal{L}_{\mathrm{seg}}$ is the binary cross entropy. + +# 3.2. Slot Decoders + +The goal of the Slot Decoder is to map the slot representation $(S^t, W^t)$ to a 2D feature map $F^t$ for the reconstruction space. As shown in Figure 2 (middle), we propose four choices for the slot decoder. + +Linear Decoder directly maps the slot features $S^t$ to their corresponding positions based on the attention mask $W^t$ : + +$$ +F _ {\text {l i n e a r}} ^ {t} (\mathbf {x}) = \frac {\sum_ {i = 1} ^ {K} S _ {i} ^ {t} (\mathbf {x}) W _ {i , \mathbf {x}} ^ {t}}{\sum_ {i = 1} ^ {K} W _ {i , \mathbf {x}} ^ {t}}, \tag {3} +$$ + +where $\mathbf{x}$ is an arbitrary 2D position. + +CNN Decoder further adds two convolutional layers to the 2D feature map formed by Equation 3: + +$$ +F _ {\mathrm {C N N}} ^ {t} = \operatorname {C N N} \left(\frac {\sum_ {i = 1} ^ {K} S _ {i} ^ {t} W _ {i , :} ^ {t}}{\sum_ {i = 1} ^ {K} W _ {i , :} ^ {t}}\right). \tag {4} +$$ + +Transformer Decoder decodes the feature by querying the slot representation with a 2D positional embedding through a transformer decoder: + +$$ +F _ {\text {t r a n s f o r m e r}} ^ {t} = \operatorname {T r a n s f o r m e r} \left(P, S ^ {t}, S ^ {t}\right), \tag {5} +$$ + +Algorithm 1: Perceiver decoder +Perceiver $(S^t,P)$ .. $S^t\to \mathrm{Norm}(S^t)$ $\hat{S}^t =$ SelfAttention $(S^t) + S^t$ $\hat{S}^t\rightarrow \mathrm{Norm}(\hat{S}^t)$ $\tilde{S}^t = \mathrm{MLP}(\hat{S}^t) + \hat{S}^t$ $\tilde{S}^t\rightarrow \mathrm{Norm}(\tilde{S}^t)\quad P\rightarrow \mathrm{Norm}(P)$ $F_{\mathrm{Perceiver}}^{t} =$ CrossAttention $(P,\tilde{S}^{t},\tilde{S}^{t})$ + +where the query $P\in \mathbb{R}^{N\times d_p}$ is a learnable positional embedding. + +Compared with the previous two linear decoders, the transformer decoder further considers the global connections between the slot features and the input query, so that it will form a more powerful feature map. However, an obvious limitation is that the transformer decoder applies self-attention to the input positional query, which is 1) redundant since the positional embedding itself is learnable, and 2) not computationally efficient therefore limits the scalability of the whole model. To solve this limitation, we further propose the perceiver decoder. + +Perceiver Decoder is inspired by [29], which designs a computationally efficient architecture to handle arbitrary outputs in addition to arbitrary inputs. Different from the original architecture, we only add a self-attention layer to the slot representations, followed by a cross-attention layer for the output and positional embedding queries. The whole procedure is illustrated in Algorithm 1. + +After forming the 2D feature map $F^t$ , we decode it to a reconstruction space with a CNN-based decoder, except for when vector-quantized space is used. As we will show next, an additional decoder is redundant in this case. + +# 3.3. Reconstruction Space + +There are four choices of reconstruction space in our framework. The plain RGB space contains the most information but also is the most complex to resolve the object/background ambiguity. The flow and depth spaces, as shown in Figure 2 (right), are more structured. Reconstructing in these spaces makes the grouping problem easier. However, there are still shortcomings, e.g. non-moving objects are not captured in the flow space and depth cannot easily distinguish between-objects that are near each other. Moreover, these two spaces are not as informative as the RGB space. In addition, we introduce the VQ-space, which is end-to-end trainable and is both structured and informative. We describe the VQ-space below. + +Vector-quantized reconstruction space. Different from the other three reconstruction spaces, here we do not predict the reconstruction directly, but rather supervise the feature map $F^{t}$ to match the latent embedding space of VQ-VAE. + +In particular, following [58], we define a latent embedding space $E$ as the set of $M$ vectors $e_i$ of dimension $d_{vq}$ , $E = \{e_i\in \mathbb{R}^{d_{vq}}|i = 1,2,\dots ,M\}$ . Given an input image $I^t$ , VQ-VAE first processes it with an encoder to get the output $z_e^t = \mathrm{Encoder}_{\mathrm{VQ}}(I^t)$ , and then the discrete latent variables $z$ are calculated by a nearest neighborhood search among the discrete feature set $E$ : + +$$ +z _ {q} ^ {t} (\mathbf {x}) = e _ {k}, \quad \text {w h e r e} k = \operatorname {a r g m i n} _ {j} \| z _ {e} ^ {t} (\mathbf {x}) - e _ {j} \| _ {2}, \tag {6} +$$ + +and $\mathbf{x}$ is an arbitrary 2D position. The reconstructed image $\hat{I}^t$ is then decoded by $\hat{I}^t = \mathrm{Decoder}_{\mathrm{VQ}}(z_q^t)$ . The objective of VQ-VAE is: + +$$ +\mathcal {L} _ {\mathrm {V Q V A E}} = \log P \left(I ^ {t} \mid z _ {q} ^ {t}\right) + \left\| s g \left[ z _ {e} ^ {t} \right] - z _ {q} ^ {t} \right\| _ {2} + \left\| s g \left[ z _ {q} ^ {t} \right] - z _ {e} ^ {t} \right\| _ {2}, \tag {7} +$$ + +where $sg[\cdot]$ is the stop-gradient operation. + +Then we use the quantized feature map $z_{q}^{t}$ as the target signal for the slot feature map $F^{t}$ . The final objective of VQ-VAE and the VQ reconstruction is: + +$$ +\mathcal {L} _ {\mathrm {V Q}} = \mathcal {L} _ {\mathrm {V Q V A E}} + \| s g [ F ^ {t} ] - z _ {q} ^ {t} \| _ {2} + \| s g [ z _ {q} ^ {t} ] - F ^ {t} \| _ {2}. \tag {8} +$$ + +Motion-guided token representation. The last term in Equation 8, $||sg[F^t] - z_q^t||_2$ , enables the motion signal from slot learning to jointly optimize the token space through the output of the slot decoder. Furthermore, the token representation and the motion cues build a connection linked by the slot learning, thus enabling the emergence of interpretable object-specific mid-level features of tokens. In addition, reconstructing in a more compact token space also benefits the model by better utilizing the motion signal to achieve an improved slot representation and temporal consistency. + +# 3.4. Optimization + +Token contrastive constraint. The goal of reconstructing in the VQ-space is that it is more compact and of lower variation compared with the RGB space. To make the VQ-space more structured, we add an optional contrastive constraint below to the vector space, which increases the independence between latent vectors: + +$$ +\mathcal {L} _ {\text {c o n t r a s t i v e}} = \left| \left| \mathbb {I} - \operatorname {s o f t m a x} \left(E \cdot E ^ {T}\right) \right| \right|, \tag {9} +$$ + +where $\mathbb{I}$ is the identity matrix and $E\in \mathbb{R}^{N\times d_{vq}}$ is the matrix of the feature embedding space $S$ . + +The final loss function is a combination of the reconstruction objective, the motion objective, and the optional contrastive constraint: + +$$ +\mathcal {L} = \lambda \mathcal {L} _ {\text {m o t i o n}} + \mathcal {L} _ {\text {r e c o n}} + \lambda_ {c} \mathbb {1} _ {\mathrm {V Q}} \mathcal {L} _ {\text {c o n t r a s t i v e}}, \tag {10} +$$ + +where $\lambda$ and $\lambda_{c}$ are weighting factors and $\mathbb{1}_{\mathrm{VQ}}$ is an indicator function. For the reconstruction loss, we set $\mathcal{L}_{\mathrm{recon}} = \mathcal{L}_{\mathrm{VQ}}$ when performing reconstruction in the VQ-space. Otherwise, we use an $L_{2}$ loss for reconstruction in the other three spaces. + +![](images/58fe986f8c7ec3db75011d693a1a648516c53d1132b86d6cd3e8f5323688aa63.jpg) +Figure 3. Frame samples from the video datasets used in our experiments. MOVi [21] (left) is a multi-object video dataset created by simulating rigid body dynamics. TRI-PD [4] (middle) is a collection of photo-realistic, synthetic driving videos. KITTI [19] (right) is a real-world benchmark with city driving scenes. + +![](images/77f1baf06ed23c0c31053cf62f6c4478f10efd24de4c939d4d0f0fa7443bca4c.jpg) + +![](images/5e8f6aa14e950aa5c277583a1f42bf4622fe94107069b50b2ba639b9c6d2ee04.jpg) + +# 4. Experimental Evaluation + +# 4.1. Experimental Settings + +Benchmarks. We evaluate our approach on three popular benchmarks with different complexity (Figure 3). + +MOVi [21] is a synthetic multi-object video dataset, created by simulating rigid body dynamics. We use this benchmark to ablate and analyze our model architecture. Following previous works [15, 53], we use the most complex subset, MOVi-E, for evaluation. MOVi-E contains both moving and static objects (maximum 20 objects) with linear random camera motion. The resolution of this dataset is $128 \times 128$ and each video contains 24 frames with a sample rate of 12 FPS. We use standard train and test split for MOVi-E. + +TRI-PD [4,25] is a synthetic dataset based on street driving scenarios collected by using a state-of-the-art synthetic data generation service [1]. The dataset comes with a collection of accurate annotations (e.g. for flow or depth), which we use to ablate the impact of additional information and its quality on various models. Each video in TRI-PD is 10 seconds captured at 20 FPS. There are 924 videos in the training set and 51 videos in the test set. We crop and resize each frame to the resolution of $480 \times 968$ . + +KITTI [19] is a real-world benchmark with city driving scenes. We use the whole 151 KITTI videos for training and the instance segmentation subsets with 200 single-frame images for evaluation. The frames are resized to $368 \times 1248$ . + +Baselines. We compare our methods against the most recent learning-based object discovery models. In particular, SAVi [36] and SAVi++ [15] are two direct extensions of the original slot attention [39] using optical flow and depth to facilitate object-centric learning. Alternatively, STEVE [53] uses a more powerful transformer decoder to enhance the object discovery performance. Bao et al. [4], Karazija et al. [33] utilize motion cues to guide object discovery. In addition, in the supplementary, we compare our approach to [51], which is a concurrent work that relies on ImageNet [12] pretraining [9]. For [4, 15, 36, 53], we use either the official code or public implementation (see supplementary for details). Due to the lack of implementation, we reuse reported results + +![](images/88da2141d7738868c42fb1955ea8e0fc4db66be8716a02f1b177e7e3c52a3781.jpg) +Figure 4. Object discovery in different reconstruction spaces on the evaluation set of MOVi-E. Compared with the other reconstruction spaces, the model trained using the VQ-space demonstrates a more accurate performance and better temporal consistency. + +with the two very recent approaches [33, 51]. + +Evaluation Metrics. Following prior works, we use the Foreground Adjusted Rand Index (FG. ARI) to evaluate the performance of the models, which captures how well the predicted segmentation masks match ground-truth masks in a permutation-invariant fashion. Notice that FG. ARI is measured in terms of the whole video, so that the temporal consistency of the slots is considered with this metric. + +Implementation Details. We use the same ResNet-18 ConvGRU encoder backbone for all the compared methods following [4]. We set the number of slots as 24 for MOVi-E and 45 for PD and KITTI based on the maximal number of objects in these datasets. On TRI-PD and KITTI, we further downsample the images by 4 for SAVi and SAVi++ to fit the GPU memory. Notice that, the other methods also produce downsampled slot masks, leading to the evaluation under the same resolution for all the methods. All the models are trained for 500 epochs using Adam [34]. During training, we train all the models with a randomly sampled video sequence of a fixed length (6, 5, 5 for MOVi-E, TRI-PD, and KITTI respectively). During the inference time, all the models are evaluated frame by frame until the end. For the FG. ARI measurement on TRI-PD, we discard any instance labels covering an area of $0.5\%$ or less of the first sampled video frame following [15]. We also evaluate FG. ARI at 5 FPS due to memory limitation. We use batch size 64 for MOVi-E, and 8 for TRI-PD and KITTI. + +Our method and Bao et al. [4] require motion segmentation signals. We apply [11], a powerful method pre-trained on the toy FlyingThings3D dataset [40], taking the ground-truth flow (MOVi-E) or RAFT [57] flow as the input. For SAVi and $\mathrm{SAV_i + + }$ , we do not use the first-frame bounding box supervision for a fair comparison. We generate the optical flow and depth annotation for KITTI using two state-of-the-art methods, RAFT [57] and VIDAR [25]. We also use them for ablations on TRI-PD. More details about the hyper-parameters, training scheme, and annotation generation are provided in the supplementary. + +
MotionSpaceDecoderFG. ARI
XVQLinear14.6
XVQLinear-CNN16.3
XVQTransformer40.1
XVQPerceiver52.4
XRGBPerceiver49.2
XFlowPerceiver36.3
XDepthPerceiver18.4
XFlow + DepthPerceiver38.0
XVQ (flow)Perceiver44.6
+ +# 4.2. Object Discovery on MOVi + +We use MOVi-E dataset to ablate our model with different choices of slot decoder, and reconstruction spaces, aiming to elucidate their roles in object discovery. We also compare the performance with state-of-the-art models in this section. + +Set up. We first fix the reconstruction space as the VQ-space and ablate the decoder architectures (first 4 lines in Table 1), and then we fix the decoder architecture with the best component and reconstruct in different reconstruction space (Line 5 to 9 in Table 1). Since SAVi++ [15] argues that reconstructing in the combined space of flow and depth is the key to their success, we add a variant $Flow + Depth$ , which also reconstructs in this combined space. Furthermore, as VQ-VAE can be trained to reconstruct the image in the flow space as well, we also report a $VQ$ (flow) variant for completeness. We do not include motion cues for the architecture analysis. A more comprehensive ablation is reported in the supplementary. + +Decoder analysis. By comparing the results on the first four lines in Table 1, similar to [53], we find that the capacity of the decoder indeed plays a key role in the object discovery performance. A simple linear decoder fails, but adding an additional CNN decoder can bring limited performance gains. Introducing powerful transformer-based decoders, on the other hand, greatly improves the object discovery capabilities of the model. Between the two transformer variants, the more advanced perceiver decoder achieves better results, while being more computationally efficient. + +Reconstruction space analysis. For the four reconstruction spaces, the VQ-space yields the best object discovery performance. We analyze the learned slots and token representations in Figure 4 and make the following discoveries. Firstly, the model trained using the VQ-space better separates the objects from the background and shows stronger temporal consistency. This is due to the more structured, compact, and lower-variance reconstruction space provided + +Table 1. Model architecture analysis with different choices of the slot decoder and reconstruction space on MOVi-E. Perceiver decoder + VQ reconstruction space yields the best performance, indicating that (1) the capacity of the decoder plays a key role in object discovery; (2) learnable, vector-quantized reconstruction space outperforms fixed alternatives like depth or flow. + +
ModelFG. ARI
Bao et al.51.6
STEVE51.2
SAVi28.1
SAVi++31.7
MoTok (no motion)52.4
MoTok63.8
SAVi (G.T. box)53.4
SAVi++ (G.T. box)84.1
+ +Table 2. Comparison to the state-of-the-art object discovery approaches on the validation sets of MOVi-E using FG. ARI. Our approach outperforms all the recent methods with the help of motion cues and vector quantization. + +by the quantized features, which simplifies the task of grouping, compared to the raw RGB space. + +Secondly, by comparing Flow and $\text{Flow+Depth}$ variants, we find that the latter indeed carries more information, allowing the model to better group the objects, which is consistent with [15]. However, interestingly, we also find that given a sufficiently strong decoder, reconstructing in depth or flow space does not bring further improvements compared to RGB, and can even decrease the performance. Finally, comparing Flow and $\text{VQ}(\text{flow})$ , we find that although adding the quantization improves performance significantly, this variant still lags behind reconstructing in the raw RGB space, reinforcing our previous conclusions. + +Comparison to the state of the art. In Table 2, we compare our approach to the state-of-the-art methods. Our model achieves the best performance among them. Additionally, compared to SAVi and SAVi++, even our variants that reconstructs in the same space outperform them, indicating that the perceiver decoder and single-stage decoding strategy are the optimal choices. Our model also outperforms STEVE, due to the fact that, in contrast to DVAE, the codebook in VQ-VAE can be jointly optimized with slot learning (see Equation 8). Compared with Bao et al., our model shows better performance even without motion cues used by that method, but further incorporating them into our approach allows it to achieve top results. + +Notice that, we compare with the variants of SAVi and SAVi++ that do not rely on ground-truth bounding boxes at test time. Their published numbers are included at the bottom of the table for reference, but they are not comparable to the other unsupervised methods. + +# 4.3. Object Discovery with Realistic Driving Videos + +Set up. Firstly, following [4], we capitalize on the ground-truth annotations available in TRI-PD, and evaluate the importance of the quality of external signals for various approaches. In particular, we separately evaluate using ground-truth (GT) flow, depth, and motion segmentation, as well as estimating those using state-of-the-art algorithms [11,25,57]. In addition, we report a variant STEVE- $m$ for which we add + +![](images/1e22679eb8d0f042d634ca737a3968d2d9f3f7d2ca9bcca49bb70b68632f8324.jpg) +Figure 5. Reconstruction, slot representation, and token masks for [4], STEVE-m [53], and our method. We use ground-truth motion cues for supervision, and visualize top-10 masks excluding background slots. MoTok effectively leverages the synergy between motion and tokenization, enabling the emergence of interpretable object-specific mid-level features, which simplify the problem of object discovery. + +
ModelSignalGPU Mem. (GB)FG. ARI
GTEST
Bao et al.Motion21.350.434.7
SAViFlow17.317.914.2
SAVi++Depth + Flow17.518.415.3
STEVE-66.712.5-
STEVE-mMotion68.245.732.2
MoTokMotion10.960.655.0
+ +Table 3. Object discovery evaluation of different models on TRI-PD dataset. GT denotes the ground-truth guidance given to each model and EST denotes the estimated cues. State-of-the-art models fail to work without motion cues; introducing token representation helps our model better utilize the motion signal. + +the same motion segmentation supervision as to [4] and our approach for a fair comparison. + +From Table 3, we can observe that our method strongly outperforms all the baselines. Additionally, we find that even with the ground-truth flow or depth, SAVI and SAVI++ do not work well for realistic videos with complex background and crowded scenes. Without motion cues, even equipped with a powerful transformer decoder, STEVE fails to work in this challenging setting as well. Finally, our model achieves better performance compared to Bao et al. with both GT and estimated motion segments, indicating its better robustness. + +Impact of motion-guided tokens: In Figure 5, we visualize the RGB reconstruction, slot representation, and token masks (if possible) for Bao et al., STEVE-m, and our MoTok. There are several key observations. Firstly, compared with STEVE-m, MoTok enables the emergence of more interpretable mid-level features, even though both models are trained with motion cues and tokenization. This is due to the interplay between object and slot discovery in our end-to-end framework, whereas STEVE-m trains DVAE separately from the rest of the model. + +Secondly, our model achieves the best RGB reconstruction result, even for the latter frames. The high quality of the reconstruction indicates that the model can better take advantage of the appearance signal to optimize the slot representation. Finally, our model demonstrates stronger temporal consistency compared with the baselines, thanks to the struc + +![](images/75203d917d95230b77c51b97a02aa668a0e9ff61b7d6c1d31c443472f0b2aa52.jpg) +Figure 6. Visualization of the learned tokens for vanilla VQ-VAE and MoTok. The proposed MoTok model shows a better alignment with semantic categories. + +tured reconstruction space. To sum up, our architecture effectively leverages the synergy between motion cues and tokenization, enabling the emergence of interpretable mid-level features, which greatly simplifies the task of object discovery. + +Interpretable tokens. To better illustrate the interpretability of motion-guided tokens, we measured their alignment with the ground-truth semantic labels using cluster purity [50]. Vanilla VQ-VAE (no motion guidance) achieves a purity of 40.1, whereas ours reaches 65.7 (chance performance: 38.5). Qualitative comparison in Figure 6 clearly demonstrates our better semantic alignment. + +Scalability to realistic videos: We additionally report the GPU memory usage per batch in Table 3. The per-slot decoding strategy and the use of transformer decoder limit the scalability of the other baselines. In comparison, our model achieves the best performance with the lowest memory consumption, showing a great generalization capability to real-world videos, thanks to the efficient perceiver decoder and the single-shot decoding strategy. + +Impact of objectives. Finally, we ablate the objective design of our model with the GT annotations in Table 4. We build three variants of our model, one without the additional token contrastive constraint defined by Equation 10 (no contrastive), the second without the motion cues (no motion), and the last trained with ground-truth instance masks for all objects to indicate the performance upper bound $(^{*})$ . By comparing the performance of these variants, we make the following observations: (1) in realistic synthetic videos even the strongest architectures fail to resolve the object/background ambiguity in the absence of motion cues, demonstrating the difficulty of object discovery tasks in the real world; (2) + +
ModelFG. ARI
MoTok60.6
MoTok (no contrastive)57.4
MoTok (no motion)15.6
MoTok*64.7
+ +Table 4. Ablation study on different learning signals. MoTok* is our model trained with ground-truth instance segmentation. The motion cue and the contrastive constraint help improve the capability of the proposed model, almost reaching the performance upper bound (MoTok*). + +
ModelFG. ARI
Bao et al. [4]47.1
SAVi20.0
SAVi++23.9
STEVE11.9
Karazija et al. [33]50.8
Karazija et al. [33] (WL)51.9
MoTok64.4
+ +Table 5. Evaluation of object discovery in the real-world KITTI dataset. Karazija et al. [33] (WL) is a variant trained with their proposed wrapping loss. Our model achieves state-of-the-art in real-world object discovery. + +contrastive constraint facilitates diversity during vector quantization, increasing the information content of each token, which helps reduce the object/background ambiguity; (3) the performance of our model is close to the upper bound, indicating the effectiveness of the motion-guided tokens. + +# 4.4. Object Discovery in the Real World + +Set up. To train the object discovery model on the real-world KITTI benchmark, we first pre-train all the models on PD using ground-truth flow and depth and then fine-tune them using estimated annotations on KITTI. + +The comparisons on KITTI are shown in Table 5 and Figure 7. We find that, consistent with our observations on TRI-PD, without motion guidance, both more powerful decoders (STEVE) and the simpler reconstruction space (SAVi and $\mathrm{SAV_i + + }$ ) fail to work. However, our model still captures good dynamic object segmentation results. Compared to Bao et al., we still achieve a better FG. ARI score and a more solid segmentation mask, indicating the benefit of the model design and the motion-guided tokens. Finally, we outperform the very recent approach of Karazija et al. [33] even when they include an additional warping loss objective. + +# 4.5. Limitations and Future Work + +Object/background ambiguity in the real world. Even after successfully parsing the world into object and background slots, separating those from each other in an unsupervised way is an open challenge. Promising directions include temporal contrastive learning [5, 28], unsupervised clustering [8, 59], and utilizing geometric cues [65, 66]. + +Slot drift. From Figure 5, we notice that the final RGB reconstruction quality decreases with longer videos, which indicates that the distribution of the slot representation has shifted in the latter frames. This issue leads to degraded temporal consistency and ambiguous object masks. Better training schemes and model improvements, such as adding data augmentation during training [15], replacing GRU units with long-short memory units [27], and enforcing temporal + +![](images/ebf2ce38a48df3be5a2aedea02d2731cdfc99ad2e300a320cf7a8a6dd50a7f60.jpg) +Figure 7. Top-10 slot visualizations of all the compared methods on the evaluation set of TRI-PD and KITTI datasets. The samples on KITTI are cropped for better visualization. We discard the masks taking more than $20\%$ of the whole pixels. Both the more powerful decoders (STEVE) and the simpler reconstruction space (SAVi and $\mathrm{SAV_i + + }$ ) fail to work in realistic driving scenes. Our model produces the most accurate masks and captures more objects. + +consistency in the token space can help. + +Better measuring object discovery in the real world. Measuring the performance of object discovery in the real world is challenging. As noted in [15], one slot may re-bind to another object after the tracked object moves out of the frame and a new object enters the scene. The FG. ARI score is not suitable for such scenarios but currently, there are no better metrics for object discovery which is able to tackle the re-binding issue. Developing a novel metric for this problem could have a significant impact on the community. + +# 5. Conclusion + +This work proposes MoTok, a unified architecture that leverages motion-guided tokenization for object discovery. By jointly training the slot representation with motion cues and vector quantization, our model enables the emergence of interpretable mid-level features which simplifies the problem of object discovery. Comprehensive evaluation on both synthetic and real-world benchmarks shows that with sufficient capacity of the slot decoder, motion guidance alleviates the need for labels, optical flow, or depth decoding, thanks to tokenization, achieving state-of-the-art results. + +Acknowledgements. We thank Dian Chen, Alexei Efros, and Andrew Owens for their valuable comments. This research was supported by Toyota Research Institute. YXW was supported in part by NSF Grant 2106825, NIFA Award 2020-67021-32799, and the NCSA Fellows program. + +# References + +[1] Parallel domain. https://paralleldomain.com/, November 2021. 5 +[2] Pablo Arbeláez, Jordi Pont-Tuset, Jonathan T Barron, Ferran Marques, and Jitendra Malik. Multiscale combinatorial grouping. In CVPR, 2014. 2 +[3] Nicolas Ballas, Li Yao, Chris Pal, and Aaron Courville. Delving deeper into convolutional networks for learning video representations. In ICLR, 2016. 3 +[4] Zhipeng Bao, Pavel Tokmakov, Allan Jabri, Yu-Xiong Wang, Adrien Gaidon, and Martial Hebert. Discovering objects that can move. In CVPR, 2022. 1, 2, 3, 5, 6, 7, 8 +[5] Zhangxing Bian, Allan Jabri, Alexei A Efros, and Andrew Owens. Learning pixel trajectories with multiscale contrastive random walks. In CVPR, 2022. 8 +[6] Christopher P Burgess, Loic Matthew, Nicholas Watters, Rishabh Kabra, Irina Higgins, Matt Botvinick, and Alexander Lerchner. MONet: Unsupervised scene decomposition and representation. arXiv preprint arXiv:1901.11390, 2019. 1, 2 +[7] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 1, 2 +[8] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features. In ECCV, 2018. 8 +[9] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 5 +[10] Chang Chen, Fei Deng, and Sungjin Ahn. Roots: Object-centric representation and rendering of 3D scenes. JMLR, 2021. 2 +[11] Achal Dave, Pavel Tokmakov, and Deva Ramanan. Towards segmenting anything that moves. In ICCV Workshops, 2019. 1, 5, 6 +[12] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 5 +[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 2 +[14] Yilun Du, Kevin Smith, Tomer Ulman, Joshua Tenenbaum, and Jiajun Wu. Unsupervised discovery of 3D physical objects from video. In ICLR, 2021. 2 +[15] Gamaleldin F Elsayed, Aravindh Mahendran, Sjoerd van Steenkiste, Klaus Greff, Michael C Mozer, and Thomas Kipf. Savi++: Towards end-to-end object-centric learning from real-world videos. In NeurIPS, 2022. 1, 2, 5, 6, 8 +[16] Martin Engelcke, Adam R Kosiorek, Oiwi Parker Jones, and Ingmar Posner. GENESIS: Generative scene inference and sampling with object-centric latent representations. In ICLR, 2020. 1, 2 + +[17] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In CVPR, 2021. 2 +[18] Pedro F Felzenszwalb and Daniel P Huttenlocher. Efficient graph-based image segmentation. IJCV, 2004. 2 +[19] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In CVPR, 2012. 5 +[20] Robert Gray. Vector quantization. IEEE Assp Magazine, 1984. 2 +[21] Klaus Greff, Francois Belletti, Lucas Beyer, Carl Doersch, Yilun Du, Daniel Duckworth, David J Fleet, Dan Gnanapragasam, Florian Golemo, Charles Herrmann, et al. Kubric: A scalable dataset generator. In CVPR, 2022. 5 +[22] Klaus Greff, Raphaël Lopez Kaufman, Rishabh Kabra, Nick Watters, Christopher Burgess, Daniel Zoran, Loic Matthey, Matthew Botvinick, and Alexander Lerchner. Multi-object representation learning with iterative variational inference. In ICML, 2019. 1, 2 +[23] Klaus Greff, Antti Rasmus, Mathias Berglund, Tele Hao, Harri Valpola, and Jürgen Schmidhuber. Tagger: Deep unsupervised perceptual grouping. In NeurIPS, 2016. 2 +[24] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In CVPR, 2022. 2 +[25] Rares Guizilini, Vitor abd Ambrus, Sudeep Pillai, Allan Raventos, and Adrien Gaidon. 3d packing for self-supervised monocular depth estimation. In CVPR, 2020. 5, 6 +[26] Paul Henderson and Christoph H Lampert. Unsupervised object-centric video generation and decomposition in 3D. In NeurIPS, 2020. 2 +[27] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation, 1997. 8 +[28] Allan Jabri, Andrew Owens, and Alexei Efros. Space-time correspondence as a contrastive random walk. NeurIPS, 2020. 8 +[29] Andrew Jaegle, Sebastian Borgeaud, Jean-Baptiste Alayrac, Carl Doersch, Catalin Ionescu, David Ding, Skanda Koppula, Daniel Zoran, Andrew Brock, Evan Shelhamer, et al. Perceiver io: A general architecture for structured inputs & outputs. In ICLR, 2022. 2, 4 +[30] Andrew Jaegle, Felix Gimeno, Andy Brock, Oriol Vinyals, Andrew Zisserman, and Joao Carreira. Perceiver: General perception with iterative attention. In ICML, 2021. 2 +[31] Jindong Jiang, Sepehr Janghorbani, Gerard De Melo, and Sungjin Ahn. SCALOR: Generative world models with scalable object representations. In ICLR, 2020. 2 +[32] Daniel Kahneman, Anne Treisman, and Brian J Gibbs. The reviewing of object files: Object-specific integration of information. Cognitive psychology, 1992. 1 +[33] Laurynas Karazija, Subhabrataand Choudhury, Iro Laina, Christian Rupprecht, and Andrea Vedaldi. Unsupervised multi-object segmentation by predicting probable motion patterns. In NeurIPS, 2022. 2, 5, 8 +[34] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 5 + +[35] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. In ICLR, 2014. 2 +[36] Thomas Kipf, Gamaleldin F Elsayed, Aravindh Mahendran, Austin Stone, Sara Sabour, Georg Heigold, Rico Jonschkowski, Alexey Dosovitskiy, and Klaus Greff. Conditional object-centric learning from video. In ICLR, 2022. 1, 2, 5 +[37] Kurt Koffka. Principles of Gestalt psychology. Routledge, 2013. 2 +[38] Zhixuan Lin, Yi-Fu Wu, Skand Vishwanath Peri, Weihao Sun, Gautam Singh, Fei Deng, Jindong Jiang, and Sungjin Ahn. Space: Unsupervised object-oriented scene representation via spatial attention and decomposition. In ICLR, 2020. 1, 2 +[39] Francesco Locatello, Dirk Weissenborn, Thomas Unterthiner, Aravindh Mahendran, Georg Heigold, Jakob Uszkoreit, Alexey Dosovitskiy, and Thomas Kipf. Object-centric learning with slot attention. In NeurIPS, 2020. 1, 2, 5 +[40] Nikolaus Mayer, Eddy Ilg, Philip Hausser, Philipp Fischer, Daniel Cremers, Alexey Dosovitskiy, and Thomas Brox. A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In CVPR, 2016. 5 +[41] Tim Meinhardt, Alexander Kirillov, Laura Leal-Taixe, and Christoph Feichtenhofer. Trackformer: Multi-object tracking with transformers. In CVPR, 2022. 2 +[42] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2 +[43] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, 2021. 1, 2 +[44] Ali Razavi, Aaron Van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. NeurIPS, 2019. 2 +[45] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster R-CNN: Towards real-time object detection with region proposal networks. In NeurIPS, 2015. 1 +[46] Danilo Jimenez Rezende, Shakir Mohamed, and Daan Wierstra. Stochastic backpropagation and approximate inference in deep generative models. In ICML, 2014. 2 +[47] Jason Tyler Rolfe. Discrete variational autoencoders. In ICLR, 2017. 1, 2 +[48] Mehdi SM Sajjadi, Daniel Duckworth, Aravindh Mahendran, Sjoerd van Steenkiste, Filip Pavetic, Mario Lucic, Leonidas J Guibas, Klaus Greff, and Thomas Kipf. Object scene representation transformer. In NeurIPS, 2022. 2 +[49] Mehdi SM Sajjadi, Henning Meyer, Etienne Pot, Urs Bergmann, Klaus Greff, Noha Radwan, Suhani Vora, Mario Lucic, Daniel Duckworth, Alexey Dosovitskiy, et al. Scene representation transformer: Geometry-free novel view synthesis through set-latent scene representations. In CVPR, 2022. 2 +[50] Hinrich Schütze, Christopher D Manning, and Prabhakar Raghavan. Introduction to information retrieval. Cambridge University Press Cambridge, 2008. 7 +[51] Maximilian Seitzer, Max Horn, Andrii Zadaianchuk, Dominik Zietlow, Tianjun Xiao, Carl-Johann Simon-Gabriel, Tong He, Zheng Zhang, Bernhard Scholkopf, Thomas Brox, et al. + +Bridging the gap to real-world object-centric learning. In ICLR, 2023. 2, 5 +[52] Gautam Singh, Fei Deng, and Sungjin Ahn. Illiterate dall-e learns to compose. In ICLR, 2021. 1, 2 +[53] Gautam Singh, Yi-Fu Wu, and Sungjin Ahn. Simple unsupervised object-centric learning for complex and naturalistic videos. In NeurlPS, 2022. 1, 2, 5, 6, 7 +[54] Elizabeth S Spelke and Katherine D Kinzler. Core knowledge. Developmental science, 2007. 1 +[55] Karl Stelzner, Kristian Kersting, and Adam R Kosiorek. Decomposing 3D scenes into objects via unsupervised volume segmentation. arXiv preprint arXiv:2104.01148, 2021. 2 +[56] Chen Sun, Austin Myers, Carl Vondrick, Kevin Murphy, and Cordelia Schmid. Videobert: A joint model for video and language representation learning. In CVPR, 2019. 2 +[57] Zachary Teed and Jia Deng. RAFT: Recurrent all-pairs field transforms for optical flow. In ECCV, 2020. 5, 6 +[58] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. NeurIPS, 2017. 2, 4 +[59] Wouter Van Gansbeke, Simon Vandenhende, Stamatios Georgoulis, and Luc Van Gool. Unsupervised semantic segmentation by contrasting object mask proposals. In ICCV, 2021. 8 +[60] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 2017. 1, 2 +[61] Rishi Veerapaneni, John D Co-Reyes, Michael Chang, Michael Janner, Chelsea Finn, Jiajun Wu, Joshua Tenenbaum, and Sergey Levine. Entity abstraction in visual model-based reinforcement learning. In CoRL, 2020. 2 +[62] Ziyi Wu, Nikita Dvornik, Klaus Greff, Thomas Kipf, and Animesh Garg. Slotformer: Unsupervised visual dynamics simulation with object-centric models. In ICLR, 2023. 2 +[63] Wilson Yan, Yunzhi Zhang, Pieter Abbeel, and Aravind Srinivas. Videogpt: Video generation using vq-vae and transformers. arXiv preprint arXiv:2104.10157, 2021. 2 +[64] Peiyu Yu, Sirui Xie, Xiaojian Ma, Yixin Zhu, Ying Nian Wu, and Song-Chun Zhu. Unsupervised foreground extraction via deep region competition. In NeurIPS, 2021. 2 +[65] Jiahui Zhang, Dawei Sun, Zixin Luo, Anbang Yao, Lei Zhou, Tianwei Shen, Yurong Chen, Long Quan, and Honggen Liao. Learning two-view correspondences and geometry using order-aware network. In ICCV, 2019. 8 +[66] Brady Zhou and Philipp Krahenbuhl. Cross-view transformers for real-time map-view semantic segmentation. In CVPR, 2022. 8 +[67] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In ICLR, 2021. 2 \ No newline at end of file diff --git a/2023/Object Discovery From Motion-Guided Tokens/images.zip b/2023/Object Discovery From Motion-Guided Tokens/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7d54ce9eb662ca63d1d98ea9dd79b1adb1058a6c --- /dev/null +++ b/2023/Object Discovery From Motion-Guided Tokens/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:59d91f5c24538d610379397835bdc841f0da200c3f5f809b3b0d35f4abdfa2a1 +size 473872 diff --git a/2023/Object Discovery From Motion-Guided Tokens/layout.json b/2023/Object Discovery From Motion-Guided Tokens/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..231704ec5ed82bf8e7a39b707cd631068a8d7b3b --- /dev/null +++ b/2023/Object Discovery From Motion-Guided Tokens/layout.json @@ -0,0 +1,9346 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 153, + 103, + 441, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 103, + 441, + 121 + ], + "spans": [ + { + "bbox": [ + 153, + 103, + 441, + 121 + ], + "type": "text", + "content": "Object Discovery from Motion-Guided Tokens" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 148, + 142, + 444, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 142, + 444, + 185 + ], + "spans": [ + { + "bbox": [ + 148, + 142, + 444, + 185 + ], + "type": "text", + "content": "Zhipeng Bao\\*,1 Pavel Tokmakov2 Yu-Xiong Wang3 Adrien Gaidon2 Martial Hebert1 CMU 2Toyota Research Institute 3UIUC" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 77, + 188, + 198, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 188, + 198, + 198 + ], + "spans": [ + { + "bbox": [ + 77, + 188, + 198, + 198 + ], + "type": "text", + "content": "{zbao, hebert}@cs.cmu.edu" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 217, + 188, + 418, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 188, + 418, + 199 + ], + "spans": [ + { + "bbox": [ + 217, + 188, + 418, + 199 + ], + "type": "text", + "content": "{pavel).(tokmakov, adrien).(gaidon}@tri.global" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 437, + 189, + 515, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 437, + 189, + 515, + 198 + ], + "spans": [ + { + "bbox": [ + 437, + 189, + 515, + 198 + ], + "type": "text", + "content": "yxw@illinois.edu" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "spans": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 252, + 290, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 252, + 290, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 252, + 290, + 456 + ], + "type": "text", + "content": "Object discovery - separating objects from the background without manual labels - is a fundamental open challenge in computer vision. Previous methods struggle to go beyond clustering of low-level cues, whether handcrafted (e.g., color, texture) or learned (e.g., from auto-encoders). In this work, we augment the auto-encoder representation learning framework with two key components: motion-guidance and mid-level feature tokenization. Although both have been separately investigated, we introduce a new transformer decoder showing that their benefits can compound thanks to motion-guided vector quantization. We show that our architecture effectively leverages the synergy between motion and tokenization, improving upon the state of the art on both synthetic and real datasets. Our approach enables the emergence of interpretable object-specific mid-level features, demonstrating the benefits of motion-guidance (no labeling) and quantization (interpretability, memory efficiency)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 468, + 128, + 481 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 468, + 128, + 481 + ], + "spans": [ + { + "bbox": [ + 47, + 468, + 128, + 481 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 489, + 288, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 489, + 288, + 632 + ], + "spans": [ + { + "bbox": [ + 46, + 489, + 288, + 632 + ], + "type": "text", + "content": "Objects are central in human and computer vision. In the former, they are a fundamental primitive used to decompose the complexity of the visual world into an actionable representation. This abstraction in turn enables higher-level cognitive abilities, such as casual reasoning and planning [32, 54]. In computer vision, object detection has achieved remarkable progress [7, 45] and is now an essential component in many applications (e.g., driving, robotics). However, these models require a large amount of manual labels from a fixed vocabulary of categories. Consequently, learning unsupervised, object-centric representations is an important step towards scaling up computer vision to the real world." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 633, + 288, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 633, + 288, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 633, + 288, + 693 + ], + "type": "text", + "content": "This topic has received renewed attention recently thanks to structured generative networks with iterative inference over a fixed set of variables [6, 16, 22, 38, 39]. These methods cluster pixels in the feature space of an auto-encoder, exhibiting behavior similar to grouping based on low-level cues," + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 309, + 228, + 425, + 286 + ], + "blocks": [ + { + "bbox": [ + 309, + 228, + 425, + 286 + ], + "lines": [ + { + "bbox": [ + 309, + 228, + 425, + 286 + ], + "spans": [ + { + "bbox": [ + 309, + 228, + 425, + 286 + ], + "type": "image", + "image_path": "bd50f31a004d67ebf426562ab8a6f0cdc614f4ea730ecdde0bf8fb70e446e90c.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 334, + 287, + 395, + 294 + ], + "lines": [ + { + "bbox": [ + 334, + 287, + 395, + 294 + ], + "spans": [ + { + "bbox": [ + 334, + 287, + 395, + 294 + ], + "type": "text", + "content": "Slot Representation" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 427, + 228, + 544, + 286 + ], + "blocks": [ + { + "bbox": [ + 427, + 228, + 544, + 286 + ], + "lines": [ + { + "bbox": [ + 427, + 228, + 544, + 286 + ], + "spans": [ + { + "bbox": [ + 427, + 228, + 544, + 286 + ], + "type": "image", + "image_path": "00bb9122d6004180f5f3a421aff54e947c80f7e98bbf1f829bbf45e00202e82f.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 450, + 287, + 518, + 295 + ], + "lines": [ + { + "bbox": [ + 450, + 287, + 518, + 295 + ], + "spans": [ + { + "bbox": [ + 450, + 287, + 518, + 295 + ], + "type": "text", + "content": "Token Representation" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 298, + 546, + 342 + ], + "lines": [ + { + "bbox": [ + 305, + 298, + 546, + 342 + ], + "spans": [ + { + "bbox": [ + 305, + 298, + 546, + 342 + ], + "type": "text", + "content": "Figure 1. TRI-PD dataset results: (left) top-10 foreground slot segments produced by our approach; (right) corresponding token representations. Compared to raw images, tokens in our framework present a more structured and compact space for reconstruction." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 354, + 547, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 354, + 547, + 389 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 547, + 389 + ], + "type": "text", + "content": "such as color or texture. Hence, they are restricted to toy images with colored geometric shapes on a plain background, and fail on more complex realistic scenes [4]." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 390, + 548, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 390, + 548, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 390, + 548, + 641 + ], + "type": "text", + "content": "Two main types of works attempt to address this shortcoming. The first family of methods sets out to simplify the grouping problem by introducing more structure into the output space, e.g., reconstructing optical flow [36] or depth [15]. They, however, require supervision, either in the form of known poses [36] or ground truth bounding boxes [15], veering away from the unsupervised goal of object discovery. In contrast, Bao et al. [4] resolve the object-background ambiguity by explicitly integrating an unsupervised motion segmentation algorithm [11] into the pipeline, showing substantial progress on realistic scenes. The second main direction to improve object discovery focuses on improving the decoder part of auto-encoding architectures [52, 53], replacing convolutional decoders with transformers [43, 60] combined with discrete variational auto-encoders (DVAE) [47] to reduce memory footprint. These more sophisticated architectures improve performance without additional supervision, including on real-world sequences. However, these methods are evaluated with different protocols (metrics, datasets) and therefore no clear architectural principles have emerged yet for unsupervised object discovery." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 642, + 548, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 548, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 548, + 713 + ], + "type": "text", + "content": "In this work, we introduce a novel architecture, Motion-guided Tokens (MoTok), based on the combination of two fundamental structural principles: motion and discretization. We define objects as discrete entities that might have an independent motion. As prior works have shown encouraging results thanks to unsupervised motion guidance and better" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 189, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 189, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 189, + 712 + ], + "type": "text", + "content": "*Work done during an internship at TRI" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22972" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 156 + ], + "type": "text", + "content": "transformer-based decoders, we propose to leverage motion to guide tokenization, the vector quantization process at the heart of attention mechanisms in transformer architectures (See Figure 1). In addition, to comprehensively evaluate the contributions of prior works, we ablate key design choices proposed in the past, such as the decoder architecture and reconstruction space, in a unified framework." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 157, + 290, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 157, + 290, + 324 + ], + "spans": [ + { + "bbox": [ + 46, + 157, + 290, + 324 + ], + "type": "text", + "content": "Our key contributions are as follows. (1) We introduce a novel auto-encoder architecture, MoTok, for unsupervised video object discovery with a new transformer decoder leveraging unsupervised motion-guided tokenization. (2) Our results on real and synthetic datasets show that with sufficient capacity in the decoder, motion guidance alleviates the need for labels, optical flow, or depth decoding thanks to tokenization, improving upon the state of the art. (3) We show that our motion-guided tokens map to interpretable mid-level features, going beyond typical clusters of low-level features, thus explaining why our method scales to challenging realistic videos. Our code, models, and synthetic data are made available at https://github.com/zpbao/MoTok/." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 332, + 134, + 345 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 332, + 134, + 345 + ], + "spans": [ + { + "bbox": [ + 47, + 332, + 134, + 345 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 353, + 288, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 353, + 288, + 400 + ], + "spans": [ + { + "bbox": [ + 46, + 353, + 288, + 400 + ], + "type": "text", + "content": "In this work, we tackle the object discovery problem in realistic videos with object-centric representation by capitalizing on motion-guided tokens. We review the most relevant works in these areas below." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 403, + 288, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 403, + 288, + 486 + ], + "spans": [ + { + "bbox": [ + 46, + 403, + 288, + 486 + ], + "type": "text", + "content": "Object Discovery tackles the problem of separating objects from background without manual labels [4]. Classic computer vision methods use appearance-based perceptual grouping to parse a scene into object-like regions [2, 18, 37]. Among them, [2] first propose a multiscale fast normalized cuts algorithm and achieved robust object segmentation performance for static images." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 487, + 290, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 487, + 290, + 642 + ], + "spans": [ + { + "bbox": [ + 46, + 487, + 290, + 642 + ], + "type": "text", + "content": "Recently, the topic of object discovery has experienced renewed attention under the name of unsupervised object-centric representation learning. A wide variety of learning-based methods has been introduced [6, 16, 22, 23, 31, 38, 39, 52, 61, 64], usually with an encoder-decoder architecture [35, 46]. These methods aim to learn compositional feature representations, e.g., a set of variables that can bind to objects in an image [6, 16, 22, 23, 39, 52, 61, 64], or a video [4, 15, 31, 33, 36, 38, 48, 51, 53, 62]. Among them, [39] first formulate the SlotAttention framework, which is used to bind a set of variables, called slots, to image locations. The slots are then decoded individually and combined to reconstruct the image." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "type": "text", + "content": "However, without additional constraints, such methods tend to converge to pixel grouping based on low-level cues, such as color, and do not generalize to realistic images or videos with complex backgrounds. To address this limitation, [36] and [15] extend the slot concept from static images to videos via reconstructing in the optical flow or depth space" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": "respectively. The intuition behind these methods is that this space provides stronger cues to separate the objects from the background. Separately, Bao et al. [4] use motion cues to guide the slots to find moving objects and then generalize to all the objects that can move." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 136, + 547, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 136, + 547, + 233 + ], + "spans": [ + { + "bbox": [ + 304, + 136, + 547, + 233 + ], + "type": "text", + "content": "In another line of work, Singh at al. [53] show that a combination of a more powerful transformer decoder [13] and discrete variational auto-encoder [47] can enhance the object discovery performance. Different from these works, we propose a unified architecture for object discovery that is flexible to different choices of decoders and reconstruction space. We also introduce the vector quantized features as an additional reconstruction space with motion guidance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 236, + 547, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 236, + 547, + 320 + ], + "spans": [ + { + "bbox": [ + 304, + 236, + 547, + 320 + ], + "type": "text", + "content": "Finally, several recent works also leverage 3D geometry as inductive biases to enforce the learning-based models to focus on object-like regions [10, 14, 26, 55]. Though these methods remain limited to the toy, synthetic environments, the underlying geometric priors are orthogonal to our approach and have a great potential to be combined with our proposed method as a future direction." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 324, + 548, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 324, + 548, + 553 + ], + "spans": [ + { + "bbox": [ + 304, + 324, + 548, + 553 + ], + "type": "text", + "content": "Vector quantization is originally developed for data compression in signal processing [20]. More recently, [58] propose the Vector-Quantized Variational Autoencoder (VQ-VAE), which learns a discrete representation of images, and models their distribution aggressively. This technique has motivated a series of investigations in solving different computer vision tasks including image synthesis [17, 24, 44], video generation [56, 63], and language modeling [24, 42, 43], to name a few. In this work, we adopt the vector quantization technique as a mid-level representation in our object discovery framework. The intuition is that reconstructing in a structured, low-dimensional feature space rather than the high-variance color space should simplify the task of resolving the object and background ambiguity. In comparison, [53] also introduce a mid-level feature space for object discovery in videos. However, they use DVAE, not VQ-VAE, and are motivated primarily by improving training efficiency. Moreover, they did not explore combing vector quantization with motion cues in an end-to-end trainable framework." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 558, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 548, + 715 + ], + "type": "text", + "content": "Transformers are originally proposed for sequence-to-sequence modeling in natural language processing [60]. They use a multi-head attention mechanism, instead of the recurrent memory units, to aggregate information from the input. Recently, vision transformer (ViT) [13] and its derivations have achieved state-of-the-art in several visual tasks [7,29,30,41,48,49,67]. Among them, Perceiver IO [29] design a computationally efficient architecture to handle arbitrary outputs in addition to arbitrary inputs, greatly enhancing the generalizability of the transformer architecture. We employ a variant of a Perceiver module as a decoder in our object discovery framework to combine high representational power with computational efficiency." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22973" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 73, + 545, + 194 + ], + "blocks": [ + { + "bbox": [ + 50, + 73, + 545, + 194 + ], + "lines": [ + { + "bbox": [ + 50, + 73, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 50, + 73, + 545, + 194 + ], + "type": "image", + "image_path": "2bf1afd0f9ca6580390c5f700aebeeff549fe9f4019eac6dc97d1852ab074670.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 201, + 547, + 236 + ], + "lines": [ + { + "bbox": [ + 46, + 201, + 547, + 236 + ], + "spans": [ + { + "bbox": [ + 46, + 201, + 547, + 236 + ], + "type": "text", + "content": "Figure 2. Model architecture of the proposed Motion-guided Tokens (MoTok) framework. MoTok is a unified framework for video object discovery that is flexible with different choices of decoders and reconstruction spaces. Our framework effectively leverages the synergy between motion and tokenization, and enables the emergence of interpretable object-specific mid-level features." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 247, + 103, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 247, + 103, + 258 + ], + "spans": [ + { + "bbox": [ + 47, + 247, + 103, + 258 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 267, + 290, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 267, + 290, + 374 + ], + "spans": [ + { + "bbox": [ + 46, + 267, + 290, + 374 + ], + "type": "text", + "content": "We now explain the proposed Motion-guided Tokens (Mo-Tok) framework in detail. Its architecture is shown in Figure 2. We first introduce a motion-guided slot learning framework [4] in Section 3.1. We then describe the slot decoder units in Section 3.2, which decode the slot features into a reconstruction space. In Section 3.3, we describe the choices of reconstruction space and explain the vector-quantized reconstruction space in detail. Finally, we demonstrate how we optimize the model in Section 3.4." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 383, + 274, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 383, + 274, + 396 + ], + "spans": [ + { + "bbox": [ + 47, + 383, + 274, + 396 + ], + "type": "text", + "content": "3.1. Preliminary: Motion-Guided Slot Learning" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 402, + 288, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 402, + 288, + 509 + ], + "spans": [ + { + "bbox": [ + 46, + 402, + 288, + 509 + ], + "type": "text", + "content": "Our object-centric representation learning module is derived from [4] (Figure 2, left). Concretely, given a sequence of video frames " + }, + { + "bbox": [ + 46, + 402, + 288, + 509 + ], + "type": "inline_equation", + "content": "\\{I^1, I^2, \\dots, I^T\\}" + }, + { + "bbox": [ + 46, + 402, + 288, + 509 + ], + "type": "text", + "content": ", we first process each frame through an encoder CNN to obtain an individual frame representation " + }, + { + "bbox": [ + 46, + 402, + 288, + 509 + ], + "type": "inline_equation", + "content": "H^t = f_{\\mathrm{enc}}(I^t)" + }, + { + "bbox": [ + 46, + 402, + 288, + 509 + ], + "type": "text", + "content": ". These individual representations are aggregated by a spatio-temporal Convolutional Gated Recurrent Unit (ConvGRU) [3] to obtain video encoding via " + }, + { + "bbox": [ + 46, + 402, + 288, + 509 + ], + "type": "inline_equation", + "content": "H'^t = \\mathrm{ConvGRU}(R^{t-1}, H^t)" + }, + { + "bbox": [ + 46, + 402, + 288, + 509 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 402, + 288, + 509 + ], + "type": "inline_equation", + "content": "R^{t-1} \\in \\mathbb{R}^{h' \\times w' \\times d_{\\mathrm{inp}}}" + }, + { + "bbox": [ + 46, + 402, + 288, + 509 + ], + "type": "text", + "content": " is the recurrent memory state." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 510, + 288, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 288, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 288, + 617 + ], + "type": "text", + "content": "Next, we perform a single attention operation with " + }, + { + "bbox": [ + 46, + 510, + 288, + 617 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 510, + 288, + 617 + ], + "type": "text", + "content": " slots to directly compute the slot state " + }, + { + "bbox": [ + 46, + 510, + 288, + 617 + ], + "type": "inline_equation", + "content": "S^t = W^t v(H'^t)" + }, + { + "bbox": [ + 46, + 510, + 288, + 617 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 510, + 288, + 617 + ], + "type": "inline_equation", + "content": "W^t \\in \\mathbb{R}^{K \\times N}" + }, + { + "bbox": [ + 46, + 510, + 288, + 617 + ], + "type": "text", + "content": " is the attention matrix, " + }, + { + "bbox": [ + 46, + 510, + 288, + 617 + ], + "type": "inline_equation", + "content": "N = h' \\times w'" + }, + { + "bbox": [ + 46, + 510, + 288, + 617 + ], + "type": "text", + "content": " indicates the flapped shape dimension, and " + }, + { + "bbox": [ + 46, + 510, + 288, + 617 + ], + "type": "inline_equation", + "content": "v(\\cdot)" + }, + { + "bbox": [ + 46, + 510, + 288, + 617 + ], + "type": "text", + "content": " is the value embedding function. " + }, + { + "bbox": [ + 46, + 510, + 288, + 617 + ], + "type": "inline_equation", + "content": "W^t" + }, + { + "bbox": [ + 46, + 510, + 288, + 617 + ], + "type": "text", + "content": " is computed using the slot state in the previous frame " + }, + { + "bbox": [ + 46, + 510, + 288, + 617 + ], + "type": "inline_equation", + "content": "S^{t-1}" + }, + { + "bbox": [ + 46, + 510, + 288, + 617 + ], + "type": "text", + "content": " and the input feature " + }, + { + "bbox": [ + 46, + 510, + 288, + 617 + ], + "type": "inline_equation", + "content": "H'^t" + }, + { + "bbox": [ + 46, + 510, + 288, + 617 + ], + "type": "text", + "content": ". For the first frame, we use a learnable initial state " + }, + { + "bbox": [ + 46, + 510, + 288, + 617 + ], + "type": "inline_equation", + "content": "S^0" + }, + { + "bbox": [ + 46, + 510, + 288, + 617 + ], + "type": "text", + "content": ". At the same time, for each slot " + }, + { + "bbox": [ + 46, + 510, + 288, + 617 + ], + "type": "inline_equation", + "content": "s_i^t" + }, + { + "bbox": [ + 46, + 510, + 288, + 617 + ], + "type": "text", + "content": ", we can also obtain the attention mask " + }, + { + "bbox": [ + 46, + 510, + 288, + 617 + ], + "type": "inline_equation", + "content": "W_{i,i}^t" + }, + { + "bbox": [ + 46, + 510, + 288, + 617 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "content": "In [4], a motion cue is also added to guide the slots to find moving objects. A set of sparse, instance-level motion segmentation masks " + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{M} = \\{M^1,M^2,\\dots,M^T\\}" + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "content": " is assumed to be provided with every video, with " + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "inline_equation", + "content": "M^{t} = \\{m_{1},m_{2},\\dots,m_{C^{t}}\\}" + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "inline_equation", + "content": "C^t" + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "content": " is the number of moving objects that were successfully segmented in frame " + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "inline_equation", + "content": "m_j\\in \\{0,1\\}^{h'\\times w'}" + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "content": " is a binary mask. The attention mask " + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "inline_equation", + "content": "W^{t}\\in \\mathbb{R}^{K\\times N}" + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "content": ", is then supervised with the motion segments. " + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "inline_equation", + "content": "M^t" + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "content": " is also considered" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 247, + 545, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 247, + 545, + 272 + ], + "spans": [ + { + "bbox": [ + 305, + 247, + 545, + 272 + ], + "type": "text", + "content": "as a set of length " + }, + { + "bbox": [ + 305, + 247, + 545, + 272 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 305, + 247, + 545, + 272 + ], + "type": "text", + "content": " padded with " + }, + { + "bbox": [ + 305, + 247, + 545, + 272 + ], + "type": "inline_equation", + "content": "\\emptyset" + }, + { + "bbox": [ + 305, + 247, + 545, + 272 + ], + "type": "text", + "content": " (no object), and a bipartite matching is found between them with the lowest cost:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 351, + 278, + 546, + 311 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 278, + 546, + 311 + ], + "spans": [ + { + "bbox": [ + 351, + 278, + 546, + 311 + ], + "type": "interline_equation", + "content": "\\hat {\\sigma} = \\underset {\\sigma} {\\arg \\min } \\sum_ {i = 1} ^ {K} \\mathcal {L} _ {\\operatorname {s e g}} \\left(m _ {i}, W _ {:, \\sigma (i)} ^ {t}\\right), \\tag {1}", + "image_path": "a4a2904c8b0aafa7250a79f462a3a70cd6b8bbd71d4c40a4609d83a2741bdbe2.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 318, + 545, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 318, + 545, + 367 + ], + "spans": [ + { + "bbox": [ + 305, + 318, + 545, + 367 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 318, + 545, + 367 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{seg}}(m_i,W_{:, \\sigma(i)}^t)" + }, + { + "bbox": [ + 305, + 318, + 545, + 367 + ], + "type": "text", + "content": " is the segmentation loss between the motion mask " + }, + { + "bbox": [ + 305, + 318, + 545, + 367 + ], + "type": "inline_equation", + "content": "m_i" + }, + { + "bbox": [ + 305, + 318, + 545, + 367 + ], + "type": "text", + "content": " and the attention map of the slot with index " + }, + { + "bbox": [ + 305, + 318, + 545, + 367 + ], + "type": "inline_equation", + "content": "\\sigma(i)" + }, + { + "bbox": [ + 305, + 318, + 545, + 367 + ], + "type": "text", + "content": ". Once the assignment " + }, + { + "bbox": [ + 305, + 318, + 545, + 367 + ], + "type": "inline_equation", + "content": "\\hat{\\sigma}" + }, + { + "bbox": [ + 305, + 318, + 545, + 367 + ], + "type": "text", + "content": " has been computed, the final motion supervision objective is defined as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 339, + 373, + 545, + 405 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 373, + 545, + 405 + ], + "spans": [ + { + "bbox": [ + 339, + 373, + 545, + 405 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {m o t i o n}} = \\sum_ {i = 1} ^ {K} \\mathbb {1} _ {\\{m _ {i} \\neq \\emptyset \\}} \\mathcal {L} _ {\\text {s e g}} \\left(m _ {i}, W _ {:, \\hat {\\sigma} (i)} ^ {t}\\right), \\tag {2}", + "image_path": "36f51ec66db0add597ed4ff2dd392c74b3049207917a04c3cba10777d041a31b.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 411, + 545, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 411, + 545, + 436 + ], + "spans": [ + { + "bbox": [ + 305, + 411, + 545, + 436 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 411, + 545, + 436 + ], + "type": "inline_equation", + "content": "\\mathbb{1}_{\\{m_i \\neq \\emptyset\\}}" + }, + { + "bbox": [ + 305, + 411, + 545, + 436 + ], + "type": "text", + "content": " denotes that the loss is only computed for the matched slots and " + }, + { + "bbox": [ + 305, + 411, + 545, + 436 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{seg}}" + }, + { + "bbox": [ + 305, + 411, + 545, + 436 + ], + "type": "text", + "content": " is the binary cross entropy." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 442, + 392, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 442, + 392, + 453 + ], + "spans": [ + { + "bbox": [ + 306, + 442, + 392, + 453 + ], + "type": "text", + "content": "3.2. Slot Decoders" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 460, + 547, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 460, + 547, + 507 + ], + "spans": [ + { + "bbox": [ + 305, + 460, + 547, + 507 + ], + "type": "text", + "content": "The goal of the Slot Decoder is to map the slot representation " + }, + { + "bbox": [ + 305, + 460, + 547, + 507 + ], + "type": "inline_equation", + "content": "(S^t, W^t)" + }, + { + "bbox": [ + 305, + 460, + 547, + 507 + ], + "type": "text", + "content": " to a 2D feature map " + }, + { + "bbox": [ + 305, + 460, + 547, + 507 + ], + "type": "inline_equation", + "content": "F^t" + }, + { + "bbox": [ + 305, + 460, + 547, + 507 + ], + "type": "text", + "content": " for the reconstruction space. As shown in Figure 2 (middle), we propose four choices for the slot decoder." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 508, + 545, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 508, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 305, + 508, + 545, + 533 + ], + "type": "text", + "content": "Linear Decoder directly maps the slot features " + }, + { + "bbox": [ + 305, + 508, + 545, + 533 + ], + "type": "inline_equation", + "content": "S^t" + }, + { + "bbox": [ + 305, + 508, + 545, + 533 + ], + "type": "text", + "content": " to their corresponding positions based on the attention mask " + }, + { + "bbox": [ + 305, + 508, + 545, + 533 + ], + "type": "inline_equation", + "content": "W^t" + }, + { + "bbox": [ + 305, + 508, + 545, + 533 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 359, + 538, + 545, + 571 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 538, + 545, + 571 + ], + "spans": [ + { + "bbox": [ + 359, + 538, + 545, + 571 + ], + "type": "interline_equation", + "content": "F _ {\\text {l i n e a r}} ^ {t} (\\mathbf {x}) = \\frac {\\sum_ {i = 1} ^ {K} S _ {i} ^ {t} (\\mathbf {x}) W _ {i , \\mathbf {x}} ^ {t}}{\\sum_ {i = 1} ^ {K} W _ {i , \\mathbf {x}} ^ {t}}, \\tag {3}", + "image_path": "031a89a0d3fd918ef92ea9f04ab4bf4181e0ad8bff6d555d9ba422b5f31d8304.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 578, + 451, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 578, + 451, + 589 + ], + "spans": [ + { + "bbox": [ + 305, + 578, + 451, + 589 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 578, + 451, + 589 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 305, + 578, + 451, + 589 + ], + "type": "text", + "content": " is an arbitrary 2D position." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 589, + 545, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 589, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 305, + 589, + 545, + 613 + ], + "type": "text", + "content": "CNN Decoder further adds two convolutional layers to the 2D feature map formed by Equation 3:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 357, + 618, + 545, + 652 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 357, + 618, + 545, + 652 + ], + "spans": [ + { + "bbox": [ + 357, + 618, + 545, + 652 + ], + "type": "interline_equation", + "content": "F _ {\\mathrm {C N N}} ^ {t} = \\operatorname {C N N} \\left(\\frac {\\sum_ {i = 1} ^ {K} S _ {i} ^ {t} W _ {i , :} ^ {t}}{\\sum_ {i = 1} ^ {K} W _ {i , :} ^ {t}}\\right). \\tag {4}", + "image_path": "8da21f6498bb6e16cc6b9c9c096a89287f9f4a998e3a12ef8f76d25c7c539a45.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 658, + 545, + 692 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 658, + 545, + 692 + ], + "spans": [ + { + "bbox": [ + 305, + 658, + 545, + 692 + ], + "type": "text", + "content": "Transformer Decoder decodes the feature by querying the slot representation with a 2D positional embedding through a transformer decoder:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 345, + 700, + 545, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 700, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 345, + 700, + 545, + 715 + ], + "type": "interline_equation", + "content": "F _ {\\text {t r a n s f o r m e r}} ^ {t} = \\operatorname {T r a n s f o r m e r} \\left(P, S ^ {t}, S ^ {t}\\right), \\tag {5}", + "image_path": "f646971d59a876532c1c516eb4b9795b0455bf73a3d778695a42faf7d0898587.jpg" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22974" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 48, + 88, + 287, + 177 + ], + "blocks": [ + { + "bbox": [ + 53, + 76, + 185, + 86 + ], + "lines": [ + { + "bbox": [ + 53, + 76, + 185, + 86 + ], + "spans": [ + { + "bbox": [ + 53, + 76, + 185, + 86 + ], + "type": "text", + "content": "Algorithm 1: Perceiver decoder" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 48, + 88, + 287, + 177 + ], + "lines": [ + { + "bbox": [ + 48, + 88, + 287, + 177 + ], + "spans": [ + { + "bbox": [ + 48, + 88, + 287, + 177 + ], + "type": "text", + "content": "Perceiver " + }, + { + "bbox": [ + 48, + 88, + 287, + 177 + ], + "type": "inline_equation", + "content": "(S^t,P)" + }, + { + "bbox": [ + 48, + 88, + 287, + 177 + ], + "type": "text", + "content": " .. " + }, + { + "bbox": [ + 48, + 88, + 287, + 177 + ], + "type": "inline_equation", + "content": "S^t\\to \\mathrm{Norm}(S^t)" + }, + { + "bbox": [ + 48, + 88, + 287, + 177 + ], + "type": "inline_equation", + "content": "\\hat{S}^t =" + }, + { + "bbox": [ + 48, + 88, + 287, + 177 + ], + "type": "text", + "content": " SelfAttention " + }, + { + "bbox": [ + 48, + 88, + 287, + 177 + ], + "type": "inline_equation", + "content": "(S^t) + S^t" + }, + { + "bbox": [ + 48, + 88, + 287, + 177 + ], + "type": "inline_equation", + "content": "\\hat{S}^t\\rightarrow \\mathrm{Norm}(\\hat{S}^t)" + }, + { + "bbox": [ + 48, + 88, + 287, + 177 + ], + "type": "inline_equation", + "content": "\\tilde{S}^t = \\mathrm{MLP}(\\hat{S}^t) + \\hat{S}^t" + }, + { + "bbox": [ + 48, + 88, + 287, + 177 + ], + "type": "inline_equation", + "content": "\\tilde{S}^t\\rightarrow \\mathrm{Norm}(\\tilde{S}^t)\\quad P\\rightarrow \\mathrm{Norm}(P)" + }, + { + "bbox": [ + 48, + 88, + 287, + 177 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{Perceiver}}^{t} =" + }, + { + "bbox": [ + 48, + 88, + 287, + 177 + ], + "type": "text", + "content": " CrossAttention " + }, + { + "bbox": [ + 48, + 88, + 287, + 177 + ], + "type": "inline_equation", + "content": "(P,\\tilde{S}^{t},\\tilde{S}^{t})" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "code_body" + } + ], + "index": 1, + "sub_type": "algorithm" + }, + { + "bbox": [ + 46, + 201, + 288, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 201, + 288, + 224 + ], + "spans": [ + { + "bbox": [ + 46, + 201, + 288, + 224 + ], + "type": "text", + "content": "where the query " + }, + { + "bbox": [ + 46, + 201, + 288, + 224 + ], + "type": "inline_equation", + "content": "P\\in \\mathbb{R}^{N\\times d_p}" + }, + { + "bbox": [ + 46, + 201, + 288, + 224 + ], + "type": "text", + "content": " is a learnable positional embedding." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 226, + 288, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 226, + 288, + 345 + ], + "spans": [ + { + "bbox": [ + 46, + 226, + 288, + 345 + ], + "type": "text", + "content": "Compared with the previous two linear decoders, the transformer decoder further considers the global connections between the slot features and the input query, so that it will form a more powerful feature map. However, an obvious limitation is that the transformer decoder applies self-attention to the input positional query, which is 1) redundant since the positional embedding itself is learnable, and 2) not computationally efficient therefore limits the scalability of the whole model. To solve this limitation, we further propose the perceiver decoder." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 346, + 287, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 346, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 46, + 346, + 287, + 430 + ], + "type": "text", + "content": "Perceiver Decoder is inspired by [29], which designs a computationally efficient architecture to handle arbitrary outputs in addition to arbitrary inputs. Different from the original architecture, we only add a self-attention layer to the slot representations, followed by a cross-attention layer for the output and positional embedding queries. The whole procedure is illustrated in Algorithm 1." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 430, + 288, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 430, + 288, + 478 + ], + "spans": [ + { + "bbox": [ + 46, + 430, + 288, + 478 + ], + "type": "text", + "content": "After forming the 2D feature map " + }, + { + "bbox": [ + 46, + 430, + 288, + 478 + ], + "type": "inline_equation", + "content": "F^t" + }, + { + "bbox": [ + 46, + 430, + 288, + 478 + ], + "type": "text", + "content": ", we decode it to a reconstruction space with a CNN-based decoder, except for when vector-quantized space is used. As we will show next, an additional decoder is redundant in this case." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 488, + 171, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 488, + 171, + 502 + ], + "spans": [ + { + "bbox": [ + 47, + 488, + 171, + 502 + ], + "type": "text", + "content": "3.3. Reconstruction Space" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 507, + 289, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 507, + 289, + 662 + ], + "spans": [ + { + "bbox": [ + 46, + 507, + 289, + 662 + ], + "type": "text", + "content": "There are four choices of reconstruction space in our framework. The plain RGB space contains the most information but also is the most complex to resolve the object/background ambiguity. The flow and depth spaces, as shown in Figure 2 (right), are more structured. Reconstructing in these spaces makes the grouping problem easier. However, there are still shortcomings, e.g. non-moving objects are not captured in the flow space and depth cannot easily distinguish between-objects that are near each other. Moreover, these two spaces are not as informative as the RGB space. In addition, we introduce the VQ-space, which is end-to-end trainable and is both structured and informative. We describe the VQ-space below." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": "Vector-quantized reconstruction space. Different from the other three reconstruction spaces, here we do not predict the reconstruction directly, but rather supervise the feature map " + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "inline_equation", + "content": "F^{t}" + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": " to match the latent embedding space of VQ-VAE." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "type": "text", + "content": "In particular, following [58], we define a latent embedding space " + }, + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "type": "text", + "content": " as the set of " + }, + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "type": "text", + "content": " vectors " + }, + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "type": "inline_equation", + "content": "e_i" + }, + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "type": "text", + "content": " of dimension " + }, + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "type": "inline_equation", + "content": "d_{vq}" + }, + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "type": "inline_equation", + "content": "E = \\{e_i\\in \\mathbb{R}^{d_{vq}}|i = 1,2,\\dots ,M\\}" + }, + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "type": "text", + "content": ". Given an input image " + }, + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "type": "inline_equation", + "content": "I^t" + }, + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "type": "text", + "content": ", VQ-VAE first processes it with an encoder to get the output " + }, + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "type": "inline_equation", + "content": "z_e^t = \\mathrm{Encoder}_{\\mathrm{VQ}}(I^t)" + }, + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "type": "text", + "content": ", and then the discrete latent variables " + }, + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "type": "text", + "content": " are calculated by a nearest neighborhood search among the discrete feature set " + }, + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 304, + 72, + 547, + 156 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 318, + 163, + 547, + 179 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 163, + 547, + 179 + ], + "spans": [ + { + "bbox": [ + 318, + 163, + 547, + 179 + ], + "type": "interline_equation", + "content": "z _ {q} ^ {t} (\\mathbf {x}) = e _ {k}, \\quad \\text {w h e r e} k = \\operatorname {a r g m i n} _ {j} \\| z _ {e} ^ {t} (\\mathbf {x}) - e _ {j} \\| _ {2}, \\tag {6}", + "image_path": "ab3e6d50996f410178d08f2b7eac70b2af1bce64a617b6564bb206393c8570c6.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 186, + 547, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 186, + 547, + 221 + ], + "spans": [ + { + "bbox": [ + 304, + 186, + 547, + 221 + ], + "type": "text", + "content": "and " + }, + { + "bbox": [ + 304, + 186, + 547, + 221 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 186, + 547, + 221 + ], + "type": "text", + "content": " is an arbitrary 2D position. The reconstructed image " + }, + { + "bbox": [ + 304, + 186, + 547, + 221 + ], + "type": "inline_equation", + "content": "\\hat{I}^t" + }, + { + "bbox": [ + 304, + 186, + 547, + 221 + ], + "type": "text", + "content": " is then decoded by " + }, + { + "bbox": [ + 304, + 186, + 547, + 221 + ], + "type": "inline_equation", + "content": "\\hat{I}^t = \\mathrm{Decoder}_{\\mathrm{VQ}}(z_q^t)" + }, + { + "bbox": [ + 304, + 186, + 547, + 221 + ], + "type": "text", + "content": ". The objective of VQ-VAE is:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 310, + 228, + 546, + 242 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 228, + 546, + 242 + ], + "spans": [ + { + "bbox": [ + 310, + 228, + 546, + 242 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {V Q V A E}} = \\log P \\left(I ^ {t} \\mid z _ {q} ^ {t}\\right) + \\left\\| s g \\left[ z _ {e} ^ {t} \\right] - z _ {q} ^ {t} \\right\\| _ {2} + \\left\\| s g \\left[ z _ {q} ^ {t} \\right] - z _ {e} ^ {t} \\right\\| _ {2}, \\tag {7}", + "image_path": "36d280143d61b13c1bea70b738448215d19aa4eb757b18489047f006987c0fa4.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 247, + 476, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 247, + 476, + 259 + ], + "spans": [ + { + "bbox": [ + 304, + 247, + 476, + 259 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 247, + 476, + 259 + ], + "type": "inline_equation", + "content": "sg[\\cdot]" + }, + { + "bbox": [ + 304, + 247, + 476, + 259 + ], + "type": "text", + "content": " is the stop-gradient operation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 259, + 547, + 296 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 259, + 547, + 296 + ], + "spans": [ + { + "bbox": [ + 304, + 259, + 547, + 296 + ], + "type": "text", + "content": "Then we use the quantized feature map " + }, + { + "bbox": [ + 304, + 259, + 547, + 296 + ], + "type": "inline_equation", + "content": "z_{q}^{t}" + }, + { + "bbox": [ + 304, + 259, + 547, + 296 + ], + "type": "text", + "content": " as the target signal for the slot feature map " + }, + { + "bbox": [ + 304, + 259, + 547, + 296 + ], + "type": "inline_equation", + "content": "F^{t}" + }, + { + "bbox": [ + 304, + 259, + 547, + 296 + ], + "type": "text", + "content": ". The final objective of VQ-VAE and the VQ reconstruction is:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 310, + 303, + 546, + 319 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 303, + 546, + 319 + ], + "spans": [ + { + "bbox": [ + 310, + 303, + 546, + 319 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {V Q}} = \\mathcal {L} _ {\\mathrm {V Q V A E}} + \\| s g [ F ^ {t} ] - z _ {q} ^ {t} \\| _ {2} + \\| s g [ z _ {q} ^ {t} ] - F ^ {t} \\| _ {2}. \\tag {8}", + "image_path": "cc3bceb1f469560b76c778fa085bc78b73e5e4f5f33ec82ae74ef91a86c17531.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 327, + 547, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 327, + 547, + 448 + ], + "spans": [ + { + "bbox": [ + 304, + 327, + 547, + 448 + ], + "type": "text", + "content": "Motion-guided token representation. The last term in Equation 8, " + }, + { + "bbox": [ + 304, + 327, + 547, + 448 + ], + "type": "inline_equation", + "content": "||sg[F^t] - z_q^t||_2" + }, + { + "bbox": [ + 304, + 327, + 547, + 448 + ], + "type": "text", + "content": ", enables the motion signal from slot learning to jointly optimize the token space through the output of the slot decoder. Furthermore, the token representation and the motion cues build a connection linked by the slot learning, thus enabling the emergence of interpretable object-specific mid-level features of tokens. In addition, reconstructing in a more compact token space also benefits the model by better utilizing the motion signal to achieve an improved slot representation and temporal consistency." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 454, + 391, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 454, + 391, + 467 + ], + "spans": [ + { + "bbox": [ + 306, + 454, + 391, + 467 + ], + "type": "text", + "content": "3.4. Optimization" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 475, + 547, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 475, + 547, + 545 + ], + "spans": [ + { + "bbox": [ + 304, + 475, + 547, + 545 + ], + "type": "text", + "content": "Token contrastive constraint. The goal of reconstructing in the VQ-space is that it is more compact and of lower variation compared with the RGB space. To make the VQ-space more structured, we add an optional contrastive constraint below to the vector space, which increases the independence between latent vectors:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 345, + 554, + 546, + 568 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 554, + 546, + 568 + ], + "spans": [ + { + "bbox": [ + 345, + 554, + 546, + 568 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {c o n t r a s t i v e}} = \\left| \\left| \\mathbb {I} - \\operatorname {s o f t m a x} \\left(E \\cdot E ^ {T}\\right) \\right| \\right|, \\tag {9}", + "image_path": "21327ddfe8acef7763efa7d97eabd3a4a3540ab6c256e7ea89b28d8655210012.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 575, + 545, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 575, + 545, + 599 + ], + "spans": [ + { + "bbox": [ + 304, + 575, + 545, + 599 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 575, + 545, + 599 + ], + "type": "inline_equation", + "content": "\\mathbb{I}" + }, + { + "bbox": [ + 304, + 575, + 545, + 599 + ], + "type": "text", + "content": " is the identity matrix and " + }, + { + "bbox": [ + 304, + 575, + 545, + 599 + ], + "type": "inline_equation", + "content": "E\\in \\mathbb{R}^{N\\times d_{vq}}" + }, + { + "bbox": [ + 304, + 575, + 545, + 599 + ], + "type": "text", + "content": " is the matrix of the feature embedding space " + }, + { + "bbox": [ + 304, + 575, + 545, + 599 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 575, + 545, + 599 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 600, + 547, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 600, + 547, + 635 + ], + "spans": [ + { + "bbox": [ + 304, + 600, + 547, + 635 + ], + "type": "text", + "content": "The final loss function is a combination of the reconstruction objective, the motion objective, and the optional contrastive constraint:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 324, + 644, + 546, + 658 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 644, + 546, + 658 + ], + "spans": [ + { + "bbox": [ + 324, + 644, + 546, + 658 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\lambda \\mathcal {L} _ {\\text {m o t i o n}} + \\mathcal {L} _ {\\text {r e c o n}} + \\lambda_ {c} \\mathbb {1} _ {\\mathrm {V Q}} \\mathcal {L} _ {\\text {c o n t r a s t i v e}}, \\tag {10}", + "image_path": "a8fdd541d486883ee97d63b7aa925d6682ca5257e44763c3f4f2830ae9898894.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_{c}" + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": " are weighting factors and " + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathbb{1}_{\\mathrm{VQ}}" + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": " is an indicator function. For the reconstruction loss, we set " + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{recon}} = \\mathcal{L}_{\\mathrm{VQ}}" + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": " when performing reconstruction in the VQ-space. Otherwise, we use an " + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": " loss for reconstruction in the other three spaces." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22975" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 72, + 124, + 152 + ], + "blocks": [ + { + "bbox": [ + 50, + 72, + 124, + 152 + ], + "lines": [ + { + "bbox": [ + 50, + 72, + 124, + 152 + ], + "spans": [ + { + "bbox": [ + 50, + 72, + 124, + 152 + ], + "type": "image", + "image_path": "58fe986f8c7ec3db75011d693a1a648516c53d1132b86d6cd3e8f5323688aa63.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 154, + 288, + 210 + ], + "lines": [ + { + "bbox": [ + 46, + 154, + 288, + 210 + ], + "spans": [ + { + "bbox": [ + 46, + 154, + 288, + 210 + ], + "type": "text", + "content": "Figure 3. Frame samples from the video datasets used in our experiments. MOVi [21] (left) is a multi-object video dataset created by simulating rigid body dynamics. TRI-PD [4] (middle) is a collection of photo-realistic, synthetic driving videos. KITTI [19] (right) is a real-world benchmark with city driving scenes." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 129, + 72, + 201, + 152 + ], + "blocks": [ + { + "bbox": [ + 129, + 72, + 201, + 152 + ], + "lines": [ + { + "bbox": [ + 129, + 72, + 201, + 152 + ], + "spans": [ + { + "bbox": [ + 129, + 72, + 201, + 152 + ], + "type": "image", + "image_path": "77f1baf06ed23c0c31053cf62f6c4478f10efd24de4c939d4d0f0fa7443bca4c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 207, + 72, + 284, + 152 + ], + "blocks": [ + { + "bbox": [ + 207, + 72, + 284, + 152 + ], + "lines": [ + { + "bbox": [ + 207, + 72, + 284, + 152 + ], + "spans": [ + { + "bbox": [ + 207, + 72, + 284, + 152 + ], + "type": "image", + "image_path": "5e8f6aa14e950aa5c277583a1f42bf4622fe94107069b50b2ba639b9c6d2ee04.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 223, + 190, + 236 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 223, + 190, + 236 + ], + "spans": [ + { + "bbox": [ + 47, + 223, + 190, + 236 + ], + "type": "text", + "content": "4. Experimental Evaluation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 243, + 173, + 257 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 243, + 173, + 257 + ], + "spans": [ + { + "bbox": [ + 47, + 243, + 173, + 257 + ], + "type": "text", + "content": "4.1. Experimental Settings" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 265, + 287, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 265, + 287, + 289 + ], + "spans": [ + { + "bbox": [ + 46, + 265, + 287, + 289 + ], + "type": "text", + "content": "Benchmarks. We evaluate our approach on three popular benchmarks with different complexity (Figure 3)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 290, + 287, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 290, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 46, + 290, + 287, + 396 + ], + "type": "text", + "content": "MOVi [21] is a synthetic multi-object video dataset, created by simulating rigid body dynamics. We use this benchmark to ablate and analyze our model architecture. Following previous works [15, 53], we use the most complex subset, MOVi-E, for evaluation. MOVi-E contains both moving and static objects (maximum 20 objects) with linear random camera motion. The resolution of this dataset is " + }, + { + "bbox": [ + 46, + 290, + 287, + 396 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 46, + 290, + 287, + 396 + ], + "type": "text", + "content": " and each video contains 24 frames with a sample rate of 12 FPS. We use standard train and test split for MOVi-E." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 399, + 287, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 399, + 287, + 505 + ], + "spans": [ + { + "bbox": [ + 46, + 399, + 287, + 505 + ], + "type": "text", + "content": "TRI-PD [4,25] is a synthetic dataset based on street driving scenarios collected by using a state-of-the-art synthetic data generation service [1]. The dataset comes with a collection of accurate annotations (e.g. for flow or depth), which we use to ablate the impact of additional information and its quality on various models. Each video in TRI-PD is 10 seconds captured at 20 FPS. There are 924 videos in the training set and 51 videos in the test set. We crop and resize each frame to the resolution of " + }, + { + "bbox": [ + 46, + 399, + 287, + 505 + ], + "type": "inline_equation", + "content": "480 \\times 968" + }, + { + "bbox": [ + 46, + 399, + 287, + 505 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 507, + 287, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 507, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 46, + 507, + 287, + 555 + ], + "type": "text", + "content": "KITTI [19] is a real-world benchmark with city driving scenes. We use the whole 151 KITTI videos for training and the instance segmentation subsets with 200 single-frame images for evaluation. The frames are resized to " + }, + { + "bbox": [ + 46, + 507, + 287, + 555 + ], + "type": "inline_equation", + "content": "368 \\times 1248" + }, + { + "bbox": [ + 46, + 507, + 287, + 555 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 558, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 288, + 713 + ], + "type": "text", + "content": "Baselines. We compare our methods against the most recent learning-based object discovery models. In particular, SAVi [36] and SAVi++ [15] are two direct extensions of the original slot attention [39] using optical flow and depth to facilitate object-centric learning. Alternatively, STEVE [53] uses a more powerful transformer decoder to enhance the object discovery performance. Bao et al. [4], Karazija et al. [33] utilize motion cues to guide object discovery. In addition, in the supplementary, we compare our approach to [51], which is a concurrent work that relies on ImageNet [12] pretraining [9]. For [4, 15, 36, 53], we use either the official code or public implementation (see supplementary for details). Due to the lack of implementation, we reuse reported results" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 307, + 70, + 545, + 163 + ], + "blocks": [ + { + "bbox": [ + 307, + 70, + 545, + 163 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 545, + 163 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 545, + 163 + ], + "type": "image", + "image_path": "88da2141d7738868c42fb1955ea8e0fc4db66be8716a02f1b177e7e3c52a3781.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 164, + 545, + 209 + ], + "lines": [ + { + "bbox": [ + 305, + 164, + 545, + 209 + ], + "spans": [ + { + "bbox": [ + 305, + 164, + 545, + 209 + ], + "type": "text", + "content": "Figure 4. Object discovery in different reconstruction spaces on the evaluation set of MOVi-E. Compared with the other reconstruction spaces, the model trained using the VQ-space demonstrates a more accurate performance and better temporal consistency." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 236, + 487, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 236, + 487, + 247 + ], + "spans": [ + { + "bbox": [ + 305, + 236, + 487, + 247 + ], + "type": "text", + "content": "with the two very recent approaches [33, 51]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 255, + 545, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 255, + 545, + 339 + ], + "spans": [ + { + "bbox": [ + 304, + 255, + 545, + 339 + ], + "type": "text", + "content": "Evaluation Metrics. Following prior works, we use the Foreground Adjusted Rand Index (FG. ARI) to evaluate the performance of the models, which captures how well the predicted segmentation masks match ground-truth masks in a permutation-invariant fashion. Notice that FG. ARI is measured in terms of the whole video, so that the temporal consistency of the slots is considered with this metric." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 347, + 546, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 347, + 546, + 574 + ], + "spans": [ + { + "bbox": [ + 304, + 347, + 546, + 574 + ], + "type": "text", + "content": "Implementation Details. We use the same ResNet-18 ConvGRU encoder backbone for all the compared methods following [4]. We set the number of slots as 24 for MOVi-E and 45 for PD and KITTI based on the maximal number of objects in these datasets. On TRI-PD and KITTI, we further downsample the images by 4 for SAVi and SAVi++ to fit the GPU memory. Notice that, the other methods also produce downsampled slot masks, leading to the evaluation under the same resolution for all the methods. All the models are trained for 500 epochs using Adam [34]. During training, we train all the models with a randomly sampled video sequence of a fixed length (6, 5, 5 for MOVi-E, TRI-PD, and KITTI respectively). During the inference time, all the models are evaluated frame by frame until the end. For the FG. ARI measurement on TRI-PD, we discard any instance labels covering an area of " + }, + { + "bbox": [ + 304, + 347, + 546, + 574 + ], + "type": "inline_equation", + "content": "0.5\\%" + }, + { + "bbox": [ + 304, + 347, + 546, + 574 + ], + "type": "text", + "content": " or less of the first sampled video frame following [15]. We also evaluate FG. ARI at 5 FPS due to memory limitation. We use batch size 64 for MOVi-E, and 8 for TRI-PD and KITTI." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "text", + "content": "Our method and Bao et al. [4] require motion segmentation signals. We apply [11], a powerful method pre-trained on the toy FlyingThings3D dataset [40], taking the ground-truth flow (MOVi-E) or RAFT [57] flow as the input. For SAVi and " + }, + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{SAV_i + + }" + }, + { + "bbox": [ + 304, + 582, + 547, + 713 + ], + "type": "text", + "content": ", we do not use the first-frame bounding box supervision for a fair comparison. We generate the optical flow and depth annotation for KITTI using two state-of-the-art methods, RAFT [57] and VIDAR [25]. We also use them for ablations on TRI-PD. More details about the hyper-parameters, training scheme, and annotation generation are provided in the supplementary." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "22976" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 72, + 70, + 260, + 174 + ], + "blocks": [ + { + "bbox": [ + 72, + 70, + 260, + 174 + ], + "lines": [ + { + "bbox": [ + 72, + 70, + 260, + 174 + ], + "spans": [ + { + "bbox": [ + 72, + 70, + 260, + 174 + ], + "type": "table", + "html": "
MotionSpaceDecoderFG. ARI
XVQLinear14.6
XVQLinear-CNN16.3
XVQTransformer40.1
XVQPerceiver52.4
XRGBPerceiver49.2
XFlowPerceiver36.3
XDepthPerceiver18.4
XFlow + DepthPerceiver38.0
XVQ (flow)Perceiver44.6
", + "image_path": "68e7b927209bef216bf007d74f96bc38a4cc2ecce589d06941c406cddc3c4387.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 266, + 196, + 280 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 266, + 196, + 280 + ], + "spans": [ + { + "bbox": [ + 47, + 266, + 196, + 280 + ], + "type": "text", + "content": "4.2. Object Discovery on MOVi" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 286, + 287, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 286, + 287, + 334 + ], + "spans": [ + { + "bbox": [ + 46, + 286, + 287, + 334 + ], + "type": "text", + "content": "We use MOVi-E dataset to ablate our model with different choices of slot decoder, and reconstruction spaces, aiming to elucidate their roles in object discovery. We also compare the performance with state-of-the-art models in this section." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 337, + 288, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 337, + 288, + 492 + ], + "spans": [ + { + "bbox": [ + 46, + 337, + 288, + 492 + ], + "type": "text", + "content": "Set up. We first fix the reconstruction space as the VQ-space and ablate the decoder architectures (first 4 lines in Table 1), and then we fix the decoder architecture with the best component and reconstruct in different reconstruction space (Line 5 to 9 in Table 1). Since SAVi++ [15] argues that reconstructing in the combined space of flow and depth is the key to their success, we add a variant " + }, + { + "bbox": [ + 46, + 337, + 288, + 492 + ], + "type": "inline_equation", + "content": "Flow + Depth" + }, + { + "bbox": [ + 46, + 337, + 288, + 492 + ], + "type": "text", + "content": ", which also reconstructs in this combined space. Furthermore, as VQ-VAE can be trained to reconstruct the image in the flow space as well, we also report a " + }, + { + "bbox": [ + 46, + 337, + 288, + 492 + ], + "type": "inline_equation", + "content": "VQ" + }, + { + "bbox": [ + 46, + 337, + 288, + 492 + ], + "type": "text", + "content": " (flow) variant for completeness. We do not include motion cues for the architecture analysis. A more comprehensive ablation is reported in the supplementary." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 495, + 287, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 495, + 287, + 615 + ], + "spans": [ + { + "bbox": [ + 46, + 495, + 287, + 615 + ], + "type": "text", + "content": "Decoder analysis. By comparing the results on the first four lines in Table 1, similar to [53], we find that the capacity of the decoder indeed plays a key role in the object discovery performance. A simple linear decoder fails, but adding an additional CNN decoder can bring limited performance gains. Introducing powerful transformer-based decoders, on the other hand, greatly improves the object discovery capabilities of the model. Between the two transformer variants, the more advanced perceiver decoder achieves better results, while being more computationally efficient." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "content": "Reconstruction space analysis. For the four reconstruction spaces, the VQ-space yields the best object discovery performance. We analyze the learned slots and token representations in Figure 4 and make the following discoveries. Firstly, the model trained using the VQ-space better separates the objects from the background and shows stronger temporal consistency. This is due to the more structured, compact, and lower-variance reconstruction space provided" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 360, + 70, + 489, + 171 + ], + "blocks": [ + { + "bbox": [ + 46, + 178, + 288, + 245 + ], + "lines": [ + { + "bbox": [ + 46, + 178, + 288, + 245 + ], + "spans": [ + { + "bbox": [ + 46, + 178, + 288, + 245 + ], + "type": "text", + "content": "Table 1. Model architecture analysis with different choices of the slot decoder and reconstruction space on MOVi-E. Perceiver decoder + VQ reconstruction space yields the best performance, indicating that (1) the capacity of the decoder plays a key role in object discovery; (2) learnable, vector-quantized reconstruction space outperforms fixed alternatives like depth or flow." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 360, + 70, + 489, + 171 + ], + "lines": [ + { + "bbox": [ + 360, + 70, + 489, + 171 + ], + "spans": [ + { + "bbox": [ + 360, + 70, + 489, + 171 + ], + "type": "table", + "html": "
ModelFG. ARI
Bao et al.51.6
STEVE51.2
SAVi28.1
SAVi++31.7
MoTok (no motion)52.4
MoTok63.8
SAVi (G.T. box)53.4
SAVi++ (G.T. box)84.1
", + "image_path": "78987071883cdceba9c7ac07a6b1fa9355cdeef9b64ea16d3da18fc038d70020.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 175, + 547, + 220 + ], + "lines": [ + { + "bbox": [ + 304, + 175, + 547, + 220 + ], + "spans": [ + { + "bbox": [ + 304, + 175, + 547, + 220 + ], + "type": "text", + "content": "Table 2. Comparison to the state-of-the-art object discovery approaches on the validation sets of MOVi-E using FG. ARI. Our approach outperforms all the recent methods with the help of motion cues and vector quantization." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 230, + 547, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 230, + 547, + 253 + ], + "spans": [ + { + "bbox": [ + 304, + 230, + 547, + 253 + ], + "type": "text", + "content": "by the quantized features, which simplifies the task of grouping, compared to the raw RGB space." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 254, + 547, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 254, + 547, + 384 + ], + "spans": [ + { + "bbox": [ + 304, + 254, + 547, + 384 + ], + "type": "text", + "content": "Secondly, by comparing Flow and " + }, + { + "bbox": [ + 304, + 254, + 547, + 384 + ], + "type": "inline_equation", + "content": "\\text{Flow+Depth}" + }, + { + "bbox": [ + 304, + 254, + 547, + 384 + ], + "type": "text", + "content": " variants, we find that the latter indeed carries more information, allowing the model to better group the objects, which is consistent with [15]. However, interestingly, we also find that given a sufficiently strong decoder, reconstructing in depth or flow space does not bring further improvements compared to RGB, and can even decrease the performance. Finally, comparing Flow and " + }, + { + "bbox": [ + 304, + 254, + 547, + 384 + ], + "type": "inline_equation", + "content": "\\text{VQ}(\\text{flow})" + }, + { + "bbox": [ + 304, + 254, + 547, + 384 + ], + "type": "text", + "content": ", we find that although adding the quantization improves performance significantly, this variant still lags behind reconstructing in the raw RGB space, reinforcing our previous conclusions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 387, + 547, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 387, + 547, + 542 + ], + "spans": [ + { + "bbox": [ + 304, + 387, + 547, + 542 + ], + "type": "text", + "content": "Comparison to the state of the art. In Table 2, we compare our approach to the state-of-the-art methods. Our model achieves the best performance among them. Additionally, compared to SAVi and SAVi++, even our variants that reconstructs in the same space outperform them, indicating that the perceiver decoder and single-stage decoding strategy are the optimal choices. Our model also outperforms STEVE, due to the fact that, in contrast to DVAE, the codebook in VQ-VAE can be jointly optimized with slot learning (see Equation 8). Compared with Bao et al., our model shows better performance even without motion cues used by that method, but further incorporating them into our approach allows it to achieve top results." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 543, + 547, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 543, + 547, + 602 + ], + "spans": [ + { + "bbox": [ + 304, + 543, + 547, + 602 + ], + "type": "text", + "content": "Notice that, we compare with the variants of SAVi and SAVi++ that do not rely on ground-truth bounding boxes at test time. Their published numbers are included at the bottom of the table for reference, but they are not comparable to the other unsupervised methods." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 609, + 545, + 622 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 609, + 545, + 622 + ], + "spans": [ + { + "bbox": [ + 305, + 609, + 545, + 622 + ], + "type": "text", + "content": "4.3. Object Discovery with Realistic Driving Videos" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "type": "text", + "content": "Set up. Firstly, following [4], we capitalize on the ground-truth annotations available in TRI-PD, and evaluate the importance of the quality of external signals for various approaches. In particular, we separately evaluate using ground-truth (GT) flow, depth, and motion segmentation, as well as estimating those using state-of-the-art algorithms [11,25,57]. In addition, we report a variant STEVE-" + }, + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 629, + 547, + 713 + ], + "type": "text", + "content": " for which we add" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "22977" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 70, + 545, + 181 + ], + "blocks": [ + { + "bbox": [ + 51, + 70, + 545, + 181 + ], + "lines": [ + { + "bbox": [ + 51, + 70, + 545, + 181 + ], + "spans": [ + { + "bbox": [ + 51, + 70, + 545, + 181 + ], + "type": "image", + "image_path": "1e22679eb8d0f042d634ca737a3968d2d9f3f7d2ca9bcca49bb70b68632f8324.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 182, + 546, + 216 + ], + "lines": [ + { + "bbox": [ + 46, + 182, + 546, + 216 + ], + "spans": [ + { + "bbox": [ + 46, + 182, + 546, + 216 + ], + "type": "text", + "content": "Figure 5. Reconstruction, slot representation, and token masks for [4], STEVE-m [53], and our method. We use ground-truth motion cues for supervision, and visualize top-10 masks excluding background slots. MoTok effectively leverages the synergy between motion and tokenization, enabling the emergence of interpretable object-specific mid-level features, which simplify the problem of object discovery." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 72, + 221, + 260, + 299 + ], + "blocks": [ + { + "bbox": [ + 72, + 221, + 260, + 299 + ], + "lines": [ + { + "bbox": [ + 72, + 221, + 260, + 299 + ], + "spans": [ + { + "bbox": [ + 72, + 221, + 260, + 299 + ], + "type": "table", + "html": "
ModelSignalGPU Mem. (GB)FG. ARI
GTEST
Bao et al.Motion21.350.434.7
SAViFlow17.317.914.2
SAVi++Depth + Flow17.518.415.3
STEVE-66.712.5-
STEVE-mMotion68.245.732.2
MoTokMotion10.960.655.0
", + "image_path": "5bd0306cb78f4e2d829440cb2142cf17ecb273bd11b6cc46b5c31a07f2f78dc5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 303, + 288, + 358 + ], + "lines": [ + { + "bbox": [ + 46, + 303, + 288, + 358 + ], + "spans": [ + { + "bbox": [ + 46, + 303, + 288, + 358 + ], + "type": "text", + "content": "Table 3. Object discovery evaluation of different models on TRI-PD dataset. GT denotes the ground-truth guidance given to each model and EST denotes the estimated cues. State-of-the-art models fail to work without motion cues; introducing token representation helps our model better utilize the motion signal." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 383, + 287, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 383, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 46, + 383, + 287, + 407 + ], + "type": "text", + "content": "the same motion segmentation supervision as to [4] and our approach for a fair comparison." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 409, + 287, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 409, + 287, + 516 + ], + "spans": [ + { + "bbox": [ + 46, + 409, + 287, + 516 + ], + "type": "text", + "content": "From Table 3, we can observe that our method strongly outperforms all the baselines. Additionally, we find that even with the ground-truth flow or depth, SAVI and SAVI++ do not work well for realistic videos with complex background and crowded scenes. Without motion cues, even equipped with a powerful transformer decoder, STEVE fails to work in this challenging setting as well. Finally, our model achieves better performance compared to Bao et al. with both GT and estimated motion segments, indicating its better robustness." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 520, + 288, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 520, + 288, + 639 + ], + "spans": [ + { + "bbox": [ + 46, + 520, + 288, + 639 + ], + "type": "text", + "content": "Impact of motion-guided tokens: In Figure 5, we visualize the RGB reconstruction, slot representation, and token masks (if possible) for Bao et al., STEVE-m, and our MoTok. There are several key observations. Firstly, compared with STEVE-m, MoTok enables the emergence of more interpretable mid-level features, even though both models are trained with motion cues and tokenization. This is due to the interplay between object and slot discovery in our end-to-end framework, whereas STEVE-m trains DVAE separately from the rest of the model." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 641, + 288, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 712 + ], + "type": "text", + "content": "Secondly, our model achieves the best RGB reconstruction result, even for the latter frames. The high quality of the reconstruction indicates that the model can better take advantage of the appearance signal to optimize the slot representation. Finally, our model demonstrates stronger temporal consistency compared with the baselines, thanks to the struc" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 308, + 222, + 543, + 269 + ], + "blocks": [ + { + "bbox": [ + 308, + 222, + 543, + 269 + ], + "lines": [ + { + "bbox": [ + 308, + 222, + 543, + 269 + ], + "spans": [ + { + "bbox": [ + 308, + 222, + 543, + 269 + ], + "type": "image", + "image_path": "75203d917d95230b77c51b97a02aa668a0e9ff61b7d6c1d31c443472f0b2aa52.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 270, + 546, + 304 + ], + "lines": [ + { + "bbox": [ + 306, + 270, + 546, + 304 + ], + "spans": [ + { + "bbox": [ + 306, + 270, + 546, + 304 + ], + "type": "text", + "content": "Figure 6. Visualization of the learned tokens for vanilla VQ-VAE and MoTok. The proposed MoTok model shows a better alignment with semantic categories." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 322, + 546, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 322, + 546, + 381 + ], + "spans": [ + { + "bbox": [ + 304, + 322, + 546, + 381 + ], + "type": "text", + "content": "tured reconstruction space. To sum up, our architecture effectively leverages the synergy between motion cues and tokenization, enabling the emergence of interpretable mid-level features, which greatly simplifies the task of object discovery." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 384, + 546, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 384, + 546, + 468 + ], + "spans": [ + { + "bbox": [ + 304, + 384, + 546, + 468 + ], + "type": "text", + "content": "Interpretable tokens. To better illustrate the interpretability of motion-guided tokens, we measured their alignment with the ground-truth semantic labels using cluster purity [50]. Vanilla VQ-VAE (no motion guidance) achieves a purity of 40.1, whereas ours reaches 65.7 (chance performance: 38.5). Qualitative comparison in Figure 6 clearly demonstrates our better semantic alignment." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 472, + 546, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 546, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 546, + 567 + ], + "type": "text", + "content": "Scalability to realistic videos: We additionally report the GPU memory usage per batch in Table 3. The per-slot decoding strategy and the use of transformer decoder limit the scalability of the other baselines. In comparison, our model achieves the best performance with the lowest memory consumption, showing a great generalization capability to real-world videos, thanks to the efficient perceiver decoder and the single-shot decoding strategy." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": "Impact of objectives. Finally, we ablate the objective design of our model with the GT annotations in Table 4. We build three variants of our model, one without the additional token contrastive constraint defined by Equation 10 (no contrastive), the second without the motion cues (no motion), and the last trained with ground-truth instance masks for all objects to indicate the performance upper bound " + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "inline_equation", + "content": "(^{*})" + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": ". By comparing the performance of these variants, we make the following observations: (1) in realistic synthetic videos even the strongest architectures fail to resolve the object/background ambiguity in the absence of motion cues, demonstrating the difficulty of object discovery tasks in the real world; (2)" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22978" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 156, + 112 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 156, + 112 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 156, + 112 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 156, + 112 + ], + "type": "table", + "html": "
ModelFG. ARI
MoTok60.6
MoTok (no contrastive)57.4
MoTok (no motion)15.6
MoTok*64.7
", + "image_path": "47e9175bfed9f7976a7e9220e04369752f643bdd669cc0a0d60fc44da9ec8280.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 166, + 70, + 286, + 142 + ], + "blocks": [ + { + "bbox": [ + 47, + 116, + 157, + 226 + ], + "lines": [ + { + "bbox": [ + 47, + 116, + 157, + 226 + ], + "spans": [ + { + "bbox": [ + 47, + 116, + 157, + 226 + ], + "type": "text", + "content": "Table 4. Ablation study on different learning signals. MoTok* is our model trained with ground-truth instance segmentation. The motion cue and the contrastive constraint help improve the capability of the proposed model, almost reaching the performance upper bound (MoTok*)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 166, + 70, + 286, + 142 + ], + "lines": [ + { + "bbox": [ + 166, + 70, + 286, + 142 + ], + "spans": [ + { + "bbox": [ + 166, + 70, + 286, + 142 + ], + "type": "table", + "html": "
ModelFG. ARI
Bao et al. [4]47.1
SAVi20.0
SAVi++23.9
STEVE11.9
Karazija et al. [33]50.8
Karazija et al. [33] (WL)51.9
MoTok64.4
", + "image_path": "6e7fbb38f7d6eef07001296af5be85c46bcd1f51989dbb349b476088a5129ceb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 162, + 146, + 288, + 224 + ], + "lines": [ + { + "bbox": [ + 162, + 146, + 288, + 224 + ], + "spans": [ + { + "bbox": [ + 162, + 146, + 288, + 224 + ], + "type": "text", + "content": "Table 5. Evaluation of object discovery in the real-world KITTI dataset. Karazija et al. [33] (WL) is a variant trained with their proposed wrapping loss. Our model achieves state-of-the-art in real-world object discovery." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 247, + 289, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 247, + 289, + 308 + ], + "spans": [ + { + "bbox": [ + 46, + 247, + 289, + 308 + ], + "type": "text", + "content": "contrastive constraint facilitates diversity during vector quantization, increasing the information content of each token, which helps reduce the object/background ambiguity; (3) the performance of our model is close to the upper bound, indicating the effectiveness of the motion-guided tokens." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 315, + 236, + 328 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 315, + 236, + 328 + ], + "spans": [ + { + "bbox": [ + 47, + 315, + 236, + 328 + ], + "type": "text", + "content": "4.4. Object Discovery in the Real World" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 335, + 287, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 335, + 287, + 383 + ], + "spans": [ + { + "bbox": [ + 46, + 335, + 287, + 383 + ], + "type": "text", + "content": "Set up. To train the object discovery model on the real-world KITTI benchmark, we first pre-train all the models on PD using ground-truth flow and depth and then fine-tune them using estimated annotations on KITTI." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 384, + 289, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 384, + 289, + 515 + ], + "spans": [ + { + "bbox": [ + 46, + 384, + 289, + 515 + ], + "type": "text", + "content": "The comparisons on KITTI are shown in Table 5 and Figure 7. We find that, consistent with our observations on TRI-PD, without motion guidance, both more powerful decoders (STEVE) and the simpler reconstruction space (SAVi and " + }, + { + "bbox": [ + 46, + 384, + 289, + 515 + ], + "type": "inline_equation", + "content": "\\mathrm{SAV_i + + }" + }, + { + "bbox": [ + 46, + 384, + 289, + 515 + ], + "type": "text", + "content": ") fail to work. However, our model still captures good dynamic object segmentation results. Compared to Bao et al., we still achieve a better FG. ARI score and a more solid segmentation mask, indicating the benefit of the model design and the motion-guided tokens. Finally, we outperform the very recent approach of Karazija et al. [33] even when they include an additional warping loss objective." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 523, + 209, + 534 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 523, + 209, + 534 + ], + "spans": [ + { + "bbox": [ + 47, + 523, + 209, + 534 + ], + "type": "text", + "content": "4.5. Limitations and Future Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 543, + 287, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 543, + 287, + 615 + ], + "spans": [ + { + "bbox": [ + 46, + 543, + 287, + 615 + ], + "type": "text", + "content": "Object/background ambiguity in the real world. Even after successfully parsing the world into object and background slots, separating those from each other in an unsupervised way is an open challenge. Promising directions include temporal contrastive learning [5, 28], unsupervised clustering [8, 59], and utilizing geometric cues [65, 66]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": "Slot drift. From Figure 5, we notice that the final RGB reconstruction quality decreases with longer videos, which indicates that the distribution of the slot representation has shifted in the latter frames. This issue leads to degraded temporal consistency and ambiguous object masks. Better training schemes and model improvements, such as adding data augmentation during training [15], replacing GRU units with long-short memory units [27], and enforcing temporal" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 309, + 70, + 545, + 266 + ], + "blocks": [ + { + "bbox": [ + 309, + 70, + 545, + 266 + ], + "lines": [ + { + "bbox": [ + 309, + 70, + 545, + 266 + ], + "spans": [ + { + "bbox": [ + 309, + 70, + 545, + 266 + ], + "type": "image", + "image_path": "ebf2ce38a48df3be5a2aedea02d2731cdfc99ad2e300a320cf7a8a6dd50a7f60.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 276, + 547, + 354 + ], + "lines": [ + { + "bbox": [ + 304, + 276, + 547, + 354 + ], + "spans": [ + { + "bbox": [ + 304, + 276, + 547, + 354 + ], + "type": "text", + "content": "Figure 7. Top-10 slot visualizations of all the compared methods on the evaluation set of TRI-PD and KITTI datasets. The samples on KITTI are cropped for better visualization. We discard the masks taking more than " + }, + { + "bbox": [ + 304, + 276, + 547, + 354 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 304, + 276, + 547, + 354 + ], + "type": "text", + "content": " of the whole pixels. Both the more powerful decoders (STEVE) and the simpler reconstruction space (SAVi and " + }, + { + "bbox": [ + 304, + 276, + 547, + 354 + ], + "type": "inline_equation", + "content": "\\mathrm{SAV_i + + }" + }, + { + "bbox": [ + 304, + 276, + 547, + 354 + ], + "type": "text", + "content": ") fail to work in realistic driving scenes. Our model produces the most accurate masks and captures more objects." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 369, + 468, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 369, + 468, + 381 + ], + "spans": [ + { + "bbox": [ + 305, + 369, + 468, + 381 + ], + "type": "text", + "content": "consistency in the token space can help." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 385, + 547, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 385, + 547, + 494 + ], + "spans": [ + { + "bbox": [ + 304, + 385, + 547, + 494 + ], + "type": "text", + "content": "Better measuring object discovery in the real world. Measuring the performance of object discovery in the real world is challenging. As noted in [15], one slot may re-bind to another object after the tracked object moves out of the frame and a new object enters the scene. The FG. ARI score is not suitable for such scenarios but currently, there are no better metrics for object discovery which is able to tackle the re-binding issue. Developing a novel metric for this problem could have a significant impact on the community." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 509, + 378, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 509, + 378, + 521 + ], + "spans": [ + { + "bbox": [ + 306, + 509, + 378, + 521 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 529, + 547, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 529, + 547, + 649 + ], + "spans": [ + { + "bbox": [ + 304, + 529, + 547, + 649 + ], + "type": "text", + "content": "This work proposes MoTok, a unified architecture that leverages motion-guided tokenization for object discovery. By jointly training the slot representation with motion cues and vector quantization, our model enables the emergence of interpretable mid-level features which simplifies the problem of object discovery. Comprehensive evaluation on both synthetic and real-world benchmarks shows that with sufficient capacity of the slot decoder, motion guidance alleviates the need for labels, optical flow, or depth decoding, thanks to tokenization, achieving state-of-the-art results." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "type": "text", + "content": "Acknowledgements. We thank Dian Chen, Alexei Efros, and Andrew Owens for their valuable comments. This research was supported by Toyota Research Institute. YXW was supported in part by NSF Grant 2106825, NIFA Award 2020-67021-32799, and the NCSA Fellows program." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22979" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 92, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 53, + 92, + 287, + 113 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 92, + 287, + 113 + ], + "spans": [ + { + "bbox": [ + 53, + 92, + 287, + 113 + ], + "type": "text", + "content": "[1] Parallel domain. https://paralleldomain.com/, November 2021. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 116, + 288, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 116, + 288, + 148 + ], + "spans": [ + { + "bbox": [ + 53, + 116, + 288, + 148 + ], + "type": "text", + "content": "[2] Pablo Arbeláez, Jordi Pont-Tuset, Jonathan T Barron, Ferran Marques, and Jitendra Malik. Multiscale combinatorial grouping. In CVPR, 2014. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 150, + 288, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 150, + 288, + 184 + ], + "spans": [ + { + "bbox": [ + 53, + 150, + 288, + 184 + ], + "type": "text", + "content": "[3] Nicolas Ballas, Li Yao, Chris Pal, and Aaron Courville. Delving deeper into convolutional networks for learning video representations. In ICLR, 2016. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 186, + 288, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 186, + 288, + 218 + ], + "spans": [ + { + "bbox": [ + 53, + 186, + 288, + 218 + ], + "type": "text", + "content": "[4] Zhipeng Bao, Pavel Tokmakov, Allan Jabri, Yu-Xiong Wang, Adrien Gaidon, and Martial Hebert. Discovering objects that can move. In CVPR, 2022. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 220, + 288, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 220, + 288, + 253 + ], + "spans": [ + { + "bbox": [ + 53, + 220, + 288, + 253 + ], + "type": "text", + "content": "[5] Zhangxing Bian, Allan Jabri, Alexei A Efros, and Andrew Owens. Learning pixel trajectories with multiscale contrastive random walks. In CVPR, 2022. 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 255, + 288, + 299 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 255, + 288, + 299 + ], + "spans": [ + { + "bbox": [ + 53, + 255, + 288, + 299 + ], + "type": "text", + "content": "[6] Christopher P Burgess, Loic Matthew, Nicholas Watters, Rishabh Kabra, Irina Higgins, Matt Botvinick, and Alexander Lerchner. MONet: Unsupervised scene decomposition and representation. arXiv preprint arXiv:1901.11390, 2019. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 301, + 288, + 344 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 301, + 288, + 344 + ], + "spans": [ + { + "bbox": [ + 53, + 301, + 288, + 344 + ], + "type": "text", + "content": "[7] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 346, + 288, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 346, + 288, + 379 + ], + "spans": [ + { + "bbox": [ + 53, + 346, + 288, + 379 + ], + "type": "text", + "content": "[8] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features. In ECCV, 2018. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 381, + 288, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 381, + 288, + 424 + ], + "spans": [ + { + "bbox": [ + 53, + 381, + 288, + 424 + ], + "type": "text", + "content": "[9] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 427, + 288, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 427, + 288, + 460 + ], + "spans": [ + { + "bbox": [ + 48, + 427, + 288, + 460 + ], + "type": "text", + "content": "[10] Chang Chen, Fei Deng, and Sungjin Ahn. Roots: Object-centric representation and rendering of 3D scenes. JMLR, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 462, + 288, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 462, + 288, + 495 + ], + "spans": [ + { + "bbox": [ + 48, + 462, + 288, + 495 + ], + "type": "text", + "content": "[11] Achal Dave, Pavel Tokmakov, and Deva Ramanan. Towards segmenting anything that moves. In ICCV Workshops, 2019. 1, 5, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 497, + 288, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 497, + 288, + 529 + ], + "spans": [ + { + "bbox": [ + 48, + 497, + 288, + 529 + ], + "type": "text", + "content": "[12] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 532, + 288, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 532, + 288, + 586 + ], + "spans": [ + { + "bbox": [ + 48, + 532, + 288, + 586 + ], + "type": "text", + "content": "[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 589, + 288, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 589, + 288, + 621 + ], + "spans": [ + { + "bbox": [ + 48, + 589, + 288, + 621 + ], + "type": "text", + "content": "[14] Yilun Du, Kevin Smith, Tomer Ulman, Joshua Tenenbaum, and Jiajun Wu. Unsupervised discovery of 3D physical objects from video. In ICLR, 2021. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 624, + 288, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 288, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 288, + 667 + ], + "type": "text", + "content": "[15] Gamaleldin F Elsayed, Aravindh Mahendran, Sjoerd van Steenkiste, Klaus Greff, Michael C Mozer, and Thomas Kipf. Savi++: Towards end-to-end object-centric learning from real-world videos. In NeurIPS, 2022. 1, 2, 5, 6, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "type": "text", + "content": "[16] Martin Engelcke, Adam R Kosiorek, Oiwi Parker Jones, and Ingmar Posner. GENESIS: Generative scene inference and sampling with object-centric latent representations. In ICLR, 2020. 1, 2" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 36, + "blocks": [ + { + "bbox": [ + 307, + 73, + 546, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 546, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 546, + 106 + ], + "type": "text", + "content": "[17] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In CVPR, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 107, + 546, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 546, + 129 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 546, + 129 + ], + "type": "text", + "content": "[18] Pedro F Felzenszwalb and Daniel P Huttenlocher. Efficient graph-based image segmentation. IJCV, 2004. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 130, + 546, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 130, + 546, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 130, + 546, + 162 + ], + "type": "text", + "content": "[19] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In CVPR, 2012. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 164, + 547, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 164, + 547, + 185 + ], + "spans": [ + { + "bbox": [ + 307, + 164, + 547, + 185 + ], + "type": "text", + "content": "[20] Robert Gray. Vector quantization. IEEE Assp Magazine, 1984. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 186, + 547, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 547, + 229 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 547, + 229 + ], + "type": "text", + "content": "[21] Klaus Greff, Francois Belletti, Lucas Beyer, Carl Doersch, Yilun Du, Daniel Duckworth, David J Fleet, Dan Gnanapragasam, Florian Golemo, Charles Herrmann, et al. Kubric: A scalable dataset generator. In CVPR, 2022. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 231, + 546, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 231, + 546, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 231, + 546, + 285 + ], + "type": "text", + "content": "[22] Klaus Greff, Raphaël Lopez Kaufman, Rishabh Kabra, Nick Watters, Christopher Burgess, Daniel Zoran, Loic Matthey, Matthew Botvinick, and Alexander Lerchner. Multi-object representation learning with iterative variational inference. In ICML, 2019. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 287, + 546, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 287, + 546, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 287, + 546, + 319 + ], + "type": "text", + "content": "[23] Klaus Greff, Antti Rasmus, Mathias Berglund, Tele Hao, Harri Valpola, and Jürgen Schmidhuber. Tagger: Deep unsupervised perceptual grouping. In NeurIPS, 2016. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 321, + 546, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 321, + 546, + 363 + ], + "spans": [ + { + "bbox": [ + 307, + 321, + 546, + 363 + ], + "type": "text", + "content": "[24] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In CVPR, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 365, + 546, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 546, + 398 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 546, + 398 + ], + "type": "text", + "content": "[25] Rares Guizilini, Vitor abd Ambrus, Sudeep Pillai, Allan Raventos, and Adrien Gaidon. 3d packing for self-supervised monocular depth estimation. In CVPR, 2020. 5, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 399, + 546, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 546, + 431 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 546, + 431 + ], + "type": "text", + "content": "[26] Paul Henderson and Christoph H Lampert. Unsupervised object-centric video generation and decomposition in 3D. In NeurIPS, 2020. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 433, + 546, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 433, + 546, + 454 + ], + "spans": [ + { + "bbox": [ + 307, + 433, + 546, + 454 + ], + "type": "text", + "content": "[27] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation, 1997. 8" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 456, + 546, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 456, + 546, + 487 + ], + "spans": [ + { + "bbox": [ + 307, + 456, + 546, + 487 + ], + "type": "text", + "content": "[28] Allan Jabri, Andrew Owens, and Alexei Efros. Space-time correspondence as a contrastive random walk. NeurIPS, 2020. 8" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 489, + 546, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 489, + 546, + 544 + ], + "spans": [ + { + "bbox": [ + 307, + 489, + 546, + 544 + ], + "type": "text", + "content": "[29] Andrew Jaegle, Sebastian Borgeaud, Jean-Baptiste Alayrac, Carl Doersch, Catalin Ionescu, David Ding, Skanda Koppula, Daniel Zoran, Andrew Brock, Evan Shelhamer, et al. Perceiver io: A general architecture for structured inputs & outputs. In ICLR, 2022. 2, 4" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 545, + 546, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 545, + 546, + 578 + ], + "spans": [ + { + "bbox": [ + 307, + 545, + 546, + 578 + ], + "type": "text", + "content": "[30] Andrew Jaegle, Felix Gimeno, Andy Brock, Oriol Vinyals, Andrew Zisserman, and Joao Carreira. Perceiver: General perception with iterative attention. In ICML, 2021. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 579, + 546, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 579, + 546, + 612 + ], + "spans": [ + { + "bbox": [ + 307, + 579, + 546, + 612 + ], + "type": "text", + "content": "[31] Jindong Jiang, Sepehr Janghorbani, Gerard De Melo, and Sungjin Ahn. SCALOR: Generative world models with scalable object representations. In ICLR, 2020. 2" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 613, + 546, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 546, + 645 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 546, + 645 + ], + "type": "text", + "content": "[32] Daniel Kahneman, Anne Treisman, and Brian J Gibbs. The reviewing of object files: Object-specific integration of information. Cognitive psychology, 1992. 1" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 307, + 647, + 546, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 546, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 546, + 689 + ], + "type": "text", + "content": "[33] Laurynas Karazija, Subhabrataand Choudhury, Iro Laina, Christian Rupprecht, and Andrea Vedaldi. Unsupervised multi-object segmentation by predicting probable motion patterns. In NeurIPS, 2022. 2, 5, 8" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 307, + 691, + 546, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 546, + 713 + ], + "type": "text", + "content": "[34] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 5" + } + ] + } + ], + "index": 35 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "22980" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 73, + 287, + 94 + ], + "spans": [ + { + "bbox": [ + 48, + 73, + 287, + 94 + ], + "type": "text", + "content": "[35] Diederik P Kingma and Max Welling. Auto-encoding variational bayes. In ICLR, 2014. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 288, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 288, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 288, + 140 + ], + "type": "text", + "content": "[36] Thomas Kipf, Gamaleldin F Elsayed, Aravindh Mahendran, Austin Stone, Sara Sabour, Georg Heigold, Rico Jonschkowski, Alexey Dosovitskiy, and Klaus Greff. Conditional object-centric learning from video. In ICLR, 2022. 1, 2, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 288, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 288, + 161 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 288, + 161 + ], + "type": "text", + "content": "[37] Kurt Koffka. Principles of Gestalt psychology. Routledge, 2013. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 163, + 288, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 288, + 207 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 288, + 207 + ], + "type": "text", + "content": "[38] Zhixuan Lin, Yi-Fu Wu, Skand Vishwanath Peri, Weihao Sun, Gautam Singh, Fei Deng, Jindong Jiang, and Sungjin Ahn. Space: Unsupervised object-oriented scene representation via spatial attention and decomposition. In ICLR, 2020. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 209, + 288, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 288, + 252 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 288, + 252 + ], + "type": "text", + "content": "[39] Francesco Locatello, Dirk Weissenborn, Thomas Unterthiner, Aravindh Mahendran, Georg Heigold, Jakob Uszkoreit, Alexey Dosovitskiy, and Thomas Kipf. Object-centric learning with slot attention. In NeurIPS, 2020. 1, 2, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 254, + 288, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 254, + 288, + 297 + ], + "spans": [ + { + "bbox": [ + 48, + 254, + 288, + 297 + ], + "type": "text", + "content": "[40] Nikolaus Mayer, Eddy Ilg, Philip Hausser, Philipp Fischer, Daniel Cremers, Alexey Dosovitskiy, and Thomas Brox. A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In CVPR, 2016. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 298, + 287, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 298, + 287, + 330 + ], + "spans": [ + { + "bbox": [ + 48, + 298, + 287, + 330 + ], + "type": "text", + "content": "[41] Tim Meinhardt, Alexander Kirillov, Laura Leal-Taixe, and Christoph Feichtenhofer. Trackformer: Multi-object tracking with transformers. In CVPR, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 332, + 288, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 332, + 288, + 374 + ], + "spans": [ + { + "bbox": [ + 48, + 332, + 288, + 374 + ], + "type": "text", + "content": "[42] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 376, + 288, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 376, + 288, + 409 + ], + "spans": [ + { + "bbox": [ + 48, + 376, + 288, + 409 + ], + "type": "text", + "content": "[43] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, 2021. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 411, + 288, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 411, + 288, + 442 + ], + "spans": [ + { + "bbox": [ + 48, + 411, + 288, + 442 + ], + "type": "text", + "content": "[44] Ali Razavi, Aaron Van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. NeurIPS, 2019. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 445, + 288, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 445, + 288, + 477 + ], + "spans": [ + { + "bbox": [ + 48, + 445, + 288, + 477 + ], + "type": "text", + "content": "[45] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster R-CNN: Towards real-time object detection with region proposal networks. In NeurIPS, 2015. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 479, + 288, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 479, + 288, + 510 + ], + "spans": [ + { + "bbox": [ + 48, + 479, + 288, + 510 + ], + "type": "text", + "content": "[46] Danilo Jimenez Rezende, Shakir Mohamed, and Daan Wierstra. Stochastic backpropagation and approximate inference in deep generative models. In ICML, 2014. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 512, + 288, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 512, + 288, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 512, + 288, + 533 + ], + "type": "text", + "content": "[47] Jason Tyler Rolfe. Discrete variational autoencoders. In ICLR, 2017. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 535, + 288, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 535, + 288, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 535, + 288, + 578 + ], + "type": "text", + "content": "[48] Mehdi SM Sajjadi, Daniel Duckworth, Aravindh Mahendran, Sjoerd van Steenkiste, Filip Pavetic, Mario Lucic, Leonidas J Guibas, Klaus Greff, and Thomas Kipf. Object scene representation transformer. In NeurIPS, 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 580, + 288, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 580, + 288, + 643 + ], + "spans": [ + { + "bbox": [ + 48, + 580, + 288, + 643 + ], + "type": "text", + "content": "[49] Mehdi SM Sajjadi, Henning Meyer, Etienne Pot, Urs Bergmann, Klaus Greff, Noha Radwan, Suhani Vora, Mario Lucic, Daniel Duckworth, Alexey Dosovitskiy, et al. Scene representation transformer: Geometry-free novel view synthesis through set-latent scene representations. In CVPR, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 647, + 288, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 288, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 288, + 679 + ], + "type": "text", + "content": "[50] Hinrich Schütze, Christopher D Manning, and Prabhakar Raghavan. Introduction to information retrieval. Cambridge University Press Cambridge, 2008. 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 681, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 288, + 713 + ], + "type": "text", + "content": "[51] Maximilian Seitzer, Max Horn, Andrii Zadaianchuk, Dominik Zietlow, Tianjun Xiao, Carl-Johann Simon-Gabriel, Tong He, Zheng Zhang, Bernhard Scholkopf, Thomas Brox, et al." + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 625 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "type": "text", + "content": "Bridging the gap to real-world object-centric learning. In ICLR, 2023. 2, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 96, + 545, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 118 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 118 + ], + "type": "text", + "content": "[52] Gautam Singh, Fei Deng, and Sungjin Ahn. Illiterate dall-e learns to compose. In ICLR, 2021. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 119, + 545, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 119, + 545, + 151 + ], + "spans": [ + { + "bbox": [ + 308, + 119, + 545, + 151 + ], + "type": "text", + "content": "[53] Gautam Singh, Yi-Fu Wu, and Sungjin Ahn. Simple unsupervised object-centric learning for complex and naturalistic videos. In NeurlPS, 2022. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 152, + 545, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 152, + 545, + 175 + ], + "spans": [ + { + "bbox": [ + 308, + 152, + 545, + 175 + ], + "type": "text", + "content": "[54] Elizabeth S Spelke and Katherine D Kinzler. Core knowledge. Developmental science, 2007. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 176, + 545, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 176, + 545, + 208 + ], + "spans": [ + { + "bbox": [ + 308, + 176, + 545, + 208 + ], + "type": "text", + "content": "[55] Karl Stelzner, Kristian Kersting, and Adam R Kosiorek. Decomposing 3D scenes into objects via unsupervised volume segmentation. arXiv preprint arXiv:2104.01148, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 209, + 545, + 242 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 209, + 545, + 242 + ], + "spans": [ + { + "bbox": [ + 308, + 209, + 545, + 242 + ], + "type": "text", + "content": "[56] Chen Sun, Austin Myers, Carl Vondrick, Kevin Murphy, and Cordelia Schmid. Videobert: A joint model for video and language representation learning. In CVPR, 2019. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 243, + 545, + 265 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 243, + 545, + 265 + ], + "spans": [ + { + "bbox": [ + 308, + 243, + 545, + 265 + ], + "type": "text", + "content": "[57] Zachary Teed and Jia Deng. RAFT: Recurrent all-pairs field transforms for optical flow. In ECCV, 2020. 5, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 266, + 545, + 288 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 266, + 545, + 288 + ], + "spans": [ + { + "bbox": [ + 308, + 266, + 545, + 288 + ], + "type": "text", + "content": "[58] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. NeurIPS, 2017. 2, 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 289, + 545, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 289, + 545, + 332 + ], + "spans": [ + { + "bbox": [ + 308, + 289, + 545, + 332 + ], + "type": "text", + "content": "[59] Wouter Van Gansbeke, Simon Vandenhende, Stamatios Georgoulis, and Luc Van Gool. Unsupervised semantic segmentation by contrasting object mask proposals. In ICCV, 2021. 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 334, + 545, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 334, + 545, + 367 + ], + "spans": [ + { + "bbox": [ + 308, + 334, + 545, + 367 + ], + "type": "text", + "content": "[60] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 2017. 1, 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 369, + 545, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 369, + 545, + 411 + ], + "spans": [ + { + "bbox": [ + 308, + 369, + 545, + 411 + ], + "type": "text", + "content": "[61] Rishi Veerapaneni, John D Co-Reyes, Michael Chang, Michael Janner, Chelsea Finn, Jiajun Wu, Joshua Tenenbaum, and Sergey Levine. Entity abstraction in visual model-based reinforcement learning. In CoRL, 2020. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 413, + 545, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 413, + 545, + 445 + ], + "spans": [ + { + "bbox": [ + 308, + 413, + 545, + 445 + ], + "type": "text", + "content": "[62] Ziyi Wu, Nikita Dvornik, Klaus Greff, Thomas Kipf, and Animesh Garg. Slotformer: Unsupervised visual dynamics simulation with object-centric models. In ICLR, 2023. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 447, + 545, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 447, + 545, + 479 + ], + "spans": [ + { + "bbox": [ + 308, + 447, + 545, + 479 + ], + "type": "text", + "content": "[63] Wilson Yan, Yunzhi Zhang, Pieter Abbeel, and Aravind Srinivas. Videogpt: Video generation using vq-vae and transformers. arXiv preprint arXiv:2104.10157, 2021. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 480, + 545, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 480, + 545, + 513 + ], + "spans": [ + { + "bbox": [ + 308, + 480, + 545, + 513 + ], + "type": "text", + "content": "[64] Peiyu Yu, Sirui Xie, Xiaojian Ma, Yixin Zhu, Ying Nian Wu, and Song-Chun Zhu. Unsupervised foreground extraction via deep region competition. In NeurIPS, 2021. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 514, + 545, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 514, + 545, + 558 + ], + "spans": [ + { + "bbox": [ + 308, + 514, + 545, + 558 + ], + "type": "text", + "content": "[65] Jiahui Zhang, Dawei Sun, Zixin Luo, Anbang Yao, Lei Zhou, Tianwei Shen, Yurong Chen, Long Quan, and Honggen Liao. Learning two-view correspondences and geometry using order-aware network. In ICCV, 2019. 8" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 308, + 559, + 545, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 559, + 545, + 591 + ], + "spans": [ + { + "bbox": [ + 308, + 559, + 545, + 591 + ], + "type": "text", + "content": "[66] Brady Zhou and Philipp Krahenbuhl. Cross-view transformers for real-time map-view semantic segmentation. In CVPR, 2022. 8" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 308, + 593, + 545, + 625 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 593, + 545, + 625 + ], + "spans": [ + { + "bbox": [ + 308, + 593, + 545, + 625 + ], + "type": "text", + "content": "[67] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In ICLR, 2021. 2" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22981" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Object Pop-Up_ Can We Infer 3D Objects and Their Poses From Human Interactions Alone_/965b59c8-d0c6-413b-852b-108c8ff0bbe7_content_list.json b/2023/Object Pop-Up_ Can We Infer 3D Objects and Their Poses From Human Interactions Alone_/965b59c8-d0c6-413b-852b-108c8ff0bbe7_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d801643121d56a53710dffb444d63b1754589012 --- /dev/null +++ b/2023/Object Pop-Up_ Can We Infer 3D Objects and Their Poses From Human Interactions Alone_/965b59c8-d0c6-413b-852b-108c8ff0bbe7_content_list.json @@ -0,0 +1,1705 @@ +[ + { + "type": "text", + "text": "Object pop-up: Can we infer 3D objects and their poses from human interactions alone?", + "text_level": 1, + "bbox": [ + 138, + 130, + 831, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ilya A. Petrov $^{1}$ , Riccardo Marin $^{1}$ , Julian Chibane $^{2,1}$ , and Gerard Pons-Moll $^{1,2,3}$", + "bbox": [ + 174, + 214, + 792, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1University of Tübingen, Germany", + "bbox": [ + 379, + 251, + 589, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Max Planck Institute for Informatics, Saarland Informatics Campus, Germany", + "bbox": [ + 250, + 268, + 717, + 282 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ Tübingen AI Center, Germany", + "bbox": [ + 390, + 282, + 578, + 296 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{i.petrov, gerard.pons-moll}@uni-tuebingen.de, riccardo.marin@mnf.uni-tuebingen.de, jchibane@mpi-inf.mpg.de", + "bbox": [ + 120, + 296, + 846, + 325 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ff13a4932e654648edde69665a0a1ece8f97f4f484679df15926cc86bf559f08.jpg", + "image_caption": [ + "Ref. Image (not used)", + "Figure 1. Human-Centric prediction of Human-Object Interaction. We explore Human-Object interaction when only the cues of the human pose interacting with an unobserved object are available ('Input'). We propose a neural model that, for the first time, can infer the location of the object ('Result') from such input. This is possible even when our subject simulates the interactions with the object ('Reference Image')." + ], + "image_footnote": [], + "bbox": [ + 76, + 373, + 148, + 500 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9474c6239a1356b0689de49886b91f9d190d57f1d5b798e781cb5e9cf9f13e72.jpg", + "image_caption": [ + "Point Cloud" + ], + "image_footnote": [], + "bbox": [ + 153, + 373, + 207, + 496 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3e19da7a40c2e74032a938a644e3074e03779b64a7d0f74b90a186e261b867d1.jpg", + "image_caption": [ + "Input", + "Result" + ], + "image_footnote": [], + "bbox": [ + 217, + 371, + 267, + 496 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2c894457d371750b0e0f25f1c6cc6cf2acbb5299eacb2470cecdfe7c2e6fb90e.jpg", + "image_caption": [ + "Ref. Image (not used)" + ], + "image_footnote": [], + "bbox": [ + 281, + 373, + 351, + 496 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/687583a36d7ec1f7205d5617528687ea5e78f37b287b39c9564d1f9ebc48adef.jpg", + "image_caption": [ + "Input Point Cloud" + ], + "image_footnote": [], + "bbox": [ + 354, + 373, + 410, + 494 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c37c453829a71c4bbbae8edf24650c9088648d2070ce1f4c0ae5c96f174f4732.jpg", + "image_caption": [ + "Result" + ], + "image_footnote": [], + "bbox": [ + 421, + 373, + 473, + 494 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2f165549bdb8c9d5784a7e54f9ee99befc57a75d6c225308e185dfe332560873.jpg", + "image_caption": [ + "Ref. Image (not used)" + ], + "image_footnote": [], + "bbox": [ + 480, + 373, + 552, + 494 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f1a5f590815891caba96603958c61bec4c0c439fbbb66442e690b26e42a3faf5.jpg", + "image_caption": [ + "Input Point Cloud" + ], + "image_footnote": [], + "bbox": [ + 557, + 373, + 612, + 493 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b5fafb11d446414b848dee849492adba0c04e5691f29e584de0588b62a338c88.jpg", + "image_caption": [ + "Result" + ], + "image_footnote": [], + "bbox": [ + 624, + 373, + 678, + 493 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/357160551435c931f39da0ac68f573519ff364f3a5eb11d57676e5a434aa0325.jpg", + "image_caption": [ + "Ref. Image (not used)" + ], + "image_footnote": [], + "bbox": [ + 686, + 373, + 756, + 494 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a705fbebadecf78663cb2592ebd8c3817e4b807d45e166317e7d15bb6e9d5ddc.jpg", + "image_caption": [ + "Input Point Cloud" + ], + "image_footnote": [], + "bbox": [ + 756, + 373, + 810, + 496 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3b6324fa122e1b48406fb033e003f6a071802a025becb7d5d6eeb8f72be88555.jpg", + "image_caption": [ + "Result" + ], + "image_footnote": [], + "bbox": [ + 821, + 373, + 872, + 496 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 582, + 313, + 598 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The intimate entanglement between objects affordances and human poses is of large interest, among others, for behavioural sciences, cognitive psychology, and Computer Vision communities. In recent years, the latter has developed several object-centric approaches: starting from items, learning pipelines synthesizing human poses and dynamics in a realistic way, satisfying both geometrical and functional expectations. However, the inverse perspective is significantly less explored: Can we infer 3D objects and their poses from human interactions alone? Our investigation follows this direction, showing that a generic 3D human point cloud is enough to pop up an unobserved object, even when the user is just imitating a functionality (e.g., looking through a binocular) without involving a tangible counterpart. We validate our method qualitatively and quantitatively, with synthetic data and sequences acquired for the task, showing applicability for XR/VR.", + "bbox": [ + 73, + 626, + 472, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 501, + 582, + 630, + 597 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Complex interactions with the world are among the unique skills distinguishing humans from other living beings. Even though our perception might be imperfect (we cannot hear ultrasonic sounds or see ultraviolet light [49]), our cognitive representation is enriched with a functional perspective, i.e., potential ways of interacting with objects or, as introduced by Gibson and colleagues, the affordance of the objects [19]. Several behavioural studies confirmed the centrality of this concept [2, 9, 43], which plays a fundamental role also for kids' development [2]. Computer Vision is well-aware that the function of an object complements its appearance [21], and exploited this in tasks like human and object reconstruction [62]. Previous literature approaches the interaction analysis from an object perspective (i.e., given an object, analyze the human interaction) [8,22,63], building object-centric priors [73], generating realistic grasps given the object [13, 34], reconstructing hand-object interactions [8,22,63]. Namely, objects induce functionality, so a human interaction (e.g., a mug suggests", + "bbox": [ + 496, + 613, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "4726", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/075157ef91344df3f099866a5b5b90e89f72a1e3ef09ee021a731fc3b175eb17.jpg", + "image_caption": [ + "Figure 2. Human-Centric vs. Object-Centric inference. A common perspective in the prior work is to infer affordances or human pose starting from an object. We explore the inverse, Human-Centric perspective of the human-object interaction relationship: given the human, we are interested in predicting the object." + ], + "image_footnote": [], + "bbox": [ + 89, + 87, + 450, + 270 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "a drinking action; an handle a grasping one). For the first time, our work reverts the perspective, suggesting that analyzing human motion and behaviour is naturally a human-centric problem (i.e., given a human interaction, what kind of functionality is it suggesting, Fig. 2). Moving the first step in this new research direction, we pose a fundamental question: Can we infer 3D objects and their poses from human interactions alone?", + "bbox": [ + 75, + 371, + 468, + 491 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "At first sight, the problem seems particularly hard and significantly under-constrained since several geometries might fit the same action. However, the human body complements this information in several ways: physical relations, characteristic poses, or body dynamics serve as valuable proxies for the involved functionality, as suggested by Fig. 1. Such hints are so powerful that we can easily imagine the kind of object and its location, even if such an object does not exist. Furthermore, even if the given pose might fit several possible solutions, our mind naturally comes to the most natural one indicated by the observed behaviour. It also suggests that solely focusing on the contact region (an approach often preferred by previous works) is insufficient in this new viewpoint. The reification principle of Gestalt psychology [37] highlights that \"the whole\" arises from \"the parts\" and their relation. Similarly, in Fig. 3, the hand grasps in B) pop up a binocular in our mind because we naturally consider the relationship with the other body parts.", + "bbox": [ + 75, + 492, + 470, + 779 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Finally, moving to a human-centric perspective in human-object interaction is critical for human studies and daily-life applications. Modern systems for AR/VR [24], and digital interaction [26] are both centered on humans, often manipulating objects that do not have a real-world counterpart. Learning to decode an object from human behaviour enables unprecedented applications.", + "bbox": [ + 75, + 779, + 468, + 885 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To answer our question, we deploy a first straightforward", + "bbox": [ + 96, + 885, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/7b03322c0e42245b8afa641d5659f3be172a5958ea7935a2c65c6e9bd893563d.jpg", + "image_caption": [ + "A)" + ], + "image_footnote": [], + "bbox": [ + 509, + 95, + 666, + 210 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a531002a2af6fdd8cb7e29cbf6fbc4abab31fc24793fb85281ef03ff919bcbe0.jpg", + "image_caption": [ + "B)", + "Figure 3. Reification principle. Gestalt theory suggests our perception considers the whole before the single parts (A). This served as inspiration for our work: the object can arise considering the body parts as a whole (B)." + ], + "image_footnote": [], + "bbox": [ + 669, + 98, + 867, + 213 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "and effective pipeline to \"pop up\" a rigid object from a 3D human point cloud. Starting from the input human point cloud and a class, we train an end-to-end pipeline to infer object location. In case a temporal sequence of point clouds is available, we suggest post-processing to avoid jittering and inconsistencies in the predictions, showing the relevance of this information to handle ambiguous poses. We show promising results on previously unaddressed tasks in digital and real-world scenes. Finally, our method allows us to analyze different features of human behaviour, highlight their contribution to object retrieval, and point to exciting directions for future works.", + "bbox": [ + 496, + 316, + 890, + 496 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our main contributions are:", + "bbox": [ + 517, + 498, + 785, + 512 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We formulate a novel problem, changing the perspective taken by previous works in the field, and open to a yet unexplored research direction;", + "2. We introduce a method capable of predicting the object starting from an input human point cloud;", + "3. We analyze different components of the human-object relationship: the contribution of different pieces of interactions (hands, body, time sequence), the point-wise saliency of input points, and the confusion produced by objects with similar functions." + ], + "bbox": [ + 511, + 522, + 890, + 693 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "All the code will be made publicly available.", + "bbox": [ + 519, + 703, + 815, + 717 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 731, + 640, + 747 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Object Functionality", + "text_level": 1, + "bbox": [ + 500, + 756, + 694, + 773 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the core of human perception of objects, functionality complements physical appearance, enhancing our perception. Gibson [19] introduced the idea that humans use affordances of objects for perception. Affordance can be defined as: \"an intrinsic property of an object, allowing an action to be performed with the object\" [35].", + "bbox": [ + 496, + 779, + 890, + 869 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "From a Computer Vision perspective, object functionality supports several tasks such as scene analysis [21], object", + "bbox": [ + 496, + 869, + 890, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "4727", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "classification [17, 35], object properties inferring [71], and it is also possible to learn object-specific human interaction models from 2D images [62]. These works suggest an intimate entanglement between the action and the object itself.", + "bbox": [ + 75, + 90, + 470, + 151 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Human-Object interaction", + "text_level": 1, + "bbox": [ + 76, + 160, + 316, + 176 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Modelling environment-aware humans and their interactions in 3D is one of the most recent challenges to creating virtual humans. We see two main lines of work there: hand-focused and full-body ones.", + "bbox": [ + 75, + 183, + 468, + 244 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Hand-Object Interaction. Several works tackle the problem of human interaction, focusing solely on the hands [10]. This has been done starting from 2D [8, 15, 18, 22, 25, 29, 30, 34, 61, 63], 2.5D [3, 5], and 3D data [4, 34, 55]. Particularly promising seems the application of well-designed priors for the motion [73]. The class of objects involved in these works are mainly limited to graspable ones. We argue that interactions involving different body parts are common in everyday life, more attractive from an applicative perspective, and more challenging. Moreover, full-body context is crucial in reconstructing even grasp interactions since body pose contains information on an object's properties, e.g. pounding with a hammer affects the whole posture to support the action.", + "bbox": [ + 75, + 261, + 468, + 473 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Fully-Body Interaction. On this line, several works focus on the interaction between a human and a scene [12,28, 31,38,57,64,65,69]. Also, in this case, priors can be used to regularize the motion [48]. Several datasets are also available to study the interaction between a human and a single object. For example, recent BEHAVE [1], GRAB [55], and InterCap [32] capture full-body interactions with diverse objects. Works address the task of humans interactions reconstruction from different kinds of data sources like single image [59, 67], video [16, 50, 60], and multiview capturing [1, 33, 51], and synthetization of them as well [7,27,41,54,56,58,66,70]. However, in all these works, we observe a general object-centric perspective: given a scene or an object, they aim to recreate the humans interacting with them. We argue that the significantly less explored complementary one has more concrete applications in daily life, especially in VR/XR contexts where the human is central to the system.", + "bbox": [ + 75, + 491, + 468, + 762 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Human-centric perspective", + "text_level": 1, + "bbox": [ + 76, + 771, + 321, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While the general trend mainly focuses on the surrounding environment and objects, there is a growing interest and availability of egocentric tools for humans [23, 68] also interacting with objects [24,36,39]. They provide a subjective view and a valuable paradigm for several applications, like letting a user interact with objects in the digital world. Recent works also involve more sophisticated devices [52,53],", + "bbox": [ + 75, + 794, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "while they are still far from applicability. In a similar direction to our work, others propose to recover objects arrangements in a room starting from the human motion [42], to hallucinate a coherent 2D image from a human pose [6], or predicting physical properties of the objects (e.g., the weight of a box) from human joints [71]. While the principle inspires us, our study significantly differs: we focus on object pose and its spatial relation with humans, starting solely from unordered point clouds.", + "bbox": [ + 496, + 90, + 890, + 227 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 239, + 589, + 255 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This section describes our setting and the main components of our methodology, both at inference and training time. An overview of our pipeline can be found in Fig. 4.", + "bbox": [ + 496, + 265, + 890, + 311 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Object Pop-Up", + "text_level": 1, + "bbox": [ + 500, + 319, + 651, + 335 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Input. Our method starts from a single human point cloud $\\mathbf{P} \\in \\mathbb{R}^{N_P \\times 3}$ with $N_P$ points and a hot-encoded object class $c$ . The input point cloud can result from 3D/4D scans, IMUs template fitting, or any other shape-from-X approach. Regardless of the point cloud source, we remark that no further information is used apart from the 3D coordinates of the points. We represent each object as 1500 key points $\\mathbf{K}_c$ uniformly sampled on the template mesh.", + "bbox": [ + 496, + 342, + 890, + 464 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Object Center. Training a model to predict an object pose from a human point cloud poses several challenges. Such a task requires the network to understand the location of different body parts and their subtle relations while jointly developing a sense of its spatial relationship with the human. Empirically, we observed that this is only feasible by carefully deconstructing the problem and designing different features to ease the learning process. As the first step to decompose this problem, we train a PointNet++ architecture [11,45] to predict the object center $\\mathbf{o}_P$ starting from $\\mathbf{P}$ . At training time, this is supervised with an L2 loss against the ground truth center $\\widehat{\\mathbf{o}}_P$ :", + "bbox": [ + 496, + 482, + 890, + 662 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\mathbf {o} _ {P}} (\\mathbf {P}) = \\left\\| \\mathbf {o} _ {P} - \\widehat {\\mathbf {o}} _ {P} \\right\\| _ {2} ^ {2}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 609, + 672, + 890, + 690 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Solving this task provides a good initialization for the object pose, and we move the key points associated with the input class $c$ to the predicted center $\\mathbf{o}_P$ . Also, the features $\\mathbf{F}_o \\in \\mathbb{R}^{512}$ extracted by the network encode import information on the whole human body.", + "bbox": [ + 496, + 700, + 890, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Local Neighbourhood. The center prediction module can be further exploited using the nearest human regions. Intuitively, considering the closest body parts is essential to infer a contact relationship, but also the influence of body parts not directly in touch with the object (e.g., head orientation while using binoculars). To learn these connections, we consider the centered key points together with", + "bbox": [ + 496, + 794, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "4728", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/13bc0c9288897809a7841d0b5a428d093b1e2e2973554020caa5b64564b8f270.jpg", + "image_caption": [ + "Figure 4. Object pop-up. Our method predicts the position of an object, starting only from an input point cloud and an object class. It relies on a careful problem decomposition in several sub-tasks, extracting features that involve the entire human body and relations between body parts near the object." + ], + "image_footnote": [], + "bbox": [ + 109, + 83, + 861, + 247 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the 3000 closest points of the input human point cloud: $\\mathbf{P}_L = KNN(\\mathbf{P},\\mathbf{o}_P)$ . We pass these two sets as a unique point cloud to a PointNet++ network, obtaining a new set of per-point features $\\mathbf{F}_{P_L}\\in \\mathbb{R}^{128}$ .", + "bbox": [ + 75, + 306, + 470, + 368 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Object displacement. To predict the object's final position, we empirically observed that directly predicting a rotation and a translation is not a good solution. Inspired by recent works that suggest a point-wise offset prediction to recover 3D human shapes [14], we apply a similar approach to our task. Our goal is to predict a point-wise shift $\\mathbf{S}_K$ for the $\\mathbf{K}_c$ vertices to align them to the target pose. We append the features $\\mathbf{F}_o$ , $\\mathbf{F}_{P_L}$ , the one-hot encoding of the object class $c$ , and a positional encoding to the centered key points, and we pass them to a decoder. At training time, we consider the following loss:", + "bbox": [ + 75, + 383, + 468, + 551 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {o f f} \\left(\\mathbf {K} _ {c}, \\mathbf {F} _ {o}, \\mathbf {F} _ {P _ {L}}, c\\right) = \\left\\| \\mathbf {S} _ {K} - \\widehat {\\mathbf {S}} _ {K} \\right\\| _ {F} ^ {2}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 137, + 556, + 468, + 575 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The network is then trained end-to-end using:", + "bbox": [ + 96, + 582, + 398, + 597 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL = L _ {\\mathbf {o} _ {P}} + \\alpha L _ {\\text {o f f}}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 204, + 604, + 468, + 622 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The weighting coefficient is $\\alpha = 10$", + "bbox": [ + 76, + 628, + 320, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Template fitting", + "text_level": 1, + "bbox": [ + 76, + 651, + 236, + 667 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Procrustes alignment. The point-wise offset produced by the network potentially distorts the key points structure in a non-rigid way. To recover the desired global rigid transformation, we rely on a Procrustes alignment [20]. This procedure takes as input two point clouds and returns the rotation $\\mathbf{R}$ and the translation $\\mathbf{t}$ to minimize the L2 distances of the points:", + "bbox": [ + 75, + 674, + 468, + 779 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nP (\\cdot , \\cdot) \\rightarrow (\\mathbf {R}, \\mathbf {t}). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 214, + 780, + 468, + 796 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We apply this to the template key points and their configuration obtained with our network:", + "bbox": [ + 76, + 800, + 468, + 830 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nP \\left(\\mathbf {K} _ {c}, \\mathbf {K} _ {c} + \\mathbf {o} _ {P} + \\mathbf {S} _ {K}\\right)\\rightarrow (\\mathbf {R}, \\mathbf {t}). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 155, + 838, + 468, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, we recover the desired object pose as:", + "bbox": [ + 96, + 862, + 401, + 877 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {T} ^ {\\prime} = \\mathbf {R T} _ {c} + \\mathbf {t} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 220, + 883, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Time Smoothing. While our pipeline is designed to work with a single point cloud as input, considering the temporal evolution of interaction is often crucial, shaping the context of the individual poses. If a temporal sequence of point clouds is available, we provide a post-processing smoothing technique to take advantage of this further information. After running our method for each frame, we smooth the centering prediction across the sequence using a Gaussian kernel. Later, we will discuss a variation of our approach that also predicts the object class. In that case, we consider the most frequent class prediction over the whole set of frames to fix a class for the sequence.", + "bbox": [ + 496, + 306, + 893, + 488 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 498, + 505, + 633, + 522 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this Section, we will describe the datasets used for training and testing our method. Then, we will present the baseline and the evaluation metrics. Finally, we will provide validation of our method as well as an analysis of its extended version that allows object class prediction.", + "bbox": [ + 496, + 531, + 892, + 607 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Datasets We jointly train on the union of BEHAVE [1] and GRAB [55], obtaining a set of:", + "bbox": [ + 498, + 632, + 890, + 662 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "15 subjects, first 8 subjects from the GRAB dataset and 7 subjects from the official training part of BEHAV;", + "- 40 different classes of objects, including all 20 objects from the BEHAVE dataset and 20 selected objects from the GRAB dataset;" + ], + "bbox": [ + 517, + 675, + 890, + 765 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We downsample training sequences of GRAB and BEHAVE to $10fps$ . To evaluate our method, we select subjects 9 and 10 from the GRAB dataset and downsample the sequences to $30fps$ . For the BEHAVE dataset, we use the official test part, which includes all sequences at $1fps$ with subject 3 and part of the sequences with subjects 4, 5. As an input, we use point clouds with 9000 points sampled uniformly over the SMPL-H [44] meshes. We refer to raw", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4729", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/39a8bcf771f32d10ed2892f9c60885239100293155c442332040d7348f12592a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsGRABBEHAVEBEHAVE-Raw
Ec↓Ev2v↓Ec↓Ev2v↓Ec↓Ev2v↓
NN0.03620.14450.08020.3445--
Ours0.02370.09430.06630.29000.08060.3143
", + "bbox": [ + 263, + 88, + 709, + 146 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Comparison with the baseline. Our method significantly outperforms the baseline, even though the baseline uses the vertex order as additional information. Moreover, the baseline method does not generalize to point clouds with an arbitrary number of vertices.", + "bbox": [ + 75, + 157, + 893, + 188 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "point clouds from the BEHAVE dataset used in our experiments as BEHAVE-Raw. We use point clouds that are fused from 4 Kinect sensors and subsample $90k$ points from them.", + "bbox": [ + 75, + 199, + 470, + 246 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Data augmentation. During training, to simulate errors in the center prediction, we randomly translate and rotate the object around the ground-truth center $\\widehat{\\mathbf{o}}_P$ .", + "bbox": [ + 75, + 276, + 470, + 324 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation details. We implement our method using PyTorch framework and use Nvidia RTX3090 GPU for training and evaluation. The model is trained using Adam optimizer for 60 epochs with a learning rate of $1e^{-4}$ , which decays 10 times after 30-th and 40-th epochs. For the first 20 epochs of training, we use ground-truth object center $\\widehat{\\mathbf{o}}_P$ instead of predicted $\\mathbf{o}_P$ to select local neighborhood $\\mathbf{F}_{P_L}$ to warm up the local PointNet++ encoder.", + "bbox": [ + 75, + 354, + 468, + 474 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Nearest-Neighbor Baseline. Since we are the first to tackle this task and no competitors are available, we propose a simple while informative baseline. Given the input point cloud, we recover the most similar in the training dataset in an L2 sense. Then, we recover the object handled by that subject and pose it in space in the same way. This baseline demonstrates that the task is non-trivial and the generalization to unseen poses and subjects of our method. Also, this baseline requires that the target point cloud and the ones in the training set share the same number of points. Hence, if the input point cloud is a raw scan, this baseline is not applicable. Our method, instead, does not rely on this assumption and is more general.", + "bbox": [ + 75, + 506, + 470, + 704 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Object classification. In our research, we also investigate the possibility of incorporating class prediction inside the network training. This task is significantly difficult at a single-frame level since an isolated pose often does not suggest a clear functionality. However, including this step is interesting to analyze the interaction and the nature of the network confusion. Hence, we modify our method by adding a decoder module that takes the global features $\\mathbf{F}_o$ and the local ones $\\mathbf{F}_{P_L}$ as input to predict the object class. Then, we add to the training a simple cross-entropy loss between the predicted class and the ground truth one $\\widehat{C}_P$ .", + "bbox": [ + 75, + 734, + 470, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Metrics", + "text_level": 1, + "bbox": [ + 500, + 199, + 596, + 214 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In qualitative experiments, we use three main metrics to evaluate our results. In the tables, we report the average error across the considered test samples.", + "bbox": [ + 498, + 222, + 893, + 268 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Vertex-to-vertex. In most cases, our resulting object and the target one share the same number of vertices. Hence, we can compute the error between our prediction and the ground truth $\\widehat{\\mathbf{T}}'$ as a point-to-point error:", + "bbox": [ + 498, + 287, + 892, + 348 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nE _ {v 2 v} = \\left\\| \\mathbf {T} ^ {\\prime} - \\widehat {\\mathbf {T}} ^ {\\prime} \\right\\| _ {F}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 358, + 892, + 378 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "When such error is computed only between the object centers, we will refer to it as $E_{c}$ .", + "bbox": [ + 498, + 388, + 890, + 420 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Chamfer distance. When we evaluate the network that also predicts the class, target objects and selected templates might not share the same number of vertices. In that case, as a metric we use bi-directional Chamfer distance:", + "bbox": [ + 498, + 438, + 893, + 500 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nE _ {c h} = \\frac {1}{\\| \\mathbf {T} ^ {\\prime} \\|} \\sum_ {x \\in \\mathbf {T} ^ {\\prime}} \\min _ {y \\in \\widehat {\\mathbf {T}} ^ {\\prime}} \\| x - y \\| _ {2} + \\frac {1}{\\| \\widehat {\\mathbf {T}} ^ {\\prime} \\|} \\sum_ {y \\in \\widehat {\\mathbf {T}} ^ {\\prime}} \\min _ {x \\in \\mathbf {T} ^ {\\prime}} \\| y - x \\| _ {2} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 507, + 890, + 560 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Classification Accuracy. In case we use our network to predict the object class, we measure our misclassification error in terms of accuracy.", + "bbox": [ + 498, + 580, + 890, + 626 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Object pose Evaluation", + "text_level": 1, + "bbox": [ + 500, + 635, + 715, + 652 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In Tab. 1, we report the quantitative evaluation on the test set of the datasets, comparing our method to the baseline. Our approach significantly outperforms the baseline, even if this latter exploits the points order information. We obtain the most significant margin on the GRAB dataset, where objects are small and mainly involve hands, showing the precision of our method. The baseline cannot be applied on BEHAVE-Raw since it does not share the same number of vertices as the training set, while our method shows only a limited performance decrease, pointing to generalization also to point clouds coming from different sources. We report qualitative results of our method on GRAB (first two rows) and BEHAVE (last row) in Fig. 5, and on BEHAVE-Raw in Fig. 6. Finally, to further evaluate the generalization of our method to unseen poses and subjects, we also considered point clouds obtained by an egocentric pipeline.", + "bbox": [ + 496, + 659, + 893, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "4730", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ff078a73221b39123d11e20a8864b6816720a635944a6a5f07dcc1af96a469bd.jpg", + "image_caption": [ + "Figure 5. Qualitative results. Results of our method on GRAB (first and second rows) and BEHAVE (third row) datasets." + ], + "image_footnote": [], + "bbox": [ + 81, + 87, + 890, + 683 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this scenario, we record a user motion with an XSens system [46], retarget the pose to an SMPL+H model, and obtain the point cloud by sampling the resulting mesh. Outputs of our method can be observed in Fig. 1 and Fig. 7. Generalizing to unseen poses and subjects acquired with wearable systems prone to measurement errors opens several exciting applications for VR/XR contexts.", + "bbox": [ + 75, + 710, + 470, + 815 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Human Affordance analysis", + "text_level": 1, + "bbox": [ + 76, + 830, + 328, + 847 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We use our method to analyze human-object interaction considering three key factors: changes in the input information, the points saliency, and confusion in classification.", + "bbox": [ + 75, + 854, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Input information. Since we are curious to analyze how information is encoded in the inputs, we consider three more scenarios:", + "bbox": [ + 498, + 710, + 890, + 755 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Hands: We generate the input data using the MANO [47] hand model annotations included in GRAB dataset. Hence, the model should infer the pose of the object without relying on other body parts.", + "- SMPL: We consider all the points of the subject, but the ones from the hands are in the rest pose provided by SMPL [40]. In this way, we analyze how much the network captures the information of the finger pose." + ], + "bbox": [ + 517, + 767, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "4731", + "bbox": [ + 482, + 945, + 513, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/c2e7c468e0c5d55fcb5d29431530726921dbb8b302b418aca884ac8d3668161a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsGRABBEHAVEBEHAVE-Raw
Ec ↓Ev2v ↓Ec ↓Ev2v ↓Ec ↓Ev2v ↓
Ours, hands0.30150.7358----
Ours, SMPL0.02450.10090.06670.28770.08430.3205
Ours, SMPLH0.02370.09430.06630.29000.08060.3143
Ours, SMPLH+T0.02350.09830.06860.28800.08230.3161
", + "bbox": [ + 238, + 88, + 733, + 172 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Human Affordance. We use our pipeline to explore how different inputs affect the object pose prediction. Using full body provides richer features for object recovery.", + "bbox": [ + 75, + 185, + 893, + 213 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e346cad04e13a3ab85f8e14b6239ebad852988c6653614dae40f3b10b6ebcc36.jpg", + "image_caption": [ + "Figure 6. BEHAVE-Raw. Results of our method considering as input raw point clouds from BEHAVE dataset. Despite the amount of noise and occlusions, our method is able to generalize and performs reliably." + ], + "image_footnote": [], + "bbox": [ + 83, + 215, + 890, + 361 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2ab01b3ce5b89b9a309e2a7421656945da492d5b268444a505f1d9a1eff979dc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 89, + 402, + 872, + 545 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c6677345b13a7e64ad2cf0e8caf3b72a03378454766f907d656201d0446aa01d.jpg", + "image_caption": [ + "Figure 7. Human from IMUs. Results of our method from a motion sequence acquired with IMUs. Even if the subject is unseen and the motion is subject to noise in the IMUs sensors, our method produces stable results.", + "Figure 8. Saliency. Point cloud saliency computed for different objects, rendered from two perspectives. The contact region is relevant for all the interactions, while the network also focuses on the feet and head regions. All the predictions are results of our method." + ], + "image_footnote": [], + "bbox": [ + 101, + 580, + 872, + 720 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "- SMPLH+T: We consider all the points of the subject with the finger correctly posed (as in our method), and also we apply the temporal smoothing outlined in Sec. 3. Temporal information contextualizes the interactions and regularizes the predictions throughout the sequence.", + "bbox": [ + 94, + 767, + 472, + 859 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For the first two cases, we train ad-hoc networks. Results of this analysis are reported in Tab. 2. We notice that using", + "bbox": [ + 75, + 869, + 472, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "only hands is not sufficient to achieve satisfactory performance on the GRAB dataset. However, introducing more contextual information helps refine these results, unavailing information on other body parts.", + "bbox": [ + 496, + 767, + 893, + 829 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Points saliency. As further evidence that object interaction involves different body parts, we conducted a study to discover what input points are crucial for the network.", + "bbox": [ + 496, + 854, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "4732", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6c0bed31d233ed9d7fb2b05d4cf60f882659a2a69509444f956284e6d6ef5e57.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodGRABBEHAVE
Ec↓Ech↓Acc.↑Ec↓Ech↓Acc.↑
NN, SMPLH0.04040.193815.420.08800.387321.48
Ours, SMPLH0.02900.138720.800.07280.304624.16
Ours, SMPLH + T0.02630.138430.150.07220.306949.28
", + "bbox": [ + 245, + 88, + 723, + 160 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Object classification. Quantitative results of models with object class prediction. Introducing temporal information has a dramatic impact on object classification.", + "bbox": [ + 75, + 170, + 893, + 200 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6a16398cfdb9213cc50694e80b3efab326b67572229f4118376c82855fbd64d8.jpg", + "image_caption": [ + "Figure 9. Confusion Matrix. The confusion matrix for a subset of the predicted classes. Objects with similar functionality are often confused (Coffemug vs Cup), while ones that suggest characteristic human poses (Dorknoob, Hammer) are well separated." + ], + "image_footnote": [], + "bbox": [ + 122, + 213, + 411, + 412 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We follow a recent protocol to find 3D point cloud saliency [72]: we cast the input through the network, we compute a loss on the output (in our case, the one in Eq. (2)), and we modify the input using the backpropagated gradient. This procedure is performed iteratively, and we refer to [72] for the details. We report the results of this study in Fig. 8, where points modified by the procedure are highlighted in red. We find the results of this analysis fascinating. As expected, the contact region is always essential to infer the correct object location. However, feet play a crucial role in all the reported cases, since they provide information about the human position and the consequent pose of the body. Another highlighted region is the head: different orientations give clues about object location and body posture.", + "bbox": [ + 75, + 487, + 468, + 699 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Confusion in classification. As a final analysis, we explore the learning object classification during the training. We jointly train a further MLP module that takes as an input $\\mathbf{F}_o$ and predicts the object class, using a cross-entropy loss. When a time sequence of point clouds is available, we exploit it by selecting the class with the highest score across the frames and applying it to the whole sequence. We report results in Tab. 3. Our experiment suggests that the task is challenging while, given the number of classes (40), we still consider our results a promising first step. Also, we notice that temporal smoothing significantly helps classification accuracy. Temporal context disambiguate poses", + "bbox": [ + 75, + 719, + 468, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "without a clear functionality. In Fig. 9, we report the confusion matrix for a subset of the classes for our method. The misclassification mainly arises from interactions of objects with similar functionality.", + "bbox": [ + 498, + 226, + 890, + 287 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusions", + "text_level": 1, + "bbox": [ + 500, + 306, + 625, + 321 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we have addressed a novel and inspiring problem that changes the perspective on object-human interaction. Our proposed model is simple, carefully designed, and inspired by behavioural studies. We collected evidence of the method's effectiveness on a large set of object classes and empirically proved its generalization on noisy and different inputs. Finally, our analysis of human affordance is unprecedented, showing that human-object interaction can also involve body parts distant from the object and pointing to interesting relations useful for applications and subsequent works.", + "bbox": [ + 498, + 333, + 890, + 500 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations and Future Works. As the first exploration in this direction, our study enables several future possibilities. In this work, the temporal information is only used after the training procedure. Incorporating this information can create other patterns, further improving the results' quality. Our empirical evidence suggests that class prediction requires further investigation and more sophisticated techniques, like specialized attention mechanisms. Finally, we do not consider sequences that involve long (e.g., hours-long) and complex (e.g., multi-objects) interactions, which are difficult to capture. We hope our work can foster the community to collect such datasets.", + "bbox": [ + 496, + 529, + 890, + 710 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements Special thanks to the RVH team and reviewers, their feedback helped improve the manuscript. This work is funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) - 409792180 (EmmyNoether Programme, project: Real Virtual Humans) and the German Federal Ministry of Education and Research (BMBF): Tübingen AI Center, FKZ: 01IS18039A. G. Pons-Moll is a member of the Machine Learning Cluster of Excellence, EXC number 2064/1 - Project number 390727645. The authors thank the International Max Planck Research School for Intelligent Systems (IMPRS-IS) for supporting I. Petrov. R. Marin is supported by an Alexander von Humboldt Foundation Research Fellowship.", + "bbox": [ + 496, + 734, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "4733", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Bharat Lal Bhatnagar, Xianghui Xie, Ilya A. Petrov, Cristian Sminchisescu, Christian Theobalt, and Gerard Pons-Moll. BEHAV: Dataset and Method for Tracking Human Object Interactions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15935-15946, 2022. 3, 4", + "[2] Amy E Booth and Sandra Waxman. Object names and object functions serve as cues to categories for infants. Developmental psychology, 38(6):948, 2002. 1", + "[3] Samarth Brahmbhatt, Cusuh Ham, Charles C Kemp, and James Hays. Contactdb: Analyzing and predicting grasp contact via thermal imaging. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8709-8719, 2019. 3", + "[4] Samarth Brahmbhatt, Ankur Handa, James Hays, and Dieter Fox. Contactgrasp: Functional multi-finger grasp synthesis from contact. In 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 2386-2393. IEEE, 2019. 3", + "[5] Samarth Brahmbhatt, Chengcheng Tang, Christopher D Twigg, Charles C Kemp, and James Hays. Contactpose: A dataset of grasps with object contact and hand pose. In European Conference on Computer Vision, pages 361-378. Springer, 2020. 3", + "[6] Tim Brooks and Alexei A Efros. Hallucinating pose-compatible scenes. In European Conference on Computer Vision, pages 510-528. Springer, 2022. 3", + "[7] Zhe Cao, Hang Gao, Karttikeya Mangalam, Qi-Zhi Cai, Minh Vo, and Jitendra Malik. Long-term human motion prediction with scene context. In European Conference on Computer Vision, pages 387-404. Springer, 2020. 3", + "[8] Zhe Cao, Ilija Radosavovic, Angjoo Kanazawa, and Jitendra Malik. Reconstructing Hand-Object Interactions in the Wild. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12417–12426, 2021. 1, 3", + "[9] Laura A Carlson-Radvansky, Eric S Covey, and Kathleen M Lattanzi. “what” effects on “where”: Functional influences on spatial relations. *Psychological Science*, 10(6):516–521, 1999. 1", + "[10] Yu-Wei Chao, Wei Yang, Yu Xiang, Pavlo Molchanov, Ankur Handa, Jonathan Tremblay, Yashraj S Narang, Karl Van Wyk, Umar Iqbal, Stan Birchfield, et al. Dexycb: A benchmark for capturing hand grasping of objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9044–9053, 2021. 3", + "[11] R. Qi Charles, Hao Su, Mo Kaichun, and Leonidas J. Guibas. PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 77-85, July 2017. 3", + "[12] Yixin Chen, Siyuan Huang, Tao Yuan, Siyuan Qi, Yixin Zhu, and Song-Chun Zhu. Holistic++ scene understanding: Single-view 3d holistic scene parsing and human pose estimation with human-object interaction and physical commonsense. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8648-8657, 2019. 3" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Sammy Christen, Muhammed Kocabas, Emre Aksan, Jemin Hwangbo, Jie Song, and Otmar Hilliges. D-Grasp: Physically Plausible Dynamic Grasp Synthesis for Hand-Object Interactions. arXiv:2112.03028 [cs], Mar. 2022. 1", + "[14] Enric Corona, Gerard Pons-Moll, Guillem Alenyà, and Francesc Moreno-Noguer. Learned vertex descent: A new direction for 3d human model fitting. arXiv preprint arXiv:2205.06254, 2022. 4", + "[15] Enric Corona, Albert Pumarola, Guillem Alenya, Francesc Moreno-Noguer, and Grégory Rogez. Ganhand: Predicting human grasp affordances in multi-object scenes. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5031-5041, 2020. 3", + "[16] Rishabh Dabral, Soshi Shimada, Arjun Jain, Christian Theobalt, and Vladislav Golyanik. Gravity-Aware Monocular 3D Human-Object Reconstruction. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 12345–12354, Oct. 2021. 3", + "[17] Shengheng Deng, Xun Xu, Chaozheng Wu, Ke Chen, and Kui Jia. 3d affordancenet: A benchmark for visual object affordance understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1778-1787, 2021. 3", + "[18] Kiana Ehsani, Shubham Tulsiani, Saurabh Gupta, Ali Farhadi, and Abhinav Gupta. Use the force, luke! learning to predict physical forces by simulating effects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 224-233, 2020. 3", + "[19] Eleanor J Gibson. The concept of affordances in development: The renascence of functionalism. In *The concept of development: The Minnesota symposia on child psychology*. Vol, volume 15, pages 55–81, 1982. 1, 2", + "[20] John C Gower. Generalized procrustes analysis. Psychometrika, 40(1):33-51, 1975. 4", + "[21] Helmu t Grabner, Juergen Gall, and Luc Van Gool. What makes a chair a chair? In CVPR 2011, pages 1529-1536, Colorado Springs, CO, USA, June 2011. IEEE. 1, 2", + "[22] Patrick Grady, Chengcheng Tang, Christopher D. Twigg, Minh Vo, Samarth Brahmbhatt, and Charles C. Kemp. ContactOpt: Optimizing Contact to Improve Grasps. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1471-1481, June 2021. 1, 3", + "[23] Vladimir Guzov, Aymen Mir, Torsten Sattler, and Gerard Pons-Moll. Human poseitioning system (hps): 3d human pose estimation and self-localization in large scenes from body-mounted sensors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4318-4329, 2021. 3", + "[24] Vladimir Guzov, Torsten Sattler, and Gerard Pons-Moll. Visually plausible human-object interaction capture from wearable sensors. In arXiv, 2022. 2, 3", + "[25] Shreyas Hampali, Sayan Deb Sarkar, Mahdi Rad, and Vincent Lepetit. Keypoint transformer: Solving joint identification in challenging hands and object interactions for accurate 3d pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11090-11100, 2022. 3" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "4734", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Gregor Harih and Bojan Dolsak. Tool-handle design based on a digital human hand model. International Journal of Industrial Ergonomics, 43(4):288-295, July 2013. 2", + "[27] Mohamed Hassan, Duygu Ceylan, Ruben Villegas, Jun Saito, Jimei Yang, Yi Zhou, and Michael J Black. Stochastic scene-aware motion prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11374-11384, 2021. 3", + "[28] Mohamed Hassan, Vasileios Choutas, Dimitrios Tzionas, and Michael J Black. Resolving 3d human pose ambiguities with 3d scene constraints. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2282-2292, 2019. 3", + "[29] Yana Hasson, Gül Varol, Cordelia Schmid, and Ivan Laptev. Towards unconstrained joint hand-object reconstruction from rgb videos. In 2021 International Conference on 3D Vision (3DV), pages 659-668. IEEE, 2021. 3", + "[30] Yana Hasson, Gul Varol, Dimitrios Tzionas, Igor Kalevatykh, Michael J Black, Ivan Laptev, and Cordelia Schmid. Learning joint reconstruction of hands and manipulated objects. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11807-11816, 2019. 3", + "[31] Chun-Hao P. Huang, Hongwei Yi, Markus Höschle, Matvey Safroshkin, Tsvetelina Alexiadis, Senya Polikovsky, Daniel Scharstein, and Michael J. Black. Capturing and inferring dense full-body human-scene contact. In IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), pages 13274–13285, June 2022. 3", + "[32] Yinghao Huang, Omid Taheri, Michael J Black, and Dimitrios Tzionas. Intercap: Joint markerless 3d tracking of humans and objects in interaction. In DAGM German Conference on Pattern Recognition, pages 281-299. Springer, 2022. 3", + "[33] Yuheng Jiang, Suyi Jiang, Guoxing Sun, Zhuo Su, Kaiwen Guo, Minye Wu, Jingyi Yu, and Lan Xu. Neuralhofusion: Neural volumetric rendering under human-object interactions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6155-6165, 2022. 3", + "[34] Korrawe Karunratanakul, Jinlong Yang, Yan Zhang, Michael J. Black, Krikamol Muandet, and Siyu Tang. Grasping Field: Learning Implicit Representations for Human Grasps. In 2020 International Conference on 3D Vision (3DV), pages 333-344, Nov. 2020. 1, 3", + "[35] Hedvig Kjellström, Javier Romero, and Danica Kragić. Visual object-action recognition: Inferring object affordances from human demonstration. Computer Vision and Image Understanding, 115(1):81-90, Jan. 2011. 2, 3", + "[36] Taein Kwon, Bugra Tekin, Jan Stuhmer, Federica Bogo, and Marc Pollefeys. H2o: Two hands manipulating objects for first person interaction recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10138-10148, 2021. 3", + "[37] Steven M Lehar. The world in your head: A gestalt view of the mechanism of conscious experience. Psychology Press, 2003. 2" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Zhi Li, Soshi Shimada, Bernt Schiele, Christian Theobalt, and Vladislav Golyanik. Mocapdeform: Monocular 3d human motion capture in deformable scenes. In International Conference on 3D Vision (3DV), 2022. 3", + "[39] Yunze Liu, Yun Liu, Che Jiang, Kangbo Lyu, Weikang Wan, Hao Shen, Boqiang Liang, Zhoujie Fu, He Wang, and Li Yi. Hoi4d: A 4d egocentric dataset for category-level human-object interaction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21013-21022, 2022. 3", + "[40] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multiperson linear model. ACM transactions on graphics (TOG), 34(6):1-16, 2015. 6", + "[41] Priyanka Mandikal and Kristen Grauman. Learning dexterous grasping with object-centric visual affordances. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 6169-6176. IEEE, 2021. 3", + "[42] Yinyu Nie, Angela Dai, Xiaoguang Han, and Matthias Nießner. Pose2room: understanding 3d scenes from human activities. In European Conference on Computer Vision, pages 425-443. Springer, 2022. 3", + "[43] Lisa M Oakes and Kelly L Madole. Function revisited: How infants construe functional features in their representation of objects. In Advances in child development and behavior, volume 36, pages 135-185. Elsevier, 2008. 1", + "[44] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3d hands, face, and body from a single image. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10975-10985, 2019. 4", + "[45] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. PointNet++: Deep Hierarchical Feature Learning on Point Sets in a Metric Space. In Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc., 2017. 3", + "[46] Daniel Roetenberg, Henk Luinge, and Per Slycke. Moven: Full 6 dof human motion tracking using miniature inertial sensors. Xsen Technologies, December, 2007. 6", + "[47] Javier Romero, Dimitrios Tzionas, and Michael J. Black. Embodied hands: Modeling and capturing hands and bodies together. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 36(6), Nov. 2017. 6", + "[48] Manolis Savva, Angel X Chang, Pat Hanrahan, Matthew Fisher, and Matthias Nießner. Pigraphs: learning interaction snapshots from observations. ACM Transactions on Graphics (TOG), 35(4):1-12, 2016. 3", + "[49] Colin G. Scanes. Chapter 1 - animal perception including differences with humans. In Colin G. Scanes and Samia R. Toukhsati, editors, Animals and Human Society, pages 1-11. Academic Press, 2018. 1", + "[50] Zhuo Su, Lan Xu, Dawei Zhong, Zhong Li, Fan Deng, Shuxue Quan, and Lu Fang. Robustfusion: Robust volumetric performance reconstruction under human-object interactions from monocular rgbd stream. arXiv preprint arXiv:2104.14837, 2021. 3" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "4735", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[51] Guoxing Sun, Xin Chen, Yizhang Chen, Anqi Pang, Pei Lin, Yuheng Jiang, Lan Xu, Jingyi Yu, and Jingya Wang. Neural free-viewpoint performance rendering under complex human-object interactions. In Proceedings of the 29th ACM International Conference on Multimedia, pages 4651-4660, 2021. 3", + "[52] Zhongda Sun, Minglu Zhu, Xuechuan Shan, and Chengkuo Lee. Augmented tactile-perception and haptic-feedback rings as human-machine interfaces aiming for immersive interactions. Nature communications, 13(1):1-13, 2022. 3", + "[53] Subramanian Sundaram, Petr Kellnhofer, Yunzhu Li, Jun-Yan Zhu, Antonio Torralba, and Wojciech Matusik. Learning the signatures of the human grasp using a scalable tactile glove. Nature, 569(7758):698-702, May 2019. 3", + "[54] Omid Taheri, Vasileios Choutas, Michael J. Black, and Dimitrios Tzionas. GOAL: Generating 4D whole-body motion for hand-object grasping. In IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), pages 13263-13273, June 2022. 3", + "[55] Omid Taheri, Nima Ghorbani, Michael J. Black, and Dimitrios Tzionas. GRAB: A Dataset of Whole-Body Human Grasping of Objects. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, Computer Vision - ECCV 2020, volume 12349, pages 581-600. Springer International Publishing, Cham, 2020. 3, 4", + "[56] Weilin Wan, Lei Yang, Lingjie Liu, Zhuoying Zhang, Ruixing Jia, Yi-King Choi, Jia Pan, Christian Theobalt, Taku Komura, and Wenping Wang. Learn to predict how humans manipulate large-sized objects from interactive motions. IEEE Robotics and Automation Letters, 7(2):4702-4709, 2022. 3", + "[57] Zhenzhen Weng and Serena Yeung. Holistic 3d human and scene mesh estimation from single view images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 334-343, 2021. 3", + "[58] Yan Wu, Jiahao Wang, Yan Zhang, Siwei Zhang, Otmar Hilliges, Fisher Yu, and Siyu Tang. Saga: Stochastic whole-body grasping with contact. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 3", + "[59] Xianghui Xie, Bharat Lal Bhatnagar, and Gerard Pons-Moll. Chore: Contact, human and object reconstruction from a single rgb image. In European Conference on Computer Vision (ECCV). Springer, October 2022. 3", + "[60] Xianghui Xie, Bharat Lal Bhatnagar, and Gerard Pons-Moll. Visibility aware human-object interaction tracking from single rgb camera. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2023. 3", + "[61] Lixin Yang, Xinyu Zhan, Kailin Li, Wenqiang Xu, Jiefeng Li, and Cewu Lu. Cpf: Learning a contact potential field to model the hand-object interaction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11097-11106, 2021. 3", + "[62] Bangpeng Yao, Jiayuan Ma, and Li Fei-Fei. Discovering Object Functionality. In Proceedings of the IEEE International Conference on Computer Vision, pages 2512-2519, 2013. 1, 3", + "[63] Yufei Ye, Abhinav Gupta, and Shubham Tulsiani. What's in your hands? 3d reconstruction of generic objects in hands." + ], + "bbox": [ + 78, + 90, + 468, + 901 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3895-3905, 2022. 1, 3", + "[64] Hongwei Yi, Chun-Hao P. Huang, Dimitrios Tzionas, Muhammed Kocabas, Mohamed Hassan, Siyu Tang, Justus Thies, and Michael J. Black. Human-Aware Object Placement for Visual Environment Reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3959-3970, 2022. 3", + "[65] Andrei Zanfir, Elisabella Maroiniu, and Cristian Sminchisescu. Monocular 3d pose and shape estimation of multiple people in natural scenes-the importance of multiple scene constraints. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2148-2157, 2018. 3", + "[66] Jason Y Zhang, Panna Felsen, Angjoo Kanazawa, and Jitendra Malik. Predicting 3d human dynamics from video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7114-7123, 2019. 3", + "[67] Jason Y Zhang, Sam Pepose, Hanbyul Joo, Deva Ramanan, Jitendra Malik, and Angjoo Kanazawa. Perceiving 3d human-object spatial arrangements from a single image in the wild. In European conference on computer vision, pages 34-51. Springer, 2020. 3", + "[68] Siwei Zhang, Qianli Ma, Yan Zhang, Zhiyin Qian, Marc Pollefeys, Federica Bogo, and Siyu Tang. Egobody: Human body shape, motion and social interactions from head-mounted devices. arXiv preprint arXiv:2112.07642, 2021. 3", + "[69] Siwei Zhang, Yan Zhang, Federica Bogo, Marc Pollefeys, and Siyu Tang. Learning motion priors for 4d human body capture in 3d scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11343-11353, 2021. 3", + "[70] Xiaohan Zhang, Bharat Lal Bhatnagar, Sebastian Starke, Vladimir Guzov, and Gerard Pons-Moll. Couch: Towards controllable human-chair interactions. In European Conference on Computer Vision (ECCV). Springer, October 2022. 3", + "[71] Qian Zheng, Weikai Wu, Hanting Pan, Niloy Mitra, Daniel Cohen-Or, and Hui Huang. Inferring object properties from human interaction and transferring them to new motions. Computational Visual Media, 7(3):375–392, Sept. 2021. 3", + "[72] Tianhang Zheng, Changyou Chen, Junsong Yuan, Bo Li, and Kui Ren. Pointcloud saliency maps. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1598-1606, 2019. 8", + "[73] Keyang Zhou, Bharat Lal Bhatnagar, Jan Eric Lenssen, and Gerard Pons-Moll. TOCH: Spatio-Temporal Object-to-Hand Correspondence for Motion Refinement, July 2022. 1, 3" + ], + "bbox": [ + 503, + 92, + 890, + 796 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "4736", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Object Pop-Up_ Can We Infer 3D Objects and Their Poses From Human Interactions Alone_/965b59c8-d0c6-413b-852b-108c8ff0bbe7_model.json b/2023/Object Pop-Up_ Can We Infer 3D Objects and Their Poses From Human Interactions Alone_/965b59c8-d0c6-413b-852b-108c8ff0bbe7_model.json new file mode 100644 index 0000000000000000000000000000000000000000..9366a92b4ef03e28843d8e770f606642766add5d --- /dev/null +++ b/2023/Object Pop-Up_ Can We Infer 3D Objects and Their Poses From Human Interactions Alone_/965b59c8-d0c6-413b-852b-108c8ff0bbe7_model.json @@ -0,0 +1,2664 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.139, + 0.131, + 0.833, + 0.175 + ], + "angle": 0, + "content": "Object pop-up: Can we infer 3D objects and their poses from human interactions alone?" + }, + { + "type": "text", + "bbox": [ + 0.175, + 0.215, + 0.794, + 0.235 + ], + "angle": 0, + "content": "Ilya A. Petrov\\(^{1}\\), Riccardo Marin\\(^{1}\\), Julian Chibane\\(^{2,1}\\), and Gerard Pons-Moll\\(^{1,2,3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.38, + 0.252, + 0.59, + 0.269 + ], + "angle": 0, + "content": "1University of Tübingen, Germany" + }, + { + "type": "text", + "bbox": [ + 0.251, + 0.269, + 0.718, + 0.283 + ], + "angle": 0, + "content": "\\(^{2}\\)Max Planck Institute for Informatics, Saarland Informatics Campus, Germany" + }, + { + "type": "text", + "bbox": [ + 0.392, + 0.283, + 0.58, + 0.297 + ], + "angle": 0, + "content": "\\(^{3}\\)Tübingen AI Center, Germany" + }, + { + "type": "text", + "bbox": [ + 0.122, + 0.297, + 0.848, + 0.326 + ], + "angle": 0, + "content": "{i.petrov, gerard.pons-moll}@uni-tuebingen.de, riccardo.marin@mnf.uni-tuebingen.de, jchibane@mpi-inf.mpg.de" + }, + { + "type": "image_caption", + "bbox": [ + 0.084, + 0.353, + 0.14, + 0.372 + ], + "angle": 0, + "content": "Ref. Image (not used)" + }, + { + "type": "image", + "bbox": [ + 0.078, + 0.374, + 0.149, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.169, + 0.354, + 0.196, + 0.362 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image_caption", + "bbox": [ + 0.152, + 0.363, + 0.209, + 0.372 + ], + "angle": 0, + "content": "Point Cloud" + }, + { + "type": "image", + "bbox": [ + 0.154, + 0.374, + 0.209, + 0.497 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.229, + 0.36, + 0.261, + 0.37 + ], + "angle": 0, + "content": "Result" + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.372, + 0.269, + 0.497 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.29, + 0.353, + 0.343, + 0.372 + ], + "angle": 0, + "content": "Ref. Image (not used)" + }, + { + "type": "image", + "bbox": [ + 0.282, + 0.374, + 0.352, + 0.497 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.355, + 0.353, + 0.415, + 0.372 + ], + "angle": 0, + "content": "Input Point Cloud" + }, + { + "type": "image", + "bbox": [ + 0.356, + 0.374, + 0.411, + 0.496 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.436, + 0.36, + 0.468, + 0.37 + ], + "angle": 0, + "content": "Result" + }, + { + "type": "image", + "bbox": [ + 0.422, + 0.374, + 0.474, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.493, + 0.353, + 0.545, + 0.372 + ], + "angle": 0, + "content": "Ref. Image (not used)" + }, + { + "type": "image", + "bbox": [ + 0.482, + 0.374, + 0.553, + 0.496 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.56, + 0.353, + 0.619, + 0.372 + ], + "angle": 0, + "content": "Input Point Cloud" + }, + { + "type": "image", + "bbox": [ + 0.558, + 0.374, + 0.614, + 0.494 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.636, + 0.36, + 0.67, + 0.37 + ], + "angle": 0, + "content": "Result" + }, + { + "type": "image", + "bbox": [ + 0.625, + 0.374, + 0.679, + 0.494 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.694, + 0.353, + 0.748, + 0.372 + ], + "angle": 0, + "content": "Ref. Image (not used)" + }, + { + "type": "image", + "bbox": [ + 0.687, + 0.374, + 0.757, + 0.496 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.765, + 0.353, + 0.822, + 0.372 + ], + "angle": 0, + "content": "Input Point Cloud" + }, + { + "type": "image", + "bbox": [ + 0.758, + 0.374, + 0.811, + 0.497 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.828, + 0.353, + 0.865, + 0.372 + ], + "angle": 0, + "content": "Result" + }, + { + "type": "image", + "bbox": [ + 0.822, + 0.374, + 0.873, + 0.497 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.51, + 0.895, + 0.566 + ], + "angle": 0, + "content": "Figure 1. Human-Centric prediction of Human-Object Interaction. We explore Human-Object interaction when only the cues of the human pose interacting with an unobserved object are available ('Input'). We propose a neural model that, for the first time, can infer the location of the object ('Result') from such input. This is possible even when our subject simulates the interactions with the object ('Reference Image')." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.583, + 0.314, + 0.599 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.074, + 0.627, + 0.473, + 0.884 + ], + "angle": 0, + "content": "The intimate entanglement between objects affordances and human poses is of large interest, among others, for behavioural sciences, cognitive psychology, and Computer Vision communities. In recent years, the latter has developed several object-centric approaches: starting from items, learning pipelines synthesizing human poses and dynamics in a realistic way, satisfying both geometrical and functional expectations. However, the inverse perspective is significantly less explored: Can we infer 3D objects and their poses from human interactions alone? Our investigation follows this direction, showing that a generic 3D human point cloud is enough to pop up an unobserved object, even when the user is just imitating a functionality (e.g., looking through a binocular) without involving a tangible counterpart. We validate our method qualitatively and quantitatively, with synthetic data and sequences acquired for the task, showing applicability for XR/VR." + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.583, + 0.631, + 0.598 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Complex interactions with the world are among the unique skills distinguishing humans from other living beings. Even though our perception might be imperfect (we cannot hear ultrasonic sounds or see ultraviolet light [49]), our cognitive representation is enriched with a functional perspective, i.e., potential ways of interacting with objects or, as introduced by Gibson and colleagues, the affordance of the objects [19]. Several behavioural studies confirmed the centrality of this concept [2, 9, 43], which plays a fundamental role also for kids' development [2]. Computer Vision is well-aware that the function of an object complements its appearance [21], and exploited this in tasks like human and object reconstruction [62]. Previous literature approaches the interaction analysis from an object perspective (i.e., given an object, analyze the human interaction) [8,22,63], building object-centric priors [73], generating realistic grasps given the object [13, 34], reconstructing hand-object interactions [8,22,63]. Namely, objects induce functionality, so a human interaction (e.g., a mug suggests" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4726" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.088, + 0.452, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.287, + 0.471, + 0.358 + ], + "angle": 0, + "content": "Figure 2. Human-Centric vs. Object-Centric inference. A common perspective in the prior work is to infer affordances or human pose starting from an object. We explore the inverse, Human-Centric perspective of the human-object interaction relationship: given the human, we are interested in predicting the object." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.372, + 0.47, + 0.492 + ], + "angle": 0, + "content": "a drinking action; an handle a grasping one). For the first time, our work reverts the perspective, suggesting that analyzing human motion and behaviour is naturally a human-centric problem (i.e., given a human interaction, what kind of functionality is it suggesting, Fig. 2). Moving the first step in this new research direction, we pose a fundamental question: Can we infer 3D objects and their poses from human interactions alone?" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.493, + 0.471, + 0.78 + ], + "angle": 0, + "content": "At first sight, the problem seems particularly hard and significantly under-constrained since several geometries might fit the same action. However, the human body complements this information in several ways: physical relations, characteristic poses, or body dynamics serve as valuable proxies for the involved functionality, as suggested by Fig. 1. Such hints are so powerful that we can easily imagine the kind of object and its location, even if such an object does not exist. Furthermore, even if the given pose might fit several possible solutions, our mind naturally comes to the most natural one indicated by the observed behaviour. It also suggests that solely focusing on the contact region (an approach often preferred by previous works) is insufficient in this new viewpoint. The reification principle of Gestalt psychology [37] highlights that \"the whole\" arises from \"the parts\" and their relation. Similarly, in Fig. 3, the hand grasps in B) pop up a binocular in our mind because we naturally consider the relationship with the other body parts." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.47, + 0.886 + ], + "angle": 0, + "content": "Finally, moving to a human-centric perspective in human-object interaction is critical for human studies and daily-life applications. Modern systems for AR/VR [24], and digital interaction [26] are both centered on humans, often manipulating objects that do not have a real-world counterpart. Learning to decode an object from human behaviour enables unprecedented applications." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.47, + 0.901 + ], + "angle": 0, + "content": "To answer our question, we deploy a first straightforward" + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.097, + 0.668, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.584, + 0.216, + 0.603, + 0.229 + ], + "angle": 0, + "content": "A)" + }, + { + "type": "image", + "bbox": [ + 0.671, + 0.099, + 0.868, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.774, + 0.216, + 0.792, + 0.228 + ], + "angle": 0, + "content": "B)" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.235, + 0.892, + 0.291 + ], + "angle": 0, + "content": "Figure 3. Reification principle. Gestalt theory suggests our perception considers the whole before the single parts (A). This served as inspiration for our work: the object can arise considering the body parts as a whole (B)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.317, + 0.892, + 0.497 + ], + "angle": 0, + "content": "and effective pipeline to \"pop up\" a rigid object from a 3D human point cloud. Starting from the input human point cloud and a class, we train an end-to-end pipeline to infer object location. In case a temporal sequence of point clouds is available, we suggest post-processing to avoid jittering and inconsistencies in the predictions, showing the relevance of this information to handle ambiguous poses. We show promising results on previously unaddressed tasks in digital and real-world scenes. Finally, our method allows us to analyze different features of human behaviour, highlight their contribution to object retrieval, and point to exciting directions for future works." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.499, + 0.786, + 0.513 + ], + "angle": 0, + "content": "In summary, our main contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.523, + 0.892, + 0.569 + ], + "angle": 0, + "content": "1. We formulate a novel problem, changing the perspective taken by previous works in the field, and open to a yet unexplored research direction;" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.578, + 0.892, + 0.609 + ], + "angle": 0, + "content": "2. We introduce a method capable of predicting the object starting from an input human point cloud;" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.618, + 0.892, + 0.694 + ], + "angle": 0, + "content": "3. We analyze different components of the human-object relationship: the contribution of different pieces of interactions (hands, body, time sequence), the point-wise saliency of input points, and the confusion produced by objects with similar functions." + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.523, + 0.892, + 0.694 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.704, + 0.816, + 0.718 + ], + "angle": 0, + "content": "All the code will be made publicly available." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.732, + 0.641, + 0.748 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.757, + 0.696, + 0.774 + ], + "angle": 0, + "content": "2.1. Object Functionality" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.871 + ], + "angle": 0, + "content": "In the core of human perception of objects, functionality complements physical appearance, enhancing our perception. Gibson [19] introduced the idea that humans use affordances of objects for perception. Affordance can be defined as: \"an intrinsic property of an object, allowing an action to be performed with the object\" [35]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "From a Computer Vision perspective, object functionality supports several tasks such as scene analysis [21], object" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4727" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.152 + ], + "angle": 0, + "content": "classification [17, 35], object properties inferring [71], and it is also possible to learn object-specific human interaction models from 2D images [62]. These works suggest an intimate entanglement between the action and the object itself." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.161, + 0.318, + 0.177 + ], + "angle": 0, + "content": "2.2. Human-Object interaction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.184, + 0.469, + 0.245 + ], + "angle": 0, + "content": "Modelling environment-aware humans and their interactions in 3D is one of the most recent challenges to creating virtual humans. We see two main lines of work there: hand-focused and full-body ones." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.262, + 0.47, + 0.474 + ], + "angle": 0, + "content": "Hand-Object Interaction. Several works tackle the problem of human interaction, focusing solely on the hands [10]. This has been done starting from 2D [8, 15, 18, 22, 25, 29, 30, 34, 61, 63], 2.5D [3, 5], and 3D data [4, 34, 55]. Particularly promising seems the application of well-designed priors for the motion [73]. The class of objects involved in these works are mainly limited to graspable ones. We argue that interactions involving different body parts are common in everyday life, more attractive from an applicative perspective, and more challenging. Moreover, full-body context is crucial in reconstructing even grasp interactions since body pose contains information on an object's properties, e.g. pounding with a hammer affects the whole posture to support the action." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.492, + 0.47, + 0.763 + ], + "angle": 0, + "content": "Fully-Body Interaction. On this line, several works focus on the interaction between a human and a scene [12,28, 31,38,57,64,65,69]. Also, in this case, priors can be used to regularize the motion [48]. Several datasets are also available to study the interaction between a human and a single object. For example, recent BEHAVE [1], GRAB [55], and InterCap [32] capture full-body interactions with diverse objects. Works address the task of humans interactions reconstruction from different kinds of data sources like single image [59, 67], video [16, 50, 60], and multiview capturing [1, 33, 51], and synthetization of them as well [7,27,41,54,56,58,66,70]. However, in all these works, we observe a general object-centric perspective: given a scene or an object, they aim to recreate the humans interacting with them. We argue that the significantly less explored complementary one has more concrete applications in daily life, especially in VR/XR contexts where the human is central to the system." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.772, + 0.322, + 0.789 + ], + "angle": 0, + "content": "2.3. Human-centric perspective" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.469, + 0.901 + ], + "angle": 0, + "content": "While the general trend mainly focuses on the surrounding environment and objects, there is a growing interest and availability of egocentric tools for humans [23, 68] also interacting with objects [24,36,39]. They provide a subjective view and a valuable paradigm for several applications, like letting a user interact with objects in the digital world. Recent works also involve more sophisticated devices [52,53]," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.228 + ], + "angle": 0, + "content": "while they are still far from applicability. In a similar direction to our work, others propose to recover objects arrangements in a room starting from the human motion [42], to hallucinate a coherent 2D image from a human pose [6], or predicting physical properties of the objects (e.g., the weight of a box) from human joints [71]. While the principle inspires us, our study significantly differs: we focus on object pose and its spatial relation with humans, starting solely from unordered point clouds." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.241, + 0.591, + 0.256 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.266, + 0.892, + 0.312 + ], + "angle": 0, + "content": "This section describes our setting and the main components of our methodology, both at inference and training time. An overview of our pipeline can be found in Fig. 4." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.32, + 0.652, + 0.337 + ], + "angle": 0, + "content": "3.1. Object Pop-Up" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.343, + 0.892, + 0.465 + ], + "angle": 0, + "content": "Input. Our method starts from a single human point cloud \\(\\mathbf{P} \\in \\mathbb{R}^{N_P \\times 3}\\) with \\(N_P\\) points and a hot-encoded object class \\(c\\). The input point cloud can result from 3D/4D scans, IMUs template fitting, or any other shape-from-X approach. Regardless of the point cloud source, we remark that no further information is used apart from the 3D coordinates of the points. We represent each object as 1500 key points \\(\\mathbf{K}_c\\) uniformly sampled on the template mesh." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.483, + 0.892, + 0.664 + ], + "angle": 0, + "content": "Object Center. Training a model to predict an object pose from a human point cloud poses several challenges. Such a task requires the network to understand the location of different body parts and their subtle relations while jointly developing a sense of its spatial relationship with the human. Empirically, we observed that this is only feasible by carefully deconstructing the problem and designing different features to ease the learning process. As the first step to decompose this problem, we train a PointNet++ architecture [11,45] to predict the object center \\(\\mathbf{o}_P\\) starting from \\(\\mathbf{P}\\). At training time, this is supervised with an L2 loss against the ground truth center \\(\\widehat{\\mathbf{o}}_P\\):" + }, + { + "type": "equation", + "bbox": [ + 0.611, + 0.674, + 0.891, + 0.691 + ], + "angle": 0, + "content": "\\[\nL _ {\\mathbf {o} _ {P}} (\\mathbf {P}) = \\left\\| \\mathbf {o} _ {P} - \\widehat {\\mathbf {o}} _ {P} \\right\\| _ {2} ^ {2}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.702, + 0.892, + 0.779 + ], + "angle": 0, + "content": "Solving this task provides a good initialization for the object pose, and we move the key points associated with the input class \\( c \\) to the predicted center \\( \\mathbf{o}_P \\). Also, the features \\( \\mathbf{F}_o \\in \\mathbb{R}^{512} \\) extracted by the network encode import information on the whole human body." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Local Neighbourhood. The center prediction module can be further exploited using the nearest human regions. Intuitively, considering the closest body parts is essential to infer a contact relationship, but also the influence of body parts not directly in touch with the object (e.g., head orientation while using binoculars). To learn these connections, we consider the centered key points together with" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4728" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.111, + 0.084, + 0.862, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.258, + 0.895, + 0.3 + ], + "angle": 0, + "content": "Figure 4. Object pop-up. Our method predicts the position of an object, starting only from an input point cloud and an object class. It relies on a careful problem decomposition in several sub-tasks, extracting features that involve the entire human body and relations between body parts near the object." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.307, + 0.471, + 0.369 + ], + "angle": 0, + "content": "the 3000 closest points of the input human point cloud: \\(\\mathbf{P}_L = KNN(\\mathbf{P},\\mathbf{o}_P)\\). We pass these two sets as a unique point cloud to a PointNet++ network, obtaining a new set of per-point features \\(\\mathbf{F}_{P_L}\\in \\mathbb{R}^{128}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.385, + 0.47, + 0.552 + ], + "angle": 0, + "content": "Object displacement. To predict the object's final position, we empirically observed that directly predicting a rotation and a translation is not a good solution. Inspired by recent works that suggest a point-wise offset prediction to recover 3D human shapes [14], we apply a similar approach to our task. Our goal is to predict a point-wise shift \\(\\mathbf{S}_K\\) for the \\(\\mathbf{K}_c\\) vertices to align them to the target pose. We append the features \\(\\mathbf{F}_o\\), \\(\\mathbf{F}_{P_L}\\), the one-hot encoding of the object class \\(c\\), and a positional encoding to the centered key points, and we pass them to a decoder. At training time, we consider the following loss:" + }, + { + "type": "equation", + "bbox": [ + 0.138, + 0.557, + 0.47, + 0.577 + ], + "angle": 0, + "content": "\\[\nL _ {o f f} \\left(\\mathbf {K} _ {c}, \\mathbf {F} _ {o}, \\mathbf {F} _ {P _ {L}}, c\\right) = \\left\\| \\mathbf {S} _ {K} - \\widehat {\\mathbf {S}} _ {K} \\right\\| _ {F} ^ {2}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.583, + 0.4, + 0.598 + ], + "angle": 0, + "content": "The network is then trained end-to-end using:" + }, + { + "type": "equation", + "bbox": [ + 0.205, + 0.606, + 0.47, + 0.623 + ], + "angle": 0, + "content": "\\[\nL = L _ {\\mathbf {o} _ {P}} + \\alpha L _ {\\text {o f f}}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.629, + 0.321, + 0.644 + ], + "angle": 0, + "content": "The weighting coefficient is \\(\\alpha = 10\\)" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.652, + 0.238, + 0.669 + ], + "angle": 0, + "content": "3.2. Template fitting" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.469, + 0.78 + ], + "angle": 0, + "content": "Procrustes alignment. The point-wise offset produced by the network potentially distorts the key points structure in a non-rigid way. To recover the desired global rigid transformation, we rely on a Procrustes alignment [20]. This procedure takes as input two point clouds and returns the rotation \\(\\mathbf{R}\\) and the translation \\(\\mathbf{t}\\) to minimize the L2 distances of the points:" + }, + { + "type": "equation", + "bbox": [ + 0.215, + 0.781, + 0.469, + 0.797 + ], + "angle": 0, + "content": "\\[\nP (\\cdot , \\cdot) \\rightarrow (\\mathbf {R}, \\mathbf {t}). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.801, + 0.469, + 0.831 + ], + "angle": 0, + "content": "We apply this to the template key points and their configuration obtained with our network:" + }, + { + "type": "equation", + "bbox": [ + 0.156, + 0.839, + 0.469, + 0.856 + ], + "angle": 0, + "content": "\\[\nP \\left(\\mathbf {K} _ {c}, \\mathbf {K} _ {c} + \\mathbf {o} _ {P} + \\mathbf {S} _ {K}\\right)\\rightarrow (\\mathbf {R}, \\mathbf {t}). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.863, + 0.402, + 0.878 + ], + "angle": 0, + "content": "Finally, we recover the desired object pose as:" + }, + { + "type": "equation", + "bbox": [ + 0.221, + 0.885, + 0.469, + 0.901 + ], + "angle": 0, + "content": "\\[\n\\mathbf {T} ^ {\\prime} = \\mathbf {R T} _ {c} + \\mathbf {t} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.307, + 0.895, + 0.489 + ], + "angle": 0, + "content": "Time Smoothing. While our pipeline is designed to work with a single point cloud as input, considering the temporal evolution of interaction is often crucial, shaping the context of the individual poses. If a temporal sequence of point clouds is available, we provide a post-processing smoothing technique to take advantage of this further information. After running our method for each frame, we smooth the centering prediction across the sequence using a Gaussian kernel. Later, we will discuss a variation of our approach that also predicts the object class. In that case, we consider the most frequent class prediction over the whole set of frames to fix a class for the sequence." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.506, + 0.634, + 0.523 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.532, + 0.893, + 0.608 + ], + "angle": 0, + "content": "In this Section, we will describe the datasets used for training and testing our method. Then, we will present the baseline and the evaluation metrics. Finally, we will provide validation of our method as well as an analysis of its extended version that allows object class prediction." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.633, + 0.892, + 0.663 + ], + "angle": 0, + "content": "Datasets We jointly train on the union of BEHAVE [1] and GRAB [55], obtaining a set of:" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.676, + 0.891, + 0.707 + ], + "angle": 0, + "content": "15 subjects, first 8 subjects from the GRAB dataset and 7 subjects from the official training part of BEHAV;" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.721, + 0.892, + 0.766 + ], + "angle": 0, + "content": "- 40 different classes of objects, including all 20 objects from the BEHAVE dataset and 20 selected objects from the GRAB dataset;" + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.676, + 0.892, + 0.766 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We downsample training sequences of GRAB and BEHAVE to \\(10fps\\). To evaluate our method, we select subjects 9 and 10 from the GRAB dataset and downsample the sequences to \\(30fps\\). For the BEHAVE dataset, we use the official test part, which includes all sequences at \\(1fps\\) with subject 3 and part of the sequences with subjects 4, 5. As an input, we use point clouds with 9000 points sampled uniformly over the SMPL-H [44] meshes. We refer to raw" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4729" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.264, + 0.089, + 0.71, + 0.147 + ], + "angle": 0, + "content": "
MethodsGRABBEHAVEBEHAVE-Raw
Ec↓Ev2v↓Ec↓Ev2v↓Ec↓Ev2v↓
NN0.03620.14450.08020.3445--
Ours0.02370.09430.06630.29000.08060.3143
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.158, + 0.895, + 0.189 + ], + "angle": 0, + "content": "Table 1. Comparison with the baseline. Our method significantly outperforms the baseline, even though the baseline uses the vertex order as additional information. Moreover, the baseline method does not generalize to point clouds with an arbitrary number of vertices." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.2, + 0.471, + 0.247 + ], + "angle": 0, + "content": "point clouds from the BEHAVE dataset used in our experiments as BEHAVE-Raw. We use point clouds that are fused from 4 Kinect sensors and subsample \\(90k\\) points from them." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.277, + 0.471, + 0.325 + ], + "angle": 0, + "content": "Data augmentation. During training, to simulate errors in the center prediction, we randomly translate and rotate the object around the ground-truth center \\(\\widehat{\\mathbf{o}}_P\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.355, + 0.47, + 0.476 + ], + "angle": 0, + "content": "Implementation details. We implement our method using PyTorch framework and use Nvidia RTX3090 GPU for training and evaluation. The model is trained using Adam optimizer for 60 epochs with a learning rate of \\(1e^{-4}\\), which decays 10 times after 30-th and 40-th epochs. For the first 20 epochs of training, we use ground-truth object center \\(\\widehat{\\mathbf{o}}_P\\) instead of predicted \\(\\mathbf{o}_P\\) to select local neighborhood \\(\\mathbf{F}_{P_L}\\) to warm up the local PointNet++ encoder." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.507, + 0.471, + 0.705 + ], + "angle": 0, + "content": "Nearest-Neighbor Baseline. Since we are the first to tackle this task and no competitors are available, we propose a simple while informative baseline. Given the input point cloud, we recover the most similar in the training dataset in an L2 sense. Then, we recover the object handled by that subject and pose it in space in the same way. This baseline demonstrates that the task is non-trivial and the generalization to unseen poses and subjects of our method. Also, this baseline requires that the target point cloud and the ones in the training set share the same number of points. Hence, if the input point cloud is a raw scan, this baseline is not applicable. Our method, instead, does not rely on this assumption and is more general." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Object classification. In our research, we also investigate the possibility of incorporating class prediction inside the network training. This task is significantly difficult at a single-frame level since an isolated pose often does not suggest a clear functionality. However, including this step is interesting to analyze the interaction and the nature of the network confusion. Hence, we modify our method by adding a decoder module that takes the global features \\(\\mathbf{F}_o\\) and the local ones \\(\\mathbf{F}_{P_L}\\) as input to predict the object class. Then, we add to the training a simple cross-entropy loss between the predicted class and the ground truth one \\(\\widehat{C}_P\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.2, + 0.597, + 0.215 + ], + "angle": 0, + "content": "4.1. Metrics" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.223, + 0.894, + 0.269 + ], + "angle": 0, + "content": "In qualitative experiments, we use three main metrics to evaluate our results. In the tables, we report the average error across the considered test samples." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.288, + 0.893, + 0.349 + ], + "angle": 0, + "content": "Vertex-to-vertex. In most cases, our resulting object and the target one share the same number of vertices. Hence, we can compute the error between our prediction and the ground truth \\(\\widehat{\\mathbf{T}}'\\) as a point-to-point error:" + }, + { + "type": "equation", + "bbox": [ + 0.623, + 0.359, + 0.893, + 0.379 + ], + "angle": 0, + "content": "\\[\nE _ {v 2 v} = \\left\\| \\mathbf {T} ^ {\\prime} - \\widehat {\\mathbf {T}} ^ {\\prime} \\right\\| _ {F}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.39, + 0.892, + 0.421 + ], + "angle": 0, + "content": "When such error is computed only between the object centers, we will refer to it as \\( E_{c} \\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.439, + 0.894, + 0.5 + ], + "angle": 0, + "content": "Chamfer distance. When we evaluate the network that also predicts the class, target objects and selected templates might not share the same number of vertices. In that case, as a metric we use bi-directional Chamfer distance:" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.508, + 0.892, + 0.561 + ], + "angle": 0, + "content": "\\[\nE _ {c h} = \\frac {1}{\\| \\mathbf {T} ^ {\\prime} \\|} \\sum_ {x \\in \\mathbf {T} ^ {\\prime}} \\min _ {y \\in \\widehat {\\mathbf {T}} ^ {\\prime}} \\| x - y \\| _ {2} + \\frac {1}{\\| \\widehat {\\mathbf {T}} ^ {\\prime} \\|} \\sum_ {y \\in \\widehat {\\mathbf {T}} ^ {\\prime}} \\min _ {x \\in \\mathbf {T} ^ {\\prime}} \\| y - x \\| _ {2} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.581, + 0.892, + 0.627 + ], + "angle": 0, + "content": "Classification Accuracy. In case we use our network to predict the object class, we measure our misclassification error in terms of accuracy." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.636, + 0.717, + 0.653 + ], + "angle": 0, + "content": "4.2. Object pose Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.895, + 0.903 + ], + "angle": 0, + "content": "In Tab. 1, we report the quantitative evaluation on the test set of the datasets, comparing our method to the baseline. Our approach significantly outperforms the baseline, even if this latter exploits the points order information. We obtain the most significant margin on the GRAB dataset, where objects are small and mainly involve hands, showing the precision of our method. The baseline cannot be applied on BEHAVE-Raw since it does not share the same number of vertices as the training set, while our method shows only a limited performance decrease, pointing to generalization also to point clouds coming from different sources. We report qualitative results of our method on GRAB (first two rows) and BEHAVE (last row) in Fig. 5, and on BEHAVE-Raw in Fig. 6. Finally, to further evaluate the generalization of our method to unseen poses and subjects, we also considered point clouds obtained by an egocentric pipeline." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "4730" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.083, + 0.088, + 0.891, + 0.684 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.122, + 0.689, + 0.846, + 0.703 + ], + "angle": 0, + "content": "Figure 5. Qualitative results. Results of our method on GRAB (first and second rows) and BEHAVE (third row) datasets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.711, + 0.471, + 0.816 + ], + "angle": 0, + "content": "In this scenario, we record a user motion with an XSens system [46], retarget the pose to an SMPL+H model, and obtain the point cloud by sampling the resulting mesh. Outputs of our method can be observed in Fig. 1 and Fig. 7. Generalizing to unseen poses and subjects acquired with wearable systems prone to measurement errors opens several exciting applications for VR/XR contexts." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.831, + 0.33, + 0.848 + ], + "angle": 0, + "content": "4.3. Human Affordance analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.903 + ], + "angle": 0, + "content": "We use our method to analyze human-object interaction considering three key factors: changes in the input information, the points saliency, and confusion in classification." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.711, + 0.892, + 0.756 + ], + "angle": 0, + "content": "Input information. Since we are curious to analyze how information is encoded in the inputs, we consider three more scenarios:" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.768, + 0.894, + 0.83 + ], + "angle": 0, + "content": "- Hands: We generate the input data using the MANO [47] hand model annotations included in GRAB dataset. Hence, the model should infer the pose of the object without relying on other body parts." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.84, + 0.895, + 0.903 + ], + "angle": 0, + "content": "- SMPL: We consider all the points of the subject, but the ones from the hands are in the rest pose provided by SMPL [40]. In this way, we analyze how much the network captures the information of the finger pose." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.768, + 0.895, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.514, + 0.957 + ], + "angle": 0, + "content": "4731" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.24, + 0.089, + 0.734, + 0.174 + ], + "angle": 0, + "content": "
MethodsGRABBEHAVEBEHAVE-Raw
Ec ↓Ev2v ↓Ec ↓Ev2v ↓Ec ↓Ev2v ↓
Ours, hands0.30150.7358----
Ours, SMPL0.02450.10090.06670.28770.08430.3205
Ours, SMPLH0.02370.09430.06630.29000.08060.3143
Ours, SMPLH+T0.02350.09830.06860.28800.08230.3161
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.186, + 0.895, + 0.214 + ], + "angle": 0, + "content": "Table 2. Human Affordance. We use our pipeline to explore how different inputs affect the object pose prediction. Using full body provides richer features for object recovery." + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.216, + 0.891, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.369, + 0.897, + 0.398 + ], + "angle": 0, + "content": "Figure 6. BEHAVE-Raw. Results of our method considering as input raw point clouds from BEHAVE dataset. Despite the amount of noise and occlusions, our method is able to generalize and performs reliably." + }, + { + "type": "image", + "bbox": [ + 0.091, + 0.404, + 0.874, + 0.546 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.552, + 0.895, + 0.581 + ], + "angle": 0, + "content": "Figure 7. Human from IMUs. Results of our method from a motion sequence acquired with IMUs. Even if the subject is unseen and the motion is subject to noise in the IMUs sensors, our method produces stable results." + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.582, + 0.874, + 0.721 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.726, + 0.895, + 0.756 + ], + "angle": 0, + "content": "Figure 8. Saliency. Point cloud saliency computed for different objects, rendered from two perspectives. The contact region is relevant for all the interactions, while the network also focuses on the feet and head regions. All the predictions are results of our method." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.768, + 0.473, + 0.86 + ], + "angle": 0, + "content": "- SMPLH+T: We consider all the points of the subject with the finger correctly posed (as in our method), and also we apply the temporal smoothing outlined in Sec. 3. Temporal information contextualizes the interactions and regularizes the predictions throughout the sequence." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.473, + 0.903 + ], + "angle": 0, + "content": "For the first two cases, we train ad-hoc networks. Results of this analysis are reported in Tab. 2. We notice that using" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.768, + 0.895, + 0.83 + ], + "angle": 0, + "content": "only hands is not sufficient to achieve satisfactory performance on the GRAB dataset. However, introducing more contextual information helps refine these results, unavailing information on other body parts." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Points saliency. As further evidence that object interaction involves different body parts, we conducted a study to discover what input points are crucial for the network." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "4732" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.246, + 0.089, + 0.725, + 0.161 + ], + "angle": 0, + "content": "
MethodGRABBEHAVE
Ec↓Ech↓Acc.↑Ec↓Ech↓Acc.↑
NN, SMPLH0.04040.193815.420.08800.387321.48
Ours, SMPLH0.02900.138720.800.07280.304624.16
Ours, SMPLH + T0.02630.138430.150.07220.306949.28
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.171, + 0.895, + 0.202 + ], + "angle": 0, + "content": "Table 3. Object classification. Quantitative results of models with object class prediction. Introducing temporal information has a dramatic impact on object classification." + }, + { + "type": "image", + "bbox": [ + 0.124, + 0.214, + 0.413, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.423, + 0.471, + 0.48 + ], + "angle": 0, + "content": "Figure 9. Confusion Matrix. The confusion matrix for a subset of the predicted classes. Objects with similar functionality are often confused (Coffemug vs Cup), while ones that suggest characteristic human poses (Dorknoob, Hammer) are well separated." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.488, + 0.47, + 0.7 + ], + "angle": 0, + "content": "We follow a recent protocol to find 3D point cloud saliency [72]: we cast the input through the network, we compute a loss on the output (in our case, the one in Eq. (2)), and we modify the input using the backpropagated gradient. This procedure is performed iteratively, and we refer to [72] for the details. We report the results of this study in Fig. 8, where points modified by the procedure are highlighted in red. We find the results of this analysis fascinating. As expected, the contact region is always essential to infer the correct object location. However, feet play a crucial role in all the reported cases, since they provide information about the human position and the consequent pose of the body. Another highlighted region is the head: different orientations give clues about object location and body posture." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Confusion in classification. As a final analysis, we explore the learning object classification during the training. We jointly train a further MLP module that takes as an input \\(\\mathbf{F}_o\\) and predicts the object class, using a cross-entropy loss. When a time sequence of point clouds is available, we exploit it by selecting the class with the highest score across the frames and applying it to the whole sequence. We report results in Tab. 3. Our experiment suggests that the task is challenging while, given the number of classes (40), we still consider our results a promising first step. Also, we notice that temporal smoothing significantly helps classification accuracy. Temporal context disambiguate poses" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.227, + 0.892, + 0.288 + ], + "angle": 0, + "content": "without a clear functionality. In Fig. 9, we report the confusion matrix for a subset of the classes for our method. The misclassification mainly arises from interactions of objects with similar functionality." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.307, + 0.627, + 0.323 + ], + "angle": 0, + "content": "5. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.334, + 0.892, + 0.501 + ], + "angle": 0, + "content": "In this work, we have addressed a novel and inspiring problem that changes the perspective on object-human interaction. Our proposed model is simple, carefully designed, and inspired by behavioural studies. We collected evidence of the method's effectiveness on a large set of object classes and empirically proved its generalization on noisy and different inputs. Finally, our analysis of human affordance is unprecedented, showing that human-object interaction can also involve body parts distant from the object and pointing to interesting relations useful for applications and subsequent works." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.53, + 0.892, + 0.711 + ], + "angle": 0, + "content": "Limitations and Future Works. As the first exploration in this direction, our study enables several future possibilities. In this work, the temporal information is only used after the training procedure. Incorporating this information can create other patterns, further improving the results' quality. Our empirical evidence suggests that class prediction requires further investigation and more sophisticated techniques, like specialized attention mechanisms. Finally, we do not consider sequences that involve long (e.g., hours-long) and complex (e.g., multi-objects) interactions, which are difficult to capture. We hope our work can foster the community to collect such datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.736, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgements Special thanks to the RVH team and reviewers, their feedback helped improve the manuscript. This work is funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) - 409792180 (EmmyNoether Programme, project: Real Virtual Humans) and the German Federal Ministry of Education and Research (BMBF): Tübingen AI Center, FKZ: 01IS18039A. G. Pons-Moll is a member of the Machine Learning Cluster of Excellence, EXC number 2064/1 - Project number 390727645. The authors thank the International Max Planck Research School for Intelligent Systems (IMPRS-IS) for supporting I. Petrov. R. Marin is supported by an Alexander von Humboldt Foundation Research Fellowship." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4733" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.198 + ], + "angle": 0, + "content": "[1] Bharat Lal Bhatnagar, Xianghui Xie, Ilya A. Petrov, Cristian Sminchisescu, Christian Theobalt, and Gerard Pons-Moll. BEHAV: Dataset and Method for Tracking Human Object Interactions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15935-15946, 2022. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.2, + 0.471, + 0.241 + ], + "angle": 0, + "content": "[2] Amy E Booth and Sandra Waxman. Object names and object functions serve as cues to categories for infants. Developmental psychology, 38(6):948, 2002. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.243, + 0.471, + 0.311 + ], + "angle": 0, + "content": "[3] Samarth Brahmbhatt, Cusuh Ham, Charles C Kemp, and James Hays. Contactdb: Analyzing and predicting grasp contact via thermal imaging. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8709-8719, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.313, + 0.471, + 0.38 + ], + "angle": 0, + "content": "[4] Samarth Brahmbhatt, Ankur Handa, James Hays, and Dieter Fox. Contactgrasp: Functional multi-finger grasp synthesis from contact. In 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 2386-2393. IEEE, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.382, + 0.471, + 0.451 + ], + "angle": 0, + "content": "[5] Samarth Brahmbhatt, Chengcheng Tang, Christopher D Twigg, Charles C Kemp, and James Hays. Contactpose: A dataset of grasps with object contact and hand pose. In European Conference on Computer Vision, pages 361-378. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.453, + 0.471, + 0.494 + ], + "angle": 0, + "content": "[6] Tim Brooks and Alexei A Efros. Hallucinating pose-compatible scenes. In European Conference on Computer Vision, pages 510-528. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.496, + 0.471, + 0.55 + ], + "angle": 0, + "content": "[7] Zhe Cao, Hang Gao, Karttikeya Mangalam, Qi-Zhi Cai, Minh Vo, and Jitendra Malik. Long-term human motion prediction with scene context. In European Conference on Computer Vision, pages 387-404. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.552, + 0.471, + 0.607 + ], + "angle": 0, + "content": "[8] Zhe Cao, Ilija Radosavovic, Angjoo Kanazawa, and Jitendra Malik. Reconstructing Hand-Object Interactions in the Wild. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12417–12426, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.608, + 0.471, + 0.661 + ], + "angle": 0, + "content": "[9] Laura A Carlson-Radvansky, Eric S Covey, and Kathleen M Lattanzi. “what” effects on “where”: Functional influences on spatial relations. *Psychological Science*, 10(6):516–521, 1999. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.664, + 0.47, + 0.747 + ], + "angle": 0, + "content": "[10] Yu-Wei Chao, Wei Yang, Yu Xiang, Pavlo Molchanov, Ankur Handa, Jonathan Tremblay, Yashraj S Narang, Karl Van Wyk, Umar Iqbal, Stan Birchfield, et al. Dexycb: A benchmark for capturing hand grasping of objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9044–9053, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.748, + 0.47, + 0.815 + ], + "angle": 0, + "content": "[11] R. Qi Charles, Hao Su, Mo Kaichun, and Leonidas J. Guibas. PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 77-85, July 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[12] Yixin Chen, Siyuan Huang, Tao Yuan, Siyuan Qi, Yixin Zhu, and Song-Chun Zhu. Holistic++ scene understanding: Single-view 3d holistic scene parsing and human pose estimation with human-object interaction and physical commonsense. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8648-8657, 2019. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[13] Sammy Christen, Muhammed Kocabas, Emre Aksan, Jemin Hwangbo, Jie Song, and Otmar Hilliges. D-Grasp: Physically Plausible Dynamic Grasp Synthesis for Hand-Object Interactions. arXiv:2112.03028 [cs], Mar. 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[14] Enric Corona, Gerard Pons-Moll, Guillem Alenyà, and Francesc Moreno-Noguer. Learned vertex descent: A new direction for 3d human model fitting. arXiv preprint arXiv:2205.06254, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.892, + 0.276 + ], + "angle": 0, + "content": "[15] Enric Corona, Albert Pumarola, Guillem Alenya, Francesc Moreno-Noguer, and Grégory Rogez. Ganhand: Predicting human grasp affordances in multi-object scenes. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5031-5041, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.278, + 0.892, + 0.346 + ], + "angle": 0, + "content": "[16] Rishabh Dabral, Soshi Shimada, Arjun Jain, Christian Theobalt, and Vladislav Golyanik. Gravity-Aware Monocular 3D Human-Object Reconstruction. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 12345–12354, Oct. 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.349, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[17] Shengheng Deng, Xun Xu, Chaozheng Wu, Ke Chen, and Kui Jia. 3d affordancenet: A benchmark for visual object affordance understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1778-1787, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.42, + 0.892, + 0.488 + ], + "angle": 0, + "content": "[18] Kiana Ehsani, Shubham Tulsiani, Saurabh Gupta, Ali Farhadi, and Abhinav Gupta. Use the force, luke! learning to predict physical forces by simulating effects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 224-233, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.49, + 0.892, + 0.545 + ], + "angle": 0, + "content": "[19] Eleanor J Gibson. The concept of affordances in development: The renascence of functionalism. In *The concept of development: The Minnesota symposia on child psychology*. Vol, volume 15, pages 55–81, 1982. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.548, + 0.892, + 0.574 + ], + "angle": 0, + "content": "[20] John C Gower. Generalized procrustes analysis. Psychometrika, 40(1):33-51, 1975. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.577, + 0.892, + 0.617 + ], + "angle": 0, + "content": "[21] Helmu t Grabner, Juergen Gall, and Luc Van Gool. What makes a chair a chair? In CVPR 2011, pages 1529-1536, Colorado Springs, CO, USA, June 2011. IEEE. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.62, + 0.892, + 0.688 + ], + "angle": 0, + "content": "[22] Patrick Grady, Chengcheng Tang, Christopher D. Twigg, Minh Vo, Samarth Brahmbhatt, and Charles C. Kemp. ContactOpt: Optimizing Contact to Improve Grasps. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1471-1481, June 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.691, + 0.892, + 0.773 + ], + "angle": 0, + "content": "[23] Vladimir Guzov, Aymen Mir, Torsten Sattler, and Gerard Pons-Moll. Human poseitioning system (hps): 3d human pose estimation and self-localization in large scenes from body-mounted sensors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4318-4329, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.776, + 0.892, + 0.815 + ], + "angle": 0, + "content": "[24] Vladimir Guzov, Torsten Sattler, and Gerard Pons-Moll. Visually plausible human-object interaction capture from wearable sensors. In arXiv, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.818, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[25] Shreyas Hampali, Sayan Deb Sarkar, Mahdi Rad, and Vincent Lepetit. Keypoint transformer: Solving joint identification in challenging hands and object interactions for accurate 3d pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11090-11100, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4734" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.134 + ], + "angle": 0, + "content": "[26] Gregor Harih and Bojan Dolsak. Tool-handle design based on a digital human hand model. International Journal of Industrial Ergonomics, 43(4):288-295, July 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.136, + 0.471, + 0.207 + ], + "angle": 0, + "content": "[27] Mohamed Hassan, Duygu Ceylan, Ruben Villegas, Jun Saito, Jimei Yang, Yi Zhou, and Michael J Black. Stochastic scene-aware motion prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11374-11384, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.208, + 0.471, + 0.275 + ], + "angle": 0, + "content": "[28] Mohamed Hassan, Vasileios Choutas, Dimitrios Tzionas, and Michael J Black. Resolving 3d human pose ambiguities with 3d scene constraints. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2282-2292, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.278, + 0.471, + 0.333 + ], + "angle": 0, + "content": "[29] Yana Hasson, Gül Varol, Cordelia Schmid, and Ivan Laptev. Towards unconstrained joint hand-object reconstruction from rgb videos. In 2021 International Conference on 3D Vision (3DV), pages 659-668. IEEE, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.335, + 0.471, + 0.417 + ], + "angle": 0, + "content": "[30] Yana Hasson, Gul Varol, Dimitrios Tzionas, Igor Kalevatykh, Michael J Black, Ivan Laptev, and Cordelia Schmid. Learning joint reconstruction of hands and manipulated objects. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11807-11816, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.42, + 0.471, + 0.503 + ], + "angle": 0, + "content": "[31] Chun-Hao P. Huang, Hongwei Yi, Markus Höschle, Matvey Safroshkin, Tsvetelina Alexiadis, Senya Polikovsky, Daniel Scharstein, and Michael J. Black. Capturing and inferring dense full-body human-scene contact. In IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), pages 13274–13285, June 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.505, + 0.471, + 0.572 + ], + "angle": 0, + "content": "[32] Yinghao Huang, Omid Taheri, Michael J Black, and Dimitrios Tzionas. Intercap: Joint markerless 3d tracking of humans and objects in interaction. In DAGM German Conference on Pattern Recognition, pages 281-299. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.576, + 0.471, + 0.658 + ], + "angle": 0, + "content": "[33] Yuheng Jiang, Suyi Jiang, Guoxing Sun, Zhuo Su, Kaiwen Guo, Minye Wu, Jingyi Yu, and Lan Xu. Neuralhofusion: Neural volumetric rendering under human-object interactions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6155-6165, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.661, + 0.471, + 0.73 + ], + "angle": 0, + "content": "[34] Korrawe Karunratanakul, Jinlong Yang, Yan Zhang, Michael J. Black, Krikamol Muandet, and Siyu Tang. Grasping Field: Learning Implicit Representations for Human Grasps. In 2020 International Conference on 3D Vision (3DV), pages 333-344, Nov. 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.732, + 0.471, + 0.787 + ], + "angle": 0, + "content": "[35] Hedvig Kjellström, Javier Romero, and Danica Kragić. Visual object-action recognition: Inferring object affordances from human demonstration. Computer Vision and Image Understanding, 115(1):81-90, Jan. 2011. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.471, + 0.857 + ], + "angle": 0, + "content": "[36] Taein Kwon, Bugra Tekin, Jan Stuhmer, Federica Bogo, and Marc Pollefeys. H2o: Two hands manipulating objects for first person interaction recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10138-10148, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.471, + 0.9 + ], + "angle": 0, + "content": "[37] Steven M Lehar. The world in your head: A gestalt view of the mechanism of conscious experience. Psychology Press, 2003. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[38] Zhi Li, Soshi Shimada, Bernt Schiele, Christian Theobalt, and Vladislav Golyanik. Mocapdeform: Monocular 3d human motion capture in deformable scenes. In International Conference on 3D Vision (3DV), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[39] Yunze Liu, Yun Liu, Che Jiang, Kangbo Lyu, Weikang Wan, Hao Shen, Boqiang Liang, Zhoujie Fu, He Wang, and Li Yi. Hoi4d: A 4d egocentric dataset for category-level human-object interaction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21013-21022, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.234, + 0.892, + 0.289 + ], + "angle": 0, + "content": "[40] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multiperson linear model. ACM transactions on graphics (TOG), 34(6):1-16, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.292, + 0.892, + 0.347 + ], + "angle": 0, + "content": "[41] Priyanka Mandikal and Kristen Grauman. Learning dexterous grasping with object-centric visual affordances. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 6169-6176. IEEE, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.349, + 0.892, + 0.404 + ], + "angle": 0, + "content": "[42] Yinyu Nie, Angela Dai, Xiaoguang Han, and Matthias Nießner. Pose2room: understanding 3d scenes from human activities. In European Conference on Computer Vision, pages 425-443. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.406, + 0.892, + 0.46 + ], + "angle": 0, + "content": "[43] Lisa M Oakes and Kelly L Madole. Function revisited: How infants construe functional features in their representation of objects. In Advances in child development and behavior, volume 36, pages 135-185. Elsevier, 2008. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.463, + 0.892, + 0.545 + ], + "angle": 0, + "content": "[44] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3d hands, face, and body from a single image. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10975-10985, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.548, + 0.892, + 0.615 + ], + "angle": 0, + "content": "[45] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. PointNet++: Deep Hierarchical Feature Learning on Point Sets in a Metric Space. In Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc., 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.618, + 0.892, + 0.659 + ], + "angle": 0, + "content": "[46] Daniel Roetenberg, Henk Luinge, and Per Slycke. Moven: Full 6 dof human motion tracking using miniature inertial sensors. Xsen Technologies, December, 2007. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.661, + 0.892, + 0.716 + ], + "angle": 0, + "content": "[47] Javier Romero, Dimitrios Tzionas, and Michael J. Black. Embodied hands: Modeling and capturing hands and bodies together. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 36(6), Nov. 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.719, + 0.892, + 0.773 + ], + "angle": 0, + "content": "[48] Manolis Savva, Angel X Chang, Pat Hanrahan, Matthew Fisher, and Matthias Nießner. Pigraphs: learning interaction snapshots from observations. ACM Transactions on Graphics (TOG), 35(4):1-12, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.776, + 0.892, + 0.829 + ], + "angle": 0, + "content": "[49] Colin G. Scanes. Chapter 1 - animal perception including differences with humans. In Colin G. Scanes and Samia R. Toukhsati, editors, Animals and Human Society, pages 1-11. Academic Press, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.832, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[50] Zhuo Su, Lan Xu, Dawei Zhong, Zhong Li, Fan Deng, Shuxue Quan, and Lu Fang. Robustfusion: Robust volumetric performance reconstruction under human-object interactions from monocular rgbd stream. arXiv preprint arXiv:2104.14837, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4735" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.174 + ], + "angle": 0, + "content": "[51] Guoxing Sun, Xin Chen, Yizhang Chen, Anqi Pang, Pei Lin, Yuheng Jiang, Lan Xu, Jingyi Yu, and Jingya Wang. Neural free-viewpoint performance rendering under complex human-object interactions. In Proceedings of the 29th ACM International Conference on Multimedia, pages 4651-4660, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.178, + 0.47, + 0.233 + ], + "angle": 0, + "content": "[52] Zhongda Sun, Minglu Zhu, Xuechuan Shan, and Chengkuo Lee. Augmented tactile-perception and haptic-feedback rings as human-machine interfaces aiming for immersive interactions. Nature communications, 13(1):1-13, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.234, + 0.47, + 0.29 + ], + "angle": 0, + "content": "[53] Subramanian Sundaram, Petr Kellnhofer, Yunzhu Li, Jun-Yan Zhu, Antonio Torralba, and Wojciech Matusik. Learning the signatures of the human grasp using a scalable tactile glove. Nature, 569(7758):698-702, May 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.291, + 0.47, + 0.359 + ], + "angle": 0, + "content": "[54] Omid Taheri, Vasileios Choutas, Michael J. Black, and Dimitrios Tzionas. GOAL: Generating 4D whole-body motion for hand-object grasping. In IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), pages 13263-13273, June 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.362, + 0.47, + 0.445 + ], + "angle": 0, + "content": "[55] Omid Taheri, Nima Ghorbani, Michael J. Black, and Dimitrios Tzionas. GRAB: A Dataset of Whole-Body Human Grasping of Objects. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, Computer Vision - ECCV 2020, volume 12349, pages 581-600. Springer International Publishing, Cham, 2020. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.447, + 0.47, + 0.516 + ], + "angle": 0, + "content": "[56] Weilin Wan, Lei Yang, Lingjie Liu, Zhuoying Zhang, Ruixing Jia, Yi-King Choi, Jia Pan, Christian Theobalt, Taku Komura, and Wenping Wang. Learn to predict how humans manipulate large-sized objects from interactive motions. IEEE Robotics and Automation Letters, 7(2):4702-4709, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.518, + 0.47, + 0.573 + ], + "angle": 0, + "content": "[57] Zhenzhen Weng and Serena Yeung. Holistic 3d human and scene mesh estimation from single view images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 334-343, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.575, + 0.47, + 0.63 + ], + "angle": 0, + "content": "[58] Yan Wu, Jiahao Wang, Yan Zhang, Siwei Zhang, Otmar Hilliges, Fisher Yu, and Siyu Tang. Saga: Stochastic whole-body grasping with contact. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.632, + 0.469, + 0.687 + ], + "angle": 0, + "content": "[59] Xianghui Xie, Bharat Lal Bhatnagar, and Gerard Pons-Moll. Chore: Contact, human and object reconstruction from a single rgb image. In European Conference on Computer Vision (ECCV). Springer, October 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.689, + 0.47, + 0.743 + ], + "angle": 0, + "content": "[60] Xianghui Xie, Bharat Lal Bhatnagar, and Gerard Pons-Moll. Visibility aware human-object interaction tracking from single rgb camera. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.746, + 0.47, + 0.814 + ], + "angle": 0, + "content": "[61] Lixin Yang, Xinyu Zhan, Kailin Li, Wenqiang Xu, Jiefeng Li, and Cewu Lu. Cpf: Learning a contact potential field to model the hand-object interaction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11097-11106, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.47, + 0.87 + ], + "angle": 0, + "content": "[62] Bangpeng Yao, Jiayuan Ma, and Li Fei-Fei. Discovering Object Functionality. In Proceedings of the IEEE International Conference on Computer Vision, pages 2512-2519, 2013. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.902 + ], + "angle": 0, + "content": "[63] Yufei Ye, Abhinav Gupta, and Shubham Tulsiani. What's in your hands? 3d reconstruction of generic objects in hands." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.133 + ], + "angle": 0, + "content": "In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3895-3905, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.136, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[64] Hongwei Yi, Chun-Hao P. Huang, Dimitrios Tzionas, Muhammed Kocabas, Mohamed Hassan, Siyu Tang, Justus Thies, and Michael J. Black. Human-Aware Object Placement for Visual Environment Reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3959-3970, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.221, + 0.892, + 0.302 + ], + "angle": 0, + "content": "[65] Andrei Zanfir, Elisabella Maroiniu, and Cristian Sminchisescu. Monocular 3d pose and shape estimation of multiple people in natural scenes-the importance of multiple scene constraints. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2148-2157, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.304, + 0.892, + 0.359 + ], + "angle": 0, + "content": "[66] Jason Y Zhang, Panna Felsen, Angjoo Kanazawa, and Jitendra Malik. Predicting 3d human dynamics from video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7114-7123, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.361, + 0.892, + 0.43 + ], + "angle": 0, + "content": "[67] Jason Y Zhang, Sam Pepose, Hanbyul Joo, Deva Ramanan, Jitendra Malik, and Angjoo Kanazawa. Perceiving 3d human-object spatial arrangements from a single image in the wild. In European conference on computer vision, pages 34-51. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.432, + 0.892, + 0.498 + ], + "angle": 0, + "content": "[68] Siwei Zhang, Qianli Ma, Yan Zhang, Zhiyin Qian, Marc Pollefeys, Federica Bogo, and Siyu Tang. Egobody: Human body shape, motion and social interactions from head-mounted devices. arXiv preprint arXiv:2112.07642, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.502, + 0.892, + 0.57 + ], + "angle": 0, + "content": "[69] Siwei Zhang, Yan Zhang, Federica Bogo, Marc Pollefeys, and Siyu Tang. Learning motion priors for 4d human body capture in 3d scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11343-11353, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.572, + 0.892, + 0.639 + ], + "angle": 0, + "content": "[70] Xiaohan Zhang, Bharat Lal Bhatnagar, Sebastian Starke, Vladimir Guzov, and Gerard Pons-Moll. Couch: Towards controllable human-chair interactions. In European Conference on Computer Vision (ECCV). Springer, October 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.643, + 0.892, + 0.698 + ], + "angle": 0, + "content": "[71] Qian Zheng, Weikai Wu, Hanting Pan, Niloy Mitra, Daniel Cohen-Or, and Hui Huang. Inferring object properties from human interaction and transferring them to new motions. Computational Visual Media, 7(3):375–392, Sept. 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.7, + 0.892, + 0.755 + ], + "angle": 0, + "content": "[72] Tianhang Zheng, Changyou Chen, Junsong Yuan, Bo Li, and Kui Ren. Pointcloud saliency maps. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1598-1606, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.757, + 0.892, + 0.797 + ], + "angle": 0, + "content": "[73] Keyang Zhou, Bharat Lal Bhatnagar, Jan Eric Lenssen, and Gerard Pons-Moll. TOCH: Spatio-Temporal Object-to-Hand Correspondence for Motion Refinement, July 2022. 1, 3" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.797 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4736" + } + ] +] \ No newline at end of file diff --git a/2023/Object Pop-Up_ Can We Infer 3D Objects and Their Poses From Human Interactions Alone_/965b59c8-d0c6-413b-852b-108c8ff0bbe7_origin.pdf b/2023/Object Pop-Up_ Can We Infer 3D Objects and Their Poses From Human Interactions Alone_/965b59c8-d0c6-413b-852b-108c8ff0bbe7_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..1d0170ea164f6944bb36029a7d28f4d6b3a10bed --- /dev/null +++ b/2023/Object Pop-Up_ Can We Infer 3D Objects and Their Poses From Human Interactions Alone_/965b59c8-d0c6-413b-852b-108c8ff0bbe7_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d2ce6e753040b0fb2d2286d2a863dedcf3a5e25f6ad8472479196e4039fa4599 +size 6698861 diff --git a/2023/Object Pop-Up_ Can We Infer 3D Objects and Their Poses From Human Interactions Alone_/full.md b/2023/Object Pop-Up_ Can We Infer 3D Objects and Their Poses From Human Interactions Alone_/full.md new file mode 100644 index 0000000000000000000000000000000000000000..26ff85858ac7f7e0ed6dd2bc66dc004be22cd245 --- /dev/null +++ b/2023/Object Pop-Up_ Can We Infer 3D Objects and Their Poses From Human Interactions Alone_/full.md @@ -0,0 +1,355 @@ +# Object pop-up: Can we infer 3D objects and their poses from human interactions alone? + +Ilya A. Petrov $^{1}$ , Riccardo Marin $^{1}$ , Julian Chibane $^{2,1}$ , and Gerard Pons-Moll $^{1,2,3}$ + +1University of Tübingen, Germany + +$^{2}$ Max Planck Institute for Informatics, Saarland Informatics Campus, Germany + +$^{3}$ Tübingen AI Center, Germany + +{i.petrov, gerard.pons-moll}@uni-tuebingen.de, riccardo.marin@mnf.uni-tuebingen.de, jchibane@mpi-inf.mpg.de + +![](images/ff13a4932e654648edde69665a0a1ece8f97f4f484679df15926cc86bf559f08.jpg) +Ref. Image (not used) +Figure 1. Human-Centric prediction of Human-Object Interaction. We explore Human-Object interaction when only the cues of the human pose interacting with an unobserved object are available ('Input'). We propose a neural model that, for the first time, can infer the location of the object ('Result') from such input. This is possible even when our subject simulates the interactions with the object ('Reference Image'). + +![](images/9474c6239a1356b0689de49886b91f9d190d57f1d5b798e781cb5e9cf9f13e72.jpg) +Point Cloud + +![](images/3e19da7a40c2e74032a938a644e3074e03779b64a7d0f74b90a186e261b867d1.jpg) +Input +Result + +![](images/2c894457d371750b0e0f25f1c6cc6cf2acbb5299eacb2470cecdfe7c2e6fb90e.jpg) +Ref. Image (not used) + +![](images/687583a36d7ec1f7205d5617528687ea5e78f37b287b39c9564d1f9ebc48adef.jpg) +Input Point Cloud + +![](images/c37c453829a71c4bbbae8edf24650c9088648d2070ce1f4c0ae5c96f174f4732.jpg) +Result + +![](images/2f165549bdb8c9d5784a7e54f9ee99befc57a75d6c225308e185dfe332560873.jpg) +Ref. Image (not used) + +![](images/f1a5f590815891caba96603958c61bec4c0c439fbbb66442e690b26e42a3faf5.jpg) +Input Point Cloud + +![](images/b5fafb11d446414b848dee849492adba0c04e5691f29e584de0588b62a338c88.jpg) +Result + +![](images/357160551435c931f39da0ac68f573519ff364f3a5eb11d57676e5a434aa0325.jpg) +Ref. Image (not used) + +![](images/a705fbebadecf78663cb2592ebd8c3817e4b807d45e166317e7d15bb6e9d5ddc.jpg) +Input Point Cloud + +![](images/3b6324fa122e1b48406fb033e003f6a071802a025becb7d5d6eeb8f72be88555.jpg) +Result + +# Abstract + +The intimate entanglement between objects affordances and human poses is of large interest, among others, for behavioural sciences, cognitive psychology, and Computer Vision communities. In recent years, the latter has developed several object-centric approaches: starting from items, learning pipelines synthesizing human poses and dynamics in a realistic way, satisfying both geometrical and functional expectations. However, the inverse perspective is significantly less explored: Can we infer 3D objects and their poses from human interactions alone? Our investigation follows this direction, showing that a generic 3D human point cloud is enough to pop up an unobserved object, even when the user is just imitating a functionality (e.g., looking through a binocular) without involving a tangible counterpart. We validate our method qualitatively and quantitatively, with synthetic data and sequences acquired for the task, showing applicability for XR/VR. + +# 1. Introduction + +Complex interactions with the world are among the unique skills distinguishing humans from other living beings. Even though our perception might be imperfect (we cannot hear ultrasonic sounds or see ultraviolet light [49]), our cognitive representation is enriched with a functional perspective, i.e., potential ways of interacting with objects or, as introduced by Gibson and colleagues, the affordance of the objects [19]. Several behavioural studies confirmed the centrality of this concept [2, 9, 43], which plays a fundamental role also for kids' development [2]. Computer Vision is well-aware that the function of an object complements its appearance [21], and exploited this in tasks like human and object reconstruction [62]. Previous literature approaches the interaction analysis from an object perspective (i.e., given an object, analyze the human interaction) [8,22,63], building object-centric priors [73], generating realistic grasps given the object [13, 34], reconstructing hand-object interactions [8,22,63]. Namely, objects induce functionality, so a human interaction (e.g., a mug suggests + +![](images/075157ef91344df3f099866a5b5b90e89f72a1e3ef09ee021a731fc3b175eb17.jpg) +Figure 2. Human-Centric vs. Object-Centric inference. A common perspective in the prior work is to infer affordances or human pose starting from an object. We explore the inverse, Human-Centric perspective of the human-object interaction relationship: given the human, we are interested in predicting the object. + +a drinking action; an handle a grasping one). For the first time, our work reverts the perspective, suggesting that analyzing human motion and behaviour is naturally a human-centric problem (i.e., given a human interaction, what kind of functionality is it suggesting, Fig. 2). Moving the first step in this new research direction, we pose a fundamental question: Can we infer 3D objects and their poses from human interactions alone? + +At first sight, the problem seems particularly hard and significantly under-constrained since several geometries might fit the same action. However, the human body complements this information in several ways: physical relations, characteristic poses, or body dynamics serve as valuable proxies for the involved functionality, as suggested by Fig. 1. Such hints are so powerful that we can easily imagine the kind of object and its location, even if such an object does not exist. Furthermore, even if the given pose might fit several possible solutions, our mind naturally comes to the most natural one indicated by the observed behaviour. It also suggests that solely focusing on the contact region (an approach often preferred by previous works) is insufficient in this new viewpoint. The reification principle of Gestalt psychology [37] highlights that "the whole" arises from "the parts" and their relation. Similarly, in Fig. 3, the hand grasps in B) pop up a binocular in our mind because we naturally consider the relationship with the other body parts. + +Finally, moving to a human-centric perspective in human-object interaction is critical for human studies and daily-life applications. Modern systems for AR/VR [24], and digital interaction [26] are both centered on humans, often manipulating objects that do not have a real-world counterpart. Learning to decode an object from human behaviour enables unprecedented applications. + +To answer our question, we deploy a first straightforward + +![](images/7b03322c0e42245b8afa641d5659f3be172a5958ea7935a2c65c6e9bd893563d.jpg) +A) + +![](images/a531002a2af6fdd8cb7e29cbf6fbc4abab31fc24793fb85281ef03ff919bcbe0.jpg) +B) +Figure 3. Reification principle. Gestalt theory suggests our perception considers the whole before the single parts (A). This served as inspiration for our work: the object can arise considering the body parts as a whole (B). + +and effective pipeline to "pop up" a rigid object from a 3D human point cloud. Starting from the input human point cloud and a class, we train an end-to-end pipeline to infer object location. In case a temporal sequence of point clouds is available, we suggest post-processing to avoid jittering and inconsistencies in the predictions, showing the relevance of this information to handle ambiguous poses. We show promising results on previously unaddressed tasks in digital and real-world scenes. Finally, our method allows us to analyze different features of human behaviour, highlight their contribution to object retrieval, and point to exciting directions for future works. + +In summary, our main contributions are: + +1. We formulate a novel problem, changing the perspective taken by previous works in the field, and open to a yet unexplored research direction; +2. We introduce a method capable of predicting the object starting from an input human point cloud; +3. We analyze different components of the human-object relationship: the contribution of different pieces of interactions (hands, body, time sequence), the point-wise saliency of input points, and the confusion produced by objects with similar functions. + +All the code will be made publicly available. + +# 2. Related Work + +# 2.1. Object Functionality + +In the core of human perception of objects, functionality complements physical appearance, enhancing our perception. Gibson [19] introduced the idea that humans use affordances of objects for perception. Affordance can be defined as: "an intrinsic property of an object, allowing an action to be performed with the object" [35]. + +From a Computer Vision perspective, object functionality supports several tasks such as scene analysis [21], object + +classification [17, 35], object properties inferring [71], and it is also possible to learn object-specific human interaction models from 2D images [62]. These works suggest an intimate entanglement between the action and the object itself. + +# 2.2. Human-Object interaction + +Modelling environment-aware humans and their interactions in 3D is one of the most recent challenges to creating virtual humans. We see two main lines of work there: hand-focused and full-body ones. + +Hand-Object Interaction. Several works tackle the problem of human interaction, focusing solely on the hands [10]. This has been done starting from 2D [8, 15, 18, 22, 25, 29, 30, 34, 61, 63], 2.5D [3, 5], and 3D data [4, 34, 55]. Particularly promising seems the application of well-designed priors for the motion [73]. The class of objects involved in these works are mainly limited to graspable ones. We argue that interactions involving different body parts are common in everyday life, more attractive from an applicative perspective, and more challenging. Moreover, full-body context is crucial in reconstructing even grasp interactions since body pose contains information on an object's properties, e.g. pounding with a hammer affects the whole posture to support the action. + +Fully-Body Interaction. On this line, several works focus on the interaction between a human and a scene [12,28, 31,38,57,64,65,69]. Also, in this case, priors can be used to regularize the motion [48]. Several datasets are also available to study the interaction between a human and a single object. For example, recent BEHAVE [1], GRAB [55], and InterCap [32] capture full-body interactions with diverse objects. Works address the task of humans interactions reconstruction from different kinds of data sources like single image [59, 67], video [16, 50, 60], and multiview capturing [1, 33, 51], and synthetization of them as well [7,27,41,54,56,58,66,70]. However, in all these works, we observe a general object-centric perspective: given a scene or an object, they aim to recreate the humans interacting with them. We argue that the significantly less explored complementary one has more concrete applications in daily life, especially in VR/XR contexts where the human is central to the system. + +# 2.3. Human-centric perspective + +While the general trend mainly focuses on the surrounding environment and objects, there is a growing interest and availability of egocentric tools for humans [23, 68] also interacting with objects [24,36,39]. They provide a subjective view and a valuable paradigm for several applications, like letting a user interact with objects in the digital world. Recent works also involve more sophisticated devices [52,53], + +while they are still far from applicability. In a similar direction to our work, others propose to recover objects arrangements in a room starting from the human motion [42], to hallucinate a coherent 2D image from a human pose [6], or predicting physical properties of the objects (e.g., the weight of a box) from human joints [71]. While the principle inspires us, our study significantly differs: we focus on object pose and its spatial relation with humans, starting solely from unordered point clouds. + +# 3. Method + +This section describes our setting and the main components of our methodology, both at inference and training time. An overview of our pipeline can be found in Fig. 4. + +# 3.1. Object Pop-Up + +Input. Our method starts from a single human point cloud $\mathbf{P} \in \mathbb{R}^{N_P \times 3}$ with $N_P$ points and a hot-encoded object class $c$ . The input point cloud can result from 3D/4D scans, IMUs template fitting, or any other shape-from-X approach. Regardless of the point cloud source, we remark that no further information is used apart from the 3D coordinates of the points. We represent each object as 1500 key points $\mathbf{K}_c$ uniformly sampled on the template mesh. + +Object Center. Training a model to predict an object pose from a human point cloud poses several challenges. Such a task requires the network to understand the location of different body parts and their subtle relations while jointly developing a sense of its spatial relationship with the human. Empirically, we observed that this is only feasible by carefully deconstructing the problem and designing different features to ease the learning process. As the first step to decompose this problem, we train a PointNet++ architecture [11,45] to predict the object center $\mathbf{o}_P$ starting from $\mathbf{P}$ . At training time, this is supervised with an L2 loss against the ground truth center $\widehat{\mathbf{o}}_P$ : + +$$ +L _ {\mathbf {o} _ {P}} (\mathbf {P}) = \left\| \mathbf {o} _ {P} - \widehat {\mathbf {o}} _ {P} \right\| _ {2} ^ {2}. \tag {1} +$$ + +Solving this task provides a good initialization for the object pose, and we move the key points associated with the input class $c$ to the predicted center $\mathbf{o}_P$ . Also, the features $\mathbf{F}_o \in \mathbb{R}^{512}$ extracted by the network encode import information on the whole human body. + +Local Neighbourhood. The center prediction module can be further exploited using the nearest human regions. Intuitively, considering the closest body parts is essential to infer a contact relationship, but also the influence of body parts not directly in touch with the object (e.g., head orientation while using binoculars). To learn these connections, we consider the centered key points together with + +![](images/13bc0c9288897809a7841d0b5a428d093b1e2e2973554020caa5b64564b8f270.jpg) +Figure 4. Object pop-up. Our method predicts the position of an object, starting only from an input point cloud and an object class. It relies on a careful problem decomposition in several sub-tasks, extracting features that involve the entire human body and relations between body parts near the object. + +the 3000 closest points of the input human point cloud: $\mathbf{P}_L = KNN(\mathbf{P},\mathbf{o}_P)$ . We pass these two sets as a unique point cloud to a PointNet++ network, obtaining a new set of per-point features $\mathbf{F}_{P_L}\in \mathbb{R}^{128}$ . + +Object displacement. To predict the object's final position, we empirically observed that directly predicting a rotation and a translation is not a good solution. Inspired by recent works that suggest a point-wise offset prediction to recover 3D human shapes [14], we apply a similar approach to our task. Our goal is to predict a point-wise shift $\mathbf{S}_K$ for the $\mathbf{K}_c$ vertices to align them to the target pose. We append the features $\mathbf{F}_o$ , $\mathbf{F}_{P_L}$ , the one-hot encoding of the object class $c$ , and a positional encoding to the centered key points, and we pass them to a decoder. At training time, we consider the following loss: + +$$ +L _ {o f f} \left(\mathbf {K} _ {c}, \mathbf {F} _ {o}, \mathbf {F} _ {P _ {L}}, c\right) = \left\| \mathbf {S} _ {K} - \widehat {\mathbf {S}} _ {K} \right\| _ {F} ^ {2}. \tag {2} +$$ + +The network is then trained end-to-end using: + +$$ +L = L _ {\mathbf {o} _ {P}} + \alpha L _ {\text {o f f}}. \tag {3} +$$ + +The weighting coefficient is $\alpha = 10$ + +# 3.2. Template fitting + +Procrustes alignment. The point-wise offset produced by the network potentially distorts the key points structure in a non-rigid way. To recover the desired global rigid transformation, we rely on a Procrustes alignment [20]. This procedure takes as input two point clouds and returns the rotation $\mathbf{R}$ and the translation $\mathbf{t}$ to minimize the L2 distances of the points: + +$$ +P (\cdot , \cdot) \rightarrow (\mathbf {R}, \mathbf {t}). \tag {4} +$$ + +We apply this to the template key points and their configuration obtained with our network: + +$$ +P \left(\mathbf {K} _ {c}, \mathbf {K} _ {c} + \mathbf {o} _ {P} + \mathbf {S} _ {K}\right)\rightarrow (\mathbf {R}, \mathbf {t}). \tag {5} +$$ + +Finally, we recover the desired object pose as: + +$$ +\mathbf {T} ^ {\prime} = \mathbf {R T} _ {c} + \mathbf {t} \tag {6} +$$ + +Time Smoothing. While our pipeline is designed to work with a single point cloud as input, considering the temporal evolution of interaction is often crucial, shaping the context of the individual poses. If a temporal sequence of point clouds is available, we provide a post-processing smoothing technique to take advantage of this further information. After running our method for each frame, we smooth the centering prediction across the sequence using a Gaussian kernel. Later, we will discuss a variation of our approach that also predicts the object class. In that case, we consider the most frequent class prediction over the whole set of frames to fix a class for the sequence. + +# 4. Experiments + +In this Section, we will describe the datasets used for training and testing our method. Then, we will present the baseline and the evaluation metrics. Finally, we will provide validation of our method as well as an analysis of its extended version that allows object class prediction. + +Datasets We jointly train on the union of BEHAVE [1] and GRAB [55], obtaining a set of: + +15 subjects, first 8 subjects from the GRAB dataset and 7 subjects from the official training part of BEHAV; +- 40 different classes of objects, including all 20 objects from the BEHAVE dataset and 20 selected objects from the GRAB dataset; + +We downsample training sequences of GRAB and BEHAVE to $10fps$ . To evaluate our method, we select subjects 9 and 10 from the GRAB dataset and downsample the sequences to $30fps$ . For the BEHAVE dataset, we use the official test part, which includes all sequences at $1fps$ with subject 3 and part of the sequences with subjects 4, 5. As an input, we use point clouds with 9000 points sampled uniformly over the SMPL-H [44] meshes. We refer to raw + +
MethodsGRABBEHAVEBEHAVE-Raw
Ec↓Ev2v↓Ec↓Ev2v↓Ec↓Ev2v↓
NN0.03620.14450.08020.3445--
Ours0.02370.09430.06630.29000.08060.3143
+ +Table 1. Comparison with the baseline. Our method significantly outperforms the baseline, even though the baseline uses the vertex order as additional information. Moreover, the baseline method does not generalize to point clouds with an arbitrary number of vertices. + +point clouds from the BEHAVE dataset used in our experiments as BEHAVE-Raw. We use point clouds that are fused from 4 Kinect sensors and subsample $90k$ points from them. + +Data augmentation. During training, to simulate errors in the center prediction, we randomly translate and rotate the object around the ground-truth center $\widehat{\mathbf{o}}_P$ . + +Implementation details. We implement our method using PyTorch framework and use Nvidia RTX3090 GPU for training and evaluation. The model is trained using Adam optimizer for 60 epochs with a learning rate of $1e^{-4}$ , which decays 10 times after 30-th and 40-th epochs. For the first 20 epochs of training, we use ground-truth object center $\widehat{\mathbf{o}}_P$ instead of predicted $\mathbf{o}_P$ to select local neighborhood $\mathbf{F}_{P_L}$ to warm up the local PointNet++ encoder. + +Nearest-Neighbor Baseline. Since we are the first to tackle this task and no competitors are available, we propose a simple while informative baseline. Given the input point cloud, we recover the most similar in the training dataset in an L2 sense. Then, we recover the object handled by that subject and pose it in space in the same way. This baseline demonstrates that the task is non-trivial and the generalization to unseen poses and subjects of our method. Also, this baseline requires that the target point cloud and the ones in the training set share the same number of points. Hence, if the input point cloud is a raw scan, this baseline is not applicable. Our method, instead, does not rely on this assumption and is more general. + +Object classification. In our research, we also investigate the possibility of incorporating class prediction inside the network training. This task is significantly difficult at a single-frame level since an isolated pose often does not suggest a clear functionality. However, including this step is interesting to analyze the interaction and the nature of the network confusion. Hence, we modify our method by adding a decoder module that takes the global features $\mathbf{F}_o$ and the local ones $\mathbf{F}_{P_L}$ as input to predict the object class. Then, we add to the training a simple cross-entropy loss between the predicted class and the ground truth one $\widehat{C}_P$ . + +# 4.1. Metrics + +In qualitative experiments, we use three main metrics to evaluate our results. In the tables, we report the average error across the considered test samples. + +Vertex-to-vertex. In most cases, our resulting object and the target one share the same number of vertices. Hence, we can compute the error between our prediction and the ground truth $\widehat{\mathbf{T}}'$ as a point-to-point error: + +$$ +E _ {v 2 v} = \left\| \mathbf {T} ^ {\prime} - \widehat {\mathbf {T}} ^ {\prime} \right\| _ {F}. \tag {7} +$$ + +When such error is computed only between the object centers, we will refer to it as $E_{c}$ . + +Chamfer distance. When we evaluate the network that also predicts the class, target objects and selected templates might not share the same number of vertices. In that case, as a metric we use bi-directional Chamfer distance: + +$$ +E _ {c h} = \frac {1}{\| \mathbf {T} ^ {\prime} \|} \sum_ {x \in \mathbf {T} ^ {\prime}} \min _ {y \in \widehat {\mathbf {T}} ^ {\prime}} \| x - y \| _ {2} + \frac {1}{\| \widehat {\mathbf {T}} ^ {\prime} \|} \sum_ {y \in \widehat {\mathbf {T}} ^ {\prime}} \min _ {x \in \mathbf {T} ^ {\prime}} \| y - x \| _ {2} \tag {8} +$$ + +Classification Accuracy. In case we use our network to predict the object class, we measure our misclassification error in terms of accuracy. + +# 4.2. Object pose Evaluation + +In Tab. 1, we report the quantitative evaluation on the test set of the datasets, comparing our method to the baseline. Our approach significantly outperforms the baseline, even if this latter exploits the points order information. We obtain the most significant margin on the GRAB dataset, where objects are small and mainly involve hands, showing the precision of our method. The baseline cannot be applied on BEHAVE-Raw since it does not share the same number of vertices as the training set, while our method shows only a limited performance decrease, pointing to generalization also to point clouds coming from different sources. We report qualitative results of our method on GRAB (first two rows) and BEHAVE (last row) in Fig. 5, and on BEHAVE-Raw in Fig. 6. Finally, to further evaluate the generalization of our method to unseen poses and subjects, we also considered point clouds obtained by an egocentric pipeline. + +![](images/ff078a73221b39123d11e20a8864b6816720a635944a6a5f07dcc1af96a469bd.jpg) +Figure 5. Qualitative results. Results of our method on GRAB (first and second rows) and BEHAVE (third row) datasets. + +In this scenario, we record a user motion with an XSens system [46], retarget the pose to an SMPL+H model, and obtain the point cloud by sampling the resulting mesh. Outputs of our method can be observed in Fig. 1 and Fig. 7. Generalizing to unseen poses and subjects acquired with wearable systems prone to measurement errors opens several exciting applications for VR/XR contexts. + +# 4.3. Human Affordance analysis + +We use our method to analyze human-object interaction considering three key factors: changes in the input information, the points saliency, and confusion in classification. + +Input information. Since we are curious to analyze how information is encoded in the inputs, we consider three more scenarios: + +- Hands: We generate the input data using the MANO [47] hand model annotations included in GRAB dataset. Hence, the model should infer the pose of the object without relying on other body parts. +- SMPL: We consider all the points of the subject, but the ones from the hands are in the rest pose provided by SMPL [40]. In this way, we analyze how much the network captures the information of the finger pose. + +
MethodsGRABBEHAVEBEHAVE-Raw
Ec ↓Ev2v ↓Ec ↓Ev2v ↓Ec ↓Ev2v ↓
Ours, hands0.30150.7358----
Ours, SMPL0.02450.10090.06670.28770.08430.3205
Ours, SMPLH0.02370.09430.06630.29000.08060.3143
Ours, SMPLH+T0.02350.09830.06860.28800.08230.3161
+ +Table 2. Human Affordance. We use our pipeline to explore how different inputs affect the object pose prediction. Using full body provides richer features for object recovery. + +![](images/e346cad04e13a3ab85f8e14b6239ebad852988c6653614dae40f3b10b6ebcc36.jpg) +Figure 6. BEHAVE-Raw. Results of our method considering as input raw point clouds from BEHAVE dataset. Despite the amount of noise and occlusions, our method is able to generalize and performs reliably. + +![](images/2ab01b3ce5b89b9a309e2a7421656945da492d5b268444a505f1d9a1eff979dc.jpg) + +![](images/c6677345b13a7e64ad2cf0e8caf3b72a03378454766f907d656201d0446aa01d.jpg) +Figure 7. Human from IMUs. Results of our method from a motion sequence acquired with IMUs. Even if the subject is unseen and the motion is subject to noise in the IMUs sensors, our method produces stable results. +Figure 8. Saliency. Point cloud saliency computed for different objects, rendered from two perspectives. The contact region is relevant for all the interactions, while the network also focuses on the feet and head regions. All the predictions are results of our method. + +- SMPLH+T: We consider all the points of the subject with the finger correctly posed (as in our method), and also we apply the temporal smoothing outlined in Sec. 3. Temporal information contextualizes the interactions and regularizes the predictions throughout the sequence. + +For the first two cases, we train ad-hoc networks. Results of this analysis are reported in Tab. 2. We notice that using + +only hands is not sufficient to achieve satisfactory performance on the GRAB dataset. However, introducing more contextual information helps refine these results, unavailing information on other body parts. + +Points saliency. As further evidence that object interaction involves different body parts, we conducted a study to discover what input points are crucial for the network. + +
MethodGRABBEHAVE
Ec↓Ech↓Acc.↑Ec↓Ech↓Acc.↑
NN, SMPLH0.04040.193815.420.08800.387321.48
Ours, SMPLH0.02900.138720.800.07280.304624.16
Ours, SMPLH + T0.02630.138430.150.07220.306949.28
+ +Table 3. Object classification. Quantitative results of models with object class prediction. Introducing temporal information has a dramatic impact on object classification. + +![](images/6a16398cfdb9213cc50694e80b3efab326b67572229f4118376c82855fbd64d8.jpg) +Figure 9. Confusion Matrix. The confusion matrix for a subset of the predicted classes. Objects with similar functionality are often confused (Coffemug vs Cup), while ones that suggest characteristic human poses (Dorknoob, Hammer) are well separated. + +We follow a recent protocol to find 3D point cloud saliency [72]: we cast the input through the network, we compute a loss on the output (in our case, the one in Eq. (2)), and we modify the input using the backpropagated gradient. This procedure is performed iteratively, and we refer to [72] for the details. We report the results of this study in Fig. 8, where points modified by the procedure are highlighted in red. We find the results of this analysis fascinating. As expected, the contact region is always essential to infer the correct object location. However, feet play a crucial role in all the reported cases, since they provide information about the human position and the consequent pose of the body. Another highlighted region is the head: different orientations give clues about object location and body posture. + +Confusion in classification. As a final analysis, we explore the learning object classification during the training. We jointly train a further MLP module that takes as an input $\mathbf{F}_o$ and predicts the object class, using a cross-entropy loss. When a time sequence of point clouds is available, we exploit it by selecting the class with the highest score across the frames and applying it to the whole sequence. We report results in Tab. 3. Our experiment suggests that the task is challenging while, given the number of classes (40), we still consider our results a promising first step. Also, we notice that temporal smoothing significantly helps classification accuracy. Temporal context disambiguate poses + +without a clear functionality. In Fig. 9, we report the confusion matrix for a subset of the classes for our method. The misclassification mainly arises from interactions of objects with similar functionality. + +# 5. Conclusions + +In this work, we have addressed a novel and inspiring problem that changes the perspective on object-human interaction. Our proposed model is simple, carefully designed, and inspired by behavioural studies. We collected evidence of the method's effectiveness on a large set of object classes and empirically proved its generalization on noisy and different inputs. Finally, our analysis of human affordance is unprecedented, showing that human-object interaction can also involve body parts distant from the object and pointing to interesting relations useful for applications and subsequent works. + +Limitations and Future Works. As the first exploration in this direction, our study enables several future possibilities. In this work, the temporal information is only used after the training procedure. Incorporating this information can create other patterns, further improving the results' quality. Our empirical evidence suggests that class prediction requires further investigation and more sophisticated techniques, like specialized attention mechanisms. Finally, we do not consider sequences that involve long (e.g., hours-long) and complex (e.g., multi-objects) interactions, which are difficult to capture. We hope our work can foster the community to collect such datasets. + +Acknowledgements Special thanks to the RVH team and reviewers, their feedback helped improve the manuscript. This work is funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) - 409792180 (EmmyNoether Programme, project: Real Virtual Humans) and the German Federal Ministry of Education and Research (BMBF): Tübingen AI Center, FKZ: 01IS18039A. G. Pons-Moll is a member of the Machine Learning Cluster of Excellence, EXC number 2064/1 - Project number 390727645. The authors thank the International Max Planck Research School for Intelligent Systems (IMPRS-IS) for supporting I. Petrov. R. Marin is supported by an Alexander von Humboldt Foundation Research Fellowship. + +# References + +[1] Bharat Lal Bhatnagar, Xianghui Xie, Ilya A. Petrov, Cristian Sminchisescu, Christian Theobalt, and Gerard Pons-Moll. BEHAV: Dataset and Method for Tracking Human Object Interactions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15935-15946, 2022. 3, 4 +[2] Amy E Booth and Sandra Waxman. Object names and object functions serve as cues to categories for infants. Developmental psychology, 38(6):948, 2002. 1 +[3] Samarth Brahmbhatt, Cusuh Ham, Charles C Kemp, and James Hays. Contactdb: Analyzing and predicting grasp contact via thermal imaging. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8709-8719, 2019. 3 +[4] Samarth Brahmbhatt, Ankur Handa, James Hays, and Dieter Fox. Contactgrasp: Functional multi-finger grasp synthesis from contact. In 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 2386-2393. IEEE, 2019. 3 +[5] Samarth Brahmbhatt, Chengcheng Tang, Christopher D Twigg, Charles C Kemp, and James Hays. Contactpose: A dataset of grasps with object contact and hand pose. In European Conference on Computer Vision, pages 361-378. Springer, 2020. 3 +[6] Tim Brooks and Alexei A Efros. Hallucinating pose-compatible scenes. In European Conference on Computer Vision, pages 510-528. Springer, 2022. 3 +[7] Zhe Cao, Hang Gao, Karttikeya Mangalam, Qi-Zhi Cai, Minh Vo, and Jitendra Malik. Long-term human motion prediction with scene context. In European Conference on Computer Vision, pages 387-404. Springer, 2020. 3 +[8] Zhe Cao, Ilija Radosavovic, Angjoo Kanazawa, and Jitendra Malik. Reconstructing Hand-Object Interactions in the Wild. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12417–12426, 2021. 1, 3 +[9] Laura A Carlson-Radvansky, Eric S Covey, and Kathleen M Lattanzi. “what” effects on “where”: Functional influences on spatial relations. *Psychological Science*, 10(6):516–521, 1999. 1 +[10] Yu-Wei Chao, Wei Yang, Yu Xiang, Pavlo Molchanov, Ankur Handa, Jonathan Tremblay, Yashraj S Narang, Karl Van Wyk, Umar Iqbal, Stan Birchfield, et al. Dexycb: A benchmark for capturing hand grasping of objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9044–9053, 2021. 3 +[11] R. Qi Charles, Hao Su, Mo Kaichun, and Leonidas J. Guibas. PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 77-85, July 2017. 3 +[12] Yixin Chen, Siyuan Huang, Tao Yuan, Siyuan Qi, Yixin Zhu, and Song-Chun Zhu. Holistic++ scene understanding: Single-view 3d holistic scene parsing and human pose estimation with human-object interaction and physical commonsense. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8648-8657, 2019. 3 + +[13] Sammy Christen, Muhammed Kocabas, Emre Aksan, Jemin Hwangbo, Jie Song, and Otmar Hilliges. D-Grasp: Physically Plausible Dynamic Grasp Synthesis for Hand-Object Interactions. arXiv:2112.03028 [cs], Mar. 2022. 1 +[14] Enric Corona, Gerard Pons-Moll, Guillem Alenyà, and Francesc Moreno-Noguer. Learned vertex descent: A new direction for 3d human model fitting. arXiv preprint arXiv:2205.06254, 2022. 4 +[15] Enric Corona, Albert Pumarola, Guillem Alenya, Francesc Moreno-Noguer, and Grégory Rogez. Ganhand: Predicting human grasp affordances in multi-object scenes. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5031-5041, 2020. 3 +[16] Rishabh Dabral, Soshi Shimada, Arjun Jain, Christian Theobalt, and Vladislav Golyanik. Gravity-Aware Monocular 3D Human-Object Reconstruction. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 12345–12354, Oct. 2021. 3 +[17] Shengheng Deng, Xun Xu, Chaozheng Wu, Ke Chen, and Kui Jia. 3d affordancenet: A benchmark for visual object affordance understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1778-1787, 2021. 3 +[18] Kiana Ehsani, Shubham Tulsiani, Saurabh Gupta, Ali Farhadi, and Abhinav Gupta. Use the force, luke! learning to predict physical forces by simulating effects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 224-233, 2020. 3 +[19] Eleanor J Gibson. The concept of affordances in development: The renascence of functionalism. In *The concept of development: The Minnesota symposia on child psychology*. Vol, volume 15, pages 55–81, 1982. 1, 2 +[20] John C Gower. Generalized procrustes analysis. Psychometrika, 40(1):33-51, 1975. 4 +[21] Helmu t Grabner, Juergen Gall, and Luc Van Gool. What makes a chair a chair? In CVPR 2011, pages 1529-1536, Colorado Springs, CO, USA, June 2011. IEEE. 1, 2 +[22] Patrick Grady, Chengcheng Tang, Christopher D. Twigg, Minh Vo, Samarth Brahmbhatt, and Charles C. Kemp. ContactOpt: Optimizing Contact to Improve Grasps. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1471-1481, June 2021. 1, 3 +[23] Vladimir Guzov, Aymen Mir, Torsten Sattler, and Gerard Pons-Moll. Human poseitioning system (hps): 3d human pose estimation and self-localization in large scenes from body-mounted sensors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4318-4329, 2021. 3 +[24] Vladimir Guzov, Torsten Sattler, and Gerard Pons-Moll. Visually plausible human-object interaction capture from wearable sensors. In arXiv, 2022. 2, 3 +[25] Shreyas Hampali, Sayan Deb Sarkar, Mahdi Rad, and Vincent Lepetit. Keypoint transformer: Solving joint identification in challenging hands and object interactions for accurate 3d pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11090-11100, 2022. 3 + +[26] Gregor Harih and Bojan Dolsak. Tool-handle design based on a digital human hand model. International Journal of Industrial Ergonomics, 43(4):288-295, July 2013. 2 +[27] Mohamed Hassan, Duygu Ceylan, Ruben Villegas, Jun Saito, Jimei Yang, Yi Zhou, and Michael J Black. Stochastic scene-aware motion prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11374-11384, 2021. 3 +[28] Mohamed Hassan, Vasileios Choutas, Dimitrios Tzionas, and Michael J Black. Resolving 3d human pose ambiguities with 3d scene constraints. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2282-2292, 2019. 3 +[29] Yana Hasson, Gül Varol, Cordelia Schmid, and Ivan Laptev. Towards unconstrained joint hand-object reconstruction from rgb videos. In 2021 International Conference on 3D Vision (3DV), pages 659-668. IEEE, 2021. 3 +[30] Yana Hasson, Gul Varol, Dimitrios Tzionas, Igor Kalevatykh, Michael J Black, Ivan Laptev, and Cordelia Schmid. Learning joint reconstruction of hands and manipulated objects. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11807-11816, 2019. 3 +[31] Chun-Hao P. Huang, Hongwei Yi, Markus Höschle, Matvey Safroshkin, Tsvetelina Alexiadis, Senya Polikovsky, Daniel Scharstein, and Michael J. Black. Capturing and inferring dense full-body human-scene contact. In IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), pages 13274–13285, June 2022. 3 +[32] Yinghao Huang, Omid Taheri, Michael J Black, and Dimitrios Tzionas. Intercap: Joint markerless 3d tracking of humans and objects in interaction. In DAGM German Conference on Pattern Recognition, pages 281-299. Springer, 2022. 3 +[33] Yuheng Jiang, Suyi Jiang, Guoxing Sun, Zhuo Su, Kaiwen Guo, Minye Wu, Jingyi Yu, and Lan Xu. Neuralhofusion: Neural volumetric rendering under human-object interactions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6155-6165, 2022. 3 +[34] Korrawe Karunratanakul, Jinlong Yang, Yan Zhang, Michael J. Black, Krikamol Muandet, and Siyu Tang. Grasping Field: Learning Implicit Representations for Human Grasps. In 2020 International Conference on 3D Vision (3DV), pages 333-344, Nov. 2020. 1, 3 +[35] Hedvig Kjellström, Javier Romero, and Danica Kragić. Visual object-action recognition: Inferring object affordances from human demonstration. Computer Vision and Image Understanding, 115(1):81-90, Jan. 2011. 2, 3 +[36] Taein Kwon, Bugra Tekin, Jan Stuhmer, Federica Bogo, and Marc Pollefeys. H2o: Two hands manipulating objects for first person interaction recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10138-10148, 2021. 3 +[37] Steven M Lehar. The world in your head: A gestalt view of the mechanism of conscious experience. Psychology Press, 2003. 2 + +[38] Zhi Li, Soshi Shimada, Bernt Schiele, Christian Theobalt, and Vladislav Golyanik. Mocapdeform: Monocular 3d human motion capture in deformable scenes. In International Conference on 3D Vision (3DV), 2022. 3 +[39] Yunze Liu, Yun Liu, Che Jiang, Kangbo Lyu, Weikang Wan, Hao Shen, Boqiang Liang, Zhoujie Fu, He Wang, and Li Yi. Hoi4d: A 4d egocentric dataset for category-level human-object interaction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21013-21022, 2022. 3 +[40] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multiperson linear model. ACM transactions on graphics (TOG), 34(6):1-16, 2015. 6 +[41] Priyanka Mandikal and Kristen Grauman. Learning dexterous grasping with object-centric visual affordances. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 6169-6176. IEEE, 2021. 3 +[42] Yinyu Nie, Angela Dai, Xiaoguang Han, and Matthias Nießner. Pose2room: understanding 3d scenes from human activities. In European Conference on Computer Vision, pages 425-443. Springer, 2022. 3 +[43] Lisa M Oakes and Kelly L Madole. Function revisited: How infants construe functional features in their representation of objects. In Advances in child development and behavior, volume 36, pages 135-185. Elsevier, 2008. 1 +[44] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3d hands, face, and body from a single image. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10975-10985, 2019. 4 +[45] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. PointNet++: Deep Hierarchical Feature Learning on Point Sets in a Metric Space. In Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc., 2017. 3 +[46] Daniel Roetenberg, Henk Luinge, and Per Slycke. Moven: Full 6 dof human motion tracking using miniature inertial sensors. Xsen Technologies, December, 2007. 6 +[47] Javier Romero, Dimitrios Tzionas, and Michael J. Black. Embodied hands: Modeling and capturing hands and bodies together. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 36(6), Nov. 2017. 6 +[48] Manolis Savva, Angel X Chang, Pat Hanrahan, Matthew Fisher, and Matthias Nießner. Pigraphs: learning interaction snapshots from observations. ACM Transactions on Graphics (TOG), 35(4):1-12, 2016. 3 +[49] Colin G. Scanes. Chapter 1 - animal perception including differences with humans. In Colin G. Scanes and Samia R. Toukhsati, editors, Animals and Human Society, pages 1-11. Academic Press, 2018. 1 +[50] Zhuo Su, Lan Xu, Dawei Zhong, Zhong Li, Fan Deng, Shuxue Quan, and Lu Fang. Robustfusion: Robust volumetric performance reconstruction under human-object interactions from monocular rgbd stream. arXiv preprint arXiv:2104.14837, 2021. 3 + +[51] Guoxing Sun, Xin Chen, Yizhang Chen, Anqi Pang, Pei Lin, Yuheng Jiang, Lan Xu, Jingyi Yu, and Jingya Wang. Neural free-viewpoint performance rendering under complex human-object interactions. In Proceedings of the 29th ACM International Conference on Multimedia, pages 4651-4660, 2021. 3 +[52] Zhongda Sun, Minglu Zhu, Xuechuan Shan, and Chengkuo Lee. Augmented tactile-perception and haptic-feedback rings as human-machine interfaces aiming for immersive interactions. Nature communications, 13(1):1-13, 2022. 3 +[53] Subramanian Sundaram, Petr Kellnhofer, Yunzhu Li, Jun-Yan Zhu, Antonio Torralba, and Wojciech Matusik. Learning the signatures of the human grasp using a scalable tactile glove. Nature, 569(7758):698-702, May 2019. 3 +[54] Omid Taheri, Vasileios Choutas, Michael J. Black, and Dimitrios Tzionas. GOAL: Generating 4D whole-body motion for hand-object grasping. In IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), pages 13263-13273, June 2022. 3 +[55] Omid Taheri, Nima Ghorbani, Michael J. Black, and Dimitrios Tzionas. GRAB: A Dataset of Whole-Body Human Grasping of Objects. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, Computer Vision - ECCV 2020, volume 12349, pages 581-600. Springer International Publishing, Cham, 2020. 3, 4 +[56] Weilin Wan, Lei Yang, Lingjie Liu, Zhuoying Zhang, Ruixing Jia, Yi-King Choi, Jia Pan, Christian Theobalt, Taku Komura, and Wenping Wang. Learn to predict how humans manipulate large-sized objects from interactive motions. IEEE Robotics and Automation Letters, 7(2):4702-4709, 2022. 3 +[57] Zhenzhen Weng and Serena Yeung. Holistic 3d human and scene mesh estimation from single view images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 334-343, 2021. 3 +[58] Yan Wu, Jiahao Wang, Yan Zhang, Siwei Zhang, Otmar Hilliges, Fisher Yu, and Siyu Tang. Saga: Stochastic whole-body grasping with contact. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 3 +[59] Xianghui Xie, Bharat Lal Bhatnagar, and Gerard Pons-Moll. Chore: Contact, human and object reconstruction from a single rgb image. In European Conference on Computer Vision (ECCV). Springer, October 2022. 3 +[60] Xianghui Xie, Bharat Lal Bhatnagar, and Gerard Pons-Moll. Visibility aware human-object interaction tracking from single rgb camera. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2023. 3 +[61] Lixin Yang, Xinyu Zhan, Kailin Li, Wenqiang Xu, Jiefeng Li, and Cewu Lu. Cpf: Learning a contact potential field to model the hand-object interaction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11097-11106, 2021. 3 +[62] Bangpeng Yao, Jiayuan Ma, and Li Fei-Fei. Discovering Object Functionality. In Proceedings of the IEEE International Conference on Computer Vision, pages 2512-2519, 2013. 1, 3 +[63] Yufei Ye, Abhinav Gupta, and Shubham Tulsiani. What's in your hands? 3d reconstruction of generic objects in hands. + +In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3895-3905, 2022. 1, 3 +[64] Hongwei Yi, Chun-Hao P. Huang, Dimitrios Tzionas, Muhammed Kocabas, Mohamed Hassan, Siyu Tang, Justus Thies, and Michael J. Black. Human-Aware Object Placement for Visual Environment Reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3959-3970, 2022. 3 +[65] Andrei Zanfir, Elisabella Maroiniu, and Cristian Sminchisescu. Monocular 3d pose and shape estimation of multiple people in natural scenes-the importance of multiple scene constraints. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2148-2157, 2018. 3 +[66] Jason Y Zhang, Panna Felsen, Angjoo Kanazawa, and Jitendra Malik. Predicting 3d human dynamics from video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7114-7123, 2019. 3 +[67] Jason Y Zhang, Sam Pepose, Hanbyul Joo, Deva Ramanan, Jitendra Malik, and Angjoo Kanazawa. Perceiving 3d human-object spatial arrangements from a single image in the wild. In European conference on computer vision, pages 34-51. Springer, 2020. 3 +[68] Siwei Zhang, Qianli Ma, Yan Zhang, Zhiyin Qian, Marc Pollefeys, Federica Bogo, and Siyu Tang. Egobody: Human body shape, motion and social interactions from head-mounted devices. arXiv preprint arXiv:2112.07642, 2021. 3 +[69] Siwei Zhang, Yan Zhang, Federica Bogo, Marc Pollefeys, and Siyu Tang. Learning motion priors for 4d human body capture in 3d scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11343-11353, 2021. 3 +[70] Xiaohan Zhang, Bharat Lal Bhatnagar, Sebastian Starke, Vladimir Guzov, and Gerard Pons-Moll. Couch: Towards controllable human-chair interactions. In European Conference on Computer Vision (ECCV). Springer, October 2022. 3 +[71] Qian Zheng, Weikai Wu, Hanting Pan, Niloy Mitra, Daniel Cohen-Or, and Hui Huang. Inferring object properties from human interaction and transferring them to new motions. Computational Visual Media, 7(3):375–392, Sept. 2021. 3 +[72] Tianhang Zheng, Changyou Chen, Junsong Yuan, Bo Li, and Kui Ren. Pointcloud saliency maps. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1598-1606, 2019. 8 +[73] Keyang Zhou, Bharat Lal Bhatnagar, Jan Eric Lenssen, and Gerard Pons-Moll. TOCH: Spatio-Temporal Object-to-Hand Correspondence for Motion Refinement, July 2022. 1, 3 \ No newline at end of file diff --git a/2023/Object Pop-Up_ Can We Infer 3D Objects and Their Poses From Human Interactions Alone_/images.zip b/2023/Object Pop-Up_ Can We Infer 3D Objects and Their Poses From Human Interactions Alone_/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1b3049e8e12832621aa5e12da28c08d76abe9711 --- /dev/null +++ b/2023/Object Pop-Up_ Can We Infer 3D Objects and Their Poses From Human Interactions Alone_/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:63e00e7514909e83eae192bd52670b09b673c41b1e24dd1d746163558cb2bcae +size 551090 diff --git a/2023/Object Pop-Up_ Can We Infer 3D Objects and Their Poses From Human Interactions Alone_/layout.json b/2023/Object Pop-Up_ Can We Infer 3D Objects and Their Poses From Human Interactions Alone_/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..05e54cdfc0e18bc6823225172c7a8b109921b28b --- /dev/null +++ b/2023/Object Pop-Up_ Can We Infer 3D Objects and Their Poses From Human Interactions Alone_/layout.json @@ -0,0 +1,8955 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 85, + 103, + 509, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 103, + 509, + 138 + ], + "spans": [ + { + "bbox": [ + 85, + 103, + 509, + 138 + ], + "type": "text", + "content": "Object pop-up: Can we infer 3D objects and their poses from human interactions alone?" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 107, + 170, + 485, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 170, + 485, + 186 + ], + "spans": [ + { + "bbox": [ + 107, + 170, + 485, + 186 + ], + "type": "text", + "content": "Ilya A. Petrov" + }, + { + "bbox": [ + 107, + 170, + 485, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 107, + 170, + 485, + 186 + ], + "type": "text", + "content": ", Riccardo Marin" + }, + { + "bbox": [ + 107, + 170, + 485, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 107, + 170, + 485, + 186 + ], + "type": "text", + "content": ", Julian Chibane" + }, + { + "bbox": [ + 107, + 170, + 485, + 186 + ], + "type": "inline_equation", + "content": "^{2,1}" + }, + { + "bbox": [ + 107, + 170, + 485, + 186 + ], + "type": "text", + "content": ", and Gerard Pons-Moll" + }, + { + "bbox": [ + 107, + 170, + 485, + 186 + ], + "type": "inline_equation", + "content": "^{1,2,3}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 232, + 199, + 361, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 232, + 199, + 361, + 213 + ], + "spans": [ + { + "bbox": [ + 232, + 199, + 361, + 213 + ], + "type": "text", + "content": "1University of Tübingen, Germany" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 153, + 213, + 439, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 213, + 439, + 224 + ], + "spans": [ + { + "bbox": [ + 153, + 213, + 439, + 224 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 153, + 213, + 439, + 224 + ], + "type": "text", + "content": "Max Planck Institute for Informatics, Saarland Informatics Campus, Germany" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 239, + 224, + 354, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 239, + 224, + 354, + 235 + ], + "spans": [ + { + "bbox": [ + 239, + 224, + 354, + 235 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 239, + 224, + 354, + 235 + ], + "type": "text", + "content": "Tübingen AI Center, Germany" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 74, + 235, + 518, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 235, + 518, + 258 + ], + "spans": [ + { + "bbox": [ + 74, + 235, + 518, + 258 + ], + "type": "text", + "content": "{i.petrov, gerard.pons-moll}@uni-tuebingen.de, riccardo.marin@mnf.uni-tuebingen.de, jchibane@mpi-inf.mpg.de" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 47, + 296, + 91, + 396 + ], + "blocks": [ + { + "bbox": [ + 51, + 279, + 85, + 294 + ], + "lines": [ + { + "bbox": [ + 51, + 279, + 85, + 294 + ], + "spans": [ + { + "bbox": [ + 51, + 279, + 85, + 294 + ], + "type": "text", + "content": "Ref. Image (not used)" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 47, + 296, + 91, + 396 + ], + "lines": [ + { + "bbox": [ + 47, + 296, + 91, + 396 + ], + "spans": [ + { + "bbox": [ + 47, + 296, + 91, + 396 + ], + "type": "image", + "image_path": "ff13a4932e654648edde69665a0a1ece8f97f4f484679df15926cc86bf559f08.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 403, + 547, + 448 + ], + "lines": [ + { + "bbox": [ + 45, + 403, + 547, + 448 + ], + "spans": [ + { + "bbox": [ + 45, + 403, + 547, + 448 + ], + "type": "text", + "content": "Figure 1. Human-Centric prediction of Human-Object Interaction. We explore Human-Object interaction when only the cues of the human pose interacting with an unobserved object are available ('Input'). We propose a neural model that, for the first time, can infer the location of the object ('Result') from such input. This is possible even when our subject simulates the interactions with the object ('Reference Image')." + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 94, + 296, + 127, + 393 + ], + "blocks": [ + { + "bbox": [ + 93, + 287, + 127, + 294 + ], + "lines": [ + { + "bbox": [ + 93, + 287, + 127, + 294 + ], + "spans": [ + { + "bbox": [ + 93, + 287, + 127, + 294 + ], + "type": "text", + "content": "Point Cloud" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 94, + 296, + 127, + 393 + ], + "lines": [ + { + "bbox": [ + 94, + 296, + 127, + 393 + ], + "spans": [ + { + "bbox": [ + 94, + 296, + 127, + 393 + ], + "type": "image", + "image_path": "9474c6239a1356b0689de49886b91f9d190d57f1d5b798e781cb5e9cf9f13e72.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 133, + 294, + 164, + 393 + ], + "blocks": [ + { + "bbox": [ + 103, + 280, + 119, + 286 + ], + "lines": [ + { + "bbox": [ + 103, + 280, + 119, + 286 + ], + "spans": [ + { + "bbox": [ + 103, + 280, + 119, + 286 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 140, + 285, + 159, + 293 + ], + "lines": [ + { + "bbox": [ + 140, + 285, + 159, + 293 + ], + "spans": [ + { + "bbox": [ + 140, + 285, + 159, + 293 + ], + "type": "text", + "content": "Result" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 133, + 294, + 164, + 393 + ], + "lines": [ + { + "bbox": [ + 133, + 294, + 164, + 393 + ], + "spans": [ + { + "bbox": [ + 133, + 294, + 164, + 393 + ], + "type": "image", + "image_path": "3e19da7a40c2e74032a938a644e3074e03779b64a7d0f74b90a186e261b867d1.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 172, + 296, + 215, + 393 + ], + "blocks": [ + { + "bbox": [ + 177, + 279, + 209, + 294 + ], + "lines": [ + { + "bbox": [ + 177, + 279, + 209, + 294 + ], + "spans": [ + { + "bbox": [ + 177, + 279, + 209, + 294 + ], + "type": "text", + "content": "Ref. Image (not used)" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 172, + 296, + 215, + 393 + ], + "lines": [ + { + "bbox": [ + 172, + 296, + 215, + 393 + ], + "spans": [ + { + "bbox": [ + 172, + 296, + 215, + 393 + ], + "type": "image", + "image_path": "2c894457d371750b0e0f25f1c6cc6cf2acbb5299eacb2470cecdfe7c2e6fb90e.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 217, + 296, + 251, + 392 + ], + "blocks": [ + { + "bbox": [ + 217, + 279, + 253, + 294 + ], + "lines": [ + { + "bbox": [ + 217, + 279, + 253, + 294 + ], + "spans": [ + { + "bbox": [ + 217, + 279, + 253, + 294 + ], + "type": "text", + "content": "Input Point Cloud" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 217, + 296, + 251, + 392 + ], + "lines": [ + { + "bbox": [ + 217, + 296, + 251, + 392 + ], + "spans": [ + { + "bbox": [ + 217, + 296, + 251, + 392 + ], + "type": "image", + "image_path": "687583a36d7ec1f7205d5617528687ea5e78f37b287b39c9564d1f9ebc48adef.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 258, + 296, + 290, + 392 + ], + "blocks": [ + { + "bbox": [ + 266, + 285, + 286, + 293 + ], + "lines": [ + { + "bbox": [ + 266, + 285, + 286, + 293 + ], + "spans": [ + { + "bbox": [ + 266, + 285, + 286, + 293 + ], + "type": "text", + "content": "Result" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 258, + 296, + 290, + 392 + ], + "lines": [ + { + "bbox": [ + 258, + 296, + 290, + 392 + ], + "spans": [ + { + "bbox": [ + 258, + 296, + 290, + 392 + ], + "type": "image", + "image_path": "c37c453829a71c4bbbae8edf24650c9088648d2070ce1f4c0ae5c96f174f4732.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 294, + 296, + 338, + 392 + ], + "blocks": [ + { + "bbox": [ + 301, + 279, + 333, + 294 + ], + "lines": [ + { + "bbox": [ + 301, + 279, + 333, + 294 + ], + "spans": [ + { + "bbox": [ + 301, + 279, + 333, + 294 + ], + "type": "text", + "content": "Ref. Image (not used)" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 294, + 296, + 338, + 392 + ], + "lines": [ + { + "bbox": [ + 294, + 296, + 338, + 392 + ], + "spans": [ + { + "bbox": [ + 294, + 296, + 338, + 392 + ], + "type": "image", + "image_path": "2f165549bdb8c9d5784a7e54f9ee99befc57a75d6c225308e185dfe332560873.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 341, + 296, + 375, + 391 + ], + "blocks": [ + { + "bbox": [ + 342, + 279, + 378, + 294 + ], + "lines": [ + { + "bbox": [ + 342, + 279, + 378, + 294 + ], + "spans": [ + { + "bbox": [ + 342, + 279, + 378, + 294 + ], + "type": "text", + "content": "Input Point Cloud" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 341, + 296, + 375, + 391 + ], + "lines": [ + { + "bbox": [ + 341, + 296, + 375, + 391 + ], + "spans": [ + { + "bbox": [ + 341, + 296, + 375, + 391 + ], + "type": "image", + "image_path": "f1a5f590815891caba96603958c61bec4c0c439fbbb66442e690b26e42a3faf5.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 382, + 296, + 415, + 391 + ], + "blocks": [ + { + "bbox": [ + 389, + 285, + 410, + 293 + ], + "lines": [ + { + "bbox": [ + 389, + 285, + 410, + 293 + ], + "spans": [ + { + "bbox": [ + 389, + 285, + 410, + 293 + ], + "type": "text", + "content": "Result" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 382, + 296, + 415, + 391 + ], + "lines": [ + { + "bbox": [ + 382, + 296, + 415, + 391 + ], + "spans": [ + { + "bbox": [ + 382, + 296, + 415, + 391 + ], + "type": "image", + "image_path": "b5fafb11d446414b848dee849492adba0c04e5691f29e584de0588b62a338c88.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 420, + 296, + 463, + 392 + ], + "blocks": [ + { + "bbox": [ + 424, + 279, + 457, + 294 + ], + "lines": [ + { + "bbox": [ + 424, + 279, + 457, + 294 + ], + "spans": [ + { + "bbox": [ + 424, + 279, + 457, + 294 + ], + "type": "text", + "content": "Ref. Image (not used)" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 420, + 296, + 463, + 392 + ], + "lines": [ + { + "bbox": [ + 420, + 296, + 463, + 392 + ], + "spans": [ + { + "bbox": [ + 420, + 296, + 463, + 392 + ], + "type": "image", + "image_path": "357160551435c931f39da0ac68f573519ff364f3a5eb11d57676e5a434aa0325.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 463, + 296, + 496, + 393 + ], + "blocks": [ + { + "bbox": [ + 468, + 279, + 503, + 294 + ], + "lines": [ + { + "bbox": [ + 468, + 279, + 503, + 294 + ], + "spans": [ + { + "bbox": [ + 468, + 279, + 503, + 294 + ], + "type": "text", + "content": "Input Point Cloud" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 463, + 296, + 496, + 393 + ], + "lines": [ + { + "bbox": [ + 463, + 296, + 496, + 393 + ], + "spans": [ + { + "bbox": [ + 463, + 296, + 496, + 393 + ], + "type": "image", + "image_path": "a705fbebadecf78663cb2592ebd8c3817e4b807d45e166317e7d15bb6e9d5ddc.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 503, + 296, + 534, + 393 + ], + "blocks": [ + { + "bbox": [ + 506, + 279, + 529, + 294 + ], + "lines": [ + { + "bbox": [ + 506, + 279, + 529, + 294 + ], + "spans": [ + { + "bbox": [ + 506, + 279, + 529, + 294 + ], + "type": "text", + "content": "Result" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 503, + 296, + 534, + 393 + ], + "lines": [ + { + "bbox": [ + 503, + 296, + 534, + 393 + ], + "spans": [ + { + "bbox": [ + 503, + 296, + 534, + 393 + ], + "type": "image", + "image_path": "3b6324fa122e1b48406fb033e003f6a071802a025becb7d5d6eeb8f72be88555.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "bbox": [ + 143, + 461, + 192, + 474 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 461, + 192, + 474 + ], + "spans": [ + { + "bbox": [ + 143, + 461, + 192, + 474 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 45, + 496, + 289, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 496, + 289, + 700 + ], + "spans": [ + { + "bbox": [ + 45, + 496, + 289, + 700 + ], + "type": "text", + "content": "The intimate entanglement between objects affordances and human poses is of large interest, among others, for behavioural sciences, cognitive psychology, and Computer Vision communities. In recent years, the latter has developed several object-centric approaches: starting from items, learning pipelines synthesizing human poses and dynamics in a realistic way, satisfying both geometrical and functional expectations. However, the inverse perspective is significantly less explored: Can we infer 3D objects and their poses from human interactions alone? Our investigation follows this direction, showing that a generic 3D human point cloud is enough to pop up an unobserved object, even when the user is just imitating a functionality (e.g., looking through a binocular) without involving a tangible counterpart. We validate our method qualitatively and quantitatively, with synthetic data and sequences acquired for the task, showing applicability for XR/VR." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 307, + 461, + 386, + 473 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 461, + 386, + 473 + ], + "spans": [ + { + "bbox": [ + 307, + 461, + 386, + 473 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 304, + 486, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 547, + 715 + ], + "type": "text", + "content": "Complex interactions with the world are among the unique skills distinguishing humans from other living beings. Even though our perception might be imperfect (we cannot hear ultrasonic sounds or see ultraviolet light [49]), our cognitive representation is enriched with a functional perspective, i.e., potential ways of interacting with objects or, as introduced by Gibson and colleagues, the affordance of the objects [19]. Several behavioural studies confirmed the centrality of this concept [2, 9, 43], which plays a fundamental role also for kids' development [2]. Computer Vision is well-aware that the function of an object complements its appearance [21], and exploited this in tasks like human and object reconstruction [62]. Previous literature approaches the interaction analysis from an object perspective (i.e., given an object, analyze the human interaction) [8,22,63], building object-centric priors [73], generating realistic grasps given the object [13, 34], reconstructing hand-object interactions [8,22,63]. Namely, objects induce functionality, so a human interaction (e.g., a mug suggests" + } + ] + } + ], + "index": 37 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4726" + } + ] + } + ], + "index": 38 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 69, + 276, + 214 + ], + "blocks": [ + { + "bbox": [ + 55, + 69, + 276, + 214 + ], + "lines": [ + { + "bbox": [ + 55, + 69, + 276, + 214 + ], + "spans": [ + { + "bbox": [ + 55, + 69, + 276, + 214 + ], + "type": "image", + "image_path": "075157ef91344df3f099866a5b5b90e89f72a1e3ef09ee021a731fc3b175eb17.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 227, + 288, + 283 + ], + "lines": [ + { + "bbox": [ + 46, + 227, + 288, + 283 + ], + "spans": [ + { + "bbox": [ + 46, + 227, + 288, + 283 + ], + "type": "text", + "content": "Figure 2. Human-Centric vs. Object-Centric inference. A common perspective in the prior work is to infer affordances or human pose starting from an object. We explore the inverse, Human-Centric perspective of the human-object interaction relationship: given the human, we are interested in predicting the object." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 294, + 287, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 294, + 287, + 389 + ], + "spans": [ + { + "bbox": [ + 46, + 294, + 287, + 389 + ], + "type": "text", + "content": "a drinking action; an handle a grasping one). For the first time, our work reverts the perspective, suggesting that analyzing human motion and behaviour is naturally a human-centric problem (i.e., given a human interaction, what kind of functionality is it suggesting, Fig. 2). Moving the first step in this new research direction, we pose a fundamental question: Can we infer 3D objects and their poses from human interactions alone?" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 390, + 288, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 390, + 288, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 390, + 288, + 617 + ], + "type": "text", + "content": "At first sight, the problem seems particularly hard and significantly under-constrained since several geometries might fit the same action. However, the human body complements this information in several ways: physical relations, characteristic poses, or body dynamics serve as valuable proxies for the involved functionality, as suggested by Fig. 1. Such hints are so powerful that we can easily imagine the kind of object and its location, even if such an object does not exist. Furthermore, even if the given pose might fit several possible solutions, our mind naturally comes to the most natural one indicated by the observed behaviour. It also suggests that solely focusing on the contact region (an approach often preferred by previous works) is insufficient in this new viewpoint. The reification principle of Gestalt psychology [37] highlights that \"the whole\" arises from \"the parts\" and their relation. Similarly, in Fig. 3, the hand grasps in B) pop up a binocular in our mind because we naturally consider the relationship with the other body parts." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 617, + 287, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 701 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 701 + ], + "type": "text", + "content": "Finally, moving to a human-centric perspective in human-object interaction is critical for human studies and daily-life applications. Modern systems for AR/VR [24], and digital interaction [26] are both centered on humans, often manipulating objects that do not have a real-world counterpart. Learning to decode an object from human behaviour enables unprecedented applications." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "text", + "content": "To answer our question, we deploy a first straightforward" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 312, + 76, + 408, + 167 + ], + "blocks": [ + { + "bbox": [ + 312, + 76, + 408, + 167 + ], + "lines": [ + { + "bbox": [ + 312, + 76, + 408, + 167 + ], + "spans": [ + { + "bbox": [ + 312, + 76, + 408, + 167 + ], + "type": "image", + "image_path": "7b03322c0e42245b8afa641d5659f3be172a5958ea7935a2c65c6e9bd893563d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 357, + 171, + 369, + 181 + ], + "lines": [ + { + "bbox": [ + 357, + 171, + 369, + 181 + ], + "spans": [ + { + "bbox": [ + 357, + 171, + 369, + 181 + ], + "type": "text", + "content": "A)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 410, + 78, + 531, + 169 + ], + "blocks": [ + { + "bbox": [ + 410, + 78, + 531, + 169 + ], + "lines": [ + { + "bbox": [ + 410, + 78, + 531, + 169 + ], + "spans": [ + { + "bbox": [ + 410, + 78, + 531, + 169 + ], + "type": "image", + "image_path": "a531002a2af6fdd8cb7e29cbf6fbc4abab31fc24793fb85281ef03ff919bcbe0.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 473, + 171, + 484, + 180 + ], + "lines": [ + { + "bbox": [ + 473, + 171, + 484, + 180 + ], + "spans": [ + { + "bbox": [ + 473, + 171, + 484, + 180 + ], + "type": "text", + "content": "B)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 186, + 545, + 230 + ], + "lines": [ + { + "bbox": [ + 304, + 186, + 545, + 230 + ], + "spans": [ + { + "bbox": [ + 304, + 186, + 545, + 230 + ], + "type": "text", + "content": "Figure 3. Reification principle. Gestalt theory suggests our perception considers the whole before the single parts (A). This served as inspiration for our work: the object can arise considering the body parts as a whole (B)." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 251, + 545, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 251, + 545, + 393 + ], + "spans": [ + { + "bbox": [ + 304, + 251, + 545, + 393 + ], + "type": "text", + "content": "and effective pipeline to \"pop up\" a rigid object from a 3D human point cloud. Starting from the input human point cloud and a class, we train an end-to-end pipeline to infer object location. In case a temporal sequence of point clouds is available, we suggest post-processing to avoid jittering and inconsistencies in the predictions, showing the relevance of this information to handle ambiguous poses. We show promising results on previously unaddressed tasks in digital and real-world scenes. Finally, our method allows us to analyze different features of human behaviour, highlight their contribution to object retrieval, and point to exciting directions for future works." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 317, + 395, + 481, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 395, + 481, + 406 + ], + "spans": [ + { + "bbox": [ + 317, + 395, + 481, + 406 + ], + "type": "text", + "content": "In summary, our main contributions are:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 414, + 545, + 549 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 313, + 414, + 545, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 414, + 545, + 450 + ], + "spans": [ + { + "bbox": [ + 313, + 414, + 545, + 450 + ], + "type": "text", + "content": "1. We formulate a novel problem, changing the perspective taken by previous works in the field, and open to a yet unexplored research direction;" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 457, + 545, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 457, + 545, + 482 + ], + "spans": [ + { + "bbox": [ + 313, + 457, + 545, + 482 + ], + "type": "text", + "content": "2. We introduce a method capable of predicting the object starting from an input human point cloud;" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 489, + 545, + 549 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 489, + 545, + 549 + ], + "spans": [ + { + "bbox": [ + 313, + 489, + 545, + 549 + ], + "type": "text", + "content": "3. We analyze different components of the human-object relationship: the contribution of different pieces of interactions (hands, body, time sequence), the point-wise saliency of input points, and the confusion produced by objects with similar functions." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 318, + 557, + 499, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 557, + 499, + 568 + ], + "spans": [ + { + "bbox": [ + 318, + 557, + 499, + 568 + ], + "type": "text", + "content": "All the code will be made publicly available." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 579, + 392, + 592 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 579, + 392, + 592 + ], + "spans": [ + { + "bbox": [ + 306, + 579, + 392, + 592 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 599, + 425, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 599, + 425, + 613 + ], + "spans": [ + { + "bbox": [ + 306, + 599, + 425, + 613 + ], + "type": "text", + "content": "2.1. Object Functionality" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 617, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 689 + ], + "type": "text", + "content": "In the core of human perception of objects, functionality complements physical appearance, enhancing our perception. Gibson [19] introduced the idea that humans use affordances of objects for perception. Affordance can be defined as: \"an intrinsic property of an object, allowing an action to be performed with the object\" [35]." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "content": "From a Computer Vision perspective, object functionality supports several tasks such as scene analysis [21], object" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4727" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 120 + ], + "type": "text", + "content": "classification [17, 35], object properties inferring [71], and it is also possible to learn object-specific human interaction models from 2D images [62]. These works suggest an intimate entanglement between the action and the object itself." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 127, + 194, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 127, + 194, + 140 + ], + "spans": [ + { + "bbox": [ + 47, + 127, + 194, + 140 + ], + "type": "text", + "content": "2.2. Human-Object interaction" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 145, + 287, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 145, + 287, + 194 + ], + "spans": [ + { + "bbox": [ + 46, + 145, + 287, + 194 + ], + "type": "text", + "content": "Modelling environment-aware humans and their interactions in 3D is one of the most recent challenges to creating virtual humans. We see two main lines of work there: hand-focused and full-body ones." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 207, + 287, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 207, + 287, + 375 + ], + "spans": [ + { + "bbox": [ + 46, + 207, + 287, + 375 + ], + "type": "text", + "content": "Hand-Object Interaction. Several works tackle the problem of human interaction, focusing solely on the hands [10]. This has been done starting from 2D [8, 15, 18, 22, 25, 29, 30, 34, 61, 63], 2.5D [3, 5], and 3D data [4, 34, 55]. Particularly promising seems the application of well-designed priors for the motion [73]. The class of objects involved in these works are mainly limited to graspable ones. We argue that interactions involving different body parts are common in everyday life, more attractive from an applicative perspective, and more challenging. Moreover, full-body context is crucial in reconstructing even grasp interactions since body pose contains information on an object's properties, e.g. pounding with a hammer affects the whole posture to support the action." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 389, + 287, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 389, + 287, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 389, + 287, + 604 + ], + "type": "text", + "content": "Fully-Body Interaction. On this line, several works focus on the interaction between a human and a scene [12,28, 31,38,57,64,65,69]. Also, in this case, priors can be used to regularize the motion [48]. Several datasets are also available to study the interaction between a human and a single object. For example, recent BEHAVE [1], GRAB [55], and InterCap [32] capture full-body interactions with diverse objects. Works address the task of humans interactions reconstruction from different kinds of data sources like single image [59, 67], video [16, 50, 60], and multiview capturing [1, 33, 51], and synthetization of them as well [7,27,41,54,56,58,66,70]. However, in all these works, we observe a general object-centric perspective: given a scene or an object, they aim to recreate the humans interacting with them. We argue that the significantly less explored complementary one has more concrete applications in daily life, especially in VR/XR contexts where the human is central to the system." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 611, + 197, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 611, + 197, + 624 + ], + "spans": [ + { + "bbox": [ + 47, + 611, + 197, + 624 + ], + "type": "text", + "content": "2.3. Human-centric perspective" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "content": "While the general trend mainly focuses on the surrounding environment and objects, there is a growing interest and availability of egocentric tools for humans [23, 68] also interacting with objects [24,36,39]. They provide a subjective view and a valuable paradigm for several applications, like letting a user interact with objects in the digital world. Recent works also involve more sophisticated devices [52,53]," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 180 + ], + "type": "text", + "content": "while they are still far from applicability. In a similar direction to our work, others propose to recover objects arrangements in a room starting from the human motion [42], to hallucinate a coherent 2D image from a human pose [6], or predicting physical properties of the objects (e.g., the weight of a box) from human joints [71]. While the principle inspires us, our study significantly differs: we focus on object pose and its spatial relation with humans, starting solely from unordered point clouds." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 190, + 361, + 202 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 190, + 361, + 202 + ], + "spans": [ + { + "bbox": [ + 306, + 190, + 361, + 202 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 210, + 545, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 210, + 545, + 247 + ], + "spans": [ + { + "bbox": [ + 304, + 210, + 545, + 247 + ], + "type": "text", + "content": "This section describes our setting and the main components of our methodology, both at inference and training time. An overview of our pipeline can be found in Fig. 4." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 253, + 399, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 253, + 399, + 266 + ], + "spans": [ + { + "bbox": [ + 306, + 253, + 399, + 266 + ], + "type": "text", + "content": "3.1. Object Pop-Up" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 271, + 545, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 271, + 545, + 368 + ], + "spans": [ + { + "bbox": [ + 304, + 271, + 545, + 368 + ], + "type": "text", + "content": "Input. Our method starts from a single human point cloud " + }, + { + "bbox": [ + 304, + 271, + 545, + 368 + ], + "type": "inline_equation", + "content": "\\mathbf{P} \\in \\mathbb{R}^{N_P \\times 3}" + }, + { + "bbox": [ + 304, + 271, + 545, + 368 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 304, + 271, + 545, + 368 + ], + "type": "inline_equation", + "content": "N_P" + }, + { + "bbox": [ + 304, + 271, + 545, + 368 + ], + "type": "text", + "content": " points and a hot-encoded object class " + }, + { + "bbox": [ + 304, + 271, + 545, + 368 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 271, + 545, + 368 + ], + "type": "text", + "content": ". The input point cloud can result from 3D/4D scans, IMUs template fitting, or any other shape-from-X approach. Regardless of the point cloud source, we remark that no further information is used apart from the 3D coordinates of the points. We represent each object as 1500 key points " + }, + { + "bbox": [ + 304, + 271, + 545, + 368 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_c" + }, + { + "bbox": [ + 304, + 271, + 545, + 368 + ], + "type": "text", + "content": " uniformly sampled on the template mesh." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 382, + 545, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 382, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 304, + 382, + 545, + 525 + ], + "type": "text", + "content": "Object Center. Training a model to predict an object pose from a human point cloud poses several challenges. Such a task requires the network to understand the location of different body parts and their subtle relations while jointly developing a sense of its spatial relationship with the human. Empirically, we observed that this is only feasible by carefully deconstructing the problem and designing different features to ease the learning process. As the first step to decompose this problem, we train a PointNet++ architecture [11,45] to predict the object center " + }, + { + "bbox": [ + 304, + 382, + 545, + 525 + ], + "type": "inline_equation", + "content": "\\mathbf{o}_P" + }, + { + "bbox": [ + 304, + 382, + 545, + 525 + ], + "type": "text", + "content": " starting from " + }, + { + "bbox": [ + 304, + 382, + 545, + 525 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 304, + 382, + 545, + 525 + ], + "type": "text", + "content": ". At training time, this is supervised with an L2 loss against the ground truth center " + }, + { + "bbox": [ + 304, + 382, + 545, + 525 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathbf{o}}_P" + }, + { + "bbox": [ + 304, + 382, + 545, + 525 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 373, + 533, + 545, + 547 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 533, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 373, + 533, + 545, + 547 + ], + "type": "interline_equation", + "content": "L _ {\\mathbf {o} _ {P}} (\\mathbf {P}) = \\left\\| \\mathbf {o} _ {P} - \\widehat {\\mathbf {o}} _ {P} \\right\\| _ {2} ^ {2}. \\tag {1}", + "image_path": "7fcb1e5151b3446f75481e837831b1c741aa541606cc8cbe65d564251339257d.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 555, + 545, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 555, + 545, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 555, + 545, + 616 + ], + "type": "text", + "content": "Solving this task provides a good initialization for the object pose, and we move the key points associated with the input class " + }, + { + "bbox": [ + 304, + 555, + 545, + 616 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 555, + 545, + 616 + ], + "type": "text", + "content": " to the predicted center " + }, + { + "bbox": [ + 304, + 555, + 545, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{o}_P" + }, + { + "bbox": [ + 304, + 555, + 545, + 616 + ], + "type": "text", + "content": ". Also, the features " + }, + { + "bbox": [ + 304, + 555, + 545, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_o \\in \\mathbb{R}^{512}" + }, + { + "bbox": [ + 304, + 555, + 545, + 616 + ], + "type": "text", + "content": " extracted by the network encode import information on the whole human body." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": "Local Neighbourhood. The center prediction module can be further exploited using the nearest human regions. Intuitively, considering the closest body parts is essential to infer a contact relationship, but also the influence of body parts not directly in touch with the object (e.g., head orientation while using binoculars). To learn these connections, we consider the centered key points together with" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4728" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 66, + 527, + 196 + ], + "blocks": [ + { + "bbox": [ + 67, + 66, + 527, + 196 + ], + "lines": [ + { + "bbox": [ + 67, + 66, + 527, + 196 + ], + "spans": [ + { + "bbox": [ + 67, + 66, + 527, + 196 + ], + "type": "image", + "image_path": "13bc0c9288897809a7841d0b5a428d093b1e2e2973554020caa5b64564b8f270.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 204, + 547, + 237 + ], + "lines": [ + { + "bbox": [ + 46, + 204, + 547, + 237 + ], + "spans": [ + { + "bbox": [ + 46, + 204, + 547, + 237 + ], + "type": "text", + "content": "Figure 4. Object pop-up. Our method predicts the position of an object, starting only from an input point cloud and an object class. It relies on a careful problem decomposition in several sub-tasks, extracting features that involve the entire human body and relations between body parts near the object." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 243, + 288, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 243, + 288, + 292 + ], + "spans": [ + { + "bbox": [ + 46, + 243, + 288, + 292 + ], + "type": "text", + "content": "the 3000 closest points of the input human point cloud: " + }, + { + "bbox": [ + 46, + 243, + 288, + 292 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_L = KNN(\\mathbf{P},\\mathbf{o}_P)" + }, + { + "bbox": [ + 46, + 243, + 288, + 292 + ], + "type": "text", + "content": ". We pass these two sets as a unique point cloud to a PointNet++ network, obtaining a new set of per-point features " + }, + { + "bbox": [ + 46, + 243, + 288, + 292 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{P_L}\\in \\mathbb{R}^{128}" + }, + { + "bbox": [ + 46, + 243, + 288, + 292 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 304, + 287, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 304, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 46, + 304, + 287, + 437 + ], + "type": "text", + "content": "Object displacement. To predict the object's final position, we empirically observed that directly predicting a rotation and a translation is not a good solution. Inspired by recent works that suggest a point-wise offset prediction to recover 3D human shapes [14], we apply a similar approach to our task. Our goal is to predict a point-wise shift " + }, + { + "bbox": [ + 46, + 304, + 287, + 437 + ], + "type": "inline_equation", + "content": "\\mathbf{S}_K" + }, + { + "bbox": [ + 46, + 304, + 287, + 437 + ], + "type": "text", + "content": " for the " + }, + { + "bbox": [ + 46, + 304, + 287, + 437 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_c" + }, + { + "bbox": [ + 46, + 304, + 287, + 437 + ], + "type": "text", + "content": " vertices to align them to the target pose. We append the features " + }, + { + "bbox": [ + 46, + 304, + 287, + 437 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_o" + }, + { + "bbox": [ + 46, + 304, + 287, + 437 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 304, + 287, + 437 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{P_L}" + }, + { + "bbox": [ + 46, + 304, + 287, + 437 + ], + "type": "text", + "content": ", the one-hot encoding of the object class " + }, + { + "bbox": [ + 46, + 304, + 287, + 437 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 304, + 287, + 437 + ], + "type": "text", + "content": ", and a positional encoding to the centered key points, and we pass them to a decoder. At training time, we consider the following loss:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 84, + 441, + 287, + 456 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 441, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 84, + 441, + 287, + 456 + ], + "type": "interline_equation", + "content": "L _ {o f f} \\left(\\mathbf {K} _ {c}, \\mathbf {F} _ {o}, \\mathbf {F} _ {P _ {L}}, c\\right) = \\left\\| \\mathbf {S} _ {K} - \\widehat {\\mathbf {S}} _ {K} \\right\\| _ {F} ^ {2}. \\tag {2}", + "image_path": "9ff130da039d5bd4b2eb5cf57518936a1ffd4cff44f5f2e12117a59b2c572747.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 461, + 244, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 461, + 244, + 473 + ], + "spans": [ + { + "bbox": [ + 59, + 461, + 244, + 473 + ], + "type": "text", + "content": "The network is then trained end-to-end using:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 125, + 479, + 287, + 493 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 479, + 287, + 493 + ], + "spans": [ + { + "bbox": [ + 125, + 479, + 287, + 493 + ], + "type": "interline_equation", + "content": "L = L _ {\\mathbf {o} _ {P}} + \\alpha L _ {\\text {o f f}}. \\tag {3}", + "image_path": "db2fc062092536becc7a95cd4d1217c5b5aa883c967ead77362f65e9395d8298.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 498, + 196, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 498, + 196, + 510 + ], + "spans": [ + { + "bbox": [ + 47, + 498, + 196, + 510 + ], + "type": "text", + "content": "The weighting coefficient is " + }, + { + "bbox": [ + 47, + 498, + 196, + 510 + ], + "type": "inline_equation", + "content": "\\alpha = 10" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 516, + 145, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 516, + 145, + 529 + ], + "spans": [ + { + "bbox": [ + 47, + 516, + 145, + 529 + ], + "type": "text", + "content": "3.2. Template fitting" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 534, + 287, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 287, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 287, + 617 + ], + "type": "text", + "content": "Procrustes alignment. The point-wise offset produced by the network potentially distorts the key points structure in a non-rigid way. To recover the desired global rigid transformation, we rely on a Procrustes alignment [20]. This procedure takes as input two point clouds and returns the rotation " + }, + { + "bbox": [ + 46, + 534, + 287, + 617 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 46, + 534, + 287, + 617 + ], + "type": "text", + "content": " and the translation " + }, + { + "bbox": [ + 46, + 534, + 287, + 617 + ], + "type": "inline_equation", + "content": "\\mathbf{t}" + }, + { + "bbox": [ + 46, + 534, + 287, + 617 + ], + "type": "text", + "content": " to minimize the L2 distances of the points:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 131, + 618, + 287, + 631 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 618, + 287, + 631 + ], + "spans": [ + { + "bbox": [ + 131, + 618, + 287, + 631 + ], + "type": "interline_equation", + "content": "P (\\cdot , \\cdot) \\rightarrow (\\mathbf {R}, \\mathbf {t}). \\tag {4}", + "image_path": "26e8b665f2dc8d2d9e790d5acf582077c0ae76bf8b729b43699921be65590c7d.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 634, + 287, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 634, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 47, + 634, + 287, + 658 + ], + "type": "text", + "content": "We apply this to the template key points and their configuration obtained with our network:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 95, + 664, + 287, + 677 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 664, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 95, + 664, + 287, + 677 + ], + "type": "interline_equation", + "content": "P \\left(\\mathbf {K} _ {c}, \\mathbf {K} _ {c} + \\mathbf {o} _ {P} + \\mathbf {S} _ {K}\\right)\\rightarrow (\\mathbf {R}, \\mathbf {t}). \\tag {5}", + "image_path": "1a79a17b64a9b7eca6395d9b869c050da87a04261359333180a8b11d8fc46afd.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 59, + 683, + 246, + 695 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 683, + 246, + 695 + ], + "spans": [ + { + "bbox": [ + 59, + 683, + 246, + 695 + ], + "type": "text", + "content": "Finally, we recover the desired object pose as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 135, + 700, + 287, + 713 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 700, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 135, + 700, + 287, + 713 + ], + "type": "interline_equation", + "content": "\\mathbf {T} ^ {\\prime} = \\mathbf {R T} _ {c} + \\mathbf {t} \\tag {6}", + "image_path": "5b2026945d8be0ad1cd181c5164ac2045d780c6867720a3841f59c0cb066d896.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 243, + 547, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 243, + 547, + 387 + ], + "spans": [ + { + "bbox": [ + 304, + 243, + 547, + 387 + ], + "type": "text", + "content": "Time Smoothing. While our pipeline is designed to work with a single point cloud as input, considering the temporal evolution of interaction is often crucial, shaping the context of the individual poses. If a temporal sequence of point clouds is available, we provide a post-processing smoothing technique to take advantage of this further information. After running our method for each frame, we smooth the centering prediction across the sequence using a Gaussian kernel. Later, we will discuss a variation of our approach that also predicts the object class. In that case, we consider the most frequent class prediction over the whole set of frames to fix a class for the sequence." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 400, + 388, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 400, + 388, + 414 + ], + "spans": [ + { + "bbox": [ + 305, + 400, + 388, + 414 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 421, + 546, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 421, + 546, + 481 + ], + "spans": [ + { + "bbox": [ + 304, + 421, + 546, + 481 + ], + "type": "text", + "content": "In this Section, we will describe the datasets used for training and testing our method. Then, we will present the baseline and the evaluation metrics. Finally, we will provide validation of our method as well as an analysis of its extended version that allows object class prediction." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 501, + 545, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 501, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 305, + 501, + 545, + 525 + ], + "type": "text", + "content": "Datasets We jointly train on the union of BEHAVE [1] and GRAB [55], obtaining a set of:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 535, + 545, + 606 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 317, + 535, + 545, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 535, + 545, + 559 + ], + "spans": [ + { + "bbox": [ + 317, + 535, + 545, + 559 + ], + "type": "text", + "content": "15 subjects, first 8 subjects from the GRAB dataset and 7 subjects from the official training part of BEHAV;" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 571, + 545, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 571, + 545, + 606 + ], + "spans": [ + { + "bbox": [ + 317, + 571, + 545, + 606 + ], + "type": "text", + "content": "- 40 different classes of objects, including all 20 objects from the BEHAVE dataset and 20 selected objects from the GRAB dataset;" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "We downsample training sequences of GRAB and BEHAVE to " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "10fps" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": ". To evaluate our method, we select subjects 9 and 10 from the GRAB dataset and downsample the sequences to " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "30fps" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": ". For the BEHAVE dataset, we use the official test part, which includes all sequences at " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "1fps" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": " with subject 3 and part of the sequences with subjects 4, 5. As an input, we use point clouds with 9000 points sampled uniformly over the SMPL-H [44] meshes. We refer to raw" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4729" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 161, + 70, + 434, + 116 + ], + "blocks": [ + { + "bbox": [ + 161, + 70, + 434, + 116 + ], + "lines": [ + { + "bbox": [ + 161, + 70, + 434, + 116 + ], + "spans": [ + { + "bbox": [ + 161, + 70, + 434, + 116 + ], + "type": "table", + "html": "
MethodsGRABBEHAVEBEHAVE-Raw
Ec↓Ev2v↓Ec↓Ev2v↓Ec↓Ev2v↓
NN0.03620.14450.08020.3445--
Ours0.02370.09430.06630.29000.08060.3143
", + "image_path": "39a8bcf771f32d10ed2892f9c60885239100293155c442332040d7348f12592a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 125, + 547, + 149 + ], + "lines": [ + { + "bbox": [ + 46, + 125, + 547, + 149 + ], + "spans": [ + { + "bbox": [ + 46, + 125, + 547, + 149 + ], + "type": "text", + "content": "Table 1. Comparison with the baseline. Our method significantly outperforms the baseline, even though the baseline uses the vertex order as additional information. Moreover, the baseline method does not generalize to point clouds with an arbitrary number of vertices." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 158, + 288, + 195 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 158, + 288, + 195 + ], + "spans": [ + { + "bbox": [ + 46, + 158, + 288, + 195 + ], + "type": "text", + "content": "point clouds from the BEHAVE dataset used in our experiments as BEHAVE-Raw. We use point clouds that are fused from 4 Kinect sensors and subsample " + }, + { + "bbox": [ + 46, + 158, + 288, + 195 + ], + "type": "inline_equation", + "content": "90k" + }, + { + "bbox": [ + 46, + 158, + 288, + 195 + ], + "type": "text", + "content": " points from them." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 219, + 288, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 219, + 288, + 257 + ], + "spans": [ + { + "bbox": [ + 46, + 219, + 288, + 257 + ], + "type": "text", + "content": "Data augmentation. During training, to simulate errors in the center prediction, we randomly translate and rotate the object around the ground-truth center " + }, + { + "bbox": [ + 46, + 219, + 288, + 257 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathbf{o}}_P" + }, + { + "bbox": [ + 46, + 219, + 288, + 257 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 281, + 287, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 281, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 46, + 281, + 287, + 376 + ], + "type": "text", + "content": "Implementation details. We implement our method using PyTorch framework and use Nvidia RTX3090 GPU for training and evaluation. The model is trained using Adam optimizer for 60 epochs with a learning rate of " + }, + { + "bbox": [ + 46, + 281, + 287, + 376 + ], + "type": "inline_equation", + "content": "1e^{-4}" + }, + { + "bbox": [ + 46, + 281, + 287, + 376 + ], + "type": "text", + "content": ", which decays 10 times after 30-th and 40-th epochs. For the first 20 epochs of training, we use ground-truth object center " + }, + { + "bbox": [ + 46, + 281, + 287, + 376 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathbf{o}}_P" + }, + { + "bbox": [ + 46, + 281, + 287, + 376 + ], + "type": "text", + "content": " instead of predicted " + }, + { + "bbox": [ + 46, + 281, + 287, + 376 + ], + "type": "inline_equation", + "content": "\\mathbf{o}_P" + }, + { + "bbox": [ + 46, + 281, + 287, + 376 + ], + "type": "text", + "content": " to select local neighborhood " + }, + { + "bbox": [ + 46, + 281, + 287, + 376 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{P_L}" + }, + { + "bbox": [ + 46, + 281, + 287, + 376 + ], + "type": "text", + "content": " to warm up the local PointNet++ encoder." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 401, + 288, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 401, + 288, + 558 + ], + "spans": [ + { + "bbox": [ + 46, + 401, + 288, + 558 + ], + "type": "text", + "content": "Nearest-Neighbor Baseline. Since we are the first to tackle this task and no competitors are available, we propose a simple while informative baseline. Given the input point cloud, we recover the most similar in the training dataset in an L2 sense. Then, we recover the object handled by that subject and pose it in space in the same way. This baseline demonstrates that the task is non-trivial and the generalization to unseen poses and subjects of our method. Also, this baseline requires that the target point cloud and the ones in the training set share the same number of points. Hence, if the input point cloud is a raw scan, this baseline is not applicable. Our method, instead, does not rely on this assumption and is more general." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": "Object classification. In our research, we also investigate the possibility of incorporating class prediction inside the network training. This task is significantly difficult at a single-frame level since an isolated pose often does not suggest a clear functionality. However, including this step is interesting to analyze the interaction and the nature of the network confusion. Hence, we modify our method by adding a decoder module that takes the global features " + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_o" + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": " and the local ones " + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{P_L}" + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": " as input to predict the object class. Then, we add to the training a simple cross-entropy loss between the predicted class and the ground truth one " + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\widehat{C}_P" + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 158, + 365, + 170 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 158, + 365, + 170 + ], + "spans": [ + { + "bbox": [ + 306, + 158, + 365, + 170 + ], + "type": "text", + "content": "4.1. Metrics" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 176, + 547, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 176, + 547, + 213 + ], + "spans": [ + { + "bbox": [ + 305, + 176, + 547, + 213 + ], + "type": "text", + "content": "In qualitative experiments, we use three main metrics to evaluate our results. In the tables, we report the average error across the considered test samples." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 228, + 546, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 228, + 546, + 276 + ], + "spans": [ + { + "bbox": [ + 305, + 228, + 546, + 276 + ], + "type": "text", + "content": "Vertex-to-vertex. In most cases, our resulting object and the target one share the same number of vertices. Hence, we can compute the error between our prediction and the ground truth " + }, + { + "bbox": [ + 305, + 228, + 546, + 276 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathbf{T}}'" + }, + { + "bbox": [ + 305, + 228, + 546, + 276 + ], + "type": "text", + "content": " as a point-to-point error:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 381, + 284, + 546, + 300 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 284, + 546, + 300 + ], + "spans": [ + { + "bbox": [ + 381, + 284, + 546, + 300 + ], + "type": "interline_equation", + "content": "E _ {v 2 v} = \\left\\| \\mathbf {T} ^ {\\prime} - \\widehat {\\mathbf {T}} ^ {\\prime} \\right\\| _ {F}. \\tag {7}", + "image_path": "568b98e9700f9cdeff3b192c57824ad02a95a8b0b249003ba73d90021c673a38.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 308, + 545, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 308, + 545, + 333 + ], + "spans": [ + { + "bbox": [ + 305, + 308, + 545, + 333 + ], + "type": "text", + "content": "When such error is computed only between the object centers, we will refer to it as " + }, + { + "bbox": [ + 305, + 308, + 545, + 333 + ], + "type": "inline_equation", + "content": "E_{c}" + }, + { + "bbox": [ + 305, + 308, + 545, + 333 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 347, + 547, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 347, + 547, + 396 + ], + "spans": [ + { + "bbox": [ + 305, + 347, + 547, + 396 + ], + "type": "text", + "content": "Chamfer distance. When we evaluate the network that also predicts the class, target objects and selected templates might not share the same number of vertices. In that case, as a metric we use bi-directional Chamfer distance:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 402, + 545, + 444 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 402, + 545, + 444 + ], + "spans": [ + { + "bbox": [ + 306, + 402, + 545, + 444 + ], + "type": "interline_equation", + "content": "E _ {c h} = \\frac {1}{\\| \\mathbf {T} ^ {\\prime} \\|} \\sum_ {x \\in \\mathbf {T} ^ {\\prime}} \\min _ {y \\in \\widehat {\\mathbf {T}} ^ {\\prime}} \\| x - y \\| _ {2} + \\frac {1}{\\| \\widehat {\\mathbf {T}} ^ {\\prime} \\|} \\sum_ {y \\in \\widehat {\\mathbf {T}} ^ {\\prime}} \\min _ {x \\in \\mathbf {T} ^ {\\prime}} \\| y - x \\| _ {2} \\tag {8}", + "image_path": "4ec924150090ac92e74f991e57ba713cc81ef091e03754ff048a3f6c4d3642d8.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 460, + 545, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 460, + 545, + 496 + ], + "spans": [ + { + "bbox": [ + 305, + 460, + 545, + 496 + ], + "type": "text", + "content": "Classification Accuracy. In case we use our network to predict the object class, we measure our misclassification error in terms of accuracy." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 503, + 438, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 503, + 438, + 517 + ], + "spans": [ + { + "bbox": [ + 306, + 503, + 438, + 517 + ], + "type": "text", + "content": "4.2. Object pose Evaluation" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 522, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 547, + 715 + ], + "type": "text", + "content": "In Tab. 1, we report the quantitative evaluation on the test set of the datasets, comparing our method to the baseline. Our approach significantly outperforms the baseline, even if this latter exploits the points order information. We obtain the most significant margin on the GRAB dataset, where objects are small and mainly involve hands, showing the precision of our method. The baseline cannot be applied on BEHAVE-Raw since it does not share the same number of vertices as the training set, while our method shows only a limited performance decrease, pointing to generalization also to point clouds coming from different sources. We report qualitative results of our method on GRAB (first two rows) and BEHAVE (last row) in Fig. 5, and on BEHAVE-Raw in Fig. 6. Finally, to further evaluate the generalization of our method to unseen poses and subjects, we also considered point clouds obtained by an egocentric pipeline." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "4730" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 69, + 545, + 541 + ], + "blocks": [ + { + "bbox": [ + 50, + 69, + 545, + 541 + ], + "lines": [ + { + "bbox": [ + 50, + 69, + 545, + 541 + ], + "spans": [ + { + "bbox": [ + 50, + 69, + 545, + 541 + ], + "type": "image", + "image_path": "ff078a73221b39123d11e20a8864b6816720a635944a6a5f07dcc1af96a469bd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 74, + 545, + 517, + 556 + ], + "lines": [ + { + "bbox": [ + 74, + 545, + 517, + 556 + ], + "spans": [ + { + "bbox": [ + 74, + 545, + 517, + 556 + ], + "type": "text", + "content": "Figure 5. Qualitative results. Results of our method on GRAB (first and second rows) and BEHAVE (third row) datasets." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 563, + 288, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 563, + 288, + 646 + ], + "spans": [ + { + "bbox": [ + 46, + 563, + 288, + 646 + ], + "type": "text", + "content": "In this scenario, we record a user motion with an XSens system [46], retarget the pose to an SMPL+H model, and obtain the point cloud by sampling the resulting mesh. Outputs of our method can be observed in Fig. 1 and Fig. 7. Generalizing to unseen poses and subjects acquired with wearable systems prone to measurement errors opens several exciting applications for VR/XR contexts." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 658, + 201, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 201, + 671 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 201, + 671 + ], + "type": "text", + "content": "4.3. Human Affordance analysis" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": "We use our method to analyze human-object interaction considering three key factors: changes in the input information, the points saliency, and confusion in classification." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 563, + 545, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 563, + 545, + 598 + ], + "spans": [ + { + "bbox": [ + 305, + 563, + 545, + 598 + ], + "type": "text", + "content": "Input information. Since we are curious to analyze how information is encoded in the inputs, we consider three more scenarios:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 317, + 608, + 547, + 715 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 317, + 608, + 547, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 608, + 547, + 657 + ], + "spans": [ + { + "bbox": [ + 317, + 608, + 547, + 657 + ], + "type": "text", + "content": "- Hands: We generate the input data using the MANO [47] hand model annotations included in GRAB dataset. Hence, the model should infer the pose of the object without relying on other body parts." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 665, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 665, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 317, + 665, + 547, + 715 + ], + "type": "text", + "content": "- SMPL: We consider all the points of the subject, but the ones from the hands are in the rest pose provided by SMPL [40]. In this way, we analyze how much the network captures the information of the finger pose." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "text", + "content": "4731" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 146, + 70, + 449, + 137 + ], + "blocks": [ + { + "bbox": [ + 146, + 70, + 449, + 137 + ], + "lines": [ + { + "bbox": [ + 146, + 70, + 449, + 137 + ], + "spans": [ + { + "bbox": [ + 146, + 70, + 449, + 137 + ], + "type": "table", + "html": "
MethodsGRABBEHAVEBEHAVE-Raw
Ec ↓Ev2v ↓Ec ↓Ev2v ↓Ec ↓Ev2v ↓
Ours, hands0.30150.7358----
Ours, SMPL0.02450.10090.06670.28770.08430.3205
Ours, SMPLH0.02370.09430.06630.29000.08060.3143
Ours, SMPLH+T0.02350.09830.06860.28800.08230.3161
", + "image_path": "c2e7c468e0c5d55fcb5d29431530726921dbb8b302b418aca884ac8d3668161a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 147, + 547, + 169 + ], + "lines": [ + { + "bbox": [ + 46, + 147, + 547, + 169 + ], + "spans": [ + { + "bbox": [ + 46, + 147, + 547, + 169 + ], + "type": "text", + "content": "Table 2. Human Affordance. We use our pipeline to explore how different inputs affect the object pose prediction. Using full body provides richer features for object recovery." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 51, + 171, + 545, + 286 + ], + "blocks": [ + { + "bbox": [ + 51, + 171, + 545, + 286 + ], + "lines": [ + { + "bbox": [ + 51, + 171, + 545, + 286 + ], + "spans": [ + { + "bbox": [ + 51, + 171, + 545, + 286 + ], + "type": "image", + "image_path": "e346cad04e13a3ab85f8e14b6239ebad852988c6653614dae40f3b10b6ebcc36.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 292, + 548, + 315 + ], + "lines": [ + { + "bbox": [ + 46, + 292, + 548, + 315 + ], + "spans": [ + { + "bbox": [ + 46, + 292, + 548, + 315 + ], + "type": "text", + "content": "Figure 6. BEHAVE-Raw. Results of our method considering as input raw point clouds from BEHAVE dataset. Despite the amount of noise and occlusions, our method is able to generalize and performs reliably." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 55, + 319, + 534, + 432 + ], + "blocks": [ + { + "bbox": [ + 55, + 319, + 534, + 432 + ], + "lines": [ + { + "bbox": [ + 55, + 319, + 534, + 432 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 534, + 432 + ], + "type": "image", + "image_path": "2ab01b3ce5b89b9a309e2a7421656945da492d5b268444a505f1d9a1eff979dc.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 62, + 460, + 534, + 571 + ], + "blocks": [ + { + "bbox": [ + 46, + 437, + 547, + 460 + ], + "lines": [ + { + "bbox": [ + 46, + 437, + 547, + 460 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 547, + 460 + ], + "type": "text", + "content": "Figure 7. Human from IMUs. Results of our method from a motion sequence acquired with IMUs. Even if the subject is unseen and the motion is subject to noise in the IMUs sensors, our method produces stable results." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 62, + 460, + 534, + 571 + ], + "lines": [ + { + "bbox": [ + 62, + 460, + 534, + 571 + ], + "spans": [ + { + "bbox": [ + 62, + 460, + 534, + 571 + ], + "type": "image", + "image_path": "c6677345b13a7e64ad2cf0e8caf3b72a03378454766f907d656201d0446aa01d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 574, + 547, + 598 + ], + "lines": [ + { + "bbox": [ + 46, + 574, + 547, + 598 + ], + "spans": [ + { + "bbox": [ + 46, + 574, + 547, + 598 + ], + "type": "text", + "content": "Figure 8. Saliency. Point cloud saliency computed for different objects, rendered from two perspectives. The contact region is relevant for all the interactions, while the network also focuses on the feet and head regions. All the predictions are results of our method." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 58, + 608, + 289, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 608, + 289, + 681 + ], + "spans": [ + { + "bbox": [ + 58, + 608, + 289, + 681 + ], + "type": "text", + "content": "- SMPLH+T: We consider all the points of the subject with the finger correctly posed (as in our method), and also we apply the temporal smoothing outlined in Sec. 3. Temporal information contextualizes the interactions and regularizes the predictions throughout the sequence." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 689, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 289, + 715 + ], + "type": "text", + "content": "For the first two cases, we train ad-hoc networks. Results of this analysis are reported in Tab. 2. We notice that using" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 608, + 547, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 608, + 547, + 657 + ], + "spans": [ + { + "bbox": [ + 304, + 608, + 547, + 657 + ], + "type": "text", + "content": "only hands is not sufficient to achieve satisfactory performance on the GRAB dataset. However, introducing more contextual information helps refine these results, unavailing information on other body parts." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "content": "Points saliency. As further evidence that object interaction involves different body parts, we conducted a study to discover what input points are crucial for the network." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "4732" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 150, + 70, + 443, + 127 + ], + "blocks": [ + { + "bbox": [ + 150, + 70, + 443, + 127 + ], + "lines": [ + { + "bbox": [ + 150, + 70, + 443, + 127 + ], + "spans": [ + { + "bbox": [ + 150, + 70, + 443, + 127 + ], + "type": "table", + "html": "
MethodGRABBEHAVE
Ec↓Ech↓Acc.↑Ec↓Ech↓Acc.↑
NN, SMPLH0.04040.193815.420.08800.387321.48
Ours, SMPLH0.02900.138720.800.07280.304624.16
Ours, SMPLH + T0.02630.138430.150.07220.306949.28
", + "image_path": "6c0bed31d233ed9d7fb2b05d4cf60f882659a2a69509444f956284e6d6ef5e57.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 135, + 547, + 159 + ], + "lines": [ + { + "bbox": [ + 46, + 135, + 547, + 159 + ], + "spans": [ + { + "bbox": [ + 46, + 135, + 547, + 159 + ], + "type": "text", + "content": "Table 3. Object classification. Quantitative results of models with object class prediction. Introducing temporal information has a dramatic impact on object classification." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 75, + 169, + 252, + 327 + ], + "blocks": [ + { + "bbox": [ + 75, + 169, + 252, + 327 + ], + "lines": [ + { + "bbox": [ + 75, + 169, + 252, + 327 + ], + "spans": [ + { + "bbox": [ + 75, + 169, + 252, + 327 + ], + "type": "image", + "image_path": "6a16398cfdb9213cc50694e80b3efab326b67572229f4118376c82855fbd64d8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 335, + 288, + 380 + ], + "lines": [ + { + "bbox": [ + 46, + 335, + 288, + 380 + ], + "spans": [ + { + "bbox": [ + 46, + 335, + 288, + 380 + ], + "type": "text", + "content": "Figure 9. Confusion Matrix. The confusion matrix for a subset of the predicted classes. Objects with similar functionality are often confused (Coffemug vs Cup), while ones that suggest characteristic human poses (Dorknoob, Hammer) are well separated." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 386, + 287, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 386, + 287, + 554 + ], + "spans": [ + { + "bbox": [ + 46, + 386, + 287, + 554 + ], + "type": "text", + "content": "We follow a recent protocol to find 3D point cloud saliency [72]: we cast the input through the network, we compute a loss on the output (in our case, the one in Eq. (2)), and we modify the input using the backpropagated gradient. This procedure is performed iteratively, and we refer to [72] for the details. We report the results of this study in Fig. 8, where points modified by the procedure are highlighted in red. We find the results of this analysis fascinating. As expected, the contact region is always essential to infer the correct object location. However, feet play a crucial role in all the reported cases, since they provide information about the human position and the consequent pose of the body. Another highlighted region is the head: different orientations give clues about object location and body posture." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 570, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 715 + ], + "type": "text", + "content": "Confusion in classification. As a final analysis, we explore the learning object classification during the training. We jointly train a further MLP module that takes as an input " + }, + { + "bbox": [ + 46, + 570, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_o" + }, + { + "bbox": [ + 46, + 570, + 287, + 715 + ], + "type": "text", + "content": " and predicts the object class, using a cross-entropy loss. When a time sequence of point clouds is available, we exploit it by selecting the class with the highest score across the frames and applying it to the whole sequence. We report results in Tab. 3. Our experiment suggests that the task is challenging while, given the number of classes (40), we still consider our results a promising first step. Also, we notice that temporal smoothing significantly helps classification accuracy. Temporal context disambiguate poses" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 179, + 545, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 179, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 305, + 179, + 545, + 228 + ], + "type": "text", + "content": "without a clear functionality. In Fig. 9, we report the confusion matrix for a subset of the classes for our method. The misclassification mainly arises from interactions of objects with similar functionality." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 243, + 383, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 243, + 383, + 255 + ], + "spans": [ + { + "bbox": [ + 306, + 243, + 383, + 255 + ], + "type": "text", + "content": "5. Conclusions" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 264, + 545, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 264, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 305, + 264, + 545, + 396 + ], + "type": "text", + "content": "In this work, we have addressed a novel and inspiring problem that changes the perspective on object-human interaction. Our proposed model is simple, carefully designed, and inspired by behavioural studies. We collected evidence of the method's effectiveness on a large set of object classes and empirically proved its generalization on noisy and different inputs. Finally, our analysis of human affordance is unprecedented, showing that human-object interaction can also involve body parts distant from the object and pointing to interesting relations useful for applications and subsequent works." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 419, + 545, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 419, + 545, + 563 + ], + "spans": [ + { + "bbox": [ + 304, + 419, + 545, + 563 + ], + "type": "text", + "content": "Limitations and Future Works. As the first exploration in this direction, our study enables several future possibilities. In this work, the temporal information is only used after the training procedure. Incorporating this information can create other patterns, further improving the results' quality. Our empirical evidence suggests that class prediction requires further investigation and more sophisticated techniques, like specialized attention mechanisms. Finally, we do not consider sequences that involve long (e.g., hours-long) and complex (e.g., multi-objects) interactions, which are difficult to capture. We hope our work can foster the community to collect such datasets." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgements Special thanks to the RVH team and reviewers, their feedback helped improve the manuscript. This work is funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) - 409792180 (EmmyNoether Programme, project: Real Virtual Humans) and the German Federal Ministry of Education and Research (BMBF): Tübingen AI Center, FKZ: 01IS18039A. G. Pons-Moll is a member of the Machine Learning Cluster of Excellence, EXC number 2064/1 - Project number 390727645. The authors thank the International Max Planck Research School for Intelligent Systems (IMPRS-IS) for supporting I. Petrov. R. Marin is supported by an Alexander von Humboldt Foundation Research Fellowship." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4733" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 156 + ], + "type": "text", + "content": "[1] Bharat Lal Bhatnagar, Xianghui Xie, Ilya A. Petrov, Cristian Sminchisescu, Christian Theobalt, and Gerard Pons-Moll. BEHAV: Dataset and Method for Tracking Human Object Interactions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15935-15946, 2022. 3, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 158, + 288, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 158, + 288, + 190 + ], + "spans": [ + { + "bbox": [ + 53, + 158, + 288, + 190 + ], + "type": "text", + "content": "[2] Amy E Booth and Sandra Waxman. Object names and object functions serve as cues to categories for infants. Developmental psychology, 38(6):948, 2002. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 192, + 288, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 192, + 288, + 246 + ], + "spans": [ + { + "bbox": [ + 53, + 192, + 288, + 246 + ], + "type": "text", + "content": "[3] Samarth Brahmbhatt, Cusuh Ham, Charles C Kemp, and James Hays. Contactdb: Analyzing and predicting grasp contact via thermal imaging. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8709-8719, 2019. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 247, + 288, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 247, + 288, + 300 + ], + "spans": [ + { + "bbox": [ + 53, + 247, + 288, + 300 + ], + "type": "text", + "content": "[4] Samarth Brahmbhatt, Ankur Handa, James Hays, and Dieter Fox. Contactgrasp: Functional multi-finger grasp synthesis from contact. In 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 2386-2393. IEEE, 2019. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 302, + 288, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 302, + 288, + 357 + ], + "spans": [ + { + "bbox": [ + 53, + 302, + 288, + 357 + ], + "type": "text", + "content": "[5] Samarth Brahmbhatt, Chengcheng Tang, Christopher D Twigg, Charles C Kemp, and James Hays. Contactpose: A dataset of grasps with object contact and hand pose. In European Conference on Computer Vision, pages 361-378. Springer, 2020. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 358, + 288, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 358, + 288, + 391 + ], + "spans": [ + { + "bbox": [ + 53, + 358, + 288, + 391 + ], + "type": "text", + "content": "[6] Tim Brooks and Alexei A Efros. Hallucinating pose-compatible scenes. In European Conference on Computer Vision, pages 510-528. Springer, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 392, + 288, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 392, + 288, + 435 + ], + "spans": [ + { + "bbox": [ + 53, + 392, + 288, + 435 + ], + "type": "text", + "content": "[7] Zhe Cao, Hang Gao, Karttikeya Mangalam, Qi-Zhi Cai, Minh Vo, and Jitendra Malik. Long-term human motion prediction with scene context. In European Conference on Computer Vision, pages 387-404. Springer, 2020. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 437, + 288, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 437, + 288, + 480 + ], + "spans": [ + { + "bbox": [ + 53, + 437, + 288, + 480 + ], + "type": "text", + "content": "[8] Zhe Cao, Ilija Radosavovic, Angjoo Kanazawa, and Jitendra Malik. Reconstructing Hand-Object Interactions in the Wild. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12417–12426, 2021. 1, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 481, + 288, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 481, + 288, + 523 + ], + "spans": [ + { + "bbox": [ + 53, + 481, + 288, + 523 + ], + "type": "text", + "content": "[9] Laura A Carlson-Radvansky, Eric S Covey, and Kathleen M Lattanzi. “what” effects on “where”: Functional influences on spatial relations. *Psychological Science*, 10(6):516–521, 1999. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 525, + 287, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 287, + 591 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 287, + 591 + ], + "type": "text", + "content": "[10] Yu-Wei Chao, Wei Yang, Yu Xiang, Pavlo Molchanov, Ankur Handa, Jonathan Tremblay, Yashraj S Narang, Karl Van Wyk, Umar Iqbal, Stan Birchfield, et al. Dexycb: A benchmark for capturing hand grasping of objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9044–9053, 2021. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 592, + 287, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 287, + 645 + ], + "type": "text", + "content": "[11] R. Qi Charles, Hao Su, Mo Kaichun, and Leonidas J. Guibas. PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 77-85, July 2017. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 647, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 713 + ], + "type": "text", + "content": "[12] Yixin Chen, Siyuan Huang, Tao Yuan, Siyuan Qi, Yixin Zhu, and Song-Chun Zhu. Holistic++ scene understanding: Single-view 3d holistic scene parsing and human pose estimation with human-object interaction and physical commonsense. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8648-8657, 2019. 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[13] Sammy Christen, Muhammed Kocabas, Emre Aksan, Jemin Hwangbo, Jie Song, and Otmar Hilliges. D-Grasp: Physically Plausible Dynamic Grasp Synthesis for Hand-Object Interactions. arXiv:2112.03028 [cs], Mar. 2022. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 161 + ], + "type": "text", + "content": "[14] Enric Corona, Gerard Pons-Moll, Guillem Alenyà, and Francesc Moreno-Noguer. Learned vertex descent: A new direction for 3d human model fitting. arXiv preprint arXiv:2205.06254, 2022. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 163, + 545, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 218 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 218 + ], + "type": "text", + "content": "[15] Enric Corona, Albert Pumarola, Guillem Alenya, Francesc Moreno-Noguer, and Grégory Rogez. Ganhand: Predicting human grasp affordances in multi-object scenes. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5031-5041, 2020. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 220, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 545, + 274 + ], + "type": "text", + "content": "[16] Rishabh Dabral, Soshi Shimada, Arjun Jain, Christian Theobalt, and Vladislav Golyanik. Gravity-Aware Monocular 3D Human-Object Reconstruction. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 12345–12354, Oct. 2021. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 276, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 276, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 307, + 276, + 545, + 330 + ], + "type": "text", + "content": "[17] Shengheng Deng, Xun Xu, Chaozheng Wu, Ke Chen, and Kui Jia. 3d affordancenet: A benchmark for visual object affordance understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1778-1787, 2021. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 332, + 545, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 332, + 545, + 386 + ], + "spans": [ + { + "bbox": [ + 307, + 332, + 545, + 386 + ], + "type": "text", + "content": "[18] Kiana Ehsani, Shubham Tulsiani, Saurabh Gupta, Ali Farhadi, and Abhinav Gupta. Use the force, luke! learning to predict physical forces by simulating effects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 224-233, 2020. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 388, + 545, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 545, + 431 + ], + "type": "text", + "content": "[19] Eleanor J Gibson. The concept of affordances in development: The renascence of functionalism. In *The concept of development: The Minnesota symposia on child psychology*. Vol, volume 15, pages 55–81, 1982. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 434, + 545, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 434, + 545, + 454 + ], + "spans": [ + { + "bbox": [ + 307, + 434, + 545, + 454 + ], + "type": "text", + "content": "[20] John C Gower. Generalized procrustes analysis. Psychometrika, 40(1):33-51, 1975. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 456, + 545, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 456, + 545, + 488 + ], + "spans": [ + { + "bbox": [ + 307, + 456, + 545, + 488 + ], + "type": "text", + "content": "[21] Helmu t Grabner, Juergen Gall, and Luc Van Gool. What makes a chair a chair? In CVPR 2011, pages 1529-1536, Colorado Springs, CO, USA, June 2011. IEEE. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 491, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 491, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 307, + 491, + 545, + 544 + ], + "type": "text", + "content": "[22] Patrick Grady, Chengcheng Tang, Christopher D. Twigg, Minh Vo, Samarth Brahmbhatt, and Charles C. Kemp. ContactOpt: Optimizing Contact to Improve Grasps. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1471-1481, June 2021. 1, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 547, + 545, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 547, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 307, + 547, + 545, + 612 + ], + "type": "text", + "content": "[23] Vladimir Guzov, Aymen Mir, Torsten Sattler, and Gerard Pons-Moll. Human poseitioning system (hps): 3d human pose estimation and self-localization in large scenes from body-mounted sensors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4318-4329, 2021. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 614, + 545, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 614, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 307, + 614, + 545, + 645 + ], + "type": "text", + "content": "[24] Vladimir Guzov, Torsten Sattler, and Gerard Pons-Moll. Visually plausible human-object interaction capture from wearable sensors. In arXiv, 2022. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 647, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 712 + ], + "type": "text", + "content": "[25] Shreyas Hampali, Sayan Deb Sarkar, Mahdi Rad, and Vincent Lepetit. Keypoint transformer: Solving joint identification in challenging hands and object interactions for accurate 3d pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11090-11100, 2022. 3" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "4734" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 106 + ], + "type": "text", + "content": "[26] Gregor Harih and Bojan Dolsak. Tool-handle design based on a digital human hand model. International Journal of Industrial Ergonomics, 43(4):288-295, July 2013. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 107, + 288, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 107, + 288, + 163 + ], + "spans": [ + { + "bbox": [ + 49, + 107, + 288, + 163 + ], + "type": "text", + "content": "[27] Mohamed Hassan, Duygu Ceylan, Ruben Villegas, Jun Saito, Jimei Yang, Yi Zhou, and Michael J Black. Stochastic scene-aware motion prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11374-11384, 2021. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 164, + 288, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 164, + 288, + 217 + ], + "spans": [ + { + "bbox": [ + 49, + 164, + 288, + 217 + ], + "type": "text", + "content": "[28] Mohamed Hassan, Vasileios Choutas, Dimitrios Tzionas, and Michael J Black. Resolving 3d human pose ambiguities with 3d scene constraints. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2282-2292, 2019. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 220, + 288, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 220, + 288, + 263 + ], + "spans": [ + { + "bbox": [ + 49, + 220, + 288, + 263 + ], + "type": "text", + "content": "[29] Yana Hasson, Gül Varol, Cordelia Schmid, and Ivan Laptev. Towards unconstrained joint hand-object reconstruction from rgb videos. In 2021 International Conference on 3D Vision (3DV), pages 659-668. IEEE, 2021. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 265, + 288, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 265, + 288, + 330 + ], + "spans": [ + { + "bbox": [ + 49, + 265, + 288, + 330 + ], + "type": "text", + "content": "[30] Yana Hasson, Gul Varol, Dimitrios Tzionas, Igor Kalevatykh, Michael J Black, Ivan Laptev, and Cordelia Schmid. Learning joint reconstruction of hands and manipulated objects. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11807-11816, 2019. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 332, + 288, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 332, + 288, + 398 + ], + "spans": [ + { + "bbox": [ + 48, + 332, + 288, + 398 + ], + "type": "text", + "content": "[31] Chun-Hao P. Huang, Hongwei Yi, Markus Höschle, Matvey Safroshkin, Tsvetelina Alexiadis, Senya Polikovsky, Daniel Scharstein, and Michael J. Black. Capturing and inferring dense full-body human-scene contact. In IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), pages 13274–13285, June 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 399, + 288, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 288, + 453 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 288, + 453 + ], + "type": "text", + "content": "[32] Yinghao Huang, Omid Taheri, Michael J Black, and Dimitrios Tzionas. Intercap: Joint markerless 3d tracking of humans and objects in interaction. In DAGM German Conference on Pattern Recognition, pages 281-299. Springer, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 456, + 288, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 288, + 521 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 288, + 521 + ], + "type": "text", + "content": "[33] Yuheng Jiang, Suyi Jiang, Guoxing Sun, Zhuo Su, Kaiwen Guo, Minye Wu, Jingyi Yu, and Lan Xu. Neuralhofusion: Neural volumetric rendering under human-object interactions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6155-6165, 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 523, + 288, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 523, + 288, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 523, + 288, + 578 + ], + "type": "text", + "content": "[34] Korrawe Karunratanakul, Jinlong Yang, Yan Zhang, Michael J. Black, Krikamol Muandet, and Siyu Tang. Grasping Field: Learning Implicit Representations for Human Grasps. In 2020 International Conference on 3D Vision (3DV), pages 333-344, Nov. 2020. 1, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 579, + 288, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 579, + 288, + 623 + ], + "spans": [ + { + "bbox": [ + 48, + 579, + 288, + 623 + ], + "type": "text", + "content": "[35] Hedvig Kjellström, Javier Romero, and Danica Kragić. Visual object-action recognition: Inferring object affordances from human demonstration. Computer Vision and Image Understanding, 115(1):81-90, Jan. 2011. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 624, + 288, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 288, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 288, + 678 + ], + "type": "text", + "content": "[36] Taein Kwon, Bugra Tekin, Jan Stuhmer, Federica Bogo, and Marc Pollefeys. H2o: Two hands manipulating objects for first person interaction recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10138-10148, 2021. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 681, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 288, + 712 + ], + "type": "text", + "content": "[37] Steven M Lehar. The world in your head: A gestalt view of the mechanism of conscious experience. Psychology Press, 2003. 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[38] Zhi Li, Soshi Shimada, Bernt Schiele, Christian Theobalt, and Vladislav Golyanik. Mocapdeform: Monocular 3d human motion capture in deformable scenes. In International Conference on 3D Vision (3DV), 2022. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 118, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 183 + ], + "type": "text", + "content": "[39] Yunze Liu, Yun Liu, Che Jiang, Kangbo Lyu, Weikang Wan, Hao Shen, Boqiang Liang, Zhoujie Fu, He Wang, and Li Yi. Hoi4d: A 4d egocentric dataset for category-level human-object interaction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21013-21022, 2022. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 185, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 185, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 185, + 545, + 228 + ], + "type": "text", + "content": "[40] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multiperson linear model. ACM transactions on graphics (TOG), 34(6):1-16, 2015. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 231, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 231, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 231, + 545, + 274 + ], + "type": "text", + "content": "[41] Priyanka Mandikal and Kristen Grauman. Learning dexterous grasping with object-centric visual affordances. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 6169-6176. IEEE, 2021. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 276, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 276, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 276, + 545, + 319 + ], + "type": "text", + "content": "[42] Yinyu Nie, Angela Dai, Xiaoguang Han, and Matthias Nießner. Pose2room: understanding 3d scenes from human activities. In European Conference on Computer Vision, pages 425-443. Springer, 2022. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 321, + 545, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 321, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 307, + 321, + 545, + 364 + ], + "type": "text", + "content": "[43] Lisa M Oakes and Kelly L Madole. Function revisited: How infants construe functional features in their representation of objects. In Advances in child development and behavior, volume 36, pages 135-185. Elsevier, 2008. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 366, + 545, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 366, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 307, + 366, + 545, + 431 + ], + "type": "text", + "content": "[44] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed AA Osman, Dimitrios Tzionas, and Michael J Black. Expressive body capture: 3d hands, face, and body from a single image. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10975-10985, 2019. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 434, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 434, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 307, + 434, + 545, + 487 + ], + "type": "text", + "content": "[45] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. PointNet++: Deep Hierarchical Feature Learning on Point Sets in a Metric Space. In Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc., 2017. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 489, + 545, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 489, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 307, + 489, + 545, + 521 + ], + "type": "text", + "content": "[46] Daniel Roetenberg, Henk Luinge, and Per Slycke. Moven: Full 6 dof human motion tracking using miniature inertial sensors. Xsen Technologies, December, 2007. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 523, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 523, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 307, + 523, + 545, + 567 + ], + "type": "text", + "content": "[47] Javier Romero, Dimitrios Tzionas, and Michael J. Black. Embodied hands: Modeling and capturing hands and bodies together. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia), 36(6), Nov. 2017. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 569, + 545, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 569, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 307, + 569, + 545, + 612 + ], + "type": "text", + "content": "[48] Manolis Savva, Angel X Chang, Pat Hanrahan, Matthew Fisher, and Matthias Nießner. Pigraphs: learning interaction snapshots from observations. ACM Transactions on Graphics (TOG), 35(4):1-12, 2016. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 614, + 545, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 614, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 307, + 614, + 545, + 656 + ], + "type": "text", + "content": "[49] Colin G. Scanes. Chapter 1 - animal perception including differences with humans. In Colin G. Scanes and Samia R. Toukhsati, editors, Animals and Human Society, pages 1-11. Academic Press, 2018. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 545, + 712 + ], + "type": "text", + "content": "[50] Zhuo Su, Lan Xu, Dawei Zhong, Zhong Li, Fan Deng, Shuxue Quan, and Lu Fang. Robustfusion: Robust volumetric performance reconstruction under human-object interactions from monocular rgbd stream. arXiv preprint arXiv:2104.14837, 2021. 3" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "4735" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 137 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 137 + ], + "type": "text", + "content": "[51] Guoxing Sun, Xin Chen, Yizhang Chen, Anqi Pang, Pei Lin, Yuheng Jiang, Lan Xu, Jingyi Yu, and Jingya Wang. Neural free-viewpoint performance rendering under complex human-object interactions. In Proceedings of the 29th ACM International Conference on Multimedia, pages 4651-4660, 2021. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 140, + 287, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 287, + 184 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 287, + 184 + ], + "type": "text", + "content": "[52] Zhongda Sun, Minglu Zhu, Xuechuan Shan, and Chengkuo Lee. Augmented tactile-perception and haptic-feedback rings as human-machine interfaces aiming for immersive interactions. Nature communications, 13(1):1-13, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 185, + 287, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 185, + 287, + 229 + ], + "spans": [ + { + "bbox": [ + 48, + 185, + 287, + 229 + ], + "type": "text", + "content": "[53] Subramanian Sundaram, Petr Kellnhofer, Yunzhu Li, Jun-Yan Zhu, Antonio Torralba, and Wojciech Matusik. Learning the signatures of the human grasp using a scalable tactile glove. Nature, 569(7758):698-702, May 2019. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 230, + 287, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 230, + 287, + 284 + ], + "spans": [ + { + "bbox": [ + 48, + 230, + 287, + 284 + ], + "type": "text", + "content": "[54] Omid Taheri, Vasileios Choutas, Michael J. Black, and Dimitrios Tzionas. GOAL: Generating 4D whole-body motion for hand-object grasping. In IEEE/CVF Conf. on Computer Vision and Pattern Recognition (CVPR), pages 13263-13273, June 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 286, + 287, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 286, + 287, + 352 + ], + "spans": [ + { + "bbox": [ + 48, + 286, + 287, + 352 + ], + "type": "text", + "content": "[55] Omid Taheri, Nima Ghorbani, Michael J. Black, and Dimitrios Tzionas. GRAB: A Dataset of Whole-Body Human Grasping of Objects. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, Computer Vision - ECCV 2020, volume 12349, pages 581-600. Springer International Publishing, Cham, 2020. 3, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 354, + 287, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 354, + 287, + 408 + ], + "spans": [ + { + "bbox": [ + 48, + 354, + 287, + 408 + ], + "type": "text", + "content": "[56] Weilin Wan, Lei Yang, Lingjie Liu, Zhuoying Zhang, Ruixing Jia, Yi-King Choi, Jia Pan, Christian Theobalt, Taku Komura, and Wenping Wang. Learn to predict how humans manipulate large-sized objects from interactive motions. IEEE Robotics and Automation Letters, 7(2):4702-4709, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 410, + 287, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 410, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 48, + 410, + 287, + 453 + ], + "type": "text", + "content": "[57] Zhenzhen Weng and Serena Yeung. Holistic 3d human and scene mesh estimation from single view images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 334-343, 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 455, + 287, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 455, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 48, + 455, + 287, + 498 + ], + "type": "text", + "content": "[58] Yan Wu, Jiahao Wang, Yan Zhang, Siwei Zhang, Otmar Hilliges, Fisher Yu, and Siyu Tang. Saga: Stochastic whole-body grasping with contact. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 500, + 287, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 500, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 48, + 500, + 287, + 544 + ], + "type": "text", + "content": "[59] Xianghui Xie, Bharat Lal Bhatnagar, and Gerard Pons-Moll. Chore: Contact, human and object reconstruction from a single rgb image. In European Conference on Computer Vision (ECCV). Springer, October 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 545, + 287, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 545, + 287, + 588 + ], + "spans": [ + { + "bbox": [ + 48, + 545, + 287, + 588 + ], + "type": "text", + "content": "[60] Xianghui Xie, Bharat Lal Bhatnagar, and Gerard Pons-Moll. Visibility aware human-object interaction tracking from single rgb camera. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 590, + 287, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 644 + ], + "type": "text", + "content": "[61] Lixin Yang, Xinyu Zhan, Kailin Li, Wenqiang Xu, Jiefeng Li, and Cewu Lu. Cpf: Learning a contact potential field to model the hand-object interaction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11097-11106, 2021. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "text", + "content": "[62] Bangpeng Yao, Jiayuan Ma, and Li Fei-Fei. Discovering Object Functionality. In Proceedings of the IEEE International Conference on Computer Vision, pages 2512-2519, 2013. 1, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 714 + ], + "type": "text", + "content": "[63] Yufei Ye, Abhinav Gupta, and Shubham Tulsiani. What's in your hands? 3d reconstruction of generic objects in hands." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 631 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 105 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 105 + ], + "type": "text", + "content": "In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3895-3905, 2022. 1, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 107, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 107, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 308, + 107, + 545, + 173 + ], + "type": "text", + "content": "[64] Hongwei Yi, Chun-Hao P. Huang, Dimitrios Tzionas, Muhammed Kocabas, Mohamed Hassan, Siyu Tang, Justus Thies, and Michael J. Black. Human-Aware Object Placement for Visual Environment Reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3959-3970, 2022. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 175, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 175, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 308, + 175, + 545, + 239 + ], + "type": "text", + "content": "[65] Andrei Zanfir, Elisabella Maroiniu, and Cristian Sminchisescu. Monocular 3d pose and shape estimation of multiple people in natural scenes-the importance of multiple scene constraints. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2148-2157, 2018. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 240, + 545, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 240, + 545, + 284 + ], + "spans": [ + { + "bbox": [ + 308, + 240, + 545, + 284 + ], + "type": "text", + "content": "[66] Jason Y Zhang, Panna Felsen, Angjoo Kanazawa, and Jitendra Malik. Predicting 3d human dynamics from video. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7114-7123, 2019. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 285, + 545, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 285, + 545, + 340 + ], + "spans": [ + { + "bbox": [ + 308, + 285, + 545, + 340 + ], + "type": "text", + "content": "[67] Jason Y Zhang, Sam Pepose, Hanbyul Joo, Deva Ramanan, Jitendra Malik, and Angjoo Kanazawa. Perceiving 3d human-object spatial arrangements from a single image in the wild. In European conference on computer vision, pages 34-51. Springer, 2020. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 342, + 545, + 394 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 342, + 545, + 394 + ], + "spans": [ + { + "bbox": [ + 308, + 342, + 545, + 394 + ], + "type": "text", + "content": "[68] Siwei Zhang, Qianli Ma, Yan Zhang, Zhiyin Qian, Marc Pollefeys, Federica Bogo, and Siyu Tang. Egobody: Human body shape, motion and social interactions from head-mounted devices. arXiv preprint arXiv:2112.07642, 2021. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 397, + 545, + 451 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 397, + 545, + 451 + ], + "spans": [ + { + "bbox": [ + 308, + 397, + 545, + 451 + ], + "type": "text", + "content": "[69] Siwei Zhang, Yan Zhang, Federica Bogo, Marc Pollefeys, and Siyu Tang. Learning motion priors for 4d human body capture in 3d scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11343-11353, 2021. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 453, + 545, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 453, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 308, + 453, + 545, + 506 + ], + "type": "text", + "content": "[70] Xiaohan Zhang, Bharat Lal Bhatnagar, Sebastian Starke, Vladimir Guzov, and Gerard Pons-Moll. Couch: Towards controllable human-chair interactions. In European Conference on Computer Vision (ECCV). Springer, October 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 509, + 545, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 509, + 545, + 552 + ], + "spans": [ + { + "bbox": [ + 308, + 509, + 545, + 552 + ], + "type": "text", + "content": "[71] Qian Zheng, Weikai Wu, Hanting Pan, Niloy Mitra, Daniel Cohen-Or, and Hui Huang. Inferring object properties from human interaction and transferring them to new motions. Computational Visual Media, 7(3):375–392, Sept. 2021. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 554, + 545, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 554, + 545, + 597 + ], + "spans": [ + { + "bbox": [ + 308, + 554, + 545, + 597 + ], + "type": "text", + "content": "[72] Tianhang Zheng, Changyou Chen, Junsong Yuan, Bo Li, and Kui Ren. Pointcloud saliency maps. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1598-1606, 2019. 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 599, + 545, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 599, + 545, + 631 + ], + "spans": [ + { + "bbox": [ + 308, + 599, + 545, + 631 + ], + "type": "text", + "content": "[73] Keyang Zhou, Bharat Lal Bhatnagar, Jan Eric Lenssen, and Gerard Pons-Moll. TOCH: Spatio-Temporal Object-to-Hand Correspondence for Motion Refinement, July 2022. 1, 3" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "4736" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Object Pose Estimation With Statistical Guarantees_ Conformal Keypoint Detection and Geometric Uncertainty Propagation/a86330a1-e4d7-4fd5-a6ce-7c82335913b8_content_list.json b/2023/Object Pose Estimation With Statistical Guarantees_ Conformal Keypoint Detection and Geometric Uncertainty Propagation/a86330a1-e4d7-4fd5-a6ce-7c82335913b8_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..900287f533a9b6ed235842c9ae63a41dd7214e6d --- /dev/null +++ b/2023/Object Pose Estimation With Statistical Guarantees_ Conformal Keypoint Detection and Geometric Uncertainty Propagation/a86330a1-e4d7-4fd5-a6ce-7c82335913b8_content_list.json @@ -0,0 +1,1803 @@ +[ + { + "type": "text", + "text": "Object Pose Estimation with Statistical Guarantees: Conformal Keypoint Detection and Geometric Uncertainty Propagation", + "text_level": 1, + "bbox": [ + 120, + 66, + 841, + 111 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Heng Yang and Marco Pavone \nNVIDIA Research", + "bbox": [ + 362, + 118, + 606, + 152 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d1ae1164c492c11124d70d6c68b4c42cff9b335a3dc17c948e1bfc7c6ac8a89b.jpg", + "image_caption": [ + "Figure 1. Probabilistically correct object pose estimation. Given (a) an input image and heatmap detections of the object semantic keypoints, our method first conformalizes the heatmaps into (b) circular or elliptical prediction sets that guarantee probabilistic coverage of the groundtruth keypoints (e.g., $90\\%$ ). Our method then propagates the uncertainty in the keypoints to the object pose, forming (c) a Pose UncleRtainty SEt (PURSE) that contains the groundtruth pose with the same probability. We develop RANdom SAmple averaGing (RANSAG) to sample from PURSE and generate (d) an average pose, and apply semidefinite relaxation to compute (e) worst-case error bounds: the blue duck attains the worst rotation error w.r.t. the (average pose) yellow duck; the red duck attains the worst translation error. Code available: https://github.com/NVlabs/ConformalKeypoint." + ], + "image_footnote": [], + "bbox": [ + 81, + 160, + 890, + 315 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 431, + 313, + 446 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The two-stage object pose estimation paradigm first detects semantic keypoints on the image and then estimates the 6D pose by minimizing reprojection errors. Despite performing well on standard benchmarks, existing techniques offer no provable guarantees on the quality and uncertainty of the estimation. In this paper, we inject two fundamental changes, namely conformal keypoint detection and geometric uncertainty propagation, into the two-stage paradigm and propose the first pose estimator that endows an estimation with provable and computable worst-case error bounds. On one hand, conformal keypoint detection applies the statistical machinery of inductive conformal prediction to convert heuristic keypoint detections into circular or elliptical prediction sets that cover the groundtruth keypoints with a user-specified marginal probability (e.g., $90\\%$ ). Geometric uncertainty propagation, on the other, propagates the geometric constraints on the keypoints to the 6D object pose, leading to a Pose UnceRtainty SEt (PURSE) that guarantees coverage of the groundtruth pose with the same probability. The PURSE, however, is a nonconvex set that does not directly lead to estimated poses and uncertainties. Therefore, we develop RANdom SAmple averaGing (RANSAG) to compute an average pose and apply semidefinite relaxation to upper bound the worst-case errors between the average pose and the groundtruth. On the LineMOD Occlusion dataset we demonstrate: (i) the PURSE covers the groundtruth with valid probabilities; (ii) the worst-case error bounds provide correct uncertainty quantification; and (iii) the average pose achieves better or similar accuracy as representative methods based on sparse keypoints.", + "bbox": [ + 73, + 450, + 473, + 905 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 501, + 431, + 630, + 446 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Estimating object poses from images is a fundamental problem in computer vision and finds extensive applications in augmented reality [42], autonomous driving [80], robotic manipulation [60], and space robotics [19]. One of the most popular paradigms for object pose estimation is a two-stage pipeline [20, 71, 72, 79, 81, 85, 89, 101], where the first stage detects (semantic) keypoints of the objects on the image, and the second stage computes the object pose by solving an optimization known as Perspective- $n$ -Points (PnP) that minimizes reprojection errors of the detected keypoints.", + "bbox": [ + 496, + 459, + 892, + 611 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Safety-critical applications call for provably correct computer vision algorithms. Existing algorithms in the two-stage paradigm (reviewed in Section 2), however, provide few performance guarantees on the quality of the estimated poses, due to three challenges. (C1) It is difficult to ensure the detected keypoints (typically from neural networks) are close to the groundtruth keypoints. In practice, the first stage often outputs keypoints that are arbitrarily wrong, known as outliers. (C2) Robust estimation is employed in the second stage to reject outliers, leading to nonconvex optimizations. Fast heuristics such as RANSAC [26] are widely adopted to find an approximate solution but they cannot guarantee global optimality and often fail without notice. (C3) There is no provably correct uncertainty quantification of the estimation, notably, a formal worst-case error bound between the estimation and the groundtruth. Though recent work [98] proposed convex relaxations to certify global optimality of RANSAC and addressed (C2), it cannot ensure correct estimation as the optimal pose may be far away from", + "bbox": [ + 496, + 613, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "8947", + "bbox": [ + 482, + 946, + 514, + 958 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "the correct pose when the keypoints are unreliable.", + "bbox": [ + 76, + 90, + 413, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Contributions. We propose a two-stage object pose estimation framework with statistical guarantees, illustrated in Fig. 1. Given an input image, we assume a neural network [71] is available to generate heatmap predictions of the object keypoints (Fig. 1(a)). Our framework then proceeds in two stages, namely conformal keypoint detection (Section 4) and geometric uncertainty propagation (Section 5). We first apply the statistical machinery of inductive conformal prediction (introduced in Section 3), with nonconformity functions inspired by the design of residual functions in classical geometric vision [39], to conformalize the heatmaps into circular or elliptical prediction sets—one for each keypoint—that guarantee coverage of the groundtruth keypoints with a user-specified marginal probability (Fig. 1(b)). This provides a simple and general methodology to bound the keypoint prediction errors (i.e., addressing (C1)). Given the keypoint prediction sets, we reformulate the constraints (enforced by the prediction sets) on the keypoints as constraints on the object pose, leading to a Pose UnceRtainty SET (PURSE) that guarantees coverage of the groundtruth pose with the same probability. Fig. 1(c) plots the boundary of an example PURSE (roll, pitch, raw angles for the rotation, and Euclidean coordinates for the translation). The PURSE, however, is an abstract nonconvex set that does not directly admit estimated poses and uncertainty. Therefore, we develop RANdom SAmple averaGing (RANSAG) to compute an average pose (Fig. 1(d)) and employ semidefinite relaxations to upper bound the worst-case rotation and translation errors between the average pose and the groundtruth (Fig. 1(e)). This gives rise to the first kind of computable worst-case probabilistic error bounds for object pose estimation (i.e., addressing (C3)). Our PURSE methodology has connections to the framework of unknown-but-bounded noise estimation in control theory [63], with special provisions to derive the bounds in a statistically principled way and enable efficient computation.", + "bbox": [ + 75, + 109, + 472, + 654 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We test our framework on the LineMOD Occlusion (LM-O) dataset [11] to verify the correctness of the theory (Section 6). First, we empirically show that the PURSE indeed contains the groundtruth pose according to the user-specified probability. Second, we demonstrate the correctness of the worst-case error bounds: when the PURSE contains the groundtruth, our bounds are always larger than, and in many cases close to, the actual errors between the average pose and the groundtruth pose. Third, we benchmark the accuracy of the average pose (coming from RANSAG) with representative two-stage pipelines based on sparse keypoints (e.g., PVNet [72]) and show that the average pose achieves better or similar accuracy.", + "bbox": [ + 75, + 655, + 468, + 852 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Limitations. A drawback of our approach, and conformal prediction in general, is that the size of the prediction sets depends on the nonconformity function (whose design", + "bbox": [ + 76, + 854, + 468, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "can be an art) and may be conservative. Our experiments suggest the bounds are loose when the keypoint prediction sets are large (e.g., giving $180^{\\circ}$ rotation bound). We discuss challenges and opportunities in tightening the bounds.", + "bbox": [ + 496, + 90, + 890, + 152 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 498, + 165, + 640, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image-based object pose estimation. We categorize object pose estimation into two paradigms: single-stage and two-stage. The latter first detects 2D-3D correspondences and then estimates the object pose via solving a PnP problem, while the former produces poses without intermediate correspondences. (i) Single-stage. Early methods perform pose estimation via template matching [29, 32, 36]. Recently, deep learning-based approaches such as PoseNet [41] and PoseCNN [95] applied CNNs to directly regress poses. A major challenge of pose regression is the nonlinearity of 3D rotations, and motivated formulating regression as classification [84, 87, 90] or designing better rotation representations [47, 102]. It is also popular to predict multiple pose hypotheses followed by voting [55, 62, 86]. (ii) Two-stage. Early research used handcrafted features [50, 56, 78] to establish 2D-3D correspondences and focused on developing algorithms for solving PnP. Notable algorithms include the minimal solver P3P [27, 46] and variants of the nonminimal solver PnP [45, 51, 68, 97]. Outliers (i.e., wrong correspondences) motivated robust estimation based on RANSAC [26], graduated non-convexity [9, 10, 96], branch-and-bound [16, 38, 52], or semidefinite relaxations [98]. Unreliable correspondences soon became the bottleneck and learned correspondences have been predominant. Learned correspondences can be sparse or dense. Sparse methods define a handful of keypoints and predict locations of the keypoints via direct regression [74, 89], probabilistic heatmap [67, 71], or voting [72]. Dense methods [12, 34, 54, 70, 93, 101] regress for each object pixel the coordinates of its corresponding 3D point. Recent literature focuses on end-to-end training via differentiating PnP [13, 15, 20, 21, 37]. Both single-stage and two-stage methods perform well on standard benchmarks [35], but a crucial feature that is missing, especially when deploying computer vision algorithms in safety-critical applications, is that these methods do not provide provably correct uncertainty quantification and formal error bounds w.r.t. the groundtruth (for either the correspondences or the poses). In this paper, we provide rigorous guarantees by applying conformal prediction to an existing keypoint detection method (the heatmap [71]) and leveraging old and new techniques in computer vision to derive formal error bounds.", + "bbox": [ + 496, + 191, + 893, + 839 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Conformal prediction in computer vision. Conformal prediction [92] is a statistical machinery that offers provably correct finite-sample uncertainty quantification without assumptions on the data distribution or the prediction model", + "bbox": [ + 496, + 839, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "8948", + "bbox": [ + 482, + 946, + 514, + 958 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "(i.e., offering a set prediction, instead of a point prediction, that guarantees probabilistic coverage of the groundtruth). Inductive conformal prediction [69] is the most popular variant of conformal prediction because it does not require retraining of the prediction models [1, 2, 49]. Applying conformal prediction to computer vision, however, is still in its infancy. Existing works focus on image classification [1, 76], tumor segmentation [3, 8, 94], and bounding box detection [2, 23, 53], which are classification or low-dimensional regression problems. Inspired by these works, our unique contributions in this paper are: (i) we apply conformal prediction to keypoints detection, a high-dimensional regression problem; (ii) we design new nonconformity functions and discuss their connections with classical geometric vision; and (iii) we develop algorithms that propagate the uncertainty after conformal prediction to form prediction sets of 6D poses, which are nonlinear and nonconvex manifold objects.", + "bbox": [ + 75, + 90, + 472, + 362 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Performance guarantees. Pose estimation from 2D-2D, 2D-3D, and 3D-3D correspondences are foundational problems in computer vision textbooks [6, 31, 58, 88] and typically boil down to formulating and solving mathematical optimization problems. Benchmarking on simulated and real datasets has been a widely adopted standard for testing different formulations and solvers. However, empirical performance can be misleading without theoretical guarantees. A striking fact is that, though error analysis is an important topic in applied math [17, 24, 43] and control theory [61, 63, 82], there is very limited literature in computer vision that reason about worst-case estimation errors between the optimal solution and the groundtruth. A popular heuristic relies on the inverse of the Hessian at an optimal solution, which provides the Cramer-Rao lower bound on the covariance of the solution (for linear regression this coincides with the covariance) [88, Section B.6] and thus cannot upper bound the estimation errors. Recent works [18, 77, 100] derived error bounds for a few geometric vision problems. However, the bounds either depend on uncheckable assumptions and cannot be computed [77, 100], or build on machinery (e.g., sum-of-squares proof [5, 64]) that only applies to estimators based on moment relaxations [18], which are still computationally expensive in practice [99]. In this paper, we develop the first kind of efficiently computable error bounds that only require the assumption of exchangeability (which comes from conformal prediction). We justify this assumption on our test dataset and numerically show our bounds can be tight for a subset of the test problems.", + "bbox": [ + 75, + 363, + 472, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Inductive Conformal Prediction", + "text_level": 1, + "bbox": [ + 76, + 829, + 369, + 845 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a set $\\{z_i = (x_i, y_i)\\}_{i=1}^l$ with observation $x_i \\in \\mathcal{X}$ and label $y_i \\in \\mathcal{Y}$ such that each $z_i \\in \\mathcal{Z} := \\mathcal{X} \\times \\mathcal{Y}$ is drawn i.i.d. from an unknown distribution on $\\mathcal{Z}$ , inductive confor", + "bbox": [ + 75, + 854, + 470, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "mal prediction (ICP) provides a set prediction $F^{\\epsilon}(x) \\subseteq \\mathcal{V}$ , parameterized by an error rate $0 < \\epsilon < 1$ , such that given a new sample $z_{l + 1} = (x_{l + 1}, y_{l + 1})$ satisfying an exchangeability condition (elaborated in Theorem 1), we have", + "bbox": [ + 496, + 90, + 890, + 151 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {P} \\left[ y _ {l + 1} \\in F ^ {\\epsilon} \\left(x _ {l + 1}\\right) \\right] \\geq 1 - \\epsilon , \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 159, + 890, + 176 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "i.e., the prediction set $F^{\\epsilon}$ guarantees to contain the true label $y_{l + 1}$ with probability at least $1 - \\epsilon$ .", + "bbox": [ + 496, + 183, + 890, + 213 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Training. We start by dividing the dataset into a proper training set $\\{z_1, \\ldots, z_m\\}$ and a calibration set $\\{z_{m+1}, \\ldots, z_l\\}$ . We shorthand $n = l - m$ as the size of the calibration set. We learn a prediction function $f: \\mathcal{X} \\to \\tilde{\\mathcal{V}}$ from the proper training set using any architecture, which allows us to fully exploit the power of modern deep learning. The prediction space $\\tilde{\\mathcal{V}}$ can be the same as the label space $\\mathcal{V}$ , or can contain auxiliary information such as a heuristic notion of uncertainty (e.g., softmax scores in classification or a heatmap in the case of keypoint detection).", + "bbox": [ + 496, + 214, + 892, + 364 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Conformal calibration. We define a nonconformity function $S: \\mathcal{Z}^m \\times \\mathcal{Z} \\to \\mathbb{R}$ to measure how well a given sample $z = (x, y)$ conforms to the proper training set. A popular instance of $S$ leverages the learned prediction $f$ :", + "bbox": [ + 496, + 364, + 893, + 426 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nS \\left(\\left\\{z _ {1}, \\dots , z _ {m} \\right\\}, (x, y)\\right) \\stackrel {{e. g.}} {{=}} r (y, f (x)), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 545, + 434, + 890, + 452 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $r: \\mathcal{Y} \\times \\tilde{\\mathcal{Y}} \\to \\mathbb{R}$ is a measure of disagreement between the label $y$ and the prediction $f(x)$ . For example, consider $\\mathcal{Y} = \\tilde{\\mathcal{Y}} = \\mathbb{R}$ , one can design $r(y, f(x)) = |y - f(x)|$ : if $(x, y)$ poorly conforms to the training set, $f$ will incur large errors. While the function $S$ can be arbitrary (e.g., a learnable neural network [83]), (2) is a convenient definition since $f$ is implicitly dependent on $\\{z_i\\}_{i=1}^m$ and $r$ can incorporate domain-specific knowledge. We then compute the nonconformity scores on the calibration set as $\\alpha_i = r(y_i, f(x_i)), i = m+1,\\ldots,l$ , and sort them in nonincreasing order $\\alpha_{\\pi(1)} \\geq \\dots \\geq \\alpha_{\\pi(n)}$ , where $\\pi(i) \\in \\{m+1,\\ldots,l\\}$ is an index permutation.", + "bbox": [ + 496, + 458, + 893, + 641 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Conformal prediction. Given a new observation $x_{l+1}$ (with an unknown $y_{l+1}$ ) and a user-specified $\\epsilon \\in (0, 1)$ , we compute the inductive conformal prediction (ICP) set as", + "bbox": [ + 496, + 641, + 893, + 686 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nF ^ {\\epsilon} \\left(x _ {l + 1}\\right) = \\left\\{y \\in \\mathcal {Y} \\mid \\alpha^ {y} \\leq \\alpha_ {\\pi (\\lfloor (n + 1) \\epsilon \\rfloor)} \\right\\}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 693, + 890, + 712 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\alpha^y = r(y, f(x_{l+1}))$ is the nonconformity score of the new sample when fixing the true label to be $y$ . In other words, the ICP set (3) outputs the set of all labels that make the nonconformity score of the new sample no greater than $\\alpha_{\\pi(\\lfloor n + 1 \\epsilon \\rfloor)} - \\text{the } \\lfloor (n + 1) \\epsilon \\rfloor$ -th largest nonconformity score in the calibration set. We have the following result stating the probabilistic coverage of the ICP set (3).", + "bbox": [ + 496, + 718, + 893, + 825 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Theorem 1 (Validity of ICP Coverage [48, 91, 92]). If $z_{m+1}, \\ldots, z_l, z_{l+1} = (x_{l+1}, y_{l+1})$ are exchangeable, i.e., their distribution is invariant under permutation, then", + "bbox": [ + 496, + 830, + 893, + 877 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n1 - \\epsilon \\leq \\mathbb {P} \\left[ y _ {l + 1} \\in F ^ {\\epsilon} \\left(x _ {l + 1}\\right) \\right] \\leq 1 - \\epsilon + 1 / (n + 1) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 883, + 890, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "8949", + "bbox": [ + 480, + 946, + 514, + 958 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/79351fc3e0ba435b18162af2bc294190b15a6316fcc94715c40aed14afceb2b4.jpg", + "image_caption": [ + "Figure 2. Beta distribution of the conditional coverage in (5) with $\\epsilon = 0.1$ and different $n$ . Notice how the conditional probability becomes more concentrated around $1 - \\epsilon$ when $n$ increases." + ], + "image_footnote": [], + "bbox": [ + 158, + 75, + 390, + 224 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "for any $\\epsilon \\in (0,1)$ . Furthermore, when conditioned on the calibration set, calling $h = \\lfloor (n + 1)\\epsilon \\rfloor$ , we have", + "bbox": [ + 73, + 273, + 468, + 306 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {P} \\left[ y _ {l + 1} \\in F ^ {\\epsilon} (x _ {l + 1}) \\mid \\{z _ {m + 1}, \\dots , z _ {l} \\} \\right] \\sim \\mathrm {B e t a} (n + 1 - h, h). (5)\n$$\n", + "text_format": "latex", + "bbox": [ + 66, + 316, + 468, + 335 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A few remarks are in order about Theorem 1. First, asking $z_{m+1}, \\ldots, z_l, z_{l+1}$ to be exchangeable is weaker than asking them to be independent. However, this assumption typically fails when the calibration set is a single video sequence, where the image frames $\\{z_{m+1}, \\ldots, z_l\\}$ are temporally correlated [57]. Fortunately, as we detail in Section 6, the way the LineMOD Occlusion dataset [11] was collected makes the exchangeability condition easily satisfied, which also suggests best practices to make the exchangeability condition hold in computer vision. Second, the lower bound in (4) can be intuitively proved because under exchangeability, $\\alpha_{l+1} := r(y_{l+1}, f(x_{l+1}))$ – the nonconformity score of the new sample with the true label– is exchangeable with the nonconformity scores of the calibration samples, and hence equally likely to fall in anywhere between the scores $\\{\\alpha_{\\pi(i)}\\}_{i=1}^n$ . Consequently, $\\mathbb{P}[y_{l+1} \\in F^\\epsilon(x_{l+1})] = \\mathbb{P}\\left[\\alpha_{l+1} \\leq \\alpha_{\\pi(\\lfloor n+1\\epsilon \\rfloor)}\\right] = 1 - \\lfloor(n+1)\\epsilon \\rfloor / (n+1) \\geq 1 - \\epsilon$ . The upper bound in (4) states that $1 - \\epsilon$ is not overly conservative (indeed tight if $n$ is large). Lastly, the probabilistic guarantee in (4) is marginal over the randomness of the calibration set, meaning if one chooses an infinite number of calibration sets, the average empirical coverage will converge to $1 - \\epsilon$ . This, however, implies that the empirical coverage given one calibration set is a random variable that fluctuates as the Beta distribution (5). Fig. 2 plots the Beta distribution at $\\epsilon = 0.1$ with different sizes of the calibration set. We observe that as $n$ increases the empirical coverage becomes more concentrated at $1 - \\epsilon$ . Our experiments show that even with a small ( $n = 200$ ) calibration set, the empirical coverage is close to, and mostly higher than, $1 - \\epsilon$ .", + "bbox": [ + 75, + 345, + 472, + 803 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Conformal Keypoint Detection", + "text_level": 1, + "bbox": [ + 75, + 814, + 359, + 832 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we apply the ICP framework in Section 3 to the problem of semantic keypoint detection.", + "bbox": [ + 75, + 839, + 468, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Setup. Denote by $x \\in \\mathbb{R}^{H \\times W \\times 3}$ an RGB image picturing an object, by $\\pmb{y} = (y_1, \\dots, y_K) \\in \\mathbb{R}^2 \\times \\dots \\times \\mathbb{R}^2 \\coloneqq \\mathcal{Y}$ .", + "bbox": [ + 75, + 869, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the groundtruth locations of $K$ semantic keypoints of the object. We partition a given dataset $\\{z_i := (x_i, y_i)\\}_{i=1}^l$ into a proper training set (of size $m$ ) and a calibration set (of size $n$ ). We follow the three steps in Section 3 to perform ICP.", + "bbox": [ + 496, + 90, + 890, + 151 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Training. We choose the heatmap approach in [71, 79] as the prediction function: given an image $x$ , [79] outputs a set of heatmaps $\\pmb{f}(x) = (f(x)_1, \\dots, f(x)_K)$ , where each $f(x)_k \\in \\Delta^{HW} := \\{v \\in \\mathbb{R}_+^{HW} \\mid \\sum_i^{HW} v_i = 1\\}$ predicts the probability distribution of the $k$ -th keypoint lying on each pixel of the image. For convenience, we use $q^j \\in \\mathbb{R}^2$ to denote the $j$ -th pixel location in $x$ and $f(x)_k^j \\in \\mathbb{R}_+$ to denote the probability of the $k$ -th keypoint lying on $q^j$ . Let $\\sigma_k$ be the index permutation that sorts $f(x)_k$ in nonincreasing order, i.e., $f(x)_k^{\\sigma_k(1)} \\geq \\dots \\geq f(x)_k^{\\sigma_k(HW)}$ . As we will soon show, choosing the heatmap approach leads to simple and intuitive designs of the nonconformity function.", + "bbox": [ + 496, + 152, + 892, + 335 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Conformal calibration. We design the following nonconformity function", + "bbox": [ + 496, + 337, + 890, + 367 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nr (\\boldsymbol {y}, \\boldsymbol {f} (x)) = \\max \\left\\{\\phi \\left(y _ {k}, f (x) _ {k}\\right) \\right\\} _ {k = 1} ^ {K} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 377, + 890, + 397 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "that uses $\\phi$ to score each keypoint and then selects the maximum score. This design considers the worst keypoint detection performance of $f$ . We provide two designs of $\\phi$ below.", + "bbox": [ + 496, + 407, + 890, + 452 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "(a) Peak. Shorthand $p_k = f(x)_k^{\\sigma_k(1)}$ as the peak probability in the $k$ -th heatmap and $q_k = q^{\\sigma_k(1)}$ as the pixel location attaining the peak probability, we design", + "bbox": [ + 496, + 453, + 890, + 501 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\phi_ {\\text {p e a k}} \\left(y _ {k}, f (x) _ {k}\\right) = p _ {k} \\| y _ {k} - q _ {k} \\| \\quad (\\text {p e a k})\n$$\n", + "text_format": "latex", + "bbox": [ + 583, + 512, + 890, + 530 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "which computes the error between the true keypoint location $y_{k}$ and the most probable keypoint location $q_{k}$ and scales the error by the peak probability $p_{k}$ . $\\phi_{\\mathrm{peak}}$ describes nonconformity because it becomes larger when the network $f$ is confidently wrong (both $\\| y_{k} - q_{k}\\|$ and $p_k$ are large), implying the sample is highly nonconforming.", + "bbox": [ + 496, + 540, + 890, + 630 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "(b) Covariance. Let $\\bar{q}_k = \\sum_{j=1}^{J} f(x)_k^{\\sigma_k(j)} q^{\\sigma_k(j)}$ be the expected location of the top- $J$ most likely detections for the $k$ -th keypoint, and $\\Sigma_k = \\sum_{j=1}^{J} f(x)_k^{\\sigma_k(j)} \\cdot (q^{\\sigma_k(j)} - \\bar{q}_k)(q^{\\sigma_k(j)} - \\bar{q}_k)^{\\top}$ as the covariance, we design", + "bbox": [ + 496, + 631, + 890, + 700 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\phi_ {\\operatorname {c o v}} \\left(y _ {k}, f (x) _ {k}\\right) = \\left(y _ {k} - \\bar {q} _ {k}\\right) ^ {\\top} \\Sigma_ {k} ^ {- 1} \\left(y _ {k} - \\bar {q} _ {k}\\right) \\quad (\\operatorname {c o v})\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 709, + 890, + 729 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "which computes the squared Mahalanobis distance [59] from the groundtruth $y_{k}$ to the top- $J$ keypoint detections (represented by the mean $\\bar{q}_{k}$ and covariance $\\Sigma_{k}$ ).2 A larger Mahalanobis distance indicates more abnormality of the heatmap $f(x)_k$ (compared to the groundtruth $y_{k}$ ) [28], and hence implies higher nonconformity.", + "bbox": [ + 496, + 738, + 890, + 830 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "1The heatmap in the original paper [71] is not a valid probability distribution as it contains negative values and do not sum up to 1. We remove the negative values and normalize it to be a valid probability distribution.", + "bbox": [ + 496, + 838, + 890, + 875 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "2We only choose the top- $J$ $(J = 100)$ most likely detections on the heatmap because the heatmap can be quite noisy in practice.", + "bbox": [ + 500, + 875, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "8950", + "bbox": [ + 482, + 946, + 514, + 958 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Using the nonconformity function (6) with (peak) or (cov), we compute the nonconformity scores of the calibration set and sort them as: $\\alpha_{\\pi(1)} \\geq \\dots \\geq \\alpha_{\\pi(n)}$ .", + "bbox": [ + 76, + 90, + 468, + 137 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Conformal prediction. Given an error rate $\\epsilon \\in (0,1)$ , we first find $\\alpha_{\\pi (\\lfloor (n + 1)\\epsilon \\rfloor)}$ . Then, according to the ICP set definition (3) and our nonconformity function (6), we output the ICP set for a new $x_{l + 1}$ as", + "bbox": [ + 76, + 137, + 468, + 198 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} F ^ {\\epsilon} (x _ {l + 1}) \\\\ = \\left\\{\\boldsymbol {y} \\in \\mathcal {Y} \\mid \\max \\left\\{\\phi \\left(y _ {k}, f \\left(x _ {l + 1}\\right)\\right) \\right\\} _ {k = 1} ^ {K} \\leq \\alpha_ {\\pi (\\lfloor (n + 1) \\epsilon \\rfloor)} \\right\\} \\\\ = \\left\\{\\boldsymbol {y} \\in \\mathcal {Y} \\mid \\phi \\left(y _ {k}, f \\left(x _ {l + 1}\\right)\\right) \\leq \\alpha_ {\\pi (\\lfloor (n + 1) \\epsilon \\rfloor)}, \\forall k \\right\\}, \\tag {7} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 212, + 468, + 267 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where we used $\\max \\{\\phi_1,\\ldots ,\\phi_K\\} \\leq \\alpha$ if and only if $\\phi_{k}\\leq \\alpha$ for any $k$ . Insert (peak) into (7), we have $F_{\\mathrm{ball}}^{\\epsilon}(x_{l + 1})$ as", + "bbox": [ + 76, + 279, + 468, + 311 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{\\boldsymbol {y} \\in \\mathcal {Y} \\mid \\| y _ {k} - q _ {l + 1, k} \\| \\leq \\frac {\\alpha_ {\\pi (\\lfloor (n + 1) \\epsilon \\rfloor)}}{p _ {l + 1 , k}}, \\forall k \\right\\}, \\quad (\\text {b a l l})\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 323, + 468, + 357 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "which defines -for the $k$ -th keypoint- a ball centered at $q_{l + 1,k}$ (the most likely detection) with a radius inversely proportional to $p_{l + 1,k}$ and proportional to $\\alpha_{\\pi (\\lfloor (n + 1)\\epsilon \\rfloor)}$ . Similarly, insert (cov) into (7), we have $F_{\\mathrm{ellipse}}^{\\epsilon}(x_{l + 1})$ as", + "bbox": [ + 76, + 369, + 468, + 431 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{\\boldsymbol {y} \\in \\mathcal {Y} \\mid (y _ {k} - \\bar {q} _ {l + 1, k}) ^ {\\mathsf {T}} \\frac {\\sum_ {l + 1 , k} ^ {- 1}}{\\alpha_ {\\pi (\\lfloor (n + 1) \\epsilon \\rfloor)} (y _ {k} - \\bar {q} _ {l + 1 , k})} \\leq 1, \\forall k \\right\\},\n$$\n", + "text_format": "latex", + "bbox": [ + 60, + 444, + 478, + 496 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "which defines -for the $k$ -th keypoint-- an ellipse centered at $\\bar{q}_{l+1,k}$ (the expected location of the top- $J$ detections) with an area proportional to $\\operatorname*{det}(\\Sigma_{l+1,k})$ and $\\alpha_{\\pi(\\lfloor n+1\\epsilon \\rfloor)}$ .3 From (ball) and (ellipse), we observe that the prediction sets become larger when (i) the heatmaps are uncertain, i.e., the peak probability is low or the covariance matrix has large determinant; and (ii) the heatmaps perform poorly on the calibration set, leading to a large $\\alpha_{\\pi(\\lfloor n+1\\epsilon \\rfloor)}$ .", + "bbox": [ + 75, + 497, + 468, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Connections to geometric vision. Our nonconformity function bears similarity to the residual function in geometric vision [4, 22, 31]. For example, the (peak) and (cov) functions are similar to the (weighted) reprojection error [31], and the \"max\" in (6) can be connected to seminal work on optimizing the $\\ell_{\\infty}$ norm [39].", + "bbox": [ + 75, + 619, + 468, + 710 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Outlier-robust nonconformity? One potential issue of the nonconformity function (6) is that a single outlier can inflate the score and the calibration quantile $\\alpha_{\\pi (\\lfloor (n + 1)\\epsilon \\rfloor)}$ and lead to conservative prediction sets (e.g., when $\\pmb{f}$ predicts $K - 1$ keypoints perfectly but misses one keypoint). A potential remedy in geometric vision is to use robust cost functions [7,9,96]. Therefore, a natural question is whether \"robustifying\" the nonconformity function (6) can lead to better prediction sets. Here we focus on only robustifying $\\phi$ in (6) and provide a negative answer.", + "bbox": [ + 75, + 710, + 470, + 863 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Proposition 2 (Invariance of ICP). Let $\\rho : \\mathbb{R}_+ \\mapsto \\mathbb{R}_+$ be any monotonically increasing function. Fixing the calibration set and error rate $\\epsilon$ , the nonconformity function", + "bbox": [ + 500, + 90, + 890, + 137 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nr _ {\\rho} (\\boldsymbol {y}, \\boldsymbol {f} (x)) = \\max \\left\\{\\rho \\left(\\phi \\left(y _ {k}, f (x) _ {k}\\right)\\right) \\right\\} _ {k = 1} ^ {K} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 147, + 890, + 165 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "leads to the same ICP set as (6).", + "bbox": [ + 500, + 176, + 714, + 191 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The proof of Proposition 2 is presented in Supplementary Material. We conclude that common robust costs, such as $\\ell_1$ , Huber, Geman-McClure, and Barron's adaptive kernel [7,9] (which are monotonically increasing on $[0, +\\infty]$ ) cannot change the ICP sets by robustifying the individual score $\\phi$ . However, it remains an open question whether changing the \"max\" operation in (6) can give rise to better ICP sets. For instance, replacing \"max\" with \" $\\sum$ in (6)\" and using the Geman-McClure robust cost $\\rho(\\phi) = \\frac{\\phi^2}{1 + \\phi^2}$ with $\\phi = \\phi_{\\mathrm{peak}}$ results in the following ICP set", + "bbox": [ + 496, + 199, + 890, + 354 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{\\boldsymbol {y} \\in \\mathcal {Y} \\mid \\sum_ {k = 1} ^ {K} \\frac {p _ {k} ^ {2} \\| y _ {k} - q _ {k} \\| ^ {2}}{1 + p _ {k} ^ {2} \\| y _ {k} - q _ {k} \\| ^ {2}} \\leq \\alpha_ {\\pi (\\lfloor (n + 1) \\epsilon \\rfloor)} \\right\\} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 364, + 890, + 407 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "that does not admit a geometric interpretation that is as simple and intuitive as the (ball) and (ellipse) sets introduced before. In fact, it is indeed the simplicity of (ball) and (ellipse) that enables us to propagate the uncertainty in keypoints to the object pose, as we will show in the next section.", + "bbox": [ + 496, + 417, + 890, + 494 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Geometric Uncertainty Propagation", + "text_level": 1, + "bbox": [ + 500, + 506, + 823, + 523 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Conformalizing the heatmaps gives us prediction sets that guarantee probabilistic coverage of the true keypoints. We unify the prediction sets (ball) and (ellipse) as", + "bbox": [ + 496, + 531, + 890, + 578 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF ^ {\\epsilon} (x) = \\left\\{\\boldsymbol {y} \\in \\mathcal {Y} \\mid \\left(y _ {k} - \\mu_ {k}\\right) ^ {\\top} \\Lambda_ {k} \\left(y _ {k} - \\mu_ {k}\\right) \\leq 1, \\forall k \\right\\}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 587, + 898, + 606 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mu_{k} = q_{l + 1,k},\\Lambda_{k} = \\frac{p_{l + 1,k}^{2}}{\\alpha_{\\pi(\\lfloor(n + 1)\\epsilon \\rfloor)}^{2}}\\mathbf{I}_{2}$ for (ball), $\\mu_{k} =$ $\\bar{q}_{l + 1,k},\\Lambda_k = \\frac{\\sum_{l + 1,k}^{-1}}{\\alpha_\\pi(\\lfloor(n + 1)\\epsilon \\rfloor)}$ for (ellipse), and we omit the subscript $l + 1$ for simplicity.", + "bbox": [ + 496, + 616, + 890, + 681 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Why not uncertainty-aware PnP? A popular way to estimate pose from (10) is to solve an uncertainty-aware PnP", + "bbox": [ + 498, + 683, + 890, + 713 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {(R, t) \\in \\mathrm {S E} (3)} \\quad \\sum_ {k = 1} ^ {K} \\left(y _ {k} - \\mu_ {k}\\right) ^ {\\mathsf {T}} \\Lambda_ {k} \\left(y _ {k} - \\mu_ {k}\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 534, + 723, + 825, + 763 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\text {s u b j e c t} \\quad y _ {k} = \\Pi \\left(R Y _ {k} + t\\right), k = 1, \\dots , K \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 767, + 890, + 782 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $Y_{k} \\in \\mathbb{R}^{3}, k = 1, \\dots, K$ are the 3D object keypoints and $\\Pi(\\cdot)$ denotes the camera projection. We challenge this approach and point out its two drawbacks. First, it is difficult to solve (11) to global optimality due to (i) the nonconvex SE(3) constraint and (ii) the rational polynomial appearing in $\\Pi(\\cdot)$ . The best known approach to solve (11) relies on either branch-and-bound [68] or local optimization.", + "bbox": [ + 496, + 794, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "3The area of $(x - \\mu)^{\\mathsf{T}}A(x - \\mu)\\leq 1$ is proportional to $\\operatorname *{det}(A^{-1})$", + "bbox": [ + 93, + 885, + 455, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "8951", + "bbox": [ + 482, + 946, + 513, + 959 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Second, solving (11) typically outputs a single optimal pose without uncertainty quantification. Are there other poses that attain similar costs as the optimal pose? How close is the optimal pose to the groundtruth pose? These questions remain not answered in the literature.", + "bbox": [ + 75, + 90, + 468, + 165 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Pose UnceRtainty SET (PURSE). We propose to, instead of solving a PnP problem similar to (11), directly propagate the uncertainty in the ICP sets to the object pose.", + "bbox": [ + 75, + 167, + 468, + 213 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Proposition 3 (PURSE). Let $s_{\\mathrm{gt}} = [\\mathrm{vec}(R_{\\mathrm{gt}})^{\\top}; t_{\\mathrm{gt}}^{\\top}]^{\\top}$ be the groundtruth object pose (that lies in front of the camera). Then, the groundtruth keypoints $\\mathbf{y} = (y_1, \\ldots, y_K)$ belong to the ICP set $F^{\\epsilon}(x)$ in (10) if and only if $s_{\\mathrm{gt}}$ belongs to the following pose uncertainty set", + "bbox": [ + 73, + 223, + 468, + 301 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nS ^ {\\epsilon} = \\left\\{s \\in \\operatorname {S E} (3) \\left| \\begin{array}{c} s ^ {\\mathsf {T}} A _ {k} s \\leq 0, k = 1, \\ldots , K \\\\ b _ {k} ^ {\\mathsf {T}} s > 0, k = 1, \\ldots , K \\end{array} \\right. \\right\\}, (\\text {P U R S E})\n$$\n", + "text_format": "latex", + "bbox": [ + 73, + 311, + 468, + 347 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $A_{k}\\in \\mathbb{S}^{12},b_{k}\\in \\mathbb{R}^{12},k = 1,\\ldots ,K$ are constant matrices dependent on $\\mu_k,\\Lambda_k,Y_k$ and camera intrinsics.", + "bbox": [ + 75, + 358, + 468, + 390 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The detailed proof for Proposition 3 is algebraically involved and postponed to the Supplementary Material. The high-level intuition is, however, straightforward: we plug in $y_{k} = \\Pi (RY_{k} + t)$ into (10) and obtain $K$ quadratic inequalities of the form $s^{\\mathsf{T}}A_{k}s\\leq 0$ . The linear inequalities $b_{k}^{\\top}s > 0$ are added to enforce the (transformed) 3D keypoints lie in front of the camera. Proposition 3 implies, if we are $1 - \\epsilon$ confident the groundtruth keypoints can be anywhere inside $F^{\\epsilon}(x)$ , then we should also be confident any pose in (PURSE) can be the groundtruth. Viewing pose estimation as a set estimation with guaranteed probabilistic coverage of the groundtruth is fundamentally different from viewing it as computing a single pose from (11) that is (hopefully) close enough to the groundtruth.", + "bbox": [ + 75, + 401, + 468, + 612 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "RANdom SExample averaGing (RANSAG). Verifying if a given pose belongs to the PURSE is straightforward via checking the inequalities in (PURSE). However, the PURSE does not directly give us estimated poses. Therefore, we propose an efficient sampling algorithm called RANSDOM SExample averaGing (RANSAG) that is analogous to RANSAC [26] and leverages the minimal solver P3P [27], presented in Algorithm 1. The intuition is that, though it is difficult to sample directly in PURSE due to the (nonconvex) constraints, it is easy to sample from the keypoint prediction set (10) due to its simple geometry (balls and ellipses). Thus, at each iteration (line 3) RANSAG samples three keypoints (line 4-5), solves the P3P inverse problem, and accept the poses that belong to the (PURSE) (line 6). RANSAG typically returns around 100 valid samples with $T = 1000$ trials. However, in difficult cases (e.g., when $S^{\\epsilon}$ is small or even empty) it is possible to obtain zero samples ( $S = \\emptyset$ ). In this situation, RANSAG samples $\\lfloor T / 20 \\rfloor$ (default 50) poses without checking if they belong to the PURSE, via sampling", + "bbox": [ + 75, + 613, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1: RANdom SAmple averaGing" + ], + "code_body": "1 Input: an ICP set $F^{\\epsilon}(x)$ (10) and its corresponding (PURSE) $S^{\\epsilon}$ ; maximum trials $T$ ; initial $\\hat{S} = \\emptyset$ ; \n2 Output: sample poses $S \\subset \\mathrm{SE}(3)$ in PURSE, and an average pose $\\bar{s} \\in \\mathrm{SE}(3)$ ; \n3 for $\\tau \\gets 1$ to $T$ do \n4 Sample $\\{k_1, k_2, k_3\\}$ from $[K] (k_1 \\neq k_2 \\neq k_3)$ ; \n5 Sample $\\hat{y}_{k_i}, i = 1, 2, 3$ from $\\{y \\in \\mathbb{R}^2 \\mid (y - \\mu_{k_i})^\\top \\Lambda_{k_i} (y - \\mu_{k_i}) \\leq 1\\}$ ; \n6 $\\hat{S} \\gets \\hat{S} \\cup (S^{\\epsilon} \\cap \\mathsf{P3P}(\\{\\hat{y}_{k_i} \\leftrightarrow Y_{k_i}\\}_{i=1}^3))$ ; \n7 end \n8 $S = \\hat{S}$ ; \n9 if $\\hat{S} = \\emptyset$ then \n10 for $\\tau \\gets 1$ to $\\lfloor T/20 \\rfloor$ do \n11 Sample $\\hat{y}_k, k = 1, \\dots, K$ from $F^{\\epsilon}(x)$ ; \n12 $\\hat{S} \\gets \\hat{S} \\cup \\mathsf{PnP}(\\{\\hat{y}_k \\leftrightarrow Y_k\\}_{k=1}^K)$ ; \n13 end \n14 end \n15 $\\bar{R} = \\operatorname{proj}_{\\mathrm{SO}(3)}(\\sum_{(R_j, *) \\in \\hat{S}} R_j)$ ; \n16 $\\bar{t} = \\frac{1}{|\\hat{S}|} \\sum_{(*, t_j) \\in \\hat{S}} t_j$ ; \n17 return: $S, \\bar{s} = (\\bar{R}, \\bar{t})$", + "bbox": [ + 506, + 112, + 866, + 468 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$K$ keypoints and solving PnP (line 9-12). After obtaining a set of poses, RANSAG performs rotation averaging (line 14) and translation averaging (line 15) to obtain an average pose $\\bar{s}$ . Note that RANSAG does not check if $\\bar{s}$ lies in the PURSE.", + "bbox": [ + 498, + 479, + 890, + 541 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Worst-case error bounds. To upper bound the errors between the average pose $\\bar{s}$ and the groundtruth $(R_{\\mathrm{gt}},t_{\\mathrm{gt}})$ , we maximize the squared pose-to-PURSE distance:", + "bbox": [ + 498, + 542, + 890, + 588 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nd _ {\\epsilon , \\lambda} ^ {2} = \\max _ {(R, t) \\in S ^ {\\epsilon}} \\lambda \\| R - \\bar {R} \\| _ {\\mathrm {F}} ^ {2} + (1 - \\lambda) \\| t - \\bar {t} \\| ^ {2} \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 601, + 890, + 626 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "given $\\lambda \\in [0,1]$ . Particularly, we compute two cases $\\lambda = 1$ (the maximum rotation distance) and $\\lambda = 0$ (the maximum translation distance). Proposition 3 states the groundtruth $(R_{\\mathrm{gt}},t_{\\mathrm{gt}})$ lies in $S^{\\epsilon}$ with $1 - \\epsilon$ probability, hence", + "bbox": [ + 498, + 641, + 890, + 703 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\| \\bar {R} - R _ {\\mathrm {g t}} \\right\\| _ {\\mathrm {F}} \\leq d _ {\\epsilon , 1}, \\quad \\left\\| \\bar {t} - t _ {\\mathrm {g t}} \\right\\| \\leq d _ {\\epsilon , 0} \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 545, + 715, + 890, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "holds with probability $1 - \\epsilon$ .", + "bbox": [ + 498, + 747, + 692, + 762 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Computing the bounds. Problem (12) is nonconvex due to the constraints of the (PURSE) $S^{\\epsilon}$ . We relax the nonconvex problem (12) into a convex semidefinite program (SDP) and employ off-the-shelf solvers to optimize", + "bbox": [ + 496, + 763, + 890, + 825 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "Here we switch from P3P to PnP because PnP uses all $K$ keypoints and there is less ambiguity in its solution.", + "bbox": [ + 500, + 838, + 890, + 863 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "In Algorithm 1 we use rotation averaging with the Chordal distance metric. The user is free to choose other single rotation averaging algorithms with different distance metrics [30].", + "bbox": [ + 500, + 863, + 890, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "8952", + "bbox": [ + 482, + 946, + 514, + 958 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the SDP [14, 40, 98]. Two possible outcomes can happen: (i) the optimal SDP value coincides with the optimal value of (12). The relaxation is said to be exact and one can extract an optimal solution of (12) from the SDP, or (ii) the relaxation is not exact, but the optimal SDP value still provides an upper bound for the optimal value of (12). Therefore, we either exactly compute $d_{\\epsilon, \\lambda}^2$ or find an upper bound, both can bound the worst-case error (cf. (13)).", + "bbox": [ + 75, + 90, + 472, + 213 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We end with a remark about computing tighter bounds.", + "bbox": [ + 96, + 214, + 460, + 229 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Remark 4 (Best Worst-case Error Bounds). (12) can be used to bound errors for all possible pose estimators (e.g., from PnP (11)). What is the best estimator that attains the smallest error bounds? This boils down to solving", + "bbox": [ + 73, + 236, + 468, + 297 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {(\\bar {R}, \\bar {t}) \\in \\mathrm {S E} (3)} \\left[ \\max _ {(R, t) \\in S ^ {\\epsilon}} \\lambda \\| R - \\bar {R} \\| _ {\\mathrm {F}} ^ {2} + (1 - \\lambda) \\| t - \\bar {t} \\| ^ {2} \\right] \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 78, + 305, + 468, + 339 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "whose solution is known as the Chebyshev center [25, 63] of the PURSE $S^{\\epsilon}$ . Unfortunately, problem (14) is more challenging than (12) and there is no efficient algorithm to solve it to global optimality. In the Supplementary Material, we evaluate the worst-case error bounds for multiple $(\\bar{R},\\bar{t})$ samples, select the smallest bounds, and compare them with those of the average pose. An interesting future research direction is to explore differentiable optimization [73] or bilevel polynomial optimization [66] to solve (14).", + "bbox": [ + 75, + 345, + 472, + 483 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6. Experiments", + "text_level": 1, + "bbox": [ + 76, + 489, + 209, + 506 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We test our approach on the LineMOD Occlusion (LM-O) dataset [11] to (i) justify the exchangeability assumption (Theorem 1) and suggest best practices for applying conformal prediction; (ii) evaluate the empirical coverage of the PURSE and verify the correctness of Theorem 1, and (iii) compute the worst-case error bounds and demonstrate tightness or looseness. We also (iv) show that the average pose achieves better or similar accuracy as other approaches.", + "bbox": [ + 75, + 513, + 468, + 635 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Implementation and runtime. We set $T = 1000$ in RANSAG; use OpenGV [44] for P3P and PnP; and add a redundant $\\| t \\| \\leq 5$ in (PURSE) to ensure bounded translation. All procedures are implemented in Python except SDP relaxations are implemented in Matlab. The runtime of RANSAG is comparable to RANSAC and below one second. The runtime of computing (12) via SDPs is around 8 seconds on a workstation with 2.2GHz AMD CPUs. The (second-order) SDP relaxations are almost always exact.", + "bbox": [ + 75, + 635, + 470, + 772 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "6We omit the technical details and refer the interested reader to [98, Section 2] for a pragmatic introduction to SDP relaxations. In practice, we use the code provided by [98] in https://github.com/MIT-SPARK/CertifiablyRobustPerception, apply a second-order SDP relaxation to (12), and use MOSEK [65] to solve the SDP (in about 8 seconds). Solving a first-order SDP relaxation of (12) takes about 0.1 second but yields looser bounds.", + "The PURSE can potentially be empty, leading to infeasibility of problem (12). In such cases, empirically the SDP solver returns \"PRIMAL_INFEASIBLE\" (red squares lying on the $y$ -axis of Fig. 3)." + ], + "bbox": [ + 75, + 777, + 470, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Dataset and exchangeability. The LM-O dataset contains 1214 test images capturing 8 different objects on a table, of which 200 images were chosen by BOP19'20 [35]. We use the 200 images for calibration and the entire 1214 images for testing. As mentioned in Section 3, if the dataset was collected as a single video sequence under natural motion (e.g., a straight line), then the exchangeability assumption would fail. However, [33] described the data collection:", + "bbox": [ + 496, + 90, + 890, + 212 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In order to guarantee a well distributed pose space sampling of the dataset pictures, we uniformly divided the upper hemisphere of the objects into equally distant pieces and took at most one image per piece. As a result, our sequences provide uniformly distributed views ...", + "bbox": [ + 529, + 217, + 859, + 308 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "which indicates the 1214 images are independent (cf. [33, Figs. 5-6]) and therefore exchangeable. This demonstrates a good example for data collection –to equally divide the parameter space and collect one observation per division– so the guarantees offered by conformal prediction are valid.", + "bbox": [ + 496, + 313, + 890, + 388 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Empirical coverage. Our approach conformalizes the heatmaps [79] as (ball) or (ellipse). The implementation8 of [79] uses either groundtruth or Faster RCNN [75] bounding boxes, giving four variants of our approach: groundtruth box plus (ball) or (ellipse) (labels: gt-ball, gt-ellipse), and Faster RCNN box plus (ball) or (ellipse) (labels: frcnn-ball, frcnn-ellipse). Fig. 3 left column shows the empirical coverage (i.e., the percentage of images whose groundtruth poses lie in (PURSE)) of all four variants with $\\epsilon = 0.1$ and $\\epsilon = 0.4$ . We see the empirical coverage is around $90\\%$ when $\\epsilon = 0.1$ and around $60\\%$ when $\\epsilon = 0.4$ , for all 8 objects. Though the empirical coverage can deviate from $1 - \\epsilon$ , it generally stays within $\\pm 5\\%$ and mostly goes above $1 - \\epsilon$ , which is encouraging given that our calibration set only has size $n = 200$ . Fig. 1 (b) plots examples of the prediction sets. More examples are shown in the Supplementary Material.", + "bbox": [ + 496, + 390, + 892, + 631 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Worst-case error bounds. Fig. 3 middle column plots the worst-case rotation error bound ( $x$ -axis) vs. the actual rotation error between the average pose and the groundtruth ( $y$ -axis) for our approach using the gt-ball setup (results for gt-ellipse, fccnn-ball and fccnn-ellipse are similar and provided in the Supplementary Material). First, when the PURSE covers the groundtruth (blue circles), the rotation error bound is always larger than the actual error (i.e., the blue circles never cross the $y = x$ diagonal). Second, when the error rate is increased from $\\epsilon = 0.1$ to $\\epsilon = 0.4$ , we observe a shift of the blue circles towards $y = x$ , indicating the error bounds get tightened. Third, our bounds are reasonably tight for most test images (i.e., the bottom-left cluster of blue circles) especially when $\\epsilon = 0.4$ . However, they can become overly conservative (i.e., the line of blue circles on the right-side boundary) due to the keypoint prediction", + "bbox": [ + 496, + 632, + 893, + 875 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "8https://github.com/yufu-wang/6D_Pose", + "bbox": [ + 514, + 886, + 812, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "8953", + "bbox": [ + 482, + 946, + 514, + 958 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5392020343fc77e977bd7590c0f66a74ae84943612cdc7e9359143bb8f8b4b54.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 143, + 59, + 395, + 210 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/123b08c508a933f19417c16de49223b14183b92242c903bc165ad0128a2100f2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 408, + 59, + 612, + 210 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2f692da27de3c202e07df39842aaa3063addc2abc6f296526eff7fff967a4682.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 625, + 59, + 823, + 212 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c578ce8ac9470503174d157a3a2f01951c6b87afe521f504aa2651772a3c4613.jpg", + "image_caption": [ + "Figure 3. Empirical coverage (left) and worst-case error bounds (middle: rotation, right: translation). Top: $\\epsilon = 0.1$ , bottom: $\\epsilon = 0.4$ . For middle and right columns, $x$ -axis represents the worst-case error bounds computed from (12), $y$ -axis represents the actual error between average pose and groundtruth pose. The area below the diagonal $y = x$ indicates correctness of the bounds (i.e., bound $\\geq$ error), and points that are closer to the diagonal from below indicate tighter bounds (perfect if precisely lie on the diagonal). Blue circles plot cases where the PURSE covers the groundtruth pose and red squares plot cases were the PURSE does not cover the groundtruth. Notice that blue circles never cross the diagonal and our bounds are correct when the PURSE contains the pose (which holds with $1 - \\epsilon$ marginal probability)." + ], + "image_footnote": [], + "bbox": [ + 140, + 212, + 395, + 364 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a3d9c202b27254808b3d11653350e7617a92cd01cb9bfe52aaea13e91c8a093e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 411, + 214, + 609, + 364 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1cf8fb534a62b66a438eac51813d3b1fad772041567f93f5d37be085727eca48.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 625, + 214, + 821, + 364 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/37c172e1d97c06f734819ccdbca9fe08a041f9078197ad9289296a4ed9d2b515.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
objectsBaselines (results adapted from [72])Conformalized heatmap
Tekin [89]PoseCNN [95]Oberweger [67]PVNet [72]gt-ballgt-ellipsefrcnn-ballfrcnn-ellipse
ε = 0.1ε = 0.4ε = 0.1ε = 0.4ε = 0.1ε = 0.4ε = 0.1
ape7.0134.669.669.1477.7079.5279.2679.8870.2071.0168.84
can11.2015.1082.6086.0973.4175.9775.8178.1367.5269.8167.69
cat3.6210.4065.1065.1287.3690.5989.5490.1174.9580.2368.98
duck5.0731.8061.4061.4482.7183.0884.0283.5579.3080.6280.06
driller1.407.4073.8073.0679.3282.5481.2282.0458.4865.9258.06
eggbox-1.9013.108.43000.090.18000
glue4.7013.8054.9055.3756.4971.0871.6972.9330.0347.1841.96
holepuncher8.2623.1066.4069.8481.6582.8983.2284.3074.9677.8576.28
average6.1617.2060.9061.0667.3370.7170.6171.3956.9361.5857.73
", + "bbox": [ + 116, + 450, + 849, + 594 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 1. Success rates of baseline methods and our conformalized heatmap (using the average pose) based on the 2D projection metric (i.e., a pose estimation is considered successful if the average 2D reprojection error is below 5 pixels).", + "bbox": [ + 75, + 595, + 888, + 625 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "sets become too large. Fig. 3 right column plots similar results for the translation. The Supplementary Material gives a more detailed analysis of this conservatism, wherein we also solve (12) for multiple samples computed by RANSAG, choose the minimum bound, and compare them with those obtained for the average pose (cf. Remark 4).", + "bbox": [ + 75, + 628, + 468, + 718 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Accuracy of the average pose. We compare the accuracy of our average pose with other methods according to the 2D projection metric (an estimation is correct if the mean reprojection error is below 5 pixels). Table 1 shows: (i) our average pose achieves significantly better success rates when using groundtruth bounding boxes, and similar success rates when using Faster RCNN; (ii) the accuracy of the average pose increases when $\\epsilon$ increases.", + "bbox": [ + 75, + 719, + 468, + 840 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusions", + "text_level": 1, + "bbox": [ + 76, + 848, + 204, + 863 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We applied inductive conformal prediction to conformalize heatmap predictions as circular or elliptical prediction", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "sets that guarantee probabilistic coverage of the groundtruth keypoints, propagated the uncertainty in keypoints to the object pose to form a PURSE, designed RANSAG to sample from PURSE and compute an average pose, and used SDP relaxations to bound worst-case estimation errors. We validated our theory on the LineMOD Occlusion dataset. Future research will investigate better nonconformity functions, and applications to other vision problems.", + "bbox": [ + 496, + 628, + 890, + 750 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 763, + 660, + 780 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We thank Yufu Wang and Kostas Daniilidis for providing the heatmap implementation to detect semantic keypoints; Rachel Luo for discussing the exchangeability assumption in conformal prediction; Luca Carlone for pointing out related work on unknown-but-bounded noise (set membership) estimation in control theory, and anonymous reviewers for providing valuable feedback.", + "bbox": [ + 496, + 787, + 890, + 893 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8954", + "bbox": [ + 482, + 946, + 514, + 958 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Anastasios Angelopoulos, Stephen Bates, Jitendra Malik, and Michael I Jordan. Uncertainty sets for image classifiers using conformal prediction. In Intl. Conf. on Learning Representations (ICLR), 2021. 3", + "[2] Anastasios N Angelopoulos and Stephen Bates. A gentle introduction to conformal prediction and distribution-free uncertainty quantification. arXiv preprint arXiv:2107.07511, 2021. 3", + "[3] Anastasios N Angelopoulos, Stephen Bates, Adam Fisch, Lihua Lei, and Tal Schuster. Conformal risk control. arXiv preprint arXiv:2208.02814, 2022. 3", + "[4] Pasquale Antonante, Vasileios Tzoumas, Heng Yang, and Luca Carlone. Outlier-robust estimation: Hardness, minimally tuned algorithms, and applications. IEEE Trans. Robotics, 38(1):281-301, 2021. 5", + "[5] Boaz Barak and David Steurer. Proofs, beliefs, and algorithms through the lens of sum-of-squares. Course notes: http://www-sumofsquares.org/public/index.html, 1, 2016. 3", + "[6] Timothy D Barfoot. State estimation for robotics. Cambridge University Press, 2017. 3", + "[7] Jonathan T Barron. A general and adaptive robust loss function. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4331-4339, 2019. 5", + "[8] Stephen Bates, Anastasios Angelopoulos, Lihua Lei, Jitendra Malik, and Michael Jordan. Distribution-free, risk-controlling prediction sets. Journal of the ACM (JACM), 68(6):1-34, 2021. 3", + "[9] Michael J Black and Anand Rangarajan. On the unification of line processes, outlier rejection, and robust statistics with applications in early vision. *Intl. J. of Computer Vision*, 19(1):57-91, 1996. 2, 5", + "[10] Andrew Blake and Andrew Zisserman. Visual reconstruction. MIT press, 1987. 2", + "[11] Eric Brachmann, Alexander Krull, Frank Michel, Stefan Gumhold, Jamie Shotton, and Carsten Rother. Learning 6d object pose estimation using 3d object coordinates. In European Conf. on Computer Vision (ECCV), pages 536-551. Springer, 2014. 2, 4, 7", + "[12] Eric Brachmann, Frank Michel, Alexander Krull, Michael Ying Yang, Stefan Gumhold, et al. Uncertainty-driven 6d pose estimation of objects and scenes from a single rgb image. In IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 3364-3372, 2016. 2", + "[13] Eric Brachmann and Carsten Rother. Learning less is more-6d camera localization via 3d surface regression. In IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 4654–4662, 2018. 2", + "[14] Jesus Briales, Laurent Kneip, and Javier Gonzalez-Jimenez. A certifiably globally optimal solution to the non-minimal relative pose problem. In IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 145-154, 2018. 7", + "[15] Dylan Campbell, Liu Liu, and Stephen Gould. Solving the blind perspective-n-point problem end-to-end with robust" + ], + "bbox": [ + 86, + 114, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "differentiable geometric optimization. In European Conf. on Computer Vision (ECCV), pages 244-261. Springer, 2020. 2", + "[16] Dylan Campbell, Lars Petersson, Laurent Kneip, and Hong-dong Li. Globally-optimal inlier set maximisation for simultaneous camera pose and feature correspondence. In Intl. Conf. on Computer Vision (ICCV), pages 1-10, 2017. 2", + "[17] Emmanuel J Candes, Justin K Romberg, and Terence Tao. Stable signal recovery from incomplete and inaccurate measurements. Communications on Pure and Applied Mathematics: A Journal Issued by the Courant Institute of Mathematical Sciences, 59(8):1207-1223, 2006. 3", + "[18] Luca Carlone. Estimation contracts for outlier-robust geometric perception. arXiv preprint arXiv:2208.10521, 2022.3", + "[19] Bo Chen, Jiewei Cao, Alvaro Parra, and Tat-Jun Chin. Satellite pose estimation with deep landmark regression and nonlinear pose refinement. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 0–0, 2019. 1", + "[20] Bo Chen, Alvaro Parra, Jiewei Cao, Nan Li, and Tat-Jun Chin. End-to-end learnable geometric vision by backpropagating pnp optimization. In IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 8100-8109, 2020. 1, 2", + "[21] Hansheng Chen, Pichao Wang, Fan Wang, Wei Tian, Lu Xiong, and Hao Li. Epro-pnp: Generalized end-to-end probabilistic perspective-n-points for monocular object pose estimation. In IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 2781-2790, 2022. 2", + "[22] Tat-Jun Chin, Zhipeng Cai, and Frank Neumann. Robust fitting in computer vision: Easy or hard? In European Conf. on Computer Vision (ECCV), pages 701-716, 2018. 5", + "[23] Florence de Grancey, Jean-Luc Adam, Lucian Alecu, Sébastien Gerchinovitz, Franck Mamalet, and David Vigouroux. Object detection with probabilistic guarantees: A conformal prediction approach. In International Conference on Computer Safety, Reliability, and Security, pages 316-329. Springer, 2022. 3", + "[24] Ilias Diakonikolas, Daniel M Kane, Sushrut Karmalkar, Ankit Pensia, and Thanasis Pittas. List-decodable sparse mean estimation via difference-of-pairs filtering. arXiv preprint arXiv:2206.05245, 2022.3", + "[25] Yonina C Eldar, Amir Beck, and Marc Teboulle. A minimax chebyshev estimator for bounded error estimation. IEEE transactions on signal processing, 56(4):1388-1397, 2008. 7", + "[26] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 1, 2, 6", + "[27] Xiao-Shan Gao, Xiao-Rong Hou, Jianliang Tang, and Hang-Fei Cheng. Complete solution classification for the perspective-three-point problem. IEEE Trans. Pattern Anal. Machine Intell., 25(8):930–943, 2003. 2, 6" + ], + "bbox": [ + 509, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "8955", + "bbox": [ + 482, + 946, + 514, + 958 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Myung Geun Kim. Multivariate outliers and decompositions of mahalanobis distance. Communications in statistics-theory and methods, 29(7):1511-1526, 2000. 4", + "[29] Chunhui Gu and Xiaofeng Ren. Discriminative mixture-of-templates for viewpoint classification. In European Conf. on Computer Vision (ECCV), pages 408–421. Springer, 2010. 2", + "[30] Richard Hartley, Jochen Trumpf, Yuchao Dai, and Hong-dong Li. Rotation averaging. International journal of computer vision, 103:Intl. J. of Computer Vision, 2013. 6", + "[31] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. Cambridge university press, 2003. 3, 5", + "[32] Stefan Hinterstoisser, Cedric Cagniart, Slobodan Ilic, Peter Sturm, Nassir Navab, Pascal Fua, and Vincent Lepetit. Gradient response maps for real-time detection of textureless objects. IEEE Trans. Pattern Anal. Machine Intell., 34(5):876-888, 2011. 2", + "[33] Stefan Hinterstoisser, Vincent Lepetit, Slobodan Ilic, Stefan Holzer, Gary Bradski, Kurt Konolige, and Nassir Navab. Model based training, detection and pose estimation of texture-less 3d objects in heavily cluttered scenes. In Asian conference on computer vision, pages 548-562. Springer, 2012. 7", + "[34] Tomas Hodan, Daniel Barath, and Jiri Matas. Epos: Estimating 6d pose of objects with symmetries. In IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 11703-11712, 2020. 2", + "[35] Tomas Hodan, Frank Michel, Eric Brachmann, Wadim Kehl, Anders GlentBuch, Dirk Kraft, Bertram Drost, Joel Vidal, Stephan Ihrke, Xenophon Zabulis, et al. Bop: Benchmark for 6d object pose estimation. In European Conf. on Computer Vision (ECCV), pages 19-34, 2018. 2, 7", + "[36] Daniel P Huttenlocher, Gregory A. Klanderman, and William J Rucklidge. Comparing images using the hausdorff distance. IEEE Trans. Pattern Anal. Machine Intell., 15(9):850-863, 1993. 2", + "[37] Shun Iwase, Xingyu Liu, Rawal Khirodkar, Rio Yokota, and Kris M Kitani. Repose: Fast 6d object pose refinement via deep texture rendering. In Intl. Conf. on Computer Vision (ICCV), pages 3303-3312, 2021. 2", + "[38] Yanmei Jiao, Yue Wang, Bo Fu, Qimeng Tan, Lei Chen, Minhang Wang, Shoudong Huang, and Rong Xiong. Globally optimal consensus maximization for robust visual inertial localization in point and line map. In IEEE/RSJ Intl. Conf. on Intelligent Robots and Systems (IROS), pages 4631-4638. IEEE, 2020. 2", + "[39] Fredrik Kahl and Richard Hartley. Multiple-view geometry under the 1 infinity norm. IEEE Trans. Pattern Anal. Machine Intell., 30(9):1603-1617, 2008. 2, 5", + "[40] Fredrik Kahl and Didier Henrion. Globally optimal estimates for geometric reconstruction problems. *Intl. J. of Computer Vision*, 74(1):3-15, 2007. 7", + "[41] Alex Kendall, Matthew Grimes, and Roberto Cipolla. Posenet: A convolutional network for real-time 6-dof camera relocalization. In Intl. Conf. on Computer Vision (ICCV), pages 2938-2946, 2015. 2" + ], + "bbox": [ + 86, + 92, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[42] Georg Klein and David Murray. Parallel tracking and mapping for small ar workspaces. In 2007 6th IEEE and ACM international symposium on mixed and augmented reality, pages 225-234. IEEE, 2007. 1", + "[43] Adam Klivans, Pravesh K Kothari, and Raghu Meka. Efficient algorithms for outlier-robust regression. In Conference On Learning Theory, pages 1420–1430. PMLR, 2018. 3", + "[44] Laurent Kneip and Paul Furgale. Opengv: A unified and generalized approach to real-time calibrated geometric vision. In IEEE Intl. Conf. on Robotics and Automation (ICRA), pages 1-8. IEEE, 2014. 7", + "[45] Laurent Kneip, Hongdong Li, and Yongduek Seo. Upnp: An optimal o(n) solution to the absolute pose problem with universal applicability. In European Conf. on Computer Vision (ECCV), pages 127-142. Springer, 2014. 2", + "[46] Laurent Kneip, Davide Scaramuzza, and Roland Siegwart. A novel parametrization of the perspective-three-point problem for a direct computation of absolute camera position and orientation. In IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 2969-2976. IEEE, 2011. 2", + "[47] Yann Labbe, Justin Carpentier, Mathieu Aubry, and Josef Sivic. Cosypos: Consistent multi-view multi-object 6d pose estimation. In European Conf. on Computer Vision (ECCV), pages 574-591. Springer, 2020. 2", + "[48] Jing Lei, Max G'Sell, Alessandro Rinaldo, Ryan J Tibshirani, and Larry Wasserman. Distribution-free predictive inference for regression. Journal of the American Statistical Association, 113(523):1094-1111, 2018. 3", + "[49] Jing Lei, James Robins, and Larry Wasserman. Distribution-free prediction sets. Journal of the American Statistical Association, 108(501):278-287, 2013. 3", + "[50] Vincent Lepetit, Pascal Fua, et al. Monocular model-based 3d tracking of rigid objects: A survey. Foundations and Trends in Computer Graphics and Vision, 1(1):1-89, 2005. 2", + "[51] Vincent Lepetit, Francesc Moreno-Noguer, and Pascal Fua. Epnp: An accurate o(n) solution to the pnp problem. *Intl. J. of Computer Vision*, 81(2):155-166, 2009. 2", + "[52] Hongdong Li. Consensus set maximization with guaranteed global optimality for robust geometry estimation. In Intl. Conf. on Computer Vision (ICCV), pages 1074-1080. IEEE, 2009. 2", + "[53] Shuo Li, Sangdon Park, Xiayan Ji, Insup Lee, and Osbert Bastani. Towards pac multi-object detection and tracking. arXiv preprint arXiv:2204.07482, 2022. 3", + "[54] Zhigang Li, Gu Wang, and Xiangyang Ji. Cdpn: Coordinates-based disentangled pose network for real-time rgb-based 6-dof object pose estimation. In Intl. Conf. on Computer Vision (ICCV), pages 7678-7687, 2019. 2", + "[55] Joerg Liebelt, Cordelia Schmid, and Klaus Schertler. Independent object class detection using 3d feature maps. In IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 1-8. IEEE, 2008. 2", + "[56] David G Lowe. Object recognition from local scale-invariant features. In Intl. Conf. on Computer Vision (ICCV), volume 2, pages 1150-1157. IEEE, 1999. 2" + ], + "bbox": [ + 509, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "8956", + "bbox": [ + 482, + 946, + 514, + 958 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[57] Rachel Luo, Shengjia Zhao, Jonathan Kuck, Boris Ivanovic, Silvio Savarese, Edward Schmerling, and Marco Pavone. Sample-efficient safety assurances using conformal prediction. arXiv preprint arXiv:2109.14082, 2021. 4", + "[58] Yi Ma, Stefano Soatto, Jana Košecka, and Shankar Sastry. An invitation to 3-d vision: from images to geometric models, volume 26. Springer, 2004. 3", + "[59] Prasanta Chandra Mahalanobis. On the generalized distance in statistics. National Institute of Science of India, 1936. 4", + "[60] Lucas Manuelli, Wei Gao, Peter Florence, and Russ Tedrake. kpm: Keypoint affordances for category-level robotic manipulation. In The International Symposium of Robotics Research, pages 132-157. Springer, 2019. 1", + "[61] Maria Cecilia Mazzaro and Mario Sznaier. A set-membership approach to blind identification. In IEEE Conf. on Decision and Control (CDC), volume 5, pages 5176-5181. IEEE, 2004. 3", + "[62] Frank Michel, Alexander Kirillov, Eric Brachmann, Alexander Krull, Stefan Gumhold, Bogdan Savchynskyy, and Carsten Rother. Global hypothesis generation for 6d object pose estimation. In IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 462-471, 2017. 2", + "[63] Mario Milanese and Antonio Vicino. Optimal estimation theory for dynamic systems with set membership uncertainty: An overview. Automatica, 27(6):997-1009, 1991. 2, 3, 7", + "[64] Ankur Moitra. Sum of squares in theoretical computer science. Sum of Squares: Theory and Applications, 77:83, 2020. 3", + "[65] MOSEK ApS. The MOSEK optimization toolbox for MATLAB manual. Version 8.1., 2017. 7", + "[66] Jiawang Nie, Li Wang, and Jane J Ye. Bilevel polynomial programs and semidefinite relaxation methods. SIAM Journal on Optimization, 27(3):1728-1757, 2017. 7", + "[67] Markus Oberweger, Mahdi Rad, and Vincent Lepetit. Making deep heatmaps robust to partial occlusions for 3d object pose estimation. In European Conf. on Computer Vision (ECCV), pages 119-134, 2018. 2, 8", + "[68] Carl Olsson, Fredrik Kahl, and Magnus Oskarsson. Optimal estimation of perspective camera pose. In 18th International Conference on Pattern Recognition (ICPR'06), volume 2, pages 5-8. IEEE, 2006. 2, 5", + "[69] Harris Papadopoulos. Inductive conformal prediction: Theory and application to neural networks. In Tools in artificial intelligence. CiteSeer, 2008. 3", + "[70] Kiru Park, Timothy Patten, and Markus Vincze. Pix2pose: Pixel-wise coordinate regression of objects for 6d pose estimation. In Intl. Conf. on Computer Vision (ICCV), pages 7668-7677, 2019. 2", + "[71] Georgios Pavlakos, Xiaowei Zhou, Aaron Chan, Konstantinos G Derpanis, and Kostas Daniilidis. 6-DoF object pose from semantic keypoints. In IEEE Intl. Conf. on Robotics and Automation (ICRA), pages 2011–2018. IEEE, 2017. 1, 2, 4", + "[72] Sida Peng, Yuan Liu, Qixing Huang, Xiaowei Zhou, and Hujun Bao. Pvnet: Pixel-wise voting network for 6dof" + ], + "bbox": [ + 86, + 92, + 470, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "pose estimation. IEEE Trans. Pattern Anal. Machine Intell., 2022. 1, 2, 8", + "[73] Luis Pineda, Taosha Fan, Maurizio Monge, Shobha Venkataraman, Paloma Sodhi, Ricky TQ Chen, Joseph Ortiz, Daniel DeTone, Austin Wang, Stuart Anderson, Jing Dong, Brandon Amos, and Mustafa Mukadam. Theseus: A Library for Differentiable Nonlinear Optimization. In Advances in Neural Information Processing Systems (NIPS), 2022. 7", + "[74] Mahdi Rad and Vincent Lepetit. Bb8: A scalable, accurate, robust to partial occlusion method for predicting the 3d poses of challenging objects without using depth. In Proceedings of the IEEE international conference on computer vision, pages 3828-3836, 2017. 2", + "[75] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in Neural Information Processing Systems (NIPS), 28, 2015. 7", + "[76] Yaniv Romano, Matteo Sesia, and Emmanuel Candes. Classification with valid and adaptive coverage. In Advances in Neural Information Processing Systems (NIPS), volume 33, pages 3581-3591, 2020. 3", + "[77] David M Rosen, Luca Carlone, Afonso S Bandeira, and John J Leonard. Se-sync: A certifiably correct algorithm for synchronization over the special euclidean group. *Intl. J. of Robotics Research*, 38(2-3):95–125, 2019. 3", + "[78] Fred Rothganger, Svetlana Lazebnik, Cordelia Schmid, and Jean Ponce. 3d object modeling and recognition using local affine-invariant image descriptors and multi-view spatial constraints. *Intl. J. of Computer Vision*, 66(3):231-259, 2006. 2", + "[79] Karl Schmeckpeper, Philip R Osteen, Yufu Wang, Georgios Pavlakos, Kenneth Chaney, Wyatt Jordan, Xiaowei Zhou, Konstantinos G Derpanis, and Kostas Daniilidis. Semantic keypoint-based pose estimation from single RGB frames. J. of Field Robotics, 2022. 1, 4, 7", + "[80] Jingnan Shi, Heng Yang, and Luca Carlone. Optimal pose and shape estimation for category-level 3d object perception. In Robotics: Science and Systems (RSS), 2021. 1", + "[81] Jingnan Shi, Heng Yang, and Luca Carlone. Optimal and robust category-level perception: Object pose and shape estimation from 2d and 3d semantic keypoints. arXiv preprint arXiv:2206.12498, 2022. 1", + "[82] Torsten Söderström. Errors-in-variables methods in system identification. Automatica, 43(6):939–958, 2007. 3", + "[83] David Stutz, Ali Taylan Cemgil, Arnaud Doucet, et al. Learning optimal conformal classifiers. In Intl. Conf. on Learning Representations (ICLR), 2022. 3", + "[84] Hao Su, Charles R Qi, Yangyan Li, and Leonidas J Guibas. Render for cnn: Viewpoint estimation in images using cnns trained with rendered 3d model views. In Intl. Conf. on Computer Vision (ICCV), pages 2686-2694, 2015. 2", + "[85] Jiaming Sun, Zihao Wang, Siyu Zhang, Xingyi He, Hongcheng Zhao, Guofeng Zhang, and Xiaowei Zhou. Onepose: One-shot object pose estimation without cad models. In IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 6825-6834, 2022. 1" + ], + "bbox": [ + 509, + 92, + 890, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "8957", + "bbox": [ + 482, + 946, + 514, + 958 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[86] Min Sun, Gary Bradski, Bing-Xin Xu, and Silvio Savarese. Depth-encoded hough voting for joint object detection and shape recovery. In European Conf. on Computer Vision (ECCV), pages 658-671. Springer, 2010. 2", + "[87] Martin Sundermeyer, Zoltan-Csaba Marton, Maximilian Durner, Manuel Brucker, and Rudolph Triebel. Implicit 3d orientation learning for 6d object detection from rgb images. In European Conf. on Computer Vision (ECCV), pages 699-715, 2018. 2", + "[88] Richard Szeliski. Computer vision: algorithms and applications. Springer Nature, 2022. 3", + "[89] Bugra Tekin, Sudipta N Sinha, and Pascal Fua. Real-time seamless single shot 6d object pose prediction. In IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 292-301, 2018. 1, 2, 8", + "[90] Shubham Tulsiani and Jitendra Malik. Viewpoints and keypoints. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1510-1519, 2015. 2", + "[91] Vladimir Vovk. Conditional validity of inductive conformal predictors. In Asian conference on machine learning, pages 475-490. PMLR, 2012. 3", + "[92] Vladimir Vovk, Alexander Gammerman, and Glenn Shafer. Algorithmic learning in a random world. Springer Science & Business Media, 2005. 2, 3", + "[93] Gu Wang, Fabian Manhardt, Federico Tombari, and Xi-angyang Ji. Gdr-net: Geometry-guided direct regression network for monocular 6d object pose estimation. In IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 16611-16621, 2021. 2", + "[94] Håkan Wieslander, Philip J Harrison, Gabriel Skogberg, Sonya Jackson, Markus Fridén, Johan Karlsson, Ola Spjuth, and Carolina Wahlby. Deep learning with conformal prediction for hierarchical analysis of large-scale whole-slide tissue images. IEEE journal of biomedical and health informatics, 25(2):371-380, 2020. 3", + "[95] Yu Xiang, Tanner Schmidt, Venkatraman Narayanan, and Dieter Fox. Posecnn: A convolutional neural network for 6d object pose estimation in cluttered scenes. In Robotics: Science and Systems (RSS), 2018. 2, 8", + "[96] Heng Yang, Pasquale Antonante, Vasileios Tzoumas, and Luca Carlone. Graduated non-convexity for robust spatial perception: From non-minimal solvers to global outlier rejection. IEEE Robotics and Automation Letters, 5(2):1127-1134, 2020. 2, 5", + "[97] Heng Yang and Luca Carlone. In perfect shape: Certifiably optimal 3d shape reconstruction from 2d landmarks. In IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 621-630, 2020. 2", + "[98] Heng Yang and Luca Carlone. Certifiably optimal outlier-robust geometric perception: Semidefinite relaxations and scalable global optimization. IEEE Trans. Pattern Anal. Machine Intell., 2022. 1, 2, 7", + "[99] Heng Yang, Ling Liang, Luca Carlone, and Kim-Chuan Toh. An inexact projected gradient method with rounding and lifting by nonlinear programming for solving rank-one semidefinite relaxation of polynomial optimization. Mathematical Programming, pages 1-64, 2022. 3" + ], + "bbox": [ + 86, + 90, + 470, + 900 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[100] Heng Yang, Jingnan Shi, and Luca Carlone. Teaser: Fast and certifiable point cloud registration. IEEE Trans. Robotics, 37(2):314-333, 2020. 3", + "[101] Sergey Zakharov, Ivan Shugurov, and Slobodan Ilic. Dpod: 6d pose object detector and refiner. In Intl. Conf. on Computer Vision (ICCV), pages 1941-1950, 2019. 1, 2", + "[102] Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. In IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 5745-5753, 2019. 2" + ], + "bbox": [ + 501, + 90, + 890, + 233 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "8958", + "bbox": [ + 482, + 946, + 514, + 958 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2023/Object Pose Estimation With Statistical Guarantees_ Conformal Keypoint Detection and Geometric Uncertainty Propagation/a86330a1-e4d7-4fd5-a6ce-7c82335913b8_model.json b/2023/Object Pose Estimation With Statistical Guarantees_ Conformal Keypoint Detection and Geometric Uncertainty Propagation/a86330a1-e4d7-4fd5-a6ce-7c82335913b8_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7c4bddd28c90384ec4886f056afb76f503cd165e --- /dev/null +++ b/2023/Object Pose Estimation With Statistical Guarantees_ Conformal Keypoint Detection and Geometric Uncertainty Propagation/a86330a1-e4d7-4fd5-a6ce-7c82335913b8_model.json @@ -0,0 +1,2875 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.122, + 0.067, + 0.843, + 0.112 + ], + "angle": 0, + "content": "Object Pose Estimation with Statistical Guarantees: Conformal Keypoint Detection and Geometric Uncertainty Propagation" + }, + { + "type": "text", + "bbox": [ + 0.364, + 0.119, + 0.607, + 0.153 + ], + "angle": 0, + "content": "Heng Yang and Marco Pavone \nNVIDIA Research" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.161, + 0.891, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.327, + 0.893, + 0.425 + ], + "angle": 0, + "content": "Figure 1. Probabilistically correct object pose estimation. Given (a) an input image and heatmap detections of the object semantic keypoints, our method first conformalizes the heatmaps into (b) circular or elliptical prediction sets that guarantee probabilistic coverage of the groundtruth keypoints (e.g., \\(90\\%\\)). Our method then propagates the uncertainty in the keypoints to the object pose, forming (c) a Pose UncleRtainty SEt (PURSE) that contains the groundtruth pose with the same probability. We develop RANdom SAmple averaGing (RANSAG) to sample from PURSE and generate (d) an average pose, and apply semidefinite relaxation to compute (e) worst-case error bounds: the blue duck attains the worst rotation error w.r.t. the (average pose) yellow duck; the red duck attains the worst translation error. Code available: https://github.com/NVlabs/ConformalKeypoint." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.432, + 0.314, + 0.447 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.452, + 0.474, + 0.906 + ], + "angle": 0, + "content": "The two-stage object pose estimation paradigm first detects semantic keypoints on the image and then estimates the 6D pose by minimizing reprojection errors. Despite performing well on standard benchmarks, existing techniques offer no provable guarantees on the quality and uncertainty of the estimation. In this paper, we inject two fundamental changes, namely conformal keypoint detection and geometric uncertainty propagation, into the two-stage paradigm and propose the first pose estimator that endows an estimation with provable and computable worst-case error bounds. On one hand, conformal keypoint detection applies the statistical machinery of inductive conformal prediction to convert heuristic keypoint detections into circular or elliptical prediction sets that cover the groundtruth keypoints with a user-specified marginal probability (e.g., \\(90\\%\\)). Geometric uncertainty propagation, on the other, propagates the geometric constraints on the keypoints to the 6D object pose, leading to a Pose UnceRtainty SEt (PURSE) that guarantees coverage of the groundtruth pose with the same probability. The PURSE, however, is a nonconvex set that does not directly lead to estimated poses and uncertainties. Therefore, we develop RANdom SAmple averaGing (RANSAG) to compute an average pose and apply semidefinite relaxation to upper bound the worst-case errors between the average pose and the groundtruth. On the LineMOD Occlusion dataset we demonstrate: (i) the PURSE covers the groundtruth with valid probabilities; (ii) the worst-case error bounds provide correct uncertainty quantification; and (iii) the average pose achieves better or similar accuracy as representative methods based on sparse keypoints." + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.432, + 0.631, + 0.448 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.46, + 0.893, + 0.612 + ], + "angle": 0, + "content": "Estimating object poses from images is a fundamental problem in computer vision and finds extensive applications in augmented reality [42], autonomous driving [80], robotic manipulation [60], and space robotics [19]. One of the most popular paradigms for object pose estimation is a two-stage pipeline [20, 71, 72, 79, 81, 85, 89, 101], where the first stage detects (semantic) keypoints of the objects on the image, and the second stage computes the object pose by solving an optimization known as Perspective-\\(n\\)-Points (PnP) that minimizes reprojection errors of the detected keypoints." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Safety-critical applications call for provably correct computer vision algorithms. Existing algorithms in the two-stage paradigm (reviewed in Section 2), however, provide few performance guarantees on the quality of the estimated poses, due to three challenges. (C1) It is difficult to ensure the detected keypoints (typically from neural networks) are close to the groundtruth keypoints. In practice, the first stage often outputs keypoints that are arbitrarily wrong, known as outliers. (C2) Robust estimation is employed in the second stage to reject outliers, leading to nonconvex optimizations. Fast heuristics such as RANSAC [26] are widely adopted to find an approximate solution but they cannot guarantee global optimality and often fail without notice. (C3) There is no provably correct uncertainty quantification of the estimation, notably, a formal worst-case error bound between the estimation and the groundtruth. Though recent work [98] proposed convex relaxations to certify global optimality of RANSAC and addressed (C2), it cannot ensure correct estimation as the optimal pose may be far away from" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.947, + 0.516, + 0.959 + ], + "angle": 0, + "content": "8947" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.414, + 0.106 + ], + "angle": 0, + "content": "the correct pose when the keypoints are unreliable." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.111, + 0.473, + 0.655 + ], + "angle": 0, + "content": "Contributions. We propose a two-stage object pose estimation framework with statistical guarantees, illustrated in Fig. 1. Given an input image, we assume a neural network [71] is available to generate heatmap predictions of the object keypoints (Fig. 1(a)). Our framework then proceeds in two stages, namely conformal keypoint detection (Section 4) and geometric uncertainty propagation (Section 5). We first apply the statistical machinery of inductive conformal prediction (introduced in Section 3), with nonconformity functions inspired by the design of residual functions in classical geometric vision [39], to conformalize the heatmaps into circular or elliptical prediction sets—one for each keypoint—that guarantee coverage of the groundtruth keypoints with a user-specified marginal probability (Fig. 1(b)). This provides a simple and general methodology to bound the keypoint prediction errors (i.e., addressing (C1)). Given the keypoint prediction sets, we reformulate the constraints (enforced by the prediction sets) on the keypoints as constraints on the object pose, leading to a Pose UnceRtainty SET (PURSE) that guarantees coverage of the groundtruth pose with the same probability. Fig. 1(c) plots the boundary of an example PURSE (roll, pitch, raw angles for the rotation, and Euclidean coordinates for the translation). The PURSE, however, is an abstract nonconvex set that does not directly admit estimated poses and uncertainty. Therefore, we develop RANdom SAmple averaGing (RANSAG) to compute an average pose (Fig. 1(d)) and employ semidefinite relaxations to upper bound the worst-case rotation and translation errors between the average pose and the groundtruth (Fig. 1(e)). This gives rise to the first kind of computable worst-case probabilistic error bounds for object pose estimation (i.e., addressing (C3)). Our PURSE methodology has connections to the framework of unknown-but-bounded noise estimation in control theory [63], with special provisions to derive the bounds in a statistically principled way and enable efficient computation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.656, + 0.47, + 0.853 + ], + "angle": 0, + "content": "We test our framework on the LineMOD Occlusion (LM-O) dataset [11] to verify the correctness of the theory (Section 6). First, we empirically show that the PURSE indeed contains the groundtruth pose according to the user-specified probability. Second, we demonstrate the correctness of the worst-case error bounds: when the PURSE contains the groundtruth, our bounds are always larger than, and in many cases close to, the actual errors between the average pose and the groundtruth pose. Third, we benchmark the accuracy of the average pose (coming from RANSAG) with representative two-stage pipelines based on sparse keypoints (e.g., PVNet [72]) and show that the average pose achieves better or similar accuracy." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Limitations. A drawback of our approach, and conformal prediction in general, is that the size of the prediction sets depends on the nonconformity function (whose design" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.153 + ], + "angle": 0, + "content": "can be an art) and may be conservative. Our experiments suggest the bounds are loose when the keypoint prediction sets are large (e.g., giving \\(180^{\\circ}\\) rotation bound). We discuss challenges and opportunities in tightening the bounds." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.166, + 0.642, + 0.182 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.192, + 0.895, + 0.84 + ], + "angle": 0, + "content": "Image-based object pose estimation. We categorize object pose estimation into two paradigms: single-stage and two-stage. The latter first detects 2D-3D correspondences and then estimates the object pose via solving a PnP problem, while the former produces poses without intermediate correspondences. (i) Single-stage. Early methods perform pose estimation via template matching [29, 32, 36]. Recently, deep learning-based approaches such as PoseNet [41] and PoseCNN [95] applied CNNs to directly regress poses. A major challenge of pose regression is the nonlinearity of 3D rotations, and motivated formulating regression as classification [84, 87, 90] or designing better rotation representations [47, 102]. It is also popular to predict multiple pose hypotheses followed by voting [55, 62, 86]. (ii) Two-stage. Early research used handcrafted features [50, 56, 78] to establish 2D-3D correspondences and focused on developing algorithms for solving PnP. Notable algorithms include the minimal solver P3P [27, 46] and variants of the nonminimal solver PnP [45, 51, 68, 97]. Outliers (i.e., wrong correspondences) motivated robust estimation based on RANSAC [26], graduated non-convexity [9, 10, 96], branch-and-bound [16, 38, 52], or semidefinite relaxations [98]. Unreliable correspondences soon became the bottleneck and learned correspondences have been predominant. Learned correspondences can be sparse or dense. Sparse methods define a handful of keypoints and predict locations of the keypoints via direct regression [74, 89], probabilistic heatmap [67, 71], or voting [72]. Dense methods [12, 34, 54, 70, 93, 101] regress for each object pixel the coordinates of its corresponding 3D point. Recent literature focuses on end-to-end training via differentiating PnP [13, 15, 20, 21, 37]. Both single-stage and two-stage methods perform well on standard benchmarks [35], but a crucial feature that is missing, especially when deploying computer vision algorithms in safety-critical applications, is that these methods do not provide provably correct uncertainty quantification and formal error bounds w.r.t. the groundtruth (for either the correspondences or the poses). In this paper, we provide rigorous guarantees by applying conformal prediction to an existing keypoint detection method (the heatmap [71]) and leveraging old and new techniques in computer vision to derive formal error bounds." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Conformal prediction in computer vision. Conformal prediction [92] is a statistical machinery that offers provably correct finite-sample uncertainty quantification without assumptions on the data distribution or the prediction model" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.948, + 0.516, + 0.959 + ], + "angle": 0, + "content": "8948" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.363 + ], + "angle": 0, + "content": "(i.e., offering a set prediction, instead of a point prediction, that guarantees probabilistic coverage of the groundtruth). Inductive conformal prediction [69] is the most popular variant of conformal prediction because it does not require retraining of the prediction models [1, 2, 49]. Applying conformal prediction to computer vision, however, is still in its infancy. Existing works focus on image classification [1, 76], tumor segmentation [3, 8, 94], and bounding box detection [2, 23, 53], which are classification or low-dimensional regression problems. Inspired by these works, our unique contributions in this paper are: (i) we apply conformal prediction to keypoints detection, a high-dimensional regression problem; (ii) we design new nonconformity functions and discuss their connections with classical geometric vision; and (iii) we develop algorithms that propagate the uncertainty after conformal prediction to form prediction sets of 6D poses, which are nonlinear and nonconvex manifold objects." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.364, + 0.473, + 0.818 + ], + "angle": 0, + "content": "Performance guarantees. Pose estimation from 2D-2D, 2D-3D, and 3D-3D correspondences are foundational problems in computer vision textbooks [6, 31, 58, 88] and typically boil down to formulating and solving mathematical optimization problems. Benchmarking on simulated and real datasets has been a widely adopted standard for testing different formulations and solvers. However, empirical performance can be misleading without theoretical guarantees. A striking fact is that, though error analysis is an important topic in applied math [17, 24, 43] and control theory [61, 63, 82], there is very limited literature in computer vision that reason about worst-case estimation errors between the optimal solution and the groundtruth. A popular heuristic relies on the inverse of the Hessian at an optimal solution, which provides the Cramer-Rao lower bound on the covariance of the solution (for linear regression this coincides with the covariance) [88, Section B.6] and thus cannot upper bound the estimation errors. Recent works [18, 77, 100] derived error bounds for a few geometric vision problems. However, the bounds either depend on uncheckable assumptions and cannot be computed [77, 100], or build on machinery (e.g., sum-of-squares proof [5, 64]) that only applies to estimators based on moment relaxations [18], which are still computationally expensive in practice [99]. In this paper, we develop the first kind of efficiently computable error bounds that only require the assumption of exchangeability (which comes from conformal prediction). We justify this assumption on our test dataset and numerically show our bounds can be tight for a subset of the test problems." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.83, + 0.37, + 0.847 + ], + "angle": 0, + "content": "3. Inductive Conformal Prediction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.855, + 0.472, + 0.901 + ], + "angle": 0, + "content": "Given a set \\(\\{z_i = (x_i, y_i)\\}_{i=1}^l\\) with observation \\(x_i \\in \\mathcal{X}\\) and label \\(y_i \\in \\mathcal{Y}\\) such that each \\(z_i \\in \\mathcal{Z} := \\mathcal{X} \\times \\mathcal{Y}\\) is drawn i.i.d. from an unknown distribution on \\(\\mathcal{Z}\\), inductive confor" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.091, + 0.892, + 0.152 + ], + "angle": 0, + "content": "mal prediction (ICP) provides a set prediction \\( F^{\\epsilon}(x) \\subseteq \\mathcal{V} \\), parameterized by an error rate \\( 0 < \\epsilon < 1 \\), such that given a new sample \\( z_{l + 1} = (x_{l + 1}, y_{l + 1}) \\) satisfying an exchangeability condition (elaborated in Theorem 1), we have" + }, + { + "type": "equation", + "bbox": [ + 0.578, + 0.16, + 0.892, + 0.178 + ], + "angle": 0, + "content": "\\[\n\\mathbb {P} \\left[ y _ {l + 1} \\in F ^ {\\epsilon} \\left(x _ {l + 1}\\right) \\right] \\geq 1 - \\epsilon , \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.184, + 0.892, + 0.214 + ], + "angle": 0, + "content": "i.e., the prediction set \\( F^{\\epsilon} \\) guarantees to contain the true label \\( y_{l + 1} \\) with probability at least \\( 1 - \\epsilon \\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.215, + 0.893, + 0.366 + ], + "angle": 0, + "content": "Training. We start by dividing the dataset into a proper training set \\(\\{z_1, \\ldots, z_m\\}\\) and a calibration set \\(\\{z_{m+1}, \\ldots, z_l\\}\\). We shorthand \\(n = l - m\\) as the size of the calibration set. We learn a prediction function \\(f: \\mathcal{X} \\to \\tilde{\\mathcal{V}}\\) from the proper training set using any architecture, which allows us to fully exploit the power of modern deep learning. The prediction space \\(\\tilde{\\mathcal{V}}\\) can be the same as the label space \\(\\mathcal{V}\\), or can contain auxiliary information such as a heuristic notion of uncertainty (e.g., softmax scores in classification or a heatmap in the case of keypoint detection)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.366, + 0.894, + 0.427 + ], + "angle": 0, + "content": "Conformal calibration. We define a nonconformity function \\( S: \\mathcal{Z}^m \\times \\mathcal{Z} \\to \\mathbb{R} \\) to measure how well a given sample \\( z = (x, y) \\) conforms to the proper training set. A popular instance of \\( S \\) leverages the learned prediction \\( f \\):" + }, + { + "type": "equation", + "bbox": [ + 0.547, + 0.435, + 0.892, + 0.453 + ], + "angle": 0, + "content": "\\[\nS \\left(\\left\\{z _ {1}, \\dots , z _ {m} \\right\\}, (x, y)\\right) \\stackrel {{e. g.}} {{=}} r (y, f (x)), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.459, + 0.895, + 0.642 + ], + "angle": 0, + "content": "where \\(r: \\mathcal{Y} \\times \\tilde{\\mathcal{Y}} \\to \\mathbb{R}\\) is a measure of disagreement between the label \\(y\\) and the prediction \\(f(x)\\). For example, consider \\(\\mathcal{Y} = \\tilde{\\mathcal{Y}} = \\mathbb{R}\\), one can design \\(r(y, f(x)) = |y - f(x)|\\): if \\((x, y)\\) poorly conforms to the training set, \\(f\\) will incur large errors. While the function \\(S\\) can be arbitrary (e.g., a learnable neural network [83]), (2) is a convenient definition since \\(f\\) is implicitly dependent on \\(\\{z_i\\}_{i=1}^m\\) and \\(r\\) can incorporate domain-specific knowledge. We then compute the nonconformity scores on the calibration set as \\(\\alpha_i = r(y_i, f(x_i)), i = m+1,\\ldots,l\\), and sort them in nonincreasing order \\(\\alpha_{\\pi(1)} \\geq \\dots \\geq \\alpha_{\\pi(n)}\\), where \\(\\pi(i) \\in \\{m+1,\\ldots,l\\}\\) is an index permutation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.642, + 0.894, + 0.687 + ], + "angle": 0, + "content": "Conformal prediction. Given a new observation \\( x_{l+1} \\) (with an unknown \\( y_{l+1} \\)) and a user-specified \\( \\epsilon \\in (0, 1) \\), we compute the inductive conformal prediction (ICP) set as" + }, + { + "type": "equation", + "bbox": [ + 0.533, + 0.694, + 0.892, + 0.713 + ], + "angle": 0, + "content": "\\[\nF ^ {\\epsilon} \\left(x _ {l + 1}\\right) = \\left\\{y \\in \\mathcal {Y} \\mid \\alpha^ {y} \\leq \\alpha_ {\\pi (\\lfloor (n + 1) \\epsilon \\rfloor)} \\right\\}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.719, + 0.894, + 0.826 + ], + "angle": 0, + "content": "where \\(\\alpha^y = r(y, f(x_{l+1}))\\) is the nonconformity score of the new sample when fixing the true label to be \\(y\\). In other words, the ICP set (3) outputs the set of all labels that make the nonconformity score of the new sample no greater than \\(\\alpha_{\\pi(\\lfloor n + 1 \\epsilon \\rfloor)} - \\text{the } \\lfloor (n + 1) \\epsilon \\rfloor\\)-th largest nonconformity score in the calibration set. We have the following result stating the probabilistic coverage of the ICP set (3)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.832, + 0.895, + 0.878 + ], + "angle": 0, + "content": "Theorem 1 (Validity of ICP Coverage [48, 91, 92]). If \\( z_{m+1}, \\ldots, z_l, z_{l+1} = (x_{l+1}, y_{l+1}) \\) are exchangeable, i.e., their distribution is invariant under permutation, then" + }, + { + "type": "equation", + "bbox": [ + 0.51, + 0.885, + 0.892, + 0.903 + ], + "angle": 0, + "content": "\\[\n1 - \\epsilon \\leq \\mathbb {P} \\left[ y _ {l + 1} \\in F ^ {\\epsilon} \\left(x _ {l + 1}\\right) \\right] \\leq 1 - \\epsilon + 1 / (n + 1) \\tag {4}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.947, + 0.516, + 0.959 + ], + "angle": 0, + "content": "8949" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.159, + 0.076, + 0.391, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.23, + 0.47, + 0.273 + ], + "angle": 0, + "content": "Figure 2. Beta distribution of the conditional coverage in (5) with \\(\\epsilon = 0.1\\) and different \\(n\\). Notice how the conditional probability becomes more concentrated around \\(1 - \\epsilon\\) when \\(n\\) increases." + }, + { + "type": "text", + "bbox": [ + 0.074, + 0.275, + 0.47, + 0.307 + ], + "angle": 0, + "content": "for any \\(\\epsilon \\in (0,1)\\). Furthermore, when conditioned on the calibration set, calling \\(h = \\lfloor (n + 1)\\epsilon \\rfloor\\), we have" + }, + { + "type": "equation", + "bbox": [ + 0.067, + 0.317, + 0.47, + 0.336 + ], + "angle": 0, + "content": "\\[\n\\mathbb {P} \\left[ y _ {l + 1} \\in F ^ {\\epsilon} (x _ {l + 1}) \\mid \\{z _ {m + 1}, \\dots , z _ {l} \\} \\right] \\sim \\mathrm {B e t a} (n + 1 - h, h). (5)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.346, + 0.473, + 0.804 + ], + "angle": 0, + "content": "A few remarks are in order about Theorem 1. First, asking \\( z_{m+1}, \\ldots, z_l, z_{l+1} \\) to be exchangeable is weaker than asking them to be independent. However, this assumption typically fails when the calibration set is a single video sequence, where the image frames \\( \\{z_{m+1}, \\ldots, z_l\\} \\) are temporally correlated [57]. Fortunately, as we detail in Section 6, the way the LineMOD Occlusion dataset [11] was collected makes the exchangeability condition easily satisfied, which also suggests best practices to make the exchangeability condition hold in computer vision. Second, the lower bound in (4) can be intuitively proved because under exchangeability, \\( \\alpha_{l+1} := r(y_{l+1}, f(x_{l+1})) \\) – the nonconformity score of the new sample with the true label– is exchangeable with the nonconformity scores of the calibration samples, and hence equally likely to fall in anywhere between the scores \\( \\{\\alpha_{\\pi(i)}\\}_{i=1}^n \\). Consequently, \\( \\mathbb{P}[y_{l+1} \\in F^\\epsilon(x_{l+1})] = \\mathbb{P}\\left[\\alpha_{l+1} \\leq \\alpha_{\\pi(\\lfloor n+1\\epsilon \\rfloor)}\\right] = 1 - \\lfloor(n+1)\\epsilon \\rfloor / (n+1) \\geq 1 - \\epsilon \\). The upper bound in (4) states that \\( 1 - \\epsilon \\) is not overly conservative (indeed tight if \\( n \\) is large). Lastly, the probabilistic guarantee in (4) is marginal over the randomness of the calibration set, meaning if one chooses an infinite number of calibration sets, the average empirical coverage will converge to \\( 1 - \\epsilon \\). This, however, implies that the empirical coverage given one calibration set is a random variable that fluctuates as the Beta distribution (5). Fig. 2 plots the Beta distribution at \\( \\epsilon = 0.1 \\) with different sizes of the calibration set. We observe that as \\( n \\) increases the empirical coverage becomes more concentrated at \\( 1 - \\epsilon \\). Our experiments show that even with a small (\\( n = 200 \\)) calibration set, the empirical coverage is close to, and mostly higher than, \\( 1 - \\epsilon \\)." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.815, + 0.36, + 0.833 + ], + "angle": 0, + "content": "4. Conformal Keypoint Detection" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.469, + 0.87 + ], + "angle": 0, + "content": "In this section, we apply the ICP framework in Section 3 to the problem of semantic keypoint detection." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.87, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Setup. Denote by \\(x \\in \\mathbb{R}^{H \\times W \\times 3}\\) an RGB image picturing an object, by \\(\\pmb{y} = (y_1, \\dots, y_K) \\in \\mathbb{R}^2 \\times \\dots \\times \\mathbb{R}^2 \\coloneqq \\mathcal{Y}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "the groundtruth locations of \\( K \\) semantic keypoints of the object. We partition a given dataset \\( \\{z_i := (x_i, y_i)\\}_{i=1}^l \\) into a proper training set (of size \\( m \\)) and a calibration set (of size \\( n \\)). We follow the three steps in Section 3 to perform ICP." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.153, + 0.893, + 0.337 + ], + "angle": 0, + "content": "Training. We choose the heatmap approach in [71, 79] as the prediction function: given an image \\(x\\), [79] outputs a set of heatmaps \\(\\pmb{f}(x) = (f(x)_1, \\dots, f(x)_K)\\), where each \\(f(x)_k \\in \\Delta^{HW} := \\{v \\in \\mathbb{R}_+^{HW} \\mid \\sum_i^{HW} v_i = 1\\}\\) predicts the probability distribution of the \\(k\\)-th keypoint lying on each pixel of the image. For convenience, we use \\(q^j \\in \\mathbb{R}^2\\) to denote the \\(j\\)-th pixel location in \\(x\\) and \\(f(x)_k^j \\in \\mathbb{R}_+\\) to denote the probability of the \\(k\\)-th keypoint lying on \\(q^j\\). Let \\(\\sigma_k\\) be the index permutation that sorts \\(f(x)_k\\) in nonincreasing order, i.e., \\(f(x)_k^{\\sigma_k(1)} \\geq \\dots \\geq f(x)_k^{\\sigma_k(HW)}\\). As we will soon show, choosing the heatmap approach leads to simple and intuitive designs of the nonconformity function." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.338, + 0.892, + 0.368 + ], + "angle": 0, + "content": "Conformal calibration. We design the following nonconformity function" + }, + { + "type": "equation", + "bbox": [ + 0.551, + 0.378, + 0.892, + 0.398 + ], + "angle": 0, + "content": "\\[\nr (\\boldsymbol {y}, \\boldsymbol {f} (x)) = \\max \\left\\{\\phi \\left(y _ {k}, f (x) _ {k}\\right) \\right\\} _ {k = 1} ^ {K} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.408, + 0.892, + 0.453 + ], + "angle": 0, + "content": "that uses \\(\\phi\\) to score each keypoint and then selects the maximum score. This design considers the worst keypoint detection performance of \\(f\\). We provide two designs of \\(\\phi\\) below." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.454, + 0.892, + 0.502 + ], + "angle": 0, + "content": "(a) Peak. Shorthand \\( p_k = f(x)_k^{\\sigma_k(1)} \\) as the peak probability in the \\( k \\)-th heatmap and \\( q_k = q^{\\sigma_k(1)} \\) as the pixel location attaining the peak probability, we design" + }, + { + "type": "equation", + "bbox": [ + 0.584, + 0.513, + 0.892, + 0.531 + ], + "angle": 0, + "content": "\\[\n\\phi_ {\\text {p e a k}} \\left(y _ {k}, f (x) _ {k}\\right) = p _ {k} \\| y _ {k} - q _ {k} \\| \\quad (\\text {p e a k})\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.541, + 0.892, + 0.631 + ], + "angle": 0, + "content": "which computes the error between the true keypoint location \\( y_{k} \\) and the most probable keypoint location \\( q_{k} \\) and scales the error by the peak probability \\( p_{k} \\). \\( \\phi_{\\mathrm{peak}} \\) describes nonconformity because it becomes larger when the network \\( f \\) is confidently wrong (both \\( \\| y_{k} - q_{k}\\| \\) and \\( p_k \\) are large), implying the sample is highly nonconforming." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.632, + 0.892, + 0.701 + ], + "angle": 0, + "content": "(b) Covariance. Let \\(\\bar{q}_k = \\sum_{j=1}^{J} f(x)_k^{\\sigma_k(j)} q^{\\sigma_k(j)}\\) be the expected location of the top-\\(J\\) most likely detections for the \\(k\\)-th keypoint, and \\(\\Sigma_k = \\sum_{j=1}^{J} f(x)_k^{\\sigma_k(j)} \\cdot (q^{\\sigma_k(j)} - \\bar{q}_k)(q^{\\sigma_k(j)} - \\bar{q}_k)^{\\top}\\) as the covariance, we design" + }, + { + "type": "equation", + "bbox": [ + 0.529, + 0.71, + 0.892, + 0.73 + ], + "angle": 0, + "content": "\\[\n\\phi_ {\\operatorname {c o v}} \\left(y _ {k}, f (x) _ {k}\\right) = \\left(y _ {k} - \\bar {q} _ {k}\\right) ^ {\\top} \\Sigma_ {k} ^ {- 1} \\left(y _ {k} - \\bar {q} _ {k}\\right) \\quad (\\operatorname {c o v})\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.739, + 0.892, + 0.831 + ], + "angle": 0, + "content": "which computes the squared Mahalanobis distance [59] from the groundtruth \\( y_{k} \\) to the top-\\( J \\) keypoint detections (represented by the mean \\( \\bar{q}_{k} \\) and covariance \\( \\Sigma_{k} \\)).2 A larger Mahalanobis distance indicates more abnormality of the heatmap \\( f(x)_k \\) (compared to the groundtruth \\( y_{k} \\)) [28], and hence implies higher nonconformity." + }, + { + "type": "page_footnote", + "bbox": [ + 0.498, + 0.839, + 0.892, + 0.876 + ], + "angle": 0, + "content": "1The heatmap in the original paper [71] is not a valid probability distribution as it contains negative values and do not sum up to 1. We remove the negative values and normalize it to be a valid probability distribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.876, + 0.892, + 0.901 + ], + "angle": 0, + "content": "2We only choose the top- \\(J\\) \\((J = 100)\\) most likely detections on the heatmap because the heatmap can be quite noisy in practice." + }, + { + "type": "list", + "bbox": [ + 0.498, + 0.839, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.947, + 0.516, + 0.959 + ], + "angle": 0, + "content": "8950" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.138 + ], + "angle": 0, + "content": "Using the nonconformity function (6) with (peak) or (cov), we compute the nonconformity scores of the calibration set and sort them as: \\(\\alpha_{\\pi(1)} \\geq \\dots \\geq \\alpha_{\\pi(n)}\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.138, + 0.47, + 0.199 + ], + "angle": 0, + "content": "Conformal prediction. Given an error rate \\(\\epsilon \\in (0,1)\\), we first find \\(\\alpha_{\\pi (\\lfloor (n + 1)\\epsilon \\rfloor)}\\). Then, according to the ICP set definition (3) and our nonconformity function (6), we output the ICP set for a new \\(x_{l + 1}\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.078, + 0.213, + 0.469, + 0.268 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} F ^ {\\epsilon} (x _ {l + 1}) \\\\ = \\left\\{\\boldsymbol {y} \\in \\mathcal {Y} \\mid \\max \\left\\{\\phi \\left(y _ {k}, f \\left(x _ {l + 1}\\right)\\right) \\right\\} _ {k = 1} ^ {K} \\leq \\alpha_ {\\pi (\\lfloor (n + 1) \\epsilon \\rfloor)} \\right\\} \\\\ = \\left\\{\\boldsymbol {y} \\in \\mathcal {Y} \\mid \\phi \\left(y _ {k}, f \\left(x _ {l + 1}\\right)\\right) \\leq \\alpha_ {\\pi (\\lfloor (n + 1) \\epsilon \\rfloor)}, \\forall k \\right\\}, \\tag {7} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.28, + 0.469, + 0.312 + ], + "angle": 0, + "content": "where we used \\(\\max \\{\\phi_1,\\ldots ,\\phi_K\\} \\leq \\alpha\\) if and only if \\(\\phi_{k}\\leq \\alpha\\) for any \\(k\\). Insert (peak) into (7), we have \\(F_{\\mathrm{ball}}^{\\epsilon}(x_{l + 1})\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.098, + 0.324, + 0.47, + 0.358 + ], + "angle": 0, + "content": "\\[\n\\left\\{\\boldsymbol {y} \\in \\mathcal {Y} \\mid \\| y _ {k} - q _ {l + 1, k} \\| \\leq \\frac {\\alpha_ {\\pi (\\lfloor (n + 1) \\epsilon \\rfloor)}}{p _ {l + 1 , k}}, \\forall k \\right\\}, \\quad (\\text {b a l l})\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.37, + 0.469, + 0.433 + ], + "angle": 0, + "content": "which defines -for the \\(k\\) -th keypoint- a ball centered at \\(q_{l + 1,k}\\) (the most likely detection) with a radius inversely proportional to \\(p_{l + 1,k}\\) and proportional to \\(\\alpha_{\\pi (\\lfloor (n + 1)\\epsilon \\rfloor)}\\). Similarly, insert (cov) into (7), we have \\(F_{\\mathrm{ellipse}}^{\\epsilon}(x_{l + 1})\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.061, + 0.445, + 0.48, + 0.497 + ], + "angle": 0, + "content": "\\[\n\\left\\{\\boldsymbol {y} \\in \\mathcal {Y} \\mid (y _ {k} - \\bar {q} _ {l + 1, k}) ^ {\\mathsf {T}} \\frac {\\sum_ {l + 1 , k} ^ {- 1}}{\\alpha_ {\\pi (\\lfloor (n + 1) \\epsilon \\rfloor)} (y _ {k} - \\bar {q} _ {l + 1 , k})} \\leq 1, \\forall k \\right\\},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.498, + 0.47, + 0.621 + ], + "angle": 0, + "content": "which defines -for the \\(k\\)-th keypoint-- an ellipse centered at \\(\\bar{q}_{l+1,k}\\) (the expected location of the top-\\(J\\) detections) with an area proportional to \\(\\operatorname*{det}(\\Sigma_{l+1,k})\\) and \\(\\alpha_{\\pi(\\lfloor n+1\\epsilon \\rfloor)}\\).3 From (ball) and (ellipse), we observe that the prediction sets become larger when (i) the heatmaps are uncertain, i.e., the peak probability is low or the covariance matrix has large determinant; and (ii) the heatmaps perform poorly on the calibration set, leading to a large \\(\\alpha_{\\pi(\\lfloor n+1\\epsilon \\rfloor)}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.621, + 0.47, + 0.711 + ], + "angle": 0, + "content": "Connections to geometric vision. Our nonconformity function bears similarity to the residual function in geometric vision [4, 22, 31]. For example, the (peak) and (cov) functions are similar to the (weighted) reprojection error [31], and the \"max\" in (6) can be connected to seminal work on optimizing the \\(\\ell_{\\infty}\\) norm [39]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.711, + 0.471, + 0.864 + ], + "angle": 0, + "content": "Outlier-robust nonconformity? One potential issue of the nonconformity function (6) is that a single outlier can inflate the score and the calibration quantile \\(\\alpha_{\\pi (\\lfloor (n + 1)\\epsilon \\rfloor)}\\) and lead to conservative prediction sets (e.g., when \\(\\pmb{f}\\) predicts \\(K - 1\\) keypoints perfectly but misses one keypoint). A potential remedy in geometric vision is to use robust cost functions [7,9,96]. Therefore, a natural question is whether \"robustifying\" the nonconformity function (6) can lead to better prediction sets. Here we focus on only robustifying \\(\\phi\\) in (6) and provide a negative answer." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.892, + 0.138 + ], + "angle": 0, + "content": "Proposition 2 (Invariance of ICP). Let \\(\\rho : \\mathbb{R}_+ \\mapsto \\mathbb{R}_+\\) be any monotonically increasing function. Fixing the calibration set and error rate \\(\\epsilon\\), the nonconformity function" + }, + { + "type": "equation", + "bbox": [ + 0.538, + 0.148, + 0.892, + 0.166 + ], + "angle": 0, + "content": "\\[\nr _ {\\rho} (\\boldsymbol {y}, \\boldsymbol {f} (x)) = \\max \\left\\{\\rho \\left(\\phi \\left(y _ {k}, f (x) _ {k}\\right)\\right) \\right\\} _ {k = 1} ^ {K} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.177, + 0.715, + 0.192 + ], + "angle": 0, + "content": "leads to the same ICP set as (6)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.2, + 0.892, + 0.356 + ], + "angle": 0, + "content": "The proof of Proposition 2 is presented in Supplementary Material. We conclude that common robust costs, such as \\(\\ell_1\\), Huber, Geman-McClure, and Barron's adaptive kernel [7,9] (which are monotonically increasing on \\([0, +\\infty]\\)) cannot change the ICP sets by robustifying the individual score \\(\\phi\\). However, it remains an open question whether changing the \"max\" operation in (6) can give rise to better ICP sets. For instance, replacing \"max\" with \"\\(\\sum\\) in (6)\" and using the Geman-McClure robust cost \\(\\rho(\\phi) = \\frac{\\phi^2}{1 + \\phi^2}\\) with \\(\\phi = \\phi_{\\mathrm{peak}}\\) results in the following ICP set" + }, + { + "type": "equation", + "bbox": [ + 0.51, + 0.366, + 0.892, + 0.409 + ], + "angle": 0, + "content": "\\[\n\\left\\{\\boldsymbol {y} \\in \\mathcal {Y} \\mid \\sum_ {k = 1} ^ {K} \\frac {p _ {k} ^ {2} \\| y _ {k} - q _ {k} \\| ^ {2}}{1 + p _ {k} ^ {2} \\| y _ {k} - q _ {k} \\| ^ {2}} \\leq \\alpha_ {\\pi (\\lfloor (n + 1) \\epsilon \\rfloor)} \\right\\} \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.418, + 0.892, + 0.495 + ], + "angle": 0, + "content": "that does not admit a geometric interpretation that is as simple and intuitive as the (ball) and (ellipse) sets introduced before. In fact, it is indeed the simplicity of (ball) and (ellipse) that enables us to propagate the uncertainty in keypoints to the object pose, as we will show in the next section." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.507, + 0.825, + 0.525 + ], + "angle": 0, + "content": "5. Geometric Uncertainty Propagation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.532, + 0.892, + 0.579 + ], + "angle": 0, + "content": "Conformalizing the heatmaps gives us prediction sets that guarantee probabilistic coverage of the true keypoints. We unify the prediction sets (ball) and (ellipse) as" + }, + { + "type": "equation", + "bbox": [ + 0.5, + 0.588, + 0.899, + 0.607 + ], + "angle": 0, + "content": "\\[\nF ^ {\\epsilon} (x) = \\left\\{\\boldsymbol {y} \\in \\mathcal {Y} \\mid \\left(y _ {k} - \\mu_ {k}\\right) ^ {\\top} \\Lambda_ {k} \\left(y _ {k} - \\mu_ {k}\\right) \\leq 1, \\forall k \\right\\}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.617, + 0.892, + 0.683 + ], + "angle": 0, + "content": "where \\(\\mu_{k} = q_{l + 1,k},\\Lambda_{k} = \\frac{p_{l + 1,k}^{2}}{\\alpha_{\\pi(\\lfloor(n + 1)\\epsilon \\rfloor)}^{2}}\\mathbf{I}_{2}\\) for (ball), \\(\\mu_{k} =\\) \\(\\bar{q}_{l + 1,k},\\Lambda_k = \\frac{\\sum_{l + 1,k}^{-1}}{\\alpha_\\pi(\\lfloor(n + 1)\\epsilon \\rfloor)}\\) for (ellipse), and we omit the subscript \\(l + 1\\) for simplicity." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.684, + 0.892, + 0.714 + ], + "angle": 0, + "content": "Why not uncertainty-aware PnP? A popular way to estimate pose from (10) is to solve an uncertainty-aware PnP" + }, + { + "type": "equation", + "bbox": [ + 0.535, + 0.724, + 0.826, + 0.765 + ], + "angle": 0, + "content": "\\[\n\\min _ {(R, t) \\in \\mathrm {S E} (3)} \\quad \\sum_ {k = 1} ^ {K} \\left(y _ {k} - \\mu_ {k}\\right) ^ {\\mathsf {T}} \\Lambda_ {k} \\left(y _ {k} - \\mu_ {k}\\right)\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.537, + 0.768, + 0.892, + 0.784 + ], + "angle": 0, + "content": "\\[\n\\text {s u b j e c t} \\quad y _ {k} = \\Pi \\left(R Y _ {k} + t\\right), k = 1, \\dots , K \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where \\( Y_{k} \\in \\mathbb{R}^{3}, k = 1, \\dots, K \\) are the 3D object keypoints and \\( \\Pi(\\cdot) \\) denotes the camera projection. We challenge this approach and point out its two drawbacks. First, it is difficult to solve (11) to global optimality due to (i) the nonconvex SE(3) constraint and (ii) the rational polynomial appearing in \\( \\Pi(\\cdot) \\). The best known approach to solve (11) relies on either branch-and-bound [68] or local optimization." + }, + { + "type": "page_footnote", + "bbox": [ + 0.094, + 0.886, + 0.456, + 0.902 + ], + "angle": 0, + "content": "3The area of \\((x - \\mu)^{\\mathsf{T}}A(x - \\mu)\\leq 1\\) is proportional to \\(\\operatorname *{det}(A^{-1})\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.948, + 0.514, + 0.96 + ], + "angle": 0, + "content": "8951" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.166 + ], + "angle": 0, + "content": "Second, solving (11) typically outputs a single optimal pose without uncertainty quantification. Are there other poses that attain similar costs as the optimal pose? How close is the optimal pose to the groundtruth pose? These questions remain not answered in the literature." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.168, + 0.47, + 0.214 + ], + "angle": 0, + "content": "Pose UnceRtainty SET (PURSE). We propose to, instead of solving a PnP problem similar to (11), directly propagate the uncertainty in the ICP sets to the object pose." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.224, + 0.47, + 0.303 + ], + "angle": 0, + "content": "Proposition 3 (PURSE). Let \\( s_{\\mathrm{gt}} = [\\mathrm{vec}(R_{\\mathrm{gt}})^{\\top}; t_{\\mathrm{gt}}^{\\top}]^{\\top} \\) be the groundtruth object pose (that lies in front of the camera). Then, the groundtruth keypoints \\( \\mathbf{y} = (y_1, \\ldots, y_K) \\) belong to the ICP set \\( F^{\\epsilon}(x) \\) in (10) if and only if \\( s_{\\mathrm{gt}} \\) belongs to the following pose uncertainty set" + }, + { + "type": "equation", + "bbox": [ + 0.074, + 0.313, + 0.47, + 0.348 + ], + "angle": 0, + "content": "\\[\nS ^ {\\epsilon} = \\left\\{s \\in \\operatorname {S E} (3) \\left| \\begin{array}{c} s ^ {\\mathsf {T}} A _ {k} s \\leq 0, k = 1, \\ldots , K \\\\ b _ {k} ^ {\\mathsf {T}} s > 0, k = 1, \\ldots , K \\end{array} \\right. \\right\\}, (\\text {P U R S E})\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.359, + 0.47, + 0.391 + ], + "angle": 0, + "content": "where \\(A_{k}\\in \\mathbb{S}^{12},b_{k}\\in \\mathbb{R}^{12},k = 1,\\ldots ,K\\) are constant matrices dependent on \\(\\mu_k,\\Lambda_k,Y_k\\) and camera intrinsics." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.402, + 0.47, + 0.613 + ], + "angle": 0, + "content": "The detailed proof for Proposition 3 is algebraically involved and postponed to the Supplementary Material. The high-level intuition is, however, straightforward: we plug in \\( y_{k} = \\Pi (RY_{k} + t) \\) into (10) and obtain \\( K \\) quadratic inequalities of the form \\( s^{\\mathsf{T}}A_{k}s\\leq 0 \\). The linear inequalities \\( b_{k}^{\\top}s > 0 \\) are added to enforce the (transformed) 3D keypoints lie in front of the camera. Proposition 3 implies, if we are \\( 1 - \\epsilon \\) confident the groundtruth keypoints can be anywhere inside \\( F^{\\epsilon}(x) \\), then we should also be confident any pose in (PURSE) can be the groundtruth. Viewing pose estimation as a set estimation with guaranteed probabilistic coverage of the groundtruth is fundamentally different from viewing it as computing a single pose from (11) that is (hopefully) close enough to the groundtruth." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.471, + 0.902 + ], + "angle": 0, + "content": "RANdom SExample averaGing (RANSAG). Verifying if a given pose belongs to the PURSE is straightforward via checking the inequalities in (PURSE). However, the PURSE does not directly give us estimated poses. Therefore, we propose an efficient sampling algorithm called RANSDOM SExample averaGing (RANSAG) that is analogous to RANSAC [26] and leverages the minimal solver P3P [27], presented in Algorithm 1. The intuition is that, though it is difficult to sample directly in PURSE due to the (nonconvex) constraints, it is easy to sample from the keypoint prediction set (10) due to its simple geometry (balls and ellipses). Thus, at each iteration (line 3) RANSAG samples three keypoints (line 4-5), solves the P3P inverse problem, and accept the poses that belong to the (PURSE) (line 6). RANSAG typically returns around 100 valid samples with \\( T = 1000 \\) trials. However, in difficult cases (e.g., when \\( S^{\\epsilon} \\) is small or even empty) it is possible to obtain zero samples (\\( S = \\emptyset \\)). In this situation, RANSAG samples \\( \\lfloor T / 20 \\rfloor \\) (default 50) poses without checking if they belong to the PURSE, via sampling" + }, + { + "type": "code_caption", + "bbox": [ + 0.51, + 0.095, + 0.803, + 0.111 + ], + "angle": 0, + "content": "Algorithm 1: RANdom SAmple averaGing" + }, + { + "type": "algorithm", + "bbox": [ + 0.507, + 0.113, + 0.867, + 0.469 + ], + "angle": 0, + "content": "1 Input: an ICP set \\(F^{\\epsilon}(x)\\) (10) and its corresponding (PURSE) \\(S^{\\epsilon}\\); maximum trials \\(T\\); initial \\(\\hat{S} = \\emptyset\\); \n2 Output: sample poses \\(S \\subset \\mathrm{SE}(3)\\) in PURSE, and an average pose \\(\\bar{s} \\in \\mathrm{SE}(3)\\); \n3 for \\(\\tau \\gets 1\\) to \\(T\\) do \n4 Sample \\(\\{k_1, k_2, k_3\\}\\) from \\([K] (k_1 \\neq k_2 \\neq k_3)\\); \n5 Sample \\(\\hat{y}_{k_i}, i = 1, 2, 3\\) from \\(\\{y \\in \\mathbb{R}^2 \\mid (y - \\mu_{k_i})^\\top \\Lambda_{k_i} (y - \\mu_{k_i}) \\leq 1\\}\\); \n6 \\(\\hat{S} \\gets \\hat{S} \\cup (S^{\\epsilon} \\cap \\mathsf{P3P}(\\{\\hat{y}_{k_i} \\leftrightarrow Y_{k_i}\\}_{i=1}^3))\\); \n7 end \n8 \\(S = \\hat{S}\\); \n9 if \\(\\hat{S} = \\emptyset\\) then \n10 for \\(\\tau \\gets 1\\) to \\(\\lfloor T/20 \\rfloor\\) do \n11 Sample \\(\\hat{y}_k, k = 1, \\dots, K\\) from \\(F^{\\epsilon}(x)\\); \n12 \\(\\hat{S} \\gets \\hat{S} \\cup \\mathsf{PnP}(\\{\\hat{y}_k \\leftrightarrow Y_k\\}_{k=1}^K)\\); \n13 end \n14 end \n15 \\(\\bar{R} = \\operatorname{proj}_{\\mathrm{SO}(3)}(\\sum_{(R_j, *) \\in \\hat{S}} R_j)\\); \n16 \\(\\bar{t} = \\frac{1}{|\\hat{S}|} \\sum_{(*, t_j) \\in \\hat{S}} t_j\\); \n17 return: \\(S, \\bar{s} = (\\bar{R}, \\bar{t})\\)" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.481, + 0.892, + 0.542 + ], + "angle": 0, + "content": "\\(K\\) keypoints and solving PnP (line 9-12). After obtaining a set of poses, RANSAG performs rotation averaging (line 14) and translation averaging (line 15) to obtain an average pose \\(\\bar{s}\\). Note that RANSAG does not check if \\(\\bar{s}\\) lies in the PURSE." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.543, + 0.892, + 0.589 + ], + "angle": 0, + "content": "Worst-case error bounds. To upper bound the errors between the average pose \\(\\bar{s}\\) and the groundtruth \\((R_{\\mathrm{gt}},t_{\\mathrm{gt}})\\), we maximize the squared pose-to-PURSE distance:" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.602, + 0.892, + 0.627 + ], + "angle": 0, + "content": "\\[\nd _ {\\epsilon , \\lambda} ^ {2} = \\max _ {(R, t) \\in S ^ {\\epsilon}} \\lambda \\| R - \\bar {R} \\| _ {\\mathrm {F}} ^ {2} + (1 - \\lambda) \\| t - \\bar {t} \\| ^ {2} \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.642, + 0.892, + 0.704 + ], + "angle": 0, + "content": "given \\(\\lambda \\in [0,1]\\). Particularly, we compute two cases \\(\\lambda = 1\\) (the maximum rotation distance) and \\(\\lambda = 0\\) (the maximum translation distance). Proposition 3 states the groundtruth \\((R_{\\mathrm{gt}},t_{\\mathrm{gt}})\\) lies in \\(S^{\\epsilon}\\) with \\(1 - \\epsilon\\) probability, hence" + }, + { + "type": "equation", + "bbox": [ + 0.547, + 0.717, + 0.892, + 0.734 + ], + "angle": 0, + "content": "\\[\n\\left\\| \\bar {R} - R _ {\\mathrm {g t}} \\right\\| _ {\\mathrm {F}} \\leq d _ {\\epsilon , 1}, \\quad \\left\\| \\bar {t} - t _ {\\mathrm {g t}} \\right\\| \\leq d _ {\\epsilon , 0} \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.748, + 0.694, + 0.763 + ], + "angle": 0, + "content": "holds with probability \\(1 - \\epsilon\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.826 + ], + "angle": 0, + "content": "Computing the bounds. Problem (12) is nonconvex due to the constraints of the (PURSE) \\( S^{\\epsilon} \\). We relax the nonconvex problem (12) into a convex semidefinite program (SDP) and employ off-the-shelf solvers to optimize" + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.839, + 0.892, + 0.864 + ], + "angle": 0, + "content": "Here we switch from P3P to PnP because PnP uses all \\(K\\) keypoints and there is less ambiguity in its solution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.864, + 0.892, + 0.9 + ], + "angle": 0, + "content": "In Algorithm 1 we use rotation averaging with the Chordal distance metric. The user is free to choose other single rotation averaging algorithms with different distance metrics [30]." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.839, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.947, + 0.516, + 0.959 + ], + "angle": 0, + "content": "8952" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.214 + ], + "angle": 0, + "content": "the SDP [14, 40, 98]. Two possible outcomes can happen: (i) the optimal SDP value coincides with the optimal value of (12). The relaxation is said to be exact and one can extract an optimal solution of (12) from the SDP, or (ii) the relaxation is not exact, but the optimal SDP value still provides an upper bound for the optimal value of (12). Therefore, we either exactly compute \\( d_{\\epsilon, \\lambda}^2 \\) or find an upper bound, both can bound the worst-case error (cf. (13))." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.215, + 0.462, + 0.23 + ], + "angle": 0, + "content": "We end with a remark about computing tighter bounds." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.237, + 0.47, + 0.299 + ], + "angle": 0, + "content": "Remark 4 (Best Worst-case Error Bounds). (12) can be used to bound errors for all possible pose estimators (e.g., from PnP (11)). What is the best estimator that attains the smallest error bounds? This boils down to solving" + }, + { + "type": "equation", + "bbox": [ + 0.08, + 0.306, + 0.47, + 0.34 + ], + "angle": 0, + "content": "\\[\n\\min _ {(\\bar {R}, \\bar {t}) \\in \\mathrm {S E} (3)} \\left[ \\max _ {(R, t) \\in S ^ {\\epsilon}} \\lambda \\| R - \\bar {R} \\| _ {\\mathrm {F}} ^ {2} + (1 - \\lambda) \\| t - \\bar {t} \\| ^ {2} \\right] \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.347, + 0.473, + 0.484 + ], + "angle": 0, + "content": "whose solution is known as the Chebyshev center [25, 63] of the PURSE \\( S^{\\epsilon} \\). Unfortunately, problem (14) is more challenging than (12) and there is no efficient algorithm to solve it to global optimality. In the Supplementary Material, we evaluate the worst-case error bounds for multiple \\( (\\bar{R},\\bar{t}) \\) samples, select the smallest bounds, and compare them with those of the average pose. An interesting future research direction is to explore differentiable optimization [73] or bilevel polynomial optimization [66] to solve (14)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.49, + 0.21, + 0.507 + ], + "angle": 0, + "content": "6. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.515, + 0.47, + 0.636 + ], + "angle": 0, + "content": "We test our approach on the LineMOD Occlusion (LM-O) dataset [11] to (i) justify the exchangeability assumption (Theorem 1) and suggest best practices for applying conformal prediction; (ii) evaluate the empirical coverage of the PURSE and verify the correctness of Theorem 1, and (iii) compute the worst-case error bounds and demonstrate tightness or looseness. We also (iv) show that the average pose achieves better or similar accuracy as other approaches." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.636, + 0.471, + 0.773 + ], + "angle": 0, + "content": "Implementation and runtime. We set \\( T = 1000 \\) in RANSAG; use OpenGV [44] for P3P and PnP; and add a redundant \\( \\| t \\| \\leq 5 \\) in (PURSE) to ensure bounded translation. All procedures are implemented in Python except SDP relaxations are implemented in Matlab. The runtime of RANSAG is comparable to RANSAC and below one second. The runtime of computing (12) via SDPs is around 8 seconds on a workstation with 2.2GHz AMD CPUs. The (second-order) SDP relaxations are almost always exact." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.779, + 0.47, + 0.863 + ], + "angle": 0, + "content": "6We omit the technical details and refer the interested reader to [98, Section 2] for a pragmatic introduction to SDP relaxations. In practice, we use the code provided by [98] in https://github.com/MIT-SPARK/CertifiablyRobustPerception, apply a second-order SDP relaxation to (12), and use MOSEK [65] to solve the SDP (in about 8 seconds). Solving a first-order SDP relaxation of (12) takes about 0.1 second but yields looser bounds." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.863, + 0.471, + 0.901 + ], + "angle": 0, + "content": "The PURSE can potentially be empty, leading to infeasibility of problem (12). In such cases, empirically the SDP solver returns \"PRIMAL_INFEASIBLE\" (red squares lying on the \\(y\\)-axis of Fig. 3)." + }, + { + "type": "list", + "bbox": [ + 0.076, + 0.779, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.213 + ], + "angle": 0, + "content": "Dataset and exchangeability. The LM-O dataset contains 1214 test images capturing 8 different objects on a table, of which 200 images were chosen by BOP19'20 [35]. We use the 200 images for calibration and the entire 1214 images for testing. As mentioned in Section 3, if the dataset was collected as a single video sequence under natural motion (e.g., a straight line), then the exchangeability assumption would fail. However, [33] described the data collection:" + }, + { + "type": "text", + "bbox": [ + 0.531, + 0.218, + 0.861, + 0.309 + ], + "angle": 0, + "content": "In order to guarantee a well distributed pose space sampling of the dataset pictures, we uniformly divided the upper hemisphere of the objects into equally distant pieces and took at most one image per piece. As a result, our sequences provide uniformly distributed views ..." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.314, + 0.892, + 0.389 + ], + "angle": 0, + "content": "which indicates the 1214 images are independent (cf. [33, Figs. 5-6]) and therefore exchangeable. This demonstrates a good example for data collection –to equally divide the parameter space and collect one observation per division– so the guarantees offered by conformal prediction are valid." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.391, + 0.893, + 0.632 + ], + "angle": 0, + "content": "Empirical coverage. Our approach conformalizes the heatmaps [79] as (ball) or (ellipse). The implementation8 of [79] uses either groundtruth or Faster RCNN [75] bounding boxes, giving four variants of our approach: groundtruth box plus (ball) or (ellipse) (labels: gt-ball, gt-ellipse), and Faster RCNN box plus (ball) or (ellipse) (labels: frcnn-ball, frcnn-ellipse). Fig. 3 left column shows the empirical coverage (i.e., the percentage of images whose groundtruth poses lie in (PURSE)) of all four variants with \\(\\epsilon = 0.1\\) and \\(\\epsilon = 0.4\\). We see the empirical coverage is around \\(90\\%\\) when \\(\\epsilon = 0.1\\) and around \\(60\\%\\) when \\(\\epsilon = 0.4\\), for all 8 objects. Though the empirical coverage can deviate from \\(1 - \\epsilon\\), it generally stays within \\(\\pm 5\\%\\) and mostly goes above \\(1 - \\epsilon\\), which is encouraging given that our calibration set only has size \\(n = 200\\). Fig. 1 (b) plots examples of the prediction sets. More examples are shown in the Supplementary Material." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.633, + 0.895, + 0.875 + ], + "angle": 0, + "content": "Worst-case error bounds. Fig. 3 middle column plots the worst-case rotation error bound (\\(x\\)-axis) vs. the actual rotation error between the average pose and the groundtruth (\\(y\\)-axis) for our approach using the gt-ball setup (results for gt-ellipse, fccnn-ball and fccnn-ellipse are similar and provided in the Supplementary Material). First, when the PURSE covers the groundtruth (blue circles), the rotation error bound is always larger than the actual error (i.e., the blue circles never cross the \\(y = x\\) diagonal). Second, when the error rate is increased from \\(\\epsilon = 0.1\\) to \\(\\epsilon = 0.4\\), we observe a shift of the blue circles towards \\(y = x\\), indicating the error bounds get tightened. Third, our bounds are reasonably tight for most test images (i.e., the bottom-left cluster of blue circles) especially when \\(\\epsilon = 0.4\\). However, they can become overly conservative (i.e., the line of blue circles on the right-side boundary) due to the keypoint prediction" + }, + { + "type": "page_footnote", + "bbox": [ + 0.516, + 0.887, + 0.813, + 0.901 + ], + "angle": 0, + "content": "8https://github.com/yufu-wang/6D_Pose" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.947, + 0.516, + 0.959 + ], + "angle": 0, + "content": "8953" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.145, + 0.06, + 0.396, + 0.211 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.41, + 0.06, + 0.614, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.626, + 0.06, + 0.825, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.142, + 0.213, + 0.396, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.412, + 0.215, + 0.61, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.626, + 0.215, + 0.822, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.368, + 0.892, + 0.451 + ], + "angle": 0, + "content": "Figure 3. Empirical coverage (left) and worst-case error bounds (middle: rotation, right: translation). Top: \\(\\epsilon = 0.1\\), bottom: \\(\\epsilon = 0.4\\). For middle and right columns, \\(x\\)-axis represents the worst-case error bounds computed from (12), \\(y\\)-axis represents the actual error between average pose and groundtruth pose. The area below the diagonal \\(y = x\\) indicates correctness of the bounds (i.e., bound \\(\\geq\\) error), and points that are closer to the diagonal from below indicate tighter bounds (perfect if precisely lie on the diagonal). Blue circles plot cases where the PURSE covers the groundtruth pose and red squares plot cases were the PURSE does not cover the groundtruth. Notice that blue circles never cross the diagonal and our bounds are correct when the PURSE contains the pose (which holds with \\(1 - \\epsilon\\) marginal probability)." + }, + { + "type": "table", + "bbox": [ + 0.117, + 0.452, + 0.851, + 0.595 + ], + "angle": 0, + "content": "
objectsBaselines (results adapted from [72])Conformalized heatmap
Tekin [89]PoseCNN [95]Oberweger [67]PVNet [72]gt-ballgt-ellipsefrcnn-ballfrcnn-ellipse
ε = 0.1ε = 0.4ε = 0.1ε = 0.4ε = 0.1ε = 0.4ε = 0.1
ape7.0134.669.669.1477.7079.5279.2679.8870.2071.0168.84
can11.2015.1082.6086.0973.4175.9775.8178.1367.5269.8167.69
cat3.6210.4065.1065.1287.3690.5989.5490.1174.9580.2368.98
duck5.0731.8061.4061.4482.7183.0884.0283.5579.3080.6280.06
driller1.407.4073.8073.0679.3282.5481.2282.0458.4865.9258.06
eggbox-1.9013.108.43000.090.18000
glue4.7013.8054.9055.3756.4971.0871.6972.9330.0347.1841.96
holepuncher8.2623.1066.4069.8481.6582.8983.2284.3074.9677.8576.28
average6.1617.2060.9061.0667.3370.7170.6171.3956.9361.5857.73
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.596, + 0.89, + 0.625 + ], + "angle": 0, + "content": "Table 1. Success rates of baseline methods and our conformalized heatmap (using the average pose) based on the 2D projection metric (i.e., a pose estimation is considered successful if the average 2D reprojection error is below 5 pixels)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.469, + 0.719 + ], + "angle": 0, + "content": "sets become too large. Fig. 3 right column plots similar results for the translation. The Supplementary Material gives a more detailed analysis of this conservatism, wherein we also solve (12) for multiple samples computed by RANSAG, choose the minimum bound, and compare them with those obtained for the average pose (cf. Remark 4)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.47, + 0.841 + ], + "angle": 0, + "content": "Accuracy of the average pose. We compare the accuracy of our average pose with other methods according to the 2D projection metric (an estimation is correct if the mean reprojection error is below 5 pixels). Table 1 shows: (i) our average pose achieves significantly better success rates when using groundtruth bounding boxes, and similar success rates when using Faster RCNN; (ii) the accuracy of the average pose increases when \\(\\epsilon\\) increases." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.849, + 0.205, + 0.864 + ], + "angle": 0, + "content": "7. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "We applied inductive conformal prediction to conformalize heatmap predictions as circular or elliptical prediction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.892, + 0.75 + ], + "angle": 0, + "content": "sets that guarantee probabilistic coverage of the groundtruth keypoints, propagated the uncertainty in keypoints to the object pose to form a PURSE, designed RANSAG to sample from PURSE and compute an average pose, and used SDP relaxations to bound worst-case estimation errors. We validated our theory on the LineMOD Occlusion dataset. Future research will investigate better nonconformity functions, and applications to other vision problems." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.764, + 0.661, + 0.781 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.788, + 0.892, + 0.895 + ], + "angle": 0, + "content": "We thank Yufu Wang and Kostas Daniilidis for providing the heatmap implementation to detect semantic keypoints; Rachel Luo for discussing the exchangeability assumption in conformal prediction; Luca Carlone for pointing out related work on unknown-but-bounded noise (set membership) estimation in control theory, and anonymous reviewers for providing valuable feedback." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.947, + 0.516, + 0.959 + ], + "angle": 0, + "content": "8954" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.116, + 0.469, + 0.171 + ], + "angle": 0, + "content": "[1] Anastasios Angelopoulos, Stephen Bates, Jitendra Malik, and Michael I Jordan. Uncertainty sets for image classifiers using conformal prediction. In Intl. Conf. on Learning Representations (ICLR), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.173, + 0.469, + 0.227 + ], + "angle": 0, + "content": "[2] Anastasios N Angelopoulos and Stephen Bates. A gentle introduction to conformal prediction and distribution-free uncertainty quantification. arXiv preprint arXiv:2107.07511, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.231, + 0.469, + 0.271 + ], + "angle": 0, + "content": "[3] Anastasios N Angelopoulos, Stephen Bates, Adam Fisch, Lihua Lei, and Tal Schuster. Conformal risk control. arXiv preprint arXiv:2208.02814, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.274, + 0.469, + 0.327 + ], + "angle": 0, + "content": "[4] Pasquale Antonante, Vasileios Tzoumas, Heng Yang, and Luca Carlone. Outlier-robust estimation: Hardness, minimally tuned algorithms, and applications. IEEE Trans. Robotics, 38(1):281-301, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.331, + 0.469, + 0.384 + ], + "angle": 0, + "content": "[5] Boaz Barak and David Steurer. Proofs, beliefs, and algorithms through the lens of sum-of-squares. Course notes: http://www-sumofsquares.org/public/index.html, 1, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.388, + 0.469, + 0.414 + ], + "angle": 0, + "content": "[6] Timothy D Barfoot. State estimation for robotics. Cambridge University Press, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.417, + 0.469, + 0.47 + ], + "angle": 0, + "content": "[7] Jonathan T Barron. A general and adaptive robust loss function. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4331-4339, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.474, + 0.469, + 0.528 + ], + "angle": 0, + "content": "[8] Stephen Bates, Anastasios Angelopoulos, Lihua Lei, Jitendra Malik, and Michael Jordan. Distribution-free, risk-controlling prediction sets. Journal of the ACM (JACM), 68(6):1-34, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.531, + 0.469, + 0.585 + ], + "angle": 0, + "content": "[9] Michael J Black and Anand Rangarajan. On the unification of line processes, outlier rejection, and robust statistics with applications in early vision. *Intl. J. of Computer Vision*, 19(1):57-91, 1996. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.588, + 0.469, + 0.615 + ], + "angle": 0, + "content": "[10] Andrew Blake and Andrew Zisserman. Visual reconstruction. MIT press, 1987. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.618, + 0.469, + 0.686 + ], + "angle": 0, + "content": "[11] Eric Brachmann, Alexander Krull, Frank Michel, Stefan Gumhold, Jamie Shotton, and Carsten Rother. Learning 6d object pose estimation using 3d object coordinates. In European Conf. on Computer Vision (ECCV), pages 536-551. Springer, 2014. 2, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.688, + 0.469, + 0.757 + ], + "angle": 0, + "content": "[12] Eric Brachmann, Frank Michel, Alexander Krull, Michael Ying Yang, Stefan Gumhold, et al. Uncertainty-driven 6d pose estimation of objects and scenes from a single rgb image. In IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 3364-3372, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.76, + 0.469, + 0.814 + ], + "angle": 0, + "content": "[13] Eric Brachmann and Carsten Rother. Learning less is more-6d camera localization via 3d surface regression. In IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 4654–4662, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.817, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[14] Jesus Briales, Laurent Kneip, and Javier Gonzalez-Jimenez. A certifiably globally optimal solution to the non-minimal relative pose problem. In IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 145-154, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.874, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[15] Dylan Campbell, Liu Liu, and Stephen Gould. Solving the blind perspective-n-point problem end-to-end with robust" + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.116, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.541, + 0.093, + 0.892, + 0.133 + ], + "angle": 0, + "content": "differentiable geometric optimization. In European Conf. on Computer Vision (ECCV), pages 244-261. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.137, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[16] Dylan Campbell, Lars Petersson, Laurent Kneip, and Hong-dong Li. Globally-optimal inlier set maximisation for simultaneous camera pose and feature correspondence. In Intl. Conf. on Computer Vision (ICCV), pages 1-10, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.209, + 0.892, + 0.277 + ], + "angle": 0, + "content": "[17] Emmanuel J Candes, Justin K Romberg, and Terence Tao. Stable signal recovery from incomplete and inaccurate measurements. Communications on Pure and Applied Mathematics: A Journal Issued by the Courant Institute of Mathematical Sciences, 59(8):1207-1223, 2006. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.281, + 0.892, + 0.321 + ], + "angle": 0, + "content": "[18] Luca Carlone. Estimation contracts for outlier-robust geometric perception. arXiv preprint arXiv:2208.10521, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.325, + 0.892, + 0.394 + ], + "angle": 0, + "content": "[19] Bo Chen, Jiewei Cao, Alvaro Parra, and Tat-Jun Chin. Satellite pose estimation with deep landmark regression and nonlinear pose refinement. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 0–0, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.397, + 0.892, + 0.465 + ], + "angle": 0, + "content": "[20] Bo Chen, Alvaro Parra, Jiewei Cao, Nan Li, and Tat-Jun Chin. End-to-end learnable geometric vision by backpropagating pnp optimization. In IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 8100-8109, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.469, + 0.892, + 0.538 + ], + "angle": 0, + "content": "[21] Hansheng Chen, Pichao Wang, Fan Wang, Wei Tian, Lu Xiong, and Hao Li. Epro-pnp: Generalized end-to-end probabilistic perspective-n-points for monocular object pose estimation. In IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 2781-2790, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.541, + 0.892, + 0.582 + ], + "angle": 0, + "content": "[22] Tat-Jun Chin, Zhipeng Cai, and Frank Neumann. Robust fitting in computer vision: Easy or hard? In European Conf. on Computer Vision (ECCV), pages 701-716, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.585, + 0.892, + 0.668 + ], + "angle": 0, + "content": "[23] Florence de Grancey, Jean-Luc Adam, Lucian Alecu, Sébastien Gerchinovitz, Franck Mamalet, and David Vigouroux. Object detection with probabilistic guarantees: A conformal prediction approach. In International Conference on Computer Safety, Reliability, and Security, pages 316-329. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.671, + 0.892, + 0.727 + ], + "angle": 0, + "content": "[24] Ilias Diakonikolas, Daniel M Kane, Sushrut Karmalkar, Ankit Pensia, and Thanasis Pittas. List-decodable sparse mean estimation via difference-of-pairs filtering. arXiv preprint arXiv:2206.05245, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.73, + 0.892, + 0.783 + ], + "angle": 0, + "content": "[25] Yonina C Eldar, Amir Beck, and Marc Teboulle. A minimax chebyshev estimator for bounded error estimation. IEEE transactions on signal processing, 56(4):1388-1397, 2008. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.787, + 0.892, + 0.842 + ], + "angle": 0, + "content": "[26] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.846, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[27] Xiao-Shan Gao, Xiao-Rong Hou, Jianliang Tang, and Hang-Fei Cheng. Complete solution classification for the perspective-three-point problem. IEEE Trans. Pattern Anal. Machine Intell., 25(8):930–943, 2003. 2, 6" + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.948, + 0.515, + 0.959 + ], + "angle": 0, + "content": "8955" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.093, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[28] Myung Geun Kim. Multivariate outliers and decompositions of mahalanobis distance. Communications in statistics-theory and methods, 29(7):1511-1526, 2000. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.136, + 0.471, + 0.19 + ], + "angle": 0, + "content": "[29] Chunhui Gu and Xiaofeng Ren. Discriminative mixture-of-templates for viewpoint classification. In European Conf. on Computer Vision (ECCV), pages 408–421. Springer, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.192, + 0.471, + 0.234 + ], + "angle": 0, + "content": "[30] Richard Hartley, Jochen Trumpf, Yuchao Dai, and Hong-dong Li. Rotation averaging. International journal of computer vision, 103:Intl. J. of Computer Vision, 2013. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.236, + 0.471, + 0.276 + ], + "angle": 0, + "content": "[31] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. Cambridge university press, 2003. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.278, + 0.471, + 0.348 + ], + "angle": 0, + "content": "[32] Stefan Hinterstoisser, Cedric Cagniart, Slobodan Ilic, Peter Sturm, Nassir Navab, Pascal Fua, and Vincent Lepetit. Gradient response maps for real-time detection of textureless objects. IEEE Trans. Pattern Anal. Machine Intell., 34(5):876-888, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.349, + 0.471, + 0.431 + ], + "angle": 0, + "content": "[33] Stefan Hinterstoisser, Vincent Lepetit, Slobodan Ilic, Stefan Holzer, Gary Bradski, Kurt Konolige, and Nassir Navab. Model based training, detection and pose estimation of texture-less 3d objects in heavily cluttered scenes. In Asian conference on computer vision, pages 548-562. Springer, 2012. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.433, + 0.471, + 0.489 + ], + "angle": 0, + "content": "[34] Tomas Hodan, Daniel Barath, and Jiri Matas. Epos: Estimating 6d pose of objects with symmetries. In IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 11703-11712, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.491, + 0.471, + 0.56 + ], + "angle": 0, + "content": "[35] Tomas Hodan, Frank Michel, Eric Brachmann, Wadim Kehl, Anders GlentBuch, Dirk Kraft, Bertram Drost, Joel Vidal, Stephan Ihrke, Xenophon Zabulis, et al. Bop: Benchmark for 6d object pose estimation. In European Conf. on Computer Vision (ECCV), pages 19-34, 2018. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.562, + 0.471, + 0.616 + ], + "angle": 0, + "content": "[36] Daniel P Huttenlocher, Gregory A. Klanderman, and William J Rucklidge. Comparing images using the hausdorff distance. IEEE Trans. Pattern Anal. Machine Intell., 15(9):850-863, 1993. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.618, + 0.471, + 0.673 + ], + "angle": 0, + "content": "[37] Shun Iwase, Xingyu Liu, Rawal Khirodkar, Rio Yokota, and Kris M Kitani. Repose: Fast 6d object pose refinement via deep texture rendering. In Intl. Conf. on Computer Vision (ICCV), pages 3303-3312, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.675, + 0.471, + 0.757 + ], + "angle": 0, + "content": "[38] Yanmei Jiao, Yue Wang, Bo Fu, Qimeng Tan, Lei Chen, Minhang Wang, Shoudong Huang, and Rong Xiong. Globally optimal consensus maximization for robust visual inertial localization in point and line map. In IEEE/RSJ Intl. Conf. on Intelligent Robots and Systems (IROS), pages 4631-4638. IEEE, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.759, + 0.471, + 0.8 + ], + "angle": 0, + "content": "[39] Fredrik Kahl and Richard Hartley. Multiple-view geometry under the 1 infinity norm. IEEE Trans. Pattern Anal. Machine Intell., 30(9):1603-1617, 2008. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.802, + 0.471, + 0.844 + ], + "angle": 0, + "content": "[40] Fredrik Kahl and Didier Henrion. Globally optimal estimates for geometric reconstruction problems. *Intl. J. of Computer Vision*, 74(1):3-15, 2007. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.846, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[41] Alex Kendall, Matthew Grimes, and Roberto Cipolla. Posenet: A convolutional network for real-time 6-dof camera relocalization. In Intl. Conf. on Computer Vision (ICCV), pages 2938-2946, 2015. 2" + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.093, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[42] Georg Klein and David Murray. Parallel tracking and mapping for small ar workspaces. In 2007 6th IEEE and ACM international symposium on mixed and augmented reality, pages 225-234. IEEE, 2007. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.149, + 0.892, + 0.203 + ], + "angle": 0, + "content": "[43] Adam Klivans, Pravesh K Kothari, and Raghu Meka. Efficient algorithms for outlier-robust regression. In Conference On Learning Theory, pages 1420–1430. PMLR, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.205, + 0.892, + 0.259 + ], + "angle": 0, + "content": "[44] Laurent Kneip and Paul Furgale. Opengv: A unified and generalized approach to real-time calibrated geometric vision. In IEEE Intl. Conf. on Robotics and Automation (ICRA), pages 1-8. IEEE, 2014. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.261, + 0.892, + 0.315 + ], + "angle": 0, + "content": "[45] Laurent Kneip, Hongdong Li, and Yongduek Seo. Upnp: An optimal o(n) solution to the absolute pose problem with universal applicability. In European Conf. on Computer Vision (ECCV), pages 127-142. Springer, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.316, + 0.892, + 0.397 + ], + "angle": 0, + "content": "[46] Laurent Kneip, Davide Scaramuzza, and Roland Siegwart. A novel parametrization of the perspective-three-point problem for a direct computation of absolute camera position and orientation. In IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 2969-2976. IEEE, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.4, + 0.892, + 0.455 + ], + "angle": 0, + "content": "[47] Yann Labbe, Justin Carpentier, Mathieu Aubry, and Josef Sivic. Cosypos: Consistent multi-view multi-object 6d pose estimation. In European Conf. on Computer Vision (ECCV), pages 574-591. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.456, + 0.892, + 0.509 + ], + "angle": 0, + "content": "[48] Jing Lei, Max G'Sell, Alessandro Rinaldo, Ryan J Tibshirani, and Larry Wasserman. Distribution-free predictive inference for regression. Journal of the American Statistical Association, 113(523):1094-1111, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.511, + 0.892, + 0.552 + ], + "angle": 0, + "content": "[49] Jing Lei, James Robins, and Larry Wasserman. Distribution-free prediction sets. Journal of the American Statistical Association, 108(501):278-287, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.553, + 0.892, + 0.606 + ], + "angle": 0, + "content": "[50] Vincent Lepetit, Pascal Fua, et al. Monocular model-based 3d tracking of rigid objects: A survey. Foundations and Trends in Computer Graphics and Vision, 1(1):1-89, 2005. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.608, + 0.892, + 0.65 + ], + "angle": 0, + "content": "[51] Vincent Lepetit, Francesc Moreno-Noguer, and Pascal Fua. Epnp: An accurate o(n) solution to the pnp problem. *Intl. J. of Computer Vision*, 81(2):155-166, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.651, + 0.892, + 0.704 + ], + "angle": 0, + "content": "[52] Hongdong Li. Consensus set maximization with guaranteed global optimality for robust geometry estimation. In Intl. Conf. on Computer Vision (ICCV), pages 1074-1080. IEEE, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.706, + 0.892, + 0.748 + ], + "angle": 0, + "content": "[53] Shuo Li, Sangdon Park, Xiayan Ji, Insup Lee, and Osbert Bastani. Towards pac multi-object detection and tracking. arXiv preprint arXiv:2204.07482, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.749, + 0.892, + 0.803 + ], + "angle": 0, + "content": "[54] Zhigang Li, Gu Wang, and Xiangyang Ji. Cdpn: Coordinates-based disentangled pose network for real-time rgb-based 6-dof object pose estimation. In Intl. Conf. on Computer Vision (ICCV), pages 7678-7687, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.804, + 0.892, + 0.859 + ], + "angle": 0, + "content": "[55] Joerg Liebelt, Cordelia Schmid, and Klaus Schertler. Independent object class detection using 3d feature maps. In IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 1-8. IEEE, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.86, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[56] David G Lowe. Object recognition from local scale-invariant features. In Intl. Conf. on Computer Vision (ICCV), volume 2, pages 1150-1157. IEEE, 1999. 2" + }, + { + "type": "list", + "bbox": [ + 0.51, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.948, + 0.516, + 0.959 + ], + "angle": 0, + "content": "8956" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.093, + 0.47, + 0.148 + ], + "angle": 0, + "content": "[57] Rachel Luo, Shengjia Zhao, Jonathan Kuck, Boris Ivanovic, Silvio Savarese, Edward Schmerling, and Marco Pavone. Sample-efficient safety assurances using conformal prediction. arXiv preprint arXiv:2109.14082, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.149, + 0.47, + 0.191 + ], + "angle": 0, + "content": "[58] Yi Ma, Stefano Soatto, Jana Košecka, and Shankar Sastry. An invitation to 3-d vision: from images to geometric models, volume 26. Springer, 2004. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.192, + 0.47, + 0.233 + ], + "angle": 0, + "content": "[59] Prasanta Chandra Mahalanobis. On the generalized distance in statistics. National Institute of Science of India, 1936. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.235, + 0.471, + 0.29 + ], + "angle": 0, + "content": "[60] Lucas Manuelli, Wei Gao, Peter Florence, and Russ Tedrake. kpm: Keypoint affordances for category-level robotic manipulation. In The International Symposium of Robotics Research, pages 132-157. Springer, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.292, + 0.471, + 0.347 + ], + "angle": 0, + "content": "[61] Maria Cecilia Mazzaro and Mario Sznaier. A set-membership approach to blind identification. In IEEE Conf. on Decision and Control (CDC), volume 5, pages 5176-5181. IEEE, 2004. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.349, + 0.47, + 0.417 + ], + "angle": 0, + "content": "[62] Frank Michel, Alexander Kirillov, Eric Brachmann, Alexander Krull, Stefan Gumhold, Bogdan Savchynskyy, and Carsten Rother. Global hypothesis generation for 6d object pose estimation. In IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 462-471, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.419, + 0.47, + 0.473 + ], + "angle": 0, + "content": "[63] Mario Milanese and Antonio Vicino. Optimal estimation theory for dynamic systems with set membership uncertainty: An overview. Automatica, 27(6):997-1009, 1991. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.476, + 0.47, + 0.516 + ], + "angle": 0, + "content": "[64] Ankur Moitra. Sum of squares in theoretical computer science. Sum of Squares: Theory and Applications, 77:83, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.518, + 0.47, + 0.546 + ], + "angle": 0, + "content": "[65] MOSEK ApS. The MOSEK optimization toolbox for MATLAB manual. Version 8.1., 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.548, + 0.47, + 0.589 + ], + "angle": 0, + "content": "[66] Jiawang Nie, Li Wang, and Jane J Ye. Bilevel polynomial programs and semidefinite relaxation methods. SIAM Journal on Optimization, 27(3):1728-1757, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.591, + 0.47, + 0.645 + ], + "angle": 0, + "content": "[67] Markus Oberweger, Mahdi Rad, and Vincent Lepetit. Making deep heatmaps robust to partial occlusions for 3d object pose estimation. In European Conf. on Computer Vision (ECCV), pages 119-134, 2018. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.647, + 0.47, + 0.702 + ], + "angle": 0, + "content": "[68] Carl Olsson, Fredrik Kahl, and Magnus Oskarsson. Optimal estimation of perspective camera pose. In 18th International Conference on Pattern Recognition (ICPR'06), volume 2, pages 5-8. IEEE, 2006. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.704, + 0.47, + 0.744 + ], + "angle": 0, + "content": "[69] Harris Papadopoulos. Inductive conformal prediction: Theory and application to neural networks. In Tools in artificial intelligence. CiteSeer, 2008. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.746, + 0.47, + 0.801 + ], + "angle": 0, + "content": "[70] Kiru Park, Timothy Patten, and Markus Vincze. Pix2pose: Pixel-wise coordinate regression of objects for 6d pose estimation. In Intl. Conf. on Computer Vision (ICCV), pages 7668-7677, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.803, + 0.47, + 0.87 + ], + "angle": 0, + "content": "[71] Georgios Pavlakos, Xiaowei Zhou, Aaron Chan, Konstantinos G Derpanis, and Kostas Daniilidis. 6-DoF object pose from semantic keypoints. In IEEE Intl. Conf. on Robotics and Automation (ICRA), pages 2011–2018. IEEE, 2017. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.873, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[72] Sida Peng, Yuan Liu, Qixing Huang, Xiaowei Zhou, and Hujun Bao. Pvnet: Pixel-wise voting network for 6dof" + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.093, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.541, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "pose estimation. IEEE Trans. Pattern Anal. Machine Intell., 2022. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.122, + 0.892, + 0.217 + ], + "angle": 0, + "content": "[73] Luis Pineda, Taosha Fan, Maurizio Monge, Shobha Venkataraman, Paloma Sodhi, Ricky TQ Chen, Joseph Ortiz, Daniel DeTone, Austin Wang, Stuart Anderson, Jing Dong, Brandon Amos, and Mustafa Mukadam. Theseus: A Library for Differentiable Nonlinear Optimization. In Advances in Neural Information Processing Systems (NIPS), 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.221, + 0.892, + 0.29 + ], + "angle": 0, + "content": "[74] Mahdi Rad and Vincent Lepetit. Bb8: A scalable, accurate, robust to partial occlusion method for predicting the 3d poses of challenging objects without using depth. In Proceedings of the IEEE international conference on computer vision, pages 3828-3836, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.292, + 0.892, + 0.347 + ], + "angle": 0, + "content": "[75] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in Neural Information Processing Systems (NIPS), 28, 2015. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.349, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[76] Yaniv Romano, Matteo Sesia, and Emmanuel Candes. Classification with valid and adaptive coverage. In Advances in Neural Information Processing Systems (NIPS), volume 33, pages 3581-3591, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.405, + 0.892, + 0.46 + ], + "angle": 0, + "content": "[77] David M Rosen, Luca Carlone, Afonso S Bandeira, and John J Leonard. Se-sync: A certifiably correct algorithm for synchronization over the special euclidean group. *Intl. J. of Robotics Research*, 38(2-3):95–125, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.462, + 0.892, + 0.53 + ], + "angle": 0, + "content": "[78] Fred Rothganger, Svetlana Lazebnik, Cordelia Schmid, and Jean Ponce. 3d object modeling and recognition using local affine-invariant image descriptors and multi-view spatial constraints. *Intl. J. of Computer Vision*, 66(3):231-259, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.532, + 0.892, + 0.601 + ], + "angle": 0, + "content": "[79] Karl Schmeckpeper, Philip R Osteen, Yufu Wang, Georgios Pavlakos, Kenneth Chaney, Wyatt Jordan, Xiaowei Zhou, Konstantinos G Derpanis, and Kostas Daniilidis. Semantic keypoint-based pose estimation from single RGB frames. J. of Field Robotics, 2022. 1, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.603, + 0.892, + 0.645 + ], + "angle": 0, + "content": "[80] Jingnan Shi, Heng Yang, and Luca Carlone. Optimal pose and shape estimation for category-level 3d object perception. In Robotics: Science and Systems (RSS), 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.647, + 0.892, + 0.701 + ], + "angle": 0, + "content": "[81] Jingnan Shi, Heng Yang, and Luca Carlone. Optimal and robust category-level perception: Object pose and shape estimation from 2d and 3d semantic keypoints. arXiv preprint arXiv:2206.12498, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.703, + 0.892, + 0.73 + ], + "angle": 0, + "content": "[82] Torsten Söderström. Errors-in-variables methods in system identification. Automatica, 43(6):939–958, 2007. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.733, + 0.892, + 0.773 + ], + "angle": 0, + "content": "[83] David Stutz, Ali Taylan Cemgil, Arnaud Doucet, et al. Learning optimal conformal classifiers. In Intl. Conf. on Learning Representations (ICLR), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.776, + 0.892, + 0.831 + ], + "angle": 0, + "content": "[84] Hao Su, Charles R Qi, Yangyan Li, and Leonidas J Guibas. Render for cnn: Viewpoint estimation in images using cnns trained with rendered 3d model views. In Intl. Conf. on Computer Vision (ICCV), pages 2686-2694, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.833, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[85] Jiaming Sun, Zihao Wang, Siyu Zhang, Xingyi He, Hongcheng Zhao, Guofeng Zhang, and Xiaowei Zhou. Onepose: One-shot object pose estimation without cad models. In IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 6825-6834, 2022. 1" + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.948, + 0.516, + 0.959 + ], + "angle": 0, + "content": "8957" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.092, + 0.47, + 0.148 + ], + "angle": 0, + "content": "[86] Min Sun, Gary Bradski, Bing-Xin Xu, and Silvio Savarese. Depth-encoded hough voting for joint object detection and shape recovery. In European Conf. on Computer Vision (ECCV), pages 658-671. Springer, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.149, + 0.472, + 0.218 + ], + "angle": 0, + "content": "[87] Martin Sundermeyer, Zoltan-Csaba Marton, Maximilian Durner, Manuel Brucker, and Rudolph Triebel. Implicit 3d orientation learning for 6d object detection from rgb images. In European Conf. on Computer Vision (ECCV), pages 699-715, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.218, + 0.472, + 0.247 + ], + "angle": 0, + "content": "[88] Richard Szeliski. Computer vision: algorithms and applications. Springer Nature, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.247, + 0.472, + 0.302 + ], + "angle": 0, + "content": "[89] Bugra Tekin, Sudipta N Sinha, and Pascal Fua. Real-time seamless single shot 6d object pose prediction. In IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 292-301, 2018. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.302, + 0.472, + 0.357 + ], + "angle": 0, + "content": "[90] Shubham Tulsiani and Jitendra Malik. Viewpoints and keypoints. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1510-1519, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.359, + 0.472, + 0.399 + ], + "angle": 0, + "content": "[91] Vladimir Vovk. Conditional validity of inductive conformal predictors. In Asian conference on machine learning, pages 475-490. PMLR, 2012. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.4, + 0.472, + 0.44 + ], + "angle": 0, + "content": "[92] Vladimir Vovk, Alexander Gammerman, and Glenn Shafer. Algorithmic learning in a random world. Springer Science & Business Media, 2005. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.442, + 0.472, + 0.511 + ], + "angle": 0, + "content": "[93] Gu Wang, Fabian Manhardt, Federico Tombari, and Xi-angyang Ji. Gdr-net: Geometry-guided direct regression network for monocular 6d object pose estimation. In IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 16611-16621, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.511, + 0.472, + 0.594 + ], + "angle": 0, + "content": "[94] Håkan Wieslander, Philip J Harrison, Gabriel Skogberg, Sonya Jackson, Markus Fridén, Johan Karlsson, Ola Spjuth, and Carolina Wahlby. Deep learning with conformal prediction for hierarchical analysis of large-scale whole-slide tissue images. IEEE journal of biomedical and health informatics, 25(2):371-380, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.595, + 0.472, + 0.65 + ], + "angle": 0, + "content": "[95] Yu Xiang, Tanner Schmidt, Venkatraman Narayanan, and Dieter Fox. Posecnn: A convolutional neural network for 6d object pose estimation in cluttered scenes. In Robotics: Science and Systems (RSS), 2018. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.651, + 0.472, + 0.718 + ], + "angle": 0, + "content": "[96] Heng Yang, Pasquale Antonante, Vasileios Tzoumas, and Luca Carlone. Graduated non-convexity for robust spatial perception: From non-minimal solvers to global outlier rejection. IEEE Robotics and Automation Letters, 5(2):1127-1134, 2020. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.72, + 0.472, + 0.776 + ], + "angle": 0, + "content": "[97] Heng Yang and Luca Carlone. In perfect shape: Certifiably optimal 3d shape reconstruction from 2d landmarks. In IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 621-630, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.776, + 0.472, + 0.831 + ], + "angle": 0, + "content": "[98] Heng Yang and Luca Carlone. Certifiably optimal outlier-robust geometric perception: Semidefinite relaxations and scalable global optimization. IEEE Trans. Pattern Anal. Machine Intell., 2022. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.832, + 0.472, + 0.901 + ], + "angle": 0, + "content": "[99] Heng Yang, Ling Liang, Luca Carlone, and Kim-Chuan Toh. An inexact projected gradient method with rounding and lifting by nonlinear programming for solving rank-one semidefinite relaxation of polynomial optimization. Mathematical Programming, pages 1-64, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[100] Heng Yang, Jingnan Shi, and Luca Carlone. Teaser: Fast and certifiable point cloud registration. IEEE Trans. Robotics, 37(2):314-333, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.178 + ], + "angle": 0, + "content": "[101] Sergey Zakharov, Ivan Shugurov, and Slobodan Ilic. Dpod: 6d pose object detector and refiner. In Intl. Conf. on Computer Vision (ICCV), pages 1941-1950, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.234 + ], + "angle": 0, + "content": "[102] Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. In IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 5745-5753, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.948, + 0.516, + 0.959 + ], + "angle": 0, + "content": "8958" + } + ] +] \ No newline at end of file diff --git a/2023/Object Pose Estimation With Statistical Guarantees_ Conformal Keypoint Detection and Geometric Uncertainty Propagation/a86330a1-e4d7-4fd5-a6ce-7c82335913b8_origin.pdf b/2023/Object Pose Estimation With Statistical Guarantees_ Conformal Keypoint Detection and Geometric Uncertainty Propagation/a86330a1-e4d7-4fd5-a6ce-7c82335913b8_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4530212cf21cdcfdee3e806da224f064faf536d1 --- /dev/null +++ b/2023/Object Pose Estimation With Statistical Guarantees_ Conformal Keypoint Detection and Geometric Uncertainty Propagation/a86330a1-e4d7-4fd5-a6ce-7c82335913b8_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:30f711eb834661ad9a7e8bc524ce501d14734c7ff6ae51ae3cfcea55756743aa +size 7487557 diff --git a/2023/Object Pose Estimation With Statistical Guarantees_ Conformal Keypoint Detection and Geometric Uncertainty Propagation/full.md b/2023/Object Pose Estimation With Statistical Guarantees_ Conformal Keypoint Detection and Geometric Uncertainty Propagation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..3cf668ca74a76e9fbfca6aab96b6777c3cd2559d --- /dev/null +++ b/2023/Object Pose Estimation With Statistical Guarantees_ Conformal Keypoint Detection and Geometric Uncertainty Propagation/full.md @@ -0,0 +1,411 @@ +# Object Pose Estimation with Statistical Guarantees: Conformal Keypoint Detection and Geometric Uncertainty Propagation + +Heng Yang and Marco Pavone +NVIDIA Research + +![](images/d1ae1164c492c11124d70d6c68b4c42cff9b335a3dc17c948e1bfc7c6ac8a89b.jpg) +Figure 1. Probabilistically correct object pose estimation. Given (a) an input image and heatmap detections of the object semantic keypoints, our method first conformalizes the heatmaps into (b) circular or elliptical prediction sets that guarantee probabilistic coverage of the groundtruth keypoints (e.g., $90\%$ ). Our method then propagates the uncertainty in the keypoints to the object pose, forming (c) a Pose UncleRtainty SEt (PURSE) that contains the groundtruth pose with the same probability. We develop RANdom SAmple averaGing (RANSAG) to sample from PURSE and generate (d) an average pose, and apply semidefinite relaxation to compute (e) worst-case error bounds: the blue duck attains the worst rotation error w.r.t. the (average pose) yellow duck; the red duck attains the worst translation error. Code available: https://github.com/NVlabs/ConformalKeypoint. + +# Abstract + +The two-stage object pose estimation paradigm first detects semantic keypoints on the image and then estimates the 6D pose by minimizing reprojection errors. Despite performing well on standard benchmarks, existing techniques offer no provable guarantees on the quality and uncertainty of the estimation. In this paper, we inject two fundamental changes, namely conformal keypoint detection and geometric uncertainty propagation, into the two-stage paradigm and propose the first pose estimator that endows an estimation with provable and computable worst-case error bounds. On one hand, conformal keypoint detection applies the statistical machinery of inductive conformal prediction to convert heuristic keypoint detections into circular or elliptical prediction sets that cover the groundtruth keypoints with a user-specified marginal probability (e.g., $90\%$ ). Geometric uncertainty propagation, on the other, propagates the geometric constraints on the keypoints to the 6D object pose, leading to a Pose UnceRtainty SEt (PURSE) that guarantees coverage of the groundtruth pose with the same probability. The PURSE, however, is a nonconvex set that does not directly lead to estimated poses and uncertainties. Therefore, we develop RANdom SAmple averaGing (RANSAG) to compute an average pose and apply semidefinite relaxation to upper bound the worst-case errors between the average pose and the groundtruth. On the LineMOD Occlusion dataset we demonstrate: (i) the PURSE covers the groundtruth with valid probabilities; (ii) the worst-case error bounds provide correct uncertainty quantification; and (iii) the average pose achieves better or similar accuracy as representative methods based on sparse keypoints. + +# 1. Introduction + +Estimating object poses from images is a fundamental problem in computer vision and finds extensive applications in augmented reality [42], autonomous driving [80], robotic manipulation [60], and space robotics [19]. One of the most popular paradigms for object pose estimation is a two-stage pipeline [20, 71, 72, 79, 81, 85, 89, 101], where the first stage detects (semantic) keypoints of the objects on the image, and the second stage computes the object pose by solving an optimization known as Perspective- $n$ -Points (PnP) that minimizes reprojection errors of the detected keypoints. + +Safety-critical applications call for provably correct computer vision algorithms. Existing algorithms in the two-stage paradigm (reviewed in Section 2), however, provide few performance guarantees on the quality of the estimated poses, due to three challenges. (C1) It is difficult to ensure the detected keypoints (typically from neural networks) are close to the groundtruth keypoints. In practice, the first stage often outputs keypoints that are arbitrarily wrong, known as outliers. (C2) Robust estimation is employed in the second stage to reject outliers, leading to nonconvex optimizations. Fast heuristics such as RANSAC [26] are widely adopted to find an approximate solution but they cannot guarantee global optimality and often fail without notice. (C3) There is no provably correct uncertainty quantification of the estimation, notably, a formal worst-case error bound between the estimation and the groundtruth. Though recent work [98] proposed convex relaxations to certify global optimality of RANSAC and addressed (C2), it cannot ensure correct estimation as the optimal pose may be far away from + +the correct pose when the keypoints are unreliable. + +Contributions. We propose a two-stage object pose estimation framework with statistical guarantees, illustrated in Fig. 1. Given an input image, we assume a neural network [71] is available to generate heatmap predictions of the object keypoints (Fig. 1(a)). Our framework then proceeds in two stages, namely conformal keypoint detection (Section 4) and geometric uncertainty propagation (Section 5). We first apply the statistical machinery of inductive conformal prediction (introduced in Section 3), with nonconformity functions inspired by the design of residual functions in classical geometric vision [39], to conformalize the heatmaps into circular or elliptical prediction sets—one for each keypoint—that guarantee coverage of the groundtruth keypoints with a user-specified marginal probability (Fig. 1(b)). This provides a simple and general methodology to bound the keypoint prediction errors (i.e., addressing (C1)). Given the keypoint prediction sets, we reformulate the constraints (enforced by the prediction sets) on the keypoints as constraints on the object pose, leading to a Pose UnceRtainty SET (PURSE) that guarantees coverage of the groundtruth pose with the same probability. Fig. 1(c) plots the boundary of an example PURSE (roll, pitch, raw angles for the rotation, and Euclidean coordinates for the translation). The PURSE, however, is an abstract nonconvex set that does not directly admit estimated poses and uncertainty. Therefore, we develop RANdom SAmple averaGing (RANSAG) to compute an average pose (Fig. 1(d)) and employ semidefinite relaxations to upper bound the worst-case rotation and translation errors between the average pose and the groundtruth (Fig. 1(e)). This gives rise to the first kind of computable worst-case probabilistic error bounds for object pose estimation (i.e., addressing (C3)). Our PURSE methodology has connections to the framework of unknown-but-bounded noise estimation in control theory [63], with special provisions to derive the bounds in a statistically principled way and enable efficient computation. + +We test our framework on the LineMOD Occlusion (LM-O) dataset [11] to verify the correctness of the theory (Section 6). First, we empirically show that the PURSE indeed contains the groundtruth pose according to the user-specified probability. Second, we demonstrate the correctness of the worst-case error bounds: when the PURSE contains the groundtruth, our bounds are always larger than, and in many cases close to, the actual errors between the average pose and the groundtruth pose. Third, we benchmark the accuracy of the average pose (coming from RANSAG) with representative two-stage pipelines based on sparse keypoints (e.g., PVNet [72]) and show that the average pose achieves better or similar accuracy. + +Limitations. A drawback of our approach, and conformal prediction in general, is that the size of the prediction sets depends on the nonconformity function (whose design + +can be an art) and may be conservative. Our experiments suggest the bounds are loose when the keypoint prediction sets are large (e.g., giving $180^{\circ}$ rotation bound). We discuss challenges and opportunities in tightening the bounds. + +# 2. Related Work + +Image-based object pose estimation. We categorize object pose estimation into two paradigms: single-stage and two-stage. The latter first detects 2D-3D correspondences and then estimates the object pose via solving a PnP problem, while the former produces poses without intermediate correspondences. (i) Single-stage. Early methods perform pose estimation via template matching [29, 32, 36]. Recently, deep learning-based approaches such as PoseNet [41] and PoseCNN [95] applied CNNs to directly regress poses. A major challenge of pose regression is the nonlinearity of 3D rotations, and motivated formulating regression as classification [84, 87, 90] or designing better rotation representations [47, 102]. It is also popular to predict multiple pose hypotheses followed by voting [55, 62, 86]. (ii) Two-stage. Early research used handcrafted features [50, 56, 78] to establish 2D-3D correspondences and focused on developing algorithms for solving PnP. Notable algorithms include the minimal solver P3P [27, 46] and variants of the nonminimal solver PnP [45, 51, 68, 97]. Outliers (i.e., wrong correspondences) motivated robust estimation based on RANSAC [26], graduated non-convexity [9, 10, 96], branch-and-bound [16, 38, 52], or semidefinite relaxations [98]. Unreliable correspondences soon became the bottleneck and learned correspondences have been predominant. Learned correspondences can be sparse or dense. Sparse methods define a handful of keypoints and predict locations of the keypoints via direct regression [74, 89], probabilistic heatmap [67, 71], or voting [72]. Dense methods [12, 34, 54, 70, 93, 101] regress for each object pixel the coordinates of its corresponding 3D point. Recent literature focuses on end-to-end training via differentiating PnP [13, 15, 20, 21, 37]. Both single-stage and two-stage methods perform well on standard benchmarks [35], but a crucial feature that is missing, especially when deploying computer vision algorithms in safety-critical applications, is that these methods do not provide provably correct uncertainty quantification and formal error bounds w.r.t. the groundtruth (for either the correspondences or the poses). In this paper, we provide rigorous guarantees by applying conformal prediction to an existing keypoint detection method (the heatmap [71]) and leveraging old and new techniques in computer vision to derive formal error bounds. + +Conformal prediction in computer vision. Conformal prediction [92] is a statistical machinery that offers provably correct finite-sample uncertainty quantification without assumptions on the data distribution or the prediction model + +(i.e., offering a set prediction, instead of a point prediction, that guarantees probabilistic coverage of the groundtruth). Inductive conformal prediction [69] is the most popular variant of conformal prediction because it does not require retraining of the prediction models [1, 2, 49]. Applying conformal prediction to computer vision, however, is still in its infancy. Existing works focus on image classification [1, 76], tumor segmentation [3, 8, 94], and bounding box detection [2, 23, 53], which are classification or low-dimensional regression problems. Inspired by these works, our unique contributions in this paper are: (i) we apply conformal prediction to keypoints detection, a high-dimensional regression problem; (ii) we design new nonconformity functions and discuss their connections with classical geometric vision; and (iii) we develop algorithms that propagate the uncertainty after conformal prediction to form prediction sets of 6D poses, which are nonlinear and nonconvex manifold objects. + +Performance guarantees. Pose estimation from 2D-2D, 2D-3D, and 3D-3D correspondences are foundational problems in computer vision textbooks [6, 31, 58, 88] and typically boil down to formulating and solving mathematical optimization problems. Benchmarking on simulated and real datasets has been a widely adopted standard for testing different formulations and solvers. However, empirical performance can be misleading without theoretical guarantees. A striking fact is that, though error analysis is an important topic in applied math [17, 24, 43] and control theory [61, 63, 82], there is very limited literature in computer vision that reason about worst-case estimation errors between the optimal solution and the groundtruth. A popular heuristic relies on the inverse of the Hessian at an optimal solution, which provides the Cramer-Rao lower bound on the covariance of the solution (for linear regression this coincides with the covariance) [88, Section B.6] and thus cannot upper bound the estimation errors. Recent works [18, 77, 100] derived error bounds for a few geometric vision problems. However, the bounds either depend on uncheckable assumptions and cannot be computed [77, 100], or build on machinery (e.g., sum-of-squares proof [5, 64]) that only applies to estimators based on moment relaxations [18], which are still computationally expensive in practice [99]. In this paper, we develop the first kind of efficiently computable error bounds that only require the assumption of exchangeability (which comes from conformal prediction). We justify this assumption on our test dataset and numerically show our bounds can be tight for a subset of the test problems. + +# 3. Inductive Conformal Prediction + +Given a set $\{z_i = (x_i, y_i)\}_{i=1}^l$ with observation $x_i \in \mathcal{X}$ and label $y_i \in \mathcal{Y}$ such that each $z_i \in \mathcal{Z} := \mathcal{X} \times \mathcal{Y}$ is drawn i.i.d. from an unknown distribution on $\mathcal{Z}$ , inductive confor + +mal prediction (ICP) provides a set prediction $F^{\epsilon}(x) \subseteq \mathcal{V}$ , parameterized by an error rate $0 < \epsilon < 1$ , such that given a new sample $z_{l + 1} = (x_{l + 1}, y_{l + 1})$ satisfying an exchangeability condition (elaborated in Theorem 1), we have + +$$ +\mathbb {P} \left[ y _ {l + 1} \in F ^ {\epsilon} \left(x _ {l + 1}\right) \right] \geq 1 - \epsilon , \tag {1} +$$ + +i.e., the prediction set $F^{\epsilon}$ guarantees to contain the true label $y_{l + 1}$ with probability at least $1 - \epsilon$ . + +Training. We start by dividing the dataset into a proper training set $\{z_1, \ldots, z_m\}$ and a calibration set $\{z_{m+1}, \ldots, z_l\}$ . We shorthand $n = l - m$ as the size of the calibration set. We learn a prediction function $f: \mathcal{X} \to \tilde{\mathcal{V}}$ from the proper training set using any architecture, which allows us to fully exploit the power of modern deep learning. The prediction space $\tilde{\mathcal{V}}$ can be the same as the label space $\mathcal{V}$ , or can contain auxiliary information such as a heuristic notion of uncertainty (e.g., softmax scores in classification or a heatmap in the case of keypoint detection). + +Conformal calibration. We define a nonconformity function $S: \mathcal{Z}^m \times \mathcal{Z} \to \mathbb{R}$ to measure how well a given sample $z = (x, y)$ conforms to the proper training set. A popular instance of $S$ leverages the learned prediction $f$ : + +$$ +S \left(\left\{z _ {1}, \dots , z _ {m} \right\}, (x, y)\right) \stackrel {{e. g.}} {{=}} r (y, f (x)), \tag {2} +$$ + +where $r: \mathcal{Y} \times \tilde{\mathcal{Y}} \to \mathbb{R}$ is a measure of disagreement between the label $y$ and the prediction $f(x)$ . For example, consider $\mathcal{Y} = \tilde{\mathcal{Y}} = \mathbb{R}$ , one can design $r(y, f(x)) = |y - f(x)|$ : if $(x, y)$ poorly conforms to the training set, $f$ will incur large errors. While the function $S$ can be arbitrary (e.g., a learnable neural network [83]), (2) is a convenient definition since $f$ is implicitly dependent on $\{z_i\}_{i=1}^m$ and $r$ can incorporate domain-specific knowledge. We then compute the nonconformity scores on the calibration set as $\alpha_i = r(y_i, f(x_i)), i = m+1,\ldots,l$ , and sort them in nonincreasing order $\alpha_{\pi(1)} \geq \dots \geq \alpha_{\pi(n)}$ , where $\pi(i) \in \{m+1,\ldots,l\}$ is an index permutation. + +Conformal prediction. Given a new observation $x_{l+1}$ (with an unknown $y_{l+1}$ ) and a user-specified $\epsilon \in (0, 1)$ , we compute the inductive conformal prediction (ICP) set as + +$$ +F ^ {\epsilon} \left(x _ {l + 1}\right) = \left\{y \in \mathcal {Y} \mid \alpha^ {y} \leq \alpha_ {\pi (\lfloor (n + 1) \epsilon \rfloor)} \right\}, \tag {3} +$$ + +where $\alpha^y = r(y, f(x_{l+1}))$ is the nonconformity score of the new sample when fixing the true label to be $y$ . In other words, the ICP set (3) outputs the set of all labels that make the nonconformity score of the new sample no greater than $\alpha_{\pi(\lfloor n + 1 \epsilon \rfloor)} - \text{the } \lfloor (n + 1) \epsilon \rfloor$ -th largest nonconformity score in the calibration set. We have the following result stating the probabilistic coverage of the ICP set (3). + +Theorem 1 (Validity of ICP Coverage [48, 91, 92]). If $z_{m+1}, \ldots, z_l, z_{l+1} = (x_{l+1}, y_{l+1})$ are exchangeable, i.e., their distribution is invariant under permutation, then + +$$ +1 - \epsilon \leq \mathbb {P} \left[ y _ {l + 1} \in F ^ {\epsilon} \left(x _ {l + 1}\right) \right] \leq 1 - \epsilon + 1 / (n + 1) \tag {4} +$$ + +![](images/79351fc3e0ba435b18162af2bc294190b15a6316fcc94715c40aed14afceb2b4.jpg) +Figure 2. Beta distribution of the conditional coverage in (5) with $\epsilon = 0.1$ and different $n$ . Notice how the conditional probability becomes more concentrated around $1 - \epsilon$ when $n$ increases. + +for any $\epsilon \in (0,1)$ . Furthermore, when conditioned on the calibration set, calling $h = \lfloor (n + 1)\epsilon \rfloor$ , we have + +$$ +\mathbb {P} \left[ y _ {l + 1} \in F ^ {\epsilon} (x _ {l + 1}) \mid \{z _ {m + 1}, \dots , z _ {l} \} \right] \sim \mathrm {B e t a} (n + 1 - h, h). (5) +$$ + +A few remarks are in order about Theorem 1. First, asking $z_{m+1}, \ldots, z_l, z_{l+1}$ to be exchangeable is weaker than asking them to be independent. However, this assumption typically fails when the calibration set is a single video sequence, where the image frames $\{z_{m+1}, \ldots, z_l\}$ are temporally correlated [57]. Fortunately, as we detail in Section 6, the way the LineMOD Occlusion dataset [11] was collected makes the exchangeability condition easily satisfied, which also suggests best practices to make the exchangeability condition hold in computer vision. Second, the lower bound in (4) can be intuitively proved because under exchangeability, $\alpha_{l+1} := r(y_{l+1}, f(x_{l+1}))$ – the nonconformity score of the new sample with the true label– is exchangeable with the nonconformity scores of the calibration samples, and hence equally likely to fall in anywhere between the scores $\{\alpha_{\pi(i)}\}_{i=1}^n$ . Consequently, $\mathbb{P}[y_{l+1} \in F^\epsilon(x_{l+1})] = \mathbb{P}\left[\alpha_{l+1} \leq \alpha_{\pi(\lfloor n+1\epsilon \rfloor)}\right] = 1 - \lfloor(n+1)\epsilon \rfloor / (n+1) \geq 1 - \epsilon$ . The upper bound in (4) states that $1 - \epsilon$ is not overly conservative (indeed tight if $n$ is large). Lastly, the probabilistic guarantee in (4) is marginal over the randomness of the calibration set, meaning if one chooses an infinite number of calibration sets, the average empirical coverage will converge to $1 - \epsilon$ . This, however, implies that the empirical coverage given one calibration set is a random variable that fluctuates as the Beta distribution (5). Fig. 2 plots the Beta distribution at $\epsilon = 0.1$ with different sizes of the calibration set. We observe that as $n$ increases the empirical coverage becomes more concentrated at $1 - \epsilon$ . Our experiments show that even with a small ( $n = 200$ ) calibration set, the empirical coverage is close to, and mostly higher than, $1 - \epsilon$ . + +# 4. Conformal Keypoint Detection + +In this section, we apply the ICP framework in Section 3 to the problem of semantic keypoint detection. + +Setup. Denote by $x \in \mathbb{R}^{H \times W \times 3}$ an RGB image picturing an object, by $\pmb{y} = (y_1, \dots, y_K) \in \mathbb{R}^2 \times \dots \times \mathbb{R}^2 \coloneqq \mathcal{Y}$ . + +the groundtruth locations of $K$ semantic keypoints of the object. We partition a given dataset $\{z_i := (x_i, y_i)\}_{i=1}^l$ into a proper training set (of size $m$ ) and a calibration set (of size $n$ ). We follow the three steps in Section 3 to perform ICP. + +Training. We choose the heatmap approach in [71, 79] as the prediction function: given an image $x$ , [79] outputs a set of heatmaps $\pmb{f}(x) = (f(x)_1, \dots, f(x)_K)$ , where each $f(x)_k \in \Delta^{HW} := \{v \in \mathbb{R}_+^{HW} \mid \sum_i^{HW} v_i = 1\}$ predicts the probability distribution of the $k$ -th keypoint lying on each pixel of the image. For convenience, we use $q^j \in \mathbb{R}^2$ to denote the $j$ -th pixel location in $x$ and $f(x)_k^j \in \mathbb{R}_+$ to denote the probability of the $k$ -th keypoint lying on $q^j$ . Let $\sigma_k$ be the index permutation that sorts $f(x)_k$ in nonincreasing order, i.e., $f(x)_k^{\sigma_k(1)} \geq \dots \geq f(x)_k^{\sigma_k(HW)}$ . As we will soon show, choosing the heatmap approach leads to simple and intuitive designs of the nonconformity function. + +Conformal calibration. We design the following nonconformity function + +$$ +r (\boldsymbol {y}, \boldsymbol {f} (x)) = \max \left\{\phi \left(y _ {k}, f (x) _ {k}\right) \right\} _ {k = 1} ^ {K} \tag {6} +$$ + +that uses $\phi$ to score each keypoint and then selects the maximum score. This design considers the worst keypoint detection performance of $f$ . We provide two designs of $\phi$ below. + +(a) Peak. Shorthand $p_k = f(x)_k^{\sigma_k(1)}$ as the peak probability in the $k$ -th heatmap and $q_k = q^{\sigma_k(1)}$ as the pixel location attaining the peak probability, we design + +$$ +\phi_ {\text {p e a k}} \left(y _ {k}, f (x) _ {k}\right) = p _ {k} \| y _ {k} - q _ {k} \| \quad (\text {p e a k}) +$$ + +which computes the error between the true keypoint location $y_{k}$ and the most probable keypoint location $q_{k}$ and scales the error by the peak probability $p_{k}$ . $\phi_{\mathrm{peak}}$ describes nonconformity because it becomes larger when the network $f$ is confidently wrong (both $\| y_{k} - q_{k}\|$ and $p_k$ are large), implying the sample is highly nonconforming. + +(b) Covariance. Let $\bar{q}_k = \sum_{j=1}^{J} f(x)_k^{\sigma_k(j)} q^{\sigma_k(j)}$ be the expected location of the top- $J$ most likely detections for the $k$ -th keypoint, and $\Sigma_k = \sum_{j=1}^{J} f(x)_k^{\sigma_k(j)} \cdot (q^{\sigma_k(j)} - \bar{q}_k)(q^{\sigma_k(j)} - \bar{q}_k)^{\top}$ as the covariance, we design + +$$ +\phi_ {\operatorname {c o v}} \left(y _ {k}, f (x) _ {k}\right) = \left(y _ {k} - \bar {q} _ {k}\right) ^ {\top} \Sigma_ {k} ^ {- 1} \left(y _ {k} - \bar {q} _ {k}\right) \quad (\operatorname {c o v}) +$$ + +which computes the squared Mahalanobis distance [59] from the groundtruth $y_{k}$ to the top- $J$ keypoint detections (represented by the mean $\bar{q}_{k}$ and covariance $\Sigma_{k}$ ).2 A larger Mahalanobis distance indicates more abnormality of the heatmap $f(x)_k$ (compared to the groundtruth $y_{k}$ ) [28], and hence implies higher nonconformity. + +Using the nonconformity function (6) with (peak) or (cov), we compute the nonconformity scores of the calibration set and sort them as: $\alpha_{\pi(1)} \geq \dots \geq \alpha_{\pi(n)}$ . + +Conformal prediction. Given an error rate $\epsilon \in (0,1)$ , we first find $\alpha_{\pi (\lfloor (n + 1)\epsilon \rfloor)}$ . Then, according to the ICP set definition (3) and our nonconformity function (6), we output the ICP set for a new $x_{l + 1}$ as + +$$ +\begin{array}{l} F ^ {\epsilon} (x _ {l + 1}) \\ = \left\{\boldsymbol {y} \in \mathcal {Y} \mid \max \left\{\phi \left(y _ {k}, f \left(x _ {l + 1}\right)\right) \right\} _ {k = 1} ^ {K} \leq \alpha_ {\pi (\lfloor (n + 1) \epsilon \rfloor)} \right\} \\ = \left\{\boldsymbol {y} \in \mathcal {Y} \mid \phi \left(y _ {k}, f \left(x _ {l + 1}\right)\right) \leq \alpha_ {\pi (\lfloor (n + 1) \epsilon \rfloor)}, \forall k \right\}, \tag {7} \\ \end{array} +$$ + +where we used $\max \{\phi_1,\ldots ,\phi_K\} \leq \alpha$ if and only if $\phi_{k}\leq \alpha$ for any $k$ . Insert (peak) into (7), we have $F_{\mathrm{ball}}^{\epsilon}(x_{l + 1})$ as + +$$ +\left\{\boldsymbol {y} \in \mathcal {Y} \mid \| y _ {k} - q _ {l + 1, k} \| \leq \frac {\alpha_ {\pi (\lfloor (n + 1) \epsilon \rfloor)}}{p _ {l + 1 , k}}, \forall k \right\}, \quad (\text {b a l l}) +$$ + +which defines -for the $k$ -th keypoint- a ball centered at $q_{l + 1,k}$ (the most likely detection) with a radius inversely proportional to $p_{l + 1,k}$ and proportional to $\alpha_{\pi (\lfloor (n + 1)\epsilon \rfloor)}$ . Similarly, insert (cov) into (7), we have $F_{\mathrm{ellipse}}^{\epsilon}(x_{l + 1})$ as + +$$ +\left\{\boldsymbol {y} \in \mathcal {Y} \mid (y _ {k} - \bar {q} _ {l + 1, k}) ^ {\mathsf {T}} \frac {\sum_ {l + 1 , k} ^ {- 1}}{\alpha_ {\pi (\lfloor (n + 1) \epsilon \rfloor)} (y _ {k} - \bar {q} _ {l + 1 , k})} \leq 1, \forall k \right\}, +$$ + +which defines -for the $k$ -th keypoint-- an ellipse centered at $\bar{q}_{l+1,k}$ (the expected location of the top- $J$ detections) with an area proportional to $\operatorname*{det}(\Sigma_{l+1,k})$ and $\alpha_{\pi(\lfloor n+1\epsilon \rfloor)}$ .3 From (ball) and (ellipse), we observe that the prediction sets become larger when (i) the heatmaps are uncertain, i.e., the peak probability is low or the covariance matrix has large determinant; and (ii) the heatmaps perform poorly on the calibration set, leading to a large $\alpha_{\pi(\lfloor n+1\epsilon \rfloor)}$ . + +Connections to geometric vision. Our nonconformity function bears similarity to the residual function in geometric vision [4, 22, 31]. For example, the (peak) and (cov) functions are similar to the (weighted) reprojection error [31], and the "max" in (6) can be connected to seminal work on optimizing the $\ell_{\infty}$ norm [39]. + +Outlier-robust nonconformity? One potential issue of the nonconformity function (6) is that a single outlier can inflate the score and the calibration quantile $\alpha_{\pi (\lfloor (n + 1)\epsilon \rfloor)}$ and lead to conservative prediction sets (e.g., when $\pmb{f}$ predicts $K - 1$ keypoints perfectly but misses one keypoint). A potential remedy in geometric vision is to use robust cost functions [7,9,96]. Therefore, a natural question is whether "robustifying" the nonconformity function (6) can lead to better prediction sets. Here we focus on only robustifying $\phi$ in (6) and provide a negative answer. + +Proposition 2 (Invariance of ICP). Let $\rho : \mathbb{R}_+ \mapsto \mathbb{R}_+$ be any monotonically increasing function. Fixing the calibration set and error rate $\epsilon$ , the nonconformity function + +$$ +r _ {\rho} (\boldsymbol {y}, \boldsymbol {f} (x)) = \max \left\{\rho \left(\phi \left(y _ {k}, f (x) _ {k}\right)\right) \right\} _ {k = 1} ^ {K} \tag {8} +$$ + +leads to the same ICP set as (6). + +The proof of Proposition 2 is presented in Supplementary Material. We conclude that common robust costs, such as $\ell_1$ , Huber, Geman-McClure, and Barron's adaptive kernel [7,9] (which are monotonically increasing on $[0, +\infty]$ ) cannot change the ICP sets by robustifying the individual score $\phi$ . However, it remains an open question whether changing the "max" operation in (6) can give rise to better ICP sets. For instance, replacing "max" with " $\sum$ in (6)" and using the Geman-McClure robust cost $\rho(\phi) = \frac{\phi^2}{1 + \phi^2}$ with $\phi = \phi_{\mathrm{peak}}$ results in the following ICP set + +$$ +\left\{\boldsymbol {y} \in \mathcal {Y} \mid \sum_ {k = 1} ^ {K} \frac {p _ {k} ^ {2} \| y _ {k} - q _ {k} \| ^ {2}}{1 + p _ {k} ^ {2} \| y _ {k} - q _ {k} \| ^ {2}} \leq \alpha_ {\pi (\lfloor (n + 1) \epsilon \rfloor)} \right\} \tag {9} +$$ + +that does not admit a geometric interpretation that is as simple and intuitive as the (ball) and (ellipse) sets introduced before. In fact, it is indeed the simplicity of (ball) and (ellipse) that enables us to propagate the uncertainty in keypoints to the object pose, as we will show in the next section. + +# 5. Geometric Uncertainty Propagation + +Conformalizing the heatmaps gives us prediction sets that guarantee probabilistic coverage of the true keypoints. We unify the prediction sets (ball) and (ellipse) as + +$$ +F ^ {\epsilon} (x) = \left\{\boldsymbol {y} \in \mathcal {Y} \mid \left(y _ {k} - \mu_ {k}\right) ^ {\top} \Lambda_ {k} \left(y _ {k} - \mu_ {k}\right) \leq 1, \forall k \right\}, \tag {10} +$$ + +where $\mu_{k} = q_{l + 1,k},\Lambda_{k} = \frac{p_{l + 1,k}^{2}}{\alpha_{\pi(\lfloor(n + 1)\epsilon \rfloor)}^{2}}\mathbf{I}_{2}$ for (ball), $\mu_{k} =$ $\bar{q}_{l + 1,k},\Lambda_k = \frac{\sum_{l + 1,k}^{-1}}{\alpha_\pi(\lfloor(n + 1)\epsilon \rfloor)}$ for (ellipse), and we omit the subscript $l + 1$ for simplicity. + +Why not uncertainty-aware PnP? A popular way to estimate pose from (10) is to solve an uncertainty-aware PnP + +$$ +\min _ {(R, t) \in \mathrm {S E} (3)} \quad \sum_ {k = 1} ^ {K} \left(y _ {k} - \mu_ {k}\right) ^ {\mathsf {T}} \Lambda_ {k} \left(y _ {k} - \mu_ {k}\right) +$$ + +$$ +\text {s u b j e c t} \quad y _ {k} = \Pi \left(R Y _ {k} + t\right), k = 1, \dots , K \tag {11} +$$ + +where $Y_{k} \in \mathbb{R}^{3}, k = 1, \dots, K$ are the 3D object keypoints and $\Pi(\cdot)$ denotes the camera projection. We challenge this approach and point out its two drawbacks. First, it is difficult to solve (11) to global optimality due to (i) the nonconvex SE(3) constraint and (ii) the rational polynomial appearing in $\Pi(\cdot)$ . The best known approach to solve (11) relies on either branch-and-bound [68] or local optimization. + +Second, solving (11) typically outputs a single optimal pose without uncertainty quantification. Are there other poses that attain similar costs as the optimal pose? How close is the optimal pose to the groundtruth pose? These questions remain not answered in the literature. + +Pose UnceRtainty SET (PURSE). We propose to, instead of solving a PnP problem similar to (11), directly propagate the uncertainty in the ICP sets to the object pose. + +Proposition 3 (PURSE). Let $s_{\mathrm{gt}} = [\mathrm{vec}(R_{\mathrm{gt}})^{\top}; t_{\mathrm{gt}}^{\top}]^{\top}$ be the groundtruth object pose (that lies in front of the camera). Then, the groundtruth keypoints $\mathbf{y} = (y_1, \ldots, y_K)$ belong to the ICP set $F^{\epsilon}(x)$ in (10) if and only if $s_{\mathrm{gt}}$ belongs to the following pose uncertainty set + +$$ +S ^ {\epsilon} = \left\{s \in \operatorname {S E} (3) \left| \begin{array}{c} s ^ {\mathsf {T}} A _ {k} s \leq 0, k = 1, \ldots , K \\ b _ {k} ^ {\mathsf {T}} s > 0, k = 1, \ldots , K \end{array} \right. \right\}, (\text {P U R S E}) +$$ + +where $A_{k}\in \mathbb{S}^{12},b_{k}\in \mathbb{R}^{12},k = 1,\ldots ,K$ are constant matrices dependent on $\mu_k,\Lambda_k,Y_k$ and camera intrinsics. + +The detailed proof for Proposition 3 is algebraically involved and postponed to the Supplementary Material. The high-level intuition is, however, straightforward: we plug in $y_{k} = \Pi (RY_{k} + t)$ into (10) and obtain $K$ quadratic inequalities of the form $s^{\mathsf{T}}A_{k}s\leq 0$ . The linear inequalities $b_{k}^{\top}s > 0$ are added to enforce the (transformed) 3D keypoints lie in front of the camera. Proposition 3 implies, if we are $1 - \epsilon$ confident the groundtruth keypoints can be anywhere inside $F^{\epsilon}(x)$ , then we should also be confident any pose in (PURSE) can be the groundtruth. Viewing pose estimation as a set estimation with guaranteed probabilistic coverage of the groundtruth is fundamentally different from viewing it as computing a single pose from (11) that is (hopefully) close enough to the groundtruth. + +RANdom SExample averaGing (RANSAG). Verifying if a given pose belongs to the PURSE is straightforward via checking the inequalities in (PURSE). However, the PURSE does not directly give us estimated poses. Therefore, we propose an efficient sampling algorithm called RANSDOM SExample averaGing (RANSAG) that is analogous to RANSAC [26] and leverages the minimal solver P3P [27], presented in Algorithm 1. The intuition is that, though it is difficult to sample directly in PURSE due to the (nonconvex) constraints, it is easy to sample from the keypoint prediction set (10) due to its simple geometry (balls and ellipses). Thus, at each iteration (line 3) RANSAG samples three keypoints (line 4-5), solves the P3P inverse problem, and accept the poses that belong to the (PURSE) (line 6). RANSAG typically returns around 100 valid samples with $T = 1000$ trials. However, in difficult cases (e.g., when $S^{\epsilon}$ is small or even empty) it is possible to obtain zero samples ( $S = \emptyset$ ). In this situation, RANSAG samples $\lfloor T / 20 \rfloor$ (default 50) poses without checking if they belong to the PURSE, via sampling + +Algorithm 1: RANdom SAmple averaGing +1 Input: an ICP set $F^{\epsilon}(x)$ (10) and its corresponding (PURSE) $S^{\epsilon}$ ; maximum trials $T$ ; initial $\hat{S} = \emptyset$ ; +2 Output: sample poses $S \subset \mathrm{SE}(3)$ in PURSE, and an average pose $\bar{s} \in \mathrm{SE}(3)$ ; +3 for $\tau \gets 1$ to $T$ do +4 Sample $\{k_1, k_2, k_3\}$ from $[K] (k_1 \neq k_2 \neq k_3)$ ; +5 Sample $\hat{y}_{k_i}, i = 1, 2, 3$ from $\{y \in \mathbb{R}^2 \mid (y - \mu_{k_i})^\top \Lambda_{k_i} (y - \mu_{k_i}) \leq 1\}$ ; +6 $\hat{S} \gets \hat{S} \cup (S^{\epsilon} \cap \mathsf{P3P}(\{\hat{y}_{k_i} \leftrightarrow Y_{k_i}\}_{i=1}^3))$ ; +7 end +8 $S = \hat{S}$ ; +9 if $\hat{S} = \emptyset$ then +10 for $\tau \gets 1$ to $\lfloor T/20 \rfloor$ do +11 Sample $\hat{y}_k, k = 1, \dots, K$ from $F^{\epsilon}(x)$ ; +12 $\hat{S} \gets \hat{S} \cup \mathsf{PnP}(\{\hat{y}_k \leftrightarrow Y_k\}_{k=1}^K)$ ; +13 end +14 end +15 $\bar{R} = \operatorname{proj}_{\mathrm{SO}(3)}(\sum_{(R_j, *) \in \hat{S}} R_j)$ ; +16 $\bar{t} = \frac{1}{|\hat{S}|} \sum_{(*, t_j) \in \hat{S}} t_j$ ; +17 return: $S, \bar{s} = (\bar{R}, \bar{t})$ + +$K$ keypoints and solving PnP (line 9-12). After obtaining a set of poses, RANSAG performs rotation averaging (line 14) and translation averaging (line 15) to obtain an average pose $\bar{s}$ . Note that RANSAG does not check if $\bar{s}$ lies in the PURSE. + +Worst-case error bounds. To upper bound the errors between the average pose $\bar{s}$ and the groundtruth $(R_{\mathrm{gt}},t_{\mathrm{gt}})$ , we maximize the squared pose-to-PURSE distance: + +$$ +d _ {\epsilon , \lambda} ^ {2} = \max _ {(R, t) \in S ^ {\epsilon}} \lambda \| R - \bar {R} \| _ {\mathrm {F}} ^ {2} + (1 - \lambda) \| t - \bar {t} \| ^ {2} \tag {12} +$$ + +given $\lambda \in [0,1]$ . Particularly, we compute two cases $\lambda = 1$ (the maximum rotation distance) and $\lambda = 0$ (the maximum translation distance). Proposition 3 states the groundtruth $(R_{\mathrm{gt}},t_{\mathrm{gt}})$ lies in $S^{\epsilon}$ with $1 - \epsilon$ probability, hence + +$$ +\left\| \bar {R} - R _ {\mathrm {g t}} \right\| _ {\mathrm {F}} \leq d _ {\epsilon , 1}, \quad \left\| \bar {t} - t _ {\mathrm {g t}} \right\| \leq d _ {\epsilon , 0} \tag {13} +$$ + +holds with probability $1 - \epsilon$ . + +Computing the bounds. Problem (12) is nonconvex due to the constraints of the (PURSE) $S^{\epsilon}$ . We relax the nonconvex problem (12) into a convex semidefinite program (SDP) and employ off-the-shelf solvers to optimize + +the SDP [14, 40, 98]. Two possible outcomes can happen: (i) the optimal SDP value coincides with the optimal value of (12). The relaxation is said to be exact and one can extract an optimal solution of (12) from the SDP, or (ii) the relaxation is not exact, but the optimal SDP value still provides an upper bound for the optimal value of (12). Therefore, we either exactly compute $d_{\epsilon, \lambda}^2$ or find an upper bound, both can bound the worst-case error (cf. (13)). + +We end with a remark about computing tighter bounds. + +Remark 4 (Best Worst-case Error Bounds). (12) can be used to bound errors for all possible pose estimators (e.g., from PnP (11)). What is the best estimator that attains the smallest error bounds? This boils down to solving + +$$ +\min _ {(\bar {R}, \bar {t}) \in \mathrm {S E} (3)} \left[ \max _ {(R, t) \in S ^ {\epsilon}} \lambda \| R - \bar {R} \| _ {\mathrm {F}} ^ {2} + (1 - \lambda) \| t - \bar {t} \| ^ {2} \right] \tag {14} +$$ + +whose solution is known as the Chebyshev center [25, 63] of the PURSE $S^{\epsilon}$ . Unfortunately, problem (14) is more challenging than (12) and there is no efficient algorithm to solve it to global optimality. In the Supplementary Material, we evaluate the worst-case error bounds for multiple $(\bar{R},\bar{t})$ samples, select the smallest bounds, and compare them with those of the average pose. An interesting future research direction is to explore differentiable optimization [73] or bilevel polynomial optimization [66] to solve (14). + +# 6. Experiments + +We test our approach on the LineMOD Occlusion (LM-O) dataset [11] to (i) justify the exchangeability assumption (Theorem 1) and suggest best practices for applying conformal prediction; (ii) evaluate the empirical coverage of the PURSE and verify the correctness of Theorem 1, and (iii) compute the worst-case error bounds and demonstrate tightness or looseness. We also (iv) show that the average pose achieves better or similar accuracy as other approaches. + +Implementation and runtime. We set $T = 1000$ in RANSAG; use OpenGV [44] for P3P and PnP; and add a redundant $\| t \| \leq 5$ in (PURSE) to ensure bounded translation. All procedures are implemented in Python except SDP relaxations are implemented in Matlab. The runtime of RANSAG is comparable to RANSAC and below one second. The runtime of computing (12) via SDPs is around 8 seconds on a workstation with 2.2GHz AMD CPUs. The (second-order) SDP relaxations are almost always exact. + +6We omit the technical details and refer the interested reader to [98, Section 2] for a pragmatic introduction to SDP relaxations. In practice, we use the code provided by [98] in https://github.com/MIT-SPARK/CertifiablyRobustPerception, apply a second-order SDP relaxation to (12), and use MOSEK [65] to solve the SDP (in about 8 seconds). Solving a first-order SDP relaxation of (12) takes about 0.1 second but yields looser bounds. +The PURSE can potentially be empty, leading to infeasibility of problem (12). In such cases, empirically the SDP solver returns "PRIMAL_INFEASIBLE" (red squares lying on the $y$ -axis of Fig. 3). + +Dataset and exchangeability. The LM-O dataset contains 1214 test images capturing 8 different objects on a table, of which 200 images were chosen by BOP19'20 [35]. We use the 200 images for calibration and the entire 1214 images for testing. As mentioned in Section 3, if the dataset was collected as a single video sequence under natural motion (e.g., a straight line), then the exchangeability assumption would fail. However, [33] described the data collection: + +In order to guarantee a well distributed pose space sampling of the dataset pictures, we uniformly divided the upper hemisphere of the objects into equally distant pieces and took at most one image per piece. As a result, our sequences provide uniformly distributed views ... + +which indicates the 1214 images are independent (cf. [33, Figs. 5-6]) and therefore exchangeable. This demonstrates a good example for data collection –to equally divide the parameter space and collect one observation per division– so the guarantees offered by conformal prediction are valid. + +Empirical coverage. Our approach conformalizes the heatmaps [79] as (ball) or (ellipse). The implementation8 of [79] uses either groundtruth or Faster RCNN [75] bounding boxes, giving four variants of our approach: groundtruth box plus (ball) or (ellipse) (labels: gt-ball, gt-ellipse), and Faster RCNN box plus (ball) or (ellipse) (labels: frcnn-ball, frcnn-ellipse). Fig. 3 left column shows the empirical coverage (i.e., the percentage of images whose groundtruth poses lie in (PURSE)) of all four variants with $\epsilon = 0.1$ and $\epsilon = 0.4$ . We see the empirical coverage is around $90\%$ when $\epsilon = 0.1$ and around $60\%$ when $\epsilon = 0.4$ , for all 8 objects. Though the empirical coverage can deviate from $1 - \epsilon$ , it generally stays within $\pm 5\%$ and mostly goes above $1 - \epsilon$ , which is encouraging given that our calibration set only has size $n = 200$ . Fig. 1 (b) plots examples of the prediction sets. More examples are shown in the Supplementary Material. + +Worst-case error bounds. Fig. 3 middle column plots the worst-case rotation error bound ( $x$ -axis) vs. the actual rotation error between the average pose and the groundtruth ( $y$ -axis) for our approach using the gt-ball setup (results for gt-ellipse, fccnn-ball and fccnn-ellipse are similar and provided in the Supplementary Material). First, when the PURSE covers the groundtruth (blue circles), the rotation error bound is always larger than the actual error (i.e., the blue circles never cross the $y = x$ diagonal). Second, when the error rate is increased from $\epsilon = 0.1$ to $\epsilon = 0.4$ , we observe a shift of the blue circles towards $y = x$ , indicating the error bounds get tightened. Third, our bounds are reasonably tight for most test images (i.e., the bottom-left cluster of blue circles) especially when $\epsilon = 0.4$ . However, they can become overly conservative (i.e., the line of blue circles on the right-side boundary) due to the keypoint prediction + +![](images/5392020343fc77e977bd7590c0f66a74ae84943612cdc7e9359143bb8f8b4b54.jpg) + +![](images/123b08c508a933f19417c16de49223b14183b92242c903bc165ad0128a2100f2.jpg) + +![](images/2f692da27de3c202e07df39842aaa3063addc2abc6f296526eff7fff967a4682.jpg) + +![](images/c578ce8ac9470503174d157a3a2f01951c6b87afe521f504aa2651772a3c4613.jpg) +Figure 3. Empirical coverage (left) and worst-case error bounds (middle: rotation, right: translation). Top: $\epsilon = 0.1$ , bottom: $\epsilon = 0.4$ . For middle and right columns, $x$ -axis represents the worst-case error bounds computed from (12), $y$ -axis represents the actual error between average pose and groundtruth pose. The area below the diagonal $y = x$ indicates correctness of the bounds (i.e., bound $\geq$ error), and points that are closer to the diagonal from below indicate tighter bounds (perfect if precisely lie on the diagonal). Blue circles plot cases where the PURSE covers the groundtruth pose and red squares plot cases were the PURSE does not cover the groundtruth. Notice that blue circles never cross the diagonal and our bounds are correct when the PURSE contains the pose (which holds with $1 - \epsilon$ marginal probability). + +![](images/a3d9c202b27254808b3d11653350e7617a92cd01cb9bfe52aaea13e91c8a093e.jpg) + +![](images/1cf8fb534a62b66a438eac51813d3b1fad772041567f93f5d37be085727eca48.jpg) + +
objectsBaselines (results adapted from [72])Conformalized heatmap
Tekin [89]PoseCNN [95]Oberweger [67]PVNet [72]gt-ballgt-ellipsefrcnn-ballfrcnn-ellipse
ε = 0.1ε = 0.4ε = 0.1ε = 0.4ε = 0.1ε = 0.4ε = 0.1
ape7.0134.669.669.1477.7079.5279.2679.8870.2071.0168.84
can11.2015.1082.6086.0973.4175.9775.8178.1367.5269.8167.69
cat3.6210.4065.1065.1287.3690.5989.5490.1174.9580.2368.98
duck5.0731.8061.4061.4482.7183.0884.0283.5579.3080.6280.06
driller1.407.4073.8073.0679.3282.5481.2282.0458.4865.9258.06
eggbox-1.9013.108.43000.090.18000
glue4.7013.8054.9055.3756.4971.0871.6972.9330.0347.1841.96
holepuncher8.2623.1066.4069.8481.6582.8983.2284.3074.9677.8576.28
average6.1617.2060.9061.0667.3370.7170.6171.3956.9361.5857.73
+ +Table 1. Success rates of baseline methods and our conformalized heatmap (using the average pose) based on the 2D projection metric (i.e., a pose estimation is considered successful if the average 2D reprojection error is below 5 pixels). + +sets become too large. Fig. 3 right column plots similar results for the translation. The Supplementary Material gives a more detailed analysis of this conservatism, wherein we also solve (12) for multiple samples computed by RANSAG, choose the minimum bound, and compare them with those obtained for the average pose (cf. Remark 4). + +Accuracy of the average pose. We compare the accuracy of our average pose with other methods according to the 2D projection metric (an estimation is correct if the mean reprojection error is below 5 pixels). Table 1 shows: (i) our average pose achieves significantly better success rates when using groundtruth bounding boxes, and similar success rates when using Faster RCNN; (ii) the accuracy of the average pose increases when $\epsilon$ increases. + +# 7. Conclusions + +We applied inductive conformal prediction to conformalize heatmap predictions as circular or elliptical prediction + +sets that guarantee probabilistic coverage of the groundtruth keypoints, propagated the uncertainty in keypoints to the object pose to form a PURSE, designed RANSAG to sample from PURSE and compute an average pose, and used SDP relaxations to bound worst-case estimation errors. We validated our theory on the LineMOD Occlusion dataset. Future research will investigate better nonconformity functions, and applications to other vision problems. + +# Acknowledgement + +We thank Yufu Wang and Kostas Daniilidis for providing the heatmap implementation to detect semantic keypoints; Rachel Luo for discussing the exchangeability assumption in conformal prediction; Luca Carlone for pointing out related work on unknown-but-bounded noise (set membership) estimation in control theory, and anonymous reviewers for providing valuable feedback. + +# References + +[1] Anastasios Angelopoulos, Stephen Bates, Jitendra Malik, and Michael I Jordan. Uncertainty sets for image classifiers using conformal prediction. In Intl. Conf. on Learning Representations (ICLR), 2021. 3 +[2] Anastasios N Angelopoulos and Stephen Bates. A gentle introduction to conformal prediction and distribution-free uncertainty quantification. arXiv preprint arXiv:2107.07511, 2021. 3 +[3] Anastasios N Angelopoulos, Stephen Bates, Adam Fisch, Lihua Lei, and Tal Schuster. Conformal risk control. arXiv preprint arXiv:2208.02814, 2022. 3 +[4] Pasquale Antonante, Vasileios Tzoumas, Heng Yang, and Luca Carlone. Outlier-robust estimation: Hardness, minimally tuned algorithms, and applications. IEEE Trans. Robotics, 38(1):281-301, 2021. 5 +[5] Boaz Barak and David Steurer. Proofs, beliefs, and algorithms through the lens of sum-of-squares. Course notes: http://www-sumofsquares.org/public/index.html, 1, 2016. 3 +[6] Timothy D Barfoot. State estimation for robotics. Cambridge University Press, 2017. 3 +[7] Jonathan T Barron. A general and adaptive robust loss function. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4331-4339, 2019. 5 +[8] Stephen Bates, Anastasios Angelopoulos, Lihua Lei, Jitendra Malik, and Michael Jordan. Distribution-free, risk-controlling prediction sets. Journal of the ACM (JACM), 68(6):1-34, 2021. 3 +[9] Michael J Black and Anand Rangarajan. On the unification of line processes, outlier rejection, and robust statistics with applications in early vision. *Intl. J. of Computer Vision*, 19(1):57-91, 1996. 2, 5 +[10] Andrew Blake and Andrew Zisserman. Visual reconstruction. MIT press, 1987. 2 +[11] Eric Brachmann, Alexander Krull, Frank Michel, Stefan Gumhold, Jamie Shotton, and Carsten Rother. Learning 6d object pose estimation using 3d object coordinates. In European Conf. on Computer Vision (ECCV), pages 536-551. Springer, 2014. 2, 4, 7 +[12] Eric Brachmann, Frank Michel, Alexander Krull, Michael Ying Yang, Stefan Gumhold, et al. Uncertainty-driven 6d pose estimation of objects and scenes from a single rgb image. In IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 3364-3372, 2016. 2 +[13] Eric Brachmann and Carsten Rother. Learning less is more-6d camera localization via 3d surface regression. In IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 4654–4662, 2018. 2 +[14] Jesus Briales, Laurent Kneip, and Javier Gonzalez-Jimenez. A certifiably globally optimal solution to the non-minimal relative pose problem. In IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 145-154, 2018. 7 +[15] Dylan Campbell, Liu Liu, and Stephen Gould. Solving the blind perspective-n-point problem end-to-end with robust + +differentiable geometric optimization. In European Conf. on Computer Vision (ECCV), pages 244-261. Springer, 2020. 2 +[16] Dylan Campbell, Lars Petersson, Laurent Kneip, and Hong-dong Li. Globally-optimal inlier set maximisation for simultaneous camera pose and feature correspondence. In Intl. Conf. on Computer Vision (ICCV), pages 1-10, 2017. 2 +[17] Emmanuel J Candes, Justin K Romberg, and Terence Tao. Stable signal recovery from incomplete and inaccurate measurements. Communications on Pure and Applied Mathematics: A Journal Issued by the Courant Institute of Mathematical Sciences, 59(8):1207-1223, 2006. 3 +[18] Luca Carlone. Estimation contracts for outlier-robust geometric perception. arXiv preprint arXiv:2208.10521, 2022.3 +[19] Bo Chen, Jiewei Cao, Alvaro Parra, and Tat-Jun Chin. Satellite pose estimation with deep landmark regression and nonlinear pose refinement. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 0–0, 2019. 1 +[20] Bo Chen, Alvaro Parra, Jiewei Cao, Nan Li, and Tat-Jun Chin. End-to-end learnable geometric vision by backpropagating pnp optimization. In IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 8100-8109, 2020. 1, 2 +[21] Hansheng Chen, Pichao Wang, Fan Wang, Wei Tian, Lu Xiong, and Hao Li. Epro-pnp: Generalized end-to-end probabilistic perspective-n-points for monocular object pose estimation. In IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 2781-2790, 2022. 2 +[22] Tat-Jun Chin, Zhipeng Cai, and Frank Neumann. Robust fitting in computer vision: Easy or hard? In European Conf. on Computer Vision (ECCV), pages 701-716, 2018. 5 +[23] Florence de Grancey, Jean-Luc Adam, Lucian Alecu, Sébastien Gerchinovitz, Franck Mamalet, and David Vigouroux. Object detection with probabilistic guarantees: A conformal prediction approach. In International Conference on Computer Safety, Reliability, and Security, pages 316-329. Springer, 2022. 3 +[24] Ilias Diakonikolas, Daniel M Kane, Sushrut Karmalkar, Ankit Pensia, and Thanasis Pittas. List-decodable sparse mean estimation via difference-of-pairs filtering. arXiv preprint arXiv:2206.05245, 2022.3 +[25] Yonina C Eldar, Amir Beck, and Marc Teboulle. A minimax chebyshev estimator for bounded error estimation. IEEE transactions on signal processing, 56(4):1388-1397, 2008. 7 +[26] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 1, 2, 6 +[27] Xiao-Shan Gao, Xiao-Rong Hou, Jianliang Tang, and Hang-Fei Cheng. Complete solution classification for the perspective-three-point problem. IEEE Trans. Pattern Anal. Machine Intell., 25(8):930–943, 2003. 2, 6 + +[28] Myung Geun Kim. Multivariate outliers and decompositions of mahalanobis distance. Communications in statistics-theory and methods, 29(7):1511-1526, 2000. 4 +[29] Chunhui Gu and Xiaofeng Ren. Discriminative mixture-of-templates for viewpoint classification. In European Conf. on Computer Vision (ECCV), pages 408–421. Springer, 2010. 2 +[30] Richard Hartley, Jochen Trumpf, Yuchao Dai, and Hong-dong Li. Rotation averaging. International journal of computer vision, 103:Intl. J. of Computer Vision, 2013. 6 +[31] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. Cambridge university press, 2003. 3, 5 +[32] Stefan Hinterstoisser, Cedric Cagniart, Slobodan Ilic, Peter Sturm, Nassir Navab, Pascal Fua, and Vincent Lepetit. Gradient response maps for real-time detection of textureless objects. IEEE Trans. Pattern Anal. Machine Intell., 34(5):876-888, 2011. 2 +[33] Stefan Hinterstoisser, Vincent Lepetit, Slobodan Ilic, Stefan Holzer, Gary Bradski, Kurt Konolige, and Nassir Navab. Model based training, detection and pose estimation of texture-less 3d objects in heavily cluttered scenes. In Asian conference on computer vision, pages 548-562. Springer, 2012. 7 +[34] Tomas Hodan, Daniel Barath, and Jiri Matas. Epos: Estimating 6d pose of objects with symmetries. In IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 11703-11712, 2020. 2 +[35] Tomas Hodan, Frank Michel, Eric Brachmann, Wadim Kehl, Anders GlentBuch, Dirk Kraft, Bertram Drost, Joel Vidal, Stephan Ihrke, Xenophon Zabulis, et al. Bop: Benchmark for 6d object pose estimation. In European Conf. on Computer Vision (ECCV), pages 19-34, 2018. 2, 7 +[36] Daniel P Huttenlocher, Gregory A. Klanderman, and William J Rucklidge. Comparing images using the hausdorff distance. IEEE Trans. Pattern Anal. Machine Intell., 15(9):850-863, 1993. 2 +[37] Shun Iwase, Xingyu Liu, Rawal Khirodkar, Rio Yokota, and Kris M Kitani. Repose: Fast 6d object pose refinement via deep texture rendering. In Intl. Conf. on Computer Vision (ICCV), pages 3303-3312, 2021. 2 +[38] Yanmei Jiao, Yue Wang, Bo Fu, Qimeng Tan, Lei Chen, Minhang Wang, Shoudong Huang, and Rong Xiong. Globally optimal consensus maximization for robust visual inertial localization in point and line map. In IEEE/RSJ Intl. Conf. on Intelligent Robots and Systems (IROS), pages 4631-4638. IEEE, 2020. 2 +[39] Fredrik Kahl and Richard Hartley. Multiple-view geometry under the 1 infinity norm. IEEE Trans. Pattern Anal. Machine Intell., 30(9):1603-1617, 2008. 2, 5 +[40] Fredrik Kahl and Didier Henrion. Globally optimal estimates for geometric reconstruction problems. *Intl. J. of Computer Vision*, 74(1):3-15, 2007. 7 +[41] Alex Kendall, Matthew Grimes, and Roberto Cipolla. Posenet: A convolutional network for real-time 6-dof camera relocalization. In Intl. Conf. on Computer Vision (ICCV), pages 2938-2946, 2015. 2 + +[42] Georg Klein and David Murray. Parallel tracking and mapping for small ar workspaces. In 2007 6th IEEE and ACM international symposium on mixed and augmented reality, pages 225-234. IEEE, 2007. 1 +[43] Adam Klivans, Pravesh K Kothari, and Raghu Meka. Efficient algorithms for outlier-robust regression. In Conference On Learning Theory, pages 1420–1430. PMLR, 2018. 3 +[44] Laurent Kneip and Paul Furgale. Opengv: A unified and generalized approach to real-time calibrated geometric vision. In IEEE Intl. Conf. on Robotics and Automation (ICRA), pages 1-8. IEEE, 2014. 7 +[45] Laurent Kneip, Hongdong Li, and Yongduek Seo. Upnp: An optimal o(n) solution to the absolute pose problem with universal applicability. In European Conf. on Computer Vision (ECCV), pages 127-142. Springer, 2014. 2 +[46] Laurent Kneip, Davide Scaramuzza, and Roland Siegwart. A novel parametrization of the perspective-three-point problem for a direct computation of absolute camera position and orientation. In IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 2969-2976. IEEE, 2011. 2 +[47] Yann Labbe, Justin Carpentier, Mathieu Aubry, and Josef Sivic. Cosypos: Consistent multi-view multi-object 6d pose estimation. In European Conf. on Computer Vision (ECCV), pages 574-591. Springer, 2020. 2 +[48] Jing Lei, Max G'Sell, Alessandro Rinaldo, Ryan J Tibshirani, and Larry Wasserman. Distribution-free predictive inference for regression. Journal of the American Statistical Association, 113(523):1094-1111, 2018. 3 +[49] Jing Lei, James Robins, and Larry Wasserman. Distribution-free prediction sets. Journal of the American Statistical Association, 108(501):278-287, 2013. 3 +[50] Vincent Lepetit, Pascal Fua, et al. Monocular model-based 3d tracking of rigid objects: A survey. Foundations and Trends in Computer Graphics and Vision, 1(1):1-89, 2005. 2 +[51] Vincent Lepetit, Francesc Moreno-Noguer, and Pascal Fua. Epnp: An accurate o(n) solution to the pnp problem. *Intl. J. of Computer Vision*, 81(2):155-166, 2009. 2 +[52] Hongdong Li. Consensus set maximization with guaranteed global optimality for robust geometry estimation. In Intl. Conf. on Computer Vision (ICCV), pages 1074-1080. IEEE, 2009. 2 +[53] Shuo Li, Sangdon Park, Xiayan Ji, Insup Lee, and Osbert Bastani. Towards pac multi-object detection and tracking. arXiv preprint arXiv:2204.07482, 2022. 3 +[54] Zhigang Li, Gu Wang, and Xiangyang Ji. Cdpn: Coordinates-based disentangled pose network for real-time rgb-based 6-dof object pose estimation. In Intl. Conf. on Computer Vision (ICCV), pages 7678-7687, 2019. 2 +[55] Joerg Liebelt, Cordelia Schmid, and Klaus Schertler. Independent object class detection using 3d feature maps. In IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 1-8. IEEE, 2008. 2 +[56] David G Lowe. Object recognition from local scale-invariant features. In Intl. Conf. on Computer Vision (ICCV), volume 2, pages 1150-1157. IEEE, 1999. 2 + +[57] Rachel Luo, Shengjia Zhao, Jonathan Kuck, Boris Ivanovic, Silvio Savarese, Edward Schmerling, and Marco Pavone. Sample-efficient safety assurances using conformal prediction. arXiv preprint arXiv:2109.14082, 2021. 4 +[58] Yi Ma, Stefano Soatto, Jana Košecka, and Shankar Sastry. An invitation to 3-d vision: from images to geometric models, volume 26. Springer, 2004. 3 +[59] Prasanta Chandra Mahalanobis. On the generalized distance in statistics. National Institute of Science of India, 1936. 4 +[60] Lucas Manuelli, Wei Gao, Peter Florence, and Russ Tedrake. kpm: Keypoint affordances for category-level robotic manipulation. In The International Symposium of Robotics Research, pages 132-157. Springer, 2019. 1 +[61] Maria Cecilia Mazzaro and Mario Sznaier. A set-membership approach to blind identification. In IEEE Conf. on Decision and Control (CDC), volume 5, pages 5176-5181. IEEE, 2004. 3 +[62] Frank Michel, Alexander Kirillov, Eric Brachmann, Alexander Krull, Stefan Gumhold, Bogdan Savchynskyy, and Carsten Rother. Global hypothesis generation for 6d object pose estimation. In IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 462-471, 2017. 2 +[63] Mario Milanese and Antonio Vicino. Optimal estimation theory for dynamic systems with set membership uncertainty: An overview. Automatica, 27(6):997-1009, 1991. 2, 3, 7 +[64] Ankur Moitra. Sum of squares in theoretical computer science. Sum of Squares: Theory and Applications, 77:83, 2020. 3 +[65] MOSEK ApS. The MOSEK optimization toolbox for MATLAB manual. Version 8.1., 2017. 7 +[66] Jiawang Nie, Li Wang, and Jane J Ye. Bilevel polynomial programs and semidefinite relaxation methods. SIAM Journal on Optimization, 27(3):1728-1757, 2017. 7 +[67] Markus Oberweger, Mahdi Rad, and Vincent Lepetit. Making deep heatmaps robust to partial occlusions for 3d object pose estimation. In European Conf. on Computer Vision (ECCV), pages 119-134, 2018. 2, 8 +[68] Carl Olsson, Fredrik Kahl, and Magnus Oskarsson. Optimal estimation of perspective camera pose. In 18th International Conference on Pattern Recognition (ICPR'06), volume 2, pages 5-8. IEEE, 2006. 2, 5 +[69] Harris Papadopoulos. Inductive conformal prediction: Theory and application to neural networks. In Tools in artificial intelligence. CiteSeer, 2008. 3 +[70] Kiru Park, Timothy Patten, and Markus Vincze. Pix2pose: Pixel-wise coordinate regression of objects for 6d pose estimation. In Intl. Conf. on Computer Vision (ICCV), pages 7668-7677, 2019. 2 +[71] Georgios Pavlakos, Xiaowei Zhou, Aaron Chan, Konstantinos G Derpanis, and Kostas Daniilidis. 6-DoF object pose from semantic keypoints. In IEEE Intl. Conf. on Robotics and Automation (ICRA), pages 2011–2018. IEEE, 2017. 1, 2, 4 +[72] Sida Peng, Yuan Liu, Qixing Huang, Xiaowei Zhou, and Hujun Bao. Pvnet: Pixel-wise voting network for 6dof + +pose estimation. IEEE Trans. Pattern Anal. Machine Intell., 2022. 1, 2, 8 +[73] Luis Pineda, Taosha Fan, Maurizio Monge, Shobha Venkataraman, Paloma Sodhi, Ricky TQ Chen, Joseph Ortiz, Daniel DeTone, Austin Wang, Stuart Anderson, Jing Dong, Brandon Amos, and Mustafa Mukadam. Theseus: A Library for Differentiable Nonlinear Optimization. In Advances in Neural Information Processing Systems (NIPS), 2022. 7 +[74] Mahdi Rad and Vincent Lepetit. Bb8: A scalable, accurate, robust to partial occlusion method for predicting the 3d poses of challenging objects without using depth. In Proceedings of the IEEE international conference on computer vision, pages 3828-3836, 2017. 2 +[75] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in Neural Information Processing Systems (NIPS), 28, 2015. 7 +[76] Yaniv Romano, Matteo Sesia, and Emmanuel Candes. Classification with valid and adaptive coverage. In Advances in Neural Information Processing Systems (NIPS), volume 33, pages 3581-3591, 2020. 3 +[77] David M Rosen, Luca Carlone, Afonso S Bandeira, and John J Leonard. Se-sync: A certifiably correct algorithm for synchronization over the special euclidean group. *Intl. J. of Robotics Research*, 38(2-3):95–125, 2019. 3 +[78] Fred Rothganger, Svetlana Lazebnik, Cordelia Schmid, and Jean Ponce. 3d object modeling and recognition using local affine-invariant image descriptors and multi-view spatial constraints. *Intl. J. of Computer Vision*, 66(3):231-259, 2006. 2 +[79] Karl Schmeckpeper, Philip R Osteen, Yufu Wang, Georgios Pavlakos, Kenneth Chaney, Wyatt Jordan, Xiaowei Zhou, Konstantinos G Derpanis, and Kostas Daniilidis. Semantic keypoint-based pose estimation from single RGB frames. J. of Field Robotics, 2022. 1, 4, 7 +[80] Jingnan Shi, Heng Yang, and Luca Carlone. Optimal pose and shape estimation for category-level 3d object perception. In Robotics: Science and Systems (RSS), 2021. 1 +[81] Jingnan Shi, Heng Yang, and Luca Carlone. Optimal and robust category-level perception: Object pose and shape estimation from 2d and 3d semantic keypoints. arXiv preprint arXiv:2206.12498, 2022. 1 +[82] Torsten Söderström. Errors-in-variables methods in system identification. Automatica, 43(6):939–958, 2007. 3 +[83] David Stutz, Ali Taylan Cemgil, Arnaud Doucet, et al. Learning optimal conformal classifiers. In Intl. Conf. on Learning Representations (ICLR), 2022. 3 +[84] Hao Su, Charles R Qi, Yangyan Li, and Leonidas J Guibas. Render for cnn: Viewpoint estimation in images using cnns trained with rendered 3d model views. In Intl. Conf. on Computer Vision (ICCV), pages 2686-2694, 2015. 2 +[85] Jiaming Sun, Zihao Wang, Siyu Zhang, Xingyi He, Hongcheng Zhao, Guofeng Zhang, and Xiaowei Zhou. Onepose: One-shot object pose estimation without cad models. In IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 6825-6834, 2022. 1 + +[86] Min Sun, Gary Bradski, Bing-Xin Xu, and Silvio Savarese. Depth-encoded hough voting for joint object detection and shape recovery. In European Conf. on Computer Vision (ECCV), pages 658-671. Springer, 2010. 2 +[87] Martin Sundermeyer, Zoltan-Csaba Marton, Maximilian Durner, Manuel Brucker, and Rudolph Triebel. Implicit 3d orientation learning for 6d object detection from rgb images. In European Conf. on Computer Vision (ECCV), pages 699-715, 2018. 2 +[88] Richard Szeliski. Computer vision: algorithms and applications. Springer Nature, 2022. 3 +[89] Bugra Tekin, Sudipta N Sinha, and Pascal Fua. Real-time seamless single shot 6d object pose prediction. In IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 292-301, 2018. 1, 2, 8 +[90] Shubham Tulsiani and Jitendra Malik. Viewpoints and keypoints. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1510-1519, 2015. 2 +[91] Vladimir Vovk. Conditional validity of inductive conformal predictors. In Asian conference on machine learning, pages 475-490. PMLR, 2012. 3 +[92] Vladimir Vovk, Alexander Gammerman, and Glenn Shafer. Algorithmic learning in a random world. Springer Science & Business Media, 2005. 2, 3 +[93] Gu Wang, Fabian Manhardt, Federico Tombari, and Xi-angyang Ji. Gdr-net: Geometry-guided direct regression network for monocular 6d object pose estimation. In IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 16611-16621, 2021. 2 +[94] Håkan Wieslander, Philip J Harrison, Gabriel Skogberg, Sonya Jackson, Markus Fridén, Johan Karlsson, Ola Spjuth, and Carolina Wahlby. Deep learning with conformal prediction for hierarchical analysis of large-scale whole-slide tissue images. IEEE journal of biomedical and health informatics, 25(2):371-380, 2020. 3 +[95] Yu Xiang, Tanner Schmidt, Venkatraman Narayanan, and Dieter Fox. Posecnn: A convolutional neural network for 6d object pose estimation in cluttered scenes. In Robotics: Science and Systems (RSS), 2018. 2, 8 +[96] Heng Yang, Pasquale Antonante, Vasileios Tzoumas, and Luca Carlone. Graduated non-convexity for robust spatial perception: From non-minimal solvers to global outlier rejection. IEEE Robotics and Automation Letters, 5(2):1127-1134, 2020. 2, 5 +[97] Heng Yang and Luca Carlone. In perfect shape: Certifiably optimal 3d shape reconstruction from 2d landmarks. In IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 621-630, 2020. 2 +[98] Heng Yang and Luca Carlone. Certifiably optimal outlier-robust geometric perception: Semidefinite relaxations and scalable global optimization. IEEE Trans. Pattern Anal. Machine Intell., 2022. 1, 2, 7 +[99] Heng Yang, Ling Liang, Luca Carlone, and Kim-Chuan Toh. An inexact projected gradient method with rounding and lifting by nonlinear programming for solving rank-one semidefinite relaxation of polynomial optimization. Mathematical Programming, pages 1-64, 2022. 3 + +[100] Heng Yang, Jingnan Shi, and Luca Carlone. Teaser: Fast and certifiable point cloud registration. IEEE Trans. Robotics, 37(2):314-333, 2020. 3 +[101] Sergey Zakharov, Ivan Shugurov, and Slobodan Ilic. Dpod: 6d pose object detector and refiner. In Intl. Conf. on Computer Vision (ICCV), pages 1941-1950, 2019. 1, 2 +[102] Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. In IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 5745-5753, 2019. 2 \ No newline at end of file diff --git a/2023/Object Pose Estimation With Statistical Guarantees_ Conformal Keypoint Detection and Geometric Uncertainty Propagation/images.zip b/2023/Object Pose Estimation With Statistical Guarantees_ Conformal Keypoint Detection and Geometric Uncertainty Propagation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..38ffad5c09689a67a5345c051a85aaf22a284406 --- /dev/null +++ b/2023/Object Pose Estimation With Statistical Guarantees_ Conformal Keypoint Detection and Geometric Uncertainty Propagation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:72abeeb8093ca6f16d71005469244c9337e526630ec0d1caa54dfaf06680cc00 +size 385776 diff --git a/2023/Object Pose Estimation With Statistical Guarantees_ Conformal Keypoint Detection and Geometric Uncertainty Propagation/layout.json b/2023/Object Pose Estimation With Statistical Guarantees_ Conformal Keypoint Detection and Geometric Uncertainty Propagation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..0646af73318f008ac39a65de29354847fcc1dcc8 --- /dev/null +++ b/2023/Object Pose Estimation With Statistical Guarantees_ Conformal Keypoint Detection and Geometric Uncertainty Propagation/layout.json @@ -0,0 +1,13220 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 74, + 53, + 515, + 88 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 53, + 515, + 88 + ], + "spans": [ + { + "bbox": [ + 74, + 53, + 515, + 88 + ], + "type": "text", + "content": "Object Pose Estimation with Statistical Guarantees: Conformal Keypoint Detection and Geometric Uncertainty Propagation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 222, + 94, + 371, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 222, + 94, + 371, + 121 + ], + "spans": [ + { + "bbox": [ + 222, + 94, + 371, + 121 + ], + "type": "text", + "content": "Heng Yang and Marco Pavone \nNVIDIA Research" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 50, + 127, + 545, + 250 + ], + "blocks": [ + { + "bbox": [ + 50, + 127, + 545, + 250 + ], + "lines": [ + { + "bbox": [ + 50, + 127, + 545, + 250 + ], + "spans": [ + { + "bbox": [ + 50, + 127, + 545, + 250 + ], + "type": "image", + "image_path": "d1ae1164c492c11124d70d6c68b4c42cff9b335a3dc17c948e1bfc7c6ac8a89b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 258, + 546, + 336 + ], + "lines": [ + { + "bbox": [ + 46, + 258, + 546, + 336 + ], + "spans": [ + { + "bbox": [ + 46, + 258, + 546, + 336 + ], + "type": "text", + "content": "Figure 1. Probabilistically correct object pose estimation. Given (a) an input image and heatmap detections of the object semantic keypoints, our method first conformalizes the heatmaps into (b) circular or elliptical prediction sets that guarantee probabilistic coverage of the groundtruth keypoints (e.g., " + }, + { + "bbox": [ + 46, + 258, + 546, + 336 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 46, + 258, + 546, + 336 + ], + "type": "text", + "content": "). Our method then propagates the uncertainty in the keypoints to the object pose, forming (c) a Pose UncleRtainty SEt (PURSE) that contains the groundtruth pose with the same probability. We develop RANdom SAmple averaGing (RANSAG) to sample from PURSE and generate (d) an average pose, and apply semidefinite relaxation to compute (e) worst-case error bounds: the blue duck attains the worst rotation error w.r.t. the (average pose) yellow duck; the red duck attains the worst translation error. Code available: https://github.com/NVlabs/ConformalKeypoint." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 342, + 192, + 354 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 342, + 192, + 354 + ], + "spans": [ + { + "bbox": [ + 143, + 342, + 192, + 354 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 45, + 357, + 290, + 717 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 357, + 290, + 717 + ], + "spans": [ + { + "bbox": [ + 45, + 357, + 290, + 717 + ], + "type": "text", + "content": "The two-stage object pose estimation paradigm first detects semantic keypoints on the image and then estimates the 6D pose by minimizing reprojection errors. Despite performing well on standard benchmarks, existing techniques offer no provable guarantees on the quality and uncertainty of the estimation. In this paper, we inject two fundamental changes, namely conformal keypoint detection and geometric uncertainty propagation, into the two-stage paradigm and propose the first pose estimator that endows an estimation with provable and computable worst-case error bounds. On one hand, conformal keypoint detection applies the statistical machinery of inductive conformal prediction to convert heuristic keypoint detections into circular or elliptical prediction sets that cover the groundtruth keypoints with a user-specified marginal probability (e.g., " + }, + { + "bbox": [ + 45, + 357, + 290, + 717 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 45, + 357, + 290, + 717 + ], + "type": "text", + "content": "). Geometric uncertainty propagation, on the other, propagates the geometric constraints on the keypoints to the 6D object pose, leading to a Pose UnceRtainty SEt (PURSE) that guarantees coverage of the groundtruth pose with the same probability. The PURSE, however, is a nonconvex set that does not directly lead to estimated poses and uncertainties. Therefore, we develop RANdom SAmple averaGing (RANSAG) to compute an average pose and apply semidefinite relaxation to upper bound the worst-case errors between the average pose and the groundtruth. On the LineMOD Occlusion dataset we demonstrate: (i) the PURSE covers the groundtruth with valid probabilities; (ii) the worst-case error bounds provide correct uncertainty quantification; and (iii) the average pose achieves better or similar accuracy as representative methods based on sparse keypoints." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 307, + 342, + 386, + 354 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 342, + 386, + 354 + ], + "spans": [ + { + "bbox": [ + 307, + 342, + 386, + 354 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 364, + 546, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 364, + 546, + 484 + ], + "spans": [ + { + "bbox": [ + 304, + 364, + 546, + 484 + ], + "type": "text", + "content": "Estimating object poses from images is a fundamental problem in computer vision and finds extensive applications in augmented reality [42], autonomous driving [80], robotic manipulation [60], and space robotics [19]. One of the most popular paradigms for object pose estimation is a two-stage pipeline [20, 71, 72, 79, 81, 85, 89, 101], where the first stage detects (semantic) keypoints of the objects on the image, and the second stage computes the object pose by solving an optimization known as Perspective-" + }, + { + "bbox": [ + 304, + 364, + 546, + 484 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 364, + 546, + 484 + ], + "type": "text", + "content": "-Points (PnP) that minimizes reprojection errors of the detected keypoints." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 486, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 547, + 713 + ], + "type": "text", + "content": "Safety-critical applications call for provably correct computer vision algorithms. Existing algorithms in the two-stage paradigm (reviewed in Section 2), however, provide few performance guarantees on the quality of the estimated poses, due to three challenges. (C1) It is difficult to ensure the detected keypoints (typically from neural networks) are close to the groundtruth keypoints. In practice, the first stage often outputs keypoints that are arbitrarily wrong, known as outliers. (C2) Robust estimation is employed in the second stage to reject outliers, leading to nonconvex optimizations. Fast heuristics such as RANSAC [26] are widely adopted to find an approximate solution but they cannot guarantee global optimality and often fail without notice. (C3) There is no provably correct uncertainty quantification of the estimation, notably, a formal worst-case error bound between the estimation and the groundtruth. Though recent work [98] proposed convex relaxations to certify global optimality of RANSAC and addressed (C2), it cannot ensure correct estimation as the optimal pose may be far away from" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 750, + 315, + 759 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 750, + 315, + 759 + ], + "spans": [ + { + "bbox": [ + 295, + 750, + 315, + 759 + ], + "type": "text", + "content": "8947" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 253, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 253, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 253, + 83 + ], + "type": "text", + "content": "the correct pose when the keypoints are unreliable." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 87, + 289, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 87, + 289, + 518 + ], + "spans": [ + { + "bbox": [ + 46, + 87, + 289, + 518 + ], + "type": "text", + "content": "Contributions. We propose a two-stage object pose estimation framework with statistical guarantees, illustrated in Fig. 1. Given an input image, we assume a neural network [71] is available to generate heatmap predictions of the object keypoints (Fig. 1(a)). Our framework then proceeds in two stages, namely conformal keypoint detection (Section 4) and geometric uncertainty propagation (Section 5). We first apply the statistical machinery of inductive conformal prediction (introduced in Section 3), with nonconformity functions inspired by the design of residual functions in classical geometric vision [39], to conformalize the heatmaps into circular or elliptical prediction sets—one for each keypoint—that guarantee coverage of the groundtruth keypoints with a user-specified marginal probability (Fig. 1(b)). This provides a simple and general methodology to bound the keypoint prediction errors (i.e., addressing (C1)). Given the keypoint prediction sets, we reformulate the constraints (enforced by the prediction sets) on the keypoints as constraints on the object pose, leading to a Pose UnceRtainty SET (PURSE) that guarantees coverage of the groundtruth pose with the same probability. Fig. 1(c) plots the boundary of an example PURSE (roll, pitch, raw angles for the rotation, and Euclidean coordinates for the translation). The PURSE, however, is an abstract nonconvex set that does not directly admit estimated poses and uncertainty. Therefore, we develop RANdom SAmple averaGing (RANSAG) to compute an average pose (Fig. 1(d)) and employ semidefinite relaxations to upper bound the worst-case rotation and translation errors between the average pose and the groundtruth (Fig. 1(e)). This gives rise to the first kind of computable worst-case probabilistic error bounds for object pose estimation (i.e., addressing (C3)). Our PURSE methodology has connections to the framework of unknown-but-bounded noise estimation in control theory [63], with special provisions to derive the bounds in a statistically principled way and enable efficient computation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 519, + 287, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 519, + 287, + 675 + ], + "spans": [ + { + "bbox": [ + 46, + 519, + 287, + 675 + ], + "type": "text", + "content": "We test our framework on the LineMOD Occlusion (LM-O) dataset [11] to verify the correctness of the theory (Section 6). First, we empirically show that the PURSE indeed contains the groundtruth pose according to the user-specified probability. Second, we demonstrate the correctness of the worst-case error bounds: when the PURSE contains the groundtruth, our bounds are always larger than, and in many cases close to, the actual errors between the average pose and the groundtruth pose. Third, we benchmark the accuracy of the average pose (coming from RANSAG) with representative two-stage pipelines based on sparse keypoints (e.g., PVNet [72]) and show that the average pose achieves better or similar accuracy." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 677, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 715 + ], + "type": "text", + "content": "Limitations. A drawback of our approach, and conformal prediction in general, is that the size of the prediction sets depends on the nonconformity function (whose design" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 545, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 121 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 121 + ], + "type": "text", + "content": "can be an art) and may be conservative. Our experiments suggest the bounds are loose when the keypoint prediction sets are large (e.g., giving " + }, + { + "bbox": [ + 304, + 72, + 545, + 121 + ], + "type": "inline_equation", + "content": "180^{\\circ}" + }, + { + "bbox": [ + 304, + 72, + 545, + 121 + ], + "type": "text", + "content": " rotation bound). We discuss challenges and opportunities in tightening the bounds." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 131, + 392, + 144 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 131, + 392, + 144 + ], + "spans": [ + { + "bbox": [ + 305, + 131, + 392, + 144 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 152, + 547, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 152, + 547, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 152, + 547, + 665 + ], + "type": "text", + "content": "Image-based object pose estimation. We categorize object pose estimation into two paradigms: single-stage and two-stage. The latter first detects 2D-3D correspondences and then estimates the object pose via solving a PnP problem, while the former produces poses without intermediate correspondences. (i) Single-stage. Early methods perform pose estimation via template matching [29, 32, 36]. Recently, deep learning-based approaches such as PoseNet [41] and PoseCNN [95] applied CNNs to directly regress poses. A major challenge of pose regression is the nonlinearity of 3D rotations, and motivated formulating regression as classification [84, 87, 90] or designing better rotation representations [47, 102]. It is also popular to predict multiple pose hypotheses followed by voting [55, 62, 86]. (ii) Two-stage. Early research used handcrafted features [50, 56, 78] to establish 2D-3D correspondences and focused on developing algorithms for solving PnP. Notable algorithms include the minimal solver P3P [27, 46] and variants of the nonminimal solver PnP [45, 51, 68, 97]. Outliers (i.e., wrong correspondences) motivated robust estimation based on RANSAC [26], graduated non-convexity [9, 10, 96], branch-and-bound [16, 38, 52], or semidefinite relaxations [98]. Unreliable correspondences soon became the bottleneck and learned correspondences have been predominant. Learned correspondences can be sparse or dense. Sparse methods define a handful of keypoints and predict locations of the keypoints via direct regression [74, 89], probabilistic heatmap [67, 71], or voting [72]. Dense methods [12, 34, 54, 70, 93, 101] regress for each object pixel the coordinates of its corresponding 3D point. Recent literature focuses on end-to-end training via differentiating PnP [13, 15, 20, 21, 37]. Both single-stage and two-stage methods perform well on standard benchmarks [35], but a crucial feature that is missing, especially when deploying computer vision algorithms in safety-critical applications, is that these methods do not provide provably correct uncertainty quantification and formal error bounds w.r.t. the groundtruth (for either the correspondences or the poses). In this paper, we provide rigorous guarantees by applying conformal prediction to an existing keypoint detection method (the heatmap [71]) and leveraging old and new techniques in computer vision to derive formal error bounds." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": "Conformal prediction in computer vision. Conformal prediction [92] is a statistical machinery that offers provably correct finite-sample uncertainty quantification without assumptions on the data distribution or the prediction model" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 750, + 315, + 759 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 750, + 315, + 759 + ], + "spans": [ + { + "bbox": [ + 295, + 750, + 315, + 759 + ], + "type": "text", + "content": "8948" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 287 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 287 + ], + "type": "text", + "content": "(i.e., offering a set prediction, instead of a point prediction, that guarantees probabilistic coverage of the groundtruth). Inductive conformal prediction [69] is the most popular variant of conformal prediction because it does not require retraining of the prediction models [1, 2, 49]. Applying conformal prediction to computer vision, however, is still in its infancy. Existing works focus on image classification [1, 76], tumor segmentation [3, 8, 94], and bounding box detection [2, 23, 53], which are classification or low-dimensional regression problems. Inspired by these works, our unique contributions in this paper are: (i) we apply conformal prediction to keypoints detection, a high-dimensional regression problem; (ii) we design new nonconformity functions and discuss their connections with classical geometric vision; and (iii) we develop algorithms that propagate the uncertainty after conformal prediction to form prediction sets of 6D poses, which are nonlinear and nonconvex manifold objects." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 288, + 289, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 288, + 289, + 647 + ], + "spans": [ + { + "bbox": [ + 46, + 288, + 289, + 647 + ], + "type": "text", + "content": "Performance guarantees. Pose estimation from 2D-2D, 2D-3D, and 3D-3D correspondences are foundational problems in computer vision textbooks [6, 31, 58, 88] and typically boil down to formulating and solving mathematical optimization problems. Benchmarking on simulated and real datasets has been a widely adopted standard for testing different formulations and solvers. However, empirical performance can be misleading without theoretical guarantees. A striking fact is that, though error analysis is an important topic in applied math [17, 24, 43] and control theory [61, 63, 82], there is very limited literature in computer vision that reason about worst-case estimation errors between the optimal solution and the groundtruth. A popular heuristic relies on the inverse of the Hessian at an optimal solution, which provides the Cramer-Rao lower bound on the covariance of the solution (for linear regression this coincides with the covariance) [88, Section B.6] and thus cannot upper bound the estimation errors. Recent works [18, 77, 100] derived error bounds for a few geometric vision problems. However, the bounds either depend on uncheckable assumptions and cannot be computed [77, 100], or build on machinery (e.g., sum-of-squares proof [5, 64]) that only applies to estimators based on moment relaxations [18], which are still computationally expensive in practice [99]. In this paper, we develop the first kind of efficiently computable error bounds that only require the assumption of exchangeability (which comes from conformal prediction). We justify this assumption on our test dataset and numerically show our bounds can be tight for a subset of the test problems." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 657, + 226, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 657, + 226, + 670 + ], + "spans": [ + { + "bbox": [ + 47, + 657, + 226, + 670 + ], + "type": "text", + "content": "3. Inductive Conformal Prediction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "content": "Given a set " + }, + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\{z_i = (x_i, y_i)\\}_{i=1}^l" + }, + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "content": " with observation " + }, + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "inline_equation", + "content": "x_i \\in \\mathcal{X}" + }, + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "content": " and label " + }, + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "inline_equation", + "content": "y_i \\in \\mathcal{Y}" + }, + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "content": " such that each " + }, + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "inline_equation", + "content": "z_i \\in \\mathcal{Z} := \\mathcal{X} \\times \\mathcal{Y}" + }, + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "content": " is drawn i.i.d. from an unknown distribution on " + }, + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}" + }, + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "content": ", inductive confor" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": "mal prediction (ICP) provides a set prediction " + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "F^{\\epsilon}(x) \\subseteq \\mathcal{V}" + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": ", parameterized by an error rate " + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "0 < \\epsilon < 1" + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": ", such that given a new sample " + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "z_{l + 1} = (x_{l + 1}, y_{l + 1})" + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": " satisfying an exchangeability condition (elaborated in Theorem 1), we have" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 353, + 126, + 545, + 140 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 126, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 353, + 126, + 545, + 140 + ], + "type": "interline_equation", + "content": "\\mathbb {P} \\left[ y _ {l + 1} \\in F ^ {\\epsilon} \\left(x _ {l + 1}\\right) \\right] \\geq 1 - \\epsilon , \\tag {1}", + "image_path": "5cbd76d7b497d65cfc5e2b53b7f833a0621e063e7b964ef7adda69c14a142791.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 145, + 545, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 145, + 545, + 169 + ], + "spans": [ + { + "bbox": [ + 304, + 145, + 545, + 169 + ], + "type": "text", + "content": "i.e., the prediction set " + }, + { + "bbox": [ + 304, + 145, + 545, + 169 + ], + "type": "inline_equation", + "content": "F^{\\epsilon}" + }, + { + "bbox": [ + 304, + 145, + 545, + 169 + ], + "type": "text", + "content": " guarantees to contain the true label " + }, + { + "bbox": [ + 304, + 145, + 545, + 169 + ], + "type": "inline_equation", + "content": "y_{l + 1}" + }, + { + "bbox": [ + 304, + 145, + 545, + 169 + ], + "type": "text", + "content": " with probability at least " + }, + { + "bbox": [ + 304, + 145, + 545, + 169 + ], + "type": "inline_equation", + "content": "1 - \\epsilon" + }, + { + "bbox": [ + 304, + 145, + 545, + 169 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 170, + 546, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 170, + 546, + 289 + ], + "spans": [ + { + "bbox": [ + 304, + 170, + 546, + 289 + ], + "type": "text", + "content": "Training. We start by dividing the dataset into a proper training set " + }, + { + "bbox": [ + 304, + 170, + 546, + 289 + ], + "type": "inline_equation", + "content": "\\{z_1, \\ldots, z_m\\}" + }, + { + "bbox": [ + 304, + 170, + 546, + 289 + ], + "type": "text", + "content": " and a calibration set " + }, + { + "bbox": [ + 304, + 170, + 546, + 289 + ], + "type": "inline_equation", + "content": "\\{z_{m+1}, \\ldots, z_l\\}" + }, + { + "bbox": [ + 304, + 170, + 546, + 289 + ], + "type": "text", + "content": ". We shorthand " + }, + { + "bbox": [ + 304, + 170, + 546, + 289 + ], + "type": "inline_equation", + "content": "n = l - m" + }, + { + "bbox": [ + 304, + 170, + 546, + 289 + ], + "type": "text", + "content": " as the size of the calibration set. We learn a prediction function " + }, + { + "bbox": [ + 304, + 170, + 546, + 289 + ], + "type": "inline_equation", + "content": "f: \\mathcal{X} \\to \\tilde{\\mathcal{V}}" + }, + { + "bbox": [ + 304, + 170, + 546, + 289 + ], + "type": "text", + "content": " from the proper training set using any architecture, which allows us to fully exploit the power of modern deep learning. The prediction space " + }, + { + "bbox": [ + 304, + 170, + 546, + 289 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{V}}" + }, + { + "bbox": [ + 304, + 170, + 546, + 289 + ], + "type": "text", + "content": " can be the same as the label space " + }, + { + "bbox": [ + 304, + 170, + 546, + 289 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 304, + 170, + 546, + 289 + ], + "type": "text", + "content": ", or can contain auxiliary information such as a heuristic notion of uncertainty (e.g., softmax scores in classification or a heatmap in the case of keypoint detection)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 289, + 547, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 289, + 547, + 338 + ], + "spans": [ + { + "bbox": [ + 304, + 289, + 547, + 338 + ], + "type": "text", + "content": "Conformal calibration. We define a nonconformity function " + }, + { + "bbox": [ + 304, + 289, + 547, + 338 + ], + "type": "inline_equation", + "content": "S: \\mathcal{Z}^m \\times \\mathcal{Z} \\to \\mathbb{R}" + }, + { + "bbox": [ + 304, + 289, + 547, + 338 + ], + "type": "text", + "content": " to measure how well a given sample " + }, + { + "bbox": [ + 304, + 289, + 547, + 338 + ], + "type": "inline_equation", + "content": "z = (x, y)" + }, + { + "bbox": [ + 304, + 289, + 547, + 338 + ], + "type": "text", + "content": " conforms to the proper training set. A popular instance of " + }, + { + "bbox": [ + 304, + 289, + 547, + 338 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 289, + 547, + 338 + ], + "type": "text", + "content": " leverages the learned prediction " + }, + { + "bbox": [ + 304, + 289, + 547, + 338 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 289, + 547, + 338 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 334, + 344, + 545, + 358 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 344, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 334, + 344, + 545, + 358 + ], + "type": "interline_equation", + "content": "S \\left(\\left\\{z _ {1}, \\dots , z _ {m} \\right\\}, (x, y)\\right) \\stackrel {{e. g.}} {{=}} r (y, f (x)), \\tag {2}", + "image_path": "813b12ea68f0e7f8f7e9b422d2631ea36bfde019456b1308f8c481e88d1b30be.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 363, + 547, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 363, + 547, + 508 + ], + "spans": [ + { + "bbox": [ + 304, + 363, + 547, + 508 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 363, + 547, + 508 + ], + "type": "inline_equation", + "content": "r: \\mathcal{Y} \\times \\tilde{\\mathcal{Y}} \\to \\mathbb{R}" + }, + { + "bbox": [ + 304, + 363, + 547, + 508 + ], + "type": "text", + "content": " is a measure of disagreement between the label " + }, + { + "bbox": [ + 304, + 363, + 547, + 508 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 363, + 547, + 508 + ], + "type": "text", + "content": " and the prediction " + }, + { + "bbox": [ + 304, + 363, + 547, + 508 + ], + "type": "inline_equation", + "content": "f(x)" + }, + { + "bbox": [ + 304, + 363, + 547, + 508 + ], + "type": "text", + "content": ". For example, consider " + }, + { + "bbox": [ + 304, + 363, + 547, + 508 + ], + "type": "inline_equation", + "content": "\\mathcal{Y} = \\tilde{\\mathcal{Y}} = \\mathbb{R}" + }, + { + "bbox": [ + 304, + 363, + 547, + 508 + ], + "type": "text", + "content": ", one can design " + }, + { + "bbox": [ + 304, + 363, + 547, + 508 + ], + "type": "inline_equation", + "content": "r(y, f(x)) = |y - f(x)|" + }, + { + "bbox": [ + 304, + 363, + 547, + 508 + ], + "type": "text", + "content": ": if " + }, + { + "bbox": [ + 304, + 363, + 547, + 508 + ], + "type": "inline_equation", + "content": "(x, y)" + }, + { + "bbox": [ + 304, + 363, + 547, + 508 + ], + "type": "text", + "content": " poorly conforms to the training set, " + }, + { + "bbox": [ + 304, + 363, + 547, + 508 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 363, + 547, + 508 + ], + "type": "text", + "content": " will incur large errors. While the function " + }, + { + "bbox": [ + 304, + 363, + 547, + 508 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 363, + 547, + 508 + ], + "type": "text", + "content": " can be arbitrary (e.g., a learnable neural network [83]), (2) is a convenient definition since " + }, + { + "bbox": [ + 304, + 363, + 547, + 508 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 363, + 547, + 508 + ], + "type": "text", + "content": " is implicitly dependent on " + }, + { + "bbox": [ + 304, + 363, + 547, + 508 + ], + "type": "inline_equation", + "content": "\\{z_i\\}_{i=1}^m" + }, + { + "bbox": [ + 304, + 363, + 547, + 508 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 363, + 547, + 508 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 304, + 363, + 547, + 508 + ], + "type": "text", + "content": " can incorporate domain-specific knowledge. We then compute the nonconformity scores on the calibration set as " + }, + { + "bbox": [ + 304, + 363, + 547, + 508 + ], + "type": "inline_equation", + "content": "\\alpha_i = r(y_i, f(x_i)), i = m+1,\\ldots,l" + }, + { + "bbox": [ + 304, + 363, + 547, + 508 + ], + "type": "text", + "content": ", and sort them in nonincreasing order " + }, + { + "bbox": [ + 304, + 363, + 547, + 508 + ], + "type": "inline_equation", + "content": "\\alpha_{\\pi(1)} \\geq \\dots \\geq \\alpha_{\\pi(n)}" + }, + { + "bbox": [ + 304, + 363, + 547, + 508 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 363, + 547, + 508 + ], + "type": "inline_equation", + "content": "\\pi(i) \\in \\{m+1,\\ldots,l\\}" + }, + { + "bbox": [ + 304, + 363, + 547, + 508 + ], + "type": "text", + "content": " is an index permutation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 508, + 547, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 508, + 547, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 508, + 547, + 544 + ], + "type": "text", + "content": "Conformal prediction. Given a new observation " + }, + { + "bbox": [ + 304, + 508, + 547, + 544 + ], + "type": "inline_equation", + "content": "x_{l+1}" + }, + { + "bbox": [ + 304, + 508, + 547, + 544 + ], + "type": "text", + "content": " (with an unknown " + }, + { + "bbox": [ + 304, + 508, + 547, + 544 + ], + "type": "inline_equation", + "content": "y_{l+1}" + }, + { + "bbox": [ + 304, + 508, + 547, + 544 + ], + "type": "text", + "content": ") and a user-specified " + }, + { + "bbox": [ + 304, + 508, + 547, + 544 + ], + "type": "inline_equation", + "content": "\\epsilon \\in (0, 1)" + }, + { + "bbox": [ + 304, + 508, + 547, + 544 + ], + "type": "text", + "content": ", we compute the inductive conformal prediction (ICP) set as" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 326, + 549, + 545, + 564 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 549, + 545, + 564 + ], + "spans": [ + { + "bbox": [ + 326, + 549, + 545, + 564 + ], + "type": "interline_equation", + "content": "F ^ {\\epsilon} \\left(x _ {l + 1}\\right) = \\left\\{y \\in \\mathcal {Y} \\mid \\alpha^ {y} \\leq \\alpha_ {\\pi (\\lfloor (n + 1) \\epsilon \\rfloor)} \\right\\}, \\tag {3}", + "image_path": "d382d68657d9ac0a87684a1f624ce41cce73c73b1a68c6cfc9e5edaf199d1238.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 569, + 547, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 569, + 547, + 654 + ], + "spans": [ + { + "bbox": [ + 304, + 569, + 547, + 654 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 569, + 547, + 654 + ], + "type": "inline_equation", + "content": "\\alpha^y = r(y, f(x_{l+1}))" + }, + { + "bbox": [ + 304, + 569, + 547, + 654 + ], + "type": "text", + "content": " is the nonconformity score of the new sample when fixing the true label to be " + }, + { + "bbox": [ + 304, + 569, + 547, + 654 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 569, + 547, + 654 + ], + "type": "text", + "content": ". In other words, the ICP set (3) outputs the set of all labels that make the nonconformity score of the new sample no greater than " + }, + { + "bbox": [ + 304, + 569, + 547, + 654 + ], + "type": "inline_equation", + "content": "\\alpha_{\\pi(\\lfloor n + 1 \\epsilon \\rfloor)} - \\text{the } \\lfloor (n + 1) \\epsilon \\rfloor" + }, + { + "bbox": [ + 304, + 569, + 547, + 654 + ], + "type": "text", + "content": "-th largest nonconformity score in the calibration set. We have the following result stating the probabilistic coverage of the ICP set (3)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 658, + 547, + 695 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 658, + 547, + 695 + ], + "spans": [ + { + "bbox": [ + 304, + 658, + 547, + 695 + ], + "type": "text", + "content": "Theorem 1 (Validity of ICP Coverage [48, 91, 92]). If " + }, + { + "bbox": [ + 304, + 658, + 547, + 695 + ], + "type": "inline_equation", + "content": "z_{m+1}, \\ldots, z_l, z_{l+1} = (x_{l+1}, y_{l+1})" + }, + { + "bbox": [ + 304, + 658, + 547, + 695 + ], + "type": "text", + "content": " are exchangeable, i.e., their distribution is invariant under permutation, then" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 700, + 545, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 700, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 700, + 545, + 715 + ], + "type": "interline_equation", + "content": "1 - \\epsilon \\leq \\mathbb {P} \\left[ y _ {l + 1} \\in F ^ {\\epsilon} \\left(x _ {l + 1}\\right) \\right] \\leq 1 - \\epsilon + 1 / (n + 1) \\tag {4}", + "image_path": "5f64b2d7a3787c73f8cc0928dcd423c8077ce1f57c5e73762f151bc04b835850.jpg" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 750, + 315, + 759 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 750, + 315, + 759 + ], + "spans": [ + { + "bbox": [ + 294, + 750, + 315, + 759 + ], + "type": "text", + "content": "8949" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 97, + 60, + 239, + 178 + ], + "blocks": [ + { + "bbox": [ + 97, + 60, + 239, + 178 + ], + "lines": [ + { + "bbox": [ + 97, + 60, + 239, + 178 + ], + "spans": [ + { + "bbox": [ + 97, + 60, + 239, + 178 + ], + "type": "image", + "image_path": "79351fc3e0ba435b18162af2bc294190b15a6316fcc94715c40aed14afceb2b4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 182, + 287, + 216 + ], + "lines": [ + { + "bbox": [ + 46, + 182, + 287, + 216 + ], + "spans": [ + { + "bbox": [ + 46, + 182, + 287, + 216 + ], + "type": "text", + "content": "Figure 2. Beta distribution of the conditional coverage in (5) with " + }, + { + "bbox": [ + 46, + 182, + 287, + 216 + ], + "type": "inline_equation", + "content": "\\epsilon = 0.1" + }, + { + "bbox": [ + 46, + 182, + 287, + 216 + ], + "type": "text", + "content": " and different " + }, + { + "bbox": [ + 46, + 182, + 287, + 216 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 182, + 287, + 216 + ], + "type": "text", + "content": ". Notice how the conditional probability becomes more concentrated around " + }, + { + "bbox": [ + 46, + 182, + 287, + 216 + ], + "type": "inline_equation", + "content": "1 - \\epsilon" + }, + { + "bbox": [ + 46, + 182, + 287, + 216 + ], + "type": "text", + "content": " when " + }, + { + "bbox": [ + 46, + 182, + 287, + 216 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 182, + 287, + 216 + ], + "type": "text", + "content": " increases." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 45, + 217, + 287, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 217, + 287, + 243 + ], + "spans": [ + { + "bbox": [ + 45, + 217, + 287, + 243 + ], + "type": "text", + "content": "for any " + }, + { + "bbox": [ + 45, + 217, + 287, + 243 + ], + "type": "inline_equation", + "content": "\\epsilon \\in (0,1)" + }, + { + "bbox": [ + 45, + 217, + 287, + 243 + ], + "type": "text", + "content": ". Furthermore, when conditioned on the calibration set, calling " + }, + { + "bbox": [ + 45, + 217, + 287, + 243 + ], + "type": "inline_equation", + "content": "h = \\lfloor (n + 1)\\epsilon \\rfloor" + }, + { + "bbox": [ + 45, + 217, + 287, + 243 + ], + "type": "text", + "content": ", we have" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 41, + 251, + 287, + 266 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 41, + 251, + 287, + 266 + ], + "spans": [ + { + "bbox": [ + 41, + 251, + 287, + 266 + ], + "type": "interline_equation", + "content": "\\mathbb {P} \\left[ y _ {l + 1} \\in F ^ {\\epsilon} (x _ {l + 1}) \\mid \\{z _ {m + 1}, \\dots , z _ {l} \\} \\right] \\sim \\mathrm {B e t a} (n + 1 - h, h). (5)", + "image_path": "82b0964e254497181068eb41b0c4404aba13450350ee43c285c9db237ef8d386.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 274, + 289, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 274, + 289, + 636 + ], + "spans": [ + { + "bbox": [ + 46, + 274, + 289, + 636 + ], + "type": "text", + "content": "A few remarks are in order about Theorem 1. First, asking " + }, + { + "bbox": [ + 46, + 274, + 289, + 636 + ], + "type": "inline_equation", + "content": "z_{m+1}, \\ldots, z_l, z_{l+1}" + }, + { + "bbox": [ + 46, + 274, + 289, + 636 + ], + "type": "text", + "content": " to be exchangeable is weaker than asking them to be independent. However, this assumption typically fails when the calibration set is a single video sequence, where the image frames " + }, + { + "bbox": [ + 46, + 274, + 289, + 636 + ], + "type": "inline_equation", + "content": "\\{z_{m+1}, \\ldots, z_l\\}" + }, + { + "bbox": [ + 46, + 274, + 289, + 636 + ], + "type": "text", + "content": " are temporally correlated [57]. Fortunately, as we detail in Section 6, the way the LineMOD Occlusion dataset [11] was collected makes the exchangeability condition easily satisfied, which also suggests best practices to make the exchangeability condition hold in computer vision. Second, the lower bound in (4) can be intuitively proved because under exchangeability, " + }, + { + "bbox": [ + 46, + 274, + 289, + 636 + ], + "type": "inline_equation", + "content": "\\alpha_{l+1} := r(y_{l+1}, f(x_{l+1}))" + }, + { + "bbox": [ + 46, + 274, + 289, + 636 + ], + "type": "text", + "content": " – the nonconformity score of the new sample with the true label– is exchangeable with the nonconformity scores of the calibration samples, and hence equally likely to fall in anywhere between the scores " + }, + { + "bbox": [ + 46, + 274, + 289, + 636 + ], + "type": "inline_equation", + "content": "\\{\\alpha_{\\pi(i)}\\}_{i=1}^n" + }, + { + "bbox": [ + 46, + 274, + 289, + 636 + ], + "type": "text", + "content": ". Consequently, " + }, + { + "bbox": [ + 46, + 274, + 289, + 636 + ], + "type": "inline_equation", + "content": "\\mathbb{P}[y_{l+1} \\in F^\\epsilon(x_{l+1})] = \\mathbb{P}\\left[\\alpha_{l+1} \\leq \\alpha_{\\pi(\\lfloor n+1\\epsilon \\rfloor)}\\right] = 1 - \\lfloor(n+1)\\epsilon \\rfloor / (n+1) \\geq 1 - \\epsilon" + }, + { + "bbox": [ + 46, + 274, + 289, + 636 + ], + "type": "text", + "content": ". The upper bound in (4) states that " + }, + { + "bbox": [ + 46, + 274, + 289, + 636 + ], + "type": "inline_equation", + "content": "1 - \\epsilon" + }, + { + "bbox": [ + 46, + 274, + 289, + 636 + ], + "type": "text", + "content": " is not overly conservative (indeed tight if " + }, + { + "bbox": [ + 46, + 274, + 289, + 636 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 274, + 289, + 636 + ], + "type": "text", + "content": " is large). Lastly, the probabilistic guarantee in (4) is marginal over the randomness of the calibration set, meaning if one chooses an infinite number of calibration sets, the average empirical coverage will converge to " + }, + { + "bbox": [ + 46, + 274, + 289, + 636 + ], + "type": "inline_equation", + "content": "1 - \\epsilon" + }, + { + "bbox": [ + 46, + 274, + 289, + 636 + ], + "type": "text", + "content": ". This, however, implies that the empirical coverage given one calibration set is a random variable that fluctuates as the Beta distribution (5). Fig. 2 plots the Beta distribution at " + }, + { + "bbox": [ + 46, + 274, + 289, + 636 + ], + "type": "inline_equation", + "content": "\\epsilon = 0.1" + }, + { + "bbox": [ + 46, + 274, + 289, + 636 + ], + "type": "text", + "content": " with different sizes of the calibration set. We observe that as " + }, + { + "bbox": [ + 46, + 274, + 289, + 636 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 274, + 289, + 636 + ], + "type": "text", + "content": " increases the empirical coverage becomes more concentrated at " + }, + { + "bbox": [ + 46, + 274, + 289, + 636 + ], + "type": "inline_equation", + "content": "1 - \\epsilon" + }, + { + "bbox": [ + 46, + 274, + 289, + 636 + ], + "type": "text", + "content": ". Our experiments show that even with a small (" + }, + { + "bbox": [ + 46, + 274, + 289, + 636 + ], + "type": "inline_equation", + "content": "n = 200" + }, + { + "bbox": [ + 46, + 274, + 289, + 636 + ], + "type": "text", + "content": ") calibration set, the empirical coverage is close to, and mostly higher than, " + }, + { + "bbox": [ + 46, + 274, + 289, + 636 + ], + "type": "inline_equation", + "content": "1 - \\epsilon" + }, + { + "bbox": [ + 46, + 274, + 289, + 636 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 645, + 220, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 645, + 220, + 659 + ], + "spans": [ + { + "bbox": [ + 46, + 645, + 220, + 659 + ], + "type": "text", + "content": "4. Conformal Keypoint Detection" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 665, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 689 + ], + "type": "text", + "content": "In this section, we apply the ICP framework in Section 3 to the problem of semantic keypoint detection." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": "Setup. Denote by " + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "x \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": " an RGB image picturing an object, by " + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{y} = (y_1, \\dots, y_K) \\in \\mathbb{R}^2 \\times \\dots \\times \\mathbb{R}^2 \\coloneqq \\mathcal{Y}" + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": "the groundtruth locations of " + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": " semantic keypoints of the object. We partition a given dataset " + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "\\{z_i := (x_i, y_i)\\}_{i=1}^l" + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": " into a proper training set (of size " + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": ") and a calibration set (of size " + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": "). We follow the three steps in Section 3 to perform ICP." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 121, + 546, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 121, + 546, + 266 + ], + "spans": [ + { + "bbox": [ + 304, + 121, + 546, + 266 + ], + "type": "text", + "content": "Training. We choose the heatmap approach in [71, 79] as the prediction function: given an image " + }, + { + "bbox": [ + 304, + 121, + 546, + 266 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 121, + 546, + 266 + ], + "type": "text", + "content": ", [79] outputs a set of heatmaps " + }, + { + "bbox": [ + 304, + 121, + 546, + 266 + ], + "type": "inline_equation", + "content": "\\pmb{f}(x) = (f(x)_1, \\dots, f(x)_K)" + }, + { + "bbox": [ + 304, + 121, + 546, + 266 + ], + "type": "text", + "content": ", where each " + }, + { + "bbox": [ + 304, + 121, + 546, + 266 + ], + "type": "inline_equation", + "content": "f(x)_k \\in \\Delta^{HW} := \\{v \\in \\mathbb{R}_+^{HW} \\mid \\sum_i^{HW} v_i = 1\\}" + }, + { + "bbox": [ + 304, + 121, + 546, + 266 + ], + "type": "text", + "content": " predicts the probability distribution of the " + }, + { + "bbox": [ + 304, + 121, + 546, + 266 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 121, + 546, + 266 + ], + "type": "text", + "content": "-th keypoint lying on each pixel of the image. For convenience, we use " + }, + { + "bbox": [ + 304, + 121, + 546, + 266 + ], + "type": "inline_equation", + "content": "q^j \\in \\mathbb{R}^2" + }, + { + "bbox": [ + 304, + 121, + 546, + 266 + ], + "type": "text", + "content": " to denote the " + }, + { + "bbox": [ + 304, + 121, + 546, + 266 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 121, + 546, + 266 + ], + "type": "text", + "content": "-th pixel location in " + }, + { + "bbox": [ + 304, + 121, + 546, + 266 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 121, + 546, + 266 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 121, + 546, + 266 + ], + "type": "inline_equation", + "content": "f(x)_k^j \\in \\mathbb{R}_+" + }, + { + "bbox": [ + 304, + 121, + 546, + 266 + ], + "type": "text", + "content": " to denote the probability of the " + }, + { + "bbox": [ + 304, + 121, + 546, + 266 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 121, + 546, + 266 + ], + "type": "text", + "content": "-th keypoint lying on " + }, + { + "bbox": [ + 304, + 121, + 546, + 266 + ], + "type": "inline_equation", + "content": "q^j" + }, + { + "bbox": [ + 304, + 121, + 546, + 266 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 304, + 121, + 546, + 266 + ], + "type": "inline_equation", + "content": "\\sigma_k" + }, + { + "bbox": [ + 304, + 121, + 546, + 266 + ], + "type": "text", + "content": " be the index permutation that sorts " + }, + { + "bbox": [ + 304, + 121, + 546, + 266 + ], + "type": "inline_equation", + "content": "f(x)_k" + }, + { + "bbox": [ + 304, + 121, + 546, + 266 + ], + "type": "text", + "content": " in nonincreasing order, i.e., " + }, + { + "bbox": [ + 304, + 121, + 546, + 266 + ], + "type": "inline_equation", + "content": "f(x)_k^{\\sigma_k(1)} \\geq \\dots \\geq f(x)_k^{\\sigma_k(HW)}" + }, + { + "bbox": [ + 304, + 121, + 546, + 266 + ], + "type": "text", + "content": ". As we will soon show, choosing the heatmap approach leads to simple and intuitive designs of the nonconformity function." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 267, + 545, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 267, + 545, + 291 + ], + "spans": [ + { + "bbox": [ + 304, + 267, + 545, + 291 + ], + "type": "text", + "content": "Conformal calibration. We design the following nonconformity function" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 337, + 299, + 545, + 315 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 299, + 545, + 315 + ], + "spans": [ + { + "bbox": [ + 337, + 299, + 545, + 315 + ], + "type": "interline_equation", + "content": "r (\\boldsymbol {y}, \\boldsymbol {f} (x)) = \\max \\left\\{\\phi \\left(y _ {k}, f (x) _ {k}\\right) \\right\\} _ {k = 1} ^ {K} \\tag {6}", + "image_path": "c20cc36bbd48964388d431d8b75c9304db9978d8d4790afb3f528f8d28d32b4e.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 323, + 545, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 323, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 304, + 323, + 545, + 358 + ], + "type": "text", + "content": "that uses " + }, + { + "bbox": [ + 304, + 323, + 545, + 358 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 304, + 323, + 545, + 358 + ], + "type": "text", + "content": " to score each keypoint and then selects the maximum score. This design considers the worst keypoint detection performance of " + }, + { + "bbox": [ + 304, + 323, + 545, + 358 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 323, + 545, + 358 + ], + "type": "text", + "content": ". We provide two designs of " + }, + { + "bbox": [ + 304, + 323, + 545, + 358 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 304, + 323, + 545, + 358 + ], + "type": "text", + "content": " below." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 359, + 545, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 359, + 545, + 397 + ], + "spans": [ + { + "bbox": [ + 304, + 359, + 545, + 397 + ], + "type": "text", + "content": "(a) Peak. Shorthand " + }, + { + "bbox": [ + 304, + 359, + 545, + 397 + ], + "type": "inline_equation", + "content": "p_k = f(x)_k^{\\sigma_k(1)}" + }, + { + "bbox": [ + 304, + 359, + 545, + 397 + ], + "type": "text", + "content": " as the peak probability in the " + }, + { + "bbox": [ + 304, + 359, + 545, + 397 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 359, + 545, + 397 + ], + "type": "text", + "content": "-th heatmap and " + }, + { + "bbox": [ + 304, + 359, + 545, + 397 + ], + "type": "inline_equation", + "content": "q_k = q^{\\sigma_k(1)}" + }, + { + "bbox": [ + 304, + 359, + 545, + 397 + ], + "type": "text", + "content": " as the pixel location attaining the peak probability, we design" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 357, + 406, + 545, + 420 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 357, + 406, + 545, + 420 + ], + "spans": [ + { + "bbox": [ + 357, + 406, + 545, + 420 + ], + "type": "interline_equation", + "content": "\\phi_ {\\text {p e a k}} \\left(y _ {k}, f (x) _ {k}\\right) = p _ {k} \\| y _ {k} - q _ {k} \\| \\quad (\\text {p e a k})", + "image_path": "0daba563d41c07d287f1ef248507a62f536e5659e4160fba29d44c3a2fbd8b8c.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 428, + 545, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 428, + 545, + 499 + ], + "spans": [ + { + "bbox": [ + 304, + 428, + 545, + 499 + ], + "type": "text", + "content": "which computes the error between the true keypoint location " + }, + { + "bbox": [ + 304, + 428, + 545, + 499 + ], + "type": "inline_equation", + "content": "y_{k}" + }, + { + "bbox": [ + 304, + 428, + 545, + 499 + ], + "type": "text", + "content": " and the most probable keypoint location " + }, + { + "bbox": [ + 304, + 428, + 545, + 499 + ], + "type": "inline_equation", + "content": "q_{k}" + }, + { + "bbox": [ + 304, + 428, + 545, + 499 + ], + "type": "text", + "content": " and scales the error by the peak probability " + }, + { + "bbox": [ + 304, + 428, + 545, + 499 + ], + "type": "inline_equation", + "content": "p_{k}" + }, + { + "bbox": [ + 304, + 428, + 545, + 499 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 428, + 545, + 499 + ], + "type": "inline_equation", + "content": "\\phi_{\\mathrm{peak}}" + }, + { + "bbox": [ + 304, + 428, + 545, + 499 + ], + "type": "text", + "content": " describes nonconformity because it becomes larger when the network " + }, + { + "bbox": [ + 304, + 428, + 545, + 499 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 428, + 545, + 499 + ], + "type": "text", + "content": " is confidently wrong (both " + }, + { + "bbox": [ + 304, + 428, + 545, + 499 + ], + "type": "inline_equation", + "content": "\\| y_{k} - q_{k}\\|" + }, + { + "bbox": [ + 304, + 428, + 545, + 499 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 428, + 545, + 499 + ], + "type": "inline_equation", + "content": "p_k" + }, + { + "bbox": [ + 304, + 428, + 545, + 499 + ], + "type": "text", + "content": " are large), implying the sample is highly nonconforming." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 500, + 545, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 500, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 500, + 545, + 555 + ], + "type": "text", + "content": "(b) Covariance. Let " + }, + { + "bbox": [ + 304, + 500, + 545, + 555 + ], + "type": "inline_equation", + "content": "\\bar{q}_k = \\sum_{j=1}^{J} f(x)_k^{\\sigma_k(j)} q^{\\sigma_k(j)}" + }, + { + "bbox": [ + 304, + 500, + 545, + 555 + ], + "type": "text", + "content": " be the expected location of the top-" + }, + { + "bbox": [ + 304, + 500, + 545, + 555 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 304, + 500, + 545, + 555 + ], + "type": "text", + "content": " most likely detections for the " + }, + { + "bbox": [ + 304, + 500, + 545, + 555 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 500, + 545, + 555 + ], + "type": "text", + "content": "-th keypoint, and " + }, + { + "bbox": [ + 304, + 500, + 545, + 555 + ], + "type": "inline_equation", + "content": "\\Sigma_k = \\sum_{j=1}^{J} f(x)_k^{\\sigma_k(j)} \\cdot (q^{\\sigma_k(j)} - \\bar{q}_k)(q^{\\sigma_k(j)} - \\bar{q}_k)^{\\top}" + }, + { + "bbox": [ + 304, + 500, + 545, + 555 + ], + "type": "text", + "content": " as the covariance, we design" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 323, + 562, + 545, + 578 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 562, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 323, + 562, + 545, + 578 + ], + "type": "interline_equation", + "content": "\\phi_ {\\operatorname {c o v}} \\left(y _ {k}, f (x) _ {k}\\right) = \\left(y _ {k} - \\bar {q} _ {k}\\right) ^ {\\top} \\Sigma_ {k} ^ {- 1} \\left(y _ {k} - \\bar {q} _ {k}\\right) \\quad (\\operatorname {c o v})", + "image_path": "6f7c7c98d2dd59851dea40ce864ee65f44a917f90f93a35e831b11a3b3287d1d.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 585, + 545, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 585, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 304, + 585, + 545, + 658 + ], + "type": "text", + "content": "which computes the squared Mahalanobis distance [59] from the groundtruth " + }, + { + "bbox": [ + 304, + 585, + 545, + 658 + ], + "type": "inline_equation", + "content": "y_{k}" + }, + { + "bbox": [ + 304, + 585, + 545, + 658 + ], + "type": "text", + "content": " to the top-" + }, + { + "bbox": [ + 304, + 585, + 545, + 658 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 304, + 585, + 545, + 658 + ], + "type": "text", + "content": " keypoint detections (represented by the mean " + }, + { + "bbox": [ + 304, + 585, + 545, + 658 + ], + "type": "inline_equation", + "content": "\\bar{q}_{k}" + }, + { + "bbox": [ + 304, + 585, + 545, + 658 + ], + "type": "text", + "content": " and covariance " + }, + { + "bbox": [ + 304, + 585, + 545, + 658 + ], + "type": "inline_equation", + "content": "\\Sigma_{k}" + }, + { + "bbox": [ + 304, + 585, + 545, + 658 + ], + "type": "text", + "content": ").2 A larger Mahalanobis distance indicates more abnormality of the heatmap " + }, + { + "bbox": [ + 304, + 585, + 545, + 658 + ], + "type": "inline_equation", + "content": "f(x)_k" + }, + { + "bbox": [ + 304, + 585, + 545, + 658 + ], + "type": "text", + "content": " (compared to the groundtruth " + }, + { + "bbox": [ + 304, + 585, + 545, + 658 + ], + "type": "inline_equation", + "content": "y_{k}" + }, + { + "bbox": [ + 304, + 585, + 545, + 658 + ], + "type": "text", + "content": ") [28], and hence implies higher nonconformity." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 304, + 664, + 545, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 664, + 545, + 693 + ], + "spans": [ + { + "bbox": [ + 304, + 664, + 545, + 693 + ], + "type": "text", + "content": "1The heatmap in the original paper [71] is not a valid probability distribution as it contains negative values and do not sum up to 1. We remove the negative values and normalize it to be a valid probability distribution." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "text", + "content": "2We only choose the top- " + }, + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "inline_equation", + "content": "(J = 100)" + }, + { + "bbox": [ + 306, + 693, + 545, + 713 + ], + "type": "text", + "content": " most likely detections on the heatmap because the heatmap can be quite noisy in practice." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 295, + 750, + 315, + 759 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 750, + 315, + 759 + ], + "spans": [ + { + "bbox": [ + 295, + 750, + 315, + 759 + ], + "type": "text", + "content": "8950" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "text", + "content": "Using the nonconformity function (6) with (peak) or (cov), we compute the nonconformity scores of the calibration set and sort them as: " + }, + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "inline_equation", + "content": "\\alpha_{\\pi(1)} \\geq \\dots \\geq \\alpha_{\\pi(n)}" + }, + { + "bbox": [ + 47, + 72, + 287, + 109 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 109, + 287, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 109, + 287, + 157 + ], + "spans": [ + { + "bbox": [ + 47, + 109, + 287, + 157 + ], + "type": "text", + "content": "Conformal prediction. Given an error rate " + }, + { + "bbox": [ + 47, + 109, + 287, + 157 + ], + "type": "inline_equation", + "content": "\\epsilon \\in (0,1)" + }, + { + "bbox": [ + 47, + 109, + 287, + 157 + ], + "type": "text", + "content": ", we first find " + }, + { + "bbox": [ + 47, + 109, + 287, + 157 + ], + "type": "inline_equation", + "content": "\\alpha_{\\pi (\\lfloor (n + 1)\\epsilon \\rfloor)}" + }, + { + "bbox": [ + 47, + 109, + 287, + 157 + ], + "type": "text", + "content": ". Then, according to the ICP set definition (3) and our nonconformity function (6), we output the ICP set for a new " + }, + { + "bbox": [ + 47, + 109, + 287, + 157 + ], + "type": "inline_equation", + "content": "x_{l + 1}" + }, + { + "bbox": [ + 47, + 109, + 287, + 157 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 168, + 287, + 212 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 168, + 287, + 212 + ], + "spans": [ + { + "bbox": [ + 47, + 168, + 287, + 212 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} F ^ {\\epsilon} (x _ {l + 1}) \\\\ = \\left\\{\\boldsymbol {y} \\in \\mathcal {Y} \\mid \\max \\left\\{\\phi \\left(y _ {k}, f \\left(x _ {l + 1}\\right)\\right) \\right\\} _ {k = 1} ^ {K} \\leq \\alpha_ {\\pi (\\lfloor (n + 1) \\epsilon \\rfloor)} \\right\\} \\\\ = \\left\\{\\boldsymbol {y} \\in \\mathcal {Y} \\mid \\phi \\left(y _ {k}, f \\left(x _ {l + 1}\\right)\\right) \\leq \\alpha_ {\\pi (\\lfloor (n + 1) \\epsilon \\rfloor)}, \\forall k \\right\\}, \\tag {7} \\\\ \\end{array}", + "image_path": "6fee9047edb2c6221d83aa9ac32f732f26b19763f34dc115b29e7162766e5d5d.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 221, + 287, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 221, + 287, + 247 + ], + "spans": [ + { + "bbox": [ + 47, + 221, + 287, + 247 + ], + "type": "text", + "content": "where we used " + }, + { + "bbox": [ + 47, + 221, + 287, + 247 + ], + "type": "inline_equation", + "content": "\\max \\{\\phi_1,\\ldots ,\\phi_K\\} \\leq \\alpha" + }, + { + "bbox": [ + 47, + 221, + 287, + 247 + ], + "type": "text", + "content": " if and only if " + }, + { + "bbox": [ + 47, + 221, + 287, + 247 + ], + "type": "inline_equation", + "content": "\\phi_{k}\\leq \\alpha" + }, + { + "bbox": [ + 47, + 221, + 287, + 247 + ], + "type": "text", + "content": " for any " + }, + { + "bbox": [ + 47, + 221, + 287, + 247 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 47, + 221, + 287, + 247 + ], + "type": "text", + "content": ". Insert (peak) into (7), we have " + }, + { + "bbox": [ + 47, + 221, + 287, + 247 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{ball}}^{\\epsilon}(x_{l + 1})" + }, + { + "bbox": [ + 47, + 221, + 287, + 247 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 256, + 287, + 283 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 256, + 287, + 283 + ], + "spans": [ + { + "bbox": [ + 59, + 256, + 287, + 283 + ], + "type": "interline_equation", + "content": "\\left\\{\\boldsymbol {y} \\in \\mathcal {Y} \\mid \\| y _ {k} - q _ {l + 1, k} \\| \\leq \\frac {\\alpha_ {\\pi (\\lfloor (n + 1) \\epsilon \\rfloor)}}{p _ {l + 1 , k}}, \\forall k \\right\\}, \\quad (\\text {b a l l})", + "image_path": "8aff2e7e4e2fde2702e3568ddbb432776bb11cf98a1b56034d1e1253ec5562ca.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 293, + 287, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 293, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 47, + 293, + 287, + 342 + ], + "type": "text", + "content": "which defines -for the " + }, + { + "bbox": [ + 47, + 293, + 287, + 342 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 47, + 293, + 287, + 342 + ], + "type": "text", + "content": " -th keypoint- a ball centered at " + }, + { + "bbox": [ + 47, + 293, + 287, + 342 + ], + "type": "inline_equation", + "content": "q_{l + 1,k}" + }, + { + "bbox": [ + 47, + 293, + 287, + 342 + ], + "type": "text", + "content": " (the most likely detection) with a radius inversely proportional to " + }, + { + "bbox": [ + 47, + 293, + 287, + 342 + ], + "type": "inline_equation", + "content": "p_{l + 1,k}" + }, + { + "bbox": [ + 47, + 293, + 287, + 342 + ], + "type": "text", + "content": " and proportional to " + }, + { + "bbox": [ + 47, + 293, + 287, + 342 + ], + "type": "inline_equation", + "content": "\\alpha_{\\pi (\\lfloor (n + 1)\\epsilon \\rfloor)}" + }, + { + "bbox": [ + 47, + 293, + 287, + 342 + ], + "type": "text", + "content": ". Similarly, insert (cov) into (7), we have " + }, + { + "bbox": [ + 47, + 293, + 287, + 342 + ], + "type": "inline_equation", + "content": "F_{\\mathrm{ellipse}}^{\\epsilon}(x_{l + 1})" + }, + { + "bbox": [ + 47, + 293, + 287, + 342 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 37, + 352, + 293, + 393 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 37, + 352, + 293, + 393 + ], + "spans": [ + { + "bbox": [ + 37, + 352, + 293, + 393 + ], + "type": "interline_equation", + "content": "\\left\\{\\boldsymbol {y} \\in \\mathcal {Y} \\mid (y _ {k} - \\bar {q} _ {l + 1, k}) ^ {\\mathsf {T}} \\frac {\\sum_ {l + 1 , k} ^ {- 1}}{\\alpha_ {\\pi (\\lfloor (n + 1) \\epsilon \\rfloor)} (y _ {k} - \\bar {q} _ {l + 1 , k})} \\leq 1, \\forall k \\right\\},", + "image_path": "8e938fe70e2d466777ed211f47aea0753b11b0aaa880bccac033f077efa96b4a.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 394, + 287, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 394, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 46, + 394, + 287, + 491 + ], + "type": "text", + "content": "which defines -for the " + }, + { + "bbox": [ + 46, + 394, + 287, + 491 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 394, + 287, + 491 + ], + "type": "text", + "content": "-th keypoint-- an ellipse centered at " + }, + { + "bbox": [ + 46, + 394, + 287, + 491 + ], + "type": "inline_equation", + "content": "\\bar{q}_{l+1,k}" + }, + { + "bbox": [ + 46, + 394, + 287, + 491 + ], + "type": "text", + "content": " (the expected location of the top-" + }, + { + "bbox": [ + 46, + 394, + 287, + 491 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 46, + 394, + 287, + 491 + ], + "type": "text", + "content": " detections) with an area proportional to " + }, + { + "bbox": [ + 46, + 394, + 287, + 491 + ], + "type": "inline_equation", + "content": "\\operatorname*{det}(\\Sigma_{l+1,k})" + }, + { + "bbox": [ + 46, + 394, + 287, + 491 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 394, + 287, + 491 + ], + "type": "inline_equation", + "content": "\\alpha_{\\pi(\\lfloor n+1\\epsilon \\rfloor)}" + }, + { + "bbox": [ + 46, + 394, + 287, + 491 + ], + "type": "text", + "content": ".3 From (ball) and (ellipse), we observe that the prediction sets become larger when (i) the heatmaps are uncertain, i.e., the peak probability is low or the covariance matrix has large determinant; and (ii) the heatmaps perform poorly on the calibration set, leading to a large " + }, + { + "bbox": [ + 46, + 394, + 287, + 491 + ], + "type": "inline_equation", + "content": "\\alpha_{\\pi(\\lfloor n+1\\epsilon \\rfloor)}" + }, + { + "bbox": [ + 46, + 394, + 287, + 491 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 491, + 287, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 491, + 287, + 563 + ], + "spans": [ + { + "bbox": [ + 46, + 491, + 287, + 563 + ], + "type": "text", + "content": "Connections to geometric vision. Our nonconformity function bears similarity to the residual function in geometric vision [4, 22, 31]. For example, the (peak) and (cov) functions are similar to the (weighted) reprojection error [31], and the \"max\" in (6) can be connected to seminal work on optimizing the " + }, + { + "bbox": [ + 46, + 491, + 287, + 563 + ], + "type": "inline_equation", + "content": "\\ell_{\\infty}" + }, + { + "bbox": [ + 46, + 491, + 287, + 563 + ], + "type": "text", + "content": " norm [39]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 563, + 288, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 563, + 288, + 684 + ], + "spans": [ + { + "bbox": [ + 46, + 563, + 288, + 684 + ], + "type": "text", + "content": "Outlier-robust nonconformity? One potential issue of the nonconformity function (6) is that a single outlier can inflate the score and the calibration quantile " + }, + { + "bbox": [ + 46, + 563, + 288, + 684 + ], + "type": "inline_equation", + "content": "\\alpha_{\\pi (\\lfloor (n + 1)\\epsilon \\rfloor)}" + }, + { + "bbox": [ + 46, + 563, + 288, + 684 + ], + "type": "text", + "content": " and lead to conservative prediction sets (e.g., when " + }, + { + "bbox": [ + 46, + 563, + 288, + 684 + ], + "type": "inline_equation", + "content": "\\pmb{f}" + }, + { + "bbox": [ + 46, + 563, + 288, + 684 + ], + "type": "text", + "content": " predicts " + }, + { + "bbox": [ + 46, + 563, + 288, + 684 + ], + "type": "inline_equation", + "content": "K - 1" + }, + { + "bbox": [ + 46, + 563, + 288, + 684 + ], + "type": "text", + "content": " keypoints perfectly but misses one keypoint). A potential remedy in geometric vision is to use robust cost functions [7,9,96]. Therefore, a natural question is whether \"robustifying\" the nonconformity function (6) can lead to better prediction sets. Here we focus on only robustifying " + }, + { + "bbox": [ + 46, + 563, + 288, + 684 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 46, + 563, + 288, + 684 + ], + "type": "text", + "content": " in (6) and provide a negative answer." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 72, + 545, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 109 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 109 + ], + "type": "text", + "content": "Proposition 2 (Invariance of ICP). Let " + }, + { + "bbox": [ + 306, + 72, + 545, + 109 + ], + "type": "inline_equation", + "content": "\\rho : \\mathbb{R}_+ \\mapsto \\mathbb{R}_+" + }, + { + "bbox": [ + 306, + 72, + 545, + 109 + ], + "type": "text", + "content": " be any monotonically increasing function. Fixing the calibration set and error rate " + }, + { + "bbox": [ + 306, + 72, + 545, + 109 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 306, + 72, + 545, + 109 + ], + "type": "text", + "content": ", the nonconformity function" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 329, + 117, + 545, + 131 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 117, + 545, + 131 + ], + "spans": [ + { + "bbox": [ + 329, + 117, + 545, + 131 + ], + "type": "interline_equation", + "content": "r _ {\\rho} (\\boldsymbol {y}, \\boldsymbol {f} (x)) = \\max \\left\\{\\rho \\left(\\phi \\left(y _ {k}, f (x) _ {k}\\right)\\right) \\right\\} _ {k = 1} ^ {K} \\tag {8}", + "image_path": "36ac294685869ae0a8a43410121ab41613c13c4625078167c4a1e8e31827f295.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 140, + 437, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 140, + 437, + 152 + ], + "spans": [ + { + "bbox": [ + 306, + 140, + 437, + 152 + ], + "type": "text", + "content": "leads to the same ICP set as (6)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 158, + 545, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 158, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 304, + 158, + 545, + 281 + ], + "type": "text", + "content": "The proof of Proposition 2 is presented in Supplementary Material. We conclude that common robust costs, such as " + }, + { + "bbox": [ + 304, + 158, + 545, + 281 + ], + "type": "inline_equation", + "content": "\\ell_1" + }, + { + "bbox": [ + 304, + 158, + 545, + 281 + ], + "type": "text", + "content": ", Huber, Geman-McClure, and Barron's adaptive kernel [7,9] (which are monotonically increasing on " + }, + { + "bbox": [ + 304, + 158, + 545, + 281 + ], + "type": "inline_equation", + "content": "[0, +\\infty]" + }, + { + "bbox": [ + 304, + 158, + 545, + 281 + ], + "type": "text", + "content": ") cannot change the ICP sets by robustifying the individual score " + }, + { + "bbox": [ + 304, + 158, + 545, + 281 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 304, + 158, + 545, + 281 + ], + "type": "text", + "content": ". However, it remains an open question whether changing the \"max\" operation in (6) can give rise to better ICP sets. For instance, replacing \"max\" with \"" + }, + { + "bbox": [ + 304, + 158, + 545, + 281 + ], + "type": "inline_equation", + "content": "\\sum" + }, + { + "bbox": [ + 304, + 158, + 545, + 281 + ], + "type": "text", + "content": " in (6)\" and using the Geman-McClure robust cost " + }, + { + "bbox": [ + 304, + 158, + 545, + 281 + ], + "type": "inline_equation", + "content": "\\rho(\\phi) = \\frac{\\phi^2}{1 + \\phi^2}" + }, + { + "bbox": [ + 304, + 158, + 545, + 281 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 304, + 158, + 545, + 281 + ], + "type": "inline_equation", + "content": "\\phi = \\phi_{\\mathrm{peak}}" + }, + { + "bbox": [ + 304, + 158, + 545, + 281 + ], + "type": "text", + "content": " results in the following ICP set" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 289, + 545, + 323 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 289, + 545, + 323 + ], + "spans": [ + { + "bbox": [ + 312, + 289, + 545, + 323 + ], + "type": "interline_equation", + "content": "\\left\\{\\boldsymbol {y} \\in \\mathcal {Y} \\mid \\sum_ {k = 1} ^ {K} \\frac {p _ {k} ^ {2} \\| y _ {k} - q _ {k} \\| ^ {2}}{1 + p _ {k} ^ {2} \\| y _ {k} - q _ {k} \\| ^ {2}} \\leq \\alpha_ {\\pi (\\lfloor (n + 1) \\epsilon \\rfloor)} \\right\\} \\tag {9}", + "image_path": "c779b173a0b7b99ec59d47f7478dc537bbac94b7e31b2820a67e671fc1757ce1.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 331, + 545, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 331, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 304, + 331, + 545, + 392 + ], + "type": "text", + "content": "that does not admit a geometric interpretation that is as simple and intuitive as the (ball) and (ellipse) sets introduced before. In fact, it is indeed the simplicity of (ball) and (ellipse) that enables us to propagate the uncertainty in keypoints to the object pose, as we will show in the next section." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 401, + 504, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 401, + 504, + 415 + ], + "spans": [ + { + "bbox": [ + 306, + 401, + 504, + 415 + ], + "type": "text", + "content": "5. Geometric Uncertainty Propagation" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 421, + 545, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 421, + 545, + 458 + ], + "spans": [ + { + "bbox": [ + 304, + 421, + 545, + 458 + ], + "type": "text", + "content": "Conformalizing the heatmaps gives us prediction sets that guarantee probabilistic coverage of the true keypoints. We unify the prediction sets (ball) and (ellipse) as" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 465, + 550, + 480 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 465, + 550, + 480 + ], + "spans": [ + { + "bbox": [ + 306, + 465, + 550, + 480 + ], + "type": "interline_equation", + "content": "F ^ {\\epsilon} (x) = \\left\\{\\boldsymbol {y} \\in \\mathcal {Y} \\mid \\left(y _ {k} - \\mu_ {k}\\right) ^ {\\top} \\Lambda_ {k} \\left(y _ {k} - \\mu_ {k}\\right) \\leq 1, \\forall k \\right\\}, \\tag {10}", + "image_path": "7af72881c341a8322d76bb0f52e8259b26c63bc0cf1c75a95d23a8e9d7fbc245.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 488, + 545, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 488, + 545, + 540 + ], + "spans": [ + { + "bbox": [ + 304, + 488, + 545, + 540 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 488, + 545, + 540 + ], + "type": "inline_equation", + "content": "\\mu_{k} = q_{l + 1,k},\\Lambda_{k} = \\frac{p_{l + 1,k}^{2}}{\\alpha_{\\pi(\\lfloor(n + 1)\\epsilon \\rfloor)}^{2}}\\mathbf{I}_{2}" + }, + { + "bbox": [ + 304, + 488, + 545, + 540 + ], + "type": "text", + "content": " for (ball), " + }, + { + "bbox": [ + 304, + 488, + 545, + 540 + ], + "type": "inline_equation", + "content": "\\mu_{k} =" + }, + { + "bbox": [ + 304, + 488, + 545, + 540 + ], + "type": "inline_equation", + "content": "\\bar{q}_{l + 1,k},\\Lambda_k = \\frac{\\sum_{l + 1,k}^{-1}}{\\alpha_\\pi(\\lfloor(n + 1)\\epsilon \\rfloor)}" + }, + { + "bbox": [ + 304, + 488, + 545, + 540 + ], + "type": "text", + "content": " for (ellipse), and we omit the subscript " + }, + { + "bbox": [ + 304, + 488, + 545, + 540 + ], + "type": "inline_equation", + "content": "l + 1" + }, + { + "bbox": [ + 304, + 488, + 545, + 540 + ], + "type": "text", + "content": " for simplicity." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 541, + 545, + 565 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 541, + 545, + 565 + ], + "spans": [ + { + "bbox": [ + 305, + 541, + 545, + 565 + ], + "type": "text", + "content": "Why not uncertainty-aware PnP? A popular way to estimate pose from (10) is to solve an uncertainty-aware PnP" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 327, + 573, + 505, + 605 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 573, + 505, + 605 + ], + "spans": [ + { + "bbox": [ + 327, + 573, + 505, + 605 + ], + "type": "interline_equation", + "content": "\\min _ {(R, t) \\in \\mathrm {S E} (3)} \\quad \\sum_ {k = 1} ^ {K} \\left(y _ {k} - \\mu_ {k}\\right) ^ {\\mathsf {T}} \\Lambda_ {k} \\left(y _ {k} - \\mu_ {k}\\right)", + "image_path": "b23eba0f1a2099b62472911768ef8fbbd12f199ac57daad9dacda8bfdb9815d2.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 328, + 608, + 545, + 620 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 608, + 545, + 620 + ], + "spans": [ + { + "bbox": [ + 328, + 608, + 545, + 620 + ], + "type": "interline_equation", + "content": "\\text {s u b j e c t} \\quad y _ {k} = \\Pi \\left(R Y _ {k} + t\\right), k = 1, \\dots , K \\tag {11}", + "image_path": "f4ec11a87e24aea19252f9f57322bcf9d73d7f9e1ecc8c6d50dc35c176177092.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "Y_{k} \\in \\mathbb{R}^{3}, k = 1, \\dots, K" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": " are the 3D object keypoints and " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\Pi(\\cdot)" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": " denotes the camera projection. We challenge this approach and point out its two drawbacks. First, it is difficult to solve (11) to global optimality due to (i) the nonconvex SE(3) constraint and (ii) the rational polynomial appearing in " + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\Pi(\\cdot)" + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": ". The best known approach to solve (11) relies on either branch-and-bound [68] or local optimization." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 57, + 701, + 279, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 701, + 279, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 701, + 279, + 714 + ], + "type": "text", + "content": "3The area of " + }, + { + "bbox": [ + 57, + 701, + 279, + 714 + ], + "type": "inline_equation", + "content": "(x - \\mu)^{\\mathsf{T}}A(x - \\mu)\\leq 1" + }, + { + "bbox": [ + 57, + 701, + 279, + 714 + ], + "type": "text", + "content": " is proportional to " + }, + { + "bbox": [ + 57, + 701, + 279, + 714 + ], + "type": "inline_equation", + "content": "\\operatorname *{det}(A^{-1})" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 295, + 750, + 314, + 760 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 750, + 314, + 760 + ], + "spans": [ + { + "bbox": [ + 295, + 750, + 314, + 760 + ], + "type": "text", + "content": "8951" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 131 + ], + "type": "text", + "content": "Second, solving (11) typically outputs a single optimal pose without uncertainty quantification. Are there other poses that attain similar costs as the optimal pose? How close is the optimal pose to the groundtruth pose? These questions remain not answered in the literature." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 133, + 287, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 133, + 287, + 169 + ], + "spans": [ + { + "bbox": [ + 46, + 133, + 287, + 169 + ], + "type": "text", + "content": "Pose UnceRtainty SET (PURSE). We propose to, instead of solving a PnP problem similar to (11), directly propagate the uncertainty in the ICP sets to the object pose." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 45, + 177, + 287, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 177, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 45, + 177, + 287, + 239 + ], + "type": "text", + "content": "Proposition 3 (PURSE). Let " + }, + { + "bbox": [ + 45, + 177, + 287, + 239 + ], + "type": "inline_equation", + "content": "s_{\\mathrm{gt}} = [\\mathrm{vec}(R_{\\mathrm{gt}})^{\\top}; t_{\\mathrm{gt}}^{\\top}]^{\\top}" + }, + { + "bbox": [ + 45, + 177, + 287, + 239 + ], + "type": "text", + "content": " be the groundtruth object pose (that lies in front of the camera). Then, the groundtruth keypoints " + }, + { + "bbox": [ + 45, + 177, + 287, + 239 + ], + "type": "inline_equation", + "content": "\\mathbf{y} = (y_1, \\ldots, y_K)" + }, + { + "bbox": [ + 45, + 177, + 287, + 239 + ], + "type": "text", + "content": " belong to the ICP set " + }, + { + "bbox": [ + 45, + 177, + 287, + 239 + ], + "type": "inline_equation", + "content": "F^{\\epsilon}(x)" + }, + { + "bbox": [ + 45, + 177, + 287, + 239 + ], + "type": "text", + "content": " in (10) if and only if " + }, + { + "bbox": [ + 45, + 177, + 287, + 239 + ], + "type": "inline_equation", + "content": "s_{\\mathrm{gt}}" + }, + { + "bbox": [ + 45, + 177, + 287, + 239 + ], + "type": "text", + "content": " belongs to the following pose uncertainty set" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 247, + 287, + 275 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 247, + 287, + 275 + ], + "spans": [ + { + "bbox": [ + 45, + 247, + 287, + 275 + ], + "type": "interline_equation", + "content": "S ^ {\\epsilon} = \\left\\{s \\in \\operatorname {S E} (3) \\left| \\begin{array}{c} s ^ {\\mathsf {T}} A _ {k} s \\leq 0, k = 1, \\ldots , K \\\\ b _ {k} ^ {\\mathsf {T}} s > 0, k = 1, \\ldots , K \\end{array} \\right. \\right\\}, (\\text {P U R S E})", + "image_path": "dfceac438e611d3f57e52d70af33002310abd5933ccbff4bb1443bbf7a885cee.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 284, + 287, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 284, + 287, + 309 + ], + "spans": [ + { + "bbox": [ + 46, + 284, + 287, + 309 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 284, + 287, + 309 + ], + "type": "inline_equation", + "content": "A_{k}\\in \\mathbb{S}^{12},b_{k}\\in \\mathbb{R}^{12},k = 1,\\ldots ,K" + }, + { + "bbox": [ + 46, + 284, + 287, + 309 + ], + "type": "text", + "content": " are constant matrices dependent on " + }, + { + "bbox": [ + 46, + 284, + 287, + 309 + ], + "type": "inline_equation", + "content": "\\mu_k,\\Lambda_k,Y_k" + }, + { + "bbox": [ + 46, + 284, + 287, + 309 + ], + "type": "text", + "content": " and camera intrinsics." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 318, + 287, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 318, + 287, + 485 + ], + "spans": [ + { + "bbox": [ + 46, + 318, + 287, + 485 + ], + "type": "text", + "content": "The detailed proof for Proposition 3 is algebraically involved and postponed to the Supplementary Material. The high-level intuition is, however, straightforward: we plug in " + }, + { + "bbox": [ + 46, + 318, + 287, + 485 + ], + "type": "inline_equation", + "content": "y_{k} = \\Pi (RY_{k} + t)" + }, + { + "bbox": [ + 46, + 318, + 287, + 485 + ], + "type": "text", + "content": " into (10) and obtain " + }, + { + "bbox": [ + 46, + 318, + 287, + 485 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 318, + 287, + 485 + ], + "type": "text", + "content": " quadratic inequalities of the form " + }, + { + "bbox": [ + 46, + 318, + 287, + 485 + ], + "type": "inline_equation", + "content": "s^{\\mathsf{T}}A_{k}s\\leq 0" + }, + { + "bbox": [ + 46, + 318, + 287, + 485 + ], + "type": "text", + "content": ". The linear inequalities " + }, + { + "bbox": [ + 46, + 318, + 287, + 485 + ], + "type": "inline_equation", + "content": "b_{k}^{\\top}s > 0" + }, + { + "bbox": [ + 46, + 318, + 287, + 485 + ], + "type": "text", + "content": " are added to enforce the (transformed) 3D keypoints lie in front of the camera. Proposition 3 implies, if we are " + }, + { + "bbox": [ + 46, + 318, + 287, + 485 + ], + "type": "inline_equation", + "content": "1 - \\epsilon" + }, + { + "bbox": [ + 46, + 318, + 287, + 485 + ], + "type": "text", + "content": " confident the groundtruth keypoints can be anywhere inside " + }, + { + "bbox": [ + 46, + 318, + 287, + 485 + ], + "type": "inline_equation", + "content": "F^{\\epsilon}(x)" + }, + { + "bbox": [ + 46, + 318, + 287, + 485 + ], + "type": "text", + "content": ", then we should also be confident any pose in (PURSE) can be the groundtruth. Viewing pose estimation as a set estimation with guaranteed probabilistic coverage of the groundtruth is fundamentally different from viewing it as computing a single pose from (11) that is (hopefully) close enough to the groundtruth." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 486, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 288, + 714 + ], + "type": "text", + "content": "RANdom SExample averaGing (RANSAG). Verifying if a given pose belongs to the PURSE is straightforward via checking the inequalities in (PURSE). However, the PURSE does not directly give us estimated poses. Therefore, we propose an efficient sampling algorithm called RANSDOM SExample averaGing (RANSAG) that is analogous to RANSAC [26] and leverages the minimal solver P3P [27], presented in Algorithm 1. The intuition is that, though it is difficult to sample directly in PURSE due to the (nonconvex) constraints, it is easy to sample from the keypoint prediction set (10) due to its simple geometry (balls and ellipses). Thus, at each iteration (line 3) RANSAG samples three keypoints (line 4-5), solves the P3P inverse problem, and accept the poses that belong to the (PURSE) (line 6). RANSAG typically returns around 100 valid samples with " + }, + { + "bbox": [ + 46, + 486, + 288, + 714 + ], + "type": "inline_equation", + "content": "T = 1000" + }, + { + "bbox": [ + 46, + 486, + 288, + 714 + ], + "type": "text", + "content": " trials. However, in difficult cases (e.g., when " + }, + { + "bbox": [ + 46, + 486, + 288, + 714 + ], + "type": "inline_equation", + "content": "S^{\\epsilon}" + }, + { + "bbox": [ + 46, + 486, + 288, + 714 + ], + "type": "text", + "content": " is small or even empty) it is possible to obtain zero samples (" + }, + { + "bbox": [ + 46, + 486, + 288, + 714 + ], + "type": "inline_equation", + "content": "S = \\emptyset" + }, + { + "bbox": [ + 46, + 486, + 288, + 714 + ], + "type": "text", + "content": "). In this situation, RANSAG samples " + }, + { + "bbox": [ + 46, + 486, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\lfloor T / 20 \\rfloor" + }, + { + "bbox": [ + 46, + 486, + 288, + 714 + ], + "type": "text", + "content": " (default 50) poses without checking if they belong to the PURSE, via sampling" + } + ] + } + ], + "index": 6 + }, + { + "type": "code", + "bbox": [ + 310, + 89, + 530, + 371 + ], + "blocks": [ + { + "bbox": [ + 312, + 75, + 491, + 87 + ], + "lines": [ + { + "bbox": [ + 312, + 75, + 491, + 87 + ], + "spans": [ + { + "bbox": [ + 312, + 75, + 491, + 87 + ], + "type": "text", + "content": "Algorithm 1: RANdom SAmple averaGing" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 310, + 89, + 530, + 371 + ], + "lines": [ + { + "bbox": [ + 310, + 89, + 530, + 371 + ], + "spans": [ + { + "bbox": [ + 310, + 89, + 530, + 371 + ], + "type": "text", + "content": "1 Input: an ICP set " + }, + { + "bbox": [ + 310, + 89, + 530, + 371 + ], + "type": "inline_equation", + "content": "F^{\\epsilon}(x)" + }, + { + "bbox": [ + 310, + 89, + 530, + 371 + ], + "type": "text", + "content": " (10) and its corresponding (PURSE) " + }, + { + "bbox": [ + 310, + 89, + 530, + 371 + ], + "type": "inline_equation", + "content": "S^{\\epsilon}" + }, + { + "bbox": [ + 310, + 89, + 530, + 371 + ], + "type": "text", + "content": "; maximum trials " + }, + { + "bbox": [ + 310, + 89, + 530, + 371 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 310, + 89, + 530, + 371 + ], + "type": "text", + "content": "; initial " + }, + { + "bbox": [ + 310, + 89, + 530, + 371 + ], + "type": "inline_equation", + "content": "\\hat{S} = \\emptyset" + }, + { + "bbox": [ + 310, + 89, + 530, + 371 + ], + "type": "text", + "content": "; \n2 Output: sample poses " + }, + { + "bbox": [ + 310, + 89, + 530, + 371 + ], + "type": "inline_equation", + "content": "S \\subset \\mathrm{SE}(3)" + }, + { + "bbox": [ + 310, + 89, + 530, + 371 + ], + "type": "text", + "content": " in PURSE, and an average pose " + }, + { + "bbox": [ + 310, + 89, + 530, + 371 + ], + "type": "inline_equation", + "content": "\\bar{s} \\in \\mathrm{SE}(3)" + }, + { + "bbox": [ + 310, + 89, + 530, + 371 + ], + "type": "text", + "content": "; \n3 for " + }, + { + "bbox": [ + 310, + 89, + 530, + 371 + ], + "type": "inline_equation", + "content": "\\tau \\gets 1" + }, + { + "bbox": [ + 310, + 89, + 530, + 371 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 310, + 89, + 530, + 371 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 310, + 89, + 530, + 371 + ], + "type": "text", + "content": " do \n4 Sample " + }, + { + "bbox": [ + 310, + 89, + 530, + 371 + ], + "type": "inline_equation", + "content": "\\{k_1, k_2, k_3\\}" + }, + { + "bbox": [ + 310, + 89, + 530, + 371 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 310, + 89, + 530, + 371 + ], + "type": "inline_equation", + "content": "[K] (k_1 \\neq k_2 \\neq k_3)" + }, + { + "bbox": [ + 310, + 89, + 530, + 371 + ], + "type": "text", + "content": "; \n5 Sample " + }, + { + "bbox": [ + 310, + 89, + 530, + 371 + ], + "type": "inline_equation", + "content": "\\hat{y}_{k_i}, i = 1, 2, 3" + }, + { + "bbox": [ + 310, + 89, + 530, + 371 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 310, + 89, + 530, + 371 + ], + "type": "inline_equation", + "content": "\\{y \\in \\mathbb{R}^2 \\mid (y - \\mu_{k_i})^\\top \\Lambda_{k_i} (y - \\mu_{k_i}) \\leq 1\\}" + }, + { + "bbox": [ + 310, + 89, + 530, + 371 + ], + "type": "text", + "content": "; \n6 " + }, + { + "bbox": [ + 310, + 89, + 530, + 371 + ], + "type": "inline_equation", + "content": "\\hat{S} \\gets \\hat{S} \\cup (S^{\\epsilon} \\cap \\mathsf{P3P}(\\{\\hat{y}_{k_i} \\leftrightarrow Y_{k_i}\\}_{i=1}^3))" + }, + { + "bbox": [ + 310, + 89, + 530, + 371 + ], + "type": "text", + "content": "; \n7 end \n8 " + }, + { + "bbox": [ + 310, + 89, + 530, + 371 + ], + "type": "inline_equation", + "content": "S = \\hat{S}" + }, + { + "bbox": [ + 310, + 89, + 530, + 371 + ], + "type": "text", + "content": "; \n9 if " + }, + { + "bbox": [ + 310, + 89, + 530, + 371 + ], + "type": "inline_equation", + "content": "\\hat{S} = \\emptyset" + }, + { + "bbox": [ + 310, + 89, + 530, + 371 + ], + "type": "text", + "content": " then \n10 for " + }, + { + "bbox": [ + 310, + 89, + 530, + 371 + ], + "type": "inline_equation", + "content": "\\tau \\gets 1" + }, + { + "bbox": [ + 310, + 89, + 530, + 371 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 310, + 89, + 530, + 371 + ], + "type": "inline_equation", + "content": "\\lfloor T/20 \\rfloor" + }, + { + "bbox": [ + 310, + 89, + 530, + 371 + ], + "type": "text", + "content": " do \n11 Sample " + }, + { + "bbox": [ + 310, + 89, + 530, + 371 + ], + "type": "inline_equation", + "content": "\\hat{y}_k, k = 1, \\dots, K" + }, + { + "bbox": [ + 310, + 89, + 530, + 371 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 310, + 89, + 530, + 371 + ], + "type": "inline_equation", + "content": "F^{\\epsilon}(x)" + }, + { + "bbox": [ + 310, + 89, + 530, + 371 + ], + "type": "text", + "content": "; \n12 " + }, + { + "bbox": [ + 310, + 89, + 530, + 371 + ], + "type": "inline_equation", + "content": "\\hat{S} \\gets \\hat{S} \\cup \\mathsf{PnP}(\\{\\hat{y}_k \\leftrightarrow Y_k\\}_{k=1}^K)" + }, + { + "bbox": [ + 310, + 89, + 530, + 371 + ], + "type": "text", + "content": "; \n13 end \n14 end \n15 " + }, + { + "bbox": [ + 310, + 89, + 530, + 371 + ], + "type": "inline_equation", + "content": "\\bar{R} = \\operatorname{proj}_{\\mathrm{SO}(3)}(\\sum_{(R_j, *) \\in \\hat{S}} R_j)" + }, + { + "bbox": [ + 310, + 89, + 530, + 371 + ], + "type": "text", + "content": "; \n16 " + }, + { + "bbox": [ + 310, + 89, + 530, + 371 + ], + "type": "inline_equation", + "content": "\\bar{t} = \\frac{1}{|\\hat{S}|} \\sum_{(*, t_j) \\in \\hat{S}} t_j" + }, + { + "bbox": [ + 310, + 89, + 530, + 371 + ], + "type": "text", + "content": "; \n17 return: " + }, + { + "bbox": [ + 310, + 89, + 530, + 371 + ], + "type": "inline_equation", + "content": "S, \\bar{s} = (\\bar{R}, \\bar{t})" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "code_body" + } + ], + "index": 8, + "sub_type": "algorithm" + }, + { + "bbox": [ + 305, + 380, + 545, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 380, + 545, + 429 + ], + "spans": [ + { + "bbox": [ + 305, + 380, + 545, + 429 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 305, + 380, + 545, + 429 + ], + "type": "text", + "content": " keypoints and solving PnP (line 9-12). After obtaining a set of poses, RANSAG performs rotation averaging (line 14) and translation averaging (line 15) to obtain an average pose " + }, + { + "bbox": [ + 305, + 380, + 545, + 429 + ], + "type": "inline_equation", + "content": "\\bar{s}" + }, + { + "bbox": [ + 305, + 380, + 545, + 429 + ], + "type": "text", + "content": ". Note that RANSAG does not check if " + }, + { + "bbox": [ + 305, + 380, + 545, + 429 + ], + "type": "inline_equation", + "content": "\\bar{s}" + }, + { + "bbox": [ + 305, + 380, + 545, + 429 + ], + "type": "text", + "content": " lies in the PURSE." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 430, + 545, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 430, + 545, + 466 + ], + "spans": [ + { + "bbox": [ + 305, + 430, + 545, + 466 + ], + "type": "text", + "content": "Worst-case error bounds. To upper bound the errors between the average pose " + }, + { + "bbox": [ + 305, + 430, + 545, + 466 + ], + "type": "inline_equation", + "content": "\\bar{s}" + }, + { + "bbox": [ + 305, + 430, + 545, + 466 + ], + "type": "text", + "content": " and the groundtruth " + }, + { + "bbox": [ + 305, + 430, + 545, + 466 + ], + "type": "inline_equation", + "content": "(R_{\\mathrm{gt}},t_{\\mathrm{gt}})" + }, + { + "bbox": [ + 305, + 430, + 545, + 466 + ], + "type": "text", + "content": ", we maximize the squared pose-to-PURSE distance:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 321, + 476, + 545, + 496 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 476, + 545, + 496 + ], + "spans": [ + { + "bbox": [ + 321, + 476, + 545, + 496 + ], + "type": "interline_equation", + "content": "d _ {\\epsilon , \\lambda} ^ {2} = \\max _ {(R, t) \\in S ^ {\\epsilon}} \\lambda \\| R - \\bar {R} \\| _ {\\mathrm {F}} ^ {2} + (1 - \\lambda) \\| t - \\bar {t} \\| ^ {2} \\tag {12}", + "image_path": "1df194c110a6dffca77af725571139f945221672f7acbf3ee08e5931428671cd.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 508, + 545, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 508, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 305, + 508, + 545, + 557 + ], + "type": "text", + "content": "given " + }, + { + "bbox": [ + 305, + 508, + 545, + 557 + ], + "type": "inline_equation", + "content": "\\lambda \\in [0,1]" + }, + { + "bbox": [ + 305, + 508, + 545, + 557 + ], + "type": "text", + "content": ". Particularly, we compute two cases " + }, + { + "bbox": [ + 305, + 508, + 545, + 557 + ], + "type": "inline_equation", + "content": "\\lambda = 1" + }, + { + "bbox": [ + 305, + 508, + 545, + 557 + ], + "type": "text", + "content": " (the maximum rotation distance) and " + }, + { + "bbox": [ + 305, + 508, + 545, + 557 + ], + "type": "inline_equation", + "content": "\\lambda = 0" + }, + { + "bbox": [ + 305, + 508, + 545, + 557 + ], + "type": "text", + "content": " (the maximum translation distance). Proposition 3 states the groundtruth " + }, + { + "bbox": [ + 305, + 508, + 545, + 557 + ], + "type": "inline_equation", + "content": "(R_{\\mathrm{gt}},t_{\\mathrm{gt}})" + }, + { + "bbox": [ + 305, + 508, + 545, + 557 + ], + "type": "text", + "content": " lies in " + }, + { + "bbox": [ + 305, + 508, + 545, + 557 + ], + "type": "inline_equation", + "content": "S^{\\epsilon}" + }, + { + "bbox": [ + 305, + 508, + 545, + 557 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 305, + 508, + 545, + 557 + ], + "type": "inline_equation", + "content": "1 - \\epsilon" + }, + { + "bbox": [ + 305, + 508, + 545, + 557 + ], + "type": "text", + "content": " probability, hence" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 334, + 567, + 545, + 581 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 567, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 334, + 567, + 545, + 581 + ], + "type": "interline_equation", + "content": "\\left\\| \\bar {R} - R _ {\\mathrm {g t}} \\right\\| _ {\\mathrm {F}} \\leq d _ {\\epsilon , 1}, \\quad \\left\\| \\bar {t} - t _ {\\mathrm {g t}} \\right\\| \\leq d _ {\\epsilon , 0} \\tag {13}", + "image_path": "f5d444aeb3c80d4f027dc27072521496c3d98bca6ea52489f0882a9859a70328.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 592, + 424, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 592, + 424, + 604 + ], + "spans": [ + { + "bbox": [ + 305, + 592, + 424, + 604 + ], + "type": "text", + "content": "holds with probability " + }, + { + "bbox": [ + 305, + 592, + 424, + 604 + ], + "type": "inline_equation", + "content": "1 - \\epsilon" + }, + { + "bbox": [ + 305, + 592, + 424, + 604 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 605, + 545, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 654 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 654 + ], + "type": "text", + "content": "Computing the bounds. Problem (12) is nonconvex due to the constraints of the (PURSE) " + }, + { + "bbox": [ + 304, + 605, + 545, + 654 + ], + "type": "inline_equation", + "content": "S^{\\epsilon}" + }, + { + "bbox": [ + 304, + 605, + 545, + 654 + ], + "type": "text", + "content": ". We relax the nonconvex problem (12) into a convex semidefinite program (SDP) and employ off-the-shelf solvers to optimize" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 306, + 664, + 545, + 684 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 664, + 545, + 684 + ], + "spans": [ + { + "bbox": [ + 306, + 664, + 545, + 684 + ], + "type": "text", + "content": "Here we switch from P3P to PnP because PnP uses all " + }, + { + "bbox": [ + 306, + 664, + 545, + 684 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 306, + 664, + 545, + 684 + ], + "type": "text", + "content": " keypoints and there is less ambiguity in its solution." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 684, + 545, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 684, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 306, + 684, + 545, + 712 + ], + "type": "text", + "content": "In Algorithm 1 we use rotation averaging with the Chordal distance metric. The user is free to choose other single rotation averaging algorithms with different distance metrics [30]." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 295, + 750, + 315, + 759 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 750, + 315, + 759 + ], + "spans": [ + { + "bbox": [ + 295, + 750, + 315, + 759 + ], + "type": "text", + "content": "8952" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 169 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 169 + ], + "type": "text", + "content": "the SDP [14, 40, 98]. Two possible outcomes can happen: (i) the optimal SDP value coincides with the optimal value of (12). The relaxation is said to be exact and one can extract an optimal solution of (12) from the SDP, or (ii) the relaxation is not exact, but the optimal SDP value still provides an upper bound for the optimal value of (12). Therefore, we either exactly compute " + }, + { + "bbox": [ + 46, + 72, + 289, + 169 + ], + "type": "inline_equation", + "content": "d_{\\epsilon, \\lambda}^2" + }, + { + "bbox": [ + 46, + 72, + 289, + 169 + ], + "type": "text", + "content": " or find an upper bound, both can bound the worst-case error (cf. (13))." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 59, + 170, + 282, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 170, + 282, + 182 + ], + "spans": [ + { + "bbox": [ + 59, + 170, + 282, + 182 + ], + "type": "text", + "content": "We end with a remark about computing tighter bounds." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 45, + 187, + 287, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 187, + 287, + 236 + ], + "spans": [ + { + "bbox": [ + 45, + 187, + 287, + 236 + ], + "type": "text", + "content": "Remark 4 (Best Worst-case Error Bounds). (12) can be used to bound errors for all possible pose estimators (e.g., from PnP (11)). What is the best estimator that attains the smallest error bounds? This boils down to solving" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 242, + 287, + 269 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 242, + 287, + 269 + ], + "spans": [ + { + "bbox": [ + 48, + 242, + 287, + 269 + ], + "type": "interline_equation", + "content": "\\min _ {(\\bar {R}, \\bar {t}) \\in \\mathrm {S E} (3)} \\left[ \\max _ {(R, t) \\in S ^ {\\epsilon}} \\lambda \\| R - \\bar {R} \\| _ {\\mathrm {F}} ^ {2} + (1 - \\lambda) \\| t - \\bar {t} \\| ^ {2} \\right] \\tag {14}", + "image_path": "910592c6d052f096ceabb42bf49f60dfa4883560d1d431681f379d377b766c47.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 274, + 289, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 274, + 289, + 383 + ], + "spans": [ + { + "bbox": [ + 46, + 274, + 289, + 383 + ], + "type": "text", + "content": "whose solution is known as the Chebyshev center [25, 63] of the PURSE " + }, + { + "bbox": [ + 46, + 274, + 289, + 383 + ], + "type": "inline_equation", + "content": "S^{\\epsilon}" + }, + { + "bbox": [ + 46, + 274, + 289, + 383 + ], + "type": "text", + "content": ". Unfortunately, problem (14) is more challenging than (12) and there is no efficient algorithm to solve it to global optimality. In the Supplementary Material, we evaluate the worst-case error bounds for multiple " + }, + { + "bbox": [ + 46, + 274, + 289, + 383 + ], + "type": "inline_equation", + "content": "(\\bar{R},\\bar{t})" + }, + { + "bbox": [ + 46, + 274, + 289, + 383 + ], + "type": "text", + "content": " samples, select the smallest bounds, and compare them with those of the average pose. An interesting future research direction is to explore differentiable optimization [73] or bilevel polynomial optimization [66] to solve (14)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 388, + 128, + 401 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 388, + 128, + 401 + ], + "spans": [ + { + "bbox": [ + 47, + 388, + 128, + 401 + ], + "type": "text", + "content": "6. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 407, + 287, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 407, + 287, + 503 + ], + "spans": [ + { + "bbox": [ + 46, + 407, + 287, + 503 + ], + "type": "text", + "content": "We test our approach on the LineMOD Occlusion (LM-O) dataset [11] to (i) justify the exchangeability assumption (Theorem 1) and suggest best practices for applying conformal prediction; (ii) evaluate the empirical coverage of the PURSE and verify the correctness of Theorem 1, and (iii) compute the worst-case error bounds and demonstrate tightness or looseness. We also (iv) show that the average pose achieves better or similar accuracy as other approaches." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 503, + 288, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 503, + 288, + 612 + ], + "spans": [ + { + "bbox": [ + 46, + 503, + 288, + 612 + ], + "type": "text", + "content": "Implementation and runtime. We set " + }, + { + "bbox": [ + 46, + 503, + 288, + 612 + ], + "type": "inline_equation", + "content": "T = 1000" + }, + { + "bbox": [ + 46, + 503, + 288, + 612 + ], + "type": "text", + "content": " in RANSAG; use OpenGV [44] for P3P and PnP; and add a redundant " + }, + { + "bbox": [ + 46, + 503, + 288, + 612 + ], + "type": "inline_equation", + "content": "\\| t \\| \\leq 5" + }, + { + "bbox": [ + 46, + 503, + 288, + 612 + ], + "type": "text", + "content": " in (PURSE) to ensure bounded translation. All procedures are implemented in Python except SDP relaxations are implemented in Matlab. The runtime of RANSAG is comparable to RANSAC and below one second. The runtime of computing (12) via SDPs is around 8 seconds on a workstation with 2.2GHz AMD CPUs. The (second-order) SDP relaxations are almost always exact." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 616, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 46, + 616, + 287, + 683 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 616, + 287, + 683 + ], + "spans": [ + { + "bbox": [ + 46, + 616, + 287, + 683 + ], + "type": "text", + "content": "6We omit the technical details and refer the interested reader to [98, Section 2] for a pragmatic introduction to SDP relaxations. In practice, we use the code provided by [98] in https://github.com/MIT-SPARK/CertifiablyRobustPerception, apply a second-order SDP relaxation to (12), and use MOSEK [65] to solve the SDP (in about 8 seconds). Solving a first-order SDP relaxation of (12) takes about 0.1 second but yields looser bounds." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 683, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 683, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 683, + 288, + 713 + ], + "type": "text", + "content": "The PURSE can potentially be empty, leading to infeasibility of problem (12). In such cases, empirically the SDP solver returns \"PRIMAL_INFEASIBLE\" (red squares lying on the " + }, + { + "bbox": [ + 46, + 683, + 288, + 713 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 46, + 683, + 288, + 713 + ], + "type": "text", + "content": "-axis of Fig. 3)." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": "Dataset and exchangeability. The LM-O dataset contains 1214 test images capturing 8 different objects on a table, of which 200 images were chosen by BOP19'20 [35]. We use the 200 images for calibration and the entire 1214 images for testing. As mentioned in Section 3, if the dataset was collected as a single video sequence under natural motion (e.g., a straight line), then the exchangeability assumption would fail. However, [33] described the data collection:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 324, + 172, + 526, + 244 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 172, + 526, + 244 + ], + "spans": [ + { + "bbox": [ + 324, + 172, + 526, + 244 + ], + "type": "text", + "content": "In order to guarantee a well distributed pose space sampling of the dataset pictures, we uniformly divided the upper hemisphere of the objects into equally distant pieces and took at most one image per piece. As a result, our sequences provide uniformly distributed views ..." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 248, + 545, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 248, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 304, + 248, + 545, + 308 + ], + "type": "text", + "content": "which indicates the 1214 images are independent (cf. [33, Figs. 5-6]) and therefore exchangeable. This demonstrates a good example for data collection –to equally divide the parameter space and collect one observation per division– so the guarantees offered by conformal prediction are valid." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 309, + 546, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 309, + 546, + 500 + ], + "spans": [ + { + "bbox": [ + 304, + 309, + 546, + 500 + ], + "type": "text", + "content": "Empirical coverage. Our approach conformalizes the heatmaps [79] as (ball) or (ellipse). The implementation8 of [79] uses either groundtruth or Faster RCNN [75] bounding boxes, giving four variants of our approach: groundtruth box plus (ball) or (ellipse) (labels: gt-ball, gt-ellipse), and Faster RCNN box plus (ball) or (ellipse) (labels: frcnn-ball, frcnn-ellipse). Fig. 3 left column shows the empirical coverage (i.e., the percentage of images whose groundtruth poses lie in (PURSE)) of all four variants with " + }, + { + "bbox": [ + 304, + 309, + 546, + 500 + ], + "type": "inline_equation", + "content": "\\epsilon = 0.1" + }, + { + "bbox": [ + 304, + 309, + 546, + 500 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 309, + 546, + 500 + ], + "type": "inline_equation", + "content": "\\epsilon = 0.4" + }, + { + "bbox": [ + 304, + 309, + 546, + 500 + ], + "type": "text", + "content": ". We see the empirical coverage is around " + }, + { + "bbox": [ + 304, + 309, + 546, + 500 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 304, + 309, + 546, + 500 + ], + "type": "text", + "content": " when " + }, + { + "bbox": [ + 304, + 309, + 546, + 500 + ], + "type": "inline_equation", + "content": "\\epsilon = 0.1" + }, + { + "bbox": [ + 304, + 309, + 546, + 500 + ], + "type": "text", + "content": " and around " + }, + { + "bbox": [ + 304, + 309, + 546, + 500 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 304, + 309, + 546, + 500 + ], + "type": "text", + "content": " when " + }, + { + "bbox": [ + 304, + 309, + 546, + 500 + ], + "type": "inline_equation", + "content": "\\epsilon = 0.4" + }, + { + "bbox": [ + 304, + 309, + 546, + 500 + ], + "type": "text", + "content": ", for all 8 objects. Though the empirical coverage can deviate from " + }, + { + "bbox": [ + 304, + 309, + 546, + 500 + ], + "type": "inline_equation", + "content": "1 - \\epsilon" + }, + { + "bbox": [ + 304, + 309, + 546, + 500 + ], + "type": "text", + "content": ", it generally stays within " + }, + { + "bbox": [ + 304, + 309, + 546, + 500 + ], + "type": "inline_equation", + "content": "\\pm 5\\%" + }, + { + "bbox": [ + 304, + 309, + 546, + 500 + ], + "type": "text", + "content": " and mostly goes above " + }, + { + "bbox": [ + 304, + 309, + 546, + 500 + ], + "type": "inline_equation", + "content": "1 - \\epsilon" + }, + { + "bbox": [ + 304, + 309, + 546, + 500 + ], + "type": "text", + "content": ", which is encouraging given that our calibration set only has size " + }, + { + "bbox": [ + 304, + 309, + 546, + 500 + ], + "type": "inline_equation", + "content": "n = 200" + }, + { + "bbox": [ + 304, + 309, + 546, + 500 + ], + "type": "text", + "content": ". Fig. 1 (b) plots examples of the prediction sets. More examples are shown in the Supplementary Material." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 501, + 547, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 501, + 547, + 693 + ], + "spans": [ + { + "bbox": [ + 304, + 501, + 547, + 693 + ], + "type": "text", + "content": "Worst-case error bounds. Fig. 3 middle column plots the worst-case rotation error bound (" + }, + { + "bbox": [ + 304, + 501, + 547, + 693 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 501, + 547, + 693 + ], + "type": "text", + "content": "-axis) vs. the actual rotation error between the average pose and the groundtruth (" + }, + { + "bbox": [ + 304, + 501, + 547, + 693 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 501, + 547, + 693 + ], + "type": "text", + "content": "-axis) for our approach using the gt-ball setup (results for gt-ellipse, fccnn-ball and fccnn-ellipse are similar and provided in the Supplementary Material). First, when the PURSE covers the groundtruth (blue circles), the rotation error bound is always larger than the actual error (i.e., the blue circles never cross the " + }, + { + "bbox": [ + 304, + 501, + 547, + 693 + ], + "type": "inline_equation", + "content": "y = x" + }, + { + "bbox": [ + 304, + 501, + 547, + 693 + ], + "type": "text", + "content": " diagonal). Second, when the error rate is increased from " + }, + { + "bbox": [ + 304, + 501, + 547, + 693 + ], + "type": "inline_equation", + "content": "\\epsilon = 0.1" + }, + { + "bbox": [ + 304, + 501, + 547, + 693 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 501, + 547, + 693 + ], + "type": "inline_equation", + "content": "\\epsilon = 0.4" + }, + { + "bbox": [ + 304, + 501, + 547, + 693 + ], + "type": "text", + "content": ", we observe a shift of the blue circles towards " + }, + { + "bbox": [ + 304, + 501, + 547, + 693 + ], + "type": "inline_equation", + "content": "y = x" + }, + { + "bbox": [ + 304, + 501, + 547, + 693 + ], + "type": "text", + "content": ", indicating the error bounds get tightened. Third, our bounds are reasonably tight for most test images (i.e., the bottom-left cluster of blue circles) especially when " + }, + { + "bbox": [ + 304, + 501, + 547, + 693 + ], + "type": "inline_equation", + "content": "\\epsilon = 0.4" + }, + { + "bbox": [ + 304, + 501, + 547, + 693 + ], + "type": "text", + "content": ". However, they can become overly conservative (i.e., the line of blue circles on the right-side boundary) due to the keypoint prediction" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 315, + 702, + 497, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 702, + 497, + 713 + ], + "spans": [ + { + "bbox": [ + 315, + 702, + 497, + 713 + ], + "type": "text", + "content": "8https://github.com/yufu-wang/6D_Pose" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 295, + 750, + 315, + 759 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 750, + 315, + 759 + ], + "spans": [ + { + "bbox": [ + 295, + 750, + 315, + 759 + ], + "type": "text", + "content": "8953" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 88, + 47, + 242, + 167 + ], + "blocks": [ + { + "bbox": [ + 88, + 47, + 242, + 167 + ], + "lines": [ + { + "bbox": [ + 88, + 47, + 242, + 167 + ], + "spans": [ + { + "bbox": [ + 88, + 47, + 242, + 167 + ], + "type": "image", + "image_path": "5392020343fc77e977bd7590c0f66a74ae84943612cdc7e9359143bb8f8b4b54.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 250, + 47, + 375, + 167 + ], + "blocks": [ + { + "bbox": [ + 250, + 47, + 375, + 167 + ], + "lines": [ + { + "bbox": [ + 250, + 47, + 375, + 167 + ], + "spans": [ + { + "bbox": [ + 250, + 47, + 375, + 167 + ], + "type": "image", + "image_path": "123b08c508a933f19417c16de49223b14183b92242c903bc165ad0128a2100f2.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 383, + 47, + 504, + 168 + ], + "blocks": [ + { + "bbox": [ + 383, + 47, + 504, + 168 + ], + "lines": [ + { + "bbox": [ + 383, + 47, + 504, + 168 + ], + "spans": [ + { + "bbox": [ + 383, + 47, + 504, + 168 + ], + "type": "image", + "image_path": "2f692da27de3c202e07df39842aaa3063addc2abc6f296526eff7fff967a4682.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 86, + 168, + 242, + 289 + ], + "blocks": [ + { + "bbox": [ + 86, + 168, + 242, + 289 + ], + "lines": [ + { + "bbox": [ + 86, + 168, + 242, + 289 + ], + "spans": [ + { + "bbox": [ + 86, + 168, + 242, + 289 + ], + "type": "image", + "image_path": "c578ce8ac9470503174d157a3a2f01951c6b87afe521f504aa2651772a3c4613.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 291, + 545, + 357 + ], + "lines": [ + { + "bbox": [ + 45, + 291, + 545, + 357 + ], + "spans": [ + { + "bbox": [ + 45, + 291, + 545, + 357 + ], + "type": "text", + "content": "Figure 3. Empirical coverage (left) and worst-case error bounds (middle: rotation, right: translation). Top: " + }, + { + "bbox": [ + 45, + 291, + 545, + 357 + ], + "type": "inline_equation", + "content": "\\epsilon = 0.1" + }, + { + "bbox": [ + 45, + 291, + 545, + 357 + ], + "type": "text", + "content": ", bottom: " + }, + { + "bbox": [ + 45, + 291, + 545, + 357 + ], + "type": "inline_equation", + "content": "\\epsilon = 0.4" + }, + { + "bbox": [ + 45, + 291, + 545, + 357 + ], + "type": "text", + "content": ". For middle and right columns, " + }, + { + "bbox": [ + 45, + 291, + 545, + 357 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 45, + 291, + 545, + 357 + ], + "type": "text", + "content": "-axis represents the worst-case error bounds computed from (12), " + }, + { + "bbox": [ + 45, + 291, + 545, + 357 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 45, + 291, + 545, + 357 + ], + "type": "text", + "content": "-axis represents the actual error between average pose and groundtruth pose. The area below the diagonal " + }, + { + "bbox": [ + 45, + 291, + 545, + 357 + ], + "type": "inline_equation", + "content": "y = x" + }, + { + "bbox": [ + 45, + 291, + 545, + 357 + ], + "type": "text", + "content": " indicates correctness of the bounds (i.e., bound " + }, + { + "bbox": [ + 45, + 291, + 545, + 357 + ], + "type": "inline_equation", + "content": "\\geq" + }, + { + "bbox": [ + 45, + 291, + 545, + 357 + ], + "type": "text", + "content": " error), and points that are closer to the diagonal from below indicate tighter bounds (perfect if precisely lie on the diagonal). Blue circles plot cases where the PURSE covers the groundtruth pose and red squares plot cases were the PURSE does not cover the groundtruth. Notice that blue circles never cross the diagonal and our bounds are correct when the PURSE contains the pose (which holds with " + }, + { + "bbox": [ + 45, + 291, + 545, + 357 + ], + "type": "inline_equation", + "content": "1 - \\epsilon" + }, + { + "bbox": [ + 45, + 291, + 545, + 357 + ], + "type": "text", + "content": " marginal probability)." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 252, + 170, + 373, + 289 + ], + "blocks": [ + { + "bbox": [ + 252, + 170, + 373, + 289 + ], + "lines": [ + { + "bbox": [ + 252, + 170, + 373, + 289 + ], + "spans": [ + { + "bbox": [ + 252, + 170, + 373, + 289 + ], + "type": "image", + "image_path": "a3d9c202b27254808b3d11653350e7617a92cd01cb9bfe52aaea13e91c8a093e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 383, + 170, + 503, + 289 + ], + "blocks": [ + { + "bbox": [ + 383, + 170, + 503, + 289 + ], + "lines": [ + { + "bbox": [ + 383, + 170, + 503, + 289 + ], + "spans": [ + { + "bbox": [ + 383, + 170, + 503, + 289 + ], + "type": "image", + "image_path": "1cf8fb534a62b66a438eac51813d3b1fad772041567f93f5d37be085727eca48.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 71, + 357, + 520, + 471 + ], + "blocks": [ + { + "bbox": [ + 71, + 357, + 520, + 471 + ], + "lines": [ + { + "bbox": [ + 71, + 357, + 520, + 471 + ], + "spans": [ + { + "bbox": [ + 71, + 357, + 520, + 471 + ], + "type": "table", + "html": "
objectsBaselines (results adapted from [72])Conformalized heatmap
Tekin [89]PoseCNN [95]Oberweger [67]PVNet [72]gt-ballgt-ellipsefrcnn-ballfrcnn-ellipse
ε = 0.1ε = 0.4ε = 0.1ε = 0.4ε = 0.1ε = 0.4ε = 0.1
ape7.0134.669.669.1477.7079.5279.2679.8870.2071.0168.84
can11.2015.1082.6086.0973.4175.9775.8178.1367.5269.8167.69
cat3.6210.4065.1065.1287.3690.5989.5490.1174.9580.2368.98
duck5.0731.8061.4061.4482.7183.0884.0283.5579.3080.6280.06
driller1.407.4073.8073.0679.3282.5481.2282.0458.4865.9258.06
eggbox-1.9013.108.43000.090.18000
glue4.7013.8054.9055.3756.4971.0871.6972.9330.0347.1841.96
holepuncher8.2623.1066.4069.8481.6582.8983.2284.3074.9677.8576.28
average6.1617.2060.9061.0667.3370.7170.6171.3956.9361.5857.73
", + "image_path": "37c172e1d97c06f734819ccdbca9fe08a041f9078197ad9289296a4ed9d2b515.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 472, + 544, + 495 + ], + "lines": [ + { + "bbox": [ + 46, + 472, + 544, + 495 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 544, + 495 + ], + "type": "text", + "content": "Table 1. Success rates of baseline methods and our conformalized heatmap (using the average pose) based on the 2D projection metric (i.e., a pose estimation is considered successful if the average 2D reprojection error is below 5 pixels)." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 498, + 287, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 569 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 569 + ], + "type": "text", + "content": "sets become too large. Fig. 3 right column plots similar results for the translation. The Supplementary Material gives a more detailed analysis of this conservatism, wherein we also solve (12) for multiple samples computed by RANSAG, choose the minimum bound, and compare them with those obtained for the average pose (cf. Remark 4)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 570, + 287, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 666 + ], + "type": "text", + "content": "Accuracy of the average pose. We compare the accuracy of our average pose with other methods according to the 2D projection metric (an estimation is correct if the mean reprojection error is below 5 pixels). Table 1 shows: (i) our average pose achieves significantly better success rates when using groundtruth bounding boxes, and similar success rates when using Faster RCNN; (ii) the accuracy of the average pose increases when " + }, + { + "bbox": [ + 46, + 570, + 287, + 666 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 46, + 570, + 287, + 666 + ], + "type": "text", + "content": " increases." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 672, + 125, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 672, + 125, + 684 + ], + "spans": [ + { + "bbox": [ + 47, + 672, + 125, + 684 + ], + "type": "text", + "content": "7. Conclusions" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "We applied inductive conformal prediction to conformalize heatmap predictions as circular or elliptical prediction" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 498, + 545, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 594 + ], + "type": "text", + "content": "sets that guarantee probabilistic coverage of the groundtruth keypoints, propagated the uncertainty in keypoints to the object pose to form a PURSE, designed RANSAG to sample from PURSE and compute an average pose, and used SDP relaxations to bound worst-case estimation errors. We validated our theory on the LineMOD Occlusion dataset. Future research will investigate better nonconformity functions, and applications to other vision problems." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 605, + 404, + 618 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 605, + 404, + 618 + ], + "spans": [ + { + "bbox": [ + 306, + 605, + 404, + 618 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 624, + 545, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 624, + 545, + 708 + ], + "spans": [ + { + "bbox": [ + 304, + 624, + 545, + 708 + ], + "type": "text", + "content": "We thank Yufu Wang and Kostas Daniilidis for providing the heatmap implementation to detect semantic keypoints; Rachel Luo for discussing the exchangeability assumption in conformal prediction; Luca Carlone for pointing out related work on unknown-but-bounded noise (set membership) estimation in control theory, and anonymous reviewers for providing valuable feedback." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 750, + 315, + 759 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 750, + 315, + 759 + ], + "spans": [ + { + "bbox": [ + 295, + 750, + 315, + 759 + ], + "type": "text", + "content": "8954" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 91, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 58, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 58, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Anastasios Angelopoulos, Stephen Bates, Jitendra Malik, and Michael I Jordan. Uncertainty sets for image classifiers using conformal prediction. In Intl. Conf. on Learning Representations (ICLR), 2021. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 137, + 287, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 137, + 287, + 179 + ], + "spans": [ + { + "bbox": [ + 58, + 137, + 287, + 179 + ], + "type": "text", + "content": "[2] Anastasios N Angelopoulos and Stephen Bates. A gentle introduction to conformal prediction and distribution-free uncertainty quantification. arXiv preprint arXiv:2107.07511, 2021. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 182, + 287, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 182, + 287, + 214 + ], + "spans": [ + { + "bbox": [ + 58, + 182, + 287, + 214 + ], + "type": "text", + "content": "[3] Anastasios N Angelopoulos, Stephen Bates, Adam Fisch, Lihua Lei, and Tal Schuster. Conformal risk control. arXiv preprint arXiv:2208.02814, 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 217, + 287, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 217, + 287, + 258 + ], + "spans": [ + { + "bbox": [ + 58, + 217, + 287, + 258 + ], + "type": "text", + "content": "[4] Pasquale Antonante, Vasileios Tzoumas, Heng Yang, and Luca Carlone. Outlier-robust estimation: Hardness, minimally tuned algorithms, and applications. IEEE Trans. Robotics, 38(1):281-301, 2021. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 262, + 287, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 262, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 58, + 262, + 287, + 304 + ], + "type": "text", + "content": "[5] Boaz Barak and David Steurer. Proofs, beliefs, and algorithms through the lens of sum-of-squares. Course notes: http://www-sumofsquares.org/public/index.html, 1, 2016. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 58, + 307, + 287, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 307, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 58, + 307, + 287, + 327 + ], + "type": "text", + "content": "[6] Timothy D Barfoot. State estimation for robotics. Cambridge University Press, 2017. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 58, + 330, + 287, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 330, + 287, + 372 + ], + "spans": [ + { + "bbox": [ + 58, + 330, + 287, + 372 + ], + "type": "text", + "content": "[7] Jonathan T Barron. A general and adaptive robust loss function. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4331-4339, 2019. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 58, + 375, + 287, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 375, + 287, + 418 + ], + "spans": [ + { + "bbox": [ + 58, + 375, + 287, + 418 + ], + "type": "text", + "content": "[8] Stephen Bates, Anastasios Angelopoulos, Lihua Lei, Jitendra Malik, and Michael Jordan. Distribution-free, risk-controlling prediction sets. Journal of the ACM (JACM), 68(6):1-34, 2021. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 58, + 420, + 287, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 420, + 287, + 463 + ], + "spans": [ + { + "bbox": [ + 58, + 420, + 287, + 463 + ], + "type": "text", + "content": "[9] Michael J Black and Anand Rangarajan. On the unification of line processes, outlier rejection, and robust statistics with applications in early vision. *Intl. J. of Computer Vision*, 19(1):57-91, 1996. 2, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 465, + 287, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 465, + 287, + 487 + ], + "spans": [ + { + "bbox": [ + 53, + 465, + 287, + 487 + ], + "type": "text", + "content": "[10] Andrew Blake and Andrew Zisserman. Visual reconstruction. MIT press, 1987. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 489, + 287, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 489, + 287, + 543 + ], + "spans": [ + { + "bbox": [ + 53, + 489, + 287, + 543 + ], + "type": "text", + "content": "[11] Eric Brachmann, Alexander Krull, Frank Michel, Stefan Gumhold, Jamie Shotton, and Carsten Rother. Learning 6d object pose estimation using 3d object coordinates. In European Conf. on Computer Vision (ECCV), pages 536-551. Springer, 2014. 2, 4, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 544, + 287, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 544, + 287, + 599 + ], + "spans": [ + { + "bbox": [ + 53, + 544, + 287, + 599 + ], + "type": "text", + "content": "[12] Eric Brachmann, Frank Michel, Alexander Krull, Michael Ying Yang, Stefan Gumhold, et al. Uncertainty-driven 6d pose estimation of objects and scenes from a single rgb image. In IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 3364-3372, 2016. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 601, + 287, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 601, + 287, + 644 + ], + "spans": [ + { + "bbox": [ + 53, + 601, + 287, + 644 + ], + "type": "text", + "content": "[13] Eric Brachmann and Carsten Rother. Learning less is more-6d camera localization via 3d surface regression. In IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 4654–4662, 2018. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 53, + 647, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 647, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 53, + 647, + 287, + 689 + ], + "type": "text", + "content": "[14] Jesus Briales, Laurent Kneip, and Javier Gonzalez-Jimenez. A certifiably globally optimal solution to the non-minimal relative pose problem. In IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 145-154, 2018. 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 53, + 692, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 692, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 53, + 692, + 287, + 713 + ], + "type": "text", + "content": "[15] Dylan Campbell, Liu Liu, and Stephen Gould. Solving the blind perspective-n-point problem end-to-end with robust" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 331, + 73, + 545, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 73, + 545, + 105 + ], + "spans": [ + { + "bbox": [ + 331, + 73, + 545, + 105 + ], + "type": "text", + "content": "differentiable geometric optimization. In European Conf. on Computer Vision (ECCV), pages 244-261. Springer, 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 108, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 108, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 312, + 108, + 545, + 162 + ], + "type": "text", + "content": "[16] Dylan Campbell, Lars Petersson, Laurent Kneip, and Hong-dong Li. Globally-optimal inlier set maximisation for simultaneous camera pose and feature correspondence. In Intl. Conf. on Computer Vision (ICCV), pages 1-10, 2017. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 165, + 545, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 165, + 545, + 219 + ], + "spans": [ + { + "bbox": [ + 312, + 165, + 545, + 219 + ], + "type": "text", + "content": "[17] Emmanuel J Candes, Justin K Romberg, and Terence Tao. Stable signal recovery from incomplete and inaccurate measurements. Communications on Pure and Applied Mathematics: A Journal Issued by the Courant Institute of Mathematical Sciences, 59(8):1207-1223, 2006. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 222, + 545, + 254 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 222, + 545, + 254 + ], + "spans": [ + { + "bbox": [ + 312, + 222, + 545, + 254 + ], + "type": "text", + "content": "[18] Luca Carlone. Estimation contracts for outlier-robust geometric perception. arXiv preprint arXiv:2208.10521, 2022.3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 257, + 545, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 257, + 545, + 312 + ], + "spans": [ + { + "bbox": [ + 312, + 257, + 545, + 312 + ], + "type": "text", + "content": "[19] Bo Chen, Jiewei Cao, Alvaro Parra, and Tat-Jun Chin. Satellite pose estimation with deep landmark regression and nonlinear pose refinement. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 0–0, 2019. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 314, + 545, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 314, + 545, + 368 + ], + "spans": [ + { + "bbox": [ + 312, + 314, + 545, + 368 + ], + "type": "text", + "content": "[20] Bo Chen, Alvaro Parra, Jiewei Cao, Nan Li, and Tat-Jun Chin. End-to-end learnable geometric vision by backpropagating pnp optimization. In IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 8100-8109, 2020. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 371, + 545, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 371, + 545, + 426 + ], + "spans": [ + { + "bbox": [ + 312, + 371, + 545, + 426 + ], + "type": "text", + "content": "[21] Hansheng Chen, Pichao Wang, Fan Wang, Wei Tian, Lu Xiong, and Hao Li. Epro-pnp: Generalized end-to-end probabilistic perspective-n-points for monocular object pose estimation. In IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 2781-2790, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 428, + 545, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 428, + 545, + 460 + ], + "spans": [ + { + "bbox": [ + 312, + 428, + 545, + 460 + ], + "type": "text", + "content": "[22] Tat-Jun Chin, Zhipeng Cai, and Frank Neumann. Robust fitting in computer vision: Easy or hard? In European Conf. on Computer Vision (ECCV), pages 701-716, 2018. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 463, + 545, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 463, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 312, + 463, + 545, + 529 + ], + "type": "text", + "content": "[23] Florence de Grancey, Jean-Luc Adam, Lucian Alecu, Sébastien Gerchinovitz, Franck Mamalet, and David Vigouroux. Object detection with probabilistic guarantees: A conformal prediction approach. In International Conference on Computer Safety, Reliability, and Security, pages 316-329. Springer, 2022. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 531, + 545, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 531, + 545, + 575 + ], + "spans": [ + { + "bbox": [ + 312, + 531, + 545, + 575 + ], + "type": "text", + "content": "[24] Ilias Diakonikolas, Daniel M Kane, Sushrut Karmalkar, Ankit Pensia, and Thanasis Pittas. List-decodable sparse mean estimation via difference-of-pairs filtering. arXiv preprint arXiv:2206.05245, 2022.3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 578, + 545, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 578, + 545, + 620 + ], + "spans": [ + { + "bbox": [ + 312, + 578, + 545, + 620 + ], + "type": "text", + "content": "[25] Yonina C Eldar, Amir Beck, and Marc Teboulle. A minimax chebyshev estimator for bounded error estimation. IEEE transactions on signal processing, 56(4):1388-1397, 2008. 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 312, + 623, + 545, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 623, + 545, + 666 + ], + "spans": [ + { + "bbox": [ + 312, + 623, + 545, + 666 + ], + "type": "text", + "content": "[26] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 1, 2, 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 312, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 312, + 670, + 545, + 712 + ], + "type": "text", + "content": "[27] Xiao-Shan Gao, Xiao-Rong Hou, Jianliang Tang, and Hang-Fei Cheng. Complete solution classification for the perspective-three-point problem. IEEE Trans. Pattern Anal. Machine Intell., 25(8):930–943, 2003. 2, 6" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 750, + 315, + 759 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 750, + 315, + 759 + ], + "spans": [ + { + "bbox": [ + 295, + 750, + 315, + 759 + ], + "type": "text", + "content": "8955" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 73, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 73, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 73, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 53, + 73, + 287, + 106 + ], + "type": "text", + "content": "[28] Myung Geun Kim. Multivariate outliers and decompositions of mahalanobis distance. Communications in statistics-theory and methods, 29(7):1511-1526, 2000. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 107, + 288, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 107, + 288, + 150 + ], + "spans": [ + { + "bbox": [ + 53, + 107, + 288, + 150 + ], + "type": "text", + "content": "[29] Chunhui Gu and Xiaofeng Ren. Discriminative mixture-of-templates for viewpoint classification. In European Conf. on Computer Vision (ECCV), pages 408–421. Springer, 2010. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 152, + 288, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 152, + 288, + 185 + ], + "spans": [ + { + "bbox": [ + 53, + 152, + 288, + 185 + ], + "type": "text", + "content": "[30] Richard Hartley, Jochen Trumpf, Yuchao Dai, and Hong-dong Li. Rotation averaging. International journal of computer vision, 103:Intl. J. of Computer Vision, 2013. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 186, + 288, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 186, + 288, + 218 + ], + "spans": [ + { + "bbox": [ + 53, + 186, + 288, + 218 + ], + "type": "text", + "content": "[31] Richard Hartley and Andrew Zisserman. Multiple view geometry in computer vision. Cambridge university press, 2003. 3, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 220, + 288, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 220, + 288, + 275 + ], + "spans": [ + { + "bbox": [ + 53, + 220, + 288, + 275 + ], + "type": "text", + "content": "[32] Stefan Hinterstoisser, Cedric Cagniart, Slobodan Ilic, Peter Sturm, Nassir Navab, Pascal Fua, and Vincent Lepetit. Gradient response maps for real-time detection of textureless objects. IEEE Trans. Pattern Anal. Machine Intell., 34(5):876-888, 2011. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 276, + 288, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 276, + 288, + 341 + ], + "spans": [ + { + "bbox": [ + 53, + 276, + 288, + 341 + ], + "type": "text", + "content": "[33] Stefan Hinterstoisser, Vincent Lepetit, Slobodan Ilic, Stefan Holzer, Gary Bradski, Kurt Konolige, and Nassir Navab. Model based training, detection and pose estimation of texture-less 3d objects in heavily cluttered scenes. In Asian conference on computer vision, pages 548-562. Springer, 2012. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 342, + 288, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 342, + 288, + 387 + ], + "spans": [ + { + "bbox": [ + 53, + 342, + 288, + 387 + ], + "type": "text", + "content": "[34] Tomas Hodan, Daniel Barath, and Jiri Matas. Epos: Estimating 6d pose of objects with symmetries. In IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 11703-11712, 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 388, + 288, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 388, + 288, + 443 + ], + "spans": [ + { + "bbox": [ + 53, + 388, + 288, + 443 + ], + "type": "text", + "content": "[35] Tomas Hodan, Frank Michel, Eric Brachmann, Wadim Kehl, Anders GlentBuch, Dirk Kraft, Bertram Drost, Joel Vidal, Stephan Ihrke, Xenophon Zabulis, et al. Bop: Benchmark for 6d object pose estimation. In European Conf. on Computer Vision (ECCV), pages 19-34, 2018. 2, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 445, + 288, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 445, + 288, + 487 + ], + "spans": [ + { + "bbox": [ + 53, + 445, + 288, + 487 + ], + "type": "text", + "content": "[36] Daniel P Huttenlocher, Gregory A. Klanderman, and William J Rucklidge. Comparing images using the hausdorff distance. IEEE Trans. Pattern Anal. Machine Intell., 15(9):850-863, 1993. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 489, + 288, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 489, + 288, + 533 + ], + "spans": [ + { + "bbox": [ + 53, + 489, + 288, + 533 + ], + "type": "text", + "content": "[37] Shun Iwase, Xingyu Liu, Rawal Khirodkar, Rio Yokota, and Kris M Kitani. Repose: Fast 6d object pose refinement via deep texture rendering. In Intl. Conf. on Computer Vision (ICCV), pages 3303-3312, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 534, + 288, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 534, + 288, + 599 + ], + "spans": [ + { + "bbox": [ + 53, + 534, + 288, + 599 + ], + "type": "text", + "content": "[38] Yanmei Jiao, Yue Wang, Bo Fu, Qimeng Tan, Lei Chen, Minhang Wang, Shoudong Huang, and Rong Xiong. Globally optimal consensus maximization for robust visual inertial localization in point and line map. In IEEE/RSJ Intl. Conf. on Intelligent Robots and Systems (IROS), pages 4631-4638. IEEE, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 601, + 288, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 601, + 288, + 633 + ], + "spans": [ + { + "bbox": [ + 53, + 601, + 288, + 633 + ], + "type": "text", + "content": "[39] Fredrik Kahl and Richard Hartley. Multiple-view geometry under the 1 infinity norm. IEEE Trans. Pattern Anal. Machine Intell., 30(9):1603-1617, 2008. 2, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 635, + 288, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 635, + 288, + 668 + ], + "spans": [ + { + "bbox": [ + 53, + 635, + 288, + 668 + ], + "type": "text", + "content": "[40] Fredrik Kahl and Didier Henrion. Globally optimal estimates for geometric reconstruction problems. *Intl. J. of Computer Vision*, 74(1):3-15, 2007. 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 670, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 670, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 53, + 670, + 288, + 713 + ], + "type": "text", + "content": "[41] Alex Kendall, Matthew Grimes, and Roberto Cipolla. Posenet: A convolutional network for real-time 6-dof camera relocalization. In Intl. Conf. on Computer Vision (ICCV), pages 2938-2946, 2015. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 312, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 312, + 73, + 545, + 116 + ], + "type": "text", + "content": "[42] Georg Klein and David Murray. Parallel tracking and mapping for small ar workspaces. In 2007 6th IEEE and ACM international symposium on mixed and augmented reality, pages 225-234. IEEE, 2007. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 118, + 545, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 118, + 545, + 160 + ], + "spans": [ + { + "bbox": [ + 312, + 118, + 545, + 160 + ], + "type": "text", + "content": "[43] Adam Klivans, Pravesh K Kothari, and Raghu Meka. Efficient algorithms for outlier-robust regression. In Conference On Learning Theory, pages 1420–1430. PMLR, 2018. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 162, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 162, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 312, + 162, + 545, + 205 + ], + "type": "text", + "content": "[44] Laurent Kneip and Paul Furgale. Opengv: A unified and generalized approach to real-time calibrated geometric vision. In IEEE Intl. Conf. on Robotics and Automation (ICRA), pages 1-8. IEEE, 2014. 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 206, + 545, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 206, + 545, + 249 + ], + "spans": [ + { + "bbox": [ + 312, + 206, + 545, + 249 + ], + "type": "text", + "content": "[45] Laurent Kneip, Hongdong Li, and Yongduek Seo. Upnp: An optimal o(n) solution to the absolute pose problem with universal applicability. In European Conf. on Computer Vision (ECCV), pages 127-142. Springer, 2014. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 250, + 545, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 250, + 545, + 314 + ], + "spans": [ + { + "bbox": [ + 312, + 250, + 545, + 314 + ], + "type": "text", + "content": "[46] Laurent Kneip, Davide Scaramuzza, and Roland Siegwart. A novel parametrization of the perspective-three-point problem for a direct computation of absolute camera position and orientation. In IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 2969-2976. IEEE, 2011. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 316, + 545, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 316, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 312, + 316, + 545, + 360 + ], + "type": "text", + "content": "[47] Yann Labbe, Justin Carpentier, Mathieu Aubry, and Josef Sivic. Cosypos: Consistent multi-view multi-object 6d pose estimation. In European Conf. on Computer Vision (ECCV), pages 574-591. Springer, 2020. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 361, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 361, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 312, + 361, + 545, + 403 + ], + "type": "text", + "content": "[48] Jing Lei, Max G'Sell, Alessandro Rinaldo, Ryan J Tibshirani, and Larry Wasserman. Distribution-free predictive inference for regression. Journal of the American Statistical Association, 113(523):1094-1111, 2018. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 404, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 404, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 312, + 404, + 545, + 437 + ], + "type": "text", + "content": "[49] Jing Lei, James Robins, and Larry Wasserman. Distribution-free prediction sets. Journal of the American Statistical Association, 108(501):278-287, 2013. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 437, + 545, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 437, + 545, + 479 + ], + "spans": [ + { + "bbox": [ + 312, + 437, + 545, + 479 + ], + "type": "text", + "content": "[50] Vincent Lepetit, Pascal Fua, et al. Monocular model-based 3d tracking of rigid objects: A survey. Foundations and Trends in Computer Graphics and Vision, 1(1):1-89, 2005. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 481, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 481, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 312, + 481, + 545, + 514 + ], + "type": "text", + "content": "[51] Vincent Lepetit, Francesc Moreno-Noguer, and Pascal Fua. Epnp: An accurate o(n) solution to the pnp problem. *Intl. J. of Computer Vision*, 81(2):155-166, 2009. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 515, + 545, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 515, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 312, + 515, + 545, + 557 + ], + "type": "text", + "content": "[52] Hongdong Li. Consensus set maximization with guaranteed global optimality for robust geometry estimation. In Intl. Conf. on Computer Vision (ICCV), pages 1074-1080. IEEE, 2009. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 559, + 545, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 559, + 545, + 592 + ], + "spans": [ + { + "bbox": [ + 312, + 559, + 545, + 592 + ], + "type": "text", + "content": "[53] Shuo Li, Sangdon Park, Xiayan Ji, Insup Lee, and Osbert Bastani. Towards pac multi-object detection and tracking. arXiv preprint arXiv:2204.07482, 2022. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 593, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 593, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 312, + 593, + 545, + 635 + ], + "type": "text", + "content": "[54] Zhigang Li, Gu Wang, and Xiangyang Ji. Cdpn: Coordinates-based disentangled pose network for real-time rgb-based 6-dof object pose estimation. In Intl. Conf. on Computer Vision (ICCV), pages 7678-7687, 2019. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 312, + 636, + 545, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 636, + 545, + 680 + ], + "spans": [ + { + "bbox": [ + 312, + 636, + 545, + 680 + ], + "type": "text", + "content": "[55] Joerg Liebelt, Cordelia Schmid, and Klaus Schertler. Independent object class detection using 3d feature maps. In IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 1-8. IEEE, 2008. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 312, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 681, + 545, + 713 + ], + "type": "text", + "content": "[56] David G Lowe. Object recognition from local scale-invariant features. In Intl. Conf. on Computer Vision (ICCV), volume 2, pages 1150-1157. IEEE, 1999. 2" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 750, + 315, + 759 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 750, + 315, + 759 + ], + "spans": [ + { + "bbox": [ + 295, + 750, + 315, + 759 + ], + "type": "text", + "content": "8956" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 73, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 73, + 287, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 73, + 287, + 117 + ], + "spans": [ + { + "bbox": [ + 53, + 73, + 287, + 117 + ], + "type": "text", + "content": "[57] Rachel Luo, Shengjia Zhao, Jonathan Kuck, Boris Ivanovic, Silvio Savarese, Edward Schmerling, and Marco Pavone. Sample-efficient safety assurances using conformal prediction. arXiv preprint arXiv:2109.14082, 2021. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 118, + 287, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 118, + 287, + 151 + ], + "spans": [ + { + "bbox": [ + 53, + 118, + 287, + 151 + ], + "type": "text", + "content": "[58] Yi Ma, Stefano Soatto, Jana Košecka, and Shankar Sastry. An invitation to 3-d vision: from images to geometric models, volume 26. Springer, 2004. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 152, + 287, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 152, + 287, + 184 + ], + "spans": [ + { + "bbox": [ + 53, + 152, + 287, + 184 + ], + "type": "text", + "content": "[59] Prasanta Chandra Mahalanobis. On the generalized distance in statistics. National Institute of Science of India, 1936. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 186, + 288, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 186, + 288, + 229 + ], + "spans": [ + { + "bbox": [ + 53, + 186, + 288, + 229 + ], + "type": "text", + "content": "[60] Lucas Manuelli, Wei Gao, Peter Florence, and Russ Tedrake. kpm: Keypoint affordances for category-level robotic manipulation. In The International Symposium of Robotics Research, pages 132-157. Springer, 2019. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 231, + 288, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 231, + 288, + 274 + ], + "spans": [ + { + "bbox": [ + 53, + 231, + 288, + 274 + ], + "type": "text", + "content": "[61] Maria Cecilia Mazzaro and Mario Sznaier. A set-membership approach to blind identification. In IEEE Conf. on Decision and Control (CDC), volume 5, pages 5176-5181. IEEE, 2004. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 276, + 287, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 276, + 287, + 330 + ], + "spans": [ + { + "bbox": [ + 53, + 276, + 287, + 330 + ], + "type": "text", + "content": "[62] Frank Michel, Alexander Kirillov, Eric Brachmann, Alexander Krull, Stefan Gumhold, Bogdan Savchynskyy, and Carsten Rother. Global hypothesis generation for 6d object pose estimation. In IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 462-471, 2017. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 331, + 287, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 331, + 287, + 374 + ], + "spans": [ + { + "bbox": [ + 53, + 331, + 287, + 374 + ], + "type": "text", + "content": "[63] Mario Milanese and Antonio Vicino. Optimal estimation theory for dynamic systems with set membership uncertainty: An overview. Automatica, 27(6):997-1009, 1991. 2, 3, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 376, + 287, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 376, + 287, + 408 + ], + "spans": [ + { + "bbox": [ + 53, + 376, + 287, + 408 + ], + "type": "text", + "content": "[64] Ankur Moitra. Sum of squares in theoretical computer science. Sum of Squares: Theory and Applications, 77:83, 2020. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 410, + 287, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 410, + 287, + 432 + ], + "spans": [ + { + "bbox": [ + 53, + 410, + 287, + 432 + ], + "type": "text", + "content": "[65] MOSEK ApS. The MOSEK optimization toolbox for MATLAB manual. Version 8.1., 2017. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 434, + 287, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 434, + 287, + 466 + ], + "spans": [ + { + "bbox": [ + 53, + 434, + 287, + 466 + ], + "type": "text", + "content": "[66] Jiawang Nie, Li Wang, and Jane J Ye. Bilevel polynomial programs and semidefinite relaxation methods. SIAM Journal on Optimization, 27(3):1728-1757, 2017. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 468, + 287, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 468, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 53, + 468, + 287, + 510 + ], + "type": "text", + "content": "[67] Markus Oberweger, Mahdi Rad, and Vincent Lepetit. Making deep heatmaps robust to partial occlusions for 3d object pose estimation. In European Conf. on Computer Vision (ECCV), pages 119-134, 2018. 2, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 512, + 287, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 512, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 53, + 512, + 287, + 555 + ], + "type": "text", + "content": "[68] Carl Olsson, Fredrik Kahl, and Magnus Oskarsson. Optimal estimation of perspective camera pose. In 18th International Conference on Pattern Recognition (ICPR'06), volume 2, pages 5-8. IEEE, 2006. 2, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 557, + 287, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 557, + 287, + 589 + ], + "spans": [ + { + "bbox": [ + 53, + 557, + 287, + 589 + ], + "type": "text", + "content": "[69] Harris Papadopoulos. Inductive conformal prediction: Theory and application to neural networks. In Tools in artificial intelligence. CiteSeer, 2008. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 590, + 287, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 590, + 287, + 634 + ], + "spans": [ + { + "bbox": [ + 53, + 590, + 287, + 634 + ], + "type": "text", + "content": "[70] Kiru Park, Timothy Patten, and Markus Vincze. Pix2pose: Pixel-wise coordinate regression of objects for 6d pose estimation. In Intl. Conf. on Computer Vision (ICCV), pages 7668-7677, 2019. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 53, + 635, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 635, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 53, + 635, + 287, + 689 + ], + "type": "text", + "content": "[71] Georgios Pavlakos, Xiaowei Zhou, Aaron Chan, Konstantinos G Derpanis, and Kostas Daniilidis. 6-DoF object pose from semantic keypoints. In IEEE Intl. Conf. on Robotics and Automation (ICRA), pages 2011–2018. IEEE, 2017. 1, 2, 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 53, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 53, + 691, + 287, + 713 + ], + "type": "text", + "content": "[72] Sida Peng, Yuan Liu, Qixing Huang, Xiaowei Zhou, and Hujun Bao. Pvnet: Pixel-wise voting network for 6dof" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 331, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 331, + 73, + 545, + 95 + ], + "type": "text", + "content": "pose estimation. IEEE Trans. Pattern Anal. Machine Intell., 2022. 1, 2, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 96, + 545, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 96, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 312, + 96, + 545, + 171 + ], + "type": "text", + "content": "[73] Luis Pineda, Taosha Fan, Maurizio Monge, Shobha Venkataraman, Paloma Sodhi, Ricky TQ Chen, Joseph Ortiz, Daniel DeTone, Austin Wang, Stuart Anderson, Jing Dong, Brandon Amos, and Mustafa Mukadam. Theseus: A Library for Differentiable Nonlinear Optimization. In Advances in Neural Information Processing Systems (NIPS), 2022. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 175, + 545, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 175, + 545, + 229 + ], + "spans": [ + { + "bbox": [ + 312, + 175, + 545, + 229 + ], + "type": "text", + "content": "[74] Mahdi Rad and Vincent Lepetit. Bb8: A scalable, accurate, robust to partial occlusion method for predicting the 3d poses of challenging objects without using depth. In Proceedings of the IEEE international conference on computer vision, pages 3828-3836, 2017. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 231, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 231, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 312, + 231, + 545, + 274 + ], + "type": "text", + "content": "[75] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in Neural Information Processing Systems (NIPS), 28, 2015. 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 276, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 276, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 312, + 276, + 545, + 319 + ], + "type": "text", + "content": "[76] Yaniv Romano, Matteo Sesia, and Emmanuel Candes. Classification with valid and adaptive coverage. In Advances in Neural Information Processing Systems (NIPS), volume 33, pages 3581-3591, 2020. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 320, + 545, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 320, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 312, + 320, + 545, + 364 + ], + "type": "text", + "content": "[77] David M Rosen, Luca Carlone, Afonso S Bandeira, and John J Leonard. Se-sync: A certifiably correct algorithm for synchronization over the special euclidean group. *Intl. J. of Robotics Research*, 38(2-3):95–125, 2019. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 365, + 545, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 365, + 545, + 419 + ], + "spans": [ + { + "bbox": [ + 312, + 365, + 545, + 419 + ], + "type": "text", + "content": "[78] Fred Rothganger, Svetlana Lazebnik, Cordelia Schmid, and Jean Ponce. 3d object modeling and recognition using local affine-invariant image descriptors and multi-view spatial constraints. *Intl. J. of Computer Vision*, 66(3):231-259, 2006. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 421, + 545, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 421, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 312, + 421, + 545, + 475 + ], + "type": "text", + "content": "[79] Karl Schmeckpeper, Philip R Osteen, Yufu Wang, Georgios Pavlakos, Kenneth Chaney, Wyatt Jordan, Xiaowei Zhou, Konstantinos G Derpanis, and Kostas Daniilidis. Semantic keypoint-based pose estimation from single RGB frames. J. of Field Robotics, 2022. 1, 4, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 477, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 477, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 312, + 477, + 545, + 510 + ], + "type": "text", + "content": "[80] Jingnan Shi, Heng Yang, and Luca Carlone. Optimal pose and shape estimation for category-level 3d object perception. In Robotics: Science and Systems (RSS), 2021. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 512, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 512, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 312, + 512, + 545, + 555 + ], + "type": "text", + "content": "[81] Jingnan Shi, Heng Yang, and Luca Carlone. Optimal and robust category-level perception: Object pose and shape estimation from 2d and 3d semantic keypoints. arXiv preprint arXiv:2206.12498, 2022. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 556, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 556, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 312, + 556, + 545, + 578 + ], + "type": "text", + "content": "[82] Torsten Söderström. Errors-in-variables methods in system identification. Automatica, 43(6):939–958, 2007. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 312, + 580, + 545, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 580, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 312, + 580, + 545, + 612 + ], + "type": "text", + "content": "[83] David Stutz, Ali Taylan Cemgil, Arnaud Doucet, et al. Learning optimal conformal classifiers. In Intl. Conf. on Learning Representations (ICLR), 2022. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 312, + 614, + 545, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 614, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 312, + 614, + 545, + 658 + ], + "type": "text", + "content": "[84] Hao Su, Charles R Qi, Yangyan Li, and Leonidas J Guibas. Render for cnn: Viewpoint estimation in images using cnns trained with rendered 3d model views. In Intl. Conf. on Computer Vision (ICCV), pages 2686-2694, 2015. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 312, + 659, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 659, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 659, + 545, + 713 + ], + "type": "text", + "content": "[85] Jiaming Sun, Zihao Wang, Siyu Zhang, Xingyi He, Hongcheng Zhao, Guofeng Zhang, and Xiaowei Zhou. Onepose: One-shot object pose estimation without cad models. In IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 6825-6834, 2022. 1" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 750, + 315, + 759 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 750, + 315, + 759 + ], + "spans": [ + { + "bbox": [ + 295, + 750, + 315, + 759 + ], + "type": "text", + "content": "8957" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 72, + 287, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 72, + 287, + 117 + ], + "spans": [ + { + "bbox": [ + 53, + 72, + 287, + 117 + ], + "type": "text", + "content": "[86] Min Sun, Gary Bradski, Bing-Xin Xu, and Silvio Savarese. Depth-encoded hough voting for joint object detection and shape recovery. In European Conf. on Computer Vision (ECCV), pages 658-671. Springer, 2010. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 118, + 288, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 118, + 288, + 172 + ], + "spans": [ + { + "bbox": [ + 53, + 118, + 288, + 172 + ], + "type": "text", + "content": "[87] Martin Sundermeyer, Zoltan-Csaba Marton, Maximilian Durner, Manuel Brucker, and Rudolph Triebel. Implicit 3d orientation learning for 6d object detection from rgb images. In European Conf. on Computer Vision (ECCV), pages 699-715, 2018. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 172, + 288, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 172, + 288, + 195 + ], + "spans": [ + { + "bbox": [ + 53, + 172, + 288, + 195 + ], + "type": "text", + "content": "[88] Richard Szeliski. Computer vision: algorithms and applications. Springer Nature, 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 195, + 288, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 195, + 288, + 239 + ], + "spans": [ + { + "bbox": [ + 53, + 195, + 288, + 239 + ], + "type": "text", + "content": "[89] Bugra Tekin, Sudipta N Sinha, and Pascal Fua. Real-time seamless single shot 6d object pose prediction. In IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 292-301, 2018. 1, 2, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 239, + 288, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 239, + 288, + 282 + ], + "spans": [ + { + "bbox": [ + 53, + 239, + 288, + 282 + ], + "type": "text", + "content": "[90] Shubham Tulsiani and Jitendra Malik. Viewpoints and keypoints. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1510-1519, 2015. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 284, + 288, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 284, + 288, + 316 + ], + "spans": [ + { + "bbox": [ + 53, + 284, + 288, + 316 + ], + "type": "text", + "content": "[91] Vladimir Vovk. Conditional validity of inductive conformal predictors. In Asian conference on machine learning, pages 475-490. PMLR, 2012. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 316, + 288, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 316, + 288, + 348 + ], + "spans": [ + { + "bbox": [ + 53, + 316, + 288, + 348 + ], + "type": "text", + "content": "[92] Vladimir Vovk, Alexander Gammerman, and Glenn Shafer. Algorithmic learning in a random world. Springer Science & Business Media, 2005. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 350, + 288, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 350, + 288, + 404 + ], + "spans": [ + { + "bbox": [ + 53, + 350, + 288, + 404 + ], + "type": "text", + "content": "[93] Gu Wang, Fabian Manhardt, Federico Tombari, and Xi-angyang Ji. Gdr-net: Geometry-guided direct regression network for monocular 6d object pose estimation. In IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 16611-16621, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 404, + 288, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 404, + 288, + 470 + ], + "spans": [ + { + "bbox": [ + 53, + 404, + 288, + 470 + ], + "type": "text", + "content": "[94] Håkan Wieslander, Philip J Harrison, Gabriel Skogberg, Sonya Jackson, Markus Fridén, Johan Karlsson, Ola Spjuth, and Carolina Wahlby. Deep learning with conformal prediction for hierarchical analysis of large-scale whole-slide tissue images. IEEE journal of biomedical and health informatics, 25(2):371-380, 2020. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 471, + 288, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 471, + 288, + 514 + ], + "spans": [ + { + "bbox": [ + 53, + 471, + 288, + 514 + ], + "type": "text", + "content": "[95] Yu Xiang, Tanner Schmidt, Venkatraman Narayanan, and Dieter Fox. Posecnn: A convolutional neural network for 6d object pose estimation in cluttered scenes. In Robotics: Science and Systems (RSS), 2018. 2, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 515, + 288, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 515, + 288, + 568 + ], + "spans": [ + { + "bbox": [ + 53, + 515, + 288, + 568 + ], + "type": "text", + "content": "[96] Heng Yang, Pasquale Antonante, Vasileios Tzoumas, and Luca Carlone. Graduated non-convexity for robust spatial perception: From non-minimal solvers to global outlier rejection. IEEE Robotics and Automation Letters, 5(2):1127-1134, 2020. 2, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 570, + 288, + 614 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 570, + 288, + 614 + ], + "spans": [ + { + "bbox": [ + 53, + 570, + 288, + 614 + ], + "type": "text", + "content": "[97] Heng Yang and Luca Carlone. In perfect shape: Certifiably optimal 3d shape reconstruction from 2d landmarks. In IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 621-630, 2020. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 614, + 288, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 614, + 288, + 658 + ], + "spans": [ + { + "bbox": [ + 53, + 614, + 288, + 658 + ], + "type": "text", + "content": "[98] Heng Yang and Luca Carlone. Certifiably optimal outlier-robust geometric perception: Semidefinite relaxations and scalable global optimization. IEEE Trans. Pattern Anal. Machine Intell., 2022. 1, 2, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 658, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 658, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 53, + 658, + 288, + 713 + ], + "type": "text", + "content": "[99] Heng Yang, Ling Liang, Luca Carlone, and Kim-Chuan Toh. An inexact projected gradient method with rounding and lifting by nonlinear programming for solving rank-one semidefinite relaxation of polynomial optimization. Mathematical Programming, pages 1-64, 2022. 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 545, + 185 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 307, + 72, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 545, + 106 + ], + "type": "text", + "content": "[100] Heng Yang, Jingnan Shi, and Luca Carlone. Teaser: Fast and certifiable point cloud registration. IEEE Trans. Robotics, 37(2):314-333, 2020. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 140 + ], + "type": "text", + "content": "[101] Sergey Zakharov, Ivan Shugurov, and Slobodan Ilic. Dpod: 6d pose object detector and refiner. In Intl. Conf. on Computer Vision (ICCV), pages 1941-1950, 2019. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 141, + 545, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 185 + ], + "type": "text", + "content": "[102] Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. In IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 5745-5753, 2019. 2" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 750, + 315, + 759 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 750, + 315, + 759 + ], + "spans": [ + { + "bbox": [ + 295, + 750, + 315, + 759 + ], + "type": "text", + "content": "8958" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Object-Aware Distillation Pyramid for Open-Vocabulary Object Detection/b1741226-bc11-487a-861d-b993de9bcc68_content_list.json b/2023/Object-Aware Distillation Pyramid for Open-Vocabulary Object Detection/b1741226-bc11-487a-861d-b993de9bcc68_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..0eeead57ca0ef3f0b9839ef58d1bed8d818063fe --- /dev/null +++ b/2023/Object-Aware Distillation Pyramid for Open-Vocabulary Object Detection/b1741226-bc11-487a-861d-b993de9bcc68_content_list.json @@ -0,0 +1,2007 @@ +[ + { + "type": "text", + "text": "Object-Aware Distillation Pyramid for Open-Vocabulary Object Detection", + "text_level": 1, + "bbox": [ + 109, + 130, + 857, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Luting Wang $^{1,3}$ Yi Liu $^{1,3}$ Penghui Du $^{1,3}$ Zihan Ding $^{1,3}$ Yue Liao $^{1,3*}$ Qiaosong Qi $^{2}$ Biaolong Chen $^{2}$ Si Liu $^{1,3}$", + "bbox": [ + 192, + 178, + 781, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Institute of Artificial Intelligence, Beihang University $^{2}$ Alibaba Group $^{3}$ Hangzhou Innovation Institute, Beihang University", + "bbox": [ + 197, + 215, + 772, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 286, + 313, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Open-vocabulary object detection aims to provide object detectors trained on a fixed set of object categories with the generalizability to detect objects described by arbitrary text queries. Previous methods adopt knowledge distillation to extract knowledge from Pretrained Vision-and-Language Models (PVLMs) and transfer it to detectors. However, due to the non-adaptive proposal cropping and single-level feature mimicking processes, they suffer from information destruction during knowledge extraction and inefficient knowledge transfer. To remedy these limitations, we propose an Object-Aware Distillation Pyramid (OADP) framework, including an Object-Aware Knowledge Extraction (OAKE) module and a Distillation Pyramid (DP) mechanism. When extracting object knowledge from PVLMs, the former adaptively transforms object proposals and adopts object-aware mask attention to obtain precise and complete knowledge of objects. The latter introduces global and block distillation for more comprehensive knowledge transfer to compensate for the missing relation information in object distillation. Extensive experiments show that our method achieves significant improvement compared to current methods. Especially on the MS-COCO dataset, our OADP framework reaches 35.6 $mAP_{50}^{N}$ , surpassing the current state-of-the-art method by 3.3 $mAP_{50}^{N}$ . Code is released at https://github.com/LutingWang/OADP.", + "bbox": [ + 73, + 316, + 473, + 696 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 78, + 715, + 209, + 732 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Open-vocabulary object detection (OVD) [49] aims to endow object detectors with the generalizability to detect open categories including both base and novel categories where only the former are annotated in the training phase. Pretrained Vision-and-Language Models (PVLMs, e.g., CLIP [32] and ALIGN [19]) have witnessed great progress in recent years, and Knowledge Distillation (KD) [17] has led to a wave of unprecedented advances transferring the zero-shot visual recognition ability from PVLMs to detec", + "bbox": [ + 75, + 741, + 468, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f129ed84a80be33e500770427197ac312505311d4f7f87e5ade25a9f87040dba.jpg", + "image_caption": [ + "Center Crop w/o Transform" + ], + "image_footnote": [], + "bbox": [ + 540, + 300, + 687, + 354 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5228dc72d6ab1ef6293d0a7752bc4b9b9a8a9c1973784a611771e7aa5f9d2750.jpg", + "image_caption": [ + "Center Crop w/ Transform" + ], + "image_footnote": [], + "bbox": [ + 705, + 300, + 777, + 354 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f838741004b1ec6c35d1ec0fe4d9f86148c99278434d41de32a2c71a52c1160b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 781, + 300, + 852, + 354 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d3b767bc3fb8f2db92240c9e23cb954fe9cabdd153cc576188271973516498cd.jpg", + "image_caption": [ + "(a) Knowledge Extraction" + ], + "image_footnote": [], + "bbox": [ + 539, + 375, + 625, + 492 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "", + "bbox": [ + 539, + 398, + 625, + 492 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/01f9825d24dda904ba1654551663d3702772b6a5aa22b52a6fd00883dd5ec49b.jpg", + "image_caption": [ + "Student" + ], + "image_footnote": [], + "bbox": [ + 651, + 391, + 720, + 492 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b6729b616c8ea716e49410ba4b18a42276deb0b2e524fc488b373ac3d0050e6a.jpg", + "image_caption": [ + "(b) Knowledge Transfer" + ], + "image_footnote": [], + "bbox": [ + 705, + 398, + 736, + 492 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2d0d5c68704061f8e03882585b6c9fc7bad957d475c4b07cd6f94678ce9b6587.jpg", + "image_caption": [ + "Object KD" + ], + "image_footnote": [], + "bbox": [ + 797, + 391, + 825, + 410 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4729854dcc7e885a0b7d0dab4693f1a2e757bd8523712119aefc2d77779e248f.jpg", + "image_caption": [ + "Figure 1. An overview of our OADP framework. (a) Directly applying center crop on proposals may throw informative object parts away, resulting in ambiguous image regions. In contrast, our OAKE module extracts complete objects and reduces the influence of surrounding distractors. (b) Our DP mechanism includes global, block, and object KD to achieve effective knowledge transfer." + ], + "image_footnote": [], + "bbox": [ + 764, + 426, + 848, + 492 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tors [12, 25, 28, 29, 48, 53]. KD typically comprises two essential steps, i.e., knowledge extraction and then knowledge transfer. A common practice in OVD is to crop objects with class-agnostic proposals and use the teacher (e.g., CLIP visual encoder) to extract knowledge of the proposals. The knowledge is then transferred to the detector (e.g., Mask R-CNN [15]) via feature mimicking.", + "bbox": [ + 496, + 611, + 893, + 717 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite significant development, we argue that conventional approaches still have two main limitations: 1) Dilemma between comprehensiveness and purity during knowledge extraction. As proposals have diverse aspect ratios, the fixed center crop strategy to square them may cut out object parts (fig. 1a). Enlarging those proposals via resizing function may alleviate this problem, but additional surrounding distractors may confuse the teacher to extract accurate proposal knowledge. 2) Missing global scene understanding during knowledge transfer. Conventional approaches merely concentrate on object-level knowledge transfer by directly mimicking the teacher's features of indi", + "bbox": [ + 496, + 719, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author (liaoyue.ai@gmail.com)", + "bbox": [ + 94, + 886, + 351, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "11186", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "vidual proposals. As a result, the student cannot fully grasp the contextual characteristics describing the interweaving of different objects. In light of the above discussions, we propose an Object-Aware Distillation Pyramid (OADP) framework to excavate the teacher's knowledge accurately and effectively transfer the knowledge to the student.", + "bbox": [ + 75, + 90, + 470, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To preserve the complete information of proposals while extracting their CLIP image embeddings, we propose an Object-Aware Knowledge Extraction (OAKE) module. Concretely, given a proposal, we square it with an adaptive resizing function to avoid destroying the object structure and involve object information as much as possible. However, the resizing process inevitably introduces environmental context, which may contain some distractors that confuse the teacher. Therefore, we propose to utilize an object token [OBJ] whose interaction manner during the forward process is almost the same as the class token [CLS] except that it only attends to patch tokens covered by the original proposal. In this way, the extracted embeddings contain precise and complete knowledge of the proposal object.", + "bbox": [ + 75, + 181, + 468, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To facilitate complete and effective knowledge transfer, we propose a Distillation Pyramid (DP) mechanism (fig. 1b). As previous works only adopt object distillation to align the feature space of detectors and PVLMs, the relation between different objects is neglected. Therefore, we propose global and block distillation to compensate for the missing relation information in object distillation. For global distillation, we optimize the $\\mathcal{L}_1$ distance between the detector backbone and the CLIP visual encoder so that the detector learns to encode rich semantics implied in the image scene. However, the CLIP visual encoder is prone to ignore background information, which may also be valuable for detection. Therefore, we take a finer step to divide the input image into several blocks and optimize the $\\mathcal{L}_1$ distance between the block embeddings of the detector and the CLIP image encoder. Overall, the above three distillation modules constitute a hierarchical distillation pyramid, allowing for the transfer of more diversified knowledge from CLIP to the detectors.", + "bbox": [ + 75, + 393, + 470, + 679 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We demonstrate the superiority of our OADP framework on MS-COCO [27] and LVIS [14] datasets. On MS-COCO, it improves the state-of-the-art results of $\\mathrm{mAP}_{50}^{\\mathrm{N}}$ from 32.3 to 35.6. On the LVIS dataset, our OADP framework reaches $21.9\\mathrm{AP_r}$ on the object detection task and $21.7\\mathrm{AP_r}$ on the instance segmentation task, leading the former methods by more than $1.1\\mathrm{AP_r}$ and $1.9\\mathrm{AP_r}$ respectively.", + "bbox": [ + 75, + 680, + 468, + 787 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 797, + 218, + 813 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Knowledge Distillation for Object Detection. KD [17,37] is a technology that helps train compact student models under the supervision of powerful teacher models. Chen et al. [5] apply KD to object detection by implementing feature-based and response-based loss for Faster R-CNN.", + "bbox": [ + 75, + 824, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Li et al. [23] apply $\\mathcal{L}_2$ loss on features sampled by student proposals. FGFI [40] only distills foreground regions near the object anchors. DeFeat [13] distills the foreground and background regions simultaneously with different factors. GID [7] distills regions where the student and teacher perform differently. G-DetKD [46] proposes a general distillation framework for object detectors. FKD [54] distills the attention map to emphasize the changeable areas. FGD [43] proposes focal and global distillation for comprehensive knowledge transfer. Compared to these detection KD methods [9, 21, 30, 44, 50], our work concentrates on knowledge transfer from PVLMs to detectors to enable open-vocabulary detection.", + "bbox": [ + 496, + 90, + 890, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Open-Vocabulary Detection. OVD [4, 20, 29, 45] aims to train a model that can detect objects of arbitrary categories, even if the categories are not seen during training. OVR-CNN [49] is the seminal work that proposes this problem and achieves great performance using image captions as well as bounding box annotations. With the prevalence of PVLMs [19,32], ViLD [12] proposes to distill the open-vocabulary knowledge from CLIP to the detector. DetPro [8] improves upon ViLD with prompt optimization. RegionCLIP [53] develops a pretraining strategy to learn region-text alignment. Detic [55] adopts weak supervisions to jointly train the detector. GLIP [22] pretrains on massive image-text pairs in a self-training fashion by unifying the detection and grounding tasks. HierKD [28] proposes instance- and global-level distillation for one-stage detectors. OV-DETR [48] turns DETR into an open-vocabulary detector with conditional binary matching. VL-PLM [52] leverages pseudo labels on novel categories to augment the detector. PB-OVD [11] generates pseudo labels based on the image captions. PromptDet [10] establishes a scalable pipeline with regional prompt learning and self-training. In this paper, we propose an OADP framework focusing on comprehensive object knowledge extraction and effective knowledge transfer.", + "bbox": [ + 496, + 295, + 890, + 657 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. OVD Benchmarks", + "text_level": 1, + "bbox": [ + 500, + 671, + 679, + 686 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "According to the training data, we summarize the existing OVD methods into four types of benchmarks: Vanilla OVD (V-OVD), Caption-based OVD (C-OVD), Generalized OVD (G-OVD), and Weakly Supervised OVD (WSOVD). All benchmarks rely on instance-level annotations and large-scale image-text pairs to learn OVD. Some of them use more types of data, as shown in tab. 1. For clarity, we define base categories as those included in the instance-level annotations, and novel categories are the others.", + "bbox": [ + 496, + 696, + 890, + 833 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "V-OVD [4, 8, 12, 20, 22, 29, 45, 53] is a pure OVD benchmark setting, which requires the detector only to train in an object detection dataset with fixed categories set. Any information about the novel categories is unavailable, but unan-", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "11187", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/0e96446e5d04b49c6f352817abfbccfe40aacc005ae1a4c6a58fa2ee2fc7d515.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
BenchmarkCaptionCategory PriorImage Label
V-OVD
C-OVD✓
G-OVD✓
WS-OVD✓✓✓
", + "bbox": [ + 111, + 88, + 436, + 167 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 1. Summary of OVD benchmarks. \"Caption\": in-domain captions like COCO-Captions. \"Category Prior\": human priors on novel categories. \"Image Label\": image-level category labels.", + "bbox": [ + 76, + 176, + 468, + 219 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "notated data is allowed. A common practice for this benchmark is to learn open vocabulary knowledge from imagetext pairs and transfer the knowledge to detectors through transfer learning or knowledge distillation. V-OVD is similar to ZSD [1,33,42,57], except that V-OVD relies on largescale image-text pairs to acquire open-vocabulary knowledge. Recently, V-OVD has attracted more and more researchers with the development of PVLMs.", + "bbox": [ + 75, + 228, + 468, + 349 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "C-OVD [3, 11, 28, 49] adds additional image caption annotation to the V-OVD benchmark. Note that by image caption data, we refer to the in-domain captions of the instance-level annotations, e.g., COCO-Captions [6], instead of the large-scale image-text pairs, e.g., CC3M [39] and CLIP400M [32]. The in-domain captions enrich the instance-level annotations and imply a distribution of potential novel categories. Compared with the V-OVD benchmark, C-OVD requires slightly more annotations and is expected to perform better.", + "bbox": [ + 75, + 354, + 468, + 505 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "G-OVD [10,48,52] introduces human priors on novel categories to the V-OVD benchmark. Intuitively, if some novel categories are far more likely to appear during inference, it would be beneficial to prepare for them during training. Most existing methods assume that all the dataset's category names (including the novels) are known to the detectors during training. Therefore, the performance of G-OVD methods may not be fairly comparable with V-OVD and C-OVD methods. A typical solution is to generate instance-level pseudo annotations for the categories.", + "bbox": [ + 75, + 510, + 468, + 661 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "WS-OVD [55] further takes advantage of image-level category labels beyond G-OVD. Similar to Weakly Supervised Detection (WSD) [2,47], the image-level category labels reflect the presence of the base and novel categories in each image. Thus, the annotation cost is far more than the benchmarks above. In this case, WS-OVD methods have the greatest potential to push the limit of OVD further.", + "bbox": [ + 75, + 666, + 468, + 772 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Object-Aware Distillation Pyramid", + "text_level": 1, + "bbox": [ + 76, + 784, + 393, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We first briefly review the task definition of OVD and the architecture of Faster R-CNN in sec. 4.1. Then, we present the overview of our OADP framework in sec. 4.2. Sec. 4.3 and sec. 4.4 introduce the OAKE module and the DP mechanism in detail. Finally, in sec. 4.5, we demonstrate the procedure to generate pseudo labels based on OAKE.", + "bbox": [ + 75, + 809, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Preliminaries", + "text_level": 1, + "bbox": [ + 500, + 90, + 638, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We represent traditional object detection datasets as $\\mathcal{D} = \\{(\\mathbf{I}_i, \\mathcal{O}_i)\\}_{i=1}^{|\\mathcal{D}|}$ , where $\\mathbf{I}_i$ is the $i^{\\text{th}}$ image and $\\mathcal{O}_i = \\{o_{ij}\\}_{j=1}^{|\\mathcal{O}_i|}$ is the corresponding set of annotated objects. Each object $o$ is a pair of object bounding box $b \\in \\mathbb{R}^4$ and category $y \\in \\mathcal{C}$ where $\\mathcal{C}$ is the category space of the dataset. We denote the training and validation datasets as $\\mathcal{D}^T$ and $\\mathcal{D}^V$ , respectively.", + "bbox": [ + 498, + 113, + 890, + 205 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "By the convention of OVD, we refer to the category space of $\\mathcal{D}^T$ and $\\mathcal{D}^V$ as $\\mathcal{C}^B$ and $\\mathcal{C}$ respectively. Normally, $\\mathcal{C}^B\\subset \\mathcal{C}$ . Categories in $\\mathcal{C}^B$ are called base categories, and those that only appear in $\\mathcal{D}^V$ are called novel categories. The novel category space is denoted as $\\mathcal{C}^N = \\mathcal{C}\\setminus \\mathcal{C}^B\\neq \\emptyset$ . For each category $c\\in \\mathcal{C}$ , we use a pretrained text encoder $\\mathcal{T}$ to encode its semantic embedding $t_c\\in \\mathbb{R}^d$ . Specifically, we use a trainable embedding $t_{\\mathrm{bg}}\\in \\mathbb{R}^{d}$ to represent the background class $\\mathrm{bg}$ .", + "bbox": [ + 498, + 207, + 890, + 340 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Since our work is based on Faster R-CNN [36], we briefly recap its framework. Given an image $\\mathbf{I} \\in \\mathbb{R}^{H \\times W \\times 3}$ , the backbone (including FPN [26]) encodes a set of hierarchical feature maps $\\mathcal{F} = \\{\\mathbf{F}_2, \\mathbf{F}_3, \\dots, \\mathbf{F}_6\\}$ and the Region Proposal Network (RPN) generates a set of proposals $\\mathcal{P} \\subset \\mathbb{R}^4$ . Then the R-CNN head performs RoI Align on $\\mathcal{F}$ to extract proposal embeddings $\\mathcal{E} = \\{e_p\\}_{p \\in \\mathcal{P}} \\subset \\mathbb{R}^d$ . The logit of a proposal $p$ being of category $c$ can be defined as:", + "bbox": [ + 498, + 340, + 890, + 453 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nl (p, c) = \\frac {e _ {p} \\cdot t _ {c}}{\\| e _ {p} \\| \\cdot \\| t _ {c} \\|}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 625, + 460, + 890, + 489 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\cdot$ is the dot product, $t_c$ is the category embedding of $c$ . For simplicity, we ignore the temperature $\\tau$ in CLIP [32]. The probability of a proposal $p$ belonging to category $c \\in \\mathcal{C} \\cup \\{\\mathrm{bg}\\}$ is:", + "bbox": [ + 498, + 496, + 890, + 553 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nP c (p, c) = \\frac {\\exp (l (p , c))}{\\sum_ {c ^ {\\prime} \\in C \\cup \\{b g \\}} \\exp (l (p , c ^ {\\prime}))}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 583, + 561, + 890, + 601 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "During training, each proposal $p$ is assigned a category label $y_{p} \\in \\mathcal{C}^{B} \\cup \\{\\text{bog}\\}$ . The R-CNN loss is defined as:", + "bbox": [ + 498, + 609, + 888, + 640 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = - \\sum_ {p \\in P} \\log P _ {\\mathcal {C} ^ {B}} (p, y _ {p}). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 609, + 648, + 890, + 678 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For simplicity, we ignore the regression term in $\\mathcal{L}$ .", + "bbox": [ + 498, + 686, + 831, + 702 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.2. Overview of OADP", + "text_level": 1, + "bbox": [ + 500, + 710, + 684, + 724 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To inject open-vocabulary concepts into Faster R-CNN, we propose an Object-Aware Distillation Pyramid (OADP) framework (fig. 2), which first extracts knowledge from CLIP [32] and then transfers it to the detector through knowledge distillation (KD) [17]. Specifically, we propose an Object-Aware Knowledge Extraction (OAKE) module, which inserts an [OBJ] token into the frozen CLIP visual encoder $\\mathcal{V}$ to extract informative knowledge from expanded region proposals selectively. For more effective knowledge transfer, we propose a Distillation Pyramid (DP) mechanism comprising an object distillation module $\\mathcal{M}^{\\mathcal{O}}$ , a block distillation module $\\mathcal{M}^{\\mathcal{B}}$ , and a global distillation module", + "bbox": [ + 496, + 732, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "11188", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0c0968e16e217028ce68e52945294b5691c0f4f6114f09380418e759946ad3ef.jpg", + "image_caption": [ + "Figure 2. Illustration of our OADP training pipeline. We adopt a pyramid architecture comprising three distillation modules: global, block, and object. Given an image $\\mathbf{I}$ , RPN generates proposals $\\mathcal{P}$ . For object distillation, RoI Align and Object Head are applied for proposal embeddings $\\mathcal{E}^O$ . To extract complete and pure object knowledge from CLIP, we crop the image regions $\\mathbf{I}^{\\mathcal{O}}$ based on the transformed proposals $\\mathcal{P}'$ and feed them to $L$ layers of masked attention, where an extra [OBJ] token (yellow) attends to the patches covered by the original proposal. For global and block distillation, GAP and block pooling are used before the corresponding heads to extract the global and block embeddings $(\\mathcal{E}^B$ and $e^{G}$ ). The teacher embeddings $\\tilde{\\mathcal{E}}^B$ and $\\tilde{e}^{G}$ are extracted via CLIP from $\\mathbf{I}$ and $\\mathcal{I}^B$ respectively." + ], + "image_footnote": [], + "bbox": [ + 119, + 89, + 857, + 329 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\mathcal{M}^G$ . The losses of the three modules are denoted as $\\mathcal{L}^O$ , $\\mathcal{L}^B$ , and $\\mathcal{L}^G$ , respectively. The total training loss is:", + "bbox": [ + 76, + 435, + 470, + 465 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {\\text {a l l}} = \\mathcal {L} + w ^ {O} \\cdot \\mathcal {L} ^ {O} + w ^ {B} \\cdot \\mathcal {L} ^ {B} + w ^ {G} \\cdot \\mathcal {L} ^ {G}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 470, + 468, + 489 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{L}$ is the R-CNN loss as defined in eq. (3); $w^{O}, w^{B}$ , and $w^{G}$ are loss weights.", + "bbox": [ + 75, + 497, + 468, + 527 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We follow the inference pipeline of ViLD-ensemble [12] and use $\\mathcal{M}^O$ to calibrate $P_{\\mathcal{C}}(p,c)$ . Similar to the R-CNN head, $\\mathcal{M}^O$ extracts the proposal embeddings $\\mathcal{E}^O = \\{e_p^O\\}_{p\\in \\mathcal{P}}\\subset \\mathbb{R}^d$ and computes the logits:", + "bbox": [ + 76, + 527, + 468, + 585 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nl ^ {O} (p, c) = \\frac {e _ {p} ^ {O} \\cdot t _ {c}}{\\left\\| e _ {p} ^ {O} \\right\\| \\cdot \\left\\| t _ {c} \\right\\|}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 595, + 468, + 627 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nP _ {c} ^ {O} (p, c) = \\frac {\\exp (l ^ {O} (p , c))}{\\sum_ {c ^ {\\prime} \\in C} \\exp (l ^ {O} (p , c ^ {\\prime}))}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 628, + 468, + 670 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $t_c$ is the category embedding of $c$ . The calibrated probability $P_{\\mathcal{C}}^{\\mathrm{cal}}(p,c)$ is:", + "bbox": [ + 75, + 678, + 468, + 709 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nP _ {\\mathcal {C}} ^ {\\text {c a l}} (p, c) = \\left\\{ \\begin{array}{l l} \\left(P _ {\\mathcal {C}} (p, c)\\right) ^ {\\lambda} \\cdot \\left(P _ {\\mathcal {C}} ^ {O} (p, c)\\right) ^ {(1 - \\lambda)}, & c \\in \\mathcal {C} ^ {B} \\\\ \\left(P _ {\\mathcal {C}} (p, c)\\right) ^ {(1 - \\lambda)} \\cdot \\left(P _ {\\mathcal {C}} ^ {O} (p, c)\\right) ^ {\\lambda}, & c \\in \\mathcal {C} ^ {N} \\\\ 1 - \\sum_ {c ^ {\\prime} \\in \\mathcal {C}} P _ {\\mathcal {C}} (p, c ^ {\\prime}), & c = b g \\end{array} \\right. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 715, + 468, + 771 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\lambda$ is set to $2/3$ . Note that the block and global distillation modules are not used during the inference phase, so the computation cost of our OADP framework is the same as ViLD-ensemble.", + "bbox": [ + 75, + 779, + 468, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.3. Object Distillation", + "text_level": 1, + "bbox": [ + 76, + 849, + 256, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The object distillation module $\\mathcal{M}^O$ aims to transfer the object-level knowledge from CLIP [32] to the detector. For", + "bbox": [ + 75, + 871, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "each proposal $p\\in \\mathcal{P}$ $\\mathcal{M}^O$ motivates the detector to extract a proposal embedding $e_p^O$ that resembles the corresponding embedding $\\tilde{e}_p^O$ extracted by the CLIP visual encoder $\\nu$ ..", + "bbox": [ + 498, + 435, + 893, + 483 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {O} = \\mathcal {L} _ {1} \\left(\\mathcal {E} ^ {O}, \\tilde {\\mathcal {E}} ^ {O}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 633, + 494, + 892, + 511 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\tilde{\\mathcal{E}}^O = \\{\\tilde{e}_p^O\\}_{p\\in \\mathcal{P}}\\subset \\mathbb{R}^d$ denotes the proposal embeddings extracted by $\\nu$ . Naturally, the quality of $\\tilde{\\mathcal{E}}^O$ affects the accuracy of $P_{\\mathcal{C}}^{O}(p,c)$ to a large extent. However, current approaches only yield sub-optimal $\\tilde{\\mathcal{E}}^O$ due to the dilemma between information comprehensiveness and less noise. For example, when non-square proposal regions are directly passed to CLIP, the center crop operation in $\\nu$ will cut out informative parts of an object, leading to incomplete structural knowledge about the object. On the other hand, if the proposals are squared or enlarged, the proposal regions will contain more ambient contexts, which may corrupt the proposal embeddings.", + "bbox": [ + 496, + 522, + 893, + 708 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To acquire more accurate $\\tilde{\\mathcal{E}}^O$ , we propose an Object-Aware Knowledge Extraction (OAKE) module, where the proposals are first transformed and then encoded with a modified version of $\\mathcal{V}$ . Given a proposal $p \\in \\mathcal{P}$ , the transformed proposal $p'$ is a square with side length $s = \\sqrt{r \\times p_h \\times p_w}$ , where $r$ is a constant scale ratio, $p_h$ and $p_w$ are the height and width of $p$ . The center of $p'$ is initially the same as $p$ but may be translated if $p'$ exceeds the image boundaries. All transformed proposals constitute $\\mathcal{P}'$ .", + "bbox": [ + 496, + 708, + 893, + 844 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While $\\mathcal{P}'$ contributes to the comprehensiveness of object knowledge, it also includes surrounding distractors. To further suppress the contextual noise, we introduce an [OBJ] token into $\\mathcal{V}$ and substitute original attention layers with", + "bbox": [ + 496, + 844, + 893, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "11189", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "masked attention layers. The modified version of $\\mathcal{V}$ is denoted as $\\mathcal{V}'$ . The input of $\\mathcal{V}'$ is a set of image regions $\\mathbf{I}^{\\mathcal{O}} = \\{\\mathbf{I}_{p'}\\}_{p' \\in \\mathcal{P}'}$ , where $p'$ is a transformed proposal and $\\mathbf{I}_{p'}$ is the cropped image region of $p'$ . Same as $\\mathcal{V}$ , each $\\mathbf{I}_{p'}$ is first mapped to a sequence of tokens $\\mathbf{X} \\in \\mathbb{R}^{N_x \\times d_x}$ , where $\\mathbf{X}_{1:N_x - 1}$ are the patch tokens of $\\mathbf{I}_{p'}$ and $\\mathbf{X}_{N_x}$ is the [CLS] token. We then augment $\\mathbf{X}$ with an [OBJ] token:", + "bbox": [ + 76, + 90, + 468, + 191 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {X} ^ {\\prime} = [ \\mathbf {X}; x _ {[ \\mathrm {O B J} ]} ] \\in \\mathbb {R} ^ {(N _ {x} + 1) \\times d _ {x}}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 194, + 468, + 212 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since the [OBJ] token serves a similar purpose as the [CLS] token, we initialize $x_{[\\mathrm{OBJ}]} = \\mathbf{X}_{N_x}$ . To regulate the interaction between [OBJ] and the other tokens, we construct a mask $m \\in \\mathbb{R}^{N_x}$ , such that:", + "bbox": [ + 76, + 217, + 468, + 273 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nm _ {i} = \\left\\{ \\begin{array}{l l} \\mathbb {1} \\left\\{\\text {t h e} i \\text {- t h p a t c h o v e r l a p s w i t h} p \\right\\}, & i < N _ {x} \\\\ 0, & i = N _ {x} \\end{array} \\right. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 277, + 468, + 314 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbb{1}\\{\\cdot\\}$ is the indicator function. Intuitively, eq. (10) means that [OBJ] only attends to the patch tokens that are covered by the original proposal $p$ . To maintain the original attentions among $\\mathbf{X}$ in $\\mathcal{V}$ , our attention mask $\\mathbf{M}$ is constructed as follows:", + "bbox": [ + 76, + 319, + 468, + 388 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {M} = \\left[ \\begin{array}{c c} \\mathbf {1} ^ {N _ {x} \\times N _ {x}} & \\mathbf {0} ^ {N _ {x}} \\\\ m & 1 \\end{array} \\right] \\in \\{0, 1 \\} ^ {(N _ {x} + 1) \\times (N _ {x} + 1)}. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 391, + 468, + 422 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Suppose $\\mathcal{V}$ has $L$ attention layers (ignoring FFNs), $\\mathcal{V}'$ is defined as follows:", + "bbox": [ + 76, + 426, + 468, + 455 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {X} ^ {l} = \\left\\{ \\begin{array}{l l} \\sigma \\left(\\log \\mathbf {M} + \\mathbf {Q} ^ {l} \\left(\\mathbf {K} ^ {l}\\right) ^ {\\top}\\right) \\mathbf {V} ^ {l} + \\mathbf {X} ^ {l - 1}, & 0 < l \\leq L \\\\ \\mathbf {X} ^ {\\prime}, & l = 0 \\end{array} \\right. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 459, + 468, + 508 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{Q}^l$ , $\\mathbf{K}^l$ , and $\\mathbf{V}^l$ are linear transformations of $\\mathbf{X}^{l-1}$ , $\\sigma$ is Softmax function. Finally, instead of [CLS], we take the output of [OBJ] as the proposal embedding, i.e., $\\tilde{e}_p^O = \\mathbf{X}_{N_x+1}^L$ . Iterating over the proposals $\\mathcal{P}$ , we obtain a set of accurate proposal embeddings $\\tilde{\\mathcal{E}}^O$ .", + "bbox": [ + 76, + 510, + 468, + 590 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Global and Block Distillation", + "text_level": 1, + "bbox": [ + 76, + 598, + 336, + 612 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "While the object distillation module $\\mathcal{M}^O$ aligns the proposal embeddings $\\mathcal{E}^O$ with the optimized CLIP embeddings $\\tilde{\\mathcal{E}}^O$ , the detector lacks a comprehensive understanding of the relation between different proposals. Therefore, we propose a global distillation module $\\mathcal{M}^G$ , which transfers the knowledge of the entire image from the CLIP visual encoder $\\mathcal{V}$ to the detector:", + "bbox": [ + 76, + 619, + 468, + 726 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ne ^ {G} = f (\\operatorname {G A P} \\left(\\mathbf {F} _ {6}\\right)) \\in \\mathbb {R} ^ {d},\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 732, + 372, + 750 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {e} ^ {G} = \\mathcal {V} (\\mathbf {I}) \\in \\mathbb {R} ^ {d},\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 751, + 316, + 768 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {G} = \\mathcal {L} _ {1} (e ^ {G}, \\tilde {e} ^ {G}),\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 770, + 321, + 787 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $f(\\cdot)$ is a linear transformation function, $\\mathrm{GAP}(\\cdot)$ is the global average pooling, and $\\mathbf{I}$ is the input image.", + "bbox": [ + 76, + 794, + 468, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Due to the existence of Human Reporting Bias [31], CLIP is prone to ignore non-salient information in an image, e.g., the background or prominent attributes of objects. Such information may be valuable for dense prediction tasks like detection [13]. Thus, we propose a block", + "bbox": [ + 75, + 825, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "distillation module $\\mathcal{M}^B$ to complement the missing knowledge in $\\mathcal{M}^G$ . We evenly divide the input image into several blocks $\\mathcal{B}\\subset \\mathbb{R}^4$ via a partition function $g(\\cdot)$ and denote the corresponding image regions as $\\mathbf{I}^{\\mathcal{B}}$ . The size of each block is fixed to $R\\times R$ , where $R$ denotes the input resolution of $\\mathcal{V}$ . In this way, the resize and center crop operations in $\\mathcal{V}$ will not take effect, thus avoiding information loss when using $\\mathcal{V}$ to encode the block embeddings $\\tilde{\\mathcal{E}}^{B} = \\{\\tilde{e}_{b}^{B}\\}_{b\\in \\mathcal{B}}\\subset \\mathbb{R}^{d}$ .", + "bbox": [ + 496, + 90, + 890, + 212 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "On the student side, we apply block pooling and a block head to extract the block embeddings $\\mathcal{E}^B = \\{e_b^O\\}_{b\\in \\mathcal{B}}\\subset$ $\\mathbb{R}^d$ . The block pooling is a combination of the block partition function $g(\\cdot)$ and RoI Align. The loss of our proposed block distillation is defined as:", + "bbox": [ + 496, + 210, + 890, + 282 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {B} = \\mathcal {L} _ {1} \\left(\\mathcal {E} ^ {B}, \\tilde {\\mathcal {E}} ^ {B}\\right). \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 633, + 292, + 890, + 311 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Compared with $\\mathcal{L}^G$ , $\\mathcal{L}^B$ distills the knowledge of each block with the same weights. Therefore, the ignored information in $\\mathcal{M}^G$ can be compensated by $\\mathcal{M}^B$ . Note that neither $\\mathcal{M}^G$ nor $\\mathcal{M}^B$ is used during inference.", + "bbox": [ + 496, + 323, + 890, + 383 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.5. Pseudo Label Generation", + "text_level": 1, + "bbox": [ + 498, + 396, + 730, + 410 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To investigate the performance of our OADP framework under the G-OVD benchmark (refer to sec. 3), we propose to generate pseudo labels with our modified $\\nu'$ . Given a proposal $p$ , we first extract the proposal embedding $\\tilde{e}_p^O$ as described in sec. 4.3. Then, the probability of $p$ belonging to category $c \\in \\mathcal{C}$ is given as:", + "bbox": [ + 496, + 419, + 890, + 503 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nl ^ {\\mathrm {P L}} (p, c) = \\frac {\\tilde {e} _ {p} ^ {O} \\cdot t _ {c}}{\\| \\tilde {e} _ {p} ^ {O} \\| \\cdot \\| t _ {c} \\|}, \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 583, + 513, + 890, + 547 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nP _ {\\mathcal {C}} ^ {\\mathrm {P L}} (p, c) = \\frac {\\exp \\left(l ^ {\\mathrm {P L}} (p , c)\\right)}{\\sum_ {c ^ {\\prime} \\in C} \\exp \\left(l ^ {\\mathrm {P L}} \\left(p , c ^ {\\prime}\\right)\\right)}, \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 549, + 890, + 589 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $t_c$ is the category embedding of $c$ . Since $P_{\\mathcal{C}}^{\\mathrm{PL}}(p,c)$ does not reflect the localization quality of $p$ [12], we define the confidence score $S_{\\mathcal{C}}(p,c)$ as:", + "bbox": [ + 496, + 602, + 890, + 646 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nS _ {\\mathcal {C}} (p, c) = P _ {\\mathcal {C}} ^ {\\mathrm {P L}} (p, c) ^ {\\gamma} \\cdot o _ {p} ^ {(1 - \\gamma)}, \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 598, + 657, + 890, + 675 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{C}$ includes both base and novel categories, $o_p\\in [0,1]$ is the objectness score of $p$ ,and $\\gamma$ is a constant balancing factor. $S_{\\mathcal{C}}(p,c)$ reflects the probability that $p$ precisely locates an instance of category $c$ . Finally, we apply class-wise NMS on the novel categories to obtain the pseudo labels.", + "bbox": [ + 496, + 686, + 890, + 762 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Note that the Softmax operation in eq. (15) is performed over all categories $\\mathcal{C}$ , even though the pseudo labels do not include instances on base categories. Such a design effectively suppresses false positives in the pseudo labels.", + "bbox": [ + 496, + 763, + 890, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The proposals to be labeled are extracted via an RPN model pretrained on $\\mathcal{D}^T$ , which only contains annotations of the base categories. ViLD [12] demonstrates that the generalization ability of $\\mathcal{P}$ is strong enough to recall most objects of the novel categories.", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "11190", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/6cdb26dc0bf3ce40b15379e5541a549e4b63d7fb7d48ff5c252859b05241e6fb.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
BenchmarkMethod\\( \\mathrm{mAP}_{50}^{N} \\)\\( \\mathrm{mAP}_{50}^{B} \\)\\( \\mathrm{mAP}_{50} \\)
ZSDSB [1]0.329.224.9
DELO [56]3.413.813.0
PL [34]4.135.927.9
V-OVDViLD [12]27.659.551.3
RegionCLIP* [53]14.252.842.7
OADP (Ours)30.053.347.2
C-OVDOVR-CNN [49]22.846.039.9
HierKD [28]20.351.343.2
RegionCLIP [53]26.854.847.5
LocOV [3]28.651.345.7
PB-OVD [11]29.144.440.4
G-OVDOV-DETR [48]29.461.052.7
VL-PLM [52]32.354.048.3
OADP (Ours)35.655.850.5
WSDWSDNN [2]19.719.619.6
Cap2Det [47]20.320.120.1
WS-OVDDetic [55]27.847.145.0
", + "bbox": [ + 81, + 88, + 467, + 366 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/769e17adecf90f81771d59279ee0357b4a0b532be19a41cdbfe85c1781a0a1b5.jpg", + "table_caption": [ + "Table 2. Comparison with other state-of-the-art methods on the OV-COCO dataset. Methods are grouped by the benchmark they use. \"ZSD\" and \"WSD\" stand for Zero-Shot Detection and Weakly Supervised Detection. \"V-OVD\", \"C-OVD\", \"G-OVD\", and \"WS-OVD\" are introduced in sec. 3. \"RegionCLIP*\" indicates a model without refinement using COCO-Captions." + ], + "table_footnote": [], + "table_body": "
MethodObject DetectionInstance Segmentation
APrAPcAPfAPAPrAPcAPfAP
ViLD [12]16.726.534.227.816.624.630.325.5
DetPro [8]20.827.832.428.419.825.628.925.9
OV-DETR [48]----17.425.032.526.6
OADP (Ours)21.928.432.028.721.726.329.026.6
", + "bbox": [ + 81, + 477, + 467, + 558 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Comparison with other state-of-the-art methods on the OV-LVIS dataset.", + "bbox": [ + 75, + 569, + 467, + 595 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 609, + 209, + 627 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we first introduce the detailed experiment setup, including the datasets, evaluation metrics, and implementation details. We then evaluate the performance of our proposed OADP framework and analyze the results compared to the state-of-the-art approaches.", + "bbox": [ + 75, + 636, + 468, + 712 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Datasets", + "text_level": 1, + "bbox": [ + 76, + 724, + 179, + 739 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Experiments are mainly conducted under the open-vocabulary COCO (OV-COCO) setting [49], where the MSCOco 2017 dataset [27] is manually divided into 48 base categories and 17 novel categories. The training dataset contains 107,761 images, and the validation dataset contains 4,836 images. We report the $\\mathrm{mAP}_{50}^{\\mathrm{N}}$ , $\\mathrm{mAP}_{50}^{\\mathrm{B}}$ , and $\\mathrm{mAP}_{50}$ metrics, i.e., the mAP at IoU threshold 0.5 for novel, base, and all categories. $\\mathrm{mAP}_{50}^{\\mathrm{N}}$ is the main metric. Some experiments are conducted under the open-vocabulary LVIS (OV-LVIS) [12] setting, where the 337 rare categories in", + "bbox": [ + 75, + 750, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "LVIS [14] are treated as novel categories, and the other 866 are base categories. Metrics for the OV-LVIS setting are $\\mathrm{AP_r}$ , $\\mathrm{AP_c}$ , $\\mathrm{AP_f}$ , and AP, i.e., the mAP for rare (novel), common, frequent, and all categories. Both object detection and instance segmentation metrics are reported.", + "bbox": [ + 496, + 90, + 890, + 167 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Implementation Details", + "text_level": 1, + "bbox": [ + 498, + 176, + 718, + 193 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training is conducted on 8 V-100 GPUs with batch size 16 in total. We use stochastic gradient descent (SGD) optimizer with 0.02 initial learning rate, 0.9 momentum, and $2.5 \\times 10^{-5}$ weight decay. The student backbone is ResNet50 [16]. Following DetPro [8], we adopt the ViT-B/32 CLIP [32] as the teacher and initialize the student backbone using SoCo [41]. The loss weights $w^{O}$ , $w^{B}$ , and $w^{G}$ are set to 0.5, 0.25, and 0.25 respectively. Under the OV-COCO setting, we train the detector for 40,000 iterations. At the $32,000^{\\mathrm{th}}$ iteration, the learning rate is divided by 10. For OV-LVIS, we use 2x (24 epochs) training schedule, where the learning rate is divided by 10 at the $16^{\\mathrm{th}}$ and $22^{\\mathrm{th}}$ epochs.", + "bbox": [ + 496, + 200, + 890, + 397 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Main Results", + "text_level": 1, + "bbox": [ + 498, + 407, + 638, + 422 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare our OADP framework with the other state-of-the-art OVD methods. As described in sec. 3, we categorize existing OVD methods by the benchmark they belong to. For completeness, we include two related benchmarks: Zero-Shot Detection (ZSD) [1, 18, 24, 35, 51] and Weakly Supervised Detection (WSD) [2, 47].", + "bbox": [ + 496, + 431, + 890, + 521 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our method mainly focuses on the V-OVD benchmark and the G-OVD benchmark. As shown in tab. 2, our OADP framework achieves $30.0\\mathrm{mAP}_{50}^{\\mathrm{N}}$ on the V-OVD benchmark. RegionCLIP* [53] uses CLIP [32] as the pretrained weight, thus adhering to the V-OVD benchmark. Some V-OVD methods [4, 20, 22, 29, 45] are not included because they rely on large-scale detection and image-text datasets and cannot be compared fairly. While C-OVD is not our primary concern, we include the corresponding methods for reference and report their performance when only COCO Captions [6] is available. Under such constraint, the performance of the caption-based methods is relatively lower than the V-OVD methods, even if additional caption data is used. Moreover, our OADP framework is perpendicular to the caption-based methods and has the potential to achieve higher performance using captions.", + "bbox": [ + 496, + 522, + 890, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For the G-OVD benchmark, we generate pseudo labels for novel categories as described in sec. 4.5. The pseudo labels are then merged with the instance-level annotations for base categories. Training our OADP framework on the mixed dataset yields $35.6\\mathrm{mAP}_{50}^{\\mathrm{N}}$ , surpassing the previous SOTA method VL-PLM [52] by $3.3\\mathrm{mAP}_{50}^{\\mathrm{N}}$ . Since PromptDet [10] relies on an external dataset (LAION-400M [38]) and uses smaller image size $(640\\times 640)$ , the result 26.6 $\\mathrm{mAP}_{50}^{\\mathrm{N}}$ is not listed in tab. 2 for fairness.", + "bbox": [ + 496, + 765, + 890, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "11191", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/f8fc86724805de1789226abfd70db7f1b1c6cebe9e0b8ba32107f890b9e19b38.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Global DistillationBlock DistillationObject DistillationmAPNovel \\( \\mathrm{mAP}_{50} \\)\\( \\mathrm{mAP}_{75} \\)mAPBase \\( \\mathrm{mAP}_{50} \\)\\( \\mathrm{mAP}_{75} \\)mAPAll \\( \\mathrm{mAP}_{50} \\)\\( \\mathrm{mAP}_{75} \\)
✓13.3224.9912.3531.8750.2934.0327.0243.6728.36
13.5125.7212.3632.8251.8935.3127.7745.0429.31
✓14.5727.2513.1734.4553.6037.2029.2546.7131.06
✓15.4927.2315.2535.9955.9638.5730.6348.4532.47
✓✓13.5026.4912.5032.1951.2533.9427.3044.7828.33
✓✓15.4728.8014.6234.0854.2936.2829.2147.6230.61
✓✓15.9229.0115.6435.3055.4537.8830.2348.5332.06
✓✓✓16.2129.9515.4733.3353.2635.4728.8547.1730.24
", + "bbox": [ + 120, + 88, + 851, + 241 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tab. 3 shows the comparison between our method and the other state-of-the-art methods on the OV-LVIS dataset. Most of the methods under the OV-LVIS setting adhere to the V-OVD benchmark, so we conduct experiments on the V-OVD benchmark only. For the object detection task, our OADP framework achieves $21.9\\mathrm{AP_r}$ , surpassing DetPro [8] by $1.1\\mathrm{AP_r}$ . We also report the performance of the instance segmentation task, which achieves $21.7\\mathrm{AP_r}$ and is $1.9\\mathrm{AP_r}$ higher than the previous SOTA method.", + "bbox": [ + 75, + 290, + 472, + 426 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 436, + 230, + 454 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct ablation studies on the OV-COCO dataset to evaluate the effectiveness of each component in our proposed OADP framework.", + "bbox": [ + 75, + 460, + 470, + 507 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "OADP. Tab. 4 shows the effectiveness of the distillation modules in our OADP framework. The first row is our re-implemented ViLD-ensemble [12]. Due to the expensive training cost of ViLD, the performance $24.99\\mathrm{mAP}_{50}^{\\mathrm{N}}$ is far below the official $27.60\\mathrm{mAP}_{50}^{\\mathrm{N}}$ . Nevertheless, with our proposed distillation pyramid, we can eventually surpass ViLD. The $2^{\\mathrm{nd}}$ to $4^{\\mathrm{th}}$ row in tab. 4 adds $\\mathcal{M}^{G}$ , $\\mathcal{M}^{B}$ , and $\\mathcal{M}^{O}$ to the baseline respectively. The global distillation module brings a $0.73\\mathrm{mAP}_{50}^{\\mathrm{N}}$ gain, while the other two bring $2.26\\mathrm{mAP}_{50}^{\\mathrm{N}}$ and $2.24\\mathrm{mAP}_{50}^{\\mathrm{N}}$ gain. Note that by adding $\\mathcal{M}^{O}$ to the baseline, we remove the original image head in ViLD-ensemble. Therefore, the $2.24\\mathrm{mAP}_{50}^{\\mathrm{N}}$ gain results from the OAKE module instead of the distillation operation. The $5^{\\mathrm{th}}$ row in tab. 4 adds $\\mathcal{M}^{G}$ and $\\mathcal{M}^{B}$ together. While the 26.49 performance is higher than sole $\\mathcal{M}^{G}$ , it is slightly lower than the $27.25\\mathrm{mAP}_{50}^{\\mathrm{N}}$ of $\\mathcal{M}^{B}$ . However, along with the object distillation module $\\mathcal{M}^{O}$ , $\\mathcal{M}^{G}$ , and $\\mathcal{M}^{B}$ achieves $28.80\\mathrm{mAP}_{50}^{\\mathrm{N}}$ and $29.01\\mathrm{mAP}_{50}^{\\mathrm{N}}$ , suggesting that $\\mathcal{M}^{G}$ and $\\mathcal{M}^{B}$ have a similar function in transferring the global scene knowledge from CLIP to the detector. Using all three modules together, we achieve $29.95\\mathrm{mAP}_{50}^{\\mathrm{N}}$ .", + "bbox": [ + 75, + 513, + 472, + 833 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "OAKE. We demonstrate the effectiveness of our OAKE module in tab. 5. Given the ground truth bounding boxes, we use different strategies to crop their image regions. (1) Baseline: $1 \\times$ crop; (2) ViLD*: $1 \\times$ and $1.5 \\times$ crop; (3)", + "bbox": [ + 75, + 839, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b573387a6afe54729b22c19d599f979ad0adb590654393d0af27c60cdd3451bb.jpg", + "table_caption": [ + "Table 4. Ablation study of the Global, Block, and Object Distillation modules in the OADP framework. The baseline is our re-implemented ViLD-ensemble model." + ], + "table_footnote": [], + "table_body": "
MethodMacro PrecisionWeighted Precision
w/o OAKEw/ OAKEw/o OAKEw/ OAKE
Baseline58.08-62.04-
ViLD*63.36-65.91-
MBS61.7063.8364.8165.82
Fixed49.0764.5351.4969.75
Adaptive51.6466.0955.8568.68
", + "bbox": [ + 503, + 287, + 890, + 400 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. Ablation study of OAKE module. \"ViLD*\" indicates our re-implementation of multi-scale region embedding. \"MBS\", \"Fixed\", and \"Adaptive\" are three transforming strategies.", + "bbox": [ + 498, + 409, + 892, + 452 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "MBS: the minimum bounding square of the original bounding box; (4) Fixed: $224 \\times 224$ bounding square; (5) Adaptive: adaptively enlarge the bounding square. For the above strategies, we use CLIP to directly extract embeddings for their image regions (\"w/o mask\"). Alternatively, we can utilize the modified CLIP visual encoder mentioned in sec. 4.3 (\"w/ mask\"). Finally, we classify these embeddings by calculating their similarity with category embeddings. To evaluate the performance, we compute \"Macro Precision\" (precision for each category independently with equal weights) and \"Weighted Precision\" (weights depending on the number of bounding boxes in each class).", + "bbox": [ + 496, + 465, + 890, + 648 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in the $1^{\\mathrm{st}}$ and $3^{\\mathrm{rd}}$ columns, \"Fixed\" and \"Adaptive\" strategies bring performance drops as they crop a larger bounding square compared to other strategies (e.g., \"MBS\") which may introduce additional surrounding distractors that confuse the CLIP visual encoder. However, with our object-aware CLIP visual encoder, the performances of the two strategies boost significantly and suppress others. It validates that the CLIP visual encoder can focus on the proposal object with our mask attention mechanism to extract accurate knowledge.", + "bbox": [ + 496, + 648, + 893, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Pseudo Label. We follow VL-PLM [52] to adopt the COCO-ZS setting for our ablation studies of the pseudo label. Both mAP and the average per-image number of PLs (#PL) on novel categories are used as metrics to evaluate the quality of the pseudo label. The baseline method directly uses [CLS] token of CLIP visual encoder to extract", + "bbox": [ + 496, + 809, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "11192", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/35b4f78da566dc64c599b1e10fc62a60da20cc0abb4fab7f1e7f26bc86167a66.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 116, + 88, + 276, + 170 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/79c362e78c9985073a685a082e3656578966290e80990a17d3df561a9895a488.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 277, + 88, + 437, + 170 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0690b25cafa5b8cecca7acb3ec5121def590defb004a492ed002e3ca50413d75.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 439, + 88, + 581, + 170 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/38e1a2e284f4805d2f556213ea90b11f59603d4a72110a57094ae10488f750a0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 586, + 88, + 728, + 170 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/765ceed1ccebeaa508ec89ca131d7923f7719bd59c6afff9afd697df17041a56.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 732, + 88, + 888, + 170 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6f5222ffa005bfa522e551d5c1be9dde1d18c991d05aff23b4d9ac8b0ceeda7b.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 116, + 172, + 276, + 252 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/260fbe17168e458edb23266bc33efa504d7b678d10111bb28346accd3a2b477e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 277, + 172, + 437, + 252 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d4189346658dbe9318f68e1471279fd1e1012d4859fc3eb03d86a6a6e69741d7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 439, + 172, + 581, + 252 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/45f3a3f6d7c052ae5c2781a452073556a900ee1a9f9720c0c90bdaaf380e9c82.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 586, + 172, + 728, + 252 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6b2667397cb062187c8f4e4d9794483346b94a6b5791220c1efb4b0697422d10.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 732, + 172, + 890, + 252 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7298ee1d89fd8e1d9b9d9d7ccfc9a7cc29767e9e0b085ee89e663f7239edc5b6.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 116, + 253, + 276, + 335 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b56b68e9d9a4d1641dc2c628af8c5c270e47eb3fe63f72b28ffc1fb87269c6bb.jpg", + "image_caption": [ + "Figure 3. Visualization of activation patterns from different detectors. (a) Pseudo labels (green) and ground truth annotations (blue) for each image. (b) The baseline detector. (c) Our OADP. The intensity of the feature response increases from blue to red." + ], + "image_footnote": [], + "bbox": [ + 277, + 253, + 437, + 335 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1be3077e0e899a9331ef3d6219013c9ea13a985c9ab3c0b37904bf28ef8c0be2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 439, + 253, + 581, + 335 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c9b9b262e5c4cd7dc393afb090bf446b9fe68bc9425360d68b22196d1a76a806.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 586, + 253, + 728, + 335 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/be5132276b97f020e47b2952e407f49eb84748935e1f616d19141c4b4b52aa1e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 732, + 253, + 890, + 335 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f20c9e2d253e914e6315c46d485f0914d5486e679c483193336b20e6fecf105a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Method\\( \\mathbf{{mAP}}^{\\mathbf{{PL}}} \\)\\( \\mathbf{{mAP}}_{50}^{\\mathbf{{PL}}} \\)\\( \\mathbf{{mAP}}_{75}^{\\mathbf{{PL}}} \\)#PL
Baseline5.012.13.1100
ViLD*17.828.118.6100
Ours19.029.919.9100
Baseline3.99.32.56.53
VL-PLM [52]-25.3-4.26
Ours17.426.518.64.14
", + "bbox": [ + 89, + 383, + 457, + 513 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Ablation study of pseudo labels. \"#PL\" is the number of pseudo labels per image.", + "bbox": [ + 75, + 523, + 470, + 551 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "proposal embeddings from the original proposal, and it relies merely on the classification score to sort proposals. The poor detection accuracy in $1^{\\text{st}}$ and $4^{\\text{th}}$ rows of tab. 6 show that without the objectness score, the baseline method can not accurately localize objects. We re-implement multiscale region embedding of ViLD [12] with a geometric mean of CLIP classification score and objectiveness score, i.e., \"ViLD*\". We adopt an adaptive transform strategy for proposals and regard the output of [OBJ] as proposal embedding. The score fusion strategy is described in sec. 4.5, where $\\gamma$ is 0.3. Our method achieves the highest mAP when the number of PLs is sufficient. VL-PLM [52] adopts a multi-scale region embedding method similar to ViLD [12] except for an arithmetic mean of classification score and objections score. When the pseudo labels are filtered with a higher confidence threshold, our method still has a significant advantage compared to VL-PLM [52] (26.50 mAPPL compared to 25.30 mAPPL).", + "bbox": [ + 75, + 564, + 468, + 835 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.5. Visualization", + "text_level": 1, + "bbox": [ + 76, + 845, + 212, + 861 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fig. 3 (a) visualizes our generated PLs on novel categories in green with ground truth boxes of base categories in", + "bbox": [ + 75, + 869, + 468, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "blue. We try our best to ensure the accuracy of PLs as much as possible to be fewer and more precise. These green PLs demonstrate that our proposal embeddings can clearly distinguish novel objects from base ones. Correspondingly, we also show activation maps from baseline (b) and our detector (c) in fig. 3. Taking the $2^{\\mathrm{nd}}$ column as an example, the activation map of our detector accurately highlights more area of novel objects, i.e. \"cup\", with our distillation pyramid mechanism. Therefore, the backbone of OADP generates more informative feature maps, which further help detect novel objects.", + "bbox": [ + 496, + 387, + 892, + 551 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 570, + 617, + 585 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we reconsider the way of knowledge extraction and knowledge transfer in existing KD-based OVD methods and propose an Object-Aware Distillation Pyramid (OADP) framework. To preserve complete and purified object representation in proposals during knowledge extraction, we propose an Object-Aware Knowledge Extraction (OAKE) module to adaptively transform proposals and extract precise object knowledge. A Distillation Pyramid (DP) mechanism is proposed to transfer contextual knowledge about the relation of different objects for better scene understanding. Experiments show that our OADP outperforms previous methods on two popular OVD benchmarks.", + "bbox": [ + 496, + 597, + 890, + 779 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. This research is supported by the National Key R&D Program of China (2022ZD0115502), National Natural Science Foundation of China (NO. 62122010), Zhejiang Provincial Natural Science Foundation of China under Grant No. LDT23F02022F02, Key Research and Development Program of Zhejiang Province under Grant 2022C01082.", + "bbox": [ + 496, + 795, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "11193", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Ankan Bansal, Karan Sikka, Gaurav Sharma, Rama Chellappa, and Ajay Divakaran. Zero-Shot Object Detection. In ECCV, 2018. 3, 6", + "[2] Hakan Bilen and Andrea Vedaldi. Weakly Supervised Deep Detection Networks. In CVPR, 2016. 3, 6", + "[3] Maria A. Bravo, Sudhanshu Mittal, and Thomas Brox. Localized Vision-Language Matching for Open-vocabulary Object Detection. In GCRP, 2022. 3, 6", + "[4] Zhaowei Cai, Gukyeong Kwon, Avinash Ravichandran, Erhan Bas, Zhuowen Tu, Rahul Bhotika, and Stefano Soatto. X-DETR: A Versatile Architecture for Instance-wise Vision-Language Tasks. In ECCV, 2022. 2, 6", + "[5] Guobin Chen, Wongun Choi, Xiang Yu, Tony Han, and Manmohan Chandraker. Learning efficient object detection models with knowledge distillation. In NeurIPS, 2017. 2", + "[6] Xinlei Chen, Hao Fang, Tsung-yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dollar, and C Lawrence Zitnick. Microsoft COCO Captions: Data Collection and Evaluation Server. arXiv preprint arXiv:1504.00325, 2015. 3, 6", + "[7] Xing Dai, Zeren Jiang, Zhao Wu, Yiping Bao, Zhicheng Wang, Si Liu, and Erjin Zhou. General Instance Distillation for Object Detection. In CVPR, 2021. 2", + "[8] Yu Du, Fangyun Wei, Zihe Zhang, Miaojing Shi, Yue Gao, and Guoqi Li. Learning to Prompt for Open-Vocabulary Object Detection with Vision-Language Model. In CVPR, 2022. 2, 6, 7", + "[9] Zhixing Du, Rui Zhang, Ming Chang, Xishan Zhang, Shaoli Liu, Tianshi Chen, and Yunji Chen. Distilling Object Detectors with Feature Richness. In NeurIPS, 2021. 2", + "[10] Chengjian Feng, Yujie Zhong, Zequn Jie, Xiangxiang Chu, Haibing Ren, Xiaolin Wei, Weidi Xie, and Lin Ma. PromptDet: Expand Your Detector Vocabulary with Uncurated Images. In ECCV, 2022. 2, 3, 6", + "[11] Mingfei Gao, Chen Xing, Juan Carlos Niebles, Junnan Li, Ran Xu, Wenhao Liu, and Caiming Xiong. Open Vocabulary Object Detection with Pseudo Bounding-Box Labels. In ECCV, 2022. 2, 3, 6", + "[12] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary Object Detection via Vision and Language Knowledge Distillation. In ICLR, 2022. 1, 2, 4, 5, 6, 7, 8", + "[13] Jianyuan Guo, Kai Han, Yunhe Wang, Han Wu, Xing-hao Chen, Chunjing Xu, and Chang Xu. Distilling" + ], + "bbox": [ + 78, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Object Detectors via Decoupled Features. In CVPR, 2021. 2, 5", + "[14] Agrim Gupta, Piotr Dollar, and Ross Girshick. LVIS: A Dataset for Large Vocabulary Instance Segmentation. In CVPR, 2019. 2, 6", + "[15] Kaiming He, Georgia Gkioxari, Piotr Dollar, and Ross Girshick. Mask R-CNN. TPAMI, 42(2), 2020. 1", + "[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In CVPR, 2016. 6", + "[17] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the Knowledge in a Neural Network. In NeurIPS, 2014. 1, 2, 3", + "[18] Peiliang Huang, Junwei Han, De Cheng, and Dingwen Zhang. Robust Region Feature Synthesizer for Zero-Shot Object Detection. In CVPR, 2022. 6", + "[19] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yunhsuan Sung, Zhen Li, and Tom Duerig. Scaling Up Visual and Vision-Language Representation Learning With Noisy Text Supervision. In ICML, 2021. 1, 2", + "[20] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. MDETR - Modulated Detection for End-to-End Multi-Modal Understanding. In ICCV, 2021. 2, 6", + "[21] Zijian Kang, Peizhen Zhang, Xiangyu Zhang, Jian Sun, and Nanning Zheng. Instance-Conditional Knowledge Distillation for Object Detection. In NeurIPS, 2021. 2", + "[22] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, KaiWei Chang, and Jianfeng Gao. Grounded Language-Image Pre-training. In CVPR, 2022. 2, 6", + "[23] Quanquan Li, Shengying Jin, and Junjie Yan. Mimicking Very Efficient Network for Object Detection. In CVPR, 2017. 2", + "[24] Zhihui Li, Lina Yao, Xiaqin Zhang, Xianzhi Wang, Salil Kanhere, and Huaxiang Zhang. Zero-Shot Object Detection with Textual Descriptions. AAAI, 33(1), 2019. 6", + "[25] Yue Liao, Aixi Zhang, Miao Lu, Yongliang Wang, Xiaobo Li, and Si Liu. Gen-vlkt: Simplify association and enhance interaction understanding for hoi detection. In CVPR, pages 20091-20100, 2022. 1", + "[26] Tsung-Yi Lin, Piotr Dollar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature Pyramid Networks for Object Detection. In CVPR, 2017. 3" + ], + "bbox": [ + 501, + 90, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "11194", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Tsung-Yi Lin, Michael Maire, Serge Belongie, Lubomir Bourdev, Ross Girshick, James Hays, Pietro Perona, Deva Ramanan, C. Lawrence Zitnick, and Piotr Dóllár. Microsoft COCO: Common Objects in Context. In ECCV, 2014. 2, 6", + "[28] Zongyang Ma, Guan Luo, Jin Gao, Liang Li, Yuxin Chen, Shaoru Wang, Congxuan Zhang, and Weiming Hu. Open-Vocabulary One-Stage Detection with Hierarchical Visual-Language Knowledge Distillation. In CVPR, 2022. 1, 2, 3, 6", + "[29] Matthias Minderer, Alexey Gritsenko, Austin Stone, Maxim Neumann, Dirk Weissenborn, Alexey Dosovitskiy, Aravindh Mahendran, Anurag Arnab, Mostafa Dehghani, Zhuoran Shen, Xiao Wang, Xiaohua Zhai, Thomas Kipf, and Neil Houlsby. Simple Open-Vocabulary Object Detection with Vision Transformers. In ECCV, 2022. 1, 2, 6", + "[30] Seyed Iman Mirzadeh, Mehrdad Farajtabar, Ang Li, Nir Levine, Akihiro Matsukawa, and Hassan Ghasemzadeh. Improved Knowledge Distillation via Teacher Assistant. AAAI, 34(04), 2020. 2", + "[31] Ishan Misra, C. Lawrence Zitnick, Margaret Mitchell, and Ross Girshick. Seeing through the Human Reporting Bias: Visual Classifiers from Noisy Human-Centric Labels. In CVPR, 2016. 5", + "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning Transferable Visual Models From Natural Language Supervision. In ICML, volume 139, 2021. 1, 2, 3, 4, 6", + "[33] Shafin Rahman, Salman Khan, and Nick Barnes. Transductive Learning for Zero-Shot Object Detection. In ICCV, 2019. 3", + "[34] Shafin Rahman, Salman Khan, and Nick Barnes. Improved Visual-Semantic Alignment for Zero-Shot Object Detection. AAAI, 34(07), 2020. 6", + "[35] Shafin Rahman, Salman H. Khan, and Fatih Porikli. Zero-Shot Object Detection: Joint Recognition and Localization of Novel Concepts. IJCV, 128(12), 2020. 6", + "[36] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks. TPAMI, 39(6), 2017. 3", + "[37] Adriana Romero, Nicolas Ballas, Samira Ebrahimi Kahou, Antoine Chassang, Carlo Gatta, and Yoshua Bengio. FitNets: Hints for Thin Deep Nets. In ICLR, 2015. 2", + "[38] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. LAION-400M: Open Dataset of CLIP-Filtered 400 Million Image-Text Pairs. arXiv preprint arXiv:2111.02114, 2021.6", + "[39] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual Captions: A Cleaned, Hypernymed, Image Alt-text Dataset For Automatic Image Captioning. In ACL, 2018. 3", + "[40] Tao Wang, Li Yuan, Xiaopeng Zhang, and Jiashi Feng. Distilling Object Detectors With Fine-Grained Feature Imitation. In CVPR, 2019. 2", + "[41] Fangyun Wei, Yue Gao, Zhirong Wu, Han Hu, and Stephen Lin. Aligning Pretraining for Detection via Object-Level Contrastive Learning. In NeurIPS, volume 27, 2021. 6", + "[42] Caixia Yan, Xiaojun Chang, Minnan Luo, Huan Liu, Xiaoqin Zhang, and Qinghua Zheng. Semantics-Guided Contrastive Network for Zero-Shot Object detection. TPAMI, 2022. 3", + "[43] Zhendong Yang, Zhe Li, Xiaohu Jiang, Yuan Gong, Zehuan Yuan, Danpei Zhao, and Chun Yuan. Focal and Global Knowledge Distillation for Detectors. In CVPR, 2022. 2", + "[44] Zhendong Yang, Zhe Li, Mingqi Shao, Dachuan Shi, Zehuan Yuan, and Chun Yuan. Masked Generative Distillation. In ECCV, 2022. 2", + "[45] Lewei Yao, Jianhua Han, Youpeng Wen, Xiaodan Liang, Dan Xu, Wei Zhang, Zhenguo Li, Chunjing Xu, and Hang Xu. DetCLIP: Dictionary-Enriched Visual-Concept Paralleled Pre-training for Open-world Detection. In NeurIPS, 2022. 2, 6", + "[46] Lewei Yao, Renjie Pi, Hang Xu, Wei Zhang, Zhenguo Li, and Tong Zhang. G-DetKD: Towards General Distillation Framework for Object Detectors via Contrastive and Semantic-guided Feature Imitation. In ICCV, 2021. 2", + "[47] Keren Ye, Mingda Zhang, Adriana Kovashka, Wei Li, Danfeng Qin, and Jesse Berent. Cap2Det: Learning to Amplify Weak Caption Supervision for Object Detection. In ICCV, 2019. 3, 6", + "[48] Yuhang Zang, Wei Li, Kaiyang Zhou, Chen Huang, and Chen Change Loy. Open-Vocabulary DETR with Conditional Matching. ECCV, 2022. 1, 2, 3, 6", + "[49] Alireza Zareian, Kevin Dela Rosa, Derek Hao Hu, and Shih-Fu Chang. Open-Vocabulary Object Detection Using Captions. In CVPR, 2021. 1, 2, 3, 6", + "[50] Peizhen Zhang, Zijian Kang, Tong Yang, Xiangyu Zhang, Nanning Zheng, and Jian Sun. LGD: Label-Guided Self-Distillation for Object Detection. AAAI, 36(3), 2022. 2" + ], + "bbox": [ + 503, + 90, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "11195", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[51] Shizhen Zhao, Changxin Gao, Yuanjie Shao, Lerenhan Li, Changqian Yu, Zhong Ji, and Nong Sang. GT-Net: Generative Transfer Network for Zero-Shot Object Detection. AAAI, 34(07), 2020. 6", + "[52] Shiyu Zhao, Zhixing Zhang, Samuel Schulter, Long Zhao, Vijay Kumar B. G, Anastasis Stathopoulos, Manmohan Chandraker, and Dimitris Metaxas. Exploiting Unlabeled Data with Vision and Language Models for Object Detection. In ECCV, 2022. 2, 3, 6, 7, 8", + "[53] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Liunian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, and Jianfeng Gao. RegionCLIP: Region-based Language-Image Pretraining. In CVPR, 2022. 1, 2, 6", + "[54] Chunting Zhou, Graham Neubig, and Jiatao Gu. Improve Object Detection with Feature-based Knowledge Distillation: Towards Accurate and Efficient Detectors. In ICLR, 2021. 2", + "[55] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Phillip Krahenbuhl, and Ishan Misra. Detecting Twenty-thousand Classes using Image-level Supervision. In ECCV, 2022. 2, 3, 6", + "[56] Pengkai Zhu, Hanxiao Wang, and Venkatesh Saligrama. Don't Even Look Once: Synthesizing Features for Zero-Shot Detection. In CVPR, 2020. 6", + "[57] Pengkai Zhu, Hanxiao Wang, and Venkatesh Saligrama. Zero Shot Detection. TCSVT, 30(4), 2020. 3" + ], + "bbox": [ + 78, + 90, + 470, + 549 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11196", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Object-Aware Distillation Pyramid for Open-Vocabulary Object Detection/b1741226-bc11-487a-861d-b993de9bcc68_model.json b/2023/Object-Aware Distillation Pyramid for Open-Vocabulary Object Detection/b1741226-bc11-487a-861d-b993de9bcc68_model.json new file mode 100644 index 0000000000000000000000000000000000000000..76db41ca0f38229f6471117a2031ee7d0487784b --- /dev/null +++ b/2023/Object-Aware Distillation Pyramid for Open-Vocabulary Object Detection/b1741226-bc11-487a-861d-b993de9bcc68_model.json @@ -0,0 +1,2631 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.111, + 0.131, + 0.859, + 0.154 + ], + "angle": 0, + "content": "Object-Aware Distillation Pyramid for Open-Vocabulary Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.193, + 0.179, + 0.782, + 0.217 + ], + "angle": 0, + "content": "Luting Wang\\(^{1,3}\\) Yi Liu\\(^{1,3}\\) Penghui Du\\(^{1,3}\\) Zihan Ding\\(^{1,3}\\) Yue Liao\\(^{1,3*}\\) Qiaosong Qi\\(^{2}\\) Biaolong Chen\\(^{2}\\) Si Liu\\(^{1,3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.199, + 0.217, + 0.773, + 0.253 + ], + "angle": 0, + "content": "\\(^{1}\\)Institute of Artificial Intelligence, Beihang University \\(^{2}\\)Alibaba Group \\(^{3}\\)Hangzhou Innovation Institute, Beihang University" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.287, + 0.314, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.318, + 0.474, + 0.697 + ], + "angle": 0, + "content": "Open-vocabulary object detection aims to provide object detectors trained on a fixed set of object categories with the generalizability to detect objects described by arbitrary text queries. Previous methods adopt knowledge distillation to extract knowledge from Pretrained Vision-and-Language Models (PVLMs) and transfer it to detectors. However, due to the non-adaptive proposal cropping and single-level feature mimicking processes, they suffer from information destruction during knowledge extraction and inefficient knowledge transfer. To remedy these limitations, we propose an Object-Aware Distillation Pyramid (OADP) framework, including an Object-Aware Knowledge Extraction (OAKE) module and a Distillation Pyramid (DP) mechanism. When extracting object knowledge from PVLMs, the former adaptively transforms object proposals and adopts object-aware mask attention to obtain precise and complete knowledge of objects. The latter introduces global and block distillation for more comprehensive knowledge transfer to compensate for the missing relation information in object distillation. Extensive experiments show that our method achieves significant improvement compared to current methods. Especially on the MS-COCO dataset, our OADP framework reaches 35.6 \\( mAP_{50}^{N} \\), surpassing the current state-of-the-art method by 3.3 \\( mAP_{50}^{N} \\). Code is released at https://github.com/LutingWang/OADP." + }, + { + "type": "title", + "bbox": [ + 0.079, + 0.717, + 0.21, + 0.733 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.742, + 0.47, + 0.879 + ], + "angle": 0, + "content": "Open-vocabulary object detection (OVD) [49] aims to endow object detectors with the generalizability to detect open categories including both base and novel categories where only the former are annotated in the training phase. Pretrained Vision-and-Language Models (PVLMs, e.g., CLIP [32] and ALIGN [19]) have witnessed great progress in recent years, and Knowledge Distillation (KD) [17] has led to a wave of unprecedented advances transferring the zero-shot visual recognition ability from PVLMs to detec" + }, + { + "type": "image_caption", + "bbox": [ + 0.541, + 0.288, + 0.681, + 0.3 + ], + "angle": 0, + "content": "Center Crop w/o Transform" + }, + { + "type": "image", + "bbox": [ + 0.541, + 0.301, + 0.688, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.715, + 0.289, + 0.844, + 0.3 + ], + "angle": 0, + "content": "Center Crop w/ Transform" + }, + { + "type": "image", + "bbox": [ + 0.707, + 0.301, + 0.779, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.782, + 0.301, + 0.853, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.628, + 0.362, + 0.766, + 0.374 + ], + "angle": 0, + "content": "(a) Knowledge Extraction" + }, + { + "type": "image", + "bbox": [ + 0.54, + 0.376, + 0.626, + 0.493 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.651, + 0.379, + 0.714, + 0.389 + ], + "angle": 0, + "content": "Student" + }, + { + "type": "image", + "bbox": [ + 0.54, + 0.4, + 0.626, + 0.493 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.652, + 0.392, + 0.721, + 0.493 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.707, + 0.4, + 0.738, + 0.493 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.799, + 0.392, + 0.826, + 0.411 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.779, + 0.413, + 0.835, + 0.422 + ], + "angle": 0, + "content": "Object KD" + }, + { + "type": "image", + "bbox": [ + 0.765, + 0.427, + 0.849, + 0.493 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.631, + 0.496, + 0.758, + 0.508 + ], + "angle": 0, + "content": "(b) Knowledge Transfer" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.513, + 0.895, + 0.597 + ], + "angle": 0, + "content": "Figure 1. An overview of our OADP framework. (a) Directly applying center crop on proposals may throw informative object parts away, resulting in ambiguous image regions. In contrast, our OAKE module extracts complete objects and reduces the influence of surrounding distractors. (b) Our DP mechanism includes global, block, and object KD to achieve effective knowledge transfer." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.612, + 0.894, + 0.718 + ], + "angle": 0, + "content": "tors [12, 25, 28, 29, 48, 53]. KD typically comprises two essential steps, i.e., knowledge extraction and then knowledge transfer. A common practice in OVD is to crop objects with class-agnostic proposals and use the teacher (e.g., CLIP visual encoder) to extract knowledge of the proposals. The knowledge is then transferred to the detector (e.g., Mask R-CNN [15]) via feature mimicking." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Despite significant development, we argue that conventional approaches still have two main limitations: 1) Dilemma between comprehensiveness and purity during knowledge extraction. As proposals have diverse aspect ratios, the fixed center crop strategy to square them may cut out object parts (fig. 1a). Enlarging those proposals via resizing function may alleviate this problem, but additional surrounding distractors may confuse the teacher to extract accurate proposal knowledge. 2) Missing global scene understanding during knowledge transfer. Conventional approaches merely concentrate on object-level knowledge transfer by directly mimicking the teacher's features of indi" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.352, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author (liaoyue.ai@gmail.com)" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11186" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.182 + ], + "angle": 0, + "content": "vidual proposals. As a result, the student cannot fully grasp the contextual characteristics describing the interweaving of different objects. In light of the above discussions, we propose an Object-Aware Distillation Pyramid (OADP) framework to excavate the teacher's knowledge accurately and effectively transfer the knowledge to the student." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.183, + 0.47, + 0.394 + ], + "angle": 0, + "content": "To preserve the complete information of proposals while extracting their CLIP image embeddings, we propose an Object-Aware Knowledge Extraction (OAKE) module. Concretely, given a proposal, we square it with an adaptive resizing function to avoid destroying the object structure and involve object information as much as possible. However, the resizing process inevitably introduces environmental context, which may contain some distractors that confuse the teacher. Therefore, we propose to utilize an object token [OBJ] whose interaction manner during the forward process is almost the same as the class token [CLS] except that it only attends to patch tokens covered by the original proposal. In this way, the extracted embeddings contain precise and complete knowledge of the proposal object." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.394, + 0.471, + 0.68 + ], + "angle": 0, + "content": "To facilitate complete and effective knowledge transfer, we propose a Distillation Pyramid (DP) mechanism (fig. 1b). As previous works only adopt object distillation to align the feature space of detectors and PVLMs, the relation between different objects is neglected. Therefore, we propose global and block distillation to compensate for the missing relation information in object distillation. For global distillation, we optimize the \\(\\mathcal{L}_1\\) distance between the detector backbone and the CLIP visual encoder so that the detector learns to encode rich semantics implied in the image scene. However, the CLIP visual encoder is prone to ignore background information, which may also be valuable for detection. Therefore, we take a finer step to divide the input image into several blocks and optimize the \\(\\mathcal{L}_1\\) distance between the block embeddings of the detector and the CLIP image encoder. Overall, the above three distillation modules constitute a hierarchical distillation pyramid, allowing for the transfer of more diversified knowledge from CLIP to the detectors." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.681, + 0.47, + 0.788 + ], + "angle": 0, + "content": "We demonstrate the superiority of our OADP framework on MS-COCO [27] and LVIS [14] datasets. On MS-COCO, it improves the state-of-the-art results of \\(\\mathrm{mAP}_{50}^{\\mathrm{N}}\\) from 32.3 to 35.6. On the LVIS dataset, our OADP framework reaches \\(21.9\\mathrm{AP_r}\\) on the object detection task and \\(21.7\\mathrm{AP_r}\\) on the instance segmentation task, leading the former methods by more than \\(1.1\\mathrm{AP_r}\\) and \\(1.9\\mathrm{AP_r}\\) respectively." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.798, + 0.22, + 0.814 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Knowledge Distillation for Object Detection. KD [17,37] is a technology that helps train compact student models under the supervision of powerful teacher models. Chen et al. [5] apply KD to object detection by implementing feature-based and response-based loss for Faster R-CNN." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.289 + ], + "angle": 0, + "content": "Li et al. [23] apply \\(\\mathcal{L}_2\\) loss on features sampled by student proposals. FGFI [40] only distills foreground regions near the object anchors. DeFeat [13] distills the foreground and background regions simultaneously with different factors. GID [7] distills regions where the student and teacher perform differently. G-DetKD [46] proposes a general distillation framework for object detectors. FKD [54] distills the attention map to emphasize the changeable areas. FGD [43] proposes focal and global distillation for comprehensive knowledge transfer. Compared to these detection KD methods [9, 21, 30, 44, 50], our work concentrates on knowledge transfer from PVLMs to detectors to enable open-vocabulary detection." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.296, + 0.892, + 0.658 + ], + "angle": 0, + "content": "Open-Vocabulary Detection. OVD [4, 20, 29, 45] aims to train a model that can detect objects of arbitrary categories, even if the categories are not seen during training. OVR-CNN [49] is the seminal work that proposes this problem and achieves great performance using image captions as well as bounding box annotations. With the prevalence of PVLMs [19,32], ViLD [12] proposes to distill the open-vocabulary knowledge from CLIP to the detector. DetPro [8] improves upon ViLD with prompt optimization. RegionCLIP [53] develops a pretraining strategy to learn region-text alignment. Detic [55] adopts weak supervisions to jointly train the detector. GLIP [22] pretrains on massive image-text pairs in a self-training fashion by unifying the detection and grounding tasks. HierKD [28] proposes instance- and global-level distillation for one-stage detectors. OV-DETR [48] turns DETR into an open-vocabulary detector with conditional binary matching. VL-PLM [52] leverages pseudo labels on novel categories to augment the detector. PB-OVD [11] generates pseudo labels based on the image captions. PromptDet [10] establishes a scalable pipeline with regional prompt learning and self-training. In this paper, we propose an OADP framework focusing on comprehensive object knowledge extraction and effective knowledge transfer." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.672, + 0.68, + 0.687 + ], + "angle": 0, + "content": "3. OVD Benchmarks" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.697, + 0.892, + 0.834 + ], + "angle": 0, + "content": "According to the training data, we summarize the existing OVD methods into four types of benchmarks: Vanilla OVD (V-OVD), Caption-based OVD (C-OVD), Generalized OVD (G-OVD), and Weakly Supervised OVD (WSOVD). All benchmarks rely on instance-level annotations and large-scale image-text pairs to learn OVD. Some of them use more types of data, as shown in tab. 1. For clarity, we define base categories as those included in the instance-level annotations, and novel categories are the others." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "V-OVD [4, 8, 12, 20, 22, 29, 45, 53] is a pure OVD benchmark setting, which requires the detector only to train in an object detection dataset with fixed categories set. Any information about the novel categories is unavailable, but unan-" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "11187" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.112, + 0.089, + 0.437, + 0.168 + ], + "angle": 0, + "content": "
BenchmarkCaptionCategory PriorImage Label
V-OVD
C-OVD✓
G-OVD✓
WS-OVD✓✓✓
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.177, + 0.47, + 0.22 + ], + "angle": 0, + "content": "Table 1. Summary of OVD benchmarks. \"Caption\": in-domain captions like COCO-Captions. \"Category Prior\": human priors on novel categories. \"Image Label\": image-level category labels." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.229, + 0.469, + 0.35 + ], + "angle": 0, + "content": "notated data is allowed. A common practice for this benchmark is to learn open vocabulary knowledge from imagetext pairs and transfer the knowledge to detectors through transfer learning or knowledge distillation. V-OVD is similar to ZSD [1,33,42,57], except that V-OVD relies on largescale image-text pairs to acquire open-vocabulary knowledge. Recently, V-OVD has attracted more and more researchers with the development of PVLMs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.355, + 0.469, + 0.506 + ], + "angle": 0, + "content": "C-OVD [3, 11, 28, 49] adds additional image caption annotation to the V-OVD benchmark. Note that by image caption data, we refer to the in-domain captions of the instance-level annotations, e.g., COCO-Captions [6], instead of the large-scale image-text pairs, e.g., CC3M [39] and CLIP400M [32]. The in-domain captions enrich the instance-level annotations and imply a distribution of potential novel categories. Compared with the V-OVD benchmark, C-OVD requires slightly more annotations and is expected to perform better." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.511, + 0.469, + 0.662 + ], + "angle": 0, + "content": "G-OVD [10,48,52] introduces human priors on novel categories to the V-OVD benchmark. Intuitively, if some novel categories are far more likely to appear during inference, it would be beneficial to prepare for them during training. Most existing methods assume that all the dataset's category names (including the novels) are known to the detectors during training. Therefore, the performance of G-OVD methods may not be fairly comparable with V-OVD and C-OVD methods. A typical solution is to generate instance-level pseudo annotations for the categories." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.667, + 0.469, + 0.773 + ], + "angle": 0, + "content": "WS-OVD [55] further takes advantage of image-level category labels beyond G-OVD. Similar to Weakly Supervised Detection (WSD) [2,47], the image-level category labels reflect the presence of the base and novel categories in each image. Thus, the annotation cost is far more than the benchmarks above. In this case, WS-OVD methods have the greatest potential to push the limit of OVD further." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.785, + 0.394, + 0.803 + ], + "angle": 0, + "content": "4. Object-Aware Distillation Pyramid" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.469, + 0.901 + ], + "angle": 0, + "content": "We first briefly review the task definition of OVD and the architecture of Faster R-CNN in sec. 4.1. Then, we present the overview of our OADP framework in sec. 4.2. Sec. 4.3 and sec. 4.4 introduce the OAKE module and the DP mechanism in detail. Finally, in sec. 4.5, we demonstrate the procedure to generate pseudo labels based on OAKE." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.64, + 0.105 + ], + "angle": 0, + "content": "4.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.114, + 0.892, + 0.207 + ], + "angle": 0, + "content": "We represent traditional object detection datasets as \\(\\mathcal{D} = \\{(\\mathbf{I}_i, \\mathcal{O}_i)\\}_{i=1}^{|\\mathcal{D}|}\\), where \\(\\mathbf{I}_i\\) is the \\(i^{\\text{th}}\\) image and \\(\\mathcal{O}_i = \\{o_{ij}\\}_{j=1}^{|\\mathcal{O}_i|}\\) is the corresponding set of annotated objects. Each object \\(o\\) is a pair of object bounding box \\(b \\in \\mathbb{R}^4\\) and category \\(y \\in \\mathcal{C}\\) where \\(\\mathcal{C}\\) is the category space of the dataset. We denote the training and validation datasets as \\(\\mathcal{D}^T\\) and \\(\\mathcal{D}^V\\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.208, + 0.892, + 0.342 + ], + "angle": 0, + "content": "By the convention of OVD, we refer to the category space of \\(\\mathcal{D}^T\\) and \\(\\mathcal{D}^V\\) as \\(\\mathcal{C}^B\\) and \\(\\mathcal{C}\\) respectively. Normally, \\(\\mathcal{C}^B\\subset \\mathcal{C}\\). Categories in \\(\\mathcal{C}^B\\) are called base categories, and those that only appear in \\(\\mathcal{D}^V\\) are called novel categories. The novel category space is denoted as \\(\\mathcal{C}^N = \\mathcal{C}\\setminus \\mathcal{C}^B\\neq \\emptyset\\). For each category \\(c\\in \\mathcal{C}\\), we use a pretrained text encoder \\(\\mathcal{T}\\) to encode its semantic embedding \\(t_c\\in \\mathbb{R}^d\\). Specifically, we use a trainable embedding \\(t_{\\mathrm{bg}}\\in \\mathbb{R}^{d}\\) to represent the background class \\(\\mathrm{bg}\\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.342, + 0.892, + 0.454 + ], + "angle": 0, + "content": "Since our work is based on Faster R-CNN [36], we briefly recap its framework. Given an image \\(\\mathbf{I} \\in \\mathbb{R}^{H \\times W \\times 3}\\), the backbone (including FPN [26]) encodes a set of hierarchical feature maps \\(\\mathcal{F} = \\{\\mathbf{F}_2, \\mathbf{F}_3, \\dots, \\mathbf{F}_6\\}\\) and the Region Proposal Network (RPN) generates a set of proposals \\(\\mathcal{P} \\subset \\mathbb{R}^4\\). Then the R-CNN head performs RoI Align on \\(\\mathcal{F}\\) to extract proposal embeddings \\(\\mathcal{E} = \\{e_p\\}_{p \\in \\mathcal{P}} \\subset \\mathbb{R}^d\\). The logit of a proposal \\(p\\) being of category \\(c\\) can be defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.627, + 0.462, + 0.892, + 0.491 + ], + "angle": 0, + "content": "\\[\nl (p, c) = \\frac {e _ {p} \\cdot t _ {c}}{\\| e _ {p} \\| \\cdot \\| t _ {c} \\|}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.497, + 0.892, + 0.554 + ], + "angle": 0, + "content": "where \\(\\cdot\\) is the dot product, \\(t_c\\) is the category embedding of \\(c\\). For simplicity, we ignore the temperature \\(\\tau\\) in CLIP [32]. The probability of a proposal \\(p\\) belonging to category \\(c \\in \\mathcal{C} \\cup \\{\\mathrm{bg}\\}\\) is:" + }, + { + "type": "equation", + "bbox": [ + 0.584, + 0.562, + 0.892, + 0.602 + ], + "angle": 0, + "content": "\\[\nP c (p, c) = \\frac {\\exp (l (p , c))}{\\sum_ {c ^ {\\prime} \\in C \\cup \\{b g \\}} \\exp (l (p , c ^ {\\prime}))}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.611, + 0.89, + 0.641 + ], + "angle": 0, + "content": "During training, each proposal \\( p \\) is assigned a category label \\( y_{p} \\in \\mathcal{C}^{B} \\cup \\{\\text{bog}\\} \\). The R-CNN loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.611, + 0.649, + 0.892, + 0.679 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = - \\sum_ {p \\in P} \\log P _ {\\mathcal {C} ^ {B}} (p, y _ {p}). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.687, + 0.833, + 0.703 + ], + "angle": 0, + "content": "For simplicity, we ignore the regression term in \\(\\mathcal{L}\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.712, + 0.685, + 0.726 + ], + "angle": 0, + "content": "4.2. Overview of OADP" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.733, + 0.892, + 0.901 + ], + "angle": 0, + "content": "To inject open-vocabulary concepts into Faster R-CNN, we propose an Object-Aware Distillation Pyramid (OADP) framework (fig. 2), which first extracts knowledge from CLIP [32] and then transfers it to the detector through knowledge distillation (KD) [17]. Specifically, we propose an Object-Aware Knowledge Extraction (OAKE) module, which inserts an [OBJ] token into the frozen CLIP visual encoder \\(\\mathcal{V}\\) to extract informative knowledge from expanded region proposals selectively. For more effective knowledge transfer, we propose a Distillation Pyramid (DP) mechanism comprising an object distillation module \\(\\mathcal{M}^{\\mathcal{O}}\\), a block distillation module \\(\\mathcal{M}^{\\mathcal{B}}\\), and a global distillation module" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11188" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.12, + 0.09, + 0.859, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.341, + 0.896, + 0.427 + ], + "angle": 0, + "content": "Figure 2. Illustration of our OADP training pipeline. We adopt a pyramid architecture comprising three distillation modules: global, block, and object. Given an image \\(\\mathbf{I}\\), RPN generates proposals \\(\\mathcal{P}\\). For object distillation, RoI Align and Object Head are applied for proposal embeddings \\(\\mathcal{E}^O\\). To extract complete and pure object knowledge from CLIP, we crop the image regions \\(\\mathbf{I}^{\\mathcal{O}}\\) based on the transformed proposals \\(\\mathcal{P}'\\) and feed them to \\(L\\) layers of masked attention, where an extra [OBJ] token (yellow) attends to the patches covered by the original proposal. For global and block distillation, GAP and block pooling are used before the corresponding heads to extract the global and block embeddings \\((\\mathcal{E}^B\\) and \\(e^{G}\\)). The teacher embeddings \\(\\tilde{\\mathcal{E}}^B\\) and \\(\\tilde{e}^{G}\\) are extracted via CLIP from \\(\\mathbf{I}\\) and \\(\\mathcal{I}^B\\) respectively." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.436, + 0.471, + 0.467 + ], + "angle": 0, + "content": "\\(\\mathcal{M}^G\\). The losses of the three modules are denoted as \\(\\mathcal{L}^O\\), \\(\\mathcal{L}^B\\), and \\(\\mathcal{L}^G\\), respectively. The total training loss is:" + }, + { + "type": "equation", + "bbox": [ + 0.136, + 0.472, + 0.47, + 0.49 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {\\text {a l l}} = \\mathcal {L} + w ^ {O} \\cdot \\mathcal {L} ^ {O} + w ^ {B} \\cdot \\mathcal {L} ^ {B} + w ^ {G} \\cdot \\mathcal {L} ^ {G}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.498, + 0.47, + 0.529 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}\\) is the R-CNN loss as defined in eq. (3); \\(w^{O}, w^{B}\\), and \\(w^{G}\\) are loss weights." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.529, + 0.47, + 0.587 + ], + "angle": 0, + "content": "We follow the inference pipeline of ViLD-ensemble [12] and use \\(\\mathcal{M}^O\\) to calibrate \\(P_{\\mathcal{C}}(p,c)\\). Similar to the R-CNN head, \\(\\mathcal{M}^O\\) extracts the proposal embeddings \\(\\mathcal{E}^O = \\{e_p^O\\}_{p\\in \\mathcal{P}}\\subset \\mathbb{R}^d\\) and computes the logits:" + }, + { + "type": "equation", + "bbox": [ + 0.163, + 0.596, + 0.47, + 0.628 + ], + "angle": 0, + "content": "\\[\nl ^ {O} (p, c) = \\frac {e _ {p} ^ {O} \\cdot t _ {c}}{\\left\\| e _ {p} ^ {O} \\right\\| \\cdot \\left\\| t _ {c} \\right\\|}, \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.158, + 0.63, + 0.47, + 0.671 + ], + "angle": 0, + "content": "\\[\nP _ {c} ^ {O} (p, c) = \\frac {\\exp (l ^ {O} (p , c))}{\\sum_ {c ^ {\\prime} \\in C} \\exp (l ^ {O} (p , c ^ {\\prime}))}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.679, + 0.47, + 0.71 + ], + "angle": 0, + "content": "where \\(t_c\\) is the category embedding of \\(c\\). The calibrated probability \\(P_{\\mathcal{C}}^{\\mathrm{cal}}(p,c)\\) is:" + }, + { + "type": "equation", + "bbox": [ + 0.092, + 0.717, + 0.47, + 0.772 + ], + "angle": 0, + "content": "\\[\nP _ {\\mathcal {C}} ^ {\\text {c a l}} (p, c) = \\left\\{ \\begin{array}{l l} \\left(P _ {\\mathcal {C}} (p, c)\\right) ^ {\\lambda} \\cdot \\left(P _ {\\mathcal {C}} ^ {O} (p, c)\\right) ^ {(1 - \\lambda)}, & c \\in \\mathcal {C} ^ {B} \\\\ \\left(P _ {\\mathcal {C}} (p, c)\\right) ^ {(1 - \\lambda)} \\cdot \\left(P _ {\\mathcal {C}} ^ {O} (p, c)\\right) ^ {\\lambda}, & c \\in \\mathcal {C} ^ {N} \\\\ 1 - \\sum_ {c ^ {\\prime} \\in \\mathcal {C}} P _ {\\mathcal {C}} (p, c ^ {\\prime}), & c = b g \\end{array} \\right. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.47, + 0.84 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is set to \\(2/3\\). Note that the block and global distillation modules are not used during the inference phase, so the computation cost of our OADP framework is the same as ViLD-ensemble." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.85, + 0.258, + 0.866 + ], + "angle": 0, + "content": "4.3. Object Distillation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.872, + 0.471, + 0.902 + ], + "angle": 0, + "content": "The object distillation module \\(\\mathcal{M}^O\\) aims to transfer the object-level knowledge from CLIP [32] to the detector. For" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.436, + 0.894, + 0.484 + ], + "angle": 0, + "content": "each proposal \\(p\\in \\mathcal{P}\\) \\(\\mathcal{M}^O\\) motivates the detector to extract a proposal embedding \\(e_p^O\\) that resembles the corresponding embedding \\(\\tilde{e}_p^O\\) extracted by the CLIP visual encoder \\(\\nu\\) .." + }, + { + "type": "equation", + "bbox": [ + 0.634, + 0.495, + 0.893, + 0.512 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {O} = \\mathcal {L} _ {1} \\left(\\mathcal {E} ^ {O}, \\tilde {\\mathcal {E}} ^ {O}\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.523, + 0.895, + 0.709 + ], + "angle": 0, + "content": "where \\(\\tilde{\\mathcal{E}}^O = \\{\\tilde{e}_p^O\\}_{p\\in \\mathcal{P}}\\subset \\mathbb{R}^d\\) denotes the proposal embeddings extracted by \\(\\nu\\). Naturally, the quality of \\(\\tilde{\\mathcal{E}}^O\\) affects the accuracy of \\(P_{\\mathcal{C}}^{O}(p,c)\\) to a large extent. However, current approaches only yield sub-optimal \\(\\tilde{\\mathcal{E}}^O\\) due to the dilemma between information comprehensiveness and less noise. For example, when non-square proposal regions are directly passed to CLIP, the center crop operation in \\(\\nu\\) will cut out informative parts of an object, leading to incomplete structural knowledge about the object. On the other hand, if the proposals are squared or enlarged, the proposal regions will contain more ambient contexts, which may corrupt the proposal embeddings." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.709, + 0.895, + 0.845 + ], + "angle": 0, + "content": "To acquire more accurate \\(\\tilde{\\mathcal{E}}^O\\), we propose an Object-Aware Knowledge Extraction (OAKE) module, where the proposals are first transformed and then encoded with a modified version of \\(\\mathcal{V}\\). Given a proposal \\(p \\in \\mathcal{P}\\), the transformed proposal \\(p'\\) is a square with side length \\(s = \\sqrt{r \\times p_h \\times p_w}\\), where \\(r\\) is a constant scale ratio, \\(p_h\\) and \\(p_w\\) are the height and width of \\(p\\). The center of \\(p'\\) is initially the same as \\(p\\) but may be translated if \\(p'\\) exceeds the image boundaries. All transformed proposals constitute \\(\\mathcal{P}'\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.845, + 0.895, + 0.901 + ], + "angle": 0, + "content": "While \\(\\mathcal{P}'\\) contributes to the comprehensiveness of object knowledge, it also includes surrounding distractors. To further suppress the contextual noise, we introduce an [OBJ] token into \\(\\mathcal{V}\\) and substitute original attention layers with" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11189" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.192 + ], + "angle": 0, + "content": "masked attention layers. The modified version of \\(\\mathcal{V}\\) is denoted as \\(\\mathcal{V}'\\). The input of \\(\\mathcal{V}'\\) is a set of image regions \\(\\mathbf{I}^{\\mathcal{O}} = \\{\\mathbf{I}_{p'}\\}_{p' \\in \\mathcal{P}'}\\), where \\(p'\\) is a transformed proposal and \\(\\mathbf{I}_{p'}\\) is the cropped image region of \\(p'\\). Same as \\(\\mathcal{V}\\), each \\(\\mathbf{I}_{p'}\\) is first mapped to a sequence of tokens \\(\\mathbf{X} \\in \\mathbb{R}^{N_x \\times d_x}\\), where \\(\\mathbf{X}_{1:N_x - 1}\\) are the patch tokens of \\(\\mathbf{I}_{p'}\\) and \\(\\mathbf{X}_{N_x}\\) is the [CLS] token. We then augment \\(\\mathbf{X}\\) with an [OBJ] token:" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.195, + 0.469, + 0.213 + ], + "angle": 0, + "content": "\\[\n\\mathbf {X} ^ {\\prime} = [ \\mathbf {X}; x _ {[ \\mathrm {O B J} ]} ] \\in \\mathbb {R} ^ {(N _ {x} + 1) \\times d _ {x}}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.218, + 0.47, + 0.275 + ], + "angle": 0, + "content": "Since the [OBJ] token serves a similar purpose as the [CLS] token, we initialize \\( x_{[\\mathrm{OBJ}]} = \\mathbf{X}_{N_x} \\). To regulate the interaction between [OBJ] and the other tokens, we construct a mask \\( m \\in \\mathbb{R}^{N_x} \\), such that:" + }, + { + "type": "equation", + "bbox": [ + 0.101, + 0.279, + 0.47, + 0.315 + ], + "angle": 0, + "content": "\\[\nm _ {i} = \\left\\{ \\begin{array}{l l} \\mathbb {1} \\left\\{\\text {t h e} i \\text {- t h p a t c h o v e r l a p s w i t h} p \\right\\}, & i < N _ {x} \\\\ 0, & i = N _ {x} \\end{array} \\right. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.32, + 0.469, + 0.389 + ], + "angle": 0, + "content": "where \\(\\mathbb{1}\\{\\cdot\\}\\) is the indicator function. Intuitively, eq. (10) means that [OBJ] only attends to the patch tokens that are covered by the original proposal \\(p\\). To maintain the original attentions among \\(\\mathbf{X}\\) in \\(\\mathcal{V}\\), our attention mask \\(\\mathbf{M}\\) is constructed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.108, + 0.392, + 0.469, + 0.423 + ], + "angle": 0, + "content": "\\[\n\\mathbf {M} = \\left[ \\begin{array}{c c} \\mathbf {1} ^ {N _ {x} \\times N _ {x}} & \\mathbf {0} ^ {N _ {x}} \\\\ m & 1 \\end{array} \\right] \\in \\{0, 1 \\} ^ {(N _ {x} + 1) \\times (N _ {x} + 1)}. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.428, + 0.47, + 0.456 + ], + "angle": 0, + "content": "Suppose \\(\\mathcal{V}\\) has \\(L\\) attention layers (ignoring FFNs), \\(\\mathcal{V}'\\) is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.092, + 0.46, + 0.469, + 0.51 + ], + "angle": 0, + "content": "\\[\n\\mathbf {X} ^ {l} = \\left\\{ \\begin{array}{l l} \\sigma \\left(\\log \\mathbf {M} + \\mathbf {Q} ^ {l} \\left(\\mathbf {K} ^ {l}\\right) ^ {\\top}\\right) \\mathbf {V} ^ {l} + \\mathbf {X} ^ {l - 1}, & 0 < l \\leq L \\\\ \\mathbf {X} ^ {\\prime}, & l = 0 \\end{array} \\right. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.511, + 0.47, + 0.591 + ], + "angle": 0, + "content": "where \\(\\mathbf{Q}^l\\), \\(\\mathbf{K}^l\\), and \\(\\mathbf{V}^l\\) are linear transformations of \\(\\mathbf{X}^{l-1}\\), \\(\\sigma\\) is Softmax function. Finally, instead of [CLS], we take the output of [OBJ] as the proposal embedding, i.e., \\(\\tilde{e}_p^O = \\mathbf{X}_{N_x+1}^L\\). Iterating over the proposals \\(\\mathcal{P}\\), we obtain a set of accurate proposal embeddings \\(\\tilde{\\mathcal{E}}^O\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.599, + 0.338, + 0.613 + ], + "angle": 0, + "content": "4.4. Global and Block Distillation" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.621, + 0.47, + 0.727 + ], + "angle": 0, + "content": "While the object distillation module \\(\\mathcal{M}^O\\) aligns the proposal embeddings \\(\\mathcal{E}^O\\) with the optimized CLIP embeddings \\(\\tilde{\\mathcal{E}}^O\\), the detector lacks a comprehensive understanding of the relation between different proposals. Therefore, we propose a global distillation module \\(\\mathcal{M}^G\\), which transfers the knowledge of the entire image from the CLIP visual encoder \\(\\mathcal{V}\\) to the detector:" + }, + { + "type": "equation", + "bbox": [ + 0.172, + 0.733, + 0.373, + 0.751 + ], + "angle": 0, + "content": "\\[\ne ^ {G} = f (\\operatorname {G A P} \\left(\\mathbf {F} _ {6}\\right)) \\in \\mathbb {R} ^ {d},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.175, + 0.752, + 0.318, + 0.769 + ], + "angle": 0, + "content": "\\[\n\\tilde {e} ^ {G} = \\mathcal {V} (\\mathbf {I}) \\in \\mathbb {R} ^ {d},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.173, + 0.771, + 0.323, + 0.789 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {G} = \\mathcal {L} _ {1} (e ^ {G}, \\tilde {e} ^ {G}),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.795, + 0.469, + 0.825 + ], + "angle": 0, + "content": "where \\( f(\\cdot) \\) is a linear transformation function, \\( \\mathrm{GAP}(\\cdot) \\) is the global average pooling, and \\( \\mathbf{I} \\) is the input image." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Due to the existence of Human Reporting Bias [31], CLIP is prone to ignore non-salient information in an image, e.g., the background or prominent attributes of objects. Such information may be valuable for dense prediction tasks like detection [13]. Thus, we propose a block" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.091, + 0.892, + 0.213 + ], + "angle": 0, + "content": "distillation module \\(\\mathcal{M}^B\\) to complement the missing knowledge in \\(\\mathcal{M}^G\\). We evenly divide the input image into several blocks \\(\\mathcal{B}\\subset \\mathbb{R}^4\\) via a partition function \\(g(\\cdot)\\) and denote the corresponding image regions as \\(\\mathbf{I}^{\\mathcal{B}}\\). The size of each block is fixed to \\(R\\times R\\), where \\(R\\) denotes the input resolution of \\(\\mathcal{V}\\). In this way, the resize and center crop operations in \\(\\mathcal{V}\\) will not take effect, thus avoiding information loss when using \\(\\mathcal{V}\\) to encode the block embeddings \\(\\tilde{\\mathcal{E}}^{B} = \\{\\tilde{e}_{b}^{B}\\}_{b\\in \\mathcal{B}}\\subset \\mathbb{R}^{d}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.212, + 0.892, + 0.283 + ], + "angle": 0, + "content": "On the student side, we apply block pooling and a block head to extract the block embeddings \\(\\mathcal{E}^B = \\{e_b^O\\}_{b\\in \\mathcal{B}}\\subset\\) \\(\\mathbb{R}^d\\) . The block pooling is a combination of the block partition function \\(g(\\cdot)\\) and RoI Align. The loss of our proposed block distillation is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.634, + 0.294, + 0.892, + 0.312 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {B} = \\mathcal {L} _ {1} \\left(\\mathcal {E} ^ {B}, \\tilde {\\mathcal {E}} ^ {B}\\right). \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.324, + 0.892, + 0.385 + ], + "angle": 0, + "content": "Compared with \\(\\mathcal{L}^G\\), \\(\\mathcal{L}^B\\) distills the knowledge of each block with the same weights. Therefore, the ignored information in \\(\\mathcal{M}^G\\) can be compensated by \\(\\mathcal{M}^B\\). Note that neither \\(\\mathcal{M}^G\\) nor \\(\\mathcal{M}^B\\) is used during inference." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.397, + 0.731, + 0.411 + ], + "angle": 0, + "content": "4.5. Pseudo Label Generation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.42, + 0.892, + 0.504 + ], + "angle": 0, + "content": "To investigate the performance of our OADP framework under the G-OVD benchmark (refer to sec. 3), we propose to generate pseudo labels with our modified \\(\\nu'\\). Given a proposal \\(p\\), we first extract the proposal embedding \\(\\tilde{e}_p^O\\) as described in sec. 4.3. Then, the probability of \\(p\\) belonging to category \\(c \\in \\mathcal{C}\\) is given as:" + }, + { + "type": "equation", + "bbox": [ + 0.584, + 0.515, + 0.891, + 0.548 + ], + "angle": 0, + "content": "\\[\nl ^ {\\mathrm {P L}} (p, c) = \\frac {\\tilde {e} _ {p} ^ {O} \\cdot t _ {c}}{\\| \\tilde {e} _ {p} ^ {O} \\| \\cdot \\| t _ {c} \\|}, \\tag {14}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.579, + 0.55, + 0.891, + 0.59 + ], + "angle": 0, + "content": "\\[\nP _ {\\mathcal {C}} ^ {\\mathrm {P L}} (p, c) = \\frac {\\exp \\left(l ^ {\\mathrm {P L}} (p , c)\\right)}{\\sum_ {c ^ {\\prime} \\in C} \\exp \\left(l ^ {\\mathrm {P L}} \\left(p , c ^ {\\prime}\\right)\\right)}, \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.603, + 0.891, + 0.647 + ], + "angle": 0, + "content": "where \\(t_c\\) is the category embedding of \\(c\\). Since \\(P_{\\mathcal{C}}^{\\mathrm{PL}}(p,c)\\) does not reflect the localization quality of \\(p\\) [12], we define the confidence score \\(S_{\\mathcal{C}}(p,c)\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.599, + 0.658, + 0.891, + 0.676 + ], + "angle": 0, + "content": "\\[\nS _ {\\mathcal {C}} (p, c) = P _ {\\mathcal {C}} ^ {\\mathrm {P L}} (p, c) ^ {\\gamma} \\cdot o _ {p} ^ {(1 - \\gamma)}, \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.688, + 0.892, + 0.763 + ], + "angle": 0, + "content": "where \\(\\mathcal{C}\\) includes both base and novel categories, \\(o_p\\in [0,1]\\) is the objectness score of \\(p\\) ,and \\(\\gamma\\) is a constant balancing factor. \\(S_{\\mathcal{C}}(p,c)\\) reflects the probability that \\(p\\) precisely locates an instance of category \\(c\\) . Finally, we apply class-wise NMS on the novel categories to obtain the pseudo labels." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.764, + 0.892, + 0.825 + ], + "angle": 0, + "content": "Note that the Softmax operation in eq. (15) is performed over all categories \\(\\mathcal{C}\\), even though the pseudo labels do not include instances on base categories. Such a design effectively suppresses false positives in the pseudo labels." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The proposals to be labeled are extracted via an RPN model pretrained on \\(\\mathcal{D}^T\\), which only contains annotations of the base categories. ViLD [12] demonstrates that the generalization ability of \\(\\mathcal{P}\\) is strong enough to recall most objects of the novel categories." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11190" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.468, + 0.367 + ], + "angle": 0, + "content": "
BenchmarkMethod\\( \\mathrm{mAP}_{50}^{N} \\)\\( \\mathrm{mAP}_{50}^{B} \\)\\( \\mathrm{mAP}_{50} \\)
ZSDSB [1]0.329.224.9
DELO [56]3.413.813.0
PL [34]4.135.927.9
V-OVDViLD [12]27.659.551.3
RegionCLIP* [53]14.252.842.7
OADP (Ours)30.053.347.2
C-OVDOVR-CNN [49]22.846.039.9
HierKD [28]20.351.343.2
RegionCLIP [53]26.854.847.5
LocOV [3]28.651.345.7
PB-OVD [11]29.144.440.4
G-OVDOV-DETR [48]29.461.052.7
VL-PLM [52]32.354.048.3
OADP (Ours)35.655.850.5
WSDWSDNN [2]19.719.619.6
Cap2Det [47]20.320.120.1
WS-OVDDetic [55]27.847.145.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.378, + 0.47, + 0.461 + ], + "angle": 0, + "content": "Table 2. Comparison with other state-of-the-art methods on the OV-COCO dataset. Methods are grouped by the benchmark they use. \"ZSD\" and \"WSD\" stand for Zero-Shot Detection and Weakly Supervised Detection. \"V-OVD\", \"C-OVD\", \"G-OVD\", and \"WS-OVD\" are introduced in sec. 3. \"RegionCLIP*\" indicates a model without refinement using COCO-Captions." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.478, + 0.468, + 0.559 + ], + "angle": 0, + "content": "
MethodObject DetectionInstance Segmentation
APrAPcAPfAPAPrAPcAPfAP
ViLD [12]16.726.534.227.816.624.630.325.5
DetPro [8]20.827.832.428.419.825.628.925.9
OV-DETR [48]----17.425.032.526.6
OADP (Ours)21.928.432.028.721.726.329.026.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.57, + 0.468, + 0.596 + ], + "angle": 0, + "content": "Table 3. Comparison with other state-of-the-art methods on the OV-LVIS dataset." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.61, + 0.21, + 0.628 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.637, + 0.47, + 0.713 + ], + "angle": 0, + "content": "In this section, we first introduce the detailed experiment setup, including the datasets, evaluation metrics, and implementation details. We then evaluate the performance of our proposed OADP framework and analyze the results compared to the state-of-the-art approaches." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.726, + 0.18, + 0.74 + ], + "angle": 0, + "content": "5.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Experiments are mainly conducted under the open-vocabulary COCO (OV-COCO) setting [49], where the MSCOco 2017 dataset [27] is manually divided into 48 base categories and 17 novel categories. The training dataset contains 107,761 images, and the validation dataset contains 4,836 images. We report the \\(\\mathrm{mAP}_{50}^{\\mathrm{N}}\\), \\(\\mathrm{mAP}_{50}^{\\mathrm{B}}\\), and \\(\\mathrm{mAP}_{50}\\) metrics, i.e., the mAP at IoU threshold 0.5 for novel, base, and all categories. \\(\\mathrm{mAP}_{50}^{\\mathrm{N}}\\) is the main metric. Some experiments are conducted under the open-vocabulary LVIS (OV-LVIS) [12] setting, where the 337 rare categories in" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "LVIS [14] are treated as novel categories, and the other 866 are base categories. Metrics for the OV-LVIS setting are \\(\\mathrm{AP_r}\\), \\(\\mathrm{AP_c}\\), \\(\\mathrm{AP_f}\\), and AP, i.e., the mAP for rare (novel), common, frequent, and all categories. Both object detection and instance segmentation metrics are reported." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.178, + 0.719, + 0.194 + ], + "angle": 0, + "content": "5.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.201, + 0.892, + 0.398 + ], + "angle": 0, + "content": "Training is conducted on 8 V-100 GPUs with batch size 16 in total. We use stochastic gradient descent (SGD) optimizer with 0.02 initial learning rate, 0.9 momentum, and \\(2.5 \\times 10^{-5}\\) weight decay. The student backbone is ResNet50 [16]. Following DetPro [8], we adopt the ViT-B/32 CLIP [32] as the teacher and initialize the student backbone using SoCo [41]. The loss weights \\(w^{O}\\), \\(w^{B}\\), and \\(w^{G}\\) are set to 0.5, 0.25, and 0.25 respectively. Under the OV-COCO setting, we train the detector for 40,000 iterations. At the \\(32,000^{\\mathrm{th}}\\) iteration, the learning rate is divided by 10. For OV-LVIS, we use 2x (24 epochs) training schedule, where the learning rate is divided by 10 at the \\(16^{\\mathrm{th}}\\) and \\(22^{\\mathrm{th}}\\) epochs." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.408, + 0.64, + 0.423 + ], + "angle": 0, + "content": "5.3. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.432, + 0.892, + 0.522 + ], + "angle": 0, + "content": "We compare our OADP framework with the other state-of-the-art OVD methods. As described in sec. 3, we categorize existing OVD methods by the benchmark they belong to. For completeness, we include two related benchmarks: Zero-Shot Detection (ZSD) [1, 18, 24, 35, 51] and Weakly Supervised Detection (WSD) [2, 47]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.523, + 0.892, + 0.765 + ], + "angle": 0, + "content": "Our method mainly focuses on the V-OVD benchmark and the G-OVD benchmark. As shown in tab. 2, our OADP framework achieves \\(30.0\\mathrm{mAP}_{50}^{\\mathrm{N}}\\) on the V-OVD benchmark. RegionCLIP* [53] uses CLIP [32] as the pretrained weight, thus adhering to the V-OVD benchmark. Some V-OVD methods [4, 20, 22, 29, 45] are not included because they rely on large-scale detection and image-text datasets and cannot be compared fairly. While C-OVD is not our primary concern, we include the corresponding methods for reference and report their performance when only COCO Captions [6] is available. Under such constraint, the performance of the caption-based methods is relatively lower than the V-OVD methods, even if additional caption data is used. Moreover, our OADP framework is perpendicular to the caption-based methods and has the potential to achieve higher performance using captions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.766, + 0.892, + 0.902 + ], + "angle": 0, + "content": "For the G-OVD benchmark, we generate pseudo labels for novel categories as described in sec. 4.5. The pseudo labels are then merged with the instance-level annotations for base categories. Training our OADP framework on the mixed dataset yields \\(35.6\\mathrm{mAP}_{50}^{\\mathrm{N}}\\), surpassing the previous SOTA method VL-PLM [52] by \\(3.3\\mathrm{mAP}_{50}^{\\mathrm{N}}\\). Since PromptDet [10] relies on an external dataset (LAION-400M [38]) and uses smaller image size \\((640\\times 640)\\), the result 26.6 \\(\\mathrm{mAP}_{50}^{\\mathrm{N}}\\) is not listed in tab. 2 for fairness." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "11191" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.122, + 0.089, + 0.852, + 0.242 + ], + "angle": 0, + "content": "
Global DistillationBlock DistillationObject DistillationmAPNovel \\( \\mathrm{mAP}_{50} \\)\\( \\mathrm{mAP}_{75} \\)mAPBase \\( \\mathrm{mAP}_{50} \\)\\( \\mathrm{mAP}_{75} \\)mAPAll \\( \\mathrm{mAP}_{50} \\)\\( \\mathrm{mAP}_{75} \\)
✓13.3224.9912.3531.8750.2934.0327.0243.6728.36
13.5125.7212.3632.8251.8935.3127.7745.0429.31
✓14.5727.2513.1734.4553.6037.2029.2546.7131.06
✓15.4927.2315.2535.9955.9638.5730.6348.4532.47
✓✓13.5026.4912.5032.1951.2533.9427.3044.7828.33
✓✓15.4728.8014.6234.0854.2936.2829.2147.6230.61
✓✓15.9229.0115.6435.3055.4537.8830.2348.5332.06
✓✓✓16.2129.9515.4733.3353.2635.4728.8547.1730.24
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.251, + 0.895, + 0.279 + ], + "angle": 0, + "content": "Table 4. Ablation study of the Global, Block, and Object Distillation modules in the OADP framework. The baseline is our re-implemented ViLD-ensemble model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.291, + 0.473, + 0.427 + ], + "angle": 0, + "content": "Tab. 3 shows the comparison between our method and the other state-of-the-art methods on the OV-LVIS dataset. Most of the methods under the OV-LVIS setting adhere to the V-OVD benchmark, so we conduct experiments on the V-OVD benchmark only. For the object detection task, our OADP framework achieves \\(21.9\\mathrm{AP_r}\\), surpassing DetPro [8] by \\(1.1\\mathrm{AP_r}\\). We also report the performance of the instance segmentation task, which achieves \\(21.7\\mathrm{AP_r}\\) and is \\(1.9\\mathrm{AP_r}\\) higher than the previous SOTA method." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.438, + 0.232, + 0.455 + ], + "angle": 0, + "content": "5.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.462, + 0.471, + 0.508 + ], + "angle": 0, + "content": "We conduct ablation studies on the OV-COCO dataset to evaluate the effectiveness of each component in our proposed OADP framework." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.515, + 0.473, + 0.834 + ], + "angle": 0, + "content": "OADP. Tab. 4 shows the effectiveness of the distillation modules in our OADP framework. The first row is our re-implemented ViLD-ensemble [12]. Due to the expensive training cost of ViLD, the performance \\(24.99\\mathrm{mAP}_{50}^{\\mathrm{N}}\\) is far below the official \\(27.60\\mathrm{mAP}_{50}^{\\mathrm{N}}\\). Nevertheless, with our proposed distillation pyramid, we can eventually surpass ViLD. The \\(2^{\\mathrm{nd}}\\) to \\(4^{\\mathrm{th}}\\) row in tab. 4 adds \\(\\mathcal{M}^{G}\\), \\(\\mathcal{M}^{B}\\), and \\(\\mathcal{M}^{O}\\) to the baseline respectively. The global distillation module brings a \\(0.73\\mathrm{mAP}_{50}^{\\mathrm{N}}\\) gain, while the other two bring \\(2.26\\mathrm{mAP}_{50}^{\\mathrm{N}}\\) and \\(2.24\\mathrm{mAP}_{50}^{\\mathrm{N}}\\) gain. Note that by adding \\(\\mathcal{M}^{O}\\) to the baseline, we remove the original image head in ViLD-ensemble. Therefore, the \\(2.24\\mathrm{mAP}_{50}^{\\mathrm{N}}\\) gain results from the OAKE module instead of the distillation operation. The \\(5^{\\mathrm{th}}\\) row in tab. 4 adds \\(\\mathcal{M}^{G}\\) and \\(\\mathcal{M}^{B}\\) together. While the 26.49 performance is higher than sole \\(\\mathcal{M}^{G}\\), it is slightly lower than the \\(27.25\\mathrm{mAP}_{50}^{\\mathrm{N}}\\) of \\(\\mathcal{M}^{B}\\). However, along with the object distillation module \\(\\mathcal{M}^{O}\\), \\(\\mathcal{M}^{G}\\), and \\(\\mathcal{M}^{B}\\) achieves \\(28.80\\mathrm{mAP}_{50}^{\\mathrm{N}}\\) and \\(29.01\\mathrm{mAP}_{50}^{\\mathrm{N}}\\), suggesting that \\(\\mathcal{M}^{G}\\) and \\(\\mathcal{M}^{B}\\) have a similar function in transferring the global scene knowledge from CLIP to the detector. Using all three modules together, we achieve \\(29.95\\mathrm{mAP}_{50}^{\\mathrm{N}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.472, + 0.903 + ], + "angle": 0, + "content": "OAKE. We demonstrate the effectiveness of our OAKE module in tab. 5. Given the ground truth bounding boxes, we use different strategies to crop their image regions. (1) Baseline: \\(1 \\times\\) crop; (2) ViLD*: \\(1 \\times\\) and \\(1.5 \\times\\) crop; (3)" + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.288, + 0.891, + 0.401 + ], + "angle": 0, + "content": "
MethodMacro PrecisionWeighted Precision
w/o OAKEw/ OAKEw/o OAKEw/ OAKE
Baseline58.08-62.04-
ViLD*63.36-65.91-
MBS61.7063.8364.8165.82
Fixed49.0764.5351.4969.75
Adaptive51.6466.0955.8568.68
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.41, + 0.893, + 0.453 + ], + "angle": 0, + "content": "Table 5. Ablation study of OAKE module. \"ViLD*\" indicates our re-implementation of multi-scale region embedding. \"MBS\", \"Fixed\", and \"Adaptive\" are three transforming strategies." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.467, + 0.892, + 0.649 + ], + "angle": 0, + "content": "MBS: the minimum bounding square of the original bounding box; (4) Fixed: \\(224 \\times 224\\) bounding square; (5) Adaptive: adaptively enlarge the bounding square. For the above strategies, we use CLIP to directly extract embeddings for their image regions (\"w/o mask\"). Alternatively, we can utilize the modified CLIP visual encoder mentioned in sec. 4.3 (\"w/ mask\"). Finally, we classify these embeddings by calculating their similarity with category embeddings. To evaluate the performance, we compute \"Macro Precision\" (precision for each category independently with equal weights) and \"Weighted Precision\" (weights depending on the number of bounding boxes in each class)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.65, + 0.894, + 0.802 + ], + "angle": 0, + "content": "As shown in the \\(1^{\\mathrm{st}}\\) and \\(3^{\\mathrm{rd}}\\) columns, \"Fixed\" and \"Adaptive\" strategies bring performance drops as they crop a larger bounding square compared to other strategies (e.g., \"MBS\") which may introduce additional surrounding distractors that confuse the CLIP visual encoder. However, with our object-aware CLIP visual encoder, the performances of the two strategies boost significantly and suppress others. It validates that the CLIP visual encoder can focus on the proposal object with our mask attention mechanism to extract accurate knowledge." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Pseudo Label. We follow VL-PLM [52] to adopt the COCO-ZS setting for our ablation studies of the pseudo label. Both mAP and the average per-image number of PLs (#PL) on novel categories are used as metrics to evaluate the quality of the pseudo label. The baseline method directly uses [CLS] token of CLIP visual encoder to extract" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "11192" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.124, + 0.11, + 0.137 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.117, + 0.089, + 0.277, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.279, + 0.089, + 0.439, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.441, + 0.089, + 0.583, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.587, + 0.089, + 0.73, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.733, + 0.089, + 0.89, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.207, + 0.11, + 0.22 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.117, + 0.173, + 0.277, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.278, + 0.173, + 0.439, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.441, + 0.173, + 0.582, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.587, + 0.173, + 0.73, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.733, + 0.173, + 0.892, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.291, + 0.11, + 0.303 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.117, + 0.255, + 0.277, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.278, + 0.255, + 0.439, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.441, + 0.255, + 0.582, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.587, + 0.255, + 0.73, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.733, + 0.255, + 0.892, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.348, + 0.894, + 0.377 + ], + "angle": 0, + "content": "Figure 3. Visualization of activation patterns from different detectors. (a) Pseudo labels (green) and ground truth annotations (blue) for each image. (b) The baseline detector. (c) Our OADP. The intensity of the feature response increases from blue to red." + }, + { + "type": "table", + "bbox": [ + 0.091, + 0.385, + 0.458, + 0.514 + ], + "angle": 0, + "content": "
Method\\( \\mathbf{{mAP}}^{\\mathbf{{PL}}} \\)\\( \\mathbf{{mAP}}_{50}^{\\mathbf{{PL}}} \\)\\( \\mathbf{{mAP}}_{75}^{\\mathbf{{PL}}} \\)#PL
Baseline5.012.13.1100
ViLD*17.828.118.6100
Ours19.029.919.9100
Baseline3.99.32.56.53
VL-PLM [52]-25.3-4.26
Ours17.426.518.64.14
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.524, + 0.471, + 0.553 + ], + "angle": 0, + "content": "Table 6. Ablation study of pseudo labels. \"#PL\" is the number of pseudo labels per image." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.565, + 0.47, + 0.837 + ], + "angle": 0, + "content": "proposal embeddings from the original proposal, and it relies merely on the classification score to sort proposals. The poor detection accuracy in \\(1^{\\text{st}}\\) and \\(4^{\\text{th}}\\) rows of tab. 6 show that without the objectness score, the baseline method can not accurately localize objects. We re-implement multiscale region embedding of ViLD [12] with a geometric mean of CLIP classification score and objectiveness score, i.e., \"ViLD*\". We adopt an adaptive transform strategy for proposals and regard the output of [OBJ] as proposal embedding. The score fusion strategy is described in sec. 4.5, where \\(\\gamma\\) is 0.3. Our method achieves the highest mAP when the number of PLs is sufficient. VL-PLM [52] adopts a multi-scale region embedding method similar to ViLD [12] except for an arithmetic mean of classification score and objections score. When the pseudo labels are filtered with a higher confidence threshold, our method still has a significant advantage compared to VL-PLM [52] (26.50 mAPPL compared to 25.30 mAPPL)." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.847, + 0.214, + 0.862 + ], + "angle": 0, + "content": "5.5. Visualization" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Fig. 3 (a) visualizes our generated PLs on novel categories in green with ground truth boxes of base categories in" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.388, + 0.893, + 0.553 + ], + "angle": 0, + "content": "blue. We try our best to ensure the accuracy of PLs as much as possible to be fewer and more precise. These green PLs demonstrate that our proposal embeddings can clearly distinguish novel objects from base ones. Correspondingly, we also show activation maps from baseline (b) and our detector (c) in fig. 3. Taking the \\(2^{\\mathrm{nd}}\\) column as an example, the activation map of our detector accurately highlights more area of novel objects, i.e. \"cup\", with our distillation pyramid mechanism. Therefore, the backbone of OADP generates more informative feature maps, which further help detect novel objects." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.571, + 0.619, + 0.587 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.598, + 0.892, + 0.78 + ], + "angle": 0, + "content": "In this paper, we reconsider the way of knowledge extraction and knowledge transfer in existing KD-based OVD methods and propose an Object-Aware Distillation Pyramid (OADP) framework. To preserve complete and purified object representation in proposals during knowledge extraction, we propose an Object-Aware Knowledge Extraction (OAKE) module to adaptively transform proposals and extract precise object knowledge. A Distillation Pyramid (DP) mechanism is proposed to transfer contextual knowledge about the relation of different objects for better scene understanding. Experiments show that our OADP outperforms previous methods on two popular OVD benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Acknowledgement. This research is supported by the National Key R&D Program of China (2022ZD0115502), National Natural Science Foundation of China (NO. 62122010), Zhejiang Provincial Natural Science Foundation of China under Grant No. LDT23F02022F02, Key Research and Development Program of Zhejiang Province under Grant 2022C01082." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11193" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.116, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[1] Ankan Bansal, Karan Sikka, Gaurav Sharma, Rama Chellappa, and Ajay Divakaran. Zero-Shot Object Detection. In ECCV, 2018. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.165, + 0.471, + 0.195 + ], + "angle": 0, + "content": "[2] Hakan Bilen and Andrea Vedaldi. Weakly Supervised Deep Detection Networks. In CVPR, 2016. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.199, + 0.471, + 0.259 + ], + "angle": 0, + "content": "[3] Maria A. Bravo, Sudhanshu Mittal, and Thomas Brox. Localized Vision-Language Matching for Open-vocabulary Object Detection. In GCRP, 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.263, + 0.471, + 0.338 + ], + "angle": 0, + "content": "[4] Zhaowei Cai, Gukyeong Kwon, Avinash Ravichandran, Erhan Bas, Zhuowen Tu, Rahul Bhotika, and Stefano Soatto. X-DETR: A Versatile Architecture for Instance-wise Vision-Language Tasks. In ECCV, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.342, + 0.471, + 0.402 + ], + "angle": 0, + "content": "[5] Guobin Chen, Wongun Choi, Xiang Yu, Tony Han, and Manmohan Chandraker. Learning efficient object detection models with knowledge distillation. In NeurIPS, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.406, + 0.471, + 0.482 + ], + "angle": 0, + "content": "[6] Xinlei Chen, Hao Fang, Tsung-yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dollar, and C Lawrence Zitnick. Microsoft COCO Captions: Data Collection and Evaluation Server. arXiv preprint arXiv:1504.00325, 2015. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.485, + 0.471, + 0.545 + ], + "angle": 0, + "content": "[7] Xing Dai, Zeren Jiang, Zhao Wu, Yiping Bao, Zhicheng Wang, Si Liu, and Erjin Zhou. General Instance Distillation for Object Detection. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.55, + 0.471, + 0.61 + ], + "angle": 0, + "content": "[8] Yu Du, Fangyun Wei, Zihe Zhang, Miaojing Shi, Yue Gao, and Guoqi Li. Learning to Prompt for Open-Vocabulary Object Detection with Vision-Language Model. In CVPR, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.614, + 0.471, + 0.674 + ], + "angle": 0, + "content": "[9] Zhixing Du, Rui Zhang, Ming Chang, Xishan Zhang, Shaoli Liu, Tianshi Chen, and Yunji Chen. Distilling Object Detectors with Feature Richness. In NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.679, + 0.47, + 0.739 + ], + "angle": 0, + "content": "[10] Chengjian Feng, Yujie Zhong, Zequn Jie, Xiangxiang Chu, Haibing Ren, Xiaolin Wei, Weidi Xie, and Lin Ma. PromptDet: Expand Your Detector Vocabulary with Uncurated Images. In ECCV, 2022. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.743, + 0.47, + 0.803 + ], + "angle": 0, + "content": "[11] Mingfei Gao, Chen Xing, Juan Carlos Niebles, Junnan Li, Ran Xu, Wenhao Liu, and Caiming Xiong. Open Vocabulary Object Detection with Pseudo Bounding-Box Labels. In ECCV, 2022. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.807, + 0.47, + 0.867 + ], + "angle": 0, + "content": "[12] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary Object Detection via Vision and Language Knowledge Distillation. In ICLR, 2022. 1, 2, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.871, + 0.47, + 0.902 + ], + "angle": 0, + "content": "[13] Jianyuan Guo, Kai Han, Yunhe Wang, Han Wu, Xing-hao Chen, Chunjing Xu, and Chang Xu. Distilling" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.471, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.536, + 0.092, + 0.892, + 0.121 + ], + "angle": 0, + "content": "Object Detectors via Decoupled Features. In CVPR, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.127, + 0.892, + 0.171 + ], + "angle": 0, + "content": "[14] Agrim Gupta, Piotr Dollar, and Ross Girshick. LVIS: A Dataset for Large Vocabulary Instance Segmentation. In CVPR, 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.178, + 0.892, + 0.208 + ], + "angle": 0, + "content": "[15] Kaiming He, Georgia Gkioxari, Piotr Dollar, and Ross Girshick. Mask R-CNN. TPAMI, 42(2), 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.214, + 0.892, + 0.258 + ], + "angle": 0, + "content": "[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In CVPR, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.264, + 0.892, + 0.309 + ], + "angle": 0, + "content": "[17] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the Knowledge in a Neural Network. In NeurIPS, 2014. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.314, + 0.892, + 0.36 + ], + "angle": 0, + "content": "[18] Peiliang Huang, Junwei Han, De Cheng, and Dingwen Zhang. Robust Region Feature Synthesizer for Zero-Shot Object Detection. In CVPR, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.366, + 0.892, + 0.442 + ], + "angle": 0, + "content": "[19] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yunhsuan Sung, Zhen Li, and Tom Duerig. Scaling Up Visual and Vision-Language Representation Learning With Noisy Text Supervision. In ICML, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.446, + 0.892, + 0.507 + ], + "angle": 0, + "content": "[20] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. MDETR - Modulated Detection for End-to-End Multi-Modal Understanding. In ICCV, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.512, + 0.892, + 0.572 + ], + "angle": 0, + "content": "[21] Zijian Kang, Peizhen Zhang, Xiangyu Zhang, Jian Sun, and Nanning Zheng. Instance-Conditional Knowledge Distillation for Object Detection. In NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.578, + 0.892, + 0.654 + ], + "angle": 0, + "content": "[22] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, KaiWei Chang, and Jianfeng Gao. Grounded Language-Image Pre-training. In CVPR, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.659, + 0.892, + 0.703 + ], + "angle": 0, + "content": "[23] Quanquan Li, Shengying Jin, and Junjie Yan. Mimicking Very Efficient Network for Object Detection. In CVPR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.709, + 0.892, + 0.769 + ], + "angle": 0, + "content": "[24] Zhihui Li, Lina Yao, Xiaqin Zhang, Xianzhi Wang, Salil Kanhere, and Huaxiang Zhang. Zero-Shot Object Detection with Textual Descriptions. AAAI, 33(1), 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.775, + 0.892, + 0.836 + ], + "angle": 0, + "content": "[25] Yue Liao, Aixi Zhang, Miao Lu, Yongliang Wang, Xiaobo Li, and Si Liu. Gen-vlkt: Simplify association and enhance interaction understanding for hoi detection. In CVPR, pages 20091-20100, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.841, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[26] Tsung-Yi Lin, Piotr Dollar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature Pyramid Networks for Object Detection. In CVPR, 2017. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "11194" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.091, + 0.47, + 0.167 + ], + "angle": 0, + "content": "[27] Tsung-Yi Lin, Michael Maire, Serge Belongie, Lubomir Bourdev, Ross Girshick, James Hays, Pietro Perona, Deva Ramanan, C. Lawrence Zitnick, and Piotr Dóllár. Microsoft COCO: Common Objects in Context. In ECCV, 2014. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.171, + 0.47, + 0.245 + ], + "angle": 0, + "content": "[28] Zongyang Ma, Guan Luo, Jin Gao, Liang Li, Yuxin Chen, Shaoru Wang, Congxuan Zhang, and Weiming Hu. Open-Vocabulary One-Stage Detection with Hierarchical Visual-Language Knowledge Distillation. In CVPR, 2022. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.25, + 0.468, + 0.355 + ], + "angle": 0, + "content": "[29] Matthias Minderer, Alexey Gritsenko, Austin Stone, Maxim Neumann, Dirk Weissenborn, Alexey Dosovitskiy, Aravindh Mahendran, Anurag Arnab, Mostafa Dehghani, Zhuoran Shen, Xiao Wang, Xiaohua Zhai, Thomas Kipf, and Neil Houlsby. Simple Open-Vocabulary Object Detection with Vision Transformers. In ECCV, 2022. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.36, + 0.468, + 0.419 + ], + "angle": 0, + "content": "[30] Seyed Iman Mirzadeh, Mehrdad Farajtabar, Ang Li, Nir Levine, Akihiro Matsukawa, and Hassan Ghasemzadeh. Improved Knowledge Distillation via Teacher Assistant. AAAI, 34(04), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.424, + 0.468, + 0.482 + ], + "angle": 0, + "content": "[31] Ishan Misra, C. Lawrence Zitnick, Margaret Mitchell, and Ross Girshick. Seeing through the Human Reporting Bias: Visual Classifiers from Noisy Human-Centric Labels. In CVPR, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.487, + 0.468, + 0.577 + ], + "angle": 0, + "content": "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning Transferable Visual Models From Natural Language Supervision. In ICML, volume 139, 2021. 1, 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.582, + 0.468, + 0.625 + ], + "angle": 0, + "content": "[33] Shafin Rahman, Salman Khan, and Nick Barnes. Transductive Learning for Zero-Shot Object Detection. In ICCV, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.631, + 0.468, + 0.674 + ], + "angle": 0, + "content": "[34] Shafin Rahman, Salman Khan, and Nick Barnes. Improved Visual-Semantic Alignment for Zero-Shot Object Detection. AAAI, 34(07), 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.68, + 0.468, + 0.737 + ], + "angle": 0, + "content": "[35] Shafin Rahman, Salman H. Khan, and Fatih Porikli. Zero-Shot Object Detection: Joint Recognition and Localization of Novel Concepts. IJCV, 128(12), 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.744, + 0.468, + 0.803 + ], + "angle": 0, + "content": "[36] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks. TPAMI, 39(6), 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.807, + 0.468, + 0.866 + ], + "angle": 0, + "content": "[37] Adriana Romero, Nicolas Ballas, Samira Ebrahimi Kahou, Antoine Chassang, Carlo Gatta, and Yoshua Bengio. FitNets: Hints for Thin Deep Nets. In ICLR, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.872, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[38] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.091, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.536, + 0.092, + 0.892, + 0.15 + ], + "angle": 0, + "content": "Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. LAION-400M: Open Dataset of CLIP-Filtered 400 Million Image-Text Pairs. arXiv preprint arXiv:2111.02114, 2021.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.157, + 0.892, + 0.217 + ], + "angle": 0, + "content": "[39] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual Captions: A Cleaned, Hypernymed, Image Alt-text Dataset For Automatic Image Captioning. In ACL, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.222, + 0.892, + 0.266 + ], + "angle": 0, + "content": "[40] Tao Wang, Li Yuan, Xiaopeng Zhang, and Jiashi Feng. Distilling Object Detectors With Fine-Grained Feature Imitation. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.272, + 0.892, + 0.33 + ], + "angle": 0, + "content": "[41] Fangyun Wei, Yue Gao, Zhirong Wu, Han Hu, and Stephen Lin. Aligning Pretraining for Detection via Object-Level Contrastive Learning. In NeurIPS, volume 27, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.337, + 0.892, + 0.395 + ], + "angle": 0, + "content": "[42] Caixia Yan, Xiaojun Chang, Minnan Luo, Huan Liu, Xiaoqin Zhang, and Qinghua Zheng. Semantics-Guided Contrastive Network for Zero-Shot Object detection. TPAMI, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.402, + 0.892, + 0.46 + ], + "angle": 0, + "content": "[43] Zhendong Yang, Zhe Li, Xiaohu Jiang, Yuan Gong, Zehuan Yuan, Danpei Zhao, and Chun Yuan. Focal and Global Knowledge Distillation for Detectors. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.467, + 0.892, + 0.51 + ], + "angle": 0, + "content": "[44] Zhendong Yang, Zhe Li, Mingqi Shao, Dachuan Shi, Zehuan Yuan, and Chun Yuan. Masked Generative Distillation. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.516, + 0.892, + 0.591 + ], + "angle": 0, + "content": "[45] Lewei Yao, Jianhua Han, Youpeng Wen, Xiaodan Liang, Dan Xu, Wei Zhang, Zhenguo Li, Chunjing Xu, and Hang Xu. DetCLIP: Dictionary-Enriched Visual-Concept Paralleled Pre-training for Open-world Detection. In NeurIPS, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.597, + 0.892, + 0.669 + ], + "angle": 0, + "content": "[46] Lewei Yao, Renjie Pi, Hang Xu, Wei Zhang, Zhenguo Li, and Tong Zhang. G-DetKD: Towards General Distillation Framework for Object Detectors via Contrastive and Semantic-guided Feature Imitation. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.676, + 0.892, + 0.735 + ], + "angle": 0, + "content": "[47] Keren Ye, Mingda Zhang, Adriana Kovashka, Wei Li, Danfeng Qin, and Jesse Berent. Cap2Det: Learning to Amplify Weak Caption Supervision for Object Detection. In ICCV, 2019. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.742, + 0.892, + 0.786 + ], + "angle": 0, + "content": "[48] Yuhang Zang, Wei Li, Kaiyang Zhou, Chen Huang, and Chen Change Loy. Open-Vocabulary DETR with Conditional Matching. ECCV, 2022. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.792, + 0.892, + 0.836 + ], + "angle": 0, + "content": "[49] Alireza Zareian, Kevin Dela Rosa, Derek Hao Hu, and Shih-Fu Chang. Open-Vocabulary Object Detection Using Captions. In CVPR, 2021. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.841, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[50] Peizhen Zhang, Zijian Kang, Tong Yang, Xiangyu Zhang, Nanning Zheng, and Jian Sun. LGD: Label-Guided Self-Distillation for Object Detection. AAAI, 36(3), 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.092, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11195" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.091, + 0.47, + 0.152 + ], + "angle": 0, + "content": "[51] Shizhen Zhao, Changxin Gao, Yuanjie Shao, Lerenhan Li, Changqian Yu, Zhong Ji, and Nong Sang. GT-Net: Generative Transfer Network for Zero-Shot Object Detection. AAAI, 34(07), 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.156, + 0.472, + 0.246 + ], + "angle": 0, + "content": "[52] Shiyu Zhao, Zhixing Zhang, Samuel Schulter, Long Zhao, Vijay Kumar B. G, Anastasis Stathopoulos, Manmohan Chandraker, and Dimitris Metaxas. Exploiting Unlabeled Data with Vision and Language Models for Object Detection. In ECCV, 2022. 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.25, + 0.472, + 0.325 + ], + "angle": 0, + "content": "[53] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Liunian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, and Jianfeng Gao. RegionCLIP: Region-based Language-Image Pretraining. In CVPR, 2022. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.329, + 0.472, + 0.389 + ], + "angle": 0, + "content": "[54] Chunting Zhou, Graham Neubig, and Jiatao Gu. Improve Object Detection with Feature-based Knowledge Distillation: Towards Accurate and Efficient Detectors. In ICLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.393, + 0.472, + 0.453 + ], + "angle": 0, + "content": "[55] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Phillip Krahenbuhl, and Ishan Misra. Detecting Twenty-thousand Classes using Image-level Supervision. In ECCV, 2022. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.458, + 0.472, + 0.502 + ], + "angle": 0, + "content": "[56] Pengkai Zhu, Hanxiao Wang, and Venkatesh Saligrama. Don't Even Look Once: Synthesizing Features for Zero-Shot Detection. In CVPR, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.507, + 0.472, + 0.55 + ], + "angle": 0, + "content": "[57] Pengkai Zhu, Hanxiao Wang, and Venkatesh Saligrama. Zero Shot Detection. TCSVT, 30(4), 2020. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.091, + 0.472, + 0.55 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11196" + } + ] +] \ No newline at end of file diff --git a/2023/Object-Aware Distillation Pyramid for Open-Vocabulary Object Detection/b1741226-bc11-487a-861d-b993de9bcc68_origin.pdf b/2023/Object-Aware Distillation Pyramid for Open-Vocabulary Object Detection/b1741226-bc11-487a-861d-b993de9bcc68_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..71cb8f5cc5ec8fb4c8357de34f134f8ff9b98e84 --- /dev/null +++ b/2023/Object-Aware Distillation Pyramid for Open-Vocabulary Object Detection/b1741226-bc11-487a-861d-b993de9bcc68_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:69f83a41618761d1c2d5cfb8aa9324407f289b943ff8a0132d88c9c01f4f53b1 +size 2402321 diff --git a/2023/Object-Aware Distillation Pyramid for Open-Vocabulary Object Detection/full.md b/2023/Object-Aware Distillation Pyramid for Open-Vocabulary Object Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..21974da14e3507198946653396832dd23c5e4320 --- /dev/null +++ b/2023/Object-Aware Distillation Pyramid for Open-Vocabulary Object Detection/full.md @@ -0,0 +1,405 @@ +# Object-Aware Distillation Pyramid for Open-Vocabulary Object Detection + +Luting Wang $^{1,3}$ Yi Liu $^{1,3}$ Penghui Du $^{1,3}$ Zihan Ding $^{1,3}$ Yue Liao $^{1,3*}$ Qiaosong Qi $^{2}$ Biaolong Chen $^{2}$ Si Liu $^{1,3}$ + +$^{1}$ Institute of Artificial Intelligence, Beihang University $^{2}$ Alibaba Group $^{3}$ Hangzhou Innovation Institute, Beihang University + +# Abstract + +Open-vocabulary object detection aims to provide object detectors trained on a fixed set of object categories with the generalizability to detect objects described by arbitrary text queries. Previous methods adopt knowledge distillation to extract knowledge from Pretrained Vision-and-Language Models (PVLMs) and transfer it to detectors. However, due to the non-adaptive proposal cropping and single-level feature mimicking processes, they suffer from information destruction during knowledge extraction and inefficient knowledge transfer. To remedy these limitations, we propose an Object-Aware Distillation Pyramid (OADP) framework, including an Object-Aware Knowledge Extraction (OAKE) module and a Distillation Pyramid (DP) mechanism. When extracting object knowledge from PVLMs, the former adaptively transforms object proposals and adopts object-aware mask attention to obtain precise and complete knowledge of objects. The latter introduces global and block distillation for more comprehensive knowledge transfer to compensate for the missing relation information in object distillation. Extensive experiments show that our method achieves significant improvement compared to current methods. Especially on the MS-COCO dataset, our OADP framework reaches 35.6 $mAP_{50}^{N}$ , surpassing the current state-of-the-art method by 3.3 $mAP_{50}^{N}$ . Code is released at https://github.com/LutingWang/OADP. + +# 1. Introduction + +Open-vocabulary object detection (OVD) [49] aims to endow object detectors with the generalizability to detect open categories including both base and novel categories where only the former are annotated in the training phase. Pretrained Vision-and-Language Models (PVLMs, e.g., CLIP [32] and ALIGN [19]) have witnessed great progress in recent years, and Knowledge Distillation (KD) [17] has led to a wave of unprecedented advances transferring the zero-shot visual recognition ability from PVLMs to detec + +![](images/f129ed84a80be33e500770427197ac312505311d4f7f87e5ade25a9f87040dba.jpg) +Center Crop w/o Transform + +![](images/5228dc72d6ab1ef6293d0a7752bc4b9b9a8a9c1973784a611771e7aa5f9d2750.jpg) +Center Crop w/ Transform + +![](images/f838741004b1ec6c35d1ec0fe4d9f86148c99278434d41de32a2c71a52c1160b.jpg) + +![](images/d3b767bc3fb8f2db92240c9e23cb954fe9cabdd153cc576188271973516498cd.jpg) +(a) Knowledge Extraction + +![](images/01f9825d24dda904ba1654551663d3702772b6a5aa22b52a6fd00883dd5ec49b.jpg) +Student + +![](images/b6729b616c8ea716e49410ba4b18a42276deb0b2e524fc488b373ac3d0050e6a.jpg) +(b) Knowledge Transfer + +![](images/2d0d5c68704061f8e03882585b6c9fc7bad957d475c4b07cd6f94678ce9b6587.jpg) +Object KD + +![](images/4729854dcc7e885a0b7d0dab4693f1a2e757bd8523712119aefc2d77779e248f.jpg) +Figure 1. An overview of our OADP framework. (a) Directly applying center crop on proposals may throw informative object parts away, resulting in ambiguous image regions. In contrast, our OAKE module extracts complete objects and reduces the influence of surrounding distractors. (b) Our DP mechanism includes global, block, and object KD to achieve effective knowledge transfer. + +tors [12, 25, 28, 29, 48, 53]. KD typically comprises two essential steps, i.e., knowledge extraction and then knowledge transfer. A common practice in OVD is to crop objects with class-agnostic proposals and use the teacher (e.g., CLIP visual encoder) to extract knowledge of the proposals. The knowledge is then transferred to the detector (e.g., Mask R-CNN [15]) via feature mimicking. + +Despite significant development, we argue that conventional approaches still have two main limitations: 1) Dilemma between comprehensiveness and purity during knowledge extraction. As proposals have diverse aspect ratios, the fixed center crop strategy to square them may cut out object parts (fig. 1a). Enlarging those proposals via resizing function may alleviate this problem, but additional surrounding distractors may confuse the teacher to extract accurate proposal knowledge. 2) Missing global scene understanding during knowledge transfer. Conventional approaches merely concentrate on object-level knowledge transfer by directly mimicking the teacher's features of indi + +vidual proposals. As a result, the student cannot fully grasp the contextual characteristics describing the interweaving of different objects. In light of the above discussions, we propose an Object-Aware Distillation Pyramid (OADP) framework to excavate the teacher's knowledge accurately and effectively transfer the knowledge to the student. + +To preserve the complete information of proposals while extracting their CLIP image embeddings, we propose an Object-Aware Knowledge Extraction (OAKE) module. Concretely, given a proposal, we square it with an adaptive resizing function to avoid destroying the object structure and involve object information as much as possible. However, the resizing process inevitably introduces environmental context, which may contain some distractors that confuse the teacher. Therefore, we propose to utilize an object token [OBJ] whose interaction manner during the forward process is almost the same as the class token [CLS] except that it only attends to patch tokens covered by the original proposal. In this way, the extracted embeddings contain precise and complete knowledge of the proposal object. + +To facilitate complete and effective knowledge transfer, we propose a Distillation Pyramid (DP) mechanism (fig. 1b). As previous works only adopt object distillation to align the feature space of detectors and PVLMs, the relation between different objects is neglected. Therefore, we propose global and block distillation to compensate for the missing relation information in object distillation. For global distillation, we optimize the $\mathcal{L}_1$ distance between the detector backbone and the CLIP visual encoder so that the detector learns to encode rich semantics implied in the image scene. However, the CLIP visual encoder is prone to ignore background information, which may also be valuable for detection. Therefore, we take a finer step to divide the input image into several blocks and optimize the $\mathcal{L}_1$ distance between the block embeddings of the detector and the CLIP image encoder. Overall, the above three distillation modules constitute a hierarchical distillation pyramid, allowing for the transfer of more diversified knowledge from CLIP to the detectors. + +We demonstrate the superiority of our OADP framework on MS-COCO [27] and LVIS [14] datasets. On MS-COCO, it improves the state-of-the-art results of $\mathrm{mAP}_{50}^{\mathrm{N}}$ from 32.3 to 35.6. On the LVIS dataset, our OADP framework reaches $21.9\mathrm{AP_r}$ on the object detection task and $21.7\mathrm{AP_r}$ on the instance segmentation task, leading the former methods by more than $1.1\mathrm{AP_r}$ and $1.9\mathrm{AP_r}$ respectively. + +# 2. Related Work + +Knowledge Distillation for Object Detection. KD [17,37] is a technology that helps train compact student models under the supervision of powerful teacher models. Chen et al. [5] apply KD to object detection by implementing feature-based and response-based loss for Faster R-CNN. + +Li et al. [23] apply $\mathcal{L}_2$ loss on features sampled by student proposals. FGFI [40] only distills foreground regions near the object anchors. DeFeat [13] distills the foreground and background regions simultaneously with different factors. GID [7] distills regions where the student and teacher perform differently. G-DetKD [46] proposes a general distillation framework for object detectors. FKD [54] distills the attention map to emphasize the changeable areas. FGD [43] proposes focal and global distillation for comprehensive knowledge transfer. Compared to these detection KD methods [9, 21, 30, 44, 50], our work concentrates on knowledge transfer from PVLMs to detectors to enable open-vocabulary detection. + +Open-Vocabulary Detection. OVD [4, 20, 29, 45] aims to train a model that can detect objects of arbitrary categories, even if the categories are not seen during training. OVR-CNN [49] is the seminal work that proposes this problem and achieves great performance using image captions as well as bounding box annotations. With the prevalence of PVLMs [19,32], ViLD [12] proposes to distill the open-vocabulary knowledge from CLIP to the detector. DetPro [8] improves upon ViLD with prompt optimization. RegionCLIP [53] develops a pretraining strategy to learn region-text alignment. Detic [55] adopts weak supervisions to jointly train the detector. GLIP [22] pretrains on massive image-text pairs in a self-training fashion by unifying the detection and grounding tasks. HierKD [28] proposes instance- and global-level distillation for one-stage detectors. OV-DETR [48] turns DETR into an open-vocabulary detector with conditional binary matching. VL-PLM [52] leverages pseudo labels on novel categories to augment the detector. PB-OVD [11] generates pseudo labels based on the image captions. PromptDet [10] establishes a scalable pipeline with regional prompt learning and self-training. In this paper, we propose an OADP framework focusing on comprehensive object knowledge extraction and effective knowledge transfer. + +# 3. OVD Benchmarks + +According to the training data, we summarize the existing OVD methods into four types of benchmarks: Vanilla OVD (V-OVD), Caption-based OVD (C-OVD), Generalized OVD (G-OVD), and Weakly Supervised OVD (WSOVD). All benchmarks rely on instance-level annotations and large-scale image-text pairs to learn OVD. Some of them use more types of data, as shown in tab. 1. For clarity, we define base categories as those included in the instance-level annotations, and novel categories are the others. + +V-OVD [4, 8, 12, 20, 22, 29, 45, 53] is a pure OVD benchmark setting, which requires the detector only to train in an object detection dataset with fixed categories set. Any information about the novel categories is unavailable, but unan- + +
BenchmarkCaptionCategory PriorImage Label
V-OVD
C-OVD✓
G-OVD✓
WS-OVD✓✓✓
+ +Table 1. Summary of OVD benchmarks. "Caption": in-domain captions like COCO-Captions. "Category Prior": human priors on novel categories. "Image Label": image-level category labels. + +notated data is allowed. A common practice for this benchmark is to learn open vocabulary knowledge from imagetext pairs and transfer the knowledge to detectors through transfer learning or knowledge distillation. V-OVD is similar to ZSD [1,33,42,57], except that V-OVD relies on largescale image-text pairs to acquire open-vocabulary knowledge. Recently, V-OVD has attracted more and more researchers with the development of PVLMs. + +C-OVD [3, 11, 28, 49] adds additional image caption annotation to the V-OVD benchmark. Note that by image caption data, we refer to the in-domain captions of the instance-level annotations, e.g., COCO-Captions [6], instead of the large-scale image-text pairs, e.g., CC3M [39] and CLIP400M [32]. The in-domain captions enrich the instance-level annotations and imply a distribution of potential novel categories. Compared with the V-OVD benchmark, C-OVD requires slightly more annotations and is expected to perform better. + +G-OVD [10,48,52] introduces human priors on novel categories to the V-OVD benchmark. Intuitively, if some novel categories are far more likely to appear during inference, it would be beneficial to prepare for them during training. Most existing methods assume that all the dataset's category names (including the novels) are known to the detectors during training. Therefore, the performance of G-OVD methods may not be fairly comparable with V-OVD and C-OVD methods. A typical solution is to generate instance-level pseudo annotations for the categories. + +WS-OVD [55] further takes advantage of image-level category labels beyond G-OVD. Similar to Weakly Supervised Detection (WSD) [2,47], the image-level category labels reflect the presence of the base and novel categories in each image. Thus, the annotation cost is far more than the benchmarks above. In this case, WS-OVD methods have the greatest potential to push the limit of OVD further. + +# 4. Object-Aware Distillation Pyramid + +We first briefly review the task definition of OVD and the architecture of Faster R-CNN in sec. 4.1. Then, we present the overview of our OADP framework in sec. 4.2. Sec. 4.3 and sec. 4.4 introduce the OAKE module and the DP mechanism in detail. Finally, in sec. 4.5, we demonstrate the procedure to generate pseudo labels based on OAKE. + +# 4.1. Preliminaries + +We represent traditional object detection datasets as $\mathcal{D} = \{(\mathbf{I}_i, \mathcal{O}_i)\}_{i=1}^{|\mathcal{D}|}$ , where $\mathbf{I}_i$ is the $i^{\text{th}}$ image and $\mathcal{O}_i = \{o_{ij}\}_{j=1}^{|\mathcal{O}_i|}$ is the corresponding set of annotated objects. Each object $o$ is a pair of object bounding box $b \in \mathbb{R}^4$ and category $y \in \mathcal{C}$ where $\mathcal{C}$ is the category space of the dataset. We denote the training and validation datasets as $\mathcal{D}^T$ and $\mathcal{D}^V$ , respectively. + +By the convention of OVD, we refer to the category space of $\mathcal{D}^T$ and $\mathcal{D}^V$ as $\mathcal{C}^B$ and $\mathcal{C}$ respectively. Normally, $\mathcal{C}^B\subset \mathcal{C}$ . Categories in $\mathcal{C}^B$ are called base categories, and those that only appear in $\mathcal{D}^V$ are called novel categories. The novel category space is denoted as $\mathcal{C}^N = \mathcal{C}\setminus \mathcal{C}^B\neq \emptyset$ . For each category $c\in \mathcal{C}$ , we use a pretrained text encoder $\mathcal{T}$ to encode its semantic embedding $t_c\in \mathbb{R}^d$ . Specifically, we use a trainable embedding $t_{\mathrm{bg}}\in \mathbb{R}^{d}$ to represent the background class $\mathrm{bg}$ . + +Since our work is based on Faster R-CNN [36], we briefly recap its framework. Given an image $\mathbf{I} \in \mathbb{R}^{H \times W \times 3}$ , the backbone (including FPN [26]) encodes a set of hierarchical feature maps $\mathcal{F} = \{\mathbf{F}_2, \mathbf{F}_3, \dots, \mathbf{F}_6\}$ and the Region Proposal Network (RPN) generates a set of proposals $\mathcal{P} \subset \mathbb{R}^4$ . Then the R-CNN head performs RoI Align on $\mathcal{F}$ to extract proposal embeddings $\mathcal{E} = \{e_p\}_{p \in \mathcal{P}} \subset \mathbb{R}^d$ . The logit of a proposal $p$ being of category $c$ can be defined as: + +$$ +l (p, c) = \frac {e _ {p} \cdot t _ {c}}{\| e _ {p} \| \cdot \| t _ {c} \|}, \tag {1} +$$ + +where $\cdot$ is the dot product, $t_c$ is the category embedding of $c$ . For simplicity, we ignore the temperature $\tau$ in CLIP [32]. The probability of a proposal $p$ belonging to category $c \in \mathcal{C} \cup \{\mathrm{bg}\}$ is: + +$$ +P c (p, c) = \frac {\exp (l (p , c))}{\sum_ {c ^ {\prime} \in C \cup \{b g \}} \exp (l (p , c ^ {\prime}))}. \tag {2} +$$ + +During training, each proposal $p$ is assigned a category label $y_{p} \in \mathcal{C}^{B} \cup \{\text{bog}\}$ . The R-CNN loss is defined as: + +$$ +\mathcal {L} = - \sum_ {p \in P} \log P _ {\mathcal {C} ^ {B}} (p, y _ {p}). \tag {3} +$$ + +For simplicity, we ignore the regression term in $\mathcal{L}$ . + +# 4.2. Overview of OADP + +To inject open-vocabulary concepts into Faster R-CNN, we propose an Object-Aware Distillation Pyramid (OADP) framework (fig. 2), which first extracts knowledge from CLIP [32] and then transfers it to the detector through knowledge distillation (KD) [17]. Specifically, we propose an Object-Aware Knowledge Extraction (OAKE) module, which inserts an [OBJ] token into the frozen CLIP visual encoder $\mathcal{V}$ to extract informative knowledge from expanded region proposals selectively. For more effective knowledge transfer, we propose a Distillation Pyramid (DP) mechanism comprising an object distillation module $\mathcal{M}^{\mathcal{O}}$ , a block distillation module $\mathcal{M}^{\mathcal{B}}$ , and a global distillation module + +![](images/0c0968e16e217028ce68e52945294b5691c0f4f6114f09380418e759946ad3ef.jpg) +Figure 2. Illustration of our OADP training pipeline. We adopt a pyramid architecture comprising three distillation modules: global, block, and object. Given an image $\mathbf{I}$ , RPN generates proposals $\mathcal{P}$ . For object distillation, RoI Align and Object Head are applied for proposal embeddings $\mathcal{E}^O$ . To extract complete and pure object knowledge from CLIP, we crop the image regions $\mathbf{I}^{\mathcal{O}}$ based on the transformed proposals $\mathcal{P}'$ and feed them to $L$ layers of masked attention, where an extra [OBJ] token (yellow) attends to the patches covered by the original proposal. For global and block distillation, GAP and block pooling are used before the corresponding heads to extract the global and block embeddings $(\mathcal{E}^B$ and $e^{G}$ ). The teacher embeddings $\tilde{\mathcal{E}}^B$ and $\tilde{e}^{G}$ are extracted via CLIP from $\mathbf{I}$ and $\mathcal{I}^B$ respectively. + +$\mathcal{M}^G$ . The losses of the three modules are denoted as $\mathcal{L}^O$ , $\mathcal{L}^B$ , and $\mathcal{L}^G$ , respectively. The total training loss is: + +$$ +\mathcal {L} ^ {\text {a l l}} = \mathcal {L} + w ^ {O} \cdot \mathcal {L} ^ {O} + w ^ {B} \cdot \mathcal {L} ^ {B} + w ^ {G} \cdot \mathcal {L} ^ {G}, \tag {4} +$$ + +where $\mathcal{L}$ is the R-CNN loss as defined in eq. (3); $w^{O}, w^{B}$ , and $w^{G}$ are loss weights. + +We follow the inference pipeline of ViLD-ensemble [12] and use $\mathcal{M}^O$ to calibrate $P_{\mathcal{C}}(p,c)$ . Similar to the R-CNN head, $\mathcal{M}^O$ extracts the proposal embeddings $\mathcal{E}^O = \{e_p^O\}_{p\in \mathcal{P}}\subset \mathbb{R}^d$ and computes the logits: + +$$ +l ^ {O} (p, c) = \frac {e _ {p} ^ {O} \cdot t _ {c}}{\left\| e _ {p} ^ {O} \right\| \cdot \left\| t _ {c} \right\|}, \tag {5} +$$ + +$$ +P _ {c} ^ {O} (p, c) = \frac {\exp (l ^ {O} (p , c))}{\sum_ {c ^ {\prime} \in C} \exp (l ^ {O} (p , c ^ {\prime}))}, \tag {6} +$$ + +where $t_c$ is the category embedding of $c$ . The calibrated probability $P_{\mathcal{C}}^{\mathrm{cal}}(p,c)$ is: + +$$ +P _ {\mathcal {C}} ^ {\text {c a l}} (p, c) = \left\{ \begin{array}{l l} \left(P _ {\mathcal {C}} (p, c)\right) ^ {\lambda} \cdot \left(P _ {\mathcal {C}} ^ {O} (p, c)\right) ^ {(1 - \lambda)}, & c \in \mathcal {C} ^ {B} \\ \left(P _ {\mathcal {C}} (p, c)\right) ^ {(1 - \lambda)} \cdot \left(P _ {\mathcal {C}} ^ {O} (p, c)\right) ^ {\lambda}, & c \in \mathcal {C} ^ {N} \\ 1 - \sum_ {c ^ {\prime} \in \mathcal {C}} P _ {\mathcal {C}} (p, c ^ {\prime}), & c = b g \end{array} \right. \tag {7} +$$ + +where $\lambda$ is set to $2/3$ . Note that the block and global distillation modules are not used during the inference phase, so the computation cost of our OADP framework is the same as ViLD-ensemble. + +# 4.3. Object Distillation + +The object distillation module $\mathcal{M}^O$ aims to transfer the object-level knowledge from CLIP [32] to the detector. For + +each proposal $p\in \mathcal{P}$ $\mathcal{M}^O$ motivates the detector to extract a proposal embedding $e_p^O$ that resembles the corresponding embedding $\tilde{e}_p^O$ extracted by the CLIP visual encoder $\nu$ .. + +$$ +\mathcal {L} ^ {O} = \mathcal {L} _ {1} \left(\mathcal {E} ^ {O}, \tilde {\mathcal {E}} ^ {O}\right), \tag {8} +$$ + +where $\tilde{\mathcal{E}}^O = \{\tilde{e}_p^O\}_{p\in \mathcal{P}}\subset \mathbb{R}^d$ denotes the proposal embeddings extracted by $\nu$ . Naturally, the quality of $\tilde{\mathcal{E}}^O$ affects the accuracy of $P_{\mathcal{C}}^{O}(p,c)$ to a large extent. However, current approaches only yield sub-optimal $\tilde{\mathcal{E}}^O$ due to the dilemma between information comprehensiveness and less noise. For example, when non-square proposal regions are directly passed to CLIP, the center crop operation in $\nu$ will cut out informative parts of an object, leading to incomplete structural knowledge about the object. On the other hand, if the proposals are squared or enlarged, the proposal regions will contain more ambient contexts, which may corrupt the proposal embeddings. + +To acquire more accurate $\tilde{\mathcal{E}}^O$ , we propose an Object-Aware Knowledge Extraction (OAKE) module, where the proposals are first transformed and then encoded with a modified version of $\mathcal{V}$ . Given a proposal $p \in \mathcal{P}$ , the transformed proposal $p'$ is a square with side length $s = \sqrt{r \times p_h \times p_w}$ , where $r$ is a constant scale ratio, $p_h$ and $p_w$ are the height and width of $p$ . The center of $p'$ is initially the same as $p$ but may be translated if $p'$ exceeds the image boundaries. All transformed proposals constitute $\mathcal{P}'$ . + +While $\mathcal{P}'$ contributes to the comprehensiveness of object knowledge, it also includes surrounding distractors. To further suppress the contextual noise, we introduce an [OBJ] token into $\mathcal{V}$ and substitute original attention layers with + +masked attention layers. The modified version of $\mathcal{V}$ is denoted as $\mathcal{V}'$ . The input of $\mathcal{V}'$ is a set of image regions $\mathbf{I}^{\mathcal{O}} = \{\mathbf{I}_{p'}\}_{p' \in \mathcal{P}'}$ , where $p'$ is a transformed proposal and $\mathbf{I}_{p'}$ is the cropped image region of $p'$ . Same as $\mathcal{V}$ , each $\mathbf{I}_{p'}$ is first mapped to a sequence of tokens $\mathbf{X} \in \mathbb{R}^{N_x \times d_x}$ , where $\mathbf{X}_{1:N_x - 1}$ are the patch tokens of $\mathbf{I}_{p'}$ and $\mathbf{X}_{N_x}$ is the [CLS] token. We then augment $\mathbf{X}$ with an [OBJ] token: + +$$ +\mathbf {X} ^ {\prime} = [ \mathbf {X}; x _ {[ \mathrm {O B J} ]} ] \in \mathbb {R} ^ {(N _ {x} + 1) \times d _ {x}}. \tag {9} +$$ + +Since the [OBJ] token serves a similar purpose as the [CLS] token, we initialize $x_{[\mathrm{OBJ}]} = \mathbf{X}_{N_x}$ . To regulate the interaction between [OBJ] and the other tokens, we construct a mask $m \in \mathbb{R}^{N_x}$ , such that: + +$$ +m _ {i} = \left\{ \begin{array}{l l} \mathbb {1} \left\{\text {t h e} i \text {- t h p a t c h o v e r l a p s w i t h} p \right\}, & i < N _ {x} \\ 0, & i = N _ {x} \end{array} \right. \tag {10} +$$ + +where $\mathbb{1}\{\cdot\}$ is the indicator function. Intuitively, eq. (10) means that [OBJ] only attends to the patch tokens that are covered by the original proposal $p$ . To maintain the original attentions among $\mathbf{X}$ in $\mathcal{V}$ , our attention mask $\mathbf{M}$ is constructed as follows: + +$$ +\mathbf {M} = \left[ \begin{array}{c c} \mathbf {1} ^ {N _ {x} \times N _ {x}} & \mathbf {0} ^ {N _ {x}} \\ m & 1 \end{array} \right] \in \{0, 1 \} ^ {(N _ {x} + 1) \times (N _ {x} + 1)}. \tag {11} +$$ + +Suppose $\mathcal{V}$ has $L$ attention layers (ignoring FFNs), $\mathcal{V}'$ is defined as follows: + +$$ +\mathbf {X} ^ {l} = \left\{ \begin{array}{l l} \sigma \left(\log \mathbf {M} + \mathbf {Q} ^ {l} \left(\mathbf {K} ^ {l}\right) ^ {\top}\right) \mathbf {V} ^ {l} + \mathbf {X} ^ {l - 1}, & 0 < l \leq L \\ \mathbf {X} ^ {\prime}, & l = 0 \end{array} \right. \tag {12} +$$ + +where $\mathbf{Q}^l$ , $\mathbf{K}^l$ , and $\mathbf{V}^l$ are linear transformations of $\mathbf{X}^{l-1}$ , $\sigma$ is Softmax function. Finally, instead of [CLS], we take the output of [OBJ] as the proposal embedding, i.e., $\tilde{e}_p^O = \mathbf{X}_{N_x+1}^L$ . Iterating over the proposals $\mathcal{P}$ , we obtain a set of accurate proposal embeddings $\tilde{\mathcal{E}}^O$ . + +# 4.4. Global and Block Distillation + +While the object distillation module $\mathcal{M}^O$ aligns the proposal embeddings $\mathcal{E}^O$ with the optimized CLIP embeddings $\tilde{\mathcal{E}}^O$ , the detector lacks a comprehensive understanding of the relation between different proposals. Therefore, we propose a global distillation module $\mathcal{M}^G$ , which transfers the knowledge of the entire image from the CLIP visual encoder $\mathcal{V}$ to the detector: + +$$ +e ^ {G} = f (\operatorname {G A P} \left(\mathbf {F} _ {6}\right)) \in \mathbb {R} ^ {d}, +$$ + +$$ +\tilde {e} ^ {G} = \mathcal {V} (\mathbf {I}) \in \mathbb {R} ^ {d}, +$$ + +$$ +\mathcal {L} ^ {G} = \mathcal {L} _ {1} (e ^ {G}, \tilde {e} ^ {G}), +$$ + +where $f(\cdot)$ is a linear transformation function, $\mathrm{GAP}(\cdot)$ is the global average pooling, and $\mathbf{I}$ is the input image. + +Due to the existence of Human Reporting Bias [31], CLIP is prone to ignore non-salient information in an image, e.g., the background or prominent attributes of objects. Such information may be valuable for dense prediction tasks like detection [13]. Thus, we propose a block + +distillation module $\mathcal{M}^B$ to complement the missing knowledge in $\mathcal{M}^G$ . We evenly divide the input image into several blocks $\mathcal{B}\subset \mathbb{R}^4$ via a partition function $g(\cdot)$ and denote the corresponding image regions as $\mathbf{I}^{\mathcal{B}}$ . The size of each block is fixed to $R\times R$ , where $R$ denotes the input resolution of $\mathcal{V}$ . In this way, the resize and center crop operations in $\mathcal{V}$ will not take effect, thus avoiding information loss when using $\mathcal{V}$ to encode the block embeddings $\tilde{\mathcal{E}}^{B} = \{\tilde{e}_{b}^{B}\}_{b\in \mathcal{B}}\subset \mathbb{R}^{d}$ . + +On the student side, we apply block pooling and a block head to extract the block embeddings $\mathcal{E}^B = \{e_b^O\}_{b\in \mathcal{B}}\subset$ $\mathbb{R}^d$ . The block pooling is a combination of the block partition function $g(\cdot)$ and RoI Align. The loss of our proposed block distillation is defined as: + +$$ +\mathcal {L} ^ {B} = \mathcal {L} _ {1} \left(\mathcal {E} ^ {B}, \tilde {\mathcal {E}} ^ {B}\right). \tag {13} +$$ + +Compared with $\mathcal{L}^G$ , $\mathcal{L}^B$ distills the knowledge of each block with the same weights. Therefore, the ignored information in $\mathcal{M}^G$ can be compensated by $\mathcal{M}^B$ . Note that neither $\mathcal{M}^G$ nor $\mathcal{M}^B$ is used during inference. + +# 4.5. Pseudo Label Generation + +To investigate the performance of our OADP framework under the G-OVD benchmark (refer to sec. 3), we propose to generate pseudo labels with our modified $\nu'$ . Given a proposal $p$ , we first extract the proposal embedding $\tilde{e}_p^O$ as described in sec. 4.3. Then, the probability of $p$ belonging to category $c \in \mathcal{C}$ is given as: + +$$ +l ^ {\mathrm {P L}} (p, c) = \frac {\tilde {e} _ {p} ^ {O} \cdot t _ {c}}{\| \tilde {e} _ {p} ^ {O} \| \cdot \| t _ {c} \|}, \tag {14} +$$ + +$$ +P _ {\mathcal {C}} ^ {\mathrm {P L}} (p, c) = \frac {\exp \left(l ^ {\mathrm {P L}} (p , c)\right)}{\sum_ {c ^ {\prime} \in C} \exp \left(l ^ {\mathrm {P L}} \left(p , c ^ {\prime}\right)\right)}, \tag {15} +$$ + +where $t_c$ is the category embedding of $c$ . Since $P_{\mathcal{C}}^{\mathrm{PL}}(p,c)$ does not reflect the localization quality of $p$ [12], we define the confidence score $S_{\mathcal{C}}(p,c)$ as: + +$$ +S _ {\mathcal {C}} (p, c) = P _ {\mathcal {C}} ^ {\mathrm {P L}} (p, c) ^ {\gamma} \cdot o _ {p} ^ {(1 - \gamma)}, \tag {16} +$$ + +where $\mathcal{C}$ includes both base and novel categories, $o_p\in [0,1]$ is the objectness score of $p$ ,and $\gamma$ is a constant balancing factor. $S_{\mathcal{C}}(p,c)$ reflects the probability that $p$ precisely locates an instance of category $c$ . Finally, we apply class-wise NMS on the novel categories to obtain the pseudo labels. + +Note that the Softmax operation in eq. (15) is performed over all categories $\mathcal{C}$ , even though the pseudo labels do not include instances on base categories. Such a design effectively suppresses false positives in the pseudo labels. + +The proposals to be labeled are extracted via an RPN model pretrained on $\mathcal{D}^T$ , which only contains annotations of the base categories. ViLD [12] demonstrates that the generalization ability of $\mathcal{P}$ is strong enough to recall most objects of the novel categories. + +
BenchmarkMethod\( \mathrm{mAP}_{50}^{N} \)\( \mathrm{mAP}_{50}^{B} \)\( \mathrm{mAP}_{50} \)
ZSDSB [1]0.329.224.9
DELO [56]3.413.813.0
PL [34]4.135.927.9
V-OVDViLD [12]27.659.551.3
RegionCLIP* [53]14.252.842.7
OADP (Ours)30.053.347.2
C-OVDOVR-CNN [49]22.846.039.9
HierKD [28]20.351.343.2
RegionCLIP [53]26.854.847.5
LocOV [3]28.651.345.7
PB-OVD [11]29.144.440.4
G-OVDOV-DETR [48]29.461.052.7
VL-PLM [52]32.354.048.3
OADP (Ours)35.655.850.5
WSDWSDNN [2]19.719.619.6
Cap2Det [47]20.320.120.1
WS-OVDDetic [55]27.847.145.0
+ +Table 2. Comparison with other state-of-the-art methods on the OV-COCO dataset. Methods are grouped by the benchmark they use. "ZSD" and "WSD" stand for Zero-Shot Detection and Weakly Supervised Detection. "V-OVD", "C-OVD", "G-OVD", and "WS-OVD" are introduced in sec. 3. "RegionCLIP*" indicates a model without refinement using COCO-Captions. + +
MethodObject DetectionInstance Segmentation
APrAPcAPfAPAPrAPcAPfAP
ViLD [12]16.726.534.227.816.624.630.325.5
DetPro [8]20.827.832.428.419.825.628.925.9
OV-DETR [48]----17.425.032.526.6
OADP (Ours)21.928.432.028.721.726.329.026.6
+ +Table 3. Comparison with other state-of-the-art methods on the OV-LVIS dataset. + +# 5. Experiments + +In this section, we first introduce the detailed experiment setup, including the datasets, evaluation metrics, and implementation details. We then evaluate the performance of our proposed OADP framework and analyze the results compared to the state-of-the-art approaches. + +# 5.1. Datasets + +Experiments are mainly conducted under the open-vocabulary COCO (OV-COCO) setting [49], where the MSCOco 2017 dataset [27] is manually divided into 48 base categories and 17 novel categories. The training dataset contains 107,761 images, and the validation dataset contains 4,836 images. We report the $\mathrm{mAP}_{50}^{\mathrm{N}}$ , $\mathrm{mAP}_{50}^{\mathrm{B}}$ , and $\mathrm{mAP}_{50}$ metrics, i.e., the mAP at IoU threshold 0.5 for novel, base, and all categories. $\mathrm{mAP}_{50}^{\mathrm{N}}$ is the main metric. Some experiments are conducted under the open-vocabulary LVIS (OV-LVIS) [12] setting, where the 337 rare categories in + +LVIS [14] are treated as novel categories, and the other 866 are base categories. Metrics for the OV-LVIS setting are $\mathrm{AP_r}$ , $\mathrm{AP_c}$ , $\mathrm{AP_f}$ , and AP, i.e., the mAP for rare (novel), common, frequent, and all categories. Both object detection and instance segmentation metrics are reported. + +# 5.2. Implementation Details + +Training is conducted on 8 V-100 GPUs with batch size 16 in total. We use stochastic gradient descent (SGD) optimizer with 0.02 initial learning rate, 0.9 momentum, and $2.5 \times 10^{-5}$ weight decay. The student backbone is ResNet50 [16]. Following DetPro [8], we adopt the ViT-B/32 CLIP [32] as the teacher and initialize the student backbone using SoCo [41]. The loss weights $w^{O}$ , $w^{B}$ , and $w^{G}$ are set to 0.5, 0.25, and 0.25 respectively. Under the OV-COCO setting, we train the detector for 40,000 iterations. At the $32,000^{\mathrm{th}}$ iteration, the learning rate is divided by 10. For OV-LVIS, we use 2x (24 epochs) training schedule, where the learning rate is divided by 10 at the $16^{\mathrm{th}}$ and $22^{\mathrm{th}}$ epochs. + +# 5.3. Main Results + +We compare our OADP framework with the other state-of-the-art OVD methods. As described in sec. 3, we categorize existing OVD methods by the benchmark they belong to. For completeness, we include two related benchmarks: Zero-Shot Detection (ZSD) [1, 18, 24, 35, 51] and Weakly Supervised Detection (WSD) [2, 47]. + +Our method mainly focuses on the V-OVD benchmark and the G-OVD benchmark. As shown in tab. 2, our OADP framework achieves $30.0\mathrm{mAP}_{50}^{\mathrm{N}}$ on the V-OVD benchmark. RegionCLIP* [53] uses CLIP [32] as the pretrained weight, thus adhering to the V-OVD benchmark. Some V-OVD methods [4, 20, 22, 29, 45] are not included because they rely on large-scale detection and image-text datasets and cannot be compared fairly. While C-OVD is not our primary concern, we include the corresponding methods for reference and report their performance when only COCO Captions [6] is available. Under such constraint, the performance of the caption-based methods is relatively lower than the V-OVD methods, even if additional caption data is used. Moreover, our OADP framework is perpendicular to the caption-based methods and has the potential to achieve higher performance using captions. + +For the G-OVD benchmark, we generate pseudo labels for novel categories as described in sec. 4.5. The pseudo labels are then merged with the instance-level annotations for base categories. Training our OADP framework on the mixed dataset yields $35.6\mathrm{mAP}_{50}^{\mathrm{N}}$ , surpassing the previous SOTA method VL-PLM [52] by $3.3\mathrm{mAP}_{50}^{\mathrm{N}}$ . Since PromptDet [10] relies on an external dataset (LAION-400M [38]) and uses smaller image size $(640\times 640)$ , the result 26.6 $\mathrm{mAP}_{50}^{\mathrm{N}}$ is not listed in tab. 2 for fairness. + +
Global DistillationBlock DistillationObject DistillationmAPNovel \( \mathrm{mAP}_{50} \)\( \mathrm{mAP}_{75} \)mAPBase \( \mathrm{mAP}_{50} \)\( \mathrm{mAP}_{75} \)mAPAll \( \mathrm{mAP}_{50} \)\( \mathrm{mAP}_{75} \)
✓13.3224.9912.3531.8750.2934.0327.0243.6728.36
13.5125.7212.3632.8251.8935.3127.7745.0429.31
✓14.5727.2513.1734.4553.6037.2029.2546.7131.06
✓15.4927.2315.2535.9955.9638.5730.6348.4532.47
✓✓13.5026.4912.5032.1951.2533.9427.3044.7828.33
✓✓15.4728.8014.6234.0854.2936.2829.2147.6230.61
✓✓15.9229.0115.6435.3055.4537.8830.2348.5332.06
✓✓✓16.2129.9515.4733.3353.2635.4728.8547.1730.24
+ +Tab. 3 shows the comparison between our method and the other state-of-the-art methods on the OV-LVIS dataset. Most of the methods under the OV-LVIS setting adhere to the V-OVD benchmark, so we conduct experiments on the V-OVD benchmark only. For the object detection task, our OADP framework achieves $21.9\mathrm{AP_r}$ , surpassing DetPro [8] by $1.1\mathrm{AP_r}$ . We also report the performance of the instance segmentation task, which achieves $21.7\mathrm{AP_r}$ and is $1.9\mathrm{AP_r}$ higher than the previous SOTA method. + +# 5.4. Ablation Study + +We conduct ablation studies on the OV-COCO dataset to evaluate the effectiveness of each component in our proposed OADP framework. + +OADP. Tab. 4 shows the effectiveness of the distillation modules in our OADP framework. The first row is our re-implemented ViLD-ensemble [12]. Due to the expensive training cost of ViLD, the performance $24.99\mathrm{mAP}_{50}^{\mathrm{N}}$ is far below the official $27.60\mathrm{mAP}_{50}^{\mathrm{N}}$ . Nevertheless, with our proposed distillation pyramid, we can eventually surpass ViLD. The $2^{\mathrm{nd}}$ to $4^{\mathrm{th}}$ row in tab. 4 adds $\mathcal{M}^{G}$ , $\mathcal{M}^{B}$ , and $\mathcal{M}^{O}$ to the baseline respectively. The global distillation module brings a $0.73\mathrm{mAP}_{50}^{\mathrm{N}}$ gain, while the other two bring $2.26\mathrm{mAP}_{50}^{\mathrm{N}}$ and $2.24\mathrm{mAP}_{50}^{\mathrm{N}}$ gain. Note that by adding $\mathcal{M}^{O}$ to the baseline, we remove the original image head in ViLD-ensemble. Therefore, the $2.24\mathrm{mAP}_{50}^{\mathrm{N}}$ gain results from the OAKE module instead of the distillation operation. The $5^{\mathrm{th}}$ row in tab. 4 adds $\mathcal{M}^{G}$ and $\mathcal{M}^{B}$ together. While the 26.49 performance is higher than sole $\mathcal{M}^{G}$ , it is slightly lower than the $27.25\mathrm{mAP}_{50}^{\mathrm{N}}$ of $\mathcal{M}^{B}$ . However, along with the object distillation module $\mathcal{M}^{O}$ , $\mathcal{M}^{G}$ , and $\mathcal{M}^{B}$ achieves $28.80\mathrm{mAP}_{50}^{\mathrm{N}}$ and $29.01\mathrm{mAP}_{50}^{\mathrm{N}}$ , suggesting that $\mathcal{M}^{G}$ and $\mathcal{M}^{B}$ have a similar function in transferring the global scene knowledge from CLIP to the detector. Using all three modules together, we achieve $29.95\mathrm{mAP}_{50}^{\mathrm{N}}$ . + +OAKE. We demonstrate the effectiveness of our OAKE module in tab. 5. Given the ground truth bounding boxes, we use different strategies to crop their image regions. (1) Baseline: $1 \times$ crop; (2) ViLD*: $1 \times$ and $1.5 \times$ crop; (3) + +Table 4. Ablation study of the Global, Block, and Object Distillation modules in the OADP framework. The baseline is our re-implemented ViLD-ensemble model. + +
MethodMacro PrecisionWeighted Precision
w/o OAKEw/ OAKEw/o OAKEw/ OAKE
Baseline58.08-62.04-
ViLD*63.36-65.91-
MBS61.7063.8364.8165.82
Fixed49.0764.5351.4969.75
Adaptive51.6466.0955.8568.68
+ +Table 5. Ablation study of OAKE module. "ViLD*" indicates our re-implementation of multi-scale region embedding. "MBS", "Fixed", and "Adaptive" are three transforming strategies. + +MBS: the minimum bounding square of the original bounding box; (4) Fixed: $224 \times 224$ bounding square; (5) Adaptive: adaptively enlarge the bounding square. For the above strategies, we use CLIP to directly extract embeddings for their image regions ("w/o mask"). Alternatively, we can utilize the modified CLIP visual encoder mentioned in sec. 4.3 ("w/ mask"). Finally, we classify these embeddings by calculating their similarity with category embeddings. To evaluate the performance, we compute "Macro Precision" (precision for each category independently with equal weights) and "Weighted Precision" (weights depending on the number of bounding boxes in each class). + +As shown in the $1^{\mathrm{st}}$ and $3^{\mathrm{rd}}$ columns, "Fixed" and "Adaptive" strategies bring performance drops as they crop a larger bounding square compared to other strategies (e.g., "MBS") which may introduce additional surrounding distractors that confuse the CLIP visual encoder. However, with our object-aware CLIP visual encoder, the performances of the two strategies boost significantly and suppress others. It validates that the CLIP visual encoder can focus on the proposal object with our mask attention mechanism to extract accurate knowledge. + +Pseudo Label. We follow VL-PLM [52] to adopt the COCO-ZS setting for our ablation studies of the pseudo label. Both mAP and the average per-image number of PLs (#PL) on novel categories are used as metrics to evaluate the quality of the pseudo label. The baseline method directly uses [CLS] token of CLIP visual encoder to extract + +![](images/35b4f78da566dc64c599b1e10fc62a60da20cc0abb4fab7f1e7f26bc86167a66.jpg) +(a) + +![](images/79c362e78c9985073a685a082e3656578966290e80990a17d3df561a9895a488.jpg) + +![](images/0690b25cafa5b8cecca7acb3ec5121def590defb004a492ed002e3ca50413d75.jpg) + +![](images/38e1a2e284f4805d2f556213ea90b11f59603d4a72110a57094ae10488f750a0.jpg) + +![](images/765ceed1ccebeaa508ec89ca131d7923f7719bd59c6afff9afd697df17041a56.jpg) + +![](images/6f5222ffa005bfa522e551d5c1be9dde1d18c991d05aff23b4d9ac8b0ceeda7b.jpg) +(b) + +![](images/260fbe17168e458edb23266bc33efa504d7b678d10111bb28346accd3a2b477e.jpg) + +![](images/d4189346658dbe9318f68e1471279fd1e1012d4859fc3eb03d86a6a6e69741d7.jpg) + +![](images/45f3a3f6d7c052ae5c2781a452073556a900ee1a9f9720c0c90bdaaf380e9c82.jpg) + +![](images/6b2667397cb062187c8f4e4d9794483346b94a6b5791220c1efb4b0697422d10.jpg) + +![](images/7298ee1d89fd8e1d9b9d9d7ccfc9a7cc29767e9e0b085ee89e663f7239edc5b6.jpg) +(c) + +![](images/b56b68e9d9a4d1641dc2c628af8c5c270e47eb3fe63f72b28ffc1fb87269c6bb.jpg) +Figure 3. Visualization of activation patterns from different detectors. (a) Pseudo labels (green) and ground truth annotations (blue) for each image. (b) The baseline detector. (c) Our OADP. The intensity of the feature response increases from blue to red. + +![](images/1be3077e0e899a9331ef3d6219013c9ea13a985c9ab3c0b37904bf28ef8c0be2.jpg) + +![](images/c9b9b262e5c4cd7dc393afb090bf446b9fe68bc9425360d68b22196d1a76a806.jpg) + +![](images/be5132276b97f020e47b2952e407f49eb84748935e1f616d19141c4b4b52aa1e.jpg) + +
Method\( \mathbf{{mAP}}^{\mathbf{{PL}}} \)\( \mathbf{{mAP}}_{50}^{\mathbf{{PL}}} \)\( \mathbf{{mAP}}_{75}^{\mathbf{{PL}}} \)#PL
Baseline5.012.13.1100
ViLD*17.828.118.6100
Ours19.029.919.9100
Baseline3.99.32.56.53
VL-PLM [52]-25.3-4.26
Ours17.426.518.64.14
+ +Table 6. Ablation study of pseudo labels. "#PL" is the number of pseudo labels per image. + +proposal embeddings from the original proposal, and it relies merely on the classification score to sort proposals. The poor detection accuracy in $1^{\text{st}}$ and $4^{\text{th}}$ rows of tab. 6 show that without the objectness score, the baseline method can not accurately localize objects. We re-implement multiscale region embedding of ViLD [12] with a geometric mean of CLIP classification score and objectiveness score, i.e., "ViLD*". We adopt an adaptive transform strategy for proposals and regard the output of [OBJ] as proposal embedding. The score fusion strategy is described in sec. 4.5, where $\gamma$ is 0.3. Our method achieves the highest mAP when the number of PLs is sufficient. VL-PLM [52] adopts a multi-scale region embedding method similar to ViLD [12] except for an arithmetic mean of classification score and objections score. When the pseudo labels are filtered with a higher confidence threshold, our method still has a significant advantage compared to VL-PLM [52] (26.50 mAPPL compared to 25.30 mAPPL). + +# 5.5. Visualization + +Fig. 3 (a) visualizes our generated PLs on novel categories in green with ground truth boxes of base categories in + +blue. We try our best to ensure the accuracy of PLs as much as possible to be fewer and more precise. These green PLs demonstrate that our proposal embeddings can clearly distinguish novel objects from base ones. Correspondingly, we also show activation maps from baseline (b) and our detector (c) in fig. 3. Taking the $2^{\mathrm{nd}}$ column as an example, the activation map of our detector accurately highlights more area of novel objects, i.e. "cup", with our distillation pyramid mechanism. Therefore, the backbone of OADP generates more informative feature maps, which further help detect novel objects. + +# 6. Conclusion + +In this paper, we reconsider the way of knowledge extraction and knowledge transfer in existing KD-based OVD methods and propose an Object-Aware Distillation Pyramid (OADP) framework. To preserve complete and purified object representation in proposals during knowledge extraction, we propose an Object-Aware Knowledge Extraction (OAKE) module to adaptively transform proposals and extract precise object knowledge. A Distillation Pyramid (DP) mechanism is proposed to transfer contextual knowledge about the relation of different objects for better scene understanding. Experiments show that our OADP outperforms previous methods on two popular OVD benchmarks. + +Acknowledgement. This research is supported by the National Key R&D Program of China (2022ZD0115502), National Natural Science Foundation of China (NO. 62122010), Zhejiang Provincial Natural Science Foundation of China under Grant No. LDT23F02022F02, Key Research and Development Program of Zhejiang Province under Grant 2022C01082. + +# References + +[1] Ankan Bansal, Karan Sikka, Gaurav Sharma, Rama Chellappa, and Ajay Divakaran. Zero-Shot Object Detection. In ECCV, 2018. 3, 6 +[2] Hakan Bilen and Andrea Vedaldi. Weakly Supervised Deep Detection Networks. In CVPR, 2016. 3, 6 +[3] Maria A. Bravo, Sudhanshu Mittal, and Thomas Brox. Localized Vision-Language Matching for Open-vocabulary Object Detection. In GCRP, 2022. 3, 6 +[4] Zhaowei Cai, Gukyeong Kwon, Avinash Ravichandran, Erhan Bas, Zhuowen Tu, Rahul Bhotika, and Stefano Soatto. X-DETR: A Versatile Architecture for Instance-wise Vision-Language Tasks. In ECCV, 2022. 2, 6 +[5] Guobin Chen, Wongun Choi, Xiang Yu, Tony Han, and Manmohan Chandraker. Learning efficient object detection models with knowledge distillation. In NeurIPS, 2017. 2 +[6] Xinlei Chen, Hao Fang, Tsung-yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dollar, and C Lawrence Zitnick. Microsoft COCO Captions: Data Collection and Evaluation Server. arXiv preprint arXiv:1504.00325, 2015. 3, 6 +[7] Xing Dai, Zeren Jiang, Zhao Wu, Yiping Bao, Zhicheng Wang, Si Liu, and Erjin Zhou. General Instance Distillation for Object Detection. In CVPR, 2021. 2 +[8] Yu Du, Fangyun Wei, Zihe Zhang, Miaojing Shi, Yue Gao, and Guoqi Li. Learning to Prompt for Open-Vocabulary Object Detection with Vision-Language Model. In CVPR, 2022. 2, 6, 7 +[9] Zhixing Du, Rui Zhang, Ming Chang, Xishan Zhang, Shaoli Liu, Tianshi Chen, and Yunji Chen. Distilling Object Detectors with Feature Richness. In NeurIPS, 2021. 2 +[10] Chengjian Feng, Yujie Zhong, Zequn Jie, Xiangxiang Chu, Haibing Ren, Xiaolin Wei, Weidi Xie, and Lin Ma. PromptDet: Expand Your Detector Vocabulary with Uncurated Images. In ECCV, 2022. 2, 3, 6 +[11] Mingfei Gao, Chen Xing, Juan Carlos Niebles, Junnan Li, Ran Xu, Wenhao Liu, and Caiming Xiong. Open Vocabulary Object Detection with Pseudo Bounding-Box Labels. In ECCV, 2022. 2, 3, 6 +[12] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary Object Detection via Vision and Language Knowledge Distillation. In ICLR, 2022. 1, 2, 4, 5, 6, 7, 8 +[13] Jianyuan Guo, Kai Han, Yunhe Wang, Han Wu, Xing-hao Chen, Chunjing Xu, and Chang Xu. Distilling + +Object Detectors via Decoupled Features. In CVPR, 2021. 2, 5 +[14] Agrim Gupta, Piotr Dollar, and Ross Girshick. LVIS: A Dataset for Large Vocabulary Instance Segmentation. In CVPR, 2019. 2, 6 +[15] Kaiming He, Georgia Gkioxari, Piotr Dollar, and Ross Girshick. Mask R-CNN. TPAMI, 42(2), 2020. 1 +[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In CVPR, 2016. 6 +[17] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the Knowledge in a Neural Network. In NeurIPS, 2014. 1, 2, 3 +[18] Peiliang Huang, Junwei Han, De Cheng, and Dingwen Zhang. Robust Region Feature Synthesizer for Zero-Shot Object Detection. In CVPR, 2022. 6 +[19] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yunhsuan Sung, Zhen Li, and Tom Duerig. Scaling Up Visual and Vision-Language Representation Learning With Noisy Text Supervision. In ICML, 2021. 1, 2 +[20] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. MDETR - Modulated Detection for End-to-End Multi-Modal Understanding. In ICCV, 2021. 2, 6 +[21] Zijian Kang, Peizhen Zhang, Xiangyu Zhang, Jian Sun, and Nanning Zheng. Instance-Conditional Knowledge Distillation for Object Detection. In NeurIPS, 2021. 2 +[22] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, KaiWei Chang, and Jianfeng Gao. Grounded Language-Image Pre-training. In CVPR, 2022. 2, 6 +[23] Quanquan Li, Shengying Jin, and Junjie Yan. Mimicking Very Efficient Network for Object Detection. In CVPR, 2017. 2 +[24] Zhihui Li, Lina Yao, Xiaqin Zhang, Xianzhi Wang, Salil Kanhere, and Huaxiang Zhang. Zero-Shot Object Detection with Textual Descriptions. AAAI, 33(1), 2019. 6 +[25] Yue Liao, Aixi Zhang, Miao Lu, Yongliang Wang, Xiaobo Li, and Si Liu. Gen-vlkt: Simplify association and enhance interaction understanding for hoi detection. In CVPR, pages 20091-20100, 2022. 1 +[26] Tsung-Yi Lin, Piotr Dollar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature Pyramid Networks for Object Detection. In CVPR, 2017. 3 + +[27] Tsung-Yi Lin, Michael Maire, Serge Belongie, Lubomir Bourdev, Ross Girshick, James Hays, Pietro Perona, Deva Ramanan, C. Lawrence Zitnick, and Piotr Dóllár. Microsoft COCO: Common Objects in Context. In ECCV, 2014. 2, 6 +[28] Zongyang Ma, Guan Luo, Jin Gao, Liang Li, Yuxin Chen, Shaoru Wang, Congxuan Zhang, and Weiming Hu. Open-Vocabulary One-Stage Detection with Hierarchical Visual-Language Knowledge Distillation. In CVPR, 2022. 1, 2, 3, 6 +[29] Matthias Minderer, Alexey Gritsenko, Austin Stone, Maxim Neumann, Dirk Weissenborn, Alexey Dosovitskiy, Aravindh Mahendran, Anurag Arnab, Mostafa Dehghani, Zhuoran Shen, Xiao Wang, Xiaohua Zhai, Thomas Kipf, and Neil Houlsby. Simple Open-Vocabulary Object Detection with Vision Transformers. In ECCV, 2022. 1, 2, 6 +[30] Seyed Iman Mirzadeh, Mehrdad Farajtabar, Ang Li, Nir Levine, Akihiro Matsukawa, and Hassan Ghasemzadeh. Improved Knowledge Distillation via Teacher Assistant. AAAI, 34(04), 2020. 2 +[31] Ishan Misra, C. Lawrence Zitnick, Margaret Mitchell, and Ross Girshick. Seeing through the Human Reporting Bias: Visual Classifiers from Noisy Human-Centric Labels. In CVPR, 2016. 5 +[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning Transferable Visual Models From Natural Language Supervision. In ICML, volume 139, 2021. 1, 2, 3, 4, 6 +[33] Shafin Rahman, Salman Khan, and Nick Barnes. Transductive Learning for Zero-Shot Object Detection. In ICCV, 2019. 3 +[34] Shafin Rahman, Salman Khan, and Nick Barnes. Improved Visual-Semantic Alignment for Zero-Shot Object Detection. AAAI, 34(07), 2020. 6 +[35] Shafin Rahman, Salman H. Khan, and Fatih Porikli. Zero-Shot Object Detection: Joint Recognition and Localization of Novel Concepts. IJCV, 128(12), 2020. 6 +[36] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks. TPAMI, 39(6), 2017. 3 +[37] Adriana Romero, Nicolas Ballas, Samira Ebrahimi Kahou, Antoine Chassang, Carlo Gatta, and Yoshua Bengio. FitNets: Hints for Thin Deep Nets. In ICLR, 2015. 2 +[38] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush + +Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. LAION-400M: Open Dataset of CLIP-Filtered 400 Million Image-Text Pairs. arXiv preprint arXiv:2111.02114, 2021.6 +[39] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual Captions: A Cleaned, Hypernymed, Image Alt-text Dataset For Automatic Image Captioning. In ACL, 2018. 3 +[40] Tao Wang, Li Yuan, Xiaopeng Zhang, and Jiashi Feng. Distilling Object Detectors With Fine-Grained Feature Imitation. In CVPR, 2019. 2 +[41] Fangyun Wei, Yue Gao, Zhirong Wu, Han Hu, and Stephen Lin. Aligning Pretraining for Detection via Object-Level Contrastive Learning. In NeurIPS, volume 27, 2021. 6 +[42] Caixia Yan, Xiaojun Chang, Minnan Luo, Huan Liu, Xiaoqin Zhang, and Qinghua Zheng. Semantics-Guided Contrastive Network for Zero-Shot Object detection. TPAMI, 2022. 3 +[43] Zhendong Yang, Zhe Li, Xiaohu Jiang, Yuan Gong, Zehuan Yuan, Danpei Zhao, and Chun Yuan. Focal and Global Knowledge Distillation for Detectors. In CVPR, 2022. 2 +[44] Zhendong Yang, Zhe Li, Mingqi Shao, Dachuan Shi, Zehuan Yuan, and Chun Yuan. Masked Generative Distillation. In ECCV, 2022. 2 +[45] Lewei Yao, Jianhua Han, Youpeng Wen, Xiaodan Liang, Dan Xu, Wei Zhang, Zhenguo Li, Chunjing Xu, and Hang Xu. DetCLIP: Dictionary-Enriched Visual-Concept Paralleled Pre-training for Open-world Detection. In NeurIPS, 2022. 2, 6 +[46] Lewei Yao, Renjie Pi, Hang Xu, Wei Zhang, Zhenguo Li, and Tong Zhang. G-DetKD: Towards General Distillation Framework for Object Detectors via Contrastive and Semantic-guided Feature Imitation. In ICCV, 2021. 2 +[47] Keren Ye, Mingda Zhang, Adriana Kovashka, Wei Li, Danfeng Qin, and Jesse Berent. Cap2Det: Learning to Amplify Weak Caption Supervision for Object Detection. In ICCV, 2019. 3, 6 +[48] Yuhang Zang, Wei Li, Kaiyang Zhou, Chen Huang, and Chen Change Loy. Open-Vocabulary DETR with Conditional Matching. ECCV, 2022. 1, 2, 3, 6 +[49] Alireza Zareian, Kevin Dela Rosa, Derek Hao Hu, and Shih-Fu Chang. Open-Vocabulary Object Detection Using Captions. In CVPR, 2021. 1, 2, 3, 6 +[50] Peizhen Zhang, Zijian Kang, Tong Yang, Xiangyu Zhang, Nanning Zheng, and Jian Sun. LGD: Label-Guided Self-Distillation for Object Detection. AAAI, 36(3), 2022. 2 + +[51] Shizhen Zhao, Changxin Gao, Yuanjie Shao, Lerenhan Li, Changqian Yu, Zhong Ji, and Nong Sang. GT-Net: Generative Transfer Network for Zero-Shot Object Detection. AAAI, 34(07), 2020. 6 +[52] Shiyu Zhao, Zhixing Zhang, Samuel Schulter, Long Zhao, Vijay Kumar B. G, Anastasis Stathopoulos, Manmohan Chandraker, and Dimitris Metaxas. Exploiting Unlabeled Data with Vision and Language Models for Object Detection. In ECCV, 2022. 2, 3, 6, 7, 8 +[53] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Liunian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, and Jianfeng Gao. RegionCLIP: Region-based Language-Image Pretraining. In CVPR, 2022. 1, 2, 6 +[54] Chunting Zhou, Graham Neubig, and Jiatao Gu. Improve Object Detection with Feature-based Knowledge Distillation: Towards Accurate and Efficient Detectors. In ICLR, 2021. 2 +[55] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Phillip Krahenbuhl, and Ishan Misra. Detecting Twenty-thousand Classes using Image-level Supervision. In ECCV, 2022. 2, 3, 6 +[56] Pengkai Zhu, Hanxiao Wang, and Venkatesh Saligrama. Don't Even Look Once: Synthesizing Features for Zero-Shot Detection. In CVPR, 2020. 6 +[57] Pengkai Zhu, Hanxiao Wang, and Venkatesh Saligrama. Zero Shot Detection. TCSVT, 30(4), 2020. 3 \ No newline at end of file diff --git a/2023/Object-Aware Distillation Pyramid for Open-Vocabulary Object Detection/images.zip b/2023/Object-Aware Distillation Pyramid for Open-Vocabulary Object Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..3b96694d8d35003c68af876ad4d6bac3e6dc1de9 --- /dev/null +++ b/2023/Object-Aware Distillation Pyramid for Open-Vocabulary Object Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:64ee955a6babf230286e8062b00e1217a7acc3fa2252c594ea866838f15ebbe8 +size 570508 diff --git a/2023/Object-Aware Distillation Pyramid for Open-Vocabulary Object Detection/layout.json b/2023/Object-Aware Distillation Pyramid for Open-Vocabulary Object Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f984ad56c8e43f4179c206ea45aef3153325d9fb --- /dev/null +++ b/2023/Object-Aware Distillation Pyramid for Open-Vocabulary Object Detection/layout.json @@ -0,0 +1,13328 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 67, + 103, + 525, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 103, + 525, + 121 + ], + "spans": [ + { + "bbox": [ + 67, + 103, + 525, + 121 + ], + "type": "text", + "content": "Object-Aware Distillation Pyramid for Open-Vocabulary Object Detection" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 118, + 141, + 478, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 141, + 478, + 171 + ], + "spans": [ + { + "bbox": [ + 118, + 141, + 478, + 171 + ], + "type": "text", + "content": "Luting Wang" + }, + { + "bbox": [ + 118, + 141, + 478, + 171 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 118, + 141, + 478, + 171 + ], + "type": "text", + "content": " Yi Liu" + }, + { + "bbox": [ + 118, + 141, + 478, + 171 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 118, + 141, + 478, + 171 + ], + "type": "text", + "content": " Penghui Du" + }, + { + "bbox": [ + 118, + 141, + 478, + 171 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 118, + 141, + 478, + 171 + ], + "type": "text", + "content": " Zihan Ding" + }, + { + "bbox": [ + 118, + 141, + 478, + 171 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 118, + 141, + 478, + 171 + ], + "type": "text", + "content": " Yue Liao" + }, + { + "bbox": [ + 118, + 141, + 478, + 171 + ], + "type": "inline_equation", + "content": "^{1,3*}" + }, + { + "bbox": [ + 118, + 141, + 478, + 171 + ], + "type": "text", + "content": " Qiaosong Qi" + }, + { + "bbox": [ + 118, + 141, + 478, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 118, + 141, + 478, + 171 + ], + "type": "text", + "content": " Biaolong Chen" + }, + { + "bbox": [ + 118, + 141, + 478, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 118, + 141, + 478, + 171 + ], + "type": "text", + "content": " Si Liu" + }, + { + "bbox": [ + 118, + 141, + 478, + 171 + ], + "type": "inline_equation", + "content": "^{1,3}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 121, + 171, + 473, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 171, + 473, + 200 + ], + "spans": [ + { + "bbox": [ + 121, + 171, + 473, + 200 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 121, + 171, + 473, + 200 + ], + "type": "text", + "content": "Institute of Artificial Intelligence, Beihang University " + }, + { + "bbox": [ + 121, + 171, + 473, + 200 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 121, + 171, + 473, + 200 + ], + "type": "text", + "content": "Alibaba Group " + }, + { + "bbox": [ + 121, + 171, + 473, + 200 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 121, + 171, + 473, + 200 + ], + "type": "text", + "content": "Hangzhou Innovation Institute, Beihang University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "spans": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 45, + 251, + 290, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 251, + 290, + 552 + ], + "spans": [ + { + "bbox": [ + 45, + 251, + 290, + 552 + ], + "type": "text", + "content": "Open-vocabulary object detection aims to provide object detectors trained on a fixed set of object categories with the generalizability to detect objects described by arbitrary text queries. Previous methods adopt knowledge distillation to extract knowledge from Pretrained Vision-and-Language Models (PVLMs) and transfer it to detectors. However, due to the non-adaptive proposal cropping and single-level feature mimicking processes, they suffer from information destruction during knowledge extraction and inefficient knowledge transfer. To remedy these limitations, we propose an Object-Aware Distillation Pyramid (OADP) framework, including an Object-Aware Knowledge Extraction (OAKE) module and a Distillation Pyramid (DP) mechanism. When extracting object knowledge from PVLMs, the former adaptively transforms object proposals and adopts object-aware mask attention to obtain precise and complete knowledge of objects. The latter introduces global and block distillation for more comprehensive knowledge transfer to compensate for the missing relation information in object distillation. Extensive experiments show that our method achieves significant improvement compared to current methods. Especially on the MS-COCO dataset, our OADP framework reaches 35.6 " + }, + { + "bbox": [ + 45, + 251, + 290, + 552 + ], + "type": "inline_equation", + "content": "mAP_{50}^{N}" + }, + { + "bbox": [ + 45, + 251, + 290, + 552 + ], + "type": "text", + "content": ", surpassing the current state-of-the-art method by 3.3 " + }, + { + "bbox": [ + 45, + 251, + 290, + 552 + ], + "type": "inline_equation", + "content": "mAP_{50}^{N}" + }, + { + "bbox": [ + 45, + 251, + 290, + 552 + ], + "type": "text", + "content": ". Code is released at https://github.com/LutingWang/OADP." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 567, + 128, + 580 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 128, + 580 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 128, + 580 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 587, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 587, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 587, + 287, + 696 + ], + "type": "text", + "content": "Open-vocabulary object detection (OVD) [49] aims to endow object detectors with the generalizability to detect open categories including both base and novel categories where only the former are annotated in the training phase. Pretrained Vision-and-Language Models (PVLMs, e.g., CLIP [32] and ALIGN [19]) have witnessed great progress in recent years, and Knowledge Distillation (KD) [17] has led to a wave of unprecedented advances transferring the zero-shot visual recognition ability from PVLMs to detec" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 331, + 238, + 421, + 281 + ], + "blocks": [ + { + "bbox": [ + 331, + 228, + 416, + 237 + ], + "lines": [ + { + "bbox": [ + 331, + 228, + 416, + 237 + ], + "spans": [ + { + "bbox": [ + 331, + 228, + 416, + 237 + ], + "type": "text", + "content": "Center Crop w/o Transform" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 331, + 238, + 421, + 281 + ], + "lines": [ + { + "bbox": [ + 331, + 238, + 421, + 281 + ], + "spans": [ + { + "bbox": [ + 331, + 238, + 421, + 281 + ], + "type": "image", + "image_path": "f129ed84a80be33e500770427197ac312505311d4f7f87e5ade25a9f87040dba.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 432, + 238, + 476, + 281 + ], + "blocks": [ + { + "bbox": [ + 437, + 228, + 516, + 237 + ], + "lines": [ + { + "bbox": [ + 437, + 228, + 516, + 237 + ], + "spans": [ + { + "bbox": [ + 437, + 228, + 516, + 237 + ], + "type": "text", + "content": "Center Crop w/ Transform" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 432, + 238, + 476, + 281 + ], + "lines": [ + { + "bbox": [ + 432, + 238, + 476, + 281 + ], + "spans": [ + { + "bbox": [ + 432, + 238, + 476, + 281 + ], + "type": "image", + "image_path": "5228dc72d6ab1ef6293d0a7752bc4b9b9a8a9c1973784a611771e7aa5f9d2750.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 478, + 238, + 522, + 281 + ], + "blocks": [ + { + "bbox": [ + 478, + 238, + 522, + 281 + ], + "lines": [ + { + "bbox": [ + 478, + 238, + 522, + 281 + ], + "spans": [ + { + "bbox": [ + 478, + 238, + 522, + 281 + ], + "type": "image", + "image_path": "f838741004b1ec6c35d1ec0fe4d9f86148c99278434d41de32a2c71a52c1160b.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 330, + 297, + 383, + 390 + ], + "blocks": [ + { + "bbox": [ + 384, + 286, + 468, + 296 + ], + "lines": [ + { + "bbox": [ + 384, + 286, + 468, + 296 + ], + "spans": [ + { + "bbox": [ + 384, + 286, + 468, + 296 + ], + "type": "text", + "content": "(a) Knowledge Extraction" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 330, + 297, + 383, + 390 + ], + "lines": [ + { + "bbox": [ + 330, + 297, + 383, + 390 + ], + "spans": [ + { + "bbox": [ + 330, + 297, + 383, + 390 + ], + "type": "image", + "image_path": "d3b767bc3fb8f2db92240c9e23cb954fe9cabdd153cc576188271973516498cd.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 330, + 316, + 383, + 390 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 316, + 383, + 390 + ], + "spans": [ + { + "bbox": [ + 330, + 316, + 383, + 390 + ], + "type": "image", + "image_path": "2cd17fa81d4cb14ab39115d505262d1f198fc0853372f0ee24de31b80859f983.jpg" + } + ] + } + ], + "index": 17, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 399, + 310, + 441, + 390 + ], + "blocks": [ + { + "bbox": [ + 398, + 300, + 436, + 308 + ], + "lines": [ + { + "bbox": [ + 398, + 300, + 436, + 308 + ], + "spans": [ + { + "bbox": [ + 398, + 300, + 436, + 308 + ], + "type": "text", + "content": "Student" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 399, + 310, + 441, + 390 + ], + "lines": [ + { + "bbox": [ + 399, + 310, + 441, + 390 + ], + "spans": [ + { + "bbox": [ + 399, + 310, + 441, + 390 + ], + "type": "image", + "image_path": "01f9825d24dda904ba1654551663d3702772b6a5aa22b52a6fd00883dd5ec49b.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 432, + 316, + 451, + 390 + ], + "blocks": [ + { + "bbox": [ + 432, + 316, + 451, + 390 + ], + "lines": [ + { + "bbox": [ + 432, + 316, + 451, + 390 + ], + "spans": [ + { + "bbox": [ + 432, + 316, + 451, + 390 + ], + "type": "image", + "image_path": "b6729b616c8ea716e49410ba4b18a42276deb0b2e524fc488b373ac3d0050e6a.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 386, + 392, + 463, + 402 + ], + "lines": [ + { + "bbox": [ + 386, + 392, + 463, + 402 + ], + "spans": [ + { + "bbox": [ + 386, + 392, + 463, + 402 + ], + "type": "text", + "content": "(b) Knowledge Transfer" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 488, + 310, + 505, + 325 + ], + "blocks": [ + { + "bbox": [ + 488, + 310, + 505, + 325 + ], + "lines": [ + { + "bbox": [ + 488, + 310, + 505, + 325 + ], + "spans": [ + { + "bbox": [ + 488, + 310, + 505, + 325 + ], + "type": "image", + "image_path": "2d0d5c68704061f8e03882585b6c9fc7bad957d475c4b07cd6f94678ce9b6587.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 476, + 327, + 511, + 334 + ], + "lines": [ + { + "bbox": [ + 476, + 327, + 511, + 334 + ], + "spans": [ + { + "bbox": [ + 476, + 327, + 511, + 334 + ], + "type": "text", + "content": "Object KD" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 468, + 338, + 519, + 390 + ], + "blocks": [ + { + "bbox": [ + 468, + 338, + 519, + 390 + ], + "lines": [ + { + "bbox": [ + 468, + 338, + 519, + 390 + ], + "spans": [ + { + "bbox": [ + 468, + 338, + 519, + 390 + ], + "type": "image", + "image_path": "4729854dcc7e885a0b7d0dab4693f1a2e757bd8523712119aefc2d77779e248f.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 406, + 547, + 472 + ], + "lines": [ + { + "bbox": [ + 304, + 406, + 547, + 472 + ], + "spans": [ + { + "bbox": [ + 304, + 406, + 547, + 472 + ], + "type": "text", + "content": "Figure 1. An overview of our OADP framework. (a) Directly applying center crop on proposals may throw informative object parts away, resulting in ambiguous image regions. In contrast, our OAKE module extracts complete objects and reduces the influence of surrounding distractors. (b) Our DP mechanism includes global, block, and object KD to achieve effective knowledge transfer." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 484, + 547, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 484, + 547, + 568 + ], + "spans": [ + { + "bbox": [ + 304, + 484, + 547, + 568 + ], + "type": "text", + "content": "tors [12, 25, 28, 29, 48, 53]. KD typically comprises two essential steps, i.e., knowledge extraction and then knowledge transfer. A common practice in OVD is to crop objects with class-agnostic proposals and use the teacher (e.g., CLIP visual encoder) to extract knowledge of the proposals. The knowledge is then transferred to the detector (e.g., Mask R-CNN [15]) via feature mimicking." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": "Despite significant development, we argue that conventional approaches still have two main limitations: 1) Dilemma between comprehensiveness and purity during knowledge extraction. As proposals have diverse aspect ratios, the fixed center crop strategy to square them may cut out object parts (fig. 1a). Enlarging those proposals via resizing function may alleviate this problem, but additional surrounding distractors may confuse the teacher to extract accurate proposal knowledge. 2) Missing global scene understanding during knowledge transfer. Conventional approaches merely concentrate on object-level knowledge transfer by directly mimicking the teacher's features of indi" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 215, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 215, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 215, + 713 + ], + "type": "text", + "content": "*Corresponding author (liaoyue.ai@gmail.com)" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11186" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "text", + "content": "vidual proposals. As a result, the student cannot fully grasp the contextual characteristics describing the interweaving of different objects. In light of the above discussions, we propose an Object-Aware Distillation Pyramid (OADP) framework to excavate the teacher's knowledge accurately and effectively transfer the knowledge to the student." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 144, + 287, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 144, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 144, + 287, + 312 + ], + "type": "text", + "content": "To preserve the complete information of proposals while extracting their CLIP image embeddings, we propose an Object-Aware Knowledge Extraction (OAKE) module. Concretely, given a proposal, we square it with an adaptive resizing function to avoid destroying the object structure and involve object information as much as possible. However, the resizing process inevitably introduces environmental context, which may contain some distractors that confuse the teacher. Therefore, we propose to utilize an object token [OBJ] whose interaction manner during the forward process is almost the same as the class token [CLS] except that it only attends to patch tokens covered by the original proposal. In this way, the extracted embeddings contain precise and complete knowledge of the proposal object." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 312, + 288, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 312, + 288, + 538 + ], + "spans": [ + { + "bbox": [ + 46, + 312, + 288, + 538 + ], + "type": "text", + "content": "To facilitate complete and effective knowledge transfer, we propose a Distillation Pyramid (DP) mechanism (fig. 1b). As previous works only adopt object distillation to align the feature space of detectors and PVLMs, the relation between different objects is neglected. Therefore, we propose global and block distillation to compensate for the missing relation information in object distillation. For global distillation, we optimize the " + }, + { + "bbox": [ + 46, + 312, + 288, + 538 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_1" + }, + { + "bbox": [ + 46, + 312, + 288, + 538 + ], + "type": "text", + "content": " distance between the detector backbone and the CLIP visual encoder so that the detector learns to encode rich semantics implied in the image scene. However, the CLIP visual encoder is prone to ignore background information, which may also be valuable for detection. Therefore, we take a finer step to divide the input image into several blocks and optimize the " + }, + { + "bbox": [ + 46, + 312, + 288, + 538 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_1" + }, + { + "bbox": [ + 46, + 312, + 288, + 538 + ], + "type": "text", + "content": " distance between the block embeddings of the detector and the CLIP image encoder. Overall, the above three distillation modules constitute a hierarchical distillation pyramid, allowing for the transfer of more diversified knowledge from CLIP to the detectors." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 539, + 287, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 539, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 46, + 539, + 287, + 624 + ], + "type": "text", + "content": "We demonstrate the superiority of our OADP framework on MS-COCO [27] and LVIS [14] datasets. On MS-COCO, it improves the state-of-the-art results of " + }, + { + "bbox": [ + 46, + 539, + 287, + 624 + ], + "type": "inline_equation", + "content": "\\mathrm{mAP}_{50}^{\\mathrm{N}}" + }, + { + "bbox": [ + 46, + 539, + 287, + 624 + ], + "type": "text", + "content": " from 32.3 to 35.6. On the LVIS dataset, our OADP framework reaches " + }, + { + "bbox": [ + 46, + 539, + 287, + 624 + ], + "type": "inline_equation", + "content": "21.9\\mathrm{AP_r}" + }, + { + "bbox": [ + 46, + 539, + 287, + 624 + ], + "type": "text", + "content": " on the object detection task and " + }, + { + "bbox": [ + 46, + 539, + 287, + 624 + ], + "type": "inline_equation", + "content": "21.7\\mathrm{AP_r}" + }, + { + "bbox": [ + 46, + 539, + 287, + 624 + ], + "type": "text", + "content": " on the instance segmentation task, leading the former methods by more than " + }, + { + "bbox": [ + 46, + 539, + 287, + 624 + ], + "type": "inline_equation", + "content": "1.1\\mathrm{AP_r}" + }, + { + "bbox": [ + 46, + 539, + 287, + 624 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 539, + 287, + 624 + ], + "type": "inline_equation", + "content": "1.9\\mathrm{AP_r}" + }, + { + "bbox": [ + 46, + 539, + 287, + 624 + ], + "type": "text", + "content": " respectively." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 632, + 134, + 644 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 632, + 134, + 644 + ], + "spans": [ + { + "bbox": [ + 47, + 632, + 134, + 644 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": "Knowledge Distillation for Object Detection. KD [17,37] is a technology that helps train compact student models under the supervision of powerful teacher models. Chen et al. [5] apply KD to object detection by implementing feature-based and response-based loss for Faster R-CNN." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "text", + "content": "Li et al. [23] apply " + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_2" + }, + { + "bbox": [ + 304, + 72, + 545, + 228 + ], + "type": "text", + "content": " loss on features sampled by student proposals. FGFI [40] only distills foreground regions near the object anchors. DeFeat [13] distills the foreground and background regions simultaneously with different factors. GID [7] distills regions where the student and teacher perform differently. G-DetKD [46] proposes a general distillation framework for object detectors. FKD [54] distills the attention map to emphasize the changeable areas. FGD [43] proposes focal and global distillation for comprehensive knowledge transfer. Compared to these detection KD methods [9, 21, 30, 44, 50], our work concentrates on knowledge transfer from PVLMs to detectors to enable open-vocabulary detection." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 234, + 545, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 234, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 234, + 545, + 521 + ], + "type": "text", + "content": "Open-Vocabulary Detection. OVD [4, 20, 29, 45] aims to train a model that can detect objects of arbitrary categories, even if the categories are not seen during training. OVR-CNN [49] is the seminal work that proposes this problem and achieves great performance using image captions as well as bounding box annotations. With the prevalence of PVLMs [19,32], ViLD [12] proposes to distill the open-vocabulary knowledge from CLIP to the detector. DetPro [8] improves upon ViLD with prompt optimization. RegionCLIP [53] develops a pretraining strategy to learn region-text alignment. Detic [55] adopts weak supervisions to jointly train the detector. GLIP [22] pretrains on massive image-text pairs in a self-training fashion by unifying the detection and grounding tasks. HierKD [28] proposes instance- and global-level distillation for one-stage detectors. OV-DETR [48] turns DETR into an open-vocabulary detector with conditional binary matching. VL-PLM [52] leverages pseudo labels on novel categories to augment the detector. PB-OVD [11] generates pseudo labels based on the image captions. PromptDet [10] establishes a scalable pipeline with regional prompt learning and self-training. In this paper, we propose an OADP framework focusing on comprehensive object knowledge extraction and effective knowledge transfer." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 532, + 416, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 532, + 416, + 544 + ], + "spans": [ + { + "bbox": [ + 306, + 532, + 416, + 544 + ], + "type": "text", + "content": "3. OVD Benchmarks" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 552, + 545, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 552, + 545, + 660 + ], + "spans": [ + { + "bbox": [ + 304, + 552, + 545, + 660 + ], + "type": "text", + "content": "According to the training data, we summarize the existing OVD methods into four types of benchmarks: Vanilla OVD (V-OVD), Caption-based OVD (C-OVD), Generalized OVD (G-OVD), and Weakly Supervised OVD (WSOVD). All benchmarks rely on instance-level annotations and large-scale image-text pairs to learn OVD. Some of them use more types of data, as shown in tab. 1. For clarity, we define base categories as those included in the instance-level annotations, and novel categories are the others." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "V-OVD [4, 8, 12, 20, 22, 29, 45, 53] is a pure OVD benchmark setting, which requires the detector only to train in an object detection dataset with fixed categories set. Any information about the novel categories is unavailable, but unan-" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "11187" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 68, + 70, + 267, + 133 + ], + "blocks": [ + { + "bbox": [ + 68, + 70, + 267, + 133 + ], + "lines": [ + { + "bbox": [ + 68, + 70, + 267, + 133 + ], + "spans": [ + { + "bbox": [ + 68, + 70, + 267, + 133 + ], + "type": "table", + "html": "
BenchmarkCaptionCategory PriorImage Label
V-OVD
C-OVD✓
G-OVD✓
WS-OVD✓✓✓
", + "image_path": "0e96446e5d04b49c6f352817abfbccfe40aacc005ae1a4c6a58fa2ee2fc7d515.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 140, + 287, + 174 + ], + "lines": [ + { + "bbox": [ + 47, + 140, + 287, + 174 + ], + "spans": [ + { + "bbox": [ + 47, + 140, + 287, + 174 + ], + "type": "text", + "content": "Table 1. Summary of OVD benchmarks. \"Caption\": in-domain captions like COCO-Captions. \"Category Prior\": human priors on novel categories. \"Image Label\": image-level category labels." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 181, + 287, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 181, + 287, + 277 + ], + "spans": [ + { + "bbox": [ + 46, + 181, + 287, + 277 + ], + "type": "text", + "content": "notated data is allowed. A common practice for this benchmark is to learn open vocabulary knowledge from imagetext pairs and transfer the knowledge to detectors through transfer learning or knowledge distillation. V-OVD is similar to ZSD [1,33,42,57], except that V-OVD relies on largescale image-text pairs to acquire open-vocabulary knowledge. Recently, V-OVD has attracted more and more researchers with the development of PVLMs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 281, + 287, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 281, + 287, + 400 + ], + "spans": [ + { + "bbox": [ + 46, + 281, + 287, + 400 + ], + "type": "text", + "content": "C-OVD [3, 11, 28, 49] adds additional image caption annotation to the V-OVD benchmark. Note that by image caption data, we refer to the in-domain captions of the instance-level annotations, e.g., COCO-Captions [6], instead of the large-scale image-text pairs, e.g., CC3M [39] and CLIP400M [32]. The in-domain captions enrich the instance-level annotations and imply a distribution of potential novel categories. Compared with the V-OVD benchmark, C-OVD requires slightly more annotations and is expected to perform better." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 404, + 287, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 404, + 287, + 524 + ], + "spans": [ + { + "bbox": [ + 46, + 404, + 287, + 524 + ], + "type": "text", + "content": "G-OVD [10,48,52] introduces human priors on novel categories to the V-OVD benchmark. Intuitively, if some novel categories are far more likely to appear during inference, it would be beneficial to prepare for them during training. Most existing methods assume that all the dataset's category names (including the novels) are known to the detectors during training. Therefore, the performance of G-OVD methods may not be fairly comparable with V-OVD and C-OVD methods. A typical solution is to generate instance-level pseudo annotations for the categories." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 528, + 287, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 528, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 46, + 528, + 287, + 612 + ], + "type": "text", + "content": "WS-OVD [55] further takes advantage of image-level category labels beyond G-OVD. Similar to Weakly Supervised Detection (WSD) [2,47], the image-level category labels reflect the presence of the base and novel categories in each image. Thus, the annotation cost is far more than the benchmarks above. In this case, WS-OVD methods have the greatest potential to push the limit of OVD further." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 621, + 241, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 621, + 241, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 621, + 241, + 635 + ], + "type": "text", + "content": "4. Object-Aware Distillation Pyramid" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 641, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 287, + 713 + ], + "type": "text", + "content": "We first briefly review the task definition of OVD and the architecture of Faster R-CNN in sec. 4.1. Then, we present the overview of our OADP framework in sec. 4.2. Sec. 4.3 and sec. 4.4 introduce the OAKE module and the DP mechanism in detail. Finally, in sec. 4.5, we demonstrate the procedure to generate pseudo labels based on OAKE." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 72, + 391, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 391, + 83 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 391, + 83 + ], + "type": "text", + "content": "4.1. Preliminaries" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 90, + 545, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 90, + 545, + 163 + ], + "spans": [ + { + "bbox": [ + 305, + 90, + 545, + 163 + ], + "type": "text", + "content": "We represent traditional object detection datasets as " + }, + { + "bbox": [ + 305, + 90, + 545, + 163 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{(\\mathbf{I}_i, \\mathcal{O}_i)\\}_{i=1}^{|\\mathcal{D}|}" + }, + { + "bbox": [ + 305, + 90, + 545, + 163 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 305, + 90, + 545, + 163 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_i" + }, + { + "bbox": [ + 305, + 90, + 545, + 163 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 305, + 90, + 545, + 163 + ], + "type": "inline_equation", + "content": "i^{\\text{th}}" + }, + { + "bbox": [ + 305, + 90, + 545, + 163 + ], + "type": "text", + "content": " image and " + }, + { + "bbox": [ + 305, + 90, + 545, + 163 + ], + "type": "inline_equation", + "content": "\\mathcal{O}_i = \\{o_{ij}\\}_{j=1}^{|\\mathcal{O}_i|}" + }, + { + "bbox": [ + 305, + 90, + 545, + 163 + ], + "type": "text", + "content": " is the corresponding set of annotated objects. Each object " + }, + { + "bbox": [ + 305, + 90, + 545, + 163 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 305, + 90, + 545, + 163 + ], + "type": "text", + "content": " is a pair of object bounding box " + }, + { + "bbox": [ + 305, + 90, + 545, + 163 + ], + "type": "inline_equation", + "content": "b \\in \\mathbb{R}^4" + }, + { + "bbox": [ + 305, + 90, + 545, + 163 + ], + "type": "text", + "content": " and category " + }, + { + "bbox": [ + 305, + 90, + 545, + 163 + ], + "type": "inline_equation", + "content": "y \\in \\mathcal{C}" + }, + { + "bbox": [ + 305, + 90, + 545, + 163 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 305, + 90, + 545, + 163 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 305, + 90, + 545, + 163 + ], + "type": "text", + "content": " is the category space of the dataset. We denote the training and validation datasets as " + }, + { + "bbox": [ + 305, + 90, + 545, + 163 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^T" + }, + { + "bbox": [ + 305, + 90, + 545, + 163 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 90, + 545, + 163 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^V" + }, + { + "bbox": [ + 305, + 90, + 545, + 163 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 164, + 545, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 164, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 305, + 164, + 545, + 270 + ], + "type": "text", + "content": "By the convention of OVD, we refer to the category space of " + }, + { + "bbox": [ + 305, + 164, + 545, + 270 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^T" + }, + { + "bbox": [ + 305, + 164, + 545, + 270 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 164, + 545, + 270 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^V" + }, + { + "bbox": [ + 305, + 164, + 545, + 270 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 305, + 164, + 545, + 270 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^B" + }, + { + "bbox": [ + 305, + 164, + 545, + 270 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 164, + 545, + 270 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 305, + 164, + 545, + 270 + ], + "type": "text", + "content": " respectively. Normally, " + }, + { + "bbox": [ + 305, + 164, + 545, + 270 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^B\\subset \\mathcal{C}" + }, + { + "bbox": [ + 305, + 164, + 545, + 270 + ], + "type": "text", + "content": ". Categories in " + }, + { + "bbox": [ + 305, + 164, + 545, + 270 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^B" + }, + { + "bbox": [ + 305, + 164, + 545, + 270 + ], + "type": "text", + "content": " are called base categories, and those that only appear in " + }, + { + "bbox": [ + 305, + 164, + 545, + 270 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^V" + }, + { + "bbox": [ + 305, + 164, + 545, + 270 + ], + "type": "text", + "content": " are called novel categories. The novel category space is denoted as " + }, + { + "bbox": [ + 305, + 164, + 545, + 270 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^N = \\mathcal{C}\\setminus \\mathcal{C}^B\\neq \\emptyset" + }, + { + "bbox": [ + 305, + 164, + 545, + 270 + ], + "type": "text", + "content": ". For each category " + }, + { + "bbox": [ + 305, + 164, + 545, + 270 + ], + "type": "inline_equation", + "content": "c\\in \\mathcal{C}" + }, + { + "bbox": [ + 305, + 164, + 545, + 270 + ], + "type": "text", + "content": ", we use a pretrained text encoder " + }, + { + "bbox": [ + 305, + 164, + 545, + 270 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 305, + 164, + 545, + 270 + ], + "type": "text", + "content": " to encode its semantic embedding " + }, + { + "bbox": [ + 305, + 164, + 545, + 270 + ], + "type": "inline_equation", + "content": "t_c\\in \\mathbb{R}^d" + }, + { + "bbox": [ + 305, + 164, + 545, + 270 + ], + "type": "text", + "content": ". Specifically, we use a trainable embedding " + }, + { + "bbox": [ + 305, + 164, + 545, + 270 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{bg}}\\in \\mathbb{R}^{d}" + }, + { + "bbox": [ + 305, + 164, + 545, + 270 + ], + "type": "text", + "content": " to represent the background class " + }, + { + "bbox": [ + 305, + 164, + 545, + 270 + ], + "type": "inline_equation", + "content": "\\mathrm{bg}" + }, + { + "bbox": [ + 305, + 164, + 545, + 270 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 270, + 545, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 270, + 545, + 359 + ], + "spans": [ + { + "bbox": [ + 305, + 270, + 545, + 359 + ], + "type": "text", + "content": "Since our work is based on Faster R-CNN [36], we briefly recap its framework. Given an image " + }, + { + "bbox": [ + 305, + 270, + 545, + 359 + ], + "type": "inline_equation", + "content": "\\mathbf{I} \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 305, + 270, + 545, + 359 + ], + "type": "text", + "content": ", the backbone (including FPN [26]) encodes a set of hierarchical feature maps " + }, + { + "bbox": [ + 305, + 270, + 545, + 359 + ], + "type": "inline_equation", + "content": "\\mathcal{F} = \\{\\mathbf{F}_2, \\mathbf{F}_3, \\dots, \\mathbf{F}_6\\}" + }, + { + "bbox": [ + 305, + 270, + 545, + 359 + ], + "type": "text", + "content": " and the Region Proposal Network (RPN) generates a set of proposals " + }, + { + "bbox": [ + 305, + 270, + 545, + 359 + ], + "type": "inline_equation", + "content": "\\mathcal{P} \\subset \\mathbb{R}^4" + }, + { + "bbox": [ + 305, + 270, + 545, + 359 + ], + "type": "text", + "content": ". Then the R-CNN head performs RoI Align on " + }, + { + "bbox": [ + 305, + 270, + 545, + 359 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 305, + 270, + 545, + 359 + ], + "type": "text", + "content": " to extract proposal embeddings " + }, + { + "bbox": [ + 305, + 270, + 545, + 359 + ], + "type": "inline_equation", + "content": "\\mathcal{E} = \\{e_p\\}_{p \\in \\mathcal{P}} \\subset \\mathbb{R}^d" + }, + { + "bbox": [ + 305, + 270, + 545, + 359 + ], + "type": "text", + "content": ". The logit of a proposal " + }, + { + "bbox": [ + 305, + 270, + 545, + 359 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 305, + 270, + 545, + 359 + ], + "type": "text", + "content": " being of category " + }, + { + "bbox": [ + 305, + 270, + 545, + 359 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 305, + 270, + 545, + 359 + ], + "type": "text", + "content": " can be defined as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 383, + 365, + 545, + 388 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 365, + 545, + 388 + ], + "spans": [ + { + "bbox": [ + 383, + 365, + 545, + 388 + ], + "type": "interline_equation", + "content": "l (p, c) = \\frac {e _ {p} \\cdot t _ {c}}{\\| e _ {p} \\| \\cdot \\| t _ {c} \\|}, \\tag {1}", + "image_path": "608d2b496be387cf31c3672b086738e9a3e23236761906641db6b1ce0822ee3a.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 393, + 545, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 393, + 545, + 438 + ], + "spans": [ + { + "bbox": [ + 305, + 393, + 545, + 438 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 393, + 545, + 438 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 305, + 393, + 545, + 438 + ], + "type": "text", + "content": " is the dot product, " + }, + { + "bbox": [ + 305, + 393, + 545, + 438 + ], + "type": "inline_equation", + "content": "t_c" + }, + { + "bbox": [ + 305, + 393, + 545, + 438 + ], + "type": "text", + "content": " is the category embedding of " + }, + { + "bbox": [ + 305, + 393, + 545, + 438 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 305, + 393, + 545, + 438 + ], + "type": "text", + "content": ". For simplicity, we ignore the temperature " + }, + { + "bbox": [ + 305, + 393, + 545, + 438 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 305, + 393, + 545, + 438 + ], + "type": "text", + "content": " in CLIP [32]. The probability of a proposal " + }, + { + "bbox": [ + 305, + 393, + 545, + 438 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 305, + 393, + 545, + 438 + ], + "type": "text", + "content": " belonging to category " + }, + { + "bbox": [ + 305, + 393, + 545, + 438 + ], + "type": "inline_equation", + "content": "c \\in \\mathcal{C} \\cup \\{\\mathrm{bg}\\}" + }, + { + "bbox": [ + 305, + 393, + 545, + 438 + ], + "type": "text", + "content": " is:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 357, + 445, + 545, + 476 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 357, + 445, + 545, + 476 + ], + "spans": [ + { + "bbox": [ + 357, + 445, + 545, + 476 + ], + "type": "interline_equation", + "content": "P c (p, c) = \\frac {\\exp (l (p , c))}{\\sum_ {c ^ {\\prime} \\in C \\cup \\{b g \\}} \\exp (l (p , c ^ {\\prime}))}. \\tag {2}", + "image_path": "3653267fbdb3f9af425faa69617b52ef69a524f681d6deec453c1d638e459694.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 483, + 544, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 483, + 544, + 507 + ], + "spans": [ + { + "bbox": [ + 305, + 483, + 544, + 507 + ], + "type": "text", + "content": "During training, each proposal " + }, + { + "bbox": [ + 305, + 483, + 544, + 507 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 305, + 483, + 544, + 507 + ], + "type": "text", + "content": " is assigned a category label " + }, + { + "bbox": [ + 305, + 483, + 544, + 507 + ], + "type": "inline_equation", + "content": "y_{p} \\in \\mathcal{C}^{B} \\cup \\{\\text{bog}\\}" + }, + { + "bbox": [ + 305, + 483, + 544, + 507 + ], + "type": "text", + "content": ". The R-CNN loss is defined as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 373, + 514, + 545, + 537 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 514, + 545, + 537 + ], + "spans": [ + { + "bbox": [ + 373, + 514, + 545, + 537 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = - \\sum_ {p \\in P} \\log P _ {\\mathcal {C} ^ {B}} (p, y _ {p}). \\tag {3}", + "image_path": "be56e1a18811a17cf002f0d7fe908d6c789abdace22097e4ec0af4aa3db47f95.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 544, + 509, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 544, + 509, + 556 + ], + "spans": [ + { + "bbox": [ + 305, + 544, + 509, + 556 + ], + "type": "text", + "content": "For simplicity, we ignore the regression term in " + }, + { + "bbox": [ + 305, + 544, + 509, + 556 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 305, + 544, + 509, + 556 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 563, + 419, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 563, + 419, + 574 + ], + "spans": [ + { + "bbox": [ + 306, + 563, + 419, + 574 + ], + "type": "text", + "content": "4.2. Overview of OADP" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 580, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 580, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 580, + 545, + 713 + ], + "type": "text", + "content": "To inject open-vocabulary concepts into Faster R-CNN, we propose an Object-Aware Distillation Pyramid (OADP) framework (fig. 2), which first extracts knowledge from CLIP [32] and then transfers it to the detector through knowledge distillation (KD) [17]. Specifically, we propose an Object-Aware Knowledge Extraction (OAKE) module, which inserts an [OBJ] token into the frozen CLIP visual encoder " + }, + { + "bbox": [ + 304, + 580, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 304, + 580, + 545, + 713 + ], + "type": "text", + "content": " to extract informative knowledge from expanded region proposals selectively. For more effective knowledge transfer, we propose a Distillation Pyramid (DP) mechanism comprising an object distillation module " + }, + { + "bbox": [ + 304, + 580, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{M}^{\\mathcal{O}}" + }, + { + "bbox": [ + 304, + 580, + 545, + 713 + ], + "type": "text", + "content": ", a block distillation module " + }, + { + "bbox": [ + 304, + 580, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{M}^{\\mathcal{B}}" + }, + { + "bbox": [ + 304, + 580, + 545, + 713 + ], + "type": "text", + "content": ", and a global distillation module" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11188" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 73, + 71, + 525, + 261 + ], + "blocks": [ + { + "bbox": [ + 73, + 71, + 525, + 261 + ], + "lines": [ + { + "bbox": [ + 73, + 71, + 525, + 261 + ], + "spans": [ + { + "bbox": [ + 73, + 71, + 525, + 261 + ], + "type": "image", + "image_path": "0c0968e16e217028ce68e52945294b5691c0f4f6114f09380418e759946ad3ef.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 270, + 548, + 338 + ], + "lines": [ + { + "bbox": [ + 46, + 270, + 548, + 338 + ], + "spans": [ + { + "bbox": [ + 46, + 270, + 548, + 338 + ], + "type": "text", + "content": "Figure 2. Illustration of our OADP training pipeline. We adopt a pyramid architecture comprising three distillation modules: global, block, and object. Given an image " + }, + { + "bbox": [ + 46, + 270, + 548, + 338 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 46, + 270, + 548, + 338 + ], + "type": "text", + "content": ", RPN generates proposals " + }, + { + "bbox": [ + 46, + 270, + 548, + 338 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 46, + 270, + 548, + 338 + ], + "type": "text", + "content": ". For object distillation, RoI Align and Object Head are applied for proposal embeddings " + }, + { + "bbox": [ + 46, + 270, + 548, + 338 + ], + "type": "inline_equation", + "content": "\\mathcal{E}^O" + }, + { + "bbox": [ + 46, + 270, + 548, + 338 + ], + "type": "text", + "content": ". To extract complete and pure object knowledge from CLIP, we crop the image regions " + }, + { + "bbox": [ + 46, + 270, + 548, + 338 + ], + "type": "inline_equation", + "content": "\\mathbf{I}^{\\mathcal{O}}" + }, + { + "bbox": [ + 46, + 270, + 548, + 338 + ], + "type": "text", + "content": " based on the transformed proposals " + }, + { + "bbox": [ + 46, + 270, + 548, + 338 + ], + "type": "inline_equation", + "content": "\\mathcal{P}'" + }, + { + "bbox": [ + 46, + 270, + 548, + 338 + ], + "type": "text", + "content": " and feed them to " + }, + { + "bbox": [ + 46, + 270, + 548, + 338 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 46, + 270, + 548, + 338 + ], + "type": "text", + "content": " layers of masked attention, where an extra [OBJ] token (yellow) attends to the patches covered by the original proposal. For global and block distillation, GAP and block pooling are used before the corresponding heads to extract the global and block embeddings " + }, + { + "bbox": [ + 46, + 270, + 548, + 338 + ], + "type": "inline_equation", + "content": "(\\mathcal{E}^B" + }, + { + "bbox": [ + 46, + 270, + 548, + 338 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 270, + 548, + 338 + ], + "type": "inline_equation", + "content": "e^{G}" + }, + { + "bbox": [ + 46, + 270, + 548, + 338 + ], + "type": "text", + "content": "). The teacher embeddings " + }, + { + "bbox": [ + 46, + 270, + 548, + 338 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{E}}^B" + }, + { + "bbox": [ + 46, + 270, + 548, + 338 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 270, + 548, + 338 + ], + "type": "inline_equation", + "content": "\\tilde{e}^{G}" + }, + { + "bbox": [ + 46, + 270, + 548, + 338 + ], + "type": "text", + "content": " are extracted via CLIP from " + }, + { + "bbox": [ + 46, + 270, + 548, + 338 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 46, + 270, + 548, + 338 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 270, + 548, + 338 + ], + "type": "inline_equation", + "content": "\\mathcal{I}^B" + }, + { + "bbox": [ + 46, + 270, + 548, + 338 + ], + "type": "text", + "content": " respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 345, + 288, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 345, + 288, + 369 + ], + "spans": [ + { + "bbox": [ + 47, + 345, + 288, + 369 + ], + "type": "inline_equation", + "content": "\\mathcal{M}^G" + }, + { + "bbox": [ + 47, + 345, + 288, + 369 + ], + "type": "text", + "content": ". The losses of the three modules are denoted as " + }, + { + "bbox": [ + 47, + 345, + 288, + 369 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^O" + }, + { + "bbox": [ + 47, + 345, + 288, + 369 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 345, + 288, + 369 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^B" + }, + { + "bbox": [ + 47, + 345, + 288, + 369 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 345, + 288, + 369 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^G" + }, + { + "bbox": [ + 47, + 345, + 288, + 369 + ], + "type": "text", + "content": ", respectively. The total training loss is:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 83, + 373, + 287, + 388 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 373, + 287, + 388 + ], + "spans": [ + { + "bbox": [ + 83, + 373, + 287, + 388 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {\\text {a l l}} = \\mathcal {L} + w ^ {O} \\cdot \\mathcal {L} ^ {O} + w ^ {B} \\cdot \\mathcal {L} ^ {B} + w ^ {G} \\cdot \\mathcal {L} ^ {G}, \\tag {4}", + "image_path": "9ab9f7fb40c5a7ef8356d7638cc93880594f094240cb65ba5eaa68552778d87e.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 394, + 287, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 394, + 287, + 418 + ], + "spans": [ + { + "bbox": [ + 46, + 394, + 287, + 418 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 394, + 287, + 418 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 46, + 394, + 287, + 418 + ], + "type": "text", + "content": " is the R-CNN loss as defined in eq. (3); " + }, + { + "bbox": [ + 46, + 394, + 287, + 418 + ], + "type": "inline_equation", + "content": "w^{O}, w^{B}" + }, + { + "bbox": [ + 46, + 394, + 287, + 418 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 394, + 287, + 418 + ], + "type": "inline_equation", + "content": "w^{G}" + }, + { + "bbox": [ + 46, + 394, + 287, + 418 + ], + "type": "text", + "content": " are loss weights." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 418, + 287, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 418, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 47, + 418, + 287, + 464 + ], + "type": "text", + "content": "We follow the inference pipeline of ViLD-ensemble [12] and use " + }, + { + "bbox": [ + 47, + 418, + 287, + 464 + ], + "type": "inline_equation", + "content": "\\mathcal{M}^O" + }, + { + "bbox": [ + 47, + 418, + 287, + 464 + ], + "type": "text", + "content": " to calibrate " + }, + { + "bbox": [ + 47, + 418, + 287, + 464 + ], + "type": "inline_equation", + "content": "P_{\\mathcal{C}}(p,c)" + }, + { + "bbox": [ + 47, + 418, + 287, + 464 + ], + "type": "text", + "content": ". Similar to the R-CNN head, " + }, + { + "bbox": [ + 47, + 418, + 287, + 464 + ], + "type": "inline_equation", + "content": "\\mathcal{M}^O" + }, + { + "bbox": [ + 47, + 418, + 287, + 464 + ], + "type": "text", + "content": " extracts the proposal embeddings " + }, + { + "bbox": [ + 47, + 418, + 287, + 464 + ], + "type": "inline_equation", + "content": "\\mathcal{E}^O = \\{e_p^O\\}_{p\\in \\mathcal{P}}\\subset \\mathbb{R}^d" + }, + { + "bbox": [ + 47, + 418, + 287, + 464 + ], + "type": "text", + "content": " and computes the logits:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 99, + 472, + 287, + 497 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 472, + 287, + 497 + ], + "spans": [ + { + "bbox": [ + 99, + 472, + 287, + 497 + ], + "type": "interline_equation", + "content": "l ^ {O} (p, c) = \\frac {e _ {p} ^ {O} \\cdot t _ {c}}{\\left\\| e _ {p} ^ {O} \\right\\| \\cdot \\left\\| t _ {c} \\right\\|}, \\tag {5}", + "image_path": "034b22103697d3ef21a5961a5c46d2a5069841dd67dab506020084dbbd72449b.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 96, + 498, + 287, + 531 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 498, + 287, + 531 + ], + "spans": [ + { + "bbox": [ + 96, + 498, + 287, + 531 + ], + "type": "interline_equation", + "content": "P _ {c} ^ {O} (p, c) = \\frac {\\exp (l ^ {O} (p , c))}{\\sum_ {c ^ {\\prime} \\in C} \\exp (l ^ {O} (p , c ^ {\\prime}))}, \\tag {6}", + "image_path": "77093b4b9f2ffd34d7a290b1331d272f0440c266dd8f98bcad909a07a52ab1f4.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 537, + 287, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 537, + 287, + 562 + ], + "spans": [ + { + "bbox": [ + 46, + 537, + 287, + 562 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 537, + 287, + 562 + ], + "type": "inline_equation", + "content": "t_c" + }, + { + "bbox": [ + 46, + 537, + 287, + 562 + ], + "type": "text", + "content": " is the category embedding of " + }, + { + "bbox": [ + 46, + 537, + 287, + 562 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 537, + 287, + 562 + ], + "type": "text", + "content": ". The calibrated probability " + }, + { + "bbox": [ + 46, + 537, + 287, + 562 + ], + "type": "inline_equation", + "content": "P_{\\mathcal{C}}^{\\mathrm{cal}}(p,c)" + }, + { + "bbox": [ + 46, + 537, + 287, + 562 + ], + "type": "text", + "content": " is:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 567, + 287, + 611 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 567, + 287, + 611 + ], + "spans": [ + { + "bbox": [ + 56, + 567, + 287, + 611 + ], + "type": "interline_equation", + "content": "P _ {\\mathcal {C}} ^ {\\text {c a l}} (p, c) = \\left\\{ \\begin{array}{l l} \\left(P _ {\\mathcal {C}} (p, c)\\right) ^ {\\lambda} \\cdot \\left(P _ {\\mathcal {C}} ^ {O} (p, c)\\right) ^ {(1 - \\lambda)}, & c \\in \\mathcal {C} ^ {B} \\\\ \\left(P _ {\\mathcal {C}} (p, c)\\right) ^ {(1 - \\lambda)} \\cdot \\left(P _ {\\mathcal {C}} ^ {O} (p, c)\\right) ^ {\\lambda}, & c \\in \\mathcal {C} ^ {N} \\\\ 1 - \\sum_ {c ^ {\\prime} \\in \\mathcal {C}} P _ {\\mathcal {C}} (p, c ^ {\\prime}), & c = b g \\end{array} \\right. \\tag {7}", + "image_path": "c4bf5a88f0f0e6c848db508aff478f5b7f1326788c32de00c0a123f92c92abea.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 617, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 617, + 287, + 665 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 617, + 287, + 665 + ], + "type": "text", + "content": " is set to " + }, + { + "bbox": [ + 46, + 617, + 287, + 665 + ], + "type": "inline_equation", + "content": "2/3" + }, + { + "bbox": [ + 46, + 617, + 287, + 665 + ], + "type": "text", + "content": ". Note that the block and global distillation modules are not used during the inference phase, so the computation cost of our OADP framework is the same as ViLD-ensemble." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 673, + 157, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 673, + 157, + 685 + ], + "spans": [ + { + "bbox": [ + 47, + 673, + 157, + 685 + ], + "type": "text", + "content": "4.3. Object Distillation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 690, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 690, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 690, + 288, + 714 + ], + "type": "text", + "content": "The object distillation module " + }, + { + "bbox": [ + 46, + 690, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{M}^O" + }, + { + "bbox": [ + 46, + 690, + 288, + 714 + ], + "type": "text", + "content": " aims to transfer the object-level knowledge from CLIP [32] to the detector. For" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 345, + 547, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 345, + 547, + 383 + ], + "spans": [ + { + "bbox": [ + 305, + 345, + 547, + 383 + ], + "type": "text", + "content": "each proposal " + }, + { + "bbox": [ + 305, + 345, + 547, + 383 + ], + "type": "inline_equation", + "content": "p\\in \\mathcal{P}" + }, + { + "bbox": [ + 305, + 345, + 547, + 383 + ], + "type": "inline_equation", + "content": "\\mathcal{M}^O" + }, + { + "bbox": [ + 305, + 345, + 547, + 383 + ], + "type": "text", + "content": " motivates the detector to extract a proposal embedding " + }, + { + "bbox": [ + 305, + 345, + 547, + 383 + ], + "type": "inline_equation", + "content": "e_p^O" + }, + { + "bbox": [ + 305, + 345, + 547, + 383 + ], + "type": "text", + "content": " that resembles the corresponding embedding " + }, + { + "bbox": [ + 305, + 345, + 547, + 383 + ], + "type": "inline_equation", + "content": "\\tilde{e}_p^O" + }, + { + "bbox": [ + 305, + 345, + 547, + 383 + ], + "type": "text", + "content": " extracted by the CLIP visual encoder " + }, + { + "bbox": [ + 305, + 345, + 547, + 383 + ], + "type": "inline_equation", + "content": "\\nu" + }, + { + "bbox": [ + 305, + 345, + 547, + 383 + ], + "type": "text", + "content": " .." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 388, + 392, + 546, + 405 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 392, + 546, + 405 + ], + "spans": [ + { + "bbox": [ + 388, + 392, + 546, + 405 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {O} = \\mathcal {L} _ {1} \\left(\\mathcal {E} ^ {O}, \\tilde {\\mathcal {E}} ^ {O}\\right), \\tag {8}", + "image_path": "d0976eb2e815249c4d7d4b87e1327484d225574dc1d47a5eb3c95c66c91bbfac.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 414, + 547, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 414, + 547, + 561 + ], + "spans": [ + { + "bbox": [ + 304, + 414, + 547, + 561 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 414, + 547, + 561 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{E}}^O = \\{\\tilde{e}_p^O\\}_{p\\in \\mathcal{P}}\\subset \\mathbb{R}^d" + }, + { + "bbox": [ + 304, + 414, + 547, + 561 + ], + "type": "text", + "content": " denotes the proposal embeddings extracted by " + }, + { + "bbox": [ + 304, + 414, + 547, + 561 + ], + "type": "inline_equation", + "content": "\\nu" + }, + { + "bbox": [ + 304, + 414, + 547, + 561 + ], + "type": "text", + "content": ". Naturally, the quality of " + }, + { + "bbox": [ + 304, + 414, + 547, + 561 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{E}}^O" + }, + { + "bbox": [ + 304, + 414, + 547, + 561 + ], + "type": "text", + "content": " affects the accuracy of " + }, + { + "bbox": [ + 304, + 414, + 547, + 561 + ], + "type": "inline_equation", + "content": "P_{\\mathcal{C}}^{O}(p,c)" + }, + { + "bbox": [ + 304, + 414, + 547, + 561 + ], + "type": "text", + "content": " to a large extent. However, current approaches only yield sub-optimal " + }, + { + "bbox": [ + 304, + 414, + 547, + 561 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{E}}^O" + }, + { + "bbox": [ + 304, + 414, + 547, + 561 + ], + "type": "text", + "content": " due to the dilemma between information comprehensiveness and less noise. For example, when non-square proposal regions are directly passed to CLIP, the center crop operation in " + }, + { + "bbox": [ + 304, + 414, + 547, + 561 + ], + "type": "inline_equation", + "content": "\\nu" + }, + { + "bbox": [ + 304, + 414, + 547, + 561 + ], + "type": "text", + "content": " will cut out informative parts of an object, leading to incomplete structural knowledge about the object. On the other hand, if the proposals are squared or enlarged, the proposal regions will contain more ambient contexts, which may corrupt the proposal embeddings." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 561, + 547, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 561, + 547, + 669 + ], + "spans": [ + { + "bbox": [ + 304, + 561, + 547, + 669 + ], + "type": "text", + "content": "To acquire more accurate " + }, + { + "bbox": [ + 304, + 561, + 547, + 669 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{E}}^O" + }, + { + "bbox": [ + 304, + 561, + 547, + 669 + ], + "type": "text", + "content": ", we propose an Object-Aware Knowledge Extraction (OAKE) module, where the proposals are first transformed and then encoded with a modified version of " + }, + { + "bbox": [ + 304, + 561, + 547, + 669 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 304, + 561, + 547, + 669 + ], + "type": "text", + "content": ". Given a proposal " + }, + { + "bbox": [ + 304, + 561, + 547, + 669 + ], + "type": "inline_equation", + "content": "p \\in \\mathcal{P}" + }, + { + "bbox": [ + 304, + 561, + 547, + 669 + ], + "type": "text", + "content": ", the transformed proposal " + }, + { + "bbox": [ + 304, + 561, + 547, + 669 + ], + "type": "inline_equation", + "content": "p'" + }, + { + "bbox": [ + 304, + 561, + 547, + 669 + ], + "type": "text", + "content": " is a square with side length " + }, + { + "bbox": [ + 304, + 561, + 547, + 669 + ], + "type": "inline_equation", + "content": "s = \\sqrt{r \\times p_h \\times p_w}" + }, + { + "bbox": [ + 304, + 561, + 547, + 669 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 561, + 547, + 669 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 304, + 561, + 547, + 669 + ], + "type": "text", + "content": " is a constant scale ratio, " + }, + { + "bbox": [ + 304, + 561, + 547, + 669 + ], + "type": "inline_equation", + "content": "p_h" + }, + { + "bbox": [ + 304, + 561, + 547, + 669 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 561, + 547, + 669 + ], + "type": "inline_equation", + "content": "p_w" + }, + { + "bbox": [ + 304, + 561, + 547, + 669 + ], + "type": "text", + "content": " are the height and width of " + }, + { + "bbox": [ + 304, + 561, + 547, + 669 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 561, + 547, + 669 + ], + "type": "text", + "content": ". The center of " + }, + { + "bbox": [ + 304, + 561, + 547, + 669 + ], + "type": "inline_equation", + "content": "p'" + }, + { + "bbox": [ + 304, + 561, + 547, + 669 + ], + "type": "text", + "content": " is initially the same as " + }, + { + "bbox": [ + 304, + 561, + 547, + 669 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 561, + 547, + 669 + ], + "type": "text", + "content": " but may be translated if " + }, + { + "bbox": [ + 304, + 561, + 547, + 669 + ], + "type": "inline_equation", + "content": "p'" + }, + { + "bbox": [ + 304, + 561, + 547, + 669 + ], + "type": "text", + "content": " exceeds the image boundaries. All transformed proposals constitute " + }, + { + "bbox": [ + 304, + 561, + 547, + 669 + ], + "type": "inline_equation", + "content": "\\mathcal{P}'" + }, + { + "bbox": [ + 304, + 561, + 547, + 669 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 669, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 669, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 669, + 547, + 713 + ], + "type": "text", + "content": "While " + }, + { + "bbox": [ + 304, + 669, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{P}'" + }, + { + "bbox": [ + 304, + 669, + 547, + 713 + ], + "type": "text", + "content": " contributes to the comprehensiveness of object knowledge, it also includes surrounding distractors. To further suppress the contextual noise, we introduce an [OBJ] token into " + }, + { + "bbox": [ + 304, + 669, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 304, + 669, + 547, + 713 + ], + "type": "text", + "content": " and substitute original attention layers with" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11189" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 152 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 152 + ], + "type": "text", + "content": "masked attention layers. The modified version of " + }, + { + "bbox": [ + 47, + 72, + 287, + 152 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 47, + 72, + 287, + 152 + ], + "type": "text", + "content": " is denoted as " + }, + { + "bbox": [ + 47, + 72, + 287, + 152 + ], + "type": "inline_equation", + "content": "\\mathcal{V}'" + }, + { + "bbox": [ + 47, + 72, + 287, + 152 + ], + "type": "text", + "content": ". The input of " + }, + { + "bbox": [ + 47, + 72, + 287, + 152 + ], + "type": "inline_equation", + "content": "\\mathcal{V}'" + }, + { + "bbox": [ + 47, + 72, + 287, + 152 + ], + "type": "text", + "content": " is a set of image regions " + }, + { + "bbox": [ + 47, + 72, + 287, + 152 + ], + "type": "inline_equation", + "content": "\\mathbf{I}^{\\mathcal{O}} = \\{\\mathbf{I}_{p'}\\}_{p' \\in \\mathcal{P}'}" + }, + { + "bbox": [ + 47, + 72, + 287, + 152 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 72, + 287, + 152 + ], + "type": "inline_equation", + "content": "p'" + }, + { + "bbox": [ + 47, + 72, + 287, + 152 + ], + "type": "text", + "content": " is a transformed proposal and " + }, + { + "bbox": [ + 47, + 72, + 287, + 152 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{p'}" + }, + { + "bbox": [ + 47, + 72, + 287, + 152 + ], + "type": "text", + "content": " is the cropped image region of " + }, + { + "bbox": [ + 47, + 72, + 287, + 152 + ], + "type": "inline_equation", + "content": "p'" + }, + { + "bbox": [ + 47, + 72, + 287, + 152 + ], + "type": "text", + "content": ". Same as " + }, + { + "bbox": [ + 47, + 72, + 287, + 152 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 47, + 72, + 287, + 152 + ], + "type": "text", + "content": ", each " + }, + { + "bbox": [ + 47, + 72, + 287, + 152 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{p'}" + }, + { + "bbox": [ + 47, + 72, + 287, + 152 + ], + "type": "text", + "content": " is first mapped to a sequence of tokens " + }, + { + "bbox": [ + 47, + 72, + 287, + 152 + ], + "type": "inline_equation", + "content": "\\mathbf{X} \\in \\mathbb{R}^{N_x \\times d_x}" + }, + { + "bbox": [ + 47, + 72, + 287, + 152 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 72, + 287, + 152 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{1:N_x - 1}" + }, + { + "bbox": [ + 47, + 72, + 287, + 152 + ], + "type": "text", + "content": " are the patch tokens of " + }, + { + "bbox": [ + 47, + 72, + 287, + 152 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{p'}" + }, + { + "bbox": [ + 47, + 72, + 287, + 152 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 72, + 287, + 152 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{N_x}" + }, + { + "bbox": [ + 47, + 72, + 287, + 152 + ], + "type": "text", + "content": " is the [CLS] token. We then augment " + }, + { + "bbox": [ + 47, + 72, + 287, + 152 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 47, + 72, + 287, + 152 + ], + "type": "text", + "content": " with an [OBJ] token:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 102, + 154, + 287, + 168 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 154, + 287, + 168 + ], + "spans": [ + { + "bbox": [ + 102, + 154, + 287, + 168 + ], + "type": "interline_equation", + "content": "\\mathbf {X} ^ {\\prime} = [ \\mathbf {X}; x _ {[ \\mathrm {O B J} ]} ] \\in \\mathbb {R} ^ {(N _ {x} + 1) \\times d _ {x}}. \\tag {9}", + "image_path": "22d5c1e3b6c70793a84647b5b0ccc889b8bc2e543cfebf4b1f57ad570ac63d43.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 172, + 287, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 172, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 47, + 172, + 287, + 217 + ], + "type": "text", + "content": "Since the [OBJ] token serves a similar purpose as the [CLS] token, we initialize " + }, + { + "bbox": [ + 47, + 172, + 287, + 217 + ], + "type": "inline_equation", + "content": "x_{[\\mathrm{OBJ}]} = \\mathbf{X}_{N_x}" + }, + { + "bbox": [ + 47, + 172, + 287, + 217 + ], + "type": "text", + "content": ". To regulate the interaction between [OBJ] and the other tokens, we construct a mask " + }, + { + "bbox": [ + 47, + 172, + 287, + 217 + ], + "type": "inline_equation", + "content": "m \\in \\mathbb{R}^{N_x}" + }, + { + "bbox": [ + 47, + 172, + 287, + 217 + ], + "type": "text", + "content": ", such that:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 220, + 287, + 249 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 220, + 287, + 249 + ], + "spans": [ + { + "bbox": [ + 61, + 220, + 287, + 249 + ], + "type": "interline_equation", + "content": "m _ {i} = \\left\\{ \\begin{array}{l l} \\mathbb {1} \\left\\{\\text {t h e} i \\text {- t h p a t c h o v e r l a p s w i t h} p \\right\\}, & i < N _ {x} \\\\ 0, & i = N _ {x} \\end{array} \\right. \\tag {10}", + "image_path": "5aa93aa8d854b9b1bec7fa3811eb66a21159d6b32bd9a96e155a6f061bce09ab.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 253, + 287, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 253, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 47, + 253, + 287, + 308 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 253, + 287, + 308 + ], + "type": "inline_equation", + "content": "\\mathbb{1}\\{\\cdot\\}" + }, + { + "bbox": [ + 47, + 253, + 287, + 308 + ], + "type": "text", + "content": " is the indicator function. Intuitively, eq. (10) means that [OBJ] only attends to the patch tokens that are covered by the original proposal " + }, + { + "bbox": [ + 47, + 253, + 287, + 308 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 47, + 253, + 287, + 308 + ], + "type": "text", + "content": ". To maintain the original attentions among " + }, + { + "bbox": [ + 47, + 253, + 287, + 308 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 47, + 253, + 287, + 308 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 47, + 253, + 287, + 308 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 47, + 253, + 287, + 308 + ], + "type": "text", + "content": ", our attention mask " + }, + { + "bbox": [ + 47, + 253, + 287, + 308 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 47, + 253, + 287, + 308 + ], + "type": "text", + "content": " is constructed as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 66, + 310, + 287, + 335 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 310, + 287, + 335 + ], + "spans": [ + { + "bbox": [ + 66, + 310, + 287, + 335 + ], + "type": "interline_equation", + "content": "\\mathbf {M} = \\left[ \\begin{array}{c c} \\mathbf {1} ^ {N _ {x} \\times N _ {x}} & \\mathbf {0} ^ {N _ {x}} \\\\ m & 1 \\end{array} \\right] \\in \\{0, 1 \\} ^ {(N _ {x} + 1) \\times (N _ {x} + 1)}. \\tag {11}", + "image_path": "b1c1f2a190c66fff8cb45cfe7cfca874c07e7b216625a97335331d2c95483f3d.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 338, + 287, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 338, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 47, + 338, + 287, + 361 + ], + "type": "text", + "content": "Suppose " + }, + { + "bbox": [ + 47, + 338, + 287, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 47, + 338, + 287, + 361 + ], + "type": "text", + "content": " has " + }, + { + "bbox": [ + 47, + 338, + 287, + 361 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 47, + 338, + 287, + 361 + ], + "type": "text", + "content": " attention layers (ignoring FFNs), " + }, + { + "bbox": [ + 47, + 338, + 287, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{V}'" + }, + { + "bbox": [ + 47, + 338, + 287, + 361 + ], + "type": "text", + "content": " is defined as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 364, + 287, + 403 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 364, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 56, + 364, + 287, + 403 + ], + "type": "interline_equation", + "content": "\\mathbf {X} ^ {l} = \\left\\{ \\begin{array}{l l} \\sigma \\left(\\log \\mathbf {M} + \\mathbf {Q} ^ {l} \\left(\\mathbf {K} ^ {l}\\right) ^ {\\top}\\right) \\mathbf {V} ^ {l} + \\mathbf {X} ^ {l - 1}, & 0 < l \\leq L \\\\ \\mathbf {X} ^ {\\prime}, & l = 0 \\end{array} \\right. \\tag {12}", + "image_path": "cf148250ccc1e5b2ba6e111cd35ca391b45333f297a2e12712bc2f2748130114.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 404, + 287, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 404, + 287, + 468 + ], + "spans": [ + { + "bbox": [ + 47, + 404, + 287, + 468 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 404, + 287, + 468 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}^l" + }, + { + "bbox": [ + 47, + 404, + 287, + 468 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 404, + 287, + 468 + ], + "type": "inline_equation", + "content": "\\mathbf{K}^l" + }, + { + "bbox": [ + 47, + 404, + 287, + 468 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 404, + 287, + 468 + ], + "type": "inline_equation", + "content": "\\mathbf{V}^l" + }, + { + "bbox": [ + 47, + 404, + 287, + 468 + ], + "type": "text", + "content": " are linear transformations of " + }, + { + "bbox": [ + 47, + 404, + 287, + 468 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^{l-1}" + }, + { + "bbox": [ + 47, + 404, + 287, + 468 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 404, + 287, + 468 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 47, + 404, + 287, + 468 + ], + "type": "text", + "content": " is Softmax function. Finally, instead of [CLS], we take the output of [OBJ] as the proposal embedding, i.e., " + }, + { + "bbox": [ + 47, + 404, + 287, + 468 + ], + "type": "inline_equation", + "content": "\\tilde{e}_p^O = \\mathbf{X}_{N_x+1}^L" + }, + { + "bbox": [ + 47, + 404, + 287, + 468 + ], + "type": "text", + "content": ". Iterating over the proposals " + }, + { + "bbox": [ + 47, + 404, + 287, + 468 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 47, + 404, + 287, + 468 + ], + "type": "text", + "content": ", we obtain a set of accurate proposal embeddings " + }, + { + "bbox": [ + 47, + 404, + 287, + 468 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{E}}^O" + }, + { + "bbox": [ + 47, + 404, + 287, + 468 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 474, + 206, + 485 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 474, + 206, + 485 + ], + "spans": [ + { + "bbox": [ + 47, + 474, + 206, + 485 + ], + "type": "text", + "content": "4.4. Global and Block Distillation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 491, + 287, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 491, + 287, + 575 + ], + "spans": [ + { + "bbox": [ + 47, + 491, + 287, + 575 + ], + "type": "text", + "content": "While the object distillation module " + }, + { + "bbox": [ + 47, + 491, + 287, + 575 + ], + "type": "inline_equation", + "content": "\\mathcal{M}^O" + }, + { + "bbox": [ + 47, + 491, + 287, + 575 + ], + "type": "text", + "content": " aligns the proposal embeddings " + }, + { + "bbox": [ + 47, + 491, + 287, + 575 + ], + "type": "inline_equation", + "content": "\\mathcal{E}^O" + }, + { + "bbox": [ + 47, + 491, + 287, + 575 + ], + "type": "text", + "content": " with the optimized CLIP embeddings " + }, + { + "bbox": [ + 47, + 491, + 287, + 575 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{E}}^O" + }, + { + "bbox": [ + 47, + 491, + 287, + 575 + ], + "type": "text", + "content": ", the detector lacks a comprehensive understanding of the relation between different proposals. Therefore, we propose a global distillation module " + }, + { + "bbox": [ + 47, + 491, + 287, + 575 + ], + "type": "inline_equation", + "content": "\\mathcal{M}^G" + }, + { + "bbox": [ + 47, + 491, + 287, + 575 + ], + "type": "text", + "content": ", which transfers the knowledge of the entire image from the CLIP visual encoder " + }, + { + "bbox": [ + 47, + 491, + 287, + 575 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 47, + 491, + 287, + 575 + ], + "type": "text", + "content": " to the detector:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 105, + 580, + 228, + 594 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 580, + 228, + 594 + ], + "spans": [ + { + "bbox": [ + 105, + 580, + 228, + 594 + ], + "type": "interline_equation", + "content": "e ^ {G} = f (\\operatorname {G A P} \\left(\\mathbf {F} _ {6}\\right)) \\in \\mathbb {R} ^ {d},", + "image_path": "eec1a985aaa13b428f0a91a5a6f05928175d0dfc2fc83b4e51e009905ec9a466.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 107, + 595, + 194, + 609 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 595, + 194, + 609 + ], + "spans": [ + { + "bbox": [ + 107, + 595, + 194, + 609 + ], + "type": "interline_equation", + "content": "\\tilde {e} ^ {G} = \\mathcal {V} (\\mathbf {I}) \\in \\mathbb {R} ^ {d},", + "image_path": "7fbd0d637e9e88bcde9d994eb581c6a2f54705e50d8672c6d42c2bf6cfff453d.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 610, + 197, + 624 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 610, + 197, + 624 + ], + "spans": [ + { + "bbox": [ + 105, + 610, + 197, + 624 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {G} = \\mathcal {L} _ {1} (e ^ {G}, \\tilde {e} ^ {G}),", + "image_path": "c881ad5a091045a5d5a18e3f26db38d3a10b9b952fde4c992c09a20dce57dccf.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 629, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 629, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 47, + 629, + 287, + 653 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 629, + 287, + 653 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 47, + 629, + 287, + 653 + ], + "type": "text", + "content": " is a linear transformation function, " + }, + { + "bbox": [ + 47, + 629, + 287, + 653 + ], + "type": "inline_equation", + "content": "\\mathrm{GAP}(\\cdot)" + }, + { + "bbox": [ + 47, + 629, + 287, + 653 + ], + "type": "text", + "content": " is the global average pooling, and " + }, + { + "bbox": [ + 47, + 629, + 287, + 653 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 47, + 629, + 287, + 653 + ], + "type": "text", + "content": " is the input image." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "text", + "content": "Due to the existence of Human Reporting Bias [31], CLIP is prone to ignore non-salient information in an image, e.g., the background or prominent attributes of objects. Such information may be valuable for dense prediction tasks like detection [13]. Thus, we propose a block" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": "distillation module " + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "\\mathcal{M}^B" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": " to complement the missing knowledge in " + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "\\mathcal{M}^G" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": ". We evenly divide the input image into several blocks " + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "\\mathcal{B}\\subset \\mathbb{R}^4" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": " via a partition function " + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "g(\\cdot)" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": " and denote the corresponding image regions as " + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "\\mathbf{I}^{\\mathcal{B}}" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": ". The size of each block is fixed to " + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "R\\times R" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": " denotes the input resolution of " + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": ". In this way, the resize and center crop operations in " + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": " will not take effect, thus avoiding information loss when using " + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": " to encode the block embeddings " + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{E}}^{B} = \\{\\tilde{e}_{b}^{B}\\}_{b\\in \\mathcal{B}}\\subset \\mathbb{R}^{d}" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 167, + 545, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 167, + 545, + 224 + ], + "spans": [ + { + "bbox": [ + 304, + 167, + 545, + 224 + ], + "type": "text", + "content": "On the student side, we apply block pooling and a block head to extract the block embeddings " + }, + { + "bbox": [ + 304, + 167, + 545, + 224 + ], + "type": "inline_equation", + "content": "\\mathcal{E}^B = \\{e_b^O\\}_{b\\in \\mathcal{B}}\\subset" + }, + { + "bbox": [ + 304, + 167, + 545, + 224 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^d" + }, + { + "bbox": [ + 304, + 167, + 545, + 224 + ], + "type": "text", + "content": " . The block pooling is a combination of the block partition function " + }, + { + "bbox": [ + 304, + 167, + 545, + 224 + ], + "type": "inline_equation", + "content": "g(\\cdot)" + }, + { + "bbox": [ + 304, + 167, + 545, + 224 + ], + "type": "text", + "content": " and RoI Align. The loss of our proposed block distillation is defined as:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 388, + 232, + 545, + 247 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 232, + 545, + 247 + ], + "spans": [ + { + "bbox": [ + 388, + 232, + 545, + 247 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {B} = \\mathcal {L} _ {1} \\left(\\mathcal {E} ^ {B}, \\tilde {\\mathcal {E}} ^ {B}\\right). \\tag {13}", + "image_path": "4dafc68aad3654aee9a902a3ef12b7e20f8ae905e5ccdca857b772da3e0814cc.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 256, + 545, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 256, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 304, + 256, + 545, + 304 + ], + "type": "text", + "content": "Compared with " + }, + { + "bbox": [ + 304, + 256, + 545, + 304 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^G" + }, + { + "bbox": [ + 304, + 256, + 545, + 304 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 256, + 545, + 304 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^B" + }, + { + "bbox": [ + 304, + 256, + 545, + 304 + ], + "type": "text", + "content": " distills the knowledge of each block with the same weights. Therefore, the ignored information in " + }, + { + "bbox": [ + 304, + 256, + 545, + 304 + ], + "type": "inline_equation", + "content": "\\mathcal{M}^G" + }, + { + "bbox": [ + 304, + 256, + 545, + 304 + ], + "type": "text", + "content": " can be compensated by " + }, + { + "bbox": [ + 304, + 256, + 545, + 304 + ], + "type": "inline_equation", + "content": "\\mathcal{M}^B" + }, + { + "bbox": [ + 304, + 256, + 545, + 304 + ], + "type": "text", + "content": ". Note that neither " + }, + { + "bbox": [ + 304, + 256, + 545, + 304 + ], + "type": "inline_equation", + "content": "\\mathcal{M}^G" + }, + { + "bbox": [ + 304, + 256, + 545, + 304 + ], + "type": "text", + "content": " nor " + }, + { + "bbox": [ + 304, + 256, + 545, + 304 + ], + "type": "inline_equation", + "content": "\\mathcal{M}^B" + }, + { + "bbox": [ + 304, + 256, + 545, + 304 + ], + "type": "text", + "content": " is used during inference." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 314, + 447, + 325 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 314, + 447, + 325 + ], + "spans": [ + { + "bbox": [ + 305, + 314, + 447, + 325 + ], + "type": "text", + "content": "4.5. Pseudo Label Generation" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 332, + 545, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 332, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 304, + 332, + 545, + 399 + ], + "type": "text", + "content": "To investigate the performance of our OADP framework under the G-OVD benchmark (refer to sec. 3), we propose to generate pseudo labels with our modified " + }, + { + "bbox": [ + 304, + 332, + 545, + 399 + ], + "type": "inline_equation", + "content": "\\nu'" + }, + { + "bbox": [ + 304, + 332, + 545, + 399 + ], + "type": "text", + "content": ". Given a proposal " + }, + { + "bbox": [ + 304, + 332, + 545, + 399 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 332, + 545, + 399 + ], + "type": "text", + "content": ", we first extract the proposal embedding " + }, + { + "bbox": [ + 304, + 332, + 545, + 399 + ], + "type": "inline_equation", + "content": "\\tilde{e}_p^O" + }, + { + "bbox": [ + 304, + 332, + 545, + 399 + ], + "type": "text", + "content": " as described in sec. 4.3. Then, the probability of " + }, + { + "bbox": [ + 304, + 332, + 545, + 399 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 332, + 545, + 399 + ], + "type": "text", + "content": " belonging to category " + }, + { + "bbox": [ + 304, + 332, + 545, + 399 + ], + "type": "inline_equation", + "content": "c \\in \\mathcal{C}" + }, + { + "bbox": [ + 304, + 332, + 545, + 399 + ], + "type": "text", + "content": " is given as:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 357, + 407, + 545, + 434 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 357, + 407, + 545, + 434 + ], + "spans": [ + { + "bbox": [ + 357, + 407, + 545, + 434 + ], + "type": "interline_equation", + "content": "l ^ {\\mathrm {P L}} (p, c) = \\frac {\\tilde {e} _ {p} ^ {O} \\cdot t _ {c}}{\\| \\tilde {e} _ {p} ^ {O} \\| \\cdot \\| t _ {c} \\|}, \\tag {14}", + "image_path": "39f4362adcd8a8b51ec194163a07a15b75b3be84b230a548b785e62008994af5.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 354, + 435, + 545, + 467 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 435, + 545, + 467 + ], + "spans": [ + { + "bbox": [ + 354, + 435, + 545, + 467 + ], + "type": "interline_equation", + "content": "P _ {\\mathcal {C}} ^ {\\mathrm {P L}} (p, c) = \\frac {\\exp \\left(l ^ {\\mathrm {P L}} (p , c)\\right)}{\\sum_ {c ^ {\\prime} \\in C} \\exp \\left(l ^ {\\mathrm {P L}} \\left(p , c ^ {\\prime}\\right)\\right)}, \\tag {15}", + "image_path": "5b1dc76d67dc6d2b8b30696f0ca44a24e4383415e15e351df4a30d36e41a8755.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 477, + 545, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 477, + 545, + 512 + ], + "spans": [ + { + "bbox": [ + 304, + 477, + 545, + 512 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 477, + 545, + 512 + ], + "type": "inline_equation", + "content": "t_c" + }, + { + "bbox": [ + 304, + 477, + 545, + 512 + ], + "type": "text", + "content": " is the category embedding of " + }, + { + "bbox": [ + 304, + 477, + 545, + 512 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 477, + 545, + 512 + ], + "type": "text", + "content": ". Since " + }, + { + "bbox": [ + 304, + 477, + 545, + 512 + ], + "type": "inline_equation", + "content": "P_{\\mathcal{C}}^{\\mathrm{PL}}(p,c)" + }, + { + "bbox": [ + 304, + 477, + 545, + 512 + ], + "type": "text", + "content": " does not reflect the localization quality of " + }, + { + "bbox": [ + 304, + 477, + 545, + 512 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 477, + 545, + 512 + ], + "type": "text", + "content": " [12], we define the confidence score " + }, + { + "bbox": [ + 304, + 477, + 545, + 512 + ], + "type": "inline_equation", + "content": "S_{\\mathcal{C}}(p,c)" + }, + { + "bbox": [ + 304, + 477, + 545, + 512 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 366, + 521, + 545, + 535 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 366, + 521, + 545, + 535 + ], + "spans": [ + { + "bbox": [ + 366, + 521, + 545, + 535 + ], + "type": "interline_equation", + "content": "S _ {\\mathcal {C}} (p, c) = P _ {\\mathcal {C}} ^ {\\mathrm {P L}} (p, c) ^ {\\gamma} \\cdot o _ {p} ^ {(1 - \\gamma)}, \\tag {16}", + "image_path": "fe2957acb4dd07ec4ded6f8be773d2749ec0c86e8b6920a76f1c3e8063b680b7.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 544, + 545, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 544, + 545, + 604 + ], + "spans": [ + { + "bbox": [ + 304, + 544, + 545, + 604 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 544, + 545, + 604 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 304, + 544, + 545, + 604 + ], + "type": "text", + "content": " includes both base and novel categories, " + }, + { + "bbox": [ + 304, + 544, + 545, + 604 + ], + "type": "inline_equation", + "content": "o_p\\in [0,1]" + }, + { + "bbox": [ + 304, + 544, + 545, + 604 + ], + "type": "text", + "content": " is the objectness score of " + }, + { + "bbox": [ + 304, + 544, + 545, + 604 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 544, + 545, + 604 + ], + "type": "text", + "content": " ,and " + }, + { + "bbox": [ + 304, + 544, + 545, + 604 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 304, + 544, + 545, + 604 + ], + "type": "text", + "content": " is a constant balancing factor. " + }, + { + "bbox": [ + 304, + 544, + 545, + 604 + ], + "type": "inline_equation", + "content": "S_{\\mathcal{C}}(p,c)" + }, + { + "bbox": [ + 304, + 544, + 545, + 604 + ], + "type": "text", + "content": " reflects the probability that " + }, + { + "bbox": [ + 304, + 544, + 545, + 604 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 544, + 545, + 604 + ], + "type": "text", + "content": " precisely locates an instance of category " + }, + { + "bbox": [ + 304, + 544, + 545, + 604 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 544, + 545, + 604 + ], + "type": "text", + "content": " . Finally, we apply class-wise NMS on the novel categories to obtain the pseudo labels." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 605, + 545, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 653 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 653 + ], + "type": "text", + "content": "Note that the Softmax operation in eq. (15) is performed over all categories " + }, + { + "bbox": [ + 304, + 605, + 545, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 304, + 605, + 545, + 653 + ], + "type": "text", + "content": ", even though the pseudo labels do not include instances on base categories. Such a design effectively suppresses false positives in the pseudo labels." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "The proposals to be labeled are extracted via an RPN model pretrained on " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^T" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": ", which only contains annotations of the base categories. ViLD [12] demonstrates that the generalization ability of " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": " is strong enough to recall most objects of the novel categories." + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11190" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 286, + 290 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 286, + 290 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 286, + 290 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 286, + 290 + ], + "type": "table", + "html": "
BenchmarkMethod\\( \\mathrm{mAP}_{50}^{N} \\)\\( \\mathrm{mAP}_{50}^{B} \\)\\( \\mathrm{mAP}_{50} \\)
ZSDSB [1]0.329.224.9
DELO [56]3.413.813.0
PL [34]4.135.927.9
V-OVDViLD [12]27.659.551.3
RegionCLIP* [53]14.252.842.7
OADP (Ours)30.053.347.2
C-OVDOVR-CNN [49]22.846.039.9
HierKD [28]20.351.343.2
RegionCLIP [53]26.854.847.5
LocOV [3]28.651.345.7
PB-OVD [11]29.144.440.4
G-OVDOV-DETR [48]29.461.052.7
VL-PLM [52]32.354.048.3
OADP (Ours)35.655.850.5
WSDWSDNN [2]19.719.619.6
Cap2Det [47]20.320.120.1
WS-OVDDetic [55]27.847.145.0
", + "image_path": "6cdb26dc0bf3ce40b15379e5541a549e4b63d7fb7d48ff5c252859b05241e6fb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 378, + 286, + 442 + ], + "blocks": [ + { + "bbox": [ + 46, + 299, + 287, + 365 + ], + "lines": [ + { + "bbox": [ + 46, + 299, + 287, + 365 + ], + "spans": [ + { + "bbox": [ + 46, + 299, + 287, + 365 + ], + "type": "text", + "content": "Table 2. Comparison with other state-of-the-art methods on the OV-COCO dataset. Methods are grouped by the benchmark they use. \"ZSD\" and \"WSD\" stand for Zero-Shot Detection and Weakly Supervised Detection. \"V-OVD\", \"C-OVD\", \"G-OVD\", and \"WS-OVD\" are introduced in sec. 3. \"RegionCLIP*\" indicates a model without refinement using COCO-Captions." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 378, + 286, + 442 + ], + "lines": [ + { + "bbox": [ + 50, + 378, + 286, + 442 + ], + "spans": [ + { + "bbox": [ + 50, + 378, + 286, + 442 + ], + "type": "table", + "html": "
MethodObject DetectionInstance Segmentation
APrAPcAPfAPAPrAPcAPfAP
ViLD [12]16.726.534.227.816.624.630.325.5
DetPro [8]20.827.832.428.419.825.628.925.9
OV-DETR [48]----17.425.032.526.6
OADP (Ours)21.928.432.028.721.726.329.026.6
", + "image_path": "769e17adecf90f81771d59279ee0357b4a0b532be19a41cdbfe85c1781a0a1b5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 451, + 286, + 472 + ], + "lines": [ + { + "bbox": [ + 46, + 451, + 286, + 472 + ], + "spans": [ + { + "bbox": [ + 46, + 451, + 286, + 472 + ], + "type": "text", + "content": "Table 3. Comparison with other state-of-the-art methods on the OV-LVIS dataset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 483, + 128, + 497 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 483, + 128, + 497 + ], + "spans": [ + { + "bbox": [ + 47, + 483, + 128, + 497 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 504, + 287, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 504, + 287, + 564 + ], + "spans": [ + { + "bbox": [ + 46, + 504, + 287, + 564 + ], + "type": "text", + "content": "In this section, we first introduce the detailed experiment setup, including the datasets, evaluation metrics, and implementation details. We then evaluate the performance of our proposed OADP framework and analyze the results compared to the state-of-the-art approaches." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 574, + 110, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 574, + 110, + 586 + ], + "spans": [ + { + "bbox": [ + 47, + 574, + 110, + 586 + ], + "type": "text", + "content": "5.1. Datasets" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "content": "Experiments are mainly conducted under the open-vocabulary COCO (OV-COCO) setting [49], where the MSCOco 2017 dataset [27] is manually divided into 48 base categories and 17 novel categories. The training dataset contains 107,761 images, and the validation dataset contains 4,836 images. We report the " + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{mAP}_{50}^{\\mathrm{N}}" + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{mAP}_{50}^{\\mathrm{B}}" + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{mAP}_{50}" + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "content": " metrics, i.e., the mAP at IoU threshold 0.5 for novel, base, and all categories. " + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{mAP}_{50}^{\\mathrm{N}}" + }, + { + "bbox": [ + 46, + 594, + 288, + 715 + ], + "type": "text", + "content": " is the main metric. Some experiments are conducted under the open-vocabulary LVIS (OV-LVIS) [12] setting, where the 337 rare categories in" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": "LVIS [14] are treated as novel categories, and the other 866 are base categories. Metrics for the OV-LVIS setting are " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "\\mathrm{AP_r}" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "\\mathrm{AP_c}" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "\\mathrm{AP_f}" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": ", and AP, i.e., the mAP for rare (novel), common, frequent, and all categories. Both object detection and instance segmentation metrics are reported." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 140, + 440, + 153 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 140, + 440, + 153 + ], + "spans": [ + { + "bbox": [ + 305, + 140, + 440, + 153 + ], + "type": "text", + "content": "5.2. Implementation Details" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 159, + 545, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 159, + 545, + 315 + ], + "spans": [ + { + "bbox": [ + 304, + 159, + 545, + 315 + ], + "type": "text", + "content": "Training is conducted on 8 V-100 GPUs with batch size 16 in total. We use stochastic gradient descent (SGD) optimizer with 0.02 initial learning rate, 0.9 momentum, and " + }, + { + "bbox": [ + 304, + 159, + 545, + 315 + ], + "type": "inline_equation", + "content": "2.5 \\times 10^{-5}" + }, + { + "bbox": [ + 304, + 159, + 545, + 315 + ], + "type": "text", + "content": " weight decay. The student backbone is ResNet50 [16]. Following DetPro [8], we adopt the ViT-B/32 CLIP [32] as the teacher and initialize the student backbone using SoCo [41]. The loss weights " + }, + { + "bbox": [ + 304, + 159, + 545, + 315 + ], + "type": "inline_equation", + "content": "w^{O}" + }, + { + "bbox": [ + 304, + 159, + 545, + 315 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 159, + 545, + 315 + ], + "type": "inline_equation", + "content": "w^{B}" + }, + { + "bbox": [ + 304, + 159, + 545, + 315 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 159, + 545, + 315 + ], + "type": "inline_equation", + "content": "w^{G}" + }, + { + "bbox": [ + 304, + 159, + 545, + 315 + ], + "type": "text", + "content": " are set to 0.5, 0.25, and 0.25 respectively. Under the OV-COCO setting, we train the detector for 40,000 iterations. At the " + }, + { + "bbox": [ + 304, + 159, + 545, + 315 + ], + "type": "inline_equation", + "content": "32,000^{\\mathrm{th}}" + }, + { + "bbox": [ + 304, + 159, + 545, + 315 + ], + "type": "text", + "content": " iteration, the learning rate is divided by 10. For OV-LVIS, we use 2x (24 epochs) training schedule, where the learning rate is divided by 10 at the " + }, + { + "bbox": [ + 304, + 159, + 545, + 315 + ], + "type": "inline_equation", + "content": "16^{\\mathrm{th}}" + }, + { + "bbox": [ + 304, + 159, + 545, + 315 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 159, + 545, + 315 + ], + "type": "inline_equation", + "content": "22^{\\mathrm{th}}" + }, + { + "bbox": [ + 304, + 159, + 545, + 315 + ], + "type": "text", + "content": " epochs." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 323, + 391, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 323, + 391, + 335 + ], + "spans": [ + { + "bbox": [ + 305, + 323, + 391, + 335 + ], + "type": "text", + "content": "5.3. Main Results" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 342, + 545, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 342, + 545, + 413 + ], + "spans": [ + { + "bbox": [ + 304, + 342, + 545, + 413 + ], + "type": "text", + "content": "We compare our OADP framework with the other state-of-the-art OVD methods. As described in sec. 3, we categorize existing OVD methods by the benchmark they belong to. For completeness, we include two related benchmarks: Zero-Shot Detection (ZSD) [1, 18, 24, 35, 51] and Weakly Supervised Detection (WSD) [2, 47]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 414, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 414, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 414, + 545, + 605 + ], + "type": "text", + "content": "Our method mainly focuses on the V-OVD benchmark and the G-OVD benchmark. As shown in tab. 2, our OADP framework achieves " + }, + { + "bbox": [ + 304, + 414, + 545, + 605 + ], + "type": "inline_equation", + "content": "30.0\\mathrm{mAP}_{50}^{\\mathrm{N}}" + }, + { + "bbox": [ + 304, + 414, + 545, + 605 + ], + "type": "text", + "content": " on the V-OVD benchmark. RegionCLIP* [53] uses CLIP [32] as the pretrained weight, thus adhering to the V-OVD benchmark. Some V-OVD methods [4, 20, 22, 29, 45] are not included because they rely on large-scale detection and image-text datasets and cannot be compared fairly. While C-OVD is not our primary concern, we include the corresponding methods for reference and report their performance when only COCO Captions [6] is available. Under such constraint, the performance of the caption-based methods is relatively lower than the V-OVD methods, even if additional caption data is used. Moreover, our OADP framework is perpendicular to the caption-based methods and has the potential to achieve higher performance using captions." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 606, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 606, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 606, + 545, + 714 + ], + "type": "text", + "content": "For the G-OVD benchmark, we generate pseudo labels for novel categories as described in sec. 4.5. The pseudo labels are then merged with the instance-level annotations for base categories. Training our OADP framework on the mixed dataset yields " + }, + { + "bbox": [ + 304, + 606, + 545, + 714 + ], + "type": "inline_equation", + "content": "35.6\\mathrm{mAP}_{50}^{\\mathrm{N}}" + }, + { + "bbox": [ + 304, + 606, + 545, + 714 + ], + "type": "text", + "content": ", surpassing the previous SOTA method VL-PLM [52] by " + }, + { + "bbox": [ + 304, + 606, + 545, + 714 + ], + "type": "inline_equation", + "content": "3.3\\mathrm{mAP}_{50}^{\\mathrm{N}}" + }, + { + "bbox": [ + 304, + 606, + 545, + 714 + ], + "type": "text", + "content": ". Since PromptDet [10] relies on an external dataset (LAION-400M [38]) and uses smaller image size " + }, + { + "bbox": [ + 304, + 606, + 545, + 714 + ], + "type": "inline_equation", + "content": "(640\\times 640)" + }, + { + "bbox": [ + 304, + 606, + 545, + 714 + ], + "type": "text", + "content": ", the result 26.6 " + }, + { + "bbox": [ + 304, + 606, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{mAP}_{50}^{\\mathrm{N}}" + }, + { + "bbox": [ + 304, + 606, + 545, + 714 + ], + "type": "text", + "content": " is not listed in tab. 2 for fairness." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11191" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 74, + 70, + 521, + 191 + ], + "blocks": [ + { + "bbox": [ + 74, + 70, + 521, + 191 + ], + "lines": [ + { + "bbox": [ + 74, + 70, + 521, + 191 + ], + "spans": [ + { + "bbox": [ + 74, + 70, + 521, + 191 + ], + "type": "table", + "html": "
Global DistillationBlock DistillationObject DistillationmAPNovel \\( \\mathrm{mAP}_{50} \\)\\( \\mathrm{mAP}_{75} \\)mAPBase \\( \\mathrm{mAP}_{50} \\)\\( \\mathrm{mAP}_{75} \\)mAPAll \\( \\mathrm{mAP}_{50} \\)\\( \\mathrm{mAP}_{75} \\)
✓13.3224.9912.3531.8750.2934.0327.0243.6728.36
13.5125.7212.3632.8251.8935.3127.7745.0429.31
✓14.5727.2513.1734.4553.6037.2029.2546.7131.06
✓15.4927.2315.2535.9955.9638.5730.6348.4532.47
✓✓13.5026.4912.5032.1951.2533.9427.3044.7828.33
✓✓15.4728.8014.6234.0854.2936.2829.2147.6230.61
✓✓15.9229.0115.6435.3055.4537.8830.2348.5332.06
✓✓✓16.2129.9515.4733.3353.2635.4728.8547.1730.24
", + "image_path": "f8fc86724805de1789226abfd70db7f1b1c6cebe9e0b8ba32107f890b9e19b38.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 230, + 289, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 230, + 289, + 338 + ], + "spans": [ + { + "bbox": [ + 46, + 230, + 289, + 338 + ], + "type": "text", + "content": "Tab. 3 shows the comparison between our method and the other state-of-the-art methods on the OV-LVIS dataset. Most of the methods under the OV-LVIS setting adhere to the V-OVD benchmark, so we conduct experiments on the V-OVD benchmark only. For the object detection task, our OADP framework achieves " + }, + { + "bbox": [ + 46, + 230, + 289, + 338 + ], + "type": "inline_equation", + "content": "21.9\\mathrm{AP_r}" + }, + { + "bbox": [ + 46, + 230, + 289, + 338 + ], + "type": "text", + "content": ", surpassing DetPro [8] by " + }, + { + "bbox": [ + 46, + 230, + 289, + 338 + ], + "type": "inline_equation", + "content": "1.1\\mathrm{AP_r}" + }, + { + "bbox": [ + 46, + 230, + 289, + 338 + ], + "type": "text", + "content": ". We also report the performance of the instance segmentation task, which achieves " + }, + { + "bbox": [ + 46, + 230, + 289, + 338 + ], + "type": "inline_equation", + "content": "21.7\\mathrm{AP_r}" + }, + { + "bbox": [ + 46, + 230, + 289, + 338 + ], + "type": "text", + "content": " and is " + }, + { + "bbox": [ + 46, + 230, + 289, + 338 + ], + "type": "inline_equation", + "content": "1.9\\mathrm{AP_r}" + }, + { + "bbox": [ + 46, + 230, + 289, + 338 + ], + "type": "text", + "content": " higher than the previous SOTA method." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 346, + 141, + 360 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 346, + 141, + 360 + ], + "spans": [ + { + "bbox": [ + 47, + 346, + 141, + 360 + ], + "type": "text", + "content": "5.4. Ablation Study" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 365, + 288, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 365, + 288, + 402 + ], + "spans": [ + { + "bbox": [ + 46, + 365, + 288, + 402 + ], + "type": "text", + "content": "We conduct ablation studies on the OV-COCO dataset to evaluate the effectiveness of each component in our proposed OADP framework." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 407, + 289, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 407, + 289, + 660 + ], + "spans": [ + { + "bbox": [ + 46, + 407, + 289, + 660 + ], + "type": "text", + "content": "OADP. Tab. 4 shows the effectiveness of the distillation modules in our OADP framework. The first row is our re-implemented ViLD-ensemble [12]. Due to the expensive training cost of ViLD, the performance " + }, + { + "bbox": [ + 46, + 407, + 289, + 660 + ], + "type": "inline_equation", + "content": "24.99\\mathrm{mAP}_{50}^{\\mathrm{N}}" + }, + { + "bbox": [ + 46, + 407, + 289, + 660 + ], + "type": "text", + "content": " is far below the official " + }, + { + "bbox": [ + 46, + 407, + 289, + 660 + ], + "type": "inline_equation", + "content": "27.60\\mathrm{mAP}_{50}^{\\mathrm{N}}" + }, + { + "bbox": [ + 46, + 407, + 289, + 660 + ], + "type": "text", + "content": ". Nevertheless, with our proposed distillation pyramid, we can eventually surpass ViLD. The " + }, + { + "bbox": [ + 46, + 407, + 289, + 660 + ], + "type": "inline_equation", + "content": "2^{\\mathrm{nd}}" + }, + { + "bbox": [ + 46, + 407, + 289, + 660 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 407, + 289, + 660 + ], + "type": "inline_equation", + "content": "4^{\\mathrm{th}}" + }, + { + "bbox": [ + 46, + 407, + 289, + 660 + ], + "type": "text", + "content": " row in tab. 4 adds " + }, + { + "bbox": [ + 46, + 407, + 289, + 660 + ], + "type": "inline_equation", + "content": "\\mathcal{M}^{G}" + }, + { + "bbox": [ + 46, + 407, + 289, + 660 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 407, + 289, + 660 + ], + "type": "inline_equation", + "content": "\\mathcal{M}^{B}" + }, + { + "bbox": [ + 46, + 407, + 289, + 660 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 407, + 289, + 660 + ], + "type": "inline_equation", + "content": "\\mathcal{M}^{O}" + }, + { + "bbox": [ + 46, + 407, + 289, + 660 + ], + "type": "text", + "content": " to the baseline respectively. The global distillation module brings a " + }, + { + "bbox": [ + 46, + 407, + 289, + 660 + ], + "type": "inline_equation", + "content": "0.73\\mathrm{mAP}_{50}^{\\mathrm{N}}" + }, + { + "bbox": [ + 46, + 407, + 289, + 660 + ], + "type": "text", + "content": " gain, while the other two bring " + }, + { + "bbox": [ + 46, + 407, + 289, + 660 + ], + "type": "inline_equation", + "content": "2.26\\mathrm{mAP}_{50}^{\\mathrm{N}}" + }, + { + "bbox": [ + 46, + 407, + 289, + 660 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 407, + 289, + 660 + ], + "type": "inline_equation", + "content": "2.24\\mathrm{mAP}_{50}^{\\mathrm{N}}" + }, + { + "bbox": [ + 46, + 407, + 289, + 660 + ], + "type": "text", + "content": " gain. Note that by adding " + }, + { + "bbox": [ + 46, + 407, + 289, + 660 + ], + "type": "inline_equation", + "content": "\\mathcal{M}^{O}" + }, + { + "bbox": [ + 46, + 407, + 289, + 660 + ], + "type": "text", + "content": " to the baseline, we remove the original image head in ViLD-ensemble. Therefore, the " + }, + { + "bbox": [ + 46, + 407, + 289, + 660 + ], + "type": "inline_equation", + "content": "2.24\\mathrm{mAP}_{50}^{\\mathrm{N}}" + }, + { + "bbox": [ + 46, + 407, + 289, + 660 + ], + "type": "text", + "content": " gain results from the OAKE module instead of the distillation operation. The " + }, + { + "bbox": [ + 46, + 407, + 289, + 660 + ], + "type": "inline_equation", + "content": "5^{\\mathrm{th}}" + }, + { + "bbox": [ + 46, + 407, + 289, + 660 + ], + "type": "text", + "content": " row in tab. 4 adds " + }, + { + "bbox": [ + 46, + 407, + 289, + 660 + ], + "type": "inline_equation", + "content": "\\mathcal{M}^{G}" + }, + { + "bbox": [ + 46, + 407, + 289, + 660 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 407, + 289, + 660 + ], + "type": "inline_equation", + "content": "\\mathcal{M}^{B}" + }, + { + "bbox": [ + 46, + 407, + 289, + 660 + ], + "type": "text", + "content": " together. While the 26.49 performance is higher than sole " + }, + { + "bbox": [ + 46, + 407, + 289, + 660 + ], + "type": "inline_equation", + "content": "\\mathcal{M}^{G}" + }, + { + "bbox": [ + 46, + 407, + 289, + 660 + ], + "type": "text", + "content": ", it is slightly lower than the " + }, + { + "bbox": [ + 46, + 407, + 289, + 660 + ], + "type": "inline_equation", + "content": "27.25\\mathrm{mAP}_{50}^{\\mathrm{N}}" + }, + { + "bbox": [ + 46, + 407, + 289, + 660 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 46, + 407, + 289, + 660 + ], + "type": "inline_equation", + "content": "\\mathcal{M}^{B}" + }, + { + "bbox": [ + 46, + 407, + 289, + 660 + ], + "type": "text", + "content": ". However, along with the object distillation module " + }, + { + "bbox": [ + 46, + 407, + 289, + 660 + ], + "type": "inline_equation", + "content": "\\mathcal{M}^{O}" + }, + { + "bbox": [ + 46, + 407, + 289, + 660 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 407, + 289, + 660 + ], + "type": "inline_equation", + "content": "\\mathcal{M}^{G}" + }, + { + "bbox": [ + 46, + 407, + 289, + 660 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 407, + 289, + 660 + ], + "type": "inline_equation", + "content": "\\mathcal{M}^{B}" + }, + { + "bbox": [ + 46, + 407, + 289, + 660 + ], + "type": "text", + "content": " achieves " + }, + { + "bbox": [ + 46, + 407, + 289, + 660 + ], + "type": "inline_equation", + "content": "28.80\\mathrm{mAP}_{50}^{\\mathrm{N}}" + }, + { + "bbox": [ + 46, + 407, + 289, + 660 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 407, + 289, + 660 + ], + "type": "inline_equation", + "content": "29.01\\mathrm{mAP}_{50}^{\\mathrm{N}}" + }, + { + "bbox": [ + 46, + 407, + 289, + 660 + ], + "type": "text", + "content": ", suggesting that " + }, + { + "bbox": [ + 46, + 407, + 289, + 660 + ], + "type": "inline_equation", + "content": "\\mathcal{M}^{G}" + }, + { + "bbox": [ + 46, + 407, + 289, + 660 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 407, + 289, + 660 + ], + "type": "inline_equation", + "content": "\\mathcal{M}^{B}" + }, + { + "bbox": [ + 46, + 407, + 289, + 660 + ], + "type": "text", + "content": " have a similar function in transferring the global scene knowledge from CLIP to the detector. Using all three modules together, we achieve " + }, + { + "bbox": [ + 46, + 407, + 289, + 660 + ], + "type": "inline_equation", + "content": "29.95\\mathrm{mAP}_{50}^{\\mathrm{N}}" + }, + { + "bbox": [ + 46, + 407, + 289, + 660 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": "OAKE. We demonstrate the effectiveness of our OAKE module in tab. 5. Given the ground truth bounding boxes, we use different strategies to crop their image regions. (1) Baseline: " + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "inline_equation", + "content": "1 \\times" + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": " crop; (2) ViLD*: " + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "inline_equation", + "content": "1 \\times" + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "inline_equation", + "content": "1.5 \\times" + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": " crop; (3)" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 308, + 228, + 545, + 317 + ], + "blocks": [ + { + "bbox": [ + 46, + 198, + 547, + 220 + ], + "lines": [ + { + "bbox": [ + 46, + 198, + 547, + 220 + ], + "spans": [ + { + "bbox": [ + 46, + 198, + 547, + 220 + ], + "type": "text", + "content": "Table 4. Ablation study of the Global, Block, and Object Distillation modules in the OADP framework. The baseline is our re-implemented ViLD-ensemble model." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 228, + 545, + 317 + ], + "lines": [ + { + "bbox": [ + 308, + 228, + 545, + 317 + ], + "spans": [ + { + "bbox": [ + 308, + 228, + 545, + 317 + ], + "type": "table", + "html": "
MethodMacro PrecisionWeighted Precision
w/o OAKEw/ OAKEw/o OAKEw/ OAKE
Baseline58.08-62.04-
ViLD*63.36-65.91-
MBS61.7063.8364.8165.82
Fixed49.0764.5351.4969.75
Adaptive51.6466.0955.8568.68
", + "image_path": "b573387a6afe54729b22c19d599f979ad0adb590654393d0af27c60cdd3451bb.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 324, + 546, + 358 + ], + "lines": [ + { + "bbox": [ + 305, + 324, + 546, + 358 + ], + "spans": [ + { + "bbox": [ + 305, + 324, + 546, + 358 + ], + "type": "text", + "content": "Table 5. Ablation study of OAKE module. \"ViLD*\" indicates our re-implementation of multi-scale region embedding. \"MBS\", \"Fixed\", and \"Adaptive\" are three transforming strategies." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 369, + 545, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 369, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 304, + 369, + 545, + 514 + ], + "type": "text", + "content": "MBS: the minimum bounding square of the original bounding box; (4) Fixed: " + }, + { + "bbox": [ + 304, + 369, + 545, + 514 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 304, + 369, + 545, + 514 + ], + "type": "text", + "content": " bounding square; (5) Adaptive: adaptively enlarge the bounding square. For the above strategies, we use CLIP to directly extract embeddings for their image regions (\"w/o mask\"). Alternatively, we can utilize the modified CLIP visual encoder mentioned in sec. 4.3 (\"w/ mask\"). Finally, we classify these embeddings by calculating their similarity with category embeddings. To evaluate the performance, we compute \"Macro Precision\" (precision for each category independently with equal weights) and \"Weighted Precision\" (weights depending on the number of bounding boxes in each class)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 514, + 547, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 514, + 547, + 635 + ], + "spans": [ + { + "bbox": [ + 304, + 514, + 547, + 635 + ], + "type": "text", + "content": "As shown in the " + }, + { + "bbox": [ + 304, + 514, + 547, + 635 + ], + "type": "inline_equation", + "content": "1^{\\mathrm{st}}" + }, + { + "bbox": [ + 304, + 514, + 547, + 635 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 514, + 547, + 635 + ], + "type": "inline_equation", + "content": "3^{\\mathrm{rd}}" + }, + { + "bbox": [ + 304, + 514, + 547, + 635 + ], + "type": "text", + "content": " columns, \"Fixed\" and \"Adaptive\" strategies bring performance drops as they crop a larger bounding square compared to other strategies (e.g., \"MBS\") which may introduce additional surrounding distractors that confuse the CLIP visual encoder. However, with our object-aware CLIP visual encoder, the performances of the two strategies boost significantly and suppress others. It validates that the CLIP visual encoder can focus on the proposal object with our mask attention mechanism to extract accurate knowledge." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "content": "Pseudo Label. We follow VL-PLM [52] to adopt the COCO-ZS setting for our ablation studies of the pseudo label. Both mAP and the average per-image number of PLs (#PL) on novel categories are used as metrics to evaluate the quality of the pseudo label. The baseline method directly uses [CLS] token of CLIP visual encoder to extract" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "11192" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 71, + 70, + 169, + 135 + ], + "blocks": [ + { + "bbox": [ + 55, + 98, + 67, + 108 + ], + "lines": [ + { + "bbox": [ + 55, + 98, + 67, + 108 + ], + "spans": [ + { + "bbox": [ + 55, + 98, + 67, + 108 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 71, + 70, + 169, + 135 + ], + "lines": [ + { + "bbox": [ + 71, + 70, + 169, + 135 + ], + "spans": [ + { + "bbox": [ + 71, + 70, + 169, + 135 + ], + "type": "image", + "image_path": "35b4f78da566dc64c599b1e10fc62a60da20cc0abb4fab7f1e7f26bc86167a66.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 170, + 70, + 268, + 135 + ], + "blocks": [ + { + "bbox": [ + 170, + 70, + 268, + 135 + ], + "lines": [ + { + "bbox": [ + 170, + 70, + 268, + 135 + ], + "spans": [ + { + "bbox": [ + 170, + 70, + 268, + 135 + ], + "type": "image", + "image_path": "79c362e78c9985073a685a082e3656578966290e80990a17d3df561a9895a488.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 269, + 70, + 356, + 135 + ], + "blocks": [ + { + "bbox": [ + 269, + 70, + 356, + 135 + ], + "lines": [ + { + "bbox": [ + 269, + 70, + 356, + 135 + ], + "spans": [ + { + "bbox": [ + 269, + 70, + 356, + 135 + ], + "type": "image", + "image_path": "0690b25cafa5b8cecca7acb3ec5121def590defb004a492ed002e3ca50413d75.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 359, + 70, + 446, + 135 + ], + "blocks": [ + { + "bbox": [ + 359, + 70, + 446, + 135 + ], + "lines": [ + { + "bbox": [ + 359, + 70, + 446, + 135 + ], + "spans": [ + { + "bbox": [ + 359, + 70, + 446, + 135 + ], + "type": "image", + "image_path": "38e1a2e284f4805d2f556213ea90b11f59603d4a72110a57094ae10488f750a0.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 448, + 70, + 544, + 135 + ], + "blocks": [ + { + "bbox": [ + 448, + 70, + 544, + 135 + ], + "lines": [ + { + "bbox": [ + 448, + 70, + 544, + 135 + ], + "spans": [ + { + "bbox": [ + 448, + 70, + 544, + 135 + ], + "type": "image", + "image_path": "765ceed1ccebeaa508ec89ca131d7923f7719bd59c6afff9afd697df17041a56.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 71, + 137, + 169, + 200 + ], + "blocks": [ + { + "bbox": [ + 55, + 163, + 67, + 174 + ], + "lines": [ + { + "bbox": [ + 55, + 163, + 67, + 174 + ], + "spans": [ + { + "bbox": [ + 55, + 163, + 67, + 174 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 71, + 137, + 169, + 200 + ], + "lines": [ + { + "bbox": [ + 71, + 137, + 169, + 200 + ], + "spans": [ + { + "bbox": [ + 71, + 137, + 169, + 200 + ], + "type": "image", + "image_path": "6f5222ffa005bfa522e551d5c1be9dde1d18c991d05aff23b4d9ac8b0ceeda7b.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 170, + 137, + 268, + 200 + ], + "blocks": [ + { + "bbox": [ + 170, + 137, + 268, + 200 + ], + "lines": [ + { + "bbox": [ + 170, + 137, + 268, + 200 + ], + "spans": [ + { + "bbox": [ + 170, + 137, + 268, + 200 + ], + "type": "image", + "image_path": "260fbe17168e458edb23266bc33efa504d7b678d10111bb28346accd3a2b477e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 269, + 137, + 356, + 200 + ], + "blocks": [ + { + "bbox": [ + 269, + 137, + 356, + 200 + ], + "lines": [ + { + "bbox": [ + 269, + 137, + 356, + 200 + ], + "spans": [ + { + "bbox": [ + 269, + 137, + 356, + 200 + ], + "type": "image", + "image_path": "d4189346658dbe9318f68e1471279fd1e1012d4859fc3eb03d86a6a6e69741d7.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 359, + 137, + 446, + 200 + ], + "blocks": [ + { + "bbox": [ + 359, + 137, + 446, + 200 + ], + "lines": [ + { + "bbox": [ + 359, + 137, + 446, + 200 + ], + "spans": [ + { + "bbox": [ + 359, + 137, + 446, + 200 + ], + "type": "image", + "image_path": "45f3a3f6d7c052ae5c2781a452073556a900ee1a9f9720c0c90bdaaf380e9c82.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 448, + 137, + 545, + 200 + ], + "blocks": [ + { + "bbox": [ + 448, + 137, + 545, + 200 + ], + "lines": [ + { + "bbox": [ + 448, + 137, + 545, + 200 + ], + "spans": [ + { + "bbox": [ + 448, + 137, + 545, + 200 + ], + "type": "image", + "image_path": "6b2667397cb062187c8f4e4d9794483346b94a6b5791220c1efb4b0697422d10.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 71, + 201, + 169, + 266 + ], + "blocks": [ + { + "bbox": [ + 55, + 230, + 67, + 239 + ], + "lines": [ + { + "bbox": [ + 55, + 230, + 67, + 239 + ], + "spans": [ + { + "bbox": [ + 55, + 230, + 67, + 239 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 71, + 201, + 169, + 266 + ], + "lines": [ + { + "bbox": [ + 71, + 201, + 169, + 266 + ], + "spans": [ + { + "bbox": [ + 71, + 201, + 169, + 266 + ], + "type": "image", + "image_path": "7298ee1d89fd8e1d9b9d9d7ccfc9a7cc29767e9e0b085ee89e663f7239edc5b6.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 170, + 201, + 268, + 266 + ], + "blocks": [ + { + "bbox": [ + 170, + 201, + 268, + 266 + ], + "lines": [ + { + "bbox": [ + 170, + 201, + 268, + 266 + ], + "spans": [ + { + "bbox": [ + 170, + 201, + 268, + 266 + ], + "type": "image", + "image_path": "b56b68e9d9a4d1641dc2c628af8c5c270e47eb3fe63f72b28ffc1fb87269c6bb.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 275, + 547, + 298 + ], + "lines": [ + { + "bbox": [ + 46, + 275, + 547, + 298 + ], + "spans": [ + { + "bbox": [ + 46, + 275, + 547, + 298 + ], + "type": "text", + "content": "Figure 3. Visualization of activation patterns from different detectors. (a) Pseudo labels (green) and ground truth annotations (blue) for each image. (b) The baseline detector. (c) Our OADP. The intensity of the feature response increases from blue to red." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 269, + 201, + 356, + 266 + ], + "blocks": [ + { + "bbox": [ + 269, + 201, + 356, + 266 + ], + "lines": [ + { + "bbox": [ + 269, + 201, + 356, + 266 + ], + "spans": [ + { + "bbox": [ + 269, + 201, + 356, + 266 + ], + "type": "image", + "image_path": "1be3077e0e899a9331ef3d6219013c9ea13a985c9ab3c0b37904bf28ef8c0be2.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 359, + 201, + 446, + 266 + ], + "blocks": [ + { + "bbox": [ + 359, + 201, + 446, + 266 + ], + "lines": [ + { + "bbox": [ + 359, + 201, + 446, + 266 + ], + "spans": [ + { + "bbox": [ + 359, + 201, + 446, + 266 + ], + "type": "image", + "image_path": "c9b9b262e5c4cd7dc393afb090bf446b9fe68bc9425360d68b22196d1a76a806.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 448, + 201, + 545, + 266 + ], + "blocks": [ + { + "bbox": [ + 448, + 201, + 545, + 266 + ], + "lines": [ + { + "bbox": [ + 448, + 201, + 545, + 266 + ], + "spans": [ + { + "bbox": [ + 448, + 201, + 545, + 266 + ], + "type": "image", + "image_path": "be5132276b97f020e47b2952e407f49eb84748935e1f616d19141c4b4b52aa1e.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "table", + "bbox": [ + 55, + 304, + 280, + 407 + ], + "blocks": [ + { + "bbox": [ + 55, + 304, + 280, + 407 + ], + "lines": [ + { + "bbox": [ + 55, + 304, + 280, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 304, + 280, + 407 + ], + "type": "table", + "html": "
Method\\( \\mathbf{{mAP}}^{\\mathbf{{PL}}} \\)\\( \\mathbf{{mAP}}_{50}^{\\mathbf{{PL}}} \\)\\( \\mathbf{{mAP}}_{75}^{\\mathbf{{PL}}} \\)#PL
Baseline5.012.13.1100
ViLD*17.828.118.6100
Ours19.029.919.9100
Baseline3.99.32.56.53
VL-PLM [52]-25.3-4.26
Ours17.426.518.64.14
", + "image_path": "f20c9e2d253e914e6315c46d485f0914d5486e679c483193336b20e6fecf105a.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "table_body" + } + ], + "index": 19 + }, + { + "bbox": [ + 46, + 415, + 288, + 437 + ], + "lines": [ + { + "bbox": [ + 46, + 415, + 288, + 437 + ], + "spans": [ + { + "bbox": [ + 46, + 415, + 288, + 437 + ], + "type": "text", + "content": "Table 6. Ablation study of pseudo labels. \"#PL\" is the number of pseudo labels per image." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 447, + 287, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 447, + 287, + 662 + ], + "spans": [ + { + "bbox": [ + 46, + 447, + 287, + 662 + ], + "type": "text", + "content": "proposal embeddings from the original proposal, and it relies merely on the classification score to sort proposals. The poor detection accuracy in " + }, + { + "bbox": [ + 46, + 447, + 287, + 662 + ], + "type": "inline_equation", + "content": "1^{\\text{st}}" + }, + { + "bbox": [ + 46, + 447, + 287, + 662 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 447, + 287, + 662 + ], + "type": "inline_equation", + "content": "4^{\\text{th}}" + }, + { + "bbox": [ + 46, + 447, + 287, + 662 + ], + "type": "text", + "content": " rows of tab. 6 show that without the objectness score, the baseline method can not accurately localize objects. We re-implement multiscale region embedding of ViLD [12] with a geometric mean of CLIP classification score and objectiveness score, i.e., \"ViLD*\". We adopt an adaptive transform strategy for proposals and regard the output of [OBJ] as proposal embedding. The score fusion strategy is described in sec. 4.5, where " + }, + { + "bbox": [ + 46, + 447, + 287, + 662 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 447, + 287, + 662 + ], + "type": "text", + "content": " is 0.3. Our method achieves the highest mAP when the number of PLs is sufficient. VL-PLM [52] adopts a multi-scale region embedding method similar to ViLD [12] except for an arithmetic mean of classification score and objections score. When the pseudo labels are filtered with a higher confidence threshold, our method still has a significant advantage compared to VL-PLM [52] (26.50 mAPPL compared to 25.30 mAPPL)." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 47, + 670, + 130, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 130, + 682 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 130, + 682 + ], + "type": "text", + "content": "5.5. Visualization" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": "Fig. 3 (a) visualizes our generated PLs on novel categories in green with ground truth boxes of base categories in" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 307, + 546, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 307, + 546, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 307, + 546, + 437 + ], + "type": "text", + "content": "blue. We try our best to ensure the accuracy of PLs as much as possible to be fewer and more precise. These green PLs demonstrate that our proposal embeddings can clearly distinguish novel objects from base ones. Correspondingly, we also show activation maps from baseline (b) and our detector (c) in fig. 3. Taking the " + }, + { + "bbox": [ + 304, + 307, + 546, + 437 + ], + "type": "inline_equation", + "content": "2^{\\mathrm{nd}}" + }, + { + "bbox": [ + 304, + 307, + 546, + 437 + ], + "type": "text", + "content": " column as an example, the activation map of our detector accurately highlights more area of novel objects, i.e. \"cup\", with our distillation pyramid mechanism. Therefore, the backbone of OADP generates more informative feature maps, which further help detect novel objects." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 306, + 452, + 378, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 452, + 378, + 464 + ], + "spans": [ + { + "bbox": [ + 306, + 452, + 378, + 464 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 473, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 473, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 473, + 545, + 617 + ], + "type": "text", + "content": "In this paper, we reconsider the way of knowledge extraction and knowledge transfer in existing KD-based OVD methods and propose an Object-Aware Distillation Pyramid (OADP) framework. To preserve complete and purified object representation in proposals during knowledge extraction, we propose an Object-Aware Knowledge Extraction (OAKE) module to adaptively transform proposals and extract precise object knowledge. A Distillation Pyramid (DP) mechanism is proposed to transfer contextual knowledge about the relation of different objects for better scene understanding. Experiments show that our OADP outperforms previous methods on two popular OVD benchmarks." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 630, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 712 + ], + "type": "text", + "content": "Acknowledgement. This research is supported by the National Key R&D Program of China (2022ZD0115502), National Natural Science Foundation of China (NO. 62122010), Zhejiang Provincial Natural Science Foundation of China under Grant No. LDT23F02022F02, Key Research and Development Program of Zhejiang Province under Grant 2022C01082." + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "11193" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 127 + ], + "type": "text", + "content": "[1] Ankan Bansal, Karan Sikka, Gaurav Sharma, Rama Chellappa, and Ajay Divakaran. Zero-Shot Object Detection. In ECCV, 2018. 3, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 130, + 288, + 154 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 130, + 288, + 154 + ], + "spans": [ + { + "bbox": [ + 53, + 130, + 288, + 154 + ], + "type": "text", + "content": "[2] Hakan Bilen and Andrea Vedaldi. Weakly Supervised Deep Detection Networks. In CVPR, 2016. 3, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 157, + 288, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 157, + 288, + 205 + ], + "spans": [ + { + "bbox": [ + 54, + 157, + 288, + 205 + ], + "type": "text", + "content": "[3] Maria A. Bravo, Sudhanshu Mittal, and Thomas Brox. Localized Vision-Language Matching for Open-vocabulary Object Detection. In GCRP, 2022. 3, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 208, + 288, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 208, + 288, + 267 + ], + "spans": [ + { + "bbox": [ + 54, + 208, + 288, + 267 + ], + "type": "text", + "content": "[4] Zhaowei Cai, Gukyeong Kwon, Avinash Ravichandran, Erhan Bas, Zhuowen Tu, Rahul Bhotika, and Stefano Soatto. X-DETR: A Versatile Architecture for Instance-wise Vision-Language Tasks. In ECCV, 2022. 2, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 270, + 288, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 270, + 288, + 318 + ], + "spans": [ + { + "bbox": [ + 54, + 270, + 288, + 318 + ], + "type": "text", + "content": "[5] Guobin Chen, Wongun Choi, Xiang Yu, Tony Han, and Manmohan Chandraker. Learning efficient object detection models with knowledge distillation. In NeurIPS, 2017. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 321, + 288, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 321, + 288, + 381 + ], + "spans": [ + { + "bbox": [ + 54, + 321, + 288, + 381 + ], + "type": "text", + "content": "[6] Xinlei Chen, Hao Fang, Tsung-yi Lin, Ramakrishna Vedantam, Saurabh Gupta, Piotr Dollar, and C Lawrence Zitnick. Microsoft COCO Captions: Data Collection and Evaluation Server. arXiv preprint arXiv:1504.00325, 2015. 3, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 384, + 288, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 384, + 288, + 431 + ], + "spans": [ + { + "bbox": [ + 54, + 384, + 288, + 431 + ], + "type": "text", + "content": "[7] Xing Dai, Zeren Jiang, Zhao Wu, Yiping Bao, Zhicheng Wang, Si Liu, and Erjin Zhou. General Instance Distillation for Object Detection. In CVPR, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 435, + 288, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 435, + 288, + 483 + ], + "spans": [ + { + "bbox": [ + 54, + 435, + 288, + 483 + ], + "type": "text", + "content": "[8] Yu Du, Fangyun Wei, Zihe Zhang, Miaojing Shi, Yue Gao, and Guoqi Li. Learning to Prompt for Open-Vocabulary Object Detection with Vision-Language Model. In CVPR, 2022. 2, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 486, + 288, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 486, + 288, + 533 + ], + "spans": [ + { + "bbox": [ + 54, + 486, + 288, + 533 + ], + "type": "text", + "content": "[9] Zhixing Du, Rui Zhang, Ming Chang, Xishan Zhang, Shaoli Liu, Tianshi Chen, and Yunji Chen. Distilling Object Detectors with Feature Richness. In NeurIPS, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 537, + 287, + 585 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 537, + 287, + 585 + ], + "spans": [ + { + "bbox": [ + 48, + 537, + 287, + 585 + ], + "type": "text", + "content": "[10] Chengjian Feng, Yujie Zhong, Zequn Jie, Xiangxiang Chu, Haibing Ren, Xiaolin Wei, Weidi Xie, and Lin Ma. PromptDet: Expand Your Detector Vocabulary with Uncurated Images. In ECCV, 2022. 2, 3, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 588, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 588, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 588, + 287, + 635 + ], + "type": "text", + "content": "[11] Mingfei Gao, Chen Xing, Juan Carlos Niebles, Junnan Li, Ran Xu, Wenhao Liu, and Caiming Xiong. Open Vocabulary Object Detection with Pseudo Bounding-Box Labels. In ECCV, 2022. 2, 3, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 639, + 287, + 686 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 639, + 287, + 686 + ], + "spans": [ + { + "bbox": [ + 48, + 639, + 287, + 686 + ], + "type": "text", + "content": "[12] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary Object Detection via Vision and Language Knowledge Distillation. In ICLR, 2022. 1, 2, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 689, + 287, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 689, + 287, + 714 + ], + "type": "text", + "content": "[13] Jianyuan Guo, Kai Han, Yunhe Wang, Han Wu, Xing-hao Chen, Chunjing Xu, and Chang Xu. Distilling" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 328, + 72, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 328, + 72, + 545, + 95 + ], + "type": "text", + "content": "Object Detectors via Decoupled Features. In CVPR, 2021. 2, 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 100, + 545, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 100, + 545, + 135 + ], + "spans": [ + { + "bbox": [ + 307, + 100, + 545, + 135 + ], + "type": "text", + "content": "[14] Agrim Gupta, Piotr Dollar, and Ross Girshick. LVIS: A Dataset for Large Vocabulary Instance Segmentation. In CVPR, 2019. 2, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 140, + 545, + 164 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 140, + 545, + 164 + ], + "spans": [ + { + "bbox": [ + 307, + 140, + 545, + 164 + ], + "type": "text", + "content": "[15] Kaiming He, Georgia Gkioxari, Piotr Dollar, and Ross Girshick. Mask R-CNN. TPAMI, 42(2), 2020. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 169, + 545, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 169, + 545, + 204 + ], + "spans": [ + { + "bbox": [ + 307, + 169, + 545, + 204 + ], + "type": "text", + "content": "[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In CVPR, 2016. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 209, + 545, + 244 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 209, + 545, + 244 + ], + "spans": [ + { + "bbox": [ + 307, + 209, + 545, + 244 + ], + "type": "text", + "content": "[17] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the Knowledge in a Neural Network. In NeurIPS, 2014. 1, 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 248, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 248, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 248, + 545, + 285 + ], + "type": "text", + "content": "[18] Peiliang Huang, Junwei Han, De Cheng, and Dingwen Zhang. Robust Region Feature Synthesizer for Zero-Shot Object Detection. In CVPR, 2022. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 289, + 545, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 289, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 307, + 289, + 545, + 350 + ], + "type": "text", + "content": "[19] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yunhsuan Sung, Zhen Li, and Tom Duerig. Scaling Up Visual and Vision-Language Representation Learning With Noisy Text Supervision. In ICML, 2021. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 353, + 545, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 353, + 545, + 401 + ], + "spans": [ + { + "bbox": [ + 307, + 353, + 545, + 401 + ], + "type": "text", + "content": "[20] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. MDETR - Modulated Detection for End-to-End Multi-Modal Understanding. In ICCV, 2021. 2, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 405, + 545, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 405, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 307, + 405, + 545, + 453 + ], + "type": "text", + "content": "[21] Zijian Kang, Peizhen Zhang, Xiangyu Zhang, Jian Sun, and Nanning Zheng. Instance-Conditional Knowledge Distillation for Object Detection. In NeurIPS, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 457, + 545, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 457, + 545, + 517 + ], + "spans": [ + { + "bbox": [ + 307, + 457, + 545, + 517 + ], + "type": "text", + "content": "[22] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, KaiWei Chang, and Jianfeng Gao. Grounded Language-Image Pre-training. In CVPR, 2022. 2, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 521, + 545, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 521, + 545, + 556 + ], + "spans": [ + { + "bbox": [ + 307, + 521, + 545, + 556 + ], + "type": "text", + "content": "[23] Quanquan Li, Shengying Jin, and Junjie Yan. Mimicking Very Efficient Network for Object Detection. In CVPR, 2017. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 561, + 545, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 561, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 307, + 561, + 545, + 609 + ], + "type": "text", + "content": "[24] Zhihui Li, Lina Yao, Xiaqin Zhang, Xianzhi Wang, Salil Kanhere, and Huaxiang Zhang. Zero-Shot Object Detection with Textual Descriptions. AAAI, 33(1), 2019. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 613, + 545, + 662 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 545, + 662 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 545, + 662 + ], + "type": "text", + "content": "[25] Yue Liao, Aixi Zhang, Miao Lu, Yongliang Wang, Xiaobo Li, and Si Liu. Gen-vlkt: Simplify association and enhance interaction understanding for hoi detection. In CVPR, pages 20091-20100, 2022. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 666, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 666, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 666, + 545, + 712 + ], + "type": "text", + "content": "[26] Tsung-Yi Lin, Piotr Dollar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature Pyramid Networks for Object Detection. In CVPR, 2017. 3" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "11194" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 132 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 132 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 132 + ], + "type": "text", + "content": "[27] Tsung-Yi Lin, Michael Maire, Serge Belongie, Lubomir Bourdev, Ross Girshick, James Hays, Pietro Perona, Deva Ramanan, C. Lawrence Zitnick, and Piotr Dóllár. Microsoft COCO: Common Objects in Context. In ECCV, 2014. 2, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 135, + 287, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 135, + 287, + 194 + ], + "spans": [ + { + "bbox": [ + 48, + 135, + 287, + 194 + ], + "type": "text", + "content": "[28] Zongyang Ma, Guan Luo, Jin Gao, Liang Li, Yuxin Chen, Shaoru Wang, Congxuan Zhang, and Weiming Hu. Open-Vocabulary One-Stage Detection with Hierarchical Visual-Language Knowledge Distillation. In CVPR, 2022. 1, 2, 3, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 198, + 286, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 198, + 286, + 281 + ], + "spans": [ + { + "bbox": [ + 48, + 198, + 286, + 281 + ], + "type": "text", + "content": "[29] Matthias Minderer, Alexey Gritsenko, Austin Stone, Maxim Neumann, Dirk Weissenborn, Alexey Dosovitskiy, Aravindh Mahendran, Anurag Arnab, Mostafa Dehghani, Zhuoran Shen, Xiao Wang, Xiaohua Zhai, Thomas Kipf, and Neil Houlsby. Simple Open-Vocabulary Object Detection with Vision Transformers. In ECCV, 2022. 1, 2, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 285, + 286, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 285, + 286, + 331 + ], + "spans": [ + { + "bbox": [ + 48, + 285, + 286, + 331 + ], + "type": "text", + "content": "[30] Seyed Iman Mirzadeh, Mehrdad Farajtabar, Ang Li, Nir Levine, Akihiro Matsukawa, and Hassan Ghasemzadeh. Improved Knowledge Distillation via Teacher Assistant. AAAI, 34(04), 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 335, + 286, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 335, + 286, + 381 + ], + "spans": [ + { + "bbox": [ + 48, + 335, + 286, + 381 + ], + "type": "text", + "content": "[31] Ishan Misra, C. Lawrence Zitnick, Margaret Mitchell, and Ross Girshick. Seeing through the Human Reporting Bias: Visual Classifiers from Noisy Human-Centric Labels. In CVPR, 2016. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 385, + 286, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 385, + 286, + 456 + ], + "spans": [ + { + "bbox": [ + 48, + 385, + 286, + 456 + ], + "type": "text", + "content": "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning Transferable Visual Models From Natural Language Supervision. In ICML, volume 139, 2021. 1, 2, 3, 4, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 460, + 286, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 460, + 286, + 495 + ], + "spans": [ + { + "bbox": [ + 48, + 460, + 286, + 495 + ], + "type": "text", + "content": "[33] Shafin Rahman, Salman Khan, and Nick Barnes. Transductive Learning for Zero-Shot Object Detection. In ICCV, 2019. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 499, + 286, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 499, + 286, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 499, + 286, + 533 + ], + "type": "text", + "content": "[34] Shafin Rahman, Salman Khan, and Nick Barnes. Improved Visual-Semantic Alignment for Zero-Shot Object Detection. AAAI, 34(07), 2020. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 538, + 286, + 583 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 538, + 286, + 583 + ], + "spans": [ + { + "bbox": [ + 48, + 538, + 286, + 583 + ], + "type": "text", + "content": "[35] Shafin Rahman, Salman H. Khan, and Fatih Porikli. Zero-Shot Object Detection: Joint Recognition and Localization of Novel Concepts. IJCV, 128(12), 2020. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 589, + 286, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 589, + 286, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 589, + 286, + 635 + ], + "type": "text", + "content": "[36] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks. TPAMI, 39(6), 2017. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 639, + 286, + 685 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 639, + 286, + 685 + ], + "spans": [ + { + "bbox": [ + 48, + 639, + 286, + 685 + ], + "type": "text", + "content": "[37] Adriana Romero, Nicolas Ballas, Samira Ebrahimi Kahou, Antoine Chassang, Carlo Gatta, and Yoshua Bengio. FitNets: Hints for Thin Deep Nets. In ICLR, 2015. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 690, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 690, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 690, + 286, + 713 + ], + "type": "text", + "content": "[38] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 72, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 328, + 72, + 545, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 72, + 545, + 118 + ], + "spans": [ + { + "bbox": [ + 328, + 72, + 545, + 118 + ], + "type": "text", + "content": "Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. LAION-400M: Open Dataset of CLIP-Filtered 400 Million Image-Text Pairs. arXiv preprint arXiv:2111.02114, 2021.6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 308, + 124, + 545, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 124, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 308, + 124, + 545, + 171 + ], + "type": "text", + "content": "[39] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual Captions: A Cleaned, Hypernymed, Image Alt-text Dataset For Automatic Image Captioning. In ACL, 2018. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 175, + 545, + 210 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 175, + 545, + 210 + ], + "spans": [ + { + "bbox": [ + 308, + 175, + 545, + 210 + ], + "type": "text", + "content": "[40] Tao Wang, Li Yuan, Xiaopeng Zhang, and Jiashi Feng. Distilling Object Detectors With Fine-Grained Feature Imitation. In CVPR, 2019. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 215, + 545, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 215, + 545, + 261 + ], + "spans": [ + { + "bbox": [ + 308, + 215, + 545, + 261 + ], + "type": "text", + "content": "[41] Fangyun Wei, Yue Gao, Zhirong Wu, Han Hu, and Stephen Lin. Aligning Pretraining for Detection via Object-Level Contrastive Learning. In NeurIPS, volume 27, 2021. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 266, + 545, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 266, + 545, + 312 + ], + "spans": [ + { + "bbox": [ + 308, + 266, + 545, + 312 + ], + "type": "text", + "content": "[42] Caixia Yan, Xiaojun Chang, Minnan Luo, Huan Liu, Xiaoqin Zhang, and Qinghua Zheng. Semantics-Guided Contrastive Network for Zero-Shot Object detection. TPAMI, 2022. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 318, + 545, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 318, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 308, + 318, + 545, + 364 + ], + "type": "text", + "content": "[43] Zhendong Yang, Zhe Li, Xiaohu Jiang, Yuan Gong, Zehuan Yuan, Danpei Zhao, and Chun Yuan. Focal and Global Knowledge Distillation for Detectors. In CVPR, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 369, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 369, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 308, + 369, + 545, + 403 + ], + "type": "text", + "content": "[44] Zhendong Yang, Zhe Li, Mingqi Shao, Dachuan Shi, Zehuan Yuan, and Chun Yuan. Masked Generative Distillation. In ECCV, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 408, + 545, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 408, + 545, + 468 + ], + "spans": [ + { + "bbox": [ + 308, + 408, + 545, + 468 + ], + "type": "text", + "content": "[45] Lewei Yao, Jianhua Han, Youpeng Wen, Xiaodan Liang, Dan Xu, Wei Zhang, Zhenguo Li, Chunjing Xu, and Hang Xu. DetCLIP: Dictionary-Enriched Visual-Concept Paralleled Pre-training for Open-world Detection. In NeurIPS, 2022. 2, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 472, + 545, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 472, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 308, + 472, + 545, + 529 + ], + "type": "text", + "content": "[46] Lewei Yao, Renjie Pi, Hang Xu, Wei Zhang, Zhenguo Li, and Tong Zhang. G-DetKD: Towards General Distillation Framework for Object Detectors via Contrastive and Semantic-guided Feature Imitation. In ICCV, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 535, + 545, + 582 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 535, + 545, + 582 + ], + "spans": [ + { + "bbox": [ + 308, + 535, + 545, + 582 + ], + "type": "text", + "content": "[47] Keren Ye, Mingda Zhang, Adriana Kovashka, Wei Li, Danfeng Qin, and Jesse Berent. Cap2Det: Learning to Amplify Weak Caption Supervision for Object Detection. In ICCV, 2019. 3, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 587, + 545, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 587, + 545, + 622 + ], + "spans": [ + { + "bbox": [ + 308, + 587, + 545, + 622 + ], + "type": "text", + "content": "[48] Yuhang Zang, Wei Li, Kaiyang Zhou, Chen Huang, and Chen Change Loy. Open-Vocabulary DETR with Conditional Matching. ECCV, 2022. 1, 2, 3, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 627, + 545, + 662 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 627, + 545, + 662 + ], + "spans": [ + { + "bbox": [ + 308, + 627, + 545, + 662 + ], + "type": "text", + "content": "[49] Alireza Zareian, Kevin Dela Rosa, Derek Hao Hu, and Shih-Fu Chang. Open-Vocabulary Object Detection Using Captions. In CVPR, 2021. 1, 2, 3, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 666, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 666, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 666, + 545, + 712 + ], + "type": "text", + "content": "[50] Peizhen Zhang, Zijian Kang, Tong Yang, Xiangyu Zhang, Nanning Zheng, and Jian Sun. LGD: Label-Guided Self-Distillation for Object Detection. AAAI, 36(3), 2022. 2" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "11195" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 435 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 120 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 120 + ], + "type": "text", + "content": "[51] Shizhen Zhao, Changxin Gao, Yuanjie Shao, Lerenhan Li, Changqian Yu, Zhong Ji, and Nong Sang. GT-Net: Generative Transfer Network for Zero-Shot Object Detection. AAAI, 34(07), 2020. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 123, + 288, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 123, + 288, + 194 + ], + "spans": [ + { + "bbox": [ + 48, + 123, + 288, + 194 + ], + "type": "text", + "content": "[52] Shiyu Zhao, Zhixing Zhang, Samuel Schulter, Long Zhao, Vijay Kumar B. G, Anastasis Stathopoulos, Manmohan Chandraker, and Dimitris Metaxas. Exploiting Unlabeled Data with Vision and Language Models for Object Detection. In ECCV, 2022. 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 198, + 288, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 198, + 288, + 257 + ], + "spans": [ + { + "bbox": [ + 48, + 198, + 288, + 257 + ], + "type": "text", + "content": "[53] Yiwu Zhong, Jianwei Yang, Pengchuan Zhang, Chunyuan Li, Noel Codella, Liunian Harold Li, Luowei Zhou, Xiyang Dai, Lu Yuan, Yin Li, and Jianfeng Gao. RegionCLIP: Region-based Language-Image Pretraining. In CVPR, 2022. 1, 2, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 260, + 288, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 260, + 288, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 260, + 288, + 308 + ], + "type": "text", + "content": "[54] Chunting Zhou, Graham Neubig, and Jiatao Gu. Improve Object Detection with Feature-based Knowledge Distillation: Towards Accurate and Efficient Detectors. In ICLR, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 311, + 288, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 311, + 288, + 358 + ], + "spans": [ + { + "bbox": [ + 48, + 311, + 288, + 358 + ], + "type": "text", + "content": "[55] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Phillip Krahenbuhl, and Ishan Misra. Detecting Twenty-thousand Classes using Image-level Supervision. In ECCV, 2022. 2, 3, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 362, + 288, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 362, + 288, + 397 + ], + "spans": [ + { + "bbox": [ + 48, + 362, + 288, + 397 + ], + "type": "text", + "content": "[56] Pengkai Zhu, Hanxiao Wang, and Venkatesh Saligrama. Don't Even Look Once: Synthesizing Features for Zero-Shot Detection. In CVPR, 2020. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 401, + 288, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 401, + 288, + 435 + ], + "spans": [ + { + "bbox": [ + 48, + 401, + 288, + 435 + ], + "type": "text", + "content": "[57] Pengkai Zhu, Hanxiao Wang, and Venkatesh Saligrama. Zero Shot Detection. TCSVT, 30(4), 2020. 3" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "11196" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Object-Goal Visual Navigation via Effective Exploration of Relations Among Historical Navigation States/02126feb-2cfd-431e-b8b7-664314fe1580_content_list.json b/2023/Object-Goal Visual Navigation via Effective Exploration of Relations Among Historical Navigation States/02126feb-2cfd-431e-b8b7-664314fe1580_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..27f2b4ff2b185ca897bad218c001c6697fc3f329 --- /dev/null +++ b/2023/Object-Goal Visual Navigation via Effective Exploration of Relations Among Historical Navigation States/02126feb-2cfd-431e-b8b7-664314fe1580_content_list.json @@ -0,0 +1,1533 @@ +[ + { + "type": "text", + "text": "Object-Goal Visual Navigation via Effective Exploration of Relations among Historical Navigation States", + "text_level": 1, + "bbox": [ + 133, + 130, + 834, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Heming Du $^{1,3}$ , Lincheng Li $^{2}$ , Zi Huang $^{3}$ , Xin Yu $^{3*}$", + "bbox": [ + 285, + 202, + 686, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Australian National University $^{2}$ Netease Fuxi AI Lab $^{3}$ The University of Queensland", + "bbox": [ + 135, + 220, + 828, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 273, + 313, + 289 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Object-goal visual navigation aims at steering an agent toward an object via a series of moving steps. Previous works mainly focus on learning informative visual representations for navigation, but overlook the impacts of navigation states on the effectiveness and efficiency of navigation. We observe that high relevance among navigation states will cause navigation inefficiency or failure for existing methods. In this paper, we present a History-inspired Navigation Policy Learning (HiNL) framework to estimate navigation states effectively by exploring relationships among historical navigation states. In HiNL, we propose a History-aware State Estimation (HaSE) module to alleviate the impacts of dominant historical states on the current state estimation. Meanwhile, HaSE also encourages an agent to be alert to the current observation changes, thus enabling the agent to make valid actions. Furthermore, we design a History-based State Regularization (HbSR) to explicitly suppress the correlation among navigation states in training. As a result, our agent can update states more effectively while reducing the correlations among navigation states. Experiments on the artificial platform AI2-THOR (i.e., iTHOR and RoboTHOR) demonstrate that HiNL significantly outperforms state-of-the-art methods on both Success Rate and SPL in unseen testing environments.", + "bbox": [ + 75, + 306, + 473, + 670 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 698, + 209, + 714 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Object-goal visual navigation is to direct an agent to move consecutively toward an object of a specific category. Without knowing the environment map beforehand, at each navigation step, an agent first needs to represent its visual observations, then estimate its navigation states from the visual representations and the preceding states, and at last predict the corresponding action. Therefore, to achieve an effective and efficient navigation system, learning instructive visual representations and navigation states is critical.", + "bbox": [ + 75, + 724, + 468, + 859 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Prevailing visual navigation works [13, 14, 50] focus on", + "bbox": [ + 96, + 861, + 468, + 876 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/286b18ce66bcf8d30701be3925c4d94da35fa6563ff7b8baa454b1b2542f5b00.jpg", + "image_caption": [], + "image_footnote": [ + "(a) Demonstration of inefficient action predictions caused by highlycorrelated navigation states. Our agent is stuck by an obstacle, i.e., lowprofile sofa, and repeatedly predicts an invalid action, i.e., MoveAhead." + ], + "bbox": [ + 542, + 275, + 856, + 428 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/92d327d72606d350f666c12ad46524fbb9736254950a97c10325cdd95ba2947c.jpg", + "image_caption": [ + "Figure 1. Motivation of our proposed History-inspired Navigation Learning (HiNL) framework." + ], + "image_footnote": [ + "(b) Demonstration of the correlation coefficients among navigation states trained in two manners. Navigation states estimated via LSTM are highly-relevant. In contrast, our HiNL produce low-correlated navigation states." + ], + "bbox": [ + 553, + 469, + 833, + 565 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "extracting informative visual representations, while some methods [13, 49] adjust navigation policy during inference. All these approaches commonly employ recurrent neural networks (e.g., LSTM) to estimate navigation states. However, we observe that the navigation states of existing methods [13, 14, 49] exhibit high relevance, as demonstrated in Figure 1b, and the highly-correlated navigation states would lead to inefficient navigation policy (i.e., failure to respond to observation changes rapidly). For instance, as shown in Figure 1a, an agent is stuck by the low-profile sofa and fails to take proper actions to circumvent the obstacle. Hence, we aim to endow an agent with the capability of updating its navigation states effectively while avoiding producing highly-correlated states.", + "bbox": [ + 496, + 671, + 890, + 882 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we propose a History-inspired Navigation", + "bbox": [ + 517, + 885, + 890, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 94, + 887, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "2563", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Learning (HiNL) framework to obtain informative navigation states by exploiting the relationships among historical navigation states. HiNL consists of two novel components: (i) a History-aware State Estimation (HaSE) module, and (ii) a History-based State Regularization (HbSR). Here, our HaSE module is designed to generate a state that can be promptly updated according to visual observations. Specifically, HaSE first analyzes the correlations among historical navigation states and then eliminates the influence of dominant historical states on the current state estimation. As a result, an agent is able to predict navigation states which can dynamically react to the current visual observations and then make sensible navigation actions.", + "bbox": [ + 75, + 90, + 472, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Furthermore, existing reinforcement learning-based object-goal navigation systems [13, 14, 49] often assume the navigation state transition exhibits the first-order Markov property. This would allow the emergence of high correlations among navigation states, leading to inferior navigation policy. To address this issue, we introduce an explicit constraint on the correlation among all the states, namely History-based State Regularization (HbSR). To be specific, HbSR enforces to relevance (i.e., correlations) between a state and all its preceding states (except its previous state) to be low. Here, we do not constrain states of two consecutive steps because temporally close states generally have relevance in practice considering the navigation continuity. After training with our HbSR, the correlations among the navigation states become much lower (see Figure 1b). This pheromone also indicates HiNL effectively updates states. Hence, our navigation system can respond to observation changes adaptively.", + "bbox": [ + 75, + 289, + 472, + 561 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To demonstrate the superiority of HiNL, we conduct experiments in the widely-adopted artificial environment iTHOR [26] and RoboTHOR [11]. HiNL outperforms the state-of-the-art by a large margin. To be specific, we improve the Success Rate (SR) from $72.2\\%$ to $80.1\\%$ and Success weighted by Path Length (SPL) from 0.449 to 0.498 in iTHOR. Overall, our major contributions are summarized as follows:", + "bbox": [ + 75, + 564, + 472, + 683 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a History-inspired Navigation Policy (HiNL) framework to effectively estimate navigation states by utilizing historical states.", + "- We design a History-aware State Estimation (HaSE) to eliminate dominant historical states in the current state estimation. Therefore, the agent reduces the impact of distant navigation states on the state estimation, and thus reacts dynamically to the observation changes.", + "- We introduce a History-based State Regularization (HbSR) to explicitly constrain the correlations among navigation states. By doing this, the agent can effectively update navigation states with low relevance." + ], + "bbox": [ + 96, + 694, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 500, + 89, + 650, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Visual navigation. Traditional works [3, 4, 32] often leverage an entire environment map for navigation and divide the task into three parts: mapping, localization, and path planning. However, environment maps are generally unavailable in unseen environments. Dissanayake et al. [12] adopt simultaneous localization and mapping (SLAM) to infer robot positions. Campari et al. [5] learn agent states via a Taskonomy model bank [52], but they need an RGB-D sensor to construct an online map during navigation.", + "bbox": [ + 496, + 116, + 893, + 253 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, due to significant advancements in deep learning [16,22,45,51], reinforcement learning-based navigation methods [28, 29, 31, 33, 54] take visual observations as inputs and predict navigation actions. Vision-Language Navigation (VLN) approaches [9, 10, 15, 39, 40] steer an agent to the target based on its visual observations and navigation guidance in natural language. Similar to VLN, point-goal visual navigation methods [44,48] aim at driving an agent to a given point with step-wise directional indications. Moreover, audio-visual navigation methods [7, 8, 18] utilize additional audio signals to move a robot to the target position. Al-Halah et al. [1] propose a transfer learning model for multiple navigation tasks by embedding various navigation goals, e.g., image, sketch, and audio.", + "bbox": [ + 496, + 255, + 893, + 467 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our work falls in the field of object-goal visual navigation [27, 31, 36, 47, 50]. However, existing object-goal navigation methods mainly focus on representing visual features comprehensively while we investigate the impact of navigation states on navigation performance. Wortsman et al. [49] exploit word embedding (i.e., GloVe embedding [35]) to represent the target category and introduce a meta network that mimics a reward function during inference. Du et al. [13] introduce an object relation graph, dubbed ORG, to encode visual observations. They also design a tentative policy for deadlock avoidance and adjust the navigation policy in unseen testing environments. The Hierarchical Object-to-Zone (HOZ) graph [53] offers coarse-to-fine guidance based on real-time updates. Additionally, VT-Net [14] incorporates object and region features with location cues, and EmbCLIP [24] leverages the contrastive language image pretraining encoder for visual navigation tasks.", + "bbox": [ + 496, + 469, + 895, + 739 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Correlation Modeling in Reinforcement Learning. Several methods [37, 43] explore correlations in hidden Markov models for inverse reinforcement learning (IRL). For action prediction, Hester et al. [20] propose Texlore to model correlations within the transition dynamics via a random forest. Vsovsic et al. [42] introduce a Bayesian approach to learn policy from demonstrations of experts by capturing correlations among actions. Alt et al. [2] design a Bayesian learning framework to establish temporal and spatial correlations among actions. Furthermore,", + "bbox": [ + 496, + 750, + 893, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2564", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/8f68e97d22c1b7e7a9f7bd6f7d1ab1c50e06d13186a68d125343093ad5edd1ea.jpg", + "image_caption": [ + "Figure 2. Our History-inspired Navigation Policy Learning (HiNL) framework. HiNL takes visual representations as input and outputs navigation actions. HiNL involves two innovative parts: a History-aware State Estimation (HaSE) module and a History-based State Regularization (HbSR). HaSE is proposed to estimate navigation states that can reflect current observation changes from the perspective of network design, while HbSR is designed to enforce the informativeness of states from the view of the training objective. Both of them help to achieve effective and efficient navigation policy." + ], + "image_footnote": [], + "bbox": [ + 112, + 104, + 867, + 265 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Sermanet [41] propose a self-supervised TCN for learning robotic behaviors and representations from unlabeled multi-viewpoint videos. TCN uses a metric learning loss to create viewpoint-invariant representations that capture relationships between end-effectors.", + "bbox": [ + 75, + 368, + 470, + 444 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Different from previous methods that establish correlations among actions, our HiNL aims to eliminate high relevance among navigation states. To the best of our knowledge, our work HiNL is the first attempt to explicitly leverage the correlations among navigation states for object-goal navigation and achieves superior navigation performance.", + "bbox": [ + 75, + 445, + 468, + 537 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Proposed Methods", + "text_level": 1, + "bbox": [ + 76, + 551, + 259, + 569 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this work, we propose a History-inspired Navigation Policy (HiNL) framework to estimate navigation states effectively. As demonstrated in Figure 2, in each step, the agent first adopts a visual feature extractor to process the current RGB observation. Then, the agent estimates the current navigation state via a History-aware State Estimation (HaSE) module and predicts a navigation action. At last, we introduce a History-based State Regularization (HbSR) to constrain the correlations among navigation states during training, thus achieving low-correlated navigation states and highly effective navigation policy.", + "bbox": [ + 75, + 577, + 470, + 744 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Task Definition and Setup", + "text_level": 1, + "bbox": [ + 76, + 755, + 313, + 772 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Following previous works [13, 14, 49, 53], in the objectgoal visual navigation task, an RGB monocular camera is the only sensor available to an agent. Prior knowledge of the entire environment, such as the topological map and 3D meshes, is unknown to the agent. An environment is composed of discrete grids, and the agent selects one of six actions, i.e., MoveAhead, RotateLeft, RotateRight, LookUp, LookDown, Done. Specif", + "bbox": [ + 75, + 779, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ically, the grid size is 0.25 meters, and the rotation angles of RotateLeft/RotateRight and LookUp/LookDown are $45^{\\circ}$ and $30^{\\circ}$ , respectively.", + "bbox": [ + 496, + 368, + 890, + 412 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "At the start of each episode, the environment randomly chooses one of the available object categories $T \\in \\{Sink, \\dots, Microwave\\}$ as the target class. Then, the agent is teleported to a random start position $s = \\{x, y, \\theta_h, \\theta_v\\}$ , where $x$ and $y$ stand for the coordinates, and $\\theta_h$ and $\\theta_v$ represent the horizontal angle and the vertical angle for the view of the agent. The agent aims to predict an action distribution based on a learned navigation policy and operates the action with the highest probability for navigation. Then, the agent moves to the next navigation step if it selects a moving action or terminates the episode if it selects the action Done. After finishing the episode, we regard the agent as successful when the following three conditions are met simultaneously: (i) the agent selects the ending action Done within the maximum step length of an episode; (ii) there is at least one instance of the specified category in the view of the agent; (iii) the distance between the visible target and the agent is less than the threshold, i.e., 1.5 meters. Otherwise, the episode is considered as a failure case.", + "bbox": [ + 496, + 414, + 892, + 700 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Visual Representation Extraction", + "text_level": 1, + "bbox": [ + 500, + 709, + 792, + 726 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Similar to previous works [13, 49], we first adopt a deep neural network, i.e., ResNet [19], to extract global features from an observed RGB image. Meanwhile, a detection module, e.g., Fast-RCNN [38] and DETR [6], is introduced to extract object features. Then, the global and object features are merged into a visual representation $\\boldsymbol{v}_t$ by graph neural network [13, 53] or transformer $[14]^1$ . For convenience, we adopt $\\boldsymbol{v}_t$ to denote the visual feature extracted", + "bbox": [ + 496, + 732, + 890, + 854 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "1In our experiments, we employ both CNN-based [13] and Transformer-based [14] visual representations, demonstrating our framework HiNL is agnostic against various visual representations.", + "bbox": [ + 498, + 862, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "2565", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/08a353a65aee560784829d4132136abe13f714d9d14284a6fa76579c3665ebbf.jpg", + "image_caption": [ + "Figure 3. Illustration of History-aware State Estimation (HaSE). The agent first adopts a Historical State Guidance Extraction module to estimate a state guidance that indicates the dominant historical states. Then, the agent extracts a state innovation that reflects incremental information of current observations, via a State-induced Perceptron module. At last, the agent generates the current navigation state estimation by fusing the state innovation and guidance in a History-aware Fusion module." + ], + "image_footnote": [], + "bbox": [ + 117, + 95, + 851, + 198 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "from the RGB observation at time step $t$ .", + "bbox": [ + 76, + 268, + 349, + 282 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. History-aware State Estimation", + "text_level": 1, + "bbox": [ + 76, + 291, + 356, + 306 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our History-aware State Estimation (HaSE) estimates the navigation states by eliminating the influence of dominant historical states and focusing on the present observation changes. As demonstrated in Figure 3, after obtaining the visual representation $\\boldsymbol{v}_t$ , we design a Historical State Guidance Extraction module to predict a state guidance $\\boldsymbol{g}_t$ indicating dominant historical states at time step $t$ . Then, we introduce a State-induced Perceptron module to obtain the current state innovation $\\boldsymbol{i}_t$ , which encodes the new information brought by the current visual observations and the last step action. Next, the history-aware guidance $\\boldsymbol{g}_t$ is employed to eliminate the dominant historical states and facilitate the estimation of the current navigation state $\\boldsymbol{s}_t$ via our proposed History-aware Fusion module.", + "bbox": [ + 75, + 314, + 470, + 526 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Historical State Guidance Extraction. In order to eliminate the dominant historical states in the current state estimation, we extract guidance for the dominant parts by exploring the relationships among historical navigation states. The agent roughly approximates a prior state $s_t'$ based on the last state $s_{t-1}$ , action $a_{t-1}$ , and the current visual representation $v_t$ . The prior state $s_t'$ is encoded with the last navigation state and the current observation, and thus we refer $s_t'$ to the query about the dominant historical states. Let FFN be a feed-forward network, $s_t'$ is formulated as follows,", + "bbox": [ + 75, + 532, + 468, + 698 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {s} _ {t} ^ {\\prime} = \\sigma (\\operatorname {F F N} \\left(\\boldsymbol {s} _ {t - 1}\\right) + \\operatorname {F F N} \\left(\\left[ \\boldsymbol {v} _ {t}; \\boldsymbol {a} _ {t - 1} \\right]\\right)), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 708, + 468, + 726 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $[;]$ and $\\sigma$ represent the concatenation operation and sigmoid function, respectively.", + "bbox": [ + 75, + 734, + 468, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After obtaining $s_t'$ , the agent memorizes a set of the historical navigation states $\\mathbb{S}_{t - \\tau : t - 1} = \\{s_{t - \\tau}, \\dots, s_{t - 1}\\}$ in the last $\\tau$ steps. In order to explore the temporal relationships among historical navigation states, $\\mathbb{S}_{t - \\tau : t - 1}$ are composed additively with temporal embeddings. Concretely, we adopt absolute temporal embedding [46] based on the timestamp at which a state is estimated. By doing this, the agent perceives not only the temporal orders among the last $\\tau$ navigation states, but also the length of the current", + "bbox": [ + 75, + 763, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "episode. Therefore, the agent can exploit temporal information to eliminate the distant navigation states in extracting the guidance. In addition, we analyze the impacts of different temporal embedding methods (i.e., no/learnable/relative temporal embedding) in Table 2.", + "bbox": [ + 496, + 268, + 890, + 343 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Afterward, we pass $s_t'$ and $\\mathbb{S}_{t - \\tau : t - 1}$ through the historical state decoder, i.e., a cross-attention module. We refer to the output $z_t$ as the dominant historical states, and calculate it as follows,", + "bbox": [ + 496, + 343, + 890, + 402 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {z} _ {\\boldsymbol {t}} = \\operatorname {D e c o d e r} _ {\\mathrm {h i s}} \\left(\\boldsymbol {s} _ {\\boldsymbol {t}} ^ {\\prime}, \\mathbb {S} _ {t - \\tau : t - 1}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 412, + 890, + 429 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathrm{Decoder}_{\\mathrm{his}}(\\cdot, \\cdot)$ indicates the historical state decoder, the first argument for query and the other for key/values. Then, we formulate the guidance $g_{t}$ as the difference between $s_{t-1}$ and $z_{t}$ . Therefore, we compute $g_{t}$ as follows,", + "bbox": [ + 496, + 438, + 890, + 500 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {g} _ {t} = \\sigma \\left(\\boldsymbol {s} _ {t - 1} - \\boldsymbol {z} _ {t}\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 624, + 513, + 890, + 529 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In addition, we set $\\tau$ as 5 in our experiments and demonstrate the impacts of different $\\tau$ in Figure 5b.", + "bbox": [ + 496, + 534, + 890, + 565 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "State-induced Perceptron. Apart from eliminating the dominant historical states, we also aim to encourage the agent to emphasize the observed changes in the current state estimation. Therefore, we extract a state innovation $i_{t}$ based on $v_{t}$ to encode the observed changes at the current step. To be specific, we projects $v_{t}, s_{t-1}$ and $a_{t-1}$ to $i_{t}$ . Inspired by LSTM [21], we compute $i_{t}$ as,", + "bbox": [ + 496, + 570, + 890, + 676 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {i} _ {t} = \\tanh (\\operatorname {F F N} \\left(\\boldsymbol {s} _ {t - 1}\\right) + \\operatorname {F F N} ([ \\boldsymbol {v} _ {t}; \\boldsymbol {a} _ {t - 1} ])) \\odot \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 539, + 684, + 890, + 708 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\sigma (\\operatorname {F F N} \\left(\\boldsymbol {s} _ {\\boldsymbol {t} - \\mathbf {1}}\\right) + \\operatorname {F F N} ([ \\boldsymbol {v} _ {\\boldsymbol {t}}; \\boldsymbol {a} _ {\\boldsymbol {t} - \\mathbf {1}} ])),\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 703, + 815, + 720 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where tanh represents hyperbolic tangent function, and $\\odot$ stands for the element-wise product.", + "bbox": [ + 500, + 728, + 890, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "History-aware Fusion. After obtaining the guidance $g_{t}$ and state innovation $i_{t}$ , we aim to estimate the current navigation state $s_{t}$ . In order to eliminate the influence of dominant navigation states, the agent first merges $s_{t-1}$ and $g_{t}$ by multiplication. Then the agent adds $i_{t}$ to the result of multiplication to incorporate the observation changes. Formally, we formulate the current navigation state $s_{t}$ estimation as follows,", + "bbox": [ + 496, + 763, + 890, + 883 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {s} _ {t} = \\tanh \\left(\\boldsymbol {s} _ {t - 1} \\odot \\boldsymbol {g} _ {t} + \\boldsymbol {i} _ {t}\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 598, + 885, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "2566", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2b9bc7c5947ecaf5265841f8714f8173c6bdf25d7fd47c5518cc3d60154f60cf.jpg", + "image_caption": [ + "Figure 4. Illustration of History-based State Regularization (HbSR). (a) First-order Markov state transition that is widely adopted by LSTM based navigation policy learning. Those methods will produce highly-correlated among states. (b) Our HbSR constrains higher-order relations among states, and leads to low-correlated states. As the states are more informative (i.e., less correlated), our navigation policy is more effective and efficiency." + ], + "image_footnote": [], + "bbox": [ + 99, + 89, + 867, + 191 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Benefiting from HaSE, the agent can effectively update the navigation states by being alert to the present observation and neglecting the influence of distant states.", + "bbox": [ + 75, + 262, + 468, + 306 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. History-based State Regularization", + "text_level": 1, + "bbox": [ + 76, + 316, + 382, + 333 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Previous reinforcement learning-based object-goal visual navigation methods are designed based on the first-order Markov assumption, while there are no explicit constraints on the correlations among navigation states. We observe that high relevance among navigation states exhibit in the experiments of previous methods [13, 14] and this is also one of the reasons that prior works yield failure cases. For instance, an agent is stuck in front of an object rather than adopting a rotation action or trapped in a loop. From these failure cases, we can see that the navigation states are highly correlated as the adopted actions are periodically the same. In order to avoid inefficient navigation caused by highly-relevant navigation states, we aim to exert an explicit constraint on the correlations among navigation states.", + "bbox": [ + 75, + 339, + 468, + 549 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We introduce a novel History-based State Regularization (HbSR) to reduce the correlations among navigation states from the perspective of the training objective. In order to measure the correlations among navigation states, we adopt the Pearson correlation coefficient [17]. To be specific, after the agent terminates an episode with $T$ steps, we obtain a set of navigation states $\\mathbb{S}_{0:T} = \\{s_0,\\dots,s_T\\}$ . Given two navigation states $s_i$ and $s_j$ , we first compute the covariance of the two states as follows,", + "bbox": [ + 75, + 551, + 468, + 686 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {C o v} \\left(\\boldsymbol {s} _ {i}, \\boldsymbol {s} _ {j}\\right) = \\mathbb {E} \\left(\\left(\\boldsymbol {s} _ {i} - \\mathbb {E} \\left(\\boldsymbol {s} _ {i}\\right)\\right) \\left(\\boldsymbol {s} _ {j} - \\mathbb {E} \\left(\\boldsymbol {s} _ {j}\\right)\\right)\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 116, + 698, + 468, + 715 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then, we compute the correlation coefficient $\\rho$ for two states based on the covariance formulated as,", + "bbox": [ + 76, + 724, + 468, + 753 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\rho \\left(\\boldsymbol {s} _ {i}, \\boldsymbol {s} _ {j}\\right) = \\frac {\\operatorname {C o v} \\left(\\boldsymbol {s} _ {i} , \\boldsymbol {s} _ {j}\\right)}{\\sqrt {\\operatorname {V a r} \\left(\\boldsymbol {s} _ {i}\\right) \\operatorname {V a r} \\left(\\boldsymbol {s} _ {j}\\right)}}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 160, + 763, + 468, + 800 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\operatorname{Var}(s_i)$ denotes the variance of $s_i$ .", + "bbox": [ + 76, + 810, + 354, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The correlation coefficient is expressed in a range $[-1, + 1]$ . Two states are highly relevant when their correlation is close to either 1 or $-1$ , whereas $\\rho = 0$ indicates irrelevant between two states. Therefore, in order to eliminate high relevance among states, we enforce the correla", + "bbox": [ + 76, + 825, + 468, + 898 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "tions of states to be close to 0. In particular, since inefficient navigation is generally caused by positive correlations of navigation states, we only apply the proposed regularization on correlations with positive values. According to the first-order Markov property, the present state $s_t$ depends on its immediately preceding state $s_{t-1}$ . Thus, we define a regularization-free threshold $\\xi$ . Moreover, in HbSR, if a temporal interval between two states is less than the threshold $\\xi$ , we do not apply the correlation regularization among consecutive states. This is because navigation has continuity and neighboring states will inherently have relevance while distant states should be irrelevant. Hence, our HbSR $l^{hsr}$ is expressed as,", + "bbox": [ + 496, + 262, + 890, + 458 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nl ^ {h s r} = \\mathbb {E} _ {\\boldsymbol {s} \\in \\mathbb {S}} \\left(\\sum_ {t = \\xi + 1} ^ {T} \\sum_ {i = 0} ^ {t - \\xi - 1} \\max \\left(0, \\rho \\left(\\boldsymbol {s} _ {t}, \\boldsymbol {s} _ {i}\\right)\\right)\\right). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 549, + 464, + 890, + 507 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Training Details", + "text_level": 1, + "bbox": [ + 500, + 522, + 661, + 537 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We employ the Asynchronous Advantage Actor-Critic (A3C) Algorithm [30] to train the agent, similar to previous works [13, 14, 49, 53]. We adopt two loss objectives, i.e., policy loss $l_{\\text{policy}}$ and value loss $l_{\\text{value}}$ , for training the navigation policy and value estimation networks, respectively. Furthermore, since the calculation of correlation coefficients is differentiable [34], we can employ the loss $l^{hsr}$ of HbSR in training. Formally, the total loss $L$ is formulated as the summation of $l_{\\text{policy}}$ , $l_{\\text{value}}$ and $l^{hsr}$ ,", + "bbox": [ + 496, + 545, + 890, + 681 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nl = \\lambda_ {0} l ^ {\\text {p o l i c y}} + \\lambda_ {1} l ^ {\\text {v a l u e}} + \\lambda_ {2} l ^ {\\text {h s r}}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 686, + 890, + 705 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_0, \\lambda_1$ and $\\lambda_2$ are weighted factors to balance three losses. In our experiments, we set these three weighted factors as 1, 0.5 and 1, respectively.", + "bbox": [ + 496, + 710, + 890, + 757 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 768, + 633, + 786 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To validate the effectiveness and efficiency of the proposed HiNL, we conduct extensive experiments on two artificial environments, i.e., iTHOR [26] and RoboTHOR [11].", + "bbox": [ + 496, + 794, + 890, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Protocols and Experimental details", + "text_level": 1, + "bbox": [ + 500, + 847, + 803, + 863 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset. We adopt iTHOR [26], an artificial environment within the AI2-THOR framework, to evaluate our method.", + "bbox": [ + 500, + 869, + 890, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "2567", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/d79a42b10661fd52a12b712556f8447ee68f0ecb9f154c987134af0355a0a8db.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodiTHOR [26]RoboTHOR [11]
AllL>5AllL>5
SuccessSPLSuccessSPLSuccessSPLSuccessSPL
Random8.0±1.33.6±0.60.3±0.10.1±0.14.0±1.01.6±0.40.2±0.10.1±0.1
WE33.0±3.514.7±1.821.4±3.011.7±1.97.0±1.22.2±0.64.7±0.91.7±0.3
SP [50]35.1±1.315.5±1.122.2±2.711.4±1.610.9±0.54.2±0.57.2±0.63.2±0.5
SAVN [49]40.8±1.216.1±0.528.7±1.513.9±0.510.2±1.03.9±0.96.9±0.82.8±0.6
ORG [13]65.3±0.737.5±0.154.8±1.036.1±0.145.4±0.821.2±1.239.6±0.718.5±0.8
ORG+TPN [13]69.3±1.239.4±1.060.7±1.338.6±1.147.8±1.223.1±0.942.9±0.920.3±1.0
HOZ [53]70.6±1.740.0±1.262.8±1.739.2±0.644.3±1.819.8±1.337.2±1.215.4±0.8
TCN [41]75.1±0.847.7±0.965.3±1.442.1±0.844.4±1.626.4±1.538.4±1.122.0±1.2
VTNet [14]72.2±1.044.9±0.763.4±1.144.0±0.953.2±1.127.5±1.747.0±0.823.3±0.7
HiNL †71.2±0.838.9±0.964.1±0.637.0±0.556.1±1.326.4±0.851.8±0.923.9±0.7
HiNL ‡80.1±1.449.8±1.974.6±1.747.6±1.460.6±1.030.8±1.256.2±0.826.6±0.9
", + "bbox": [ + 96, + 88, + 874, + 320 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Comparison with existing methods. We report the average Success rate (%) and SPL (%) in iTHOR [26] and RoboTHOR [11] as well as their variances in subscripts by repeating experiments five times. $L > 5$ represents the episodes that require at least 5 steps. † and ‡ indicate using CNN-based and Transformer-based visual extractors (i.e., ORG [13] and VTNet [14]), respectively.", + "bbox": [ + 75, + 329, + 893, + 375 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The environment includes four types of scenes, i.e., kitchen, living room, bedroom, and bathroom. Each type of scene consists of 30 rooms with different furniture items and placements. Following [13, 14, 53], we use 22 categories as the target classes and ensure that there are at least four targets in each room. We choose the first 20 out of 30 rooms per scene as the training set and equally divide the remaining 40 rooms into the validation and test set.", + "bbox": [ + 75, + 386, + 470, + 507 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Furthermore, we conduct experiments in RoboTHOR [11], the other synthetic environment in the AI2-THOR framework. Different from iTHOR, each scene in RoboTHOR is separated by several clapboards. RoboTHOR consists of 89 scenes, while only 75 apartments are published for train/val (60 for training and 15 for validation). Therefore, we select the first 55 of 60 original training scenes as our training dataset and use the rest 5 scenes as the validation set. Then, we adopt the remaining 15 original validation apartments as our test set. We report the test result of the model with the highest success rate on the validation set for both iTHOR and RoboTHOR.", + "bbox": [ + 75, + 507, + 468, + 688 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation metrics. We adopt two evaluation metrics, i.e., success rate and Success weighted by Path Length (SPL), to assess our method performance. Success rate is introduced to evaluate the navigation effectiveness and calculated by $\\frac{1}{N}\\sum_{n = 0}^{N}S_{n}$ , where $N$ is the number of episodes and $S_{n}$ represents a binary success indicator of the $n$ -th episode. Meanwhile, SPL measures the efficiency of navigation trajectories. Let $L_{n}$ be the length of the $n$ -th episode and $L_{opt}$ indicate the length of the optimal path. SPL is formulated as $\\frac{1}{N}\\sum_{n = 0}^{N}S_{n}\\frac{L_{n}}{\\max(L_{n},L_{opt})}$ .", + "bbox": [ + 75, + 695, + 470, + 851 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details. Following [14], we train the VT for 20 epochs with the supervision of expert experience. Then, we train the navigation policy for 2M episodes in to-", + "bbox": [ + 75, + 854, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tal with 32 asynchronous agents. In making a robot learn to approach the target, the penalty is set to $1e^{-3}$ for every step that passes prior to arrival, and thus an agent is encouraged to navigate as quickly as possible. Meanwhile, to encourage the agent complete navigation effectively, we offer a large reward of 5 when the agent finishes a trajectory successfully. We adopt Adam [25] with a learning rate $10^{-4}$ . Our codes will be publicly related for reproducibility.", + "bbox": [ + 496, + 386, + 893, + 508 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparison with Existing Methods", + "text_level": 1, + "bbox": [ + 498, + 518, + 808, + 536 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.1 Competing Methods", + "text_level": 1, + "bbox": [ + 500, + 542, + 696, + 558 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare our method with the following ones: Random policy. An agent makes decisions based on a uniform action probability. Thus, the movements and stops of the agent are random. Scene Prior (SP) [50] utilizes the scene prior knowledge and category associations for navigation with a graph neural network on the FastText database [23]. Word Embedding (WE) makes use of GloVe embedding [35] to signify the target category instead of relying on detection. From trial and error, an agent establishes the connection between object appearances and GloVe embeddings. Self-adaptive Visual Navigation (SAVN) [49] presents a meta reinforcement learning method to assist an agent in adapting to unseen environments during inference. Object Relationship Graph (ORG) [13] proposes a visual representation learning method to encode relationships among categories via a graph and design a tentative policy network (TPN) to escape from deadlocks during testing. Hierarchical Object-to-Zone (HOZ) [53] introduces a graph to guide an agent in a coarse-to-fine manner and a online-learning mechanism to update the graph. Time-Contrastive Networks (TCN) [41] presents a self-supervised method, deriving robotic behaviors and representations from unlabeled", + "bbox": [ + 496, + 568, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "2568", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/96713046f49535485a148b09d71a2a753335d7eb0f23e82333d0546b96d1daf5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodAllL>5
SuccessSPLSuccessSPL
HaSEHiNL † w/o HaSE68.137.860.335.9
HiNL †71.238.964.137.0
HiNL ‡ w/o HaSE77.945.771.643.5
HiNL ‡80.149.874.647.6
HiNL ‡ w/ tem.78.539.871.239.1
HiNL ‡ w/ lea. tem.79.145.873.143.8
HiNL ‡ w/ rel. tem.77.344.169.441.6
HiNL ‡80.149.874.647.6
HbSRHiNL † w/o HbSR66.838.157.835.8
HiNL †71.238.964.137.0
HiNL ‡ w/o HbSR75.446.967.043.8
HiNL ‡80.149.874.647.6
HiNL ‡ neg. & pos.75.446.967.143.8
HiNL ‡80.149.874.647.6
", + "bbox": [ + 81, + 88, + 465, + 337 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Ablation study on different components of HiNL in iTHOR [26] environment. † and ‡ indicate using CNN-based and Transformer-based visual extractors (i.e., ORG [13] and VT-Net [14]), respectively. HiNL w/o tem. indicates HiNL without temporal embedding in HaSE, while HiNL w/ lea. tem. and HiNL w/ rel. tem. represent the adoption of learnable and relative temporal embedding in HiNL, respectively. HiNL neg. & pos. stands for regularizing both positive and negative correlations toward 0.", + "bbox": [ + 75, + 349, + 467, + 460 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "videos captured across multiple viewpoints, leveraging metric learning loss to establish viewpoint-invariant representations that encapsulate end-effector relationships. Visual Transformer for Navigation (VTNet) [14] explores the spatial correlations among objects and observation regions for navigation via a transformer-based network.", + "bbox": [ + 75, + 474, + 467, + 565 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2.2 Quantitative Results", + "text_level": 1, + "bbox": [ + 76, + 574, + 272, + 588 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To demonstrate the superiority of our proposed learning framework, we select two different visual representation extractors: one is based on graph network (i.e., ORG [13]), and the other is based on the transformer (i.e., VT [14]). Table 1 indicates that both HiNL $\\dagger$ and HiNL $\\ddagger$ achieve superior results compared with their original methods (i.e., ORG and VTNet). This demonstrates that given different visual representations, HiNL can improve navigation performance constantly by explicitly exploring the relationships among historical states.", + "bbox": [ + 75, + 598, + 467, + 747 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To be specific, HiNL $\\ddagger$ with the transformer-based extractor surpasses the existing methods by a large margin on both success rate $(+7.9\\% / + 7.4\\%)$ and SPL $(+4.9\\% / + 3.3\\%)$ in iTHOR [26]/RoboTHOR [11]. Adopting ORG as the visual representation extractor, HiNL $\\dagger$ achieves significant improvement over the original ORG by nearly $10\\%$ in success rate in both environments. VTNet and ORG only adopt LSTM to merge well-designed visual representations into navigation states without any explicit constraints on states. The correlations among navigation states are not exploited.", + "bbox": [ + 75, + 750, + 467, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "This experiment suggests that our HiNL leads to instructive navigation states, and thus significantly improves the effectiveness and efficiency of our navigation system.", + "bbox": [ + 498, + 90, + 890, + 136 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As indicated in Table 1, we observe that HiNL $\\dagger$ significantly outperforms ORG+TPN [13]. Although TPN employs an auxiliary network to fine-tune the initial navigation policy network when the agent is stuck in deadlocks during inference, TPN does not guarantee navigation states to be low relevant. Consequently, an agent might be trapped in looping actions, such as moving with rotating, because of the high correlations among navigation states. HiNL solves this issue by reducing the correlations among navigation states, and thus achieves proper actions for navigation. Furthermore, adopting TCN-based loss (i.e., TCN) yields inferior performance with $5\\%$ decrease in success rate compared to HiNL. This is because TCN tries to pull temporally neighboring features closer while consecutive navigation states might be very different due to a rotation action. Thus, TCN may overly emphasize the reliance among neighboring states, thus leading to inferior state estimation. In addition, our method achieves more than two times the success rate and SPL than SAVN in both environments. Benefiting from well-designed visual representations and lowering relevance among navigation states, HiNL achieves better navigation performance in testing even without adjusting the navigation policy. This also implies that our navigation states significantly improve the effectiveness and efficiency of navigation in unseen environments.", + "bbox": [ + 496, + 137, + 890, + 513 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2.3 Case Study", + "text_level": 1, + "bbox": [ + 500, + 523, + 633, + 537 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 5a illustrates trajectories of three navigation episodes proceeded by VTNet [14] and HiNL $\\ddagger$ . In the first case, VTNet fails to find the target and is stuck in the environment, as seen in the first column. On the contrary, HiNL $\\ddagger$ navigates toward the object successfully. This implies that using HiNL, our agents can react rapidly to the observed changes. In the second case, VTNet moves forward while rotating and thus leads to inefficient navigation. In contrast, the navigation system trained with HiNL avoids rotating and approaches the target with fewer steps. This demonstrates that our HiNL improves navigation efficiency by reducing the correlations among states. In the final case, VTNet moves forward consistently until it reaches the wall. Benefiting from our HaSE, the agent can get rid of dominant historical states and react rapidly to the observed changes.", + "bbox": [ + 496, + 547, + 890, + 773 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 786, + 653, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impacts of different components. To illustrate the impacts of different components, we conduct experiments on adopting one of the proposed components (i.e., HaSE and HbSR). Furthermore, we report the navigation performance with different visual representation extractors, i.e., ORG and VT. As indicated in Table 2, both HiNL $\\dagger$ and", + "bbox": [ + 496, + 810, + 890, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "2569", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cdeace07492ab6745fbb0a8855c160ce4cad1f47817895e77376c8826c92988b.jpg", + "image_caption": [ + "Figure 5. (a) Visual results in the simulated environments iTHOR. The arrow color changes (i.e., from blue to red) represent the navigation progress (i.e., from beginning to end). Black arrows indicate rotations. The target objects are highlighted by the green/red boxes, where a green box stands for a success episode and a red box represents a failure case. (b) Impacts of different historical state length $\\tau \\in [2,10]$ . (c) Results of different regularization-free states $\\xi \\in [1,8]$ . We set $\\tau = 5$ and $\\xi = 1$ and highlight them by the black vertical dotted lines in (b) and (c)." + ], + "image_footnote": [], + "bbox": [ + 84, + 85, + 599, + 277 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/432a840037cb021724dc457044c1bad233f0286c843ae753da6e3c10bb90da0c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 633, + 74, + 870, + 277 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "HiNL $\\ddagger$ achieve gains in navigation performance on success rate $(+2.9\\% / + 2.2\\%)$ and SPL $(+0.9\\% / + 4.1\\%)$ by introducing HaSE. Furthermore, HiNL $\\dagger$ and HiNL $\\ddagger$ outperform models without HbSR (i.e., HiNL $\\dagger/\\mathrm{HiNL}\\ddagger$ w/o HbSR), respectively. This manifests that both of our proposed components improve the performance of navigation.", + "bbox": [ + 75, + 368, + 468, + 460 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impacts of different temporal embedding in HaSE. Table 2 compares different temporal embedding methods, (i.e., w/o temporal embedding, w/ relative temporal embedding and w/ absolute temporal embedding (HiNL)) and highlights the effectiveness of incorporating temporal embedding in historical states. HiNL with absolute temporal embedding significantly improves success rates and SPL on iTHOR compared to no embedding (HiNL $\\ddagger$ w/o tem.), learnable embedding (HiNL $\\ddagger$ w/ lea. tem.), and relative embedding (HiNL $\\ddagger$ w/ rel. tem.). This improvement, especially in navigation efficiency, stems from providing the agent with the current episode length.", + "bbox": [ + 75, + 465, + 468, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Comparison of different $\\tau$ in HaSE. To illustrate the influence of different historical state lengths, we conduct experiments with $\\tau \\in [2,10]$ . As demonstrated in Figure 5b, we observe that as historical states become too many, our agent may fail to converge to an optimal policy. On the other hand, estimating navigation states with few historical states do not have sufficient capacity to effectively update states. The agent reaches the peak of success rate and SPL when it adopts 5 historical states in HaSE.", + "bbox": [ + 75, + 652, + 468, + 789 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impacts of different $\\xi$ in HbSR. To study the impacts of adopting different regularization-free thresholds $\\xi$ , we conduct experiments with $\\xi \\in [1,8]$ in iTHOR. As seen in Figure 5c, increasing $\\xi$ drops performance clearly. Both success rate (80% to 76%) and SPL (49.8% to 45.8%) decrease when $\\xi$ increases from 1 to 8. Since two navigation states generally have high correlations when they are close", + "bbox": [ + 75, + 795, + 468, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "in temporal, raising $\\xi$ indicates ignoring constraints on these correlations. As a result, the effects of HbSR are compromised, and the navigation performance is dropped.", + "bbox": [ + 498, + 368, + 890, + 414 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Comparison of regularization on different values. To analyze the impacts of regularization on different values, we demonstrate the comparisons of applying HbSR on (i) both negative and positive correlations and (ii) only positive correlations, as indicated in Table 2. Compared to HiNL trained on regularizing the correlations based on both positive and negative values (HiNL $\\ddagger$ neg. & pos.), HiNL improves performance by neglecting the negative correlations. This comparison suggests that inefficient navigation is generally caused by positively correlated navigation states.", + "bbox": [ + 498, + 416, + 892, + 568 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 580, + 617, + 595 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we proposed an effective navigation state learning approach for visual navigation, named History-inspired Navigation Policy Learning (HiNL). Our HiNL involves two innovate parts: a history-aware state estimation module (HaSE) and a history-based state regularization (HbSR). HbSR eliminates the negative impacts of dominant historical states and enables current states to reflect newly observed visual information. HbSR significantly reduces the correlations among navigation states, thus leading to more informative states as well as effective actions. Extensive results demonstrate that our HiNL significantly improves navigation performance.", + "bbox": [ + 496, + 606, + 890, + 787 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 800, + 660, + 816 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This research is funded in part by ARC-Discovery grant (DP220100800 to XY), ARC-DECRA grant (DE230100477 to XY) and ARC Centres of Excellence grant (CE200100025). We thank all anonymous reviewers and ACs for their constructive suggestions.", + "bbox": [ + 498, + 825, + 890, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "2570", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Ziad Al-Halah, Santhosh Kumar Ramakrishnan, and Kristen Grauman. Zero experience required: Plug & play modular transfer learning for semantic visual navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17031-17041, 2022. 2", + "[2] Bastian Alt, Adrian Šošić, and Heinz Koeppl. Correlation priors for reinforcement learning. Advances in Neural Information Processing Systems, 32, 2019. 2", + "[3] Johann Borenstein and Yoram Koren. Real-time obstacle avoidance for fast mobile robots. IEEE Transactions on systems, Man, and Cybernetics, 19(5):1179-1187, 1989. 2", + "[4] Johann Borenstein and Yoram Koren. The vector field histogram-fast obstacle avoidance for mobile robots. IEEE transactions on robotics and automation, 7(3):278-288, 1991. 2", + "[5] Tommaso Campari, Leonardo Lamanna, Paolo Traverso, Luciano Serafini, and Lamberto Ballan. Online learning of reusable abstract models for object goal navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14870-14879, 2022. 2", + "[6] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 3", + "[7] Changan Chen, Ziad Al-Halah, and Kristen Grauman. Semantic audio-visual navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15516–15525, 2021. 2", + "[8] Changan Chen, Unnat Jain, Carl Schissler, Sebastia Vicenc Amengual Gari, Ziad Al-Halah, Vamsi Krishna Ithapu, Philip Robinson, and Kristen Grauman. Soundspaces: Audio-visual navigation in 3d environments. In European Conference on Computer Vision, pages 17–36. Springer, 2020. 2", + "[9] Jinyu Chen, Chen Gao, Erli Meng, Qiong Zhang, and Si Liu. Reinforced structured state-evolution for vision-language navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15450-15459, 2022. 2", + "[10] Kevin Chen, Juan Pablo de Vicente, Gabriel Sepulveda, Fei Xia, Alvaro Soto, Marynel Vázquez, and Silvio Savarese. A behavioral approach to visual navigation with graph localization networks. arXiv preprint arXiv:1903.00445, 2019. 2", + "[11] Matt Deitke, Winson Han, Alvaro Herrasti, Aniruddha Kembhavi, Eric Kolve, Roozbeh Mottaghi, Jordi Salvador, Dustin Schwenk, Eli VanderBilt, Matthew Wallingford, et al. Robothor: An open simulation-to-real embodied ai platform. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3164-3174, 2020. 2, 5, 6, 7", + "[12] MWM Gamini Dissanayake, Paul Newman, Steve Clark, Hugh F Durrant-Whyte, and Michael Csorba. A solution to the simultaneous localization and map building (slam) problem. IEEE Transactions on robotics and automation, 17(3):229-241, 2001. 2" + ], + "bbox": [ + 78, + 114, + 468, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Heming Du, Xin Yu, and Liang Zheng. Learning object relation graph and tentative policy for visual navigation. arXiv preprint arXiv:2007.11018, 2020. 1, 2, 3, 5, 6, 7", + "[14] Heming Du, Xin Yu, and Liang Zheng. Vtnet: Visual transformer network for object goal navigation. arXiv preprint arXiv:2105.09447, 2021. 1, 2, 3, 5, 6, 7", + "[15] Kuan Fang, Alexander Toshev, Li Fei-Fei, and Silvio Savarese. Scene memory transformer for embodied agents in long-horizon tasks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 538–547, 2019. 2", + "[16] Xiaoyu Feng, Heming Du, Yueqi Duan, Yongpan Liu, and Hehe Fan. Sefformer: Structure embedding transformer for 3d object detection. arXiv preprint arXiv:2209.01745, 2022.2", + "[17] David Freedman, Robert Pisani, and Roger Purves. Statistics (international student edition). Pisani, R. Purves, 4th edn. WW Norton & Company, New York, 2007. 5", + "[18] Chuang Gan, Yiwei Zhang, Jiajun Wu, Boqing Gong, and Joshua B Tenenbaum. Look, listen, and act: Towards audio-visual embodied navigation. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 9701–9707. IEEE, 2020. 2", + "[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3", + "[20] Todd Hester and Peter Stone. Texlore: real-time sample-efficient reinforcement learning for robots. Machine learning, 90(3):385-429, 2013. 2", + "[21] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation, 9(8):1735-1780, 1997. 4", + "[22] Meihuizi Jia, Lei Shen, Xin Shen, Lejian Liao, Meng Chen, Xiaodong He, Zhendong Chen, and Jiaqi Li. Mnerqq: An end-to-end mrc framework for multimodal named entity recognition with query grounding. arXiv preprint arXiv:2211.14739, 2022. 2", + "[23] Armand Joulin, Edouard Grave, Piotr Bojanowski, and Tomas Mikolov. Bag of tricks for efficient text classification. arXiv preprint arXiv:1607.01759, 2016. 6", + "[24] Apoorv Khandelwal, Luca Weihs, Roozbeh Mottaghi, and Aniruddha Kembhavi. Simple but effective: Clip embeddings for embodied ai. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14829-14838, 2022. 2", + "[25] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6", + "[26] Eric Kolve, Roozbeh Mottaghi, Winson Han, Eli VanderBilt, Luca Weihs, Alvaro Herrasti, Daniel Gordon, Yuke Zhu, Abhinav Gupta, and Ali Farhadi. AI2-THOR: An Interactive 3D Environment for Visual AI. arXiv, 2017. 2, 5, 6, 7", + "[27] Bar Mayo, Tamir Hazan, and Ayellet Tal. Visual navigation with spatial attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16898-16907, 2021. 2" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "2571", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Lina Mezghan, Sainbayar Sukhbaatar, Thibaut Lavril, Oleksandr Maksymets, Dhruv Batra, Piotr Bojanowski, and Karteek Alahari. Memory-augmented reinforcement learning for image-goal navigation. In 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 3316-3323. IEEE, 2022. 2", + "[29] Piotr Mirowski, Razvan Pascanu, Fabio Viola, Hubert Soyer, Andrew J Ballard, Andrea Banino, Misha Denil, Ross Goroshin, Laurent Sifre, Koray Kavukcuoglu, et al. Learning to navigate in complex environments. arXiv preprint arXiv:1611.03673, 2016. 2", + "[30] Volodymyr Mnih, Adria Puigdomenech Badia, Mehdi Mirza, Alex Graves, Timothy Lillicrap, Tim Harley, David Silver, and Koray Kavukcuoglu. Asynchronous methods for deep reinforcement learning. In International conference on machine learning, pages 1928-1937, 2016. 5", + "[31] Arsalan Mousavian, Alexander Toshev, Marek Fiser, Jana Košecka, Ayzaan Wahid, and James Davidson. Visual representations for semantic target driven navigation. In 2019 International Conference on Robotics and Automation (ICRA), pages 8846-8852. IEEE, 2019. 2", + "[32] Giuseppe Oriolo, Marilena Vendittelli, and Giovanni Ulivi. On-line map building and navigation for autonomous mobile robots. In Proceedings of 1995 IEEE International Conference on Robotics and Automation, volume 3, pages 2900-2906. IEEE, 1995. 2", + "[33] Emilio Parisotto and Ruslan Salakhutdinov. Neural map: Structured memory for deep reinforcement learning. In International Conference on Learning Representations, 2018. 2", + "[34] Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. Automatic differentiation in pytorch. 2017. 5", + "[35] Jeffrey Pennington, Richard Socher, and Christopher D Manning. Glove: Global vectors for word representation. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 1532-1543, 2014. 2, 6", + "[36] Santhosh Kumar Ramakrishnan, Devendra Singh Chaplot, Ziad Al-Halah, Jitendra Malik, and Kristen Grauman. Poni: Potential functions for objectgoal navigation with interaction-free learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18890-18900, 2022. 2", + "[37] Pravesh Ranchod, Benjamin Rosman, and George Konidaris. Nonparametric bayesian reward segmentation for skill discovery using inverse reinforcement learning. In 2015 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 471-477. IEEE, 2015. 2", + "[38] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 3", + "[39] Nikolay Savinov, Alexey Dosovitskiy, and Vladlen Koltun. Semi-parametric topological memory for navigation. arXiv preprint arXiv:1803.00653, 2018. 2" + ], + "bbox": [ + 78, + 90, + 470, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Gabriel Sepulveda, Juan Carlos Niebles, and Alvaro Soto. A deep learning based behavioral approach to indoor autonomous navigation. In 2018 IEEE International Conference on Robotics and Automation (ICRA), pages 4646-4653. IEEE, 2018. 2", + "[41] Pierre Sermanet, Corey Lynch, Yevgen Chebotar, Jasmine Hsu, Eric Jang, Stefan Schaal, Sergey Levine, and Google Brain. Time-contrastive networks: Self-supervised learning from video. In 2018 IEEE international conference on robotics and automation (ICRA), pages 1134–1141. IEEE, 2018. 3, 6", + "[42] Adrian Šošić, Abdelhak M Zoubir, and Heinz Koeppl. A bayesian approach to policy recognition and state representation learning. IEEE transactions on pattern analysis and machine intelligence, 40(6):1295-1308, 2017. 2", + "[43] Adrian Šošic, Abdelhak M Zoubir, Elmar Rueckert, Jan Peters, and Heinz Koeppl. Inverse reinforcement learning via nonparametric spatio-temporal subgoal modeling. Journal of Machine Learning Research, 19(69):1-45, 2018. 2", + "[44] Tianqi Tang, Heming Du, Xin Yu, and Yi Yang. Monocular camera-based point-goal navigation by learning depth channel and cross-modality pyramid fusion. 2022. 2", + "[45] Yurun Tian, Xin Yu, Bin Fan, Fuchao Wu, Huub Heijnen, and Vassileios Balntas. Sosnet: Second order similarity regularization for local descriptor learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11016-11025, 2019. 2", + "[46] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 4", + "[47] Su Wang, Ceslee Montgomery, Jordi Orbay, Vighnesh Birodkar, Aleksandra Faust, Izzeddin Gur, Natasha Jaques, Austin Waters, Jason Baldridge, and Peter Anderson. Less is more: Generating grounded navigation instructions from landmarks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15428-15438, 2022. 2", + "[48] Erik Wijmans, Abhishek Kadian, Ari Morcos, Stefan Lee, Irfan Essa, Devi Parikh, Manolis Savva, and Dhruv Batra. Dd-ppo: Learning near-perfect pointgoal navigators from 2.5 billion frames. arXiv preprint arXiv:1911.00357, 2019. 2", + "[49] Mitchell Wortsman, Kiana Ehsani, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Learning to learn how to learn: Self-adaptive visual navigation using meta-learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6750-6759, 2019. 1, 2, 3, 5, 6", + "[50] Wei Yang, Xiaolong Wang, Ali Farhadi, Abhinav Gupta, and Roozbeh Mottaghi. Visual semantic navigation using scene priors. arXiv preprint arXiv:1810.06543, 2018. 1, 2, 6", + "[51] Xin Yu and Fatih Porikli. Ultra-resolving face images by discriminative generative networks. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part V, pages 318-333. Springer, 2016. 2", + "[52] Amir R Zamir, Alexander Sax, William Shen, Leonidas J Guibas, Jitendra Malik, and Silvio Savarese. Taskonomy:" + ], + "bbox": [ + 501, + 92, + 892, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "2572", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Disentangling task transfer learning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3712-3722, 2018. 2", + "[53] Sixian Zhang, Xinhang Song, Yubing Bai, Weijie Li, Yakui Chu, and Shuqiang Jiang. Hierarchical object-to-zone graph for object navigation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15130-15140, 2021. 2, 3, 5, 6", + "[54] Yuke Zhu, Roozbeh Mottaghi, Eric Kolve, Joseph J Lim, Abhinav Gupta, Li Fei-Fei, and Ali Farhadi. Target-driven visual navigation in indoor scenes using deep reinforcement learning. In 2017 IEEE international conference on robotics and automation (ICRA), pages 3357-3364. IEEE, 2017. 2" + ], + "bbox": [ + 78, + 90, + 468, + 273 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "2573", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Object-Goal Visual Navigation via Effective Exploration of Relations Among Historical Navigation States/02126feb-2cfd-431e-b8b7-664314fe1580_model.json b/2023/Object-Goal Visual Navigation via Effective Exploration of Relations Among Historical Navigation States/02126feb-2cfd-431e-b8b7-664314fe1580_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8d73992b08dff5337b032ca5b61abdae161e8f7a --- /dev/null +++ b/2023/Object-Goal Visual Navigation via Effective Exploration of Relations Among Historical Navigation States/02126feb-2cfd-431e-b8b7-664314fe1580_model.json @@ -0,0 +1,2136 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.134, + 0.131, + 0.836, + 0.177 + ], + "angle": 0, + "content": "Object-Goal Visual Navigation via Effective Exploration of Relations among Historical Navigation States" + }, + { + "type": "text", + "bbox": [ + 0.287, + 0.203, + 0.687, + 0.221 + ], + "angle": 0, + "content": "Heming Du\\(^{1,3}\\), Lincheng Li\\(^{2}\\), Zi Huang\\(^{3}\\), Xin Yu\\(^{3*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.136, + 0.221, + 0.829, + 0.24 + ], + "angle": 0, + "content": "\\(^{1}\\) Australian National University \\(^{2}\\) Netease Fuxi AI Lab \\(^{3}\\) The University of Queensland" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.274, + 0.314, + 0.29 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.307, + 0.474, + 0.671 + ], + "angle": 0, + "content": "Object-goal visual navigation aims at steering an agent toward an object via a series of moving steps. Previous works mainly focus on learning informative visual representations for navigation, but overlook the impacts of navigation states on the effectiveness and efficiency of navigation. We observe that high relevance among navigation states will cause navigation inefficiency or failure for existing methods. In this paper, we present a History-inspired Navigation Policy Learning (HiNL) framework to estimate navigation states effectively by exploring relationships among historical navigation states. In HiNL, we propose a History-aware State Estimation (HaSE) module to alleviate the impacts of dominant historical states on the current state estimation. Meanwhile, HaSE also encourages an agent to be alert to the current observation changes, thus enabling the agent to make valid actions. Furthermore, we design a History-based State Regularization (HbSR) to explicitly suppress the correlation among navigation states in training. As a result, our agent can update states more effectively while reducing the correlations among navigation states. Experiments on the artificial platform AI2-THOR (i.e., iTHOR and RoboTHOR) demonstrate that HiNL significantly outperforms state-of-the-art methods on both Success Rate and SPL in unseen testing environments." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.699, + 0.21, + 0.715 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.725, + 0.47, + 0.861 + ], + "angle": 0, + "content": "Object-goal visual navigation is to direct an agent to move consecutively toward an object of a specific category. Without knowing the environment map beforehand, at each navigation step, an agent first needs to represent its visual observations, then estimate its navigation states from the visual representations and the preceding states, and at last predict the corresponding action. Therefore, to achieve an effective and efficient navigation system, learning instructive visual representations and navigation states is critical." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.862, + 0.469, + 0.877 + ], + "angle": 0, + "content": "Prevailing visual navigation works [13, 14, 50] focus on" + }, + { + "type": "image", + "bbox": [ + 0.543, + 0.276, + 0.857, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.499, + 0.432, + 0.892, + 0.468 + ], + "angle": 0, + "content": "(a) Demonstration of inefficient action predictions caused by highlycorrelated navigation states. Our agent is stuck by an obstacle, i.e., lowprofile sofa, and repeatedly predicts an invalid action, i.e., MoveAhead." + }, + { + "type": "image", + "bbox": [ + 0.554, + 0.47, + 0.834, + 0.566 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.5, + 0.567, + 0.892, + 0.604 + ], + "angle": 0, + "content": "(b) Demonstration of the correlation coefficients among navigation states trained in two manners. Navigation states estimated via LSTM are highly-relevant. In contrast, our HiNL produce low-correlated navigation states." + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.612, + 0.892, + 0.64 + ], + "angle": 0, + "content": "Figure 1. Motivation of our proposed History-inspired Navigation Learning (HiNL) framework." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.672, + 0.892, + 0.883 + ], + "angle": 0, + "content": "extracting informative visual representations, while some methods [13, 49] adjust navigation policy during inference. All these approaches commonly employ recurrent neural networks (e.g., LSTM) to estimate navigation states. However, we observe that the navigation states of existing methods [13, 14, 49] exhibit high relevance, as demonstrated in Figure 1b, and the highly-correlated navigation states would lead to inefficient navigation policy (i.e., failure to respond to observation changes rapidly). For instance, as shown in Figure 1a, an agent is stuck by the low-profile sofa and fails to take proper actions to circumvent the obstacle. Hence, we aim to endow an agent with the capability of updating its navigation states effectively while avoiding producing highly-correlated states." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "In this work, we propose a History-inspired Navigation" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.888, + 0.221, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2563" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.288 + ], + "angle": 0, + "content": "Learning (HiNL) framework to obtain informative navigation states by exploiting the relationships among historical navigation states. HiNL consists of two novel components: (i) a History-aware State Estimation (HaSE) module, and (ii) a History-based State Regularization (HbSR). Here, our HaSE module is designed to generate a state that can be promptly updated according to visual observations. Specifically, HaSE first analyzes the correlations among historical navigation states and then eliminates the influence of dominant historical states on the current state estimation. As a result, an agent is able to predict navigation states which can dynamically react to the current visual observations and then make sensible navigation actions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.29, + 0.473, + 0.563 + ], + "angle": 0, + "content": "Furthermore, existing reinforcement learning-based object-goal navigation systems [13, 14, 49] often assume the navigation state transition exhibits the first-order Markov property. This would allow the emergence of high correlations among navigation states, leading to inferior navigation policy. To address this issue, we introduce an explicit constraint on the correlation among all the states, namely History-based State Regularization (HbSR). To be specific, HbSR enforces to relevance (i.e., correlations) between a state and all its preceding states (except its previous state) to be low. Here, we do not constrain states of two consecutive steps because temporally close states generally have relevance in practice considering the navigation continuity. After training with our HbSR, the correlations among the navigation states become much lower (see Figure 1b). This pheromone also indicates HiNL effectively updates states. Hence, our navigation system can respond to observation changes adaptively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.565, + 0.473, + 0.684 + ], + "angle": 0, + "content": "To demonstrate the superiority of HiNL, we conduct experiments in the widely-adopted artificial environment iTHOR [26] and RoboTHOR [11]. HiNL outperforms the state-of-the-art by a large margin. To be specific, we improve the Success Rate (SR) from \\(72.2\\%\\) to \\(80.1\\%\\) and Success weighted by Path Length (SPL) from 0.449 to 0.498 in iTHOR. Overall, our major contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.695, + 0.47, + 0.74 + ], + "angle": 0, + "content": "- We propose a History-inspired Navigation Policy (HiNL) framework to effectively estimate navigation states by utilizing historical states." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.753, + 0.47, + 0.829 + ], + "angle": 0, + "content": "- We design a History-aware State Estimation (HaSE) to eliminate dominant historical states in the current state estimation. Therefore, the agent reduces the impact of distant navigation states on the state estimation, and thus reacts dynamically to the observation changes." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.84, + 0.47, + 0.901 + ], + "angle": 0, + "content": "- We introduce a History-based State Regularization (HbSR) to explicitly constrain the correlations among navigation states. By doing this, the agent can effectively update navigation states with low relevance." + }, + { + "type": "list", + "bbox": [ + 0.097, + 0.695, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.09, + 0.651, + 0.107 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.117, + 0.895, + 0.254 + ], + "angle": 0, + "content": "Visual navigation. Traditional works [3, 4, 32] often leverage an entire environment map for navigation and divide the task into three parts: mapping, localization, and path planning. However, environment maps are generally unavailable in unseen environments. Dissanayake et al. [12] adopt simultaneous localization and mapping (SLAM) to infer robot positions. Campari et al. [5] learn agent states via a Taskonomy model bank [52], but they need an RGB-D sensor to construct an online map during navigation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.256, + 0.895, + 0.468 + ], + "angle": 0, + "content": "Recently, due to significant advancements in deep learning [16,22,45,51], reinforcement learning-based navigation methods [28, 29, 31, 33, 54] take visual observations as inputs and predict navigation actions. Vision-Language Navigation (VLN) approaches [9, 10, 15, 39, 40] steer an agent to the target based on its visual observations and navigation guidance in natural language. Similar to VLN, point-goal visual navigation methods [44,48] aim at driving an agent to a given point with step-wise directional indications. Moreover, audio-visual navigation methods [7, 8, 18] utilize additional audio signals to move a robot to the target position. Al-Halah et al. [1] propose a transfer learning model for multiple navigation tasks by embedding various navigation goals, e.g., image, sketch, and audio." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.47, + 0.896, + 0.741 + ], + "angle": 0, + "content": "Our work falls in the field of object-goal visual navigation [27, 31, 36, 47, 50]. However, existing object-goal navigation methods mainly focus on representing visual features comprehensively while we investigate the impact of navigation states on navigation performance. Wortsman et al. [49] exploit word embedding (i.e., GloVe embedding [35]) to represent the target category and introduce a meta network that mimics a reward function during inference. Du et al. [13] introduce an object relation graph, dubbed ORG, to encode visual observations. They also design a tentative policy for deadlock avoidance and adjust the navigation policy in unseen testing environments. The Hierarchical Object-to-Zone (HOZ) graph [53] offers coarse-to-fine guidance based on real-time updates. Additionally, VT-Net [14] incorporates object and region features with location cues, and EmbCLIP [24] leverages the contrastive language image pretraining encoder for visual navigation tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Correlation Modeling in Reinforcement Learning. Several methods [37, 43] explore correlations in hidden Markov models for inverse reinforcement learning (IRL). For action prediction, Hester et al. [20] propose Texlore to model correlations within the transition dynamics via a random forest. Vsovsic et al. [42] introduce a Bayesian approach to learn policy from demonstrations of experts by capturing correlations among actions. Alt et al. [2] design a Bayesian learning framework to establish temporal and spatial correlations among actions. Furthermore," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2564" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.114, + 0.105, + 0.868, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.272, + 0.895, + 0.345 + ], + "angle": 0, + "content": "Figure 2. Our History-inspired Navigation Policy Learning (HiNL) framework. HiNL takes visual representations as input and outputs navigation actions. HiNL involves two innovative parts: a History-aware State Estimation (HaSE) module and a History-based State Regularization (HbSR). HaSE is proposed to estimate navigation states that can reflect current observation changes from the perspective of network design, while HbSR is designed to enforce the informativeness of states from the view of the training objective. Both of them help to achieve effective and efficient navigation policy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.369, + 0.471, + 0.445 + ], + "angle": 0, + "content": "Sermanet [41] propose a self-supervised TCN for learning robotic behaviors and representations from unlabeled multi-viewpoint videos. TCN uses a metric learning loss to create viewpoint-invariant representations that capture relationships between end-effectors." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.446, + 0.47, + 0.538 + ], + "angle": 0, + "content": "Different from previous methods that establish correlations among actions, our HiNL aims to eliminate high relevance among navigation states. To the best of our knowledge, our work HiNL is the first attempt to explicitly leverage the correlations among navigation states for object-goal navigation and achieves superior navigation performance." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.552, + 0.26, + 0.57 + ], + "angle": 0, + "content": "3. Proposed Methods" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.578, + 0.471, + 0.746 + ], + "angle": 0, + "content": "In this work, we propose a History-inspired Navigation Policy (HiNL) framework to estimate navigation states effectively. As demonstrated in Figure 2, in each step, the agent first adopts a visual feature extractor to process the current RGB observation. Then, the agent estimates the current navigation state via a History-aware State Estimation (HaSE) module and predicts a navigation action. At last, we introduce a History-based State Regularization (HbSR) to constrain the correlations among navigation states during training, thus achieving low-correlated navigation states and highly effective navigation policy." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.756, + 0.314, + 0.773 + ], + "angle": 0, + "content": "3.1. Task Definition and Setup" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Following previous works [13, 14, 49, 53], in the objectgoal visual navigation task, an RGB monocular camera is the only sensor available to an agent. Prior knowledge of the entire environment, such as the topological map and 3D meshes, is unknown to the agent. An environment is composed of discrete grids, and the agent selects one of six actions, i.e., MoveAhead, RotateLeft, RotateRight, LookUp, LookDown, Done. Specif" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.369, + 0.892, + 0.414 + ], + "angle": 0, + "content": "ically, the grid size is 0.25 meters, and the rotation angles of RotateLeft/RotateRight and LookUp/LookDown are \\(45^{\\circ}\\) and \\(30^{\\circ}\\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.415, + 0.893, + 0.701 + ], + "angle": 0, + "content": "At the start of each episode, the environment randomly chooses one of the available object categories \\( T \\in \\{Sink, \\dots, Microwave\\} \\) as the target class. Then, the agent is teleported to a random start position \\( s = \\{x, y, \\theta_h, \\theta_v\\} \\), where \\( x \\) and \\( y \\) stand for the coordinates, and \\( \\theta_h \\) and \\( \\theta_v \\) represent the horizontal angle and the vertical angle for the view of the agent. The agent aims to predict an action distribution based on a learned navigation policy and operates the action with the highest probability for navigation. Then, the agent moves to the next navigation step if it selects a moving action or terminates the episode if it selects the action Done. After finishing the episode, we regard the agent as successful when the following three conditions are met simultaneously: (i) the agent selects the ending action Done within the maximum step length of an episode; (ii) there is at least one instance of the specified category in the view of the agent; (iii) the distance between the visible target and the agent is less than the threshold, i.e., 1.5 meters. Otherwise, the episode is considered as a failure case." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.71, + 0.794, + 0.727 + ], + "angle": 0, + "content": "3.2. Visual Representation Extraction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.733, + 0.892, + 0.856 + ], + "angle": 0, + "content": "Similar to previous works [13, 49], we first adopt a deep neural network, i.e., ResNet [19], to extract global features from an observed RGB image. Meanwhile, a detection module, e.g., Fast-RCNN [38] and DETR [6], is introduced to extract object features. Then, the global and object features are merged into a visual representation \\( \\boldsymbol{v}_t \\) by graph neural network [13, 53] or transformer \\( [14]^1 \\). For convenience, we adopt \\( \\boldsymbol{v}_t \\) to denote the visual feature extracted" + }, + { + "type": "page_footnote", + "bbox": [ + 0.499, + 0.863, + 0.892, + 0.901 + ], + "angle": 0, + "content": "1In our experiments, we employ both CNN-based [13] and Transformer-based [14] visual representations, demonstrating our framework HiNL is agnostic against various visual representations." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2565" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.096, + 0.852, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.199, + 0.894, + 0.255 + ], + "angle": 0, + "content": "Figure 3. Illustration of History-aware State Estimation (HaSE). The agent first adopts a Historical State Guidance Extraction module to estimate a state guidance that indicates the dominant historical states. Then, the agent extracts a state innovation that reflects incremental information of current observations, via a State-induced Perceptron module. At last, the agent generates the current navigation state estimation by fusing the state innovation and guidance in a History-aware Fusion module." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.269, + 0.351, + 0.284 + ], + "angle": 0, + "content": "from the RGB observation at time step \\( t \\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.292, + 0.357, + 0.308 + ], + "angle": 0, + "content": "3.3. History-aware State Estimation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.315, + 0.471, + 0.527 + ], + "angle": 0, + "content": "Our History-aware State Estimation (HaSE) estimates the navigation states by eliminating the influence of dominant historical states and focusing on the present observation changes. As demonstrated in Figure 3, after obtaining the visual representation \\( \\boldsymbol{v}_t \\), we design a Historical State Guidance Extraction module to predict a state guidance \\( \\boldsymbol{g}_t \\) indicating dominant historical states at time step \\( t \\). Then, we introduce a State-induced Perceptron module to obtain the current state innovation \\( \\boldsymbol{i}_t \\), which encodes the new information brought by the current visual observations and the last step action. Next, the history-aware guidance \\( \\boldsymbol{g}_t \\) is employed to eliminate the dominant historical states and facilitate the estimation of the current navigation state \\( \\boldsymbol{s}_t \\) via our proposed History-aware Fusion module." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.534, + 0.47, + 0.699 + ], + "angle": 0, + "content": "Historical State Guidance Extraction. In order to eliminate the dominant historical states in the current state estimation, we extract guidance for the dominant parts by exploring the relationships among historical navigation states. The agent roughly approximates a prior state \\( s_t' \\) based on the last state \\( s_{t-1} \\), action \\( a_{t-1} \\), and the current visual representation \\( v_t \\). The prior state \\( s_t' \\) is encoded with the last navigation state and the current observation, and thus we refer \\( s_t' \\) to the query about the dominant historical states. Let FFN be a feed-forward network, \\( s_t' \\) is formulated as follows," + }, + { + "type": "equation", + "bbox": [ + 0.132, + 0.709, + 0.47, + 0.727 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {s} _ {t} ^ {\\prime} = \\sigma (\\operatorname {F F N} \\left(\\boldsymbol {s} _ {t - 1}\\right) + \\operatorname {F F N} \\left(\\left[ \\boldsymbol {v} _ {t}; \\boldsymbol {a} _ {t - 1} \\right]\\right)), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.47, + 0.765 + ], + "angle": 0, + "content": "where \\([;]\\) and \\(\\sigma\\) represent the concatenation operation and sigmoid function, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.902 + ], + "angle": 0, + "content": "After obtaining \\( s_t' \\), the agent memorizes a set of the historical navigation states \\( \\mathbb{S}_{t - \\tau : t - 1} = \\{s_{t - \\tau}, \\dots, s_{t - 1}\\} \\) in the last \\( \\tau \\) steps. In order to explore the temporal relationships among historical navigation states, \\( \\mathbb{S}_{t - \\tau : t - 1} \\) are composed additively with temporal embeddings. Concretely, we adopt absolute temporal embedding [46] based on the timestamp at which a state is estimated. By doing this, the agent perceives not only the temporal orders among the last \\( \\tau \\) navigation states, but also the length of the current" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.269, + 0.892, + 0.344 + ], + "angle": 0, + "content": "episode. Therefore, the agent can exploit temporal information to eliminate the distant navigation states in extracting the guidance. In addition, we analyze the impacts of different temporal embedding methods (i.e., no/learnable/relative temporal embedding) in Table 2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.344, + 0.892, + 0.404 + ], + "angle": 0, + "content": "Afterward, we pass \\( s_t' \\) and \\( \\mathbb{S}_{t - \\tau : t - 1} \\) through the historical state decoder, i.e., a cross-attention module. We refer to the output \\( z_t \\) as the dominant historical states, and calculate it as follows," + }, + { + "type": "equation", + "bbox": [ + 0.588, + 0.413, + 0.892, + 0.43 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {z} _ {\\boldsymbol {t}} = \\operatorname {D e c o d e r} _ {\\mathrm {h i s}} \\left(\\boldsymbol {s} _ {\\boldsymbol {t}} ^ {\\prime}, \\mathbb {S} _ {t - \\tau : t - 1}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.439, + 0.892, + 0.5 + ], + "angle": 0, + "content": "where \\(\\mathrm{Decoder}_{\\mathrm{his}}(\\cdot, \\cdot)\\) indicates the historical state decoder, the first argument for query and the other for key/values. Then, we formulate the guidance \\(g_{t}\\) as the difference between \\(s_{t-1}\\) and \\(z_{t}\\). Therefore, we compute \\(g_{t}\\) as follows," + }, + { + "type": "equation", + "bbox": [ + 0.625, + 0.515, + 0.891, + 0.53 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {g} _ {t} = \\sigma \\left(\\boldsymbol {s} _ {t - 1} - \\boldsymbol {z} _ {t}\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.535, + 0.892, + 0.566 + ], + "angle": 0, + "content": "In addition, we set \\(\\tau\\) as 5 in our experiments and demonstrate the impacts of different \\(\\tau\\) in Figure 5b." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.571, + 0.892, + 0.678 + ], + "angle": 0, + "content": "State-induced Perceptron. Apart from eliminating the dominant historical states, we also aim to encourage the agent to emphasize the observed changes in the current state estimation. Therefore, we extract a state innovation \\( i_{t} \\) based on \\( v_{t} \\) to encode the observed changes at the current step. To be specific, we projects \\( v_{t}, s_{t-1} \\) and \\( a_{t-1} \\) to \\( i_{t} \\). Inspired by LSTM [21], we compute \\( i_{t} \\) as," + }, + { + "type": "equation", + "bbox": [ + 0.54, + 0.685, + 0.892, + 0.709 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {i} _ {t} = \\tanh (\\operatorname {F F N} \\left(\\boldsymbol {s} _ {t - 1}\\right) + \\operatorname {F F N} ([ \\boldsymbol {v} _ {t}; \\boldsymbol {a} _ {t - 1} ])) \\odot \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.575, + 0.704, + 0.816, + 0.721 + ], + "angle": 0, + "content": "\\[\n\\sigma (\\operatorname {F F N} \\left(\\boldsymbol {s} _ {\\boldsymbol {t} - \\mathbf {1}}\\right) + \\operatorname {F F N} ([ \\boldsymbol {v} _ {\\boldsymbol {t}}; \\boldsymbol {a} _ {\\boldsymbol {t} - \\mathbf {1}} ])),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.729, + 0.892, + 0.76 + ], + "angle": 0, + "content": "where tanh represents hyperbolic tangent function, and \\(\\odot\\) stands for the element-wise product." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.884 + ], + "angle": 0, + "content": "History-aware Fusion. After obtaining the guidance \\( g_{t} \\) and state innovation \\( i_{t} \\), we aim to estimate the current navigation state \\( s_{t} \\). In order to eliminate the influence of dominant navigation states, the agent first merges \\( s_{t-1} \\) and \\( g_{t} \\) by multiplication. Then the agent adds \\( i_{t} \\) to the result of multiplication to incorporate the observation changes. Formally, we formulate the current navigation state \\( s_{t} \\) estimation as follows," + }, + { + "type": "equation", + "bbox": [ + 0.599, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {s} _ {t} = \\tanh \\left(\\boldsymbol {s} _ {t - 1} \\odot \\boldsymbol {g} _ {t} + \\boldsymbol {i} _ {t}\\right). \\tag {5}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2566" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.09, + 0.868, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.193, + 0.894, + 0.248 + ], + "angle": 0, + "content": "Figure 4. Illustration of History-based State Regularization (HbSR). (a) First-order Markov state transition that is widely adopted by LSTM based navigation policy learning. Those methods will produce highly-correlated among states. (b) Our HbSR constrains higher-order relations among states, and leads to low-correlated states. As the states are more informative (i.e., less correlated), our navigation policy is more effective and efficiency." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.263, + 0.47, + 0.308 + ], + "angle": 0, + "content": "Benefiting from HaSE, the agent can effectively update the navigation states by being alert to the present observation and neglecting the influence of distant states." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.317, + 0.383, + 0.334 + ], + "angle": 0, + "content": "3.4. History-based State Regularization" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.34, + 0.47, + 0.55 + ], + "angle": 0, + "content": "Previous reinforcement learning-based object-goal visual navigation methods are designed based on the first-order Markov assumption, while there are no explicit constraints on the correlations among navigation states. We observe that high relevance among navigation states exhibit in the experiments of previous methods [13, 14] and this is also one of the reasons that prior works yield failure cases. For instance, an agent is stuck in front of an object rather than adopting a rotation action or trapped in a loop. From these failure cases, we can see that the navigation states are highly correlated as the adopted actions are periodically the same. In order to avoid inefficient navigation caused by highly-relevant navigation states, we aim to exert an explicit constraint on the correlations among navigation states." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.552, + 0.47, + 0.687 + ], + "angle": 0, + "content": "We introduce a novel History-based State Regularization (HbSR) to reduce the correlations among navigation states from the perspective of the training objective. In order to measure the correlations among navigation states, we adopt the Pearson correlation coefficient [17]. To be specific, after the agent terminates an episode with \\( T \\) steps, we obtain a set of navigation states \\( \\mathbb{S}_{0:T} = \\{s_0,\\dots,s_T\\} \\). Given two navigation states \\( s_i \\) and \\( s_j \\), we first compute the covariance of the two states as follows," + }, + { + "type": "equation", + "bbox": [ + 0.117, + 0.699, + 0.469, + 0.716 + ], + "angle": 0, + "content": "\\[\n\\operatorname {C o v} \\left(\\boldsymbol {s} _ {i}, \\boldsymbol {s} _ {j}\\right) = \\mathbb {E} \\left(\\left(\\boldsymbol {s} _ {i} - \\mathbb {E} \\left(\\boldsymbol {s} _ {i}\\right)\\right) \\left(\\boldsymbol {s} _ {j} - \\mathbb {E} \\left(\\boldsymbol {s} _ {j}\\right)\\right)\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.726, + 0.469, + 0.755 + ], + "angle": 0, + "content": "Then, we compute the correlation coefficient \\(\\rho\\) for two states based on the covariance formulated as," + }, + { + "type": "equation", + "bbox": [ + 0.161, + 0.765, + 0.469, + 0.801 + ], + "angle": 0, + "content": "\\[\n\\rho \\left(\\boldsymbol {s} _ {i}, \\boldsymbol {s} _ {j}\\right) = \\frac {\\operatorname {C o v} \\left(\\boldsymbol {s} _ {i} , \\boldsymbol {s} _ {j}\\right)}{\\sqrt {\\operatorname {V a r} \\left(\\boldsymbol {s} _ {i}\\right) \\operatorname {V a r} \\left(\\boldsymbol {s} _ {j}\\right)}}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.811, + 0.356, + 0.825 + ], + "angle": 0, + "content": "where \\(\\operatorname{Var}(s_i)\\) denotes the variance of \\(s_i\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.826, + 0.469, + 0.9 + ], + "angle": 0, + "content": "The correlation coefficient is expressed in a range \\([-1, + 1]\\). Two states are highly relevant when their correlation is close to either 1 or \\(-1\\), whereas \\(\\rho = 0\\) indicates irrelevant between two states. Therefore, in order to eliminate high relevance among states, we enforce the correla" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.263, + 0.892, + 0.459 + ], + "angle": 0, + "content": "tions of states to be close to 0. In particular, since inefficient navigation is generally caused by positive correlations of navigation states, we only apply the proposed regularization on correlations with positive values. According to the first-order Markov property, the present state \\( s_t \\) depends on its immediately preceding state \\( s_{t-1} \\). Thus, we define a regularization-free threshold \\( \\xi \\). Moreover, in HbSR, if a temporal interval between two states is less than the threshold \\( \\xi \\), we do not apply the correlation regularization among consecutive states. This is because navigation has continuity and neighboring states will inherently have relevance while distant states should be irrelevant. Hence, our HbSR \\( l^{hsr} \\) is expressed as," + }, + { + "type": "equation", + "bbox": [ + 0.55, + 0.465, + 0.892, + 0.508 + ], + "angle": 0, + "content": "\\[\nl ^ {h s r} = \\mathbb {E} _ {\\boldsymbol {s} \\in \\mathbb {S}} \\left(\\sum_ {t = \\xi + 1} ^ {T} \\sum_ {i = 0} ^ {t - \\xi - 1} \\max \\left(0, \\rho \\left(\\boldsymbol {s} _ {t}, \\boldsymbol {s} _ {i}\\right)\\right)\\right). \\tag {8}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.523, + 0.662, + 0.539 + ], + "angle": 0, + "content": "3.5. Training Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.546, + 0.892, + 0.683 + ], + "angle": 0, + "content": "We employ the Asynchronous Advantage Actor-Critic (A3C) Algorithm [30] to train the agent, similar to previous works [13, 14, 49, 53]. We adopt two loss objectives, i.e., policy loss \\( l_{\\text{policy}} \\) and value loss \\( l_{\\text{value}} \\), for training the navigation policy and value estimation networks, respectively. Furthermore, since the calculation of correlation coefficients is differentiable [34], we can employ the loss \\( l^{hsr} \\) of HbSR in training. Formally, the total loss \\( L \\) is formulated as the summation of \\( l_{\\text{policy}} \\), \\( l_{\\text{value}} \\) and \\( l^{hsr} \\)," + }, + { + "type": "equation", + "bbox": [ + 0.579, + 0.688, + 0.891, + 0.706 + ], + "angle": 0, + "content": "\\[\nl = \\lambda_ {0} l ^ {\\text {p o l i c y}} + \\lambda_ {1} l ^ {\\text {v a l u e}} + \\lambda_ {2} l ^ {\\text {h s r}}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.712, + 0.891, + 0.758 + ], + "angle": 0, + "content": "where \\(\\lambda_0, \\lambda_1\\) and \\(\\lambda_2\\) are weighted factors to balance three losses. In our experiments, we set these three weighted factors as 1, 0.5 and 1, respectively." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.77, + 0.634, + 0.787 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.891, + 0.84 + ], + "angle": 0, + "content": "To validate the effectiveness and efficiency of the proposed HiNL, we conduct extensive experiments on two artificial environments, i.e., iTHOR [26] and RoboTHOR [11]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.848, + 0.805, + 0.864 + ], + "angle": 0, + "content": "4.1. Protocols and Experimental details" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Dataset. We adopt iTHOR [26], an artificial environment within the AI2-THOR framework, to evaluate our method." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2567" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.098, + 0.089, + 0.875, + 0.321 + ], + "angle": 0, + "content": "
MethodiTHOR [26]RoboTHOR [11]
AllL>5AllL>5
SuccessSPLSuccessSPLSuccessSPLSuccessSPL
Random8.0±1.33.6±0.60.3±0.10.1±0.14.0±1.01.6±0.40.2±0.10.1±0.1
WE33.0±3.514.7±1.821.4±3.011.7±1.97.0±1.22.2±0.64.7±0.91.7±0.3
SP [50]35.1±1.315.5±1.122.2±2.711.4±1.610.9±0.54.2±0.57.2±0.63.2±0.5
SAVN [49]40.8±1.216.1±0.528.7±1.513.9±0.510.2±1.03.9±0.96.9±0.82.8±0.6
ORG [13]65.3±0.737.5±0.154.8±1.036.1±0.145.4±0.821.2±1.239.6±0.718.5±0.8
ORG+TPN [13]69.3±1.239.4±1.060.7±1.338.6±1.147.8±1.223.1±0.942.9±0.920.3±1.0
HOZ [53]70.6±1.740.0±1.262.8±1.739.2±0.644.3±1.819.8±1.337.2±1.215.4±0.8
TCN [41]75.1±0.847.7±0.965.3±1.442.1±0.844.4±1.626.4±1.538.4±1.122.0±1.2
VTNet [14]72.2±1.044.9±0.763.4±1.144.0±0.953.2±1.127.5±1.747.0±0.823.3±0.7
HiNL †71.2±0.838.9±0.964.1±0.637.0±0.556.1±1.326.4±0.851.8±0.923.9±0.7
HiNL ‡80.1±1.449.8±1.974.6±1.747.6±1.460.6±1.030.8±1.256.2±0.826.6±0.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.33, + 0.895, + 0.375 + ], + "angle": 0, + "content": "Table 1. Comparison with existing methods. We report the average Success rate (%) and SPL (%) in iTHOR [26] and RoboTHOR [11] as well as their variances in subscripts by repeating experiments five times. \\( L > 5 \\) represents the episodes that require at least 5 steps. † and ‡ indicate using CNN-based and Transformer-based visual extractors (i.e., ORG [13] and VTNet [14]), respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.387, + 0.471, + 0.508 + ], + "angle": 0, + "content": "The environment includes four types of scenes, i.e., kitchen, living room, bedroom, and bathroom. Each type of scene consists of 30 rooms with different furniture items and placements. Following [13, 14, 53], we use 22 categories as the target classes and ensure that there are at least four targets in each room. We choose the first 20 out of 30 rooms per scene as the training set and equally divide the remaining 40 rooms into the validation and test set." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.508, + 0.47, + 0.689 + ], + "angle": 0, + "content": "Furthermore, we conduct experiments in RoboTHOR [11], the other synthetic environment in the AI2-THOR framework. Different from iTHOR, each scene in RoboTHOR is separated by several clapboards. RoboTHOR consists of 89 scenes, while only 75 apartments are published for train/val (60 for training and 15 for validation). Therefore, we select the first 55 of 60 original training scenes as our training dataset and use the rest 5 scenes as the validation set. Then, we adopt the remaining 15 original validation apartments as our test set. We report the test result of the model with the highest success rate on the validation set for both iTHOR and RoboTHOR." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.696, + 0.471, + 0.852 + ], + "angle": 0, + "content": "Evaluation metrics. We adopt two evaluation metrics, i.e., success rate and Success weighted by Path Length (SPL), to assess our method performance. Success rate is introduced to evaluate the navigation effectiveness and calculated by \\(\\frac{1}{N}\\sum_{n = 0}^{N}S_{n}\\), where \\(N\\) is the number of episodes and \\(S_{n}\\) represents a binary success indicator of the \\(n\\)-th episode. Meanwhile, SPL measures the efficiency of navigation trajectories. Let \\(L_{n}\\) be the length of the \\(n\\)-th episode and \\(L_{opt}\\) indicate the length of the optimal path. SPL is formulated as \\(\\frac{1}{N}\\sum_{n = 0}^{N}S_{n}\\frac{L_{n}}{\\max(L_{n},L_{opt})}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Implementation details. Following [14], we train the VT for 20 epochs with the supervision of expert experience. Then, we train the navigation policy for 2M episodes in to-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.387, + 0.895, + 0.51 + ], + "angle": 0, + "content": "tal with 32 asynchronous agents. In making a robot learn to approach the target, the penalty is set to \\(1e^{-3}\\) for every step that passes prior to arrival, and thus an agent is encouraged to navigate as quickly as possible. Meanwhile, to encourage the agent complete navigation effectively, we offer a large reward of 5 when the agent finishes a trajectory successfully. We adopt Adam [25] with a learning rate \\(10^{-4}\\). Our codes will be publicly related for reproducibility." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.52, + 0.81, + 0.537 + ], + "angle": 0, + "content": "4.2. Comparison with Existing Methods" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.544, + 0.697, + 0.559 + ], + "angle": 0, + "content": "4.2.1 Competing Methods" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.569, + 0.895, + 0.903 + ], + "angle": 0, + "content": "We compare our method with the following ones: Random policy. An agent makes decisions based on a uniform action probability. Thus, the movements and stops of the agent are random. Scene Prior (SP) [50] utilizes the scene prior knowledge and category associations for navigation with a graph neural network on the FastText database [23]. Word Embedding (WE) makes use of GloVe embedding [35] to signify the target category instead of relying on detection. From trial and error, an agent establishes the connection between object appearances and GloVe embeddings. Self-adaptive Visual Navigation (SAVN) [49] presents a meta reinforcement learning method to assist an agent in adapting to unseen environments during inference. Object Relationship Graph (ORG) [13] proposes a visual representation learning method to encode relationships among categories via a graph and design a tentative policy network (TPN) to escape from deadlocks during testing. Hierarchical Object-to-Zone (HOZ) [53] introduces a graph to guide an agent in a coarse-to-fine manner and a online-learning mechanism to update the graph. Time-Contrastive Networks (TCN) [41] presents a self-supervised method, deriving robotic behaviors and representations from unlabeled" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2568" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.467, + 0.338 + ], + "angle": 0, + "content": "
MethodAllL>5
SuccessSPLSuccessSPL
HaSEHiNL † w/o HaSE68.137.860.335.9
HiNL †71.238.964.137.0
HiNL ‡ w/o HaSE77.945.771.643.5
HiNL ‡80.149.874.647.6
HiNL ‡ w/ tem.78.539.871.239.1
HiNL ‡ w/ lea. tem.79.145.873.143.8
HiNL ‡ w/ rel. tem.77.344.169.441.6
HiNL ‡80.149.874.647.6
HbSRHiNL † w/o HbSR66.838.157.835.8
HiNL †71.238.964.137.0
HiNL ‡ w/o HbSR75.446.967.043.8
HiNL ‡80.149.874.647.6
HiNL ‡ neg. & pos.75.446.967.143.8
HiNL ‡80.149.874.647.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.35, + 0.468, + 0.461 + ], + "angle": 0, + "content": "Table 2. Ablation study on different components of HiNL in iTHOR [26] environment. † and ‡ indicate using CNN-based and Transformer-based visual extractors (i.e., ORG [13] and VT-Net [14]), respectively. HiNL w/o tem. indicates HiNL without temporal embedding in HaSE, while HiNL w/ lea. tem. and HiNL w/ rel. tem. represent the adoption of learnable and relative temporal embedding in HiNL, respectively. HiNL neg. & pos. stands for regularizing both positive and negative correlations toward 0." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.476, + 0.468, + 0.566 + ], + "angle": 0, + "content": "videos captured across multiple viewpoints, leveraging metric learning loss to establish viewpoint-invariant representations that encapsulate end-effector relationships. Visual Transformer for Navigation (VTNet) [14] explores the spatial correlations among objects and observation regions for navigation via a transformer-based network." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.575, + 0.274, + 0.589 + ], + "angle": 0, + "content": "4.2.2 Quantitative Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.468, + 0.748 + ], + "angle": 0, + "content": "To demonstrate the superiority of our proposed learning framework, we select two different visual representation extractors: one is based on graph network (i.e., ORG [13]), and the other is based on the transformer (i.e., VT [14]). Table 1 indicates that both HiNL \\(\\dagger\\) and HiNL \\(\\ddagger\\) achieve superior results compared with their original methods (i.e., ORG and VTNet). This demonstrates that given different visual representations, HiNL can improve navigation performance constantly by explicitly exploring the relationships among historical states." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.468, + 0.901 + ], + "angle": 0, + "content": "To be specific, HiNL \\(\\ddagger\\) with the transformer-based extractor surpasses the existing methods by a large margin on both success rate \\((+7.9\\% / + 7.4\\%)\\) and SPL \\((+4.9\\% / + 3.3\\%)\\) in iTHOR [26]/RoboTHOR [11]. Adopting ORG as the visual representation extractor, HiNL \\(\\dagger\\) achieves significant improvement over the original ORG by nearly \\(10\\%\\) in success rate in both environments. VTNet and ORG only adopt LSTM to merge well-designed visual representations into navigation states without any explicit constraints on states. The correlations among navigation states are not exploited." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.891, + 0.137 + ], + "angle": 0, + "content": "This experiment suggests that our HiNL leads to instructive navigation states, and thus significantly improves the effectiveness and efficiency of our navigation system." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.138, + 0.892, + 0.514 + ], + "angle": 0, + "content": "As indicated in Table 1, we observe that HiNL \\(\\dagger\\) significantly outperforms ORG+TPN [13]. Although TPN employs an auxiliary network to fine-tune the initial navigation policy network when the agent is stuck in deadlocks during inference, TPN does not guarantee navigation states to be low relevant. Consequently, an agent might be trapped in looping actions, such as moving with rotating, because of the high correlations among navigation states. HiNL solves this issue by reducing the correlations among navigation states, and thus achieves proper actions for navigation. Furthermore, adopting TCN-based loss (i.e., TCN) yields inferior performance with \\(5\\%\\) decrease in success rate compared to HiNL. This is because TCN tries to pull temporally neighboring features closer while consecutive navigation states might be very different due to a rotation action. Thus, TCN may overly emphasize the reliance among neighboring states, thus leading to inferior state estimation. In addition, our method achieves more than two times the success rate and SPL than SAVN in both environments. Benefiting from well-designed visual representations and lowering relevance among navigation states, HiNL achieves better navigation performance in testing even without adjusting the navigation policy. This also implies that our navigation states significantly improve the effectiveness and efficiency of navigation in unseen environments." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.524, + 0.634, + 0.539 + ], + "angle": 0, + "content": "4.2.3 Case Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.549, + 0.892, + 0.775 + ], + "angle": 0, + "content": "Figure 5a illustrates trajectories of three navigation episodes proceeded by VTNet [14] and HiNL \\(\\ddagger\\). In the first case, VTNet fails to find the target and is stuck in the environment, as seen in the first column. On the contrary, HiNL \\(\\ddagger\\) navigates toward the object successfully. This implies that using HiNL, our agents can react rapidly to the observed changes. In the second case, VTNet moves forward while rotating and thus leads to inefficient navigation. In contrast, the navigation system trained with HiNL avoids rotating and approaches the target with fewer steps. This demonstrates that our HiNL improves navigation efficiency by reducing the correlations among states. In the final case, VTNet moves forward consistently until it reaches the wall. Benefiting from our HaSE, the agent can get rid of dominant historical states and react rapidly to the observed changes." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.787, + 0.655, + 0.803 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Impacts of different components. To illustrate the impacts of different components, we conduct experiments on adopting one of the proposed components (i.e., HaSE and HbSR). Furthermore, we report the navigation performance with different visual representation extractors, i.e., ORG and VT. As indicated in Table 2, both HiNL \\(\\dagger\\) and" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2569" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.086, + 0.6, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.634, + 0.075, + 0.871, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.292, + 0.894, + 0.361 + ], + "angle": 0, + "content": "Figure 5. (a) Visual results in the simulated environments iTHOR. The arrow color changes (i.e., from blue to red) represent the navigation progress (i.e., from beginning to end). Black arrows indicate rotations. The target objects are highlighted by the green/red boxes, where a green box stands for a success episode and a red box represents a failure case. (b) Impacts of different historical state length \\(\\tau \\in [2,10]\\). (c) Results of different regularization-free states \\(\\xi \\in [1,8]\\). We set \\(\\tau = 5\\) and \\(\\xi = 1\\) and highlight them by the black vertical dotted lines in (b) and (c)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.369, + 0.47, + 0.462 + ], + "angle": 0, + "content": "HiNL \\(\\ddagger\\) achieve gains in navigation performance on success rate \\((+2.9\\% / + 2.2\\%)\\) and SPL \\((+0.9\\% / + 4.1\\%)\\) by introducing HaSE. Furthermore, HiNL \\(\\dagger\\) and HiNL \\(\\ddagger\\) outperform models without HbSR (i.e., HiNL \\(\\dagger/\\mathrm{HiNL}\\ddagger\\) w/o HbSR), respectively. This manifests that both of our proposed components improve the performance of navigation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.466, + 0.47, + 0.649 + ], + "angle": 0, + "content": "Impacts of different temporal embedding in HaSE. Table 2 compares different temporal embedding methods, (i.e., w/o temporal embedding, w/ relative temporal embedding and w/ absolute temporal embedding (HiNL)) and highlights the effectiveness of incorporating temporal embedding in historical states. HiNL with absolute temporal embedding significantly improves success rates and SPL on iTHOR compared to no embedding (HiNL \\(\\ddagger\\) w/o tem.), learnable embedding (HiNL \\(\\ddagger\\) w/ lea. tem.), and relative embedding (HiNL \\(\\ddagger\\) w/ rel. tem.). This improvement, especially in navigation efficiency, stems from providing the agent with the current episode length." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.653, + 0.47, + 0.79 + ], + "angle": 0, + "content": "Comparison of different \\(\\tau\\) in HaSE. To illustrate the influence of different historical state lengths, we conduct experiments with \\(\\tau \\in [2,10]\\). As demonstrated in Figure 5b, we observe that as historical states become too many, our agent may fail to converge to an optimal policy. On the other hand, estimating navigation states with few historical states do not have sufficient capacity to effectively update states. The agent reaches the peak of success rate and SPL when it adopts 5 historical states in HaSE." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Impacts of different \\(\\xi\\) in HbSR. To study the impacts of adopting different regularization-free thresholds \\(\\xi\\), we conduct experiments with \\(\\xi \\in [1,8]\\) in iTHOR. As seen in Figure 5c, increasing \\(\\xi\\) drops performance clearly. Both success rate (80% to 76%) and SPL (49.8% to 45.8%) decrease when \\(\\xi\\) increases from 1 to 8. Since two navigation states generally have high correlations when they are close" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.369, + 0.892, + 0.415 + ], + "angle": 0, + "content": "in temporal, raising \\(\\xi\\) indicates ignoring constraints on these correlations. As a result, the effects of HbSR are compromised, and the navigation performance is dropped." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.417, + 0.893, + 0.569 + ], + "angle": 0, + "content": "Comparison of regularization on different values. To analyze the impacts of regularization on different values, we demonstrate the comparisons of applying HbSR on (i) both negative and positive correlations and (ii) only positive correlations, as indicated in Table 2. Compared to HiNL trained on regularizing the correlations based on both positive and negative values (HiNL \\(\\ddagger\\) neg. & pos.), HiNL improves performance by neglecting the negative correlations. This comparison suggests that inefficient navigation is generally caused by positively correlated navigation states." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.581, + 0.619, + 0.597 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.607, + 0.892, + 0.789 + ], + "angle": 0, + "content": "In this paper, we proposed an effective navigation state learning approach for visual navigation, named History-inspired Navigation Policy Learning (HiNL). Our HiNL involves two innovate parts: a history-aware state estimation module (HaSE) and a history-based state regularization (HbSR). HbSR eliminates the negative impacts of dominant historical states and enables current states to reflect newly observed visual information. HbSR significantly reduces the correlations among navigation states, thus leading to more informative states as well as effective actions. Extensive results demonstrate that our HiNL significantly improves navigation performance." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.801, + 0.661, + 0.817 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.826, + 0.892, + 0.902 + ], + "angle": 0, + "content": "This research is funded in part by ARC-Discovery grant (DP220100800 to XY), ARC-DECRA grant (DE230100477 to XY) and ARC Centres of Excellence grant (CE200100025). We thank all anonymous reviewers and ACs for their constructive suggestions." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2570" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Ziad Al-Halah, Santhosh Kumar Ramakrishnan, and Kristen Grauman. Zero experience required: Plug & play modular transfer learning for semantic visual navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17031-17041, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.185, + 0.47, + 0.227 + ], + "angle": 0, + "content": "[2] Bastian Alt, Adrian Šošić, and Heinz Koeppl. Correlation priors for reinforcement learning. Advances in Neural Information Processing Systems, 32, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.228, + 0.47, + 0.269 + ], + "angle": 0, + "content": "[3] Johann Borenstein and Yoram Koren. Real-time obstacle avoidance for fast mobile robots. IEEE Transactions on systems, Man, and Cybernetics, 19(5):1179-1187, 1989. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.271, + 0.47, + 0.325 + ], + "angle": 0, + "content": "[4] Johann Borenstein and Yoram Koren. The vector field histogram-fast obstacle avoidance for mobile robots. IEEE transactions on robotics and automation, 7(3):278-288, 1991. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.327, + 0.47, + 0.396 + ], + "angle": 0, + "content": "[5] Tommaso Campari, Leonardo Lamanna, Paolo Traverso, Luciano Serafini, and Lamberto Ballan. Online learning of reusable abstract models for object goal navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14870-14879, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.398, + 0.47, + 0.453 + ], + "angle": 0, + "content": "[6] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.454, + 0.47, + 0.509 + ], + "angle": 0, + "content": "[7] Changan Chen, Ziad Al-Halah, and Kristen Grauman. Semantic audio-visual navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15516–15525, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.51, + 0.47, + 0.591 + ], + "angle": 0, + "content": "[8] Changan Chen, Unnat Jain, Carl Schissler, Sebastia Vicenc Amengual Gari, Ziad Al-Halah, Vamsi Krishna Ithapu, Philip Robinson, and Kristen Grauman. Soundspaces: Audio-visual navigation in 3d environments. In European Conference on Computer Vision, pages 17–36. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.593, + 0.47, + 0.661 + ], + "angle": 0, + "content": "[9] Jinyu Chen, Chen Gao, Erli Meng, Qiong Zhang, and Si Liu. Reinforced structured state-evolution for vision-language navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15450-15459, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.664, + 0.47, + 0.732 + ], + "angle": 0, + "content": "[10] Kevin Chen, Juan Pablo de Vicente, Gabriel Sepulveda, Fei Xia, Alvaro Soto, Marynel Vázquez, and Silvio Savarese. A behavioral approach to visual navigation with graph localization networks. arXiv preprint arXiv:1903.00445, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.734, + 0.47, + 0.83 + ], + "angle": 0, + "content": "[11] Matt Deitke, Winson Han, Alvaro Herrasti, Aniruddha Kembhavi, Eric Kolve, Roozbeh Mottaghi, Jordi Salvador, Dustin Schwenk, Eli VanderBilt, Matthew Wallingford, et al. Robothor: An open simulation-to-real embodied ai platform. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3164-3174, 2020. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[12] MWM Gamini Dissanayake, Paul Newman, Steve Clark, Hugh F Durrant-Whyte, and Michael Csorba. A solution to the simultaneous localization and map building (slam) problem. IEEE Transactions on robotics and automation, 17(3):229-241, 2001. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "[13] Heming Du, Xin Yu, and Liang Zheng. Learning object relation graph and tentative policy for visual navigation. arXiv preprint arXiv:2007.11018, 2020. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.892, + 0.178 + ], + "angle": 0, + "content": "[14] Heming Du, Xin Yu, and Liang Zheng. Vtnet: Visual transformer network for object goal navigation. arXiv preprint arXiv:2105.09447, 2021. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.18, + 0.892, + 0.248 + ], + "angle": 0, + "content": "[15] Kuan Fang, Alexander Toshev, Li Fei-Fei, and Silvio Savarese. Scene memory transformer for embodied agents in long-horizon tasks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 538–547, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.251, + 0.892, + 0.306 + ], + "angle": 0, + "content": "[16] Xiaoyu Feng, Heming Du, Yueqi Duan, Yongpan Liu, and Hehe Fan. Sefformer: Structure embedding transformer for 3d object detection. arXiv preprint arXiv:2209.01745, 2022.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.309, + 0.892, + 0.352 + ], + "angle": 0, + "content": "[17] David Freedman, Robert Pisani, and Roger Purves. Statistics (international student edition). Pisani, R. Purves, 4th edn. WW Norton & Company, New York, 2007. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.354, + 0.892, + 0.422 + ], + "angle": 0, + "content": "[18] Chuang Gan, Yiwei Zhang, Jiajun Wu, Boqing Gong, and Joshua B Tenenbaum. Look, listen, and act: Towards audio-visual embodied navigation. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 9701–9707. IEEE, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.424, + 0.892, + 0.481 + ], + "angle": 0, + "content": "[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.483, + 0.892, + 0.525 + ], + "angle": 0, + "content": "[20] Todd Hester and Peter Stone. Texlore: real-time sample-efficient reinforcement learning for robots. Machine learning, 90(3):385-429, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.527, + 0.892, + 0.554 + ], + "angle": 0, + "content": "[21] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation, 9(8):1735-1780, 1997. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.556, + 0.892, + 0.625 + ], + "angle": 0, + "content": "[22] Meihuizi Jia, Lei Shen, Xin Shen, Lejian Liao, Meng Chen, Xiaodong He, Zhendong Chen, and Jiaqi Li. Mnerqq: An end-to-end mrc framework for multimodal named entity recognition with query grounding. arXiv preprint arXiv:2211.14739, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.628, + 0.892, + 0.67 + ], + "angle": 0, + "content": "[23] Armand Joulin, Edouard Grave, Piotr Bojanowski, and Tomas Mikolov. Bag of tricks for efficient text classification. arXiv preprint arXiv:1607.01759, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.672, + 0.892, + 0.74 + ], + "angle": 0, + "content": "[24] Apoorv Khandelwal, Luca Weihs, Roozbeh Mottaghi, and Aniruddha Kembhavi. Simple but effective: Clip embeddings for embodied ai. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14829-14838, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.743, + 0.892, + 0.785 + ], + "angle": 0, + "content": "[25] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.788, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[26] Eric Kolve, Roozbeh Mottaghi, Winson Han, Eli VanderBilt, Luca Weihs, Alvaro Herrasti, Daniel Gordon, Yuke Zhu, Abhinav Gupta, and Ali Farhadi. AI2-THOR: An Interactive 3D Environment for Visual AI. arXiv, 2017. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[27] Bar Mayo, Tamir Hazan, and Ayellet Tal. Visual navigation with spatial attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16898-16907, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "2571" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.175 + ], + "angle": 0, + "content": "[28] Lina Mezghan, Sainbayar Sukhbaatar, Thibaut Lavril, Oleksandr Maksymets, Dhruv Batra, Piotr Bojanowski, and Karteek Alahari. Memory-augmented reinforcement learning for image-goal navigation. In 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 3316-3323. IEEE, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.178, + 0.472, + 0.247 + ], + "angle": 0, + "content": "[29] Piotr Mirowski, Razvan Pascanu, Fabio Viola, Hubert Soyer, Andrew J Ballard, Andrea Banino, Misha Denil, Ross Goroshin, Laurent Sifre, Koray Kavukcuoglu, et al. Learning to navigate in complex environments. arXiv preprint arXiv:1611.03673, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.249, + 0.472, + 0.319 + ], + "angle": 0, + "content": "[30] Volodymyr Mnih, Adria Puigdomenech Badia, Mehdi Mirza, Alex Graves, Timothy Lillicrap, Tim Harley, David Silver, and Koray Kavukcuoglu. Asynchronous methods for deep reinforcement learning. In International conference on machine learning, pages 1928-1937, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.32, + 0.472, + 0.389 + ], + "angle": 0, + "content": "[31] Arsalan Mousavian, Alexander Toshev, Marek Fiser, Jana Košecka, Ayzaan Wahid, and James Davidson. Visual representations for semantic target driven navigation. In 2019 International Conference on Robotics and Automation (ICRA), pages 8846-8852. IEEE, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.39, + 0.472, + 0.459 + ], + "angle": 0, + "content": "[32] Giuseppe Oriolo, Marilena Vendittelli, and Giovanni Ulivi. On-line map building and navigation for autonomous mobile robots. In Proceedings of 1995 IEEE International Conference on Robotics and Automation, volume 3, pages 2900-2906. IEEE, 1995. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.461, + 0.472, + 0.516 + ], + "angle": 0, + "content": "[33] Emilio Parisotto and Ruslan Salakhutdinov. Neural map: Structured memory for deep reinforcement learning. In International Conference on Learning Representations, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.518, + 0.472, + 0.574 + ], + "angle": 0, + "content": "[34] Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. Automatic differentiation in pytorch. 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.576, + 0.472, + 0.644 + ], + "angle": 0, + "content": "[35] Jeffrey Pennington, Richard Socher, and Christopher D Manning. Glove: Global vectors for word representation. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 1532-1543, 2014. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.647, + 0.472, + 0.73 + ], + "angle": 0, + "content": "[36] Santhosh Kumar Ramakrishnan, Devendra Singh Chaplot, Ziad Al-Halah, Jitendra Malik, and Kristen Grauman. Poni: Potential functions for objectgoal navigation with interaction-free learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18890-18900, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.732, + 0.472, + 0.801 + ], + "angle": 0, + "content": "[37] Pravesh Ranchod, Benjamin Rosman, and George Konidaris. Nonparametric bayesian reward segmentation for skill discovery using inverse reinforcement learning. In 2015 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 471-477. IEEE, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.472, + 0.858 + ], + "angle": 0, + "content": "[38] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.472, + 0.902 + ], + "angle": 0, + "content": "[39] Nikolay Savinov, Alexey Dosovitskiy, and Vladlen Koltun. Semi-parametric topological memory for navigation. arXiv preprint arXiv:1803.00653, 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.161 + ], + "angle": 0, + "content": "[40] Gabriel Sepulveda, Juan Carlos Niebles, and Alvaro Soto. A deep learning based behavioral approach to indoor autonomous navigation. In 2018 IEEE International Conference on Robotics and Automation (ICRA), pages 4646-4653. IEEE, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.893, + 0.245 + ], + "angle": 0, + "content": "[41] Pierre Sermanet, Corey Lynch, Yevgen Chebotar, Jasmine Hsu, Eric Jang, Stefan Schaal, Sergey Levine, and Google Brain. Time-contrastive networks: Self-supervised learning from video. In 2018 IEEE international conference on robotics and automation (ICRA), pages 1134–1141. IEEE, 2018. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.247, + 0.893, + 0.302 + ], + "angle": 0, + "content": "[42] Adrian Šošić, Abdelhak M Zoubir, and Heinz Koeppl. A bayesian approach to policy recognition and state representation learning. IEEE transactions on pattern analysis and machine intelligence, 40(6):1295-1308, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.302, + 0.893, + 0.357 + ], + "angle": 0, + "content": "[43] Adrian Šošic, Abdelhak M Zoubir, Elmar Rueckert, Jan Peters, and Heinz Koeppl. Inverse reinforcement learning via nonparametric spatio-temporal subgoal modeling. Journal of Machine Learning Research, 19(69):1-45, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.358, + 0.893, + 0.399 + ], + "angle": 0, + "content": "[44] Tianqi Tang, Heming Du, Xin Yu, and Yi Yang. Monocular camera-based point-goal navigation by learning depth channel and cross-modality pyramid fusion. 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.4, + 0.893, + 0.469 + ], + "angle": 0, + "content": "[45] Yurun Tian, Xin Yu, Bin Fan, Fuchao Wu, Huub Heijnen, and Vassileios Balntas. Sosnet: Second order similarity regularization for local descriptor learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11016-11025, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.47, + 0.893, + 0.525 + ], + "angle": 0, + "content": "[46] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.526, + 0.893, + 0.621 + ], + "angle": 0, + "content": "[47] Su Wang, Ceslee Montgomery, Jordi Orbay, Vighnesh Birodkar, Aleksandra Faust, Izzeddin Gur, Natasha Jaques, Austin Waters, Jason Baldridge, and Peter Anderson. Less is more: Generating grounded navigation instructions from landmarks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15428-15438, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.623, + 0.893, + 0.678 + ], + "angle": 0, + "content": "[48] Erik Wijmans, Abhishek Kadian, Ari Morcos, Stefan Lee, Irfan Essa, Devi Parikh, Manolis Savva, and Dhruv Batra. Dd-ppo: Learning near-perfect pointgoal navigators from 2.5 billion frames. arXiv preprint arXiv:1911.00357, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.679, + 0.893, + 0.761 + ], + "angle": 0, + "content": "[49] Mitchell Wortsman, Kiana Ehsani, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Learning to learn how to learn: Self-adaptive visual navigation using meta-learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6750-6759, 2019. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.762, + 0.893, + 0.803 + ], + "angle": 0, + "content": "[50] Wei Yang, Xiaolong Wang, Ali Farhadi, Abhinav Gupta, and Roozbeh Mottaghi. Visual semantic navigation using scene priors. arXiv preprint arXiv:1810.06543, 2018. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.804, + 0.893, + 0.873 + ], + "angle": 0, + "content": "[51] Xin Yu and Fatih Porikli. Ultra-resolving face images by discriminative generative networks. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part V, pages 318-333. Springer, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.874, + 0.893, + 0.902 + ], + "angle": 0, + "content": "[52] Amir R Zamir, Alexander Sax, William Shen, Leonidas J Guibas, Jitendra Malik, and Silvio Savarese. Taskonomy:" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2572" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.469, + 0.134 + ], + "angle": 0, + "content": "Disentangling task transfer learning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3712-3722, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.469, + 0.205 + ], + "angle": 0, + "content": "[53] Sixian Zhang, Xinhang Song, Yubing Bai, Weijie Li, Yakui Chu, and Shuqiang Jiang. Hierarchical object-to-zone graph for object navigation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15130-15140, 2021. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.469, + 0.275 + ], + "angle": 0, + "content": "[54] Yuke Zhu, Roozbeh Mottaghi, Eric Kolve, Joseph J Lim, Abhinav Gupta, Li Fei-Fei, and Ali Farhadi. Target-driven visual navigation in indoor scenes using deep reinforcement learning. In 2017 IEEE international conference on robotics and automation (ICRA), pages 3357-3364. IEEE, 2017. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.956 + ], + "angle": 0, + "content": "2573" + } + ] +] \ No newline at end of file diff --git a/2023/Object-Goal Visual Navigation via Effective Exploration of Relations Among Historical Navigation States/02126feb-2cfd-431e-b8b7-664314fe1580_origin.pdf b/2023/Object-Goal Visual Navigation via Effective Exploration of Relations Among Historical Navigation States/02126feb-2cfd-431e-b8b7-664314fe1580_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..1daab51ffddcb73f489b9b965c1743e3c916f1f0 --- /dev/null +++ b/2023/Object-Goal Visual Navigation via Effective Exploration of Relations Among Historical Navigation States/02126feb-2cfd-431e-b8b7-664314fe1580_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bad98b20edf52f675c103a3df67955ed0d8464f8d9dd679b5583b9b394978c9d +size 3257375 diff --git a/2023/Object-Goal Visual Navigation via Effective Exploration of Relations Among Historical Navigation States/full.md b/2023/Object-Goal Visual Navigation via Effective Exploration of Relations Among Historical Navigation States/full.md new file mode 100644 index 0000000000000000000000000000000000000000..6f2969db8e2beae8751657cb21982ba11cb9befe --- /dev/null +++ b/2023/Object-Goal Visual Navigation via Effective Exploration of Relations Among Historical Navigation States/full.md @@ -0,0 +1,302 @@ +# Object-Goal Visual Navigation via Effective Exploration of Relations among Historical Navigation States + +Heming Du $^{1,3}$ , Lincheng Li $^{2}$ , Zi Huang $^{3}$ , Xin Yu $^{3*}$ + +$^{1}$ Australian National University $^{2}$ Netease Fuxi AI Lab $^{3}$ The University of Queensland + +# Abstract + +Object-goal visual navigation aims at steering an agent toward an object via a series of moving steps. Previous works mainly focus on learning informative visual representations for navigation, but overlook the impacts of navigation states on the effectiveness and efficiency of navigation. We observe that high relevance among navigation states will cause navigation inefficiency or failure for existing methods. In this paper, we present a History-inspired Navigation Policy Learning (HiNL) framework to estimate navigation states effectively by exploring relationships among historical navigation states. In HiNL, we propose a History-aware State Estimation (HaSE) module to alleviate the impacts of dominant historical states on the current state estimation. Meanwhile, HaSE also encourages an agent to be alert to the current observation changes, thus enabling the agent to make valid actions. Furthermore, we design a History-based State Regularization (HbSR) to explicitly suppress the correlation among navigation states in training. As a result, our agent can update states more effectively while reducing the correlations among navigation states. Experiments on the artificial platform AI2-THOR (i.e., iTHOR and RoboTHOR) demonstrate that HiNL significantly outperforms state-of-the-art methods on both Success Rate and SPL in unseen testing environments. + +# 1. Introduction + +Object-goal visual navigation is to direct an agent to move consecutively toward an object of a specific category. Without knowing the environment map beforehand, at each navigation step, an agent first needs to represent its visual observations, then estimate its navigation states from the visual representations and the preceding states, and at last predict the corresponding action. Therefore, to achieve an effective and efficient navigation system, learning instructive visual representations and navigation states is critical. + +Prevailing visual navigation works [13, 14, 50] focus on + +![](images/286b18ce66bcf8d30701be3925c4d94da35fa6563ff7b8baa454b1b2542f5b00.jpg) +(a) Demonstration of inefficient action predictions caused by highlycorrelated navigation states. Our agent is stuck by an obstacle, i.e., lowprofile sofa, and repeatedly predicts an invalid action, i.e., MoveAhead. + +Figure 1. Motivation of our proposed History-inspired Navigation Learning (HiNL) framework. +![](images/92d327d72606d350f666c12ad46524fbb9736254950a97c10325cdd95ba2947c.jpg) +(b) Demonstration of the correlation coefficients among navigation states trained in two manners. Navigation states estimated via LSTM are highly-relevant. In contrast, our HiNL produce low-correlated navigation states. + +extracting informative visual representations, while some methods [13, 49] adjust navigation policy during inference. All these approaches commonly employ recurrent neural networks (e.g., LSTM) to estimate navigation states. However, we observe that the navigation states of existing methods [13, 14, 49] exhibit high relevance, as demonstrated in Figure 1b, and the highly-correlated navigation states would lead to inefficient navigation policy (i.e., failure to respond to observation changes rapidly). For instance, as shown in Figure 1a, an agent is stuck by the low-profile sofa and fails to take proper actions to circumvent the obstacle. Hence, we aim to endow an agent with the capability of updating its navigation states effectively while avoiding producing highly-correlated states. + +In this work, we propose a History-inspired Navigation + +Learning (HiNL) framework to obtain informative navigation states by exploiting the relationships among historical navigation states. HiNL consists of two novel components: (i) a History-aware State Estimation (HaSE) module, and (ii) a History-based State Regularization (HbSR). Here, our HaSE module is designed to generate a state that can be promptly updated according to visual observations. Specifically, HaSE first analyzes the correlations among historical navigation states and then eliminates the influence of dominant historical states on the current state estimation. As a result, an agent is able to predict navigation states which can dynamically react to the current visual observations and then make sensible navigation actions. + +Furthermore, existing reinforcement learning-based object-goal navigation systems [13, 14, 49] often assume the navigation state transition exhibits the first-order Markov property. This would allow the emergence of high correlations among navigation states, leading to inferior navigation policy. To address this issue, we introduce an explicit constraint on the correlation among all the states, namely History-based State Regularization (HbSR). To be specific, HbSR enforces to relevance (i.e., correlations) between a state and all its preceding states (except its previous state) to be low. Here, we do not constrain states of two consecutive steps because temporally close states generally have relevance in practice considering the navigation continuity. After training with our HbSR, the correlations among the navigation states become much lower (see Figure 1b). This pheromone also indicates HiNL effectively updates states. Hence, our navigation system can respond to observation changes adaptively. + +To demonstrate the superiority of HiNL, we conduct experiments in the widely-adopted artificial environment iTHOR [26] and RoboTHOR [11]. HiNL outperforms the state-of-the-art by a large margin. To be specific, we improve the Success Rate (SR) from $72.2\%$ to $80.1\%$ and Success weighted by Path Length (SPL) from 0.449 to 0.498 in iTHOR. Overall, our major contributions are summarized as follows: + +- We propose a History-inspired Navigation Policy (HiNL) framework to effectively estimate navigation states by utilizing historical states. +- We design a History-aware State Estimation (HaSE) to eliminate dominant historical states in the current state estimation. Therefore, the agent reduces the impact of distant navigation states on the state estimation, and thus reacts dynamically to the observation changes. +- We introduce a History-based State Regularization (HbSR) to explicitly constrain the correlations among navigation states. By doing this, the agent can effectively update navigation states with low relevance. + +# 2. Related Works + +Visual navigation. Traditional works [3, 4, 32] often leverage an entire environment map for navigation and divide the task into three parts: mapping, localization, and path planning. However, environment maps are generally unavailable in unseen environments. Dissanayake et al. [12] adopt simultaneous localization and mapping (SLAM) to infer robot positions. Campari et al. [5] learn agent states via a Taskonomy model bank [52], but they need an RGB-D sensor to construct an online map during navigation. + +Recently, due to significant advancements in deep learning [16,22,45,51], reinforcement learning-based navigation methods [28, 29, 31, 33, 54] take visual observations as inputs and predict navigation actions. Vision-Language Navigation (VLN) approaches [9, 10, 15, 39, 40] steer an agent to the target based on its visual observations and navigation guidance in natural language. Similar to VLN, point-goal visual navigation methods [44,48] aim at driving an agent to a given point with step-wise directional indications. Moreover, audio-visual navigation methods [7, 8, 18] utilize additional audio signals to move a robot to the target position. Al-Halah et al. [1] propose a transfer learning model for multiple navigation tasks by embedding various navigation goals, e.g., image, sketch, and audio. + +Our work falls in the field of object-goal visual navigation [27, 31, 36, 47, 50]. However, existing object-goal navigation methods mainly focus on representing visual features comprehensively while we investigate the impact of navigation states on navigation performance. Wortsman et al. [49] exploit word embedding (i.e., GloVe embedding [35]) to represent the target category and introduce a meta network that mimics a reward function during inference. Du et al. [13] introduce an object relation graph, dubbed ORG, to encode visual observations. They also design a tentative policy for deadlock avoidance and adjust the navigation policy in unseen testing environments. The Hierarchical Object-to-Zone (HOZ) graph [53] offers coarse-to-fine guidance based on real-time updates. Additionally, VT-Net [14] incorporates object and region features with location cues, and EmbCLIP [24] leverages the contrastive language image pretraining encoder for visual navigation tasks. + +Correlation Modeling in Reinforcement Learning. Several methods [37, 43] explore correlations in hidden Markov models for inverse reinforcement learning (IRL). For action prediction, Hester et al. [20] propose Texlore to model correlations within the transition dynamics via a random forest. Vsovsic et al. [42] introduce a Bayesian approach to learn policy from demonstrations of experts by capturing correlations among actions. Alt et al. [2] design a Bayesian learning framework to establish temporal and spatial correlations among actions. Furthermore, + +![](images/8f68e97d22c1b7e7a9f7bd6f7d1ab1c50e06d13186a68d125343093ad5edd1ea.jpg) +Figure 2. Our History-inspired Navigation Policy Learning (HiNL) framework. HiNL takes visual representations as input and outputs navigation actions. HiNL involves two innovative parts: a History-aware State Estimation (HaSE) module and a History-based State Regularization (HbSR). HaSE is proposed to estimate navigation states that can reflect current observation changes from the perspective of network design, while HbSR is designed to enforce the informativeness of states from the view of the training objective. Both of them help to achieve effective and efficient navigation policy. + +Sermanet [41] propose a self-supervised TCN for learning robotic behaviors and representations from unlabeled multi-viewpoint videos. TCN uses a metric learning loss to create viewpoint-invariant representations that capture relationships between end-effectors. + +Different from previous methods that establish correlations among actions, our HiNL aims to eliminate high relevance among navigation states. To the best of our knowledge, our work HiNL is the first attempt to explicitly leverage the correlations among navigation states for object-goal navigation and achieves superior navigation performance. + +# 3. Proposed Methods + +In this work, we propose a History-inspired Navigation Policy (HiNL) framework to estimate navigation states effectively. As demonstrated in Figure 2, in each step, the agent first adopts a visual feature extractor to process the current RGB observation. Then, the agent estimates the current navigation state via a History-aware State Estimation (HaSE) module and predicts a navigation action. At last, we introduce a History-based State Regularization (HbSR) to constrain the correlations among navigation states during training, thus achieving low-correlated navigation states and highly effective navigation policy. + +# 3.1. Task Definition and Setup + +Following previous works [13, 14, 49, 53], in the objectgoal visual navigation task, an RGB monocular camera is the only sensor available to an agent. Prior knowledge of the entire environment, such as the topological map and 3D meshes, is unknown to the agent. An environment is composed of discrete grids, and the agent selects one of six actions, i.e., MoveAhead, RotateLeft, RotateRight, LookUp, LookDown, Done. Specif + +ically, the grid size is 0.25 meters, and the rotation angles of RotateLeft/RotateRight and LookUp/LookDown are $45^{\circ}$ and $30^{\circ}$ , respectively. + +At the start of each episode, the environment randomly chooses one of the available object categories $T \in \{Sink, \dots, Microwave\}$ as the target class. Then, the agent is teleported to a random start position $s = \{x, y, \theta_h, \theta_v\}$ , where $x$ and $y$ stand for the coordinates, and $\theta_h$ and $\theta_v$ represent the horizontal angle and the vertical angle for the view of the agent. The agent aims to predict an action distribution based on a learned navigation policy and operates the action with the highest probability for navigation. Then, the agent moves to the next navigation step if it selects a moving action or terminates the episode if it selects the action Done. After finishing the episode, we regard the agent as successful when the following three conditions are met simultaneously: (i) the agent selects the ending action Done within the maximum step length of an episode; (ii) there is at least one instance of the specified category in the view of the agent; (iii) the distance between the visible target and the agent is less than the threshold, i.e., 1.5 meters. Otherwise, the episode is considered as a failure case. + +# 3.2. Visual Representation Extraction + +Similar to previous works [13, 49], we first adopt a deep neural network, i.e., ResNet [19], to extract global features from an observed RGB image. Meanwhile, a detection module, e.g., Fast-RCNN [38] and DETR [6], is introduced to extract object features. Then, the global and object features are merged into a visual representation $\boldsymbol{v}_t$ by graph neural network [13, 53] or transformer $[14]^1$ . For convenience, we adopt $\boldsymbol{v}_t$ to denote the visual feature extracted + +![](images/08a353a65aee560784829d4132136abe13f714d9d14284a6fa76579c3665ebbf.jpg) +Figure 3. Illustration of History-aware State Estimation (HaSE). The agent first adopts a Historical State Guidance Extraction module to estimate a state guidance that indicates the dominant historical states. Then, the agent extracts a state innovation that reflects incremental information of current observations, via a State-induced Perceptron module. At last, the agent generates the current navigation state estimation by fusing the state innovation and guidance in a History-aware Fusion module. + +from the RGB observation at time step $t$ . + +# 3.3. History-aware State Estimation + +Our History-aware State Estimation (HaSE) estimates the navigation states by eliminating the influence of dominant historical states and focusing on the present observation changes. As demonstrated in Figure 3, after obtaining the visual representation $\boldsymbol{v}_t$ , we design a Historical State Guidance Extraction module to predict a state guidance $\boldsymbol{g}_t$ indicating dominant historical states at time step $t$ . Then, we introduce a State-induced Perceptron module to obtain the current state innovation $\boldsymbol{i}_t$ , which encodes the new information brought by the current visual observations and the last step action. Next, the history-aware guidance $\boldsymbol{g}_t$ is employed to eliminate the dominant historical states and facilitate the estimation of the current navigation state $\boldsymbol{s}_t$ via our proposed History-aware Fusion module. + +Historical State Guidance Extraction. In order to eliminate the dominant historical states in the current state estimation, we extract guidance for the dominant parts by exploring the relationships among historical navigation states. The agent roughly approximates a prior state $s_t'$ based on the last state $s_{t-1}$ , action $a_{t-1}$ , and the current visual representation $v_t$ . The prior state $s_t'$ is encoded with the last navigation state and the current observation, and thus we refer $s_t'$ to the query about the dominant historical states. Let FFN be a feed-forward network, $s_t'$ is formulated as follows, + +$$ +\boldsymbol {s} _ {t} ^ {\prime} = \sigma (\operatorname {F F N} \left(\boldsymbol {s} _ {t - 1}\right) + \operatorname {F F N} \left(\left[ \boldsymbol {v} _ {t}; \boldsymbol {a} _ {t - 1} \right]\right)), \tag {1} +$$ + +where $[;]$ and $\sigma$ represent the concatenation operation and sigmoid function, respectively. + +After obtaining $s_t'$ , the agent memorizes a set of the historical navigation states $\mathbb{S}_{t - \tau : t - 1} = \{s_{t - \tau}, \dots, s_{t - 1}\}$ in the last $\tau$ steps. In order to explore the temporal relationships among historical navigation states, $\mathbb{S}_{t - \tau : t - 1}$ are composed additively with temporal embeddings. Concretely, we adopt absolute temporal embedding [46] based on the timestamp at which a state is estimated. By doing this, the agent perceives not only the temporal orders among the last $\tau$ navigation states, but also the length of the current + +episode. Therefore, the agent can exploit temporal information to eliminate the distant navigation states in extracting the guidance. In addition, we analyze the impacts of different temporal embedding methods (i.e., no/learnable/relative temporal embedding) in Table 2. + +Afterward, we pass $s_t'$ and $\mathbb{S}_{t - \tau : t - 1}$ through the historical state decoder, i.e., a cross-attention module. We refer to the output $z_t$ as the dominant historical states, and calculate it as follows, + +$$ +\boldsymbol {z} _ {\boldsymbol {t}} = \operatorname {D e c o d e r} _ {\mathrm {h i s}} \left(\boldsymbol {s} _ {\boldsymbol {t}} ^ {\prime}, \mathbb {S} _ {t - \tau : t - 1}\right), \tag {2} +$$ + +where $\mathrm{Decoder}_{\mathrm{his}}(\cdot, \cdot)$ indicates the historical state decoder, the first argument for query and the other for key/values. Then, we formulate the guidance $g_{t}$ as the difference between $s_{t-1}$ and $z_{t}$ . Therefore, we compute $g_{t}$ as follows, + +$$ +\boldsymbol {g} _ {t} = \sigma \left(\boldsymbol {s} _ {t - 1} - \boldsymbol {z} _ {t}\right). \tag {3} +$$ + +In addition, we set $\tau$ as 5 in our experiments and demonstrate the impacts of different $\tau$ in Figure 5b. + +State-induced Perceptron. Apart from eliminating the dominant historical states, we also aim to encourage the agent to emphasize the observed changes in the current state estimation. Therefore, we extract a state innovation $i_{t}$ based on $v_{t}$ to encode the observed changes at the current step. To be specific, we projects $v_{t}, s_{t-1}$ and $a_{t-1}$ to $i_{t}$ . Inspired by LSTM [21], we compute $i_{t}$ as, + +$$ +\boldsymbol {i} _ {t} = \tanh (\operatorname {F F N} \left(\boldsymbol {s} _ {t - 1}\right) + \operatorname {F F N} ([ \boldsymbol {v} _ {t}; \boldsymbol {a} _ {t - 1} ])) \odot \tag {4} +$$ + +$$ +\sigma (\operatorname {F F N} \left(\boldsymbol {s} _ {\boldsymbol {t} - \mathbf {1}}\right) + \operatorname {F F N} ([ \boldsymbol {v} _ {\boldsymbol {t}}; \boldsymbol {a} _ {\boldsymbol {t} - \mathbf {1}} ])), +$$ + +where tanh represents hyperbolic tangent function, and $\odot$ stands for the element-wise product. + +History-aware Fusion. After obtaining the guidance $g_{t}$ and state innovation $i_{t}$ , we aim to estimate the current navigation state $s_{t}$ . In order to eliminate the influence of dominant navigation states, the agent first merges $s_{t-1}$ and $g_{t}$ by multiplication. Then the agent adds $i_{t}$ to the result of multiplication to incorporate the observation changes. Formally, we formulate the current navigation state $s_{t}$ estimation as follows, + +$$ +\boldsymbol {s} _ {t} = \tanh \left(\boldsymbol {s} _ {t - 1} \odot \boldsymbol {g} _ {t} + \boldsymbol {i} _ {t}\right). \tag {5} +$$ + +![](images/2b9bc7c5947ecaf5265841f8714f8173c6bdf25d7fd47c5518cc3d60154f60cf.jpg) +Figure 4. Illustration of History-based State Regularization (HbSR). (a) First-order Markov state transition that is widely adopted by LSTM based navigation policy learning. Those methods will produce highly-correlated among states. (b) Our HbSR constrains higher-order relations among states, and leads to low-correlated states. As the states are more informative (i.e., less correlated), our navigation policy is more effective and efficiency. + +Benefiting from HaSE, the agent can effectively update the navigation states by being alert to the present observation and neglecting the influence of distant states. + +# 3.4. History-based State Regularization + +Previous reinforcement learning-based object-goal visual navigation methods are designed based on the first-order Markov assumption, while there are no explicit constraints on the correlations among navigation states. We observe that high relevance among navigation states exhibit in the experiments of previous methods [13, 14] and this is also one of the reasons that prior works yield failure cases. For instance, an agent is stuck in front of an object rather than adopting a rotation action or trapped in a loop. From these failure cases, we can see that the navigation states are highly correlated as the adopted actions are periodically the same. In order to avoid inefficient navigation caused by highly-relevant navigation states, we aim to exert an explicit constraint on the correlations among navigation states. + +We introduce a novel History-based State Regularization (HbSR) to reduce the correlations among navigation states from the perspective of the training objective. In order to measure the correlations among navigation states, we adopt the Pearson correlation coefficient [17]. To be specific, after the agent terminates an episode with $T$ steps, we obtain a set of navigation states $\mathbb{S}_{0:T} = \{s_0,\dots,s_T\}$ . Given two navigation states $s_i$ and $s_j$ , we first compute the covariance of the two states as follows, + +$$ +\operatorname {C o v} \left(\boldsymbol {s} _ {i}, \boldsymbol {s} _ {j}\right) = \mathbb {E} \left(\left(\boldsymbol {s} _ {i} - \mathbb {E} \left(\boldsymbol {s} _ {i}\right)\right) \left(\boldsymbol {s} _ {j} - \mathbb {E} \left(\boldsymbol {s} _ {j}\right)\right)\right). \tag {6} +$$ + +Then, we compute the correlation coefficient $\rho$ for two states based on the covariance formulated as, + +$$ +\rho \left(\boldsymbol {s} _ {i}, \boldsymbol {s} _ {j}\right) = \frac {\operatorname {C o v} \left(\boldsymbol {s} _ {i} , \boldsymbol {s} _ {j}\right)}{\sqrt {\operatorname {V a r} \left(\boldsymbol {s} _ {i}\right) \operatorname {V a r} \left(\boldsymbol {s} _ {j}\right)}}, \tag {7} +$$ + +where $\operatorname{Var}(s_i)$ denotes the variance of $s_i$ . + +The correlation coefficient is expressed in a range $[-1, + 1]$ . Two states are highly relevant when their correlation is close to either 1 or $-1$ , whereas $\rho = 0$ indicates irrelevant between two states. Therefore, in order to eliminate high relevance among states, we enforce the correla + +tions of states to be close to 0. In particular, since inefficient navigation is generally caused by positive correlations of navigation states, we only apply the proposed regularization on correlations with positive values. According to the first-order Markov property, the present state $s_t$ depends on its immediately preceding state $s_{t-1}$ . Thus, we define a regularization-free threshold $\xi$ . Moreover, in HbSR, if a temporal interval between two states is less than the threshold $\xi$ , we do not apply the correlation regularization among consecutive states. This is because navigation has continuity and neighboring states will inherently have relevance while distant states should be irrelevant. Hence, our HbSR $l^{hsr}$ is expressed as, + +$$ +l ^ {h s r} = \mathbb {E} _ {\boldsymbol {s} \in \mathbb {S}} \left(\sum_ {t = \xi + 1} ^ {T} \sum_ {i = 0} ^ {t - \xi - 1} \max \left(0, \rho \left(\boldsymbol {s} _ {t}, \boldsymbol {s} _ {i}\right)\right)\right). \tag {8} +$$ + +# 3.5. Training Details + +We employ the Asynchronous Advantage Actor-Critic (A3C) Algorithm [30] to train the agent, similar to previous works [13, 14, 49, 53]. We adopt two loss objectives, i.e., policy loss $l_{\text{policy}}$ and value loss $l_{\text{value}}$ , for training the navigation policy and value estimation networks, respectively. Furthermore, since the calculation of correlation coefficients is differentiable [34], we can employ the loss $l^{hsr}$ of HbSR in training. Formally, the total loss $L$ is formulated as the summation of $l_{\text{policy}}$ , $l_{\text{value}}$ and $l^{hsr}$ , + +$$ +l = \lambda_ {0} l ^ {\text {p o l i c y}} + \lambda_ {1} l ^ {\text {v a l u e}} + \lambda_ {2} l ^ {\text {h s r}}, \tag {9} +$$ + +where $\lambda_0, \lambda_1$ and $\lambda_2$ are weighted factors to balance three losses. In our experiments, we set these three weighted factors as 1, 0.5 and 1, respectively. + +# 4. Experiments + +To validate the effectiveness and efficiency of the proposed HiNL, we conduct extensive experiments on two artificial environments, i.e., iTHOR [26] and RoboTHOR [11]. + +# 4.1. Protocols and Experimental details + +Dataset. We adopt iTHOR [26], an artificial environment within the AI2-THOR framework, to evaluate our method. + +
MethodiTHOR [26]RoboTHOR [11]
AllL>5AllL>5
SuccessSPLSuccessSPLSuccessSPLSuccessSPL
Random8.0±1.33.6±0.60.3±0.10.1±0.14.0±1.01.6±0.40.2±0.10.1±0.1
WE33.0±3.514.7±1.821.4±3.011.7±1.97.0±1.22.2±0.64.7±0.91.7±0.3
SP [50]35.1±1.315.5±1.122.2±2.711.4±1.610.9±0.54.2±0.57.2±0.63.2±0.5
SAVN [49]40.8±1.216.1±0.528.7±1.513.9±0.510.2±1.03.9±0.96.9±0.82.8±0.6
ORG [13]65.3±0.737.5±0.154.8±1.036.1±0.145.4±0.821.2±1.239.6±0.718.5±0.8
ORG+TPN [13]69.3±1.239.4±1.060.7±1.338.6±1.147.8±1.223.1±0.942.9±0.920.3±1.0
HOZ [53]70.6±1.740.0±1.262.8±1.739.2±0.644.3±1.819.8±1.337.2±1.215.4±0.8
TCN [41]75.1±0.847.7±0.965.3±1.442.1±0.844.4±1.626.4±1.538.4±1.122.0±1.2
VTNet [14]72.2±1.044.9±0.763.4±1.144.0±0.953.2±1.127.5±1.747.0±0.823.3±0.7
HiNL †71.2±0.838.9±0.964.1±0.637.0±0.556.1±1.326.4±0.851.8±0.923.9±0.7
HiNL ‡80.1±1.449.8±1.974.6±1.747.6±1.460.6±1.030.8±1.256.2±0.826.6±0.9
+ +Table 1. Comparison with existing methods. We report the average Success rate (%) and SPL (%) in iTHOR [26] and RoboTHOR [11] as well as their variances in subscripts by repeating experiments five times. $L > 5$ represents the episodes that require at least 5 steps. † and ‡ indicate using CNN-based and Transformer-based visual extractors (i.e., ORG [13] and VTNet [14]), respectively. + +The environment includes four types of scenes, i.e., kitchen, living room, bedroom, and bathroom. Each type of scene consists of 30 rooms with different furniture items and placements. Following [13, 14, 53], we use 22 categories as the target classes and ensure that there are at least four targets in each room. We choose the first 20 out of 30 rooms per scene as the training set and equally divide the remaining 40 rooms into the validation and test set. + +Furthermore, we conduct experiments in RoboTHOR [11], the other synthetic environment in the AI2-THOR framework. Different from iTHOR, each scene in RoboTHOR is separated by several clapboards. RoboTHOR consists of 89 scenes, while only 75 apartments are published for train/val (60 for training and 15 for validation). Therefore, we select the first 55 of 60 original training scenes as our training dataset and use the rest 5 scenes as the validation set. Then, we adopt the remaining 15 original validation apartments as our test set. We report the test result of the model with the highest success rate on the validation set for both iTHOR and RoboTHOR. + +Evaluation metrics. We adopt two evaluation metrics, i.e., success rate and Success weighted by Path Length (SPL), to assess our method performance. Success rate is introduced to evaluate the navigation effectiveness and calculated by $\frac{1}{N}\sum_{n = 0}^{N}S_{n}$ , where $N$ is the number of episodes and $S_{n}$ represents a binary success indicator of the $n$ -th episode. Meanwhile, SPL measures the efficiency of navigation trajectories. Let $L_{n}$ be the length of the $n$ -th episode and $L_{opt}$ indicate the length of the optimal path. SPL is formulated as $\frac{1}{N}\sum_{n = 0}^{N}S_{n}\frac{L_{n}}{\max(L_{n},L_{opt})}$ . + +Implementation details. Following [14], we train the VT for 20 epochs with the supervision of expert experience. Then, we train the navigation policy for 2M episodes in to- + +tal with 32 asynchronous agents. In making a robot learn to approach the target, the penalty is set to $1e^{-3}$ for every step that passes prior to arrival, and thus an agent is encouraged to navigate as quickly as possible. Meanwhile, to encourage the agent complete navigation effectively, we offer a large reward of 5 when the agent finishes a trajectory successfully. We adopt Adam [25] with a learning rate $10^{-4}$ . Our codes will be publicly related for reproducibility. + +# 4.2. Comparison with Existing Methods + +# 4.2.1 Competing Methods + +We compare our method with the following ones: Random policy. An agent makes decisions based on a uniform action probability. Thus, the movements and stops of the agent are random. Scene Prior (SP) [50] utilizes the scene prior knowledge and category associations for navigation with a graph neural network on the FastText database [23]. Word Embedding (WE) makes use of GloVe embedding [35] to signify the target category instead of relying on detection. From trial and error, an agent establishes the connection between object appearances and GloVe embeddings. Self-adaptive Visual Navigation (SAVN) [49] presents a meta reinforcement learning method to assist an agent in adapting to unseen environments during inference. Object Relationship Graph (ORG) [13] proposes a visual representation learning method to encode relationships among categories via a graph and design a tentative policy network (TPN) to escape from deadlocks during testing. Hierarchical Object-to-Zone (HOZ) [53] introduces a graph to guide an agent in a coarse-to-fine manner and a online-learning mechanism to update the graph. Time-Contrastive Networks (TCN) [41] presents a self-supervised method, deriving robotic behaviors and representations from unlabeled + +
MethodAllL>5
SuccessSPLSuccessSPL
HaSEHiNL † w/o HaSE68.137.860.335.9
HiNL †71.238.964.137.0
HiNL ‡ w/o HaSE77.945.771.643.5
HiNL ‡80.149.874.647.6
HiNL ‡ w/ tem.78.539.871.239.1
HiNL ‡ w/ lea. tem.79.145.873.143.8
HiNL ‡ w/ rel. tem.77.344.169.441.6
HiNL ‡80.149.874.647.6
HbSRHiNL † w/o HbSR66.838.157.835.8
HiNL †71.238.964.137.0
HiNL ‡ w/o HbSR75.446.967.043.8
HiNL ‡80.149.874.647.6
HiNL ‡ neg. & pos.75.446.967.143.8
HiNL ‡80.149.874.647.6
+ +Table 2. Ablation study on different components of HiNL in iTHOR [26] environment. † and ‡ indicate using CNN-based and Transformer-based visual extractors (i.e., ORG [13] and VT-Net [14]), respectively. HiNL w/o tem. indicates HiNL without temporal embedding in HaSE, while HiNL w/ lea. tem. and HiNL w/ rel. tem. represent the adoption of learnable and relative temporal embedding in HiNL, respectively. HiNL neg. & pos. stands for regularizing both positive and negative correlations toward 0. + +videos captured across multiple viewpoints, leveraging metric learning loss to establish viewpoint-invariant representations that encapsulate end-effector relationships. Visual Transformer for Navigation (VTNet) [14] explores the spatial correlations among objects and observation regions for navigation via a transformer-based network. + +# 4.2.2 Quantitative Results + +To demonstrate the superiority of our proposed learning framework, we select two different visual representation extractors: one is based on graph network (i.e., ORG [13]), and the other is based on the transformer (i.e., VT [14]). Table 1 indicates that both HiNL $\dagger$ and HiNL $\ddagger$ achieve superior results compared with their original methods (i.e., ORG and VTNet). This demonstrates that given different visual representations, HiNL can improve navigation performance constantly by explicitly exploring the relationships among historical states. + +To be specific, HiNL $\ddagger$ with the transformer-based extractor surpasses the existing methods by a large margin on both success rate $(+7.9\% / + 7.4\%)$ and SPL $(+4.9\% / + 3.3\%)$ in iTHOR [26]/RoboTHOR [11]. Adopting ORG as the visual representation extractor, HiNL $\dagger$ achieves significant improvement over the original ORG by nearly $10\%$ in success rate in both environments. VTNet and ORG only adopt LSTM to merge well-designed visual representations into navigation states without any explicit constraints on states. The correlations among navigation states are not exploited. + +This experiment suggests that our HiNL leads to instructive navigation states, and thus significantly improves the effectiveness and efficiency of our navigation system. + +As indicated in Table 1, we observe that HiNL $\dagger$ significantly outperforms ORG+TPN [13]. Although TPN employs an auxiliary network to fine-tune the initial navigation policy network when the agent is stuck in deadlocks during inference, TPN does not guarantee navigation states to be low relevant. Consequently, an agent might be trapped in looping actions, such as moving with rotating, because of the high correlations among navigation states. HiNL solves this issue by reducing the correlations among navigation states, and thus achieves proper actions for navigation. Furthermore, adopting TCN-based loss (i.e., TCN) yields inferior performance with $5\%$ decrease in success rate compared to HiNL. This is because TCN tries to pull temporally neighboring features closer while consecutive navigation states might be very different due to a rotation action. Thus, TCN may overly emphasize the reliance among neighboring states, thus leading to inferior state estimation. In addition, our method achieves more than two times the success rate and SPL than SAVN in both environments. Benefiting from well-designed visual representations and lowering relevance among navigation states, HiNL achieves better navigation performance in testing even without adjusting the navigation policy. This also implies that our navigation states significantly improve the effectiveness and efficiency of navigation in unseen environments. + +# 4.2.3 Case Study + +Figure 5a illustrates trajectories of three navigation episodes proceeded by VTNet [14] and HiNL $\ddagger$ . In the first case, VTNet fails to find the target and is stuck in the environment, as seen in the first column. On the contrary, HiNL $\ddagger$ navigates toward the object successfully. This implies that using HiNL, our agents can react rapidly to the observed changes. In the second case, VTNet moves forward while rotating and thus leads to inefficient navigation. In contrast, the navigation system trained with HiNL avoids rotating and approaches the target with fewer steps. This demonstrates that our HiNL improves navigation efficiency by reducing the correlations among states. In the final case, VTNet moves forward consistently until it reaches the wall. Benefiting from our HaSE, the agent can get rid of dominant historical states and react rapidly to the observed changes. + +# 4.3. Ablation Study + +Impacts of different components. To illustrate the impacts of different components, we conduct experiments on adopting one of the proposed components (i.e., HaSE and HbSR). Furthermore, we report the navigation performance with different visual representation extractors, i.e., ORG and VT. As indicated in Table 2, both HiNL $\dagger$ and + +![](images/cdeace07492ab6745fbb0a8855c160ce4cad1f47817895e77376c8826c92988b.jpg) +Figure 5. (a) Visual results in the simulated environments iTHOR. The arrow color changes (i.e., from blue to red) represent the navigation progress (i.e., from beginning to end). Black arrows indicate rotations. The target objects are highlighted by the green/red boxes, where a green box stands for a success episode and a red box represents a failure case. (b) Impacts of different historical state length $\tau \in [2,10]$ . (c) Results of different regularization-free states $\xi \in [1,8]$ . We set $\tau = 5$ and $\xi = 1$ and highlight them by the black vertical dotted lines in (b) and (c). + +![](images/432a840037cb021724dc457044c1bad233f0286c843ae753da6e3c10bb90da0c.jpg) + +HiNL $\ddagger$ achieve gains in navigation performance on success rate $(+2.9\% / + 2.2\%)$ and SPL $(+0.9\% / + 4.1\%)$ by introducing HaSE. Furthermore, HiNL $\dagger$ and HiNL $\ddagger$ outperform models without HbSR (i.e., HiNL $\dagger/\mathrm{HiNL}\ddagger$ w/o HbSR), respectively. This manifests that both of our proposed components improve the performance of navigation. + +Impacts of different temporal embedding in HaSE. Table 2 compares different temporal embedding methods, (i.e., w/o temporal embedding, w/ relative temporal embedding and w/ absolute temporal embedding (HiNL)) and highlights the effectiveness of incorporating temporal embedding in historical states. HiNL with absolute temporal embedding significantly improves success rates and SPL on iTHOR compared to no embedding (HiNL $\ddagger$ w/o tem.), learnable embedding (HiNL $\ddagger$ w/ lea. tem.), and relative embedding (HiNL $\ddagger$ w/ rel. tem.). This improvement, especially in navigation efficiency, stems from providing the agent with the current episode length. + +Comparison of different $\tau$ in HaSE. To illustrate the influence of different historical state lengths, we conduct experiments with $\tau \in [2,10]$ . As demonstrated in Figure 5b, we observe that as historical states become too many, our agent may fail to converge to an optimal policy. On the other hand, estimating navigation states with few historical states do not have sufficient capacity to effectively update states. The agent reaches the peak of success rate and SPL when it adopts 5 historical states in HaSE. + +Impacts of different $\xi$ in HbSR. To study the impacts of adopting different regularization-free thresholds $\xi$ , we conduct experiments with $\xi \in [1,8]$ in iTHOR. As seen in Figure 5c, increasing $\xi$ drops performance clearly. Both success rate (80% to 76%) and SPL (49.8% to 45.8%) decrease when $\xi$ increases from 1 to 8. Since two navigation states generally have high correlations when they are close + +in temporal, raising $\xi$ indicates ignoring constraints on these correlations. As a result, the effects of HbSR are compromised, and the navigation performance is dropped. + +Comparison of regularization on different values. To analyze the impacts of regularization on different values, we demonstrate the comparisons of applying HbSR on (i) both negative and positive correlations and (ii) only positive correlations, as indicated in Table 2. Compared to HiNL trained on regularizing the correlations based on both positive and negative values (HiNL $\ddagger$ neg. & pos.), HiNL improves performance by neglecting the negative correlations. This comparison suggests that inefficient navigation is generally caused by positively correlated navigation states. + +# 5. Conclusion + +In this paper, we proposed an effective navigation state learning approach for visual navigation, named History-inspired Navigation Policy Learning (HiNL). Our HiNL involves two innovate parts: a history-aware state estimation module (HaSE) and a history-based state regularization (HbSR). HbSR eliminates the negative impacts of dominant historical states and enables current states to reflect newly observed visual information. HbSR significantly reduces the correlations among navigation states, thus leading to more informative states as well as effective actions. Extensive results demonstrate that our HiNL significantly improves navigation performance. + +# Acknowledgement + +This research is funded in part by ARC-Discovery grant (DP220100800 to XY), ARC-DECRA grant (DE230100477 to XY) and ARC Centres of Excellence grant (CE200100025). We thank all anonymous reviewers and ACs for their constructive suggestions. + +# References + +[1] Ziad Al-Halah, Santhosh Kumar Ramakrishnan, and Kristen Grauman. Zero experience required: Plug & play modular transfer learning for semantic visual navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17031-17041, 2022. 2 +[2] Bastian Alt, Adrian Šošić, and Heinz Koeppl. Correlation priors for reinforcement learning. Advances in Neural Information Processing Systems, 32, 2019. 2 +[3] Johann Borenstein and Yoram Koren. Real-time obstacle avoidance for fast mobile robots. IEEE Transactions on systems, Man, and Cybernetics, 19(5):1179-1187, 1989. 2 +[4] Johann Borenstein and Yoram Koren. The vector field histogram-fast obstacle avoidance for mobile robots. IEEE transactions on robotics and automation, 7(3):278-288, 1991. 2 +[5] Tommaso Campari, Leonardo Lamanna, Paolo Traverso, Luciano Serafini, and Lamberto Ballan. Online learning of reusable abstract models for object goal navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14870-14879, 2022. 2 +[6] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 3 +[7] Changan Chen, Ziad Al-Halah, and Kristen Grauman. Semantic audio-visual navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15516–15525, 2021. 2 +[8] Changan Chen, Unnat Jain, Carl Schissler, Sebastia Vicenc Amengual Gari, Ziad Al-Halah, Vamsi Krishna Ithapu, Philip Robinson, and Kristen Grauman. Soundspaces: Audio-visual navigation in 3d environments. In European Conference on Computer Vision, pages 17–36. Springer, 2020. 2 +[9] Jinyu Chen, Chen Gao, Erli Meng, Qiong Zhang, and Si Liu. Reinforced structured state-evolution for vision-language navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15450-15459, 2022. 2 +[10] Kevin Chen, Juan Pablo de Vicente, Gabriel Sepulveda, Fei Xia, Alvaro Soto, Marynel Vázquez, and Silvio Savarese. A behavioral approach to visual navigation with graph localization networks. arXiv preprint arXiv:1903.00445, 2019. 2 +[11] Matt Deitke, Winson Han, Alvaro Herrasti, Aniruddha Kembhavi, Eric Kolve, Roozbeh Mottaghi, Jordi Salvador, Dustin Schwenk, Eli VanderBilt, Matthew Wallingford, et al. Robothor: An open simulation-to-real embodied ai platform. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3164-3174, 2020. 2, 5, 6, 7 +[12] MWM Gamini Dissanayake, Paul Newman, Steve Clark, Hugh F Durrant-Whyte, and Michael Csorba. A solution to the simultaneous localization and map building (slam) problem. IEEE Transactions on robotics and automation, 17(3):229-241, 2001. 2 + +[13] Heming Du, Xin Yu, and Liang Zheng. Learning object relation graph and tentative policy for visual navigation. arXiv preprint arXiv:2007.11018, 2020. 1, 2, 3, 5, 6, 7 +[14] Heming Du, Xin Yu, and Liang Zheng. Vtnet: Visual transformer network for object goal navigation. arXiv preprint arXiv:2105.09447, 2021. 1, 2, 3, 5, 6, 7 +[15] Kuan Fang, Alexander Toshev, Li Fei-Fei, and Silvio Savarese. Scene memory transformer for embodied agents in long-horizon tasks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 538–547, 2019. 2 +[16] Xiaoyu Feng, Heming Du, Yueqi Duan, Yongpan Liu, and Hehe Fan. Sefformer: Structure embedding transformer for 3d object detection. arXiv preprint arXiv:2209.01745, 2022.2 +[17] David Freedman, Robert Pisani, and Roger Purves. Statistics (international student edition). Pisani, R. Purves, 4th edn. WW Norton & Company, New York, 2007. 5 +[18] Chuang Gan, Yiwei Zhang, Jiajun Wu, Boqing Gong, and Joshua B Tenenbaum. Look, listen, and act: Towards audio-visual embodied navigation. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 9701–9707. IEEE, 2020. 2 +[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3 +[20] Todd Hester and Peter Stone. Texlore: real-time sample-efficient reinforcement learning for robots. Machine learning, 90(3):385-429, 2013. 2 +[21] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation, 9(8):1735-1780, 1997. 4 +[22] Meihuizi Jia, Lei Shen, Xin Shen, Lejian Liao, Meng Chen, Xiaodong He, Zhendong Chen, and Jiaqi Li. Mnerqq: An end-to-end mrc framework for multimodal named entity recognition with query grounding. arXiv preprint arXiv:2211.14739, 2022. 2 +[23] Armand Joulin, Edouard Grave, Piotr Bojanowski, and Tomas Mikolov. Bag of tricks for efficient text classification. arXiv preprint arXiv:1607.01759, 2016. 6 +[24] Apoorv Khandelwal, Luca Weihs, Roozbeh Mottaghi, and Aniruddha Kembhavi. Simple but effective: Clip embeddings for embodied ai. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14829-14838, 2022. 2 +[25] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6 +[26] Eric Kolve, Roozbeh Mottaghi, Winson Han, Eli VanderBilt, Luca Weihs, Alvaro Herrasti, Daniel Gordon, Yuke Zhu, Abhinav Gupta, and Ali Farhadi. AI2-THOR: An Interactive 3D Environment for Visual AI. arXiv, 2017. 2, 5, 6, 7 +[27] Bar Mayo, Tamir Hazan, and Ayellet Tal. Visual navigation with spatial attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16898-16907, 2021. 2 + +[28] Lina Mezghan, Sainbayar Sukhbaatar, Thibaut Lavril, Oleksandr Maksymets, Dhruv Batra, Piotr Bojanowski, and Karteek Alahari. Memory-augmented reinforcement learning for image-goal navigation. In 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 3316-3323. IEEE, 2022. 2 +[29] Piotr Mirowski, Razvan Pascanu, Fabio Viola, Hubert Soyer, Andrew J Ballard, Andrea Banino, Misha Denil, Ross Goroshin, Laurent Sifre, Koray Kavukcuoglu, et al. Learning to navigate in complex environments. arXiv preprint arXiv:1611.03673, 2016. 2 +[30] Volodymyr Mnih, Adria Puigdomenech Badia, Mehdi Mirza, Alex Graves, Timothy Lillicrap, Tim Harley, David Silver, and Koray Kavukcuoglu. Asynchronous methods for deep reinforcement learning. In International conference on machine learning, pages 1928-1937, 2016. 5 +[31] Arsalan Mousavian, Alexander Toshev, Marek Fiser, Jana Košecka, Ayzaan Wahid, and James Davidson. Visual representations for semantic target driven navigation. In 2019 International Conference on Robotics and Automation (ICRA), pages 8846-8852. IEEE, 2019. 2 +[32] Giuseppe Oriolo, Marilena Vendittelli, and Giovanni Ulivi. On-line map building and navigation for autonomous mobile robots. In Proceedings of 1995 IEEE International Conference on Robotics and Automation, volume 3, pages 2900-2906. IEEE, 1995. 2 +[33] Emilio Parisotto and Ruslan Salakhutdinov. Neural map: Structured memory for deep reinforcement learning. In International Conference on Learning Representations, 2018. 2 +[34] Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. Automatic differentiation in pytorch. 2017. 5 +[35] Jeffrey Pennington, Richard Socher, and Christopher D Manning. Glove: Global vectors for word representation. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 1532-1543, 2014. 2, 6 +[36] Santhosh Kumar Ramakrishnan, Devendra Singh Chaplot, Ziad Al-Halah, Jitendra Malik, and Kristen Grauman. Poni: Potential functions for objectgoal navigation with interaction-free learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18890-18900, 2022. 2 +[37] Pravesh Ranchod, Benjamin Rosman, and George Konidaris. Nonparametric bayesian reward segmentation for skill discovery using inverse reinforcement learning. In 2015 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 471-477. IEEE, 2015. 2 +[38] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 3 +[39] Nikolay Savinov, Alexey Dosovitskiy, and Vladlen Koltun. Semi-parametric topological memory for navigation. arXiv preprint arXiv:1803.00653, 2018. 2 + +[40] Gabriel Sepulveda, Juan Carlos Niebles, and Alvaro Soto. A deep learning based behavioral approach to indoor autonomous navigation. In 2018 IEEE International Conference on Robotics and Automation (ICRA), pages 4646-4653. IEEE, 2018. 2 +[41] Pierre Sermanet, Corey Lynch, Yevgen Chebotar, Jasmine Hsu, Eric Jang, Stefan Schaal, Sergey Levine, and Google Brain. Time-contrastive networks: Self-supervised learning from video. In 2018 IEEE international conference on robotics and automation (ICRA), pages 1134–1141. IEEE, 2018. 3, 6 +[42] Adrian Šošić, Abdelhak M Zoubir, and Heinz Koeppl. A bayesian approach to policy recognition and state representation learning. IEEE transactions on pattern analysis and machine intelligence, 40(6):1295-1308, 2017. 2 +[43] Adrian Šošic, Abdelhak M Zoubir, Elmar Rueckert, Jan Peters, and Heinz Koeppl. Inverse reinforcement learning via nonparametric spatio-temporal subgoal modeling. Journal of Machine Learning Research, 19(69):1-45, 2018. 2 +[44] Tianqi Tang, Heming Du, Xin Yu, and Yi Yang. Monocular camera-based point-goal navigation by learning depth channel and cross-modality pyramid fusion. 2022. 2 +[45] Yurun Tian, Xin Yu, Bin Fan, Fuchao Wu, Huub Heijnen, and Vassileios Balntas. Sosnet: Second order similarity regularization for local descriptor learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11016-11025, 2019. 2 +[46] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 4 +[47] Su Wang, Ceslee Montgomery, Jordi Orbay, Vighnesh Birodkar, Aleksandra Faust, Izzeddin Gur, Natasha Jaques, Austin Waters, Jason Baldridge, and Peter Anderson. Less is more: Generating grounded navigation instructions from landmarks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15428-15438, 2022. 2 +[48] Erik Wijmans, Abhishek Kadian, Ari Morcos, Stefan Lee, Irfan Essa, Devi Parikh, Manolis Savva, and Dhruv Batra. Dd-ppo: Learning near-perfect pointgoal navigators from 2.5 billion frames. arXiv preprint arXiv:1911.00357, 2019. 2 +[49] Mitchell Wortsman, Kiana Ehsani, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Learning to learn how to learn: Self-adaptive visual navigation using meta-learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6750-6759, 2019. 1, 2, 3, 5, 6 +[50] Wei Yang, Xiaolong Wang, Ali Farhadi, Abhinav Gupta, and Roozbeh Mottaghi. Visual semantic navigation using scene priors. arXiv preprint arXiv:1810.06543, 2018. 1, 2, 6 +[51] Xin Yu and Fatih Porikli. Ultra-resolving face images by discriminative generative networks. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part V, pages 318-333. Springer, 2016. 2 +[52] Amir R Zamir, Alexander Sax, William Shen, Leonidas J Guibas, Jitendra Malik, and Silvio Savarese. Taskonomy: + +Disentangling task transfer learning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3712-3722, 2018. 2 +[53] Sixian Zhang, Xinhang Song, Yubing Bai, Weijie Li, Yakui Chu, and Shuqiang Jiang. Hierarchical object-to-zone graph for object navigation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15130-15140, 2021. 2, 3, 5, 6 +[54] Yuke Zhu, Roozbeh Mottaghi, Eric Kolve, Joseph J Lim, Abhinav Gupta, Li Fei-Fei, and Ali Farhadi. Target-driven visual navigation in indoor scenes using deep reinforcement learning. In 2017 IEEE international conference on robotics and automation (ICRA), pages 3357-3364. IEEE, 2017. 2 \ No newline at end of file diff --git a/2023/Object-Goal Visual Navigation via Effective Exploration of Relations Among Historical Navigation States/images.zip b/2023/Object-Goal Visual Navigation via Effective Exploration of Relations Among Historical Navigation States/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..04d86aa818d44a29f4150cf3e4b2db64e2de0d3f --- /dev/null +++ b/2023/Object-Goal Visual Navigation via Effective Exploration of Relations Among Historical Navigation States/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fd0dc286d20cdce00423f5c67cc9be1c52ee8bddd284122c4fec8f49a6a0fcef +size 480598 diff --git a/2023/Object-Goal Visual Navigation via Effective Exploration of Relations Among Historical Navigation States/layout.json b/2023/Object-Goal Visual Navigation via Effective Exploration of Relations Among Historical Navigation States/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4b3353351785adbb8e7675517db936a3072d6e46 --- /dev/null +++ b/2023/Object-Goal Visual Navigation via Effective Exploration of Relations Among Historical Navigation States/layout.json @@ -0,0 +1,9129 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 82, + 103, + 511, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 103, + 511, + 140 + ], + "spans": [ + { + "bbox": [ + 82, + 103, + 511, + 140 + ], + "type": "text", + "content": "Object-Goal Visual Navigation via Effective Exploration of Relations among Historical Navigation States" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 175, + 160, + 420, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 160, + 420, + 175 + ], + "spans": [ + { + "bbox": [ + 175, + 160, + 420, + 175 + ], + "type": "text", + "content": "Heming Du" + }, + { + "bbox": [ + 175, + 160, + 420, + 175 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 175, + 160, + 420, + 175 + ], + "type": "text", + "content": ", Lincheng Li" + }, + { + "bbox": [ + 175, + 160, + 420, + 175 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 175, + 160, + 420, + 175 + ], + "type": "text", + "content": ", Zi Huang" + }, + { + "bbox": [ + 175, + 160, + 420, + 175 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 175, + 160, + 420, + 175 + ], + "type": "text", + "content": ", Xin Yu" + }, + { + "bbox": [ + 175, + 160, + 420, + 175 + ], + "type": "inline_equation", + "content": "^{3*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 83, + 175, + 507, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 175, + 507, + 190 + ], + "spans": [ + { + "bbox": [ + 83, + 175, + 507, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 83, + 175, + 507, + 190 + ], + "type": "text", + "content": " Australian National University " + }, + { + "bbox": [ + 83, + 175, + 507, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 83, + 175, + 507, + 190 + ], + "type": "text", + "content": " Netease Fuxi AI Lab " + }, + { + "bbox": [ + 83, + 175, + 507, + 190 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 83, + 175, + 507, + 190 + ], + "type": "text", + "content": " The University of Queensland" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 217, + 192, + 229 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 217, + 192, + 229 + ], + "spans": [ + { + "bbox": [ + 143, + 217, + 192, + 229 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 243, + 290, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 243, + 290, + 531 + ], + "spans": [ + { + "bbox": [ + 46, + 243, + 290, + 531 + ], + "type": "text", + "content": "Object-goal visual navigation aims at steering an agent toward an object via a series of moving steps. Previous works mainly focus on learning informative visual representations for navigation, but overlook the impacts of navigation states on the effectiveness and efficiency of navigation. We observe that high relevance among navigation states will cause navigation inefficiency or failure for existing methods. In this paper, we present a History-inspired Navigation Policy Learning (HiNL) framework to estimate navigation states effectively by exploring relationships among historical navigation states. In HiNL, we propose a History-aware State Estimation (HaSE) module to alleviate the impacts of dominant historical states on the current state estimation. Meanwhile, HaSE also encourages an agent to be alert to the current observation changes, thus enabling the agent to make valid actions. Furthermore, we design a History-based State Regularization (HbSR) to explicitly suppress the correlation among navigation states in training. As a result, our agent can update states more effectively while reducing the correlations among navigation states. Experiments on the artificial platform AI2-THOR (i.e., iTHOR and RoboTHOR) demonstrate that HiNL significantly outperforms state-of-the-art methods on both Success Rate and SPL in unseen testing environments." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 553, + 128, + 566 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 553, + 128, + 566 + ], + "spans": [ + { + "bbox": [ + 47, + 553, + 128, + 566 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 574, + 287, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 574, + 287, + 681 + ], + "spans": [ + { + "bbox": [ + 46, + 574, + 287, + 681 + ], + "type": "text", + "content": "Object-goal visual navigation is to direct an agent to move consecutively toward an object of a specific category. Without knowing the environment map beforehand, at each navigation step, an agent first needs to represent its visual observations, then estimate its navigation states from the visual representations and the preceding states, and at last predict the corresponding action. Therefore, to achieve an effective and efficient navigation system, learning instructive visual representations and navigation states is critical." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 59, + 682, + 287, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 682, + 287, + 694 + ], + "spans": [ + { + "bbox": [ + 59, + 682, + 287, + 694 + ], + "type": "text", + "content": "Prevailing visual navigation works [13, 14, 50] focus on" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 332, + 218, + 524, + 339 + ], + "blocks": [ + { + "bbox": [ + 332, + 218, + 524, + 339 + ], + "lines": [ + { + "bbox": [ + 332, + 218, + 524, + 339 + ], + "spans": [ + { + "bbox": [ + 332, + 218, + 524, + 339 + ], + "type": "image", + "image_path": "286b18ce66bcf8d30701be3925c4d94da35fa6563ff7b8baa454b1b2542f5b00.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 342, + 545, + 370 + ], + "lines": [ + { + "bbox": [ + 305, + 342, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 305, + 342, + 545, + 370 + ], + "type": "text", + "content": "(a) Demonstration of inefficient action predictions caused by highlycorrelated navigation states. Our agent is stuck by an obstacle, i.e., lowprofile sofa, and repeatedly predicts an invalid action, i.e., MoveAhead." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 339, + 372, + 510, + 448 + ], + "blocks": [ + { + "bbox": [ + 339, + 372, + 510, + 448 + ], + "lines": [ + { + "bbox": [ + 339, + 372, + 510, + 448 + ], + "spans": [ + { + "bbox": [ + 339, + 372, + 510, + 448 + ], + "type": "image", + "image_path": "92d327d72606d350f666c12ad46524fbb9736254950a97c10325cdd95ba2947c.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 449, + 545, + 478 + ], + "lines": [ + { + "bbox": [ + 306, + 449, + 545, + 478 + ], + "spans": [ + { + "bbox": [ + 306, + 449, + 545, + 478 + ], + "type": "text", + "content": "(b) Demonstration of the correlation coefficients among navigation states trained in two manners. Navigation states estimated via LSTM are highly-relevant. In contrast, our HiNL produce low-correlated navigation states." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 306, + 484, + 545, + 506 + ], + "lines": [ + { + "bbox": [ + 306, + 484, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 306, + 484, + 545, + 506 + ], + "type": "text", + "content": "Figure 1. Motivation of our proposed History-inspired Navigation Learning (HiNL) framework." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 532, + 545, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 532, + 545, + 699 + ], + "spans": [ + { + "bbox": [ + 304, + 532, + 545, + 699 + ], + "type": "text", + "content": "extracting informative visual representations, while some methods [13, 49] adjust navigation policy during inference. All these approaches commonly employ recurrent neural networks (e.g., LSTM) to estimate navigation states. However, we observe that the navigation states of existing methods [13, 14, 49] exhibit high relevance, as demonstrated in Figure 1b, and the highly-correlated navigation states would lead to inefficient navigation policy (i.e., failure to respond to observation changes rapidly). For instance, as shown in Figure 1a, an agent is stuck by the low-profile sofa and fails to take proper actions to circumvent the obstacle. Hence, we aim to endow an agent with the capability of updating its navigation states effectively while avoiding producing highly-correlated states." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 317, + 701, + 545, + 714 + ], + "type": "text", + "content": "In this work, we propose a History-inspired Navigation" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 703, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 703, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 703, + 135, + 712 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2563" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": "Learning (HiNL) framework to obtain informative navigation states by exploiting the relationships among historical navigation states. HiNL consists of two novel components: (i) a History-aware State Estimation (HaSE) module, and (ii) a History-based State Regularization (HbSR). Here, our HaSE module is designed to generate a state that can be promptly updated according to visual observations. Specifically, HaSE first analyzes the correlations among historical navigation states and then eliminates the influence of dominant historical states on the current state estimation. As a result, an agent is able to predict navigation states which can dynamically react to the current visual observations and then make sensible navigation actions." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 229, + 289, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 229, + 289, + 445 + ], + "spans": [ + { + "bbox": [ + 46, + 229, + 289, + 445 + ], + "type": "text", + "content": "Furthermore, existing reinforcement learning-based object-goal navigation systems [13, 14, 49] often assume the navigation state transition exhibits the first-order Markov property. This would allow the emergence of high correlations among navigation states, leading to inferior navigation policy. To address this issue, we introduce an explicit constraint on the correlation among all the states, namely History-based State Regularization (HbSR). To be specific, HbSR enforces to relevance (i.e., correlations) between a state and all its preceding states (except its previous state) to be low. Here, we do not constrain states of two consecutive steps because temporally close states generally have relevance in practice considering the navigation continuity. After training with our HbSR, the correlations among the navigation states become much lower (see Figure 1b). This pheromone also indicates HiNL effectively updates states. Hence, our navigation system can respond to observation changes adaptively." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 447, + 289, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 447, + 289, + 541 + ], + "spans": [ + { + "bbox": [ + 46, + 447, + 289, + 541 + ], + "type": "text", + "content": "To demonstrate the superiority of HiNL, we conduct experiments in the widely-adopted artificial environment iTHOR [26] and RoboTHOR [11]. HiNL outperforms the state-of-the-art by a large margin. To be specific, we improve the Success Rate (SR) from " + }, + { + "bbox": [ + 46, + 447, + 289, + 541 + ], + "type": "inline_equation", + "content": "72.2\\%" + }, + { + "bbox": [ + 46, + 447, + 289, + 541 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 447, + 289, + 541 + ], + "type": "inline_equation", + "content": "80.1\\%" + }, + { + "bbox": [ + 46, + 447, + 289, + 541 + ], + "type": "text", + "content": " and Success weighted by Path Length (SPL) from 0.449 to 0.498 in iTHOR. Overall, our major contributions are summarized as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 550, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 59, + 550, + 287, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 550, + 287, + 586 + ], + "spans": [ + { + "bbox": [ + 59, + 550, + 287, + 586 + ], + "type": "text", + "content": "- We propose a History-inspired Navigation Policy (HiNL) framework to effectively estimate navigation states by utilizing historical states." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 596, + 287, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 596, + 287, + 656 + ], + "spans": [ + { + "bbox": [ + 59, + 596, + 287, + 656 + ], + "type": "text", + "content": "- We design a History-aware State Estimation (HaSE) to eliminate dominant historical states in the current state estimation. Therefore, the agent reduces the impact of distant navigation states on the state estimation, and thus reacts dynamically to the observation changes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 665, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 665, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 665, + 287, + 713 + ], + "type": "text", + "content": "- We introduce a History-based State Regularization (HbSR) to explicitly constrain the correlations among navigation states. By doing this, the agent can effectively update navigation states with low relevance." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 71, + 398, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 398, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 398, + 84 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 92, + 547, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 92, + 547, + 201 + ], + "spans": [ + { + "bbox": [ + 304, + 92, + 547, + 201 + ], + "type": "text", + "content": "Visual navigation. Traditional works [3, 4, 32] often leverage an entire environment map for navigation and divide the task into three parts: mapping, localization, and path planning. However, environment maps are generally unavailable in unseen environments. Dissanayake et al. [12] adopt simultaneous localization and mapping (SLAM) to infer robot positions. Campari et al. [5] learn agent states via a Taskonomy model bank [52], but they need an RGB-D sensor to construct an online map during navigation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 202, + 547, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 202, + 547, + 370 + ], + "spans": [ + { + "bbox": [ + 304, + 202, + 547, + 370 + ], + "type": "text", + "content": "Recently, due to significant advancements in deep learning [16,22,45,51], reinforcement learning-based navigation methods [28, 29, 31, 33, 54] take visual observations as inputs and predict navigation actions. Vision-Language Navigation (VLN) approaches [9, 10, 15, 39, 40] steer an agent to the target based on its visual observations and navigation guidance in natural language. Similar to VLN, point-goal visual navigation methods [44,48] aim at driving an agent to a given point with step-wise directional indications. Moreover, audio-visual navigation methods [7, 8, 18] utilize additional audio signals to move a robot to the target position. Al-Halah et al. [1] propose a transfer learning model for multiple navigation tasks by embedding various navigation goals, e.g., image, sketch, and audio." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 372, + 548, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 372, + 548, + 586 + ], + "spans": [ + { + "bbox": [ + 304, + 372, + 548, + 586 + ], + "type": "text", + "content": "Our work falls in the field of object-goal visual navigation [27, 31, 36, 47, 50]. However, existing object-goal navigation methods mainly focus on representing visual features comprehensively while we investigate the impact of navigation states on navigation performance. Wortsman et al. [49] exploit word embedding (i.e., GloVe embedding [35]) to represent the target category and introduce a meta network that mimics a reward function during inference. Du et al. [13] introduce an object relation graph, dubbed ORG, to encode visual observations. They also design a tentative policy for deadlock avoidance and adjust the navigation policy in unseen testing environments. The Hierarchical Object-to-Zone (HOZ) graph [53] offers coarse-to-fine guidance based on real-time updates. Additionally, VT-Net [14] incorporates object and region features with location cues, and EmbCLIP [24] leverages the contrastive language image pretraining encoder for visual navigation tasks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 713 + ], + "type": "text", + "content": "Correlation Modeling in Reinforcement Learning. Several methods [37, 43] explore correlations in hidden Markov models for inverse reinforcement learning (IRL). For action prediction, Hester et al. [20] propose Texlore to model correlations within the transition dynamics via a random forest. Vsovsic et al. [42] introduce a Bayesian approach to learn policy from demonstrations of experts by capturing correlations among actions. Alt et al. [2] design a Bayesian learning framework to establish temporal and spatial correlations among actions. Furthermore," + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2564" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 83, + 531, + 210 + ], + "blocks": [ + { + "bbox": [ + 69, + 83, + 531, + 210 + ], + "lines": [ + { + "bbox": [ + 69, + 83, + 531, + 210 + ], + "spans": [ + { + "bbox": [ + 69, + 83, + 531, + 210 + ], + "type": "image", + "image_path": "8f68e97d22c1b7e7a9f7bd6f7d1ab1c50e06d13186a68d125343093ad5edd1ea.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 215, + 547, + 273 + ], + "lines": [ + { + "bbox": [ + 45, + 215, + 547, + 273 + ], + "spans": [ + { + "bbox": [ + 45, + 215, + 547, + 273 + ], + "type": "text", + "content": "Figure 2. Our History-inspired Navigation Policy Learning (HiNL) framework. HiNL takes visual representations as input and outputs navigation actions. HiNL involves two innovative parts: a History-aware State Estimation (HaSE) module and a History-based State Regularization (HbSR). HaSE is proposed to estimate navigation states that can reflect current observation changes from the perspective of network design, while HbSR is designed to enforce the informativeness of states from the view of the training objective. Both of them help to achieve effective and efficient navigation policy." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 292, + 288, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 292, + 288, + 352 + ], + "spans": [ + { + "bbox": [ + 46, + 292, + 288, + 352 + ], + "type": "text", + "content": "Sermanet [41] propose a self-supervised TCN for learning robotic behaviors and representations from unlabeled multi-viewpoint videos. TCN uses a metric learning loss to create viewpoint-invariant representations that capture relationships between end-effectors." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 353, + 287, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 353, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 46, + 353, + 287, + 426 + ], + "type": "text", + "content": "Different from previous methods that establish correlations among actions, our HiNL aims to eliminate high relevance among navigation states. To the best of our knowledge, our work HiNL is the first attempt to explicitly leverage the correlations among navigation states for object-goal navigation and achieves superior navigation performance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 437, + 159, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 437, + 159, + 451 + ], + "spans": [ + { + "bbox": [ + 47, + 437, + 159, + 451 + ], + "type": "text", + "content": "3. Proposed Methods" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 457, + 288, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 457, + 288, + 590 + ], + "spans": [ + { + "bbox": [ + 46, + 457, + 288, + 590 + ], + "type": "text", + "content": "In this work, we propose a History-inspired Navigation Policy (HiNL) framework to estimate navigation states effectively. As demonstrated in Figure 2, in each step, the agent first adopts a visual feature extractor to process the current RGB observation. Then, the agent estimates the current navigation state via a History-aware State Estimation (HaSE) module and predicts a navigation action. At last, we introduce a History-based State Regularization (HbSR) to constrain the correlations among navigation states during training, thus achieving low-correlated navigation states and highly effective navigation policy." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 598, + 192, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 598, + 192, + 612 + ], + "spans": [ + { + "bbox": [ + 47, + 598, + 192, + 612 + ], + "type": "text", + "content": "3.1. Task Definition and Setup" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 715 + ], + "type": "text", + "content": "Following previous works [13, 14, 49, 53], in the objectgoal visual navigation task, an RGB monocular camera is the only sensor available to an agent. Prior knowledge of the entire environment, such as the topological map and 3D meshes, is unknown to the agent. An environment is composed of discrete grids, and the agent selects one of six actions, i.e., MoveAhead, RotateLeft, RotateRight, LookUp, LookDown, Done. Specif" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 292, + 545, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 292, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 304, + 292, + 545, + 327 + ], + "type": "text", + "content": "ically, the grid size is 0.25 meters, and the rotation angles of RotateLeft/RotateRight and LookUp/LookDown are " + }, + { + "bbox": [ + 304, + 292, + 545, + 327 + ], + "type": "inline_equation", + "content": "45^{\\circ}" + }, + { + "bbox": [ + 304, + 292, + 545, + 327 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 292, + 545, + 327 + ], + "type": "inline_equation", + "content": "30^{\\circ}" + }, + { + "bbox": [ + 304, + 292, + 545, + 327 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 328, + 546, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 328, + 546, + 555 + ], + "spans": [ + { + "bbox": [ + 304, + 328, + 546, + 555 + ], + "type": "text", + "content": "At the start of each episode, the environment randomly chooses one of the available object categories " + }, + { + "bbox": [ + 304, + 328, + 546, + 555 + ], + "type": "inline_equation", + "content": "T \\in \\{Sink, \\dots, Microwave\\}" + }, + { + "bbox": [ + 304, + 328, + 546, + 555 + ], + "type": "text", + "content": " as the target class. Then, the agent is teleported to a random start position " + }, + { + "bbox": [ + 304, + 328, + 546, + 555 + ], + "type": "inline_equation", + "content": "s = \\{x, y, \\theta_h, \\theta_v\\}" + }, + { + "bbox": [ + 304, + 328, + 546, + 555 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 328, + 546, + 555 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 304, + 328, + 546, + 555 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 328, + 546, + 555 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 328, + 546, + 555 + ], + "type": "text", + "content": " stand for the coordinates, and " + }, + { + "bbox": [ + 304, + 328, + 546, + 555 + ], + "type": "inline_equation", + "content": "\\theta_h" + }, + { + "bbox": [ + 304, + 328, + 546, + 555 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 328, + 546, + 555 + ], + "type": "inline_equation", + "content": "\\theta_v" + }, + { + "bbox": [ + 304, + 328, + 546, + 555 + ], + "type": "text", + "content": " represent the horizontal angle and the vertical angle for the view of the agent. The agent aims to predict an action distribution based on a learned navigation policy and operates the action with the highest probability for navigation. Then, the agent moves to the next navigation step if it selects a moving action or terminates the episode if it selects the action Done. After finishing the episode, we regard the agent as successful when the following three conditions are met simultaneously: (i) the agent selects the ending action Done within the maximum step length of an episode; (ii) there is at least one instance of the specified category in the view of the agent; (iii) the distance between the visible target and the agent is less than the threshold, i.e., 1.5 meters. Otherwise, the episode is considered as a failure case." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 562, + 485, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 562, + 485, + 575 + ], + "spans": [ + { + "bbox": [ + 306, + 562, + 485, + 575 + ], + "type": "text", + "content": "3.2. Visual Representation Extraction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 580, + 545, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 580, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 580, + 545, + 677 + ], + "type": "text", + "content": "Similar to previous works [13, 49], we first adopt a deep neural network, i.e., ResNet [19], to extract global features from an observed RGB image. Meanwhile, a detection module, e.g., Fast-RCNN [38] and DETR [6], is introduced to extract object features. Then, the global and object features are merged into a visual representation " + }, + { + "bbox": [ + 304, + 580, + 545, + 677 + ], + "type": "inline_equation", + "content": "\\boldsymbol{v}_t" + }, + { + "bbox": [ + 304, + 580, + 545, + 677 + ], + "type": "text", + "content": " by graph neural network [13, 53] or transformer " + }, + { + "bbox": [ + 304, + 580, + 545, + 677 + ], + "type": "inline_equation", + "content": "[14]^1" + }, + { + "bbox": [ + 304, + 580, + 545, + 677 + ], + "type": "text", + "content": ". For convenience, we adopt " + }, + { + "bbox": [ + 304, + 580, + 545, + 677 + ], + "type": "inline_equation", + "content": "\\boldsymbol{v}_t" + }, + { + "bbox": [ + 304, + 580, + 545, + 677 + ], + "type": "text", + "content": " to denote the visual feature extracted" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 305, + 683, + 545, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 683, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 683, + 545, + 713 + ], + "type": "text", + "content": "1In our experiments, we employ both CNN-based [13] and Transformer-based [14] visual representations, demonstrating our framework HiNL is agnostic against various visual representations." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2565" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 76, + 521, + 157 + ], + "blocks": [ + { + "bbox": [ + 72, + 76, + 521, + 157 + ], + "lines": [ + { + "bbox": [ + 72, + 76, + 521, + 157 + ], + "spans": [ + { + "bbox": [ + 72, + 76, + 521, + 157 + ], + "type": "image", + "image_path": "08a353a65aee560784829d4132136abe13f714d9d14284a6fa76579c3665ebbf.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 157, + 547, + 201 + ], + "lines": [ + { + "bbox": [ + 46, + 157, + 547, + 201 + ], + "spans": [ + { + "bbox": [ + 46, + 157, + 547, + 201 + ], + "type": "text", + "content": "Figure 3. Illustration of History-aware State Estimation (HaSE). The agent first adopts a Historical State Guidance Extraction module to estimate a state guidance that indicates the dominant historical states. Then, the agent extracts a state innovation that reflects incremental information of current observations, via a State-induced Perceptron module. At last, the agent generates the current navigation state estimation by fusing the state innovation and guidance in a History-aware Fusion module." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 213, + 214, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 213, + 214, + 224 + ], + "spans": [ + { + "bbox": [ + 47, + 213, + 214, + 224 + ], + "type": "text", + "content": "from the RGB observation at time step " + }, + { + "bbox": [ + 47, + 213, + 214, + 224 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 213, + 214, + 224 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 231, + 218, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 231, + 218, + 243 + ], + "spans": [ + { + "bbox": [ + 47, + 231, + 218, + 243 + ], + "type": "text", + "content": "3.3. History-aware State Estimation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 249, + 288, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 249, + 288, + 417 + ], + "spans": [ + { + "bbox": [ + 46, + 249, + 288, + 417 + ], + "type": "text", + "content": "Our History-aware State Estimation (HaSE) estimates the navigation states by eliminating the influence of dominant historical states and focusing on the present observation changes. As demonstrated in Figure 3, after obtaining the visual representation " + }, + { + "bbox": [ + 46, + 249, + 288, + 417 + ], + "type": "inline_equation", + "content": "\\boldsymbol{v}_t" + }, + { + "bbox": [ + 46, + 249, + 288, + 417 + ], + "type": "text", + "content": ", we design a Historical State Guidance Extraction module to predict a state guidance " + }, + { + "bbox": [ + 46, + 249, + 288, + 417 + ], + "type": "inline_equation", + "content": "\\boldsymbol{g}_t" + }, + { + "bbox": [ + 46, + 249, + 288, + 417 + ], + "type": "text", + "content": " indicating dominant historical states at time step " + }, + { + "bbox": [ + 46, + 249, + 288, + 417 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 249, + 288, + 417 + ], + "type": "text", + "content": ". Then, we introduce a State-induced Perceptron module to obtain the current state innovation " + }, + { + "bbox": [ + 46, + 249, + 288, + 417 + ], + "type": "inline_equation", + "content": "\\boldsymbol{i}_t" + }, + { + "bbox": [ + 46, + 249, + 288, + 417 + ], + "type": "text", + "content": ", which encodes the new information brought by the current visual observations and the last step action. Next, the history-aware guidance " + }, + { + "bbox": [ + 46, + 249, + 288, + 417 + ], + "type": "inline_equation", + "content": "\\boldsymbol{g}_t" + }, + { + "bbox": [ + 46, + 249, + 288, + 417 + ], + "type": "text", + "content": " is employed to eliminate the dominant historical states and facilitate the estimation of the current navigation state " + }, + { + "bbox": [ + 46, + 249, + 288, + 417 + ], + "type": "inline_equation", + "content": "\\boldsymbol{s}_t" + }, + { + "bbox": [ + 46, + 249, + 288, + 417 + ], + "type": "text", + "content": " via our proposed History-aware Fusion module." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 422, + 287, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 422, + 287, + 553 + ], + "spans": [ + { + "bbox": [ + 46, + 422, + 287, + 553 + ], + "type": "text", + "content": "Historical State Guidance Extraction. In order to eliminate the dominant historical states in the current state estimation, we extract guidance for the dominant parts by exploring the relationships among historical navigation states. The agent roughly approximates a prior state " + }, + { + "bbox": [ + 46, + 422, + 287, + 553 + ], + "type": "inline_equation", + "content": "s_t'" + }, + { + "bbox": [ + 46, + 422, + 287, + 553 + ], + "type": "text", + "content": " based on the last state " + }, + { + "bbox": [ + 46, + 422, + 287, + 553 + ], + "type": "inline_equation", + "content": "s_{t-1}" + }, + { + "bbox": [ + 46, + 422, + 287, + 553 + ], + "type": "text", + "content": ", action " + }, + { + "bbox": [ + 46, + 422, + 287, + 553 + ], + "type": "inline_equation", + "content": "a_{t-1}" + }, + { + "bbox": [ + 46, + 422, + 287, + 553 + ], + "type": "text", + "content": ", and the current visual representation " + }, + { + "bbox": [ + 46, + 422, + 287, + 553 + ], + "type": "inline_equation", + "content": "v_t" + }, + { + "bbox": [ + 46, + 422, + 287, + 553 + ], + "type": "text", + "content": ". The prior state " + }, + { + "bbox": [ + 46, + 422, + 287, + 553 + ], + "type": "inline_equation", + "content": "s_t'" + }, + { + "bbox": [ + 46, + 422, + 287, + 553 + ], + "type": "text", + "content": " is encoded with the last navigation state and the current observation, and thus we refer " + }, + { + "bbox": [ + 46, + 422, + 287, + 553 + ], + "type": "inline_equation", + "content": "s_t'" + }, + { + "bbox": [ + 46, + 422, + 287, + 553 + ], + "type": "text", + "content": " to the query about the dominant historical states. Let FFN be a feed-forward network, " + }, + { + "bbox": [ + 46, + 422, + 287, + 553 + ], + "type": "inline_equation", + "content": "s_t'" + }, + { + "bbox": [ + 46, + 422, + 287, + 553 + ], + "type": "text", + "content": " is formulated as follows," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 80, + 561, + 287, + 575 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 561, + 287, + 575 + ], + "spans": [ + { + "bbox": [ + 80, + 561, + 287, + 575 + ], + "type": "interline_equation", + "content": "\\boldsymbol {s} _ {t} ^ {\\prime} = \\sigma (\\operatorname {F F N} \\left(\\boldsymbol {s} _ {t - 1}\\right) + \\operatorname {F F N} \\left(\\left[ \\boldsymbol {v} _ {t}; \\boldsymbol {a} _ {t - 1} \\right]\\right)), \\tag {1}", + "image_path": "d604b43a77426fd784b107d01f51525246aee047de89a1740a6795665d0ded3d.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 582, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 605 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 582, + 287, + 605 + ], + "type": "inline_equation", + "content": "[;]" + }, + { + "bbox": [ + 46, + 582, + 287, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 582, + 287, + 605 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 46, + 582, + 287, + 605 + ], + "type": "text", + "content": " represent the concatenation operation and sigmoid function, respectively." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "text", + "content": "After obtaining " + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "inline_equation", + "content": "s_t'" + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "text", + "content": ", the agent memorizes a set of the historical navigation states " + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbb{S}_{t - \\tau : t - 1} = \\{s_{t - \\tau}, \\dots, s_{t - 1}\\}" + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "text", + "content": " in the last " + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "text", + "content": " steps. In order to explore the temporal relationships among historical navigation states, " + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathbb{S}_{t - \\tau : t - 1}" + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "text", + "content": " are composed additively with temporal embeddings. Concretely, we adopt absolute temporal embedding [46] based on the timestamp at which a state is estimated. By doing this, the agent perceives not only the temporal orders among the last " + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "text", + "content": " navigation states, but also the length of the current" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 213, + 545, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 213, + 545, + 272 + ], + "spans": [ + { + "bbox": [ + 304, + 213, + 545, + 272 + ], + "type": "text", + "content": "episode. Therefore, the agent can exploit temporal information to eliminate the distant navigation states in extracting the guidance. In addition, we analyze the impacts of different temporal embedding methods (i.e., no/learnable/relative temporal embedding) in Table 2." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 272, + 545, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 272, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 304, + 272, + 545, + 319 + ], + "type": "text", + "content": "Afterward, we pass " + }, + { + "bbox": [ + 304, + 272, + 545, + 319 + ], + "type": "inline_equation", + "content": "s_t'" + }, + { + "bbox": [ + 304, + 272, + 545, + 319 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 272, + 545, + 319 + ], + "type": "inline_equation", + "content": "\\mathbb{S}_{t - \\tau : t - 1}" + }, + { + "bbox": [ + 304, + 272, + 545, + 319 + ], + "type": "text", + "content": " through the historical state decoder, i.e., a cross-attention module. We refer to the output " + }, + { + "bbox": [ + 304, + 272, + 545, + 319 + ], + "type": "inline_equation", + "content": "z_t" + }, + { + "bbox": [ + 304, + 272, + 545, + 319 + ], + "type": "text", + "content": " as the dominant historical states, and calculate it as follows," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 359, + 327, + 545, + 340 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 327, + 545, + 340 + ], + "spans": [ + { + "bbox": [ + 359, + 327, + 545, + 340 + ], + "type": "interline_equation", + "content": "\\boldsymbol {z} _ {\\boldsymbol {t}} = \\operatorname {D e c o d e r} _ {\\mathrm {h i s}} \\left(\\boldsymbol {s} _ {\\boldsymbol {t}} ^ {\\prime}, \\mathbb {S} _ {t - \\tau : t - 1}\\right), \\tag {2}", + "image_path": "6764bac36266c0cec8d71a5e5a480477931f3574c76fae4d8ba830acdc10b0a5.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 347, + 545, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 347, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 304, + 347, + 545, + 396 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 347, + 545, + 396 + ], + "type": "inline_equation", + "content": "\\mathrm{Decoder}_{\\mathrm{his}}(\\cdot, \\cdot)" + }, + { + "bbox": [ + 304, + 347, + 545, + 396 + ], + "type": "text", + "content": " indicates the historical state decoder, the first argument for query and the other for key/values. Then, we formulate the guidance " + }, + { + "bbox": [ + 304, + 347, + 545, + 396 + ], + "type": "inline_equation", + "content": "g_{t}" + }, + { + "bbox": [ + 304, + 347, + 545, + 396 + ], + "type": "text", + "content": " as the difference between " + }, + { + "bbox": [ + 304, + 347, + 545, + 396 + ], + "type": "inline_equation", + "content": "s_{t-1}" + }, + { + "bbox": [ + 304, + 347, + 545, + 396 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 347, + 545, + 396 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 304, + 347, + 545, + 396 + ], + "type": "text", + "content": ". Therefore, we compute " + }, + { + "bbox": [ + 304, + 347, + 545, + 396 + ], + "type": "inline_equation", + "content": "g_{t}" + }, + { + "bbox": [ + 304, + 347, + 545, + 396 + ], + "type": "text", + "content": " as follows," + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 382, + 407, + 545, + 419 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 407, + 545, + 419 + ], + "spans": [ + { + "bbox": [ + 382, + 407, + 545, + 419 + ], + "type": "interline_equation", + "content": "\\boldsymbol {g} _ {t} = \\sigma \\left(\\boldsymbol {s} _ {t - 1} - \\boldsymbol {z} _ {t}\\right). \\tag {3}", + "image_path": "29e9ad396f0555c2b1e4cb3551e4b1bc6c595bc96e3769f9418fcfa965933556.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 423, + 545, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 423, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 304, + 423, + 545, + 448 + ], + "type": "text", + "content": "In addition, we set " + }, + { + "bbox": [ + 304, + 423, + 545, + 448 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 423, + 545, + 448 + ], + "type": "text", + "content": " as 5 in our experiments and demonstrate the impacts of different " + }, + { + "bbox": [ + 304, + 423, + 545, + 448 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 304, + 423, + 545, + 448 + ], + "type": "text", + "content": " in Figure 5b." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 452, + 545, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 452, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 304, + 452, + 545, + 536 + ], + "type": "text", + "content": "State-induced Perceptron. Apart from eliminating the dominant historical states, we also aim to encourage the agent to emphasize the observed changes in the current state estimation. Therefore, we extract a state innovation " + }, + { + "bbox": [ + 304, + 452, + 545, + 536 + ], + "type": "inline_equation", + "content": "i_{t}" + }, + { + "bbox": [ + 304, + 452, + 545, + 536 + ], + "type": "text", + "content": " based on " + }, + { + "bbox": [ + 304, + 452, + 545, + 536 + ], + "type": "inline_equation", + "content": "v_{t}" + }, + { + "bbox": [ + 304, + 452, + 545, + 536 + ], + "type": "text", + "content": " to encode the observed changes at the current step. To be specific, we projects " + }, + { + "bbox": [ + 304, + 452, + 545, + 536 + ], + "type": "inline_equation", + "content": "v_{t}, s_{t-1}" + }, + { + "bbox": [ + 304, + 452, + 545, + 536 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 452, + 545, + 536 + ], + "type": "inline_equation", + "content": "a_{t-1}" + }, + { + "bbox": [ + 304, + 452, + 545, + 536 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 452, + 545, + 536 + ], + "type": "inline_equation", + "content": "i_{t}" + }, + { + "bbox": [ + 304, + 452, + 545, + 536 + ], + "type": "text", + "content": ". Inspired by LSTM [21], we compute " + }, + { + "bbox": [ + 304, + 452, + 545, + 536 + ], + "type": "inline_equation", + "content": "i_{t}" + }, + { + "bbox": [ + 304, + 452, + 545, + 536 + ], + "type": "text", + "content": " as," + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 330, + 542, + 545, + 561 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 542, + 545, + 561 + ], + "spans": [ + { + "bbox": [ + 330, + 542, + 545, + 561 + ], + "type": "interline_equation", + "content": "\\boldsymbol {i} _ {t} = \\tanh (\\operatorname {F F N} \\left(\\boldsymbol {s} _ {t - 1}\\right) + \\operatorname {F F N} ([ \\boldsymbol {v} _ {t}; \\boldsymbol {a} _ {t - 1} ])) \\odot \\tag {4}", + "image_path": "ec4461094dcfdd772e234b8a5c163be788bb8e17b6a05a9efdad5b1f2470bb7c.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 351, + 557, + 499, + 571 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 557, + 499, + 571 + ], + "spans": [ + { + "bbox": [ + 351, + 557, + 499, + 571 + ], + "type": "interline_equation", + "content": "\\sigma (\\operatorname {F F N} \\left(\\boldsymbol {s} _ {\\boldsymbol {t} - \\mathbf {1}}\\right) + \\operatorname {F F N} ([ \\boldsymbol {v} _ {\\boldsymbol {t}}; \\boldsymbol {a} _ {\\boldsymbol {t} - \\mathbf {1}} ])),", + "image_path": "3a9250076c0d620b2b42fa23fd0344414d5d1a736042dbb3ba55e57524d40b50.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 577, + 545, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 577, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 306, + 577, + 545, + 601 + ], + "type": "text", + "content": "where tanh represents hyperbolic tangent function, and " + }, + { + "bbox": [ + 306, + 577, + 545, + 601 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 306, + 577, + 545, + 601 + ], + "type": "text", + "content": " stands for the element-wise product." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 605, + 545, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 700 + ], + "type": "text", + "content": "History-aware Fusion. After obtaining the guidance " + }, + { + "bbox": [ + 304, + 605, + 545, + 700 + ], + "type": "inline_equation", + "content": "g_{t}" + }, + { + "bbox": [ + 304, + 605, + 545, + 700 + ], + "type": "text", + "content": " and state innovation " + }, + { + "bbox": [ + 304, + 605, + 545, + 700 + ], + "type": "inline_equation", + "content": "i_{t}" + }, + { + "bbox": [ + 304, + 605, + 545, + 700 + ], + "type": "text", + "content": ", we aim to estimate the current navigation state " + }, + { + "bbox": [ + 304, + 605, + 545, + 700 + ], + "type": "inline_equation", + "content": "s_{t}" + }, + { + "bbox": [ + 304, + 605, + 545, + 700 + ], + "type": "text", + "content": ". In order to eliminate the influence of dominant navigation states, the agent first merges " + }, + { + "bbox": [ + 304, + 605, + 545, + 700 + ], + "type": "inline_equation", + "content": "s_{t-1}" + }, + { + "bbox": [ + 304, + 605, + 545, + 700 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 605, + 545, + 700 + ], + "type": "inline_equation", + "content": "g_{t}" + }, + { + "bbox": [ + 304, + 605, + 545, + 700 + ], + "type": "text", + "content": " by multiplication. Then the agent adds " + }, + { + "bbox": [ + 304, + 605, + 545, + 700 + ], + "type": "inline_equation", + "content": "i_{t}" + }, + { + "bbox": [ + 304, + 605, + 545, + 700 + ], + "type": "text", + "content": " to the result of multiplication to incorporate the observation changes. Formally, we formulate the current navigation state " + }, + { + "bbox": [ + 304, + 605, + 545, + 700 + ], + "type": "inline_equation", + "content": "s_{t}" + }, + { + "bbox": [ + 304, + 605, + 545, + 700 + ], + "type": "text", + "content": " estimation as follows," + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 366, + 701, + 545, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 366, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 366, + 701, + 545, + 714 + ], + "type": "interline_equation", + "content": "\\boldsymbol {s} _ {t} = \\tanh \\left(\\boldsymbol {s} _ {t - 1} \\odot \\boldsymbol {g} _ {t} + \\boldsymbol {i} _ {t}\\right). \\tag {5}", + "image_path": "caa75ceba4c04c8b13152d76bfb754068ab57e6d7c5a267452ec3ab92e60b71f.jpg" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2566" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 71, + 531, + 152 + ], + "blocks": [ + { + "bbox": [ + 61, + 71, + 531, + 152 + ], + "lines": [ + { + "bbox": [ + 61, + 71, + 531, + 152 + ], + "spans": [ + { + "bbox": [ + 61, + 71, + 531, + 152 + ], + "type": "image", + "image_path": "2b9bc7c5947ecaf5265841f8714f8173c6bdf25d7fd47c5518cc3d60154f60cf.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 45, + 152, + 547, + 196 + ], + "lines": [ + { + "bbox": [ + 45, + 152, + 547, + 196 + ], + "spans": [ + { + "bbox": [ + 45, + 152, + 547, + 196 + ], + "type": "text", + "content": "Figure 4. Illustration of History-based State Regularization (HbSR). (a) First-order Markov state transition that is widely adopted by LSTM based navigation policy learning. Those methods will produce highly-correlated among states. (b) Our HbSR constrains higher-order relations among states, and leads to low-correlated states. As the states are more informative (i.e., less correlated), our navigation policy is more effective and efficiency." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 208, + 287, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 208, + 287, + 243 + ], + "spans": [ + { + "bbox": [ + 46, + 208, + 287, + 243 + ], + "type": "text", + "content": "Benefiting from HaSE, the agent can effectively update the navigation states by being alert to the present observation and neglecting the influence of distant states." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 251, + 234, + 264 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 251, + 234, + 264 + ], + "spans": [ + { + "bbox": [ + 47, + 251, + 234, + 264 + ], + "type": "text", + "content": "3.4. History-based State Regularization" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 269, + 287, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 269, + 287, + 435 + ], + "spans": [ + { + "bbox": [ + 46, + 269, + 287, + 435 + ], + "type": "text", + "content": "Previous reinforcement learning-based object-goal visual navigation methods are designed based on the first-order Markov assumption, while there are no explicit constraints on the correlations among navigation states. We observe that high relevance among navigation states exhibit in the experiments of previous methods [13, 14] and this is also one of the reasons that prior works yield failure cases. For instance, an agent is stuck in front of an object rather than adopting a rotation action or trapped in a loop. From these failure cases, we can see that the navigation states are highly correlated as the adopted actions are periodically the same. In order to avoid inefficient navigation caused by highly-relevant navigation states, we aim to exert an explicit constraint on the correlations among navigation states." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 437, + 287, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 287, + 544 + ], + "type": "text", + "content": "We introduce a novel History-based State Regularization (HbSR) to reduce the correlations among navigation states from the perspective of the training objective. In order to measure the correlations among navigation states, we adopt the Pearson correlation coefficient [17]. To be specific, after the agent terminates an episode with " + }, + { + "bbox": [ + 46, + 437, + 287, + 544 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 437, + 287, + 544 + ], + "type": "text", + "content": " steps, we obtain a set of navigation states " + }, + { + "bbox": [ + 46, + 437, + 287, + 544 + ], + "type": "inline_equation", + "content": "\\mathbb{S}_{0:T} = \\{s_0,\\dots,s_T\\}" + }, + { + "bbox": [ + 46, + 437, + 287, + 544 + ], + "type": "text", + "content": ". Given two navigation states " + }, + { + "bbox": [ + 46, + 437, + 287, + 544 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 46, + 437, + 287, + 544 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 437, + 287, + 544 + ], + "type": "inline_equation", + "content": "s_j" + }, + { + "bbox": [ + 46, + 437, + 287, + 544 + ], + "type": "text", + "content": ", we first compute the covariance of the two states as follows," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 71, + 553, + 287, + 567 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 553, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 71, + 553, + 287, + 567 + ], + "type": "interline_equation", + "content": "\\operatorname {C o v} \\left(\\boldsymbol {s} _ {i}, \\boldsymbol {s} _ {j}\\right) = \\mathbb {E} \\left(\\left(\\boldsymbol {s} _ {i} - \\mathbb {E} \\left(\\boldsymbol {s} _ {i}\\right)\\right) \\left(\\boldsymbol {s} _ {j} - \\mathbb {E} \\left(\\boldsymbol {s} _ {j}\\right)\\right)\\right). \\tag {6}", + "image_path": "a2f4f774bce4e11724b16dca75f6f41ac6d03c1bc8d3a6d53af7d66a5aa4f4d9.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 574, + 287, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 574, + 287, + 597 + ], + "spans": [ + { + "bbox": [ + 47, + 574, + 287, + 597 + ], + "type": "text", + "content": "Then, we compute the correlation coefficient " + }, + { + "bbox": [ + 47, + 574, + 287, + 597 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 47, + 574, + 287, + 597 + ], + "type": "text", + "content": " for two states based on the covariance formulated as," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 98, + 605, + 287, + 634 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 98, + 605, + 287, + 634 + ], + "spans": [ + { + "bbox": [ + 98, + 605, + 287, + 634 + ], + "type": "interline_equation", + "content": "\\rho \\left(\\boldsymbol {s} _ {i}, \\boldsymbol {s} _ {j}\\right) = \\frac {\\operatorname {C o v} \\left(\\boldsymbol {s} _ {i} , \\boldsymbol {s} _ {j}\\right)}{\\sqrt {\\operatorname {V a r} \\left(\\boldsymbol {s} _ {i}\\right) \\operatorname {V a r} \\left(\\boldsymbol {s} _ {j}\\right)}}, \\tag {7}", + "image_path": "15d317669729fd823c97fcca1c5dfd85999fd32843eb604ddac992fb00f0df70.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 642, + 217, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 642, + 217, + 653 + ], + "spans": [ + { + "bbox": [ + 47, + 642, + 217, + 653 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 642, + 217, + 653 + ], + "type": "inline_equation", + "content": "\\operatorname{Var}(s_i)" + }, + { + "bbox": [ + 47, + 642, + 217, + 653 + ], + "type": "text", + "content": " denotes the variance of " + }, + { + "bbox": [ + 47, + 642, + 217, + 653 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 47, + 642, + 217, + 653 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 654, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 654, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 654, + 287, + 712 + ], + "type": "text", + "content": "The correlation coefficient is expressed in a range " + }, + { + "bbox": [ + 47, + 654, + 287, + 712 + ], + "type": "inline_equation", + "content": "[-1, + 1]" + }, + { + "bbox": [ + 47, + 654, + 287, + 712 + ], + "type": "text", + "content": ". Two states are highly relevant when their correlation is close to either 1 or " + }, + { + "bbox": [ + 47, + 654, + 287, + 712 + ], + "type": "inline_equation", + "content": "-1" + }, + { + "bbox": [ + 47, + 654, + 287, + 712 + ], + "type": "text", + "content": ", whereas " + }, + { + "bbox": [ + 47, + 654, + 287, + 712 + ], + "type": "inline_equation", + "content": "\\rho = 0" + }, + { + "bbox": [ + 47, + 654, + 287, + 712 + ], + "type": "text", + "content": " indicates irrelevant between two states. Therefore, in order to eliminate high relevance among states, we enforce the correla" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 208, + 545, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 208, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 304, + 208, + 545, + 363 + ], + "type": "text", + "content": "tions of states to be close to 0. In particular, since inefficient navigation is generally caused by positive correlations of navigation states, we only apply the proposed regularization on correlations with positive values. According to the first-order Markov property, the present state " + }, + { + "bbox": [ + 304, + 208, + 545, + 363 + ], + "type": "inline_equation", + "content": "s_t" + }, + { + "bbox": [ + 304, + 208, + 545, + 363 + ], + "type": "text", + "content": " depends on its immediately preceding state " + }, + { + "bbox": [ + 304, + 208, + 545, + 363 + ], + "type": "inline_equation", + "content": "s_{t-1}" + }, + { + "bbox": [ + 304, + 208, + 545, + 363 + ], + "type": "text", + "content": ". Thus, we define a regularization-free threshold " + }, + { + "bbox": [ + 304, + 208, + 545, + 363 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 304, + 208, + 545, + 363 + ], + "type": "text", + "content": ". Moreover, in HbSR, if a temporal interval between two states is less than the threshold " + }, + { + "bbox": [ + 304, + 208, + 545, + 363 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 304, + 208, + 545, + 363 + ], + "type": "text", + "content": ", we do not apply the correlation regularization among consecutive states. This is because navigation has continuity and neighboring states will inherently have relevance while distant states should be irrelevant. Hence, our HbSR " + }, + { + "bbox": [ + 304, + 208, + 545, + 363 + ], + "type": "inline_equation", + "content": "l^{hsr}" + }, + { + "bbox": [ + 304, + 208, + 545, + 363 + ], + "type": "text", + "content": " is expressed as," + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 336, + 368, + 545, + 402 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 368, + 545, + 402 + ], + "spans": [ + { + "bbox": [ + 336, + 368, + 545, + 402 + ], + "type": "interline_equation", + "content": "l ^ {h s r} = \\mathbb {E} _ {\\boldsymbol {s} \\in \\mathbb {S}} \\left(\\sum_ {t = \\xi + 1} ^ {T} \\sum_ {i = 0} ^ {t - \\xi - 1} \\max \\left(0, \\rho \\left(\\boldsymbol {s} _ {t}, \\boldsymbol {s} _ {i}\\right)\\right)\\right). \\tag {8}", + "image_path": "7477345fb8036fb0cadb18f182a33226903e63fa8ebac5a0922abb226e303b71.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 414, + 405, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 414, + 405, + 426 + ], + "spans": [ + { + "bbox": [ + 306, + 414, + 405, + 426 + ], + "type": "text", + "content": "3.5. Training Details" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 432, + 545, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 432, + 545, + 540 + ], + "spans": [ + { + "bbox": [ + 304, + 432, + 545, + 540 + ], + "type": "text", + "content": "We employ the Asynchronous Advantage Actor-Critic (A3C) Algorithm [30] to train the agent, similar to previous works [13, 14, 49, 53]. We adopt two loss objectives, i.e., policy loss " + }, + { + "bbox": [ + 304, + 432, + 545, + 540 + ], + "type": "inline_equation", + "content": "l_{\\text{policy}}" + }, + { + "bbox": [ + 304, + 432, + 545, + 540 + ], + "type": "text", + "content": " and value loss " + }, + { + "bbox": [ + 304, + 432, + 545, + 540 + ], + "type": "inline_equation", + "content": "l_{\\text{value}}" + }, + { + "bbox": [ + 304, + 432, + 545, + 540 + ], + "type": "text", + "content": ", for training the navigation policy and value estimation networks, respectively. Furthermore, since the calculation of correlation coefficients is differentiable [34], we can employ the loss " + }, + { + "bbox": [ + 304, + 432, + 545, + 540 + ], + "type": "inline_equation", + "content": "l^{hsr}" + }, + { + "bbox": [ + 304, + 432, + 545, + 540 + ], + "type": "text", + "content": " of HbSR in training. Formally, the total loss " + }, + { + "bbox": [ + 304, + 432, + 545, + 540 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 304, + 432, + 545, + 540 + ], + "type": "text", + "content": " is formulated as the summation of " + }, + { + "bbox": [ + 304, + 432, + 545, + 540 + ], + "type": "inline_equation", + "content": "l_{\\text{policy}}" + }, + { + "bbox": [ + 304, + 432, + 545, + 540 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 432, + 545, + 540 + ], + "type": "inline_equation", + "content": "l_{\\text{value}}" + }, + { + "bbox": [ + 304, + 432, + 545, + 540 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 432, + 545, + 540 + ], + "type": "inline_equation", + "content": "l^{hsr}" + }, + { + "bbox": [ + 304, + 432, + 545, + 540 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 354, + 544, + 545, + 559 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 544, + 545, + 559 + ], + "spans": [ + { + "bbox": [ + 354, + 544, + 545, + 559 + ], + "type": "interline_equation", + "content": "l = \\lambda_ {0} l ^ {\\text {p o l i c y}} + \\lambda_ {1} l ^ {\\text {v a l u e}} + \\lambda_ {2} l ^ {\\text {h s r}}, \\tag {9}", + "image_path": "ff647a7501fe9dc4ccc7cfec946496f1541c7f52ddb1383524beb84fbdd00e9d.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 563, + 545, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 563, + 545, + 600 + ], + "spans": [ + { + "bbox": [ + 304, + 563, + 545, + 600 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 563, + 545, + 600 + ], + "type": "inline_equation", + "content": "\\lambda_0, \\lambda_1" + }, + { + "bbox": [ + 304, + 563, + 545, + 600 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 563, + 545, + 600 + ], + "type": "inline_equation", + "content": "\\lambda_2" + }, + { + "bbox": [ + 304, + 563, + 545, + 600 + ], + "type": "text", + "content": " are weighted factors to balance three losses. In our experiments, we set these three weighted factors as 1, 0.5 and 1, respectively." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 609, + 388, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 609, + 388, + 623 + ], + "spans": [ + { + "bbox": [ + 306, + 609, + 388, + 623 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 629, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 665 + ], + "type": "text", + "content": "To validate the effectiveness and efficiency of the proposed HiNL, we conduct extensive experiments on two artificial environments, i.e., iTHOR [26] and RoboTHOR [11]." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 671, + 492, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 671, + 492, + 684 + ], + "spans": [ + { + "bbox": [ + 306, + 671, + 492, + 684 + ], + "type": "text", + "content": "4.1. Protocols and Experimental details" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 689, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 712 + ], + "type": "text", + "content": "Dataset. We adopt iTHOR [26], an artificial environment within the AI2-THOR framework, to evaluate our method." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2567" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 70, + 535, + 254 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 535, + 254 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 535, + 254 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 535, + 254 + ], + "type": "table", + "html": "
MethodiTHOR [26]RoboTHOR [11]
AllL>5AllL>5
SuccessSPLSuccessSPLSuccessSPLSuccessSPL
Random8.0±1.33.6±0.60.3±0.10.1±0.14.0±1.01.6±0.40.2±0.10.1±0.1
WE33.0±3.514.7±1.821.4±3.011.7±1.97.0±1.22.2±0.64.7±0.91.7±0.3
SP [50]35.1±1.315.5±1.122.2±2.711.4±1.610.9±0.54.2±0.57.2±0.63.2±0.5
SAVN [49]40.8±1.216.1±0.528.7±1.513.9±0.510.2±1.03.9±0.96.9±0.82.8±0.6
ORG [13]65.3±0.737.5±0.154.8±1.036.1±0.145.4±0.821.2±1.239.6±0.718.5±0.8
ORG+TPN [13]69.3±1.239.4±1.060.7±1.338.6±1.147.8±1.223.1±0.942.9±0.920.3±1.0
HOZ [53]70.6±1.740.0±1.262.8±1.739.2±0.644.3±1.819.8±1.337.2±1.215.4±0.8
TCN [41]75.1±0.847.7±0.965.3±1.442.1±0.844.4±1.626.4±1.538.4±1.122.0±1.2
VTNet [14]72.2±1.044.9±0.763.4±1.144.0±0.953.2±1.127.5±1.747.0±0.823.3±0.7
HiNL †71.2±0.838.9±0.964.1±0.637.0±0.556.1±1.326.4±0.851.8±0.923.9±0.7
HiNL ‡80.1±1.449.8±1.974.6±1.747.6±1.460.6±1.030.8±1.256.2±0.826.6±0.9
", + "image_path": "d79a42b10661fd52a12b712556f8447ee68f0ecb9f154c987134af0355a0a8db.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 261, + 547, + 297 + ], + "lines": [ + { + "bbox": [ + 46, + 261, + 547, + 297 + ], + "spans": [ + { + "bbox": [ + 46, + 261, + 547, + 297 + ], + "type": "text", + "content": "Table 1. Comparison with existing methods. We report the average Success rate (%) and SPL (%) in iTHOR [26] and RoboTHOR [11] as well as their variances in subscripts by repeating experiments five times. " + }, + { + "bbox": [ + 46, + 261, + 547, + 297 + ], + "type": "inline_equation", + "content": "L > 5" + }, + { + "bbox": [ + 46, + 261, + 547, + 297 + ], + "type": "text", + "content": " represents the episodes that require at least 5 steps. † and ‡ indicate using CNN-based and Transformer-based visual extractors (i.e., ORG [13] and VTNet [14]), respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 306, + 288, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 306, + 288, + 402 + ], + "spans": [ + { + "bbox": [ + 46, + 306, + 288, + 402 + ], + "type": "text", + "content": "The environment includes four types of scenes, i.e., kitchen, living room, bedroom, and bathroom. Each type of scene consists of 30 rooms with different furniture items and placements. Following [13, 14, 53], we use 22 categories as the target classes and ensure that there are at least four targets in each room. We choose the first 20 out of 30 rooms per scene as the training set and equally divide the remaining 40 rooms into the validation and test set." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 402, + 287, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 402, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 46, + 402, + 287, + 545 + ], + "type": "text", + "content": "Furthermore, we conduct experiments in RoboTHOR [11], the other synthetic environment in the AI2-THOR framework. Different from iTHOR, each scene in RoboTHOR is separated by several clapboards. RoboTHOR consists of 89 scenes, while only 75 apartments are published for train/val (60 for training and 15 for validation). Therefore, we select the first 55 of 60 original training scenes as our training dataset and use the rest 5 scenes as the validation set. Then, we adopt the remaining 15 original validation apartments as our test set. We report the test result of the model with the highest success rate on the validation set for both iTHOR and RoboTHOR." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 551, + 288, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 551, + 288, + 674 + ], + "spans": [ + { + "bbox": [ + 46, + 551, + 288, + 674 + ], + "type": "text", + "content": "Evaluation metrics. We adopt two evaluation metrics, i.e., success rate and Success weighted by Path Length (SPL), to assess our method performance. Success rate is introduced to evaluate the navigation effectiveness and calculated by " + }, + { + "bbox": [ + 46, + 551, + 288, + 674 + ], + "type": "inline_equation", + "content": "\\frac{1}{N}\\sum_{n = 0}^{N}S_{n}" + }, + { + "bbox": [ + 46, + 551, + 288, + 674 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 551, + 288, + 674 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 551, + 288, + 674 + ], + "type": "text", + "content": " is the number of episodes and " + }, + { + "bbox": [ + 46, + 551, + 288, + 674 + ], + "type": "inline_equation", + "content": "S_{n}" + }, + { + "bbox": [ + 46, + 551, + 288, + 674 + ], + "type": "text", + "content": " represents a binary success indicator of the " + }, + { + "bbox": [ + 46, + 551, + 288, + 674 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 551, + 288, + 674 + ], + "type": "text", + "content": "-th episode. Meanwhile, SPL measures the efficiency of navigation trajectories. Let " + }, + { + "bbox": [ + 46, + 551, + 288, + 674 + ], + "type": "inline_equation", + "content": "L_{n}" + }, + { + "bbox": [ + 46, + 551, + 288, + 674 + ], + "type": "text", + "content": " be the length of the " + }, + { + "bbox": [ + 46, + 551, + 288, + 674 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 551, + 288, + 674 + ], + "type": "text", + "content": "-th episode and " + }, + { + "bbox": [ + 46, + 551, + 288, + 674 + ], + "type": "inline_equation", + "content": "L_{opt}" + }, + { + "bbox": [ + 46, + 551, + 288, + 674 + ], + "type": "text", + "content": " indicate the length of the optimal path. SPL is formulated as " + }, + { + "bbox": [ + 46, + 551, + 288, + 674 + ], + "type": "inline_equation", + "content": "\\frac{1}{N}\\sum_{n = 0}^{N}S_{n}\\frac{L_{n}}{\\max(L_{n},L_{opt})}" + }, + { + "bbox": [ + 46, + 551, + 288, + 674 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": "Implementation details. Following [14], we train the VT for 20 epochs with the supervision of expert experience. Then, we train the navigation policy for 2M episodes in to-" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 306, + 547, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 306, + 547, + 403 + ], + "spans": [ + { + "bbox": [ + 304, + 306, + 547, + 403 + ], + "type": "text", + "content": "tal with 32 asynchronous agents. In making a robot learn to approach the target, the penalty is set to " + }, + { + "bbox": [ + 304, + 306, + 547, + 403 + ], + "type": "inline_equation", + "content": "1e^{-3}" + }, + { + "bbox": [ + 304, + 306, + 547, + 403 + ], + "type": "text", + "content": " for every step that passes prior to arrival, and thus an agent is encouraged to navigate as quickly as possible. Meanwhile, to encourage the agent complete navigation effectively, we offer a large reward of 5 when the agent finishes a trajectory successfully. We adopt Adam [25] with a learning rate " + }, + { + "bbox": [ + 304, + 306, + 547, + 403 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 304, + 306, + 547, + 403 + ], + "type": "text", + "content": ". Our codes will be publicly related for reproducibility." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 411, + 495, + 425 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 411, + 495, + 425 + ], + "spans": [ + { + "bbox": [ + 305, + 411, + 495, + 425 + ], + "type": "text", + "content": "4.2. Comparison with Existing Methods" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 430, + 426, + 442 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 430, + 426, + 442 + ], + "spans": [ + { + "bbox": [ + 306, + 430, + 426, + 442 + ], + "type": "text", + "content": "4.2.1 Competing Methods" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 450, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 450, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 547, + 715 + ], + "type": "text", + "content": "We compare our method with the following ones: Random policy. An agent makes decisions based on a uniform action probability. Thus, the movements and stops of the agent are random. Scene Prior (SP) [50] utilizes the scene prior knowledge and category associations for navigation with a graph neural network on the FastText database [23]. Word Embedding (WE) makes use of GloVe embedding [35] to signify the target category instead of relying on detection. From trial and error, an agent establishes the connection between object appearances and GloVe embeddings. Self-adaptive Visual Navigation (SAVN) [49] presents a meta reinforcement learning method to assist an agent in adapting to unseen environments during inference. Object Relationship Graph (ORG) [13] proposes a visual representation learning method to encode relationships among categories via a graph and design a tentative policy network (TPN) to escape from deadlocks during testing. Hierarchical Object-to-Zone (HOZ) [53] introduces a graph to guide an agent in a coarse-to-fine manner and a online-learning mechanism to update the graph. Time-Contrastive Networks (TCN) [41] presents a self-supervised method, deriving robotic behaviors and representations from unlabeled" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2568" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 285, + 267 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 285, + 267 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 285, + 267 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 285, + 267 + ], + "type": "table", + "html": "
MethodAllL>5
SuccessSPLSuccessSPL
HaSEHiNL † w/o HaSE68.137.860.335.9
HiNL †71.238.964.137.0
HiNL ‡ w/o HaSE77.945.771.643.5
HiNL ‡80.149.874.647.6
HiNL ‡ w/ tem.78.539.871.239.1
HiNL ‡ w/ lea. tem.79.145.873.143.8
HiNL ‡ w/ rel. tem.77.344.169.441.6
HiNL ‡80.149.874.647.6
HbSRHiNL † w/o HbSR66.838.157.835.8
HiNL †71.238.964.137.0
HiNL ‡ w/o HbSR75.446.967.043.8
HiNL ‡80.149.874.647.6
HiNL ‡ neg. & pos.75.446.967.143.8
HiNL ‡80.149.874.647.6
", + "image_path": "96713046f49535485a148b09d71a2a753335d7eb0f23e82333d0546b96d1daf5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 277, + 286, + 365 + ], + "lines": [ + { + "bbox": [ + 46, + 277, + 286, + 365 + ], + "spans": [ + { + "bbox": [ + 46, + 277, + 286, + 365 + ], + "type": "text", + "content": "Table 2. Ablation study on different components of HiNL in iTHOR [26] environment. † and ‡ indicate using CNN-based and Transformer-based visual extractors (i.e., ORG [13] and VT-Net [14]), respectively. HiNL w/o tem. indicates HiNL without temporal embedding in HaSE, while HiNL w/ lea. tem. and HiNL w/ rel. tem. represent the adoption of learnable and relative temporal embedding in HiNL, respectively. HiNL neg. & pos. stands for regularizing both positive and negative correlations toward 0." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 376, + 286, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 376, + 286, + 448 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 286, + 448 + ], + "type": "text", + "content": "videos captured across multiple viewpoints, leveraging metric learning loss to establish viewpoint-invariant representations that encapsulate end-effector relationships. Visual Transformer for Navigation (VTNet) [14] explores the spatial correlations among objects and observation regions for navigation via a transformer-based network." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 455, + 167, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 455, + 167, + 466 + ], + "spans": [ + { + "bbox": [ + 47, + 455, + 167, + 466 + ], + "type": "text", + "content": "4.2.2 Quantitative Results" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 474, + 286, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 286, + 592 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 286, + 592 + ], + "type": "text", + "content": "To demonstrate the superiority of our proposed learning framework, we select two different visual representation extractors: one is based on graph network (i.e., ORG [13]), and the other is based on the transformer (i.e., VT [14]). Table 1 indicates that both HiNL " + }, + { + "bbox": [ + 46, + 474, + 286, + 592 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 46, + 474, + 286, + 592 + ], + "type": "text", + "content": " and HiNL " + }, + { + "bbox": [ + 46, + 474, + 286, + 592 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 46, + 474, + 286, + 592 + ], + "type": "text", + "content": " achieve superior results compared with their original methods (i.e., ORG and VTNet). This demonstrates that given different visual representations, HiNL can improve navigation performance constantly by explicitly exploring the relationships among historical states." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "content": "To be specific, HiNL " + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "content": " with the transformer-based extractor surpasses the existing methods by a large margin on both success rate " + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "inline_equation", + "content": "(+7.9\\% / + 7.4\\%)" + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "content": " and SPL " + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "inline_equation", + "content": "(+4.9\\% / + 3.3\\%)" + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "content": " in iTHOR [26]/RoboTHOR [11]. Adopting ORG as the visual representation extractor, HiNL " + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "content": " achieves significant improvement over the original ORG by nearly " + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "content": " in success rate in both environments. VTNet and ORG only adopt LSTM to merge well-designed visual representations into navigation states without any explicit constraints on states. The correlations among navigation states are not exploited." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 108 + ], + "type": "text", + "content": "This experiment suggests that our HiNL leads to instructive navigation states, and thus significantly improves the effectiveness and efficiency of our navigation system." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 109, + 545, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 109, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 304, + 109, + 545, + 407 + ], + "type": "text", + "content": "As indicated in Table 1, we observe that HiNL " + }, + { + "bbox": [ + 304, + 109, + 545, + 407 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 304, + 109, + 545, + 407 + ], + "type": "text", + "content": " significantly outperforms ORG+TPN [13]. Although TPN employs an auxiliary network to fine-tune the initial navigation policy network when the agent is stuck in deadlocks during inference, TPN does not guarantee navigation states to be low relevant. Consequently, an agent might be trapped in looping actions, such as moving with rotating, because of the high correlations among navigation states. HiNL solves this issue by reducing the correlations among navigation states, and thus achieves proper actions for navigation. Furthermore, adopting TCN-based loss (i.e., TCN) yields inferior performance with " + }, + { + "bbox": [ + 304, + 109, + 545, + 407 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 304, + 109, + 545, + 407 + ], + "type": "text", + "content": " decrease in success rate compared to HiNL. This is because TCN tries to pull temporally neighboring features closer while consecutive navigation states might be very different due to a rotation action. Thus, TCN may overly emphasize the reliance among neighboring states, thus leading to inferior state estimation. In addition, our method achieves more than two times the success rate and SPL than SAVN in both environments. Benefiting from well-designed visual representations and lowering relevance among navigation states, HiNL achieves better navigation performance in testing even without adjusting the navigation policy. This also implies that our navigation states significantly improve the effectiveness and efficiency of navigation in unseen environments." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 415, + 388, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 415, + 388, + 426 + ], + "spans": [ + { + "bbox": [ + 306, + 415, + 388, + 426 + ], + "type": "text", + "content": "4.2.3 Case Study" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 434, + 545, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 434, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 304, + 434, + 545, + 613 + ], + "type": "text", + "content": "Figure 5a illustrates trajectories of three navigation episodes proceeded by VTNet [14] and HiNL " + }, + { + "bbox": [ + 304, + 434, + 545, + 613 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 304, + 434, + 545, + 613 + ], + "type": "text", + "content": ". In the first case, VTNet fails to find the target and is stuck in the environment, as seen in the first column. On the contrary, HiNL " + }, + { + "bbox": [ + 304, + 434, + 545, + 613 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 304, + 434, + 545, + 613 + ], + "type": "text", + "content": " navigates toward the object successfully. This implies that using HiNL, our agents can react rapidly to the observed changes. In the second case, VTNet moves forward while rotating and thus leads to inefficient navigation. In contrast, the navigation system trained with HiNL avoids rotating and approaches the target with fewer steps. This demonstrates that our HiNL improves navigation efficiency by reducing the correlations among states. In the final case, VTNet moves forward consistently until it reaches the wall. Benefiting from our HaSE, the agent can get rid of dominant historical states and react rapidly to the observed changes." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 623, + 400, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 623, + 400, + 635 + ], + "spans": [ + { + "bbox": [ + 306, + 623, + 400, + 635 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 642, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 712 + ], + "type": "text", + "content": "Impacts of different components. To illustrate the impacts of different components, we conduct experiments on adopting one of the proposed components (i.e., HaSE and HbSR). Furthermore, we report the navigation performance with different visual representation extractors, i.e., ORG and VT. As indicated in Table 2, both HiNL " + }, + { + "bbox": [ + 304, + 642, + 545, + 712 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 304, + 642, + 545, + 712 + ], + "type": "text", + "content": " and" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2569" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 68, + 367, + 220 + ], + "blocks": [ + { + "bbox": [ + 52, + 68, + 367, + 220 + ], + "lines": [ + { + "bbox": [ + 52, + 68, + 367, + 220 + ], + "spans": [ + { + "bbox": [ + 52, + 68, + 367, + 220 + ], + "type": "image", + "image_path": "cdeace07492ab6745fbb0a8855c160ce4cad1f47817895e77376c8826c92988b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 231, + 547, + 285 + ], + "lines": [ + { + "bbox": [ + 46, + 231, + 547, + 285 + ], + "spans": [ + { + "bbox": [ + 46, + 231, + 547, + 285 + ], + "type": "text", + "content": "Figure 5. (a) Visual results in the simulated environments iTHOR. The arrow color changes (i.e., from blue to red) represent the navigation progress (i.e., from beginning to end). Black arrows indicate rotations. The target objects are highlighted by the green/red boxes, where a green box stands for a success episode and a red box represents a failure case. (b) Impacts of different historical state length " + }, + { + "bbox": [ + 46, + 231, + 547, + 285 + ], + "type": "inline_equation", + "content": "\\tau \\in [2,10]" + }, + { + "bbox": [ + 46, + 231, + 547, + 285 + ], + "type": "text", + "content": ". (c) Results of different regularization-free states " + }, + { + "bbox": [ + 46, + 231, + 547, + 285 + ], + "type": "inline_equation", + "content": "\\xi \\in [1,8]" + }, + { + "bbox": [ + 46, + 231, + 547, + 285 + ], + "type": "text", + "content": ". We set " + }, + { + "bbox": [ + 46, + 231, + 547, + 285 + ], + "type": "inline_equation", + "content": "\\tau = 5" + }, + { + "bbox": [ + 46, + 231, + 547, + 285 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 231, + 547, + 285 + ], + "type": "inline_equation", + "content": "\\xi = 1" + }, + { + "bbox": [ + 46, + 231, + 547, + 285 + ], + "type": "text", + "content": " and highlight them by the black vertical dotted lines in (b) and (c)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 388, + 59, + 533, + 220 + ], + "blocks": [ + { + "bbox": [ + 388, + 59, + 533, + 220 + ], + "lines": [ + { + "bbox": [ + 388, + 59, + 533, + 220 + ], + "spans": [ + { + "bbox": [ + 388, + 59, + 533, + 220 + ], + "type": "image", + "image_path": "432a840037cb021724dc457044c1bad233f0286c843ae753da6e3c10bb90da0c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 292, + 287, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 292, + 287, + 365 + ], + "spans": [ + { + "bbox": [ + 46, + 292, + 287, + 365 + ], + "type": "text", + "content": "HiNL " + }, + { + "bbox": [ + 46, + 292, + 287, + 365 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 46, + 292, + 287, + 365 + ], + "type": "text", + "content": " achieve gains in navigation performance on success rate " + }, + { + "bbox": [ + 46, + 292, + 287, + 365 + ], + "type": "inline_equation", + "content": "(+2.9\\% / + 2.2\\%)" + }, + { + "bbox": [ + 46, + 292, + 287, + 365 + ], + "type": "text", + "content": " and SPL " + }, + { + "bbox": [ + 46, + 292, + 287, + 365 + ], + "type": "inline_equation", + "content": "(+0.9\\% / + 4.1\\%)" + }, + { + "bbox": [ + 46, + 292, + 287, + 365 + ], + "type": "text", + "content": " by introducing HaSE. Furthermore, HiNL " + }, + { + "bbox": [ + 46, + 292, + 287, + 365 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 46, + 292, + 287, + 365 + ], + "type": "text", + "content": " and HiNL " + }, + { + "bbox": [ + 46, + 292, + 287, + 365 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 46, + 292, + 287, + 365 + ], + "type": "text", + "content": " outperform models without HbSR (i.e., HiNL " + }, + { + "bbox": [ + 46, + 292, + 287, + 365 + ], + "type": "inline_equation", + "content": "\\dagger/\\mathrm{HiNL}\\ddagger" + }, + { + "bbox": [ + 46, + 292, + 287, + 365 + ], + "type": "text", + "content": " w/o HbSR), respectively. This manifests that both of our proposed components improve the performance of navigation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 369, + 287, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 369, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 46, + 369, + 287, + 514 + ], + "type": "text", + "content": "Impacts of different temporal embedding in HaSE. Table 2 compares different temporal embedding methods, (i.e., w/o temporal embedding, w/ relative temporal embedding and w/ absolute temporal embedding (HiNL)) and highlights the effectiveness of incorporating temporal embedding in historical states. HiNL with absolute temporal embedding significantly improves success rates and SPL on iTHOR compared to no embedding (HiNL " + }, + { + "bbox": [ + 46, + 369, + 287, + 514 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 46, + 369, + 287, + 514 + ], + "type": "text", + "content": " w/o tem.), learnable embedding (HiNL " + }, + { + "bbox": [ + 46, + 369, + 287, + 514 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 46, + 369, + 287, + 514 + ], + "type": "text", + "content": " w/ lea. tem.), and relative embedding (HiNL " + }, + { + "bbox": [ + 46, + 369, + 287, + 514 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 46, + 369, + 287, + 514 + ], + "type": "text", + "content": " w/ rel. tem.). This improvement, especially in navigation efficiency, stems from providing the agent with the current episode length." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 517, + 287, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 517, + 287, + 625 + ], + "spans": [ + { + "bbox": [ + 46, + 517, + 287, + 625 + ], + "type": "text", + "content": "Comparison of different " + }, + { + "bbox": [ + 46, + 517, + 287, + 625 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 517, + 287, + 625 + ], + "type": "text", + "content": " in HaSE. To illustrate the influence of different historical state lengths, we conduct experiments with " + }, + { + "bbox": [ + 46, + 517, + 287, + 625 + ], + "type": "inline_equation", + "content": "\\tau \\in [2,10]" + }, + { + "bbox": [ + 46, + 517, + 287, + 625 + ], + "type": "text", + "content": ". As demonstrated in Figure 5b, we observe that as historical states become too many, our agent may fail to converge to an optimal policy. On the other hand, estimating navigation states with few historical states do not have sufficient capacity to effectively update states. The agent reaches the peak of success rate and SPL when it adopts 5 historical states in HaSE." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 630, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 287, + 715 + ], + "type": "text", + "content": "Impacts of different " + }, + { + "bbox": [ + 46, + 630, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 46, + 630, + 287, + 715 + ], + "type": "text", + "content": " in HbSR. To study the impacts of adopting different regularization-free thresholds " + }, + { + "bbox": [ + 46, + 630, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 46, + 630, + 287, + 715 + ], + "type": "text", + "content": ", we conduct experiments with " + }, + { + "bbox": [ + 46, + 630, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\xi \\in [1,8]" + }, + { + "bbox": [ + 46, + 630, + 287, + 715 + ], + "type": "text", + "content": " in iTHOR. As seen in Figure 5c, increasing " + }, + { + "bbox": [ + 46, + 630, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 46, + 630, + 287, + 715 + ], + "type": "text", + "content": " drops performance clearly. Both success rate (80% to 76%) and SPL (49.8% to 45.8%) decrease when " + }, + { + "bbox": [ + 46, + 630, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 46, + 630, + 287, + 715 + ], + "type": "text", + "content": " increases from 1 to 8. Since two navigation states generally have high correlations when they are close" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 292, + 545, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 292, + 545, + 328 + ], + "spans": [ + { + "bbox": [ + 305, + 292, + 545, + 328 + ], + "type": "text", + "content": "in temporal, raising " + }, + { + "bbox": [ + 305, + 292, + 545, + 328 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 305, + 292, + 545, + 328 + ], + "type": "text", + "content": " indicates ignoring constraints on these correlations. As a result, the effects of HbSR are compromised, and the navigation performance is dropped." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 330, + 546, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 330, + 546, + 450 + ], + "spans": [ + { + "bbox": [ + 305, + 330, + 546, + 450 + ], + "type": "text", + "content": "Comparison of regularization on different values. To analyze the impacts of regularization on different values, we demonstrate the comparisons of applying HbSR on (i) both negative and positive correlations and (ii) only positive correlations, as indicated in Table 2. Compared to HiNL trained on regularizing the correlations based on both positive and negative values (HiNL " + }, + { + "bbox": [ + 305, + 330, + 546, + 450 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 305, + 330, + 546, + 450 + ], + "type": "text", + "content": " neg. & pos.), HiNL improves performance by neglecting the negative correlations. This comparison suggests that inefficient navigation is generally caused by positively correlated navigation states." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 460, + 378, + 472 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 460, + 378, + 472 + ], + "spans": [ + { + "bbox": [ + 306, + 460, + 378, + 472 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 480, + 545, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 480, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 304, + 480, + 545, + 624 + ], + "type": "text", + "content": "In this paper, we proposed an effective navigation state learning approach for visual navigation, named History-inspired Navigation Policy Learning (HiNL). Our HiNL involves two innovate parts: a history-aware state estimation module (HaSE) and a history-based state regularization (HbSR). HbSR eliminates the negative impacts of dominant historical states and enables current states to reflect newly observed visual information. HbSR significantly reduces the correlations among navigation states, thus leading to more informative states as well as effective actions. Extensive results demonstrate that our HiNL significantly improves navigation performance." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 634, + 404, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 634, + 404, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 634, + 404, + 647 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 654, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 654, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 654, + 545, + 714 + ], + "type": "text", + "content": "This research is funded in part by ARC-Discovery grant (DP220100800 to XY), ARC-DECRA grant (DE230100477 to XY) and ARC Centres of Excellence grant (CE200100025). We thank all anonymous reviewers and ACs for their constructive suggestions." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2570" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Ziad Al-Halah, Santhosh Kumar Ramakrishnan, and Kristen Grauman. Zero experience required: Plug & play modular transfer learning for semantic visual navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17031-17041, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 146, + 287, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 146, + 287, + 179 + ], + "spans": [ + { + "bbox": [ + 53, + 146, + 287, + 179 + ], + "type": "text", + "content": "[2] Bastian Alt, Adrian Šošić, and Heinz Koeppl. Correlation priors for reinforcement learning. Advances in Neural Information Processing Systems, 32, 2019. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 180, + 287, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 180, + 287, + 213 + ], + "spans": [ + { + "bbox": [ + 53, + 180, + 287, + 213 + ], + "type": "text", + "content": "[3] Johann Borenstein and Yoram Koren. Real-time obstacle avoidance for fast mobile robots. IEEE Transactions on systems, Man, and Cybernetics, 19(5):1179-1187, 1989. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 214, + 287, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 214, + 287, + 257 + ], + "spans": [ + { + "bbox": [ + 53, + 214, + 287, + 257 + ], + "type": "text", + "content": "[4] Johann Borenstein and Yoram Koren. The vector field histogram-fast obstacle avoidance for mobile robots. IEEE transactions on robotics and automation, 7(3):278-288, 1991. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 258, + 287, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 258, + 287, + 313 + ], + "spans": [ + { + "bbox": [ + 53, + 258, + 287, + 313 + ], + "type": "text", + "content": "[5] Tommaso Campari, Leonardo Lamanna, Paolo Traverso, Luciano Serafini, and Lamberto Ballan. Online learning of reusable abstract models for object goal navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14870-14879, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 315, + 287, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 315, + 287, + 358 + ], + "spans": [ + { + "bbox": [ + 53, + 315, + 287, + 358 + ], + "type": "text", + "content": "[6] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 359, + 287, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 359, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 53, + 359, + 287, + 403 + ], + "type": "text", + "content": "[7] Changan Chen, Ziad Al-Halah, and Kristen Grauman. Semantic audio-visual navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15516–15525, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 403, + 287, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 403, + 287, + 468 + ], + "spans": [ + { + "bbox": [ + 53, + 403, + 287, + 468 + ], + "type": "text", + "content": "[8] Changan Chen, Unnat Jain, Carl Schissler, Sebastia Vicenc Amengual Gari, Ziad Al-Halah, Vamsi Krishna Ithapu, Philip Robinson, and Kristen Grauman. Soundspaces: Audio-visual navigation in 3d environments. In European Conference on Computer Vision, pages 17–36. Springer, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 469, + 287, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 469, + 287, + 523 + ], + "spans": [ + { + "bbox": [ + 53, + 469, + 287, + 523 + ], + "type": "text", + "content": "[9] Jinyu Chen, Chen Gao, Erli Meng, Qiong Zhang, and Si Liu. Reinforced structured state-evolution for vision-language navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15450-15459, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 525, + 287, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 287, + 579 + ], + "type": "text", + "content": "[10] Kevin Chen, Juan Pablo de Vicente, Gabriel Sepulveda, Fei Xia, Alvaro Soto, Marynel Vázquez, and Silvio Savarese. A behavioral approach to visual navigation with graph localization networks. arXiv preprint arXiv:1903.00445, 2019. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 581, + 287, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 581, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 581, + 287, + 657 + ], + "type": "text", + "content": "[11] Matt Deitke, Winson Han, Alvaro Herrasti, Aniruddha Kembhavi, Eric Kolve, Roozbeh Mottaghi, Jordi Salvador, Dustin Schwenk, Eli VanderBilt, Matthew Wallingford, et al. Robothor: An open simulation-to-real embodied ai platform. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3164-3174, 2020. 2, 5, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[12] MWM Gamini Dissanayake, Paul Newman, Steve Clark, Hugh F Durrant-Whyte, and Michael Csorba. A solution to the simultaneous localization and map building (slam) problem. IEEE Transactions on robotics and automation, 17(3):229-241, 2001. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[13] Heming Du, Xin Yu, and Liang Zheng. Learning object relation graph and tentative policy for visual navigation. arXiv preprint arXiv:2007.11018, 2020. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 108, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 545, + 140 + ], + "type": "text", + "content": "[14] Heming Du, Xin Yu, and Liang Zheng. Vtnet: Visual transformer network for object goal navigation. arXiv preprint arXiv:2105.09447, 2021. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 142, + 545, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 142, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 307, + 142, + 545, + 196 + ], + "type": "text", + "content": "[15] Kuan Fang, Alexander Toshev, Li Fei-Fei, and Silvio Savarese. Scene memory transformer for embodied agents in long-horizon tasks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 538–547, 2019. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 198, + 545, + 242 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 198, + 545, + 242 + ], + "spans": [ + { + "bbox": [ + 307, + 198, + 545, + 242 + ], + "type": "text", + "content": "[16] Xiaoyu Feng, Heming Du, Yueqi Duan, Yongpan Liu, and Hehe Fan. Sefformer: Structure embedding transformer for 3d object detection. arXiv preprint arXiv:2209.01745, 2022.2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 244, + 545, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 244, + 545, + 278 + ], + "spans": [ + { + "bbox": [ + 307, + 244, + 545, + 278 + ], + "type": "text", + "content": "[17] David Freedman, Robert Pisani, and Roger Purves. Statistics (international student edition). Pisani, R. Purves, 4th edn. WW Norton & Company, New York, 2007. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 280, + 545, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 280, + 545, + 334 + ], + "spans": [ + { + "bbox": [ + 307, + 280, + 545, + 334 + ], + "type": "text", + "content": "[18] Chuang Gan, Yiwei Zhang, Jiajun Wu, Boqing Gong, and Joshua B Tenenbaum. Look, listen, and act: Towards audio-visual embodied navigation. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 9701–9707. IEEE, 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 335, + 545, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 335, + 545, + 380 + ], + "spans": [ + { + "bbox": [ + 307, + 335, + 545, + 380 + ], + "type": "text", + "content": "[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 382, + 545, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 382, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 307, + 382, + 545, + 415 + ], + "type": "text", + "content": "[20] Todd Hester and Peter Stone. Texlore: real-time sample-efficient reinforcement learning for robots. Machine learning, 90(3):385-429, 2013. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 417, + 545, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 417, + 545, + 438 + ], + "spans": [ + { + "bbox": [ + 307, + 417, + 545, + 438 + ], + "type": "text", + "content": "[21] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation, 9(8):1735-1780, 1997. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 440, + 545, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 440, + 545, + 495 + ], + "spans": [ + { + "bbox": [ + 307, + 440, + 545, + 495 + ], + "type": "text", + "content": "[22] Meihuizi Jia, Lei Shen, Xin Shen, Lejian Liao, Meng Chen, Xiaodong He, Zhendong Chen, and Jiaqi Li. Mnerqq: An end-to-end mrc framework for multimodal named entity recognition with query grounding. arXiv preprint arXiv:2211.14739, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 497, + 545, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 497, + 545, + 530 + ], + "spans": [ + { + "bbox": [ + 307, + 497, + 545, + 530 + ], + "type": "text", + "content": "[23] Armand Joulin, Edouard Grave, Piotr Bojanowski, and Tomas Mikolov. Bag of tricks for efficient text classification. arXiv preprint arXiv:1607.01759, 2016. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 532, + 545, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 532, + 545, + 586 + ], + "spans": [ + { + "bbox": [ + 307, + 532, + 545, + 586 + ], + "type": "text", + "content": "[24] Apoorv Khandelwal, Luca Weihs, Roozbeh Mottaghi, and Aniruddha Kembhavi. Simple but effective: Clip embeddings for embodied ai. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14829-14838, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 588, + 545, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 588, + 545, + 621 + ], + "spans": [ + { + "bbox": [ + 307, + 588, + 545, + 621 + ], + "type": "text", + "content": "[25] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 624, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 667 + ], + "type": "text", + "content": "[26] Eric Kolve, Roozbeh Mottaghi, Winson Han, Eli VanderBilt, Luca Weihs, Alvaro Herrasti, Daniel Gordon, Yuke Zhu, Abhinav Gupta, and Ali Farhadi. AI2-THOR: An Interactive 3D Environment for Visual AI. arXiv, 2017. 2, 5, 6, 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 545, + 712 + ], + "type": "text", + "content": "[27] Bar Mayo, Tamir Hazan, and Ayellet Tal. Visual navigation with spatial attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16898-16907, 2021. 2" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "2571" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 138 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 138 + ], + "type": "text", + "content": "[28] Lina Mezghan, Sainbayar Sukhbaatar, Thibaut Lavril, Oleksandr Maksymets, Dhruv Batra, Piotr Bojanowski, and Karteek Alahari. Memory-augmented reinforcement learning for image-goal navigation. In 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 3316-3323. IEEE, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 140, + 288, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 140, + 288, + 195 + ], + "spans": [ + { + "bbox": [ + 48, + 140, + 288, + 195 + ], + "type": "text", + "content": "[29] Piotr Mirowski, Razvan Pascanu, Fabio Viola, Hubert Soyer, Andrew J Ballard, Andrea Banino, Misha Denil, Ross Goroshin, Laurent Sifre, Koray Kavukcuoglu, et al. Learning to navigate in complex environments. arXiv preprint arXiv:1611.03673, 2016. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 197, + 288, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 197, + 288, + 252 + ], + "spans": [ + { + "bbox": [ + 48, + 197, + 288, + 252 + ], + "type": "text", + "content": "[30] Volodymyr Mnih, Adria Puigdomenech Badia, Mehdi Mirza, Alex Graves, Timothy Lillicrap, Tim Harley, David Silver, and Koray Kavukcuoglu. Asynchronous methods for deep reinforcement learning. In International conference on machine learning, pages 1928-1937, 2016. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 253, + 288, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 253, + 288, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 253, + 288, + 308 + ], + "type": "text", + "content": "[31] Arsalan Mousavian, Alexander Toshev, Marek Fiser, Jana Košecka, Ayzaan Wahid, and James Davidson. Visual representations for semantic target driven navigation. In 2019 International Conference on Robotics and Automation (ICRA), pages 8846-8852. IEEE, 2019. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 308, + 288, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 308, + 288, + 363 + ], + "spans": [ + { + "bbox": [ + 48, + 308, + 288, + 363 + ], + "type": "text", + "content": "[32] Giuseppe Oriolo, Marilena Vendittelli, and Giovanni Ulivi. On-line map building and navigation for autonomous mobile robots. In Proceedings of 1995 IEEE International Conference on Robotics and Automation, volume 3, pages 2900-2906. IEEE, 1995. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 365, + 288, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 365, + 288, + 408 + ], + "spans": [ + { + "bbox": [ + 48, + 365, + 288, + 408 + ], + "type": "text", + "content": "[33] Emilio Parisotto and Ruslan Salakhutdinov. Neural map: Structured memory for deep reinforcement learning. In International Conference on Learning Representations, 2018. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 410, + 288, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 410, + 288, + 454 + ], + "spans": [ + { + "bbox": [ + 48, + 410, + 288, + 454 + ], + "type": "text", + "content": "[34] Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. Automatic differentiation in pytorch. 2017. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 456, + 288, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 288, + 510 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 288, + 510 + ], + "type": "text", + "content": "[35] Jeffrey Pennington, Richard Socher, and Christopher D Manning. Glove: Global vectors for word representation. In Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP), pages 1532-1543, 2014. 2, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 512, + 288, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 512, + 288, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 512, + 288, + 578 + ], + "type": "text", + "content": "[36] Santhosh Kumar Ramakrishnan, Devendra Singh Chaplot, Ziad Al-Halah, Jitendra Malik, and Kristen Grauman. Poni: Potential functions for objectgoal navigation with interaction-free learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18890-18900, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 579, + 288, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 579, + 288, + 634 + ], + "spans": [ + { + "bbox": [ + 48, + 579, + 288, + 634 + ], + "type": "text", + "content": "[37] Pravesh Ranchod, Benjamin Rosman, and George Konidaris. Nonparametric bayesian reward segmentation for skill discovery using inverse reinforcement learning. In 2015 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 471-477. IEEE, 2015. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 635, + 288, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 288, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 288, + 679 + ], + "type": "text", + "content": "[38] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 681, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 288, + 714 + ], + "type": "text", + "content": "[39] Nikolay Savinov, Alexey Dosovitskiy, and Vladlen Koltun. Semi-parametric topological memory for navigation. arXiv preprint arXiv:1803.00653, 2018. 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 546, + 714 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 307, + 73, + 546, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 546, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 546, + 127 + ], + "type": "text", + "content": "[40] Gabriel Sepulveda, Juan Carlos Niebles, and Alvaro Soto. A deep learning based behavioral approach to indoor autonomous navigation. In 2018 IEEE International Conference on Robotics and Automation (ICRA), pages 4646-4653. IEEE, 2018. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 129, + 546, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 546, + 194 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 546, + 194 + ], + "type": "text", + "content": "[41] Pierre Sermanet, Corey Lynch, Yevgen Chebotar, Jasmine Hsu, Eric Jang, Stefan Schaal, Sergey Levine, and Google Brain. Time-contrastive networks: Self-supervised learning from video. In 2018 IEEE international conference on robotics and automation (ICRA), pages 1134–1141. IEEE, 2018. 3, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 195, + 546, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 195, + 546, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 195, + 546, + 239 + ], + "type": "text", + "content": "[42] Adrian Šošić, Abdelhak M Zoubir, and Heinz Koeppl. A bayesian approach to policy recognition and state representation learning. IEEE transactions on pattern analysis and machine intelligence, 40(6):1295-1308, 2017. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 239, + 546, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 239, + 546, + 282 + ], + "spans": [ + { + "bbox": [ + 307, + 239, + 546, + 282 + ], + "type": "text", + "content": "[43] Adrian Šošic, Abdelhak M Zoubir, Elmar Rueckert, Jan Peters, and Heinz Koeppl. Inverse reinforcement learning via nonparametric spatio-temporal subgoal modeling. Journal of Machine Learning Research, 19(69):1-45, 2018. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 283, + 546, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 283, + 546, + 316 + ], + "spans": [ + { + "bbox": [ + 307, + 283, + 546, + 316 + ], + "type": "text", + "content": "[44] Tianqi Tang, Heming Du, Xin Yu, and Yi Yang. Monocular camera-based point-goal navigation by learning depth channel and cross-modality pyramid fusion. 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 316, + 546, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 316, + 546, + 371 + ], + "spans": [ + { + "bbox": [ + 307, + 316, + 546, + 371 + ], + "type": "text", + "content": "[45] Yurun Tian, Xin Yu, Bin Fan, Fuchao Wu, Huub Heijnen, and Vassileios Balntas. Sosnet: Second order similarity regularization for local descriptor learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11016-11025, 2019. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 372, + 546, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 372, + 546, + 415 + ], + "spans": [ + { + "bbox": [ + 307, + 372, + 546, + 415 + ], + "type": "text", + "content": "[46] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 416, + 546, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 416, + 546, + 491 + ], + "spans": [ + { + "bbox": [ + 307, + 416, + 546, + 491 + ], + "type": "text", + "content": "[47] Su Wang, Ceslee Montgomery, Jordi Orbay, Vighnesh Birodkar, Aleksandra Faust, Izzeddin Gur, Natasha Jaques, Austin Waters, Jason Baldridge, and Peter Anderson. Less is more: Generating grounded navigation instructions from landmarks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15428-15438, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 493, + 546, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 493, + 546, + 536 + ], + "spans": [ + { + "bbox": [ + 307, + 493, + 546, + 536 + ], + "type": "text", + "content": "[48] Erik Wijmans, Abhishek Kadian, Ari Morcos, Stefan Lee, Irfan Essa, Devi Parikh, Manolis Savva, and Dhruv Batra. Dd-ppo: Learning near-perfect pointgoal navigators from 2.5 billion frames. arXiv preprint arXiv:1911.00357, 2019. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 537, + 546, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 537, + 546, + 602 + ], + "spans": [ + { + "bbox": [ + 307, + 537, + 546, + 602 + ], + "type": "text", + "content": "[49] Mitchell Wortsman, Kiana Ehsani, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Learning to learn how to learn: Self-adaptive visual navigation using meta-learning. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6750-6759, 2019. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 603, + 546, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 603, + 546, + 635 + ], + "spans": [ + { + "bbox": [ + 307, + 603, + 546, + 635 + ], + "type": "text", + "content": "[50] Wei Yang, Xiaolong Wang, Ali Farhadi, Abhinav Gupta, and Roozbeh Mottaghi. Visual semantic navigation using scene priors. arXiv preprint arXiv:1810.06543, 2018. 1, 2, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 636, + 546, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 636, + 546, + 691 + ], + "spans": [ + { + "bbox": [ + 307, + 636, + 546, + 691 + ], + "type": "text", + "content": "[51] Xin Yu and Fatih Porikli. Ultra-resolving face images by discriminative generative networks. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part V, pages 318-333. Springer, 2016. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 692, + 546, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 692, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 692, + 546, + 714 + ], + "type": "text", + "content": "[52] Amir R Zamir, Alexander Sax, William Shen, Leonidas J Guibas, Jitendra Malik, and Silvio Savarese. Taskonomy:" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2572" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 217 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 106 + ], + "type": "text", + "content": "Disentangling task transfer learning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3712-3722, 2018. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 162 + ], + "type": "text", + "content": "[53] Sixian Zhang, Xinhang Song, Yubing Bai, Weijie Li, Yakui Chu, and Shuqiang Jiang. Hierarchical object-to-zone graph for object navigation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15130-15140, 2021. 2, 3, 5, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 287, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 287, + 217 + ], + "type": "text", + "content": "[54] Yuke Zhu, Roozbeh Mottaghi, Eric Kolve, Joseph J Lim, Abhinav Gupta, Li Fei-Fei, and Ali Farhadi. Target-driven visual navigation in indoor scenes using deep reinforcement learning. In 2017 IEEE international conference on robotics and automation (ICRA), pages 3357-3364. IEEE, 2017. 2" + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "2573" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/ObjectMatch_ Robust Registration Using Canonical Object Correspondences/9d0104e4-40f2-4a9d-bf9d-8065e49cc737_content_list.json b/2023/ObjectMatch_ Robust Registration Using Canonical Object Correspondences/9d0104e4-40f2-4a9d-bf9d-8065e49cc737_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9246c804650e18f40668e98443b7603d16c0a536 --- /dev/null +++ b/2023/ObjectMatch_ Robust Registration Using Canonical Object Correspondences/9d0104e4-40f2-4a9d-bf9d-8065e49cc737_content_list.json @@ -0,0 +1,1828 @@ +[ + { + "type": "text", + "text": "ObjectMatch: Robust Registration using Canonical Object Correspondences", + "text_level": 1, + "bbox": [ + 98, + 130, + 870, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Can Gumeli", + "bbox": [ + 238, + 180, + 339, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Angela Dai", + "bbox": [ + 416, + 181, + 511, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Matthias Nießner", + "bbox": [ + 589, + 181, + 730, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Technical University of Munich", + "text_level": 1, + "bbox": [ + 354, + 215, + 609, + 233 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d3f06c5d45558df7016439e082690d03636262a3120456f1cc122b9de7dc3222.jpg", + "image_caption": [ + "Low-overlap Frames with Matching Objects" + ], + "image_footnote": [], + "bbox": [ + 96, + 287, + 274, + 391 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2c3cef79c96c5e79dbb46a94649ef866e5264b69ef2f493c581ba90d9950a3c4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 277, + 287, + 454, + 390 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4357d973bceeaaad23d31aa645eee790d5fe4d7bf49d5a225e95cfd010db473f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 454, + 332, + 475, + 347 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/17c4fb6bd2b65901cfbd33acbe69dcaae383bb72f40ea24e4f9fe3fe8e95608f.jpg", + "image_caption": [ + "Canonical Correspondences" + ], + "image_footnote": [], + "bbox": [ + 475, + 287, + 705, + 390 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fba06f9104b57c26157b6040acc7258a375f26936281e12abbee5750171b2a18.jpg", + "image_caption": [ + "Camera and Object Poses", + "Figure 1. Modern camera pose estimation relies on feature matching between overlapping frames - in this work, we present ObjectMatch to find correspondences between frames with little or no overlap by predicting semantic mappings through canonical object correspondences. The images above share no direct overlap, yet our method establishes indirect correspondences, thus enabling a successful registration." + ], + "image_footnote": [], + "bbox": [ + 728, + 284, + 874, + 392 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 493, + 312, + 508 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We present ObjectMatch1, a semantic and object-centric camera pose estimator for RGB-D SLAM pipelines. Modern camera pose estimators rely on direct correspondences of overlapping regions between frames; however, they cannot align camera frames with little or no overlap. In this work, we propose to leverage indirect correspondences obtained via semantic object identification. For instance, when an object is seen from the front in one frame and from the back in another frame, we can provide additional pose constraints through canonical object correspondences. We first propose a neural network to predict such correspondences on a per-pixel level, which we then combine in our energy formulation with state-of-the-art keypoint matching solved with a joint Gauss-Newton optimization. In a pairwise setting, our method improves registration recall of state-of-the-art feature matching, including from $24\\%$ to $45\\%$ in pairs with $10\\%$ or less inter-frame overlap. In registering RGB-D sequences, our method outperforms cutting-edge SLAM baselines in challenging, low-frame-rate scenarios, achieving more than $35\\%$ reduction in trajectory error in multiple scenes.", + "bbox": [ + 75, + 523, + 473, + 843 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 501, + 455, + 630, + 472 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "RGB-D registration and 3D SLAM has been a fundamental task in computer vision, with significant study and enabling many applications in mixed reality, robotics, and content creation. Central to both state-of-the-art traditional and learning-based camera pose estimation is establishing correspondences between points in input frames. However, correspondence estimation remains quite challenging when there is little or no overlap between frames.", + "bbox": [ + 496, + 484, + 890, + 606 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In contrast, humans can easily localize across these challenging scenarios by leveraging additional semantic knowledge – in particular, by further localizing at the level of objects and identifying matching objects between views. For instance, when observing a chair from the back and the side (e.g., in Figure 1), view overlap is minimal (or even no view overlap), resulting in failed registration from keypoint matching. However, the semantic knowledge of the chair and its object pose nonetheless enables humans to estimate the poses from which the front and side views were taken. Thus, we propose to take a new perspective on camera pose estimation and imbue camera registration with awareness of such semantic correspondences between objects for robust performance in these challenging scenarios.", + "bbox": [ + 496, + 609, + 892, + 821 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To this end, we propose ObjectMatch, a new paradigm for camera pose estimation leveraging canonical object correspondences in tandem with local keypoint correspondences between views. This enables significantly more robust registration under a variety of challenging scenar-", + "bbox": [ + 496, + 824, + 890, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1https://cangumeli.github.io/ObjectMatch/", + "bbox": [ + 94, + 886, + 418, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "13082", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ios, including low view overlap. For a sequence of input frames, ObjectMatch learns to semantically identify objects across frames, enabling a compact, global parameterization of 9-DoF object poses. Object correspondences are established through predicting normalized object coordinates [43], dense correspondences from object pixels to a canonically oriented space for each object. We then formulate a joint camera and object pose optimization that constrains object correspondences indirectly, operating irrespective of the shared visibility of image regions. Our approach is complementary to state-of-the-art SLAM methods, and we leverage our energy formulation to complement state-of-the-art keypoint matching [10, 11, 36] in a joint Gauss-Newton optimization.", + "bbox": [ + 75, + 90, + 472, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our method outperforms strong baselines in both pairwise registration and registration of RGB-D frame sequences. In pairwise registration of challenging ScanNet [9] image pairs, we improve pose recall from $24\\%$ to $45\\%$ when the overlap is below $10\\%$ . On sequence registration of room-scale RGB-D scenes, our method outperforms various strong baselines in difficult, low-frame-rate settings in several TUM-RGBD [39] and ScanNet [9] scenes, reducing the trajectory error by more than $35\\%$ in multiple challenging scenes.", + "bbox": [ + 75, + 301, + 472, + 453 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To sum up, our main contributions include:", + "bbox": [ + 96, + 454, + 383, + 468 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- An object-centric camera pose estimator that can handle low-overlap frame sets via indirect, canonical object correspondences established with predicted dense, per-pixel normalized object coordinates.", + "- A joint energy formulation that leverages semantic object identification and dense, normalized object coordinates corresponding to canonical object geometries.", + "- Our semantic grounding of object correspondences enables significantly more robust registration in low-overlap and low-frame-rate cases. ObjectMatch improves over state of the art from $24\\%$ to $45\\%$ registration recall of $\\leq 10\\%$ overlap frame pairs and achieves over $35\\%$ trajectory error reduction in several challenging sequences." + ], + "bbox": [ + 94, + 472, + 468, + 694 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 709, + 218, + 724 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "RGB-D Registration and SLAM. In recent years, there have been many advances in indoor RGB-D reconstruction. Earlier RGB-D fusion approaches focus on frame-to-model camera tracking [20, 29]. To handle the loop closures better, more recent SLAM systems introduce explicit strategies or global optimization methods for handling loop closures through global optimization [5,10,28,37,44] to fix tracking errors. More recently, deep learning techniques have been applied to registration and SLAM scenarios, with methods ranging from geometric point cloud registration [6, 19, 31] as well as neural field based SLAM techniques [18, 40, 48].", + "bbox": [ + 75, + 734, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite all the successes in RGB-D SLAM and registration, the task is still challenging since incomplete loop closures observed via low-overlap frames cannot be handled, and most SLAM methods require a very high overlap between consecutive frames to track cameras accurately.", + "bbox": [ + 496, + 90, + 890, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Feature Matching. Modern RGB(-D) camera pose estimators rely on a feature-matching backbone. Classical global registration techniques [5,46] use FPFH features [34] over point cloud fragments. On the other hand, many global RGB-D SLAM techniques rely on sparse color features [10, 28]. While being successful in many scenarios, conventional feature matching often fails when the interframe overlap is low. Therefore, deep learning techniques have been utilized for predicting overlapping regions based on geometry or color. On the geometric side, Deep Global Registration [6] predicts overlapping point features using nearest neighbor search over learned geometric features [7]. Methods such as PREDATOR [19] and Geometric Transformer [31] use attention mechanisms to target overlapping regions for registration. In the domain of color features, SuperPoint and SuperGlue [11, 36] build a formative approach in GNN-based keypoint feature matching. Methods such as LoFTR [42] introduce more dense and accurate sub-pixel level matching. Despite being very successful in handling wide-baseline scenarios, learned feature matching still requires a significant amount of shared visibility and geometric overlap.", + "bbox": [ + 496, + 174, + 892, + 507 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Camera Pose Estimation with Semantic Cues. Several methods have been developed to incorporate semantic priors to improve low-overlap registration. PlaneMatch [38] proposed coplanarity priors for handling loop closures and low-overlap cases. Our method instead leverages object-centric constraints, exploiting the power of semantic object recognition. Another related direction is feature hallucination by leveraging image and object semantics. NeurHal [13] focuses on correspondence hallucination using image inpainting and outpainting, formulating a PnP optimization over hallucinated matches. Virtual Correspondence (VC) [24] introduces a human-centric approach that leverages hallucinated object volumes to form virtual epipolar constraints between frames. In contrast, we use indirect instead of direct correspondences that do not require hallucinated object volumes or image regions. Furthermore, our method works on a diverse set of furniture categories while VC focuses on humans. Pioneered by SLAM++ [35], there is also a rich literature of object-centric SLAM solutions, e.g., [26, 41]. Such SLAM methods leverage local, per-frame poses of objects to establish constraints; instead, we develop a global object pose optimization that is more robust against occluded viewpoints.", + "bbox": [ + 496, + 515, + 892, + 863 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Object Pose Estimation using Normalized Object Coordinates. 3D object pose estimation has been widely stud", + "bbox": [ + 498, + 869, + 890, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "13083", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/5fd461f8ba293bf3aa6fcb79e1190e4e440fb2ab3150d5e8e43c1ff99c53e9de.jpg", + "image_caption": [ + "Figure 2. Overview of our approach to incorporate object correspondence grounding in global pose estimation. From a set of input RGB-D frames, ObjectMatch predicts object instances for each frame with dense normalized object correspondences. The predicted object instances are used to identify objects across frames, forming indirect object correspondences. We combine object correspondences with SuperGlue [11,36] keypoint matches in a joint energy optimization that yields both camera and object poses in a global registration." + ], + "image_footnote": [], + "bbox": [ + 89, + 88, + 877, + 237 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ied from RGB and RGB-D inputs. Normalized Object Coordinate Space (NOCS) [43] was proposed to form dense correspondences from input RGB-D frames to canonical object geometries, enabling better generalization than direct regression of object poses. End2End SOCs [2] formulated a NOC-based approach for CAD retrieval and alignment to 3D scans, using a differentiable Procrustes algorithm. To enable CAD alignment to single RGB images, ROCA [14] leveraged NOC space in combination with predicted depths, formulating a robust differentiable Procrustes optimization [14]. Seeing Behind Objects [27] further leveraged NOC correspondences both to obtain local object poses and object completion for re-identification for RGB-D multi-object tracking. Wide Disparity Re-localization [26] uses the NOC predictions from [43] to construct an object-level map for re-localization in SLAM. In contrast to these approaches that focus on individual object poses, we use NOC correspondences directly in a multi-frame, global camera, and object pose optimization.", + "bbox": [ + 75, + 335, + 472, + 623 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 635, + 169, + 650 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Setup", + "text_level": 1, + "bbox": [ + 76, + 660, + 228, + 676 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given $K$ RGB-D frames $\\{(I_1^c, I_1^d), \\dots, (I_K^c, I_K^d)\\}$ , we aim to optimize their 6-DoF camera poses $T_c = \\{T_2, \\dots, T_K\\}$ , assuming the first frame is the reference, i.e., $T_1 = \\mathbb{I}$ . A 6-DoF camera pose $T_i$ is represented by Euler angles $\\gamma$ and translations $t, T_i = (\\gamma_x, \\gamma_y, \\gamma_z, t_x, t_y, t_z)$ .", + "bbox": [ + 76, + 681, + 468, + 758 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We also parameterize global, 9-DoF object poses, $\\bar{T}_o = (\\gamma_x,\\gamma_y,\\gamma_z,t_x,t_y,t_z,s_x,s_y,s_z)$ , comprising 6-DoF angles and translations, and 3-DoF anisotropic scales $s$ .", + "bbox": [ + 76, + 758, + 468, + 803 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We formulate a joint energy optimization of the form:", + "bbox": [ + 96, + 804, + 452, + 819 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nT ^ {*}, \\bar {T} ^ {*} = \\operatorname {a r g m i n} _ {T, \\bar {T}} \\left(\\mathrm {E} _ {c} (T, M) + \\mathrm {E} _ {o} (T, \\bar {T}, N)\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 828, + 468, + 847 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $M$ are inter-frame feature matches and $N$ are intra-frame canonical object correspondences established with normalized object coordinates (NOCs) that densely map", + "bbox": [ + 76, + 854, + 468, + 901 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/08e5f55989e5efa801d13fecfce2a35ccdde2a703ae97a58438d71d085dc9799.jpg", + "image_caption": [ + "Figure 3. Multi-modal object recognition and NOC prediction. Our ResNet-FPN [16, 21] backbone takes color, reversed jet colored depth [12], and $2\\mathrm{x}$ downsampled colored 3D depth normals, and produces multi-scale features by averaging different input encodings. From the obtained features, our method recognizes objects and predicts NOCs for each object, based on a Mask-RCNN [15]-style prediction." + ], + "image_footnote": [], + "bbox": [ + 511, + 310, + 877, + 481 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "to the canonical space of an object in $[-0.5, 0.5]^3$ . $\\mathrm{E}_c$ is the feature-matching energy function, and $\\mathrm{E}_o$ is our object-centric energy function. Since robust feature matching and optimization are readily available off the shelf [10, 11, 31, 36], our method focuses on building the object-centric energy $\\mathrm{E}_o$ . To this end, we need two function approximators, realized via deep neural networks: (1) a learned model for object recognition and NOC prediction, and (2) a learned model for object identification. The realization of these networks is described in Sections 3.2 and 3.3, respectively, and energy function Eq. 1 in Section 3.4. An overview of our approach is visualized in Figure 2.", + "bbox": [ + 496, + 609, + 892, + 790 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Predicting Object Correspondences", + "text_level": 1, + "bbox": [ + 500, + 801, + 808, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To obtain object constraints in our final optimization, we recognize objects via object detection and instance segmentation, and predict object correspondences as dense NOCs [43] for each object, as shown in Figure 3. We build on a Mask-RCNN [15] with ResNet50-FPN [16, 21] back", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "13084", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "bone, pre-trained on ImageNet [33] and COCO [22].", + "bbox": [ + 75, + 90, + 424, + 106 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To input the depth of the RGB-D frames, we propose a modified, multi-modal ResNet50-FPN [12, 16, 21] backbone. Our backbone takes $480 \\times 640$ color, $480 \\times 640$ reverse jet-colored depth, and $240 \\times 320$ colored depth normals as input. We average the resulting FPN features to a single feature pyramid:", + "bbox": [ + 75, + 106, + 468, + 196 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nG = \\frac {\\operatorname {F P N} ^ {c} \\left(I ^ {c}\\right) + \\operatorname {F P N} ^ {d} \\left(I ^ {d}\\right) + \\operatorname {U} \\left(\\operatorname {F P N} ^ {\\mathrm {n}} \\left(I ^ {n}\\right)\\right)}{3}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 205, + 468, + 239 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $I^{c}, I^{d}, I^{n}$ are color, depth, and normal images, $\\mathrm{FPN}^{c}, \\mathrm{FPN}^{d}, \\mathrm{FPN}^{n}$ are the corresponding ResNet50-FPN backbones, and U is an upsampling operator to match the normal features' spatial size with others. This enables fine-tuning the pre-trained bounding box, class, and instance segmentation heads of Mask-RCNN [15], while also exploiting depth information. We use symmetrically structured FPNs, all pre-trained on ImageNet and COCO as initialization, but without any parameter sharing.", + "bbox": [ + 75, + 247, + 468, + 383 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To obtain object correspondences, we establish mappings from detected objects to their canonical spaces, in the form of dense NOCs. That is, for each pixel in the object's instance mask, we predict 3D points $P_{o}^{\\mathrm{noc}}$ using a fully convolutional network:", + "bbox": [ + 75, + 383, + 468, + 458 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nP _ {o} ^ {\\text {n o c}} = \\operatorname {F C N} \\left(G _ {o}\\right), p \\in [ - 0. 5, 0. 5 ] ^ {3} \\forall p \\in P _ {o} ^ {\\text {n o c}}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 469, + 468, + 487 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We optimize an $\\ell_1$ loss $L_{\\mathrm{noc}}$ using ground-truth NOCs $P^{\\mathrm{noc - gt}}$", + "bbox": [ + 76, + 496, + 468, + 523 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {n o c}} = \\sum_ {o} \\sum_ {i} \\left\\| P _ {o, i} ^ {\\text {n o c}} - P _ {o} ^ {\\text {n o c - g t}} \\right\\| _ {1}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 523, + 468, + 554 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since symmetric objects [1, 43] induce ambiguities in NOCs (e.g., a round table), we classify symmetry type of objects (round, square, rectangle, non-symmetric), $c_{\\mathrm{sym}} = \\mathrm{MLP}^{\\mathrm{sym}}(G_o)$ , optimized using a cross-entropy loss $L_{\\mathrm{sym}}$ . We also make $L_{\\mathrm{noc}}$ symmetry aware, taking the minimum $\\ell_1$ difference over the set of correct NOCs [43]. We use non-symmetric objects during inference to avoid inconsistent NOCs across views.", + "bbox": [ + 75, + 559, + 468, + 679 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In addition to NOCs, we also regress anisotropic 3D object scales $s_o$ using a fully connected network, $s_o = \\mathrm{MLP}^{\\mathrm{scale}}(G_o)$ , and optimize $s_o$ with an $\\ell_1$ loss $L_{\\mathrm{scale}}$ . The object scale enables holistic object pose understanding within each frame and helps to filter potential object matches across views using scale consistency.", + "bbox": [ + 75, + 680, + 468, + 770 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, to make our NOC-depth correspondences least-squares friendly for our desired optimization, we also introduce a per-frame differentiable Procrustes objective $L_{\\mathrm{proc}}$ , using a differentiable Kabsch solver [32] to obtain local object rotations and translations:", + "bbox": [ + 75, + 771, + 468, + 845 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nR _ {o} ^ {*}, t _ {o} ^ {*} = \\operatorname {a r g m i n} _ {R _ {o}, t _ {o}} \\left(\\sum_ {i} \\left\\| R _ {o} \\left(P _ {o, i} ^ {\\text {n o c}} \\odot s _ {o}\\right) + t _ {o} - P _ {o, i} ^ {\\text {d e p t h}} \\right\\| _ {2} ^ {2}\\right) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 856, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/402ff96d4437294d38869e400320ad0ad7e7e2754222176c013a76f84600eec3.jpg", + "image_caption": [ + "Figure 4. Our foreground/background metric-learning encoder for object matching, inspired by re-OBJ [3]. Using the detected and segmented objects from the model in Section 3.2, we encode foreground and background regions of objects, using light-weight, multi-modal ResNet18 encoders on the RGB-D features." + ], + "image_footnote": [], + "bbox": [ + 506, + 90, + 883, + 236 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "for each object $o$ , where $P_{o}^{\\mathrm{depth}}$ are back-projected input RGB-D depths corresponding to the object's predicted instance mask in its region of interest, and $\\odot$ denotes element-wise multiplication. We train the local object poses with", + "bbox": [ + 498, + 334, + 890, + 397 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {p r o c}} = w _ {r} \\sum_ {o} | | R _ {o} ^ {*} - R _ {o} ^ {\\mathrm {g t}} | | _ {1} + w _ {t} \\sum_ {o} | | t _ {o} ^ {*} - t _ {o} ^ {\\mathrm {g t}} | | _ {2} ^ {2}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 406, + 890, + 438 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our full loss used for training is then", + "bbox": [ + 519, + 448, + 764, + 463 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL = L _ {m} + w _ {n} L _ {\\text {n o c}} + w _ {s} L _ {\\text {s c a l e}} + w _ {\\text {s y m}} L _ {\\text {s y m}} + w _ {p} L _ {\\text {p r o c}}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 508, + 474, + 890, + 491 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $L_{m}$ is the sum of Mask-RCNN losses [15] and $w$ are scalar weights balancing additional losses.", + "bbox": [ + 498, + 502, + 890, + 532 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Implementation. We use the augmented 400k ScanNet train image split for training [9, 14], with Scan2CAD labels of the most common 9 furniture categories following [1, 14, 25]. We train a standard Detector2 Mask-RCNN pipeline [15, 45] with 1k warm-up iterations, 0.003 base learning rate, and learning rate decays at 60k and 100k iterations with a total of 120k training iterations.", + "bbox": [ + 498, + 536, + 890, + 641 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Matching Object Instances", + "text_level": 1, + "bbox": [ + 500, + 650, + 743, + 667 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In our global pose optimization formulation, the relation between frames is formed via a global identification of objects across frames. To enable such identification without any heuristic spatial assumptions, we train a simple metric learner that ranks object semantic similarities across frames. Our object matching is shown in Figure 4.", + "bbox": [ + 496, + 674, + 890, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We characterize objects for matching by their respective RGB-D features in the object instance mask, in addition to the global context in the 5-times upscaled object bounding box. All inputs are resized to $224 \\times 224$ and input to a lightweight ResNet18 [16] backbone pre-trained on ImageNet [33].", + "bbox": [ + 496, + 763, + 890, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Similar to object detection, we employ two backbones for color and colored depth inputs. We omit normal input in object matching, as it empirically did not provide any", + "bbox": [ + 498, + 854, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "13085", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "benefit. For each input modality, we train two ResNet18 backbones for masked and inverse-masked crops, namely foreground (object) and background (context) encodings,", + "bbox": [ + 75, + 90, + 468, + 137 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ne = \\operatorname {M L P} \\left(\\left[ \\mathrm {R N} ^ {c} \\left(F _ {c}\\right), \\mathrm {R N} ^ {c} \\left(B _ {c}\\right) \\right] + \\left[ \\mathrm {R N} ^ {d} \\left(F _ {d}\\right), \\mathrm {R N} ^ {d} \\left(B _ {d}\\right) \\right]\\right) \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 83, + 143, + 468, + 176 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where RN are ResNet18s, MLP is a fully connected network, $F$ , $B$ are foreground and background crops for color (c) and depth (d), and $e$ is the object embedding vector.", + "bbox": [ + 76, + 176, + 468, + 222 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given an anchor object embedding $e_a$ , a correctly matching object embedding $e_p$ , and a negative example $e_n$ , we train our metric learning using a triplet margin loss [17]:", + "bbox": [ + 76, + 223, + 468, + 268 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {t r i}} = \\max \\left(\\mathrm {d} \\left(e _ {a}, e _ {p}\\right) - \\mathrm {d} \\left(e _ {a}, e _ {n}\\right) + 1. 0, 0\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 127, + 277, + 468, + 294 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $d$ is the $\\ell_2$ distance.", + "bbox": [ + 76, + 304, + 251, + 318 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We only consider triplets from the same category, as the object recognition pipeline provides classification. At inference time, we match the best instances using the Hungarian algorithm and apply a threshold $d(e_i, e_j) < \\alpha$ for the matching object pairs from the same class. This semantic matching can be scaled to multiple frames via, e.g., object tracking with re-identification, or in our case, a simple pose graph optimization over frame pairs.", + "bbox": [ + 75, + 319, + 468, + 439 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation. We implement the identification network using PyTorch [30] and train it on ScanNet data [9]. We train the network for 100k iterations with a batch size of 8 using a momentum optimizer with a learning rate 1e-4 and momentum 0.9.", + "bbox": [ + 75, + 443, + 468, + 517 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Energy Optimization", + "text_level": 1, + "bbox": [ + 76, + 527, + 274, + 542 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We realize the joint energy minimization in Eq. 1 using keypoint and NOC constraints.", + "bbox": [ + 75, + 550, + 468, + 579 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Using the predicted NOC constraints with back-projected depths, we can re-write the $\\mathrm{E}_o$ in Eq. 1 as:", + "bbox": [ + 75, + 580, + 468, + 611 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {E} _ {o} (T, \\bar {T}, P ^ {\\text {d e p t h}}, P ^ {\\text {n o c}}) = \\sum_ {o} \\sum_ {c} \\sum_ {k} \\| T _ {c} P _ {o, c, k} ^ {\\text {d e p t h}} - \\bar {T} _ {o} P _ {o, c, k} ^ {\\text {n o c}} \\| _ {2} ^ {2} \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 76, + 619, + 472, + 664 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\bar{T}$ , $T$ represent 9-DoF and 6-DoF camera and object transformations, respectively, and subscripts $o$ , $c$ , $k$ correspond to objects, cameras (frames), and points (pixels) within the frames, respectively. Here, object indices are determined by the object identification, and global object poses $\\bar{T}$ indirectly constrain frames to each other without any explicit inter-frame feature matching.", + "bbox": [ + 75, + 664, + 468, + 768 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In many cases, object constraints alone may not be sufficient to optimize the camera pose (e.g., frames may not share matching objects together). However, our object-based optimization is fully complementary to classical feature matching, and we thus formulate our objective in combination with feature-matching constraints $\\mathbf{E}_c$ :", + "bbox": [ + 75, + 771, + 468, + 861 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {E} _ {c} (T, P ^ {\\text {d e p t h}}) = \\sum_ {i} \\sum_ {j} \\sum_ {m, n} | | T _ {i} P _ {i, m} ^ {\\text {d e p t h}} - T _ {j} P _ {j, n} ^ {\\text {d e p t h}} | | _ {2} ^ {2} \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 869, + 468, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our method is agnostic to the feature matcher, whether classical or learning-based. In this work, we experiment with two different keypoint matching systems to realize $\\mathrm{E}_c$ , namely SuperGlue [11, 36] and Geometric Transformer [31], both offering state-of-the-art indoor feature matching.", + "bbox": [ + 498, + 90, + 890, + 166 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With both object and feature-matching constraints, we realize the desired joint energy formulation as", + "bbox": [ + 498, + 167, + 890, + 196 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nT ^ {*}, \\bar {T} ^ {*} = \\operatorname {a r g m i n} _ {T, \\bar {T}} \\left(w _ {c} \\mathrm {E} _ {c} + w _ {o} \\mathrm {E} _ {o}\\right). \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 205, + 890, + 222 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $w_{c}, w_{o}$ weight feature-matching, and object energies.", + "bbox": [ + 498, + 229, + 890, + 244 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since non-linear least squares problems can be sensitive to outliers, we additionally employ outlier removal. Similar to BundleFusion [10], we apply Kabsch filtering to both intra-frame and keypoint constraints, using the matching depth-NOC and depth-depth correspondences, respectively. That is, we iteratively solve an orthogonal Procrustes problem and only keep the correspondences that have lower optimization errors. We use a liberal $20\\mathrm{cm}$ threshold to handle wide-baseline frames. Objects and inter-frame matches are rejected if the number of NOCs is $< 15$ and the number of keypoints is $< 5$ , respectively.", + "bbox": [ + 496, + 244, + 890, + 411 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To solve this least-squares problem, we use a Gauss-Newton optimizer. To handle outliers during this global optimization, we remove $>15cm$ error residuals during optimization. As global optimization can produce coarse results, we further apply an ICP refinement.", + "bbox": [ + 496, + 411, + 890, + 486 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To handle SLAM-style sequence registration, we use a state-of-the-art global pose graph optimization [5, 47] over pairwise frame registration, using a single hierarchy level for simplicity. However, our method could be scaled to multi-frame optimization via object tracking and subdivided into multiple hierarchy levels to handle very large-scale scenes.", + "bbox": [ + 496, + 487, + 890, + 590 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 604, + 632, + 621 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Pairwise Registration", + "text_level": 1, + "bbox": [ + 500, + 628, + 700, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Dataset. We evaluate pairwise camera pose estimation results on 1569 challenging and diverse frame pairs from the validation set of ScanNet [9]. Unlike previous works [19, 36], our evaluation includes a significant number of low-overlap pairs, with over $15\\%$ of pairs having $\\leq 10\\%$ overlap.", + "bbox": [ + 496, + 652, + 890, + 743 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Optimization. To establish object correspondences across frame pairs in Eq. 10, we use the top-1 matching object with embedding distance threshold 0.05. In the absence of both keypoint matches and object matches, we consider object matches with distance $< 0.15$ . Only objects of the same class label and predicted maximum scale ratio $< 1.5$ can match. We determine the best matching object based on the number of NOC-depth constraints after the filtering described in Section 3.4. We combine it with keypoint constraints and refine the results with ICP.", + "bbox": [ + 496, + 750, + 890, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "13086", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Classical Baselines. We compare our method against various hand-crafted feature-matching and registration baselines. These baselines include BundleFusion keypoint optimization with SIFT descriptors and Kabsch filtering [10, 23], (SIFT + BF), and two geometric global registration baselines, Fast Global Registration (Fast GR) [46] and Global Registration (GR) [5], based on approximate and exact RANSAC over FPFH [34] features, respectively. As a post-processing step, all methods are refined using ICP. We use Open3D [47] implementations for global registration baselines as well as ICP post-processing of all methods. Our custom implementation of SIFT + BF uses the SIFT implementation of OpenCV [4].", + "bbox": [ + 76, + 90, + 472, + 287 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Learned Baselines. We also compare with pose estimators using learned models for feature or object matching: Geometric Transformer (GeoTrans) [31], which performs dense geometric feature matching, and SG + BF which leverages the powerful learned SuperGlue [36] and SuperPoint [11] feature matching approaches in combination with BundleFusion [10] Kabsch filtering and 3D Gauss-Newton optimization. We also use our network predictions to create a 3D object tracking baseline, Object Track, mimicking tracking-based object-SLAM systems. Object Track obtains local object poses and object matches from Sections 3.2 and 3.3, but instead uses relative local object poses to get the camera pose instead of a global energy optimization. All methods are refined using ICP as a post-processing step.", + "bbox": [ + 76, + 294, + 472, + 521 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. We use the Pose Recall metric for pose evaluation, following previous RGB-D pairwise registration works [6, 19]. To comprehensively capture performance, we employ several thresholds for absolute translation (in cm) and rotation (in °). Since absolute translation is more difficult than rotation, its thresholds are twice the angle thresholds, following [6].", + "bbox": [ + 76, + 525, + 470, + 633 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative Results. In Table 1 and Table 2, we evaluate on ScanNet [9] validation frame pairs, measuring recall in various thresholds and overlap levels. Combined with state-of-the-art feature matching in a joint optimization, our method outperforms both classical and learning-based methods. Furthermore, the gap increases with lower overlap, as shown in Table 2 and Figure 6, since our method can leverage object-based correspondences to estimate alignment in the absence of keypoint matches. We also show that our global optimization (Ours (w/o keypoints)) significantly outperforms naive object-based tracking (Object Track), demonstrating the efficacy of our global optimization formulation. Our approach is designed to complement keypoint matching, notably improving the state of the art in combination with state-of-the-art geometric and RGB-D feature matching.", + "bbox": [ + 76, + 637, + 472, + 878 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Analysis of Performance by Overlap. In Table 2 and", + "bbox": [ + 76, + 885, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9cfccc32790ad4679ee2bd4e20401e3ba5785621353cd5934e538fa5e306ca31.jpg", + "image_caption": [ + "NOCs of Top-1 Match" + ], + "image_footnote": [], + "bbox": [ + 516, + 107, + 764, + 181 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/aae1864fcbbfbbb16200eba2f316e34e1333cf68d68533dfdcc063a4fc80b427.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 181, + 763, + 256 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5e0ddfb0b0661574f7e42f78b245d399f194b0a22f8d27e0907bc5be33bd3749.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 258, + 763, + 332 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4f6d79e6f9bb0c7daaf728dfa343a8264343b45aaf5b9b601720d48f13b3fc73.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 334, + 761, + 407 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/98858c7490497f8d59f5fc3b36be9a2527dfd29ba4729fbaf06e26849334577c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 410, + 761, + 484 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/51b93a94c1fd6952d789c2402d7fa5511b9427d76f97310574e952ede3dd4fca.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 486, + 763, + 560 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2c42047b9fec4bd00946536866713f105e93ff83637d1c6732d1d28dcea880d8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 561, + 763, + 636 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/53f9ace2a1b7a4da49930ca4fdd768a56fc8669dae68c3cfd25a9278f2faf2e5.jpg", + "image_caption": [ + "Figure 5. Low-overlap registration on ScanNet [9], where traditional feature matching fails. Predicted NOC correspondences are visualized, along with object box and camera poses of the left and right images in green and blue, respectively." + ], + "image_footnote": [], + "bbox": [ + 517, + 638, + 763, + 710 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9d1e9c4da8637be8c38427dc11bb77f3c7d08ec70909acbce7aab359430a47bd.jpg", + "image_caption": [ + "Registration" + ], + "image_footnote": [], + "bbox": [ + 764, + 107, + 870, + 181 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/68865de6741342b2bb7fc162ae1b181069c1995ec38897caf92bb9cccf83a88d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 763, + 185, + 870, + 253 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e4830fff8a54603d09f95da77b25b63df1446127b4b3feb9a94a373211a8818c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 764, + 258, + 864, + 332 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b36974f6f162f2d9568da7dcbaf7c449602704d4c1b128b1f3ba285bc9aef9dd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 766, + 337, + 857, + 407 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/242be60d142942c8ef0bababb8327e87f8f07153a8791633fe3f48b0a43c6056.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 766, + 414, + 867, + 484 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3ea99be09070dccc218cf16ea93cb5422577435b1920255cdd9cb61207302a2a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 764, + 494, + 864, + 560 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ae863170fd7256e6b7adb669a803c7cba9509a1542f9890621d9e3591d6a7594.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 766, + 563, + 875, + 635 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/bc2838c48f0271e24329e4d7d63266202f03b2acbf61030549174569ce6c0f71.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 774, + 643, + 875, + 710 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Figure 6, we show $30\\mathrm{cm}$ , $15^{\\circ}$ recall performance at different geometric overlap rates, using the radius-based geometric overlap percentage measurement from [19] with $1\\mathrm{cm}$ threshold. While all methods' performances decrease with overlap, our method retains significantly more robustness. Classical keypoint matching (SIFT + BF, Fast GR,", + "bbox": [ + 498, + 809, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "13087", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e8edc373384b7b333a0c213a65d5682773f680f28796862a345d230c37bc5084.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodPose Recall by Threshold
5°, 10cm10°, 20cm15°, 30cm
SIFT + BF [10, 23]16.8318.4220.40
Fast GR [34, 46]39.0742.9644.68
GR [5]43.0847.6750.03
Object Track37.4143.2145.19
GeoTrans [31]67.1174.5776.93
SG + BF [10, 11, 36]71.1380.1181.71
Ours (w/o keypoints)56.4163.6766.99
Ours (w/ GeoTrans)68.3276.8079.35
Ours (w/ SG + BF)74.2584.5887.06
", + "bbox": [ + 89, + 88, + 452, + 236 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/fec60ddaa3b53f7c5ae76095c1830e1f990f8687fdd370fea4f811d842b723c5.jpg", + "table_caption": [ + "Table 1. Pose Recall results on ScanNet [9] validation images. Combined with state-of-the-art feature matching, our method outperforms various classical and learning-based baselines. Our approach complements both Geometric Transformer [31] (Ours (w/ GeoTrans)) and SuperGlue [36] (Ours (w/ SG + BF)) feature matches, notably improving pose recall." + ], + "table_footnote": [], + "table_body": "
MethodRecall by Overlap %
≤ 10(10,30)≥ 30
SIFT + BF [10,23]1.302.2033.59
Fast GR [34,46]0.0013.9769.04
GR [5,34]0.6520.3674.62
Object Track8.4431.5458.86
GeoTrans [31]22.0864.2793.11
SG + BF [10,11,36]24.0373.4595.95
Ours (w/o keypoints)32.4753.2980.31
Ours (w/ GeoTrans)35.0665.4794.42
Ours (w/ SG + BF)45.4581.4497.16
", + "bbox": [ + 109, + 337, + 433, + 494 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Recall at $15^{\\circ}$ , $30\\mathrm{cm}$ by overlap percentage. Our method significantly outperforms strong baselines in challenging low-overlap frame pairs, almost doubling recall for overlap $\\leq 10\\%$ .", + "bbox": [ + 75, + 500, + 470, + 542 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d970bedea8200bb3a2326b3b67f1fe102e89862b86d5356040df18c8774cddbe.jpg", + "image_caption": [ + "Figure 6. Change of pose performance by overlap percentage on ScanNet [9] validation pairs. Our method is significantly more robust against decreasing overlap compared to various classical and learned baselines." + ], + "image_footnote": [], + "bbox": [ + 106, + 558, + 439, + 753 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "SIFT) suffer strongly in low-overlap scenarios, which are challenging for hand-crafted descriptors. The learned approaches of GeoTrans and SG+BF maintain some performance with decreasing overlap but still suffer strongly. Our", + "bbox": [ + 75, + 839, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "incorporation of object-based reasoning effectively complements keypoint matching while significantly improving robustness to low overlap, obtaining a performance improvement from $24.03\\%$ to $45.45\\%$ for overlap $\\leq 10\\%$ .", + "bbox": [ + 496, + 90, + 890, + 151 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative Results. Figure 5 shows examples of low-overlap registration on ScanNet [9] frame pairs. Due to the minimal overlap, traditional keypoint matching cannot find sufficient correspondences, while our object grounding enables accurate camera pose estimation through indirect object constraints.", + "bbox": [ + 496, + 156, + 890, + 247 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Registration of RGB-D SLAM Sequences", + "text_level": 1, + "bbox": [ + 498, + 255, + 852, + 272 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We further demonstrate the effectiveness of our method in an RGB-D SLAM setting, registering sequences of frames in room-scale scenes. In particular, we evaluate challenging, low-frame-rate settings that reflect common non-expert capture such as fast camera motion.", + "bbox": [ + 496, + 279, + 890, + 353 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For optimization, we use a global pose graph optimization over pairwise registration similar to the multi-way global registration of [5, 6, 47], as it offers an off-the-shelf, simple, and robust solution. We refer to our supplementary material for further optimization details.", + "bbox": [ + 496, + 354, + 890, + 430 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Datasets and Evaluation Metrics. We evaluate our method on TUM-RGBD [39] and ScanNet [9] scenes. For ScanNet, we use a set of 12 scenes from the validation and test sets, having a wide range of sizes, environments, and camera trajectories. We sample every 30th frame for TUM-RGBD (1Hz) and every 20th frame in ScanNet (1.5Hz). All methods use the same hyperparameters for each dataset, respectively. To evaluate registration quality, we use the standard root mean squared trajectory error used for TUM-RGBD evaluation [39].", + "bbox": [ + 496, + 435, + 890, + 585 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Baselines. We implement various popular feature descriptors from cutting-edge SLAM systems in global pose graph optimization setups and thoroughly tune the hyperparameters of each method to support low-frame-rate sequences. We compare against Redwood [5], a global multi-way pose graph that uses FPFH [34] features. We also deploy robust RGB-D SIFT features used in BundleFusion [10, 23] in a global pose graph optimization, creating the SIFT + BF baseline. Finally, we use the state-of-the-art SuperGlue matches of SuperPoint features with BundleFusion's sparse 3D optimization, SG + BF, as described in Section 4.1.", + "bbox": [ + 496, + 592, + 890, + 772 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative Results. In Table 3, we compare our method to state-of-the-art traditional and learned SLAM approaches. Across both TUM-RGBD and ScanNet validation and test scenes [9, 39], our object-grounded approach enables improved camera pose estimation. In particular, for the larger-scale ScanNet scenes, ObjectMatch achieves notable reductions in trajectory error, of $38\\%$ , $49\\%$ , $50\\%$ , and $44\\%$ in scenes 0423, 0430, 0718, and 0773.", + "bbox": [ + 496, + 779, + 890, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "13088", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ce8f08b2ef19c6461899ee6dfecfb9cf6ebbcb2117ee279f9e07c4c866a052ff.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodTUM Scene @ 1HzScanNet Scene @ 1.5Hz
fr1/deskfr1/xyzfr2/xyzfr2/deskfr3/office001100810169020704230430046104940718076007730805
SIFT + BF [10,23]30.603.492.73157.0238.01222.49108.03126.57104.64160.69248.0020.9033.9881.4838.0322.4786.80
Redwood [5,34,47]3.101.971.88140.17175.64164.0150.9647.2447.1788.43156.7610.479.77133.536.5543.19253.39
SG + BF [10,11,36]3.011.921.815.305.1413.886.778.095.627.6735.973.317.1613.426.3712.4312.98
Ours (w/ SG + BF)3.011.921.815.264.9212.535.736.855.324.7518.232.676.856.705.887.0010.86
", + "bbox": [ + 81, + 88, + 883, + 148 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. ATE RMSE values (cm) on room-scale TUM-RGBD [39] and ScanNet [9] test scenes. We evaluate at challenging 1 FPS and 1.5 FPS by sampling every 30th and 20th frame from TUM-RGBD and ScanNet, respectively. ObjectMatch outperforms strong baselines leveraging classical and learning-based SLAM systems.", + "bbox": [ + 76, + 152, + 892, + 196 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c64e4018aceca06ceec3712dd2fdac8ab735dda3de5ffca06545abbeb9227de9.jpg", + "image_caption": [ + "Figure 7. Qualitative comparison for registration of ScanNet scenes [9], in comparison with only feature matching SG + BF [10, 11, 36]. All results are visualized by volumetric fusion of the respective RGB-D frames with the estimated poses. Our optimized poses produce more accurate, clean, and consistent reconstruction, due to incorporating the object-based information in our global optimization. Example loop closures with matching object correspondences are shown on the right." + ], + "image_footnote": [], + "bbox": [ + 81, + 220, + 460, + 542 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Qualitative Results. In Figure 7, we show ScanNet [9] scene reconstructions using estimated camera poses. RGB-D frames are fused using a scalable TSDF volumetric fusion [8,47]. Our approach complements state-of-the-art feature matching [10, 11, 36], enabling more accurate and consistent reconstructions using our indirect correspondences, in comparison to feature matching alone. We refer to the supplemental for additional qualitative sequence registration results.", + "bbox": [ + 75, + 689, + 468, + 825 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations and Future Work. ObjectMatch shows the capability of incorporating object semantics into global pose optimization; however, various limitations remain. In particular, our approach leverages higher-level seman", + "bbox": [ + 75, + 839, + 468, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tic correspondence given by objects for registration; however, in scenarios where object matches do not exist in frame views, we cannot leverage this constraint and instead use only feature matches. Further leveraging background information regarding shared structures seen from different views would provide additional semantic correspondence information to further complement registration. Additionally, while estimating canonical correspondences with objects can provide significant view information, object alignments can be somewhat coarse, while reasoning at a finer-grained level of object parts would provide more precise reasoning. Finally, we believe future work can reduce our method's system complexity by adopting a joint multiframe SLAM optimization with fewer hyperparameters as well as a unified end-to-end architecture for learning object correspondences and identification together.", + "bbox": [ + 496, + 220, + 893, + 464 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 476, + 617, + 491 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have presented ObjectMatch which introduces a new paradigm for incorporating object semantics into camera pose registration. We leverage indirect, canonical object correspondences established with normalized object coordinates for camera pose optimization aspects in RGB-D registration and SLAM. To obtain these correspondences, we propose two multi-modal neural networks for object recognition with normalized object coordinate prediction and object identification. ObjectMatch operates in tandem with state-of-the-art feature matching in a joint Gauss-Newton optimization, with its object grounding enabling registration to handle frame pairs with very low to no shared visibility. As a result, ObjectMatch significantly improves the state-of-the-art feature matching in both pairwise and sequence registration of RGB-D frames, particularly in the challenging low-overlap regime. Overall, we hope our method opens up new possibilities in the context of leveraging semantic information for camera pose estimation.", + "bbox": [ + 496, + 501, + 890, + 773 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This work was supported by the ERC Starting Grant Scan2CAD (804724), the Bavarian State Ministry of Science and the Arts coordinated by the Bavarian Research Institute for Digital Transformation (bidt), the German Research Foundation (DFG) Grant \"Making Machine Learning on Static and Dynamic 3D Data Practical\", and the German Research Foundation (DFG) Research Unit \"Learning and Simulation in Visual Computing.\"", + "bbox": [ + 496, + 779, + 892, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "13089", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Armen Avetisyan, Manuel Dahnert, Angela Dai, Manolis Savva, Angel X Chang, and Matthias Nießner. Scan2cad: Learning cad model alignment in rgb-d scans. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 2614-2623, 2019. 4", + "[2] Armen Avetisyan, Angela Dai, and Matthias Nießner. End-to-end cad model retrieval and 9dof alignment in 3d scans. In Proceedings of the IEEE/CVF International Conference on computer vision, pages 2551-2560, 2019. 3", + "[3] Vaibhav Bansal, Stuart James, and Alessio Del Bue. re-object: Jointly learning the foreground and background for object instance re-identification. In International Conference on Image Analysis and Processing, pages 402–413. Springer, 2019. 4", + "[4] G. Bradski. The OpenCV Library. Dr. Dobb's Journal of Software Tools, 2000. 6", + "[5] Sungjoon Choi, Qian-Yi Zhou, and Vladlen Koltun. Robust reconstruction of indoor scenes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5556-5565, 2015. 2, 5, 6, 7, 8", + "[6] Christopher Choy, Wei Dong, and Vladlen Koltun. Deep global registration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2514-2523, 2020. 2, 6, 7", + "[7] Christopher Choy, Jaesik Park, and Vladlen Koltun. Fully convolutional geometric features. In ICCV, 2019. 2", + "[8] Brian Curless and Marc Levoy. A volumetric method for building complex models from range images. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, pages 303-312. ACM, 1996. 8", + "[9] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 2, 4, 5, 6, 7, 8", + "[10] Angela Dai, Matthias Nießner, Michael Zollhöfer, Shahram Izadi, and Christian Theobalt. Bundlefusion: Real-time globally consistent 3d reconstruction using on-the-fly surface reintegration. ACM Transactions on Graphics (ToG), 36(4):1, 2017. 2, 3, 5, 6, 7, 8", + "[11] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 224-236, 2018. 2, 3, 5, 6, 7, 8", + "[12] Andreas Eitel, Jost Tobias Springenberg, Luciano Spinello, Martin Riedmiller, and Wolfram Burgard. Multimodal deep learning for robust rgb-d object recognition. In 2015 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 681-687. IEEE, 2015. 3, 4", + "[13] Hugo Germain, Vincent Lepetit, and Guillaume Bourmaud. Visual correspondence hallucination. In International Conference on Learning Representations, 2022. 2", + "[14] Can Gümeli, Angela Dai, and Matthias Nießner. Roca: Robust cad model retrieval and alignment from a single image. In Proceedings of the IEEE/CVF Conference on Computer" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Vision and Pattern Recognition, pages 4022-4031, 2022. 3, 4", + "[15] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 3, 4", + "[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3, 4", + "[17] Elad Hoffer and Nir Ailon. Deep metric learning using triplet network. In International workshop on similarity-based pattern recognition, pages 84-92. Springer, 2015. 5", + "[18] Jiahui Huang, Shi-Sheng Huang, Haoxuan Song, and ShiMin Hu. Di-fusion: Online implicit 3d reconstruction with deep priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8932-8941, 2021. 2", + "[19] Shengyu Huang, Zan Gojcic, Mikhail Usvyatsov, Andreas Wieser, and Konrad Schindler. Predator: Registration of 3d point clouds with low overlap. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4267-4276, June 2021. 2, 5, 6", + "[20] Shahram Izadi, David Kim, Otmar Hilliges, David Molyneaux, Richard Newcombe, Pushmeet Kohli, Jamie Shotton, Steve Hodges, Dustin Freeman, Andrew Davison, et al. Kinectfusion: real-time 3d reconstruction and interaction using a moving depth camera. In Proceedings of the 24th annual ACM symposium on User interface software and technology, pages 559-568, 2011. 2", + "[21] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2117-2125, 2017. 3, 4", + "[22] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 4", + "[23] David G Lowe. Object recognition from local scale-invariant features. In Proceedings of the seventh IEEE international conference on computer vision, volume 2, pages 1150-1157. IEEE, 1999. 6, 7, 8", + "[24] Wei-Chiu Ma, Anqi Joyce Yang, Shenlong Wang, Raquel Urtasun, and Antonio Torralba. Virtual correspondence: Humans as a cue for extreme-view geometry. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15924-15934, 2022. 2", + "[25] Kevis-Kokitsi Maninis, Stefan Popov, Matthias Niesser, and Vittorio Ferrari. Vid2cad: Cad model alignment using multiview constraints from videos. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 4", + "[26] Yuhang Ming, Xingrui Yang, and Andrew Calway. Object-augmented rgb-d slam for wide-disparity relocalisation. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 2203-2209. IEEE, 2021. 2, 3" + ], + "bbox": [ + 503, + 92, + 893, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "13090", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Norman Muller, Yu-Shiang Wong, Niloy J Mitra, Angela Dai, and Matthias Nießner. Seeing behind objects for 3d multi-object tracking in rgb-d sequences. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6071-6080, 2021. 3", + "[28] Raul Mur-Artal and Juan D Tardós. Orb-slam2: An open-source slam system for monocular, stereo, and rgb-d cameras. IEEE transactions on robotics, 33(5):1255-1262, 2017. 2", + "[29] Matthias Nießner, Michael Zollhöfer, Shahram Izadi, and Marc Stamminger. Real-time 3d reconstruction at scale using voxel hashing. ACM Transactions on Graphics (ToG), 32(6):1-11, 2013. 2", + "[30] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems 32, pages 8024-8035. Curran Associates, Inc., 2019. 5", + "[31] Zheng Qin, Hao Yu, Changjian Wang, Yulan Guo, Yuxing Peng, and Kai Xu. Geometric transformer for fast and robust point cloud registration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11143-11152, 2022. 2, 3, 5, 6, 7", + "[32] Nikhila Ravi, Jeremy Reizenstein, David Novotny, Taylor Gordon, Wan-Yen Lo, Justin Johnson, and Georgia Gkioxari. Accelerating 3d deep learning with pytorch3d. arXiv:2007.08501, 2020.4", + "[33] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115(3):211-252, 2015. 4", + "[34] Radu Bogdan Rusu, Nico Blodow, and Michael Beetz. Fast point feature histograms (fpfh) for 3d registration. In 2009 IEEE international conference on robotics and automation, pages 3212-3217. IEEE, 2009. 2, 6, 7, 8", + "[35] Renato F Salas-Moreno, Richard A Newcombe, Hauke Strasdat, Paul HJ Kelly, and Andrew J Davison. Slam++: Simultaneous localisation and mapping at the level of objects. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1352-1359, 2013. 2", + "[36] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4938–4947, 2020. 2, 3, 5, 6, 7, 8", + "[37] Thomas Schops, Torsten Sattler, and Marc Pollefeys. Bad slam: Bundle adjusted direct rgb-d slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 134-144, 2019. 2", + "[38] Yifei Shi, Kai Xu, Matthias Niessner, Szymon Rusinkiewicz, and Thomas Funkhouser. Planematch: Patch coplanarity prediction for robust rgb-d reconstruction. In Proceedings of the" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "European Conference on Computer Vision (ECCV), pages 750-766, 2018. 2", + "[39] J. Sturm, N. Engelhard, F. Endres, W. Burgard, and D. Cremers. A benchmark for the evaluation of rgb-d slam systems. In Proc. of the International Conference on Intelligent Robot Systems (IROS), Oct. 2012. 2, 7, 8", + "[40] Edgar Sucar, Shikun Liu, Joseph Ortiz, and Andrew J Davison. imap: Implicit mapping and positioning in real-time. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6229-6238, 2021. 2", + "[41] Edgar Sucar, Kentaro Wada, and Andrew Davison. NodeSLAM: Neural object descriptors for multi-view shape reconstruction. In Proceedings of the International Conference on 3D Vision (3DV), 2020. 2", + "[42] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8922-8931, 2021. 2", + "[43] He Wang, Srinath Sridhar, Jingwei Huang, Julien Valentin, Shuran Song, and Leonidas J Guibas. Normalized object coordinate space for category-level 6d object pose and size estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2642-2651, 2019. 2, 3, 4", + "[44] T Whelan, S Leutenegger, RF. Salas-Moreno, B Glocker, and AJ. Davison. Elasticfusion: Dense slam without a pose graph. In Robotics: Science and Systems, 2015. 2", + "[45] Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. Detector2. https://github.com/facebookresearch/detectron2, 2019.4", + "[46] Qian-Yi Zhou, Jaesik Park, and Vladlen Koltun. Fast global registration. In European conference on computer vision, pages 766-782. Springer, 2016. 2, 6, 7", + "[47] Qian-Yi Zhou, Jaesik Park, and Vladlen Koltun. Open3D: A modern library for 3D data processing. arXiv:1801.09847, 2018. 5, 6, 7, 8", + "[48] Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hujun Bao, Zhaopeng Cui, Martin R Oswald, and Marc Pollefeys. Nice-slam: Neural implicit scalable encoding for slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12786-12796, 2022. 2" + ], + "bbox": [ + 501, + 92, + 893, + 698 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "13091", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/ObjectMatch_ Robust Registration Using Canonical Object Correspondences/9d0104e4-40f2-4a9d-bf9d-8065e49cc737_model.json b/2023/ObjectMatch_ Robust Registration Using Canonical Object Correspondences/9d0104e4-40f2-4a9d-bf9d-8065e49cc737_model.json new file mode 100644 index 0000000000000000000000000000000000000000..af4e57f1c0b05d8d3ebdaf513cd0fc52feba245f --- /dev/null +++ b/2023/ObjectMatch_ Robust Registration Using Canonical Object Correspondences/9d0104e4-40f2-4a9d-bf9d-8065e49cc737_model.json @@ -0,0 +1,2398 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.099, + 0.131, + 0.872, + 0.153 + ], + "angle": 0, + "content": "ObjectMatch: Robust Registration using Canonical Object Correspondences" + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.181, + 0.34, + 0.198 + ], + "angle": 0, + "content": "Can Gumeli" + }, + { + "type": "text", + "bbox": [ + 0.418, + 0.182, + 0.512, + 0.199 + ], + "angle": 0, + "content": "Angela Dai" + }, + { + "type": "text", + "bbox": [ + 0.591, + 0.182, + 0.731, + 0.198 + ], + "angle": 0, + "content": "Matthias Nießner" + }, + { + "type": "title", + "bbox": [ + 0.356, + 0.217, + 0.61, + 0.234 + ], + "angle": 0, + "content": "Technical University of Munich" + }, + { + "type": "image_caption", + "bbox": [ + 0.147, + 0.269, + 0.414, + 0.284 + ], + "angle": 0, + "content": "Low-overlap Frames with Matching Objects" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.288, + 0.275, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.279, + 0.288, + 0.455, + 0.391 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.455, + 0.333, + 0.476, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.506, + 0.269, + 0.678, + 0.283 + ], + "angle": 0, + "content": "Canonical Correspondences" + }, + { + "type": "image", + "bbox": [ + 0.476, + 0.288, + 0.707, + 0.391 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.722, + 0.269, + 0.88, + 0.283 + ], + "angle": 0, + "content": "Camera and Object Poses" + }, + { + "type": "image", + "bbox": [ + 0.729, + 0.285, + 0.875, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.403, + 0.893, + 0.445 + ], + "angle": 0, + "content": "Figure 1. Modern camera pose estimation relies on feature matching between overlapping frames - in this work, we present ObjectMatch to find correspondences between frames with little or no overlap by predicting semantic mappings through canonical object correspondences. The images above share no direct overlap, yet our method establishes indirect correspondences, thus enabling a successful registration." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.494, + 0.313, + 0.509 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.524, + 0.474, + 0.844 + ], + "angle": 0, + "content": "We present ObjectMatch1, a semantic and object-centric camera pose estimator for RGB-D SLAM pipelines. Modern camera pose estimators rely on direct correspondences of overlapping regions between frames; however, they cannot align camera frames with little or no overlap. In this work, we propose to leverage indirect correspondences obtained via semantic object identification. For instance, when an object is seen from the front in one frame and from the back in another frame, we can provide additional pose constraints through canonical object correspondences. We first propose a neural network to predict such correspondences on a per-pixel level, which we then combine in our energy formulation with state-of-the-art keypoint matching solved with a joint Gauss-Newton optimization. In a pairwise setting, our method improves registration recall of state-of-the-art feature matching, including from \\(24\\%\\) to \\(45\\%\\) in pairs with \\(10\\%\\) or less inter-frame overlap. In registering RGB-D sequences, our method outperforms cutting-edge SLAM baselines in challenging, low-frame-rate scenarios, achieving more than \\(35\\%\\) reduction in trajectory error in multiple scenes." + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.457, + 0.631, + 0.473 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.486, + 0.892, + 0.607 + ], + "angle": 0, + "content": "RGB-D registration and 3D SLAM has been a fundamental task in computer vision, with significant study and enabling many applications in mixed reality, robotics, and content creation. Central to both state-of-the-art traditional and learning-based camera pose estimation is establishing correspondences between points in input frames. However, correspondence estimation remains quite challenging when there is little or no overlap between frames." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.61, + 0.893, + 0.822 + ], + "angle": 0, + "content": "In contrast, humans can easily localize across these challenging scenarios by leveraging additional semantic knowledge – in particular, by further localizing at the level of objects and identifying matching objects between views. For instance, when observing a chair from the back and the side (e.g., in Figure 1), view overlap is minimal (or even no view overlap), resulting in failed registration from keypoint matching. However, the semantic knowledge of the chair and its object pose nonetheless enables humans to estimate the poses from which the front and side views were taken. Thus, we propose to take a new perspective on camera pose estimation and imbue camera registration with awareness of such semantic correspondences between objects for robust performance in these challenging scenarios." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "To this end, we propose ObjectMatch, a new paradigm for camera pose estimation leveraging canonical object correspondences in tandem with local keypoint correspondences between views. This enables significantly more robust registration under a variety of challenging scenar-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.419, + 0.9 + ], + "angle": 0, + "content": "1https://cangumeli.github.io/ObjectMatch/" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13082" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.302 + ], + "angle": 0, + "content": "ios, including low view overlap. For a sequence of input frames, ObjectMatch learns to semantically identify objects across frames, enabling a compact, global parameterization of 9-DoF object poses. Object correspondences are established through predicting normalized object coordinates [43], dense correspondences from object pixels to a canonically oriented space for each object. We then formulate a joint camera and object pose optimization that constrains object correspondences indirectly, operating irrespective of the shared visibility of image regions. Our approach is complementary to state-of-the-art SLAM methods, and we leverage our energy formulation to complement state-of-the-art keypoint matching [10, 11, 36] in a joint Gauss-Newton optimization." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.303, + 0.473, + 0.454 + ], + "angle": 0, + "content": "Our method outperforms strong baselines in both pairwise registration and registration of RGB-D frame sequences. In pairwise registration of challenging ScanNet [9] image pairs, we improve pose recall from \\(24\\%\\) to \\(45\\%\\) when the overlap is below \\(10\\%\\). On sequence registration of room-scale RGB-D scenes, our method outperforms various strong baselines in difficult, low-frame-rate settings in several TUM-RGBD [39] and ScanNet [9] scenes, reducing the trajectory error by more than \\(35\\%\\) in multiple challenging scenes." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.455, + 0.384, + 0.469 + ], + "angle": 0, + "content": "To sum up, our main contributions include:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.473, + 0.469, + 0.534 + ], + "angle": 0, + "content": "- An object-centric camera pose estimator that can handle low-overlap frame sets via indirect, canonical object correspondences established with predicted dense, per-pixel normalized object coordinates." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.538, + 0.469, + 0.584 + ], + "angle": 0, + "content": "- A joint energy formulation that leverages semantic object identification and dense, normalized object coordinates corresponding to canonical object geometries." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.589, + 0.469, + 0.695 + ], + "angle": 0, + "content": "- Our semantic grounding of object correspondences enables significantly more robust registration in low-overlap and low-frame-rate cases. ObjectMatch improves over state of the art from \\(24\\%\\) to \\(45\\%\\) registration recall of \\(\\leq 10\\%\\) overlap frame pairs and achieves over \\(35\\%\\) trajectory error reduction in several challenging sequences." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.473, + 0.469, + 0.695 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.71, + 0.22, + 0.725 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.473, + 0.903 + ], + "angle": 0, + "content": "RGB-D Registration and SLAM. In recent years, there have been many advances in indoor RGB-D reconstruction. Earlier RGB-D fusion approaches focus on frame-to-model camera tracking [20, 29]. To handle the loop closures better, more recent SLAM systems introduce explicit strategies or global optimization methods for handling loop closures through global optimization [5,10,28,37,44] to fix tracking errors. More recently, deep learning techniques have been applied to registration and SLAM scenarios, with methods ranging from geometric point cloud registration [6, 19, 31] as well as neural field based SLAM techniques [18, 40, 48]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "Despite all the successes in RGB-D SLAM and registration, the task is still challenging since incomplete loop closures observed via low-overlap frames cannot be handled, and most SLAM methods require a very high overlap between consecutive frames to track cameras accurately." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.175, + 0.893, + 0.508 + ], + "angle": 0, + "content": "Feature Matching. Modern RGB(-D) camera pose estimators rely on a feature-matching backbone. Classical global registration techniques [5,46] use FPFH features [34] over point cloud fragments. On the other hand, many global RGB-D SLAM techniques rely on sparse color features [10, 28]. While being successful in many scenarios, conventional feature matching often fails when the interframe overlap is low. Therefore, deep learning techniques have been utilized for predicting overlapping regions based on geometry or color. On the geometric side, Deep Global Registration [6] predicts overlapping point features using nearest neighbor search over learned geometric features [7]. Methods such as PREDATOR [19] and Geometric Transformer [31] use attention mechanisms to target overlapping regions for registration. In the domain of color features, SuperPoint and SuperGlue [11, 36] build a formative approach in GNN-based keypoint feature matching. Methods such as LoFTR [42] introduce more dense and accurate sub-pixel level matching. Despite being very successful in handling wide-baseline scenarios, learned feature matching still requires a significant amount of shared visibility and geometric overlap." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.516, + 0.893, + 0.864 + ], + "angle": 0, + "content": "Camera Pose Estimation with Semantic Cues. Several methods have been developed to incorporate semantic priors to improve low-overlap registration. PlaneMatch [38] proposed coplanarity priors for handling loop closures and low-overlap cases. Our method instead leverages object-centric constraints, exploiting the power of semantic object recognition. Another related direction is feature hallucination by leveraging image and object semantics. NeurHal [13] focuses on correspondence hallucination using image inpainting and outpainting, formulating a PnP optimization over hallucinated matches. Virtual Correspondence (VC) [24] introduces a human-centric approach that leverages hallucinated object volumes to form virtual epipolar constraints between frames. In contrast, we use indirect instead of direct correspondences that do not require hallucinated object volumes or image regions. Furthermore, our method works on a diverse set of furniture categories while VC focuses on humans. Pioneered by SLAM++ [35], there is also a rich literature of object-centric SLAM solutions, e.g., [26, 41]. Such SLAM methods leverage local, per-frame poses of objects to establish constraints; instead, we develop a global object pose optimization that is more robust against occluded viewpoints." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Object Pose Estimation using Normalized Object Coordinates. 3D object pose estimation has been widely stud" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13083" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.089, + 0.878, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.243, + 0.895, + 0.3 + ], + "angle": 0, + "content": "Figure 2. Overview of our approach to incorporate object correspondence grounding in global pose estimation. From a set of input RGB-D frames, ObjectMatch predicts object instances for each frame with dense normalized object correspondences. The predicted object instances are used to identify objects across frames, forming indirect object correspondences. We combine object correspondences with SuperGlue [11,36] keypoint matches in a joint energy optimization that yields both camera and object poses in a global registration." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.336, + 0.473, + 0.624 + ], + "angle": 0, + "content": "ied from RGB and RGB-D inputs. Normalized Object Coordinate Space (NOCS) [43] was proposed to form dense correspondences from input RGB-D frames to canonical object geometries, enabling better generalization than direct regression of object poses. End2End SOCs [2] formulated a NOC-based approach for CAD retrieval and alignment to 3D scans, using a differentiable Procrustes algorithm. To enable CAD alignment to single RGB images, ROCA [14] leveraged NOC space in combination with predicted depths, formulating a robust differentiable Procrustes optimization [14]. Seeing Behind Objects [27] further leveraged NOC correspondences both to obtain local object poses and object completion for re-identification for RGB-D multi-object tracking. Wide Disparity Re-localization [26] uses the NOC predictions from [43] to construct an object-level map for re-localization in SLAM. In contrast to these approaches that focus on individual object poses, we use NOC correspondences directly in a multi-frame, global camera, and object pose optimization." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.636, + 0.17, + 0.651 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.661, + 0.229, + 0.677 + ], + "angle": 0, + "content": "3.1. Problem Setup" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.683, + 0.47, + 0.759 + ], + "angle": 0, + "content": "Given \\(K\\) RGB-D frames \\(\\{(I_1^c, I_1^d), \\dots, (I_K^c, I_K^d)\\}\\), we aim to optimize their 6-DoF camera poses \\(T_c = \\{T_2, \\dots, T_K\\}\\), assuming the first frame is the reference, i.e., \\(T_1 = \\mathbb{I}\\). A 6-DoF camera pose \\(T_i\\) is represented by Euler angles \\(\\gamma\\) and translations \\(t, T_i = (\\gamma_x, \\gamma_y, \\gamma_z, t_x, t_y, t_z)\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.76, + 0.469, + 0.804 + ], + "angle": 0, + "content": "We also parameterize global, 9-DoF object poses, \\(\\bar{T}_o = (\\gamma_x,\\gamma_y,\\gamma_z,t_x,t_y,t_z,s_x,s_y,s_z)\\), comprising 6-DoF angles and translations, and 3-DoF anisotropic scales \\(s\\)." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.805, + 0.454, + 0.82 + ], + "angle": 0, + "content": "We formulate a joint energy optimization of the form:" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.829, + 0.47, + 0.848 + ], + "angle": 0, + "content": "\\[\nT ^ {*}, \\bar {T} ^ {*} = \\operatorname {a r g m i n} _ {T, \\bar {T}} \\left(\\mathrm {E} _ {c} (T, M) + \\mathrm {E} _ {o} (T, \\bar {T}, N)\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.47, + 0.902 + ], + "angle": 0, + "content": "where \\(M\\) are inter-frame feature matches and \\(N\\) are intra-frame canonical object correspondences established with normalized object coordinates (NOCs) that densely map" + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.311, + 0.878, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.486, + 0.892, + 0.584 + ], + "angle": 0, + "content": "Figure 3. Multi-modal object recognition and NOC prediction. Our ResNet-FPN [16, 21] backbone takes color, reversed jet colored depth [12], and \\(2\\mathrm{x}\\) downsampled colored 3D depth normals, and produces multi-scale features by averaging different input encodings. From the obtained features, our method recognizes objects and predicts NOCs for each object, based on a Mask-RCNN [15]-style prediction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.61, + 0.893, + 0.791 + ], + "angle": 0, + "content": "to the canonical space of an object in \\([-0.5, 0.5]^3\\). \\( \\mathrm{E}_c \\) is the feature-matching energy function, and \\( \\mathrm{E}_o \\) is our object-centric energy function. Since robust feature matching and optimization are readily available off the shelf [10, 11, 31, 36], our method focuses on building the object-centric energy \\( \\mathrm{E}_o \\). To this end, we need two function approximators, realized via deep neural networks: (1) a learned model for object recognition and NOC prediction, and (2) a learned model for object identification. The realization of these networks is described in Sections 3.2 and 3.3, respectively, and energy function Eq. 1 in Section 3.4. An overview of our approach is visualized in Figure 2." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.802, + 0.809, + 0.818 + ], + "angle": 0, + "content": "3.2. Predicting Object Correspondences" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "To obtain object constraints in our final optimization, we recognize objects via object detection and instance segmentation, and predict object correspondences as dense NOCs [43] for each object, as shown in Figure 3. We build on a Mask-RCNN [15] with ResNet50-FPN [16, 21] back" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13084" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.426, + 0.107 + ], + "angle": 0, + "content": "bone, pre-trained on ImageNet [33] and COCO [22]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.107, + 0.47, + 0.198 + ], + "angle": 0, + "content": "To input the depth of the RGB-D frames, we propose a modified, multi-modal ResNet50-FPN [12, 16, 21] backbone. Our backbone takes \\(480 \\times 640\\) color, \\(480 \\times 640\\) reverse jet-colored depth, and \\(240 \\times 320\\) colored depth normals as input. We average the resulting FPN features to a single feature pyramid:" + }, + { + "type": "equation", + "bbox": [ + 0.101, + 0.206, + 0.469, + 0.24 + ], + "angle": 0, + "content": "\\[\nG = \\frac {\\operatorname {F P N} ^ {c} \\left(I ^ {c}\\right) + \\operatorname {F P N} ^ {d} \\left(I ^ {d}\\right) + \\operatorname {U} \\left(\\operatorname {F P N} ^ {\\mathrm {n}} \\left(I ^ {n}\\right)\\right)}{3}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.248, + 0.469, + 0.384 + ], + "angle": 0, + "content": "where \\( I^{c}, I^{d}, I^{n} \\) are color, depth, and normal images, \\( \\mathrm{FPN}^{c}, \\mathrm{FPN}^{d}, \\mathrm{FPN}^{n} \\) are the corresponding ResNet50-FPN backbones, and U is an upsampling operator to match the normal features' spatial size with others. This enables fine-tuning the pre-trained bounding box, class, and instance segmentation heads of Mask-RCNN [15], while also exploiting depth information. We use symmetrically structured FPNs, all pre-trained on ImageNet and COCO as initialization, but without any parameter sharing." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.385, + 0.469, + 0.459 + ], + "angle": 0, + "content": "To obtain object correspondences, we establish mappings from detected objects to their canonical spaces, in the form of dense NOCs. That is, for each pixel in the object's instance mask, we predict 3D points \\( P_{o}^{\\mathrm{noc}} \\) using a fully convolutional network:" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.47, + 0.469, + 0.488 + ], + "angle": 0, + "content": "\\[\nP _ {o} ^ {\\text {n o c}} = \\operatorname {F C N} \\left(G _ {o}\\right), p \\in [ - 0. 5, 0. 5 ] ^ {3} \\forall p \\in P _ {o} ^ {\\text {n o c}}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.497, + 0.469, + 0.525 + ], + "angle": 0, + "content": "We optimize an \\(\\ell_1\\) loss \\(L_{\\mathrm{noc}}\\) using ground-truth NOCs \\(P^{\\mathrm{noc - gt}}\\)" + }, + { + "type": "equation", + "bbox": [ + 0.158, + 0.525, + 0.469, + 0.555 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {n o c}} = \\sum_ {o} \\sum_ {i} \\left\\| P _ {o, i} ^ {\\text {n o c}} - P _ {o} ^ {\\text {n o c - g t}} \\right\\| _ {1}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.56, + 0.469, + 0.68 + ], + "angle": 0, + "content": "Since symmetric objects [1, 43] induce ambiguities in NOCs (e.g., a round table), we classify symmetry type of objects (round, square, rectangle, non-symmetric), \\( c_{\\mathrm{sym}} = \\mathrm{MLP}^{\\mathrm{sym}}(G_o) \\), optimized using a cross-entropy loss \\( L_{\\mathrm{sym}} \\). We also make \\( L_{\\mathrm{noc}} \\) symmetry aware, taking the minimum \\( \\ell_1 \\) difference over the set of correct NOCs [43]. We use non-symmetric objects during inference to avoid inconsistent NOCs across views." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.681, + 0.469, + 0.771 + ], + "angle": 0, + "content": "In addition to NOCs, we also regress anisotropic 3D object scales \\( s_o \\) using a fully connected network, \\( s_o = \\mathrm{MLP}^{\\mathrm{scale}}(G_o) \\), and optimize \\( s_o \\) with an \\( \\ell_1 \\) loss \\( L_{\\mathrm{scale}} \\). The object scale enables holistic object pose understanding within each frame and helps to filter potential object matches across views using scale consistency." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.772, + 0.469, + 0.847 + ], + "angle": 0, + "content": "Finally, to make our NOC-depth correspondences least-squares friendly for our desired optimization, we also introduce a per-frame differentiable Procrustes objective \\( L_{\\mathrm{proc}} \\), using a differentiable Kabsch solver [32] to obtain local object rotations and translations:" + }, + { + "type": "equation", + "bbox": [ + 0.078, + 0.857, + 0.469, + 0.901 + ], + "angle": 0, + "content": "\\[\nR _ {o} ^ {*}, t _ {o} ^ {*} = \\operatorname {a r g m i n} _ {R _ {o}, t _ {o}} \\left(\\sum_ {i} \\left\\| R _ {o} \\left(P _ {o, i} ^ {\\text {n o c}} \\odot s _ {o}\\right) + t _ {o} - P _ {o, i} ^ {\\text {d e p t h}} \\right\\| _ {2} ^ {2}\\right) \\tag {5}\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.091, + 0.885, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.241, + 0.892, + 0.31 + ], + "angle": 0, + "content": "Figure 4. Our foreground/background metric-learning encoder for object matching, inspired by re-OBJ [3]. Using the detected and segmented objects from the model in Section 3.2, we encode foreground and background regions of objects, using light-weight, multi-modal ResNet18 encoders on the RGB-D features." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.335, + 0.892, + 0.398 + ], + "angle": 0, + "content": "for each object \\(o\\), where \\(P_{o}^{\\mathrm{depth}}\\) are back-projected input RGB-D depths corresponding to the object's predicted instance mask in its region of interest, and \\(\\odot\\) denotes element-wise multiplication. We train the local object poses with" + }, + { + "type": "equation", + "bbox": [ + 0.516, + 0.407, + 0.891, + 0.439 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {p r o c}} = w _ {r} \\sum_ {o} | | R _ {o} ^ {*} - R _ {o} ^ {\\mathrm {g t}} | | _ {1} + w _ {t} \\sum_ {o} | | t _ {o} ^ {*} - t _ {o} ^ {\\mathrm {g t}} | | _ {2} ^ {2}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.449, + 0.766, + 0.464 + ], + "angle": 0, + "content": "Our full loss used for training is then" + }, + { + "type": "equation", + "bbox": [ + 0.509, + 0.475, + 0.891, + 0.492 + ], + "angle": 0, + "content": "\\[\nL = L _ {m} + w _ {n} L _ {\\text {n o c}} + w _ {s} L _ {\\text {s c a l e}} + w _ {\\text {s y m}} L _ {\\text {s y m}} + w _ {p} L _ {\\text {p r o c}}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.503, + 0.891, + 0.533 + ], + "angle": 0, + "content": "where \\( L_{m} \\) is the sum of Mask-RCNN losses [15] and \\( w \\) are scalar weights balancing additional losses." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.537, + 0.892, + 0.642 + ], + "angle": 0, + "content": "Implementation. We use the augmented 400k ScanNet train image split for training [9, 14], with Scan2CAD labels of the most common 9 furniture categories following [1, 14, 25]. We train a standard Detector2 Mask-RCNN pipeline [15, 45] with 1k warm-up iterations, 0.003 base learning rate, and learning rate decays at 60k and 100k iterations with a total of 120k training iterations." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.651, + 0.744, + 0.668 + ], + "angle": 0, + "content": "3.3. Matching Object Instances" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.892, + 0.765 + ], + "angle": 0, + "content": "In our global pose optimization formulation, the relation between frames is formed via a global identification of objects across frames. To enable such identification without any heuristic spatial assumptions, we train a simple metric learner that ranks object semantic similarities across frames. Our object matching is shown in Figure 4." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.856 + ], + "angle": 0, + "content": "We characterize objects for matching by their respective RGB-D features in the object instance mask, in addition to the global context in the 5-times upscaled object bounding box. All inputs are resized to \\(224 \\times 224\\) and input to a lightweight ResNet18 [16] backbone pre-trained on ImageNet [33]." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Similar to object detection, we employ two backbones for color and colored depth inputs. We omit normal input in object matching, as it empirically did not provide any" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13085" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.138 + ], + "angle": 0, + "content": "benefit. For each input modality, we train two ResNet18 backbones for masked and inverse-masked crops, namely foreground (object) and background (context) encodings," + }, + { + "type": "equation", + "bbox": [ + 0.084, + 0.145, + 0.469, + 0.177 + ], + "angle": 0, + "content": "\\[\ne = \\operatorname {M L P} \\left(\\left[ \\mathrm {R N} ^ {c} \\left(F _ {c}\\right), \\mathrm {R N} ^ {c} \\left(B _ {c}\\right) \\right] + \\left[ \\mathrm {R N} ^ {d} \\left(F _ {d}\\right), \\mathrm {R N} ^ {d} \\left(B _ {d}\\right) \\right]\\right) \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.178, + 0.469, + 0.223 + ], + "angle": 0, + "content": "where RN are ResNet18s, MLP is a fully connected network, \\( F \\), \\( B \\) are foreground and background crops for color (c) and depth (d), and \\( e \\) is the object embedding vector." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.224, + 0.469, + 0.269 + ], + "angle": 0, + "content": "Given an anchor object embedding \\( e_a \\), a correctly matching object embedding \\( e_p \\), and a negative example \\( e_n \\), we train our metric learning using a triplet margin loss [17]:" + }, + { + "type": "equation", + "bbox": [ + 0.129, + 0.278, + 0.469, + 0.295 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {t r i}} = \\max \\left(\\mathrm {d} \\left(e _ {a}, e _ {p}\\right) - \\mathrm {d} \\left(e _ {a}, e _ {n}\\right) + 1. 0, 0\\right), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.305, + 0.253, + 0.319 + ], + "angle": 0, + "content": "where \\(d\\) is the \\(\\ell_2\\) distance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.32, + 0.469, + 0.44 + ], + "angle": 0, + "content": "We only consider triplets from the same category, as the object recognition pipeline provides classification. At inference time, we match the best instances using the Hungarian algorithm and apply a threshold \\( d(e_i, e_j) < \\alpha \\) for the matching object pairs from the same class. This semantic matching can be scaled to multiple frames via, e.g., object tracking with re-identification, or in our case, a simple pose graph optimization over frame pairs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.444, + 0.469, + 0.518 + ], + "angle": 0, + "content": "Implementation. We implement the identification network using PyTorch [30] and train it on ScanNet data [9]. We train the network for 100k iterations with a batch size of 8 using a momentum optimizer with a learning rate 1e-4 and momentum 0.9." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.528, + 0.276, + 0.544 + ], + "angle": 0, + "content": "3.4. Energy Optimization" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.551, + 0.469, + 0.58 + ], + "angle": 0, + "content": "We realize the joint energy minimization in Eq. 1 using keypoint and NOC constraints." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.581, + 0.469, + 0.612 + ], + "angle": 0, + "content": "Using the predicted NOC constraints with back-projected depths, we can re-write the \\( \\mathrm{E}_o \\) in Eq. 1 as:" + }, + { + "type": "equation", + "bbox": [ + 0.077, + 0.621, + 0.473, + 0.665 + ], + "angle": 0, + "content": "\\[\n\\mathrm {E} _ {o} (T, \\bar {T}, P ^ {\\text {d e p t h}}, P ^ {\\text {n o c}}) = \\sum_ {o} \\sum_ {c} \\sum_ {k} \\| T _ {c} P _ {o, c, k} ^ {\\text {d e p t h}} - \\bar {T} _ {o} P _ {o, c, k} ^ {\\text {n o c}} \\| _ {2} ^ {2} \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.665, + 0.469, + 0.77 + ], + "angle": 0, + "content": "where \\(\\bar{T}\\), \\(T\\) represent 9-DoF and 6-DoF camera and object transformations, respectively, and subscripts \\(o\\), \\(c\\), \\(k\\) correspond to objects, cameras (frames), and points (pixels) within the frames, respectively. Here, object indices are determined by the object identification, and global object poses \\(\\bar{T}\\) indirectly constrain frames to each other without any explicit inter-frame feature matching." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.772, + 0.469, + 0.862 + ], + "angle": 0, + "content": "In many cases, object constraints alone may not be sufficient to optimize the camera pose (e.g., frames may not share matching objects together). However, our object-based optimization is fully complementary to classical feature matching, and we thus formulate our objective in combination with feature-matching constraints \\(\\mathbf{E}_c\\):" + }, + { + "type": "equation", + "bbox": [ + 0.085, + 0.87, + 0.469, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\mathrm {E} _ {c} (T, P ^ {\\text {d e p t h}}) = \\sum_ {i} \\sum_ {j} \\sum_ {m, n} | | T _ {i} P _ {i, m} ^ {\\text {d e p t h}} - T _ {j} P _ {j, n} ^ {\\text {d e p t h}} | | _ {2} ^ {2} \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.167 + ], + "angle": 0, + "content": "Our method is agnostic to the feature matcher, whether classical or learning-based. In this work, we experiment with two different keypoint matching systems to realize \\( \\mathrm{E}_c \\), namely SuperGlue [11, 36] and Geometric Transformer [31], both offering state-of-the-art indoor feature matching." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.168, + 0.892, + 0.197 + ], + "angle": 0, + "content": "With both object and feature-matching constraints, we realize the desired joint energy formulation as" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.206, + 0.891, + 0.223 + ], + "angle": 0, + "content": "\\[\nT ^ {*}, \\bar {T} ^ {*} = \\operatorname {a r g m i n} _ {T, \\bar {T}} \\left(w _ {c} \\mathrm {E} _ {c} + w _ {o} \\mathrm {E} _ {o}\\right). \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.231, + 0.891, + 0.246 + ], + "angle": 0, + "content": "where \\(w_{c}, w_{o}\\) weight feature-matching, and object energies." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.246, + 0.892, + 0.412 + ], + "angle": 0, + "content": "Since non-linear least squares problems can be sensitive to outliers, we additionally employ outlier removal. Similar to BundleFusion [10], we apply Kabsch filtering to both intra-frame and keypoint constraints, using the matching depth-NOC and depth-depth correspondences, respectively. That is, we iteratively solve an orthogonal Procrustes problem and only keep the correspondences that have lower optimization errors. We use a liberal \\(20\\mathrm{cm}\\) threshold to handle wide-baseline frames. Objects and inter-frame matches are rejected if the number of NOCs is \\(< 15\\) and the number of keypoints is \\(< 5\\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.412, + 0.892, + 0.487 + ], + "angle": 0, + "content": "To solve this least-squares problem, we use a Gauss-Newton optimizer. To handle outliers during this global optimization, we remove \\(>15cm\\) error residuals during optimization. As global optimization can produce coarse results, we further apply an ICP refinement." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.488, + 0.892, + 0.591 + ], + "angle": 0, + "content": "To handle SLAM-style sequence registration, we use a state-of-the-art global pose graph optimization [5, 47] over pairwise frame registration, using a single hierarchy level for simplicity. However, our method could be scaled to multi-frame optimization via object tracking and subdivided into multiple hierarchy levels to handle very large-scale scenes." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.606, + 0.633, + 0.622 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.63, + 0.702, + 0.645 + ], + "angle": 0, + "content": "4.1. Pairwise Registration" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.653, + 0.892, + 0.744 + ], + "angle": 0, + "content": "Evaluation Dataset. We evaluate pairwise camera pose estimation results on 1569 challenging and diverse frame pairs from the validation set of ScanNet [9]. Unlike previous works [19, 36], our evaluation includes a significant number of low-overlap pairs, with over \\(15\\%\\) of pairs having \\(\\leq 10\\%\\) overlap." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Optimization. To establish object correspondences across frame pairs in Eq. 10, we use the top-1 matching object with embedding distance threshold 0.05. In the absence of both keypoint matches and object matches, we consider object matches with distance \\( < 0.15 \\). Only objects of the same class label and predicted maximum scale ratio \\( < 1.5 \\) can match. We determine the best matching object based on the number of NOC-depth constraints after the filtering described in Section 3.4. We combine it with keypoint constraints and refine the results with ICP." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13086" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.289 + ], + "angle": 0, + "content": "Classical Baselines. We compare our method against various hand-crafted feature-matching and registration baselines. These baselines include BundleFusion keypoint optimization with SIFT descriptors and Kabsch filtering [10, 23], (SIFT + BF), and two geometric global registration baselines, Fast Global Registration (Fast GR) [46] and Global Registration (GR) [5], based on approximate and exact RANSAC over FPFH [34] features, respectively. As a post-processing step, all methods are refined using ICP. We use Open3D [47] implementations for global registration baselines as well as ICP post-processing of all methods. Our custom implementation of SIFT + BF uses the SIFT implementation of OpenCV [4]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.295, + 0.473, + 0.522 + ], + "angle": 0, + "content": "Learned Baselines. We also compare with pose estimators using learned models for feature or object matching: Geometric Transformer (GeoTrans) [31], which performs dense geometric feature matching, and SG + BF which leverages the powerful learned SuperGlue [36] and SuperPoint [11] feature matching approaches in combination with BundleFusion [10] Kabsch filtering and 3D Gauss-Newton optimization. We also use our network predictions to create a 3D object tracking baseline, Object Track, mimicking tracking-based object-SLAM systems. Object Track obtains local object poses and object matches from Sections 3.2 and 3.3, but instead uses relative local object poses to get the camera pose instead of a global energy optimization. All methods are refined using ICP as a post-processing step." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.526, + 0.472, + 0.634 + ], + "angle": 0, + "content": "Evaluation Metrics. We use the Pose Recall metric for pose evaluation, following previous RGB-D pairwise registration works [6, 19]. To comprehensively capture performance, we employ several thresholds for absolute translation (in cm) and rotation (in °). Since absolute translation is more difficult than rotation, its thresholds are twice the angle thresholds, following [6]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.638, + 0.473, + 0.88 + ], + "angle": 0, + "content": "Quantitative Results. In Table 1 and Table 2, we evaluate on ScanNet [9] validation frame pairs, measuring recall in various thresholds and overlap levels. Combined with state-of-the-art feature matching in a joint optimization, our method outperforms both classical and learning-based methods. Furthermore, the gap increases with lower overlap, as shown in Table 2 and Figure 6, since our method can leverage object-based correspondences to estimate alignment in the absence of keypoint matches. We also show that our global optimization (Ours (w/o keypoints)) significantly outperforms naive object-based tracking (Object Track), demonstrating the efficacy of our global optimization formulation. Our approach is designed to complement keypoint matching, notably improving the state of the art in combination with state-of-the-art geometric and RGB-D feature matching." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.886, + 0.472, + 0.902 + ], + "angle": 0, + "content": "Analysis of Performance by Overlap. In Table 2 and" + }, + { + "type": "image_caption", + "bbox": [ + 0.569, + 0.092, + 0.701, + 0.105 + ], + "angle": 0, + "content": "NOCs of Top-1 Match" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.108, + 0.765, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.183, + 0.764, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.259, + 0.764, + 0.333 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.335, + 0.763, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.411, + 0.763, + 0.485 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.487, + 0.764, + 0.561 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.563, + 0.764, + 0.637 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.639, + 0.764, + 0.712 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.778, + 0.092, + 0.855, + 0.106 + ], + "angle": 0, + "content": "Registration" + }, + { + "type": "image", + "bbox": [ + 0.766, + 0.108, + 0.871, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.764, + 0.186, + 0.871, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.766, + 0.259, + 0.866, + 0.333 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.767, + 0.338, + 0.859, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.767, + 0.415, + 0.868, + 0.485 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.766, + 0.495, + 0.866, + 0.561 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.767, + 0.564, + 0.877, + 0.636 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.775, + 0.644, + 0.877, + 0.712 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.718, + 0.894, + 0.775 + ], + "angle": 0, + "content": "Figure 5. Low-overlap registration on ScanNet [9], where traditional feature matching fails. Predicted NOC correspondences are visualized, along with object box and camera poses of the left and right images in green and blue, respectively." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.81, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Figure 6, we show \\(30\\mathrm{cm}\\), \\(15^{\\circ}\\) recall performance at different geometric overlap rates, using the radius-based geometric overlap percentage measurement from [19] with \\(1\\mathrm{cm}\\) threshold. While all methods' performances decrease with overlap, our method retains significantly more robustness. Classical keypoint matching (SIFT + BF, Fast GR," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13087" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.091, + 0.089, + 0.454, + 0.237 + ], + "angle": 0, + "content": "
MethodPose Recall by Threshold
5°, 10cm10°, 20cm15°, 30cm
SIFT + BF [10, 23]16.8318.4220.40
Fast GR [34, 46]39.0742.9644.68
GR [5]43.0847.6750.03
Object Track37.4143.2145.19
GeoTrans [31]67.1174.5776.93
SG + BF [10, 11, 36]71.1380.1181.71
Ours (w/o keypoints)56.4163.6766.99
Ours (w/ GeoTrans)68.3276.8079.35
Ours (w/ SG + BF)74.2584.5887.06
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.24, + 0.472, + 0.325 + ], + "angle": 0, + "content": "Table 1. Pose Recall results on ScanNet [9] validation images. Combined with state-of-the-art feature matching, our method outperforms various classical and learning-based baselines. Our approach complements both Geometric Transformer [31] (Ours (w/ GeoTrans)) and SuperGlue [36] (Ours (w/ SG + BF)) feature matches, notably improving pose recall." + }, + { + "type": "table", + "bbox": [ + 0.111, + 0.338, + 0.434, + 0.496 + ], + "angle": 0, + "content": "
MethodRecall by Overlap %
≤ 10(10,30)≥ 30
SIFT + BF [10,23]1.302.2033.59
Fast GR [34,46]0.0013.9769.04
GR [5,34]0.6520.3674.62
Object Track8.4431.5458.86
GeoTrans [31]22.0864.2793.11
SG + BF [10,11,36]24.0373.4595.95
Ours (w/o keypoints)32.4753.2980.31
Ours (w/ GeoTrans)35.0665.4794.42
Ours (w/ SG + BF)45.4581.4497.16
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.5, + 0.472, + 0.543 + ], + "angle": 0, + "content": "Table 2. Recall at \\(15^{\\circ}\\), \\(30\\mathrm{cm}\\) by overlap percentage. Our method significantly outperforms strong baselines in challenging low-overlap frame pairs, almost doubling recall for overlap \\(\\leq 10\\%\\)." + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.559, + 0.441, + 0.754 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.757, + 0.47, + 0.811 + ], + "angle": 0, + "content": "Figure 6. Change of pose performance by overlap percentage on ScanNet [9] validation pairs. Our method is significantly more robust against decreasing overlap compared to various classical and learned baselines." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.903 + ], + "angle": 0, + "content": "SIFT) suffer strongly in low-overlap scenarios, which are challenging for hand-crafted descriptors. The learned approaches of GeoTrans and SG+BF maintain some performance with decreasing overlap but still suffer strongly. Our" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "incorporation of object-based reasoning effectively complements keypoint matching while significantly improving robustness to low overlap, obtaining a performance improvement from \\(24.03\\%\\) to \\(45.45\\%\\) for overlap \\(\\leq 10\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.157, + 0.892, + 0.248 + ], + "angle": 0, + "content": "Qualitative Results. Figure 5 shows examples of low-overlap registration on ScanNet [9] frame pairs. Due to the minimal overlap, traditional keypoint matching cannot find sufficient correspondences, while our object grounding enables accurate camera pose estimation through indirect object constraints." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.256, + 0.854, + 0.273 + ], + "angle": 0, + "content": "4.2. Registration of RGB-D SLAM Sequences" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.28, + 0.892, + 0.354 + ], + "angle": 0, + "content": "We further demonstrate the effectiveness of our method in an RGB-D SLAM setting, registering sequences of frames in room-scale scenes. In particular, we evaluate challenging, low-frame-rate settings that reflect common non-expert capture such as fast camera motion." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.355, + 0.892, + 0.431 + ], + "angle": 0, + "content": "For optimization, we use a global pose graph optimization over pairwise registration similar to the multi-way global registration of [5, 6, 47], as it offers an off-the-shelf, simple, and robust solution. We refer to our supplementary material for further optimization details." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.436, + 0.892, + 0.587 + ], + "angle": 0, + "content": "Datasets and Evaluation Metrics. We evaluate our method on TUM-RGBD [39] and ScanNet [9] scenes. For ScanNet, we use a set of 12 scenes from the validation and test sets, having a wide range of sizes, environments, and camera trajectories. We sample every 30th frame for TUM-RGBD (1Hz) and every 20th frame in ScanNet (1.5Hz). All methods use the same hyperparameters for each dataset, respectively. To evaluate registration quality, we use the standard root mean squared trajectory error used for TUM-RGBD evaluation [39]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.593, + 0.892, + 0.773 + ], + "angle": 0, + "content": "Baselines. We implement various popular feature descriptors from cutting-edge SLAM systems in global pose graph optimization setups and thoroughly tune the hyperparameters of each method to support low-frame-rate sequences. We compare against Redwood [5], a global multi-way pose graph that uses FPFH [34] features. We also deploy robust RGB-D SIFT features used in BundleFusion [10, 23] in a global pose graph optimization, creating the SIFT + BF baseline. Finally, we use the state-of-the-art SuperGlue matches of SuperPoint features with BundleFusion's sparse 3D optimization, SG + BF, as described in Section 4.1." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.9 + ], + "angle": 0, + "content": "Quantitative Results. In Table 3, we compare our method to state-of-the-art traditional and learned SLAM approaches. Across both TUM-RGBD and ScanNet validation and test scenes [9, 39], our object-grounded approach enables improved camera pose estimation. In particular, for the larger-scale ScanNet scenes, ObjectMatch achieves notable reductions in trajectory error, of \\(38\\%\\), \\(49\\%\\), \\(50\\%\\), and \\(44\\%\\) in scenes 0423, 0430, 0718, and 0773." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "13088" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.089, + 0.885, + 0.149 + ], + "angle": 0, + "content": "
MethodTUM Scene @ 1HzScanNet Scene @ 1.5Hz
fr1/deskfr1/xyzfr2/xyzfr2/deskfr3/office001100810169020704230430046104940718076007730805
SIFT + BF [10,23]30.603.492.73157.0238.01222.49108.03126.57104.64160.69248.0020.9033.9881.4838.0322.4786.80
Redwood [5,34,47]3.101.971.88140.17175.64164.0150.9647.2447.1788.43156.7610.479.77133.536.5543.19253.39
SG + BF [10,11,36]3.011.921.815.305.1413.886.778.095.627.6735.973.317.1613.426.3712.4312.98
Ours (w/ SG + BF)3.011.921.815.264.9212.535.736.855.324.7518.232.676.856.705.887.0010.86
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.154, + 0.893, + 0.197 + ], + "angle": 0, + "content": "Table 3. ATE RMSE values (cm) on room-scale TUM-RGBD [39] and ScanNet [9] test scenes. We evaluate at challenging 1 FPS and 1.5 FPS by sampling every 30th and 20th frame from TUM-RGBD and ScanNet, respectively. ObjectMatch outperforms strong baselines leveraging classical and learning-based SLAM systems." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.221, + 0.462, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.548, + 0.47, + 0.66 + ], + "angle": 0, + "content": "Figure 7. Qualitative comparison for registration of ScanNet scenes [9], in comparison with only feature matching SG + BF [10, 11, 36]. All results are visualized by volumetric fusion of the respective RGB-D frames with the estimated poses. Our optimized poses produce more accurate, clean, and consistent reconstruction, due to incorporating the object-based information in our global optimization. Example loop closures with matching object correspondences are shown on the right." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.69, + 0.47, + 0.826 + ], + "angle": 0, + "content": "Qualitative Results. In Figure 7, we show ScanNet [9] scene reconstructions using estimated camera poses. RGB-D frames are fused using a scalable TSDF volumetric fusion [8,47]. Our approach complements state-of-the-art feature matching [10, 11, 36], enabling more accurate and consistent reconstructions using our indirect correspondences, in comparison to feature matching alone. We refer to the supplemental for additional qualitative sequence registration results." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Limitations and Future Work. ObjectMatch shows the capability of incorporating object semantics into global pose optimization; however, various limitations remain. In particular, our approach leverages higher-level seman" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.222, + 0.895, + 0.465 + ], + "angle": 0, + "content": "tic correspondence given by objects for registration; however, in scenarios where object matches do not exist in frame views, we cannot leverage this constraint and instead use only feature matches. Further leveraging background information regarding shared structures seen from different views would provide additional semantic correspondence information to further complement registration. Additionally, while estimating canonical correspondences with objects can provide significant view information, object alignments can be somewhat coarse, while reasoning at a finer-grained level of object parts would provide more precise reasoning. Finally, we believe future work can reduce our method's system complexity by adopting a joint multiframe SLAM optimization with fewer hyperparameters as well as a unified end-to-end architecture for learning object correspondences and identification together." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.477, + 0.619, + 0.492 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.502, + 0.892, + 0.775 + ], + "angle": 0, + "content": "We have presented ObjectMatch which introduces a new paradigm for incorporating object semantics into camera pose registration. We leverage indirect, canonical object correspondences established with normalized object coordinates for camera pose optimization aspects in RGB-D registration and SLAM. To obtain these correspondences, we propose two multi-modal neural networks for object recognition with normalized object coordinate prediction and object identification. ObjectMatch operates in tandem with state-of-the-art feature matching in a joint Gauss-Newton optimization, with its object grounding enabling registration to handle frame pairs with very low to no shared visibility. As a result, ObjectMatch significantly improves the state-of-the-art feature matching in both pairwise and sequence registration of RGB-D frames, particularly in the challenging low-overlap regime. Overall, we hope our method opens up new possibilities in the context of leveraging semantic information for camera pose estimation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Acknowledgements. This work was supported by the ERC Starting Grant Scan2CAD (804724), the Bavarian State Ministry of Science and the Arts coordinated by the Bavarian Research Institute for Digital Transformation (bidt), the German Research Foundation (DFG) Grant \"Making Machine Learning on Static and Dynamic 3D Data Practical\", and the German Research Foundation (DFG) Research Unit \"Learning and Simulation in Visual Computing.\"" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13089" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Armen Avetisyan, Manuel Dahnert, Angela Dai, Manolis Savva, Angel X Chang, and Matthias Nießner. Scan2cad: Learning cad model alignment in rgb-d scans. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 2614-2623, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.185, + 0.472, + 0.24 + ], + "angle": 0, + "content": "[2] Armen Avetisyan, Angela Dai, and Matthias Nießner. End-to-end cad model retrieval and 9dof alignment in 3d scans. In Proceedings of the IEEE/CVF International Conference on computer vision, pages 2551-2560, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.242, + 0.47, + 0.309 + ], + "angle": 0, + "content": "[3] Vaibhav Bansal, Stuart James, and Alessio Del Bue. re-object: Jointly learning the foreground and background for object instance re-identification. In International Conference on Image Analysis and Processing, pages 402–413. Springer, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.311, + 0.472, + 0.339 + ], + "angle": 0, + "content": "[4] G. Bradski. The OpenCV Library. Dr. Dobb's Journal of Software Tools, 2000. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.341, + 0.47, + 0.395 + ], + "angle": 0, + "content": "[5] Sungjoon Choi, Qian-Yi Zhou, and Vladlen Koltun. Robust reconstruction of indoor scenes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5556-5565, 2015. 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.397, + 0.47, + 0.451 + ], + "angle": 0, + "content": "[6] Christopher Choy, Wei Dong, and Vladlen Koltun. Deep global registration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2514-2523, 2020. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.453, + 0.469, + 0.479 + ], + "angle": 0, + "content": "[7] Christopher Choy, Jaesik Park, and Vladlen Koltun. Fully convolutional geometric features. In ICCV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.481, + 0.47, + 0.536 + ], + "angle": 0, + "content": "[8] Brian Curless and Marc Levoy. A volumetric method for building complex models from range images. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, pages 303-312. ACM, 1996. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.538, + 0.47, + 0.606 + ], + "angle": 0, + "content": "[9] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 2, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.607, + 0.47, + 0.675 + ], + "angle": 0, + "content": "[10] Angela Dai, Matthias Nießner, Michael Zollhöfer, Shahram Izadi, and Christian Theobalt. Bundlefusion: Real-time globally consistent 3d reconstruction using on-the-fly surface reintegration. ACM Transactions on Graphics (ToG), 36(4):1, 2017. 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.677, + 0.47, + 0.745 + ], + "angle": 0, + "content": "[11] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 224-236, 2018. 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.747, + 0.47, + 0.816 + ], + "angle": 0, + "content": "[12] Andreas Eitel, Jost Tobias Springenberg, Luciano Spinello, Martin Riedmiller, and Wolfram Burgard. Multimodal deep learning for robust rgb-d object recognition. In 2015 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 681-687. IEEE, 2015. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.817, + 0.47, + 0.859 + ], + "angle": 0, + "content": "[13] Hugo Germain, Vincent Lepetit, and Guillaume Bourmaud. Visual correspondence hallucination. In International Conference on Learning Representations, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[14] Can Gümeli, Angela Dai, and Matthias Nießner. Roca: Robust cad model retrieval and alignment from a single image. In Proceedings of the IEEE/CVF Conference on Computer" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.119 + ], + "angle": 0, + "content": "Vision and Pattern Recognition, pages 4022-4031, 2022. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.121, + 0.894, + 0.175 + ], + "angle": 0, + "content": "[15] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.177, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.233, + 0.892, + 0.274 + ], + "angle": 0, + "content": "[17] Elad Hoffer and Nir Ailon. Deep metric learning using triplet network. In International workshop on similarity-based pattern recognition, pages 84-92. Springer, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.275, + 0.892, + 0.342 + ], + "angle": 0, + "content": "[18] Jiahui Huang, Shi-Sheng Huang, Haoxuan Song, and ShiMin Hu. Di-fusion: Online implicit 3d reconstruction with deep priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8932-8941, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.344, + 0.892, + 0.413 + ], + "angle": 0, + "content": "[19] Shengyu Huang, Zan Gojcic, Mikhail Usvyatsov, Andreas Wieser, and Konrad Schindler. Predator: Registration of 3d point clouds with low overlap. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4267-4276, June 2021. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.414, + 0.892, + 0.51 + ], + "angle": 0, + "content": "[20] Shahram Izadi, David Kim, Otmar Hilliges, David Molyneaux, Richard Newcombe, Pushmeet Kohli, Jamie Shotton, Steve Hodges, Dustin Freeman, Andrew Davison, et al. Kinectfusion: real-time 3d reconstruction and interaction using a moving depth camera. In Proceedings of the 24th annual ACM symposium on User interface software and technology, pages 559-568, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.511, + 0.892, + 0.58 + ], + "angle": 0, + "content": "[21] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2117-2125, 2017. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.581, + 0.892, + 0.649 + ], + "angle": 0, + "content": "[22] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.65, + 0.892, + 0.705 + ], + "angle": 0, + "content": "[23] David G Lowe. Object recognition from local scale-invariant features. In Proceedings of the seventh IEEE international conference on computer vision, volume 2, pages 1150-1157. IEEE, 1999. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.706, + 0.894, + 0.775 + ], + "angle": 0, + "content": "[24] Wei-Chiu Ma, Anqi Joyce Yang, Shenlong Wang, Raquel Urtasun, and Antonio Torralba. Virtual correspondence: Humans as a cue for extreme-view geometry. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15924-15934, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.776, + 0.892, + 0.831 + ], + "angle": 0, + "content": "[25] Kevis-Kokitsi Maninis, Stefan Popov, Matthias Niesser, and Vittorio Ferrari. Vid2cad: Cad model alignment using multiview constraints from videos. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[26] Yuhang Ming, Xingrui Yang, and Andrew Calway. Object-augmented rgb-d slam for wide-disparity relocalisation. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 2203-2209. IEEE, 2021. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.894, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13090" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.161 + ], + "angle": 0, + "content": "[27] Norman Muller, Yu-Shiang Wong, Niloy J Mitra, Angela Dai, and Matthias Nießner. Seeing behind objects for 3d multi-object tracking in rgb-d sequences. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6071-6080, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.164, + 0.472, + 0.219 + ], + "angle": 0, + "content": "[28] Raul Mur-Artal and Juan D Tardós. Orb-slam2: An open-source slam system for monocular, stereo, and rgb-d cameras. IEEE transactions on robotics, 33(5):1255-1262, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.221, + 0.472, + 0.276 + ], + "angle": 0, + "content": "[29] Matthias Nießner, Michael Zollhöfer, Shahram Izadi, and Marc Stamminger. Real-time 3d reconstruction at scale using voxel hashing. ACM Transactions on Graphics (ToG), 32(6):1-11, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.278, + 0.472, + 0.403 + ], + "angle": 0, + "content": "[30] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems 32, pages 8024-8035. Curran Associates, Inc., 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.405, + 0.472, + 0.473 + ], + "angle": 0, + "content": "[31] Zheng Qin, Hao Yu, Changjian Wang, Yulan Guo, Yuxing Peng, and Kai Xu. Geometric transformer for fast and robust point cloud registration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11143-11152, 2022. 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.476, + 0.472, + 0.53 + ], + "angle": 0, + "content": "[32] Nikhila Ravi, Jeremy Reizenstein, David Novotny, Taylor Gordon, Wan-Yen Lo, Justin Johnson, and Georgia Gkioxari. Accelerating 3d deep learning with pytorch3d. arXiv:2007.08501, 2020.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.532, + 0.472, + 0.601 + ], + "angle": 0, + "content": "[33] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115(3):211-252, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.603, + 0.472, + 0.659 + ], + "angle": 0, + "content": "[34] Radu Bogdan Rusu, Nico Blodow, and Michael Beetz. Fast point feature histograms (fpfh) for 3d registration. In 2009 IEEE international conference on robotics and automation, pages 3212-3217. IEEE, 2009. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.661, + 0.472, + 0.73 + ], + "angle": 0, + "content": "[35] Renato F Salas-Moreno, Richard A Newcombe, Hauke Strasdat, Paul HJ Kelly, and Andrew J Davison. Slam++: Simultaneous localisation and mapping at the level of objects. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1352-1359, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.732, + 0.472, + 0.8 + ], + "angle": 0, + "content": "[36] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4938–4947, 2020. 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.472, + 0.858 + ], + "angle": 0, + "content": "[37] Thomas Schops, Torsten Sattler, and Marc Pollefeys. Bad slam: Bundle adjusted direct rgb-d slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 134-144, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.472, + 0.901 + ], + "angle": 0, + "content": "[38] Yifei Shi, Kai Xu, Matthias Niessner, Szymon Rusinkiewicz, and Thomas Funkhouser. Planematch: Patch coplanarity prediction for robust rgb-d reconstruction. In Proceedings of the" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "European Conference on Computer Vision (ECCV), pages 750-766, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.894, + 0.177 + ], + "angle": 0, + "content": "[39] J. Sturm, N. Engelhard, F. Endres, W. Burgard, and D. Cremers. A benchmark for the evaluation of rgb-d slam systems. In Proc. of the International Conference on Intelligent Robot Systems (IROS), Oct. 2012. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.234 + ], + "angle": 0, + "content": "[40] Edgar Sucar, Shikun Liu, Joseph Ortiz, and Andrew J Davison. imap: Implicit mapping and positioning in real-time. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6229-6238, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.235, + 0.892, + 0.29 + ], + "angle": 0, + "content": "[41] Edgar Sucar, Kentaro Wada, and Andrew Davison. NodeSLAM: Neural object descriptors for multi-view shape reconstruction. In Proceedings of the International Conference on 3D Vision (3DV), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.292, + 0.892, + 0.36 + ], + "angle": 0, + "content": "[42] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8922-8931, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.362, + 0.892, + 0.445 + ], + "angle": 0, + "content": "[43] He Wang, Srinath Sridhar, Jingwei Huang, Julien Valentin, Shuran Song, and Leonidas J Guibas. Normalized object coordinate space for category-level 6d object pose and size estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2642-2651, 2019. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.446, + 0.892, + 0.488 + ], + "angle": 0, + "content": "[44] T Whelan, S Leutenegger, RF. Salas-Moreno, B Glocker, and AJ. Davison. Elasticfusion: Dense slam without a pose graph. In Robotics: Science and Systems, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.489, + 0.892, + 0.53 + ], + "angle": 0, + "content": "[45] Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. Detector2. https://github.com/facebookresearch/detectron2, 2019.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.532, + 0.892, + 0.573 + ], + "angle": 0, + "content": "[46] Qian-Yi Zhou, Jaesik Park, and Vladlen Koltun. Fast global registration. In European conference on computer vision, pages 766-782. Springer, 2016. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.575, + 0.892, + 0.615 + ], + "angle": 0, + "content": "[47] Qian-Yi Zhou, Jaesik Park, and Vladlen Koltun. Open3D: A modern library for 3D data processing. arXiv:1801.09847, 2018. 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.617, + 0.892, + 0.699 + ], + "angle": 0, + "content": "[48] Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hujun Bao, Zhaopeng Cui, Martin R Oswald, and Marc Pollefeys. Nice-slam: Neural implicit scalable encoding for slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12786-12796, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.699 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "13091" + } + ] +] \ No newline at end of file diff --git a/2023/ObjectMatch_ Robust Registration Using Canonical Object Correspondences/9d0104e4-40f2-4a9d-bf9d-8065e49cc737_origin.pdf b/2023/ObjectMatch_ Robust Registration Using Canonical Object Correspondences/9d0104e4-40f2-4a9d-bf9d-8065e49cc737_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4cd61931a955839383b7ba9caedebe5d2b1f97bd --- /dev/null +++ b/2023/ObjectMatch_ Robust Registration Using Canonical Object Correspondences/9d0104e4-40f2-4a9d-bf9d-8065e49cc737_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:902bdaa207dba06290b69257102654466fdf39ac548a833038cbded703c726f1 +size 4109012 diff --git a/2023/ObjectMatch_ Robust Registration Using Canonical Object Correspondences/full.md b/2023/ObjectMatch_ Robust Registration Using Canonical Object Correspondences/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c83b07e27481fe368fada20082ec60d08dae4b10 --- /dev/null +++ b/2023/ObjectMatch_ Robust Registration Using Canonical Object Correspondences/full.md @@ -0,0 +1,359 @@ +# ObjectMatch: Robust Registration using Canonical Object Correspondences + +Can Gumeli + +Angela Dai + +Matthias Nießner + +# Technical University of Munich + +![](images/d3f06c5d45558df7016439e082690d03636262a3120456f1cc122b9de7dc3222.jpg) +Low-overlap Frames with Matching Objects + +![](images/2c3cef79c96c5e79dbb46a94649ef866e5264b69ef2f493c581ba90d9950a3c4.jpg) + +![](images/4357d973bceeaaad23d31aa645eee790d5fe4d7bf49d5a225e95cfd010db473f.jpg) + +![](images/17c4fb6bd2b65901cfbd33acbe69dcaae383bb72f40ea24e4f9fe3fe8e95608f.jpg) +Canonical Correspondences + +![](images/fba06f9104b57c26157b6040acc7258a375f26936281e12abbee5750171b2a18.jpg) +Camera and Object Poses +Figure 1. Modern camera pose estimation relies on feature matching between overlapping frames - in this work, we present ObjectMatch to find correspondences between frames with little or no overlap by predicting semantic mappings through canonical object correspondences. The images above share no direct overlap, yet our method establishes indirect correspondences, thus enabling a successful registration. + +# Abstract + +We present ObjectMatch1, a semantic and object-centric camera pose estimator for RGB-D SLAM pipelines. Modern camera pose estimators rely on direct correspondences of overlapping regions between frames; however, they cannot align camera frames with little or no overlap. In this work, we propose to leverage indirect correspondences obtained via semantic object identification. For instance, when an object is seen from the front in one frame and from the back in another frame, we can provide additional pose constraints through canonical object correspondences. We first propose a neural network to predict such correspondences on a per-pixel level, which we then combine in our energy formulation with state-of-the-art keypoint matching solved with a joint Gauss-Newton optimization. In a pairwise setting, our method improves registration recall of state-of-the-art feature matching, including from $24\%$ to $45\%$ in pairs with $10\%$ or less inter-frame overlap. In registering RGB-D sequences, our method outperforms cutting-edge SLAM baselines in challenging, low-frame-rate scenarios, achieving more than $35\%$ reduction in trajectory error in multiple scenes. + +# 1. Introduction + +RGB-D registration and 3D SLAM has been a fundamental task in computer vision, with significant study and enabling many applications in mixed reality, robotics, and content creation. Central to both state-of-the-art traditional and learning-based camera pose estimation is establishing correspondences between points in input frames. However, correspondence estimation remains quite challenging when there is little or no overlap between frames. + +In contrast, humans can easily localize across these challenging scenarios by leveraging additional semantic knowledge – in particular, by further localizing at the level of objects and identifying matching objects between views. For instance, when observing a chair from the back and the side (e.g., in Figure 1), view overlap is minimal (or even no view overlap), resulting in failed registration from keypoint matching. However, the semantic knowledge of the chair and its object pose nonetheless enables humans to estimate the poses from which the front and side views were taken. Thus, we propose to take a new perspective on camera pose estimation and imbue camera registration with awareness of such semantic correspondences between objects for robust performance in these challenging scenarios. + +To this end, we propose ObjectMatch, a new paradigm for camera pose estimation leveraging canonical object correspondences in tandem with local keypoint correspondences between views. This enables significantly more robust registration under a variety of challenging scenar- + +ios, including low view overlap. For a sequence of input frames, ObjectMatch learns to semantically identify objects across frames, enabling a compact, global parameterization of 9-DoF object poses. Object correspondences are established through predicting normalized object coordinates [43], dense correspondences from object pixels to a canonically oriented space for each object. We then formulate a joint camera and object pose optimization that constrains object correspondences indirectly, operating irrespective of the shared visibility of image regions. Our approach is complementary to state-of-the-art SLAM methods, and we leverage our energy formulation to complement state-of-the-art keypoint matching [10, 11, 36] in a joint Gauss-Newton optimization. + +Our method outperforms strong baselines in both pairwise registration and registration of RGB-D frame sequences. In pairwise registration of challenging ScanNet [9] image pairs, we improve pose recall from $24\%$ to $45\%$ when the overlap is below $10\%$ . On sequence registration of room-scale RGB-D scenes, our method outperforms various strong baselines in difficult, low-frame-rate settings in several TUM-RGBD [39] and ScanNet [9] scenes, reducing the trajectory error by more than $35\%$ in multiple challenging scenes. + +To sum up, our main contributions include: + +- An object-centric camera pose estimator that can handle low-overlap frame sets via indirect, canonical object correspondences established with predicted dense, per-pixel normalized object coordinates. +- A joint energy formulation that leverages semantic object identification and dense, normalized object coordinates corresponding to canonical object geometries. +- Our semantic grounding of object correspondences enables significantly more robust registration in low-overlap and low-frame-rate cases. ObjectMatch improves over state of the art from $24\%$ to $45\%$ registration recall of $\leq 10\%$ overlap frame pairs and achieves over $35\%$ trajectory error reduction in several challenging sequences. + +# 2. Related Work + +RGB-D Registration and SLAM. In recent years, there have been many advances in indoor RGB-D reconstruction. Earlier RGB-D fusion approaches focus on frame-to-model camera tracking [20, 29]. To handle the loop closures better, more recent SLAM systems introduce explicit strategies or global optimization methods for handling loop closures through global optimization [5,10,28,37,44] to fix tracking errors. More recently, deep learning techniques have been applied to registration and SLAM scenarios, with methods ranging from geometric point cloud registration [6, 19, 31] as well as neural field based SLAM techniques [18, 40, 48]. + +Despite all the successes in RGB-D SLAM and registration, the task is still challenging since incomplete loop closures observed via low-overlap frames cannot be handled, and most SLAM methods require a very high overlap between consecutive frames to track cameras accurately. + +Feature Matching. Modern RGB(-D) camera pose estimators rely on a feature-matching backbone. Classical global registration techniques [5,46] use FPFH features [34] over point cloud fragments. On the other hand, many global RGB-D SLAM techniques rely on sparse color features [10, 28]. While being successful in many scenarios, conventional feature matching often fails when the interframe overlap is low. Therefore, deep learning techniques have been utilized for predicting overlapping regions based on geometry or color. On the geometric side, Deep Global Registration [6] predicts overlapping point features using nearest neighbor search over learned geometric features [7]. Methods such as PREDATOR [19] and Geometric Transformer [31] use attention mechanisms to target overlapping regions for registration. In the domain of color features, SuperPoint and SuperGlue [11, 36] build a formative approach in GNN-based keypoint feature matching. Methods such as LoFTR [42] introduce more dense and accurate sub-pixel level matching. Despite being very successful in handling wide-baseline scenarios, learned feature matching still requires a significant amount of shared visibility and geometric overlap. + +Camera Pose Estimation with Semantic Cues. Several methods have been developed to incorporate semantic priors to improve low-overlap registration. PlaneMatch [38] proposed coplanarity priors for handling loop closures and low-overlap cases. Our method instead leverages object-centric constraints, exploiting the power of semantic object recognition. Another related direction is feature hallucination by leveraging image and object semantics. NeurHal [13] focuses on correspondence hallucination using image inpainting and outpainting, formulating a PnP optimization over hallucinated matches. Virtual Correspondence (VC) [24] introduces a human-centric approach that leverages hallucinated object volumes to form virtual epipolar constraints between frames. In contrast, we use indirect instead of direct correspondences that do not require hallucinated object volumes or image regions. Furthermore, our method works on a diverse set of furniture categories while VC focuses on humans. Pioneered by SLAM++ [35], there is also a rich literature of object-centric SLAM solutions, e.g., [26, 41]. Such SLAM methods leverage local, per-frame poses of objects to establish constraints; instead, we develop a global object pose optimization that is more robust against occluded viewpoints. + +Object Pose Estimation using Normalized Object Coordinates. 3D object pose estimation has been widely stud + +![](images/5fd461f8ba293bf3aa6fcb79e1190e4e440fb2ab3150d5e8e43c1ff99c53e9de.jpg) +Figure 2. Overview of our approach to incorporate object correspondence grounding in global pose estimation. From a set of input RGB-D frames, ObjectMatch predicts object instances for each frame with dense normalized object correspondences. The predicted object instances are used to identify objects across frames, forming indirect object correspondences. We combine object correspondences with SuperGlue [11,36] keypoint matches in a joint energy optimization that yields both camera and object poses in a global registration. + +ied from RGB and RGB-D inputs. Normalized Object Coordinate Space (NOCS) [43] was proposed to form dense correspondences from input RGB-D frames to canonical object geometries, enabling better generalization than direct regression of object poses. End2End SOCs [2] formulated a NOC-based approach for CAD retrieval and alignment to 3D scans, using a differentiable Procrustes algorithm. To enable CAD alignment to single RGB images, ROCA [14] leveraged NOC space in combination with predicted depths, formulating a robust differentiable Procrustes optimization [14]. Seeing Behind Objects [27] further leveraged NOC correspondences both to obtain local object poses and object completion for re-identification for RGB-D multi-object tracking. Wide Disparity Re-localization [26] uses the NOC predictions from [43] to construct an object-level map for re-localization in SLAM. In contrast to these approaches that focus on individual object poses, we use NOC correspondences directly in a multi-frame, global camera, and object pose optimization. + +# 3. Method + +# 3.1. Problem Setup + +Given $K$ RGB-D frames $\{(I_1^c, I_1^d), \dots, (I_K^c, I_K^d)\}$ , we aim to optimize their 6-DoF camera poses $T_c = \{T_2, \dots, T_K\}$ , assuming the first frame is the reference, i.e., $T_1 = \mathbb{I}$ . A 6-DoF camera pose $T_i$ is represented by Euler angles $\gamma$ and translations $t, T_i = (\gamma_x, \gamma_y, \gamma_z, t_x, t_y, t_z)$ . + +We also parameterize global, 9-DoF object poses, $\bar{T}_o = (\gamma_x,\gamma_y,\gamma_z,t_x,t_y,t_z,s_x,s_y,s_z)$ , comprising 6-DoF angles and translations, and 3-DoF anisotropic scales $s$ . + +We formulate a joint energy optimization of the form: + +$$ +T ^ {*}, \bar {T} ^ {*} = \operatorname {a r g m i n} _ {T, \bar {T}} \left(\mathrm {E} _ {c} (T, M) + \mathrm {E} _ {o} (T, \bar {T}, N)\right) \tag {1} +$$ + +where $M$ are inter-frame feature matches and $N$ are intra-frame canonical object correspondences established with normalized object coordinates (NOCs) that densely map + +![](images/08e5f55989e5efa801d13fecfce2a35ccdde2a703ae97a58438d71d085dc9799.jpg) +Figure 3. Multi-modal object recognition and NOC prediction. Our ResNet-FPN [16, 21] backbone takes color, reversed jet colored depth [12], and $2\mathrm{x}$ downsampled colored 3D depth normals, and produces multi-scale features by averaging different input encodings. From the obtained features, our method recognizes objects and predicts NOCs for each object, based on a Mask-RCNN [15]-style prediction. + +to the canonical space of an object in $[-0.5, 0.5]^3$ . $\mathrm{E}_c$ is the feature-matching energy function, and $\mathrm{E}_o$ is our object-centric energy function. Since robust feature matching and optimization are readily available off the shelf [10, 11, 31, 36], our method focuses on building the object-centric energy $\mathrm{E}_o$ . To this end, we need two function approximators, realized via deep neural networks: (1) a learned model for object recognition and NOC prediction, and (2) a learned model for object identification. The realization of these networks is described in Sections 3.2 and 3.3, respectively, and energy function Eq. 1 in Section 3.4. An overview of our approach is visualized in Figure 2. + +# 3.2. Predicting Object Correspondences + +To obtain object constraints in our final optimization, we recognize objects via object detection and instance segmentation, and predict object correspondences as dense NOCs [43] for each object, as shown in Figure 3. We build on a Mask-RCNN [15] with ResNet50-FPN [16, 21] back + +bone, pre-trained on ImageNet [33] and COCO [22]. + +To input the depth of the RGB-D frames, we propose a modified, multi-modal ResNet50-FPN [12, 16, 21] backbone. Our backbone takes $480 \times 640$ color, $480 \times 640$ reverse jet-colored depth, and $240 \times 320$ colored depth normals as input. We average the resulting FPN features to a single feature pyramid: + +$$ +G = \frac {\operatorname {F P N} ^ {c} \left(I ^ {c}\right) + \operatorname {F P N} ^ {d} \left(I ^ {d}\right) + \operatorname {U} \left(\operatorname {F P N} ^ {\mathrm {n}} \left(I ^ {n}\right)\right)}{3}, \tag {2} +$$ + +where $I^{c}, I^{d}, I^{n}$ are color, depth, and normal images, $\mathrm{FPN}^{c}, \mathrm{FPN}^{d}, \mathrm{FPN}^{n}$ are the corresponding ResNet50-FPN backbones, and U is an upsampling operator to match the normal features' spatial size with others. This enables fine-tuning the pre-trained bounding box, class, and instance segmentation heads of Mask-RCNN [15], while also exploiting depth information. We use symmetrically structured FPNs, all pre-trained on ImageNet and COCO as initialization, but without any parameter sharing. + +To obtain object correspondences, we establish mappings from detected objects to their canonical spaces, in the form of dense NOCs. That is, for each pixel in the object's instance mask, we predict 3D points $P_{o}^{\mathrm{noc}}$ using a fully convolutional network: + +$$ +P _ {o} ^ {\text {n o c}} = \operatorname {F C N} \left(G _ {o}\right), p \in [ - 0. 5, 0. 5 ] ^ {3} \forall p \in P _ {o} ^ {\text {n o c}}. \tag {3} +$$ + +We optimize an $\ell_1$ loss $L_{\mathrm{noc}}$ using ground-truth NOCs $P^{\mathrm{noc - gt}}$ + +$$ +L _ {\text {n o c}} = \sum_ {o} \sum_ {i} \left\| P _ {o, i} ^ {\text {n o c}} - P _ {o} ^ {\text {n o c - g t}} \right\| _ {1}. \tag {4} +$$ + +Since symmetric objects [1, 43] induce ambiguities in NOCs (e.g., a round table), we classify symmetry type of objects (round, square, rectangle, non-symmetric), $c_{\mathrm{sym}} = \mathrm{MLP}^{\mathrm{sym}}(G_o)$ , optimized using a cross-entropy loss $L_{\mathrm{sym}}$ . We also make $L_{\mathrm{noc}}$ symmetry aware, taking the minimum $\ell_1$ difference over the set of correct NOCs [43]. We use non-symmetric objects during inference to avoid inconsistent NOCs across views. + +In addition to NOCs, we also regress anisotropic 3D object scales $s_o$ using a fully connected network, $s_o = \mathrm{MLP}^{\mathrm{scale}}(G_o)$ , and optimize $s_o$ with an $\ell_1$ loss $L_{\mathrm{scale}}$ . The object scale enables holistic object pose understanding within each frame and helps to filter potential object matches across views using scale consistency. + +Finally, to make our NOC-depth correspondences least-squares friendly for our desired optimization, we also introduce a per-frame differentiable Procrustes objective $L_{\mathrm{proc}}$ , using a differentiable Kabsch solver [32] to obtain local object rotations and translations: + +$$ +R _ {o} ^ {*}, t _ {o} ^ {*} = \operatorname {a r g m i n} _ {R _ {o}, t _ {o}} \left(\sum_ {i} \left\| R _ {o} \left(P _ {o, i} ^ {\text {n o c}} \odot s _ {o}\right) + t _ {o} - P _ {o, i} ^ {\text {d e p t h}} \right\| _ {2} ^ {2}\right) \tag {5} +$$ + +![](images/402ff96d4437294d38869e400320ad0ad7e7e2754222176c013a76f84600eec3.jpg) +Figure 4. Our foreground/background metric-learning encoder for object matching, inspired by re-OBJ [3]. Using the detected and segmented objects from the model in Section 3.2, we encode foreground and background regions of objects, using light-weight, multi-modal ResNet18 encoders on the RGB-D features. + +for each object $o$ , where $P_{o}^{\mathrm{depth}}$ are back-projected input RGB-D depths corresponding to the object's predicted instance mask in its region of interest, and $\odot$ denotes element-wise multiplication. We train the local object poses with + +$$ +L _ {\text {p r o c}} = w _ {r} \sum_ {o} | | R _ {o} ^ {*} - R _ {o} ^ {\mathrm {g t}} | | _ {1} + w _ {t} \sum_ {o} | | t _ {o} ^ {*} - t _ {o} ^ {\mathrm {g t}} | | _ {2} ^ {2}. \tag {6} +$$ + +Our full loss used for training is then + +$$ +L = L _ {m} + w _ {n} L _ {\text {n o c}} + w _ {s} L _ {\text {s c a l e}} + w _ {\text {s y m}} L _ {\text {s y m}} + w _ {p} L _ {\text {p r o c}}, \tag {7} +$$ + +where $L_{m}$ is the sum of Mask-RCNN losses [15] and $w$ are scalar weights balancing additional losses. + +Implementation. We use the augmented 400k ScanNet train image split for training [9, 14], with Scan2CAD labels of the most common 9 furniture categories following [1, 14, 25]. We train a standard Detector2 Mask-RCNN pipeline [15, 45] with 1k warm-up iterations, 0.003 base learning rate, and learning rate decays at 60k and 100k iterations with a total of 120k training iterations. + +# 3.3. Matching Object Instances + +In our global pose optimization formulation, the relation between frames is formed via a global identification of objects across frames. To enable such identification without any heuristic spatial assumptions, we train a simple metric learner that ranks object semantic similarities across frames. Our object matching is shown in Figure 4. + +We characterize objects for matching by their respective RGB-D features in the object instance mask, in addition to the global context in the 5-times upscaled object bounding box. All inputs are resized to $224 \times 224$ and input to a lightweight ResNet18 [16] backbone pre-trained on ImageNet [33]. + +Similar to object detection, we employ two backbones for color and colored depth inputs. We omit normal input in object matching, as it empirically did not provide any + +benefit. For each input modality, we train two ResNet18 backbones for masked and inverse-masked crops, namely foreground (object) and background (context) encodings, + +$$ +e = \operatorname {M L P} \left(\left[ \mathrm {R N} ^ {c} \left(F _ {c}\right), \mathrm {R N} ^ {c} \left(B _ {c}\right) \right] + \left[ \mathrm {R N} ^ {d} \left(F _ {d}\right), \mathrm {R N} ^ {d} \left(B _ {d}\right) \right]\right) \tag {8} +$$ + +where RN are ResNet18s, MLP is a fully connected network, $F$ , $B$ are foreground and background crops for color (c) and depth (d), and $e$ is the object embedding vector. + +Given an anchor object embedding $e_a$ , a correctly matching object embedding $e_p$ , and a negative example $e_n$ , we train our metric learning using a triplet margin loss [17]: + +$$ +L _ {\text {t r i}} = \max \left(\mathrm {d} \left(e _ {a}, e _ {p}\right) - \mathrm {d} \left(e _ {a}, e _ {n}\right) + 1. 0, 0\right), \tag {9} +$$ + +where $d$ is the $\ell_2$ distance. + +We only consider triplets from the same category, as the object recognition pipeline provides classification. At inference time, we match the best instances using the Hungarian algorithm and apply a threshold $d(e_i, e_j) < \alpha$ for the matching object pairs from the same class. This semantic matching can be scaled to multiple frames via, e.g., object tracking with re-identification, or in our case, a simple pose graph optimization over frame pairs. + +Implementation. We implement the identification network using PyTorch [30] and train it on ScanNet data [9]. We train the network for 100k iterations with a batch size of 8 using a momentum optimizer with a learning rate 1e-4 and momentum 0.9. + +# 3.4. Energy Optimization + +We realize the joint energy minimization in Eq. 1 using keypoint and NOC constraints. + +Using the predicted NOC constraints with back-projected depths, we can re-write the $\mathrm{E}_o$ in Eq. 1 as: + +$$ +\mathrm {E} _ {o} (T, \bar {T}, P ^ {\text {d e p t h}}, P ^ {\text {n o c}}) = \sum_ {o} \sum_ {c} \sum_ {k} \| T _ {c} P _ {o, c, k} ^ {\text {d e p t h}} - \bar {T} _ {o} P _ {o, c, k} ^ {\text {n o c}} \| _ {2} ^ {2} \tag {10} +$$ + +where $\bar{T}$ , $T$ represent 9-DoF and 6-DoF camera and object transformations, respectively, and subscripts $o$ , $c$ , $k$ correspond to objects, cameras (frames), and points (pixels) within the frames, respectively. Here, object indices are determined by the object identification, and global object poses $\bar{T}$ indirectly constrain frames to each other without any explicit inter-frame feature matching. + +In many cases, object constraints alone may not be sufficient to optimize the camera pose (e.g., frames may not share matching objects together). However, our object-based optimization is fully complementary to classical feature matching, and we thus formulate our objective in combination with feature-matching constraints $\mathbf{E}_c$ : + +$$ +\mathrm {E} _ {c} (T, P ^ {\text {d e p t h}}) = \sum_ {i} \sum_ {j} \sum_ {m, n} | | T _ {i} P _ {i, m} ^ {\text {d e p t h}} - T _ {j} P _ {j, n} ^ {\text {d e p t h}} | | _ {2} ^ {2} \tag {11} +$$ + +Our method is agnostic to the feature matcher, whether classical or learning-based. In this work, we experiment with two different keypoint matching systems to realize $\mathrm{E}_c$ , namely SuperGlue [11, 36] and Geometric Transformer [31], both offering state-of-the-art indoor feature matching. + +With both object and feature-matching constraints, we realize the desired joint energy formulation as + +$$ +T ^ {*}, \bar {T} ^ {*} = \operatorname {a r g m i n} _ {T, \bar {T}} \left(w _ {c} \mathrm {E} _ {c} + w _ {o} \mathrm {E} _ {o}\right). \tag {12} +$$ + +where $w_{c}, w_{o}$ weight feature-matching, and object energies. + +Since non-linear least squares problems can be sensitive to outliers, we additionally employ outlier removal. Similar to BundleFusion [10], we apply Kabsch filtering to both intra-frame and keypoint constraints, using the matching depth-NOC and depth-depth correspondences, respectively. That is, we iteratively solve an orthogonal Procrustes problem and only keep the correspondences that have lower optimization errors. We use a liberal $20\mathrm{cm}$ threshold to handle wide-baseline frames. Objects and inter-frame matches are rejected if the number of NOCs is $< 15$ and the number of keypoints is $< 5$ , respectively. + +To solve this least-squares problem, we use a Gauss-Newton optimizer. To handle outliers during this global optimization, we remove $>15cm$ error residuals during optimization. As global optimization can produce coarse results, we further apply an ICP refinement. + +To handle SLAM-style sequence registration, we use a state-of-the-art global pose graph optimization [5, 47] over pairwise frame registration, using a single hierarchy level for simplicity. However, our method could be scaled to multi-frame optimization via object tracking and subdivided into multiple hierarchy levels to handle very large-scale scenes. + +# 4. Experiments + +# 4.1. Pairwise Registration + +Evaluation Dataset. We evaluate pairwise camera pose estimation results on 1569 challenging and diverse frame pairs from the validation set of ScanNet [9]. Unlike previous works [19, 36], our evaluation includes a significant number of low-overlap pairs, with over $15\%$ of pairs having $\leq 10\%$ overlap. + +Optimization. To establish object correspondences across frame pairs in Eq. 10, we use the top-1 matching object with embedding distance threshold 0.05. In the absence of both keypoint matches and object matches, we consider object matches with distance $< 0.15$ . Only objects of the same class label and predicted maximum scale ratio $< 1.5$ can match. We determine the best matching object based on the number of NOC-depth constraints after the filtering described in Section 3.4. We combine it with keypoint constraints and refine the results with ICP. + +Classical Baselines. We compare our method against various hand-crafted feature-matching and registration baselines. These baselines include BundleFusion keypoint optimization with SIFT descriptors and Kabsch filtering [10, 23], (SIFT + BF), and two geometric global registration baselines, Fast Global Registration (Fast GR) [46] and Global Registration (GR) [5], based on approximate and exact RANSAC over FPFH [34] features, respectively. As a post-processing step, all methods are refined using ICP. We use Open3D [47] implementations for global registration baselines as well as ICP post-processing of all methods. Our custom implementation of SIFT + BF uses the SIFT implementation of OpenCV [4]. + +Learned Baselines. We also compare with pose estimators using learned models for feature or object matching: Geometric Transformer (GeoTrans) [31], which performs dense geometric feature matching, and SG + BF which leverages the powerful learned SuperGlue [36] and SuperPoint [11] feature matching approaches in combination with BundleFusion [10] Kabsch filtering and 3D Gauss-Newton optimization. We also use our network predictions to create a 3D object tracking baseline, Object Track, mimicking tracking-based object-SLAM systems. Object Track obtains local object poses and object matches from Sections 3.2 and 3.3, but instead uses relative local object poses to get the camera pose instead of a global energy optimization. All methods are refined using ICP as a post-processing step. + +Evaluation Metrics. We use the Pose Recall metric for pose evaluation, following previous RGB-D pairwise registration works [6, 19]. To comprehensively capture performance, we employ several thresholds for absolute translation (in cm) and rotation (in °). Since absolute translation is more difficult than rotation, its thresholds are twice the angle thresholds, following [6]. + +Quantitative Results. In Table 1 and Table 2, we evaluate on ScanNet [9] validation frame pairs, measuring recall in various thresholds and overlap levels. Combined with state-of-the-art feature matching in a joint optimization, our method outperforms both classical and learning-based methods. Furthermore, the gap increases with lower overlap, as shown in Table 2 and Figure 6, since our method can leverage object-based correspondences to estimate alignment in the absence of keypoint matches. We also show that our global optimization (Ours (w/o keypoints)) significantly outperforms naive object-based tracking (Object Track), demonstrating the efficacy of our global optimization formulation. Our approach is designed to complement keypoint matching, notably improving the state of the art in combination with state-of-the-art geometric and RGB-D feature matching. + +Analysis of Performance by Overlap. In Table 2 and + +![](images/9cfccc32790ad4679ee2bd4e20401e3ba5785621353cd5934e538fa5e306ca31.jpg) +NOCs of Top-1 Match + +![](images/aae1864fcbbfbbb16200eba2f316e34e1333cf68d68533dfdcc063a4fc80b427.jpg) + +![](images/5e0ddfb0b0661574f7e42f78b245d399f194b0a22f8d27e0907bc5be33bd3749.jpg) + +![](images/4f6d79e6f9bb0c7daaf728dfa343a8264343b45aaf5b9b601720d48f13b3fc73.jpg) + +![](images/98858c7490497f8d59f5fc3b36be9a2527dfd29ba4729fbaf06e26849334577c.jpg) + +![](images/51b93a94c1fd6952d789c2402d7fa5511b9427d76f97310574e952ede3dd4fca.jpg) + +![](images/2c42047b9fec4bd00946536866713f105e93ff83637d1c6732d1d28dcea880d8.jpg) + +![](images/53f9ace2a1b7a4da49930ca4fdd768a56fc8669dae68c3cfd25a9278f2faf2e5.jpg) +Figure 5. Low-overlap registration on ScanNet [9], where traditional feature matching fails. Predicted NOC correspondences are visualized, along with object box and camera poses of the left and right images in green and blue, respectively. + +![](images/9d1e9c4da8637be8c38427dc11bb77f3c7d08ec70909acbce7aab359430a47bd.jpg) +Registration + +![](images/68865de6741342b2bb7fc162ae1b181069c1995ec38897caf92bb9cccf83a88d.jpg) + +![](images/e4830fff8a54603d09f95da77b25b63df1446127b4b3feb9a94a373211a8818c.jpg) + +![](images/b36974f6f162f2d9568da7dcbaf7c449602704d4c1b128b1f3ba285bc9aef9dd.jpg) + +![](images/242be60d142942c8ef0bababb8327e87f8f07153a8791633fe3f48b0a43c6056.jpg) + +![](images/3ea99be09070dccc218cf16ea93cb5422577435b1920255cdd9cb61207302a2a.jpg) + +![](images/ae863170fd7256e6b7adb669a803c7cba9509a1542f9890621d9e3591d6a7594.jpg) + +![](images/bc2838c48f0271e24329e4d7d63266202f03b2acbf61030549174569ce6c0f71.jpg) + +Figure 6, we show $30\mathrm{cm}$ , $15^{\circ}$ recall performance at different geometric overlap rates, using the radius-based geometric overlap percentage measurement from [19] with $1\mathrm{cm}$ threshold. While all methods' performances decrease with overlap, our method retains significantly more robustness. Classical keypoint matching (SIFT + BF, Fast GR, + +
MethodPose Recall by Threshold
5°, 10cm10°, 20cm15°, 30cm
SIFT + BF [10, 23]16.8318.4220.40
Fast GR [34, 46]39.0742.9644.68
GR [5]43.0847.6750.03
Object Track37.4143.2145.19
GeoTrans [31]67.1174.5776.93
SG + BF [10, 11, 36]71.1380.1181.71
Ours (w/o keypoints)56.4163.6766.99
Ours (w/ GeoTrans)68.3276.8079.35
Ours (w/ SG + BF)74.2584.5887.06
+ +Table 1. Pose Recall results on ScanNet [9] validation images. Combined with state-of-the-art feature matching, our method outperforms various classical and learning-based baselines. Our approach complements both Geometric Transformer [31] (Ours (w/ GeoTrans)) and SuperGlue [36] (Ours (w/ SG + BF)) feature matches, notably improving pose recall. + +
MethodRecall by Overlap %
≤ 10(10,30)≥ 30
SIFT + BF [10,23]1.302.2033.59
Fast GR [34,46]0.0013.9769.04
GR [5,34]0.6520.3674.62
Object Track8.4431.5458.86
GeoTrans [31]22.0864.2793.11
SG + BF [10,11,36]24.0373.4595.95
Ours (w/o keypoints)32.4753.2980.31
Ours (w/ GeoTrans)35.0665.4794.42
Ours (w/ SG + BF)45.4581.4497.16
+ +Table 2. Recall at $15^{\circ}$ , $30\mathrm{cm}$ by overlap percentage. Our method significantly outperforms strong baselines in challenging low-overlap frame pairs, almost doubling recall for overlap $\leq 10\%$ . + +![](images/d970bedea8200bb3a2326b3b67f1fe102e89862b86d5356040df18c8774cddbe.jpg) +Figure 6. Change of pose performance by overlap percentage on ScanNet [9] validation pairs. Our method is significantly more robust against decreasing overlap compared to various classical and learned baselines. + +SIFT) suffer strongly in low-overlap scenarios, which are challenging for hand-crafted descriptors. The learned approaches of GeoTrans and SG+BF maintain some performance with decreasing overlap but still suffer strongly. Our + +incorporation of object-based reasoning effectively complements keypoint matching while significantly improving robustness to low overlap, obtaining a performance improvement from $24.03\%$ to $45.45\%$ for overlap $\leq 10\%$ . + +Qualitative Results. Figure 5 shows examples of low-overlap registration on ScanNet [9] frame pairs. Due to the minimal overlap, traditional keypoint matching cannot find sufficient correspondences, while our object grounding enables accurate camera pose estimation through indirect object constraints. + +# 4.2. Registration of RGB-D SLAM Sequences + +We further demonstrate the effectiveness of our method in an RGB-D SLAM setting, registering sequences of frames in room-scale scenes. In particular, we evaluate challenging, low-frame-rate settings that reflect common non-expert capture such as fast camera motion. + +For optimization, we use a global pose graph optimization over pairwise registration similar to the multi-way global registration of [5, 6, 47], as it offers an off-the-shelf, simple, and robust solution. We refer to our supplementary material for further optimization details. + +Datasets and Evaluation Metrics. We evaluate our method on TUM-RGBD [39] and ScanNet [9] scenes. For ScanNet, we use a set of 12 scenes from the validation and test sets, having a wide range of sizes, environments, and camera trajectories. We sample every 30th frame for TUM-RGBD (1Hz) and every 20th frame in ScanNet (1.5Hz). All methods use the same hyperparameters for each dataset, respectively. To evaluate registration quality, we use the standard root mean squared trajectory error used for TUM-RGBD evaluation [39]. + +Baselines. We implement various popular feature descriptors from cutting-edge SLAM systems in global pose graph optimization setups and thoroughly tune the hyperparameters of each method to support low-frame-rate sequences. We compare against Redwood [5], a global multi-way pose graph that uses FPFH [34] features. We also deploy robust RGB-D SIFT features used in BundleFusion [10, 23] in a global pose graph optimization, creating the SIFT + BF baseline. Finally, we use the state-of-the-art SuperGlue matches of SuperPoint features with BundleFusion's sparse 3D optimization, SG + BF, as described in Section 4.1. + +Quantitative Results. In Table 3, we compare our method to state-of-the-art traditional and learned SLAM approaches. Across both TUM-RGBD and ScanNet validation and test scenes [9, 39], our object-grounded approach enables improved camera pose estimation. In particular, for the larger-scale ScanNet scenes, ObjectMatch achieves notable reductions in trajectory error, of $38\%$ , $49\%$ , $50\%$ , and $44\%$ in scenes 0423, 0430, 0718, and 0773. + +
MethodTUM Scene @ 1HzScanNet Scene @ 1.5Hz
fr1/deskfr1/xyzfr2/xyzfr2/deskfr3/office001100810169020704230430046104940718076007730805
SIFT + BF [10,23]30.603.492.73157.0238.01222.49108.03126.57104.64160.69248.0020.9033.9881.4838.0322.4786.80
Redwood [5,34,47]3.101.971.88140.17175.64164.0150.9647.2447.1788.43156.7610.479.77133.536.5543.19253.39
SG + BF [10,11,36]3.011.921.815.305.1413.886.778.095.627.6735.973.317.1613.426.3712.4312.98
Ours (w/ SG + BF)3.011.921.815.264.9212.535.736.855.324.7518.232.676.856.705.887.0010.86
+ +Table 3. ATE RMSE values (cm) on room-scale TUM-RGBD [39] and ScanNet [9] test scenes. We evaluate at challenging 1 FPS and 1.5 FPS by sampling every 30th and 20th frame from TUM-RGBD and ScanNet, respectively. ObjectMatch outperforms strong baselines leveraging classical and learning-based SLAM systems. + +![](images/c64e4018aceca06ceec3712dd2fdac8ab735dda3de5ffca06545abbeb9227de9.jpg) +Figure 7. Qualitative comparison for registration of ScanNet scenes [9], in comparison with only feature matching SG + BF [10, 11, 36]. All results are visualized by volumetric fusion of the respective RGB-D frames with the estimated poses. Our optimized poses produce more accurate, clean, and consistent reconstruction, due to incorporating the object-based information in our global optimization. Example loop closures with matching object correspondences are shown on the right. + +Qualitative Results. In Figure 7, we show ScanNet [9] scene reconstructions using estimated camera poses. RGB-D frames are fused using a scalable TSDF volumetric fusion [8,47]. Our approach complements state-of-the-art feature matching [10, 11, 36], enabling more accurate and consistent reconstructions using our indirect correspondences, in comparison to feature matching alone. We refer to the supplemental for additional qualitative sequence registration results. + +Limitations and Future Work. ObjectMatch shows the capability of incorporating object semantics into global pose optimization; however, various limitations remain. In particular, our approach leverages higher-level seman + +tic correspondence given by objects for registration; however, in scenarios where object matches do not exist in frame views, we cannot leverage this constraint and instead use only feature matches. Further leveraging background information regarding shared structures seen from different views would provide additional semantic correspondence information to further complement registration. Additionally, while estimating canonical correspondences with objects can provide significant view information, object alignments can be somewhat coarse, while reasoning at a finer-grained level of object parts would provide more precise reasoning. Finally, we believe future work can reduce our method's system complexity by adopting a joint multiframe SLAM optimization with fewer hyperparameters as well as a unified end-to-end architecture for learning object correspondences and identification together. + +# 5. Conclusion + +We have presented ObjectMatch which introduces a new paradigm for incorporating object semantics into camera pose registration. We leverage indirect, canonical object correspondences established with normalized object coordinates for camera pose optimization aspects in RGB-D registration and SLAM. To obtain these correspondences, we propose two multi-modal neural networks for object recognition with normalized object coordinate prediction and object identification. ObjectMatch operates in tandem with state-of-the-art feature matching in a joint Gauss-Newton optimization, with its object grounding enabling registration to handle frame pairs with very low to no shared visibility. As a result, ObjectMatch significantly improves the state-of-the-art feature matching in both pairwise and sequence registration of RGB-D frames, particularly in the challenging low-overlap regime. Overall, we hope our method opens up new possibilities in the context of leveraging semantic information for camera pose estimation. + +Acknowledgements. This work was supported by the ERC Starting Grant Scan2CAD (804724), the Bavarian State Ministry of Science and the Arts coordinated by the Bavarian Research Institute for Digital Transformation (bidt), the German Research Foundation (DFG) Grant "Making Machine Learning on Static and Dynamic 3D Data Practical", and the German Research Foundation (DFG) Research Unit "Learning and Simulation in Visual Computing." + +# References + +[1] Armen Avetisyan, Manuel Dahnert, Angela Dai, Manolis Savva, Angel X Chang, and Matthias Nießner. Scan2cad: Learning cad model alignment in rgb-d scans. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 2614-2623, 2019. 4 +[2] Armen Avetisyan, Angela Dai, and Matthias Nießner. End-to-end cad model retrieval and 9dof alignment in 3d scans. In Proceedings of the IEEE/CVF International Conference on computer vision, pages 2551-2560, 2019. 3 +[3] Vaibhav Bansal, Stuart James, and Alessio Del Bue. re-object: Jointly learning the foreground and background for object instance re-identification. In International Conference on Image Analysis and Processing, pages 402–413. Springer, 2019. 4 +[4] G. Bradski. The OpenCV Library. Dr. Dobb's Journal of Software Tools, 2000. 6 +[5] Sungjoon Choi, Qian-Yi Zhou, and Vladlen Koltun. Robust reconstruction of indoor scenes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5556-5565, 2015. 2, 5, 6, 7, 8 +[6] Christopher Choy, Wei Dong, and Vladlen Koltun. Deep global registration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2514-2523, 2020. 2, 6, 7 +[7] Christopher Choy, Jaesik Park, and Vladlen Koltun. Fully convolutional geometric features. In ICCV, 2019. 2 +[8] Brian Curless and Marc Levoy. A volumetric method for building complex models from range images. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, pages 303-312. ACM, 1996. 8 +[9] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 2, 4, 5, 6, 7, 8 +[10] Angela Dai, Matthias Nießner, Michael Zollhöfer, Shahram Izadi, and Christian Theobalt. Bundlefusion: Real-time globally consistent 3d reconstruction using on-the-fly surface reintegration. ACM Transactions on Graphics (ToG), 36(4):1, 2017. 2, 3, 5, 6, 7, 8 +[11] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 224-236, 2018. 2, 3, 5, 6, 7, 8 +[12] Andreas Eitel, Jost Tobias Springenberg, Luciano Spinello, Martin Riedmiller, and Wolfram Burgard. Multimodal deep learning for robust rgb-d object recognition. In 2015 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 681-687. IEEE, 2015. 3, 4 +[13] Hugo Germain, Vincent Lepetit, and Guillaume Bourmaud. Visual correspondence hallucination. In International Conference on Learning Representations, 2022. 2 +[14] Can Gümeli, Angela Dai, and Matthias Nießner. Roca: Robust cad model retrieval and alignment from a single image. In Proceedings of the IEEE/CVF Conference on Computer + +Vision and Pattern Recognition, pages 4022-4031, 2022. 3, 4 +[15] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 3, 4 +[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3, 4 +[17] Elad Hoffer and Nir Ailon. Deep metric learning using triplet network. In International workshop on similarity-based pattern recognition, pages 84-92. Springer, 2015. 5 +[18] Jiahui Huang, Shi-Sheng Huang, Haoxuan Song, and ShiMin Hu. Di-fusion: Online implicit 3d reconstruction with deep priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8932-8941, 2021. 2 +[19] Shengyu Huang, Zan Gojcic, Mikhail Usvyatsov, Andreas Wieser, and Konrad Schindler. Predator: Registration of 3d point clouds with low overlap. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4267-4276, June 2021. 2, 5, 6 +[20] Shahram Izadi, David Kim, Otmar Hilliges, David Molyneaux, Richard Newcombe, Pushmeet Kohli, Jamie Shotton, Steve Hodges, Dustin Freeman, Andrew Davison, et al. Kinectfusion: real-time 3d reconstruction and interaction using a moving depth camera. In Proceedings of the 24th annual ACM symposium on User interface software and technology, pages 559-568, 2011. 2 +[21] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2117-2125, 2017. 3, 4 +[22] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 4 +[23] David G Lowe. Object recognition from local scale-invariant features. In Proceedings of the seventh IEEE international conference on computer vision, volume 2, pages 1150-1157. IEEE, 1999. 6, 7, 8 +[24] Wei-Chiu Ma, Anqi Joyce Yang, Shenlong Wang, Raquel Urtasun, and Antonio Torralba. Virtual correspondence: Humans as a cue for extreme-view geometry. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15924-15934, 2022. 2 +[25] Kevis-Kokitsi Maninis, Stefan Popov, Matthias Niesser, and Vittorio Ferrari. Vid2cad: Cad model alignment using multiview constraints from videos. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 4 +[26] Yuhang Ming, Xingrui Yang, and Andrew Calway. Object-augmented rgb-d slam for wide-disparity relocalisation. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 2203-2209. IEEE, 2021. 2, 3 + +[27] Norman Muller, Yu-Shiang Wong, Niloy J Mitra, Angela Dai, and Matthias Nießner. Seeing behind objects for 3d multi-object tracking in rgb-d sequences. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6071-6080, 2021. 3 +[28] Raul Mur-Artal and Juan D Tardós. Orb-slam2: An open-source slam system for monocular, stereo, and rgb-d cameras. IEEE transactions on robotics, 33(5):1255-1262, 2017. 2 +[29] Matthias Nießner, Michael Zollhöfer, Shahram Izadi, and Marc Stamminger. Real-time 3d reconstruction at scale using voxel hashing. ACM Transactions on Graphics (ToG), 32(6):1-11, 2013. 2 +[30] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems 32, pages 8024-8035. Curran Associates, Inc., 2019. 5 +[31] Zheng Qin, Hao Yu, Changjian Wang, Yulan Guo, Yuxing Peng, and Kai Xu. Geometric transformer for fast and robust point cloud registration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11143-11152, 2022. 2, 3, 5, 6, 7 +[32] Nikhila Ravi, Jeremy Reizenstein, David Novotny, Taylor Gordon, Wan-Yen Lo, Justin Johnson, and Georgia Gkioxari. Accelerating 3d deep learning with pytorch3d. arXiv:2007.08501, 2020.4 +[33] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115(3):211-252, 2015. 4 +[34] Radu Bogdan Rusu, Nico Blodow, and Michael Beetz. Fast point feature histograms (fpfh) for 3d registration. In 2009 IEEE international conference on robotics and automation, pages 3212-3217. IEEE, 2009. 2, 6, 7, 8 +[35] Renato F Salas-Moreno, Richard A Newcombe, Hauke Strasdat, Paul HJ Kelly, and Andrew J Davison. Slam++: Simultaneous localisation and mapping at the level of objects. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1352-1359, 2013. 2 +[36] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4938–4947, 2020. 2, 3, 5, 6, 7, 8 +[37] Thomas Schops, Torsten Sattler, and Marc Pollefeys. Bad slam: Bundle adjusted direct rgb-d slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 134-144, 2019. 2 +[38] Yifei Shi, Kai Xu, Matthias Niessner, Szymon Rusinkiewicz, and Thomas Funkhouser. Planematch: Patch coplanarity prediction for robust rgb-d reconstruction. In Proceedings of the + +European Conference on Computer Vision (ECCV), pages 750-766, 2018. 2 +[39] J. Sturm, N. Engelhard, F. Endres, W. Burgard, and D. Cremers. A benchmark for the evaluation of rgb-d slam systems. In Proc. of the International Conference on Intelligent Robot Systems (IROS), Oct. 2012. 2, 7, 8 +[40] Edgar Sucar, Shikun Liu, Joseph Ortiz, and Andrew J Davison. imap: Implicit mapping and positioning in real-time. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6229-6238, 2021. 2 +[41] Edgar Sucar, Kentaro Wada, and Andrew Davison. NodeSLAM: Neural object descriptors for multi-view shape reconstruction. In Proceedings of the International Conference on 3D Vision (3DV), 2020. 2 +[42] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8922-8931, 2021. 2 +[43] He Wang, Srinath Sridhar, Jingwei Huang, Julien Valentin, Shuran Song, and Leonidas J Guibas. Normalized object coordinate space for category-level 6d object pose and size estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2642-2651, 2019. 2, 3, 4 +[44] T Whelan, S Leutenegger, RF. Salas-Moreno, B Glocker, and AJ. Davison. Elasticfusion: Dense slam without a pose graph. In Robotics: Science and Systems, 2015. 2 +[45] Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. Detector2. https://github.com/facebookresearch/detectron2, 2019.4 +[46] Qian-Yi Zhou, Jaesik Park, and Vladlen Koltun. Fast global registration. In European conference on computer vision, pages 766-782. Springer, 2016. 2, 6, 7 +[47] Qian-Yi Zhou, Jaesik Park, and Vladlen Koltun. Open3D: A modern library for 3D data processing. arXiv:1801.09847, 2018. 5, 6, 7, 8 +[48] Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hujun Bao, Zhaopeng Cui, Martin R Oswald, and Marc Pollefeys. Nice-slam: Neural implicit scalable encoding for slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12786-12796, 2022. 2 \ No newline at end of file diff --git a/2023/ObjectMatch_ Robust Registration Using Canonical Object Correspondences/images.zip b/2023/ObjectMatch_ Robust Registration Using Canonical Object Correspondences/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..9b3ca13ecbeaacb93feab9cd5c05c9d04c1eccbe --- /dev/null +++ b/2023/ObjectMatch_ Robust Registration Using Canonical Object Correspondences/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d27829a6ff7f1ecfc193475ccd8b462118271873f3c30a08b6b45cd13505584c +size 604334 diff --git a/2023/ObjectMatch_ Robust Registration Using Canonical Object Correspondences/layout.json b/2023/ObjectMatch_ Robust Registration Using Canonical Object Correspondences/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5c696a6a842b6197291bcc6ae93ed3e728318610 --- /dev/null +++ b/2023/ObjectMatch_ Robust Registration Using Canonical Object Correspondences/layout.json @@ -0,0 +1,9362 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 60, + 103, + 533, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 103, + 533, + 121 + ], + "spans": [ + { + "bbox": [ + 60, + 103, + 533, + 121 + ], + "type": "text", + "content": "ObjectMatch: Robust Registration using Canonical Object Correspondences" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 146, + 143, + 208, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 143, + 208, + 156 + ], + "spans": [ + { + "bbox": [ + 146, + 143, + 208, + 156 + ], + "type": "text", + "content": "Can Gumeli" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 255, + 144, + 313, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 144, + 313, + 157 + ], + "spans": [ + { + "bbox": [ + 255, + 144, + 313, + 157 + ], + "type": "text", + "content": "Angela Dai" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 361, + 144, + 447, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 144, + 447, + 156 + ], + "spans": [ + { + "bbox": [ + 361, + 144, + 447, + 156 + ], + "type": "text", + "content": "Matthias Nießner" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 217, + 171, + 373, + 185 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 171, + 373, + 185 + ], + "spans": [ + { + "bbox": [ + 217, + 171, + 373, + 185 + ], + "type": "text", + "content": "Technical University of Munich" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 59, + 228, + 168, + 310 + ], + "blocks": [ + { + "bbox": [ + 89, + 213, + 253, + 224 + ], + "lines": [ + { + "bbox": [ + 89, + 213, + 253, + 224 + ], + "spans": [ + { + "bbox": [ + 89, + 213, + 253, + 224 + ], + "type": "text", + "content": "Low-overlap Frames with Matching Objects" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 228, + 168, + 310 + ], + "lines": [ + { + "bbox": [ + 59, + 228, + 168, + 310 + ], + "spans": [ + { + "bbox": [ + 59, + 228, + 168, + 310 + ], + "type": "image", + "image_path": "d3f06c5d45558df7016439e082690d03636262a3120456f1cc122b9de7dc3222.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 170, + 228, + 278, + 309 + ], + "blocks": [ + { + "bbox": [ + 170, + 228, + 278, + 309 + ], + "lines": [ + { + "bbox": [ + 170, + 228, + 278, + 309 + ], + "spans": [ + { + "bbox": [ + 170, + 228, + 278, + 309 + ], + "type": "image", + "image_path": "2c3cef79c96c5e79dbb46a94649ef866e5264b69ef2f493c581ba90d9950a3c4.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 278, + 263, + 291, + 275 + ], + "blocks": [ + { + "bbox": [ + 278, + 263, + 291, + 275 + ], + "lines": [ + { + "bbox": [ + 278, + 263, + 291, + 275 + ], + "spans": [ + { + "bbox": [ + 278, + 263, + 291, + 275 + ], + "type": "image", + "image_path": "4357d973bceeaaad23d31aa645eee790d5fe4d7bf49d5a225e95cfd010db473f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 291, + 228, + 432, + 309 + ], + "blocks": [ + { + "bbox": [ + 309, + 213, + 414, + 224 + ], + "lines": [ + { + "bbox": [ + 309, + 213, + 414, + 224 + ], + "spans": [ + { + "bbox": [ + 309, + 213, + 414, + 224 + ], + "type": "text", + "content": "Canonical Correspondences" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 291, + 228, + 432, + 309 + ], + "lines": [ + { + "bbox": [ + 291, + 228, + 432, + 309 + ], + "spans": [ + { + "bbox": [ + 291, + 228, + 432, + 309 + ], + "type": "image", + "image_path": "17c4fb6bd2b65901cfbd33acbe69dcaae383bb72f40ea24e4f9fe3fe8e95608f.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 446, + 225, + 535, + 311 + ], + "blocks": [ + { + "bbox": [ + 441, + 213, + 538, + 224 + ], + "lines": [ + { + "bbox": [ + 441, + 213, + 538, + 224 + ], + "spans": [ + { + "bbox": [ + 441, + 213, + 538, + 224 + ], + "type": "text", + "content": "Camera and Object Poses" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 446, + 225, + 535, + 311 + ], + "lines": [ + { + "bbox": [ + 446, + 225, + 535, + 311 + ], + "spans": [ + { + "bbox": [ + 446, + 225, + 535, + 311 + ], + "type": "image", + "image_path": "fba06f9104b57c26157b6040acc7258a375f26936281e12abbee5750171b2a18.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 319, + 546, + 352 + ], + "lines": [ + { + "bbox": [ + 46, + 319, + 546, + 352 + ], + "spans": [ + { + "bbox": [ + 46, + 319, + 546, + 352 + ], + "type": "text", + "content": "Figure 1. Modern camera pose estimation relies on feature matching between overlapping frames - in this work, we present ObjectMatch to find correspondences between frames with little or no overlap by predicting semantic mappings through canonical object correspondences. The images above share no direct overlap, yet our method establishes indirect correspondences, thus enabling a successful registration." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 143, + 391, + 191, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 391, + 191, + 403 + ], + "spans": [ + { + "bbox": [ + 143, + 391, + 191, + 403 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 415, + 290, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 415, + 290, + 668 + ], + "spans": [ + { + "bbox": [ + 46, + 415, + 290, + 668 + ], + "type": "text", + "content": "We present ObjectMatch1, a semantic and object-centric camera pose estimator for RGB-D SLAM pipelines. Modern camera pose estimators rely on direct correspondences of overlapping regions between frames; however, they cannot align camera frames with little or no overlap. In this work, we propose to leverage indirect correspondences obtained via semantic object identification. For instance, when an object is seen from the front in one frame and from the back in another frame, we can provide additional pose constraints through canonical object correspondences. We first propose a neural network to predict such correspondences on a per-pixel level, which we then combine in our energy formulation with state-of-the-art keypoint matching solved with a joint Gauss-Newton optimization. In a pairwise setting, our method improves registration recall of state-of-the-art feature matching, including from " + }, + { + "bbox": [ + 46, + 415, + 290, + 668 + ], + "type": "inline_equation", + "content": "24\\%" + }, + { + "bbox": [ + 46, + 415, + 290, + 668 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 415, + 290, + 668 + ], + "type": "inline_equation", + "content": "45\\%" + }, + { + "bbox": [ + 46, + 415, + 290, + 668 + ], + "type": "text", + "content": " in pairs with " + }, + { + "bbox": [ + 46, + 415, + 290, + 668 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 415, + 290, + 668 + ], + "type": "text", + "content": " or less inter-frame overlap. In registering RGB-D sequences, our method outperforms cutting-edge SLAM baselines in challenging, low-frame-rate scenarios, achieving more than " + }, + { + "bbox": [ + 46, + 415, + 290, + 668 + ], + "type": "inline_equation", + "content": "35\\%" + }, + { + "bbox": [ + 46, + 415, + 290, + 668 + ], + "type": "text", + "content": " reduction in trajectory error in multiple scenes." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 361, + 386, + 374 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 361, + 386, + 374 + ], + "spans": [ + { + "bbox": [ + 307, + 361, + 386, + 374 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 384, + 545, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 384, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 304, + 384, + 545, + 480 + ], + "type": "text", + "content": "RGB-D registration and 3D SLAM has been a fundamental task in computer vision, with significant study and enabling many applications in mixed reality, robotics, and content creation. Central to both state-of-the-art traditional and learning-based camera pose estimation is establishing correspondences between points in input frames. However, correspondence estimation remains quite challenging when there is little or no overlap between frames." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 483, + 546, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 483, + 546, + 651 + ], + "spans": [ + { + "bbox": [ + 304, + 483, + 546, + 651 + ], + "type": "text", + "content": "In contrast, humans can easily localize across these challenging scenarios by leveraging additional semantic knowledge – in particular, by further localizing at the level of objects and identifying matching objects between views. For instance, when observing a chair from the back and the side (e.g., in Figure 1), view overlap is minimal (or even no view overlap), resulting in failed registration from keypoint matching. However, the semantic knowledge of the chair and its object pose nonetheless enables humans to estimate the poses from which the front and side views were taken. Thus, we propose to take a new perspective on camera pose estimation and imbue camera registration with awareness of such semantic correspondences between objects for robust performance in these challenging scenarios." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 545, + 713 + ], + "type": "text", + "content": "To this end, we propose ObjectMatch, a new paradigm for camera pose estimation leveraging canonical object correspondences in tandem with local keypoint correspondences between views. This enables significantly more robust registration under a variety of challenging scenar-" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 256, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 256, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 256, + 712 + ], + "type": "text", + "content": "1https://cangumeli.github.io/ObjectMatch/" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13082" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 239 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 239 + ], + "type": "text", + "content": "ios, including low view overlap. For a sequence of input frames, ObjectMatch learns to semantically identify objects across frames, enabling a compact, global parameterization of 9-DoF object poses. Object correspondences are established through predicting normalized object coordinates [43], dense correspondences from object pixels to a canonically oriented space for each object. We then formulate a joint camera and object pose optimization that constrains object correspondences indirectly, operating irrespective of the shared visibility of image regions. Our approach is complementary to state-of-the-art SLAM methods, and we leverage our energy formulation to complement state-of-the-art keypoint matching [10, 11, 36] in a joint Gauss-Newton optimization." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 239, + 289, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 239, + 289, + 359 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 289, + 359 + ], + "type": "text", + "content": "Our method outperforms strong baselines in both pairwise registration and registration of RGB-D frame sequences. In pairwise registration of challenging ScanNet [9] image pairs, we improve pose recall from " + }, + { + "bbox": [ + 46, + 239, + 289, + 359 + ], + "type": "inline_equation", + "content": "24\\%" + }, + { + "bbox": [ + 46, + 239, + 289, + 359 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 239, + 289, + 359 + ], + "type": "inline_equation", + "content": "45\\%" + }, + { + "bbox": [ + 46, + 239, + 289, + 359 + ], + "type": "text", + "content": " when the overlap is below " + }, + { + "bbox": [ + 46, + 239, + 289, + 359 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 239, + 289, + 359 + ], + "type": "text", + "content": ". On sequence registration of room-scale RGB-D scenes, our method outperforms various strong baselines in difficult, low-frame-rate settings in several TUM-RGBD [39] and ScanNet [9] scenes, reducing the trajectory error by more than " + }, + { + "bbox": [ + 46, + 239, + 289, + 359 + ], + "type": "inline_equation", + "content": "35\\%" + }, + { + "bbox": [ + 46, + 239, + 289, + 359 + ], + "type": "text", + "content": " in multiple challenging scenes." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 59, + 360, + 235, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 360, + 235, + 371 + ], + "spans": [ + { + "bbox": [ + 59, + 360, + 235, + 371 + ], + "type": "text", + "content": "To sum up, our main contributions include:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 374, + 287, + 550 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 58, + 374, + 287, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 374, + 287, + 422 + ], + "spans": [ + { + "bbox": [ + 58, + 374, + 287, + 422 + ], + "type": "text", + "content": "- An object-centric camera pose estimator that can handle low-overlap frame sets via indirect, canonical object correspondences established with predicted dense, per-pixel normalized object coordinates." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 426, + 287, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 426, + 287, + 462 + ], + "spans": [ + { + "bbox": [ + 58, + 426, + 287, + 462 + ], + "type": "text", + "content": "- A joint energy formulation that leverages semantic object identification and dense, normalized object coordinates corresponding to canonical object geometries." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 466, + 287, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 466, + 287, + 550 + ], + "spans": [ + { + "bbox": [ + 58, + 466, + 287, + 550 + ], + "type": "text", + "content": "- Our semantic grounding of object correspondences enables significantly more robust registration in low-overlap and low-frame-rate cases. ObjectMatch improves over state of the art from " + }, + { + "bbox": [ + 58, + 466, + 287, + 550 + ], + "type": "inline_equation", + "content": "24\\%" + }, + { + "bbox": [ + 58, + 466, + 287, + 550 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 58, + 466, + 287, + 550 + ], + "type": "inline_equation", + "content": "45\\%" + }, + { + "bbox": [ + 58, + 466, + 287, + 550 + ], + "type": "text", + "content": " registration recall of " + }, + { + "bbox": [ + 58, + 466, + 287, + 550 + ], + "type": "inline_equation", + "content": "\\leq 10\\%" + }, + { + "bbox": [ + 58, + 466, + 287, + 550 + ], + "type": "text", + "content": " overlap frame pairs and achieves over " + }, + { + "bbox": [ + 58, + 466, + 287, + 550 + ], + "type": "inline_equation", + "content": "35\\%" + }, + { + "bbox": [ + 58, + 466, + 287, + 550 + ], + "type": "text", + "content": " trajectory error reduction in several challenging sequences." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 562, + 134, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 562, + 134, + 574 + ], + "spans": [ + { + "bbox": [ + 47, + 562, + 134, + 574 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 289, + 715 + ], + "type": "text", + "content": "RGB-D Registration and SLAM. In recent years, there have been many advances in indoor RGB-D reconstruction. Earlier RGB-D fusion approaches focus on frame-to-model camera tracking [20, 29]. To handle the loop closures better, more recent SLAM systems introduce explicit strategies or global optimization methods for handling loop closures through global optimization [5,10,28,37,44] to fix tracking errors. More recently, deep learning techniques have been applied to registration and SLAM scenarios, with methods ranging from geometric point cloud registration [6, 19, 31] as well as neural field based SLAM techniques [18, 40, 48]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": "Despite all the successes in RGB-D SLAM and registration, the task is still challenging since incomplete loop closures observed via low-overlap frames cannot be handled, and most SLAM methods require a very high overlap between consecutive frames to track cameras accurately." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 138, + 546, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 138, + 546, + 402 + ], + "spans": [ + { + "bbox": [ + 304, + 138, + 546, + 402 + ], + "type": "text", + "content": "Feature Matching. Modern RGB(-D) camera pose estimators rely on a feature-matching backbone. Classical global registration techniques [5,46] use FPFH features [34] over point cloud fragments. On the other hand, many global RGB-D SLAM techniques rely on sparse color features [10, 28]. While being successful in many scenarios, conventional feature matching often fails when the interframe overlap is low. Therefore, deep learning techniques have been utilized for predicting overlapping regions based on geometry or color. On the geometric side, Deep Global Registration [6] predicts overlapping point features using nearest neighbor search over learned geometric features [7]. Methods such as PREDATOR [19] and Geometric Transformer [31] use attention mechanisms to target overlapping regions for registration. In the domain of color features, SuperPoint and SuperGlue [11, 36] build a formative approach in GNN-based keypoint feature matching. Methods such as LoFTR [42] introduce more dense and accurate sub-pixel level matching. Despite being very successful in handling wide-baseline scenarios, learned feature matching still requires a significant amount of shared visibility and geometric overlap." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 408, + 546, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 408, + 546, + 684 + ], + "spans": [ + { + "bbox": [ + 304, + 408, + 546, + 684 + ], + "type": "text", + "content": "Camera Pose Estimation with Semantic Cues. Several methods have been developed to incorporate semantic priors to improve low-overlap registration. PlaneMatch [38] proposed coplanarity priors for handling loop closures and low-overlap cases. Our method instead leverages object-centric constraints, exploiting the power of semantic object recognition. Another related direction is feature hallucination by leveraging image and object semantics. NeurHal [13] focuses on correspondence hallucination using image inpainting and outpainting, formulating a PnP optimization over hallucinated matches. Virtual Correspondence (VC) [24] introduces a human-centric approach that leverages hallucinated object volumes to form virtual epipolar constraints between frames. In contrast, we use indirect instead of direct correspondences that do not require hallucinated object volumes or image regions. Furthermore, our method works on a diverse set of furniture categories while VC focuses on humans. Pioneered by SLAM++ [35], there is also a rich literature of object-centric SLAM solutions, e.g., [26, 41]. Such SLAM methods leverage local, per-frame poses of objects to establish constraints; instead, we develop a global object pose optimization that is more robust against occluded viewpoints." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 714 + ], + "type": "text", + "content": "Object Pose Estimation using Normalized Object Coordinates. 3D object pose estimation has been widely stud" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13083" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 70, + 537, + 188 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 537, + 188 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 537, + 188 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 537, + 188 + ], + "type": "image", + "image_path": "5fd461f8ba293bf3aa6fcb79e1190e4e440fb2ab3150d5e8e43c1ff99c53e9de.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 192, + 547, + 237 + ], + "lines": [ + { + "bbox": [ + 46, + 192, + 547, + 237 + ], + "spans": [ + { + "bbox": [ + 46, + 192, + 547, + 237 + ], + "type": "text", + "content": "Figure 2. Overview of our approach to incorporate object correspondence grounding in global pose estimation. From a set of input RGB-D frames, ObjectMatch predicts object instances for each frame with dense normalized object correspondences. The predicted object instances are used to identify objects across frames, forming indirect object correspondences. We combine object correspondences with SuperGlue [11,36] keypoint matches in a joint energy optimization that yields both camera and object poses in a global registration." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 266, + 289, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 289, + 494 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 289, + 494 + ], + "type": "text", + "content": "ied from RGB and RGB-D inputs. Normalized Object Coordinate Space (NOCS) [43] was proposed to form dense correspondences from input RGB-D frames to canonical object geometries, enabling better generalization than direct regression of object poses. End2End SOCs [2] formulated a NOC-based approach for CAD retrieval and alignment to 3D scans, using a differentiable Procrustes algorithm. To enable CAD alignment to single RGB images, ROCA [14] leveraged NOC space in combination with predicted depths, formulating a robust differentiable Procrustes optimization [14]. Seeing Behind Objects [27] further leveraged NOC correspondences both to obtain local object poses and object completion for re-identification for RGB-D multi-object tracking. Wide Disparity Re-localization [26] uses the NOC predictions from [43] to construct an object-level map for re-localization in SLAM. In contrast to these approaches that focus on individual object poses, we use NOC correspondences directly in a multi-frame, global camera, and object pose optimization." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 503, + 104, + 515 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 503, + 104, + 515 + ], + "spans": [ + { + "bbox": [ + 47, + 503, + 104, + 515 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 523, + 140, + 536 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 523, + 140, + 536 + ], + "spans": [ + { + "bbox": [ + 47, + 523, + 140, + 536 + ], + "type": "text", + "content": "3.1. Problem Setup" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 540, + 287, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 540, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 47, + 540, + 287, + 601 + ], + "type": "text", + "content": "Given " + }, + { + "bbox": [ + 47, + 540, + 287, + 601 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 47, + 540, + 287, + 601 + ], + "type": "text", + "content": " RGB-D frames " + }, + { + "bbox": [ + 47, + 540, + 287, + 601 + ], + "type": "inline_equation", + "content": "\\{(I_1^c, I_1^d), \\dots, (I_K^c, I_K^d)\\}" + }, + { + "bbox": [ + 47, + 540, + 287, + 601 + ], + "type": "text", + "content": ", we aim to optimize their 6-DoF camera poses " + }, + { + "bbox": [ + 47, + 540, + 287, + 601 + ], + "type": "inline_equation", + "content": "T_c = \\{T_2, \\dots, T_K\\}" + }, + { + "bbox": [ + 47, + 540, + 287, + 601 + ], + "type": "text", + "content": ", assuming the first frame is the reference, i.e., " + }, + { + "bbox": [ + 47, + 540, + 287, + 601 + ], + "type": "inline_equation", + "content": "T_1 = \\mathbb{I}" + }, + { + "bbox": [ + 47, + 540, + 287, + 601 + ], + "type": "text", + "content": ". A 6-DoF camera pose " + }, + { + "bbox": [ + 47, + 540, + 287, + 601 + ], + "type": "inline_equation", + "content": "T_i" + }, + { + "bbox": [ + 47, + 540, + 287, + 601 + ], + "type": "text", + "content": " is represented by Euler angles " + }, + { + "bbox": [ + 47, + 540, + 287, + 601 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 47, + 540, + 287, + 601 + ], + "type": "text", + "content": " and translations " + }, + { + "bbox": [ + 47, + 540, + 287, + 601 + ], + "type": "inline_equation", + "content": "t, T_i = (\\gamma_x, \\gamma_y, \\gamma_z, t_x, t_y, t_z)" + }, + { + "bbox": [ + 47, + 540, + 287, + 601 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 601, + 287, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 601, + 287, + 636 + ], + "spans": [ + { + "bbox": [ + 47, + 601, + 287, + 636 + ], + "type": "text", + "content": "We also parameterize global, 9-DoF object poses, " + }, + { + "bbox": [ + 47, + 601, + 287, + 636 + ], + "type": "inline_equation", + "content": "\\bar{T}_o = (\\gamma_x,\\gamma_y,\\gamma_z,t_x,t_y,t_z,s_x,s_y,s_z)" + }, + { + "bbox": [ + 47, + 601, + 287, + 636 + ], + "type": "text", + "content": ", comprising 6-DoF angles and translations, and 3-DoF anisotropic scales " + }, + { + "bbox": [ + 47, + 601, + 287, + 636 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 47, + 601, + 287, + 636 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 59, + 637, + 277, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 637, + 277, + 649 + ], + "spans": [ + { + "bbox": [ + 59, + 637, + 277, + 649 + ], + "type": "text", + "content": "We formulate a joint energy optimization of the form:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 656, + 287, + 671 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 656, + 287, + 671 + ], + "spans": [ + { + "bbox": [ + 62, + 656, + 287, + 671 + ], + "type": "interline_equation", + "content": "T ^ {*}, \\bar {T} ^ {*} = \\operatorname {a r g m i n} _ {T, \\bar {T}} \\left(\\mathrm {E} _ {c} (T, M) + \\mathrm {E} _ {o} (T, \\bar {T}, N)\\right) \\tag {1}", + "image_path": "f3121c871a6830beca75c137cefe3b289f121be3e56bca35fab8862881cd4019.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "content": " are inter-frame feature matches and " + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "content": " are intra-frame canonical object correspondences established with normalized object coordinates (NOCs) that densely map" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 313, + 246, + 537, + 381 + ], + "blocks": [ + { + "bbox": [ + 313, + 246, + 537, + 381 + ], + "lines": [ + { + "bbox": [ + 313, + 246, + 537, + 381 + ], + "spans": [ + { + "bbox": [ + 313, + 246, + 537, + 381 + ], + "type": "image", + "image_path": "08e5f55989e5efa801d13fecfce2a35ccdde2a703ae97a58438d71d085dc9799.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 384, + 545, + 462 + ], + "lines": [ + { + "bbox": [ + 304, + 384, + 545, + 462 + ], + "spans": [ + { + "bbox": [ + 304, + 384, + 545, + 462 + ], + "type": "text", + "content": "Figure 3. Multi-modal object recognition and NOC prediction. Our ResNet-FPN [16, 21] backbone takes color, reversed jet colored depth [12], and " + }, + { + "bbox": [ + 304, + 384, + 545, + 462 + ], + "type": "inline_equation", + "content": "2\\mathrm{x}" + }, + { + "bbox": [ + 304, + 384, + 545, + 462 + ], + "type": "text", + "content": " downsampled colored 3D depth normals, and produces multi-scale features by averaging different input encodings. From the obtained features, our method recognizes objects and predicts NOCs for each object, based on a Mask-RCNN [15]-style prediction." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 483, + 546, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 483, + 546, + 626 + ], + "spans": [ + { + "bbox": [ + 304, + 483, + 546, + 626 + ], + "type": "text", + "content": "to the canonical space of an object in " + }, + { + "bbox": [ + 304, + 483, + 546, + 626 + ], + "type": "inline_equation", + "content": "[-0.5, 0.5]^3" + }, + { + "bbox": [ + 304, + 483, + 546, + 626 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 483, + 546, + 626 + ], + "type": "inline_equation", + "content": "\\mathrm{E}_c" + }, + { + "bbox": [ + 304, + 483, + 546, + 626 + ], + "type": "text", + "content": " is the feature-matching energy function, and " + }, + { + "bbox": [ + 304, + 483, + 546, + 626 + ], + "type": "inline_equation", + "content": "\\mathrm{E}_o" + }, + { + "bbox": [ + 304, + 483, + 546, + 626 + ], + "type": "text", + "content": " is our object-centric energy function. Since robust feature matching and optimization are readily available off the shelf [10, 11, 31, 36], our method focuses on building the object-centric energy " + }, + { + "bbox": [ + 304, + 483, + 546, + 626 + ], + "type": "inline_equation", + "content": "\\mathrm{E}_o" + }, + { + "bbox": [ + 304, + 483, + 546, + 626 + ], + "type": "text", + "content": ". To this end, we need two function approximators, realized via deep neural networks: (1) a learned model for object recognition and NOC prediction, and (2) a learned model for object identification. The realization of these networks is described in Sections 3.2 and 3.3, respectively, and energy function Eq. 1 in Section 3.4. An overview of our approach is visualized in Figure 2." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 635, + 495, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 635, + 495, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 635, + 495, + 647 + ], + "type": "text", + "content": "3.2. Predicting Object Correspondences" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "To obtain object constraints in our final optimization, we recognize objects via object detection and instance segmentation, and predict object correspondences as dense NOCs [43] for each object, as shown in Figure 3. We build on a Mask-RCNN [15] with ResNet50-FPN [16, 21] back" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13084" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 260, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 260, + 84 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 260, + 84 + ], + "type": "text", + "content": "bone, pre-trained on ImageNet [33] and COCO [22]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 84, + 287, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 84, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 84, + 287, + 156 + ], + "type": "text", + "content": "To input the depth of the RGB-D frames, we propose a modified, multi-modal ResNet50-FPN [12, 16, 21] backbone. Our backbone takes " + }, + { + "bbox": [ + 46, + 84, + 287, + 156 + ], + "type": "inline_equation", + "content": "480 \\times 640" + }, + { + "bbox": [ + 46, + 84, + 287, + 156 + ], + "type": "text", + "content": " color, " + }, + { + "bbox": [ + 46, + 84, + 287, + 156 + ], + "type": "inline_equation", + "content": "480 \\times 640" + }, + { + "bbox": [ + 46, + 84, + 287, + 156 + ], + "type": "text", + "content": " reverse jet-colored depth, and " + }, + { + "bbox": [ + 46, + 84, + 287, + 156 + ], + "type": "inline_equation", + "content": "240 \\times 320" + }, + { + "bbox": [ + 46, + 84, + 287, + 156 + ], + "type": "text", + "content": " colored depth normals as input. We average the resulting FPN features to a single feature pyramid:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 163, + 287, + 190 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 163, + 287, + 190 + ], + "spans": [ + { + "bbox": [ + 61, + 163, + 287, + 190 + ], + "type": "interline_equation", + "content": "G = \\frac {\\operatorname {F P N} ^ {c} \\left(I ^ {c}\\right) + \\operatorname {F P N} ^ {d} \\left(I ^ {d}\\right) + \\operatorname {U} \\left(\\operatorname {F P N} ^ {\\mathrm {n}} \\left(I ^ {n}\\right)\\right)}{3}, \\tag {2}", + "image_path": "5320f44833e7d13925d30ccb2272c5306fbbd81c6e1b92f9c5e7aee686b8660f.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 196, + 287, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 196, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 46, + 196, + 287, + 304 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 196, + 287, + 304 + ], + "type": "inline_equation", + "content": "I^{c}, I^{d}, I^{n}" + }, + { + "bbox": [ + 46, + 196, + 287, + 304 + ], + "type": "text", + "content": " are color, depth, and normal images, " + }, + { + "bbox": [ + 46, + 196, + 287, + 304 + ], + "type": "inline_equation", + "content": "\\mathrm{FPN}^{c}, \\mathrm{FPN}^{d}, \\mathrm{FPN}^{n}" + }, + { + "bbox": [ + 46, + 196, + 287, + 304 + ], + "type": "text", + "content": " are the corresponding ResNet50-FPN backbones, and U is an upsampling operator to match the normal features' spatial size with others. This enables fine-tuning the pre-trained bounding box, class, and instance segmentation heads of Mask-RCNN [15], while also exploiting depth information. We use symmetrically structured FPNs, all pre-trained on ImageNet and COCO as initialization, but without any parameter sharing." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 304, + 287, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 304, + 287, + 363 + ], + "spans": [ + { + "bbox": [ + 46, + 304, + 287, + 363 + ], + "type": "text", + "content": "To obtain object correspondences, we establish mappings from detected objects to their canonical spaces, in the form of dense NOCs. That is, for each pixel in the object's instance mask, we predict 3D points " + }, + { + "bbox": [ + 46, + 304, + 287, + 363 + ], + "type": "inline_equation", + "content": "P_{o}^{\\mathrm{noc}}" + }, + { + "bbox": [ + 46, + 304, + 287, + 363 + ], + "type": "text", + "content": " using a fully convolutional network:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 372, + 287, + 386 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 372, + 287, + 386 + ], + "spans": [ + { + "bbox": [ + 62, + 372, + 287, + 386 + ], + "type": "interline_equation", + "content": "P _ {o} ^ {\\text {n o c}} = \\operatorname {F C N} \\left(G _ {o}\\right), p \\in [ - 0. 5, 0. 5 ] ^ {3} \\forall p \\in P _ {o} ^ {\\text {n o c}}. \\tag {3}", + "image_path": "76e0871f08ae8135118af82ac42caa5d8a04dfc3992601adf280389e4fac0a74.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 393, + 287, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 393, + 287, + 415 + ], + "spans": [ + { + "bbox": [ + 47, + 393, + 287, + 415 + ], + "type": "text", + "content": "We optimize an " + }, + { + "bbox": [ + 47, + 393, + 287, + 415 + ], + "type": "inline_equation", + "content": "\\ell_1" + }, + { + "bbox": [ + 47, + 393, + 287, + 415 + ], + "type": "text", + "content": " loss " + }, + { + "bbox": [ + 47, + 393, + 287, + 415 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{noc}}" + }, + { + "bbox": [ + 47, + 393, + 287, + 415 + ], + "type": "text", + "content": " using ground-truth NOCs " + }, + { + "bbox": [ + 47, + 393, + 287, + 415 + ], + "type": "inline_equation", + "content": "P^{\\mathrm{noc - gt}}" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 96, + 415, + 287, + 439 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 415, + 287, + 439 + ], + "spans": [ + { + "bbox": [ + 96, + 415, + 287, + 439 + ], + "type": "interline_equation", + "content": "L _ {\\text {n o c}} = \\sum_ {o} \\sum_ {i} \\left\\| P _ {o, i} ^ {\\text {n o c}} - P _ {o} ^ {\\text {n o c - g t}} \\right\\| _ {1}. \\tag {4}", + "image_path": "1572021eb1e211c7efcf36726efd8f77a0b1d9e1bbf2eb9efb29d1a33eb24021.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 443, + 287, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 443, + 287, + 538 + ], + "spans": [ + { + "bbox": [ + 46, + 443, + 287, + 538 + ], + "type": "text", + "content": "Since symmetric objects [1, 43] induce ambiguities in NOCs (e.g., a round table), we classify symmetry type of objects (round, square, rectangle, non-symmetric), " + }, + { + "bbox": [ + 46, + 443, + 287, + 538 + ], + "type": "inline_equation", + "content": "c_{\\mathrm{sym}} = \\mathrm{MLP}^{\\mathrm{sym}}(G_o)" + }, + { + "bbox": [ + 46, + 443, + 287, + 538 + ], + "type": "text", + "content": ", optimized using a cross-entropy loss " + }, + { + "bbox": [ + 46, + 443, + 287, + 538 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{sym}}" + }, + { + "bbox": [ + 46, + 443, + 287, + 538 + ], + "type": "text", + "content": ". We also make " + }, + { + "bbox": [ + 46, + 443, + 287, + 538 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{noc}}" + }, + { + "bbox": [ + 46, + 443, + 287, + 538 + ], + "type": "text", + "content": " symmetry aware, taking the minimum " + }, + { + "bbox": [ + 46, + 443, + 287, + 538 + ], + "type": "inline_equation", + "content": "\\ell_1" + }, + { + "bbox": [ + 46, + 443, + 287, + 538 + ], + "type": "text", + "content": " difference over the set of correct NOCs [43]. We use non-symmetric objects during inference to avoid inconsistent NOCs across views." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 539, + 287, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 539, + 287, + 610 + ], + "spans": [ + { + "bbox": [ + 46, + 539, + 287, + 610 + ], + "type": "text", + "content": "In addition to NOCs, we also regress anisotropic 3D object scales " + }, + { + "bbox": [ + 46, + 539, + 287, + 610 + ], + "type": "inline_equation", + "content": "s_o" + }, + { + "bbox": [ + 46, + 539, + 287, + 610 + ], + "type": "text", + "content": " using a fully connected network, " + }, + { + "bbox": [ + 46, + 539, + 287, + 610 + ], + "type": "inline_equation", + "content": "s_o = \\mathrm{MLP}^{\\mathrm{scale}}(G_o)" + }, + { + "bbox": [ + 46, + 539, + 287, + 610 + ], + "type": "text", + "content": ", and optimize " + }, + { + "bbox": [ + 46, + 539, + 287, + 610 + ], + "type": "inline_equation", + "content": "s_o" + }, + { + "bbox": [ + 46, + 539, + 287, + 610 + ], + "type": "text", + "content": " with an " + }, + { + "bbox": [ + 46, + 539, + 287, + 610 + ], + "type": "inline_equation", + "content": "\\ell_1" + }, + { + "bbox": [ + 46, + 539, + 287, + 610 + ], + "type": "text", + "content": " loss " + }, + { + "bbox": [ + 46, + 539, + 287, + 610 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{scale}}" + }, + { + "bbox": [ + 46, + 539, + 287, + 610 + ], + "type": "text", + "content": ". The object scale enables holistic object pose understanding within each frame and helps to filter potential object matches across views using scale consistency." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 611, + 287, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 611, + 287, + 670 + ], + "spans": [ + { + "bbox": [ + 46, + 611, + 287, + 670 + ], + "type": "text", + "content": "Finally, to make our NOC-depth correspondences least-squares friendly for our desired optimization, we also introduce a per-frame differentiable Procrustes objective " + }, + { + "bbox": [ + 46, + 611, + 287, + 670 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{proc}}" + }, + { + "bbox": [ + 46, + 611, + 287, + 670 + ], + "type": "text", + "content": ", using a differentiable Kabsch solver [32] to obtain local object rotations and translations:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 678, + 287, + 713 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 678, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 678, + 287, + 713 + ], + "type": "interline_equation", + "content": "R _ {o} ^ {*}, t _ {o} ^ {*} = \\operatorname {a r g m i n} _ {R _ {o}, t _ {o}} \\left(\\sum_ {i} \\left\\| R _ {o} \\left(P _ {o, i} ^ {\\text {n o c}} \\odot s _ {o}\\right) + t _ {o} - P _ {o, i} ^ {\\text {d e p t h}} \\right\\| _ {2} ^ {2}\\right) \\tag {5}", + "image_path": "82b4b57bd0a3287e87effaed8b602f9c850805e997810192712eaab2c32ab855.jpg" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 310, + 72, + 541, + 187 + ], + "blocks": [ + { + "bbox": [ + 310, + 72, + 541, + 187 + ], + "lines": [ + { + "bbox": [ + 310, + 72, + 541, + 187 + ], + "spans": [ + { + "bbox": [ + 310, + 72, + 541, + 187 + ], + "type": "image", + "image_path": "402ff96d4437294d38869e400320ad0ad7e7e2754222176c013a76f84600eec3.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 190, + 545, + 245 + ], + "lines": [ + { + "bbox": [ + 305, + 190, + 545, + 245 + ], + "spans": [ + { + "bbox": [ + 305, + 190, + 545, + 245 + ], + "type": "text", + "content": "Figure 4. Our foreground/background metric-learning encoder for object matching, inspired by re-OBJ [3]. Using the detected and segmented objects from the model in Section 3.2, we encode foreground and background regions of objects, using light-weight, multi-modal ResNet18 encoders on the RGB-D features." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 265, + 545, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 265, + 545, + 315 + ], + "spans": [ + { + "bbox": [ + 305, + 265, + 545, + 315 + ], + "type": "text", + "content": "for each object " + }, + { + "bbox": [ + 305, + 265, + 545, + 315 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 305, + 265, + 545, + 315 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 305, + 265, + 545, + 315 + ], + "type": "inline_equation", + "content": "P_{o}^{\\mathrm{depth}}" + }, + { + "bbox": [ + 305, + 265, + 545, + 315 + ], + "type": "text", + "content": " are back-projected input RGB-D depths corresponding to the object's predicted instance mask in its region of interest, and " + }, + { + "bbox": [ + 305, + 265, + 545, + 315 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 305, + 265, + 545, + 315 + ], + "type": "text", + "content": " denotes element-wise multiplication. We train the local object poses with" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 315, + 322, + 545, + 347 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 322, + 545, + 347 + ], + "spans": [ + { + "bbox": [ + 315, + 322, + 545, + 347 + ], + "type": "interline_equation", + "content": "L _ {\\text {p r o c}} = w _ {r} \\sum_ {o} | | R _ {o} ^ {*} - R _ {o} ^ {\\mathrm {g t}} | | _ {1} + w _ {t} \\sum_ {o} | | t _ {o} ^ {*} - t _ {o} ^ {\\mathrm {g t}} | | _ {2} ^ {2}. \\tag {6}", + "image_path": "75a794ea578955a8116deb72aee62d4d7e9ebdd2efca02de3ccc618be832d9ac.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 318, + 355, + 468, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 355, + 468, + 367 + ], + "spans": [ + { + "bbox": [ + 318, + 355, + 468, + 367 + ], + "type": "text", + "content": "Our full loss used for training is then" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 311, + 376, + 545, + 389 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 376, + 545, + 389 + ], + "spans": [ + { + "bbox": [ + 311, + 376, + 545, + 389 + ], + "type": "interline_equation", + "content": "L = L _ {m} + w _ {n} L _ {\\text {n o c}} + w _ {s} L _ {\\text {s c a l e}} + w _ {\\text {s y m}} L _ {\\text {s y m}} + w _ {p} L _ {\\text {p r o c}}, \\tag {7}", + "image_path": "7e83ec113b61509e9a6176998d3a8d95ccda9e9a9af0053d29b58e06c6e55e02.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 398, + 545, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 398, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 305, + 398, + 545, + 422 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 398, + 545, + 422 + ], + "type": "inline_equation", + "content": "L_{m}" + }, + { + "bbox": [ + 305, + 398, + 545, + 422 + ], + "type": "text", + "content": " is the sum of Mask-RCNN losses [15] and " + }, + { + "bbox": [ + 305, + 398, + 545, + 422 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 305, + 398, + 545, + 422 + ], + "type": "text", + "content": " are scalar weights balancing additional losses." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 425, + 545, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 425, + 545, + 508 + ], + "spans": [ + { + "bbox": [ + 305, + 425, + 545, + 508 + ], + "type": "text", + "content": "Implementation. We use the augmented 400k ScanNet train image split for training [9, 14], with Scan2CAD labels of the most common 9 furniture categories following [1, 14, 25]. We train a standard Detector2 Mask-RCNN pipeline [15, 45] with 1k warm-up iterations, 0.003 base learning rate, and learning rate decays at 60k and 100k iterations with a total of 120k training iterations." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 515, + 455, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 515, + 455, + 529 + ], + "spans": [ + { + "bbox": [ + 306, + 515, + 455, + 529 + ], + "type": "text", + "content": "3.3. Matching Object Instances" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 534, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 545, + 605 + ], + "type": "text", + "content": "In our global pose optimization formulation, the relation between frames is formed via a global identification of objects across frames. To enable such identification without any heuristic spatial assumptions, we train a simple metric learner that ranks object semantic similarities across frames. Our object matching is shown in Figure 4." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 605, + 545, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 677 + ], + "type": "text", + "content": "We characterize objects for matching by their respective RGB-D features in the object instance mask, in addition to the global context in the 5-times upscaled object bounding box. All inputs are resized to " + }, + { + "bbox": [ + 304, + 605, + 545, + 677 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 304, + 605, + 545, + 677 + ], + "type": "text", + "content": " and input to a lightweight ResNet18 [16] backbone pre-trained on ImageNet [33]." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 305, + 677, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 714 + ], + "type": "text", + "content": "Similar to object detection, we employ two backbones for color and colored depth inputs. We omit normal input in object matching, as it empirically did not provide any" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13085" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 109 + ], + "type": "text", + "content": "benefit. For each input modality, we train two ResNet18 backbones for masked and inverse-masked crops, namely foreground (object) and background (context) encodings," + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 51, + 114, + 287, + 140 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 51, + 114, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 51, + 114, + 287, + 140 + ], + "type": "interline_equation", + "content": "e = \\operatorname {M L P} \\left(\\left[ \\mathrm {R N} ^ {c} \\left(F _ {c}\\right), \\mathrm {R N} ^ {c} \\left(B _ {c}\\right) \\right] + \\left[ \\mathrm {R N} ^ {d} \\left(F _ {d}\\right), \\mathrm {R N} ^ {d} \\left(B _ {d}\\right) \\right]\\right) \\tag {8}", + "image_path": "1173fc93016985189ba62d8a364873deca2dc2472c39d8937761f99404ba0345.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 140, + 287, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 140, + 287, + 176 + ], + "spans": [ + { + "bbox": [ + 47, + 140, + 287, + 176 + ], + "type": "text", + "content": "where RN are ResNet18s, MLP is a fully connected network, " + }, + { + "bbox": [ + 47, + 140, + 287, + 176 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 47, + 140, + 287, + 176 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 140, + 287, + 176 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 47, + 140, + 287, + 176 + ], + "type": "text", + "content": " are foreground and background crops for color (c) and depth (d), and " + }, + { + "bbox": [ + 47, + 140, + 287, + 176 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 47, + 140, + 287, + 176 + ], + "type": "text", + "content": " is the object embedding vector." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 177, + 287, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 177, + 287, + 213 + ], + "spans": [ + { + "bbox": [ + 47, + 177, + 287, + 213 + ], + "type": "text", + "content": "Given an anchor object embedding " + }, + { + "bbox": [ + 47, + 177, + 287, + 213 + ], + "type": "inline_equation", + "content": "e_a" + }, + { + "bbox": [ + 47, + 177, + 287, + 213 + ], + "type": "text", + "content": ", a correctly matching object embedding " + }, + { + "bbox": [ + 47, + 177, + 287, + 213 + ], + "type": "inline_equation", + "content": "e_p" + }, + { + "bbox": [ + 47, + 177, + 287, + 213 + ], + "type": "text", + "content": ", and a negative example " + }, + { + "bbox": [ + 47, + 177, + 287, + 213 + ], + "type": "inline_equation", + "content": "e_n" + }, + { + "bbox": [ + 47, + 177, + 287, + 213 + ], + "type": "text", + "content": ", we train our metric learning using a triplet margin loss [17]:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 78, + 220, + 287, + 233 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 220, + 287, + 233 + ], + "spans": [ + { + "bbox": [ + 78, + 220, + 287, + 233 + ], + "type": "interline_equation", + "content": "L _ {\\text {t r i}} = \\max \\left(\\mathrm {d} \\left(e _ {a}, e _ {p}\\right) - \\mathrm {d} \\left(e _ {a}, e _ {n}\\right) + 1. 0, 0\\right), \\tag {9}", + "image_path": "3b22081e92663fb21efe8a8b1868f580864a390e2ba8fa06b864cb90b392ac4d.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 241, + 154, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 241, + 154, + 252 + ], + "spans": [ + { + "bbox": [ + 47, + 241, + 154, + 252 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 241, + 154, + 252 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 47, + 241, + 154, + 252 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 47, + 241, + 154, + 252 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 47, + 241, + 154, + 252 + ], + "type": "text", + "content": " distance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 253, + 287, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 253, + 287, + 348 + ], + "spans": [ + { + "bbox": [ + 46, + 253, + 287, + 348 + ], + "type": "text", + "content": "We only consider triplets from the same category, as the object recognition pipeline provides classification. At inference time, we match the best instances using the Hungarian algorithm and apply a threshold " + }, + { + "bbox": [ + 46, + 253, + 287, + 348 + ], + "type": "inline_equation", + "content": "d(e_i, e_j) < \\alpha" + }, + { + "bbox": [ + 46, + 253, + 287, + 348 + ], + "type": "text", + "content": " for the matching object pairs from the same class. This semantic matching can be scaled to multiple frames via, e.g., object tracking with re-identification, or in our case, a simple pose graph optimization over frame pairs." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 351, + 287, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 351, + 287, + 410 + ], + "spans": [ + { + "bbox": [ + 46, + 351, + 287, + 410 + ], + "type": "text", + "content": "Implementation. We implement the identification network using PyTorch [30] and train it on ScanNet data [9]. We train the network for 100k iterations with a batch size of 8 using a momentum optimizer with a learning rate 1e-4 and momentum 0.9." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 418, + 168, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 418, + 168, + 430 + ], + "spans": [ + { + "bbox": [ + 47, + 418, + 168, + 430 + ], + "type": "text", + "content": "3.4. Energy Optimization" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 436, + 287, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 436, + 287, + 459 + ], + "spans": [ + { + "bbox": [ + 46, + 436, + 287, + 459 + ], + "type": "text", + "content": "We realize the joint energy minimization in Eq. 1 using keypoint and NOC constraints." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 460, + 287, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 460, + 287, + 484 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 287, + 484 + ], + "type": "text", + "content": "Using the predicted NOC constraints with back-projected depths, we can re-write the " + }, + { + "bbox": [ + 46, + 460, + 287, + 484 + ], + "type": "inline_equation", + "content": "\\mathrm{E}_o" + }, + { + "bbox": [ + 46, + 460, + 287, + 484 + ], + "type": "text", + "content": " in Eq. 1 as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 491, + 289, + 526 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 491, + 289, + 526 + ], + "spans": [ + { + "bbox": [ + 47, + 491, + 289, + 526 + ], + "type": "interline_equation", + "content": "\\mathrm {E} _ {o} (T, \\bar {T}, P ^ {\\text {d e p t h}}, P ^ {\\text {n o c}}) = \\sum_ {o} \\sum_ {c} \\sum_ {k} \\| T _ {c} P _ {o, c, k} ^ {\\text {d e p t h}} - \\bar {T} _ {o} P _ {o, c, k} ^ {\\text {n o c}} \\| _ {2} ^ {2} \\tag {10}", + "image_path": "aca8be2aa14b748b9bca71619b2c8832ad0f91946b7e8f374fcaa4495cd62956.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 526, + 287, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 526, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 46, + 526, + 287, + 609 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 526, + 287, + 609 + ], + "type": "inline_equation", + "content": "\\bar{T}" + }, + { + "bbox": [ + 46, + 526, + 287, + 609 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 526, + 287, + 609 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 46, + 526, + 287, + 609 + ], + "type": "text", + "content": " represent 9-DoF and 6-DoF camera and object transformations, respectively, and subscripts " + }, + { + "bbox": [ + 46, + 526, + 287, + 609 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 46, + 526, + 287, + 609 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 526, + 287, + 609 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 526, + 287, + 609 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 526, + 287, + 609 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 526, + 287, + 609 + ], + "type": "text", + "content": " correspond to objects, cameras (frames), and points (pixels) within the frames, respectively. Here, object indices are determined by the object identification, and global object poses " + }, + { + "bbox": [ + 46, + 526, + 287, + 609 + ], + "type": "inline_equation", + "content": "\\bar{T}" + }, + { + "bbox": [ + 46, + 526, + 287, + 609 + ], + "type": "text", + "content": " indirectly constrain frames to each other without any explicit inter-frame feature matching." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 611, + 287, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 611, + 287, + 682 + ], + "spans": [ + { + "bbox": [ + 46, + 611, + 287, + 682 + ], + "type": "text", + "content": "In many cases, object constraints alone may not be sufficient to optimize the camera pose (e.g., frames may not share matching objects together). However, our object-based optimization is fully complementary to classical feature matching, and we thus formulate our objective in combination with feature-matching constraints " + }, + { + "bbox": [ + 46, + 611, + 287, + 682 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_c" + }, + { + "bbox": [ + 46, + 611, + 287, + 682 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 52, + 689, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 52, + 689, + 287, + 715 + ], + "type": "interline_equation", + "content": "\\mathrm {E} _ {c} (T, P ^ {\\text {d e p t h}}) = \\sum_ {i} \\sum_ {j} \\sum_ {m, n} | | T _ {i} P _ {i, m} ^ {\\text {d e p t h}} - T _ {j} P _ {j, n} ^ {\\text {d e p t h}} | | _ {2} ^ {2} \\tag {11}", + "image_path": "38c974fb345c11d563cc4cd805e23409b77384ce72838bfb71fed074f1ed730c.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 72, + 545, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 132 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 132 + ], + "type": "text", + "content": "Our method is agnostic to the feature matcher, whether classical or learning-based. In this work, we experiment with two different keypoint matching systems to realize " + }, + { + "bbox": [ + 305, + 72, + 545, + 132 + ], + "type": "inline_equation", + "content": "\\mathrm{E}_c" + }, + { + "bbox": [ + 305, + 72, + 545, + 132 + ], + "type": "text", + "content": ", namely SuperGlue [11, 36] and Geometric Transformer [31], both offering state-of-the-art indoor feature matching." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 133, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 133, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 305, + 133, + 545, + 156 + ], + "type": "text", + "content": "With both object and feature-matching constraints, we realize the desired joint energy formulation as" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 348, + 163, + 545, + 176 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 163, + 545, + 176 + ], + "spans": [ + { + "bbox": [ + 348, + 163, + 545, + 176 + ], + "type": "interline_equation", + "content": "T ^ {*}, \\bar {T} ^ {*} = \\operatorname {a r g m i n} _ {T, \\bar {T}} \\left(w _ {c} \\mathrm {E} _ {c} + w _ {o} \\mathrm {E} _ {o}\\right). \\tag {12}", + "image_path": "eb04a27edfdc8d0c95bd4541bff901ecb4fe241b809e1a29ab7940b567eea4be.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 182, + 545, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 182, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 305, + 182, + 545, + 194 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 182, + 545, + 194 + ], + "type": "inline_equation", + "content": "w_{c}, w_{o}" + }, + { + "bbox": [ + 305, + 182, + 545, + 194 + ], + "type": "text", + "content": " weight feature-matching, and object energies." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 194, + 545, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 194, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 304, + 194, + 545, + 326 + ], + "type": "text", + "content": "Since non-linear least squares problems can be sensitive to outliers, we additionally employ outlier removal. Similar to BundleFusion [10], we apply Kabsch filtering to both intra-frame and keypoint constraints, using the matching depth-NOC and depth-depth correspondences, respectively. That is, we iteratively solve an orthogonal Procrustes problem and only keep the correspondences that have lower optimization errors. We use a liberal " + }, + { + "bbox": [ + 304, + 194, + 545, + 326 + ], + "type": "inline_equation", + "content": "20\\mathrm{cm}" + }, + { + "bbox": [ + 304, + 194, + 545, + 326 + ], + "type": "text", + "content": " threshold to handle wide-baseline frames. Objects and inter-frame matches are rejected if the number of NOCs is " + }, + { + "bbox": [ + 304, + 194, + 545, + 326 + ], + "type": "inline_equation", + "content": "< 15" + }, + { + "bbox": [ + 304, + 194, + 545, + 326 + ], + "type": "text", + "content": " and the number of keypoints is " + }, + { + "bbox": [ + 304, + 194, + 545, + 326 + ], + "type": "inline_equation", + "content": "< 5" + }, + { + "bbox": [ + 304, + 194, + 545, + 326 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 326, + 545, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 326, + 545, + 385 + ], + "spans": [ + { + "bbox": [ + 304, + 326, + 545, + 385 + ], + "type": "text", + "content": "To solve this least-squares problem, we use a Gauss-Newton optimizer. To handle outliers during this global optimization, we remove " + }, + { + "bbox": [ + 304, + 326, + 545, + 385 + ], + "type": "inline_equation", + "content": ">15cm" + }, + { + "bbox": [ + 304, + 326, + 545, + 385 + ], + "type": "text", + "content": " error residuals during optimization. As global optimization can produce coarse results, we further apply an ICP refinement." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 386, + 545, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 386, + 545, + 468 + ], + "spans": [ + { + "bbox": [ + 304, + 386, + 545, + 468 + ], + "type": "text", + "content": "To handle SLAM-style sequence registration, we use a state-of-the-art global pose graph optimization [5, 47] over pairwise frame registration, using a single hierarchy level for simplicity. However, our method could be scaled to multi-frame optimization via object tracking and subdivided into multiple hierarchy levels to handle very large-scale scenes." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 479, + 387, + 492 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 479, + 387, + 492 + ], + "spans": [ + { + "bbox": [ + 306, + 479, + 387, + 492 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 498, + 429, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 498, + 429, + 510 + ], + "spans": [ + { + "bbox": [ + 306, + 498, + 429, + 510 + ], + "type": "text", + "content": "4.1. Pairwise Registration" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 517, + 545, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 517, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 304, + 517, + 545, + 589 + ], + "type": "text", + "content": "Evaluation Dataset. We evaluate pairwise camera pose estimation results on 1569 challenging and diverse frame pairs from the validation set of ScanNet [9]. Unlike previous works [19, 36], our evaluation includes a significant number of low-overlap pairs, with over " + }, + { + "bbox": [ + 304, + 517, + 545, + 589 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 304, + 517, + 545, + 589 + ], + "type": "text", + "content": " of pairs having " + }, + { + "bbox": [ + 304, + 517, + 545, + 589 + ], + "type": "inline_equation", + "content": "\\leq 10\\%" + }, + { + "bbox": [ + 304, + 517, + 545, + 589 + ], + "type": "text", + "content": " overlap." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 594, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 712 + ], + "type": "text", + "content": "Optimization. To establish object correspondences across frame pairs in Eq. 10, we use the top-1 matching object with embedding distance threshold 0.05. In the absence of both keypoint matches and object matches, we consider object matches with distance " + }, + { + "bbox": [ + 304, + 594, + 545, + 712 + ], + "type": "inline_equation", + "content": "< 0.15" + }, + { + "bbox": [ + 304, + 594, + 545, + 712 + ], + "type": "text", + "content": ". Only objects of the same class label and predicted maximum scale ratio " + }, + { + "bbox": [ + 304, + 594, + 545, + 712 + ], + "type": "inline_equation", + "content": "< 1.5" + }, + { + "bbox": [ + 304, + 594, + 545, + 712 + ], + "type": "text", + "content": " can match. We determine the best matching object based on the number of NOC-depth constraints after the filtering described in Section 3.4. We combine it with keypoint constraints and refine the results with ICP." + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13086" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 228 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 228 + ], + "type": "text", + "content": "Classical Baselines. We compare our method against various hand-crafted feature-matching and registration baselines. These baselines include BundleFusion keypoint optimization with SIFT descriptors and Kabsch filtering [10, 23], (SIFT + BF), and two geometric global registration baselines, Fast Global Registration (Fast GR) [46] and Global Registration (GR) [5], based on approximate and exact RANSAC over FPFH [34] features, respectively. As a post-processing step, all methods are refined using ICP. We use Open3D [47] implementations for global registration baselines as well as ICP post-processing of all methods. Our custom implementation of SIFT + BF uses the SIFT implementation of OpenCV [4]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 233, + 289, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 233, + 289, + 413 + ], + "spans": [ + { + "bbox": [ + 47, + 233, + 289, + 413 + ], + "type": "text", + "content": "Learned Baselines. We also compare with pose estimators using learned models for feature or object matching: Geometric Transformer (GeoTrans) [31], which performs dense geometric feature matching, and SG + BF which leverages the powerful learned SuperGlue [36] and SuperPoint [11] feature matching approaches in combination with BundleFusion [10] Kabsch filtering and 3D Gauss-Newton optimization. We also use our network predictions to create a 3D object tracking baseline, Object Track, mimicking tracking-based object-SLAM systems. Object Track obtains local object poses and object matches from Sections 3.2 and 3.3, but instead uses relative local object poses to get the camera pose instead of a global energy optimization. All methods are refined using ICP as a post-processing step." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 416, + 288, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 416, + 288, + 502 + ], + "spans": [ + { + "bbox": [ + 47, + 416, + 288, + 502 + ], + "type": "text", + "content": "Evaluation Metrics. We use the Pose Recall metric for pose evaluation, following previous RGB-D pairwise registration works [6, 19]. To comprehensively capture performance, we employ several thresholds for absolute translation (in cm) and rotation (in °). Since absolute translation is more difficult than rotation, its thresholds are twice the angle thresholds, following [6]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 505, + 289, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 505, + 289, + 696 + ], + "spans": [ + { + "bbox": [ + 47, + 505, + 289, + 696 + ], + "type": "text", + "content": "Quantitative Results. In Table 1 and Table 2, we evaluate on ScanNet [9] validation frame pairs, measuring recall in various thresholds and overlap levels. Combined with state-of-the-art feature matching in a joint optimization, our method outperforms both classical and learning-based methods. Furthermore, the gap increases with lower overlap, as shown in Table 2 and Figure 6, since our method can leverage object-based correspondences to estimate alignment in the absence of keypoint matches. We also show that our global optimization (Ours (w/o keypoints)) significantly outperforms naive object-based tracking (Object Track), demonstrating the efficacy of our global optimization formulation. Our approach is designed to complement keypoint matching, notably improving the state of the art in combination with state-of-the-art geometric and RGB-D feature matching." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 701, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 701, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 701, + 288, + 714 + ], + "type": "text", + "content": "Analysis of Performance by Overlap. In Table 2 and" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 316, + 85, + 468, + 144 + ], + "blocks": [ + { + "bbox": [ + 348, + 72, + 429, + 83 + ], + "lines": [ + { + "bbox": [ + 348, + 72, + 429, + 83 + ], + "spans": [ + { + "bbox": [ + 348, + 72, + 429, + 83 + ], + "type": "text", + "content": "NOCs of Top-1 Match" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 316, + 85, + 468, + 144 + ], + "lines": [ + { + "bbox": [ + 316, + 85, + 468, + 144 + ], + "spans": [ + { + "bbox": [ + 316, + 85, + 468, + 144 + ], + "type": "image", + "image_path": "9cfccc32790ad4679ee2bd4e20401e3ba5785621353cd5934e538fa5e306ca31.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 317, + 144, + 467, + 203 + ], + "blocks": [ + { + "bbox": [ + 317, + 144, + 467, + 203 + ], + "lines": [ + { + "bbox": [ + 317, + 144, + 467, + 203 + ], + "spans": [ + { + "bbox": [ + 317, + 144, + 467, + 203 + ], + "type": "image", + "image_path": "aae1864fcbbfbbb16200eba2f316e34e1333cf68d68533dfdcc063a4fc80b427.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 317, + 205, + 467, + 263 + ], + "blocks": [ + { + "bbox": [ + 317, + 205, + 467, + 263 + ], + "lines": [ + { + "bbox": [ + 317, + 205, + 467, + 263 + ], + "spans": [ + { + "bbox": [ + 317, + 205, + 467, + 263 + ], + "type": "image", + "image_path": "5e0ddfb0b0661574f7e42f78b245d399f194b0a22f8d27e0907bc5be33bd3749.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 317, + 265, + 466, + 323 + ], + "blocks": [ + { + "bbox": [ + 317, + 265, + 466, + 323 + ], + "lines": [ + { + "bbox": [ + 317, + 265, + 466, + 323 + ], + "spans": [ + { + "bbox": [ + 317, + 265, + 466, + 323 + ], + "type": "image", + "image_path": "4f6d79e6f9bb0c7daaf728dfa343a8264343b45aaf5b9b601720d48f13b3fc73.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 317, + 325, + 466, + 384 + ], + "blocks": [ + { + "bbox": [ + 317, + 325, + 466, + 384 + ], + "lines": [ + { + "bbox": [ + 317, + 325, + 466, + 384 + ], + "spans": [ + { + "bbox": [ + 317, + 325, + 466, + 384 + ], + "type": "image", + "image_path": "98858c7490497f8d59f5fc3b36be9a2527dfd29ba4729fbaf06e26849334577c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 317, + 385, + 467, + 444 + ], + "blocks": [ + { + "bbox": [ + 317, + 385, + 467, + 444 + ], + "lines": [ + { + "bbox": [ + 317, + 385, + 467, + 444 + ], + "spans": [ + { + "bbox": [ + 317, + 385, + 467, + 444 + ], + "type": "image", + "image_path": "51b93a94c1fd6952d789c2402d7fa5511b9427d76f97310574e952ede3dd4fca.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 317, + 445, + 467, + 504 + ], + "blocks": [ + { + "bbox": [ + 317, + 445, + 467, + 504 + ], + "lines": [ + { + "bbox": [ + 317, + 445, + 467, + 504 + ], + "spans": [ + { + "bbox": [ + 317, + 445, + 467, + 504 + ], + "type": "image", + "image_path": "2c42047b9fec4bd00946536866713f105e93ff83637d1c6732d1d28dcea880d8.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 317, + 506, + 467, + 563 + ], + "blocks": [ + { + "bbox": [ + 317, + 506, + 467, + 563 + ], + "lines": [ + { + "bbox": [ + 317, + 506, + 467, + 563 + ], + "spans": [ + { + "bbox": [ + 317, + 506, + 467, + 563 + ], + "type": "image", + "image_path": "53f9ace2a1b7a4da49930ca4fdd768a56fc8669dae68c3cfd25a9278f2faf2e5.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 568, + 547, + 613 + ], + "lines": [ + { + "bbox": [ + 305, + 568, + 547, + 613 + ], + "spans": [ + { + "bbox": [ + 305, + 568, + 547, + 613 + ], + "type": "text", + "content": "Figure 5. Low-overlap registration on ScanNet [9], where traditional feature matching fails. Predicted NOC correspondences are visualized, along with object box and camera poses of the left and right images in green and blue, respectively." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 468, + 85, + 533, + 144 + ], + "blocks": [ + { + "bbox": [ + 476, + 72, + 523, + 83 + ], + "lines": [ + { + "bbox": [ + 476, + 72, + 523, + 83 + ], + "spans": [ + { + "bbox": [ + 476, + 72, + 523, + 83 + ], + "type": "text", + "content": "Registration" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 468, + 85, + 533, + 144 + ], + "lines": [ + { + "bbox": [ + 468, + 85, + 533, + 144 + ], + "spans": [ + { + "bbox": [ + 468, + 85, + 533, + 144 + ], + "type": "image", + "image_path": "9d1e9c4da8637be8c38427dc11bb77f3c7d08ec70909acbce7aab359430a47bd.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 467, + 147, + 533, + 201 + ], + "blocks": [ + { + "bbox": [ + 467, + 147, + 533, + 201 + ], + "lines": [ + { + "bbox": [ + 467, + 147, + 533, + 201 + ], + "spans": [ + { + "bbox": [ + 467, + 147, + 533, + 201 + ], + "type": "image", + "image_path": "68865de6741342b2bb7fc162ae1b181069c1995ec38897caf92bb9cccf83a88d.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 468, + 205, + 529, + 263 + ], + "blocks": [ + { + "bbox": [ + 468, + 205, + 529, + 263 + ], + "lines": [ + { + "bbox": [ + 468, + 205, + 529, + 263 + ], + "spans": [ + { + "bbox": [ + 468, + 205, + 529, + 263 + ], + "type": "image", + "image_path": "e4830fff8a54603d09f95da77b25b63df1446127b4b3feb9a94a373211a8818c.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 469, + 267, + 525, + 323 + ], + "blocks": [ + { + "bbox": [ + 469, + 267, + 525, + 323 + ], + "lines": [ + { + "bbox": [ + 469, + 267, + 525, + 323 + ], + "spans": [ + { + "bbox": [ + 469, + 267, + 525, + 323 + ], + "type": "image", + "image_path": "b36974f6f162f2d9568da7dcbaf7c449602704d4c1b128b1f3ba285bc9aef9dd.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 469, + 328, + 531, + 384 + ], + "blocks": [ + { + "bbox": [ + 469, + 328, + 531, + 384 + ], + "lines": [ + { + "bbox": [ + 469, + 328, + 531, + 384 + ], + "spans": [ + { + "bbox": [ + 469, + 328, + 531, + 384 + ], + "type": "image", + "image_path": "242be60d142942c8ef0bababb8327e87f8f07153a8791633fe3f48b0a43c6056.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 468, + 392, + 529, + 444 + ], + "blocks": [ + { + "bbox": [ + 468, + 392, + 529, + 444 + ], + "lines": [ + { + "bbox": [ + 468, + 392, + 529, + 444 + ], + "spans": [ + { + "bbox": [ + 468, + 392, + 529, + 444 + ], + "type": "image", + "image_path": "3ea99be09070dccc218cf16ea93cb5422577435b1920255cdd9cb61207302a2a.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 469, + 446, + 536, + 503 + ], + "blocks": [ + { + "bbox": [ + 469, + 446, + 536, + 503 + ], + "lines": [ + { + "bbox": [ + 469, + 446, + 536, + 503 + ], + "spans": [ + { + "bbox": [ + 469, + 446, + 536, + 503 + ], + "type": "image", + "image_path": "ae863170fd7256e6b7adb669a803c7cba9509a1542f9890621d9e3591d6a7594.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 474, + 510, + 536, + 563 + ], + "blocks": [ + { + "bbox": [ + 474, + 510, + 536, + 563 + ], + "lines": [ + { + "bbox": [ + 474, + 510, + 536, + 563 + ], + "spans": [ + { + "bbox": [ + 474, + 510, + 536, + 563 + ], + "type": "image", + "image_path": "bc2838c48f0271e24329e4d7d63266202f03b2acbf61030549174569ce6c0f71.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "bbox": [ + 305, + 641, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 641, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 641, + 547, + 715 + ], + "type": "text", + "content": "Figure 6, we show " + }, + { + "bbox": [ + 305, + 641, + 547, + 715 + ], + "type": "inline_equation", + "content": "30\\mathrm{cm}" + }, + { + "bbox": [ + 305, + 641, + 547, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 641, + 547, + 715 + ], + "type": "inline_equation", + "content": "15^{\\circ}" + }, + { + "bbox": [ + 305, + 641, + 547, + 715 + ], + "type": "text", + "content": " recall performance at different geometric overlap rates, using the radius-based geometric overlap percentage measurement from [19] with " + }, + { + "bbox": [ + 305, + 641, + 547, + 715 + ], + "type": "inline_equation", + "content": "1\\mathrm{cm}" + }, + { + "bbox": [ + 305, + 641, + 547, + 715 + ], + "type": "text", + "content": " threshold. While all methods' performances decrease with overlap, our method retains significantly more robustness. Classical keypoint matching (SIFT + BF, Fast GR," + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13087" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 55, + 70, + 277, + 187 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 277, + 187 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 277, + 187 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 277, + 187 + ], + "type": "table", + "html": "
MethodPose Recall by Threshold
5°, 10cm10°, 20cm15°, 30cm
SIFT + BF [10, 23]16.8318.4220.40
Fast GR [34, 46]39.0742.9644.68
GR [5]43.0847.6750.03
Object Track37.4143.2145.19
GeoTrans [31]67.1174.5776.93
SG + BF [10, 11, 36]71.1380.1181.71
Ours (w/o keypoints)56.4163.6766.99
Ours (w/ GeoTrans)68.3276.8079.35
Ours (w/ SG + BF)74.2584.5887.06
", + "image_path": "e8edc373384b7b333a0c213a65d5682773f680f28796862a345d230c37bc5084.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 67, + 267, + 265, + 392 + ], + "blocks": [ + { + "bbox": [ + 46, + 190, + 288, + 257 + ], + "lines": [ + { + "bbox": [ + 46, + 190, + 288, + 257 + ], + "spans": [ + { + "bbox": [ + 46, + 190, + 288, + 257 + ], + "type": "text", + "content": "Table 1. Pose Recall results on ScanNet [9] validation images. Combined with state-of-the-art feature matching, our method outperforms various classical and learning-based baselines. Our approach complements both Geometric Transformer [31] (Ours (w/ GeoTrans)) and SuperGlue [36] (Ours (w/ SG + BF)) feature matches, notably improving pose recall." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 67, + 267, + 265, + 392 + ], + "lines": [ + { + "bbox": [ + 67, + 267, + 265, + 392 + ], + "spans": [ + { + "bbox": [ + 67, + 267, + 265, + 392 + ], + "type": "table", + "html": "
MethodRecall by Overlap %
≤ 10(10,30)≥ 30
SIFT + BF [10,23]1.302.2033.59
Fast GR [34,46]0.0013.9769.04
GR [5,34]0.6520.3674.62
Object Track8.4431.5458.86
GeoTrans [31]22.0864.2793.11
SG + BF [10,11,36]24.0373.4595.95
Ours (w/o keypoints)32.4753.2980.31
Ours (w/ GeoTrans)35.0665.4794.42
Ours (w/ SG + BF)45.4581.4497.16
", + "image_path": "fec60ddaa3b53f7c5ae76095c1830e1f990f8687fdd370fea4f811d842b723c5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 396, + 288, + 430 + ], + "lines": [ + { + "bbox": [ + 46, + 396, + 288, + 430 + ], + "spans": [ + { + "bbox": [ + 46, + 396, + 288, + 430 + ], + "type": "text", + "content": "Table 2. Recall at " + }, + { + "bbox": [ + 46, + 396, + 288, + 430 + ], + "type": "inline_equation", + "content": "15^{\\circ}" + }, + { + "bbox": [ + 46, + 396, + 288, + 430 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 396, + 288, + 430 + ], + "type": "inline_equation", + "content": "30\\mathrm{cm}" + }, + { + "bbox": [ + 46, + 396, + 288, + 430 + ], + "type": "text", + "content": " by overlap percentage. Our method significantly outperforms strong baselines in challenging low-overlap frame pairs, almost doubling recall for overlap " + }, + { + "bbox": [ + 46, + 396, + 288, + 430 + ], + "type": "inline_equation", + "content": "\\leq 10\\%" + }, + { + "bbox": [ + 46, + 396, + 288, + 430 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 65, + 442, + 269, + 597 + ], + "blocks": [ + { + "bbox": [ + 65, + 442, + 269, + 597 + ], + "lines": [ + { + "bbox": [ + 65, + 442, + 269, + 597 + ], + "spans": [ + { + "bbox": [ + 65, + 442, + 269, + 597 + ], + "type": "image", + "image_path": "d970bedea8200bb3a2326b3b67f1fe102e89862b86d5356040df18c8774cddbe.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 599, + 287, + 642 + ], + "lines": [ + { + "bbox": [ + 46, + 599, + 287, + 642 + ], + "spans": [ + { + "bbox": [ + 46, + 599, + 287, + 642 + ], + "type": "text", + "content": "Figure 6. Change of pose performance by overlap percentage on ScanNet [9] validation pairs. Our method is significantly more robust against decreasing overlap compared to various classical and learned baselines." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": "SIFT) suffer strongly in low-overlap scenarios, which are challenging for hand-crafted descriptors. The learned approaches of GeoTrans and SG+BF maintain some performance with decreasing overlap but still suffer strongly. Our" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": "incorporation of object-based reasoning effectively complements keypoint matching while significantly improving robustness to low overlap, obtaining a performance improvement from " + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "24.03\\%" + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "45.45\\%" + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": " for overlap " + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "inline_equation", + "content": "\\leq 10\\%" + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 124, + 545, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 124, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 304, + 124, + 545, + 196 + ], + "type": "text", + "content": "Qualitative Results. Figure 5 shows examples of low-overlap registration on ScanNet [9] frame pairs. Due to the minimal overlap, traditional keypoint matching cannot find sufficient correspondences, while our object grounding enables accurate camera pose estimation through indirect object constraints." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 202, + 522, + 216 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 202, + 522, + 216 + ], + "spans": [ + { + "bbox": [ + 305, + 202, + 522, + 216 + ], + "type": "text", + "content": "4.2. Registration of RGB-D SLAM Sequences" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 221, + 545, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 221, + 545, + 280 + ], + "spans": [ + { + "bbox": [ + 304, + 221, + 545, + 280 + ], + "type": "text", + "content": "We further demonstrate the effectiveness of our method in an RGB-D SLAM setting, registering sequences of frames in room-scale scenes. In particular, we evaluate challenging, low-frame-rate settings that reflect common non-expert capture such as fast camera motion." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 281, + 545, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 281, + 545, + 341 + ], + "spans": [ + { + "bbox": [ + 304, + 281, + 545, + 341 + ], + "type": "text", + "content": "For optimization, we use a global pose graph optimization over pairwise registration similar to the multi-way global registration of [5, 6, 47], as it offers an off-the-shelf, simple, and robust solution. We refer to our supplementary material for further optimization details." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 345, + 545, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 345, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 304, + 345, + 545, + 464 + ], + "type": "text", + "content": "Datasets and Evaluation Metrics. We evaluate our method on TUM-RGBD [39] and ScanNet [9] scenes. For ScanNet, we use a set of 12 scenes from the validation and test sets, having a wide range of sizes, environments, and camera trajectories. We sample every 30th frame for TUM-RGBD (1Hz) and every 20th frame in ScanNet (1.5Hz). All methods use the same hyperparameters for each dataset, respectively. To evaluate registration quality, we use the standard root mean squared trajectory error used for TUM-RGBD evaluation [39]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 469, + 545, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 469, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 304, + 469, + 545, + 612 + ], + "type": "text", + "content": "Baselines. We implement various popular feature descriptors from cutting-edge SLAM systems in global pose graph optimization setups and thoroughly tune the hyperparameters of each method to support low-frame-rate sequences. We compare against Redwood [5], a global multi-way pose graph that uses FPFH [34] features. We also deploy robust RGB-D SIFT features used in BundleFusion [10, 23] in a global pose graph optimization, creating the SIFT + BF baseline. Finally, we use the state-of-the-art SuperGlue matches of SuperPoint features with BundleFusion's sparse 3D optimization, SG + BF, as described in Section 4.1." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 617, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 712 + ], + "type": "text", + "content": "Quantitative Results. In Table 3, we compare our method to state-of-the-art traditional and learned SLAM approaches. Across both TUM-RGBD and ScanNet validation and test scenes [9, 39], our object-grounded approach enables improved camera pose estimation. In particular, for the larger-scale ScanNet scenes, ObjectMatch achieves notable reductions in trajectory error, of " + }, + { + "bbox": [ + 304, + 617, + 545, + 712 + ], + "type": "inline_equation", + "content": "38\\%" + }, + { + "bbox": [ + 304, + 617, + 545, + 712 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 617, + 545, + 712 + ], + "type": "inline_equation", + "content": "49\\%" + }, + { + "bbox": [ + 304, + 617, + 545, + 712 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 617, + 545, + 712 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 304, + 617, + 545, + 712 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 617, + 545, + 712 + ], + "type": "inline_equation", + "content": "44\\%" + }, + { + "bbox": [ + 304, + 617, + 545, + 712 + ], + "type": "text", + "content": " in scenes 0423, 0430, 0718, and 0773." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13088" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 70, + 541, + 118 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 541, + 118 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 541, + 118 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 541, + 118 + ], + "type": "table", + "html": "
MethodTUM Scene @ 1HzScanNet Scene @ 1.5Hz
fr1/deskfr1/xyzfr2/xyzfr2/deskfr3/office001100810169020704230430046104940718076007730805
SIFT + BF [10,23]30.603.492.73157.0238.01222.49108.03126.57104.64160.69248.0020.9033.9881.4838.0322.4786.80
Redwood [5,34,47]3.101.971.88140.17175.64164.0150.9647.2447.1788.43156.7610.479.77133.536.5543.19253.39
SG + BF [10,11,36]3.011.921.815.305.1413.886.778.095.627.6735.973.317.1613.426.3712.4312.98
Ours (w/ SG + BF)3.011.921.815.264.9212.535.736.855.324.7518.232.676.856.705.887.0010.86
", + "image_path": "ce8f08b2ef19c6461899ee6dfecfb9cf6ebbcb2117ee279f9e07c4c866a052ff.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 121, + 546, + 156 + ], + "lines": [ + { + "bbox": [ + 47, + 121, + 546, + 156 + ], + "spans": [ + { + "bbox": [ + 47, + 121, + 546, + 156 + ], + "type": "text", + "content": "Table 3. ATE RMSE values (cm) on room-scale TUM-RGBD [39] and ScanNet [9] test scenes. We evaluate at challenging 1 FPS and 1.5 FPS by sampling every 30th and 20th frame from TUM-RGBD and ScanNet, respectively. ObjectMatch outperforms strong baselines leveraging classical and learning-based SLAM systems." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 50, + 175, + 282, + 430 + ], + "blocks": [ + { + "bbox": [ + 50, + 175, + 282, + 430 + ], + "lines": [ + { + "bbox": [ + 50, + 175, + 282, + 430 + ], + "spans": [ + { + "bbox": [ + 50, + 175, + 282, + 430 + ], + "type": "image", + "image_path": "c64e4018aceca06ceec3712dd2fdac8ab735dda3de5ffca06545abbeb9227de9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 434, + 287, + 522 + ], + "lines": [ + { + "bbox": [ + 46, + 434, + 287, + 522 + ], + "spans": [ + { + "bbox": [ + 46, + 434, + 287, + 522 + ], + "type": "text", + "content": "Figure 7. Qualitative comparison for registration of ScanNet scenes [9], in comparison with only feature matching SG + BF [10, 11, 36]. All results are visualized by volumetric fusion of the respective RGB-D frames with the estimated poses. Our optimized poses produce more accurate, clean, and consistent reconstruction, due to incorporating the object-based information in our global optimization. Example loop closures with matching object correspondences are shown on the right." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 546, + 287, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 546, + 287, + 654 + ], + "spans": [ + { + "bbox": [ + 46, + 546, + 287, + 654 + ], + "type": "text", + "content": "Qualitative Results. In Figure 7, we show ScanNet [9] scene reconstructions using estimated camera poses. RGB-D frames are fused using a scalable TSDF volumetric fusion [8,47]. Our approach complements state-of-the-art feature matching [10, 11, 36], enabling more accurate and consistent reconstructions using our indirect correspondences, in comparison to feature matching alone. We refer to the supplemental for additional qualitative sequence registration results." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 287, + 715 + ], + "type": "text", + "content": "Limitations and Future Work. ObjectMatch shows the capability of incorporating object semantics into global pose optimization; however, various limitations remain. In particular, our approach leverages higher-level seman" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 175, + 547, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 175, + 547, + 368 + ], + "spans": [ + { + "bbox": [ + 304, + 175, + 547, + 368 + ], + "type": "text", + "content": "tic correspondence given by objects for registration; however, in scenarios where object matches do not exist in frame views, we cannot leverage this constraint and instead use only feature matches. Further leveraging background information regarding shared structures seen from different views would provide additional semantic correspondence information to further complement registration. Additionally, while estimating canonical correspondences with objects can provide significant view information, object alignments can be somewhat coarse, while reasoning at a finer-grained level of object parts would provide more precise reasoning. Finally, we believe future work can reduce our method's system complexity by adopting a joint multiframe SLAM optimization with fewer hyperparameters as well as a unified end-to-end architecture for learning object correspondences and identification together." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 377, + 378, + 389 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 377, + 378, + 389 + ], + "spans": [ + { + "bbox": [ + 306, + 377, + 378, + 389 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 397, + 545, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 397, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 304, + 397, + 545, + 613 + ], + "type": "text", + "content": "We have presented ObjectMatch which introduces a new paradigm for incorporating object semantics into camera pose registration. We leverage indirect, canonical object correspondences established with normalized object coordinates for camera pose optimization aspects in RGB-D registration and SLAM. To obtain these correspondences, we propose two multi-modal neural networks for object recognition with normalized object coordinate prediction and object identification. ObjectMatch operates in tandem with state-of-the-art feature matching in a joint Gauss-Newton optimization, with its object grounding enabling registration to handle frame pairs with very low to no shared visibility. As a result, ObjectMatch significantly improves the state-of-the-art feature matching in both pairwise and sequence registration of RGB-D frames, particularly in the challenging low-overlap regime. Overall, we hope our method opens up new possibilities in the context of leveraging semantic information for camera pose estimation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 617, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 546, + 714 + ], + "type": "text", + "content": "Acknowledgements. This work was supported by the ERC Starting Grant Scan2CAD (804724), the Bavarian State Ministry of Science and the Arts coordinated by the Bavarian Research Institute for Digital Transformation (bidt), the German Research Foundation (DFG) Grant \"Making Machine Learning on Static and Dynamic 3D Data Practical\", and the German Research Foundation (DFG) Research Unit \"Learning and Simulation in Visual Computing.\"" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13089" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Armen Avetisyan, Manuel Dahnert, Angela Dai, Manolis Savva, Angel X Chang, and Matthias Nießner. Scan2cad: Learning cad model alignment in rgb-d scans. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 2614-2623, 2019. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 146, + 288, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 146, + 288, + 190 + ], + "spans": [ + { + "bbox": [ + 53, + 146, + 288, + 190 + ], + "type": "text", + "content": "[2] Armen Avetisyan, Angela Dai, and Matthias Nießner. End-to-end cad model retrieval and 9dof alignment in 3d scans. In Proceedings of the IEEE/CVF International Conference on computer vision, pages 2551-2560, 2019. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 191, + 287, + 244 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 191, + 287, + 244 + ], + "spans": [ + { + "bbox": [ + 53, + 191, + 287, + 244 + ], + "type": "text", + "content": "[3] Vaibhav Bansal, Stuart James, and Alessio Del Bue. re-object: Jointly learning the foreground and background for object instance re-identification. In International Conference on Image Analysis and Processing, pages 402–413. Springer, 2019. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 246, + 288, + 268 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 246, + 288, + 268 + ], + "spans": [ + { + "bbox": [ + 53, + 246, + 288, + 268 + ], + "type": "text", + "content": "[4] G. Bradski. The OpenCV Library. Dr. Dobb's Journal of Software Tools, 2000. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 270, + 287, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 270, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 53, + 270, + 287, + 312 + ], + "type": "text", + "content": "[5] Sungjoon Choi, Qian-Yi Zhou, and Vladlen Koltun. Robust reconstruction of indoor scenes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5556-5565, 2015. 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 314, + 287, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 314, + 287, + 357 + ], + "spans": [ + { + "bbox": [ + 53, + 314, + 287, + 357 + ], + "type": "text", + "content": "[6] Christopher Choy, Wei Dong, and Vladlen Koltun. Deep global registration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2514-2523, 2020. 2, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 358, + 287, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 358, + 287, + 379 + ], + "spans": [ + { + "bbox": [ + 53, + 358, + 287, + 379 + ], + "type": "text", + "content": "[7] Christopher Choy, Jaesik Park, and Vladlen Koltun. Fully convolutional geometric features. In ICCV, 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 380, + 287, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 287, + 424 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 287, + 424 + ], + "type": "text", + "content": "[8] Brian Curless and Marc Levoy. A volumetric method for building complex models from range images. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, pages 303-312. ACM, 1996. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 426, + 287, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 426, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 53, + 426, + 287, + 479 + ], + "type": "text", + "content": "[9] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 2, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 480, + 287, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 480, + 287, + 534 + ], + "spans": [ + { + "bbox": [ + 48, + 480, + 287, + 534 + ], + "type": "text", + "content": "[10] Angela Dai, Matthias Nießner, Michael Zollhöfer, Shahram Izadi, and Christian Theobalt. Bundlefusion: Real-time globally consistent 3d reconstruction using on-the-fly surface reintegration. ACM Transactions on Graphics (ToG), 36(4):1, 2017. 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 536, + 287, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 287, + 590 + ], + "type": "text", + "content": "[11] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 224-236, 2018. 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 591, + 287, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 591, + 287, + 646 + ], + "spans": [ + { + "bbox": [ + 48, + 591, + 287, + 646 + ], + "type": "text", + "content": "[12] Andreas Eitel, Jost Tobias Springenberg, Luciano Spinello, Martin Riedmiller, and Wolfram Burgard. Multimodal deep learning for robust rgb-d object recognition. In 2015 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 681-687. IEEE, 2015. 3, 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 647, + 287, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 680 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 680 + ], + "type": "text", + "content": "[13] Hugo Germain, Vincent Lepetit, and Guillaume Bourmaud. Visual correspondence hallucination. In International Conference on Learning Representations, 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "text", + "content": "[14] Can Gümeli, Angela Dai, and Matthias Nießner. Roca: Robust cad model retrieval and alignment from a single image. In Proceedings of the IEEE/CVF Conference on Computer" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "type": "text", + "content": "Vision and Pattern Recognition, pages 4022-4031, 2022. 3, 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 95, + 547, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 95, + 547, + 138 + ], + "spans": [ + { + "bbox": [ + 308, + 95, + 547, + 138 + ], + "type": "text", + "content": "[15] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 3, 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 140, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 140, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 308, + 140, + 545, + 183 + ], + "type": "text", + "content": "[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3, 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 184, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 184, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 308, + 184, + 545, + 217 + ], + "type": "text", + "content": "[17] Elad Hoffer and Nir Ailon. Deep metric learning using triplet network. In International workshop on similarity-based pattern recognition, pages 84-92. Springer, 2015. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 217, + 545, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 217, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 308, + 217, + 545, + 270 + ], + "type": "text", + "content": "[18] Jiahui Huang, Shi-Sheng Huang, Haoxuan Song, and ShiMin Hu. Di-fusion: Online implicit 3d reconstruction with deep priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8932-8941, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 272, + 545, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 272, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 308, + 272, + 545, + 327 + ], + "type": "text", + "content": "[19] Shengyu Huang, Zan Gojcic, Mikhail Usvyatsov, Andreas Wieser, and Konrad Schindler. Predator: Registration of 3d point clouds with low overlap. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4267-4276, June 2021. 2, 5, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 327, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 327, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 308, + 327, + 545, + 403 + ], + "type": "text", + "content": "[20] Shahram Izadi, David Kim, Otmar Hilliges, David Molyneaux, Richard Newcombe, Pushmeet Kohli, Jamie Shotton, Steve Hodges, Dustin Freeman, Andrew Davison, et al. Kinectfusion: real-time 3d reconstruction and interaction using a moving depth camera. In Proceedings of the 24th annual ACM symposium on User interface software and technology, pages 559-568, 2011. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 404, + 545, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 404, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 308, + 404, + 545, + 459 + ], + "type": "text", + "content": "[21] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2117-2125, 2017. 3, 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 460, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 460, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 308, + 460, + 545, + 514 + ], + "type": "text", + "content": "[22] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 514, + 545, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 514, + 545, + 558 + ], + "spans": [ + { + "bbox": [ + 308, + 514, + 545, + 558 + ], + "type": "text", + "content": "[23] David G Lowe. Object recognition from local scale-invariant features. In Proceedings of the seventh IEEE international conference on computer vision, volume 2, pages 1150-1157. IEEE, 1999. 6, 7, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 559, + 547, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 559, + 547, + 613 + ], + "spans": [ + { + "bbox": [ + 308, + 559, + 547, + 613 + ], + "type": "text", + "content": "[24] Wei-Chiu Ma, Anqi Joyce Yang, Shenlong Wang, Raquel Urtasun, and Antonio Torralba. Virtual correspondence: Humans as a cue for extreme-view geometry. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15924-15934, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 614, + 545, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 614, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 308, + 614, + 545, + 658 + ], + "type": "text", + "content": "[25] Kevis-Kokitsi Maninis, Stefan Popov, Matthias Niesser, and Vittorio Ferrari. Vid2cad: Cad model alignment using multiview constraints from videos. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "text", + "content": "[26] Yuhang Ming, Xingrui Yang, and Andrew Calway. Object-augmented rgb-d slam for wide-disparity relocalisation. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 2203-2209. IEEE, 2021. 2, 3" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13090" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "type": "text", + "content": "[27] Norman Muller, Yu-Shiang Wong, Niloy J Mitra, Angela Dai, and Matthias Nießner. Seeing behind objects for 3d multi-object tracking in rgb-d sequences. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6071-6080, 2021. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 288, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 288, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 288, + 173 + ], + "type": "text", + "content": "[28] Raul Mur-Artal and Juan D Tardós. Orb-slam2: An open-source slam system for monocular, stereo, and rgb-d cameras. IEEE transactions on robotics, 33(5):1255-1262, 2017. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 175, + 288, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 288, + 218 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 288, + 218 + ], + "type": "text", + "content": "[29] Matthias Nießner, Michael Zollhöfer, Shahram Izadi, and Marc Stamminger. Real-time 3d reconstruction at scale using voxel hashing. ACM Transactions on Graphics (ToG), 32(6):1-11, 2013. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 220, + 288, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 220, + 288, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 220, + 288, + 319 + ], + "type": "text", + "content": "[30] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems 32, pages 8024-8035. Curran Associates, Inc., 2019. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 320, + 288, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 320, + 288, + 374 + ], + "spans": [ + { + "bbox": [ + 48, + 320, + 288, + 374 + ], + "type": "text", + "content": "[31] Zheng Qin, Hao Yu, Changjian Wang, Yulan Guo, Yuxing Peng, and Kai Xu. Geometric transformer for fast and robust point cloud registration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11143-11152, 2022. 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 376, + 288, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 376, + 288, + 419 + ], + "spans": [ + { + "bbox": [ + 48, + 376, + 288, + 419 + ], + "type": "text", + "content": "[32] Nikhila Ravi, Jeremy Reizenstein, David Novotny, Taylor Gordon, Wan-Yen Lo, Justin Johnson, and Georgia Gkioxari. Accelerating 3d deep learning with pytorch3d. arXiv:2007.08501, 2020.4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 421, + 288, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 421, + 288, + 475 + ], + "spans": [ + { + "bbox": [ + 48, + 421, + 288, + 475 + ], + "type": "text", + "content": "[33] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision, 115(3):211-252, 2015. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 477, + 288, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 477, + 288, + 521 + ], + "spans": [ + { + "bbox": [ + 48, + 477, + 288, + 521 + ], + "type": "text", + "content": "[34] Radu Bogdan Rusu, Nico Blodow, and Michael Beetz. Fast point feature histograms (fpfh) for 3d registration. In 2009 IEEE international conference on robotics and automation, pages 3212-3217. IEEE, 2009. 2, 6, 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 523, + 288, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 523, + 288, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 523, + 288, + 578 + ], + "type": "text", + "content": "[35] Renato F Salas-Moreno, Richard A Newcombe, Hauke Strasdat, Paul HJ Kelly, and Andrew J Davison. Slam++: Simultaneous localisation and mapping at the level of objects. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1352-1359, 2013. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 579, + 288, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 579, + 288, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 579, + 288, + 633 + ], + "type": "text", + "content": "[36] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4938–4947, 2020. 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 635, + 288, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 288, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 288, + 679 + ], + "type": "text", + "content": "[37] Thomas Schops, Torsten Sattler, and Marc Pollefeys. Bad slam: Bundle adjusted direct rgb-d slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 134-144, 2019. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 681, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 288, + 713 + ], + "type": "text", + "content": "[38] Yifei Shi, Kai Xu, Matthias Niessner, Szymon Rusinkiewicz, and Thomas Funkhouser. Planematch: Patch coplanarity prediction for robust rgb-d reconstruction. In Proceedings of the" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 553 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "European Conference on Computer Vision (ECCV), pages 750-766, 2018. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 96, + 547, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 547, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 547, + 140 + ], + "type": "text", + "content": "[39] J. Sturm, N. Engelhard, F. Endres, W. Burgard, and D. Cremers. A benchmark for the evaluation of rgb-d slam systems. In Proc. of the International Conference on Intelligent Robot Systems (IROS), Oct. 2012. 2, 7, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 141, + 545, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 185 + ], + "type": "text", + "content": "[40] Edgar Sucar, Shikun Liu, Joseph Ortiz, and Andrew J Davison. imap: Implicit mapping and positioning in real-time. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6229-6238, 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 186, + 545, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 229 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 229 + ], + "type": "text", + "content": "[41] Edgar Sucar, Kentaro Wada, and Andrew Davison. NodeSLAM: Neural object descriptors for multi-view shape reconstruction. In Proceedings of the International Conference on 3D Vision (3DV), 2020. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 231, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 231, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 308, + 231, + 545, + 285 + ], + "type": "text", + "content": "[42] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8922-8931, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 286, + 545, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 286, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 308, + 286, + 545, + 352 + ], + "type": "text", + "content": "[43] He Wang, Srinath Sridhar, Jingwei Huang, Julien Valentin, Shuran Song, and Leonidas J Guibas. Normalized object coordinate space for category-level 6d object pose and size estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2642-2651, 2019. 2, 3, 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 353, + 545, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 353, + 545, + 386 + ], + "spans": [ + { + "bbox": [ + 308, + 353, + 545, + 386 + ], + "type": "text", + "content": "[44] T Whelan, S Leutenegger, RF. Salas-Moreno, B Glocker, and AJ. Davison. Elasticfusion: Dense slam without a pose graph. In Robotics: Science and Systems, 2015. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 387, + 545, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 387, + 545, + 419 + ], + "spans": [ + { + "bbox": [ + 308, + 387, + 545, + 419 + ], + "type": "text", + "content": "[45] Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. Detector2. https://github.com/facebookresearch/detectron2, 2019.4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 421, + 545, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 421, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 308, + 421, + 545, + 453 + ], + "type": "text", + "content": "[46] Qian-Yi Zhou, Jaesik Park, and Vladlen Koltun. Fast global registration. In European conference on computer vision, pages 766-782. Springer, 2016. 2, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 455, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 455, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 308, + 455, + 545, + 487 + ], + "type": "text", + "content": "[47] Qian-Yi Zhou, Jaesik Park, and Vladlen Koltun. Open3D: A modern library for 3D data processing. arXiv:1801.09847, 2018. 5, 6, 7, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 488, + 545, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 488, + 545, + 553 + ], + "spans": [ + { + "bbox": [ + 308, + 488, + 545, + 553 + ], + "type": "text", + "content": "[48] Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hujun Bao, Zhaopeng Cui, Martin R Oswald, and Marc Pollefeys. Nice-slam: Neural implicit scalable encoding for slam. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12786-12796, 2022. 2" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "13091" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/ObjectStitch_ Object Compositing With Diffusion Model/69193070-f3e5-4dd6-8743-b15c03b3e677_content_list.json b/2023/ObjectStitch_ Object Compositing With Diffusion Model/69193070-f3e5-4dd6-8743-b15c03b3e677_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..184b7ccf73b646dc19e6c98b2f846eb60c9f7b0c --- /dev/null +++ b/2023/ObjectStitch_ Object Compositing With Diffusion Model/69193070-f3e5-4dd6-8743-b15c03b3e677_content_list.json @@ -0,0 +1,1504 @@ +[ + { + "type": "text", + "text": "ObjectStitch: Object Compositing with Diffusion Model", + "text_level": 1, + "bbox": [ + 200, + 130, + 769, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yizhi Song $^{1*}$ , Zhifei Zhang $^{2}$ , Zhe Lin $^{2}$ , Scott Cohen $^{2}$ , Brian Price $^{2}$ , Jianming Zhang $^{2}$ , Soo Ye Kim $^{2}$ , Daniel Aliaga $^{1}$ Purdue University $^{1}$ , Adobe Research $^{2}$", + "bbox": [ + 218, + 179, + 750, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}\\{ \\text{song630,aliaga} \\} @$ urdue.edu, $^{2}\\{ \\text{zhang,zlin,scohen,bprice, jianmzha,sooyek} \\} @$ adobe.com", + "bbox": [ + 117, + 234, + 844, + 252 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/22213a7b75764b52e031010cdd4d202a63f5ed7ff1196f9033550863a307d181.jpg", + "image_caption": [ + "(a) copy-paste", + "Figure 1. Example results of object compositing with (a) copy-and-paste scheme, (b) traditional compositing pipeline and (c) ours (Object-Stitch). Traditional compositing pipeline is done with best possible off-the-shelf models including foreground/background color harmonization [19,47], poisson blending [34], and shadow synthesis [42]. ObjectStitch achieves more realistic results, and can address geometry correction, harmonization, shadow generation, and view synthesis all-in-one while preserving similar appearance to the reference object." + ], + "image_footnote": [], + "bbox": [ + 83, + 316, + 215, + 525 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cfd5085ea9132eb738ebc7a687b25f8531b86b2607301cc275614fd8d0fc339e.jpg", + "image_caption": [ + "(b) traditional compositing pipeline" + ], + "image_footnote": [], + "bbox": [ + 217, + 316, + 349, + 419 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/588dbed269c6232e2cee162781a76218dfafae91d8f5ccc6ae21ed7a2750ced8.jpg", + "image_caption": [ + "(c) ours" + ], + "image_footnote": [], + "bbox": [ + 349, + 316, + 483, + 419 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5b6e3e861229a4d5e838350ba8389cc3452f4f619b8265e23db05db68dae1317.jpg", + "image_caption": [ + "(a) copy-paste" + ], + "image_footnote": [], + "bbox": [ + 486, + 316, + 620, + 419 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/235043b1565a429c9e4344fbe0911f02383e936001c2df3ddd3c7482d2f6ceec.jpg", + "image_caption": [ + "(b) traditional compositing pipeline" + ], + "image_footnote": [], + "bbox": [ + 620, + 316, + 754, + 419 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ac24a6ff2324588107bb51f60953fff77c50e285fc565a9ec760772b3f6ef057.jpg", + "image_caption": [ + "(c) ours" + ], + "image_footnote": [], + "bbox": [ + 754, + 316, + 887, + 419 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 614, + 313, + 630 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Object compositing based on 2D images is a challenging problem since it typically involves multiple processing stages such as color harmonization, geometry correction and shadow generation to generate realistic results. Furthermore, annotating training data pairs for compositing requires substantial manual effort from professionals, and is hardly scalable. Thus, with the recent advances in generative models, in this work, we propose a self-supervised framework for object compositing by leveraging the power of conditional diffusion models. Our framework can holistically address the object compositing task in a unified model, transforming the viewpoint, geometry, color and shadow of the generated object while requiring no manual labeling. To preserve the input object's characteristics, we introduce a content adaptor that helps to maintain categori", + "bbox": [ + 73, + 647, + 473, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "cal semantics and object appearance. A data augmentation method is further adopted to improve the fidelity of the generator. Our method outperforms relevant baselines in both realism and faithfulness of the synthesized result images in a user study on various real-world images.", + "bbox": [ + 496, + 602, + 893, + 678 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 708, + 632, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Image compositing is an essential task in image editing that aims to insert an object from a given image into another image in a realistic way. Conventionally, many sub-tasks are involved in composing an object to a new scene, including color harmonization [6, 7, 19, 51], relighting [52], and shadow generation [16, 29, 43] in order to naturally blend the object into the new image. As shown in Tab. 1, most previous methods [6, 7, 16, 19, 28, 43] focus on a single sub-task required for image compositing. Consequently, they must be appropriately combined to obtain a composite image where the input object is re-synthesized to have the", + "bbox": [ + 496, + 734, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Work done during internship at Adobe.", + "bbox": [ + 99, + 886, + 320, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "18310", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/7bc0c3bed694450bbe9b4af35643afcf845469858136727e06fd223be8b25e6d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodGeometryLightShadowView
ST-GAN [28]✓XXX
SSH [19]X✓XX
DCCF [51]X✓XX
SSN [43]XX✓X
SGRNet [16]XX✓X
GCC-GAN [5]✓✓XX
Ours✓✓✓✓
", + "bbox": [ + 91, + 88, + 450, + 220 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Table 1. Prior works only focus on one or two aspects of object compositing, and they cannot synthesize novel views. In contrast, our model can address all perspectives as listed.", + "bbox": [ + 75, + 224, + 468, + 267 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "color, lighting and shadow that is consistent with the background scene. As shown in Fig. 1, results produced in this way still look unnatural, partly due to the viewpoint of the inserted object being different from the overall background.", + "bbox": [ + 75, + 287, + 467, + 349 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Harmonizing the geometry and synthesizing novel views have often been overlooked in 2D image compositing, which require an accurate understanding of both the geometry of the object and the background scene from 2D images. Previous works [15, 21, 22] handle 3D object compositing with explicit background information such as lighting positions and depth. More recently, [28] utilize GANs to estimate the homography of the object. However, this method is limited to placing furniture in indoor scenes. In this paper, we propose a generic image compositing method that is able to harmonize the geometry of the input object along with color, lighting and shadow with the background image using a diffusion-based generative model.", + "bbox": [ + 75, + 351, + 467, + 547 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In recent years, generative models such as GANs [4, 10, 17, 20] and diffusion models [1, 14, 30, 32, 33, 38, 39] have shown great potential in synthesizing realistic images. In particular, diffusion model-based frameworks are versatile and outperform various prior methods in image editing [1, 23, 32] and other applications [11, 31, 35]. However, most image editing diffusion models focus on using text inputs to manipulate images [1, 3, 9, 23, 33], which is insufficient for image compositing as verbal representations cannot fully capture the details or preserve the identity and appearance of a given object image. There have been recent works [23, 39] focusing on generating diverse contexts while preserving the key features of the object; however, these models are designed for a different task than object compositing. Furthermore, [23] requires fine-tuning the model for each input object and [39] also needs to be finetuned on multiple images of the same object. Therefore they are limited for general object compositing.", + "bbox": [ + 75, + 549, + 467, + 823 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we leverage diffusion models to simultaneously handle multiple aspects of image compositing such as color harmonization, relighting, geometry correction and shadow generation. With image guidance rather than text guidance, we aim to preserve the identity and appearance of", + "bbox": [ + 75, + 824, + 467, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the original object in the generated composite image.", + "bbox": [ + 498, + 90, + 849, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Specifically, our model synthesizes a composite image given (i) a source object, (ii) a target background image, and (iii) a bounding box specifying the location to insert the object. The proposed framework consists of a content adaptor and a generator module: the content adaptor is designed to extract a representation from the input object containing both high-level semantics and low-level details such as color and shape; the generator module preserves the background scene while improving the generation quality and versatility. Our framework is trained in a fully self-supervised manner and no task-specific labeling is required at any point during training. Moreover, various data augmentation techniques are applied to further improve the fidelity and realism of the output. We evaluate our proposed method on a real-world dataset closely simulating real use cases for image compositing.", + "bbox": [ + 496, + 107, + 890, + 348 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are summarized as follows:", + "bbox": [ + 517, + 348, + 821, + 362 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We present the first diffusion model-based framework for generative object compositing that can handle multiple aspects of compositing such as viewpoint, geometry, lighting and shadow.", + "- We propose a content adaptor module which learns a descriptive multi-modal embedding from images, enabling image guidance for diffusion models.", + "- Our framework is trained in a self-supervised manner without any task-specific annotations, employing data augmentation techniques to improve the fidelity of generation.", + "- We collect a high-resolution real-world dataset for object compositing with diverse images, containing manually annotated object scales and locations." + ], + "bbox": [ + 517, + 367, + 890, + 579 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 594, + 640, + 609 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Image Compositing", + "text_level": 1, + "bbox": [ + 500, + 619, + 686, + 636 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image compositing is a challenging task in reference-guided image editing, where the object in a given foreground image is to be inserted into a background image. The generated composite image is expected to look realistic, with the appearance of the original object being preserved. Prior works often focus on an individual aspect of this problem, such as geometric correction [2, 28], image harmonization [6, 7, 51], image matting [50] and shadow generation [29].", + "bbox": [ + 496, + 643, + 890, + 779 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Lin et al. [28] propose an iterative GAN-based framework to correct geometric inconsistencies between object and background images. In their model, STNs [18] are integrated into the generator network to predict a series of warp updates. Although their work can improve the realism of the scene geometry in the composite image, it is limited to inserting furniture into indoor scenes. Azadi et al. [2] focus on binary composition, using a composition-decomposition", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "18311", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "network to capture interactions between a pair of objects.", + "bbox": [ + 76, + 90, + 454, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Image harmonization aims at minimizing the inconsistency between the input object and the background image. Traditional methods [24, 46] usually concentrate on obtaining color statistics and then transfer this information between foreground and background. Recent works seek to solve this problem via deep neural networks. [6, 7] reformulate image harmonization as a domain adaptation problem, while [19] converts it to a representation fusing problem and utilizes self-supervised training.", + "bbox": [ + 75, + 107, + 467, + 241 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Shadow synthesis is an effect that is often overlooked in previous image compositing methods although it is essential for generating realistic composite images. SGRNet [16] divides this task into a shadow mask generation stage and a shadow filling stage. SSN [43], focusing on soft shadows, predicts an ambient occlusion map as cue for shadow generation.", + "bbox": [ + 75, + 243, + 467, + 348 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "There are also works simultaneously addressing multiple sub-problems of image compositing. Chen et al. [5] develop a system including multiple network modules to produce plausible images with geometric, color and boundary consistency. However, the pose of the generated object is constrained by the mask input and the model cannot generalize to non-rigid objects such as animals. Our proposed model also handles the image composition problem in a unified manner, generating foreground objects which are harmonious and geometrically consistent with the background while synthesizing novel views and shadows.", + "bbox": [ + 75, + 349, + 467, + 515 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Guided Image Synthesis", + "text_level": 1, + "bbox": [ + 76, + 527, + 299, + 544 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recent years have witnessed great advances in diffusion models. Diffusion models are a family of deep generative models based on several predominant works [14, 44, 45], defined with two Markov chains. In the forward process, they gradually add noise to the data, and in the reverse process, they learn to recover the data from noise. Following the growth of research in this area, diffusion models have shown their potential in a great number of applications, such as text-to-image generation [9, 33, 38], image editing [1, 23, 32] and guided image synthesis [30].", + "bbox": [ + 75, + 551, + 467, + 703 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Text-to-image synthesis has been a popular research topic over the past few years. Stable Diffusion [38] made a significant contribution to this task and it reduces the computational cost by applying the diffusion model in the latent space. Other works explore ways of more flexible and controllable text-driven image editing. Avrahami et al. [1] design an algorithm in their Blended Diffusion framework to fuse noised versions of the input with the local text-guided diffusion latent, generating smooth transition. Similarly, GLIDE [33] is also capable of handling text-guided inpainting after fine-tuning on this specific task. Some methods give users more straightforward control on images. SDEdit [32] allows stroke painting in images, blending user input", + "bbox": [ + 75, + 704, + 467, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "into the image by adding noise to the input and denoising through a stochastic differential equation. SDG [30] injects Semantic Diffusion Guidance at each iteration of the generation and can provide multi-modal guidance.", + "bbox": [ + 496, + 90, + 890, + 150 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While the versatility and generation quality of diffusion models have been repeatedly demonstrated, preserving object appearance remains a very challenging problem in image editing. Ruiz et al. [39] address this problem by optimizing their model using a reconstruction loss and a class-specific prior preservation loss. Though their approach can preserve the details of the object, multiple images of the same object are required to fine-tune their model on this target object. Kawar et al. [23] handle this issue by interpolating the initial text embedding and the optimized embedding for reconstruction, but their method is mainly designed for applying edits on objects, not for generating the same object in different contexts. In this paper, we propose the first generative object compositing framework based on diffusion models, which generates the harmonized and geometrically-corrected subject with a novel view and shadow. Furthermore, the characteristics of the original object is preserved in the synthesized composite image.", + "bbox": [ + 496, + 151, + 890, + 422 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Proposed Method", + "text_level": 1, + "bbox": [ + 500, + 436, + 671, + 454 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We define the generative object compositing problem as: Given an input triplet $(I_o, I_{bg}, M)$ that consists of an object image $I_o \\in \\mathbb{R}^{H_s \\times W_s \\times 3}$ , a background image $I_{bg} \\in \\mathbb{R}^{H_t \\times W_t \\times 3}$ , and its associated binary mask $M \\in \\mathbb{R}^{H_t \\times W_t \\times 1}$ with the desired object location set to 0 and the rest to 1, the goal is to compose the input object into the masked area $I_{bg} \\otimes M$ . $M$ is considered as a soft constraint of the location and scale of the composited object. The output generated image should look realistic, while the appearance of the object is preserved. Our problem setting is different from text-guided image generation and inpainting in that the condition input is a reference object image rather than a text prompt. Inspired by the success of text-guided diffusion models [37, 38, 40], which inject text conditioning into the diffusion architecture, we design our method for generative compositing to leverage such pretrained models.", + "bbox": [ + 496, + 462, + 890, + 703 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "An overview of our framework is shown in Fig. 2. It consists of an object image encoder extracting semantic features from the object and a conditional diffusion generator. To leverage the power of pretrained text-to-image diffusion models, we introduce a content adaptor that can bridge the gap between the object encoder and conditional generator by transforming a sequence of visual tokens to a sequence of text tokens to overcome the domain gap between image and text. This design further helps to preserve the object appearance. We propose a two-stage training process: in the first stage, the content adaptor is trained on large image/text pairs to maintain high-level semantics of the object; in the second stage, it is trained in the context of diffusion", + "bbox": [ + 496, + 704, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "18312", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/07cd73ec57e3862f79bb3a725a0012186561a38f2767aac1c67b8d92bf5e52a4.jpg", + "image_caption": [ + "Figure 2. System pipeline. Our framework consists of a content adaptor and a generator (a pretrained text-to-image diffusion model). The input image $I_{o}$ is fed into a ViT and the adaptor which produces a descriptive embedding. At the same time the background image $I_{bg}$ is taken as input by the diffusion model. At each iteration during the denoising stage, we apply the mask $M$ on the generated image $I_{out}$ , so that the generator only denoises the masked area $I_{out} \\otimes M$ ." + ], + "image_footnote": [], + "bbox": [ + 98, + 88, + 867, + 241 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "generator to encode key identity features of the object by encouraging the visual reconstruction of the object in the original image. Lastly, the generator module is fine-tuned on the embedding produced by the adaptor through cross attention blocks. All stages our trained in a self-supervised manner to avoid expensive annotation for obtaining object compositing training data.", + "bbox": [ + 75, + 313, + 473, + 420 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Generator", + "text_level": 1, + "bbox": [ + 76, + 441, + 194, + 455 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As depicted in Fig. 2, we leverage a pretrained text-to-image diffusion model architecture and modify it for the compositing task by (i) introducing a mask in the input, and (ii) adjusting the U-Net input to contain the original background image outside the hole and noise inside the hole with mask blending. In order to condition the model on the guidance embedding $E$ , an attention mechanism is applied as:", + "bbox": [ + 75, + 468, + 470, + 575 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {S o f t m a x} \\left(\\frac {\\left(\\boldsymbol {W} _ {Q} E _ {x}\\right) \\left(\\boldsymbol {W} _ {K} E\\right) ^ {T}}{\\sqrt {d}}\\right) \\boldsymbol {W} _ {V} E = \\boldsymbol {A} \\boldsymbol {V} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 604, + 468, + 641 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $E_{x}$ is an intermediate representation of the denoising autoencoder and $W_{Q} \\in \\mathbb{R}^{d \\times d_{x}}$ , $W_{K} \\in \\mathbb{R}^{d \\times d_{e}}$ and $W_{V} \\in \\mathbb{R}^{d \\times d_{e}}$ are embedding matrices. The background outside the mask area should be perfectly preserved for generative object compositing. Thus, we use the input mask $M$ for blending the background image $I_{bg}$ with the generated image $I_{out}$ . As a result, the generator only denoises the masked area $I_{out} \\otimes M$ . To use this model on our task, a straightforward way (a baseline) would be to apply image captioning on the object image and feeding the resulting caption directly to the diffusion model as the condition. However, text embedding cannot capture fine grained visual details. Therefore, in order to directly leverage the given object image, we introduce the content adaptor to transform the visual features from a pretrained visual encoder to text features (tokens) to use as conditioning for the generator.", + "bbox": [ + 75, + 657, + 472, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/af112cf73253444a877283eaf3708010be628359e7496df8e70d58a36a5dbea1.jpg", + "image_caption": [ + "Figure 3. Structure of the Content Adaptor. In the first stage, it is trained on a large dataset of image-caption pairs to learn multimodal sequential embeddings containing high-level semantics. In the second stage, it is fine-tuned under the diffusion framework to learn to encode identity features in adaptive embedding." + ], + "image_footnote": [], + "bbox": [ + 511, + 310, + 877, + 518 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Content Adaptor", + "text_level": 1, + "bbox": [ + 500, + 606, + 668, + 623 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To prevent the loss of key identity information, we use an image encoder instead of a text encoder to produce the embedding from the input object image. However, the image embedding cannot be effectively utilized by the diffusion model for two reasons:", + "bbox": [ + 496, + 631, + 893, + 705 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- The image embedding $\\widetilde{E}$ and the text embedding $E$ are from different domains. Since the diffusion model was trained on $E$ , it cannot generate meaningful contents from image embedding sequence;", + "- A mismatch in the dimensions of $\\widetilde{E} \\in \\mathbb{R}^{k \\times 257 \\times 1024}$ and $E \\in \\mathbb{R}^{k \\times 77 \\times 768}$ , where $k$ is batch-size." + ], + "bbox": [ + 517, + 710, + 890, + 801 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Therefore, based on the above observations, we develop a sequence-to-sequence translator architecture as shown in Fig. 3. Given an image-caption pair as an input tuple $(I,t)$ , we employ two pretrained ViT-L/14 encoders $C_t$ , $C_i$ from CLIP [36] to produce text embedding $E = C_t(t)$ and image embedding $\\widetilde{E} = C_i(I)$ , respectively. The adaptor $T$", + "bbox": [ + 496, + 808, + 893, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "18313", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2c40277be9bbef09e48d76f757a40fa7941352eabd768d0de4773d2a5d6c5f0d.jpg", + "image_caption": [ + "Figure 4. Illustration of our synthetic data generation and data augmentation scheme. The top row shows the data generation process including perspective warping, random rotation, and random color shifting. The original image is used as both the input background and ground truth, while the perturbed object is fed into the adaptor. The bottom row shows crop and shift augmentations, which help to improve the generation quality and preserve object details." + ], + "image_footnote": [], + "bbox": [ + 106, + 88, + 439, + 263 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "consists of three components: a 1D convolutional layer, attention blocks [48] and an MLP, where the 1D convolution modifies the length of the embedding from 257 to 77, the MLP maps the embedding dimension from 1024 to 768, and the attention blocks bridge the gap between text domain and image domain. We design a two-stage optimization method to train this module, which is explained in Sec. 3.3.2.", + "bbox": [ + 75, + 376, + 470, + 482 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Self-supervised Framework", + "text_level": 1, + "bbox": [ + 76, + 492, + 326, + 508 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "There is no publicly available image compositing training dataset with annotations which is sufficient for training a diffusion model, and it is extremely challenging to manually annotate such data. Therefore, we propose a self-supervised training scheme and a synthetic data generation approach that simulate real-world scenarios. We also introduce a data augmentation method to enrich the training data as well as improve the robustness of our model.", + "bbox": [ + 75, + 515, + 468, + 636 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.1 Data Generation and Augmentation", + "text_level": 1, + "bbox": [ + 76, + 650, + 382, + 664 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training data generation. We collect our synthetic training data from Pixabay and use an object instance segmentation model [26] to predict panoptic segmentation and classification labels. We first filter the dataset by removing objects with very small or large sizes. Then, we apply spatial and color perturbations to simulate many real use-case scenarios where the input image and background image have different scene geometry and lighting conditions.", + "bbox": [ + 75, + 672, + 468, + 794 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The top row of Fig. 4 illustrates this process. Inspired by [8], we randomly perturb the four points of the object bounding box to apply projective transformation, followed by a random rotation within the range $[- \\theta, \\theta]$ ( $\\theta = 20^{\\circ}$ ) and color perturbation. The segmentation mask (perturbed in the same way as the image) is used to extract the object.", + "bbox": [ + 75, + 795, + 468, + 883 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This data synthesis pipeline is fully controllable and can", + "bbox": [ + 96, + 885, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "also be employed as data augmentation during training. Another key advantage is that it is free of manual labeling, since the original image is used as the ground truth. We use the bounding box as the mask as it not only fully covers the object, but also extends to its neighboring area (providing room for shadow generation). We find that it is flexible enough for the model to apply spatial transformations, synthesize novel views and generate shadows and reflection.", + "bbox": [ + 496, + 90, + 890, + 212 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Real-world evaluation data. Given that there is no existing dataset specifically for our task, we manually collect a new dataset, closely simulating real-world use cases, as an evaluation benchmark for object compositing. The dataset consists of 503 pairs of common objects (including both rigid and non-rigid objects such as vehicles and wildlife) and diverse background images (covering both indoor and outdoor scenes). It also contains challenging cases where there is a large mismatch of lighting conditions or viewpoints between foreground and background. The dataset images are collected from Pixabay. The labeling procedure closely simulate the real scenarios where the input object is placed at a target location in the background image and then scaled at the user's will. The compositing region is determined as a loose bounding box around the object.", + "bbox": [ + 496, + 215, + 892, + 443 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Data augmentation. Inspired by [25], we introduce random shift and crop augmentations during training, while ensuring that the foreground object is always contained in the crop window. This process is illustrated in the bottom row of Fig. 4. Applying this augmentation method for both training and inference results in a notable improvement in the realism of the generated results. Quantitative results are provided in Sec. 4.4.", + "bbox": [ + 496, + 446, + 893, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.2 Training", + "text_level": 1, + "bbox": [ + 500, + 579, + 614, + 594 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Content adaptor pretraining. We first pretrain the content adaptor to keep the semantics of the object by mapping the image embedding to text embedding. At this first stage, we optimize the content adaptor on a sequence-to-sequence translation task, which learns to project the image embedding into a multi-modal space. During training, $C_t$ , $C_i$ are frozen, and the translator is trained on 3,203,338 image-caption pairs from a filtered LAION dataset [41].", + "bbox": [ + 496, + 603, + 890, + 723 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given the input image embedding $\\vec{E}$ , we use the text embedding $E$ as target, the objective function of this translation task is defined as:", + "bbox": [ + 496, + 724, + 890, + 768 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {d i s t}} = \\| T (\\tilde {E}) - E \\| _ {1}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 780, + 890, + 797 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $T(\\cdot)$ is the content adaptor.", + "bbox": [ + 500, + 810, + 723, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "However, the multi-modal embedding obtained solely from this first stage optimization mostly carry high-level semantics of the input image without much of texture details. Hence, we further refine $\\widetilde{E}$ to obtain $\\hat{E}$ , the adaptive embedding, for better appearance preservation.", + "bbox": [ + 496, + 825, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "18314", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Content adaptor fine-tuning. We further optimize the content adaptor to produce an adaptive embedding which maintains instance-level properties of the object. After pretraining, we insert the content adaptor to the pretrained diffusion framework by feeding the adaptive embedding as context to the attention blocks. Then, the diffusion model is frozen and the adaptor is trained using:", + "bbox": [ + 76, + 90, + 472, + 196 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {a d a p t}} = \\mathbb {E} _ {T, \\epsilon \\sim \\mathcal {N} (0, 1)} [ \\| \\epsilon - \\epsilon_ {\\theta} (I _ {t} \\circ M, t, T (\\widetilde {E})) \\| _ {2} ^ {2} ], \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 205, + 468, + 226 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where the content adaptor $T(\\cdot)$ is optimized. It is trained on our synthetic dataset filtered from Pixabay, containing 467,280 foreground and background image pairs for training and 25,960 pairs for validation.", + "bbox": [ + 76, + 234, + 468, + 296 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Generator fine-tuning. After the aforementioned two-stage training of the content adaptor is completed, we freeze the content adaptor and train the generator module after initializing the text-to-image diffusion model with pretrained weights. Crop and shift augmentations are applied in this process. Based on the Latent Diffusion model [38], the objective loss in our generator module is defined by:", + "bbox": [ + 75, + 300, + 468, + 405 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {g e n} = \\mathbb {E} _ {\\hat {E}, \\epsilon \\sim \\mathcal {N} (0, 1)} [ \\| \\epsilon - \\epsilon_ {\\theta} (I _ {t} \\circ M, t, \\hat {E}) \\| _ {2} ^ {2} ], \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 116, + 415, + 468, + 435 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $I$ is the input image, $I_{t}$ is a noisy version of $I$ at timestep $t$ , and $\\epsilon_{\\theta}$ denotes the denoising model which is optimized. In order to adapt the text-guided generation task to the setting of image compositing, we apply the input mask on the image at every time-step.", + "bbox": [ + 76, + 444, + 468, + 521 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 532, + 209, + 549 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Training Details", + "text_level": 1, + "bbox": [ + 76, + 556, + 238, + 573 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The first training stage of the content adaptor takes 15 epochs with a learning rate of $10^{-4}$ and batch size of 2048. The image size processed by ViT is $224 \\times 224$ . To set up the pipeline, we employed the pretrained image encoder and text encoder from CLIP [36]. The second training stage of the content adaptor takes 13 epochs with the learning rate of $2 \\times 10^{-5}$ and batch size of 512. We resize the images to 512 $\\times$ 512 and adapt the pretrained Stable Diffusion model [38] to our task. After the above process, the generator is trained for 20 epochs with the learning rate of $4 \\times 10^{-5}$ and batch size of 576. The input image size is the same as that in the second training stage. All training stages are conducted on 8 A100 GPUs with Adam optimizer.", + "bbox": [ + 75, + 580, + 468, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Quantitative Evaluation", + "text_level": 1, + "bbox": [ + 76, + 786, + 299, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To obtain quantitative results on a real dataset that simulates real-world use cases, where users drag and drop an object to a background image, we conduct a user study to compare to three baselines, i.e., copy-and-paste, pretrained stable diffusion model fine-tuned on BLIP [27], and a joint stable diffusion model integrating both BLIP", + "bbox": [ + 75, + 810, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7762c2ac838214911580c6ca9dcb9182697c5f410c8a7d739d199a38570455ec.jpg", + "image_caption": [ + "Pick from image A and image B the one that you think looks more realistic." + ], + "image_footnote": [], + "bbox": [ + 516, + 112, + 879, + 186 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d8c93f8d424c09f5d59ffa0015cf45ac8908a35a86fbd674ee179dacb39c85e1.jpg", + "image_caption": [ + "Which generated object do you think is most likely to be the same one from the guidance?", + "Figure 5. User study results. We conduct side-by-side comparisons between our method and one of baseline methods to quantify the generation quality in terms of realism and appearance preservation. The results show that our method outperforms baselines." + ], + "image_footnote": [], + "bbox": [ + 517, + 220, + 859, + 266 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and SDEdit [32]. Given the fact that there is no existing diffusion-based method addressing the same problem, we train the baselines on the same pretrained diffusion model using the same self-supervised training scheme and dataset as ours. To the best of our knowledge, after adapting to our setting, they are the closest baselines to our task. In the BLIP baseline, the content adaptor is replaced by BLIP, a state-of-the-art captioning model. The text embedding obtained from the BLIP caption is fed into the generator module. For the SDEdit baseline, we keep the BLIP model which will significantly improve the generation fidelity of a pretrained diffusion model. We set the noise strength to 0.65 in SDEdit, striking a balance between realism and faithfulness. Details are provided in the supplementary material.", + "bbox": [ + 496, + 332, + 890, + 544 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We perform a user study using our real testing dataset, which consists of 503 object-background pairs. For each example, we set up a side-by-side comparison between the results from one of the baselines and our method, respectively. The subject will be asked two questions: 1) which is more realistic and 2) which generated object is more close to the given object? We collected 1,494 votes from more than 170 users. As shown in Fig. 5, our method outperforms the other baselines for both questions. The higher preference rates demonstrate the effectiveness of the content adaptor and data augmentation method in improving generation fidelity. It also shows that our content adaptor is capable of maintaining details and attributes of the input objects, while not being constrained by the input when harmonizing objects to the background. In contrast, although SDEdit is better at preserving appearance than BLIP, it is constrained by the original pose, shape, and other attributes of the object, making the generated object less coherent with the background.", + "bbox": [ + 496, + 545, + 892, + 832 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Qualitative Evaluation", + "text_level": 1, + "bbox": [ + 500, + 845, + 709, + 861 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Fig. 6, we compare our method to various baselines. In the first two rows of Fig. 6, we place the same object", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "18315", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/82d0bf282400d201f0171608e61f21fa777e64396600d4c3d2837c47a876cca9.jpg", + "image_caption": [ + "Object", + "Target Image", + "Copy-and-paste", + "BLIP", + "SDEdit", + "Ours", + "Figure 6. Qualitative comparison on the real-world evaluation dataset. Our method can appropriately correct the viewpoint and geometry while harmonizing the object with the new background, generating realistic composite results. Moreover, lighting and shadows are naturally adjusted in the composite image. Compared to baselines using text guidance, our method can better preserve the object appearance." + ], + "image_footnote": [], + "bbox": [ + 80, + 109, + 890, + 843 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "18316", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b3f52d7ecf7a5f8e1a5fb9f08fd7e277ec235ca2bc55845860541dc7914ad9ac.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
AdaptorOptimizationAugmentationBLIP [27]FID ↓CLIP text score ↑CLIP image score ↑
✓✓✓✗15.433129.859497.0625
✗✓✓✗18.825429.828196.0000
✓✗✓✗16.138129.796996.8125
✓✓✗✗18.369829.781396.1875
✗✗✓✓17.900029.718895.8125
", + "bbox": [ + 145, + 88, + 821, + 186 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 2. Ablation study. We evaluate the effectiveness of the components listed in the first four columns: 1) whether the content adaptor is used; 2) whether the adaptor is optimized for appearance preservation; 3) whether crop augmentation is utilized in training; 4) if the adaptor module is replaced by BLIP to predict text embedding from an image. We use FID to assess the generation fidelity and modified CLIP score (explained in Sec. 4.4) to measure similarity of the guidance and the predicted image.", + "bbox": [ + 75, + 191, + 893, + 250 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "in two scenes with notable differences in the geometry. Results show that our model is capable of correcting geometric inconsistencies while preserving the characteristics of the object. The limitations of the baselines are also shown in Fig. 6: BLIP often fails to preserve the appearance; SDEdit preserves texture and pose better than the first baseline, but often cannot make appropriate geometry transformations.", + "bbox": [ + 75, + 262, + 472, + 369 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 378, + 230, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Tab. 2, we demonstrate the effectiveness of each key component in our ObjectStitch framework by ablating each of them. We use FID [13] and a modified CLIP [12, 36] score, which measure the fidelity and semantic matching performance between given and generated objects. Instead of directly adopting the CLIP score, which semantically matches the prompt and the image, we modify it to CLIP text score and CLIP image score that are defined as follows:", + "bbox": [ + 75, + 401, + 470, + 523 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {C} _ {t x t} = E \\left[ s \\cdot f \\left(I _ {p r e d}\\right) \\cdot g \\left(B \\left(I _ {g t}\\right)\\right) \\right], \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 550, + 468, + 566 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {C} _ {i m g} = E [ s \\cdot f (I _ {p r e d}) \\cdot f (I _ {g t}) ], \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 569, + 468, + 585 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "where $B(\\cdot)$ is pretrained BLIP [27], and $s$ is a logit scale.", + "bbox": [ + 76, + 597, + 455, + 612 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For the ablation models, we first remove the content adaptor to evaluate its effect. In order to keep the embedding dimension fed to the generator, we removed the attention blocks from the content adaptor. We also drop the bias and activation function of the last linear layer, such that in the adaptation process, the model will only learn a non-adaptive embedding dictionary. Without the attention layers, the model has difficulty bridging the domain gap between text and image, resulting in the degradation of generation fidelity as shown in the second row of Tab. 2.", + "bbox": [ + 75, + 613, + 468, + 763 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The third row in Tab. 2 ablates the second training stage of the content adaptor by simply skipping it. Without this process, the model fails to learn descriptive embedding, which encodes instance-level features of the guidance object, leading to a drop in appearance preservation.", + "bbox": [ + 75, + 763, + 468, + 839 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Data augmentation is another key element for better performance, i.e., cropping and shift augmentation increases the mask area, thus improving the generation quality in detail. The fourth row in Tab. 2 illustrates that with the ab", + "bbox": [ + 75, + 840, + 468, + 898 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "sence of data augmentation, there is a notable drop in realism, and the generation carries less accurate details.", + "bbox": [ + 498, + 262, + 890, + 292 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To further assess the importance of the content adaptor module, we replace it with a pretrained BLIP model, which transfers the reference image to a text embedding. As shown in the last row in Tab. 2, this results in the lowest CLIP image score, indicating that high-level semantics represented by text are not sufficient for maintaining object identity, highlighting the importance of our content adaptor.", + "bbox": [ + 496, + 294, + 892, + 401 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion and Future Work", + "text_level": 1, + "bbox": [ + 498, + 417, + 769, + 434 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose the first diffusion-based approach to tackle object compositing. Traditional methods require many steps including harmonization, geometry/view adjustment, and shadow generation. By contrast, our method directly achieves realistic compositing. We construct our model from a text-to-image generation network and introduce a novel content adaptor module. Furthermore, we present a fully self-supervised framework to train our model with synthetic data generation and augmentation. We show our superior performance over baseline methods through user studies on real-world examples.", + "bbox": [ + 496, + 445, + 892, + 611 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Since we give the first attempt to handle this challenging task, our method has several limitations. It lacks control over the appearance preservation of the synthesized object. Potential solutions are training the visual encoder on pairs of images with the same object, or training it jointly with the UNet. Another limitation is caused by masking the output image, which prohibits our model from generating global effects, i.e., the shadow can only be synthesized within the mask. To address this issue, we may need to synthesize complete background images and pair them with objects, and train with a conditional generation framework. We also observed artifacts around the object in the background. This can be improved by training the model to also predict an instance mask [49]. We will improve this in our future work.", + "bbox": [ + 496, + 613, + 893, + 825 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 500, + 844, + 668, + 859 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work is partially supported by NSF grants 1835739 and 2106717. We also thank Zhiquan Wang for discussion.", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "18317", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18208-18218, 2022. 2, 3", + "[2] Samaneh Azadi, Deepak Pathak, Sayna Ebrahimi, and Trevor Darrell. Compositional gan: Learning image-conditional binary composition. International Journal of Computer Vision, 128(10):2570-2585, 2020. 2", + "[3] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, et al. edifi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 2", + "[4] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. arXiv preprint arXiv:1809.11096, 2018. 2", + "[5] Bor-Chun Chen and Andrew Kae. Toward realistic image compositing with adversarial learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8415-8424, 2019. 2, 3", + "[6] Wenyan Cong, Li Niu, Jianfu Zhang, Jing Liang, and Liqing Zhang. Bargainnet: Background-guided domain translation for image harmonization. In 2021 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2021. 1, 2, 3", + "[7] Wenyan Cong, Jianfu Zhang, Li Niu, Liu Liu, Zhixin Ling, Weiyuan Li, and Liqing Zhang. Dovenet: Deep image harmonization via domain verification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8394-8403, 2020. 1, 2, 3", + "[8] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Deep image homography estimation. arXiv preprint arXiv:1606.03798, 2016. 5", + "[9] Zhida Feng, Zhenyu Zhang, Xintong Yu, Yewei Fang, Lanxin Li, Xuyi Chen, Yuxiang Lu, Jiaxiang Liu, Weichong Yin, Shikun Feng, et al. Ernie-vilg 2.0: Improving text-to-image diffusion model with knowledge-enhanced mixture-of-denoising-experts. arXiv preprint arXiv:2210.15257, 2022.2.3", + "[10] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020. 2", + "[11] Liu He, Yijuan Lu, John Corring, Dinei Florencio, and Cha Zhang. Diffusion-based document layout generation. arXiv preprint arXiv:2303.10787, 2023. 2", + "[12] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021. 8", + "[13] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 8" + ], + "bbox": [ + 76, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 2, 3", + "[15] Yannick Hold-Geoffroy, Kalyan Sunkavalli, Sunil Hadap, Emiliano Gambaretto, and Jean-François Lalonde. Deep outdoor illumination estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7312–7321, 2017. 2", + "[16] Yan Hong, Li Niu, and Jianfu Zhang. Shadow generation for composite image in real-world scenes. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 914-922, 2022. 1, 2, 3", + "[17] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 2", + "[18] Max Jaderberg, Karen Simonyan, Andrew Zisserman, et al. Spatial transformer networks. Advances in neural information processing systems, 28, 2015. 2", + "[19] Yifan Jiang, He Zhang, Jianming Zhang, Yilin Wang, Zhe Lin, Kalyan Sunkavalli, Simon Chen, Sohrab Amirghodsi, Sarah Kong, and Zhangyang Wang. Ssh: A self-supervised framework for image harmonization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4832-4841, 2021. 1, 2, 3", + "[20] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 2", + "[21] Kevin Karsch, Varsha Hedau, David Forsyth, and Derek Hoiem. Rendering synthetic objects into legacy photographs. ACM Transactions on Graphics (TOG), 30(6):1-12, 2011. 2", + "[22] Kevin Karsch, Kalyan Sunkavalli, Sunil Hadap, Nathan Carr, Hailin Jin, Rafael Fonte, Michael Sittig, and David Forsyth. Automatic scene inference for 3d object compositing. ACM Transactions on Graphics (TOG), 33(3):1-15, 2014. 2", + "[23] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagc: Text-based real image editing with diffusion models. arXiv preprint arXiv:2210.09276, 2022. 2, 3", + "[24] Jean-Francois Lalonde and Alexei A Efros. Using color compatibility for assessing image realism. In 2007 IEEE 11th International Conference on Computer Vision, pages 1-8. IEEE, 2007. 3", + "[25] Seung Hoon Lee, Seunghyun Lee, and Byung Cheol Song. Vision transformer for small-size datasets. arXiv preprint arXiv:2112.13492, 2021. 5", + "[26] Youngwan Lee and Jongyoul Park. Centermask: Realtime anchor-free instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13906-13915, 2020. 5", + "[27] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. arXiv preprint arXiv:2201.12086, 2022. 6, 8", + "[28] Chen-Hsuan Lin, Ersin Yumer, Oliver Wang, Eli Shechtman, and Simon Lucey. St-gan: Spatial transformer generative" + ], + "bbox": [ + 501, + 92, + 893, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "18318", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "adversarial networks for image compositing. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 9455-9464, 2018. 1, 2", + "[29] Daquan Liu, Chengjiang Long, Hongpan Zhang, Hanning Yu, Xinzhi Dong, and Chunxia Xiao. Arshadowgan: Shadow generative adversarial network for augmented reality in single light scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8139-8148, 2020. 1, 2", + "[30] Xihui Liu, Dong Huk Park, Samaneh Azadi, Gong Zhang, Arman Chopikyan, Yuxiao Hu, Humphrey Shi, Anna Rohrbach, and Trevor Darrell. More control for free! image synthesis with semantic diffusion guidance. arXiv preprint arXiv:2112.05744, 2021. 2, 3", + "[31] Shitong Luo and Wei Hu. Diffusion probabilistic models for 3d point cloud generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2837-2845, 2021. 2", + "[32] Chenlin Meng, Yang Song, Jiaming Song, Jiajun Wu, JunYan Zhu, and Stefano Ermon. Sdedit: Image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 2, 3, 6", + "[33] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 2, 3", + "[34] Patrick Pérez, Michel Gangnet, and Andrew Blake. Poisson image editing. In ACM SIGGRAPH 2003 Papers, pages 313-318. 2003. 1", + "[35] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 2", + "[36] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 4, 6, 8", + "[37] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 3", + "[38] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2, 3, 6", + "[39] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. arXiv preprint arXiv:2208.12242, 2022. 2, 3", + "[40] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image" + ], + "bbox": [ + 78, + 92, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022.3", + "[41] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 5", + "[42] Yichen Sheng, Yifan Liu, Jianming Zhang, Wei Yin, A Cengiz Oztireli, He Zhang, Zhe Lin, Eli Shechtman, and Bedrich Benes. Controllable shadow generation using pixel height maps. In European Conference on Computer Vision, pages 240-256. Springer, 2022. 1", + "[43] Yichen Sheng, Jianming Zhang, and Bedrich Benes. Ssn: Soft shadow network for image compositing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4380-4390, 2021. 1, 2, 3", + "[44] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256-2265. PMLR, 2015. 3", + "[45] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in Neural Information Processing Systems, 32, 2019. 3", + "[46] Kalyan Sunkavalli, Micah K Johnson, Wojciech Matusik, and Hanspeter Pfister. Multi-scale image harmonization. ACM Transactions on Graphics (TOG), 29(4):1-10, 2010. 3", + "[47] Yi-Hsuan Tsai, Xiaohui Shen, Zhe Lin, Kalyan Sunkavalli, Xin Lu, and Ming-Hsuan Yang. Deep image harmonization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3789-3797, 2017. 1", + "[48] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 5", + "[49] Shaoan Xie, Zhifei Zhang, Zhe Lin, Tobias Hinz, and Kun Zhang. Smartbrush: Text and shape guided object inpainting with diffusion model. arXiv preprint arXiv:2212.05034, 2022.8", + "[50] Ning Xu, Brian Price, Scott Cohen, and Thomas Huang. Deep image matting. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2970-2979, 2017. 2", + "[51] Ben Xue, Shenghui Ran, Quan Chen, Rongfei Jia, Binqiang Zhao, and Xing Tang. Dccf: Deep comprehensible color filter learning framework for high-resolution image harmonization. arXiv preprint arXiv:2207.04788, 2022. 1, 2", + "[52] Hao Zhou, Sunil Hadap, Kalyan Sunkavalli, and David W Jacobs. Deep single-image portrait relighting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7194-7202, 2019. 1" + ], + "bbox": [ + 503, + 92, + 890, + 825 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "18319", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/ObjectStitch_ Object Compositing With Diffusion Model/69193070-f3e5-4dd6-8743-b15c03b3e677_model.json b/2023/ObjectStitch_ Object Compositing With Diffusion Model/69193070-f3e5-4dd6-8743-b15c03b3e677_model.json new file mode 100644 index 0000000000000000000000000000000000000000..60290d48404e58e06ea7cae9baff32c75417cc9c --- /dev/null +++ b/2023/ObjectStitch_ Object Compositing With Diffusion Model/69193070-f3e5-4dd6-8743-b15c03b3e677_model.json @@ -0,0 +1,2244 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.202, + 0.131, + 0.77, + 0.154 + ], + "angle": 0, + "content": "ObjectStitch: Object Compositing with Diffusion Model" + }, + { + "type": "text", + "bbox": [ + 0.219, + 0.18, + 0.751, + 0.234 + ], + "angle": 0, + "content": "Yizhi Song\\(^{1*}\\), Zhifei Zhang\\(^{2}\\), Zhe Lin\\(^{2}\\), Scott Cohen\\(^{2}\\), Brian Price\\(^{2}\\), Jianming Zhang\\(^{2}\\), Soo Ye Kim\\(^{2}\\), Daniel Aliaga\\(^{1}\\) Purdue University\\(^{1}\\), Adobe Research\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.119, + 0.236, + 0.846, + 0.253 + ], + "angle": 0, + "content": "\\(^{1}\\{ \\text{song630,aliaga} \\} @\\)urdue.edu,\\(^{2}\\{ \\text{zhang,zlin,scohen,bprice, jianmzha,sooyek} \\} @\\)adobe.com" + }, + { + "type": "image_caption", + "bbox": [ + 0.105, + 0.293, + 0.197, + 0.308 + ], + "angle": 0, + "content": "(a) copy-paste" + }, + { + "type": "image_caption", + "bbox": [ + 0.219, + 0.287, + 0.348, + 0.314 + ], + "angle": 0, + "content": "(b) traditional compositing pipeline" + }, + { + "type": "image_caption", + "bbox": [ + 0.393, + 0.294, + 0.443, + 0.308 + ], + "angle": 0, + "content": "(c) ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.51, + 0.294, + 0.602, + 0.308 + ], + "angle": 0, + "content": "(a) copy-paste" + }, + { + "type": "image_caption", + "bbox": [ + 0.623, + 0.287, + 0.752, + 0.314 + ], + "angle": 0, + "content": "(b) traditional compositing pipeline" + }, + { + "type": "image_caption", + "bbox": [ + 0.797, + 0.294, + 0.846, + 0.308 + ], + "angle": 0, + "content": "(c) ours" + }, + { + "type": "image", + "bbox": [ + 0.084, + 0.317, + 0.217, + 0.526 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.218, + 0.317, + 0.351, + 0.42 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.351, + 0.317, + 0.484, + 0.42 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.488, + 0.317, + 0.622, + 0.42 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.622, + 0.318, + 0.755, + 0.42 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.755, + 0.318, + 0.888, + 0.42 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.532, + 0.894, + 0.588 + ], + "angle": 0, + "content": "Figure 1. Example results of object compositing with (a) copy-and-paste scheme, (b) traditional compositing pipeline and (c) ours (Object-Stitch). Traditional compositing pipeline is done with best possible off-the-shelf models including foreground/background color harmonization [19,47], poisson blending [34], and shadow synthesis [42]. ObjectStitch achieves more realistic results, and can address geometry correction, harmonization, shadow generation, and view synthesis all-in-one while preserving similar appearance to the reference object." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.616, + 0.314, + 0.631 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.648, + 0.474, + 0.875 + ], + "angle": 0, + "content": "Object compositing based on 2D images is a challenging problem since it typically involves multiple processing stages such as color harmonization, geometry correction and shadow generation to generate realistic results. Furthermore, annotating training data pairs for compositing requires substantial manual effort from professionals, and is hardly scalable. Thus, with the recent advances in generative models, in this work, we propose a self-supervised framework for object compositing by leveraging the power of conditional diffusion models. Our framework can holistically address the object compositing task in a unified model, transforming the viewpoint, geometry, color and shadow of the generated object while requiring no manual labeling. To preserve the input object's characteristics, we introduce a content adaptor that helps to maintain categori" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.603, + 0.895, + 0.679 + ], + "angle": 0, + "content": "cal semantics and object appearance. A data augmentation method is further adopted to improve the fidelity of the generator. Our method outperforms relevant baselines in both realism and faithfulness of the synthesized result images in a user study on various real-world images." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.709, + 0.633, + 0.725 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Image compositing is an essential task in image editing that aims to insert an object from a given image into another image in a realistic way. Conventionally, many sub-tasks are involved in composing an object to a new scene, including color harmonization [6, 7, 19, 51], relighting [52], and shadow generation [16, 29, 43] in order to naturally blend the object into the new image. As shown in Tab. 1, most previous methods [6, 7, 16, 19, 28, 43] focus on a single sub-task required for image compositing. Consequently, they must be appropriately combined to obtain a composite image where the input object is re-synthesized to have the" + }, + { + "type": "page_footnote", + "bbox": [ + 0.101, + 0.887, + 0.321, + 0.901 + ], + "angle": 0, + "content": "* Work done during internship at Adobe." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18310" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.089, + 0.451, + 0.222 + ], + "angle": 0, + "content": "
MethodGeometryLightShadowView
ST-GAN [28]✓XXX
SSH [19]X✓XX
DCCF [51]X✓XX
SSN [43]XX✓X
SGRNet [16]XX✓X
GCC-GAN [5]✓✓XX
Ours✓✓✓✓
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.226, + 0.47, + 0.268 + ], + "angle": 0, + "content": "Table 1. Prior works only focus on one or two aspects of object compositing, and they cannot synthesize novel views. In contrast, our model can address all perspectives as listed." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.288, + 0.468, + 0.35 + ], + "angle": 0, + "content": "color, lighting and shadow that is consistent with the background scene. As shown in Fig. 1, results produced in this way still look unnatural, partly due to the viewpoint of the inserted object being different from the overall background." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.352, + 0.468, + 0.548 + ], + "angle": 0, + "content": "Harmonizing the geometry and synthesizing novel views have often been overlooked in 2D image compositing, which require an accurate understanding of both the geometry of the object and the background scene from 2D images. Previous works [15, 21, 22] handle 3D object compositing with explicit background information such as lighting positions and depth. More recently, [28] utilize GANs to estimate the homography of the object. However, this method is limited to placing furniture in indoor scenes. In this paper, we propose a generic image compositing method that is able to harmonize the geometry of the input object along with color, lighting and shadow with the background image using a diffusion-based generative model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.55, + 0.468, + 0.824 + ], + "angle": 0, + "content": "In recent years, generative models such as GANs [4, 10, 17, 20] and diffusion models [1, 14, 30, 32, 33, 38, 39] have shown great potential in synthesizing realistic images. In particular, diffusion model-based frameworks are versatile and outperform various prior methods in image editing [1, 23, 32] and other applications [11, 31, 35]. However, most image editing diffusion models focus on using text inputs to manipulate images [1, 3, 9, 23, 33], which is insufficient for image compositing as verbal representations cannot fully capture the details or preserve the identity and appearance of a given object image. There have been recent works [23, 39] focusing on generating diverse contexts while preserving the key features of the object; however, these models are designed for a different task than object compositing. Furthermore, [23] requires fine-tuning the model for each input object and [39] also needs to be finetuned on multiple images of the same object. Therefore they are limited for general object compositing." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.468, + 0.903 + ], + "angle": 0, + "content": "In this work, we leverage diffusion models to simultaneously handle multiple aspects of image compositing such as color harmonization, relighting, geometry correction and shadow generation. With image guidance rather than text guidance, we aim to preserve the identity and appearance of" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.85, + 0.107 + ], + "angle": 0, + "content": "the original object in the generated composite image." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.108, + 0.892, + 0.349 + ], + "angle": 0, + "content": "Specifically, our model synthesizes a composite image given (i) a source object, (ii) a target background image, and (iii) a bounding box specifying the location to insert the object. The proposed framework consists of a content adaptor and a generator module: the content adaptor is designed to extract a representation from the input object containing both high-level semantics and low-level details such as color and shape; the generator module preserves the background scene while improving the generation quality and versatility. Our framework is trained in a fully self-supervised manner and no task-specific labeling is required at any point during training. Moreover, various data augmentation techniques are applied to further improve the fidelity and realism of the output. We evaluate our proposed method on a real-world dataset closely simulating real use cases for image compositing." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.349, + 0.822, + 0.363 + ], + "angle": 0, + "content": "Our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.368, + 0.892, + 0.429 + ], + "angle": 0, + "content": "- We present the first diffusion model-based framework for generative object compositing that can handle multiple aspects of compositing such as viewpoint, geometry, lighting and shadow." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.429, + 0.892, + 0.473 + ], + "angle": 0, + "content": "- We propose a content adaptor module which learns a descriptive multi-modal embedding from images, enabling image guidance for diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.474, + 0.892, + 0.534 + ], + "angle": 0, + "content": "- Our framework is trained in a self-supervised manner without any task-specific annotations, employing data augmentation techniques to improve the fidelity of generation." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.535, + 0.892, + 0.58 + ], + "angle": 0, + "content": "- We collect a high-resolution real-world dataset for object compositing with diverse images, containing manually annotated object scales and locations." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.368, + 0.892, + 0.58 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.595, + 0.642, + 0.611 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.62, + 0.687, + 0.637 + ], + "angle": 0, + "content": "2.1. Image Compositing" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.892, + 0.78 + ], + "angle": 0, + "content": "Image compositing is a challenging task in reference-guided image editing, where the object in a given foreground image is to be inserted into a background image. The generated composite image is expected to look realistic, with the appearance of the original object being preserved. Prior works often focus on an individual aspect of this problem, such as geometric correction [2, 28], image harmonization [6, 7, 51], image matting [50] and shadow generation [29]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Lin et al. [28] propose an iterative GAN-based framework to correct geometric inconsistencies between object and background images. In their model, STNs [18] are integrated into the generator network to predict a series of warp updates. Although their work can improve the realism of the scene geometry in the composite image, it is limited to inserting furniture into indoor scenes. Azadi et al. [2] focus on binary composition, using a composition-decomposition" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "18311" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.455, + 0.106 + ], + "angle": 0, + "content": "network to capture interactions between a pair of objects." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.108, + 0.468, + 0.242 + ], + "angle": 0, + "content": "Image harmonization aims at minimizing the inconsistency between the input object and the background image. Traditional methods [24, 46] usually concentrate on obtaining color statistics and then transfer this information between foreground and background. Recent works seek to solve this problem via deep neural networks. [6, 7] reformulate image harmonization as a domain adaptation problem, while [19] converts it to a representation fusing problem and utilizes self-supervised training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.244, + 0.468, + 0.349 + ], + "angle": 0, + "content": "Shadow synthesis is an effect that is often overlooked in previous image compositing methods although it is essential for generating realistic composite images. SGRNet [16] divides this task into a shadow mask generation stage and a shadow filling stage. SSN [43], focusing on soft shadows, predicts an ambient occlusion map as cue for shadow generation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.351, + 0.468, + 0.516 + ], + "angle": 0, + "content": "There are also works simultaneously addressing multiple sub-problems of image compositing. Chen et al. [5] develop a system including multiple network modules to produce plausible images with geometric, color and boundary consistency. However, the pose of the generated object is constrained by the mask input and the model cannot generalize to non-rigid objects such as animals. Our proposed model also handles the image composition problem in a unified manner, generating foreground objects which are harmonious and geometrically consistent with the background while synthesizing novel views and shadows." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.529, + 0.3, + 0.545 + ], + "angle": 0, + "content": "2.2. Guided Image Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.553, + 0.468, + 0.704 + ], + "angle": 0, + "content": "Recent years have witnessed great advances in diffusion models. Diffusion models are a family of deep generative models based on several predominant works [14, 44, 45], defined with two Markov chains. In the forward process, they gradually add noise to the data, and in the reverse process, they learn to recover the data from noise. Following the growth of research in this area, diffusion models have shown their potential in a great number of applications, such as text-to-image generation [9, 33, 38], image editing [1, 23, 32] and guided image synthesis [30]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Text-to-image synthesis has been a popular research topic over the past few years. Stable Diffusion [38] made a significant contribution to this task and it reduces the computational cost by applying the diffusion model in the latent space. Other works explore ways of more flexible and controllable text-driven image editing. Avrahami et al. [1] design an algorithm in their Blended Diffusion framework to fuse noised versions of the input with the local text-guided diffusion latent, generating smooth transition. Similarly, GLIDE [33] is also capable of handling text-guided inpainting after fine-tuning on this specific task. Some methods give users more straightforward control on images. SDEdit [32] allows stroke painting in images, blending user input" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.151 + ], + "angle": 0, + "content": "into the image by adding noise to the input and denoising through a stochastic differential equation. SDG [30] injects Semantic Diffusion Guidance at each iteration of the generation and can provide multi-modal guidance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.152, + 0.892, + 0.424 + ], + "angle": 0, + "content": "While the versatility and generation quality of diffusion models have been repeatedly demonstrated, preserving object appearance remains a very challenging problem in image editing. Ruiz et al. [39] address this problem by optimizing their model using a reconstruction loss and a class-specific prior preservation loss. Though their approach can preserve the details of the object, multiple images of the same object are required to fine-tune their model on this target object. Kawar et al. [23] handle this issue by interpolating the initial text embedding and the optimized embedding for reconstruction, but their method is mainly designed for applying edits on objects, not for generating the same object in different contexts. In this paper, we propose the first generative object compositing framework based on diffusion models, which generates the harmonized and geometrically-corrected subject with a novel view and shadow. Furthermore, the characteristics of the original object is preserved in the synthesized composite image." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.438, + 0.673, + 0.455 + ], + "angle": 0, + "content": "3. Proposed Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.463, + 0.892, + 0.704 + ], + "angle": 0, + "content": "We define the generative object compositing problem as: Given an input triplet \\((I_o, I_{bg}, M)\\) that consists of an object image \\(I_o \\in \\mathbb{R}^{H_s \\times W_s \\times 3}\\), a background image \\(I_{bg} \\in \\mathbb{R}^{H_t \\times W_t \\times 3}\\), and its associated binary mask \\(M \\in \\mathbb{R}^{H_t \\times W_t \\times 1}\\) with the desired object location set to 0 and the rest to 1, the goal is to compose the input object into the masked area \\(I_{bg} \\otimes M\\). \\(M\\) is considered as a soft constraint of the location and scale of the composited object. The output generated image should look realistic, while the appearance of the object is preserved. Our problem setting is different from text-guided image generation and inpainting in that the condition input is a reference object image rather than a text prompt. Inspired by the success of text-guided diffusion models [37, 38, 40], which inject text conditioning into the diffusion architecture, we design our method for generative compositing to leverage such pretrained models." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.901 + ], + "angle": 0, + "content": "An overview of our framework is shown in Fig. 2. It consists of an object image encoder extracting semantic features from the object and a conditional diffusion generator. To leverage the power of pretrained text-to-image diffusion models, we introduce a content adaptor that can bridge the gap between the object encoder and conditional generator by transforming a sequence of visual tokens to a sequence of text tokens to overcome the domain gap between image and text. This design further helps to preserve the object appearance. We propose a two-stage training process: in the first stage, the content adaptor is trained on large image/text pairs to maintain high-level semantics of the object; in the second stage, it is trained in the context of diffusion" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "18312" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.089, + 0.868, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.244, + 0.895, + 0.3 + ], + "angle": 0, + "content": "Figure 2. System pipeline. Our framework consists of a content adaptor and a generator (a pretrained text-to-image diffusion model). The input image \\( I_{o} \\) is fed into a ViT and the adaptor which produces a descriptive embedding. At the same time the background image \\( I_{bg} \\) is taken as input by the diffusion model. At each iteration during the denoising stage, we apply the mask \\( M \\) on the generated image \\( I_{out} \\), so that the generator only denoises the masked area \\( I_{out} \\otimes M \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.314, + 0.475, + 0.421 + ], + "angle": 0, + "content": "generator to encode key identity features of the object by encouraging the visual reconstruction of the object in the original image. Lastly, the generator module is fine-tuned on the embedding produced by the adaptor through cross attention blocks. All stages our trained in a self-supervised manner to avoid expensive annotation for obtaining object compositing training data." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.442, + 0.196, + 0.456 + ], + "angle": 0, + "content": "3.1. Generator" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.469, + 0.471, + 0.577 + ], + "angle": 0, + "content": "As depicted in Fig. 2, we leverage a pretrained text-to-image diffusion model architecture and modify it for the compositing task by (i) introducing a mask in the input, and (ii) adjusting the U-Net input to contain the original background image outside the hole and noise inside the hole with mask blending. In order to condition the model on the guidance embedding \\( E \\), an attention mechanism is applied as:" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.606, + 0.47, + 0.642 + ], + "angle": 0, + "content": "\\[\n\\operatorname {S o f t m a x} \\left(\\frac {\\left(\\boldsymbol {W} _ {Q} E _ {x}\\right) \\left(\\boldsymbol {W} _ {K} E\\right) ^ {T}}{\\sqrt {d}}\\right) \\boldsymbol {W} _ {V} E = \\boldsymbol {A} \\boldsymbol {V} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.473, + 0.902 + ], + "angle": 0, + "content": "where \\( E_{x} \\) is an intermediate representation of the denoising autoencoder and \\( W_{Q} \\in \\mathbb{R}^{d \\times d_{x}} \\), \\( W_{K} \\in \\mathbb{R}^{d \\times d_{e}} \\) and \\( W_{V} \\in \\mathbb{R}^{d \\times d_{e}} \\) are embedding matrices. The background outside the mask area should be perfectly preserved for generative object compositing. Thus, we use the input mask \\( M \\) for blending the background image \\( I_{bg} \\) with the generated image \\( I_{out} \\). As a result, the generator only denoises the masked area \\( I_{out} \\otimes M \\). To use this model on our task, a straightforward way (a baseline) would be to apply image captioning on the object image and feeding the resulting caption directly to the diffusion model as the condition. However, text embedding cannot capture fine grained visual details. Therefore, in order to directly leverage the given object image, we introduce the content adaptor to transform the visual features from a pretrained visual encoder to text features (tokens) to use as conditioning for the generator." + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.311, + 0.878, + 0.52 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.521, + 0.895, + 0.592 + ], + "angle": 0, + "content": "Figure 3. Structure of the Content Adaptor. In the first stage, it is trained on a large dataset of image-caption pairs to learn multimodal sequential embeddings containing high-level semantics. In the second stage, it is fine-tuned under the diffusion framework to learn to encode identity features in adaptive embedding." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.607, + 0.669, + 0.624 + ], + "angle": 0, + "content": "3.2. Content Adaptor" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.632, + 0.894, + 0.707 + ], + "angle": 0, + "content": "To prevent the loss of key identity information, we use an image encoder instead of a text encoder to produce the embedding from the input object image. However, the image embedding cannot be effectively utilized by the diffusion model for two reasons:" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.712, + 0.892, + 0.773 + ], + "angle": 0, + "content": "- The image embedding \\(\\widetilde{E}\\) and the text embedding \\(E\\) are from different domains. Since the diffusion model was trained on \\(E\\), it cannot generate meaningful contents from image embedding sequence;" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.772, + 0.891, + 0.802 + ], + "angle": 0, + "content": "- A mismatch in the dimensions of \\(\\widetilde{E} \\in \\mathbb{R}^{k \\times 257 \\times 1024}\\) and \\(E \\in \\mathbb{R}^{k \\times 77 \\times 768}\\), where \\(k\\) is batch-size." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.712, + 0.892, + 0.802 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.809, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Therefore, based on the above observations, we develop a sequence-to-sequence translator architecture as shown in Fig. 3. Given an image-caption pair as an input tuple \\((I,t)\\), we employ two pretrained ViT-L/14 encoders \\(C_t\\), \\(C_i\\) from CLIP [36] to produce text embedding \\(E = C_t(t)\\) and image embedding \\(\\widetilde{E} = C_i(I)\\), respectively. The adaptor \\(T\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18313" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.089, + 0.44, + 0.265 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.266, + 0.473, + 0.364 + ], + "angle": 0, + "content": "Figure 4. Illustration of our synthetic data generation and data augmentation scheme. The top row shows the data generation process including perspective warping, random rotation, and random color shifting. The original image is used as both the input background and ground truth, while the perturbed object is fed into the adaptor. The bottom row shows crop and shift augmentations, which help to improve the generation quality and preserve object details." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.377, + 0.471, + 0.483 + ], + "angle": 0, + "content": "consists of three components: a 1D convolutional layer, attention blocks [48] and an MLP, where the 1D convolution modifies the length of the embedding from 257 to 77, the MLP maps the embedding dimension from 1024 to 768, and the attention blocks bridge the gap between text domain and image domain. We design a two-stage optimization method to train this module, which is explained in Sec. 3.3.2." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.493, + 0.327, + 0.509 + ], + "angle": 0, + "content": "3.3. Self-supervised Framework" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.516, + 0.47, + 0.637 + ], + "angle": 0, + "content": "There is no publicly available image compositing training dataset with annotations which is sufficient for training a diffusion model, and it is extremely challenging to manually annotate such data. Therefore, we propose a self-supervised training scheme and a synthetic data generation approach that simulate real-world scenarios. We also introduce a data augmentation method to enrich the training data as well as improve the robustness of our model." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.651, + 0.383, + 0.665 + ], + "angle": 0, + "content": "3.3.1 Data Generation and Augmentation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.674, + 0.47, + 0.795 + ], + "angle": 0, + "content": "Training data generation. We collect our synthetic training data from Pixabay and use an object instance segmentation model [26] to predict panoptic segmentation and classification labels. We first filter the dataset by removing objects with very small or large sizes. Then, we apply spatial and color perturbations to simulate many real use-case scenarios where the input image and background image have different scene geometry and lighting conditions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.47, + 0.885 + ], + "angle": 0, + "content": "The top row of Fig. 4 illustrates this process. Inspired by [8], we randomly perturb the four points of the object bounding box to apply projective transformation, followed by a random rotation within the range \\([- \\theta, \\theta]\\) (\\(\\theta = 20^{\\circ}\\)) and color perturbation. The segmentation mask (perturbed in the same way as the image) is used to extract the object." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.47, + 0.901 + ], + "angle": 0, + "content": "This data synthesis pipeline is fully controllable and can" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.213 + ], + "angle": 0, + "content": "also be employed as data augmentation during training. Another key advantage is that it is free of manual labeling, since the original image is used as the ground truth. We use the bounding box as the mask as it not only fully covers the object, but also extends to its neighboring area (providing room for shadow generation). We find that it is flexible enough for the model to apply spatial transformations, synthesize novel views and generate shadows and reflection." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.216, + 0.893, + 0.444 + ], + "angle": 0, + "content": "Real-world evaluation data. Given that there is no existing dataset specifically for our task, we manually collect a new dataset, closely simulating real-world use cases, as an evaluation benchmark for object compositing. The dataset consists of 503 pairs of common objects (including both rigid and non-rigid objects such as vehicles and wildlife) and diverse background images (covering both indoor and outdoor scenes). It also contains challenging cases where there is a large mismatch of lighting conditions or viewpoints between foreground and background. The dataset images are collected from Pixabay. The labeling procedure closely simulate the real scenarios where the input object is placed at a target location in the background image and then scaled at the user's will. The compositing region is determined as a loose bounding box around the object." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.447, + 0.894, + 0.568 + ], + "angle": 0, + "content": "Data augmentation. Inspired by [25], we introduce random shift and crop augmentations during training, while ensuring that the foreground object is always contained in the crop window. This process is illustrated in the bottom row of Fig. 4. Applying this augmentation method for both training and inference results in a notable improvement in the realism of the generated results. Quantitative results are provided in Sec. 4.4." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.58, + 0.616, + 0.595 + ], + "angle": 0, + "content": "3.3.2 Training" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.604, + 0.892, + 0.724 + ], + "angle": 0, + "content": "Content adaptor pretraining. We first pretrain the content adaptor to keep the semantics of the object by mapping the image embedding to text embedding. At this first stage, we optimize the content adaptor on a sequence-to-sequence translation task, which learns to project the image embedding into a multi-modal space. During training, \\( C_t \\), \\( C_i \\) are frozen, and the translator is trained on 3,203,338 image-caption pairs from a filtered LAION dataset [41]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.725, + 0.892, + 0.769 + ], + "angle": 0, + "content": "Given the input image embedding \\(\\vec{E}\\), we use the text embedding \\(E\\) as target, the objective function of this translation task is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.616, + 0.781, + 0.892, + 0.799 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {d i s t}} = \\| T (\\tilde {E}) - E \\| _ {1}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.811, + 0.725, + 0.825 + ], + "angle": 0, + "content": "where \\( T(\\cdot) \\) is the content adaptor." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.901 + ], + "angle": 0, + "content": "However, the multi-modal embedding obtained solely from this first stage optimization mostly carry high-level semantics of the input image without much of texture details. Hence, we further refine \\(\\widetilde{E}\\) to obtain \\(\\hat{E}\\), the adaptive embedding, for better appearance preservation." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18314" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.198 + ], + "angle": 0, + "content": "Content adaptor fine-tuning. We further optimize the content adaptor to produce an adaptive embedding which maintains instance-level properties of the object. After pretraining, we insert the content adaptor to the pretrained diffusion framework by feeding the adaptive embedding as context to the attention blocks. Then, the diffusion model is frozen and the adaptor is trained using:" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.207, + 0.47, + 0.227 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {a d a p t}} = \\mathbb {E} _ {T, \\epsilon \\sim \\mathcal {N} (0, 1)} [ \\| \\epsilon - \\epsilon_ {\\theta} (I _ {t} \\circ M, t, T (\\widetilde {E})) \\| _ {2} ^ {2} ], \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.236, + 0.47, + 0.297 + ], + "angle": 0, + "content": "where the content adaptor \\( T(\\cdot) \\) is optimized. It is trained on our synthetic dataset filtered from Pixabay, containing 467,280 foreground and background image pairs for training and 25,960 pairs for validation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.301, + 0.469, + 0.406 + ], + "angle": 0, + "content": "Generator fine-tuning. After the aforementioned two-stage training of the content adaptor is completed, we freeze the content adaptor and train the generator module after initializing the text-to-image diffusion model with pretrained weights. Crop and shift augmentations are applied in this process. Based on the Latent Diffusion model [38], the objective loss in our generator module is defined by:" + }, + { + "type": "equation", + "bbox": [ + 0.117, + 0.416, + 0.47, + 0.436 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {g e n} = \\mathbb {E} _ {\\hat {E}, \\epsilon \\sim \\mathcal {N} (0, 1)} [ \\| \\epsilon - \\epsilon_ {\\theta} (I _ {t} \\circ M, t, \\hat {E}) \\| _ {2} ^ {2} ], \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.445, + 0.47, + 0.522 + ], + "angle": 0, + "content": "where \\(I\\) is the input image, \\(I_{t}\\) is a noisy version of \\(I\\) at timestep \\(t\\), and \\(\\epsilon_{\\theta}\\) denotes the denoising model which is optimized. In order to adapt the text-guided generation task to the setting of image compositing, we apply the input mask on the image at every time-step." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.534, + 0.21, + 0.55 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.558, + 0.239, + 0.574 + ], + "angle": 0, + "content": "4.1. Training Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.582, + 0.47, + 0.779 + ], + "angle": 0, + "content": "The first training stage of the content adaptor takes 15 epochs with a learning rate of \\(10^{-4}\\) and batch size of 2048. The image size processed by ViT is \\(224 \\times 224\\). To set up the pipeline, we employed the pretrained image encoder and text encoder from CLIP [36]. The second training stage of the content adaptor takes 13 epochs with the learning rate of \\(2 \\times 10^{-5}\\) and batch size of 512. We resize the images to 512 \\(\\times\\) 512 and adapt the pretrained Stable Diffusion model [38] to our task. After the above process, the generator is trained for 20 epochs with the learning rate of \\(4 \\times 10^{-5}\\) and batch size of 576. The input image size is the same as that in the second training stage. All training stages are conducted on 8 A100 GPUs with Adam optimizer." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.787, + 0.3, + 0.803 + ], + "angle": 0, + "content": "4.2. Quantitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.47, + 0.902 + ], + "angle": 0, + "content": "To obtain quantitative results on a real dataset that simulates real-world use cases, where users drag and drop an object to a background image, we conduct a user study to compare to three baselines, i.e., copy-and-paste, pretrained stable diffusion model fine-tuned on BLIP [27], and a joint stable diffusion model integrating both BLIP" + }, + { + "type": "image_caption", + "bbox": [ + 0.541, + 0.09, + 0.822, + 0.111 + ], + "angle": 0, + "content": "Pick from image A and image B the one that you think looks more realistic." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.113, + 0.88, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.545, + 0.197, + 0.82, + 0.219 + ], + "angle": 0, + "content": "Which generated object do you think is most likely to be the same one from the guidance?" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.221, + 0.86, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.268, + 0.892, + 0.323 + ], + "angle": 0, + "content": "Figure 5. User study results. We conduct side-by-side comparisons between our method and one of baseline methods to quantify the generation quality in terms of realism and appearance preservation. The results show that our method outperforms baselines." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.333, + 0.892, + 0.545 + ], + "angle": 0, + "content": "and SDEdit [32]. Given the fact that there is no existing diffusion-based method addressing the same problem, we train the baselines on the same pretrained diffusion model using the same self-supervised training scheme and dataset as ours. To the best of our knowledge, after adapting to our setting, they are the closest baselines to our task. In the BLIP baseline, the content adaptor is replaced by BLIP, a state-of-the-art captioning model. The text embedding obtained from the BLIP caption is fed into the generator module. For the SDEdit baseline, we keep the BLIP model which will significantly improve the generation fidelity of a pretrained diffusion model. We set the noise strength to 0.65 in SDEdit, striking a balance between realism and faithfulness. Details are provided in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.546, + 0.893, + 0.833 + ], + "angle": 0, + "content": "We perform a user study using our real testing dataset, which consists of 503 object-background pairs. For each example, we set up a side-by-side comparison between the results from one of the baselines and our method, respectively. The subject will be asked two questions: 1) which is more realistic and 2) which generated object is more close to the given object? We collected 1,494 votes from more than 170 users. As shown in Fig. 5, our method outperforms the other baselines for both questions. The higher preference rates demonstrate the effectiveness of the content adaptor and data augmentation method in improving generation fidelity. It also shows that our content adaptor is capable of maintaining details and attributes of the input objects, while not being constrained by the input when harmonizing objects to the background. In contrast, although SDEdit is better at preserving appearance than BLIP, it is constrained by the original pose, shape, and other attributes of the object, making the generated object less coherent with the background." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.846, + 0.71, + 0.862 + ], + "angle": 0, + "content": "4.3. Qualitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In Fig. 6, we compare our method to various baselines. In the first two rows of Fig. 6, we place the same object" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18315" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.125, + 0.095, + 0.162, + 0.106 + ], + "angle": 0, + "content": "Object" + }, + { + "type": "image_caption", + "bbox": [ + 0.244, + 0.095, + 0.317, + 0.106 + ], + "angle": 0, + "content": "Target Image" + }, + { + "type": "image_caption", + "bbox": [ + 0.375, + 0.095, + 0.463, + 0.106 + ], + "angle": 0, + "content": "Copy-and-paste" + }, + { + "type": "image_caption", + "bbox": [ + 0.539, + 0.095, + 0.566, + 0.105 + ], + "angle": 0, + "content": "BLIP" + }, + { + "type": "image_caption", + "bbox": [ + 0.668, + 0.095, + 0.708, + 0.105 + ], + "angle": 0, + "content": "SDEdit" + }, + { + "type": "image_caption", + "bbox": [ + 0.807, + 0.095, + 0.835, + 0.105 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.111, + 0.891, + 0.844 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.079, + 0.853, + 0.893, + 0.896 + ], + "angle": 0, + "content": "Figure 6. Qualitative comparison on the real-world evaluation dataset. Our method can appropriately correct the viewpoint and geometry while harmonizing the object with the new background, generating realistic composite results. Moreover, lighting and shadows are naturally adjusted in the composite image. Compared to baselines using text guidance, our method can better preserve the object appearance." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18316" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.146, + 0.089, + 0.823, + 0.188 + ], + "angle": 0, + "content": "
AdaptorOptimizationAugmentationBLIP [27]FID ↓CLIP text score ↑CLIP image score ↑
✓✓✓✗15.433129.859497.0625
✗✓✓✗18.825429.828196.0000
✓✗✓✗16.138129.796996.8125
✓✓✗✗18.369829.781396.1875
✗✗✓✓17.900029.718895.8125
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.193, + 0.895, + 0.251 + ], + "angle": 0, + "content": "Table 2. Ablation study. We evaluate the effectiveness of the components listed in the first four columns: 1) whether the content adaptor is used; 2) whether the adaptor is optimized for appearance preservation; 3) whether crop augmentation is utilized in training; 4) if the adaptor module is replaced by BLIP to predict text embedding from an image. We use FID to assess the generation fidelity and modified CLIP score (explained in Sec. 4.4) to measure similarity of the guidance and the predicted image." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.263, + 0.473, + 0.37 + ], + "angle": 0, + "content": "in two scenes with notable differences in the geometry. Results show that our model is capable of correcting geometric inconsistencies while preserving the characteristics of the object. The limitations of the baselines are also shown in Fig. 6: BLIP often fails to preserve the appearance; SDEdit preserves texture and pose better than the first baseline, but often cannot make appropriate geometry transformations." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.379, + 0.231, + 0.395 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.402, + 0.471, + 0.525 + ], + "angle": 0, + "content": "In Tab. 2, we demonstrate the effectiveness of each key component in our ObjectStitch framework by ablating each of them. We use FID [13] and a modified CLIP [12, 36] score, which measure the fidelity and semantic matching performance between given and generated objects. Instead of directly adopting the CLIP score, which semantically matches the prompt and the image, we modify it to CLIP text score and CLIP image score that are defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.152, + 0.551, + 0.469, + 0.568 + ], + "angle": 0, + "content": "\\[\n\\mathcal {C} _ {t x t} = E \\left[ s \\cdot f \\left(I _ {p r e d}\\right) \\cdot g \\left(B \\left(I _ {g t}\\right)\\right) \\right], \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.57, + 0.469, + 0.587 + ], + "angle": 0, + "content": "\\[\n\\mathcal {C} _ {i m g} = E [ s \\cdot f (I _ {p r e d}) \\cdot f (I _ {g t}) ], \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.598, + 0.456, + 0.613 + ], + "angle": 0, + "content": "where \\(B(\\cdot)\\) is pretrained BLIP [27], and \\(s\\) is a logit scale." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.469, + 0.764 + ], + "angle": 0, + "content": "For the ablation models, we first remove the content adaptor to evaluate its effect. In order to keep the embedding dimension fed to the generator, we removed the attention blocks from the content adaptor. We also drop the bias and activation function of the last linear layer, such that in the adaptation process, the model will only learn a non-adaptive embedding dictionary. Without the attention layers, the model has difficulty bridging the domain gap between text and image, resulting in the degradation of generation fidelity as shown in the second row of Tab. 2." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.469, + 0.84 + ], + "angle": 0, + "content": "The third row in Tab. 2 ablates the second training stage of the content adaptor by simply skipping it. Without this process, the model fails to learn descriptive embedding, which encodes instance-level features of the guidance object, leading to a drop in appearance preservation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.841, + 0.469, + 0.9 + ], + "angle": 0, + "content": "Data augmentation is another key element for better performance, i.e., cropping and shift augmentation increases the mask area, thus improving the generation quality in detail. The fourth row in Tab. 2 illustrates that with the ab" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.263, + 0.892, + 0.293 + ], + "angle": 0, + "content": "sence of data augmentation, there is a notable drop in realism, and the generation carries less accurate details." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.295, + 0.893, + 0.402 + ], + "angle": 0, + "content": "To further assess the importance of the content adaptor module, we replace it with a pretrained BLIP model, which transfers the reference image to a text embedding. As shown in the last row in Tab. 2, this results in the lowest CLIP image score, indicating that high-level semantics represented by text are not sufficient for maintaining object identity, highlighting the importance of our content adaptor." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.419, + 0.771, + 0.435 + ], + "angle": 0, + "content": "5. Conclusion and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.446, + 0.893, + 0.612 + ], + "angle": 0, + "content": "We propose the first diffusion-based approach to tackle object compositing. Traditional methods require many steps including harmonization, geometry/view adjustment, and shadow generation. By contrast, our method directly achieves realistic compositing. We construct our model from a text-to-image generation network and introduce a novel content adaptor module. Furthermore, we present a fully self-supervised framework to train our model with synthetic data generation and augmentation. We show our superior performance over baseline methods through user studies on real-world examples." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.894, + 0.826 + ], + "angle": 0, + "content": "Since we give the first attempt to handle this challenging task, our method has several limitations. It lacks control over the appearance preservation of the synthesized object. Potential solutions are training the visual encoder on pairs of images with the same object, or training it jointly with the UNet. Another limitation is caused by masking the output image, which prohibits our model from generating global effects, i.e., the shadow can only be synthesized within the mask. To address this issue, we may need to synthesize complete background images and pair them with objects, and train with a conditional generation framework. We also observed artifacts around the object in the background. This can be improved by training the model to also predict an instance mask [49]. We will improve this in our future work." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.845, + 0.669, + 0.861 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "This work is partially supported by NSF grants 1835739 and 2106717. We also thank Zhiquan Wang for discussion." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18317" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18208-18218, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.472, + 0.228 + ], + "angle": 0, + "content": "[2] Samaneh Azadi, Deepak Pathak, Sayna Ebrahimi, and Trevor Darrell. Compositional gan: Learning image-conditional binary composition. International Journal of Computer Vision, 128(10):2570-2585, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.231, + 0.471, + 0.3 + ], + "angle": 0, + "content": "[3] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, et al. edifi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.302, + 0.47, + 0.344 + ], + "angle": 0, + "content": "[4] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. arXiv preprint arXiv:1809.11096, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.346, + 0.472, + 0.401 + ], + "angle": 0, + "content": "[5] Bor-Chun Chen and Andrew Kae. Toward realistic image compositing with adversarial learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8415-8424, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.403, + 0.47, + 0.471 + ], + "angle": 0, + "content": "[6] Wenyan Cong, Li Niu, Jianfu Zhang, Jing Liang, and Liqing Zhang. Bargainnet: Background-guided domain translation for image harmonization. In 2021 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2021. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.474, + 0.472, + 0.544 + ], + "angle": 0, + "content": "[7] Wenyan Cong, Jianfu Zhang, Li Niu, Liu Liu, Zhixin Ling, Weiyuan Li, and Liqing Zhang. Dovenet: Deep image harmonization via domain verification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8394-8403, 2020. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.546, + 0.47, + 0.586 + ], + "angle": 0, + "content": "[8] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Deep image homography estimation. arXiv preprint arXiv:1606.03798, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.589, + 0.47, + 0.67 + ], + "angle": 0, + "content": "[9] Zhida Feng, Zhenyu Zhang, Xintong Yu, Yewei Fang, Lanxin Li, Xuyi Chen, Yuxiang Lu, Jiaxiang Liu, Weichong Yin, Shikun Feng, et al. Ernie-vilg 2.0: Improving text-to-image diffusion model with knowledge-enhanced mixture-of-denoising-experts. arXiv preprint arXiv:2210.15257, 2022.2.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.673, + 0.47, + 0.729 + ], + "angle": 0, + "content": "[10] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.731, + 0.47, + 0.772 + ], + "angle": 0, + "content": "[11] Liu He, Yijuan Lu, John Corring, Dinei Florencio, and Cha Zhang. Diffusion-based document layout generation. arXiv preprint arXiv:2303.10787, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.775, + 0.47, + 0.829 + ], + "angle": 0, + "content": "[12] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.832, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[13] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 8" + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.116, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[14] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.894, + 0.204 + ], + "angle": 0, + "content": "[15] Yannick Hold-Geoffroy, Kalyan Sunkavalli, Sunil Hadap, Emiliano Gambaretto, and Jean-François Lalonde. Deep outdoor illumination estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7312–7321, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.205, + 0.895, + 0.259 + ], + "angle": 0, + "content": "[16] Yan Hong, Li Niu, and Jianfu Zhang. Shadow generation for composite image in real-world scenes. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 914-922, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.261, + 0.894, + 0.327 + ], + "angle": 0, + "content": "[17] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.33, + 0.892, + 0.371 + ], + "angle": 0, + "content": "[18] Max Jaderberg, Karen Simonyan, Andrew Zisserman, et al. Spatial transformer networks. Advances in neural information processing systems, 28, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.372, + 0.892, + 0.455 + ], + "angle": 0, + "content": "[19] Yifan Jiang, He Zhang, Jianming Zhang, Yilin Wang, Zhe Lin, Kalyan Sunkavalli, Simon Chen, Sohrab Amirghodsi, Sarah Kong, and Zhangyang Wang. Ssh: A self-supervised framework for image harmonization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4832-4841, 2021. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.456, + 0.892, + 0.511 + ], + "angle": 0, + "content": "[20] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.511, + 0.892, + 0.552 + ], + "angle": 0, + "content": "[21] Kevin Karsch, Varsha Hedau, David Forsyth, and Derek Hoiem. Rendering synthetic objects into legacy photographs. ACM Transactions on Graphics (TOG), 30(6):1-12, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.553, + 0.892, + 0.608 + ], + "angle": 0, + "content": "[22] Kevin Karsch, Kalyan Sunkavalli, Sunil Hadap, Nathan Carr, Hailin Jin, Rafael Fonte, Michael Sittig, and David Forsyth. Automatic scene inference for 3d object compositing. ACM Transactions on Graphics (TOG), 33(3):1-15, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.609, + 0.892, + 0.663 + ], + "angle": 0, + "content": "[23] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagc: Text-based real image editing with diffusion models. arXiv preprint arXiv:2210.09276, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.665, + 0.892, + 0.718 + ], + "angle": 0, + "content": "[24] Jean-Francois Lalonde and Alexei A Efros. Using color compatibility for assessing image realism. In 2007 IEEE 11th International Conference on Computer Vision, pages 1-8. IEEE, 2007. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.72, + 0.892, + 0.761 + ], + "angle": 0, + "content": "[25] Seung Hoon Lee, Seunghyun Lee, and Byung Cheol Song. Vision transformer for small-size datasets. arXiv preprint arXiv:2112.13492, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.762, + 0.894, + 0.817 + ], + "angle": 0, + "content": "[26] Youngwan Lee and Jongyoul Park. Centermask: Realtime anchor-free instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13906-13915, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.818, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[27] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. arXiv preprint arXiv:2201.12086, 2022. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[28] Chen-Hsuan Lin, Ersin Yumer, Oliver Wang, Eli Shechtman, and Simon Lucey. St-gan: Spatial transformer generative" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.895, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18318" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.093, + 0.469, + 0.134 + ], + "angle": 0, + "content": "adversarial networks for image compositing. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 9455-9464, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.469, + 0.219 + ], + "angle": 0, + "content": "[29] Daquan Liu, Chengjiang Long, Hongpan Zhang, Hanning Yu, Xinzhi Dong, and Chunxia Xiao. Arshadowgan: Shadow generative adversarial network for augmented reality in single light scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8139-8148, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.221, + 0.469, + 0.289 + ], + "angle": 0, + "content": "[30] Xihui Liu, Dong Huk Park, Samaneh Azadi, Gong Zhang, Arman Chopikyan, Yuxiao Hu, Humphrey Shi, Anna Rohrbach, and Trevor Darrell. More control for free! image synthesis with semantic diffusion guidance. arXiv preprint arXiv:2112.05744, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.291, + 0.469, + 0.347 + ], + "angle": 0, + "content": "[31] Shitong Luo and Wei Hu. Diffusion probabilistic models for 3d point cloud generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2837-2845, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.349, + 0.469, + 0.403 + ], + "angle": 0, + "content": "[32] Chenlin Meng, Yang Song, Jiaming Song, Jiajun Wu, JunYan Zhu, and Stefano Ermon. Sdedit: Image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.406, + 0.469, + 0.473 + ], + "angle": 0, + "content": "[33] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.476, + 0.469, + 0.516 + ], + "angle": 0, + "content": "[34] Patrick Pérez, Michel Gangnet, and Andrew Blake. Poisson image editing. In ACM SIGGRAPH 2003 Papers, pages 313-318. 2003. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.519, + 0.469, + 0.56 + ], + "angle": 0, + "content": "[35] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.563, + 0.469, + 0.645 + ], + "angle": 0, + "content": "[36] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 4, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.647, + 0.469, + 0.714 + ], + "angle": 0, + "content": "[37] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.718, + 0.469, + 0.786 + ], + "angle": 0, + "content": "[38] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.789, + 0.469, + 0.844 + ], + "angle": 0, + "content": "[39] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. arXiv preprint arXiv:2208.12242, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.846, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[40] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[41] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.192, + 0.892, + 0.261 + ], + "angle": 0, + "content": "[42] Yichen Sheng, Yifan Liu, Jianming Zhang, Wei Yin, A Cengiz Oztireli, He Zhang, Zhe Lin, Eli Shechtman, and Bedrich Benes. Controllable shadow generation using pixel height maps. In European Conference on Computer Vision, pages 240-256. Springer, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.263, + 0.892, + 0.318 + ], + "angle": 0, + "content": "[43] Yichen Sheng, Jianming Zhang, and Bedrich Benes. Ssn: Soft shadow network for image compositing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4380-4390, 2021. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.32, + 0.892, + 0.386 + ], + "angle": 0, + "content": "[44] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256-2265. PMLR, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.389, + 0.892, + 0.431 + ], + "angle": 0, + "content": "[45] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in Neural Information Processing Systems, 32, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.433, + 0.892, + 0.486 + ], + "angle": 0, + "content": "[46] Kalyan Sunkavalli, Micah K Johnson, Wojciech Matusik, and Hanspeter Pfister. Multi-scale image harmonization. ACM Transactions on Graphics (TOG), 29(4):1-10, 2010. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.489, + 0.892, + 0.544 + ], + "angle": 0, + "content": "[47] Yi-Hsuan Tsai, Xiaohui Shen, Zhe Lin, Kalyan Sunkavalli, Xin Lu, and Ming-Hsuan Yang. Deep image harmonization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3789-3797, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.546, + 0.892, + 0.601 + ], + "angle": 0, + "content": "[48] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.602, + 0.892, + 0.655 + ], + "angle": 0, + "content": "[49] Shaoan Xie, Zhifei Zhang, Zhe Lin, Tobias Hinz, and Kun Zhang. Smartbrush: Text and shape guided object inpainting with diffusion model. arXiv preprint arXiv:2212.05034, 2022.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.659, + 0.892, + 0.713 + ], + "angle": 0, + "content": "[50] Ning Xu, Brian Price, Scott Cohen, and Thomas Huang. Deep image matting. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2970-2979, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.716, + 0.892, + 0.77 + ], + "angle": 0, + "content": "[51] Ben Xue, Shenghui Ran, Quan Chen, Rongfei Jia, Binqiang Zhao, and Xing Tang. Dccf: Deep comprehensible color filter learning framework for high-resolution image harmonization. arXiv preprint arXiv:2207.04788, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.772, + 0.892, + 0.827 + ], + "angle": 0, + "content": "[52] Hao Zhou, Sunil Hadap, Kalyan Sunkavalli, and David W Jacobs. Deep single-image portrait relighting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7194-7202, 2019. 1" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.827 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18319" + } + ] +] \ No newline at end of file diff --git a/2023/ObjectStitch_ Object Compositing With Diffusion Model/69193070-f3e5-4dd6-8743-b15c03b3e677_origin.pdf b/2023/ObjectStitch_ Object Compositing With Diffusion Model/69193070-f3e5-4dd6-8743-b15c03b3e677_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0e191c5574861c9772b4f87dab2086456707d556 --- /dev/null +++ b/2023/ObjectStitch_ Object Compositing With Diffusion Model/69193070-f3e5-4dd6-8743-b15c03b3e677_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:80309744b847d8802b96c40d4d4083b945f402ca4fc08b2cfa0b81477f4e7ef5 +size 6616545 diff --git a/2023/ObjectStitch_ Object Compositing With Diffusion Model/full.md b/2023/ObjectStitch_ Object Compositing With Diffusion Model/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5f3be464d5b33bc4cc2d9df14f6224c45bea4a92 --- /dev/null +++ b/2023/ObjectStitch_ Object Compositing With Diffusion Model/full.md @@ -0,0 +1,303 @@ +# ObjectStitch: Object Compositing with Diffusion Model + +Yizhi Song $^{1*}$ , Zhifei Zhang $^{2}$ , Zhe Lin $^{2}$ , Scott Cohen $^{2}$ , Brian Price $^{2}$ , Jianming Zhang $^{2}$ , Soo Ye Kim $^{2}$ , Daniel Aliaga $^{1}$ Purdue University $^{1}$ , Adobe Research $^{2}$ + +$^{1}\{ \text{song630,aliaga} \} @$ urdue.edu, $^{2}\{ \text{zhang,zlin,scohen,bprice, jianmzha,sooyek} \} @$ adobe.com + +![](images/22213a7b75764b52e031010cdd4d202a63f5ed7ff1196f9033550863a307d181.jpg) +(a) copy-paste +Figure 1. Example results of object compositing with (a) copy-and-paste scheme, (b) traditional compositing pipeline and (c) ours (Object-Stitch). Traditional compositing pipeline is done with best possible off-the-shelf models including foreground/background color harmonization [19,47], poisson blending [34], and shadow synthesis [42]. ObjectStitch achieves more realistic results, and can address geometry correction, harmonization, shadow generation, and view synthesis all-in-one while preserving similar appearance to the reference object. + +![](images/cfd5085ea9132eb738ebc7a687b25f8531b86b2607301cc275614fd8d0fc339e.jpg) +(b) traditional compositing pipeline + +![](images/588dbed269c6232e2cee162781a76218dfafae91d8f5ccc6ae21ed7a2750ced8.jpg) +(c) ours + +![](images/5b6e3e861229a4d5e838350ba8389cc3452f4f619b8265e23db05db68dae1317.jpg) +(a) copy-paste + +![](images/235043b1565a429c9e4344fbe0911f02383e936001c2df3ddd3c7482d2f6ceec.jpg) +(b) traditional compositing pipeline + +![](images/ac24a6ff2324588107bb51f60953fff77c50e285fc565a9ec760772b3f6ef057.jpg) +(c) ours + +# Abstract + +Object compositing based on 2D images is a challenging problem since it typically involves multiple processing stages such as color harmonization, geometry correction and shadow generation to generate realistic results. Furthermore, annotating training data pairs for compositing requires substantial manual effort from professionals, and is hardly scalable. Thus, with the recent advances in generative models, in this work, we propose a self-supervised framework for object compositing by leveraging the power of conditional diffusion models. Our framework can holistically address the object compositing task in a unified model, transforming the viewpoint, geometry, color and shadow of the generated object while requiring no manual labeling. To preserve the input object's characteristics, we introduce a content adaptor that helps to maintain categori + +cal semantics and object appearance. A data augmentation method is further adopted to improve the fidelity of the generator. Our method outperforms relevant baselines in both realism and faithfulness of the synthesized result images in a user study on various real-world images. + +# 1. Introduction + +Image compositing is an essential task in image editing that aims to insert an object from a given image into another image in a realistic way. Conventionally, many sub-tasks are involved in composing an object to a new scene, including color harmonization [6, 7, 19, 51], relighting [52], and shadow generation [16, 29, 43] in order to naturally blend the object into the new image. As shown in Tab. 1, most previous methods [6, 7, 16, 19, 28, 43] focus on a single sub-task required for image compositing. Consequently, they must be appropriately combined to obtain a composite image where the input object is re-synthesized to have the + +
MethodGeometryLightShadowView
ST-GAN [28]✓XXX
SSH [19]X✓XX
DCCF [51]X✓XX
SSN [43]XX✓X
SGRNet [16]XX✓X
GCC-GAN [5]✓✓XX
Ours✓✓✓✓
+ +Table 1. Prior works only focus on one or two aspects of object compositing, and they cannot synthesize novel views. In contrast, our model can address all perspectives as listed. + +color, lighting and shadow that is consistent with the background scene. As shown in Fig. 1, results produced in this way still look unnatural, partly due to the viewpoint of the inserted object being different from the overall background. + +Harmonizing the geometry and synthesizing novel views have often been overlooked in 2D image compositing, which require an accurate understanding of both the geometry of the object and the background scene from 2D images. Previous works [15, 21, 22] handle 3D object compositing with explicit background information such as lighting positions and depth. More recently, [28] utilize GANs to estimate the homography of the object. However, this method is limited to placing furniture in indoor scenes. In this paper, we propose a generic image compositing method that is able to harmonize the geometry of the input object along with color, lighting and shadow with the background image using a diffusion-based generative model. + +In recent years, generative models such as GANs [4, 10, 17, 20] and diffusion models [1, 14, 30, 32, 33, 38, 39] have shown great potential in synthesizing realistic images. In particular, diffusion model-based frameworks are versatile and outperform various prior methods in image editing [1, 23, 32] and other applications [11, 31, 35]. However, most image editing diffusion models focus on using text inputs to manipulate images [1, 3, 9, 23, 33], which is insufficient for image compositing as verbal representations cannot fully capture the details or preserve the identity and appearance of a given object image. There have been recent works [23, 39] focusing on generating diverse contexts while preserving the key features of the object; however, these models are designed for a different task than object compositing. Furthermore, [23] requires fine-tuning the model for each input object and [39] also needs to be finetuned on multiple images of the same object. Therefore they are limited for general object compositing. + +In this work, we leverage diffusion models to simultaneously handle multiple aspects of image compositing such as color harmonization, relighting, geometry correction and shadow generation. With image guidance rather than text guidance, we aim to preserve the identity and appearance of + +the original object in the generated composite image. + +Specifically, our model synthesizes a composite image given (i) a source object, (ii) a target background image, and (iii) a bounding box specifying the location to insert the object. The proposed framework consists of a content adaptor and a generator module: the content adaptor is designed to extract a representation from the input object containing both high-level semantics and low-level details such as color and shape; the generator module preserves the background scene while improving the generation quality and versatility. Our framework is trained in a fully self-supervised manner and no task-specific labeling is required at any point during training. Moreover, various data augmentation techniques are applied to further improve the fidelity and realism of the output. We evaluate our proposed method on a real-world dataset closely simulating real use cases for image compositing. + +Our contributions are summarized as follows: + +- We present the first diffusion model-based framework for generative object compositing that can handle multiple aspects of compositing such as viewpoint, geometry, lighting and shadow. +- We propose a content adaptor module which learns a descriptive multi-modal embedding from images, enabling image guidance for diffusion models. +- Our framework is trained in a self-supervised manner without any task-specific annotations, employing data augmentation techniques to improve the fidelity of generation. +- We collect a high-resolution real-world dataset for object compositing with diverse images, containing manually annotated object scales and locations. + +# 2. Related Work + +# 2.1. Image Compositing + +Image compositing is a challenging task in reference-guided image editing, where the object in a given foreground image is to be inserted into a background image. The generated composite image is expected to look realistic, with the appearance of the original object being preserved. Prior works often focus on an individual aspect of this problem, such as geometric correction [2, 28], image harmonization [6, 7, 51], image matting [50] and shadow generation [29]. + +Lin et al. [28] propose an iterative GAN-based framework to correct geometric inconsistencies between object and background images. In their model, STNs [18] are integrated into the generator network to predict a series of warp updates. Although their work can improve the realism of the scene geometry in the composite image, it is limited to inserting furniture into indoor scenes. Azadi et al. [2] focus on binary composition, using a composition-decomposition + +network to capture interactions between a pair of objects. + +Image harmonization aims at minimizing the inconsistency between the input object and the background image. Traditional methods [24, 46] usually concentrate on obtaining color statistics and then transfer this information between foreground and background. Recent works seek to solve this problem via deep neural networks. [6, 7] reformulate image harmonization as a domain adaptation problem, while [19] converts it to a representation fusing problem and utilizes self-supervised training. + +Shadow synthesis is an effect that is often overlooked in previous image compositing methods although it is essential for generating realistic composite images. SGRNet [16] divides this task into a shadow mask generation stage and a shadow filling stage. SSN [43], focusing on soft shadows, predicts an ambient occlusion map as cue for shadow generation. + +There are also works simultaneously addressing multiple sub-problems of image compositing. Chen et al. [5] develop a system including multiple network modules to produce plausible images with geometric, color and boundary consistency. However, the pose of the generated object is constrained by the mask input and the model cannot generalize to non-rigid objects such as animals. Our proposed model also handles the image composition problem in a unified manner, generating foreground objects which are harmonious and geometrically consistent with the background while synthesizing novel views and shadows. + +# 2.2. Guided Image Synthesis + +Recent years have witnessed great advances in diffusion models. Diffusion models are a family of deep generative models based on several predominant works [14, 44, 45], defined with two Markov chains. In the forward process, they gradually add noise to the data, and in the reverse process, they learn to recover the data from noise. Following the growth of research in this area, diffusion models have shown their potential in a great number of applications, such as text-to-image generation [9, 33, 38], image editing [1, 23, 32] and guided image synthesis [30]. + +Text-to-image synthesis has been a popular research topic over the past few years. Stable Diffusion [38] made a significant contribution to this task and it reduces the computational cost by applying the diffusion model in the latent space. Other works explore ways of more flexible and controllable text-driven image editing. Avrahami et al. [1] design an algorithm in their Blended Diffusion framework to fuse noised versions of the input with the local text-guided diffusion latent, generating smooth transition. Similarly, GLIDE [33] is also capable of handling text-guided inpainting after fine-tuning on this specific task. Some methods give users more straightforward control on images. SDEdit [32] allows stroke painting in images, blending user input + +into the image by adding noise to the input and denoising through a stochastic differential equation. SDG [30] injects Semantic Diffusion Guidance at each iteration of the generation and can provide multi-modal guidance. + +While the versatility and generation quality of diffusion models have been repeatedly demonstrated, preserving object appearance remains a very challenging problem in image editing. Ruiz et al. [39] address this problem by optimizing their model using a reconstruction loss and a class-specific prior preservation loss. Though their approach can preserve the details of the object, multiple images of the same object are required to fine-tune their model on this target object. Kawar et al. [23] handle this issue by interpolating the initial text embedding and the optimized embedding for reconstruction, but their method is mainly designed for applying edits on objects, not for generating the same object in different contexts. In this paper, we propose the first generative object compositing framework based on diffusion models, which generates the harmonized and geometrically-corrected subject with a novel view and shadow. Furthermore, the characteristics of the original object is preserved in the synthesized composite image. + +# 3. Proposed Method + +We define the generative object compositing problem as: Given an input triplet $(I_o, I_{bg}, M)$ that consists of an object image $I_o \in \mathbb{R}^{H_s \times W_s \times 3}$ , a background image $I_{bg} \in \mathbb{R}^{H_t \times W_t \times 3}$ , and its associated binary mask $M \in \mathbb{R}^{H_t \times W_t \times 1}$ with the desired object location set to 0 and the rest to 1, the goal is to compose the input object into the masked area $I_{bg} \otimes M$ . $M$ is considered as a soft constraint of the location and scale of the composited object. The output generated image should look realistic, while the appearance of the object is preserved. Our problem setting is different from text-guided image generation and inpainting in that the condition input is a reference object image rather than a text prompt. Inspired by the success of text-guided diffusion models [37, 38, 40], which inject text conditioning into the diffusion architecture, we design our method for generative compositing to leverage such pretrained models. + +An overview of our framework is shown in Fig. 2. It consists of an object image encoder extracting semantic features from the object and a conditional diffusion generator. To leverage the power of pretrained text-to-image diffusion models, we introduce a content adaptor that can bridge the gap between the object encoder and conditional generator by transforming a sequence of visual tokens to a sequence of text tokens to overcome the domain gap between image and text. This design further helps to preserve the object appearance. We propose a two-stage training process: in the first stage, the content adaptor is trained on large image/text pairs to maintain high-level semantics of the object; in the second stage, it is trained in the context of diffusion + +![](images/07cd73ec57e3862f79bb3a725a0012186561a38f2767aac1c67b8d92bf5e52a4.jpg) +Figure 2. System pipeline. Our framework consists of a content adaptor and a generator (a pretrained text-to-image diffusion model). The input image $I_{o}$ is fed into a ViT and the adaptor which produces a descriptive embedding. At the same time the background image $I_{bg}$ is taken as input by the diffusion model. At each iteration during the denoising stage, we apply the mask $M$ on the generated image $I_{out}$ , so that the generator only denoises the masked area $I_{out} \otimes M$ . + +generator to encode key identity features of the object by encouraging the visual reconstruction of the object in the original image. Lastly, the generator module is fine-tuned on the embedding produced by the adaptor through cross attention blocks. All stages our trained in a self-supervised manner to avoid expensive annotation for obtaining object compositing training data. + +# 3.1. Generator + +As depicted in Fig. 2, we leverage a pretrained text-to-image diffusion model architecture and modify it for the compositing task by (i) introducing a mask in the input, and (ii) adjusting the U-Net input to contain the original background image outside the hole and noise inside the hole with mask blending. In order to condition the model on the guidance embedding $E$ , an attention mechanism is applied as: + +$$ +\operatorname {S o f t m a x} \left(\frac {\left(\boldsymbol {W} _ {Q} E _ {x}\right) \left(\boldsymbol {W} _ {K} E\right) ^ {T}}{\sqrt {d}}\right) \boldsymbol {W} _ {V} E = \boldsymbol {A} \boldsymbol {V} \tag {1} +$$ + +where $E_{x}$ is an intermediate representation of the denoising autoencoder and $W_{Q} \in \mathbb{R}^{d \times d_{x}}$ , $W_{K} \in \mathbb{R}^{d \times d_{e}}$ and $W_{V} \in \mathbb{R}^{d \times d_{e}}$ are embedding matrices. The background outside the mask area should be perfectly preserved for generative object compositing. Thus, we use the input mask $M$ for blending the background image $I_{bg}$ with the generated image $I_{out}$ . As a result, the generator only denoises the masked area $I_{out} \otimes M$ . To use this model on our task, a straightforward way (a baseline) would be to apply image captioning on the object image and feeding the resulting caption directly to the diffusion model as the condition. However, text embedding cannot capture fine grained visual details. Therefore, in order to directly leverage the given object image, we introduce the content adaptor to transform the visual features from a pretrained visual encoder to text features (tokens) to use as conditioning for the generator. + +![](images/af112cf73253444a877283eaf3708010be628359e7496df8e70d58a36a5dbea1.jpg) +Figure 3. Structure of the Content Adaptor. In the first stage, it is trained on a large dataset of image-caption pairs to learn multimodal sequential embeddings containing high-level semantics. In the second stage, it is fine-tuned under the diffusion framework to learn to encode identity features in adaptive embedding. + +# 3.2. Content Adaptor + +To prevent the loss of key identity information, we use an image encoder instead of a text encoder to produce the embedding from the input object image. However, the image embedding cannot be effectively utilized by the diffusion model for two reasons: + +- The image embedding $\widetilde{E}$ and the text embedding $E$ are from different domains. Since the diffusion model was trained on $E$ , it cannot generate meaningful contents from image embedding sequence; +- A mismatch in the dimensions of $\widetilde{E} \in \mathbb{R}^{k \times 257 \times 1024}$ and $E \in \mathbb{R}^{k \times 77 \times 768}$ , where $k$ is batch-size. + +Therefore, based on the above observations, we develop a sequence-to-sequence translator architecture as shown in Fig. 3. Given an image-caption pair as an input tuple $(I,t)$ , we employ two pretrained ViT-L/14 encoders $C_t$ , $C_i$ from CLIP [36] to produce text embedding $E = C_t(t)$ and image embedding $\widetilde{E} = C_i(I)$ , respectively. The adaptor $T$ + +![](images/2c40277be9bbef09e48d76f757a40fa7941352eabd768d0de4773d2a5d6c5f0d.jpg) +Figure 4. Illustration of our synthetic data generation and data augmentation scheme. The top row shows the data generation process including perspective warping, random rotation, and random color shifting. The original image is used as both the input background and ground truth, while the perturbed object is fed into the adaptor. The bottom row shows crop and shift augmentations, which help to improve the generation quality and preserve object details. + +consists of three components: a 1D convolutional layer, attention blocks [48] and an MLP, where the 1D convolution modifies the length of the embedding from 257 to 77, the MLP maps the embedding dimension from 1024 to 768, and the attention blocks bridge the gap between text domain and image domain. We design a two-stage optimization method to train this module, which is explained in Sec. 3.3.2. + +# 3.3. Self-supervised Framework + +There is no publicly available image compositing training dataset with annotations which is sufficient for training a diffusion model, and it is extremely challenging to manually annotate such data. Therefore, we propose a self-supervised training scheme and a synthetic data generation approach that simulate real-world scenarios. We also introduce a data augmentation method to enrich the training data as well as improve the robustness of our model. + +# 3.3.1 Data Generation and Augmentation + +Training data generation. We collect our synthetic training data from Pixabay and use an object instance segmentation model [26] to predict panoptic segmentation and classification labels. We first filter the dataset by removing objects with very small or large sizes. Then, we apply spatial and color perturbations to simulate many real use-case scenarios where the input image and background image have different scene geometry and lighting conditions. + +The top row of Fig. 4 illustrates this process. Inspired by [8], we randomly perturb the four points of the object bounding box to apply projective transformation, followed by a random rotation within the range $[- \theta, \theta]$ ( $\theta = 20^{\circ}$ ) and color perturbation. The segmentation mask (perturbed in the same way as the image) is used to extract the object. + +This data synthesis pipeline is fully controllable and can + +also be employed as data augmentation during training. Another key advantage is that it is free of manual labeling, since the original image is used as the ground truth. We use the bounding box as the mask as it not only fully covers the object, but also extends to its neighboring area (providing room for shadow generation). We find that it is flexible enough for the model to apply spatial transformations, synthesize novel views and generate shadows and reflection. + +Real-world evaluation data. Given that there is no existing dataset specifically for our task, we manually collect a new dataset, closely simulating real-world use cases, as an evaluation benchmark for object compositing. The dataset consists of 503 pairs of common objects (including both rigid and non-rigid objects such as vehicles and wildlife) and diverse background images (covering both indoor and outdoor scenes). It also contains challenging cases where there is a large mismatch of lighting conditions or viewpoints between foreground and background. The dataset images are collected from Pixabay. The labeling procedure closely simulate the real scenarios where the input object is placed at a target location in the background image and then scaled at the user's will. The compositing region is determined as a loose bounding box around the object. + +Data augmentation. Inspired by [25], we introduce random shift and crop augmentations during training, while ensuring that the foreground object is always contained in the crop window. This process is illustrated in the bottom row of Fig. 4. Applying this augmentation method for both training and inference results in a notable improvement in the realism of the generated results. Quantitative results are provided in Sec. 4.4. + +# 3.3.2 Training + +Content adaptor pretraining. We first pretrain the content adaptor to keep the semantics of the object by mapping the image embedding to text embedding. At this first stage, we optimize the content adaptor on a sequence-to-sequence translation task, which learns to project the image embedding into a multi-modal space. During training, $C_t$ , $C_i$ are frozen, and the translator is trained on 3,203,338 image-caption pairs from a filtered LAION dataset [41]. + +Given the input image embedding $\vec{E}$ , we use the text embedding $E$ as target, the objective function of this translation task is defined as: + +$$ +\mathcal {L} _ {\text {d i s t}} = \| T (\tilde {E}) - E \| _ {1}, \tag {2} +$$ + +where $T(\cdot)$ is the content adaptor. + +However, the multi-modal embedding obtained solely from this first stage optimization mostly carry high-level semantics of the input image without much of texture details. Hence, we further refine $\widetilde{E}$ to obtain $\hat{E}$ , the adaptive embedding, for better appearance preservation. + +Content adaptor fine-tuning. We further optimize the content adaptor to produce an adaptive embedding which maintains instance-level properties of the object. After pretraining, we insert the content adaptor to the pretrained diffusion framework by feeding the adaptive embedding as context to the attention blocks. Then, the diffusion model is frozen and the adaptor is trained using: + +$$ +\mathcal {L} _ {\text {a d a p t}} = \mathbb {E} _ {T, \epsilon \sim \mathcal {N} (0, 1)} [ \| \epsilon - \epsilon_ {\theta} (I _ {t} \circ M, t, T (\widetilde {E})) \| _ {2} ^ {2} ], \tag {3} +$$ + +where the content adaptor $T(\cdot)$ is optimized. It is trained on our synthetic dataset filtered from Pixabay, containing 467,280 foreground and background image pairs for training and 25,960 pairs for validation. + +Generator fine-tuning. After the aforementioned two-stage training of the content adaptor is completed, we freeze the content adaptor and train the generator module after initializing the text-to-image diffusion model with pretrained weights. Crop and shift augmentations are applied in this process. Based on the Latent Diffusion model [38], the objective loss in our generator module is defined by: + +$$ +\mathcal {L} _ {g e n} = \mathbb {E} _ {\hat {E}, \epsilon \sim \mathcal {N} (0, 1)} [ \| \epsilon - \epsilon_ {\theta} (I _ {t} \circ M, t, \hat {E}) \| _ {2} ^ {2} ], \tag {4} +$$ + +where $I$ is the input image, $I_{t}$ is a noisy version of $I$ at timestep $t$ , and $\epsilon_{\theta}$ denotes the denoising model which is optimized. In order to adapt the text-guided generation task to the setting of image compositing, we apply the input mask on the image at every time-step. + +# 4. Experiments + +# 4.1. Training Details + +The first training stage of the content adaptor takes 15 epochs with a learning rate of $10^{-4}$ and batch size of 2048. The image size processed by ViT is $224 \times 224$ . To set up the pipeline, we employed the pretrained image encoder and text encoder from CLIP [36]. The second training stage of the content adaptor takes 13 epochs with the learning rate of $2 \times 10^{-5}$ and batch size of 512. We resize the images to 512 $\times$ 512 and adapt the pretrained Stable Diffusion model [38] to our task. After the above process, the generator is trained for 20 epochs with the learning rate of $4 \times 10^{-5}$ and batch size of 576. The input image size is the same as that in the second training stage. All training stages are conducted on 8 A100 GPUs with Adam optimizer. + +# 4.2. Quantitative Evaluation + +To obtain quantitative results on a real dataset that simulates real-world use cases, where users drag and drop an object to a background image, we conduct a user study to compare to three baselines, i.e., copy-and-paste, pretrained stable diffusion model fine-tuned on BLIP [27], and a joint stable diffusion model integrating both BLIP + +![](images/7762c2ac838214911580c6ca9dcb9182697c5f410c8a7d739d199a38570455ec.jpg) +Pick from image A and image B the one that you think looks more realistic. + +![](images/d8c93f8d424c09f5d59ffa0015cf45ac8908a35a86fbd674ee179dacb39c85e1.jpg) +Which generated object do you think is most likely to be the same one from the guidance? +Figure 5. User study results. We conduct side-by-side comparisons between our method and one of baseline methods to quantify the generation quality in terms of realism and appearance preservation. The results show that our method outperforms baselines. + +and SDEdit [32]. Given the fact that there is no existing diffusion-based method addressing the same problem, we train the baselines on the same pretrained diffusion model using the same self-supervised training scheme and dataset as ours. To the best of our knowledge, after adapting to our setting, they are the closest baselines to our task. In the BLIP baseline, the content adaptor is replaced by BLIP, a state-of-the-art captioning model. The text embedding obtained from the BLIP caption is fed into the generator module. For the SDEdit baseline, we keep the BLIP model which will significantly improve the generation fidelity of a pretrained diffusion model. We set the noise strength to 0.65 in SDEdit, striking a balance between realism and faithfulness. Details are provided in the supplementary material. + +We perform a user study using our real testing dataset, which consists of 503 object-background pairs. For each example, we set up a side-by-side comparison between the results from one of the baselines and our method, respectively. The subject will be asked two questions: 1) which is more realistic and 2) which generated object is more close to the given object? We collected 1,494 votes from more than 170 users. As shown in Fig. 5, our method outperforms the other baselines for both questions. The higher preference rates demonstrate the effectiveness of the content adaptor and data augmentation method in improving generation fidelity. It also shows that our content adaptor is capable of maintaining details and attributes of the input objects, while not being constrained by the input when harmonizing objects to the background. In contrast, although SDEdit is better at preserving appearance than BLIP, it is constrained by the original pose, shape, and other attributes of the object, making the generated object less coherent with the background. + +# 4.3. Qualitative Evaluation + +In Fig. 6, we compare our method to various baselines. In the first two rows of Fig. 6, we place the same object + +![](images/82d0bf282400d201f0171608e61f21fa777e64396600d4c3d2837c47a876cca9.jpg) +Object +Target Image +Copy-and-paste +BLIP +SDEdit +Ours +Figure 6. Qualitative comparison on the real-world evaluation dataset. Our method can appropriately correct the viewpoint and geometry while harmonizing the object with the new background, generating realistic composite results. Moreover, lighting and shadows are naturally adjusted in the composite image. Compared to baselines using text guidance, our method can better preserve the object appearance. + +
AdaptorOptimizationAugmentationBLIP [27]FID ↓CLIP text score ↑CLIP image score ↑
✓✓✓✗15.433129.859497.0625
✗✓✓✗18.825429.828196.0000
✓✗✓✗16.138129.796996.8125
✓✓✗✗18.369829.781396.1875
✗✗✓✓17.900029.718895.8125
+ +Table 2. Ablation study. We evaluate the effectiveness of the components listed in the first four columns: 1) whether the content adaptor is used; 2) whether the adaptor is optimized for appearance preservation; 3) whether crop augmentation is utilized in training; 4) if the adaptor module is replaced by BLIP to predict text embedding from an image. We use FID to assess the generation fidelity and modified CLIP score (explained in Sec. 4.4) to measure similarity of the guidance and the predicted image. + +in two scenes with notable differences in the geometry. Results show that our model is capable of correcting geometric inconsistencies while preserving the characteristics of the object. The limitations of the baselines are also shown in Fig. 6: BLIP often fails to preserve the appearance; SDEdit preserves texture and pose better than the first baseline, but often cannot make appropriate geometry transformations. + +# 4.4. Ablation Study + +In Tab. 2, we demonstrate the effectiveness of each key component in our ObjectStitch framework by ablating each of them. We use FID [13] and a modified CLIP [12, 36] score, which measure the fidelity and semantic matching performance between given and generated objects. Instead of directly adopting the CLIP score, which semantically matches the prompt and the image, we modify it to CLIP text score and CLIP image score that are defined as follows: + +$$ +\mathcal {C} _ {t x t} = E \left[ s \cdot f \left(I _ {p r e d}\right) \cdot g \left(B \left(I _ {g t}\right)\right) \right], \tag {5} +$$ + +$$ +\mathcal {C} _ {i m g} = E [ s \cdot f (I _ {p r e d}) \cdot f (I _ {g t}) ], \tag {6} +$$ + +where $B(\cdot)$ is pretrained BLIP [27], and $s$ is a logit scale. + +For the ablation models, we first remove the content adaptor to evaluate its effect. In order to keep the embedding dimension fed to the generator, we removed the attention blocks from the content adaptor. We also drop the bias and activation function of the last linear layer, such that in the adaptation process, the model will only learn a non-adaptive embedding dictionary. Without the attention layers, the model has difficulty bridging the domain gap between text and image, resulting in the degradation of generation fidelity as shown in the second row of Tab. 2. + +The third row in Tab. 2 ablates the second training stage of the content adaptor by simply skipping it. Without this process, the model fails to learn descriptive embedding, which encodes instance-level features of the guidance object, leading to a drop in appearance preservation. + +Data augmentation is another key element for better performance, i.e., cropping and shift augmentation increases the mask area, thus improving the generation quality in detail. The fourth row in Tab. 2 illustrates that with the ab + +sence of data augmentation, there is a notable drop in realism, and the generation carries less accurate details. + +To further assess the importance of the content adaptor module, we replace it with a pretrained BLIP model, which transfers the reference image to a text embedding. As shown in the last row in Tab. 2, this results in the lowest CLIP image score, indicating that high-level semantics represented by text are not sufficient for maintaining object identity, highlighting the importance of our content adaptor. + +# 5. Conclusion and Future Work + +We propose the first diffusion-based approach to tackle object compositing. Traditional methods require many steps including harmonization, geometry/view adjustment, and shadow generation. By contrast, our method directly achieves realistic compositing. We construct our model from a text-to-image generation network and introduce a novel content adaptor module. Furthermore, we present a fully self-supervised framework to train our model with synthetic data generation and augmentation. We show our superior performance over baseline methods through user studies on real-world examples. + +Since we give the first attempt to handle this challenging task, our method has several limitations. It lacks control over the appearance preservation of the synthesized object. Potential solutions are training the visual encoder on pairs of images with the same object, or training it jointly with the UNet. Another limitation is caused by masking the output image, which prohibits our model from generating global effects, i.e., the shadow can only be synthesized within the mask. To address this issue, we may need to synthesize complete background images and pair them with objects, and train with a conditional generation framework. We also observed artifacts around the object in the background. This can be improved by training the model to also predict an instance mask [49]. We will improve this in our future work. + +# Acknowledgements + +This work is partially supported by NSF grants 1835739 and 2106717. We also thank Zhiquan Wang for discussion. + +# References + +[1] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18208-18218, 2022. 2, 3 +[2] Samaneh Azadi, Deepak Pathak, Sayna Ebrahimi, and Trevor Darrell. Compositional gan: Learning image-conditional binary composition. International Journal of Computer Vision, 128(10):2570-2585, 2020. 2 +[3] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, et al. edifi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 2 +[4] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. arXiv preprint arXiv:1809.11096, 2018. 2 +[5] Bor-Chun Chen and Andrew Kae. Toward realistic image compositing with adversarial learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8415-8424, 2019. 2, 3 +[6] Wenyan Cong, Li Niu, Jianfu Zhang, Jing Liang, and Liqing Zhang. Bargainnet: Background-guided domain translation for image harmonization. In 2021 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2021. 1, 2, 3 +[7] Wenyan Cong, Jianfu Zhang, Li Niu, Liu Liu, Zhixin Ling, Weiyuan Li, and Liqing Zhang. Dovenet: Deep image harmonization via domain verification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8394-8403, 2020. 1, 2, 3 +[8] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Deep image homography estimation. arXiv preprint arXiv:1606.03798, 2016. 5 +[9] Zhida Feng, Zhenyu Zhang, Xintong Yu, Yewei Fang, Lanxin Li, Xuyi Chen, Yuxiang Lu, Jiaxiang Liu, Weichong Yin, Shikun Feng, et al. Ernie-vilg 2.0: Improving text-to-image diffusion model with knowledge-enhanced mixture-of-denoising-experts. arXiv preprint arXiv:2210.15257, 2022.2.3 +[10] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020. 2 +[11] Liu He, Yijuan Lu, John Corring, Dinei Florencio, and Cha Zhang. Diffusion-based document layout generation. arXiv preprint arXiv:2303.10787, 2023. 2 +[12] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021. 8 +[13] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 8 + +[14] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 2, 3 +[15] Yannick Hold-Geoffroy, Kalyan Sunkavalli, Sunil Hadap, Emiliano Gambaretto, and Jean-François Lalonde. Deep outdoor illumination estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7312–7321, 2017. 2 +[16] Yan Hong, Li Niu, and Jianfu Zhang. Shadow generation for composite image in real-world scenes. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 914-922, 2022. 1, 2, 3 +[17] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 2 +[18] Max Jaderberg, Karen Simonyan, Andrew Zisserman, et al. Spatial transformer networks. Advances in neural information processing systems, 28, 2015. 2 +[19] Yifan Jiang, He Zhang, Jianming Zhang, Yilin Wang, Zhe Lin, Kalyan Sunkavalli, Simon Chen, Sohrab Amirghodsi, Sarah Kong, and Zhangyang Wang. Ssh: A self-supervised framework for image harmonization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4832-4841, 2021. 1, 2, 3 +[20] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 2 +[21] Kevin Karsch, Varsha Hedau, David Forsyth, and Derek Hoiem. Rendering synthetic objects into legacy photographs. ACM Transactions on Graphics (TOG), 30(6):1-12, 2011. 2 +[22] Kevin Karsch, Kalyan Sunkavalli, Sunil Hadap, Nathan Carr, Hailin Jin, Rafael Fonte, Michael Sittig, and David Forsyth. Automatic scene inference for 3d object compositing. ACM Transactions on Graphics (TOG), 33(3):1-15, 2014. 2 +[23] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagc: Text-based real image editing with diffusion models. arXiv preprint arXiv:2210.09276, 2022. 2, 3 +[24] Jean-Francois Lalonde and Alexei A Efros. Using color compatibility for assessing image realism. In 2007 IEEE 11th International Conference on Computer Vision, pages 1-8. IEEE, 2007. 3 +[25] Seung Hoon Lee, Seunghyun Lee, and Byung Cheol Song. Vision transformer for small-size datasets. arXiv preprint arXiv:2112.13492, 2021. 5 +[26] Youngwan Lee and Jongyoul Park. Centermask: Realtime anchor-free instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13906-13915, 2020. 5 +[27] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. arXiv preprint arXiv:2201.12086, 2022. 6, 8 +[28] Chen-Hsuan Lin, Ersin Yumer, Oliver Wang, Eli Shechtman, and Simon Lucey. St-gan: Spatial transformer generative + +adversarial networks for image compositing. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 9455-9464, 2018. 1, 2 +[29] Daquan Liu, Chengjiang Long, Hongpan Zhang, Hanning Yu, Xinzhi Dong, and Chunxia Xiao. Arshadowgan: Shadow generative adversarial network for augmented reality in single light scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8139-8148, 2020. 1, 2 +[30] Xihui Liu, Dong Huk Park, Samaneh Azadi, Gong Zhang, Arman Chopikyan, Yuxiao Hu, Humphrey Shi, Anna Rohrbach, and Trevor Darrell. More control for free! image synthesis with semantic diffusion guidance. arXiv preprint arXiv:2112.05744, 2021. 2, 3 +[31] Shitong Luo and Wei Hu. Diffusion probabilistic models for 3d point cloud generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2837-2845, 2021. 2 +[32] Chenlin Meng, Yang Song, Jiaming Song, Jiajun Wu, JunYan Zhu, and Stefano Ermon. Sdedit: Image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 2, 3, 6 +[33] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 2, 3 +[34] Patrick Pérez, Michel Gangnet, and Andrew Blake. Poisson image editing. In ACM SIGGRAPH 2003 Papers, pages 313-318. 2003. 1 +[35] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 2 +[36] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 4, 6, 8 +[37] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 3 +[38] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2, 3, 6 +[39] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. arXiv preprint arXiv:2208.12242, 2022. 2, 3 +[40] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image + +diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022.3 +[41] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 5 +[42] Yichen Sheng, Yifan Liu, Jianming Zhang, Wei Yin, A Cengiz Oztireli, He Zhang, Zhe Lin, Eli Shechtman, and Bedrich Benes. Controllable shadow generation using pixel height maps. In European Conference on Computer Vision, pages 240-256. Springer, 2022. 1 +[43] Yichen Sheng, Jianming Zhang, and Bedrich Benes. Ssn: Soft shadow network for image compositing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4380-4390, 2021. 1, 2, 3 +[44] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256-2265. PMLR, 2015. 3 +[45] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in Neural Information Processing Systems, 32, 2019. 3 +[46] Kalyan Sunkavalli, Micah K Johnson, Wojciech Matusik, and Hanspeter Pfister. Multi-scale image harmonization. ACM Transactions on Graphics (TOG), 29(4):1-10, 2010. 3 +[47] Yi-Hsuan Tsai, Xiaohui Shen, Zhe Lin, Kalyan Sunkavalli, Xin Lu, and Ming-Hsuan Yang. Deep image harmonization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3789-3797, 2017. 1 +[48] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 5 +[49] Shaoan Xie, Zhifei Zhang, Zhe Lin, Tobias Hinz, and Kun Zhang. Smartbrush: Text and shape guided object inpainting with diffusion model. arXiv preprint arXiv:2212.05034, 2022.8 +[50] Ning Xu, Brian Price, Scott Cohen, and Thomas Huang. Deep image matting. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2970-2979, 2017. 2 +[51] Ben Xue, Shenghui Ran, Quan Chen, Rongfei Jia, Binqiang Zhao, and Xing Tang. Dccf: Deep comprehensible color filter learning framework for high-resolution image harmonization. arXiv preprint arXiv:2207.04788, 2022. 1, 2 +[52] Hao Zhou, Sunil Hadap, Kalyan Sunkavalli, and David W Jacobs. Deep single-image portrait relighting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7194-7202, 2019. 1 \ No newline at end of file diff --git a/2023/ObjectStitch_ Object Compositing With Diffusion Model/images.zip b/2023/ObjectStitch_ Object Compositing With Diffusion Model/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b27475dc4f5ea79f29e0b10de5b85062023dd848 --- /dev/null +++ b/2023/ObjectStitch_ Object Compositing With Diffusion Model/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:118610711b53543f6d7afed44e9a6942135e4dffc5e71af8b873d82168256e9d +size 701941 diff --git a/2023/ObjectStitch_ Object Compositing With Diffusion Model/layout.json b/2023/ObjectStitch_ Object Compositing With Diffusion Model/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4a236d0114c94f855268e1f06dc99a98699440ab --- /dev/null +++ b/2023/ObjectStitch_ Object Compositing With Diffusion Model/layout.json @@ -0,0 +1,8134 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 123, + 103, + 471, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 103, + 471, + 121 + ], + "spans": [ + { + "bbox": [ + 123, + 103, + 471, + 121 + ], + "type": "text", + "content": "ObjectStitch: Object Compositing with Diffusion Model" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 134, + 142, + 459, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 142, + 459, + 185 + ], + "spans": [ + { + "bbox": [ + 134, + 142, + 459, + 185 + ], + "type": "text", + "content": "Yizhi Song" + }, + { + "bbox": [ + 134, + 142, + 459, + 185 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 134, + 142, + 459, + 185 + ], + "type": "text", + "content": ", Zhifei Zhang" + }, + { + "bbox": [ + 134, + 142, + 459, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 134, + 142, + 459, + 185 + ], + "type": "text", + "content": ", Zhe Lin" + }, + { + "bbox": [ + 134, + 142, + 459, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 134, + 142, + 459, + 185 + ], + "type": "text", + "content": ", Scott Cohen" + }, + { + "bbox": [ + 134, + 142, + 459, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 134, + 142, + 459, + 185 + ], + "type": "text", + "content": ", Brian Price" + }, + { + "bbox": [ + 134, + 142, + 459, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 134, + 142, + 459, + 185 + ], + "type": "text", + "content": ", Jianming Zhang" + }, + { + "bbox": [ + 134, + 142, + 459, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 134, + 142, + 459, + 185 + ], + "type": "text", + "content": ", Soo Ye Kim" + }, + { + "bbox": [ + 134, + 142, + 459, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 134, + 142, + 459, + 185 + ], + "type": "text", + "content": ", Daniel Aliaga" + }, + { + "bbox": [ + 134, + 142, + 459, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 134, + 142, + 459, + 185 + ], + "type": "text", + "content": " Purdue University" + }, + { + "bbox": [ + 134, + 142, + 459, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 134, + 142, + 459, + 185 + ], + "type": "text", + "content": ", Adobe Research" + }, + { + "bbox": [ + 134, + 142, + 459, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 72, + 186, + 517, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 186, + 517, + 200 + ], + "spans": [ + { + "bbox": [ + 72, + 186, + 517, + 200 + ], + "type": "inline_equation", + "content": "^{1}\\{ \\text{song630,aliaga} \\} @" + }, + { + "bbox": [ + 72, + 186, + 517, + 200 + ], + "type": "text", + "content": "urdue.edu," + }, + { + "bbox": [ + 72, + 186, + 517, + 200 + ], + "type": "inline_equation", + "content": "^{2}\\{ \\text{zhang,zlin,scohen,bprice, jianmzha,sooyek} \\} @" + }, + { + "bbox": [ + 72, + 186, + 517, + 200 + ], + "type": "text", + "content": "adobe.com" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 51, + 251, + 132, + 416 + ], + "blocks": [ + { + "bbox": [ + 64, + 232, + 120, + 243 + ], + "lines": [ + { + "bbox": [ + 64, + 232, + 120, + 243 + ], + "spans": [ + { + "bbox": [ + 64, + 232, + 120, + 243 + ], + "type": "text", + "content": "(a) copy-paste" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 51, + 251, + 132, + 416 + ], + "lines": [ + { + "bbox": [ + 51, + 251, + 132, + 416 + ], + "spans": [ + { + "bbox": [ + 51, + 251, + 132, + 416 + ], + "type": "image", + "image_path": "22213a7b75764b52e031010cdd4d202a63f5ed7ff1196f9033550863a307d181.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 421, + 547, + 465 + ], + "lines": [ + { + "bbox": [ + 46, + 421, + 547, + 465 + ], + "spans": [ + { + "bbox": [ + 46, + 421, + 547, + 465 + ], + "type": "text", + "content": "Figure 1. Example results of object compositing with (a) copy-and-paste scheme, (b) traditional compositing pipeline and (c) ours (Object-Stitch). Traditional compositing pipeline is done with best possible off-the-shelf models including foreground/background color harmonization [19,47], poisson blending [34], and shadow synthesis [42]. ObjectStitch achieves more realistic results, and can address geometry correction, harmonization, shadow generation, and view synthesis all-in-one while preserving similar appearance to the reference object." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 133, + 251, + 214, + 332 + ], + "blocks": [ + { + "bbox": [ + 134, + 227, + 212, + 248 + ], + "lines": [ + { + "bbox": [ + 134, + 227, + 212, + 248 + ], + "spans": [ + { + "bbox": [ + 134, + 227, + 212, + 248 + ], + "type": "text", + "content": "(b) traditional compositing pipeline" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 133, + 251, + 214, + 332 + ], + "lines": [ + { + "bbox": [ + 133, + 251, + 214, + 332 + ], + "spans": [ + { + "bbox": [ + 133, + 251, + 214, + 332 + ], + "type": "image", + "image_path": "cfd5085ea9132eb738ebc7a687b25f8531b86b2607301cc275614fd8d0fc339e.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 214, + 251, + 296, + 332 + ], + "blocks": [ + { + "bbox": [ + 240, + 232, + 271, + 243 + ], + "lines": [ + { + "bbox": [ + 240, + 232, + 271, + 243 + ], + "spans": [ + { + "bbox": [ + 240, + 232, + 271, + 243 + ], + "type": "text", + "content": "(c) ours" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 214, + 251, + 296, + 332 + ], + "lines": [ + { + "bbox": [ + 214, + 251, + 296, + 332 + ], + "spans": [ + { + "bbox": [ + 214, + 251, + 296, + 332 + ], + "type": "image", + "image_path": "588dbed269c6232e2cee162781a76218dfafae91d8f5ccc6ae21ed7a2750ced8.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 298, + 251, + 380, + 332 + ], + "blocks": [ + { + "bbox": [ + 312, + 232, + 368, + 243 + ], + "lines": [ + { + "bbox": [ + 312, + 232, + 368, + 243 + ], + "spans": [ + { + "bbox": [ + 312, + 232, + 368, + 243 + ], + "type": "text", + "content": "(a) copy-paste" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 298, + 251, + 380, + 332 + ], + "lines": [ + { + "bbox": [ + 298, + 251, + 380, + 332 + ], + "spans": [ + { + "bbox": [ + 298, + 251, + 380, + 332 + ], + "type": "image", + "image_path": "5b6e3e861229a4d5e838350ba8389cc3452f4f619b8265e23db05db68dae1317.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 380, + 251, + 462, + 332 + ], + "blocks": [ + { + "bbox": [ + 381, + 227, + 460, + 248 + ], + "lines": [ + { + "bbox": [ + 381, + 227, + 460, + 248 + ], + "spans": [ + { + "bbox": [ + 381, + 227, + 460, + 248 + ], + "type": "text", + "content": "(b) traditional compositing pipeline" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 380, + 251, + 462, + 332 + ], + "lines": [ + { + "bbox": [ + 380, + 251, + 462, + 332 + ], + "spans": [ + { + "bbox": [ + 380, + 251, + 462, + 332 + ], + "type": "image", + "image_path": "235043b1565a429c9e4344fbe0911f02383e936001c2df3ddd3c7482d2f6ceec.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 462, + 251, + 543, + 332 + ], + "blocks": [ + { + "bbox": [ + 487, + 232, + 517, + 243 + ], + "lines": [ + { + "bbox": [ + 487, + 232, + 517, + 243 + ], + "spans": [ + { + "bbox": [ + 487, + 232, + 517, + 243 + ], + "type": "text", + "content": "(c) ours" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 462, + 251, + 543, + 332 + ], + "lines": [ + { + "bbox": [ + 462, + 251, + 543, + 332 + ], + "spans": [ + { + "bbox": [ + 462, + 251, + 543, + 332 + ], + "type": "image", + "image_path": "ac24a6ff2324588107bb51f60953fff77c50e285fc565a9ec760772b3f6ef057.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 143, + 487, + 192, + 499 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 487, + 192, + 499 + ], + "spans": [ + { + "bbox": [ + 143, + 487, + 192, + 499 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 45, + 513, + 290, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 513, + 290, + 693 + ], + "spans": [ + { + "bbox": [ + 45, + 513, + 290, + 693 + ], + "type": "text", + "content": "Object compositing based on 2D images is a challenging problem since it typically involves multiple processing stages such as color harmonization, geometry correction and shadow generation to generate realistic results. Furthermore, annotating training data pairs for compositing requires substantial manual effort from professionals, and is hardly scalable. Thus, with the recent advances in generative models, in this work, we propose a self-supervised framework for object compositing by leveraging the power of conditional diffusion models. Our framework can holistically address the object compositing task in a unified model, transforming the viewpoint, geometry, color and shadow of the generated object while requiring no manual labeling. To preserve the input object's characteristics, we introduce a content adaptor that helps to maintain categori" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 477, + 547, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 477, + 547, + 537 + ], + "spans": [ + { + "bbox": [ + 304, + 477, + 547, + 537 + ], + "type": "text", + "content": "cal semantics and object appearance. A data augmentation method is further adopted to improve the fidelity of the generator. Our method outperforms relevant baselines in both realism and faithfulness of the synthesized result images in a user study on various real-world images." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 306, + 561, + 387, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 561, + 387, + 574 + ], + "spans": [ + { + "bbox": [ + 306, + 561, + 387, + 574 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": "Image compositing is an essential task in image editing that aims to insert an object from a given image into another image in a realistic way. Conventionally, many sub-tasks are involved in composing an object to a new scene, including color harmonization [6, 7, 19, 51], relighting [52], and shadow generation [16, 29, 43] in order to naturally blend the object into the new image. As shown in Tab. 1, most previous methods [6, 7, 16, 19, 28, 43] focus on a single sub-task required for image compositing. Consequently, they must be appropriately combined to obtain a composite image where the input object is re-synthesized to have the" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 702, + 196, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 702, + 196, + 713 + ], + "spans": [ + { + "bbox": [ + 61, + 702, + 196, + 713 + ], + "type": "text", + "content": "* Work done during internship at Adobe." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18310" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 276, + 175 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 276, + 175 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 276, + 175 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 276, + 175 + ], + "type": "table", + "html": "
MethodGeometryLightShadowView
ST-GAN [28]✓XXX
SSH [19]X✓XX
DCCF [51]X✓XX
SSN [43]XX✓X
SGRNet [16]XX✓X
GCC-GAN [5]✓✓XX
Ours✓✓✓✓
", + "image_path": "7bc0c3bed694450bbe9b4af35643afcf845469858136727e06fd223be8b25e6d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 178, + 287, + 212 + ], + "lines": [ + { + "bbox": [ + 46, + 178, + 287, + 212 + ], + "spans": [ + { + "bbox": [ + 46, + 178, + 287, + 212 + ], + "type": "text", + "content": "Table 1. Prior works only focus on one or two aspects of object compositing, and they cannot synthesize novel views. In contrast, our model can address all perspectives as listed." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 228, + 286, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 228, + 286, + 277 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 286, + 277 + ], + "type": "text", + "content": "color, lighting and shadow that is consistent with the background scene. As shown in Fig. 1, results produced in this way still look unnatural, partly due to the viewpoint of the inserted object being different from the overall background." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 278, + 286, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 278, + 286, + 434 + ], + "spans": [ + { + "bbox": [ + 46, + 278, + 286, + 434 + ], + "type": "text", + "content": "Harmonizing the geometry and synthesizing novel views have often been overlooked in 2D image compositing, which require an accurate understanding of both the geometry of the object and the background scene from 2D images. Previous works [15, 21, 22] handle 3D object compositing with explicit background information such as lighting positions and depth. More recently, [28] utilize GANs to estimate the homography of the object. However, this method is limited to placing furniture in indoor scenes. In this paper, we propose a generic image compositing method that is able to harmonize the geometry of the input object along with color, lighting and shadow with the background image using a diffusion-based generative model." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 435, + 286, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 435, + 286, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 435, + 286, + 652 + ], + "type": "text", + "content": "In recent years, generative models such as GANs [4, 10, 17, 20] and diffusion models [1, 14, 30, 32, 33, 38, 39] have shown great potential in synthesizing realistic images. In particular, diffusion model-based frameworks are versatile and outperform various prior methods in image editing [1, 23, 32] and other applications [11, 31, 35]. However, most image editing diffusion models focus on using text inputs to manipulate images [1, 3, 9, 23, 33], which is insufficient for image compositing as verbal representations cannot fully capture the details or preserve the identity and appearance of a given object image. There have been recent works [23, 39] focusing on generating diverse contexts while preserving the key features of the object; however, these models are designed for a different task than object compositing. Furthermore, [23] requires fine-tuning the model for each input object and [39] also needs to be finetuned on multiple images of the same object. Therefore they are limited for general object compositing." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 653, + 286, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 286, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 286, + 715 + ], + "type": "text", + "content": "In this work, we leverage diffusion models to simultaneously handle multiple aspects of image compositing such as color harmonization, relighting, geometry correction and shadow generation. With image guidance rather than text guidance, we aim to preserve the identity and appearance of" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 72, + 520, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 520, + 84 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 520, + 84 + ], + "type": "text", + "content": "the original object in the generated composite image." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 85, + 545, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 85, + 545, + 276 + ], + "spans": [ + { + "bbox": [ + 304, + 85, + 545, + 276 + ], + "type": "text", + "content": "Specifically, our model synthesizes a composite image given (i) a source object, (ii) a target background image, and (iii) a bounding box specifying the location to insert the object. The proposed framework consists of a content adaptor and a generator module: the content adaptor is designed to extract a representation from the input object containing both high-level semantics and low-level details such as color and shape; the generator module preserves the background scene while improving the generation quality and versatility. Our framework is trained in a fully self-supervised manner and no task-specific labeling is required at any point during training. Moreover, various data augmentation techniques are applied to further improve the fidelity and realism of the output. We evaluate our proposed method on a real-world dataset closely simulating real use cases for image compositing." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 317, + 276, + 503, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 276, + 503, + 287 + ], + "spans": [ + { + "bbox": [ + 317, + 276, + 503, + 287 + ], + "type": "text", + "content": "Our contributions are summarized as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 317, + 291, + 545, + 459 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 317, + 291, + 545, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 291, + 545, + 339 + ], + "spans": [ + { + "bbox": [ + 317, + 291, + 545, + 339 + ], + "type": "text", + "content": "- We present the first diffusion model-based framework for generative object compositing that can handle multiple aspects of compositing such as viewpoint, geometry, lighting and shadow." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 317, + 339, + 545, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 339, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 317, + 339, + 545, + 374 + ], + "type": "text", + "content": "- We propose a content adaptor module which learns a descriptive multi-modal embedding from images, enabling image guidance for diffusion models." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 317, + 375, + 545, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 375, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 317, + 375, + 545, + 422 + ], + "type": "text", + "content": "- Our framework is trained in a self-supervised manner without any task-specific annotations, employing data augmentation techniques to improve the fidelity of generation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 317, + 423, + 545, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 423, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 317, + 423, + 545, + 459 + ], + "type": "text", + "content": "- We collect a high-resolution real-world dataset for object compositing with diverse images, containing manually annotated object scales and locations." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 471, + 392, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 471, + 392, + 483 + ], + "spans": [ + { + "bbox": [ + 306, + 471, + 392, + 483 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 491, + 420, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 491, + 420, + 504 + ], + "spans": [ + { + "bbox": [ + 306, + 491, + 420, + 504 + ], + "type": "text", + "content": "2.1. Image Compositing" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 510, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 545, + 617 + ], + "type": "text", + "content": "Image compositing is a challenging task in reference-guided image editing, where the object in a given foreground image is to be inserted into a background image. The generated composite image is expected to look realistic, with the appearance of the original object being preserved. Prior works often focus on an individual aspect of this problem, such as geometric correction [2, 28], image harmonization [6, 7, 51], image matting [50] and shadow generation [29]." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "Lin et al. [28] propose an iterative GAN-based framework to correct geometric inconsistencies between object and background images. In their model, STNs [18] are integrated into the generator network to predict a series of warp updates. Although their work can improve the realism of the scene geometry in the composite image, it is limited to inserting furniture into indoor scenes. Azadi et al. [2] focus on binary composition, using a composition-decomposition" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18311" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 278, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 278, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 278, + 83 + ], + "type": "text", + "content": "network to capture interactions between a pair of objects." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 85, + 286, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 85, + 286, + 191 + ], + "spans": [ + { + "bbox": [ + 46, + 85, + 286, + 191 + ], + "type": "text", + "content": "Image harmonization aims at minimizing the inconsistency between the input object and the background image. Traditional methods [24, 46] usually concentrate on obtaining color statistics and then transfer this information between foreground and background. Recent works seek to solve this problem via deep neural networks. [6, 7] reformulate image harmonization as a domain adaptation problem, while [19] converts it to a representation fusing problem and utilizes self-supervised training." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 193, + 286, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 193, + 286, + 276 + ], + "spans": [ + { + "bbox": [ + 46, + 193, + 286, + 276 + ], + "type": "text", + "content": "Shadow synthesis is an effect that is often overlooked in previous image compositing methods although it is essential for generating realistic composite images. SGRNet [16] divides this task into a shadow mask generation stage and a shadow filling stage. SSN [43], focusing on soft shadows, predicts an ambient occlusion map as cue for shadow generation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 277, + 286, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 277, + 286, + 408 + ], + "spans": [ + { + "bbox": [ + 46, + 277, + 286, + 408 + ], + "type": "text", + "content": "There are also works simultaneously addressing multiple sub-problems of image compositing. Chen et al. [5] develop a system including multiple network modules to produce plausible images with geometric, color and boundary consistency. However, the pose of the generated object is constrained by the mask input and the model cannot generalize to non-rigid objects such as animals. Our proposed model also handles the image composition problem in a unified manner, generating foreground objects which are harmonious and geometrically consistent with the background while synthesizing novel views and shadows." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 418, + 183, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 418, + 183, + 431 + ], + "spans": [ + { + "bbox": [ + 47, + 418, + 183, + 431 + ], + "type": "text", + "content": "2.2. Guided Image Synthesis" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 437, + 286, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 286, + 557 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 286, + 557 + ], + "type": "text", + "content": "Recent years have witnessed great advances in diffusion models. Diffusion models are a family of deep generative models based on several predominant works [14, 44, 45], defined with two Markov chains. In the forward process, they gradually add noise to the data, and in the reverse process, they learn to recover the data from noise. Following the growth of research in this area, diffusion models have shown their potential in a great number of applications, such as text-to-image generation [9, 33, 38], image editing [1, 23, 32] and guided image synthesis [30]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 558, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 286, + 713 + ], + "type": "text", + "content": "Text-to-image synthesis has been a popular research topic over the past few years. Stable Diffusion [38] made a significant contribution to this task and it reduces the computational cost by applying the diffusion model in the latent space. Other works explore ways of more flexible and controllable text-driven image editing. Avrahami et al. [1] design an algorithm in their Blended Diffusion framework to fuse noised versions of the input with the local text-guided diffusion latent, generating smooth transition. Similarly, GLIDE [33] is also capable of handling text-guided inpainting after fine-tuning on this specific task. Some methods give users more straightforward control on images. SDEdit [32] allows stroke painting in images, blending user input" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 119 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 119 + ], + "type": "text", + "content": "into the image by adding noise to the input and denoising through a stochastic differential equation. SDG [30] injects Semantic Diffusion Guidance at each iteration of the generation and can provide multi-modal guidance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 120, + 545, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 120, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 304, + 120, + 545, + 335 + ], + "type": "text", + "content": "While the versatility and generation quality of diffusion models have been repeatedly demonstrated, preserving object appearance remains a very challenging problem in image editing. Ruiz et al. [39] address this problem by optimizing their model using a reconstruction loss and a class-specific prior preservation loss. Though their approach can preserve the details of the object, multiple images of the same object are required to fine-tune their model on this target object. Kawar et al. [23] handle this issue by interpolating the initial text embedding and the optimized embedding for reconstruction, but their method is mainly designed for applying edits on objects, not for generating the same object in different contexts. In this paper, we propose the first generative object compositing framework based on diffusion models, which generates the harmonized and geometrically-corrected subject with a novel view and shadow. Furthermore, the characteristics of the original object is preserved in the synthesized composite image." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 346, + 411, + 360 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 346, + 411, + 360 + ], + "spans": [ + { + "bbox": [ + 306, + 346, + 411, + 360 + ], + "type": "text", + "content": "3. Proposed Method" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 366, + 545, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 366, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 304, + 366, + 545, + 557 + ], + "type": "text", + "content": "We define the generative object compositing problem as: Given an input triplet " + }, + { + "bbox": [ + 304, + 366, + 545, + 557 + ], + "type": "inline_equation", + "content": "(I_o, I_{bg}, M)" + }, + { + "bbox": [ + 304, + 366, + 545, + 557 + ], + "type": "text", + "content": " that consists of an object image " + }, + { + "bbox": [ + 304, + 366, + 545, + 557 + ], + "type": "inline_equation", + "content": "I_o \\in \\mathbb{R}^{H_s \\times W_s \\times 3}" + }, + { + "bbox": [ + 304, + 366, + 545, + 557 + ], + "type": "text", + "content": ", a background image " + }, + { + "bbox": [ + 304, + 366, + 545, + 557 + ], + "type": "inline_equation", + "content": "I_{bg} \\in \\mathbb{R}^{H_t \\times W_t \\times 3}" + }, + { + "bbox": [ + 304, + 366, + 545, + 557 + ], + "type": "text", + "content": ", and its associated binary mask " + }, + { + "bbox": [ + 304, + 366, + 545, + 557 + ], + "type": "inline_equation", + "content": "M \\in \\mathbb{R}^{H_t \\times W_t \\times 1}" + }, + { + "bbox": [ + 304, + 366, + 545, + 557 + ], + "type": "text", + "content": " with the desired object location set to 0 and the rest to 1, the goal is to compose the input object into the masked area " + }, + { + "bbox": [ + 304, + 366, + 545, + 557 + ], + "type": "inline_equation", + "content": "I_{bg} \\otimes M" + }, + { + "bbox": [ + 304, + 366, + 545, + 557 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 366, + 545, + 557 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 366, + 545, + 557 + ], + "type": "text", + "content": " is considered as a soft constraint of the location and scale of the composited object. The output generated image should look realistic, while the appearance of the object is preserved. Our problem setting is different from text-guided image generation and inpainting in that the condition input is a reference object image rather than a text prompt. Inspired by the success of text-guided diffusion models [37, 38, 40], which inject text conditioning into the diffusion architecture, we design our method for generative compositing to leverage such pretrained models." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 713 + ], + "type": "text", + "content": "An overview of our framework is shown in Fig. 2. It consists of an object image encoder extracting semantic features from the object and a conditional diffusion generator. To leverage the power of pretrained text-to-image diffusion models, we introduce a content adaptor that can bridge the gap between the object encoder and conditional generator by transforming a sequence of visual tokens to a sequence of text tokens to overcome the domain gap between image and text. This design further helps to preserve the object appearance. We propose a two-stage training process: in the first stage, the content adaptor is trained on large image/text pairs to maintain high-level semantics of the object; in the second stage, it is trained in the context of diffusion" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18312" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 70, + 531, + 191 + ], + "blocks": [ + { + "bbox": [ + 60, + 70, + 531, + 191 + ], + "lines": [ + { + "bbox": [ + 60, + 70, + 531, + 191 + ], + "spans": [ + { + "bbox": [ + 60, + 70, + 531, + 191 + ], + "type": "image", + "image_path": "07cd73ec57e3862f79bb3a725a0012186561a38f2767aac1c67b8d92bf5e52a4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 193, + 547, + 237 + ], + "lines": [ + { + "bbox": [ + 46, + 193, + 547, + 237 + ], + "spans": [ + { + "bbox": [ + 46, + 193, + 547, + 237 + ], + "type": "text", + "content": "Figure 2. System pipeline. Our framework consists of a content adaptor and a generator (a pretrained text-to-image diffusion model). The input image " + }, + { + "bbox": [ + 46, + 193, + 547, + 237 + ], + "type": "inline_equation", + "content": "I_{o}" + }, + { + "bbox": [ + 46, + 193, + 547, + 237 + ], + "type": "text", + "content": " is fed into a ViT and the adaptor which produces a descriptive embedding. At the same time the background image " + }, + { + "bbox": [ + 46, + 193, + 547, + 237 + ], + "type": "inline_equation", + "content": "I_{bg}" + }, + { + "bbox": [ + 46, + 193, + 547, + 237 + ], + "type": "text", + "content": " is taken as input by the diffusion model. At each iteration during the denoising stage, we apply the mask " + }, + { + "bbox": [ + 46, + 193, + 547, + 237 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 193, + 547, + 237 + ], + "type": "text", + "content": " on the generated image " + }, + { + "bbox": [ + 46, + 193, + 547, + 237 + ], + "type": "inline_equation", + "content": "I_{out}" + }, + { + "bbox": [ + 46, + 193, + 547, + 237 + ], + "type": "text", + "content": ", so that the generator only denoises the masked area " + }, + { + "bbox": [ + 46, + 193, + 547, + 237 + ], + "type": "inline_equation", + "content": "I_{out} \\otimes M" + }, + { + "bbox": [ + 46, + 193, + 547, + 237 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 248, + 290, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 248, + 290, + 333 + ], + "spans": [ + { + "bbox": [ + 46, + 248, + 290, + 333 + ], + "type": "text", + "content": "generator to encode key identity features of the object by encouraging the visual reconstruction of the object in the original image. Lastly, the generator module is fine-tuned on the embedding produced by the adaptor through cross attention blocks. All stages our trained in a self-supervised manner to avoid expensive annotation for obtaining object compositing training data." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 350, + 119, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 350, + 119, + 361 + ], + "spans": [ + { + "bbox": [ + 47, + 350, + 119, + 361 + ], + "type": "text", + "content": "3.1. Generator" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 371, + 288, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 371, + 288, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 371, + 288, + 456 + ], + "type": "text", + "content": "As depicted in Fig. 2, we leverage a pretrained text-to-image diffusion model architecture and modify it for the compositing task by (i) introducing a mask in the input, and (ii) adjusting the U-Net input to contain the original background image outside the hole and noise inside the hole with mask blending. In order to condition the model on the guidance embedding " + }, + { + "bbox": [ + 46, + 371, + 288, + 456 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 46, + 371, + 288, + 456 + ], + "type": "text", + "content": ", an attention mechanism is applied as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 479, + 287, + 508 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 479, + 287, + 508 + ], + "spans": [ + { + "bbox": [ + 62, + 479, + 287, + 508 + ], + "type": "interline_equation", + "content": "\\operatorname {S o f t m a x} \\left(\\frac {\\left(\\boldsymbol {W} _ {Q} E _ {x}\\right) \\left(\\boldsymbol {W} _ {K} E\\right) ^ {T}}{\\sqrt {d}}\\right) \\boldsymbol {W} _ {V} E = \\boldsymbol {A} \\boldsymbol {V} \\tag {1}", + "image_path": "0c9234aab5b3f35e198e16740fcfaed448bd0ae11ef80795a9987c66f7b8d1ab.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 521, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 289, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 521, + 289, + 714 + ], + "type": "inline_equation", + "content": "E_{x}" + }, + { + "bbox": [ + 46, + 521, + 289, + 714 + ], + "type": "text", + "content": " is an intermediate representation of the denoising autoencoder and " + }, + { + "bbox": [ + 46, + 521, + 289, + 714 + ], + "type": "inline_equation", + "content": "W_{Q} \\in \\mathbb{R}^{d \\times d_{x}}" + }, + { + "bbox": [ + 46, + 521, + 289, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 521, + 289, + 714 + ], + "type": "inline_equation", + "content": "W_{K} \\in \\mathbb{R}^{d \\times d_{e}}" + }, + { + "bbox": [ + 46, + 521, + 289, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 521, + 289, + 714 + ], + "type": "inline_equation", + "content": "W_{V} \\in \\mathbb{R}^{d \\times d_{e}}" + }, + { + "bbox": [ + 46, + 521, + 289, + 714 + ], + "type": "text", + "content": " are embedding matrices. The background outside the mask area should be perfectly preserved for generative object compositing. Thus, we use the input mask " + }, + { + "bbox": [ + 46, + 521, + 289, + 714 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 46, + 521, + 289, + 714 + ], + "type": "text", + "content": " for blending the background image " + }, + { + "bbox": [ + 46, + 521, + 289, + 714 + ], + "type": "inline_equation", + "content": "I_{bg}" + }, + { + "bbox": [ + 46, + 521, + 289, + 714 + ], + "type": "text", + "content": " with the generated image " + }, + { + "bbox": [ + 46, + 521, + 289, + 714 + ], + "type": "inline_equation", + "content": "I_{out}" + }, + { + "bbox": [ + 46, + 521, + 289, + 714 + ], + "type": "text", + "content": ". As a result, the generator only denoises the masked area " + }, + { + "bbox": [ + 46, + 521, + 289, + 714 + ], + "type": "inline_equation", + "content": "I_{out} \\otimes M" + }, + { + "bbox": [ + 46, + 521, + 289, + 714 + ], + "type": "text", + "content": ". To use this model on our task, a straightforward way (a baseline) would be to apply image captioning on the object image and feeding the resulting caption directly to the diffusion model as the condition. However, text embedding cannot capture fine grained visual details. Therefore, in order to directly leverage the given object image, we introduce the content adaptor to transform the visual features from a pretrained visual encoder to text features (tokens) to use as conditioning for the generator." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 313, + 246, + 537, + 411 + ], + "blocks": [ + { + "bbox": [ + 313, + 246, + 537, + 411 + ], + "lines": [ + { + "bbox": [ + 313, + 246, + 537, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 246, + 537, + 411 + ], + "type": "image", + "image_path": "af112cf73253444a877283eaf3708010be628359e7496df8e70d58a36a5dbea1.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 412, + 547, + 468 + ], + "lines": [ + { + "bbox": [ + 304, + 412, + 547, + 468 + ], + "spans": [ + { + "bbox": [ + 304, + 412, + 547, + 468 + ], + "type": "text", + "content": "Figure 3. Structure of the Content Adaptor. In the first stage, it is trained on a large dataset of image-caption pairs to learn multimodal sequential embeddings containing high-level semantics. In the second stage, it is fine-tuned under the diffusion framework to learn to encode identity features in adaptive embedding." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 480, + 409, + 494 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 480, + 409, + 494 + ], + "spans": [ + { + "bbox": [ + 306, + 480, + 409, + 494 + ], + "type": "text", + "content": "3.2. Content Adaptor" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 500, + 547, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 500, + 547, + 559 + ], + "spans": [ + { + "bbox": [ + 304, + 500, + 547, + 559 + ], + "type": "text", + "content": "To prevent the loss of key identity information, we use an image encoder instead of a text encoder to produce the embedding from the input object image. However, the image embedding cannot be effectively utilized by the diffusion model for two reasons:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 317, + 563, + 545, + 635 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 317, + 563, + 545, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 563, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 317, + 563, + 545, + 612 + ], + "type": "text", + "content": "- The image embedding " + }, + { + "bbox": [ + 317, + 563, + 545, + 612 + ], + "type": "inline_equation", + "content": "\\widetilde{E}" + }, + { + "bbox": [ + 317, + 563, + 545, + 612 + ], + "type": "text", + "content": " and the text embedding " + }, + { + "bbox": [ + 317, + 563, + 545, + 612 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 317, + 563, + 545, + 612 + ], + "type": "text", + "content": " are from different domains. Since the diffusion model was trained on " + }, + { + "bbox": [ + 317, + 563, + 545, + 612 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 317, + 563, + 545, + 612 + ], + "type": "text", + "content": ", it cannot generate meaningful contents from image embedding sequence;" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 317, + 611, + 545, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 611, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 317, + 611, + 545, + 635 + ], + "type": "text", + "content": "- A mismatch in the dimensions of " + }, + { + "bbox": [ + 317, + 611, + 545, + 635 + ], + "type": "inline_equation", + "content": "\\widetilde{E} \\in \\mathbb{R}^{k \\times 257 \\times 1024}" + }, + { + "bbox": [ + 317, + 611, + 545, + 635 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 317, + 611, + 545, + 635 + ], + "type": "inline_equation", + "content": "E \\in \\mathbb{R}^{k \\times 77 \\times 768}" + }, + { + "bbox": [ + 317, + 611, + 545, + 635 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 317, + 611, + 545, + 635 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 317, + 611, + 545, + 635 + ], + "type": "text", + "content": " is batch-size." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 304, + 640, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 640, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 640, + 547, + 715 + ], + "type": "text", + "content": "Therefore, based on the above observations, we develop a sequence-to-sequence translator architecture as shown in Fig. 3. Given an image-caption pair as an input tuple " + }, + { + "bbox": [ + 304, + 640, + 547, + 715 + ], + "type": "inline_equation", + "content": "(I,t)" + }, + { + "bbox": [ + 304, + 640, + 547, + 715 + ], + "type": "text", + "content": ", we employ two pretrained ViT-L/14 encoders " + }, + { + "bbox": [ + 304, + 640, + 547, + 715 + ], + "type": "inline_equation", + "content": "C_t" + }, + { + "bbox": [ + 304, + 640, + 547, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 640, + 547, + 715 + ], + "type": "inline_equation", + "content": "C_i" + }, + { + "bbox": [ + 304, + 640, + 547, + 715 + ], + "type": "text", + "content": " from CLIP [36] to produce text embedding " + }, + { + "bbox": [ + 304, + 640, + 547, + 715 + ], + "type": "inline_equation", + "content": "E = C_t(t)" + }, + { + "bbox": [ + 304, + 640, + 547, + 715 + ], + "type": "text", + "content": " and image embedding " + }, + { + "bbox": [ + 304, + 640, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\widetilde{E} = C_i(I)" + }, + { + "bbox": [ + 304, + 640, + 547, + 715 + ], + "type": "text", + "content": ", respectively. The adaptor " + }, + { + "bbox": [ + 304, + 640, + 547, + 715 + ], + "type": "inline_equation", + "content": "T" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18313" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 70, + 269, + 209 + ], + "blocks": [ + { + "bbox": [ + 65, + 70, + 269, + 209 + ], + "lines": [ + { + "bbox": [ + 65, + 70, + 269, + 209 + ], + "spans": [ + { + "bbox": [ + 65, + 70, + 269, + 209 + ], + "type": "image", + "image_path": "2c40277be9bbef09e48d76f757a40fa7941352eabd768d0de4773d2a5d6c5f0d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 210, + 289, + 288 + ], + "lines": [ + { + "bbox": [ + 46, + 210, + 289, + 288 + ], + "spans": [ + { + "bbox": [ + 46, + 210, + 289, + 288 + ], + "type": "text", + "content": "Figure 4. Illustration of our synthetic data generation and data augmentation scheme. The top row shows the data generation process including perspective warping, random rotation, and random color shifting. The original image is used as both the input background and ground truth, while the perturbed object is fed into the adaptor. The bottom row shows crop and shift augmentations, which help to improve the generation quality and preserve object details." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 298, + 288, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 298, + 288, + 382 + ], + "spans": [ + { + "bbox": [ + 46, + 298, + 288, + 382 + ], + "type": "text", + "content": "consists of three components: a 1D convolutional layer, attention blocks [48] and an MLP, where the 1D convolution modifies the length of the embedding from 257 to 77, the MLP maps the embedding dimension from 1024 to 768, and the attention blocks bridge the gap between text domain and image domain. We design a two-stage optimization method to train this module, which is explained in Sec. 3.3.2." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 390, + 200, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 390, + 200, + 403 + ], + "spans": [ + { + "bbox": [ + 47, + 390, + 200, + 403 + ], + "type": "text", + "content": "3.3. Self-supervised Framework" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 408, + 287, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 408, + 287, + 504 + ], + "spans": [ + { + "bbox": [ + 46, + 408, + 287, + 504 + ], + "type": "text", + "content": "There is no publicly available image compositing training dataset with annotations which is sufficient for training a diffusion model, and it is extremely challenging to manually annotate such data. Therefore, we propose a self-supervised training scheme and a synthetic data generation approach that simulate real-world scenarios. We also introduce a data augmentation method to enrich the training data as well as improve the robustness of our model." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 515, + 234, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 515, + 234, + 526 + ], + "spans": [ + { + "bbox": [ + 47, + 515, + 234, + 526 + ], + "type": "text", + "content": "3.3.1 Data Generation and Augmentation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 533, + 287, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 287, + 629 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 287, + 629 + ], + "type": "text", + "content": "Training data generation. We collect our synthetic training data from Pixabay and use an object instance segmentation model [26] to predict panoptic segmentation and classification labels. We first filter the dataset by removing objects with very small or large sizes. Then, we apply spatial and color perturbations to simulate many real use-case scenarios where the input image and background image have different scene geometry and lighting conditions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 630, + 287, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 287, + 700 + ], + "type": "text", + "content": "The top row of Fig. 4 illustrates this process. Inspired by [8], we randomly perturb the four points of the object bounding box to apply projective transformation, followed by a random rotation within the range " + }, + { + "bbox": [ + 46, + 630, + 287, + 700 + ], + "type": "inline_equation", + "content": "[- \\theta, \\theta]" + }, + { + "bbox": [ + 46, + 630, + 287, + 700 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 46, + 630, + 287, + 700 + ], + "type": "inline_equation", + "content": "\\theta = 20^{\\circ}" + }, + { + "bbox": [ + 46, + 630, + 287, + 700 + ], + "type": "text", + "content": ") and color perturbation. The segmentation mask (perturbed in the same way as the image) is used to extract the object." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 287, + 713 + ], + "type": "text", + "content": "This data synthesis pipeline is fully controllable and can" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": "also be employed as data augmentation during training. Another key advantage is that it is free of manual labeling, since the original image is used as the ground truth. We use the bounding box as the mask as it not only fully covers the object, but also extends to its neighboring area (providing room for shadow generation). We find that it is flexible enough for the model to apply spatial transformations, synthesize novel views and generate shadows and reflection." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 171, + 546, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 171, + 546, + 351 + ], + "spans": [ + { + "bbox": [ + 304, + 171, + 546, + 351 + ], + "type": "text", + "content": "Real-world evaluation data. Given that there is no existing dataset specifically for our task, we manually collect a new dataset, closely simulating real-world use cases, as an evaluation benchmark for object compositing. The dataset consists of 503 pairs of common objects (including both rigid and non-rigid objects such as vehicles and wildlife) and diverse background images (covering both indoor and outdoor scenes). It also contains challenging cases where there is a large mismatch of lighting conditions or viewpoints between foreground and background. The dataset images are collected from Pixabay. The labeling procedure closely simulate the real scenarios where the input object is placed at a target location in the background image and then scaled at the user's will. The compositing region is determined as a loose bounding box around the object." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 354, + 547, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 354, + 547, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 547, + 449 + ], + "type": "text", + "content": "Data augmentation. Inspired by [25], we introduce random shift and crop augmentations during training, while ensuring that the foreground object is always contained in the crop window. This process is illustrated in the bottom row of Fig. 4. Applying this augmentation method for both training and inference results in a notable improvement in the realism of the generated results. Quantitative results are provided in Sec. 4.4." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 459, + 376, + 471 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 459, + 376, + 471 + ], + "spans": [ + { + "bbox": [ + 306, + 459, + 376, + 471 + ], + "type": "text", + "content": "3.3.2 Training" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 478, + 545, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 478, + 545, + 573 + ], + "spans": [ + { + "bbox": [ + 304, + 478, + 545, + 573 + ], + "type": "text", + "content": "Content adaptor pretraining. We first pretrain the content adaptor to keep the semantics of the object by mapping the image embedding to text embedding. At this first stage, we optimize the content adaptor on a sequence-to-sequence translation task, which learns to project the image embedding into a multi-modal space. During training, " + }, + { + "bbox": [ + 304, + 478, + 545, + 573 + ], + "type": "inline_equation", + "content": "C_t" + }, + { + "bbox": [ + 304, + 478, + 545, + 573 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 478, + 545, + 573 + ], + "type": "inline_equation", + "content": "C_i" + }, + { + "bbox": [ + 304, + 478, + 545, + 573 + ], + "type": "text", + "content": " are frozen, and the translator is trained on 3,203,338 image-caption pairs from a filtered LAION dataset [41]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 574, + 545, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 574, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 304, + 574, + 545, + 609 + ], + "type": "text", + "content": "Given the input image embedding " + }, + { + "bbox": [ + 304, + 574, + 545, + 609 + ], + "type": "inline_equation", + "content": "\\vec{E}" + }, + { + "bbox": [ + 304, + 574, + 545, + 609 + ], + "type": "text", + "content": ", we use the text embedding " + }, + { + "bbox": [ + 304, + 574, + 545, + 609 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 304, + 574, + 545, + 609 + ], + "type": "text", + "content": " as target, the objective function of this translation task is defined as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 376, + 618, + 545, + 632 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 618, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 376, + 618, + 545, + 632 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {d i s t}} = \\| T (\\tilde {E}) - E \\| _ {1}, \\tag {2}", + "image_path": "e1820f105e77f1d907bd9040a5ed5de4364642f9ee6e2d51ddd275b3b7abd7b1.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 642, + 443, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 642, + 443, + 653 + ], + "spans": [ + { + "bbox": [ + 306, + 642, + 443, + 653 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 642, + 443, + 653 + ], + "type": "inline_equation", + "content": "T(\\cdot)" + }, + { + "bbox": [ + 306, + 642, + 443, + 653 + ], + "type": "text", + "content": " is the content adaptor." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": "However, the multi-modal embedding obtained solely from this first stage optimization mostly carry high-level semantics of the input image without much of texture details. Hence, we further refine " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\widetilde{E}" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": " to obtain " + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\hat{E}" + }, + { + "bbox": [ + 304, + 654, + 545, + 713 + ], + "type": "text", + "content": ", the adaptive embedding, for better appearance preservation." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18314" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 156 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 156 + ], + "type": "text", + "content": "Content adaptor fine-tuning. We further optimize the content adaptor to produce an adaptive embedding which maintains instance-level properties of the object. After pretraining, we insert the content adaptor to the pretrained diffusion framework by feeding the adaptive embedding as context to the attention blocks. Then, the diffusion model is frozen and the adaptor is trained using:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 163, + 287, + 179 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 163, + 287, + 179 + ], + "spans": [ + { + "bbox": [ + 55, + 163, + 287, + 179 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {a d a p t}} = \\mathbb {E} _ {T, \\epsilon \\sim \\mathcal {N} (0, 1)} [ \\| \\epsilon - \\epsilon_ {\\theta} (I _ {t} \\circ M, t, T (\\widetilde {E})) \\| _ {2} ^ {2} ], \\tag {3}", + "image_path": "2fd89b6bfd75ba30d05b6e738967afbdbcab1956c2a6c85fd9ce75e984fa5f7a.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 186, + 287, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 186, + 287, + 235 + ], + "spans": [ + { + "bbox": [ + 47, + 186, + 287, + 235 + ], + "type": "text", + "content": "where the content adaptor " + }, + { + "bbox": [ + 47, + 186, + 287, + 235 + ], + "type": "inline_equation", + "content": "T(\\cdot)" + }, + { + "bbox": [ + 47, + 186, + 287, + 235 + ], + "type": "text", + "content": " is optimized. It is trained on our synthetic dataset filtered from Pixabay, containing 467,280 foreground and background image pairs for training and 25,960 pairs for validation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 238, + 287, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 238, + 287, + 321 + ], + "spans": [ + { + "bbox": [ + 46, + 238, + 287, + 321 + ], + "type": "text", + "content": "Generator fine-tuning. After the aforementioned two-stage training of the content adaptor is completed, we freeze the content adaptor and train the generator module after initializing the text-to-image diffusion model with pretrained weights. Crop and shift augmentations are applied in this process. Based on the Latent Diffusion model [38], the objective loss in our generator module is defined by:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 71, + 329, + 287, + 345 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 329, + 287, + 345 + ], + "spans": [ + { + "bbox": [ + 71, + 329, + 287, + 345 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {g e n} = \\mathbb {E} _ {\\hat {E}, \\epsilon \\sim \\mathcal {N} (0, 1)} [ \\| \\epsilon - \\epsilon_ {\\theta} (I _ {t} \\circ M, t, \\hat {E}) \\| _ {2} ^ {2} ], \\tag {4}", + "image_path": "44f64ccbe591afc7695935cef3e3d9fe843854fcca983d49ec4c89606dfc7ad7.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 352, + 287, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 352, + 287, + 413 + ], + "spans": [ + { + "bbox": [ + 47, + 352, + 287, + 413 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 352, + 287, + 413 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 47, + 352, + 287, + 413 + ], + "type": "text", + "content": " is the input image, " + }, + { + "bbox": [ + 47, + 352, + 287, + 413 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 47, + 352, + 287, + 413 + ], + "type": "text", + "content": " is a noisy version of " + }, + { + "bbox": [ + 47, + 352, + 287, + 413 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 47, + 352, + 287, + 413 + ], + "type": "text", + "content": " at timestep " + }, + { + "bbox": [ + 47, + 352, + 287, + 413 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 352, + 287, + 413 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 352, + 287, + 413 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 47, + 352, + 287, + 413 + ], + "type": "text", + "content": " denotes the denoising model which is optimized. In order to adapt the text-guided generation task to the setting of image compositing, we apply the input mask on the image at every time-step." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 422, + 128, + 435 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 422, + 128, + 435 + ], + "spans": [ + { + "bbox": [ + 47, + 422, + 128, + 435 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 441, + 146, + 454 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 441, + 146, + 454 + ], + "spans": [ + { + "bbox": [ + 47, + 441, + 146, + 454 + ], + "type": "text", + "content": "4.1. Training Details" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 460, + 287, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 460, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 460, + 287, + 616 + ], + "type": "text", + "content": "The first training stage of the content adaptor takes 15 epochs with a learning rate of " + }, + { + "bbox": [ + 46, + 460, + 287, + 616 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 46, + 460, + 287, + 616 + ], + "type": "text", + "content": " and batch size of 2048. The image size processed by ViT is " + }, + { + "bbox": [ + 46, + 460, + 287, + 616 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 46, + 460, + 287, + 616 + ], + "type": "text", + "content": ". To set up the pipeline, we employed the pretrained image encoder and text encoder from CLIP [36]. The second training stage of the content adaptor takes 13 epochs with the learning rate of " + }, + { + "bbox": [ + 46, + 460, + 287, + 616 + ], + "type": "inline_equation", + "content": "2 \\times 10^{-5}" + }, + { + "bbox": [ + 46, + 460, + 287, + 616 + ], + "type": "text", + "content": " and batch size of 512. We resize the images to 512 " + }, + { + "bbox": [ + 46, + 460, + 287, + 616 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 46, + 460, + 287, + 616 + ], + "type": "text", + "content": " 512 and adapt the pretrained Stable Diffusion model [38] to our task. After the above process, the generator is trained for 20 epochs with the learning rate of " + }, + { + "bbox": [ + 46, + 460, + 287, + 616 + ], + "type": "inline_equation", + "content": "4 \\times 10^{-5}" + }, + { + "bbox": [ + 46, + 460, + 287, + 616 + ], + "type": "text", + "content": " and batch size of 576. The input image size is the same as that in the second training stage. All training stages are conducted on 8 A100 GPUs with Adam optimizer." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 623, + 183, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 623, + 183, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 623, + 183, + 635 + ], + "type": "text", + "content": "4.2. Quantitative Evaluation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "content": "To obtain quantitative results on a real dataset that simulates real-world use cases, where users drag and drop an object to a background image, we conduct a user study to compare to three baselines, i.e., copy-and-paste, pretrained stable diffusion model fine-tuned on BLIP [27], and a joint stable diffusion model integrating both BLIP" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 316, + 89, + 538, + 148 + ], + "blocks": [ + { + "bbox": [ + 331, + 71, + 503, + 87 + ], + "lines": [ + { + "bbox": [ + 331, + 71, + 503, + 87 + ], + "spans": [ + { + "bbox": [ + 331, + 71, + 503, + 87 + ], + "type": "text", + "content": "Pick from image A and image B the one that you think looks more realistic." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 316, + 89, + 538, + 148 + ], + "lines": [ + { + "bbox": [ + 316, + 89, + 538, + 148 + ], + "spans": [ + { + "bbox": [ + 316, + 89, + 538, + 148 + ], + "type": "image", + "image_path": "7762c2ac838214911580c6ca9dcb9182697c5f410c8a7d739d199a38570455ec.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 317, + 175, + 526, + 211 + ], + "blocks": [ + { + "bbox": [ + 333, + 156, + 501, + 173 + ], + "lines": [ + { + "bbox": [ + 333, + 156, + 501, + 173 + ], + "spans": [ + { + "bbox": [ + 333, + 156, + 501, + 173 + ], + "type": "text", + "content": "Which generated object do you think is most likely to be the same one from the guidance?" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 317, + 175, + 526, + 211 + ], + "lines": [ + { + "bbox": [ + 317, + 175, + 526, + 211 + ], + "spans": [ + { + "bbox": [ + 317, + 175, + 526, + 211 + ], + "type": "image", + "image_path": "d8c93f8d424c09f5d59ffa0015cf45ac8908a35a86fbd674ee179dacb39c85e1.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 212, + 545, + 255 + ], + "lines": [ + { + "bbox": [ + 305, + 212, + 545, + 255 + ], + "spans": [ + { + "bbox": [ + 305, + 212, + 545, + 255 + ], + "type": "text", + "content": "Figure 5. User study results. We conduct side-by-side comparisons between our method and one of baseline methods to quantify the generation quality in terms of realism and appearance preservation. The results show that our method outperforms baselines." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 263, + 545, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 263, + 545, + 431 + ], + "spans": [ + { + "bbox": [ + 304, + 263, + 545, + 431 + ], + "type": "text", + "content": "and SDEdit [32]. Given the fact that there is no existing diffusion-based method addressing the same problem, we train the baselines on the same pretrained diffusion model using the same self-supervised training scheme and dataset as ours. To the best of our knowledge, after adapting to our setting, they are the closest baselines to our task. In the BLIP baseline, the content adaptor is replaced by BLIP, a state-of-the-art captioning model. The text embedding obtained from the BLIP caption is fed into the generator module. For the SDEdit baseline, we keep the BLIP model which will significantly improve the generation fidelity of a pretrained diffusion model. We set the noise strength to 0.65 in SDEdit, striking a balance between realism and faithfulness. Details are provided in the supplementary material." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 432, + 546, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 432, + 546, + 659 + ], + "spans": [ + { + "bbox": [ + 304, + 432, + 546, + 659 + ], + "type": "text", + "content": "We perform a user study using our real testing dataset, which consists of 503 object-background pairs. For each example, we set up a side-by-side comparison between the results from one of the baselines and our method, respectively. The subject will be asked two questions: 1) which is more realistic and 2) which generated object is more close to the given object? We collected 1,494 votes from more than 170 users. As shown in Fig. 5, our method outperforms the other baselines for both questions. The higher preference rates demonstrate the effectiveness of the content adaptor and data augmentation method in improving generation fidelity. It also shows that our content adaptor is capable of maintaining details and attributes of the input objects, while not being constrained by the input when harmonizing objects to the background. In contrast, although SDEdit is better at preserving appearance than BLIP, it is constrained by the original pose, shape, and other attributes of the object, making the generated object less coherent with the background." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 670, + 434, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 670, + 434, + 682 + ], + "spans": [ + { + "bbox": [ + 306, + 670, + 434, + 682 + ], + "type": "text", + "content": "4.3. Qualitative Evaluation" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "In Fig. 6, we compare our method to various baselines. In the first two rows of Fig. 6, we place the same object" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18315" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 87, + 545, + 668 + ], + "blocks": [ + { + "bbox": [ + 76, + 75, + 99, + 83 + ], + "lines": [ + { + "bbox": [ + 76, + 75, + 99, + 83 + ], + "spans": [ + { + "bbox": [ + 76, + 75, + 99, + 83 + ], + "type": "text", + "content": "Object" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 149, + 75, + 194, + 83 + ], + "lines": [ + { + "bbox": [ + 149, + 75, + 194, + 83 + ], + "spans": [ + { + "bbox": [ + 149, + 75, + 194, + 83 + ], + "type": "text", + "content": "Target Image" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 229, + 75, + 283, + 83 + ], + "lines": [ + { + "bbox": [ + 229, + 75, + 283, + 83 + ], + "spans": [ + { + "bbox": [ + 229, + 75, + 283, + 83 + ], + "type": "text", + "content": "Copy-and-paste" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 329, + 75, + 346, + 83 + ], + "lines": [ + { + "bbox": [ + 329, + 75, + 346, + 83 + ], + "spans": [ + { + "bbox": [ + 329, + 75, + 346, + 83 + ], + "type": "text", + "content": "BLIP" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 408, + 75, + 433, + 83 + ], + "lines": [ + { + "bbox": [ + 408, + 75, + 433, + 83 + ], + "spans": [ + { + "bbox": [ + 408, + 75, + 433, + 83 + ], + "type": "text", + "content": "SDEdit" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 493, + 75, + 511, + 83 + ], + "lines": [ + { + "bbox": [ + 493, + 75, + 511, + 83 + ], + "spans": [ + { + "bbox": [ + 493, + 75, + 511, + 83 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 49, + 87, + 545, + 668 + ], + "lines": [ + { + "bbox": [ + 49, + 87, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 49, + 87, + 545, + 668 + ], + "type": "image", + "image_path": "82d0bf282400d201f0171608e61f21fa777e64396600d4c3d2837c47a876cca9.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 48, + 675, + 546, + 709 + ], + "lines": [ + { + "bbox": [ + 48, + 675, + 546, + 709 + ], + "spans": [ + { + "bbox": [ + 48, + 675, + 546, + 709 + ], + "type": "text", + "content": "Figure 6. Qualitative comparison on the real-world evaluation dataset. Our method can appropriately correct the viewpoint and geometry while harmonizing the object with the new background, generating realistic composite results. Moreover, lighting and shadows are naturally adjusted in the composite image. Compared to baselines using text guidance, our method can better preserve the object appearance." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "18316" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 89, + 70, + 503, + 148 + ], + "blocks": [ + { + "bbox": [ + 89, + 70, + 503, + 148 + ], + "lines": [ + { + "bbox": [ + 89, + 70, + 503, + 148 + ], + "spans": [ + { + "bbox": [ + 89, + 70, + 503, + 148 + ], + "type": "table", + "html": "
AdaptorOptimizationAugmentationBLIP [27]FID ↓CLIP text score ↑CLIP image score ↑
✓✓✓✗15.433129.859497.0625
✗✓✓✗18.825429.828196.0000
✓✗✓✗16.138129.796996.8125
✓✓✗✗18.369829.781396.1875
✗✗✓✓17.900029.718895.8125
", + "image_path": "b3f52d7ecf7a5f8e1a5fb9f08fd7e277ec235ca2bc55845860541dc7914ad9ac.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 152, + 547, + 198 + ], + "lines": [ + { + "bbox": [ + 46, + 152, + 547, + 198 + ], + "spans": [ + { + "bbox": [ + 46, + 152, + 547, + 198 + ], + "type": "text", + "content": "Table 2. Ablation study. We evaluate the effectiveness of the components listed in the first four columns: 1) whether the content adaptor is used; 2) whether the adaptor is optimized for appearance preservation; 3) whether crop augmentation is utilized in training; 4) if the adaptor module is replaced by BLIP to predict text embedding from an image. We use FID to assess the generation fidelity and modified CLIP score (explained in Sec. 4.4) to measure similarity of the guidance and the predicted image." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 208, + 289, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 208, + 289, + 293 + ], + "spans": [ + { + "bbox": [ + 46, + 208, + 289, + 293 + ], + "type": "text", + "content": "in two scenes with notable differences in the geometry. Results show that our model is capable of correcting geometric inconsistencies while preserving the characteristics of the object. The limitations of the baselines are also shown in Fig. 6: BLIP often fails to preserve the appearance; SDEdit preserves texture and pose better than the first baseline, but often cannot make appropriate geometry transformations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 300, + 141, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 300, + 141, + 312 + ], + "spans": [ + { + "bbox": [ + 47, + 300, + 141, + 312 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 318, + 288, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 318, + 288, + 415 + ], + "spans": [ + { + "bbox": [ + 46, + 318, + 288, + 415 + ], + "type": "text", + "content": "In Tab. 2, we demonstrate the effectiveness of each key component in our ObjectStitch framework by ablating each of them. We use FID [13] and a modified CLIP [12, 36] score, which measure the fidelity and semantic matching performance between given and generated objects. Instead of directly adopting the CLIP score, which semantically matches the prompt and the image, we modify it to CLIP text score and CLIP image score that are defined as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 93, + 436, + 287, + 449 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 436, + 287, + 449 + ], + "spans": [ + { + "bbox": [ + 93, + 436, + 287, + 449 + ], + "type": "interline_equation", + "content": "\\mathcal {C} _ {t x t} = E \\left[ s \\cdot f \\left(I _ {p r e d}\\right) \\cdot g \\left(B \\left(I _ {g t}\\right)\\right) \\right], \\tag {5}", + "image_path": "a132a71399c00ece01dce722c1dfdf0475aae8234b154e48555f57680bf4219c.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 91, + 451, + 287, + 464 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 451, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 91, + 451, + 287, + 464 + ], + "type": "interline_equation", + "content": "\\mathcal {C} _ {i m g} = E [ s \\cdot f (I _ {p r e d}) \\cdot f (I _ {g t}) ], \\tag {6}", + "image_path": "cbaeda237d24e0a01e287111607156240d259fb3c9d8754863ceff62686a445c.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 473, + 279, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 473, + 279, + 485 + ], + "spans": [ + { + "bbox": [ + 47, + 473, + 279, + 485 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 473, + 279, + 485 + ], + "type": "inline_equation", + "content": "B(\\cdot)" + }, + { + "bbox": [ + 47, + 473, + 279, + 485 + ], + "type": "text", + "content": " is pretrained BLIP [27], and " + }, + { + "bbox": [ + 47, + 473, + 279, + 485 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 47, + 473, + 279, + 485 + ], + "type": "text", + "content": " is a logit scale." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 486, + 287, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 287, + 605 + ], + "type": "text", + "content": "For the ablation models, we first remove the content adaptor to evaluate its effect. In order to keep the embedding dimension fed to the generator, we removed the attention blocks from the content adaptor. We also drop the bias and activation function of the last linear layer, such that in the adaptation process, the model will only learn a non-adaptive embedding dictionary. Without the attention layers, the model has difficulty bridging the domain gap between text and image, resulting in the degradation of generation fidelity as shown in the second row of Tab. 2." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 605, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 665 + ], + "type": "text", + "content": "The third row in Tab. 2 ablates the second training stage of the content adaptor by simply skipping it. Without this process, the model fails to learn descriptive embedding, which encodes instance-level features of the guidance object, leading to a drop in appearance preservation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 666, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 666, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 666, + 287, + 712 + ], + "type": "text", + "content": "Data augmentation is another key element for better performance, i.e., cropping and shift augmentation increases the mask area, thus improving the generation quality in detail. The fourth row in Tab. 2 illustrates that with the ab" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 208, + 545, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 208, + 545, + 232 + ], + "spans": [ + { + "bbox": [ + 305, + 208, + 545, + 232 + ], + "type": "text", + "content": "sence of data augmentation, there is a notable drop in realism, and the generation carries less accurate details." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 233, + 546, + 318 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 233, + 546, + 318 + ], + "spans": [ + { + "bbox": [ + 304, + 233, + 546, + 318 + ], + "type": "text", + "content": "To further assess the importance of the content adaptor module, we replace it with a pretrained BLIP model, which transfers the reference image to a text embedding. As shown in the last row in Tab. 2, this results in the lowest CLIP image score, indicating that high-level semantics represented by text are not sufficient for maintaining object identity, highlighting the importance of our content adaptor." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 331, + 471, + 344 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 331, + 471, + 344 + ], + "spans": [ + { + "bbox": [ + 305, + 331, + 471, + 344 + ], + "type": "text", + "content": "5. Conclusion and Future Work" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 353, + 546, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 353, + 546, + 484 + ], + "spans": [ + { + "bbox": [ + 304, + 353, + 546, + 484 + ], + "type": "text", + "content": "We propose the first diffusion-based approach to tackle object compositing. Traditional methods require many steps including harmonization, geometry/view adjustment, and shadow generation. By contrast, our method directly achieves realistic compositing. We construct our model from a text-to-image generation network and introduce a novel content adaptor module. Furthermore, we present a fully self-supervised framework to train our model with synthetic data generation and augmentation. We show our superior performance over baseline methods through user studies on real-world examples." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 486, + 547, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 547, + 654 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 547, + 654 + ], + "type": "text", + "content": "Since we give the first attempt to handle this challenging task, our method has several limitations. It lacks control over the appearance preservation of the synthesized object. Potential solutions are training the visual encoder on pairs of images with the same object, or training it jointly with the UNet. Another limitation is caused by masking the output image, which prohibits our model from generating global effects, i.e., the shadow can only be synthesized within the mask. To address this issue, we may need to synthesize complete background images and pair them with objects, and train with a conditional generation framework. We also observed artifacts around the object in the background. This can be improved by training the model to also predict an instance mask [49]. We will improve this in our future work." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 669, + 409, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 669, + 409, + 681 + ], + "spans": [ + { + "bbox": [ + 306, + 669, + 409, + 681 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "This work is partially supported by NSF grants 1835739 and 2106717. We also thank Zhiquan Wang for discussion." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18317" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Omri Avrahami, Dani Lischinski, and Ohad Fried. Blended diffusion for text-driven editing of natural images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18208-18218, 2022. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 288, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 288, + 180 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 288, + 180 + ], + "type": "text", + "content": "[2] Samaneh Azadi, Deepak Pathak, Sayna Ebrahimi, and Trevor Darrell. Compositional gan: Learning image-conditional binary composition. International Journal of Computer Vision, 128(10):2570-2585, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 182, + 288, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 182, + 288, + 237 + ], + "spans": [ + { + "bbox": [ + 53, + 182, + 288, + 237 + ], + "type": "text", + "content": "[3] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, et al. edifi: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 239, + 287, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 239, + 287, + 272 + ], + "spans": [ + { + "bbox": [ + 53, + 239, + 287, + 272 + ], + "type": "text", + "content": "[4] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. arXiv preprint arXiv:1809.11096, 2018. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 274, + 288, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 274, + 288, + 317 + ], + "spans": [ + { + "bbox": [ + 53, + 274, + 288, + 317 + ], + "type": "text", + "content": "[5] Bor-Chun Chen and Andrew Kae. Toward realistic image compositing with adversarial learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8415-8424, 2019. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 319, + 287, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 319, + 287, + 373 + ], + "spans": [ + { + "bbox": [ + 53, + 319, + 287, + 373 + ], + "type": "text", + "content": "[6] Wenyan Cong, Li Niu, Jianfu Zhang, Jing Liang, and Liqing Zhang. Bargainnet: Background-guided domain translation for image harmonization. In 2021 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2021. 1, 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 375, + 288, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 375, + 288, + 430 + ], + "spans": [ + { + "bbox": [ + 53, + 375, + 288, + 430 + ], + "type": "text", + "content": "[7] Wenyan Cong, Jianfu Zhang, Li Niu, Liu Liu, Zhixin Ling, Weiyuan Li, and Liqing Zhang. Dovenet: Deep image harmonization via domain verification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8394-8403, 2020. 1, 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 432, + 287, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 432, + 287, + 464 + ], + "spans": [ + { + "bbox": [ + 53, + 432, + 287, + 464 + ], + "type": "text", + "content": "[8] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Deep image homography estimation. arXiv preprint arXiv:1606.03798, 2016. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 466, + 287, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 466, + 287, + 530 + ], + "spans": [ + { + "bbox": [ + 53, + 466, + 287, + 530 + ], + "type": "text", + "content": "[9] Zhida Feng, Zhenyu Zhang, Xintong Yu, Yewei Fang, Lanxin Li, Xuyi Chen, Yuxiang Lu, Jiaxiang Liu, Weichong Yin, Shikun Feng, et al. Ernie-vilg 2.0: Improving text-to-image diffusion model with knowledge-enhanced mixture-of-denoising-experts. arXiv preprint arXiv:2210.15257, 2022.2.3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 533, + 287, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 533, + 287, + 577 + ], + "spans": [ + { + "bbox": [ + 47, + 533, + 287, + 577 + ], + "type": "text", + "content": "[10] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139–144, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 578, + 287, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 578, + 287, + 611 + ], + "spans": [ + { + "bbox": [ + 47, + 578, + 287, + 611 + ], + "type": "text", + "content": "[11] Liu He, Yijuan Lu, John Corring, Dinei Florencio, and Cha Zhang. Diffusion-based document layout generation. arXiv preprint arXiv:2303.10787, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 613, + 287, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 613, + 287, + 656 + ], + "spans": [ + { + "bbox": [ + 47, + 613, + 287, + 656 + ], + "type": "text", + "content": "[12] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021. 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 287, + 712 + ], + "type": "text", + "content": "[13] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 8" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[14] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 107, + 547, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 547, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 547, + 161 + ], + "type": "text", + "content": "[15] Yannick Hold-Geoffroy, Kalyan Sunkavalli, Sunil Hadap, Emiliano Gambaretto, and Jean-François Lalonde. Deep outdoor illumination estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7312–7321, 2017. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 162, + 547, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 162, + 547, + 205 + ], + "spans": [ + { + "bbox": [ + 307, + 162, + 547, + 205 + ], + "type": "text", + "content": "[16] Yan Hong, Li Niu, and Jianfu Zhang. Shadow generation for composite image in real-world scenes. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 914-922, 2022. 1, 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 206, + 547, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 206, + 547, + 258 + ], + "spans": [ + { + "bbox": [ + 307, + 206, + 547, + 258 + ], + "type": "text", + "content": "[17] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1125-1134, 2017. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 261, + 545, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 261, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 307, + 261, + 545, + 293 + ], + "type": "text", + "content": "[18] Max Jaderberg, Karen Simonyan, Andrew Zisserman, et al. Spatial transformer networks. Advances in neural information processing systems, 28, 2015. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 294, + 545, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 294, + 545, + 360 + ], + "spans": [ + { + "bbox": [ + 307, + 294, + 545, + 360 + ], + "type": "text", + "content": "[19] Yifan Jiang, He Zhang, Jianming Zhang, Yilin Wang, Zhe Lin, Kalyan Sunkavalli, Simon Chen, Sohrab Amirghodsi, Sarah Kong, and Zhangyang Wang. Ssh: A self-supervised framework for image harmonization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4832-4841, 2021. 1, 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 361, + 545, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 361, + 545, + 404 + ], + "spans": [ + { + "bbox": [ + 307, + 361, + 545, + 404 + ], + "type": "text", + "content": "[20] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 404, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 404, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 307, + 404, + 545, + 437 + ], + "type": "text", + "content": "[21] Kevin Karsch, Varsha Hedau, David Forsyth, and Derek Hoiem. Rendering synthetic objects into legacy photographs. ACM Transactions on Graphics (TOG), 30(6):1-12, 2011. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 437, + 545, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 437, + 545, + 481 + ], + "spans": [ + { + "bbox": [ + 307, + 437, + 545, + 481 + ], + "type": "text", + "content": "[22] Kevin Karsch, Kalyan Sunkavalli, Sunil Hadap, Nathan Carr, Hailin Jin, Rafael Fonte, Michael Sittig, and David Forsyth. Automatic scene inference for 3d object compositing. ACM Transactions on Graphics (TOG), 33(3):1-15, 2014. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 482, + 545, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 482, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 307, + 482, + 545, + 525 + ], + "type": "text", + "content": "[23] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagc: Text-based real image editing with diffusion models. arXiv preprint arXiv:2210.09276, 2022. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 526, + 545, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 526, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 307, + 526, + 545, + 568 + ], + "type": "text", + "content": "[24] Jean-Francois Lalonde and Alexei A Efros. Using color compatibility for assessing image realism. In 2007 IEEE 11th International Conference on Computer Vision, pages 1-8. IEEE, 2007. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 570, + 545, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 570, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 307, + 570, + 545, + 602 + ], + "type": "text", + "content": "[25] Seung Hoon Lee, Seunghyun Lee, and Byung Cheol Song. Vision transformer for small-size datasets. arXiv preprint arXiv:2112.13492, 2021. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 603, + 547, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 603, + 547, + 647 + ], + "spans": [ + { + "bbox": [ + 307, + 603, + 547, + 647 + ], + "type": "text", + "content": "[26] Youngwan Lee and Jongyoul Park. Centermask: Realtime anchor-free instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13906-13915, 2020. 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 647, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 690 + ], + "type": "text", + "content": "[27] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. arXiv preprint arXiv:2201.12086, 2022. 6, 8" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 691, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 714 + ], + "type": "text", + "content": "[28] Chen-Hsuan Lin, Ersin Yumer, Oliver Wang, Eli Shechtman, and Simon Lucey. St-gan: Spatial transformer generative" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18318" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 67, + 73, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 73, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 67, + 73, + 287, + 106 + ], + "type": "text", + "content": "adversarial networks for image compositing. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 9455-9464, 2018. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 173 + ], + "type": "text", + "content": "[29] Daquan Liu, Chengjiang Long, Hongpan Zhang, Hanning Yu, Xinzhi Dong, and Chunxia Xiao. Arshadowgan: Shadow generative adversarial network for augmented reality in single light scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8139-8148, 2020. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 175, + 287, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 175, + 287, + 228 + ], + "spans": [ + { + "bbox": [ + 49, + 175, + 287, + 228 + ], + "type": "text", + "content": "[30] Xihui Liu, Dong Huk Park, Samaneh Azadi, Gong Zhang, Arman Chopikyan, Yuxiao Hu, Humphrey Shi, Anna Rohrbach, and Trevor Darrell. More control for free! image synthesis with semantic diffusion guidance. arXiv preprint arXiv:2112.05744, 2021. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 230, + 287, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 230, + 287, + 274 + ], + "spans": [ + { + "bbox": [ + 49, + 230, + 287, + 274 + ], + "type": "text", + "content": "[31] Shitong Luo and Wei Hu. Diffusion probabilistic models for 3d point cloud generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2837-2845, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 276, + 287, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 276, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 49, + 276, + 287, + 319 + ], + "type": "text", + "content": "[32] Chenlin Meng, Yang Song, Jiaming Song, Jiajun Wu, JunYan Zhu, and Stefano Ermon. Sdedit: Image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 2, 3, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 321, + 287, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 321, + 287, + 374 + ], + "spans": [ + { + "bbox": [ + 49, + 321, + 287, + 374 + ], + "type": "text", + "content": "[33] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 376, + 287, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 376, + 287, + 408 + ], + "spans": [ + { + "bbox": [ + 49, + 376, + 287, + 408 + ], + "type": "text", + "content": "[34] Patrick Pérez, Michel Gangnet, and Andrew Blake. Poisson image editing. In ACM SIGGRAPH 2003 Papers, pages 313-318. 2003. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 411, + 287, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 411, + 287, + 443 + ], + "spans": [ + { + "bbox": [ + 49, + 411, + 287, + 443 + ], + "type": "text", + "content": "[35] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 445, + 287, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 445, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 49, + 445, + 287, + 510 + ], + "type": "text", + "content": "[36] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 4, 6, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 512, + 287, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 512, + 287, + 565 + ], + "spans": [ + { + "bbox": [ + 49, + 512, + 287, + 565 + ], + "type": "text", + "content": "[37] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 568, + 287, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 568, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 49, + 568, + 287, + 622 + ], + "type": "text", + "content": "[38] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2, 3, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 624, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 624, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 49, + 624, + 287, + 668 + ], + "type": "text", + "content": "[39] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. arXiv preprint arXiv:2208.12242, 2022. 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 49, + 670, + 287, + 713 + ], + "type": "text", + "content": "[40] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 654 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022.3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 96, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 150 + ], + "type": "text", + "content": "[41] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 152, + 545, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 152, + 545, + 206 + ], + "spans": [ + { + "bbox": [ + 308, + 152, + 545, + 206 + ], + "type": "text", + "content": "[42] Yichen Sheng, Yifan Liu, Jianming Zhang, Wei Yin, A Cengiz Oztireli, He Zhang, Zhe Lin, Eli Shechtman, and Bedrich Benes. Controllable shadow generation using pixel height maps. In European Conference on Computer Vision, pages 240-256. Springer, 2022. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 208, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 208, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 308, + 208, + 545, + 251 + ], + "type": "text", + "content": "[43] Yichen Sheng, Jianming Zhang, and Bedrich Benes. Ssn: Soft shadow network for image compositing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4380-4390, 2021. 1, 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 253, + 545, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 253, + 545, + 305 + ], + "spans": [ + { + "bbox": [ + 308, + 253, + 545, + 305 + ], + "type": "text", + "content": "[44] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256-2265. PMLR, 2015. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 308, + 545, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 308, + 545, + 341 + ], + "spans": [ + { + "bbox": [ + 308, + 308, + 545, + 341 + ], + "type": "text", + "content": "[45] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in Neural Information Processing Systems, 32, 2019. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 342, + 545, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 342, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 308, + 342, + 545, + 384 + ], + "type": "text", + "content": "[46] Kalyan Sunkavalli, Micah K Johnson, Wojciech Matusik, and Hanspeter Pfister. Multi-scale image harmonization. ACM Transactions on Graphics (TOG), 29(4):1-10, 2010. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 387, + 545, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 387, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 308, + 387, + 545, + 430 + ], + "type": "text", + "content": "[47] Yi-Hsuan Tsai, Xiaohui Shen, Zhe Lin, Kalyan Sunkavalli, Xin Lu, and Ming-Hsuan Yang. Deep image harmonization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3789-3797, 2017. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 432, + 545, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 432, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 308, + 432, + 545, + 475 + ], + "type": "text", + "content": "[48] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 476, + 545, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 476, + 545, + 518 + ], + "spans": [ + { + "bbox": [ + 308, + 476, + 545, + 518 + ], + "type": "text", + "content": "[49] Shaoan Xie, Zhifei Zhang, Zhe Lin, Tobias Hinz, and Kun Zhang. Smartbrush: Text and shape guided object inpainting with diffusion model. arXiv preprint arXiv:2212.05034, 2022.8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 521, + 545, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 521, + 545, + 564 + ], + "spans": [ + { + "bbox": [ + 308, + 521, + 545, + 564 + ], + "type": "text", + "content": "[50] Ning Xu, Brian Price, Scott Cohen, and Thomas Huang. Deep image matting. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2970-2979, 2017. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 567, + 545, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 567, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 308, + 567, + 545, + 609 + ], + "type": "text", + "content": "[51] Ben Xue, Shenghui Ran, Quan Chen, Rongfei Jia, Binqiang Zhao, and Xing Tang. Dccf: Deep comprehensible color filter learning framework for high-resolution image harmonization. arXiv preprint arXiv:2207.04788, 2022. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 611, + 545, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 611, + 545, + 654 + ], + "spans": [ + { + "bbox": [ + 308, + 611, + 545, + 654 + ], + "type": "text", + "content": "[52] Hao Zhou, Sunil Hadap, Kalyan Sunkavalli, and David W Jacobs. Deep single-image portrait relighting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7194-7202, 2019. 1" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "18319" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Observation-Centric SORT_ Rethinking SORT for Robust Multi-Object Tracking/b4c4cc9d-52d9-4722-825a-050e1d27cf7f_content_list.json b/2023/Observation-Centric SORT_ Rethinking SORT for Robust Multi-Object Tracking/b4c4cc9d-52d9-4722-825a-050e1d27cf7f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..62e6a492ff482476b2f1cb3bf4ea4b78e4cae1c6 --- /dev/null +++ b/2023/Observation-Centric SORT_ Rethinking SORT for Robust Multi-Object Tracking/b4c4cc9d-52d9-4722-825a-050e1d27cf7f_content_list.json @@ -0,0 +1,1565 @@ +[ + { + "type": "text", + "text": "Observation-Centric SORT: Rethinking SORT for Robust Multi-Object Tracking", + "text_level": 1, + "bbox": [ + 218, + 130, + 751, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jinkun Cao $^{1}$ Jiangmiao Pang $^{2}$ Xinshuo Weng $^{3}$ Rawal Khirodkar $^{1}$ Kris Kitani $^{1}$ $^{1}$ Carnegie Mellon University $^{2}$ Shanghai AI Laboratory $^{3}$ Nvidia", + "bbox": [ + 138, + 202, + 830, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 273, + 313, + 289 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kalman filter (KF) based methods for multi-object tracking (MOT) make an assumption that objects move linearly. While this assumption is acceptable for very short periods of occlusion, linear estimates of motion for prolonged time can be highly inaccurate. Moreover, when there is no measurement available to update Kalman filter parameters, the standard convention is to trust the priori state estimations for posteriori update. This leads to the accumulation of errors during a period of occlusion. The error causes significant motion direction variance in practice. In this work, we show that a basic Kalman filter can still obtain state-of-the-art tracking performance if proper care is taken to fix the noise accumulated during occlusion. Instead of relying only on the linear state estimate (i.e., estimation-centric approach), we use object observations (i.e., the measurements by object detector) to compute a virtual trajectory over the occlusion period to fix the error accumulation of filter parameters. This allows more time steps to correct errors accumulated during occlusion. We name our method Observation-Centric SORT (OC-SORT). It remains Simple, Online, and Real-Time but improves robustness during occlusion and non-linear motion. Given off-the-shelf detections as input, OC-SORT runs at $700+$ FPS on a single CPU. It achieves state-of-the-art on multiple datasets, including MOT17, MOT20, KITTI, head tracking, and especially DanceTrack where the object motion is highly non-linear. The code and models are available at https://github.com/noahcao/OC_SORT.", + "bbox": [ + 76, + 305, + 473, + 729 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 753, + 209, + 768 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We aim to develop a motion model-based multi-object tracking (MOT) method that is robust to occlusion and nonlinear motion. Most existing motion model-based algorithms assume that the tracking targets have a constant velocity within a time interval, which is called the linear motion assumption. This assumption breaks in many practical scenarios, but it still works because when the time interval is small enough, the object's motion can be reasonably ap-", + "bbox": [ + 75, + 779, + 468, + 902 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5a0838cea78f29b37d62ae49f4eb3b995510ba9830aa17a2fce3e36949e545bf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 271, + 594, + 376 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bfb6f1f89cbd0d313250928330757f39596ecd085dcca59a94ee799b5a3b1769.jpg", + "image_caption": [ + "(a) SORT" + ], + "image_footnote": [], + "bbox": [ + 596, + 271, + 691, + 376 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6b3834b6036c88f9cf10417ad67ee7e3226beb078ebbf1d6d7e4eb7d4d4cbcb8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 272, + 787, + 376 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8b044f0ea30ee14e9c9801d8db562cf61f06ab5e620e555afc0d9608dbf18400.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 790, + 272, + 883, + 376 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3a16e705f0b020b43a39f9919535d2fc7dbcd1042668bbe2cd15f5a1ad57864c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 391, + 594, + 494 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5e9472c633d3813f1c0e4d0d38b6044a4cc65696afeaf1271b2e371ca652d7ff.jpg", + "image_caption": [ + "(b) The proposed OC-SORT" + ], + "image_footnote": [], + "bbox": [ + 596, + 391, + 689, + 493 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b92475af28c88b8ee65bab550c8ef4291a7393e219ed5930f5d523310ac6e6d3.jpg", + "image_caption": [ + "Figure 1. Samples from the results on DanceTrack [54]. SORT and OC-SORT use the same detection results. On the third frame, SORT encounters an ID switch for the backflip target while OC-SORT tracks it consistently." + ], + "image_footnote": [], + "bbox": [ + 694, + 391, + 787, + 493 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/86515da1892e887e1bbb101381a5768a76a55a57a3f9b1affbd9e1d6d1ff1352.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 790, + 391, + 883, + 493 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "proximated as linear. In this work, we are motivated by the fact that most of the errors from motion model-based tracking methods occur when occlusion and non-linear motion happen together. To mitigate the adverse effects caused, we first rethink current motion models and recognize some limitations. Then, we propose addressing them for more robust tracking performance, especially in occlusion.", + "bbox": [ + 496, + 611, + 893, + 717 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As the main branch of motion model-based tracking, filtering-based methods assume a transition function to predict the state of objects on future time steps, which are called state \"estimations\". Besides estimations, they leverage an observation model, such as an object detector, to derive the state measurements of target objects, also called \"observations\". Observations usually serve as auxiliary information to help update the posteriori parameters of the filter. The trajectories are still extended by the state estimations. Among this line of work, the most widely used one is SORT [3], which uses a Kalman filter (KF) to estimate object states and a linear motion function as the transition", + "bbox": [ + 496, + 719, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "9686", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "function between time steps. However, SORT shows insufficient tracking robustness when the object motion is nonlinear, and no observations are available when updating the filter posteriori parameters.", + "bbox": [ + 78, + 90, + 467, + 150 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we recognize three limitations of SORT. First, although the high frame rate is the key to approximating the object motion as linear, it also amplifies the model's sensitivity to the noise of state estimations. Specifically, between consecutive frames of a high frame-rate video, we demonstrate that the noise of displacement of the object can be of the same magnitude as the actual object displacement, leading to the estimated object velocity by KF suffering from a significant variance. Also, the noise in the velocity estimate will accumulate into the position estimate by the transition process. Second, the noise of state estimations by KF is accumulated along the time when there is no observation available in the update stage of KF. We show that the error accumulates very fast with respect to the time of the target object's being untracked. The noise's influence on the velocity direction often makes the track lost again even after re-association. Last, given the development of modern detectors, the object state by detections usually has lower variance than the state estimations propagated along time steps by a fixed transition function in filters. However, SORT is designed to prolong the object trajectories by state estimations instead of observations.", + "bbox": [ + 80, + 154, + 467, + 484 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To relieve the negative effect of these limitations, we propose two main innovations in this work. First, we design a module to use object state observations to reduce the accumulated error during the track's being lost in a backcheck fashion. To be precise, besides the traditional stages of predict and update, we add a stage of re-update to correct the accumulated error. The re-update is triggered when a track is re-activated by associating to an observation after a period of being untracked. The re-update uses virtual observations on the historical time steps to prevent error accumulation. The virtual observations come from a trajectory generated using the last-seen observation before untracked and the latest observation re-activating this track as anchors. We name it Observation-centric Re-Update (ORU).", + "bbox": [ + 80, + 488, + 467, + 698 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Besides ORU, the assumption of linear motion provides the consistency of the object motion direction. But this cue is hard to be used in SORT's association because of the heavy noise in direction estimation. But we propose an observation-centric manner to incorporate the direction consistency of tracks in the cost matrix for the association. We name it Observation-Centric Momentum (OCM). We also provide analytical justification for the noise of velocity direction estimation in practice.", + "bbox": [ + 80, + 702, + 467, + 837 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The proposed method, named as Observation-Centric SORT or OC-SORT in short, remains simple, online, real-time and significantly improves robustness over occlusion and non-linear motion. Our contributions are summarized", + "bbox": [ + 78, + 840, + 467, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "as the following:", + "bbox": [ + 504, + 92, + 612, + 106 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We recognize, analytically and empirically, three limitations of SORT, i.e. sensitivity to the noise of state estimations, error accumulation over time, and being estimation-centric;", + "2. We propose OC-SORT for tracking under occlusion and non-linear motion by fixing SORT's limitations. It achieves state-of-the-art performance on multiple datasets in an online and real-time fashion." + ], + "bbox": [ + 514, + 111, + 888, + 233 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 504, + 261, + 645, + 275 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Motion Models. Many modern MOT algorithms [3, 11, 63, 70, 73] use motion models. Typically, these motion models use Bayesian estimation [34] to predict the next state by maximizing a posterior estimation. As one of the most classic motion models, Kalman filter (KF) [30] is a recursive Bayes filter that follows a typical predict-update cycle. The true state is assumed to be an unobserved Markov process, and the measurements are observations from a hidden Markov model [44]. Given that the linear motion assumption limits KF, follow-up works like Extended KF [52] and Unscented KF [28] were proposed to handle non-linear motion with first-order and third-order Taylor approximation. However, they still rely on approximating the Gaussian prior assumed by KF and require motion pattern assumption. On the other hand, particle filters [22] solve the non-linear motion by sampling-based posterior estimation but require exponential order of computation. Therefore, these variants of Kalman filter and particle filters are rarely adopted in the visual multi-object tracking and the mostly adopted motion model is still based on Kalman filter [3].", + "bbox": [ + 503, + 289, + 888, + 590 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multi-object Tracking. As a classic computer vision task, visual multi-object tracking is traditionally approached from probabilistic perspectives, e.g. joint probabilistic association [1]. And modern video object tracking is usually built upon modern object detectors [46, 48, 74]. SORT [3] adopts the Kalman filter for motion-based multi-object tracking given observations from deep detectors. DeepSORT [63] further introduces deep visual features [23, 51] into object association under the framework of SORT. Re-identification-based object association[42, 63, 71] has also become popular since then but falls short when scenes are crowded and objects are represented coarsely (e.g. enclosed by bounding boxes), or object appearance is not distinguishable. More recently, transformers [58] have been introduced to MOT [8, 39, 55, 69] to learn deep representations from both visual information and object trajectories. However, their performance still has a significant gap between state-of-the-art tracking-by-detection methods in terms of both accuracy and time efficiency.", + "bbox": [ + 503, + 613, + 888, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "9687", + "bbox": [ + 483, + 946, + 513, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Rethink the Limitations of SORT", + "text_level": 1, + "bbox": [ + 76, + 89, + 382, + 106 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we review Kalman filter and SORT [3]. We recognize some of their limitations, which are significant with occlusion and non-linear object motion. We are motivated to improve tracking robustness by fixing them.", + "bbox": [ + 76, + 114, + 468, + 176 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 76, + 183, + 217, + 199 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Kalman filter (KF) [30] is a linear estimator for dynamical systems discretized in the time domain. KF only requires the state estimations on the previous time step and the current measurement to estimate the target state on the next time step. The filter maintains two variables, the posteriori state estimate $\\mathbf{x}$ , and the posteriori estimate covariance matrix $\\mathbf{P}$ . In the task of object tracking, we describe the KF process with the state transition model $\\mathbf{F}$ , the observation model $\\mathbf{H}$ , the process noise $\\mathbf{Q}$ , and the observation noise $\\mathbf{R}$ . At each step $t$ , given observations $\\mathbf{z}_t$ , KF works in an alternation of predict and update stages:", + "bbox": [ + 76, + 207, + 472, + 375 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\text {p r e d i c t} \\left\\{ \\begin{array}{l} \\hat {\\mathbf {x}} _ {t | t - 1} = \\mathbf {F} _ {t} \\hat {\\mathbf {x}} _ {t - 1 | t - 1} \\\\ \\mathbf {P} _ {t | t - 1} = \\mathbf {F} _ {t} \\mathbf {P} _ {t - 1 | t - 1} \\mathbf {F} _ {t} ^ {\\top} + \\mathbf {Q} _ {t} \\end{array} , \\right. \\\\ \\text {u p d a t e} \\left\\{ \\begin{array}{l} \\mathbf {K} _ {t} = \\mathbf {P} _ {t | t - 1} \\mathbf {H} _ {t} ^ {\\top} \\left(\\mathbf {H} _ {t} \\mathbf {P} _ {t | t - 1} \\mathbf {H} _ {t} ^ {\\top} + \\mathbf {R} _ {t}\\right) ^ {- 1} \\quad (1) \\\\ \\hat {\\mathbf {x}} _ {t | t} = \\hat {\\mathbf {x}} _ {t | t - 1} + \\mathbf {K} _ {t} (\\mathbf {z} _ {t} - \\mathbf {H} _ {t} \\hat {\\mathbf {x}} _ {t | t - 1}) \\\\ \\mathbf {P} _ {t | t} = (\\mathbf {I} - \\mathbf {K} _ {t} \\mathbf {H} _ {t}) \\mathbf {P} _ {t | t - 1} \\end{array} . \\right. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 381, + 468, + 481 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The stage of predict is to derive the state estimations on the next time step $t$ . Given a measurement of target states on the next step $t$ , the stage of update aims to update the posteriori parameters in KF. Because the measurement comes from the observation model $\\mathbf{H}$ , it is also called \"observation\" in many scenarios.", + "bbox": [ + 76, + 488, + 468, + 577 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "SORT [3] is a multi-object tracker built upon KF. The KF's state $\\mathbf{x}$ in SORT is defined as $\\mathbf{x} = [u,v,s,r,\\dot{u},\\dot{v},\\dot{s}]^{\\top}$ , where $(u,v)$ is the 2D coordinates of the object center in the image. $s$ is the bounding box scale (area) and $r$ is the bounding box aspect ratio. The aspect ratio $r$ is assumed to be constant. The other three variables, $\\dot{u},\\dot{v}$ and $\\dot{s}$ are the corresponding time derivatives. The observation is a bounding box $\\mathbf{z} = [u,v,w,h,c]^{\\top}$ with object center position $(u,v)$ , object width $w$ , and height $h$ and the detection confidence $c$ respectively. SORT assumes linear motion as the transition model $\\mathbf{F}$ which leads to the state estimation as", + "bbox": [ + 76, + 578, + 468, + 744 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nu _ {t + 1} = u _ {t} + \\dot {u} _ {t} \\Delta t, \\quad v _ {t + 1} = v _ {t} + \\dot {v} _ {t} \\Delta t. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 753, + 468, + 771 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To leverage KF (Eq 1) in SORT for visual MOT, the stage of predict corresponds to estimating the object position on the next video frame. And the observations used for the update stage usually come from a detection model. The update stage is to update Kalman filter parameters and does not directly edit the tracking outcomes.", + "bbox": [ + 76, + 779, + 468, + 869 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "When the time difference between two steps is constant during the transition, $e.g.$ , the video frame rate is constant,", + "bbox": [ + 76, + 869, + 468, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "we can set $\\Delta t = 1$ . When the video frame rate is high, SORT works well even when the object motion is non-linear globally, (e.g. dancing, fencing, wrestling) because the motion of the target object can be well approximated as linear within short time intervals. However, in practice, observations are often absent on some time steps, e.g. the target object is occluded in multi-object tracking. In such cases, we cannot update the KF parameters by the update operation as in Eq. 1 anymore. SORT uses the priori estimations directly as posterior. We call this \"dummy update\", namely", + "bbox": [ + 496, + 90, + 890, + 243 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {x}} _ {t \\mid t} = \\hat {\\mathbf {x}} _ {t \\mid t - 1}, \\mathbf {P} _ {t \\mid t} = \\mathbf {P} _ {t \\mid t - 1}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 253, + 890, + 272 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The philosophy behind such a design is to trust estimations when no observations are available to supervise them. We thus call the tracking algorithms following this scheme \"estimation-centric\". However, we will see that this estimation-centric mechanism can cause trouble when nonlinear motion and occlusion happen together.", + "bbox": [ + 496, + 281, + 890, + 372 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Limitations of SORT", + "text_level": 1, + "bbox": [ + 500, + 381, + 697, + 396 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we identify three main limitations of SORT which are connected. This analysis lays the foundation of our proposed method.", + "bbox": [ + 496, + 404, + 890, + 450 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2.1 Sensitive to State Noise", + "text_level": 1, + "bbox": [ + 500, + 460, + 714, + 474 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Now we show that SORT is sensitive to the noise from KF's state estimations. To begin with, we assume that the estimated object center position follows $u \\sim \\mathcal{N}(\\mu_u, \\sigma_u^2)$ and $v \\sim \\mathcal{N}(\\mu_v, \\sigma_v^2)$ , where $(\\mu_u, \\mu_v)$ is the underlying true position. Then, if we assume that the state noises are independent on different steps, by Eq.2, the object speed between two time steps, $t \\longrightarrow t + \\Delta t$ , is", + "bbox": [ + 496, + 483, + 890, + 590 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\dot {u} = \\frac {u _ {t + \\Delta t} - u _ {t}}{\\Delta t}, \\quad \\dot {v} = \\frac {v _ {t + \\Delta t} - v _ {t}}{\\Delta t}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 601, + 890, + 628 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "making the noise of estimated speed $\\delta_{\\dot{u}}\\sim \\mathcal{N}(0,\\frac{2\\sigma_u^2}{(\\Delta t)^2})$ $\\delta_{\\dot{v}}\\sim \\mathcal{N}(0,\\frac{2\\sigma_v^2}{(\\Delta t)^2})$ . Therefore, a small $\\Delta t$ will amplify the noise. This suggests that SORT will suffer from the heavy noise of velocity estimation on high-frame-rate videos. The analysis above is simplified from the reality. In practice, velocity won't be determined by the state on future time steps. For a more strict analysis, please refer to Appendix G.", + "bbox": [ + 496, + 632, + 890, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Moreover, for most multi-object tracking scenarios, the target object displacement is only a few pixels between consecutive frames. For instance, the average displacement is 1.93 pixels and 0.65 pixels along the image width and height for the MOT17 [41] training dataset. In such a case, even if the estimated position has a shift of only a single pixel, it causes a significant variation in the estimated speed. In general, the variance of the speed estimation can be of the same magnitude as the speed itself or even greater. This will not make a massive impact as the shift is only of few pixels", + "bbox": [ + 496, + 750, + 890, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "9688", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d9a8507f837ba33dcae203cc6bf03fba0cfc3bcadb76306f639c7a4ecec00b36.jpg", + "image_caption": [ + "Figure 2. The pipeline of our proposed OC-SORT. The red boxes are detections, orange boxes are active tracks, blue boxes are untracked tracks, and dashed boxes are the estimates from KF. During association, OCM is used to add the velocity consistency cost. The target #1 is lost on the frame t+1 because of occlusion. But on the next frame, it is recovered by referring to its observation of the frame t by OCR. It being re-tracked triggers ORU from t to t+2 for the parameters of its KF." + ], + "image_footnote": [], + "bbox": [ + 127, + 88, + 851, + 328 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "from the ground truth on the next time step and the observations, whose variance is independent of the time, will be able to fix the noise when updating the posteriori parameters. However, we find that such a high sensitivity to state noise introduces significant problems in practice after being amplified by the error accumulation across multiple time steps when no observation is available for KF update.", + "bbox": [ + 75, + 410, + 470, + 516 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.2 Temporal Error Magnification", + "text_level": 1, + "bbox": [ + 76, + 518, + 344, + 534 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For analysis above in Eq. 4, we assume the noise of the object state is i.i.d on different time steps (this is a simplified version, a more detailed analysis is provided in Appendix G). This is reasonable for object detections but not for the estimations from KF. This is because KF's estimations always rely on its estimations on previous time steps. The effect is usually minor because KF can use observation in update to prevent the posteriori state estimation and covariance, i.e. $\\hat{\\mathbf{x}}_{t|t}$ and $\\mathbf{P}_{t|t}$ , deviating from the true value too far away. However, when no observations are provided to KF, it cannot use observation to update its parameters. Then it has to follow Eq. 3 to prolong the estimated trajectory to the next time step. Consider a track is occluded on the time steps between $t$ and $t + T$ and the noise of speed estimate follows $\\delta_{\\dot{u}_t} \\sim \\mathcal{N}(0,2\\sigma_u^2)$ , $\\delta_{\\dot{v}_t} \\sim \\mathcal{N}(0,2\\sigma_v^2)$ for SORT. On the step $t + T$ , state estimation would be", + "bbox": [ + 75, + 542, + 468, + 785 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nu _ {t + T} = u _ {t} + T \\dot {u} _ {t}, \\quad v _ {t + T} = v _ {t} + T \\dot {v} _ {t}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 791, + 468, + 806 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "whose noise follows $\\delta_{u_{t + T}}\\sim \\mathcal{N}(0,2T^2\\sigma_u^2)$ and $\\delta_{v_{t + T}}\\sim$ $\\mathcal{N}(0,2T^2\\sigma_v^2)$ . So without the observations, the estimation from the linear motion assumption of KF results in a fast error accumulation with respect to time. Given $\\sigma_v$ and $\\sigma_{u}$ is of the same magnitude as object displacement between consecutive frames, the noise of final object position", + "bbox": [ + 75, + 809, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$(u_{t + T}, v_{t + T})$ is of the same magnitude as the object size. For instance, the size of pedestrians close to the camera on MOT17 is around $50 \\times 300$ pixels. So even assuming the variance of position estimation is only 1 pixel, 10-frame occlusion can accumulate a shift in final position estimation as large as the object size. Such error magnification leads to a major accumulation of errors when the scenes are crowded.", + "bbox": [ + 496, + 409, + 890, + 515 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.3 Estimation-Centric", + "text_level": 1, + "bbox": [ + 498, + 521, + 689, + 535 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The aforementioned limitations come from a fundamental property of SORT that it follows KF to be estimation-centric. It allows update without the existence of observations and purely trusts the estimations. A key difference between state estimations and observations is that we can assume that the observations by an object detector in each frame are affected by i.i.d. noise $\\delta_{\\mathbf{z}} \\sim \\mathcal{N}(0, \\sigma'^2)$ while the noise in state estimations can be accumulated along the hidden Markov process. Moreover, modern object detectors use powerful object visual features [48, 51]. It makes that, even on a single frame, it is usually safe to assume $\\sigma' < \\sigma_u$ and $\\sigma' < \\sigma_v$ because the object localization is more accurate by detection than from the state estimations through linear motion assumption. Combined with the previously mentioned two limitations, being estimation-centric makes SORT suffer from heavy noise when there is occlusion and the object motion is not perfectly linear.", + "bbox": [ + 496, + 540, + 890, + 797 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Observation-Centric SORT", + "text_level": 1, + "bbox": [ + 498, + 814, + 756, + 829 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we introduce the proposed Observation-Centric SORT (OC-SORT). To address the limitations of SORT discussed above, we use the momentum of the object moving into the association stage and develop a", + "bbox": [ + 496, + 840, + 893, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "9689", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1bbcd5172d3c920aac4e5ce8a63f11b55c0a4fdeb4743beb0bedc75c26af596e.jpg", + "image_caption": [ + "Figure 3. Example of how Observation-centric Re-Update (ORU) reduces the error accumulation when a track is broken. The target is occluded between the second and the third time step and the tracker finds it back at the third step. Yellow boxes are the state observations by the detector. White stars are the estimated centers without ORU. Yellow stars are the estimated centers fixed by ORU. The gray star on the fourth step is the estimated center without ORU and fails to match observations." + ], + "image_footnote": [], + "bbox": [ + 158, + 88, + 810, + 239 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "pipeline with less noise and more robustness over occlusion and non-linear motion. The key is to design the tracker as observation-centric instead of estimation-centric. If a track is recovered from being untracked, we use an Observation-centric Re-Update (ORU) strategy to counter the accumulated error during the untracked period. OC-SORT also adds an Observation-Centric Momentum (OCM) term in the association cost. Please refer to Algorithm 1 in Appendix for the pseudo-code of OC-SORT. The pipeline is shown in Fig. 2.", + "bbox": [ + 75, + 318, + 473, + 470 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Observation-centric Re-Update (ORU)", + "text_level": 1, + "bbox": [ + 76, + 477, + 410, + 494 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In practice, even if an object can be associated again by SORT after a period of being untracked, it is probably lost again because its KF parameters have already deviated far away from the correct due to the temporal error magnification. To alleviate this problem, we propose Observation-centric Re-Update (ORU) to reduce the accumulated error. Once a track is associated with an observation again after a period of being untracked (\"re-activation\"), we backcheck the period of its being lost and re-update the parameters of KF. The re-update is based on \"observations\" from a virtual trajectory. The virtual trajectory is generated referring to the observations on the steps starting and ending the untracked period. For example, by denoting the last-seen observation before being untracked as $\\mathbf{z}_{t_1}$ and the observation triggering the re-association as $\\mathbf{z}_{t_2}$ , the virtual trajectory is denoted as", + "bbox": [ + 75, + 500, + 473, + 729 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathbf {z}} _ {t} = \\operatorname {T r a j} _ {\\text {v i r t u a l}} \\left(\\mathbf {z} _ {t _ {1}}, \\mathbf {z} _ {t _ {2}}, t\\right), t _ {1} < t < t _ {2}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 137, + 742, + 468, + 758 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then, along the trajectory of $\\tilde{\\mathbf{z}}_t(t_1 < t < t_2)$ , we run the loop of predict and re-update. The re-update operation is", + "bbox": [ + 76, + 763, + 468, + 794 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nr e - u p d a t e \\left\\{ \\begin{array}{l} \\mathbf {K} _ {t} = \\mathbf {P} _ {t | t - 1} \\mathbf {H} _ {t} ^ {\\top} \\left(\\mathbf {H} _ {t} \\mathbf {P} _ {t | t - 1} \\mathbf {H} _ {t} ^ {\\top} + \\mathbf {R} _ {t}\\right) ^ {- 1} \\\\ \\hat {\\mathbf {x}} _ {t | t} = \\hat {\\mathbf {x}} _ {t | t - 1} + \\mathbf {K} _ {t} \\left(\\tilde {\\mathbf {z}} _ {t} - \\mathbf {H} _ {t} \\hat {\\mathbf {x}} _ {t | t - 1}\\right) \\\\ \\mathbf {P} _ {t | t} = (\\mathbf {I} - \\mathbf {K} _ {t} \\mathbf {H} _ {t}) \\mathbf {P} _ {t | t - 1} \\end{array} \\right. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 800, + 468, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As the observations on the virtual trajectory match the motion pattern anchored by the last-seen and the latest asso", + "bbox": [ + 76, + 869, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/dd82417e88e92bedf836ce19b2d8d0998876de1b60b1d7e712f34cb9dcf768ff.jpg", + "image_caption": [ + "Figure 4. Calculation of motion direction difference in OCM. The green line indicates an existing track and the dots are the observations on it. The red dots are the new observations to be associated. The blue link and the yellow link form the directions of $\\theta^{\\mathrm{track}}$ and $\\theta^{\\mathrm{intention}}$ respectively. The included angle is the difference of direction $\\Delta \\theta$ ." + ], + "image_footnote": [], + "bbox": [ + 555, + 319, + 849, + 398 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ciated real observations, the update will not suffer from the error accumulated through the dummy update anymore. We call the proposed process Observation-centric Re-Update. It serves as an independent stage outside the predict-update loop and is triggered only a track is re-activated from a period of having no observations.", + "bbox": [ + 496, + 513, + 893, + 604 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Observation-Centric Momentum (OCM)", + "text_level": 1, + "bbox": [ + 500, + 617, + 849, + 633 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In a reasonably short time interval, we can approximate the motion as linear. And the linear motion assumption also asks for consistent motion direction. But the noise prevents us from leveraging the consistency of direction. To be precise, to determine the motion direction, we need the object state on two steps with a time difference $\\Delta t$ . If $\\Delta t$ is small, the velocity noise would be significant because of the estimation's sensitivity to state noise. If $\\Delta t$ is big, the noise of direction estimation can also be significant because of the temporal error magnification and the failure of linear motion assumption. As state observations have no problem of temporal error magnification that state estimations suffer from, we propose to use observations instead of estimations to reduce the noise of motion direction calculation and introduce the term of velocity consistency to help the association.", + "bbox": [ + 496, + 642, + 893, + 882 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With the new term, given $N$ existing tracks and $M$ de", + "bbox": [ + 519, + 885, + 890, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "9690", + "bbox": [ + 482, + 944, + 514, + 957 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "tectons on the new-coming time step, the association cost matrix is formulated as", + "bbox": [ + 76, + 90, + 468, + 119 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nC (\\hat {\\mathbf {X}}, \\mathbf {Z}) = C _ {\\mathrm {I o U}} (\\hat {\\mathbf {X}}, \\mathbf {Z}) + \\lambda C _ {v} (\\mathcal {Z}, \\mathbf {Z}), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 130, + 468, + 148 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\hat{\\mathbf{X}}\\in \\mathbb{R}^{N\\times 7}$ is the set of object state estimations and $\\mathbf{Z}\\in \\mathbb{R}^{M\\times 5}$ is the set of observations on the new time step. $\\lambda$ is a weighting factor. $\\mathcal{Z}$ contains the trajectory of observations of all existing tracks. $C_{\\mathrm{IoU}}(\\cdot ,\\cdot)$ calculates the negative pairwise IoU (Intersection over Union) and $C_v(\\cdot ,\\cdot)$ calculates the consistency between the directions of i) linking two observations on an existing track $(\\theta^{\\mathrm{track}})$ and ii) linking a track's historical observation and a new observation $(\\theta^{\\mathrm{intention}})$ . $C_v$ contains all pairs of $\\Delta \\theta = |\\theta^{\\mathrm{track}} - \\theta^{\\mathrm{intention}}|$ . In our implementation, we calculate the motion direction in radians, namely $\\theta = \\arctan \\left(\\frac{v_1 - v_2}{u_1 - u_2}\\right)$ where $(u_{1},v_{1})$ and $(u_{2},v_{2})$ are the observations on two different time steps. The calculation is also illustrated in Figure 4.", + "bbox": [ + 76, + 157, + 468, + 354 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Following the assumptions of noise distribution mentioned before, we can derive a closed-form probability density function of the distribution of the noise in the direction estimation. The derivation is explained in detail in Appendix A. By analyzing the property of this distribution, we reach a conclusion that, under the linear-motion model, the scale of the noise of direction estimation is negatively correlated to the time difference between the two observation points, i.e. $\\Delta t$ . This suggests increasing $\\Delta t$ to achieve a low-noisy estimation of $\\theta$ . However, the assumption of linear motion typically holds only when $\\Delta t$ is small enough. Therefore, the choice of $\\Delta t$ requires a trade-off.", + "bbox": [ + 75, + 356, + 468, + 536 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Besides ORU and OCM, we also find it empirically helpful to check a track's last presence to recover it from being lost. We thus apply a heuristic Observation-Centric Recovery (OCR) technique. OCR will start a second attempt of associating between the last observation of unmatched tracks to the unmatched observations after the usual association stage. It can handle the case of an object stopping or being occluded for a short time interval.", + "bbox": [ + 75, + 537, + 468, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 671, + 207, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 76, + 696, + 266, + 713 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We evaluate our method on multiple multi-object tracking datasets including MOT17 [41], MOT20 [14], KITTI [20], DanceTrack [54] and CroHD [56]. MOT17 [41] and MOT20 [14] are for pedestrian tracking, where targets mostly move linearly, while scenes in MOT20 are more crowded. KITTI [20] is for pedestrian and car tracking with a relatively low frame rate of 10FPS. CroHD is a dataset for head tracking in the crowd and the results on it are included in the appendix. DanceTrack [54] is a recently proposed dataset for human tracking. For the data in DanceTrack, object localization is easy, but the object motion is highly non-linear. Furthermore, the objects", + "bbox": [ + 75, + 719, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "have a close appearance, severe occlusion, and frequent crossovers. Considering our goal is to improve tracking robustness under occlusion and non-linear object motion, we would emphasize the comparison on DanceTrack.", + "bbox": [ + 496, + 90, + 890, + 151 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementations. For a fair comparison, we directly apply the object detections from existing baselines. For MOT17, MOT20, and DanceTrack, we use the publicly availableYOLOX [19] detector weights by ByteTrack [70]. For KITTI [20], we use the detections from PermaTrack [57] publicly available in the official release1. For ORU, we generate the virtual trajectory during occlusion with the constant-velocity assumption. Therefore, Eq. 6 is adopted as $\\tilde{\\mathbf{z}}_t = \\mathbf{z}_{t_1} + \\frac{t - t_1}{t_2 - t_1} (\\mathbf{z}_{t_2} - \\mathbf{z}_{t_1}), t_1 < t < t_2$ . For OCM, the velocity direction is calculated using the observations three time steps apart, i.e. $\\Delta t = 3$ . The direction difference is measured by the absolute difference of angles in radians. We set $\\lambda = 0.2$ in Eq. 8. Following the common practice of SORT, we set the detection confidence threshold at 0.4 for MOT20 and 0.6 for other datasets. The IoU threshold during association is 0.3.", + "bbox": [ + 496, + 152, + 892, + 393 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics. We adopt HOTA [37] as the main metric as it maintains a proper balance between the accuracy of object detection and association [37]. We also emphasize AssA to evaluate the association performance. IDF1 is also used for association performance evaluation. Other metrics we report, such as MOTA, are highly related to detection performance. It is fair to use these metrics only when all methods use the same detections for tracking, which is referred to as \"public tracking\" as reported in Appendix C.", + "bbox": [ + 496, + 395, + 892, + 532 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Benchmark Results", + "text_level": 1, + "bbox": [ + 500, + 544, + 686, + 559 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Here we report the benchmark results on multiple datasets. We put all methods that use the shared detection results in the blue blocks at the bottom of each table.", + "bbox": [ + 496, + 569, + 890, + 613 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "MOT17 and MOT20. We report OC-SORT's performance on MOT17 and MOT20 in Table 1 and Table 2 using private detections. To make a fair comparison, we use the same detection as ByteTrack [70]. OC-SORT achieves performance comparable to other state-of-the-art methods. Our gains are especially significant in MOT20 under severe pedestrian occlusion, setting a state-of-the-art HOTA of 62.1. As our method is designed to be simple for better generalization, we do not use adaptive detection thresholds as in ByteTrack. Also, ByteTrack uses more detections of low-confidence to achieve higher MOTA scores but we keep the detection confidence threshold the same as on other datasets, which is the common practice in the community. We inherit the linear interpolation on the two datasets by baseline methods for a fair comparison. To more clearly discard the variance from the detector, we also perform public tracking on MOT17 and MOT20, which is reported in Table 12 and Table 13 in", + "bbox": [ + 496, + 614, + 892, + 872 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "1 https://github.com/TRI-ML/permatrack/", + "bbox": [ + 517, + 886, + 736, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "9691", + "bbox": [ + 480, + 944, + 513, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/709af639570817f2467cf2173b10c89036a78010c02a753d91a891cef84f7273.jpg", + "table_caption": [ + "Table 1. Results on MOT17-test with the private detections. ByteTrack and OC-SORT share detections." + ], + "table_footnote": [], + "table_body": "
TrackerHOTA↑MOTA↑IDF1↑FP(104)↓FN(104)↓IDs↓Frag↓AssA↑AssR↑
FairMOT [71]59.373.772.32.7511.73,3038,07358.063.6
TransCt [67]54.573.262.22.3112.44,6149,51949.754.2
TransTrk [55]54.175.263.55.028.643,6034,87247.957.1
GRTU [60]62.074.975.03.2010.81,8121,82462.165.8
QDTrack [42]53.968.766.32.6614.73,3788,09152.757.2
MOTR [69]57.271.968.42.1113.62,1153,89755.859.2
PermaTr [57]55.573.868.92.9011.53,6996,13253.159.8
TransMOT [12]61.776.775.13.629.322,3467,71959.966.5
GTR [75]59.175.371.52.6811.02,859-61.6-
DST-Tracker [8]60.175.272.32.4211.02,729-62.1-
MeMOT [5]56.972.569.02.7211.52,724-55.2-
UniCom [68]61.777.275.55.017.335,379---
ByteTrack [70]63.180.377.32.558.372,1962,27762.068.2
OC-SORT63.278.077.51.5110.81,9502,04063.267.5
", + "bbox": [ + 166, + 106, + 802, + 276 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/54f30aece2e2af5d3c3d9cb24809c9f30470ef059b75340736299c92aeb0851c.jpg", + "table_caption": [ + "Table 2. Results on MOT20-test with private detections. ByteTrack and OC-SORT share detections." + ], + "table_footnote": [], + "table_body": "
TrackerHOTA↑MOTA↑IDF1↑FP(104)↓FN(104)↓IDs↓Frag↓AssA↑AssR↑
FairMOT [71]54.661.867.310.38.895,2437,87454.760.7
TransCt [67]43.558.549.66.4214.64,6959,58137.045.1
Semi-TCL [35]55.365.270.16.1211.54,1398,50856.360.9
CSTrack [36]54.066.668.62.5414.43,1967,63254.057.6
GSDT [61]53.667.167.53.1913.53,1319,87552.758.5
TransMOT [12]61.977.575.23.428.081,6152,42160.166.3
MeMOT [5]54.163.766.14.7913.81,938-55.0-
ByteTrack [70]61.377.875.22.628.761,2231,46059.666.2
OC-SORT62.175.575.91.8010.89131,19862.067.5
", + "bbox": [ + 163, + 305, + 810, + 422 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f1de8a177c5db9e420995daf7203362ce12946acd3e005091247b770da9edf36.jpg", + "table_caption": [ + "Table 3. Results on DanceTrack test set. Methods in the blue block share the same detections." + ], + "table_footnote": [], + "table_body": "
TrackerHOTA↑DetA↑AssA↑MOTA↑IDF1↑
CenterTrack [73]41.878.122.686.835.7
FairMOT [71]39.766.723.882.240.8
QDTrack [42]45.772.129.283.044.8
TransTrk[55]45.575.927.588.445.2
TraDes [64]43.374.525.486.241.2
MOTR [69]54.273.540.279.751.5
GTR [75]48.072.531.984.750.3
DST-Tracker [8]51.972.334.684.951.0
SORT [3]47.972.031.291.850.8
DeepSORT [63]45.671.029.787.847.9
ByteTrack [70]47.371.631.489.552.5
OC-SORT54.680.440.289.654.6
OC-SORT + Linear Interp55.180.440.492.254.9
", + "bbox": [ + 78, + 472, + 470, + 628 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Appendix C. OC-SORT still outperforms the existing state-of-the-art in public tracking settings.", + "bbox": [ + 76, + 641, + 468, + 671 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "DanceTrack. To evaluate OC-SORT under challenging non-linear object motion, we report results on the DanceTrack in Table 3. OC-SORT sets a new state-of-the-art, outperforming the baselines by a great margin under nonlinear object motions. We compare the tracking results of SORT and OC-SORT under extreme non-linear situations in Fig.1 and more samples are available in Fig.8 in Appendix E. We also visualize the output trajectories by OC-SORT and SORT on randomly selected DanceTrack video clips in Fig.9 in Appendix E. For multi-object tracking in occlusion and non-linear motion, the results on DanceTrack are strong evidence of the effectiveness of OC-SORT.", + "bbox": [ + 75, + 672, + 468, + 852 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "KITTI. In Table 4 we report the results on the KITTI dataset. For a fair comparison, we adopt the detector weights by PermaTr [57] and report its performance in the", + "bbox": [ + 76, + 854, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "table as well. We run OC-SORT given the shared detections. As initializing SORT's track requires continuous tracking across several frames (\"minimum hits\"), we observe that the results not recorded during the track initialization make a significant difference. To address this problem, we perform offline head padding (HP) post-processing by writing these entries back after finishing the online tracking stage. The results of the car category on KITTI show an essential shortcoming of the default implementation version of OC-SORT that it chooses the IoU matching for the association. When the objects move fast or the frame rate is low, the IoU of bounding boxes between consecutive frames can be very low or even zero. This issue does not come from the intrinsic design of OC-SORT and is widely observed when using IoU as the association cue. Adding other cues [49, 72, 73] and appearance similarity [38, 63] have been demonstrated [63] effective to solve this. In contrast to the relatively inferior car tracking performance, OC-SORT improves pedestrian tracking performance to a new state-of-the-art. Using the same detections, OC-SORT achieves a large performance gap over PermaTr with $10\\mathrm{x}$ faster speed.", + "bbox": [ + 496, + 441, + 890, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The results on multiple benchmarks have demonstrated the effectiveness and efficiency of OC-SORT. We note that we use a shared parameter stack across datasets. Carefully tuning the parameters can probably further boost the performance. For example, the adaptive detection threshold is proven useful in previous work [70]. Besides the association accuracy, we also care about the inference speed. Given off-the-shelf detections, OC-SORT runs at 793 FPS on an Intel i9-9980XE CPU @ 3.00GHz. Therefore, OC-SORT", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "9692", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7ffd7bdbb19f9faf4509658e09cfa58b4a54afb411813b890f9e2e8d8c8c6e41.jpg", + "table_caption": [ + "Table 4. Results on KITTI-test. Our method uses the same detections as PermaTr [57]" + ], + "table_footnote": [], + "table_body": "
CarPedestrian
TrackerHOTA↑MOTA↑AssA↑IDs↓Frag↓HOTA↑MOTA↑AssA↑IDs↓Frag↓
IMMDP [65]68.6682.7569.76211181-----
SMAT [21]71.8883.6472.13198294-----
TrackMPNN [45]72.3087.3370.6348123739.4052.1035.45626669
MPNTrack [4]-----45.2646.2347.283971,078
CenterTr [73]73.0288.8371.1825422740.3553.8436.93425618
LGM [59]73.1487.6072.31448164-----
TuSimple [11]71.5586.3171.1129221845.8857.6147.62246651
PermaTr [57]77.4290.8577.6627527147.4365.0543.66483703
OC-SORT74.6487.8174.5225731852.9562.0057.81181598
OC-SORT + HP76.5490.2876.3925028054.6965.1459.08184609
", + "bbox": [ + 163, + 106, + 803, + 250 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/919728821b524f919edd360715daf917f4edf8e1619d10990f6e61b4a60b146c.jpg", + "table_caption": [ + "Table 5. Ablation on MOT17-val and DanceTrack-val." + ], + "table_footnote": [], + "table_body": "
MOT17-valDanceTrack-val
ORUOCMOCRHOTA↑AssA↑IDF1↑HOTA↑AssA↑IDF1↑
64.966.876.947.831.048.3
✓66.368.077.248.532.249.8
✓✓66.469.077.852.135.050.6
✓✓✓66.568.977.752.135.351.6
", + "bbox": [ + 80, + 289, + 462, + 375 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/1c821a059a12aa9fe1c7d0924727d1434f579bade38ee37fb0bbc6d72aafb627.jpg", + "table_caption": [ + "Table 6. Ablation on the trajectory hypothesis in ORU." + ], + "table_footnote": [], + "table_body": "
MOT17-valDanceTrack-val
HOTA↑AssA↑IDF1↑HOTA↑AssA↑IDF1↑
Const. Speed66.568.977.752.135.351.6
GPR63.165.275.749.533.749.6
Linear Regression64.366.576.049.333.449.2
Const. Acceleration66.267.977.451.334.850.9
", + "bbox": [ + 78, + 393, + 468, + 477 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/96f8a51f9d665cc96a5527ae69491c475340e714dadccc54b60e35cd9ad3d990.jpg", + "table_caption": [ + "Table 7. Influence from the value of ${\\Delta t}$ in OCM." + ], + "table_footnote": [], + "table_body": "
MOT17-valDanceTrack-val
HOTA↑AssA↑IDF1↑HOTA↑AssA↑IDF1↑
Δt = 166.167.576.951.334.351.3
Δt = 266.368.077.352.235.451.4
Δt = 366.568.977.752.135.351.6
Δt = 666.067.576.952.135.451.8
", + "bbox": [ + 89, + 500, + 450, + 583 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "can still run in an online and real-time fashion.", + "bbox": [ + 76, + 592, + 385, + 606 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 618, + 230, + 635 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Component Ablation. We ablate the contribution of proposed modules on the validation sets of MOT17 and DanceTrack in Table 5. The splitting of the MOT17 validation set follows a popular convention [73]. The results demonstrate the effectiveness of the proposed modules in OC-SORT. The results show that the performance gain from ORU is significant on both datasets but OCM only shows significant help on DanceTrack dataset where object motion is more complicated and the occlusion is heavy. The ablation study proves the effectiveness of our proposed method to improve tracking robustness in occlusion and non-linear motion.", + "bbox": [ + 75, + 642, + 468, + 808 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Virtual Trajectory in ORU. For simplicity, we follow the naive hypothesis of constant speed to generate a virtual trajectory in ORU. There are other alternatives like constant acceleration, regression-based fitting such as Linear Regression (LR) or Gaussian Process Regression (GPR), and Near Constant Acceleration Model (NCAM) [27]. The results of", + "bbox": [ + 75, + 809, + 470, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "comparing these choices are shown in Table 6. For GPR, we use the RBF kernel [10] $k(\\mathbf{x},\\mathbf{x}^{\\prime}) = \\exp \\left(-\\frac{||\\mathbf{x} - \\mathbf{x}^{\\prime}||^{2}}{50}\\right)$ . We provide more studies on the kernel configuration in Appendix B. The results show that local hypotheses such as Constant Speed/Acceleration perform much better than global hypotheses such as LR and GPR. This is probably because, as virtual trajectory generation happens in an online fashion, it is hard to get a reliable fit using only limited data points on historical time steps.", + "bbox": [ + 496, + 273, + 890, + 416 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "$\\Delta t$ in OCM. There is a trade-off when choosing the time difference $\\Delta t$ in OCM (Section 4). A large $\\Delta t$ decreases the noise of velocity estimation, but is also likely to discourage approximating object motion as linear. Therefore, we study the influence of varying $\\Delta t$ in Table 7. Our results agree with our analysis that increasing $\\Delta t$ from $\\Delta t = 1$ can boost performance. But increasing $\\Delta t$ higher than a best-practice value instead hurts the performance because of the difficulty of maintaining the approximation of linear motion.", + "bbox": [ + 496, + 417, + 892, + 554 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 563, + 617, + 579 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We analyze the popular motion-based tracker SORT and recognize its intrinsic limitations from using Kalman filter. These limitations significantly hurt tracking accuracy when the tracker fails to gain observations for supervision - likely caused by unreliable detectors, occlusion, or fast and non-linear target object motion. To address these issues, we propose Observation-Centric SORT (OC-SORT). OC-SORT is more robust to occlusion and non-linear object motion while keeping simple, online, and real-time. In our experiments on diverse datasets, OC-SORT significantly outperforms the state-of-the-art. The gain is especially significant for multi-object tracking under occlusion and non-linear object motion.", + "bbox": [ + 496, + 582, + 890, + 777 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. We thank David Held, Deva Ramanan, and Siddharth Ancha for the discussion about the theoretical modeling of OC-SORT. We thank Yuda Song for the help with the analysis of OCM error distribution. We also would like to thank Zhengyi Luo, Yuda Song, and Erica Weng for the detailed feedback on the paper writing. We also thank Yifu Zhang for sharing his experience with ByteTrack. This project was sponsored in part by NSF NRI award 2024173.", + "bbox": [ + 496, + 779, + 892, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "9693", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 106 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Yaakov Bar-Shalom, Fred Daum, and Jim Huang. The probabilistic data association filter. IEEE Control Systems Magazine, 29(6):82-100, 2009.", + "[2] Sumit Basu, Irfan Essa, and Alex Pentland. Motion regularization for model-based head tracking. In Proceedings of 13th International Conference on Pattern Recognition, volume 3, pages 611-616. IEEE, 1996.", + "[3] Alex Bewley, Zongyuan Ge, Lionel Ott, Fabio Ramos, and Ben Upcroft. Simple online and realtime tracking. In 2016 IEEE international conference on image processing (ICIP), pages 3464-3468. IEEE, 2016.", + "[4] Guillem Brasó and Laura Leal-Taixe. Learning a neural solver for multiple object tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6247-6257, 2020.", + "[5] Jiarui Cai, Mingze Xu, Wei Li, Yuanjun Xiong, Wei Xia, Zhuowen Tu, and Stefano Soatto. Memot: multi-object tracking with memory. In CVPR, 2022.", + "[6] Jinkun Cao, Jiangmiao Pang, Xinshuo Weng, Rawal Khirodkar, and Kris M Kitani. Object tracking by hierarchical part-whole attention.", + "[7] Jinkun Cao, Hongyang Tang, Hao-Shu Fang, Xiaoyong Shen, Cewu Lu, and Yu-Wing Tai. Cross-domain adaptation for animal pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9498-9507, 2019.", + "[8] Jinkun Cao, Hao Wu, and Kris Kitani. Track targets by dense spatio-temporal position encoding. arXiv preprint arXiv:2210.09455, 2022.", + "[9] Ming-Fang Chang, John Lambert, Patsorn Sangkloy, Jagjeet Singh, Slawomir Bak, Andrew Hartnett, De Wang, Peter Carr, Simon Lucey, Deva Ramanan, et al. Argoverse: 3d tracking and forecasting with rich maps. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8748-8757, 2019.", + "[10] Yin-Wen Chang, Cho-Jui Hsieh, Kai-Wei Chang, Michael Ringgaard, and Chih-Jen Lin. Training and testing low-degree polynomial data mappings via linearsvm. Journal of Machine Learning Research, 11(4), 2010.", + "[11] Wongun Choi. Near-online multi-target tracking with aggregated local flow descriptor. In Proceedings of the IEEE international conference on computer vision, pages 3029-3037, 2015.", + "[12] Peng Chu, Jiang Wang, Quanzeng You, Haibin Ling, and Zicheng Liu. Transmot: Spatial-temporal graph transformer for multiple object tracking. arXiv preprint arXiv:2104.00194, 2021.", + "[13] Peng Dai, Renliang Weng, Wongun Choi, Changshui Zhang, Zhangping He, and Wei Ding. Learning a proposal classifier for multiple object tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2443-2452, 2021.", + "[14] Patrick Dendorfer, Hamid Rezatofighi, Anton Milan, Javen Shi, Daniel Cremers, Ian Reid, Stefan Roth, Konrad Schindler, and Laura Leal-Taixe. Mot20: A benchmark for multi object tracking in crowded scenes. arXiv preprint arXiv:2003.09003, 2020.", + "[15] David Duvenaud. Automatic model construction with Gaus" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "sian processes. PhD thesis, University of Cambridge, 2014.", + "[16] Hao-Shu Fang, Jinkun Cao, Yu-Wing Tai, and Cewu Lu. Pairwise body-part attention for recognizing human-object interactions. In Proceedings of the European conference on computer vision (ECCV), pages 51-67, 2018.", + "[17] Juergen Gall, Angela Yao, Nima Razavi, Luc Van Gool, and Victor Lempitsky. Hough forests for object detection, tracking, and action recognition. IEEE transactions on pattern analysis and machine intelligence, 33(11):2188-2202, 2011.", + "[18] Damien Garreau, Wittawat Jitkrittum, and Motonobu Kanagawa. Large sample analysis of the median heuristic. arXiv preprint arXiv:1707.07269, 2017.", + "[19] Zheng Ge, Songtao Liu, Feng Wang, Zeming Li, and Jian Sun. Yolox: Exceeding yolo series in 2021. arXiv preprint arXiv:2107.08430, 2021.", + "[20] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. The International Journal of Robotics Research, 32(11):1231-1237, 2013.", + "[21] Nicolas Franco Gonzalez, Andres Ospina, and Philippe Calvez. Smat: Smart multiple affinity metrics for multiple object tracking. In International Conference on Image Analysis and Recognition, pages 48-62. Springer, 2020.", + "[22] Fredrik Gustafsson, Fredrik Gunnarsson, Niclas Bergman, Urban Forssell, Jonas Jansson, Rickard Karlsson, and P-J Nordlund. Particle filters for positioning, navigation, and tracking. IEEE Transactions on signal processing, 50(2):425-437, 2002.", + "[23] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016.", + "[24] David V Hinkley. On the ratio of two correlated normal random variables. Biometrika, 56(3):635-639, 1969.", + "[25] Andrea Hornakova, Roberto Henschel, Bodo Rosenhahn, and Paul Swoboda. Lifted disjoint paths with application in multiple object tracking. In International Conference on Machine Learning, pages 4364-4375. PMLR, 2020.", + "[26] Andrea Hornakova, Timo Kaiser, Paul Swoboda, Michal Rolinek, Bodo Rosenhahn, and Roberto Henschel. Making higher order mot scalable: An efficient approximate solver for lifted disjoint paths. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6330-6340, 2021.", + "[27] Andrew H Jazwinski. Stochastic processes and filtering theory. Courier Corporation, 2007.", + "[28] Simon J Julier and Jeffrey K Uhlmann. New extension of the kalman filter to nonlinear systems. In Signal processing, sensor fusion, and target recognition VI, volume 3068, pages 182-193. International Society for Optics and Photonics, 1997.", + "[29] Simon J Julier and Jeffrey K Uhlmann. Unscented filtering and nonlinear estimation. Proceedings of the IEEE, 92(3):401-422, 2004.", + "[30] Rudolf Emil Kalman et al. Contributions to the theory of optimal control. Bol. soc. mat. mexicana, 5(2):102-119, 1960.", + "[31] Kris M Kitani, Brian D Ziebart, James Andrew Bagnell, and Martial Hebert. Activity forecasting. In European conference on computer vision, pages 201-214. Springer, 2012." + ], + "bbox": [ + 503, + 92, + 890, + 888 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9694", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[32] Jonathan Ko and Dieter Fox. Gp-bayesfilters: Bayesian filtering using gaussian process prediction and observation models. Autonomous Robots, 27(1):75-90, 2009.", + "[33] Parth Kothari, Sven Kreiss, and Alexandre Alahi. Human trajectory forecasting in crowds: A deep learning perspective. IEEE Transactions on Intelligent Transportation Systems, 2021.", + "[34] Erich L Lehmann and George Casella. Theory of point estimation. Springer Science & Business Media, 2006.", + "[35] Wei Li, Yuanjun Xiong, Shuo Yang, Mingze Xu, Yongxin Wang, and Wei Xia. Semi-tcl: Semi-supervised track contrastive representation learning. arXiv preprint arXiv:2107.02396, 2021.", + "[36] Chao Liang, Zhipeng Zhang, Yi Lu, Xue Zhou, Bing Li, Xiyong Ye, and Jianxiao Zou. Rethinking the competition between detection and reid in multi-object tracking. arXiv preprint arXiv:2010.12138, 2020.", + "[37] Jonathon Luiten, Aljosa Osep, Patrick Dendorfer, Philip Torr, Andreas Geiger, Laura Leal-Taixe, and Bastian Leibe. Hota: A higher order metric for evaluating multi-object tracking. International journal of computer vision, 129(2):548-578, 2021.", + "[38] Gerard Maggiolino, Adnan Ahmad, Jinkun Cao, and Kris Kitani. Deep oc-sort: Multi-pedestrian tracking by adaptive re-identification. arXiv preprint arXiv:2302.11813, 2023.", + "[39] Tim Meinhardt, Alexander Kirillov, Laura Leal-Taixe, and Christoph Feichtenhofer. Trackformer: Multi-object tracking with transformers. arXiv preprint arXiv:2101.02702, 2021.", + "[40] Stan Melax, Leonid Keselman, and Sterling Orsten. Dynamics based 3d skeletal hand tracking. In Proceedings of the ACM SIGGRAPH Symposium on Interactive 3D Graphics and Games, pages 184-184, 2013.", + "[41] Anton Milan, Laura Leal-Taixe, Ian Reid, Stefan Roth, and Konrad Schindler. Mot16: A benchmark for multi-object tracking. arXiv preprint arXiv:1603.00831, 2016.", + "[42] Jiangmiao Pang, Linlu Qiu, Xia Li, Haofeng Chen, Qi Li, Trevor Darrell, and Fisher Yu. Quasi-dense similarity learning for multiple object tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 164-173, 2021.", + "[43] Dezhi Peng, Zikai Sun, Zirong Chen, Zirui Cai, Lele Xie, and Lianwen Jin. Detecting heads using feature refine net and cascaded multi-scale architecture. arXiv preprint arXiv:1803.09256, 2018.", + "[44] Lawrence Rabiner and Biinghwang Juang. An introduction to hidden markov models. *iiee assp magazine*, 3(1):4-16, 1986.", + "[45] Akshay Rangesh, Pranav Maheshwari, Mez Gebre, Siddhesh Mhatre, Vahid Ramezani, and Mohan M Trivedi. Trackmpnn: A message passing graph neural architecture for multi-object tracking. arXiv preprint arXiv:2101.04206, 2021.", + "[46] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 779-788, 2016.", + "[47] Steven Reece and Stephen Roberts. An introduction to gaussian processes for the kalman filter expert. In 2010 13th In-" + ], + "bbox": [ + 78, + 92, + 467, + 888 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ternational Conference on Information Fusion, pages 1-9. IEEE, 2010.", + "[48] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015.", + "[49] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 658-666, 2019.", + "[50] Toby Sharp, Cem Keskin, Duncan Robertson, Jonathan Taylor, Jamie Shotton, David Kim, Christoph Rhemann, Ido Leichter, Alon Vinnikov, Yichen Wei, et al. Accurate, robust, and flexible real-time hand tracking. In Proceedings of the 33rd annual ACM conference on human factors in computing systems, pages 3633-3642, 2015.", + "[51] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014.", + "[52] Gerald L Smith, Stanley F Schmidt, and Leonard A McGee. Application of statistical filter theory to the optimal estimation of position and velocity on board a circumlunar vehicle. National Aeronautics and Space Administration, 1962.", + "[53] Daniel Stadler and Jurgen Beyerer. Improving multiple pedestrian tracking by track management and occlusion handling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10958-10967, 2021.", + "[54] Peize Sun, Jinkun Cao, Yi Jiang, Zehuan Yuan, Song Bai, Kris Kitani, and Ping Luo. Dancetrack: Multi-object tracking in uniform appearance and diverse motion. arXiv preprint arXiv:2111.14690, 2021.", + "[55] Peize Sun, Jinkun Cao, Yi Jiang, Rufeng Zhang, Enze Xie, Zehuan Yuan, Changhu Wang, and Ping Luo. Transtrack: Multiple object tracking with transformer. arXiv preprint arXiv:2012.15460, 2020.", + "[56] Ramana Sundararaman, Cedric De Almeida Braga, Eric Marchand, and Julien Pette. Tracking pedestrian heads in dense crowd. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3865-3875, 2021.", + "[57] Pavel Tokmakov, Jie Li, Wolfram Burgard, and Adrien Gaidon. Learning to track with object permanence. In ICCV, pages 10860-10869, 2021.", + "[58] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017.", + "[59] Gaoang Wang, Renshu Gu, Zuozhu Liu, Weijie Hu, Mingli Song, and Jenq-Neng Hwang. Track without appearance: Learn box and tracklet embedding with local and global motion patterns for vehicle tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9876-9886, 2021.", + "[60] Shuai Wang, Hao Sheng, Yang Zhang, Yubin Wu, and Zhang Xiong. A general recurrent tracking framework without real data. In Proceedings of the IEEE/CVF International Confer-" + ], + "bbox": [ + 503, + 93, + 890, + 892 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "9695", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ence on Computer Vision, pages 13219-13228, 2021.", + "[61] Yongxin Wang, Kris Kitani, and Xinshuo Weng. Joint object detection and multi-object tracking with graph neural networks. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 13708-13715. IEEE, 2021.", + "[62] Christopher Williams and Carl Rasmussen. Gaussian processes for regression. Advances in neural information processing systems, 8, 1995.", + "[63] Nicolai Wojke, Alex Bewley, and Dietrich Paulus. Simple online and realtime tracking with a deep association metric. In 2017 IEEE international conference on image processing (ICIP), pages 3645-3649. IEEE, 2017.", + "[64] Jialian Wu, Jiale Cao, Liangchen Song, Yu Wang, Ming Yang, and Junsong Yuan. Track to detect and segment: An online multi-object tracker. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12352-12361, 2021.", + "[65] Yu Xiang, Alexandre Alahi, and Silvio Savarese. Learning to track: Online multi-object tracking by decision making. In Proceedings of the IEEE international conference on computer vision, pages 4705-4713, 2015.", + "[66] Yuliang Xiu, Jiefeng Li, Haoyu Wang, Yinghong Fang, and Cewu Lu. Pose flow: Efficient online pose tracking. arXiv preprint arXiv:1802.00977, 2018.", + "[67] Yihong Xu, Yutong Ban, Guillaume Delorme, Chuang Gan, Daniela Rus, and Xavier Alameda-Pineda. Transcenter: Transformers with dense queries for multiple-object tracking. arXiv preprint arXiv:2103.15145, 2021.", + "[68] Bin Yan, Yi Jiang, Peize Sun, Dong Wang, Zehuan Yuan, Ping Luo, and Huchuan Lu. Towards grand unification of object tracking. In ECCV. Springer, 2022.", + "[69] Fangao Zeng and et al. Motr: End-to-end multiple-object tracking with transformer. In ECCV. Springer, 2022.", + "[70] Yifu Zhang, Peize Sun, Yi Jiang, Dongdong Yu, Zehuan Yuan, Ping Luo, Wenyu Liu, and Xinggang Wang. Byetetrack: Multi-object tracking by associating every detection box. arXiv preprint arXiv:2110.06864, 2021.", + "[71] Yifu Zhang, Chunyu Wang, Xinggang Wang, Wenjun Zeng, and Wenyu Liu. Fairmot: On the fairness of detection and re-identification in multiple object tracking. International Journal of Computer Vision, 129(11):3069-3087, 2021.", + "[72] Zhaohui Zheng, Ping Wang, Wei Liu, Jinze Li, Rongguang Ye, and Dongwei Ren. Distance-iou loss: Faster and better learning for bounding box regression. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 12993-13000, 2020.", + "[73] Xingyi Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Tracking objects as points. In ECCV, pages 474-490. Springer, 2020.", + "[74] Xingyi Zhou, Dequan Wang, and Philipp Kähenbuhl. Objects as points. arXiv preprint arXiv:1904.07850, 2019.", + "[75] Xingyi Zhou, Tianwei Yin, Vladlen Koltun, and Philipp Krahenbuhl. Global tracking transformers. In CVPR, 2022." + ], + "bbox": [ + 78, + 92, + 468, + 806 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "9696", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Observation-Centric SORT_ Rethinking SORT for Robust Multi-Object Tracking/b4c4cc9d-52d9-4722-825a-050e1d27cf7f_model.json b/2023/Observation-Centric SORT_ Rethinking SORT for Robust Multi-Object Tracking/b4c4cc9d-52d9-4722-825a-050e1d27cf7f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a1086bd21e5f00b1e75d1462b77fba65ed1946be --- /dev/null +++ b/2023/Observation-Centric SORT_ Rethinking SORT for Robust Multi-Object Tracking/b4c4cc9d-52d9-4722-825a-050e1d27cf7f_model.json @@ -0,0 +1,2422 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.219, + 0.131, + 0.753, + 0.177 + ], + "angle": 0, + "content": "Observation-Centric SORT: Rethinking SORT for Robust Multi-Object Tracking" + }, + { + "type": "text", + "bbox": [ + 0.139, + 0.203, + 0.831, + 0.24 + ], + "angle": 0, + "content": "Jinkun Cao\\(^{1}\\) Jiangmiao Pang\\(^{2}\\) Xinshuo Weng\\(^{3}\\) Rawal Khirodkar\\(^{1}\\) Kris Kitani\\(^{1}\\) \n\\(^{1}\\) Carnegie Mellon University \\(^{2}\\) Shanghai AI Laboratory \\(^{3}\\) Nvidia" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.274, + 0.314, + 0.29 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.306, + 0.474, + 0.73 + ], + "angle": 0, + "content": "Kalman filter (KF) based methods for multi-object tracking (MOT) make an assumption that objects move linearly. While this assumption is acceptable for very short periods of occlusion, linear estimates of motion for prolonged time can be highly inaccurate. Moreover, when there is no measurement available to update Kalman filter parameters, the standard convention is to trust the priori state estimations for posteriori update. This leads to the accumulation of errors during a period of occlusion. The error causes significant motion direction variance in practice. In this work, we show that a basic Kalman filter can still obtain state-of-the-art tracking performance if proper care is taken to fix the noise accumulated during occlusion. Instead of relying only on the linear state estimate (i.e., estimation-centric approach), we use object observations (i.e., the measurements by object detector) to compute a virtual trajectory over the occlusion period to fix the error accumulation of filter parameters. This allows more time steps to correct errors accumulated during occlusion. We name our method Observation-Centric SORT (OC-SORT). It remains Simple, Online, and Real-Time but improves robustness during occlusion and non-linear motion. Given off-the-shelf detections as input, OC-SORT runs at \\(700+\\) FPS on a single CPU. It achieves state-of-the-art on multiple datasets, including MOT17, MOT20, KITTI, head tracking, and especially DanceTrack where the object motion is highly non-linear. The code and models are available at https://github.com/noahcao/OC_SORT." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.755, + 0.21, + 0.77 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.47, + 0.903 + ], + "angle": 0, + "content": "We aim to develop a motion model-based multi-object tracking (MOT) method that is robust to occlusion and nonlinear motion. Most existing motion model-based algorithms assume that the tracking targets have a constant velocity within a time interval, which is called the linear motion assumption. This assumption breaks in many practical scenarios, but it still works because when the time interval is small enough, the object's motion can be reasonably ap-" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.272, + 0.596, + 0.377 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.598, + 0.272, + 0.692, + 0.377 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.695, + 0.273, + 0.788, + 0.377 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.792, + 0.273, + 0.885, + 0.377 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.667, + 0.378, + 0.72, + 0.39 + ], + "angle": 0, + "content": "(a) SORT" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.392, + 0.595, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.598, + 0.392, + 0.691, + 0.494 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.695, + 0.392, + 0.788, + 0.494 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.792, + 0.392, + 0.885, + 0.494 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.617, + 0.496, + 0.769, + 0.509 + ], + "angle": 0, + "content": "(b) The proposed OC-SORT" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.521, + 0.895, + 0.582 + ], + "angle": 0, + "content": "Figure 1. Samples from the results on DanceTrack [54]. SORT and OC-SORT use the same detection results. On the third frame, SORT encounters an ID switch for the backflip target while OC-SORT tracks it consistently." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.612, + 0.894, + 0.718 + ], + "angle": 0, + "content": "proximated as linear. In this work, we are motivated by the fact that most of the errors from motion model-based tracking methods occur when occlusion and non-linear motion happen together. To mitigate the adverse effects caused, we first rethink current motion models and recognize some limitations. Then, we propose addressing them for more robust tracking performance, especially in occlusion." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.895, + 0.901 + ], + "angle": 0, + "content": "As the main branch of motion model-based tracking, filtering-based methods assume a transition function to predict the state of objects on future time steps, which are called state \"estimations\". Besides estimations, they leverage an observation model, such as an object detector, to derive the state measurements of target objects, also called \"observations\". Observations usually serve as auxiliary information to help update the posteriori parameters of the filter. The trajectories are still extended by the state estimations. Among this line of work, the most widely used one is SORT [3], which uses a Kalman filter (KF) to estimate object states and a linear motion function as the transition" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9686" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.151 + ], + "angle": 0, + "content": "function between time steps. However, SORT shows insufficient tracking robustness when the object motion is nonlinear, and no observations are available when updating the filter posteriori parameters." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.155, + 0.468, + 0.485 + ], + "angle": 0, + "content": "In this work, we recognize three limitations of SORT. First, although the high frame rate is the key to approximating the object motion as linear, it also amplifies the model's sensitivity to the noise of state estimations. Specifically, between consecutive frames of a high frame-rate video, we demonstrate that the noise of displacement of the object can be of the same magnitude as the actual object displacement, leading to the estimated object velocity by KF suffering from a significant variance. Also, the noise in the velocity estimate will accumulate into the position estimate by the transition process. Second, the noise of state estimations by KF is accumulated along the time when there is no observation available in the update stage of KF. We show that the error accumulates very fast with respect to the time of the target object's being untracked. The noise's influence on the velocity direction often makes the track lost again even after re-association. Last, given the development of modern detectors, the object state by detections usually has lower variance than the state estimations propagated along time steps by a fixed transition function in filters. However, SORT is designed to prolong the object trajectories by state estimations instead of observations." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.489, + 0.468, + 0.699 + ], + "angle": 0, + "content": "To relieve the negative effect of these limitations, we propose two main innovations in this work. First, we design a module to use object state observations to reduce the accumulated error during the track's being lost in a backcheck fashion. To be precise, besides the traditional stages of predict and update, we add a stage of re-update to correct the accumulated error. The re-update is triggered when a track is re-activated by associating to an observation after a period of being untracked. The re-update uses virtual observations on the historical time steps to prevent error accumulation. The virtual observations come from a trajectory generated using the last-seen observation before untracked and the latest observation re-activating this track as anchors. We name it Observation-centric Re-Update (ORU)." + }, + { + "type": "text", + "bbox": [ + 0.081, + 0.703, + 0.468, + 0.838 + ], + "angle": 0, + "content": "Besides ORU, the assumption of linear motion provides the consistency of the object motion direction. But this cue is hard to be used in SORT's association because of the heavy noise in direction estimation. But we propose an observation-centric manner to incorporate the direction consistency of tracks in the cost matrix for the association. We name it Observation-Centric Momentum (OCM). We also provide analytical justification for the noise of velocity direction estimation in practice." + }, + { + "type": "text", + "bbox": [ + 0.08, + 0.841, + 0.468, + 0.899 + ], + "angle": 0, + "content": "The proposed method, named as Observation-Centric SORT or OC-SORT in short, remains simple, online, real-time and significantly improves robustness over occlusion and non-linear motion. Our contributions are summarized" + }, + { + "type": "text", + "bbox": [ + 0.505, + 0.093, + 0.613, + 0.107 + ], + "angle": 0, + "content": "as the following:" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.112, + 0.89, + 0.171 + ], + "angle": 0, + "content": "1. We recognize, analytically and empirically, three limitations of SORT, i.e. sensitivity to the noise of state estimations, error accumulation over time, and being estimation-centric;" + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.175, + 0.89, + 0.234 + ], + "angle": 0, + "content": "2. We propose OC-SORT for tracking under occlusion and non-linear motion by fixing SORT's limitations. It achieves state-of-the-art performance on multiple datasets in an online and real-time fashion." + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.112, + 0.89, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.505, + 0.262, + 0.647, + 0.276 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.29, + 0.89, + 0.591 + ], + "angle": 0, + "content": "Motion Models. Many modern MOT algorithms [3, 11, 63, 70, 73] use motion models. Typically, these motion models use Bayesian estimation [34] to predict the next state by maximizing a posterior estimation. As one of the most classic motion models, Kalman filter (KF) [30] is a recursive Bayes filter that follows a typical predict-update cycle. The true state is assumed to be an unobserved Markov process, and the measurements are observations from a hidden Markov model [44]. Given that the linear motion assumption limits KF, follow-up works like Extended KF [52] and Unscented KF [28] were proposed to handle non-linear motion with first-order and third-order Taylor approximation. However, they still rely on approximating the Gaussian prior assumed by KF and require motion pattern assumption. On the other hand, particle filters [22] solve the non-linear motion by sampling-based posterior estimation but require exponential order of computation. Therefore, these variants of Kalman filter and particle filters are rarely adopted in the visual multi-object tracking and the mostly adopted motion model is still based on Kalman filter [3]." + }, + { + "type": "text", + "bbox": [ + 0.504, + 0.614, + 0.89, + 0.899 + ], + "angle": 0, + "content": "Multi-object Tracking. As a classic computer vision task, visual multi-object tracking is traditionally approached from probabilistic perspectives, e.g. joint probabilistic association [1]. And modern video object tracking is usually built upon modern object detectors [46, 48, 74]. SORT [3] adopts the Kalman filter for motion-based multi-object tracking given observations from deep detectors. DeepSORT [63] further introduces deep visual features [23, 51] into object association under the framework of SORT. Re-identification-based object association[42, 63, 71] has also become popular since then but falls short when scenes are crowded and objects are represented coarsely (e.g. enclosed by bounding boxes), or object appearance is not distinguishable. More recently, transformers [58] have been introduced to MOT [8, 39, 55, 69] to learn deep representations from both visual information and object trajectories. However, their performance still has a significant gap between state-of-the-art tracking-by-detection methods in terms of both accuracy and time efficiency." + }, + { + "type": "page_number", + "bbox": [ + 0.485, + 0.947, + 0.514, + 0.956 + ], + "angle": 0, + "content": "9687" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.09, + 0.383, + 0.107 + ], + "angle": 0, + "content": "3. Rethink the Limitations of SORT" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.115, + 0.47, + 0.178 + ], + "angle": 0, + "content": "In this section, we review Kalman filter and SORT [3]. We recognize some of their limitations, which are significant with occlusion and non-linear object motion. We are motivated to improve tracking robustness by fixing them." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.184, + 0.218, + 0.2 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.208, + 0.473, + 0.375 + ], + "angle": 0, + "content": "Kalman filter (KF) [30] is a linear estimator for dynamical systems discretized in the time domain. KF only requires the state estimations on the previous time step and the current measurement to estimate the target state on the next time step. The filter maintains two variables, the posteriori state estimate \\(\\mathbf{x}\\), and the posteriori estimate covariance matrix \\(\\mathbf{P}\\). In the task of object tracking, we describe the KF process with the state transition model \\(\\mathbf{F}\\), the observation model \\(\\mathbf{H}\\), the process noise \\(\\mathbf{Q}\\), and the observation noise \\(\\mathbf{R}\\). At each step \\(t\\), given observations \\(\\mathbf{z}_t\\), KF works in an alternation of predict and update stages:" + }, + { + "type": "equation", + "bbox": [ + 0.085, + 0.382, + 0.47, + 0.482 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\text {p r e d i c t} \\left\\{ \\begin{array}{l} \\hat {\\mathbf {x}} _ {t | t - 1} = \\mathbf {F} _ {t} \\hat {\\mathbf {x}} _ {t - 1 | t - 1} \\\\ \\mathbf {P} _ {t | t - 1} = \\mathbf {F} _ {t} \\mathbf {P} _ {t - 1 | t - 1} \\mathbf {F} _ {t} ^ {\\top} + \\mathbf {Q} _ {t} \\end{array} , \\right. \\\\ \\text {u p d a t e} \\left\\{ \\begin{array}{l} \\mathbf {K} _ {t} = \\mathbf {P} _ {t | t - 1} \\mathbf {H} _ {t} ^ {\\top} \\left(\\mathbf {H} _ {t} \\mathbf {P} _ {t | t - 1} \\mathbf {H} _ {t} ^ {\\top} + \\mathbf {R} _ {t}\\right) ^ {- 1} \\quad (1) \\\\ \\hat {\\mathbf {x}} _ {t | t} = \\hat {\\mathbf {x}} _ {t | t - 1} + \\mathbf {K} _ {t} (\\mathbf {z} _ {t} - \\mathbf {H} _ {t} \\hat {\\mathbf {x}} _ {t | t - 1}) \\\\ \\mathbf {P} _ {t | t} = (\\mathbf {I} - \\mathbf {K} _ {t} \\mathbf {H} _ {t}) \\mathbf {P} _ {t | t - 1} \\end{array} . \\right. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.489, + 0.47, + 0.578 + ], + "angle": 0, + "content": "The stage of predict is to derive the state estimations on the next time step \\( t \\). Given a measurement of target states on the next step \\( t \\), the stage of update aims to update the posteriori parameters in KF. Because the measurement comes from the observation model \\( \\mathbf{H} \\), it is also called \"observation\" in many scenarios." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.579, + 0.47, + 0.745 + ], + "angle": 0, + "content": "SORT [3] is a multi-object tracker built upon KF. The KF's state \\(\\mathbf{x}\\) in SORT is defined as \\(\\mathbf{x} = [u,v,s,r,\\dot{u},\\dot{v},\\dot{s}]^{\\top}\\), where \\((u,v)\\) is the 2D coordinates of the object center in the image. \\(s\\) is the bounding box scale (area) and \\(r\\) is the bounding box aspect ratio. The aspect ratio \\(r\\) is assumed to be constant. The other three variables, \\(\\dot{u},\\dot{v}\\) and \\(\\dot{s}\\) are the corresponding time derivatives. The observation is a bounding box \\(\\mathbf{z} = [u,v,w,h,c]^{\\top}\\) with object center position \\((u,v)\\), object width \\(w\\), and height \\(h\\) and the detection confidence \\(c\\) respectively. SORT assumes linear motion as the transition model \\(\\mathbf{F}\\) which leads to the state estimation as" + }, + { + "type": "equation", + "bbox": [ + 0.136, + 0.755, + 0.469, + 0.772 + ], + "angle": 0, + "content": "\\[\nu _ {t + 1} = u _ {t} + \\dot {u} _ {t} \\Delta t, \\quad v _ {t + 1} = v _ {t} + \\dot {v} _ {t} \\Delta t. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.78, + 0.469, + 0.87 + ], + "angle": 0, + "content": "To leverage KF (Eq 1) in SORT for visual MOT, the stage of predict corresponds to estimating the object position on the next video frame. And the observations used for the update stage usually come from a detection model. The update stage is to update Kalman filter parameters and does not directly edit the tracking outcomes." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.902 + ], + "angle": 0, + "content": "When the time difference between two steps is constant during the transition, \\(e.g.\\), the video frame rate is constant," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.244 + ], + "angle": 0, + "content": "we can set \\(\\Delta t = 1\\). When the video frame rate is high, SORT works well even when the object motion is non-linear globally, (e.g. dancing, fencing, wrestling) because the motion of the target object can be well approximated as linear within short time intervals. However, in practice, observations are often absent on some time steps, e.g. the target object is occluded in multi-object tracking. In such cases, we cannot update the KF parameters by the update operation as in Eq. 1 anymore. SORT uses the priori estimations directly as posterior. We call this \"dummy update\", namely" + }, + { + "type": "equation", + "bbox": [ + 0.597, + 0.255, + 0.892, + 0.273 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {x}} _ {t \\mid t} = \\hat {\\mathbf {x}} _ {t \\mid t - 1}, \\mathbf {P} _ {t \\mid t} = \\mathbf {P} _ {t \\mid t - 1}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.282, + 0.892, + 0.373 + ], + "angle": 0, + "content": "The philosophy behind such a design is to trust estimations when no observations are available to supervise them. We thus call the tracking algorithms following this scheme \"estimation-centric\". However, we will see that this estimation-centric mechanism can cause trouble when nonlinear motion and occlusion happen together." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.382, + 0.698, + 0.397 + ], + "angle": 0, + "content": "3.2. Limitations of SORT" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.405, + 0.892, + 0.451 + ], + "angle": 0, + "content": "In this section, we identify three main limitations of SORT which are connected. This analysis lays the foundation of our proposed method." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.461, + 0.715, + 0.476 + ], + "angle": 0, + "content": "3.2.1 Sensitive to State Noise" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.484, + 0.892, + 0.591 + ], + "angle": 0, + "content": "Now we show that SORT is sensitive to the noise from KF's state estimations. To begin with, we assume that the estimated object center position follows \\( u \\sim \\mathcal{N}(\\mu_u, \\sigma_u^2) \\) and \\( v \\sim \\mathcal{N}(\\mu_v, \\sigma_v^2) \\), where \\( (\\mu_u, \\mu_v) \\) is the underlying true position. Then, if we assume that the state noises are independent on different steps, by Eq.2, the object speed between two time steps, \\( t \\longrightarrow t + \\Delta t \\), is" + }, + { + "type": "equation", + "bbox": [ + 0.554, + 0.602, + 0.892, + 0.63 + ], + "angle": 0, + "content": "\\[\n\\dot {u} = \\frac {u _ {t + \\Delta t} - u _ {t}}{\\Delta t}, \\quad \\dot {v} = \\frac {v _ {t + \\Delta t} - v _ {t}}{\\Delta t}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.633, + 0.892, + 0.75 + ], + "angle": 0, + "content": "making the noise of estimated speed \\(\\delta_{\\dot{u}}\\sim \\mathcal{N}(0,\\frac{2\\sigma_u^2}{(\\Delta t)^2})\\) \\(\\delta_{\\dot{v}}\\sim \\mathcal{N}(0,\\frac{2\\sigma_v^2}{(\\Delta t)^2})\\) . Therefore, a small \\(\\Delta t\\) will amplify the noise. This suggests that SORT will suffer from the heavy noise of velocity estimation on high-frame-rate videos. The analysis above is simplified from the reality. In practice, velocity won't be determined by the state on future time steps. For a more strict analysis, please refer to Appendix G." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Moreover, for most multi-object tracking scenarios, the target object displacement is only a few pixels between consecutive frames. For instance, the average displacement is 1.93 pixels and 0.65 pixels along the image width and height for the MOT17 [41] training dataset. In such a case, even if the estimated position has a shift of only a single pixel, it causes a significant variation in the estimated speed. In general, the variance of the speed estimation can be of the same magnitude as the speed itself or even greater. This will not make a massive impact as the shift is only of few pixels" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9688" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.129, + 0.089, + 0.852, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.334, + 0.895, + 0.396 + ], + "angle": 0, + "content": "Figure 2. The pipeline of our proposed OC-SORT. The red boxes are detections, orange boxes are active tracks, blue boxes are untracked tracks, and dashed boxes are the estimates from KF. During association, OCM is used to add the velocity consistency cost. The target #1 is lost on the frame t+1 because of occlusion. But on the next frame, it is recovered by referring to its observation of the frame t by OCR. It being re-tracked triggers ORU from t to t+2 for the parameters of its KF." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.411, + 0.471, + 0.517 + ], + "angle": 0, + "content": "from the ground truth on the next time step and the observations, whose variance is independent of the time, will be able to fix the noise when updating the posteriori parameters. However, we find that such a high sensitivity to state noise introduces significant problems in practice after being amplified by the error accumulation across multiple time steps when no observation is available for KF update." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.52, + 0.345, + 0.535 + ], + "angle": 0, + "content": "3.2.2 Temporal Error Magnification" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.543, + 0.47, + 0.786 + ], + "angle": 0, + "content": "For analysis above in Eq. 4, we assume the noise of the object state is i.i.d on different time steps (this is a simplified version, a more detailed analysis is provided in Appendix G). This is reasonable for object detections but not for the estimations from KF. This is because KF's estimations always rely on its estimations on previous time steps. The effect is usually minor because KF can use observation in update to prevent the posteriori state estimation and covariance, i.e. \\(\\hat{\\mathbf{x}}_{t|t}\\) and \\(\\mathbf{P}_{t|t}\\), deviating from the true value too far away. However, when no observations are provided to KF, it cannot use observation to update its parameters. Then it has to follow Eq. 3 to prolong the estimated trajectory to the next time step. Consider a track is occluded on the time steps between \\(t\\) and \\(t + T\\) and the noise of speed estimate follows \\(\\delta_{\\dot{u}_t} \\sim \\mathcal{N}(0,2\\sigma_u^2)\\), \\(\\delta_{\\dot{v}_t} \\sim \\mathcal{N}(0,2\\sigma_v^2)\\) for SORT. On the step \\(t + T\\), state estimation would be" + }, + { + "type": "equation", + "bbox": [ + 0.132, + 0.792, + 0.469, + 0.807 + ], + "angle": 0, + "content": "\\[\nu _ {t + T} = u _ {t} + T \\dot {u} _ {t}, \\quad v _ {t + T} = v _ {t} + T \\dot {v} _ {t}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.471, + 0.902 + ], + "angle": 0, + "content": "whose noise follows \\(\\delta_{u_{t + T}}\\sim \\mathcal{N}(0,2T^2\\sigma_u^2)\\) and \\(\\delta_{v_{t + T}}\\sim\\) \\(\\mathcal{N}(0,2T^2\\sigma_v^2)\\). So without the observations, the estimation from the linear motion assumption of KF results in a fast error accumulation with respect to time. Given \\(\\sigma_v\\) and \\(\\sigma_{u}\\) is of the same magnitude as object displacement between consecutive frames, the noise of final object position" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.41, + 0.892, + 0.516 + ], + "angle": 0, + "content": "\\((u_{t + T}, v_{t + T})\\) is of the same magnitude as the object size. For instance, the size of pedestrians close to the camera on MOT17 is around \\(50 \\times 300\\) pixels. So even assuming the variance of position estimation is only 1 pixel, 10-frame occlusion can accumulate a shift in final position estimation as large as the object size. Such error magnification leads to a major accumulation of errors when the scenes are crowded." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.522, + 0.691, + 0.536 + ], + "angle": 0, + "content": "3.2.3 Estimation-Centric" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.541, + 0.892, + 0.798 + ], + "angle": 0, + "content": "The aforementioned limitations come from a fundamental property of SORT that it follows KF to be estimation-centric. It allows update without the existence of observations and purely trusts the estimations. A key difference between state estimations and observations is that we can assume that the observations by an object detector in each frame are affected by i.i.d. noise \\(\\delta_{\\mathbf{z}} \\sim \\mathcal{N}(0, \\sigma'^2)\\) while the noise in state estimations can be accumulated along the hidden Markov process. Moreover, modern object detectors use powerful object visual features [48, 51]. It makes that, even on a single frame, it is usually safe to assume \\(\\sigma' < \\sigma_u\\) and \\(\\sigma' < \\sigma_v\\) because the object localization is more accurate by detection than from the state estimations through linear motion assumption. Combined with the previously mentioned two limitations, being estimation-centric makes SORT suffer from heavy noise when there is occlusion and the object motion is not perfectly linear." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.815, + 0.757, + 0.83 + ], + "angle": 0, + "content": "4. Observation-Centric SORT" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.841, + 0.894, + 0.902 + ], + "angle": 0, + "content": "In this section, we introduce the proposed Observation-Centric SORT (OC-SORT). To address the limitations of SORT discussed above, we use the momentum of the object moving into the association stage and develop a" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9689" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.16, + 0.089, + 0.812, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.243, + 0.895, + 0.306 + ], + "angle": 0, + "content": "Figure 3. Example of how Observation-centric Re-Update (ORU) reduces the error accumulation when a track is broken. The target is occluded between the second and the third time step and the tracker finds it back at the third step. Yellow boxes are the state observations by the detector. White stars are the estimated centers without ORU. Yellow stars are the estimated centers fixed by ORU. The gray star on the fourth step is the estimated center without ORU and fails to match observations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.319, + 0.475, + 0.472 + ], + "angle": 0, + "content": "pipeline with less noise and more robustness over occlusion and non-linear motion. The key is to design the tracker as observation-centric instead of estimation-centric. If a track is recovered from being untracked, we use an Observation-centric Re-Update (ORU) strategy to counter the accumulated error during the untracked period. OC-SORT also adds an Observation-Centric Momentum (OCM) term in the association cost. Please refer to Algorithm 1 in Appendix for the pseudo-code of OC-SORT. The pipeline is shown in Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.478, + 0.411, + 0.495 + ], + "angle": 0, + "content": "4.1. Observation-centric Re-Update (ORU)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.501, + 0.474, + 0.73 + ], + "angle": 0, + "content": "In practice, even if an object can be associated again by SORT after a period of being untracked, it is probably lost again because its KF parameters have already deviated far away from the correct due to the temporal error magnification. To alleviate this problem, we propose Observation-centric Re-Update (ORU) to reduce the accumulated error. Once a track is associated with an observation again after a period of being untracked (\"re-activation\"), we backcheck the period of its being lost and re-update the parameters of KF. The re-update is based on \"observations\" from a virtual trajectory. The virtual trajectory is generated referring to the observations on the steps starting and ending the untracked period. For example, by denoting the last-seen observation before being untracked as \\(\\mathbf{z}_{t_1}\\) and the observation triggering the re-association as \\(\\mathbf{z}_{t_2}\\), the virtual trajectory is denoted as" + }, + { + "type": "equation", + "bbox": [ + 0.138, + 0.743, + 0.47, + 0.76 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathbf {z}} _ {t} = \\operatorname {T r a j} _ {\\text {v i r t u a l}} \\left(\\mathbf {z} _ {t _ {1}}, \\mathbf {z} _ {t _ {2}}, t\\right), t _ {1} < t < t _ {2}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.764, + 0.47, + 0.795 + ], + "angle": 0, + "content": "Then, along the trajectory of \\(\\tilde{\\mathbf{z}}_t(t_1 < t < t_2)\\), we run the loop of predict and re-update. The re-update operation is" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.801, + 0.47, + 0.87 + ], + "angle": 0, + "content": "\\[\nr e - u p d a t e \\left\\{ \\begin{array}{l} \\mathbf {K} _ {t} = \\mathbf {P} _ {t | t - 1} \\mathbf {H} _ {t} ^ {\\top} \\left(\\mathbf {H} _ {t} \\mathbf {P} _ {t | t - 1} \\mathbf {H} _ {t} ^ {\\top} + \\mathbf {R} _ {t}\\right) ^ {- 1} \\\\ \\hat {\\mathbf {x}} _ {t | t} = \\hat {\\mathbf {x}} _ {t | t - 1} + \\mathbf {K} _ {t} \\left(\\tilde {\\mathbf {z}} _ {t} - \\mathbf {H} _ {t} \\hat {\\mathbf {x}} _ {t | t - 1}\\right) \\\\ \\mathbf {P} _ {t | t} = (\\mathbf {I} - \\mathbf {K} _ {t} \\mathbf {H} _ {t}) \\mathbf {P} _ {t | t - 1} \\end{array} \\right. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.471, + 0.902 + ], + "angle": 0, + "content": "As the observations on the virtual trajectory match the motion pattern anchored by the last-seen and the latest asso" + }, + { + "type": "image", + "bbox": [ + 0.557, + 0.32, + 0.851, + 0.4 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.411, + 0.895, + 0.504 + ], + "angle": 0, + "content": "Figure 4. Calculation of motion direction difference in OCM. The green line indicates an existing track and the dots are the observations on it. The red dots are the new observations to be associated. The blue link and the yellow link form the directions of \\(\\theta^{\\mathrm{track}}\\) and \\(\\theta^{\\mathrm{intention}}\\) respectively. The included angle is the difference of direction \\(\\Delta \\theta\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.514, + 0.895, + 0.606 + ], + "angle": 0, + "content": "ciated real observations, the update will not suffer from the error accumulated through the dummy update anymore. We call the proposed process Observation-centric Re-Update. It serves as an independent stage outside the predict-update loop and is triggered only a track is re-activated from a period of having no observations." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.618, + 0.85, + 0.635 + ], + "angle": 0, + "content": "4.2. Observation-Centric Momentum (OCM)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.643, + 0.895, + 0.883 + ], + "angle": 0, + "content": "In a reasonably short time interval, we can approximate the motion as linear. And the linear motion assumption also asks for consistent motion direction. But the noise prevents us from leveraging the consistency of direction. To be precise, to determine the motion direction, we need the object state on two steps with a time difference \\(\\Delta t\\). If \\(\\Delta t\\) is small, the velocity noise would be significant because of the estimation's sensitivity to state noise. If \\(\\Delta t\\) is big, the noise of direction estimation can also be significant because of the temporal error magnification and the failure of linear motion assumption. As state observations have no problem of temporal error magnification that state estimations suffer from, we propose to use observations instead of estimations to reduce the noise of motion direction calculation and introduce the term of velocity consistency to help the association." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "With the new term, given \\(N\\) existing tracks and \\(M\\) de" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.958 + ], + "angle": 0, + "content": "9690" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.121 + ], + "angle": 0, + "content": "tectons on the new-coming time step, the association cost matrix is formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.131, + 0.469, + 0.15 + ], + "angle": 0, + "content": "\\[\nC (\\hat {\\mathbf {X}}, \\mathbf {Z}) = C _ {\\mathrm {I o U}} (\\hat {\\mathbf {X}}, \\mathbf {Z}) + \\lambda C _ {v} (\\mathcal {Z}, \\mathbf {Z}), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.159, + 0.469, + 0.356 + ], + "angle": 0, + "content": "where \\(\\hat{\\mathbf{X}}\\in \\mathbb{R}^{N\\times 7}\\) is the set of object state estimations and \\(\\mathbf{Z}\\in \\mathbb{R}^{M\\times 5}\\) is the set of observations on the new time step. \\(\\lambda\\) is a weighting factor. \\(\\mathcal{Z}\\) contains the trajectory of observations of all existing tracks. \\(C_{\\mathrm{IoU}}(\\cdot ,\\cdot)\\) calculates the negative pairwise IoU (Intersection over Union) and \\(C_v(\\cdot ,\\cdot)\\) calculates the consistency between the directions of i) linking two observations on an existing track \\((\\theta^{\\mathrm{track}})\\) and ii) linking a track's historical observation and a new observation \\((\\theta^{\\mathrm{intention}})\\). \\(C_v\\) contains all pairs of \\(\\Delta \\theta = |\\theta^{\\mathrm{track}} - \\theta^{\\mathrm{intention}}|\\). In our implementation, we calculate the motion direction in radians, namely \\(\\theta = \\arctan \\left(\\frac{v_1 - v_2}{u_1 - u_2}\\right)\\) where \\((u_{1},v_{1})\\) and \\((u_{2},v_{2})\\) are the observations on two different time steps. The calculation is also illustrated in Figure 4." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.357, + 0.469, + 0.537 + ], + "angle": 0, + "content": "Following the assumptions of noise distribution mentioned before, we can derive a closed-form probability density function of the distribution of the noise in the direction estimation. The derivation is explained in detail in Appendix A. By analyzing the property of this distribution, we reach a conclusion that, under the linear-motion model, the scale of the noise of direction estimation is negatively correlated to the time difference between the two observation points, i.e. \\(\\Delta t\\). This suggests increasing \\(\\Delta t\\) to achieve a low-noisy estimation of \\(\\theta\\). However, the assumption of linear motion typically holds only when \\(\\Delta t\\) is small enough. Therefore, the choice of \\(\\Delta t\\) requires a trade-off." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.538, + 0.469, + 0.658 + ], + "angle": 0, + "content": "Besides ORU and OCM, we also find it empirically helpful to check a track's last presence to recover it from being lost. We thus apply a heuristic Observation-Centric Recovery (OCR) technique. OCR will start a second attempt of associating between the last observation of unmatched tracks to the unmatched observations after the usual association stage. It can handle the case of an object stopping or being occluded for a short time interval." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.672, + 0.209, + 0.688 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.697, + 0.267, + 0.714 + ], + "angle": 0, + "content": "5.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Datasets. We evaluate our method on multiple multi-object tracking datasets including MOT17 [41], MOT20 [14], KITTI [20], DanceTrack [54] and CroHD [56]. MOT17 [41] and MOT20 [14] are for pedestrian tracking, where targets mostly move linearly, while scenes in MOT20 are more crowded. KITTI [20] is for pedestrian and car tracking with a relatively low frame rate of 10FPS. CroHD is a dataset for head tracking in the crowd and the results on it are included in the appendix. DanceTrack [54] is a recently proposed dataset for human tracking. For the data in DanceTrack, object localization is easy, but the object motion is highly non-linear. Furthermore, the objects" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "have a close appearance, severe occlusion, and frequent crossovers. Considering our goal is to improve tracking robustness under occlusion and non-linear object motion, we would emphasize the comparison on DanceTrack." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.153, + 0.893, + 0.394 + ], + "angle": 0, + "content": "Implementations. For a fair comparison, we directly apply the object detections from existing baselines. For MOT17, MOT20, and DanceTrack, we use the publicly availableYOLOX [19] detector weights by ByteTrack [70]. For KITTI [20], we use the detections from PermaTrack [57] publicly available in the official release1. For ORU, we generate the virtual trajectory during occlusion with the constant-velocity assumption. Therefore, Eq. 6 is adopted as \\(\\tilde{\\mathbf{z}}_t = \\mathbf{z}_{t_1} + \\frac{t - t_1}{t_2 - t_1} (\\mathbf{z}_{t_2} - \\mathbf{z}_{t_1}), t_1 < t < t_2\\). For OCM, the velocity direction is calculated using the observations three time steps apart, i.e. \\(\\Delta t = 3\\). The direction difference is measured by the absolute difference of angles in radians. We set \\(\\lambda = 0.2\\) in Eq. 8. Following the common practice of SORT, we set the detection confidence threshold at 0.4 for MOT20 and 0.6 for other datasets. The IoU threshold during association is 0.3." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.396, + 0.893, + 0.533 + ], + "angle": 0, + "content": "Metrics. We adopt HOTA [37] as the main metric as it maintains a proper balance between the accuracy of object detection and association [37]. We also emphasize AssA to evaluate the association performance. IDF1 is also used for association performance evaluation. Other metrics we report, such as MOTA, are highly related to detection performance. It is fair to use these metrics only when all methods use the same detections for tracking, which is referred to as \"public tracking\" as reported in Appendix C." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.545, + 0.687, + 0.56 + ], + "angle": 0, + "content": "5.2. Benchmark Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.57, + 0.892, + 0.614 + ], + "angle": 0, + "content": "Here we report the benchmark results on multiple datasets. We put all methods that use the shared detection results in the blue blocks at the bottom of each table." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.616, + 0.893, + 0.873 + ], + "angle": 0, + "content": "MOT17 and MOT20. We report OC-SORT's performance on MOT17 and MOT20 in Table 1 and Table 2 using private detections. To make a fair comparison, we use the same detection as ByteTrack [70]. OC-SORT achieves performance comparable to other state-of-the-art methods. Our gains are especially significant in MOT20 under severe pedestrian occlusion, setting a state-of-the-art HOTA of 62.1. As our method is designed to be simple for better generalization, we do not use adaptive detection thresholds as in ByteTrack. Also, ByteTrack uses more detections of low-confidence to achieve higher MOTA scores but we keep the detection confidence threshold the same as on other datasets, which is the common practice in the community. We inherit the linear interpolation on the two datasets by baseline methods for a fair comparison. To more clearly discard the variance from the detector, we also perform public tracking on MOT17 and MOT20, which is reported in Table 12 and Table 13 in" + }, + { + "type": "page_footnote", + "bbox": [ + 0.518, + 0.887, + 0.738, + 0.901 + ], + "angle": 0, + "content": "1 https://github.com/TRI-ML/permatrack/" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "9691" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.144, + 0.089, + 0.825, + 0.105 + ], + "angle": 0, + "content": "Table 1. Results on MOT17-test with the private detections. ByteTrack and OC-SORT share detections." + }, + { + "type": "table", + "bbox": [ + 0.168, + 0.107, + 0.803, + 0.277 + ], + "angle": 0, + "content": "
TrackerHOTA↑MOTA↑IDF1↑FP(104)↓FN(104)↓IDs↓Frag↓AssA↑AssR↑
FairMOT [71]59.373.772.32.7511.73,3038,07358.063.6
TransCt [67]54.573.262.22.3112.44,6149,51949.754.2
TransTrk [55]54.175.263.55.028.643,6034,87247.957.1
GRTU [60]62.074.975.03.2010.81,8121,82462.165.8
QDTrack [42]53.968.766.32.6614.73,3788,09152.757.2
MOTR [69]57.271.968.42.1113.62,1153,89755.859.2
PermaTr [57]55.573.868.92.9011.53,6996,13253.159.8
TransMOT [12]61.776.775.13.629.322,3467,71959.966.5
GTR [75]59.175.371.52.6811.02,859-61.6-
DST-Tracker [8]60.175.272.32.4211.02,729-62.1-
MeMOT [5]56.972.569.02.7211.52,724-55.2-
UniCom [68]61.777.275.55.017.335,379---
ByteTrack [70]63.180.377.32.558.372,1962,27762.068.2
OC-SORT63.278.077.51.5110.81,9502,04063.267.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.157, + 0.289, + 0.813, + 0.304 + ], + "angle": 0, + "content": "Table 2. Results on MOT20-test with private detections. ByteTrack and OC-SORT share detections." + }, + { + "type": "table", + "bbox": [ + 0.164, + 0.306, + 0.812, + 0.424 + ], + "angle": 0, + "content": "
TrackerHOTA↑MOTA↑IDF1↑FP(104)↓FN(104)↓IDs↓Frag↓AssA↑AssR↑
FairMOT [71]54.661.867.310.38.895,2437,87454.760.7
TransCt [67]43.558.549.66.4214.64,6959,58137.045.1
Semi-TCL [35]55.365.270.16.1211.54,1398,50856.360.9
CSTrack [36]54.066.668.62.5414.43,1967,63254.057.6
GSDT [61]53.667.167.53.1913.53,1319,87552.758.5
TransMOT [12]61.977.575.23.428.081,6152,42160.166.3
MeMOT [5]54.163.766.14.7913.81,938-55.0-
ByteTrack [70]61.377.875.22.628.761,2231,46059.666.2
OC-SORT62.175.575.91.8010.89131,19862.067.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.441, + 0.47, + 0.469 + ], + "angle": 0, + "content": "Table 3. Results on DanceTrack test set. Methods in the blue block share the same detections." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.473, + 0.472, + 0.63 + ], + "angle": 0, + "content": "
TrackerHOTA↑DetA↑AssA↑MOTA↑IDF1↑
CenterTrack [73]41.878.122.686.835.7
FairMOT [71]39.766.723.882.240.8
QDTrack [42]45.772.129.283.044.8
TransTrk[55]45.575.927.588.445.2
TraDes [64]43.374.525.486.241.2
MOTR [69]54.273.540.279.751.5
GTR [75]48.072.531.984.750.3
DST-Tracker [8]51.972.334.684.951.0
SORT [3]47.972.031.291.850.8
DeepSORT [63]45.671.029.787.847.9
ByteTrack [70]47.371.631.489.552.5
OC-SORT54.680.440.289.654.6
OC-SORT + Linear Interp55.180.440.492.254.9
" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.642, + 0.469, + 0.672 + ], + "angle": 0, + "content": "Appendix C. OC-SORT still outperforms the existing state-of-the-art in public tracking settings." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.673, + 0.469, + 0.853 + ], + "angle": 0, + "content": "DanceTrack. To evaluate OC-SORT under challenging non-linear object motion, we report results on the DanceTrack in Table 3. OC-SORT sets a new state-of-the-art, outperforming the baselines by a great margin under nonlinear object motions. We compare the tracking results of SORT and OC-SORT under extreme non-linear situations in Fig.1 and more samples are available in Fig.8 in Appendix E. We also visualize the output trajectories by OC-SORT and SORT on randomly selected DanceTrack video clips in Fig.9 in Appendix E. For multi-object tracking in occlusion and non-linear motion, the results on DanceTrack are strong evidence of the effectiveness of OC-SORT." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "KITTI. In Table 4 we report the results on the KITTI dataset. For a fair comparison, we adopt the detector weights by PermaTr [57] and report its performance in the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.442, + 0.892, + 0.76 + ], + "angle": 0, + "content": "table as well. We run OC-SORT given the shared detections. As initializing SORT's track requires continuous tracking across several frames (\"minimum hits\"), we observe that the results not recorded during the track initialization make a significant difference. To address this problem, we perform offline head padding (HP) post-processing by writing these entries back after finishing the online tracking stage. The results of the car category on KITTI show an essential shortcoming of the default implementation version of OC-SORT that it chooses the IoU matching for the association. When the objects move fast or the frame rate is low, the IoU of bounding boxes between consecutive frames can be very low or even zero. This issue does not come from the intrinsic design of OC-SORT and is widely observed when using IoU as the association cue. Adding other cues [49, 72, 73] and appearance similarity [38, 63] have been demonstrated [63] effective to solve this. In contrast to the relatively inferior car tracking performance, OC-SORT improves pedestrian tracking performance to a new state-of-the-art. Using the same detections, OC-SORT achieves a large performance gap over PermaTr with \\(10\\mathrm{x}\\) faster speed." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The results on multiple benchmarks have demonstrated the effectiveness and efficiency of OC-SORT. We note that we use a shared parameter stack across datasets. Carefully tuning the parameters can probably further boost the performance. For example, the adaptive detection threshold is proven useful in previous work [70]. Besides the association accuracy, we also care about the inference speed. Given off-the-shelf detections, OC-SORT runs at 793 FPS on an Intel i9-9980XE CPU @ 3.00GHz. Therefore, OC-SORT" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9692" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.2, + 0.09, + 0.768, + 0.105 + ], + "angle": 0, + "content": "Table 4. Results on KITTI-test. Our method uses the same detections as PermaTr [57]" + }, + { + "type": "table", + "bbox": [ + 0.164, + 0.107, + 0.805, + 0.25 + ], + "angle": 0, + "content": "
CarPedestrian
TrackerHOTA↑MOTA↑AssA↑IDs↓Frag↓HOTA↑MOTA↑AssA↑IDs↓Frag↓
IMMDP [65]68.6682.7569.76211181-----
SMAT [21]71.8883.6472.13198294-----
TrackMPNN [45]72.3087.3370.6348123739.4052.1035.45626669
MPNTrack [4]-----45.2646.2347.283971,078
CenterTr [73]73.0288.8371.1825422740.3553.8436.93425618
LGM [59]73.1487.6072.31448164-----
TuSimple [11]71.5586.3171.1129221845.8857.6147.62246651
PermaTr [57]77.4290.8577.6627527147.4365.0543.66483703
OC-SORT74.6487.8174.5225731852.9562.0057.81181598
OC-SORT + HP76.5490.2876.3925028054.6965.1459.08184609
" + }, + { + "type": "table_caption", + "bbox": [ + 0.092, + 0.274, + 0.454, + 0.288 + ], + "angle": 0, + "content": "Table 5. Ablation on MOT17-val and DanceTrack-val." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.29, + 0.464, + 0.375 + ], + "angle": 0, + "content": "
MOT17-valDanceTrack-val
ORUOCMOCRHOTA↑AssA↑IDF1↑HOTA↑AssA↑IDF1↑
64.966.876.947.831.048.3
✓66.368.077.248.532.249.8
✓✓66.469.077.852.135.050.6
✓✓✓66.568.977.752.135.351.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.092, + 0.378, + 0.456, + 0.393 + ], + "angle": 0, + "content": "Table 6. Ablation on the trajectory hypothesis in ORU." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.395, + 0.47, + 0.478 + ], + "angle": 0, + "content": "
MOT17-valDanceTrack-val
HOTA↑AssA↑IDF1↑HOTA↑AssA↑IDF1↑
Const. Speed66.568.977.752.135.351.6
GPR63.165.275.749.533.749.6
Linear Regression64.366.576.049.333.449.2
Const. Acceleration66.267.977.451.334.850.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.11, + 0.483, + 0.436, + 0.497 + ], + "angle": 0, + "content": "Table 7. Influence from the value of \\( {\\Delta t} \\) in OCM." + }, + { + "type": "table", + "bbox": [ + 0.09, + 0.5, + 0.452, + 0.584 + ], + "angle": 0, + "content": "
MOT17-valDanceTrack-val
HOTA↑AssA↑IDF1↑HOTA↑AssA↑IDF1↑
Δt = 166.167.576.951.334.351.3
Δt = 266.368.077.352.235.451.4
Δt = 366.568.977.752.135.351.6
Δt = 666.067.576.952.135.451.8
" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.593, + 0.387, + 0.607 + ], + "angle": 0, + "content": "can still run in an online and real-time fashion." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.619, + 0.231, + 0.636 + ], + "angle": 0, + "content": "5.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.643, + 0.47, + 0.809 + ], + "angle": 0, + "content": "Component Ablation. We ablate the contribution of proposed modules on the validation sets of MOT17 and DanceTrack in Table 5. The splitting of the MOT17 validation set follows a popular convention [73]. The results demonstrate the effectiveness of the proposed modules in OC-SORT. The results show that the performance gain from ORU is significant on both datasets but OCM only shows significant help on DanceTrack dataset where object motion is more complicated and the occlusion is heavy. The ablation study proves the effectiveness of our proposed method to improve tracking robustness in occlusion and non-linear motion." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Virtual Trajectory in ORU. For simplicity, we follow the naive hypothesis of constant speed to generate a virtual trajectory in ORU. There are other alternatives like constant acceleration, regression-based fitting such as Linear Regression (LR) or Gaussian Process Regression (GPR), and Near Constant Acceleration Model (NCAM) [27]. The results of" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.275, + 0.892, + 0.417 + ], + "angle": 0, + "content": "comparing these choices are shown in Table 6. For GPR, we use the RBF kernel [10] \\( k(\\mathbf{x},\\mathbf{x}^{\\prime}) = \\exp \\left(-\\frac{||\\mathbf{x} - \\mathbf{x}^{\\prime}||^{2}}{50}\\right) \\). We provide more studies on the kernel configuration in Appendix B. The results show that local hypotheses such as Constant Speed/Acceleration perform much better than global hypotheses such as LR and GPR. This is probably because, as virtual trajectory generation happens in an online fashion, it is hard to get a reliable fit using only limited data points on historical time steps." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.418, + 0.893, + 0.555 + ], + "angle": 0, + "content": "\\(\\Delta t\\) in OCM. There is a trade-off when choosing the time difference \\(\\Delta t\\) in OCM (Section 4). A large \\(\\Delta t\\) decreases the noise of velocity estimation, but is also likely to discourage approximating object motion as linear. Therefore, we study the influence of varying \\(\\Delta t\\) in Table 7. Our results agree with our analysis that increasing \\(\\Delta t\\) from \\(\\Delta t = 1\\) can boost performance. But increasing \\(\\Delta t\\) higher than a best-practice value instead hurts the performance because of the difficulty of maintaining the approximation of linear motion." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.564, + 0.619, + 0.58 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.583, + 0.892, + 0.779 + ], + "angle": 0, + "content": "We analyze the popular motion-based tracker SORT and recognize its intrinsic limitations from using Kalman filter. These limitations significantly hurt tracking accuracy when the tracker fails to gain observations for supervision - likely caused by unreliable detectors, occlusion, or fast and non-linear target object motion. To address these issues, we propose Observation-Centric SORT (OC-SORT). OC-SORT is more robust to occlusion and non-linear object motion while keeping simple, online, and real-time. In our experiments on diverse datasets, OC-SORT significantly outperforms the state-of-the-art. The gain is especially significant for multi-object tracking under occlusion and non-linear object motion." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Acknowledgement. We thank David Held, Deva Ramanan, and Siddharth Ancha for the discussion about the theoretical modeling of OC-SORT. We thank Yuda Song for the help with the analysis of OCM error distribution. We also would like to thank Zhengyi Luo, Yuda Song, and Erica Weng for the detailed feedback on the paper writing. We also thank Yifu Zhang for sharing his experience with ByteTrack. This project was sponsored in part by NSF NRI award 2024173." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9693" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.09, + 0.176, + 0.107 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.156 + ], + "angle": 0, + "content": "[1] Yaakov Bar-Shalom, Fred Daum, and Jim Huang. The probabilistic data association filter. IEEE Control Systems Magazine, 29(6):82-100, 2009." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.157, + 0.472, + 0.21 + ], + "angle": 0, + "content": "[2] Sumit Basu, Irfan Essa, and Alex Pentland. Motion regularization for model-based head tracking. In Proceedings of 13th International Conference on Pattern Recognition, volume 3, pages 611-616. IEEE, 1996." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.21, + 0.471, + 0.264 + ], + "angle": 0, + "content": "[3] Alex Bewley, Zongyuan Ge, Lionel Ott, Fabio Ramos, and Ben Upcroft. Simple online and realtime tracking. In 2016 IEEE international conference on image processing (ICIP), pages 3464-3468. IEEE, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.265, + 0.472, + 0.319 + ], + "angle": 0, + "content": "[4] Guillem Brasó and Laura Leal-Taixe. Learning a neural solver for multiple object tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6247-6257, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.319, + 0.471, + 0.359 + ], + "angle": 0, + "content": "[5] Jiarui Cai, Mingze Xu, Wei Li, Yuanjun Xiong, Wei Xia, Zhuowen Tu, and Stefano Soatto. Memot: multi-object tracking with memory. In CVPR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.359, + 0.471, + 0.399 + ], + "angle": 0, + "content": "[6] Jinkun Cao, Jiangmiao Pang, Xinshuo Weng, Rawal Khirodkar, and Kris M Kitani. Object tracking by hierarchical part-whole attention." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.399, + 0.471, + 0.467 + ], + "angle": 0, + "content": "[7] Jinkun Cao, Hongyang Tang, Hao-Shu Fang, Xiaoyong Shen, Cewu Lu, and Yu-Wing Tai. Cross-domain adaptation for animal pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9498-9507, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.467, + 0.471, + 0.507 + ], + "angle": 0, + "content": "[8] Jinkun Cao, Hao Wu, and Kris Kitani. Track targets by dense spatio-temporal position encoding. arXiv preprint arXiv:2210.09455, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.508, + 0.471, + 0.589 + ], + "angle": 0, + "content": "[9] Ming-Fang Chang, John Lambert, Patsorn Sangkloy, Jagjeet Singh, Slawomir Bak, Andrew Hartnett, De Wang, Peter Carr, Simon Lucey, Deva Ramanan, et al. Argoverse: 3d tracking and forecasting with rich maps. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8748-8757, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.589, + 0.471, + 0.643 + ], + "angle": 0, + "content": "[10] Yin-Wen Chang, Cho-Jui Hsieh, Kai-Wei Chang, Michael Ringgaard, and Chih-Jen Lin. Training and testing low-degree polynomial data mappings via linearsvm. Journal of Machine Learning Research, 11(4), 2010." + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.643, + 0.471, + 0.697 + ], + "angle": 0, + "content": "[11] Wongun Choi. Near-online multi-target tracking with aggregated local flow descriptor. In Proceedings of the IEEE international conference on computer vision, pages 3029-3037, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.697, + 0.471, + 0.751 + ], + "angle": 0, + "content": "[12] Peng Chu, Jiang Wang, Quanzeng You, Haibin Ling, and Zicheng Liu. Transmot: Spatial-temporal graph transformer for multiple object tracking. arXiv preprint arXiv:2104.00194, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.752, + 0.471, + 0.82 + ], + "angle": 0, + "content": "[13] Peng Dai, Renliang Weng, Wongun Choi, Changshui Zhang, Zhangping He, and Wei Ding. Learning a proposal classifier for multiple object tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2443-2452, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.82, + 0.471, + 0.887 + ], + "angle": 0, + "content": "[14] Patrick Dendorfer, Hamid Rezatofighi, Anton Milan, Javen Shi, Daniel Cremers, Ian Reid, Stefan Roth, Konrad Schindler, and Laura Leal-Taixe. Mot20: A benchmark for multi object tracking in crowded scenes. arXiv preprint arXiv:2003.09003, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.887, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[15] David Duvenaud. Automatic model construction with Gaus" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.888, + 0.107 + ], + "angle": 0, + "content": "sian processes. PhD thesis, University of Cambridge, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.107, + 0.892, + 0.161 + ], + "angle": 0, + "content": "[16] Hao-Shu Fang, Jinkun Cao, Yu-Wing Tai, and Cewu Lu. Pairwise body-part attention for recognizing human-object interactions. In Proceedings of the European conference on computer vision (ECCV), pages 51-67, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.161, + 0.892, + 0.215 + ], + "angle": 0, + "content": "[17] Juergen Gall, Angela Yao, Nima Razavi, Luc Van Gool, and Victor Lempitsky. Hough forests for object detection, tracking, and action recognition. IEEE transactions on pattern analysis and machine intelligence, 33(11):2188-2202, 2011." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.215, + 0.892, + 0.255 + ], + "angle": 0, + "content": "[18] Damien Garreau, Wittawat Jitkrittum, and Motonobu Kanagawa. Large sample analysis of the median heuristic. arXiv preprint arXiv:1707.07269, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.255, + 0.892, + 0.295 + ], + "angle": 0, + "content": "[19] Zheng Ge, Songtao Liu, Feng Wang, Zeming Li, and Jian Sun. Yolox: Exceeding yolo series in 2021. arXiv preprint arXiv:2107.08430, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.295, + 0.892, + 0.349 + ], + "angle": 0, + "content": "[20] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. The International Journal of Robotics Research, 32(11):1231-1237, 2013." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.349, + 0.892, + 0.404 + ], + "angle": 0, + "content": "[21] Nicolas Franco Gonzalez, Andres Ospina, and Philippe Calvez. Smat: Smart multiple affinity metrics for multiple object tracking. In International Conference on Image Analysis and Recognition, pages 48-62. Springer, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.404, + 0.892, + 0.471 + ], + "angle": 0, + "content": "[22] Fredrik Gustafsson, Fredrik Gunnarsson, Niclas Bergman, Urban Forssell, Jonas Jansson, Rickard Karlsson, and P-J Nordlund. Particle filters for positioning, navigation, and tracking. IEEE Transactions on signal processing, 50(2):425-437, 2002." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.471, + 0.892, + 0.526 + ], + "angle": 0, + "content": "[23] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.526, + 0.892, + 0.552 + ], + "angle": 0, + "content": "[24] David V Hinkley. On the ratio of two correlated normal random variables. Biometrika, 56(3):635-639, 1969." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.552, + 0.892, + 0.606 + ], + "angle": 0, + "content": "[25] Andrea Hornakova, Roberto Henschel, Bodo Rosenhahn, and Paul Swoboda. Lifted disjoint paths with application in multiple object tracking. In International Conference on Machine Learning, pages 4364-4375. PMLR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.606, + 0.892, + 0.687 + ], + "angle": 0, + "content": "[26] Andrea Hornakova, Timo Kaiser, Paul Swoboda, Michal Rolinek, Bodo Rosenhahn, and Roberto Henschel. Making higher order mot scalable: An efficient approximate solver for lifted disjoint paths. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6330-6340, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.687, + 0.892, + 0.714 + ], + "angle": 0, + "content": "[27] Andrew H Jazwinski. Stochastic processes and filtering theory. Courier Corporation, 2007." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.714, + 0.892, + 0.781 + ], + "angle": 0, + "content": "[28] Simon J Julier and Jeffrey K Uhlmann. New extension of the kalman filter to nonlinear systems. In Signal processing, sensor fusion, and target recognition VI, volume 3068, pages 182-193. International Society for Optics and Photonics, 1997." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.781, + 0.892, + 0.822 + ], + "angle": 0, + "content": "[29] Simon J Julier and Jeffrey K Uhlmann. Unscented filtering and nonlinear estimation. Proceedings of the IEEE, 92(3):401-422, 2004." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.822, + 0.892, + 0.848 + ], + "angle": 0, + "content": "[30] Rudolf Emil Kalman et al. Contributions to the theory of optimal control. Bol. soc. mat. mexicana, 5(2):102-119, 1960." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.848, + 0.892, + 0.89 + ], + "angle": 0, + "content": "[31] Kris M Kitani, Brian D Ziebart, James Andrew Bagnell, and Martial Hebert. Activity forecasting. In European conference on computer vision, pages 201-214. Springer, 2012." + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.89 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9694" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.093, + 0.468, + 0.133 + ], + "angle": 0, + "content": "[32] Jonathan Ko and Dieter Fox. Gp-bayesfilters: Bayesian filtering using gaussian process prediction and observation models. Autonomous Robots, 27(1):75-90, 2009." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.134, + 0.468, + 0.187 + ], + "angle": 0, + "content": "[33] Parth Kothari, Sven Kreiss, and Alexandre Alahi. Human trajectory forecasting in crowds: A deep learning perspective. IEEE Transactions on Intelligent Transportation Systems, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.188, + 0.468, + 0.214 + ], + "angle": 0, + "content": "[34] Erich L Lehmann and George Casella. Theory of point estimation. Springer Science & Business Media, 2006." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.215, + 0.468, + 0.268 + ], + "angle": 0, + "content": "[35] Wei Li, Yuanjun Xiong, Shuo Yang, Mingze Xu, Yongxin Wang, and Wei Xia. Semi-tcl: Semi-supervised track contrastive representation learning. arXiv preprint arXiv:2107.02396, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.269, + 0.468, + 0.322 + ], + "angle": 0, + "content": "[36] Chao Liang, Zhipeng Zhang, Yi Lu, Xue Zhou, Bing Li, Xiyong Ye, and Jianxiao Zou. Rethinking the competition between detection and reid in multi-object tracking. arXiv preprint arXiv:2010.12138, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.323, + 0.468, + 0.389 + ], + "angle": 0, + "content": "[37] Jonathon Luiten, Aljosa Osep, Patrick Dendorfer, Philip Torr, Andreas Geiger, Laura Leal-Taixe, and Bastian Leibe. Hota: A higher order metric for evaluating multi-object tracking. International journal of computer vision, 129(2):548-578, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.391, + 0.468, + 0.431 + ], + "angle": 0, + "content": "[38] Gerard Maggiolino, Adnan Ahmad, Jinkun Cao, and Kris Kitani. Deep oc-sort: Multi-pedestrian tracking by adaptive re-identification. arXiv preprint arXiv:2302.11813, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.431, + 0.468, + 0.484 + ], + "angle": 0, + "content": "[39] Tim Meinhardt, Alexander Kirillov, Laura Leal-Taixe, and Christoph Feichtenhofer. Trackformer: Multi-object tracking with transformers. arXiv preprint arXiv:2101.02702, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.485, + 0.468, + 0.538 + ], + "angle": 0, + "content": "[40] Stan Melax, Leonid Keselman, and Sterling Orsten. Dynamics based 3d skeletal hand tracking. In Proceedings of the ACM SIGGRAPH Symposium on Interactive 3D Graphics and Games, pages 184-184, 2013." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.539, + 0.468, + 0.579 + ], + "angle": 0, + "content": "[41] Anton Milan, Laura Leal-Taixe, Ian Reid, Stefan Roth, and Konrad Schindler. Mot16: A benchmark for multi-object tracking. arXiv preprint arXiv:1603.00831, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.58, + 0.468, + 0.647 + ], + "angle": 0, + "content": "[42] Jiangmiao Pang, Linlu Qiu, Xia Li, Haofeng Chen, Qi Li, Trevor Darrell, and Fisher Yu. Quasi-dense similarity learning for multiple object tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 164-173, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.648, + 0.468, + 0.7 + ], + "angle": 0, + "content": "[43] Dezhi Peng, Zikai Sun, Zirong Chen, Zirui Cai, Lele Xie, and Lianwen Jin. Detecting heads using feature refine net and cascaded multi-scale architecture. arXiv preprint arXiv:1803.09256, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.701, + 0.468, + 0.74 + ], + "angle": 0, + "content": "[44] Lawrence Rabiner and Biinghwang Juang. An introduction to hidden markov models. *iiee assp magazine*, 3(1):4-16, 1986." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.741, + 0.468, + 0.808 + ], + "angle": 0, + "content": "[45] Akshay Rangesh, Pranav Maheshwari, Mez Gebre, Siddhesh Mhatre, Vahid Ramezani, and Mohan M Trivedi. Trackmpnn: A message passing graph neural architecture for multi-object tracking. arXiv preprint arXiv:2101.04206, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.809, + 0.468, + 0.864 + ], + "angle": 0, + "content": "[46] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 779-788, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.864, + 0.468, + 0.89 + ], + "angle": 0, + "content": "[47] Steven Reece and Stephen Roberts. An introduction to gaussian processes for the kalman filter expert. In 2010 13th In-" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.468, + 0.89 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.094, + 0.89, + 0.119 + ], + "angle": 0, + "content": "ternational Conference on Information Fusion, pages 1-9. IEEE, 2010." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.121, + 0.892, + 0.174 + ], + "angle": 0, + "content": "[48] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.175, + 0.892, + 0.255 + ], + "angle": 0, + "content": "[49] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 658-666, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.256, + 0.892, + 0.338 + ], + "angle": 0, + "content": "[50] Toby Sharp, Cem Keskin, Duncan Robertson, Jonathan Taylor, Jamie Shotton, David Kim, Christoph Rhemann, Ido Leichter, Alon Vinnikov, Yichen Wei, et al. Accurate, robust, and flexible real-time hand tracking. In Proceedings of the 33rd annual ACM conference on human factors in computing systems, pages 3633-3642, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.339, + 0.892, + 0.378 + ], + "angle": 0, + "content": "[51] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.379, + 0.892, + 0.432 + ], + "angle": 0, + "content": "[52] Gerald L Smith, Stanley F Schmidt, and Leonard A McGee. Application of statistical filter theory to the optimal estimation of position and velocity on board a circumlunar vehicle. National Aeronautics and Space Administration, 1962." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.433, + 0.892, + 0.499 + ], + "angle": 0, + "content": "[53] Daniel Stadler and Jurgen Beyerer. Improving multiple pedestrian tracking by track management and occlusion handling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10958-10967, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.5, + 0.892, + 0.554 + ], + "angle": 0, + "content": "[54] Peize Sun, Jinkun Cao, Yi Jiang, Zehuan Yuan, Song Bai, Kris Kitani, and Ping Luo. Dancetrack: Multi-object tracking in uniform appearance and diverse motion. arXiv preprint arXiv:2111.14690, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.555, + 0.892, + 0.607 + ], + "angle": 0, + "content": "[55] Peize Sun, Jinkun Cao, Yi Jiang, Rufeng Zhang, Enze Xie, Zehuan Yuan, Changhu Wang, and Ping Luo. Transtrack: Multiple object tracking with transformer. arXiv preprint arXiv:2012.15460, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.608, + 0.892, + 0.675 + ], + "angle": 0, + "content": "[56] Ramana Sundararaman, Cedric De Almeida Braga, Eric Marchand, and Julien Pette. Tracking pedestrian heads in dense crowd. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3865-3875, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.676, + 0.892, + 0.717 + ], + "angle": 0, + "content": "[57] Pavel Tokmakov, Jie Li, Wolfram Burgard, and Adrien Gaidon. Learning to track with object permanence. In ICCV, pages 10860-10869, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.717, + 0.892, + 0.77 + ], + "angle": 0, + "content": "[58] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.771, + 0.892, + 0.852 + ], + "angle": 0, + "content": "[59] Gaoang Wang, Renshu Gu, Zuozhu Liu, Weijie Hu, Mingli Song, and Jenq-Neng Hwang. Track without appearance: Learn box and tracklet embedding with local and global motion patterns for vehicle tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9876-9886, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.853, + 0.892, + 0.893 + ], + "angle": 0, + "content": "[60] Shuai Wang, Hao Sheng, Yang Zhang, Yubin Wu, and Zhang Xiong. A general recurrent tracking framework without real data. In Proceedings of the IEEE/CVF International Confer-" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.094, + 0.892, + 0.893 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.957 + ], + "angle": 0, + "content": "9695" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.093, + 0.428, + 0.106 + ], + "angle": 0, + "content": "ence on Computer Vision, pages 13219-13228, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.107, + 0.47, + 0.16 + ], + "angle": 0, + "content": "[61] Yongxin Wang, Kris Kitani, and Xinshuo Weng. Joint object detection and multi-object tracking with graph neural networks. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 13708-13715. IEEE, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.16, + 0.47, + 0.2 + ], + "angle": 0, + "content": "[62] Christopher Williams and Carl Rasmussen. Gaussian processes for regression. Advances in neural information processing systems, 8, 1995." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.2, + 0.47, + 0.255 + ], + "angle": 0, + "content": "[63] Nicolai Wojke, Alex Bewley, and Dietrich Paulus. Simple online and realtime tracking with a deep association metric. In 2017 IEEE international conference on image processing (ICIP), pages 3645-3649. IEEE, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.255, + 0.47, + 0.322 + ], + "angle": 0, + "content": "[64] Jialian Wu, Jiale Cao, Liangchen Song, Yu Wang, Ming Yang, and Junsong Yuan. Track to detect and segment: An online multi-object tracker. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12352-12361, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.322, + 0.47, + 0.377 + ], + "angle": 0, + "content": "[65] Yu Xiang, Alexandre Alahi, and Silvio Savarese. Learning to track: Online multi-object tracking by decision making. In Proceedings of the IEEE international conference on computer vision, pages 4705-4713, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.377, + 0.47, + 0.417 + ], + "angle": 0, + "content": "[66] Yuliang Xiu, Jiefeng Li, Haoyu Wang, Yinghong Fang, and Cewu Lu. Pose flow: Efficient online pose tracking. arXiv preprint arXiv:1802.00977, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.417, + 0.47, + 0.471 + ], + "angle": 0, + "content": "[67] Yihong Xu, Yutong Ban, Guillaume Delorme, Chuang Gan, Daniela Rus, and Xavier Alameda-Pineda. Transcenter: Transformers with dense queries for multiple-object tracking. arXiv preprint arXiv:2103.15145, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.471, + 0.47, + 0.512 + ], + "angle": 0, + "content": "[68] Bin Yan, Yi Jiang, Peize Sun, Dong Wang, Zehuan Yuan, Ping Luo, and Huchuan Lu. Towards grand unification of object tracking. In ECCV. Springer, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.512, + 0.47, + 0.538 + ], + "angle": 0, + "content": "[69] Fangao Zeng and et al. Motr: End-to-end multiple-object tracking with transformer. In ECCV. Springer, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.538, + 0.47, + 0.592 + ], + "angle": 0, + "content": "[70] Yifu Zhang, Peize Sun, Yi Jiang, Dongdong Yu, Zehuan Yuan, Ping Luo, Wenyu Liu, and Xinggang Wang. Byetetrack: Multi-object tracking by associating every detection box. arXiv preprint arXiv:2110.06864, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.592, + 0.47, + 0.646 + ], + "angle": 0, + "content": "[71] Yifu Zhang, Chunyu Wang, Xinggang Wang, Wenjun Zeng, and Wenyu Liu. Fairmot: On the fairness of detection and re-identification in multiple object tracking. International Journal of Computer Vision, 129(11):3069-3087, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.646, + 0.47, + 0.713 + ], + "angle": 0, + "content": "[72] Zhaohui Zheng, Ping Wang, Wei Liu, Jinze Li, Rongguang Ye, and Dongwei Ren. Distance-iou loss: Faster and better learning for bounding box regression. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 12993-13000, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.713, + 0.47, + 0.755 + ], + "angle": 0, + "content": "[73] Xingyi Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Tracking objects as points. In ECCV, pages 474-490. Springer, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.755, + 0.47, + 0.781 + ], + "angle": 0, + "content": "[74] Xingyi Zhou, Dequan Wang, and Philipp Kähenbuhl. Objects as points. arXiv preprint arXiv:1904.07850, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.781, + 0.47, + 0.807 + ], + "angle": 0, + "content": "[75] Xingyi Zhou, Tianwei Yin, Vladlen Koltun, and Philipp Krahenbuhl. Global tracking transformers. In CVPR, 2022." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.093, + 0.47, + 0.807 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9696" + } + ] +] \ No newline at end of file diff --git a/2023/Observation-Centric SORT_ Rethinking SORT for Robust Multi-Object Tracking/b4c4cc9d-52d9-4722-825a-050e1d27cf7f_origin.pdf b/2023/Observation-Centric SORT_ Rethinking SORT for Robust Multi-Object Tracking/b4c4cc9d-52d9-4722-825a-050e1d27cf7f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..9495613fab5aec2201713b6b21119273ab693c4b --- /dev/null +++ b/2023/Observation-Centric SORT_ Rethinking SORT for Robust Multi-Object Tracking/b4c4cc9d-52d9-4722-825a-050e1d27cf7f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:de885907f30a8cad0cd03135d2e42ba4e908374b6e9eaf0360d2d1052abb96da +size 318017 diff --git a/2023/Observation-Centric SORT_ Rethinking SORT for Robust Multi-Object Tracking/full.md b/2023/Observation-Centric SORT_ Rethinking SORT for Robust Multi-Object Tracking/full.md new file mode 100644 index 0000000000000000000000000000000000000000..902ef5b231cee045700ee33ea9c6effb84c1e4d9 --- /dev/null +++ b/2023/Observation-Centric SORT_ Rethinking SORT for Robust Multi-Object Tracking/full.md @@ -0,0 +1,331 @@ +# Observation-Centric SORT: Rethinking SORT for Robust Multi-Object Tracking + +Jinkun Cao $^{1}$ Jiangmiao Pang $^{2}$ Xinshuo Weng $^{3}$ Rawal Khirodkar $^{1}$ Kris Kitani $^{1}$ $^{1}$ Carnegie Mellon University $^{2}$ Shanghai AI Laboratory $^{3}$ Nvidia + +# Abstract + +Kalman filter (KF) based methods for multi-object tracking (MOT) make an assumption that objects move linearly. While this assumption is acceptable for very short periods of occlusion, linear estimates of motion for prolonged time can be highly inaccurate. Moreover, when there is no measurement available to update Kalman filter parameters, the standard convention is to trust the priori state estimations for posteriori update. This leads to the accumulation of errors during a period of occlusion. The error causes significant motion direction variance in practice. In this work, we show that a basic Kalman filter can still obtain state-of-the-art tracking performance if proper care is taken to fix the noise accumulated during occlusion. Instead of relying only on the linear state estimate (i.e., estimation-centric approach), we use object observations (i.e., the measurements by object detector) to compute a virtual trajectory over the occlusion period to fix the error accumulation of filter parameters. This allows more time steps to correct errors accumulated during occlusion. We name our method Observation-Centric SORT (OC-SORT). It remains Simple, Online, and Real-Time but improves robustness during occlusion and non-linear motion. Given off-the-shelf detections as input, OC-SORT runs at $700+$ FPS on a single CPU. It achieves state-of-the-art on multiple datasets, including MOT17, MOT20, KITTI, head tracking, and especially DanceTrack where the object motion is highly non-linear. The code and models are available at https://github.com/noahcao/OC_SORT. + +# 1. Introduction + +We aim to develop a motion model-based multi-object tracking (MOT) method that is robust to occlusion and nonlinear motion. Most existing motion model-based algorithms assume that the tracking targets have a constant velocity within a time interval, which is called the linear motion assumption. This assumption breaks in many practical scenarios, but it still works because when the time interval is small enough, the object's motion can be reasonably ap- + +![](images/5a0838cea78f29b37d62ae49f4eb3b995510ba9830aa17a2fce3e36949e545bf.jpg) + +![](images/bfb6f1f89cbd0d313250928330757f39596ecd085dcca59a94ee799b5a3b1769.jpg) +(a) SORT + +![](images/6b3834b6036c88f9cf10417ad67ee7e3226beb078ebbf1d6d7e4eb7d4d4cbcb8.jpg) + +![](images/8b044f0ea30ee14e9c9801d8db562cf61f06ab5e620e555afc0d9608dbf18400.jpg) + +![](images/3a16e705f0b020b43a39f9919535d2fc7dbcd1042668bbe2cd15f5a1ad57864c.jpg) + +![](images/5e9472c633d3813f1c0e4d0d38b6044a4cc65696afeaf1271b2e371ca652d7ff.jpg) +(b) The proposed OC-SORT + +![](images/b92475af28c88b8ee65bab550c8ef4291a7393e219ed5930f5d523310ac6e6d3.jpg) +Figure 1. Samples from the results on DanceTrack [54]. SORT and OC-SORT use the same detection results. On the third frame, SORT encounters an ID switch for the backflip target while OC-SORT tracks it consistently. + +![](images/86515da1892e887e1bbb101381a5768a76a55a57a3f9b1affbd9e1d6d1ff1352.jpg) + +proximated as linear. In this work, we are motivated by the fact that most of the errors from motion model-based tracking methods occur when occlusion and non-linear motion happen together. To mitigate the adverse effects caused, we first rethink current motion models and recognize some limitations. Then, we propose addressing them for more robust tracking performance, especially in occlusion. + +As the main branch of motion model-based tracking, filtering-based methods assume a transition function to predict the state of objects on future time steps, which are called state "estimations". Besides estimations, they leverage an observation model, such as an object detector, to derive the state measurements of target objects, also called "observations". Observations usually serve as auxiliary information to help update the posteriori parameters of the filter. The trajectories are still extended by the state estimations. Among this line of work, the most widely used one is SORT [3], which uses a Kalman filter (KF) to estimate object states and a linear motion function as the transition + +function between time steps. However, SORT shows insufficient tracking robustness when the object motion is nonlinear, and no observations are available when updating the filter posteriori parameters. + +In this work, we recognize three limitations of SORT. First, although the high frame rate is the key to approximating the object motion as linear, it also amplifies the model's sensitivity to the noise of state estimations. Specifically, between consecutive frames of a high frame-rate video, we demonstrate that the noise of displacement of the object can be of the same magnitude as the actual object displacement, leading to the estimated object velocity by KF suffering from a significant variance. Also, the noise in the velocity estimate will accumulate into the position estimate by the transition process. Second, the noise of state estimations by KF is accumulated along the time when there is no observation available in the update stage of KF. We show that the error accumulates very fast with respect to the time of the target object's being untracked. The noise's influence on the velocity direction often makes the track lost again even after re-association. Last, given the development of modern detectors, the object state by detections usually has lower variance than the state estimations propagated along time steps by a fixed transition function in filters. However, SORT is designed to prolong the object trajectories by state estimations instead of observations. + +To relieve the negative effect of these limitations, we propose two main innovations in this work. First, we design a module to use object state observations to reduce the accumulated error during the track's being lost in a backcheck fashion. To be precise, besides the traditional stages of predict and update, we add a stage of re-update to correct the accumulated error. The re-update is triggered when a track is re-activated by associating to an observation after a period of being untracked. The re-update uses virtual observations on the historical time steps to prevent error accumulation. The virtual observations come from a trajectory generated using the last-seen observation before untracked and the latest observation re-activating this track as anchors. We name it Observation-centric Re-Update (ORU). + +Besides ORU, the assumption of linear motion provides the consistency of the object motion direction. But this cue is hard to be used in SORT's association because of the heavy noise in direction estimation. But we propose an observation-centric manner to incorporate the direction consistency of tracks in the cost matrix for the association. We name it Observation-Centric Momentum (OCM). We also provide analytical justification for the noise of velocity direction estimation in practice. + +The proposed method, named as Observation-Centric SORT or OC-SORT in short, remains simple, online, real-time and significantly improves robustness over occlusion and non-linear motion. Our contributions are summarized + +as the following: + +1. We recognize, analytically and empirically, three limitations of SORT, i.e. sensitivity to the noise of state estimations, error accumulation over time, and being estimation-centric; +2. We propose OC-SORT for tracking under occlusion and non-linear motion by fixing SORT's limitations. It achieves state-of-the-art performance on multiple datasets in an online and real-time fashion. + +# 2. Related Works + +Motion Models. Many modern MOT algorithms [3, 11, 63, 70, 73] use motion models. Typically, these motion models use Bayesian estimation [34] to predict the next state by maximizing a posterior estimation. As one of the most classic motion models, Kalman filter (KF) [30] is a recursive Bayes filter that follows a typical predict-update cycle. The true state is assumed to be an unobserved Markov process, and the measurements are observations from a hidden Markov model [44]. Given that the linear motion assumption limits KF, follow-up works like Extended KF [52] and Unscented KF [28] were proposed to handle non-linear motion with first-order and third-order Taylor approximation. However, they still rely on approximating the Gaussian prior assumed by KF and require motion pattern assumption. On the other hand, particle filters [22] solve the non-linear motion by sampling-based posterior estimation but require exponential order of computation. Therefore, these variants of Kalman filter and particle filters are rarely adopted in the visual multi-object tracking and the mostly adopted motion model is still based on Kalman filter [3]. + +Multi-object Tracking. As a classic computer vision task, visual multi-object tracking is traditionally approached from probabilistic perspectives, e.g. joint probabilistic association [1]. And modern video object tracking is usually built upon modern object detectors [46, 48, 74]. SORT [3] adopts the Kalman filter for motion-based multi-object tracking given observations from deep detectors. DeepSORT [63] further introduces deep visual features [23, 51] into object association under the framework of SORT. Re-identification-based object association[42, 63, 71] has also become popular since then but falls short when scenes are crowded and objects are represented coarsely (e.g. enclosed by bounding boxes), or object appearance is not distinguishable. More recently, transformers [58] have been introduced to MOT [8, 39, 55, 69] to learn deep representations from both visual information and object trajectories. However, their performance still has a significant gap between state-of-the-art tracking-by-detection methods in terms of both accuracy and time efficiency. + +# 3. Rethink the Limitations of SORT + +In this section, we review Kalman filter and SORT [3]. We recognize some of their limitations, which are significant with occlusion and non-linear object motion. We are motivated to improve tracking robustness by fixing them. + +# 3.1. Preliminaries + +Kalman filter (KF) [30] is a linear estimator for dynamical systems discretized in the time domain. KF only requires the state estimations on the previous time step and the current measurement to estimate the target state on the next time step. The filter maintains two variables, the posteriori state estimate $\mathbf{x}$ , and the posteriori estimate covariance matrix $\mathbf{P}$ . In the task of object tracking, we describe the KF process with the state transition model $\mathbf{F}$ , the observation model $\mathbf{H}$ , the process noise $\mathbf{Q}$ , and the observation noise $\mathbf{R}$ . At each step $t$ , given observations $\mathbf{z}_t$ , KF works in an alternation of predict and update stages: + +$$ +\begin{array}{l} \text {p r e d i c t} \left\{ \begin{array}{l} \hat {\mathbf {x}} _ {t | t - 1} = \mathbf {F} _ {t} \hat {\mathbf {x}} _ {t - 1 | t - 1} \\ \mathbf {P} _ {t | t - 1} = \mathbf {F} _ {t} \mathbf {P} _ {t - 1 | t - 1} \mathbf {F} _ {t} ^ {\top} + \mathbf {Q} _ {t} \end{array} , \right. \\ \text {u p d a t e} \left\{ \begin{array}{l} \mathbf {K} _ {t} = \mathbf {P} _ {t | t - 1} \mathbf {H} _ {t} ^ {\top} \left(\mathbf {H} _ {t} \mathbf {P} _ {t | t - 1} \mathbf {H} _ {t} ^ {\top} + \mathbf {R} _ {t}\right) ^ {- 1} \quad (1) \\ \hat {\mathbf {x}} _ {t | t} = \hat {\mathbf {x}} _ {t | t - 1} + \mathbf {K} _ {t} (\mathbf {z} _ {t} - \mathbf {H} _ {t} \hat {\mathbf {x}} _ {t | t - 1}) \\ \mathbf {P} _ {t | t} = (\mathbf {I} - \mathbf {K} _ {t} \mathbf {H} _ {t}) \mathbf {P} _ {t | t - 1} \end{array} . \right. \\ \end{array} +$$ + +The stage of predict is to derive the state estimations on the next time step $t$ . Given a measurement of target states on the next step $t$ , the stage of update aims to update the posteriori parameters in KF. Because the measurement comes from the observation model $\mathbf{H}$ , it is also called "observation" in many scenarios. + +SORT [3] is a multi-object tracker built upon KF. The KF's state $\mathbf{x}$ in SORT is defined as $\mathbf{x} = [u,v,s,r,\dot{u},\dot{v},\dot{s}]^{\top}$ , where $(u,v)$ is the 2D coordinates of the object center in the image. $s$ is the bounding box scale (area) and $r$ is the bounding box aspect ratio. The aspect ratio $r$ is assumed to be constant. The other three variables, $\dot{u},\dot{v}$ and $\dot{s}$ are the corresponding time derivatives. The observation is a bounding box $\mathbf{z} = [u,v,w,h,c]^{\top}$ with object center position $(u,v)$ , object width $w$ , and height $h$ and the detection confidence $c$ respectively. SORT assumes linear motion as the transition model $\mathbf{F}$ which leads to the state estimation as + +$$ +u _ {t + 1} = u _ {t} + \dot {u} _ {t} \Delta t, \quad v _ {t + 1} = v _ {t} + \dot {v} _ {t} \Delta t. \tag {2} +$$ + +To leverage KF (Eq 1) in SORT for visual MOT, the stage of predict corresponds to estimating the object position on the next video frame. And the observations used for the update stage usually come from a detection model. The update stage is to update Kalman filter parameters and does not directly edit the tracking outcomes. + +When the time difference between two steps is constant during the transition, $e.g.$ , the video frame rate is constant, + +we can set $\Delta t = 1$ . When the video frame rate is high, SORT works well even when the object motion is non-linear globally, (e.g. dancing, fencing, wrestling) because the motion of the target object can be well approximated as linear within short time intervals. However, in practice, observations are often absent on some time steps, e.g. the target object is occluded in multi-object tracking. In such cases, we cannot update the KF parameters by the update operation as in Eq. 1 anymore. SORT uses the priori estimations directly as posterior. We call this "dummy update", namely + +$$ +\hat {\mathbf {x}} _ {t \mid t} = \hat {\mathbf {x}} _ {t \mid t - 1}, \mathbf {P} _ {t \mid t} = \mathbf {P} _ {t \mid t - 1}. \tag {3} +$$ + +The philosophy behind such a design is to trust estimations when no observations are available to supervise them. We thus call the tracking algorithms following this scheme "estimation-centric". However, we will see that this estimation-centric mechanism can cause trouble when nonlinear motion and occlusion happen together. + +# 3.2. Limitations of SORT + +In this section, we identify three main limitations of SORT which are connected. This analysis lays the foundation of our proposed method. + +# 3.2.1 Sensitive to State Noise + +Now we show that SORT is sensitive to the noise from KF's state estimations. To begin with, we assume that the estimated object center position follows $u \sim \mathcal{N}(\mu_u, \sigma_u^2)$ and $v \sim \mathcal{N}(\mu_v, \sigma_v^2)$ , where $(\mu_u, \mu_v)$ is the underlying true position. Then, if we assume that the state noises are independent on different steps, by Eq.2, the object speed between two time steps, $t \longrightarrow t + \Delta t$ , is + +$$ +\dot {u} = \frac {u _ {t + \Delta t} - u _ {t}}{\Delta t}, \quad \dot {v} = \frac {v _ {t + \Delta t} - v _ {t}}{\Delta t}, \tag {4} +$$ + +making the noise of estimated speed $\delta_{\dot{u}}\sim \mathcal{N}(0,\frac{2\sigma_u^2}{(\Delta t)^2})$ $\delta_{\dot{v}}\sim \mathcal{N}(0,\frac{2\sigma_v^2}{(\Delta t)^2})$ . Therefore, a small $\Delta t$ will amplify the noise. This suggests that SORT will suffer from the heavy noise of velocity estimation on high-frame-rate videos. The analysis above is simplified from the reality. In practice, velocity won't be determined by the state on future time steps. For a more strict analysis, please refer to Appendix G. + +Moreover, for most multi-object tracking scenarios, the target object displacement is only a few pixels between consecutive frames. For instance, the average displacement is 1.93 pixels and 0.65 pixels along the image width and height for the MOT17 [41] training dataset. In such a case, even if the estimated position has a shift of only a single pixel, it causes a significant variation in the estimated speed. In general, the variance of the speed estimation can be of the same magnitude as the speed itself or even greater. This will not make a massive impact as the shift is only of few pixels + +![](images/d9a8507f837ba33dcae203cc6bf03fba0cfc3bcadb76306f639c7a4ecec00b36.jpg) +Figure 2. The pipeline of our proposed OC-SORT. The red boxes are detections, orange boxes are active tracks, blue boxes are untracked tracks, and dashed boxes are the estimates from KF. During association, OCM is used to add the velocity consistency cost. The target #1 is lost on the frame t+1 because of occlusion. But on the next frame, it is recovered by referring to its observation of the frame t by OCR. It being re-tracked triggers ORU from t to t+2 for the parameters of its KF. + +from the ground truth on the next time step and the observations, whose variance is independent of the time, will be able to fix the noise when updating the posteriori parameters. However, we find that such a high sensitivity to state noise introduces significant problems in practice after being amplified by the error accumulation across multiple time steps when no observation is available for KF update. + +# 3.2.2 Temporal Error Magnification + +For analysis above in Eq. 4, we assume the noise of the object state is i.i.d on different time steps (this is a simplified version, a more detailed analysis is provided in Appendix G). This is reasonable for object detections but not for the estimations from KF. This is because KF's estimations always rely on its estimations on previous time steps. The effect is usually minor because KF can use observation in update to prevent the posteriori state estimation and covariance, i.e. $\hat{\mathbf{x}}_{t|t}$ and $\mathbf{P}_{t|t}$ , deviating from the true value too far away. However, when no observations are provided to KF, it cannot use observation to update its parameters. Then it has to follow Eq. 3 to prolong the estimated trajectory to the next time step. Consider a track is occluded on the time steps between $t$ and $t + T$ and the noise of speed estimate follows $\delta_{\dot{u}_t} \sim \mathcal{N}(0,2\sigma_u^2)$ , $\delta_{\dot{v}_t} \sim \mathcal{N}(0,2\sigma_v^2)$ for SORT. On the step $t + T$ , state estimation would be + +$$ +u _ {t + T} = u _ {t} + T \dot {u} _ {t}, \quad v _ {t + T} = v _ {t} + T \dot {v} _ {t}, \tag {5} +$$ + +whose noise follows $\delta_{u_{t + T}}\sim \mathcal{N}(0,2T^2\sigma_u^2)$ and $\delta_{v_{t + T}}\sim$ $\mathcal{N}(0,2T^2\sigma_v^2)$ . So without the observations, the estimation from the linear motion assumption of KF results in a fast error accumulation with respect to time. Given $\sigma_v$ and $\sigma_{u}$ is of the same magnitude as object displacement between consecutive frames, the noise of final object position + +$(u_{t + T}, v_{t + T})$ is of the same magnitude as the object size. For instance, the size of pedestrians close to the camera on MOT17 is around $50 \times 300$ pixels. So even assuming the variance of position estimation is only 1 pixel, 10-frame occlusion can accumulate a shift in final position estimation as large as the object size. Such error magnification leads to a major accumulation of errors when the scenes are crowded. + +# 3.2.3 Estimation-Centric + +The aforementioned limitations come from a fundamental property of SORT that it follows KF to be estimation-centric. It allows update without the existence of observations and purely trusts the estimations. A key difference between state estimations and observations is that we can assume that the observations by an object detector in each frame are affected by i.i.d. noise $\delta_{\mathbf{z}} \sim \mathcal{N}(0, \sigma'^2)$ while the noise in state estimations can be accumulated along the hidden Markov process. Moreover, modern object detectors use powerful object visual features [48, 51]. It makes that, even on a single frame, it is usually safe to assume $\sigma' < \sigma_u$ and $\sigma' < \sigma_v$ because the object localization is more accurate by detection than from the state estimations through linear motion assumption. Combined with the previously mentioned two limitations, being estimation-centric makes SORT suffer from heavy noise when there is occlusion and the object motion is not perfectly linear. + +# 4. Observation-Centric SORT + +In this section, we introduce the proposed Observation-Centric SORT (OC-SORT). To address the limitations of SORT discussed above, we use the momentum of the object moving into the association stage and develop a + +![](images/1bbcd5172d3c920aac4e5ce8a63f11b55c0a4fdeb4743beb0bedc75c26af596e.jpg) +Figure 3. Example of how Observation-centric Re-Update (ORU) reduces the error accumulation when a track is broken. The target is occluded between the second and the third time step and the tracker finds it back at the third step. Yellow boxes are the state observations by the detector. White stars are the estimated centers without ORU. Yellow stars are the estimated centers fixed by ORU. The gray star on the fourth step is the estimated center without ORU and fails to match observations. + +pipeline with less noise and more robustness over occlusion and non-linear motion. The key is to design the tracker as observation-centric instead of estimation-centric. If a track is recovered from being untracked, we use an Observation-centric Re-Update (ORU) strategy to counter the accumulated error during the untracked period. OC-SORT also adds an Observation-Centric Momentum (OCM) term in the association cost. Please refer to Algorithm 1 in Appendix for the pseudo-code of OC-SORT. The pipeline is shown in Fig. 2. + +# 4.1. Observation-centric Re-Update (ORU) + +In practice, even if an object can be associated again by SORT after a period of being untracked, it is probably lost again because its KF parameters have already deviated far away from the correct due to the temporal error magnification. To alleviate this problem, we propose Observation-centric Re-Update (ORU) to reduce the accumulated error. Once a track is associated with an observation again after a period of being untracked ("re-activation"), we backcheck the period of its being lost and re-update the parameters of KF. The re-update is based on "observations" from a virtual trajectory. The virtual trajectory is generated referring to the observations on the steps starting and ending the untracked period. For example, by denoting the last-seen observation before being untracked as $\mathbf{z}_{t_1}$ and the observation triggering the re-association as $\mathbf{z}_{t_2}$ , the virtual trajectory is denoted as + +$$ +\tilde {\mathbf {z}} _ {t} = \operatorname {T r a j} _ {\text {v i r t u a l}} \left(\mathbf {z} _ {t _ {1}}, \mathbf {z} _ {t _ {2}}, t\right), t _ {1} < t < t _ {2}. \tag {6} +$$ + +Then, along the trajectory of $\tilde{\mathbf{z}}_t(t_1 < t < t_2)$ , we run the loop of predict and re-update. The re-update operation is + +$$ +r e - u p d a t e \left\{ \begin{array}{l} \mathbf {K} _ {t} = \mathbf {P} _ {t | t - 1} \mathbf {H} _ {t} ^ {\top} \left(\mathbf {H} _ {t} \mathbf {P} _ {t | t - 1} \mathbf {H} _ {t} ^ {\top} + \mathbf {R} _ {t}\right) ^ {- 1} \\ \hat {\mathbf {x}} _ {t | t} = \hat {\mathbf {x}} _ {t | t - 1} + \mathbf {K} _ {t} \left(\tilde {\mathbf {z}} _ {t} - \mathbf {H} _ {t} \hat {\mathbf {x}} _ {t | t - 1}\right) \\ \mathbf {P} _ {t | t} = (\mathbf {I} - \mathbf {K} _ {t} \mathbf {H} _ {t}) \mathbf {P} _ {t | t - 1} \end{array} \right. \tag {7} +$$ + +As the observations on the virtual trajectory match the motion pattern anchored by the last-seen and the latest asso + +![](images/dd82417e88e92bedf836ce19b2d8d0998876de1b60b1d7e712f34cb9dcf768ff.jpg) +Figure 4. Calculation of motion direction difference in OCM. The green line indicates an existing track and the dots are the observations on it. The red dots are the new observations to be associated. The blue link and the yellow link form the directions of $\theta^{\mathrm{track}}$ and $\theta^{\mathrm{intention}}$ respectively. The included angle is the difference of direction $\Delta \theta$ . + +ciated real observations, the update will not suffer from the error accumulated through the dummy update anymore. We call the proposed process Observation-centric Re-Update. It serves as an independent stage outside the predict-update loop and is triggered only a track is re-activated from a period of having no observations. + +# 4.2. Observation-Centric Momentum (OCM) + +In a reasonably short time interval, we can approximate the motion as linear. And the linear motion assumption also asks for consistent motion direction. But the noise prevents us from leveraging the consistency of direction. To be precise, to determine the motion direction, we need the object state on two steps with a time difference $\Delta t$ . If $\Delta t$ is small, the velocity noise would be significant because of the estimation's sensitivity to state noise. If $\Delta t$ is big, the noise of direction estimation can also be significant because of the temporal error magnification and the failure of linear motion assumption. As state observations have no problem of temporal error magnification that state estimations suffer from, we propose to use observations instead of estimations to reduce the noise of motion direction calculation and introduce the term of velocity consistency to help the association. + +With the new term, given $N$ existing tracks and $M$ de + +tectons on the new-coming time step, the association cost matrix is formulated as + +$$ +C (\hat {\mathbf {X}}, \mathbf {Z}) = C _ {\mathrm {I o U}} (\hat {\mathbf {X}}, \mathbf {Z}) + \lambda C _ {v} (\mathcal {Z}, \mathbf {Z}), \tag {8} +$$ + +where $\hat{\mathbf{X}}\in \mathbb{R}^{N\times 7}$ is the set of object state estimations and $\mathbf{Z}\in \mathbb{R}^{M\times 5}$ is the set of observations on the new time step. $\lambda$ is a weighting factor. $\mathcal{Z}$ contains the trajectory of observations of all existing tracks. $C_{\mathrm{IoU}}(\cdot ,\cdot)$ calculates the negative pairwise IoU (Intersection over Union) and $C_v(\cdot ,\cdot)$ calculates the consistency between the directions of i) linking two observations on an existing track $(\theta^{\mathrm{track}})$ and ii) linking a track's historical observation and a new observation $(\theta^{\mathrm{intention}})$ . $C_v$ contains all pairs of $\Delta \theta = |\theta^{\mathrm{track}} - \theta^{\mathrm{intention}}|$ . In our implementation, we calculate the motion direction in radians, namely $\theta = \arctan \left(\frac{v_1 - v_2}{u_1 - u_2}\right)$ where $(u_{1},v_{1})$ and $(u_{2},v_{2})$ are the observations on two different time steps. The calculation is also illustrated in Figure 4. + +Following the assumptions of noise distribution mentioned before, we can derive a closed-form probability density function of the distribution of the noise in the direction estimation. The derivation is explained in detail in Appendix A. By analyzing the property of this distribution, we reach a conclusion that, under the linear-motion model, the scale of the noise of direction estimation is negatively correlated to the time difference between the two observation points, i.e. $\Delta t$ . This suggests increasing $\Delta t$ to achieve a low-noisy estimation of $\theta$ . However, the assumption of linear motion typically holds only when $\Delta t$ is small enough. Therefore, the choice of $\Delta t$ requires a trade-off. + +Besides ORU and OCM, we also find it empirically helpful to check a track's last presence to recover it from being lost. We thus apply a heuristic Observation-Centric Recovery (OCR) technique. OCR will start a second attempt of associating between the last observation of unmatched tracks to the unmatched observations after the usual association stage. It can handle the case of an object stopping or being occluded for a short time interval. + +# 5. Experiments + +# 5.1. Experimental Setup + +Datasets. We evaluate our method on multiple multi-object tracking datasets including MOT17 [41], MOT20 [14], KITTI [20], DanceTrack [54] and CroHD [56]. MOT17 [41] and MOT20 [14] are for pedestrian tracking, where targets mostly move linearly, while scenes in MOT20 are more crowded. KITTI [20] is for pedestrian and car tracking with a relatively low frame rate of 10FPS. CroHD is a dataset for head tracking in the crowd and the results on it are included in the appendix. DanceTrack [54] is a recently proposed dataset for human tracking. For the data in DanceTrack, object localization is easy, but the object motion is highly non-linear. Furthermore, the objects + +have a close appearance, severe occlusion, and frequent crossovers. Considering our goal is to improve tracking robustness under occlusion and non-linear object motion, we would emphasize the comparison on DanceTrack. + +Implementations. For a fair comparison, we directly apply the object detections from existing baselines. For MOT17, MOT20, and DanceTrack, we use the publicly availableYOLOX [19] detector weights by ByteTrack [70]. For KITTI [20], we use the detections from PermaTrack [57] publicly available in the official release1. For ORU, we generate the virtual trajectory during occlusion with the constant-velocity assumption. Therefore, Eq. 6 is adopted as $\tilde{\mathbf{z}}_t = \mathbf{z}_{t_1} + \frac{t - t_1}{t_2 - t_1} (\mathbf{z}_{t_2} - \mathbf{z}_{t_1}), t_1 < t < t_2$ . For OCM, the velocity direction is calculated using the observations three time steps apart, i.e. $\Delta t = 3$ . The direction difference is measured by the absolute difference of angles in radians. We set $\lambda = 0.2$ in Eq. 8. Following the common practice of SORT, we set the detection confidence threshold at 0.4 for MOT20 and 0.6 for other datasets. The IoU threshold during association is 0.3. + +Metrics. We adopt HOTA [37] as the main metric as it maintains a proper balance between the accuracy of object detection and association [37]. We also emphasize AssA to evaluate the association performance. IDF1 is also used for association performance evaluation. Other metrics we report, such as MOTA, are highly related to detection performance. It is fair to use these metrics only when all methods use the same detections for tracking, which is referred to as "public tracking" as reported in Appendix C. + +# 5.2. Benchmark Results + +Here we report the benchmark results on multiple datasets. We put all methods that use the shared detection results in the blue blocks at the bottom of each table. + +MOT17 and MOT20. We report OC-SORT's performance on MOT17 and MOT20 in Table 1 and Table 2 using private detections. To make a fair comparison, we use the same detection as ByteTrack [70]. OC-SORT achieves performance comparable to other state-of-the-art methods. Our gains are especially significant in MOT20 under severe pedestrian occlusion, setting a state-of-the-art HOTA of 62.1. As our method is designed to be simple for better generalization, we do not use adaptive detection thresholds as in ByteTrack. Also, ByteTrack uses more detections of low-confidence to achieve higher MOTA scores but we keep the detection confidence threshold the same as on other datasets, which is the common practice in the community. We inherit the linear interpolation on the two datasets by baseline methods for a fair comparison. To more clearly discard the variance from the detector, we also perform public tracking on MOT17 and MOT20, which is reported in Table 12 and Table 13 in + +Table 1. Results on MOT17-test with the private detections. ByteTrack and OC-SORT share detections. + +
TrackerHOTA↑MOTA↑IDF1↑FP(104)↓FN(104)↓IDs↓Frag↓AssA↑AssR↑
FairMOT [71]59.373.772.32.7511.73,3038,07358.063.6
TransCt [67]54.573.262.22.3112.44,6149,51949.754.2
TransTrk [55]54.175.263.55.028.643,6034,87247.957.1
GRTU [60]62.074.975.03.2010.81,8121,82462.165.8
QDTrack [42]53.968.766.32.6614.73,3788,09152.757.2
MOTR [69]57.271.968.42.1113.62,1153,89755.859.2
PermaTr [57]55.573.868.92.9011.53,6996,13253.159.8
TransMOT [12]61.776.775.13.629.322,3467,71959.966.5
GTR [75]59.175.371.52.6811.02,859-61.6-
DST-Tracker [8]60.175.272.32.4211.02,729-62.1-
MeMOT [5]56.972.569.02.7211.52,724-55.2-
UniCom [68]61.777.275.55.017.335,379---
ByteTrack [70]63.180.377.32.558.372,1962,27762.068.2
OC-SORT63.278.077.51.5110.81,9502,04063.267.5
+ +Table 2. Results on MOT20-test with private detections. ByteTrack and OC-SORT share detections. + +
TrackerHOTA↑MOTA↑IDF1↑FP(104)↓FN(104)↓IDs↓Frag↓AssA↑AssR↑
FairMOT [71]54.661.867.310.38.895,2437,87454.760.7
TransCt [67]43.558.549.66.4214.64,6959,58137.045.1
Semi-TCL [35]55.365.270.16.1211.54,1398,50856.360.9
CSTrack [36]54.066.668.62.5414.43,1967,63254.057.6
GSDT [61]53.667.167.53.1913.53,1319,87552.758.5
TransMOT [12]61.977.575.23.428.081,6152,42160.166.3
MeMOT [5]54.163.766.14.7913.81,938-55.0-
ByteTrack [70]61.377.875.22.628.761,2231,46059.666.2
OC-SORT62.175.575.91.8010.89131,19862.067.5
+ +Table 3. Results on DanceTrack test set. Methods in the blue block share the same detections. + +
TrackerHOTA↑DetA↑AssA↑MOTA↑IDF1↑
CenterTrack [73]41.878.122.686.835.7
FairMOT [71]39.766.723.882.240.8
QDTrack [42]45.772.129.283.044.8
TransTrk[55]45.575.927.588.445.2
TraDes [64]43.374.525.486.241.2
MOTR [69]54.273.540.279.751.5
GTR [75]48.072.531.984.750.3
DST-Tracker [8]51.972.334.684.951.0
SORT [3]47.972.031.291.850.8
DeepSORT [63]45.671.029.787.847.9
ByteTrack [70]47.371.631.489.552.5
OC-SORT54.680.440.289.654.6
OC-SORT + Linear Interp55.180.440.492.254.9
+ +Appendix C. OC-SORT still outperforms the existing state-of-the-art in public tracking settings. + +DanceTrack. To evaluate OC-SORT under challenging non-linear object motion, we report results on the DanceTrack in Table 3. OC-SORT sets a new state-of-the-art, outperforming the baselines by a great margin under nonlinear object motions. We compare the tracking results of SORT and OC-SORT under extreme non-linear situations in Fig.1 and more samples are available in Fig.8 in Appendix E. We also visualize the output trajectories by OC-SORT and SORT on randomly selected DanceTrack video clips in Fig.9 in Appendix E. For multi-object tracking in occlusion and non-linear motion, the results on DanceTrack are strong evidence of the effectiveness of OC-SORT. + +KITTI. In Table 4 we report the results on the KITTI dataset. For a fair comparison, we adopt the detector weights by PermaTr [57] and report its performance in the + +table as well. We run OC-SORT given the shared detections. As initializing SORT's track requires continuous tracking across several frames ("minimum hits"), we observe that the results not recorded during the track initialization make a significant difference. To address this problem, we perform offline head padding (HP) post-processing by writing these entries back after finishing the online tracking stage. The results of the car category on KITTI show an essential shortcoming of the default implementation version of OC-SORT that it chooses the IoU matching for the association. When the objects move fast or the frame rate is low, the IoU of bounding boxes between consecutive frames can be very low or even zero. This issue does not come from the intrinsic design of OC-SORT and is widely observed when using IoU as the association cue. Adding other cues [49, 72, 73] and appearance similarity [38, 63] have been demonstrated [63] effective to solve this. In contrast to the relatively inferior car tracking performance, OC-SORT improves pedestrian tracking performance to a new state-of-the-art. Using the same detections, OC-SORT achieves a large performance gap over PermaTr with $10\mathrm{x}$ faster speed. + +The results on multiple benchmarks have demonstrated the effectiveness and efficiency of OC-SORT. We note that we use a shared parameter stack across datasets. Carefully tuning the parameters can probably further boost the performance. For example, the adaptive detection threshold is proven useful in previous work [70]. Besides the association accuracy, we also care about the inference speed. Given off-the-shelf detections, OC-SORT runs at 793 FPS on an Intel i9-9980XE CPU @ 3.00GHz. Therefore, OC-SORT + +Table 4. Results on KITTI-test. Our method uses the same detections as PermaTr [57] + +
CarPedestrian
TrackerHOTA↑MOTA↑AssA↑IDs↓Frag↓HOTA↑MOTA↑AssA↑IDs↓Frag↓
IMMDP [65]68.6682.7569.76211181-----
SMAT [21]71.8883.6472.13198294-----
TrackMPNN [45]72.3087.3370.6348123739.4052.1035.45626669
MPNTrack [4]-----45.2646.2347.283971,078
CenterTr [73]73.0288.8371.1825422740.3553.8436.93425618
LGM [59]73.1487.6072.31448164-----
TuSimple [11]71.5586.3171.1129221845.8857.6147.62246651
PermaTr [57]77.4290.8577.6627527147.4365.0543.66483703
OC-SORT74.6487.8174.5225731852.9562.0057.81181598
OC-SORT + HP76.5490.2876.3925028054.6965.1459.08184609
+ +Table 5. Ablation on MOT17-val and DanceTrack-val. + +
MOT17-valDanceTrack-val
ORUOCMOCRHOTA↑AssA↑IDF1↑HOTA↑AssA↑IDF1↑
64.966.876.947.831.048.3
✓66.368.077.248.532.249.8
✓✓66.469.077.852.135.050.6
✓✓✓66.568.977.752.135.351.6
+ +Table 6. Ablation on the trajectory hypothesis in ORU. + +
MOT17-valDanceTrack-val
HOTA↑AssA↑IDF1↑HOTA↑AssA↑IDF1↑
Const. Speed66.568.977.752.135.351.6
GPR63.165.275.749.533.749.6
Linear Regression64.366.576.049.333.449.2
Const. Acceleration66.267.977.451.334.850.9
+ +Table 7. Influence from the value of ${\Delta t}$ in OCM. + +
MOT17-valDanceTrack-val
HOTA↑AssA↑IDF1↑HOTA↑AssA↑IDF1↑
Δt = 166.167.576.951.334.351.3
Δt = 266.368.077.352.235.451.4
Δt = 366.568.977.752.135.351.6
Δt = 666.067.576.952.135.451.8
+ +can still run in an online and real-time fashion. + +# 5.3. Ablation Study + +Component Ablation. We ablate the contribution of proposed modules on the validation sets of MOT17 and DanceTrack in Table 5. The splitting of the MOT17 validation set follows a popular convention [73]. The results demonstrate the effectiveness of the proposed modules in OC-SORT. The results show that the performance gain from ORU is significant on both datasets but OCM only shows significant help on DanceTrack dataset where object motion is more complicated and the occlusion is heavy. The ablation study proves the effectiveness of our proposed method to improve tracking robustness in occlusion and non-linear motion. + +Virtual Trajectory in ORU. For simplicity, we follow the naive hypothesis of constant speed to generate a virtual trajectory in ORU. There are other alternatives like constant acceleration, regression-based fitting such as Linear Regression (LR) or Gaussian Process Regression (GPR), and Near Constant Acceleration Model (NCAM) [27]. The results of + +comparing these choices are shown in Table 6. For GPR, we use the RBF kernel [10] $k(\mathbf{x},\mathbf{x}^{\prime}) = \exp \left(-\frac{||\mathbf{x} - \mathbf{x}^{\prime}||^{2}}{50}\right)$ . We provide more studies on the kernel configuration in Appendix B. The results show that local hypotheses such as Constant Speed/Acceleration perform much better than global hypotheses such as LR and GPR. This is probably because, as virtual trajectory generation happens in an online fashion, it is hard to get a reliable fit using only limited data points on historical time steps. + +$\Delta t$ in OCM. There is a trade-off when choosing the time difference $\Delta t$ in OCM (Section 4). A large $\Delta t$ decreases the noise of velocity estimation, but is also likely to discourage approximating object motion as linear. Therefore, we study the influence of varying $\Delta t$ in Table 7. Our results agree with our analysis that increasing $\Delta t$ from $\Delta t = 1$ can boost performance. But increasing $\Delta t$ higher than a best-practice value instead hurts the performance because of the difficulty of maintaining the approximation of linear motion. + +# 6. Conclusion + +We analyze the popular motion-based tracker SORT and recognize its intrinsic limitations from using Kalman filter. These limitations significantly hurt tracking accuracy when the tracker fails to gain observations for supervision - likely caused by unreliable detectors, occlusion, or fast and non-linear target object motion. To address these issues, we propose Observation-Centric SORT (OC-SORT). OC-SORT is more robust to occlusion and non-linear object motion while keeping simple, online, and real-time. In our experiments on diverse datasets, OC-SORT significantly outperforms the state-of-the-art. The gain is especially significant for multi-object tracking under occlusion and non-linear object motion. + +Acknowledgement. We thank David Held, Deva Ramanan, and Siddharth Ancha for the discussion about the theoretical modeling of OC-SORT. We thank Yuda Song for the help with the analysis of OCM error distribution. We also would like to thank Zhengyi Luo, Yuda Song, and Erica Weng for the detailed feedback on the paper writing. We also thank Yifu Zhang for sharing his experience with ByteTrack. This project was sponsored in part by NSF NRI award 2024173. + +# References + +[1] Yaakov Bar-Shalom, Fred Daum, and Jim Huang. The probabilistic data association filter. IEEE Control Systems Magazine, 29(6):82-100, 2009. +[2] Sumit Basu, Irfan Essa, and Alex Pentland. Motion regularization for model-based head tracking. In Proceedings of 13th International Conference on Pattern Recognition, volume 3, pages 611-616. IEEE, 1996. +[3] Alex Bewley, Zongyuan Ge, Lionel Ott, Fabio Ramos, and Ben Upcroft. Simple online and realtime tracking. In 2016 IEEE international conference on image processing (ICIP), pages 3464-3468. IEEE, 2016. +[4] Guillem Brasó and Laura Leal-Taixe. Learning a neural solver for multiple object tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6247-6257, 2020. +[5] Jiarui Cai, Mingze Xu, Wei Li, Yuanjun Xiong, Wei Xia, Zhuowen Tu, and Stefano Soatto. Memot: multi-object tracking with memory. In CVPR, 2022. +[6] Jinkun Cao, Jiangmiao Pang, Xinshuo Weng, Rawal Khirodkar, and Kris M Kitani. Object tracking by hierarchical part-whole attention. +[7] Jinkun Cao, Hongyang Tang, Hao-Shu Fang, Xiaoyong Shen, Cewu Lu, and Yu-Wing Tai. Cross-domain adaptation for animal pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9498-9507, 2019. +[8] Jinkun Cao, Hao Wu, and Kris Kitani. Track targets by dense spatio-temporal position encoding. arXiv preprint arXiv:2210.09455, 2022. +[9] Ming-Fang Chang, John Lambert, Patsorn Sangkloy, Jagjeet Singh, Slawomir Bak, Andrew Hartnett, De Wang, Peter Carr, Simon Lucey, Deva Ramanan, et al. Argoverse: 3d tracking and forecasting with rich maps. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8748-8757, 2019. +[10] Yin-Wen Chang, Cho-Jui Hsieh, Kai-Wei Chang, Michael Ringgaard, and Chih-Jen Lin. Training and testing low-degree polynomial data mappings via linearsvm. Journal of Machine Learning Research, 11(4), 2010. +[11] Wongun Choi. Near-online multi-target tracking with aggregated local flow descriptor. In Proceedings of the IEEE international conference on computer vision, pages 3029-3037, 2015. +[12] Peng Chu, Jiang Wang, Quanzeng You, Haibin Ling, and Zicheng Liu. Transmot: Spatial-temporal graph transformer for multiple object tracking. arXiv preprint arXiv:2104.00194, 2021. +[13] Peng Dai, Renliang Weng, Wongun Choi, Changshui Zhang, Zhangping He, and Wei Ding. Learning a proposal classifier for multiple object tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2443-2452, 2021. +[14] Patrick Dendorfer, Hamid Rezatofighi, Anton Milan, Javen Shi, Daniel Cremers, Ian Reid, Stefan Roth, Konrad Schindler, and Laura Leal-Taixe. Mot20: A benchmark for multi object tracking in crowded scenes. arXiv preprint arXiv:2003.09003, 2020. +[15] David Duvenaud. Automatic model construction with Gaus + +sian processes. PhD thesis, University of Cambridge, 2014. +[16] Hao-Shu Fang, Jinkun Cao, Yu-Wing Tai, and Cewu Lu. Pairwise body-part attention for recognizing human-object interactions. In Proceedings of the European conference on computer vision (ECCV), pages 51-67, 2018. +[17] Juergen Gall, Angela Yao, Nima Razavi, Luc Van Gool, and Victor Lempitsky. Hough forests for object detection, tracking, and action recognition. IEEE transactions on pattern analysis and machine intelligence, 33(11):2188-2202, 2011. +[18] Damien Garreau, Wittawat Jitkrittum, and Motonobu Kanagawa. Large sample analysis of the median heuristic. arXiv preprint arXiv:1707.07269, 2017. +[19] Zheng Ge, Songtao Liu, Feng Wang, Zeming Li, and Jian Sun. Yolox: Exceeding yolo series in 2021. arXiv preprint arXiv:2107.08430, 2021. +[20] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. The International Journal of Robotics Research, 32(11):1231-1237, 2013. +[21] Nicolas Franco Gonzalez, Andres Ospina, and Philippe Calvez. Smat: Smart multiple affinity metrics for multiple object tracking. In International Conference on Image Analysis and Recognition, pages 48-62. Springer, 2020. +[22] Fredrik Gustafsson, Fredrik Gunnarsson, Niclas Bergman, Urban Forssell, Jonas Jansson, Rickard Karlsson, and P-J Nordlund. Particle filters for positioning, navigation, and tracking. IEEE Transactions on signal processing, 50(2):425-437, 2002. +[23] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. +[24] David V Hinkley. On the ratio of two correlated normal random variables. Biometrika, 56(3):635-639, 1969. +[25] Andrea Hornakova, Roberto Henschel, Bodo Rosenhahn, and Paul Swoboda. Lifted disjoint paths with application in multiple object tracking. In International Conference on Machine Learning, pages 4364-4375. PMLR, 2020. +[26] Andrea Hornakova, Timo Kaiser, Paul Swoboda, Michal Rolinek, Bodo Rosenhahn, and Roberto Henschel. Making higher order mot scalable: An efficient approximate solver for lifted disjoint paths. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6330-6340, 2021. +[27] Andrew H Jazwinski. Stochastic processes and filtering theory. Courier Corporation, 2007. +[28] Simon J Julier and Jeffrey K Uhlmann. New extension of the kalman filter to nonlinear systems. In Signal processing, sensor fusion, and target recognition VI, volume 3068, pages 182-193. International Society for Optics and Photonics, 1997. +[29] Simon J Julier and Jeffrey K Uhlmann. Unscented filtering and nonlinear estimation. Proceedings of the IEEE, 92(3):401-422, 2004. +[30] Rudolf Emil Kalman et al. Contributions to the theory of optimal control. Bol. soc. mat. mexicana, 5(2):102-119, 1960. +[31] Kris M Kitani, Brian D Ziebart, James Andrew Bagnell, and Martial Hebert. Activity forecasting. In European conference on computer vision, pages 201-214. Springer, 2012. + +[32] Jonathan Ko and Dieter Fox. Gp-bayesfilters: Bayesian filtering using gaussian process prediction and observation models. Autonomous Robots, 27(1):75-90, 2009. +[33] Parth Kothari, Sven Kreiss, and Alexandre Alahi. Human trajectory forecasting in crowds: A deep learning perspective. IEEE Transactions on Intelligent Transportation Systems, 2021. +[34] Erich L Lehmann and George Casella. Theory of point estimation. Springer Science & Business Media, 2006. +[35] Wei Li, Yuanjun Xiong, Shuo Yang, Mingze Xu, Yongxin Wang, and Wei Xia. Semi-tcl: Semi-supervised track contrastive representation learning. arXiv preprint arXiv:2107.02396, 2021. +[36] Chao Liang, Zhipeng Zhang, Yi Lu, Xue Zhou, Bing Li, Xiyong Ye, and Jianxiao Zou. Rethinking the competition between detection and reid in multi-object tracking. arXiv preprint arXiv:2010.12138, 2020. +[37] Jonathon Luiten, Aljosa Osep, Patrick Dendorfer, Philip Torr, Andreas Geiger, Laura Leal-Taixe, and Bastian Leibe. Hota: A higher order metric for evaluating multi-object tracking. International journal of computer vision, 129(2):548-578, 2021. +[38] Gerard Maggiolino, Adnan Ahmad, Jinkun Cao, and Kris Kitani. Deep oc-sort: Multi-pedestrian tracking by adaptive re-identification. arXiv preprint arXiv:2302.11813, 2023. +[39] Tim Meinhardt, Alexander Kirillov, Laura Leal-Taixe, and Christoph Feichtenhofer. Trackformer: Multi-object tracking with transformers. arXiv preprint arXiv:2101.02702, 2021. +[40] Stan Melax, Leonid Keselman, and Sterling Orsten. Dynamics based 3d skeletal hand tracking. In Proceedings of the ACM SIGGRAPH Symposium on Interactive 3D Graphics and Games, pages 184-184, 2013. +[41] Anton Milan, Laura Leal-Taixe, Ian Reid, Stefan Roth, and Konrad Schindler. Mot16: A benchmark for multi-object tracking. arXiv preprint arXiv:1603.00831, 2016. +[42] Jiangmiao Pang, Linlu Qiu, Xia Li, Haofeng Chen, Qi Li, Trevor Darrell, and Fisher Yu. Quasi-dense similarity learning for multiple object tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 164-173, 2021. +[43] Dezhi Peng, Zikai Sun, Zirong Chen, Zirui Cai, Lele Xie, and Lianwen Jin. Detecting heads using feature refine net and cascaded multi-scale architecture. arXiv preprint arXiv:1803.09256, 2018. +[44] Lawrence Rabiner and Biinghwang Juang. An introduction to hidden markov models. *iiee assp magazine*, 3(1):4-16, 1986. +[45] Akshay Rangesh, Pranav Maheshwari, Mez Gebre, Siddhesh Mhatre, Vahid Ramezani, and Mohan M Trivedi. Trackmpnn: A message passing graph neural architecture for multi-object tracking. arXiv preprint arXiv:2101.04206, 2021. +[46] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 779-788, 2016. +[47] Steven Reece and Stephen Roberts. An introduction to gaussian processes for the kalman filter expert. In 2010 13th In- + +ternational Conference on Information Fusion, pages 1-9. IEEE, 2010. +[48] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015. +[49] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 658-666, 2019. +[50] Toby Sharp, Cem Keskin, Duncan Robertson, Jonathan Taylor, Jamie Shotton, David Kim, Christoph Rhemann, Ido Leichter, Alon Vinnikov, Yichen Wei, et al. Accurate, robust, and flexible real-time hand tracking. In Proceedings of the 33rd annual ACM conference on human factors in computing systems, pages 3633-3642, 2015. +[51] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. +[52] Gerald L Smith, Stanley F Schmidt, and Leonard A McGee. Application of statistical filter theory to the optimal estimation of position and velocity on board a circumlunar vehicle. National Aeronautics and Space Administration, 1962. +[53] Daniel Stadler and Jurgen Beyerer. Improving multiple pedestrian tracking by track management and occlusion handling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10958-10967, 2021. +[54] Peize Sun, Jinkun Cao, Yi Jiang, Zehuan Yuan, Song Bai, Kris Kitani, and Ping Luo. Dancetrack: Multi-object tracking in uniform appearance and diverse motion. arXiv preprint arXiv:2111.14690, 2021. +[55] Peize Sun, Jinkun Cao, Yi Jiang, Rufeng Zhang, Enze Xie, Zehuan Yuan, Changhu Wang, and Ping Luo. Transtrack: Multiple object tracking with transformer. arXiv preprint arXiv:2012.15460, 2020. +[56] Ramana Sundararaman, Cedric De Almeida Braga, Eric Marchand, and Julien Pette. Tracking pedestrian heads in dense crowd. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3865-3875, 2021. +[57] Pavel Tokmakov, Jie Li, Wolfram Burgard, and Adrien Gaidon. Learning to track with object permanence. In ICCV, pages 10860-10869, 2021. +[58] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. +[59] Gaoang Wang, Renshu Gu, Zuozhu Liu, Weijie Hu, Mingli Song, and Jenq-Neng Hwang. Track without appearance: Learn box and tracklet embedding with local and global motion patterns for vehicle tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9876-9886, 2021. +[60] Shuai Wang, Hao Sheng, Yang Zhang, Yubin Wu, and Zhang Xiong. A general recurrent tracking framework without real data. In Proceedings of the IEEE/CVF International Confer- + +ence on Computer Vision, pages 13219-13228, 2021. +[61] Yongxin Wang, Kris Kitani, and Xinshuo Weng. Joint object detection and multi-object tracking with graph neural networks. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 13708-13715. IEEE, 2021. +[62] Christopher Williams and Carl Rasmussen. Gaussian processes for regression. Advances in neural information processing systems, 8, 1995. +[63] Nicolai Wojke, Alex Bewley, and Dietrich Paulus. Simple online and realtime tracking with a deep association metric. In 2017 IEEE international conference on image processing (ICIP), pages 3645-3649. IEEE, 2017. +[64] Jialian Wu, Jiale Cao, Liangchen Song, Yu Wang, Ming Yang, and Junsong Yuan. Track to detect and segment: An online multi-object tracker. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12352-12361, 2021. +[65] Yu Xiang, Alexandre Alahi, and Silvio Savarese. Learning to track: Online multi-object tracking by decision making. In Proceedings of the IEEE international conference on computer vision, pages 4705-4713, 2015. +[66] Yuliang Xiu, Jiefeng Li, Haoyu Wang, Yinghong Fang, and Cewu Lu. Pose flow: Efficient online pose tracking. arXiv preprint arXiv:1802.00977, 2018. +[67] Yihong Xu, Yutong Ban, Guillaume Delorme, Chuang Gan, Daniela Rus, and Xavier Alameda-Pineda. Transcenter: Transformers with dense queries for multiple-object tracking. arXiv preprint arXiv:2103.15145, 2021. +[68] Bin Yan, Yi Jiang, Peize Sun, Dong Wang, Zehuan Yuan, Ping Luo, and Huchuan Lu. Towards grand unification of object tracking. In ECCV. Springer, 2022. +[69] Fangao Zeng and et al. Motr: End-to-end multiple-object tracking with transformer. In ECCV. Springer, 2022. +[70] Yifu Zhang, Peize Sun, Yi Jiang, Dongdong Yu, Zehuan Yuan, Ping Luo, Wenyu Liu, and Xinggang Wang. Byetetrack: Multi-object tracking by associating every detection box. arXiv preprint arXiv:2110.06864, 2021. +[71] Yifu Zhang, Chunyu Wang, Xinggang Wang, Wenjun Zeng, and Wenyu Liu. Fairmot: On the fairness of detection and re-identification in multiple object tracking. International Journal of Computer Vision, 129(11):3069-3087, 2021. +[72] Zhaohui Zheng, Ping Wang, Wei Liu, Jinze Li, Rongguang Ye, and Dongwei Ren. Distance-iou loss: Faster and better learning for bounding box regression. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 12993-13000, 2020. +[73] Xingyi Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Tracking objects as points. In ECCV, pages 474-490. Springer, 2020. +[74] Xingyi Zhou, Dequan Wang, and Philipp Kähenbuhl. Objects as points. arXiv preprint arXiv:1904.07850, 2019. +[75] Xingyi Zhou, Tianwei Yin, Vladlen Koltun, and Philipp Krahenbuhl. Global tracking transformers. In CVPR, 2022. \ No newline at end of file diff --git a/2023/Observation-Centric SORT_ Rethinking SORT for Robust Multi-Object Tracking/images.zip b/2023/Observation-Centric SORT_ Rethinking SORT for Robust Multi-Object Tracking/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..aa28375b75b829c3060208bb95071d8b0ffc8aaa --- /dev/null +++ b/2023/Observation-Centric SORT_ Rethinking SORT for Robust Multi-Object Tracking/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e894baa10a1b54bfc99f80116c0dd4951b8f0c04c25e0a16940deb0bf7b56bb6 +size 536715 diff --git a/2023/Observation-Centric SORT_ Rethinking SORT for Robust Multi-Object Tracking/layout.json b/2023/Observation-Centric SORT_ Rethinking SORT for Robust Multi-Object Tracking/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..884e8bffd1c5db36a996bb9cea3db286fa7b8e65 --- /dev/null +++ b/2023/Observation-Centric SORT_ Rethinking SORT for Robust Multi-Object Tracking/layout.json @@ -0,0 +1,9428 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 134, + 103, + 460, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 103, + 460, + 140 + ], + "spans": [ + { + "bbox": [ + 134, + 103, + 460, + 140 + ], + "type": "text", + "content": "Observation-Centric SORT: Rethinking SORT for Robust Multi-Object Tracking" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 85, + 160, + 508, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 160, + 508, + 190 + ], + "spans": [ + { + "bbox": [ + 85, + 160, + 508, + 190 + ], + "type": "text", + "content": "Jinkun Cao" + }, + { + "bbox": [ + 85, + 160, + 508, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 85, + 160, + 508, + 190 + ], + "type": "text", + "content": " Jiangmiao Pang" + }, + { + "bbox": [ + 85, + 160, + 508, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 85, + 160, + 508, + 190 + ], + "type": "text", + "content": " Xinshuo Weng" + }, + { + "bbox": [ + 85, + 160, + 508, + 190 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 85, + 160, + 508, + 190 + ], + "type": "text", + "content": " Rawal Khirodkar" + }, + { + "bbox": [ + 85, + 160, + 508, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 85, + 160, + 508, + 190 + ], + "type": "text", + "content": " Kris Kitani" + }, + { + "bbox": [ + 85, + 160, + 508, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 85, + 160, + 508, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 85, + 160, + 508, + 190 + ], + "type": "text", + "content": " Carnegie Mellon University " + }, + { + "bbox": [ + 85, + 160, + 508, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 85, + 160, + 508, + 190 + ], + "type": "text", + "content": " Shanghai AI Laboratory " + }, + { + "bbox": [ + 85, + 160, + 508, + 190 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 85, + 160, + 508, + 190 + ], + "type": "text", + "content": " Nvidia" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 217, + 192, + 229 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 217, + 192, + 229 + ], + "spans": [ + { + "bbox": [ + 143, + 217, + 192, + 229 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 242, + 290, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 242, + 290, + 578 + ], + "spans": [ + { + "bbox": [ + 47, + 242, + 290, + 578 + ], + "type": "text", + "content": "Kalman filter (KF) based methods for multi-object tracking (MOT) make an assumption that objects move linearly. While this assumption is acceptable for very short periods of occlusion, linear estimates of motion for prolonged time can be highly inaccurate. Moreover, when there is no measurement available to update Kalman filter parameters, the standard convention is to trust the priori state estimations for posteriori update. This leads to the accumulation of errors during a period of occlusion. The error causes significant motion direction variance in practice. In this work, we show that a basic Kalman filter can still obtain state-of-the-art tracking performance if proper care is taken to fix the noise accumulated during occlusion. Instead of relying only on the linear state estimate (i.e., estimation-centric approach), we use object observations (i.e., the measurements by object detector) to compute a virtual trajectory over the occlusion period to fix the error accumulation of filter parameters. This allows more time steps to correct errors accumulated during occlusion. We name our method Observation-Centric SORT (OC-SORT). It remains Simple, Online, and Real-Time but improves robustness during occlusion and non-linear motion. Given off-the-shelf detections as input, OC-SORT runs at " + }, + { + "bbox": [ + 47, + 242, + 290, + 578 + ], + "type": "inline_equation", + "content": "700+" + }, + { + "bbox": [ + 47, + 242, + 290, + 578 + ], + "type": "text", + "content": " FPS on a single CPU. It achieves state-of-the-art on multiple datasets, including MOT17, MOT20, KITTI, head tracking, and especially DanceTrack where the object motion is highly non-linear. The code and models are available at https://github.com/noahcao/OC_SORT." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 597, + 128, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 597, + 128, + 609 + ], + "spans": [ + { + "bbox": [ + 47, + 597, + 128, + 609 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 617, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 715 + ], + "type": "text", + "content": "We aim to develop a motion model-based multi-object tracking (MOT) method that is robust to occlusion and nonlinear motion. Most existing motion model-based algorithms assume that the tracking targets have a constant velocity within a time interval, which is called the linear motion assumption. This assumption breaks in many practical scenarios, but it still works because when the time interval is small enough, the object's motion can be reasonably ap-" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 306, + 215, + 364, + 298 + ], + "blocks": [ + { + "bbox": [ + 306, + 215, + 364, + 298 + ], + "lines": [ + { + "bbox": [ + 306, + 215, + 364, + 298 + ], + "spans": [ + { + "bbox": [ + 306, + 215, + 364, + 298 + ], + "type": "image", + "image_path": "5a0838cea78f29b37d62ae49f4eb3b995510ba9830aa17a2fce3e36949e545bf.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 365, + 215, + 423, + 298 + ], + "blocks": [ + { + "bbox": [ + 365, + 215, + 423, + 298 + ], + "lines": [ + { + "bbox": [ + 365, + 215, + 423, + 298 + ], + "spans": [ + { + "bbox": [ + 365, + 215, + 423, + 298 + ], + "type": "image", + "image_path": "bfb6f1f89cbd0d313250928330757f39596ecd085dcca59a94ee799b5a3b1769.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 408, + 299, + 440, + 308 + ], + "lines": [ + { + "bbox": [ + 408, + 299, + 440, + 308 + ], + "spans": [ + { + "bbox": [ + 408, + 299, + 440, + 308 + ], + "type": "text", + "content": "(a) SORT" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 425, + 216, + 482, + 298 + ], + "blocks": [ + { + "bbox": [ + 425, + 216, + 482, + 298 + ], + "lines": [ + { + "bbox": [ + 425, + 216, + 482, + 298 + ], + "spans": [ + { + "bbox": [ + 425, + 216, + 482, + 298 + ], + "type": "image", + "image_path": "6b3834b6036c88f9cf10417ad67ee7e3226beb078ebbf1d6d7e4eb7d4d4cbcb8.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 484, + 216, + 541, + 298 + ], + "blocks": [ + { + "bbox": [ + 484, + 216, + 541, + 298 + ], + "lines": [ + { + "bbox": [ + 484, + 216, + 541, + 298 + ], + "spans": [ + { + "bbox": [ + 484, + 216, + 541, + 298 + ], + "type": "image", + "image_path": "8b044f0ea30ee14e9c9801d8db562cf61f06ab5e620e555afc0d9608dbf18400.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 307, + 310, + 364, + 392 + ], + "blocks": [ + { + "bbox": [ + 307, + 310, + 364, + 392 + ], + "lines": [ + { + "bbox": [ + 307, + 310, + 364, + 392 + ], + "spans": [ + { + "bbox": [ + 307, + 310, + 364, + 392 + ], + "type": "image", + "image_path": "3a16e705f0b020b43a39f9919535d2fc7dbcd1042668bbe2cd15f5a1ad57864c.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 365, + 310, + 422, + 391 + ], + "blocks": [ + { + "bbox": [ + 365, + 310, + 422, + 391 + ], + "lines": [ + { + "bbox": [ + 365, + 310, + 422, + 391 + ], + "spans": [ + { + "bbox": [ + 365, + 310, + 422, + 391 + ], + "type": "image", + "image_path": "5e9472c633d3813f1c0e4d0d38b6044a4cc65696afeaf1271b2e371ca652d7ff.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 377, + 392, + 470, + 403 + ], + "lines": [ + { + "bbox": [ + 377, + 392, + 470, + 403 + ], + "spans": [ + { + "bbox": [ + 377, + 392, + 470, + 403 + ], + "type": "text", + "content": "(b) The proposed OC-SORT" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 425, + 310, + 482, + 391 + ], + "blocks": [ + { + "bbox": [ + 425, + 310, + 482, + 391 + ], + "lines": [ + { + "bbox": [ + 425, + 310, + 482, + 391 + ], + "spans": [ + { + "bbox": [ + 425, + 310, + 482, + 391 + ], + "type": "image", + "image_path": "b92475af28c88b8ee65bab550c8ef4291a7393e219ed5930f5d523310ac6e6d3.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 412, + 547, + 460 + ], + "lines": [ + { + "bbox": [ + 305, + 412, + 547, + 460 + ], + "spans": [ + { + "bbox": [ + 305, + 412, + 547, + 460 + ], + "type": "text", + "content": "Figure 1. Samples from the results on DanceTrack [54]. SORT and OC-SORT use the same detection results. On the third frame, SORT encounters an ID switch for the backflip target while OC-SORT tracks it consistently." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 484, + 310, + 541, + 391 + ], + "blocks": [ + { + "bbox": [ + 484, + 310, + 541, + 391 + ], + "lines": [ + { + "bbox": [ + 484, + 310, + 541, + 391 + ], + "spans": [ + { + "bbox": [ + 484, + 310, + 541, + 391 + ], + "type": "image", + "image_path": "86515da1892e887e1bbb101381a5768a76a55a57a3f9b1affbd9e1d6d1ff1352.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 484, + 547, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 484, + 547, + 568 + ], + "spans": [ + { + "bbox": [ + 304, + 484, + 547, + 568 + ], + "type": "text", + "content": "proximated as linear. In this work, we are motivated by the fact that most of the errors from motion model-based tracking methods occur when occlusion and non-linear motion happen together. To mitigate the adverse effects caused, we first rethink current motion models and recognize some limitations. Then, we propose addressing them for more robust tracking performance, especially in occlusion." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 547, + 713 + ], + "type": "text", + "content": "As the main branch of motion model-based tracking, filtering-based methods assume a transition function to predict the state of objects on future time steps, which are called state \"estimations\". Besides estimations, they leverage an observation model, such as an object detector, to derive the state measurements of target objects, also called \"observations\". Observations usually serve as auxiliary information to help update the posteriori parameters of the filter. The trajectories are still extended by the state estimations. Among this line of work, the most widely used one is SORT [3], which uses a Kalman filter (KF) to estimate object states and a linear motion function as the transition" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9686" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 286, + 119 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 286, + 119 + ], + "type": "text", + "content": "function between time steps. However, SORT shows insufficient tracking robustness when the object motion is nonlinear, and no observations are available when updating the filter posteriori parameters." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 122, + 286, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 122, + 286, + 384 + ], + "spans": [ + { + "bbox": [ + 49, + 122, + 286, + 384 + ], + "type": "text", + "content": "In this work, we recognize three limitations of SORT. First, although the high frame rate is the key to approximating the object motion as linear, it also amplifies the model's sensitivity to the noise of state estimations. Specifically, between consecutive frames of a high frame-rate video, we demonstrate that the noise of displacement of the object can be of the same magnitude as the actual object displacement, leading to the estimated object velocity by KF suffering from a significant variance. Also, the noise in the velocity estimate will accumulate into the position estimate by the transition process. Second, the noise of state estimations by KF is accumulated along the time when there is no observation available in the update stage of KF. We show that the error accumulates very fast with respect to the time of the target object's being untracked. The noise's influence on the velocity direction often makes the track lost again even after re-association. Last, given the development of modern detectors, the object state by detections usually has lower variance than the state estimations propagated along time steps by a fixed transition function in filters. However, SORT is designed to prolong the object trajectories by state estimations instead of observations." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 387, + 286, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 387, + 286, + 553 + ], + "spans": [ + { + "bbox": [ + 49, + 387, + 286, + 553 + ], + "type": "text", + "content": "To relieve the negative effect of these limitations, we propose two main innovations in this work. First, we design a module to use object state observations to reduce the accumulated error during the track's being lost in a backcheck fashion. To be precise, besides the traditional stages of predict and update, we add a stage of re-update to correct the accumulated error. The re-update is triggered when a track is re-activated by associating to an observation after a period of being untracked. The re-update uses virtual observations on the historical time steps to prevent error accumulation. The virtual observations come from a trajectory generated using the last-seen observation before untracked and the latest observation re-activating this track as anchors. We name it Observation-centric Re-Update (ORU)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 556, + 286, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 556, + 286, + 663 + ], + "spans": [ + { + "bbox": [ + 49, + 556, + 286, + 663 + ], + "type": "text", + "content": "Besides ORU, the assumption of linear motion provides the consistency of the object motion direction. But this cue is hard to be used in SORT's association because of the heavy noise in direction estimation. But we propose an observation-centric manner to incorporate the direction consistency of tracks in the cost matrix for the association. We name it Observation-Centric Momentum (OCM). We also provide analytical justification for the noise of velocity direction estimation in practice." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 666, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 666, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 666, + 286, + 712 + ], + "type": "text", + "content": "The proposed method, named as Observation-Centric SORT or OC-SORT in short, remains simple, online, real-time and significantly improves robustness over occlusion and non-linear motion. Our contributions are summarized" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 309, + 73, + 375, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 73, + 375, + 84 + ], + "spans": [ + { + "bbox": [ + 309, + 73, + 375, + 84 + ], + "type": "text", + "content": "as the following:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 315, + 88, + 544, + 185 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 316, + 88, + 544, + 135 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 88, + 544, + 135 + ], + "spans": [ + { + "bbox": [ + 316, + 88, + 544, + 135 + ], + "type": "text", + "content": "1. We recognize, analytically and empirically, three limitations of SORT, i.e. sensitivity to the noise of state estimations, error accumulation over time, and being estimation-centric;" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 315, + 138, + 544, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 138, + 544, + 185 + ], + "spans": [ + { + "bbox": [ + 315, + 138, + 544, + 185 + ], + "type": "text", + "content": "2. We propose OC-SORT for tracking under occlusion and non-linear motion by fixing SORT's limitations. It achieves state-of-the-art performance on multiple datasets in an online and real-time fashion." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 309, + 207, + 395, + 218 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 309, + 207, + 395, + 218 + ], + "spans": [ + { + "bbox": [ + 309, + 207, + 395, + 218 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 308, + 229, + 544, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 229, + 544, + 468 + ], + "spans": [ + { + "bbox": [ + 308, + 229, + 544, + 468 + ], + "type": "text", + "content": "Motion Models. Many modern MOT algorithms [3, 11, 63, 70, 73] use motion models. Typically, these motion models use Bayesian estimation [34] to predict the next state by maximizing a posterior estimation. As one of the most classic motion models, Kalman filter (KF) [30] is a recursive Bayes filter that follows a typical predict-update cycle. The true state is assumed to be an unobserved Markov process, and the measurements are observations from a hidden Markov model [44]. Given that the linear motion assumption limits KF, follow-up works like Extended KF [52] and Unscented KF [28] were proposed to handle non-linear motion with first-order and third-order Taylor approximation. However, they still rely on approximating the Gaussian prior assumed by KF and require motion pattern assumption. On the other hand, particle filters [22] solve the non-linear motion by sampling-based posterior estimation but require exponential order of computation. Therefore, these variants of Kalman filter and particle filters are rarely adopted in the visual multi-object tracking and the mostly adopted motion model is still based on Kalman filter [3]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 308, + 486, + 544, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 486, + 544, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 486, + 544, + 712 + ], + "type": "text", + "content": "Multi-object Tracking. As a classic computer vision task, visual multi-object tracking is traditionally approached from probabilistic perspectives, e.g. joint probabilistic association [1]. And modern video object tracking is usually built upon modern object detectors [46, 48, 74]. SORT [3] adopts the Kalman filter for motion-based multi-object tracking given observations from deep detectors. DeepSORT [63] further introduces deep visual features [23, 51] into object association under the framework of SORT. Re-identification-based object association[42, 63, 71] has also become popular since then but falls short when scenes are crowded and objects are represented coarsely (e.g. enclosed by bounding boxes), or object appearance is not distinguishable. More recently, transformers [58] have been introduced to MOT [8, 39, 55, 69] to learn deep representations from both visual information and object trajectories. However, their performance still has a significant gap between state-of-the-art tracking-by-detection methods in terms of both accuracy and time efficiency." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 750, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 750, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 750, + 314, + 757 + ], + "type": "text", + "content": "9687" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 71, + 234, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 71, + 234, + 84 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 234, + 84 + ], + "type": "text", + "content": "3. Rethink the Limitations of SORT" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 91, + 287, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 91, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 47, + 91, + 287, + 140 + ], + "type": "text", + "content": "In this section, we review Kalman filter and SORT [3]. We recognize some of their limitations, which are significant with occlusion and non-linear object motion. We are motivated to improve tracking robustness by fixing them." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 145, + 133, + 158 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 145, + 133, + 158 + ], + "spans": [ + { + "bbox": [ + 47, + 145, + 133, + 158 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 164, + 289, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 164, + 289, + 297 + ], + "spans": [ + { + "bbox": [ + 47, + 164, + 289, + 297 + ], + "type": "text", + "content": "Kalman filter (KF) [30] is a linear estimator for dynamical systems discretized in the time domain. KF only requires the state estimations on the previous time step and the current measurement to estimate the target state on the next time step. The filter maintains two variables, the posteriori state estimate " + }, + { + "bbox": [ + 47, + 164, + 289, + 297 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 47, + 164, + 289, + 297 + ], + "type": "text", + "content": ", and the posteriori estimate covariance matrix " + }, + { + "bbox": [ + 47, + 164, + 289, + 297 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 47, + 164, + 289, + 297 + ], + "type": "text", + "content": ". In the task of object tracking, we describe the KF process with the state transition model " + }, + { + "bbox": [ + 47, + 164, + 289, + 297 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 47, + 164, + 289, + 297 + ], + "type": "text", + "content": ", the observation model " + }, + { + "bbox": [ + 47, + 164, + 289, + 297 + ], + "type": "inline_equation", + "content": "\\mathbf{H}" + }, + { + "bbox": [ + 47, + 164, + 289, + 297 + ], + "type": "text", + "content": ", the process noise " + }, + { + "bbox": [ + 47, + 164, + 289, + 297 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 47, + 164, + 289, + 297 + ], + "type": "text", + "content": ", and the observation noise " + }, + { + "bbox": [ + 47, + 164, + 289, + 297 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 47, + 164, + 289, + 297 + ], + "type": "text", + "content": ". At each step " + }, + { + "bbox": [ + 47, + 164, + 289, + 297 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 164, + 289, + 297 + ], + "type": "text", + "content": ", given observations " + }, + { + "bbox": [ + 47, + 164, + 289, + 297 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t" + }, + { + "bbox": [ + 47, + 164, + 289, + 297 + ], + "type": "text", + "content": ", KF works in an alternation of predict and update stages:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 52, + 302, + 287, + 381 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 302, + 287, + 381 + ], + "spans": [ + { + "bbox": [ + 52, + 302, + 287, + 381 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\text {p r e d i c t} \\left\\{ \\begin{array}{l} \\hat {\\mathbf {x}} _ {t | t - 1} = \\mathbf {F} _ {t} \\hat {\\mathbf {x}} _ {t - 1 | t - 1} \\\\ \\mathbf {P} _ {t | t - 1} = \\mathbf {F} _ {t} \\mathbf {P} _ {t - 1 | t - 1} \\mathbf {F} _ {t} ^ {\\top} + \\mathbf {Q} _ {t} \\end{array} , \\right. \\\\ \\text {u p d a t e} \\left\\{ \\begin{array}{l} \\mathbf {K} _ {t} = \\mathbf {P} _ {t | t - 1} \\mathbf {H} _ {t} ^ {\\top} \\left(\\mathbf {H} _ {t} \\mathbf {P} _ {t | t - 1} \\mathbf {H} _ {t} ^ {\\top} + \\mathbf {R} _ {t}\\right) ^ {- 1} \\quad (1) \\\\ \\hat {\\mathbf {x}} _ {t | t} = \\hat {\\mathbf {x}} _ {t | t - 1} + \\mathbf {K} _ {t} (\\mathbf {z} _ {t} - \\mathbf {H} _ {t} \\hat {\\mathbf {x}} _ {t | t - 1}) \\\\ \\mathbf {P} _ {t | t} = (\\mathbf {I} - \\mathbf {K} _ {t} \\mathbf {H} _ {t}) \\mathbf {P} _ {t | t - 1} \\end{array} . \\right. \\\\ \\end{array}", + "image_path": "8073bcb30cd394abd9110f226821f7d374ecb1bca42e5a959a86ce582b6d5391.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 387, + 287, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 387, + 287, + 457 + ], + "spans": [ + { + "bbox": [ + 47, + 387, + 287, + 457 + ], + "type": "text", + "content": "The stage of predict is to derive the state estimations on the next time step " + }, + { + "bbox": [ + 47, + 387, + 287, + 457 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 387, + 287, + 457 + ], + "type": "text", + "content": ". Given a measurement of target states on the next step " + }, + { + "bbox": [ + 47, + 387, + 287, + 457 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 47, + 387, + 287, + 457 + ], + "type": "text", + "content": ", the stage of update aims to update the posteriori parameters in KF. Because the measurement comes from the observation model " + }, + { + "bbox": [ + 47, + 387, + 287, + 457 + ], + "type": "inline_equation", + "content": "\\mathbf{H}" + }, + { + "bbox": [ + 47, + 387, + 287, + 457 + ], + "type": "text", + "content": ", it is also called \"observation\" in many scenarios." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 458, + 287, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 458, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 47, + 458, + 287, + 590 + ], + "type": "text", + "content": "SORT [3] is a multi-object tracker built upon KF. The KF's state " + }, + { + "bbox": [ + 47, + 458, + 287, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 47, + 458, + 287, + 590 + ], + "type": "text", + "content": " in SORT is defined as " + }, + { + "bbox": [ + 47, + 458, + 287, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{x} = [u,v,s,r,\\dot{u},\\dot{v},\\dot{s}]^{\\top}" + }, + { + "bbox": [ + 47, + 458, + 287, + 590 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 458, + 287, + 590 + ], + "type": "inline_equation", + "content": "(u,v)" + }, + { + "bbox": [ + 47, + 458, + 287, + 590 + ], + "type": "text", + "content": " is the 2D coordinates of the object center in the image. " + }, + { + "bbox": [ + 47, + 458, + 287, + 590 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 47, + 458, + 287, + 590 + ], + "type": "text", + "content": " is the bounding box scale (area) and " + }, + { + "bbox": [ + 47, + 458, + 287, + 590 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 47, + 458, + 287, + 590 + ], + "type": "text", + "content": " is the bounding box aspect ratio. The aspect ratio " + }, + { + "bbox": [ + 47, + 458, + 287, + 590 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 47, + 458, + 287, + 590 + ], + "type": "text", + "content": " is assumed to be constant. The other three variables, " + }, + { + "bbox": [ + 47, + 458, + 287, + 590 + ], + "type": "inline_equation", + "content": "\\dot{u},\\dot{v}" + }, + { + "bbox": [ + 47, + 458, + 287, + 590 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 458, + 287, + 590 + ], + "type": "inline_equation", + "content": "\\dot{s}" + }, + { + "bbox": [ + 47, + 458, + 287, + 590 + ], + "type": "text", + "content": " are the corresponding time derivatives. The observation is a bounding box " + }, + { + "bbox": [ + 47, + 458, + 287, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{z} = [u,v,w,h,c]^{\\top}" + }, + { + "bbox": [ + 47, + 458, + 287, + 590 + ], + "type": "text", + "content": " with object center position " + }, + { + "bbox": [ + 47, + 458, + 287, + 590 + ], + "type": "inline_equation", + "content": "(u,v)" + }, + { + "bbox": [ + 47, + 458, + 287, + 590 + ], + "type": "text", + "content": ", object width " + }, + { + "bbox": [ + 47, + 458, + 287, + 590 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 47, + 458, + 287, + 590 + ], + "type": "text", + "content": ", and height " + }, + { + "bbox": [ + 47, + 458, + 287, + 590 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 47, + 458, + 287, + 590 + ], + "type": "text", + "content": " and the detection confidence " + }, + { + "bbox": [ + 47, + 458, + 287, + 590 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 47, + 458, + 287, + 590 + ], + "type": "text", + "content": " respectively. SORT assumes linear motion as the transition model " + }, + { + "bbox": [ + 47, + 458, + 287, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 47, + 458, + 287, + 590 + ], + "type": "text", + "content": " which leads to the state estimation as" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 83, + 597, + 287, + 611 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 597, + 287, + 611 + ], + "spans": [ + { + "bbox": [ + 83, + 597, + 287, + 611 + ], + "type": "interline_equation", + "content": "u _ {t + 1} = u _ {t} + \\dot {u} _ {t} \\Delta t, \\quad v _ {t + 1} = v _ {t} + \\dot {v} _ {t} \\Delta t. \\tag {2}", + "image_path": "950ce0cb8911bc48d7eec0bc0d3f3e993d9a1f17d5f41437edafb48f0e38d268.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 617, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 617, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 47, + 617, + 287, + 689 + ], + "type": "text", + "content": "To leverage KF (Eq 1) in SORT for visual MOT, the stage of predict corresponds to estimating the object position on the next video frame. And the observations used for the update stage usually come from a detection model. The update stage is to update Kalman filter parameters and does not directly edit the tracking outcomes." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "When the time difference between two steps is constant during the transition, " + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "e.g." + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": ", the video frame rate is constant," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "text", + "content": "we can set " + }, + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "inline_equation", + "content": "\\Delta t = 1" + }, + { + "bbox": [ + 304, + 72, + 545, + 193 + ], + "type": "text", + "content": ". When the video frame rate is high, SORT works well even when the object motion is non-linear globally, (e.g. dancing, fencing, wrestling) because the motion of the target object can be well approximated as linear within short time intervals. However, in practice, observations are often absent on some time steps, e.g. the target object is occluded in multi-object tracking. In such cases, we cannot update the KF parameters by the update operation as in Eq. 1 anymore. SORT uses the priori estimations directly as posterior. We call this \"dummy update\", namely" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 365, + 201, + 545, + 216 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 201, + 545, + 216 + ], + "spans": [ + { + "bbox": [ + 365, + 201, + 545, + 216 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {x}} _ {t \\mid t} = \\hat {\\mathbf {x}} _ {t \\mid t - 1}, \\mathbf {P} _ {t \\mid t} = \\mathbf {P} _ {t \\mid t - 1}. \\tag {3}", + "image_path": "7e4d1e543ac62f5cb6bc0a47b88af657624e5133e54d0aff168c957e05b3c27c.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 223, + 545, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 223, + 545, + 295 + ], + "spans": [ + { + "bbox": [ + 304, + 223, + 545, + 295 + ], + "type": "text", + "content": "The philosophy behind such a design is to trust estimations when no observations are available to supervise them. We thus call the tracking algorithms following this scheme \"estimation-centric\". However, we will see that this estimation-centric mechanism can cause trouble when nonlinear motion and occlusion happen together." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 302, + 427, + 314 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 302, + 427, + 314 + ], + "spans": [ + { + "bbox": [ + 306, + 302, + 427, + 314 + ], + "type": "text", + "content": "3.2. Limitations of SORT" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 320, + 545, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 320, + 545, + 357 + ], + "spans": [ + { + "bbox": [ + 304, + 320, + 545, + 357 + ], + "type": "text", + "content": "In this section, we identify three main limitations of SORT which are connected. This analysis lays the foundation of our proposed method." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 365, + 437, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 365, + 437, + 376 + ], + "spans": [ + { + "bbox": [ + 306, + 365, + 437, + 376 + ], + "type": "text", + "content": "3.2.1 Sensitive to State Noise" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 383, + 545, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 383, + 545, + 468 + ], + "spans": [ + { + "bbox": [ + 304, + 383, + 545, + 468 + ], + "type": "text", + "content": "Now we show that SORT is sensitive to the noise from KF's state estimations. To begin with, we assume that the estimated object center position follows " + }, + { + "bbox": [ + 304, + 383, + 545, + 468 + ], + "type": "inline_equation", + "content": "u \\sim \\mathcal{N}(\\mu_u, \\sigma_u^2)" + }, + { + "bbox": [ + 304, + 383, + 545, + 468 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 383, + 545, + 468 + ], + "type": "inline_equation", + "content": "v \\sim \\mathcal{N}(\\mu_v, \\sigma_v^2)" + }, + { + "bbox": [ + 304, + 383, + 545, + 468 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 383, + 545, + 468 + ], + "type": "inline_equation", + "content": "(\\mu_u, \\mu_v)" + }, + { + "bbox": [ + 304, + 383, + 545, + 468 + ], + "type": "text", + "content": " is the underlying true position. Then, if we assume that the state noises are independent on different steps, by Eq.2, the object speed between two time steps, " + }, + { + "bbox": [ + 304, + 383, + 545, + 468 + ], + "type": "inline_equation", + "content": "t \\longrightarrow t + \\Delta t" + }, + { + "bbox": [ + 304, + 383, + 545, + 468 + ], + "type": "text", + "content": ", is" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 339, + 476, + 545, + 498 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 476, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 339, + 476, + 545, + 498 + ], + "type": "interline_equation", + "content": "\\dot {u} = \\frac {u _ {t + \\Delta t} - u _ {t}}{\\Delta t}, \\quad \\dot {v} = \\frac {v _ {t + \\Delta t} - v _ {t}}{\\Delta t}, \\tag {4}", + "image_path": "4c1ca31ebb0e30d71a9605a206a2a4db8b5ceb7e1a293613670366cb896bb820.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 501, + 545, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 501, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 501, + 545, + 594 + ], + "type": "text", + "content": "making the noise of estimated speed " + }, + { + "bbox": [ + 304, + 501, + 545, + 594 + ], + "type": "inline_equation", + "content": "\\delta_{\\dot{u}}\\sim \\mathcal{N}(0,\\frac{2\\sigma_u^2}{(\\Delta t)^2})" + }, + { + "bbox": [ + 304, + 501, + 545, + 594 + ], + "type": "inline_equation", + "content": "\\delta_{\\dot{v}}\\sim \\mathcal{N}(0,\\frac{2\\sigma_v^2}{(\\Delta t)^2})" + }, + { + "bbox": [ + 304, + 501, + 545, + 594 + ], + "type": "text", + "content": " . Therefore, a small " + }, + { + "bbox": [ + 304, + 501, + 545, + 594 + ], + "type": "inline_equation", + "content": "\\Delta t" + }, + { + "bbox": [ + 304, + 501, + 545, + 594 + ], + "type": "text", + "content": " will amplify the noise. This suggests that SORT will suffer from the heavy noise of velocity estimation on high-frame-rate videos. The analysis above is simplified from the reality. In practice, velocity won't be determined by the state on future time steps. For a more strict analysis, please refer to Appendix G." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 594, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 714 + ], + "type": "text", + "content": "Moreover, for most multi-object tracking scenarios, the target object displacement is only a few pixels between consecutive frames. For instance, the average displacement is 1.93 pixels and 0.65 pixels along the image width and height for the MOT17 [41] training dataset. In such a case, even if the estimated position has a shift of only a single pixel, it causes a significant variation in the estimated speed. In general, the variance of the speed estimation can be of the same magnitude as the speed itself or even greater. This will not make a massive impact as the shift is only of few pixels" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9688" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 78, + 70, + 521, + 260 + ], + "blocks": [ + { + "bbox": [ + 78, + 70, + 521, + 260 + ], + "lines": [ + { + "bbox": [ + 78, + 70, + 521, + 260 + ], + "spans": [ + { + "bbox": [ + 78, + 70, + 521, + 260 + ], + "type": "image", + "image_path": "d9a8507f837ba33dcae203cc6bf03fba0cfc3bcadb76306f639c7a4ecec00b36.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 264, + 547, + 313 + ], + "lines": [ + { + "bbox": [ + 46, + 264, + 547, + 313 + ], + "spans": [ + { + "bbox": [ + 46, + 264, + 547, + 313 + ], + "type": "text", + "content": "Figure 2. The pipeline of our proposed OC-SORT. The red boxes are detections, orange boxes are active tracks, blue boxes are untracked tracks, and dashed boxes are the estimates from KF. During association, OCM is used to add the velocity consistency cost. The target #1 is lost on the frame t+1 because of occlusion. But on the next frame, it is recovered by referring to its observation of the frame t by OCR. It being re-tracked triggers ORU from t to t+2 for the parameters of its KF." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 325, + 288, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 325, + 288, + 409 + ], + "spans": [ + { + "bbox": [ + 46, + 325, + 288, + 409 + ], + "type": "text", + "content": "from the ground truth on the next time step and the observations, whose variance is independent of the time, will be able to fix the noise when updating the posteriori parameters. However, we find that such a high sensitivity to state noise introduces significant problems in practice after being amplified by the error accumulation across multiple time steps when no observation is available for KF update." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 411, + 211, + 423 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 411, + 211, + 423 + ], + "spans": [ + { + "bbox": [ + 47, + 411, + 211, + 423 + ], + "type": "text", + "content": "3.2.2 Temporal Error Magnification" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 430, + 287, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 430, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 46, + 430, + 287, + 622 + ], + "type": "text", + "content": "For analysis above in Eq. 4, we assume the noise of the object state is i.i.d on different time steps (this is a simplified version, a more detailed analysis is provided in Appendix G). This is reasonable for object detections but not for the estimations from KF. This is because KF's estimations always rely on its estimations on previous time steps. The effect is usually minor because KF can use observation in update to prevent the posteriori state estimation and covariance, i.e. " + }, + { + "bbox": [ + 46, + 430, + 287, + 622 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_{t|t}" + }, + { + "bbox": [ + 46, + 430, + 287, + 622 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 430, + 287, + 622 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{t|t}" + }, + { + "bbox": [ + 46, + 430, + 287, + 622 + ], + "type": "text", + "content": ", deviating from the true value too far away. However, when no observations are provided to KF, it cannot use observation to update its parameters. Then it has to follow Eq. 3 to prolong the estimated trajectory to the next time step. Consider a track is occluded on the time steps between " + }, + { + "bbox": [ + 46, + 430, + 287, + 622 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 430, + 287, + 622 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 430, + 287, + 622 + ], + "type": "inline_equation", + "content": "t + T" + }, + { + "bbox": [ + 46, + 430, + 287, + 622 + ], + "type": "text", + "content": " and the noise of speed estimate follows " + }, + { + "bbox": [ + 46, + 430, + 287, + 622 + ], + "type": "inline_equation", + "content": "\\delta_{\\dot{u}_t} \\sim \\mathcal{N}(0,2\\sigma_u^2)" + }, + { + "bbox": [ + 46, + 430, + 287, + 622 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 430, + 287, + 622 + ], + "type": "inline_equation", + "content": "\\delta_{\\dot{v}_t} \\sim \\mathcal{N}(0,2\\sigma_v^2)" + }, + { + "bbox": [ + 46, + 430, + 287, + 622 + ], + "type": "text", + "content": " for SORT. On the step " + }, + { + "bbox": [ + 46, + 430, + 287, + 622 + ], + "type": "inline_equation", + "content": "t + T" + }, + { + "bbox": [ + 46, + 430, + 287, + 622 + ], + "type": "text", + "content": ", state estimation would be" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 80, + 627, + 287, + 639 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 627, + 287, + 639 + ], + "spans": [ + { + "bbox": [ + 80, + 627, + 287, + 639 + ], + "type": "interline_equation", + "content": "u _ {t + T} = u _ {t} + T \\dot {u} _ {t}, \\quad v _ {t + T} = v _ {t} + T \\dot {v} _ {t}, \\tag {5}", + "image_path": "b65157ccb23c486df5fc2288b64c3c4598c3fa5acd32bb29adb76c0582d521a7.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": "whose noise follows " + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\delta_{u_{t + T}}\\sim \\mathcal{N}(0,2T^2\\sigma_u^2)" + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\delta_{v_{t + T}}\\sim" + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(0,2T^2\\sigma_v^2)" + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": ". So without the observations, the estimation from the linear motion assumption of KF results in a fast error accumulation with respect to time. Given " + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\sigma_v" + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\sigma_{u}" + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": " is of the same magnitude as object displacement between consecutive frames, the noise of final object position" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 324, + 545, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 324, + 545, + 408 + ], + "spans": [ + { + "bbox": [ + 304, + 324, + 545, + 408 + ], + "type": "inline_equation", + "content": "(u_{t + T}, v_{t + T})" + }, + { + "bbox": [ + 304, + 324, + 545, + 408 + ], + "type": "text", + "content": " is of the same magnitude as the object size. For instance, the size of pedestrians close to the camera on MOT17 is around " + }, + { + "bbox": [ + 304, + 324, + 545, + 408 + ], + "type": "inline_equation", + "content": "50 \\times 300" + }, + { + "bbox": [ + 304, + 324, + 545, + 408 + ], + "type": "text", + "content": " pixels. So even assuming the variance of position estimation is only 1 pixel, 10-frame occlusion can accumulate a shift in final position estimation as large as the object size. Such error magnification leads to a major accumulation of errors when the scenes are crowded." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 413, + 422, + 424 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 413, + 422, + 424 + ], + "spans": [ + { + "bbox": [ + 305, + 413, + 422, + 424 + ], + "type": "text", + "content": "3.2.3 Estimation-Centric" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 428, + 545, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 428, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 304, + 428, + 545, + 632 + ], + "type": "text", + "content": "The aforementioned limitations come from a fundamental property of SORT that it follows KF to be estimation-centric. It allows update without the existence of observations and purely trusts the estimations. A key difference between state estimations and observations is that we can assume that the observations by an object detector in each frame are affected by i.i.d. noise " + }, + { + "bbox": [ + 304, + 428, + 545, + 632 + ], + "type": "inline_equation", + "content": "\\delta_{\\mathbf{z}} \\sim \\mathcal{N}(0, \\sigma'^2)" + }, + { + "bbox": [ + 304, + 428, + 545, + 632 + ], + "type": "text", + "content": " while the noise in state estimations can be accumulated along the hidden Markov process. Moreover, modern object detectors use powerful object visual features [48, 51]. It makes that, even on a single frame, it is usually safe to assume " + }, + { + "bbox": [ + 304, + 428, + 545, + 632 + ], + "type": "inline_equation", + "content": "\\sigma' < \\sigma_u" + }, + { + "bbox": [ + 304, + 428, + 545, + 632 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 428, + 545, + 632 + ], + "type": "inline_equation", + "content": "\\sigma' < \\sigma_v" + }, + { + "bbox": [ + 304, + 428, + 545, + 632 + ], + "type": "text", + "content": " because the object localization is more accurate by detection than from the state estimations through linear motion assumption. Combined with the previously mentioned two limitations, being estimation-centric makes SORT suffer from heavy noise when there is occlusion and the object motion is not perfectly linear." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 645, + 463, + 657 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 645, + 463, + 657 + ], + "spans": [ + { + "bbox": [ + 305, + 645, + 463, + 657 + ], + "type": "text", + "content": "4. Observation-Centric SORT" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 666, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 666, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 547, + 714 + ], + "type": "text", + "content": "In this section, we introduce the proposed Observation-Centric SORT (OC-SORT). To address the limitations of SORT discussed above, we use the momentum of the object moving into the association stage and develop a" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9689" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 97, + 70, + 496, + 190 + ], + "blocks": [ + { + "bbox": [ + 97, + 70, + 496, + 190 + ], + "lines": [ + { + "bbox": [ + 97, + 70, + 496, + 190 + ], + "spans": [ + { + "bbox": [ + 97, + 70, + 496, + 190 + ], + "type": "image", + "image_path": "1bbcd5172d3c920aac4e5ce8a63f11b55c0a4fdeb4743beb0bedc75c26af596e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 192, + 547, + 242 + ], + "lines": [ + { + "bbox": [ + 46, + 192, + 547, + 242 + ], + "spans": [ + { + "bbox": [ + 46, + 192, + 547, + 242 + ], + "type": "text", + "content": "Figure 3. Example of how Observation-centric Re-Update (ORU) reduces the error accumulation when a track is broken. The target is occluded between the second and the third time step and the tracker finds it back at the third step. Yellow boxes are the state observations by the detector. White stars are the estimated centers without ORU. Yellow stars are the estimated centers fixed by ORU. The gray star on the fourth step is the estimated center without ORU and fails to match observations." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 252, + 290, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 252, + 290, + 373 + ], + "spans": [ + { + "bbox": [ + 46, + 252, + 290, + 373 + ], + "type": "text", + "content": "pipeline with less noise and more robustness over occlusion and non-linear motion. The key is to design the tracker as observation-centric instead of estimation-centric. If a track is recovered from being untracked, we use an Observation-centric Re-Update (ORU) strategy to counter the accumulated error during the untracked period. OC-SORT also adds an Observation-Centric Momentum (OCM) term in the association cost. Please refer to Algorithm 1 in Appendix for the pseudo-code of OC-SORT. The pipeline is shown in Fig. 2." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 378, + 251, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 378, + 251, + 392 + ], + "spans": [ + { + "bbox": [ + 47, + 378, + 251, + 392 + ], + "type": "text", + "content": "4.1. Observation-centric Re-Update (ORU)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 396, + 290, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 396, + 290, + 578 + ], + "spans": [ + { + "bbox": [ + 46, + 396, + 290, + 578 + ], + "type": "text", + "content": "In practice, even if an object can be associated again by SORT after a period of being untracked, it is probably lost again because its KF parameters have already deviated far away from the correct due to the temporal error magnification. To alleviate this problem, we propose Observation-centric Re-Update (ORU) to reduce the accumulated error. Once a track is associated with an observation again after a period of being untracked (\"re-activation\"), we backcheck the period of its being lost and re-update the parameters of KF. The re-update is based on \"observations\" from a virtual trajectory. The virtual trajectory is generated referring to the observations on the steps starting and ending the untracked period. For example, by denoting the last-seen observation before being untracked as " + }, + { + "bbox": [ + 46, + 396, + 290, + 578 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{t_1}" + }, + { + "bbox": [ + 46, + 396, + 290, + 578 + ], + "type": "text", + "content": " and the observation triggering the re-association as " + }, + { + "bbox": [ + 46, + 396, + 290, + 578 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{t_2}" + }, + { + "bbox": [ + 46, + 396, + 290, + 578 + ], + "type": "text", + "content": ", the virtual trajectory is denoted as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 84, + 588, + 287, + 601 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 588, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 84, + 588, + 287, + 601 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathbf {z}} _ {t} = \\operatorname {T r a j} _ {\\text {v i r t u a l}} \\left(\\mathbf {z} _ {t _ {1}}, \\mathbf {z} _ {t _ {2}}, t\\right), t _ {1} < t < t _ {2}. \\tag {6}", + "image_path": "247d1fa414f8e0665760b4275480bb5d19cd5c7b67516002903d3c7c4cdfce13.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 605, + 287, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 605, + 287, + 629 + ], + "spans": [ + { + "bbox": [ + 47, + 605, + 287, + 629 + ], + "type": "text", + "content": "Then, along the trajectory of " + }, + { + "bbox": [ + 47, + 605, + 287, + 629 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{z}}_t(t_1 < t < t_2)" + }, + { + "bbox": [ + 47, + 605, + 287, + 629 + ], + "type": "text", + "content": ", we run the loop of predict and re-update. The re-update operation is" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 634, + 287, + 689 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 634, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 634, + 287, + 689 + ], + "type": "interline_equation", + "content": "r e - u p d a t e \\left\\{ \\begin{array}{l} \\mathbf {K} _ {t} = \\mathbf {P} _ {t | t - 1} \\mathbf {H} _ {t} ^ {\\top} \\left(\\mathbf {H} _ {t} \\mathbf {P} _ {t | t - 1} \\mathbf {H} _ {t} ^ {\\top} + \\mathbf {R} _ {t}\\right) ^ {- 1} \\\\ \\hat {\\mathbf {x}} _ {t | t} = \\hat {\\mathbf {x}} _ {t | t - 1} + \\mathbf {K} _ {t} \\left(\\tilde {\\mathbf {z}} _ {t} - \\mathbf {H} _ {t} \\hat {\\mathbf {x}} _ {t | t - 1}\\right) \\\\ \\mathbf {P} _ {t | t} = (\\mathbf {I} - \\mathbf {K} _ {t} \\mathbf {H} _ {t}) \\mathbf {P} _ {t | t - 1} \\end{array} \\right. \\tag {7}", + "image_path": "f87fa0d9787b5dd6c4d1b9d10531f3a09e6b0891e37f9c9c4c54fbe455a75fad.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "content": "As the observations on the virtual trajectory match the motion pattern anchored by the last-seen and the latest asso" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 340, + 253, + 520, + 316 + ], + "blocks": [ + { + "bbox": [ + 340, + 253, + 520, + 316 + ], + "lines": [ + { + "bbox": [ + 340, + 253, + 520, + 316 + ], + "spans": [ + { + "bbox": [ + 340, + 253, + 520, + 316 + ], + "type": "image", + "image_path": "dd82417e88e92bedf836ce19b2d8d0998876de1b60b1d7e712f34cb9dcf768ff.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 325, + 547, + 399 + ], + "lines": [ + { + "bbox": [ + 304, + 325, + 547, + 399 + ], + "spans": [ + { + "bbox": [ + 304, + 325, + 547, + 399 + ], + "type": "text", + "content": "Figure 4. Calculation of motion direction difference in OCM. The green line indicates an existing track and the dots are the observations on it. The red dots are the new observations to be associated. The blue link and the yellow link form the directions of " + }, + { + "bbox": [ + 304, + 325, + 547, + 399 + ], + "type": "inline_equation", + "content": "\\theta^{\\mathrm{track}}" + }, + { + "bbox": [ + 304, + 325, + 547, + 399 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 325, + 547, + 399 + ], + "type": "inline_equation", + "content": "\\theta^{\\mathrm{intention}}" + }, + { + "bbox": [ + 304, + 325, + 547, + 399 + ], + "type": "text", + "content": " respectively. The included angle is the difference of direction " + }, + { + "bbox": [ + 304, + 325, + 547, + 399 + ], + "type": "inline_equation", + "content": "\\Delta \\theta" + }, + { + "bbox": [ + 304, + 325, + 547, + 399 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 407, + 547, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 407, + 547, + 479 + ], + "spans": [ + { + "bbox": [ + 304, + 407, + 547, + 479 + ], + "type": "text", + "content": "ciated real observations, the update will not suffer from the error accumulated through the dummy update anymore. We call the proposed process Observation-centric Re-Update. It serves as an independent stage outside the predict-update loop and is triggered only a track is re-activated from a period of having no observations." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 489, + 520, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 489, + 520, + 502 + ], + "spans": [ + { + "bbox": [ + 306, + 489, + 520, + 502 + ], + "type": "text", + "content": "4.2. Observation-Centric Momentum (OCM)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 509, + 547, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 509, + 547, + 699 + ], + "spans": [ + { + "bbox": [ + 304, + 509, + 547, + 699 + ], + "type": "text", + "content": "In a reasonably short time interval, we can approximate the motion as linear. And the linear motion assumption also asks for consistent motion direction. But the noise prevents us from leveraging the consistency of direction. To be precise, to determine the motion direction, we need the object state on two steps with a time difference " + }, + { + "bbox": [ + 304, + 509, + 547, + 699 + ], + "type": "inline_equation", + "content": "\\Delta t" + }, + { + "bbox": [ + 304, + 509, + 547, + 699 + ], + "type": "text", + "content": ". If " + }, + { + "bbox": [ + 304, + 509, + 547, + 699 + ], + "type": "inline_equation", + "content": "\\Delta t" + }, + { + "bbox": [ + 304, + 509, + 547, + 699 + ], + "type": "text", + "content": " is small, the velocity noise would be significant because of the estimation's sensitivity to state noise. If " + }, + { + "bbox": [ + 304, + 509, + 547, + 699 + ], + "type": "inline_equation", + "content": "\\Delta t" + }, + { + "bbox": [ + 304, + 509, + 547, + 699 + ], + "type": "text", + "content": " is big, the noise of direction estimation can also be significant because of the temporal error magnification and the failure of linear motion assumption. As state observations have no problem of temporal error magnification that state estimations suffer from, we propose to use observations instead of estimations to reduce the noise of motion direction calculation and introduce the term of velocity consistency to help the association." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 318, + 701, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 318, + 701, + 545, + 714 + ], + "type": "text", + "content": "With the new term, given " + }, + { + "bbox": [ + 318, + 701, + 545, + 714 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 318, + 701, + 545, + 714 + ], + "type": "text", + "content": " existing tracks and " + }, + { + "bbox": [ + 318, + 701, + 545, + 714 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 318, + 701, + 545, + 714 + ], + "type": "text", + "content": " de" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "text", + "content": "9690" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 95 + ], + "type": "text", + "content": "tectons on the new-coming time step, the association cost matrix is formulated as" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 87, + 103, + 287, + 118 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 103, + 287, + 118 + ], + "spans": [ + { + "bbox": [ + 87, + 103, + 287, + 118 + ], + "type": "interline_equation", + "content": "C (\\hat {\\mathbf {X}}, \\mathbf {Z}) = C _ {\\mathrm {I o U}} (\\hat {\\mathbf {X}}, \\mathbf {Z}) + \\lambda C _ {v} (\\mathcal {Z}, \\mathbf {Z}), \\tag {8}", + "image_path": "823161ba10bae20f8488daaa6fabbca1c3edd6b903714ee9f743044b41e18a43.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 125, + 287, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 125, + 287, + 281 + ], + "spans": [ + { + "bbox": [ + 47, + 125, + 287, + 281 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 125, + 287, + 281 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{X}}\\in \\mathbb{R}^{N\\times 7}" + }, + { + "bbox": [ + 47, + 125, + 287, + 281 + ], + "type": "text", + "content": " is the set of object state estimations and " + }, + { + "bbox": [ + 47, + 125, + 287, + 281 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}\\in \\mathbb{R}^{M\\times 5}" + }, + { + "bbox": [ + 47, + 125, + 287, + 281 + ], + "type": "text", + "content": " is the set of observations on the new time step. " + }, + { + "bbox": [ + 47, + 125, + 287, + 281 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 47, + 125, + 287, + 281 + ], + "type": "text", + "content": " is a weighting factor. " + }, + { + "bbox": [ + 47, + 125, + 287, + 281 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}" + }, + { + "bbox": [ + 47, + 125, + 287, + 281 + ], + "type": "text", + "content": " contains the trajectory of observations of all existing tracks. " + }, + { + "bbox": [ + 47, + 125, + 287, + 281 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{IoU}}(\\cdot ,\\cdot)" + }, + { + "bbox": [ + 47, + 125, + 287, + 281 + ], + "type": "text", + "content": " calculates the negative pairwise IoU (Intersection over Union) and " + }, + { + "bbox": [ + 47, + 125, + 287, + 281 + ], + "type": "inline_equation", + "content": "C_v(\\cdot ,\\cdot)" + }, + { + "bbox": [ + 47, + 125, + 287, + 281 + ], + "type": "text", + "content": " calculates the consistency between the directions of i) linking two observations on an existing track " + }, + { + "bbox": [ + 47, + 125, + 287, + 281 + ], + "type": "inline_equation", + "content": "(\\theta^{\\mathrm{track}})" + }, + { + "bbox": [ + 47, + 125, + 287, + 281 + ], + "type": "text", + "content": " and ii) linking a track's historical observation and a new observation " + }, + { + "bbox": [ + 47, + 125, + 287, + 281 + ], + "type": "inline_equation", + "content": "(\\theta^{\\mathrm{intention}})" + }, + { + "bbox": [ + 47, + 125, + 287, + 281 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 47, + 125, + 287, + 281 + ], + "type": "inline_equation", + "content": "C_v" + }, + { + "bbox": [ + 47, + 125, + 287, + 281 + ], + "type": "text", + "content": " contains all pairs of " + }, + { + "bbox": [ + 47, + 125, + 287, + 281 + ], + "type": "inline_equation", + "content": "\\Delta \\theta = |\\theta^{\\mathrm{track}} - \\theta^{\\mathrm{intention}}|" + }, + { + "bbox": [ + 47, + 125, + 287, + 281 + ], + "type": "text", + "content": ". In our implementation, we calculate the motion direction in radians, namely " + }, + { + "bbox": [ + 47, + 125, + 287, + 281 + ], + "type": "inline_equation", + "content": "\\theta = \\arctan \\left(\\frac{v_1 - v_2}{u_1 - u_2}\\right)" + }, + { + "bbox": [ + 47, + 125, + 287, + 281 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 47, + 125, + 287, + 281 + ], + "type": "inline_equation", + "content": "(u_{1},v_{1})" + }, + { + "bbox": [ + 47, + 125, + 287, + 281 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 125, + 287, + 281 + ], + "type": "inline_equation", + "content": "(u_{2},v_{2})" + }, + { + "bbox": [ + 47, + 125, + 287, + 281 + ], + "type": "text", + "content": " are the observations on two different time steps. The calculation is also illustrated in Figure 4." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 282, + 287, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 282, + 287, + 425 + ], + "spans": [ + { + "bbox": [ + 46, + 282, + 287, + 425 + ], + "type": "text", + "content": "Following the assumptions of noise distribution mentioned before, we can derive a closed-form probability density function of the distribution of the noise in the direction estimation. The derivation is explained in detail in Appendix A. By analyzing the property of this distribution, we reach a conclusion that, under the linear-motion model, the scale of the noise of direction estimation is negatively correlated to the time difference between the two observation points, i.e. " + }, + { + "bbox": [ + 46, + 282, + 287, + 425 + ], + "type": "inline_equation", + "content": "\\Delta t" + }, + { + "bbox": [ + 46, + 282, + 287, + 425 + ], + "type": "text", + "content": ". This suggests increasing " + }, + { + "bbox": [ + 46, + 282, + 287, + 425 + ], + "type": "inline_equation", + "content": "\\Delta t" + }, + { + "bbox": [ + 46, + 282, + 287, + 425 + ], + "type": "text", + "content": " to achieve a low-noisy estimation of " + }, + { + "bbox": [ + 46, + 282, + 287, + 425 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 282, + 287, + 425 + ], + "type": "text", + "content": ". However, the assumption of linear motion typically holds only when " + }, + { + "bbox": [ + 46, + 282, + 287, + 425 + ], + "type": "inline_equation", + "content": "\\Delta t" + }, + { + "bbox": [ + 46, + 282, + 287, + 425 + ], + "type": "text", + "content": " is small enough. Therefore, the choice of " + }, + { + "bbox": [ + 46, + 282, + 287, + 425 + ], + "type": "inline_equation", + "content": "\\Delta t" + }, + { + "bbox": [ + 46, + 282, + 287, + 425 + ], + "type": "text", + "content": " requires a trade-off." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 426, + 287, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 287, + 521 + ], + "type": "text", + "content": "Besides ORU and OCM, we also find it empirically helpful to check a track's last presence to recover it from being lost. We thus apply a heuristic Observation-Centric Recovery (OCR) technique. OCR will start a second attempt of associating between the last observation of unmatched tracks to the unmatched observations after the usual association stage. It can handle the case of an object stopping or being occluded for a short time interval." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 532, + 127, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 532, + 127, + 544 + ], + "spans": [ + { + "bbox": [ + 47, + 532, + 127, + 544 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 552, + 163, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 552, + 163, + 565 + ], + "spans": [ + { + "bbox": [ + 47, + 552, + 163, + 565 + ], + "type": "text", + "content": "5.1. Experimental Setup" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 713 + ], + "type": "text", + "content": "Datasets. We evaluate our method on multiple multi-object tracking datasets including MOT17 [41], MOT20 [14], KITTI [20], DanceTrack [54] and CroHD [56]. MOT17 [41] and MOT20 [14] are for pedestrian tracking, where targets mostly move linearly, while scenes in MOT20 are more crowded. KITTI [20] is for pedestrian and car tracking with a relatively low frame rate of 10FPS. CroHD is a dataset for head tracking in the crowd and the results on it are included in the appendix. DanceTrack [54] is a recently proposed dataset for human tracking. For the data in DanceTrack, object localization is easy, but the object motion is highly non-linear. Furthermore, the objects" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": "have a close appearance, severe occlusion, and frequent crossovers. Considering our goal is to improve tracking robustness under occlusion and non-linear object motion, we would emphasize the comparison on DanceTrack." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 121, + 546, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 121, + 546, + 312 + ], + "spans": [ + { + "bbox": [ + 304, + 121, + 546, + 312 + ], + "type": "text", + "content": "Implementations. For a fair comparison, we directly apply the object detections from existing baselines. For MOT17, MOT20, and DanceTrack, we use the publicly availableYOLOX [19] detector weights by ByteTrack [70]. For KITTI [20], we use the detections from PermaTrack [57] publicly available in the official release1. For ORU, we generate the virtual trajectory during occlusion with the constant-velocity assumption. Therefore, Eq. 6 is adopted as " + }, + { + "bbox": [ + 304, + 121, + 546, + 312 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{z}}_t = \\mathbf{z}_{t_1} + \\frac{t - t_1}{t_2 - t_1} (\\mathbf{z}_{t_2} - \\mathbf{z}_{t_1}), t_1 < t < t_2" + }, + { + "bbox": [ + 304, + 121, + 546, + 312 + ], + "type": "text", + "content": ". For OCM, the velocity direction is calculated using the observations three time steps apart, i.e. " + }, + { + "bbox": [ + 304, + 121, + 546, + 312 + ], + "type": "inline_equation", + "content": "\\Delta t = 3" + }, + { + "bbox": [ + 304, + 121, + 546, + 312 + ], + "type": "text", + "content": ". The direction difference is measured by the absolute difference of angles in radians. We set " + }, + { + "bbox": [ + 304, + 121, + 546, + 312 + ], + "type": "inline_equation", + "content": "\\lambda = 0.2" + }, + { + "bbox": [ + 304, + 121, + 546, + 312 + ], + "type": "text", + "content": " in Eq. 8. Following the common practice of SORT, we set the detection confidence threshold at 0.4 for MOT20 and 0.6 for other datasets. The IoU threshold during association is 0.3." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 313, + 546, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 313, + 546, + 422 + ], + "spans": [ + { + "bbox": [ + 304, + 313, + 546, + 422 + ], + "type": "text", + "content": "Metrics. We adopt HOTA [37] as the main metric as it maintains a proper balance between the accuracy of object detection and association [37]. We also emphasize AssA to evaluate the association performance. IDF1 is also used for association performance evaluation. Other metrics we report, such as MOTA, are highly related to detection performance. It is fair to use these metrics only when all methods use the same detections for tracking, which is referred to as \"public tracking\" as reported in Appendix C." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 431, + 420, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 431, + 420, + 443 + ], + "spans": [ + { + "bbox": [ + 306, + 431, + 420, + 443 + ], + "type": "text", + "content": "5.2. Benchmark Results" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 451, + 545, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 451, + 545, + 486 + ], + "spans": [ + { + "bbox": [ + 304, + 451, + 545, + 486 + ], + "type": "text", + "content": "Here we report the benchmark results on multiple datasets. We put all methods that use the shared detection results in the blue blocks at the bottom of each table." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 487, + 546, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 487, + 546, + 691 + ], + "spans": [ + { + "bbox": [ + 304, + 487, + 546, + 691 + ], + "type": "text", + "content": "MOT17 and MOT20. We report OC-SORT's performance on MOT17 and MOT20 in Table 1 and Table 2 using private detections. To make a fair comparison, we use the same detection as ByteTrack [70]. OC-SORT achieves performance comparable to other state-of-the-art methods. Our gains are especially significant in MOT20 under severe pedestrian occlusion, setting a state-of-the-art HOTA of 62.1. As our method is designed to be simple for better generalization, we do not use adaptive detection thresholds as in ByteTrack. Also, ByteTrack uses more detections of low-confidence to achieve higher MOTA scores but we keep the detection confidence threshold the same as on other datasets, which is the common practice in the community. We inherit the linear interpolation on the two datasets by baseline methods for a fair comparison. To more clearly discard the variance from the detector, we also perform public tracking on MOT17 and MOT20, which is reported in Table 12 and Table 13 in" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 317, + 702, + 451, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 702, + 451, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 702, + 451, + 713 + ], + "type": "text", + "content": "1 https://github.com/TRI-ML/permatrack/" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 314, + 757 + ], + "type": "text", + "content": "9691" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 102, + 84, + 491, + 219 + ], + "blocks": [ + { + "bbox": [ + 88, + 70, + 504, + 83 + ], + "lines": [ + { + "bbox": [ + 88, + 70, + 504, + 83 + ], + "spans": [ + { + "bbox": [ + 88, + 70, + 504, + 83 + ], + "type": "text", + "content": "Table 1. Results on MOT17-test with the private detections. ByteTrack and OC-SORT share detections." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 102, + 84, + 491, + 219 + ], + "lines": [ + { + "bbox": [ + 102, + 84, + 491, + 219 + ], + "spans": [ + { + "bbox": [ + 102, + 84, + 491, + 219 + ], + "type": "table", + "html": "
TrackerHOTA↑MOTA↑IDF1↑FP(104)↓FN(104)↓IDs↓Frag↓AssA↑AssR↑
FairMOT [71]59.373.772.32.7511.73,3038,07358.063.6
TransCt [67]54.573.262.22.3112.44,6149,51949.754.2
TransTrk [55]54.175.263.55.028.643,6034,87247.957.1
GRTU [60]62.074.975.03.2010.81,8121,82462.165.8
QDTrack [42]53.968.766.32.6614.73,3788,09152.757.2
MOTR [69]57.271.968.42.1113.62,1153,89755.859.2
PermaTr [57]55.573.868.92.9011.53,6996,13253.159.8
TransMOT [12]61.776.775.13.629.322,3467,71959.966.5
GTR [75]59.175.371.52.6811.02,859-61.6-
DST-Tracker [8]60.175.272.32.4211.02,729-62.1-
MeMOT [5]56.972.569.02.7211.52,724-55.2-
UniCom [68]61.777.275.55.017.335,379---
ByteTrack [70]63.180.377.32.558.372,1962,27762.068.2
OC-SORT63.278.077.51.5110.81,9502,04063.267.5
", + "image_path": "709af639570817f2467cf2173b10c89036a78010c02a753d91a891cef84f7273.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 100, + 242, + 496, + 335 + ], + "blocks": [ + { + "bbox": [ + 96, + 228, + 497, + 240 + ], + "lines": [ + { + "bbox": [ + 96, + 228, + 497, + 240 + ], + "spans": [ + { + "bbox": [ + 96, + 228, + 497, + 240 + ], + "type": "text", + "content": "Table 2. Results on MOT20-test with private detections. ByteTrack and OC-SORT share detections." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 100, + 242, + 496, + 335 + ], + "lines": [ + { + "bbox": [ + 100, + 242, + 496, + 335 + ], + "spans": [ + { + "bbox": [ + 100, + 242, + 496, + 335 + ], + "type": "table", + "html": "
TrackerHOTA↑MOTA↑IDF1↑FP(104)↓FN(104)↓IDs↓Frag↓AssA↑AssR↑
FairMOT [71]54.661.867.310.38.895,2437,87454.760.7
TransCt [67]43.558.549.66.4214.64,6959,58137.045.1
Semi-TCL [35]55.365.270.16.1211.54,1398,50856.360.9
CSTrack [36]54.066.668.62.5414.43,1967,63254.057.6
GSDT [61]53.667.167.53.1913.53,1319,87552.758.5
TransMOT [12]61.977.575.23.428.081,6152,42160.166.3
MeMOT [5]54.163.766.14.7913.81,938-55.0-
ByteTrack [70]61.377.875.22.628.761,2231,46059.666.2
OC-SORT62.175.575.91.8010.89131,19862.067.5
", + "image_path": "54f30aece2e2af5d3c3d9cb24809c9f30470ef059b75340736299c92aeb0851c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 48, + 374, + 288, + 498 + ], + "blocks": [ + { + "bbox": [ + 47, + 349, + 287, + 371 + ], + "lines": [ + { + "bbox": [ + 47, + 349, + 287, + 371 + ], + "spans": [ + { + "bbox": [ + 47, + 349, + 287, + 371 + ], + "type": "text", + "content": "Table 3. Results on DanceTrack test set. Methods in the blue block share the same detections." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 374, + 288, + 498 + ], + "lines": [ + { + "bbox": [ + 48, + 374, + 288, + 498 + ], + "spans": [ + { + "bbox": [ + 48, + 374, + 288, + 498 + ], + "type": "table", + "html": "
TrackerHOTA↑DetA↑AssA↑MOTA↑IDF1↑
CenterTrack [73]41.878.122.686.835.7
FairMOT [71]39.766.723.882.240.8
QDTrack [42]45.772.129.283.044.8
TransTrk[55]45.575.927.588.445.2
TraDes [64]43.374.525.486.241.2
MOTR [69]54.273.540.279.751.5
GTR [75]48.072.531.984.750.3
DST-Tracker [8]51.972.334.684.951.0
SORT [3]47.972.031.291.850.8
DeepSORT [63]45.671.029.787.847.9
ByteTrack [70]47.371.631.489.552.5
OC-SORT54.680.440.289.654.6
OC-SORT + Linear Interp55.180.440.492.254.9
", + "image_path": "f1de8a177c5db9e420995daf7203362ce12946acd3e005091247b770da9edf36.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 508, + 287, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 508, + 287, + 532 + ], + "spans": [ + { + "bbox": [ + 47, + 508, + 287, + 532 + ], + "type": "text", + "content": "Appendix C. OC-SORT still outperforms the existing state-of-the-art in public tracking settings." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 533, + 287, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 287, + 675 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 287, + 675 + ], + "type": "text", + "content": "DanceTrack. To evaluate OC-SORT under challenging non-linear object motion, we report results on the DanceTrack in Table 3. OC-SORT sets a new state-of-the-art, outperforming the baselines by a great margin under nonlinear object motions. We compare the tracking results of SORT and OC-SORT under extreme non-linear situations in Fig.1 and more samples are available in Fig.8 in Appendix E. We also visualize the output trajectories by OC-SORT and SORT on randomly selected DanceTrack video clips in Fig.9 in Appendix E. For multi-object tracking in occlusion and non-linear motion, the results on DanceTrack are strong evidence of the effectiveness of OC-SORT." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 713 + ], + "type": "text", + "content": "KITTI. In Table 4 we report the results on the KITTI dataset. For a fair comparison, we adopt the detector weights by PermaTr [57] and report its performance in the" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 350, + 545, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 350, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 304, + 350, + 545, + 601 + ], + "type": "text", + "content": "table as well. We run OC-SORT given the shared detections. As initializing SORT's track requires continuous tracking across several frames (\"minimum hits\"), we observe that the results not recorded during the track initialization make a significant difference. To address this problem, we perform offline head padding (HP) post-processing by writing these entries back after finishing the online tracking stage. The results of the car category on KITTI show an essential shortcoming of the default implementation version of OC-SORT that it chooses the IoU matching for the association. When the objects move fast or the frame rate is low, the IoU of bounding boxes between consecutive frames can be very low or even zero. This issue does not come from the intrinsic design of OC-SORT and is widely observed when using IoU as the association cue. Adding other cues [49, 72, 73] and appearance similarity [38, 63] have been demonstrated [63] effective to solve this. In contrast to the relatively inferior car tracking performance, OC-SORT improves pedestrian tracking performance to a new state-of-the-art. Using the same detections, OC-SORT achieves a large performance gap over PermaTr with " + }, + { + "bbox": [ + 304, + 350, + 545, + 601 + ], + "type": "inline_equation", + "content": "10\\mathrm{x}" + }, + { + "bbox": [ + 304, + 350, + 545, + 601 + ], + "type": "text", + "content": " faster speed." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "The results on multiple benchmarks have demonstrated the effectiveness and efficiency of OC-SORT. We note that we use a shared parameter stack across datasets. Carefully tuning the parameters can probably further boost the performance. For example, the adaptive detection threshold is proven useful in previous work [70]. Besides the association accuracy, we also care about the inference speed. Given off-the-shelf detections, OC-SORT runs at 793 FPS on an Intel i9-9980XE CPU @ 3.00GHz. Therefore, OC-SORT" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9692" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 100, + 84, + 492, + 198 + ], + "blocks": [ + { + "bbox": [ + 122, + 71, + 470, + 83 + ], + "lines": [ + { + "bbox": [ + 122, + 71, + 470, + 83 + ], + "spans": [ + { + "bbox": [ + 122, + 71, + 470, + 83 + ], + "type": "text", + "content": "Table 4. Results on KITTI-test. Our method uses the same detections as PermaTr [57]" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 100, + 84, + 492, + 198 + ], + "lines": [ + { + "bbox": [ + 100, + 84, + 492, + 198 + ], + "spans": [ + { + "bbox": [ + 100, + 84, + 492, + 198 + ], + "type": "table", + "html": "
CarPedestrian
TrackerHOTA↑MOTA↑AssA↑IDs↓Frag↓HOTA↑MOTA↑AssA↑IDs↓Frag↓
IMMDP [65]68.6682.7569.76211181-----
SMAT [21]71.8883.6472.13198294-----
TrackMPNN [45]72.3087.3370.6348123739.4052.1035.45626669
MPNTrack [4]-----45.2646.2347.283971,078
CenterTr [73]73.0288.8371.1825422740.3553.8436.93425618
LGM [59]73.1487.6072.31448164-----
TuSimple [11]71.5586.3171.1129221845.8857.6147.62246651
PermaTr [57]77.4290.8577.6627527147.4365.0543.66483703
OC-SORT74.6487.8174.5225731852.9562.0057.81181598
OC-SORT + HP76.5490.2876.3925028054.6965.1459.08184609
", + "image_path": "7ffd7bdbb19f9faf4509658e09cfa58b4a54afb411813b890f9e2e8d8c8c6e41.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 49, + 229, + 283, + 297 + ], + "blocks": [ + { + "bbox": [ + 56, + 217, + 277, + 228 + ], + "lines": [ + { + "bbox": [ + 56, + 217, + 277, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 277, + 228 + ], + "type": "text", + "content": "Table 5. Ablation on MOT17-val and DanceTrack-val." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 229, + 283, + 297 + ], + "lines": [ + { + "bbox": [ + 49, + 229, + 283, + 297 + ], + "spans": [ + { + "bbox": [ + 49, + 229, + 283, + 297 + ], + "type": "table", + "html": "
MOT17-valDanceTrack-val
ORUOCMOCRHOTA↑AssA↑IDF1↑HOTA↑AssA↑IDF1↑
64.966.876.947.831.048.3
✓66.368.077.248.532.249.8
✓✓66.469.077.852.135.050.6
✓✓✓66.568.977.752.135.351.6
", + "image_path": "919728821b524f919edd360715daf917f4edf8e1619d10990f6e61b4a60b146c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 48, + 312, + 287, + 378 + ], + "blocks": [ + { + "bbox": [ + 56, + 299, + 279, + 311 + ], + "lines": [ + { + "bbox": [ + 56, + 299, + 279, + 311 + ], + "spans": [ + { + "bbox": [ + 56, + 299, + 279, + 311 + ], + "type": "text", + "content": "Table 6. Ablation on the trajectory hypothesis in ORU." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 312, + 287, + 378 + ], + "lines": [ + { + "bbox": [ + 48, + 312, + 287, + 378 + ], + "spans": [ + { + "bbox": [ + 48, + 312, + 287, + 378 + ], + "type": "table", + "html": "
MOT17-valDanceTrack-val
HOTA↑AssA↑IDF1↑HOTA↑AssA↑IDF1↑
Const. Speed66.568.977.752.135.351.6
GPR63.165.275.749.533.749.6
Linear Regression64.366.576.049.333.449.2
Const. Acceleration66.267.977.451.334.850.9
", + "image_path": "1c821a059a12aa9fe1c7d0924727d1434f579bade38ee37fb0bbc6d72aafb627.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 55, + 396, + 276, + 462 + ], + "blocks": [ + { + "bbox": [ + 67, + 382, + 266, + 393 + ], + "lines": [ + { + "bbox": [ + 67, + 382, + 266, + 393 + ], + "spans": [ + { + "bbox": [ + 67, + 382, + 266, + 393 + ], + "type": "text", + "content": "Table 7. Influence from the value of " + }, + { + "bbox": [ + 67, + 382, + 266, + 393 + ], + "type": "inline_equation", + "content": "{\\Delta t}" + }, + { + "bbox": [ + 67, + 382, + 266, + 393 + ], + "type": "text", + "content": " in OCM." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 55, + 396, + 276, + 462 + ], + "lines": [ + { + "bbox": [ + 55, + 396, + 276, + 462 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 276, + 462 + ], + "type": "table", + "html": "
MOT17-valDanceTrack-val
HOTA↑AssA↑IDF1↑HOTA↑AssA↑IDF1↑
Δt = 166.167.576.951.334.351.3
Δt = 266.368.077.352.235.451.4
Δt = 366.568.977.752.135.351.6
Δt = 666.067.576.952.135.451.8
", + "image_path": "96f8a51f9d665cc96a5527ae69491c475340e714dadccc54b60e35cd9ad3d990.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 469, + 236, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 469, + 236, + 480 + ], + "spans": [ + { + "bbox": [ + 47, + 469, + 236, + 480 + ], + "type": "text", + "content": "can still run in an online and real-time fashion." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 490, + 141, + 503 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 490, + 141, + 503 + ], + "spans": [ + { + "bbox": [ + 47, + 490, + 141, + 503 + ], + "type": "text", + "content": "5.3. Ablation Study" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 509, + 287, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 509, + 287, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 509, + 287, + 640 + ], + "type": "text", + "content": "Component Ablation. We ablate the contribution of proposed modules on the validation sets of MOT17 and DanceTrack in Table 5. The splitting of the MOT17 validation set follows a popular convention [73]. The results demonstrate the effectiveness of the proposed modules in OC-SORT. The results show that the performance gain from ORU is significant on both datasets but OCM only shows significant help on DanceTrack dataset where object motion is more complicated and the occlusion is heavy. The ablation study proves the effectiveness of our proposed method to improve tracking robustness in occlusion and non-linear motion." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 714 + ], + "type": "text", + "content": "Virtual Trajectory in ORU. For simplicity, we follow the naive hypothesis of constant speed to generate a virtual trajectory in ORU. There are other alternatives like constant acceleration, regression-based fitting such as Linear Regression (LR) or Gaussian Process Regression (GPR), and Near Constant Acceleration Model (NCAM) [27]. The results of" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 217, + 545, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 217, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 304, + 217, + 545, + 330 + ], + "type": "text", + "content": "comparing these choices are shown in Table 6. For GPR, we use the RBF kernel [10] " + }, + { + "bbox": [ + 304, + 217, + 545, + 330 + ], + "type": "inline_equation", + "content": "k(\\mathbf{x},\\mathbf{x}^{\\prime}) = \\exp \\left(-\\frac{||\\mathbf{x} - \\mathbf{x}^{\\prime}||^{2}}{50}\\right)" + }, + { + "bbox": [ + 304, + 217, + 545, + 330 + ], + "type": "text", + "content": ". We provide more studies on the kernel configuration in Appendix B. The results show that local hypotheses such as Constant Speed/Acceleration perform much better than global hypotheses such as LR and GPR. This is probably because, as virtual trajectory generation happens in an online fashion, it is hard to get a reliable fit using only limited data points on historical time steps." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 331, + 546, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 331, + 546, + 439 + ], + "spans": [ + { + "bbox": [ + 304, + 331, + 546, + 439 + ], + "type": "inline_equation", + "content": "\\Delta t" + }, + { + "bbox": [ + 304, + 331, + 546, + 439 + ], + "type": "text", + "content": " in OCM. There is a trade-off when choosing the time difference " + }, + { + "bbox": [ + 304, + 331, + 546, + 439 + ], + "type": "inline_equation", + "content": "\\Delta t" + }, + { + "bbox": [ + 304, + 331, + 546, + 439 + ], + "type": "text", + "content": " in OCM (Section 4). A large " + }, + { + "bbox": [ + 304, + 331, + 546, + 439 + ], + "type": "inline_equation", + "content": "\\Delta t" + }, + { + "bbox": [ + 304, + 331, + 546, + 439 + ], + "type": "text", + "content": " decreases the noise of velocity estimation, but is also likely to discourage approximating object motion as linear. Therefore, we study the influence of varying " + }, + { + "bbox": [ + 304, + 331, + 546, + 439 + ], + "type": "inline_equation", + "content": "\\Delta t" + }, + { + "bbox": [ + 304, + 331, + 546, + 439 + ], + "type": "text", + "content": " in Table 7. Our results agree with our analysis that increasing " + }, + { + "bbox": [ + 304, + 331, + 546, + 439 + ], + "type": "inline_equation", + "content": "\\Delta t" + }, + { + "bbox": [ + 304, + 331, + 546, + 439 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 304, + 331, + 546, + 439 + ], + "type": "inline_equation", + "content": "\\Delta t = 1" + }, + { + "bbox": [ + 304, + 331, + 546, + 439 + ], + "type": "text", + "content": " can boost performance. But increasing " + }, + { + "bbox": [ + 304, + 331, + 546, + 439 + ], + "type": "inline_equation", + "content": "\\Delta t" + }, + { + "bbox": [ + 304, + 331, + 546, + 439 + ], + "type": "text", + "content": " higher than a best-practice value instead hurts the performance because of the difficulty of maintaining the approximation of linear motion." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 446, + 378, + 459 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 446, + 378, + 459 + ], + "spans": [ + { + "bbox": [ + 306, + 446, + 378, + 459 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 461, + 545, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 461, + 545, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 461, + 545, + 616 + ], + "type": "text", + "content": "We analyze the popular motion-based tracker SORT and recognize its intrinsic limitations from using Kalman filter. These limitations significantly hurt tracking accuracy when the tracker fails to gain observations for supervision - likely caused by unreliable detectors, occlusion, or fast and non-linear target object motion. To address these issues, we propose Observation-Centric SORT (OC-SORT). OC-SORT is more robust to occlusion and non-linear object motion while keeping simple, online, and real-time. In our experiments on diverse datasets, OC-SORT significantly outperforms the state-of-the-art. The gain is especially significant for multi-object tracking under occlusion and non-linear object motion." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 617, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 546, + 714 + ], + "type": "text", + "content": "Acknowledgement. We thank David Held, Deva Ramanan, and Siddharth Ancha for the discussion about the theoretical modeling of OC-SORT. We thank Yuda Song for the help with the analysis of OCM error distribution. We also would like to thank Zhengyi Luo, Yuda Song, and Erica Weng for the detailed feedback on the paper writing. We also thank Yifu Zhang for sharing his experience with ByteTrack. This project was sponsored in part by NSF NRI award 2024173." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9693" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 84 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "text", + "content": "[1] Yaakov Bar-Shalom, Fred Daum, and Jim Huang. The probabilistic data association filter. IEEE Control Systems Magazine, 29(6):82-100, 2009." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 124, + 288, + 166 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 124, + 288, + 166 + ], + "spans": [ + { + "bbox": [ + 53, + 124, + 288, + 166 + ], + "type": "text", + "content": "[2] Sumit Basu, Irfan Essa, and Alex Pentland. Motion regularization for model-based head tracking. In Proceedings of 13th International Conference on Pattern Recognition, volume 3, pages 611-616. IEEE, 1996." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 166, + 288, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 166, + 288, + 209 + ], + "spans": [ + { + "bbox": [ + 53, + 166, + 288, + 209 + ], + "type": "text", + "content": "[3] Alex Bewley, Zongyuan Ge, Lionel Ott, Fabio Ramos, and Ben Upcroft. Simple online and realtime tracking. In 2016 IEEE international conference on image processing (ICIP), pages 3464-3468. IEEE, 2016." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 209, + 288, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 209, + 288, + 252 + ], + "spans": [ + { + "bbox": [ + 53, + 209, + 288, + 252 + ], + "type": "text", + "content": "[4] Guillem Brasó and Laura Leal-Taixe. Learning a neural solver for multiple object tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6247-6257, 2020." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 252, + 288, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 252, + 288, + 284 + ], + "spans": [ + { + "bbox": [ + 53, + 252, + 288, + 284 + ], + "type": "text", + "content": "[5] Jiarui Cai, Mingze Xu, Wei Li, Yuanjun Xiong, Wei Xia, Zhuowen Tu, and Stefano Soatto. Memot: multi-object tracking with memory. In CVPR, 2022." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 284, + 288, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 284, + 288, + 316 + ], + "spans": [ + { + "bbox": [ + 53, + 284, + 288, + 316 + ], + "type": "text", + "content": "[6] Jinkun Cao, Jiangmiao Pang, Xinshuo Weng, Rawal Khirodkar, and Kris M Kitani. Object tracking by hierarchical part-whole attention." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 316, + 288, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 316, + 288, + 369 + ], + "spans": [ + { + "bbox": [ + 53, + 316, + 288, + 369 + ], + "type": "text", + "content": "[7] Jinkun Cao, Hongyang Tang, Hao-Shu Fang, Xiaoyong Shen, Cewu Lu, and Yu-Wing Tai. Cross-domain adaptation for animal pose estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9498-9507, 2019." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 369, + 288, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 369, + 288, + 401 + ], + "spans": [ + { + "bbox": [ + 53, + 369, + 288, + 401 + ], + "type": "text", + "content": "[8] Jinkun Cao, Hao Wu, and Kris Kitani. Track targets by dense spatio-temporal position encoding. arXiv preprint arXiv:2210.09455, 2022." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 402, + 288, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 402, + 288, + 466 + ], + "spans": [ + { + "bbox": [ + 53, + 402, + 288, + 466 + ], + "type": "text", + "content": "[9] Ming-Fang Chang, John Lambert, Patsorn Sangkloy, Jagjeet Singh, Slawomir Bak, Andrew Hartnett, De Wang, Peter Carr, Simon Lucey, Deva Ramanan, et al. Argoverse: 3d tracking and forecasting with rich maps. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8748-8757, 2019." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 466, + 288, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 466, + 288, + 509 + ], + "spans": [ + { + "bbox": [ + 48, + 466, + 288, + 509 + ], + "type": "text", + "content": "[10] Yin-Wen Chang, Cho-Jui Hsieh, Kai-Wei Chang, Michael Ringgaard, and Chih-Jen Lin. Training and testing low-degree polynomial data mappings via linearsvm. Journal of Machine Learning Research, 11(4), 2010." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 509, + 288, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 509, + 288, + 552 + ], + "spans": [ + { + "bbox": [ + 48, + 509, + 288, + 552 + ], + "type": "text", + "content": "[11] Wongun Choi. Near-online multi-target tracking with aggregated local flow descriptor. In Proceedings of the IEEE international conference on computer vision, pages 3029-3037, 2015." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 552, + 288, + 594 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 552, + 288, + 594 + ], + "spans": [ + { + "bbox": [ + 48, + 552, + 288, + 594 + ], + "type": "text", + "content": "[12] Peng Chu, Jiang Wang, Quanzeng You, Haibin Ling, and Zicheng Liu. Transmot: Spatial-temporal graph transformer for multiple object tracking. arXiv preprint arXiv:2104.00194, 2021." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 595, + 288, + 649 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 595, + 288, + 649 + ], + "spans": [ + { + "bbox": [ + 48, + 595, + 288, + 649 + ], + "type": "text", + "content": "[13] Peng Dai, Renliang Weng, Wongun Choi, Changshui Zhang, Zhangping He, and Wei Ding. Learning a proposal classifier for multiple object tracking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2443-2452, 2021." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 649, + 288, + 702 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 649, + 288, + 702 + ], + "spans": [ + { + "bbox": [ + 48, + 649, + 288, + 702 + ], + "type": "text", + "content": "[14] Patrick Dendorfer, Hamid Rezatofighi, Anton Milan, Javen Shi, Daniel Cremers, Ian Reid, Stefan Roth, Konrad Schindler, and Laura Leal-Taixe. Mot20: A benchmark for multi object tracking in crowded scenes. arXiv preprint arXiv:2003.09003, 2020." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 702, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 702, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 702, + 288, + 713 + ], + "type": "text", + "content": "[15] David Duvenaud. Automatic model construction with Gaus" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 704 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 326, + 73, + 543, + 84 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 543, + 84 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 543, + 84 + ], + "type": "text", + "content": "sian processes. PhD thesis, University of Cambridge, 2014." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 84, + 545, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 84, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 308, + 84, + 545, + 127 + ], + "type": "text", + "content": "[16] Hao-Shu Fang, Jinkun Cao, Yu-Wing Tai, and Cewu Lu. Pairwise body-part attention for recognizing human-object interactions. In Proceedings of the European conference on computer vision (ECCV), pages 51-67, 2018." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 127, + 545, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 127, + 545, + 170 + ], + "spans": [ + { + "bbox": [ + 308, + 127, + 545, + 170 + ], + "type": "text", + "content": "[17] Juergen Gall, Angela Yao, Nima Razavi, Luc Van Gool, and Victor Lempitsky. Hough forests for object detection, tracking, and action recognition. IEEE transactions on pattern analysis and machine intelligence, 33(11):2188-2202, 2011." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 170, + 545, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 170, + 545, + 201 + ], + "spans": [ + { + "bbox": [ + 308, + 170, + 545, + 201 + ], + "type": "text", + "content": "[18] Damien Garreau, Wittawat Jitkrittum, and Motonobu Kanagawa. Large sample analysis of the median heuristic. arXiv preprint arXiv:1707.07269, 2017." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 201, + 545, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 201, + 545, + 233 + ], + "spans": [ + { + "bbox": [ + 308, + 201, + 545, + 233 + ], + "type": "text", + "content": "[19] Zheng Ge, Songtao Liu, Feng Wang, Zeming Li, and Jian Sun. Yolox: Exceeding yolo series in 2021. arXiv preprint arXiv:2107.08430, 2021." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 233, + 545, + 276 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 233, + 545, + 276 + ], + "spans": [ + { + "bbox": [ + 308, + 233, + 545, + 276 + ], + "type": "text", + "content": "[20] Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset. The International Journal of Robotics Research, 32(11):1231-1237, 2013." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 276, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 276, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 308, + 276, + 545, + 319 + ], + "type": "text", + "content": "[21] Nicolas Franco Gonzalez, Andres Ospina, and Philippe Calvez. Smat: Smart multiple affinity metrics for multiple object tracking. In International Conference on Image Analysis and Recognition, pages 48-62. Springer, 2020." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 319, + 545, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 319, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 308, + 319, + 545, + 373 + ], + "type": "text", + "content": "[22] Fredrik Gustafsson, Fredrik Gunnarsson, Niclas Bergman, Urban Forssell, Jonas Jansson, Rickard Karlsson, and P-J Nordlund. Particle filters for positioning, navigation, and tracking. IEEE Transactions on signal processing, 50(2):425-437, 2002." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 373, + 545, + 416 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 373, + 545, + 416 + ], + "spans": [ + { + "bbox": [ + 308, + 373, + 545, + 416 + ], + "type": "text", + "content": "[23] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 416, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 416, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 308, + 416, + 545, + 437 + ], + "type": "text", + "content": "[24] David V Hinkley. On the ratio of two correlated normal random variables. Biometrika, 56(3):635-639, 1969." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 437, + 545, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 437, + 545, + 479 + ], + "spans": [ + { + "bbox": [ + 308, + 437, + 545, + 479 + ], + "type": "text", + "content": "[25] Andrea Hornakova, Roberto Henschel, Bodo Rosenhahn, and Paul Swoboda. Lifted disjoint paths with application in multiple object tracking. In International Conference on Machine Learning, pages 4364-4375. PMLR, 2020." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 479, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 479, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 308, + 479, + 545, + 544 + ], + "type": "text", + "content": "[26] Andrea Hornakova, Timo Kaiser, Paul Swoboda, Michal Rolinek, Bodo Rosenhahn, and Roberto Henschel. Making higher order mot scalable: An efficient approximate solver for lifted disjoint paths. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6330-6340, 2021." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 544, + 545, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 544, + 545, + 565 + ], + "spans": [ + { + "bbox": [ + 308, + 544, + 545, + 565 + ], + "type": "text", + "content": "[27] Andrew H Jazwinski. Stochastic processes and filtering theory. Courier Corporation, 2007." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 565, + 545, + 618 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 565, + 545, + 618 + ], + "spans": [ + { + "bbox": [ + 308, + 565, + 545, + 618 + ], + "type": "text", + "content": "[28] Simon J Julier and Jeffrey K Uhlmann. New extension of the kalman filter to nonlinear systems. In Signal processing, sensor fusion, and target recognition VI, volume 3068, pages 182-193. International Society for Optics and Photonics, 1997." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 618, + 545, + 651 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 618, + 545, + 651 + ], + "spans": [ + { + "bbox": [ + 308, + 618, + 545, + 651 + ], + "type": "text", + "content": "[29] Simon J Julier and Jeffrey K Uhlmann. Unscented filtering and nonlinear estimation. Proceedings of the IEEE, 92(3):401-422, 2004." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 651, + 545, + 671 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 651, + 545, + 671 + ], + "spans": [ + { + "bbox": [ + 308, + 651, + 545, + 671 + ], + "type": "text", + "content": "[30] Rudolf Emil Kalman et al. Contributions to the theory of optimal control. Bol. soc. mat. mexicana, 5(2):102-119, 1960." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 308, + 671, + 545, + 704 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 671, + 545, + 704 + ], + "spans": [ + { + "bbox": [ + 308, + 671, + 545, + 704 + ], + "type": "text", + "content": "[31] Kris M Kitani, Brian D Ziebart, James Andrew Bagnell, and Martial Hebert. Activity forecasting. In European conference on computer vision, pages 201-214. Springer, 2012." + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9694" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 286, + 704 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 48, + 73, + 286, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 73, + 286, + 105 + ], + "spans": [ + { + "bbox": [ + 48, + 73, + 286, + 105 + ], + "type": "text", + "content": "[32] Jonathan Ko and Dieter Fox. Gp-bayesfilters: Bayesian filtering using gaussian process prediction and observation models. Autonomous Robots, 27(1):75-90, 2009." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 106, + 286, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 106, + 286, + 148 + ], + "spans": [ + { + "bbox": [ + 49, + 106, + 286, + 148 + ], + "type": "text", + "content": "[33] Parth Kothari, Sven Kreiss, and Alexandre Alahi. Human trajectory forecasting in crowds: A deep learning perspective. IEEE Transactions on Intelligent Transportation Systems, 2021." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 148, + 286, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 148, + 286, + 169 + ], + "spans": [ + { + "bbox": [ + 49, + 148, + 286, + 169 + ], + "type": "text", + "content": "[34] Erich L Lehmann and George Casella. Theory of point estimation. Springer Science & Business Media, 2006." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 170, + 286, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 170, + 286, + 212 + ], + "spans": [ + { + "bbox": [ + 49, + 170, + 286, + 212 + ], + "type": "text", + "content": "[35] Wei Li, Yuanjun Xiong, Shuo Yang, Mingze Xu, Yongxin Wang, and Wei Xia. Semi-tcl: Semi-supervised track contrastive representation learning. arXiv preprint arXiv:2107.02396, 2021." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 213, + 286, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 213, + 286, + 255 + ], + "spans": [ + { + "bbox": [ + 49, + 213, + 286, + 255 + ], + "type": "text", + "content": "[36] Chao Liang, Zhipeng Zhang, Yi Lu, Xue Zhou, Bing Li, Xiyong Ye, and Jianxiao Zou. Rethinking the competition between detection and reid in multi-object tracking. arXiv preprint arXiv:2010.12138, 2020." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 255, + 286, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 255, + 286, + 308 + ], + "spans": [ + { + "bbox": [ + 49, + 255, + 286, + 308 + ], + "type": "text", + "content": "[37] Jonathon Luiten, Aljosa Osep, Patrick Dendorfer, Philip Torr, Andreas Geiger, Laura Leal-Taixe, and Bastian Leibe. Hota: A higher order metric for evaluating multi-object tracking. International journal of computer vision, 129(2):548-578, 2021." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 309, + 286, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 309, + 286, + 341 + ], + "spans": [ + { + "bbox": [ + 49, + 309, + 286, + 341 + ], + "type": "text", + "content": "[38] Gerard Maggiolino, Adnan Ahmad, Jinkun Cao, and Kris Kitani. Deep oc-sort: Multi-pedestrian tracking by adaptive re-identification. arXiv preprint arXiv:2302.11813, 2023." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 341, + 286, + 383 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 341, + 286, + 383 + ], + "spans": [ + { + "bbox": [ + 49, + 341, + 286, + 383 + ], + "type": "text", + "content": "[39] Tim Meinhardt, Alexander Kirillov, Laura Leal-Taixe, and Christoph Feichtenhofer. Trackformer: Multi-object tracking with transformers. arXiv preprint arXiv:2101.02702, 2021." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 384, + 286, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 384, + 286, + 426 + ], + "spans": [ + { + "bbox": [ + 49, + 384, + 286, + 426 + ], + "type": "text", + "content": "[40] Stan Melax, Leonid Keselman, and Sterling Orsten. Dynamics based 3d skeletal hand tracking. In Proceedings of the ACM SIGGRAPH Symposium on Interactive 3D Graphics and Games, pages 184-184, 2013." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 426, + 286, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 426, + 286, + 458 + ], + "spans": [ + { + "bbox": [ + 49, + 426, + 286, + 458 + ], + "type": "text", + "content": "[41] Anton Milan, Laura Leal-Taixe, Ian Reid, Stefan Roth, and Konrad Schindler. Mot16: A benchmark for multi-object tracking. arXiv preprint arXiv:1603.00831, 2016." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 459, + 286, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 459, + 286, + 512 + ], + "spans": [ + { + "bbox": [ + 49, + 459, + 286, + 512 + ], + "type": "text", + "content": "[42] Jiangmiao Pang, Linlu Qiu, Xia Li, Haofeng Chen, Qi Li, Trevor Darrell, and Fisher Yu. Quasi-dense similarity learning for multiple object tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 164-173, 2021." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 513, + 286, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 513, + 286, + 554 + ], + "spans": [ + { + "bbox": [ + 49, + 513, + 286, + 554 + ], + "type": "text", + "content": "[43] Dezhi Peng, Zikai Sun, Zirong Chen, Zirui Cai, Lele Xie, and Lianwen Jin. Detecting heads using feature refine net and cascaded multi-scale architecture. arXiv preprint arXiv:1803.09256, 2018." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 555, + 286, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 555, + 286, + 586 + ], + "spans": [ + { + "bbox": [ + 49, + 555, + 286, + 586 + ], + "type": "text", + "content": "[44] Lawrence Rabiner and Biinghwang Juang. An introduction to hidden markov models. *iiee assp magazine*, 3(1):4-16, 1986." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 49, + 586, + 286, + 639 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 586, + 286, + 639 + ], + "spans": [ + { + "bbox": [ + 49, + 586, + 286, + 639 + ], + "type": "text", + "content": "[45] Akshay Rangesh, Pranav Maheshwari, Mez Gebre, Siddhesh Mhatre, Vahid Ramezani, and Mohan M Trivedi. Trackmpnn: A message passing graph neural architecture for multi-object tracking. arXiv preprint arXiv:2101.04206, 2021." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 49, + 640, + 286, + 684 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 640, + 286, + 684 + ], + "spans": [ + { + "bbox": [ + 49, + 640, + 286, + 684 + ], + "type": "text", + "content": "[46] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 779-788, 2016." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 49, + 684, + 286, + 704 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 684, + 286, + 704 + ], + "spans": [ + { + "bbox": [ + 49, + 684, + 286, + 704 + ], + "type": "text", + "content": "[47] Steven Reece and Stephen Roberts. An introduction to gaussian processes for the kalman filter expert. In 2010 13th In-" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 74, + 545, + 707 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 327, + 74, + 544, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 74, + 544, + 94 + ], + "spans": [ + { + "bbox": [ + 327, + 74, + 544, + 94 + ], + "type": "text", + "content": "ternational Conference on Information Fusion, pages 1-9. IEEE, 2010." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 95, + 545, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 95, + 545, + 137 + ], + "spans": [ + { + "bbox": [ + 308, + 95, + 545, + 137 + ], + "type": "text", + "content": "[48] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural information processing systems, 28, 2015." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 138, + 545, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 138, + 545, + 201 + ], + "spans": [ + { + "bbox": [ + 308, + 138, + 545, + 201 + ], + "type": "text", + "content": "[49] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 658-666, 2019." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 202, + 545, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 202, + 545, + 267 + ], + "spans": [ + { + "bbox": [ + 308, + 202, + 545, + 267 + ], + "type": "text", + "content": "[50] Toby Sharp, Cem Keskin, Duncan Robertson, Jonathan Taylor, Jamie Shotton, David Kim, Christoph Rhemann, Ido Leichter, Alon Vinnikov, Yichen Wei, et al. Accurate, robust, and flexible real-time hand tracking. In Proceedings of the 33rd annual ACM conference on human factors in computing systems, pages 3633-3642, 2015." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 268, + 545, + 299 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 268, + 545, + 299 + ], + "spans": [ + { + "bbox": [ + 308, + 268, + 545, + 299 + ], + "type": "text", + "content": "[51] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 300, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 300, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 308, + 300, + 545, + 342 + ], + "type": "text", + "content": "[52] Gerald L Smith, Stanley F Schmidt, and Leonard A McGee. Application of statistical filter theory to the optimal estimation of position and velocity on board a circumlunar vehicle. National Aeronautics and Space Administration, 1962." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 342, + 545, + 395 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 342, + 545, + 395 + ], + "spans": [ + { + "bbox": [ + 308, + 342, + 545, + 395 + ], + "type": "text", + "content": "[53] Daniel Stadler and Jurgen Beyerer. Improving multiple pedestrian tracking by track management and occlusion handling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10958-10967, 2021." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 396, + 545, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 396, + 545, + 438 + ], + "spans": [ + { + "bbox": [ + 308, + 396, + 545, + 438 + ], + "type": "text", + "content": "[54] Peize Sun, Jinkun Cao, Yi Jiang, Zehuan Yuan, Song Bai, Kris Kitani, and Ping Luo. Dancetrack: Multi-object tracking in uniform appearance and diverse motion. arXiv preprint arXiv:2111.14690, 2021." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 439, + 545, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 439, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 308, + 439, + 545, + 480 + ], + "type": "text", + "content": "[55] Peize Sun, Jinkun Cao, Yi Jiang, Rufeng Zhang, Enze Xie, Zehuan Yuan, Changhu Wang, and Ping Luo. Transtrack: Multiple object tracking with transformer. arXiv preprint arXiv:2012.15460, 2020." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 481, + 545, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 481, + 545, + 534 + ], + "spans": [ + { + "bbox": [ + 308, + 481, + 545, + 534 + ], + "type": "text", + "content": "[56] Ramana Sundararaman, Cedric De Almeida Braga, Eric Marchand, and Julien Pette. Tracking pedestrian heads in dense crowd. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3865-3875, 2021." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 535, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 535, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 308, + 535, + 545, + 567 + ], + "type": "text", + "content": "[57] Pavel Tokmakov, Jie Li, Wolfram Burgard, and Adrien Gaidon. Learning to track with object permanence. In ICCV, pages 10860-10869, 2021." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 567, + 545, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 567, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 308, + 567, + 545, + 609 + ], + "type": "text", + "content": "[58] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 610, + 545, + 674 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 610, + 545, + 674 + ], + "spans": [ + { + "bbox": [ + 308, + 610, + 545, + 674 + ], + "type": "text", + "content": "[59] Gaoang Wang, Renshu Gu, Zuozhu Liu, Weijie Hu, Mingli Song, and Jenq-Neng Hwang. Track without appearance: Learn box and tracklet embedding with local and global motion patterns for vehicle tracking. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9876-9886, 2021." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 675, + 545, + 707 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 675, + 545, + 707 + ], + "spans": [ + { + "bbox": [ + 308, + 675, + 545, + 707 + ], + "type": "text", + "content": "[60] Shuai Wang, Hao Sheng, Yang Zhang, Yubin Wu, and Zhang Xiong. A general recurrent tracking framework without real data. In Proceedings of the IEEE/CVF International Confer-" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9695" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 73, + 287, + 639 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 67, + 73, + 261, + 83 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 73, + 261, + 83 + ], + "spans": [ + { + "bbox": [ + 67, + 73, + 261, + 83 + ], + "type": "text", + "content": "ence on Computer Vision, pages 13219-13228, 2021." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 84, + 287, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 84, + 287, + 126 + ], + "spans": [ + { + "bbox": [ + 48, + 84, + 287, + 126 + ], + "type": "text", + "content": "[61] Yongxin Wang, Kris Kitani, and Xinshuo Weng. Joint object detection and multi-object tracking with graph neural networks. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 13708-13715. IEEE, 2021." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 126, + 287, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 126, + 287, + 158 + ], + "spans": [ + { + "bbox": [ + 48, + 126, + 287, + 158 + ], + "type": "text", + "content": "[62] Christopher Williams and Carl Rasmussen. Gaussian processes for regression. Advances in neural information processing systems, 8, 1995." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 158, + 287, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 158, + 287, + 201 + ], + "spans": [ + { + "bbox": [ + 48, + 158, + 287, + 201 + ], + "type": "text", + "content": "[63] Nicolai Wojke, Alex Bewley, and Dietrich Paulus. Simple online and realtime tracking with a deep association metric. In 2017 IEEE international conference on image processing (ICIP), pages 3645-3649. IEEE, 2017." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 201, + 287, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 201, + 287, + 255 + ], + "spans": [ + { + "bbox": [ + 48, + 201, + 287, + 255 + ], + "type": "text", + "content": "[64] Jialian Wu, Jiale Cao, Liangchen Song, Yu Wang, Ming Yang, and Junsong Yuan. Track to detect and segment: An online multi-object tracker. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12352-12361, 2021." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 255, + 287, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 255, + 287, + 298 + ], + "spans": [ + { + "bbox": [ + 48, + 255, + 287, + 298 + ], + "type": "text", + "content": "[65] Yu Xiang, Alexandre Alahi, and Silvio Savarese. Learning to track: Online multi-object tracking by decision making. In Proceedings of the IEEE international conference on computer vision, pages 4705-4713, 2015." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 298, + 287, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 298, + 287, + 330 + ], + "spans": [ + { + "bbox": [ + 48, + 298, + 287, + 330 + ], + "type": "text", + "content": "[66] Yuliang Xiu, Jiefeng Li, Haoyu Wang, Yinghong Fang, and Cewu Lu. Pose flow: Efficient online pose tracking. arXiv preprint arXiv:1802.00977, 2018." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 330, + 287, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 330, + 287, + 373 + ], + "spans": [ + { + "bbox": [ + 48, + 330, + 287, + 373 + ], + "type": "text", + "content": "[67] Yihong Xu, Yutong Ban, Guillaume Delorme, Chuang Gan, Daniela Rus, and Xavier Alameda-Pineda. Transcenter: Transformers with dense queries for multiple-object tracking. arXiv preprint arXiv:2103.15145, 2021." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 373, + 287, + 405 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 373, + 287, + 405 + ], + "spans": [ + { + "bbox": [ + 48, + 373, + 287, + 405 + ], + "type": "text", + "content": "[68] Bin Yan, Yi Jiang, Peize Sun, Dong Wang, Zehuan Yuan, Ping Luo, and Huchuan Lu. Towards grand unification of object tracking. In ECCV. Springer, 2022." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 405, + 287, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 405, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 48, + 405, + 287, + 426 + ], + "type": "text", + "content": "[69] Fangao Zeng and et al. Motr: End-to-end multiple-object tracking with transformer. In ECCV. Springer, 2022." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 426, + 287, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 426, + 287, + 468 + ], + "spans": [ + { + "bbox": [ + 48, + 426, + 287, + 468 + ], + "type": "text", + "content": "[70] Yifu Zhang, Peize Sun, Yi Jiang, Dongdong Yu, Zehuan Yuan, Ping Luo, Wenyu Liu, and Xinggang Wang. Byetetrack: Multi-object tracking by associating every detection box. arXiv preprint arXiv:2110.06864, 2021." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 468, + 287, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 468, + 287, + 511 + ], + "spans": [ + { + "bbox": [ + 48, + 468, + 287, + 511 + ], + "type": "text", + "content": "[71] Yifu Zhang, Chunyu Wang, Xinggang Wang, Wenjun Zeng, and Wenyu Liu. Fairmot: On the fairness of detection and re-identification in multiple object tracking. International Journal of Computer Vision, 129(11):3069-3087, 2021." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 511, + 287, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 511, + 287, + 564 + ], + "spans": [ + { + "bbox": [ + 48, + 511, + 287, + 564 + ], + "type": "text", + "content": "[72] Zhaohui Zheng, Ping Wang, Wei Liu, Jinze Li, Rongguang Ye, and Dongwei Ren. Distance-iou loss: Faster and better learning for bounding box regression. In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 12993-13000, 2020." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 564, + 287, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 564, + 287, + 597 + ], + "spans": [ + { + "bbox": [ + 48, + 564, + 287, + 597 + ], + "type": "text", + "content": "[73] Xingyi Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Tracking objects as points. In ECCV, pages 474-490. Springer, 2020." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 597, + 287, + 618 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 597, + 287, + 618 + ], + "spans": [ + { + "bbox": [ + 48, + 597, + 287, + 618 + ], + "type": "text", + "content": "[74] Xingyi Zhou, Dequan Wang, and Philipp Kähenbuhl. Objects as points. arXiv preprint arXiv:1904.07850, 2019." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 618, + 287, + 639 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 618, + 287, + 639 + ], + "spans": [ + { + "bbox": [ + 48, + 618, + 287, + 639 + ], + "type": "text", + "content": "[75] Xingyi Zhou, Tianwei Yin, Vladlen Koltun, and Philipp Krahenbuhl. Global tracking transformers. In CVPR, 2022." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9696" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/OcTr_ Octree-Based Transformer for 3D Object Detection/b6928c01-d853-4d0d-9874-89b3af0f167a_content_list.json b/2023/OcTr_ Octree-Based Transformer for 3D Object Detection/b6928c01-d853-4d0d-9874-89b3af0f167a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..727db2c21ea436fb0bc0603c2ed81b8db1915775 --- /dev/null +++ b/2023/OcTr_ Octree-Based Transformer for 3D Object Detection/b6928c01-d853-4d0d-9874-89b3af0f167a_content_list.json @@ -0,0 +1,1686 @@ +[ + { + "type": "text", + "text": "OcTr: Octree-based Transformer for 3D Object Detection", + "text_level": 1, + "bbox": [ + 192, + 130, + 777, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chao Zhou $^{1,2}$ , Yanan Zhang $^{1,2}$ , Jiaxin Chen $^{2}$ , Di Huang $^{1,2,3*}$", + "bbox": [ + 250, + 179, + 723, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1State Key Laboratory of Software Development Environment, Beihang University, Beijing, China", + "bbox": [ + 98, + 198, + 872, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ School of Computer Science and Engineering, Beihang University, Beijing, China", + "bbox": [ + 158, + 215, + 810, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ Hangzhou Innovation Institute, Beihang University, Hangzhou, China", + "bbox": [ + 207, + 233, + 763, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{zhouchaobeing, zhangyanan, jiaxinchen, dhuang}@buaa.edu.cn", + "bbox": [ + 220, + 253, + 741, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 304, + 312, + 319 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A key challenge for LiDAR-based 3D object detection is to capture sufficient features from large scale 3D scenes especially for distant or/and occluded objects. Albeit recent efforts made by Transformers with the long sequence modeling capability, they fail to properly balance the accuracy and efficiency, suffering from inadequate receptive fields or coarse-grained holistic correlations. In this paper, we propose an Octree-based Transformer, named OcTr, to address this issue. It first constructs a dynamic octree on the hierarchical feature pyramid through conducting self-attention on the top level and then recursively propagates to the level below restricted by the octants, which captures rich global context in a coarse-to-fine manner while maintaining the computational complexity under control. Furthermore, for enhanced foreground perception, we propose a hybrid positional embedding, composed of the semantic-aware positional embedding and attention mask, to fully exploit semantic and geometry clues. Extensive experiments are conducted on the Waymo Open Dataset and KITTI Dataset, and OcTr reaches newly state-of-the-art results.", + "bbox": [ + 75, + 335, + 473, + 638 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 667, + 207, + 683 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D object detection from point clouds has received extensive attention during the past decade for its ability to provide accurate and stable recognition and localization in autonomous driving perception systems. In this task, feature learning plays a very fundamental and crucial role; yet it is rather challenging due to not only the disordered and sparse nature of data sampling, but also to insufficient acquisition under occlusion or at a distance. To address this issue, many methods have been proposed, which can be taxonomized into two major classes, i.e. grid-based and point-based. The former first regularize point clouds into multi-view images or voxels and then apply 2D or 3D CNNs to build shape rep-", + "bbox": [ + 75, + 694, + 468, + 876 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2b31e2f8605ec3d3cb51f157ddcc16cee3ae8767cbf4da1de378779044aa5931.jpg", + "image_caption": [ + "(1) Fixed Pattern Limited Receptive Field", + "(2) Set Proxy Limited Representation" + ], + "image_footnote": [], + "bbox": [ + 509, + 314, + 883, + 433 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9ffd3a5a130671d8edff22494faaded510bca63d3781060532966f5ddbaef2e4.jpg", + "image_caption": [ + "Figure 1. Illustration of three sparsification strategies of attention matrices. Fixed pattern (1) narrows receptive fields and set proxy (2) discards elaborate correlations. The proposed octree construction (3) keeps the global receptive field in a coarse-grained manner while maintaining fine-grained representations." + ], + "image_footnote": [], + "bbox": [ + 534, + 454, + 859, + 571 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "resentations [4, 52], while the latter directly conduct MLP based networks such as PointNet++ [33] and DGCNN [50] on original points for geometry description [32, 40, 42, 60]. Unfortunately, they fail to capture necessary context information through the small receptive fields in the deep models, leading to limited results.", + "bbox": [ + 496, + 686, + 890, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Witnessing the recent success of Transformers in NLP, many studies have investigated and extended such architectures for 3D vision [24, 29, 58, 61]. Transformers are reputed to model long-range dependencies, delivering global receptive fields, and to be suitable for scattered inputs of arbitrary sizes. Meanwhile, in contrast to those static weights that are learned in convolutions, Transformers dynamically aggregate the input features according to the relationships", + "bbox": [ + 496, + 779, + 892, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*indicates the corresponding author.", + "bbox": [ + 94, + 887, + 287, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "5166", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "between tokens. Regarding the case in 3D object detection, compared to point-based Transformers [11,61], voxel-based ones show the superiority in efficiency. However, they tend to suffer heavy computations when dealing with large scale scenes because of the quadratic complexity of Transformers, with the underlying dilemma between the grid size and the grid amount in voxelization. Taking the KITTI dataset as an example, it is unrealistic for Transformers to operate on the feature map with the spatial shape of $200 \\times 176 \\times 5$ , which is commonly adopted in most of the detection heads [38,46,52,55].", + "bbox": [ + 75, + 90, + 472, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "More recently, there have appeared an influx of efficient self-attention model variants that attempt to tackle long sequences as input. They generally sparsify the attention matrix by fixed patterns [7, 23, 34], learned patterns [21, 45] or a combination of them [1, 56]. Fixed patterns chunk the input sequence into blocks of local windows or dilation windows, whilst learned patterns determine a notion of token relevance and eliminate or cluster outliers. Specific to 3D object detection from point clouds, VoTr [24] modifies self-attention with pre-defined patterns including local windows and stride dilation ones in a sparse query manner, and the dilation mechanism enlarges the receptive field by sampling attending tokens in a radius. SST [9] splits input tokens into non-overlapping patterns in a block-wise way and enables window shifting to capture cross-window correlation. Despite some improvements reported, they both only achieve bigger local receptive fields rather than the expected global ones, and computations still increase rapidly with the expansion of receptive fields.", + "bbox": [ + 75, + 260, + 472, + 547 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Another alternative on self-attention is to take advantage of a proxy memory bank which has the access to the entire sequence tokens [1, 2, 56]. By using a small number of induced proxies to compress the whole sequence, it diffuses the global context efficiently. VoxSet [12] adapts Set Transformer [19] to 3D object detection and exploits an induced set to model a set-to-set point cloud translation. With the help of the compressed global proxies and Conv-FFN, it obtains a global receptive field; nevertheless, as they admit, it is sub-optimal to set only a few latent codes as proxies for a large 3D scene, prone to impairing the representation of different point cloud structures and their correlations. Therefore, there remains space for a stronger solution.", + "bbox": [ + 75, + 550, + 472, + 747 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we present a novel Transformer network, namely Octree-based Transformer (OcTr), for 3D object detection. We firstly devise an octree-based learnable sparse pattern, i.e. OctAttn, which meticulously and efficiently encodes point clouds of scenes as shown in Fig. 1. The OctAttn module constructs a feature pyramid by gathering and applies self-attention to the top level of the feature pyramid to select the most relevant tokens, which are deemed as the octants to be divided in the subsequent. When propagating to the level below, the key/value inputs are restricted by the", + "bbox": [ + 75, + 750, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "octants from the top. Through recursively conducting this process, OctAttn captures rich global context features by a global receptive field in a coarse-to-fine manner while reducing the quadratic complexity of vanilla self-attention to the linear complexity. In addition, for better foreground perception, we propose a hybrid positional embedding, which consists of the semantic-aware positional embedding and attention mask, to fully exploit geometry and semantic clues. Thanks to the designs above, OcTr delivers a competitive trade-off between accuracy and efficiency.", + "bbox": [ + 496, + 90, + 893, + 242 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contribution is summarized in three-fold:", + "bbox": [ + 517, + 242, + 821, + 257 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We propose OcTr for voxel-based 3D object detection, which efficiently learns enhanced representations by a global receptive field with rich contexts.", + "2. We propose an octree-based learnable attention sparsification scheme (OctAttn) and a hybrid positional embedding combining geometry and semantics.", + "3. We carry out experiments on the Waymo Open Dataset (WOD) and the KITTI dataset and report state-of-the-art performance with significant gains on far objects." + ], + "bbox": [ + 511, + 261, + 890, + 414 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 429, + 640, + 445 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. 3D Object Detection from Point Clouds", + "text_level": 1, + "bbox": [ + 500, + 454, + 836, + 470 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "There exist two prevailing point-cloud representations in 3D object detection, i.e. point-based and voxel-based.", + "bbox": [ + 500, + 478, + 890, + 508 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The point-based methods [31,40,42,60] directly process raw point clouds in the irregular 3D space. As a pioneering attempt, F-Pointnet [32] employs instance segmentation in frustums to extract proposals. VoteNet [31] clusters objects from the surface in a deep Hough voting manner. PointRCNN [40] generates 3D RoIs with foreground segmentation and applies an RCNN-style [35] two-stage refinement. Different from PointRCNN, some existing methods [3,53,59] build a lightweight and efficient single stage 3D object detection framework. However, the current point-based methods still suffer from a large computation burden, which is not suitable for large-scale point cloud scenes.", + "bbox": [ + 496, + 508, + 893, + 688 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The voxel-based ones [4,6,13,52,55,62,63] conduct voxelization on entire point clouds to construct regular grids. VoxelNet [63] exploits the voxel feature encoding layer with 3D convolutions to extract the feature of each voxel. SECOND [52] improves the model with sparse 3D convolutions, significantly increasing the speed of both training and inference. Pointpillars [18] compacts point clouds into vertical columns and encodes them with 2D CNNs. Several recent methods [13, 38, 54] also explore merging point-based and voxel-based networks into one framework for complementary features from different representations of point clouds. Unfortunately, they all use small convolution kernels with limited receptive fields, which are not competent to capture global context that is important to 3D detection.", + "bbox": [ + 496, + 689, + 893, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "5167", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e9ca629127423699a851f636446a042a215fcb31ee294f13960afbf86af50a5f.jpg", + "image_caption": [ + "Figure 2. Framework overview of the proposed Octree-based Transformer (OcTr) model." + ], + "image_footnote": [], + "bbox": [ + 86, + 92, + 885, + 244 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Transformer in 3D Vision", + "text_level": 1, + "bbox": [ + 76, + 301, + 310, + 316 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Inspired by the great success of the self-attention mechanism in NLP [47] and CV [8, 23], Transformers have been adapted to 3D vision for their ability to capture long-range dependencies. For instance, Point Transformer [61] brings in vector attention that modulates individual feature channels for point cloud classification and segmentation; PCT [11] presents offset-attention with the implicit Laplace operator and normalization refinement which is more suitable for point cloud learning. To address the high latency, some methods [30, 57] adopt voxels or patches for acceleration.", + "bbox": [ + 75, + 325, + 468, + 474 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For 3D detection, 3DETR [28] treats and predicts bounding boxes as sequences in an end-to-end manner. CT3D [37] leverages a channel-wise Transformer architecture to refine the RoI head. To learn context-aware representations, some studies [9,12,24,29] introduce Transformers into a point- or voxel-based encoder. Pointformer [29] stacks local, global and local-global Transformers based on the point-based encoder; VoTr [24] exploits dilated attention with fast query to enlarge receptive fields; VoxSet [12] builds an induced point set as proxies of global context and applies point-to-point translation using voxels as mediums; and SST [9] embraces single strides without down-sampling and conducts window attention combining with its shifted version. Even though they all expand receptive fields by diverse Transformer variants, global context is not adequately involved or efficiently utilized. In contrast, we propose Octree-based Transformer (OcTr) for voxel-based 3D object detection, achieving a true global receptive field that balances accuracy and efficiency.", + "bbox": [ + 75, + 476, + 470, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 761, + 168, + 776 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Framework", + "text_level": 1, + "bbox": [ + 76, + 786, + 204, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This sub-section describes the overall framework of the proposed OcTr model as shown in Fig. 2. Specifically, as inspired by [51], we first voxelize the point cloud into regular grids and adopt the sparse 3D convolution for patch embedding, where the girds are regarded as the \"tokens\" and are passed through the Octree Transformer Blocks", + "bbox": [ + 75, + 809, + 470, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "(OTB). Compared with the vanilla Transformer block, the self-attention module is substituted by the proposed octree-attention $OctAttn$ , which encodes global context in a more efficient way. After applying a hybrid semantic embedding on multi-scale features, we sequentially stack two OTBs, tailed by a down-sampling layer. The voxel features are then projected into the BEV view by point-wise convolutions and are passed through a multi-scale dense 2D backbone. Ultimately, an anchor-based or anchor-free RPN head is used for 3D proposal generation, and the RoI head is optional for refinement. Note that our OcTr can be adopted for most of the voxel-based detection frameworks by simply altering 3D backbones.", + "bbox": [ + 496, + 301, + 893, + 498 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Self-attention Revisit", + "text_level": 1, + "bbox": [ + 500, + 506, + 697, + 520 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "According to [47], Transformer encoder blocks typically include a multi-head self-attention (MHSA) mechanism, a feed-forward network (FFN), a normalization function, and the residual connections [14]. Given an input sequence $X$ , the principle part of MHSA is formulated as", + "bbox": [ + 496, + 529, + 893, + 604 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {M H S A} (X) = \\sum_ {h = 1} ^ {H} W _ {h} \\left[ \\sigma \\left(\\frac {X W _ {q} W _ {k} ^ {T} X ^ {T}}{\\sqrt {d}}\\right) \\cdot X W _ {v} \\right], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 611, + 890, + 652 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $h$ denotes the index of the head, and $H$ , $\\sigma$ , $W$ and $d$ are the amount of heads, softmax function, learnable weight and feature dimension, respectively. The subscripts of $q$ , $k$ and $v$ indicate query, key and value. The inputs and outputs of the MHSA module are connected by residual connectors and normalization layers. The MLP-based FFN connects its inputs/outputs in a similar manner.", + "bbox": [ + 496, + 657, + 893, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Octree Attention", + "text_level": 1, + "bbox": [ + 500, + 771, + 666, + 786 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "An octree is a multi-scale asymmetric and efficient representation for unstructured 3D data such as point clouds. To build an octree for the input point cloud, we recursively sub-divide it in the breadth-first order until the pre-defined octree depth is reached. Whether to sub-divide an octant is determined by the occupancy [36, 48], the surface approximation [44, 49] or a learning algorithm [26].", + "bbox": [ + 496, + 795, + 893, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "5168", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/96bc99e00865db3b240e60cadeb29d51d8cd8d5d934e9de16e6f190fd171ed3a.jpg", + "image_caption": [ + "Figure 3. Octree Attention (OctAttn). The OctAttn module constructs a dynamic octree on a hierarchical feature pyramid with a partitive criterion of attention scores. For illustration, an octree representation is individually pruned from the pyramid for each grid in $F$ . For example, the blue grids are the divided octants for the grids with stars. The detailed structures of OcTr-Top and OcTr-Base are shown at the right side." + ], + "image_footnote": [], + "bbox": [ + 84, + 92, + 566, + 294 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/563e90e76e17443f4be4476c9793a81ee2db92313c1c830bfac3a3aa51bc6940.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 573, + 95, + 728, + 290 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/034cdb1500198579c1ba7e2d834de7809406d207fea9c28dee8523b9c58421b1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 733, + 98, + 887, + 290 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Despite of being resource-friendly, the octree representation is discrete and non-differentiable, which motivates us to ameliorate previous cumbersome pre-processing division by introducing a novel octree-based attention mechanism. As depicted in [21, 45], the attention matrices calculated by self-attention imply the relevance of input tokens and guide feature selection. We thus prune the dense attention matrices of the multi-scale feature pyramid to sparse octree attention in an adaptive and parallel manner, namely OctAttn.", + "bbox": [ + 75, + 378, + 468, + 527 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Fig. 3, let the output feature map and coordinates of convolutional patch embedding be $F_0 \\in \\mathbb{R}^{M \\times d}$ and $I_0 \\in \\mathbb{R}^{M \\times 3}$ ; $M$ and $d$ indicate the amount of the nonempty grids in a batch and feature dimension, respectively. Based on $F_0$ , we generate a multi-scale feature pyramid as", + "bbox": [ + 76, + 532, + 468, + 608 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nC = \\left\\{F _ {n}, I _ {n} \\right\\} ^ {N}, n \\in [ 0, N), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 628, + 468, + 646 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where", + "bbox": [ + 76, + 659, + 122, + 672 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nI _ {n} = \\left\\lfloor \\frac {I _ {0}}{2 ^ {n}} \\right\\rfloor , F _ {n} = \\operatorname {B N} \\left(S _ {\\max } \\left(F _ {0}, I _ {n}\\right)\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 691, + 468, + 720 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$n$ , $N$ , BN and $S_{max}$ are the index of the level of the multiscale feature pyramid, the height of the pyramid, batch normalization [16] and the max scatter function, respectively.", + "bbox": [ + 75, + 732, + 468, + 777 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The pruning begins from the top of the pyramid. The top feature map, i.e. $F_{N - 1}$ , is reorganized to dense input tokens $\\bar{F}_{N - 1} \\in \\mathbb{R}^{B \\times m_{N - 1} \\times d}$ , where $B$ and $m_{N - 1}$ are the batch size and maximum number of non-empty voxels per batch, respectively. The voxel is padded with 0 if it is empty. As shown in Eq. (4), the MHSA takes $\\bar{F}_{N - 1}$ as input and outputs the attention score matrices $\\mathcal{A}_{N - 1} \\in \\mathbb{R}^{B \\times m_{N - 1} \\times m_{N - 1}}$ and attentive features $\\bar{F}_{N - 1}'$ as", + "bbox": [ + 75, + 780, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {A} _ {N - 1} = \\sum_ {h = 1} ^ {H} \\sigma (\\frac {\\bar {F} _ {N - 1} W _ {q} W _ {k} ^ {T} \\bar {F} _ {N - 1} ^ {T}}{\\sqrt {d}}), \\\\ \\bar {F} _ {N - 1} ^ {\\prime} = \\sum_ {h = 1} ^ {H} W _ {h} \\left[ \\sigma \\left(\\frac {\\bar {F} _ {N - 1} W _ {q} W _ {k} ^ {T} \\bar {F} _ {N - 1} ^ {T}}{\\sqrt {d}}\\right) \\cdot \\bar {F} _ {N - 1} W _ {v} \\right]. \\tag {4} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 508, + 402, + 890, + 502 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For each query token, we select the top $k$ attention scores as its most relevant token group in a row-wise way, denoted by $O_{N - 1}\\in \\mathbb{Z}^{B\\times m_{N - 1}\\times k}$ .", + "bbox": [ + 498, + 507, + 890, + 551 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "When propagating from top to bottom through the pyramid and reaching the $n$ -th level, we uniformly sample limited $K$ attending octants from the selected regions with features in $\\bar{F}_n$ and topk indices in $O_{n + 1}$ . We conduct cross-attention instead of self-attention with the dense query sequence of the $n$ -th level, $\\bar{Q}_n\\in \\mathbb{R}^{B\\times m_n\\times d}$ , and the compact sampled key/value sequence of the $n$ -th level, $\\bar{K}_n / \\bar{V}_n\\in \\mathbb{R}^{B\\times m_n\\times K\\times d}$ , which is formulated as below", + "bbox": [ + 498, + 551, + 890, + 672 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {A} _ {n} = \\sum_ {h = 1} ^ {H} \\sigma \\left(\\frac {\\bar {Q} _ {n} W _ {q} W _ {k} ^ {T} \\bar {K} _ {n} ^ {T}}{\\sqrt {d}}\\right), \\tag {5} \\\\ \\bar {F} _ {n} ^ {\\prime} = \\sum_ {h = 1} ^ {H} W _ {h} \\left[ \\sigma \\left(\\frac {\\bar {Q} _ {n} W _ {q} W _ {k} ^ {T} \\bar {K} _ {n} ^ {T}}{\\sqrt {d}}\\right) \\cdot \\bar {V} _ {n} W _ {v} \\right]. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 545, + 698, + 890, + 782 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Backing off the sampling, this can be treated as an attention mask on self-attention matrices. The above process is run recursively until reaching the bottom level of the pyramid.", + "bbox": [ + 498, + 794, + 890, + 840 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Furthermore, as the top $k$ selection is a hard decision that disables gradient back-propagation, we adopt the Gumbeltop $k$ technique [17] to perform a differentiable and continuous approximation by replacing the vanilla top $k$ selection.", + "bbox": [ + 498, + 840, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "5169", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The normalized scores used in top $k$ are derived from the distribution in Eq. (6) during training, maintaining the original ones during inference. $g,\\tau ,m_{n}$ denote the noise sampled from the Gumbel distribution, the temperature and the amount of non-empty voxels in layer $n$ , respectively.", + "bbox": [ + 75, + 90, + 468, + 167 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\np _ {i} = \\frac {\\exp \\left(\\left(\\mathcal {A} _ {n} ^ {i} + g _ {i}\\right) / \\tau\\right)}{\\sum_ {i} \\exp \\left(\\left(\\mathcal {A} _ {n} ^ {i} + g _ {i}\\right) / \\tau\\right)} \\in [ 0, 1 ]. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 179, + 468, + 228 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In order to leverage the multi-scale features in distinct spatial shapes, we concatenate them by upsampling, which is implemented by inverse indices of the scatter function, followed by a linear projection layer for aligning the input feature dimension.", + "bbox": [ + 75, + 234, + 468, + 309 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As the local context is generally critical for object detection, inspired by [7], we additionally introduce a Locally enhanced Positional Embedding (LePE) which enables local neighbor interactions on the value sequence. With submanifold sparse convolutions, we replace the residual connections in the attention mechanism with LePE.", + "bbox": [ + 75, + 311, + 468, + 401 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, OTB is formulated as follows", + "bbox": [ + 96, + 402, + 346, + 417 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {F} = \\operatorname {F C} \\left(\\left\\{F _ {N - 1} ^ {\\prime} \\mid F _ {N - 2} ^ {\\prime} \\right| \\dots \\left| F _ {0} ^ {\\prime} \\right\\}\\right) + \\operatorname {L e P E} \\left(F _ {0}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 116, + 441, + 468, + 469 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {F} ^ {\\prime} = \\mathrm {B N} (\\mathrm {F F N} (\\tilde {F})) + \\tilde {F},\n$$\n", + "text_format": "latex", + "bbox": [ + 119, + 465, + 297, + 483 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $F_{n}^{\\prime}\\in \\mathbb{R}^{m_0\\times d}$ denotes the compact tensor of the upsampled $\\bar{F}_n^\\prime$ , | indicates concatenation, and FC denotes the fully-connected layer.", + "bbox": [ + 76, + 498, + 468, + 544 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Besides, we analyze the time complexity of the octree attention as below", + "bbox": [ + 76, + 545, + 468, + 574 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {O} \\left(\\left(\\frac {M}{\\omega^ {N - 1}}\\right) ^ {2} + \\sum_ {n = 0} ^ {N - 2} \\frac {K M}{\\omega^ {n}}\\right) \\tag {8} \\\\ = \\mathcal {O} \\left(\\left(\\frac {M}{\\omega^ {N - 1}}\\right) ^ {2} + \\frac {\\omega}{\\omega - 1} K M \\left(1 - \\omega^ {1 - N}\\right)\\right), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 129, + 584, + 468, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\omega$ is the average down-sampling ratio in the sparse voxel representation.", + "bbox": [ + 76, + 669, + 468, + 699 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Semantic Positional Embedding", + "text_level": 1, + "bbox": [ + 76, + 710, + 357, + 726 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Due to the large proportion of background grids in point clouds, the attention matrices are dominated by background grid pairs, leading to a sub-optimal solution. To fully leverage the local 3D shape patterns and original voxel coordinates, we propose a hybrid positional embedding to capture both the geometry and semantic clues as displayed in Fig. 4.", + "bbox": [ + 75, + 734, + 468, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, we first segment foreground grids using the supervision from the ground truth. Segmentation scores are predicted by a sub-manifold sparse convolution branch with a sigmoid function; and the focal loss [22] is applied to balance the foreground and background.", + "bbox": [ + 75, + 825, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/899be8311de29cb25bdd3d9fb85ee5bea664f6148bd5813802b8e45d11e05ebc.jpg", + "image_caption": [ + "Figure 4. Illustration of semantic positional embedding, where an extra foreground segmentation branch is adopted with supervision and semantic scores are concatenated as absolute positional embeddings and serve as relative masks on attention matrices." + ], + "image_footnote": [], + "bbox": [ + 506, + 89, + 885, + 287 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We concatenate the center coordinate $(x, y, z)$ and the semantic score with the feature $f$ in a grid-wise manner, followed by a linear projection without the bias as below", + "bbox": [ + 498, + 364, + 890, + 410 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\operatorname {S A P E} (X) = \\operatorname {F C} _ {d + 4 \\rightarrow d} (\\{x, y, z, \\text {s c o r e} | f \\}) \\tag {9} \\\\ = \\operatorname {F C} _ {d \\rightarrow d} (f) + \\operatorname {F C} _ {4 \\rightarrow d} (x, y, z, s c o r e). \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 419, + 890, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Eq. (9) is equivalent to the absolute positional embedding (APE), thus being denoted as the Semantic APE (SAPE). The scatter function such as the mean, max and batch normalization naturally provide position and semantic information of the downsampled voxel grids, making it applicable for the multi-scale feature pyramid.", + "bbox": [ + 498, + 465, + 890, + 556 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Besides the semantic clues implicitly used in SAPE, we employ the Semantic Attention Mask (SAM) based on the segmentation scores and we mask the attention matrices to address correlations between inferior queries and superior keys in a simple yet effective way. Given a scalar attention matrix $\\mathcal{A} \\in \\mathbb{R}^{N_q \\times N_k}$ before softmax and the segmentation scores of query and key/value $S_q \\in \\mathbb{R}^{N_q}$ and $S_k \\in \\mathbb{R}^{N_k}$ , we formulate the attention matrix after softmax as", + "bbox": [ + 498, + 556, + 890, + 675 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {A} ^ {\\prime} = \\sigma (- \\Gamma \\cdot [ 1 - (S _ {q} \\geq \\delta_ {q}) (S _ {k} \\geq \\delta_ {k}) ^ {T} ] + \\mathcal {A}), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 686, + 890, + 705 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $N_{q},\\delta_{q},N_{k},\\delta_{k},\\sigma$ and $\\Gamma$ are the length and threshold of the query sequence, length and threshold of the key sequence, softmax function and infinite scalar, respectively.", + "bbox": [ + 498, + 710, + 890, + 755 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, we broadcast the semantic mask $\\mathcal{A}'$ to all heads.", + "bbox": [ + 517, + 756, + 888, + 770 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 784, + 630, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate the proposed OcTr network on the Waymo Open Dataset (WOD) and KITTI dataset, both of which are popular in 3D object detection. In this section, we introduce the benchmarks and implementation details, make comparison to the previous state-of-the-art counterparts, and ablate the key designs of OcTr.", + "bbox": [ + 498, + 809, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5170", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/aa326a65166577e124eeb1665378696b5235ce62f1edff364bf70673aa92202a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelVehicle (L1) \nmAP/mAPHVehicle (L2) \nmAP/mAPHPedes. (L1) \nmAP/mAPHPedes. (L2) \nmAP/mAPHCyclist (L1) \nmAP/mAPHCyclist (L2) \nmAP/mAPH
SECOND [52]70.96/70.3462.58/62.0265.23/54.2457.22/47.4957.13/55.6254.97/53.53
PointPillar [18]70.43/69.8362.18/61.6466.21/46.3258.18/40.6455.26/51.7553.18/49.80
PartA²Net [41]74.82/74.3265.88/65.4271.76/63.6462.53/55.3067.35/66.1565.05/63.89
PVRCNN [38]75.41/74.7467.44/66.8071.98/61.2463.70/53.9565.88/64.2563.39/61.82
CenterPoint [55]71.33/70.7663.16/62.6572.09/65.4964.27/58.2368.68/67.3966.11/64.87
LiDAR-RCNN [20]73.5/73.064.7/64.271.2/58.763.1/51.768.6/66.966.1/64.4
Voxel-RCNN [6]75.59/-66.59/--/--/--/--/-
PVRCNN++ [39]77.82/77.3269.07/68.6277.99/71.3669.92/63.7471.80/70.7169.31/68.26
SST† [9]76.22/75.7968.04/67.6481.39/74.0572.82/65.93-/--/-
PDV [15]76.85/76.3369.30/68.8174.19/65.9665.85/58.2868.71/67.5566.49/65.36
Ours78.12/77.6369.79/69.3480.76/74.3972.48/66.5272.58/71.5069.93/68.90
", + "bbox": [ + 84, + 89, + 885, + 296 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/adabb445440801418caab8cf123aa46ffc2ae67de1068a9e8cb8a05703c31e95.jpg", + "table_caption": [ + "Table 1. Performance on WOD with 202 validation sequences for vehicle (IoU=0.7), pedestrian (IoU=0.5) and cyclist (IoU=0.5), using 20% samples for training. All the results are achieved by the models simultaneously trained for 3 classes on single frames, except the ones of the model marked by $\\dagger$ , which is only trained for a single class. Refer to Supp. B for the results trained with 100% samples." + ], + "table_footnote": [], + "table_body": "
ModelmAP3D(L1)@ Vehicle
Overall0-30m30m-50m50m-inf
PV-RCNN [38]70.3091.9269.2142.17
Voxel-RCNN [6]75.5992.4974.0953.15
VoTR-TSD [24]74.9592.2873.3651.09
CT3D [37]76.3092.5175.0755.36
Pyramid_PV [25]76.3092.6774.9154.54
PDV [15]76.8593.1375.4954.75
VoxSeT [12]77.8292.7877.2154.41
Ours78.8292.9977.6658.02
ModelmAP3D(L2)@ Vehicle
Overall0-30m30-50m50m-inf
PV-RCNN [38]65.3691.5865.1336.46
Voxel-RCNN [6]66.5991.7467.8940.80
CT3D [37]69.0491.7668.9342.60
PDV [15]69.3092.4169.3642.16
VoxSeT [12]70.2192.0570.1043.20
Ours70.5091.7871.2845.46
", + "bbox": [ + 84, + 358, + 460, + 614 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Results on the WOD validation set in different ranges for vehicle detection.", + "bbox": [ + 76, + 626, + 470, + 655 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Datasets and Implementation Details", + "text_level": 1, + "bbox": [ + 76, + 664, + 397, + 681 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "WOD [43] is a large dataset of autonomous driving scenes. It totally contains 798 training sequences with around 160K LiDAR samples and 202 validation sequences with 40K LiDAR samples, with the mean Average Precision (mAP) and mAP weighted by heading accuracy (mAPH) as evaluation metrics. There are also two levels of difficulty describing the sparsity in each bounding box, and LEVEL_1 (L1) and LEVEL_2 (L2) denote more than 5 points and 1-5 points, respectively. For detection performance along distance, it provides mAP/mAPH on 0-30m, 30m-50m and 50m-inf.", + "bbox": [ + 75, + 688, + 468, + 838 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "KITTI [10] is a widely used benchmark for 3D object detection, which includes 3,712, 3,769 and 7,518 frames for training, validation and testing, respectively. mAP is used as the official metric with 11 recall points for the val set and", + "bbox": [ + 75, + 839, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "40 for the test set, and the IoU thresholds are set to 0.7, 0.5, and 0.5 for car, pedestrian and cyclist. We use the official setting in all experiments.", + "bbox": [ + 498, + 359, + 890, + 407 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details The total loss for optimizing the overall two-stage detection is formulated as Eq. (11), where $\\mathcal{L}_{rcnn}$ can be omitted if there is no RoI head. Refer to Supp. A.1 for more information.", + "bbox": [ + 498, + 407, + 890, + 468 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {d e t}} = \\mathcal {L} _ {\\text {r p n}} + \\mathcal {L} _ {\\text {r c n n}} + \\mathcal {L} _ {\\text {s e g}}. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 493, + 890, + 511 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Results on WOD", + "text_level": 1, + "bbox": [ + 500, + 523, + 666, + 540 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The results on the validation set are displayed in Tab. 1, and it can be seen that we achieve new state-of-the-art performance on all the three classes. In particular, for pedestrian, we outperform the baseline model PV-RCNN++ [39] by $2.77\\% / 2.56\\%$ in terms of both L1 and L2 mAP, which indicates the effectiveness of the proposed model in handling hard examples.", + "bbox": [ + 496, + 550, + 890, + 656 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In comparison with other Transformer-based models, we focus on vehicle since the counterparts only report the performance on it. As Tab. 2 shows, our OcTr achieves the best mAP among all these convolution- and Transformer-based backbones. It also outperforms the Transformer-based detection head network CT3D [37] by $2.52\\%$ and $1.46\\%$ in L1 and L2 mAP. Regarding the accuracies at different distances, OcTr ranks the first place in the range of $30\\mathrm{m} - 50\\mathrm{m}$ and $50\\mathrm{m}$ inf, which surpasses the previous best by $0.45\\%$ , $2.66\\%$ in L1 mAP and $1.18\\%$ , $2.26\\%$ in L2 mAP respectively. It clearly illustrates that OcTr has the advantage in capturing long-range fine-grained context, which facilitates dealing with objects far away. Indeed, far objects generally have much more sparse points than near ones and heavily rely on context for detection, thus benefiting more from OcTr. Refer to Fig. 5 for visualization.", + "bbox": [ + 496, + 659, + 892, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "5171", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/0f624f54e106775c42a50cd91a9d0e082bcde6aea6890eeb0d93afa061e25c71.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Model\\( \\mathbf{{mAP}}_{3D}@\\mathbf{{Car}} \\)\\( \\mathbf{{mAP}}_{3D}@\\mathbf{{Pedestrian}} \\)\\( \\mathbf{{mAP}}_{3D}@\\mathbf{{Cyclist}} \\)
EasyMod.HardEasyMod.HardEasyMod.Hard
SECOND [52]88.6178.6277.2256.5552.9847.7380.5867.1563.10
PointPillars [18]88.4677.2874.6557.7552.2947.9080.0462.6159.52
VoTR [24]87.8678.2776.93------
VoxSeT [12]88.4578.4877.0760.6254.7450.3984.0768.1165.14
Ours88.4378.5777.1661.4957.1752.3585.2970.4466.17
", + "bbox": [ + 183, + 88, + 787, + 200 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e8f19e2377cbed17b02badadd11a980b387c88b6c587f88e0d407dcd220af016.jpg", + "table_caption": [ + "Table 3. Results of the single-stage models on the KITTI val set. All the models adopt the same anchor-based region proposal network as the detection head. \"Mod.\" denotes the moderate difficulty level." + ], + "table_footnote": [], + "table_body": "
Model\\( \\mathbf{mAP}_{3D}@\\mathbf{Car on test} \\)\\( \\mathbf{mAP}_{3D}@\\mathbf{Car on val} \\)
EasyMod.HardMeanEasyMod.HardMean
SECOND [52]83.3472.5565.8273.9088.6178.6277.2281.48
PointPillars [18]82.5874.3168.9975.2986.6276.0668.9177.20
STD [54]87.9579.7175.0980.9289.7079.8079.3082.93
SA-SSD [13]88.7579.7974.1680.9090.1579.9178.7882.95
3DSSD [53]88.3679.5774.5580.8389.7179.4578.6782.61
PV-RCNN [38]90.2581.4376.8282.8389.3583.6978.7083.91
Voxel-RCNN [6]90.9081.6277.0683.1989.4184.5278.9384.29
CT3D [37]87.8381.7777.1682.2589.5486.0678.9984.86
VoTR-TSD [24]89.9082.0979.1483.7189.0484.0478.6883.92
VoxSeT [12]88.5382.0677.4682.6889.2186.7178.5684.83
Focals Conv [4]90.5582.2877.5983.4789.5284.9379.1884.54
Ours90.8882.6477.7783.7689.8086.9779.2885.35
", + "bbox": [ + 202, + 252, + 767, + 470 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3a431abd997cf8e5ddd0d4ccf1d80084e77d3fee64bbd45943cf38316cb5a4b3.jpg", + "table_caption": [ + "Table 4. Comparison to the state-of-the-art models on the KITTI test and val sets. \"Mod.\" and \"Mean\" denote the moderate difficulty level and the average mAP for the three levels, respectively. The best results are bolded and the second best ones are underlined." + ], + "table_footnote": [ + "Table 5. Results of extensions to different representative detectors on the WOD validation set." + ], + "table_body": "
DetectorVeh. mAP (L1/L2)Pedes. mAP (L1/L2)
SECOND [52]70.96/62.5865.23/57.22
Ours73.28/65.0568.08/60.36
PV-RCNN [38]75.41/67.4471.98/63.70
Ours76.77/68.3173.22/64.30
PV-RCNN++ [39]77.82/69.0777.99/69.92
Ours78.01/69.6080.75/72.45
", + "bbox": [ + 84, + 525, + 460, + 628 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Results on KITTI", + "text_level": 1, + "bbox": [ + 76, + 672, + 250, + 690 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The performance of the single-stage detectors is shown in Tab. 3. We take SECOND [52], a commonly used anchor-based model, as the baseline, and compare OcTr with another two advanced Transformer-based variants VoxSet [12] and VoTR [24]. We can see that OcTr achieves comparable results with SECOND and VoxSeT in car, while it significantly outperforms all the counterparts and reports the best performance both in pedestrian and cyclist. Benefiting from the large receptive field and fine-grained global context, it exceeds SECOND by $4.19\\%$ and $3.29\\%$ for pedestrians and cyclists respectively, where hard samples often appear.", + "bbox": [ + 75, + 700, + 468, + 867 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We summarize the performance of the two-stage models on the KITTI test set in Tab. 4. With the help of the", + "bbox": [ + 75, + 869, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/dcd1d85cd2b4203735b4e4d9a3462c7a7c8dab24efae73b7e1b53949726c2ee4.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
AttentionVeh. mAP (L1/L2)Pedes. mAP (L1/L2)
Ours (OctAttn)73.3/65.168.1/60.4
Performer [5]71.4/63.665.7/57.9
ACT [27]71.7/63.564.3/56.1
VoTr [24]69.4/61.565.0/57.0
Nearest K68.2/59.864.9/56.7
", + "bbox": [ + 506, + 525, + 885, + 619 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 6. Ablation on various attention mechanisms and sampling patterns on the WOD validation set.", + "bbox": [ + 498, + 630, + 890, + 659 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "multi-scale backbone features and rich global context, OcTr reaches a leading mAP in car at the moderate level, surpassing the state-of-the-art Focals-Conv [4] by $0.36\\%$ . We also evaluate OcTr on the KITTI val set, and OcTr again delivers the best performance in the average score, outperforming the second-best by $0.49\\%$ . One can observe that we rank the best or the second best in all the cases.", + "bbox": [ + 496, + 667, + 893, + 776 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation study", + "text_level": 1, + "bbox": [ + 500, + 786, + 650, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Scalability on various detectors As summarized in Tab. 5, we conduct experiments on three different and representative detectors, SECOND (single-stage, anchor-based), PV-RCNN (two-stage, anchor-based) and PV-RCNN++ (two-stage, anchor-free). Regardless of the number of stages or region proposal network, we acquire sound improvements", + "bbox": [ + 496, + 809, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "5172", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/201e72619ea3ded3f1161cba651e73c6c396b7cf0becd41e85393538b4488ad3.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
LEPESAPESAMVeh. mAP (L1/L2)Pedes. mAP (L1/L2)
71.35/63.3065.75/57.89
✓72.34/64.3266.56/58.62
✓✓72.64/64.4666.62/58.83
✓✓72.86/64.4067.79/59.90
✓✓✓73.28/65.0568.08/60.36
", + "bbox": [ + 84, + 88, + 460, + 170 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/1bab0f751e46f54d24c2060df348a88747ee74a6d686184af3ed38b5bb780eca.jpg", + "table_caption": [ + "Table 7. Ablation on semantic positional embedding on the WOD validation set." + ], + "table_footnote": [], + "table_body": "
topk numberVeh. mAP (L1/L2)Pedes. mAP (L1/L2)
170.38/62.2064.19/56.43
472.58/64.4266.21/58.42
873.28/65.0568.08/60.36
1673.25/65.0167.89/60.10
", + "bbox": [ + 84, + 220, + 460, + 303 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/cddb82516b0acf59e26496f81274296bc54fb3072ac36b7e71c04040619e417f.jpg", + "table_caption": [ + "Table 8. Results of different $k$ values on the WOD validation set." + ], + "table_footnote": [], + "table_body": "
Method#Param. (M)Latency (ms)Memory (GB)
SECOND [52]5.3482.3
VoTR-SSD [24]4.8673.0
VoxSeT-SSD [12]3.0373.6
OcTr-SSD2.9642.5
", + "bbox": [ + 84, + 340, + 465, + 412 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 9. Resource costs of different backbones with single-stage detectors on the KITTI dataset, test on GTX2080Ti.", + "bbox": [ + 76, + 424, + 467, + 450 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "compared to the baselines of the sparse convolution backbones, highlighting its scalability.", + "bbox": [ + 75, + 477, + 467, + 507 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation on OctAttn We carry out additional experiments to make apple-to/apple comparison with several representative linear Transformer methods, including Performer [5] (kernel-based linear attention), ACT [27] (cluster-based linear attention), VoTr [24] (fixed patterns) and the Nearest-K strategy. Tab. 6 lists the results, and our OctAttn clearly performs the best, showing its ability.", + "bbox": [ + 75, + 507, + 467, + 612 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation on semantic positional embedding We individually evaluate the contributions of LePE, SAPE and SAM with SECOND as the baseline detector on WOD in Tab. 7. By incorporating LePE, the L1/L2 performance is boosted by $0.99\\% / 1.02\\%$ and $0.81\\% / 0.73\\%$ on vehicle and pedestrian, illustrating its necessity. Furthermore, we separately verify the validity of SAPE and SAM. With semantic clues, we observe that SAPE increases by $0.3\\%$ L1 mAP on vehicle, while SAM provides an L1/L2 mAP improvement of $0.52\\% / 0.08\\%$ and $1.23\\% / 1.28\\%$ on vehicle and pedestrian. Finally, we simultaneously apply SAPE and SAM and construct the full model of OcTr, which gains $0.94\\% / 0.73\\%$ and $1.52\\% / 1.74\\%$ L1/L2 mAP on vehicle and pedestrian, showing its impact.", + "bbox": [ + 75, + 613, + 467, + 824 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Influence by topk Tab. 8 shows the performance of OcTr with various values of $k$ . As we can see, the performance improves when $k$ becomes larger but quickly saturates. Due to the redundancy of the scanned scenes, we argue that only a few tokens with high relevance need to be subdivided to", + "bbox": [ + 75, + 824, + 467, + 898 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "embrace fine-grained features.", + "bbox": [ + 498, + 90, + 702, + 104 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Analysis on model complexity We compare OcTr with two recent Transformer-based models in terms of resource cost by keeping the same detection head. Tab. 9 shows that with the learnable octree attention mechanism, OcTr consistently maintains less model parameters and less memory occupancies than the counterparts. Regarding the inference speed, VoxSeT runs faster, but it should be noted that VoxSeT inputs with pillars which discard the height dimension, leading to inferior results. As for VoTR, we deliver a mild improvement in efficiency while bringing a large gain in performance.", + "bbox": [ + 496, + 106, + 890, + 271 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/772b9d9f91728cd492ba4862a309cb6ecbe9349aec506fc8fbb76e755af67f38.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 286, + 888, + 393 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/359cc85ce0005bf04fb6f6291db3c1b6ed59be462b73ca75830975ca33e96e37.jpg", + "image_caption": [ + "Figure 5. Visualization of results by OcTr on the WOD validation split. Blue/red indicates predicted/ground-truth bounding boxes" + ], + "image_footnote": [], + "bbox": [ + 501, + 393, + 883, + 511 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 575, + 617, + 590 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper proposes a novel voxel-based approach to 3D object detection, namely OcTr. It aims to balance the fine-grained global representation and efficiency with acceptable resource costs. To this end, we propose a learned sparsification attention mechanism, OctAttn, which adaptively prunes the octants from the multi-scale feature pyramid in a top-to-bottom manner. Furthermore, we adopt a hybrid semantic-aware positional embedding based on foreground segmentation. Extensive experiments are conducted on WOD and KITTI, and OcTr reaches the state-of-the-art performance, validating its effectiveness.", + "bbox": [ + 496, + 602, + 890, + 768 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 500, + 784, + 650, + 800 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work is partly supported by the National Natural Science Foundation of China (No. 62022011 and No. 62202034), the Research Program of State Key Laboratory of Software Development Environment (SKLSDE-2021ZX-04), and the Fundamental Research Funds for the Central Universities.", + "bbox": [ + 496, + 809, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "5173", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Joshua Ainslie, Santiago Ontanon, Chris Alberti, Vaclav Cvicek, Zachary Fisher, Philip Pham, Anirudh Ravula, Sumit Sanghai, Qifan Wang, and Li Yang. Etc: Encoding long and structured inputs in transformers. In EMNLP, 2020. 2", + "[2] Iz Beltagy, Matthew E Peters, and Arman Cohan. Longformer: The long-document transformer. arXiv preprint arXiv:2004.05150, 2020. 2", + "[3] Chen Chen, Zhe Chen, Jing Zhang, and Dacheng Tao. Sasa: Semantics-augmented set abstraction for point-based 3d object detection. In AAAI, volume 1, 2022. 2", + "[4] Yukang Chen, Yanwei Li, Xiangyu Zhang, Jian Sun, and Jiaya Jia. Focal sparse convolutional networks for 3d object detection. In CVPR, pages 5428-5437, 2022. 1, 2, 7", + "[5] Krzysztof Marcin Choromanski, Valerii Likhosherstov, David Dohan, Xingyou Song, Andreea Gane, Tamás Sarlós, Peter Hawkins, Jared Quincy Davis, Afroz Mohiuddin, Lukasz Kaiser, David Benjamin Belanger, Lucy J. Colwell, and Adrian Weller. Rethinking attention with performers. In ICLR, 2021. 7, 8", + "[6] Jiajun Deng, Shaoshuai Shi, Peiwei Li, Wengang Zhou, Yanyong Zhang, and Houqiang Li. Voxel r-cnn: Towards high performance voxel-based 3d object detection. In AAAI, volume 35, pages 1201-1209, 2021. 2, 6, 7", + "[7] Xiaoyi Dong, Jianmin Bao, Dongdong Chen, Weiming Zhang, Nenghai Yu, Lu Yuan, Dong Chen, and Baining Guo. Cswin transformer: A general vision transformer backbone with cross-shaped windows. In CVPR, pages 12124-12134, 2022. 2, 5", + "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2020. 3", + "[9] Lue Fan, Ziqi Pang, Tianyuan Zhang, Yu-Xiong Wang, Hang Zhao, Feng Wang, Naiyan Wang, and Zhaoxiang Zhang. Embracing single stride 3d object detector with sparse transformer. In CVPR, pages 8458-8468, 2022. 2, 3, 6", + "[10] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In CVPR, pages 3354-3361. IEEE, 2012. 6", + "[11] Meng-Hao Guo, Jun-Xiong Cai, Zheng-Ning Liu, Tai-Jiang Mu, Ralph R Martin, and Shi-Min Hu. Pct: Point cloud transformer. CVM, 7(2):187-199, 2021. 2, 3", + "[12] Chenhang He, Ruihuang Li, Shuai Li, and Lei Zhang. Voxel set transformer: A set-to-set approach to 3d object detection from point clouds. In CVPR, pages 8417-8427, 2022. 2, 3, 6, 7, 8", + "[13] Chenhang He, Hui Zeng, Jianqiang Huang, Xian-Sheng Hua, and Lei Zhang. Structure aware single-stage 3d object detection from point cloud. In CVPR, pages 11873-11882, 2020. 2, 7", + "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 3" + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Jordan SK Hu. Point density-aware voxels for lidar 3d object detection. In CVPR, pages 8469-8478, 2022. 6", + "[16] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In ICML, pages 448-456. PMLR, 2015. 4", + "[17] Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax. In ICLR, 2016. 4", + "[18] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In CVPR, pages 12697-12705, 2019. 2, 6, 7", + "[19] Juho Lee, Yoonho Lee, Jungtaek Kim, Adam Kosiorek, Seungjin Choi, and Yee Whye Teh. Set transformer: A framework for attention-based permutation-invariant neural networks. In ICML, pages 3744-3753. PMLR, 2019. 2", + "[20] Zhichao Li, Feng Wang, and Naiyan Wang. Lidar r-cnn: An efficient and universal 3d object detector. In CVPR, pages 7546-7555, 2021. 6", + "[21] Youwei Liang, Chongjian Ge, Zhan Tong, Yibing Song, Jue Wang, and Pengtao Xie. Not all patches are what you need: Expediting vision transformers via token reorganizations. In ICLR, 2022. 2, 4", + "[22] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In ICCV, pages 2980-2988, 2017. 5", + "[23] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 10012-10022, 2021. 2, 3", + "[24] Jiageng Mao. Voxel transformer for 3d object detection. In ICCV, pages 3164-3173, 2021. 1, 2, 3, 6, 7, 8", + "[25] Jiageng Mao, Minzhe Niu, Haoyue Bai, Xiaodan Liang, Hang Xu, and Chunjing Xu. Pyramid r-cnn: Towards better performance and adaptability for 3d object detection. In ICCV, pages 2723-2732, 2021. 6", + "[26] Julien NP Martel, David B Lindell, Connor Z Lin, Eric R Chan, Marco Monteiro, and Gordon Wetzstein. Acorn: Adaptive coordinate networks for neural scene representation. In ACM SIGGRAPH, 2021. 3", + "[27] Zheng Minghang, Gao Peng, Zhang Renrui, Li Kunchang, Li Hongsheng, and Dong Hao. End-to-end object detection with adaptive clustering transformer. In BMVC, page 226, 2021. 7, 8", + "[28] Ishan Misra, Rohit Girdhar, and Armand Joulin. An end-to-end transformer model for 3d object detection. In ICCV, pages 2906-2917, 2021. 3", + "[29] Xuran Pan, Zhuofan Xia, Shiji Song, Li Erran Li, and Gao Huang. 3d object detection with pointformer. In CVPR, pages 7463-7472, 2021. 1, 3", + "[30] Chunghyun Park, Yoonwoo Jeong, Minsu Cho, and Jaesik Park. Fast point transformer. In CVPR, pages 16949-16958, 2022. 3", + "[31] Charles R Qi, Or Litany, Kaiming He, and Leonidas J Guibas. Deep hough voting for 3d object detection in point clouds. In ICCV, pages 9277-9286, 2019. 2", + "[32] Charles R Qi, Wei Liu, Chenxia Wu, Hao Su, and Leonidas J Guibas. Frustum pointnets for 3d object detection from rgb-d data. In CVPR, pages 918-927, 2018. 1, 2" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "5174", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[33] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In NeurIPS, volume 30, 2017. 1", + "[34] Jiezhong Qiu, Hao Ma, Omer Levy, Scott Wen-tau Yih, Sinong Wang, and Jie Tang. Blockwise self-attention for long document understanding. arXiv preprint arXiv:1911.02972, 2019. 2", + "[35] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In NeurIPS, volume 28, 2015. 2", + "[36] Gernot Riegler, Ali Osman Ulusoy, and Andreas Geiger. Octnet: Learning deep 3d representations at high resolutions. In CVPR, pages 3577-3586, 2017. 3", + "[37] Hualian Sheng, Sijia Cai, Yuan Liu, Bing Deng, Jianqiang Huang, Xian-Sheng Hua, and Min-Jian Zhao. Improving 3d object detection with channel-wise transformer. In ICCV, pages 2743-2752, 2021. 3, 6, 7", + "[38] Shaoshuai Shi, Chaoxu Guo, Li Jiang, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. Pv-rcnn: Point-voxel feature set abstraction for 3d object detection. In CVPR, pages 10529–10538, 2020. 2, 6, 7", + "[39] Shaoshuai Shi, Li Jiang, Jiajun Deng, Zhe Wang, Chaoxu Guo, Jianping Shi, Xiaogang Wang, and Hongsheng Li. Pvrccnn++: Point-voxel feature set abstraction with local vector representation for 3d object detection. IEEE TPAMI, 2021. 6, 7", + "[40] Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. Pointr-cnn: 3d object proposal generation and detection from point cloud. In CVPR, pages 770-779, 2019. 1, 2", + "[41] Shaoshuai Shi, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. From points to parts: 3d object detection from point cloud with part-aware and part-aggregation network. IEEE TPAMI, 43(8):2647-2664, 2020. 6", + "[42] Weijing Shi and Raj Rajkumar. Point-gnn: Graph neural network for 3d object detection in a point cloud. In CVPR, pages 1711-1719, 2020. 1, 2", + "[43] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In CVPR, pages 2446-2454, 2020. 6", + "[44] Jia-Heng Tang, Weikai Chen, Jie Yang, Bo Wang, Songrun Liu, Bo Yang, and Lin Gao. Octfield: Hierarchical implicit functions for 3d modeling. In NeurIPS, 2021. 3", + "[45] Shitao Tang, Jiahui Zhang, Siyu Zhu, and Ping Tan. Quadtree attention for vision transformers. In ICLR, 2022. 2, 4", + "[46] OpenPCDet Development Team. Openpcdet: An opensource toolbox for 3d object detection from point clouds. https://github.com/open-mmlab/OpenPCDet, 2020.2", + "[47] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Ilia Polosukhin. Attention is all you need. In NeurIPS, volume 30, 2017. 3", + "[48] Peng-Shuai Wang, Yang Liu, Yu-Xiao Guo, Chun-Yu Sun, and Xin Tong. O-cnn: Octree-based convolutional neural networks for 3d shape analysis. ACM TOG, 36(4):1-11, 2017. 3" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[49] Peng-Shuai Wang, Chun-Yu Sun, Yang Liu, and Xin Tong. Adaptive o-cnn: A patch-based deep representation of 3d shapes. ACM TOG, 37(6):1-11, 2018. 3", + "[50] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph cnn for learning on point clouds. ACM TOG, 38(5):1-12, 2019. 1", + "[51] Tete Xiao, Mannat Singh, Eric Mintun, Trevor Darrell, Piotr Dollár, and Ross Girshick. Early convolutions help transformers see better. In NeurIPS, volume 34, pages 30392-30400, 2021. 3", + "[52] Yan Yan, Yuxing Mao, and Bo Li. Second: Sparsely embedded convolutional detection. Sensors, 18(10):3337, 2018. 1, 2, 6, 7, 8", + "[53] Zetong Yang, Yanan Sun, Shu Liu, and Jiaya Jia. 3dssd: Point-based 3d single stage object detector. In CVPR, pages 11040-11048, 2020. 2, 7", + "[54] Zetong Yang, Yanan Sun, Shu Liu, Xiaoyong Shen, and Jiaya Jia. Std: Sparse-to-dense 3d object detector for point cloud. In CVPR, pages 1951-1960, 2019. 2, 7", + "[55] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3d object detection and tracking. In CVPR, pages 11784-11793, 2021. 2, 6", + "[56] Manzil Zaheer, Guru Guruganesh, Kumar Avinava Dubey, Joshua Ainslie, Chris Alberti, Santiago Ontanon, Philip Pham, Anirudh Ravula, Qifan Wang, Li Yang, et al. Big bird: Transformers for longer sequences. In NeurIPS, volume 33, pages 17283-17297, 2020. 2", + "[57] Cheng Zhang, Haocheng Wan, Xinyi Shen, and Zizhao Wu. Patchformer: An efficient point transformer with patch attention. In CVPR, pages 11799-11808, 2022. 3", + "[58] Yanan Zhang, Jiaxin Chen, and Di Huang. Cat-det: Contrastively augmented transformer for multi-modal 3d object detection. In CVPR, pages 908-917, 2022. 1", + "[59] Yifan Zhang, Qingyong Hu, Guoquan Xu, Yanxin Ma, Jianwei Wan, and Yulan Guo. Not all points are equal: Learning highly efficient point-based detectors for 3d lidar point clouds. In CVPR, pages 18953-18962, 2022. 2", + "[60] Yanan Zhang, Di Huang, and Yunhong Wang. Pc-rgnn: Point cloud completion and graph neural network for 3d object detection. In AAAI, volume 35, pages 3430-3437, 2021. 1, 2", + "[61] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In ICCV, pages 16259-16268, 2021. 1, 2, 3", + "[62] Wu Zheng, Weiliang Tang, Sijin Chen, Li Jiang, and Chi-Wing Fu. Cia-ssd: Confident iou-aware single-stage object detector from point cloud. In AAAI, volume 35, pages 3555–3562, 2021. 2", + "[63] Yin Zhou and Oncel Tuzel. Voxelnet: End-to-end learning for point cloud based 3d object detection. In CVPR, pages 4490-4499, 2018. 2" + ], + "bbox": [ + 501, + 92, + 890, + 828 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "5175", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/OcTr_ Octree-Based Transformer for 3D Object Detection/b6928c01-d853-4d0d-9874-89b3af0f167a_model.json b/2023/OcTr_ Octree-Based Transformer for 3D Object Detection/b6928c01-d853-4d0d-9874-89b3af0f167a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..1ba68ddf5ff5c75b68094cbd8062a55071454bce --- /dev/null +++ b/2023/OcTr_ Octree-Based Transformer for 3D Object Detection/b6928c01-d853-4d0d-9874-89b3af0f167a_model.json @@ -0,0 +1,2398 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.193, + 0.131, + 0.778, + 0.152 + ], + "angle": 0, + "content": "OcTr: Octree-based Transformer for 3D Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.25, + 0.18, + 0.724, + 0.199 + ], + "angle": 0, + "content": "Chao Zhou\\(^{1,2}\\), Yanan Zhang\\(^{1,2}\\), Jiaxin Chen\\(^{2}\\), Di Huang\\(^{1,2,3*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.099, + 0.199, + 0.873, + 0.217 + ], + "angle": 0, + "content": "1State Key Laboratory of Software Development Environment, Beihang University, Beijing, China" + }, + { + "type": "text", + "bbox": [ + 0.159, + 0.217, + 0.812, + 0.234 + ], + "angle": 0, + "content": "\\(^{2}\\)School of Computer Science and Engineering, Beihang University, Beijing, China" + }, + { + "type": "text", + "bbox": [ + 0.208, + 0.234, + 0.764, + 0.252 + ], + "angle": 0, + "content": "\\(^{3}\\)Hangzhou Innovation Institute, Beihang University, Hangzhou, China" + }, + { + "type": "text", + "bbox": [ + 0.221, + 0.255, + 0.742, + 0.269 + ], + "angle": 0, + "content": "{zhouchaobeing, zhangyanan, jiaxinchen, dhuang}@buaa.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.305, + 0.313, + 0.32 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.337, + 0.474, + 0.64 + ], + "angle": 0, + "content": "A key challenge for LiDAR-based 3D object detection is to capture sufficient features from large scale 3D scenes especially for distant or/and occluded objects. Albeit recent efforts made by Transformers with the long sequence modeling capability, they fail to properly balance the accuracy and efficiency, suffering from inadequate receptive fields or coarse-grained holistic correlations. In this paper, we propose an Octree-based Transformer, named OcTr, to address this issue. It first constructs a dynamic octree on the hierarchical feature pyramid through conducting self-attention on the top level and then recursively propagates to the level below restricted by the octants, which captures rich global context in a coarse-to-fine manner while maintaining the computational complexity under control. Furthermore, for enhanced foreground perception, we propose a hybrid positional embedding, composed of the semantic-aware positional embedding and attention mask, to fully exploit semantic and geometry clues. Extensive experiments are conducted on the Waymo Open Dataset and KITTI Dataset, and OcTr reaches newly state-of-the-art results." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.669, + 0.208, + 0.684 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.695, + 0.47, + 0.877 + ], + "angle": 0, + "content": "3D object detection from point clouds has received extensive attention during the past decade for its ability to provide accurate and stable recognition and localization in autonomous driving perception systems. In this task, feature learning plays a very fundamental and crucial role; yet it is rather challenging due to not only the disordered and sparse nature of data sampling, but also to insufficient acquisition under occlusion or at a distance. To address this issue, many methods have been proposed, which can be taxonomized into two major classes, i.e. grid-based and point-based. The former first regularize point clouds into multi-view images or voxels and then apply 2D or 3D CNNs to build shape rep-" + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.315, + 0.885, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.539, + 0.435, + 0.712, + 0.445 + ], + "angle": 0, + "content": "(1) Fixed Pattern Limited Receptive Field" + }, + { + "type": "image_caption", + "bbox": [ + 0.734, + 0.436, + 0.885, + 0.447 + ], + "angle": 0, + "content": "(2) Set Proxy Limited Representation" + }, + { + "type": "image", + "bbox": [ + 0.535, + 0.455, + 0.86, + 0.572 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.589, + 0.892, + 0.659 + ], + "angle": 0, + "content": "Figure 1. Illustration of three sparsification strategies of attention matrices. Fixed pattern (1) narrows receptive fields and set proxy (2) discards elaborate correlations. The proposed octree construction (3) keeps the global receptive field in a coarse-grained manner while maintaining fine-grained representations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.688, + 0.892, + 0.779 + ], + "angle": 0, + "content": "resentations [4, 52], while the latter directly conduct MLP based networks such as PointNet++ [33] and DGCNN [50] on original points for geometry description [32, 40, 42, 60]. Unfortunately, they fail to capture necessary context information through the small receptive fields in the deep models, leading to limited results." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.893, + 0.903 + ], + "angle": 0, + "content": "Witnessing the recent success of Transformers in NLP, many studies have investigated and extended such architectures for 3D vision [24, 29, 58, 61]. Transformers are reputed to model long-range dependencies, delivering global receptive fields, and to be suitable for scattered inputs of arbitrary sizes. Meanwhile, in contrast to those static weights that are learned in convolutions, Transformers dynamically aggregate the input features according to the relationships" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.888, + 0.288, + 0.9 + ], + "angle": 0, + "content": "*indicates the corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5166" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.258 + ], + "angle": 0, + "content": "between tokens. Regarding the case in 3D object detection, compared to point-based Transformers [11,61], voxel-based ones show the superiority in efficiency. However, they tend to suffer heavy computations when dealing with large scale scenes because of the quadratic complexity of Transformers, with the underlying dilemma between the grid size and the grid amount in voxelization. Taking the KITTI dataset as an example, it is unrealistic for Transformers to operate on the feature map with the spatial shape of \\(200 \\times 176 \\times 5\\), which is commonly adopted in most of the detection heads [38,46,52,55]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.261, + 0.473, + 0.548 + ], + "angle": 0, + "content": "More recently, there have appeared an influx of efficient self-attention model variants that attempt to tackle long sequences as input. They generally sparsify the attention matrix by fixed patterns [7, 23, 34], learned patterns [21, 45] or a combination of them [1, 56]. Fixed patterns chunk the input sequence into blocks of local windows or dilation windows, whilst learned patterns determine a notion of token relevance and eliminate or cluster outliers. Specific to 3D object detection from point clouds, VoTr [24] modifies self-attention with pre-defined patterns including local windows and stride dilation ones in a sparse query manner, and the dilation mechanism enlarges the receptive field by sampling attending tokens in a radius. SST [9] splits input tokens into non-overlapping patterns in a block-wise way and enables window shifting to capture cross-window correlation. Despite some improvements reported, they both only achieve bigger local receptive fields rather than the expected global ones, and computations still increase rapidly with the expansion of receptive fields." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.551, + 0.473, + 0.748 + ], + "angle": 0, + "content": "Another alternative on self-attention is to take advantage of a proxy memory bank which has the access to the entire sequence tokens [1, 2, 56]. By using a small number of induced proxies to compress the whole sequence, it diffuses the global context efficiently. VoxSet [12] adapts Set Transformer [19] to 3D object detection and exploits an induced set to model a set-to-set point cloud translation. With the help of the compressed global proxies and Conv-FFN, it obtains a global receptive field; nevertheless, as they admit, it is sub-optimal to set only a few latent codes as proxies for a large 3D scene, prone to impairing the representation of different point cloud structures and their correlations. Therefore, there remains space for a stronger solution." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.473, + 0.903 + ], + "angle": 0, + "content": "In this paper, we present a novel Transformer network, namely Octree-based Transformer (OcTr), for 3D object detection. We firstly devise an octree-based learnable sparse pattern, i.e. OctAttn, which meticulously and efficiently encodes point clouds of scenes as shown in Fig. 1. The OctAttn module constructs a feature pyramid by gathering and applies self-attention to the top level of the feature pyramid to select the most relevant tokens, which are deemed as the octants to be divided in the subsequent. When propagating to the level below, the key/value inputs are restricted by the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.243 + ], + "angle": 0, + "content": "octants from the top. Through recursively conducting this process, OctAttn captures rich global context features by a global receptive field in a coarse-to-fine manner while reducing the quadratic complexity of vanilla self-attention to the linear complexity. In addition, for better foreground perception, we propose a hybrid positional embedding, which consists of the semantic-aware positional embedding and attention mask, to fully exploit geometry and semantic clues. Thanks to the designs above, OcTr delivers a competitive trade-off between accuracy and efficiency." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.243, + 0.822, + 0.258 + ], + "angle": 0, + "content": "Our contribution is summarized in three-fold:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.262, + 0.892, + 0.308 + ], + "angle": 0, + "content": "1. We propose OcTr for voxel-based 3D object detection, which efficiently learns enhanced representations by a global receptive field with rich contexts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.316, + 0.891, + 0.361 + ], + "angle": 0, + "content": "2. We propose an octree-based learnable attention sparsification scheme (OctAttn) and a hybrid positional embedding combining geometry and semantics." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.369, + 0.892, + 0.415 + ], + "angle": 0, + "content": "3. We carry out experiments on the Waymo Open Dataset (WOD) and the KITTI dataset and report state-of-the-art performance with significant gains on far objects." + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.262, + 0.892, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.43, + 0.642, + 0.446 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.455, + 0.838, + 0.471 + ], + "angle": 0, + "content": "2.1. 3D Object Detection from Point Clouds" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.479, + 0.892, + 0.509 + ], + "angle": 0, + "content": "There exist two prevailing point-cloud representations in 3D object detection, i.e. point-based and voxel-based." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.51, + 0.895, + 0.689 + ], + "angle": 0, + "content": "The point-based methods [31,40,42,60] directly process raw point clouds in the irregular 3D space. As a pioneering attempt, F-Pointnet [32] employs instance segmentation in frustums to extract proposals. VoteNet [31] clusters objects from the surface in a deep Hough voting manner. PointRCNN [40] generates 3D RoIs with foreground segmentation and applies an RCNN-style [35] two-stage refinement. Different from PointRCNN, some existing methods [3,53,59] build a lightweight and efficient single stage 3D object detection framework. However, the current point-based methods still suffer from a large computation burden, which is not suitable for large-scale point cloud scenes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.895, + 0.902 + ], + "angle": 0, + "content": "The voxel-based ones [4,6,13,52,55,62,63] conduct voxelization on entire point clouds to construct regular grids. VoxelNet [63] exploits the voxel feature encoding layer with 3D convolutions to extract the feature of each voxel. SECOND [52] improves the model with sparse 3D convolutions, significantly increasing the speed of both training and inference. Pointpillars [18] compacts point clouds into vertical columns and encodes them with 2D CNNs. Several recent methods [13, 38, 54] also explore merging point-based and voxel-based networks into one framework for complementary features from different representations of point clouds. Unfortunately, they all use small convolution kernels with limited receptive fields, which are not competent to capture global context that is important to 3D detection." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5167" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.093, + 0.887, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.22, + 0.262, + 0.748, + 0.276 + ], + "angle": 0, + "content": "Figure 2. Framework overview of the proposed Octree-based Transformer (OcTr) model." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.302, + 0.312, + 0.318 + ], + "angle": 0, + "content": "2.2. Transformer in 3D Vision" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.326, + 0.47, + 0.476 + ], + "angle": 0, + "content": "Inspired by the great success of the self-attention mechanism in NLP [47] and CV [8, 23], Transformers have been adapted to 3D vision for their ability to capture long-range dependencies. For instance, Point Transformer [61] brings in vector attention that modulates individual feature channels for point cloud classification and segmentation; PCT [11] presents offset-attention with the implicit Laplace operator and normalization refinement which is more suitable for point cloud learning. To address the high latency, some methods [30, 57] adopt voxels or patches for acceleration." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.477, + 0.471, + 0.751 + ], + "angle": 0, + "content": "For 3D detection, 3DETR [28] treats and predicts bounding boxes as sequences in an end-to-end manner. CT3D [37] leverages a channel-wise Transformer architecture to refine the RoI head. To learn context-aware representations, some studies [9,12,24,29] introduce Transformers into a point- or voxel-based encoder. Pointformer [29] stacks local, global and local-global Transformers based on the point-based encoder; VoTr [24] exploits dilated attention with fast query to enlarge receptive fields; VoxSet [12] builds an induced point set as proxies of global context and applies point-to-point translation using voxels as mediums; and SST [9] embraces single strides without down-sampling and conducts window attention combining with its shifted version. Even though they all expand receptive fields by diverse Transformer variants, global context is not adequately involved or efficiently utilized. In contrast, we propose Octree-based Transformer (OcTr) for voxel-based 3D object detection, achieving a true global receptive field that balances accuracy and efficiency." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.762, + 0.169, + 0.777 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.787, + 0.205, + 0.802 + ], + "angle": 0, + "content": "3.1. Framework" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.471, + 0.901 + ], + "angle": 0, + "content": "This sub-section describes the overall framework of the proposed OcTr model as shown in Fig. 2. Specifically, as inspired by [51], we first voxelize the point cloud into regular grids and adopt the sparse 3D convolution for patch embedding, where the girds are regarded as the \"tokens\" and are passed through the Octree Transformer Blocks" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.303, + 0.895, + 0.499 + ], + "angle": 0, + "content": "(OTB). Compared with the vanilla Transformer block, the self-attention module is substituted by the proposed octree-attention \\( OctAttn \\), which encodes global context in a more efficient way. After applying a hybrid semantic embedding on multi-scale features, we sequentially stack two OTBs, tailed by a down-sampling layer. The voxel features are then projected into the BEV view by point-wise convolutions and are passed through a multi-scale dense 2D backbone. Ultimately, an anchor-based or anchor-free RPN head is used for 3D proposal generation, and the RoI head is optional for refinement. Note that our OcTr can be adopted for most of the voxel-based detection frameworks by simply altering 3D backbones." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.507, + 0.698, + 0.521 + ], + "angle": 0, + "content": "3.2. Self-attention Revisit" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.53, + 0.894, + 0.606 + ], + "angle": 0, + "content": "According to [47], Transformer encoder blocks typically include a multi-head self-attention (MHSA) mechanism, a feed-forward network (FFN), a normalization function, and the residual connections [14]. Given an input sequence \\(X\\), the principle part of MHSA is formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.516, + 0.612, + 0.892, + 0.653 + ], + "angle": 0, + "content": "\\[\n\\operatorname {M H S A} (X) = \\sum_ {h = 1} ^ {H} W _ {h} \\left[ \\sigma \\left(\\frac {X W _ {q} W _ {k} ^ {T} X ^ {T}}{\\sqrt {d}}\\right) \\cdot X W _ {v} \\right], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.894, + 0.765 + ], + "angle": 0, + "content": "where \\( h \\) denotes the index of the head, and \\( H \\), \\( \\sigma \\), \\( W \\) and \\( d \\) are the amount of heads, softmax function, learnable weight and feature dimension, respectively. The subscripts of \\( q \\), \\( k \\) and \\( v \\) indicate query, key and value. The inputs and outputs of the MHSA module are connected by residual connectors and normalization layers. The MLP-based FFN connects its inputs/outputs in a similar manner." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.772, + 0.667, + 0.787 + ], + "angle": 0, + "content": "3.3. Octree Attention" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.894, + 0.901 + ], + "angle": 0, + "content": "An octree is a multi-scale asymmetric and efficient representation for unstructured 3D data such as point clouds. To build an octree for the input point cloud, we recursively sub-divide it in the breadth-first order until the pre-defined octree depth is reached. Whether to sub-divide an octant is determined by the occupancy [36, 48], the surface approximation [44, 49] or a learning algorithm [26]." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5168" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.093, + 0.568, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.575, + 0.097, + 0.73, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.735, + 0.099, + 0.888, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.312, + 0.894, + 0.368 + ], + "angle": 0, + "content": "Figure 3. Octree Attention (OctAttn). The OctAttn module constructs a dynamic octree on a hierarchical feature pyramid with a partitive criterion of attention scores. For illustration, an octree representation is individually pruned from the pyramid for each grid in \\( F \\). For example, the blue grids are the divided octants for the grids with stars. The detailed structures of OcTr-Top and OcTr-Base are shown at the right side." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.38, + 0.47, + 0.529 + ], + "angle": 0, + "content": "Despite of being resource-friendly, the octree representation is discrete and non-differentiable, which motivates us to ameliorate previous cumbersome pre-processing division by introducing a novel octree-based attention mechanism. As depicted in [21, 45], the attention matrices calculated by self-attention imply the relevance of input tokens and guide feature selection. We thus prune the dense attention matrices of the multi-scale feature pyramid to sparse octree attention in an adaptive and parallel manner, namely OctAttn." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.533, + 0.47, + 0.609 + ], + "angle": 0, + "content": "As shown in Fig. 3, let the output feature map and coordinates of convolutional patch embedding be \\( F_0 \\in \\mathbb{R}^{M \\times d} \\) and \\( I_0 \\in \\mathbb{R}^{M \\times 3} \\); \\( M \\) and \\( d \\) indicate the amount of the nonempty grids in a batch and feature dimension, respectively. Based on \\( F_0 \\), we generate a multi-scale feature pyramid as" + }, + { + "type": "equation", + "bbox": [ + 0.176, + 0.629, + 0.47, + 0.647 + ], + "angle": 0, + "content": "\\[\nC = \\left\\{F _ {n}, I _ {n} \\right\\} ^ {N}, n \\in [ 0, N), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.66, + 0.124, + 0.673 + ], + "angle": 0, + "content": "where" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.692, + 0.47, + 0.722 + ], + "angle": 0, + "content": "\\[\nI _ {n} = \\left\\lfloor \\frac {I _ {0}}{2 ^ {n}} \\right\\rfloor , F _ {n} = \\operatorname {B N} \\left(S _ {\\max } \\left(F _ {0}, I _ {n}\\right)\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.733, + 0.469, + 0.779 + ], + "angle": 0, + "content": "\\(n\\), \\(N\\), BN and \\(S_{max}\\) are the index of the level of the multiscale feature pyramid, the height of the pyramid, batch normalization [16] and the max scatter function, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.47, + 0.904 + ], + "angle": 0, + "content": "The pruning begins from the top of the pyramid. The top feature map, i.e. \\(F_{N - 1}\\), is reorganized to dense input tokens \\(\\bar{F}_{N - 1} \\in \\mathbb{R}^{B \\times m_{N - 1} \\times d}\\), where \\(B\\) and \\(m_{N - 1}\\) are the batch size and maximum number of non-empty voxels per batch, respectively. The voxel is padded with 0 if it is empty. As shown in Eq. (4), the MHSA takes \\(\\bar{F}_{N - 1}\\) as input and outputs the attention score matrices \\(\\mathcal{A}_{N - 1} \\in \\mathbb{R}^{B \\times m_{N - 1} \\times m_{N - 1}}\\) and attentive features \\(\\bar{F}_{N - 1}'\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.509, + 0.404, + 0.892, + 0.503 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {A} _ {N - 1} = \\sum_ {h = 1} ^ {H} \\sigma (\\frac {\\bar {F} _ {N - 1} W _ {q} W _ {k} ^ {T} \\bar {F} _ {N - 1} ^ {T}}{\\sqrt {d}}), \\\\ \\bar {F} _ {N - 1} ^ {\\prime} = \\sum_ {h = 1} ^ {H} W _ {h} \\left[ \\sigma \\left(\\frac {\\bar {F} _ {N - 1} W _ {q} W _ {k} ^ {T} \\bar {F} _ {N - 1} ^ {T}}{\\sqrt {d}}\\right) \\cdot \\bar {F} _ {N - 1} W _ {v} \\right]. \\tag {4} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.508, + 0.892, + 0.553 + ], + "angle": 0, + "content": "For each query token, we select the top \\(k\\) attention scores as its most relevant token group in a row-wise way, denoted by \\(O_{N - 1}\\in \\mathbb{Z}^{B\\times m_{N - 1}\\times k}\\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.553, + 0.892, + 0.674 + ], + "angle": 0, + "content": "When propagating from top to bottom through the pyramid and reaching the \\(n\\)-th level, we uniformly sample limited \\(K\\) attending octants from the selected regions with features in \\(\\bar{F}_n\\) and topk indices in \\(O_{n + 1}\\). We conduct cross-attention instead of self-attention with the dense query sequence of the \\(n\\)-th level, \\(\\bar{Q}_n\\in \\mathbb{R}^{B\\times m_n\\times d}\\), and the compact sampled key/value sequence of the \\(n\\)-th level, \\(\\bar{K}_n / \\bar{V}_n\\in \\mathbb{R}^{B\\times m_n\\times K\\times d}\\), which is formulated as below" + }, + { + "type": "equation", + "bbox": [ + 0.546, + 0.699, + 0.892, + 0.784 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {A} _ {n} = \\sum_ {h = 1} ^ {H} \\sigma \\left(\\frac {\\bar {Q} _ {n} W _ {q} W _ {k} ^ {T} \\bar {K} _ {n} ^ {T}}{\\sqrt {d}}\\right), \\tag {5} \\\\ \\bar {F} _ {n} ^ {\\prime} = \\sum_ {h = 1} ^ {H} W _ {h} \\left[ \\sigma \\left(\\frac {\\bar {Q} _ {n} W _ {q} W _ {k} ^ {T} \\bar {K} _ {n} ^ {T}}{\\sqrt {d}}\\right) \\cdot \\bar {V} _ {n} W _ {v} \\right]. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.795, + 0.892, + 0.841 + ], + "angle": 0, + "content": "Backing off the sampling, this can be treated as an attention mask on self-attention matrices. The above process is run recursively until reaching the bottom level of the pyramid." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.841, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Furthermore, as the top \\(k\\) selection is a hard decision that disables gradient back-propagation, we adopt the Gumbeltop \\(k\\) technique [17] to perform a differentiable and continuous approximation by replacing the vanilla top \\(k\\) selection." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5169" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.168 + ], + "angle": 0, + "content": "The normalized scores used in top \\(k\\) are derived from the distribution in Eq. (6) during training, maintaining the original ones during inference. \\(g,\\tau ,m_{n}\\) denote the noise sampled from the Gumbel distribution, the temperature and the amount of non-empty voxels in layer \\(n\\), respectively." + }, + { + "type": "equation", + "bbox": [ + 0.139, + 0.18, + 0.469, + 0.229 + ], + "angle": 0, + "content": "\\[\np _ {i} = \\frac {\\exp \\left(\\left(\\mathcal {A} _ {n} ^ {i} + g _ {i}\\right) / \\tau\\right)}{\\sum_ {i} \\exp \\left(\\left(\\mathcal {A} _ {n} ^ {i} + g _ {i}\\right) / \\tau\\right)} \\in [ 0, 1 ]. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.236, + 0.469, + 0.31 + ], + "angle": 0, + "content": "In order to leverage the multi-scale features in distinct spatial shapes, we concatenate them by upsampling, which is implemented by inverse indices of the scatter function, followed by a linear projection layer for aligning the input feature dimension." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.312, + 0.469, + 0.402 + ], + "angle": 0, + "content": "As the local context is generally critical for object detection, inspired by [7], we additionally introduce a Locally enhanced Positional Embedding (LePE) which enables local neighbor interactions on the value sequence. With submanifold sparse convolutions, we replace the residual connections in the attention mechanism with LePE." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.404, + 0.348, + 0.418 + ], + "angle": 0, + "content": "Finally, OTB is formulated as follows" + }, + { + "type": "equation", + "bbox": [ + 0.117, + 0.443, + 0.469, + 0.47 + ], + "angle": 0, + "content": "\\[\n\\tilde {F} = \\operatorname {F C} \\left(\\left\\{F _ {N - 1} ^ {\\prime} \\mid F _ {N - 2} ^ {\\prime} \\right| \\dots \\left| F _ {0} ^ {\\prime} \\right\\}\\right) + \\operatorname {L e P E} \\left(F _ {0}\\right), \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.12, + 0.466, + 0.298, + 0.484 + ], + "angle": 0, + "content": "\\[\n\\tilde {F} ^ {\\prime} = \\mathrm {B N} (\\mathrm {F F N} (\\tilde {F})) + \\tilde {F},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.499, + 0.469, + 0.545 + ], + "angle": 0, + "content": "where \\(F_{n}^{\\prime}\\in \\mathbb{R}^{m_0\\times d}\\) denotes the compact tensor of the upsampled \\(\\bar{F}_n^\\prime\\) , | indicates concatenation, and FC denotes the fully-connected layer." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.546, + 0.469, + 0.575 + ], + "angle": 0, + "content": "Besides, we analyze the time complexity of the octree attention as below" + }, + { + "type": "equation", + "bbox": [ + 0.13, + 0.585, + 0.469, + 0.658 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {O} \\left(\\left(\\frac {M}{\\omega^ {N - 1}}\\right) ^ {2} + \\sum_ {n = 0} ^ {N - 2} \\frac {K M}{\\omega^ {n}}\\right) \\tag {8} \\\\ = \\mathcal {O} \\left(\\left(\\frac {M}{\\omega^ {N - 1}}\\right) ^ {2} + \\frac {\\omega}{\\omega - 1} K M \\left(1 - \\omega^ {1 - N}\\right)\\right), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.67, + 0.469, + 0.7 + ], + "angle": 0, + "content": "where \\(\\omega\\) is the average down-sampling ratio in the sparse voxel representation." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.711, + 0.358, + 0.727 + ], + "angle": 0, + "content": "3.4. Semantic Positional Embedding" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.469, + 0.825 + ], + "angle": 0, + "content": "Due to the large proportion of background grids in point clouds, the attention matrices are dominated by background grid pairs, leading to a sub-optimal solution. To fully leverage the local 3D shape patterns and original voxel coordinates, we propose a hybrid positional embedding to capture both the geometry and semantic clues as displayed in Fig. 4." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Specifically, we first segment foreground grids using the supervision from the ground truth. Segmentation scores are predicted by a sub-manifold sparse convolution branch with a sigmoid function; and the focal loss [22] is applied to balance the foreground and background." + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.09, + 0.887, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.301, + 0.892, + 0.357 + ], + "angle": 0, + "content": "Figure 4. Illustration of semantic positional embedding, where an extra foreground segmentation branch is adopted with supervision and semantic scores are concatenated as absolute positional embeddings and serve as relative masks on attention matrices." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.365, + 0.892, + 0.411 + ], + "angle": 0, + "content": "We concatenate the center coordinate \\((x, y, z)\\) and the semantic score with the feature \\(f\\) in a grid-wise manner, followed by a linear projection without the bias as below" + }, + { + "type": "equation", + "bbox": [ + 0.518, + 0.42, + 0.892, + 0.456 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\operatorname {S A P E} (X) = \\operatorname {F C} _ {d + 4 \\rightarrow d} (\\{x, y, z, \\text {s c o r e} | f \\}) \\tag {9} \\\\ = \\operatorname {F C} _ {d \\rightarrow d} (f) + \\operatorname {F C} _ {4 \\rightarrow d} (x, y, z, s c o r e). \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.466, + 0.892, + 0.557 + ], + "angle": 0, + "content": "Eq. (9) is equivalent to the absolute positional embedding (APE), thus being denoted as the Semantic APE (SAPE). The scatter function such as the mean, max and batch normalization naturally provide position and semantic information of the downsampled voxel grids, making it applicable for the multi-scale feature pyramid." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.557, + 0.892, + 0.676 + ], + "angle": 0, + "content": "Besides the semantic clues implicitly used in SAPE, we employ the Semantic Attention Mask (SAM) based on the segmentation scores and we mask the attention matrices to address correlations between inferior queries and superior keys in a simple yet effective way. Given a scalar attention matrix \\(\\mathcal{A} \\in \\mathbb{R}^{N_q \\times N_k}\\) before softmax and the segmentation scores of query and key/value \\(S_q \\in \\mathbb{R}^{N_q}\\) and \\(S_k \\in \\mathbb{R}^{N_k}\\), we formulate the attention matrix after softmax as" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.687, + 0.891, + 0.706 + ], + "angle": 0, + "content": "\\[\n\\mathcal {A} ^ {\\prime} = \\sigma (- \\Gamma \\cdot [ 1 - (S _ {q} \\geq \\delta_ {q}) (S _ {k} \\geq \\delta_ {k}) ^ {T} ] + \\mathcal {A}), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.711, + 0.891, + 0.756 + ], + "angle": 0, + "content": "where \\(N_{q},\\delta_{q},N_{k},\\delta_{k},\\sigma\\) and \\(\\Gamma\\) are the length and threshold of the query sequence, length and threshold of the key sequence, softmax function and infinite scalar, respectively." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.757, + 0.889, + 0.771 + ], + "angle": 0, + "content": "Finally, we broadcast the semantic mask \\(\\mathcal{A}'\\) to all heads." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.785, + 0.632, + 0.802 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.81, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We evaluate the proposed OcTr network on the Waymo Open Dataset (WOD) and KITTI dataset, both of which are popular in 3D object detection. In this section, we introduce the benchmarks and implementation details, make comparison to the previous state-of-the-art counterparts, and ablate the key designs of OcTr." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5170" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.086, + 0.09, + 0.887, + 0.297 + ], + "angle": 0, + "content": "
ModelVehicle (L1) \nmAP/mAPHVehicle (L2) \nmAP/mAPHPedes. (L1) \nmAP/mAPHPedes. (L2) \nmAP/mAPHCyclist (L1) \nmAP/mAPHCyclist (L2) \nmAP/mAPH
SECOND [52]70.96/70.3462.58/62.0265.23/54.2457.22/47.4957.13/55.6254.97/53.53
PointPillar [18]70.43/69.8362.18/61.6466.21/46.3258.18/40.6455.26/51.7553.18/49.80
PartA²Net [41]74.82/74.3265.88/65.4271.76/63.6462.53/55.3067.35/66.1565.05/63.89
PVRCNN [38]75.41/74.7467.44/66.8071.98/61.2463.70/53.9565.88/64.2563.39/61.82
CenterPoint [55]71.33/70.7663.16/62.6572.09/65.4964.27/58.2368.68/67.3966.11/64.87
LiDAR-RCNN [20]73.5/73.064.7/64.271.2/58.763.1/51.768.6/66.966.1/64.4
Voxel-RCNN [6]75.59/-66.59/--/--/--/--/-
PVRCNN++ [39]77.82/77.3269.07/68.6277.99/71.3669.92/63.7471.80/70.7169.31/68.26
SST† [9]76.22/75.7968.04/67.6481.39/74.0572.82/65.93-/--/-
PDV [15]76.85/76.3369.30/68.8174.19/65.9665.85/58.2868.71/67.5566.49/65.36
Ours78.12/77.6369.79/69.3480.76/74.3972.48/66.5272.58/71.5069.93/68.90
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.307, + 0.895, + 0.352 + ], + "angle": 0, + "content": "Table 1. Performance on WOD with 202 validation sequences for vehicle (IoU=0.7), pedestrian (IoU=0.5) and cyclist (IoU=0.5), using 20% samples for training. All the results are achieved by the models simultaneously trained for 3 classes on single frames, except the ones of the model marked by \\(\\dagger\\), which is only trained for a single class. Refer to Supp. B for the results trained with 100% samples." + }, + { + "type": "table", + "bbox": [ + 0.086, + 0.359, + 0.462, + 0.616 + ], + "angle": 0, + "content": "
ModelmAP3D(L1)@ Vehicle
Overall0-30m30m-50m50m-inf
PV-RCNN [38]70.3091.9269.2142.17
Voxel-RCNN [6]75.5992.4974.0953.15
VoTR-TSD [24]74.9592.2873.3651.09
CT3D [37]76.3092.5175.0755.36
Pyramid_PV [25]76.3092.6774.9154.54
PDV [15]76.8593.1375.4954.75
VoxSeT [12]77.8292.7877.2154.41
Ours78.8292.9977.6658.02
ModelmAP3D(L2)@ Vehicle
Overall0-30m30-50m50m-inf
PV-RCNN [38]65.3691.5865.1336.46
Voxel-RCNN [6]66.5991.7467.8940.80
CT3D [37]69.0491.7668.9342.60
PDV [15]69.3092.4169.3642.16
VoxSeT [12]70.2192.0570.1043.20
Ours70.5091.7871.2845.46
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.627, + 0.472, + 0.656 + ], + "angle": 0, + "content": "Table 2. Results on the WOD validation set in different ranges for vehicle detection." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.665, + 0.398, + 0.682 + ], + "angle": 0, + "content": "4.1. Datasets and Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.47, + 0.839 + ], + "angle": 0, + "content": "WOD [43] is a large dataset of autonomous driving scenes. It totally contains 798 training sequences with around 160K LiDAR samples and 202 validation sequences with 40K LiDAR samples, with the mean Average Precision (mAP) and mAP weighted by heading accuracy (mAPH) as evaluation metrics. There are also two levels of difficulty describing the sparsity in each bounding box, and LEVEL_1 (L1) and LEVEL_2 (L2) denote more than 5 points and 1-5 points, respectively. For detection performance along distance, it provides mAP/mAPH on 0-30m, 30m-50m and 50m-inf." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.902 + ], + "angle": 0, + "content": "KITTI [10] is a widely used benchmark for 3D object detection, which includes 3,712, 3,769 and 7,518 frames for training, validation and testing, respectively. mAP is used as the official metric with 11 recall points for the val set and" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.361, + 0.892, + 0.408 + ], + "angle": 0, + "content": "40 for the test set, and the IoU thresholds are set to 0.7, 0.5, and 0.5 for car, pedestrian and cyclist. We use the official setting in all experiments." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.409, + 0.892, + 0.469 + ], + "angle": 0, + "content": "Implementation Details The total loss for optimizing the overall two-stage detection is formulated as Eq. (11), where \\(\\mathcal{L}_{rcnn}\\) can be omitted if there is no RoI head. Refer to Supp. A.1 for more information." + }, + { + "type": "equation", + "bbox": [ + 0.593, + 0.494, + 0.892, + 0.512 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {d e t}} = \\mathcal {L} _ {\\text {r p n}} + \\mathcal {L} _ {\\text {r c n n}} + \\mathcal {L} _ {\\text {s e g}}. \\tag {11}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.525, + 0.667, + 0.541 + ], + "angle": 0, + "content": "4.2. Results on WOD" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.551, + 0.892, + 0.657 + ], + "angle": 0, + "content": "The results on the validation set are displayed in Tab. 1, and it can be seen that we achieve new state-of-the-art performance on all the three classes. In particular, for pedestrian, we outperform the baseline model PV-RCNN++ [39] by \\(2.77\\% / 2.56\\%\\) in terms of both L1 and L2 mAP, which indicates the effectiveness of the proposed model in handling hard examples." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.893, + 0.901 + ], + "angle": 0, + "content": "In comparison with other Transformer-based models, we focus on vehicle since the counterparts only report the performance on it. As Tab. 2 shows, our OcTr achieves the best mAP among all these convolution- and Transformer-based backbones. It also outperforms the Transformer-based detection head network CT3D [37] by \\(2.52\\%\\) and \\(1.46\\%\\) in L1 and L2 mAP. Regarding the accuracies at different distances, OcTr ranks the first place in the range of \\(30\\mathrm{m} - 50\\mathrm{m}\\) and \\(50\\mathrm{m}\\) inf, which surpasses the previous best by \\(0.45\\%\\), \\(2.66\\%\\) in L1 mAP and \\(1.18\\%\\), \\(2.26\\%\\) in L2 mAP respectively. It clearly illustrates that OcTr has the advantage in capturing long-range fine-grained context, which facilitates dealing with objects far away. Indeed, far objects generally have much more sparse points than near ones and heavily rely on context for detection, thus benefiting more from OcTr. Refer to Fig. 5 for visualization." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "5171" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.184, + 0.089, + 0.788, + 0.201 + ], + "angle": 0, + "content": "
Model\\( \\mathbf{{mAP}}_{3D}@\\mathbf{{Car}} \\)\\( \\mathbf{{mAP}}_{3D}@\\mathbf{{Pedestrian}} \\)\\( \\mathbf{{mAP}}_{3D}@\\mathbf{{Cyclist}} \\)
EasyMod.HardEasyMod.HardEasyMod.Hard
SECOND [52]88.6178.6277.2256.5552.9847.7380.5867.1563.10
PointPillars [18]88.4677.2874.6557.7552.2947.9080.0462.6159.52
VoTR [24]87.8678.2776.93------
VoxSeT [12]88.4578.4877.0760.6254.7450.3984.0768.1165.14
Ours88.4378.5777.1661.4957.1752.3585.2970.4466.17
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.21, + 0.895, + 0.24 + ], + "angle": 0, + "content": "Table 3. Results of the single-stage models on the KITTI val set. All the models adopt the same anchor-based region proposal network as the detection head. \"Mod.\" denotes the moderate difficulty level." + }, + { + "type": "table", + "bbox": [ + 0.204, + 0.253, + 0.769, + 0.471 + ], + "angle": 0, + "content": "
Model\\( \\mathbf{mAP}_{3D}@\\mathbf{Car on test} \\)\\( \\mathbf{mAP}_{3D}@\\mathbf{Car on val} \\)
EasyMod.HardMeanEasyMod.HardMean
SECOND [52]83.3472.5565.8273.9088.6178.6277.2281.48
PointPillars [18]82.5874.3168.9975.2986.6276.0668.9177.20
STD [54]87.9579.7175.0980.9289.7079.8079.3082.93
SA-SSD [13]88.7579.7974.1680.9090.1579.9178.7882.95
3DSSD [53]88.3679.5774.5580.8389.7179.4578.6782.61
PV-RCNN [38]90.2581.4376.8282.8389.3583.6978.7083.91
Voxel-RCNN [6]90.9081.6277.0683.1989.4184.5278.9384.29
CT3D [37]87.8381.7777.1682.2589.5486.0678.9984.86
VoTR-TSD [24]89.9082.0979.1483.7189.0484.0478.6883.92
VoxSeT [12]88.5382.0677.4682.6889.2186.7178.5684.83
Focals Conv [4]90.5582.2877.5983.4789.5284.9379.1884.54
Ours90.8882.6477.7783.7689.8086.9779.2885.35
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.481, + 0.893, + 0.511 + ], + "angle": 0, + "content": "Table 4. Comparison to the state-of-the-art models on the KITTI test and val sets. \"Mod.\" and \"Mean\" denote the moderate difficulty level and the average mAP for the three levels, respectively. The best results are bolded and the second best ones are underlined." + }, + { + "type": "table", + "bbox": [ + 0.085, + 0.526, + 0.462, + 0.629 + ], + "angle": 0, + "content": "
DetectorVeh. mAP (L1/L2)Pedes. mAP (L1/L2)
SECOND [52]70.96/62.5865.23/57.22
Ours73.28/65.0568.08/60.36
PV-RCNN [38]75.41/67.4471.98/63.70
Ours76.77/68.3173.22/64.30
PV-RCNN++ [39]77.82/69.0777.99/69.92
Ours78.01/69.6080.75/72.45
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.077, + 0.631, + 0.47, + 0.659 + ], + "angle": 0, + "content": "Table 5. Results of extensions to different representative detectors on the WOD validation set." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.674, + 0.251, + 0.691 + ], + "angle": 0, + "content": "4.3. Results on KITTI" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.702, + 0.47, + 0.868 + ], + "angle": 0, + "content": "The performance of the single-stage detectors is shown in Tab. 3. We take SECOND [52], a commonly used anchor-based model, as the baseline, and compare OcTr with another two advanced Transformer-based variants VoxSet [12] and VoTR [24]. We can see that OcTr achieves comparable results with SECOND and VoxSeT in car, while it significantly outperforms all the counterparts and reports the best performance both in pedestrian and cyclist. Benefiting from the large receptive field and fine-grained global context, it exceeds SECOND by \\(4.19\\%\\) and \\(3.29\\%\\) for pedestrians and cyclists respectively, where hard samples often appear." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.471, + 0.903 + ], + "angle": 0, + "content": "We summarize the performance of the two-stage models on the KITTI test set in Tab. 4. With the help of the" + }, + { + "type": "table", + "bbox": [ + 0.508, + 0.526, + 0.887, + 0.621 + ], + "angle": 0, + "content": "
AttentionVeh. mAP (L1/L2)Pedes. mAP (L1/L2)
Ours (OctAttn)73.3/65.168.1/60.4
Performer [5]71.4/63.665.7/57.9
ACT [27]71.7/63.564.3/56.1
VoTr [24]69.4/61.565.0/57.0
Nearest K68.2/59.864.9/56.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.631, + 0.892, + 0.66 + ], + "angle": 0, + "content": "Table 6. Ablation on various attention mechanisms and sampling patterns on the WOD validation set." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.669, + 0.894, + 0.777 + ], + "angle": 0, + "content": "multi-scale backbone features and rich global context, OcTr reaches a leading mAP in car at the moderate level, surpassing the state-of-the-art Focals-Conv [4] by \\(0.36\\%\\). We also evaluate OcTr on the KITTI val set, and OcTr again delivers the best performance in the average score, outperforming the second-best by \\(0.49\\%\\). One can observe that we rank the best or the second best in all the cases." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.787, + 0.651, + 0.803 + ], + "angle": 0, + "content": "4.4. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Scalability on various detectors As summarized in Tab. 5, we conduct experiments on three different and representative detectors, SECOND (single-stage, anchor-based), PV-RCNN (two-stage, anchor-based) and PV-RCNN++ (two-stage, anchor-free). Regardless of the number of stages or region proposal network, we acquire sound improvements" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "5172" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.085, + 0.089, + 0.462, + 0.171 + ], + "angle": 0, + "content": "
LEPESAPESAMVeh. mAP (L1/L2)Pedes. mAP (L1/L2)
71.35/63.3065.75/57.89
✓72.34/64.3266.56/58.62
✓✓72.64/64.4666.62/58.83
✓✓72.86/64.4067.79/59.90
✓✓✓73.28/65.0568.08/60.36
" + }, + { + "type": "table_caption", + "bbox": [ + 0.078, + 0.183, + 0.468, + 0.21 + ], + "angle": 0, + "content": "Table 7. Ablation on semantic positional embedding on the WOD validation set." + }, + { + "type": "table", + "bbox": [ + 0.085, + 0.222, + 0.462, + 0.304 + ], + "angle": 0, + "content": "
topk numberVeh. mAP (L1/L2)Pedes. mAP (L1/L2)
170.38/62.2064.19/56.43
472.58/64.4266.21/58.42
873.28/65.0568.08/60.36
1673.25/65.0167.89/60.10
" + }, + { + "type": "table_caption", + "bbox": [ + 0.08, + 0.315, + 0.467, + 0.328 + ], + "angle": 0, + "content": "Table 8. Results of different \\( k \\) values on the WOD validation set." + }, + { + "type": "table", + "bbox": [ + 0.085, + 0.342, + 0.466, + 0.414 + ], + "angle": 0, + "content": "
Method#Param. (M)Latency (ms)Memory (GB)
SECOND [52]5.3482.3
VoTR-SSD [24]4.8673.0
VoxSeT-SSD [12]3.0373.6
OcTr-SSD2.9642.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.425, + 0.468, + 0.452 + ], + "angle": 0, + "content": "Table 9. Resource costs of different backbones with single-stage detectors on the KITTI dataset, test on GTX2080Ti." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.478, + 0.468, + 0.508 + ], + "angle": 0, + "content": "compared to the baselines of the sparse convolution backbones, highlighting its scalability." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.508, + 0.468, + 0.613 + ], + "angle": 0, + "content": "Ablation on OctAttn We carry out additional experiments to make apple-to/apple comparison with several representative linear Transformer methods, including Performer [5] (kernel-based linear attention), ACT [27] (cluster-based linear attention), VoTr [24] (fixed patterns) and the Nearest-K strategy. Tab. 6 lists the results, and our OctAttn clearly performs the best, showing its ability." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.468, + 0.825 + ], + "angle": 0, + "content": "Ablation on semantic positional embedding We individually evaluate the contributions of LePE, SAPE and SAM with SECOND as the baseline detector on WOD in Tab. 7. By incorporating LePE, the L1/L2 performance is boosted by \\(0.99\\% / 1.02\\%\\) and \\(0.81\\% / 0.73\\%\\) on vehicle and pedestrian, illustrating its necessity. Furthermore, we separately verify the validity of SAPE and SAM. With semantic clues, we observe that SAPE increases by \\(0.3\\%\\) L1 mAP on vehicle, while SAM provides an L1/L2 mAP improvement of \\(0.52\\% / 0.08\\%\\) and \\(1.23\\% / 1.28\\%\\) on vehicle and pedestrian. Finally, we simultaneously apply SAPE and SAM and construct the full model of OcTr, which gains \\(0.94\\% / 0.73\\%\\) and \\(1.52\\% / 1.74\\%\\) L1/L2 mAP on vehicle and pedestrian, showing its impact." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.468, + 0.9 + ], + "angle": 0, + "content": "Influence by topk Tab. 8 shows the performance of OcTr with various values of \\( k \\). As we can see, the performance improves when \\( k \\) becomes larger but quickly saturates. Due to the redundancy of the scanned scenes, we argue that only a few tokens with high relevance need to be subdivided to" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.704, + 0.106 + ], + "angle": 0, + "content": "embrace fine-grained features." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.107, + 0.892, + 0.272 + ], + "angle": 0, + "content": "Analysis on model complexity We compare OcTr with two recent Transformer-based models in terms of resource cost by keeping the same detection head. Tab. 9 shows that with the learnable octree attention mechanism, OcTr consistently maintains less model parameters and less memory occupancies than the counterparts. Regarding the inference speed, VoxSeT runs faster, but it should be noted that VoxSeT inputs with pillars which discard the height dimension, leading to inferior results. As for VoTR, we deliver a mild improvement in efficiency while bringing a large gain in performance." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.287, + 0.89, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.395, + 0.885, + 0.512 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.523, + 0.892, + 0.553 + ], + "angle": 0, + "content": "Figure 5. Visualization of results by OcTr on the WOD validation split. Blue/red indicates predicted/ground-truth bounding boxes" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.577, + 0.618, + 0.592 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.603, + 0.892, + 0.769 + ], + "angle": 0, + "content": "This paper proposes a novel voxel-based approach to 3D object detection, namely OcTr. It aims to balance the fine-grained global representation and efficiency with acceptable resource costs. To this end, we propose a learned sparsification attention mechanism, OctAttn, which adaptively prunes the octants from the multi-scale feature pyramid in a top-to-bottom manner. Furthermore, we adopt a hybrid semantic-aware positional embedding based on foreground segmentation. Extensive experiments are conducted on WOD and KITTI, and OcTr reaches the state-of-the-art performance, validating its effectiveness." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.785, + 0.651, + 0.801 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.9 + ], + "angle": 0, + "content": "This work is partly supported by the National Natural Science Foundation of China (No. 62022011 and No. 62202034), the Research Program of State Key Laboratory of Software Development Environment (SKLSDE-2021ZX-04), and the Fundamental Research Funds for the Central Universities." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5173" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Joshua Ainslie, Santiago Ontanon, Chris Alberti, Vaclav Cvicek, Zachary Fisher, Philip Pham, Anirudh Ravula, Sumit Sanghai, Qifan Wang, and Li Yang. Etc: Encoding long and structured inputs in transformers. In EMNLP, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.187, + 0.472, + 0.227 + ], + "angle": 0, + "content": "[2] Iz Beltagy, Matthew E Peters, and Arman Cohan. Longformer: The long-document transformer. arXiv preprint arXiv:2004.05150, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.23, + 0.47, + 0.272 + ], + "angle": 0, + "content": "[3] Chen Chen, Zhe Chen, Jing Zhang, and Dacheng Tao. Sasa: Semantics-augmented set abstraction for point-based 3d object detection. In AAAI, volume 1, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.274, + 0.469, + 0.316 + ], + "angle": 0, + "content": "[4] Yukang Chen, Yanwei Li, Xiangyu Zhang, Jian Sun, and Jiaya Jia. Focal sparse convolutional networks for 3d object detection. In CVPR, pages 5428-5437, 2022. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.317, + 0.469, + 0.399 + ], + "angle": 0, + "content": "[5] Krzysztof Marcin Choromanski, Valerii Likhosherstov, David Dohan, Xingyou Song, Andreea Gane, Tamás Sarlós, Peter Hawkins, Jared Quincy Davis, Afroz Mohiuddin, Lukasz Kaiser, David Benjamin Belanger, Lucy J. Colwell, and Adrian Weller. Rethinking attention with performers. In ICLR, 2021. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.402, + 0.469, + 0.457 + ], + "angle": 0, + "content": "[6] Jiajun Deng, Shaoshuai Shi, Peiwei Li, Wengang Zhou, Yanyong Zhang, and Houqiang Li. Voxel r-cnn: Towards high performance voxel-based 3d object detection. In AAAI, volume 35, pages 1201-1209, 2021. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.459, + 0.469, + 0.527 + ], + "angle": 0, + "content": "[7] Xiaoyi Dong, Jianmin Bao, Dongdong Chen, Weiming Zhang, Nenghai Yu, Lu Yuan, Dong Chen, and Baining Guo. Cswin transformer: A general vision transformer backbone with cross-shaped windows. In CVPR, pages 12124-12134, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.53, + 0.469, + 0.599 + ], + "angle": 0, + "content": "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.601, + 0.469, + 0.657 + ], + "angle": 0, + "content": "[9] Lue Fan, Ziqi Pang, Tianyuan Zhang, Yu-Xiong Wang, Hang Zhao, Feng Wang, Naiyan Wang, and Zhaoxiang Zhang. Embracing single stride 3d object detector with sparse transformer. In CVPR, pages 8458-8468, 2022. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.659, + 0.469, + 0.699 + ], + "angle": 0, + "content": "[10] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In CVPR, pages 3354-3361. IEEE, 2012. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.702, + 0.469, + 0.742 + ], + "angle": 0, + "content": "[11] Meng-Hao Guo, Jun-Xiong Cai, Zheng-Ning Liu, Tai-Jiang Mu, Ralph R Martin, and Shi-Min Hu. Pct: Point cloud transformer. CVM, 7(2):187-199, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.745, + 0.469, + 0.8 + ], + "angle": 0, + "content": "[12] Chenhang He, Ruihuang Li, Shuai Li, and Lei Zhang. Voxel set transformer: A set-to-set approach to 3d object detection from point clouds. In CVPR, pages 8417-8427, 2022. 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.469, + 0.857 + ], + "angle": 0, + "content": "[13] Chenhang He, Hui Zeng, Jianqiang Huang, Xian-Sheng Hua, and Lei Zhang. Structure aware single-stage 3d object detection from point cloud. In CVPR, pages 11873-11882, 2020. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "[15] Jordan SK Hu. Point density-aware voxels for lidar 3d object detection. In CVPR, pages 8469-8478, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.121, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[16] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In ICML, pages 448-456. PMLR, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[17] Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax. In ICLR, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.191, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[18] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In CVPR, pages 12697-12705, 2019. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.246, + 0.892, + 0.302 + ], + "angle": 0, + "content": "[19] Juho Lee, Yoonho Lee, Jungtaek Kim, Adam Kosiorek, Seungjin Choi, and Yee Whye Teh. Set transformer: A framework for attention-based permutation-invariant neural networks. In ICML, pages 3744-3753. PMLR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.303, + 0.892, + 0.343 + ], + "angle": 0, + "content": "[20] Zhichao Li, Feng Wang, and Naiyan Wang. Lidar r-cnn: An efficient and universal 3d object detector. In CVPR, pages 7546-7555, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.344, + 0.892, + 0.398 + ], + "angle": 0, + "content": "[21] Youwei Liang, Chongjian Ge, Zhan Tong, Yibing Song, Jue Wang, and Pengtao Xie. Not all patches are what you need: Expediting vision transformers via token reorganizations. In ICLR, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.4, + 0.892, + 0.44 + ], + "angle": 0, + "content": "[22] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In ICCV, pages 2980-2988, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.442, + 0.892, + 0.496 + ], + "angle": 0, + "content": "[23] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 10012-10022, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.498, + 0.892, + 0.525 + ], + "angle": 0, + "content": "[24] Jiageng Mao. Voxel transformer for 3d object detection. In ICCV, pages 3164-3173, 2021. 1, 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.526, + 0.892, + 0.58 + ], + "angle": 0, + "content": "[25] Jiageng Mao, Minzhe Niu, Haoyue Bai, Xiaodan Liang, Hang Xu, and Chunjing Xu. Pyramid r-cnn: Towards better performance and adaptability for 3d object detection. In ICCV, pages 2723-2732, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.581, + 0.892, + 0.635 + ], + "angle": 0, + "content": "[26] Julien NP Martel, David B Lindell, Connor Z Lin, Eric R Chan, Marco Monteiro, and Gordon Wetzstein. Acorn: Adaptive coordinate networks for neural scene representation. In ACM SIGGRAPH, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.637, + 0.892, + 0.691 + ], + "angle": 0, + "content": "[27] Zheng Minghang, Gao Peng, Zhang Renrui, Li Kunchang, Li Hongsheng, and Dong Hao. End-to-end object detection with adaptive clustering transformer. In BMVC, page 226, 2021. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.692, + 0.892, + 0.734 + ], + "angle": 0, + "content": "[28] Ishan Misra, Rohit Girdhar, and Armand Joulin. An end-to-end transformer model for 3d object detection. In ICCV, pages 2906-2917, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.735, + 0.892, + 0.775 + ], + "angle": 0, + "content": "[29] Xuran Pan, Zhuofan Xia, Shiji Song, Li Erran Li, and Gao Huang. 3d object detection with pointformer. In CVPR, pages 7463-7472, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.776, + 0.892, + 0.816 + ], + "angle": 0, + "content": "[30] Chunghyun Park, Yoonwoo Jeong, Minsu Cho, and Jaesik Park. Fast point transformer. In CVPR, pages 16949-16958, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.818, + 0.892, + 0.859 + ], + "angle": 0, + "content": "[31] Charles R Qi, Or Litany, Kaiming He, and Leonidas J Guibas. Deep hough voting for 3d object detection in point clouds. In ICCV, pages 9277-9286, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.86, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[32] Charles R Qi, Wei Liu, Chenxia Wu, Hao Su, and Leonidas J Guibas. Frustum pointnets for 3d object detection from rgb-d data. In CVPR, pages 918-927, 2018. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5174" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[33] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In NeurIPS, volume 30, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.134, + 0.47, + 0.189 + ], + "angle": 0, + "content": "[34] Jiezhong Qiu, Hao Ma, Omer Levy, Scott Wen-tau Yih, Sinong Wang, and Jie Tang. Blockwise self-attention for long document understanding. arXiv preprint arXiv:1911.02972, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.19, + 0.469, + 0.232 + ], + "angle": 0, + "content": "[35] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In NeurIPS, volume 28, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.233, + 0.469, + 0.273 + ], + "angle": 0, + "content": "[36] Gernot Riegler, Ali Osman Ulusoy, and Andreas Geiger. Octnet: Learning deep 3d representations at high resolutions. In CVPR, pages 3577-3586, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.274, + 0.469, + 0.328 + ], + "angle": 0, + "content": "[37] Hualian Sheng, Sijia Cai, Yuan Liu, Bing Deng, Jianqiang Huang, Xian-Sheng Hua, and Min-Jian Zhao. Improving 3d object detection with channel-wise transformer. In ICCV, pages 2743-2752, 2021. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.33, + 0.469, + 0.385 + ], + "angle": 0, + "content": "[38] Shaoshuai Shi, Chaoxu Guo, Li Jiang, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. Pv-rcnn: Point-voxel feature set abstraction for 3d object detection. In CVPR, pages 10529–10538, 2020. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.386, + 0.469, + 0.453 + ], + "angle": 0, + "content": "[39] Shaoshuai Shi, Li Jiang, Jiajun Deng, Zhe Wang, Chaoxu Guo, Jianping Shi, Xiaogang Wang, and Hongsheng Li. Pvrccnn++: Point-voxel feature set abstraction with local vector representation for 3d object detection. IEEE TPAMI, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.455, + 0.469, + 0.496 + ], + "angle": 0, + "content": "[40] Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. Pointr-cnn: 3d object proposal generation and detection from point cloud. In CVPR, pages 770-779, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.497, + 0.469, + 0.551 + ], + "angle": 0, + "content": "[41] Shaoshuai Shi, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. From points to parts: 3d object detection from point cloud with part-aware and part-aggregation network. IEEE TPAMI, 43(8):2647-2664, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.553, + 0.469, + 0.593 + ], + "angle": 0, + "content": "[42] Weijing Shi and Raj Rajkumar. Point-gnn: Graph neural network for 3d object detection in a point cloud. In CVPR, pages 1711-1719, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.595, + 0.469, + 0.663 + ], + "angle": 0, + "content": "[43] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In CVPR, pages 2446-2454, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.664, + 0.469, + 0.705 + ], + "angle": 0, + "content": "[44] Jia-Heng Tang, Weikai Chen, Jie Yang, Bo Wang, Songrun Liu, Bo Yang, and Lin Gao. Octfield: Hierarchical implicit functions for 3d modeling. In NeurIPS, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.706, + 0.469, + 0.733 + ], + "angle": 0, + "content": "[45] Shitao Tang, Jiahui Zhang, Siyu Zhu, and Ping Tan. Quadtree attention for vision transformers. In ICLR, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.735, + 0.469, + 0.787 + ], + "angle": 0, + "content": "[46] OpenPCDet Development Team. Openpcdet: An opensource toolbox for 3d object detection from point clouds. https://github.com/open-mmlab/OpenPCDet, 2020.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.469, + 0.843 + ], + "angle": 0, + "content": "[47] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Ilia Polosukhin. Attention is all you need. In NeurIPS, volume 30, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.469, + 0.899 + ], + "angle": 0, + "content": "[48] Peng-Shuai Wang, Yang Liu, Yu-Xiao Guo, Chun-Yu Sun, and Xin Tong. O-cnn: Octree-based convolutional neural networks for 3d shape analysis. ACM TOG, 36(4):1-11, 2017. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[49] Peng-Shuai Wang, Chun-Yu Sun, Yang Liu, and Xin Tong. Adaptive o-cnn: A patch-based deep representation of 3d shapes. ACM TOG, 37(6):1-11, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.189 + ], + "angle": 0, + "content": "[50] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph cnn for learning on point clouds. ACM TOG, 38(5):1-12, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.246 + ], + "angle": 0, + "content": "[51] Tete Xiao, Mannat Singh, Eric Mintun, Trevor Darrell, Piotr Dollár, and Ross Girshick. Early convolutions help transformers see better. In NeurIPS, volume 34, pages 30392-30400, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.249, + 0.892, + 0.289 + ], + "angle": 0, + "content": "[52] Yan Yan, Yuxing Mao, and Bo Li. Second: Sparsely embedded convolutional detection. Sensors, 18(10):3337, 2018. 1, 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.291, + 0.892, + 0.332 + ], + "angle": 0, + "content": "[53] Zetong Yang, Yanan Sun, Shu Liu, and Jiaya Jia. 3dssd: Point-based 3d single stage object detector. In CVPR, pages 11040-11048, 2020. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.334, + 0.892, + 0.375 + ], + "angle": 0, + "content": "[54] Zetong Yang, Yanan Sun, Shu Liu, Xiaoyong Shen, and Jiaya Jia. Std: Sparse-to-dense 3d object detector for point cloud. In CVPR, pages 1951-1960, 2019. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.377, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[55] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3d object detection and tracking. In CVPR, pages 11784-11793, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.42, + 0.892, + 0.489 + ], + "angle": 0, + "content": "[56] Manzil Zaheer, Guru Guruganesh, Kumar Avinava Dubey, Joshua Ainslie, Chris Alberti, Santiago Ontanon, Philip Pham, Anirudh Ravula, Qifan Wang, Li Yang, et al. Big bird: Transformers for longer sequences. In NeurIPS, volume 33, pages 17283-17297, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.49, + 0.892, + 0.532 + ], + "angle": 0, + "content": "[57] Cheng Zhang, Haocheng Wan, Xinyi Shen, and Zizhao Wu. Patchformer: An efficient point transformer with patch attention. In CVPR, pages 11799-11808, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.533, + 0.892, + 0.574 + ], + "angle": 0, + "content": "[58] Yanan Zhang, Jiaxin Chen, and Di Huang. Cat-det: Contrastively augmented transformer for multi-modal 3d object detection. In CVPR, pages 908-917, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.576, + 0.892, + 0.631 + ], + "angle": 0, + "content": "[59] Yifan Zhang, Qingyong Hu, Guoquan Xu, Yanxin Ma, Jianwei Wan, and Yulan Guo. Not all points are equal: Learning highly efficient point-based detectors for 3d lidar point clouds. In CVPR, pages 18953-18962, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.633, + 0.892, + 0.686 + ], + "angle": 0, + "content": "[60] Yanan Zhang, Di Huang, and Yunhong Wang. Pc-rgnn: Point cloud completion and graph neural network for 3d object detection. In AAAI, volume 35, pages 3430-3437, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.689, + 0.892, + 0.73 + ], + "angle": 0, + "content": "[61] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In ICCV, pages 16259-16268, 2021. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.732, + 0.892, + 0.786 + ], + "angle": 0, + "content": "[62] Wu Zheng, Weiliang Tang, Sijin Chen, Li Jiang, and Chi-Wing Fu. Cia-ssd: Confident iou-aware single-stage object detector from point cloud. In AAAI, volume 35, pages 3555–3562, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.789, + 0.892, + 0.829 + ], + "angle": 0, + "content": "[63] Yin Zhou and Oncel Tuzel. Voxelnet: End-to-end learning for point cloud based 3d object detection. In CVPR, pages 4490-4499, 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.829 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5175" + } + ] +] \ No newline at end of file diff --git a/2023/OcTr_ Octree-Based Transformer for 3D Object Detection/b6928c01-d853-4d0d-9874-89b3af0f167a_origin.pdf b/2023/OcTr_ Octree-Based Transformer for 3D Object Detection/b6928c01-d853-4d0d-9874-89b3af0f167a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c9caf563bf8fdc389ac406cea321a8a02d99a1e0 --- /dev/null +++ b/2023/OcTr_ Octree-Based Transformer for 3D Object Detection/b6928c01-d853-4d0d-9874-89b3af0f167a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6b98aadf908a87708a15dda5a157db830b53519e68d49c0a59822a354326fd27 +size 1183425 diff --git a/2023/OcTr_ Octree-Based Transformer for 3D Object Detection/full.md b/2023/OcTr_ Octree-Based Transformer for 3D Object Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7f796024ea89a0c08b8fc55382b42f62517f6303 --- /dev/null +++ b/2023/OcTr_ Octree-Based Transformer for 3D Object Detection/full.md @@ -0,0 +1,357 @@ +# OcTr: Octree-based Transformer for 3D Object Detection + +Chao Zhou $^{1,2}$ , Yanan Zhang $^{1,2}$ , Jiaxin Chen $^{2}$ , Di Huang $^{1,2,3*}$ + +1State Key Laboratory of Software Development Environment, Beihang University, Beijing, China + +$^{2}$ School of Computer Science and Engineering, Beihang University, Beijing, China + +$^{3}$ Hangzhou Innovation Institute, Beihang University, Hangzhou, China + +{zhouchaobeing, zhangyanan, jiaxinchen, dhuang}@buaa.edu.cn + +# Abstract + +A key challenge for LiDAR-based 3D object detection is to capture sufficient features from large scale 3D scenes especially for distant or/and occluded objects. Albeit recent efforts made by Transformers with the long sequence modeling capability, they fail to properly balance the accuracy and efficiency, suffering from inadequate receptive fields or coarse-grained holistic correlations. In this paper, we propose an Octree-based Transformer, named OcTr, to address this issue. It first constructs a dynamic octree on the hierarchical feature pyramid through conducting self-attention on the top level and then recursively propagates to the level below restricted by the octants, which captures rich global context in a coarse-to-fine manner while maintaining the computational complexity under control. Furthermore, for enhanced foreground perception, we propose a hybrid positional embedding, composed of the semantic-aware positional embedding and attention mask, to fully exploit semantic and geometry clues. Extensive experiments are conducted on the Waymo Open Dataset and KITTI Dataset, and OcTr reaches newly state-of-the-art results. + +# 1. Introduction + +3D object detection from point clouds has received extensive attention during the past decade for its ability to provide accurate and stable recognition and localization in autonomous driving perception systems. In this task, feature learning plays a very fundamental and crucial role; yet it is rather challenging due to not only the disordered and sparse nature of data sampling, but also to insufficient acquisition under occlusion or at a distance. To address this issue, many methods have been proposed, which can be taxonomized into two major classes, i.e. grid-based and point-based. The former first regularize point clouds into multi-view images or voxels and then apply 2D or 3D CNNs to build shape rep- + +![](images/2b31e2f8605ec3d3cb51f157ddcc16cee3ae8767cbf4da1de378779044aa5931.jpg) +(1) Fixed Pattern Limited Receptive Field +(2) Set Proxy Limited Representation + +![](images/9ffd3a5a130671d8edff22494faaded510bca63d3781060532966f5ddbaef2e4.jpg) +Figure 1. Illustration of three sparsification strategies of attention matrices. Fixed pattern (1) narrows receptive fields and set proxy (2) discards elaborate correlations. The proposed octree construction (3) keeps the global receptive field in a coarse-grained manner while maintaining fine-grained representations. + +resentations [4, 52], while the latter directly conduct MLP based networks such as PointNet++ [33] and DGCNN [50] on original points for geometry description [32, 40, 42, 60]. Unfortunately, they fail to capture necessary context information through the small receptive fields in the deep models, leading to limited results. + +Witnessing the recent success of Transformers in NLP, many studies have investigated and extended such architectures for 3D vision [24, 29, 58, 61]. Transformers are reputed to model long-range dependencies, delivering global receptive fields, and to be suitable for scattered inputs of arbitrary sizes. Meanwhile, in contrast to those static weights that are learned in convolutions, Transformers dynamically aggregate the input features according to the relationships + +between tokens. Regarding the case in 3D object detection, compared to point-based Transformers [11,61], voxel-based ones show the superiority in efficiency. However, they tend to suffer heavy computations when dealing with large scale scenes because of the quadratic complexity of Transformers, with the underlying dilemma between the grid size and the grid amount in voxelization. Taking the KITTI dataset as an example, it is unrealistic for Transformers to operate on the feature map with the spatial shape of $200 \times 176 \times 5$ , which is commonly adopted in most of the detection heads [38,46,52,55]. + +More recently, there have appeared an influx of efficient self-attention model variants that attempt to tackle long sequences as input. They generally sparsify the attention matrix by fixed patterns [7, 23, 34], learned patterns [21, 45] or a combination of them [1, 56]. Fixed patterns chunk the input sequence into blocks of local windows or dilation windows, whilst learned patterns determine a notion of token relevance and eliminate or cluster outliers. Specific to 3D object detection from point clouds, VoTr [24] modifies self-attention with pre-defined patterns including local windows and stride dilation ones in a sparse query manner, and the dilation mechanism enlarges the receptive field by sampling attending tokens in a radius. SST [9] splits input tokens into non-overlapping patterns in a block-wise way and enables window shifting to capture cross-window correlation. Despite some improvements reported, they both only achieve bigger local receptive fields rather than the expected global ones, and computations still increase rapidly with the expansion of receptive fields. + +Another alternative on self-attention is to take advantage of a proxy memory bank which has the access to the entire sequence tokens [1, 2, 56]. By using a small number of induced proxies to compress the whole sequence, it diffuses the global context efficiently. VoxSet [12] adapts Set Transformer [19] to 3D object detection and exploits an induced set to model a set-to-set point cloud translation. With the help of the compressed global proxies and Conv-FFN, it obtains a global receptive field; nevertheless, as they admit, it is sub-optimal to set only a few latent codes as proxies for a large 3D scene, prone to impairing the representation of different point cloud structures and their correlations. Therefore, there remains space for a stronger solution. + +In this paper, we present a novel Transformer network, namely Octree-based Transformer (OcTr), for 3D object detection. We firstly devise an octree-based learnable sparse pattern, i.e. OctAttn, which meticulously and efficiently encodes point clouds of scenes as shown in Fig. 1. The OctAttn module constructs a feature pyramid by gathering and applies self-attention to the top level of the feature pyramid to select the most relevant tokens, which are deemed as the octants to be divided in the subsequent. When propagating to the level below, the key/value inputs are restricted by the + +octants from the top. Through recursively conducting this process, OctAttn captures rich global context features by a global receptive field in a coarse-to-fine manner while reducing the quadratic complexity of vanilla self-attention to the linear complexity. In addition, for better foreground perception, we propose a hybrid positional embedding, which consists of the semantic-aware positional embedding and attention mask, to fully exploit geometry and semantic clues. Thanks to the designs above, OcTr delivers a competitive trade-off between accuracy and efficiency. + +Our contribution is summarized in three-fold: + +1. We propose OcTr for voxel-based 3D object detection, which efficiently learns enhanced representations by a global receptive field with rich contexts. +2. We propose an octree-based learnable attention sparsification scheme (OctAttn) and a hybrid positional embedding combining geometry and semantics. +3. We carry out experiments on the Waymo Open Dataset (WOD) and the KITTI dataset and report state-of-the-art performance with significant gains on far objects. + +# 2. Related Work + +# 2.1. 3D Object Detection from Point Clouds + +There exist two prevailing point-cloud representations in 3D object detection, i.e. point-based and voxel-based. + +The point-based methods [31,40,42,60] directly process raw point clouds in the irregular 3D space. As a pioneering attempt, F-Pointnet [32] employs instance segmentation in frustums to extract proposals. VoteNet [31] clusters objects from the surface in a deep Hough voting manner. PointRCNN [40] generates 3D RoIs with foreground segmentation and applies an RCNN-style [35] two-stage refinement. Different from PointRCNN, some existing methods [3,53,59] build a lightweight and efficient single stage 3D object detection framework. However, the current point-based methods still suffer from a large computation burden, which is not suitable for large-scale point cloud scenes. + +The voxel-based ones [4,6,13,52,55,62,63] conduct voxelization on entire point clouds to construct regular grids. VoxelNet [63] exploits the voxel feature encoding layer with 3D convolutions to extract the feature of each voxel. SECOND [52] improves the model with sparse 3D convolutions, significantly increasing the speed of both training and inference. Pointpillars [18] compacts point clouds into vertical columns and encodes them with 2D CNNs. Several recent methods [13, 38, 54] also explore merging point-based and voxel-based networks into one framework for complementary features from different representations of point clouds. Unfortunately, they all use small convolution kernels with limited receptive fields, which are not competent to capture global context that is important to 3D detection. + +![](images/e9ca629127423699a851f636446a042a215fcb31ee294f13960afbf86af50a5f.jpg) +Figure 2. Framework overview of the proposed Octree-based Transformer (OcTr) model. + +# 2.2. Transformer in 3D Vision + +Inspired by the great success of the self-attention mechanism in NLP [47] and CV [8, 23], Transformers have been adapted to 3D vision for their ability to capture long-range dependencies. For instance, Point Transformer [61] brings in vector attention that modulates individual feature channels for point cloud classification and segmentation; PCT [11] presents offset-attention with the implicit Laplace operator and normalization refinement which is more suitable for point cloud learning. To address the high latency, some methods [30, 57] adopt voxels or patches for acceleration. + +For 3D detection, 3DETR [28] treats and predicts bounding boxes as sequences in an end-to-end manner. CT3D [37] leverages a channel-wise Transformer architecture to refine the RoI head. To learn context-aware representations, some studies [9,12,24,29] introduce Transformers into a point- or voxel-based encoder. Pointformer [29] stacks local, global and local-global Transformers based on the point-based encoder; VoTr [24] exploits dilated attention with fast query to enlarge receptive fields; VoxSet [12] builds an induced point set as proxies of global context and applies point-to-point translation using voxels as mediums; and SST [9] embraces single strides without down-sampling and conducts window attention combining with its shifted version. Even though they all expand receptive fields by diverse Transformer variants, global context is not adequately involved or efficiently utilized. In contrast, we propose Octree-based Transformer (OcTr) for voxel-based 3D object detection, achieving a true global receptive field that balances accuracy and efficiency. + +# 3. Method + +# 3.1. Framework + +This sub-section describes the overall framework of the proposed OcTr model as shown in Fig. 2. Specifically, as inspired by [51], we first voxelize the point cloud into regular grids and adopt the sparse 3D convolution for patch embedding, where the girds are regarded as the "tokens" and are passed through the Octree Transformer Blocks + +(OTB). Compared with the vanilla Transformer block, the self-attention module is substituted by the proposed octree-attention $OctAttn$ , which encodes global context in a more efficient way. After applying a hybrid semantic embedding on multi-scale features, we sequentially stack two OTBs, tailed by a down-sampling layer. The voxel features are then projected into the BEV view by point-wise convolutions and are passed through a multi-scale dense 2D backbone. Ultimately, an anchor-based or anchor-free RPN head is used for 3D proposal generation, and the RoI head is optional for refinement. Note that our OcTr can be adopted for most of the voxel-based detection frameworks by simply altering 3D backbones. + +# 3.2. Self-attention Revisit + +According to [47], Transformer encoder blocks typically include a multi-head self-attention (MHSA) mechanism, a feed-forward network (FFN), a normalization function, and the residual connections [14]. Given an input sequence $X$ , the principle part of MHSA is formulated as + +$$ +\operatorname {M H S A} (X) = \sum_ {h = 1} ^ {H} W _ {h} \left[ \sigma \left(\frac {X W _ {q} W _ {k} ^ {T} X ^ {T}}{\sqrt {d}}\right) \cdot X W _ {v} \right], \tag {1} +$$ + +where $h$ denotes the index of the head, and $H$ , $\sigma$ , $W$ and $d$ are the amount of heads, softmax function, learnable weight and feature dimension, respectively. The subscripts of $q$ , $k$ and $v$ indicate query, key and value. The inputs and outputs of the MHSA module are connected by residual connectors and normalization layers. The MLP-based FFN connects its inputs/outputs in a similar manner. + +# 3.3. Octree Attention + +An octree is a multi-scale asymmetric and efficient representation for unstructured 3D data such as point clouds. To build an octree for the input point cloud, we recursively sub-divide it in the breadth-first order until the pre-defined octree depth is reached. Whether to sub-divide an octant is determined by the occupancy [36, 48], the surface approximation [44, 49] or a learning algorithm [26]. + +![](images/96bc99e00865db3b240e60cadeb29d51d8cd8d5d934e9de16e6f190fd171ed3a.jpg) +Figure 3. Octree Attention (OctAttn). The OctAttn module constructs a dynamic octree on a hierarchical feature pyramid with a partitive criterion of attention scores. For illustration, an octree representation is individually pruned from the pyramid for each grid in $F$ . For example, the blue grids are the divided octants for the grids with stars. The detailed structures of OcTr-Top and OcTr-Base are shown at the right side. + +![](images/563e90e76e17443f4be4476c9793a81ee2db92313c1c830bfac3a3aa51bc6940.jpg) + +![](images/034cdb1500198579c1ba7e2d834de7809406d207fea9c28dee8523b9c58421b1.jpg) + +Despite of being resource-friendly, the octree representation is discrete and non-differentiable, which motivates us to ameliorate previous cumbersome pre-processing division by introducing a novel octree-based attention mechanism. As depicted in [21, 45], the attention matrices calculated by self-attention imply the relevance of input tokens and guide feature selection. We thus prune the dense attention matrices of the multi-scale feature pyramid to sparse octree attention in an adaptive and parallel manner, namely OctAttn. + +As shown in Fig. 3, let the output feature map and coordinates of convolutional patch embedding be $F_0 \in \mathbb{R}^{M \times d}$ and $I_0 \in \mathbb{R}^{M \times 3}$ ; $M$ and $d$ indicate the amount of the nonempty grids in a batch and feature dimension, respectively. Based on $F_0$ , we generate a multi-scale feature pyramid as + +$$ +C = \left\{F _ {n}, I _ {n} \right\} ^ {N}, n \in [ 0, N), \tag {2} +$$ + +where + +$$ +I _ {n} = \left\lfloor \frac {I _ {0}}{2 ^ {n}} \right\rfloor , F _ {n} = \operatorname {B N} \left(S _ {\max } \left(F _ {0}, I _ {n}\right)\right), \tag {3} +$$ + +$n$ , $N$ , BN and $S_{max}$ are the index of the level of the multiscale feature pyramid, the height of the pyramid, batch normalization [16] and the max scatter function, respectively. + +The pruning begins from the top of the pyramid. The top feature map, i.e. $F_{N - 1}$ , is reorganized to dense input tokens $\bar{F}_{N - 1} \in \mathbb{R}^{B \times m_{N - 1} \times d}$ , where $B$ and $m_{N - 1}$ are the batch size and maximum number of non-empty voxels per batch, respectively. The voxel is padded with 0 if it is empty. As shown in Eq. (4), the MHSA takes $\bar{F}_{N - 1}$ as input and outputs the attention score matrices $\mathcal{A}_{N - 1} \in \mathbb{R}^{B \times m_{N - 1} \times m_{N - 1}}$ and attentive features $\bar{F}_{N - 1}'$ as + +$$ +\begin{array}{l} \mathcal {A} _ {N - 1} = \sum_ {h = 1} ^ {H} \sigma (\frac {\bar {F} _ {N - 1} W _ {q} W _ {k} ^ {T} \bar {F} _ {N - 1} ^ {T}}{\sqrt {d}}), \\ \bar {F} _ {N - 1} ^ {\prime} = \sum_ {h = 1} ^ {H} W _ {h} \left[ \sigma \left(\frac {\bar {F} _ {N - 1} W _ {q} W _ {k} ^ {T} \bar {F} _ {N - 1} ^ {T}}{\sqrt {d}}\right) \cdot \bar {F} _ {N - 1} W _ {v} \right]. \tag {4} \\ \end{array} +$$ + +For each query token, we select the top $k$ attention scores as its most relevant token group in a row-wise way, denoted by $O_{N - 1}\in \mathbb{Z}^{B\times m_{N - 1}\times k}$ . + +When propagating from top to bottom through the pyramid and reaching the $n$ -th level, we uniformly sample limited $K$ attending octants from the selected regions with features in $\bar{F}_n$ and topk indices in $O_{n + 1}$ . We conduct cross-attention instead of self-attention with the dense query sequence of the $n$ -th level, $\bar{Q}_n\in \mathbb{R}^{B\times m_n\times d}$ , and the compact sampled key/value sequence of the $n$ -th level, $\bar{K}_n / \bar{V}_n\in \mathbb{R}^{B\times m_n\times K\times d}$ , which is formulated as below + +$$ +\begin{array}{l} \mathcal {A} _ {n} = \sum_ {h = 1} ^ {H} \sigma \left(\frac {\bar {Q} _ {n} W _ {q} W _ {k} ^ {T} \bar {K} _ {n} ^ {T}}{\sqrt {d}}\right), \tag {5} \\ \bar {F} _ {n} ^ {\prime} = \sum_ {h = 1} ^ {H} W _ {h} \left[ \sigma \left(\frac {\bar {Q} _ {n} W _ {q} W _ {k} ^ {T} \bar {K} _ {n} ^ {T}}{\sqrt {d}}\right) \cdot \bar {V} _ {n} W _ {v} \right]. \\ \end{array} +$$ + +Backing off the sampling, this can be treated as an attention mask on self-attention matrices. The above process is run recursively until reaching the bottom level of the pyramid. + +Furthermore, as the top $k$ selection is a hard decision that disables gradient back-propagation, we adopt the Gumbeltop $k$ technique [17] to perform a differentiable and continuous approximation by replacing the vanilla top $k$ selection. + +The normalized scores used in top $k$ are derived from the distribution in Eq. (6) during training, maintaining the original ones during inference. $g,\tau ,m_{n}$ denote the noise sampled from the Gumbel distribution, the temperature and the amount of non-empty voxels in layer $n$ , respectively. + +$$ +p _ {i} = \frac {\exp \left(\left(\mathcal {A} _ {n} ^ {i} + g _ {i}\right) / \tau\right)}{\sum_ {i} \exp \left(\left(\mathcal {A} _ {n} ^ {i} + g _ {i}\right) / \tau\right)} \in [ 0, 1 ]. \tag {6} +$$ + +In order to leverage the multi-scale features in distinct spatial shapes, we concatenate them by upsampling, which is implemented by inverse indices of the scatter function, followed by a linear projection layer for aligning the input feature dimension. + +As the local context is generally critical for object detection, inspired by [7], we additionally introduce a Locally enhanced Positional Embedding (LePE) which enables local neighbor interactions on the value sequence. With submanifold sparse convolutions, we replace the residual connections in the attention mechanism with LePE. + +Finally, OTB is formulated as follows + +$$ +\tilde {F} = \operatorname {F C} \left(\left\{F _ {N - 1} ^ {\prime} \mid F _ {N - 2} ^ {\prime} \right| \dots \left| F _ {0} ^ {\prime} \right\}\right) + \operatorname {L e P E} \left(F _ {0}\right), \tag {7} +$$ + +$$ +\tilde {F} ^ {\prime} = \mathrm {B N} (\mathrm {F F N} (\tilde {F})) + \tilde {F}, +$$ + +where $F_{n}^{\prime}\in \mathbb{R}^{m_0\times d}$ denotes the compact tensor of the upsampled $\bar{F}_n^\prime$ , | indicates concatenation, and FC denotes the fully-connected layer. + +Besides, we analyze the time complexity of the octree attention as below + +$$ +\begin{array}{l} \mathcal {O} \left(\left(\frac {M}{\omega^ {N - 1}}\right) ^ {2} + \sum_ {n = 0} ^ {N - 2} \frac {K M}{\omega^ {n}}\right) \tag {8} \\ = \mathcal {O} \left(\left(\frac {M}{\omega^ {N - 1}}\right) ^ {2} + \frac {\omega}{\omega - 1} K M \left(1 - \omega^ {1 - N}\right)\right), \\ \end{array} +$$ + +where $\omega$ is the average down-sampling ratio in the sparse voxel representation. + +# 3.4. Semantic Positional Embedding + +Due to the large proportion of background grids in point clouds, the attention matrices are dominated by background grid pairs, leading to a sub-optimal solution. To fully leverage the local 3D shape patterns and original voxel coordinates, we propose a hybrid positional embedding to capture both the geometry and semantic clues as displayed in Fig. 4. + +Specifically, we first segment foreground grids using the supervision from the ground truth. Segmentation scores are predicted by a sub-manifold sparse convolution branch with a sigmoid function; and the focal loss [22] is applied to balance the foreground and background. + +![](images/899be8311de29cb25bdd3d9fb85ee5bea664f6148bd5813802b8e45d11e05ebc.jpg) +Figure 4. Illustration of semantic positional embedding, where an extra foreground segmentation branch is adopted with supervision and semantic scores are concatenated as absolute positional embeddings and serve as relative masks on attention matrices. + +We concatenate the center coordinate $(x, y, z)$ and the semantic score with the feature $f$ in a grid-wise manner, followed by a linear projection without the bias as below + +$$ +\begin{array}{l} \operatorname {S A P E} (X) = \operatorname {F C} _ {d + 4 \rightarrow d} (\{x, y, z, \text {s c o r e} | f \}) \tag {9} \\ = \operatorname {F C} _ {d \rightarrow d} (f) + \operatorname {F C} _ {4 \rightarrow d} (x, y, z, s c o r e). \\ \end{array} +$$ + +Eq. (9) is equivalent to the absolute positional embedding (APE), thus being denoted as the Semantic APE (SAPE). The scatter function such as the mean, max and batch normalization naturally provide position and semantic information of the downsampled voxel grids, making it applicable for the multi-scale feature pyramid. + +Besides the semantic clues implicitly used in SAPE, we employ the Semantic Attention Mask (SAM) based on the segmentation scores and we mask the attention matrices to address correlations between inferior queries and superior keys in a simple yet effective way. Given a scalar attention matrix $\mathcal{A} \in \mathbb{R}^{N_q \times N_k}$ before softmax and the segmentation scores of query and key/value $S_q \in \mathbb{R}^{N_q}$ and $S_k \in \mathbb{R}^{N_k}$ , we formulate the attention matrix after softmax as + +$$ +\mathcal {A} ^ {\prime} = \sigma (- \Gamma \cdot [ 1 - (S _ {q} \geq \delta_ {q}) (S _ {k} \geq \delta_ {k}) ^ {T} ] + \mathcal {A}), \tag {10} +$$ + +where $N_{q},\delta_{q},N_{k},\delta_{k},\sigma$ and $\Gamma$ are the length and threshold of the query sequence, length and threshold of the key sequence, softmax function and infinite scalar, respectively. + +Finally, we broadcast the semantic mask $\mathcal{A}'$ to all heads. + +# 4. Experiments + +We evaluate the proposed OcTr network on the Waymo Open Dataset (WOD) and KITTI dataset, both of which are popular in 3D object detection. In this section, we introduce the benchmarks and implementation details, make comparison to the previous state-of-the-art counterparts, and ablate the key designs of OcTr. + +
ModelVehicle (L1) +mAP/mAPHVehicle (L2) +mAP/mAPHPedes. (L1) +mAP/mAPHPedes. (L2) +mAP/mAPHCyclist (L1) +mAP/mAPHCyclist (L2) +mAP/mAPH
SECOND [52]70.96/70.3462.58/62.0265.23/54.2457.22/47.4957.13/55.6254.97/53.53
PointPillar [18]70.43/69.8362.18/61.6466.21/46.3258.18/40.6455.26/51.7553.18/49.80
PartA²Net [41]74.82/74.3265.88/65.4271.76/63.6462.53/55.3067.35/66.1565.05/63.89
PVRCNN [38]75.41/74.7467.44/66.8071.98/61.2463.70/53.9565.88/64.2563.39/61.82
CenterPoint [55]71.33/70.7663.16/62.6572.09/65.4964.27/58.2368.68/67.3966.11/64.87
LiDAR-RCNN [20]73.5/73.064.7/64.271.2/58.763.1/51.768.6/66.966.1/64.4
Voxel-RCNN [6]75.59/-66.59/--/--/--/--/-
PVRCNN++ [39]77.82/77.3269.07/68.6277.99/71.3669.92/63.7471.80/70.7169.31/68.26
SST† [9]76.22/75.7968.04/67.6481.39/74.0572.82/65.93-/--/-
PDV [15]76.85/76.3369.30/68.8174.19/65.9665.85/58.2868.71/67.5566.49/65.36
Ours78.12/77.6369.79/69.3480.76/74.3972.48/66.5272.58/71.5069.93/68.90
+ +Table 1. Performance on WOD with 202 validation sequences for vehicle (IoU=0.7), pedestrian (IoU=0.5) and cyclist (IoU=0.5), using 20% samples for training. All the results are achieved by the models simultaneously trained for 3 classes on single frames, except the ones of the model marked by $\dagger$ , which is only trained for a single class. Refer to Supp. B for the results trained with 100% samples. + +
ModelmAP3D(L1)@ Vehicle
Overall0-30m30m-50m50m-inf
PV-RCNN [38]70.3091.9269.2142.17
Voxel-RCNN [6]75.5992.4974.0953.15
VoTR-TSD [24]74.9592.2873.3651.09
CT3D [37]76.3092.5175.0755.36
Pyramid_PV [25]76.3092.6774.9154.54
PDV [15]76.8593.1375.4954.75
VoxSeT [12]77.8292.7877.2154.41
Ours78.8292.9977.6658.02
ModelmAP3D(L2)@ Vehicle
Overall0-30m30-50m50m-inf
PV-RCNN [38]65.3691.5865.1336.46
Voxel-RCNN [6]66.5991.7467.8940.80
CT3D [37]69.0491.7668.9342.60
PDV [15]69.3092.4169.3642.16
VoxSeT [12]70.2192.0570.1043.20
Ours70.5091.7871.2845.46
+ +Table 2. Results on the WOD validation set in different ranges for vehicle detection. + +# 4.1. Datasets and Implementation Details + +WOD [43] is a large dataset of autonomous driving scenes. It totally contains 798 training sequences with around 160K LiDAR samples and 202 validation sequences with 40K LiDAR samples, with the mean Average Precision (mAP) and mAP weighted by heading accuracy (mAPH) as evaluation metrics. There are also two levels of difficulty describing the sparsity in each bounding box, and LEVEL_1 (L1) and LEVEL_2 (L2) denote more than 5 points and 1-5 points, respectively. For detection performance along distance, it provides mAP/mAPH on 0-30m, 30m-50m and 50m-inf. + +KITTI [10] is a widely used benchmark for 3D object detection, which includes 3,712, 3,769 and 7,518 frames for training, validation and testing, respectively. mAP is used as the official metric with 11 recall points for the val set and + +40 for the test set, and the IoU thresholds are set to 0.7, 0.5, and 0.5 for car, pedestrian and cyclist. We use the official setting in all experiments. + +Implementation Details The total loss for optimizing the overall two-stage detection is formulated as Eq. (11), where $\mathcal{L}_{rcnn}$ can be omitted if there is no RoI head. Refer to Supp. A.1 for more information. + +$$ +\mathcal {L} _ {\text {d e t}} = \mathcal {L} _ {\text {r p n}} + \mathcal {L} _ {\text {r c n n}} + \mathcal {L} _ {\text {s e g}}. \tag {11} +$$ + +# 4.2. Results on WOD + +The results on the validation set are displayed in Tab. 1, and it can be seen that we achieve new state-of-the-art performance on all the three classes. In particular, for pedestrian, we outperform the baseline model PV-RCNN++ [39] by $2.77\% / 2.56\%$ in terms of both L1 and L2 mAP, which indicates the effectiveness of the proposed model in handling hard examples. + +In comparison with other Transformer-based models, we focus on vehicle since the counterparts only report the performance on it. As Tab. 2 shows, our OcTr achieves the best mAP among all these convolution- and Transformer-based backbones. It also outperforms the Transformer-based detection head network CT3D [37] by $2.52\%$ and $1.46\%$ in L1 and L2 mAP. Regarding the accuracies at different distances, OcTr ranks the first place in the range of $30\mathrm{m} - 50\mathrm{m}$ and $50\mathrm{m}$ inf, which surpasses the previous best by $0.45\%$ , $2.66\%$ in L1 mAP and $1.18\%$ , $2.26\%$ in L2 mAP respectively. It clearly illustrates that OcTr has the advantage in capturing long-range fine-grained context, which facilitates dealing with objects far away. Indeed, far objects generally have much more sparse points than near ones and heavily rely on context for detection, thus benefiting more from OcTr. Refer to Fig. 5 for visualization. + +
Model\( \mathbf{{mAP}}_{3D}@\mathbf{{Car}} \)\( \mathbf{{mAP}}_{3D}@\mathbf{{Pedestrian}} \)\( \mathbf{{mAP}}_{3D}@\mathbf{{Cyclist}} \)
EasyMod.HardEasyMod.HardEasyMod.Hard
SECOND [52]88.6178.6277.2256.5552.9847.7380.5867.1563.10
PointPillars [18]88.4677.2874.6557.7552.2947.9080.0462.6159.52
VoTR [24]87.8678.2776.93------
VoxSeT [12]88.4578.4877.0760.6254.7450.3984.0768.1165.14
Ours88.4378.5777.1661.4957.1752.3585.2970.4466.17
+ +Table 3. Results of the single-stage models on the KITTI val set. All the models adopt the same anchor-based region proposal network as the detection head. "Mod." denotes the moderate difficulty level. + +
Model\( \mathbf{mAP}_{3D}@\mathbf{Car on test} \)\( \mathbf{mAP}_{3D}@\mathbf{Car on val} \)
EasyMod.HardMeanEasyMod.HardMean
SECOND [52]83.3472.5565.8273.9088.6178.6277.2281.48
PointPillars [18]82.5874.3168.9975.2986.6276.0668.9177.20
STD [54]87.9579.7175.0980.9289.7079.8079.3082.93
SA-SSD [13]88.7579.7974.1680.9090.1579.9178.7882.95
3DSSD [53]88.3679.5774.5580.8389.7179.4578.6782.61
PV-RCNN [38]90.2581.4376.8282.8389.3583.6978.7083.91
Voxel-RCNN [6]90.9081.6277.0683.1989.4184.5278.9384.29
CT3D [37]87.8381.7777.1682.2589.5486.0678.9984.86
VoTR-TSD [24]89.9082.0979.1483.7189.0484.0478.6883.92
VoxSeT [12]88.5382.0677.4682.6889.2186.7178.5684.83
Focals Conv [4]90.5582.2877.5983.4789.5284.9379.1884.54
Ours90.8882.6477.7783.7689.8086.9779.2885.35
+ +Table 4. Comparison to the state-of-the-art models on the KITTI test and val sets. "Mod." and "Mean" denote the moderate difficulty level and the average mAP for the three levels, respectively. The best results are bolded and the second best ones are underlined. + +
DetectorVeh. mAP (L1/L2)Pedes. mAP (L1/L2)
SECOND [52]70.96/62.5865.23/57.22
Ours73.28/65.0568.08/60.36
PV-RCNN [38]75.41/67.4471.98/63.70
Ours76.77/68.3173.22/64.30
PV-RCNN++ [39]77.82/69.0777.99/69.92
Ours78.01/69.6080.75/72.45
+ +Table 5. Results of extensions to different representative detectors on the WOD validation set. + +# 4.3. Results on KITTI + +The performance of the single-stage detectors is shown in Tab. 3. We take SECOND [52], a commonly used anchor-based model, as the baseline, and compare OcTr with another two advanced Transformer-based variants VoxSet [12] and VoTR [24]. We can see that OcTr achieves comparable results with SECOND and VoxSeT in car, while it significantly outperforms all the counterparts and reports the best performance both in pedestrian and cyclist. Benefiting from the large receptive field and fine-grained global context, it exceeds SECOND by $4.19\%$ and $3.29\%$ for pedestrians and cyclists respectively, where hard samples often appear. + +We summarize the performance of the two-stage models on the KITTI test set in Tab. 4. With the help of the + +
AttentionVeh. mAP (L1/L2)Pedes. mAP (L1/L2)
Ours (OctAttn)73.3/65.168.1/60.4
Performer [5]71.4/63.665.7/57.9
ACT [27]71.7/63.564.3/56.1
VoTr [24]69.4/61.565.0/57.0
Nearest K68.2/59.864.9/56.7
+ +Table 6. Ablation on various attention mechanisms and sampling patterns on the WOD validation set. + +multi-scale backbone features and rich global context, OcTr reaches a leading mAP in car at the moderate level, surpassing the state-of-the-art Focals-Conv [4] by $0.36\%$ . We also evaluate OcTr on the KITTI val set, and OcTr again delivers the best performance in the average score, outperforming the second-best by $0.49\%$ . One can observe that we rank the best or the second best in all the cases. + +# 4.4. Ablation study + +Scalability on various detectors As summarized in Tab. 5, we conduct experiments on three different and representative detectors, SECOND (single-stage, anchor-based), PV-RCNN (two-stage, anchor-based) and PV-RCNN++ (two-stage, anchor-free). Regardless of the number of stages or region proposal network, we acquire sound improvements + +
LEPESAPESAMVeh. mAP (L1/L2)Pedes. mAP (L1/L2)
71.35/63.3065.75/57.89
✓72.34/64.3266.56/58.62
✓✓72.64/64.4666.62/58.83
✓✓72.86/64.4067.79/59.90
✓✓✓73.28/65.0568.08/60.36
+ +Table 7. Ablation on semantic positional embedding on the WOD validation set. + +
topk numberVeh. mAP (L1/L2)Pedes. mAP (L1/L2)
170.38/62.2064.19/56.43
472.58/64.4266.21/58.42
873.28/65.0568.08/60.36
1673.25/65.0167.89/60.10
+ +Table 8. Results of different $k$ values on the WOD validation set. + +
Method#Param. (M)Latency (ms)Memory (GB)
SECOND [52]5.3482.3
VoTR-SSD [24]4.8673.0
VoxSeT-SSD [12]3.0373.6
OcTr-SSD2.9642.5
+ +Table 9. Resource costs of different backbones with single-stage detectors on the KITTI dataset, test on GTX2080Ti. + +compared to the baselines of the sparse convolution backbones, highlighting its scalability. + +Ablation on OctAttn We carry out additional experiments to make apple-to/apple comparison with several representative linear Transformer methods, including Performer [5] (kernel-based linear attention), ACT [27] (cluster-based linear attention), VoTr [24] (fixed patterns) and the Nearest-K strategy. Tab. 6 lists the results, and our OctAttn clearly performs the best, showing its ability. + +Ablation on semantic positional embedding We individually evaluate the contributions of LePE, SAPE and SAM with SECOND as the baseline detector on WOD in Tab. 7. By incorporating LePE, the L1/L2 performance is boosted by $0.99\% / 1.02\%$ and $0.81\% / 0.73\%$ on vehicle and pedestrian, illustrating its necessity. Furthermore, we separately verify the validity of SAPE and SAM. With semantic clues, we observe that SAPE increases by $0.3\%$ L1 mAP on vehicle, while SAM provides an L1/L2 mAP improvement of $0.52\% / 0.08\%$ and $1.23\% / 1.28\%$ on vehicle and pedestrian. Finally, we simultaneously apply SAPE and SAM and construct the full model of OcTr, which gains $0.94\% / 0.73\%$ and $1.52\% / 1.74\%$ L1/L2 mAP on vehicle and pedestrian, showing its impact. + +Influence by topk Tab. 8 shows the performance of OcTr with various values of $k$ . As we can see, the performance improves when $k$ becomes larger but quickly saturates. Due to the redundancy of the scanned scenes, we argue that only a few tokens with high relevance need to be subdivided to + +embrace fine-grained features. + +Analysis on model complexity We compare OcTr with two recent Transformer-based models in terms of resource cost by keeping the same detection head. Tab. 9 shows that with the learnable octree attention mechanism, OcTr consistently maintains less model parameters and less memory occupancies than the counterparts. Regarding the inference speed, VoxSeT runs faster, but it should be noted that VoxSeT inputs with pillars which discard the height dimension, leading to inferior results. As for VoTR, we deliver a mild improvement in efficiency while bringing a large gain in performance. + +![](images/772b9d9f91728cd492ba4862a309cb6ecbe9349aec506fc8fbb76e755af67f38.jpg) + +![](images/359cc85ce0005bf04fb6f6291db3c1b6ed59be462b73ca75830975ca33e96e37.jpg) +Figure 5. Visualization of results by OcTr on the WOD validation split. Blue/red indicates predicted/ground-truth bounding boxes + +# 5. Conclusion + +This paper proposes a novel voxel-based approach to 3D object detection, namely OcTr. It aims to balance the fine-grained global representation and efficiency with acceptable resource costs. To this end, we propose a learned sparsification attention mechanism, OctAttn, which adaptively prunes the octants from the multi-scale feature pyramid in a top-to-bottom manner. Furthermore, we adopt a hybrid semantic-aware positional embedding based on foreground segmentation. Extensive experiments are conducted on WOD and KITTI, and OcTr reaches the state-of-the-art performance, validating its effectiveness. + +# Acknowledgment + +This work is partly supported by the National Natural Science Foundation of China (No. 62022011 and No. 62202034), the Research Program of State Key Laboratory of Software Development Environment (SKLSDE-2021ZX-04), and the Fundamental Research Funds for the Central Universities. + +# References + +[1] Joshua Ainslie, Santiago Ontanon, Chris Alberti, Vaclav Cvicek, Zachary Fisher, Philip Pham, Anirudh Ravula, Sumit Sanghai, Qifan Wang, and Li Yang. Etc: Encoding long and structured inputs in transformers. In EMNLP, 2020. 2 +[2] Iz Beltagy, Matthew E Peters, and Arman Cohan. Longformer: The long-document transformer. arXiv preprint arXiv:2004.05150, 2020. 2 +[3] Chen Chen, Zhe Chen, Jing Zhang, and Dacheng Tao. Sasa: Semantics-augmented set abstraction for point-based 3d object detection. In AAAI, volume 1, 2022. 2 +[4] Yukang Chen, Yanwei Li, Xiangyu Zhang, Jian Sun, and Jiaya Jia. Focal sparse convolutional networks for 3d object detection. In CVPR, pages 5428-5437, 2022. 1, 2, 7 +[5] Krzysztof Marcin Choromanski, Valerii Likhosherstov, David Dohan, Xingyou Song, Andreea Gane, Tamás Sarlós, Peter Hawkins, Jared Quincy Davis, Afroz Mohiuddin, Lukasz Kaiser, David Benjamin Belanger, Lucy J. Colwell, and Adrian Weller. Rethinking attention with performers. In ICLR, 2021. 7, 8 +[6] Jiajun Deng, Shaoshuai Shi, Peiwei Li, Wengang Zhou, Yanyong Zhang, and Houqiang Li. Voxel r-cnn: Towards high performance voxel-based 3d object detection. In AAAI, volume 35, pages 1201-1209, 2021. 2, 6, 7 +[7] Xiaoyi Dong, Jianmin Bao, Dongdong Chen, Weiming Zhang, Nenghai Yu, Lu Yuan, Dong Chen, and Baining Guo. Cswin transformer: A general vision transformer backbone with cross-shaped windows. In CVPR, pages 12124-12134, 2022. 2, 5 +[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2020. 3 +[9] Lue Fan, Ziqi Pang, Tianyuan Zhang, Yu-Xiong Wang, Hang Zhao, Feng Wang, Naiyan Wang, and Zhaoxiang Zhang. Embracing single stride 3d object detector with sparse transformer. In CVPR, pages 8458-8468, 2022. 2, 3, 6 +[10] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In CVPR, pages 3354-3361. IEEE, 2012. 6 +[11] Meng-Hao Guo, Jun-Xiong Cai, Zheng-Ning Liu, Tai-Jiang Mu, Ralph R Martin, and Shi-Min Hu. Pct: Point cloud transformer. CVM, 7(2):187-199, 2021. 2, 3 +[12] Chenhang He, Ruihuang Li, Shuai Li, and Lei Zhang. Voxel set transformer: A set-to-set approach to 3d object detection from point clouds. In CVPR, pages 8417-8427, 2022. 2, 3, 6, 7, 8 +[13] Chenhang He, Hui Zeng, Jianqiang Huang, Xian-Sheng Hua, and Lei Zhang. Structure aware single-stage 3d object detection from point cloud. In CVPR, pages 11873-11882, 2020. 2, 7 +[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 3 + +[15] Jordan SK Hu. Point density-aware voxels for lidar 3d object detection. In CVPR, pages 8469-8478, 2022. 6 +[16] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In ICML, pages 448-456. PMLR, 2015. 4 +[17] Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax. In ICLR, 2016. 4 +[18] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In CVPR, pages 12697-12705, 2019. 2, 6, 7 +[19] Juho Lee, Yoonho Lee, Jungtaek Kim, Adam Kosiorek, Seungjin Choi, and Yee Whye Teh. Set transformer: A framework for attention-based permutation-invariant neural networks. In ICML, pages 3744-3753. PMLR, 2019. 2 +[20] Zhichao Li, Feng Wang, and Naiyan Wang. Lidar r-cnn: An efficient and universal 3d object detector. In CVPR, pages 7546-7555, 2021. 6 +[21] Youwei Liang, Chongjian Ge, Zhan Tong, Yibing Song, Jue Wang, and Pengtao Xie. Not all patches are what you need: Expediting vision transformers via token reorganizations. In ICLR, 2022. 2, 4 +[22] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In ICCV, pages 2980-2988, 2017. 5 +[23] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 10012-10022, 2021. 2, 3 +[24] Jiageng Mao. Voxel transformer for 3d object detection. In ICCV, pages 3164-3173, 2021. 1, 2, 3, 6, 7, 8 +[25] Jiageng Mao, Minzhe Niu, Haoyue Bai, Xiaodan Liang, Hang Xu, and Chunjing Xu. Pyramid r-cnn: Towards better performance and adaptability for 3d object detection. In ICCV, pages 2723-2732, 2021. 6 +[26] Julien NP Martel, David B Lindell, Connor Z Lin, Eric R Chan, Marco Monteiro, and Gordon Wetzstein. Acorn: Adaptive coordinate networks for neural scene representation. In ACM SIGGRAPH, 2021. 3 +[27] Zheng Minghang, Gao Peng, Zhang Renrui, Li Kunchang, Li Hongsheng, and Dong Hao. End-to-end object detection with adaptive clustering transformer. In BMVC, page 226, 2021. 7, 8 +[28] Ishan Misra, Rohit Girdhar, and Armand Joulin. An end-to-end transformer model for 3d object detection. In ICCV, pages 2906-2917, 2021. 3 +[29] Xuran Pan, Zhuofan Xia, Shiji Song, Li Erran Li, and Gao Huang. 3d object detection with pointformer. In CVPR, pages 7463-7472, 2021. 1, 3 +[30] Chunghyun Park, Yoonwoo Jeong, Minsu Cho, and Jaesik Park. Fast point transformer. In CVPR, pages 16949-16958, 2022. 3 +[31] Charles R Qi, Or Litany, Kaiming He, and Leonidas J Guibas. Deep hough voting for 3d object detection in point clouds. In ICCV, pages 9277-9286, 2019. 2 +[32] Charles R Qi, Wei Liu, Chenxia Wu, Hao Su, and Leonidas J Guibas. Frustum pointnets for 3d object detection from rgb-d data. In CVPR, pages 918-927, 2018. 1, 2 + +[33] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In NeurIPS, volume 30, 2017. 1 +[34] Jiezhong Qiu, Hao Ma, Omer Levy, Scott Wen-tau Yih, Sinong Wang, and Jie Tang. Blockwise self-attention for long document understanding. arXiv preprint arXiv:1911.02972, 2019. 2 +[35] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In NeurIPS, volume 28, 2015. 2 +[36] Gernot Riegler, Ali Osman Ulusoy, and Andreas Geiger. Octnet: Learning deep 3d representations at high resolutions. In CVPR, pages 3577-3586, 2017. 3 +[37] Hualian Sheng, Sijia Cai, Yuan Liu, Bing Deng, Jianqiang Huang, Xian-Sheng Hua, and Min-Jian Zhao. Improving 3d object detection with channel-wise transformer. In ICCV, pages 2743-2752, 2021. 3, 6, 7 +[38] Shaoshuai Shi, Chaoxu Guo, Li Jiang, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. Pv-rcnn: Point-voxel feature set abstraction for 3d object detection. In CVPR, pages 10529–10538, 2020. 2, 6, 7 +[39] Shaoshuai Shi, Li Jiang, Jiajun Deng, Zhe Wang, Chaoxu Guo, Jianping Shi, Xiaogang Wang, and Hongsheng Li. Pvrccnn++: Point-voxel feature set abstraction with local vector representation for 3d object detection. IEEE TPAMI, 2021. 6, 7 +[40] Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. Pointr-cnn: 3d object proposal generation and detection from point cloud. In CVPR, pages 770-779, 2019. 1, 2 +[41] Shaoshuai Shi, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. From points to parts: 3d object detection from point cloud with part-aware and part-aggregation network. IEEE TPAMI, 43(8):2647-2664, 2020. 6 +[42] Weijing Shi and Raj Rajkumar. Point-gnn: Graph neural network for 3d object detection in a point cloud. In CVPR, pages 1711-1719, 2020. 1, 2 +[43] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In CVPR, pages 2446-2454, 2020. 6 +[44] Jia-Heng Tang, Weikai Chen, Jie Yang, Bo Wang, Songrun Liu, Bo Yang, and Lin Gao. Octfield: Hierarchical implicit functions for 3d modeling. In NeurIPS, 2021. 3 +[45] Shitao Tang, Jiahui Zhang, Siyu Zhu, and Ping Tan. Quadtree attention for vision transformers. In ICLR, 2022. 2, 4 +[46] OpenPCDet Development Team. Openpcdet: An opensource toolbox for 3d object detection from point clouds. https://github.com/open-mmlab/OpenPCDet, 2020.2 +[47] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Ilia Polosukhin. Attention is all you need. In NeurIPS, volume 30, 2017. 3 +[48] Peng-Shuai Wang, Yang Liu, Yu-Xiao Guo, Chun-Yu Sun, and Xin Tong. O-cnn: Octree-based convolutional neural networks for 3d shape analysis. ACM TOG, 36(4):1-11, 2017. 3 + +[49] Peng-Shuai Wang, Chun-Yu Sun, Yang Liu, and Xin Tong. Adaptive o-cnn: A patch-based deep representation of 3d shapes. ACM TOG, 37(6):1-11, 2018. 3 +[50] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph cnn for learning on point clouds. ACM TOG, 38(5):1-12, 2019. 1 +[51] Tete Xiao, Mannat Singh, Eric Mintun, Trevor Darrell, Piotr Dollár, and Ross Girshick. Early convolutions help transformers see better. In NeurIPS, volume 34, pages 30392-30400, 2021. 3 +[52] Yan Yan, Yuxing Mao, and Bo Li. Second: Sparsely embedded convolutional detection. Sensors, 18(10):3337, 2018. 1, 2, 6, 7, 8 +[53] Zetong Yang, Yanan Sun, Shu Liu, and Jiaya Jia. 3dssd: Point-based 3d single stage object detector. In CVPR, pages 11040-11048, 2020. 2, 7 +[54] Zetong Yang, Yanan Sun, Shu Liu, Xiaoyong Shen, and Jiaya Jia. Std: Sparse-to-dense 3d object detector for point cloud. In CVPR, pages 1951-1960, 2019. 2, 7 +[55] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3d object detection and tracking. In CVPR, pages 11784-11793, 2021. 2, 6 +[56] Manzil Zaheer, Guru Guruganesh, Kumar Avinava Dubey, Joshua Ainslie, Chris Alberti, Santiago Ontanon, Philip Pham, Anirudh Ravula, Qifan Wang, Li Yang, et al. Big bird: Transformers for longer sequences. In NeurIPS, volume 33, pages 17283-17297, 2020. 2 +[57] Cheng Zhang, Haocheng Wan, Xinyi Shen, and Zizhao Wu. Patchformer: An efficient point transformer with patch attention. In CVPR, pages 11799-11808, 2022. 3 +[58] Yanan Zhang, Jiaxin Chen, and Di Huang. Cat-det: Contrastively augmented transformer for multi-modal 3d object detection. In CVPR, pages 908-917, 2022. 1 +[59] Yifan Zhang, Qingyong Hu, Guoquan Xu, Yanxin Ma, Jianwei Wan, and Yulan Guo. Not all points are equal: Learning highly efficient point-based detectors for 3d lidar point clouds. In CVPR, pages 18953-18962, 2022. 2 +[60] Yanan Zhang, Di Huang, and Yunhong Wang. Pc-rgnn: Point cloud completion and graph neural network for 3d object detection. In AAAI, volume 35, pages 3430-3437, 2021. 1, 2 +[61] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In ICCV, pages 16259-16268, 2021. 1, 2, 3 +[62] Wu Zheng, Weiliang Tang, Sijin Chen, Li Jiang, and Chi-Wing Fu. Cia-ssd: Confident iou-aware single-stage object detector from point cloud. In AAAI, volume 35, pages 3555–3562, 2021. 2 +[63] Yin Zhou and Oncel Tuzel. Voxelnet: End-to-end learning for point cloud based 3d object detection. In CVPR, pages 4490-4499, 2018. 2 \ No newline at end of file diff --git a/2023/OcTr_ Octree-Based Transformer for 3D Object Detection/images.zip b/2023/OcTr_ Octree-Based Transformer for 3D Object Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..282fbf4bcd5f282f940725b16d6a2981b3e333a6 --- /dev/null +++ b/2023/OcTr_ Octree-Based Transformer for 3D Object Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5c12376c9dabd6c135809b35ee5f68c9012fdce736688d3cf963d652978703e9 +size 861898 diff --git a/2023/OcTr_ Octree-Based Transformer for 3D Object Detection/layout.json b/2023/OcTr_ Octree-Based Transformer for 3D Object Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..1ac025a5c131248c6ed411bc4da0ffb92816f0a9 --- /dev/null +++ b/2023/OcTr_ Octree-Based Transformer for 3D Object Detection/layout.json @@ -0,0 +1,9046 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 118, + 103, + 476, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 103, + 476, + 120 + ], + "spans": [ + { + "bbox": [ + 118, + 103, + 476, + 120 + ], + "type": "text", + "content": "OcTr: Octree-based Transformer for 3D Object Detection" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 153, + 142, + 443, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 142, + 443, + 157 + ], + "spans": [ + { + "bbox": [ + 153, + 142, + 443, + 157 + ], + "type": "text", + "content": "Chao Zhou" + }, + { + "bbox": [ + 153, + 142, + 443, + 157 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 153, + 142, + 443, + 157 + ], + "type": "text", + "content": ", Yanan Zhang" + }, + { + "bbox": [ + 153, + 142, + 443, + 157 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 153, + 142, + 443, + 157 + ], + "type": "text", + "content": ", Jiaxin Chen" + }, + { + "bbox": [ + 153, + 142, + 443, + 157 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 153, + 142, + 443, + 157 + ], + "type": "text", + "content": ", Di Huang" + }, + { + "bbox": [ + 153, + 142, + 443, + 157 + ], + "type": "inline_equation", + "content": "^{1,2,3*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 60, + 157, + 534, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 157, + 534, + 171 + ], + "spans": [ + { + "bbox": [ + 60, + 157, + 534, + 171 + ], + "type": "text", + "content": "1State Key Laboratory of Software Development Environment, Beihang University, Beijing, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 97, + 171, + 496, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 171, + 496, + 185 + ], + "spans": [ + { + "bbox": [ + 97, + 171, + 496, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 97, + 171, + 496, + 185 + ], + "type": "text", + "content": "School of Computer Science and Engineering, Beihang University, Beijing, China" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 127, + 185, + 467, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 185, + 467, + 199 + ], + "spans": [ + { + "bbox": [ + 127, + 185, + 467, + 199 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 127, + 185, + 467, + 199 + ], + "type": "text", + "content": "Hangzhou Innovation Institute, Beihang University, Hangzhou, China" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 135, + 201, + 454, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 201, + 454, + 213 + ], + "spans": [ + { + "bbox": [ + 135, + 201, + 454, + 213 + ], + "type": "text", + "content": "{zhouchaobeing, zhangyanan, jiaxinchen, dhuang}@buaa.edu.cn" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 143, + 241, + 191, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 241, + 191, + 253 + ], + "spans": [ + { + "bbox": [ + 143, + 241, + 191, + 253 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 266, + 290, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 290, + 506 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 290, + 506 + ], + "type": "text", + "content": "A key challenge for LiDAR-based 3D object detection is to capture sufficient features from large scale 3D scenes especially for distant or/and occluded objects. Albeit recent efforts made by Transformers with the long sequence modeling capability, they fail to properly balance the accuracy and efficiency, suffering from inadequate receptive fields or coarse-grained holistic correlations. In this paper, we propose an Octree-based Transformer, named OcTr, to address this issue. It first constructs a dynamic octree on the hierarchical feature pyramid through conducting self-attention on the top level and then recursively propagates to the level below restricted by the octants, which captures rich global context in a coarse-to-fine manner while maintaining the computational complexity under control. Furthermore, for enhanced foreground perception, we propose a hybrid positional embedding, composed of the semantic-aware positional embedding and attention mask, to fully exploit semantic and geometry clues. Extensive experiments are conducted on the Waymo Open Dataset and KITTI Dataset, and OcTr reaches newly state-of-the-art results." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 529, + 127, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 529, + 127, + 541 + ], + "spans": [ + { + "bbox": [ + 47, + 529, + 127, + 541 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 550, + 287, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 550, + 287, + 694 + ], + "spans": [ + { + "bbox": [ + 46, + 550, + 287, + 694 + ], + "type": "text", + "content": "3D object detection from point clouds has received extensive attention during the past decade for its ability to provide accurate and stable recognition and localization in autonomous driving perception systems. In this task, feature learning plays a very fundamental and crucial role; yet it is rather challenging due to not only the disordered and sparse nature of data sampling, but also to insufficient acquisition under occlusion or at a distance. To address this issue, many methods have been proposed, which can be taxonomized into two major classes, i.e. grid-based and point-based. The former first regularize point clouds into multi-view images or voxels and then apply 2D or 3D CNNs to build shape rep-" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 312, + 249, + 541, + 343 + ], + "blocks": [ + { + "bbox": [ + 312, + 249, + 541, + 343 + ], + "lines": [ + { + "bbox": [ + 312, + 249, + 541, + 343 + ], + "spans": [ + { + "bbox": [ + 312, + 249, + 541, + 343 + ], + "type": "image", + "image_path": "2b31e2f8605ec3d3cb51f157ddcc16cee3ae8767cbf4da1de378779044aa5931.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 329, + 344, + 435, + 352 + ], + "lines": [ + { + "bbox": [ + 329, + 344, + 435, + 352 + ], + "spans": [ + { + "bbox": [ + 329, + 344, + 435, + 352 + ], + "type": "text", + "content": "(1) Fixed Pattern Limited Receptive Field" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 449, + 345, + 541, + 354 + ], + "lines": [ + { + "bbox": [ + 449, + 345, + 541, + 354 + ], + "spans": [ + { + "bbox": [ + 449, + 345, + 541, + 354 + ], + "type": "text", + "content": "(2) Set Proxy Limited Representation" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 327, + 360, + 526, + 453 + ], + "blocks": [ + { + "bbox": [ + 327, + 360, + 526, + 453 + ], + "lines": [ + { + "bbox": [ + 327, + 360, + 526, + 453 + ], + "spans": [ + { + "bbox": [ + 327, + 360, + 526, + 453 + ], + "type": "image", + "image_path": "9ffd3a5a130671d8edff22494faaded510bca63d3781060532966f5ddbaef2e4.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 466, + 545, + 521 + ], + "lines": [ + { + "bbox": [ + 305, + 466, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 305, + 466, + 545, + 521 + ], + "type": "text", + "content": "Figure 1. Illustration of three sparsification strategies of attention matrices. Fixed pattern (1) narrows receptive fields and set proxy (2) discards elaborate correlations. The proposed octree construction (3) keeps the global receptive field in a coarse-grained manner while maintaining fine-grained representations." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 544, + 545, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 544, + 545, + 616 + ], + "spans": [ + { + "bbox": [ + 304, + 544, + 545, + 616 + ], + "type": "text", + "content": "resentations [4, 52], while the latter directly conduct MLP based networks such as PointNet++ [33] and DGCNN [50] on original points for geometry description [32, 40, 42, 60]. Unfortunately, they fail to capture necessary context information through the small receptive fields in the deep models, leading to limited results." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 617, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 546, + 715 + ], + "type": "text", + "content": "Witnessing the recent success of Transformers in NLP, many studies have investigated and extended such architectures for 3D vision [24, 29, 58, 61]. Transformers are reputed to model long-range dependencies, delivering global receptive fields, and to be suitable for scattered inputs of arbitrary sizes. Meanwhile, in contrast to those static weights that are learned in convolutions, Transformers dynamically aggregate the input features according to the relationships" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 703, + 176, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 703, + 176, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 703, + 176, + 712 + ], + "type": "text", + "content": "*indicates the corresponding author." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5166" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": "between tokens. Regarding the case in 3D object detection, compared to point-based Transformers [11,61], voxel-based ones show the superiority in efficiency. However, they tend to suffer heavy computations when dealing with large scale scenes because of the quadratic complexity of Transformers, with the underlying dilemma between the grid size and the grid amount in voxelization. Taking the KITTI dataset as an example, it is unrealistic for Transformers to operate on the feature map with the spatial shape of " + }, + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "inline_equation", + "content": "200 \\times 176 \\times 5" + }, + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": ", which is commonly adopted in most of the detection heads [38,46,52,55]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 206, + 289, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 206, + 289, + 434 + ], + "spans": [ + { + "bbox": [ + 46, + 206, + 289, + 434 + ], + "type": "text", + "content": "More recently, there have appeared an influx of efficient self-attention model variants that attempt to tackle long sequences as input. They generally sparsify the attention matrix by fixed patterns [7, 23, 34], learned patterns [21, 45] or a combination of them [1, 56]. Fixed patterns chunk the input sequence into blocks of local windows or dilation windows, whilst learned patterns determine a notion of token relevance and eliminate or cluster outliers. Specific to 3D object detection from point clouds, VoTr [24] modifies self-attention with pre-defined patterns including local windows and stride dilation ones in a sparse query manner, and the dilation mechanism enlarges the receptive field by sampling attending tokens in a radius. SST [9] splits input tokens into non-overlapping patterns in a block-wise way and enables window shifting to capture cross-window correlation. Despite some improvements reported, they both only achieve bigger local receptive fields rather than the expected global ones, and computations still increase rapidly with the expansion of receptive fields." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 436, + 289, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 436, + 289, + 592 + ], + "spans": [ + { + "bbox": [ + 46, + 436, + 289, + 592 + ], + "type": "text", + "content": "Another alternative on self-attention is to take advantage of a proxy memory bank which has the access to the entire sequence tokens [1, 2, 56]. By using a small number of induced proxies to compress the whole sequence, it diffuses the global context efficiently. VoxSet [12] adapts Set Transformer [19] to 3D object detection and exploits an induced set to model a set-to-set point cloud translation. With the help of the compressed global proxies and Conv-FFN, it obtains a global receptive field; nevertheless, as they admit, it is sub-optimal to set only a few latent codes as proxies for a large 3D scene, prone to impairing the representation of different point cloud structures and their correlations. Therefore, there remains space for a stronger solution." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": "In this paper, we present a novel Transformer network, namely Octree-based Transformer (OcTr), for 3D object detection. We firstly devise an octree-based learnable sparse pattern, i.e. OctAttn, which meticulously and efficiently encodes point clouds of scenes as shown in Fig. 1. The OctAttn module constructs a feature pyramid by gathering and applies self-attention to the top level of the feature pyramid to select the most relevant tokens, which are deemed as the octants to be divided in the subsequent. When propagating to the level below, the key/value inputs are restricted by the" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 547, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 192 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 192 + ], + "type": "text", + "content": "octants from the top. Through recursively conducting this process, OctAttn captures rich global context features by a global receptive field in a coarse-to-fine manner while reducing the quadratic complexity of vanilla self-attention to the linear complexity. In addition, for better foreground perception, we propose a hybrid positional embedding, which consists of the semantic-aware positional embedding and attention mask, to fully exploit geometry and semantic clues. Thanks to the designs above, OcTr delivers a competitive trade-off between accuracy and efficiency." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 317, + 192, + 503, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 192, + 503, + 204 + ], + "spans": [ + { + "bbox": [ + 317, + 192, + 503, + 204 + ], + "type": "text", + "content": "Our contribution is summarized in three-fold:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 207, + 545, + 328 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 313, + 207, + 545, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 207, + 545, + 243 + ], + "spans": [ + { + "bbox": [ + 313, + 207, + 545, + 243 + ], + "type": "text", + "content": "1. We propose OcTr for voxel-based 3D object detection, which efficiently learns enhanced representations by a global receptive field with rich contexts." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 250, + 545, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 250, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 313, + 250, + 545, + 285 + ], + "type": "text", + "content": "2. We propose an octree-based learnable attention sparsification scheme (OctAttn) and a hybrid positional embedding combining geometry and semantics." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 292, + 545, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 292, + 545, + 328 + ], + "spans": [ + { + "bbox": [ + 313, + 292, + 545, + 328 + ], + "type": "text", + "content": "3. We carry out experiments on the Waymo Open Dataset (WOD) and the KITTI dataset and report state-of-the-art performance with significant gains on far objects." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 340, + 392, + 353 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 340, + 392, + 353 + ], + "spans": [ + { + "bbox": [ + 306, + 340, + 392, + 353 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 360, + 512, + 373 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 360, + 512, + 373 + ], + "spans": [ + { + "bbox": [ + 306, + 360, + 512, + 373 + ], + "type": "text", + "content": "2.1. 3D Object Detection from Point Clouds" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 379, + 545, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 379, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 306, + 379, + 545, + 403 + ], + "type": "text", + "content": "There exist two prevailing point-cloud representations in 3D object detection, i.e. point-based and voxel-based." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 403, + 547, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 403, + 547, + 545 + ], + "spans": [ + { + "bbox": [ + 304, + 403, + 547, + 545 + ], + "type": "text", + "content": "The point-based methods [31,40,42,60] directly process raw point clouds in the irregular 3D space. As a pioneering attempt, F-Pointnet [32] employs instance segmentation in frustums to extract proposals. VoteNet [31] clusters objects from the surface in a deep Hough voting manner. PointRCNN [40] generates 3D RoIs with foreground segmentation and applies an RCNN-style [35] two-stage refinement. Different from PointRCNN, some existing methods [3,53,59] build a lightweight and efficient single stage 3D object detection framework. However, the current point-based methods still suffer from a large computation burden, which is not suitable for large-scale point cloud scenes." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 546, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 547, + 714 + ], + "type": "text", + "content": "The voxel-based ones [4,6,13,52,55,62,63] conduct voxelization on entire point clouds to construct regular grids. VoxelNet [63] exploits the voxel feature encoding layer with 3D convolutions to extract the feature of each voxel. SECOND [52] improves the model with sparse 3D convolutions, significantly increasing the speed of both training and inference. Pointpillars [18] compacts point clouds into vertical columns and encodes them with 2D CNNs. Several recent methods [13, 38, 54] also explore merging point-based and voxel-based networks into one framework for complementary features from different representations of point clouds. Unfortunately, they all use small convolution kernels with limited receptive fields, which are not competent to capture global context that is important to 3D detection." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5167" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 73, + 542, + 194 + ], + "blocks": [ + { + "bbox": [ + 53, + 73, + 542, + 194 + ], + "lines": [ + { + "bbox": [ + 53, + 73, + 542, + 194 + ], + "spans": [ + { + "bbox": [ + 53, + 73, + 542, + 194 + ], + "type": "image", + "image_path": "e9ca629127423699a851f636446a042a215fcb31ee294f13960afbf86af50a5f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 134, + 207, + 457, + 218 + ], + "lines": [ + { + "bbox": [ + 134, + 207, + 457, + 218 + ], + "spans": [ + { + "bbox": [ + 134, + 207, + 457, + 218 + ], + "type": "text", + "content": "Figure 2. Framework overview of the proposed Octree-based Transformer (OcTr) model." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 239, + 190, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 239, + 190, + 251 + ], + "spans": [ + { + "bbox": [ + 47, + 239, + 190, + 251 + ], + "type": "text", + "content": "2.2. Transformer in 3D Vision" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 258, + 287, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 258, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 46, + 258, + 287, + 376 + ], + "type": "text", + "content": "Inspired by the great success of the self-attention mechanism in NLP [47] and CV [8, 23], Transformers have been adapted to 3D vision for their ability to capture long-range dependencies. For instance, Point Transformer [61] brings in vector attention that modulates individual feature channels for point cloud classification and segmentation; PCT [11] presents offset-attention with the implicit Laplace operator and normalization refinement which is more suitable for point cloud learning. To address the high latency, some methods [30, 57] adopt voxels or patches for acceleration." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 377, + 288, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 377, + 288, + 594 + ], + "spans": [ + { + "bbox": [ + 46, + 377, + 288, + 594 + ], + "type": "text", + "content": "For 3D detection, 3DETR [28] treats and predicts bounding boxes as sequences in an end-to-end manner. CT3D [37] leverages a channel-wise Transformer architecture to refine the RoI head. To learn context-aware representations, some studies [9,12,24,29] introduce Transformers into a point- or voxel-based encoder. Pointformer [29] stacks local, global and local-global Transformers based on the point-based encoder; VoTr [24] exploits dilated attention with fast query to enlarge receptive fields; VoxSet [12] builds an induced point set as proxies of global context and applies point-to-point translation using voxels as mediums; and SST [9] embraces single strides without down-sampling and conducts window attention combining with its shifted version. Even though they all expand receptive fields by diverse Transformer variants, global context is not adequately involved or efficiently utilized. In contrast, we propose Octree-based Transformer (OcTr) for voxel-based 3D object detection, achieving a true global receptive field that balances accuracy and efficiency." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 603, + 103, + 615 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 603, + 103, + 615 + ], + "spans": [ + { + "bbox": [ + 47, + 603, + 103, + 615 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 623, + 125, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 623, + 125, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 623, + 125, + 635 + ], + "type": "text", + "content": "3.1. Framework" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 641, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 713 + ], + "type": "text", + "content": "This sub-section describes the overall framework of the proposed OcTr model as shown in Fig. 2. Specifically, as inspired by [51], we first voxelize the point cloud into regular grids and adopt the sparse 3D convolution for patch embedding, where the girds are regarded as the \"tokens\" and are passed through the Octree Transformer Blocks" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 239, + 547, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 239, + 547, + 395 + ], + "spans": [ + { + "bbox": [ + 304, + 239, + 547, + 395 + ], + "type": "text", + "content": "(OTB). Compared with the vanilla Transformer block, the self-attention module is substituted by the proposed octree-attention " + }, + { + "bbox": [ + 304, + 239, + 547, + 395 + ], + "type": "inline_equation", + "content": "OctAttn" + }, + { + "bbox": [ + 304, + 239, + 547, + 395 + ], + "type": "text", + "content": ", which encodes global context in a more efficient way. After applying a hybrid semantic embedding on multi-scale features, we sequentially stack two OTBs, tailed by a down-sampling layer. The voxel features are then projected into the BEV view by point-wise convolutions and are passed through a multi-scale dense 2D backbone. Ultimately, an anchor-based or anchor-free RPN head is used for 3D proposal generation, and the RoI head is optional for refinement. Note that our OcTr can be adopted for most of the voxel-based detection frameworks by simply altering 3D backbones." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 401, + 427, + 412 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 401, + 427, + 412 + ], + "spans": [ + { + "bbox": [ + 306, + 401, + 427, + 412 + ], + "type": "text", + "content": "3.2. Self-attention Revisit" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 419, + 547, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 419, + 547, + 479 + ], + "spans": [ + { + "bbox": [ + 304, + 419, + 547, + 479 + ], + "type": "text", + "content": "According to [47], Transformer encoder blocks typically include a multi-head self-attention (MHSA) mechanism, a feed-forward network (FFN), a normalization function, and the residual connections [14]. Given an input sequence " + }, + { + "bbox": [ + 304, + 419, + 547, + 479 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 304, + 419, + 547, + 479 + ], + "type": "text", + "content": ", the principle part of MHSA is formulated as" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 315, + 484, + 545, + 517 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 484, + 545, + 517 + ], + "spans": [ + { + "bbox": [ + 315, + 484, + 545, + 517 + ], + "type": "interline_equation", + "content": "\\operatorname {M H S A} (X) = \\sum_ {h = 1} ^ {H} W _ {h} \\left[ \\sigma \\left(\\frac {X W _ {q} W _ {k} ^ {T} X ^ {T}}{\\sqrt {d}}\\right) \\cdot X W _ {v} \\right], \\tag {1}", + "image_path": "25d91144ad12ffc4fdf5a392db771660989482930142928f9681d89acbd04aea.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 521, + 547, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 547, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 547, + 605 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 521, + 547, + 605 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 304, + 521, + 547, + 605 + ], + "type": "text", + "content": " denotes the index of the head, and " + }, + { + "bbox": [ + 304, + 521, + 547, + 605 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 304, + 521, + 547, + 605 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 521, + 547, + 605 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 304, + 521, + 547, + 605 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 521, + 547, + 605 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 304, + 521, + 547, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 521, + 547, + 605 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 521, + 547, + 605 + ], + "type": "text", + "content": " are the amount of heads, softmax function, learnable weight and feature dimension, respectively. The subscripts of " + }, + { + "bbox": [ + 304, + 521, + 547, + 605 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 304, + 521, + 547, + 605 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 521, + 547, + 605 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 521, + 547, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 521, + 547, + 605 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 304, + 521, + 547, + 605 + ], + "type": "text", + "content": " indicate query, key and value. The inputs and outputs of the MHSA module are connected by residual connectors and normalization layers. The MLP-based FFN connects its inputs/outputs in a similar manner." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 611, + 408, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 611, + 408, + 623 + ], + "spans": [ + { + "bbox": [ + 306, + 611, + 408, + 623 + ], + "type": "text", + "content": "3.3. Octree Attention" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "type": "text", + "content": "An octree is a multi-scale asymmetric and efficient representation for unstructured 3D data such as point clouds. To build an octree for the input point cloud, we recursively sub-divide it in the breadth-first order until the pre-defined octree depth is reached. Whether to sub-divide an octant is determined by the occupancy [36, 48], the surface approximation [44, 49] or a learning algorithm [26]." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5168" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 73, + 347, + 233 + ], + "blocks": [ + { + "bbox": [ + 52, + 73, + 347, + 233 + ], + "lines": [ + { + "bbox": [ + 52, + 73, + 347, + 233 + ], + "spans": [ + { + "bbox": [ + 52, + 73, + 347, + 233 + ], + "type": "image", + "image_path": "96bc99e00865db3b240e60cadeb29d51d8cd8d5d934e9de16e6f190fd171ed3a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 247, + 547, + 291 + ], + "lines": [ + { + "bbox": [ + 46, + 247, + 547, + 291 + ], + "spans": [ + { + "bbox": [ + 46, + 247, + 547, + 291 + ], + "type": "text", + "content": "Figure 3. Octree Attention (OctAttn). The OctAttn module constructs a dynamic octree on a hierarchical feature pyramid with a partitive criterion of attention scores. For illustration, an octree representation is individually pruned from the pyramid for each grid in " + }, + { + "bbox": [ + 46, + 247, + 547, + 291 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 46, + 247, + 547, + 291 + ], + "type": "text", + "content": ". For example, the blue grids are the divided octants for the grids with stars. The detailed structures of OcTr-Top and OcTr-Base are shown at the right side." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 351, + 76, + 446, + 230 + ], + "blocks": [ + { + "bbox": [ + 351, + 76, + 446, + 230 + ], + "lines": [ + { + "bbox": [ + 351, + 76, + 446, + 230 + ], + "spans": [ + { + "bbox": [ + 351, + 76, + 446, + 230 + ], + "type": "image", + "image_path": "563e90e76e17443f4be4476c9793a81ee2db92313c1c830bfac3a3aa51bc6940.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 449, + 78, + 543, + 230 + ], + "blocks": [ + { + "bbox": [ + 449, + 78, + 543, + 230 + ], + "lines": [ + { + "bbox": [ + 449, + 78, + 543, + 230 + ], + "spans": [ + { + "bbox": [ + 449, + 78, + 543, + 230 + ], + "type": "image", + "image_path": "034cdb1500198579c1ba7e2d834de7809406d207fea9c28dee8523b9c58421b1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 300, + 287, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 300, + 287, + 418 + ], + "spans": [ + { + "bbox": [ + 46, + 300, + 287, + 418 + ], + "type": "text", + "content": "Despite of being resource-friendly, the octree representation is discrete and non-differentiable, which motivates us to ameliorate previous cumbersome pre-processing division by introducing a novel octree-based attention mechanism. As depicted in [21, 45], the attention matrices calculated by self-attention imply the relevance of input tokens and guide feature selection. We thus prune the dense attention matrices of the multi-scale feature pyramid to sparse octree attention in an adaptive and parallel manner, namely OctAttn." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 422, + 287, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 422, + 287, + 482 + ], + "spans": [ + { + "bbox": [ + 47, + 422, + 287, + 482 + ], + "type": "text", + "content": "As shown in Fig. 3, let the output feature map and coordinates of convolutional patch embedding be " + }, + { + "bbox": [ + 47, + 422, + 287, + 482 + ], + "type": "inline_equation", + "content": "F_0 \\in \\mathbb{R}^{M \\times d}" + }, + { + "bbox": [ + 47, + 422, + 287, + 482 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 422, + 287, + 482 + ], + "type": "inline_equation", + "content": "I_0 \\in \\mathbb{R}^{M \\times 3}" + }, + { + "bbox": [ + 47, + 422, + 287, + 482 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 47, + 422, + 287, + 482 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 47, + 422, + 287, + 482 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 422, + 287, + 482 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 47, + 422, + 287, + 482 + ], + "type": "text", + "content": " indicate the amount of the nonempty grids in a batch and feature dimension, respectively. Based on " + }, + { + "bbox": [ + 47, + 422, + 287, + 482 + ], + "type": "inline_equation", + "content": "F_0" + }, + { + "bbox": [ + 47, + 422, + 287, + 482 + ], + "type": "text", + "content": ", we generate a multi-scale feature pyramid as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 107, + 498, + 287, + 512 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 498, + 287, + 512 + ], + "spans": [ + { + "bbox": [ + 107, + 498, + 287, + 512 + ], + "type": "interline_equation", + "content": "C = \\left\\{F _ {n}, I _ {n} \\right\\} ^ {N}, n \\in [ 0, N), \\tag {2}", + "image_path": "f3736b90b0b9116ccb75796f4d7aadb9ee419ae21612b86396d56481ffc90481.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 522, + 75, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 522, + 75, + 533 + ], + "spans": [ + { + "bbox": [ + 47, + 522, + 75, + 533 + ], + "type": "text", + "content": "where" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 85, + 548, + 287, + 571 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 548, + 287, + 571 + ], + "spans": [ + { + "bbox": [ + 85, + 548, + 287, + 571 + ], + "type": "interline_equation", + "content": "I _ {n} = \\left\\lfloor \\frac {I _ {0}}{2 ^ {n}} \\right\\rfloor , F _ {n} = \\operatorname {B N} \\left(S _ {\\max } \\left(F _ {0}, I _ {n}\\right)\\right), \\tag {3}", + "image_path": "8013afcfec42a3d47127a8906026ee7228d73b206ee2b1968b8e19fb47805ead.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 580, + 287, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 580, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 580, + 287, + 616 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 580, + 287, + 616 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 580, + 287, + 616 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 580, + 287, + 616 + ], + "type": "text", + "content": ", BN and " + }, + { + "bbox": [ + 46, + 580, + 287, + 616 + ], + "type": "inline_equation", + "content": "S_{max}" + }, + { + "bbox": [ + 46, + 580, + 287, + 616 + ], + "type": "text", + "content": " are the index of the level of the multiscale feature pyramid, the height of the pyramid, batch normalization [16] and the max scatter function, respectively." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 618, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 287, + 715 + ], + "type": "text", + "content": "The pruning begins from the top of the pyramid. The top feature map, i.e. " + }, + { + "bbox": [ + 46, + 618, + 287, + 715 + ], + "type": "inline_equation", + "content": "F_{N - 1}" + }, + { + "bbox": [ + 46, + 618, + 287, + 715 + ], + "type": "text", + "content": ", is reorganized to dense input tokens " + }, + { + "bbox": [ + 46, + 618, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\bar{F}_{N - 1} \\in \\mathbb{R}^{B \\times m_{N - 1} \\times d}" + }, + { + "bbox": [ + 46, + 618, + 287, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 618, + 287, + 715 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 46, + 618, + 287, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 618, + 287, + 715 + ], + "type": "inline_equation", + "content": "m_{N - 1}" + }, + { + "bbox": [ + 46, + 618, + 287, + 715 + ], + "type": "text", + "content": " are the batch size and maximum number of non-empty voxels per batch, respectively. The voxel is padded with 0 if it is empty. As shown in Eq. (4), the MHSA takes " + }, + { + "bbox": [ + 46, + 618, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\bar{F}_{N - 1}" + }, + { + "bbox": [ + 46, + 618, + 287, + 715 + ], + "type": "text", + "content": " as input and outputs the attention score matrices " + }, + { + "bbox": [ + 46, + 618, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{N - 1} \\in \\mathbb{R}^{B \\times m_{N - 1} \\times m_{N - 1}}" + }, + { + "bbox": [ + 46, + 618, + 287, + 715 + ], + "type": "text", + "content": " and attentive features " + }, + { + "bbox": [ + 46, + 618, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\bar{F}_{N - 1}'" + }, + { + "bbox": [ + 46, + 618, + 287, + 715 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 311, + 319, + 545, + 398 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 319, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 311, + 319, + 545, + 398 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {A} _ {N - 1} = \\sum_ {h = 1} ^ {H} \\sigma (\\frac {\\bar {F} _ {N - 1} W _ {q} W _ {k} ^ {T} \\bar {F} _ {N - 1} ^ {T}}{\\sqrt {d}}), \\\\ \\bar {F} _ {N - 1} ^ {\\prime} = \\sum_ {h = 1} ^ {H} W _ {h} \\left[ \\sigma \\left(\\frac {\\bar {F} _ {N - 1} W _ {q} W _ {k} ^ {T} \\bar {F} _ {N - 1} ^ {T}}{\\sqrt {d}}\\right) \\cdot \\bar {F} _ {N - 1} W _ {v} \\right]. \\tag {4} \\\\ \\end{array}", + "image_path": "533612fb7ba6a97b2b6953e32327e75a934788440ed1d97981e63364edd7b867.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 402, + 545, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 402, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 305, + 402, + 545, + 437 + ], + "type": "text", + "content": "For each query token, we select the top " + }, + { + "bbox": [ + 305, + 402, + 545, + 437 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 305, + 402, + 545, + 437 + ], + "type": "text", + "content": " attention scores as its most relevant token group in a row-wise way, denoted by " + }, + { + "bbox": [ + 305, + 402, + 545, + 437 + ], + "type": "inline_equation", + "content": "O_{N - 1}\\in \\mathbb{Z}^{B\\times m_{N - 1}\\times k}" + }, + { + "bbox": [ + 305, + 402, + 545, + 437 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 437, + 545, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 437, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 305, + 437, + 545, + 533 + ], + "type": "text", + "content": "When propagating from top to bottom through the pyramid and reaching the " + }, + { + "bbox": [ + 305, + 437, + 545, + 533 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 305, + 437, + 545, + 533 + ], + "type": "text", + "content": "-th level, we uniformly sample limited " + }, + { + "bbox": [ + 305, + 437, + 545, + 533 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 305, + 437, + 545, + 533 + ], + "type": "text", + "content": " attending octants from the selected regions with features in " + }, + { + "bbox": [ + 305, + 437, + 545, + 533 + ], + "type": "inline_equation", + "content": "\\bar{F}_n" + }, + { + "bbox": [ + 305, + 437, + 545, + 533 + ], + "type": "text", + "content": " and topk indices in " + }, + { + "bbox": [ + 305, + 437, + 545, + 533 + ], + "type": "inline_equation", + "content": "O_{n + 1}" + }, + { + "bbox": [ + 305, + 437, + 545, + 533 + ], + "type": "text", + "content": ". We conduct cross-attention instead of self-attention with the dense query sequence of the " + }, + { + "bbox": [ + 305, + 437, + 545, + 533 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 305, + 437, + 545, + 533 + ], + "type": "text", + "content": "-th level, " + }, + { + "bbox": [ + 305, + 437, + 545, + 533 + ], + "type": "inline_equation", + "content": "\\bar{Q}_n\\in \\mathbb{R}^{B\\times m_n\\times d}" + }, + { + "bbox": [ + 305, + 437, + 545, + 533 + ], + "type": "text", + "content": ", and the compact sampled key/value sequence of the " + }, + { + "bbox": [ + 305, + 437, + 545, + 533 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 305, + 437, + 545, + 533 + ], + "type": "text", + "content": "-th level, " + }, + { + "bbox": [ + 305, + 437, + 545, + 533 + ], + "type": "inline_equation", + "content": "\\bar{K}_n / \\bar{V}_n\\in \\mathbb{R}^{B\\times m_n\\times K\\times d}" + }, + { + "bbox": [ + 305, + 437, + 545, + 533 + ], + "type": "text", + "content": ", which is formulated as below" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 334, + 553, + 545, + 620 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 553, + 545, + 620 + ], + "spans": [ + { + "bbox": [ + 334, + 553, + 545, + 620 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {A} _ {n} = \\sum_ {h = 1} ^ {H} \\sigma \\left(\\frac {\\bar {Q} _ {n} W _ {q} W _ {k} ^ {T} \\bar {K} _ {n} ^ {T}}{\\sqrt {d}}\\right), \\tag {5} \\\\ \\bar {F} _ {n} ^ {\\prime} = \\sum_ {h = 1} ^ {H} W _ {h} \\left[ \\sigma \\left(\\frac {\\bar {Q} _ {n} W _ {q} W _ {k} ^ {T} \\bar {K} _ {n} ^ {T}}{\\sqrt {d}}\\right) \\cdot \\bar {V} _ {n} W _ {v} \\right]. \\\\ \\end{array}", + "image_path": "3953ba2e6e251e2ff354f496c341c41a5da8d1a540b0234488e7e33af6ee039f.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 629, + 545, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 629, + 545, + 666 + ], + "spans": [ + { + "bbox": [ + 305, + 629, + 545, + 666 + ], + "type": "text", + "content": "Backing off the sampling, this can be treated as an attention mask on self-attention matrices. The above process is run recursively until reaching the bottom level of the pyramid." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 666, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 666, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 666, + 545, + 713 + ], + "type": "text", + "content": "Furthermore, as the top " + }, + { + "bbox": [ + 305, + 666, + 545, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 305, + 666, + 545, + 713 + ], + "type": "text", + "content": " selection is a hard decision that disables gradient back-propagation, we adopt the Gumbeltop " + }, + { + "bbox": [ + 305, + 666, + 545, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 305, + 666, + 545, + 713 + ], + "type": "text", + "content": " technique [17] to perform a differentiable and continuous approximation by replacing the vanilla top " + }, + { + "bbox": [ + 305, + 666, + 545, + 713 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 305, + 666, + 545, + 713 + ], + "type": "text", + "content": " selection." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5169" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "content": "The normalized scores used in top " + }, + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "content": " are derived from the distribution in Eq. (6) during training, maintaining the original ones during inference. " + }, + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "inline_equation", + "content": "g,\\tau ,m_{n}" + }, + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "content": " denote the noise sampled from the Gumbel distribution, the temperature and the amount of non-empty voxels in layer " + }, + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 85, + 142, + 287, + 181 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 142, + 287, + 181 + ], + "spans": [ + { + "bbox": [ + 85, + 142, + 287, + 181 + ], + "type": "interline_equation", + "content": "p _ {i} = \\frac {\\exp \\left(\\left(\\mathcal {A} _ {n} ^ {i} + g _ {i}\\right) / \\tau\\right)}{\\sum_ {i} \\exp \\left(\\left(\\mathcal {A} _ {n} ^ {i} + g _ {i}\\right) / \\tau\\right)} \\in [ 0, 1 ]. \\tag {6}", + "image_path": "6a0ae1d1f127b99f4d154b0949be57e20916701e10447c5a292100f5670a02be.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 186, + 287, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 186, + 287, + 245 + ], + "spans": [ + { + "bbox": [ + 46, + 186, + 287, + 245 + ], + "type": "text", + "content": "In order to leverage the multi-scale features in distinct spatial shapes, we concatenate them by upsampling, which is implemented by inverse indices of the scatter function, followed by a linear projection layer for aligning the input feature dimension." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 247, + 287, + 318 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 247, + 287, + 318 + ], + "spans": [ + { + "bbox": [ + 46, + 247, + 287, + 318 + ], + "type": "text", + "content": "As the local context is generally critical for object detection, inspired by [7], we additionally introduce a Locally enhanced Positional Embedding (LePE) which enables local neighbor interactions on the value sequence. With submanifold sparse convolutions, we replace the residual connections in the attention mechanism with LePE." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 319, + 212, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 319, + 212, + 331 + ], + "spans": [ + { + "bbox": [ + 59, + 319, + 212, + 331 + ], + "type": "text", + "content": "Finally, OTB is formulated as follows" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 71, + 350, + 287, + 372 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 350, + 287, + 372 + ], + "spans": [ + { + "bbox": [ + 71, + 350, + 287, + 372 + ], + "type": "interline_equation", + "content": "\\tilde {F} = \\operatorname {F C} \\left(\\left\\{F _ {N - 1} ^ {\\prime} \\mid F _ {N - 2} ^ {\\prime} \\right| \\dots \\left| F _ {0} ^ {\\prime} \\right\\}\\right) + \\operatorname {L e P E} \\left(F _ {0}\\right), \\tag {7}", + "image_path": "82275fc4790352e6c55feaefb53f0987d142ed1c3d052a78aceafbc51180e16a.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 73, + 369, + 182, + 383 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 369, + 182, + 383 + ], + "spans": [ + { + "bbox": [ + 73, + 369, + 182, + 383 + ], + "type": "interline_equation", + "content": "\\tilde {F} ^ {\\prime} = \\mathrm {B N} (\\mathrm {F F N} (\\tilde {F})) + \\tilde {F},", + "image_path": "1a34bd1fcc1846d51e0eb97b41712136e0f6efdd442b7628fdecc718260cc0af.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 395, + 287, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 395, + 287, + 431 + ], + "spans": [ + { + "bbox": [ + 47, + 395, + 287, + 431 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 395, + 287, + 431 + ], + "type": "inline_equation", + "content": "F_{n}^{\\prime}\\in \\mathbb{R}^{m_0\\times d}" + }, + { + "bbox": [ + 47, + 395, + 287, + 431 + ], + "type": "text", + "content": " denotes the compact tensor of the upsampled " + }, + { + "bbox": [ + 47, + 395, + 287, + 431 + ], + "type": "inline_equation", + "content": "\\bar{F}_n^\\prime" + }, + { + "bbox": [ + 47, + 395, + 287, + 431 + ], + "type": "text", + "content": " , | indicates concatenation, and FC denotes the fully-connected layer." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 432, + 287, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 432, + 287, + 455 + ], + "spans": [ + { + "bbox": [ + 47, + 432, + 287, + 455 + ], + "type": "text", + "content": "Besides, we analyze the time complexity of the octree attention as below" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 79, + 463, + 287, + 521 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 463, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 79, + 463, + 287, + 521 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {O} \\left(\\left(\\frac {M}{\\omega^ {N - 1}}\\right) ^ {2} + \\sum_ {n = 0} ^ {N - 2} \\frac {K M}{\\omega^ {n}}\\right) \\tag {8} \\\\ = \\mathcal {O} \\left(\\left(\\frac {M}{\\omega^ {N - 1}}\\right) ^ {2} + \\frac {\\omega}{\\omega - 1} K M \\left(1 - \\omega^ {1 - N}\\right)\\right), \\\\ \\end{array}", + "image_path": "0e1063ca2d9ef9dd004081d26149d8261509b71dab072322e66b2a00a448c433.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 530, + 287, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 530, + 287, + 554 + ], + "spans": [ + { + "bbox": [ + 47, + 530, + 287, + 554 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 530, + 287, + 554 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 47, + 530, + 287, + 554 + ], + "type": "text", + "content": " is the average down-sampling ratio in the sparse voxel representation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 563, + 219, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 563, + 219, + 575 + ], + "spans": [ + { + "bbox": [ + 47, + 563, + 219, + 575 + ], + "type": "text", + "content": "3.4. Semantic Positional Embedding" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 582, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 287, + 653 + ], + "type": "text", + "content": "Due to the large proportion of background grids in point clouds, the attention matrices are dominated by background grid pairs, leading to a sub-optimal solution. To fully leverage the local 3D shape patterns and original voxel coordinates, we propose a hybrid positional embedding to capture both the geometry and semantic clues as displayed in Fig. 4." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "text", + "content": "Specifically, we first segment foreground grids using the supervision from the ground truth. Segmentation scores are predicted by a sub-manifold sparse convolution branch with a sigmoid function; and the focal loss [22] is applied to balance the foreground and background." + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 310, + 71, + 542, + 228 + ], + "blocks": [ + { + "bbox": [ + 310, + 71, + 542, + 228 + ], + "lines": [ + { + "bbox": [ + 310, + 71, + 542, + 228 + ], + "spans": [ + { + "bbox": [ + 310, + 71, + 542, + 228 + ], + "type": "image", + "image_path": "899be8311de29cb25bdd3d9fb85ee5bea664f6148bd5813802b8e45d11e05ebc.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 238, + 545, + 282 + ], + "lines": [ + { + "bbox": [ + 305, + 238, + 545, + 282 + ], + "spans": [ + { + "bbox": [ + 305, + 238, + 545, + 282 + ], + "type": "text", + "content": "Figure 4. Illustration of semantic positional embedding, where an extra foreground segmentation branch is adopted with supervision and semantic scores are concatenated as absolute positional embeddings and serve as relative masks on attention matrices." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 289, + 545, + 325 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 289, + 545, + 325 + ], + "spans": [ + { + "bbox": [ + 305, + 289, + 545, + 325 + ], + "type": "text", + "content": "We concatenate the center coordinate " + }, + { + "bbox": [ + 305, + 289, + 545, + 325 + ], + "type": "inline_equation", + "content": "(x, y, z)" + }, + { + "bbox": [ + 305, + 289, + 545, + 325 + ], + "type": "text", + "content": " and the semantic score with the feature " + }, + { + "bbox": [ + 305, + 289, + 545, + 325 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 305, + 289, + 545, + 325 + ], + "type": "text", + "content": " in a grid-wise manner, followed by a linear projection without the bias as below" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 332, + 545, + 361 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 332, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 317, + 332, + 545, + 361 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\operatorname {S A P E} (X) = \\operatorname {F C} _ {d + 4 \\rightarrow d} (\\{x, y, z, \\text {s c o r e} | f \\}) \\tag {9} \\\\ = \\operatorname {F C} _ {d \\rightarrow d} (f) + \\operatorname {F C} _ {4 \\rightarrow d} (x, y, z, s c o r e). \\\\ \\end{array}", + "image_path": "256a0c66ed156583c3d6f452747094e80dd983f2010f0b58164a87b9fd8c6b2b.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 369, + 545, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 369, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 305, + 369, + 545, + 441 + ], + "type": "text", + "content": "Eq. (9) is equivalent to the absolute positional embedding (APE), thus being denoted as the Semantic APE (SAPE). The scatter function such as the mean, max and batch normalization naturally provide position and semantic information of the downsampled voxel grids, making it applicable for the multi-scale feature pyramid." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 441, + 545, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 441, + 545, + 535 + ], + "spans": [ + { + "bbox": [ + 305, + 441, + 545, + 535 + ], + "type": "text", + "content": "Besides the semantic clues implicitly used in SAPE, we employ the Semantic Attention Mask (SAM) based on the segmentation scores and we mask the attention matrices to address correlations between inferior queries and superior keys in a simple yet effective way. Given a scalar attention matrix " + }, + { + "bbox": [ + 305, + 441, + 545, + 535 + ], + "type": "inline_equation", + "content": "\\mathcal{A} \\in \\mathbb{R}^{N_q \\times N_k}" + }, + { + "bbox": [ + 305, + 441, + 545, + 535 + ], + "type": "text", + "content": " before softmax and the segmentation scores of query and key/value " + }, + { + "bbox": [ + 305, + 441, + 545, + 535 + ], + "type": "inline_equation", + "content": "S_q \\in \\mathbb{R}^{N_q}" + }, + { + "bbox": [ + 305, + 441, + 545, + 535 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 441, + 545, + 535 + ], + "type": "inline_equation", + "content": "S_k \\in \\mathbb{R}^{N_k}" + }, + { + "bbox": [ + 305, + 441, + 545, + 535 + ], + "type": "text", + "content": ", we formulate the attention matrix after softmax as" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 318, + 544, + 545, + 559 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 544, + 545, + 559 + ], + "spans": [ + { + "bbox": [ + 318, + 544, + 545, + 559 + ], + "type": "interline_equation", + "content": "\\mathcal {A} ^ {\\prime} = \\sigma (- \\Gamma \\cdot [ 1 - (S _ {q} \\geq \\delta_ {q}) (S _ {k} \\geq \\delta_ {k}) ^ {T} ] + \\mathcal {A}), \\tag {10}", + "image_path": "f1669b194ebf1d6632e2bc12caa9a97a5fa1d950e041707d936678a450c563be.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 563, + 545, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 563, + 545, + 598 + ], + "spans": [ + { + "bbox": [ + 305, + 563, + 545, + 598 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 563, + 545, + 598 + ], + "type": "inline_equation", + "content": "N_{q},\\delta_{q},N_{k},\\delta_{k},\\sigma" + }, + { + "bbox": [ + 305, + 563, + 545, + 598 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 563, + 545, + 598 + ], + "type": "inline_equation", + "content": "\\Gamma" + }, + { + "bbox": [ + 305, + 563, + 545, + 598 + ], + "type": "text", + "content": " are the length and threshold of the query sequence, length and threshold of the key sequence, softmax function and infinite scalar, respectively." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 599, + 544, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 599, + 544, + 610 + ], + "spans": [ + { + "bbox": [ + 317, + 599, + 544, + 610 + ], + "type": "text", + "content": "Finally, we broadcast the semantic mask " + }, + { + "bbox": [ + 317, + 599, + 544, + 610 + ], + "type": "inline_equation", + "content": "\\mathcal{A}'" + }, + { + "bbox": [ + 317, + 599, + 544, + 610 + ], + "type": "text", + "content": " to all heads." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 306, + 621, + 386, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 621, + 386, + 635 + ], + "spans": [ + { + "bbox": [ + 306, + 621, + 386, + 635 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 305, + 641, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 641, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 641, + 545, + 713 + ], + "type": "text", + "content": "We evaluate the proposed OcTr network on the Waymo Open Dataset (WOD) and KITTI dataset, both of which are popular in 3D object detection. In this section, we introduce the benchmarks and implementation details, make comparison to the previous state-of-the-art counterparts, and ablate the key designs of OcTr." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5170" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 52, + 71, + 542, + 235 + ], + "blocks": [ + { + "bbox": [ + 52, + 71, + 542, + 235 + ], + "lines": [ + { + "bbox": [ + 52, + 71, + 542, + 235 + ], + "spans": [ + { + "bbox": [ + 52, + 71, + 542, + 235 + ], + "type": "table", + "html": "
ModelVehicle (L1) \nmAP/mAPHVehicle (L2) \nmAP/mAPHPedes. (L1) \nmAP/mAPHPedes. (L2) \nmAP/mAPHCyclist (L1) \nmAP/mAPHCyclist (L2) \nmAP/mAPH
SECOND [52]70.96/70.3462.58/62.0265.23/54.2457.22/47.4957.13/55.6254.97/53.53
PointPillar [18]70.43/69.8362.18/61.6466.21/46.3258.18/40.6455.26/51.7553.18/49.80
PartA²Net [41]74.82/74.3265.88/65.4271.76/63.6462.53/55.3067.35/66.1565.05/63.89
PVRCNN [38]75.41/74.7467.44/66.8071.98/61.2463.70/53.9565.88/64.2563.39/61.82
CenterPoint [55]71.33/70.7663.16/62.6572.09/65.4964.27/58.2368.68/67.3966.11/64.87
LiDAR-RCNN [20]73.5/73.064.7/64.271.2/58.763.1/51.768.6/66.966.1/64.4
Voxel-RCNN [6]75.59/-66.59/--/--/--/--/-
PVRCNN++ [39]77.82/77.3269.07/68.6277.99/71.3669.92/63.7471.80/70.7169.31/68.26
SST† [9]76.22/75.7968.04/67.6481.39/74.0572.82/65.93-/--/-
PDV [15]76.85/76.3369.30/68.8174.19/65.9665.85/58.2868.71/67.5566.49/65.36
Ours78.12/77.6369.79/69.3480.76/74.3972.48/66.5272.58/71.5069.93/68.90
", + "image_path": "aa326a65166577e124eeb1665378696b5235ce62f1edff364bf70673aa92202a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 52, + 284, + 282, + 487 + ], + "blocks": [ + { + "bbox": [ + 46, + 243, + 547, + 278 + ], + "lines": [ + { + "bbox": [ + 46, + 243, + 547, + 278 + ], + "spans": [ + { + "bbox": [ + 46, + 243, + 547, + 278 + ], + "type": "text", + "content": "Table 1. Performance on WOD with 202 validation sequences for vehicle (IoU=0.7), pedestrian (IoU=0.5) and cyclist (IoU=0.5), using 20% samples for training. All the results are achieved by the models simultaneously trained for 3 classes on single frames, except the ones of the model marked by " + }, + { + "bbox": [ + 46, + 243, + 547, + 278 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 46, + 243, + 547, + 278 + ], + "type": "text", + "content": ", which is only trained for a single class. Refer to Supp. B for the results trained with 100% samples." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 52, + 284, + 282, + 487 + ], + "lines": [ + { + "bbox": [ + 52, + 284, + 282, + 487 + ], + "spans": [ + { + "bbox": [ + 52, + 284, + 282, + 487 + ], + "type": "table", + "html": "
ModelmAP3D(L1)@ Vehicle
Overall0-30m30m-50m50m-inf
PV-RCNN [38]70.3091.9269.2142.17
Voxel-RCNN [6]75.5992.4974.0953.15
VoTR-TSD [24]74.9592.2873.3651.09
CT3D [37]76.3092.5175.0755.36
Pyramid_PV [25]76.3092.6774.9154.54
PDV [15]76.8593.1375.4954.75
VoxSeT [12]77.8292.7877.2154.41
Ours78.8292.9977.6658.02
ModelmAP3D(L2)@ Vehicle
Overall0-30m30-50m50m-inf
PV-RCNN [38]65.3691.5865.1336.46
Voxel-RCNN [6]66.5991.7467.8940.80
CT3D [37]69.0491.7668.9342.60
PDV [15]69.3092.4169.3642.16
VoxSeT [12]70.2192.0570.1043.20
Ours70.5091.7871.2845.46
", + "image_path": "adabb445440801418caab8cf123aa46ffc2ae67de1068a9e8cb8a05703c31e95.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 496, + 288, + 519 + ], + "lines": [ + { + "bbox": [ + 47, + 496, + 288, + 519 + ], + "spans": [ + { + "bbox": [ + 47, + 496, + 288, + 519 + ], + "type": "text", + "content": "Table 2. Results on the WOD validation set in different ranges for vehicle detection." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 526, + 243, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 526, + 243, + 540 + ], + "spans": [ + { + "bbox": [ + 47, + 526, + 243, + 540 + ], + "type": "text", + "content": "4.1. Datasets and Implementation Details" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 545, + 287, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 287, + 664 + ], + "type": "text", + "content": "WOD [43] is a large dataset of autonomous driving scenes. It totally contains 798 training sequences with around 160K LiDAR samples and 202 validation sequences with 40K LiDAR samples, with the mean Average Precision (mAP) and mAP weighted by heading accuracy (mAPH) as evaluation metrics. There are also two levels of difficulty describing the sparsity in each bounding box, and LEVEL_1 (L1) and LEVEL_2 (L2) denote more than 5 points and 1-5 points, respectively. For detection performance along distance, it provides mAP/mAPH on 0-30m, 30m-50m and 50m-inf." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": "KITTI [10] is a widely used benchmark for 3D object detection, which includes 3,712, 3,769 and 7,518 frames for training, validation and testing, respectively. mAP is used as the official metric with 11 recall points for the val set and" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 285, + 545, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 285, + 545, + 323 + ], + "spans": [ + { + "bbox": [ + 305, + 285, + 545, + 323 + ], + "type": "text", + "content": "40 for the test set, and the IoU thresholds are set to 0.7, 0.5, and 0.5 for car, pedestrian and cyclist. We use the official setting in all experiments." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 323, + 545, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 323, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 305, + 323, + 545, + 371 + ], + "type": "text", + "content": "Implementation Details The total loss for optimizing the overall two-stage detection is formulated as Eq. (11), where " + }, + { + "bbox": [ + 305, + 323, + 545, + 371 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rcnn}" + }, + { + "bbox": [ + 305, + 323, + 545, + 371 + ], + "type": "text", + "content": " can be omitted if there is no RoI head. Refer to Supp. A.1 for more information." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 362, + 391, + 545, + 405 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 391, + 545, + 405 + ], + "spans": [ + { + "bbox": [ + 362, + 391, + 545, + 405 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {d e t}} = \\mathcal {L} _ {\\text {r p n}} + \\mathcal {L} _ {\\text {r c n n}} + \\mathcal {L} _ {\\text {s e g}}. \\tag {11}", + "image_path": "fa97003e915fdf4bba5c74afb82547a7378efcd01a7a862fdeba4cca632dd838.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 415, + 408, + 428 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 415, + 408, + 428 + ], + "spans": [ + { + "bbox": [ + 306, + 415, + 408, + 428 + ], + "type": "text", + "content": "4.2. Results on WOD" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 436, + 545, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 436, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 304, + 436, + 545, + 520 + ], + "type": "text", + "content": "The results on the validation set are displayed in Tab. 1, and it can be seen that we achieve new state-of-the-art performance on all the three classes. In particular, for pedestrian, we outperform the baseline model PV-RCNN++ [39] by " + }, + { + "bbox": [ + 304, + 436, + 545, + 520 + ], + "type": "inline_equation", + "content": "2.77\\% / 2.56\\%" + }, + { + "bbox": [ + 304, + 436, + 545, + 520 + ], + "type": "text", + "content": " in terms of both L1 and L2 mAP, which indicates the effectiveness of the proposed model in handling hard examples." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 522, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 546, + 713 + ], + "type": "text", + "content": "In comparison with other Transformer-based models, we focus on vehicle since the counterparts only report the performance on it. As Tab. 2 shows, our OcTr achieves the best mAP among all these convolution- and Transformer-based backbones. It also outperforms the Transformer-based detection head network CT3D [37] by " + }, + { + "bbox": [ + 304, + 522, + 546, + 713 + ], + "type": "inline_equation", + "content": "2.52\\%" + }, + { + "bbox": [ + 304, + 522, + 546, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 522, + 546, + 713 + ], + "type": "inline_equation", + "content": "1.46\\%" + }, + { + "bbox": [ + 304, + 522, + 546, + 713 + ], + "type": "text", + "content": " in L1 and L2 mAP. Regarding the accuracies at different distances, OcTr ranks the first place in the range of " + }, + { + "bbox": [ + 304, + 522, + 546, + 713 + ], + "type": "inline_equation", + "content": "30\\mathrm{m} - 50\\mathrm{m}" + }, + { + "bbox": [ + 304, + 522, + 546, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 522, + 546, + 713 + ], + "type": "inline_equation", + "content": "50\\mathrm{m}" + }, + { + "bbox": [ + 304, + 522, + 546, + 713 + ], + "type": "text", + "content": " inf, which surpasses the previous best by " + }, + { + "bbox": [ + 304, + 522, + 546, + 713 + ], + "type": "inline_equation", + "content": "0.45\\%" + }, + { + "bbox": [ + 304, + 522, + 546, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 522, + 546, + 713 + ], + "type": "inline_equation", + "content": "2.66\\%" + }, + { + "bbox": [ + 304, + 522, + 546, + 713 + ], + "type": "text", + "content": " in L1 mAP and " + }, + { + "bbox": [ + 304, + 522, + 546, + 713 + ], + "type": "inline_equation", + "content": "1.18\\%" + }, + { + "bbox": [ + 304, + 522, + 546, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 522, + 546, + 713 + ], + "type": "inline_equation", + "content": "2.26\\%" + }, + { + "bbox": [ + 304, + 522, + 546, + 713 + ], + "type": "text", + "content": " in L2 mAP respectively. It clearly illustrates that OcTr has the advantage in capturing long-range fine-grained context, which facilitates dealing with objects far away. Indeed, far objects generally have much more sparse points than near ones and heavily rely on context for detection, thus benefiting more from OcTr. Refer to Fig. 5 for visualization." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "5171" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 112, + 70, + 482, + 159 + ], + "blocks": [ + { + "bbox": [ + 112, + 70, + 482, + 159 + ], + "lines": [ + { + "bbox": [ + 112, + 70, + 482, + 159 + ], + "spans": [ + { + "bbox": [ + 112, + 70, + 482, + 159 + ], + "type": "table", + "html": "
Model\\( \\mathbf{{mAP}}_{3D}@\\mathbf{{Car}} \\)\\( \\mathbf{{mAP}}_{3D}@\\mathbf{{Pedestrian}} \\)\\( \\mathbf{{mAP}}_{3D}@\\mathbf{{Cyclist}} \\)
EasyMod.HardEasyMod.HardEasyMod.Hard
SECOND [52]88.6178.6277.2256.5552.9847.7380.5867.1563.10
PointPillars [18]88.4677.2874.6557.7552.2947.9080.0462.6159.52
VoTR [24]87.8678.2776.93------
VoxSeT [12]88.4578.4877.0760.6254.7450.3984.0768.1165.14
Ours88.4378.5777.1661.4957.1752.3585.2970.4466.17
", + "image_path": "0f624f54e106775c42a50cd91a9d0e082bcde6aea6890eeb0d93afa061e25c71.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 124, + 200, + 470, + 373 + ], + "blocks": [ + { + "bbox": [ + 46, + 166, + 547, + 190 + ], + "lines": [ + { + "bbox": [ + 46, + 166, + 547, + 190 + ], + "spans": [ + { + "bbox": [ + 46, + 166, + 547, + 190 + ], + "type": "text", + "content": "Table 3. Results of the single-stage models on the KITTI val set. All the models adopt the same anchor-based region proposal network as the detection head. \"Mod.\" denotes the moderate difficulty level." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 124, + 200, + 470, + 373 + ], + "lines": [ + { + "bbox": [ + 124, + 200, + 470, + 373 + ], + "spans": [ + { + "bbox": [ + 124, + 200, + 470, + 373 + ], + "type": "table", + "html": "
Model\\( \\mathbf{mAP}_{3D}@\\mathbf{Car on test} \\)\\( \\mathbf{mAP}_{3D}@\\mathbf{Car on val} \\)
EasyMod.HardMeanEasyMod.HardMean
SECOND [52]83.3472.5565.8273.9088.6178.6277.2281.48
PointPillars [18]82.5874.3168.9975.2986.6276.0668.9177.20
STD [54]87.9579.7175.0980.9289.7079.8079.3082.93
SA-SSD [13]88.7579.7974.1680.9090.1579.9178.7882.95
3DSSD [53]88.3679.5774.5580.8389.7179.4578.6782.61
PV-RCNN [38]90.2581.4376.8282.8389.3583.6978.7083.91
Voxel-RCNN [6]90.9081.6277.0683.1989.4184.5278.9384.29
CT3D [37]87.8381.7777.1682.2589.5486.0678.9984.86
VoTR-TSD [24]89.9082.0979.1483.7189.0484.0478.6883.92
VoxSeT [12]88.5382.0677.4682.6889.2186.7178.5684.83
Focals Conv [4]90.5582.2877.5983.4789.5284.9379.1884.54
Ours90.8882.6477.7783.7689.8086.9779.2885.35
", + "image_path": "e8f19e2377cbed17b02badadd11a980b387c88b6c587f88e0d407dcd220af016.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 52, + 416, + 282, + 498 + ], + "blocks": [ + { + "bbox": [ + 46, + 380, + 546, + 404 + ], + "lines": [ + { + "bbox": [ + 46, + 380, + 546, + 404 + ], + "spans": [ + { + "bbox": [ + 46, + 380, + 546, + 404 + ], + "type": "text", + "content": "Table 4. Comparison to the state-of-the-art models on the KITTI test and val sets. \"Mod.\" and \"Mean\" denote the moderate difficulty level and the average mAP for the three levels, respectively. The best results are bolded and the second best ones are underlined." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 52, + 416, + 282, + 498 + ], + "lines": [ + { + "bbox": [ + 52, + 416, + 282, + 498 + ], + "spans": [ + { + "bbox": [ + 52, + 416, + 282, + 498 + ], + "type": "table", + "html": "
DetectorVeh. mAP (L1/L2)Pedes. mAP (L1/L2)
SECOND [52]70.96/62.5865.23/57.22
Ours73.28/65.0568.08/60.36
PV-RCNN [38]75.41/67.4471.98/63.70
Ours76.77/68.3173.22/64.30
PV-RCNN++ [39]77.82/69.0777.99/69.92
Ours78.01/69.6080.75/72.45
", + "image_path": "3a431abd997cf8e5ddd0d4ccf1d80084e77d3fee64bbd45943cf38316cb5a4b3.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 47, + 499, + 287, + 521 + ], + "lines": [ + { + "bbox": [ + 47, + 499, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 47, + 499, + 287, + 521 + ], + "type": "text", + "content": "Table 5. Results of extensions to different representative detectors on the WOD validation set." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 533, + 153, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 533, + 153, + 547 + ], + "spans": [ + { + "bbox": [ + 47, + 533, + 153, + 547 + ], + "type": "text", + "content": "4.3. Results on KITTI" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 555, + 287, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 555, + 287, + 687 + ], + "spans": [ + { + "bbox": [ + 46, + 555, + 287, + 687 + ], + "type": "text", + "content": "The performance of the single-stage detectors is shown in Tab. 3. We take SECOND [52], a commonly used anchor-based model, as the baseline, and compare OcTr with another two advanced Transformer-based variants VoxSet [12] and VoTR [24]. We can see that OcTr achieves comparable results with SECOND and VoxSeT in car, while it significantly outperforms all the counterparts and reports the best performance both in pedestrian and cyclist. Benefiting from the large receptive field and fine-grained global context, it exceeds SECOND by " + }, + { + "bbox": [ + 46, + 555, + 287, + 687 + ], + "type": "inline_equation", + "content": "4.19\\%" + }, + { + "bbox": [ + 46, + 555, + 287, + 687 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 555, + 287, + 687 + ], + "type": "inline_equation", + "content": "3.29\\%" + }, + { + "bbox": [ + 46, + 555, + 287, + 687 + ], + "type": "text", + "content": " for pedestrians and cyclists respectively, where hard samples often appear." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "text", + "content": "We summarize the performance of the two-stage models on the KITTI test set in Tab. 4. With the help of the" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 310, + 416, + 542, + 491 + ], + "blocks": [ + { + "bbox": [ + 310, + 416, + 542, + 491 + ], + "lines": [ + { + "bbox": [ + 310, + 416, + 542, + 491 + ], + "spans": [ + { + "bbox": [ + 310, + 416, + 542, + 491 + ], + "type": "table", + "html": "
AttentionVeh. mAP (L1/L2)Pedes. mAP (L1/L2)
Ours (OctAttn)73.3/65.168.1/60.4
Performer [5]71.4/63.665.7/57.9
ACT [27]71.7/63.564.3/56.1
VoTr [24]69.4/61.565.0/57.0
Nearest K68.2/59.864.9/56.7
", + "image_path": "dcd1d85cd2b4203735b4e4d9a3462c7a7c8dab24efae73b7e1b53949726c2ee4.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 499, + 545, + 522 + ], + "lines": [ + { + "bbox": [ + 305, + 499, + 545, + 522 + ], + "spans": [ + { + "bbox": [ + 305, + 499, + 545, + 522 + ], + "type": "text", + "content": "Table 6. Ablation on various attention mechanisms and sampling patterns on the WOD validation set." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 529, + 547, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 529, + 547, + 615 + ], + "spans": [ + { + "bbox": [ + 304, + 529, + 547, + 615 + ], + "type": "text", + "content": "multi-scale backbone features and rich global context, OcTr reaches a leading mAP in car at the moderate level, surpassing the state-of-the-art Focals-Conv [4] by " + }, + { + "bbox": [ + 304, + 529, + 547, + 615 + ], + "type": "inline_equation", + "content": "0.36\\%" + }, + { + "bbox": [ + 304, + 529, + 547, + 615 + ], + "type": "text", + "content": ". We also evaluate OcTr on the KITTI val set, and OcTr again delivers the best performance in the average score, outperforming the second-best by " + }, + { + "bbox": [ + 304, + 529, + 547, + 615 + ], + "type": "inline_equation", + "content": "0.49\\%" + }, + { + "bbox": [ + 304, + 529, + 547, + 615 + ], + "type": "text", + "content": ". One can observe that we rank the best or the second best in all the cases." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 623, + 398, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 623, + 398, + 635 + ], + "spans": [ + { + "bbox": [ + 306, + 623, + 398, + 635 + ], + "type": "text", + "content": "4.4. Ablation study" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 715 + ], + "type": "text", + "content": "Scalability on various detectors As summarized in Tab. 5, we conduct experiments on three different and representative detectors, SECOND (single-stage, anchor-based), PV-RCNN (two-stage, anchor-based) and PV-RCNN++ (two-stage, anchor-free). Regardless of the number of stages or region proposal network, we acquire sound improvements" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "5172" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 52, + 70, + 282, + 135 + ], + "blocks": [ + { + "bbox": [ + 52, + 70, + 282, + 135 + ], + "lines": [ + { + "bbox": [ + 52, + 70, + 282, + 135 + ], + "spans": [ + { + "bbox": [ + 52, + 70, + 282, + 135 + ], + "type": "table", + "html": "
LEPESAPESAMVeh. mAP (L1/L2)Pedes. mAP (L1/L2)
71.35/63.3065.75/57.89
✓72.34/64.3266.56/58.62
✓✓72.64/64.4666.62/58.83
✓✓72.86/64.4067.79/59.90
✓✓✓73.28/65.0568.08/60.36
", + "image_path": "201e72619ea3ded3f1161cba651e73c6c396b7cf0becd41e85393538b4488ad3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 52, + 175, + 282, + 240 + ], + "blocks": [ + { + "bbox": [ + 47, + 144, + 286, + 166 + ], + "lines": [ + { + "bbox": [ + 47, + 144, + 286, + 166 + ], + "spans": [ + { + "bbox": [ + 47, + 144, + 286, + 166 + ], + "type": "text", + "content": "Table 7. Ablation on semantic positional embedding on the WOD validation set." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 52, + 175, + 282, + 240 + ], + "lines": [ + { + "bbox": [ + 52, + 175, + 282, + 240 + ], + "spans": [ + { + "bbox": [ + 52, + 175, + 282, + 240 + ], + "type": "table", + "html": "
topk numberVeh. mAP (L1/L2)Pedes. mAP (L1/L2)
170.38/62.2064.19/56.43
472.58/64.4266.21/58.42
873.28/65.0568.08/60.36
1673.25/65.0167.89/60.10
", + "image_path": "1bab0f751e46f54d24c2060df348a88747ee74a6d686184af3ed38b5bb780eca.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 52, + 270, + 285, + 327 + ], + "blocks": [ + { + "bbox": [ + 48, + 249, + 285, + 259 + ], + "lines": [ + { + "bbox": [ + 48, + 249, + 285, + 259 + ], + "spans": [ + { + "bbox": [ + 48, + 249, + 285, + 259 + ], + "type": "text", + "content": "Table 8. Results of different " + }, + { + "bbox": [ + 48, + 249, + 285, + 259 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 48, + 249, + 285, + 259 + ], + "type": "text", + "content": " values on the WOD validation set." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 52, + 270, + 285, + 327 + ], + "lines": [ + { + "bbox": [ + 52, + 270, + 285, + 327 + ], + "spans": [ + { + "bbox": [ + 52, + 270, + 285, + 327 + ], + "type": "table", + "html": "
Method#Param. (M)Latency (ms)Memory (GB)
SECOND [52]5.3482.3
VoTR-SSD [24]4.8673.0
VoxSeT-SSD [12]3.0373.6
OcTr-SSD2.9642.5
", + "image_path": "cddb82516b0acf59e26496f81274296bc54fb3072ac36b7e71c04040619e417f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 336, + 286, + 357 + ], + "lines": [ + { + "bbox": [ + 47, + 336, + 286, + 357 + ], + "spans": [ + { + "bbox": [ + 47, + 336, + 286, + 357 + ], + "type": "text", + "content": "Table 9. Resource costs of different backbones with single-stage detectors on the KITTI dataset, test on GTX2080Ti." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 378, + 286, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 378, + 286, + 402 + ], + "spans": [ + { + "bbox": [ + 46, + 378, + 286, + 402 + ], + "type": "text", + "content": "compared to the baselines of the sparse convolution backbones, highlighting its scalability." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 402, + 286, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 402, + 286, + 485 + ], + "spans": [ + { + "bbox": [ + 46, + 402, + 286, + 485 + ], + "type": "text", + "content": "Ablation on OctAttn We carry out additional experiments to make apple-to/apple comparison with several representative linear Transformer methods, including Performer [5] (kernel-based linear attention), ACT [27] (cluster-based linear attention), VoTr [24] (fixed patterns) and the Nearest-K strategy. Tab. 6 lists the results, and our OctAttn clearly performs the best, showing its ability." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 486, + 286, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 286, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 286, + 653 + ], + "type": "text", + "content": "Ablation on semantic positional embedding We individually evaluate the contributions of LePE, SAPE and SAM with SECOND as the baseline detector on WOD in Tab. 7. By incorporating LePE, the L1/L2 performance is boosted by " + }, + { + "bbox": [ + 46, + 486, + 286, + 653 + ], + "type": "inline_equation", + "content": "0.99\\% / 1.02\\%" + }, + { + "bbox": [ + 46, + 486, + 286, + 653 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 486, + 286, + 653 + ], + "type": "inline_equation", + "content": "0.81\\% / 0.73\\%" + }, + { + "bbox": [ + 46, + 486, + 286, + 653 + ], + "type": "text", + "content": " on vehicle and pedestrian, illustrating its necessity. Furthermore, we separately verify the validity of SAPE and SAM. With semantic clues, we observe that SAPE increases by " + }, + { + "bbox": [ + 46, + 486, + 286, + 653 + ], + "type": "inline_equation", + "content": "0.3\\%" + }, + { + "bbox": [ + 46, + 486, + 286, + 653 + ], + "type": "text", + "content": " L1 mAP on vehicle, while SAM provides an L1/L2 mAP improvement of " + }, + { + "bbox": [ + 46, + 486, + 286, + 653 + ], + "type": "inline_equation", + "content": "0.52\\% / 0.08\\%" + }, + { + "bbox": [ + 46, + 486, + 286, + 653 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 486, + 286, + 653 + ], + "type": "inline_equation", + "content": "1.23\\% / 1.28\\%" + }, + { + "bbox": [ + 46, + 486, + 286, + 653 + ], + "type": "text", + "content": " on vehicle and pedestrian. Finally, we simultaneously apply SAPE and SAM and construct the full model of OcTr, which gains " + }, + { + "bbox": [ + 46, + 486, + 286, + 653 + ], + "type": "inline_equation", + "content": "0.94\\% / 0.73\\%" + }, + { + "bbox": [ + 46, + 486, + 286, + 653 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 486, + 286, + 653 + ], + "type": "inline_equation", + "content": "1.52\\% / 1.74\\%" + }, + { + "bbox": [ + 46, + 486, + 286, + 653 + ], + "type": "text", + "content": " L1/L2 mAP on vehicle and pedestrian, showing its impact." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 653, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 286, + 712 + ], + "type": "text", + "content": "Influence by topk Tab. 8 shows the performance of OcTr with various values of " + }, + { + "bbox": [ + 46, + 653, + 286, + 712 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 653, + 286, + 712 + ], + "type": "text", + "content": ". As we can see, the performance improves when " + }, + { + "bbox": [ + 46, + 653, + 286, + 712 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 653, + 286, + 712 + ], + "type": "text", + "content": " becomes larger but quickly saturates. Due to the redundancy of the scanned scenes, we argue that only a few tokens with high relevance need to be subdivided to" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 72, + 430, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 430, + 83 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 430, + 83 + ], + "type": "text", + "content": "embrace fine-grained features." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 84, + 545, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 84, + 545, + 215 + ], + "spans": [ + { + "bbox": [ + 304, + 84, + 545, + 215 + ], + "type": "text", + "content": "Analysis on model complexity We compare OcTr with two recent Transformer-based models in terms of resource cost by keeping the same detection head. Tab. 9 shows that with the learnable octree attention mechanism, OcTr consistently maintains less model parameters and less memory occupancies than the counterparts. Regarding the inference speed, VoxSeT runs faster, but it should be noted that VoxSeT inputs with pillars which discard the height dimension, leading to inferior results. As for VoTR, we deliver a mild improvement in efficiency while bringing a large gain in performance." + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 307, + 227, + 544, + 312 + ], + "blocks": [ + { + "bbox": [ + 307, + 227, + 544, + 312 + ], + "lines": [ + { + "bbox": [ + 307, + 227, + 544, + 312 + ], + "spans": [ + { + "bbox": [ + 307, + 227, + 544, + 312 + ], + "type": "image", + "image_path": "772b9d9f91728cd492ba4862a309cb6ecbe9349aec506fc8fbb76e755af67f38.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 307, + 312, + 541, + 405 + ], + "blocks": [ + { + "bbox": [ + 307, + 312, + 541, + 405 + ], + "lines": [ + { + "bbox": [ + 307, + 312, + 541, + 405 + ], + "spans": [ + { + "bbox": [ + 307, + 312, + 541, + 405 + ], + "type": "image", + "image_path": "359cc85ce0005bf04fb6f6291db3c1b6ed59be462b73ca75830975ca33e96e37.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 414, + 545, + 437 + ], + "lines": [ + { + "bbox": [ + 305, + 414, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 305, + 414, + 545, + 437 + ], + "type": "text", + "content": "Figure 5. Visualization of results by OcTr on the WOD validation split. Blue/red indicates predicted/ground-truth bounding boxes" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 456, + 378, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 456, + 378, + 468 + ], + "spans": [ + { + "bbox": [ + 306, + 456, + 378, + 468 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 477, + 545, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 477, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 304, + 477, + 545, + 609 + ], + "type": "text", + "content": "This paper proposes a novel voxel-based approach to 3D object detection, namely OcTr. It aims to balance the fine-grained global representation and efficiency with acceptable resource costs. To this end, we propose a learned sparsification attention mechanism, OctAttn, which adaptively prunes the octants from the multi-scale feature pyramid in a top-to-bottom manner. Furthermore, we adopt a hybrid semantic-aware positional embedding based on foreground segmentation. Extensive experiments are conducted on WOD and KITTI, and OcTr reaches the state-of-the-art performance, validating its effectiveness." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 621, + 398, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 621, + 398, + 634 + ], + "spans": [ + { + "bbox": [ + 306, + 621, + 398, + 634 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 641, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 712 + ], + "type": "text", + "content": "This work is partly supported by the National Natural Science Foundation of China (No. 62022011 and No. 62202034), the Research Program of State Key Laboratory of Software Development Environment (SKLSDE-2021ZX-04), and the Fundamental Research Funds for the Central Universities." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5173" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Joshua Ainslie, Santiago Ontanon, Chris Alberti, Vaclav Cvicek, Zachary Fisher, Philip Pham, Anirudh Ravula, Sumit Sanghai, Qifan Wang, and Li Yang. Etc: Encoding long and structured inputs in transformers. In EMNLP, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 148, + 288, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 288, + 179 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 288, + 179 + ], + "type": "text", + "content": "[2] Iz Beltagy, Matthew E Peters, and Arman Cohan. Longformer: The long-document transformer. arXiv preprint arXiv:2004.05150, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 182, + 287, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 182, + 287, + 215 + ], + "spans": [ + { + "bbox": [ + 53, + 182, + 287, + 215 + ], + "type": "text", + "content": "[3] Chen Chen, Zhe Chen, Jing Zhang, and Dacheng Tao. Sasa: Semantics-augmented set abstraction for point-based 3d object detection. In AAAI, volume 1, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 217, + 287, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 217, + 287, + 250 + ], + "spans": [ + { + "bbox": [ + 53, + 217, + 287, + 250 + ], + "type": "text", + "content": "[4] Yukang Chen, Yanwei Li, Xiangyu Zhang, Jian Sun, and Jiaya Jia. Focal sparse convolutional networks for 3d object detection. In CVPR, pages 5428-5437, 2022. 1, 2, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 251, + 287, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 251, + 287, + 316 + ], + "spans": [ + { + "bbox": [ + 53, + 251, + 287, + 316 + ], + "type": "text", + "content": "[5] Krzysztof Marcin Choromanski, Valerii Likhosherstov, David Dohan, Xingyou Song, Andreea Gane, Tamás Sarlós, Peter Hawkins, Jared Quincy Davis, Afroz Mohiuddin, Lukasz Kaiser, David Benjamin Belanger, Lucy J. Colwell, and Adrian Weller. Rethinking attention with performers. In ICLR, 2021. 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 318, + 287, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 318, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 53, + 318, + 287, + 361 + ], + "type": "text", + "content": "[6] Jiajun Deng, Shaoshuai Shi, Peiwei Li, Wengang Zhou, Yanyong Zhang, and Houqiang Li. Voxel r-cnn: Towards high performance voxel-based 3d object detection. In AAAI, volume 35, pages 1201-1209, 2021. 2, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 363, + 287, + 417 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 363, + 287, + 417 + ], + "spans": [ + { + "bbox": [ + 53, + 363, + 287, + 417 + ], + "type": "text", + "content": "[7] Xiaoyi Dong, Jianmin Bao, Dongdong Chen, Weiming Zhang, Nenghai Yu, Lu Yuan, Dong Chen, and Baining Guo. Cswin transformer: A general vision transformer backbone with cross-shaped windows. In CVPR, pages 12124-12134, 2022. 2, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 419, + 287, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 419, + 287, + 474 + ], + "spans": [ + { + "bbox": [ + 53, + 419, + 287, + 474 + ], + "type": "text", + "content": "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2020. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 475, + 287, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 475, + 287, + 520 + ], + "spans": [ + { + "bbox": [ + 53, + 475, + 287, + 520 + ], + "type": "text", + "content": "[9] Lue Fan, Ziqi Pang, Tianyuan Zhang, Yu-Xiong Wang, Hang Zhao, Feng Wang, Naiyan Wang, and Zhaoxiang Zhang. Embracing single stride 3d object detector with sparse transformer. In CVPR, pages 8458-8468, 2022. 2, 3, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 521, + 287, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 521, + 287, + 553 + ], + "spans": [ + { + "bbox": [ + 48, + 521, + 287, + 553 + ], + "type": "text", + "content": "[10] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In CVPR, pages 3354-3361. IEEE, 2012. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 555, + 287, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 555, + 287, + 587 + ], + "spans": [ + { + "bbox": [ + 48, + 555, + 287, + 587 + ], + "type": "text", + "content": "[11] Meng-Hao Guo, Jun-Xiong Cai, Zheng-Ning Liu, Tai-Jiang Mu, Ralph R Martin, and Shi-Min Hu. Pct: Point cloud transformer. CVM, 7(2):187-199, 2021. 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 590, + 287, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 633 + ], + "type": "text", + "content": "[12] Chenhang He, Ruihuang Li, Shuai Li, and Lei Zhang. Voxel set transformer: A set-to-set approach to 3d object detection from point clouds. In CVPR, pages 8417-8427, 2022. 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "type": "text", + "content": "[13] Chenhang He, Hui Zeng, Jianqiang Huang, Xian-Sheng Hua, and Lei Zhang. Structure aware single-stage 3d object detection from point cloud. In CVPR, pages 11873-11882, 2020. 2, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 3" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "type": "text", + "content": "[15] Jordan SK Hu. Point density-aware voxels for lidar 3d object detection. In CVPR, pages 8469-8478, 2022. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 95, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 95, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 95, + 545, + 128 + ], + "type": "text", + "content": "[16] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In ICML, pages 448-456. PMLR, 2015. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 129, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 150 + ], + "type": "text", + "content": "[17] Eric Jang, Shixiang Gu, and Ben Poole. Categorical reparameterization with gumbel-softmax. In ICLR, 2016. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 151, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 151, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 307, + 151, + 545, + 194 + ], + "type": "text", + "content": "[18] Alex H Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. In CVPR, pages 12697-12705, 2019. 2, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 194, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 194, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 194, + 545, + 239 + ], + "type": "text", + "content": "[19] Juho Lee, Yoonho Lee, Jungtaek Kim, Adam Kosiorek, Seungjin Choi, and Yee Whye Teh. Set transformer: A framework for attention-based permutation-invariant neural networks. In ICML, pages 3744-3753. PMLR, 2019. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 239, + 545, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 239, + 545, + 271 + ], + "spans": [ + { + "bbox": [ + 307, + 239, + 545, + 271 + ], + "type": "text", + "content": "[20] Zhichao Li, Feng Wang, and Naiyan Wang. Lidar r-cnn: An efficient and universal 3d object detector. In CVPR, pages 7546-7555, 2021. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 272, + 545, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 272, + 545, + 315 + ], + "spans": [ + { + "bbox": [ + 307, + 272, + 545, + 315 + ], + "type": "text", + "content": "[21] Youwei Liang, Chongjian Ge, Zhan Tong, Yibing Song, Jue Wang, and Pengtao Xie. Not all patches are what you need: Expediting vision transformers via token reorganizations. In ICLR, 2022. 2, 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 316, + 545, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 316, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 307, + 316, + 545, + 348 + ], + "type": "text", + "content": "[22] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In ICCV, pages 2980-2988, 2017. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 350, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 350, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 307, + 350, + 545, + 392 + ], + "type": "text", + "content": "[23] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 10012-10022, 2021. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 394, + 545, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 394, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 307, + 394, + 545, + 415 + ], + "type": "text", + "content": "[24] Jiageng Mao. Voxel transformer for 3d object detection. In ICCV, pages 3164-3173, 2021. 1, 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 416, + 545, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 416, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 307, + 416, + 545, + 459 + ], + "type": "text", + "content": "[25] Jiageng Mao, Minzhe Niu, Haoyue Bai, Xiaodan Liang, Hang Xu, and Chunjing Xu. Pyramid r-cnn: Towards better performance and adaptability for 3d object detection. In ICCV, pages 2723-2732, 2021. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 460, + 545, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 460, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 307, + 460, + 545, + 502 + ], + "type": "text", + "content": "[26] Julien NP Martel, David B Lindell, Connor Z Lin, Eric R Chan, Marco Monteiro, and Gordon Wetzstein. Acorn: Adaptive coordinate networks for neural scene representation. In ACM SIGGRAPH, 2021. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 504, + 545, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 504, + 545, + 547 + ], + "spans": [ + { + "bbox": [ + 307, + 504, + 545, + 547 + ], + "type": "text", + "content": "[27] Zheng Minghang, Gao Peng, Zhang Renrui, Li Kunchang, Li Hongsheng, and Dong Hao. End-to-end object detection with adaptive clustering transformer. In BMVC, page 226, 2021. 7, 8" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 548, + 545, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 548, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 307, + 548, + 545, + 581 + ], + "type": "text", + "content": "[28] Ishan Misra, Rohit Girdhar, and Armand Joulin. An end-to-end transformer model for 3d object detection. In ICCV, pages 2906-2917, 2021. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 582, + 545, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 582, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 307, + 582, + 545, + 613 + ], + "type": "text", + "content": "[29] Xuran Pan, Zhuofan Xia, Shiji Song, Li Erran Li, and Gao Huang. 3d object detection with pointformer. In CVPR, pages 7463-7472, 2021. 1, 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 614, + 545, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 614, + 545, + 646 + ], + "spans": [ + { + "bbox": [ + 307, + 614, + 545, + 646 + ], + "type": "text", + "content": "[30] Chunghyun Park, Yoonwoo Jeong, Minsu Cho, and Jaesik Park. Fast point transformer. In CVPR, pages 16949-16958, 2022. 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 647, + 545, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 680 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 680 + ], + "type": "text", + "content": "[31] Charles R Qi, Or Litany, Kaiming He, and Leonidas J Guibas. Deep hough voting for 3d object detection in point clouds. In ICCV, pages 9277-9286, 2019. 2" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 681, + 545, + 713 + ], + "type": "text", + "content": "[32] Charles R Qi, Wei Liu, Chenxia Wu, Hao Su, and Leonidas J Guibas. Frustum pointnets for 3d object detection from rgb-d data. In CVPR, pages 918-927, 2018. 1, 2" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5174" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[33] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In NeurIPS, volume 30, 2017. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 287, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 287, + 149 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 287, + 149 + ], + "type": "text", + "content": "[34] Jiezhong Qiu, Hao Ma, Omer Levy, Scott Wen-tau Yih, Sinong Wang, and Jie Tang. Blockwise self-attention for long document understanding. arXiv preprint arXiv:1911.02972, 2019. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 150, + 287, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 150, + 287, + 183 + ], + "spans": [ + { + "bbox": [ + 48, + 150, + 287, + 183 + ], + "type": "text", + "content": "[35] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In NeurIPS, volume 28, 2015. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 184, + 287, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 184, + 287, + 216 + ], + "spans": [ + { + "bbox": [ + 48, + 184, + 287, + 216 + ], + "type": "text", + "content": "[36] Gernot Riegler, Ali Osman Ulusoy, and Andreas Geiger. Octnet: Learning deep 3d representations at high resolutions. In CVPR, pages 3577-3586, 2017. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 217, + 287, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 217, + 287, + 259 + ], + "spans": [ + { + "bbox": [ + 48, + 217, + 287, + 259 + ], + "type": "text", + "content": "[37] Hualian Sheng, Sijia Cai, Yuan Liu, Bing Deng, Jianqiang Huang, Xian-Sheng Hua, and Min-Jian Zhao. Improving 3d object detection with channel-wise transformer. In ICCV, pages 2743-2752, 2021. 3, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 261, + 287, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 261, + 287, + 304 + ], + "spans": [ + { + "bbox": [ + 48, + 261, + 287, + 304 + ], + "type": "text", + "content": "[38] Shaoshuai Shi, Chaoxu Guo, Li Jiang, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. Pv-rcnn: Point-voxel feature set abstraction for 3d object detection. In CVPR, pages 10529–10538, 2020. 2, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 305, + 287, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 305, + 287, + 358 + ], + "spans": [ + { + "bbox": [ + 48, + 305, + 287, + 358 + ], + "type": "text", + "content": "[39] Shaoshuai Shi, Li Jiang, Jiajun Deng, Zhe Wang, Chaoxu Guo, Jianping Shi, Xiaogang Wang, and Hongsheng Li. Pvrccnn++: Point-voxel feature set abstraction with local vector representation for 3d object detection. IEEE TPAMI, 2021. 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 360, + 287, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 360, + 287, + 392 + ], + "spans": [ + { + "bbox": [ + 48, + 360, + 287, + 392 + ], + "type": "text", + "content": "[40] Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. Pointr-cnn: 3d object proposal generation and detection from point cloud. In CVPR, pages 770-779, 2019. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 393, + 287, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 393, + 287, + 436 + ], + "spans": [ + { + "bbox": [ + 48, + 393, + 287, + 436 + ], + "type": "text", + "content": "[41] Shaoshuai Shi, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. From points to parts: 3d object detection from point cloud with part-aware and part-aggregation network. IEEE TPAMI, 43(8):2647-2664, 2020. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 437, + 287, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 437, + 287, + 469 + ], + "spans": [ + { + "bbox": [ + 48, + 437, + 287, + 469 + ], + "type": "text", + "content": "[42] Weijing Shi and Raj Rajkumar. Point-gnn: Graph neural network for 3d object detection in a point cloud. In CVPR, pages 1711-1719, 2020. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 471, + 287, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 471, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 48, + 471, + 287, + 525 + ], + "type": "text", + "content": "[43] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In CVPR, pages 2446-2454, 2020. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 525, + 287, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 287, + 558 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 287, + 558 + ], + "type": "text", + "content": "[44] Jia-Heng Tang, Weikai Chen, Jie Yang, Bo Wang, Songrun Liu, Bo Yang, and Lin Gao. Octfield: Hierarchical implicit functions for 3d modeling. In NeurIPS, 2021. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 559, + 287, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 559, + 287, + 580 + ], + "spans": [ + { + "bbox": [ + 48, + 559, + 287, + 580 + ], + "type": "text", + "content": "[45] Shitao Tang, Jiahui Zhang, Siyu Zhu, and Ping Tan. Quadtree attention for vision transformers. In ICLR, 2022. 2, 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 582, + 287, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 582, + 287, + 623 + ], + "spans": [ + { + "bbox": [ + 48, + 582, + 287, + 623 + ], + "type": "text", + "content": "[46] OpenPCDet Development Team. Openpcdet: An opensource toolbox for 3d object detection from point clouds. https://github.com/open-mmlab/OpenPCDet, 2020.2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "type": "text", + "content": "[47] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Ilia Polosukhin. Attention is all you need. In NeurIPS, volume 30, 2017. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "type": "text", + "content": "[48] Peng-Shuai Wang, Yang Liu, Yu-Xiao Guo, Chun-Yu Sun, and Xin Tong. O-cnn: Octree-based convolutional neural networks for 3d shape analysis. ACM TOG, 36(4):1-11, 2017. 3" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 656 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[49] Peng-Shuai Wang, Chun-Yu Sun, Yang Liu, and Xin Tong. Adaptive o-cnn: A patch-based deep representation of 3d shapes. ACM TOG, 37(6):1-11, 2018. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 107, + 545, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 149 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 149 + ], + "type": "text", + "content": "[50] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph cnn for learning on point clouds. ACM TOG, 38(5):1-12, 2019. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 152, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 194 + ], + "type": "text", + "content": "[51] Tete Xiao, Mannat Singh, Eric Mintun, Trevor Darrell, Piotr Dollár, and Ross Girshick. Early convolutions help transformers see better. In NeurIPS, volume 34, pages 30392-30400, 2021. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 197, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 197, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 197, + 545, + 228 + ], + "type": "text", + "content": "[52] Yan Yan, Yuxing Mao, and Bo Li. Second: Sparsely embedded convolutional detection. Sensors, 18(10):3337, 2018. 1, 2, 6, 7, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 230, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 230, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 230, + 545, + 262 + ], + "type": "text", + "content": "[53] Zetong Yang, Yanan Sun, Shu Liu, and Jiaya Jia. 3dssd: Point-based 3d single stage object detector. In CVPR, pages 11040-11048, 2020. 2, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 264, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 264, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 307, + 264, + 545, + 297 + ], + "type": "text", + "content": "[54] Zetong Yang, Yanan Sun, Shu Liu, Xiaoyong Shen, and Jiaya Jia. Std: Sparse-to-dense 3d object detector for point cloud. In CVPR, pages 1951-1960, 2019. 2, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 298, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 298, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 307, + 298, + 545, + 330 + ], + "type": "text", + "content": "[55] Tianwei Yin, Xingyi Zhou, and Philipp Krahenbuhl. Center-based 3d object detection and tracking. In CVPR, pages 11784-11793, 2021. 2, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 332, + 545, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 332, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 307, + 332, + 545, + 387 + ], + "type": "text", + "content": "[56] Manzil Zaheer, Guru Guruganesh, Kumar Avinava Dubey, Joshua Ainslie, Chris Alberti, Santiago Ontanon, Philip Pham, Anirudh Ravula, Qifan Wang, Li Yang, et al. Big bird: Transformers for longer sequences. In NeurIPS, volume 33, pages 17283-17297, 2020. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 388, + 545, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 545, + 421 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 545, + 421 + ], + "type": "text", + "content": "[57] Cheng Zhang, Haocheng Wan, Xinyi Shen, and Zizhao Wu. Patchformer: An efficient point transformer with patch attention. In CVPR, pages 11799-11808, 2022. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 422, + 545, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 422, + 545, + 454 + ], + "spans": [ + { + "bbox": [ + 307, + 422, + 545, + 454 + ], + "type": "text", + "content": "[58] Yanan Zhang, Jiaxin Chen, and Di Huang. Cat-det: Contrastively augmented transformer for multi-modal 3d object detection. In CVPR, pages 908-917, 2022. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 456, + 545, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 456, + 545, + 499 + ], + "spans": [ + { + "bbox": [ + 307, + 456, + 545, + 499 + ], + "type": "text", + "content": "[59] Yifan Zhang, Qingyong Hu, Guoquan Xu, Yanxin Ma, Jianwei Wan, and Yulan Guo. Not all points are equal: Learning highly efficient point-based detectors for 3d lidar point clouds. In CVPR, pages 18953-18962, 2022. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 501, + 545, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 501, + 545, + 543 + ], + "spans": [ + { + "bbox": [ + 307, + 501, + 545, + 543 + ], + "type": "text", + "content": "[60] Yanan Zhang, Di Huang, and Yunhong Wang. Pc-rgnn: Point cloud completion and graph neural network for 3d object detection. In AAAI, volume 35, pages 3430-3437, 2021. 1, 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 545, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 545, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 307, + 545, + 545, + 578 + ], + "type": "text", + "content": "[61] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In ICCV, pages 16259-16268, 2021. 1, 2, 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 579, + 545, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 579, + 545, + 622 + ], + "spans": [ + { + "bbox": [ + 307, + 579, + 545, + 622 + ], + "type": "text", + "content": "[62] Wu Zheng, Weiliang Tang, Sijin Chen, Li Jiang, and Chi-Wing Fu. Cia-ssd: Confident iou-aware single-stage object detector from point cloud. In AAAI, volume 35, pages 3555–3562, 2021. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 624, + 545, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 656 + ], + "type": "text", + "content": "[63] Yin Zhou and Oncel Tuzel. Voxelnet: End-to-end learning for point cloud based 3d object detection. In CVPR, pages 4490-4499, 2018. 2" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5175" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Occlusion-Free Scene Recovery via Neural Radiance Fields/223bdb66-a260-46a1-86eb-891a595d848a_content_list.json b/2023/Occlusion-Free Scene Recovery via Neural Radiance Fields/223bdb66-a260-46a1-86eb-891a595d848a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8cfc51eb514b3ef5b9f94787f64712b713ed108d --- /dev/null +++ b/2023/Occlusion-Free Scene Recovery via Neural Radiance Fields/223bdb66-a260-46a1-86eb-891a595d848a_content_list.json @@ -0,0 +1,2081 @@ +[ + { + "type": "text", + "text": "Occlusion-Free Scene Recovery via Neural Radiance Fields", + "text_level": 1, + "bbox": [ + 186, + 130, + 782, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chengxuan Zhu $^{1,2}$ Renjie Wan $^{3}$ Yunkai Tang $^{1,2}$ Boxin Shi $^{1,2*}$", + "bbox": [ + 220, + 179, + 751, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ National Engineering Research Center of Visual Technology, School of Computer Science, Peking University $^{2}$ National Key Laboratory for Multimedia Information Processing, School of Computer Science, Peking University $^{3}$ Department of Computer Science, Hong Kong Baptist University {peterzhu, shiboxin}@pku.edu.cn, renjiewan@hkbu.edu.hk, tangyunkai@stu.pku.edu.cn", + "bbox": [ + 125, + 199, + 839, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 304, + 313, + 320 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Our everyday lives are filled with occlusions that we strive to see through. By aggregating desired background information from different viewpoints, we can easily eliminate such occlusions without any external occlusion-free supervision. Though several occlusion removal methods have been proposed to empower machine vision systems with such ability, their performances are still unsatisfactory due to reliance on external supervision. We propose a novel method for occlusion removal by directly building a mapping between position and viewing angles and the corresponding occlusion-free scene details leveraging Neural Radiance Fields (NeRF). We also develop an effective scheme to jointly optimize camera parameters and scene reconstruction when occlusions are present. An additional depth constraint is applied to supervise the entire optimization without labeled external data for training. The experimental results on existing and newly collected datasets validate the effectiveness of our method. Our project page: https://freebutuselessssoul.github.io/occnerf.", + "bbox": [ + 76, + 335, + 473, + 623 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 667, + 209, + 684 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Neural Radiance Fields (NeRF) are capable of learning the scene representation implicitly from a set of 2D images, yet not every scene is favored by observers. Many undesirable occlusions in our world obscure details that are essential to our understanding of the world. In general, such obstructions range from water droplets and scribbles on a piece of glass, to fences or any objects occluding the desired scenes (e.g., a statue closer to the camera in a landmark scene). How to apply computational methods to exclude them from the scene representation is of great interest.", + "bbox": [ + 75, + 693, + 468, + 845 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Occlusion removal (e.g., [29]) is the direct solution to achieve this goal. However, explicit occlusion removal may", + "bbox": [ + 76, + 845, + 468, + 876 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "oversmooth essential details necessary for clearly observing the desired background scenes. In addition, current methods mainly depend on external occlusion-free supervision (e.g., fence removal [4], raindrop removal [22]) to develop the reliable capability in removing certain types of occlusions. Once encountering a new scenario with unseen occlusion types beyond their training data, these methods might show degraded performances. To handle more diverse types of occlusions, generic constraints from multiple viewpoints are widely adopted [4, 10, 12, 14, 19, 27, 29] via mimicking our human vision systems, who can easily piece together the desired background scenes by looking at them from different viewpoints. But the majority of these methods just consider viewpoints as a prior in relation to spatial correlations. Their backbones still rely on external training data with corresponding ground truth for optimization, which still does not fundamentally alleviate the difficulty of handling diverse occlusions in the real world.", + "bbox": [ + 496, + 305, + 893, + 577 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "An occlusion-free world can be progressively aggregated by seeing its occluded part from different viewing directions to reveal occlusions previously unobservable in each single perspective, as illustrated in the left part of Fig. 1 (the fan is occluded in the target view). Since NeRF [18] employs an implicit representation to map viewpoints to pixels, one may come to the naive solution of directly constructing a NeRF which is optimized across multiple viewpoints. However, the vanilla NeRF [18] representing the scene as a whole is not able to treat occlusion and background scenes distinctively, and as long as the occlusion remains static, NeRF is designed to faithfully reconstruct its presence. Meanwhile, many NeRF variants can decompose the whole scene into different components (e.g., NeRF-W [17], Ha-NeRF [3], NeRFRen [7]), but they cannot handle the real-world static occlusions. This is because NeRF-W [17] and Ha-NeRF [3] rely on the inconsistency of undesired components across different views to achieve such separation, which is difficult to be observed in a continuous 3D world. On the other hand, NeRFRen [7] only works in separating the transmission and reflection compo", + "bbox": [ + 496, + 583, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 94, + 886, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "20722", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "nents caused by semi-transparent planar glass, which is incapable of handling opaque occlusions in the real world.", + "bbox": [ + 75, + 90, + 467, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Another problem comes from NeRF's reliance on camera parameters pre-computed by COLMAP [23], because handcrafted features extracted and matched using COLMAP [23] are for the whole scene, and are incapable in distinguishing between undesired occlusions and the desired background. When the features from occlusion dominate the matching process, the obtained camera parameters cannot faithfully model the spatial correlation of background scenes across multiple viewpoints. Besides, COLMAP [23] is not a stable option for pose estimation in the real world [28]. The existence of occlusions may prevent it from working properly, making the occlusion-free scene representation infeasible.", + "bbox": [ + 75, + 121, + 467, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we aim at seeing through the occluded scenes by developing an occlusion-free scene representation without considering specific occlusion types, based on which we can render any occlusion-free images from desired viewpoints. Our method first maps viewing angles and their corresponding scene details by leveraging NeRF. We then introduce a depth constraint to probe the occluded areas by measuring the depth of occlusion and background, by assuming that occlusions are always in the foreground with closer distance. During the scene modeling process, a pose refinement scheme is further introduced to refine the camera pose with the features of the background scene. As outlined in Fig. 1, our pipeline contains three modules to achieve the above goals: 1) a scene reconstruction module to represent the whole scene using NeRF (with occlusions), 2) a cost volume construction module to gather information from neighboring views as guidance (to indicate where occlusions are), and 3) a selective supervision scheme to constrain another NeRF on the desired background information (occlusions removed), and our contributions can be summarized as follows:", + "bbox": [ + 75, + 316, + 467, + 632 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- an occlusion-free representation without relying on any external prior as supervisory knowledge;", + "- a joint optimization of pose refinement and scene reconstruction by effective multi-view feature fusion;", + "- a selective supervision scheme to probe the occluded areas guided by the scene depth information." + ], + "bbox": [ + 94, + 633, + 467, + 726 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Based on the experiments with a dataset containing diverse types of occlusions, the proposed method can eliminate occlusions including scribbles and water droplets on a piece of glass, fences, and even irregular-shaped statues without relying on any external supervisions.", + "bbox": [ + 75, + 734, + 467, + 809 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 821, + 217, + 835 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. NeRF and its variants", + "text_level": 1, + "bbox": [ + 76, + 847, + 281, + 861 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "NeRF becomes a popular choice for implicit volumetric scene representation [18]. By mapping the point and current", + "bbox": [ + 76, + 869, + 467, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "viewing direction to its color and density using a neural network and a differentiable rendering scheme, it provides an effective way to learn from multiple viewing angles. Such ability has been explored to complement the missing depth for grasping transparent objects [9]. Based on NeRF, a series of methods have been proposed to tackle the problems like scene understanding [33] and reasoning [26]. Besides, by considering specific physical priors, NeRF-based frameworks also show promising performance in learning clear representation from challenging scenarios [7,8,16]. Recent NeRF variants can even reconstruct the scene from input with various perturbations. For example, NeRF-W [17] and Ha-NeRF [3] can separate the transient objects from the whole scene. However, they target occlusions that are inconsistent throughout the image set and become ineffective when such occlusions remain static and consistent.", + "bbox": [ + 496, + 90, + 890, + 330 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Since NeRF relies on pre-computed camera parameters for scene representation, it fails when the pre-computation is not feasible. Some methods (e.g., NeRF-- [28] and BARF [13]) have been proposed to alleviate this issue by optimizing the camera parameters along with the scene representation. However, how to avoid the interference from undesired scene during the camera parameter optimization remains a problem to be solved.", + "bbox": [ + 496, + 335, + 890, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Occlusion removal", + "text_level": 1, + "bbox": [ + 500, + 477, + 679, + 491 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Traditional methods in occlusion removal address this problem by propagating neighboring pixels via anisotropic diffusion [2] or solving differential equations [1]. Recently, there are more approaches solving this problem in a learning-based manner [11, 34]. A recent trend is to employ the edge information [21] or segmentation masks [24] as a stronger prior to alleviate the ill-posedness of this problem.", + "bbox": [ + 496, + 503, + 890, + 609 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Researchers in this area have been aware of the importance of removing occlusions from multiple viewpoints. Traditional methods try to take the benefits of multiple viewpoints by computing a disparity map [10], dense flow field [29], or visual parallax [19]. Using a light field camera that records the spatial and angular information of light rays in the space is also helpful in removing occlusions [25]. However, these methods also pose restrictions on their inputs, such as using stereo image pairs [10], requiring the background to be visible in at least one of the input views [25, 29], or requesting obviously relative motion between occlusions and backgrounds [19]. It is also a recent trend to apply deep learning in guiding network optimization with fewer constraints on input. For example, some methods seek to use the temporal information [4] or the optical flow [14] from video frames to guide the occlusion-free recovery process. However, these methods are confined to small movements of the camera, partially due to the difficulty in 3D scene representation.", + "bbox": [ + 496, + 613, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "20723", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a0275a5899b29a71caf6f47a0dedd3e6df3c6d8c135b8ed9e8512222ce1b4c0d.jpg", + "image_caption": [ + "Figure 1. Pipeline of the proposed method. By taking the positional encoding [18] of position and viewing direction, the \"scene MLP\" (colored in blue) models the scene with occlusion faithfully, and the \"background MLP\" (colored in green) models the background only. To this end, a cost volume is constructed using images captured from neighboring reference views, guiding the joint optimization of scene MLP and camera parameters, as we discuss in Sec. 4.1. To supervise the training of background MLP and remove occlusions, we propose to aggregate the information from scene MLP about whether the output of background NeRF should be similar to the observed color, by learning a supervision mask from the weights along the ray using a \"mask MLP\" (colored in amber), as explained in Sec. 4.2. The reference and target views constitute the family of neighboring images whose features are warped to construct the cost volume." + ], + "image_footnote": [], + "bbox": [ + 106, + 88, + 869, + 280 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminary", + "text_level": 1, + "bbox": [ + 76, + 400, + 202, + 417 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our goal is to generate an occlusion-free scene representation given a collection of $N$ images $\\mathcal{I} = \\{I_1, I_2, \\ldots, I_N\\}$ captured from $N$ different viewpoints, where the $i$ -th viewpoint is defined by the extrinsic matrix $[\\mathbf{R}_i | \\mathbf{t}_i]$ and the camera is defined by a shared intrinsic matrix $\\mathbf{K}$ . Together they form the camera matrices $\\mathcal{P} = \\{\\mathbf{P}_1, \\mathbf{P}_2, \\ldots, \\mathbf{P}_N\\}$ , where $\\mathbf{P}_i = \\mathbf{K}[\\mathbf{R}_i | \\mathbf{t}_i]$ , as illustrated in the left part of Fig. 1.", + "bbox": [ + 75, + 425, + 468, + 530 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As we follow the approach of NeRF [18], we briefly introduce it in this section for self-contained purposes. NeRF uses a neural network $F_{\\Theta}$ to construct a radiance field, which maps the 3D coordinate of a point $\\mathbf{x} = (x,y,z)$ and viewing direction $\\mathbf{d} = (\\theta ,\\phi)$ into the point's radiance color $\\mathbf{c} = (r,g,b)$ and density $\\sigma$ . Then, the volume rendering scheme is applied to render an image. For each training image and every pixel $\\mathbf{p} = (u,v)$ , a ray is emitted from the camera position $\\mathbf{o}$ at direction", + "bbox": [ + 75, + 531, + 468, + 667 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {d} _ {i} (u, v) = \\mathbf {R} _ {i} \\left[ \\frac {u - W / 2}{f}, - \\frac {v - H / 2}{f}, - 1 \\right] ^ {\\top}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 672, + 468, + 710 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $W, H,$ and $f$ are the width, height, and focal length of the input image, respectively. Marching through the ray $\\mathbf{r}(t) = \\mathbf{o} + t\\mathbf{d}$ and aggregating the colors and densities provide us the final color", + "bbox": [ + 75, + 715, + 468, + 773 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {I} _ {i} (\\mathbf {p}) = \\int_ {t _ {n}} ^ {t _ {f}} T (t) \\sigma (\\mathbf {r} (t)) \\mathbf {c} (\\mathbf {r} (t), \\mathbf {d} _ {i}) d t, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 780, + 468, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $T(t) = \\exp(-\\int_{t_n}^t \\sigma(\\mathbf{r}(s)) ds)$ stands for the accumulated transmittance along the ray, or intuitively, the probability of the emitted radiance of $\\mathbf{r}(t)$ to reach the camera without being blocked by points in between. $t_n$ and $t_f$ are the far end and the near end of rendering respectively.", + "bbox": [ + 75, + 823, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To make the integral tractable, points are sampled along the ray, and Eq. (2) can be rewritten into", + "bbox": [ + 498, + 402, + 890, + 431 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {I} _ {i} (\\mathbf {p}) = \\sum_ {k} ^ {N _ {s}} T _ {k} \\left(1 - \\exp \\left(- \\sigma_ {k} \\delta_ {k}\\right)\\right) \\mathbf {c} \\left(\\mathbf {r} _ {k}, \\mathbf {d}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 440, + 890, + 481 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nT _ {k} = \\exp \\left(- \\sum_ {j = 1} ^ {k - 1} \\sigma_ {j} \\delta_ {j}\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 484, + 890, + 527 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For the convenience of notation, $\\sigma (\\mathbf{r}_k)$ is abbreviated to $\\sigma_{k}$ and $\\delta_{k}$ is defined by $t_{k + 1} - t_k$ ; $N_{s}$ is the number of points sampled on the ray, $\\mathbf{r}_k = \\mathbf{o} + t_k\\mathbf{d}$ denotes the position of $k$ -th sampling point, and $t_k$ refers to the depth of the $k$ -th sampling point seen from the target view.", + "bbox": [ + 496, + 537, + 890, + 612 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "NeRF optimizes the scene representation neural network by minimizing the photometric loss $\\mathcal{L} = \\sum_{i=1}^{N} \\|I_i - \\hat{I}_i\\|_2$ , which is simply the difference between input images $\\mathcal{I}$ and corresponding predicted images $\\hat{\\mathcal{I}}$ . The optimization process can be formulated as:", + "bbox": [ + 496, + 612, + 890, + 686 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\Theta^ {*} = \\arg \\min _ {\\Theta} \\mathcal {L} (\\hat {\\mathcal {I}} | \\mathcal {I}, \\mathcal {P}). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 695, + 890, + 719 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Despite NeRF's strong capability in representing a scene from multiple viewpoints, the occlusions pose challenges to it from two aspects: 1) How to selectively aggregate background information from multiple viewpoints and 2) how to accurately estimate the camera pose of desired background in a continuous space.", + "bbox": [ + 496, + 726, + 890, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Proposed approach", + "text_level": 1, + "bbox": [ + 498, + 829, + 687, + 847 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We focus on how to address the two challenges above when using NeRF to see through occlusion from multiple viewpoints and our pipeline is illustrated in Fig. 1. In", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "20724", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Sec. 4.1, we first describe our joint optimization framework to accurately estimate the camera pose and reconstruct the scene with occlusions using a scene MLP. Then, in Sec. 4.2, we introduce how to selectively aggregate background information from multiple viewpoints based on the rectified background pose using a mask MLP. Under the guidance of mask MLP, we finally train a background MLP to reconstruct occlusion-free scenes. Note that both the scene MLP and the background MLP map the positional encoding of position and viewing direction into color and density, as described in Sec. 3.", + "bbox": [ + 75, + 90, + 472, + 257 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Scene reconstruction with cost volume", + "text_level": 1, + "bbox": [ + 76, + 267, + 406, + 282 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "NeRF and most of its variants use COLMAP [23] as a silver bullet to estimate camera poses for the scene representation, while COLMAP [23] is not an ideal setting for building an occlusion-free representation. The extracted local features cannot be distinguished between undesired occlusion and the desired background, so the matching process can lead to an erroneous correspondence, or even directly fail to estimate the camera parameters. Some recent works try to address the dependency on COLMAP [23] by jointly optimizing camera parameters and the neural radiance fields with the photometric loss [28], achieving comparable performance to COLMAP [23].", + "bbox": [ + 75, + 291, + 468, + 470 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, only using photometric loss does not solve the fundamental problem, because during training, it only uses one view at a time, namely the target view itself. This makes the system ignorant of the 3D scene as a whole, and thus still cannot tell occlusions from the desired background. Besides, in the early stage of training when NeRF can hardly output meaningful information, the photometric loss cannot constrain camera extrinsics to converge to the correct rotation and translation. This can lead to a sub-optimal solution.", + "bbox": [ + 75, + 473, + 468, + 608 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To make our method aware of occlusions and background, we design a cost volume loss inspired by multiview stereo solutions [6, 30]. It encourages the system to model the location corresponding to each pixel at the correct depth, by utilizing the feature consistency of the same point seen from neighboring reference views. Next, we will introduce how our cost volume loss is constructed.", + "bbox": [ + 75, + 608, + 468, + 714 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Feature extraction. A 2D map $F_{i}$ is extracted for every input image $I_{i}$ by using a pre-trained VGG-19 network $\\eta$ as a feature encoder as $F_{i} = \\eta (I_{i})$ . For the convenience of notation, we define $F_{i}(\\mathbf{p})$ as the feature vector at 2D location $\\mathbf{p}$ . When $\\mathbf{p}$ falls off the grid, bilinear interpolation is performed to acquire the feature vector. The extracted 2D features represent local appearance, so intuitively the feature of the same point of interest across neighboring views are assumed to be similar. On the other hand, if two feature vectors vary greatly, it is unlikely that they describe the same area.", + "bbox": [ + 75, + 734, + 470, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Feature volume construction. To construct a feature volume at the $j$ -th viewpoint (target view), the image features from neighboring reference views should be warped with respect to the depth. The warping from the $i$ -th view to the $j$ -th view can be formulated as a $3 \\times 3$ matrix", + "bbox": [ + 496, + 90, + 890, + 167 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {H} _ {i j} (z) = \\mathbf {K R} _ {i} (\\mathbf {I d} - \\frac {\\left(\\mathbf {R} _ {j} ^ {\\top} \\mathbf {t} _ {j} - \\mathbf {R} _ {i} ^ {\\top} \\mathbf {t} _ {i}\\right) \\mathbf {n} _ {j} ^ {\\top}}{z} \\cdot \\mathbf {R} _ {j}) \\mathbf {R} _ {j} ^ {\\top} \\mathbf {K} ^ {- 1}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 500, + 174, + 890, + 219 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{Id}$ is the identity matrix, and $\\mathbf{n}_j$ is the principle axis of the $j$ -th view. This matrix describes a mapping that the point at the 2D coordinate of $\\mathbf{p}$ and the depth of $z$ in warped feature map corresponds to the point at $\\mathbf{H}_{ij}(z)\\mathbf{p}$ in the original feature map. Applying the warping matrix to the feature map $F_{i}$ derives a family of warped image feature maps:", + "bbox": [ + 496, + 220, + 890, + 311 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF _ {i, j, z} (\\mathbf {p}) = F _ {i} \\left(\\mathbf {H} _ {i j} (z) \\mathbf {p}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 320, + 890, + 338 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "which is warped from $F_{i}$ , seen from the $j$ -th view, and at the depth of $z$ .", + "bbox": [ + 496, + 347, + 890, + 376 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We then construct a cost volume at the points which we sample using the warped feature maps. Based on the intuition that similar features mean more likelihood of the same area, a large variance of warped features indicates that features belonging to different areas are warped to the current point. Thus, the variance metric can be used as a clue for scene reconstruction as", + "bbox": [ + 496, + 377, + 890, + 482 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {V} _ {j} (u, v, z) = \\operatorname {V a r} _ {i} \\left(F _ {i, j, z} ([ u, v, 1 ] ^ {\\top})\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 491, + 890, + 511 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\operatorname{Var}_i(\\cdot)$ denotes the variance across $M$ selected neighboring views.", + "bbox": [ + 496, + 518, + 890, + 550 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Joint optimization. To apply the cost volume as supervision for NeRF training, we consider the weight of the $k$ -th sampled point", + "bbox": [ + 496, + 566, + 890, + 613 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nW _ {k} = T _ {k} \\left(1 - \\exp (- \\sigma_ {k})\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 623, + 890, + 640 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "which refers to the contribution of the point at $\\mathbf{r}_t$ to the result $\\hat{I}_j(u,v)$ . We build the scene MLP to reconstruct the scene with occlusion, to provide an intermediate scene representation for further occlusion removal. For a high-quality reconstruction, the scene MLP should be discouraged to assign the points that are inconsistent across neighboring views with a large weight. Based on the observations above, we construct the cost volume loss as", + "bbox": [ + 496, + 648, + 890, + 768 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {c o s t}} = \\sum_ {\\mathbf {p} = (u, v)} \\sum_ {i = 1} ^ {N _ {s}} \\mathbf {V} _ {j} (u, v, t _ {i}) W (u, v, t _ {i}). \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 777, + 890, + 821 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Meanwhile, we also apply the self-consistency loss aiming at scene reconstruction as", + "bbox": [ + 496, + 830, + 890, + 861 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {s c}} = \\sum_ {\\mathbf {p}} \\| \\hat {I} _ {i} (\\mathbf {p}) - I _ {i} (\\mathbf {p}) \\| _ {2} ^ {2}. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 598, + 869, + 890, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "20725", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "So far, we have explained the design of scene reconstruction scheme where occlusions exist, including the key component of cost volume, as depicted in the blue box of Fig. 1. Together, they provide a more robust estimation of the scene and the camera poses when occlusions are present.", + "bbox": [ + 76, + 90, + 468, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Selective supervision scheme", + "text_level": 1, + "bbox": [ + 76, + 185, + 333, + 200 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Existing image-based occlusion removal methods [14, 27] can function properly when a large amount of labeled data is available, by considering the occlusion removal as a mapping from samples with occlusions to their occlusion-free counterparts. Our method is grounded on a totally different underlying logic. By simulating human eyes to observe a specified scenes, our method can aggregate a continuous occlusion-free scene representation from different viewpoints. Then, we can render any occlusion-free viewpoints from this learned scene representation. Such mechanism enables a novel way to remove occlusion in an unsupervised manner. The key problem then becomes how to focus on the desired background only during scene aggregation.", + "bbox": [ + 75, + 210, + 470, + 422 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We achieve this goal based on an intuitive observation: a majority of undesired occlusions always appear in the foreground. However, directly applying depth as supervision is not feasible, since the depth does not necessarily correlate with the presence of occlusion. To this end, we propose to use bidirectional depth inconsistency as a clue. If the expected ray termination depth seen from the camera and that from the other end of the ray show great difference, the ray should pass through some foreground occlusion. Note that \"termination\" in rendering refers to the point that no point behind it contributes to the rendered ray. This assumption follows the physical law, and does not rely on any external prior or excessive training data. Note that this prior is opposite to the shell-shaped geometry assumption by NeRFReN [7], which assumes that the density along the ray peaks out at the surface.", + "bbox": [ + 75, + 426, + 468, + 667 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In implementing this, we design a mask MLP $P_{\\Psi}$ to predict the occluding likelihood $\\tilde{P} (\\mathbf{r})$ based on the weights along the ray, formulated as:", + "bbox": [ + 75, + 670, + 468, + 715 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {P} (\\mathbf {r}) = P _ {\\Psi} \\left(\\left\\{W _ {k} \\right\\} _ {k = 1} ^ {N _ {s}}\\right). \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 733, + 468, + 758 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Based on the intuition discussed above, the mask MLP is trained based on the bidirectional depth inconsistency. The expected termination depth $\\overrightarrow{D}$ and reversed termination depth $\\overline{D}$ is defined by", + "bbox": [ + 75, + 777, + 468, + 844 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\overrightarrow {D} (\\mathbf {r}) = \\sum_ {k = 1} ^ {N _ {s}} W _ {k} t _ {k} = \\sum_ {k = 1} ^ {N _ {s}} T _ {k} \\left(1 - \\exp \\left(- \\sigma_ {k} \\delta_ {k}\\right)\\right) t _ {k}, \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 861, + 468, + 904 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/eb4e446a46cf45c4e43ad3874b15d415d7df0fcad6e26a8386951670036e3d5f.jpg", + "image_caption": [ + "(a) $\\overrightarrow{D}$" + ], + "image_footnote": [], + "bbox": [ + 498, + 87, + 625, + 147 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/92b80fe52eafdc54c201cd30093437f4cdd9ad0e04ba217fc012b95e90715ff0.jpg", + "image_caption": [ + "(b) $\\Delta_D$" + ], + "image_footnote": [], + "bbox": [ + 629, + 88, + 758, + 146 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/12c6a3ddacd7b4ff1b789953d4b44b49af049e21d6624b176308a75dab88893b.jpg", + "image_caption": [ + "(c) $P$", + "Figure 2. An illustration of the depth and mask used for selective supervision (without losing generality, we show the example for fence, and this applies to other types of occlusions as well): (a) estimated depth $(\\vec{D})$ of the scene with occlusion, (b) the bidirectional depth difference $(\\Delta_D)$ , and (c) the supervision mask $(P)$ . Note that the white regions in (c) denote the desired background details that need supervision and undesired foreground occlusions are marked in black. Please visit our project page for animated results." + ], + "image_footnote": [], + "bbox": [ + 759, + 87, + 890, + 147 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and", + "bbox": [ + 500, + 327, + 529, + 340 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\overleftarrow {D} (\\mathbf {r}) = \\sum_ {k = 1} ^ {N _ {s}} \\overleftarrow {W} _ {k} t _ {k} = \\sum_ {k = 1} ^ {N _ {s}} \\overleftarrow {T} _ {k} \\left(1 - \\exp \\left(- \\sigma_ {k} \\delta_ {k}\\right)\\right) t _ {k}, \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 348, + 890, + 388 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "respectively, where $\\overline{T}_k = \\exp (-\\sum_{j > k}\\sigma_j\\delta_j)$ is the reversed accumulated transmittance. An example of $\\vec{D}$ can be found in Fig. 2(a). We train the mask MLP based on the prior described before, that $P(\\mathbf{r})$ correlates with $\\Delta_D(\\mathbf{r}) = \\overleftarrow{D} -\\overrightarrow{D}$ , by", + "bbox": [ + 496, + 400, + 890, + 488 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {m a s k}} = - \\cos \\left(\\left\\{P (\\mathbf {r}) - \\alpha \\right\\} _ {\\mathbf {r} \\in \\mathcal {R}}, \\left\\{\\Delta_ {D} (\\mathbf {r}) - \\overline {{\\Delta_ {D}}} \\right\\} _ {\\mathbf {r} \\in \\mathcal {R}}\\right), \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 494, + 890, + 526 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{R}$ is the patch of rays for training, $\\overline{\\Delta_D}$ is the mean value of $\\Delta_D$ across the entire input view, and $\\cos (\\mathbf{a},\\mathbf{b})$ is the cosine similarity between a and b. We show an example of $\\Delta_D$ in Fig. 2(b). $\\alpha$ is a pre-defined threshold for the binary classification of $\\tilde{P}$ , deriving a binary supervision mask $P$ by", + "bbox": [ + 496, + 527, + 890, + 616 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nP (\\mathbf {r}) = \\left\\{ \\begin{array}{l l} 0, & \\tilde {P} (\\mathbf {r}) < \\alpha \\\\ 1, & \\tilde {P} (\\mathbf {r}) \\geq \\alpha \\end{array} \\right.. \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 616, + 890, + 650 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Fig. 2(c), the obtained binary mask $P$ can clearly distinguish the desired background to be kept and the undesired occlusion to be removed.", + "bbox": [ + 496, + 654, + 890, + 698 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We further train a background MLP which reconstructs the scene belonging to the background only. Specifically, we only supervise the rendered color of background MLP to be similar to the input image in the background region where $P(\\mathbf{r}) = 1$ .", + "bbox": [ + 496, + 699, + 890, + 775 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {b g}} = \\sum_ {\\mathbf {p}} \\left\\| \\left(\\left(\\hat {I} _ {i} ^ {\\prime} (\\mathbf {p}) - I _ {i} (\\mathbf {p})\\right) \\odot P (\\mathbf {r}) \\right\\| _ {2} ^ {2} \\right. \\tag {17} \\\\ + \\sum_ {\\mathbf {p}} \\left\\| (\\vec {D} ^ {\\prime} (\\mathbf {r}) - \\vec {D} (\\mathbf {r})) \\odot P (\\mathbf {r}) \\right\\| _ {1}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 782, + 890, + 857 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\hat{I}_i^{\\prime}(\\mathbf{p})$ and $\\vec{D}^{\\prime}(\\mathbf{r})$ denote the rendered color and expected ray termination depth of the background MLP.", + "bbox": [ + 496, + 867, + 890, + 901 + ], + "page_idx": 4 + }, + { + "type": "footer", + "text": "20726", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With background MLP holding the occlusion-free scene representation and the scene MLP containing all the information including occlusions, we have finally gathered all the pieces of the puzzle.", + "bbox": [ + 76, + 90, + 468, + 151 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Implementation details", + "text_level": 1, + "bbox": [ + 76, + 159, + 290, + 175 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We implement our method using PyTorch. The scene MLP and the background MLP, each having 256 channels, consist of 6 layers and 8 layers respectively, which predict the colors and densities corresponding to the whole scene and the background separately. The mask MLP is a 2-layer MLP with 64 channels. When constructing cost volume, $M = 4$ neighboring views are involved to calculate variance in Eq. (8). In the training phase, the scene MLP is jointly optimized with camera parameters of each scene, after being trained coarsely to fit the scene from COLMAP [23] camera pose or identity matrix initialization. All of the trainable parameters are optimized per scene. We train scene MLP for the first $20\\%$ of iterations, and then train the system end-to-end.", + "bbox": [ + 75, + 183, + 468, + 393 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "On each iteration, we evenly sample 128 points along each ray. A batch contains random 1024 pixels in one of the images. We use the Adam optimizer with defaults values $\\beta_{1} = 0.999$ , $\\beta_{2} = 0.9$ , $\\epsilon = 10^{-8}$ , and a learning rate $10^{-3}$ that decays following the cosine scheduler [15] during the optimization. If initialized with identity matrix, the model is trained for about 150K iterations. If a coarse COLMAP [23] estimation is available, our model converges after about 50K iterations.", + "bbox": [ + 75, + 393, + 470, + 529 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 542, + 209, + 559 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset. To validate the effectiveness of our method, we collect an evaluation dataset containing 10 different scenes, covering various types of occlusions. The dataset not only contains multiple sparse viewpoints of occlusion scenes collected by ourselves, but also samples selected from existing datasets [29,31], to increase the diversity from different capturing conditions. Specifically, FENCE1, SCRIBBLE2 and RAINDROP are adopted from [29], and WIRE2 is adopted from [31]. The rest 6 of the scenes are captured by ourselves using a Sony $\\alpha 7$ III camera or an iPhone 12. For easy reference, we name each set of data using the occlusion type, as shown in Fig. 3.", + "bbox": [ + 75, + 568, + 468, + 750 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For our newly captured samples, we take 20 to 60 images with different poses for each scene, where $85\\%$ of the images are used for training our NeRF-based method, and the rest $15\\%$ for testing. This is a commonly adopted setting by many NeRF-based methods [7, 8, 18]. To facilitate the quantitative evaluation, one of newly captured scene is with ground truth, namely SCRIBBLE1. For this scene, we place a piece of glass with scribbles between the desired background scene and the camera. Then, a multi-view sparse image set is acquired by moving cameras in front of the", + "bbox": [ + 75, + 750, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0e5ca29bd3105841712d3adbb3e1ee3ba95cff7e81b5d343337c1189007680a0.jpg", + "image_caption": [ + "(a) FENCE1" + ], + "image_footnote": [], + "bbox": [ + 506, + 87, + 584, + 123 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e02365c55347ee7aa9fde4c72d4b709ec0293cce250e439385f8c46bf02bc6b7.jpg", + "image_caption": [ + "(b) FENCE2" + ], + "image_footnote": [], + "bbox": [ + 588, + 88, + 665, + 123 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c64a71041da4182430496bf7ad9638810e88de1c92464a50711c22f010c29a89.jpg", + "image_caption": [ + "(c) FENCE3", + "(d) SCRIBBLE1(e) SCRIBBLE2" + ], + "image_footnote": [], + "bbox": [ + 666, + 88, + 743, + 123 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/cf4538b29a225a63eb2c3871ea55d442076a44cf815f4e39dec1b081a9af6161.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 746, + 88, + 823, + 123 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f1c24814e9e2aed511e37de84e89b5403e6c14817360a740a704bcc4f18efdba.jpg", + "image_caption": [ + "(f) SCRIBBLE3 (g) RAINDROP" + ], + "image_footnote": [], + "bbox": [ + 506, + 137, + 584, + 171 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3b77cbb4c76bfa9bea5693ea7c6d14a848223677acd413721ef64a759818c063.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 588, + 138, + 665, + 171 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6ad0eb2b3708ac578af25ec332fbd713543174fa5daf73d6c08a5756ceb28d91.jpg", + "image_caption": [ + "(h) STATUE", + "Figure 3. From (a) to (j), we provide one sample image for each scene in our evaluation dataset. The names of the scenes are listed below the image indicating occlusion type." + ], + "image_footnote": [], + "bbox": [ + 666, + 138, + 743, + 171 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9ae9ce1288252b7c3c4299c12b574b290c70653e9a688a5c635d8a9669d481c0.jpg", + "image_caption": [ + "(i) WIRE1" + ], + "image_footnote": [], + "bbox": [ + 746, + 137, + 823, + 171 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a5bc464626ba89cfe590c2a67dfe6094684991291969b717cce5138f9104a535.jpg", + "image_caption": [ + "(j) WIRE2" + ], + "image_footnote": [], + "bbox": [ + 826, + 138, + 903, + 171 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "glass. To obtain the ground truth for validation, we remove the glass while taking the first and last view of the scene, so that these two views can come with a ground truth.", + "bbox": [ + 498, + 260, + 892, + 306 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9977cba7993c2e8977713ddf7ba885844ffa332a310d1e00eade623c2e95e617.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 509, + 316, + 638, + 375 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/63c631172c06aa5518842d4a94993172fae483e8130829ce6ca5c120248b41b2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 640, + 318, + 717, + 375 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5d0a1c60494e6e082406595c259a21fbee1cafc2a85b3fb4f030974222bd988b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 718, + 318, + 897, + 375 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c8ba6650eee6059e3e2faa5540dafae4517a1dc3f5adca1c70fdb2c90021de4a.jpg", + "image_caption": [ + "(b) PWC-Net+NeRF [14, 18]" + ], + "image_footnote": [], + "bbox": [ + 509, + 391, + 638, + 448 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/af24594123f5cc33ade3c64b25eed1b0cb1da24d45284641be34d5530108166a.jpg", + "image_caption": [ + "(a) Vanilla NeRF [18]" + ], + "image_footnote": [], + "bbox": [ + 640, + 391, + 717, + 448 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/92823486a0237d025b5e1787158b617e0ddcb52154196aa3b3038b766ba51e6d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 718, + 391, + 897, + 448 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2e37d99c93a2ef58e37cd5f64baf04b40fba2498bd2046fda159be2fcd3833f8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 509, + 464, + 638, + 520 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/88db9acb742ca029104ff629418cecd601dbe789367c1a37f2941f53779f7daf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 640, + 464, + 717, + 520 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/92bd0903dcd9d070ce2168b1ca3e7e06853233f030f550a99e043c80cd291da3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 718, + 464, + 897, + 520 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6daed4086598c7e2a322fecd5428c7fc7ed1b117e8599e596ead58853120b9d3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 509, + 536, + 638, + 593 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8d82aa719327b598dca1ea1f571e3e2a01b1e789d7d65c754439f4463c249963.jpg", + "image_caption": [ + "(c) Ha-NeRF [3]", + "(d) NeRF-W [17]" + ], + "image_footnote": [], + "bbox": [ + 640, + 536, + 717, + 593 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/20e7c9a3736e00a234af0ff74905fa352cbdfababa451d1857dbe5492ae1ad54.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 718, + 536, + 897, + 593 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e54464a78a2ef751d1954c47cfd1c8456072d5b55a3c3871216a34f5d81ac8a9.jpg", + "image_caption": [ + "Figure 4. Visual quality comparisons for occlusion removal. From left to right, we show animated novel view synthesis on FENCE1, STATUE, and WIRE1. From (a) to (e), we show results obtained by (a) vanilla NeRF [18], (b) PWC-Net + NeRF [14, 18], (c) Ha-NeRF [3], (d) NeRF-W [17], and (e) our method. Please visit our project page for animated results. Some results show stronger variation over the animation due to failure in extracting consistent background." + ], + "image_footnote": [], + "bbox": [ + 509, + 608, + 638, + 665 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c9da8cded0f8478f5cf866e92d83f853313fc7622468c87462f59d95a8af4cdc.jpg", + "image_caption": [ + "(e) Our method" + ], + "image_footnote": [], + "bbox": [ + 640, + 609, + 717, + 665 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/732cd18bbef128dd1f5623c456a7de2af32866f6ad2c456bf0cf8776884a62dd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 718, + 609, + 897, + 665 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. To our knowledge, there is no similar work that construct an occlusion-free scene based on NeRF representation. We design three baselines for comparison by considering both the contributions from occlusion removal and NeRF to verify our performance advantage. 1) A state-of", + "bbox": [ + 496, + 825, + 892, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "20727", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/14f17c09f2df42f6f9916e6e9d9ec3b7ff253d79be50b29b5cd6c75c6ad6971b.jpg", + "table_caption": [ + "Table 1. Quantitative comparison results. $\\uparrow (\\downarrow)$ indicates larger (smaller) values are better, and bold font indicates the best results." + ], + "table_footnote": [], + "table_body": "
MethodPSNR↑SSIM↑LPIPS↓
PWC-Net+NeRF [14,18]18.250.790.45
NeRF-W [17]11.060.450.61
Ha-NeRF [3]15.410.700.41
Proposed19.390.830.37
", + "bbox": [ + 78, + 122, + 470, + 204 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dda4b2d506b49a7a9d5204b5a6a4484b967e6a89ac4d7d7316fc9814eb44b471.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 215, + 187, + 273 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d9b881336a0ad94df19a9a180ddd72abfda0e3ce6d65325a603801ef59b3bcdf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 189, + 215, + 336, + 273 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d0efb166d2b78e6326e50be32c8e55cb89d94149659bf44b36f4a5712e42012d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 336, + 215, + 467, + 273 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cf7fccab72517023724dcd45ddf735aa9232a87481d760efb4746185e41685e0.jpg", + "image_caption": [ + "(a)", + "(d)", + "Figure 5. Visual quality comparisons for occlusion removal on SCRIBBLE1 scene. (a) is one image in the validation set with occlusion and (f) is ground truth of this view) with results from: (b) results obtained by Ha-NeRF [3], (c) results obtained by NeRF-W [17], (d) results obtained by occlusion removal [14] + NeRF [18], and (e) our method." + ], + "image_footnote": [], + "bbox": [ + 80, + 289, + 187, + 345 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/55634547e8c938be4d0af2152bf4e49abc91c15ecbac6b4a3ee7e2aff6fbcf5f.jpg", + "image_caption": [ + "(b)", + "(e)" + ], + "image_footnote": [], + "bbox": [ + 189, + 289, + 336, + 345 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/aae097cf4956e10dd9833088135f0d97fb6542d569e218daf35766aeec7f88df.jpg", + "image_caption": [ + "(c)", + "(f)" + ], + "image_footnote": [], + "bbox": [ + 336, + 289, + 467, + 345 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the-art occlusion removal method for image sequences or videos (referred to as PWC-Net [14] based on the main pillar of the work) + NeRF [18]: In this setting, PWC-Net [14] serves as a pre-processing tool to eliminate occlusion before training the NeRF [18]. 2) NeRF-W [17]: This is a method designed for separating a complete scene into two components. 3) Ha-NeRF [3]: This is another method that decomposes the scenes into distinct parts. We compare those settings with our method on our evaluation dataset, as listed in Fig. 3. For fairness of comparison, we evenly sample 128 points along the ray without applying importance sampling. We also keep the trainable parameters in each baseline approximately the same with the proposed method.", + "bbox": [ + 75, + 472, + 468, + 667 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation methodology. We conduct both qualitative and quantitative evaluations on our dataset. For quantitative evaluation, we compare reconstruction quality of the estimated occlusion-free images for viewpoints with corresponding ground truth by PSNR, SSIM, and LPIPS [32].", + "bbox": [ + 75, + 686, + 468, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Qualitative evaluation", + "text_level": 1, + "bbox": [ + 76, + 771, + 282, + 787 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The qualitative results of the rendered novel occlusion-free views are presented in Figs. 4 and 5. In Fig. 4, we use the scenes of FENCE1, STATUE and WIRE1 in our dataset to show that our method can reliably reconstruct novel background views, and it works no matter whether the occlusions cover a large area or have an irregular shape. In Fig. 5, we take one frame in SCRIBBLE1 with ground truth to show", + "bbox": [ + 75, + 794, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a72071363551b10fae0b0e46b7797e3f73808acb0a43cea0c9afb5cf3af8cf03.jpg", + "image_caption": [ + "(a)", + "Figure 6. Visual quality comparisons for occlusion removal on RAINDROP scene, with results from (a) an image-based raindrop removal method [22], (b) an image-based occlusion removal method [14], and (c) our method. Our method outperforms the baselines in the highlighted red boxes." + ], + "image_footnote": [], + "bbox": [ + 506, + 89, + 630, + 258 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/969a8f43eee4fc0ac4fb4c07c19c257abeee61c15007018b759ddd91e14842ae.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 630, + 89, + 758, + 258 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f8c4977a011adc2138b3b87bc242faadf36c5521f0dcfc707e82aade4b072453.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 759, + 89, + 883, + 258 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the fidelity of our results.", + "bbox": [ + 500, + 366, + 668, + 380 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "PWC-Net + NeRF [14, 18] only achieves comparable results in FENCE1 where the occlusions are regularly shaped, but shows poor artifacts on unseen types of occlusions, such as the irregularly-shaped STATUE and WIRE1, as shown in Fig. 4(b). This shows their reliance on training data. Besides, the occlusion removal method can lead to an oversmoothing issue which undermines the reconstruction accuracy. Meanwhile, the results from Ha-NeRF [3] and NeRF-W [17] in Figs. 4(c) and 4(d) either shows degenerated details or still preserves occlusions, which verifies our claim that they cannot handle the consistent occlusions observed in a continuous image sequence. The results from our method in Fig. 4(e) clearly remove these occlusions with different shapes and distributions.", + "bbox": [ + 496, + 381, + 890, + 592 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our method relies on an accurate separation between foreground and background, but the assumption may be violated if obstructions are semi-transparent. We conduct experiments on the RAINDROP scene in our validation dataset to evaluate the robustness of our method under such scenario. With the results displayed in Fig. 6, our method can also effectively remove the raindrop in this example. It achieves competitive or even better results in some regions when compared with a method specifically designed for raindrop removal [22] as well as PWC-Net [14] designed to remove occlusions including raindrop.", + "bbox": [ + 496, + 594, + 890, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Quantitative evaluation", + "text_level": 1, + "bbox": [ + 500, + 771, + 715, + 787 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The quantitative results of the occlusion-free novel view in Tab. 1 validate the observation in Figs. 4 and 5. Higher PSNR values show that our method can render occlusion-free novel views and recover the color information with higher accuracy. Higher SSIM values indicate that our method can preserve the structural information with high-frequency details. Lower LPIPS values show that recovered", + "bbox": [ + 496, + 794, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "20728", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/125cc2b668633ea76581f9e024f833d6ae82776d17bf55f25f1dcc9adecff15b.jpg", + "table_caption": [ + "Table 2. Ablation study results aggregated by the types of occlusions. We compare the rendered results of the scene MLP with NeRF [18] using COLMAP [23] as camera parameter estimation and NeRF-- [28] initialized with COLMAP camera parameters." + ], + "table_footnote": [], + "table_body": "
MetricsPSNR ↑ / SSIM ↑ / LPIPS↓
MethodCOLMAP [23]NeRF-- [28]Scene MLP
DatasetFENCE{1,2,3}24.01 / 0.73 / 0.3818.67 / 0.54 / 0.5125.46 / 0.72 / 0.35
SCRIBBLE{1,2,3}22.70 / 0.78 / 0.3117.98 / 0.52 / 0.4623.85 / 0.84 / 0.34
WIRES{1,2}26.21 / 0.93 / 0.2022.28 / 0.88 / 0.2534.44 / 0.95 / 0.23
RAINDROP27.69 / 0.87 / 0.2627.78 / 0.83 / 0.2528.69 / 0.86 / 0.25
STATUE27.53 / 0.84 / 0.3318.49 / 0.56 / 0.4927.08 / 0.80 / 0.37
", + "bbox": [ + 169, + 128, + 802, + 242 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "images by our method better aligns with human perception.", + "bbox": [ + 76, + 267, + 467, + 282 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 292, + 230, + 309 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our method consists of two major components: joint optimization for pose refinement and a selective supervision strategy to eliminate occlusions. Since the selective supervision scheme largely bases on the intermediate information of scene reconstruction MLP, a high-quality reconstruction of the scene with occlusion is crucial to our performance.", + "bbox": [ + 75, + 316, + 467, + 407 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In the ablation study, we first replace pose refinement with COLMAP [23]. From the results shown in Fig. 7(b), the occlusions largely remain, because the fixed erroneous camera parameters makes distinguishing scene components from scene MLP infeasible.", + "bbox": [ + 75, + 407, + 467, + 482 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We also explore the results without the selective supervision scheme, while maintaining pose refinement, namely only output of the scene MLP. In this setting, our method cannot remove occlusions, but it can reconstruct the scene containing occlusions with a high fidelity, as shown in Fig. 7(c). It can also reconstruct the scene with higher fidelity than NeRF based on COLMAP [23], as shown quantitatively in Tab. 2. Our complete model shows the best results for majority of scenes in our dataset.", + "bbox": [ + 75, + 484, + 467, + 619 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bf5ee8ca67981648d22d756febd1b56846ecbb79837458b809d70a1cbfc438ef.jpg", + "image_caption": [ + "(a)", + "Figure 7. Visual comparisons for the ablation study on the FENCE1 scene with results from: (a) our complete method, (b) the method without joint optimization, and (c) the method without selective supervisions. Please visit our project page for animated results." + ], + "image_footnote": [], + "bbox": [ + 84, + 633, + 209, + 689 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d31575ea6e7a88fb837fe24bd44153cdd1d6e2f395605b26d9aded90c44f98d1.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 210, + 633, + 336, + 689 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1dfb77ab3bbda8f496972a0ad9f7dd965e64cb88cfe070a2b127fde2291c88e2.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 336, + 633, + 464, + 689 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusions", + "text_level": 1, + "bbox": [ + 76, + 799, + 204, + 814 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, grounded on Neural Radiance Fields, we construct an occlusion-free scene representation, which is capable of rendering occlusion-free images with desired viewpoints. By directly using the viewpoints as a part of the input, our method is able to aggregate occluded details", + "bbox": [ + 75, + 825, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1c7c7da4d783196c6225b0bffd6047fb0b3026b1d59cf2f64fda5258cd56bebf.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 508, + 258, + 633, + 323 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1bc9fc9c305487a108c62887045dfe07e7fb6643ce2b5fe2e6b6dac8edbbb849.jpg", + "image_caption": [ + "(b)", + "Figure 8. A failure scene of the proposed method, when reflection is present as occlusion. (a) input sample, (b) results of the scene MLP, (c) failed reflection removal with the background MLP." + ], + "image_footnote": [], + "bbox": [ + 637, + 258, + 761, + 323 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7d15b9f0b0151f2ba1d655aef24dfa0fdb904335e48fe224fb04d46cafae02e7.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 764, + 258, + 888, + 323 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "from multiple viewpoints without relying on any external guidance. We propose a pose refinement scheme to ensure robust training when the camera poses cannot be accurately estimated by COLMAP [23]. To focus on modeling of desired background scenes within a specific viewpoint, we further introduce a depth constraint to probe the occluded areas by measuring the depth of occlusion and background. Experimental results show that our method is capable of achieving promising results under a variety of scenarios.", + "bbox": [ + 498, + 406, + 890, + 542 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. Despite the promising performance of our proposed method, several limitations are still to be addressed in our future study. First, our method clearly assumes the undesired layers are in the foreground. It cannot disambiguate the cases where details from different layers are intertwined, such as scenes with reflection, which is especially challenging due to its semi-transparency and widespread presence. We show a failure case of the proposed method in a scene with reflection as occlusion in Fig. 8. Besides, since several variants of NeRF have been proposed for faster convergence and more accurate reconstruction [5,20], advanced NeRF variants may boost the performance of the proposed method. In addition, we also hope to reduce the number of images required for joint optimization of camera parameters and NeRF in our future work.", + "bbox": [ + 498, + 560, + 890, + 787 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. This work is supported by National Natural Science Foundation of China under Grant No. 62136001, 62088102. Renjie Wan is supported by the Blue Sky Research Fund of HKBU under Grant No. BSRF/21-22/16 and Guangdong Basic and Applied Basic Research Foundation under Grant No. 2022A1515110692.", + "bbox": [ + 498, + 805, + 890, + 895 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "20729", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Coloma Ballester, Marcelo Bertalmio, Vicent Caseles, Guillermo Sapiro, and Joan Verdera. Filling-in by joint interpolation of vector fields and gray levels. IEEE Transactions on Image Processing, 2001. 2", + "[2] Marcelo Bertalmio, Guillermo Sapiro, Vincent Caseles, and Coloma Ballester. Image inpainting. In Proc. of ACM SIGGRAPH, 2000. 2", + "[3] Xingyu Chen, Qi Zhang, Xiaoyu Li, Yue Chen, Ying Feng, Xuan Wang, and Jue Wang. Hallucinated neural radiance fields in the wild. In CVPR, 2022. 1, 2, 6, 7", + "[4] Chen Du, Byeongkeun Kang, Zheng Xu, Ji Dai, and Truong Nguyen. Accurate and efficient video de-fencing using convolutional neural networks and temporal information. In Proc. of International Conference on Multimedia and Expo, 2018. 1, 2", + "[5] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In CVPR, 2022. 8", + "[6] Xiaodong Gu, Zhiwen Fan, Siyu Zhu, Zuozhuo Dai, Feitong Tan, and Ping Tan. Cascade cost volume for high-resolution multi-view stereo and stereo matching. In CVPR, 2020. 4", + "[7] Yuan-Chen Guo, Di Kang, Linchao Bao, Yu He, and Song-Hai Zhang. NeRFReN: Neural radiance fields with reflections. In CVPR, 2022. 1, 2, 5, 6", + "[8] Xin Huang, Qi Zhang, Ying Feng, Hongdong Li, Xuan Wang, and Qing Wang. HDR-NeRF: High dynamic range neural radiance fields. In CVPR, 2022. 2, 6", + "[9] Jeffrey Ichnowski, Yahav Avigal, Justin Kerr, and Ken Goldberg. Dex-NeRF: Using a neural radiance field to grasp transparent objects. In Proc. of Conference on Robot Learning, 2021. 2", + "[10] Sankaraganesh Jonna, Sukla Satapathy, and Rajiv R Sahay. Stereo image de-fencing using smartphones. In Proc. of International Conference on Acoustics, Speech, and Signal Processing, 2017. 1, 2", + "[11] Jingyuan Li, Ning Wang, Lefei Zhang, Bo Du, and Dacheng Tao. Recurrent feature reasoning for image inpainting. In CVPR, 2020. 2", + "[12] Xiaoyu Li, Bo Zhang, Jing Liao, and Pedro V Sander. Let's see clearly: Contaminant artifact removal for moving cameras. In Proc. of International Conference on Computer Vision, 2021. 1", + "[13] Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. BARF: Bundle-adjusting neural radiance fields. In CVPR, 2021. 2", + "[14] Yu-Lun Liu, Wei-Sheng Lai, Ming-Hsuan Yang, Yung-Yu Chuang, and Jia-Bin Huang. Learning to see through obstructions. In CVPR, 2020. 1, 2, 5, 6, 7", + "[15] Ilya Loshchilov and Frank Hutter. SGDR: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 6", + "[16] Li Ma, Xiaoyu Li, Jing Liao, Qi Zhang, Xuan Wang, Jue Wang, and Pedro V Sander. Deblur-NeRF: Neural radiance fields from blurry images. In CVPR, 2022. 2" + ], + "bbox": [ + 78, + 116, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[17] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. NeRF in the Wild: Neural radiance fields for unconstrained photo collections. In CVPR, 2021. 1, 2, 6, 7", + "[18] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In Proc. of European Conference on Computer Vision, 2020. 1, 2, 3, 6, 7, 8", + "[19] Yadong Mu, Wei Liu, and Shuicheng Yan. Video de-fencing. IEEE Transactions on Circuits and Systems for Video Technology, 2014. 1, 2", + "[20] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. TOG, 2022. 8", + "[21] Kamyar Nazeri, Eric Ng, Tony Joseph, Faisal Z Qureshi, and Mehran Ebrahimi. EdgeConnect: Generative image inpainting with adversarial edge learning. arXiv preprint arXiv:1901.00212, 2019. 2", + "[22] Rui Qian, Robby T Tan, Wenhan Yang, Jiajun Su, and Jiaying Liu. Attentive generative adversarial network for raindrop removal from a single image. In CVPR, 2018. 1, 7", + "[23] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In CVPR, 2016. 2, 4, 6, 8", + "[24] Yuhang Song, Chao Yang, Yeji Shen, Peng Wang, Qin Huang, and C-C Jay Kuo. SPG-Net: Segmentation prediction and guidance network for image inpainting. In Proc. of British Machine Vision Conference, 2018. 2", + "[25] V. Vaish, M. Levoy, R. Szeliski, C.L. Zitnick, and Sing Bing Kang. Reconstructing occluded surfaces using synthetic apertures: Stereo, focus and robust measures. In CVPR, 2006. 2", + "[26] Suhani Vora, Noha Radwan, Klaus Greff, Henning Meyer, Kyle Genova, Mehdi SM Sajjadi, Etienne Pot, Andrea Tagliasacchi, and Daniel Duckworth. NeSF: Neural semantic fields for generalizable semantic segmentation of 3d scenes. arXiv preprint arXiv:2111.13260, 2021. 2", + "[27] Renjie Wan, Boxin Shi, Haoliang Li, Ling-Yu Duan, and Alex C Kot. Face image reflection removal. International Journal of Computer Vision, 2021. 1, 5", + "[28] Zirui Wang, Shangzhe Wu, Weidi Xie, Min Chen, and Victor Adrian Prisacariu. NeRF---: Neural radiance fields without known camera parameters. arXiv preprint arXiv:2102.07064, 2021. 2, 4, 8", + "[29] Tianfan Xue, Michael Rubinstein, Ce Liu, and William T Freeman. A computational approach for obstruction-free photography. TOG, 2015. 1, 2, 6", + "[30] Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. MVSNet: Depth inference for unstructured multiview stereo. In Proc. of European Conference on Computer Vision, 2018. 4", + "[31] Renjiao Yi, Jue Wang, and Ping Tan. Automatic fence segmentation in videos of dynamic scenes. In CVPR, 2016. 6", + "[32] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 7" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "20730", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[33] Shuaifeng Zhi, Tristan Laidlow, Stefan Leutenegger, and Andrew J Davison. In-place scene labelling and understanding with implicit scene representation. In Proc. of International Conference on Computer Vision, 2021. 2", + "[34] Manyu Zhu, Dongliang He, Xin Li, Chao Li, Fu Li, Xiao Liu, Errui Ding, and Zhaoxiang Zhang. Image inpainting by end-to-end cascaded refinement with mask awareness. IEEE Transactions on Image Processing, 2021. 2" + ], + "bbox": [ + 78, + 90, + 468, + 204 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "20731", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Occlusion-Free Scene Recovery via Neural Radiance Fields/223bdb66-a260-46a1-86eb-891a595d848a_model.json b/2023/Occlusion-Free Scene Recovery via Neural Radiance Fields/223bdb66-a260-46a1-86eb-891a595d848a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3e06b20acacf417a126553288abba60d3e32f629 --- /dev/null +++ b/2023/Occlusion-Free Scene Recovery via Neural Radiance Fields/223bdb66-a260-46a1-86eb-891a595d848a_model.json @@ -0,0 +1,2717 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.187, + 0.131, + 0.783, + 0.152 + ], + "angle": 0, + "content": "Occlusion-Free Scene Recovery via Neural Radiance Fields" + }, + { + "type": "text", + "bbox": [ + 0.222, + 0.18, + 0.752, + 0.2 + ], + "angle": 0, + "content": "Chengxuan Zhu\\(^{1,2}\\) Renjie Wan\\(^{3}\\) Yunkai Tang\\(^{1,2}\\) Boxin Shi\\(^{1,2*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.126, + 0.2, + 0.841, + 0.27 + ], + "angle": 0, + "content": "\\(^{1}\\)National Engineering Research Center of Visual Technology, School of Computer Science, Peking University \\(^{2}\\)National Key Laboratory for Multimedia Information Processing, School of Computer Science, Peking University \\(^{3}\\)Department of Computer Science, Hong Kong Baptist University {peterzhu, shiboxin}@pku.edu.cn, renjiewan@hkbu.edu.hk, tangyunkai@stu.pku.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.305, + 0.314, + 0.321 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.337, + 0.474, + 0.624 + ], + "angle": 0, + "content": "Our everyday lives are filled with occlusions that we strive to see through. By aggregating desired background information from different viewpoints, we can easily eliminate such occlusions without any external occlusion-free supervision. Though several occlusion removal methods have been proposed to empower machine vision systems with such ability, their performances are still unsatisfactory due to reliance on external supervision. We propose a novel method for occlusion removal by directly building a mapping between position and viewing angles and the corresponding occlusion-free scene details leveraging Neural Radiance Fields (NeRF). We also develop an effective scheme to jointly optimize camera parameters and scene reconstruction when occlusions are present. An additional depth constraint is applied to supervise the entire optimization without labeled external data for training. The experimental results on existing and newly collected datasets validate the effectiveness of our method. Our project page: https://freebutuselessssoul.github.io/occnerf." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.669, + 0.21, + 0.685 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.694, + 0.47, + 0.846 + ], + "angle": 0, + "content": "Neural Radiance Fields (NeRF) are capable of learning the scene representation implicitly from a set of 2D images, yet not every scene is favored by observers. Many undesirable occlusions in our world obscure details that are essential to our understanding of the world. In general, such obstructions range from water droplets and scribbles on a piece of glass, to fences or any objects occluding the desired scenes (e.g., a statue closer to the camera in a landmark scene). How to apply computational methods to exclude them from the scene representation is of great interest." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.847, + 0.47, + 0.877 + ], + "angle": 0, + "content": "Occlusion removal (e.g., [29]) is the direct solution to achieve this goal. However, explicit occlusion removal may" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.306, + 0.895, + 0.578 + ], + "angle": 0, + "content": "oversmooth essential details necessary for clearly observing the desired background scenes. In addition, current methods mainly depend on external occlusion-free supervision (e.g., fence removal [4], raindrop removal [22]) to develop the reliable capability in removing certain types of occlusions. Once encountering a new scenario with unseen occlusion types beyond their training data, these methods might show degraded performances. To handle more diverse types of occlusions, generic constraints from multiple viewpoints are widely adopted [4, 10, 12, 14, 19, 27, 29] via mimicking our human vision systems, who can easily piece together the desired background scenes by looking at them from different viewpoints. But the majority of these methods just consider viewpoints as a prior in relation to spatial correlations. Their backbones still rely on external training data with corresponding ground truth for optimization, which still does not fundamentally alleviate the difficulty of handling diverse occlusions in the real world." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.895, + 0.903 + ], + "angle": 0, + "content": "An occlusion-free world can be progressively aggregated by seeing its occluded part from different viewing directions to reveal occlusions previously unobservable in each single perspective, as illustrated in the left part of Fig. 1 (the fan is occluded in the target view). Since NeRF [18] employs an implicit representation to map viewpoints to pixels, one may come to the naive solution of directly constructing a NeRF which is optimized across multiple viewpoints. However, the vanilla NeRF [18] representing the scene as a whole is not able to treat occlusion and background scenes distinctively, and as long as the occlusion remains static, NeRF is designed to faithfully reconstruct its presence. Meanwhile, many NeRF variants can decompose the whole scene into different components (e.g., NeRF-W [17], Ha-NeRF [3], NeRFRen [7]), but they cannot handle the real-world static occlusions. This is because NeRF-W [17] and Ha-NeRF [3] rely on the inconsistency of undesired components across different views to achieve such separation, which is difficult to be observed in a continuous 3D world. On the other hand, NeRFRen [7] only works in separating the transmission and reflection compo" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.222, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20722" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.468, + 0.121 + ], + "angle": 0, + "content": "nents caused by semi-transparent planar glass, which is incapable of handling opaque occlusions in the real world." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.122, + 0.468, + 0.317 + ], + "angle": 0, + "content": "Another problem comes from NeRF's reliance on camera parameters pre-computed by COLMAP [23], because handcrafted features extracted and matched using COLMAP [23] are for the whole scene, and are incapable in distinguishing between undesired occlusions and the desired background. When the features from occlusion dominate the matching process, the obtained camera parameters cannot faithfully model the spatial correlation of background scenes across multiple viewpoints. Besides, COLMAP [23] is not a stable option for pose estimation in the real world [28]. The existence of occlusions may prevent it from working properly, making the occlusion-free scene representation infeasible." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.318, + 0.468, + 0.633 + ], + "angle": 0, + "content": "In this paper, we aim at seeing through the occluded scenes by developing an occlusion-free scene representation without considering specific occlusion types, based on which we can render any occlusion-free images from desired viewpoints. Our method first maps viewing angles and their corresponding scene details by leveraging NeRF. We then introduce a depth constraint to probe the occluded areas by measuring the depth of occlusion and background, by assuming that occlusions are always in the foreground with closer distance. During the scene modeling process, a pose refinement scheme is further introduced to refine the camera pose with the features of the background scene. As outlined in Fig. 1, our pipeline contains three modules to achieve the above goals: 1) a scene reconstruction module to represent the whole scene using NeRF (with occlusions), 2) a cost volume construction module to gather information from neighboring views as guidance (to indicate where occlusions are), and 3) a selective supervision scheme to constrain another NeRF on the desired background information (occlusions removed), and our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.635, + 0.468, + 0.665 + ], + "angle": 0, + "content": "- an occlusion-free representation without relying on any external prior as supervisory knowledge;" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.667, + 0.468, + 0.695 + ], + "angle": 0, + "content": "- a joint optimization of pose refinement and scene reconstruction by effective multi-view feature fusion;" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.698, + 0.468, + 0.727 + ], + "angle": 0, + "content": "- a selective supervision scheme to probe the occluded areas guided by the scene depth information." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.635, + 0.468, + 0.727 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.468, + 0.81 + ], + "angle": 0, + "content": "Based on the experiments with a dataset containing diverse types of occlusions, the proposed method can eliminate occlusions including scribbles and water droplets on a piece of glass, fences, and even irregular-shaped statues without relying on any external supervisions." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.822, + 0.218, + 0.837 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.848, + 0.282, + 0.862 + ], + "angle": 0, + "content": "2.1. NeRF and its variants" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.468, + 0.901 + ], + "angle": 0, + "content": "NeRF becomes a popular choice for implicit volumetric scene representation [18]. By mapping the point and current" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.332 + ], + "angle": 0, + "content": "viewing direction to its color and density using a neural network and a differentiable rendering scheme, it provides an effective way to learn from multiple viewing angles. Such ability has been explored to complement the missing depth for grasping transparent objects [9]. Based on NeRF, a series of methods have been proposed to tackle the problems like scene understanding [33] and reasoning [26]. Besides, by considering specific physical priors, NeRF-based frameworks also show promising performance in learning clear representation from challenging scenarios [7,8,16]. Recent NeRF variants can even reconstruct the scene from input with various perturbations. For example, NeRF-W [17] and Ha-NeRF [3] can separate the transient objects from the whole scene. However, they target occlusions that are inconsistent throughout the image set and become ineffective when such occlusions remain static and consistent." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.337, + 0.892, + 0.457 + ], + "angle": 0, + "content": "Since NeRF relies on pre-computed camera parameters for scene representation, it fails when the pre-computation is not feasible. Some methods (e.g., NeRF-- [28] and BARF [13]) have been proposed to alleviate this issue by optimizing the camera parameters along with the scene representation. However, how to avoid the interference from undesired scene during the camera parameter optimization remains a problem to be solved." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.478, + 0.68, + 0.492 + ], + "angle": 0, + "content": "2.2. Occlusion removal" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.505, + 0.892, + 0.611 + ], + "angle": 0, + "content": "Traditional methods in occlusion removal address this problem by propagating neighboring pixels via anisotropic diffusion [2] or solving differential equations [1]. Recently, there are more approaches solving this problem in a learning-based manner [11, 34]. A recent trend is to employ the edge information [21] or segmentation masks [24] as a stronger prior to alleviate the ill-posedness of this problem." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Researchers in this area have been aware of the importance of removing occlusions from multiple viewpoints. Traditional methods try to take the benefits of multiple viewpoints by computing a disparity map [10], dense flow field [29], or visual parallax [19]. Using a light field camera that records the spatial and angular information of light rays in the space is also helpful in removing occlusions [25]. However, these methods also pose restrictions on their inputs, such as using stereo image pairs [10], requiring the background to be visible in at least one of the input views [25, 29], or requesting obviously relative motion between occlusions and backgrounds [19]. It is also a recent trend to apply deep learning in guiding network optimization with fewer constraints on input. For example, some methods seek to use the temporal information [4] or the optical flow [14] from video frames to guide the occlusion-free recovery process. However, these methods are confined to small movements of the camera, partially due to the difficulty in 3D scene representation." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20723" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.089, + 0.87, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.29, + 0.892, + 0.388 + ], + "angle": 0, + "content": "Figure 1. Pipeline of the proposed method. By taking the positional encoding [18] of position and viewing direction, the \"scene MLP\" (colored in blue) models the scene with occlusion faithfully, and the \"background MLP\" (colored in green) models the background only. To this end, a cost volume is constructed using images captured from neighboring reference views, guiding the joint optimization of scene MLP and camera parameters, as we discuss in Sec. 4.1. To supervise the training of background MLP and remove occlusions, we propose to aggregate the information from scene MLP about whether the output of background NeRF should be similar to the observed color, by learning a supervision mask from the weights along the ray using a \"mask MLP\" (colored in amber), as explained in Sec. 4.2. The reference and target views constitute the family of neighboring images whose features are warped to construct the cost volume." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.401, + 0.204, + 0.418 + ], + "angle": 0, + "content": "3. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.426, + 0.469, + 0.531 + ], + "angle": 0, + "content": "Our goal is to generate an occlusion-free scene representation given a collection of \\(N\\) images \\(\\mathcal{I} = \\{I_1, I_2, \\ldots, I_N\\}\\) captured from \\(N\\) different viewpoints, where the \\(i\\)-th viewpoint is defined by the extrinsic matrix \\([\\mathbf{R}_i | \\mathbf{t}_i]\\) and the camera is defined by a shared intrinsic matrix \\(\\mathbf{K}\\). Together they form the camera matrices \\(\\mathcal{P} = \\{\\mathbf{P}_1, \\mathbf{P}_2, \\ldots, \\mathbf{P}_N\\}\\), where \\(\\mathbf{P}_i = \\mathbf{K}[\\mathbf{R}_i | \\mathbf{t}_i]\\), as illustrated in the left part of Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.532, + 0.47, + 0.668 + ], + "angle": 0, + "content": "As we follow the approach of NeRF [18], we briefly introduce it in this section for self-contained purposes. NeRF uses a neural network \\( F_{\\Theta} \\) to construct a radiance field, which maps the 3D coordinate of a point \\( \\mathbf{x} = (x,y,z) \\) and viewing direction \\( \\mathbf{d} = (\\theta ,\\phi) \\) into the point's radiance color \\( \\mathbf{c} = (r,g,b) \\) and density \\( \\sigma \\). Then, the volume rendering scheme is applied to render an image. For each training image and every pixel \\( \\mathbf{p} = (u,v) \\), a ray is emitted from the camera position \\( \\mathbf{o} \\) at direction" + }, + { + "type": "equation", + "bbox": [ + 0.115, + 0.674, + 0.469, + 0.711 + ], + "angle": 0, + "content": "\\[\n\\mathbf {d} _ {i} (u, v) = \\mathbf {R} _ {i} \\left[ \\frac {u - W / 2}{f}, - \\frac {v - H / 2}{f}, - 1 \\right] ^ {\\top}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.716, + 0.469, + 0.775 + ], + "angle": 0, + "content": "where \\(W, H,\\) and \\(f\\) are the width, height, and focal length of the input image, respectively. Marching through the ray \\(\\mathbf{r}(t) = \\mathbf{o} + t\\mathbf{d}\\) and aggregating the colors and densities provide us the final color" + }, + { + "type": "equation", + "bbox": [ + 0.141, + 0.781, + 0.469, + 0.817 + ], + "angle": 0, + "content": "\\[\n\\hat {I} _ {i} (\\mathbf {p}) = \\int_ {t _ {n}} ^ {t _ {f}} T (t) \\sigma (\\mathbf {r} (t)) \\mathbf {c} (\\mathbf {r} (t), \\mathbf {d} _ {i}) d t, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.824, + 0.469, + 0.901 + ], + "angle": 0, + "content": "where \\( T(t) = \\exp(-\\int_{t_n}^t \\sigma(\\mathbf{r}(s)) ds) \\) stands for the accumulated transmittance along the ray, or intuitively, the probability of the emitted radiance of \\( \\mathbf{r}(t) \\) to reach the camera without being blocked by points in between. \\( t_n \\) and \\( t_f \\) are the far end and the near end of rendering respectively." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.403, + 0.892, + 0.432 + ], + "angle": 0, + "content": "To make the integral tractable, points are sampled along the ray, and Eq. (2) can be rewritten into" + }, + { + "type": "equation", + "bbox": [ + 0.551, + 0.441, + 0.892, + 0.482 + ], + "angle": 0, + "content": "\\[\n\\hat {I} _ {i} (\\mathbf {p}) = \\sum_ {k} ^ {N _ {s}} T _ {k} \\left(1 - \\exp \\left(- \\sigma_ {k} \\delta_ {k}\\right)\\right) \\mathbf {c} \\left(\\mathbf {r} _ {k}, \\mathbf {d}\\right), \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.486, + 0.892, + 0.529 + ], + "angle": 0, + "content": "\\[\nT _ {k} = \\exp \\left(- \\sum_ {j = 1} ^ {k - 1} \\sigma_ {j} \\delta_ {j}\\right). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.538, + 0.892, + 0.613 + ], + "angle": 0, + "content": "For the convenience of notation, \\(\\sigma (\\mathbf{r}_k)\\) is abbreviated to \\(\\sigma_{k}\\) and \\(\\delta_{k}\\) is defined by \\(t_{k + 1} - t_k\\); \\(N_{s}\\) is the number of points sampled on the ray, \\(\\mathbf{r}_k = \\mathbf{o} + t_k\\mathbf{d}\\) denotes the position of \\(k\\)-th sampling point, and \\(t_k\\) refers to the depth of the \\(k\\)-th sampling point seen from the target view." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.613, + 0.892, + 0.687 + ], + "angle": 0, + "content": "NeRF optimizes the scene representation neural network by minimizing the photometric loss \\(\\mathcal{L} = \\sum_{i=1}^{N} \\|I_i - \\hat{I}_i\\|_2\\), which is simply the difference between input images \\(\\mathcal{I}\\) and corresponding predicted images \\(\\hat{\\mathcal{I}}\\). The optimization process can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.696, + 0.891, + 0.72 + ], + "angle": 0, + "content": "\\[\n\\Theta^ {*} = \\arg \\min _ {\\Theta} \\mathcal {L} (\\hat {\\mathcal {I}} | \\mathcal {I}, \\mathcal {P}). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.727, + 0.892, + 0.818 + ], + "angle": 0, + "content": "Despite NeRF's strong capability in representing a scene from multiple viewpoints, the occlusions pose challenges to it from two aspects: 1) How to selectively aggregate background information from multiple viewpoints and 2) how to accurately estimate the camera pose of desired background in a continuous space." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.83, + 0.688, + 0.848 + ], + "angle": 0, + "content": "4. Proposed approach" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We focus on how to address the two challenges above when using NeRF to see through occlusion from multiple viewpoints and our pipeline is illustrated in Fig. 1. In" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20724" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.258 + ], + "angle": 0, + "content": "Sec. 4.1, we first describe our joint optimization framework to accurately estimate the camera pose and reconstruct the scene with occlusions using a scene MLP. Then, in Sec. 4.2, we introduce how to selectively aggregate background information from multiple viewpoints based on the rectified background pose using a mask MLP. Under the guidance of mask MLP, we finally train a background MLP to reconstruct occlusion-free scenes. Note that both the scene MLP and the background MLP map the positional encoding of position and viewing direction into color and density, as described in Sec. 3." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.268, + 0.408, + 0.284 + ], + "angle": 0, + "content": "4.1. Scene reconstruction with cost volume" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.292, + 0.47, + 0.472 + ], + "angle": 0, + "content": "NeRF and most of its variants use COLMAP [23] as a silver bullet to estimate camera poses for the scene representation, while COLMAP [23] is not an ideal setting for building an occlusion-free representation. The extracted local features cannot be distinguished between undesired occlusion and the desired background, so the matching process can lead to an erroneous correspondence, or even directly fail to estimate the camera parameters. Some recent works try to address the dependency on COLMAP [23] by jointly optimizing camera parameters and the neural radiance fields with the photometric loss [28], achieving comparable performance to COLMAP [23]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.474, + 0.47, + 0.609 + ], + "angle": 0, + "content": "However, only using photometric loss does not solve the fundamental problem, because during training, it only uses one view at a time, namely the target view itself. This makes the system ignorant of the 3D scene as a whole, and thus still cannot tell occlusions from the desired background. Besides, in the early stage of training when NeRF can hardly output meaningful information, the photometric loss cannot constrain camera extrinsics to converge to the correct rotation and translation. This can lead to a sub-optimal solution." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.609, + 0.47, + 0.715 + ], + "angle": 0, + "content": "To make our method aware of occlusions and background, we design a cost volume loss inspired by multiview stereo solutions [6, 30]. It encourages the system to model the location corresponding to each pixel at the correct depth, by utilizing the feature consistency of the same point seen from neighboring reference views. Next, we will introduce how our cost volume loss is constructed." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Feature extraction. A 2D map \\( F_{i} \\) is extracted for every input image \\( I_{i} \\) by using a pre-trained VGG-19 network \\( \\eta \\) as a feature encoder as \\( F_{i} = \\eta (I_{i}) \\). For the convenience of notation, we define \\( F_{i}(\\mathbf{p}) \\) as the feature vector at 2D location \\( \\mathbf{p} \\). When \\( \\mathbf{p} \\) falls off the grid, bilinear interpolation is performed to acquire the feature vector. The extracted 2D features represent local appearance, so intuitively the feature of the same point of interest across neighboring views are assumed to be similar. On the other hand, if two feature vectors vary greatly, it is unlikely that they describe the same area." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "Feature volume construction. To construct a feature volume at the \\(j\\)-th viewpoint (target view), the image features from neighboring reference views should be warped with respect to the depth. The warping from the \\(i\\)-th view to the \\(j\\)-th view can be formulated as a \\(3 \\times 3\\) matrix" + }, + { + "type": "equation", + "bbox": [ + 0.501, + 0.175, + 0.892, + 0.22 + ], + "angle": 0, + "content": "\\[\n\\mathbf {H} _ {i j} (z) = \\mathbf {K R} _ {i} (\\mathbf {I d} - \\frac {\\left(\\mathbf {R} _ {j} ^ {\\top} \\mathbf {t} _ {j} - \\mathbf {R} _ {i} ^ {\\top} \\mathbf {t} _ {i}\\right) \\mathbf {n} _ {j} ^ {\\top}}{z} \\cdot \\mathbf {R} _ {j}) \\mathbf {R} _ {j} ^ {\\top} \\mathbf {K} ^ {- 1}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.221, + 0.892, + 0.312 + ], + "angle": 0, + "content": "where \\(\\mathbf{Id}\\) is the identity matrix, and \\(\\mathbf{n}_j\\) is the principle axis of the \\(j\\)-th view. This matrix describes a mapping that the point at the 2D coordinate of \\(\\mathbf{p}\\) and the depth of \\(z\\) in warped feature map corresponds to the point at \\(\\mathbf{H}_{ij}(z)\\mathbf{p}\\) in the original feature map. Applying the warping matrix to the feature map \\(F_{i}\\) derives a family of warped image feature maps:" + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.321, + 0.892, + 0.339 + ], + "angle": 0, + "content": "\\[\nF _ {i, j, z} (\\mathbf {p}) = F _ {i} \\left(\\mathbf {H} _ {i j} (z) \\mathbf {p}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.348, + 0.892, + 0.377 + ], + "angle": 0, + "content": "which is warped from \\(F_{i}\\), seen from the \\(j\\)-th view, and at the depth of \\(z\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.378, + 0.892, + 0.483 + ], + "angle": 0, + "content": "We then construct a cost volume at the points which we sample using the warped feature maps. Based on the intuition that similar features mean more likelihood of the same area, a large variance of warped features indicates that features belonging to different areas are warped to the current point. Thus, the variance metric can be used as a clue for scene reconstruction as" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.492, + 0.892, + 0.512 + ], + "angle": 0, + "content": "\\[\n\\mathbf {V} _ {j} (u, v, z) = \\operatorname {V a r} _ {i} \\left(F _ {i, j, z} ([ u, v, 1 ] ^ {\\top})\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.52, + 0.892, + 0.551 + ], + "angle": 0, + "content": "where \\(\\operatorname{Var}_i(\\cdot)\\) denotes the variance across \\(M\\) selected neighboring views." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.568, + 0.892, + 0.614 + ], + "angle": 0, + "content": "Joint optimization. To apply the cost volume as supervision for NeRF training, we consider the weight of the \\(k\\)-th sampled point" + }, + { + "type": "equation", + "bbox": [ + 0.605, + 0.624, + 0.892, + 0.641 + ], + "angle": 0, + "content": "\\[\nW _ {k} = T _ {k} \\left(1 - \\exp (- \\sigma_ {k})\\right), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.65, + 0.892, + 0.77 + ], + "angle": 0, + "content": "which refers to the contribution of the point at \\(\\mathbf{r}_t\\) to the result \\(\\hat{I}_j(u,v)\\). We build the scene MLP to reconstruct the scene with occlusion, to provide an intermediate scene representation for further occlusion removal. For a high-quality reconstruction, the scene MLP should be discouraged to assign the points that are inconsistent across neighboring views with a large weight. Based on the observations above, we construct the cost volume loss as" + }, + { + "type": "equation", + "bbox": [ + 0.538, + 0.779, + 0.892, + 0.823 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {c o s t}} = \\sum_ {\\mathbf {p} = (u, v)} \\sum_ {i = 1} ^ {N _ {s}} \\mathbf {V} _ {j} (u, v, t _ {i}) W (u, v, t _ {i}). \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.832, + 0.892, + 0.862 + ], + "angle": 0, + "content": "Meanwhile, we also apply the self-consistency loss aiming at scene reconstruction as" + }, + { + "type": "equation", + "bbox": [ + 0.599, + 0.871, + 0.892, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {s c}} = \\sum_ {\\mathbf {p}} \\| \\hat {I} _ {i} (\\mathbf {p}) - I _ {i} (\\mathbf {p}) \\| _ {2} ^ {2}. \\tag {11}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "20725" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.168 + ], + "angle": 0, + "content": "So far, we have explained the design of scene reconstruction scheme where occlusions exist, including the key component of cost volume, as depicted in the blue box of Fig. 1. Together, they provide a more robust estimation of the scene and the camera poses when occlusions are present." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.186, + 0.334, + 0.202 + ], + "angle": 0, + "content": "4.2. Selective supervision scheme" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.212, + 0.471, + 0.424 + ], + "angle": 0, + "content": "Existing image-based occlusion removal methods [14, 27] can function properly when a large amount of labeled data is available, by considering the occlusion removal as a mapping from samples with occlusions to their occlusion-free counterparts. Our method is grounded on a totally different underlying logic. By simulating human eyes to observe a specified scenes, our method can aggregate a continuous occlusion-free scene representation from different viewpoints. Then, we can render any occlusion-free viewpoints from this learned scene representation. Such mechanism enables a novel way to remove occlusion in an unsupervised manner. The key problem then becomes how to focus on the desired background only during scene aggregation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.427, + 0.47, + 0.669 + ], + "angle": 0, + "content": "We achieve this goal based on an intuitive observation: a majority of undesired occlusions always appear in the foreground. However, directly applying depth as supervision is not feasible, since the depth does not necessarily correlate with the presence of occlusion. To this end, we propose to use bidirectional depth inconsistency as a clue. If the expected ray termination depth seen from the camera and that from the other end of the ray show great difference, the ray should pass through some foreground occlusion. Note that \"termination\" in rendering refers to the point that no point behind it contributes to the rendered ray. This assumption follows the physical law, and does not rely on any external prior or excessive training data. Note that this prior is opposite to the shell-shaped geometry assumption by NeRFReN [7], which assumes that the density along the ray peaks out at the surface." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.671, + 0.469, + 0.716 + ], + "angle": 0, + "content": "In implementing this, we design a mask MLP \\(P_{\\Psi}\\) to predict the occluding likelihood \\(\\tilde{P} (\\mathbf{r})\\) based on the weights along the ray, formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.186, + 0.734, + 0.469, + 0.76 + ], + "angle": 0, + "content": "\\[\n\\tilde {P} (\\mathbf {r}) = P _ {\\Psi} \\left(\\left\\{W _ {k} \\right\\} _ {k = 1} ^ {N _ {s}}\\right). \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.779, + 0.47, + 0.845 + ], + "angle": 0, + "content": "Based on the intuition discussed above, the mask MLP is trained based on the bidirectional depth inconsistency. The expected termination depth \\(\\overrightarrow{D}\\) and reversed termination depth \\(\\overline{D}\\) is defined by" + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.862, + 0.47, + 0.905 + ], + "angle": 0, + "content": "\\[\n\\overrightarrow {D} (\\mathbf {r}) = \\sum_ {k = 1} ^ {N _ {s}} W _ {k} t _ {k} = \\sum_ {k = 1} ^ {N _ {s}} T _ {k} \\left(1 - \\exp \\left(- \\sigma_ {k} \\delta_ {k}\\right)\\right) t _ {k}, \\tag {13}\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.499, + 0.088, + 0.626, + 0.148 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.548, + 0.15, + 0.582, + 0.163 + ], + "angle": 0, + "content": "(a) \\(\\overrightarrow{D}\\)" + }, + { + "type": "image", + "bbox": [ + 0.63, + 0.089, + 0.759, + 0.147 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.671, + 0.149, + 0.715, + 0.161 + ], + "angle": 0, + "content": "(b) \\(\\Delta_D\\)" + }, + { + "type": "image", + "bbox": [ + 0.76, + 0.088, + 0.892, + 0.148 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.807, + 0.149, + 0.838, + 0.159 + ], + "angle": 0, + "content": "(c) \\(P\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.177, + 0.894, + 0.302 + ], + "angle": 0, + "content": "Figure 2. An illustration of the depth and mask used for selective supervision (without losing generality, we show the example for fence, and this applies to other types of occlusions as well): (a) estimated depth \\((\\vec{D})\\) of the scene with occlusion, (b) the bidirectional depth difference \\((\\Delta_D)\\), and (c) the supervision mask \\((P)\\). Note that the white regions in (c) denote the desired background details that need supervision and undesired foreground occlusions are marked in black. Please visit our project page for animated results." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.328, + 0.531, + 0.342 + ], + "angle": 0, + "content": "and" + }, + { + "type": "equation", + "bbox": [ + 0.512, + 0.349, + 0.892, + 0.39 + ], + "angle": 0, + "content": "\\[\n\\overleftarrow {D} (\\mathbf {r}) = \\sum_ {k = 1} ^ {N _ {s}} \\overleftarrow {W} _ {k} t _ {k} = \\sum_ {k = 1} ^ {N _ {s}} \\overleftarrow {T} _ {k} \\left(1 - \\exp \\left(- \\sigma_ {k} \\delta_ {k}\\right)\\right) t _ {k}, \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.401, + 0.892, + 0.489 + ], + "angle": 0, + "content": "respectively, where \\(\\overline{T}_k = \\exp (-\\sum_{j > k}\\sigma_j\\delta_j)\\) is the reversed accumulated transmittance. An example of \\(\\vec{D}\\) can be found in Fig. 2(a). We train the mask MLP based on the prior described before, that \\(P(\\mathbf{r})\\) correlates with \\(\\Delta_D(\\mathbf{r}) = \\overleftarrow{D} -\\overrightarrow{D}\\), by" + }, + { + "type": "equation", + "bbox": [ + 0.513, + 0.496, + 0.891, + 0.527 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {m a s k}} = - \\cos \\left(\\left\\{P (\\mathbf {r}) - \\alpha \\right\\} _ {\\mathbf {r} \\in \\mathcal {R}}, \\left\\{\\Delta_ {D} (\\mathbf {r}) - \\overline {{\\Delta_ {D}}} \\right\\} _ {\\mathbf {r} \\in \\mathcal {R}}\\right), \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.528, + 0.892, + 0.617 + ], + "angle": 0, + "content": "where \\(\\mathcal{R}\\) is the patch of rays for training, \\(\\overline{\\Delta_D}\\) is the mean value of \\(\\Delta_D\\) across the entire input view, and \\(\\cos (\\mathbf{a},\\mathbf{b})\\) is the cosine similarity between a and b. We show an example of \\(\\Delta_D\\) in Fig. 2(b). \\(\\alpha\\) is a pre-defined threshold for the binary classification of \\(\\tilde{P}\\), deriving a binary supervision mask \\(P\\) by" + }, + { + "type": "equation", + "bbox": [ + 0.6, + 0.617, + 0.891, + 0.651 + ], + "angle": 0, + "content": "\\[\nP (\\mathbf {r}) = \\left\\{ \\begin{array}{l l} 0, & \\tilde {P} (\\mathbf {r}) < \\alpha \\\\ 1, & \\tilde {P} (\\mathbf {r}) \\geq \\alpha \\end{array} \\right.. \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.655, + 0.891, + 0.699 + ], + "angle": 0, + "content": "As shown in Fig. 2(c), the obtained binary mask \\( P \\) can clearly distinguish the desired background to be kept and the undesired occlusion to be removed." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.7, + 0.892, + 0.776 + ], + "angle": 0, + "content": "We further train a background MLP which reconstructs the scene belonging to the background only. Specifically, we only supervise the rendered color of background MLP to be similar to the input image in the background region where \\( P(\\mathbf{r}) = 1 \\)." + }, + { + "type": "equation", + "bbox": [ + 0.561, + 0.783, + 0.892, + 0.858 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {b g}} = \\sum_ {\\mathbf {p}} \\left\\| \\left(\\left(\\hat {I} _ {i} ^ {\\prime} (\\mathbf {p}) - I _ {i} (\\mathbf {p})\\right) \\odot P (\\mathbf {r}) \\right\\| _ {2} ^ {2} \\right. \\tag {17} \\\\ + \\sum_ {\\mathbf {p}} \\left\\| (\\vec {D} ^ {\\prime} (\\mathbf {r}) - \\vec {D} (\\mathbf {r})) \\odot P (\\mathbf {r}) \\right\\| _ {1}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.868, + 0.891, + 0.902 + ], + "angle": 0, + "content": "where \\(\\hat{I}_i^{\\prime}(\\mathbf{p})\\) and \\(\\vec{D}^{\\prime}(\\mathbf{r})\\) denote the rendered color and expected ray termination depth of the background MLP." + }, + { + "type": "footer", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20726" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.152 + ], + "angle": 0, + "content": "With background MLP holding the occlusion-free scene representation and the scene MLP containing all the information including occlusions, we have finally gathered all the pieces of the puzzle." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.16, + 0.292, + 0.176 + ], + "angle": 0, + "content": "4.3. Implementation details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.184, + 0.47, + 0.394 + ], + "angle": 0, + "content": "We implement our method using PyTorch. The scene MLP and the background MLP, each having 256 channels, consist of 6 layers and 8 layers respectively, which predict the colors and densities corresponding to the whole scene and the background separately. The mask MLP is a 2-layer MLP with 64 channels. When constructing cost volume, \\( M = 4 \\) neighboring views are involved to calculate variance in Eq. (8). In the training phase, the scene MLP is jointly optimized with camera parameters of each scene, after being trained coarsely to fit the scene from COLMAP [23] camera pose or identity matrix initialization. All of the trainable parameters are optimized per scene. We train scene MLP for the first \\( 20\\% \\) of iterations, and then train the system end-to-end." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.395, + 0.471, + 0.53 + ], + "angle": 0, + "content": "On each iteration, we evenly sample 128 points along each ray. A batch contains random 1024 pixels in one of the images. We use the Adam optimizer with defaults values \\(\\beta_{1} = 0.999\\), \\(\\beta_{2} = 0.9\\), \\(\\epsilon = 10^{-8}\\), and a learning rate \\(10^{-3}\\) that decays following the cosine scheduler [15] during the optimization. If initialized with identity matrix, the model is trained for about 150K iterations. If a coarse COLMAP [23] estimation is available, our model converges after about 50K iterations." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.544, + 0.21, + 0.56 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.569, + 0.47, + 0.75 + ], + "angle": 0, + "content": "Dataset. To validate the effectiveness of our method, we collect an evaluation dataset containing 10 different scenes, covering various types of occlusions. The dataset not only contains multiple sparse viewpoints of occlusion scenes collected by ourselves, but also samples selected from existing datasets [29,31], to increase the diversity from different capturing conditions. Specifically, FENCE1, SCRIBBLE2 and RAINDROP are adopted from [29], and WIRE2 is adopted from [31]. The rest 6 of the scenes are captured by ourselves using a Sony \\(\\alpha 7\\) III camera or an iPhone 12. For easy reference, we name each set of data using the occlusion type, as shown in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.902 + ], + "angle": 0, + "content": "For our newly captured samples, we take 20 to 60 images with different poses for each scene, where \\(85\\%\\) of the images are used for training our NeRF-based method, and the rest \\(15\\%\\) for testing. This is a commonly adopted setting by many NeRF-based methods [7, 8, 18]. To facilitate the quantitative evaluation, one of newly captured scene is with ground truth, namely SCRIBBLE1. For this scene, we place a piece of glass with scribbles between the desired background scene and the camera. Then, a multi-view sparse image set is acquired by moving cameras in front of the" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.088, + 0.586, + 0.124 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.126, + 0.575, + 0.137 + ], + "angle": 0, + "content": "(a) FENCE1" + }, + { + "type": "image", + "bbox": [ + 0.589, + 0.089, + 0.666, + 0.124 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.593, + 0.126, + 0.656, + 0.137 + ], + "angle": 0, + "content": "(b) FENCE2" + }, + { + "type": "image", + "bbox": [ + 0.668, + 0.089, + 0.745, + 0.124 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.672, + 0.126, + 0.735, + 0.137 + ], + "angle": 0, + "content": "(c) FENCE3" + }, + { + "type": "image", + "bbox": [ + 0.748, + 0.089, + 0.825, + 0.124 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.743, + 0.126, + 0.904, + 0.137 + ], + "angle": 0, + "content": "(d) SCRIBBLE1(e) SCRIBBLE2" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.138, + 0.586, + 0.172 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.504, + 0.174, + 0.664, + 0.186 + ], + "angle": 0, + "content": "(f) SCRIBBLE3 (g) RAINDROP" + }, + { + "type": "image", + "bbox": [ + 0.589, + 0.139, + 0.666, + 0.172 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.668, + 0.139, + 0.745, + 0.172 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.672, + 0.174, + 0.736, + 0.185 + ], + "angle": 0, + "content": "(h) STATUE" + }, + { + "type": "image", + "bbox": [ + 0.748, + 0.138, + 0.825, + 0.172 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.757, + 0.174, + 0.81, + 0.185 + ], + "angle": 0, + "content": "(i) WIRE1" + }, + { + "type": "image", + "bbox": [ + 0.827, + 0.139, + 0.904, + 0.172 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.836, + 0.174, + 0.891, + 0.186 + ], + "angle": 0, + "content": "(j) WIRE2" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.194, + 0.893, + 0.236 + ], + "angle": 0, + "content": "Figure 3. From (a) to (j), we provide one sample image for each scene in our evaluation dataset. The names of the scenes are listed below the image indicating occlusion type." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.261, + 0.893, + 0.307 + ], + "angle": 0, + "content": "glass. To obtain the ground truth for validation, we remove the glass while taking the first and last view of the scene, so that these two views can come with a ground truth." + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.318, + 0.64, + 0.376 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.641, + 0.319, + 0.718, + 0.376 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.719, + 0.319, + 0.898, + 0.376 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.643, + 0.379, + 0.759, + 0.391 + ], + "angle": 0, + "content": "(a) Vanilla NeRF [18]" + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.392, + 0.64, + 0.449 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.641, + 0.392, + 0.718, + 0.449 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.719, + 0.392, + 0.898, + 0.449 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.624, + 0.452, + 0.778, + 0.463 + ], + "angle": 0, + "content": "(b) PWC-Net+NeRF [14, 18]" + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.465, + 0.64, + 0.521 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.641, + 0.465, + 0.718, + 0.521 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.719, + 0.465, + 0.898, + 0.521 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.657, + 0.524, + 0.746, + 0.535 + ], + "angle": 0, + "content": "(c) Ha-NeRF [3]" + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.537, + 0.64, + 0.594 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.641, + 0.537, + 0.718, + 0.594 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.719, + 0.537, + 0.898, + 0.594 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.655, + 0.596, + 0.748, + 0.608 + ], + "angle": 0, + "content": "(d) NeRF-W [17]" + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.609, + 0.64, + 0.666 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.641, + 0.61, + 0.718, + 0.666 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.719, + 0.61, + 0.898, + 0.666 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.66, + 0.669, + 0.743, + 0.68 + ], + "angle": 0, + "content": "(e) Our method" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.686, + 0.893, + 0.797 + ], + "angle": 0, + "content": "Figure 4. Visual quality comparisons for occlusion removal. From left to right, we show animated novel view synthesis on FENCE1, STATUE, and WIRE1. From (a) to (e), we show results obtained by (a) vanilla NeRF [18], (b) PWC-Net + NeRF [14, 18], (c) Ha-NeRF [3], (d) NeRF-W [17], and (e) our method. Please visit our project page for animated results. Some results show stronger variation over the animation due to failure in extracting consistent background." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Baselines. To our knowledge, there is no similar work that construct an occlusion-free scene based on NeRF representation. We design three baselines for comparison by considering both the contributions from occlusion removal and NeRF to verify our performance advantage. 1) A state-of" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20727" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.09, + 0.471, + 0.117 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison results. \\(\\uparrow (\\downarrow)\\) indicates larger (smaller) values are better, and bold font indicates the best results." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.123, + 0.472, + 0.205 + ], + "angle": 0, + "content": "
MethodPSNR↑SSIM↑LPIPS↓
PWC-Net+NeRF [14,18]18.250.790.45
NeRF-W [17]11.060.450.61
Ha-NeRF [3]15.410.700.41
Proposed19.390.830.37
" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.217, + 0.189, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.133, + 0.279, + 0.148, + 0.288 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.19, + 0.217, + 0.338, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.262, + 0.278, + 0.277, + 0.288 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.338, + 0.217, + 0.468, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.391, + 0.278, + 0.406, + 0.288 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.29, + 0.189, + 0.347 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.132, + 0.351, + 0.149, + 0.361 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image", + "bbox": [ + 0.19, + 0.29, + 0.338, + 0.347 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.262, + 0.351, + 0.278, + 0.361 + ], + "angle": 0, + "content": "(e)" + }, + { + "type": "image", + "bbox": [ + 0.338, + 0.29, + 0.468, + 0.347 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.391, + 0.351, + 0.406, + 0.361 + ], + "angle": 0, + "content": "(f)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.369, + 0.47, + 0.452 + ], + "angle": 0, + "content": "Figure 5. Visual quality comparisons for occlusion removal on SCRIBBLE1 scene. (a) is one image in the validation set with occlusion and (f) is ground truth of this view) with results from: (b) results obtained by Ha-NeRF [3], (c) results obtained by NeRF-W [17], (d) results obtained by occlusion removal [14] + NeRF [18], and (e) our method." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.473, + 0.469, + 0.669 + ], + "angle": 0, + "content": "the-art occlusion removal method for image sequences or videos (referred to as PWC-Net [14] based on the main pillar of the work) + NeRF [18]: In this setting, PWC-Net [14] serves as a pre-processing tool to eliminate occlusion before training the NeRF [18]. 2) NeRF-W [17]: This is a method designed for separating a complete scene into two components. 3) Ha-NeRF [3]: This is another method that decomposes the scenes into distinct parts. We compare those settings with our method on our evaluation dataset, as listed in Fig. 3. For fairness of comparison, we evenly sample 128 points along the ray without applying importance sampling. We also keep the trainable parameters in each baseline approximately the same with the proposed method." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.687, + 0.469, + 0.763 + ], + "angle": 0, + "content": "Evaluation methodology. We conduct both qualitative and quantitative evaluations on our dataset. For quantitative evaluation, we compare reconstruction quality of the estimated occlusion-free images for viewpoints with corresponding ground truth by PSNR, SSIM, and LPIPS [32]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.772, + 0.283, + 0.788 + ], + "angle": 0, + "content": "5.1. Qualitative evaluation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.469, + 0.901 + ], + "angle": 0, + "content": "The qualitative results of the rendered novel occlusion-free views are presented in Figs. 4 and 5. In Fig. 4, we use the scenes of FENCE1, STATUE and WIRE1 in our dataset to show that our method can reliably reconstruct novel background views, and it works no matter whether the occlusions cover a large area or have an irregular shape. In Fig. 5, we take one frame in SCRIBBLE1 with ground truth to show" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.09, + 0.632, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.559, + 0.264, + 0.575, + 0.275 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.632, + 0.09, + 0.759, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.684, + 0.264, + 0.7, + 0.275 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.76, + 0.09, + 0.885, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.811, + 0.264, + 0.827, + 0.275 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.281, + 0.892, + 0.351 + ], + "angle": 0, + "content": "Figure 6. Visual quality comparisons for occlusion removal on RAINDROP scene, with results from (a) an image-based raindrop removal method [22], (b) an image-based occlusion removal method [14], and (c) our method. Our method outperforms the baselines in the highlighted red boxes." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.367, + 0.669, + 0.381 + ], + "angle": 0, + "content": "the fidelity of our results." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.382, + 0.892, + 0.593 + ], + "angle": 0, + "content": "PWC-Net + NeRF [14, 18] only achieves comparable results in FENCE1 where the occlusions are regularly shaped, but shows poor artifacts on unseen types of occlusions, such as the irregularly-shaped STATUE and WIRE1, as shown in Fig. 4(b). This shows their reliance on training data. Besides, the occlusion removal method can lead to an oversmoothing issue which undermines the reconstruction accuracy. Meanwhile, the results from Ha-NeRF [3] and NeRF-W [17] in Figs. 4(c) and 4(d) either shows degenerated details or still preserves occlusions, which verifies our claim that they cannot handle the consistent occlusions observed in a continuous image sequence. The results from our method in Fig. 4(e) clearly remove these occlusions with different shapes and distributions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.595, + 0.892, + 0.76 + ], + "angle": 0, + "content": "Our method relies on an accurate separation between foreground and background, but the assumption may be violated if obstructions are semi-transparent. We conduct experiments on the RAINDROP scene in our validation dataset to evaluate the robustness of our method under such scenario. With the results displayed in Fig. 6, our method can also effectively remove the raindrop in this example. It achieves competitive or even better results in some regions when compared with a method specifically designed for raindrop removal [22] as well as PWC-Net [14] designed to remove occlusions including raindrop." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.772, + 0.716, + 0.788 + ], + "angle": 0, + "content": "5.2. Quantitative evaluation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The quantitative results of the occlusion-free novel view in Tab. 1 validate the observation in Figs. 4 and 5. Higher PSNR values show that our method can render occlusion-free novel views and recover the color information with higher accuracy. Higher SSIM values indicate that our method can preserve the structural information with high-frequency details. Lower LPIPS values show that recovered" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20728" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.892, + 0.119 + ], + "angle": 0, + "content": "Table 2. Ablation study results aggregated by the types of occlusions. We compare the rendered results of the scene MLP with NeRF [18] using COLMAP [23] as camera parameter estimation and NeRF-- [28] initialized with COLMAP camera parameters." + }, + { + "type": "table", + "bbox": [ + 0.171, + 0.129, + 0.803, + 0.243 + ], + "angle": 0, + "content": "
MetricsPSNR ↑ / SSIM ↑ / LPIPS↓
MethodCOLMAP [23]NeRF-- [28]Scene MLP
DatasetFENCE{1,2,3}24.01 / 0.73 / 0.3818.67 / 0.54 / 0.5125.46 / 0.72 / 0.35
SCRIBBLE{1,2,3}22.70 / 0.78 / 0.3117.98 / 0.52 / 0.4623.85 / 0.84 / 0.34
WIRES{1,2}26.21 / 0.93 / 0.2022.28 / 0.88 / 0.2534.44 / 0.95 / 0.23
RAINDROP27.69 / 0.87 / 0.2627.78 / 0.83 / 0.2528.69 / 0.86 / 0.25
STATUE27.53 / 0.84 / 0.3318.49 / 0.56 / 0.4927.08 / 0.80 / 0.37
" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.268, + 0.468, + 0.283 + ], + "angle": 0, + "content": "images by our method better aligns with human perception." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.294, + 0.231, + 0.31 + ], + "angle": 0, + "content": "5.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.318, + 0.468, + 0.408 + ], + "angle": 0, + "content": "Our method consists of two major components: joint optimization for pose refinement and a selective supervision strategy to eliminate occlusions. Since the selective supervision scheme largely bases on the intermediate information of scene reconstruction MLP, a high-quality reconstruction of the scene with occlusion is crucial to our performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.409, + 0.468, + 0.483 + ], + "angle": 0, + "content": "In the ablation study, we first replace pose refinement with COLMAP [23]. From the results shown in Fig. 7(b), the occlusions largely remain, because the fixed erroneous camera parameters makes distinguishing scene components from scene MLP infeasible." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.485, + 0.468, + 0.621 + ], + "angle": 0, + "content": "We also explore the results without the selective supervision scheme, while maintaining pose refinement, namely only output of the scene MLP. In this setting, our method cannot remove occlusions, but it can reconstruct the scene containing occlusions with a high fidelity, as shown in Fig. 7(c). It can also reconstruct the scene with higher fidelity than NeRF based on COLMAP [23], as shown quantitatively in Tab. 2. Our complete model shows the best results for majority of scenes in our dataset." + }, + { + "type": "image", + "bbox": [ + 0.085, + 0.634, + 0.21, + 0.69 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.138, + 0.693, + 0.154, + 0.705 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.212, + 0.634, + 0.338, + 0.69 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.264, + 0.693, + 0.281, + 0.705 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.338, + 0.634, + 0.465, + 0.69 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.39, + 0.693, + 0.407, + 0.705 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.711, + 0.468, + 0.767 + ], + "angle": 0, + "content": "Figure 7. Visual comparisons for the ablation study on the FENCE1 scene with results from: (a) our complete method, (b) the method without joint optimization, and (c) the method without selective supervisions. Please visit our project page for animated results." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.8, + 0.205, + 0.815 + ], + "angle": 0, + "content": "6. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.468, + 0.901 + ], + "angle": 0, + "content": "In this paper, grounded on Neural Radiance Fields, we construct an occlusion-free scene representation, which is capable of rendering occlusion-free images with desired viewpoints. By directly using the viewpoints as a part of the input, our method is able to aggregate occluded details" + }, + { + "type": "image", + "bbox": [ + 0.509, + 0.259, + 0.635, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.561, + 0.327, + 0.578, + 0.338 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.638, + 0.259, + 0.763, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.688, + 0.327, + 0.705, + 0.338 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.765, + 0.259, + 0.89, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.815, + 0.327, + 0.832, + 0.338 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.345, + 0.892, + 0.387 + ], + "angle": 0, + "content": "Figure 8. A failure scene of the proposed method, when reflection is present as occlusion. (a) input sample, (b) results of the scene MLP, (c) failed reflection removal with the background MLP." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.407, + 0.892, + 0.543 + ], + "angle": 0, + "content": "from multiple viewpoints without relying on any external guidance. We propose a pose refinement scheme to ensure robust training when the camera poses cannot be accurately estimated by COLMAP [23]. To focus on modeling of desired background scenes within a specific viewpoint, we further introduce a depth constraint to probe the occluded areas by measuring the depth of occlusion and background. Experimental results show that our method is capable of achieving promising results under a variety of scenarios." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.561, + 0.892, + 0.788 + ], + "angle": 0, + "content": "Limitations. Despite the promising performance of our proposed method, several limitations are still to be addressed in our future study. First, our method clearly assumes the undesired layers are in the foreground. It cannot disambiguate the cases where details from different layers are intertwined, such as scenes with reflection, which is especially challenging due to its semi-transparency and widespread presence. We show a failure case of the proposed method in a scene with reflection as occlusion in Fig. 8. Besides, since several variants of NeRF have been proposed for faster convergence and more accurate reconstruction [5,20], advanced NeRF variants may boost the performance of the proposed method. In addition, we also hope to reduce the number of images required for joint optimization of camera parameters and NeRF in our future work." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.806, + 0.892, + 0.896 + ], + "angle": 0, + "content": "Acknowledgement. This work is supported by National Natural Science Foundation of China under Grant No. 62136001, 62088102. Renjie Wan is supported by the Blue Sky Research Fund of HKBU under Grant No. BSRF/21-22/16 and Guangdong Basic and Applied Basic Research Foundation under Grant No. 2022A1515110692." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20729" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.117, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Coloma Ballester, Marcelo Bertalmio, Vicent Caseles, Guillermo Sapiro, and Joan Verdera. Filling-in by joint interpolation of vector fields and gray levels. IEEE Transactions on Image Processing, 2001. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.174, + 0.471, + 0.215 + ], + "angle": 0, + "content": "[2] Marcelo Bertalmio, Guillermo Sapiro, Vincent Caseles, and Coloma Ballester. Image inpainting. In Proc. of ACM SIGGRAPH, 2000. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.218, + 0.471, + 0.26 + ], + "angle": 0, + "content": "[3] Xingyu Chen, Qi Zhang, Xiaoyu Li, Yue Chen, Ying Feng, Xuan Wang, and Jue Wang. Hallucinated neural radiance fields in the wild. In CVPR, 2022. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.262, + 0.471, + 0.331 + ], + "angle": 0, + "content": "[4] Chen Du, Byeongkeun Kang, Zheng Xu, Ji Dai, and Truong Nguyen. Accurate and efficient video de-fencing using convolutional neural networks and temporal information. In Proc. of International Conference on Multimedia and Expo, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.334, + 0.471, + 0.375 + ], + "angle": 0, + "content": "[5] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In CVPR, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.378, + 0.471, + 0.42 + ], + "angle": 0, + "content": "[6] Xiaodong Gu, Zhiwen Fan, Siyu Zhu, Zuozhuo Dai, Feitong Tan, and Ping Tan. Cascade cost volume for high-resolution multi-view stereo and stereo matching. In CVPR, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.422, + 0.471, + 0.463 + ], + "angle": 0, + "content": "[7] Yuan-Chen Guo, Di Kang, Linchao Bao, Yu He, and Song-Hai Zhang. NeRFReN: Neural radiance fields with reflections. In CVPR, 2022. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.466, + 0.471, + 0.507 + ], + "angle": 0, + "content": "[8] Xin Huang, Qi Zhang, Ying Feng, Hongdong Li, Xuan Wang, and Qing Wang. HDR-NeRF: High dynamic range neural radiance fields. In CVPR, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.51, + 0.471, + 0.565 + ], + "angle": 0, + "content": "[9] Jeffrey Ichnowski, Yahav Avigal, Justin Kerr, and Ken Goldberg. Dex-NeRF: Using a neural radiance field to grasp transparent objects. In Proc. of Conference on Robot Learning, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.568, + 0.471, + 0.623 + ], + "angle": 0, + "content": "[10] Sankaraganesh Jonna, Sukla Satapathy, and Rajiv R Sahay. Stereo image de-fencing using smartphones. In Proc. of International Conference on Acoustics, Speech, and Signal Processing, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.626, + 0.471, + 0.667 + ], + "angle": 0, + "content": "[11] Jingyuan Li, Ning Wang, Lefei Zhang, Bo Du, and Dacheng Tao. Recurrent feature reasoning for image inpainting. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.67, + 0.471, + 0.724 + ], + "angle": 0, + "content": "[12] Xiaoyu Li, Bo Zhang, Jing Liao, and Pedro V Sander. Let's see clearly: Contaminant artifact removal for moving cameras. In Proc. of International Conference on Computer Vision, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.727, + 0.471, + 0.768 + ], + "angle": 0, + "content": "[13] Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. BARF: Bundle-adjusting neural radiance fields. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.772, + 0.471, + 0.812 + ], + "angle": 0, + "content": "[14] Yu-Lun Liu, Wei-Sheng Lai, Ming-Hsuan Yang, Yung-Yu Chuang, and Jia-Bin Huang. Learning to see through obstructions. In CVPR, 2020. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.814, + 0.471, + 0.856 + ], + "angle": 0, + "content": "[15] Ilya Loshchilov and Frank Hutter. SGDR: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[16] Li Ma, Xiaoyu Li, Jing Liao, Qi Zhang, Xuan Wang, Jue Wang, and Pedro V Sander. Deblur-NeRF: Neural radiance fields from blurry images. In CVPR, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.117, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.148 + ], + "angle": 0, + "content": "[17] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. NeRF in the Wild: Neural radiance fields for unconstrained photo collections. In CVPR, 2021. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.151, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[18] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In Proc. of European Conference on Computer Vision, 2020. 1, 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.222, + 0.892, + 0.263 + ], + "angle": 0, + "content": "[19] Yadong Mu, Wei Liu, and Shuicheng Yan. Video de-fencing. IEEE Transactions on Circuits and Systems for Video Technology, 2014. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.266, + 0.892, + 0.307 + ], + "angle": 0, + "content": "[20] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. TOG, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.31, + 0.892, + 0.363 + ], + "angle": 0, + "content": "[21] Kamyar Nazeri, Eric Ng, Tony Joseph, Faisal Z Qureshi, and Mehran Ebrahimi. EdgeConnect: Generative image inpainting with adversarial edge learning. arXiv preprint arXiv:1901.00212, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.367, + 0.892, + 0.409 + ], + "angle": 0, + "content": "[22] Rui Qian, Robby T Tan, Wenhan Yang, Jiajun Su, and Jiaying Liu. Attentive generative adversarial network for raindrop removal from a single image. In CVPR, 2018. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.411, + 0.892, + 0.437 + ], + "angle": 0, + "content": "[23] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In CVPR, 2016. 2, 4, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.44, + 0.894, + 0.495 + ], + "angle": 0, + "content": "[24] Yuhang Song, Chao Yang, Yeji Shen, Peng Wang, Qin Huang, and C-C Jay Kuo. SPG-Net: Segmentation prediction and guidance network for image inpainting. In Proc. of British Machine Vision Conference, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.498, + 0.892, + 0.552 + ], + "angle": 0, + "content": "[25] V. Vaish, M. Levoy, R. Szeliski, C.L. Zitnick, and Sing Bing Kang. Reconstructing occluded surfaces using synthetic apertures: Stereo, focus and robust measures. In CVPR, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.556, + 0.892, + 0.624 + ], + "angle": 0, + "content": "[26] Suhani Vora, Noha Radwan, Klaus Greff, Henning Meyer, Kyle Genova, Mehdi SM Sajjadi, Etienne Pot, Andrea Tagliasacchi, and Daniel Duckworth. NeSF: Neural semantic fields for generalizable semantic segmentation of 3d scenes. arXiv preprint arXiv:2111.13260, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.627, + 0.892, + 0.668 + ], + "angle": 0, + "content": "[27] Renjie Wan, Boxin Shi, Haoliang Li, Ling-Yu Duan, and Alex C Kot. Face image reflection removal. International Journal of Computer Vision, 2021. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.671, + 0.892, + 0.725 + ], + "angle": 0, + "content": "[28] Zirui Wang, Shangzhe Wu, Weidi Xie, Min Chen, and Victor Adrian Prisacariu. NeRF---: Neural radiance fields without known camera parameters. arXiv preprint arXiv:2102.07064, 2021. 2, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.728, + 0.892, + 0.77 + ], + "angle": 0, + "content": "[29] Tianfan Xue, Michael Rubinstein, Ce Liu, and William T Freeman. A computational approach for obstruction-free photography. TOG, 2015. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.772, + 0.892, + 0.827 + ], + "angle": 0, + "content": "[30] Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. MVSNet: Depth inference for unstructured multiview stereo. In Proc. of European Conference on Computer Vision, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.829, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[31] Renjiao Yi, Jue Wang, and Ping Tan. Automatic fence segmentation in videos of dynamic scenes. In CVPR, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[32] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 7" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20730" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.079, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[33] Shuaifeng Zhi, Tristan Laidlow, Stefan Leutenegger, and Andrew J Davison. In-place scene labelling and understanding with implicit scene representation. In Proc. of International Conference on Computer Vision, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.469, + 0.205 + ], + "angle": 0, + "content": "[34] Manyu Zhu, Dongliang He, Xin Li, Chao Li, Fu Li, Xiao Liu, Errui Ding, and Zhaoxiang Zhang. Image inpainting by end-to-end cascaded refinement with mask awareness. IEEE Transactions on Image Processing, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.092, + 0.47, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20731" + } + ] +] \ No newline at end of file diff --git a/2023/Occlusion-Free Scene Recovery via Neural Radiance Fields/223bdb66-a260-46a1-86eb-891a595d848a_origin.pdf b/2023/Occlusion-Free Scene Recovery via Neural Radiance Fields/223bdb66-a260-46a1-86eb-891a595d848a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b2a9142aa4fa469e9856fb082f3b7fd7ed0d8096 --- /dev/null +++ b/2023/Occlusion-Free Scene Recovery via Neural Radiance Fields/223bdb66-a260-46a1-86eb-891a595d848a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e9ed31ba3b37cc0c8335427a50f251d62d5f8f11ea8ff8b29dbcf76257fbc671 +size 4317439 diff --git a/2023/Occlusion-Free Scene Recovery via Neural Radiance Fields/full.md b/2023/Occlusion-Free Scene Recovery via Neural Radiance Fields/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c988a17587103b531b73a50b6bd49cc17565d125 --- /dev/null +++ b/2023/Occlusion-Free Scene Recovery via Neural Radiance Fields/full.md @@ -0,0 +1,417 @@ +# Occlusion-Free Scene Recovery via Neural Radiance Fields + +Chengxuan Zhu $^{1,2}$ Renjie Wan $^{3}$ Yunkai Tang $^{1,2}$ Boxin Shi $^{1,2*}$ + +$^{1}$ National Engineering Research Center of Visual Technology, School of Computer Science, Peking University $^{2}$ National Key Laboratory for Multimedia Information Processing, School of Computer Science, Peking University $^{3}$ Department of Computer Science, Hong Kong Baptist University {peterzhu, shiboxin}@pku.edu.cn, renjiewan@hkbu.edu.hk, tangyunkai@stu.pku.edu.cn + +# Abstract + +Our everyday lives are filled with occlusions that we strive to see through. By aggregating desired background information from different viewpoints, we can easily eliminate such occlusions without any external occlusion-free supervision. Though several occlusion removal methods have been proposed to empower machine vision systems with such ability, their performances are still unsatisfactory due to reliance on external supervision. We propose a novel method for occlusion removal by directly building a mapping between position and viewing angles and the corresponding occlusion-free scene details leveraging Neural Radiance Fields (NeRF). We also develop an effective scheme to jointly optimize camera parameters and scene reconstruction when occlusions are present. An additional depth constraint is applied to supervise the entire optimization without labeled external data for training. The experimental results on existing and newly collected datasets validate the effectiveness of our method. Our project page: https://freebutuselessssoul.github.io/occnerf. + +# 1. Introduction + +Neural Radiance Fields (NeRF) are capable of learning the scene representation implicitly from a set of 2D images, yet not every scene is favored by observers. Many undesirable occlusions in our world obscure details that are essential to our understanding of the world. In general, such obstructions range from water droplets and scribbles on a piece of glass, to fences or any objects occluding the desired scenes (e.g., a statue closer to the camera in a landmark scene). How to apply computational methods to exclude them from the scene representation is of great interest. + +Occlusion removal (e.g., [29]) is the direct solution to achieve this goal. However, explicit occlusion removal may + +oversmooth essential details necessary for clearly observing the desired background scenes. In addition, current methods mainly depend on external occlusion-free supervision (e.g., fence removal [4], raindrop removal [22]) to develop the reliable capability in removing certain types of occlusions. Once encountering a new scenario with unseen occlusion types beyond their training data, these methods might show degraded performances. To handle more diverse types of occlusions, generic constraints from multiple viewpoints are widely adopted [4, 10, 12, 14, 19, 27, 29] via mimicking our human vision systems, who can easily piece together the desired background scenes by looking at them from different viewpoints. But the majority of these methods just consider viewpoints as a prior in relation to spatial correlations. Their backbones still rely on external training data with corresponding ground truth for optimization, which still does not fundamentally alleviate the difficulty of handling diverse occlusions in the real world. + +An occlusion-free world can be progressively aggregated by seeing its occluded part from different viewing directions to reveal occlusions previously unobservable in each single perspective, as illustrated in the left part of Fig. 1 (the fan is occluded in the target view). Since NeRF [18] employs an implicit representation to map viewpoints to pixels, one may come to the naive solution of directly constructing a NeRF which is optimized across multiple viewpoints. However, the vanilla NeRF [18] representing the scene as a whole is not able to treat occlusion and background scenes distinctively, and as long as the occlusion remains static, NeRF is designed to faithfully reconstruct its presence. Meanwhile, many NeRF variants can decompose the whole scene into different components (e.g., NeRF-W [17], Ha-NeRF [3], NeRFRen [7]), but they cannot handle the real-world static occlusions. This is because NeRF-W [17] and Ha-NeRF [3] rely on the inconsistency of undesired components across different views to achieve such separation, which is difficult to be observed in a continuous 3D world. On the other hand, NeRFRen [7] only works in separating the transmission and reflection compo + +nents caused by semi-transparent planar glass, which is incapable of handling opaque occlusions in the real world. + +Another problem comes from NeRF's reliance on camera parameters pre-computed by COLMAP [23], because handcrafted features extracted and matched using COLMAP [23] are for the whole scene, and are incapable in distinguishing between undesired occlusions and the desired background. When the features from occlusion dominate the matching process, the obtained camera parameters cannot faithfully model the spatial correlation of background scenes across multiple viewpoints. Besides, COLMAP [23] is not a stable option for pose estimation in the real world [28]. The existence of occlusions may prevent it from working properly, making the occlusion-free scene representation infeasible. + +In this paper, we aim at seeing through the occluded scenes by developing an occlusion-free scene representation without considering specific occlusion types, based on which we can render any occlusion-free images from desired viewpoints. Our method first maps viewing angles and their corresponding scene details by leveraging NeRF. We then introduce a depth constraint to probe the occluded areas by measuring the depth of occlusion and background, by assuming that occlusions are always in the foreground with closer distance. During the scene modeling process, a pose refinement scheme is further introduced to refine the camera pose with the features of the background scene. As outlined in Fig. 1, our pipeline contains three modules to achieve the above goals: 1) a scene reconstruction module to represent the whole scene using NeRF (with occlusions), 2) a cost volume construction module to gather information from neighboring views as guidance (to indicate where occlusions are), and 3) a selective supervision scheme to constrain another NeRF on the desired background information (occlusions removed), and our contributions can be summarized as follows: + +- an occlusion-free representation without relying on any external prior as supervisory knowledge; +- a joint optimization of pose refinement and scene reconstruction by effective multi-view feature fusion; +- a selective supervision scheme to probe the occluded areas guided by the scene depth information. + +Based on the experiments with a dataset containing diverse types of occlusions, the proposed method can eliminate occlusions including scribbles and water droplets on a piece of glass, fences, and even irregular-shaped statues without relying on any external supervisions. + +# 2. Related Work + +# 2.1. NeRF and its variants + +NeRF becomes a popular choice for implicit volumetric scene representation [18]. By mapping the point and current + +viewing direction to its color and density using a neural network and a differentiable rendering scheme, it provides an effective way to learn from multiple viewing angles. Such ability has been explored to complement the missing depth for grasping transparent objects [9]. Based on NeRF, a series of methods have been proposed to tackle the problems like scene understanding [33] and reasoning [26]. Besides, by considering specific physical priors, NeRF-based frameworks also show promising performance in learning clear representation from challenging scenarios [7,8,16]. Recent NeRF variants can even reconstruct the scene from input with various perturbations. For example, NeRF-W [17] and Ha-NeRF [3] can separate the transient objects from the whole scene. However, they target occlusions that are inconsistent throughout the image set and become ineffective when such occlusions remain static and consistent. + +Since NeRF relies on pre-computed camera parameters for scene representation, it fails when the pre-computation is not feasible. Some methods (e.g., NeRF-- [28] and BARF [13]) have been proposed to alleviate this issue by optimizing the camera parameters along with the scene representation. However, how to avoid the interference from undesired scene during the camera parameter optimization remains a problem to be solved. + +# 2.2. Occlusion removal + +Traditional methods in occlusion removal address this problem by propagating neighboring pixels via anisotropic diffusion [2] or solving differential equations [1]. Recently, there are more approaches solving this problem in a learning-based manner [11, 34]. A recent trend is to employ the edge information [21] or segmentation masks [24] as a stronger prior to alleviate the ill-posedness of this problem. + +Researchers in this area have been aware of the importance of removing occlusions from multiple viewpoints. Traditional methods try to take the benefits of multiple viewpoints by computing a disparity map [10], dense flow field [29], or visual parallax [19]. Using a light field camera that records the spatial and angular information of light rays in the space is also helpful in removing occlusions [25]. However, these methods also pose restrictions on their inputs, such as using stereo image pairs [10], requiring the background to be visible in at least one of the input views [25, 29], or requesting obviously relative motion between occlusions and backgrounds [19]. It is also a recent trend to apply deep learning in guiding network optimization with fewer constraints on input. For example, some methods seek to use the temporal information [4] or the optical flow [14] from video frames to guide the occlusion-free recovery process. However, these methods are confined to small movements of the camera, partially due to the difficulty in 3D scene representation. + +![](images/a0275a5899b29a71caf6f47a0dedd3e6df3c6d8c135b8ed9e8512222ce1b4c0d.jpg) +Figure 1. Pipeline of the proposed method. By taking the positional encoding [18] of position and viewing direction, the "scene MLP" (colored in blue) models the scene with occlusion faithfully, and the "background MLP" (colored in green) models the background only. To this end, a cost volume is constructed using images captured from neighboring reference views, guiding the joint optimization of scene MLP and camera parameters, as we discuss in Sec. 4.1. To supervise the training of background MLP and remove occlusions, we propose to aggregate the information from scene MLP about whether the output of background NeRF should be similar to the observed color, by learning a supervision mask from the weights along the ray using a "mask MLP" (colored in amber), as explained in Sec. 4.2. The reference and target views constitute the family of neighboring images whose features are warped to construct the cost volume. + +# 3. Preliminary + +Our goal is to generate an occlusion-free scene representation given a collection of $N$ images $\mathcal{I} = \{I_1, I_2, \ldots, I_N\}$ captured from $N$ different viewpoints, where the $i$ -th viewpoint is defined by the extrinsic matrix $[\mathbf{R}_i | \mathbf{t}_i]$ and the camera is defined by a shared intrinsic matrix $\mathbf{K}$ . Together they form the camera matrices $\mathcal{P} = \{\mathbf{P}_1, \mathbf{P}_2, \ldots, \mathbf{P}_N\}$ , where $\mathbf{P}_i = \mathbf{K}[\mathbf{R}_i | \mathbf{t}_i]$ , as illustrated in the left part of Fig. 1. + +As we follow the approach of NeRF [18], we briefly introduce it in this section for self-contained purposes. NeRF uses a neural network $F_{\Theta}$ to construct a radiance field, which maps the 3D coordinate of a point $\mathbf{x} = (x,y,z)$ and viewing direction $\mathbf{d} = (\theta ,\phi)$ into the point's radiance color $\mathbf{c} = (r,g,b)$ and density $\sigma$ . Then, the volume rendering scheme is applied to render an image. For each training image and every pixel $\mathbf{p} = (u,v)$ , a ray is emitted from the camera position $\mathbf{o}$ at direction + +$$ +\mathbf {d} _ {i} (u, v) = \mathbf {R} _ {i} \left[ \frac {u - W / 2}{f}, - \frac {v - H / 2}{f}, - 1 \right] ^ {\top}, \tag {1} +$$ + +where $W, H,$ and $f$ are the width, height, and focal length of the input image, respectively. Marching through the ray $\mathbf{r}(t) = \mathbf{o} + t\mathbf{d}$ and aggregating the colors and densities provide us the final color + +$$ +\hat {I} _ {i} (\mathbf {p}) = \int_ {t _ {n}} ^ {t _ {f}} T (t) \sigma (\mathbf {r} (t)) \mathbf {c} (\mathbf {r} (t), \mathbf {d} _ {i}) d t, \tag {2} +$$ + +where $T(t) = \exp(-\int_{t_n}^t \sigma(\mathbf{r}(s)) ds)$ stands for the accumulated transmittance along the ray, or intuitively, the probability of the emitted radiance of $\mathbf{r}(t)$ to reach the camera without being blocked by points in between. $t_n$ and $t_f$ are the far end and the near end of rendering respectively. + +To make the integral tractable, points are sampled along the ray, and Eq. (2) can be rewritten into + +$$ +\hat {I} _ {i} (\mathbf {p}) = \sum_ {k} ^ {N _ {s}} T _ {k} \left(1 - \exp \left(- \sigma_ {k} \delta_ {k}\right)\right) \mathbf {c} \left(\mathbf {r} _ {k}, \mathbf {d}\right), \tag {3} +$$ + +$$ +T _ {k} = \exp \left(- \sum_ {j = 1} ^ {k - 1} \sigma_ {j} \delta_ {j}\right). \tag {4} +$$ + +For the convenience of notation, $\sigma (\mathbf{r}_k)$ is abbreviated to $\sigma_{k}$ and $\delta_{k}$ is defined by $t_{k + 1} - t_k$ ; $N_{s}$ is the number of points sampled on the ray, $\mathbf{r}_k = \mathbf{o} + t_k\mathbf{d}$ denotes the position of $k$ -th sampling point, and $t_k$ refers to the depth of the $k$ -th sampling point seen from the target view. + +NeRF optimizes the scene representation neural network by minimizing the photometric loss $\mathcal{L} = \sum_{i=1}^{N} \|I_i - \hat{I}_i\|_2$ , which is simply the difference between input images $\mathcal{I}$ and corresponding predicted images $\hat{\mathcal{I}}$ . The optimization process can be formulated as: + +$$ +\Theta^ {*} = \arg \min _ {\Theta} \mathcal {L} (\hat {\mathcal {I}} | \mathcal {I}, \mathcal {P}). \tag {5} +$$ + +Despite NeRF's strong capability in representing a scene from multiple viewpoints, the occlusions pose challenges to it from two aspects: 1) How to selectively aggregate background information from multiple viewpoints and 2) how to accurately estimate the camera pose of desired background in a continuous space. + +# 4. Proposed approach + +We focus on how to address the two challenges above when using NeRF to see through occlusion from multiple viewpoints and our pipeline is illustrated in Fig. 1. In + +Sec. 4.1, we first describe our joint optimization framework to accurately estimate the camera pose and reconstruct the scene with occlusions using a scene MLP. Then, in Sec. 4.2, we introduce how to selectively aggregate background information from multiple viewpoints based on the rectified background pose using a mask MLP. Under the guidance of mask MLP, we finally train a background MLP to reconstruct occlusion-free scenes. Note that both the scene MLP and the background MLP map the positional encoding of position and viewing direction into color and density, as described in Sec. 3. + +# 4.1. Scene reconstruction with cost volume + +NeRF and most of its variants use COLMAP [23] as a silver bullet to estimate camera poses for the scene representation, while COLMAP [23] is not an ideal setting for building an occlusion-free representation. The extracted local features cannot be distinguished between undesired occlusion and the desired background, so the matching process can lead to an erroneous correspondence, or even directly fail to estimate the camera parameters. Some recent works try to address the dependency on COLMAP [23] by jointly optimizing camera parameters and the neural radiance fields with the photometric loss [28], achieving comparable performance to COLMAP [23]. + +However, only using photometric loss does not solve the fundamental problem, because during training, it only uses one view at a time, namely the target view itself. This makes the system ignorant of the 3D scene as a whole, and thus still cannot tell occlusions from the desired background. Besides, in the early stage of training when NeRF can hardly output meaningful information, the photometric loss cannot constrain camera extrinsics to converge to the correct rotation and translation. This can lead to a sub-optimal solution. + +To make our method aware of occlusions and background, we design a cost volume loss inspired by multiview stereo solutions [6, 30]. It encourages the system to model the location corresponding to each pixel at the correct depth, by utilizing the feature consistency of the same point seen from neighboring reference views. Next, we will introduce how our cost volume loss is constructed. + +Feature extraction. A 2D map $F_{i}$ is extracted for every input image $I_{i}$ by using a pre-trained VGG-19 network $\eta$ as a feature encoder as $F_{i} = \eta (I_{i})$ . For the convenience of notation, we define $F_{i}(\mathbf{p})$ as the feature vector at 2D location $\mathbf{p}$ . When $\mathbf{p}$ falls off the grid, bilinear interpolation is performed to acquire the feature vector. The extracted 2D features represent local appearance, so intuitively the feature of the same point of interest across neighboring views are assumed to be similar. On the other hand, if two feature vectors vary greatly, it is unlikely that they describe the same area. + +Feature volume construction. To construct a feature volume at the $j$ -th viewpoint (target view), the image features from neighboring reference views should be warped with respect to the depth. The warping from the $i$ -th view to the $j$ -th view can be formulated as a $3 \times 3$ matrix + +$$ +\mathbf {H} _ {i j} (z) = \mathbf {K R} _ {i} (\mathbf {I d} - \frac {\left(\mathbf {R} _ {j} ^ {\top} \mathbf {t} _ {j} - \mathbf {R} _ {i} ^ {\top} \mathbf {t} _ {i}\right) \mathbf {n} _ {j} ^ {\top}}{z} \cdot \mathbf {R} _ {j}) \mathbf {R} _ {j} ^ {\top} \mathbf {K} ^ {- 1}, \tag {6} +$$ + +where $\mathbf{Id}$ is the identity matrix, and $\mathbf{n}_j$ is the principle axis of the $j$ -th view. This matrix describes a mapping that the point at the 2D coordinate of $\mathbf{p}$ and the depth of $z$ in warped feature map corresponds to the point at $\mathbf{H}_{ij}(z)\mathbf{p}$ in the original feature map. Applying the warping matrix to the feature map $F_{i}$ derives a family of warped image feature maps: + +$$ +F _ {i, j, z} (\mathbf {p}) = F _ {i} \left(\mathbf {H} _ {i j} (z) \mathbf {p}\right), \tag {7} +$$ + +which is warped from $F_{i}$ , seen from the $j$ -th view, and at the depth of $z$ . + +We then construct a cost volume at the points which we sample using the warped feature maps. Based on the intuition that similar features mean more likelihood of the same area, a large variance of warped features indicates that features belonging to different areas are warped to the current point. Thus, the variance metric can be used as a clue for scene reconstruction as + +$$ +\mathbf {V} _ {j} (u, v, z) = \operatorname {V a r} _ {i} \left(F _ {i, j, z} ([ u, v, 1 ] ^ {\top})\right), \tag {8} +$$ + +where $\operatorname{Var}_i(\cdot)$ denotes the variance across $M$ selected neighboring views. + +Joint optimization. To apply the cost volume as supervision for NeRF training, we consider the weight of the $k$ -th sampled point + +$$ +W _ {k} = T _ {k} \left(1 - \exp (- \sigma_ {k})\right), \tag {9} +$$ + +which refers to the contribution of the point at $\mathbf{r}_t$ to the result $\hat{I}_j(u,v)$ . We build the scene MLP to reconstruct the scene with occlusion, to provide an intermediate scene representation for further occlusion removal. For a high-quality reconstruction, the scene MLP should be discouraged to assign the points that are inconsistent across neighboring views with a large weight. Based on the observations above, we construct the cost volume loss as + +$$ +\mathcal {L} _ {\text {c o s t}} = \sum_ {\mathbf {p} = (u, v)} \sum_ {i = 1} ^ {N _ {s}} \mathbf {V} _ {j} (u, v, t _ {i}) W (u, v, t _ {i}). \tag {10} +$$ + +Meanwhile, we also apply the self-consistency loss aiming at scene reconstruction as + +$$ +\mathcal {L} _ {\mathrm {s c}} = \sum_ {\mathbf {p}} \| \hat {I} _ {i} (\mathbf {p}) - I _ {i} (\mathbf {p}) \| _ {2} ^ {2}. \tag {11} +$$ + +So far, we have explained the design of scene reconstruction scheme where occlusions exist, including the key component of cost volume, as depicted in the blue box of Fig. 1. Together, they provide a more robust estimation of the scene and the camera poses when occlusions are present. + +# 4.2. Selective supervision scheme + +Existing image-based occlusion removal methods [14, 27] can function properly when a large amount of labeled data is available, by considering the occlusion removal as a mapping from samples with occlusions to their occlusion-free counterparts. Our method is grounded on a totally different underlying logic. By simulating human eyes to observe a specified scenes, our method can aggregate a continuous occlusion-free scene representation from different viewpoints. Then, we can render any occlusion-free viewpoints from this learned scene representation. Such mechanism enables a novel way to remove occlusion in an unsupervised manner. The key problem then becomes how to focus on the desired background only during scene aggregation. + +We achieve this goal based on an intuitive observation: a majority of undesired occlusions always appear in the foreground. However, directly applying depth as supervision is not feasible, since the depth does not necessarily correlate with the presence of occlusion. To this end, we propose to use bidirectional depth inconsistency as a clue. If the expected ray termination depth seen from the camera and that from the other end of the ray show great difference, the ray should pass through some foreground occlusion. Note that "termination" in rendering refers to the point that no point behind it contributes to the rendered ray. This assumption follows the physical law, and does not rely on any external prior or excessive training data. Note that this prior is opposite to the shell-shaped geometry assumption by NeRFReN [7], which assumes that the density along the ray peaks out at the surface. + +In implementing this, we design a mask MLP $P_{\Psi}$ to predict the occluding likelihood $\tilde{P} (\mathbf{r})$ based on the weights along the ray, formulated as: + +$$ +\tilde {P} (\mathbf {r}) = P _ {\Psi} \left(\left\{W _ {k} \right\} _ {k = 1} ^ {N _ {s}}\right). \tag {12} +$$ + +Based on the intuition discussed above, the mask MLP is trained based on the bidirectional depth inconsistency. The expected termination depth $\overrightarrow{D}$ and reversed termination depth $\overline{D}$ is defined by + +$$ +\overrightarrow {D} (\mathbf {r}) = \sum_ {k = 1} ^ {N _ {s}} W _ {k} t _ {k} = \sum_ {k = 1} ^ {N _ {s}} T _ {k} \left(1 - \exp \left(- \sigma_ {k} \delta_ {k}\right)\right) t _ {k}, \tag {13} +$$ + +![](images/eb4e446a46cf45c4e43ad3874b15d415d7df0fcad6e26a8386951670036e3d5f.jpg) +(a) $\overrightarrow{D}$ + +![](images/92b80fe52eafdc54c201cd30093437f4cdd9ad0e04ba217fc012b95e90715ff0.jpg) +(b) $\Delta_D$ + +![](images/12c6a3ddacd7b4ff1b789953d4b44b49af049e21d6624b176308a75dab88893b.jpg) +(c) $P$ +Figure 2. An illustration of the depth and mask used for selective supervision (without losing generality, we show the example for fence, and this applies to other types of occlusions as well): (a) estimated depth $(\vec{D})$ of the scene with occlusion, (b) the bidirectional depth difference $(\Delta_D)$ , and (c) the supervision mask $(P)$ . Note that the white regions in (c) denote the desired background details that need supervision and undesired foreground occlusions are marked in black. Please visit our project page for animated results. + +and + +$$ +\overleftarrow {D} (\mathbf {r}) = \sum_ {k = 1} ^ {N _ {s}} \overleftarrow {W} _ {k} t _ {k} = \sum_ {k = 1} ^ {N _ {s}} \overleftarrow {T} _ {k} \left(1 - \exp \left(- \sigma_ {k} \delta_ {k}\right)\right) t _ {k}, \tag {14} +$$ + +respectively, where $\overline{T}_k = \exp (-\sum_{j > k}\sigma_j\delta_j)$ is the reversed accumulated transmittance. An example of $\vec{D}$ can be found in Fig. 2(a). We train the mask MLP based on the prior described before, that $P(\mathbf{r})$ correlates with $\Delta_D(\mathbf{r}) = \overleftarrow{D} -\overrightarrow{D}$ , by + +$$ +\mathcal {L} _ {\text {m a s k}} = - \cos \left(\left\{P (\mathbf {r}) - \alpha \right\} _ {\mathbf {r} \in \mathcal {R}}, \left\{\Delta_ {D} (\mathbf {r}) - \overline {{\Delta_ {D}}} \right\} _ {\mathbf {r} \in \mathcal {R}}\right), \tag {15} +$$ + +where $\mathcal{R}$ is the patch of rays for training, $\overline{\Delta_D}$ is the mean value of $\Delta_D$ across the entire input view, and $\cos (\mathbf{a},\mathbf{b})$ is the cosine similarity between a and b. We show an example of $\Delta_D$ in Fig. 2(b). $\alpha$ is a pre-defined threshold for the binary classification of $\tilde{P}$ , deriving a binary supervision mask $P$ by + +$$ +P (\mathbf {r}) = \left\{ \begin{array}{l l} 0, & \tilde {P} (\mathbf {r}) < \alpha \\ 1, & \tilde {P} (\mathbf {r}) \geq \alpha \end{array} \right.. \tag {16} +$$ + +As shown in Fig. 2(c), the obtained binary mask $P$ can clearly distinguish the desired background to be kept and the undesired occlusion to be removed. + +We further train a background MLP which reconstructs the scene belonging to the background only. Specifically, we only supervise the rendered color of background MLP to be similar to the input image in the background region where $P(\mathbf{r}) = 1$ . + +$$ +\begin{array}{l} \mathcal {L} _ {\mathrm {b g}} = \sum_ {\mathbf {p}} \left\| \left(\left(\hat {I} _ {i} ^ {\prime} (\mathbf {p}) - I _ {i} (\mathbf {p})\right) \odot P (\mathbf {r}) \right\| _ {2} ^ {2} \right. \tag {17} \\ + \sum_ {\mathbf {p}} \left\| (\vec {D} ^ {\prime} (\mathbf {r}) - \vec {D} (\mathbf {r})) \odot P (\mathbf {r}) \right\| _ {1}, \\ \end{array} +$$ + +where $\hat{I}_i^{\prime}(\mathbf{p})$ and $\vec{D}^{\prime}(\mathbf{r})$ denote the rendered color and expected ray termination depth of the background MLP. + +With background MLP holding the occlusion-free scene representation and the scene MLP containing all the information including occlusions, we have finally gathered all the pieces of the puzzle. + +# 4.3. Implementation details + +We implement our method using PyTorch. The scene MLP and the background MLP, each having 256 channels, consist of 6 layers and 8 layers respectively, which predict the colors and densities corresponding to the whole scene and the background separately. The mask MLP is a 2-layer MLP with 64 channels. When constructing cost volume, $M = 4$ neighboring views are involved to calculate variance in Eq. (8). In the training phase, the scene MLP is jointly optimized with camera parameters of each scene, after being trained coarsely to fit the scene from COLMAP [23] camera pose or identity matrix initialization. All of the trainable parameters are optimized per scene. We train scene MLP for the first $20\%$ of iterations, and then train the system end-to-end. + +On each iteration, we evenly sample 128 points along each ray. A batch contains random 1024 pixels in one of the images. We use the Adam optimizer with defaults values $\beta_{1} = 0.999$ , $\beta_{2} = 0.9$ , $\epsilon = 10^{-8}$ , and a learning rate $10^{-3}$ that decays following the cosine scheduler [15] during the optimization. If initialized with identity matrix, the model is trained for about 150K iterations. If a coarse COLMAP [23] estimation is available, our model converges after about 50K iterations. + +# 5. Experiments + +Dataset. To validate the effectiveness of our method, we collect an evaluation dataset containing 10 different scenes, covering various types of occlusions. The dataset not only contains multiple sparse viewpoints of occlusion scenes collected by ourselves, but also samples selected from existing datasets [29,31], to increase the diversity from different capturing conditions. Specifically, FENCE1, SCRIBBLE2 and RAINDROP are adopted from [29], and WIRE2 is adopted from [31]. The rest 6 of the scenes are captured by ourselves using a Sony $\alpha 7$ III camera or an iPhone 12. For easy reference, we name each set of data using the occlusion type, as shown in Fig. 3. + +For our newly captured samples, we take 20 to 60 images with different poses for each scene, where $85\%$ of the images are used for training our NeRF-based method, and the rest $15\%$ for testing. This is a commonly adopted setting by many NeRF-based methods [7, 8, 18]. To facilitate the quantitative evaluation, one of newly captured scene is with ground truth, namely SCRIBBLE1. For this scene, we place a piece of glass with scribbles between the desired background scene and the camera. Then, a multi-view sparse image set is acquired by moving cameras in front of the + +![](images/0e5ca29bd3105841712d3adbb3e1ee3ba95cff7e81b5d343337c1189007680a0.jpg) +(a) FENCE1 + +![](images/e02365c55347ee7aa9fde4c72d4b709ec0293cce250e439385f8c46bf02bc6b7.jpg) +(b) FENCE2 + +![](images/c64a71041da4182430496bf7ad9638810e88de1c92464a50711c22f010c29a89.jpg) +(c) FENCE3 +(d) SCRIBBLE1(e) SCRIBBLE2 + +![](images/cf4538b29a225a63eb2c3871ea55d442076a44cf815f4e39dec1b081a9af6161.jpg) + +![](images/f1c24814e9e2aed511e37de84e89b5403e6c14817360a740a704bcc4f18efdba.jpg) +(f) SCRIBBLE3 (g) RAINDROP + +![](images/3b77cbb4c76bfa9bea5693ea7c6d14a848223677acd413721ef64a759818c063.jpg) + +![](images/6ad0eb2b3708ac578af25ec332fbd713543174fa5daf73d6c08a5756ceb28d91.jpg) +(h) STATUE +Figure 3. From (a) to (j), we provide one sample image for each scene in our evaluation dataset. The names of the scenes are listed below the image indicating occlusion type. + +![](images/9ae9ce1288252b7c3c4299c12b574b290c70653e9a688a5c635d8a9669d481c0.jpg) +(i) WIRE1 + +![](images/a5bc464626ba89cfe590c2a67dfe6094684991291969b717cce5138f9104a535.jpg) +(j) WIRE2 + +glass. To obtain the ground truth for validation, we remove the glass while taking the first and last view of the scene, so that these two views can come with a ground truth. + +![](images/9977cba7993c2e8977713ddf7ba885844ffa332a310d1e00eade623c2e95e617.jpg) + +![](images/63c631172c06aa5518842d4a94993172fae483e8130829ce6ca5c120248b41b2.jpg) + +![](images/5d0a1c60494e6e082406595c259a21fbee1cafc2a85b3fb4f030974222bd988b.jpg) + +![](images/c8ba6650eee6059e3e2faa5540dafae4517a1dc3f5adca1c70fdb2c90021de4a.jpg) +(b) PWC-Net+NeRF [14, 18] + +![](images/af24594123f5cc33ade3c64b25eed1b0cb1da24d45284641be34d5530108166a.jpg) +(a) Vanilla NeRF [18] + +![](images/92823486a0237d025b5e1787158b617e0ddcb52154196aa3b3038b766ba51e6d.jpg) + +![](images/2e37d99c93a2ef58e37cd5f64baf04b40fba2498bd2046fda159be2fcd3833f8.jpg) + +![](images/88db9acb742ca029104ff629418cecd601dbe789367c1a37f2941f53779f7daf.jpg) + +![](images/92bd0903dcd9d070ce2168b1ca3e7e06853233f030f550a99e043c80cd291da3.jpg) + +![](images/6daed4086598c7e2a322fecd5428c7fc7ed1b117e8599e596ead58853120b9d3.jpg) + +![](images/8d82aa719327b598dca1ea1f571e3e2a01b1e789d7d65c754439f4463c249963.jpg) +(c) Ha-NeRF [3] +(d) NeRF-W [17] + +![](images/20e7c9a3736e00a234af0ff74905fa352cbdfababa451d1857dbe5492ae1ad54.jpg) + +![](images/e54464a78a2ef751d1954c47cfd1c8456072d5b55a3c3871216a34f5d81ac8a9.jpg) +Figure 4. Visual quality comparisons for occlusion removal. From left to right, we show animated novel view synthesis on FENCE1, STATUE, and WIRE1. From (a) to (e), we show results obtained by (a) vanilla NeRF [18], (b) PWC-Net + NeRF [14, 18], (c) Ha-NeRF [3], (d) NeRF-W [17], and (e) our method. Please visit our project page for animated results. Some results show stronger variation over the animation due to failure in extracting consistent background. + +![](images/c9da8cded0f8478f5cf866e92d83f853313fc7622468c87462f59d95a8af4cdc.jpg) +(e) Our method + +![](images/732cd18bbef128dd1f5623c456a7de2af32866f6ad2c456bf0cf8776884a62dd.jpg) + +Baselines. To our knowledge, there is no similar work that construct an occlusion-free scene based on NeRF representation. We design three baselines for comparison by considering both the contributions from occlusion removal and NeRF to verify our performance advantage. 1) A state-of + +Table 1. Quantitative comparison results. $\uparrow (\downarrow)$ indicates larger (smaller) values are better, and bold font indicates the best results. + +
MethodPSNR↑SSIM↑LPIPS↓
PWC-Net+NeRF [14,18]18.250.790.45
NeRF-W [17]11.060.450.61
Ha-NeRF [3]15.410.700.41
Proposed19.390.830.37
+ +![](images/dda4b2d506b49a7a9d5204b5a6a4484b967e6a89ac4d7d7316fc9814eb44b471.jpg) + +![](images/d9b881336a0ad94df19a9a180ddd72abfda0e3ce6d65325a603801ef59b3bcdf.jpg) + +![](images/d0efb166d2b78e6326e50be32c8e55cb89d94149659bf44b36f4a5712e42012d.jpg) + +![](images/cf7fccab72517023724dcd45ddf735aa9232a87481d760efb4746185e41685e0.jpg) +(a) +(d) +Figure 5. Visual quality comparisons for occlusion removal on SCRIBBLE1 scene. (a) is one image in the validation set with occlusion and (f) is ground truth of this view) with results from: (b) results obtained by Ha-NeRF [3], (c) results obtained by NeRF-W [17], (d) results obtained by occlusion removal [14] + NeRF [18], and (e) our method. + +![](images/55634547e8c938be4d0af2152bf4e49abc91c15ecbac6b4a3ee7e2aff6fbcf5f.jpg) +(b) +(e) + +![](images/aae097cf4956e10dd9833088135f0d97fb6542d569e218daf35766aeec7f88df.jpg) +(c) +(f) + +the-art occlusion removal method for image sequences or videos (referred to as PWC-Net [14] based on the main pillar of the work) + NeRF [18]: In this setting, PWC-Net [14] serves as a pre-processing tool to eliminate occlusion before training the NeRF [18]. 2) NeRF-W [17]: This is a method designed for separating a complete scene into two components. 3) Ha-NeRF [3]: This is another method that decomposes the scenes into distinct parts. We compare those settings with our method on our evaluation dataset, as listed in Fig. 3. For fairness of comparison, we evenly sample 128 points along the ray without applying importance sampling. We also keep the trainable parameters in each baseline approximately the same with the proposed method. + +Evaluation methodology. We conduct both qualitative and quantitative evaluations on our dataset. For quantitative evaluation, we compare reconstruction quality of the estimated occlusion-free images for viewpoints with corresponding ground truth by PSNR, SSIM, and LPIPS [32]. + +# 5.1. Qualitative evaluation + +The qualitative results of the rendered novel occlusion-free views are presented in Figs. 4 and 5. In Fig. 4, we use the scenes of FENCE1, STATUE and WIRE1 in our dataset to show that our method can reliably reconstruct novel background views, and it works no matter whether the occlusions cover a large area or have an irregular shape. In Fig. 5, we take one frame in SCRIBBLE1 with ground truth to show + +![](images/a72071363551b10fae0b0e46b7797e3f73808acb0a43cea0c9afb5cf3af8cf03.jpg) +(a) +Figure 6. Visual quality comparisons for occlusion removal on RAINDROP scene, with results from (a) an image-based raindrop removal method [22], (b) an image-based occlusion removal method [14], and (c) our method. Our method outperforms the baselines in the highlighted red boxes. + +![](images/969a8f43eee4fc0ac4fb4c07c19c257abeee61c15007018b759ddd91e14842ae.jpg) +(b) + +![](images/f8c4977a011adc2138b3b87bc242faadf36c5521f0dcfc707e82aade4b072453.jpg) +(c) + +the fidelity of our results. + +PWC-Net + NeRF [14, 18] only achieves comparable results in FENCE1 where the occlusions are regularly shaped, but shows poor artifacts on unseen types of occlusions, such as the irregularly-shaped STATUE and WIRE1, as shown in Fig. 4(b). This shows their reliance on training data. Besides, the occlusion removal method can lead to an oversmoothing issue which undermines the reconstruction accuracy. Meanwhile, the results from Ha-NeRF [3] and NeRF-W [17] in Figs. 4(c) and 4(d) either shows degenerated details or still preserves occlusions, which verifies our claim that they cannot handle the consistent occlusions observed in a continuous image sequence. The results from our method in Fig. 4(e) clearly remove these occlusions with different shapes and distributions. + +Our method relies on an accurate separation between foreground and background, but the assumption may be violated if obstructions are semi-transparent. We conduct experiments on the RAINDROP scene in our validation dataset to evaluate the robustness of our method under such scenario. With the results displayed in Fig. 6, our method can also effectively remove the raindrop in this example. It achieves competitive or even better results in some regions when compared with a method specifically designed for raindrop removal [22] as well as PWC-Net [14] designed to remove occlusions including raindrop. + +# 5.2. Quantitative evaluation + +The quantitative results of the occlusion-free novel view in Tab. 1 validate the observation in Figs. 4 and 5. Higher PSNR values show that our method can render occlusion-free novel views and recover the color information with higher accuracy. Higher SSIM values indicate that our method can preserve the structural information with high-frequency details. Lower LPIPS values show that recovered + +Table 2. Ablation study results aggregated by the types of occlusions. We compare the rendered results of the scene MLP with NeRF [18] using COLMAP [23] as camera parameter estimation and NeRF-- [28] initialized with COLMAP camera parameters. + +
MetricsPSNR ↑ / SSIM ↑ / LPIPS↓
MethodCOLMAP [23]NeRF-- [28]Scene MLP
DatasetFENCE{1,2,3}24.01 / 0.73 / 0.3818.67 / 0.54 / 0.5125.46 / 0.72 / 0.35
SCRIBBLE{1,2,3}22.70 / 0.78 / 0.3117.98 / 0.52 / 0.4623.85 / 0.84 / 0.34
WIRES{1,2}26.21 / 0.93 / 0.2022.28 / 0.88 / 0.2534.44 / 0.95 / 0.23
RAINDROP27.69 / 0.87 / 0.2627.78 / 0.83 / 0.2528.69 / 0.86 / 0.25
STATUE27.53 / 0.84 / 0.3318.49 / 0.56 / 0.4927.08 / 0.80 / 0.37
+ +images by our method better aligns with human perception. + +# 5.3. Ablation Study + +Our method consists of two major components: joint optimization for pose refinement and a selective supervision strategy to eliminate occlusions. Since the selective supervision scheme largely bases on the intermediate information of scene reconstruction MLP, a high-quality reconstruction of the scene with occlusion is crucial to our performance. + +In the ablation study, we first replace pose refinement with COLMAP [23]. From the results shown in Fig. 7(b), the occlusions largely remain, because the fixed erroneous camera parameters makes distinguishing scene components from scene MLP infeasible. + +We also explore the results without the selective supervision scheme, while maintaining pose refinement, namely only output of the scene MLP. In this setting, our method cannot remove occlusions, but it can reconstruct the scene containing occlusions with a high fidelity, as shown in Fig. 7(c). It can also reconstruct the scene with higher fidelity than NeRF based on COLMAP [23], as shown quantitatively in Tab. 2. Our complete model shows the best results for majority of scenes in our dataset. + +![](images/bf5ee8ca67981648d22d756febd1b56846ecbb79837458b809d70a1cbfc438ef.jpg) +(a) +Figure 7. Visual comparisons for the ablation study on the FENCE1 scene with results from: (a) our complete method, (b) the method without joint optimization, and (c) the method without selective supervisions. Please visit our project page for animated results. + +![](images/d31575ea6e7a88fb837fe24bd44153cdd1d6e2f395605b26d9aded90c44f98d1.jpg) +(b) + +![](images/1dfb77ab3bbda8f496972a0ad9f7dd965e64cb88cfe070a2b127fde2291c88e2.jpg) +(c) + +# 6. Conclusions + +In this paper, grounded on Neural Radiance Fields, we construct an occlusion-free scene representation, which is capable of rendering occlusion-free images with desired viewpoints. By directly using the viewpoints as a part of the input, our method is able to aggregate occluded details + +![](images/1c7c7da4d783196c6225b0bffd6047fb0b3026b1d59cf2f64fda5258cd56bebf.jpg) +(a) + +![](images/1bc9fc9c305487a108c62887045dfe07e7fb6643ce2b5fe2e6b6dac8edbbb849.jpg) +(b) +Figure 8. A failure scene of the proposed method, when reflection is present as occlusion. (a) input sample, (b) results of the scene MLP, (c) failed reflection removal with the background MLP. + +![](images/7d15b9f0b0151f2ba1d655aef24dfa0fdb904335e48fe224fb04d46cafae02e7.jpg) +(c) + +from multiple viewpoints without relying on any external guidance. We propose a pose refinement scheme to ensure robust training when the camera poses cannot be accurately estimated by COLMAP [23]. To focus on modeling of desired background scenes within a specific viewpoint, we further introduce a depth constraint to probe the occluded areas by measuring the depth of occlusion and background. Experimental results show that our method is capable of achieving promising results under a variety of scenarios. + +Limitations. Despite the promising performance of our proposed method, several limitations are still to be addressed in our future study. First, our method clearly assumes the undesired layers are in the foreground. It cannot disambiguate the cases where details from different layers are intertwined, such as scenes with reflection, which is especially challenging due to its semi-transparency and widespread presence. We show a failure case of the proposed method in a scene with reflection as occlusion in Fig. 8. Besides, since several variants of NeRF have been proposed for faster convergence and more accurate reconstruction [5,20], advanced NeRF variants may boost the performance of the proposed method. In addition, we also hope to reduce the number of images required for joint optimization of camera parameters and NeRF in our future work. + +Acknowledgement. This work is supported by National Natural Science Foundation of China under Grant No. 62136001, 62088102. Renjie Wan is supported by the Blue Sky Research Fund of HKBU under Grant No. BSRF/21-22/16 and Guangdong Basic and Applied Basic Research Foundation under Grant No. 2022A1515110692. + +# References + +[1] Coloma Ballester, Marcelo Bertalmio, Vicent Caseles, Guillermo Sapiro, and Joan Verdera. Filling-in by joint interpolation of vector fields and gray levels. IEEE Transactions on Image Processing, 2001. 2 +[2] Marcelo Bertalmio, Guillermo Sapiro, Vincent Caseles, and Coloma Ballester. Image inpainting. In Proc. of ACM SIGGRAPH, 2000. 2 +[3] Xingyu Chen, Qi Zhang, Xiaoyu Li, Yue Chen, Ying Feng, Xuan Wang, and Jue Wang. Hallucinated neural radiance fields in the wild. In CVPR, 2022. 1, 2, 6, 7 +[4] Chen Du, Byeongkeun Kang, Zheng Xu, Ji Dai, and Truong Nguyen. Accurate and efficient video de-fencing using convolutional neural networks and temporal information. In Proc. of International Conference on Multimedia and Expo, 2018. 1, 2 +[5] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In CVPR, 2022. 8 +[6] Xiaodong Gu, Zhiwen Fan, Siyu Zhu, Zuozhuo Dai, Feitong Tan, and Ping Tan. Cascade cost volume for high-resolution multi-view stereo and stereo matching. In CVPR, 2020. 4 +[7] Yuan-Chen Guo, Di Kang, Linchao Bao, Yu He, and Song-Hai Zhang. NeRFReN: Neural radiance fields with reflections. In CVPR, 2022. 1, 2, 5, 6 +[8] Xin Huang, Qi Zhang, Ying Feng, Hongdong Li, Xuan Wang, and Qing Wang. HDR-NeRF: High dynamic range neural radiance fields. In CVPR, 2022. 2, 6 +[9] Jeffrey Ichnowski, Yahav Avigal, Justin Kerr, and Ken Goldberg. Dex-NeRF: Using a neural radiance field to grasp transparent objects. In Proc. of Conference on Robot Learning, 2021. 2 +[10] Sankaraganesh Jonna, Sukla Satapathy, and Rajiv R Sahay. Stereo image de-fencing using smartphones. In Proc. of International Conference on Acoustics, Speech, and Signal Processing, 2017. 1, 2 +[11] Jingyuan Li, Ning Wang, Lefei Zhang, Bo Du, and Dacheng Tao. Recurrent feature reasoning for image inpainting. In CVPR, 2020. 2 +[12] Xiaoyu Li, Bo Zhang, Jing Liao, and Pedro V Sander. Let's see clearly: Contaminant artifact removal for moving cameras. In Proc. of International Conference on Computer Vision, 2021. 1 +[13] Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. BARF: Bundle-adjusting neural radiance fields. In CVPR, 2021. 2 +[14] Yu-Lun Liu, Wei-Sheng Lai, Ming-Hsuan Yang, Yung-Yu Chuang, and Jia-Bin Huang. Learning to see through obstructions. In CVPR, 2020. 1, 2, 5, 6, 7 +[15] Ilya Loshchilov and Frank Hutter. SGDR: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 6 +[16] Li Ma, Xiaoyu Li, Jing Liao, Qi Zhang, Xuan Wang, Jue Wang, and Pedro V Sander. Deblur-NeRF: Neural radiance fields from blurry images. In CVPR, 2022. 2 + +[17] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. NeRF in the Wild: Neural radiance fields for unconstrained photo collections. In CVPR, 2021. 1, 2, 6, 7 +[18] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In Proc. of European Conference on Computer Vision, 2020. 1, 2, 3, 6, 7, 8 +[19] Yadong Mu, Wei Liu, and Shuicheng Yan. Video de-fencing. IEEE Transactions on Circuits and Systems for Video Technology, 2014. 1, 2 +[20] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. TOG, 2022. 8 +[21] Kamyar Nazeri, Eric Ng, Tony Joseph, Faisal Z Qureshi, and Mehran Ebrahimi. EdgeConnect: Generative image inpainting with adversarial edge learning. arXiv preprint arXiv:1901.00212, 2019. 2 +[22] Rui Qian, Robby T Tan, Wenhan Yang, Jiajun Su, and Jiaying Liu. Attentive generative adversarial network for raindrop removal from a single image. In CVPR, 2018. 1, 7 +[23] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In CVPR, 2016. 2, 4, 6, 8 +[24] Yuhang Song, Chao Yang, Yeji Shen, Peng Wang, Qin Huang, and C-C Jay Kuo. SPG-Net: Segmentation prediction and guidance network for image inpainting. In Proc. of British Machine Vision Conference, 2018. 2 +[25] V. Vaish, M. Levoy, R. Szeliski, C.L. Zitnick, and Sing Bing Kang. Reconstructing occluded surfaces using synthetic apertures: Stereo, focus and robust measures. In CVPR, 2006. 2 +[26] Suhani Vora, Noha Radwan, Klaus Greff, Henning Meyer, Kyle Genova, Mehdi SM Sajjadi, Etienne Pot, Andrea Tagliasacchi, and Daniel Duckworth. NeSF: Neural semantic fields for generalizable semantic segmentation of 3d scenes. arXiv preprint arXiv:2111.13260, 2021. 2 +[27] Renjie Wan, Boxin Shi, Haoliang Li, Ling-Yu Duan, and Alex C Kot. Face image reflection removal. International Journal of Computer Vision, 2021. 1, 5 +[28] Zirui Wang, Shangzhe Wu, Weidi Xie, Min Chen, and Victor Adrian Prisacariu. NeRF---: Neural radiance fields without known camera parameters. arXiv preprint arXiv:2102.07064, 2021. 2, 4, 8 +[29] Tianfan Xue, Michael Rubinstein, Ce Liu, and William T Freeman. A computational approach for obstruction-free photography. TOG, 2015. 1, 2, 6 +[30] Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. MVSNet: Depth inference for unstructured multiview stereo. In Proc. of European Conference on Computer Vision, 2018. 4 +[31] Renjiao Yi, Jue Wang, and Ping Tan. Automatic fence segmentation in videos of dynamic scenes. In CVPR, 2016. 6 +[32] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 7 + +[33] Shuaifeng Zhi, Tristan Laidlow, Stefan Leutenegger, and Andrew J Davison. In-place scene labelling and understanding with implicit scene representation. In Proc. of International Conference on Computer Vision, 2021. 2 +[34] Manyu Zhu, Dongliang He, Xin Li, Chao Li, Fu Li, Xiao Liu, Errui Ding, and Zhaoxiang Zhang. Image inpainting by end-to-end cascaded refinement with mask awareness. IEEE Transactions on Image Processing, 2021. 2 \ No newline at end of file diff --git a/2023/Occlusion-Free Scene Recovery via Neural Radiance Fields/images.zip b/2023/Occlusion-Free Scene Recovery via Neural Radiance Fields/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..cbac0bd2017cc4150f13bcb0b0c1b646ce780a1e --- /dev/null +++ b/2023/Occlusion-Free Scene Recovery via Neural Radiance Fields/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a944089cdfbc995c26be740600888bc50b8a7cacf8ecfe61a36bbd4f557cbe29 +size 495624 diff --git a/2023/Occlusion-Free Scene Recovery via Neural Radiance Fields/layout.json b/2023/Occlusion-Free Scene Recovery via Neural Radiance Fields/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..246233f4ade5c7cec3c1c9ccc0c8c21f19689b54 --- /dev/null +++ b/2023/Occlusion-Free Scene Recovery via Neural Radiance Fields/layout.json @@ -0,0 +1,10694 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 114, + 103, + 479, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 103, + 479, + 120 + ], + "spans": [ + { + "bbox": [ + 114, + 103, + 479, + 120 + ], + "type": "text", + "content": "Occlusion-Free Scene Recovery via Neural Radiance Fields" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 135, + 142, + 460, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 142, + 460, + 158 + ], + "spans": [ + { + "bbox": [ + 135, + 142, + 460, + 158 + ], + "type": "text", + "content": "Chengxuan Zhu" + }, + { + "bbox": [ + 135, + 142, + 460, + 158 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 135, + 142, + 460, + 158 + ], + "type": "text", + "content": " Renjie Wan" + }, + { + "bbox": [ + 135, + 142, + 460, + 158 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 135, + 142, + 460, + 158 + ], + "type": "text", + "content": " Yunkai Tang" + }, + { + "bbox": [ + 135, + 142, + 460, + 158 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 135, + 142, + 460, + 158 + ], + "type": "text", + "content": " Boxin Shi" + }, + { + "bbox": [ + 135, + 142, + 460, + 158 + ], + "type": "inline_equation", + "content": "^{1,2*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 77, + 158, + 514, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 158, + 514, + 213 + ], + "spans": [ + { + "bbox": [ + 77, + 158, + 514, + 213 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 77, + 158, + 514, + 213 + ], + "type": "text", + "content": "National Engineering Research Center of Visual Technology, School of Computer Science, Peking University " + }, + { + "bbox": [ + 77, + 158, + 514, + 213 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 77, + 158, + 514, + 213 + ], + "type": "text", + "content": "National Key Laboratory for Multimedia Information Processing, School of Computer Science, Peking University " + }, + { + "bbox": [ + 77, + 158, + 514, + 213 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 77, + 158, + 514, + 213 + ], + "type": "text", + "content": "Department of Computer Science, Hong Kong Baptist University {peterzhu, shiboxin}@pku.edu.cn, renjiewan@hkbu.edu.hk, tangyunkai@stu.pku.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 241, + 192, + 254 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 241, + 192, + 254 + ], + "spans": [ + { + "bbox": [ + 143, + 241, + 192, + 254 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 266, + 290, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 266, + 290, + 494 + ], + "spans": [ + { + "bbox": [ + 47, + 266, + 290, + 494 + ], + "type": "text", + "content": "Our everyday lives are filled with occlusions that we strive to see through. By aggregating desired background information from different viewpoints, we can easily eliminate such occlusions without any external occlusion-free supervision. Though several occlusion removal methods have been proposed to empower machine vision systems with such ability, their performances are still unsatisfactory due to reliance on external supervision. We propose a novel method for occlusion removal by directly building a mapping between position and viewing angles and the corresponding occlusion-free scene details leveraging Neural Radiance Fields (NeRF). We also develop an effective scheme to jointly optimize camera parameters and scene reconstruction when occlusions are present. An additional depth constraint is applied to supervise the entire optimization without labeled external data for training. The experimental results on existing and newly collected datasets validate the effectiveness of our method. Our project page: https://freebutuselessssoul.github.io/occnerf." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 529, + 128, + 542 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 529, + 128, + 542 + ], + "spans": [ + { + "bbox": [ + 47, + 529, + 128, + 542 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 549, + 287, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 549, + 287, + 670 + ], + "spans": [ + { + "bbox": [ + 46, + 549, + 287, + 670 + ], + "type": "text", + "content": "Neural Radiance Fields (NeRF) are capable of learning the scene representation implicitly from a set of 2D images, yet not every scene is favored by observers. Many undesirable occlusions in our world obscure details that are essential to our understanding of the world. In general, such obstructions range from water droplets and scribbles on a piece of glass, to fences or any objects occluding the desired scenes (e.g., a statue closer to the camera in a landmark scene). How to apply computational methods to exclude them from the scene representation is of great interest." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 670, + 287, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 287, + 694 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 287, + 694 + ], + "type": "text", + "content": "Occlusion removal (e.g., [29]) is the direct solution to achieve this goal. However, explicit occlusion removal may" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 242, + 547, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 242, + 547, + 457 + ], + "spans": [ + { + "bbox": [ + 304, + 242, + 547, + 457 + ], + "type": "text", + "content": "oversmooth essential details necessary for clearly observing the desired background scenes. In addition, current methods mainly depend on external occlusion-free supervision (e.g., fence removal [4], raindrop removal [22]) to develop the reliable capability in removing certain types of occlusions. Once encountering a new scenario with unseen occlusion types beyond their training data, these methods might show degraded performances. To handle more diverse types of occlusions, generic constraints from multiple viewpoints are widely adopted [4, 10, 12, 14, 19, 27, 29] via mimicking our human vision systems, who can easily piece together the desired background scenes by looking at them from different viewpoints. But the majority of these methods just consider viewpoints as a prior in relation to spatial correlations. Their backbones still rely on external training data with corresponding ground truth for optimization, which still does not fundamentally alleviate the difficulty of handling diverse occlusions in the real world." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 462, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 547, + 715 + ], + "type": "text", + "content": "An occlusion-free world can be progressively aggregated by seeing its occluded part from different viewing directions to reveal occlusions previously unobservable in each single perspective, as illustrated in the left part of Fig. 1 (the fan is occluded in the target view). Since NeRF [18] employs an implicit representation to map viewpoints to pixels, one may come to the naive solution of directly constructing a NeRF which is optimized across multiple viewpoints. However, the vanilla NeRF [18] representing the scene as a whole is not able to treat occlusion and background scenes distinctively, and as long as the occlusion remains static, NeRF is designed to faithfully reconstruct its presence. Meanwhile, many NeRF variants can decompose the whole scene into different components (e.g., NeRF-W [17], Ha-NeRF [3], NeRFRen [7]), but they cannot handle the real-world static occlusions. This is because NeRF-W [17] and Ha-NeRF [3] rely on the inconsistency of undesired components across different views to achieve such separation, which is difficult to be observed in a continuous 3D world. On the other hand, NeRFRen [7] only works in separating the transmission and reflection compo" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20722" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 286, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 286, + 95 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 286, + 95 + ], + "type": "text", + "content": "nents caused by semi-transparent planar glass, which is incapable of handling opaque occlusions in the real world." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 96, + 286, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 96, + 286, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 96, + 286, + 251 + ], + "type": "text", + "content": "Another problem comes from NeRF's reliance on camera parameters pre-computed by COLMAP [23], because handcrafted features extracted and matched using COLMAP [23] are for the whole scene, and are incapable in distinguishing between undesired occlusions and the desired background. When the features from occlusion dominate the matching process, the obtained camera parameters cannot faithfully model the spatial correlation of background scenes across multiple viewpoints. Besides, COLMAP [23] is not a stable option for pose estimation in the real world [28]. The existence of occlusions may prevent it from working properly, making the occlusion-free scene representation infeasible." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 251, + 286, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 251, + 286, + 501 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 286, + 501 + ], + "type": "text", + "content": "In this paper, we aim at seeing through the occluded scenes by developing an occlusion-free scene representation without considering specific occlusion types, based on which we can render any occlusion-free images from desired viewpoints. Our method first maps viewing angles and their corresponding scene details by leveraging NeRF. We then introduce a depth constraint to probe the occluded areas by measuring the depth of occlusion and background, by assuming that occlusions are always in the foreground with closer distance. During the scene modeling process, a pose refinement scheme is further introduced to refine the camera pose with the features of the background scene. As outlined in Fig. 1, our pipeline contains three modules to achieve the above goals: 1) a scene reconstruction module to represent the whole scene using NeRF (with occlusions), 2) a cost volume construction module to gather information from neighboring views as guidance (to indicate where occlusions are), and 3) a selective supervision scheme to constrain another NeRF on the desired background information (occlusions removed), and our contributions can be summarized as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 502, + 286, + 575 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 58, + 502, + 286, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 502, + 286, + 526 + ], + "spans": [ + { + "bbox": [ + 58, + 502, + 286, + 526 + ], + "type": "text", + "content": "- an occlusion-free representation without relying on any external prior as supervisory knowledge;" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 528, + 286, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 528, + 286, + 550 + ], + "spans": [ + { + "bbox": [ + 58, + 528, + 286, + 550 + ], + "type": "text", + "content": "- a joint optimization of pose refinement and scene reconstruction by effective multi-view feature fusion;" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 552, + 286, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 552, + 286, + 575 + ], + "spans": [ + { + "bbox": [ + 58, + 552, + 286, + 575 + ], + "type": "text", + "content": "- a selective supervision scheme to probe the occluded areas guided by the scene depth information." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 46, + 582, + 286, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 286, + 641 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 286, + 641 + ], + "type": "text", + "content": "Based on the experiments with a dataset containing diverse types of occlusions, the proposed method can eliminate occlusions including scribbles and water droplets on a piece of glass, fences, and even irregular-shaped statues without relying on any external supervisions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 651, + 133, + 662 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 651, + 133, + 662 + ], + "spans": [ + { + "bbox": [ + 47, + 651, + 133, + 662 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 671, + 172, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 172, + 682 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 172, + 682 + ], + "type": "text", + "content": "2.1. NeRF and its variants" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 286, + 713 + ], + "type": "text", + "content": "NeRF becomes a popular choice for implicit volumetric scene representation [18]. By mapping the point and current" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 72, + 545, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 262 + ], + "type": "text", + "content": "viewing direction to its color and density using a neural network and a differentiable rendering scheme, it provides an effective way to learn from multiple viewing angles. Such ability has been explored to complement the missing depth for grasping transparent objects [9]. Based on NeRF, a series of methods have been proposed to tackle the problems like scene understanding [33] and reasoning [26]. Besides, by considering specific physical priors, NeRF-based frameworks also show promising performance in learning clear representation from challenging scenarios [7,8,16]. Recent NeRF variants can even reconstruct the scene from input with various perturbations. For example, NeRF-W [17] and Ha-NeRF [3] can separate the transient objects from the whole scene. However, they target occlusions that are inconsistent throughout the image set and become ineffective when such occlusions remain static and consistent." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 266, + 545, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 266, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 304, + 266, + 545, + 361 + ], + "type": "text", + "content": "Since NeRF relies on pre-computed camera parameters for scene representation, it fails when the pre-computation is not feasible. Some methods (e.g., NeRF-- [28] and BARF [13]) have been proposed to alleviate this issue by optimizing the camera parameters along with the scene representation. However, how to avoid the interference from undesired scene during the camera parameter optimization remains a problem to be solved." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 378, + 416, + 389 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 378, + 416, + 389 + ], + "spans": [ + { + "bbox": [ + 306, + 378, + 416, + 389 + ], + "type": "text", + "content": "2.2. Occlusion removal" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 399, + 545, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 399, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 304, + 399, + 545, + 483 + ], + "type": "text", + "content": "Traditional methods in occlusion removal address this problem by propagating neighboring pixels via anisotropic diffusion [2] or solving differential equations [1]. Recently, there are more approaches solving this problem in a learning-based manner [11, 34]. A recent trend is to employ the edge information [21] or segmentation masks [24] as a stronger prior to alleviate the ill-posedness of this problem." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 545, + 713 + ], + "type": "text", + "content": "Researchers in this area have been aware of the importance of removing occlusions from multiple viewpoints. Traditional methods try to take the benefits of multiple viewpoints by computing a disparity map [10], dense flow field [29], or visual parallax [19]. Using a light field camera that records the spatial and angular information of light rays in the space is also helpful in removing occlusions [25]. However, these methods also pose restrictions on their inputs, such as using stereo image pairs [10], requiring the background to be visible in at least one of the input views [25, 29], or requesting obviously relative motion between occlusions and backgrounds [19]. It is also a recent trend to apply deep learning in guiding network optimization with fewer constraints on input. For example, some methods seek to use the temporal information [4] or the optical flow [14] from video frames to guide the occlusion-free recovery process. However, these methods are confined to small movements of the camera, partially due to the difficulty in 3D scene representation." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20723" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 70, + 532, + 222 + ], + "blocks": [ + { + "bbox": [ + 65, + 70, + 532, + 222 + ], + "lines": [ + { + "bbox": [ + 65, + 70, + 532, + 222 + ], + "spans": [ + { + "bbox": [ + 65, + 70, + 532, + 222 + ], + "type": "image", + "image_path": "a0275a5899b29a71caf6f47a0dedd3e6df3c6d8c135b8ed9e8512222ce1b4c0d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 229, + 545, + 307 + ], + "lines": [ + { + "bbox": [ + 46, + 229, + 545, + 307 + ], + "spans": [ + { + "bbox": [ + 46, + 229, + 545, + 307 + ], + "type": "text", + "content": "Figure 1. Pipeline of the proposed method. By taking the positional encoding [18] of position and viewing direction, the \"scene MLP\" (colored in blue) models the scene with occlusion faithfully, and the \"background MLP\" (colored in green) models the background only. To this end, a cost volume is constructed using images captured from neighboring reference views, guiding the joint optimization of scene MLP and camera parameters, as we discuss in Sec. 4.1. To supervise the training of background MLP and remove occlusions, we propose to aggregate the information from scene MLP about whether the output of background NeRF should be similar to the observed color, by learning a supervision mask from the weights along the ray using a \"mask MLP\" (colored in amber), as explained in Sec. 4.2. The reference and target views constitute the family of neighboring images whose features are warped to construct the cost volume." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 317, + 124, + 331 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 317, + 124, + 331 + ], + "spans": [ + { + "bbox": [ + 47, + 317, + 124, + 331 + ], + "type": "text", + "content": "3. Preliminary" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 337, + 287, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 337, + 287, + 420 + ], + "spans": [ + { + "bbox": [ + 46, + 337, + 287, + 420 + ], + "type": "text", + "content": "Our goal is to generate an occlusion-free scene representation given a collection of " + }, + { + "bbox": [ + 46, + 337, + 287, + 420 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 337, + 287, + 420 + ], + "type": "text", + "content": " images " + }, + { + "bbox": [ + 46, + 337, + 287, + 420 + ], + "type": "inline_equation", + "content": "\\mathcal{I} = \\{I_1, I_2, \\ldots, I_N\\}" + }, + { + "bbox": [ + 46, + 337, + 287, + 420 + ], + "type": "text", + "content": " captured from " + }, + { + "bbox": [ + 46, + 337, + 287, + 420 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 337, + 287, + 420 + ], + "type": "text", + "content": " different viewpoints, where the " + }, + { + "bbox": [ + 46, + 337, + 287, + 420 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 337, + 287, + 420 + ], + "type": "text", + "content": "-th viewpoint is defined by the extrinsic matrix " + }, + { + "bbox": [ + 46, + 337, + 287, + 420 + ], + "type": "inline_equation", + "content": "[\\mathbf{R}_i | \\mathbf{t}_i]" + }, + { + "bbox": [ + 46, + 337, + 287, + 420 + ], + "type": "text", + "content": " and the camera is defined by a shared intrinsic matrix " + }, + { + "bbox": [ + 46, + 337, + 287, + 420 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 46, + 337, + 287, + 420 + ], + "type": "text", + "content": ". Together they form the camera matrices " + }, + { + "bbox": [ + 46, + 337, + 287, + 420 + ], + "type": "inline_equation", + "content": "\\mathcal{P} = \\{\\mathbf{P}_1, \\mathbf{P}_2, \\ldots, \\mathbf{P}_N\\}" + }, + { + "bbox": [ + 46, + 337, + 287, + 420 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 337, + 287, + 420 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_i = \\mathbf{K}[\\mathbf{R}_i | \\mathbf{t}_i]" + }, + { + "bbox": [ + 46, + 337, + 287, + 420 + ], + "type": "text", + "content": ", as illustrated in the left part of Fig. 1." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 421, + 287, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 421, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 421, + 287, + 529 + ], + "type": "text", + "content": "As we follow the approach of NeRF [18], we briefly introduce it in this section for self-contained purposes. NeRF uses a neural network " + }, + { + "bbox": [ + 46, + 421, + 287, + 529 + ], + "type": "inline_equation", + "content": "F_{\\Theta}" + }, + { + "bbox": [ + 46, + 421, + 287, + 529 + ], + "type": "text", + "content": " to construct a radiance field, which maps the 3D coordinate of a point " + }, + { + "bbox": [ + 46, + 421, + 287, + 529 + ], + "type": "inline_equation", + "content": "\\mathbf{x} = (x,y,z)" + }, + { + "bbox": [ + 46, + 421, + 287, + 529 + ], + "type": "text", + "content": " and viewing direction " + }, + { + "bbox": [ + 46, + 421, + 287, + 529 + ], + "type": "inline_equation", + "content": "\\mathbf{d} = (\\theta ,\\phi)" + }, + { + "bbox": [ + 46, + 421, + 287, + 529 + ], + "type": "text", + "content": " into the point's radiance color " + }, + { + "bbox": [ + 46, + 421, + 287, + 529 + ], + "type": "inline_equation", + "content": "\\mathbf{c} = (r,g,b)" + }, + { + "bbox": [ + 46, + 421, + 287, + 529 + ], + "type": "text", + "content": " and density " + }, + { + "bbox": [ + 46, + 421, + 287, + 529 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 46, + 421, + 287, + 529 + ], + "type": "text", + "content": ". Then, the volume rendering scheme is applied to render an image. For each training image and every pixel " + }, + { + "bbox": [ + 46, + 421, + 287, + 529 + ], + "type": "inline_equation", + "content": "\\mathbf{p} = (u,v)" + }, + { + "bbox": [ + 46, + 421, + 287, + 529 + ], + "type": "text", + "content": ", a ray is emitted from the camera position " + }, + { + "bbox": [ + 46, + 421, + 287, + 529 + ], + "type": "inline_equation", + "content": "\\mathbf{o}" + }, + { + "bbox": [ + 46, + 421, + 287, + 529 + ], + "type": "text", + "content": " at direction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 70, + 533, + 287, + 563 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 533, + 287, + 563 + ], + "spans": [ + { + "bbox": [ + 70, + 533, + 287, + 563 + ], + "type": "interline_equation", + "content": "\\mathbf {d} _ {i} (u, v) = \\mathbf {R} _ {i} \\left[ \\frac {u - W / 2}{f}, - \\frac {v - H / 2}{f}, - 1 \\right] ^ {\\top}, \\tag {1}", + "image_path": "0e80ac60025d1bd3e7a95564e827be7519acd53cca7958d4e51d1ae4c34c901e.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 567, + 287, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 567, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 46, + 567, + 287, + 613 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 567, + 287, + 613 + ], + "type": "inline_equation", + "content": "W, H," + }, + { + "bbox": [ + 46, + 567, + 287, + 613 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 567, + 287, + 613 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 567, + 287, + 613 + ], + "type": "text", + "content": " are the width, height, and focal length of the input image, respectively. Marching through the ray " + }, + { + "bbox": [ + 46, + 567, + 287, + 613 + ], + "type": "inline_equation", + "content": "\\mathbf{r}(t) = \\mathbf{o} + t\\mathbf{d}" + }, + { + "bbox": [ + 46, + 567, + 287, + 613 + ], + "type": "text", + "content": " and aggregating the colors and densities provide us the final color" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 86, + 618, + 287, + 647 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 618, + 287, + 647 + ], + "spans": [ + { + "bbox": [ + 86, + 618, + 287, + 647 + ], + "type": "interline_equation", + "content": "\\hat {I} _ {i} (\\mathbf {p}) = \\int_ {t _ {n}} ^ {t _ {f}} T (t) \\sigma (\\mathbf {r} (t)) \\mathbf {c} (\\mathbf {r} (t), \\mathbf {d} _ {i}) d t, \\tag {2}", + "image_path": "1b824ccc8ba95a00eaf1e4e538c090a6e585586c49c279485fd285f4648c24e6.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 652, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 652, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 652, + 287, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 652, + 287, + 713 + ], + "type": "inline_equation", + "content": "T(t) = \\exp(-\\int_{t_n}^t \\sigma(\\mathbf{r}(s)) ds)" + }, + { + "bbox": [ + 46, + 652, + 287, + 713 + ], + "type": "text", + "content": " stands for the accumulated transmittance along the ray, or intuitively, the probability of the emitted radiance of " + }, + { + "bbox": [ + 46, + 652, + 287, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{r}(t)" + }, + { + "bbox": [ + 46, + 652, + 287, + 713 + ], + "type": "text", + "content": " to reach the camera without being blocked by points in between. " + }, + { + "bbox": [ + 46, + 652, + 287, + 713 + ], + "type": "inline_equation", + "content": "t_n" + }, + { + "bbox": [ + 46, + 652, + 287, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 652, + 287, + 713 + ], + "type": "inline_equation", + "content": "t_f" + }, + { + "bbox": [ + 46, + 652, + 287, + 713 + ], + "type": "text", + "content": " are the far end and the near end of rendering respectively." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 319, + 545, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 319, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 305, + 319, + 545, + 342 + ], + "type": "text", + "content": "To make the integral tractable, points are sampled along the ray, and Eq. (2) can be rewritten into" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 337, + 349, + 545, + 381 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 349, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 337, + 349, + 545, + 381 + ], + "type": "interline_equation", + "content": "\\hat {I} _ {i} (\\mathbf {p}) = \\sum_ {k} ^ {N _ {s}} T _ {k} \\left(1 - \\exp \\left(- \\sigma_ {k} \\delta_ {k}\\right)\\right) \\mathbf {c} \\left(\\mathbf {r} _ {k}, \\mathbf {d}\\right), \\tag {3}", + "image_path": "821fe16b91a94b5e11a23df9ce5045a74a5f166cf8fe46c61c179e4ef4b57ecf.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 347, + 384, + 545, + 418 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 384, + 545, + 418 + ], + "spans": [ + { + "bbox": [ + 347, + 384, + 545, + 418 + ], + "type": "interline_equation", + "content": "T _ {k} = \\exp \\left(- \\sum_ {j = 1} ^ {k - 1} \\sigma_ {j} \\delta_ {j}\\right). \\tag {4}", + "image_path": "5ed5e57c2a95b41ef347603cae097abcecac6beabce5b2a073e7360267020ee8.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 426, + 545, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 545, + 485 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 545, + 485 + ], + "type": "text", + "content": "For the convenience of notation, " + }, + { + "bbox": [ + 304, + 426, + 545, + 485 + ], + "type": "inline_equation", + "content": "\\sigma (\\mathbf{r}_k)" + }, + { + "bbox": [ + 304, + 426, + 545, + 485 + ], + "type": "text", + "content": " is abbreviated to " + }, + { + "bbox": [ + 304, + 426, + 545, + 485 + ], + "type": "inline_equation", + "content": "\\sigma_{k}" + }, + { + "bbox": [ + 304, + 426, + 545, + 485 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 426, + 545, + 485 + ], + "type": "inline_equation", + "content": "\\delta_{k}" + }, + { + "bbox": [ + 304, + 426, + 545, + 485 + ], + "type": "text", + "content": " is defined by " + }, + { + "bbox": [ + 304, + 426, + 545, + 485 + ], + "type": "inline_equation", + "content": "t_{k + 1} - t_k" + }, + { + "bbox": [ + 304, + 426, + 545, + 485 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 304, + 426, + 545, + 485 + ], + "type": "inline_equation", + "content": "N_{s}" + }, + { + "bbox": [ + 304, + 426, + 545, + 485 + ], + "type": "text", + "content": " is the number of points sampled on the ray, " + }, + { + "bbox": [ + 304, + 426, + 545, + 485 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_k = \\mathbf{o} + t_k\\mathbf{d}" + }, + { + "bbox": [ + 304, + 426, + 545, + 485 + ], + "type": "text", + "content": " denotes the position of " + }, + { + "bbox": [ + 304, + 426, + 545, + 485 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 426, + 545, + 485 + ], + "type": "text", + "content": "-th sampling point, and " + }, + { + "bbox": [ + 304, + 426, + 545, + 485 + ], + "type": "inline_equation", + "content": "t_k" + }, + { + "bbox": [ + 304, + 426, + 545, + 485 + ], + "type": "text", + "content": " refers to the depth of the " + }, + { + "bbox": [ + 304, + 426, + 545, + 485 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 426, + 545, + 485 + ], + "type": "text", + "content": "-th sampling point seen from the target view." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 485, + 545, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 485, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 485, + 545, + 544 + ], + "type": "text", + "content": "NeRF optimizes the scene representation neural network by minimizing the photometric loss " + }, + { + "bbox": [ + 304, + 485, + 545, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{L} = \\sum_{i=1}^{N} \\|I_i - \\hat{I}_i\\|_2" + }, + { + "bbox": [ + 304, + 485, + 545, + 544 + ], + "type": "text", + "content": ", which is simply the difference between input images " + }, + { + "bbox": [ + 304, + 485, + 545, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 304, + 485, + 545, + 544 + ], + "type": "text", + "content": " and corresponding predicted images " + }, + { + "bbox": [ + 304, + 485, + 545, + 544 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{I}}" + }, + { + "bbox": [ + 304, + 485, + 545, + 544 + ], + "type": "text", + "content": ". The optimization process can be formulated as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 372, + 551, + 545, + 570 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 551, + 545, + 570 + ], + "spans": [ + { + "bbox": [ + 372, + 551, + 545, + 570 + ], + "type": "interline_equation", + "content": "\\Theta^ {*} = \\arg \\min _ {\\Theta} \\mathcal {L} (\\hat {\\mathcal {I}} | \\mathcal {I}, \\mathcal {P}). \\tag {5}", + "image_path": "b74d2c477d1881a2026f77b6dc3bb64d22a750c68f5d4ef310e94fb3aefcb42f.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 575, + 545, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 575, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 304, + 575, + 545, + 647 + ], + "type": "text", + "content": "Despite NeRF's strong capability in representing a scene from multiple viewpoints, the occlusions pose challenges to it from two aspects: 1) How to selectively aggregate background information from multiple viewpoints and 2) how to accurately estimate the camera pose of desired background in a continuous space." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 657, + 421, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 657, + 421, + 671 + ], + "spans": [ + { + "bbox": [ + 305, + 657, + 421, + 671 + ], + "type": "text", + "content": "4. Proposed approach" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "We focus on how to address the two challenges above when using NeRF to see through occlusion from multiple viewpoints and our pipeline is illustrated in Fig. 1. In" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20724" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": "Sec. 4.1, we first describe our joint optimization framework to accurately estimate the camera pose and reconstruct the scene with occlusions using a scene MLP. Then, in Sec. 4.2, we introduce how to selectively aggregate background information from multiple viewpoints based on the rectified background pose using a mask MLP. Under the guidance of mask MLP, we finally train a background MLP to reconstruct occlusion-free scenes. Note that both the scene MLP and the background MLP map the positional encoding of position and viewing direction into color and density, as described in Sec. 3." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 212, + 249, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 212, + 249, + 224 + ], + "spans": [ + { + "bbox": [ + 47, + 212, + 249, + 224 + ], + "type": "text", + "content": "4.1. Scene reconstruction with cost volume" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 231, + 287, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 231, + 287, + 373 + ], + "spans": [ + { + "bbox": [ + 46, + 231, + 287, + 373 + ], + "type": "text", + "content": "NeRF and most of its variants use COLMAP [23] as a silver bullet to estimate camera poses for the scene representation, while COLMAP [23] is not an ideal setting for building an occlusion-free representation. The extracted local features cannot be distinguished between undesired occlusion and the desired background, so the matching process can lead to an erroneous correspondence, or even directly fail to estimate the camera parameters. Some recent works try to address the dependency on COLMAP [23] by jointly optimizing camera parameters and the neural radiance fields with the photometric loss [28], achieving comparable performance to COLMAP [23]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 375, + 287, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 375, + 287, + 482 + ], + "spans": [ + { + "bbox": [ + 46, + 375, + 287, + 482 + ], + "type": "text", + "content": "However, only using photometric loss does not solve the fundamental problem, because during training, it only uses one view at a time, namely the target view itself. This makes the system ignorant of the 3D scene as a whole, and thus still cannot tell occlusions from the desired background. Besides, in the early stage of training when NeRF can hardly output meaningful information, the photometric loss cannot constrain camera extrinsics to converge to the correct rotation and translation. This can lead to a sub-optimal solution." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 482, + 287, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 482, + 287, + 566 + ], + "spans": [ + { + "bbox": [ + 46, + 482, + 287, + 566 + ], + "type": "text", + "content": "To make our method aware of occlusions and background, we design a cost volume loss inspired by multiview stereo solutions [6, 30]. It encourages the system to model the location corresponding to each pixel at the correct depth, by utilizing the feature consistency of the same point seen from neighboring reference views. Next, we will introduce how our cost volume loss is constructed." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "text", + "content": "Feature extraction. A 2D map " + }, + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "inline_equation", + "content": "F_{i}" + }, + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "text", + "content": " is extracted for every input image " + }, + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "inline_equation", + "content": "I_{i}" + }, + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "text", + "content": " by using a pre-trained VGG-19 network " + }, + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "text", + "content": " as a feature encoder as " + }, + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "inline_equation", + "content": "F_{i} = \\eta (I_{i})" + }, + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "text", + "content": ". For the convenience of notation, we define " + }, + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "inline_equation", + "content": "F_{i}(\\mathbf{p})" + }, + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "text", + "content": " as the feature vector at 2D location " + }, + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "text", + "content": ". When " + }, + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 46, + 582, + 288, + 713 + ], + "type": "text", + "content": " falls off the grid, bilinear interpolation is performed to acquire the feature vector. The extracted 2D features represent local appearance, so intuitively the feature of the same point of interest across neighboring views are assumed to be similar. On the other hand, if two feature vectors vary greatly, it is unlikely that they describe the same area." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": "Feature volume construction. To construct a feature volume at the " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": "-th viewpoint (target view), the image features from neighboring reference views should be warped with respect to the depth. The warping from the " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": "-th view to the " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": "-th view can be formulated as a " + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": " matrix" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 138, + 545, + 174 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 138, + 545, + 174 + ], + "spans": [ + { + "bbox": [ + 306, + 138, + 545, + 174 + ], + "type": "interline_equation", + "content": "\\mathbf {H} _ {i j} (z) = \\mathbf {K R} _ {i} (\\mathbf {I d} - \\frac {\\left(\\mathbf {R} _ {j} ^ {\\top} \\mathbf {t} _ {j} - \\mathbf {R} _ {i} ^ {\\top} \\mathbf {t} _ {i}\\right) \\mathbf {n} _ {j} ^ {\\top}}{z} \\cdot \\mathbf {R} _ {j}) \\mathbf {R} _ {j} ^ {\\top} \\mathbf {K} ^ {- 1}, \\tag {6}", + "image_path": "960eed59496c31691c5023a55b8faa11157584c17e9031e08fedee0754db7b50.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 175, + 545, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 175, + 545, + 247 + ], + "spans": [ + { + "bbox": [ + 304, + 175, + 545, + 247 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 175, + 545, + 247 + ], + "type": "inline_equation", + "content": "\\mathbf{Id}" + }, + { + "bbox": [ + 304, + 175, + 545, + 247 + ], + "type": "text", + "content": " is the identity matrix, and " + }, + { + "bbox": [ + 304, + 175, + 545, + 247 + ], + "type": "inline_equation", + "content": "\\mathbf{n}_j" + }, + { + "bbox": [ + 304, + 175, + 545, + 247 + ], + "type": "text", + "content": " is the principle axis of the " + }, + { + "bbox": [ + 304, + 175, + 545, + 247 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 175, + 545, + 247 + ], + "type": "text", + "content": "-th view. This matrix describes a mapping that the point at the 2D coordinate of " + }, + { + "bbox": [ + 304, + 175, + 545, + 247 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 304, + 175, + 545, + 247 + ], + "type": "text", + "content": " and the depth of " + }, + { + "bbox": [ + 304, + 175, + 545, + 247 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 304, + 175, + 545, + 247 + ], + "type": "text", + "content": " in warped feature map corresponds to the point at " + }, + { + "bbox": [ + 304, + 175, + 545, + 247 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_{ij}(z)\\mathbf{p}" + }, + { + "bbox": [ + 304, + 175, + 545, + 247 + ], + "type": "text", + "content": " in the original feature map. Applying the warping matrix to the feature map " + }, + { + "bbox": [ + 304, + 175, + 545, + 247 + ], + "type": "inline_equation", + "content": "F_{i}" + }, + { + "bbox": [ + 304, + 175, + 545, + 247 + ], + "type": "text", + "content": " derives a family of warped image feature maps:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 372, + 254, + 545, + 268 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 254, + 545, + 268 + ], + "spans": [ + { + "bbox": [ + 372, + 254, + 545, + 268 + ], + "type": "interline_equation", + "content": "F _ {i, j, z} (\\mathbf {p}) = F _ {i} \\left(\\mathbf {H} _ {i j} (z) \\mathbf {p}\\right), \\tag {7}", + "image_path": "08872f2156066cbef187cea5e02165decc1fa2f14064251af0c4a5df8971acfc.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 275, + 545, + 298 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 275, + 545, + 298 + ], + "spans": [ + { + "bbox": [ + 304, + 275, + 545, + 298 + ], + "type": "text", + "content": "which is warped from " + }, + { + "bbox": [ + 304, + 275, + 545, + 298 + ], + "type": "inline_equation", + "content": "F_{i}" + }, + { + "bbox": [ + 304, + 275, + 545, + 298 + ], + "type": "text", + "content": ", seen from the " + }, + { + "bbox": [ + 304, + 275, + 545, + 298 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 275, + 545, + 298 + ], + "type": "text", + "content": "-th view, and at the depth of " + }, + { + "bbox": [ + 304, + 275, + 545, + 298 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 304, + 275, + 545, + 298 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 299, + 545, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 299, + 545, + 382 + ], + "spans": [ + { + "bbox": [ + 304, + 299, + 545, + 382 + ], + "type": "text", + "content": "We then construct a cost volume at the points which we sample using the warped feature maps. Based on the intuition that similar features mean more likelihood of the same area, a large variance of warped features indicates that features belonging to different areas are warped to the current point. Thus, the variance metric can be used as a clue for scene reconstruction as" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 347, + 389, + 545, + 405 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 389, + 545, + 405 + ], + "spans": [ + { + "bbox": [ + 347, + 389, + 545, + 405 + ], + "type": "interline_equation", + "content": "\\mathbf {V} _ {j} (u, v, z) = \\operatorname {V a r} _ {i} \\left(F _ {i, j, z} ([ u, v, 1 ] ^ {\\top})\\right), \\tag {8}", + "image_path": "edcce0f15ff23cd4e7f269118de85f8f69e9da408915c0f42fce0a1d5cf784e1.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 411, + 545, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 411, + 545, + 436 + ], + "spans": [ + { + "bbox": [ + 304, + 411, + 545, + 436 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 411, + 545, + 436 + ], + "type": "inline_equation", + "content": "\\operatorname{Var}_i(\\cdot)" + }, + { + "bbox": [ + 304, + 411, + 545, + 436 + ], + "type": "text", + "content": " denotes the variance across " + }, + { + "bbox": [ + 304, + 411, + 545, + 436 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 411, + 545, + 436 + ], + "type": "text", + "content": " selected neighboring views." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 449, + 545, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 545, + 486 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 545, + 486 + ], + "type": "text", + "content": "Joint optimization. To apply the cost volume as supervision for NeRF training, we consider the weight of the " + }, + { + "bbox": [ + 304, + 449, + 545, + 486 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 304, + 449, + 545, + 486 + ], + "type": "text", + "content": "-th sampled point" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 370, + 494, + 545, + 507 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 494, + 545, + 507 + ], + "spans": [ + { + "bbox": [ + 370, + 494, + 545, + 507 + ], + "type": "interline_equation", + "content": "W _ {k} = T _ {k} \\left(1 - \\exp (- \\sigma_ {k})\\right), \\tag {9}", + "image_path": "6d0d9f2739e605e45ab89953f2357e7706c3d3ff11f1dd710650281508bd4332.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 514, + 545, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 514, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 304, + 514, + 545, + 609 + ], + "type": "text", + "content": "which refers to the contribution of the point at " + }, + { + "bbox": [ + 304, + 514, + 545, + 609 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_t" + }, + { + "bbox": [ + 304, + 514, + 545, + 609 + ], + "type": "text", + "content": " to the result " + }, + { + "bbox": [ + 304, + 514, + 545, + 609 + ], + "type": "inline_equation", + "content": "\\hat{I}_j(u,v)" + }, + { + "bbox": [ + 304, + 514, + 545, + 609 + ], + "type": "text", + "content": ". We build the scene MLP to reconstruct the scene with occlusion, to provide an intermediate scene representation for further occlusion removal. For a high-quality reconstruction, the scene MLP should be discouraged to assign the points that are inconsistent across neighboring views with a large weight. Based on the observations above, we construct the cost volume loss as" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 329, + 616, + 545, + 651 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 616, + 545, + 651 + ], + "spans": [ + { + "bbox": [ + 329, + 616, + 545, + 651 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {c o s t}} = \\sum_ {\\mathbf {p} = (u, v)} \\sum_ {i = 1} ^ {N _ {s}} \\mathbf {V} _ {j} (u, v, t _ {i}) W (u, v, t _ {i}). \\tag {10}", + "image_path": "3cfbac7cf1065a0926fd3562d6750db33e5a05395baa91137231a747f8467242.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 658, + 545, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 658, + 545, + 682 + ], + "spans": [ + { + "bbox": [ + 304, + 658, + 545, + 682 + ], + "type": "text", + "content": "Meanwhile, we also apply the self-consistency loss aiming at scene reconstruction as" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 366, + 689, + 545, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 366, + 689, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 366, + 689, + 545, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {s c}} = \\sum_ {\\mathbf {p}} \\| \\hat {I} _ {i} (\\mathbf {p}) - I _ {i} (\\mathbf {p}) \\| _ {2} ^ {2}. \\tag {11}", + "image_path": "e9da8b974dc760009227e832ee6c3d27a2a441dfa4be22c0b1dfb201b230000a.jpg" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "20725" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "text", + "content": "So far, we have explained the design of scene reconstruction scheme where occlusions exist, including the key component of cost volume, as depicted in the blue box of Fig. 1. Together, they provide a more robust estimation of the scene and the camera poses when occlusions are present." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 147, + 204, + 159 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 147, + 204, + 159 + ], + "spans": [ + { + "bbox": [ + 47, + 147, + 204, + 159 + ], + "type": "text", + "content": "4.2. Selective supervision scheme" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 167, + 288, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 167, + 288, + 335 + ], + "spans": [ + { + "bbox": [ + 46, + 167, + 288, + 335 + ], + "type": "text", + "content": "Existing image-based occlusion removal methods [14, 27] can function properly when a large amount of labeled data is available, by considering the occlusion removal as a mapping from samples with occlusions to their occlusion-free counterparts. Our method is grounded on a totally different underlying logic. By simulating human eyes to observe a specified scenes, our method can aggregate a continuous occlusion-free scene representation from different viewpoints. Then, we can render any occlusion-free viewpoints from this learned scene representation. Such mechanism enables a novel way to remove occlusion in an unsupervised manner. The key problem then becomes how to focus on the desired background only during scene aggregation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 338, + 287, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 338, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 338, + 287, + 529 + ], + "type": "text", + "content": "We achieve this goal based on an intuitive observation: a majority of undesired occlusions always appear in the foreground. However, directly applying depth as supervision is not feasible, since the depth does not necessarily correlate with the presence of occlusion. To this end, we propose to use bidirectional depth inconsistency as a clue. If the expected ray termination depth seen from the camera and that from the other end of the ray show great difference, the ray should pass through some foreground occlusion. Note that \"termination\" in rendering refers to the point that no point behind it contributes to the rendered ray. This assumption follows the physical law, and does not rely on any external prior or excessive training data. Note that this prior is opposite to the shell-shaped geometry assumption by NeRFReN [7], which assumes that the density along the ray peaks out at the surface." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 531, + 287, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 531, + 287, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 531, + 287, + 567 + ], + "type": "text", + "content": "In implementing this, we design a mask MLP " + }, + { + "bbox": [ + 46, + 531, + 287, + 567 + ], + "type": "inline_equation", + "content": "P_{\\Psi}" + }, + { + "bbox": [ + 46, + 531, + 287, + 567 + ], + "type": "text", + "content": " to predict the occluding likelihood " + }, + { + "bbox": [ + 46, + 531, + 287, + 567 + ], + "type": "inline_equation", + "content": "\\tilde{P} (\\mathbf{r})" + }, + { + "bbox": [ + 46, + 531, + 287, + 567 + ], + "type": "text", + "content": " based on the weights along the ray, formulated as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 113, + 581, + 287, + 601 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 581, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 113, + 581, + 287, + 601 + ], + "type": "interline_equation", + "content": "\\tilde {P} (\\mathbf {r}) = P _ {\\Psi} \\left(\\left\\{W _ {k} \\right\\} _ {k = 1} ^ {N _ {s}}\\right). \\tag {12}", + "image_path": "07cda04519607dc8c77bbba95828919ce2587998b704c998e8f636db820ae158.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 616, + 287, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 616, + 287, + 669 + ], + "spans": [ + { + "bbox": [ + 46, + 616, + 287, + 669 + ], + "type": "text", + "content": "Based on the intuition discussed above, the mask MLP is trained based on the bidirectional depth inconsistency. The expected termination depth " + }, + { + "bbox": [ + 46, + 616, + 287, + 669 + ], + "type": "inline_equation", + "content": "\\overrightarrow{D}" + }, + { + "bbox": [ + 46, + 616, + 287, + 669 + ], + "type": "text", + "content": " and reversed termination depth " + }, + { + "bbox": [ + 46, + 616, + 287, + 669 + ], + "type": "inline_equation", + "content": "\\overline{D}" + }, + { + "bbox": [ + 46, + 616, + 287, + 669 + ], + "type": "text", + "content": " is defined by" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 682, + 287, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 682, + 287, + 716 + ], + "spans": [ + { + "bbox": [ + 56, + 682, + 287, + 716 + ], + "type": "interline_equation", + "content": "\\overrightarrow {D} (\\mathbf {r}) = \\sum_ {k = 1} ^ {N _ {s}} W _ {k} t _ {k} = \\sum_ {k = 1} ^ {N _ {s}} T _ {k} \\left(1 - \\exp \\left(- \\sigma_ {k} \\delta_ {k}\\right)\\right) t _ {k}, \\tag {13}", + "image_path": "cfe6fee37c81ec72e20900b03bcfd7fe6a37ed0592d013fe9074e0f96b65a05b.jpg" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 305, + 69, + 383, + 117 + ], + "blocks": [ + { + "bbox": [ + 305, + 69, + 383, + 117 + ], + "lines": [ + { + "bbox": [ + 305, + 69, + 383, + 117 + ], + "spans": [ + { + "bbox": [ + 305, + 69, + 383, + 117 + ], + "type": "image", + "image_path": "eb4e446a46cf45c4e43ad3874b15d415d7df0fcad6e26a8386951670036e3d5f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 335, + 118, + 356, + 129 + ], + "lines": [ + { + "bbox": [ + 335, + 118, + 356, + 129 + ], + "spans": [ + { + "bbox": [ + 335, + 118, + 356, + 129 + ], + "type": "text", + "content": "(a) " + }, + { + "bbox": [ + 335, + 118, + 356, + 129 + ], + "type": "inline_equation", + "content": "\\overrightarrow{D}" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 385, + 70, + 464, + 116 + ], + "blocks": [ + { + "bbox": [ + 385, + 70, + 464, + 116 + ], + "lines": [ + { + "bbox": [ + 385, + 70, + 464, + 116 + ], + "spans": [ + { + "bbox": [ + 385, + 70, + 464, + 116 + ], + "type": "image", + "image_path": "92b80fe52eafdc54c201cd30093437f4cdd9ad0e04ba217fc012b95e90715ff0.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 410, + 118, + 437, + 127 + ], + "lines": [ + { + "bbox": [ + 410, + 118, + 437, + 127 + ], + "spans": [ + { + "bbox": [ + 410, + 118, + 437, + 127 + ], + "type": "text", + "content": "(b) " + }, + { + "bbox": [ + 410, + 118, + 437, + 127 + ], + "type": "inline_equation", + "content": "\\Delta_D" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 465, + 69, + 545, + 117 + ], + "blocks": [ + { + "bbox": [ + 465, + 69, + 545, + 117 + ], + "lines": [ + { + "bbox": [ + 465, + 69, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 465, + 69, + 545, + 117 + ], + "type": "image", + "image_path": "12c6a3ddacd7b4ff1b789953d4b44b49af049e21d6624b176308a75dab88893b.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 493, + 118, + 512, + 125 + ], + "lines": [ + { + "bbox": [ + 493, + 118, + 512, + 125 + ], + "spans": [ + { + "bbox": [ + 493, + 118, + 512, + 125 + ], + "type": "text", + "content": "(c) " + }, + { + "bbox": [ + 493, + 118, + 512, + 125 + ], + "type": "inline_equation", + "content": "P" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 140, + 547, + 239 + ], + "lines": [ + { + "bbox": [ + 304, + 140, + 547, + 239 + ], + "spans": [ + { + "bbox": [ + 304, + 140, + 547, + 239 + ], + "type": "text", + "content": "Figure 2. An illustration of the depth and mask used for selective supervision (without losing generality, we show the example for fence, and this applies to other types of occlusions as well): (a) estimated depth " + }, + { + "bbox": [ + 304, + 140, + 547, + 239 + ], + "type": "inline_equation", + "content": "(\\vec{D})" + }, + { + "bbox": [ + 304, + 140, + 547, + 239 + ], + "type": "text", + "content": " of the scene with occlusion, (b) the bidirectional depth difference " + }, + { + "bbox": [ + 304, + 140, + 547, + 239 + ], + "type": "inline_equation", + "content": "(\\Delta_D)" + }, + { + "bbox": [ + 304, + 140, + 547, + 239 + ], + "type": "text", + "content": ", and (c) the supervision mask " + }, + { + "bbox": [ + 304, + 140, + 547, + 239 + ], + "type": "inline_equation", + "content": "(P)" + }, + { + "bbox": [ + 304, + 140, + 547, + 239 + ], + "type": "text", + "content": ". Note that the white regions in (c) denote the desired background details that need supervision and undesired foreground occlusions are marked in black. Please visit our project page for animated results." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 259, + 324, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 259, + 324, + 270 + ], + "spans": [ + { + "bbox": [ + 306, + 259, + 324, + 270 + ], + "type": "text", + "content": "and" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 276, + 545, + 308 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 276, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 313, + 276, + 545, + 308 + ], + "type": "interline_equation", + "content": "\\overleftarrow {D} (\\mathbf {r}) = \\sum_ {k = 1} ^ {N _ {s}} \\overleftarrow {W} _ {k} t _ {k} = \\sum_ {k = 1} ^ {N _ {s}} \\overleftarrow {T} _ {k} \\left(1 - \\exp \\left(- \\sigma_ {k} \\delta_ {k}\\right)\\right) t _ {k}, \\tag {14}", + "image_path": "b8b36acef1ef51faaca539fe558ad64f40f87a14ec27861042657441db718ec9.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 317, + 545, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 317, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 304, + 317, + 545, + 387 + ], + "type": "text", + "content": "respectively, where " + }, + { + "bbox": [ + 304, + 317, + 545, + 387 + ], + "type": "inline_equation", + "content": "\\overline{T}_k = \\exp (-\\sum_{j > k}\\sigma_j\\delta_j)" + }, + { + "bbox": [ + 304, + 317, + 545, + 387 + ], + "type": "text", + "content": " is the reversed accumulated transmittance. An example of " + }, + { + "bbox": [ + 304, + 317, + 545, + 387 + ], + "type": "inline_equation", + "content": "\\vec{D}" + }, + { + "bbox": [ + 304, + 317, + 545, + 387 + ], + "type": "text", + "content": " can be found in Fig. 2(a). We train the mask MLP based on the prior described before, that " + }, + { + "bbox": [ + 304, + 317, + 545, + 387 + ], + "type": "inline_equation", + "content": "P(\\mathbf{r})" + }, + { + "bbox": [ + 304, + 317, + 545, + 387 + ], + "type": "text", + "content": " correlates with " + }, + { + "bbox": [ + 304, + 317, + 545, + 387 + ], + "type": "inline_equation", + "content": "\\Delta_D(\\mathbf{r}) = \\overleftarrow{D} -\\overrightarrow{D}" + }, + { + "bbox": [ + 304, + 317, + 545, + 387 + ], + "type": "text", + "content": ", by" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 392, + 545, + 417 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 392, + 545, + 417 + ], + "spans": [ + { + "bbox": [ + 313, + 392, + 545, + 417 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {m a s k}} = - \\cos \\left(\\left\\{P (\\mathbf {r}) - \\alpha \\right\\} _ {\\mathbf {r} \\in \\mathcal {R}}, \\left\\{\\Delta_ {D} (\\mathbf {r}) - \\overline {{\\Delta_ {D}}} \\right\\} _ {\\mathbf {r} \\in \\mathcal {R}}\\right), \\tag {15}", + "image_path": "ad5f3714484a5e0295feb75afa2eb9a6894599099fee68cb4685b9fb67134ecd.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 418, + 545, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 418, + 545, + 488 + ], + "spans": [ + { + "bbox": [ + 304, + 418, + 545, + 488 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 418, + 545, + 488 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + }, + { + "bbox": [ + 304, + 418, + 545, + 488 + ], + "type": "text", + "content": " is the patch of rays for training, " + }, + { + "bbox": [ + 304, + 418, + 545, + 488 + ], + "type": "inline_equation", + "content": "\\overline{\\Delta_D}" + }, + { + "bbox": [ + 304, + 418, + 545, + 488 + ], + "type": "text", + "content": " is the mean value of " + }, + { + "bbox": [ + 304, + 418, + 545, + 488 + ], + "type": "inline_equation", + "content": "\\Delta_D" + }, + { + "bbox": [ + 304, + 418, + 545, + 488 + ], + "type": "text", + "content": " across the entire input view, and " + }, + { + "bbox": [ + 304, + 418, + 545, + 488 + ], + "type": "inline_equation", + "content": "\\cos (\\mathbf{a},\\mathbf{b})" + }, + { + "bbox": [ + 304, + 418, + 545, + 488 + ], + "type": "text", + "content": " is the cosine similarity between a and b. We show an example of " + }, + { + "bbox": [ + 304, + 418, + 545, + 488 + ], + "type": "inline_equation", + "content": "\\Delta_D" + }, + { + "bbox": [ + 304, + 418, + 545, + 488 + ], + "type": "text", + "content": " in Fig. 2(b). " + }, + { + "bbox": [ + 304, + 418, + 545, + 488 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 418, + 545, + 488 + ], + "type": "text", + "content": " is a pre-defined threshold for the binary classification of " + }, + { + "bbox": [ + 304, + 418, + 545, + 488 + ], + "type": "inline_equation", + "content": "\\tilde{P}" + }, + { + "bbox": [ + 304, + 418, + 545, + 488 + ], + "type": "text", + "content": ", deriving a binary supervision mask " + }, + { + "bbox": [ + 304, + 418, + 545, + 488 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 418, + 545, + 488 + ], + "type": "text", + "content": " by" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 367, + 488, + 545, + 515 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 488, + 545, + 515 + ], + "spans": [ + { + "bbox": [ + 367, + 488, + 545, + 515 + ], + "type": "interline_equation", + "content": "P (\\mathbf {r}) = \\left\\{ \\begin{array}{l l} 0, & \\tilde {P} (\\mathbf {r}) < \\alpha \\\\ 1, & \\tilde {P} (\\mathbf {r}) \\geq \\alpha \\end{array} \\right.. \\tag {16}", + "image_path": "b70cb04342be501ecb937154e2ae9e8e9dc96982b755c682d4ded6ecf25c68ad.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 518, + 545, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 518, + 545, + 553 + ], + "spans": [ + { + "bbox": [ + 304, + 518, + 545, + 553 + ], + "type": "text", + "content": "As shown in Fig. 2(c), the obtained binary mask " + }, + { + "bbox": [ + 304, + 518, + 545, + 553 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 518, + 545, + 553 + ], + "type": "text", + "content": " can clearly distinguish the desired background to be kept and the undesired occlusion to be removed." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 554, + 545, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 554, + 545, + 614 + ], + "spans": [ + { + "bbox": [ + 304, + 554, + 545, + 614 + ], + "type": "text", + "content": "We further train a background MLP which reconstructs the scene belonging to the background only. Specifically, we only supervise the rendered color of background MLP to be similar to the input image in the background region where " + }, + { + "bbox": [ + 304, + 554, + 545, + 614 + ], + "type": "inline_equation", + "content": "P(\\mathbf{r}) = 1" + }, + { + "bbox": [ + 304, + 554, + 545, + 614 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 343, + 620, + 545, + 679 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 620, + 545, + 679 + ], + "spans": [ + { + "bbox": [ + 343, + 620, + 545, + 679 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\mathrm {b g}} = \\sum_ {\\mathbf {p}} \\left\\| \\left(\\left(\\hat {I} _ {i} ^ {\\prime} (\\mathbf {p}) - I _ {i} (\\mathbf {p})\\right) \\odot P (\\mathbf {r}) \\right\\| _ {2} ^ {2} \\right. \\tag {17} \\\\ + \\sum_ {\\mathbf {p}} \\left\\| (\\vec {D} ^ {\\prime} (\\mathbf {r}) - \\vec {D} (\\mathbf {r})) \\odot P (\\mathbf {r}) \\right\\| _ {1}, \\\\ \\end{array}", + "image_path": "32d9d3f68ad1cf594742b880d9f6d8bd17183cab966f35a96c8bc4d818300a59.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 687, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 687, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 687, + 545, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 687, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\hat{I}_i^{\\prime}(\\mathbf{p})" + }, + { + "bbox": [ + 304, + 687, + 545, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 687, + 545, + 714 + ], + "type": "inline_equation", + "content": "\\vec{D}^{\\prime}(\\mathbf{r})" + }, + { + "bbox": [ + 304, + 687, + 545, + 714 + ], + "type": "text", + "content": " denote the rendered color and expected ray termination depth of the background MLP." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20726" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 120 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 120 + ], + "type": "text", + "content": "With background MLP holding the occlusion-free scene representation and the scene MLP containing all the information including occlusions, we have finally gathered all the pieces of the puzzle." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 126, + 178, + 139 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 126, + 178, + 139 + ], + "spans": [ + { + "bbox": [ + 47, + 126, + 178, + 139 + ], + "type": "text", + "content": "4.3. Implementation details" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 145, + 287, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 145, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 145, + 287, + 312 + ], + "type": "text", + "content": "We implement our method using PyTorch. The scene MLP and the background MLP, each having 256 channels, consist of 6 layers and 8 layers respectively, which predict the colors and densities corresponding to the whole scene and the background separately. The mask MLP is a 2-layer MLP with 64 channels. When constructing cost volume, " + }, + { + "bbox": [ + 46, + 145, + 287, + 312 + ], + "type": "inline_equation", + "content": "M = 4" + }, + { + "bbox": [ + 46, + 145, + 287, + 312 + ], + "type": "text", + "content": " neighboring views are involved to calculate variance in Eq. (8). In the training phase, the scene MLP is jointly optimized with camera parameters of each scene, after being trained coarsely to fit the scene from COLMAP [23] camera pose or identity matrix initialization. All of the trainable parameters are optimized per scene. We train scene MLP for the first " + }, + { + "bbox": [ + 46, + 145, + 287, + 312 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 46, + 145, + 287, + 312 + ], + "type": "text", + "content": " of iterations, and then train the system end-to-end." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 312, + 288, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 312, + 288, + 419 + ], + "spans": [ + { + "bbox": [ + 46, + 312, + 288, + 419 + ], + "type": "text", + "content": "On each iteration, we evenly sample 128 points along each ray. A batch contains random 1024 pixels in one of the images. We use the Adam optimizer with defaults values " + }, + { + "bbox": [ + 46, + 312, + 288, + 419 + ], + "type": "inline_equation", + "content": "\\beta_{1} = 0.999" + }, + { + "bbox": [ + 46, + 312, + 288, + 419 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 312, + 288, + 419 + ], + "type": "inline_equation", + "content": "\\beta_{2} = 0.9" + }, + { + "bbox": [ + 46, + 312, + 288, + 419 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 312, + 288, + 419 + ], + "type": "inline_equation", + "content": "\\epsilon = 10^{-8}" + }, + { + "bbox": [ + 46, + 312, + 288, + 419 + ], + "type": "text", + "content": ", and a learning rate " + }, + { + "bbox": [ + 46, + 312, + 288, + 419 + ], + "type": "inline_equation", + "content": "10^{-3}" + }, + { + "bbox": [ + 46, + 312, + 288, + 419 + ], + "type": "text", + "content": " that decays following the cosine scheduler [15] during the optimization. If initialized with identity matrix, the model is trained for about 150K iterations. If a coarse COLMAP [23] estimation is available, our model converges after about 50K iterations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 430, + 128, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 430, + 128, + 443 + ], + "spans": [ + { + "bbox": [ + 47, + 430, + 128, + 443 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 450, + 287, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 450, + 287, + 594 + ], + "spans": [ + { + "bbox": [ + 46, + 450, + 287, + 594 + ], + "type": "text", + "content": "Dataset. To validate the effectiveness of our method, we collect an evaluation dataset containing 10 different scenes, covering various types of occlusions. The dataset not only contains multiple sparse viewpoints of occlusion scenes collected by ourselves, but also samples selected from existing datasets [29,31], to increase the diversity from different capturing conditions. Specifically, FENCE1, SCRIBBLE2 and RAINDROP are adopted from [29], and WIRE2 is adopted from [31]. The rest 6 of the scenes are captured by ourselves using a Sony " + }, + { + "bbox": [ + 46, + 450, + 287, + 594 + ], + "type": "inline_equation", + "content": "\\alpha 7" + }, + { + "bbox": [ + 46, + 450, + 287, + 594 + ], + "type": "text", + "content": " III camera or an iPhone 12. For easy reference, we name each set of data using the occlusion type, as shown in Fig. 3." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": "For our newly captured samples, we take 20 to 60 images with different poses for each scene, where " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "85\\%" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": " of the images are used for training our NeRF-based method, and the rest " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": " for testing. This is a commonly adopted setting by many NeRF-based methods [7, 8, 18]. To facilitate the quantitative evaluation, one of newly captured scene is with ground truth, namely SCRIBBLE1. For this scene, we place a piece of glass with scribbles between the desired background scene and the camera. Then, a multi-view sparse image set is acquired by moving cameras in front of the" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 310, + 69, + 358, + 98 + ], + "blocks": [ + { + "bbox": [ + 310, + 69, + 358, + 98 + ], + "lines": [ + { + "bbox": [ + 310, + 69, + 358, + 98 + ], + "spans": [ + { + "bbox": [ + 310, + 69, + 358, + 98 + ], + "type": "image", + "image_path": "0e5ca29bd3105841712d3adbb3e1ee3ba95cff7e81b5d343337c1189007680a0.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 99, + 351, + 108 + ], + "lines": [ + { + "bbox": [ + 313, + 99, + 351, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 99, + 351, + 108 + ], + "type": "text", + "content": "(a) FENCE1" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 360, + 70, + 407, + 98 + ], + "blocks": [ + { + "bbox": [ + 360, + 70, + 407, + 98 + ], + "lines": [ + { + "bbox": [ + 360, + 70, + 407, + 98 + ], + "spans": [ + { + "bbox": [ + 360, + 70, + 407, + 98 + ], + "type": "image", + "image_path": "e02365c55347ee7aa9fde4c72d4b709ec0293cce250e439385f8c46bf02bc6b7.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 362, + 99, + 401, + 108 + ], + "lines": [ + { + "bbox": [ + 362, + 99, + 401, + 108 + ], + "spans": [ + { + "bbox": [ + 362, + 99, + 401, + 108 + ], + "type": "text", + "content": "(b) FENCE2" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 408, + 70, + 455, + 98 + ], + "blocks": [ + { + "bbox": [ + 408, + 70, + 455, + 98 + ], + "lines": [ + { + "bbox": [ + 408, + 70, + 455, + 98 + ], + "spans": [ + { + "bbox": [ + 408, + 70, + 455, + 98 + ], + "type": "image", + "image_path": "c64a71041da4182430496bf7ad9638810e88de1c92464a50711c22f010c29a89.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 411, + 99, + 449, + 108 + ], + "lines": [ + { + "bbox": [ + 411, + 99, + 449, + 108 + ], + "spans": [ + { + "bbox": [ + 411, + 99, + 449, + 108 + ], + "type": "text", + "content": "(c) FENCE3" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 454, + 99, + 553, + 108 + ], + "lines": [ + { + "bbox": [ + 454, + 99, + 553, + 108 + ], + "spans": [ + { + "bbox": [ + 454, + 99, + 553, + 108 + ], + "type": "text", + "content": "(d) SCRIBBLE1(e) SCRIBBLE2" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 457, + 70, + 504, + 98 + ], + "blocks": [ + { + "bbox": [ + 457, + 70, + 504, + 98 + ], + "lines": [ + { + "bbox": [ + 457, + 70, + 504, + 98 + ], + "spans": [ + { + "bbox": [ + 457, + 70, + 504, + 98 + ], + "type": "image", + "image_path": "cf4538b29a225a63eb2c3871ea55d442076a44cf815f4e39dec1b081a9af6161.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 310, + 109, + 358, + 136 + ], + "blocks": [ + { + "bbox": [ + 310, + 109, + 358, + 136 + ], + "lines": [ + { + "bbox": [ + 310, + 109, + 358, + 136 + ], + "spans": [ + { + "bbox": [ + 310, + 109, + 358, + 136 + ], + "type": "image", + "image_path": "f1c24814e9e2aed511e37de84e89b5403e6c14817360a740a704bcc4f18efdba.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 308, + 137, + 406, + 147 + ], + "lines": [ + { + "bbox": [ + 308, + 137, + 406, + 147 + ], + "spans": [ + { + "bbox": [ + 308, + 137, + 406, + 147 + ], + "type": "text", + "content": "(f) SCRIBBLE3 (g) RAINDROP" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 360, + 110, + 407, + 136 + ], + "blocks": [ + { + "bbox": [ + 360, + 110, + 407, + 136 + ], + "lines": [ + { + "bbox": [ + 360, + 110, + 407, + 136 + ], + "spans": [ + { + "bbox": [ + 360, + 110, + 407, + 136 + ], + "type": "image", + "image_path": "3b77cbb4c76bfa9bea5693ea7c6d14a848223677acd413721ef64a759818c063.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 408, + 110, + 455, + 136 + ], + "blocks": [ + { + "bbox": [ + 408, + 110, + 455, + 136 + ], + "lines": [ + { + "bbox": [ + 408, + 110, + 455, + 136 + ], + "spans": [ + { + "bbox": [ + 408, + 110, + 455, + 136 + ], + "type": "image", + "image_path": "6ad0eb2b3708ac578af25ec332fbd713543174fa5daf73d6c08a5756ceb28d91.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 411, + 137, + 450, + 146 + ], + "lines": [ + { + "bbox": [ + 411, + 137, + 450, + 146 + ], + "spans": [ + { + "bbox": [ + 411, + 137, + 450, + 146 + ], + "type": "text", + "content": "(h) STATUE" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 153, + 546, + 186 + ], + "lines": [ + { + "bbox": [ + 305, + 153, + 546, + 186 + ], + "spans": [ + { + "bbox": [ + 305, + 153, + 546, + 186 + ], + "type": "text", + "content": "Figure 3. From (a) to (j), we provide one sample image for each scene in our evaluation dataset. The names of the scenes are listed below the image indicating occlusion type." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 457, + 109, + 504, + 136 + ], + "blocks": [ + { + "bbox": [ + 457, + 109, + 504, + 136 + ], + "lines": [ + { + "bbox": [ + 457, + 109, + 504, + 136 + ], + "spans": [ + { + "bbox": [ + 457, + 109, + 504, + 136 + ], + "type": "image", + "image_path": "9ae9ce1288252b7c3c4299c12b574b290c70653e9a688a5c635d8a9669d481c0.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 463, + 137, + 495, + 146 + ], + "lines": [ + { + "bbox": [ + 463, + 137, + 495, + 146 + ], + "spans": [ + { + "bbox": [ + 463, + 137, + 495, + 146 + ], + "type": "text", + "content": "(i) WIRE1" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 506, + 110, + 553, + 136 + ], + "blocks": [ + { + "bbox": [ + 506, + 110, + 553, + 136 + ], + "lines": [ + { + "bbox": [ + 506, + 110, + 553, + 136 + ], + "spans": [ + { + "bbox": [ + 506, + 110, + 553, + 136 + ], + "type": "image", + "image_path": "a5bc464626ba89cfe590c2a67dfe6094684991291969b717cce5138f9104a535.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 511, + 137, + 545, + 147 + ], + "lines": [ + { + "bbox": [ + 511, + 137, + 545, + 147 + ], + "spans": [ + { + "bbox": [ + 511, + 137, + 545, + 147 + ], + "type": "text", + "content": "(j) WIRE2" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "bbox": [ + 305, + 206, + 546, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 206, + 546, + 243 + ], + "spans": [ + { + "bbox": [ + 305, + 206, + 546, + 243 + ], + "type": "text", + "content": "glass. To obtain the ground truth for validation, we remove the glass while taking the first and last view of the scene, so that these two views can come with a ground truth." + } + ] + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 312, + 251, + 391, + 297 + ], + "blocks": [ + { + "bbox": [ + 312, + 251, + 391, + 297 + ], + "lines": [ + { + "bbox": [ + 312, + 251, + 391, + 297 + ], + "spans": [ + { + "bbox": [ + 312, + 251, + 391, + 297 + ], + "type": "image", + "image_path": "9977cba7993c2e8977713ddf7ba885844ffa332a310d1e00eade623c2e95e617.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 392, + 252, + 439, + 297 + ], + "blocks": [ + { + "bbox": [ + 392, + 252, + 439, + 297 + ], + "lines": [ + { + "bbox": [ + 392, + 252, + 439, + 297 + ], + "spans": [ + { + "bbox": [ + 392, + 252, + 439, + 297 + ], + "type": "image", + "image_path": "63c631172c06aa5518842d4a94993172fae483e8130829ce6ca5c120248b41b2.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 440, + 252, + 549, + 297 + ], + "blocks": [ + { + "bbox": [ + 440, + 252, + 549, + 297 + ], + "lines": [ + { + "bbox": [ + 440, + 252, + 549, + 297 + ], + "spans": [ + { + "bbox": [ + 440, + 252, + 549, + 297 + ], + "type": "image", + "image_path": "5d0a1c60494e6e082406595c259a21fbee1cafc2a85b3fb4f030974222bd988b.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 312, + 310, + 391, + 355 + ], + "blocks": [ + { + "bbox": [ + 312, + 310, + 391, + 355 + ], + "lines": [ + { + "bbox": [ + 312, + 310, + 391, + 355 + ], + "spans": [ + { + "bbox": [ + 312, + 310, + 391, + 355 + ], + "type": "image", + "image_path": "c8ba6650eee6059e3e2faa5540dafae4517a1dc3f5adca1c70fdb2c90021de4a.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 381, + 357, + 476, + 366 + ], + "lines": [ + { + "bbox": [ + 381, + 357, + 476, + 366 + ], + "spans": [ + { + "bbox": [ + 381, + 357, + 476, + 366 + ], + "type": "text", + "content": "(b) PWC-Net+NeRF [14, 18]" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 392, + 310, + 439, + 355 + ], + "blocks": [ + { + "bbox": [ + 393, + 300, + 464, + 309 + ], + "lines": [ + { + "bbox": [ + 393, + 300, + 464, + 309 + ], + "spans": [ + { + "bbox": [ + 393, + 300, + 464, + 309 + ], + "type": "text", + "content": "(a) Vanilla NeRF [18]" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 392, + 310, + 439, + 355 + ], + "lines": [ + { + "bbox": [ + 392, + 310, + 439, + 355 + ], + "spans": [ + { + "bbox": [ + 392, + 310, + 439, + 355 + ], + "type": "image", + "image_path": "af24594123f5cc33ade3c64b25eed1b0cb1da24d45284641be34d5530108166a.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 440, + 310, + 549, + 355 + ], + "blocks": [ + { + "bbox": [ + 440, + 310, + 549, + 355 + ], + "lines": [ + { + "bbox": [ + 440, + 310, + 549, + 355 + ], + "spans": [ + { + "bbox": [ + 440, + 310, + 549, + 355 + ], + "type": "image", + "image_path": "92823486a0237d025b5e1787158b617e0ddcb52154196aa3b3038b766ba51e6d.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 312, + 368, + 391, + 412 + ], + "blocks": [ + { + "bbox": [ + 312, + 368, + 391, + 412 + ], + "lines": [ + { + "bbox": [ + 312, + 368, + 391, + 412 + ], + "spans": [ + { + "bbox": [ + 312, + 368, + 391, + 412 + ], + "type": "image", + "image_path": "2e37d99c93a2ef58e37cd5f64baf04b40fba2498bd2046fda159be2fcd3833f8.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 392, + 368, + 439, + 412 + ], + "blocks": [ + { + "bbox": [ + 392, + 368, + 439, + 412 + ], + "lines": [ + { + "bbox": [ + 392, + 368, + 439, + 412 + ], + "spans": [ + { + "bbox": [ + 392, + 368, + 439, + 412 + ], + "type": "image", + "image_path": "88db9acb742ca029104ff629418cecd601dbe789367c1a37f2941f53779f7daf.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 440, + 368, + 549, + 412 + ], + "blocks": [ + { + "bbox": [ + 440, + 368, + 549, + 412 + ], + "lines": [ + { + "bbox": [ + 440, + 368, + 549, + 412 + ], + "spans": [ + { + "bbox": [ + 440, + 368, + 549, + 412 + ], + "type": "image", + "image_path": "92bd0903dcd9d070ce2168b1ca3e7e06853233f030f550a99e043c80cd291da3.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 312, + 425, + 391, + 470 + ], + "blocks": [ + { + "bbox": [ + 312, + 425, + 391, + 470 + ], + "lines": [ + { + "bbox": [ + 312, + 425, + 391, + 470 + ], + "spans": [ + { + "bbox": [ + 312, + 425, + 391, + 470 + ], + "type": "image", + "image_path": "6daed4086598c7e2a322fecd5428c7fc7ed1b117e8599e596ead58853120b9d3.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 392, + 425, + 439, + 470 + ], + "blocks": [ + { + "bbox": [ + 402, + 415, + 456, + 423 + ], + "lines": [ + { + "bbox": [ + 402, + 415, + 456, + 423 + ], + "spans": [ + { + "bbox": [ + 402, + 415, + 456, + 423 + ], + "type": "text", + "content": "(c) Ha-NeRF [3]" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 392, + 425, + 439, + 470 + ], + "lines": [ + { + "bbox": [ + 392, + 425, + 439, + 470 + ], + "spans": [ + { + "bbox": [ + 392, + 425, + 439, + 470 + ], + "type": "image", + "image_path": "8d82aa719327b598dca1ea1f571e3e2a01b1e789d7d65c754439f4463c249963.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 400, + 472, + 457, + 481 + ], + "lines": [ + { + "bbox": [ + 400, + 472, + 457, + 481 + ], + "spans": [ + { + "bbox": [ + 400, + 472, + 457, + 481 + ], + "type": "text", + "content": "(d) NeRF-W [17]" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_caption" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 440, + 425, + 549, + 470 + ], + "blocks": [ + { + "bbox": [ + 440, + 425, + 549, + 470 + ], + "lines": [ + { + "bbox": [ + 440, + 425, + 549, + 470 + ], + "spans": [ + { + "bbox": [ + 440, + 425, + 549, + 470 + ], + "type": "image", + "image_path": "20e7c9a3736e00a234af0ff74905fa352cbdfababa451d1857dbe5492ae1ad54.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 312, + 482, + 391, + 527 + ], + "blocks": [ + { + "bbox": [ + 312, + 482, + 391, + 527 + ], + "lines": [ + { + "bbox": [ + 312, + 482, + 391, + 527 + ], + "spans": [ + { + "bbox": [ + 312, + 482, + 391, + 527 + ], + "type": "image", + "image_path": "e54464a78a2ef751d1954c47cfd1c8456072d5b55a3c3871216a34f5d81ac8a9.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 543, + 546, + 631 + ], + "lines": [ + { + "bbox": [ + 304, + 543, + 546, + 631 + ], + "spans": [ + { + "bbox": [ + 304, + 543, + 546, + 631 + ], + "type": "text", + "content": "Figure 4. Visual quality comparisons for occlusion removal. From left to right, we show animated novel view synthesis on FENCE1, STATUE, and WIRE1. From (a) to (e), we show results obtained by (a) vanilla NeRF [18], (b) PWC-Net + NeRF [14, 18], (c) Ha-NeRF [3], (d) NeRF-W [17], and (e) our method. Please visit our project page for animated results. Some results show stronger variation over the animation due to failure in extracting consistent background." + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_caption" + } + ], + "index": 42 + }, + { + "type": "image", + "bbox": [ + 392, + 483, + 439, + 527 + ], + "blocks": [ + { + "bbox": [ + 392, + 483, + 439, + 527 + ], + "lines": [ + { + "bbox": [ + 392, + 483, + 439, + 527 + ], + "spans": [ + { + "bbox": [ + 392, + 483, + 439, + 527 + ], + "type": "image", + "image_path": "c9da8cded0f8478f5cf866e92d83f853313fc7622468c87462f59d95a8af4cdc.jpg" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 403, + 529, + 454, + 538 + ], + "lines": [ + { + "bbox": [ + 403, + 529, + 454, + 538 + ], + "spans": [ + { + "bbox": [ + 403, + 529, + 454, + 538 + ], + "type": "text", + "content": "(e) Our method" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_caption" + } + ], + "index": 43 + }, + { + "type": "image", + "bbox": [ + 440, + 483, + 549, + 527 + ], + "blocks": [ + { + "bbox": [ + 440, + 483, + 549, + 527 + ], + "lines": [ + { + "bbox": [ + 440, + 483, + 549, + 527 + ], + "spans": [ + { + "bbox": [ + 440, + 483, + 549, + 527 + ], + "type": "image", + "image_path": "732cd18bbef128dd1f5623c456a7de2af32866f6ad2c456bf0cf8776884a62dd.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + } + ], + "index": 44 + }, + { + "bbox": [ + 304, + 654, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 546, + 713 + ], + "type": "text", + "content": "Baselines. To our knowledge, there is no similar work that construct an occlusion-free scene based on NeRF representation. We design three baselines for comparison by considering both the contributions from occlusion removal and NeRF to verify our performance advantage. 1) A state-of" + } + ] + } + ], + "index": 47 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20727" + } + ] + } + ], + "index": 48 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 97, + 288, + 162 + ], + "blocks": [ + { + "bbox": [ + 47, + 71, + 288, + 92 + ], + "lines": [ + { + "bbox": [ + 47, + 71, + 288, + 92 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 288, + 92 + ], + "type": "text", + "content": "Table 1. Quantitative comparison results. " + }, + { + "bbox": [ + 47, + 71, + 288, + 92 + ], + "type": "inline_equation", + "content": "\\uparrow (\\downarrow)" + }, + { + "bbox": [ + 47, + 71, + 288, + 92 + ], + "type": "text", + "content": " indicates larger (smaller) values are better, and bold font indicates the best results." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 97, + 288, + 162 + ], + "lines": [ + { + "bbox": [ + 48, + 97, + 288, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 97, + 288, + 162 + ], + "type": "table", + "html": "
MethodPSNR↑SSIM↑LPIPS↓
PWC-Net+NeRF [14,18]18.250.790.45
NeRF-W [17]11.060.450.61
Ha-NeRF [3]15.410.700.41
Proposed19.390.830.37
", + "image_path": "14f17c09f2df42f6f9916e6e9d9ec3b7ff253d79be50b29b5cd6c75c6ad6971b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 49, + 171, + 115, + 217 + ], + "blocks": [ + { + "bbox": [ + 49, + 171, + 115, + 217 + ], + "lines": [ + { + "bbox": [ + 49, + 171, + 115, + 217 + ], + "spans": [ + { + "bbox": [ + 49, + 171, + 115, + 217 + ], + "type": "image", + "image_path": "dda4b2d506b49a7a9d5204b5a6a4484b967e6a89ac4d7d7316fc9814eb44b471.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 116, + 171, + 206, + 217 + ], + "blocks": [ + { + "bbox": [ + 116, + 171, + 206, + 217 + ], + "lines": [ + { + "bbox": [ + 116, + 171, + 206, + 217 + ], + "spans": [ + { + "bbox": [ + 116, + 171, + 206, + 217 + ], + "type": "image", + "image_path": "d9b881336a0ad94df19a9a180ddd72abfda0e3ce6d65325a603801ef59b3bcdf.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 206, + 171, + 286, + 217 + ], + "blocks": [ + { + "bbox": [ + 206, + 171, + 286, + 217 + ], + "lines": [ + { + "bbox": [ + 206, + 171, + 286, + 217 + ], + "spans": [ + { + "bbox": [ + 206, + 171, + 286, + 217 + ], + "type": "image", + "image_path": "d0efb166d2b78e6326e50be32c8e55cb89d94149659bf44b36f4a5712e42012d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 49, + 229, + 115, + 274 + ], + "blocks": [ + { + "bbox": [ + 81, + 220, + 90, + 228 + ], + "lines": [ + { + "bbox": [ + 81, + 220, + 90, + 228 + ], + "spans": [ + { + "bbox": [ + 81, + 220, + 90, + 228 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 49, + 229, + 115, + 274 + ], + "lines": [ + { + "bbox": [ + 49, + 229, + 115, + 274 + ], + "spans": [ + { + "bbox": [ + 49, + 229, + 115, + 274 + ], + "type": "image", + "image_path": "cf7fccab72517023724dcd45ddf735aa9232a87481d760efb4746185e41685e0.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 80, + 277, + 91, + 285 + ], + "lines": [ + { + "bbox": [ + 80, + 277, + 91, + 285 + ], + "spans": [ + { + "bbox": [ + 80, + 277, + 91, + 285 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 292, + 287, + 357 + ], + "lines": [ + { + "bbox": [ + 46, + 292, + 287, + 357 + ], + "spans": [ + { + "bbox": [ + 46, + 292, + 287, + 357 + ], + "type": "text", + "content": "Figure 5. Visual quality comparisons for occlusion removal on SCRIBBLE1 scene. (a) is one image in the validation set with occlusion and (f) is ground truth of this view) with results from: (b) results obtained by Ha-NeRF [3], (c) results obtained by NeRF-W [17], (d) results obtained by occlusion removal [14] + NeRF [18], and (e) our method." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 116, + 229, + 206, + 274 + ], + "blocks": [ + { + "bbox": [ + 160, + 220, + 169, + 228 + ], + "lines": [ + { + "bbox": [ + 160, + 220, + 169, + 228 + ], + "spans": [ + { + "bbox": [ + 160, + 220, + 169, + 228 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 116, + 229, + 206, + 274 + ], + "lines": [ + { + "bbox": [ + 116, + 229, + 206, + 274 + ], + "spans": [ + { + "bbox": [ + 116, + 229, + 206, + 274 + ], + "type": "image", + "image_path": "55634547e8c938be4d0af2152bf4e49abc91c15ecbac6b4a3ee7e2aff6fbcf5f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 160, + 277, + 170, + 285 + ], + "lines": [ + { + "bbox": [ + 160, + 277, + 170, + 285 + ], + "spans": [ + { + "bbox": [ + 160, + 277, + 170, + 285 + ], + "type": "text", + "content": "(e)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 206, + 229, + 286, + 274 + ], + "blocks": [ + { + "bbox": [ + 239, + 220, + 248, + 228 + ], + "lines": [ + { + "bbox": [ + 239, + 220, + 248, + 228 + ], + "spans": [ + { + "bbox": [ + 239, + 220, + 248, + 228 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 206, + 229, + 286, + 274 + ], + "lines": [ + { + "bbox": [ + 206, + 229, + 286, + 274 + ], + "spans": [ + { + "bbox": [ + 206, + 229, + 286, + 274 + ], + "type": "image", + "image_path": "aae097cf4956e10dd9833088135f0d97fb6542d569e218daf35766aeec7f88df.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 239, + 277, + 248, + 285 + ], + "lines": [ + { + "bbox": [ + 239, + 277, + 248, + 285 + ], + "spans": [ + { + "bbox": [ + 239, + 277, + 248, + 285 + ], + "type": "text", + "content": "(f)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 374, + 287, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 374, + 287, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 374, + 287, + 529 + ], + "type": "text", + "content": "the-art occlusion removal method for image sequences or videos (referred to as PWC-Net [14] based on the main pillar of the work) + NeRF [18]: In this setting, PWC-Net [14] serves as a pre-processing tool to eliminate occlusion before training the NeRF [18]. 2) NeRF-W [17]: This is a method designed for separating a complete scene into two components. 3) Ha-NeRF [3]: This is another method that decomposes the scenes into distinct parts. We compare those settings with our method on our evaluation dataset, as listed in Fig. 3. For fairness of comparison, we evenly sample 128 points along the ray without applying importance sampling. We also keep the trainable parameters in each baseline approximately the same with the proposed method." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 46, + 544, + 287, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 544, + 287, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 544, + 287, + 604 + ], + "type": "text", + "content": "Evaluation methodology. We conduct both qualitative and quantitative evaluations on our dataset. For quantitative evaluation, we compare reconstruction quality of the estimated occlusion-free images for viewpoints with corresponding ground truth by PSNR, SSIM, and LPIPS [32]." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 47, + 611, + 173, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 611, + 173, + 624 + ], + "spans": [ + { + "bbox": [ + 47, + 611, + 173, + 624 + ], + "type": "text", + "content": "5.1. Qualitative evaluation" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 287, + 713 + ], + "type": "text", + "content": "The qualitative results of the rendered novel occlusion-free views are presented in Figs. 4 and 5. In Fig. 4, we use the scenes of FENCE1, STATUE and WIRE1 in our dataset to show that our method can reliably reconstruct novel background views, and it works no matter whether the occlusions cover a large area or have an irregular shape. In Fig. 5, we take one frame in SCRIBBLE1 with ground truth to show" + } + ] + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 310, + 71, + 386, + 205 + ], + "blocks": [ + { + "bbox": [ + 310, + 71, + 386, + 205 + ], + "lines": [ + { + "bbox": [ + 310, + 71, + 386, + 205 + ], + "spans": [ + { + "bbox": [ + 310, + 71, + 386, + 205 + ], + "type": "image", + "image_path": "a72071363551b10fae0b0e46b7797e3f73808acb0a43cea0c9afb5cf3af8cf03.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 342, + 209, + 351, + 217 + ], + "lines": [ + { + "bbox": [ + 342, + 209, + 351, + 217 + ], + "spans": [ + { + "bbox": [ + 342, + 209, + 351, + 217 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 222, + 545, + 277 + ], + "lines": [ + { + "bbox": [ + 304, + 222, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 304, + 222, + 545, + 277 + ], + "type": "text", + "content": "Figure 6. Visual quality comparisons for occlusion removal on RAINDROP scene, with results from (a) an image-based raindrop removal method [22], (b) an image-based occlusion removal method [14], and (c) our method. Our method outperforms the baselines in the highlighted red boxes." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 386, + 71, + 464, + 205 + ], + "blocks": [ + { + "bbox": [ + 386, + 71, + 464, + 205 + ], + "lines": [ + { + "bbox": [ + 386, + 71, + 464, + 205 + ], + "spans": [ + { + "bbox": [ + 386, + 71, + 464, + 205 + ], + "type": "image", + "image_path": "969a8f43eee4fc0ac4fb4c07c19c257abeee61c15007018b759ddd91e14842ae.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 418, + 209, + 428, + 217 + ], + "lines": [ + { + "bbox": [ + 418, + 209, + 428, + 217 + ], + "spans": [ + { + "bbox": [ + 418, + 209, + 428, + 217 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 465, + 71, + 541, + 205 + ], + "blocks": [ + { + "bbox": [ + 465, + 71, + 541, + 205 + ], + "lines": [ + { + "bbox": [ + 465, + 71, + 541, + 205 + ], + "spans": [ + { + "bbox": [ + 465, + 71, + 541, + 205 + ], + "type": "image", + "image_path": "f8c4977a011adc2138b3b87bc242faadf36c5521f0dcfc707e82aade4b072453.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 496, + 209, + 506, + 217 + ], + "lines": [ + { + "bbox": [ + 496, + 209, + 506, + 217 + ], + "spans": [ + { + "bbox": [ + 496, + 209, + 506, + 217 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 290, + 409, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 290, + 409, + 301 + ], + "spans": [ + { + "bbox": [ + 306, + 290, + 409, + 301 + ], + "type": "text", + "content": "the fidelity of our results." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 302, + 545, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 302, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 304, + 302, + 545, + 469 + ], + "type": "text", + "content": "PWC-Net + NeRF [14, 18] only achieves comparable results in FENCE1 where the occlusions are regularly shaped, but shows poor artifacts on unseen types of occlusions, such as the irregularly-shaped STATUE and WIRE1, as shown in Fig. 4(b). This shows their reliance on training data. Besides, the occlusion removal method can lead to an oversmoothing issue which undermines the reconstruction accuracy. Meanwhile, the results from Ha-NeRF [3] and NeRF-W [17] in Figs. 4(c) and 4(d) either shows degenerated details or still preserves occlusions, which verifies our claim that they cannot handle the consistent occlusions observed in a continuous image sequence. The results from our method in Fig. 4(e) clearly remove these occlusions with different shapes and distributions." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 471, + 545, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 471, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 304, + 471, + 545, + 601 + ], + "type": "text", + "content": "Our method relies on an accurate separation between foreground and background, but the assumption may be violated if obstructions are semi-transparent. We conduct experiments on the RAINDROP scene in our validation dataset to evaluate the robustness of our method under such scenario. With the results displayed in Fig. 6, our method can also effectively remove the raindrop in this example. It achieves competitive or even better results in some regions when compared with a method specifically designed for raindrop removal [22] as well as PWC-Net [14] designed to remove occlusions including raindrop." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 306, + 611, + 438, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 611, + 438, + 624 + ], + "spans": [ + { + "bbox": [ + 306, + 611, + 438, + 624 + ], + "type": "text", + "content": "5.2. Quantitative evaluation" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": "The quantitative results of the occlusion-free novel view in Tab. 1 validate the observation in Figs. 4 and 5. Higher PSNR values show that our method can render occlusion-free novel views and recover the color information with higher accuracy. Higher SSIM values indicate that our method can preserve the structural information with high-frequency details. Lower LPIPS values show that recovered" + } + ] + } + ], + "index": 30 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20728" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 104, + 102, + 491, + 192 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 545, + 94 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 545, + 94 + ], + "type": "text", + "content": "Table 2. Ablation study results aggregated by the types of occlusions. We compare the rendered results of the scene MLP with NeRF [18] using COLMAP [23] as camera parameter estimation and NeRF-- [28] initialized with COLMAP camera parameters." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 104, + 102, + 491, + 192 + ], + "lines": [ + { + "bbox": [ + 104, + 102, + 491, + 192 + ], + "spans": [ + { + "bbox": [ + 104, + 102, + 491, + 192 + ], + "type": "table", + "html": "
MetricsPSNR ↑ / SSIM ↑ / LPIPS↓
MethodCOLMAP [23]NeRF-- [28]Scene MLP
DatasetFENCE{1,2,3}24.01 / 0.73 / 0.3818.67 / 0.54 / 0.5125.46 / 0.72 / 0.35
SCRIBBLE{1,2,3}22.70 / 0.78 / 0.3117.98 / 0.52 / 0.4623.85 / 0.84 / 0.34
WIRES{1,2}26.21 / 0.93 / 0.2022.28 / 0.88 / 0.2534.44 / 0.95 / 0.23
RAINDROP27.69 / 0.87 / 0.2627.78 / 0.83 / 0.2528.69 / 0.86 / 0.25
STATUE27.53 / 0.84 / 0.3318.49 / 0.56 / 0.4927.08 / 0.80 / 0.37
", + "image_path": "125cc2b668633ea76581f9e024f833d6ae82776d17bf55f25f1dcc9adecff15b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 212, + 286, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 212, + 286, + 224 + ], + "spans": [ + { + "bbox": [ + 47, + 212, + 286, + 224 + ], + "type": "text", + "content": "images by our method better aligns with human perception." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 232, + 141, + 245 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 232, + 141, + 245 + ], + "spans": [ + { + "bbox": [ + 47, + 232, + 141, + 245 + ], + "type": "text", + "content": "5.3. Ablation Study" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 251, + 286, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 251, + 286, + 323 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 286, + 323 + ], + "type": "text", + "content": "Our method consists of two major components: joint optimization for pose refinement and a selective supervision strategy to eliminate occlusions. Since the selective supervision scheme largely bases on the intermediate information of scene reconstruction MLP, a high-quality reconstruction of the scene with occlusion is crucial to our performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 323, + 286, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 323, + 286, + 382 + ], + "spans": [ + { + "bbox": [ + 46, + 323, + 286, + 382 + ], + "type": "text", + "content": "In the ablation study, we first replace pose refinement with COLMAP [23]. From the results shown in Fig. 7(b), the occlusions largely remain, because the fixed erroneous camera parameters makes distinguishing scene components from scene MLP infeasible." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 384, + 286, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 384, + 286, + 491 + ], + "spans": [ + { + "bbox": [ + 46, + 384, + 286, + 491 + ], + "type": "text", + "content": "We also explore the results without the selective supervision scheme, while maintaining pose refinement, namely only output of the scene MLP. In this setting, our method cannot remove occlusions, but it can reconstruct the scene containing occlusions with a high fidelity, as shown in Fig. 7(c). It can also reconstruct the scene with higher fidelity than NeRF based on COLMAP [23], as shown quantitatively in Tab. 2. Our complete model shows the best results for majority of scenes in our dataset." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 52, + 502, + 128, + 546 + ], + "blocks": [ + { + "bbox": [ + 52, + 502, + 128, + 546 + ], + "lines": [ + { + "bbox": [ + 52, + 502, + 128, + 546 + ], + "spans": [ + { + "bbox": [ + 52, + 502, + 128, + 546 + ], + "type": "image", + "image_path": "bf5ee8ca67981648d22d756febd1b56846ecbb79837458b809d70a1cbfc438ef.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 84, + 548, + 94, + 558 + ], + "lines": [ + { + "bbox": [ + 84, + 548, + 94, + 558 + ], + "spans": [ + { + "bbox": [ + 84, + 548, + 94, + 558 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 563, + 286, + 607 + ], + "lines": [ + { + "bbox": [ + 46, + 563, + 286, + 607 + ], + "spans": [ + { + "bbox": [ + 46, + 563, + 286, + 607 + ], + "type": "text", + "content": "Figure 7. Visual comparisons for the ablation study on the FENCE1 scene with results from: (a) our complete method, (b) the method without joint optimization, and (c) the method without selective supervisions. Please visit our project page for animated results." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 129, + 502, + 206, + 546 + ], + "blocks": [ + { + "bbox": [ + 129, + 502, + 206, + 546 + ], + "lines": [ + { + "bbox": [ + 129, + 502, + 206, + 546 + ], + "spans": [ + { + "bbox": [ + 129, + 502, + 206, + 546 + ], + "type": "image", + "image_path": "d31575ea6e7a88fb837fe24bd44153cdd1d6e2f395605b26d9aded90c44f98d1.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 161, + 548, + 171, + 558 + ], + "lines": [ + { + "bbox": [ + 161, + 548, + 171, + 558 + ], + "spans": [ + { + "bbox": [ + 161, + 548, + 171, + 558 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 206, + 502, + 284, + 546 + ], + "blocks": [ + { + "bbox": [ + 206, + 502, + 284, + 546 + ], + "lines": [ + { + "bbox": [ + 206, + 502, + 284, + 546 + ], + "spans": [ + { + "bbox": [ + 206, + 502, + 284, + 546 + ], + "type": "image", + "image_path": "1dfb77ab3bbda8f496972a0ad9f7dd965e64cb88cfe070a2b127fde2291c88e2.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 238, + 548, + 249, + 558 + ], + "lines": [ + { + "bbox": [ + 238, + 548, + 249, + 558 + ], + "spans": [ + { + "bbox": [ + 238, + 548, + 249, + 558 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 633, + 125, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 633, + 125, + 645 + ], + "spans": [ + { + "bbox": [ + 47, + 633, + 125, + 645 + ], + "type": "text", + "content": "6. Conclusions" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 654, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 286, + 713 + ], + "type": "text", + "content": "In this paper, grounded on Neural Radiance Fields, we construct an occlusion-free scene representation, which is capable of rendering occlusion-free images with desired viewpoints. By directly using the viewpoints as a part of the input, our method is able to aggregate occluded details" + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 311, + 205, + 388, + 256 + ], + "blocks": [ + { + "bbox": [ + 311, + 205, + 388, + 256 + ], + "lines": [ + { + "bbox": [ + 311, + 205, + 388, + 256 + ], + "spans": [ + { + "bbox": [ + 311, + 205, + 388, + 256 + ], + "type": "image", + "image_path": "1c7c7da4d783196c6225b0bffd6047fb0b3026b1d59cf2f64fda5258cd56bebf.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 343, + 258, + 353, + 267 + ], + "lines": [ + { + "bbox": [ + 343, + 258, + 353, + 267 + ], + "spans": [ + { + "bbox": [ + 343, + 258, + 353, + 267 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 390, + 205, + 466, + 256 + ], + "blocks": [ + { + "bbox": [ + 390, + 205, + 466, + 256 + ], + "lines": [ + { + "bbox": [ + 390, + 205, + 466, + 256 + ], + "spans": [ + { + "bbox": [ + 390, + 205, + 466, + 256 + ], + "type": "image", + "image_path": "1bc9fc9c305487a108c62887045dfe07e7fb6643ce2b5fe2e6b6dac8edbbb849.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 421, + 258, + 431, + 267 + ], + "lines": [ + { + "bbox": [ + 421, + 258, + 431, + 267 + ], + "spans": [ + { + "bbox": [ + 421, + 258, + 431, + 267 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 273, + 545, + 306 + ], + "lines": [ + { + "bbox": [ + 306, + 273, + 545, + 306 + ], + "spans": [ + { + "bbox": [ + 306, + 273, + 545, + 306 + ], + "type": "text", + "content": "Figure 8. A failure scene of the proposed method, when reflection is present as occlusion. (a) input sample, (b) results of the scene MLP, (c) failed reflection removal with the background MLP." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 468, + 205, + 544, + 256 + ], + "blocks": [ + { + "bbox": [ + 468, + 205, + 544, + 256 + ], + "lines": [ + { + "bbox": [ + 468, + 205, + 544, + 256 + ], + "spans": [ + { + "bbox": [ + 468, + 205, + 544, + 256 + ], + "type": "image", + "image_path": "7d15b9f0b0151f2ba1d655aef24dfa0fdb904335e48fe224fb04d46cafae02e7.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 498, + 258, + 509, + 267 + ], + "lines": [ + { + "bbox": [ + 498, + 258, + 509, + 267 + ], + "spans": [ + { + "bbox": [ + 498, + 258, + 509, + 267 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 322, + 545, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 322, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 305, + 322, + 545, + 430 + ], + "type": "text", + "content": "from multiple viewpoints without relying on any external guidance. We propose a pose refinement scheme to ensure robust training when the camera poses cannot be accurately estimated by COLMAP [23]. To focus on modeling of desired background scenes within a specific viewpoint, we further introduce a depth constraint to probe the occluded areas by measuring the depth of occlusion and background. Experimental results show that our method is capable of achieving promising results under a variety of scenarios." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 305, + 444, + 545, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 444, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 305, + 444, + 545, + 624 + ], + "type": "text", + "content": "Limitations. Despite the promising performance of our proposed method, several limitations are still to be addressed in our future study. First, our method clearly assumes the undesired layers are in the foreground. It cannot disambiguate the cases where details from different layers are intertwined, such as scenes with reflection, which is especially challenging due to its semi-transparency and widespread presence. We show a failure case of the proposed method in a scene with reflection as occlusion in Fig. 8. Besides, since several variants of NeRF have been proposed for faster convergence and more accurate reconstruction [5,20], advanced NeRF variants may boost the performance of the proposed method. In addition, we also hope to reduce the number of images required for joint optimization of camera parameters and NeRF in our future work." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 305, + 638, + 545, + 709 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 638, + 545, + 709 + ], + "spans": [ + { + "bbox": [ + 305, + 638, + 545, + 709 + ], + "type": "text", + "content": "Acknowledgement. This work is supported by National Natural Science Foundation of China under Grant No. 62136001, 62088102. Renjie Wan is supported by the Blue Sky Research Fund of HKBU under Grant No. BSRF/21-22/16 and Guangdong Basic and Applied Basic Research Foundation under Grant No. 2022A1515110692." + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20729" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 92, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 53, + 92, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 92, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 92, + 287, + 135 + ], + "type": "text", + "content": "[1] Coloma Ballester, Marcelo Bertalmio, Vicent Caseles, Guillermo Sapiro, and Joan Verdera. Filling-in by joint interpolation of vector fields and gray levels. IEEE Transactions on Image Processing, 2001. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 288, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 288, + 170 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 288, + 170 + ], + "type": "text", + "content": "[2] Marcelo Bertalmio, Guillermo Sapiro, Vincent Caseles, and Coloma Ballester. Image inpainting. In Proc. of ACM SIGGRAPH, 2000. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 172, + 288, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 172, + 288, + 205 + ], + "spans": [ + { + "bbox": [ + 53, + 172, + 288, + 205 + ], + "type": "text", + "content": "[3] Xingyu Chen, Qi Zhang, Xiaoyu Li, Yue Chen, Ying Feng, Xuan Wang, and Jue Wang. Hallucinated neural radiance fields in the wild. In CVPR, 2022. 1, 2, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 207, + 288, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 207, + 288, + 262 + ], + "spans": [ + { + "bbox": [ + 53, + 207, + 288, + 262 + ], + "type": "text", + "content": "[4] Chen Du, Byeongkeun Kang, Zheng Xu, Ji Dai, and Truong Nguyen. Accurate and efficient video de-fencing using convolutional neural networks and temporal information. In Proc. of International Conference on Multimedia and Expo, 2018. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 264, + 288, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 264, + 288, + 297 + ], + "spans": [ + { + "bbox": [ + 53, + 264, + 288, + 297 + ], + "type": "text", + "content": "[5] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In CVPR, 2022. 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 299, + 288, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 299, + 288, + 332 + ], + "spans": [ + { + "bbox": [ + 53, + 299, + 288, + 332 + ], + "type": "text", + "content": "[6] Xiaodong Gu, Zhiwen Fan, Siyu Zhu, Zuozhuo Dai, Feitong Tan, and Ping Tan. Cascade cost volume for high-resolution multi-view stereo and stereo matching. In CVPR, 2020. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 334, + 288, + 366 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 334, + 288, + 366 + ], + "spans": [ + { + "bbox": [ + 53, + 334, + 288, + 366 + ], + "type": "text", + "content": "[7] Yuan-Chen Guo, Di Kang, Linchao Bao, Yu He, and Song-Hai Zhang. NeRFReN: Neural radiance fields with reflections. In CVPR, 2022. 1, 2, 5, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 369, + 288, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 369, + 288, + 401 + ], + "spans": [ + { + "bbox": [ + 53, + 369, + 288, + 401 + ], + "type": "text", + "content": "[8] Xin Huang, Qi Zhang, Ying Feng, Hongdong Li, Xuan Wang, and Qing Wang. HDR-NeRF: High dynamic range neural radiance fields. In CVPR, 2022. 2, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 403, + 288, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 403, + 288, + 447 + ], + "spans": [ + { + "bbox": [ + 53, + 403, + 288, + 447 + ], + "type": "text", + "content": "[9] Jeffrey Ichnowski, Yahav Avigal, Justin Kerr, and Ken Goldberg. Dex-NeRF: Using a neural radiance field to grasp transparent objects. In Proc. of Conference on Robot Learning, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 449, + 288, + 493 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 449, + 288, + 493 + ], + "spans": [ + { + "bbox": [ + 48, + 449, + 288, + 493 + ], + "type": "text", + "content": "[10] Sankaraganesh Jonna, Sukla Satapathy, and Rajiv R Sahay. Stereo image de-fencing using smartphones. In Proc. of International Conference on Acoustics, Speech, and Signal Processing, 2017. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 495, + 288, + 528 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 495, + 288, + 528 + ], + "spans": [ + { + "bbox": [ + 48, + 495, + 288, + 528 + ], + "type": "text", + "content": "[11] Jingyuan Li, Ning Wang, Lefei Zhang, Bo Du, and Dacheng Tao. Recurrent feature reasoning for image inpainting. In CVPR, 2020. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 530, + 288, + 573 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 530, + 288, + 573 + ], + "spans": [ + { + "bbox": [ + 48, + 530, + 288, + 573 + ], + "type": "text", + "content": "[12] Xiaoyu Li, Bo Zhang, Jing Liao, and Pedro V Sander. Let's see clearly: Contaminant artifact removal for moving cameras. In Proc. of International Conference on Computer Vision, 2021. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 575, + 288, + 608 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 575, + 288, + 608 + ], + "spans": [ + { + "bbox": [ + 48, + 575, + 288, + 608 + ], + "type": "text", + "content": "[13] Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. BARF: Bundle-adjusting neural radiance fields. In CVPR, 2021. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 611, + 288, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 611, + 288, + 643 + ], + "spans": [ + { + "bbox": [ + 48, + 611, + 288, + 643 + ], + "type": "text", + "content": "[14] Yu-Lun Liu, Wei-Sheng Lai, Ming-Hsuan Yang, Yung-Yu Chuang, and Jia-Bin Huang. Learning to see through obstructions. In CVPR, 2020. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 644, + 288, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 644, + 288, + 677 + ], + "spans": [ + { + "bbox": [ + 48, + 644, + 288, + 677 + ], + "type": "text", + "content": "[15] Ilya Loshchilov and Frank Hutter. SGDR: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 680, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 288, + 713 + ], + "type": "text", + "content": "[16] Li Ma, Xiaoyu Li, Jing Liao, Qi Zhang, Xuan Wang, Jue Wang, and Pedro V Sander. Deblur-NeRF: Neural radiance fields from blurry images. In CVPR, 2022. 2" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "text", + "content": "[17] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. NeRF in the Wild: Neural radiance fields for unconstrained photo collections. In CVPR, 2021. 1, 2, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 119, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 119, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 119, + 545, + 173 + ], + "type": "text", + "content": "[18] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In Proc. of European Conference on Computer Vision, 2020. 1, 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 175, + 545, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 208 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 208 + ], + "type": "text", + "content": "[19] Yadong Mu, Wei Liu, and Shuicheng Yan. Video de-fencing. IEEE Transactions on Circuits and Systems for Video Technology, 2014. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 210, + 545, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 210, + 545, + 243 + ], + "spans": [ + { + "bbox": [ + 307, + 210, + 545, + 243 + ], + "type": "text", + "content": "[20] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. TOG, 2022. 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 245, + 545, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 245, + 545, + 287 + ], + "spans": [ + { + "bbox": [ + 307, + 245, + 545, + 287 + ], + "type": "text", + "content": "[21] Kamyar Nazeri, Eric Ng, Tony Joseph, Faisal Z Qureshi, and Mehran Ebrahimi. EdgeConnect: Generative image inpainting with adversarial edge learning. arXiv preprint arXiv:1901.00212, 2019. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 290, + 545, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 290, + 545, + 323 + ], + "spans": [ + { + "bbox": [ + 307, + 290, + 545, + 323 + ], + "type": "text", + "content": "[22] Rui Qian, Robby T Tan, Wenhan Yang, Jiajun Su, and Jiaying Liu. Attentive generative adversarial network for raindrop removal from a single image. In CVPR, 2018. 1, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 325, + 545, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 325, + 545, + 346 + ], + "spans": [ + { + "bbox": [ + 307, + 325, + 545, + 346 + ], + "type": "text", + "content": "[23] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In CVPR, 2016. 2, 4, 6, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 348, + 547, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 348, + 547, + 392 + ], + "spans": [ + { + "bbox": [ + 307, + 348, + 547, + 392 + ], + "type": "text", + "content": "[24] Yuhang Song, Chao Yang, Yeji Shen, Peng Wang, Qin Huang, and C-C Jay Kuo. SPG-Net: Segmentation prediction and guidance network for image inpainting. In Proc. of British Machine Vision Conference, 2018. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 394, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 394, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 307, + 394, + 545, + 437 + ], + "type": "text", + "content": "[25] V. Vaish, M. Levoy, R. Szeliski, C.L. Zitnick, and Sing Bing Kang. Reconstructing occluded surfaces using synthetic apertures: Stereo, focus and robust measures. In CVPR, 2006. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 440, + 545, + 494 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 440, + 545, + 494 + ], + "spans": [ + { + "bbox": [ + 307, + 440, + 545, + 494 + ], + "type": "text", + "content": "[26] Suhani Vora, Noha Radwan, Klaus Greff, Henning Meyer, Kyle Genova, Mehdi SM Sajjadi, Etienne Pot, Andrea Tagliasacchi, and Daniel Duckworth. NeSF: Neural semantic fields for generalizable semantic segmentation of 3d scenes. arXiv preprint arXiv:2111.13260, 2021. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 496, + 545, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 496, + 545, + 529 + ], + "spans": [ + { + "bbox": [ + 307, + 496, + 545, + 529 + ], + "type": "text", + "content": "[27] Renjie Wan, Boxin Shi, Haoliang Li, Ling-Yu Duan, and Alex C Kot. Face image reflection removal. International Journal of Computer Vision, 2021. 1, 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 531, + 545, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 531, + 545, + 574 + ], + "spans": [ + { + "bbox": [ + 307, + 531, + 545, + 574 + ], + "type": "text", + "content": "[28] Zirui Wang, Shangzhe Wu, Weidi Xie, Min Chen, and Victor Adrian Prisacariu. NeRF---: Neural radiance fields without known camera parameters. arXiv preprint arXiv:2102.07064, 2021. 2, 4, 8" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 576, + 545, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 576, + 545, + 609 + ], + "spans": [ + { + "bbox": [ + 307, + 576, + 545, + 609 + ], + "type": "text", + "content": "[29] Tianfan Xue, Michael Rubinstein, Ce Liu, and William T Freeman. A computational approach for obstruction-free photography. TOG, 2015. 1, 2, 6" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 611, + 545, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 611, + 545, + 654 + ], + "spans": [ + { + "bbox": [ + 307, + 611, + 545, + 654 + ], + "type": "text", + "content": "[30] Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. MVSNet: Depth inference for unstructured multiview stereo. In Proc. of European Conference on Computer Vision, 2018. 4" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 656, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 656, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 656, + 545, + 678 + ], + "type": "text", + "content": "[31] Renjiao Yi, Jue Wang, and Ping Tan. Automatic fence segmentation in videos of dynamic scenes. In CVPR, 2016. 6" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 545, + 713 + ], + "type": "text", + "content": "[32] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 7" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20730" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 162 + ], + "type": "list", + "angle": 0, + "index": 2, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[33] Shuaifeng Zhi, Tristan Laidlow, Stefan Leutenegger, and Andrew J Davison. In-place scene labelling and understanding with implicit scene representation. In Proc. of International Conference on Computer Vision, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 162 + ], + "type": "text", + "content": "[34] Manyu Zhu, Dongliang He, Xin Li, Chao Li, Fu Li, Xiao Liu, Errui Ding, and Zhaoxiang Zhang. Image inpainting by end-to-end cascaded refinement with mask awareness. IEEE Transactions on Image Processing, 2021. 2" + } + ] + } + ], + "index": 1 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "20731" + } + ] + } + ], + "index": 3 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Octree Guided Unoriented Surface Reconstruction/0db48127-a127-468b-be90-b27ec5813046_content_list.json b/2023/Octree Guided Unoriented Surface Reconstruction/0db48127-a127-468b-be90-b27ec5813046_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..85f2db012f57786d96df7d34184d7cd8458eacf0 --- /dev/null +++ b/2023/Octree Guided Unoriented Surface Reconstruction/0db48127-a127-468b-be90-b27ec5813046_content_list.json @@ -0,0 +1,1892 @@ +[ + { + "type": "text", + "text": "Octree Guided Unoriented Surface Reconstruction", + "text_level": 1, + "bbox": [ + 228, + 130, + 741, + 150 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chamin Hewa Koneputugodage1", + "bbox": [ + 161, + 180, + 423, + 200 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yizhak Ben-Shabat1,2", + "bbox": [ + 465, + 180, + 638, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Stephen Gould", + "bbox": [ + 676, + 180, + 805, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1The Australian National University", + "bbox": [ + 173, + 208, + 460, + 226 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Technion Israel Institute of Technology", + "bbox": [ + 480, + 208, + 800, + 227 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{chamin.hewa, stephen.gould}@anu.edu.au sitszikbs@gmail.com", + "bbox": [ + 225, + 229, + 740, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 279, + 313, + 295 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We address the problem of surface reconstruction from unoriented point clouds. Implicit neural representations (INRs) have become popular for this task, but when information relating to the inside versus outside of a shape is not available (such as shape occupancy, signed distances or surface normal orientation) optimization relies on heuristics and regularizers to recover the surface. These methods can be slow to converge and easily get stuck in local minima. We propose a two-step approach, OG-INR, where we (1) construct a discrete octree and label what is inside and outside (2) optimize for a continuous and high-fidelity shape using an INR that is initially guided by the octree's labelling. To solve for our labelling, we propose an energy function over the discrete structure and provide an efficient move-making algorithm that explores many possible labellings. Furthermore we show that we can easily inject knowledge into the discrete octree, providing a simple way to influence the result from the continuous INR. We evaluate the effectiveness of our approach on two unoriented surface reconstruction datasets and show competitive performance compared to other unoriented, and some oriented, methods. Our results show that the exploration by the move-making algorithm avoids many of the bad local minima reached by purely gradient descent optimized methods (see Figure 1).", + "bbox": [ + 75, + 313, + 473, + 676 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 707, + 209, + 723 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Surface reconstruction from 3D point clouds has been studied extensively in computer vision and computer graphics. The task requires estimating a mesh of the shape's surface from a point cloud sampled from the surface of the shape. We focus on the reconstruction of watertight 3D shapes, i.e., shapes that have a well defined interior and exterior. Such shapes are often represented as signed distance fields (SDFs) or occupancy fields, which can be efficiently encoded by a neural network [28, 30]. As these representations are fields parameterized by neural networks, they are often called neural fields. Furthermore, they implicitly rep", + "bbox": [ + 75, + 734, + 470, + 902 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5de864bc5d353c490ae1de645706df3dd2d6590d7777bb1840f60601319edfee.jpg", + "image_caption": [ + "Figure 1. Our method with a SIREN INR (OG-SIREN, left) compared to SIREN wo n (right) for two shapes from the ShapeNet dataset. Our octree guidance allows for consistent inside-outside determinism. On the other hand, SIREN wo n gets stuck in local minima from which it cannot escape (due to needing to completely change the occupancy of certain areas) creating extraneous surfaces (often called ghost geometries in the literature)." + ], + "image_footnote": [], + "bbox": [ + 503, + 277, + 895, + 397 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "resent the shape by a level set of the field, thus they are also often referred to as implicit neural representations (INRs).", + "bbox": [ + 496, + 536, + 890, + 566 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The broader class of neural fields, including INRs, have been very popular over the last few years as they can handle arbitrary topology, are memory efficient, and are continuous with potentially infinite resolution [39, 46]. Among the INRs for 3D shapes, SDFs are the most popular as they provide more useful information (distances not just occupancy) and are required for downstream graphics algorithms such as sphere tracing and approximate soft shadows [29, 35].", + "bbox": [ + 496, + 566, + 890, + 688 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "When learning an implicit representation of a shape, a major difficulty is predicting whether points in space are inside or outside the shape. Many methods require oriented surface normals for the input points or signed distances from the surface, which usually are not given by raw data from scans. While this can be estimated using the line of sight information [17] or algorithms [4,6, 15, 21], they yield noisy predictions that after postprocessing still can lead to bad results (see Section 4.4). We consider the task of unoriented surface reconstruction, where such information is not available, and only the sampled surface points are given. We demonstrate that our method performs competitively and sometimes better than oriented methods, even when they are given the ground truth (GT) normals.", + "bbox": [ + 496, + 689, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "16717", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4c2eb5865acd0522fff845a067abf9d689f85cc689a884b86aabb874c412093c.jpg", + "image_caption": [ + "Figure 2. An illustration of our method, OG-INR. Given an unoriented point cloud (left), we progressively build and label an octree around the points (middle). The octree at depths 3-7 are shown. Surface leaves of the octree (yellow) are leaves that contain points from the point cloud, other leaves are labelled as inside (blue) or outside (transparent) by minimizing an energy function. We then train an INR model to obtain an SDF, using the labelling as supervision for the initial training, after which we can extract a mesh (right)." + ], + "image_footnote": [], + "bbox": [ + 84, + 95, + 883, + 233 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We propose OG-INR, which uses a discrete representation in conjunction with the continuous INR representation (see Figure 2). Given an input unoriented point cloud, we progressively build an octree from the input points and determine which leaf nodes are surface leaves (i.e., leaves that contain a point). We also label all other leaves as inside or outside leaves (if they should be within or outside the shape respectively). To do this we minimize an energy function that trades off the watertight surface property, that every surface point should border with the inside and outside of the shape, with a minimal surface constraint. We then train standard INR architectures, initially guiding the training by the octree labelling, and show that the INR converges much faster and is less prone to large failure regions.", + "bbox": [ + 75, + 339, + 472, + 551 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main contributions are:", + "bbox": [ + 96, + 551, + 282, + 565 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce a novel method to guide the initial stage of INR training. It uses a labelled octree structure to allow the INR to converge significantly faster and alleviates the local minima problem of INRs.", + "- We propose an energy function over octree labels that captures the task of surface reconstruction. It balances the constraint of maintaining known surface regions with minimising the overall surface area.", + "- We provide an efficient move-making algorithm to optimize the energy function, which explores many inside-outside possibilities in a structured manner.", + "- Our discrete representation is easily understandable and human-modifiable, giving an intuitive method for applying changes to the resulting SDF." + ], + "bbox": [ + 94, + 566, + 468, + 784 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 799, + 218, + 814 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Unsupervised Surface Reconstruction. Representing shapes with INRs was popularized by DeepSDF [30], where a ReLU network directly models an SDF using GT SDF supervision. SAL [1] instead uses unsigned distance supervision, approximated from the input point cloud, as an", + "bbox": [ + 75, + 825, + 468, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "unsupervised and unoriented approach. SALD [2] showed that adding unoriented normal supervision improved upon this, and IGR [14] further added the eikonal constraint loss (see Equation 5) that has been widely adopted [5, 22, 34]. NSP [45] addresses the task by solving a kernel regression problem but needs input normals. PHASE [22] introduces a novel loss inspired by the phase transition problem that encourages boundary crossings at the surface points.", + "bbox": [ + 496, + 339, + 890, + 460 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, as the inputs to INRs are low dimensional, they are biased to low frequency solutions that do not recover high fidelity detail [36]. To introduce high frequencies, SIREN [34] uses sine activations and Fourier Feature Networks [36] use a fixed Fourier encoding. While being able to reconstruct high fidelity shapes with a modest number of parameters, these methods are hard to train as the high frequencies overfit quickly. A different approach is parametric encoding [29, 35], where spatial regions are represented by trainable parameters (often using an octree structure). We cover such methods in the next subsection.", + "bbox": [ + 496, + 460, + 892, + 626 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Note that IGR and SIREN can operate without normals, which we refer to by \"IGR wo n\" and \"SIREN wo n\".", + "bbox": [ + 498, + 627, + 890, + 656 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Octrees for Surface Reconstruction. Octrees were first introduced by Meagher [27] as an efficient data structure for representing and manipulating 3D data, and are now used extensively. The classic method Poisson Surface Reconstruction (PSR) [18] and its followup SPSR [19] use octrees to structure their implicit functions.", + "bbox": [ + 496, + 657, + 890, + 747 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Early works on using octrees with neural networks include HSP [16] and OGN [38], who train encoder-decoder networks to learn octrees with occupancy values, and OctNet [32] and O-CNN [40], who implement 3D CNNs that operate over octrees rather than voxel grids for shape analysis tasks. O-CNN's descendants [41, 43] build on this and adapt it to shape generation, completion and reconstruction.", + "bbox": [ + 496, + 750, + 890, + 854 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "More recently octrees have been used in INRs for representing fine details efficiently. OctField [37] and NSVF [23] both learn local implicit functions for individual octree", + "bbox": [ + 498, + 854, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "16718", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "nodes. NGLOD [35] and ACORN [26] instead learn a single INR, but store trainable parameters at octree node vertices to encode spatial regions. IMLSNet [24] uses O-CNN to predict both an octree and oriented points per octree node to represent the surface using implicit moving least squares.", + "bbox": [ + 75, + 90, + 468, + 167 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, previous octree reconstruction methods were supervised, trained on GT SDF/occupancy of either the current shape or a separate training set. Recently Dual Octree Graph CNNs [42] addressed the unsupervised setting, where they learn local features on the dual graph of an octree. However their method has parameters that need unsupervised training, and they require input normals.", + "bbox": [ + 75, + 169, + 467, + 273 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "SAP [31] truly addresses the unsupervised and unoriented task by making the PSR pipeline differentiable. Starting from an oriented point cloud of a sphere, they use Chamfer loss to guide that point cloud to a dense, oriented representation. Their method is fast and among state-of-the-art, however it can fail drastically and often does not produce a watertight mesh (see Section 4). It is also hard to adapt to different applications and losses, unlike INRs.", + "bbox": [ + 75, + 276, + 467, + 397 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Initialization for INRs. Due to the difficulty of training INRs, initialization is very important [1, 5]. However regardless of input, most methods initialize to a sphere. These include SAL [1] and its descendants [2, 14, 22], DiGS [5] and IDF [47]. This gives many important properties, such as biasing shape interior to the domain's center and satisfying the eikonal condition at the start of training, but it is problematic for shapes with sharp concavities. Our method, on the other hand, provides firm guidance for both the eikonal constraint and the shape based on the input point cloud, and thus does not bias the training toward convex objects.", + "bbox": [ + 75, + 398, + 467, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Energy Minimization. Discrete energy minimization has been used extensively in both 2D and 3D vision for segmentation and object labelling [9, 10]. An energy function encodes our preference over the joint assignment to a set of variables, e.g., binary variables representing occupancy of voxels in 3D space. The goal of energy minimization is to find an assignment to the variables with minimal cost. This is NP-hard in general and a number of approximation algorithms have been proposed [9]. One class of algorithms, known as move-making, starts with an initial assignment and iteratively evaluates changes to a subset of the variables (called a move). If the change results in lower energy then the move is accepted, otherwise the changed values are reverted. Moves are repeatedly tried until no further reduction in energy is found. Iterated Conditional Modes (ICM) [8] is the simplest example, where each move involves changing a single variable and so is very efficient, but it is prone to getting stuck in local minima. Generalized ICM improves on this by allowing multiple variables to change in a move, but it is very costly. Our method is a variant of generalized ICM specialized to our problem, where we introduce restrictions on the move space improve efficiency.", + "bbox": [ + 75, + 568, + 468, + 900 + ], + "page_idx": 2 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 Overall octree labeling algorithm" + ], + "code_body": "1: procedure LABELOCTREE $(\\chi ,d_i,k,C)$ \n2: $\\mathcal{O} =$ CreateOctree $(\\chi ,d_{i})$ \n3: Initialization(O) \n4: for $j$ in $0,\\dots ,k$ do \n5: if $j\\neq 0$ then \n6: ExpandTree(O) \n7: end if ▷ Now at depth $d = d_{i} + j$ \n8: Grow(O) \n9: for $s$ in $C$ do \n10: makeMove(O,s) \n11: end for \n12: end for \n13: return O \n14: end procedure", + "bbox": [ + 506, + 108, + 890, + 325 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. OG-INR: Octree Guided Shape INRs", + "text_level": 1, + "bbox": [ + 498, + 354, + 834, + 373 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Formulation", + "text_level": 1, + "bbox": [ + 500, + 383, + 702, + 398 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let $\\chi = \\{x_{i}\\}_{i\\in I}$ be a point cloud sampled from a surface $S = \\partial \\mathcal{V}$ of some watertight volume $\\mathcal{V} \\subset \\mathcal{D}$ where $\\mathcal{D} \\subset \\mathbb{R}^3$ is compact. We wish to reconstruct the surface $S$ . We choose to represent $S$ as the zero-level set of a signed distance function (SDF), and thus wish to train a neural network $\\Phi$ to output signed distances for all points $x \\in \\mathcal{D}$ ,", + "bbox": [ + 498, + 407, + 890, + 500 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\Phi (x, \\theta) = (- 1) ^ {o (x)} d (x, \\mathcal {S}) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 515, + 890, + 534 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\theta$ are the model parameters, $o(x)$ is the occupancy of point x (1 if inside and 0 if outside) and $d(x,S)$ is the distance from $x$ to its closest point on $S$ .", + "bbox": [ + 498, + 550, + 890, + 595 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Overview", + "text_level": 1, + "bbox": [ + 500, + 612, + 609, + 627 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We first aim to produce a rough and discrete, but globally consistent, occupancy determination for our domain by solving for an inside-outside labelling on nodes of an octree. To do this we propose an energy function that combines both the watertight surface property and the minimal surface objective. Despite this energy being intractable to solve exactly, we provide a domain-guided move-making algorithm to get low energy solutions (see Section 3.3).", + "bbox": [ + 496, + 638, + 890, + 758 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We then design additional INR loss terms to guide INRs using our octree labelling (see Section 3.4). These losses quickly guide the INR towards a rough but globally consistent representation that allows it to avoid many bad local minima. Since these loss terms are agnostic to INR architecture, our method can work with any INR model.", + "bbox": [ + 496, + 762, + 890, + 852 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As we assume our domain is compact, we first normalize the point cloud to fit within the unit sphere in 3D, and then set our domain as $\\mathcal{D} = [-1.1, 1.1]^3$ .", + "bbox": [ + 496, + 854, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "16719", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f8e6d3e2c77f83f9c633127372ee809593877ed8787d6deaacee1dc63003561e.jpg", + "image_caption": [ + "A: Depth 3, After Initialization(O)" + ], + "image_footnote": [], + "bbox": [ + 158, + 88, + 318, + 196 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d0e11320b3b5133d63e407682efb364864df0910d567f0ab41091ea9cfbbf82a.jpg", + "image_caption": [ + "B: Depth 3, After Grow(O)" + ], + "image_footnote": [], + "bbox": [ + 320, + 88, + 478, + 196 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6792545c03f90c3363a9f1c34b9830756db7c86ee42e7b34da3a1c4ad4ae3dba.jpg", + "image_caption": [ + "C: Depth 3, After all makeMove(O,s)" + ], + "image_footnote": [], + "bbox": [ + 480, + 88, + 640, + 196 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/7a1ba188c9c51b656e16c8c46a2a344d98b0f719b8159361c13834b4468f142b.jpg", + "image_caption": [ + "D: Depth 4, After all makeMove(O, s)" + ], + "image_footnote": [], + "bbox": [ + 650, + 88, + 810, + 196 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c74dd1814d045289fb252dafcc233062c15cbde10bf3e05c29cbb16e54837a3d.jpg", + "image_caption": [ + "E: Depth 5, After Grow(O)", + "Figure 3. Our algorithm demonstrated on a 2D quadtree example. The black dots are our input point cloud, the yellow squares are surface leaves, and the blue and gray squares are non-surface leaves labelled as inside and outside respectively." + ], + "image_footnote": [], + "bbox": [ + 158, + 215, + 318, + 321 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1868eaab03bd81417376bc723aef5ce934ae151b0b40a9f884cb014f065fa34b.jpg", + "image_caption": [ + "F: Depth 5, After makeMove(O,1)" + ], + "image_footnote": [], + "bbox": [ + 320, + 215, + 478, + 321 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/137f8dc3155e079b792fa48ddef725f8cf477f359953eb425859653beac36864.jpg", + "image_caption": [ + "G: Depth 5, After all makeMove(O,s)" + ], + "image_footnote": [], + "bbox": [ + 480, + 215, + 638, + 321 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/eac3eca101ede04a8fd90faad2d351a6964c8305332a7203f3f8935c26f0bec7.jpg", + "image_caption": [ + "H: Depth 6, After all makeMove(O,s)" + ], + "image_footnote": [], + "bbox": [ + 650, + 215, + 810, + 321 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Octree Construction and Labelling", + "text_level": 1, + "bbox": [ + 76, + 407, + 380, + 424 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Setup. We present our octree labelling algorithm in Algorithm 1, which takes in point cloud $\\chi$ , an initial depth to build the octree to $d_{i}$ , a number of further depth expansions $k$ (the final octree has depth $d_{f} = d_{i} + k$ ) and a set of move set sizes $C$ . We set $d = 3$ , $k = 4$ and $C = [1,2,10,10000]$ in our experiments. The algorithm returns a labelled octree $\\mathcal{O} = (\\mathcal{N},\\mathcal{L}_{ns},\\mathcal{L}_s,y)$ where", + "bbox": [ + 75, + 431, + 468, + 537 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- $\\mathcal{N}$ are the nodes of the octree. The octree is built by repeatedly expanding nodes containing or near points from $\\chi$ , so it is highly unbalanced.", + "- $\\mathcal{L} \\subset \\mathcal{N}$ are the leaf nodes, and thus $\\mathcal{L}$ partitions $\\mathcal{D}$ . Furthermore, we define surface leaves $\\mathcal{L}_s \\subset \\mathcal{L}$ to be leaves that contain points from the point cloud (colored yellow in Figure 2 and Figure 3) and all other leaves to be non-surface leaves $\\mathcal{L}_{ns} = \\mathcal{L} \\setminus \\mathcal{L}_s$ .", + "- $y$ is the binary occupancy label vector indexed by non-surface leaves $\\ell \\in \\mathcal{L}_{ns}$ denoting whether the space defined by each $\\ell$ is inside, $y_{\\ell} = 1$ , or outside, $y_{\\ell} = 0$ ." + ], + "bbox": [ + 94, + 539, + 468, + 710 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We will construct our algorithm so that the neighbours of every surface leaf $\\ell$ is at the same depth. Thus let $n_{\\ell}^{s}$ and $n_{\\ell}^{ns}$ be the number of its surface and non-surface neighbours and $n_{\\ell}^{(1)}(\\pmb{y})$ and $n_{\\ell}^{(0)}(\\pmb{y})$ be the number of inside and outside neighbors w.r.t. labelling $\\pmb{y}$ . Then $n_{\\ell}^{ns} = n_{\\ell}^{(1)}(\\pmb{y}) + n_{\\ell}^{(0)}(\\pmb{y})$ and $n_{\\ell}^{s} + n_{\\ell}^{s} = 26$ (or 8 for our 2D example in Figure 3).", + "bbox": [ + 75, + 712, + 468, + 808 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We will constantly refer to the boundary as the leaves touching the boundary of $\\mathcal{D}$ and the inside-outside border to be the non-surface leaves who have a non-surface neighbor with a different label (different occupancy value).", + "bbox": [ + 75, + 809, + 468, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We determine inside-outside occupancy based on three properties: (1) surface property: surfaces should separate", + "bbox": [ + 75, + 869, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "inside and outside regions, (2) minimal surface constraint: the total surface area of the shape's surface $S$ should be as small as possible, (3) boundary constraint: the boundary of the domain $\\mathcal{D}$ should be outside.", + "bbox": [ + 496, + 409, + 890, + 467 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Energy function. We propose an energy function over labellings $\\mathbf{y} = \\{y \\mid y_i \\in \\{0,1\\}, i \\in \\mathcal{L}_{ns}\\}$ based on the first two criteria, the third is hard-coded at the start of the algorithm. For the surface property, ideally every surface leaf would have both inside and outside neighbours, however it may only have surface leaf neighbours which contain both inside and outside regions by definition (see Figure 3 C,D,G,H, some surface leafs have no inside neighbours). For the minimal surface constraint, we only create surfaces when adjacent non-surface leaves have different labellings (note surface leaves are fixed), so we want to minimise this.", + "bbox": [ + 496, + 469, + 890, + 633 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For the surface property, we want every surface leaf $\\ell$ to have some amount of both inside and outside neighbours, where surface neighbours partially count towards both. Thus, we aim for $\\gamma_{\\ell}^{(1)} = \\gamma^{(1)} - \\eta^{(1)}n_{\\ell}^{s}$ non-surface inside neighbors and $\\gamma_{\\ell}^{(0)} = \\gamma^{(0)} - \\eta^{(0)}n_{\\ell}^{s}$ non-surface outside neighbors where $\\gamma^{(\\cdot)},\\eta^{(\\cdot)}$ are hyperparameters. Then the energy cost for a surface leaf $\\ell$ is", + "bbox": [ + 496, + 635, + 890, + 746 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nE _ {\\ell} ^ {S P} (\\boldsymbol {y}) = \\max \\left\\{\\gamma_ {\\ell} ^ {(1)} - n _ {\\ell} ^ {(1)} (\\boldsymbol {y}), \\gamma_ {\\ell} ^ {(0)} - n _ {\\ell} ^ {(0)} (\\boldsymbol {y}), 0 \\right\\}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 503, + 751, + 890, + 777 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For the minimal surface constraint, we want to minimise the total surface area between every pair of non-surface neighbors with different labellings. Then for two non-surface leaves $\\ell, \\ell'$ , our energy cost is", + "bbox": [ + 496, + 781, + 890, + 840 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nE _ {\\ell , \\ell^ {\\prime}} ^ {M S C} (\\boldsymbol {y}) = a _ {\\ell \\ell^ {\\prime}} \\left| y _ {\\ell} - y _ {\\ell^ {\\prime}} \\right| \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 845, + 890, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $a_{\\ell \\ell^{\\prime}}$ is the (surface) area in common between them (note $a_{ij} = 0$ if they are not adjacent).", + "bbox": [ + 496, + 869, + 890, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "16720", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The two criteria can be formalized as an energy function over labellings $\\mathbf{y} = \\{y_{i}\\in \\{0,1\\} \\mid i\\in \\mathcal{L}_{ns}\\}$ by", + "bbox": [ + 76, + 90, + 468, + 122 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nE (\\boldsymbol {y}) = \\sum_ {j \\in \\mathcal {L} _ {s}} E _ {j} ^ {S P} (\\boldsymbol {y}) + \\lambda \\sum_ {i, j \\in \\mathcal {L} _ {n s}} E _ {i, j} ^ {M S C} (\\boldsymbol {y}) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 122, + 130, + 468, + 165 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda$ is a hyperparameter.", + "bbox": [ + 76, + 172, + 272, + 186 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.1 Octree Creation and Labelling Algorithm", + "text_level": 1, + "bbox": [ + 76, + 205, + 419, + 220 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The energy function in Equation 4 is non-trivial to minimize. It is a discrete energy minimisation problem with hard unary terms (the boundary constraint), pairwise terms (Equation 3 which is essentially a smoothness constraint) and n-ary terms (Equation 2). The n-ary term however is not concave, so we cannot use existing methods to perform exact inference using graph-cuts [13,20].", + "bbox": [ + 75, + 229, + 468, + 334 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Instead, we employ a move-making algorithm [9] with moves specialized for our task (see Figure 3). We first hard code the boundary constraint (leaves on the boundary of the domain are outside) and initialize all other leaves to inside. To be efficient, we (1) only consider moves that change leaves from inside to outside (never the other way around), (2) require such moves include an inside leaf from the inside-outside border, (3) iteratively grow such potential moves, and (4) apply moves at multiple depths. We also have a grow stage to remove extraneous inside leaves without needing guidance from the energy function. We now describe the stages of our algorithm, depicted in Algorithm 1.", + "bbox": [ + 75, + 335, + 468, + 516 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Octree Creation. We start with a single node $n_0$ covering the whole space $\\mathcal{D}$ . We then recursively expand any node (into 8 equally sized child nodes) that contains points from $\\chi$ in it until depth $d_i$ is reached. Note that every surface leaf is at depth $d$ .", + "bbox": [ + 75, + 517, + 468, + 590 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Initialisation. We hard code our boundary condition and set every other leaf node to inside (see Figure 3 A).", + "bbox": [ + 76, + 592, + 468, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Grow Stage. We first expand any neighbor of a surface leaf that is not also at the current depth $d$ (new leaves keep the label of its parent). We repeatedly expand the outside region changing inside leaves to outside if they stick out from a plane, i.e., have at least $26/2 + 1 = 14$ outside neighbors (or $8/2 + 1 = 5$ for our 2D example in Figure 3). Finally we expand nodes that are on the inside-outsider border and are not at the current octree depth.", + "bbox": [ + 75, + 621, + 468, + 742 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Move-Making Stage. To overcome local minima we propose sets of leaves to change from inside to outside in a single move. Sets are constructed as follows. Starting with an inside leaf on the border we greedily add neighboring inside leaves to the set, selecting the neighbor which results in lowest energy, and calculate the cost of the current move set. Note that the energy can temporarily increase. We do this until there are no inside neighbors remaining or a maximum set size $s$ is reached.1 If any of these moves result", + "bbox": [ + 75, + 743, + 468, + 878 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "in a decrease in energy, we apply the move with the largest decrease, We repeat the above for different starting leaves and for multiple increasing set sizes $s\\in C$", + "bbox": [ + 500, + 90, + 890, + 135 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Figure 3 illustrates this process in 2D. After reaching labelling (E), all single variable changes ( $s = 1$ ) are considered, resulting in (F). Observe that the inside-outside border has been refined but larger regions are still incorrectly labelled. Considering larger moves as shown in (G) fixes many of these regions, with further refinements being made at deeper levels of the octree as shown in (H). We explain this 2D example in more detail in the supplemental material, including why we cannot reach a lower energy from fixing certain regions until a deeper level is reached.", + "bbox": [ + 496, + 136, + 890, + 286 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Expanding. We find that labelling at multiple depths is faster and more stable than directly labelling the largest depth. Note that expanding completely changes $\\mathcal{L}_s$ and adds more nodes to $\\mathcal{L}_{ns}$ , and thus expands $y$ . We keep the labels for previous non-surface leaves, and assign all the new non-surface leaves to inside.", + "bbox": [ + 496, + 287, + 890, + 377 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. SDF learning", + "text_level": 1, + "bbox": [ + 500, + 386, + 638, + 402 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use the octree to guide the SDF learning. Given an INR, we train it with the combination of four losses. The first two are the standard losses in the literature [1,2,5,14, 34], the surface loss to enforce that the input point cloud should lie on the zero level set and the eikonal loss [14] to enforce that the gradient has magnitude one everywhere", + "bbox": [ + 496, + 409, + 890, + 501 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {1} = \\sum_ {x \\in \\chi} | \\Phi (x, \\theta) |, \\mathcal {L} _ {2} = \\int_ {\\mathcal {D}} | \\| \\nabla_ {x} \\Phi (x, \\theta) \\| _ {2} - 1 | d x. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 501, + 510, + 890, + 546 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We now introduce two loss functions that take guidance from the labelled octree. We first produce an approximate signed distance function $\\tilde{d}_s(x)$ for finely sampled points $x\\in \\mathcal{D}$ by computing the distance to the closest input point cloud point and choose its sign based on the label of the octree leaf that contains that point (for surface leaves we choose inside). We use this as direct supervision", + "bbox": [ + 496, + 556, + 890, + 662 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {3} = \\int_ {\\mathcal {D}} | \\Phi (x, \\theta) - \\tilde {d} _ {s} (x) | \\mathrm {d} x. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 671, + 890, + 704 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We also enforce the sign of each octree leaf explicitly (again taking surface leaves to be considered as inside) at a fixed number of randomly sampled points in that leaf $S_{\\ell}$", + "bbox": [ + 496, + 710, + 890, + 757 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {4} = \\sum_ {\\ell \\in \\mathcal {L}} \\left(\\sum_ {x \\in S _ {\\ell}} \\left[ (- 1) ^ {\\left(1 - y _ {\\ell} ^ {\\prime}\\right)} \\Phi (x, \\theta) \\right] _ {> 0}\\right) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 766, + 890, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $y_{\\ell}^{\\prime} = y_{\\ell}$ for $\\ell \\in \\mathcal{L}_{ns}$ and $y_{\\ell}^{\\prime} = 1$ for $\\ell \\in \\mathcal{L}_s$ . Note that sampling a fixed number per leaf automatically adjusts our sampling to be denser in more important regions (i.e., closer to the surface) due to the octree structure.", + "bbox": [ + 496, + 816, + 890, + 878 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "1In practice we include additional early stopping criteria to avoid moves", + "bbox": [ + 94, + 886, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "that are unlikely to eventually be accepted (see supplemental for details).", + "bbox": [ + 500, + 886, + 883, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "16721", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/cf02dbfef551deea93b6c981bba572e6dcc901c4ef9e131c27ea4e8d6e2f4a3e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 106, + 90, + 225, + 172 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/59d4b6c9af85268ca638d5411996e72771ca83eae409b4bd53e4c0a35d2120c8.jpg", + "image_caption": [ + "Depth 7" + ], + "image_footnote": [], + "bbox": [ + 106, + 181, + 222, + 266 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e29c96d12c52aa202598aa68dba57b202a065e16a592c7806e70a9dc1522bc4c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 233, + 90, + 352, + 172 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/23f4291f6c796179b1eb1a1b198f038378670bfea4d4a266f6dde2bcfac05fe0.jpg", + "image_caption": [ + "Depth 6" + ], + "image_footnote": [], + "bbox": [ + 233, + 181, + 349, + 266 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a58e93455b70c98fa68ef061c4771657b3147a8f0f0bca953c8a9002ac542583.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 361, + 97, + 480, + 172 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/966fbf36fd3d67bf945d4fa9a375a4a7df7ea805117cba6fd9d7d18981d895f6.jpg", + "image_caption": [ + "SAP [31]", + "Figure 4. A comparison of surface reconstructions of shapes from the watercraft and bench classes of ShapeNet. The left two reconstructions are using OG-SIREN with the octree built to different depths." + ], + "image_footnote": [], + "bbox": [ + 361, + 181, + 477, + 266 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2627356b92797d69c7ee784fdfc3805b03fa2631acb814b35f92190b7d5e3cf5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 491, + 90, + 609, + 172 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f9c98c37b2c9267b067220a4b2543d3b5dce4e068afae9a2d3d97810903853ea.jpg", + "image_caption": [ + "DiGS [5]" + ], + "image_footnote": [], + "bbox": [ + 491, + 181, + 604, + 266 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c0de15092a2816fc81dfc4896807309f2496d9f8e205671fdccffddf100ac226.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 617, + 90, + 733, + 172 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3aedd6e7cc40a7ed6823ef2c29571a59e4f0d60ea00467ee07e45335e5a20639.jpg", + "image_caption": [ + "SAL [1]" + ], + "image_footnote": [], + "bbox": [ + 620, + 181, + 732, + 265 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c09ef78ac6c66bf30905402a33ffb4ff88416f88bcd96e96c774d9ed0ee3fe70.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 743, + 90, + 862, + 172 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1b709118308ae31c6c3b8114b850aa24fd499e2437ee5ae58a43f887c3c56f94.jpg", + "image_caption": [ + "SIREN won [34]" + ], + "image_footnote": [], + "bbox": [ + 745, + 181, + 859, + 266 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our overall loss is then the weighted combination", + "bbox": [ + 96, + 352, + 426, + 367 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} (\\Phi) = \\lambda_ {1} \\mathcal {L} _ {1} + \\lambda_ {2} \\mathcal {L} _ {2} + \\lambda_ {3} \\mathcal {L} _ {3} + \\lambda_ {4} \\mathcal {L} _ {4}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 378, + 468, + 395 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We keep $\\lambda_{1}$ and $\\lambda_{2}$ fixed throughout training as their loss terms have no bias. As the guidance in $\\mathcal{L}_3$ and $\\mathcal{L}_4$ is approximate but highly informative, we initially use large values for $\\lambda_3,\\lambda_4$ , allowing the INR to quickly fit to the rough shape, and then reduce them during training. More training details are in the supplemental material.", + "bbox": [ + 75, + 407, + 468, + 500 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Results", + "text_level": 1, + "bbox": [ + 76, + 512, + 163, + 527 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We evaluate the performance of our method on the ShapeNet [12] and Surface Reconstruction Benchmark (SRB) [7] datasets. ShapeNet has a wide variety of shapes, with thin surfaces, hollow regions and disjoint objects. However the shapes are rather uniformly sampled. We use the preprocessing and split of Neural Splines [45], which consists of 13 shape classes with 20 shapes in each shape class. For each shape it provides the ground truth mesh (preprocessed to be watertight), an input point cloud and occupancy annotation at dense points. SRB contains five shapes from simulated noisy range scans, but it is a popularly used dataset due to the shapes exhibiting many challenging features. These include large missing regions, thin and sharp regions, and small surface details. The former means that good priors in the reconstruction methods, such as smoothness priors, are very important.", + "bbox": [ + 75, + 537, + 468, + 779 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We follow the metrics used in the literature for each dataset: the squared Chamfer distance and IoU for ShapeNet and Chamfer and Hausdorff distance for SRB.", + "bbox": [ + 75, + 779, + 468, + 824 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "INRs. We demonstrate our method on two INRs, SIREN [34], and NGLOD [35] (note that NGLOD only works in a supervised setting as it has no global consistency, but our octree guidance is sufficient). Thus we name our two methods OG-SIREN and OG-NGLOD where OG stands for", + "bbox": [ + 75, + 824, + 468, + 898 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "\"Octree Guided\". To extract a mesh we first evaluate our INRs on a $256^3$ grid and then use marching cubes [25]. We use the same settings for the octree and the INRs on both datasets, and give more details in our supplemental.", + "bbox": [ + 496, + 352, + 892, + 412 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Reconstruction Performance", + "text_level": 1, + "bbox": [ + 498, + 422, + 756, + 436 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ShapeNet dataset. The results are shown in Table 1 (left) and visualizations on two shapes are shown in Figure 4. Our methods get the smallest mean squared Chamfer distance out of the unoriented methods by almost an order of magnitude, and are even better than the best performing oriented method. Likewise for mean IoU our methods are better than both unoriented and oriented methods. For median performance, our methods are comparable to the best oriented and unoriented methods. SAP [31] gets the best median IoU, but their mean is much lower than ours, indicating that in a few instances they misrepresent large regions. On the other hand, due to our move-making algorithm we reduce the likelihood of vastly incorrect labellings, and as a result achieve low standard deviations on both metrics despite the high performance.", + "bbox": [ + 496, + 445, + 892, + 671 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "SRB dataset. Table 1 (right) shows that DiGS, SAP, PHASE+FF and our two methods do comparably, with DiGS performing slightly better. Note the performance of these methods are comparable to most oriented methods. Also note that the difference between our two methods is more prominent on SRB as the large missing regions means that smoothness priors are more important. Octree guidance is NGLOD's only global supervision, as it fits locally, while SIREN is biased towards smooth shapes.", + "bbox": [ + 496, + 671, + 890, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Octree Depth Ablations", + "text_level": 1, + "bbox": [ + 500, + 816, + 718, + 832 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 4 shows the results of OG-SIREN on ShapeNet when evaluating at different depths of the octree. We include an octree IoU metric, where we compute the IoU of the labelling of the octree directly (taking surface leaves to", + "bbox": [ + 496, + 839, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "16722", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/fa83e085aa3b24c26e597ef68cae30ee989d8da1a04ba16f06860685288365d3.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodSquared Chamfer ↓IoU ↑
meanmedianstdmeanmedianstd
SPSR [19]5.44e-51.97e-55.06e-40.99260.99560.0105
IGR [14]5.12e-41.13e-42.15e-30.81020.84800.1519
SIREN [34]1.03e-45.28e-51.93e-40.82680.90970.2329
FFN [36]9.12e-58.65e-53.36e-50.82180.83960.0989
Biharmonic RBF [11]1.11e-48.97e-57.06e-50.82470.86420.1350
SVR [33]1.14e-41.04e-45.99e-50.76250.78190.1300
NSP [45]5.36e-54.06e-53.64e-50.89730.92300.0871
DiGS + n [5]2.74e-42.32e-59.90e-40.92000.97740.1992
N Est+SPSR [19]3.76e-34.37e-51.14e-20.71870.97610.3767
SIREN won [34]3.08e-42.58e-43.26e-40.30850.29520.2014
SAL [1]1.14e-32.11e-43.63e-30.40300.39440.2722
DiGS [5]1.32e-42.55e-54.73e-40.93900.97640.1262
SAP [31]4.09e-42.46e-52.60e-30.91180.99230.2002
Our OG-SIREN3.75e-52.17e-58.24e-50.96150.98710.1048
Our OG-NGLOD5.07e-52.57e-59.39e-50.95930.98700.1057
", + "bbox": [ + 89, + 93, + 560, + 343 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/84a04421c2ddb5a2135b9d0b50c465ae82a2f2ecf2f73e7644bcfa54aba1d3b8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethoddC↓dH↓
SPSR [19]0.214.69
DGP [44]0.215.18
IGR [14]0.192.99
SIREN [34]0.193.86
NSP [45]0.172.85
PHASE [22]0.162.77
DiGS + n [5]0.183.55
N Est.+SPSR [19]1.2522.59
IGR wo n [14]1.3816.33
SIREN wo n [34]0.427.67
SAL [14]0.367.47
IGR+FF [14]0.9611.06
PHASE+FF [22]0.224.96
DiGS [5]0.193.52
SAP [31]0.214.51
Our OG-SIREN0.204.06
Our OG-NGLOD0.226.03
", + "bbox": [ + 632, + 89, + 836, + 349 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/747f748de88ce733f4d841276f5451c326c502ae2a2ad88ece8d8f3443e64cb1.jpg", + "table_caption": [ + "Table 1. Surface Reconstruction results on ShapeNet [12] (left) and SRB [7] (right). Methods above the line use ground truth normal information, and methods below do not. Left: The mean, median and standard deviation of the squared Chamfer distance and IoU of all 260 shapes are reported. Right: The mean Chamfer $d_{C}$ and Hausdorff distance $d_{H}$ for all 5 shapes are reported." + ], + "table_footnote": [], + "table_body": "
MethoddC↓dH↓
SIREN w on [34]0.427.67
OG-SIREN0.204.06
SIREN (+n) [34]0.193.86
OG-SIREN + H°0.192.96
OG-SIREN + GT°0.172.44
OG-NGLOD0.226.03
OG-NGLOD + H°0.195.50
OG-NGLOD + GT°0.175.12
NGLOD+GTd [35]0.131.97
", + "bbox": [ + 166, + 426, + 380, + 575 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Results on SRB [7] when supervising INRs with varying levels of information (see Section 4.3). $\\mathrm{GT}^d$ means dense GT SDF supervision, while $\\mathrm{GT}^o$ and $\\mathrm{H}^o$ mean GT and human estimated occupancy per octree node.", + "bbox": [ + 75, + 587, + 470, + 643 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "be inside). The IoU of the octree is much lower than the IoU of the final mesh, indicating that the INR training is learning beyond the octree supervision. Furthermore as the depth increases, the octree IoU improves and so do the metrics on the final mesh, indicating that the INR training is heavily affected by the initial training guidance from the octree. See the supplemental material for a depth ablation on SRB.", + "bbox": [ + 75, + 670, + 468, + 777 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Modifying the octree", + "text_level": 1, + "bbox": [ + 76, + 786, + 272, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To show the versatility of our method, in Table 2 we demonstrate how modifying the octree labelling can affect the performance of the resulting shape. For $\\mathrm{OG - INR + GT^o}$ we set octree labels using GT occupancy. For $\\mathrm{OG - INR + H^o}$ we set octree labels using sparse human supervision: we change up to 30 large octree leaves at the lower depths (up", + "bbox": [ + 75, + 809, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "to depth 5) where the changes are obvious from a human perspective. With both octree supervision signals we get increases in performance, showing that the discrete nature of octree guidance does not introduce much bias.", + "bbox": [ + 496, + 429, + 893, + 488 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also compare to the INRs with other forms of supervision. As NGLOD cannot be used in an unsupervised setting, we provide result when trained on dense GT.", + "bbox": [ + 496, + 489, + 893, + 535 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Further Comparisons", + "text_level": 1, + "bbox": [ + 498, + 544, + 705, + 561 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The need for unoriented methods. We compare to a recent industry-standard implementation [48] of normal estimation + SPSR [19] (SPSR+N Est.). Table 1 (left) shows that for the well sampled surfaces in ShapeNet, when ground truth normals are available SPSR achieves close to perfect IoUs. However when normals need to be estimated, it has good median performance but very poor average performance, indicating the normal estimation procedure will often be quite inaccurate. Table 1 (right) shows that for point clouds with large missing regions, SPSR with ground truth normals has room for improvement, while SPSR with estimated normals performs extremely poorly.", + "bbox": [ + 496, + 568, + 892, + 750 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "While neither dataset has provided line of sight information (as would be given by sensors), a more informative use of such information than estimating rough normals would be to use it to label parts of our octree. See Section 4.3 for how additional octree supervision improves performance.", + "bbox": [ + 496, + 750, + 893, + 825 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Overall Timing Comparison. Table 3 shows the time taken for each method. Our method's gradient descent training is the fastest to converge as it needs much fewer iterations. Even when adding on the average octree building time from ShapeNet, we can see that the time taken for both", + "bbox": [ + 496, + 825, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "16723", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6ca24abf1e208c0fd895580adc7478bffe3cfeaa04815fad900f335600ede9c1.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodParametersTime per iter. (s)Num itersTime (s)Speed Up
N Est. +SPSR---42 (13 + 29)12× (· + 18×)
SIREN (wo n)264K0.052100005201×
SAL2.1M0.1751000017500.3×
DiGS264K0.1201000012000.4×
SAP120K-32003301.6×
Our OG-SIREN264K0.158600135 (40 + 95)3.9× (· + 5.5×)
Our OG-NGLOD68.7M0.17330092 (40 + 52)5.7× (· + 10×)
", + "bbox": [ + 171, + 88, + 802, + 210 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/ec1f68ec8677136d33895e116f2ed17b8b3166a5e5cc666623c7059a777613b7.jpg", + "table_caption": [ + "Table 3. Time and number of parameters for each model on ShapeNet. For methods with a precomputation stage (normal estimation or octree building), the time for both is included with the precomputation and training times in parenthesis. Speed up is measured with respect to SIREN/SIREN w o n. We ran our experiments on a single Nvidia RTX 2080 GPU and a i7-8700K CPU." + ], + "table_footnote": [], + "table_body": "
DepthMeanOctree IoU ↑ MedianStdMeanFinal Mesh (OG-SIREN)Squared Chamfer ↓Octree Build + Label Time (s)
IoU ↑ MedianStdMeanMedianStdMeanMedianStd
50.47480.47020.18540.71010.77120.25941.84e-41.35e-42.23e-42.21.62.2
60.64010.64100.16040.95030.98350.11723.93e-52.34e-58.20e-57.65.38.6
70.77570.80060.13050.96150.98710.10483.75e-52.17e-58.24e-538.119.367.7
", + "bbox": [ + 124, + 276, + 844, + 371 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Octree depth ablation on ShapeNet. Octree IoU is computed using the labelling of the leaves (surface leaves chosen as inside).", + "bbox": [ + 84, + 381, + 880, + 396 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "stages of our method is still faster than other methods, often by an order of magnitude. We implemented the octree in Cython [3] but further optimisations are possible.", + "bbox": [ + 75, + 422, + 468, + 467 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Comparison Against No Octree Guidance. Comparing SIREN wo n with and without the octree guidance we can see that the guidance drastically improves both time performance and reconstruction metrics (see Table 1,Table 3). When initially guided by our octree labelling, SIREN (wo n) converges significantly faster (92s vs 520s) and gets a 3x improvement in ShapeNet IoU. Without guidance from normals or our octree, SIREN wo n gets stuck in bad local minima where it needs to flip its inside-outsideness prediction, causing ghost geometry (see Figure 1). When comparing the two types of INR training guidance applied to SIREN wo n, ground truth normals vs our octree, we can see that the octree guidance still performs much better than SIREN on ShapeNet, while performing comparably on SRB.", + "bbox": [ + 75, + 469, + 470, + 681 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Discussion", + "text_level": 1, + "bbox": [ + 76, + 693, + 194, + 708 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. Like all unoriented methods, the octree guidance has difficulty with large missing regions, which is demonstrated heavily in SRB, and thin surfaces, which appear in SRB (Daratech) and ShapeNet (airplane and watercraft). Despite this, our methods perform competitively on shapes with these difficulties on both datasets. See the supplemental for per class breakdown and visualizations.", + "bbox": [ + 75, + 717, + 468, + 824 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Other possible limitations would be the ability to deal with noise and outliers. Note that all methods do poorly with large amounts of either corruption. For small noise the method's performance does not degrade considerably as demonstrated by the results on SRB. For larger noise, the", + "bbox": [ + 75, + 824, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "volume occupied by surface leaves would be much larger. While this would not introduce error, the octree's benefits would be reduced. For outliers there would be significant bias introduced by the octree method, which if sparse could be preprocessed out beforehand.", + "bbox": [ + 496, + 422, + 890, + 497 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Broader Impact. Faster and more accurate unoriented surface reconstruction methods will allow for accurate meshes to be constructed easily from regular scanners. Potential misuses of this technology include unauthorized replication and harmful content creation.", + "bbox": [ + 496, + 500, + 890, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 594, + 617, + 609 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have introduced an octree guided method for training implicit neural representations (INRs) for the task of unoriented surface reconstruction. We formulate the main difficulty of the task, inside-outside determinism, as an octree labelling problem with respect to an energy function that captures watertight surface properties and the minimal surface constraint. We also detail an efficient method to minimize the energy. We then use this supervision to guide the initial stage of INR training, using loss functions that can be used by any INR architecture. Our results show that applying our octree guided supervision provides significant time and reconstruction performance improvements, with competitive or better results than the current state-of-the-art.", + "bbox": [ + 496, + 621, + 890, + 818 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This work has received funding from the European Union's Horizon 2020 research and innovation programme under the Marie Sklodowska-Curie grant agreement No. 893465. S. Gould is a recipient of an ARC Future Fellowship (proj. no. LP200100421) funded by the Australian Government.", + "bbox": [ + 496, + 825, + 893, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "16724", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Matan Atzmon and Yaron Lipman. SAL: Sign Agnostic Learning of shapes from raw data. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2565-2574, 2020. 2, 3, 5, 6, 7", + "[2] Matan Atzmon and Yaron Lipman. SALD: Sign Agnostic Learning with Derivatives. In Proc. of the International Conference on Learning Representations (ICLR), 2021. 2, 3, 5", + "[3] Stefan Behnel, Robert Bradshaw, Craig Citro, Lisandro Dalcin, Dag Sverre Seljebotn, and Kurt Smith. Cython: The best of both worlds. Computing in Science & Engineering, 13(2):31-39, 2011. 8", + "[4] Yizhak Ben-Shabat and Stephen Gould. DeepFit: 3D surface fitting via neural network weighted least squares. In Proc. of the European Conference on Computer Vision (ECCV), pages 20-34. Springer, 2020. 1", + "[5] Yizhak Ben-Shabat, Chamin Hewa Koneputugodage, and Stephen Gould. Digs: Divergence guided shape implicit neural representation for unoriented point clouds. arXiv preprint arXiv:2106.10811, 2021. 2, 3, 5, 6, 7", + "[6] Yizhak Ben-Shabat, Michael Lindenbaum, and Anath Fischer. Nesti-net: Normal estimation for unstructured 3D point clouds using convolutional neural networks. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 10112-10120, 2019. 1", + "[7] Matthew Berger, Joshua A Levine, Luis Gustavo Nonato, Gabriel Taubin, and Claudio T Silva. A benchmark for surface reconstruction. ACM Trans. on Graphics (ToG), 32:1-17, 2013. 6, 7", + "[8] Julian Besag. On the statistical analysis of dirty pictures. Journal of the Royal Statistical Society: Series B (Methodological), 48(3):259-279, 1986. 3", + "[9] Andrew Blake, Pushmeet Kohli, and Carsten Rother. Markov Random Fields for Vision and Image Processing. The MIT Press, 07 2011. 3, 5", + "[10] Yuri Boykov, Olga Veksler, and Ramin Zabih. Fast approximate energy minimization via graph cuts. IEEE Transactions on pattern analysis and machine intelligence, 23(11):1222-1239, 2001. 3", + "[11] Jonathan C Carr, Richard K Beatson, Jon B Cherrie, Tim J Mitchell, W Richard Fright, Bruce C McCallum, and Tim R Evans. Reconstruction and representation of 3D objects with radial basis functions. In Proc. of the Conference on Computer Graphics and Interactive Techniques, pages 67-76, 2001. 7", + "[12] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. ShapeNet: An information-rich 3D model repository. arXiv preprint arXiv:1512.03012, 2015. 6, 7", + "[13] Stephen Gould. Max-margin learning for lower linear envelope potentials in binary markov random fields. In Proc. of the International Conference on Machine Learning (ICML). Omnipress, 2011. 5", + "[14] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit Geometric Regularization for learning shapes. In Proc. of the International Conference on" + ], + "bbox": [ + 76, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Machine Learning (ICML), volume 119, pages 3789-3799. \nPMLR, 2020. 2, 3, 5, 7", + "[15] Paul Guerrero, Yanir Kleiman, Maks Ovsjanikov, and Niloy J Mitra. PCPNet: Learning local shape properties from raw point clouds. In Computer Graphics Forum, volume 37, pages 75-85. Wiley Online Library, 2018. 1", + "[16] Christian Hane, Shubham Tulsiani, and Jitendra Malik. Hierarchical surface prediction for 3d object reconstruction. In 2017 International Conference on 3D Vision (3DV), pages 412-420. IEEE, 2017. 2", + "[17] Shahram Izadi, David Kim, Otmar Hilliges, David Molyneaux, Richard Newcombe, Pushmeet Kohli, Jamie Shotton, Steve Hodges, Dustin Freeman, Andrew Davison, and Andrew Fitzgibbon. Kinectfusion: Real-time 3d reconstruction and interaction using a moving depth camera. In UIST '11 Proceedings of the 24th annual ACM symposium on User interface software and technology, pages 559-568. ACM, October 2011. 1", + "[18] Michael Kazhdan, Matthew Bolitho, and Hugues Hoppe. Poisson surface reconstruction. In Proc. of the Eurographics Symposium on Geometry Processing (SGP), volume 7, 2006. 2", + "[19] Michael Kazhdan and Hugues Hoppe. Screened poisson surface reconstruction. ACM Trans. on Graphics (ToG), 32:1-13, 2013. 2, 7", + "[20] Pushmeet Kohli, M. Kumar, and Philip Torr. P3 & beyond: Solving energies with higher order cliques. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2007. 5", + "[21] Jan Eric Lenssen, Christian Osendorfer, and Jonathan Masci. Deep iterative surface normal estimation. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 11247-11256, 2020. 1", + "[22] Yaron Lipman. Phase transitions, distance functions, and implicit neural representations. arXiv preprint arXiv:2106.07689, 2021. 2, 3, 7", + "[23] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. Advances in Neural Information Processing Systems, 33:15651-15663, 2020. 2", + "[24] Shi-Lin Liu, Hao-Xiang Guo, Hao Pan, Peng-Shuai Wang, Xin Tong, and Yang Liu. Deep implicit moving least-squares functions for 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1788-1797, 2021. 3", + "[25] William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3D surface construction algorithm. In Proc. of the Intl. Conference on Computer Graphics and Interactive Techniques (SIGGRAPH), volume 21, pages 163-169. ACM New York, NY, USA, 1987. 6", + "[26] Julien NP Martel, David B Lindell, Connor Z Lin, Eric R Chan, Marco Monteiro, and Gordon Wetzstein. Acorn: Adaptive coordinate networks for neural scene representation. arXiv preprint arXiv:2105.02788, 2021. 3", + "[27] Donald J. Meagher. Geometric modeling using octree encoding. Comput. Graph. Image Process., 19:85, 1982. 2", + "[28] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks:" + ], + "bbox": [ + 501, + 92, + 890, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "16725", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Learning 3D reconstruction in function space. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4460-4470, 2019. 1", + "[29] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, July 2022. 1, 2", + "[30] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. DeepSDF: Learning continuous signed distance functions for shape representation. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 165-174, 2019. 1, 2", + "[31] Songyou Peng, Chiyu Jiang, Yiyi Liao, Michael Niemeyer, Marc Pollefeys, and Andreas Geiger. Shape As Points: A differentiable poisson solver. arXiv preprint arXiv:2106.03452, 2021. 3, 6, 7", + "[32] Gernot Riegler, Ali Osman Ulusoy, and Andreas Geiger. Octnet: Learning deep 3d representations at high resolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3577-3586, 2017. 2", + "[33] Bernhard Scholkopf, Joachim Giesen, and Simon Spalinger. Kernel methods for implicit surface modeling. In Advances in Neural Information Processing Systems (NeurIPS), NIPS'04, page 1193-1200, Cambridge, MA, USA, 2004. MIT Press. 7", + "[34] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. Advances in Neural Information Processing Systems (NeurIPS), 33, 2020. 2, 5, 6, 7", + "[35] Towaki Takikawa, Joey Litalien, Kangxue Yin, Karsten Kreis, Charles Loop, Derek Nowrouzezahrai, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Neural geometric level of detail: Real-time rendering with implicit 3D shapes. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 2, 3, 6, 7", + "[36] Matthew Tancik, Pratul P. Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan T. Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. Advances in Neural Information Processing Systems (NeurIPS), 2020. 2, 7", + "[37] Jia-Heng Tang, Weikai Chen, Jie Yang, Bo Wang, Songrun Liu, Bo Yang, and Lin Gao. Octfield: Hierarchical implicit functions for 3d modeling. In The Thirty-Fifth Annual Conference on Neural Information Processing Systems (NeurIPS), 2021. 2", + "[38] Maxim Tatarchenko, Alexey Dosovitskiy, and Thomas Brox. Octree generating networks: Efficient convolutional architectures for high-resolution 3d outputs. In Proceedings of the IEEE international conference on computer vision, pages 2088-2096, 2017. 2", + "[39] Ayush Tewari, Justus Thies, Ben Mildenhall, Pratul Srinivasan, Edgar Tretschk, Yifan Wang, Christoph Lassner, Vincent Sitzmann, Ricardo Martin-Brualla, Stephen Lombardi, et al. Advances in neural rendering. arXiv preprint arXiv:2111.05849, 2021. 1" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Peng-Shuai Wang, Yang Liu, Yu-Xiao Guo, Chun-Yu Sun, and Xin Tong. O-cnn: Octree-based convolutional neural networks for 3d shape analysis. ACM Transactions On Graphics (TOG), 36(4):1-11, 2017. 2", + "[41] Peng-Shuai Wang, Yang Liu, and Xin Tong. Deep octree-based cnns with output-guided skip connections for 3d shape and scene completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 266-267, 2020. 2", + "[42] Peng-Shuai Wang, Yang Liu, and Xin Tong. Dual octree graph networks for learning adaptive volumetric shape representations. ACM Transactions on Graphics (TOG), 41(4):1-15, 2022. 3", + "[43] Peng-Shuai Wang, Chun-Yu Sun, Yang Liu, and Xin Tong. Adaptive o-cnn: A patch-based deep representation of 3d shapes. ACM Transactions on Graphics (TOG), 37(6):1-11, 2018. 2", + "[44] Francis Williams, Teseo Schneider, Claudio Silva, Denis Zorin, Joan Bruna, and Daniele Panozzo. Deep Geometric Prior for surface reconstruction. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 10130-10139, 2019. 7", + "[45] Francis Williams, Matthew Trager, Joan Bruna, and Denis Zorin. Neural Splines: Fitting 3D surfaces with infinitely-wide neural networks. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 9949-9958, 2021. 2, 6, 7", + "[46] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond. Computer Graphics Forum, 2022. 1", + "[47] Wang Yifan, Lukas Rahmann, and Olga Sorkine-Hornung. Geometry-consistent neural shape representation with implicit displacement fields. arXiv preprint arXiv:2106.05187, 2021. 3", + "[48] Qian-Yi Zhou, Jaesik Park, and Vladlen Koltun. Open3D: A modern library for 3D data processing. arXiv:1801.09847, 2018. 7" + ], + "bbox": [ + 501, + 92, + 890, + 640 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "16726", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Octree Guided Unoriented Surface Reconstruction/0db48127-a127-468b-be90-b27ec5813046_model.json b/2023/Octree Guided Unoriented Surface Reconstruction/0db48127-a127-468b-be90-b27ec5813046_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b0658d2702bebab5c63a884add5cfb48e5face46 --- /dev/null +++ b/2023/Octree Guided Unoriented Surface Reconstruction/0db48127-a127-468b-be90-b27ec5813046_model.json @@ -0,0 +1,2574 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.229, + 0.131, + 0.743, + 0.151 + ], + "angle": 0, + "content": "Octree Guided Unoriented Surface Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.163, + 0.181, + 0.424, + 0.201 + ], + "angle": 0, + "content": "Chamin Hewa Koneputugodage1" + }, + { + "type": "text", + "bbox": [ + 0.466, + 0.181, + 0.639, + 0.198 + ], + "angle": 0, + "content": "Yizhak Ben-Shabat1,2" + }, + { + "type": "text", + "bbox": [ + 0.678, + 0.181, + 0.806, + 0.199 + ], + "angle": 0, + "content": "Stephen Gould" + }, + { + "type": "text", + "bbox": [ + 0.174, + 0.209, + 0.461, + 0.227 + ], + "angle": 0, + "content": "1The Australian National University" + }, + { + "type": "text", + "bbox": [ + 0.481, + 0.209, + 0.802, + 0.228 + ], + "angle": 0, + "content": "\\(^{2}\\)Technion Israel Institute of Technology" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.23, + 0.741, + 0.246 + ], + "angle": 0, + "content": "{chamin.hewa, stephen.gould}@anu.edu.au sitszikbs@gmail.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.28, + 0.314, + 0.296 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.314, + 0.474, + 0.678 + ], + "angle": 0, + "content": "We address the problem of surface reconstruction from unoriented point clouds. Implicit neural representations (INRs) have become popular for this task, but when information relating to the inside versus outside of a shape is not available (such as shape occupancy, signed distances or surface normal orientation) optimization relies on heuristics and regularizers to recover the surface. These methods can be slow to converge and easily get stuck in local minima. We propose a two-step approach, OG-INR, where we (1) construct a discrete octree and label what is inside and outside (2) optimize for a continuous and high-fidelity shape using an INR that is initially guided by the octree's labelling. To solve for our labelling, we propose an energy function over the discrete structure and provide an efficient move-making algorithm that explores many possible labellings. Furthermore we show that we can easily inject knowledge into the discrete octree, providing a simple way to influence the result from the continuous INR. We evaluate the effectiveness of our approach on two unoriented surface reconstruction datasets and show competitive performance compared to other unoriented, and some oriented, methods. Our results show that the exploration by the move-making algorithm avoids many of the bad local minima reached by purely gradient descent optimized methods (see Figure 1)." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.708, + 0.21, + 0.724 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Surface reconstruction from 3D point clouds has been studied extensively in computer vision and computer graphics. The task requires estimating a mesh of the shape's surface from a point cloud sampled from the surface of the shape. We focus on the reconstruction of watertight 3D shapes, i.e., shapes that have a well defined interior and exterior. Such shapes are often represented as signed distance fields (SDFs) or occupancy fields, which can be efficiently encoded by a neural network [28, 30]. As these representations are fields parameterized by neural networks, they are often called neural fields. Furthermore, they implicitly rep" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.278, + 0.896, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.41, + 0.895, + 0.508 + ], + "angle": 0, + "content": "Figure 1. Our method with a SIREN INR (OG-SIREN, left) compared to SIREN wo n (right) for two shapes from the ShapeNet dataset. Our octree guidance allows for consistent inside-outside determinism. On the other hand, SIREN wo n gets stuck in local minima from which it cannot escape (due to needing to completely change the occupancy of certain areas) creating extraneous surfaces (often called ghost geometries in the literature)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.537, + 0.892, + 0.567 + ], + "angle": 0, + "content": "resent the shape by a level set of the field, thus they are also often referred to as implicit neural representations (INRs)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.568, + 0.892, + 0.689 + ], + "angle": 0, + "content": "The broader class of neural fields, including INRs, have been very popular over the last few years as they can handle arbitrary topology, are memory efficient, and are continuous with potentially infinite resolution [39, 46]. Among the INRs for 3D shapes, SDFs are the most popular as they provide more useful information (distances not just occupancy) and are required for downstream graphics algorithms such as sphere tracing and approximate soft shadows [29, 35]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.895, + 0.902 + ], + "angle": 0, + "content": "When learning an implicit representation of a shape, a major difficulty is predicting whether points in space are inside or outside the shape. Many methods require oriented surface normals for the input points or signed distances from the surface, which usually are not given by raw data from scans. While this can be estimated using the line of sight information [17] or algorithms [4,6, 15, 21], they yield noisy predictions that after postprocessing still can lead to bad results (see Section 4.4). We consider the task of unoriented surface reconstruction, where such information is not available, and only the sampled surface points are given. We demonstrate that our method performs competitively and sometimes better than oriented methods, even when they are given the ground truth (GT) normals." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16717" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.096, + 0.885, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.257, + 0.895, + 0.314 + ], + "angle": 0, + "content": "Figure 2. An illustration of our method, OG-INR. Given an unoriented point cloud (left), we progressively build and label an octree around the points (middle). The octree at depths 3-7 are shown. Surface leaves of the octree (yellow) are leaves that contain points from the point cloud, other leaves are labelled as inside (blue) or outside (transparent) by minimizing an energy function. We then train an INR model to obtain an SDF, using the labelling as supervision for the initial training, after which we can extract a mesh (right)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.34, + 0.473, + 0.552 + ], + "angle": 0, + "content": "We propose OG-INR, which uses a discrete representation in conjunction with the continuous INR representation (see Figure 2). Given an input unoriented point cloud, we progressively build an octree from the input points and determine which leaf nodes are surface leaves (i.e., leaves that contain a point). We also label all other leaves as inside or outside leaves (if they should be within or outside the shape respectively). To do this we minimize an energy function that trades off the watertight surface property, that every surface point should border with the inside and outside of the shape, with a minimal surface constraint. We then train standard INR architectures, initially guiding the training by the octree labelling, and show that the INR converges much faster and is less prone to large failure regions." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.553, + 0.283, + 0.566 + ], + "angle": 0, + "content": "Our main contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.568, + 0.469, + 0.627 + ], + "angle": 0, + "content": "- We introduce a novel method to guide the initial stage of INR training. It uses a labelled octree structure to allow the INR to converge significantly faster and alleviates the local minima problem of INRs." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.63, + 0.469, + 0.689 + ], + "angle": 0, + "content": "- We propose an energy function over octree labels that captures the task of surface reconstruction. It balances the constraint of maintaining known surface regions with minimising the overall surface area." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.693, + 0.469, + 0.737 + ], + "angle": 0, + "content": "- We provide an efficient move-making algorithm to optimize the energy function, which explores many inside-outside possibilities in a structured manner." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.74, + 0.469, + 0.785 + ], + "angle": 0, + "content": "- Our discrete representation is easily understandable and human-modifiable, giving an intuitive method for applying changes to the resulting SDF." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.568, + 0.469, + 0.785 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.8, + 0.22, + 0.815 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Unsupervised Surface Reconstruction. Representing shapes with INRs was popularized by DeepSDF [30], where a ReLU network directly models an SDF using GT SDF supervision. SAL [1] instead uses unsigned distance supervision, approximated from the input point cloud, as an" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.34, + 0.892, + 0.461 + ], + "angle": 0, + "content": "unsupervised and unoriented approach. SALD [2] showed that adding unoriented normal supervision improved upon this, and IGR [14] further added the eikonal constraint loss (see Equation 5) that has been widely adopted [5, 22, 34]. NSP [45] addresses the task by solving a kernel regression problem but needs input normals. PHASE [22] introduces a novel loss inspired by the phase transition problem that encourages boundary crossings at the surface points." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.462, + 0.893, + 0.627 + ], + "angle": 0, + "content": "However, as the inputs to INRs are low dimensional, they are biased to low frequency solutions that do not recover high fidelity detail [36]. To introduce high frequencies, SIREN [34] uses sine activations and Fourier Feature Networks [36] use a fixed Fourier encoding. While being able to reconstruct high fidelity shapes with a modest number of parameters, these methods are hard to train as the high frequencies overfit quickly. A different approach is parametric encoding [29, 35], where spatial regions are represented by trainable parameters (often using an octree structure). We cover such methods in the next subsection." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.628, + 0.892, + 0.657 + ], + "angle": 0, + "content": "Note that IGR and SIREN can operate without normals, which we refer to by \"IGR wo n\" and \"SIREN wo n\"." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.892, + 0.748 + ], + "angle": 0, + "content": "Octrees for Surface Reconstruction. Octrees were first introduced by Meagher [27] as an efficient data structure for representing and manipulating 3D data, and are now used extensively. The classic method Poisson Surface Reconstruction (PSR) [18] and its followup SPSR [19] use octrees to structure their implicit functions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.856 + ], + "angle": 0, + "content": "Early works on using octrees with neural networks include HSP [16] and OGN [38], who train encoder-decoder networks to learn octrees with occupancy values, and OctNet [32] and O-CNN [40], who implement 3D CNNs that operate over octrees rather than voxel grids for shape analysis tasks. O-CNN's descendants [41, 43] build on this and adapt it to shape generation, completion and reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "More recently octrees have been used in INRs for representing fine details efficiently. OctField [37] and NSVF [23] both learn local implicit functions for individual octree" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16718" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.47, + 0.168 + ], + "angle": 0, + "content": "nodes. NGLOD [35] and ACORN [26] instead learn a single INR, but store trainable parameters at octree node vertices to encode spatial regions. IMLSNet [24] uses O-CNN to predict both an octree and oriented points per octree node to represent the surface using implicit moving least squares." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.17, + 0.468, + 0.275 + ], + "angle": 0, + "content": "However, previous octree reconstruction methods were supervised, trained on GT SDF/occupancy of either the current shape or a separate training set. Recently Dual Octree Graph CNNs [42] addressed the unsupervised setting, where they learn local features on the dual graph of an octree. However their method has parameters that need unsupervised training, and they require input normals." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.277, + 0.468, + 0.398 + ], + "angle": 0, + "content": "SAP [31] truly addresses the unsupervised and unoriented task by making the PSR pipeline differentiable. Starting from an oriented point cloud of a sphere, they use Chamfer loss to guide that point cloud to a dense, oriented representation. Their method is fast and among state-of-the-art, however it can fail drastically and often does not produce a watertight mesh (see Section 4). It is also hard to adapt to different applications and losses, unlike INRs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.4, + 0.468, + 0.567 + ], + "angle": 0, + "content": "Initialization for INRs. Due to the difficulty of training INRs, initialization is very important [1, 5]. However regardless of input, most methods initialize to a sphere. These include SAL [1] and its descendants [2, 14, 22], DiGS [5] and IDF [47]. This gives many important properties, such as biasing shape interior to the domain's center and satisfying the eikonal condition at the start of training, but it is problematic for shapes with sharp concavities. Our method, on the other hand, provides firm guidance for both the eikonal constraint and the shape based on the input point cloud, and thus does not bias the training toward convex objects." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.569, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Energy Minimization. Discrete energy minimization has been used extensively in both 2D and 3D vision for segmentation and object labelling [9, 10]. An energy function encodes our preference over the joint assignment to a set of variables, e.g., binary variables representing occupancy of voxels in 3D space. The goal of energy minimization is to find an assignment to the variables with minimal cost. This is NP-hard in general and a number of approximation algorithms have been proposed [9]. One class of algorithms, known as move-making, starts with an initial assignment and iteratively evaluates changes to a subset of the variables (called a move). If the change results in lower energy then the move is accepted, otherwise the changed values are reverted. Moves are repeatedly tried until no further reduction in energy is found. Iterated Conditional Modes (ICM) [8] is the simplest example, where each move involves changing a single variable and so is very efficient, but it is prone to getting stuck in local minima. Generalized ICM improves on this by allowing multiple variables to change in a move, but it is very costly. Our method is a variant of generalized ICM specialized to our problem, where we introduce restrictions on the move space improve efficiency." + }, + { + "type": "code_caption", + "bbox": [ + 0.501, + 0.091, + 0.811, + 0.107 + ], + "angle": 0, + "content": "Algorithm 1 Overall octree labeling algorithm" + }, + { + "type": "algorithm", + "bbox": [ + 0.507, + 0.109, + 0.892, + 0.326 + ], + "angle": 0, + "content": "1: procedure LABELOCTREE \\((\\chi ,d_i,k,C)\\) \n2: \\(\\mathcal{O} =\\) CreateOctree \\((\\chi ,d_{i})\\) \n3: Initialization(O) \n4: for \\(j\\) in \\(0,\\dots ,k\\) do \n5: if \\(j\\neq 0\\) then \n6: ExpandTree(O) \n7: end if ▷ Now at depth \\(d = d_{i} + j\\) \n8: Grow(O) \n9: for \\(s\\) in \\(C\\) do \n10: makeMove(O,s) \n11: end for \n12: end for \n13: return O \n14: end procedure" + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.356, + 0.836, + 0.374 + ], + "angle": 0, + "content": "3. OG-INR: Octree Guided Shape INRs" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.384, + 0.704, + 0.399 + ], + "angle": 0, + "content": "3.1. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.409, + 0.892, + 0.501 + ], + "angle": 0, + "content": "Let \\(\\chi = \\{x_{i}\\}_{i\\in I}\\) be a point cloud sampled from a surface \\(S = \\partial \\mathcal{V}\\) of some watertight volume \\(\\mathcal{V} \\subset \\mathcal{D}\\) where \\(\\mathcal{D} \\subset \\mathbb{R}^3\\) is compact. We wish to reconstruct the surface \\(S\\). We choose to represent \\(S\\) as the zero-level set of a signed distance function (SDF), and thus wish to train a neural network \\(\\Phi\\) to output signed distances for all points \\(x \\in \\mathcal{D}\\)," + }, + { + "type": "equation", + "bbox": [ + 0.604, + 0.516, + 0.891, + 0.535 + ], + "angle": 0, + "content": "\\[\n\\Phi (x, \\theta) = (- 1) ^ {o (x)} d (x, \\mathcal {S}) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.551, + 0.892, + 0.596 + ], + "angle": 0, + "content": "where \\(\\theta\\) are the model parameters, \\(o(x)\\) is the occupancy of point x (1 if inside and 0 if outside) and \\(d(x,S)\\) is the distance from \\(x\\) to its closest point on \\(S\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.613, + 0.61, + 0.628 + ], + "angle": 0, + "content": "3.2. Overview" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.639, + 0.892, + 0.76 + ], + "angle": 0, + "content": "We first aim to produce a rough and discrete, but globally consistent, occupancy determination for our domain by solving for an inside-outside labelling on nodes of an octree. To do this we propose an energy function that combines both the watertight surface property and the minimal surface objective. Despite this energy being intractable to solve exactly, we provide a domain-guided move-making algorithm to get low energy solutions (see Section 3.3)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.763, + 0.892, + 0.853 + ], + "angle": 0, + "content": "We then design additional INR loss terms to guide INRs using our octree labelling (see Section 3.4). These losses quickly guide the INR towards a rough but globally consistent representation that allows it to avoid many bad local minima. Since these loss terms are agnostic to INR architecture, our method can work with any INR model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "As we assume our domain is compact, we first normalize the point cloud to fit within the unit sphere in 3D, and then set our domain as \\(\\mathcal{D} = [-1.1, 1.1]^3\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16719" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.16, + 0.089, + 0.319, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.162, + 0.198, + 0.307, + 0.208 + ], + "angle": 0, + "content": "A: Depth 3, After Initialization(O)" + }, + { + "type": "image", + "bbox": [ + 0.321, + 0.089, + 0.48, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.325, + 0.199, + 0.437, + 0.209 + ], + "angle": 0, + "content": "B: Depth 3, After Grow(O)" + }, + { + "type": "image", + "bbox": [ + 0.481, + 0.089, + 0.641, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.485, + 0.199, + 0.639, + 0.209 + ], + "angle": 0, + "content": "C: Depth 3, After all makeMove(O,s)" + }, + { + "type": "image", + "bbox": [ + 0.651, + 0.089, + 0.811, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.653, + 0.199, + 0.807, + 0.209 + ], + "angle": 0, + "content": "D: Depth 4, After all makeMove(O, s)" + }, + { + "type": "image", + "bbox": [ + 0.16, + 0.216, + 0.32, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.164, + 0.326, + 0.274, + 0.336 + ], + "angle": 0, + "content": "E: Depth 5, After Grow(O)" + }, + { + "type": "image", + "bbox": [ + 0.321, + 0.216, + 0.48, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.325, + 0.326, + 0.465, + 0.336 + ], + "angle": 0, + "content": "F: Depth 5, After makeMove(O,1)" + }, + { + "type": "image", + "bbox": [ + 0.481, + 0.216, + 0.64, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.485, + 0.326, + 0.638, + 0.336 + ], + "angle": 0, + "content": "G: Depth 5, After all makeMove(O,s)" + }, + { + "type": "image", + "bbox": [ + 0.651, + 0.216, + 0.811, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.653, + 0.326, + 0.807, + 0.336 + ], + "angle": 0, + "content": "H: Depth 6, After all makeMove(O,s)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.354, + 0.892, + 0.384 + ], + "angle": 0, + "content": "Figure 3. Our algorithm demonstrated on a 2D quadtree example. The black dots are our input point cloud, the yellow squares are surface leaves, and the blue and gray squares are non-surface leaves labelled as inside and outside respectively." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.409, + 0.382, + 0.425 + ], + "angle": 0, + "content": "3.3. Octree Construction and Labelling" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.433, + 0.47, + 0.538 + ], + "angle": 0, + "content": "Setup. We present our octree labelling algorithm in Algorithm 1, which takes in point cloud \\(\\chi\\), an initial depth to build the octree to \\(d_{i}\\), a number of further depth expansions \\(k\\) (the final octree has depth \\(d_{f} = d_{i} + k\\)) and a set of move set sizes \\(C\\). We set \\(d = 3\\), \\(k = 4\\) and \\(C = [1,2,10,10000]\\) in our experiments. The algorithm returns a labelled octree \\(\\mathcal{O} = (\\mathcal{N},\\mathcal{L}_{ns},\\mathcal{L}_s,y)\\) where" + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.54, + 0.469, + 0.584 + ], + "angle": 0, + "content": "- \\(\\mathcal{N}\\) are the nodes of the octree. The octree is built by repeatedly expanding nodes containing or near points from \\(\\chi\\), so it is highly unbalanced." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.588, + 0.469, + 0.663 + ], + "angle": 0, + "content": "- \\(\\mathcal{L} \\subset \\mathcal{N}\\) are the leaf nodes, and thus \\(\\mathcal{L}\\) partitions \\(\\mathcal{D}\\). Furthermore, we define surface leaves \\(\\mathcal{L}_s \\subset \\mathcal{L}\\) to be leaves that contain points from the point cloud (colored yellow in Figure 2 and Figure 3) and all other leaves to be non-surface leaves \\(\\mathcal{L}_{ns} = \\mathcal{L} \\setminus \\mathcal{L}_s\\)." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.667, + 0.469, + 0.711 + ], + "angle": 0, + "content": "- \\( y \\) is the binary occupancy label vector indexed by non-surface leaves \\( \\ell \\in \\mathcal{L}_{ns} \\) denoting whether the space defined by each \\( \\ell \\) is inside, \\( y_{\\ell} = 1 \\), or outside, \\( y_{\\ell} = 0 \\)." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.54, + 0.469, + 0.711 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.713, + 0.469, + 0.809 + ], + "angle": 0, + "content": "We will construct our algorithm so that the neighbours of every surface leaf \\(\\ell\\) is at the same depth. Thus let \\(n_{\\ell}^{s}\\) and \\(n_{\\ell}^{ns}\\) be the number of its surface and non-surface neighbours and \\(n_{\\ell}^{(1)}(\\pmb{y})\\) and \\(n_{\\ell}^{(0)}(\\pmb{y})\\) be the number of inside and outside neighbors w.r.t. labelling \\(\\pmb{y}\\). Then \\(n_{\\ell}^{ns} = n_{\\ell}^{(1)}(\\pmb{y}) + n_{\\ell}^{(0)}(\\pmb{y})\\) and \\(n_{\\ell}^{s} + n_{\\ell}^{s} = 26\\) (or 8 for our 2D example in Figure 3)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.469, + 0.87 + ], + "angle": 0, + "content": "We will constantly refer to the boundary as the leaves touching the boundary of \\(\\mathcal{D}\\) and the inside-outside border to be the non-surface leaves who have a non-surface neighbor with a different label (different occupancy value)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.469, + 0.902 + ], + "angle": 0, + "content": "We determine inside-outside occupancy based on three properties: (1) surface property: surfaces should separate" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.41, + 0.892, + 0.468 + ], + "angle": 0, + "content": "inside and outside regions, (2) minimal surface constraint: the total surface area of the shape's surface \\(S\\) should be as small as possible, (3) boundary constraint: the boundary of the domain \\(\\mathcal{D}\\) should be outside." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.47, + 0.892, + 0.635 + ], + "angle": 0, + "content": "Energy function. We propose an energy function over labellings \\( \\mathbf{y} = \\{y \\mid y_i \\in \\{0,1\\}, i \\in \\mathcal{L}_{ns}\\} \\) based on the first two criteria, the third is hard-coded at the start of the algorithm. For the surface property, ideally every surface leaf would have both inside and outside neighbours, however it may only have surface leaf neighbours which contain both inside and outside regions by definition (see Figure 3 C,D,G,H, some surface leafs have no inside neighbours). For the minimal surface constraint, we only create surfaces when adjacent non-surface leaves have different labellings (note surface leaves are fixed), so we want to minimise this." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.636, + 0.892, + 0.747 + ], + "angle": 0, + "content": "For the surface property, we want every surface leaf \\(\\ell\\) to have some amount of both inside and outside neighbours, where surface neighbours partially count towards both. Thus, we aim for \\(\\gamma_{\\ell}^{(1)} = \\gamma^{(1)} - \\eta^{(1)}n_{\\ell}^{s}\\) non-surface inside neighbors and \\(\\gamma_{\\ell}^{(0)} = \\gamma^{(0)} - \\eta^{(0)}n_{\\ell}^{s}\\) non-surface outside neighbors where \\(\\gamma^{(\\cdot)},\\eta^{(\\cdot)}\\) are hyperparameters. Then the energy cost for a surface leaf \\(\\ell\\) is" + }, + { + "type": "equation", + "bbox": [ + 0.504, + 0.752, + 0.892, + 0.778 + ], + "angle": 0, + "content": "\\[\nE _ {\\ell} ^ {S P} (\\boldsymbol {y}) = \\max \\left\\{\\gamma_ {\\ell} ^ {(1)} - n _ {\\ell} ^ {(1)} (\\boldsymbol {y}), \\gamma_ {\\ell} ^ {(0)} - n _ {\\ell} ^ {(0)} (\\boldsymbol {y}), 0 \\right\\}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.782, + 0.892, + 0.842 + ], + "angle": 0, + "content": "For the minimal surface constraint, we want to minimise the total surface area between every pair of non-surface neighbors with different labellings. Then for two non-surface leaves \\(\\ell, \\ell'\\), our energy cost is" + }, + { + "type": "equation", + "bbox": [ + 0.606, + 0.847, + 0.892, + 0.866 + ], + "angle": 0, + "content": "\\[\nE _ {\\ell , \\ell^ {\\prime}} ^ {M S C} (\\boldsymbol {y}) = a _ {\\ell \\ell^ {\\prime}} \\left| y _ {\\ell} - y _ {\\ell^ {\\prime}} \\right| \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "where \\(a_{\\ell \\ell^{\\prime}}\\) is the (surface) area in common between them (note \\(a_{ij} = 0\\) if they are not adjacent)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16720" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.123 + ], + "angle": 0, + "content": "The two criteria can be formalized as an energy function over labellings \\( \\mathbf{y} = \\{y_{i}\\in \\{0,1\\} \\mid i\\in \\mathcal{L}_{ns}\\} \\) by" + }, + { + "type": "equation", + "bbox": [ + 0.123, + 0.131, + 0.469, + 0.166 + ], + "angle": 0, + "content": "\\[\nE (\\boldsymbol {y}) = \\sum_ {j \\in \\mathcal {L} _ {s}} E _ {j} ^ {S P} (\\boldsymbol {y}) + \\lambda \\sum_ {i, j \\in \\mathcal {L} _ {n s}} E _ {i, j} ^ {M S C} (\\boldsymbol {y}) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.173, + 0.274, + 0.188 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is a hyperparameter." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.206, + 0.421, + 0.222 + ], + "angle": 0, + "content": "3.3.1 Octree Creation and Labelling Algorithm" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.23, + 0.469, + 0.335 + ], + "angle": 0, + "content": "The energy function in Equation 4 is non-trivial to minimize. It is a discrete energy minimisation problem with hard unary terms (the boundary constraint), pairwise terms (Equation 3 which is essentially a smoothness constraint) and n-ary terms (Equation 2). The n-ary term however is not concave, so we cannot use existing methods to perform exact inference using graph-cuts [13,20]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.336, + 0.469, + 0.517 + ], + "angle": 0, + "content": "Instead, we employ a move-making algorithm [9] with moves specialized for our task (see Figure 3). We first hard code the boundary constraint (leaves on the boundary of the domain are outside) and initialize all other leaves to inside. To be efficient, we (1) only consider moves that change leaves from inside to outside (never the other way around), (2) require such moves include an inside leaf from the inside-outside border, (3) iteratively grow such potential moves, and (4) apply moves at multiple depths. We also have a grow stage to remove extraneous inside leaves without needing guidance from the energy function. We now describe the stages of our algorithm, depicted in Algorithm 1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.518, + 0.469, + 0.592 + ], + "angle": 0, + "content": "Octree Creation. We start with a single node \\( n_0 \\) covering the whole space \\( \\mathcal{D} \\). We then recursively expand any node (into 8 equally sized child nodes) that contains points from \\( \\chi \\) in it until depth \\( d_i \\) is reached. Note that every surface leaf is at depth \\( d \\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.593, + 0.469, + 0.621 + ], + "angle": 0, + "content": "Initialisation. We hard code our boundary condition and set every other leaf node to inside (see Figure 3 A)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.622, + 0.469, + 0.743 + ], + "angle": 0, + "content": "Grow Stage. We first expand any neighbor of a surface leaf that is not also at the current depth \\( d \\) (new leaves keep the label of its parent). We repeatedly expand the outside region changing inside leaves to outside if they stick out from a plane, i.e., have at least \\( 26/2 + 1 = 14 \\) outside neighbors (or \\( 8/2 + 1 = 5 \\) for our 2D example in Figure 3). Finally we expand nodes that are on the inside-outsider border and are not at the current octree depth." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.744, + 0.469, + 0.879 + ], + "angle": 0, + "content": "Move-Making Stage. To overcome local minima we propose sets of leaves to change from inside to outside in a single move. Sets are constructed as follows. Starting with an inside leaf on the border we greedily add neighboring inside leaves to the set, selecting the neighbor which results in lowest energy, and calculate the cost of the current move set. Note that the energy can temporarily increase. We do this until there are no inside neighbors remaining or a maximum set size \\( s \\) is reached.1 If any of these moves result" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.892, + 0.136 + ], + "angle": 0, + "content": "in a decrease in energy, we apply the move with the largest decrease, We repeat the above for different starting leaves and for multiple increasing set sizes \\(s\\in C\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.137, + 0.892, + 0.287 + ], + "angle": 0, + "content": "Figure 3 illustrates this process in 2D. After reaching labelling (E), all single variable changes (\\(s = 1\\)) are considered, resulting in (F). Observe that the inside-outside border has been refined but larger regions are still incorrectly labelled. Considering larger moves as shown in (G) fixes many of these regions, with further refinements being made at deeper levels of the octree as shown in (H). We explain this 2D example in more detail in the supplemental material, including why we cannot reach a lower energy from fixing certain regions until a deeper level is reached." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.288, + 0.892, + 0.378 + ], + "angle": 0, + "content": "Expanding. We find that labelling at multiple depths is faster and more stable than directly labelling the largest depth. Note that expanding completely changes \\(\\mathcal{L}_s\\) and adds more nodes to \\(\\mathcal{L}_{ns}\\), and thus expands \\(y\\). We keep the labels for previous non-surface leaves, and assign all the new non-surface leaves to inside." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.387, + 0.64, + 0.404 + ], + "angle": 0, + "content": "3.4. SDF learning" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.41, + 0.892, + 0.502 + ], + "angle": 0, + "content": "We use the octree to guide the SDF learning. Given an INR, we train it with the combination of four losses. The first two are the standard losses in the literature [1,2,5,14, 34], the surface loss to enforce that the input point cloud should lie on the zero level set and the eikonal loss [14] to enforce that the gradient has magnitude one everywhere" + }, + { + "type": "equation", + "bbox": [ + 0.503, + 0.511, + 0.892, + 0.547 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {1} = \\sum_ {x \\in \\chi} | \\Phi (x, \\theta) |, \\mathcal {L} _ {2} = \\int_ {\\mathcal {D}} | \\| \\nabla_ {x} \\Phi (x, \\theta) \\| _ {2} - 1 | d x. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.557, + 0.892, + 0.663 + ], + "angle": 0, + "content": "We now introduce two loss functions that take guidance from the labelled octree. We first produce an approximate signed distance function \\(\\tilde{d}_s(x)\\) for finely sampled points \\(x\\in \\mathcal{D}\\) by computing the distance to the closest input point cloud point and choose its sign based on the label of the octree leaf that contains that point (for surface leaves we choose inside). We use this as direct supervision" + }, + { + "type": "equation", + "bbox": [ + 0.591, + 0.672, + 0.892, + 0.705 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {3} = \\int_ {\\mathcal {D}} | \\Phi (x, \\theta) - \\tilde {d} _ {s} (x) | \\mathrm {d} x. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.712, + 0.892, + 0.758 + ], + "angle": 0, + "content": "We also enforce the sign of each octree leaf explicitly (again taking surface leaves to be considered as inside) at a fixed number of randomly sampled points in that leaf \\( S_{\\ell} \\)" + }, + { + "type": "equation", + "bbox": [ + 0.552, + 0.767, + 0.892, + 0.809 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {4} = \\sum_ {\\ell \\in \\mathcal {L}} \\left(\\sum_ {x \\in S _ {\\ell}} \\left[ (- 1) ^ {\\left(1 - y _ {\\ell} ^ {\\prime}\\right)} \\Phi (x, \\theta) \\right] _ {> 0}\\right) \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.818, + 0.892, + 0.879 + ], + "angle": 0, + "content": "where \\( y_{\\ell}^{\\prime} = y_{\\ell} \\) for \\( \\ell \\in \\mathcal{L}_{ns} \\) and \\( y_{\\ell}^{\\prime} = 1 \\) for \\( \\ell \\in \\mathcal{L}_s \\). Note that sampling a fixed number per leaf automatically adjusts our sampling to be denser in more important regions (i.e., closer to the surface) due to the octree structure." + }, + { + "type": "page_footnote", + "bbox": [ + 0.095, + 0.887, + 0.469, + 0.901 + ], + "angle": 0, + "content": "1In practice we include additional early stopping criteria to avoid moves" + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.887, + 0.884, + 0.901 + ], + "angle": 0, + "content": "that are unlikely to eventually be accepted (see supplemental for details)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16721" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.092, + 0.226, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.182, + 0.223, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.141, + 0.271, + 0.191, + 0.285 + ], + "angle": 0, + "content": "Depth 7" + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.092, + 0.353, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.182, + 0.351, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.269, + 0.271, + 0.319, + 0.285 + ], + "angle": 0, + "content": "Depth 6" + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.098, + 0.481, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.182, + 0.478, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.392, + 0.271, + 0.45, + 0.286 + ], + "angle": 0, + "content": "SAP [31]" + }, + { + "type": "image", + "bbox": [ + 0.492, + 0.092, + 0.61, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.492, + 0.182, + 0.605, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.521, + 0.271, + 0.578, + 0.285 + ], + "angle": 0, + "content": "DiGS [5]" + }, + { + "type": "image", + "bbox": [ + 0.619, + 0.092, + 0.735, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.621, + 0.182, + 0.733, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.652, + 0.271, + 0.702, + 0.285 + ], + "angle": 0, + "content": "SAL [1]" + }, + { + "type": "image", + "bbox": [ + 0.745, + 0.092, + 0.863, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.746, + 0.182, + 0.861, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.752, + 0.271, + 0.857, + 0.285 + ], + "angle": 0, + "content": "SIREN won [34]" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.298, + 0.892, + 0.327 + ], + "angle": 0, + "content": "Figure 4. A comparison of surface reconstructions of shapes from the watercraft and bench classes of ShapeNet. The left two reconstructions are using OG-SIREN with the octree built to different depths." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.353, + 0.427, + 0.368 + ], + "angle": 0, + "content": "Our overall loss is then the weighted combination" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.38, + 0.469, + 0.396 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} (\\Phi) = \\lambda_ {1} \\mathcal {L} _ {1} + \\lambda_ {2} \\mathcal {L} _ {2} + \\lambda_ {3} \\mathcal {L} _ {3} + \\lambda_ {4} \\mathcal {L} _ {4}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.408, + 0.47, + 0.501 + ], + "angle": 0, + "content": "We keep \\(\\lambda_{1}\\) and \\(\\lambda_{2}\\) fixed throughout training as their loss terms have no bias. As the guidance in \\(\\mathcal{L}_3\\) and \\(\\mathcal{L}_4\\) is approximate but highly informative, we initially use large values for \\(\\lambda_3,\\lambda_4\\), allowing the INR to quickly fit to the rough shape, and then reduce them during training. More training details are in the supplemental material." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.513, + 0.164, + 0.529 + ], + "angle": 0, + "content": "4. Results" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.539, + 0.47, + 0.78 + ], + "angle": 0, + "content": "Datasets. We evaluate the performance of our method on the ShapeNet [12] and Surface Reconstruction Benchmark (SRB) [7] datasets. ShapeNet has a wide variety of shapes, with thin surfaces, hollow regions and disjoint objects. However the shapes are rather uniformly sampled. We use the preprocessing and split of Neural Splines [45], which consists of 13 shape classes with 20 shapes in each shape class. For each shape it provides the ground truth mesh (preprocessed to be watertight), an input point cloud and occupancy annotation at dense points. SRB contains five shapes from simulated noisy range scans, but it is a popularly used dataset due to the shapes exhibiting many challenging features. These include large missing regions, thin and sharp regions, and small surface details. The former means that good priors in the reconstruction methods, such as smoothness priors, are very important." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.469, + 0.825 + ], + "angle": 0, + "content": "We follow the metrics used in the literature for each dataset: the squared Chamfer distance and IoU for ShapeNet and Chamfer and Hausdorff distance for SRB." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.469, + 0.9 + ], + "angle": 0, + "content": "INRs. We demonstrate our method on two INRs, SIREN [34], and NGLOD [35] (note that NGLOD only works in a supervised setting as it has no global consistency, but our octree guidance is sufficient). Thus we name our two methods OG-SIREN and OG-NGLOD where OG stands for" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.353, + 0.893, + 0.414 + ], + "angle": 0, + "content": "\"Octree Guided\". To extract a mesh we first evaluate our INRs on a \\(256^3\\) grid and then use marching cubes [25]. We use the same settings for the octree and the INRs on both datasets, and give more details in our supplemental." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.423, + 0.757, + 0.437 + ], + "angle": 0, + "content": "4.1. Reconstruction Performance" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.446, + 0.893, + 0.672 + ], + "angle": 0, + "content": "ShapeNet dataset. The results are shown in Table 1 (left) and visualizations on two shapes are shown in Figure 4. Our methods get the smallest mean squared Chamfer distance out of the unoriented methods by almost an order of magnitude, and are even better than the best performing oriented method. Likewise for mean IoU our methods are better than both unoriented and oriented methods. For median performance, our methods are comparable to the best oriented and unoriented methods. SAP [31] gets the best median IoU, but their mean is much lower than ours, indicating that in a few instances they misrepresent large regions. On the other hand, due to our move-making algorithm we reduce the likelihood of vastly incorrect labellings, and as a result achieve low standard deviations on both metrics despite the high performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.672, + 0.892, + 0.809 + ], + "angle": 0, + "content": "SRB dataset. Table 1 (right) shows that DiGS, SAP, PHASE+FF and our two methods do comparably, with DiGS performing slightly better. Note the performance of these methods are comparable to most oriented methods. Also note that the difference between our two methods is more prominent on SRB as the large missing regions means that smoothness priors are more important. Octree guidance is NGLOD's only global supervision, as it fits locally, while SIREN is biased towards smooth shapes." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.818, + 0.719, + 0.833 + ], + "angle": 0, + "content": "4.2. Octree Depth Ablations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Table 4 shows the results of OG-SIREN on ShapeNet when evaluating at different depths of the octree. We include an octree IoU metric, where we compute the IoU of the labelling of the octree directly (taking surface leaves to" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16722" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.091, + 0.094, + 0.561, + 0.344 + ], + "angle": 0, + "content": "
MethodSquared Chamfer ↓IoU ↑
meanmedianstdmeanmedianstd
SPSR [19]5.44e-51.97e-55.06e-40.99260.99560.0105
IGR [14]5.12e-41.13e-42.15e-30.81020.84800.1519
SIREN [34]1.03e-45.28e-51.93e-40.82680.90970.2329
FFN [36]9.12e-58.65e-53.36e-50.82180.83960.0989
Biharmonic RBF [11]1.11e-48.97e-57.06e-50.82470.86420.1350
SVR [33]1.14e-41.04e-45.99e-50.76250.78190.1300
NSP [45]5.36e-54.06e-53.64e-50.89730.92300.0871
DiGS + n [5]2.74e-42.32e-59.90e-40.92000.97740.1992
N Est+SPSR [19]3.76e-34.37e-51.14e-20.71870.97610.3767
SIREN won [34]3.08e-42.58e-43.26e-40.30850.29520.2014
SAL [1]1.14e-32.11e-43.63e-30.40300.39440.2722
DiGS [5]1.32e-42.55e-54.73e-40.93900.97640.1262
SAP [31]4.09e-42.46e-52.60e-30.91180.99230.2002
Our OG-SIREN3.75e-52.17e-58.24e-50.96150.98710.1048
Our OG-NGLOD5.07e-52.57e-59.39e-50.95930.98700.1057
" + }, + { + "type": "table", + "bbox": [ + 0.633, + 0.09, + 0.837, + 0.351 + ], + "angle": 0, + "content": "
MethoddC↓dH↓
SPSR [19]0.214.69
DGP [44]0.215.18
IGR [14]0.192.99
SIREN [34]0.193.86
NSP [45]0.172.85
PHASE [22]0.162.77
DiGS + n [5]0.183.55
N Est.+SPSR [19]1.2522.59
IGR wo n [14]1.3816.33
SIREN wo n [34]0.427.67
SAL [14]0.367.47
IGR+FF [14]0.9611.06
PHASE+FF [22]0.224.96
DiGS [5]0.193.52
SAP [31]0.214.51
Our OG-SIREN0.204.06
Our OG-NGLOD0.226.03
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.361, + 0.895, + 0.404 + ], + "angle": 0, + "content": "Table 1. Surface Reconstruction results on ShapeNet [12] (left) and SRB [7] (right). Methods above the line use ground truth normal information, and methods below do not. Left: The mean, median and standard deviation of the squared Chamfer distance and IoU of all 260 shapes are reported. Right: The mean Chamfer \\(d_{C}\\) and Hausdorff distance \\(d_{H}\\) for all 5 shapes are reported." + }, + { + "type": "table", + "bbox": [ + 0.168, + 0.427, + 0.382, + 0.577 + ], + "angle": 0, + "content": "
MethoddC↓dH↓
SIREN w on [34]0.427.67
OG-SIREN0.204.06
SIREN (+n) [34]0.193.86
OG-SIREN + H°0.192.96
OG-SIREN + GT°0.172.44
OG-NGLOD0.226.03
OG-NGLOD + H°0.195.50
OG-NGLOD + GT°0.175.12
NGLOD+GTd [35]0.131.97
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.588, + 0.472, + 0.645 + ], + "angle": 0, + "content": "Table 2. Results on SRB [7] when supervising INRs with varying levels of information (see Section 4.3). \\(\\mathrm{GT}^d\\) means dense GT SDF supervision, while \\(\\mathrm{GT}^o\\) and \\(\\mathrm{H}^o\\) mean GT and human estimated occupancy per octree node." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.671, + 0.47, + 0.779 + ], + "angle": 0, + "content": "be inside). The IoU of the octree is much lower than the IoU of the final mesh, indicating that the INR training is learning beyond the octree supervision. Furthermore as the depth increases, the octree IoU improves and so do the metrics on the final mesh, indicating that the INR training is heavily affected by the initial training guidance from the octree. See the supplemental material for a depth ablation on SRB." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.787, + 0.274, + 0.803 + ], + "angle": 0, + "content": "4.3. Modifying the octree" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.471, + 0.903 + ], + "angle": 0, + "content": "To show the versatility of our method, in Table 2 we demonstrate how modifying the octree labelling can affect the performance of the resulting shape. For \\(\\mathrm{OG - INR + GT^o}\\) we set octree labels using GT occupancy. For \\(\\mathrm{OG - INR + H^o}\\) we set octree labels using sparse human supervision: we change up to 30 large octree leaves at the lower depths (up" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.43, + 0.895, + 0.489 + ], + "angle": 0, + "content": "to depth 5) where the changes are obvious from a human perspective. With both octree supervision signals we get increases in performance, showing that the discrete nature of octree guidance does not introduce much bias." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.49, + 0.895, + 0.536 + ], + "angle": 0, + "content": "We also compare to the INRs with other forms of supervision. As NGLOD cannot be used in an unsupervised setting, we provide result when trained on dense GT." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.545, + 0.707, + 0.562 + ], + "angle": 0, + "content": "4.4. Further Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.569, + 0.893, + 0.75 + ], + "angle": 0, + "content": "The need for unoriented methods. We compare to a recent industry-standard implementation [48] of normal estimation + SPSR [19] (SPSR+N Est.). Table 1 (left) shows that for the well sampled surfaces in ShapeNet, when ground truth normals are available SPSR achieves close to perfect IoUs. However when normals need to be estimated, it has good median performance but very poor average performance, indicating the normal estimation procedure will often be quite inaccurate. Table 1 (right) shows that for point clouds with large missing regions, SPSR with ground truth normals has room for improvement, while SPSR with estimated normals performs extremely poorly." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.894, + 0.826 + ], + "angle": 0, + "content": "While neither dataset has provided line of sight information (as would be given by sensors), a more informative use of such information than estimating rough normals would be to use it to label parts of our octree. See Section 4.3 for how additional octree supervision improves performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Overall Timing Comparison. Table 3 shows the time taken for each method. Our method's gradient descent training is the fastest to converge as it needs much fewer iterations. Even when adding on the average octree building time from ShapeNet, we can see that the time taken for both" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16723" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.172, + 0.089, + 0.803, + 0.212 + ], + "angle": 0, + "content": "
MethodParametersTime per iter. (s)Num itersTime (s)Speed Up
N Est. +SPSR---42 (13 + 29)12× (· + 18×)
SIREN (wo n)264K0.052100005201×
SAL2.1M0.1751000017500.3×
DiGS264K0.1201000012000.4×
SAP120K-32003301.6×
Our OG-SIREN264K0.158600135 (40 + 95)3.9× (· + 5.5×)
Our OG-NGLOD68.7M0.17330092 (40 + 52)5.7× (· + 10×)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.222, + 0.895, + 0.266 + ], + "angle": 0, + "content": "Table 3. Time and number of parameters for each model on ShapeNet. For methods with a precomputation stage (normal estimation or octree building), the time for both is included with the precomputation and training times in parenthesis. Speed up is measured with respect to SIREN/SIREN w o n. We ran our experiments on a single Nvidia RTX 2080 GPU and a i7-8700K CPU." + }, + { + "type": "table", + "bbox": [ + 0.125, + 0.277, + 0.846, + 0.372 + ], + "angle": 0, + "content": "
DepthMeanOctree IoU ↑ MedianStdMeanFinal Mesh (OG-SIREN)Squared Chamfer ↓Octree Build + Label Time (s)
IoU ↑ MedianStdMeanMedianStdMeanMedianStd
50.47480.47020.18540.71010.77120.25941.84e-41.35e-42.23e-42.21.62.2
60.64010.64100.16040.95030.98350.11723.93e-52.34e-58.20e-57.65.38.6
70.77570.80060.13050.96150.98710.10483.75e-52.17e-58.24e-538.119.367.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.085, + 0.382, + 0.882, + 0.397 + ], + "angle": 0, + "content": "Table 4. Octree depth ablation on ShapeNet. Octree IoU is computed using the labelling of the leaves (surface leaves chosen as inside)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.423, + 0.47, + 0.468 + ], + "angle": 0, + "content": "stages of our method is still faster than other methods, often by an order of magnitude. We implemented the octree in Cython [3] but further optimisations are possible." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.47, + 0.471, + 0.683 + ], + "angle": 0, + "content": "Comparison Against No Octree Guidance. Comparing SIREN wo n with and without the octree guidance we can see that the guidance drastically improves both time performance and reconstruction metrics (see Table 1,Table 3). When initially guided by our octree labelling, SIREN (wo n) converges significantly faster (92s vs 520s) and gets a 3x improvement in ShapeNet IoU. Without guidance from normals or our octree, SIREN wo n gets stuck in bad local minima where it needs to flip its inside-outsideness prediction, causing ghost geometry (see Figure 1). When comparing the two types of INR training guidance applied to SIREN wo n, ground truth normals vs our octree, we can see that the octree guidance still performs much better than SIREN on ShapeNet, while performing comparably on SRB." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.694, + 0.196, + 0.709 + ], + "angle": 0, + "content": "4.5. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.718, + 0.47, + 0.825 + ], + "angle": 0, + "content": "Limitations. Like all unoriented methods, the octree guidance has difficulty with large missing regions, which is demonstrated heavily in SRB, and thin surfaces, which appear in SRB (Daratech) and ShapeNet (airplane and watercraft). Despite this, our methods perform competitively on shapes with these difficulties on both datasets. See the supplemental for per class breakdown and visualizations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Other possible limitations would be the ability to deal with noise and outliers. Note that all methods do poorly with large amounts of either corruption. For small noise the method's performance does not degrade considerably as demonstrated by the results on SRB. For larger noise, the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.423, + 0.892, + 0.498 + ], + "angle": 0, + "content": "volume occupied by surface leaves would be much larger. While this would not introduce error, the octree's benefits would be reduced. For outliers there would be significant bias introduced by the octree method, which if sparse could be preprocessed out beforehand." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.5, + 0.892, + 0.576 + ], + "angle": 0, + "content": "Broader Impact. Faster and more accurate unoriented surface reconstruction methods will allow for accurate meshes to be constructed easily from regular scanners. Potential misuses of this technology include unauthorized replication and harmful content creation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.595, + 0.619, + 0.611 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.622, + 0.892, + 0.819 + ], + "angle": 0, + "content": "We have introduced an octree guided method for training implicit neural representations (INRs) for the task of unoriented surface reconstruction. We formulate the main difficulty of the task, inside-outside determinism, as an octree labelling problem with respect to an energy function that captures watertight surface properties and the minimal surface constraint. We also detail an efficient method to minimize the energy. We then use this supervision to guide the initial stage of INR training, using loss functions that can be used by any INR architecture. Our results show that applying our octree guided supervision provides significant time and reconstruction performance improvements, with competitive or better results than the current state-of-the-art." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Acknowledgements. This work has received funding from the European Union's Horizon 2020 research and innovation programme under the Marie Sklodowska-Curie grant agreement No. 893465. S. Gould is a recipient of an ARC Future Fellowship (proj. no. LP200100421) funded by the Australian Government." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "16724" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Matan Atzmon and Yaron Lipman. SAL: Sign Agnostic Learning of shapes from raw data. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2565-2574, 2020. 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.172, + 0.47, + 0.213 + ], + "angle": 0, + "content": "[2] Matan Atzmon and Yaron Lipman. SALD: Sign Agnostic Learning with Derivatives. In Proc. of the International Conference on Learning Representations (ICLR), 2021. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.213, + 0.47, + 0.268 + ], + "angle": 0, + "content": "[3] Stefan Behnel, Robert Bradshaw, Craig Citro, Lisandro Dalcin, Dag Sverre Seljebotn, and Kurt Smith. Cython: The best of both worlds. Computing in Science & Engineering, 13(2):31-39, 2011. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.269, + 0.47, + 0.325 + ], + "angle": 0, + "content": "[4] Yizhak Ben-Shabat and Stephen Gould. DeepFit: 3D surface fitting via neural network weighted least squares. In Proc. of the European Conference on Computer Vision (ECCV), pages 20-34. Springer, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.326, + 0.47, + 0.381 + ], + "angle": 0, + "content": "[5] Yizhak Ben-Shabat, Chamin Hewa Koneputugodage, and Stephen Gould. Digs: Divergence guided shape implicit neural representation for unoriented point clouds. arXiv preprint arXiv:2106.10811, 2021. 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.382, + 0.47, + 0.451 + ], + "angle": 0, + "content": "[6] Yizhak Ben-Shabat, Michael Lindenbaum, and Anath Fischer. Nesti-net: Normal estimation for unstructured 3D point clouds using convolutional neural networks. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 10112-10120, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.452, + 0.47, + 0.507 + ], + "angle": 0, + "content": "[7] Matthew Berger, Joshua A Levine, Luis Gustavo Nonato, Gabriel Taubin, and Claudio T Silva. A benchmark for surface reconstruction. ACM Trans. on Graphics (ToG), 32:1-17, 2013. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.508, + 0.47, + 0.55 + ], + "angle": 0, + "content": "[8] Julian Besag. On the statistical analysis of dirty pictures. Journal of the Royal Statistical Society: Series B (Methodological), 48(3):259-279, 1986. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.55, + 0.47, + 0.592 + ], + "angle": 0, + "content": "[9] Andrew Blake, Pushmeet Kohli, and Carsten Rother. Markov Random Fields for Vision and Image Processing. The MIT Press, 07 2011. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.593, + 0.47, + 0.648 + ], + "angle": 0, + "content": "[10] Yuri Boykov, Olga Veksler, and Ramin Zabih. Fast approximate energy minimization via graph cuts. IEEE Transactions on pattern analysis and machine intelligence, 23(11):1222-1239, 2001. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.649, + 0.47, + 0.731 + ], + "angle": 0, + "content": "[11] Jonathan C Carr, Richard K Beatson, Jon B Cherrie, Tim J Mitchell, W Richard Fright, Bruce C McCallum, and Tim R Evans. Reconstruction and representation of 3D objects with radial basis functions. In Proc. of the Conference on Computer Graphics and Interactive Techniques, pages 67-76, 2001. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.732, + 0.47, + 0.802 + ], + "angle": 0, + "content": "[12] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. ShapeNet: An information-rich 3D model repository. arXiv preprint arXiv:1512.03012, 2015. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.803, + 0.471, + 0.858 + ], + "angle": 0, + "content": "[13] Stephen Gould. Max-margin learning for lower linear envelope potentials in binary markov random fields. In Proc. of the International Conference on Machine Learning (ICML). Omnipress, 2011. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.859, + 0.471, + 0.902 + ], + "angle": 0, + "content": "[14] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit Geometric Regularization for learning shapes. In Proc. of the International Conference on" + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.115, + 0.471, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "Machine Learning (ICML), volume 119, pages 3789-3799. \nPMLR, 2020. 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.12, + 0.892, + 0.177 + ], + "angle": 0, + "content": "[15] Paul Guerrero, Yanir Kleiman, Maks Ovsjanikov, and Niloy J Mitra. PCPNet: Learning local shape properties from raw point clouds. In Computer Graphics Forum, volume 37, pages 75-85. Wiley Online Library, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.177, + 0.892, + 0.231 + ], + "angle": 0, + "content": "[16] Christian Hane, Shubham Tulsiani, and Jitendra Malik. Hierarchical surface prediction for 3d object reconstruction. In 2017 International Conference on 3D Vision (3DV), pages 412-420. IEEE, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.232, + 0.892, + 0.342 + ], + "angle": 0, + "content": "[17] Shahram Izadi, David Kim, Otmar Hilliges, David Molyneaux, Richard Newcombe, Pushmeet Kohli, Jamie Shotton, Steve Hodges, Dustin Freeman, Andrew Davison, and Andrew Fitzgibbon. Kinectfusion: Real-time 3d reconstruction and interaction using a moving depth camera. In UIST '11 Proceedings of the 24th annual ACM symposium on User interface software and technology, pages 559-568. ACM, October 2011. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.343, + 0.892, + 0.397 + ], + "angle": 0, + "content": "[18] Michael Kazhdan, Matthew Bolitho, and Hugues Hoppe. Poisson surface reconstruction. In Proc. of the Eurographics Symposium on Geometry Processing (SGP), volume 7, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.399, + 0.892, + 0.44 + ], + "angle": 0, + "content": "[19] Michael Kazhdan and Hugues Hoppe. Screened poisson surface reconstruction. ACM Trans. on Graphics (ToG), 32:1-13, 2013. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.441, + 0.892, + 0.496 + ], + "angle": 0, + "content": "[20] Pushmeet Kohli, M. Kumar, and Philip Torr. P3 & beyond: Solving energies with higher order cliques. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2007. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.497, + 0.892, + 0.552 + ], + "angle": 0, + "content": "[21] Jan Eric Lenssen, Christian Osendorfer, and Jonathan Masci. Deep iterative surface normal estimation. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 11247-11256, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.553, + 0.892, + 0.593 + ], + "angle": 0, + "content": "[22] Yaron Lipman. Phase transitions, distance functions, and implicit neural representations. arXiv preprint arXiv:2106.07689, 2021. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.594, + 0.892, + 0.649 + ], + "angle": 0, + "content": "[23] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. Advances in Neural Information Processing Systems, 33:15651-15663, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.65, + 0.892, + 0.72 + ], + "angle": 0, + "content": "[24] Shi-Lin Liu, Hao-Xiang Guo, Hao Pan, Peng-Shuai Wang, Xin Tong, and Yang Liu. Deep implicit moving least-squares functions for 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1788-1797, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.72, + 0.892, + 0.788 + ], + "angle": 0, + "content": "[25] William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3D surface construction algorithm. In Proc. of the Intl. Conference on Computer Graphics and Interactive Techniques (SIGGRAPH), volume 21, pages 163-169. ACM New York, NY, USA, 1987. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.789, + 0.892, + 0.845 + ], + "angle": 0, + "content": "[26] Julien NP Martel, David B Lindell, Connor Z Lin, Eric R Chan, Marco Monteiro, and Gordon Wetzstein. Acorn: Adaptive coordinate networks for neural scene representation. arXiv preprint arXiv:2105.02788, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.892, + 0.873 + ], + "angle": 0, + "content": "[27] Donald J. Meagher. Geometric modeling using octree encoding. Comput. Graph. Image Process., 19:85, 1982. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[28] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks:" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16725" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "Learning 3D reconstruction in function space. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4460-4470, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.47, + 0.191 + ], + "angle": 0, + "content": "[29] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, July 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.193, + 0.471, + 0.262 + ], + "angle": 0, + "content": "[30] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. DeepSDF: Learning continuous signed distance functions for shape representation. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 165-174, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.264, + 0.47, + 0.319 + ], + "angle": 0, + "content": "[31] Songyou Peng, Chiyu Jiang, Yiyi Liao, Michael Niemeyer, Marc Pollefeys, and Andreas Geiger. Shape As Points: A differentiable poisson solver. arXiv preprint arXiv:2106.03452, 2021. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.321, + 0.47, + 0.377 + ], + "angle": 0, + "content": "[32] Gernot Riegler, Ali Osman Ulusoy, and Andreas Geiger. Octnet: Learning deep 3d representations at high resolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3577-3586, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.379, + 0.469, + 0.447 + ], + "angle": 0, + "content": "[33] Bernhard Scholkopf, Joachim Giesen, and Simon Spalinger. Kernel methods for implicit surface modeling. In Advances in Neural Information Processing Systems (NeurIPS), NIPS'04, page 1193-1200, Cambridge, MA, USA, 2004. MIT Press. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.449, + 0.47, + 0.517 + ], + "angle": 0, + "content": "[34] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. Advances in Neural Information Processing Systems (NeurIPS), 33, 2020. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.52, + 0.47, + 0.603 + ], + "angle": 0, + "content": "[35] Towaki Takikawa, Joey Litalien, Kangxue Yin, Karsten Kreis, Charles Loop, Derek Nowrouzezahrai, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Neural geometric level of detail: Real-time rendering with implicit 3D shapes. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.605, + 0.47, + 0.688 + ], + "angle": 0, + "content": "[36] Matthew Tancik, Pratul P. Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan T. Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. Advances in Neural Information Processing Systems (NeurIPS), 2020. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.689, + 0.47, + 0.759 + ], + "angle": 0, + "content": "[37] Jia-Heng Tang, Weikai Chen, Jie Yang, Bo Wang, Songrun Liu, Bo Yang, and Lin Gao. Octfield: Hierarchical implicit functions for 3d modeling. In The Thirty-Fifth Annual Conference on Neural Information Processing Systems (NeurIPS), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.47, + 0.829 + ], + "angle": 0, + "content": "[38] Maxim Tatarchenko, Alexey Dosovitskiy, and Thomas Brox. Octree generating networks: Efficient convolutional architectures for high-resolution 3d outputs. In Proceedings of the IEEE international conference on computer vision, pages 2088-2096, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.832, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[39] Ayush Tewari, Justus Thies, Ben Mildenhall, Pratul Srinivasan, Edgar Tretschk, Yifan Wang, Christoph Lassner, Vincent Sitzmann, Ricardo Martin-Brualla, Stephen Lombardi, et al. Advances in neural rendering. arXiv preprint arXiv:2111.05849, 2021. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[40] Peng-Shuai Wang, Yang Liu, Yu-Xiao Guo, Chun-Yu Sun, and Xin Tong. O-cnn: Octree-based convolutional neural networks for 3d shape analysis. ACM Transactions On Graphics (TOG), 36(4):1-11, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.218 + ], + "angle": 0, + "content": "[41] Peng-Shuai Wang, Yang Liu, and Xin Tong. Deep octree-based cnns with output-guided skip connections for 3d shape and scene completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 266-267, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.22, + 0.892, + 0.274 + ], + "angle": 0, + "content": "[42] Peng-Shuai Wang, Yang Liu, and Xin Tong. Dual octree graph networks for learning adaptive volumetric shape representations. ACM Transactions on Graphics (TOG), 41(4):1-15, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.276, + 0.892, + 0.33 + ], + "angle": 0, + "content": "[43] Peng-Shuai Wang, Chun-Yu Sun, Yang Liu, and Xin Tong. Adaptive o-cnn: A patch-based deep representation of 3d shapes. ACM Transactions on Graphics (TOG), 37(6):1-11, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.333, + 0.892, + 0.402 + ], + "angle": 0, + "content": "[44] Francis Williams, Teseo Schneider, Claudio Silva, Denis Zorin, Joan Bruna, and Daniele Panozzo. Deep Geometric Prior for surface reconstruction. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 10130-10139, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.404, + 0.892, + 0.471 + ], + "angle": 0, + "content": "[45] Francis Williams, Matthew Trager, Joan Bruna, and Denis Zorin. Neural Splines: Fitting 3D surfaces with infinitely-wide neural networks. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 9949-9958, 2021. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.474, + 0.892, + 0.543 + ], + "angle": 0, + "content": "[46] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond. Computer Graphics Forum, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.545, + 0.892, + 0.598 + ], + "angle": 0, + "content": "[47] Wang Yifan, Lukas Rahmann, and Olga Sorkine-Hornung. Geometry-consistent neural shape representation with implicit displacement fields. arXiv preprint arXiv:2106.05187, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.601, + 0.892, + 0.641 + ], + "angle": 0, + "content": "[48] Qian-Yi Zhou, Jaesik Park, and Vladlen Koltun. Open3D: A modern library for 3D data processing. arXiv:1801.09847, 2018. 7" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.641 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16726" + } + ] +] \ No newline at end of file diff --git a/2023/Octree Guided Unoriented Surface Reconstruction/0db48127-a127-468b-be90-b27ec5813046_origin.pdf b/2023/Octree Guided Unoriented Surface Reconstruction/0db48127-a127-468b-be90-b27ec5813046_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a8fb943f016fcf6409ed32d1147c7769218fc6a3 --- /dev/null +++ b/2023/Octree Guided Unoriented Surface Reconstruction/0db48127-a127-468b-be90-b27ec5813046_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0ddb2eaca081c99f2757d051e85e92a5be3493b1d9a924e36a8404be99a1628b +size 5321749 diff --git a/2023/Octree Guided Unoriented Surface Reconstruction/full.md b/2023/Octree Guided Unoriented Surface Reconstruction/full.md new file mode 100644 index 0000000000000000000000000000000000000000..1ed7cd05006e7e97af6dea7066caf7502fd90a5c --- /dev/null +++ b/2023/Octree Guided Unoriented Surface Reconstruction/full.md @@ -0,0 +1,383 @@ +# Octree Guided Unoriented Surface Reconstruction + +Chamin Hewa Koneputugodage1 + +Yizhak Ben-Shabat1,2 + +Stephen Gould + +1The Australian National University + +$^{2}$ Technion Israel Institute of Technology + +{chamin.hewa, stephen.gould}@anu.edu.au sitszikbs@gmail.com + +# Abstract + +We address the problem of surface reconstruction from unoriented point clouds. Implicit neural representations (INRs) have become popular for this task, but when information relating to the inside versus outside of a shape is not available (such as shape occupancy, signed distances or surface normal orientation) optimization relies on heuristics and regularizers to recover the surface. These methods can be slow to converge and easily get stuck in local minima. We propose a two-step approach, OG-INR, where we (1) construct a discrete octree and label what is inside and outside (2) optimize for a continuous and high-fidelity shape using an INR that is initially guided by the octree's labelling. To solve for our labelling, we propose an energy function over the discrete structure and provide an efficient move-making algorithm that explores many possible labellings. Furthermore we show that we can easily inject knowledge into the discrete octree, providing a simple way to influence the result from the continuous INR. We evaluate the effectiveness of our approach on two unoriented surface reconstruction datasets and show competitive performance compared to other unoriented, and some oriented, methods. Our results show that the exploration by the move-making algorithm avoids many of the bad local minima reached by purely gradient descent optimized methods (see Figure 1). + +# 1. Introduction + +Surface reconstruction from 3D point clouds has been studied extensively in computer vision and computer graphics. The task requires estimating a mesh of the shape's surface from a point cloud sampled from the surface of the shape. We focus on the reconstruction of watertight 3D shapes, i.e., shapes that have a well defined interior and exterior. Such shapes are often represented as signed distance fields (SDFs) or occupancy fields, which can be efficiently encoded by a neural network [28, 30]. As these representations are fields parameterized by neural networks, they are often called neural fields. Furthermore, they implicitly rep + +![](images/5de864bc5d353c490ae1de645706df3dd2d6590d7777bb1840f60601319edfee.jpg) +Figure 1. Our method with a SIREN INR (OG-SIREN, left) compared to SIREN wo n (right) for two shapes from the ShapeNet dataset. Our octree guidance allows for consistent inside-outside determinism. On the other hand, SIREN wo n gets stuck in local minima from which it cannot escape (due to needing to completely change the occupancy of certain areas) creating extraneous surfaces (often called ghost geometries in the literature). + +resent the shape by a level set of the field, thus they are also often referred to as implicit neural representations (INRs). + +The broader class of neural fields, including INRs, have been very popular over the last few years as they can handle arbitrary topology, are memory efficient, and are continuous with potentially infinite resolution [39, 46]. Among the INRs for 3D shapes, SDFs are the most popular as they provide more useful information (distances not just occupancy) and are required for downstream graphics algorithms such as sphere tracing and approximate soft shadows [29, 35]. + +When learning an implicit representation of a shape, a major difficulty is predicting whether points in space are inside or outside the shape. Many methods require oriented surface normals for the input points or signed distances from the surface, which usually are not given by raw data from scans. While this can be estimated using the line of sight information [17] or algorithms [4,6, 15, 21], they yield noisy predictions that after postprocessing still can lead to bad results (see Section 4.4). We consider the task of unoriented surface reconstruction, where such information is not available, and only the sampled surface points are given. We demonstrate that our method performs competitively and sometimes better than oriented methods, even when they are given the ground truth (GT) normals. + +![](images/4c2eb5865acd0522fff845a067abf9d689f85cc689a884b86aabb874c412093c.jpg) +Figure 2. An illustration of our method, OG-INR. Given an unoriented point cloud (left), we progressively build and label an octree around the points (middle). The octree at depths 3-7 are shown. Surface leaves of the octree (yellow) are leaves that contain points from the point cloud, other leaves are labelled as inside (blue) or outside (transparent) by minimizing an energy function. We then train an INR model to obtain an SDF, using the labelling as supervision for the initial training, after which we can extract a mesh (right). + +We propose OG-INR, which uses a discrete representation in conjunction with the continuous INR representation (see Figure 2). Given an input unoriented point cloud, we progressively build an octree from the input points and determine which leaf nodes are surface leaves (i.e., leaves that contain a point). We also label all other leaves as inside or outside leaves (if they should be within or outside the shape respectively). To do this we minimize an energy function that trades off the watertight surface property, that every surface point should border with the inside and outside of the shape, with a minimal surface constraint. We then train standard INR architectures, initially guiding the training by the octree labelling, and show that the INR converges much faster and is less prone to large failure regions. + +Our main contributions are: + +- We introduce a novel method to guide the initial stage of INR training. It uses a labelled octree structure to allow the INR to converge significantly faster and alleviates the local minima problem of INRs. +- We propose an energy function over octree labels that captures the task of surface reconstruction. It balances the constraint of maintaining known surface regions with minimising the overall surface area. +- We provide an efficient move-making algorithm to optimize the energy function, which explores many inside-outside possibilities in a structured manner. +- Our discrete representation is easily understandable and human-modifiable, giving an intuitive method for applying changes to the resulting SDF. + +# 2. Related Work + +Unsupervised Surface Reconstruction. Representing shapes with INRs was popularized by DeepSDF [30], where a ReLU network directly models an SDF using GT SDF supervision. SAL [1] instead uses unsigned distance supervision, approximated from the input point cloud, as an + +unsupervised and unoriented approach. SALD [2] showed that adding unoriented normal supervision improved upon this, and IGR [14] further added the eikonal constraint loss (see Equation 5) that has been widely adopted [5, 22, 34]. NSP [45] addresses the task by solving a kernel regression problem but needs input normals. PHASE [22] introduces a novel loss inspired by the phase transition problem that encourages boundary crossings at the surface points. + +However, as the inputs to INRs are low dimensional, they are biased to low frequency solutions that do not recover high fidelity detail [36]. To introduce high frequencies, SIREN [34] uses sine activations and Fourier Feature Networks [36] use a fixed Fourier encoding. While being able to reconstruct high fidelity shapes with a modest number of parameters, these methods are hard to train as the high frequencies overfit quickly. A different approach is parametric encoding [29, 35], where spatial regions are represented by trainable parameters (often using an octree structure). We cover such methods in the next subsection. + +Note that IGR and SIREN can operate without normals, which we refer to by "IGR wo n" and "SIREN wo n". + +Octrees for Surface Reconstruction. Octrees were first introduced by Meagher [27] as an efficient data structure for representing and manipulating 3D data, and are now used extensively. The classic method Poisson Surface Reconstruction (PSR) [18] and its followup SPSR [19] use octrees to structure their implicit functions. + +Early works on using octrees with neural networks include HSP [16] and OGN [38], who train encoder-decoder networks to learn octrees with occupancy values, and OctNet [32] and O-CNN [40], who implement 3D CNNs that operate over octrees rather than voxel grids for shape analysis tasks. O-CNN's descendants [41, 43] build on this and adapt it to shape generation, completion and reconstruction. + +More recently octrees have been used in INRs for representing fine details efficiently. OctField [37] and NSVF [23] both learn local implicit functions for individual octree + +nodes. NGLOD [35] and ACORN [26] instead learn a single INR, but store trainable parameters at octree node vertices to encode spatial regions. IMLSNet [24] uses O-CNN to predict both an octree and oriented points per octree node to represent the surface using implicit moving least squares. + +However, previous octree reconstruction methods were supervised, trained on GT SDF/occupancy of either the current shape or a separate training set. Recently Dual Octree Graph CNNs [42] addressed the unsupervised setting, where they learn local features on the dual graph of an octree. However their method has parameters that need unsupervised training, and they require input normals. + +SAP [31] truly addresses the unsupervised and unoriented task by making the PSR pipeline differentiable. Starting from an oriented point cloud of a sphere, they use Chamfer loss to guide that point cloud to a dense, oriented representation. Their method is fast and among state-of-the-art, however it can fail drastically and often does not produce a watertight mesh (see Section 4). It is also hard to adapt to different applications and losses, unlike INRs. + +Initialization for INRs. Due to the difficulty of training INRs, initialization is very important [1, 5]. However regardless of input, most methods initialize to a sphere. These include SAL [1] and its descendants [2, 14, 22], DiGS [5] and IDF [47]. This gives many important properties, such as biasing shape interior to the domain's center and satisfying the eikonal condition at the start of training, but it is problematic for shapes with sharp concavities. Our method, on the other hand, provides firm guidance for both the eikonal constraint and the shape based on the input point cloud, and thus does not bias the training toward convex objects. + +Energy Minimization. Discrete energy minimization has been used extensively in both 2D and 3D vision for segmentation and object labelling [9, 10]. An energy function encodes our preference over the joint assignment to a set of variables, e.g., binary variables representing occupancy of voxels in 3D space. The goal of energy minimization is to find an assignment to the variables with minimal cost. This is NP-hard in general and a number of approximation algorithms have been proposed [9]. One class of algorithms, known as move-making, starts with an initial assignment and iteratively evaluates changes to a subset of the variables (called a move). If the change results in lower energy then the move is accepted, otherwise the changed values are reverted. Moves are repeatedly tried until no further reduction in energy is found. Iterated Conditional Modes (ICM) [8] is the simplest example, where each move involves changing a single variable and so is very efficient, but it is prone to getting stuck in local minima. Generalized ICM improves on this by allowing multiple variables to change in a move, but it is very costly. Our method is a variant of generalized ICM specialized to our problem, where we introduce restrictions on the move space improve efficiency. + +Algorithm 1 Overall octree labeling algorithm +1: procedure LABELOCTREE $(\chi ,d_i,k,C)$ +2: $\mathcal{O} =$ CreateOctree $(\chi ,d_{i})$ +3: Initialization(O) +4: for $j$ in $0,\dots ,k$ do +5: if $j\neq 0$ then +6: ExpandTree(O) +7: end if ▷ Now at depth $d = d_{i} + j$ +8: Grow(O) +9: for $s$ in $C$ do +10: makeMove(O,s) +11: end for +12: end for +13: return O +14: end procedure + +# 3. OG-INR: Octree Guided Shape INRs + +# 3.1. Problem Formulation + +Let $\chi = \{x_{i}\}_{i\in I}$ be a point cloud sampled from a surface $S = \partial \mathcal{V}$ of some watertight volume $\mathcal{V} \subset \mathcal{D}$ where $\mathcal{D} \subset \mathbb{R}^3$ is compact. We wish to reconstruct the surface $S$ . We choose to represent $S$ as the zero-level set of a signed distance function (SDF), and thus wish to train a neural network $\Phi$ to output signed distances for all points $x \in \mathcal{D}$ , + +$$ +\Phi (x, \theta) = (- 1) ^ {o (x)} d (x, \mathcal {S}) \tag {1} +$$ + +where $\theta$ are the model parameters, $o(x)$ is the occupancy of point x (1 if inside and 0 if outside) and $d(x,S)$ is the distance from $x$ to its closest point on $S$ . + +# 3.2. Overview + +We first aim to produce a rough and discrete, but globally consistent, occupancy determination for our domain by solving for an inside-outside labelling on nodes of an octree. To do this we propose an energy function that combines both the watertight surface property and the minimal surface objective. Despite this energy being intractable to solve exactly, we provide a domain-guided move-making algorithm to get low energy solutions (see Section 3.3). + +We then design additional INR loss terms to guide INRs using our octree labelling (see Section 3.4). These losses quickly guide the INR towards a rough but globally consistent representation that allows it to avoid many bad local minima. Since these loss terms are agnostic to INR architecture, our method can work with any INR model. + +As we assume our domain is compact, we first normalize the point cloud to fit within the unit sphere in 3D, and then set our domain as $\mathcal{D} = [-1.1, 1.1]^3$ . + +![](images/f8e6d3e2c77f83f9c633127372ee809593877ed8787d6deaacee1dc63003561e.jpg) +A: Depth 3, After Initialization(O) + +![](images/d0e11320b3b5133d63e407682efb364864df0910d567f0ab41091ea9cfbbf82a.jpg) +B: Depth 3, After Grow(O) + +![](images/6792545c03f90c3363a9f1c34b9830756db7c86ee42e7b34da3a1c4ad4ae3dba.jpg) +C: Depth 3, After all makeMove(O,s) + +![](images/7a1ba188c9c51b656e16c8c46a2a344d98b0f719b8159361c13834b4468f142b.jpg) +D: Depth 4, After all makeMove(O, s) + +![](images/c74dd1814d045289fb252dafcc233062c15cbde10bf3e05c29cbb16e54837a3d.jpg) +E: Depth 5, After Grow(O) +Figure 3. Our algorithm demonstrated on a 2D quadtree example. The black dots are our input point cloud, the yellow squares are surface leaves, and the blue and gray squares are non-surface leaves labelled as inside and outside respectively. + +![](images/1868eaab03bd81417376bc723aef5ce934ae151b0b40a9f884cb014f065fa34b.jpg) +F: Depth 5, After makeMove(O,1) + +![](images/137f8dc3155e079b792fa48ddef725f8cf477f359953eb425859653beac36864.jpg) +G: Depth 5, After all makeMove(O,s) + +![](images/eac3eca101ede04a8fd90faad2d351a6964c8305332a7203f3f8935c26f0bec7.jpg) +H: Depth 6, After all makeMove(O,s) + +# 3.3. Octree Construction and Labelling + +Setup. We present our octree labelling algorithm in Algorithm 1, which takes in point cloud $\chi$ , an initial depth to build the octree to $d_{i}$ , a number of further depth expansions $k$ (the final octree has depth $d_{f} = d_{i} + k$ ) and a set of move set sizes $C$ . We set $d = 3$ , $k = 4$ and $C = [1,2,10,10000]$ in our experiments. The algorithm returns a labelled octree $\mathcal{O} = (\mathcal{N},\mathcal{L}_{ns},\mathcal{L}_s,y)$ where + +- $\mathcal{N}$ are the nodes of the octree. The octree is built by repeatedly expanding nodes containing or near points from $\chi$ , so it is highly unbalanced. +- $\mathcal{L} \subset \mathcal{N}$ are the leaf nodes, and thus $\mathcal{L}$ partitions $\mathcal{D}$ . Furthermore, we define surface leaves $\mathcal{L}_s \subset \mathcal{L}$ to be leaves that contain points from the point cloud (colored yellow in Figure 2 and Figure 3) and all other leaves to be non-surface leaves $\mathcal{L}_{ns} = \mathcal{L} \setminus \mathcal{L}_s$ . +- $y$ is the binary occupancy label vector indexed by non-surface leaves $\ell \in \mathcal{L}_{ns}$ denoting whether the space defined by each $\ell$ is inside, $y_{\ell} = 1$ , or outside, $y_{\ell} = 0$ . + +We will construct our algorithm so that the neighbours of every surface leaf $\ell$ is at the same depth. Thus let $n_{\ell}^{s}$ and $n_{\ell}^{ns}$ be the number of its surface and non-surface neighbours and $n_{\ell}^{(1)}(\pmb{y})$ and $n_{\ell}^{(0)}(\pmb{y})$ be the number of inside and outside neighbors w.r.t. labelling $\pmb{y}$ . Then $n_{\ell}^{ns} = n_{\ell}^{(1)}(\pmb{y}) + n_{\ell}^{(0)}(\pmb{y})$ and $n_{\ell}^{s} + n_{\ell}^{s} = 26$ (or 8 for our 2D example in Figure 3). + +We will constantly refer to the boundary as the leaves touching the boundary of $\mathcal{D}$ and the inside-outside border to be the non-surface leaves who have a non-surface neighbor with a different label (different occupancy value). + +We determine inside-outside occupancy based on three properties: (1) surface property: surfaces should separate + +inside and outside regions, (2) minimal surface constraint: the total surface area of the shape's surface $S$ should be as small as possible, (3) boundary constraint: the boundary of the domain $\mathcal{D}$ should be outside. + +Energy function. We propose an energy function over labellings $\mathbf{y} = \{y \mid y_i \in \{0,1\}, i \in \mathcal{L}_{ns}\}$ based on the first two criteria, the third is hard-coded at the start of the algorithm. For the surface property, ideally every surface leaf would have both inside and outside neighbours, however it may only have surface leaf neighbours which contain both inside and outside regions by definition (see Figure 3 C,D,G,H, some surface leafs have no inside neighbours). For the minimal surface constraint, we only create surfaces when adjacent non-surface leaves have different labellings (note surface leaves are fixed), so we want to minimise this. + +For the surface property, we want every surface leaf $\ell$ to have some amount of both inside and outside neighbours, where surface neighbours partially count towards both. Thus, we aim for $\gamma_{\ell}^{(1)} = \gamma^{(1)} - \eta^{(1)}n_{\ell}^{s}$ non-surface inside neighbors and $\gamma_{\ell}^{(0)} = \gamma^{(0)} - \eta^{(0)}n_{\ell}^{s}$ non-surface outside neighbors where $\gamma^{(\cdot)},\eta^{(\cdot)}$ are hyperparameters. Then the energy cost for a surface leaf $\ell$ is + +$$ +E _ {\ell} ^ {S P} (\boldsymbol {y}) = \max \left\{\gamma_ {\ell} ^ {(1)} - n _ {\ell} ^ {(1)} (\boldsymbol {y}), \gamma_ {\ell} ^ {(0)} - n _ {\ell} ^ {(0)} (\boldsymbol {y}), 0 \right\}. \tag {2} +$$ + +For the minimal surface constraint, we want to minimise the total surface area between every pair of non-surface neighbors with different labellings. Then for two non-surface leaves $\ell, \ell'$ , our energy cost is + +$$ +E _ {\ell , \ell^ {\prime}} ^ {M S C} (\boldsymbol {y}) = a _ {\ell \ell^ {\prime}} \left| y _ {\ell} - y _ {\ell^ {\prime}} \right| \tag {3} +$$ + +where $a_{\ell \ell^{\prime}}$ is the (surface) area in common between them (note $a_{ij} = 0$ if they are not adjacent). + +The two criteria can be formalized as an energy function over labellings $\mathbf{y} = \{y_{i}\in \{0,1\} \mid i\in \mathcal{L}_{ns}\}$ by + +$$ +E (\boldsymbol {y}) = \sum_ {j \in \mathcal {L} _ {s}} E _ {j} ^ {S P} (\boldsymbol {y}) + \lambda \sum_ {i, j \in \mathcal {L} _ {n s}} E _ {i, j} ^ {M S C} (\boldsymbol {y}) \tag {4} +$$ + +where $\lambda$ is a hyperparameter. + +# 3.3.1 Octree Creation and Labelling Algorithm + +The energy function in Equation 4 is non-trivial to minimize. It is a discrete energy minimisation problem with hard unary terms (the boundary constraint), pairwise terms (Equation 3 which is essentially a smoothness constraint) and n-ary terms (Equation 2). The n-ary term however is not concave, so we cannot use existing methods to perform exact inference using graph-cuts [13,20]. + +Instead, we employ a move-making algorithm [9] with moves specialized for our task (see Figure 3). We first hard code the boundary constraint (leaves on the boundary of the domain are outside) and initialize all other leaves to inside. To be efficient, we (1) only consider moves that change leaves from inside to outside (never the other way around), (2) require such moves include an inside leaf from the inside-outside border, (3) iteratively grow such potential moves, and (4) apply moves at multiple depths. We also have a grow stage to remove extraneous inside leaves without needing guidance from the energy function. We now describe the stages of our algorithm, depicted in Algorithm 1. + +Octree Creation. We start with a single node $n_0$ covering the whole space $\mathcal{D}$ . We then recursively expand any node (into 8 equally sized child nodes) that contains points from $\chi$ in it until depth $d_i$ is reached. Note that every surface leaf is at depth $d$ . + +Initialisation. We hard code our boundary condition and set every other leaf node to inside (see Figure 3 A). + +Grow Stage. We first expand any neighbor of a surface leaf that is not also at the current depth $d$ (new leaves keep the label of its parent). We repeatedly expand the outside region changing inside leaves to outside if they stick out from a plane, i.e., have at least $26/2 + 1 = 14$ outside neighbors (or $8/2 + 1 = 5$ for our 2D example in Figure 3). Finally we expand nodes that are on the inside-outsider border and are not at the current octree depth. + +Move-Making Stage. To overcome local minima we propose sets of leaves to change from inside to outside in a single move. Sets are constructed as follows. Starting with an inside leaf on the border we greedily add neighboring inside leaves to the set, selecting the neighbor which results in lowest energy, and calculate the cost of the current move set. Note that the energy can temporarily increase. We do this until there are no inside neighbors remaining or a maximum set size $s$ is reached.1 If any of these moves result + +in a decrease in energy, we apply the move with the largest decrease, We repeat the above for different starting leaves and for multiple increasing set sizes $s\in C$ + +Figure 3 illustrates this process in 2D. After reaching labelling (E), all single variable changes ( $s = 1$ ) are considered, resulting in (F). Observe that the inside-outside border has been refined but larger regions are still incorrectly labelled. Considering larger moves as shown in (G) fixes many of these regions, with further refinements being made at deeper levels of the octree as shown in (H). We explain this 2D example in more detail in the supplemental material, including why we cannot reach a lower energy from fixing certain regions until a deeper level is reached. + +Expanding. We find that labelling at multiple depths is faster and more stable than directly labelling the largest depth. Note that expanding completely changes $\mathcal{L}_s$ and adds more nodes to $\mathcal{L}_{ns}$ , and thus expands $y$ . We keep the labels for previous non-surface leaves, and assign all the new non-surface leaves to inside. + +# 3.4. SDF learning + +We use the octree to guide the SDF learning. Given an INR, we train it with the combination of four losses. The first two are the standard losses in the literature [1,2,5,14, 34], the surface loss to enforce that the input point cloud should lie on the zero level set and the eikonal loss [14] to enforce that the gradient has magnitude one everywhere + +$$ +\mathcal {L} _ {1} = \sum_ {x \in \chi} | \Phi (x, \theta) |, \mathcal {L} _ {2} = \int_ {\mathcal {D}} | \| \nabla_ {x} \Phi (x, \theta) \| _ {2} - 1 | d x. \tag {5} +$$ + +We now introduce two loss functions that take guidance from the labelled octree. We first produce an approximate signed distance function $\tilde{d}_s(x)$ for finely sampled points $x\in \mathcal{D}$ by computing the distance to the closest input point cloud point and choose its sign based on the label of the octree leaf that contains that point (for surface leaves we choose inside). We use this as direct supervision + +$$ +\mathcal {L} _ {3} = \int_ {\mathcal {D}} | \Phi (x, \theta) - \tilde {d} _ {s} (x) | \mathrm {d} x. \tag {6} +$$ + +We also enforce the sign of each octree leaf explicitly (again taking surface leaves to be considered as inside) at a fixed number of randomly sampled points in that leaf $S_{\ell}$ + +$$ +\mathcal {L} _ {4} = \sum_ {\ell \in \mathcal {L}} \left(\sum_ {x \in S _ {\ell}} \left[ (- 1) ^ {\left(1 - y _ {\ell} ^ {\prime}\right)} \Phi (x, \theta) \right] _ {> 0}\right) \tag {7} +$$ + +where $y_{\ell}^{\prime} = y_{\ell}$ for $\ell \in \mathcal{L}_{ns}$ and $y_{\ell}^{\prime} = 1$ for $\ell \in \mathcal{L}_s$ . Note that sampling a fixed number per leaf automatically adjusts our sampling to be denser in more important regions (i.e., closer to the surface) due to the octree structure. + +![](images/cf02dbfef551deea93b6c981bba572e6dcc901c4ef9e131c27ea4e8d6e2f4a3e.jpg) + +![](images/59d4b6c9af85268ca638d5411996e72771ca83eae409b4bd53e4c0a35d2120c8.jpg) +Depth 7 + +![](images/e29c96d12c52aa202598aa68dba57b202a065e16a592c7806e70a9dc1522bc4c.jpg) + +![](images/23f4291f6c796179b1eb1a1b198f038378670bfea4d4a266f6dde2bcfac05fe0.jpg) +Depth 6 + +![](images/a58e93455b70c98fa68ef061c4771657b3147a8f0f0bca953c8a9002ac542583.jpg) + +![](images/966fbf36fd3d67bf945d4fa9a375a4a7df7ea805117cba6fd9d7d18981d895f6.jpg) +SAP [31] +Figure 4. A comparison of surface reconstructions of shapes from the watercraft and bench classes of ShapeNet. The left two reconstructions are using OG-SIREN with the octree built to different depths. + +![](images/2627356b92797d69c7ee784fdfc3805b03fa2631acb814b35f92190b7d5e3cf5.jpg) + +![](images/f9c98c37b2c9267b067220a4b2543d3b5dce4e068afae9a2d3d97810903853ea.jpg) +DiGS [5] + +![](images/c0de15092a2816fc81dfc4896807309f2496d9f8e205671fdccffddf100ac226.jpg) + +![](images/3aedd6e7cc40a7ed6823ef2c29571a59e4f0d60ea00467ee07e45335e5a20639.jpg) +SAL [1] + +![](images/c09ef78ac6c66bf30905402a33ffb4ff88416f88bcd96e96c774d9ed0ee3fe70.jpg) + +![](images/1b709118308ae31c6c3b8114b850aa24fd499e2437ee5ae58a43f887c3c56f94.jpg) +SIREN won [34] + +Our overall loss is then the weighted combination + +$$ +\mathcal {L} (\Phi) = \lambda_ {1} \mathcal {L} _ {1} + \lambda_ {2} \mathcal {L} _ {2} + \lambda_ {3} \mathcal {L} _ {3} + \lambda_ {4} \mathcal {L} _ {4}. \tag {8} +$$ + +We keep $\lambda_{1}$ and $\lambda_{2}$ fixed throughout training as their loss terms have no bias. As the guidance in $\mathcal{L}_3$ and $\mathcal{L}_4$ is approximate but highly informative, we initially use large values for $\lambda_3,\lambda_4$ , allowing the INR to quickly fit to the rough shape, and then reduce them during training. More training details are in the supplemental material. + +# 4. Results + +Datasets. We evaluate the performance of our method on the ShapeNet [12] and Surface Reconstruction Benchmark (SRB) [7] datasets. ShapeNet has a wide variety of shapes, with thin surfaces, hollow regions and disjoint objects. However the shapes are rather uniformly sampled. We use the preprocessing and split of Neural Splines [45], which consists of 13 shape classes with 20 shapes in each shape class. For each shape it provides the ground truth mesh (preprocessed to be watertight), an input point cloud and occupancy annotation at dense points. SRB contains five shapes from simulated noisy range scans, but it is a popularly used dataset due to the shapes exhibiting many challenging features. These include large missing regions, thin and sharp regions, and small surface details. The former means that good priors in the reconstruction methods, such as smoothness priors, are very important. + +We follow the metrics used in the literature for each dataset: the squared Chamfer distance and IoU for ShapeNet and Chamfer and Hausdorff distance for SRB. + +INRs. We demonstrate our method on two INRs, SIREN [34], and NGLOD [35] (note that NGLOD only works in a supervised setting as it has no global consistency, but our octree guidance is sufficient). Thus we name our two methods OG-SIREN and OG-NGLOD where OG stands for + +"Octree Guided". To extract a mesh we first evaluate our INRs on a $256^3$ grid and then use marching cubes [25]. We use the same settings for the octree and the INRs on both datasets, and give more details in our supplemental. + +# 4.1. Reconstruction Performance + +ShapeNet dataset. The results are shown in Table 1 (left) and visualizations on two shapes are shown in Figure 4. Our methods get the smallest mean squared Chamfer distance out of the unoriented methods by almost an order of magnitude, and are even better than the best performing oriented method. Likewise for mean IoU our methods are better than both unoriented and oriented methods. For median performance, our methods are comparable to the best oriented and unoriented methods. SAP [31] gets the best median IoU, but their mean is much lower than ours, indicating that in a few instances they misrepresent large regions. On the other hand, due to our move-making algorithm we reduce the likelihood of vastly incorrect labellings, and as a result achieve low standard deviations on both metrics despite the high performance. + +SRB dataset. Table 1 (right) shows that DiGS, SAP, PHASE+FF and our two methods do comparably, with DiGS performing slightly better. Note the performance of these methods are comparable to most oriented methods. Also note that the difference between our two methods is more prominent on SRB as the large missing regions means that smoothness priors are more important. Octree guidance is NGLOD's only global supervision, as it fits locally, while SIREN is biased towards smooth shapes. + +# 4.2. Octree Depth Ablations + +Table 4 shows the results of OG-SIREN on ShapeNet when evaluating at different depths of the octree. We include an octree IoU metric, where we compute the IoU of the labelling of the octree directly (taking surface leaves to + +
MethodSquared Chamfer ↓IoU ↑
meanmedianstdmeanmedianstd
SPSR [19]5.44e-51.97e-55.06e-40.99260.99560.0105
IGR [14]5.12e-41.13e-42.15e-30.81020.84800.1519
SIREN [34]1.03e-45.28e-51.93e-40.82680.90970.2329
FFN [36]9.12e-58.65e-53.36e-50.82180.83960.0989
Biharmonic RBF [11]1.11e-48.97e-57.06e-50.82470.86420.1350
SVR [33]1.14e-41.04e-45.99e-50.76250.78190.1300
NSP [45]5.36e-54.06e-53.64e-50.89730.92300.0871
DiGS + n [5]2.74e-42.32e-59.90e-40.92000.97740.1992
N Est+SPSR [19]3.76e-34.37e-51.14e-20.71870.97610.3767
SIREN won [34]3.08e-42.58e-43.26e-40.30850.29520.2014
SAL [1]1.14e-32.11e-43.63e-30.40300.39440.2722
DiGS [5]1.32e-42.55e-54.73e-40.93900.97640.1262
SAP [31]4.09e-42.46e-52.60e-30.91180.99230.2002
Our OG-SIREN3.75e-52.17e-58.24e-50.96150.98710.1048
Our OG-NGLOD5.07e-52.57e-59.39e-50.95930.98700.1057
+ +
MethoddC↓dH↓
SPSR [19]0.214.69
DGP [44]0.215.18
IGR [14]0.192.99
SIREN [34]0.193.86
NSP [45]0.172.85
PHASE [22]0.162.77
DiGS + n [5]0.183.55
N Est.+SPSR [19]1.2522.59
IGR wo n [14]1.3816.33
SIREN wo n [34]0.427.67
SAL [14]0.367.47
IGR+FF [14]0.9611.06
PHASE+FF [22]0.224.96
DiGS [5]0.193.52
SAP [31]0.214.51
Our OG-SIREN0.204.06
Our OG-NGLOD0.226.03
+ +Table 1. Surface Reconstruction results on ShapeNet [12] (left) and SRB [7] (right). Methods above the line use ground truth normal information, and methods below do not. Left: The mean, median and standard deviation of the squared Chamfer distance and IoU of all 260 shapes are reported. Right: The mean Chamfer $d_{C}$ and Hausdorff distance $d_{H}$ for all 5 shapes are reported. + +
MethoddC↓dH↓
SIREN w on [34]0.427.67
OG-SIREN0.204.06
SIREN (+n) [34]0.193.86
OG-SIREN + H°0.192.96
OG-SIREN + GT°0.172.44
OG-NGLOD0.226.03
OG-NGLOD + H°0.195.50
OG-NGLOD + GT°0.175.12
NGLOD+GTd [35]0.131.97
+ +Table 2. Results on SRB [7] when supervising INRs with varying levels of information (see Section 4.3). $\mathrm{GT}^d$ means dense GT SDF supervision, while $\mathrm{GT}^o$ and $\mathrm{H}^o$ mean GT and human estimated occupancy per octree node. + +be inside). The IoU of the octree is much lower than the IoU of the final mesh, indicating that the INR training is learning beyond the octree supervision. Furthermore as the depth increases, the octree IoU improves and so do the metrics on the final mesh, indicating that the INR training is heavily affected by the initial training guidance from the octree. See the supplemental material for a depth ablation on SRB. + +# 4.3. Modifying the octree + +To show the versatility of our method, in Table 2 we demonstrate how modifying the octree labelling can affect the performance of the resulting shape. For $\mathrm{OG - INR + GT^o}$ we set octree labels using GT occupancy. For $\mathrm{OG - INR + H^o}$ we set octree labels using sparse human supervision: we change up to 30 large octree leaves at the lower depths (up + +to depth 5) where the changes are obvious from a human perspective. With both octree supervision signals we get increases in performance, showing that the discrete nature of octree guidance does not introduce much bias. + +We also compare to the INRs with other forms of supervision. As NGLOD cannot be used in an unsupervised setting, we provide result when trained on dense GT. + +# 4.4. Further Comparisons + +The need for unoriented methods. We compare to a recent industry-standard implementation [48] of normal estimation + SPSR [19] (SPSR+N Est.). Table 1 (left) shows that for the well sampled surfaces in ShapeNet, when ground truth normals are available SPSR achieves close to perfect IoUs. However when normals need to be estimated, it has good median performance but very poor average performance, indicating the normal estimation procedure will often be quite inaccurate. Table 1 (right) shows that for point clouds with large missing regions, SPSR with ground truth normals has room for improvement, while SPSR with estimated normals performs extremely poorly. + +While neither dataset has provided line of sight information (as would be given by sensors), a more informative use of such information than estimating rough normals would be to use it to label parts of our octree. See Section 4.3 for how additional octree supervision improves performance. + +Overall Timing Comparison. Table 3 shows the time taken for each method. Our method's gradient descent training is the fastest to converge as it needs much fewer iterations. Even when adding on the average octree building time from ShapeNet, we can see that the time taken for both + +
MethodParametersTime per iter. (s)Num itersTime (s)Speed Up
N Est. +SPSR---42 (13 + 29)12× (· + 18×)
SIREN (wo n)264K0.052100005201×
SAL2.1M0.1751000017500.3×
DiGS264K0.1201000012000.4×
SAP120K-32003301.6×
Our OG-SIREN264K0.158600135 (40 + 95)3.9× (· + 5.5×)
Our OG-NGLOD68.7M0.17330092 (40 + 52)5.7× (· + 10×)
+ +Table 3. Time and number of parameters for each model on ShapeNet. For methods with a precomputation stage (normal estimation or octree building), the time for both is included with the precomputation and training times in parenthesis. Speed up is measured with respect to SIREN/SIREN w o n. We ran our experiments on a single Nvidia RTX 2080 GPU and a i7-8700K CPU. + +
DepthMeanOctree IoU ↑ MedianStdMeanFinal Mesh (OG-SIREN)Squared Chamfer ↓Octree Build + Label Time (s)
IoU ↑ MedianStdMeanMedianStdMeanMedianStd
50.47480.47020.18540.71010.77120.25941.84e-41.35e-42.23e-42.21.62.2
60.64010.64100.16040.95030.98350.11723.93e-52.34e-58.20e-57.65.38.6
70.77570.80060.13050.96150.98710.10483.75e-52.17e-58.24e-538.119.367.7
+ +Table 4. Octree depth ablation on ShapeNet. Octree IoU is computed using the labelling of the leaves (surface leaves chosen as inside). + +stages of our method is still faster than other methods, often by an order of magnitude. We implemented the octree in Cython [3] but further optimisations are possible. + +Comparison Against No Octree Guidance. Comparing SIREN wo n with and without the octree guidance we can see that the guidance drastically improves both time performance and reconstruction metrics (see Table 1,Table 3). When initially guided by our octree labelling, SIREN (wo n) converges significantly faster (92s vs 520s) and gets a 3x improvement in ShapeNet IoU. Without guidance from normals or our octree, SIREN wo n gets stuck in bad local minima where it needs to flip its inside-outsideness prediction, causing ghost geometry (see Figure 1). When comparing the two types of INR training guidance applied to SIREN wo n, ground truth normals vs our octree, we can see that the octree guidance still performs much better than SIREN on ShapeNet, while performing comparably on SRB. + +# 4.5. Discussion + +Limitations. Like all unoriented methods, the octree guidance has difficulty with large missing regions, which is demonstrated heavily in SRB, and thin surfaces, which appear in SRB (Daratech) and ShapeNet (airplane and watercraft). Despite this, our methods perform competitively on shapes with these difficulties on both datasets. See the supplemental for per class breakdown and visualizations. + +Other possible limitations would be the ability to deal with noise and outliers. Note that all methods do poorly with large amounts of either corruption. For small noise the method's performance does not degrade considerably as demonstrated by the results on SRB. For larger noise, the + +volume occupied by surface leaves would be much larger. While this would not introduce error, the octree's benefits would be reduced. For outliers there would be significant bias introduced by the octree method, which if sparse could be preprocessed out beforehand. + +Broader Impact. Faster and more accurate unoriented surface reconstruction methods will allow for accurate meshes to be constructed easily from regular scanners. Potential misuses of this technology include unauthorized replication and harmful content creation. + +# 5. Conclusion + +We have introduced an octree guided method for training implicit neural representations (INRs) for the task of unoriented surface reconstruction. We formulate the main difficulty of the task, inside-outside determinism, as an octree labelling problem with respect to an energy function that captures watertight surface properties and the minimal surface constraint. We also detail an efficient method to minimize the energy. We then use this supervision to guide the initial stage of INR training, using loss functions that can be used by any INR architecture. Our results show that applying our octree guided supervision provides significant time and reconstruction performance improvements, with competitive or better results than the current state-of-the-art. + +Acknowledgements. This work has received funding from the European Union's Horizon 2020 research and innovation programme under the Marie Sklodowska-Curie grant agreement No. 893465. S. Gould is a recipient of an ARC Future Fellowship (proj. no. LP200100421) funded by the Australian Government. + +# References + +[1] Matan Atzmon and Yaron Lipman. SAL: Sign Agnostic Learning of shapes from raw data. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2565-2574, 2020. 2, 3, 5, 6, 7 +[2] Matan Atzmon and Yaron Lipman. SALD: Sign Agnostic Learning with Derivatives. In Proc. of the International Conference on Learning Representations (ICLR), 2021. 2, 3, 5 +[3] Stefan Behnel, Robert Bradshaw, Craig Citro, Lisandro Dalcin, Dag Sverre Seljebotn, and Kurt Smith. Cython: The best of both worlds. Computing in Science & Engineering, 13(2):31-39, 2011. 8 +[4] Yizhak Ben-Shabat and Stephen Gould. DeepFit: 3D surface fitting via neural network weighted least squares. In Proc. of the European Conference on Computer Vision (ECCV), pages 20-34. Springer, 2020. 1 +[5] Yizhak Ben-Shabat, Chamin Hewa Koneputugodage, and Stephen Gould. Digs: Divergence guided shape implicit neural representation for unoriented point clouds. arXiv preprint arXiv:2106.10811, 2021. 2, 3, 5, 6, 7 +[6] Yizhak Ben-Shabat, Michael Lindenbaum, and Anath Fischer. Nesti-net: Normal estimation for unstructured 3D point clouds using convolutional neural networks. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 10112-10120, 2019. 1 +[7] Matthew Berger, Joshua A Levine, Luis Gustavo Nonato, Gabriel Taubin, and Claudio T Silva. A benchmark for surface reconstruction. ACM Trans. on Graphics (ToG), 32:1-17, 2013. 6, 7 +[8] Julian Besag. On the statistical analysis of dirty pictures. Journal of the Royal Statistical Society: Series B (Methodological), 48(3):259-279, 1986. 3 +[9] Andrew Blake, Pushmeet Kohli, and Carsten Rother. Markov Random Fields for Vision and Image Processing. The MIT Press, 07 2011. 3, 5 +[10] Yuri Boykov, Olga Veksler, and Ramin Zabih. Fast approximate energy minimization via graph cuts. IEEE Transactions on pattern analysis and machine intelligence, 23(11):1222-1239, 2001. 3 +[11] Jonathan C Carr, Richard K Beatson, Jon B Cherrie, Tim J Mitchell, W Richard Fright, Bruce C McCallum, and Tim R Evans. Reconstruction and representation of 3D objects with radial basis functions. In Proc. of the Conference on Computer Graphics and Interactive Techniques, pages 67-76, 2001. 7 +[12] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. ShapeNet: An information-rich 3D model repository. arXiv preprint arXiv:1512.03012, 2015. 6, 7 +[13] Stephen Gould. Max-margin learning for lower linear envelope potentials in binary markov random fields. In Proc. of the International Conference on Machine Learning (ICML). Omnipress, 2011. 5 +[14] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit Geometric Regularization for learning shapes. In Proc. of the International Conference on + +Machine Learning (ICML), volume 119, pages 3789-3799. +PMLR, 2020. 2, 3, 5, 7 +[15] Paul Guerrero, Yanir Kleiman, Maks Ovsjanikov, and Niloy J Mitra. PCPNet: Learning local shape properties from raw point clouds. In Computer Graphics Forum, volume 37, pages 75-85. Wiley Online Library, 2018. 1 +[16] Christian Hane, Shubham Tulsiani, and Jitendra Malik. Hierarchical surface prediction for 3d object reconstruction. In 2017 International Conference on 3D Vision (3DV), pages 412-420. IEEE, 2017. 2 +[17] Shahram Izadi, David Kim, Otmar Hilliges, David Molyneaux, Richard Newcombe, Pushmeet Kohli, Jamie Shotton, Steve Hodges, Dustin Freeman, Andrew Davison, and Andrew Fitzgibbon. Kinectfusion: Real-time 3d reconstruction and interaction using a moving depth camera. In UIST '11 Proceedings of the 24th annual ACM symposium on User interface software and technology, pages 559-568. ACM, October 2011. 1 +[18] Michael Kazhdan, Matthew Bolitho, and Hugues Hoppe. Poisson surface reconstruction. In Proc. of the Eurographics Symposium on Geometry Processing (SGP), volume 7, 2006. 2 +[19] Michael Kazhdan and Hugues Hoppe. Screened poisson surface reconstruction. ACM Trans. on Graphics (ToG), 32:1-13, 2013. 2, 7 +[20] Pushmeet Kohli, M. Kumar, and Philip Torr. P3 & beyond: Solving energies with higher order cliques. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2007. 5 +[21] Jan Eric Lenssen, Christian Osendorfer, and Jonathan Masci. Deep iterative surface normal estimation. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 11247-11256, 2020. 1 +[22] Yaron Lipman. Phase transitions, distance functions, and implicit neural representations. arXiv preprint arXiv:2106.07689, 2021. 2, 3, 7 +[23] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. Advances in Neural Information Processing Systems, 33:15651-15663, 2020. 2 +[24] Shi-Lin Liu, Hao-Xiang Guo, Hao Pan, Peng-Shuai Wang, Xin Tong, and Yang Liu. Deep implicit moving least-squares functions for 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1788-1797, 2021. 3 +[25] William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3D surface construction algorithm. In Proc. of the Intl. Conference on Computer Graphics and Interactive Techniques (SIGGRAPH), volume 21, pages 163-169. ACM New York, NY, USA, 1987. 6 +[26] Julien NP Martel, David B Lindell, Connor Z Lin, Eric R Chan, Marco Monteiro, and Gordon Wetzstein. Acorn: Adaptive coordinate networks for neural scene representation. arXiv preprint arXiv:2105.02788, 2021. 3 +[27] Donald J. Meagher. Geometric modeling using octree encoding. Comput. Graph. Image Process., 19:85, 1982. 2 +[28] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: + +Learning 3D reconstruction in function space. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4460-4470, 2019. 1 +[29] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, July 2022. 1, 2 +[30] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. DeepSDF: Learning continuous signed distance functions for shape representation. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 165-174, 2019. 1, 2 +[31] Songyou Peng, Chiyu Jiang, Yiyi Liao, Michael Niemeyer, Marc Pollefeys, and Andreas Geiger. Shape As Points: A differentiable poisson solver. arXiv preprint arXiv:2106.03452, 2021. 3, 6, 7 +[32] Gernot Riegler, Ali Osman Ulusoy, and Andreas Geiger. Octnet: Learning deep 3d representations at high resolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3577-3586, 2017. 2 +[33] Bernhard Scholkopf, Joachim Giesen, and Simon Spalinger. Kernel methods for implicit surface modeling. In Advances in Neural Information Processing Systems (NeurIPS), NIPS'04, page 1193-1200, Cambridge, MA, USA, 2004. MIT Press. 7 +[34] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. Advances in Neural Information Processing Systems (NeurIPS), 33, 2020. 2, 5, 6, 7 +[35] Towaki Takikawa, Joey Litalien, Kangxue Yin, Karsten Kreis, Charles Loop, Derek Nowrouzezahrai, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Neural geometric level of detail: Real-time rendering with implicit 3D shapes. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 2, 3, 6, 7 +[36] Matthew Tancik, Pratul P. Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan T. Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. Advances in Neural Information Processing Systems (NeurIPS), 2020. 2, 7 +[37] Jia-Heng Tang, Weikai Chen, Jie Yang, Bo Wang, Songrun Liu, Bo Yang, and Lin Gao. Octfield: Hierarchical implicit functions for 3d modeling. In The Thirty-Fifth Annual Conference on Neural Information Processing Systems (NeurIPS), 2021. 2 +[38] Maxim Tatarchenko, Alexey Dosovitskiy, and Thomas Brox. Octree generating networks: Efficient convolutional architectures for high-resolution 3d outputs. In Proceedings of the IEEE international conference on computer vision, pages 2088-2096, 2017. 2 +[39] Ayush Tewari, Justus Thies, Ben Mildenhall, Pratul Srinivasan, Edgar Tretschk, Yifan Wang, Christoph Lassner, Vincent Sitzmann, Ricardo Martin-Brualla, Stephen Lombardi, et al. Advances in neural rendering. arXiv preprint arXiv:2111.05849, 2021. 1 + +[40] Peng-Shuai Wang, Yang Liu, Yu-Xiao Guo, Chun-Yu Sun, and Xin Tong. O-cnn: Octree-based convolutional neural networks for 3d shape analysis. ACM Transactions On Graphics (TOG), 36(4):1-11, 2017. 2 +[41] Peng-Shuai Wang, Yang Liu, and Xin Tong. Deep octree-based cnns with output-guided skip connections for 3d shape and scene completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 266-267, 2020. 2 +[42] Peng-Shuai Wang, Yang Liu, and Xin Tong. Dual octree graph networks for learning adaptive volumetric shape representations. ACM Transactions on Graphics (TOG), 41(4):1-15, 2022. 3 +[43] Peng-Shuai Wang, Chun-Yu Sun, Yang Liu, and Xin Tong. Adaptive o-cnn: A patch-based deep representation of 3d shapes. ACM Transactions on Graphics (TOG), 37(6):1-11, 2018. 2 +[44] Francis Williams, Teseo Schneider, Claudio Silva, Denis Zorin, Joan Bruna, and Daniele Panozzo. Deep Geometric Prior for surface reconstruction. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 10130-10139, 2019. 7 +[45] Francis Williams, Matthew Trager, Joan Bruna, and Denis Zorin. Neural Splines: Fitting 3D surfaces with infinitely-wide neural networks. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 9949-9958, 2021. 2, 6, 7 +[46] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond. Computer Graphics Forum, 2022. 1 +[47] Wang Yifan, Lukas Rahmann, and Olga Sorkine-Hornung. Geometry-consistent neural shape representation with implicit displacement fields. arXiv preprint arXiv:2106.05187, 2021. 3 +[48] Qian-Yi Zhou, Jaesik Park, and Vladlen Koltun. Open3D: A modern library for 3D data processing. arXiv:1801.09847, 2018. 7 \ No newline at end of file diff --git a/2023/Octree Guided Unoriented Surface Reconstruction/images.zip b/2023/Octree Guided Unoriented Surface Reconstruction/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..eff19fa6d8d80f30e510ddb228245517f740be46 --- /dev/null +++ b/2023/Octree Guided Unoriented Surface Reconstruction/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ee6d49f59e165d9ef12d5c2e633e0321bf5ac47dbf6801423bc7ae4dc6ae828a +size 541837 diff --git a/2023/Octree Guided Unoriented Surface Reconstruction/layout.json b/2023/Octree Guided Unoriented Surface Reconstruction/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..bfc2f69e9a038e0eafafe508f24d6b174adb8d4f --- /dev/null +++ b/2023/Octree Guided Unoriented Surface Reconstruction/layout.json @@ -0,0 +1,9998 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 140, + 103, + 454, + 119 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 103, + 454, + 119 + ], + "spans": [ + { + "bbox": [ + 140, + 103, + 454, + 119 + ], + "type": "text", + "content": "Octree Guided Unoriented Surface Reconstruction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 99, + 143, + 259, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 143, + 259, + 159 + ], + "spans": [ + { + "bbox": [ + 99, + 143, + 259, + 159 + ], + "type": "text", + "content": "Chamin Hewa Koneputugodage1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 285, + 143, + 391, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 285, + 143, + 391, + 156 + ], + "spans": [ + { + "bbox": [ + 285, + 143, + 391, + 156 + ], + "type": "text", + "content": "Yizhak Ben-Shabat1,2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 414, + 143, + 493, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 414, + 143, + 493, + 157 + ], + "spans": [ + { + "bbox": [ + 414, + 143, + 493, + 157 + ], + "type": "text", + "content": "Stephen Gould" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 165, + 282, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 165, + 282, + 179 + ], + "spans": [ + { + "bbox": [ + 106, + 165, + 282, + 179 + ], + "type": "text", + "content": "1The Australian National University" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 294, + 165, + 490, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 165, + 490, + 180 + ], + "spans": [ + { + "bbox": [ + 294, + 165, + 490, + 180 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 294, + 165, + 490, + 180 + ], + "type": "text", + "content": "Technion Israel Institute of Technology" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 182, + 453, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 182, + 453, + 194 + ], + "spans": [ + { + "bbox": [ + 138, + 182, + 453, + 194 + ], + "type": "text", + "content": "{chamin.hewa, stephen.gould}@anu.edu.au sitszikbs@gmail.com" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 143, + 221, + 192, + 234 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 221, + 192, + 234 + ], + "spans": [ + { + "bbox": [ + 143, + 221, + 192, + 234 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 248, + 290, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 248, + 290, + 536 + ], + "spans": [ + { + "bbox": [ + 46, + 248, + 290, + 536 + ], + "type": "text", + "content": "We address the problem of surface reconstruction from unoriented point clouds. Implicit neural representations (INRs) have become popular for this task, but when information relating to the inside versus outside of a shape is not available (such as shape occupancy, signed distances or surface normal orientation) optimization relies on heuristics and regularizers to recover the surface. These methods can be slow to converge and easily get stuck in local minima. We propose a two-step approach, OG-INR, where we (1) construct a discrete octree and label what is inside and outside (2) optimize for a continuous and high-fidelity shape using an INR that is initially guided by the octree's labelling. To solve for our labelling, we propose an energy function over the discrete structure and provide an efficient move-making algorithm that explores many possible labellings. Furthermore we show that we can easily inject knowledge into the discrete octree, providing a simple way to influence the result from the continuous INR. We evaluate the effectiveness of our approach on two unoriented surface reconstruction datasets and show competitive performance compared to other unoriented, and some oriented, methods. Our results show that the exploration by the move-making algorithm avoids many of the bad local minima reached by purely gradient descent optimized methods (see Figure 1)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 560, + 128, + 573 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 560, + 128, + 573 + ], + "spans": [ + { + "bbox": [ + 47, + 560, + 128, + 573 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 715 + ], + "type": "text", + "content": "Surface reconstruction from 3D point clouds has been studied extensively in computer vision and computer graphics. The task requires estimating a mesh of the shape's surface from a point cloud sampled from the surface of the shape. We focus on the reconstruction of watertight 3D shapes, i.e., shapes that have a well defined interior and exterior. Such shapes are often represented as signed distance fields (SDFs) or occupancy fields, which can be efficiently encoded by a neural network [28, 30]. As these representations are fields parameterized by neural networks, they are often called neural fields. Furthermore, they implicitly rep" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 308, + 220, + 548, + 315 + ], + "blocks": [ + { + "bbox": [ + 308, + 220, + 548, + 315 + ], + "lines": [ + { + "bbox": [ + 308, + 220, + 548, + 315 + ], + "spans": [ + { + "bbox": [ + 308, + 220, + 548, + 315 + ], + "type": "image", + "image_path": "5de864bc5d353c490ae1de645706df3dd2d6590d7777bb1840f60601319edfee.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 324, + 547, + 402 + ], + "lines": [ + { + "bbox": [ + 304, + 324, + 547, + 402 + ], + "spans": [ + { + "bbox": [ + 304, + 324, + 547, + 402 + ], + "type": "text", + "content": "Figure 1. Our method with a SIREN INR (OG-SIREN, left) compared to SIREN wo n (right) for two shapes from the ShapeNet dataset. Our octree guidance allows for consistent inside-outside determinism. On the other hand, SIREN wo n gets stuck in local minima from which it cannot escape (due to needing to completely change the occupancy of certain areas) creating extraneous surfaces (often called ghost geometries in the literature)." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 425, + 545, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 425, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 425, + 545, + 449 + ], + "type": "text", + "content": "resent the shape by a level set of the field, thus they are also often referred to as implicit neural representations (INRs)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 449, + 545, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 545, + 545 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 545, + 545 + ], + "type": "text", + "content": "The broader class of neural fields, including INRs, have been very popular over the last few years as they can handle arbitrary topology, are memory efficient, and are continuous with potentially infinite resolution [39, 46]. Among the INRs for 3D shapes, SDFs are the most popular as they provide more useful information (distances not just occupancy) and are required for downstream graphics algorithms such as sphere tracing and approximate soft shadows [29, 35]." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 546, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 547, + 714 + ], + "type": "text", + "content": "When learning an implicit representation of a shape, a major difficulty is predicting whether points in space are inside or outside the shape. Many methods require oriented surface normals for the input points or signed distances from the surface, which usually are not given by raw data from scans. While this can be estimated using the line of sight information [17] or algorithms [4,6, 15, 21], they yield noisy predictions that after postprocessing still can lead to bad results (see Section 4.4). We consider the task of unoriented surface reconstruction, where such information is not available, and only the sampled surface points are given. We demonstrate that our method performs competitively and sometimes better than oriented methods, even when they are given the ground truth (GT) normals." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16717" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 76, + 541, + 185 + ], + "blocks": [ + { + "bbox": [ + 52, + 76, + 541, + 185 + ], + "lines": [ + { + "bbox": [ + 52, + 76, + 541, + 185 + ], + "spans": [ + { + "bbox": [ + 52, + 76, + 541, + 185 + ], + "type": "image", + "image_path": "4c2eb5865acd0522fff845a067abf9d689f85cc689a884b86aabb874c412093c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 203, + 547, + 248 + ], + "lines": [ + { + "bbox": [ + 46, + 203, + 547, + 248 + ], + "spans": [ + { + "bbox": [ + 46, + 203, + 547, + 248 + ], + "type": "text", + "content": "Figure 2. An illustration of our method, OG-INR. Given an unoriented point cloud (left), we progressively build and label an octree around the points (middle). The octree at depths 3-7 are shown. Surface leaves of the octree (yellow) are leaves that contain points from the point cloud, other leaves are labelled as inside (blue) or outside (transparent) by minimizing an energy function. We then train an INR model to obtain an SDF, using the labelling as supervision for the initial training, after which we can extract a mesh (right)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 269, + 289, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 269, + 289, + 437 + ], + "spans": [ + { + "bbox": [ + 46, + 269, + 289, + 437 + ], + "type": "text", + "content": "We propose OG-INR, which uses a discrete representation in conjunction with the continuous INR representation (see Figure 2). Given an input unoriented point cloud, we progressively build an octree from the input points and determine which leaf nodes are surface leaves (i.e., leaves that contain a point). We also label all other leaves as inside or outside leaves (if they should be within or outside the shape respectively). To do this we minimize an energy function that trades off the watertight surface property, that every surface point should border with the inside and outside of the shape, with a minimal surface constraint. We then train standard INR architectures, initially guiding the training by the octree labelling, and show that the INR converges much faster and is less prone to large failure regions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 437, + 173, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 437, + 173, + 448 + ], + "spans": [ + { + "bbox": [ + 59, + 437, + 173, + 448 + ], + "type": "text", + "content": "Our main contributions are:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 449, + 287, + 621 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 58, + 449, + 287, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 449, + 287, + 496 + ], + "spans": [ + { + "bbox": [ + 58, + 449, + 287, + 496 + ], + "type": "text", + "content": "- We introduce a novel method to guide the initial stage of INR training. It uses a labelled octree structure to allow the INR to converge significantly faster and alleviates the local minima problem of INRs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 498, + 287, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 498, + 287, + 545 + ], + "spans": [ + { + "bbox": [ + 58, + 498, + 287, + 545 + ], + "type": "text", + "content": "- We propose an energy function over octree labels that captures the task of surface reconstruction. It balances the constraint of maintaining known surface regions with minimising the overall surface area." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 58, + 548, + 287, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 548, + 287, + 583 + ], + "spans": [ + { + "bbox": [ + 58, + 548, + 287, + 583 + ], + "type": "text", + "content": "- We provide an efficient move-making algorithm to optimize the energy function, which explores many inside-outside possibilities in a structured manner." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 58, + 586, + 287, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 586, + 287, + 621 + ], + "spans": [ + { + "bbox": [ + 58, + 586, + 287, + 621 + ], + "type": "text", + "content": "- Our discrete representation is easily understandable and human-modifiable, giving an intuitive method for applying changes to the resulting SDF." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 633, + 134, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 633, + 134, + 645 + ], + "spans": [ + { + "bbox": [ + 47, + 633, + 134, + 645 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 654, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 287, + 715 + ], + "type": "text", + "content": "Unsupervised Surface Reconstruction. Representing shapes with INRs was popularized by DeepSDF [30], where a ReLU network directly models an SDF using GT SDF supervision. SAL [1] instead uses unsigned distance supervision, approximated from the input point cloud, as an" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 269, + 545, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 269, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 304, + 269, + 545, + 365 + ], + "type": "text", + "content": "unsupervised and unoriented approach. SALD [2] showed that adding unoriented normal supervision improved upon this, and IGR [14] further added the eikonal constraint loss (see Equation 5) that has been widely adopted [5, 22, 34]. NSP [45] addresses the task by solving a kernel regression problem but needs input normals. PHASE [22] introduces a novel loss inspired by the phase transition problem that encourages boundary crossings at the surface points." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 365, + 546, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 365, + 546, + 496 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 546, + 496 + ], + "type": "text", + "content": "However, as the inputs to INRs are low dimensional, they are biased to low frequency solutions that do not recover high fidelity detail [36]. To introduce high frequencies, SIREN [34] uses sine activations and Fourier Feature Networks [36] use a fixed Fourier encoding. While being able to reconstruct high fidelity shapes with a modest number of parameters, these methods are hard to train as the high frequencies overfit quickly. A different approach is parametric encoding [29, 35], where spatial regions are represented by trainable parameters (often using an octree structure). We cover such methods in the next subsection." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 497, + 545, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 497, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 305, + 497, + 545, + 520 + ], + "type": "text", + "content": "Note that IGR and SIREN can operate without normals, which we refer to by \"IGR wo n\" and \"SIREN wo n\"." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 521, + 545, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 592 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 592 + ], + "type": "text", + "content": "Octrees for Surface Reconstruction. Octrees were first introduced by Meagher [27] as an efficient data structure for representing and manipulating 3D data, and are now used extensively. The classic method Poisson Surface Reconstruction (PSR) [18] and its followup SPSR [19] use octrees to structure their implicit functions." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 594, + 545, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 677 + ], + "type": "text", + "content": "Early works on using octrees with neural networks include HSP [16] and OGN [38], who train encoder-decoder networks to learn octrees with occupancy values, and OctNet [32] and O-CNN [40], who implement 3D CNNs that operate over octrees rather than voxel grids for shape analysis tasks. O-CNN's descendants [41, 43] build on this and adapt it to shape generation, completion and reconstruction." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": "More recently octrees have been used in INRs for representing fine details efficiently. OctField [37] and NSVF [23] both learn local implicit functions for individual octree" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16718" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 133 + ], + "type": "text", + "content": "nodes. NGLOD [35] and ACORN [26] instead learn a single INR, but store trainable parameters at octree node vertices to encode spatial regions. IMLSNet [24] uses O-CNN to predict both an octree and oriented points per octree node to represent the surface using implicit moving least squares." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 134, + 286, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 134, + 286, + 217 + ], + "spans": [ + { + "bbox": [ + 46, + 134, + 286, + 217 + ], + "type": "text", + "content": "However, previous octree reconstruction methods were supervised, trained on GT SDF/occupancy of either the current shape or a separate training set. Recently Dual Octree Graph CNNs [42] addressed the unsupervised setting, where they learn local features on the dual graph of an octree. However their method has parameters that need unsupervised training, and they require input normals." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 219, + 286, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 219, + 286, + 315 + ], + "spans": [ + { + "bbox": [ + 46, + 219, + 286, + 315 + ], + "type": "text", + "content": "SAP [31] truly addresses the unsupervised and unoriented task by making the PSR pipeline differentiable. Starting from an oriented point cloud of a sphere, they use Chamfer loss to guide that point cloud to a dense, oriented representation. Their method is fast and among state-of-the-art, however it can fail drastically and often does not produce a watertight mesh (see Section 4). It is also hard to adapt to different applications and losses, unlike INRs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 316, + 286, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 316, + 286, + 449 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 286, + 449 + ], + "type": "text", + "content": "Initialization for INRs. Due to the difficulty of training INRs, initialization is very important [1, 5]. However regardless of input, most methods initialize to a sphere. These include SAL [1] and its descendants [2, 14, 22], DiGS [5] and IDF [47]. This gives many important properties, such as biasing shape interior to the domain's center and satisfying the eikonal condition at the start of training, but it is problematic for shapes with sharp concavities. Our method, on the other hand, provides firm guidance for both the eikonal constraint and the shape based on the input point cloud, and thus does not bias the training toward convex objects." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 450, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 450, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 450, + 287, + 713 + ], + "type": "text", + "content": "Energy Minimization. Discrete energy minimization has been used extensively in both 2D and 3D vision for segmentation and object labelling [9, 10]. An energy function encodes our preference over the joint assignment to a set of variables, e.g., binary variables representing occupancy of voxels in 3D space. The goal of energy minimization is to find an assignment to the variables with minimal cost. This is NP-hard in general and a number of approximation algorithms have been proposed [9]. One class of algorithms, known as move-making, starts with an initial assignment and iteratively evaluates changes to a subset of the variables (called a move). If the change results in lower energy then the move is accepted, otherwise the changed values are reverted. Moves are repeatedly tried until no further reduction in energy is found. Iterated Conditional Modes (ICM) [8] is the simplest example, where each move involves changing a single variable and so is very efficient, but it is prone to getting stuck in local minima. Generalized ICM improves on this by allowing multiple variables to change in a move, but it is very costly. Our method is a variant of generalized ICM specialized to our problem, where we introduce restrictions on the move space improve efficiency." + } + ] + } + ], + "index": 4 + }, + { + "type": "code", + "bbox": [ + 310, + 86, + 545, + 258 + ], + "blocks": [ + { + "bbox": [ + 306, + 72, + 496, + 84 + ], + "lines": [ + { + "bbox": [ + 306, + 72, + 496, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 496, + 84 + ], + "type": "text", + "content": "Algorithm 1 Overall octree labeling algorithm" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 310, + 86, + 545, + 258 + ], + "lines": [ + { + "bbox": [ + 310, + 86, + 545, + 258 + ], + "spans": [ + { + "bbox": [ + 310, + 86, + 545, + 258 + ], + "type": "text", + "content": "1: procedure LABELOCTREE " + }, + { + "bbox": [ + 310, + 86, + 545, + 258 + ], + "type": "inline_equation", + "content": "(\\chi ,d_i,k,C)" + }, + { + "bbox": [ + 310, + 86, + 545, + 258 + ], + "type": "text", + "content": " \n2: " + }, + { + "bbox": [ + 310, + 86, + 545, + 258 + ], + "type": "inline_equation", + "content": "\\mathcal{O} =" + }, + { + "bbox": [ + 310, + 86, + 545, + 258 + ], + "type": "text", + "content": " CreateOctree " + }, + { + "bbox": [ + 310, + 86, + 545, + 258 + ], + "type": "inline_equation", + "content": "(\\chi ,d_{i})" + }, + { + "bbox": [ + 310, + 86, + 545, + 258 + ], + "type": "text", + "content": " \n3: Initialization(O) \n4: for " + }, + { + "bbox": [ + 310, + 86, + 545, + 258 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 310, + 86, + 545, + 258 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 310, + 86, + 545, + 258 + ], + "type": "inline_equation", + "content": "0,\\dots ,k" + }, + { + "bbox": [ + 310, + 86, + 545, + 258 + ], + "type": "text", + "content": " do \n5: if " + }, + { + "bbox": [ + 310, + 86, + 545, + 258 + ], + "type": "inline_equation", + "content": "j\\neq 0" + }, + { + "bbox": [ + 310, + 86, + 545, + 258 + ], + "type": "text", + "content": " then \n6: ExpandTree(O) \n7: end if ▷ Now at depth " + }, + { + "bbox": [ + 310, + 86, + 545, + 258 + ], + "type": "inline_equation", + "content": "d = d_{i} + j" + }, + { + "bbox": [ + 310, + 86, + 545, + 258 + ], + "type": "text", + "content": " \n8: Grow(O) \n9: for " + }, + { + "bbox": [ + 310, + 86, + 545, + 258 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 310, + 86, + 545, + 258 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 310, + 86, + 545, + 258 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 310, + 86, + 545, + 258 + ], + "type": "text", + "content": " do \n10: makeMove(O,s) \n11: end for \n12: end for \n13: return O \n14: end procedure" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "code_body" + } + ], + "index": 6, + "sub_type": "algorithm" + }, + { + "bbox": [ + 305, + 281, + 511, + 296 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 281, + 511, + 296 + ], + "spans": [ + { + "bbox": [ + 305, + 281, + 511, + 296 + ], + "type": "text", + "content": "3. OG-INR: Octree Guided Shape INRs" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 304, + 430, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 304, + 430, + 316 + ], + "spans": [ + { + "bbox": [ + 306, + 304, + 430, + 316 + ], + "type": "text", + "content": "3.1. Problem Formulation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 323, + 545, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 323, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 305, + 323, + 545, + 396 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 305, + 323, + 545, + 396 + ], + "type": "inline_equation", + "content": "\\chi = \\{x_{i}\\}_{i\\in I}" + }, + { + "bbox": [ + 305, + 323, + 545, + 396 + ], + "type": "text", + "content": " be a point cloud sampled from a surface " + }, + { + "bbox": [ + 305, + 323, + 545, + 396 + ], + "type": "inline_equation", + "content": "S = \\partial \\mathcal{V}" + }, + { + "bbox": [ + 305, + 323, + 545, + 396 + ], + "type": "text", + "content": " of some watertight volume " + }, + { + "bbox": [ + 305, + 323, + 545, + 396 + ], + "type": "inline_equation", + "content": "\\mathcal{V} \\subset \\mathcal{D}" + }, + { + "bbox": [ + 305, + 323, + 545, + 396 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 305, + 323, + 545, + 396 + ], + "type": "inline_equation", + "content": "\\mathcal{D} \\subset \\mathbb{R}^3" + }, + { + "bbox": [ + 305, + 323, + 545, + 396 + ], + "type": "text", + "content": " is compact. We wish to reconstruct the surface " + }, + { + "bbox": [ + 305, + 323, + 545, + 396 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 305, + 323, + 545, + 396 + ], + "type": "text", + "content": ". We choose to represent " + }, + { + "bbox": [ + 305, + 323, + 545, + 396 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 305, + 323, + 545, + 396 + ], + "type": "text", + "content": " as the zero-level set of a signed distance function (SDF), and thus wish to train a neural network " + }, + { + "bbox": [ + 305, + 323, + 545, + 396 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 305, + 323, + 545, + 396 + ], + "type": "text", + "content": " to output signed distances for all points " + }, + { + "bbox": [ + 305, + 323, + 545, + 396 + ], + "type": "inline_equation", + "content": "x \\in \\mathcal{D}" + }, + { + "bbox": [ + 305, + 323, + 545, + 396 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 369, + 408, + 545, + 423 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 408, + 545, + 423 + ], + "spans": [ + { + "bbox": [ + 369, + 408, + 545, + 423 + ], + "type": "interline_equation", + "content": "\\Phi (x, \\theta) = (- 1) ^ {o (x)} d (x, \\mathcal {S}) \\tag {1}", + "image_path": "0f71814202a927e626cf4585a2f3687dedd1cb87958dd7de4f734a740b221a3b.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 436, + 545, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 436, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 305, + 436, + 545, + 472 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 436, + 545, + 472 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 305, + 436, + 545, + 472 + ], + "type": "text", + "content": " are the model parameters, " + }, + { + "bbox": [ + 305, + 436, + 545, + 472 + ], + "type": "inline_equation", + "content": "o(x)" + }, + { + "bbox": [ + 305, + 436, + 545, + 472 + ], + "type": "text", + "content": " is the occupancy of point x (1 if inside and 0 if outside) and " + }, + { + "bbox": [ + 305, + 436, + 545, + 472 + ], + "type": "inline_equation", + "content": "d(x,S)" + }, + { + "bbox": [ + 305, + 436, + 545, + 472 + ], + "type": "text", + "content": " is the distance from " + }, + { + "bbox": [ + 305, + 436, + 545, + 472 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 305, + 436, + 545, + 472 + ], + "type": "text", + "content": " to its closest point on " + }, + { + "bbox": [ + 305, + 436, + 545, + 472 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 305, + 436, + 545, + 472 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 485, + 373, + 497 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 485, + 373, + 497 + ], + "spans": [ + { + "bbox": [ + 306, + 485, + 373, + 497 + ], + "type": "text", + "content": "3.2. Overview" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 506, + 545, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 506, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 304, + 506, + 545, + 601 + ], + "type": "text", + "content": "We first aim to produce a rough and discrete, but globally consistent, occupancy determination for our domain by solving for an inside-outside labelling on nodes of an octree. To do this we propose an energy function that combines both the watertight surface property and the minimal surface objective. Despite this energy being intractable to solve exactly, we provide a domain-guided move-making algorithm to get low energy solutions (see Section 3.3)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 604, + 545, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 604, + 545, + 675 + ], + "spans": [ + { + "bbox": [ + 304, + 604, + 545, + 675 + ], + "type": "text", + "content": "We then design additional INR loss terms to guide INRs using our octree labelling (see Section 3.4). These losses quickly guide the INR towards a rough but globally consistent representation that allows it to avoid many bad local minima. Since these loss terms are agnostic to INR architecture, our method can work with any INR model." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "As we assume our domain is compact, we first normalize the point cloud to fit within the unit sphere in 3D, and then set our domain as " + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = [-1.1, 1.1]^3" + }, + { + "bbox": [ + 304, + 677, + 545, + 713 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16719" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 97, + 70, + 195, + 156 + ], + "blocks": [ + { + "bbox": [ + 97, + 70, + 195, + 156 + ], + "lines": [ + { + "bbox": [ + 97, + 70, + 195, + 156 + ], + "spans": [ + { + "bbox": [ + 97, + 70, + 195, + 156 + ], + "type": "image", + "image_path": "f8e6d3e2c77f83f9c633127372ee809593877ed8787d6deaacee1dc63003561e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 99, + 156, + 187, + 164 + ], + "lines": [ + { + "bbox": [ + 99, + 156, + 187, + 164 + ], + "spans": [ + { + "bbox": [ + 99, + 156, + 187, + 164 + ], + "type": "text", + "content": "A: Depth 3, After Initialization(O)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 196, + 70, + 293, + 156 + ], + "blocks": [ + { + "bbox": [ + 196, + 70, + 293, + 156 + ], + "lines": [ + { + "bbox": [ + 196, + 70, + 293, + 156 + ], + "spans": [ + { + "bbox": [ + 196, + 70, + 293, + 156 + ], + "type": "image", + "image_path": "d0e11320b3b5133d63e407682efb364864df0910d567f0ab41091ea9cfbbf82a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 198, + 157, + 267, + 165 + ], + "lines": [ + { + "bbox": [ + 198, + 157, + 267, + 165 + ], + "spans": [ + { + "bbox": [ + 198, + 157, + 267, + 165 + ], + "type": "text", + "content": "B: Depth 3, After Grow(O)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 294, + 70, + 392, + 156 + ], + "blocks": [ + { + "bbox": [ + 294, + 70, + 392, + 156 + ], + "lines": [ + { + "bbox": [ + 294, + 70, + 392, + 156 + ], + "spans": [ + { + "bbox": [ + 294, + 70, + 392, + 156 + ], + "type": "image", + "image_path": "6792545c03f90c3363a9f1c34b9830756db7c86ee42e7b34da3a1c4ad4ae3dba.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 296, + 157, + 391, + 165 + ], + "lines": [ + { + "bbox": [ + 296, + 157, + 391, + 165 + ], + "spans": [ + { + "bbox": [ + 296, + 157, + 391, + 165 + ], + "type": "text", + "content": "C: Depth 3, After all makeMove(O,s)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 398, + 70, + 496, + 156 + ], + "blocks": [ + { + "bbox": [ + 398, + 70, + 496, + 156 + ], + "lines": [ + { + "bbox": [ + 398, + 70, + 496, + 156 + ], + "spans": [ + { + "bbox": [ + 398, + 70, + 496, + 156 + ], + "type": "image", + "image_path": "7a1ba188c9c51b656e16c8c46a2a344d98b0f719b8159361c13834b4468f142b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 399, + 157, + 493, + 165 + ], + "lines": [ + { + "bbox": [ + 399, + 157, + 493, + 165 + ], + "spans": [ + { + "bbox": [ + 399, + 157, + 493, + 165 + ], + "type": "text", + "content": "D: Depth 4, After all makeMove(O, s)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 97, + 171, + 195, + 255 + ], + "blocks": [ + { + "bbox": [ + 97, + 171, + 195, + 255 + ], + "lines": [ + { + "bbox": [ + 97, + 171, + 195, + 255 + ], + "spans": [ + { + "bbox": [ + 97, + 171, + 195, + 255 + ], + "type": "image", + "image_path": "c74dd1814d045289fb252dafcc233062c15cbde10bf3e05c29cbb16e54837a3d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 100, + 258, + 167, + 266 + ], + "lines": [ + { + "bbox": [ + 100, + 258, + 167, + 266 + ], + "spans": [ + { + "bbox": [ + 100, + 258, + 167, + 266 + ], + "type": "text", + "content": "E: Depth 5, After Grow(O)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 280, + 545, + 304 + ], + "lines": [ + { + "bbox": [ + 46, + 280, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 46, + 280, + 545, + 304 + ], + "type": "text", + "content": "Figure 3. Our algorithm demonstrated on a 2D quadtree example. The black dots are our input point cloud, the yellow squares are surface leaves, and the blue and gray squares are non-surface leaves labelled as inside and outside respectively." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 196, + 171, + 293, + 255 + ], + "blocks": [ + { + "bbox": [ + 196, + 171, + 293, + 255 + ], + "lines": [ + { + "bbox": [ + 196, + 171, + 293, + 255 + ], + "spans": [ + { + "bbox": [ + 196, + 171, + 293, + 255 + ], + "type": "image", + "image_path": "1868eaab03bd81417376bc723aef5ce934ae151b0b40a9f884cb014f065fa34b.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 198, + 258, + 284, + 266 + ], + "lines": [ + { + "bbox": [ + 198, + 258, + 284, + 266 + ], + "spans": [ + { + "bbox": [ + 198, + 258, + 284, + 266 + ], + "type": "text", + "content": "F: Depth 5, After makeMove(O,1)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 294, + 171, + 391, + 255 + ], + "blocks": [ + { + "bbox": [ + 294, + 171, + 391, + 255 + ], + "lines": [ + { + "bbox": [ + 294, + 171, + 391, + 255 + ], + "spans": [ + { + "bbox": [ + 294, + 171, + 391, + 255 + ], + "type": "image", + "image_path": "137f8dc3155e079b792fa48ddef725f8cf477f359953eb425859653beac36864.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 296, + 258, + 390, + 266 + ], + "lines": [ + { + "bbox": [ + 296, + 258, + 390, + 266 + ], + "spans": [ + { + "bbox": [ + 296, + 258, + 390, + 266 + ], + "type": "text", + "content": "G: Depth 5, After all makeMove(O,s)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 398, + 171, + 496, + 255 + ], + "blocks": [ + { + "bbox": [ + 398, + 171, + 496, + 255 + ], + "lines": [ + { + "bbox": [ + 398, + 171, + 496, + 255 + ], + "spans": [ + { + "bbox": [ + 398, + 171, + 496, + 255 + ], + "type": "image", + "image_path": "eac3eca101ede04a8fd90faad2d351a6964c8305332a7203f3f8935c26f0bec7.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 399, + 258, + 493, + 266 + ], + "lines": [ + { + "bbox": [ + 399, + 258, + 493, + 266 + ], + "spans": [ + { + "bbox": [ + 399, + 258, + 493, + 266 + ], + "type": "text", + "content": "H: Depth 6, After all makeMove(O,s)" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 47, + 323, + 233, + 336 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 323, + 233, + 336 + ], + "spans": [ + { + "bbox": [ + 47, + 323, + 233, + 336 + ], + "type": "text", + "content": "3.3. Octree Construction and Labelling" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 46, + 342, + 287, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 342, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 46, + 342, + 287, + 426 + ], + "type": "text", + "content": "Setup. We present our octree labelling algorithm in Algorithm 1, which takes in point cloud " + }, + { + "bbox": [ + 46, + 342, + 287, + 426 + ], + "type": "inline_equation", + "content": "\\chi" + }, + { + "bbox": [ + 46, + 342, + 287, + 426 + ], + "type": "text", + "content": ", an initial depth to build the octree to " + }, + { + "bbox": [ + 46, + 342, + 287, + 426 + ], + "type": "inline_equation", + "content": "d_{i}" + }, + { + "bbox": [ + 46, + 342, + 287, + 426 + ], + "type": "text", + "content": ", a number of further depth expansions " + }, + { + "bbox": [ + 46, + 342, + 287, + 426 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 342, + 287, + 426 + ], + "type": "text", + "content": " (the final octree has depth " + }, + { + "bbox": [ + 46, + 342, + 287, + 426 + ], + "type": "inline_equation", + "content": "d_{f} = d_{i} + k" + }, + { + "bbox": [ + 46, + 342, + 287, + 426 + ], + "type": "text", + "content": ") and a set of move set sizes " + }, + { + "bbox": [ + 46, + 342, + 287, + 426 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 46, + 342, + 287, + 426 + ], + "type": "text", + "content": ". We set " + }, + { + "bbox": [ + 46, + 342, + 287, + 426 + ], + "type": "inline_equation", + "content": "d = 3" + }, + { + "bbox": [ + 46, + 342, + 287, + 426 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 342, + 287, + 426 + ], + "type": "inline_equation", + "content": "k = 4" + }, + { + "bbox": [ + 46, + 342, + 287, + 426 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 342, + 287, + 426 + ], + "type": "inline_equation", + "content": "C = [1,2,10,10000]" + }, + { + "bbox": [ + 46, + 342, + 287, + 426 + ], + "type": "text", + "content": " in our experiments. The algorithm returns a labelled octree " + }, + { + "bbox": [ + 46, + 342, + 287, + 426 + ], + "type": "inline_equation", + "content": "\\mathcal{O} = (\\mathcal{N},\\mathcal{L}_{ns},\\mathcal{L}_s,y)" + }, + { + "bbox": [ + 46, + 342, + 287, + 426 + ], + "type": "text", + "content": " where" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 58, + 427, + 287, + 563 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 58, + 427, + 287, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 427, + 287, + 462 + ], + "spans": [ + { + "bbox": [ + 58, + 427, + 287, + 462 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 58, + 427, + 287, + 462 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 58, + 427, + 287, + 462 + ], + "type": "text", + "content": " are the nodes of the octree. The octree is built by repeatedly expanding nodes containing or near points from " + }, + { + "bbox": [ + 58, + 427, + 287, + 462 + ], + "type": "inline_equation", + "content": "\\chi" + }, + { + "bbox": [ + 58, + 427, + 287, + 462 + ], + "type": "text", + "content": ", so it is highly unbalanced." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 58, + 465, + 287, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 465, + 287, + 525 + ], + "spans": [ + { + "bbox": [ + 58, + 465, + 287, + 525 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 58, + 465, + 287, + 525 + ], + "type": "inline_equation", + "content": "\\mathcal{L} \\subset \\mathcal{N}" + }, + { + "bbox": [ + 58, + 465, + 287, + 525 + ], + "type": "text", + "content": " are the leaf nodes, and thus " + }, + { + "bbox": [ + 58, + 465, + 287, + 525 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 58, + 465, + 287, + 525 + ], + "type": "text", + "content": " partitions " + }, + { + "bbox": [ + 58, + 465, + 287, + 525 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 58, + 465, + 287, + 525 + ], + "type": "text", + "content": ". Furthermore, we define surface leaves " + }, + { + "bbox": [ + 58, + 465, + 287, + 525 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_s \\subset \\mathcal{L}" + }, + { + "bbox": [ + 58, + 465, + 287, + 525 + ], + "type": "text", + "content": " to be leaves that contain points from the point cloud (colored yellow in Figure 2 and Figure 3) and all other leaves to be non-surface leaves " + }, + { + "bbox": [ + 58, + 465, + 287, + 525 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ns} = \\mathcal{L} \\setminus \\mathcal{L}_s" + }, + { + "bbox": [ + 58, + 465, + 287, + 525 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 58, + 528, + 287, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 528, + 287, + 563 + ], + "spans": [ + { + "bbox": [ + 58, + 528, + 287, + 563 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 58, + 528, + 287, + 563 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 58, + 528, + 287, + 563 + ], + "type": "text", + "content": " is the binary occupancy label vector indexed by non-surface leaves " + }, + { + "bbox": [ + 58, + 528, + 287, + 563 + ], + "type": "inline_equation", + "content": "\\ell \\in \\mathcal{L}_{ns}" + }, + { + "bbox": [ + 58, + 528, + 287, + 563 + ], + "type": "text", + "content": " denoting whether the space defined by each " + }, + { + "bbox": [ + 58, + 528, + 287, + 563 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 58, + 528, + 287, + 563 + ], + "type": "text", + "content": " is inside, " + }, + { + "bbox": [ + 58, + 528, + 287, + 563 + ], + "type": "inline_equation", + "content": "y_{\\ell} = 1" + }, + { + "bbox": [ + 58, + 528, + 287, + 563 + ], + "type": "text", + "content": ", or outside, " + }, + { + "bbox": [ + 58, + 528, + 287, + 563 + ], + "type": "inline_equation", + "content": "y_{\\ell} = 0" + }, + { + "bbox": [ + 58, + 528, + 287, + 563 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 46, + 564, + 287, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 564, + 287, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 564, + 287, + 640 + ], + "type": "text", + "content": "We will construct our algorithm so that the neighbours of every surface leaf " + }, + { + "bbox": [ + 46, + 564, + 287, + 640 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 46, + 564, + 287, + 640 + ], + "type": "text", + "content": " is at the same depth. Thus let " + }, + { + "bbox": [ + 46, + 564, + 287, + 640 + ], + "type": "inline_equation", + "content": "n_{\\ell}^{s}" + }, + { + "bbox": [ + 46, + 564, + 287, + 640 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 564, + 287, + 640 + ], + "type": "inline_equation", + "content": "n_{\\ell}^{ns}" + }, + { + "bbox": [ + 46, + 564, + 287, + 640 + ], + "type": "text", + "content": " be the number of its surface and non-surface neighbours and " + }, + { + "bbox": [ + 46, + 564, + 287, + 640 + ], + "type": "inline_equation", + "content": "n_{\\ell}^{(1)}(\\pmb{y})" + }, + { + "bbox": [ + 46, + 564, + 287, + 640 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 564, + 287, + 640 + ], + "type": "inline_equation", + "content": "n_{\\ell}^{(0)}(\\pmb{y})" + }, + { + "bbox": [ + 46, + 564, + 287, + 640 + ], + "type": "text", + "content": " be the number of inside and outside neighbors w.r.t. labelling " + }, + { + "bbox": [ + 46, + 564, + 287, + 640 + ], + "type": "inline_equation", + "content": "\\pmb{y}" + }, + { + "bbox": [ + 46, + 564, + 287, + 640 + ], + "type": "text", + "content": ". Then " + }, + { + "bbox": [ + 46, + 564, + 287, + 640 + ], + "type": "inline_equation", + "content": "n_{\\ell}^{ns} = n_{\\ell}^{(1)}(\\pmb{y}) + n_{\\ell}^{(0)}(\\pmb{y})" + }, + { + "bbox": [ + 46, + 564, + 287, + 640 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 564, + 287, + 640 + ], + "type": "inline_equation", + "content": "n_{\\ell}^{s} + n_{\\ell}^{s} = 26" + }, + { + "bbox": [ + 46, + 564, + 287, + 640 + ], + "type": "text", + "content": " (or 8 for our 2D example in Figure 3)." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 46, + 641, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 287, + 689 + ], + "type": "text", + "content": "We will constantly refer to the boundary as the leaves touching the boundary of " + }, + { + "bbox": [ + 46, + 641, + 287, + 689 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 46, + 641, + 287, + 689 + ], + "type": "text", + "content": " and the inside-outside border to be the non-surface leaves who have a non-surface neighbor with a different label (different occupancy value)." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": "We determine inside-outside occupancy based on three properties: (1) surface property: surfaces should separate" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 324, + 545, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 324, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 304, + 324, + 545, + 370 + ], + "type": "text", + "content": "inside and outside regions, (2) minimal surface constraint: the total surface area of the shape's surface " + }, + { + "bbox": [ + 304, + 324, + 545, + 370 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 324, + 545, + 370 + ], + "type": "text", + "content": " should be as small as possible, (3) boundary constraint: the boundary of the domain " + }, + { + "bbox": [ + 304, + 324, + 545, + 370 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 304, + 324, + 545, + 370 + ], + "type": "text", + "content": " should be outside." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 372, + 545, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 372, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 304, + 372, + 545, + 502 + ], + "type": "text", + "content": "Energy function. We propose an energy function over labellings " + }, + { + "bbox": [ + 304, + 372, + 545, + 502 + ], + "type": "inline_equation", + "content": "\\mathbf{y} = \\{y \\mid y_i \\in \\{0,1\\}, i \\in \\mathcal{L}_{ns}\\}" + }, + { + "bbox": [ + 304, + 372, + 545, + 502 + ], + "type": "text", + "content": " based on the first two criteria, the third is hard-coded at the start of the algorithm. For the surface property, ideally every surface leaf would have both inside and outside neighbours, however it may only have surface leaf neighbours which contain both inside and outside regions by definition (see Figure 3 C,D,G,H, some surface leafs have no inside neighbours). For the minimal surface constraint, we only create surfaces when adjacent non-surface leaves have different labellings (note surface leaves are fixed), so we want to minimise this." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 503, + 545, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 503, + 545, + 591 + ], + "spans": [ + { + "bbox": [ + 304, + 503, + 545, + 591 + ], + "type": "text", + "content": "For the surface property, we want every surface leaf " + }, + { + "bbox": [ + 304, + 503, + 545, + 591 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 304, + 503, + 545, + 591 + ], + "type": "text", + "content": " to have some amount of both inside and outside neighbours, where surface neighbours partially count towards both. Thus, we aim for " + }, + { + "bbox": [ + 304, + 503, + 545, + 591 + ], + "type": "inline_equation", + "content": "\\gamma_{\\ell}^{(1)} = \\gamma^{(1)} - \\eta^{(1)}n_{\\ell}^{s}" + }, + { + "bbox": [ + 304, + 503, + 545, + 591 + ], + "type": "text", + "content": " non-surface inside neighbors and " + }, + { + "bbox": [ + 304, + 503, + 545, + 591 + ], + "type": "inline_equation", + "content": "\\gamma_{\\ell}^{(0)} = \\gamma^{(0)} - \\eta^{(0)}n_{\\ell}^{s}" + }, + { + "bbox": [ + 304, + 503, + 545, + 591 + ], + "type": "text", + "content": " non-surface outside neighbors where " + }, + { + "bbox": [ + 304, + 503, + 545, + 591 + ], + "type": "inline_equation", + "content": "\\gamma^{(\\cdot)},\\eta^{(\\cdot)}" + }, + { + "bbox": [ + 304, + 503, + 545, + 591 + ], + "type": "text", + "content": " are hyperparameters. Then the energy cost for a surface leaf " + }, + { + "bbox": [ + 304, + 503, + 545, + 591 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 304, + 503, + 545, + 591 + ], + "type": "text", + "content": " is" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 595, + 545, + 616 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 595, + 545, + 616 + ], + "spans": [ + { + "bbox": [ + 308, + 595, + 545, + 616 + ], + "type": "interline_equation", + "content": "E _ {\\ell} ^ {S P} (\\boldsymbol {y}) = \\max \\left\\{\\gamma_ {\\ell} ^ {(1)} - n _ {\\ell} ^ {(1)} (\\boldsymbol {y}), \\gamma_ {\\ell} ^ {(0)} - n _ {\\ell} ^ {(0)} (\\boldsymbol {y}), 0 \\right\\}. \\tag {2}", + "image_path": "b92bb2da18bec07a546dbc3aebcde4be4c1d576036bcb4cac4cf41e57031502f.jpg" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 304, + 619, + 545, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 619, + 545, + 666 + ], + "spans": [ + { + "bbox": [ + 304, + 619, + 545, + 666 + ], + "type": "text", + "content": "For the minimal surface constraint, we want to minimise the total surface area between every pair of non-surface neighbors with different labellings. Then for two non-surface leaves " + }, + { + "bbox": [ + 304, + 619, + 545, + 666 + ], + "type": "inline_equation", + "content": "\\ell, \\ell'" + }, + { + "bbox": [ + 304, + 619, + 545, + 666 + ], + "type": "text", + "content": ", our energy cost is" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 370, + 670, + 545, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 670, + 545, + 685 + ], + "spans": [ + { + "bbox": [ + 370, + 670, + 545, + 685 + ], + "type": "interline_equation", + "content": "E _ {\\ell , \\ell^ {\\prime}} ^ {M S C} (\\boldsymbol {y}) = a _ {\\ell \\ell^ {\\prime}} \\left| y _ {\\ell} - y _ {\\ell^ {\\prime}} \\right| \\tag {3}", + "image_path": "9762c25f5cc1f8ea14cb14c0b39f05d95b9f93708a417065d2d2c2f3d0e1d9d5.jpg" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "a_{\\ell \\ell^{\\prime}}" + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "content": " is the (surface) area in common between them (note " + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "inline_equation", + "content": "a_{ij} = 0" + }, + { + "bbox": [ + 304, + 689, + 545, + 714 + ], + "type": "text", + "content": " if they are not adjacent)." + } + ] + } + ], + "index": 32 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16720" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 97 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 97 + ], + "type": "text", + "content": "The two criteria can be formalized as an energy function over labellings " + }, + { + "bbox": [ + 47, + 72, + 287, + 97 + ], + "type": "inline_equation", + "content": "\\mathbf{y} = \\{y_{i}\\in \\{0,1\\} \\mid i\\in \\mathcal{L}_{ns}\\}" + }, + { + "bbox": [ + 47, + 72, + 287, + 97 + ], + "type": "text", + "content": " by" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 75, + 103, + 287, + 131 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 103, + 287, + 131 + ], + "spans": [ + { + "bbox": [ + 75, + 103, + 287, + 131 + ], + "type": "interline_equation", + "content": "E (\\boldsymbol {y}) = \\sum_ {j \\in \\mathcal {L} _ {s}} E _ {j} ^ {S P} (\\boldsymbol {y}) + \\lambda \\sum_ {i, j \\in \\mathcal {L} _ {n s}} E _ {i, j} ^ {M S C} (\\boldsymbol {y}) \\tag {4}", + "image_path": "3908299269c34343746ea7e9e1fa975de77f05399d637ea6127ce85fa0897728.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 137, + 167, + 148 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 137, + 167, + 148 + ], + "spans": [ + { + "bbox": [ + 47, + 137, + 167, + 148 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 137, + 167, + 148 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 47, + 137, + 167, + 148 + ], + "type": "text", + "content": " is a hyperparameter." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 163, + 257, + 175 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 163, + 257, + 175 + ], + "spans": [ + { + "bbox": [ + 47, + 163, + 257, + 175 + ], + "type": "text", + "content": "3.3.1 Octree Creation and Labelling Algorithm" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 182, + 287, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 182, + 287, + 265 + ], + "spans": [ + { + "bbox": [ + 46, + 182, + 287, + 265 + ], + "type": "text", + "content": "The energy function in Equation 4 is non-trivial to minimize. It is a discrete energy minimisation problem with hard unary terms (the boundary constraint), pairwise terms (Equation 3 which is essentially a smoothness constraint) and n-ary terms (Equation 2). The n-ary term however is not concave, so we cannot use existing methods to perform exact inference using graph-cuts [13,20]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 266, + 287, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 266, + 287, + 409 + ], + "spans": [ + { + "bbox": [ + 46, + 266, + 287, + 409 + ], + "type": "text", + "content": "Instead, we employ a move-making algorithm [9] with moves specialized for our task (see Figure 3). We first hard code the boundary constraint (leaves on the boundary of the domain are outside) and initialize all other leaves to inside. To be efficient, we (1) only consider moves that change leaves from inside to outside (never the other way around), (2) require such moves include an inside leaf from the inside-outside border, (3) iteratively grow such potential moves, and (4) apply moves at multiple depths. We also have a grow stage to remove extraneous inside leaves without needing guidance from the energy function. We now describe the stages of our algorithm, depicted in Algorithm 1." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 410, + 287, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 410, + 287, + 468 + ], + "spans": [ + { + "bbox": [ + 46, + 410, + 287, + 468 + ], + "type": "text", + "content": "Octree Creation. We start with a single node " + }, + { + "bbox": [ + 46, + 410, + 287, + 468 + ], + "type": "inline_equation", + "content": "n_0" + }, + { + "bbox": [ + 46, + 410, + 287, + 468 + ], + "type": "text", + "content": " covering the whole space " + }, + { + "bbox": [ + 46, + 410, + 287, + 468 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 46, + 410, + 287, + 468 + ], + "type": "text", + "content": ". We then recursively expand any node (into 8 equally sized child nodes) that contains points from " + }, + { + "bbox": [ + 46, + 410, + 287, + 468 + ], + "type": "inline_equation", + "content": "\\chi" + }, + { + "bbox": [ + 46, + 410, + 287, + 468 + ], + "type": "text", + "content": " in it until depth " + }, + { + "bbox": [ + 46, + 410, + 287, + 468 + ], + "type": "inline_equation", + "content": "d_i" + }, + { + "bbox": [ + 46, + 410, + 287, + 468 + ], + "type": "text", + "content": " is reached. Note that every surface leaf is at depth " + }, + { + "bbox": [ + 46, + 410, + 287, + 468 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 410, + 287, + 468 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 469, + 287, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 469, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 47, + 469, + 287, + 491 + ], + "type": "text", + "content": "Initialisation. We hard code our boundary condition and set every other leaf node to inside (see Figure 3 A)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 492, + 287, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 492, + 287, + 588 + ], + "spans": [ + { + "bbox": [ + 46, + 492, + 287, + 588 + ], + "type": "text", + "content": "Grow Stage. We first expand any neighbor of a surface leaf that is not also at the current depth " + }, + { + "bbox": [ + 46, + 492, + 287, + 588 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 492, + 287, + 588 + ], + "type": "text", + "content": " (new leaves keep the label of its parent). We repeatedly expand the outside region changing inside leaves to outside if they stick out from a plane, i.e., have at least " + }, + { + "bbox": [ + 46, + 492, + 287, + 588 + ], + "type": "inline_equation", + "content": "26/2 + 1 = 14" + }, + { + "bbox": [ + 46, + 492, + 287, + 588 + ], + "type": "text", + "content": " outside neighbors (or " + }, + { + "bbox": [ + 46, + 492, + 287, + 588 + ], + "type": "inline_equation", + "content": "8/2 + 1 = 5" + }, + { + "bbox": [ + 46, + 492, + 287, + 588 + ], + "type": "text", + "content": " for our 2D example in Figure 3). Finally we expand nodes that are on the inside-outsider border and are not at the current octree depth." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 589, + 287, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 589, + 287, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 589, + 287, + 696 + ], + "type": "text", + "content": "Move-Making Stage. To overcome local minima we propose sets of leaves to change from inside to outside in a single move. Sets are constructed as follows. Starting with an inside leaf on the border we greedily add neighboring inside leaves to the set, selecting the neighbor which results in lowest energy, and calculate the cost of the current move set. Note that the energy can temporarily increase. We do this until there are no inside neighbors remaining or a maximum set size " + }, + { + "bbox": [ + 46, + 589, + 287, + 696 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 589, + 287, + 696 + ], + "type": "text", + "content": " is reached.1 If any of these moves result" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 72, + 545, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 107 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 107 + ], + "type": "text", + "content": "in a decrease in energy, we apply the move with the largest decrease, We repeat the above for different starting leaves and for multiple increasing set sizes " + }, + { + "bbox": [ + 306, + 72, + 545, + 107 + ], + "type": "inline_equation", + "content": "s\\in C" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 108, + 545, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 108, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 304, + 108, + 545, + 227 + ], + "type": "text", + "content": "Figure 3 illustrates this process in 2D. After reaching labelling (E), all single variable changes (" + }, + { + "bbox": [ + 304, + 108, + 545, + 227 + ], + "type": "inline_equation", + "content": "s = 1" + }, + { + "bbox": [ + 304, + 108, + 545, + 227 + ], + "type": "text", + "content": ") are considered, resulting in (F). Observe that the inside-outside border has been refined but larger regions are still incorrectly labelled. Considering larger moves as shown in (G) fixes many of these regions, with further refinements being made at deeper levels of the octree as shown in (H). We explain this 2D example in more detail in the supplemental material, including why we cannot reach a lower energy from fixing certain regions until a deeper level is reached." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 228, + 545, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 228, + 545, + 299 + ], + "spans": [ + { + "bbox": [ + 304, + 228, + 545, + 299 + ], + "type": "text", + "content": "Expanding. We find that labelling at multiple depths is faster and more stable than directly labelling the largest depth. Note that expanding completely changes " + }, + { + "bbox": [ + 304, + 228, + 545, + 299 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_s" + }, + { + "bbox": [ + 304, + 228, + 545, + 299 + ], + "type": "text", + "content": " and adds more nodes to " + }, + { + "bbox": [ + 304, + 228, + 545, + 299 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ns}" + }, + { + "bbox": [ + 304, + 228, + 545, + 299 + ], + "type": "text", + "content": ", and thus expands " + }, + { + "bbox": [ + 304, + 228, + 545, + 299 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 228, + 545, + 299 + ], + "type": "text", + "content": ". We keep the labels for previous non-surface leaves, and assign all the new non-surface leaves to inside." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 306, + 391, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 306, + 391, + 319 + ], + "spans": [ + { + "bbox": [ + 306, + 306, + 391, + 319 + ], + "type": "text", + "content": "3.4. SDF learning" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 324, + 545, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 324, + 545, + 397 + ], + "spans": [ + { + "bbox": [ + 304, + 324, + 545, + 397 + ], + "type": "text", + "content": "We use the octree to guide the SDF learning. Given an INR, we train it with the combination of four losses. The first two are the standard losses in the literature [1,2,5,14, 34], the surface loss to enforce that the input point cloud should lie on the zero level set and the eikonal loss [14] to enforce that the gradient has magnitude one everywhere" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 404, + 545, + 433 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 404, + 545, + 433 + ], + "spans": [ + { + "bbox": [ + 307, + 404, + 545, + 433 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {1} = \\sum_ {x \\in \\chi} | \\Phi (x, \\theta) |, \\mathcal {L} _ {2} = \\int_ {\\mathcal {D}} | \\| \\nabla_ {x} \\Phi (x, \\theta) \\| _ {2} - 1 | d x. \\tag {5}", + "image_path": "443d46857863a0af07ae039b8811aa08504e28649255858795825570614cfcb9.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 441, + 545, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 441, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 304, + 441, + 545, + 525 + ], + "type": "text", + "content": "We now introduce two loss functions that take guidance from the labelled octree. We first produce an approximate signed distance function " + }, + { + "bbox": [ + 304, + 441, + 545, + 525 + ], + "type": "inline_equation", + "content": "\\tilde{d}_s(x)" + }, + { + "bbox": [ + 304, + 441, + 545, + 525 + ], + "type": "text", + "content": " for finely sampled points " + }, + { + "bbox": [ + 304, + 441, + 545, + 525 + ], + "type": "inline_equation", + "content": "x\\in \\mathcal{D}" + }, + { + "bbox": [ + 304, + 441, + 545, + 525 + ], + "type": "text", + "content": " by computing the distance to the closest input point cloud point and choose its sign based on the label of the octree leaf that contains that point (for surface leaves we choose inside). We use this as direct supervision" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 361, + 532, + 545, + 558 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 532, + 545, + 558 + ], + "spans": [ + { + "bbox": [ + 361, + 532, + 545, + 558 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {3} = \\int_ {\\mathcal {D}} | \\Phi (x, \\theta) - \\tilde {d} _ {s} (x) | \\mathrm {d} x. \\tag {6}", + "image_path": "201fefde67d3a51c3e849a18d8d78e8aea44328e65d77dff3519f1aba3774a00.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 563, + 545, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 563, + 545, + 600 + ], + "spans": [ + { + "bbox": [ + 304, + 563, + 545, + 600 + ], + "type": "text", + "content": "We also enforce the sign of each octree leaf explicitly (again taking surface leaves to be considered as inside) at a fixed number of randomly sampled points in that leaf " + }, + { + "bbox": [ + 304, + 563, + 545, + 600 + ], + "type": "inline_equation", + "content": "S_{\\ell}" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 337, + 607, + 545, + 640 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 607, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 337, + 607, + 545, + 640 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {4} = \\sum_ {\\ell \\in \\mathcal {L}} \\left(\\sum_ {x \\in S _ {\\ell}} \\left[ (- 1) ^ {\\left(1 - y _ {\\ell} ^ {\\prime}\\right)} \\Phi (x, \\theta) \\right] _ {> 0}\\right) \\tag {7}", + "image_path": "a40e359c413f7a97f2e7470aefbf48c85754850fb19634ae0ff0fc9cc763773e.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 647, + 545, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 647, + 545, + 696 + ], + "spans": [ + { + "bbox": [ + 304, + 647, + 545, + 696 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 647, + 545, + 696 + ], + "type": "inline_equation", + "content": "y_{\\ell}^{\\prime} = y_{\\ell}" + }, + { + "bbox": [ + 304, + 647, + 545, + 696 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 304, + 647, + 545, + 696 + ], + "type": "inline_equation", + "content": "\\ell \\in \\mathcal{L}_{ns}" + }, + { + "bbox": [ + 304, + 647, + 545, + 696 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 647, + 545, + 696 + ], + "type": "inline_equation", + "content": "y_{\\ell}^{\\prime} = 1" + }, + { + "bbox": [ + 304, + 647, + 545, + 696 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 304, + 647, + 545, + 696 + ], + "type": "inline_equation", + "content": "\\ell \\in \\mathcal{L}_s" + }, + { + "bbox": [ + 304, + 647, + 545, + 696 + ], + "type": "text", + "content": ". Note that sampling a fixed number per leaf automatically adjusts our sampling to be denser in more important regions (i.e., closer to the surface) due to the octree structure." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 58, + 702, + 287, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 287, + 713 + ], + "type": "text", + "content": "1In practice we include additional early stopping criteria to avoid moves" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 306, + 702, + 541, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 702, + 541, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 702, + 541, + 713 + ], + "type": "text", + "content": "that are unlikely to eventually be accepted (see supplemental for details)." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16721" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 72, + 138, + 137 + ], + "blocks": [ + { + "bbox": [ + 65, + 72, + 138, + 137 + ], + "lines": [ + { + "bbox": [ + 65, + 72, + 138, + 137 + ], + "spans": [ + { + "bbox": [ + 65, + 72, + 138, + 137 + ], + "type": "image", + "image_path": "cf02dbfef551deea93b6c981bba572e6dcc901c4ef9e131c27ea4e8d6e2f4a3e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 65, + 144, + 136, + 211 + ], + "blocks": [ + { + "bbox": [ + 65, + 144, + 136, + 211 + ], + "lines": [ + { + "bbox": [ + 65, + 144, + 136, + 211 + ], + "spans": [ + { + "bbox": [ + 65, + 144, + 136, + 211 + ], + "type": "image", + "image_path": "59d4b6c9af85268ca638d5411996e72771ca83eae409b4bd53e4c0a35d2120c8.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 86, + 214, + 116, + 225 + ], + "lines": [ + { + "bbox": [ + 86, + 214, + 116, + 225 + ], + "spans": [ + { + "bbox": [ + 86, + 214, + 116, + 225 + ], + "type": "text", + "content": "Depth 7" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 143, + 72, + 216, + 137 + ], + "blocks": [ + { + "bbox": [ + 143, + 72, + 216, + 137 + ], + "lines": [ + { + "bbox": [ + 143, + 72, + 216, + 137 + ], + "spans": [ + { + "bbox": [ + 143, + 72, + 216, + 137 + ], + "type": "image", + "image_path": "e29c96d12c52aa202598aa68dba57b202a065e16a592c7806e70a9dc1522bc4c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 143, + 144, + 214, + 211 + ], + "blocks": [ + { + "bbox": [ + 143, + 144, + 214, + 211 + ], + "lines": [ + { + "bbox": [ + 143, + 144, + 214, + 211 + ], + "spans": [ + { + "bbox": [ + 143, + 144, + 214, + 211 + ], + "type": "image", + "image_path": "23f4291f6c796179b1eb1a1b198f038378670bfea4d4a266f6dde2bcfac05fe0.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 164, + 214, + 195, + 225 + ], + "lines": [ + { + "bbox": [ + 164, + 214, + 195, + 225 + ], + "spans": [ + { + "bbox": [ + 164, + 214, + 195, + 225 + ], + "type": "text", + "content": "Depth 6" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 221, + 77, + 294, + 137 + ], + "blocks": [ + { + "bbox": [ + 221, + 77, + 294, + 137 + ], + "lines": [ + { + "bbox": [ + 221, + 77, + 294, + 137 + ], + "spans": [ + { + "bbox": [ + 221, + 77, + 294, + 137 + ], + "type": "image", + "image_path": "a58e93455b70c98fa68ef061c4771657b3147a8f0f0bca953c8a9002ac542583.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 221, + 144, + 292, + 211 + ], + "blocks": [ + { + "bbox": [ + 221, + 144, + 292, + 211 + ], + "lines": [ + { + "bbox": [ + 221, + 144, + 292, + 211 + ], + "spans": [ + { + "bbox": [ + 221, + 144, + 292, + 211 + ], + "type": "image", + "image_path": "966fbf36fd3d67bf945d4fa9a375a4a7df7ea805117cba6fd9d7d18981d895f6.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 239, + 214, + 275, + 226 + ], + "lines": [ + { + "bbox": [ + 239, + 214, + 275, + 226 + ], + "spans": [ + { + "bbox": [ + 239, + 214, + 275, + 226 + ], + "type": "text", + "content": "SAP [31]" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 236, + 545, + 258 + ], + "lines": [ + { + "bbox": [ + 46, + 236, + 545, + 258 + ], + "spans": [ + { + "bbox": [ + 46, + 236, + 545, + 258 + ], + "type": "text", + "content": "Figure 4. A comparison of surface reconstructions of shapes from the watercraft and bench classes of ShapeNet. The left two reconstructions are using OG-SIREN with the octree built to different depths." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 301, + 72, + 373, + 137 + ], + "blocks": [ + { + "bbox": [ + 301, + 72, + 373, + 137 + ], + "lines": [ + { + "bbox": [ + 301, + 72, + 373, + 137 + ], + "spans": [ + { + "bbox": [ + 301, + 72, + 373, + 137 + ], + "type": "image", + "image_path": "2627356b92797d69c7ee784fdfc3805b03fa2631acb814b35f92190b7d5e3cf5.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 301, + 144, + 370, + 211 + ], + "blocks": [ + { + "bbox": [ + 301, + 144, + 370, + 211 + ], + "lines": [ + { + "bbox": [ + 301, + 144, + 370, + 211 + ], + "spans": [ + { + "bbox": [ + 301, + 144, + 370, + 211 + ], + "type": "image", + "image_path": "f9c98c37b2c9267b067220a4b2543d3b5dce4e068afae9a2d3d97810903853ea.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 318, + 214, + 353, + 225 + ], + "lines": [ + { + "bbox": [ + 318, + 214, + 353, + 225 + ], + "spans": [ + { + "bbox": [ + 318, + 214, + 353, + 225 + ], + "type": "text", + "content": "DiGS [5]" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 378, + 72, + 449, + 137 + ], + "blocks": [ + { + "bbox": [ + 378, + 72, + 449, + 137 + ], + "lines": [ + { + "bbox": [ + 378, + 72, + 449, + 137 + ], + "spans": [ + { + "bbox": [ + 378, + 72, + 449, + 137 + ], + "type": "image", + "image_path": "c0de15092a2816fc81dfc4896807309f2496d9f8e205671fdccffddf100ac226.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 380, + 144, + 448, + 210 + ], + "blocks": [ + { + "bbox": [ + 380, + 144, + 448, + 210 + ], + "lines": [ + { + "bbox": [ + 380, + 144, + 448, + 210 + ], + "spans": [ + { + "bbox": [ + 380, + 144, + 448, + 210 + ], + "type": "image", + "image_path": "3aedd6e7cc40a7ed6823ef2c29571a59e4f0d60ea00467ee07e45335e5a20639.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 399, + 214, + 429, + 225 + ], + "lines": [ + { + "bbox": [ + 399, + 214, + 429, + 225 + ], + "spans": [ + { + "bbox": [ + 399, + 214, + 429, + 225 + ], + "type": "text", + "content": "SAL [1]" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 455, + 72, + 528, + 137 + ], + "blocks": [ + { + "bbox": [ + 455, + 72, + 528, + 137 + ], + "lines": [ + { + "bbox": [ + 455, + 72, + 528, + 137 + ], + "spans": [ + { + "bbox": [ + 455, + 72, + 528, + 137 + ], + "type": "image", + "image_path": "c09ef78ac6c66bf30905402a33ffb4ff88416f88bcd96e96c774d9ed0ee3fe70.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 456, + 144, + 526, + 211 + ], + "blocks": [ + { + "bbox": [ + 456, + 144, + 526, + 211 + ], + "lines": [ + { + "bbox": [ + 456, + 144, + 526, + 211 + ], + "spans": [ + { + "bbox": [ + 456, + 144, + 526, + 211 + ], + "type": "image", + "image_path": "1b709118308ae31c6c3b8114b850aa24fd499e2437ee5ae58a43f887c3c56f94.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 460, + 214, + 524, + 225 + ], + "lines": [ + { + "bbox": [ + 460, + 214, + 524, + 225 + ], + "spans": [ + { + "bbox": [ + 460, + 214, + 524, + 225 + ], + "type": "text", + "content": "SIREN won [34]" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 59, + 279, + 261, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 279, + 261, + 291 + ], + "spans": [ + { + "bbox": [ + 59, + 279, + 261, + 291 + ], + "type": "text", + "content": "Our overall loss is then the weighted combination" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 85, + 300, + 287, + 313 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 300, + 287, + 313 + ], + "spans": [ + { + "bbox": [ + 85, + 300, + 287, + 313 + ], + "type": "interline_equation", + "content": "\\mathcal {L} (\\Phi) = \\lambda_ {1} \\mathcal {L} _ {1} + \\lambda_ {2} \\mathcal {L} _ {2} + \\lambda_ {3} \\mathcal {L} _ {3} + \\lambda_ {4} \\mathcal {L} _ {4}. \\tag {8}", + "image_path": "8cac2b6cf7c0b36e569499cfdd5f6b71c8f08c841080a5762b86c1b2131d9eed.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 46, + 323, + 287, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 323, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 46, + 323, + 287, + 396 + ], + "type": "text", + "content": "We keep " + }, + { + "bbox": [ + 46, + 323, + 287, + 396 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 46, + 323, + 287, + 396 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 323, + 287, + 396 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 46, + 323, + 287, + 396 + ], + "type": "text", + "content": " fixed throughout training as their loss terms have no bias. As the guidance in " + }, + { + "bbox": [ + 46, + 323, + 287, + 396 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_3" + }, + { + "bbox": [ + 46, + 323, + 287, + 396 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 323, + 287, + 396 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_4" + }, + { + "bbox": [ + 46, + 323, + 287, + 396 + ], + "type": "text", + "content": " is approximate but highly informative, we initially use large values for " + }, + { + "bbox": [ + 46, + 323, + 287, + 396 + ], + "type": "inline_equation", + "content": "\\lambda_3,\\lambda_4" + }, + { + "bbox": [ + 46, + 323, + 287, + 396 + ], + "type": "text", + "content": ", allowing the INR to quickly fit to the rough shape, and then reduce them during training. More training details are in the supplemental material." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 47, + 406, + 100, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 406, + 100, + 418 + ], + "spans": [ + { + "bbox": [ + 47, + 406, + 100, + 418 + ], + "type": "text", + "content": "4. Results" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 46, + 426, + 287, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 287, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 287, + 617 + ], + "type": "text", + "content": "Datasets. We evaluate the performance of our method on the ShapeNet [12] and Surface Reconstruction Benchmark (SRB) [7] datasets. ShapeNet has a wide variety of shapes, with thin surfaces, hollow regions and disjoint objects. However the shapes are rather uniformly sampled. We use the preprocessing and split of Neural Splines [45], which consists of 13 shape classes with 20 shapes in each shape class. For each shape it provides the ground truth mesh (preprocessed to be watertight), an input point cloud and occupancy annotation at dense points. SRB contains five shapes from simulated noisy range scans, but it is a popularly used dataset due to the shapes exhibiting many challenging features. These include large missing regions, thin and sharp regions, and small surface details. The former means that good priors in the reconstruction methods, such as smoothness priors, are very important." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 46, + 617, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 653 + ], + "type": "text", + "content": "We follow the metrics used in the literature for each dataset: the squared Chamfer distance and IoU for ShapeNet and Chamfer and Hausdorff distance for SRB." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 46, + 653, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 712 + ], + "type": "text", + "content": "INRs. We demonstrate our method on two INRs, SIREN [34], and NGLOD [35] (note that NGLOD only works in a supervised setting as it has no global consistency, but our octree guidance is sufficient). Thus we name our two methods OG-SIREN and OG-NGLOD where OG stands for" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 279, + 546, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 279, + 546, + 327 + ], + "spans": [ + { + "bbox": [ + 304, + 279, + 546, + 327 + ], + "type": "text", + "content": "\"Octree Guided\". To extract a mesh we first evaluate our INRs on a " + }, + { + "bbox": [ + 304, + 279, + 546, + 327 + ], + "type": "inline_equation", + "content": "256^3" + }, + { + "bbox": [ + 304, + 279, + 546, + 327 + ], + "type": "text", + "content": " grid and then use marching cubes [25]. We use the same settings for the octree and the INRs on both datasets, and give more details in our supplemental." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 305, + 335, + 463, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 335, + 463, + 346 + ], + "spans": [ + { + "bbox": [ + 305, + 335, + 463, + 346 + ], + "type": "text", + "content": "4.1. Reconstruction Performance" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 353, + 546, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 353, + 546, + 532 + ], + "spans": [ + { + "bbox": [ + 304, + 353, + 546, + 532 + ], + "type": "text", + "content": "ShapeNet dataset. The results are shown in Table 1 (left) and visualizations on two shapes are shown in Figure 4. Our methods get the smallest mean squared Chamfer distance out of the unoriented methods by almost an order of magnitude, and are even better than the best performing oriented method. Likewise for mean IoU our methods are better than both unoriented and oriented methods. For median performance, our methods are comparable to the best oriented and unoriented methods. SAP [31] gets the best median IoU, but their mean is much lower than ours, indicating that in a few instances they misrepresent large regions. On the other hand, due to our move-making algorithm we reduce the likelihood of vastly incorrect labellings, and as a result achieve low standard deviations on both metrics despite the high performance." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 304, + 532, + 545, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 532, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 532, + 545, + 640 + ], + "type": "text", + "content": "SRB dataset. Table 1 (right) shows that DiGS, SAP, PHASE+FF and our two methods do comparably, with DiGS performing slightly better. Note the performance of these methods are comparable to most oriented methods. Also note that the difference between our two methods is more prominent on SRB as the large missing regions means that smoothness priors are more important. Octree guidance is NGLOD's only global supervision, as it fits locally, while SIREN is biased towards smooth shapes." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 306, + 647, + 440, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 647, + 440, + 659 + ], + "spans": [ + { + "bbox": [ + 306, + 647, + 440, + 659 + ], + "type": "text", + "content": "4.2. Octree Depth Ablations" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": "Table 4 shows the results of OG-SIREN on ShapeNet when evaluating at different depths of the octree. We include an octree IoU metric, where we compute the IoU of the labelling of the octree directly (taking surface leaves to" + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16722" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 55, + 74, + 343, + 272 + ], + "blocks": [ + { + "bbox": [ + 55, + 74, + 343, + 272 + ], + "lines": [ + { + "bbox": [ + 55, + 74, + 343, + 272 + ], + "spans": [ + { + "bbox": [ + 55, + 74, + 343, + 272 + ], + "type": "table", + "html": "
MethodSquared Chamfer ↓IoU ↑
meanmedianstdmeanmedianstd
SPSR [19]5.44e-51.97e-55.06e-40.99260.99560.0105
IGR [14]5.12e-41.13e-42.15e-30.81020.84800.1519
SIREN [34]1.03e-45.28e-51.93e-40.82680.90970.2329
FFN [36]9.12e-58.65e-53.36e-50.82180.83960.0989
Biharmonic RBF [11]1.11e-48.97e-57.06e-50.82470.86420.1350
SVR [33]1.14e-41.04e-45.99e-50.76250.78190.1300
NSP [45]5.36e-54.06e-53.64e-50.89730.92300.0871
DiGS + n [5]2.74e-42.32e-59.90e-40.92000.97740.1992
N Est+SPSR [19]3.76e-34.37e-51.14e-20.71870.97610.3767
SIREN won [34]3.08e-42.58e-43.26e-40.30850.29520.2014
SAL [1]1.14e-32.11e-43.63e-30.40300.39440.2722
DiGS [5]1.32e-42.55e-54.73e-40.93900.97640.1262
SAP [31]4.09e-42.46e-52.60e-30.91180.99230.2002
Our OG-SIREN3.75e-52.17e-58.24e-50.96150.98710.1048
Our OG-NGLOD5.07e-52.57e-59.39e-50.95930.98700.1057
", + "image_path": "fa83e085aa3b24c26e597ef68cae30ee989d8da1a04ba16f06860685288365d3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 387, + 71, + 512, + 277 + ], + "blocks": [ + { + "bbox": [ + 387, + 71, + 512, + 277 + ], + "lines": [ + { + "bbox": [ + 387, + 71, + 512, + 277 + ], + "spans": [ + { + "bbox": [ + 387, + 71, + 512, + 277 + ], + "type": "table", + "html": "
MethoddC↓dH↓
SPSR [19]0.214.69
DGP [44]0.215.18
IGR [14]0.192.99
SIREN [34]0.193.86
NSP [45]0.172.85
PHASE [22]0.162.77
DiGS + n [5]0.183.55
N Est.+SPSR [19]1.2522.59
IGR wo n [14]1.3816.33
SIREN wo n [34]0.427.67
SAL [14]0.367.47
IGR+FF [14]0.9611.06
PHASE+FF [22]0.224.96
DiGS [5]0.193.52
SAP [31]0.214.51
Our OG-SIREN0.204.06
Our OG-NGLOD0.226.03
", + "image_path": "84a04421c2ddb5a2135b9d0b50c465ae82a2f2ecf2f73e7644bcfa54aba1d3b8.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 102, + 338, + 233, + 456 + ], + "blocks": [ + { + "bbox": [ + 46, + 285, + 547, + 319 + ], + "lines": [ + { + "bbox": [ + 46, + 285, + 547, + 319 + ], + "spans": [ + { + "bbox": [ + 46, + 285, + 547, + 319 + ], + "type": "text", + "content": "Table 1. Surface Reconstruction results on ShapeNet [12] (left) and SRB [7] (right). Methods above the line use ground truth normal information, and methods below do not. Left: The mean, median and standard deviation of the squared Chamfer distance and IoU of all 260 shapes are reported. Right: The mean Chamfer " + }, + { + "bbox": [ + 46, + 285, + 547, + 319 + ], + "type": "inline_equation", + "content": "d_{C}" + }, + { + "bbox": [ + 46, + 285, + 547, + 319 + ], + "type": "text", + "content": " and Hausdorff distance " + }, + { + "bbox": [ + 46, + 285, + 547, + 319 + ], + "type": "inline_equation", + "content": "d_{H}" + }, + { + "bbox": [ + 46, + 285, + 547, + 319 + ], + "type": "text", + "content": " for all 5 shapes are reported." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 102, + 338, + 233, + 456 + ], + "lines": [ + { + "bbox": [ + 102, + 338, + 233, + 456 + ], + "spans": [ + { + "bbox": [ + 102, + 338, + 233, + 456 + ], + "type": "table", + "html": "
MethoddC↓dH↓
SIREN w on [34]0.427.67
OG-SIREN0.204.06
SIREN (+n) [34]0.193.86
OG-SIREN + H°0.192.96
OG-SIREN + GT°0.172.44
OG-NGLOD0.226.03
OG-NGLOD + H°0.195.50
OG-NGLOD + GT°0.175.12
NGLOD+GTd [35]0.131.97
", + "image_path": "747f748de88ce733f4d841276f5451c326c502ae2a2ad88ece8d8f3443e64cb1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 465, + 288, + 510 + ], + "lines": [ + { + "bbox": [ + 46, + 465, + 288, + 510 + ], + "spans": [ + { + "bbox": [ + 46, + 465, + 288, + 510 + ], + "type": "text", + "content": "Table 2. Results on SRB [7] when supervising INRs with varying levels of information (see Section 4.3). " + }, + { + "bbox": [ + 46, + 465, + 288, + 510 + ], + "type": "inline_equation", + "content": "\\mathrm{GT}^d" + }, + { + "bbox": [ + 46, + 465, + 288, + 510 + ], + "type": "text", + "content": " means dense GT SDF supervision, while " + }, + { + "bbox": [ + 46, + 465, + 288, + 510 + ], + "type": "inline_equation", + "content": "\\mathrm{GT}^o" + }, + { + "bbox": [ + 46, + 465, + 288, + 510 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 465, + 288, + 510 + ], + "type": "inline_equation", + "content": "\\mathrm{H}^o" + }, + { + "bbox": [ + 46, + 465, + 288, + 510 + ], + "type": "text", + "content": " mean GT and human estimated occupancy per octree node." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 531, + 287, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 531, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 531, + 287, + 616 + ], + "type": "text", + "content": "be inside). The IoU of the octree is much lower than the IoU of the final mesh, indicating that the INR training is learning beyond the octree supervision. Furthermore as the depth increases, the octree IoU improves and so do the metrics on the final mesh, indicating that the INR training is heavily affected by the initial training guidance from the octree. See the supplemental material for a depth ablation on SRB." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 623, + 167, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 623, + 167, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 623, + 167, + 635 + ], + "type": "text", + "content": "4.3. Modifying the octree" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "content": "To show the versatility of our method, in Table 2 we demonstrate how modifying the octree labelling can affect the performance of the resulting shape. For " + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{OG - INR + GT^o}" + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "content": " we set octree labels using GT occupancy. For " + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{OG - INR + H^o}" + }, + { + "bbox": [ + 46, + 641, + 288, + 715 + ], + "type": "text", + "content": " we set octree labels using sparse human supervision: we change up to 30 large octree leaves at the lower depths (up" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 340, + 547, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 340, + 547, + 387 + ], + "spans": [ + { + "bbox": [ + 304, + 340, + 547, + 387 + ], + "type": "text", + "content": "to depth 5) where the changes are obvious from a human perspective. With both octree supervision signals we get increases in performance, showing that the discrete nature of octree guidance does not introduce much bias." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 388, + 547, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 388, + 547, + 424 + ], + "spans": [ + { + "bbox": [ + 304, + 388, + 547, + 424 + ], + "type": "text", + "content": "We also compare to the INRs with other forms of supervision. As NGLOD cannot be used in an unsupervised setting, we provide result when trained on dense GT." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 431, + 432, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 431, + 432, + 445 + ], + "spans": [ + { + "bbox": [ + 305, + 431, + 432, + 445 + ], + "type": "text", + "content": "4.4. Further Comparisons" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 450, + 546, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 450, + 546, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 546, + 594 + ], + "type": "text", + "content": "The need for unoriented methods. We compare to a recent industry-standard implementation [48] of normal estimation + SPSR [19] (SPSR+N Est.). Table 1 (left) shows that for the well sampled surfaces in ShapeNet, when ground truth normals are available SPSR achieves close to perfect IoUs. However when normals need to be estimated, it has good median performance but very poor average performance, indicating the normal estimation procedure will often be quite inaccurate. Table 1 (right) shows that for point clouds with large missing regions, SPSR with ground truth normals has room for improvement, while SPSR with estimated normals performs extremely poorly." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 594, + 547, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 654 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 654 + ], + "type": "text", + "content": "While neither dataset has provided line of sight information (as would be given by sensors), a more informative use of such information than estimating rough normals would be to use it to label parts of our octree. See Section 4.3 for how additional octree supervision improves performance." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "content": "Overall Timing Comparison. Table 3 shows the time taken for each method. Our method's gradient descent training is the fastest to converge as it needs much fewer iterations. Even when adding on the average octree building time from ShapeNet, we can see that the time taken for both" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16723" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 105, + 70, + 491, + 167 + ], + "blocks": [ + { + "bbox": [ + 105, + 70, + 491, + 167 + ], + "lines": [ + { + "bbox": [ + 105, + 70, + 491, + 167 + ], + "spans": [ + { + "bbox": [ + 105, + 70, + 491, + 167 + ], + "type": "table", + "html": "
MethodParametersTime per iter. (s)Num itersTime (s)Speed Up
N Est. +SPSR---42 (13 + 29)12× (· + 18×)
SIREN (wo n)264K0.052100005201×
SAL2.1M0.1751000017500.3×
DiGS264K0.1201000012000.4×
SAP120K-32003301.6×
Our OG-SIREN264K0.158600135 (40 + 95)3.9× (· + 5.5×)
Our OG-NGLOD68.7M0.17330092 (40 + 52)5.7× (· + 10×)
", + "image_path": "6ca24abf1e208c0fd895580adc7478bffe3cfeaa04815fad900f335600ede9c1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 76, + 219, + 517, + 294 + ], + "blocks": [ + { + "bbox": [ + 46, + 175, + 547, + 210 + ], + "lines": [ + { + "bbox": [ + 46, + 175, + 547, + 210 + ], + "spans": [ + { + "bbox": [ + 46, + 175, + 547, + 210 + ], + "type": "text", + "content": "Table 3. Time and number of parameters for each model on ShapeNet. For methods with a precomputation stage (normal estimation or octree building), the time for both is included with the precomputation and training times in parenthesis. Speed up is measured with respect to SIREN/SIREN w o n. We ran our experiments on a single Nvidia RTX 2080 GPU and a i7-8700K CPU." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 76, + 219, + 517, + 294 + ], + "lines": [ + { + "bbox": [ + 76, + 219, + 517, + 294 + ], + "spans": [ + { + "bbox": [ + 76, + 219, + 517, + 294 + ], + "type": "table", + "html": "
DepthMeanOctree IoU ↑ MedianStdMeanFinal Mesh (OG-SIREN)Squared Chamfer ↓Octree Build + Label Time (s)
IoU ↑ MedianStdMeanMedianStdMeanMedianStd
50.47480.47020.18540.71010.77120.25941.84e-41.35e-42.23e-42.21.62.2
60.64010.64100.16040.95030.98350.11723.93e-52.34e-58.20e-57.65.38.6
70.77570.80060.13050.96150.98710.10483.75e-52.17e-58.24e-538.119.367.7
", + "image_path": "ec1f68ec8677136d33895e116f2ed17b8b3166a5e5cc666623c7059a777613b7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 52, + 302, + 539, + 314 + ], + "lines": [ + { + "bbox": [ + 52, + 302, + 539, + 314 + ], + "spans": [ + { + "bbox": [ + 52, + 302, + 539, + 314 + ], + "type": "text", + "content": "Table 4. Octree depth ablation on ShapeNet. Octree IoU is computed using the labelling of the leaves (surface leaves chosen as inside)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 335, + 287, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 335, + 287, + 370 + ], + "spans": [ + { + "bbox": [ + 46, + 335, + 287, + 370 + ], + "type": "text", + "content": "stages of our method is still faster than other methods, often by an order of magnitude. We implemented the octree in Cython [3] but further optimisations are possible." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 372, + 288, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 372, + 288, + 540 + ], + "spans": [ + { + "bbox": [ + 46, + 372, + 288, + 540 + ], + "type": "text", + "content": "Comparison Against No Octree Guidance. Comparing SIREN wo n with and without the octree guidance we can see that the guidance drastically improves both time performance and reconstruction metrics (see Table 1,Table 3). When initially guided by our octree labelling, SIREN (wo n) converges significantly faster (92s vs 520s) and gets a 3x improvement in ShapeNet IoU. Without guidance from normals or our octree, SIREN wo n gets stuck in bad local minima where it needs to flip its inside-outsideness prediction, causing ghost geometry (see Figure 1). When comparing the two types of INR training guidance applied to SIREN wo n, ground truth normals vs our octree, we can see that the octree guidance still performs much better than SIREN on ShapeNet, while performing comparably on SRB." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 549, + 119, + 561 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 549, + 119, + 561 + ], + "spans": [ + { + "bbox": [ + 47, + 549, + 119, + 561 + ], + "type": "text", + "content": "4.5. Discussion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 568, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 568, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 568, + 287, + 653 + ], + "type": "text", + "content": "Limitations. Like all unoriented methods, the octree guidance has difficulty with large missing regions, which is demonstrated heavily in SRB, and thin surfaces, which appear in SRB (Daratech) and ShapeNet (airplane and watercraft). Despite this, our methods perform competitively on shapes with these difficulties on both datasets. See the supplemental for per class breakdown and visualizations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": "Other possible limitations would be the ability to deal with noise and outliers. Note that all methods do poorly with large amounts of either corruption. For small noise the method's performance does not degrade considerably as demonstrated by the results on SRB. For larger noise, the" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 335, + 545, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 335, + 545, + 394 + ], + "spans": [ + { + "bbox": [ + 304, + 335, + 545, + 394 + ], + "type": "text", + "content": "volume occupied by surface leaves would be much larger. While this would not introduce error, the octree's benefits would be reduced. For outliers there would be significant bias introduced by the octree method, which if sparse could be preprocessed out beforehand." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 396, + 545, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 396, + 545, + 456 + ], + "spans": [ + { + "bbox": [ + 304, + 396, + 545, + 456 + ], + "type": "text", + "content": "Broader Impact. Faster and more accurate unoriented surface reconstruction methods will allow for accurate meshes to be constructed easily from regular scanners. Potential misuses of this technology include unauthorized replication and harmful content creation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 471, + 378, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 471, + 378, + 483 + ], + "spans": [ + { + "bbox": [ + 306, + 471, + 378, + 483 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 492, + 545, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 492, + 545, + 648 + ], + "spans": [ + { + "bbox": [ + 304, + 492, + 545, + 648 + ], + "type": "text", + "content": "We have introduced an octree guided method for training implicit neural representations (INRs) for the task of unoriented surface reconstruction. We formulate the main difficulty of the task, inside-outside determinism, as an octree labelling problem with respect to an energy function that captures watertight surface properties and the minimal surface constraint. We also detail an efficient method to minimize the energy. We then use this supervision to guide the initial stage of INR training, using loss functions that can be used by any INR architecture. Our results show that applying our octree guided supervision provides significant time and reconstruction performance improvements, with competitive or better results than the current state-of-the-art." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 715 + ], + "type": "text", + "content": "Acknowledgements. This work has received funding from the European Union's Horizon 2020 research and innovation programme under the Marie Sklodowska-Curie grant agreement No. 893465. S. Gould is a recipient of an ARC Future Fellowship (proj. no. LP200100421) funded by the Australian Government." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "16724" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Matan Atzmon and Yaron Lipman. SAL: Sign Agnostic Learning of shapes from raw data. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2565-2574, 2020. 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 136, + 287, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 136, + 287, + 168 + ], + "spans": [ + { + "bbox": [ + 53, + 136, + 287, + 168 + ], + "type": "text", + "content": "[2] Matan Atzmon and Yaron Lipman. SALD: Sign Agnostic Learning with Derivatives. In Proc. of the International Conference on Learning Representations (ICLR), 2021. 2, 3, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 168, + 287, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 168, + 287, + 212 + ], + "spans": [ + { + "bbox": [ + 53, + 168, + 287, + 212 + ], + "type": "text", + "content": "[3] Stefan Behnel, Robert Bradshaw, Craig Citro, Lisandro Dalcin, Dag Sverre Seljebotn, and Kurt Smith. Cython: The best of both worlds. Computing in Science & Engineering, 13(2):31-39, 2011. 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 213, + 287, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 213, + 287, + 257 + ], + "spans": [ + { + "bbox": [ + 53, + 213, + 287, + 257 + ], + "type": "text", + "content": "[4] Yizhak Ben-Shabat and Stephen Gould. DeepFit: 3D surface fitting via neural network weighted least squares. In Proc. of the European Conference on Computer Vision (ECCV), pages 20-34. Springer, 2020. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 258, + 287, + 301 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 258, + 287, + 301 + ], + "spans": [ + { + "bbox": [ + 53, + 258, + 287, + 301 + ], + "type": "text", + "content": "[5] Yizhak Ben-Shabat, Chamin Hewa Koneputugodage, and Stephen Gould. Digs: Divergence guided shape implicit neural representation for unoriented point clouds. arXiv preprint arXiv:2106.10811, 2021. 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 302, + 287, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 302, + 287, + 357 + ], + "spans": [ + { + "bbox": [ + 53, + 302, + 287, + 357 + ], + "type": "text", + "content": "[6] Yizhak Ben-Shabat, Michael Lindenbaum, and Anath Fischer. Nesti-net: Normal estimation for unstructured 3D point clouds using convolutional neural networks. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 10112-10120, 2019. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 357, + 287, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 357, + 287, + 401 + ], + "spans": [ + { + "bbox": [ + 53, + 357, + 287, + 401 + ], + "type": "text", + "content": "[7] Matthew Berger, Joshua A Levine, Luis Gustavo Nonato, Gabriel Taubin, and Claudio T Silva. A benchmark for surface reconstruction. ACM Trans. on Graphics (ToG), 32:1-17, 2013. 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 402, + 287, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 402, + 287, + 435 + ], + "spans": [ + { + "bbox": [ + 53, + 402, + 287, + 435 + ], + "type": "text", + "content": "[8] Julian Besag. On the statistical analysis of dirty pictures. Journal of the Royal Statistical Society: Series B (Methodological), 48(3):259-279, 1986. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 435, + 287, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 435, + 287, + 468 + ], + "spans": [ + { + "bbox": [ + 53, + 435, + 287, + 468 + ], + "type": "text", + "content": "[9] Andrew Blake, Pushmeet Kohli, and Carsten Rother. Markov Random Fields for Vision and Image Processing. The MIT Press, 07 2011. 3, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 469, + 287, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 469, + 287, + 513 + ], + "spans": [ + { + "bbox": [ + 47, + 469, + 287, + 513 + ], + "type": "text", + "content": "[10] Yuri Boykov, Olga Veksler, and Ramin Zabih. Fast approximate energy minimization via graph cuts. IEEE Transactions on pattern analysis and machine intelligence, 23(11):1222-1239, 2001. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 514, + 287, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 514, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 47, + 514, + 287, + 578 + ], + "type": "text", + "content": "[11] Jonathan C Carr, Richard K Beatson, Jon B Cherrie, Tim J Mitchell, W Richard Fright, Bruce C McCallum, and Tim R Evans. Reconstruction and representation of 3D objects with radial basis functions. In Proc. of the Conference on Computer Graphics and Interactive Techniques, pages 67-76, 2001. 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 579, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 579, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 579, + 287, + 635 + ], + "type": "text", + "content": "[12] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. ShapeNet: An information-rich 3D model repository. arXiv preprint arXiv:1512.03012, 2015. 6, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 635, + 288, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 288, + 679 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 288, + 679 + ], + "type": "text", + "content": "[13] Stephen Gould. Max-margin learning for lower linear envelope potentials in binary markov random fields. In Proc. of the International Conference on Machine Learning (ICML). Omnipress, 2011. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 680, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 680, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 680, + 288, + 714 + ], + "type": "text", + "content": "[14] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit Geometric Regularization for learning shapes. In Proc. of the International Conference on" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 714 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "Machine Learning (ICML), volume 119, pages 3789-3799. \nPMLR, 2020. 2, 3, 5, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 95, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 95, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 95, + 545, + 140 + ], + "type": "text", + "content": "[15] Paul Guerrero, Yanir Kleiman, Maks Ovsjanikov, and Niloy J Mitra. PCPNet: Learning local shape properties from raw point clouds. In Computer Graphics Forum, volume 37, pages 75-85. Wiley Online Library, 2018. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 140, + 545, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 140, + 545, + 182 + ], + "spans": [ + { + "bbox": [ + 307, + 140, + 545, + 182 + ], + "type": "text", + "content": "[16] Christian Hane, Shubham Tulsiani, and Jitendra Malik. Hierarchical surface prediction for 3d object reconstruction. In 2017 International Conference on 3D Vision (3DV), pages 412-420. IEEE, 2017. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 183, + 545, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 183, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 307, + 183, + 545, + 270 + ], + "type": "text", + "content": "[17] Shahram Izadi, David Kim, Otmar Hilliges, David Molyneaux, Richard Newcombe, Pushmeet Kohli, Jamie Shotton, Steve Hodges, Dustin Freeman, Andrew Davison, and Andrew Fitzgibbon. Kinectfusion: Real-time 3d reconstruction and interaction using a moving depth camera. In UIST '11 Proceedings of the 24th annual ACM symposium on User interface software and technology, pages 559-568. ACM, October 2011. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 271, + 545, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 271, + 545, + 314 + ], + "spans": [ + { + "bbox": [ + 307, + 271, + 545, + 314 + ], + "type": "text", + "content": "[18] Michael Kazhdan, Matthew Bolitho, and Hugues Hoppe. Poisson surface reconstruction. In Proc. of the Eurographics Symposium on Geometry Processing (SGP), volume 7, 2006. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 316, + 545, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 316, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 307, + 316, + 545, + 348 + ], + "type": "text", + "content": "[19] Michael Kazhdan and Hugues Hoppe. Screened poisson surface reconstruction. ACM Trans. on Graphics (ToG), 32:1-13, 2013. 2, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 349, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 349, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 307, + 349, + 545, + 392 + ], + "type": "text", + "content": "[20] Pushmeet Kohli, M. Kumar, and Philip Torr. P3 & beyond: Solving energies with higher order cliques. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2007. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 393, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 393, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 307, + 393, + 545, + 437 + ], + "type": "text", + "content": "[21] Jan Eric Lenssen, Christian Osendorfer, and Jonathan Masci. Deep iterative surface normal estimation. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 11247-11256, 2020. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 437, + 545, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 437, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 307, + 437, + 545, + 469 + ], + "type": "text", + "content": "[22] Yaron Lipman. Phase transitions, distance functions, and implicit neural representations. arXiv preprint arXiv:2106.07689, 2021. 2, 3, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 470, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 470, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 307, + 470, + 545, + 514 + ], + "type": "text", + "content": "[23] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. Advances in Neural Information Processing Systems, 33:15651-15663, 2020. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 514, + 545, + 570 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 514, + 545, + 570 + ], + "spans": [ + { + "bbox": [ + 307, + 514, + 545, + 570 + ], + "type": "text", + "content": "[24] Shi-Lin Liu, Hao-Xiang Guo, Hao Pan, Peng-Shuai Wang, Xin Tong, and Yang Liu. Deep implicit moving least-squares functions for 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1788-1797, 2021. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 570, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 570, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 307, + 570, + 545, + 624 + ], + "type": "text", + "content": "[25] William E Lorensen and Harvey E Cline. Marching cubes: A high resolution 3D surface construction algorithm. In Proc. of the Intl. Conference on Computer Graphics and Interactive Techniques (SIGGRAPH), volume 21, pages 163-169. ACM New York, NY, USA, 1987. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 624, + 545, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 669 + ], + "type": "text", + "content": "[26] Julien NP Martel, David B Lindell, Connor Z Lin, Eric R Chan, Marco Monteiro, and Gordon Wetzstein. Acorn: Adaptive coordinate networks for neural scene representation. arXiv preprint arXiv:2105.02788, 2021. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 669, + 545, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 545, + 691 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 545, + 691 + ], + "type": "text", + "content": "[27] Donald J. Meagher. Geometric modeling using octree encoding. Comput. Graph. Image Process., 19:85, 1982. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 691, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 714 + ], + "type": "text", + "content": "[28] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks:" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16725" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 106 + ], + "type": "text", + "content": "Learning 3D reconstruction in function space. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4460-4470, 2019. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 287, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 287, + 151 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 287, + 151 + ], + "type": "text", + "content": "[29] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, July 2022. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 288, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 288, + 207 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 288, + 207 + ], + "type": "text", + "content": "[30] Jeong Joon Park, Peter Florence, Julian Straub, Richard Newcombe, and Steven Lovegrove. DeepSDF: Learning continuous signed distance functions for shape representation. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 165-174, 2019. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 209, + 287, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 209, + 287, + 252 + ], + "spans": [ + { + "bbox": [ + 48, + 209, + 287, + 252 + ], + "type": "text", + "content": "[31] Songyou Peng, Chiyu Jiang, Yiyi Liao, Michael Niemeyer, Marc Pollefeys, and Andreas Geiger. Shape As Points: A differentiable poisson solver. arXiv preprint arXiv:2106.03452, 2021. 3, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 254, + 287, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 254, + 287, + 298 + ], + "spans": [ + { + "bbox": [ + 48, + 254, + 287, + 298 + ], + "type": "text", + "content": "[32] Gernot Riegler, Ali Osman Ulusoy, and Andreas Geiger. Octnet: Learning deep 3d representations at high resolutions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3577-3586, 2017. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 300, + 287, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 300, + 287, + 354 + ], + "spans": [ + { + "bbox": [ + 48, + 300, + 287, + 354 + ], + "type": "text", + "content": "[33] Bernhard Scholkopf, Joachim Giesen, and Simon Spalinger. Kernel methods for implicit surface modeling. In Advances in Neural Information Processing Systems (NeurIPS), NIPS'04, page 1193-1200, Cambridge, MA, USA, 2004. MIT Press. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 355, + 287, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 355, + 287, + 409 + ], + "spans": [ + { + "bbox": [ + 48, + 355, + 287, + 409 + ], + "type": "text", + "content": "[34] Vincent Sitzmann, Julien Martel, Alexander Bergman, David Lindell, and Gordon Wetzstein. Implicit neural representations with periodic activation functions. Advances in Neural Information Processing Systems (NeurIPS), 33, 2020. 2, 5, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 411, + 287, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 411, + 287, + 477 + ], + "spans": [ + { + "bbox": [ + 48, + 411, + 287, + 477 + ], + "type": "text", + "content": "[35] Towaki Takikawa, Joey Litalien, Kangxue Yin, Karsten Kreis, Charles Loop, Derek Nowrouzezahrai, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Neural geometric level of detail: Real-time rendering with implicit 3D shapes. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 479, + 287, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 479, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 48, + 479, + 287, + 544 + ], + "type": "text", + "content": "[36] Matthew Tancik, Pratul P. Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan T. Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. Advances in Neural Information Processing Systems (NeurIPS), 2020. 2, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 545, + 287, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 545, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 545, + 287, + 601 + ], + "type": "text", + "content": "[37] Jia-Heng Tang, Weikai Chen, Jie Yang, Bo Wang, Songrun Liu, Bo Yang, and Lin Gao. Octfield: Hierarchical implicit functions for 3d modeling. In The Thirty-Fifth Annual Conference on Neural Information Processing Systems (NeurIPS), 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 602, + 287, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 287, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 287, + 656 + ], + "type": "text", + "content": "[38] Maxim Tatarchenko, Alexey Dosovitskiy, and Thomas Brox. Octree generating networks: Efficient convolutional architectures for high-resolution 3d outputs. In Proceedings of the IEEE international conference on computer vision, pages 2088-2096, 2017. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[39] Ayush Tewari, Justus Thies, Ben Mildenhall, Pratul Srinivasan, Edgar Tretschk, Yifan Wang, Christoph Lassner, Vincent Sitzmann, Ricardo Martin-Brualla, Stephen Lombardi, et al. Advances in neural rendering. arXiv preprint arXiv:2111.05849, 2021. 1" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 507 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[40] Peng-Shuai Wang, Yang Liu, Yu-Xiao Guo, Chun-Yu Sun, and Xin Tong. O-cnn: Octree-based convolutional neural networks for 3d shape analysis. ACM Transactions On Graphics (TOG), 36(4):1-11, 2017. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 118, + 545, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 172 + ], + "type": "text", + "content": "[41] Peng-Shuai Wang, Yang Liu, and Xin Tong. Deep octree-based cnns with output-guided skip connections for 3d shape and scene completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 266-267, 2020. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 174, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 174, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 174, + 545, + 217 + ], + "type": "text", + "content": "[42] Peng-Shuai Wang, Yang Liu, and Xin Tong. Dual octree graph networks for learning adaptive volumetric shape representations. ACM Transactions on Graphics (TOG), 41(4):1-15, 2022. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 218, + 545, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 218, + 545, + 261 + ], + "spans": [ + { + "bbox": [ + 307, + 218, + 545, + 261 + ], + "type": "text", + "content": "[43] Peng-Shuai Wang, Chun-Yu Sun, Yang Liu, and Xin Tong. Adaptive o-cnn: A patch-based deep representation of 3d shapes. ACM Transactions on Graphics (TOG), 37(6):1-11, 2018. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 263, + 545, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 263, + 545, + 318 + ], + "spans": [ + { + "bbox": [ + 307, + 263, + 545, + 318 + ], + "type": "text", + "content": "[44] Francis Williams, Teseo Schneider, Claudio Silva, Denis Zorin, Joan Bruna, and Daniele Panozzo. Deep Geometric Prior for surface reconstruction. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 10130-10139, 2019. 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 319, + 545, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 319, + 545, + 373 + ], + "spans": [ + { + "bbox": [ + 307, + 319, + 545, + 373 + ], + "type": "text", + "content": "[45] Francis Williams, Matthew Trager, Joan Bruna, and Denis Zorin. Neural Splines: Fitting 3D surfaces with infinitely-wide neural networks. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 9949-9958, 2021. 2, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 375, + 545, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 375, + 545, + 430 + ], + "spans": [ + { + "bbox": [ + 307, + 375, + 545, + 430 + ], + "type": "text", + "content": "[46] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond. Computer Graphics Forum, 2022. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 431, + 545, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 431, + 545, + 473 + ], + "spans": [ + { + "bbox": [ + 307, + 431, + 545, + 473 + ], + "type": "text", + "content": "[47] Wang Yifan, Lukas Rahmann, and Olga Sorkine-Hornung. Geometry-consistent neural shape representation with implicit displacement fields. arXiv preprint arXiv:2106.05187, 2021. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 475, + 545, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 475, + 545, + 507 + ], + "spans": [ + { + "bbox": [ + 307, + 475, + 545, + 507 + ], + "type": "text", + "content": "[48] Qian-Yi Zhou, Jaesik Park, and Vladlen Koltun. Open3D: A modern library for 3D data processing. arXiv:1801.09847, 2018. 7" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16726" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Omni Aggregation Networks for Lightweight Image Super-Resolution/39175e81-b86e-49a9-967e-e590c3dbdaa0_content_list.json b/2023/Omni Aggregation Networks for Lightweight Image Super-Resolution/39175e81-b86e-49a9-967e-e590c3dbdaa0_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8e430b9d7fa5696189b476cce0cbb61c27c0257b --- /dev/null +++ b/2023/Omni Aggregation Networks for Lightweight Image Super-Resolution/39175e81-b86e-49a9-967e-e590c3dbdaa0_content_list.json @@ -0,0 +1,2041 @@ +[ + { + "type": "text", + "text": "Omni Aggregation Networks for Lightweight Image Super-Resolution", + "text_level": 1, + "bbox": [ + 133, + 130, + 834, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hang Wang $^{2*}$ Xuanhong Chen $^{1*}$ Bingbing Ni $^{1,2\\dagger}$ Yutian Liu $^{1}$ Jinfan Liu $^{1}$", + "bbox": [ + 166, + 179, + 795, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Shanghai Jiao Tong University, Shanghai 200240, China $^{2}$ Huawei", + "bbox": [ + 210, + 198, + 759, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "francis970625@gmail.com {chen19910528,nibingbing,stau7001,ljflnjz}@sjtu.edu.cn", + "bbox": [ + 140, + 215, + 821, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 312, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While lightweight ViT framework has made tremendous progress in image super-resolution, its uni-dimensional self-attention modeling, as well as homogeneous aggregation scheme, limit its effective receptive field (ERF) to include more comprehensive interactions from both spatial and channel dimensions. To tackle these drawbacks, this work proposes two enhanced components under a new Omni-SR architecture. First, an Omni Self-Attention (OSA) block is proposed based on dense interaction principle, which can simultaneously model pixel-interaction from both spatial and channel dimensions, mining the potential correlations across omni-axis (i.e., spatial and channel). Coupling with mainstream window partitioning strategies, OSA can achieve superior performance with compelling computational budgets. Second, a multi-scale interaction scheme is proposed to mitigate sub-optimal ERF (i.e., premature saturation) in shallow models, which facilitates local propagation and meso-/global-scale interactions, rendering an omni-scale aggregation building block. Extensive experiments demonstrate that Omni-SR achieves record-high performance on lightweight super-resolution benchmarks (e.g., 26.95dB@Urban100 ×4 with only 792K parameters). Our code is available at https://github.com/Francis0625/Omni-SR.", + "bbox": [ + 75, + 300, + 473, + 662 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 690, + 209, + 705 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Image super-resolution (SR) is a long-standing low-level problem that aims to recover high-resolution (HR) images from degraded low-resolution (LR) inputs. Recently, vision transformer [14, 51] based (i.e., ViT-based) SR frameworks [5, 31] have emerged, showing significant performance gains compared to previously dominant Convolutional Neural Networks (CNNs) [66]. However, most attempts [31] are devoted to improving the large-scale ViT-based models, while the development of lightweight ViTs (typically, less than 1M parameters) remains fraught with", + "bbox": [ + 75, + 715, + 468, + 867 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/83843f4b5574f4fcbc3b18bdae706065a44d808b3483f1033fb43f0367b2a5b6.jpg", + "image_caption": [ + "Spatial SA" + ], + "image_footnote": [], + "bbox": [ + 511, + 276, + 624, + 373 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/71a6f804e8ab1375f1e3508b309d41565c74a41c0618e0ff295d86651f9eb999.jpg", + "image_caption": [ + "Channel SA" + ], + "image_footnote": [], + "bbox": [ + 637, + 276, + 751, + 373 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8086c58f33b0a6019dc2e25b33792325b411b68b2bf8b57692f6c1666aee8eb1.jpg", + "image_caption": [ + "Omni SA" + ], + "image_footnote": [], + "bbox": [ + 763, + 276, + 877, + 373 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d30ea2a4d9bfa5b22bdcffc7e2e332265caf4a9f549ff90fd94720220b464bbc.jpg", + "image_caption": [ + "PSNR" + ], + "image_footnote": [], + "bbox": [ + 513, + 401, + 612, + 510 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/af7413c8721fbc957a54d747d3e3cbb64dfd53dd21045b42dff47fc63bd5563e.jpg", + "image_caption": [ + "Parameters (K)" + ], + "image_footnote": [], + "bbox": [ + 616, + 401, + 750, + 511 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b10e420f7e0a0818b6694462ea8c62a3c345ce4bdc0ae41e0f367c9f1b9849e6.jpg", + "image_caption": [ + "Iteration", + "Figure 1. Typical self-attention schemes [31,59] can only perform uni-dimensional (e.g., spatial-only) interactions, and." + ], + "image_footnote": [], + "bbox": [ + 750, + 401, + 874, + 511 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "difficulties. This paper focuses on boosting the restoration performance of lightweight ViT-based frameworks.", + "bbox": [ + 498, + 579, + 890, + 609 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Two difficulties hinder the development of lightweight ViT-based models: 1) Uni-dimensional aggregation operators (i.e., spatial [31] only or channel [59] only) imprisons the full potential of self-attention operators. Contemporary self-attention generally realizes the interaction between pixels by calculating the cross-covariance of the spatial direction (i.e., width and height) and exchanges context information in a channel-separated manner. This interaction scheme ignores the explicit use of channel information. However, recent evidences [59] and our practice show that self-attention in the channel dimension (i.e., computationally more compact than spatial self-attention) is also crucial in low-level tasks. 2) Homogeneous aggregation schemes (i.e., Simple hierarchical stacking of single operators, e.g., convolution, self-attention) neglect abundant texture patterns of multi-scales, which is urgently needed in SR task. Specifically, a single operator is only sensitive to information of one scale [6, 12], e.g., self-attention is sensitive to long-term information and pays little attention to lo", + "bbox": [ + 496, + 613, + 892, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal Contribution.", + "bbox": [ + 94, + 875, + 207, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author: Bingbing Ni.", + "bbox": [ + 96, + 887, + 294, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "22378", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "cal information. Additionally, stacking of homogeneous operators proves to be inefficient and suffers from premature saturation of the interaction range [8], which is reflected as a suboptimal effective receptive field. The above problem is exacerbated in lightweight models because lightweight models cannot stack enough layers.", + "bbox": [ + 75, + 90, + 468, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In order to solve the above problems and pursue higher performance, this work proposes a novel omni-dimension feature aggregation scheme called Omni Self-Attention (OSA) exploiting both spatial and channel axis information in a simultaneous manner (i.e., extends the interaction into three-dimensional space), which offers higher-order receptive field information, as shown in Figure 1. Unlike scalar-based (a group of important coefficient) channel interaction [19], OSA enables comprehensive information propagation and interaction by cascading computation of the cross-covariance matrices between spatial/channel dimensions. The proposed OSA module can be plugged into any mainstream self-attention variants (e.g., Swin [35], Halo [50]), which provides a finer granularity of important encoding (compared to the vanilla channel attention [19]), achieving a perceptible improvement in contextual aggregation capabilities. Furthermore, a multi-scale hierarchical aggregation block, named Omni-Scale Aggregation Group (i.e., OSAG for short), is presented to achieve tailored encoding of varying scales of texture patterns. Specifically, OSAG builds three cascaded aggregators: local convolution (for local details), meso self-attention (focusing on mid-scale pattern processing), and global self-attention (pursuing global context understanding), rendering an omni-scale (i.e., local-/meso-/global-scale simultaneously) feature extraction capability. Compared to the homogenized feature extraction schemes [28, 31], our OSAG is able to mine richer information producing features with higher information entropy. Coupling with the above two designs, we establish a new ViT-based framework for lightweight superresolution, called Omni-SR, which exhibits superior restoration performance as well as covers a larger interaction range while maintaining an attractive model size, i.e., $792K$ .", + "bbox": [ + 75, + 181, + 472, + 680 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We extensively experiment with the proposed framework with both qualitative and quantitative evaluations on mainstream open-source image super-resolution datasets. It is demonstrated that our framework achieves state-of-the-art performance at the lightweight model scale (e.g., $\\text{Urban100} \\times 4: 26.95\\mathrm{dB}$ , $\\text{Manga109} \\times 4: 31.50\\mathrm{dB}$ ). More importantly, compared to existing ViT-based super-solution frameworks, our framework shows superior optimization properties (e.g., convergence speed, smoother loss landscape), which endow our model with better robustness.", + "bbox": [ + 75, + 680, + 468, + 833 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 76, + 845, + 225, + 861 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image Super-resolution. CNNs have achieved remarkable success in image SR task. SRCNN [13] is the first", + "bbox": [ + 76, + 869, + 468, + 902 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0a11c3f9b6aa9678b0a36037821e2d50d310e88c6867c243403df9e178d054ae.jpg", + "image_caption": [ + "Figure 2. Illustration of spatial attention and channel attention. These typical attention paradigms only model uni-dimensional (i.e., spatial-only / channel-only) interaction." + ], + "image_footnote": [], + "bbox": [ + 501, + 88, + 694, + 220 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/da9fa7abd21172944f7bbcafab2e7f8521fdb44f49ab186ce3530fcde058bbce.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 88, + 890, + 220 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "work to introduce CNNs into the SR field. Many methods [25, 48, 66] employ skip connection to speed up network convergence and improve the reconstruction quality. Channel Attention [66] is also proposed to enhance the representation ability of the SR model. In order to obtain better reconstruction quality with limited computing resources, several methods [23,38,42,47] explore lightweight architectural design. DRCN [26] utilizes the recursive operation to reduce the number of parameters. DRRN [47] introduces global and local residual learning on the basis of DRCN to accelerate training and improve the quality of details. CARN [1] employs cascading mechanism upon a residual network. IMDN [22] proposes an information multistillation block to archive better time performance. Another line of research is to utilize model compression techniques, e.g., knowledge distillation [15, 17, 65] and neural architecture search [11]) to reduce computing costs. Recently, a series of transformer-based SR models [5,8,31,37] emerge with superior performance. Chenet al. [5] develop a pre-trained model for the low-level computer vision task using the transformer architecture. Based on Swin transformer [35], SwinIR [31] proposes a three-stage framework, refreshing the state-of-the-art of SR task. More recently, some works [5,30] explore ImageNet pre-training strategy to further enhance SR performance.", + "bbox": [ + 496, + 292, + 892, + 670 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Lightweight Vision Transformer. Due to the urgent demands for applying networks to resource-constrained devices, lightweight vision transformer [14, 51] has attracted widespread attention. Many attempts [7,9,10,37,41,43,57, 62] have been made to develop lightweight ViTs with comparable performance. A series of methods focus on combining convolutions with transformers to learn both local and global representations. For instance, LVT [57] introduces convolution in self-attention to enrich low-level features. MobileViT [41] replaces matrix multiplication in convolutions with transformer layers to learn global representations. Similarly, EdgeViTs [43] employs an information exchange bottleneck for full spatial interactions. Different from interpreting convolutions into vision transformers, LightViT [21] proposes aggregated self-attention for better", + "bbox": [ + 496, + 674, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "22379", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/118b879237fe11c15116324502a94b71f3b9cb241e4e3748c0ae1532450ec6db.jpg", + "image_caption": [ + "Figure 3. The overall architecture of the proposed Omni-SR framework and structure of OSAG and Omni Self-Attention (OSA)." + ], + "image_footnote": [], + "bbox": [ + 78, + 87, + 890, + 344 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "information aggregation. In this work, we resort to ViT architecture to achieve lightweight and accurate SR.", + "bbox": [ + 75, + 382, + 470, + 412 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 76, + 444, + 210, + 462 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Attention Mechanisms in Super-Resolution", + "text_level": 1, + "bbox": [ + 76, + 474, + 442, + 492 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Two attention paradigms are widely adopted in SR to assist in analyzing and aggregating comprehensive patterns.", + "bbox": [ + 75, + 505, + 468, + 536 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Spatial Attention. Spatial attention can be regarded as an anisotropic selection process. Spatial self-attention [37, 51] and spatial gate [10, 58] are predominantly applied. As shown in Figure 2, spatial self-attention calculates the cross-covariance along the spatial dimension, and the spatial gate generates the channel-separated masks. Neither of them can transmit information between channels.", + "bbox": [ + 75, + 541, + 468, + 647 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Channel Attention. There are two categories of channel attention, i.e., scalar-based [19] and covariance-based [59], proposed to perform channel recalibration or transmit patterns among channels. As shown in Figure 2, the former predicts a group of importance scalars to weigh different channels, while the latter computes a cross-covariance matrix to enable channel re-weighting and information transmission simultaneously. Compared to spatial attention, channel attention handles spatial dimension isotropically, and thus, the complexity is significantly reduced, which also impairs the accuracy of aggregation.", + "bbox": [ + 75, + 652, + 470, + 820 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Several attempts [44,55] have proved that both spatial attention and channel attention are beneficial for SR task and their characteristics are complementary to each other, thus integrating them in a computationally compact way would bring notable benefits in expressive capability.", + "bbox": [ + 75, + 824, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Omni Self-Attention Block", + "text_level": 1, + "bbox": [ + 500, + 382, + 741, + 398 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To mine all the correlations hidden in the latent variables, we propose a novel self-attention paradigm called Omni Self-Attention (OSA) block. Unlike existing self-attention paradigms (e.g., spatial self-attention [5, 37, 51]) that only indulge in unidimensional processing, OSA establishes the spatial and channel context simultaneously. The obtained two-dimensional relationship is highly necessary and beneficial, especially for lightweight models. On the one hand, as the network deepens, important information is scattered into different channels [19], and it is critical to deal with them in a timely manner. On the other hand, although spatial self-attention takes advantage of the channel dimension in calculating the covariance, it does not transmit the information between channels (refer to Sec. 3.1). Given the above conditions, our OSA is designed to transmit both spatial and dimensional information in a compact manner.", + "bbox": [ + 496, + 407, + 890, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The proposed OSA calculates the score matrices corresponding to the space and channel direction through sequential matrix operations and rotation, as illustrated in Figure 3. Specifically, suppose $X \\in \\mathbb{R}^{HW \\times C}$ denotes the input feature, where $H$ and $W$ are the width and height of the input, and $C$ is the channel number. Firstly, $X$ is embedded to query, key and value matrices $Q^s, K^s, V^s \\in \\mathbb{R}^{HW \\times C}$ through linearly projection. We calculate the production of query and key to get the spatial attention map of size $\\mathbb{R}^{HW \\times HW}$ . Then we perform the spatial attention to obtain the intermediate aggregated results. Note that window strategy is usually used to significantly reduce the resource overhead. Next stage, we rotate the input query and key matrices to get the transposed query and key matrices $Q^c, K^c \\in \\mathbb{R}^{C \\times HW}$ , and also rotate the value matrices to get the value matrix $V^c \\in \\mathbb{R}^{C \\times HW}$ for the subsequent channel-wise self-attention. The obtained channel-wise attention map of size $\\mathbb{R}^{C \\times C}$ models channel-wise relation", + "bbox": [ + 496, + 650, + 892, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "22380", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ships. Finally, we get the final aggregated $Y_{OSA}$ by the inverse rotation of the channel attention output $Y_{c}$ . The whole OSA process is formulated as follows:", + "bbox": [ + 76, + 90, + 470, + 133 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nQ ^ {s} = X \\cdot \\mathcal {W} _ {q}, K ^ {s} = X \\cdot \\mathcal {W} _ {k}, V ^ {s} = X \\cdot \\mathcal {W} _ {v}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 116, + 143, + 468, + 159 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nY _ {s} = \\mathcal {A} ^ {s} \\left(Q ^ {s}, K ^ {s}, V ^ {s}\\right) = \\operatorname {S o f t M a x} \\left(Q ^ {s} K ^ {s T}\\right) \\cdot V ^ {s}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 117, + 162, + 468, + 179 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nQ ^ {c} = \\mathcal {R} \\left(Q ^ {\\prime}\\right), \\quad K ^ {c} = \\mathcal {R} \\left(K ^ {\\prime}\\right), \\quad V ^ {c} = \\mathcal {R} \\left(V ^ {\\prime}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 117, + 181, + 468, + 196 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nY _ {c} = \\mathcal {A} ^ {c} \\left(Q ^ {c}, K ^ {c}, V ^ {c}\\right) = \\operatorname {S o f t M a x} \\left(K ^ {c} Q ^ {c T}\\right) \\cdot V ^ {c}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 117, + 200, + 468, + 217 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nY _ {O S A} = \\mathcal {R} ^ {- 1} \\left(Y _ {c}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 117, + 220, + 468, + 234 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{W}_q, \\mathcal{W}_k, \\mathcal{W}_v$ denote the linear projection matrices for the query, key, and value, respectively. $Q', K', V'$ are the input embedding matrices of channel-wise self-attention, which are embedded from fore spatial self-attention or or copied directly from $Q^s, K^s, V^s$ . $\\mathcal{R}(\\cdot)$ denotes the rotation operation around spatial axis and $\\mathcal{R}^{-1}(\\cdot)$ is the inverse rotation. Some normalization factors are omitted for the sake of simplicity. In particular, this design shows compelling properties that can integrate the elementwise results of two matrix operations (i.e., spatial-/channel-matrix operation), thereby enabling omni-axial interactions. Note that our proposed OSA paradigm can be a drop-in replacement of the Swin [31,35] attention block to higher performance with less parameters. Benefiting from the smaller attention map size of channel self-attention, the proposed OSA is less computationally intensive compared to the cascade shifted-window self-attention scheme in Swin.", + "bbox": [ + 75, + 247, + 468, + 503 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Discussion with other hybrid attention paradigms. Compared to previous hybrid channel and spatial attention works like CBAM [55] and BAM [44], their scalar-based attention weights only reflect the relative degree of importance, without further inter-pixel information exchange, leading to limited relation modeling capability. Several recent works [8] also incorporate channel attention with spatial Self-attention, but these attempts only resort to scalar weights for channel recalibration, while our OSA paradigm enables channel-wise interaction to mine the potential correlations in omni-axis. Performance comparison of different attention paradigms can be found in Sec. 4.4.", + "bbox": [ + 75, + 505, + 468, + 686 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Omni-Scale Aggregation Group", + "text_level": 1, + "bbox": [ + 76, + 695, + 357, + 712 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "How to utilize the proposed OSA paradigm to build a high-performance and compact network is another key topic. Although hierarchical stacking of windows-based self-attention (e.g., swin [31, 35]) has become mainstream, various works have found that the window-based paradigms are very inefficient for large-range interactions, especially for shallow networks. It is worth pointing out that large-range interaction can provide a pleasing effective receptive field, which is crucial for improving image restoration performance [37]. Unfortunately, direct global interaction is resource-prohibitive and detracts from local aggregation capabilities. Taking these points into account, we propose", + "bbox": [ + 75, + 719, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/daea26868e5c66071cd70cbfa3bee43f8f0b385a1022c93322afdc79d86fd7e6.jpg", + "image_caption": [ + "Figure 4. Illustration of omni-scale aggregation scheme. Our proposed Omni-SR contains three types of feature aggregation at local level, meso and global level, respectively." + ], + "image_footnote": [], + "bbox": [ + 522, + 90, + 870, + 156 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "an Omni-Scale Aggregation Group (i.e., OSAG for short) to pursue progressive receptive field feature aggregation with low computational complexity. As shown in Figure 3, OSAG mainly consists of three stages: local, meso and global aggregations. Specifically, a channel attention [19] enhanced inverted bottleneck [18] is introduced to fulfill the local pattern process with limited overhead. Based on the proposed OSA paradigm, we derive two instances (i.e., Meso-OSA and Global-OSA) responsible for the interaction and aggregation of meso and global information. Note that the proposed omni self-attention paradigm can be used for different purposes. Meso-OSA performs attention within a group of non-overlap patches, which restricts Meso-OSA to only focus on meso-scale pattern understanding. Global-OSA samples data point sparsely across the entire feature within an atrous manner, endowing Global-OSA with the ability to achieve global interactions at a compelling cost.", + "bbox": [ + 496, + 217, + 890, + 489 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The only difference between Meso-OSA and Global-OSA is the window partition strategy, as shown in Figure 4. In order to achieve meso-scale interaction, Meso-OSA split the input feature $X$ into non-overlapping blocks with size $P \\times P$ . Note that after window partition, the block dimensions are gathered onto the spatial dimension (i.e., -2 axis): $(H, W, C) \\to \\left(\\frac{H}{P} \\times P, \\frac{W}{P} \\times P, C\\right) \\to \\left(\\frac{HW}{P^2}, P^2, C\\right)$ . While the Global-OSA divides the input feature into a uniform $G \\times G$ grid, with each lattice having an adaptive size of $\\frac{H}{G} \\times \\frac{W}{G}$ . Similar to Meso-OSA, the grid dimension is also gathered on the spatial axis (i.e., -2 axis): $(H, W, C) \\to (G \\times \\frac{H}{G}, G \\times \\frac{W}{G}, C) \\to (G^2, \\frac{HW}{G^2}, C) \\to (\\frac{HW}{G^2}, G^2, C)$ .", + "bbox": [ + 498, + 489, + 890, + 674 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Network Architecture", + "text_level": 1, + "bbox": [ + 500, + 680, + 705, + 694 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Overall Structure. Based on the Omni Self-Attention paradigm and the Omni-Scale Aggregation Group, we further develop a lightweight Omni-SR framework to achieve high-performance image super-resolution. As shown in Figure 3, Omni-SR consists of three parts, i.e., shallow feature extraction, deep feature extraction, and image reconstruction. Specifically, given the LR input $I_{LR} \\in \\mathbb{R}^{H \\times W \\times C_{in}}$ , we first use a $3 \\times 3$ convolution $H_{SF}$ to extract shallow feature $X_0 \\in \\mathbb{R}^{H \\times W \\times C}$ as", + "bbox": [ + 498, + 700, + 890, + 827 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nX _ {0} = H _ {S F} \\left(I _ {L R}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 643, + 835, + 890, + 849 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $C_{in}$ and $C$ denote the channel number of the input and shallow feature. The convolution layer provides a simple way to convert the input from image space into", + "bbox": [ + 498, + 857, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "22381", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "high-dimensional feature space. Then we use $K$ stacked omni-scale aggregation groups (OSAG) and one $3 \\times 3$ convolution layer $H_{CONV}$ in a cascade manner to extract deep feature $F_{DF}$ . Such a process can be expressed as", + "bbox": [ + 75, + 90, + 468, + 148 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nX _ {i} = H _ {O S A G _ {i}} \\left(X _ {i - 1}\\right), \\quad i = 1, 2, \\dots , K, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 147, + 157, + 468, + 179 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nX _ {D F} = H _ {C O N V} \\left(X _ {K}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 175, + 336, + 189 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $H_{\\text{OSAG}_i}$ represents the $i$ -th OSAG, $X_1, X_2, \\ldots, X_K$ denote intermediate features. Following [31], we also use a convolutional layer at the end of feature extraction to get better feature aggregation. Finally we reconstruct the HR image $I_{HR}$ by aggregating shallow and deep features as", + "bbox": [ + 76, + 198, + 468, + 270 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nI _ {H R} = H _ {R e c} \\left(X _ {0} + X _ {D F}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 196, + 279, + 468, + 294 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $H_{Rec}(\\cdot)$ denotes the reconstruction module. In detail, PixelShuffle [46] is used to up-sample the fused feature.", + "bbox": [ + 76, + 304, + 467, + 335 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Omni-Scale Aggregation Group (OSAG). As shown in Figure 3, each OSAG contains a local convolution block (LCB), a meso-OSA block, a global-OSA block, and an ESA block [28, 34]. The whole process can be formulated as", + "bbox": [ + 76, + 351, + 468, + 421 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nX _ {r e s} = H _ {\\text {G l o b a l - O S A B} _ {i}} \\left(H _ {\\text {M e s o - O S A B} _ {i}} \\left(H _ {\\text {L C B} _ {i}} \\left(X _ {i - 1}\\right)\\right)\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 120, + 431, + 468, + 446 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nX _ {i} = H _ {E S A _ {i}} \\left(H _ {\\text {C o n v} _ {i}} \\left(X _ {\\text {r e s}} + X _ {i - 1}\\right)\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 449, + 468, + 464 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $X_{i-1}$ and $X_i$ represents the input and output feature of $i$ -th OSAG. After the mapping of convolution layers, we insert a Meso-OSAB for window-based self-attention and a Global-OSAB to enlarge the receptive field for better information aggregation. At the end of OSAG, we reserve the convolutional layer and ESA block following [28, 66].", + "bbox": [ + 76, + 474, + 468, + 565 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In specific, LCB is implemented as a stack of pointwise and depthwise convolutions with a CA module [24] between them to adaptively re-weight channel-wise features. This block aims to aggregate local contextual information as well as to increase the trainability of the network [56]. Two types of OSA blocks (i.e., Meso-OSA block and Global-OSA block) are then followed to obtain interactions from different regions. Based on different window partition strategies, Meso-OSA block seeks inner-block interaction, and Global-OSA blocks aim for global mixing. OSA blocks follow typical Transformers designs with Feedforward network (FFN) and LayerNorm [2], and the only difference is that the origin self-attention operation is replaced with our proposed OSA operator. For FFN, we adopt the GDFN proposed by Restormer [59]. Combining these individuals seamlessly, the designed OSAG enables information propagation between any pair of tokens in the feature map. We use the ESA module proposed in [28, 34] to further refine the fused feature.", + "bbox": [ + 75, + 565, + 468, + 851 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Optimization Objective. Following prior works [31, 32, 53, 67], we train the model by minimizing a standard $L_{1}$", + "bbox": [ + 76, + 869, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "loss between model prediction $\\hat{I}_{HR}$ and HR label $I_{HR}$ as follows:", + "bbox": [ + 498, + 90, + 890, + 119 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\left\\| I _ {H R} - \\hat {I} _ {H R} \\right\\| _ {1}. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 629, + 122, + 890, + 138 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 164, + 630, + 180 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 500, + 189, + 689, + 205 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets and Metrics. Following previous work [31,32, 38,49,66], DIV2K [49] and Flickr2K [49] are used as training datasets. For a fair comparison, we employ two training protocols, i.e., training with DIV2K only and training with DF2K (DIV2K + Flickr2K). Note that the model trained with DF2K is marked with small $\\dagger$ . For testing, we adopt five standard benchmark datasets: Set5 [4], Set14 [60], B100 [39], Urban100 [20] and Manga109 [40]. PSNR and SSIM [54] are adopted to evaluate the SR performance on the Y channel of the transformed YCbCr space.", + "bbox": [ + 498, + 214, + 890, + 364 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. During training, we augment the data with random horizontal flips and 90/270-degree rotations. LR images are generated by bicubic downsampling [63] from HR images. OSAG number is set to 5, and channel number of the whole network is set to 64. The attention head number and window size are set to 4 and 8 for both Meso-OSAB and Global-OSAB. We use Adam [27] optimizer to train the model with a batch size of 64 for $800\\mathrm{K}$ iterations. The initial learning rate is set to $10^{-4}$ and halved for every $200\\mathrm{k}$ iterations. In each training batch, we randomly crop LR patches of size $64\\times 64$ as input. Our method is implemented with PyTorch [45], and all experiments are conducted on one NVIDIA V100 GPU. Note that no other data augmentation (e.g., Mixup [61], RGB channel shuffle) or training skills (e.g., pre-training [30], cosine learning schedule [36]) are employed. It should be pointed out that we maintain the consistency of model parameters in the ablation study by adjusting the channels of $1\\times 1$ convolution.", + "bbox": [ + 498, + 366, + 890, + 638 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Comparison with the SOTA SR methods", + "text_level": 1, + "bbox": [ + 500, + 648, + 844, + 665 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To evaluate the effectiveness of Omni-SR, we compare our model with several advanced lightweight SR methods under a scale factor of 2/3/4. In particular, former works, VDSR [25], CARN [1], IMDN [22], EDSR [32], RFDN [33], MemNet [48], MAFFSRN [42], LatticeNet [38], RLFN [28], ESRT [37] and SwinIR [31] are introduced for comparison.", + "bbox": [ + 498, + 672, + 890, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Quantitative results. In Table 1, the quantitative comparisons of different lightweight methods are presented on five benchmark datasets. With a similar model size, the performance of our Omni-SR surpasses existing methods with a notable margin on all benchmarks. In particular, compared to other transformer architectures with comparable parameters like SwinIR [31] and ESRT [37], the proposed Omni-SR obtains the best performance. The results", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "22382", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/c2e54aa90bde62e82b487e9f121e248131537897edc599a39451d3c33c4058dd.jpg", + "table_caption": [ + "Table 1. Quantitative comparison (PSNR/SSIM) for lightweight image SR with state-of-the-art methods on benchmark datasets. The best and second-best results are marked in red and blue colors, respectively. “†” indicates that model is trained on DF2K." + ], + "table_footnote": [], + "table_body": "
MethodYearsScaleParams(K)Set5 PSNR/ SSIMSet14 PSNR/ SSIMBSD100 PSNR/ SSIMUrban100 PSNR/ SSIMManga109 PSNR/ SSIM
VDSR [25]CVPR1666636.66 / 0.954233.05 / 0.912731.90 / 0.896030.76 / 0.914037.22 / 0.9750
MemNet [48]ICCV1767837.78 / 0.959733.28 / 0.914232.08 / 0.897831.31 / 0.919537.72 / 0.9740
SRMDNF [64]CVPR18151137.79 / 0.96033.32 / 0.91532.05 / 0.898531.33 / 0.920438.07 / 0.9761
CARN [1]ECCV181,59237.76 / 0.959033.52 / 0.916632.09 / 0.897831.92 / 0.925638.36 / 0.9765
IMDN [22]MM1969438.00 / 0.960533.63 / 0.917732.19 / 0.899632.17 / 0.928338.88 / 0.9774
RFDN-L [33]ECCV2062638.08 / 0.960633.67 / 0.919032.18 / 0.899632.24 / 0.929038.95 / 0.9773
MAFFSRN [42]ECCV20×240237.97 / 0.960333.49 / 0.917032.14 / 0.899431.96 / 0.9268-/-
LatticeNet [38]ECCV2075638.15 / 0.961033.78 / 0.919332.25 / 0.900532.43 / 0.9302-/-
RLFN [28]CVPRW2252738.07 / 0.960733.72 / 0.918732.22 / 0.900032.33 / 0.9299-/-
SwinIR [31]ICCVW2187838.14 / 0.961133.86 / 0.920632.31 / 0.901232.76 / 0.934039.12 / 0.9783
Omni-SROurs77238.22 / 0.961333.98 / 0.921032.36 / 0.902033.05 / 0.936339.28 / 0.9784
Omni-SR†Ours77238.29 / 0.961734.27 / 0.923832.41 / 0.902633.30 / 0.938639.53 / 0.9792
VDSR [25]CVPR1666633.66 / 0.921329.77 / 0.831428.82 / 0.797627.14 / 0.827932.01 / 0.9340
MemNet [48]ICCV1767834.09 / 0.924830.00 / 0.835028.96 / 0.800127.56 / 0.837632.51 / 0.9369
EDSR [32]CVPRW171,55534.37 / 0.927030.28 / 0.841729.09 / 0.805228.15 / 0.852733.45 / 0.9439
SRMDNF [64]CVPR181,52834.12 / 0.925430.04 / 0.838228.97 / 0.802527.57 / 0.839833.00 / 0.9403
CARN [1]ECCV181,59234.29 / 0.925530.29 / 0.840729.06 / 0.803428.06 / 0.849333.50 / 0.9440
IMDN [22]MM1970334.36 / 0.927030.32 / 0.841729.09 / 0.804628.17 / 0.851933.61 / 0.9445
RFDN-L [33]ECCV20×363334.47 / 0.928030.35 / 0.842129.11 / 0.805328.32 / 0.854733.78 / 0.9458
MAFFSRN [42]ECCV2080734.45 / 0.927730.40 / 0.843229.13 / 0.806128.26 / 0.8552-/-
LatticeNet [38]ECCV2076534.53 / 0.928130.39 / 0.842429.15 / 0.805928.33 / 0.8538-/-
ESRT [37]CVPRW2277034.42 / 0.926830.43 / 0.843329.15 / 0.806328.46 / 0.857433.95 / 0.9455
SwinIR [31]ICCVW2188634.62 / 0.928930.54 / 0.846329.20 / 0.808228.66 / 0.862433.98 / 0.9478
Omni-SROurs78034.70 / 0.929430.57 / 0.846929.28 / 0.809428.84 / 0.865634.22 / 0.9487
Omni-SR†Ours78034.77 / 0.930430.70 / 0.848929.33 / 0.811129.12 / 0.871234.64 / 0.9507
VDSR [25]CVPR1666631.35 / 0.883828.01 / 0.767427.29 / 0.725125.18 / 0.752428.83 / 0.8870
MemNet [48]ICCV1767831.74 / 0.889328.26 / 0.772327.40 / 0.728125.50 / 0.763029.42 / 0.8942
EDSR [32]CVPRW171,51832.09 / 0.893828.58 / 0.781327.57 / 0.735726.04 / 0.784930.35 / 0.9067
SRMDNF [64]CVPR181,55231.96 / 0.892528.35 / 0.778727.49 / 0.733725.68 / 0.773130.09 / 0.9024
CARN [1]ECCV181,59232.13 / 0.893728.60 / 0.780627.58 / 0.734926.07 / 0.783730.47 / 0.9084
IMDN [22]MM1971532.21 / 0.894828.58 / 0.781127.56 / 0.735326.04 / 0.783830.45 / 0.9075
RFDN-L [33]ECCV20×464332.28 / 0.895728.61 / 0.781827.58 / 0.736326.20 / 0.788330.61 / 0.9096
MAFFSRN [42]ECCV2083032.20 / 0.895326.62 / 0.782227.59 / 0.737026.16 / 0.7887-/-
LatticeNet [38]ECCV2077732.30 / 0.896228.68 / 0.783027.62 / 0.736726.25 / 0.7873-/-
ESRT [37]CVPRW2254332.24 / 0.895228.62 / 0.781327.60 / 0.736426.17 / 0.7877-/-
SwinIR [31]ICCVW2175132.19 / 0.894728.69 / 0.783327.69 / 0.737926.39 / 0.796230.75 / 0.9100
Omni-SROurs89732.44 / 0.897628.77 / 0.785827.69 / 0.740626.47 / 0.798030.92 / 0.9151
Omni-SR†Ours79232.49 / 0.898828.78 / 0.785927.71 / 0.741526.64 / 0.801831.02 / 0.9151
Omni-SR†Ours79232.57 / 0.899328.95 / 0.789827.81 / 0.743926.95 / 0.810531.50 / 0.9192
", + "bbox": [ + 89, + 118, + 879, + 542 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/893f81de995407df99654d0b38795a6f908c601bfb716c4c33644406bacbe803.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 78, + 556, + 269, + 688 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ffe582007a99f84656e3b6d885cdc2a6f8a9cf27362c2e7952528c61d4746ecc.jpg", + "image_caption": [ + "(b)", + "Figure 5. (a) PSNR improvement of Omni-SR variants with different OSAG number (K) over the smallest Omni-SR model $(\\mathrm{K} = 1)$ for $4\\times$ SR. (b) The number of model parameters vs. PSNR of different lightweight methods on Urban100 dataset for $4\\times$ SR." + ], + "image_footnote": [], + "bbox": [ + 272, + 556, + 464, + 693 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "showcase that the omni-axis (i.e., spatial + channel) interaction introduced by OSA can effectively boost the model's contextual aggregation capabilities, which promises superior SR performance. Coupling with large training dataset DF2K, the performance can be further improved, especially on Urban100. We suppose that such a phenomenon can be ascribed to the images in Urban100 have many similar patches, and the long-term relationship introduced by", + "bbox": [ + 75, + 780, + 468, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "OSAG can bring great benefits for detail restoration. More importantly, with similar parameters, our model reduces $28\\%$ of computational complexity (Omni-SR: 36G FLOPs vs. SwinIR: 50G FLOPs @1280×720), showing its effectiveness and efficiency.", + "bbox": [ + 498, + 559, + 890, + 633 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Visual comparison. In Figure 6, we also provide a visual comparison of different lightweight SR methods at $\\times 4$ scale. We can observe that the HR images constructed by Omni-SR contain more fine-grained details, while other methods generate blurred edges or artifacts in complicated areas. For instance, in the $1st$ row, our model is able to restore the detailed texture of the wall pleasantly, which all other methods fail to restore. Visual results also validate the effectiveness of the proposed OSA paradigm, which can perform omni-axis pixel-wise interaction modeling, thus obtaining a more powerful reconstruction capability.", + "bbox": [ + 496, + 638, + 892, + 805 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Trade-off between Model Size and Performance. In experiments, we set the number of OSAG as 5 to make the model size around $800\\mathrm{K}$ for a fair comparison with other methods. We also explore model performance with smaller parameter sizes by reducing OSAG number K. As shown in Figure 5(a), compared to the smallest model variant with", + "bbox": [ + 496, + 810, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "22383", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/94d802cd6418ce37365ce9e96117e6ce93556852e979423c7d4d4fb7695d8b22.jpg", + "image_caption": [ + "Urban100: img_012.png" + ], + "image_footnote": [], + "bbox": [ + 99, + 88, + 308, + 199 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/494a73f50393065708fb86e2beeeeef07544713ef02e318bd799cdb4823f802c.jpg", + "image_caption": [ + "PSNR Reference" + ], + "image_footnote": [], + "bbox": [ + 310, + 88, + 421, + 132 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/401fb73e79fac28b9e0da2dd38d4647fc73da9a39f897c70b0fdd43f2fa8f301.jpg", + "image_caption": [ + "20.65 dB Bicubic" + ], + "image_footnote": [], + "bbox": [ + 424, + 88, + 535, + 132 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/883f3e8878b3dc8f6e6c72dd84f3a645091e2d54afd68c28b67c0d9ed14f3e28.jpg", + "image_caption": [ + "21.15 dB SRCNN [13]" + ], + "image_footnote": [], + "bbox": [ + 537, + 88, + 648, + 132 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1e5a0907028261a23d10f724de5d8d804a90b7dacce7856eb02735db56b204df.jpg", + "image_caption": [ + "21.22 dB FSRCNN [29]" + ], + "image_footnote": [], + "bbox": [ + 651, + 88, + 761, + 132 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f89c955c7a4e3a57046090ed64ecaaa0e21b351a1926ff23b69312fe596e0792.jpg", + "image_caption": [ + "21.99 dB CARN [1]" + ], + "image_footnote": [], + "bbox": [ + 763, + 88, + 875, + 132 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/77836fefd440f40b36f59a49c94431d7ba2877842db8df27ee2c56c043a021fb.jpg", + "image_caption": [ + "Urban100: img_024.png" + ], + "image_footnote": [], + "bbox": [ + 99, + 223, + 307, + 330 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ab6c47781207dc279b26c766c0b3d0a8968f71e8642f553ef7451110331dd877.jpg", + "image_caption": [ + "21.95 dB IMDN [22]" + ], + "image_footnote": [], + "bbox": [ + 310, + 155, + 421, + 199 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/02a5fb9235cbec0479716589693eedc88d20f1ca68ad0b747f0d30bf586135d2.jpg", + "image_caption": [ + "21.95 dB RLFN [28]" + ], + "image_footnote": [], + "bbox": [ + 424, + 155, + 535, + 199 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/89cf7d445d853fc9f318cfb4b06feb57bff736ef54700ddc08adc1ded5c7cd1e.jpg", + "image_caption": [ + "22.00 dB LatticeNet [38]" + ], + "image_footnote": [], + "bbox": [ + 537, + 155, + 648, + 199 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dafc61eeee66ffca2f3a71457045dd04837b4e46ef0c3ffedab0f4bbc58020fa.jpg", + "image_caption": [ + "22.34 dB SwinIR [31]" + ], + "image_footnote": [], + "bbox": [ + 651, + 155, + 761, + 199 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4ec3907b3e3aa4f900e3e98c6906c9e62be75f06cb6569e71ef5d6bfb2cf891c.jpg", + "image_caption": [ + "24.16 dB Omni-SR (Ours)" + ], + "image_footnote": [], + "bbox": [ + 764, + 155, + 875, + 199 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6dcc1cf12baab7f8aa1c5bf8c289d536f1246b2712c317b68b08601e653fe82f.jpg", + "image_caption": [ + "Set14: img_barbara.png", + "Figure 6. Visual comparison for $\\times 4$ SR methods. The patches for comparison are marked with red boxes. (Best viewed by zooming.)" + ], + "image_footnote": [], + "bbox": [ + 99, + 357, + 307, + 465 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/269c2cee48363e525646d0ac24f793ad38cf655e6d1b9a76a24d3fec01f95687.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 310, + 199, + 421, + 220 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/162f7a8fc7d6f7d23f09a1207931df688e23ed0f0aeba914e85f887459ec038f.jpg", + "image_caption": [ + "PSNR Reference" + ], + "image_footnote": [], + "bbox": [ + 310, + 220, + 421, + 266 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0e0f08caf93a4c93c799f7d427d119f3a1b87b2554d2fb1605237037c8bd7668.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 424, + 199, + 535, + 220 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/45ec03ffa2e49f379143ca7614d60053f9e0f811fd3c203d016790238e600d00.jpg", + "image_caption": [ + "16.90 dB Bicubic" + ], + "image_footnote": [], + "bbox": [ + 424, + 220, + 535, + 266 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d4a4ee7a4c389733d89e2a71ee28d9d7a9dea65d1b66edec6dec075351bf135a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 537, + 199, + 648, + 220 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/efc98eb37785400a9c767986a9120da7db0ea56c4c60f8fa2a9822dd34381cf7.jpg", + "image_caption": [ + "17.54 dB SRCNN [13]" + ], + "image_footnote": [], + "bbox": [ + 537, + 220, + 650, + 266 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/75746093afebd3051e4e2320e109478ede0540aff4821e59af273b6ab0c491c1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 651, + 199, + 761, + 220 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/796ee10c29fb3f1c5f8ea85ef79a72b6f618b87397c99523b7f1a03acde2a15e.jpg", + "image_caption": [ + "17.71 dB FSRCNN [29]" + ], + "image_footnote": [], + "bbox": [ + 651, + 220, + 761, + 266 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3390d5d4392ecc3a6302e05f7f15d37bce49a6bb4d8189486a7f587fdbfd74b0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 764, + 199, + 875, + 220 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/24294bddef8c107df51af9be3a5763d13c14f3971dde0d736128786dd077502f.jpg", + "image_caption": [ + "18.78 dB CARN [1]" + ], + "image_footnote": [], + "bbox": [ + 764, + 220, + 875, + 266 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/507e9e529c914fcca45fed3d861be593ad8c4931a1d6fff2061d2040b4ae10a3.jpg", + "image_caption": [ + "18.77 dB \n[MDN [22]]" + ], + "image_footnote": [], + "bbox": [ + 310, + 287, + 421, + 332 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/05c429fd2742ea720938c2b91ed02202c26231eee65aeda4f572292e5ce0463c.jpg", + "image_caption": [ + "19.23 dB RLFN [28]" + ], + "image_footnote": [], + "bbox": [ + 424, + 287, + 535, + 332 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c79a6db4d56adf3eab4c6412ae0c805ff9261eda98ceb04dc7f85b4f9fede4ba.jpg", + "image_caption": [ + "18.91 dB LatticeNet [38]" + ], + "image_footnote": [], + "bbox": [ + 537, + 287, + 650, + 332 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/89b2d08e02920eeb1a440525897f9db3363759839d30c9c701ac30a394f74c1a.jpg", + "image_caption": [ + "19.07 dB SwinIR [31]" + ], + "image_footnote": [], + "bbox": [ + 651, + 287, + 761, + 332 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bb801bd68add06260d770f876872149c032dfc7e783f367e180b9553bded8c1f.jpg", + "image_caption": [ + "21.18 dB Omni-SR (Ours)" + ], + "image_footnote": [], + "bbox": [ + 764, + 287, + 875, + 332 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3624e33137ec1266b291073d03192799a127f08adc46b00c2d2a0dae18128c51.jpg", + "image_caption": [ + "PSNR Reference" + ], + "image_footnote": [], + "bbox": [ + 310, + 354, + 421, + 398 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c3b48d85deeed63cf49264f906f84a21e402e04e2f791f4e32c43fdeaadfec58.jpg", + "image_caption": [ + "23.58 dB Bicubic" + ], + "image_footnote": [], + "bbox": [ + 424, + 354, + 535, + 398 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e7f906c0018864b466385ffdaa4d9b737212722f7535688b21c97690f0b078ad.jpg", + "image_caption": [ + "24.04 dB SRCNN [13]" + ], + "image_footnote": [], + "bbox": [ + 537, + 354, + 648, + 398 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d1b7d92000dba7f4252448f42df8ed241decc120487dc1578d8161f5ea35d67b.jpg", + "image_caption": [ + "24.11 dB FSRCNN [29]" + ], + "image_footnote": [], + "bbox": [ + 651, + 354, + 761, + 398 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b9987c31f09d6af8941ef0f02b263bd6e2e398ccb41239bd0a34ac1182fb7580.jpg", + "image_caption": [ + "24.48 dB CARN [1]" + ], + "image_footnote": [], + "bbox": [ + 764, + 354, + 875, + 398 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/758acc31a79474f0bb5a07acfdaba0efa7ad19c835bae55ae7a5d7f31ee328b7.jpg", + "image_caption": [ + "24.62 dB IMDN [22]" + ], + "image_footnote": [], + "bbox": [ + 310, + 421, + 421, + 465 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/daf5882753d09415d704f856b2b055023f6a358b147b5d07009e844138c0e774.jpg", + "image_caption": [ + "24.43 dB RLFN [28]" + ], + "image_footnote": [], + "bbox": [ + 424, + 421, + 535, + 465 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dc33abd478c695a711c0927226082f8e01cc8f485c764b97df6d826b745e2a89.jpg", + "image_caption": [ + "24.64 dB LatticeNet [38]" + ], + "image_footnote": [], + "bbox": [ + 537, + 421, + 648, + 465 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3d47daf6c3df1925794bd00e517e13d1381e1b9f2c13b83091fbbb776e4c2079.jpg", + "image_caption": [ + "24.45 dB SwinIR [31]" + ], + "image_footnote": [], + "bbox": [ + 651, + 421, + 761, + 465 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2f5a163e9a687d1c82a3b64614fbf786df9a85df92794e741b2c13cfc325abf4.jpg", + "image_caption": [ + "26.19 dB Omni-SR (Ours)" + ], + "image_footnote": [], + "bbox": [ + 764, + 421, + 875, + 465 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "$\\mathrm{K} = 1$ increasing the number of OSAG leads to stable performance improvements. In Figure 5(b), we present PSNR vs. parameters of different methods. It can be found that Omni-SR achieves the best results under various settings, showing its effectiveness and scalability.", + "bbox": [ + 75, + 521, + 468, + 595 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Analysis of Omni Self-Attention", + "text_level": 1, + "bbox": [ + 76, + 604, + 357, + 619 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we illustrate the optimization features of OSA and further uncover its underlying mechanism. Self-attention is a low-bias operator, which makes its optimization difficult and requires more training epochs. For this, we introduce the additional channel-wise interaction to alleviate it. In Figure 7(a), we show the loss curves of different self-attention paradigms on the DIV2K training set, including spatial self-attention, channel self-attention, and the proposed omni self-attention. We can see that our OSA presents a obviously superior convergence speed. More importantly, the performance at the final epoch is also significantly ahead of them. The above phenomenon clearly shows that our OSA has superior good optimization characteristics. Further, we delve into why channel-wise interactions lead to these improvements. We calculate the normalized entropies [52] of the hidden layer features of the network composed of the above three computational primitives. We illustrate the entropy results in Figure 7(c).", + "bbox": [ + 75, + 628, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in the figure, in all outgoing layers, our OSA-encoded features show higher entropy, indicating that our OSA encodes richer information. More information may come from various scales, and this information can help the operators to reconstruct the exact details faster. We speculate that this is the potential reason why our OSA shows better optimization properties. In addition, following previous works [8, 16], we also resort to LAM analysis. DI [16] metric can measure the furthest interaction distance of the model. From Figure 8 we can observe that Omni-SR generally has the highest max diffusion index than other methods, showing that our OSA paradigm can effectively capture long-range interactions.", + "bbox": [ + 496, + 521, + 890, + 717 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 500, + 726, + 651, + 742 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of Omni Self-Attention. The core idea of our framework is to extend the vanilla self-attention with a channel-wise relationship to build omni-axis pixel-wise interaction. Based on Omni-SR framework, we design several variant models, and their SR results are reported in Table 2. We first simply remove the channel-wise component to form a spatial-only variant $(\\mathrm{Omni - SR_{sp}})$ , and its performance is degraded by $0.13\\mathrm{dB}$ compared to the full model. Such a significant degradation justifies the importance of channel interactions. Note that Omni- $\\mathrm{SR_{sp}}$ still", + "bbox": [ + 496, + 750, + 890, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "22384", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1b026b04daf7a008b04feb302dd3557f9f9b26d00703fc2ce40473e53e2a01b4.jpg", + "image_caption": [ + "(a) Ablation studies on OSAG" + ], + "image_footnote": [], + "bbox": [ + 138, + 88, + 423, + 199 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4bdb033e383e968248d934e8e127f02ded6bc64bdeda49ea475346c3e22f556d.jpg", + "image_caption": [ + "(b) Training loss on DIV2K" + ], + "image_footnote": [], + "bbox": [ + 436, + 87, + 625, + 198 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3fa37dfb9888c6ddc4427f7895be2ebc9b9110e8f7bde7677047a0794d12b00c.jpg", + "image_caption": [ + "(c) Feature correlation" + ], + "image_footnote": [], + "bbox": [ + 640, + 89, + 823, + 198 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e385dad27e053ce2fce368849b68cc7dadc75fce15fae1142ee2c956409b8a27.jpg", + "image_caption": [ + "Figure 7. (a) Ablation studies on different components in OSAG. (b) Training loss of different attention paradigms. (c) Feature correlation analysis of different attention paradigms. All the results are reported on DIV2K dataset for $4 \\times$ SR.", + "Figure 8. The distribution of DI values for different methods on Urban100. Top-left points with show a narrow area of interest, and right-bottom points show a large area of interest." + ], + "image_footnote": [], + "bbox": [ + 102, + 270, + 444, + 391 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "outperforms SwinIR by $0.04\\mathrm{dB}@$ Urban $100\\times 4$ , which benefits from the global interaction introduced by grid window partition. Similarly, we remove the spatial self-attention component to derive a channel self-attention variant, Omni $\\mathrm{SR}_{\\mathrm{ca}}$ , and such a modification also leads to undesirable performance degradation. Besides, we use the most widely adopted channel and spatial attention configurations (i.e., SE [19] and CBAM [55]) to act as alternative operators for channel and spatial aggregation. Both substitutions (Omni $\\mathrm{SR}_{\\mathrm{SE}}$ , Omni- $\\mathrm{SR}_{\\mathrm{CBAM}}$ ) hurt PNSR performance compared to the full model. The above results show that the specific interaction paradigm (e.g., scalar-based, covariance-based) is equally important, and our channel interaction based on the covariance matrix shows great advantages.", + "bbox": [ + 75, + 460, + 468, + 672 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of Omni-Scale Aggregation Group. In Omni-SR, we propose a local-meso-global interaction scheme (i.e., OSAG) to pursue progressive feature aggregation. To investigate its effectiveness, we design three different kinds of interaction schemes based on Omni-SR framework: Separated scheme, Mixed scheme, and our fully designed Omni scheme (i.e., our proposed OSAG), and the ablation study results are shown in Figure 7(a). In the figure, we employ different words (e.g., \"Local\", \"Meso+Global\") to represent specific schemes, e.g., \"Local\" denotes using Local-Conv block to replace Meso-OSA and Global-OSA; \"Local+Global\" represents replace original cascaded Meso-OSA and Global-OSA with cascaded Local-Conv and Global-OSA. We can observe that single interaction schemes (e.g., \"Local\") perform the worst. Interestingly,", + "bbox": [ + 75, + 674, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/8b9d5ef9b1c65f2a01d3b4733d1c81e1ef80bd538d4264f6a6eb46b2ffcc694e.jpg", + "table_caption": [ + "Table 2. Ablation studies of omni self-attention on Urban100. $\\mathrm{Omni - SR}_{(*)}$ denotes different modifications. 'SA' and 'S-SA' denote spatial gate and spatial self-attention. 'CA' and 'C-SA' denote channel gate and channel self-attention. We maintain the consistency of model parameters by adjusting the channels of $1\\times 1Conv$" + ], + "table_footnote": [], + "table_body": "
ModelSAS-SACAC-SAFLOPs×2×3×4
Omni-SRsp×✓××33G32.8828.7226.51
Omni-SRSE×✓✓×34G32.8328.7126.50
Omni-SRCBAM✓×✓✓34G32.9228.7626.53
Omni-SRca×××✓33G32.6528.6026.45
Omni-SRfull×✓×✓36G33.0528.8426.64
", + "bbox": [ + 503, + 344, + 890, + 431 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the \"Global\" scheme is inferior to the \"Meso\" one due to its poor optimization properties of global self-attention [3, 35, 50]. Once two interaction operators are combined, the performance improves steadily. Among them, \"Meso+ Global\" setting achieves the second-best performance. Furthermore, combining all three interaction schemes together, we obtain the best performing scheme, i.e., \"Omni\". From the above experiments, we can infer that obvious performance gains can be obtained by introducing various-scale interactions, which also illustrates the feasibility and effectiveness of our proposed OSAG.", + "bbox": [ + 496, + 446, + 890, + 613 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 627, + 617, + 642 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we propose Omni-SR, a lightweight framework for image SR. We propose the Omni Self-attention paradigm for simultaneous spatial and channel interactions, mining all the potential correlations across omni-axis. Furthermore, we propose an omni-scale aggregation scheme to effectively enlarge the receptive fields with low computational complexity, which encodes contextual relations in a progressively hierarchical manner. Extensive experiments on public benchmark datasets and comprehensive analytical studies validate its prominent SR performance.", + "bbox": [ + 496, + 652, + 890, + 803 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgement", + "text_level": 1, + "bbox": [ + 500, + 815, + 679, + 832 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported by National Science Foundation of China (U20B2072, 61976137). This work was also partly supported by SJTU Medical Engineering Cross Research Grant YG2021ZD18.", + "bbox": [ + 496, + 839, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "22385", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Namhyuk Ahn, Byungkon Kang, and Kyung-Ah Sohn. Fast, accurate, and lightweight super-resolution with cascading residual network. In ECCV, 2018. 2, 5, 6, 7", + "[2] Lei Jimmy Ba, Jamie Ryan Kiros, and Geoffrey E. Hinton. Layer normalization. CoRR, abs/1607.06450, 2016. 5", + "[3] Irwan Bello, Barret Zoph, Quoc Le, Ashish Vaswani, and Jonathon Shlens. Attention augmented convolutional networks. In ICCV, 2019. 8", + "[4] Marco Bevilacqua, Aline Roumy, Christine Guillemot, and Marie-Line Alberi-Morel. Low-complexity single-image super-resolution based on nonnegative neighbor embedding. In BMVC, 2012. 5", + "[5] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In CVPR, 2021. 1, 2, 3", + "[6] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. arXiv:1606.00915, 2016. 1", + "[7] Xuanhong Chen, Hang Wang, and Bingbing Ni. X-volution: On the unification of convolution and self-attention. arXiv preprint arXiv:2106.02253, 2021. 2", + "[8] Xiangyu Chen, Xintao Wang, Jiantao Zhou, and Chao Dong. Activating more pixels in image super-resolution transformer. CoRR, abs/2205.04437, 2022. 2, 4, 7", + "[9] Yinpeng Chen, Xiyang Dai, Dongdong Chen, Mengchen Liu, Xiaoyi Dong, Lu Yuan, and Zicheng Liu. Mobileformer: Bridging mobilenet and transformer. In CVPR, 2022. 2", + "[10] Yinpeng Chen, Xiyang Dai, Dongdong Chen, Mengchen Liu, Xiaoyi Dong, Lu Yuan, and Zicheng Liu. Simple baselines for image restoration. In ECCV, 2022. 2, 3", + "[11] Xiangxiang Chu, Bo Zhang, Hailong Ma, Ruijun Xu, and Qingyuan Li. Fast, accurate and lightweight super-resolution with neural architecture search. In ICPR, 2020. 2", + "[12] Xiaohan Ding, Xiangyu Zhang, Jungong Han, and Guiguang Ding. Scaling up your kernels to 31x31: Revisiting large kernel design in cnns. In CVPR, 2022. 1", + "[13] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Image super-resolution using deep convolutional networks. CoRR, abs/1501.00092, 2015. 2, 7", + "[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 1, 2", + "[15] Qinquan Gao, Yan Zhao, Gen Li, and Tong Tong. Image super-resolution using knowledge distillation. In ACCV, 2018. 2", + "[16] Jinjin Gu and Chao Dong. Interpreting super-resolution networks with local attribution maps. In CVPR, 2021. 7", + "[17] Geoffrey E. Hinton, Oriol Vinyals, and Jeffrey Dean. Distilling the knowledge in a neural network. CoRR, abs/1503.02531, 2015. 2" + ], + "bbox": [ + 78, + 114, + 468, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[18] Andrew G Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andretto, and Hartwig Adam. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv:1704.04861, 2017. 4", + "[19] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In CVPR, 2018. 2, 3, 4, 8", + "[20] Jia-Bin Huang, Abhishek Singh, and Narendra Ahuja. Single image super-resolution from transformed self-exemplars. In CVPR, 2015. 5", + "[21] Tao Huang, Lang Huang, Shan You, Fei Wang, Chen Qian, and Chang Xu. Lightvit: Towards light-weight convolution-free vision transformers. CoRR, abs/2207.05557, 2022. 2", + "[22] Zheng Hui, Xinbo Gao, Yunchu Yang, and Xiumei Wang. Lightweight image super-resolution with information multistillation network. In ACMMM, 2019. 2, 5, 6, 7", + "[23] Zheng Hui, Xiumei Wang, and Xinbo Gao. Fast and accurate single image super-resolution via information distillation network. In CVPR, 2018. 2", + "[24] Forrest N Iandola, Song Han, Matthew W Moskewicz, Khalid Ashraf, William J Dally, and Kurt Keutzer. SqueezeNet: Alexnet-level accuracy with 50x fewer parameters and; 0.5 mb model size. arXiv:1602.07360, 2016. 5", + "[25] Jiwon Kim, Jung Kwon Lee, and Kyoung Mu Lee. Accurate image super-resolution using very deep convolutional networks. In CVPR, 2016. 2, 5, 6", + "[26] Jiwon Kim, Jung Kwon Lee, and Kyoung Mu Lee. Deeplyrecursive convolutional network for image super-resolution. In CVPR, 2016. 2", + "[27] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 5", + "[28] Fangyuan Kong, Mingxi Li, Songwei Liu, Ding Liu, Jingwen He, Yang Bai, Fangmin Chen, and Lean Fu. Residual local feature network for efficient super-resolution. In CVPR Workshops, 2022. 2, 5, 6, 7", + "[29] Wei-Sheng Lai, Jia-Bin Huang, Narendra Ahuja, and Ming-Hsuan Yang. Fast and accurate image super-resolution with deep laplacian pyramid networks. TPAMI, 41(11):2599-2613, 2018. 7", + "[30] Wenbo Li, Xin Lu, Shengju Qian, Jiangbo Lu, Xiangyu Zhang, and Jiaya Jia. On efficient transformer and image pre-training for low-level vision. arXiv preprint arXiv:2112.10175, 2021. 2, 5", + "[31] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In ICCV, 2021. 1, 2, 4, 5, 6, 7", + "[32] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In CVPRW, 2017. 5, 6", + "[33] Jie Liu, Jie Tang, and Gangshan Wu. Residual feature distillation network for lightweight image super-resolution. In ECCV, 2020. 5, 6", + "[34] Jie Liu, Wenjie Zhang, Yuting Tang, Jie Tang, and Gangshan Wu. Residual feature aggregation network for image superresolution. In CVPR, 2020. 5", + "[35] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer:" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "22386", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Hierarchical vision transformer using shifted windows. In ICCV, 2021. 2, 4, 8", + "[36] Ilya Loshchilov and Frank Hutter. SGDR: stochastic gradient descent with warm restarts. In *ICLR*, 2017. 5", + "[37] Zhisheng Lu, Juncheng Li, Hong Liu, Chaoyan Huang, Linlin Zhang, and Tieyong Zeng. Transformer for single image super-resolution. In CVPR Workshops, 2022. 2, 3, 4, 5, 6", + "[38] Xiaotong Luo, Yuan Xie, Yulun Zhang, Yanyun Qu, Cuihua Li, and Yun Fu. Latticenet: Towards lightweight image super-resolution with lattice block. In ECCV, 2020. 2, 5, 6, 7", + "[39] David Martin, Charless Fowlkes, Doron Tal, and Jitendra Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In ICCV. IEEE, 2001. 5", + "[40] Yusuke Matsui, Kota Ito, Yuji Aramaki, Azuma Fujimoto, Toru Ogawa, Toshihiko Yamasaki, and Kiyoharu Aizawa. Sketch-based manga retrieval using manga109 dataset. Multimedia Tools and Applications, 76(20):21811-21838, 2017. 5", + "[41] Sachin Mehta and Mohammad Rastegari. Mobilevit: Lightweight, general-purpose, and mobile-friendly vision transformer. CoRR, abs/2110.02178, 2021. 2", + "[42] Abdul Muqueet, Jiwon Hwang, Subin Yang, JungHeum Kang, Yongwoo Kim, and Sung-Ho Bae. Multi-attention based ultra lightweight image super-resolution. In ECCV, 2020. 2, 5, 6", + "[43] Junting Pan, Adrian Bulat, Fuwen Tan, Xiatian Zhu, Lukasz Dudziak, Hongsheng Li, Georgios Tzimiropoulos, and Brais Martínez. Edgevits: Competing light-weight cnns on mobile devices with vision transformers. In ECCV, 2022. 2", + "[44] Jongchan Park, Sanghyun Woo, Joon-Young Lee, and In So Kweon. Bam: Bottleneck attention module. arXiv:1807.06514, 2018. 3, 4", + "[45] Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. Automatic differentiation in pytorch. In NeurIPS Workshop, 2017. 5", + "[46] Wenzhe Shi, Jose Caballero, Ferenc Huszár, Johannes Totz, Andrew P Aitken, Rob Bishop, Daniel Rueckert, and Zehan Wang. Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In CVPR, 2016. 5", + "[47] Ying Tai, Jian Yang, and Xiaoming Liu. Image superresolution via deep recursive residual network. In CVPR, 2017. 2", + "[48] Ying Tai, Jian Yang, Xiaoming Liu, and Chunyan Xu. Memnet: A persistent memory network for image restoration. In ICCV, 2017. 2, 5, 6", + "[49] Radu Timofte, Eirikur Agustsson, Luc Van Gool, Ming-Hsuan Yang, and Lei Zhang. Ntire 2017 challenge on single image super-resolution: Methods and results. In CVPR workshops, 2017. 5", + "[50] Ashish Vaswani, Prajit Ramachandran, Aravind Srinivas, Niki Parmar, Blake Hechtman, and Jonathon Shlens. Scaling local self-attention for parameter efficient visual backbones. In CVPR, 2021. 2, 8" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[51] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NIPS, 2017. 1, 2, 3", + "[52] Pei Wang, Yijun Li, and Nuno Vasconcelos. Rethinking and improving the robustness of image style transfer. In CVPR, 2021. 7", + "[53] Xiaohang Wang, Xuanhong Chen, Bingbing Ni, Zhengyan Tong, and Hang Wang. Learning continuous depth representation via geometric spatial aggregator. In AAAI, 2023. 5", + "[54] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 5", + "[55] Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon. Cbam: Convolutional block attention module. In ECCV, 2018. 3, 4, 8", + "[56] Tete Xiao, Mannat Singh, Eric Mintun, Trevor Darrell, Piotr Dólar, and Ross B. Girshick. Early convolutions help transformers see better. In NeurIPS, 2021. 5", + "[57] Chenglin Yang, Yilin Wang, Jianming Zhang, He Zhang, Zijun Wei, Zhe Lin, and Alan L. Yuille. Lite vision transformer with enhanced self-attention. In CVPR, 2022. 2", + "[58] Jianwei Yang, Chunyuan Li, Xiyang Dai, and Jianfeng Gao. Focal modulation networks, 2022. 3", + "[59] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In CVPR, 2022. 1, 3, 5", + "[60] Roman Zeyde, Michael Elad, and Matan Protter. On single image scale-up using sparse-representations. In International conference on curves and surfaces, pages 711-730. Springer, 2010. 5", + "[61] Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In ICLR, 2018. 5", + "[62] Haokui Zhang, Wenze Hu, and Xiaoyu Wang. Parc-net: Position aware circular convolution with merits from convnets and transformer. In ECCV, 2022. 2", + "[63] Kai Zhang, Wangmeng Zuo, Shuhang Gu, and Lei Zhang. Learning deep cnn denoiser prior for image restoration. In CVPR, 2017. 5", + "[64] Kai Zhang, Wangmeng Zuo, and Lei Zhang. Learning a single convolutional super-resolution network for multiple degradations. In CVPR, 2018. 6", + "[65] Yiman Zhang, Hanting Chen, Xinghao Chen, Yiping Deng, Chunjing Xu, and Yunhe Wang. Data-free knowledge distillation for image super-resolution. In CVPR, 2021. 2", + "[66] Yulun Zhang, Kunpeng Li, Kai Li, Lichen Wang, Bineng Zhong, and Yun Fu. Image super-resolution using very deep residual channel attention networks. In ECCV, 2018. 1, 2, 5", + "[67] Yulun Zhang, Yapeng Tian, Yu Kong, Bineng Zhong, and Yun Fu. Residual dense network for image super-resolution. In CVPR, 2018. 5" + ], + "bbox": [ + 501, + 92, + 890, + 844 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "22387", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Omni Aggregation Networks for Lightweight Image Super-Resolution/39175e81-b86e-49a9-967e-e590c3dbdaa0_model.json b/2023/Omni Aggregation Networks for Lightweight Image Super-Resolution/39175e81-b86e-49a9-967e-e590c3dbdaa0_model.json new file mode 100644 index 0000000000000000000000000000000000000000..0224a423c67f95fd2b331f47f35bea4742b8f25c --- /dev/null +++ b/2023/Omni Aggregation Networks for Lightweight Image Super-Resolution/39175e81-b86e-49a9-967e-e590c3dbdaa0_model.json @@ -0,0 +1,3091 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.134, + 0.131, + 0.836, + 0.152 + ], + "angle": 0, + "content": "Omni Aggregation Networks for Lightweight Image Super-Resolution" + }, + { + "type": "text", + "bbox": [ + 0.167, + 0.18, + 0.796, + 0.199 + ], + "angle": 0, + "content": "Hang Wang\\(^{2*}\\) Xuanhong Chen\\(^{1*}\\) Bingbing Ni\\(^{1,2\\dagger}\\) Yutian Liu\\(^{1}\\) Jinfan Liu\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.212, + 0.199, + 0.76, + 0.217 + ], + "angle": 0, + "content": "\\(^{1}\\)Shanghai Jiao Tong University, Shanghai 200240, China \\(^{2}\\) Huawei" + }, + { + "type": "text", + "bbox": [ + 0.142, + 0.217, + 0.823, + 0.235 + ], + "angle": 0, + "content": "francis970625@gmail.com {chen19910528,nibingbing,stau7001,ljflnjz}@sjtu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.313, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.301, + 0.474, + 0.664 + ], + "angle": 0, + "content": "While lightweight ViT framework has made tremendous progress in image super-resolution, its uni-dimensional self-attention modeling, as well as homogeneous aggregation scheme, limit its effective receptive field (ERF) to include more comprehensive interactions from both spatial and channel dimensions. To tackle these drawbacks, this work proposes two enhanced components under a new Omni-SR architecture. First, an Omni Self-Attention (OSA) block is proposed based on dense interaction principle, which can simultaneously model pixel-interaction from both spatial and channel dimensions, mining the potential correlations across omni-axis (i.e., spatial and channel). Coupling with mainstream window partitioning strategies, OSA can achieve superior performance with compelling computational budgets. Second, a multi-scale interaction scheme is proposed to mitigate sub-optimal ERF (i.e., premature saturation) in shallow models, which facilitates local propagation and meso-/global-scale interactions, rendering an omni-scale aggregation building block. Extensive experiments demonstrate that Omni-SR achieves record-high performance on lightweight super-resolution benchmarks (e.g., 26.95dB@Urban100 ×4 with only 792K parameters). Our code is available at https://github.com/Francis0625/Omni-SR." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.691, + 0.21, + 0.707 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.716, + 0.47, + 0.868 + ], + "angle": 0, + "content": "Image super-resolution (SR) is a long-standing low-level problem that aims to recover high-resolution (HR) images from degraded low-resolution (LR) inputs. Recently, vision transformer [14, 51] based (i.e., ViT-based) SR frameworks [5, 31] have emerged, showing significant performance gains compared to previously dominant Convolutional Neural Networks (CNNs) [66]. However, most attempts [31] are devoted to improving the large-scale ViT-based models, while the development of lightweight ViTs (typically, less than 1M parameters) remains fraught with" + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.277, + 0.625, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.539, + 0.382, + 0.593, + 0.392 + ], + "angle": 0, + "content": "Spatial SA" + }, + { + "type": "image", + "bbox": [ + 0.638, + 0.277, + 0.752, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.665, + 0.381, + 0.724, + 0.392 + ], + "angle": 0, + "content": "Channel SA" + }, + { + "type": "image", + "bbox": [ + 0.764, + 0.277, + 0.879, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.799, + 0.382, + 0.846, + 0.392 + ], + "angle": 0, + "content": "Omni SA" + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.402, + 0.613, + 0.511 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.548, + 0.512, + 0.576, + 0.521 + ], + "angle": 0, + "content": "PSNR" + }, + { + "type": "image", + "bbox": [ + 0.617, + 0.402, + 0.75, + 0.512 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.671, + 0.512, + 0.737, + 0.522 + ], + "angle": 0, + "content": "Parameters (K)" + }, + { + "type": "image", + "bbox": [ + 0.75, + 0.402, + 0.875, + 0.512 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.811, + 0.512, + 0.848, + 0.522 + ], + "angle": 0, + "content": "Iteration" + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.533, + 0.892, + 0.562 + ], + "angle": 0, + "content": "Figure 1. Typical self-attention schemes [31,59] can only perform uni-dimensional (e.g., spatial-only) interactions, and." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.58, + 0.892, + 0.611 + ], + "angle": 0, + "content": "difficulties. This paper focuses on boosting the restoration performance of lightweight ViT-based frameworks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.614, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Two difficulties hinder the development of lightweight ViT-based models: 1) Uni-dimensional aggregation operators (i.e., spatial [31] only or channel [59] only) imprisons the full potential of self-attention operators. Contemporary self-attention generally realizes the interaction between pixels by calculating the cross-covariance of the spatial direction (i.e., width and height) and exchanges context information in a channel-separated manner. This interaction scheme ignores the explicit use of channel information. However, recent evidences [59] and our practice show that self-attention in the channel dimension (i.e., computationally more compact than spatial self-attention) is also crucial in low-level tasks. 2) Homogeneous aggregation schemes (i.e., Simple hierarchical stacking of single operators, e.g., convolution, self-attention) neglect abundant texture patterns of multi-scales, which is urgently needed in SR task. Specifically, a single operator is only sensitive to information of one scale [6, 12], e.g., self-attention is sensitive to long-term information and pays little attention to lo" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.875, + 0.209, + 0.888 + ], + "angle": 0, + "content": "*Equal Contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.097, + 0.888, + 0.295, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author: Bingbing Ni." + }, + { + "type": "list", + "bbox": [ + 0.096, + 0.875, + 0.295, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22378" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.182 + ], + "angle": 0, + "content": "cal information. Additionally, stacking of homogeneous operators proves to be inefficient and suffers from premature saturation of the interaction range [8], which is reflected as a suboptimal effective receptive field. The above problem is exacerbated in lightweight models because lightweight models cannot stack enough layers." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.182, + 0.473, + 0.681 + ], + "angle": 0, + "content": "In order to solve the above problems and pursue higher performance, this work proposes a novel omni-dimension feature aggregation scheme called Omni Self-Attention (OSA) exploiting both spatial and channel axis information in a simultaneous manner (i.e., extends the interaction into three-dimensional space), which offers higher-order receptive field information, as shown in Figure 1. Unlike scalar-based (a group of important coefficient) channel interaction [19], OSA enables comprehensive information propagation and interaction by cascading computation of the cross-covariance matrices between spatial/channel dimensions. The proposed OSA module can be plugged into any mainstream self-attention variants (e.g., Swin [35], Halo [50]), which provides a finer granularity of important encoding (compared to the vanilla channel attention [19]), achieving a perceptible improvement in contextual aggregation capabilities. Furthermore, a multi-scale hierarchical aggregation block, named Omni-Scale Aggregation Group (i.e., OSAG for short), is presented to achieve tailored encoding of varying scales of texture patterns. Specifically, OSAG builds three cascaded aggregators: local convolution (for local details), meso self-attention (focusing on mid-scale pattern processing), and global self-attention (pursuing global context understanding), rendering an omni-scale (i.e., local-/meso-/global-scale simultaneously) feature extraction capability. Compared to the homogenized feature extraction schemes [28, 31], our OSAG is able to mine richer information producing features with higher information entropy. Coupling with the above two designs, we establish a new ViT-based framework for lightweight superresolution, called Omni-SR, which exhibits superior restoration performance as well as covers a larger interaction range while maintaining an attractive model size, i.e., \\(792K\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.681, + 0.47, + 0.834 + ], + "angle": 0, + "content": "We extensively experiment with the proposed framework with both qualitative and quantitative evaluations on mainstream open-source image super-resolution datasets. It is demonstrated that our framework achieves state-of-the-art performance at the lightweight model scale (e.g., \\( \\text{Urban100} \\times 4: 26.95\\mathrm{dB} \\), \\( \\text{Manga109} \\times 4: 31.50\\mathrm{dB} \\)). More importantly, compared to existing ViT-based super-solution frameworks, our framework shows superior optimization properties (e.g., convergence speed, smoother loss landscape), which endow our model with better robustness." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.846, + 0.227, + 0.862 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Image Super-resolution. CNNs have achieved remarkable success in image SR task. SRCNN [13] is the first" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.089, + 0.695, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.696, + 0.089, + 0.891, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.229, + 0.893, + 0.271 + ], + "angle": 0, + "content": "Figure 2. Illustration of spatial attention and channel attention. These typical attention paradigms only model uni-dimensional (i.e., spatial-only / channel-only) interaction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.293, + 0.893, + 0.671 + ], + "angle": 0, + "content": "work to introduce CNNs into the SR field. Many methods [25, 48, 66] employ skip connection to speed up network convergence and improve the reconstruction quality. Channel Attention [66] is also proposed to enhance the representation ability of the SR model. In order to obtain better reconstruction quality with limited computing resources, several methods [23,38,42,47] explore lightweight architectural design. DRCN [26] utilizes the recursive operation to reduce the number of parameters. DRRN [47] introduces global and local residual learning on the basis of DRCN to accelerate training and improve the quality of details. CARN [1] employs cascading mechanism upon a residual network. IMDN [22] proposes an information multistillation block to archive better time performance. Another line of research is to utilize model compression techniques, e.g., knowledge distillation [15, 17, 65] and neural architecture search [11]) to reduce computing costs. Recently, a series of transformer-based SR models [5,8,31,37] emerge with superior performance. Chenet al. [5] develop a pre-trained model for the low-level computer vision task using the transformer architecture. Based on Swin transformer [35], SwinIR [31] proposes a three-stage framework, refreshing the state-of-the-art of SR task. More recently, some works [5,30] explore ImageNet pre-training strategy to further enhance SR performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Lightweight Vision Transformer. Due to the urgent demands for applying networks to resource-constrained devices, lightweight vision transformer [14, 51] has attracted widespread attention. Many attempts [7,9,10,37,41,43,57, 62] have been made to develop lightweight ViTs with comparable performance. A series of methods focus on combining convolutions with transformers to learn both local and global representations. For instance, LVT [57] introduces convolution in self-attention to enrich low-level features. MobileViT [41] replaces matrix multiplication in convolutions with transformer layers to learn global representations. Similarly, EdgeViTs [43] employs an information exchange bottleneck for full spatial interactions. Different from interpreting convolutions into vision transformers, LightViT [21] proposes aggregated self-attention for better" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22379" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.088, + 0.891, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.104, + 0.349, + 0.865, + 0.364 + ], + "angle": 0, + "content": "Figure 3. The overall architecture of the proposed Omni-SR framework and structure of OSAG and Omni Self-Attention (OSA)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.383, + 0.471, + 0.414 + ], + "angle": 0, + "content": "information aggregation. In this work, we resort to ViT architecture to achieve lightweight and accurate SR." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.445, + 0.212, + 0.463 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.476, + 0.444, + 0.493 + ], + "angle": 0, + "content": "3.1. Attention Mechanisms in Super-Resolution" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.506, + 0.47, + 0.537 + ], + "angle": 0, + "content": "Two attention paradigms are widely adopted in SR to assist in analyzing and aggregating comprehensive patterns." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.542, + 0.47, + 0.648 + ], + "angle": 0, + "content": "Spatial Attention. Spatial attention can be regarded as an anisotropic selection process. Spatial self-attention [37, 51] and spatial gate [10, 58] are predominantly applied. As shown in Figure 2, spatial self-attention calculates the cross-covariance along the spatial dimension, and the spatial gate generates the channel-separated masks. Neither of them can transmit information between channels." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.653, + 0.471, + 0.821 + ], + "angle": 0, + "content": "Channel Attention. There are two categories of channel attention, i.e., scalar-based [19] and covariance-based [59], proposed to perform channel recalibration or transmit patterns among channels. As shown in Figure 2, the former predicts a group of importance scalars to weigh different channels, while the latter computes a cross-covariance matrix to enable channel re-weighting and information transmission simultaneously. Compared to spatial attention, channel attention handles spatial dimension isotropically, and thus, the complexity is significantly reduced, which also impairs the accuracy of aggregation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Several attempts [44,55] have proved that both spatial attention and channel attention are beneficial for SR task and their characteristics are complementary to each other, thus integrating them in a computationally compact way would bring notable benefits in expressive capability." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.383, + 0.742, + 0.399 + ], + "angle": 0, + "content": "3.2. Omni Self-Attention Block" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.408, + 0.892, + 0.65 + ], + "angle": 0, + "content": "To mine all the correlations hidden in the latent variables, we propose a novel self-attention paradigm called Omni Self-Attention (OSA) block. Unlike existing self-attention paradigms (e.g., spatial self-attention [5, 37, 51]) that only indulge in unidimensional processing, OSA establishes the spatial and channel context simultaneously. The obtained two-dimensional relationship is highly necessary and beneficial, especially for lightweight models. On the one hand, as the network deepens, important information is scattered into different channels [19], and it is critical to deal with them in a timely manner. On the other hand, although spatial self-attention takes advantage of the channel dimension in calculating the covariance, it does not transmit the information between channels (refer to Sec. 3.1). Given the above conditions, our OSA is designed to transmit both spatial and dimensional information in a compact manner." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.651, + 0.893, + 0.901 + ], + "angle": 0, + "content": "The proposed OSA calculates the score matrices corresponding to the space and channel direction through sequential matrix operations and rotation, as illustrated in Figure 3. Specifically, suppose \\( X \\in \\mathbb{R}^{HW \\times C} \\) denotes the input feature, where \\( H \\) and \\( W \\) are the width and height of the input, and \\( C \\) is the channel number. Firstly, \\( X \\) is embedded to query, key and value matrices \\( Q^s, K^s, V^s \\in \\mathbb{R}^{HW \\times C} \\) through linearly projection. We calculate the production of query and key to get the spatial attention map of size \\( \\mathbb{R}^{HW \\times HW} \\). Then we perform the spatial attention to obtain the intermediate aggregated results. Note that window strategy is usually used to significantly reduce the resource overhead. Next stage, we rotate the input query and key matrices to get the transposed query and key matrices \\( Q^c, K^c \\in \\mathbb{R}^{C \\times HW} \\), and also rotate the value matrices to get the value matrix \\( V^c \\in \\mathbb{R}^{C \\times HW} \\) for the subsequent channel-wise self-attention. The obtained channel-wise attention map of size \\( \\mathbb{R}^{C \\times C} \\) models channel-wise relation" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22380" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.471, + 0.135 + ], + "angle": 0, + "content": "ships. Finally, we get the final aggregated \\( Y_{OSA} \\) by the inverse rotation of the channel attention output \\( Y_{c} \\). The whole OSA process is formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.117, + 0.144, + 0.47, + 0.16 + ], + "angle": 0, + "content": "\\[\nQ ^ {s} = X \\cdot \\mathcal {W} _ {q}, K ^ {s} = X \\cdot \\mathcal {W} _ {k}, V ^ {s} = X \\cdot \\mathcal {W} _ {v}, \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.118, + 0.163, + 0.469, + 0.18 + ], + "angle": 0, + "content": "\\[\nY _ {s} = \\mathcal {A} ^ {s} \\left(Q ^ {s}, K ^ {s}, V ^ {s}\\right) = \\operatorname {S o f t M a x} \\left(Q ^ {s} K ^ {s T}\\right) \\cdot V ^ {s}, \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.118, + 0.182, + 0.469, + 0.198 + ], + "angle": 0, + "content": "\\[\nQ ^ {c} = \\mathcal {R} \\left(Q ^ {\\prime}\\right), \\quad K ^ {c} = \\mathcal {R} \\left(K ^ {\\prime}\\right), \\quad V ^ {c} = \\mathcal {R} \\left(V ^ {\\prime}\\right), \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.118, + 0.201, + 0.469, + 0.218 + ], + "angle": 0, + "content": "\\[\nY _ {c} = \\mathcal {A} ^ {c} \\left(Q ^ {c}, K ^ {c}, V ^ {c}\\right) = \\operatorname {S o f t M a x} \\left(K ^ {c} Q ^ {c T}\\right) \\cdot V ^ {c}, \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.118, + 0.221, + 0.469, + 0.236 + ], + "angle": 0, + "content": "\\[\nY _ {O S A} = \\mathcal {R} ^ {- 1} \\left(Y _ {c}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.248, + 0.469, + 0.504 + ], + "angle": 0, + "content": "where \\(\\mathcal{W}_q, \\mathcal{W}_k, \\mathcal{W}_v\\) denote the linear projection matrices for the query, key, and value, respectively. \\(Q', K', V'\\) are the input embedding matrices of channel-wise self-attention, which are embedded from fore spatial self-attention or or copied directly from \\(Q^s, K^s, V^s\\). \\(\\mathcal{R}(\\cdot)\\) denotes the rotation operation around spatial axis and \\(\\mathcal{R}^{-1}(\\cdot)\\) is the inverse rotation. Some normalization factors are omitted for the sake of simplicity. In particular, this design shows compelling properties that can integrate the elementwise results of two matrix operations (i.e., spatial-/channel-matrix operation), thereby enabling omni-axial interactions. Note that our proposed OSA paradigm can be a drop-in replacement of the Swin [31,35] attention block to higher performance with less parameters. Benefiting from the smaller attention map size of channel self-attention, the proposed OSA is less computationally intensive compared to the cascade shifted-window self-attention scheme in Swin." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.506, + 0.47, + 0.687 + ], + "angle": 0, + "content": "Discussion with other hybrid attention paradigms. Compared to previous hybrid channel and spatial attention works like CBAM [55] and BAM [44], their scalar-based attention weights only reflect the relative degree of importance, without further inter-pixel information exchange, leading to limited relation modeling capability. Several recent works [8] also incorporate channel attention with spatial Self-attention, but these attempts only resort to scalar weights for channel recalibration, while our OSA paradigm enables channel-wise interaction to mine the potential correlations in omni-axis. Performance comparison of different attention paradigms can be found in Sec. 4.4." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.696, + 0.358, + 0.713 + ], + "angle": 0, + "content": "3.3. Omni-Scale Aggregation Group" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.47, + 0.902 + ], + "angle": 0, + "content": "How to utilize the proposed OSA paradigm to build a high-performance and compact network is another key topic. Although hierarchical stacking of windows-based self-attention (e.g., swin [31, 35]) has become mainstream, various works have found that the window-based paradigms are very inefficient for large-range interactions, especially for shallow networks. It is worth pointing out that large-range interaction can provide a pleasing effective receptive field, which is crucial for improving image restoration performance [37]. Unfortunately, direct global interaction is resource-prohibitive and detracts from local aggregation capabilities. Taking these points into account, we propose" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.091, + 0.871, + 0.157 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.166, + 0.892, + 0.208 + ], + "angle": 0, + "content": "Figure 4. Illustration of omni-scale aggregation scheme. Our proposed Omni-SR contains three types of feature aggregation at local level, meso and global level, respectively." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.218, + 0.892, + 0.49 + ], + "angle": 0, + "content": "an Omni-Scale Aggregation Group (i.e., OSAG for short) to pursue progressive receptive field feature aggregation with low computational complexity. As shown in Figure 3, OSAG mainly consists of three stages: local, meso and global aggregations. Specifically, a channel attention [19] enhanced inverted bottleneck [18] is introduced to fulfill the local pattern process with limited overhead. Based on the proposed OSA paradigm, we derive two instances (i.e., Meso-OSA and Global-OSA) responsible for the interaction and aggregation of meso and global information. Note that the proposed omni self-attention paradigm can be used for different purposes. Meso-OSA performs attention within a group of non-overlap patches, which restricts Meso-OSA to only focus on meso-scale pattern understanding. Global-OSA samples data point sparsely across the entire feature within an atrous manner, endowing Global-OSA with the ability to achieve global interactions at a compelling cost." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.491, + 0.892, + 0.675 + ], + "angle": 0, + "content": "The only difference between Meso-OSA and Global-OSA is the window partition strategy, as shown in Figure 4. In order to achieve meso-scale interaction, Meso-OSA split the input feature \\(X\\) into non-overlapping blocks with size \\(P \\times P\\). Note that after window partition, the block dimensions are gathered onto the spatial dimension (i.e., -2 axis): \\((H, W, C) \\to \\left(\\frac{H}{P} \\times P, \\frac{W}{P} \\times P, C\\right) \\to \\left(\\frac{HW}{P^2}, P^2, C\\right)\\). While the Global-OSA divides the input feature into a uniform \\(G \\times G\\) grid, with each lattice having an adaptive size of \\(\\frac{H}{G} \\times \\frac{W}{G}\\). Similar to Meso-OSA, the grid dimension is also gathered on the spatial axis (i.e., -2 axis): \\((H, W, C) \\to (G \\times \\frac{H}{G}, G \\times \\frac{W}{G}, C) \\to (G^2, \\frac{HW}{G^2}, C) \\to (\\frac{HW}{G^2}, G^2, C)\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.681, + 0.706, + 0.695 + ], + "angle": 0, + "content": "3.4. Network Architecture" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.702, + 0.892, + 0.828 + ], + "angle": 0, + "content": "Overall Structure. Based on the Omni Self-Attention paradigm and the Omni-Scale Aggregation Group, we further develop a lightweight Omni-SR framework to achieve high-performance image super-resolution. As shown in Figure 3, Omni-SR consists of three parts, i.e., shallow feature extraction, deep feature extraction, and image reconstruction. Specifically, given the LR input \\( I_{LR} \\in \\mathbb{R}^{H \\times W \\times C_{in}} \\), we first use a \\( 3 \\times 3 \\) convolution \\( H_{SF} \\) to extract shallow feature \\( X_0 \\in \\mathbb{R}^{H \\times W \\times C} \\) as" + }, + { + "type": "equation", + "bbox": [ + 0.644, + 0.836, + 0.891, + 0.851 + ], + "angle": 0, + "content": "\\[\nX _ {0} = H _ {S F} \\left(I _ {L R}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.858, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where \\( C_{in} \\) and \\( C \\) denote the channel number of the input and shallow feature. The convolution layer provides a simple way to convert the input from image space into" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22381" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.149 + ], + "angle": 0, + "content": "high-dimensional feature space. Then we use \\(K\\) stacked omni-scale aggregation groups (OSAG) and one \\(3 \\times 3\\) convolution layer \\(H_{CONV}\\) in a cascade manner to extract deep feature \\(F_{DF}\\). Such a process can be expressed as" + }, + { + "type": "equation", + "bbox": [ + 0.148, + 0.158, + 0.47, + 0.18 + ], + "angle": 0, + "content": "\\[\nX _ {i} = H _ {O S A G _ {i}} \\left(X _ {i - 1}\\right), \\quad i = 1, 2, \\dots , K, \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.211, + 0.176, + 0.338, + 0.19 + ], + "angle": 0, + "content": "\\[\nX _ {D F} = H _ {C O N V} \\left(X _ {K}\\right),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.199, + 0.469, + 0.271 + ], + "angle": 0, + "content": "where \\( H_{\\text{OSAG}_i} \\) represents the \\( i \\)-th OSAG, \\( X_1, X_2, \\ldots, X_K \\) denote intermediate features. Following [31], we also use a convolutional layer at the end of feature extraction to get better feature aggregation. Finally we reconstruct the HR image \\( I_{HR} \\) by aggregating shallow and deep features as" + }, + { + "type": "equation", + "bbox": [ + 0.197, + 0.28, + 0.469, + 0.295 + ], + "angle": 0, + "content": "\\[\nI _ {H R} = H _ {R e c} \\left(X _ {0} + X _ {D F}\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.305, + 0.468, + 0.336 + ], + "angle": 0, + "content": "where \\(H_{Rec}(\\cdot)\\) denotes the reconstruction module. In detail, PixelShuffle [46] is used to up-sample the fused feature." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.352, + 0.469, + 0.422 + ], + "angle": 0, + "content": "Omni-Scale Aggregation Group (OSAG). As shown in Figure 3, each OSAG contains a local convolution block (LCB), a meso-OSA block, a global-OSA block, and an ESA block [28, 34]. The whole process can be formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.121, + 0.432, + 0.469, + 0.448 + ], + "angle": 0, + "content": "\\[\nX _ {r e s} = H _ {\\text {G l o b a l - O S A B} _ {i}} \\left(H _ {\\text {M e s o - O S A B} _ {i}} \\left(H _ {\\text {L C B} _ {i}} \\left(X _ {i - 1}\\right)\\right)\\right), \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.162, + 0.45, + 0.469, + 0.465 + ], + "angle": 0, + "content": "\\[\nX _ {i} = H _ {E S A _ {i}} \\left(H _ {\\text {C o n v} _ {i}} \\left(X _ {\\text {r e s}} + X _ {i - 1}\\right)\\right), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.475, + 0.469, + 0.566 + ], + "angle": 0, + "content": "where \\( X_{i-1} \\) and \\( X_i \\) represents the input and output feature of \\( i \\)-th OSAG. After the mapping of convolution layers, we insert a Meso-OSAB for window-based self-attention and a Global-OSAB to enlarge the receptive field for better information aggregation. At the end of OSAG, we reserve the convolutional layer and ESA block following [28, 66]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.566, + 0.469, + 0.852 + ], + "angle": 0, + "content": "In specific, LCB is implemented as a stack of pointwise and depthwise convolutions with a CA module [24] between them to adaptively re-weight channel-wise features. This block aims to aggregate local contextual information as well as to increase the trainability of the network [56]. Two types of OSA blocks (i.e., Meso-OSA block and Global-OSA block) are then followed to obtain interactions from different regions. Based on different window partition strategies, Meso-OSA block seeks inner-block interaction, and Global-OSA blocks aim for global mixing. OSA blocks follow typical Transformers designs with Feedforward network (FFN) and LayerNorm [2], and the only difference is that the origin self-attention operation is replaced with our proposed OSA operator. For FFN, we adopt the GDFN proposed by Restormer [59]. Combining these individuals seamlessly, the designed OSAG enables information propagation between any pair of tokens in the feature map. We use the ESA module proposed in [28, 34] to further refine the fused feature." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Optimization Objective. Following prior works [31, 32, 53, 67], we train the model by minimizing a standard \\( L_{1} \\)" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.091, + 0.892, + 0.12 + ], + "angle": 0, + "content": "loss between model prediction \\(\\hat{I}_{HR}\\) and HR label \\(I_{HR}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.63, + 0.123, + 0.892, + 0.14 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\left\\| I _ {H R} - \\hat {I} _ {H R} \\right\\| _ {1}. \\tag {11}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.165, + 0.632, + 0.181 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.19, + 0.691, + 0.207 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.215, + 0.892, + 0.366 + ], + "angle": 0, + "content": "Datasets and Metrics. Following previous work [31,32, 38,49,66], DIV2K [49] and Flickr2K [49] are used as training datasets. For a fair comparison, we employ two training protocols, i.e., training with DIV2K only and training with DF2K (DIV2K + Flickr2K). Note that the model trained with DF2K is marked with small \\(\\dagger\\). For testing, we adopt five standard benchmark datasets: Set5 [4], Set14 [60], B100 [39], Urban100 [20] and Manga109 [40]. PSNR and SSIM [54] are adopted to evaluate the SR performance on the Y channel of the transformed YCbCr space." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.367, + 0.892, + 0.639 + ], + "angle": 0, + "content": "Implementation Details. During training, we augment the data with random horizontal flips and 90/270-degree rotations. LR images are generated by bicubic downsampling [63] from HR images. OSAG number is set to 5, and channel number of the whole network is set to 64. The attention head number and window size are set to 4 and 8 for both Meso-OSAB and Global-OSAB. We use Adam [27] optimizer to train the model with a batch size of 64 for \\(800\\mathrm{K}\\) iterations. The initial learning rate is set to \\(10^{-4}\\) and halved for every \\(200\\mathrm{k}\\) iterations. In each training batch, we randomly crop LR patches of size \\(64\\times 64\\) as input. Our method is implemented with PyTorch [45], and all experiments are conducted on one NVIDIA V100 GPU. Note that no other data augmentation (e.g., Mixup [61], RGB channel shuffle) or training skills (e.g., pre-training [30], cosine learning schedule [36]) are employed. It should be pointed out that we maintain the consistency of model parameters in the ablation study by adjusting the channels of \\(1\\times 1\\) convolution." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.65, + 0.845, + 0.666 + ], + "angle": 0, + "content": "4.2. Comparison with the SOTA SR methods" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.674, + 0.891, + 0.78 + ], + "angle": 0, + "content": "To evaluate the effectiveness of Omni-SR, we compare our model with several advanced lightweight SR methods under a scale factor of 2/3/4. In particular, former works, VDSR [25], CARN [1], IMDN [22], EDSR [32], RFDN [33], MemNet [48], MAFFSRN [42], LatticeNet [38], RLFN [28], ESRT [37] and SwinIR [31] are introduced for comparison." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Quantitative results. In Table 1, the quantitative comparisons of different lightweight methods are presented on five benchmark datasets. With a similar model size, the performance of our Omni-SR surpasses existing methods with a notable margin on all benchmarks. In particular, compared to other transformer architectures with comparable parameters like SwinIR [31] and ESRT [37], the proposed Omni-SR obtains the best performance. The results" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22382" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.09, + 0.892, + 0.117 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison (PSNR/SSIM) for lightweight image SR with state-of-the-art methods on benchmark datasets. The best and second-best results are marked in red and blue colors, respectively. “†” indicates that model is trained on DF2K." + }, + { + "type": "table", + "bbox": [ + 0.09, + 0.119, + 0.88, + 0.543 + ], + "angle": 0, + "content": "
MethodYearsScaleParams(K)Set5 PSNR/ SSIMSet14 PSNR/ SSIMBSD100 PSNR/ SSIMUrban100 PSNR/ SSIMManga109 PSNR/ SSIM
VDSR [25]CVPR1666636.66 / 0.954233.05 / 0.912731.90 / 0.896030.76 / 0.914037.22 / 0.9750
MemNet [48]ICCV1767837.78 / 0.959733.28 / 0.914232.08 / 0.897831.31 / 0.919537.72 / 0.9740
SRMDNF [64]CVPR18151137.79 / 0.96033.32 / 0.91532.05 / 0.898531.33 / 0.920438.07 / 0.9761
CARN [1]ECCV181,59237.76 / 0.959033.52 / 0.916632.09 / 0.897831.92 / 0.925638.36 / 0.9765
IMDN [22]MM1969438.00 / 0.960533.63 / 0.917732.19 / 0.899632.17 / 0.928338.88 / 0.9774
RFDN-L [33]ECCV2062638.08 / 0.960633.67 / 0.919032.18 / 0.899632.24 / 0.929038.95 / 0.9773
MAFFSRN [42]ECCV20×240237.97 / 0.960333.49 / 0.917032.14 / 0.899431.96 / 0.9268-/-
LatticeNet [38]ECCV2075638.15 / 0.961033.78 / 0.919332.25 / 0.900532.43 / 0.9302-/-
RLFN [28]CVPRW2252738.07 / 0.960733.72 / 0.918732.22 / 0.900032.33 / 0.9299-/-
SwinIR [31]ICCVW2187838.14 / 0.961133.86 / 0.920632.31 / 0.901232.76 / 0.934039.12 / 0.9783
Omni-SROurs77238.22 / 0.961333.98 / 0.921032.36 / 0.902033.05 / 0.936339.28 / 0.9784
Omni-SR†Ours77238.29 / 0.961734.27 / 0.923832.41 / 0.902633.30 / 0.938639.53 / 0.9792
VDSR [25]CVPR1666633.66 / 0.921329.77 / 0.831428.82 / 0.797627.14 / 0.827932.01 / 0.9340
MemNet [48]ICCV1767834.09 / 0.924830.00 / 0.835028.96 / 0.800127.56 / 0.837632.51 / 0.9369
EDSR [32]CVPRW171,55534.37 / 0.927030.28 / 0.841729.09 / 0.805228.15 / 0.852733.45 / 0.9439
SRMDNF [64]CVPR181,52834.12 / 0.925430.04 / 0.838228.97 / 0.802527.57 / 0.839833.00 / 0.9403
CARN [1]ECCV181,59234.29 / 0.925530.29 / 0.840729.06 / 0.803428.06 / 0.849333.50 / 0.9440
IMDN [22]MM1970334.36 / 0.927030.32 / 0.841729.09 / 0.804628.17 / 0.851933.61 / 0.9445
RFDN-L [33]ECCV20×363334.47 / 0.928030.35 / 0.842129.11 / 0.805328.32 / 0.854733.78 / 0.9458
MAFFSRN [42]ECCV2080734.45 / 0.927730.40 / 0.843229.13 / 0.806128.26 / 0.8552-/-
LatticeNet [38]ECCV2076534.53 / 0.928130.39 / 0.842429.15 / 0.805928.33 / 0.8538-/-
ESRT [37]CVPRW2277034.42 / 0.926830.43 / 0.843329.15 / 0.806328.46 / 0.857433.95 / 0.9455
SwinIR [31]ICCVW2188634.62 / 0.928930.54 / 0.846329.20 / 0.808228.66 / 0.862433.98 / 0.9478
Omni-SROurs78034.70 / 0.929430.57 / 0.846929.28 / 0.809428.84 / 0.865634.22 / 0.9487
Omni-SR†Ours78034.77 / 0.930430.70 / 0.848929.33 / 0.811129.12 / 0.871234.64 / 0.9507
VDSR [25]CVPR1666631.35 / 0.883828.01 / 0.767427.29 / 0.725125.18 / 0.752428.83 / 0.8870
MemNet [48]ICCV1767831.74 / 0.889328.26 / 0.772327.40 / 0.728125.50 / 0.763029.42 / 0.8942
EDSR [32]CVPRW171,51832.09 / 0.893828.58 / 0.781327.57 / 0.735726.04 / 0.784930.35 / 0.9067
SRMDNF [64]CVPR181,55231.96 / 0.892528.35 / 0.778727.49 / 0.733725.68 / 0.773130.09 / 0.9024
CARN [1]ECCV181,59232.13 / 0.893728.60 / 0.780627.58 / 0.734926.07 / 0.783730.47 / 0.9084
IMDN [22]MM1971532.21 / 0.894828.58 / 0.781127.56 / 0.735326.04 / 0.783830.45 / 0.9075
RFDN-L [33]ECCV20×464332.28 / 0.895728.61 / 0.781827.58 / 0.736326.20 / 0.788330.61 / 0.9096
MAFFSRN [42]ECCV2083032.20 / 0.895326.62 / 0.782227.59 / 0.737026.16 / 0.7887-/-
LatticeNet [38]ECCV2077732.30 / 0.896228.68 / 0.783027.62 / 0.736726.25 / 0.7873-/-
ESRT [37]CVPRW2254332.24 / 0.895228.62 / 0.781327.60 / 0.736426.17 / 0.7877-/-
SwinIR [31]ICCVW2175132.19 / 0.894728.69 / 0.783327.69 / 0.737926.39 / 0.796230.75 / 0.9100
Omni-SROurs89732.44 / 0.897628.77 / 0.785827.69 / 0.740626.47 / 0.798030.92 / 0.9151
Omni-SR†Ours79232.49 / 0.898828.78 / 0.785927.71 / 0.741526.64 / 0.801831.02 / 0.9151
Omni-SR†Ours79232.57 / 0.899328.95 / 0.789827.81 / 0.743926.95 / 0.810531.50 / 0.9192
" + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.557, + 0.271, + 0.689 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.173, + 0.695, + 0.191, + 0.706 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.557, + 0.465, + 0.694 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.374, + 0.695, + 0.393, + 0.706 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.71, + 0.47, + 0.766 + ], + "angle": 0, + "content": "Figure 5. (a) PSNR improvement of Omni-SR variants with different OSAG number (K) over the smallest Omni-SR model \\((\\mathrm{K} = 1)\\) for \\(4\\times\\) SR. (b) The number of model parameters vs. PSNR of different lightweight methods on Urban100 dataset for \\(4\\times\\) SR." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.47, + 0.902 + ], + "angle": 0, + "content": "showcase that the omni-axis (i.e., spatial + channel) interaction introduced by OSA can effectively boost the model's contextual aggregation capabilities, which promises superior SR performance. Coupling with large training dataset DF2K, the performance can be further improved, especially on Urban100. We suppose that such a phenomenon can be ascribed to the images in Urban100 have many similar patches, and the long-term relationship introduced by" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.56, + 0.892, + 0.635 + ], + "angle": 0, + "content": "OSAG can bring great benefits for detail restoration. More importantly, with similar parameters, our model reduces \\(28\\%\\) of computational complexity (Omni-SR: 36G FLOPs vs. SwinIR: 50G FLOPs @1280×720), showing its effectiveness and efficiency." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.64, + 0.893, + 0.806 + ], + "angle": 0, + "content": "Visual comparison. In Figure 6, we also provide a visual comparison of different lightweight SR methods at \\(\\times 4\\) scale. We can observe that the HR images constructed by Omni-SR contain more fine-grained details, while other methods generate blurred edges or artifacts in complicated areas. For instance, in the \\(1st\\) row, our model is able to restore the detailed texture of the wall pleasantly, which all other methods fail to restore. Visual results also validate the effectiveness of the proposed OSA paradigm, which can perform omni-axis pixel-wise interaction modeling, thus obtaining a more powerful reconstruction capability." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Trade-off between Model Size and Performance. In experiments, we set the number of OSAG as 5 to make the model size around \\(800\\mathrm{K}\\) for a fair comparison with other methods. We also explore model performance with smaller parameter sizes by reducing OSAG number K. As shown in Figure 5(a), compared to the smallest model variant with" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22383" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.089, + 0.309, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.116, + 0.204, + 0.298, + 0.219 + ], + "angle": 0, + "content": "Urban100: img_012.png" + }, + { + "type": "image", + "bbox": [ + 0.311, + 0.089, + 0.423, + 0.133 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.338, + 0.133, + 0.4, + 0.154 + ], + "angle": 0, + "content": "PSNR Reference" + }, + { + "type": "image", + "bbox": [ + 0.425, + 0.089, + 0.536, + 0.133 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.453, + 0.133, + 0.509, + 0.154 + ], + "angle": 0, + "content": "20.65 dB Bicubic" + }, + { + "type": "image", + "bbox": [ + 0.538, + 0.089, + 0.65, + 0.133 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.552, + 0.133, + 0.636, + 0.155 + ], + "angle": 0, + "content": "21.15 dB SRCNN [13]" + }, + { + "type": "image", + "bbox": [ + 0.652, + 0.089, + 0.762, + 0.133 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.665, + 0.133, + 0.752, + 0.155 + ], + "angle": 0, + "content": "21.22 dB FSRCNN [29]" + }, + { + "type": "image", + "bbox": [ + 0.764, + 0.089, + 0.876, + 0.133 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.791, + 0.133, + 0.853, + 0.155 + ], + "angle": 0, + "content": "21.99 dB CARN [1]" + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.224, + 0.308, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.116, + 0.337, + 0.298, + 0.352 + ], + "angle": 0, + "content": "Urban100: img_024.png" + }, + { + "type": "image", + "bbox": [ + 0.311, + 0.156, + 0.423, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.335, + 0.198, + 0.403, + 0.222 + ], + "angle": 0, + "content": "21.95 dB IMDN [22]" + }, + { + "type": "image", + "bbox": [ + 0.425, + 0.156, + 0.536, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.446, + 0.198, + 0.517, + 0.222 + ], + "angle": 0, + "content": "21.95 dB RLFN [28]" + }, + { + "type": "image", + "bbox": [ + 0.538, + 0.156, + 0.65, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.548, + 0.2, + 0.639, + 0.222 + ], + "angle": 0, + "content": "22.00 dB LatticeNet [38]" + }, + { + "type": "image", + "bbox": [ + 0.652, + 0.156, + 0.763, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.669, + 0.2, + 0.746, + 0.222 + ], + "angle": 0, + "content": "22.34 dB SwinIR [31]" + }, + { + "type": "image", + "bbox": [ + 0.765, + 0.156, + 0.876, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.767, + 0.2, + 0.875, + 0.222 + ], + "angle": 0, + "content": "24.16 dB Omni-SR (Ours)" + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.358, + 0.308, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.11, + 0.47, + 0.303, + 0.485 + ], + "angle": 0, + "content": "Set14: img_barbara.png" + }, + { + "type": "image", + "bbox": [ + 0.311, + 0.2, + 0.423, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.311, + 0.222, + 0.423, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.338, + 0.266, + 0.398, + 0.288 + ], + "angle": 0, + "content": "PSNR Reference" + }, + { + "type": "image", + "bbox": [ + 0.425, + 0.2, + 0.536, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.425, + 0.222, + 0.536, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.455, + 0.266, + 0.509, + 0.288 + ], + "angle": 0, + "content": "16.90 dB Bicubic" + }, + { + "type": "image", + "bbox": [ + 0.538, + 0.2, + 0.65, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.538, + 0.222, + 0.651, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.552, + 0.266, + 0.636, + 0.288 + ], + "angle": 0, + "content": "17.54 dB SRCNN [13]" + }, + { + "type": "image", + "bbox": [ + 0.652, + 0.2, + 0.763, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.652, + 0.222, + 0.762, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.665, + 0.266, + 0.751, + 0.288 + ], + "angle": 0, + "content": "17.71 dB FSRCNN [29]" + }, + { + "type": "image", + "bbox": [ + 0.765, + 0.2, + 0.876, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.765, + 0.222, + 0.876, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.79, + 0.266, + 0.853, + 0.288 + ], + "angle": 0, + "content": "18.78 dB CARN [1]" + }, + { + "type": "image", + "bbox": [ + 0.311, + 0.289, + 0.423, + 0.333 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.338, + 0.332, + 0.402, + 0.355 + ], + "angle": 0, + "content": "18.77 dB \n[MDN [22]]" + }, + { + "type": "image", + "bbox": [ + 0.425, + 0.289, + 0.536, + 0.333 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.446, + 0.332, + 0.517, + 0.355 + ], + "angle": 0, + "content": "19.23 dB RLFN [28]" + }, + { + "type": "image", + "bbox": [ + 0.538, + 0.289, + 0.651, + 0.333 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.547, + 0.332, + 0.639, + 0.355 + ], + "angle": 0, + "content": "18.91 dB LatticeNet [38]" + }, + { + "type": "image", + "bbox": [ + 0.652, + 0.289, + 0.763, + 0.333 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.669, + 0.332, + 0.746, + 0.355 + ], + "angle": 0, + "content": "19.07 dB SwinIR [31]" + }, + { + "type": "image", + "bbox": [ + 0.765, + 0.289, + 0.876, + 0.333 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.767, + 0.332, + 0.875, + 0.355 + ], + "angle": 0, + "content": "21.18 dB Omni-SR (Ours)" + }, + { + "type": "image", + "bbox": [ + 0.311, + 0.356, + 0.423, + 0.4 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.338, + 0.399, + 0.398, + 0.421 + ], + "angle": 0, + "content": "PSNR Reference" + }, + { + "type": "image", + "bbox": [ + 0.425, + 0.356, + 0.536, + 0.4 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.453, + 0.399, + 0.508, + 0.42 + ], + "angle": 0, + "content": "23.58 dB Bicubic" + }, + { + "type": "image", + "bbox": [ + 0.538, + 0.356, + 0.65, + 0.4 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.552, + 0.4, + 0.636, + 0.422 + ], + "angle": 0, + "content": "24.04 dB SRCNN [13]" + }, + { + "type": "image", + "bbox": [ + 0.652, + 0.356, + 0.763, + 0.4 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.665, + 0.4, + 0.751, + 0.422 + ], + "angle": 0, + "content": "24.11 dB FSRCNN [29]" + }, + { + "type": "image", + "bbox": [ + 0.765, + 0.356, + 0.876, + 0.4 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.79, + 0.399, + 0.853, + 0.422 + ], + "angle": 0, + "content": "24.48 dB CARN [1]" + }, + { + "type": "image", + "bbox": [ + 0.311, + 0.422, + 0.423, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.335, + 0.466, + 0.402, + 0.489 + ], + "angle": 0, + "content": "24.62 dB IMDN [22]" + }, + { + "type": "image", + "bbox": [ + 0.425, + 0.422, + 0.536, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.446, + 0.466, + 0.517, + 0.489 + ], + "angle": 0, + "content": "24.43 dB RLFN [28]" + }, + { + "type": "image", + "bbox": [ + 0.538, + 0.422, + 0.65, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.547, + 0.466, + 0.639, + 0.489 + ], + "angle": 0, + "content": "24.64 dB LatticeNet [38]" + }, + { + "type": "image", + "bbox": [ + 0.652, + 0.422, + 0.763, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.669, + 0.466, + 0.746, + 0.489 + ], + "angle": 0, + "content": "24.45 dB SwinIR [31]" + }, + { + "type": "image", + "bbox": [ + 0.765, + 0.422, + 0.876, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.767, + 0.466, + 0.875, + 0.489 + ], + "angle": 0, + "content": "26.19 dB Omni-SR (Ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.493, + 0.879, + 0.508 + ], + "angle": 0, + "content": "Figure 6. Visual comparison for \\(\\times 4\\) SR methods. The patches for comparison are marked with red boxes. (Best viewed by zooming.)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.522, + 0.47, + 0.597 + ], + "angle": 0, + "content": "\\(\\mathrm{K} = 1\\) increasing the number of OSAG leads to stable performance improvements. In Figure 5(b), we present PSNR vs. parameters of different methods. It can be found that Omni-SR achieves the best results under various settings, showing its effectiveness and scalability." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.606, + 0.358, + 0.621 + ], + "angle": 0, + "content": "4.3. Analysis of Omni Self-Attention" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.471, + 0.902 + ], + "angle": 0, + "content": "In this section, we illustrate the optimization features of OSA and further uncover its underlying mechanism. Self-attention is a low-bias operator, which makes its optimization difficult and requires more training epochs. For this, we introduce the additional channel-wise interaction to alleviate it. In Figure 7(a), we show the loss curves of different self-attention paradigms on the DIV2K training set, including spatial self-attention, channel self-attention, and the proposed omni self-attention. We can see that our OSA presents a obviously superior convergence speed. More importantly, the performance at the final epoch is also significantly ahead of them. The above phenomenon clearly shows that our OSA has superior good optimization characteristics. Further, we delve into why channel-wise interactions lead to these improvements. We calculate the normalized entropies [52] of the hidden layer features of the network composed of the above three computational primitives. We illustrate the entropy results in Figure 7(c)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.522, + 0.892, + 0.718 + ], + "angle": 0, + "content": "As shown in the figure, in all outgoing layers, our OSA-encoded features show higher entropy, indicating that our OSA encodes richer information. More information may come from various scales, and this information can help the operators to reconstruct the exact details faster. We speculate that this is the potential reason why our OSA shows better optimization properties. In addition, following previous works [8, 16], we also resort to LAM analysis. DI [16] metric can measure the furthest interaction distance of the model. From Figure 8 we can observe that Omni-SR generally has the highest max diffusion index than other methods, showing that our OSA paradigm can effectively capture long-range interactions." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.727, + 0.653, + 0.743 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Effect of Omni Self-Attention. The core idea of our framework is to extend the vanilla self-attention with a channel-wise relationship to build omni-axis pixel-wise interaction. Based on Omni-SR framework, we design several variant models, and their SR results are reported in Table 2. We first simply remove the channel-wise component to form a spatial-only variant \\((\\mathrm{Omni - SR_{sp}})\\), and its performance is degraded by \\(0.13\\mathrm{dB}\\) compared to the full model. Such a significant degradation justifies the importance of channel interactions. Note that Omni-\\(\\mathrm{SR_{sp}}\\) still" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22384" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.14, + 0.089, + 0.424, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.202, + 0.202, + 0.395, + 0.215 + ], + "angle": 0, + "content": "(a) Ablation studies on OSAG" + }, + { + "type": "image", + "bbox": [ + 0.437, + 0.088, + 0.627, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.453, + 0.202, + 0.63, + 0.216 + ], + "angle": 0, + "content": "(b) Training loss on DIV2K" + }, + { + "type": "image", + "bbox": [ + 0.642, + 0.09, + 0.824, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.677, + 0.201, + 0.82, + 0.214 + ], + "angle": 0, + "content": "(c) Feature correlation" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.223, + 0.894, + 0.252 + ], + "angle": 0, + "content": "Figure 7. (a) Ablation studies on different components in OSAG. (b) Training loss of different attention paradigms. (c) Feature correlation analysis of different attention paradigms. All the results are reported on DIV2K dataset for \\(4 \\times\\) SR." + }, + { + "type": "image", + "bbox": [ + 0.103, + 0.271, + 0.445, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.402, + 0.471, + 0.444 + ], + "angle": 0, + "content": "Figure 8. The distribution of DI values for different methods on Urban100. Top-left points with show a narrow area of interest, and right-bottom points show a large area of interest." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.461, + 0.47, + 0.673 + ], + "angle": 0, + "content": "outperforms SwinIR by \\(0.04\\mathrm{dB}@\\) Urban \\(100\\times 4\\) , which benefits from the global interaction introduced by grid window partition. Similarly, we remove the spatial self-attention component to derive a channel self-attention variant, Omni\\(\\mathrm{SR}_{\\mathrm{ca}}\\) , and such a modification also leads to undesirable performance degradation. Besides, we use the most widely adopted channel and spatial attention configurations (i.e., SE [19] and CBAM [55]) to act as alternative operators for channel and spatial aggregation. Both substitutions (Omni\\(\\mathrm{SR}_{\\mathrm{SE}}\\) , Omni-\\(\\mathrm{SR}_{\\mathrm{CBAM}}\\) ) hurt PNSR performance compared to the full model. The above results show that the specific interaction paradigm (e.g., scalar-based, covariance-based) is equally important, and our channel interaction based on the covariance matrix shows great advantages." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Effect of Omni-Scale Aggregation Group. In Omni-SR, we propose a local-meso-global interaction scheme (i.e., OSAG) to pursue progressive feature aggregation. To investigate its effectiveness, we design three different kinds of interaction schemes based on Omni-SR framework: Separated scheme, Mixed scheme, and our fully designed Omni scheme (i.e., our proposed OSAG), and the ablation study results are shown in Figure 7(a). In the figure, we employ different words (e.g., \"Local\", \"Meso+Global\") to represent specific schemes, e.g., \"Local\" denotes using Local-Conv block to replace Meso-OSA and Global-OSA; \"Local+Global\" represents replace original cascaded Meso-OSA and Global-OSA with cascaded Local-Conv and Global-OSA. We can observe that single interaction schemes (e.g., \"Local\") perform the worst. Interestingly," + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.269, + 0.892, + 0.34 + ], + "angle": 0, + "content": "Table 2. Ablation studies of omni self-attention on Urban100. \\(\\mathrm{Omni - SR}_{(*)}\\) denotes different modifications. 'SA' and 'S-SA' denote spatial gate and spatial self-attention. 'CA' and 'C-SA' denote channel gate and channel self-attention. We maintain the consistency of model parameters by adjusting the channels of \\(1\\times 1Conv\\)" + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.345, + 0.891, + 0.432 + ], + "angle": 0, + "content": "
ModelSAS-SACAC-SAFLOPs×2×3×4
Omni-SRsp×✓××33G32.8828.7226.51
Omni-SRSE×✓✓×34G32.8328.7126.50
Omni-SRCBAM✓×✓✓34G32.9228.7626.53
Omni-SRca×××✓33G32.6528.6026.45
Omni-SRfull×✓×✓36G33.0528.8426.64
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.448, + 0.892, + 0.614 + ], + "angle": 0, + "content": "the \"Global\" scheme is inferior to the \"Meso\" one due to its poor optimization properties of global self-attention [3, 35, 50]. Once two interaction operators are combined, the performance improves steadily. Among them, \"Meso+ Global\" setting achieves the second-best performance. Furthermore, combining all three interaction schemes together, we obtain the best performing scheme, i.e., \"Omni\". From the above experiments, we can infer that obvious performance gains can be obtained by introducing various-scale interactions, which also illustrates the feasibility and effectiveness of our proposed OSAG." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.628, + 0.619, + 0.643 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.653, + 0.892, + 0.804 + ], + "angle": 0, + "content": "In this work, we propose Omni-SR, a lightweight framework for image SR. We propose the Omni Self-attention paradigm for simultaneous spatial and channel interactions, mining all the potential correlations across omni-axis. Furthermore, we propose an omni-scale aggregation scheme to effectively enlarge the receptive fields with low computational complexity, which encodes contextual relations in a progressively hierarchical manner. Extensive experiments on public benchmark datasets and comprehensive analytical studies validate its prominent SR performance." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.816, + 0.681, + 0.833 + ], + "angle": 0, + "content": "6. Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.9 + ], + "angle": 0, + "content": "This work was supported by National Science Foundation of China (U20B2072, 61976137). This work was also partly supported by SJTU Medical Engineering Cross Research Grant YG2021ZD18." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22385" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.156 + ], + "angle": 0, + "content": "[1] Namhyuk Ahn, Byungkon Kang, and Kyung-Ah Sohn. Fast, accurate, and lightweight super-resolution with cascading residual network. In ECCV, 2018. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.157, + 0.468, + 0.185 + ], + "angle": 0, + "content": "[2] Lei Jimmy Ba, Jamie Ryan Kiros, and Geoffrey E. Hinton. Layer normalization. CoRR, abs/1607.06450, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.187, + 0.468, + 0.226 + ], + "angle": 0, + "content": "[3] Irwan Bello, Barret Zoph, Quoc Le, Ashish Vaswani, and Jonathon Shlens. Attention augmented convolutional networks. In ICCV, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.228, + 0.468, + 0.282 + ], + "angle": 0, + "content": "[4] Marco Bevilacqua, Aline Roumy, Christine Guillemot, and Marie-Line Alberi-Morel. Low-complexity single-image super-resolution based on nonnegative neighbor embedding. In BMVC, 2012. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.284, + 0.468, + 0.338 + ], + "angle": 0, + "content": "[5] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In CVPR, 2021. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.34, + 0.468, + 0.395 + ], + "angle": 0, + "content": "[6] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. arXiv:1606.00915, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.397, + 0.468, + 0.437 + ], + "angle": 0, + "content": "[7] Xuanhong Chen, Hang Wang, and Bingbing Ni. X-volution: On the unification of convolution and self-attention. arXiv preprint arXiv:2106.02253, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.439, + 0.468, + 0.479 + ], + "angle": 0, + "content": "[8] Xiangyu Chen, Xintao Wang, Jiantao Zhou, and Chao Dong. Activating more pixels in image super-resolution transformer. CoRR, abs/2205.04437, 2022. 2, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.481, + 0.468, + 0.534 + ], + "angle": 0, + "content": "[9] Yinpeng Chen, Xiyang Dai, Dongdong Chen, Mengchen Liu, Xiaoyi Dong, Lu Yuan, and Zicheng Liu. Mobileformer: Bridging mobilenet and transformer. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.537, + 0.468, + 0.577 + ], + "angle": 0, + "content": "[10] Yinpeng Chen, Xiyang Dai, Dongdong Chen, Mengchen Liu, Xiaoyi Dong, Lu Yuan, and Zicheng Liu. Simple baselines for image restoration. In ECCV, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.579, + 0.468, + 0.619 + ], + "angle": 0, + "content": "[11] Xiangxiang Chu, Bo Zhang, Hailong Ma, Ruijun Xu, and Qingyuan Li. Fast, accurate and lightweight super-resolution with neural architecture search. In ICPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.621, + 0.468, + 0.661 + ], + "angle": 0, + "content": "[12] Xiaohan Ding, Xiangyu Zhang, Jungong Han, and Guiguang Ding. Scaling up your kernels to 31x31: Revisiting large kernel design in cnns. In CVPR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.663, + 0.468, + 0.703 + ], + "angle": 0, + "content": "[13] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Image super-resolution using deep convolutional networks. CoRR, abs/1501.00092, 2015. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.705, + 0.468, + 0.787 + ], + "angle": 0, + "content": "[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.468, + 0.829 + ], + "angle": 0, + "content": "[15] Qinquan Gao, Yan Zhao, Gen Li, and Tong Tong. Image super-resolution using knowledge distillation. In ACCV, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.468, + 0.858 + ], + "angle": 0, + "content": "[16] Jinjin Gu and Chao Dong. Interpreting super-resolution networks with local attribution maps. In CVPR, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.468, + 0.9 + ], + "angle": 0, + "content": "[17] Geoffrey E. Hinton, Oriol Vinyals, and Jeffrey Dean. Distilling the knowledge in a neural network. CoRR, abs/1503.02531, 2015. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.161 + ], + "angle": 0, + "content": "[18] Andrew G Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andretto, and Hartwig Adam. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv:1704.04861, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[19] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In CVPR, 2018. 2, 3, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.233 + ], + "angle": 0, + "content": "[20] Jia-Bin Huang, Abhishek Singh, and Narendra Ahuja. Single image super-resolution from transformed self-exemplars. In CVPR, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.235, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[21] Tao Huang, Lang Huang, Shan You, Fei Wang, Chen Qian, and Chang Xu. Lightvit: Towards light-weight convolution-free vision transformers. CoRR, abs/2207.05557, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.277, + 0.892, + 0.318 + ], + "angle": 0, + "content": "[22] Zheng Hui, Xinbo Gao, Yunchu Yang, and Xiumei Wang. Lightweight image super-resolution with information multistillation network. In ACMMM, 2019. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.32, + 0.892, + 0.36 + ], + "angle": 0, + "content": "[23] Zheng Hui, Xiumei Wang, and Xinbo Gao. Fast and accurate single image super-resolution via information distillation network. In CVPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.362, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[24] Forrest N Iandola, Song Han, Matthew W Moskewicz, Khalid Ashraf, William J Dally, and Kurt Keutzer. SqueezeNet: Alexnet-level accuracy with 50x fewer parameters and; 0.5 mb model size. arXiv:1602.07360, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.419, + 0.892, + 0.46 + ], + "angle": 0, + "content": "[25] Jiwon Kim, Jung Kwon Lee, and Kyoung Mu Lee. Accurate image super-resolution using very deep convolutional networks. In CVPR, 2016. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.461, + 0.892, + 0.502 + ], + "angle": 0, + "content": "[26] Jiwon Kim, Jung Kwon Lee, and Kyoung Mu Lee. Deeplyrecursive convolutional network for image super-resolution. In CVPR, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.504, + 0.892, + 0.532 + ], + "angle": 0, + "content": "[27] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.533, + 0.892, + 0.588 + ], + "angle": 0, + "content": "[28] Fangyuan Kong, Mingxi Li, Songwei Liu, Ding Liu, Jingwen He, Yang Bai, Fangmin Chen, and Lean Fu. Residual local feature network for efficient super-resolution. In CVPR Workshops, 2022. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.59, + 0.892, + 0.643 + ], + "angle": 0, + "content": "[29] Wei-Sheng Lai, Jia-Bin Huang, Narendra Ahuja, and Ming-Hsuan Yang. Fast and accurate image super-resolution with deep laplacian pyramid networks. TPAMI, 41(11):2599-2613, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.646, + 0.892, + 0.701 + ], + "angle": 0, + "content": "[30] Wenbo Li, Xin Lu, Shengju Qian, Jiangbo Lu, Xiangyu Zhang, and Jiaya Jia. On efficient transformer and image pre-training for low-level vision. arXiv preprint arXiv:2112.10175, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.703, + 0.892, + 0.744 + ], + "angle": 0, + "content": "[31] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In ICCV, 2021. 1, 2, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.746, + 0.892, + 0.787 + ], + "angle": 0, + "content": "[32] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In CVPRW, 2017. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.788, + 0.892, + 0.829 + ], + "angle": 0, + "content": "[33] Jie Liu, Jie Tang, and Gangshan Wu. Residual feature distillation network for lightweight image super-resolution. In ECCV, 2020. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.831, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[34] Jie Liu, Wenjie Zhang, Yuting Tang, Jie Tang, and Gangshan Wu. Residual feature aggregation network for image superresolution. In CVPR, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[35] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer:" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "22386" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.092, + 0.469, + 0.12 + ], + "angle": 0, + "content": "Hierarchical vision transformer using shifted windows. In ICCV, 2021. 2, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.469, + 0.148 + ], + "angle": 0, + "content": "[36] Ilya Loshchilov and Frank Hutter. SGDR: stochastic gradient descent with warm restarts. In *ICLR*, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.151, + 0.469, + 0.192 + ], + "angle": 0, + "content": "[37] Zhisheng Lu, Juncheng Li, Hong Liu, Chaoyan Huang, Linlin Zhang, and Tieyong Zeng. Transformer for single image super-resolution. In CVPR Workshops, 2022. 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.194, + 0.469, + 0.247 + ], + "angle": 0, + "content": "[38] Xiaotong Luo, Yuan Xie, Yulun Zhang, Yanyun Qu, Cuihua Li, and Yun Fu. Latticenet: Towards lightweight image super-resolution with lattice block. In ECCV, 2020. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.25, + 0.469, + 0.305 + ], + "angle": 0, + "content": "[39] David Martin, Charless Fowlkes, Doron Tal, and Jitendra Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In ICCV. IEEE, 2001. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.307, + 0.469, + 0.374 + ], + "angle": 0, + "content": "[40] Yusuke Matsui, Kota Ito, Yuji Aramaki, Azuma Fujimoto, Toru Ogawa, Toshihiko Yamasaki, and Kiyoharu Aizawa. Sketch-based manga retrieval using manga109 dataset. Multimedia Tools and Applications, 76(20):21811-21838, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.378, + 0.469, + 0.418 + ], + "angle": 0, + "content": "[41] Sachin Mehta and Mohammad Rastegari. Mobilevit: Lightweight, general-purpose, and mobile-friendly vision transformer. CoRR, abs/2110.02178, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.421, + 0.469, + 0.473 + ], + "angle": 0, + "content": "[42] Abdul Muqueet, Jiwon Hwang, Subin Yang, JungHeum Kang, Yongwoo Kim, and Sung-Ho Bae. Multi-attention based ultra lightweight image super-resolution. In ECCV, 2020. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.477, + 0.469, + 0.531 + ], + "angle": 0, + "content": "[43] Junting Pan, Adrian Bulat, Fuwen Tan, Xiatian Zhu, Lukasz Dudziak, Hongsheng Li, Georgios Tzimiropoulos, and Brais Martínez. Edgevits: Competing light-weight cnns on mobile devices with vision transformers. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.534, + 0.469, + 0.574 + ], + "angle": 0, + "content": "[44] Jongchan Park, Sanghyun Woo, Joon-Young Lee, and In So Kweon. Bam: Bottleneck attention module. arXiv:1807.06514, 2018. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.577, + 0.469, + 0.631 + ], + "angle": 0, + "content": "[45] Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. Automatic differentiation in pytorch. In NeurIPS Workshop, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.633, + 0.469, + 0.7 + ], + "angle": 0, + "content": "[46] Wenzhe Shi, Jose Caballero, Ferenc Huszár, Johannes Totz, Andrew P Aitken, Rob Bishop, Daniel Rueckert, and Zehan Wang. Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In CVPR, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.704, + 0.469, + 0.743 + ], + "angle": 0, + "content": "[47] Ying Tai, Jian Yang, and Xiaoming Liu. Image superresolution via deep recursive residual network. In CVPR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.747, + 0.469, + 0.787 + ], + "angle": 0, + "content": "[48] Ying Tai, Jian Yang, Xiaoming Liu, and Chunyan Xu. Memnet: A persistent memory network for image restoration. In ICCV, 2017. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.469, + 0.843 + ], + "angle": 0, + "content": "[49] Radu Timofte, Eirikur Agustsson, Luc Van Gool, Ming-Hsuan Yang, and Lei Zhang. Ntire 2017 challenge on single image super-resolution: Methods and results. In CVPR workshops, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.899 + ], + "angle": 0, + "content": "[50] Ashish Vaswani, Prajit Ramachandran, Aravind Srinivas, Niki Parmar, Blake Hechtman, and Jonathon Shlens. Scaling local self-attention for parameter efficient visual backbones. In CVPR, 2021. 2, 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[51] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NIPS, 2017. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.892, + 0.175 + ], + "angle": 0, + "content": "[52] Pei Wang, Yijun Li, and Nuno Vasconcelos. Rethinking and improving the robustness of image style transfer. In CVPR, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.22 + ], + "angle": 0, + "content": "[53] Xiaohang Wang, Xuanhong Chen, Bingbing Ni, Zhengyan Tong, and Hang Wang. Learning continuous depth representation via geometric spatial aggregator. In AAAI, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.222, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[54] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.278, + 0.892, + 0.317 + ], + "angle": 0, + "content": "[55] Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon. Cbam: Convolutional block attention module. In ECCV, 2018. 3, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.321, + 0.892, + 0.36 + ], + "angle": 0, + "content": "[56] Tete Xiao, Mannat Singh, Eric Mintun, Trevor Darrell, Piotr Dólar, and Ross B. Girshick. Early convolutions help transformers see better. In NeurIPS, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.364, + 0.892, + 0.403 + ], + "angle": 0, + "content": "[57] Chenglin Yang, Yilin Wang, Jianming Zhang, He Zhang, Zijun Wei, Zhe Lin, and Alan L. Yuille. Lite vision transformer with enhanced self-attention. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.407, + 0.892, + 0.432 + ], + "angle": 0, + "content": "[58] Jianwei Yang, Chunyuan Li, Xiyang Dai, and Jianfeng Gao. Focal modulation networks, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.435, + 0.892, + 0.489 + ], + "angle": 0, + "content": "[59] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In CVPR, 2022. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.492, + 0.892, + 0.546 + ], + "angle": 0, + "content": "[60] Roman Zeyde, Michael Elad, and Matan Protter. On single image scale-up using sparse-representations. In International conference on curves and surfaces, pages 711-730. Springer, 2010. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.549, + 0.892, + 0.588 + ], + "angle": 0, + "content": "[61] Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In ICLR, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.591, + 0.892, + 0.631 + ], + "angle": 0, + "content": "[62] Haokui Zhang, Wenze Hu, and Xiaoyu Wang. Parc-net: Position aware circular convolution with merits from convnets and transformer. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.634, + 0.892, + 0.673 + ], + "angle": 0, + "content": "[63] Kai Zhang, Wangmeng Zuo, Shuhang Gu, and Lei Zhang. Learning deep cnn denoiser prior for image restoration. In CVPR, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.677, + 0.892, + 0.717 + ], + "angle": 0, + "content": "[64] Kai Zhang, Wangmeng Zuo, and Lei Zhang. Learning a single convolutional super-resolution network for multiple degradations. In CVPR, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.72, + 0.892, + 0.76 + ], + "angle": 0, + "content": "[65] Yiman Zhang, Hanting Chen, Xinghao Chen, Yiping Deng, Chunjing Xu, and Yunhe Wang. Data-free knowledge distillation for image super-resolution. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.763, + 0.892, + 0.802 + ], + "angle": 0, + "content": "[66] Yulun Zhang, Kunpeng Li, Kai Li, Lichen Wang, Bineng Zhong, and Yun Fu. Image super-resolution using very deep residual channel attention networks. In ECCV, 2018. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.805, + 0.892, + 0.845 + ], + "angle": 0, + "content": "[67] Yulun Zhang, Yapeng Tian, Yu Kong, Bineng Zhong, and Yun Fu. Residual dense network for image super-resolution. In CVPR, 2018. 5" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.845 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22387" + } + ] +] \ No newline at end of file diff --git a/2023/Omni Aggregation Networks for Lightweight Image Super-Resolution/39175e81-b86e-49a9-967e-e590c3dbdaa0_origin.pdf b/2023/Omni Aggregation Networks for Lightweight Image Super-Resolution/39175e81-b86e-49a9-967e-e590c3dbdaa0_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..3fcbaea6a0b7187a5b4eb0c8f78be23ebeecfc23 --- /dev/null +++ b/2023/Omni Aggregation Networks for Lightweight Image Super-Resolution/39175e81-b86e-49a9-967e-e590c3dbdaa0_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c7b56c3d5b2daa1a00e21fbc2da33a03189b5654a7a4a59d211ba84e10a38219 +size 7844837 diff --git a/2023/Omni Aggregation Networks for Lightweight Image Super-Resolution/full.md b/2023/Omni Aggregation Networks for Lightweight Image Super-Resolution/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b02c75be8d8c2c02231acb9b297c254885f383b4 --- /dev/null +++ b/2023/Omni Aggregation Networks for Lightweight Image Super-Resolution/full.md @@ -0,0 +1,428 @@ +# Omni Aggregation Networks for Lightweight Image Super-Resolution + +Hang Wang $^{2*}$ Xuanhong Chen $^{1*}$ Bingbing Ni $^{1,2\dagger}$ Yutian Liu $^{1}$ Jinfan Liu $^{1}$ + +$^{1}$ Shanghai Jiao Tong University, Shanghai 200240, China $^{2}$ Huawei + +francis970625@gmail.com {chen19910528,nibingbing,stau7001,ljflnjz}@sjtu.edu.cn + +# Abstract + +While lightweight ViT framework has made tremendous progress in image super-resolution, its uni-dimensional self-attention modeling, as well as homogeneous aggregation scheme, limit its effective receptive field (ERF) to include more comprehensive interactions from both spatial and channel dimensions. To tackle these drawbacks, this work proposes two enhanced components under a new Omni-SR architecture. First, an Omni Self-Attention (OSA) block is proposed based on dense interaction principle, which can simultaneously model pixel-interaction from both spatial and channel dimensions, mining the potential correlations across omni-axis (i.e., spatial and channel). Coupling with mainstream window partitioning strategies, OSA can achieve superior performance with compelling computational budgets. Second, a multi-scale interaction scheme is proposed to mitigate sub-optimal ERF (i.e., premature saturation) in shallow models, which facilitates local propagation and meso-/global-scale interactions, rendering an omni-scale aggregation building block. Extensive experiments demonstrate that Omni-SR achieves record-high performance on lightweight super-resolution benchmarks (e.g., 26.95dB@Urban100 ×4 with only 792K parameters). Our code is available at https://github.com/Francis0625/Omni-SR. + +# 1. Introduction + +Image super-resolution (SR) is a long-standing low-level problem that aims to recover high-resolution (HR) images from degraded low-resolution (LR) inputs. Recently, vision transformer [14, 51] based (i.e., ViT-based) SR frameworks [5, 31] have emerged, showing significant performance gains compared to previously dominant Convolutional Neural Networks (CNNs) [66]. However, most attempts [31] are devoted to improving the large-scale ViT-based models, while the development of lightweight ViTs (typically, less than 1M parameters) remains fraught with + +![](images/83843f4b5574f4fcbc3b18bdae706065a44d808b3483f1033fb43f0367b2a5b6.jpg) +Spatial SA + +![](images/71a6f804e8ab1375f1e3508b309d41565c74a41c0618e0ff295d86651f9eb999.jpg) +Channel SA + +![](images/8086c58f33b0a6019dc2e25b33792325b411b68b2bf8b57692f6c1666aee8eb1.jpg) +Omni SA + +![](images/d30ea2a4d9bfa5b22bdcffc7e2e332265caf4a9f549ff90fd94720220b464bbc.jpg) +PSNR + +![](images/af7413c8721fbc957a54d747d3e3cbb64dfd53dd21045b42dff47fc63bd5563e.jpg) +Parameters (K) + +![](images/b10e420f7e0a0818b6694462ea8c62a3c345ce4bdc0ae41e0f367c9f1b9849e6.jpg) +Iteration +Figure 1. Typical self-attention schemes [31,59] can only perform uni-dimensional (e.g., spatial-only) interactions, and. + +difficulties. This paper focuses on boosting the restoration performance of lightweight ViT-based frameworks. + +Two difficulties hinder the development of lightweight ViT-based models: 1) Uni-dimensional aggregation operators (i.e., spatial [31] only or channel [59] only) imprisons the full potential of self-attention operators. Contemporary self-attention generally realizes the interaction between pixels by calculating the cross-covariance of the spatial direction (i.e., width and height) and exchanges context information in a channel-separated manner. This interaction scheme ignores the explicit use of channel information. However, recent evidences [59] and our practice show that self-attention in the channel dimension (i.e., computationally more compact than spatial self-attention) is also crucial in low-level tasks. 2) Homogeneous aggregation schemes (i.e., Simple hierarchical stacking of single operators, e.g., convolution, self-attention) neglect abundant texture patterns of multi-scales, which is urgently needed in SR task. Specifically, a single operator is only sensitive to information of one scale [6, 12], e.g., self-attention is sensitive to long-term information and pays little attention to lo + +cal information. Additionally, stacking of homogeneous operators proves to be inefficient and suffers from premature saturation of the interaction range [8], which is reflected as a suboptimal effective receptive field. The above problem is exacerbated in lightweight models because lightweight models cannot stack enough layers. + +In order to solve the above problems and pursue higher performance, this work proposes a novel omni-dimension feature aggregation scheme called Omni Self-Attention (OSA) exploiting both spatial and channel axis information in a simultaneous manner (i.e., extends the interaction into three-dimensional space), which offers higher-order receptive field information, as shown in Figure 1. Unlike scalar-based (a group of important coefficient) channel interaction [19], OSA enables comprehensive information propagation and interaction by cascading computation of the cross-covariance matrices between spatial/channel dimensions. The proposed OSA module can be plugged into any mainstream self-attention variants (e.g., Swin [35], Halo [50]), which provides a finer granularity of important encoding (compared to the vanilla channel attention [19]), achieving a perceptible improvement in contextual aggregation capabilities. Furthermore, a multi-scale hierarchical aggregation block, named Omni-Scale Aggregation Group (i.e., OSAG for short), is presented to achieve tailored encoding of varying scales of texture patterns. Specifically, OSAG builds three cascaded aggregators: local convolution (for local details), meso self-attention (focusing on mid-scale pattern processing), and global self-attention (pursuing global context understanding), rendering an omni-scale (i.e., local-/meso-/global-scale simultaneously) feature extraction capability. Compared to the homogenized feature extraction schemes [28, 31], our OSAG is able to mine richer information producing features with higher information entropy. Coupling with the above two designs, we establish a new ViT-based framework for lightweight superresolution, called Omni-SR, which exhibits superior restoration performance as well as covers a larger interaction range while maintaining an attractive model size, i.e., $792K$ . + +We extensively experiment with the proposed framework with both qualitative and quantitative evaluations on mainstream open-source image super-resolution datasets. It is demonstrated that our framework achieves state-of-the-art performance at the lightweight model scale (e.g., $\text{Urban100} \times 4: 26.95\mathrm{dB}$ , $\text{Manga109} \times 4: 31.50\mathrm{dB}$ ). More importantly, compared to existing ViT-based super-solution frameworks, our framework shows superior optimization properties (e.g., convergence speed, smoother loss landscape), which endow our model with better robustness. + +# 2. Related Works + +Image Super-resolution. CNNs have achieved remarkable success in image SR task. SRCNN [13] is the first + +![](images/0a11c3f9b6aa9678b0a36037821e2d50d310e88c6867c243403df9e178d054ae.jpg) +Figure 2. Illustration of spatial attention and channel attention. These typical attention paradigms only model uni-dimensional (i.e., spatial-only / channel-only) interaction. + +![](images/da9fa7abd21172944f7bbcafab2e7f8521fdb44f49ab186ce3530fcde058bbce.jpg) + +work to introduce CNNs into the SR field. Many methods [25, 48, 66] employ skip connection to speed up network convergence and improve the reconstruction quality. Channel Attention [66] is also proposed to enhance the representation ability of the SR model. In order to obtain better reconstruction quality with limited computing resources, several methods [23,38,42,47] explore lightweight architectural design. DRCN [26] utilizes the recursive operation to reduce the number of parameters. DRRN [47] introduces global and local residual learning on the basis of DRCN to accelerate training and improve the quality of details. CARN [1] employs cascading mechanism upon a residual network. IMDN [22] proposes an information multistillation block to archive better time performance. Another line of research is to utilize model compression techniques, e.g., knowledge distillation [15, 17, 65] and neural architecture search [11]) to reduce computing costs. Recently, a series of transformer-based SR models [5,8,31,37] emerge with superior performance. Chenet al. [5] develop a pre-trained model for the low-level computer vision task using the transformer architecture. Based on Swin transformer [35], SwinIR [31] proposes a three-stage framework, refreshing the state-of-the-art of SR task. More recently, some works [5,30] explore ImageNet pre-training strategy to further enhance SR performance. + +Lightweight Vision Transformer. Due to the urgent demands for applying networks to resource-constrained devices, lightweight vision transformer [14, 51] has attracted widespread attention. Many attempts [7,9,10,37,41,43,57, 62] have been made to develop lightweight ViTs with comparable performance. A series of methods focus on combining convolutions with transformers to learn both local and global representations. For instance, LVT [57] introduces convolution in self-attention to enrich low-level features. MobileViT [41] replaces matrix multiplication in convolutions with transformer layers to learn global representations. Similarly, EdgeViTs [43] employs an information exchange bottleneck for full spatial interactions. Different from interpreting convolutions into vision transformers, LightViT [21] proposes aggregated self-attention for better + +![](images/118b879237fe11c15116324502a94b71f3b9cb241e4e3748c0ae1532450ec6db.jpg) +Figure 3. The overall architecture of the proposed Omni-SR framework and structure of OSAG and Omni Self-Attention (OSA). + +information aggregation. In this work, we resort to ViT architecture to achieve lightweight and accurate SR. + +# 3. Methodology + +# 3.1. Attention Mechanisms in Super-Resolution + +Two attention paradigms are widely adopted in SR to assist in analyzing and aggregating comprehensive patterns. + +Spatial Attention. Spatial attention can be regarded as an anisotropic selection process. Spatial self-attention [37, 51] and spatial gate [10, 58] are predominantly applied. As shown in Figure 2, spatial self-attention calculates the cross-covariance along the spatial dimension, and the spatial gate generates the channel-separated masks. Neither of them can transmit information between channels. + +Channel Attention. There are two categories of channel attention, i.e., scalar-based [19] and covariance-based [59], proposed to perform channel recalibration or transmit patterns among channels. As shown in Figure 2, the former predicts a group of importance scalars to weigh different channels, while the latter computes a cross-covariance matrix to enable channel re-weighting and information transmission simultaneously. Compared to spatial attention, channel attention handles spatial dimension isotropically, and thus, the complexity is significantly reduced, which also impairs the accuracy of aggregation. + +Several attempts [44,55] have proved that both spatial attention and channel attention are beneficial for SR task and their characteristics are complementary to each other, thus integrating them in a computationally compact way would bring notable benefits in expressive capability. + +# 3.2. Omni Self-Attention Block + +To mine all the correlations hidden in the latent variables, we propose a novel self-attention paradigm called Omni Self-Attention (OSA) block. Unlike existing self-attention paradigms (e.g., spatial self-attention [5, 37, 51]) that only indulge in unidimensional processing, OSA establishes the spatial and channel context simultaneously. The obtained two-dimensional relationship is highly necessary and beneficial, especially for lightweight models. On the one hand, as the network deepens, important information is scattered into different channels [19], and it is critical to deal with them in a timely manner. On the other hand, although spatial self-attention takes advantage of the channel dimension in calculating the covariance, it does not transmit the information between channels (refer to Sec. 3.1). Given the above conditions, our OSA is designed to transmit both spatial and dimensional information in a compact manner. + +The proposed OSA calculates the score matrices corresponding to the space and channel direction through sequential matrix operations and rotation, as illustrated in Figure 3. Specifically, suppose $X \in \mathbb{R}^{HW \times C}$ denotes the input feature, where $H$ and $W$ are the width and height of the input, and $C$ is the channel number. Firstly, $X$ is embedded to query, key and value matrices $Q^s, K^s, V^s \in \mathbb{R}^{HW \times C}$ through linearly projection. We calculate the production of query and key to get the spatial attention map of size $\mathbb{R}^{HW \times HW}$ . Then we perform the spatial attention to obtain the intermediate aggregated results. Note that window strategy is usually used to significantly reduce the resource overhead. Next stage, we rotate the input query and key matrices to get the transposed query and key matrices $Q^c, K^c \in \mathbb{R}^{C \times HW}$ , and also rotate the value matrices to get the value matrix $V^c \in \mathbb{R}^{C \times HW}$ for the subsequent channel-wise self-attention. The obtained channel-wise attention map of size $\mathbb{R}^{C \times C}$ models channel-wise relation + +ships. Finally, we get the final aggregated $Y_{OSA}$ by the inverse rotation of the channel attention output $Y_{c}$ . The whole OSA process is formulated as follows: + +$$ +Q ^ {s} = X \cdot \mathcal {W} _ {q}, K ^ {s} = X \cdot \mathcal {W} _ {k}, V ^ {s} = X \cdot \mathcal {W} _ {v}, \tag {1} +$$ + +$$ +Y _ {s} = \mathcal {A} ^ {s} \left(Q ^ {s}, K ^ {s}, V ^ {s}\right) = \operatorname {S o f t M a x} \left(Q ^ {s} K ^ {s T}\right) \cdot V ^ {s}, \tag {2} +$$ + +$$ +Q ^ {c} = \mathcal {R} \left(Q ^ {\prime}\right), \quad K ^ {c} = \mathcal {R} \left(K ^ {\prime}\right), \quad V ^ {c} = \mathcal {R} \left(V ^ {\prime}\right), \tag {3} +$$ + +$$ +Y _ {c} = \mathcal {A} ^ {c} \left(Q ^ {c}, K ^ {c}, V ^ {c}\right) = \operatorname {S o f t M a x} \left(K ^ {c} Q ^ {c T}\right) \cdot V ^ {c}, \tag {4} +$$ + +$$ +Y _ {O S A} = \mathcal {R} ^ {- 1} \left(Y _ {c}\right), \tag {5} +$$ + +where $\mathcal{W}_q, \mathcal{W}_k, \mathcal{W}_v$ denote the linear projection matrices for the query, key, and value, respectively. $Q', K', V'$ are the input embedding matrices of channel-wise self-attention, which are embedded from fore spatial self-attention or or copied directly from $Q^s, K^s, V^s$ . $\mathcal{R}(\cdot)$ denotes the rotation operation around spatial axis and $\mathcal{R}^{-1}(\cdot)$ is the inverse rotation. Some normalization factors are omitted for the sake of simplicity. In particular, this design shows compelling properties that can integrate the elementwise results of two matrix operations (i.e., spatial-/channel-matrix operation), thereby enabling omni-axial interactions. Note that our proposed OSA paradigm can be a drop-in replacement of the Swin [31,35] attention block to higher performance with less parameters. Benefiting from the smaller attention map size of channel self-attention, the proposed OSA is less computationally intensive compared to the cascade shifted-window self-attention scheme in Swin. + +Discussion with other hybrid attention paradigms. Compared to previous hybrid channel and spatial attention works like CBAM [55] and BAM [44], their scalar-based attention weights only reflect the relative degree of importance, without further inter-pixel information exchange, leading to limited relation modeling capability. Several recent works [8] also incorporate channel attention with spatial Self-attention, but these attempts only resort to scalar weights for channel recalibration, while our OSA paradigm enables channel-wise interaction to mine the potential correlations in omni-axis. Performance comparison of different attention paradigms can be found in Sec. 4.4. + +# 3.3. Omni-Scale Aggregation Group + +How to utilize the proposed OSA paradigm to build a high-performance and compact network is another key topic. Although hierarchical stacking of windows-based self-attention (e.g., swin [31, 35]) has become mainstream, various works have found that the window-based paradigms are very inefficient for large-range interactions, especially for shallow networks. It is worth pointing out that large-range interaction can provide a pleasing effective receptive field, which is crucial for improving image restoration performance [37]. Unfortunately, direct global interaction is resource-prohibitive and detracts from local aggregation capabilities. Taking these points into account, we propose + +![](images/daea26868e5c66071cd70cbfa3bee43f8f0b385a1022c93322afdc79d86fd7e6.jpg) +Figure 4. Illustration of omni-scale aggregation scheme. Our proposed Omni-SR contains three types of feature aggregation at local level, meso and global level, respectively. + +an Omni-Scale Aggregation Group (i.e., OSAG for short) to pursue progressive receptive field feature aggregation with low computational complexity. As shown in Figure 3, OSAG mainly consists of three stages: local, meso and global aggregations. Specifically, a channel attention [19] enhanced inverted bottleneck [18] is introduced to fulfill the local pattern process with limited overhead. Based on the proposed OSA paradigm, we derive two instances (i.e., Meso-OSA and Global-OSA) responsible for the interaction and aggregation of meso and global information. Note that the proposed omni self-attention paradigm can be used for different purposes. Meso-OSA performs attention within a group of non-overlap patches, which restricts Meso-OSA to only focus on meso-scale pattern understanding. Global-OSA samples data point sparsely across the entire feature within an atrous manner, endowing Global-OSA with the ability to achieve global interactions at a compelling cost. + +The only difference between Meso-OSA and Global-OSA is the window partition strategy, as shown in Figure 4. In order to achieve meso-scale interaction, Meso-OSA split the input feature $X$ into non-overlapping blocks with size $P \times P$ . Note that after window partition, the block dimensions are gathered onto the spatial dimension (i.e., -2 axis): $(H, W, C) \to \left(\frac{H}{P} \times P, \frac{W}{P} \times P, C\right) \to \left(\frac{HW}{P^2}, P^2, C\right)$ . While the Global-OSA divides the input feature into a uniform $G \times G$ grid, with each lattice having an adaptive size of $\frac{H}{G} \times \frac{W}{G}$ . Similar to Meso-OSA, the grid dimension is also gathered on the spatial axis (i.e., -2 axis): $(H, W, C) \to (G \times \frac{H}{G}, G \times \frac{W}{G}, C) \to (G^2, \frac{HW}{G^2}, C) \to (\frac{HW}{G^2}, G^2, C)$ . + +# 3.4. Network Architecture + +Overall Structure. Based on the Omni Self-Attention paradigm and the Omni-Scale Aggregation Group, we further develop a lightweight Omni-SR framework to achieve high-performance image super-resolution. As shown in Figure 3, Omni-SR consists of three parts, i.e., shallow feature extraction, deep feature extraction, and image reconstruction. Specifically, given the LR input $I_{LR} \in \mathbb{R}^{H \times W \times C_{in}}$ , we first use a $3 \times 3$ convolution $H_{SF}$ to extract shallow feature $X_0 \in \mathbb{R}^{H \times W \times C}$ as + +$$ +X _ {0} = H _ {S F} \left(I _ {L R}\right), \tag {6} +$$ + +where $C_{in}$ and $C$ denote the channel number of the input and shallow feature. The convolution layer provides a simple way to convert the input from image space into + +high-dimensional feature space. Then we use $K$ stacked omni-scale aggregation groups (OSAG) and one $3 \times 3$ convolution layer $H_{CONV}$ in a cascade manner to extract deep feature $F_{DF}$ . Such a process can be expressed as + +$$ +X _ {i} = H _ {O S A G _ {i}} \left(X _ {i - 1}\right), \quad i = 1, 2, \dots , K, \tag {7} +$$ + +$$ +X _ {D F} = H _ {C O N V} \left(X _ {K}\right), +$$ + +where $H_{\text{OSAG}_i}$ represents the $i$ -th OSAG, $X_1, X_2, \ldots, X_K$ denote intermediate features. Following [31], we also use a convolutional layer at the end of feature extraction to get better feature aggregation. Finally we reconstruct the HR image $I_{HR}$ by aggregating shallow and deep features as + +$$ +I _ {H R} = H _ {R e c} \left(X _ {0} + X _ {D F}\right), \tag {8} +$$ + +where $H_{Rec}(\cdot)$ denotes the reconstruction module. In detail, PixelShuffle [46] is used to up-sample the fused feature. + +Omni-Scale Aggregation Group (OSAG). As shown in Figure 3, each OSAG contains a local convolution block (LCB), a meso-OSA block, a global-OSA block, and an ESA block [28, 34]. The whole process can be formulated as + +$$ +X _ {r e s} = H _ {\text {G l o b a l - O S A B} _ {i}} \left(H _ {\text {M e s o - O S A B} _ {i}} \left(H _ {\text {L C B} _ {i}} \left(X _ {i - 1}\right)\right)\right), \tag {9} +$$ + +$$ +X _ {i} = H _ {E S A _ {i}} \left(H _ {\text {C o n v} _ {i}} \left(X _ {\text {r e s}} + X _ {i - 1}\right)\right), \tag {10} +$$ + +where $X_{i-1}$ and $X_i$ represents the input and output feature of $i$ -th OSAG. After the mapping of convolution layers, we insert a Meso-OSAB for window-based self-attention and a Global-OSAB to enlarge the receptive field for better information aggregation. At the end of OSAG, we reserve the convolutional layer and ESA block following [28, 66]. + +In specific, LCB is implemented as a stack of pointwise and depthwise convolutions with a CA module [24] between them to adaptively re-weight channel-wise features. This block aims to aggregate local contextual information as well as to increase the trainability of the network [56]. Two types of OSA blocks (i.e., Meso-OSA block and Global-OSA block) are then followed to obtain interactions from different regions. Based on different window partition strategies, Meso-OSA block seeks inner-block interaction, and Global-OSA blocks aim for global mixing. OSA blocks follow typical Transformers designs with Feedforward network (FFN) and LayerNorm [2], and the only difference is that the origin self-attention operation is replaced with our proposed OSA operator. For FFN, we adopt the GDFN proposed by Restormer [59]. Combining these individuals seamlessly, the designed OSAG enables information propagation between any pair of tokens in the feature map. We use the ESA module proposed in [28, 34] to further refine the fused feature. + +Optimization Objective. Following prior works [31, 32, 53, 67], we train the model by minimizing a standard $L_{1}$ + +loss between model prediction $\hat{I}_{HR}$ and HR label $I_{HR}$ as follows: + +$$ +\mathcal {L} = \left\| I _ {H R} - \hat {I} _ {H R} \right\| _ {1}. \tag {11} +$$ + +# 4. Experiments + +# 4.1. Experimental Setup + +Datasets and Metrics. Following previous work [31,32, 38,49,66], DIV2K [49] and Flickr2K [49] are used as training datasets. For a fair comparison, we employ two training protocols, i.e., training with DIV2K only and training with DF2K (DIV2K + Flickr2K). Note that the model trained with DF2K is marked with small $\dagger$ . For testing, we adopt five standard benchmark datasets: Set5 [4], Set14 [60], B100 [39], Urban100 [20] and Manga109 [40]. PSNR and SSIM [54] are adopted to evaluate the SR performance on the Y channel of the transformed YCbCr space. + +Implementation Details. During training, we augment the data with random horizontal flips and 90/270-degree rotations. LR images are generated by bicubic downsampling [63] from HR images. OSAG number is set to 5, and channel number of the whole network is set to 64. The attention head number and window size are set to 4 and 8 for both Meso-OSAB and Global-OSAB. We use Adam [27] optimizer to train the model with a batch size of 64 for $800\mathrm{K}$ iterations. The initial learning rate is set to $10^{-4}$ and halved for every $200\mathrm{k}$ iterations. In each training batch, we randomly crop LR patches of size $64\times 64$ as input. Our method is implemented with PyTorch [45], and all experiments are conducted on one NVIDIA V100 GPU. Note that no other data augmentation (e.g., Mixup [61], RGB channel shuffle) or training skills (e.g., pre-training [30], cosine learning schedule [36]) are employed. It should be pointed out that we maintain the consistency of model parameters in the ablation study by adjusting the channels of $1\times 1$ convolution. + +# 4.2. Comparison with the SOTA SR methods + +To evaluate the effectiveness of Omni-SR, we compare our model with several advanced lightweight SR methods under a scale factor of 2/3/4. In particular, former works, VDSR [25], CARN [1], IMDN [22], EDSR [32], RFDN [33], MemNet [48], MAFFSRN [42], LatticeNet [38], RLFN [28], ESRT [37] and SwinIR [31] are introduced for comparison. + +Quantitative results. In Table 1, the quantitative comparisons of different lightweight methods are presented on five benchmark datasets. With a similar model size, the performance of our Omni-SR surpasses existing methods with a notable margin on all benchmarks. In particular, compared to other transformer architectures with comparable parameters like SwinIR [31] and ESRT [37], the proposed Omni-SR obtains the best performance. The results + +Table 1. Quantitative comparison (PSNR/SSIM) for lightweight image SR with state-of-the-art methods on benchmark datasets. The best and second-best results are marked in red and blue colors, respectively. “†” indicates that model is trained on DF2K. + +
MethodYearsScaleParams(K)Set5 PSNR/ SSIMSet14 PSNR/ SSIMBSD100 PSNR/ SSIMUrban100 PSNR/ SSIMManga109 PSNR/ SSIM
VDSR [25]CVPR1666636.66 / 0.954233.05 / 0.912731.90 / 0.896030.76 / 0.914037.22 / 0.9750
MemNet [48]ICCV1767837.78 / 0.959733.28 / 0.914232.08 / 0.897831.31 / 0.919537.72 / 0.9740
SRMDNF [64]CVPR18151137.79 / 0.96033.32 / 0.91532.05 / 0.898531.33 / 0.920438.07 / 0.9761
CARN [1]ECCV181,59237.76 / 0.959033.52 / 0.916632.09 / 0.897831.92 / 0.925638.36 / 0.9765
IMDN [22]MM1969438.00 / 0.960533.63 / 0.917732.19 / 0.899632.17 / 0.928338.88 / 0.9774
RFDN-L [33]ECCV2062638.08 / 0.960633.67 / 0.919032.18 / 0.899632.24 / 0.929038.95 / 0.9773
MAFFSRN [42]ECCV20×240237.97 / 0.960333.49 / 0.917032.14 / 0.899431.96 / 0.9268-/-
LatticeNet [38]ECCV2075638.15 / 0.961033.78 / 0.919332.25 / 0.900532.43 / 0.9302-/-
RLFN [28]CVPRW2252738.07 / 0.960733.72 / 0.918732.22 / 0.900032.33 / 0.9299-/-
SwinIR [31]ICCVW2187838.14 / 0.961133.86 / 0.920632.31 / 0.901232.76 / 0.934039.12 / 0.9783
Omni-SROurs77238.22 / 0.961333.98 / 0.921032.36 / 0.902033.05 / 0.936339.28 / 0.9784
Omni-SR†Ours77238.29 / 0.961734.27 / 0.923832.41 / 0.902633.30 / 0.938639.53 / 0.9792
VDSR [25]CVPR1666633.66 / 0.921329.77 / 0.831428.82 / 0.797627.14 / 0.827932.01 / 0.9340
MemNet [48]ICCV1767834.09 / 0.924830.00 / 0.835028.96 / 0.800127.56 / 0.837632.51 / 0.9369
EDSR [32]CVPRW171,55534.37 / 0.927030.28 / 0.841729.09 / 0.805228.15 / 0.852733.45 / 0.9439
SRMDNF [64]CVPR181,52834.12 / 0.925430.04 / 0.838228.97 / 0.802527.57 / 0.839833.00 / 0.9403
CARN [1]ECCV181,59234.29 / 0.925530.29 / 0.840729.06 / 0.803428.06 / 0.849333.50 / 0.9440
IMDN [22]MM1970334.36 / 0.927030.32 / 0.841729.09 / 0.804628.17 / 0.851933.61 / 0.9445
RFDN-L [33]ECCV20×363334.47 / 0.928030.35 / 0.842129.11 / 0.805328.32 / 0.854733.78 / 0.9458
MAFFSRN [42]ECCV2080734.45 / 0.927730.40 / 0.843229.13 / 0.806128.26 / 0.8552-/-
LatticeNet [38]ECCV2076534.53 / 0.928130.39 / 0.842429.15 / 0.805928.33 / 0.8538-/-
ESRT [37]CVPRW2277034.42 / 0.926830.43 / 0.843329.15 / 0.806328.46 / 0.857433.95 / 0.9455
SwinIR [31]ICCVW2188634.62 / 0.928930.54 / 0.846329.20 / 0.808228.66 / 0.862433.98 / 0.9478
Omni-SROurs78034.70 / 0.929430.57 / 0.846929.28 / 0.809428.84 / 0.865634.22 / 0.9487
Omni-SR†Ours78034.77 / 0.930430.70 / 0.848929.33 / 0.811129.12 / 0.871234.64 / 0.9507
VDSR [25]CVPR1666631.35 / 0.883828.01 / 0.767427.29 / 0.725125.18 / 0.752428.83 / 0.8870
MemNet [48]ICCV1767831.74 / 0.889328.26 / 0.772327.40 / 0.728125.50 / 0.763029.42 / 0.8942
EDSR [32]CVPRW171,51832.09 / 0.893828.58 / 0.781327.57 / 0.735726.04 / 0.784930.35 / 0.9067
SRMDNF [64]CVPR181,55231.96 / 0.892528.35 / 0.778727.49 / 0.733725.68 / 0.773130.09 / 0.9024
CARN [1]ECCV181,59232.13 / 0.893728.60 / 0.780627.58 / 0.734926.07 / 0.783730.47 / 0.9084
IMDN [22]MM1971532.21 / 0.894828.58 / 0.781127.56 / 0.735326.04 / 0.783830.45 / 0.9075
RFDN-L [33]ECCV20×464332.28 / 0.895728.61 / 0.781827.58 / 0.736326.20 / 0.788330.61 / 0.9096
MAFFSRN [42]ECCV2083032.20 / 0.895326.62 / 0.782227.59 / 0.737026.16 / 0.7887-/-
LatticeNet [38]ECCV2077732.30 / 0.896228.68 / 0.783027.62 / 0.736726.25 / 0.7873-/-
ESRT [37]CVPRW2254332.24 / 0.895228.62 / 0.781327.60 / 0.736426.17 / 0.7877-/-
SwinIR [31]ICCVW2175132.19 / 0.894728.69 / 0.783327.69 / 0.737926.39 / 0.796230.75 / 0.9100
Omni-SROurs89732.44 / 0.897628.77 / 0.785827.69 / 0.740626.47 / 0.798030.92 / 0.9151
Omni-SR†Ours79232.49 / 0.898828.78 / 0.785927.71 / 0.741526.64 / 0.801831.02 / 0.9151
Omni-SR†Ours79232.57 / 0.899328.95 / 0.789827.81 / 0.743926.95 / 0.810531.50 / 0.9192
+ +![](images/893f81de995407df99654d0b38795a6f908c601bfb716c4c33644406bacbe803.jpg) +(a) + +![](images/ffe582007a99f84656e3b6d885cdc2a6f8a9cf27362c2e7952528c61d4746ecc.jpg) +(b) +Figure 5. (a) PSNR improvement of Omni-SR variants with different OSAG number (K) over the smallest Omni-SR model $(\mathrm{K} = 1)$ for $4\times$ SR. (b) The number of model parameters vs. PSNR of different lightweight methods on Urban100 dataset for $4\times$ SR. + +showcase that the omni-axis (i.e., spatial + channel) interaction introduced by OSA can effectively boost the model's contextual aggregation capabilities, which promises superior SR performance. Coupling with large training dataset DF2K, the performance can be further improved, especially on Urban100. We suppose that such a phenomenon can be ascribed to the images in Urban100 have many similar patches, and the long-term relationship introduced by + +OSAG can bring great benefits for detail restoration. More importantly, with similar parameters, our model reduces $28\%$ of computational complexity (Omni-SR: 36G FLOPs vs. SwinIR: 50G FLOPs @1280×720), showing its effectiveness and efficiency. + +Visual comparison. In Figure 6, we also provide a visual comparison of different lightweight SR methods at $\times 4$ scale. We can observe that the HR images constructed by Omni-SR contain more fine-grained details, while other methods generate blurred edges or artifacts in complicated areas. For instance, in the $1st$ row, our model is able to restore the detailed texture of the wall pleasantly, which all other methods fail to restore. Visual results also validate the effectiveness of the proposed OSA paradigm, which can perform omni-axis pixel-wise interaction modeling, thus obtaining a more powerful reconstruction capability. + +Trade-off between Model Size and Performance. In experiments, we set the number of OSAG as 5 to make the model size around $800\mathrm{K}$ for a fair comparison with other methods. We also explore model performance with smaller parameter sizes by reducing OSAG number K. As shown in Figure 5(a), compared to the smallest model variant with + +![](images/94d802cd6418ce37365ce9e96117e6ce93556852e979423c7d4d4fb7695d8b22.jpg) +Urban100: img_012.png + +![](images/494a73f50393065708fb86e2beeeeef07544713ef02e318bd799cdb4823f802c.jpg) +PSNR Reference + +![](images/401fb73e79fac28b9e0da2dd38d4647fc73da9a39f897c70b0fdd43f2fa8f301.jpg) +20.65 dB Bicubic + +![](images/883f3e8878b3dc8f6e6c72dd84f3a645091e2d54afd68c28b67c0d9ed14f3e28.jpg) +21.15 dB SRCNN [13] + +![](images/1e5a0907028261a23d10f724de5d8d804a90b7dacce7856eb02735db56b204df.jpg) +21.22 dB FSRCNN [29] + +![](images/f89c955c7a4e3a57046090ed64ecaaa0e21b351a1926ff23b69312fe596e0792.jpg) +21.99 dB CARN [1] + +![](images/77836fefd440f40b36f59a49c94431d7ba2877842db8df27ee2c56c043a021fb.jpg) +Urban100: img_024.png + +![](images/ab6c47781207dc279b26c766c0b3d0a8968f71e8642f553ef7451110331dd877.jpg) +21.95 dB IMDN [22] + +![](images/02a5fb9235cbec0479716589693eedc88d20f1ca68ad0b747f0d30bf586135d2.jpg) +21.95 dB RLFN [28] + +![](images/89cf7d445d853fc9f318cfb4b06feb57bff736ef54700ddc08adc1ded5c7cd1e.jpg) +22.00 dB LatticeNet [38] + +![](images/dafc61eeee66ffca2f3a71457045dd04837b4e46ef0c3ffedab0f4bbc58020fa.jpg) +22.34 dB SwinIR [31] + +![](images/4ec3907b3e3aa4f900e3e98c6906c9e62be75f06cb6569e71ef5d6bfb2cf891c.jpg) +24.16 dB Omni-SR (Ours) + +![](images/6dcc1cf12baab7f8aa1c5bf8c289d536f1246b2712c317b68b08601e653fe82f.jpg) +Set14: img_barbara.png +Figure 6. Visual comparison for $\times 4$ SR methods. The patches for comparison are marked with red boxes. (Best viewed by zooming.) + +![](images/269c2cee48363e525646d0ac24f793ad38cf655e6d1b9a76a24d3fec01f95687.jpg) + +![](images/162f7a8fc7d6f7d23f09a1207931df688e23ed0f0aeba914e85f887459ec038f.jpg) +PSNR Reference + +![](images/0e0f08caf93a4c93c799f7d427d119f3a1b87b2554d2fb1605237037c8bd7668.jpg) + +![](images/45ec03ffa2e49f379143ca7614d60053f9e0f811fd3c203d016790238e600d00.jpg) +16.90 dB Bicubic + +![](images/d4a4ee7a4c389733d89e2a71ee28d9d7a9dea65d1b66edec6dec075351bf135a.jpg) + +![](images/efc98eb37785400a9c767986a9120da7db0ea56c4c60f8fa2a9822dd34381cf7.jpg) +17.54 dB SRCNN [13] + +![](images/75746093afebd3051e4e2320e109478ede0540aff4821e59af273b6ab0c491c1.jpg) + +![](images/796ee10c29fb3f1c5f8ea85ef79a72b6f618b87397c99523b7f1a03acde2a15e.jpg) +17.71 dB FSRCNN [29] + +![](images/3390d5d4392ecc3a6302e05f7f15d37bce49a6bb4d8189486a7f587fdbfd74b0.jpg) + +![](images/24294bddef8c107df51af9be3a5763d13c14f3971dde0d736128786dd077502f.jpg) +18.78 dB CARN [1] + +![](images/507e9e529c914fcca45fed3d861be593ad8c4931a1d6fff2061d2040b4ae10a3.jpg) +18.77 dB +[MDN [22]] + +![](images/05c429fd2742ea720938c2b91ed02202c26231eee65aeda4f572292e5ce0463c.jpg) +19.23 dB RLFN [28] + +![](images/c79a6db4d56adf3eab4c6412ae0c805ff9261eda98ceb04dc7f85b4f9fede4ba.jpg) +18.91 dB LatticeNet [38] + +![](images/89b2d08e02920eeb1a440525897f9db3363759839d30c9c701ac30a394f74c1a.jpg) +19.07 dB SwinIR [31] + +![](images/bb801bd68add06260d770f876872149c032dfc7e783f367e180b9553bded8c1f.jpg) +21.18 dB Omni-SR (Ours) + +![](images/3624e33137ec1266b291073d03192799a127f08adc46b00c2d2a0dae18128c51.jpg) +PSNR Reference + +![](images/c3b48d85deeed63cf49264f906f84a21e402e04e2f791f4e32c43fdeaadfec58.jpg) +23.58 dB Bicubic + +![](images/e7f906c0018864b466385ffdaa4d9b737212722f7535688b21c97690f0b078ad.jpg) +24.04 dB SRCNN [13] + +![](images/d1b7d92000dba7f4252448f42df8ed241decc120487dc1578d8161f5ea35d67b.jpg) +24.11 dB FSRCNN [29] + +![](images/b9987c31f09d6af8941ef0f02b263bd6e2e398ccb41239bd0a34ac1182fb7580.jpg) +24.48 dB CARN [1] + +![](images/758acc31a79474f0bb5a07acfdaba0efa7ad19c835bae55ae7a5d7f31ee328b7.jpg) +24.62 dB IMDN [22] + +![](images/daf5882753d09415d704f856b2b055023f6a358b147b5d07009e844138c0e774.jpg) +24.43 dB RLFN [28] + +![](images/dc33abd478c695a711c0927226082f8e01cc8f485c764b97df6d826b745e2a89.jpg) +24.64 dB LatticeNet [38] + +![](images/3d47daf6c3df1925794bd00e517e13d1381e1b9f2c13b83091fbbb776e4c2079.jpg) +24.45 dB SwinIR [31] + +![](images/2f5a163e9a687d1c82a3b64614fbf786df9a85df92794e741b2c13cfc325abf4.jpg) +26.19 dB Omni-SR (Ours) + +$\mathrm{K} = 1$ increasing the number of OSAG leads to stable performance improvements. In Figure 5(b), we present PSNR vs. parameters of different methods. It can be found that Omni-SR achieves the best results under various settings, showing its effectiveness and scalability. + +# 4.3. Analysis of Omni Self-Attention + +In this section, we illustrate the optimization features of OSA and further uncover its underlying mechanism. Self-attention is a low-bias operator, which makes its optimization difficult and requires more training epochs. For this, we introduce the additional channel-wise interaction to alleviate it. In Figure 7(a), we show the loss curves of different self-attention paradigms on the DIV2K training set, including spatial self-attention, channel self-attention, and the proposed omni self-attention. We can see that our OSA presents a obviously superior convergence speed. More importantly, the performance at the final epoch is also significantly ahead of them. The above phenomenon clearly shows that our OSA has superior good optimization characteristics. Further, we delve into why channel-wise interactions lead to these improvements. We calculate the normalized entropies [52] of the hidden layer features of the network composed of the above three computational primitives. We illustrate the entropy results in Figure 7(c). + +As shown in the figure, in all outgoing layers, our OSA-encoded features show higher entropy, indicating that our OSA encodes richer information. More information may come from various scales, and this information can help the operators to reconstruct the exact details faster. We speculate that this is the potential reason why our OSA shows better optimization properties. In addition, following previous works [8, 16], we also resort to LAM analysis. DI [16] metric can measure the furthest interaction distance of the model. From Figure 8 we can observe that Omni-SR generally has the highest max diffusion index than other methods, showing that our OSA paradigm can effectively capture long-range interactions. + +# 4.4. Ablation Study + +Effect of Omni Self-Attention. The core idea of our framework is to extend the vanilla self-attention with a channel-wise relationship to build omni-axis pixel-wise interaction. Based on Omni-SR framework, we design several variant models, and their SR results are reported in Table 2. We first simply remove the channel-wise component to form a spatial-only variant $(\mathrm{Omni - SR_{sp}})$ , and its performance is degraded by $0.13\mathrm{dB}$ compared to the full model. Such a significant degradation justifies the importance of channel interactions. Note that Omni- $\mathrm{SR_{sp}}$ still + +![](images/1b026b04daf7a008b04feb302dd3557f9f9b26d00703fc2ce40473e53e2a01b4.jpg) +(a) Ablation studies on OSAG + +![](images/4bdb033e383e968248d934e8e127f02ded6bc64bdeda49ea475346c3e22f556d.jpg) +(b) Training loss on DIV2K + +![](images/3fa37dfb9888c6ddc4427f7895be2ebc9b9110e8f7bde7677047a0794d12b00c.jpg) +(c) Feature correlation + +![](images/e385dad27e053ce2fce368849b68cc7dadc75fce15fae1142ee2c956409b8a27.jpg) +Figure 7. (a) Ablation studies on different components in OSAG. (b) Training loss of different attention paradigms. (c) Feature correlation analysis of different attention paradigms. All the results are reported on DIV2K dataset for $4 \times$ SR. +Figure 8. The distribution of DI values for different methods on Urban100. Top-left points with show a narrow area of interest, and right-bottom points show a large area of interest. + +outperforms SwinIR by $0.04\mathrm{dB}@$ Urban $100\times 4$ , which benefits from the global interaction introduced by grid window partition. Similarly, we remove the spatial self-attention component to derive a channel self-attention variant, Omni $\mathrm{SR}_{\mathrm{ca}}$ , and such a modification also leads to undesirable performance degradation. Besides, we use the most widely adopted channel and spatial attention configurations (i.e., SE [19] and CBAM [55]) to act as alternative operators for channel and spatial aggregation. Both substitutions (Omni $\mathrm{SR}_{\mathrm{SE}}$ , Omni- $\mathrm{SR}_{\mathrm{CBAM}}$ ) hurt PNSR performance compared to the full model. The above results show that the specific interaction paradigm (e.g., scalar-based, covariance-based) is equally important, and our channel interaction based on the covariance matrix shows great advantages. + +Effect of Omni-Scale Aggregation Group. In Omni-SR, we propose a local-meso-global interaction scheme (i.e., OSAG) to pursue progressive feature aggregation. To investigate its effectiveness, we design three different kinds of interaction schemes based on Omni-SR framework: Separated scheme, Mixed scheme, and our fully designed Omni scheme (i.e., our proposed OSAG), and the ablation study results are shown in Figure 7(a). In the figure, we employ different words (e.g., "Local", "Meso+Global") to represent specific schemes, e.g., "Local" denotes using Local-Conv block to replace Meso-OSA and Global-OSA; "Local+Global" represents replace original cascaded Meso-OSA and Global-OSA with cascaded Local-Conv and Global-OSA. We can observe that single interaction schemes (e.g., "Local") perform the worst. Interestingly, + +Table 2. Ablation studies of omni self-attention on Urban100. $\mathrm{Omni - SR}_{(*)}$ denotes different modifications. 'SA' and 'S-SA' denote spatial gate and spatial self-attention. 'CA' and 'C-SA' denote channel gate and channel self-attention. We maintain the consistency of model parameters by adjusting the channels of $1\times 1Conv$ + +
ModelSAS-SACAC-SAFLOPs×2×3×4
Omni-SRsp×✓××33G32.8828.7226.51
Omni-SRSE×✓✓×34G32.8328.7126.50
Omni-SRCBAM✓×✓✓34G32.9228.7626.53
Omni-SRca×××✓33G32.6528.6026.45
Omni-SRfull×✓×✓36G33.0528.8426.64
+ +the "Global" scheme is inferior to the "Meso" one due to its poor optimization properties of global self-attention [3, 35, 50]. Once two interaction operators are combined, the performance improves steadily. Among them, "Meso+ Global" setting achieves the second-best performance. Furthermore, combining all three interaction schemes together, we obtain the best performing scheme, i.e., "Omni". From the above experiments, we can infer that obvious performance gains can be obtained by introducing various-scale interactions, which also illustrates the feasibility and effectiveness of our proposed OSAG. + +# 5. Conclusion + +In this work, we propose Omni-SR, a lightweight framework for image SR. We propose the Omni Self-attention paradigm for simultaneous spatial and channel interactions, mining all the potential correlations across omni-axis. Furthermore, we propose an omni-scale aggregation scheme to effectively enlarge the receptive fields with low computational complexity, which encodes contextual relations in a progressively hierarchical manner. Extensive experiments on public benchmark datasets and comprehensive analytical studies validate its prominent SR performance. + +# 6. Acknowledgement + +This work was supported by National Science Foundation of China (U20B2072, 61976137). This work was also partly supported by SJTU Medical Engineering Cross Research Grant YG2021ZD18. + +# References + +[1] Namhyuk Ahn, Byungkon Kang, and Kyung-Ah Sohn. Fast, accurate, and lightweight super-resolution with cascading residual network. In ECCV, 2018. 2, 5, 6, 7 +[2] Lei Jimmy Ba, Jamie Ryan Kiros, and Geoffrey E. Hinton. Layer normalization. CoRR, abs/1607.06450, 2016. 5 +[3] Irwan Bello, Barret Zoph, Quoc Le, Ashish Vaswani, and Jonathon Shlens. Attention augmented convolutional networks. In ICCV, 2019. 8 +[4] Marco Bevilacqua, Aline Roumy, Christine Guillemot, and Marie-Line Alberi-Morel. Low-complexity single-image super-resolution based on nonnegative neighbor embedding. In BMVC, 2012. 5 +[5] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In CVPR, 2021. 1, 2, 3 +[6] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. arXiv:1606.00915, 2016. 1 +[7] Xuanhong Chen, Hang Wang, and Bingbing Ni. X-volution: On the unification of convolution and self-attention. arXiv preprint arXiv:2106.02253, 2021. 2 +[8] Xiangyu Chen, Xintao Wang, Jiantao Zhou, and Chao Dong. Activating more pixels in image super-resolution transformer. CoRR, abs/2205.04437, 2022. 2, 4, 7 +[9] Yinpeng Chen, Xiyang Dai, Dongdong Chen, Mengchen Liu, Xiaoyi Dong, Lu Yuan, and Zicheng Liu. Mobileformer: Bridging mobilenet and transformer. In CVPR, 2022. 2 +[10] Yinpeng Chen, Xiyang Dai, Dongdong Chen, Mengchen Liu, Xiaoyi Dong, Lu Yuan, and Zicheng Liu. Simple baselines for image restoration. In ECCV, 2022. 2, 3 +[11] Xiangxiang Chu, Bo Zhang, Hailong Ma, Ruijun Xu, and Qingyuan Li. Fast, accurate and lightweight super-resolution with neural architecture search. In ICPR, 2020. 2 +[12] Xiaohan Ding, Xiangyu Zhang, Jungong Han, and Guiguang Ding. Scaling up your kernels to 31x31: Revisiting large kernel design in cnns. In CVPR, 2022. 1 +[13] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Image super-resolution using deep convolutional networks. CoRR, abs/1501.00092, 2015. 2, 7 +[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 1, 2 +[15] Qinquan Gao, Yan Zhao, Gen Li, and Tong Tong. Image super-resolution using knowledge distillation. In ACCV, 2018. 2 +[16] Jinjin Gu and Chao Dong. Interpreting super-resolution networks with local attribution maps. In CVPR, 2021. 7 +[17] Geoffrey E. Hinton, Oriol Vinyals, and Jeffrey Dean. Distilling the knowledge in a neural network. CoRR, abs/1503.02531, 2015. 2 + +[18] Andrew G Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andretto, and Hartwig Adam. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv:1704.04861, 2017. 4 +[19] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In CVPR, 2018. 2, 3, 4, 8 +[20] Jia-Bin Huang, Abhishek Singh, and Narendra Ahuja. Single image super-resolution from transformed self-exemplars. In CVPR, 2015. 5 +[21] Tao Huang, Lang Huang, Shan You, Fei Wang, Chen Qian, and Chang Xu. Lightvit: Towards light-weight convolution-free vision transformers. CoRR, abs/2207.05557, 2022. 2 +[22] Zheng Hui, Xinbo Gao, Yunchu Yang, and Xiumei Wang. Lightweight image super-resolution with information multistillation network. In ACMMM, 2019. 2, 5, 6, 7 +[23] Zheng Hui, Xiumei Wang, and Xinbo Gao. Fast and accurate single image super-resolution via information distillation network. In CVPR, 2018. 2 +[24] Forrest N Iandola, Song Han, Matthew W Moskewicz, Khalid Ashraf, William J Dally, and Kurt Keutzer. SqueezeNet: Alexnet-level accuracy with 50x fewer parameters and; 0.5 mb model size. arXiv:1602.07360, 2016. 5 +[25] Jiwon Kim, Jung Kwon Lee, and Kyoung Mu Lee. Accurate image super-resolution using very deep convolutional networks. In CVPR, 2016. 2, 5, 6 +[26] Jiwon Kim, Jung Kwon Lee, and Kyoung Mu Lee. Deeplyrecursive convolutional network for image super-resolution. In CVPR, 2016. 2 +[27] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 5 +[28] Fangyuan Kong, Mingxi Li, Songwei Liu, Ding Liu, Jingwen He, Yang Bai, Fangmin Chen, and Lean Fu. Residual local feature network for efficient super-resolution. In CVPR Workshops, 2022. 2, 5, 6, 7 +[29] Wei-Sheng Lai, Jia-Bin Huang, Narendra Ahuja, and Ming-Hsuan Yang. Fast and accurate image super-resolution with deep laplacian pyramid networks. TPAMI, 41(11):2599-2613, 2018. 7 +[30] Wenbo Li, Xin Lu, Shengju Qian, Jiangbo Lu, Xiangyu Zhang, and Jiaya Jia. On efficient transformer and image pre-training for low-level vision. arXiv preprint arXiv:2112.10175, 2021. 2, 5 +[31] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In ICCV, 2021. 1, 2, 4, 5, 6, 7 +[32] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In CVPRW, 2017. 5, 6 +[33] Jie Liu, Jie Tang, and Gangshan Wu. Residual feature distillation network for lightweight image super-resolution. In ECCV, 2020. 5, 6 +[34] Jie Liu, Wenjie Zhang, Yuting Tang, Jie Tang, and Gangshan Wu. Residual feature aggregation network for image superresolution. In CVPR, 2020. 5 +[35] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: + +Hierarchical vision transformer using shifted windows. In ICCV, 2021. 2, 4, 8 +[36] Ilya Loshchilov and Frank Hutter. SGDR: stochastic gradient descent with warm restarts. In *ICLR*, 2017. 5 +[37] Zhisheng Lu, Juncheng Li, Hong Liu, Chaoyan Huang, Linlin Zhang, and Tieyong Zeng. Transformer for single image super-resolution. In CVPR Workshops, 2022. 2, 3, 4, 5, 6 +[38] Xiaotong Luo, Yuan Xie, Yulun Zhang, Yanyun Qu, Cuihua Li, and Yun Fu. Latticenet: Towards lightweight image super-resolution with lattice block. In ECCV, 2020. 2, 5, 6, 7 +[39] David Martin, Charless Fowlkes, Doron Tal, and Jitendra Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In ICCV. IEEE, 2001. 5 +[40] Yusuke Matsui, Kota Ito, Yuji Aramaki, Azuma Fujimoto, Toru Ogawa, Toshihiko Yamasaki, and Kiyoharu Aizawa. Sketch-based manga retrieval using manga109 dataset. Multimedia Tools and Applications, 76(20):21811-21838, 2017. 5 +[41] Sachin Mehta and Mohammad Rastegari. Mobilevit: Lightweight, general-purpose, and mobile-friendly vision transformer. CoRR, abs/2110.02178, 2021. 2 +[42] Abdul Muqueet, Jiwon Hwang, Subin Yang, JungHeum Kang, Yongwoo Kim, and Sung-Ho Bae. Multi-attention based ultra lightweight image super-resolution. In ECCV, 2020. 2, 5, 6 +[43] Junting Pan, Adrian Bulat, Fuwen Tan, Xiatian Zhu, Lukasz Dudziak, Hongsheng Li, Georgios Tzimiropoulos, and Brais Martínez. Edgevits: Competing light-weight cnns on mobile devices with vision transformers. In ECCV, 2022. 2 +[44] Jongchan Park, Sanghyun Woo, Joon-Young Lee, and In So Kweon. Bam: Bottleneck attention module. arXiv:1807.06514, 2018. 3, 4 +[45] Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. Automatic differentiation in pytorch. In NeurIPS Workshop, 2017. 5 +[46] Wenzhe Shi, Jose Caballero, Ferenc Huszár, Johannes Totz, Andrew P Aitken, Rob Bishop, Daniel Rueckert, and Zehan Wang. Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In CVPR, 2016. 5 +[47] Ying Tai, Jian Yang, and Xiaoming Liu. Image superresolution via deep recursive residual network. In CVPR, 2017. 2 +[48] Ying Tai, Jian Yang, Xiaoming Liu, and Chunyan Xu. Memnet: A persistent memory network for image restoration. In ICCV, 2017. 2, 5, 6 +[49] Radu Timofte, Eirikur Agustsson, Luc Van Gool, Ming-Hsuan Yang, and Lei Zhang. Ntire 2017 challenge on single image super-resolution: Methods and results. In CVPR workshops, 2017. 5 +[50] Ashish Vaswani, Prajit Ramachandran, Aravind Srinivas, Niki Parmar, Blake Hechtman, and Jonathon Shlens. Scaling local self-attention for parameter efficient visual backbones. In CVPR, 2021. 2, 8 + +[51] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NIPS, 2017. 1, 2, 3 +[52] Pei Wang, Yijun Li, and Nuno Vasconcelos. Rethinking and improving the robustness of image style transfer. In CVPR, 2021. 7 +[53] Xiaohang Wang, Xuanhong Chen, Bingbing Ni, Zhengyan Tong, and Hang Wang. Learning continuous depth representation via geometric spatial aggregator. In AAAI, 2023. 5 +[54] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 5 +[55] Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon. Cbam: Convolutional block attention module. In ECCV, 2018. 3, 4, 8 +[56] Tete Xiao, Mannat Singh, Eric Mintun, Trevor Darrell, Piotr Dólar, and Ross B. Girshick. Early convolutions help transformers see better. In NeurIPS, 2021. 5 +[57] Chenglin Yang, Yilin Wang, Jianming Zhang, He Zhang, Zijun Wei, Zhe Lin, and Alan L. Yuille. Lite vision transformer with enhanced self-attention. In CVPR, 2022. 2 +[58] Jianwei Yang, Chunyuan Li, Xiyang Dai, and Jianfeng Gao. Focal modulation networks, 2022. 3 +[59] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In CVPR, 2022. 1, 3, 5 +[60] Roman Zeyde, Michael Elad, and Matan Protter. On single image scale-up using sparse-representations. In International conference on curves and surfaces, pages 711-730. Springer, 2010. 5 +[61] Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In ICLR, 2018. 5 +[62] Haokui Zhang, Wenze Hu, and Xiaoyu Wang. Parc-net: Position aware circular convolution with merits from convnets and transformer. In ECCV, 2022. 2 +[63] Kai Zhang, Wangmeng Zuo, Shuhang Gu, and Lei Zhang. Learning deep cnn denoiser prior for image restoration. In CVPR, 2017. 5 +[64] Kai Zhang, Wangmeng Zuo, and Lei Zhang. Learning a single convolutional super-resolution network for multiple degradations. In CVPR, 2018. 6 +[65] Yiman Zhang, Hanting Chen, Xinghao Chen, Yiping Deng, Chunjing Xu, and Yunhe Wang. Data-free knowledge distillation for image super-resolution. In CVPR, 2021. 2 +[66] Yulun Zhang, Kunpeng Li, Kai Li, Lichen Wang, Bineng Zhong, and Yun Fu. Image super-resolution using very deep residual channel attention networks. In ECCV, 2018. 1, 2, 5 +[67] Yulun Zhang, Yapeng Tian, Yu Kong, Bineng Zhong, and Yun Fu. Residual dense network for image super-resolution. In CVPR, 2018. 5 \ No newline at end of file diff --git a/2023/Omni Aggregation Networks for Lightweight Image Super-Resolution/images.zip b/2023/Omni Aggregation Networks for Lightweight Image Super-Resolution/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..598596b16459182c5069ef031fafa5d45577bbe0 --- /dev/null +++ b/2023/Omni Aggregation Networks for Lightweight Image Super-Resolution/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e506eb5b83ab5e95853ba15daff98185f148b4c457d8e6e21ad51be3574f8caf +size 965883 diff --git a/2023/Omni Aggregation Networks for Lightweight Image Super-Resolution/layout.json b/2023/Omni Aggregation Networks for Lightweight Image Super-Resolution/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3417a329fccea7efd9fcad7cd25b91bd4c13af53 --- /dev/null +++ b/2023/Omni Aggregation Networks for Lightweight Image Super-Resolution/layout.json @@ -0,0 +1,11450 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 82, + 103, + 511, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 103, + 511, + 120 + ], + "spans": [ + { + "bbox": [ + 82, + 103, + 511, + 120 + ], + "type": "text", + "content": "Omni Aggregation Networks for Lightweight Image Super-Resolution" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 102, + 142, + 487, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 142, + 487, + 157 + ], + "spans": [ + { + "bbox": [ + 102, + 142, + 487, + 157 + ], + "type": "text", + "content": "Hang Wang" + }, + { + "bbox": [ + 102, + 142, + 487, + 157 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 102, + 142, + 487, + 157 + ], + "type": "text", + "content": " Xuanhong Chen" + }, + { + "bbox": [ + 102, + 142, + 487, + 157 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 102, + 142, + 487, + 157 + ], + "type": "text", + "content": " Bingbing Ni" + }, + { + "bbox": [ + 102, + 142, + 487, + 157 + ], + "type": "inline_equation", + "content": "^{1,2\\dagger}" + }, + { + "bbox": [ + 102, + 142, + 487, + 157 + ], + "type": "text", + "content": " Yutian Liu" + }, + { + "bbox": [ + 102, + 142, + 487, + 157 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 102, + 142, + 487, + 157 + ], + "type": "text", + "content": " Jinfan Liu" + }, + { + "bbox": [ + 102, + 142, + 487, + 157 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 129, + 157, + 465, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 157, + 465, + 171 + ], + "spans": [ + { + "bbox": [ + 129, + 157, + 465, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 129, + 157, + 465, + 171 + ], + "type": "text", + "content": "Shanghai Jiao Tong University, Shanghai 200240, China " + }, + { + "bbox": [ + 129, + 157, + 465, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 129, + 157, + 465, + 171 + ], + "type": "text", + "content": " Huawei" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 86, + 171, + 503, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 171, + 503, + 186 + ], + "spans": [ + { + "bbox": [ + 86, + 171, + 503, + 186 + ], + "type": "text", + "content": "francis970625@gmail.com {chen19910528,nibingbing,stau7001,ljflnjz}@sjtu.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 213, + 191, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 191, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 191, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 238, + 290, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 238, + 290, + 525 + ], + "spans": [ + { + "bbox": [ + 46, + 238, + 290, + 525 + ], + "type": "text", + "content": "While lightweight ViT framework has made tremendous progress in image super-resolution, its uni-dimensional self-attention modeling, as well as homogeneous aggregation scheme, limit its effective receptive field (ERF) to include more comprehensive interactions from both spatial and channel dimensions. To tackle these drawbacks, this work proposes two enhanced components under a new Omni-SR architecture. First, an Omni Self-Attention (OSA) block is proposed based on dense interaction principle, which can simultaneously model pixel-interaction from both spatial and channel dimensions, mining the potential correlations across omni-axis (i.e., spatial and channel). Coupling with mainstream window partitioning strategies, OSA can achieve superior performance with compelling computational budgets. Second, a multi-scale interaction scheme is proposed to mitigate sub-optimal ERF (i.e., premature saturation) in shallow models, which facilitates local propagation and meso-/global-scale interactions, rendering an omni-scale aggregation building block. Extensive experiments demonstrate that Omni-SR achieves record-high performance on lightweight super-resolution benchmarks (e.g., 26.95dB@Urban100 ×4 with only 792K parameters). Our code is available at https://github.com/Francis0625/Omni-SR." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 547, + 128, + 559 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 547, + 128, + 559 + ], + "spans": [ + { + "bbox": [ + 47, + 547, + 128, + 559 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 567, + 287, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 567, + 287, + 687 + ], + "spans": [ + { + "bbox": [ + 46, + 567, + 287, + 687 + ], + "type": "text", + "content": "Image super-resolution (SR) is a long-standing low-level problem that aims to recover high-resolution (HR) images from degraded low-resolution (LR) inputs. Recently, vision transformer [14, 51] based (i.e., ViT-based) SR frameworks [5, 31] have emerged, showing significant performance gains compared to previously dominant Convolutional Neural Networks (CNNs) [66]. However, most attempts [31] are devoted to improving the large-scale ViT-based models, while the development of lightweight ViTs (typically, less than 1M parameters) remains fraught with" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 313, + 219, + 382, + 296 + ], + "blocks": [ + { + "bbox": [ + 313, + 219, + 382, + 296 + ], + "lines": [ + { + "bbox": [ + 313, + 219, + 382, + 296 + ], + "spans": [ + { + "bbox": [ + 313, + 219, + 382, + 296 + ], + "type": "image", + "image_path": "83843f4b5574f4fcbc3b18bdae706065a44d808b3483f1033fb43f0367b2a5b6.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 329, + 302, + 362, + 310 + ], + "lines": [ + { + "bbox": [ + 329, + 302, + 362, + 310 + ], + "spans": [ + { + "bbox": [ + 329, + 302, + 362, + 310 + ], + "type": "text", + "content": "Spatial SA" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 390, + 219, + 460, + 296 + ], + "blocks": [ + { + "bbox": [ + 390, + 219, + 460, + 296 + ], + "lines": [ + { + "bbox": [ + 390, + 219, + 460, + 296 + ], + "spans": [ + { + "bbox": [ + 390, + 219, + 460, + 296 + ], + "type": "image", + "image_path": "71a6f804e8ab1375f1e3508b309d41565c74a41c0618e0ff295d86651f9eb999.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 406, + 301, + 443, + 310 + ], + "lines": [ + { + "bbox": [ + 406, + 301, + 443, + 310 + ], + "spans": [ + { + "bbox": [ + 406, + 301, + 443, + 310 + ], + "type": "text", + "content": "Channel SA" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 467, + 219, + 537, + 296 + ], + "blocks": [ + { + "bbox": [ + 467, + 219, + 537, + 296 + ], + "lines": [ + { + "bbox": [ + 467, + 219, + 537, + 296 + ], + "spans": [ + { + "bbox": [ + 467, + 219, + 537, + 296 + ], + "type": "image", + "image_path": "8086c58f33b0a6019dc2e25b33792325b411b68b2bf8b57692f6c1666aee8eb1.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 488, + 302, + 517, + 310 + ], + "lines": [ + { + "bbox": [ + 488, + 302, + 517, + 310 + ], + "spans": [ + { + "bbox": [ + 488, + 302, + 517, + 310 + ], + "type": "text", + "content": "Omni SA" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 314, + 318, + 375, + 404 + ], + "blocks": [ + { + "bbox": [ + 314, + 318, + 375, + 404 + ], + "lines": [ + { + "bbox": [ + 314, + 318, + 375, + 404 + ], + "spans": [ + { + "bbox": [ + 314, + 318, + 375, + 404 + ], + "type": "image", + "image_path": "d30ea2a4d9bfa5b22bdcffc7e2e332265caf4a9f549ff90fd94720220b464bbc.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 335, + 405, + 352, + 412 + ], + "lines": [ + { + "bbox": [ + 335, + 405, + 352, + 412 + ], + "spans": [ + { + "bbox": [ + 335, + 405, + 352, + 412 + ], + "type": "text", + "content": "PSNR" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 377, + 318, + 459, + 405 + ], + "blocks": [ + { + "bbox": [ + 377, + 318, + 459, + 405 + ], + "lines": [ + { + "bbox": [ + 377, + 318, + 459, + 405 + ], + "spans": [ + { + "bbox": [ + 377, + 318, + 459, + 405 + ], + "type": "image", + "image_path": "af7413c8721fbc957a54d747d3e3cbb64dfd53dd21045b42dff47fc63bd5563e.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 410, + 405, + 451, + 413 + ], + "lines": [ + { + "bbox": [ + 410, + 405, + 451, + 413 + ], + "spans": [ + { + "bbox": [ + 410, + 405, + 451, + 413 + ], + "type": "text", + "content": "Parameters (K)" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 459, + 318, + 535, + 405 + ], + "blocks": [ + { + "bbox": [ + 459, + 318, + 535, + 405 + ], + "lines": [ + { + "bbox": [ + 459, + 318, + 535, + 405 + ], + "spans": [ + { + "bbox": [ + 459, + 318, + 535, + 405 + ], + "type": "image", + "image_path": "b10e420f7e0a0818b6694462ea8c62a3c345ce4bdc0ae41e0f367c9f1b9849e6.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 496, + 405, + 518, + 413 + ], + "lines": [ + { + "bbox": [ + 496, + 405, + 518, + 413 + ], + "spans": [ + { + "bbox": [ + 496, + 405, + 518, + 413 + ], + "type": "text", + "content": "Iteration" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 422, + 545, + 445 + ], + "lines": [ + { + "bbox": [ + 306, + 422, + 545, + 445 + ], + "spans": [ + { + "bbox": [ + 306, + 422, + 545, + 445 + ], + "type": "text", + "content": "Figure 1. Typical self-attention schemes [31,59] can only perform uni-dimensional (e.g., spatial-only) interactions, and." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 459, + 545, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 459, + 545, + 483 + ], + "spans": [ + { + "bbox": [ + 305, + 459, + 545, + 483 + ], + "type": "text", + "content": "difficulties. This paper focuses on boosting the restoration performance of lightweight ViT-based frameworks." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 486, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 486, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 486, + 546, + 713 + ], + "type": "text", + "content": "Two difficulties hinder the development of lightweight ViT-based models: 1) Uni-dimensional aggregation operators (i.e., spatial [31] only or channel [59] only) imprisons the full potential of self-attention operators. Contemporary self-attention generally realizes the interaction between pixels by calculating the cross-covariance of the spatial direction (i.e., width and height) and exchanges context information in a channel-separated manner. This interaction scheme ignores the explicit use of channel information. However, recent evidences [59] and our practice show that self-attention in the channel dimension (i.e., computationally more compact than spatial self-attention) is also crucial in low-level tasks. 2) Homogeneous aggregation schemes (i.e., Simple hierarchical stacking of single operators, e.g., convolution, self-attention) neglect abundant texture patterns of multi-scales, which is urgently needed in SR task. Specifically, a single operator is only sensitive to information of one scale [6, 12], e.g., self-attention is sensitive to long-term information and pays little attention to lo" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 693, + 127, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 693, + 127, + 703 + ], + "spans": [ + { + "bbox": [ + 58, + 693, + 127, + 703 + ], + "type": "text", + "content": "*Equal Contribution." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 59, + 703, + 180, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 703, + 180, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 703, + 180, + 712 + ], + "type": "text", + "content": "† Corresponding author: Bingbing Ni." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22378" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": "cal information. Additionally, stacking of homogeneous operators proves to be inefficient and suffers from premature saturation of the interaction range [8], which is reflected as a suboptimal effective receptive field. The above problem is exacerbated in lightweight models because lightweight models cannot stack enough layers." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 144, + 289, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 144, + 289, + 539 + ], + "spans": [ + { + "bbox": [ + 46, + 144, + 289, + 539 + ], + "type": "text", + "content": "In order to solve the above problems and pursue higher performance, this work proposes a novel omni-dimension feature aggregation scheme called Omni Self-Attention (OSA) exploiting both spatial and channel axis information in a simultaneous manner (i.e., extends the interaction into three-dimensional space), which offers higher-order receptive field information, as shown in Figure 1. Unlike scalar-based (a group of important coefficient) channel interaction [19], OSA enables comprehensive information propagation and interaction by cascading computation of the cross-covariance matrices between spatial/channel dimensions. The proposed OSA module can be plugged into any mainstream self-attention variants (e.g., Swin [35], Halo [50]), which provides a finer granularity of important encoding (compared to the vanilla channel attention [19]), achieving a perceptible improvement in contextual aggregation capabilities. Furthermore, a multi-scale hierarchical aggregation block, named Omni-Scale Aggregation Group (i.e., OSAG for short), is presented to achieve tailored encoding of varying scales of texture patterns. Specifically, OSAG builds three cascaded aggregators: local convolution (for local details), meso self-attention (focusing on mid-scale pattern processing), and global self-attention (pursuing global context understanding), rendering an omni-scale (i.e., local-/meso-/global-scale simultaneously) feature extraction capability. Compared to the homogenized feature extraction schemes [28, 31], our OSAG is able to mine richer information producing features with higher information entropy. Coupling with the above two designs, we establish a new ViT-based framework for lightweight superresolution, called Omni-SR, which exhibits superior restoration performance as well as covers a larger interaction range while maintaining an attractive model size, i.e., " + }, + { + "bbox": [ + 46, + 144, + 289, + 539 + ], + "type": "inline_equation", + "content": "792K" + }, + { + "bbox": [ + 46, + 144, + 289, + 539 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 539, + 287, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 539, + 287, + 660 + ], + "spans": [ + { + "bbox": [ + 46, + 539, + 287, + 660 + ], + "type": "text", + "content": "We extensively experiment with the proposed framework with both qualitative and quantitative evaluations on mainstream open-source image super-resolution datasets. It is demonstrated that our framework achieves state-of-the-art performance at the lightweight model scale (e.g., " + }, + { + "bbox": [ + 46, + 539, + 287, + 660 + ], + "type": "inline_equation", + "content": "\\text{Urban100} \\times 4: 26.95\\mathrm{dB}" + }, + { + "bbox": [ + 46, + 539, + 287, + 660 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 539, + 287, + 660 + ], + "type": "inline_equation", + "content": "\\text{Manga109} \\times 4: 31.50\\mathrm{dB}" + }, + { + "bbox": [ + 46, + 539, + 287, + 660 + ], + "type": "text", + "content": "). More importantly, compared to existing ViT-based super-solution frameworks, our framework shows superior optimization properties (e.g., convergence speed, smoother loss landscape), which endow our model with better robustness." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 670, + 138, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 138, + 682 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 138, + 682 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "content": "Image Super-resolution. CNNs have achieved remarkable success in image SR task. SRCNN [13] is the first" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 307, + 70, + 425, + 175 + ], + "blocks": [ + { + "bbox": [ + 307, + 70, + 425, + 175 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 425, + 175 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 425, + 175 + ], + "type": "image", + "image_path": "0a11c3f9b6aa9678b0a36037821e2d50d310e88c6867c243403df9e178d054ae.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 181, + 546, + 214 + ], + "lines": [ + { + "bbox": [ + 305, + 181, + 546, + 214 + ], + "spans": [ + { + "bbox": [ + 305, + 181, + 546, + 214 + ], + "type": "text", + "content": "Figure 2. Illustration of spatial attention and channel attention. These typical attention paradigms only model uni-dimensional (i.e., spatial-only / channel-only) interaction." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 425, + 70, + 545, + 175 + ], + "blocks": [ + { + "bbox": [ + 425, + 70, + 545, + 175 + ], + "lines": [ + { + "bbox": [ + 425, + 70, + 545, + 175 + ], + "spans": [ + { + "bbox": [ + 425, + 70, + 545, + 175 + ], + "type": "image", + "image_path": "da9fa7abd21172944f7bbcafab2e7f8521fdb44f49ab186ce3530fcde058bbce.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 232, + 546, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 232, + 546, + 531 + ], + "spans": [ + { + "bbox": [ + 304, + 232, + 546, + 531 + ], + "type": "text", + "content": "work to introduce CNNs into the SR field. Many methods [25, 48, 66] employ skip connection to speed up network convergence and improve the reconstruction quality. Channel Attention [66] is also proposed to enhance the representation ability of the SR model. In order to obtain better reconstruction quality with limited computing resources, several methods [23,38,42,47] explore lightweight architectural design. DRCN [26] utilizes the recursive operation to reduce the number of parameters. DRRN [47] introduces global and local residual learning on the basis of DRCN to accelerate training and improve the quality of details. CARN [1] employs cascading mechanism upon a residual network. IMDN [22] proposes an information multistillation block to archive better time performance. Another line of research is to utilize model compression techniques, e.g., knowledge distillation [15, 17, 65] and neural architecture search [11]) to reduce computing costs. Recently, a series of transformer-based SR models [5,8,31,37] emerge with superior performance. Chenet al. [5] develop a pre-trained model for the low-level computer vision task using the transformer architecture. Based on Swin transformer [35], SwinIR [31] proposes a three-stage framework, refreshing the state-of-the-art of SR task. More recently, some works [5,30] explore ImageNet pre-training strategy to further enhance SR performance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "content": "Lightweight Vision Transformer. Due to the urgent demands for applying networks to resource-constrained devices, lightweight vision transformer [14, 51] has attracted widespread attention. Many attempts [7,9,10,37,41,43,57, 62] have been made to develop lightweight ViTs with comparable performance. A series of methods focus on combining convolutions with transformers to learn both local and global representations. For instance, LVT [57] introduces convolution in self-attention to enrich low-level features. MobileViT [41] replaces matrix multiplication in convolutions with transformer layers to learn global representations. Similarly, EdgeViTs [43] employs an information exchange bottleneck for full spatial interactions. Different from interpreting convolutions into vision transformers, LightViT [21] proposes aggregated self-attention for better" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22379" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 69, + 545, + 273 + ], + "blocks": [ + { + "bbox": [ + 48, + 69, + 545, + 273 + ], + "lines": [ + { + "bbox": [ + 48, + 69, + 545, + 273 + ], + "spans": [ + { + "bbox": [ + 48, + 69, + 545, + 273 + ], + "type": "image", + "image_path": "118b879237fe11c15116324502a94b71f3b9cb241e4e3748c0ae1532450ec6db.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 63, + 276, + 529, + 288 + ], + "lines": [ + { + "bbox": [ + 63, + 276, + 529, + 288 + ], + "spans": [ + { + "bbox": [ + 63, + 276, + 529, + 288 + ], + "type": "text", + "content": "Figure 3. The overall architecture of the proposed Omni-SR framework and structure of OSAG and Omni Self-Attention (OSA)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 303, + 288, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 303, + 288, + 327 + ], + "spans": [ + { + "bbox": [ + 46, + 303, + 288, + 327 + ], + "type": "text", + "content": "information aggregation. In this work, we resort to ViT architecture to achieve lightweight and accurate SR." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 352, + 129, + 366 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 352, + 129, + 366 + ], + "spans": [ + { + "bbox": [ + 47, + 352, + 129, + 366 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 376, + 271, + 390 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 376, + 271, + 390 + ], + "spans": [ + { + "bbox": [ + 47, + 376, + 271, + 390 + ], + "type": "text", + "content": "3.1. Attention Mechanisms in Super-Resolution" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 400, + 287, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 400, + 287, + 425 + ], + "spans": [ + { + "bbox": [ + 46, + 400, + 287, + 425 + ], + "type": "text", + "content": "Two attention paradigms are widely adopted in SR to assist in analyzing and aggregating comprehensive patterns." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 429, + 287, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 429, + 287, + 513 + ], + "spans": [ + { + "bbox": [ + 46, + 429, + 287, + 513 + ], + "type": "text", + "content": "Spatial Attention. Spatial attention can be regarded as an anisotropic selection process. Spatial self-attention [37, 51] and spatial gate [10, 58] are predominantly applied. As shown in Figure 2, spatial self-attention calculates the cross-covariance along the spatial dimension, and the spatial gate generates the channel-separated masks. Neither of them can transmit information between channels." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 517, + 288, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 517, + 288, + 650 + ], + "spans": [ + { + "bbox": [ + 46, + 517, + 288, + 650 + ], + "type": "text", + "content": "Channel Attention. There are two categories of channel attention, i.e., scalar-based [19] and covariance-based [59], proposed to perform channel recalibration or transmit patterns among channels. As shown in Figure 2, the former predicts a group of importance scalars to weigh different channels, while the latter computes a cross-covariance matrix to enable channel re-weighting and information transmission simultaneously. Compared to spatial attention, channel attention handles spatial dimension isotropically, and thus, the complexity is significantly reduced, which also impairs the accuracy of aggregation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": "Several attempts [44,55] have proved that both spatial attention and channel attention are beneficial for SR task and their characteristics are complementary to each other, thus integrating them in a computationally compact way would bring notable benefits in expressive capability." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 303, + 454, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 303, + 454, + 316 + ], + "spans": [ + { + "bbox": [ + 306, + 303, + 454, + 316 + ], + "type": "text", + "content": "3.2. Omni Self-Attention Block" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 323, + 545, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 323, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 304, + 323, + 545, + 514 + ], + "type": "text", + "content": "To mine all the correlations hidden in the latent variables, we propose a novel self-attention paradigm called Omni Self-Attention (OSA) block. Unlike existing self-attention paradigms (e.g., spatial self-attention [5, 37, 51]) that only indulge in unidimensional processing, OSA establishes the spatial and channel context simultaneously. The obtained two-dimensional relationship is highly necessary and beneficial, especially for lightweight models. On the one hand, as the network deepens, important information is scattered into different channels [19], and it is critical to deal with them in a timely manner. On the other hand, although spatial self-attention takes advantage of the channel dimension in calculating the covariance, it does not transmit the information between channels (refer to Sec. 3.1). Given the above conditions, our OSA is designed to transmit both spatial and dimensional information in a compact manner." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 515, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 515, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 515, + 546, + 713 + ], + "type": "text", + "content": "The proposed OSA calculates the score matrices corresponding to the space and channel direction through sequential matrix operations and rotation, as illustrated in Figure 3. Specifically, suppose " + }, + { + "bbox": [ + 304, + 515, + 546, + 713 + ], + "type": "inline_equation", + "content": "X \\in \\mathbb{R}^{HW \\times C}" + }, + { + "bbox": [ + 304, + 515, + 546, + 713 + ], + "type": "text", + "content": " denotes the input feature, where " + }, + { + "bbox": [ + 304, + 515, + 546, + 713 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 304, + 515, + 546, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 515, + 546, + 713 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 304, + 515, + 546, + 713 + ], + "type": "text", + "content": " are the width and height of the input, and " + }, + { + "bbox": [ + 304, + 515, + 546, + 713 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 515, + 546, + 713 + ], + "type": "text", + "content": " is the channel number. Firstly, " + }, + { + "bbox": [ + 304, + 515, + 546, + 713 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 304, + 515, + 546, + 713 + ], + "type": "text", + "content": " is embedded to query, key and value matrices " + }, + { + "bbox": [ + 304, + 515, + 546, + 713 + ], + "type": "inline_equation", + "content": "Q^s, K^s, V^s \\in \\mathbb{R}^{HW \\times C}" + }, + { + "bbox": [ + 304, + 515, + 546, + 713 + ], + "type": "text", + "content": " through linearly projection. We calculate the production of query and key to get the spatial attention map of size " + }, + { + "bbox": [ + 304, + 515, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{HW \\times HW}" + }, + { + "bbox": [ + 304, + 515, + 546, + 713 + ], + "type": "text", + "content": ". Then we perform the spatial attention to obtain the intermediate aggregated results. Note that window strategy is usually used to significantly reduce the resource overhead. Next stage, we rotate the input query and key matrices to get the transposed query and key matrices " + }, + { + "bbox": [ + 304, + 515, + 546, + 713 + ], + "type": "inline_equation", + "content": "Q^c, K^c \\in \\mathbb{R}^{C \\times HW}" + }, + { + "bbox": [ + 304, + 515, + 546, + 713 + ], + "type": "text", + "content": ", and also rotate the value matrices to get the value matrix " + }, + { + "bbox": [ + 304, + 515, + 546, + 713 + ], + "type": "inline_equation", + "content": "V^c \\in \\mathbb{R}^{C \\times HW}" + }, + { + "bbox": [ + 304, + 515, + 546, + 713 + ], + "type": "text", + "content": " for the subsequent channel-wise self-attention. The obtained channel-wise attention map of size " + }, + { + "bbox": [ + 304, + 515, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{C \\times C}" + }, + { + "bbox": [ + 304, + 515, + 546, + 713 + ], + "type": "text", + "content": " models channel-wise relation" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22380" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 288, + 106 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 288, + 106 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 288, + 106 + ], + "type": "text", + "content": "ships. Finally, we get the final aggregated " + }, + { + "bbox": [ + 47, + 72, + 288, + 106 + ], + "type": "inline_equation", + "content": "Y_{OSA}" + }, + { + "bbox": [ + 47, + 72, + 288, + 106 + ], + "type": "text", + "content": " by the inverse rotation of the channel attention output " + }, + { + "bbox": [ + 47, + 72, + 288, + 106 + ], + "type": "inline_equation", + "content": "Y_{c}" + }, + { + "bbox": [ + 47, + 72, + 288, + 106 + ], + "type": "text", + "content": ". The whole OSA process is formulated as follows:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 71, + 114, + 287, + 126 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 114, + 287, + 126 + ], + "spans": [ + { + "bbox": [ + 71, + 114, + 287, + 126 + ], + "type": "interline_equation", + "content": "Q ^ {s} = X \\cdot \\mathcal {W} _ {q}, K ^ {s} = X \\cdot \\mathcal {W} _ {k}, V ^ {s} = X \\cdot \\mathcal {W} _ {v}, \\tag {1}", + "image_path": "02e27cde129bf7285c0bdb688851b46dcf9752601f0cf1eaaaa8051cddb11579.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 72, + 129, + 287, + 142 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 129, + 287, + 142 + ], + "spans": [ + { + "bbox": [ + 72, + 129, + 287, + 142 + ], + "type": "interline_equation", + "content": "Y _ {s} = \\mathcal {A} ^ {s} \\left(Q ^ {s}, K ^ {s}, V ^ {s}\\right) = \\operatorname {S o f t M a x} \\left(Q ^ {s} K ^ {s T}\\right) \\cdot V ^ {s}, \\tag {2}", + "image_path": "e3618b61908d14c039546fab2e427b764e5cd159959107425cf93f23f6392dc7.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 72, + 144, + 287, + 156 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 144, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 72, + 144, + 287, + 156 + ], + "type": "interline_equation", + "content": "Q ^ {c} = \\mathcal {R} \\left(Q ^ {\\prime}\\right), \\quad K ^ {c} = \\mathcal {R} \\left(K ^ {\\prime}\\right), \\quad V ^ {c} = \\mathcal {R} \\left(V ^ {\\prime}\\right), \\tag {3}", + "image_path": "30047d1a2a3eb2e37cad07d3695aa02f4bb4980ffe231ab2dd285a4370d93e3c.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 72, + 159, + 287, + 172 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 159, + 287, + 172 + ], + "spans": [ + { + "bbox": [ + 72, + 159, + 287, + 172 + ], + "type": "interline_equation", + "content": "Y _ {c} = \\mathcal {A} ^ {c} \\left(Q ^ {c}, K ^ {c}, V ^ {c}\\right) = \\operatorname {S o f t M a x} \\left(K ^ {c} Q ^ {c T}\\right) \\cdot V ^ {c}, \\tag {4}", + "image_path": "4ffc98af66cb61dc1b4076c4ee102730e7abed3b63754739ad3db79d44ae8cce.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 72, + 175, + 287, + 186 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 175, + 287, + 186 + ], + "spans": [ + { + "bbox": [ + 72, + 175, + 287, + 186 + ], + "type": "interline_equation", + "content": "Y _ {O S A} = \\mathcal {R} ^ {- 1} \\left(Y _ {c}\\right), \\tag {5}", + "image_path": "36722a2ffcca3a4366214e9ecffef1e590a46a6a1ded5932456500a5df8d9e7c.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 196, + 287, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 196, + 287, + 399 + ], + "spans": [ + { + "bbox": [ + 46, + 196, + 287, + 399 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 196, + 287, + 399 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_q, \\mathcal{W}_k, \\mathcal{W}_v" + }, + { + "bbox": [ + 46, + 196, + 287, + 399 + ], + "type": "text", + "content": " denote the linear projection matrices for the query, key, and value, respectively. " + }, + { + "bbox": [ + 46, + 196, + 287, + 399 + ], + "type": "inline_equation", + "content": "Q', K', V'" + }, + { + "bbox": [ + 46, + 196, + 287, + 399 + ], + "type": "text", + "content": " are the input embedding matrices of channel-wise self-attention, which are embedded from fore spatial self-attention or or copied directly from " + }, + { + "bbox": [ + 46, + 196, + 287, + 399 + ], + "type": "inline_equation", + "content": "Q^s, K^s, V^s" + }, + { + "bbox": [ + 46, + 196, + 287, + 399 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 46, + 196, + 287, + 399 + ], + "type": "inline_equation", + "content": "\\mathcal{R}(\\cdot)" + }, + { + "bbox": [ + 46, + 196, + 287, + 399 + ], + "type": "text", + "content": " denotes the rotation operation around spatial axis and " + }, + { + "bbox": [ + 46, + 196, + 287, + 399 + ], + "type": "inline_equation", + "content": "\\mathcal{R}^{-1}(\\cdot)" + }, + { + "bbox": [ + 46, + 196, + 287, + 399 + ], + "type": "text", + "content": " is the inverse rotation. Some normalization factors are omitted for the sake of simplicity. In particular, this design shows compelling properties that can integrate the elementwise results of two matrix operations (i.e., spatial-/channel-matrix operation), thereby enabling omni-axial interactions. Note that our proposed OSA paradigm can be a drop-in replacement of the Swin [31,35] attention block to higher performance with less parameters. Benefiting from the smaller attention map size of channel self-attention, the proposed OSA is less computationally intensive compared to the cascade shifted-window self-attention scheme in Swin." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 400, + 287, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 400, + 287, + 544 + ], + "spans": [ + { + "bbox": [ + 46, + 400, + 287, + 544 + ], + "type": "text", + "content": "Discussion with other hybrid attention paradigms. Compared to previous hybrid channel and spatial attention works like CBAM [55] and BAM [44], their scalar-based attention weights only reflect the relative degree of importance, without further inter-pixel information exchange, leading to limited relation modeling capability. Several recent works [8] also incorporate channel attention with spatial Self-attention, but these attempts only resort to scalar weights for channel recalibration, while our OSA paradigm enables channel-wise interaction to mine the potential correlations in omni-axis. Performance comparison of different attention paradigms can be found in Sec. 4.4." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 551, + 219, + 564 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 551, + 219, + 564 + ], + "spans": [ + { + "bbox": [ + 47, + 551, + 219, + 564 + ], + "type": "text", + "content": "3.3. Omni-Scale Aggregation Group" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "type": "text", + "content": "How to utilize the proposed OSA paradigm to build a high-performance and compact network is another key topic. Although hierarchical stacking of windows-based self-attention (e.g., swin [31, 35]) has become mainstream, various works have found that the window-based paradigms are very inefficient for large-range interactions, especially for shallow networks. It is worth pointing out that large-range interaction can provide a pleasing effective receptive field, which is crucial for improving image restoration performance [37]. Unfortunately, direct global interaction is resource-prohibitive and detracts from local aggregation capabilities. Taking these points into account, we propose" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 320, + 72, + 533, + 124 + ], + "blocks": [ + { + "bbox": [ + 320, + 72, + 533, + 124 + ], + "lines": [ + { + "bbox": [ + 320, + 72, + 533, + 124 + ], + "spans": [ + { + "bbox": [ + 320, + 72, + 533, + 124 + ], + "type": "image", + "image_path": "daea26868e5c66071cd70cbfa3bee43f8f0b385a1022c93322afdc79d86fd7e6.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 131, + 545, + 164 + ], + "lines": [ + { + "bbox": [ + 305, + 131, + 545, + 164 + ], + "spans": [ + { + "bbox": [ + 305, + 131, + 545, + 164 + ], + "type": "text", + "content": "Figure 4. Illustration of omni-scale aggregation scheme. Our proposed Omni-SR contains three types of feature aggregation at local level, meso and global level, respectively." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 172, + 545, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 172, + 545, + 388 + ], + "spans": [ + { + "bbox": [ + 304, + 172, + 545, + 388 + ], + "type": "text", + "content": "an Omni-Scale Aggregation Group (i.e., OSAG for short) to pursue progressive receptive field feature aggregation with low computational complexity. As shown in Figure 3, OSAG mainly consists of three stages: local, meso and global aggregations. Specifically, a channel attention [19] enhanced inverted bottleneck [18] is introduced to fulfill the local pattern process with limited overhead. Based on the proposed OSA paradigm, we derive two instances (i.e., Meso-OSA and Global-OSA) responsible for the interaction and aggregation of meso and global information. Note that the proposed omni self-attention paradigm can be used for different purposes. Meso-OSA performs attention within a group of non-overlap patches, which restricts Meso-OSA to only focus on meso-scale pattern understanding. Global-OSA samples data point sparsely across the entire feature within an atrous manner, endowing Global-OSA with the ability to achieve global interactions at a compelling cost." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 388, + 545, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 388, + 545, + 534 + ], + "spans": [ + { + "bbox": [ + 305, + 388, + 545, + 534 + ], + "type": "text", + "content": "The only difference between Meso-OSA and Global-OSA is the window partition strategy, as shown in Figure 4. In order to achieve meso-scale interaction, Meso-OSA split the input feature " + }, + { + "bbox": [ + 305, + 388, + 545, + 534 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 305, + 388, + 545, + 534 + ], + "type": "text", + "content": " into non-overlapping blocks with size " + }, + { + "bbox": [ + 305, + 388, + 545, + 534 + ], + "type": "inline_equation", + "content": "P \\times P" + }, + { + "bbox": [ + 305, + 388, + 545, + 534 + ], + "type": "text", + "content": ". Note that after window partition, the block dimensions are gathered onto the spatial dimension (i.e., -2 axis): " + }, + { + "bbox": [ + 305, + 388, + 545, + 534 + ], + "type": "inline_equation", + "content": "(H, W, C) \\to \\left(\\frac{H}{P} \\times P, \\frac{W}{P} \\times P, C\\right) \\to \\left(\\frac{HW}{P^2}, P^2, C\\right)" + }, + { + "bbox": [ + 305, + 388, + 545, + 534 + ], + "type": "text", + "content": ". While the Global-OSA divides the input feature into a uniform " + }, + { + "bbox": [ + 305, + 388, + 545, + 534 + ], + "type": "inline_equation", + "content": "G \\times G" + }, + { + "bbox": [ + 305, + 388, + 545, + 534 + ], + "type": "text", + "content": " grid, with each lattice having an adaptive size of " + }, + { + "bbox": [ + 305, + 388, + 545, + 534 + ], + "type": "inline_equation", + "content": "\\frac{H}{G} \\times \\frac{W}{G}" + }, + { + "bbox": [ + 305, + 388, + 545, + 534 + ], + "type": "text", + "content": ". Similar to Meso-OSA, the grid dimension is also gathered on the spatial axis (i.e., -2 axis): " + }, + { + "bbox": [ + 305, + 388, + 545, + 534 + ], + "type": "inline_equation", + "content": "(H, W, C) \\to (G \\times \\frac{H}{G}, G \\times \\frac{W}{G}, C) \\to (G^2, \\frac{HW}{G^2}, C) \\to (\\frac{HW}{G^2}, G^2, C)" + }, + { + "bbox": [ + 305, + 388, + 545, + 534 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 539, + 432, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 539, + 432, + 550 + ], + "spans": [ + { + "bbox": [ + 306, + 539, + 432, + 550 + ], + "type": "text", + "content": "3.4. Network Architecture" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 555, + 545, + 655 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 555, + 545, + 655 + ], + "spans": [ + { + "bbox": [ + 305, + 555, + 545, + 655 + ], + "type": "text", + "content": "Overall Structure. Based on the Omni Self-Attention paradigm and the Omni-Scale Aggregation Group, we further develop a lightweight Omni-SR framework to achieve high-performance image super-resolution. As shown in Figure 3, Omni-SR consists of three parts, i.e., shallow feature extraction, deep feature extraction, and image reconstruction. Specifically, given the LR input " + }, + { + "bbox": [ + 305, + 555, + 545, + 655 + ], + "type": "inline_equation", + "content": "I_{LR} \\in \\mathbb{R}^{H \\times W \\times C_{in}}" + }, + { + "bbox": [ + 305, + 555, + 545, + 655 + ], + "type": "text", + "content": ", we first use a " + }, + { + "bbox": [ + 305, + 555, + 545, + 655 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 305, + 555, + 545, + 655 + ], + "type": "text", + "content": " convolution " + }, + { + "bbox": [ + 305, + 555, + 545, + 655 + ], + "type": "inline_equation", + "content": "H_{SF}" + }, + { + "bbox": [ + 305, + 555, + 545, + 655 + ], + "type": "text", + "content": " to extract shallow feature " + }, + { + "bbox": [ + 305, + 555, + 545, + 655 + ], + "type": "inline_equation", + "content": "X_0 \\in \\mathbb{R}^{H \\times W \\times C}" + }, + { + "bbox": [ + 305, + 555, + 545, + 655 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 394, + 662, + 545, + 673 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 662, + 545, + 673 + ], + "spans": [ + { + "bbox": [ + 394, + 662, + 545, + 673 + ], + "type": "interline_equation", + "content": "X _ {0} = H _ {S F} \\left(I _ {L R}\\right), \\tag {6}", + "image_path": "addca1f1a4d15e0930b9cabc1529989468775b33d4ec4118acb1188efdcf02de.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 679, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 679, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 679, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 679, + 545, + 713 + ], + "type": "inline_equation", + "content": "C_{in}" + }, + { + "bbox": [ + 305, + 679, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 679, + 545, + 713 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 305, + 679, + 545, + 713 + ], + "type": "text", + "content": " denote the channel number of the input and shallow feature. The convolution layer provides a simple way to convert the input from image space into" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22381" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 118 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 118 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 118 + ], + "type": "text", + "content": "high-dimensional feature space. Then we use " + }, + { + "bbox": [ + 46, + 72, + 287, + 118 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 46, + 72, + 287, + 118 + ], + "type": "text", + "content": " stacked omni-scale aggregation groups (OSAG) and one " + }, + { + "bbox": [ + 46, + 72, + 287, + 118 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 46, + 72, + 287, + 118 + ], + "type": "text", + "content": " convolution layer " + }, + { + "bbox": [ + 46, + 72, + 287, + 118 + ], + "type": "inline_equation", + "content": "H_{CONV}" + }, + { + "bbox": [ + 46, + 72, + 287, + 118 + ], + "type": "text", + "content": " in a cascade manner to extract deep feature " + }, + { + "bbox": [ + 46, + 72, + 287, + 118 + ], + "type": "inline_equation", + "content": "F_{DF}" + }, + { + "bbox": [ + 46, + 72, + 287, + 118 + ], + "type": "text", + "content": ". Such a process can be expressed as" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 90, + 125, + 287, + 142 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 125, + 287, + 142 + ], + "spans": [ + { + "bbox": [ + 90, + 125, + 287, + 142 + ], + "type": "interline_equation", + "content": "X _ {i} = H _ {O S A G _ {i}} \\left(X _ {i - 1}\\right), \\quad i = 1, 2, \\dots , K, \\tag {7}", + "image_path": "23d83b277beacf08a61aec2500aebd0e3cdea89d6708e42adbe4d347735deb88.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 129, + 139, + 206, + 150 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 139, + 206, + 150 + ], + "spans": [ + { + "bbox": [ + 129, + 139, + 206, + 150 + ], + "type": "interline_equation", + "content": "X _ {D F} = H _ {C O N V} \\left(X _ {K}\\right),", + "image_path": "691a23ba134aecbe47e8f557aedf71a7507056ee3d53d14bae4fff754cebff86.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 157, + 287, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 157, + 287, + 214 + ], + "spans": [ + { + "bbox": [ + 47, + 157, + 287, + 214 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 157, + 287, + 214 + ], + "type": "inline_equation", + "content": "H_{\\text{OSAG}_i}" + }, + { + "bbox": [ + 47, + 157, + 287, + 214 + ], + "type": "text", + "content": " represents the " + }, + { + "bbox": [ + 47, + 157, + 287, + 214 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 157, + 287, + 214 + ], + "type": "text", + "content": "-th OSAG, " + }, + { + "bbox": [ + 47, + 157, + 287, + 214 + ], + "type": "inline_equation", + "content": "X_1, X_2, \\ldots, X_K" + }, + { + "bbox": [ + 47, + 157, + 287, + 214 + ], + "type": "text", + "content": " denote intermediate features. Following [31], we also use a convolutional layer at the end of feature extraction to get better feature aggregation. Finally we reconstruct the HR image " + }, + { + "bbox": [ + 47, + 157, + 287, + 214 + ], + "type": "inline_equation", + "content": "I_{HR}" + }, + { + "bbox": [ + 47, + 157, + 287, + 214 + ], + "type": "text", + "content": " by aggregating shallow and deep features as" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 120, + 221, + 287, + 233 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 221, + 287, + 233 + ], + "spans": [ + { + "bbox": [ + 120, + 221, + 287, + 233 + ], + "type": "interline_equation", + "content": "I _ {H R} = H _ {R e c} \\left(X _ {0} + X _ {D F}\\right), \\tag {8}", + "image_path": "0a9c7f613a8deaa784ad855711b6916af62080b90700cbe4de9bf585f867cfea.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 241, + 286, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 241, + 286, + 266 + ], + "spans": [ + { + "bbox": [ + 47, + 241, + 286, + 266 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 241, + 286, + 266 + ], + "type": "inline_equation", + "content": "H_{Rec}(\\cdot)" + }, + { + "bbox": [ + 47, + 241, + 286, + 266 + ], + "type": "text", + "content": " denotes the reconstruction module. In detail, PixelShuffle [46] is used to up-sample the fused feature." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 278, + 287, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 278, + 287, + 334 + ], + "spans": [ + { + "bbox": [ + 47, + 278, + 287, + 334 + ], + "type": "text", + "content": "Omni-Scale Aggregation Group (OSAG). As shown in Figure 3, each OSAG contains a local convolution block (LCB), a meso-OSA block, a global-OSA block, and an ESA block [28, 34]. The whole process can be formulated as" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 74, + 342, + 287, + 354 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 342, + 287, + 354 + ], + "spans": [ + { + "bbox": [ + 74, + 342, + 287, + 354 + ], + "type": "interline_equation", + "content": "X _ {r e s} = H _ {\\text {G l o b a l - O S A B} _ {i}} \\left(H _ {\\text {M e s o - O S A B} _ {i}} \\left(H _ {\\text {L C B} _ {i}} \\left(X _ {i - 1}\\right)\\right)\\right), \\tag {9}", + "image_path": "3b288425f638ca654d092b3db1361a3d61abb66ba7db273b143d63a27c45859a.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 99, + 356, + 287, + 368 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 356, + 287, + 368 + ], + "spans": [ + { + "bbox": [ + 99, + 356, + 287, + 368 + ], + "type": "interline_equation", + "content": "X _ {i} = H _ {E S A _ {i}} \\left(H _ {\\text {C o n v} _ {i}} \\left(X _ {\\text {r e s}} + X _ {i - 1}\\right)\\right), \\tag {10}", + "image_path": "2e00b896ff726de1c930fdc9681855c170ba74514f2438eae6d8bc3b072f5332.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 376, + 287, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 376, + 287, + 448 + ], + "spans": [ + { + "bbox": [ + 47, + 376, + 287, + 448 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 376, + 287, + 448 + ], + "type": "inline_equation", + "content": "X_{i-1}" + }, + { + "bbox": [ + 47, + 376, + 287, + 448 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 376, + 287, + 448 + ], + "type": "inline_equation", + "content": "X_i" + }, + { + "bbox": [ + 47, + 376, + 287, + 448 + ], + "type": "text", + "content": " represents the input and output feature of " + }, + { + "bbox": [ + 47, + 376, + 287, + 448 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 376, + 287, + 448 + ], + "type": "text", + "content": "-th OSAG. After the mapping of convolution layers, we insert a Meso-OSAB for window-based self-attention and a Global-OSAB to enlarge the receptive field for better information aggregation. At the end of OSAG, we reserve the convolutional layer and ESA block following [28, 66]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 448, + 287, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 448, + 287, + 674 + ], + "spans": [ + { + "bbox": [ + 46, + 448, + 287, + 674 + ], + "type": "text", + "content": "In specific, LCB is implemented as a stack of pointwise and depthwise convolutions with a CA module [24] between them to adaptively re-weight channel-wise features. This block aims to aggregate local contextual information as well as to increase the trainability of the network [56]. Two types of OSA blocks (i.e., Meso-OSA block and Global-OSA block) are then followed to obtain interactions from different regions. Based on different window partition strategies, Meso-OSA block seeks inner-block interaction, and Global-OSA blocks aim for global mixing. OSA blocks follow typical Transformers designs with Feedforward network (FFN) and LayerNorm [2], and the only difference is that the origin self-attention operation is replaced with our proposed OSA operator. For FFN, we adopt the GDFN proposed by Restormer [59]. Combining these individuals seamlessly, the designed OSAG enables information propagation between any pair of tokens in the feature map. We use the ESA module proposed in [28, 34] to further refine the fused feature." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "Optimization Objective. Following prior works [31, 32, 53, 67], we train the model by minimizing a standard " + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "inline_equation", + "content": "L_{1}" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "content": "loss between model prediction " + }, + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "inline_equation", + "content": "\\hat{I}_{HR}" + }, + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "content": " and HR label " + }, + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "inline_equation", + "content": "I_{HR}" + }, + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 385, + 97, + 545, + 110 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 385, + 97, + 545, + 110 + ], + "spans": [ + { + "bbox": [ + 385, + 97, + 545, + 110 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\left\\| I _ {H R} - \\hat {I} _ {H R} \\right\\| _ {1}. \\tag {11}", + "image_path": "b4176e0fd1b93284dd7a6249f441d53e1a33d3720c72a043d0a55f0b5e630202.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 130, + 386, + 143 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 130, + 386, + 143 + ], + "spans": [ + { + "bbox": [ + 306, + 130, + 386, + 143 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 150, + 422, + 163 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 150, + 422, + 163 + ], + "spans": [ + { + "bbox": [ + 306, + 150, + 422, + 163 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 170, + 545, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 170, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 305, + 170, + 545, + 289 + ], + "type": "text", + "content": "Datasets and Metrics. Following previous work [31,32, 38,49,66], DIV2K [49] and Flickr2K [49] are used as training datasets. For a fair comparison, we employ two training protocols, i.e., training with DIV2K only and training with DF2K (DIV2K + Flickr2K). Note that the model trained with DF2K is marked with small " + }, + { + "bbox": [ + 305, + 170, + 545, + 289 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 305, + 170, + 545, + 289 + ], + "type": "text", + "content": ". For testing, we adopt five standard benchmark datasets: Set5 [4], Set14 [60], B100 [39], Urban100 [20] and Manga109 [40]. PSNR and SSIM [54] are adopted to evaluate the SR performance on the Y channel of the transformed YCbCr space." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 290, + 545, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 290, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 305, + 290, + 545, + 506 + ], + "type": "text", + "content": "Implementation Details. During training, we augment the data with random horizontal flips and 90/270-degree rotations. LR images are generated by bicubic downsampling [63] from HR images. OSAG number is set to 5, and channel number of the whole network is set to 64. The attention head number and window size are set to 4 and 8 for both Meso-OSAB and Global-OSAB. We use Adam [27] optimizer to train the model with a batch size of 64 for " + }, + { + "bbox": [ + 305, + 290, + 545, + 506 + ], + "type": "inline_equation", + "content": "800\\mathrm{K}" + }, + { + "bbox": [ + 305, + 290, + 545, + 506 + ], + "type": "text", + "content": " iterations. The initial learning rate is set to " + }, + { + "bbox": [ + 305, + 290, + 545, + 506 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 305, + 290, + 545, + 506 + ], + "type": "text", + "content": " and halved for every " + }, + { + "bbox": [ + 305, + 290, + 545, + 506 + ], + "type": "inline_equation", + "content": "200\\mathrm{k}" + }, + { + "bbox": [ + 305, + 290, + 545, + 506 + ], + "type": "text", + "content": " iterations. In each training batch, we randomly crop LR patches of size " + }, + { + "bbox": [ + 305, + 290, + 545, + 506 + ], + "type": "inline_equation", + "content": "64\\times 64" + }, + { + "bbox": [ + 305, + 290, + 545, + 506 + ], + "type": "text", + "content": " as input. Our method is implemented with PyTorch [45], and all experiments are conducted on one NVIDIA V100 GPU. Note that no other data augmentation (e.g., Mixup [61], RGB channel shuffle) or training skills (e.g., pre-training [30], cosine learning schedule [36]) are employed. It should be pointed out that we maintain the consistency of model parameters in the ablation study by adjusting the channels of " + }, + { + "bbox": [ + 305, + 290, + 545, + 506 + ], + "type": "inline_equation", + "content": "1\\times 1" + }, + { + "bbox": [ + 305, + 290, + 545, + 506 + ], + "type": "text", + "content": " convolution." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 514, + 517, + 527 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 514, + 517, + 527 + ], + "spans": [ + { + "bbox": [ + 306, + 514, + 517, + 527 + ], + "type": "text", + "content": "4.2. Comparison with the SOTA SR methods" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 533, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 533, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 305, + 533, + 545, + 617 + ], + "type": "text", + "content": "To evaluate the effectiveness of Omni-SR, we compare our model with several advanced lightweight SR methods under a scale factor of 2/3/4. In particular, former works, VDSR [25], CARN [1], IMDN [22], EDSR [32], RFDN [33], MemNet [48], MAFFSRN [42], LatticeNet [38], RLFN [28], ESRT [37] and SwinIR [31] are introduced for comparison." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "Quantitative results. In Table 1, the quantitative comparisons of different lightweight methods are presented on five benchmark datasets. With a similar model size, the performance of our Omni-SR surpasses existing methods with a notable margin on all benchmarks. In particular, compared to other transformer architectures with comparable parameters like SwinIR [31] and ESRT [37], the proposed Omni-SR obtains the best performance. The results" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22382" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 55, + 94, + 538, + 430 + ], + "blocks": [ + { + "bbox": [ + 47, + 71, + 545, + 92 + ], + "lines": [ + { + "bbox": [ + 47, + 71, + 545, + 92 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 545, + 92 + ], + "type": "text", + "content": "Table 1. Quantitative comparison (PSNR/SSIM) for lightweight image SR with state-of-the-art methods on benchmark datasets. The best and second-best results are marked in red and blue colors, respectively. “†” indicates that model is trained on DF2K." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 55, + 94, + 538, + 430 + ], + "lines": [ + { + "bbox": [ + 55, + 94, + 538, + 430 + ], + "spans": [ + { + "bbox": [ + 55, + 94, + 538, + 430 + ], + "type": "table", + "html": "
MethodYearsScaleParams(K)Set5 PSNR/ SSIMSet14 PSNR/ SSIMBSD100 PSNR/ SSIMUrban100 PSNR/ SSIMManga109 PSNR/ SSIM
VDSR [25]CVPR1666636.66 / 0.954233.05 / 0.912731.90 / 0.896030.76 / 0.914037.22 / 0.9750
MemNet [48]ICCV1767837.78 / 0.959733.28 / 0.914232.08 / 0.897831.31 / 0.919537.72 / 0.9740
SRMDNF [64]CVPR18151137.79 / 0.96033.32 / 0.91532.05 / 0.898531.33 / 0.920438.07 / 0.9761
CARN [1]ECCV181,59237.76 / 0.959033.52 / 0.916632.09 / 0.897831.92 / 0.925638.36 / 0.9765
IMDN [22]MM1969438.00 / 0.960533.63 / 0.917732.19 / 0.899632.17 / 0.928338.88 / 0.9774
RFDN-L [33]ECCV2062638.08 / 0.960633.67 / 0.919032.18 / 0.899632.24 / 0.929038.95 / 0.9773
MAFFSRN [42]ECCV20×240237.97 / 0.960333.49 / 0.917032.14 / 0.899431.96 / 0.9268-/-
LatticeNet [38]ECCV2075638.15 / 0.961033.78 / 0.919332.25 / 0.900532.43 / 0.9302-/-
RLFN [28]CVPRW2252738.07 / 0.960733.72 / 0.918732.22 / 0.900032.33 / 0.9299-/-
SwinIR [31]ICCVW2187838.14 / 0.961133.86 / 0.920632.31 / 0.901232.76 / 0.934039.12 / 0.9783
Omni-SROurs77238.22 / 0.961333.98 / 0.921032.36 / 0.902033.05 / 0.936339.28 / 0.9784
Omni-SR†Ours77238.29 / 0.961734.27 / 0.923832.41 / 0.902633.30 / 0.938639.53 / 0.9792
VDSR [25]CVPR1666633.66 / 0.921329.77 / 0.831428.82 / 0.797627.14 / 0.827932.01 / 0.9340
MemNet [48]ICCV1767834.09 / 0.924830.00 / 0.835028.96 / 0.800127.56 / 0.837632.51 / 0.9369
EDSR [32]CVPRW171,55534.37 / 0.927030.28 / 0.841729.09 / 0.805228.15 / 0.852733.45 / 0.9439
SRMDNF [64]CVPR181,52834.12 / 0.925430.04 / 0.838228.97 / 0.802527.57 / 0.839833.00 / 0.9403
CARN [1]ECCV181,59234.29 / 0.925530.29 / 0.840729.06 / 0.803428.06 / 0.849333.50 / 0.9440
IMDN [22]MM1970334.36 / 0.927030.32 / 0.841729.09 / 0.804628.17 / 0.851933.61 / 0.9445
RFDN-L [33]ECCV20×363334.47 / 0.928030.35 / 0.842129.11 / 0.805328.32 / 0.854733.78 / 0.9458
MAFFSRN [42]ECCV2080734.45 / 0.927730.40 / 0.843229.13 / 0.806128.26 / 0.8552-/-
LatticeNet [38]ECCV2076534.53 / 0.928130.39 / 0.842429.15 / 0.805928.33 / 0.8538-/-
ESRT [37]CVPRW2277034.42 / 0.926830.43 / 0.843329.15 / 0.806328.46 / 0.857433.95 / 0.9455
SwinIR [31]ICCVW2188634.62 / 0.928930.54 / 0.846329.20 / 0.808228.66 / 0.862433.98 / 0.9478
Omni-SROurs78034.70 / 0.929430.57 / 0.846929.28 / 0.809428.84 / 0.865634.22 / 0.9487
Omni-SR†Ours78034.77 / 0.930430.70 / 0.848929.33 / 0.811129.12 / 0.871234.64 / 0.9507
VDSR [25]CVPR1666631.35 / 0.883828.01 / 0.767427.29 / 0.725125.18 / 0.752428.83 / 0.8870
MemNet [48]ICCV1767831.74 / 0.889328.26 / 0.772327.40 / 0.728125.50 / 0.763029.42 / 0.8942
EDSR [32]CVPRW171,51832.09 / 0.893828.58 / 0.781327.57 / 0.735726.04 / 0.784930.35 / 0.9067
SRMDNF [64]CVPR181,55231.96 / 0.892528.35 / 0.778727.49 / 0.733725.68 / 0.773130.09 / 0.9024
CARN [1]ECCV181,59232.13 / 0.893728.60 / 0.780627.58 / 0.734926.07 / 0.783730.47 / 0.9084
IMDN [22]MM1971532.21 / 0.894828.58 / 0.781127.56 / 0.735326.04 / 0.783830.45 / 0.9075
RFDN-L [33]ECCV20×464332.28 / 0.895728.61 / 0.781827.58 / 0.736326.20 / 0.788330.61 / 0.9096
MAFFSRN [42]ECCV2083032.20 / 0.895326.62 / 0.782227.59 / 0.737026.16 / 0.7887-/-
LatticeNet [38]ECCV2077732.30 / 0.896228.68 / 0.783027.62 / 0.736726.25 / 0.7873-/-
ESRT [37]CVPRW2254332.24 / 0.895228.62 / 0.781327.60 / 0.736426.17 / 0.7877-/-
SwinIR [31]ICCVW2175132.19 / 0.894728.69 / 0.783327.69 / 0.737926.39 / 0.796230.75 / 0.9100
Omni-SROurs89732.44 / 0.897628.77 / 0.785827.69 / 0.740626.47 / 0.798030.92 / 0.9151
Omni-SR†Ours79232.49 / 0.898828.78 / 0.785927.71 / 0.741526.64 / 0.801831.02 / 0.9151
Omni-SR†Ours79232.57 / 0.899328.95 / 0.789827.81 / 0.743926.95 / 0.810531.50 / 0.9192
", + "image_path": "c2e54aa90bde62e82b487e9f121e248131537897edc599a39451d3c33c4058dd.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 48, + 441, + 165, + 545 + ], + "blocks": [ + { + "bbox": [ + 48, + 441, + 165, + 545 + ], + "lines": [ + { + "bbox": [ + 48, + 441, + 165, + 545 + ], + "spans": [ + { + "bbox": [ + 48, + 441, + 165, + 545 + ], + "type": "image", + "image_path": "893f81de995407df99654d0b38795a6f908c601bfb716c4c33644406bacbe803.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 105, + 550, + 116, + 559 + ], + "lines": [ + { + "bbox": [ + 105, + 550, + 116, + 559 + ], + "spans": [ + { + "bbox": [ + 105, + 550, + 116, + 559 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 167, + 441, + 284, + 549 + ], + "blocks": [ + { + "bbox": [ + 167, + 441, + 284, + 549 + ], + "lines": [ + { + "bbox": [ + 167, + 441, + 284, + 549 + ], + "spans": [ + { + "bbox": [ + 167, + 441, + 284, + 549 + ], + "type": "image", + "image_path": "ffe582007a99f84656e3b6d885cdc2a6f8a9cf27362c2e7952528c61d4746ecc.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 228, + 550, + 240, + 559 + ], + "lines": [ + { + "bbox": [ + 228, + 550, + 240, + 559 + ], + "spans": [ + { + "bbox": [ + 228, + 550, + 240, + 559 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 562, + 287, + 606 + ], + "lines": [ + { + "bbox": [ + 46, + 562, + 287, + 606 + ], + "spans": [ + { + "bbox": [ + 46, + 562, + 287, + 606 + ], + "type": "text", + "content": "Figure 5. (a) PSNR improvement of Omni-SR variants with different OSAG number (K) over the smallest Omni-SR model " + }, + { + "bbox": [ + 46, + 562, + 287, + 606 + ], + "type": "inline_equation", + "content": "(\\mathrm{K} = 1)" + }, + { + "bbox": [ + 46, + 562, + 287, + 606 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 46, + 562, + 287, + 606 + ], + "type": "inline_equation", + "content": "4\\times" + }, + { + "bbox": [ + 46, + 562, + 287, + 606 + ], + "type": "text", + "content": " SR. (b) The number of model parameters vs. PSNR of different lightweight methods on Urban100 dataset for " + }, + { + "bbox": [ + 46, + 562, + 287, + 606 + ], + "type": "inline_equation", + "content": "4\\times" + }, + { + "bbox": [ + 46, + 562, + 287, + 606 + ], + "type": "text", + "content": " SR." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 287, + 714 + ], + "type": "text", + "content": "showcase that the omni-axis (i.e., spatial + channel) interaction introduced by OSA can effectively boost the model's contextual aggregation capabilities, which promises superior SR performance. Coupling with large training dataset DF2K, the performance can be further improved, especially on Urban100. We suppose that such a phenomenon can be ascribed to the images in Urban100 have many similar patches, and the long-term relationship introduced by" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 443, + 545, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 443, + 545, + 502 + ], + "spans": [ + { + "bbox": [ + 305, + 443, + 545, + 502 + ], + "type": "text", + "content": "OSAG can bring great benefits for detail restoration. More importantly, with similar parameters, our model reduces " + }, + { + "bbox": [ + 305, + 443, + 545, + 502 + ], + "type": "inline_equation", + "content": "28\\%" + }, + { + "bbox": [ + 305, + 443, + 545, + 502 + ], + "type": "text", + "content": " of computational complexity (Omni-SR: 36G FLOPs vs. SwinIR: 50G FLOPs @1280×720), showing its effectiveness and efficiency." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 506, + 546, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 506, + 546, + 638 + ], + "spans": [ + { + "bbox": [ + 304, + 506, + 546, + 638 + ], + "type": "text", + "content": "Visual comparison. In Figure 6, we also provide a visual comparison of different lightweight SR methods at " + }, + { + "bbox": [ + 304, + 506, + 546, + 638 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 304, + 506, + 546, + 638 + ], + "type": "text", + "content": " scale. We can observe that the HR images constructed by Omni-SR contain more fine-grained details, while other methods generate blurred edges or artifacts in complicated areas. For instance, in the " + }, + { + "bbox": [ + 304, + 506, + 546, + 638 + ], + "type": "inline_equation", + "content": "1st" + }, + { + "bbox": [ + 304, + 506, + 546, + 638 + ], + "type": "text", + "content": " row, our model is able to restore the detailed texture of the wall pleasantly, which all other methods fail to restore. Visual results also validate the effectiveness of the proposed OSA paradigm, which can perform omni-axis pixel-wise interaction modeling, thus obtaining a more powerful reconstruction capability." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "type": "text", + "content": "Trade-off between Model Size and Performance. In experiments, we set the number of OSAG as 5 to make the model size around " + }, + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "type": "inline_equation", + "content": "800\\mathrm{K}" + }, + { + "bbox": [ + 304, + 642, + 547, + 713 + ], + "type": "text", + "content": " for a fair comparison with other methods. We also explore model performance with smaller parameter sizes by reducing OSAG number K. As shown in Figure 5(a), compared to the smallest model variant with" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22383" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 70, + 189, + 158 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 189, + 158 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 189, + 158 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 189, + 158 + ], + "type": "image", + "image_path": "94d802cd6418ce37365ce9e96117e6ce93556852e979423c7d4d4fb7695d8b22.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 70, + 161, + 182, + 173 + ], + "lines": [ + { + "bbox": [ + 70, + 161, + 182, + 173 + ], + "spans": [ + { + "bbox": [ + 70, + 161, + 182, + 173 + ], + "type": "text", + "content": "Urban100: img_012.png" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 190, + 70, + 258, + 105 + ], + "blocks": [ + { + "bbox": [ + 190, + 70, + 258, + 105 + ], + "lines": [ + { + "bbox": [ + 190, + 70, + 258, + 105 + ], + "spans": [ + { + "bbox": [ + 190, + 70, + 258, + 105 + ], + "type": "image", + "image_path": "494a73f50393065708fb86e2beeeeef07544713ef02e318bd799cdb4823f802c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 206, + 105, + 244, + 121 + ], + "lines": [ + { + "bbox": [ + 206, + 105, + 244, + 121 + ], + "spans": [ + { + "bbox": [ + 206, + 105, + 244, + 121 + ], + "type": "text", + "content": "PSNR Reference" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 260, + 70, + 328, + 105 + ], + "blocks": [ + { + "bbox": [ + 260, + 70, + 328, + 105 + ], + "lines": [ + { + "bbox": [ + 260, + 70, + 328, + 105 + ], + "spans": [ + { + "bbox": [ + 260, + 70, + 328, + 105 + ], + "type": "image", + "image_path": "401fb73e79fac28b9e0da2dd38d4647fc73da9a39f897c70b0fdd43f2fa8f301.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 277, + 105, + 311, + 121 + ], + "lines": [ + { + "bbox": [ + 277, + 105, + 311, + 121 + ], + "spans": [ + { + "bbox": [ + 277, + 105, + 311, + 121 + ], + "type": "text", + "content": "20.65 dB Bicubic" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 329, + 70, + 397, + 105 + ], + "blocks": [ + { + "bbox": [ + 329, + 70, + 397, + 105 + ], + "lines": [ + { + "bbox": [ + 329, + 70, + 397, + 105 + ], + "spans": [ + { + "bbox": [ + 329, + 70, + 397, + 105 + ], + "type": "image", + "image_path": "883f3e8878b3dc8f6e6c72dd84f3a645091e2d54afd68c28b67c0d9ed14f3e28.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 337, + 105, + 389, + 122 + ], + "lines": [ + { + "bbox": [ + 337, + 105, + 389, + 122 + ], + "spans": [ + { + "bbox": [ + 337, + 105, + 389, + 122 + ], + "type": "text", + "content": "21.15 dB SRCNN [13]" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 399, + 70, + 466, + 105 + ], + "blocks": [ + { + "bbox": [ + 399, + 70, + 466, + 105 + ], + "lines": [ + { + "bbox": [ + 399, + 70, + 466, + 105 + ], + "spans": [ + { + "bbox": [ + 399, + 70, + 466, + 105 + ], + "type": "image", + "image_path": "1e5a0907028261a23d10f724de5d8d804a90b7dacce7856eb02735db56b204df.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 406, + 105, + 460, + 122 + ], + "lines": [ + { + "bbox": [ + 406, + 105, + 460, + 122 + ], + "spans": [ + { + "bbox": [ + 406, + 105, + 460, + 122 + ], + "type": "text", + "content": "21.22 dB FSRCNN [29]" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 467, + 70, + 536, + 105 + ], + "blocks": [ + { + "bbox": [ + 467, + 70, + 536, + 105 + ], + "lines": [ + { + "bbox": [ + 467, + 70, + 536, + 105 + ], + "spans": [ + { + "bbox": [ + 467, + 70, + 536, + 105 + ], + "type": "image", + "image_path": "f89c955c7a4e3a57046090ed64ecaaa0e21b351a1926ff23b69312fe596e0792.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 484, + 105, + 522, + 122 + ], + "lines": [ + { + "bbox": [ + 484, + 105, + 522, + 122 + ], + "spans": [ + { + "bbox": [ + 484, + 105, + 522, + 122 + ], + "type": "text", + "content": "21.99 dB CARN [1]" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 61, + 177, + 188, + 262 + ], + "blocks": [ + { + "bbox": [ + 61, + 177, + 188, + 262 + ], + "lines": [ + { + "bbox": [ + 61, + 177, + 188, + 262 + ], + "spans": [ + { + "bbox": [ + 61, + 177, + 188, + 262 + ], + "type": "image", + "image_path": "77836fefd440f40b36f59a49c94431d7ba2877842db8df27ee2c56c043a021fb.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 70, + 266, + 182, + 278 + ], + "lines": [ + { + "bbox": [ + 70, + 266, + 182, + 278 + ], + "spans": [ + { + "bbox": [ + 70, + 266, + 182, + 278 + ], + "type": "text", + "content": "Urban100: img_024.png" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 190, + 123, + 258, + 158 + ], + "blocks": [ + { + "bbox": [ + 190, + 123, + 258, + 158 + ], + "lines": [ + { + "bbox": [ + 190, + 123, + 258, + 158 + ], + "spans": [ + { + "bbox": [ + 190, + 123, + 258, + 158 + ], + "type": "image", + "image_path": "ab6c47781207dc279b26c766c0b3d0a8968f71e8642f553ef7451110331dd877.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 205, + 156, + 246, + 175 + ], + "lines": [ + { + "bbox": [ + 205, + 156, + 246, + 175 + ], + "spans": [ + { + "bbox": [ + 205, + 156, + 246, + 175 + ], + "type": "text", + "content": "21.95 dB IMDN [22]" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 260, + 123, + 328, + 158 + ], + "blocks": [ + { + "bbox": [ + 260, + 123, + 328, + 158 + ], + "lines": [ + { + "bbox": [ + 260, + 123, + 328, + 158 + ], + "spans": [ + { + "bbox": [ + 260, + 123, + 328, + 158 + ], + "type": "image", + "image_path": "02a5fb9235cbec0479716589693eedc88d20f1ca68ad0b747f0d30bf586135d2.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 272, + 156, + 316, + 175 + ], + "lines": [ + { + "bbox": [ + 272, + 156, + 316, + 175 + ], + "spans": [ + { + "bbox": [ + 272, + 156, + 316, + 175 + ], + "type": "text", + "content": "21.95 dB RLFN [28]" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 329, + 123, + 397, + 158 + ], + "blocks": [ + { + "bbox": [ + 329, + 123, + 397, + 158 + ], + "lines": [ + { + "bbox": [ + 329, + 123, + 397, + 158 + ], + "spans": [ + { + "bbox": [ + 329, + 123, + 397, + 158 + ], + "type": "image", + "image_path": "89cf7d445d853fc9f318cfb4b06feb57bff736ef54700ddc08adc1ded5c7cd1e.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 335, + 158, + 391, + 175 + ], + "lines": [ + { + "bbox": [ + 335, + 158, + 391, + 175 + ], + "spans": [ + { + "bbox": [ + 335, + 158, + 391, + 175 + ], + "type": "text", + "content": "22.00 dB LatticeNet [38]" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 399, + 123, + 466, + 158 + ], + "blocks": [ + { + "bbox": [ + 399, + 123, + 466, + 158 + ], + "lines": [ + { + "bbox": [ + 399, + 123, + 466, + 158 + ], + "spans": [ + { + "bbox": [ + 399, + 123, + 466, + 158 + ], + "type": "image", + "image_path": "dafc61eeee66ffca2f3a71457045dd04837b4e46ef0c3ffedab0f4bbc58020fa.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 409, + 158, + 456, + 175 + ], + "lines": [ + { + "bbox": [ + 409, + 158, + 456, + 175 + ], + "spans": [ + { + "bbox": [ + 409, + 158, + 456, + 175 + ], + "type": "text", + "content": "22.34 dB SwinIR [31]" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 468, + 123, + 536, + 158 + ], + "blocks": [ + { + "bbox": [ + 468, + 123, + 536, + 158 + ], + "lines": [ + { + "bbox": [ + 468, + 123, + 536, + 158 + ], + "spans": [ + { + "bbox": [ + 468, + 123, + 536, + 158 + ], + "type": "image", + "image_path": "4ec3907b3e3aa4f900e3e98c6906c9e62be75f06cb6569e71ef5d6bfb2cf891c.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 469, + 158, + 535, + 175 + ], + "lines": [ + { + "bbox": [ + 469, + 158, + 535, + 175 + ], + "spans": [ + { + "bbox": [ + 469, + 158, + 535, + 175 + ], + "type": "text", + "content": "24.16 dB Omni-SR (Ours)" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 61, + 283, + 188, + 369 + ], + "blocks": [ + { + "bbox": [ + 61, + 283, + 188, + 369 + ], + "lines": [ + { + "bbox": [ + 61, + 283, + 188, + 369 + ], + "spans": [ + { + "bbox": [ + 61, + 283, + 188, + 369 + ], + "type": "image", + "image_path": "6dcc1cf12baab7f8aa1c5bf8c289d536f1246b2712c317b68b08601e653fe82f.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 372, + 185, + 384 + ], + "lines": [ + { + "bbox": [ + 67, + 372, + 185, + 384 + ], + "spans": [ + { + "bbox": [ + 67, + 372, + 185, + 384 + ], + "type": "text", + "content": "Set14: img_barbara.png" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 390, + 537, + 402 + ], + "lines": [ + { + "bbox": [ + 55, + 390, + 537, + 402 + ], + "spans": [ + { + "bbox": [ + 55, + 390, + 537, + 402 + ], + "type": "text", + "content": "Figure 6. Visual comparison for " + }, + { + "bbox": [ + 55, + 390, + 537, + 402 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 55, + 390, + 537, + 402 + ], + "type": "text", + "content": " SR methods. The patches for comparison are marked with red boxes. (Best viewed by zooming.)" + } + ] + } + ], + "index": 71, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 190, + 158, + 258, + 175 + ], + "blocks": [ + { + "bbox": [ + 190, + 158, + 258, + 175 + ], + "lines": [ + { + "bbox": [ + 190, + 158, + 258, + 175 + ], + "spans": [ + { + "bbox": [ + 190, + 158, + 258, + 175 + ], + "type": "image", + "image_path": "269c2cee48363e525646d0ac24f793ad38cf655e6d1b9a76a24d3fec01f95687.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 190, + 175, + 258, + 211 + ], + "blocks": [ + { + "bbox": [ + 190, + 175, + 258, + 211 + ], + "lines": [ + { + "bbox": [ + 190, + 175, + 258, + 211 + ], + "spans": [ + { + "bbox": [ + 190, + 175, + 258, + 211 + ], + "type": "image", + "image_path": "162f7a8fc7d6f7d23f09a1207931df688e23ed0f0aeba914e85f887459ec038f.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 206, + 210, + 243, + 228 + ], + "lines": [ + { + "bbox": [ + 206, + 210, + 243, + 228 + ], + "spans": [ + { + "bbox": [ + 206, + 210, + 243, + 228 + ], + "type": "text", + "content": "PSNR Reference" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 260, + 158, + 328, + 175 + ], + "blocks": [ + { + "bbox": [ + 260, + 158, + 328, + 175 + ], + "lines": [ + { + "bbox": [ + 260, + 158, + 328, + 175 + ], + "spans": [ + { + "bbox": [ + 260, + 158, + 328, + 175 + ], + "type": "image", + "image_path": "0e0f08caf93a4c93c799f7d427d119f3a1b87b2554d2fb1605237037c8bd7668.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 260, + 175, + 328, + 211 + ], + "blocks": [ + { + "bbox": [ + 260, + 175, + 328, + 211 + ], + "lines": [ + { + "bbox": [ + 260, + 175, + 328, + 211 + ], + "spans": [ + { + "bbox": [ + 260, + 175, + 328, + 211 + ], + "type": "image", + "image_path": "45ec03ffa2e49f379143ca7614d60053f9e0f811fd3c203d016790238e600d00.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 278, + 210, + 311, + 228 + ], + "lines": [ + { + "bbox": [ + 278, + 210, + 311, + 228 + ], + "spans": [ + { + "bbox": [ + 278, + 210, + 311, + 228 + ], + "type": "text", + "content": "16.90 dB Bicubic" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 329, + 158, + 397, + 175 + ], + "blocks": [ + { + "bbox": [ + 329, + 158, + 397, + 175 + ], + "lines": [ + { + "bbox": [ + 329, + 158, + 397, + 175 + ], + "spans": [ + { + "bbox": [ + 329, + 158, + 397, + 175 + ], + "type": "image", + "image_path": "d4a4ee7a4c389733d89e2a71ee28d9d7a9dea65d1b66edec6dec075351bf135a.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 329, + 175, + 398, + 211 + ], + "blocks": [ + { + "bbox": [ + 329, + 175, + 398, + 211 + ], + "lines": [ + { + "bbox": [ + 329, + 175, + 398, + 211 + ], + "spans": [ + { + "bbox": [ + 329, + 175, + 398, + 211 + ], + "type": "image", + "image_path": "efc98eb37785400a9c767986a9120da7db0ea56c4c60f8fa2a9822dd34381cf7.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 337, + 210, + 389, + 228 + ], + "lines": [ + { + "bbox": [ + 337, + 210, + 389, + 228 + ], + "spans": [ + { + "bbox": [ + 337, + 210, + 389, + 228 + ], + "type": "text", + "content": "17.54 dB SRCNN [13]" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 399, + 158, + 466, + 175 + ], + "blocks": [ + { + "bbox": [ + 399, + 158, + 466, + 175 + ], + "lines": [ + { + "bbox": [ + 399, + 158, + 466, + 175 + ], + "spans": [ + { + "bbox": [ + 399, + 158, + 466, + 175 + ], + "type": "image", + "image_path": "75746093afebd3051e4e2320e109478ede0540aff4821e59af273b6ab0c491c1.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 399, + 175, + 466, + 211 + ], + "blocks": [ + { + "bbox": [ + 399, + 175, + 466, + 211 + ], + "lines": [ + { + "bbox": [ + 399, + 175, + 466, + 211 + ], + "spans": [ + { + "bbox": [ + 399, + 175, + 466, + 211 + ], + "type": "image", + "image_path": "796ee10c29fb3f1c5f8ea85ef79a72b6f618b87397c99523b7f1a03acde2a15e.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 406, + 210, + 459, + 228 + ], + "lines": [ + { + "bbox": [ + 406, + 210, + 459, + 228 + ], + "spans": [ + { + "bbox": [ + 406, + 210, + 459, + 228 + ], + "type": "text", + "content": "17.71 dB FSRCNN [29]" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 468, + 158, + 536, + 175 + ], + "blocks": [ + { + "bbox": [ + 468, + 158, + 536, + 175 + ], + "lines": [ + { + "bbox": [ + 468, + 158, + 536, + 175 + ], + "spans": [ + { + "bbox": [ + 468, + 158, + 536, + 175 + ], + "type": "image", + "image_path": "3390d5d4392ecc3a6302e05f7f15d37bce49a6bb4d8189486a7f587fdbfd74b0.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 468, + 175, + 536, + 211 + ], + "blocks": [ + { + "bbox": [ + 468, + 175, + 536, + 211 + ], + "lines": [ + { + "bbox": [ + 468, + 175, + 536, + 211 + ], + "spans": [ + { + "bbox": [ + 468, + 175, + 536, + 211 + ], + "type": "image", + "image_path": "24294bddef8c107df51af9be3a5763d13c14f3971dde0d736128786dd077502f.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 483, + 210, + 522, + 228 + ], + "lines": [ + { + "bbox": [ + 483, + 210, + 522, + 228 + ], + "spans": [ + { + "bbox": [ + 483, + 210, + 522, + 228 + ], + "type": "text", + "content": "18.78 dB CARN [1]" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 190, + 228, + 258, + 263 + ], + "blocks": [ + { + "bbox": [ + 190, + 228, + 258, + 263 + ], + "lines": [ + { + "bbox": [ + 190, + 228, + 258, + 263 + ], + "spans": [ + { + "bbox": [ + 190, + 228, + 258, + 263 + ], + "type": "image", + "image_path": "507e9e529c914fcca45fed3d861be593ad8c4931a1d6fff2061d2040b4ae10a3.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 206, + 262, + 246, + 281 + ], + "lines": [ + { + "bbox": [ + 206, + 262, + 246, + 281 + ], + "spans": [ + { + "bbox": [ + 206, + 262, + 246, + 281 + ], + "type": "text", + "content": "18.77 dB \n[MDN [22]]" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_caption" + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 260, + 228, + 328, + 263 + ], + "blocks": [ + { + "bbox": [ + 260, + 228, + 328, + 263 + ], + "lines": [ + { + "bbox": [ + 260, + 228, + 328, + 263 + ], + "spans": [ + { + "bbox": [ + 260, + 228, + 328, + 263 + ], + "type": "image", + "image_path": "05c429fd2742ea720938c2b91ed02202c26231eee65aeda4f572292e5ce0463c.jpg" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 272, + 262, + 316, + 281 + ], + "lines": [ + { + "bbox": [ + 272, + 262, + 316, + 281 + ], + "spans": [ + { + "bbox": [ + 272, + 262, + 316, + 281 + ], + "type": "text", + "content": "19.23 dB RLFN [28]" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_caption" + } + ], + "index": 43 + }, + { + "type": "image", + "bbox": [ + 329, + 228, + 398, + 263 + ], + "blocks": [ + { + "bbox": [ + 329, + 228, + 398, + 263 + ], + "lines": [ + { + "bbox": [ + 329, + 228, + 398, + 263 + ], + "spans": [ + { + "bbox": [ + 329, + 228, + 398, + 263 + ], + "type": "image", + "image_path": "c79a6db4d56adf3eab4c6412ae0c805ff9261eda98ceb04dc7f85b4f9fede4ba.jpg" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 334, + 262, + 391, + 281 + ], + "lines": [ + { + "bbox": [ + 334, + 262, + 391, + 281 + ], + "spans": [ + { + "bbox": [ + 334, + 262, + 391, + 281 + ], + "type": "text", + "content": "18.91 dB LatticeNet [38]" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_caption" + } + ], + "index": 45 + }, + { + "type": "image", + "bbox": [ + 399, + 228, + 466, + 263 + ], + "blocks": [ + { + "bbox": [ + 399, + 228, + 466, + 263 + ], + "lines": [ + { + "bbox": [ + 399, + 228, + 466, + 263 + ], + "spans": [ + { + "bbox": [ + 399, + 228, + 466, + 263 + ], + "type": "image", + "image_path": "89b2d08e02920eeb1a440525897f9db3363759839d30c9c701ac30a394f74c1a.jpg" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 409, + 262, + 456, + 281 + ], + "lines": [ + { + "bbox": [ + 409, + 262, + 456, + 281 + ], + "spans": [ + { + "bbox": [ + 409, + 262, + 456, + 281 + ], + "type": "text", + "content": "19.07 dB SwinIR [31]" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_caption" + } + ], + "index": 47 + }, + { + "type": "image", + "bbox": [ + 468, + 228, + 536, + 263 + ], + "blocks": [ + { + "bbox": [ + 468, + 228, + 536, + 263 + ], + "lines": [ + { + "bbox": [ + 468, + 228, + 536, + 263 + ], + "spans": [ + { + "bbox": [ + 468, + 228, + 536, + 263 + ], + "type": "image", + "image_path": "bb801bd68add06260d770f876872149c032dfc7e783f367e180b9553bded8c1f.jpg" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 469, + 262, + 535, + 281 + ], + "lines": [ + { + "bbox": [ + 469, + 262, + 535, + 281 + ], + "spans": [ + { + "bbox": [ + 469, + 262, + 535, + 281 + ], + "type": "text", + "content": "21.18 dB Omni-SR (Ours)" + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_caption" + } + ], + "index": 49 + }, + { + "type": "image", + "bbox": [ + 190, + 281, + 258, + 316 + ], + "blocks": [ + { + "bbox": [ + 190, + 281, + 258, + 316 + ], + "lines": [ + { + "bbox": [ + 190, + 281, + 258, + 316 + ], + "spans": [ + { + "bbox": [ + 190, + 281, + 258, + 316 + ], + "type": "image", + "image_path": "3624e33137ec1266b291073d03192799a127f08adc46b00c2d2a0dae18128c51.jpg" + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 206, + 316, + 243, + 333 + ], + "lines": [ + { + "bbox": [ + 206, + 316, + 243, + 333 + ], + "spans": [ + { + "bbox": [ + 206, + 316, + 243, + 333 + ], + "type": "text", + "content": "PSNR Reference" + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "image_caption" + } + ], + "index": 51 + }, + { + "type": "image", + "bbox": [ + 260, + 281, + 328, + 316 + ], + "blocks": [ + { + "bbox": [ + 260, + 281, + 328, + 316 + ], + "lines": [ + { + "bbox": [ + 260, + 281, + 328, + 316 + ], + "spans": [ + { + "bbox": [ + 260, + 281, + 328, + 316 + ], + "type": "image", + "image_path": "c3b48d85deeed63cf49264f906f84a21e402e04e2f791f4e32c43fdeaadfec58.jpg" + } + ] + } + ], + "index": 53, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 277, + 316, + 310, + 332 + ], + "lines": [ + { + "bbox": [ + 277, + 316, + 310, + 332 + ], + "spans": [ + { + "bbox": [ + 277, + 316, + 310, + 332 + ], + "type": "text", + "content": "23.58 dB Bicubic" + } + ] + } + ], + "index": 54, + "angle": 0, + "type": "image_caption" + } + ], + "index": 53 + }, + { + "type": "image", + "bbox": [ + 329, + 281, + 397, + 316 + ], + "blocks": [ + { + "bbox": [ + 329, + 281, + 397, + 316 + ], + "lines": [ + { + "bbox": [ + 329, + 281, + 397, + 316 + ], + "spans": [ + { + "bbox": [ + 329, + 281, + 397, + 316 + ], + "type": "image", + "image_path": "e7f906c0018864b466385ffdaa4d9b737212722f7535688b21c97690f0b078ad.jpg" + } + ] + } + ], + "index": 55, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 337, + 316, + 389, + 334 + ], + "lines": [ + { + "bbox": [ + 337, + 316, + 389, + 334 + ], + "spans": [ + { + "bbox": [ + 337, + 316, + 389, + 334 + ], + "type": "text", + "content": "24.04 dB SRCNN [13]" + } + ] + } + ], + "index": 56, + "angle": 0, + "type": "image_caption" + } + ], + "index": 55 + }, + { + "type": "image", + "bbox": [ + 399, + 281, + 466, + 316 + ], + "blocks": [ + { + "bbox": [ + 399, + 281, + 466, + 316 + ], + "lines": [ + { + "bbox": [ + 399, + 281, + 466, + 316 + ], + "spans": [ + { + "bbox": [ + 399, + 281, + 466, + 316 + ], + "type": "image", + "image_path": "d1b7d92000dba7f4252448f42df8ed241decc120487dc1578d8161f5ea35d67b.jpg" + } + ] + } + ], + "index": 57, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 406, + 316, + 459, + 334 + ], + "lines": [ + { + "bbox": [ + 406, + 316, + 459, + 334 + ], + "spans": [ + { + "bbox": [ + 406, + 316, + 459, + 334 + ], + "type": "text", + "content": "24.11 dB FSRCNN [29]" + } + ] + } + ], + "index": 58, + "angle": 0, + "type": "image_caption" + } + ], + "index": 57 + }, + { + "type": "image", + "bbox": [ + 468, + 281, + 536, + 316 + ], + "blocks": [ + { + "bbox": [ + 468, + 281, + 536, + 316 + ], + "lines": [ + { + "bbox": [ + 468, + 281, + 536, + 316 + ], + "spans": [ + { + "bbox": [ + 468, + 281, + 536, + 316 + ], + "type": "image", + "image_path": "b9987c31f09d6af8941ef0f02b263bd6e2e398ccb41239bd0a34ac1182fb7580.jpg" + } + ] + } + ], + "index": 59, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 483, + 316, + 522, + 334 + ], + "lines": [ + { + "bbox": [ + 483, + 316, + 522, + 334 + ], + "spans": [ + { + "bbox": [ + 483, + 316, + 522, + 334 + ], + "type": "text", + "content": "24.48 dB CARN [1]" + } + ] + } + ], + "index": 60, + "angle": 0, + "type": "image_caption" + } + ], + "index": 59 + }, + { + "type": "image", + "bbox": [ + 190, + 334, + 258, + 369 + ], + "blocks": [ + { + "bbox": [ + 190, + 334, + 258, + 369 + ], + "lines": [ + { + "bbox": [ + 190, + 334, + 258, + 369 + ], + "spans": [ + { + "bbox": [ + 190, + 334, + 258, + 369 + ], + "type": "image", + "image_path": "758acc31a79474f0bb5a07acfdaba0efa7ad19c835bae55ae7a5d7f31ee328b7.jpg" + } + ] + } + ], + "index": 61, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 205, + 369, + 246, + 387 + ], + "lines": [ + { + "bbox": [ + 205, + 369, + 246, + 387 + ], + "spans": [ + { + "bbox": [ + 205, + 369, + 246, + 387 + ], + "type": "text", + "content": "24.62 dB IMDN [22]" + } + ] + } + ], + "index": 62, + "angle": 0, + "type": "image_caption" + } + ], + "index": 61 + }, + { + "type": "image", + "bbox": [ + 260, + 334, + 328, + 369 + ], + "blocks": [ + { + "bbox": [ + 260, + 334, + 328, + 369 + ], + "lines": [ + { + "bbox": [ + 260, + 334, + 328, + 369 + ], + "spans": [ + { + "bbox": [ + 260, + 334, + 328, + 369 + ], + "type": "image", + "image_path": "daf5882753d09415d704f856b2b055023f6a358b147b5d07009e844138c0e774.jpg" + } + ] + } + ], + "index": 63, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 272, + 369, + 316, + 387 + ], + "lines": [ + { + "bbox": [ + 272, + 369, + 316, + 387 + ], + "spans": [ + { + "bbox": [ + 272, + 369, + 316, + 387 + ], + "type": "text", + "content": "24.43 dB RLFN [28]" + } + ] + } + ], + "index": 64, + "angle": 0, + "type": "image_caption" + } + ], + "index": 63 + }, + { + "type": "image", + "bbox": [ + 329, + 334, + 397, + 369 + ], + "blocks": [ + { + "bbox": [ + 329, + 334, + 397, + 369 + ], + "lines": [ + { + "bbox": [ + 329, + 334, + 397, + 369 + ], + "spans": [ + { + "bbox": [ + 329, + 334, + 397, + 369 + ], + "type": "image", + "image_path": "dc33abd478c695a711c0927226082f8e01cc8f485c764b97df6d826b745e2a89.jpg" + } + ] + } + ], + "index": 65, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 334, + 369, + 391, + 387 + ], + "lines": [ + { + "bbox": [ + 334, + 369, + 391, + 387 + ], + "spans": [ + { + "bbox": [ + 334, + 369, + 391, + 387 + ], + "type": "text", + "content": "24.64 dB LatticeNet [38]" + } + ] + } + ], + "index": 66, + "angle": 0, + "type": "image_caption" + } + ], + "index": 65 + }, + { + "type": "image", + "bbox": [ + 399, + 334, + 466, + 369 + ], + "blocks": [ + { + "bbox": [ + 399, + 334, + 466, + 369 + ], + "lines": [ + { + "bbox": [ + 399, + 334, + 466, + 369 + ], + "spans": [ + { + "bbox": [ + 399, + 334, + 466, + 369 + ], + "type": "image", + "image_path": "3d47daf6c3df1925794bd00e517e13d1381e1b9f2c13b83091fbbb776e4c2079.jpg" + } + ] + } + ], + "index": 67, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 409, + 369, + 456, + 387 + ], + "lines": [ + { + "bbox": [ + 409, + 369, + 456, + 387 + ], + "spans": [ + { + "bbox": [ + 409, + 369, + 456, + 387 + ], + "type": "text", + "content": "24.45 dB SwinIR [31]" + } + ] + } + ], + "index": 68, + "angle": 0, + "type": "image_caption" + } + ], + "index": 67 + }, + { + "type": "image", + "bbox": [ + 468, + 334, + 536, + 369 + ], + "blocks": [ + { + "bbox": [ + 468, + 334, + 536, + 369 + ], + "lines": [ + { + "bbox": [ + 468, + 334, + 536, + 369 + ], + "spans": [ + { + "bbox": [ + 468, + 334, + 536, + 369 + ], + "type": "image", + "image_path": "2f5a163e9a687d1c82a3b64614fbf786df9a85df92794e741b2c13cfc325abf4.jpg" + } + ] + } + ], + "index": 69, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 469, + 369, + 535, + 387 + ], + "lines": [ + { + "bbox": [ + 469, + 369, + 535, + 387 + ], + "spans": [ + { + "bbox": [ + 469, + 369, + 535, + 387 + ], + "type": "text", + "content": "26.19 dB Omni-SR (Ours)" + } + ] + } + ], + "index": 70, + "angle": 0, + "type": "image_caption" + } + ], + "index": 69 + }, + { + "bbox": [ + 46, + 413, + 287, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 413, + 287, + 472 + ], + "spans": [ + { + "bbox": [ + 46, + 413, + 287, + 472 + ], + "type": "inline_equation", + "content": "\\mathrm{K} = 1" + }, + { + "bbox": [ + 46, + 413, + 287, + 472 + ], + "type": "text", + "content": " increasing the number of OSAG leads to stable performance improvements. In Figure 5(b), we present PSNR vs. parameters of different methods. It can be found that Omni-SR achieves the best results under various settings, showing its effectiveness and scalability." + } + ] + } + ], + "index": 72 + }, + { + "bbox": [ + 47, + 479, + 219, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 479, + 219, + 491 + ], + "spans": [ + { + "bbox": [ + 47, + 479, + 219, + 491 + ], + "type": "text", + "content": "4.3. Analysis of Omni Self-Attention" + } + ] + } + ], + "index": 73 + }, + { + "bbox": [ + 46, + 498, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 288, + 714 + ], + "type": "text", + "content": "In this section, we illustrate the optimization features of OSA and further uncover its underlying mechanism. Self-attention is a low-bias operator, which makes its optimization difficult and requires more training epochs. For this, we introduce the additional channel-wise interaction to alleviate it. In Figure 7(a), we show the loss curves of different self-attention paradigms on the DIV2K training set, including spatial self-attention, channel self-attention, and the proposed omni self-attention. We can see that our OSA presents a obviously superior convergence speed. More importantly, the performance at the final epoch is also significantly ahead of them. The above phenomenon clearly shows that our OSA has superior good optimization characteristics. Further, we delve into why channel-wise interactions lead to these improvements. We calculate the normalized entropies [52] of the hidden layer features of the network composed of the above three computational primitives. We illustrate the entropy results in Figure 7(c)." + } + ] + } + ], + "index": 74 + }, + { + "bbox": [ + 304, + 413, + 545, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 413, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 304, + 413, + 545, + 568 + ], + "type": "text", + "content": "As shown in the figure, in all outgoing layers, our OSA-encoded features show higher entropy, indicating that our OSA encodes richer information. More information may come from various scales, and this information can help the operators to reconstruct the exact details faster. We speculate that this is the potential reason why our OSA shows better optimization properties. In addition, following previous works [8, 16], we also resort to LAM analysis. DI [16] metric can measure the furthest interaction distance of the model. From Figure 8 we can observe that Omni-SR generally has the highest max diffusion index than other methods, showing that our OSA paradigm can effectively capture long-range interactions." + } + ] + } + ], + "index": 75 + }, + { + "bbox": [ + 306, + 575, + 399, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 575, + 399, + 588 + ], + "spans": [ + { + "bbox": [ + 306, + 575, + 399, + 588 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 76 + }, + { + "bbox": [ + 304, + 594, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 715 + ], + "type": "text", + "content": "Effect of Omni Self-Attention. The core idea of our framework is to extend the vanilla self-attention with a channel-wise relationship to build omni-axis pixel-wise interaction. Based on Omni-SR framework, we design several variant models, and their SR results are reported in Table 2. We first simply remove the channel-wise component to form a spatial-only variant " + }, + { + "bbox": [ + 304, + 594, + 545, + 715 + ], + "type": "inline_equation", + "content": "(\\mathrm{Omni - SR_{sp}})" + }, + { + "bbox": [ + 304, + 594, + 545, + 715 + ], + "type": "text", + "content": ", and its performance is degraded by " + }, + { + "bbox": [ + 304, + 594, + 545, + 715 + ], + "type": "inline_equation", + "content": "0.13\\mathrm{dB}" + }, + { + "bbox": [ + 304, + 594, + 545, + 715 + ], + "type": "text", + "content": " compared to the full model. Such a significant degradation justifies the importance of channel interactions. Note that Omni-" + }, + { + "bbox": [ + 304, + 594, + 545, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{SR_{sp}}" + }, + { + "bbox": [ + 304, + 594, + 545, + 715 + ], + "type": "text", + "content": " still" + } + ] + } + ], + "index": 77 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22384" + } + ] + } + ], + "index": 78 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 85, + 70, + 259, + 158 + ], + "blocks": [ + { + "bbox": [ + 85, + 70, + 259, + 158 + ], + "lines": [ + { + "bbox": [ + 85, + 70, + 259, + 158 + ], + "spans": [ + { + "bbox": [ + 85, + 70, + 259, + 158 + ], + "type": "image", + "image_path": "1b026b04daf7a008b04feb302dd3557f9f9b26d00703fc2ce40473e53e2a01b4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 123, + 159, + 241, + 170 + ], + "lines": [ + { + "bbox": [ + 123, + 159, + 241, + 170 + ], + "spans": [ + { + "bbox": [ + 123, + 159, + 241, + 170 + ], + "type": "text", + "content": "(a) Ablation studies on OSAG" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 267, + 69, + 383, + 157 + ], + "blocks": [ + { + "bbox": [ + 267, + 69, + 383, + 157 + ], + "lines": [ + { + "bbox": [ + 267, + 69, + 383, + 157 + ], + "spans": [ + { + "bbox": [ + 267, + 69, + 383, + 157 + ], + "type": "image", + "image_path": "4bdb033e383e968248d934e8e127f02ded6bc64bdeda49ea475346c3e22f556d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 277, + 159, + 385, + 171 + ], + "lines": [ + { + "bbox": [ + 277, + 159, + 385, + 171 + ], + "spans": [ + { + "bbox": [ + 277, + 159, + 385, + 171 + ], + "type": "text", + "content": "(b) Training loss on DIV2K" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 392, + 71, + 504, + 157 + ], + "blocks": [ + { + "bbox": [ + 392, + 71, + 504, + 157 + ], + "lines": [ + { + "bbox": [ + 392, + 71, + 504, + 157 + ], + "spans": [ + { + "bbox": [ + 392, + 71, + 504, + 157 + ], + "type": "image", + "image_path": "3fa37dfb9888c6ddc4427f7895be2ebc9b9110e8f7bde7677047a0794d12b00c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 414, + 159, + 501, + 169 + ], + "lines": [ + { + "bbox": [ + 414, + 159, + 501, + 169 + ], + "spans": [ + { + "bbox": [ + 414, + 159, + 501, + 169 + ], + "type": "text", + "content": "(c) Feature correlation" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 63, + 214, + 272, + 310 + ], + "blocks": [ + { + "bbox": [ + 46, + 176, + 547, + 199 + ], + "lines": [ + { + "bbox": [ + 46, + 176, + 547, + 199 + ], + "spans": [ + { + "bbox": [ + 46, + 176, + 547, + 199 + ], + "type": "text", + "content": "Figure 7. (a) Ablation studies on different components in OSAG. (b) Training loss of different attention paradigms. (c) Feature correlation analysis of different attention paradigms. All the results are reported on DIV2K dataset for " + }, + { + "bbox": [ + 46, + 176, + 547, + 199 + ], + "type": "inline_equation", + "content": "4 \\times" + }, + { + "bbox": [ + 46, + 176, + 547, + 199 + ], + "type": "text", + "content": " SR." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 63, + 214, + 272, + 310 + ], + "lines": [ + { + "bbox": [ + 63, + 214, + 272, + 310 + ], + "spans": [ + { + "bbox": [ + 63, + 214, + 272, + 310 + ], + "type": "image", + "image_path": "e385dad27e053ce2fce368849b68cc7dadc75fce15fae1142ee2c956409b8a27.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 318, + 288, + 351 + ], + "lines": [ + { + "bbox": [ + 46, + 318, + 288, + 351 + ], + "spans": [ + { + "bbox": [ + 46, + 318, + 288, + 351 + ], + "type": "text", + "content": "Figure 8. The distribution of DI values for different methods on Urban100. Top-left points with show a narrow area of interest, and right-bottom points show a large area of interest." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 365, + 287, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 365, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 365, + 287, + 533 + ], + "type": "text", + "content": "outperforms SwinIR by " + }, + { + "bbox": [ + 46, + 365, + 287, + 533 + ], + "type": "inline_equation", + "content": "0.04\\mathrm{dB}@" + }, + { + "bbox": [ + 46, + 365, + 287, + 533 + ], + "type": "text", + "content": " Urban " + }, + { + "bbox": [ + 46, + 365, + 287, + 533 + ], + "type": "inline_equation", + "content": "100\\times 4" + }, + { + "bbox": [ + 46, + 365, + 287, + 533 + ], + "type": "text", + "content": " , which benefits from the global interaction introduced by grid window partition. Similarly, we remove the spatial self-attention component to derive a channel self-attention variant, Omni" + }, + { + "bbox": [ + 46, + 365, + 287, + 533 + ], + "type": "inline_equation", + "content": "\\mathrm{SR}_{\\mathrm{ca}}" + }, + { + "bbox": [ + 46, + 365, + 287, + 533 + ], + "type": "text", + "content": " , and such a modification also leads to undesirable performance degradation. Besides, we use the most widely adopted channel and spatial attention configurations (i.e., SE [19] and CBAM [55]) to act as alternative operators for channel and spatial aggregation. Both substitutions (Omni" + }, + { + "bbox": [ + 46, + 365, + 287, + 533 + ], + "type": "inline_equation", + "content": "\\mathrm{SR}_{\\mathrm{SE}}" + }, + { + "bbox": [ + 46, + 365, + 287, + 533 + ], + "type": "text", + "content": " , Omni-" + }, + { + "bbox": [ + 46, + 365, + 287, + 533 + ], + "type": "inline_equation", + "content": "\\mathrm{SR}_{\\mathrm{CBAM}}" + }, + { + "bbox": [ + 46, + 365, + 287, + 533 + ], + "type": "text", + "content": " ) hurt PNSR performance compared to the full model. The above results show that the specific interaction paradigm (e.g., scalar-based, covariance-based) is equally important, and our channel interaction based on the covariance matrix shows great advantages." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 534, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 288, + 715 + ], + "type": "text", + "content": "Effect of Omni-Scale Aggregation Group. In Omni-SR, we propose a local-meso-global interaction scheme (i.e., OSAG) to pursue progressive feature aggregation. To investigate its effectiveness, we design three different kinds of interaction schemes based on Omni-SR framework: Separated scheme, Mixed scheme, and our fully designed Omni scheme (i.e., our proposed OSAG), and the ablation study results are shown in Figure 7(a). In the figure, we employ different words (e.g., \"Local\", \"Meso+Global\") to represent specific schemes, e.g., \"Local\" denotes using Local-Conv block to replace Meso-OSA and Global-OSA; \"Local+Global\" represents replace original cascaded Meso-OSA and Global-OSA with cascaded Local-Conv and Global-OSA. We can observe that single interaction schemes (e.g., \"Local\") perform the worst. Interestingly," + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 308, + 273, + 545, + 342 + ], + "blocks": [ + { + "bbox": [ + 305, + 213, + 545, + 269 + ], + "lines": [ + { + "bbox": [ + 305, + 213, + 545, + 269 + ], + "spans": [ + { + "bbox": [ + 305, + 213, + 545, + 269 + ], + "type": "text", + "content": "Table 2. Ablation studies of omni self-attention on Urban100. " + }, + { + "bbox": [ + 305, + 213, + 545, + 269 + ], + "type": "inline_equation", + "content": "\\mathrm{Omni - SR}_{(*)}" + }, + { + "bbox": [ + 305, + 213, + 545, + 269 + ], + "type": "text", + "content": " denotes different modifications. 'SA' and 'S-SA' denote spatial gate and spatial self-attention. 'CA' and 'C-SA' denote channel gate and channel self-attention. We maintain the consistency of model parameters by adjusting the channels of " + }, + { + "bbox": [ + 305, + 213, + 545, + 269 + ], + "type": "inline_equation", + "content": "1\\times 1Conv" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 273, + 545, + 342 + ], + "lines": [ + { + "bbox": [ + 308, + 273, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 308, + 273, + 545, + 342 + ], + "type": "table", + "html": "
ModelSAS-SACAC-SAFLOPs×2×3×4
Omni-SRsp×✓××33G32.8828.7226.51
Omni-SRSE×✓✓×34G32.8328.7126.50
Omni-SRCBAM✓×✓✓34G32.9228.7626.53
Omni-SRca×××✓33G32.6528.6026.45
Omni-SRfull×✓×✓36G33.0528.8426.64
", + "image_path": "8b9d5ef9b1c65f2a01d3b4733d1c81e1ef80bd538d4264f6a6eb46b2ffcc694e.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 354, + 545, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 354, + 545, + 486 + ], + "spans": [ + { + "bbox": [ + 304, + 354, + 545, + 486 + ], + "type": "text", + "content": "the \"Global\" scheme is inferior to the \"Meso\" one due to its poor optimization properties of global self-attention [3, 35, 50]. Once two interaction operators are combined, the performance improves steadily. Among them, \"Meso+ Global\" setting achieves the second-best performance. Furthermore, combining all three interaction schemes together, we obtain the best performing scheme, i.e., \"Omni\". From the above experiments, we can infer that obvious performance gains can be obtained by introducing various-scale interactions, which also illustrates the feasibility and effectiveness of our proposed OSAG." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 497, + 378, + 509 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 497, + 378, + 509 + ], + "spans": [ + { + "bbox": [ + 306, + 497, + 378, + 509 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 517, + 545, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 517, + 545, + 636 + ], + "spans": [ + { + "bbox": [ + 304, + 517, + 545, + 636 + ], + "type": "text", + "content": "In this work, we propose Omni-SR, a lightweight framework for image SR. We propose the Omni Self-attention paradigm for simultaneous spatial and channel interactions, mining all the potential correlations across omni-axis. Furthermore, we propose an omni-scale aggregation scheme to effectively enlarge the receptive fields with low computational complexity, which encodes contextual relations in a progressively hierarchical manner. Extensive experiments on public benchmark datasets and comprehensive analytical studies validate its prominent SR performance." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 646, + 416, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 646, + 416, + 659 + ], + "spans": [ + { + "bbox": [ + 306, + 646, + 416, + 659 + ], + "type": "text", + "content": "6. Acknowledgement" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 712 + ], + "type": "text", + "content": "This work was supported by National Science Foundation of China (U20B2072, 61976137). This work was also partly supported by SJTU Medical Engineering Cross Research Grant YG2021ZD18." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22385" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "text", + "content": "[1] Namhyuk Ahn, Byungkon Kang, and Kyung-Ah Sohn. Fast, accurate, and lightweight super-resolution with cascading residual network. In ECCV, 2018. 2, 5, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 124, + 286, + 146 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 124, + 286, + 146 + ], + "spans": [ + { + "bbox": [ + 53, + 124, + 286, + 146 + ], + "type": "text", + "content": "[2] Lei Jimmy Ba, Jamie Ryan Kiros, and Geoffrey E. Hinton. Layer normalization. CoRR, abs/1607.06450, 2016. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 148, + 286, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 286, + 178 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 286, + 178 + ], + "type": "text", + "content": "[3] Irwan Bello, Barret Zoph, Quoc Le, Ashish Vaswani, and Jonathon Shlens. Attention augmented convolutional networks. In ICCV, 2019. 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 180, + 286, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 180, + 286, + 223 + ], + "spans": [ + { + "bbox": [ + 53, + 180, + 286, + 223 + ], + "type": "text", + "content": "[4] Marco Bevilacqua, Aline Roumy, Christine Guillemot, and Marie-Line Alberi-Morel. Low-complexity single-image super-resolution based on nonnegative neighbor embedding. In BMVC, 2012. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 224, + 286, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 224, + 286, + 267 + ], + "spans": [ + { + "bbox": [ + 53, + 224, + 286, + 267 + ], + "type": "text", + "content": "[5] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In CVPR, 2021. 1, 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 269, + 286, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 269, + 286, + 312 + ], + "spans": [ + { + "bbox": [ + 53, + 269, + 286, + 312 + ], + "type": "text", + "content": "[6] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. arXiv:1606.00915, 2016. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 314, + 286, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 314, + 286, + 346 + ], + "spans": [ + { + "bbox": [ + 53, + 314, + 286, + 346 + ], + "type": "text", + "content": "[7] Xuanhong Chen, Hang Wang, and Bingbing Ni. X-volution: On the unification of convolution and self-attention. arXiv preprint arXiv:2106.02253, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 347, + 286, + 379 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 347, + 286, + 379 + ], + "spans": [ + { + "bbox": [ + 53, + 347, + 286, + 379 + ], + "type": "text", + "content": "[8] Xiangyu Chen, Xintao Wang, Jiantao Zhou, and Chao Dong. Activating more pixels in image super-resolution transformer. CoRR, abs/2205.04437, 2022. 2, 4, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 380, + 286, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 286, + 422 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 286, + 422 + ], + "type": "text", + "content": "[9] Yinpeng Chen, Xiyang Dai, Dongdong Chen, Mengchen Liu, Xiaoyi Dong, Lu Yuan, and Zicheng Liu. Mobileformer: Bridging mobilenet and transformer. In CVPR, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 425, + 286, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 425, + 286, + 456 + ], + "spans": [ + { + "bbox": [ + 48, + 425, + 286, + 456 + ], + "type": "text", + "content": "[10] Yinpeng Chen, Xiyang Dai, Dongdong Chen, Mengchen Liu, Xiaoyi Dong, Lu Yuan, and Zicheng Liu. Simple baselines for image restoration. In ECCV, 2022. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 458, + 286, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 458, + 286, + 490 + ], + "spans": [ + { + "bbox": [ + 48, + 458, + 286, + 490 + ], + "type": "text", + "content": "[11] Xiangxiang Chu, Bo Zhang, Hailong Ma, Ruijun Xu, and Qingyuan Li. Fast, accurate and lightweight super-resolution with neural architecture search. In ICPR, 2020. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 491, + 286, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 491, + 286, + 523 + ], + "spans": [ + { + "bbox": [ + 48, + 491, + 286, + 523 + ], + "type": "text", + "content": "[12] Xiaohan Ding, Xiangyu Zhang, Jungong Han, and Guiguang Ding. Scaling up your kernels to 31x31: Revisiting large kernel design in cnns. In CVPR, 2022. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 525, + 286, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 286, + 556 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 286, + 556 + ], + "type": "text", + "content": "[13] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Image super-resolution using deep convolutional networks. CoRR, abs/1501.00092, 2015. 2, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 558, + 286, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 558, + 286, + 623 + ], + "spans": [ + { + "bbox": [ + 48, + 558, + 286, + 623 + ], + "type": "text", + "content": "[14] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 624, + 286, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 286, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 286, + 656 + ], + "type": "text", + "content": "[15] Qinquan Gao, Yan Zhao, Gen Li, and Tong Tong. Image super-resolution using knowledge distillation. In ACCV, 2018. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 658, + 286, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 286, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 286, + 679 + ], + "type": "text", + "content": "[16] Jinjin Gu and Chao Dong. Interpreting super-resolution networks with local attribution maps. In CVPR, 2021. 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 681, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 286, + 712 + ], + "type": "text", + "content": "[17] Geoffrey E. Hinton, Oriol Vinyals, and Jeffrey Dean. Distilling the knowledge in a neural network. CoRR, abs/1503.02531, 2015. 2" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 37, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 127 + ], + "type": "text", + "content": "[18] Andrew G Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andretto, and Hartwig Adam. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv:1704.04861, 2017. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 129, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 150 + ], + "type": "text", + "content": "[19] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In CVPR, 2018. 2, 3, 4, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 152, + 545, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 184 + ], + "type": "text", + "content": "[20] Jia-Bin Huang, Abhishek Singh, and Narendra Ahuja. Single image super-resolution from transformed self-exemplars. In CVPR, 2015. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 186, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 217 + ], + "type": "text", + "content": "[21] Tao Huang, Lang Huang, Shan You, Fei Wang, Chen Qian, and Chang Xu. Lightvit: Towards light-weight convolution-free vision transformers. CoRR, abs/2207.05557, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 219, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 219, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 219, + 545, + 251 + ], + "type": "text", + "content": "[22] Zheng Hui, Xinbo Gao, Yunchu Yang, and Xiumei Wang. Lightweight image super-resolution with information multistillation network. In ACMMM, 2019. 2, 5, 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 253, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 253, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 253, + 545, + 285 + ], + "type": "text", + "content": "[23] Zheng Hui, Xiumei Wang, and Xinbo Gao. Fast and accurate single image super-resolution via information distillation network. In CVPR, 2018. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 286, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 286, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 307, + 286, + 545, + 330 + ], + "type": "text", + "content": "[24] Forrest N Iandola, Song Han, Matthew W Moskewicz, Khalid Ashraf, William J Dally, and Kurt Keutzer. SqueezeNet: Alexnet-level accuracy with 50x fewer parameters and; 0.5 mb model size. arXiv:1602.07360, 2016. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 331, + 545, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 331, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 307, + 331, + 545, + 364 + ], + "type": "text", + "content": "[25] Jiwon Kim, Jung Kwon Lee, and Kyoung Mu Lee. Accurate image super-resolution using very deep convolutional networks. In CVPR, 2016. 2, 5, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 365, + 545, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 545, + 397 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 545, + 397 + ], + "type": "text", + "content": "[26] Jiwon Kim, Jung Kwon Lee, and Kyoung Mu Lee. Deeplyrecursive convolutional network for image super-resolution. In CVPR, 2016. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 399, + 545, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 545, + 421 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 545, + 421 + ], + "type": "text", + "content": "[27] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 422, + 545, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 422, + 545, + 465 + ], + "spans": [ + { + "bbox": [ + 307, + 422, + 545, + 465 + ], + "type": "text", + "content": "[28] Fangyuan Kong, Mingxi Li, Songwei Liu, Ding Liu, Jingwen He, Yang Bai, Fangmin Chen, and Lean Fu. Residual local feature network for efficient super-resolution. In CVPR Workshops, 2022. 2, 5, 6, 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 467, + 545, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 467, + 545, + 509 + ], + "spans": [ + { + "bbox": [ + 307, + 467, + 545, + 509 + ], + "type": "text", + "content": "[29] Wei-Sheng Lai, Jia-Bin Huang, Narendra Ahuja, and Ming-Hsuan Yang. Fast and accurate image super-resolution with deep laplacian pyramid networks. TPAMI, 41(11):2599-2613, 2018. 7" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 511, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 511, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 307, + 511, + 545, + 555 + ], + "type": "text", + "content": "[30] Wenbo Li, Xin Lu, Shengju Qian, Jiangbo Lu, Xiangyu Zhang, and Jiaya Jia. On efficient transformer and image pre-training for low-level vision. arXiv preprint arXiv:2112.10175, 2021. 2, 5" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 556, + 545, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 556, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 307, + 556, + 545, + 589 + ], + "type": "text", + "content": "[31] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In ICCV, 2021. 1, 2, 4, 5, 6, 7" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 590, + 545, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 590, + 545, + 623 + ], + "spans": [ + { + "bbox": [ + 307, + 590, + 545, + 623 + ], + "type": "text", + "content": "[32] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In CVPRW, 2017. 5, 6" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 307, + 624, + 545, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 656 + ], + "type": "text", + "content": "[33] Jie Liu, Jie Tang, and Gangshan Wu. Residual feature distillation network for lightweight image super-resolution. In ECCV, 2020. 5, 6" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 307, + 658, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 658, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 658, + 545, + 689 + ], + "type": "text", + "content": "[34] Jie Liu, Wenjie Zhang, Yuting Tang, Jie Tang, and Gangshan Wu. Residual feature aggregation network for image superresolution. In CVPR, 2020. 5" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "text", + "content": "[35] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer:" + } + ] + } + ], + "index": 36 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "22386" + } + ] + } + ], + "index": 38 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 95 + ], + "type": "text", + "content": "Hierarchical vision transformer using shifted windows. In ICCV, 2021. 2, 4, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 117 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 117 + ], + "type": "text", + "content": "[36] Ilya Loshchilov and Frank Hutter. SGDR: stochastic gradient descent with warm restarts. In *ICLR*, 2017. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 119, + 287, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 119, + 287, + 152 + ], + "spans": [ + { + "bbox": [ + 48, + 119, + 287, + 152 + ], + "type": "text", + "content": "[37] Zhisheng Lu, Juncheng Li, Hong Liu, Chaoyan Huang, Linlin Zhang, and Tieyong Zeng. Transformer for single image super-resolution. In CVPR Workshops, 2022. 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 153, + 287, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 153, + 287, + 195 + ], + "spans": [ + { + "bbox": [ + 48, + 153, + 287, + 195 + ], + "type": "text", + "content": "[38] Xiaotong Luo, Yuan Xie, Yulun Zhang, Yanyun Qu, Cuihua Li, and Yun Fu. Latticenet: Towards lightweight image super-resolution with lattice block. In ECCV, 2020. 2, 5, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 198, + 287, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 198, + 287, + 241 + ], + "spans": [ + { + "bbox": [ + 48, + 198, + 287, + 241 + ], + "type": "text", + "content": "[39] David Martin, Charless Fowlkes, Doron Tal, and Jitendra Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In ICCV. IEEE, 2001. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 243, + 287, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 243, + 287, + 296 + ], + "spans": [ + { + "bbox": [ + 48, + 243, + 287, + 296 + ], + "type": "text", + "content": "[40] Yusuke Matsui, Kota Ito, Yuji Aramaki, Azuma Fujimoto, Toru Ogawa, Toshihiko Yamasaki, and Kiyoharu Aizawa. Sketch-based manga retrieval using manga109 dataset. Multimedia Tools and Applications, 76(20):21811-21838, 2017. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 299, + 287, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 299, + 287, + 331 + ], + "spans": [ + { + "bbox": [ + 48, + 299, + 287, + 331 + ], + "type": "text", + "content": "[41] Sachin Mehta and Mohammad Rastegari. Mobilevit: Lightweight, general-purpose, and mobile-friendly vision transformer. CoRR, abs/2110.02178, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 333, + 287, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 333, + 287, + 374 + ], + "spans": [ + { + "bbox": [ + 48, + 333, + 287, + 374 + ], + "type": "text", + "content": "[42] Abdul Muqueet, Jiwon Hwang, Subin Yang, JungHeum Kang, Yongwoo Kim, and Sung-Ho Bae. Multi-attention based ultra lightweight image super-resolution. In ECCV, 2020. 2, 5, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 377, + 287, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 377, + 287, + 420 + ], + "spans": [ + { + "bbox": [ + 48, + 377, + 287, + 420 + ], + "type": "text", + "content": "[43] Junting Pan, Adrian Bulat, Fuwen Tan, Xiatian Zhu, Lukasz Dudziak, Hongsheng Li, Georgios Tzimiropoulos, and Brais Martínez. Edgevits: Competing light-weight cnns on mobile devices with vision transformers. In ECCV, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 422, + 287, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 422, + 287, + 454 + ], + "spans": [ + { + "bbox": [ + 48, + 422, + 287, + 454 + ], + "type": "text", + "content": "[44] Jongchan Park, Sanghyun Woo, Joon-Young Lee, and In So Kweon. Bam: Bottleneck attention module. arXiv:1807.06514, 2018. 3, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 456, + 287, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 287, + 499 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 287, + 499 + ], + "type": "text", + "content": "[45] Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, Edward Yang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga, and Adam Lerer. Automatic differentiation in pytorch. In NeurIPS Workshop, 2017. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 501, + 287, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 501, + 287, + 554 + ], + "spans": [ + { + "bbox": [ + 48, + 501, + 287, + 554 + ], + "type": "text", + "content": "[46] Wenzhe Shi, Jose Caballero, Ferenc Huszár, Johannes Totz, Andrew P Aitken, Rob Bishop, Daniel Rueckert, and Zehan Wang. Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In CVPR, 2016. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 557, + 287, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 557, + 287, + 588 + ], + "spans": [ + { + "bbox": [ + 48, + 557, + 287, + 588 + ], + "type": "text", + "content": "[47] Ying Tai, Jian Yang, and Xiaoming Liu. Image superresolution via deep recursive residual network. In CVPR, 2017. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 591, + 287, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 591, + 287, + 623 + ], + "spans": [ + { + "bbox": [ + 48, + 591, + 287, + 623 + ], + "type": "text", + "content": "[48] Ying Tai, Jian Yang, Xiaoming Liu, and Chunyan Xu. Memnet: A persistent memory network for image restoration. In ICCV, 2017. 2, 5, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 667 + ], + "type": "text", + "content": "[49] Radu Timofte, Eirikur Agustsson, Luc Van Gool, Ming-Hsuan Yang, and Lei Zhang. Ntire 2017 challenge on single image super-resolution: Methods and results. In CVPR workshops, 2017. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[50] Ashish Vaswani, Prajit Ramachandran, Aravind Srinivas, Niki Parmar, Blake Hechtman, and Jonathon Shlens. Scaling local self-attention for parameter efficient visual backbones. In CVPR, 2021. 2, 8" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 669 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[51] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NIPS, 2017. 1, 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 108, + 545, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 545, + 138 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 545, + 138 + ], + "type": "text", + "content": "[52] Pei Wang, Yijun Li, and Nuno Vasconcelos. Rethinking and improving the robustness of image style transfer. In CVPR, 2021. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 141, + 545, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 174 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 174 + ], + "type": "text", + "content": "[53] Xiaohang Wang, Xuanhong Chen, Bingbing Ni, Zhengyan Tong, and Hang Wang. Learning continuous depth representation via geometric spatial aggregator. In AAAI, 2023. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "type": "text", + "content": "[54] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 220, + 545, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 545, + 251 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 545, + 251 + ], + "type": "text", + "content": "[55] Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon. Cbam: Convolutional block attention module. In ECCV, 2018. 3, 4, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 254, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 254, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 254, + 545, + 285 + ], + "type": "text", + "content": "[56] Tete Xiao, Mannat Singh, Eric Mintun, Trevor Darrell, Piotr Dólar, and Ross B. Girshick. Early convolutions help transformers see better. In NeurIPS, 2021. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 288, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 288, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 288, + 545, + 319 + ], + "type": "text", + "content": "[57] Chenglin Yang, Yilin Wang, Jianming Zhang, He Zhang, Zijun Wei, Zhe Lin, and Alan L. Yuille. Lite vision transformer with enhanced self-attention. In CVPR, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 322, + 545, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 322, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 307, + 322, + 545, + 342 + ], + "type": "text", + "content": "[58] Jianwei Yang, Chunyuan Li, Xiyang Dai, and Jianfeng Gao. Focal modulation networks, 2022. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 344, + 545, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 344, + 545, + 387 + ], + "spans": [ + { + "bbox": [ + 307, + 344, + 545, + 387 + ], + "type": "text", + "content": "[59] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In CVPR, 2022. 1, 3, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 389, + 545, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 389, + 545, + 432 + ], + "spans": [ + { + "bbox": [ + 307, + 389, + 545, + 432 + ], + "type": "text", + "content": "[60] Roman Zeyde, Michael Elad, and Matan Protter. On single image scale-up using sparse-representations. In International conference on curves and surfaces, pages 711-730. Springer, 2010. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 434, + 545, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 434, + 545, + 465 + ], + "spans": [ + { + "bbox": [ + 307, + 434, + 545, + 465 + ], + "type": "text", + "content": "[61] Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In ICLR, 2018. 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 468, + 545, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 468, + 545, + 499 + ], + "spans": [ + { + "bbox": [ + 307, + 468, + 545, + 499 + ], + "type": "text", + "content": "[62] Haokui Zhang, Wenze Hu, and Xiaoyu Wang. Parc-net: Position aware circular convolution with merits from convnets and transformer. In ECCV, 2022. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 502, + 545, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 502, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 307, + 502, + 545, + 533 + ], + "type": "text", + "content": "[63] Kai Zhang, Wangmeng Zuo, Shuhang Gu, and Lei Zhang. Learning deep cnn denoiser prior for image restoration. In CVPR, 2017. 5" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 536, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 536, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 307, + 536, + 545, + 567 + ], + "type": "text", + "content": "[64] Kai Zhang, Wangmeng Zuo, and Lei Zhang. Learning a single convolutional super-resolution network for multiple degradations. In CVPR, 2018. 6" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 570, + 545, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 570, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 307, + 570, + 545, + 601 + ], + "type": "text", + "content": "[65] Yiman Zhang, Hanting Chen, Xinghao Chen, Yiping Deng, Chunjing Xu, and Yunhe Wang. Data-free knowledge distillation for image super-resolution. In CVPR, 2021. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 604, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 604, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 307, + 604, + 545, + 635 + ], + "type": "text", + "content": "[66] Yulun Zhang, Kunpeng Li, Kai Li, Lichen Wang, Bineng Zhong, and Yun Fu. Image super-resolution using very deep residual channel attention networks. In ECCV, 2018. 1, 2, 5" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 637, + 545, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 637, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 307, + 637, + 545, + 669 + ], + "type": "text", + "content": "[67] Yulun Zhang, Yapeng Tian, Yu Kong, Bineng Zhong, and Yun Fu. Residual dense network for image super-resolution. In CVPR, 2018. 5" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22387" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Omni3D_ A Large Benchmark and Model for 3D Object Detection in the Wild/2efbd142-17ac-4a1f-b83f-1ae025c49292_content_list.json b/2023/Omni3D_ A Large Benchmark and Model for 3D Object Detection in the Wild/2efbd142-17ac-4a1f-b83f-1ae025c49292_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c5e34d5f2411e0a34323810c44d3ad1cb650d6a0 --- /dev/null +++ b/2023/Omni3D_ A Large Benchmark and Model for 3D Object Detection in the Wild/2efbd142-17ac-4a1f-b83f-1ae025c49292_content_list.json @@ -0,0 +1,1739 @@ +[ + { + "type": "text", + "text": "OMNI3D: A Large Benchmark and Model for 3D Object Detection in the Wild", + "text_level": 1, + "bbox": [ + 88, + 130, + 880, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Garrick Brazil1 Abhinav Kumar2 Julian Straub1 Nikhila Ravi1 Justin Johnson1 Georgia Gkioxari3", + "bbox": [ + 196, + 179, + 746, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Meta AI $^{2}$ Michigan State University $^{3}$ Caltech", + "bbox": [ + 277, + 226, + 684, + 244 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cdc5e4b3a711848ba026a9f90d0ae7eb0318877f0b52fa7ff8fa6e6af19cd00d.jpg", + "image_caption": [ + "Figure 1. Left: We introduce OMNI3D, a benchmark for 3D object detection which is larger and more diverse than popular 3D benchmarks. Right: We propose Cube R-CNN, which generalizes to unseen datasets (e.g. COCO [45]) and outperforms prior works on existing datasets." + ], + "image_footnote": [], + "bbox": [ + 81, + 282, + 308, + 431 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b2e4cd0d07e95acf9efe8fa561fe11ac96e94ea29c63c3de6bdc344f252986b5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 316, + 286, + 496, + 429 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d7c28d109f6884ad9edc9e375242ea1b28f20535f1caa7d0f9cd691613cd76d8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 286, + 707, + 430 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c54b7910c8ec5a1763fc91285dbf156997eae3fc5de7482371add9b80e36c733.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 718, + 286, + 890, + 428 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 491, + 312, + 507 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recognizing scenes and objects in 3D from a single image is a longstanding goal of computer vision with applications in robotics and AR/VR. For 2D recognition, large datasets and scalable solutions have led to unprecedented advances. In 3D, existing benchmarks are small in size and approaches specialize in few object categories and specific domains, e.g. urban driving scenes. Motivated by the success of 2D recognition, we revisit the task of 3D object detection by introducing a large benchmark, called OMNI3D. OMNI3D re-purposes and combines existing datasets resulting in 234k images annotated with more than 3 million instances and 98 categories. 3D detection at such scale is challenging due to variations in camera intrinsic and the rich diversity of scene and object types. We propose a model, called Cube R-CNN, designed to generalize across camera and scene types with a unified approach. We show that Cube R-CNN outperforms prior works on the larger OMNI3D and existing benchmarks. Finally, we prove that OMNI3D is a powerful dataset for 3D object recognition and show that it improves single-dataset performance and can accelerate learning on new smaller datasets via pre-training.1", + "bbox": [ + 75, + 525, + 473, + 843 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 501, + 491, + 632, + 507 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Understanding objects and their properties from single images is a longstanding problem in computer vision with applications in robotics and AR/VR. In the last decade, 2D object recognition [26,40,63,64,72] has made tremendous advances toward predicting objects on the image plane with the help of large datasets [24,45]. However, the world and its objects are three dimensional laid out in 3D space. Perceiving objects in 3D from 2D visual inputs poses new challenges framed by the task of 3D object detection. Here, the goal is to estimate a 3D location and 3D extent of each object in an image in the form of a tight oriented 3D bounding box.", + "bbox": [ + 496, + 518, + 893, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Today 3D object detection is studied under two different lenses: for urban domains in the context of autonomous vehicles [7,13,50,52,57] or indoor scenes [31,36,58,73]. Despite the problem formulation being shared, methods share little insights between domains. Often approaches are tailored to work only for the domain in question. For instance, urban methods make assumptions about objects resting on a ground plane and model only yaw angles for 3D rotation. Indoor techniques may use a confined depth range (e.g. up to $6\\mathrm{m}$ in [58]). These assumptions are generally not true in the real world. Moreover, the most popular benchmarks for image-based 3D object detection are small. Indoor SUN RGB-D [70] has $10\\mathrm{k}$ images, urban KITTI [21] has $7\\mathrm{k}$ images; 2D benchmarks like COCO [45] are $20\\times$ larger.", + "bbox": [ + 496, + 688, + 895, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1We release the OMNI3D benchmark and Cube R-CNN models at https://github.com/facebookresearch/omni3d.", + "bbox": [ + 76, + 875, + 444, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "13154", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We address the absence of a general large-scale dataset for 3D object detection by introducing a large and diverse 3D benchmark called OMNI3D. OMNI3D is curated from publicly released datasets, SUN RBG-D [70], ARKitScenes [6], Hypersim [65], Objectron [2], KITTI [21] and nuScenes [9], and comprises 234k images with 3 million objects annotated with 3D boxes across 98 categories including chair, sofa, laptop, table, cup, shoes, pillow, books, car, person, etc. Sec. 3 describes the curation process which involves re-purposing the raw data and annotations from the aforementioned datasets, which originally target different applications. As shown in Fig. 1, OMNI3D is $20\\times$ larger than existing popular benchmarks used for 3D detection, SUN RGB-D and KITTI. For efficient evaluation on the large OMNI3D, we introduce a new algorithm for intersection-over-union of 3D boxes which is $450\\times$ faster than previous solutions [2]. We empirically prove the impact of OMNI3D as a large-scale dataset and show that it improves single-dataset performance by up to $5.3\\%$ AP on urban and $3.8\\%$ on indoor benchmarks.", + "bbox": [ + 76, + 90, + 472, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "On the large and diverse OMNI3D, we design a general and simple 3D object detector, called Cube R-CNN, inspired by advances in 2D and 3D recognition of recent years [22,52,64,68]. Cube R-CNN detects all objects and their 3D location, size and rotation end-to-end from a single image of any domain and for many object categories. Attributed to OMNI3D's diversity, our model shows strong generalization and outperforms prior works for indoor and urban domains with one unified model, as shown in Fig. 1. Learning from such diverse data comes with challenges as OMNI3D contains images of highly varying focal lengths which exaggerate scale-depth ambiguity (Fig. 4). We remedy this by operating on virtual depth which transforms object depth with the same virtual camera intrinsics across the dataset. An added benefit of virtual depth is that it allows the use of data augmentations (e.g. image rescaling) during training, which is a critical feature for 2D detection [12,80], and as we show, also for 3D. Our approach with one unified design outperforms prior best approaches in $\\mathrm{AP}_{\\mathrm{3D}}$ , ImVoxelNet [66] by $4.1\\%$ on indoor SUN RGB-D, GUPNet [52] by $9.5\\%$ on urban KITTI, and PGD [77] by $7.9\\%$ on OMNI3D.", + "bbox": [ + 76, + 378, + 472, + 695 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We summarize our contributions:", + "bbox": [ + 96, + 696, + 318, + 710 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce OMNI3D, a benchmark for image-based 3D object detection sourced from existing 3D datasets, which is $20 \\times$ larger than existing 3D benchmarks.", + "- We implement a new algorithm for IoU of 3D boxes, which is $450 \\times$ faster than prior solutions.", + "- We design a general-purpose baseline method, Cube R-CNN, which tackles 3D object detection for many categories and across domains with a unified approach. We propose virtual depth to eliminate the ambiguity from varying camera focal lengths in OMNI3D." + ], + "bbox": [ + 96, + 723, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 89, + 640, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Cube R-CNN and OMNI3D draw from key research advances in 2D and 3D object detection.", + "bbox": [ + 498, + 114, + 893, + 146 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2D Object Detection. Here, methods include two-stage approaches [26, 64] which predict object regions with a region proposal network (RPN) and then refine them via an MLP. Single-stage detectors [44, 47, 63, 72, 85] omit the RPN and predict regions directly from the backbone.", + "bbox": [ + 498, + 150, + 893, + 226 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D Object Detection. Monocular 3D object detectors predict 3D cuboids from single input images. There is extensive work in the urban self-driving domain where the car class is at the epicenter [13, 19, 23, 25, 30, 46, 49, 54, 57, 62, 68, 69, 76, 78, 84]. CenterNet [85] predicts 3D depth and size from fully-convolutional center features, and is extended by [14, 37, 42, 48, 50–52, 55, 83, 84, 87]. M3D-RPN [7] trains an RPN with 3D anchors, enhanced further by [8, 18, 38, 75, 88]. FCOS3D [78] extends the anchorless FCOS [72] detector to predict 3D cuboids. Its successor PGD [77] furthers the approach with probabilistic depth uncertainty. Others use pseudo depth [4, 15, 53, 59, 79, 81] and explore depth and point-based LiDAR techniques [35, 60]. Similar to ours, [11, 68, 69] add a 3D head, specialized for urban scenes and objects, on two-stage Faster R-CNN. [25, 69] augment their training by synthetically generating depth and box-fitted views, coined as virtual views or depth. In our work, virtual depth aims at addressing varying focal lengths.", + "bbox": [ + 498, + 229, + 893, + 502 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For indoor scenes, a vast line of work tackles room layout estimation [17,28,41,56]. Huang et al [31] predict 3D oriented bounding boxes for indoor objects. Factored3D [73] and 3D-RelNet [36] jointly predict object voxel shapes. Total3D [58] predicts 3D boxes and meshes by additionally training on datasets with annotated 3D shapes. ImVoxelNet [66] proposes domain-specific methods which share an underlying framework for processing volumes of 3D voxels. In contrast, we explore 3D object detection in its general form by tackling outdoor and indoor domains jointly in a single model and with a vocabulary of $5 \\times$ more categories.", + "bbox": [ + 498, + 503, + 893, + 670 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D Datasets. KITTI [21] and SUN RGB-D [70] are popular datasets for 3D object detection on urban and indoor scenes respectively. Since 2019, 3D datasets have emerged, both for indoor [2, 3, 6, 16, 65] and outdoor [9, 10, 20, 29, 32, 33, 71]. In isolation, these datasets target different tasks and applications and have unique properties and biases, e.g. object and scene types, focal length, coordinate systems, etc. In this work, we unify existing representative datasets [2, 6, 9, 21, 65, 70]. We process the raw visual data, re-purpose their annotations, and carefully curate the union of their semantic labels in order to build a coherent large-scale benchmark, called OMNI3D. OMNI3D is $20 \\times$ larger than widely-used benchmarks and notably more diverse. As such, new challenges arise stemming from the increased variance in visual domain, object rotation, size, layouts, and camera intrinsics.", + "bbox": [ + 498, + 672, + 895, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "13155", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/789db86a26df8ee5fbd79eebf2d7ad9f27b912998be6b082816bdf1e6913a452.jpg", + "image_caption": [ + "(a) Spatial statistics" + ], + "image_footnote": [], + "bbox": [ + 76, + 85, + 444, + 223 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4063bb12980d1e73898db058fa5b56c72a5b320cca0603175613ba5e926cfce7.jpg", + "image_caption": [ + "(b) 2D Scale statistics", + "Figure 2. OMNI3D analysis. (a) distribution of object centers on (top) normalized image, XY-plane, and (bottom) normalized depth, XZ-plane, (b) relative 2D object scales, and (c) category frequency." + ], + "image_footnote": [], + "bbox": [ + 446, + 106, + 656, + 219 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0e975865c5d381f2f1f79180d494a13640ed937d56f993f333c2d1dd2d6c11f5.jpg", + "image_caption": [ + "(c) Category statistics" + ], + "image_footnote": [], + "bbox": [ + 658, + 107, + 872, + 219 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. The OMNI3D Benchmark", + "text_level": 1, + "bbox": [ + 75, + 287, + 318, + 304 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The primary benchmarks for 3D object detection are small, focus on a few categories and are of a single domain. For instance, the popular KITTI [21] contains only urban scenes, has 7k images and 3 categories, with a focus on car. SUN RGB-D [70] has 10k images. The small size and lack of variance in 3D datasets is a stark difference to 2D counterparts, such as COCO [45] and LVIS [24], which have pioneered progress in 2D recognition.", + "bbox": [ + 73, + 314, + 470, + 434 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We aim to bridge the gap to 2D by introducing OMNI3D, a large-scale and diverse benchmark for image-based 3D object detection consisting of $234\\mathrm{k}$ images, 3 million labeled 3D bounding boxes, and 98 object categories. We source from recently released 3D datasets of urban (nuScenes [9] and KITTI [21]), indoor (SUN RGB-D [70], ARKitScenes [6] and Hypersim [65]), and general (Objectron [2]) scenes. Each of these datasets target different applications (e.g. point-cloud recognition or reconstruction in [6], inverse rendering in [65]), provide visual data in different forms (e.g. videos in [2, 6], rig captures in [9]) and annotate different object types. To build a coherent benchmark, we process the varying raw visual data, re-purpose their annotations to extract 3D cuboids in a unified 3D camera coordinate system, and carefully curate the final vocabulary. More details about the benchmark creation in the Appendix.", + "bbox": [ + 75, + 436, + 473, + 678 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We analyze OMNI3D and show its rich spatial and semantic properties proving it is visually diverse, similar to 2D data, and highly challenging for 3D as depicted in Fig. 2. We show the value of OMNI3D for the task of 3D object detection with extensive quantitative analysis in Sec. 5.", + "bbox": [ + 75, + 678, + 470, + 753 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Dataset Analysis", + "text_level": 1, + "bbox": [ + 75, + 767, + 240, + 782 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Splits. We split the dataset into $175\\mathrm{k} / 19\\mathrm{k} / 39\\mathrm{k}$ images for train/val/test respectively, consistent with original splits when available, and otherwise free of overlapping video sequences in splits. We denote indoor and outdoor subsets as $\\mathrm{OMN13D_{IN}}$ (SUN RGB-D, Hypersim, ARKit), and $\\mathrm{OMN13D_{OUT}}$ (KITTI, nuScenes). Objectron, with primarily close-up objects, is used only in the full OMN13D setting.", + "bbox": [ + 75, + 794, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Layout statistics. Fig. 2(a) shows the distribution of object centers onto the image plane by projecting centroids on the XY-plane (top row), and the distribution of object depths by projecting centroids onto the XZ-plane (bottom row). We find that OMNI3D's spatial distribution has a center bias, similar to 2D datasets COCO and LVIS. Fig. 2(b) depicts the relative object size distribution, defined as the square root of object area divided by image area. Objects are more likely to be small in size similar to LVIS (Fig. 6c in [24]) suggesting that OMNI3D is also challenging for 2D detection, while objects in OMNI3DOUT are noticeably smaller.", + "bbox": [ + 496, + 289, + 893, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The bottom row of Fig. 2(a) normalizes object depth in a $[0,20\\mathrm{m}]$ range, chosen for visualization and satisfies $88\\%$ of object instances; OMNI3D depth ranges as far as $300\\mathrm{m}$ . We observe that the OMNI3D depth distribution is far more diverse than SUN RGB-D and KITTI, which are biased toward near or road-side objects respectively. See Appendix for each data source distribution plot. Fig. 2(a) demonstrates OMNI3D's rich diversity in spatial distribution and depth which suffers significantly less bias than existing 3D benchmarks and is comparable in complexity to 2D datasets.", + "bbox": [ + 496, + 455, + 893, + 607 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2D and 3D correlation. A common assumption in urban scenes is that objects rest on a ground plane and appear smaller with depth. To verify if that is true generally, we compute correlations. We find that $2\\mathrm{D}y$ and $3\\mathrm{D}z$ are indeed fairly correlated in $\\mathrm{OMNI3D_{OUT}}$ at 0.524, but significantly less in $\\mathrm{OMNI3D_{IN}}$ at 0.006. Similarly, relative 2D object size $(\\sqrt{h\\cdot w})$ and $z$ correlation is 0.543 and 0.102 respectively. This confirms our claim that common assumptions in urban scenes are not generally true, making the task challenging.", + "bbox": [ + 496, + 609, + 893, + 747 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Category statistics. Fig. 2(c) plots the distribution of instances across the 98 categories of OMNI3D. The long-tail suggests that low-shot recognition in both 2D and 3D will be critical for performance. In this work, we want to focus on large-scale and diverse 3D recognition, which is comparably unexplored. We therefore filter and focus on categories with at least 1000 instances. This leaves 50 diverse categories including chair, sofa, laptop, table, books, car, truck, pedestrian and more, 19 of which have more than 10k instances. We provide more per-category details in the Appendix.", + "bbox": [ + 496, + 750, + 895, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "13156", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/48250649f1de3140a3a8dbcf9325d3359327e7a602b7399a3a0aa1aabbe7005f.jpg", + "image_caption": [ + "Figure 3. Overview. Cube R-CNN takes as input an RGB image, detects all objects in 2D and predicts their 3D cuboids $B_{3\\mathrm{D}}$ . During training, the 3D corners of the cuboids are compared against 3D ground truth with the point-cloud chamfer distance." + ], + "image_footnote": [], + "bbox": [ + 78, + 90, + 890, + 311 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 76, + 361, + 168, + 378 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our goal is to design a simple and effective model for general 3D object detection. Hence, our approach is free of domain or object specific strategies. We design our 3D object detection framework by extending Faster R-CNN [64] with a 3D object head which predicts a cuboid per each detected 2D object. We refer to our method as Cube R-CNN. Figure 3 shows an overview of our approach.", + "bbox": [ + 75, + 388, + 473, + 494 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Cube R-CNN", + "text_level": 1, + "bbox": [ + 76, + 507, + 218, + 522 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our model builds on Faster R-CNN [64], an end-to-end region-based object detection framework. Faster-RCNN consists of a backbone network, commonly a CNN, which embeds the input image into a higher-dimensional feature space. A region proposal network (RPN) predicts regions of interest (RoIs) representing object candidates in the image. A 2D box head inputs the backbone feature map and processes each RoI to predict a category and a more accurate 2D box. Faster R-CNN can be easily extended to tackle more tasks by adding task-specific heads, e.g. Mask R-CNN [26] adds a mask head to additionally output object silhouettes.", + "bbox": [ + 75, + 532, + 468, + 698 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For the task of 3D object detection, we extend Faster R-CNN by introducing a 3D detection head which predicts a 3D cuboid for each detected 2D object. Cube R-CNN extends Faster R-CNN in three ways: (a) we replace the binary classifier in RPN which predicts region objectness with a regressor that predicts IoUness, (b) we introduce a cube head which estimates the parameters to define a 3D cuboid for each detected object (similar in concept to [68]), and (c) we define a new training objective which incorporates a virtual depth for the task of 3D object detection.", + "bbox": [ + 75, + 699, + 470, + 851 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "IoUness. The role of RPN is two-fold: (a) it proposes RoIs by regressing 2D box coordinates from pre-computed anchors and (b) it classifies regions as object or not (object-", + "bbox": [ + 75, + 854, + 472, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ness). This is sensible in exhaustively labeled datasets where it can be reliably assessed if a region contains an object. However, OMNI3D combines many data sources with no guarantee that all instances of all classes are labeled. We overcome this by replacing objectness with IoUness, applied only to foreground. Similar to [34], a regressor predicts IoU between a RoI and a ground truth. Let $o$ be the predicted IoU for a RoI and $\\hat{o}$ be the 2D IoU between the region and its ground truth; we apply a binary cross-entropy (CE) loss $\\mathcal{L}_{\\mathrm{IoUness}} = \\ell_{\\mathrm{CE}}(o,\\hat{o})$ . We train on regions whose IoU exceeds 0.05 with a ground truth in order to learn IoUness from a wide range of region overlaps. Thus, the RPN training objective becomes $\\mathcal{L}_{\\mathrm{RPN}} = \\hat{o}\\cdot (\\mathcal{L}_{\\mathrm{IoUness}} + \\mathcal{L}_{\\mathrm{reg}})$ , where $\\mathcal{L}_{\\mathrm{reg}}$ is the 2D box regression loss from [64]. The loss is weighted by $\\hat{o}$ to prioritize candidates close to true objects.", + "bbox": [ + 496, + 363, + 893, + 592 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Cube Head. We extend Faster R-CNN with a new head, called cube head, to predict a 3D cuboid for each detected 2D object. The cube head inputs $7 \\times 7$ feature maps pooled from the backbone for each predicted region and feeds them to 2 fully-connected (FC) layers with 1024 hidden dimensions. All 3D estimations in the cube head are category-specific. The cube head represents a 3D cuboid with 13 parameters each predicted by a final FC layer:", + "bbox": [ + 496, + 599, + 893, + 720 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- $[u, v]$ represent the projected 3D center on the image plane relative to the 2D RoI", + "- $z \\in \\mathbb{R}_+$ is the object's center depth in meters transformed from virtual depth $z_v$ (explained below)", + "- $[ \\bar{w}, \\bar{h}, \\bar{l} ] \\in \\mathbb{R}_{+}^{3}$ are the log-normalized physical box dimensions in meters", + "- $\\mathbf{p} \\in \\mathbb{R}^6$ is the continuous 6D [86] allocentric rotation", + "- $\\mu \\in \\mathbb{R}_+$ is the predicted 3D uncertainty" + ], + "bbox": [ + 517, + 727, + 890, + 848 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The above parameters form the final 3D box in camera view coordinates for each detected 2D object. The object's 3D center $\\mathbf{X}$ is estimated from the predicted 2D projected", + "bbox": [ + 496, + 854, + 893, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "13157", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "center $[u,v]$ and depth $z$ via", + "bbox": [ + 76, + 90, + 266, + 107 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {X} (u, v, z) = \\left(\\frac {z}{f _ {x}} \\left(r _ {x} + u r _ {w} - p _ {x}\\right), \\frac {z}{f _ {y}} \\left(r _ {y} + v r _ {h} - p _ {y}\\right), z\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 114, + 468, + 148 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $[r_x, r_y, r_w, r_h]$ is the object's 2D box, $(f_x, f_y)$ are the camera's known focal lengths and $(p_x, p_y)$ the principal point. The 3D box dimensions $\\mathbf{d}$ are derived from $[\\bar{w}, \\bar{h}, \\bar{l}]$ which are log-normalized with category-specific precomputed means $(w_0, h_0, l_0)$ for width, height and length respectively, and are arranged into a diagonal matrix via", + "bbox": [ + 76, + 148, + 472, + 242 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {d} (\\bar {w}, \\bar {h}, \\bar {l}) = \\operatorname {d i a g} \\left(\\exp (\\bar {w}) w _ {0}, \\exp (\\bar {h}) h _ {0}, \\exp (\\bar {l}) l _ {0}\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 248, + 468, + 266 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, we derive the object's pose $\\mathbf{R}(\\mathbf{p})$ as a $3 \\times 3$ rotation matrix based on a 6D parameterization (2 directional vectors) of $\\mathbf{p}$ following [86] which is converted from allocentric to egocentric rotation similar to [39], defined formally in Appendix. The final 3D cuboid, defined by 8 corners, is", + "bbox": [ + 75, + 273, + 470, + 349 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nB _ {\\mathrm {3 D}} (u, v, z, \\bar {w}, \\bar {h}, \\bar {l}, \\mathbf {p}) = \\mathbf {R} (\\mathbf {p}) \\mathbf {d} (\\bar {w}, \\bar {h}, \\bar {l}) B _ {\\text {u n i t}} + \\mathbf {X} (u, v, z) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 86, + 356, + 468, + 386 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $B_{\\mathrm{unit}}$ are the 8 corners of an axis-aligned unit cube centered at $(0,0,0)$ . Lastly, $\\mu$ denotes a learned 3D uncertainty, which is mapped to a confidence at inference then joined with the classification score $s$ from the 2D box head to form the final score for the prediction, as $\\sqrt{s \\cdot \\exp(-\\mu)}$ .", + "bbox": [ + 75, + 386, + 470, + 463 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training objective. Our training objective consists of 2D losses from the RPN and 2D box head and 3D losses from the cube head. The 3D objective compares each predicted 3D cuboid with its matched ground truth via a chamfer loss, treating the 8 box corners of the 3D boxes as point clouds, namely $\\mathcal{L}_{\\mathrm{3D}}^{\\mathrm{all}} = \\ell_{\\mathrm{chamfer}}(B_{\\mathrm{3D}}, B_{\\mathrm{3D}}^{\\mathrm{gt}})$ . Note that $\\mathcal{L}_{\\mathrm{3D}}^{\\mathrm{all}}$ entangles all 3D variables via the box predictor $B_{\\mathrm{3D}}$ (Eq. 3), such that that errors in variables may be ambiguous from one another. Thus, we isolate each variable group with separate disentangled losses, following [68]. The disentangled loss for each variable group substitutes all but its variables with the ground truth from Eq. 3 to create a pseudo box prediction. For example, the disentangled loss for the projected center $[u, v]$ produces a 3D box with all but $(u, v)$ replaced with the true values and then compares to the ground truth box,", + "bbox": [ + 75, + 465, + 472, + 691 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {3 D}} ^ {(u, v)} = \\left\\| B _ {\\mathrm {3 D}} (u, v, z ^ {\\mathrm {g t}}, \\bar {w} ^ {\\mathrm {g t}}, \\bar {h} ^ {\\mathrm {g t}}, \\bar {l} ^ {\\mathrm {g t}}, \\mathbf {p} ^ {\\mathrm {g t}}) - B _ {\\mathrm {3 D}} ^ {\\mathrm {g t}} \\right\\| _ {1} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 700, + 468, + 723 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use an $L_{1}$ loss for $\\mathcal{L}_{\\mathrm{3D}}^{(u,v)}$ , $\\mathcal{L}_{\\mathrm{3D}}^{(z)}$ and $\\mathcal{L}_{\\mathrm{3D}}^{(\\bar{w},\\bar{h},\\bar{l})}$ and chamfer for $\\mathcal{L}_{\\mathrm{3D}}^{\\mathbf{P}}$ to account for cuboid symmetry such that rotation matrices of Euler angles modulo $\\pi$ produce the same noncanonical 3D cuboid. Losses in box coordinates have natural advantages over losses which directly compare variables to ground truths in that their gradients are appropriately weighted by the error as shown in [68]. The 3D objective is defined as, $\\mathcal{L}_{\\mathrm{3D}} = \\mathcal{L}_{\\mathrm{3D}}^{(u,v)} + \\mathcal{L}_{\\mathrm{3D}}^{(z)} + \\mathcal{L}_{\\mathrm{3D}}^{(\\bar{w},\\bar{h},\\bar{l})} + \\mathcal{L}_{\\mathrm{3D}}^{\\mathbf{P}} + \\mathcal{L}_{\\mathrm{3D}}^{\\mathrm{all}}$ .", + "bbox": [ + 75, + 731, + 468, + 859 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The collective training objective of Cube R-CNN is,", + "bbox": [ + 96, + 859, + 441, + 875 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\mathrm {R P N}} + \\mathcal {L} _ {\\mathrm {2 D}} + \\sqrt {2} \\cdot \\exp (- \\mu) \\cdot \\mathcal {L} _ {\\mathrm {3 D}} + \\mu \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 122, + 883, + 468, + 902 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b994ceceef479499166490ea0d206666179a0b48729edfa20c1aaf24e501295d.jpg", + "image_caption": [ + "Figure 4. Varying camera intrinsics exaggerate scale-depth ambiguity – the same object at two different depths can project to the same image, as in (a) and (b). We address this by introducing a virtual camera which is invariant to intrinsics and transforms the object's depth to a virtual depth $z_v$ such that the effective image size $H_v$ and focal length $f_v$ are consistent across images." + ], + "image_footnote": [], + "bbox": [ + 500, + 92, + 689, + 258 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2078c047e41eabfaf45470df4b46e07f69028df1b006123e33e713c23eba49d7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 93, + 890, + 258 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\mathcal{L}_{\\mathrm{RPN}}$ is the RPN loss, described above, and $\\mathcal{L}_{2\\mathrm{D}}$ is the 2D box head loss from [64]. The 3D loss is weighted by the predicted 3D uncertainty (inspired by [52]), such that the model may trade a penalty to reduce the 3D loss when uncertain. In practice, $\\mu$ is helpful for both improving the 3D ratings at inference and reducing the loss of hard samples in training.", + "bbox": [ + 496, + 359, + 893, + 452 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Virtual Depth", + "text_level": 1, + "bbox": [ + 500, + 465, + 643, + 482 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A critical part of 3D object detection is predicting an object's depth in metric units. Estimating depth from visual cues requires an implicit mapping of 2D pixels to 3D distances, which is more ambiguous if camera intrinsics vary. Prior works are able to ignore this as they primarily train on images from a single sensor. OMNI3D contains images from many sensors and thus demonstrates large variations in camera intrinsics. We design Cube R-CNN to be robust to intrinsics by predicting the object's virtual depth $z_{v}$ , which projects the metric depth $z$ to an invariant camera space.", + "bbox": [ + 496, + 489, + 893, + 641 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Virtual depth scales the depth using the (known) camera intrinsics such that the effective image size and focal length are consistent across the dataset, illustrated in Fig. 4. The effective image properties are referred to as virtual image height $H_{v}$ and virtual focal length $f_{v}$ , and are both hyperparameters. If $z$ is the true metric depth of an object from an image with height $H$ and focal length $f$ , its virtual depth is defined as $z_{v} = z \\cdot \\frac{f_{v}}{f} \\frac{H}{H_{v}}$ . The derivation is in the Appendix.", + "bbox": [ + 496, + 642, + 893, + 766 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Invariance to camera intrinsics via virtual depth also enables scale augmentations during training, since $H$ can vary without harming the consistency of $z_v$ . Data augmentations from image resizing tend to be critical for 2D models but are not often used in 3D methods [7, 50, 68] since if unaccounted for they increase scale-depth ambiguity. Virtual depth lessens such ambiguities and therefore enables powerful data augmentations in training. We empirically prove the two-fold effects of virtual depth in Sec. 5", + "bbox": [ + 496, + 765, + 893, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "13158", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a1a2d7a345fba5423768e23b6e43263027ad5e429ba8ca2ecb6e0bd1707e2621.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 88, + 225, + 172 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a2dc74297760f80c0033139fd8f2a66ca1c5c049a5b35661ec6607549405f342.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 227, + 88, + 372, + 172 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fe6baf6159df0942d2b2c58dc92dcab418722b3568ee670c0348462699d660e8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 372, + 88, + 480, + 172 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/63865173e449343cb6a235b186fef61c7eb2831b82800b145e648ab7559c8b4a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 480, + 88, + 629, + 172 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9b2f18ec20a52ec69421e2c4747a97fa3ff7ac1e6c0e44d907e19ce1840c4000.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 629, + 89, + 774, + 172 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/994acec229be8a238f2e6ae3de3084ad3eff1edf9f1e7dc9153bdae35573ce04.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 774, + 89, + 883, + 172 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1fe5ddef149919d8c5b2512ad32fd572416c68a60e7c3ed18fe48a020a463fdf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 80, + 172, + 225, + 258 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/16cf37b188a3a95b8418357b413f7e0c4a64ce5f07b4c152025a192001691b48.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 225, + 172, + 372, + 258 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7e97d86431114eced98487b0a9f167fd210658c5acd71e0788da9343891eceef.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 372, + 172, + 480, + 258 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ed95d112221cbf933ef94b625c3e9c71ae08b6c3a7e52ddfe2ab54a65a4d3904.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 480, + 172, + 629, + 258 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c5864368434eca5af31628f837fc8568b705e61a49f0d171a647acd10589eced.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 629, + 172, + 776, + 258 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/eb5549ccd6cbfcf2fdf7358d68b028373f4e940a742e804d7fc6f9e956e76eb7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 776, + 172, + 883, + 258 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/62880ba4b80097575de00d90c051d8a0059a2d7951a901860f40d392abde6529.jpg", + "image_caption": [ + "Figure 5. Cube R-CNN predictions on OMNI3D test. For each example, we show the input image, the 3D predictions overlaid on the image and a top view with a base composed of $1\\mathrm{m}\\times 1\\mathrm{m}$ tiles. See more results from images and videos in the Appendix." + ], + "image_footnote": [], + "bbox": [ + 81, + 258, + 236, + 325 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/da91209b692c3de536abce28f727b2b9b52bea4afcb0c168baf45d40b4e3e456.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 236, + 258, + 393, + 325 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5c3715f40fd44e583e3284afb2c8c619bd2bda3be2e27605386f9827378e1468.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 395, + 258, + 480, + 325 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ad691737e8c087b90e5339938ce38bf914cb8e5308914dd316de4da9443ff3e7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 480, + 258, + 638, + 325 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/78af7f71cfa0eb0c9f119b807ebafa86676166b3cf2cc4f2558e7274ebf03fb9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 638, + 258, + 795, + 325 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/884996be08dfa5c2cedafd2a7fc444c85bbcb00f1e08495b4d65bf40eab8f572.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 795, + 258, + 883, + 325 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 76, + 369, + 207, + 386 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We tackle image-based 3D object detection on OMNI3D and compare to prior best methods. We evaluate Cube R-CNN on existing 3D benchmarks with a single unified model. Finally, we prove the effectiveness of OMNI3D as a large-scale 3D object detection dataset by showing comprehensive cross-dataset generalization and its impact for pre-training.", + "bbox": [ + 75, + 400, + 470, + 491 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details. We implement Cube R-CNN using Detector2 [80] and PyTorch3D [61]. Our backbone is DLA34 [82]-FPN [43] pretrained on ImageNet [67]. We train for 128 epochs with a batch size of 192 images across 48 V100s. We use SGD with a learning rate of 0.12 which decays by a factor of 10 after $60\\%$ and $80\\%$ of training. During training, we use random data augmentation of horizontal flipping and scales $\\in [0.50, 1.25]$ , enabled by virtual depth. Virtual camera parameters are set to $f_v = H_v = 512$ . Source code and models are publicly available.", + "bbox": [ + 75, + 500, + 470, + 650 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metric. Following popular benchmarks for 2D and 3D recognition, we use average-precision (AP) as our 3D metric. Predictions are matched to ground truth by measuring their overlap using $\\mathsf{I}\\circ \\mathsf{U}_{\\mathsf{3D}}$ which computes the intersection-overunion of 3D cuboids. We compute a mean $\\mathrm{AP}_{\\mathrm{3D}}$ across all 50 categories in OMNI3D and over a range of $\\mathsf{I}\\circ \\mathsf{U}_{\\mathsf{3D}}$ thresholds $\\tau \\in [0.05,0.10,\\dots ,0.50]$ . The range of $\\tau$ is more relaxed than in 2D to account for the new dimension - see the Appendix for a comparison between 3D and 2D ranges for IoU. General 3D objects may be occluded by other objects or truncated by the image border, and can have arbitrarily small projections. Following [21], we ignore objects with high occlusion ( $>66\\%$ ) or truncation ( $>33\\%$ ), and tiny projected objects ( $< 6.25\\%$ image height). Moreover, we also report $\\mathrm{AP}_{\\mathrm{3D}}$ at varying levels of depth $d$ as near: $0 < d \\leq 10m$ , medium: $10m < d \\leq 35m$ , far: $35m < d \\leq \\infty$ .", + "bbox": [ + 75, + 657, + 472, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fast $\\mathsf{IOU}_{\\mathbf{3D}}$ . $\\mathsf{IOU}_{\\mathbf{3D}}$ compares two 3D cuboids by computing their intersection-over-union. Images usually have many objects and produce several predictions so $\\mathsf{IOU}_{\\mathbf{3D}}$ computations need to be fast. Prior implementations [21] approximate $\\mathsf{IOU}_{\\mathbf{3D}}$ by projecting 3D boxes on a ground plane and multiply the top-view 2D intersection with the box heights to compute a 3D volume. Such approximations become notably inaccurate when objects are not on a planar ground or have an arbitrary orientation (e.g. with nonzero pitch or roll). Objectron [2] provides an exact solution, but relies on external libraries [5, 74] implemented in $\\mathbf{C}++$ and is not batched. We implement a new, fast and exact algorithm which computes the intersecting shape by representing cuboids as meshes and finds face intersections – described in Appendix. Our algorithm is batched with $\\mathbf{C}++$ and CUDA support. Our implementation is $90\\times$ faster than Objectron in $\\mathbf{C}++$ , and $450\\times$ faster in CUDA. As a result, evaluation on the large OMNI3D takes only seconds instead of several hours.", + "bbox": [ + 496, + 371, + 893, + 643 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Model Performance", + "text_level": 1, + "bbox": [ + 500, + 657, + 687, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "First, we ablate the design choices of Cube R-CNN. Then, we compare to state-of-the-art methods on existing benchmarks and OMNI3D and show that it performs superior or on par with prior works which are specialized for their respective domains. Our Cube R-CNN uses a single unified design to tackle general 3D object detection across domains. Fig. 5 shows qualitative predictions on the OMNI3D test set.", + "bbox": [ + 496, + 683, + 893, + 789 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Ablations. Table 1 ablates the features of Cube R-CNN on OMNI3D. We report $\\mathrm{AP}_{3\\mathrm{D}}$ performance at single IoU thresholds (0.25 and 0.50), and for near, medium and far objects. We further train and evaluate on domain-specific subsets, OMNI3DIN $(\\mathrm{AP}_{3\\mathrm{D}}^{\\mathrm{IN}})$ and OMNI3DOUT $(\\mathrm{AP}_{3\\mathrm{D}}^{\\mathrm{OUT}})$ to show how ablation trends hold for single domain scenarios. From Table 1 we draw a few standout observations:", + "bbox": [ + 496, + 794, + 893, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "13159", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/fae1a06f10f7cdb927fd70106e23bdde063ae2755514e5f65c3d8b4ad935f682.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Cube R-CNNAP3DAP253DAP503DAPnear3DAPmed3DAPfar3DAPIN3DAPOUT3D
w/o disentangled22.624.39.226.811.88.214.530.9
w/o IoUness22.223.68.826.411.18.314.331.0
w/o L3D20.221.86.426.412.17.413.829.1
w/o scale aug.20.221.58.023.59.86.812.328.1
w/o scale | virtual19.821.27.523.48.65.712.226.0
w/o virtual depth17.318.44.422.77.97.113.230.3
w/o uncertainty μ17.218.35.420.510.87.09.125.4
ours23.324.99.527.912.18.515.031.9
", + "bbox": [ + 181, + 85, + 787, + 218 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8ad4cd841d71a02c837ea47b991d5118412c25059013b8c5ce73c1b15dedecb2.jpg", + "table_caption": [ + "Table 1. Cube R-CNN ablations. We report $\\mathrm{AP}_{3\\mathrm{D}}$ on OMNI3D, then at IoU thresholds 0.25 and 0.50, and for near, medium and far objects. We further report $\\mathrm{AP}_{3\\mathrm{D}}$ when both training and evaluating on OMNI3DIN and OMNI3DOUT subsets." + ], + "table_footnote": [ + "Table 2. OMNI3D comparison with GUPNet [52], ImVoxelNet [66], M3D-RPN [7], FCOS3D [78], PGD [77], SMOKE [50]. We apply virtual camera to the last four $(+vc)$ and show the gains in purple. $f$ denotes if a method handles variance in focal length." + ], + "table_body": "
MethodfOMNI3DOUTOMNI3D
AP3DAPNU3DAPOUT3DAP3D
M3D-RPN [7]X10.417.913.7-
SMOKE [50]X25.420.419.59.6
FCOS3D [78]X14.620.917.69.8
PGD [77]X21.426.322.911.2
GUPNet [52]✓24.520.519.9-
M3D-RPN +vc✓16.2+5.820.4+2.917.0+3.3-
ImVoxelNet [66]✓23.523.421.59.4
SMOKE +vc✓25.9+0.520.4+0.020.0+0.510.4+0.8
FCOS3D +vc✓17.8+3.225.1+4.221.3+3.710.6+0.8
PGD +vc✓22.8+1.431.2+4.926.8+3.915.4+4.2
Cube R-CNN✓36.032.731.923.3
", + "bbox": [ + 81, + 250, + 475, + 417 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Virtual depth is effective and improves $\\mathrm{AP}_{3\\mathrm{D}}$ by $+6\\%$ , most noticeable in the full OMNI3D which has the largest variances in camera intrinsics. Enabled by virtual depth, scale augmentations during training increase $\\mathrm{AP}_{3\\mathrm{D}}$ by $+3.1\\%$ on OMNI3D, $+2.7\\%$ on OMNI3D $_{\\text{IN}}$ and $+3.8\\%$ on OMNI3D $_{\\text{OUT}}$ . We find scale augmentation controlled without virtual depth to be harmful by $-2.5\\%$ (rows 6 - 7).", + "bbox": [ + 75, + 489, + 470, + 595 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3D uncertainty boosts performance by about $+6\\%$ on all OMNI3D subsets. Intuitively, it serves as a measure of confidence for the model's 3D predictions; removing it means that the model relies only on its 2D classification to score cuboids. If uncertainty is used only to scale samples in Eq. 5, but not at inference then $\\mathrm{AP}_{3\\mathrm{D}}$ still improves but by $+1.6\\%$ .", + "bbox": [ + 75, + 599, + 470, + 691 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1 also shows that the entangled $\\mathcal{L}_{3\\mathrm{D}}^{\\mathrm{all}}$ loss from Eq. 5 boosts $\\mathrm{AP}_{3\\mathrm{D}}$ by $+3.1\\%$ while disentangled losses contribute less $(+0.7\\%)$ . Replacing objectness with IoUness in the RPN head improves $\\mathrm{AP}_{3\\mathrm{D}}$ on OMNI3D by $+1.1\\%$ .", + "bbox": [ + 75, + 691, + 470, + 753 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison to other methods. We compare Cube R-CNN to prior best approaches. We choose representative state-of-the-art methods designed for the urban and indoor domains, and evaluate on our proposed OMNI3D benchmark and single-dataset benchmarks, KITTI and SUN RGB-D.", + "bbox": [ + 75, + 758, + 470, + 834 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparisons on OMNI3D. Table 2 compares Cube R-CNN to M3D-RPN [7], a single-shot 3D anchor approach, FCOS3D [78] and its follow-up PGD [77], which use a fully convolutional one-stage model, GUPNet [52], which uses", + "bbox": [ + 75, + 839, + 472, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/eaf0d3cb0ed5d6ce271a8a50f06f51debe646a62a590620fb9fbdf595f084e2a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodDAP703DAP70BEV
EasyMedHardEasyMedHard
SMOKE [50]X14.039.767.8420.8314.4912.75
ImVoxelNet [66]X17.1510.979.1525.1916.3713.58
PGD [77]X19.0511.769.3926.8916.5113.49
GUPNet [52]X22.2615.0213.1230.2921.1918.20
Cube R-CNNX23.5915.0112.5631.7021.2018.43
MonoDTR [30]✓21.9915.3912.7328.5920.3817.14
DD3D [59]✓23.1916.8714.3632.3523.4120.42
", + "bbox": [ + 504, + 250, + 898, + 368 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. KITTI leaderboard results on car for image-based approaches. D denotes methods trained with extra depth supervision.", + "bbox": [ + 498, + 369, + 893, + 398 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "2D-3D geometry and camera focal length to derive depth, SMOKE [50] which predicts 3D object center and offsets densely and ImVoxelNet [66], which uses intrinsics to unproject 2D image features to a 3D volume followed by 3D convolutions. These methods originally experiment on urban domains, so we compare on OMNI3DOUT and the full OMNI3D. We train each method using public code and report the best of multiple runs after tuning their hyper-parameters to ensure best performance. We extend all but GUPNet and ImVoxelNet with a virtual camera to handle varying intrinsics, denoted by $+vc$ . GUPNet and ImVoxelNet dynamically use per-image intrinsics, which can naturally account for camera variation. M3D-RPN and GUPNet do not support multi-node training, which makes scaling to OMNI3D difficult and are omitted. For OMNI3DOUT, we report $\\mathrm{AP}_{3\\mathrm{D}}^{\\mathrm{OUT}}$ on its test set and its subparts, $\\mathrm{AP}_{3\\mathrm{D}}^{\\mathrm{KIT}}$ and $\\mathrm{AP}_{3\\mathrm{D}}^{\\mathrm{NU}}$ .", + "bbox": [ + 496, + 412, + 893, + 654 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "From Table 2 we observe that our Cube R-CNN outperforms competitive methods on both $\\mathrm{OMNI3D_{OUT}}$ and OMNI3D. The virtual camera extensions result in performance gains of varying degree for all approaches, e.g. $+4.2\\%$ for PGD on OMNI3D, proving its impact as a general purpose feature for mixed dataset training for the task of 3D object detection. FCOS3D, SMOKE, and ImVoxelNet [66] struggle the most on OMNI3D, which is not surprising as they are typically tailored for dataset-specific model choices.", + "bbox": [ + 496, + 654, + 893, + 790 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparisons on KITTI. Table 3 shows results on KITTI's test set using their server [1]. We train Cube R-CNN on KITTI only. Cube R-CNN is a general purpose 3D object detector; its design is not tuned for the KITTI benchmark. Yet, it performs better or on par with recent best methods which are heavily tailored for KITTI and is only slightly worse to models trained with extra depth supervision.", + "bbox": [ + 496, + 795, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "13160", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/0101700579859d3a07b5b845b6a241f90684d7647b283f535e2a6e23eb350e4d.jpg", + "table_caption": [], + "table_footnote": [ + "Table 4. We compare to indoor models with common categories. For Total3D, we use oracle 2D detections to report $\\mathsf{I}\\circ \\mathsf{U}_{\\mathsf{3D}}$ ( $3^{\\mathrm{rd}}$ col). For ImVoxelNet, we report the detection metric of $\\mathrm{AP}_{\\mathrm{3D}}$ ( $4^{\\mathrm{th}}$ col)." + ], + "table_body": "
MethodTrained onI0U3DAP3D
Total3D [58]SUN RGB-D23.3-
ImVoxelNet [66]SUN RGB-D-30.6
Cube R-CNNSUN RGB-D36.234.7
Cube R-CNNOMNI3DIN37.835.4
", + "bbox": [ + 86, + 88, + 455, + 155 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "KITTI's $\\mathrm{AP}_{3\\mathrm{D}}$ is fundamentally similar to ours but sets a single $\\mathsf{I}\\circ \\mathsf{U}_{3\\mathrm{D}}$ threshold at 0.70, which behaves similar to a nearly perfect 2D threshold of 0.94 (see the Appendix). To remedy this, our $\\mathrm{AP}_{3\\mathrm{D}}$ is a mean over many thresholds, inspired by COCO [45]. In contrast, nuScenes proposes a mean AP based on 3D center distance but omits size and rotation, with settings generally tailored to the urban domain (e.g. size and rotation variations are less extreme for cars). We provide more analysis with this metric in the Appendix.", + "bbox": [ + 75, + 208, + 470, + 345 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Comparisons on SUN RGB-D. Table 4 compares Cube R-CNN to Total3D [58] and ImVoxelNet [66], two state-of-the-art methods on SUN RGB-D on 10 common categories. Total3D's public model requires 2D object boxes and categories as input. Hence, for a fair comparison, we use ground truth 2D detections as input to both our and their method, each trained using a ResNet34 [27] backbone and report mean IoU3D (3rd col). We compare to ImVoxelNet in the full 3D object detection setting and report AP3D (4th col). We compare Cube R-CNN from two respects, first trained on SUN RGB-D identical to the baselines, then trained on OMNI3DIN which subsumes SUN RGB-D. Our model outperforms Total3D by +12.9% and ImVoxelNet by +4.1% when trained on the same training set and increases the performance gap when trained on the larger OMNI3DIN.", + "bbox": [ + 75, + 347, + 470, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Zero-shot Performance. One commonly sought after property of detectors is zero-shot generalization to other datasets. Cube R-CNN outperforms all other methods at zero-shot. We find our model trained on KITTI achieves $12.7\\%$ $\\mathrm{AP}_{3\\mathrm{D}}$ on nuScenes; the second best is M3D-RPN+vc with $10.7\\%$ . Conversely, when training on nuScenes, our model achieves $20.2\\%$ on KITTI; the second best is GUPNet with $17.3\\%$ .", + "bbox": [ + 75, + 578, + 470, + 683 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2. The Impact of the OMNI3D Benchmark", + "text_level": 1, + "bbox": [ + 76, + 691, + 410, + 708 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Sec. 5.1 analyzes the performance of Cube R-CNN for the task of 3D object detection. Now, we turn to OMNI3D and its impact as a large-scale benchmark. We show two use cases of OMNI3D: (a) a universal 3D dataset which integrates smaller ones, and (b) a pre-training dataset.", + "bbox": [ + 75, + 715, + 468, + 792 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "OMNI3D as a universal dataset. We treat OMNI3D as a dataset which integrates smaller single ones and show its impact on each one. We train Cube R-CNN on OMNI3D and compare to single-dataset training in Table 5, for the indoor (left) and urban (right) domain. $\\mathrm{AP}_{\\mathrm{3D}}$ is reported on the category intersection (10 for indoor, 3 for outdoor) to ensure comparability. Training on OMNI3D and its domain-specific", + "bbox": [ + 75, + 795, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e0cc40fbdf096e9aa4486f6a78d3a405aaa34853c938ff167809c965102a0bf1.jpg", + "table_caption": [], + "table_footnote": [ + "Table 5. Cross-dataset performance on intersecting categories (10 for indoor, 3 for outdoor) for comparable cross-evaluations." + ], + "table_body": "
Trained onAPH3DAPSUN3DAPAR3DTrained onAPKIT3DAPNU3D
Hypersim15.29.57.5KITTI37.112.7
SUN5.834.713.1nuScenes20.238.6
ARKit5.914.238.6OMNI3D37.835.8
OMNI3D19.032.638.2OMNI3DOUT42.439.0
OMNI3DIN17.835.441.2
", + "bbox": [ + 504, + 87, + 883, + 167 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8fe75a37d19012db8085bce4fd4c08e79478b6133bc6adfa3132a832bcfe0cc7.jpg", + "image_caption": [ + "Figure 6. Pre-training on OMNI3D- vs. ImageNet." + ], + "image_footnote": [], + "bbox": [ + 532, + 205, + 854, + 344 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "subsets, $\\mathrm{OMN13D_{IN}}$ and $\\mathrm{OMN13D_{OUT}}$ , results in higher performance compared to single-dataset training, signifying that our large OMN13D generalizes better and should be preferred over single dataset training. ARKit sees a $+2.6\\%$ boost and KITTI $+5.3\\%$ . Except for Hypersim, the domain-specific subsets tend to perform better on their domain, which is not surprising given their distinct properties (Fig. 2).", + "bbox": [ + 496, + 375, + 893, + 481 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "OMNI3D as a pre-training dataset. Next, we demonstrate the utility of OMNI3D for pre-training. In this setting, an unseen dataset is used for finetuning from a model pre-trained on OMNI3D. The motivation is to determine how a large-scale 3D dataset could accelerate low-shot learning with minimum need for costly 3D annotations on a new dataset. We choose SUN RGB-D and KITTI as our unseen datasets given their popularity and small size. We pre-train Cube R-CNN on OMNI3D-, which removes them from OMNI3D, then finetune the models using a $\\%$ of their training data. The curves in Fig. 6 show the model quickly gains its upper-bound performance at a small fraction of the training data, when pre-trained on OMNI3D- vs. ImageNet, without any few-shot training tricks. A model finetuned on only $5\\%$ of its target can achieve $>70\\%$ of the upper-bound performance.", + "bbox": [ + 496, + 484, + 893, + 712 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 726, + 617, + 742 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose a large and diverse 3D object detection benchmark, OMNI3D, and a general purpose 3D object detector, Cube R-CNN. Models and data are publicly released. Our extensive analysis (Table 1, Fig. 5, Appendix) show the strength of general 3D object detection as well as the limitations of our method, e.g. localizing far away objects, and uncommon object types or contexts. Moreover, for accurate real-world 3D predictions, our method is limited by the assumption of known camera intrinsics, which may be minimized in future work using self-calibration techniques.", + "bbox": [ + 496, + 750, + 893, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "13161", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] KITTI server. https://www.cvlibs.net/datasets/kitti/eval_object.php?obj_benchmark=3d. Accessed: 2022-11-01. 7", + "[2] Adel Ahmadyan, Liangkai Zhang, Artsiom Ablavatski, Jianing Wei, and Matthias Grundmann. Objectron: A large scale dataset of object-centric videos in the wild with pose annotations. CVPR, 2021. 2, 3, 6", + "[3] Armen Avetisyan, Manuel Dahnert, Angela Dai, Manolis Savva, Angel X Chang, and Matthias Nießner. Scan2CAD: Learning CAD model alignment in RGB-D scans. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 2614-2623, 2019. 2", + "[4] Wentao Bao, Bin Xu, and Zhenzhong Chen. MonoFENet: Monocular 3D object detection with feature enhancement networks. TIP, 2019. 2", + "[5] Bradford Barber, David Dobkin, and Hannu Huhdanpaa. The quickhull algorithm for convex hulls. ACM Transactions on Mathematical Software (TOMS), 1996. 6", + "[6] Gilad Baruch, Zhuoyuan Chen, Afshin Dehghan, Tal Dimry, Yuri Feigin, Peter Fu, Thomas Gebauer, Brandon Joffe, Daniel Kurz, Arik Schwartz, and Elad Shulman. ARKitScenes - a diverse real-world dataset for 3D indoor scene understanding using mobile RGB-D data. In NeurIPS Datasets and Benchmarks Track, 2021. 2, 3", + "[7] Garrick Brazil and Xiaoming Liu. M3D-RPN: Monocular 3D region proposal network for object detection. In ICCV, 2019. 1, 2, 5, 7", + "[8] Garrick Brazil, Gerard Pons-Moll, Xiaoming Liu, and Bernt Schiele. Kinematic 3D object detection in monocular video. In ECCV, 2020. 2", + "[9] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuScenes: A multimodal dataset for autonomous driving. In CVPR, 2020. 2, 3", + "[10] Ming-Fang Chang, John Lambert, Patsorn Sangkloy, Jagjeet Singh, Slawomir Bak, Andrew Hartnett, De Wang, Peter Carr, Simon Lucey, Deva Ramanan, et al. Argoverse: 3D tracking and forecasting with rich maps. In CVPR, 2019. 2", + "[11] Hansheng Chen, Yuyao Huang, Wei Tian, Zhong Gao, and Lu Xiong. MonoRUn: Monocular 3D object detection by reconstruction and uncertainty propagation. In CVPR, 2021. 2", + "[12] Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jiarui Xu, Zheng Zhang, Dazhi Cheng, Chenchen Zhu, Tianheng Cheng, Qijie Zhao, Buyu Li, Xin Lu, Rui Zhu, Yue Wu, Jifeng Dai, Jingdong Wang, Jianping Shi, Wanli Ouyang, Chen Change Loy, and Dahua Lin. MMDetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155, 2019. 2", + "[13] Xiaozhi Chen, Kaustav Kundu, Ziyu Zhang, Huimin Ma, Sanja Fidler, and Raquel Urtasun. Monocular 3D object detection for autonomous driving. In CVPR, 2016. 1, 2", + "[14] Yongjian Chen, Lei Tai, Kai Sun, and Mingyang Li. MonoPair: Monocular 3D object detection using pairwise spatial relationships. In CVPR, 2020. 2" + ], + "bbox": [ + 76, + 114, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Yi-Nan Chen, Hang Dai, and Yong Ding. Pseudo-stereo for monocular 3D object detection in autonomous driving. In CVPR, 2022. 2", + "[16] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. ScanNet: Richly-annotated 3D reconstructions of indoor scenes. In CVPR, 2017. 2", + "[17] Saumitro Dasgupta, Kuan Fang, Kevin Chen, and Silvio Savarese. Delay: Robust spatial layout estimation for cluttered indoor scenes. In CVPR, 2016. 2", + "[18] Mingyu Ding, Yuqi Huo, Hongwei Yi, Zhe Wang, Jianping Shi, Zhiwu Lu, and Ping Luo. Learning depth-guided convolutions for monocular 3D object detection. In CVPRW, 2020. 2", + "[19] Huan Fu, Mingming Gong, Chaohui Wang, Kayhan Bat-manghelich, and Dacheng Tao. Deep Ordinal Regression Network for Monocular Depth Estimation. In CVPR, 2018. 2", + "[20] Nils Gählert, Nicolas Jourdan, Marius Cordts, Uwe Franke, and Joachim Denzler. Cityscapes 3D: Dataset and benchmark for 9 DOF vehicle detection. In CVPRW, 2020. 2", + "[21] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the KITTI vision benchmark suite. In CVPR, 2012. 1, 2, 3, 6", + "[22] Georgia Gkioxari, Jitendra Malik, and Justin Johnson. Mesh R-CNN. In ICCV, 2019. 2", + "[23] Jiaqi Gu, Bojian Wu, Lubin Fan, Jianqiang Huang, Shen Cao, Zhiyu Xiang, and Xian-Sheng Hua. Homography loss for monocular 3D object detection. In CVPR, 2022. 2", + "[24] Agrim Gupta, Piotr Dollar, and Ross Girshick. LVIS: A dataset for large vocabulary instance segmentation. In CVPR, 2019. 1, 3", + "[25] Chenhang He, Jianqiang Huang, Xian-Sheng Hua, and Lei Zhang. Aug3D-RPN: Improving monocular 3D object detection by synthetic images with virtual depth. arXiv preprint arXiv:2107.13269, 2021. 2", + "[26] Kaiming He, Georgia Gkioxari, Piotr Dollár, and Ross Girshick. Mask R-CNN. In ICCV, 2017. 1, 2, 4", + "[27] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 8", + "[28] Varsha Hedau, Derek Hoiem, and David Forsyth. Recovering the spatial layout of cluttered rooms. In ICCV, 2009. 2", + "[29] John Houston, Guido Zuidhof, Luca Bergamini, Yawei Ye, Long Chen, Ashesh Jain, Sammy Omari, Vladimir Iglovikov, and Peter Ondruska. One thousand and one hours: Self-driving motion prediction dataset. arXiv preprint arXiv:2006.14480, 2020. 2", + "[30] Kuan-Chih Huang, Tsung-Han Wu, Hung-Ting Su, and Winston Hsu. MonoDTR: Monocular 3D object detection with depth-aware transformer. In CVPR, 2022. 2, 7", + "[31] Siyuan Huang, Siyuan Qi, Yinxue Xiao, Yixin Zhu, Ying Nian Wu, and Song-Chun Zhu. Cooperative holistic scene understanding: Unifying 3D object, layout, and camera pose estimation. In NeurIPS, 2018. 1, 2", + "[32] Xinyu Huang, Peng Wang, Xinjing Cheng, Dingfu Zhou, Qichuan Geng, and Ruigang Yang. The apolloscape open dataset for autonomous driving and its application. TPAMI, 2019. 2" + ], + "bbox": [ + 501, + 92, + 893, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "13162", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[33] R. Kesten, M. Usman, J. Houston, T. Pandya, K. Nadhamuni, A. Ferreira, M. Yuan, B. Low, A. Jain, P. Ondruska, S. Omari, S. Shah, A. Kulkarni, A. Kazakova, C. Tao, L. Platinsky, W. Jiang, and V. Shet. Level 5 perception dataset 2020. https://level-5.global/level5/data/, 2019.2", + "[34] Dahun Kim, Tsung-Yi Lin, Anelia Angelova, In Kweon, and Weicheng Kuo. Learning open-world object proposals without learning to classify. RAL, 2022. 4", + "[35] Jason Ku, Melissa Mozifian, Jungwook Lee, Ali Harakeh, and Steven Waslander. Joint 3D proposal generation and object detection from view aggregation. In IROS, 2018. 2", + "[36] Nilesh Kulkarni, Ishan Misra, Shubham Tulsiani, and Abhinav Gupta. 3D-RelNet: Joint object and relational network for 3D prediction. In ICCV, 2019. 1, 2", + "[37] Abhinav Kumar, Garrick Brazil, Enrique Corona, Armin Parchami, and Xiaoming Liu. DEVIANT: Depth EquiVarIAnt NeTwork for monocular 3D object detection. In ECCV, 2022. 2", + "[38] Abhinav Kumar, Garrick Brazil, and Xiaoming Liu. GrooMeD-NMS: Grouped mathematically differentiable NMS for monocular 3D object detection. In CVPR, 2021. 2", + "[39] Abhijit Kundu, Yin Li, and James Rehg. 3D-RCNN: Instance-level 3D object reconstruction via render-and-compare. In CVPR, 2018. 5", + "[40] Hei Law and Jia Deng. CornerNet: Detecting objects as paired keypoints. In ECCV, 2018. 1", + "[41] David C Lee, Martial Hebert, and Takeo Kanade. Geometric reasoning for single image structure recovery. In CVPR, 2009. 2", + "[42] Peixuan Li, Huaici Zhao, Pengfei Liu, and Feidao Cao. RTM3D: Real-time monocular 3D detection from object keypoints for autonomous driving. In ECCV, 2020. 2", + "[43] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In CVPR, 2017. 6", + "[44] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In ICCV, 2017. 2", + "[45] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dolkar, and C Lawrence Zitnick. Microsoft COCO: Common objects in context. In ECCV, 2014. 1, 3, 8", + "[46] Ce Liu, Shuhang Gu, Luc Van Gool, and Radu Timofte. Deep line encoding for monocular 3D object detection and depth prediction. In BMVC, 2021. 2", + "[47] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander Berg. SSD: Single shot multibox detector. In ECCV, 2016. 2", + "[48] Xianpeng Liu, Nan Xue, and Tianfu Wu. Learning auxiliary monocular contexts helps monocular 3D object detection. In AAAI, 2022. 2", + "[49] Yuxuan Liu, Yuan Yixuan, and Ming Liu. Ground-aware monocular 3D object detection for autonomous driving. RAL, 2021. 2", + "[50] Zechen Liu, Zizhang Wu, and Roland Tóth. SMOKE: Single-stage monocular 3D object detection via keypoint estimation. In CVPRW, 2020. 1, 2, 5, 7" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[51] Zongdai Liu, Dingfu Zhou, Feixiang Lu, Jin Fang, and Liangjun Zhang. Autoshape: Real-time shape-aware monocular 3D object detection. In ICCV, 2021. 2", + "[52] Yan Lu, Xinzhu Ma, Lei Yang, Tianzhu Zhang, Yating Liu, Qi Chu, Junjie Yan, and Wanli Ouyang. Geometry uncertainty projection network for monocular 3D object detection. In ICCV, 2021. 1, 2, 5, 7", + "[53] Xinzhu Ma, Shinan Liu, Zhiyi Xia, Hongwen Zhang, Xingyu Zeng, and Wanli Ouyang. Rethinking pseudo-lidar representation. In ECCV, 2020. 2", + "[54] Xinzhu Ma, Zhihui Wang, Haojie Li, Pengbo Zhang, Wanli Ouyang, and Xin Fan. Accurate monocular 3D object detection via color-embedded 3D reconstruction for autonomous driving. In ICCV, 2019. 2", + "[55] Xinzhu Ma, Yinmin Zhang, Dan Xu, Dongzhan Zhou, Shuai Yi, Haojie Li, and Wanli Ouyang. Delving into localization errors for monocular 3D object detection. In CVPR, 2021. 2", + "[56] Arun Mallya and Svetlana Lazebnik. Learning informative edge maps for indoor scene layout prediction. In ICCV, 2015. 2", + "[57] Arsalan Mousavian, Dragomir Anguelov, John Flynn, and Jana Kosecka. 3D bounding box estimation using deep learning and geometry. In CVPR, 2017. 1, 2", + "[58] Yinyu Nie, Xiaoguang Han, Shihui Guo, Yujuan Zheng, Jian Chang, and Jian Zhang. Total3DUnderstanding: Joint layout, object pose and mesh reconstruction for indoor scenes from a single image. In CVPR, 2020. 1, 2, 8", + "[59] Dennis Park, Rares Ambrus, Vitor Guizilini, Jie Li, and Adrien Gaidon. Is pseudo-lidar needed for monocular 3D object detection? In ICCV, 2021. 2, 7", + "[60] Charles Qi, Wei Liu, Chenxia Wu, Hao Su, and Leonidas Guibas. Frustum pointnets for 3D object detection from RGB-D data. In CVPR, 2018. 2", + "[61] Nikhila Ravi, Jeremy Reizenstein, David Novotny, Taylor Gordon, Wan-Yen Lo, Justin Johnson, and Georgia Gkioxari. Accelerating 3D deep learning with PyTorch3D. arXiv:2007.08501, 2020. 6", + "[62] Cody Reading, Ali Harakeh, Julia Chae, and Steven Waslander. Categorical depth distribution network for monocular 3D object detection. In CVPR, 2021. 2", + "[63] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In CVPR, 2016. 1, 2", + "[64] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster R-CNN: Towards real-time object detection with region proposal networks. In NeurIPS, 2015. 1, 2, 4, 5", + "[65] Mike Roberts, Jason Ramapuram, Anurag Ranjan, Atulit Kumar, Miguel Bautista, Nathan Paczan, Russ Webb, and Joshua Susskind. Hypersim: A photorealistic synthetic dataset for holistic indoor scene understanding. In ICCV, 2021. 2, 3", + "[66] Danila Rukhovich, Anna Vorontsova, and Anton Konushin. ImVoxelNet: Image to voxels projection for monocular and multi-view general-purpose 3D object detection. In WACV, 2022. 2, 7, 8", + "[67] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander Berg, and Li" + ], + "bbox": [ + 501, + 92, + 892, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "13163", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. IJCV, 2015. 6", + "[68] Andrea Simonelli, Samuel Rota Bulo, Lorenzo Porzi, Manuel López-Antequera, and Peter Kontschieder. Disentangling monocular 3D object detection. In ICCV, 2019. 2, 4, 5", + "[69] Andrea Simonelli, Samuel Rota Bulo, Lorenzo Porzi, Elisa Ricci, and Peter Kontschieder. Towards generalization across depth for monocular 3D object detection. In ECCV, 2020. 2", + "[70] Shuran Song, Samuel Lichtenberg, and Jianxiong Xiao. SUN RGB-D: A RGB-D scene understanding benchmark suite. In CVPR, 2015. 1, 2, 3", + "[71] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, Vijay Vasudevan, Wei Han, Jiquan Ngiam, Hang Zhao, Aleksei Timofeev, Scott Ettinger, Maxim Krivokon, Amy Gao, Aditya Joshi, Yu Zhang, Jonathon Shlens, Zhifeng Chen, and Dragomir Anguelov. Scalability in perception for autonomous driving: Waymo open dataset. In CVPR, 2020. 2", + "[72] Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. FCOS: Fully convolutional one-stage object detection. In ICCV, 2019. 1, 2", + "[73] Shubham Tulsiani, Saurabh Gupta, David Fouhey, Alexei Efros, and Jitendra Malik. Factoring shape, pose, and layout from the 2D image of a 3D scene. In CVPR, 2018. 1, 2", + "[74] Pauli Virtanen, Ralf Gommers, Travis Oliphant, Matt Haberland, Tyler Reddy, David Cournapeau, Evgeni Burovski, Pearu Peterson, Warren Weckesser, Jonathan Bright, et al. Scipy 1.0: fundamental algorithms for scientific computing in python. Nature methods, 2020. 6", + "[75] Li Wang, Liang Du, Xiaqing Ye, Yanwei Fu, Guodong Guo, Xiangyang Xue, Jianfeng Feng, and Li Zhang. Depth-conditioned dynamic message propagation for monocular 3D object detection. In CVPR, 2021. 2", + "[76] Li Wang, Li Zhang, Yi Zhu, Zhi Zhang, Tong He, Mu Li, and Xiangyang Xue. Progressive coordinate transforms for monocular 3D object detection. In NeurIPS, 2021. 2", + "[77] Tai Wang, Zhu Xinge, Jiangmiao Pang, and Dahua Lin. Probabilistic and geometric depth: Detecting objects in perspective. In CoRL, 2022. 2, 7", + "[78] Tai Wang, Xinge Zhu, Jiangmiao Pang, and Dahua Lin. FCOS3D: Fully convolutional one-stage monocular 3D object detection. In ICCVW, 2021. 2, 7", + "[79] Yan Wang, Wei-Lun Chao, Divyansh Garg, Bharath Hariharan, Mark Campbell, and Kilian Weinberger. Pseudo-lidar from visual depth estimation: Bridging the gap in 3D object detection for autonomous driving. In CVPR, 2019. 2", + "[80] Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. Detector2. https://github.com/facebookresearch/detectron2, 2019. 2, 6", + "[81] Yurong You, Yan Wang, Wei-Lun Chao, Divyansh Garg, Geoff Pleiss, Bharath Hariharan, Mark Campbell, and Kilian Weinberger. Pseudo-lidar++: Accurate depth for 3D object detection in autonomous driving. In ICLR, 2020. 2", + "[82] Fisher Yu, Dequan Wang, Evan Shelhamer, and Trevor Darrell. Deep layer aggregation. In CVPR, 2018. 6", + "[83] Yunpeng Zhang, Jiwen Lu, and Jie Zhou. Objects are different: Flexible monocular 3D object detection. In CVPR, 2021. 2" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[84] Dingfu Zhou, Xibin Song, Yuchao Dai, Junbo Yin, Feixiang Lu, Miao Liao, Jin Fang, and Liangjun Zhang. IAFA: Instance-aware feature aggregation for 3D object detection from a single image. In ACCV, 2020. 2", + "[85] Xingyi Zhou, Dequan Wang, and Philipp Krahenbuhl. Objects as points. In arXiv preprint arXiv:1904.07850, 2019. 2", + "[86] Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. In CVPR, 2019. 4, 5", + "[87] Yunsong Zhou, Yuan He, Hongzi Zhu, Cheng Wang, Hongyang Li, and Qinhong Jiang. Monocular 3D object detection: An extrinsic parameter free approach. In CVPR, 2021. 2", + "[88] Zhikang Zou, Xiaqing Ye, Liang Du, Xianhui Cheng, Xiao Tan, Li Zhang, Jianfeng Feng, Xiangyang Xue, and Errui Ding. The devil is in the task: Exploiting reciprocal appearance-localization features for monocular 3D object detection. In ICCV, 2021. 2" + ], + "bbox": [ + 501, + 92, + 893, + 345 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "13164", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/Omni3D_ A Large Benchmark and Model for 3D Object Detection in the Wild/2efbd142-17ac-4a1f-b83f-1ae025c49292_model.json b/2023/Omni3D_ A Large Benchmark and Model for 3D Object Detection in the Wild/2efbd142-17ac-4a1f-b83f-1ae025c49292_model.json new file mode 100644 index 0000000000000000000000000000000000000000..97bc72f83a0fa862cc389c33158703f6e94560ea --- /dev/null +++ b/2023/Omni3D_ A Large Benchmark and Model for 3D Object Detection in the Wild/2efbd142-17ac-4a1f-b83f-1ae025c49292_model.json @@ -0,0 +1,2741 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.089, + 0.131, + 0.882, + 0.152 + ], + "angle": 0, + "content": "OMNI3D: A Large Benchmark and Model for 3D Object Detection in the Wild" + }, + { + "type": "text", + "bbox": [ + 0.197, + 0.18, + 0.748, + 0.221 + ], + "angle": 0, + "content": "Garrick Brazil1 Abhinav Kumar2 Julian Straub1 Nikhila Ravi1 Justin Johnson1 Georgia Gkioxari3" + }, + { + "type": "text", + "bbox": [ + 0.279, + 0.227, + 0.686, + 0.246 + ], + "angle": 0, + "content": "\\(^{1}\\)Meta AI \\(^{2}\\)Michigan State University \\(^{3}\\)Caltech" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.284, + 0.309, + 0.432 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.318, + 0.287, + 0.498, + 0.43 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.287, + 0.708, + 0.431 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.72, + 0.287, + 0.891, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.438, + 0.895, + 0.467 + ], + "angle": 0, + "content": "Figure 1. Left: We introduce OMNI3D, a benchmark for 3D object detection which is larger and more diverse than popular 3D benchmarks. Right: We propose Cube R-CNN, which generalizes to unseen datasets (e.g. COCO [45]) and outperforms prior works on existing datasets." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.492, + 0.313, + 0.508 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.526, + 0.474, + 0.844 + ], + "angle": 0, + "content": "Recognizing scenes and objects in 3D from a single image is a longstanding goal of computer vision with applications in robotics and AR/VR. For 2D recognition, large datasets and scalable solutions have led to unprecedented advances. In 3D, existing benchmarks are small in size and approaches specialize in few object categories and specific domains, e.g. urban driving scenes. Motivated by the success of 2D recognition, we revisit the task of 3D object detection by introducing a large benchmark, called OMNI3D. OMNI3D re-purposes and combines existing datasets resulting in 234k images annotated with more than 3 million instances and 98 categories. 3D detection at such scale is challenging due to variations in camera intrinsic and the rich diversity of scene and object types. We propose a model, called Cube R-CNN, designed to generalize across camera and scene types with a unified approach. We show that Cube R-CNN outperforms prior works on the larger OMNI3D and existing benchmarks. Finally, we prove that OMNI3D is a powerful dataset for 3D object recognition and show that it improves single-dataset performance and can accelerate learning on new smaller datasets via pre-training.1" + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.492, + 0.633, + 0.508 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.52, + 0.895, + 0.687 + ], + "angle": 0, + "content": "Understanding objects and their properties from single images is a longstanding problem in computer vision with applications in robotics and AR/VR. In the last decade, 2D object recognition [26,40,63,64,72] has made tremendous advances toward predicting objects on the image plane with the help of large datasets [24,45]. However, the world and its objects are three dimensional laid out in 3D space. Perceiving objects in 3D from 2D visual inputs poses new challenges framed by the task of 3D object detection. Here, the goal is to estimate a 3D location and 3D extent of each object in an image in the form of a tight oriented 3D bounding box." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.897, + 0.902 + ], + "angle": 0, + "content": "Today 3D object detection is studied under two different lenses: for urban domains in the context of autonomous vehicles [7,13,50,52,57] or indoor scenes [31,36,58,73]. Despite the problem formulation being shared, methods share little insights between domains. Often approaches are tailored to work only for the domain in question. For instance, urban methods make assumptions about objects resting on a ground plane and model only yaw angles for 3D rotation. Indoor techniques may use a confined depth range (e.g. up to \\(6\\mathrm{m}\\) in [58]). These assumptions are generally not true in the real world. Moreover, the most popular benchmarks for image-based 3D object detection are small. Indoor SUN RGB-D [70] has \\(10\\mathrm{k}\\) images, urban KITTI [21] has \\(7\\mathrm{k}\\) images; 2D benchmarks like COCO [45] are \\(20\\times\\) larger." + }, + { + "type": "page_footnote", + "bbox": [ + 0.077, + 0.875, + 0.446, + 0.901 + ], + "angle": 0, + "content": "1We release the OMNI3D benchmark and Cube R-CNN models at https://github.com/facebookresearch/omni3d." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "13154" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.379 + ], + "angle": 0, + "content": "We address the absence of a general large-scale dataset for 3D object detection by introducing a large and diverse 3D benchmark called OMNI3D. OMNI3D is curated from publicly released datasets, SUN RBG-D [70], ARKitScenes [6], Hypersim [65], Objectron [2], KITTI [21] and nuScenes [9], and comprises 234k images with 3 million objects annotated with 3D boxes across 98 categories including chair, sofa, laptop, table, cup, shoes, pillow, books, car, person, etc. Sec. 3 describes the curation process which involves re-purposing the raw data and annotations from the aforementioned datasets, which originally target different applications. As shown in Fig. 1, OMNI3D is \\(20\\times\\) larger than existing popular benchmarks used for 3D detection, SUN RGB-D and KITTI. For efficient evaluation on the large OMNI3D, we introduce a new algorithm for intersection-over-union of 3D boxes which is \\(450\\times\\) faster than previous solutions [2]. We empirically prove the impact of OMNI3D as a large-scale dataset and show that it improves single-dataset performance by up to \\(5.3\\%\\) AP on urban and \\(3.8\\%\\) on indoor benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.379, + 0.473, + 0.696 + ], + "angle": 0, + "content": "On the large and diverse OMNI3D, we design a general and simple 3D object detector, called Cube R-CNN, inspired by advances in 2D and 3D recognition of recent years [22,52,64,68]. Cube R-CNN detects all objects and their 3D location, size and rotation end-to-end from a single image of any domain and for many object categories. Attributed to OMNI3D's diversity, our model shows strong generalization and outperforms prior works for indoor and urban domains with one unified model, as shown in Fig. 1. Learning from such diverse data comes with challenges as OMNI3D contains images of highly varying focal lengths which exaggerate scale-depth ambiguity (Fig. 4). We remedy this by operating on virtual depth which transforms object depth with the same virtual camera intrinsics across the dataset. An added benefit of virtual depth is that it allows the use of data augmentations (e.g. image rescaling) during training, which is a critical feature for 2D detection [12,80], and as we show, also for 3D. Our approach with one unified design outperforms prior best approaches in \\(\\mathrm{AP}_{\\mathrm{3D}}\\), ImVoxelNet [66] by \\(4.1\\%\\) on indoor SUN RGB-D, GUPNet [52] by \\(9.5\\%\\) on urban KITTI, and PGD [77] by \\(7.9\\%\\) on OMNI3D." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.697, + 0.32, + 0.711 + ], + "angle": 0, + "content": "We summarize our contributions:" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.724, + 0.47, + 0.77 + ], + "angle": 0, + "content": "- We introduce OMNI3D, a benchmark for image-based 3D object detection sourced from existing 3D datasets, which is \\(20 \\times\\) larger than existing 3D benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.782, + 0.47, + 0.813 + ], + "angle": 0, + "content": "- We implement a new algorithm for IoU of 3D boxes, which is \\(450 \\times\\) faster than prior solutions." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.825, + 0.47, + 0.901 + ], + "angle": 0, + "content": "- We design a general-purpose baseline method, Cube R-CNN, which tackles 3D object detection for many categories and across domains with a unified approach. We propose virtual depth to eliminate the ambiguity from varying camera focal lengths in OMNI3D." + }, + { + "type": "list", + "bbox": [ + 0.097, + 0.724, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.09, + 0.642, + 0.107 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.116, + 0.895, + 0.147 + ], + "angle": 0, + "content": "Cube R-CNN and OMNI3D draw from key research advances in 2D and 3D object detection." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.151, + 0.895, + 0.227 + ], + "angle": 0, + "content": "2D Object Detection. Here, methods include two-stage approaches [26, 64] which predict object regions with a region proposal network (RPN) and then refine them via an MLP. Single-stage detectors [44, 47, 63, 72, 85] omit the RPN and predict regions directly from the backbone." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.231, + 0.895, + 0.503 + ], + "angle": 0, + "content": "3D Object Detection. Monocular 3D object detectors predict 3D cuboids from single input images. There is extensive work in the urban self-driving domain where the car class is at the epicenter [13, 19, 23, 25, 30, 46, 49, 54, 57, 62, 68, 69, 76, 78, 84]. CenterNet [85] predicts 3D depth and size from fully-convolutional center features, and is extended by [14, 37, 42, 48, 50–52, 55, 83, 84, 87]. M3D-RPN [7] trains an RPN with 3D anchors, enhanced further by [8, 18, 38, 75, 88]. FCOS3D [78] extends the anchorless FCOS [72] detector to predict 3D cuboids. Its successor PGD [77] furthers the approach with probabilistic depth uncertainty. Others use pseudo depth [4, 15, 53, 59, 79, 81] and explore depth and point-based LiDAR techniques [35, 60]. Similar to ours, [11, 68, 69] add a 3D head, specialized for urban scenes and objects, on two-stage Faster R-CNN. [25, 69] augment their training by synthetically generating depth and box-fitted views, coined as virtual views or depth. In our work, virtual depth aims at addressing varying focal lengths." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.505, + 0.895, + 0.671 + ], + "angle": 0, + "content": "For indoor scenes, a vast line of work tackles room layout estimation [17,28,41,56]. Huang et al [31] predict 3D oriented bounding boxes for indoor objects. Factored3D [73] and 3D-RelNet [36] jointly predict object voxel shapes. Total3D [58] predicts 3D boxes and meshes by additionally training on datasets with annotated 3D shapes. ImVoxelNet [66] proposes domain-specific methods which share an underlying framework for processing volumes of 3D voxels. In contrast, we explore 3D object detection in its general form by tackling outdoor and indoor domains jointly in a single model and with a vocabulary of \\(5 \\times\\) more categories." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.674, + 0.897, + 0.901 + ], + "angle": 0, + "content": "3D Datasets. KITTI [21] and SUN RGB-D [70] are popular datasets for 3D object detection on urban and indoor scenes respectively. Since 2019, 3D datasets have emerged, both for indoor [2, 3, 6, 16, 65] and outdoor [9, 10, 20, 29, 32, 33, 71]. In isolation, these datasets target different tasks and applications and have unique properties and biases, e.g. object and scene types, focal length, coordinate systems, etc. In this work, we unify existing representative datasets [2, 6, 9, 21, 65, 70]. We process the raw visual data, re-purpose their annotations, and carefully curate the union of their semantic labels in order to build a coherent large-scale benchmark, called OMNI3D. OMNI3D is \\(20 \\times\\) larger than widely-used benchmarks and notably more diverse. As such, new challenges arise stemming from the increased variance in visual domain, object rotation, size, layouts, and camera intrinsics." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "13155" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.086, + 0.445, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.202, + 0.228, + 0.333, + 0.243 + ], + "angle": 0, + "content": "(a) Spatial statistics" + }, + { + "type": "image", + "bbox": [ + 0.447, + 0.107, + 0.657, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.485, + 0.228, + 0.631, + 0.241 + ], + "angle": 0, + "content": "(b) 2D Scale statistics" + }, + { + "type": "image", + "bbox": [ + 0.66, + 0.108, + 0.874, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.707, + 0.228, + 0.851, + 0.242 + ], + "angle": 0, + "content": "(c) Category statistics" + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.246, + 0.894, + 0.274 + ], + "angle": 0, + "content": "Figure 2. OMNI3D analysis. (a) distribution of object centers on (top) normalized image, XY-plane, and (bottom) normalized depth, XZ-plane, (b) relative 2D object scales, and (c) category frequency." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.289, + 0.32, + 0.305 + ], + "angle": 0, + "content": "3. The OMNI3D Benchmark" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.315, + 0.471, + 0.435 + ], + "angle": 0, + "content": "The primary benchmarks for 3D object detection are small, focus on a few categories and are of a single domain. For instance, the popular KITTI [21] contains only urban scenes, has 7k images and 3 categories, with a focus on car. SUN RGB-D [70] has 10k images. The small size and lack of variance in 3D datasets is a stark difference to 2D counterparts, such as COCO [45] and LVIS [24], which have pioneered progress in 2D recognition." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.437, + 0.474, + 0.679 + ], + "angle": 0, + "content": "We aim to bridge the gap to 2D by introducing OMNI3D, a large-scale and diverse benchmark for image-based 3D object detection consisting of \\(234\\mathrm{k}\\) images, 3 million labeled 3D bounding boxes, and 98 object categories. We source from recently released 3D datasets of urban (nuScenes [9] and KITTI [21]), indoor (SUN RGB-D [70], ARKitScenes [6] and Hypersim [65]), and general (Objectron [2]) scenes. Each of these datasets target different applications (e.g. point-cloud recognition or reconstruction in [6], inverse rendering in [65]), provide visual data in different forms (e.g. videos in [2, 6], rig captures in [9]) and annotate different object types. To build a coherent benchmark, we process the varying raw visual data, re-purpose their annotations to extract 3D cuboids in a unified 3D camera coordinate system, and carefully curate the final vocabulary. More details about the benchmark creation in the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.679, + 0.472, + 0.755 + ], + "angle": 0, + "content": "We analyze OMNI3D and show its rich spatial and semantic properties proving it is visually diverse, similar to 2D data, and highly challenging for 3D as depicted in Fig. 2. We show the value of OMNI3D for the task of 3D object detection with extensive quantitative analysis in Sec. 5." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.768, + 0.241, + 0.784 + ], + "angle": 0, + "content": "3.1. Dataset Analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Splits. We split the dataset into \\(175\\mathrm{k} / 19\\mathrm{k} / 39\\mathrm{k}\\) images for train/val/test respectively, consistent with original splits when available, and otherwise free of overlapping video sequences in splits. We denote indoor and outdoor subsets as \\(\\mathrm{OMN13D_{IN}}\\) (SUN RGB-D, Hypersim, ARKit), and \\(\\mathrm{OMN13D_{OUT}}\\) (KITTI, nuScenes). Objectron, with primarily close-up objects, is used only in the full OMN13D setting." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.29, + 0.895, + 0.456 + ], + "angle": 0, + "content": "Layout statistics. Fig. 2(a) shows the distribution of object centers onto the image plane by projecting centroids on the XY-plane (top row), and the distribution of object depths by projecting centroids onto the XZ-plane (bottom row). We find that OMNI3D's spatial distribution has a center bias, similar to 2D datasets COCO and LVIS. Fig. 2(b) depicts the relative object size distribution, defined as the square root of object area divided by image area. Objects are more likely to be small in size similar to LVIS (Fig. 6c in [24]) suggesting that OMNI3D is also challenging for 2D detection, while objects in OMNI3DOUT are noticeably smaller." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.456, + 0.895, + 0.608 + ], + "angle": 0, + "content": "The bottom row of Fig. 2(a) normalizes object depth in a \\([0,20\\mathrm{m}]\\) range, chosen for visualization and satisfies \\(88\\%\\) of object instances; OMNI3D depth ranges as far as \\(300\\mathrm{m}\\). We observe that the OMNI3D depth distribution is far more diverse than SUN RGB-D and KITTI, which are biased toward near or road-side objects respectively. See Appendix for each data source distribution plot. Fig. 2(a) demonstrates OMNI3D's rich diversity in spatial distribution and depth which suffers significantly less bias than existing 3D benchmarks and is comparable in complexity to 2D datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.61, + 0.895, + 0.748 + ], + "angle": 0, + "content": "2D and 3D correlation. A common assumption in urban scenes is that objects rest on a ground plane and appear smaller with depth. To verify if that is true generally, we compute correlations. We find that \\(2\\mathrm{D}y\\) and \\(3\\mathrm{D}z\\) are indeed fairly correlated in \\(\\mathrm{OMNI3D_{OUT}}\\) at 0.524, but significantly less in \\(\\mathrm{OMNI3D_{IN}}\\) at 0.006. Similarly, relative 2D object size \\((\\sqrt{h\\cdot w})\\) and \\(z\\) correlation is 0.543 and 0.102 respectively. This confirms our claim that common assumptions in urban scenes are not generally true, making the task challenging." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.897, + 0.902 + ], + "angle": 0, + "content": "Category statistics. Fig. 2(c) plots the distribution of instances across the 98 categories of OMNI3D. The long-tail suggests that low-shot recognition in both 2D and 3D will be critical for performance. In this work, we want to focus on large-scale and diverse 3D recognition, which is comparably unexplored. We therefore filter and focus on categories with at least 1000 instances. This leaves 50 diverse categories including chair, sofa, laptop, table, books, car, truck, pedestrian and more, 19 of which have more than 10k instances. We provide more per-category details in the Appendix." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13156" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.092, + 0.892, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.316, + 0.895, + 0.345 + ], + "angle": 0, + "content": "Figure 3. Overview. Cube R-CNN takes as input an RGB image, detects all objects in 2D and predicts their 3D cuboids \\( B_{3\\mathrm{D}} \\). During training, the 3D corners of the cuboids are compared against 3D ground truth with the point-cloud chamfer distance." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.362, + 0.169, + 0.379 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.389, + 0.475, + 0.496 + ], + "angle": 0, + "content": "Our goal is to design a simple and effective model for general 3D object detection. Hence, our approach is free of domain or object specific strategies. We design our 3D object detection framework by extending Faster R-CNN [64] with a 3D object head which predicts a cuboid per each detected 2D object. We refer to our method as Cube R-CNN. Figure 3 shows an overview of our approach." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.508, + 0.219, + 0.523 + ], + "angle": 0, + "content": "4.1. Cube R-CNN" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.533, + 0.47, + 0.699 + ], + "angle": 0, + "content": "Our model builds on Faster R-CNN [64], an end-to-end region-based object detection framework. Faster-RCNN consists of a backbone network, commonly a CNN, which embeds the input image into a higher-dimensional feature space. A region proposal network (RPN) predicts regions of interest (RoIs) representing object candidates in the image. A 2D box head inputs the backbone feature map and processes each RoI to predict a category and a more accurate 2D box. Faster R-CNN can be easily extended to tackle more tasks by adding task-specific heads, e.g. Mask R-CNN [26] adds a mask head to additionally output object silhouettes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.7, + 0.471, + 0.852 + ], + "angle": 0, + "content": "For the task of 3D object detection, we extend Faster R-CNN by introducing a 3D detection head which predicts a 3D cuboid for each detected 2D object. Cube R-CNN extends Faster R-CNN in three ways: (a) we replace the binary classifier in RPN which predicts region objectness with a regressor that predicts IoUness, (b) we introduce a cube head which estimates the parameters to define a 3D cuboid for each detected object (similar in concept to [68]), and (c) we define a new training objective which incorporates a virtual depth for the task of 3D object detection." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.473, + 0.903 + ], + "angle": 0, + "content": "IoUness. The role of RPN is two-fold: (a) it proposes RoIs by regressing 2D box coordinates from pre-computed anchors and (b) it classifies regions as object or not (object-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.364, + 0.895, + 0.593 + ], + "angle": 0, + "content": "ness). This is sensible in exhaustively labeled datasets where it can be reliably assessed if a region contains an object. However, OMNI3D combines many data sources with no guarantee that all instances of all classes are labeled. We overcome this by replacing objectness with IoUness, applied only to foreground. Similar to [34], a regressor predicts IoU between a RoI and a ground truth. Let \\( o \\) be the predicted IoU for a RoI and \\( \\hat{o} \\) be the 2D IoU between the region and its ground truth; we apply a binary cross-entropy (CE) loss \\( \\mathcal{L}_{\\mathrm{IoUness}} = \\ell_{\\mathrm{CE}}(o,\\hat{o}) \\). We train on regions whose IoU exceeds 0.05 with a ground truth in order to learn IoUness from a wide range of region overlaps. Thus, the RPN training objective becomes \\( \\mathcal{L}_{\\mathrm{RPN}} = \\hat{o}\\cdot (\\mathcal{L}_{\\mathrm{IoUness}} + \\mathcal{L}_{\\mathrm{reg}}) \\), where \\( \\mathcal{L}_{\\mathrm{reg}} \\) is the 2D box regression loss from [64]. The loss is weighted by \\( \\hat{o} \\) to prioritize candidates close to true objects." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.6, + 0.895, + 0.722 + ], + "angle": 0, + "content": "Cube Head. We extend Faster R-CNN with a new head, called cube head, to predict a 3D cuboid for each detected 2D object. The cube head inputs \\(7 \\times 7\\) feature maps pooled from the backbone for each predicted region and feeds them to 2 fully-connected (FC) layers with 1024 hidden dimensions. All 3D estimations in the cube head are category-specific. The cube head represents a 3D cuboid with 13 parameters each predicted by a final FC layer:" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.728, + 0.891, + 0.757 + ], + "angle": 0, + "content": "- \\([u, v]\\) represent the projected 3D center on the image plane relative to the 2D RoI" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.758, + 0.892, + 0.788 + ], + "angle": 0, + "content": "- \\(z \\in \\mathbb{R}_+\\) is the object's center depth in meters transformed from virtual depth \\(z_v\\) (explained below)" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.788, + 0.891, + 0.817 + ], + "angle": 0, + "content": "- \\([ \\bar{w}, \\bar{h}, \\bar{l} ] \\in \\mathbb{R}_{+}^{3}\\) are the log-normalized physical box dimensions in meters" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.818, + 0.884, + 0.834 + ], + "angle": 0, + "content": "- \\(\\mathbf{p} \\in \\mathbb{R}^6\\) is the continuous 6D [86] allocentric rotation" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.834, + 0.795, + 0.849 + ], + "angle": 0, + "content": "- \\(\\mu \\in \\mathbb{R}_+\\) is the predicted 3D uncertainty" + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.728, + 0.892, + 0.849 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.895, + 0.903 + ], + "angle": 0, + "content": "The above parameters form the final 3D box in camera view coordinates for each detected 2D object. The object's 3D center \\(\\mathbf{X}\\) is estimated from the predicted 2D projected" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13157" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.267, + 0.108 + ], + "angle": 0, + "content": "center \\([u,v]\\) and depth \\(z\\) via" + }, + { + "type": "equation", + "bbox": [ + 0.088, + 0.115, + 0.47, + 0.15 + ], + "angle": 0, + "content": "\\[\n\\mathbf {X} (u, v, z) = \\left(\\frac {z}{f _ {x}} \\left(r _ {x} + u r _ {w} - p _ {x}\\right), \\frac {z}{f _ {y}} \\left(r _ {y} + v r _ {h} - p _ {y}\\right), z\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.15, + 0.473, + 0.243 + ], + "angle": 0, + "content": "where \\([r_x, r_y, r_w, r_h]\\) is the object's 2D box, \\((f_x, f_y)\\) are the camera's known focal lengths and \\((p_x, p_y)\\) the principal point. The 3D box dimensions \\(\\mathbf{d}\\) are derived from \\([\\bar{w}, \\bar{h}, \\bar{l}]\\) which are log-normalized with category-specific precomputed means \\((w_0, h_0, l_0)\\) for width, height and length respectively, and are arranged into a diagonal matrix via" + }, + { + "type": "equation", + "bbox": [ + 0.114, + 0.249, + 0.47, + 0.267 + ], + "angle": 0, + "content": "\\[\n\\mathbf {d} (\\bar {w}, \\bar {h}, \\bar {l}) = \\operatorname {d i a g} \\left(\\exp (\\bar {w}) w _ {0}, \\exp (\\bar {h}) h _ {0}, \\exp (\\bar {l}) l _ {0}\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.274, + 0.471, + 0.351 + ], + "angle": 0, + "content": "Finally, we derive the object's pose \\(\\mathbf{R}(\\mathbf{p})\\) as a \\(3 \\times 3\\) rotation matrix based on a 6D parameterization (2 directional vectors) of \\(\\mathbf{p}\\) following [86] which is converted from allocentric to egocentric rotation similar to [39], defined formally in Appendix. The final 3D cuboid, defined by 8 corners, is" + }, + { + "type": "equation", + "bbox": [ + 0.088, + 0.357, + 0.47, + 0.387 + ], + "angle": 0, + "content": "\\[\nB _ {\\mathrm {3 D}} (u, v, z, \\bar {w}, \\bar {h}, \\bar {l}, \\mathbf {p}) = \\mathbf {R} (\\mathbf {p}) \\mathbf {d} (\\bar {w}, \\bar {h}, \\bar {l}) B _ {\\text {u n i t}} + \\mathbf {X} (u, v, z) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.387, + 0.471, + 0.464 + ], + "angle": 0, + "content": "where \\( B_{\\mathrm{unit}} \\) are the 8 corners of an axis-aligned unit cube centered at \\( (0,0,0) \\). Lastly, \\( \\mu \\) denotes a learned 3D uncertainty, which is mapped to a confidence at inference then joined with the classification score \\( s \\) from the 2D box head to form the final score for the prediction, as \\( \\sqrt{s \\cdot \\exp(-\\mu)} \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.466, + 0.473, + 0.693 + ], + "angle": 0, + "content": "Training objective. Our training objective consists of 2D losses from the RPN and 2D box head and 3D losses from the cube head. The 3D objective compares each predicted 3D cuboid with its matched ground truth via a chamfer loss, treating the 8 box corners of the 3D boxes as point clouds, namely \\(\\mathcal{L}_{\\mathrm{3D}}^{\\mathrm{all}} = \\ell_{\\mathrm{chamfer}}(B_{\\mathrm{3D}}, B_{\\mathrm{3D}}^{\\mathrm{gt}})\\). Note that \\(\\mathcal{L}_{\\mathrm{3D}}^{\\mathrm{all}}\\) entangles all 3D variables via the box predictor \\(B_{\\mathrm{3D}}\\) (Eq. 3), such that that errors in variables may be ambiguous from one another. Thus, we isolate each variable group with separate disentangled losses, following [68]. The disentangled loss for each variable group substitutes all but its variables with the ground truth from Eq. 3 to create a pseudo box prediction. For example, the disentangled loss for the projected center \\([u, v]\\) produces a 3D box with all but \\((u, v)\\) replaced with the true values and then compares to the ground truth box," + }, + { + "type": "equation", + "bbox": [ + 0.098, + 0.702, + 0.47, + 0.724 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {3 D}} ^ {(u, v)} = \\left\\| B _ {\\mathrm {3 D}} (u, v, z ^ {\\mathrm {g t}}, \\bar {w} ^ {\\mathrm {g t}}, \\bar {h} ^ {\\mathrm {g t}}, \\bar {l} ^ {\\mathrm {g t}}, \\mathbf {p} ^ {\\mathrm {g t}}) - B _ {\\mathrm {3 D}} ^ {\\mathrm {g t}} \\right\\| _ {1} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.732, + 0.47, + 0.861 + ], + "angle": 0, + "content": "We use an \\(L_{1}\\) loss for \\(\\mathcal{L}_{\\mathrm{3D}}^{(u,v)}\\), \\(\\mathcal{L}_{\\mathrm{3D}}^{(z)}\\) and \\(\\mathcal{L}_{\\mathrm{3D}}^{(\\bar{w},\\bar{h},\\bar{l})}\\) and chamfer for \\(\\mathcal{L}_{\\mathrm{3D}}^{\\mathbf{P}}\\) to account for cuboid symmetry such that rotation matrices of Euler angles modulo \\(\\pi\\) produce the same noncanonical 3D cuboid. Losses in box coordinates have natural advantages over losses which directly compare variables to ground truths in that their gradients are appropriately weighted by the error as shown in [68]. The 3D objective is defined as, \\(\\mathcal{L}_{\\mathrm{3D}} = \\mathcal{L}_{\\mathrm{3D}}^{(u,v)} + \\mathcal{L}_{\\mathrm{3D}}^{(z)} + \\mathcal{L}_{\\mathrm{3D}}^{(\\bar{w},\\bar{h},\\bar{l})} + \\mathcal{L}_{\\mathrm{3D}}^{\\mathbf{P}} + \\mathcal{L}_{\\mathrm{3D}}^{\\mathrm{all}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.861, + 0.442, + 0.875 + ], + "angle": 0, + "content": "The collective training objective of Cube R-CNN is," + }, + { + "type": "equation", + "bbox": [ + 0.124, + 0.884, + 0.47, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\mathrm {R P N}} + \\mathcal {L} _ {\\mathrm {2 D}} + \\sqrt {2} \\cdot \\exp (- \\mu) \\cdot \\mathcal {L} _ {\\mathrm {3 D}} + \\mu \\tag {5}\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.093, + 0.691, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.094, + 0.891, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.262, + 0.895, + 0.348 + ], + "angle": 0, + "content": "Figure 4. Varying camera intrinsics exaggerate scale-depth ambiguity – the same object at two different depths can project to the same image, as in (a) and (b). We address this by introducing a virtual camera which is invariant to intrinsics and transforms the object's depth to a virtual depth \\( z_v \\) such that the effective image size \\( H_v \\) and focal length \\( f_v \\) are consistent across images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.361, + 0.895, + 0.453 + ], + "angle": 0, + "content": "\\(\\mathcal{L}_{\\mathrm{RPN}}\\) is the RPN loss, described above, and \\(\\mathcal{L}_{2\\mathrm{D}}\\) is the 2D box head loss from [64]. The 3D loss is weighted by the predicted 3D uncertainty (inspired by [52]), such that the model may trade a penalty to reduce the 3D loss when uncertain. In practice, \\(\\mu\\) is helpful for both improving the 3D ratings at inference and reducing the loss of hard samples in training." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.466, + 0.644, + 0.483 + ], + "angle": 0, + "content": "4.2. Virtual Depth" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.49, + 0.895, + 0.642 + ], + "angle": 0, + "content": "A critical part of 3D object detection is predicting an object's depth in metric units. Estimating depth from visual cues requires an implicit mapping of 2D pixels to 3D distances, which is more ambiguous if camera intrinsics vary. Prior works are able to ignore this as they primarily train on images from a single sensor. OMNI3D contains images from many sensors and thus demonstrates large variations in camera intrinsics. We design Cube R-CNN to be robust to intrinsics by predicting the object's virtual depth \\( z_{v} \\), which projects the metric depth \\( z \\) to an invariant camera space." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.643, + 0.895, + 0.767 + ], + "angle": 0, + "content": "Virtual depth scales the depth using the (known) camera intrinsics such that the effective image size and focal length are consistent across the dataset, illustrated in Fig. 4. The effective image properties are referred to as virtual image height \\( H_{v} \\) and virtual focal length \\( f_{v} \\), and are both hyperparameters. If \\( z \\) is the true metric depth of an object from an image with height \\( H \\) and focal length \\( f \\), its virtual depth is defined as \\( z_{v} = z \\cdot \\frac{f_{v}}{f} \\frac{H}{H_{v}} \\). The derivation is in the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.766, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Invariance to camera intrinsics via virtual depth also enables scale augmentations during training, since \\( H \\) can vary without harming the consistency of \\( z_v \\). Data augmentations from image resizing tend to be critical for 2D models but are not often used in 3D methods [7, 50, 68] since if unaccounted for they increase scale-depth ambiguity. Virtual depth lessens such ambiguities and therefore enables powerful data augmentations in training. We empirically prove the two-fold effects of virtual depth in Sec. 5" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13158" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.227, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.228, + 0.089, + 0.373, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.089, + 0.481, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.481, + 0.089, + 0.63, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.63, + 0.09, + 0.776, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.776, + 0.09, + 0.885, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.174, + 0.226, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.227, + 0.174, + 0.373, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.174, + 0.481, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.481, + 0.174, + 0.63, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.63, + 0.174, + 0.777, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.777, + 0.174, + 0.885, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.26, + 0.237, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.238, + 0.26, + 0.395, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.396, + 0.26, + 0.482, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.482, + 0.26, + 0.639, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.64, + 0.26, + 0.796, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.797, + 0.26, + 0.885, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.329, + 0.892, + 0.357 + ], + "angle": 0, + "content": "Figure 5. Cube R-CNN predictions on OMNI3D test. For each example, we show the input image, the 3D predictions overlaid on the image and a top view with a base composed of \\(1\\mathrm{m}\\times 1\\mathrm{m}\\) tiles. See more results from images and videos in the Appendix." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.371, + 0.208, + 0.387 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.401, + 0.471, + 0.492 + ], + "angle": 0, + "content": "We tackle image-based 3D object detection on OMNI3D and compare to prior best methods. We evaluate Cube R-CNN on existing 3D benchmarks with a single unified model. Finally, we prove the effectiveness of OMNI3D as a large-scale 3D object detection dataset by showing comprehensive cross-dataset generalization and its impact for pre-training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.5, + 0.471, + 0.651 + ], + "angle": 0, + "content": "Implementation details. We implement Cube R-CNN using Detector2 [80] and PyTorch3D [61]. Our backbone is DLA34 [82]-FPN [43] pretrained on ImageNet [67]. We train for 128 epochs with a batch size of 192 images across 48 V100s. We use SGD with a learning rate of 0.12 which decays by a factor of 10 after \\(60\\%\\) and \\(80\\%\\) of training. During training, we use random data augmentation of horizontal flipping and scales \\(\\in [0.50, 1.25]\\), enabled by virtual depth. Virtual camera parameters are set to \\(f_v = H_v = 512\\). Source code and models are publicly available." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.473, + 0.901 + ], + "angle": 0, + "content": "Metric. Following popular benchmarks for 2D and 3D recognition, we use average-precision (AP) as our 3D metric. Predictions are matched to ground truth by measuring their overlap using \\(\\mathsf{I}\\circ \\mathsf{U}_{\\mathsf{3D}}\\) which computes the intersection-overunion of 3D cuboids. We compute a mean \\(\\mathrm{AP}_{\\mathrm{3D}}\\) across all 50 categories in OMNI3D and over a range of \\(\\mathsf{I}\\circ \\mathsf{U}_{\\mathsf{3D}}\\) thresholds \\(\\tau \\in [0.05,0.10,\\dots ,0.50]\\). The range of \\(\\tau\\) is more relaxed than in 2D to account for the new dimension - see the Appendix for a comparison between 3D and 2D ranges for IoU. General 3D objects may be occluded by other objects or truncated by the image border, and can have arbitrarily small projections. Following [21], we ignore objects with high occlusion (\\(>66\\%\\)) or truncation (\\(>33\\%\\)), and tiny projected objects (\\(< 6.25\\%\\) image height). Moreover, we also report \\(\\mathrm{AP}_{\\mathrm{3D}}\\) at varying levels of depth \\(d\\) as near: \\(0 < d \\leq 10m\\), medium: \\(10m < d \\leq 35m\\), far: \\(35m < d \\leq \\infty\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.372, + 0.895, + 0.644 + ], + "angle": 0, + "content": "Fast \\(\\mathsf{IOU}_{\\mathbf{3D}}\\). \\(\\mathsf{IOU}_{\\mathbf{3D}}\\) compares two 3D cuboids by computing their intersection-over-union. Images usually have many objects and produce several predictions so \\(\\mathsf{IOU}_{\\mathbf{3D}}\\) computations need to be fast. Prior implementations [21] approximate \\(\\mathsf{IOU}_{\\mathbf{3D}}\\) by projecting 3D boxes on a ground plane and multiply the top-view 2D intersection with the box heights to compute a 3D volume. Such approximations become notably inaccurate when objects are not on a planar ground or have an arbitrary orientation (e.g. with nonzero pitch or roll). Objectron [2] provides an exact solution, but relies on external libraries [5, 74] implemented in \\(\\mathbf{C}++\\) and is not batched. We implement a new, fast and exact algorithm which computes the intersecting shape by representing cuboids as meshes and finds face intersections – described in Appendix. Our algorithm is batched with \\(\\mathbf{C}++\\) and CUDA support. Our implementation is \\(90\\times\\) faster than Objectron in \\(\\mathbf{C}++\\), and \\(450\\times\\) faster in CUDA. As a result, evaluation on the large OMNI3D takes only seconds instead of several hours." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.659, + 0.688, + 0.673 + ], + "angle": 0, + "content": "5.1. Model Performance" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.684, + 0.894, + 0.79 + ], + "angle": 0, + "content": "First, we ablate the design choices of Cube R-CNN. Then, we compare to state-of-the-art methods on existing benchmarks and OMNI3D and show that it performs superior or on par with prior works which are specialized for their respective domains. Our Cube R-CNN uses a single unified design to tackle general 3D object detection across domains. Fig. 5 shows qualitative predictions on the OMNI3D test set." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.894, + 0.9 + ], + "angle": 0, + "content": "Ablations. Table 1 ablates the features of Cube R-CNN on OMNI3D. We report \\(\\mathrm{AP}_{3\\mathrm{D}}\\) performance at single IoU thresholds (0.25 and 0.50), and for near, medium and far objects. We further train and evaluate on domain-specific subsets, OMNI3DIN \\((\\mathrm{AP}_{3\\mathrm{D}}^{\\mathrm{IN}})\\) and OMNI3DOUT \\((\\mathrm{AP}_{3\\mathrm{D}}^{\\mathrm{OUT}})\\) to show how ablation trends hold for single domain scenarios. From Table 1 we draw a few standout observations:" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13159" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.182, + 0.087, + 0.788, + 0.219 + ], + "angle": 0, + "content": "
Cube R-CNNAP3DAP253DAP503DAPnear3DAPmed3DAPfar3DAPIN3DAPOUT3D
w/o disentangled22.624.39.226.811.88.214.530.9
w/o IoUness22.223.68.826.411.18.314.331.0
w/o L3D20.221.86.426.412.17.413.829.1
w/o scale aug.20.221.58.023.59.86.812.328.1
w/o scale | virtual19.821.27.523.48.65.712.226.0
w/o virtual depth17.318.44.422.77.97.113.230.3
w/o uncertainty μ17.218.35.420.510.87.09.125.4
ours23.324.99.527.912.18.515.031.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.22, + 0.895, + 0.248 + ], + "angle": 0, + "content": "Table 1. Cube R-CNN ablations. We report \\(\\mathrm{AP}_{3\\mathrm{D}}\\) on OMNI3D, then at IoU thresholds 0.25 and 0.50, and for near, medium and far objects. We further report \\(\\mathrm{AP}_{3\\mathrm{D}}\\) when both training and evaluating on OMNI3DIN and OMNI3DOUT subsets." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.25, + 0.476, + 0.419 + ], + "angle": 0, + "content": "
MethodfOMNI3DOUTOMNI3D
AP3DAPNU3DAPOUT3DAP3D
M3D-RPN [7]X10.417.913.7-
SMOKE [50]X25.420.419.59.6
FCOS3D [78]X14.620.917.69.8
PGD [77]X21.426.322.911.2
GUPNet [52]✓24.520.519.9-
M3D-RPN +vc✓16.2+5.820.4+2.917.0+3.3-
ImVoxelNet [66]✓23.523.421.59.4
SMOKE +vc✓25.9+0.520.4+0.020.0+0.510.4+0.8
FCOS3D +vc✓17.8+3.225.1+4.221.3+3.710.6+0.8
PGD +vc✓22.8+1.431.2+4.926.8+3.915.4+4.2
Cube R-CNN✓36.032.731.923.3
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.076, + 0.419, + 0.471, + 0.475 + ], + "angle": 0, + "content": "Table 2. OMNI3D comparison with GUPNet [52], ImVoxelNet [66], M3D-RPN [7], FCOS3D [78], PGD [77], SMOKE [50]. We apply virtual camera to the last four \\((+vc)\\) and show the gains in purple. \\(f\\) denotes if a method handles variance in focal length." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.49, + 0.471, + 0.597 + ], + "angle": 0, + "content": "Virtual depth is effective and improves \\(\\mathrm{AP}_{3\\mathrm{D}}\\) by \\(+6\\%\\), most noticeable in the full OMNI3D which has the largest variances in camera intrinsics. Enabled by virtual depth, scale augmentations during training increase \\(\\mathrm{AP}_{3\\mathrm{D}}\\) by \\(+3.1\\%\\) on OMNI3D, \\(+2.7\\%\\) on OMNI3D\\(_{\\text{IN}}\\) and \\(+3.8\\%\\) on OMNI3D\\(_{\\text{OUT}}\\). We find scale augmentation controlled without virtual depth to be harmful by \\(-2.5\\%\\) (rows 6 - 7)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.601, + 0.471, + 0.692 + ], + "angle": 0, + "content": "3D uncertainty boosts performance by about \\(+6\\%\\) on all OMNI3D subsets. Intuitively, it serves as a measure of confidence for the model's 3D predictions; removing it means that the model relies only on its 2D classification to score cuboids. If uncertainty is used only to scale samples in Eq. 5, but not at inference then \\(\\mathrm{AP}_{3\\mathrm{D}}\\) still improves but by \\(+1.6\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.693, + 0.471, + 0.755 + ], + "angle": 0, + "content": "Table 1 also shows that the entangled \\(\\mathcal{L}_{3\\mathrm{D}}^{\\mathrm{all}}\\) loss from Eq. 5 boosts \\(\\mathrm{AP}_{3\\mathrm{D}}\\) by \\(+3.1\\%\\) while disentangled losses contribute less \\((+0.7\\%)\\). Replacing objectness with IoUness in the RPN head improves \\(\\mathrm{AP}_{3\\mathrm{D}}\\) on OMNI3D by \\(+1.1\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.759, + 0.471, + 0.835 + ], + "angle": 0, + "content": "Comparison to other methods. We compare Cube R-CNN to prior best approaches. We choose representative state-of-the-art methods designed for the urban and indoor domains, and evaluate on our proposed OMNI3D benchmark and single-dataset benchmarks, KITTI and SUN RGB-D." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Comparisons on OMNI3D. Table 2 compares Cube R-CNN to M3D-RPN [7], a single-shot 3D anchor approach, FCOS3D [78] and its follow-up PGD [77], which use a fully convolutional one-stage model, GUPNet [52], which uses" + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.25, + 0.899, + 0.369 + ], + "angle": 0, + "content": "
MethodDAP703DAP70BEV
EasyMedHardEasyMedHard
SMOKE [50]X14.039.767.8420.8314.4912.75
ImVoxelNet [66]X17.1510.979.1525.1916.3713.58
PGD [77]X19.0511.769.3926.8916.5113.49
GUPNet [52]X22.2615.0213.1230.2921.1918.20
Cube R-CNNX23.5915.0112.5631.7021.2018.43
MonoDTR [30]✓21.9915.3912.7328.5920.3817.14
DD3D [59]✓23.1916.8714.3632.3523.4120.42
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.37, + 0.895, + 0.4 + ], + "angle": 0, + "content": "Table 3. KITTI leaderboard results on car for image-based approaches. D denotes methods trained with extra depth supervision." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.413, + 0.895, + 0.655 + ], + "angle": 0, + "content": "2D-3D geometry and camera focal length to derive depth, SMOKE [50] which predicts 3D object center and offsets densely and ImVoxelNet [66], which uses intrinsics to unproject 2D image features to a 3D volume followed by 3D convolutions. These methods originally experiment on urban domains, so we compare on OMNI3DOUT and the full OMNI3D. We train each method using public code and report the best of multiple runs after tuning their hyper-parameters to ensure best performance. We extend all but GUPNet and ImVoxelNet with a virtual camera to handle varying intrinsics, denoted by \\(+vc\\). GUPNet and ImVoxelNet dynamically use per-image intrinsics, which can naturally account for camera variation. M3D-RPN and GUPNet do not support multi-node training, which makes scaling to OMNI3D difficult and are omitted. For OMNI3DOUT, we report \\(\\mathrm{AP}_{3\\mathrm{D}}^{\\mathrm{OUT}}\\) on its test set and its subparts, \\(\\mathrm{AP}_{3\\mathrm{D}}^{\\mathrm{KIT}}\\) and \\(\\mathrm{AP}_{3\\mathrm{D}}^{\\mathrm{NU}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.655, + 0.895, + 0.791 + ], + "angle": 0, + "content": "From Table 2 we observe that our Cube R-CNN outperforms competitive methods on both \\(\\mathrm{OMNI3D_{OUT}}\\) and OMNI3D. The virtual camera extensions result in performance gains of varying degree for all approaches, e.g. \\(+4.2\\%\\) for PGD on OMNI3D, proving its impact as a general purpose feature for mixed dataset training for the task of 3D object detection. FCOS3D, SMOKE, and ImVoxelNet [66] struggle the most on OMNI3D, which is not surprising as they are typically tailored for dataset-specific model choices." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Comparisons on KITTI. Table 3 shows results on KITTI's test set using their server [1]. We train Cube R-CNN on KITTI only. Cube R-CNN is a general purpose 3D object detector; its design is not tuned for the KITTI benchmark. Yet, it performs better or on par with recent best methods which are heavily tailored for KITTI and is only slightly worse to models trained with extra depth supervision." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13160" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.088, + 0.089, + 0.457, + 0.156 + ], + "angle": 0, + "content": "
MethodTrained onI0U3DAP3D
Total3D [58]SUN RGB-D23.3-
ImVoxelNet [66]SUN RGB-D-30.6
Cube R-CNNSUN RGB-D36.234.7
Cube R-CNNOMNI3DIN37.835.4
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.076, + 0.156, + 0.472, + 0.199 + ], + "angle": 0, + "content": "Table 4. We compare to indoor models with common categories. For Total3D, we use oracle 2D detections to report \\(\\mathsf{I}\\circ \\mathsf{U}_{\\mathsf{3D}}\\) (\\(3^{\\mathrm{rd}}\\) col). For ImVoxelNet, we report the detection metric of \\(\\mathrm{AP}_{\\mathrm{3D}}\\) (\\(4^{\\mathrm{th}}\\) col)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.209, + 0.472, + 0.346 + ], + "angle": 0, + "content": "KITTI's \\(\\mathrm{AP}_{3\\mathrm{D}}\\) is fundamentally similar to ours but sets a single \\(\\mathsf{I}\\circ \\mathsf{U}_{3\\mathrm{D}}\\) threshold at 0.70, which behaves similar to a nearly perfect 2D threshold of 0.94 (see the Appendix). To remedy this, our \\(\\mathrm{AP}_{3\\mathrm{D}}\\) is a mean over many thresholds, inspired by COCO [45]. In contrast, nuScenes proposes a mean AP based on 3D center distance but omits size and rotation, with settings generally tailored to the urban domain (e.g. size and rotation variations are less extreme for cars). We provide more analysis with this metric in the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.348, + 0.472, + 0.576 + ], + "angle": 0, + "content": "Comparisons on SUN RGB-D. Table 4 compares Cube R-CNN to Total3D [58] and ImVoxelNet [66], two state-of-the-art methods on SUN RGB-D on 10 common categories. Total3D's public model requires 2D object boxes and categories as input. Hence, for a fair comparison, we use ground truth 2D detections as input to both our and their method, each trained using a ResNet34 [27] backbone and report mean IoU3D (3rd col). We compare to ImVoxelNet in the full 3D object detection setting and report AP3D (4th col). We compare Cube R-CNN from two respects, first trained on SUN RGB-D identical to the baselines, then trained on OMNI3DIN which subsumes SUN RGB-D. Our model outperforms Total3D by +12.9% and ImVoxelNet by +4.1% when trained on the same training set and increases the performance gap when trained on the larger OMNI3DIN." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.579, + 0.472, + 0.684 + ], + "angle": 0, + "content": "Zero-shot Performance. One commonly sought after property of detectors is zero-shot generalization to other datasets. Cube R-CNN outperforms all other methods at zero-shot. We find our model trained on KITTI achieves \\(12.7\\%\\) \\(\\mathrm{AP}_{3\\mathrm{D}}\\) on nuScenes; the second best is M3D-RPN+vc with \\(10.7\\%\\). Conversely, when training on nuScenes, our model achieves \\(20.2\\%\\) on KITTI; the second best is GUPNet with \\(17.3\\%\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.693, + 0.411, + 0.709 + ], + "angle": 0, + "content": "5.2. The Impact of the OMNI3D Benchmark" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.716, + 0.47, + 0.793 + ], + "angle": 0, + "content": "Sec. 5.1 analyzes the performance of Cube R-CNN for the task of 3D object detection. Now, we turn to OMNI3D and its impact as a large-scale benchmark. We show two use cases of OMNI3D: (a) a universal 3D dataset which integrates smaller ones, and (b) a pre-training dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.471, + 0.903 + ], + "angle": 0, + "content": "OMNI3D as a universal dataset. We treat OMNI3D as a dataset which integrates smaller single ones and show its impact on each one. We train Cube R-CNN on OMNI3D and compare to single-dataset training in Table 5, for the indoor (left) and urban (right) domain. \\(\\mathrm{AP}_{\\mathrm{3D}}\\) is reported on the category intersection (10 for indoor, 3 for outdoor) to ensure comparability. Training on OMNI3D and its domain-specific" + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.088, + 0.885, + 0.168 + ], + "angle": 0, + "content": "
Trained onAPH3DAPSUN3DAPAR3DTrained onAPKIT3DAPNU3D
Hypersim15.29.57.5KITTI37.112.7
SUN5.834.713.1nuScenes20.238.6
ARKit5.914.238.6OMNI3D37.835.8
OMNI3D19.032.638.2OMNI3DOUT42.439.0
OMNI3DIN17.835.441.2
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.499, + 0.169, + 0.892, + 0.196 + ], + "angle": 0, + "content": "Table 5. Cross-dataset performance on intersecting categories (10 for indoor, 3 for outdoor) for comparable cross-evaluations." + }, + { + "type": "image", + "bbox": [ + 0.534, + 0.207, + 0.856, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.545, + 0.349, + 0.846, + 0.364 + ], + "angle": 0, + "content": "Figure 6. Pre-training on OMNI3D- vs. ImageNet." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.375, + 0.895, + 0.482 + ], + "angle": 0, + "content": "subsets, \\(\\mathrm{OMN13D_{IN}}\\) and \\(\\mathrm{OMN13D_{OUT}}\\), results in higher performance compared to single-dataset training, signifying that our large OMN13D generalizes better and should be preferred over single dataset training. ARKit sees a \\(+2.6\\%\\) boost and KITTI \\(+5.3\\%\\). Except for Hypersim, the domain-specific subsets tend to perform better on their domain, which is not surprising given their distinct properties (Fig. 2)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.485, + 0.895, + 0.713 + ], + "angle": 0, + "content": "OMNI3D as a pre-training dataset. Next, we demonstrate the utility of OMNI3D for pre-training. In this setting, an unseen dataset is used for finetuning from a model pre-trained on OMNI3D. The motivation is to determine how a large-scale 3D dataset could accelerate low-shot learning with minimum need for costly 3D annotations on a new dataset. We choose SUN RGB-D and KITTI as our unseen datasets given their popularity and small size. We pre-train Cube R-CNN on OMNI3D-, which removes them from OMNI3D, then finetune the models using a \\(\\%\\) of their training data. The curves in Fig. 6 show the model quickly gains its upper-bound performance at a small fraction of the training data, when pre-trained on OMNI3D- vs. ImageNet, without any few-shot training tricks. A model finetuned on only \\(5\\%\\) of its target can achieve \\(>70\\%\\) of the upper-bound performance." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.727, + 0.619, + 0.743 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.903 + ], + "angle": 0, + "content": "We propose a large and diverse 3D object detection benchmark, OMNI3D, and a general purpose 3D object detector, Cube R-CNN. Models and data are publicly released. Our extensive analysis (Table 1, Fig. 5, Appendix) show the strength of general 3D object detection as well as the limitations of our method, e.g. localizing far away objects, and uncommon object types or contexts. Moreover, for accurate real-world 3D predictions, our method is limited by the assumption of known camera intrinsics, which may be minimized in future work using self-calibration techniques." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "13161" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.472, + 0.156 + ], + "angle": 0, + "content": "[1] KITTI server. https://www.cvlibs.net/datasets/kitti/eval_object.php?obj_benchmark=3d. Accessed: 2022-11-01. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.157, + 0.472, + 0.212 + ], + "angle": 0, + "content": "[2] Adel Ahmadyan, Liangkai Zhang, Artsiom Ablavatski, Jianing Wei, and Matthias Grundmann. Objectron: A large scale dataset of object-centric videos in the wild with pose annotations. CVPR, 2021. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.213, + 0.472, + 0.284 + ], + "angle": 0, + "content": "[3] Armen Avetisyan, Manuel Dahnert, Angela Dai, Manolis Savva, Angel X Chang, and Matthias Nießner. Scan2CAD: Learning CAD model alignment in RGB-D scans. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 2614-2623, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.284, + 0.472, + 0.325 + ], + "angle": 0, + "content": "[4] Wentao Bao, Bin Xu, and Zhenzhong Chen. MonoFENet: Monocular 3D object detection with feature enhancement networks. TIP, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.326, + 0.472, + 0.368 + ], + "angle": 0, + "content": "[5] Bradford Barber, David Dobkin, and Hannu Huhdanpaa. The quickhull algorithm for convex hulls. ACM Transactions on Mathematical Software (TOMS), 1996. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.368, + 0.472, + 0.451 + ], + "angle": 0, + "content": "[6] Gilad Baruch, Zhuoyuan Chen, Afshin Dehghan, Tal Dimry, Yuri Feigin, Peter Fu, Thomas Gebauer, Brandon Joffe, Daniel Kurz, Arik Schwartz, and Elad Shulman. ARKitScenes - a diverse real-world dataset for 3D indoor scene understanding using mobile RGB-D data. In NeurIPS Datasets and Benchmarks Track, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.452, + 0.472, + 0.493 + ], + "angle": 0, + "content": "[7] Garrick Brazil and Xiaoming Liu. M3D-RPN: Monocular 3D region proposal network for object detection. In ICCV, 2019. 1, 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.494, + 0.472, + 0.536 + ], + "angle": 0, + "content": "[8] Garrick Brazil, Gerard Pons-Moll, Xiaoming Liu, and Bernt Schiele. Kinematic 3D object detection in monocular video. In ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.536, + 0.472, + 0.592 + ], + "angle": 0, + "content": "[9] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuScenes: A multimodal dataset for autonomous driving. In CVPR, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.593, + 0.472, + 0.649 + ], + "angle": 0, + "content": "[10] Ming-Fang Chang, John Lambert, Patsorn Sangkloy, Jagjeet Singh, Slawomir Bak, Andrew Hartnett, De Wang, Peter Carr, Simon Lucey, Deva Ramanan, et al. Argoverse: 3D tracking and forecasting with rich maps. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.649, + 0.472, + 0.704 + ], + "angle": 0, + "content": "[11] Hansheng Chen, Yuyao Huang, Wei Tian, Zhong Gao, and Lu Xiong. MonoRUn: Monocular 3D object detection by reconstruction and uncertainty propagation. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.705, + 0.472, + 0.816 + ], + "angle": 0, + "content": "[12] Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jiarui Xu, Zheng Zhang, Dazhi Cheng, Chenchen Zhu, Tianheng Cheng, Qijie Zhao, Buyu Li, Xin Lu, Rui Zhu, Yue Wu, Jifeng Dai, Jingdong Wang, Jianping Shi, Wanli Ouyang, Chen Change Loy, and Dahua Lin. MMDetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.816, + 0.472, + 0.859 + ], + "angle": 0, + "content": "[13] Xiaozhi Chen, Kaustav Kundu, Ziyu Zhang, Huimin Ma, Sanja Fidler, and Raquel Urtasun. Monocular 3D object detection for autonomous driving. In CVPR, 2016. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.078, + 0.859, + 0.472, + 0.902 + ], + "angle": 0, + "content": "[14] Yongjian Chen, Lei Tai, Kai Sun, and Mingyang Li. MonoPair: Monocular 3D object detection using pairwise spatial relationships. In CVPR, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.115, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.134 + ], + "angle": 0, + "content": "[15] Yi-Nan Chen, Hang Dai, and Yong Ding. Pseudo-stereo for monocular 3D object detection in autonomous driving. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.134, + 0.894, + 0.189 + ], + "angle": 0, + "content": "[16] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. ScanNet: Richly-annotated 3D reconstructions of indoor scenes. In CVPR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.189, + 0.894, + 0.232 + ], + "angle": 0, + "content": "[17] Saumitro Dasgupta, Kuan Fang, Kevin Chen, and Silvio Savarese. Delay: Robust spatial layout estimation for cluttered indoor scenes. In CVPR, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.232, + 0.894, + 0.286 + ], + "angle": 0, + "content": "[18] Mingyu Ding, Yuqi Huo, Hongwei Yi, Zhe Wang, Jianping Shi, Zhiwu Lu, and Ping Luo. Learning depth-guided convolutions for monocular 3D object detection. In CVPRW, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.286, + 0.893, + 0.329 + ], + "angle": 0, + "content": "[19] Huan Fu, Mingming Gong, Chaohui Wang, Kayhan Bat-manghelich, and Dacheng Tao. Deep Ordinal Regression Network for Monocular Depth Estimation. In CVPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.329, + 0.893, + 0.371 + ], + "angle": 0, + "content": "[20] Nils Gählert, Nicolas Jourdan, Marius Cordts, Uwe Franke, and Joachim Denzler. Cityscapes 3D: Dataset and benchmark for 9 DOF vehicle detection. In CVPRW, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.371, + 0.893, + 0.413 + ], + "angle": 0, + "content": "[21] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the KITTI vision benchmark suite. In CVPR, 2012. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.413, + 0.893, + 0.44 + ], + "angle": 0, + "content": "[22] Georgia Gkioxari, Jitendra Malik, and Justin Johnson. Mesh R-CNN. In ICCV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.44, + 0.893, + 0.483 + ], + "angle": 0, + "content": "[23] Jiaqi Gu, Bojian Wu, Lubin Fan, Jianqiang Huang, Shen Cao, Zhiyu Xiang, and Xian-Sheng Hua. Homography loss for monocular 3D object detection. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.483, + 0.893, + 0.524 + ], + "angle": 0, + "content": "[24] Agrim Gupta, Piotr Dollar, and Ross Girshick. LVIS: A dataset for large vocabulary instance segmentation. In CVPR, 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.524, + 0.893, + 0.58 + ], + "angle": 0, + "content": "[25] Chenhang He, Jianqiang Huang, Xian-Sheng Hua, and Lei Zhang. Aug3D-RPN: Improving monocular 3D object detection by synthetic images with virtual depth. arXiv preprint arXiv:2107.13269, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.58, + 0.893, + 0.608 + ], + "angle": 0, + "content": "[26] Kaiming He, Georgia Gkioxari, Piotr Dollár, and Ross Girshick. Mask R-CNN. In ICCV, 2017. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.608, + 0.893, + 0.649 + ], + "angle": 0, + "content": "[27] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.649, + 0.893, + 0.679 + ], + "angle": 0, + "content": "[28] Varsha Hedau, Derek Hoiem, and David Forsyth. Recovering the spatial layout of cluttered rooms. In ICCV, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.679, + 0.893, + 0.747 + ], + "angle": 0, + "content": "[29] John Houston, Guido Zuidhof, Luca Bergamini, Yawei Ye, Long Chen, Ashesh Jain, Sammy Omari, Vladimir Iglovikov, and Peter Ondruska. One thousand and one hours: Self-driving motion prediction dataset. arXiv preprint arXiv:2006.14480, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.747, + 0.893, + 0.789 + ], + "angle": 0, + "content": "[30] Kuan-Chih Huang, Tsung-Han Wu, Hung-Ting Su, and Winston Hsu. MonoDTR: Monocular 3D object detection with depth-aware transformer. In CVPR, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.789, + 0.893, + 0.845 + ], + "angle": 0, + "content": "[31] Siyuan Huang, Siyuan Qi, Yinxue Xiao, Yixin Zhu, Ying Nian Wu, and Song-Chun Zhu. Cooperative holistic scene understanding: Unifying 3D object, layout, and camera pose estimation. In NeurIPS, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.893, + 0.9 + ], + "angle": 0, + "content": "[32] Xinyu Huang, Peng Wang, Xinjing Cheng, Dingfu Zhou, Qichuan Geng, and Ruigang Yang. The apolloscape open dataset for autonomous driving and its application. TPAMI, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13162" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.161 + ], + "angle": 0, + "content": "[33] R. Kesten, M. Usman, J. Houston, T. Pandya, K. Nadhamuni, A. Ferreira, M. Yuan, B. Low, A. Jain, P. Ondruska, S. Omari, S. Shah, A. Kulkarni, A. Kazakova, C. Tao, L. Platinsky, W. Jiang, and V. Shet. Level 5 perception dataset 2020. https://level-5.global/level5/data/, 2019.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.472, + 0.204 + ], + "angle": 0, + "content": "[34] Dahun Kim, Tsung-Yi Lin, Anelia Angelova, In Kweon, and Weicheng Kuo. Learning open-world object proposals without learning to classify. RAL, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.206, + 0.472, + 0.247 + ], + "angle": 0, + "content": "[35] Jason Ku, Melissa Mozifian, Jungwook Lee, Ali Harakeh, and Steven Waslander. Joint 3D proposal generation and object detection from view aggregation. In IROS, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.249, + 0.472, + 0.29 + ], + "angle": 0, + "content": "[36] Nilesh Kulkarni, Ishan Misra, Shubham Tulsiani, and Abhinav Gupta. 3D-RelNet: Joint object and relational network for 3D prediction. In ICCV, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.291, + 0.472, + 0.345 + ], + "angle": 0, + "content": "[37] Abhinav Kumar, Garrick Brazil, Enrique Corona, Armin Parchami, and Xiaoming Liu. DEVIANT: Depth EquiVarIAnt NeTwork for monocular 3D object detection. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.347, + 0.472, + 0.389 + ], + "angle": 0, + "content": "[38] Abhinav Kumar, Garrick Brazil, and Xiaoming Liu. GrooMeD-NMS: Grouped mathematically differentiable NMS for monocular 3D object detection. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.39, + 0.472, + 0.431 + ], + "angle": 0, + "content": "[39] Abhijit Kundu, Yin Li, and James Rehg. 3D-RCNN: Instance-level 3D object reconstruction via render-and-compare. In CVPR, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.433, + 0.472, + 0.461 + ], + "angle": 0, + "content": "[40] Hei Law and Jia Deng. CornerNet: Detecting objects as paired keypoints. In ECCV, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.462, + 0.472, + 0.502 + ], + "angle": 0, + "content": "[41] David C Lee, Martial Hebert, and Takeo Kanade. Geometric reasoning for single image structure recovery. In CVPR, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.504, + 0.472, + 0.546 + ], + "angle": 0, + "content": "[42] Peixuan Li, Huaici Zhao, Pengfei Liu, and Feidao Cao. RTM3D: Real-time monocular 3D detection from object keypoints for autonomous driving. In ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.547, + 0.472, + 0.588 + ], + "angle": 0, + "content": "[43] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In CVPR, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.59, + 0.472, + 0.63 + ], + "angle": 0, + "content": "[44] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In ICCV, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.632, + 0.472, + 0.687 + ], + "angle": 0, + "content": "[45] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dolkar, and C Lawrence Zitnick. Microsoft COCO: Common objects in context. In ECCV, 2014. 1, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.689, + 0.472, + 0.73 + ], + "angle": 0, + "content": "[46] Ce Liu, Shuhang Gu, Luc Van Gool, and Radu Timofte. Deep line encoding for monocular 3D object detection and depth prediction. In BMVC, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.732, + 0.472, + 0.773 + ], + "angle": 0, + "content": "[47] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander Berg. SSD: Single shot multibox detector. In ECCV, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.774, + 0.472, + 0.814 + ], + "angle": 0, + "content": "[48] Xianpeng Liu, Nan Xue, and Tianfu Wu. Learning auxiliary monocular contexts helps monocular 3D object detection. In AAAI, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.472, + 0.857 + ], + "angle": 0, + "content": "[49] Yuxuan Liu, Yuan Yixuan, and Ming Liu. Ground-aware monocular 3D object detection for autonomous driving. RAL, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.472, + 0.9 + ], + "angle": 0, + "content": "[50] Zechen Liu, Zizhang Wu, and Roland Tóth. SMOKE: Single-stage monocular 3D object detection via keypoint estimation. In CVPRW, 2020. 1, 2, 5, 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.134 + ], + "angle": 0, + "content": "[51] Zongdai Liu, Dingfu Zhou, Feixiang Lu, Jin Fang, and Liangjun Zhang. Autoshape: Real-time shape-aware monocular 3D object detection. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.893, + 0.19 + ], + "angle": 0, + "content": "[52] Yan Lu, Xinzhu Ma, Lei Yang, Tianzhu Zhang, Yating Liu, Qi Chu, Junjie Yan, and Wanli Ouyang. Geometry uncertainty projection network for monocular 3D object detection. In ICCV, 2021. 1, 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.893, + 0.233 + ], + "angle": 0, + "content": "[53] Xinzhu Ma, Shinan Liu, Zhiyi Xia, Hongwen Zhang, Xingyu Zeng, and Wanli Ouyang. Rethinking pseudo-lidar representation. In ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.235, + 0.893, + 0.29 + ], + "angle": 0, + "content": "[54] Xinzhu Ma, Zhihui Wang, Haojie Li, Pengbo Zhang, Wanli Ouyang, and Xin Fan. Accurate monocular 3D object detection via color-embedded 3D reconstruction for autonomous driving. In ICCV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.292, + 0.893, + 0.332 + ], + "angle": 0, + "content": "[55] Xinzhu Ma, Yinmin Zhang, Dan Xu, Dongzhan Zhou, Shuai Yi, Haojie Li, and Wanli Ouyang. Delving into localization errors for monocular 3D object detection. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.334, + 0.893, + 0.373 + ], + "angle": 0, + "content": "[56] Arun Mallya and Svetlana Lazebnik. Learning informative edge maps for indoor scene layout prediction. In ICCV, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.376, + 0.893, + 0.418 + ], + "angle": 0, + "content": "[57] Arsalan Mousavian, Dragomir Anguelov, John Flynn, and Jana Kosecka. 3D bounding box estimation using deep learning and geometry. In CVPR, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.42, + 0.893, + 0.475 + ], + "angle": 0, + "content": "[58] Yinyu Nie, Xiaoguang Han, Shihui Guo, Yujuan Zheng, Jian Chang, and Jian Zhang. Total3DUnderstanding: Joint layout, object pose and mesh reconstruction for indoor scenes from a single image. In CVPR, 2020. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.476, + 0.893, + 0.517 + ], + "angle": 0, + "content": "[59] Dennis Park, Rares Ambrus, Vitor Guizilini, Jie Li, and Adrien Gaidon. Is pseudo-lidar needed for monocular 3D object detection? In ICCV, 2021. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.519, + 0.893, + 0.56 + ], + "angle": 0, + "content": "[60] Charles Qi, Wei Liu, Chenxia Wu, Hao Su, and Leonidas Guibas. Frustum pointnets for 3D object detection from RGB-D data. In CVPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.562, + 0.893, + 0.616 + ], + "angle": 0, + "content": "[61] Nikhila Ravi, Jeremy Reizenstein, David Novotny, Taylor Gordon, Wan-Yen Lo, Justin Johnson, and Georgia Gkioxari. Accelerating 3D deep learning with PyTorch3D. arXiv:2007.08501, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.618, + 0.893, + 0.659 + ], + "angle": 0, + "content": "[62] Cody Reading, Ali Harakeh, Julia Chae, and Steven Waslander. Categorical depth distribution network for monocular 3D object detection. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.661, + 0.893, + 0.702 + ], + "angle": 0, + "content": "[63] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In CVPR, 2016. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.704, + 0.893, + 0.745 + ], + "angle": 0, + "content": "[64] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster R-CNN: Towards real-time object detection with region proposal networks. In NeurIPS, 2015. 1, 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.747, + 0.893, + 0.802 + ], + "angle": 0, + "content": "[65] Mike Roberts, Jason Ramapuram, Anurag Ranjan, Atulit Kumar, Miguel Bautista, Nathan Paczan, Russ Webb, and Joshua Susskind. Hypersim: A photorealistic synthetic dataset for holistic indoor scene understanding. In ICCV, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.804, + 0.893, + 0.857 + ], + "angle": 0, + "content": "[66] Danila Rukhovich, Anna Vorontsova, and Anton Konushin. ImVoxelNet: Image to voxels projection for monocular and multi-view general-purpose 3D object detection. In WACV, 2022. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.893, + 0.901 + ], + "angle": 0, + "content": "[67] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander Berg, and Li" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13163" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.47, + 0.12 + ], + "angle": 0, + "content": "Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. IJCV, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.162 + ], + "angle": 0, + "content": "[68] Andrea Simonelli, Samuel Rota Bulo, Lorenzo Porzi, Manuel López-Antequera, and Peter Kontschieder. Disentangling monocular 3D object detection. In ICCV, 2019. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.47, + 0.204 + ], + "angle": 0, + "content": "[69] Andrea Simonelli, Samuel Rota Bulo, Lorenzo Porzi, Elisa Ricci, and Peter Kontschieder. Towards generalization across depth for monocular 3D object detection. In ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.205, + 0.47, + 0.245 + ], + "angle": 0, + "content": "[70] Shuran Song, Samuel Lichtenberg, and Jianxiong Xiao. SUN RGB-D: A RGB-D scene understanding benchmark suite. In CVPR, 2015. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.247, + 0.471, + 0.356 + ], + "angle": 0, + "content": "[71] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, Vijay Vasudevan, Wei Han, Jiquan Ngiam, Hang Zhao, Aleksei Timofeev, Scott Ettinger, Maxim Krivokon, Amy Gao, Aditya Joshi, Yu Zhang, Jonathon Shlens, Zhifeng Chen, and Dragomir Anguelov. Scalability in perception for autonomous driving: Waymo open dataset. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.358, + 0.471, + 0.398 + ], + "angle": 0, + "content": "[72] Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. FCOS: Fully convolutional one-stage object detection. In ICCV, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.4, + 0.47, + 0.44 + ], + "angle": 0, + "content": "[73] Shubham Tulsiani, Saurabh Gupta, David Fouhey, Alexei Efros, and Jitendra Malik. Factoring shape, pose, and layout from the 2D image of a 3D scene. In CVPR, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.442, + 0.471, + 0.51 + ], + "angle": 0, + "content": "[74] Pauli Virtanen, Ralf Gommers, Travis Oliphant, Matt Haberland, Tyler Reddy, David Cournapeau, Evgeni Burovski, Pearu Peterson, Warren Weckesser, Jonathan Bright, et al. Scipy 1.0: fundamental algorithms for scientific computing in python. Nature methods, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.511, + 0.471, + 0.566 + ], + "angle": 0, + "content": "[75] Li Wang, Liang Du, Xiaqing Ye, Yanwei Fu, Guodong Guo, Xiangyang Xue, Jianfeng Feng, and Li Zhang. Depth-conditioned dynamic message propagation for monocular 3D object detection. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.567, + 0.471, + 0.607 + ], + "angle": 0, + "content": "[76] Li Wang, Li Zhang, Yi Zhu, Zhi Zhang, Tong He, Mu Li, and Xiangyang Xue. Progressive coordinate transforms for monocular 3D object detection. In NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.609, + 0.471, + 0.649 + ], + "angle": 0, + "content": "[77] Tai Wang, Zhu Xinge, Jiangmiao Pang, and Dahua Lin. Probabilistic and geometric depth: Detecting objects in perspective. In CoRL, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.651, + 0.47, + 0.691 + ], + "angle": 0, + "content": "[78] Tai Wang, Xinge Zhu, Jiangmiao Pang, and Dahua Lin. FCOS3D: Fully convolutional one-stage monocular 3D object detection. In ICCVW, 2021. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.692, + 0.471, + 0.747 + ], + "angle": 0, + "content": "[79] Yan Wang, Wei-Lun Chao, Divyansh Garg, Bharath Hariharan, Mark Campbell, and Kilian Weinberger. Pseudo-lidar from visual depth estimation: Bridging the gap in 3D object detection for autonomous driving. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.748, + 0.471, + 0.788 + ], + "angle": 0, + "content": "[80] Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. Detector2. https://github.com/facebookresearch/detectron2, 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.79, + 0.471, + 0.845 + ], + "angle": 0, + "content": "[81] Yurong You, Yan Wang, Wei-Lun Chao, Divyansh Garg, Geoff Pleiss, Bharath Hariharan, Mark Campbell, and Kilian Weinberger. Pseudo-lidar++: Accurate depth for 3D object detection in autonomous driving. In ICLR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.471, + 0.873 + ], + "angle": 0, + "content": "[82] Fisher Yu, Dequan Wang, Evan Shelhamer, and Trevor Darrell. Deep layer aggregation. In CVPR, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.874, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[83] Yunpeng Zhang, Jiwen Lu, and Jie Zhou. Objects are different: Flexible monocular 3D object detection. In CVPR, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.147 + ], + "angle": 0, + "content": "[84] Dingfu Zhou, Xibin Song, Yuchao Dai, Junbo Yin, Feixiang Lu, Miao Liao, Jin Fang, and Liangjun Zhang. IAFA: Instance-aware feature aggregation for 3D object detection from a single image. In ACCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.177 + ], + "angle": 0, + "content": "[85] Xingyi Zhou, Dequan Wang, and Philipp Krahenbuhl. Objects as points. In arXiv preprint arXiv:1904.07850, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.179, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[86] Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. In CVPR, 2019. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.221, + 0.894, + 0.275 + ], + "angle": 0, + "content": "[87] Yunsong Zhou, Yuan He, Hongzi Zhu, Cheng Wang, Hongyang Li, and Qinhong Jiang. Monocular 3D object detection: An extrinsic parameter free approach. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.277, + 0.894, + 0.346 + ], + "angle": 0, + "content": "[88] Zhikang Zou, Xiaqing Ye, Liang Du, Xianhui Cheng, Xiao Tan, Li Zhang, Jianfeng Feng, Xiangyang Xue, and Errui Ding. The devil is in the task: Exploiting reciprocal appearance-localization features for monocular 3D object detection. In ICCV, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13164" + } + ] +] \ No newline at end of file diff --git a/2023/Omni3D_ A Large Benchmark and Model for 3D Object Detection in the Wild/2efbd142-17ac-4a1f-b83f-1ae025c49292_origin.pdf b/2023/Omni3D_ A Large Benchmark and Model for 3D Object Detection in the Wild/2efbd142-17ac-4a1f-b83f-1ae025c49292_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f602e94dedc53ca3d5129d0f43bbdf8cf7d771ce --- /dev/null +++ b/2023/Omni3D_ A Large Benchmark and Model for 3D Object Detection in the Wild/2efbd142-17ac-4a1f-b83f-1ae025c49292_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:897c23fd470c57c1a83d24b795aa5ed7b46d769cf54b67e59fcb427690628518 +size 7442538 diff --git a/2023/Omni3D_ A Large Benchmark and Model for 3D Object Detection in the Wild/full.md b/2023/Omni3D_ A Large Benchmark and Model for 3D Object Detection in the Wild/full.md new file mode 100644 index 0000000000000000000000000000000000000000..2cbf2879394a71fa28cf9c94436c917909d57169 --- /dev/null +++ b/2023/Omni3D_ A Large Benchmark and Model for 3D Object Detection in the Wild/full.md @@ -0,0 +1,361 @@ +# OMNI3D: A Large Benchmark and Model for 3D Object Detection in the Wild + +Garrick Brazil1 Abhinav Kumar2 Julian Straub1 Nikhila Ravi1 Justin Johnson1 Georgia Gkioxari3 + +$^{1}$ Meta AI $^{2}$ Michigan State University $^{3}$ Caltech + +![](images/cdc5e4b3a711848ba026a9f90d0ae7eb0318877f0b52fa7ff8fa6e6af19cd00d.jpg) +Figure 1. Left: We introduce OMNI3D, a benchmark for 3D object detection which is larger and more diverse than popular 3D benchmarks. Right: We propose Cube R-CNN, which generalizes to unseen datasets (e.g. COCO [45]) and outperforms prior works on existing datasets. + +![](images/b2e4cd0d07e95acf9efe8fa561fe11ac96e94ea29c63c3de6bdc344f252986b5.jpg) + +![](images/d7c28d109f6884ad9edc9e375242ea1b28f20535f1caa7d0f9cd691613cd76d8.jpg) + +![](images/c54b7910c8ec5a1763fc91285dbf156997eae3fc5de7482371add9b80e36c733.jpg) + +# Abstract + +Recognizing scenes and objects in 3D from a single image is a longstanding goal of computer vision with applications in robotics and AR/VR. For 2D recognition, large datasets and scalable solutions have led to unprecedented advances. In 3D, existing benchmarks are small in size and approaches specialize in few object categories and specific domains, e.g. urban driving scenes. Motivated by the success of 2D recognition, we revisit the task of 3D object detection by introducing a large benchmark, called OMNI3D. OMNI3D re-purposes and combines existing datasets resulting in 234k images annotated with more than 3 million instances and 98 categories. 3D detection at such scale is challenging due to variations in camera intrinsic and the rich diversity of scene and object types. We propose a model, called Cube R-CNN, designed to generalize across camera and scene types with a unified approach. We show that Cube R-CNN outperforms prior works on the larger OMNI3D and existing benchmarks. Finally, we prove that OMNI3D is a powerful dataset for 3D object recognition and show that it improves single-dataset performance and can accelerate learning on new smaller datasets via pre-training.1 + +# 1. Introduction + +Understanding objects and their properties from single images is a longstanding problem in computer vision with applications in robotics and AR/VR. In the last decade, 2D object recognition [26,40,63,64,72] has made tremendous advances toward predicting objects on the image plane with the help of large datasets [24,45]. However, the world and its objects are three dimensional laid out in 3D space. Perceiving objects in 3D from 2D visual inputs poses new challenges framed by the task of 3D object detection. Here, the goal is to estimate a 3D location and 3D extent of each object in an image in the form of a tight oriented 3D bounding box. + +Today 3D object detection is studied under two different lenses: for urban domains in the context of autonomous vehicles [7,13,50,52,57] or indoor scenes [31,36,58,73]. Despite the problem formulation being shared, methods share little insights between domains. Often approaches are tailored to work only for the domain in question. For instance, urban methods make assumptions about objects resting on a ground plane and model only yaw angles for 3D rotation. Indoor techniques may use a confined depth range (e.g. up to $6\mathrm{m}$ in [58]). These assumptions are generally not true in the real world. Moreover, the most popular benchmarks for image-based 3D object detection are small. Indoor SUN RGB-D [70] has $10\mathrm{k}$ images, urban KITTI [21] has $7\mathrm{k}$ images; 2D benchmarks like COCO [45] are $20\times$ larger. + +We address the absence of a general large-scale dataset for 3D object detection by introducing a large and diverse 3D benchmark called OMNI3D. OMNI3D is curated from publicly released datasets, SUN RBG-D [70], ARKitScenes [6], Hypersim [65], Objectron [2], KITTI [21] and nuScenes [9], and comprises 234k images with 3 million objects annotated with 3D boxes across 98 categories including chair, sofa, laptop, table, cup, shoes, pillow, books, car, person, etc. Sec. 3 describes the curation process which involves re-purposing the raw data and annotations from the aforementioned datasets, which originally target different applications. As shown in Fig. 1, OMNI3D is $20\times$ larger than existing popular benchmarks used for 3D detection, SUN RGB-D and KITTI. For efficient evaluation on the large OMNI3D, we introduce a new algorithm for intersection-over-union of 3D boxes which is $450\times$ faster than previous solutions [2]. We empirically prove the impact of OMNI3D as a large-scale dataset and show that it improves single-dataset performance by up to $5.3\%$ AP on urban and $3.8\%$ on indoor benchmarks. + +On the large and diverse OMNI3D, we design a general and simple 3D object detector, called Cube R-CNN, inspired by advances in 2D and 3D recognition of recent years [22,52,64,68]. Cube R-CNN detects all objects and their 3D location, size and rotation end-to-end from a single image of any domain and for many object categories. Attributed to OMNI3D's diversity, our model shows strong generalization and outperforms prior works for indoor and urban domains with one unified model, as shown in Fig. 1. Learning from such diverse data comes with challenges as OMNI3D contains images of highly varying focal lengths which exaggerate scale-depth ambiguity (Fig. 4). We remedy this by operating on virtual depth which transforms object depth with the same virtual camera intrinsics across the dataset. An added benefit of virtual depth is that it allows the use of data augmentations (e.g. image rescaling) during training, which is a critical feature for 2D detection [12,80], and as we show, also for 3D. Our approach with one unified design outperforms prior best approaches in $\mathrm{AP}_{\mathrm{3D}}$ , ImVoxelNet [66] by $4.1\%$ on indoor SUN RGB-D, GUPNet [52] by $9.5\%$ on urban KITTI, and PGD [77] by $7.9\%$ on OMNI3D. + +We summarize our contributions: + +- We introduce OMNI3D, a benchmark for image-based 3D object detection sourced from existing 3D datasets, which is $20 \times$ larger than existing 3D benchmarks. +- We implement a new algorithm for IoU of 3D boxes, which is $450 \times$ faster than prior solutions. +- We design a general-purpose baseline method, Cube R-CNN, which tackles 3D object detection for many categories and across domains with a unified approach. We propose virtual depth to eliminate the ambiguity from varying camera focal lengths in OMNI3D. + +# 2. Related Work + +Cube R-CNN and OMNI3D draw from key research advances in 2D and 3D object detection. + +2D Object Detection. Here, methods include two-stage approaches [26, 64] which predict object regions with a region proposal network (RPN) and then refine them via an MLP. Single-stage detectors [44, 47, 63, 72, 85] omit the RPN and predict regions directly from the backbone. + +3D Object Detection. Monocular 3D object detectors predict 3D cuboids from single input images. There is extensive work in the urban self-driving domain where the car class is at the epicenter [13, 19, 23, 25, 30, 46, 49, 54, 57, 62, 68, 69, 76, 78, 84]. CenterNet [85] predicts 3D depth and size from fully-convolutional center features, and is extended by [14, 37, 42, 48, 50–52, 55, 83, 84, 87]. M3D-RPN [7] trains an RPN with 3D anchors, enhanced further by [8, 18, 38, 75, 88]. FCOS3D [78] extends the anchorless FCOS [72] detector to predict 3D cuboids. Its successor PGD [77] furthers the approach with probabilistic depth uncertainty. Others use pseudo depth [4, 15, 53, 59, 79, 81] and explore depth and point-based LiDAR techniques [35, 60]. Similar to ours, [11, 68, 69] add a 3D head, specialized for urban scenes and objects, on two-stage Faster R-CNN. [25, 69] augment their training by synthetically generating depth and box-fitted views, coined as virtual views or depth. In our work, virtual depth aims at addressing varying focal lengths. + +For indoor scenes, a vast line of work tackles room layout estimation [17,28,41,56]. Huang et al [31] predict 3D oriented bounding boxes for indoor objects. Factored3D [73] and 3D-RelNet [36] jointly predict object voxel shapes. Total3D [58] predicts 3D boxes and meshes by additionally training on datasets with annotated 3D shapes. ImVoxelNet [66] proposes domain-specific methods which share an underlying framework for processing volumes of 3D voxels. In contrast, we explore 3D object detection in its general form by tackling outdoor and indoor domains jointly in a single model and with a vocabulary of $5 \times$ more categories. + +3D Datasets. KITTI [21] and SUN RGB-D [70] are popular datasets for 3D object detection on urban and indoor scenes respectively. Since 2019, 3D datasets have emerged, both for indoor [2, 3, 6, 16, 65] and outdoor [9, 10, 20, 29, 32, 33, 71]. In isolation, these datasets target different tasks and applications and have unique properties and biases, e.g. object and scene types, focal length, coordinate systems, etc. In this work, we unify existing representative datasets [2, 6, 9, 21, 65, 70]. We process the raw visual data, re-purpose their annotations, and carefully curate the union of their semantic labels in order to build a coherent large-scale benchmark, called OMNI3D. OMNI3D is $20 \times$ larger than widely-used benchmarks and notably more diverse. As such, new challenges arise stemming from the increased variance in visual domain, object rotation, size, layouts, and camera intrinsics. + +![](images/789db86a26df8ee5fbd79eebf2d7ad9f27b912998be6b082816bdf1e6913a452.jpg) +(a) Spatial statistics + +![](images/4063bb12980d1e73898db058fa5b56c72a5b320cca0603175613ba5e926cfce7.jpg) +(b) 2D Scale statistics +Figure 2. OMNI3D analysis. (a) distribution of object centers on (top) normalized image, XY-plane, and (bottom) normalized depth, XZ-plane, (b) relative 2D object scales, and (c) category frequency. + +![](images/0e975865c5d381f2f1f79180d494a13640ed937d56f993f333c2d1dd2d6c11f5.jpg) +(c) Category statistics + +# 3. The OMNI3D Benchmark + +The primary benchmarks for 3D object detection are small, focus on a few categories and are of a single domain. For instance, the popular KITTI [21] contains only urban scenes, has 7k images and 3 categories, with a focus on car. SUN RGB-D [70] has 10k images. The small size and lack of variance in 3D datasets is a stark difference to 2D counterparts, such as COCO [45] and LVIS [24], which have pioneered progress in 2D recognition. + +We aim to bridge the gap to 2D by introducing OMNI3D, a large-scale and diverse benchmark for image-based 3D object detection consisting of $234\mathrm{k}$ images, 3 million labeled 3D bounding boxes, and 98 object categories. We source from recently released 3D datasets of urban (nuScenes [9] and KITTI [21]), indoor (SUN RGB-D [70], ARKitScenes [6] and Hypersim [65]), and general (Objectron [2]) scenes. Each of these datasets target different applications (e.g. point-cloud recognition or reconstruction in [6], inverse rendering in [65]), provide visual data in different forms (e.g. videos in [2, 6], rig captures in [9]) and annotate different object types. To build a coherent benchmark, we process the varying raw visual data, re-purpose their annotations to extract 3D cuboids in a unified 3D camera coordinate system, and carefully curate the final vocabulary. More details about the benchmark creation in the Appendix. + +We analyze OMNI3D and show its rich spatial and semantic properties proving it is visually diverse, similar to 2D data, and highly challenging for 3D as depicted in Fig. 2. We show the value of OMNI3D for the task of 3D object detection with extensive quantitative analysis in Sec. 5. + +# 3.1. Dataset Analysis + +Splits. We split the dataset into $175\mathrm{k} / 19\mathrm{k} / 39\mathrm{k}$ images for train/val/test respectively, consistent with original splits when available, and otherwise free of overlapping video sequences in splits. We denote indoor and outdoor subsets as $\mathrm{OMN13D_{IN}}$ (SUN RGB-D, Hypersim, ARKit), and $\mathrm{OMN13D_{OUT}}$ (KITTI, nuScenes). Objectron, with primarily close-up objects, is used only in the full OMN13D setting. + +Layout statistics. Fig. 2(a) shows the distribution of object centers onto the image plane by projecting centroids on the XY-plane (top row), and the distribution of object depths by projecting centroids onto the XZ-plane (bottom row). We find that OMNI3D's spatial distribution has a center bias, similar to 2D datasets COCO and LVIS. Fig. 2(b) depicts the relative object size distribution, defined as the square root of object area divided by image area. Objects are more likely to be small in size similar to LVIS (Fig. 6c in [24]) suggesting that OMNI3D is also challenging for 2D detection, while objects in OMNI3DOUT are noticeably smaller. + +The bottom row of Fig. 2(a) normalizes object depth in a $[0,20\mathrm{m}]$ range, chosen for visualization and satisfies $88\%$ of object instances; OMNI3D depth ranges as far as $300\mathrm{m}$ . We observe that the OMNI3D depth distribution is far more diverse than SUN RGB-D and KITTI, which are biased toward near or road-side objects respectively. See Appendix for each data source distribution plot. Fig. 2(a) demonstrates OMNI3D's rich diversity in spatial distribution and depth which suffers significantly less bias than existing 3D benchmarks and is comparable in complexity to 2D datasets. + +2D and 3D correlation. A common assumption in urban scenes is that objects rest on a ground plane and appear smaller with depth. To verify if that is true generally, we compute correlations. We find that $2\mathrm{D}y$ and $3\mathrm{D}z$ are indeed fairly correlated in $\mathrm{OMNI3D_{OUT}}$ at 0.524, but significantly less in $\mathrm{OMNI3D_{IN}}$ at 0.006. Similarly, relative 2D object size $(\sqrt{h\cdot w})$ and $z$ correlation is 0.543 and 0.102 respectively. This confirms our claim that common assumptions in urban scenes are not generally true, making the task challenging. + +Category statistics. Fig. 2(c) plots the distribution of instances across the 98 categories of OMNI3D. The long-tail suggests that low-shot recognition in both 2D and 3D will be critical for performance. In this work, we want to focus on large-scale and diverse 3D recognition, which is comparably unexplored. We therefore filter and focus on categories with at least 1000 instances. This leaves 50 diverse categories including chair, sofa, laptop, table, books, car, truck, pedestrian and more, 19 of which have more than 10k instances. We provide more per-category details in the Appendix. + +![](images/48250649f1de3140a3a8dbcf9325d3359327e7a602b7399a3a0aa1aabbe7005f.jpg) +Figure 3. Overview. Cube R-CNN takes as input an RGB image, detects all objects in 2D and predicts their 3D cuboids $B_{3\mathrm{D}}$ . During training, the 3D corners of the cuboids are compared against 3D ground truth with the point-cloud chamfer distance. + +# 4. Method + +Our goal is to design a simple and effective model for general 3D object detection. Hence, our approach is free of domain or object specific strategies. We design our 3D object detection framework by extending Faster R-CNN [64] with a 3D object head which predicts a cuboid per each detected 2D object. We refer to our method as Cube R-CNN. Figure 3 shows an overview of our approach. + +# 4.1. Cube R-CNN + +Our model builds on Faster R-CNN [64], an end-to-end region-based object detection framework. Faster-RCNN consists of a backbone network, commonly a CNN, which embeds the input image into a higher-dimensional feature space. A region proposal network (RPN) predicts regions of interest (RoIs) representing object candidates in the image. A 2D box head inputs the backbone feature map and processes each RoI to predict a category and a more accurate 2D box. Faster R-CNN can be easily extended to tackle more tasks by adding task-specific heads, e.g. Mask R-CNN [26] adds a mask head to additionally output object silhouettes. + +For the task of 3D object detection, we extend Faster R-CNN by introducing a 3D detection head which predicts a 3D cuboid for each detected 2D object. Cube R-CNN extends Faster R-CNN in three ways: (a) we replace the binary classifier in RPN which predicts region objectness with a regressor that predicts IoUness, (b) we introduce a cube head which estimates the parameters to define a 3D cuboid for each detected object (similar in concept to [68]), and (c) we define a new training objective which incorporates a virtual depth for the task of 3D object detection. + +IoUness. The role of RPN is two-fold: (a) it proposes RoIs by regressing 2D box coordinates from pre-computed anchors and (b) it classifies regions as object or not (object- + +ness). This is sensible in exhaustively labeled datasets where it can be reliably assessed if a region contains an object. However, OMNI3D combines many data sources with no guarantee that all instances of all classes are labeled. We overcome this by replacing objectness with IoUness, applied only to foreground. Similar to [34], a regressor predicts IoU between a RoI and a ground truth. Let $o$ be the predicted IoU for a RoI and $\hat{o}$ be the 2D IoU between the region and its ground truth; we apply a binary cross-entropy (CE) loss $\mathcal{L}_{\mathrm{IoUness}} = \ell_{\mathrm{CE}}(o,\hat{o})$ . We train on regions whose IoU exceeds 0.05 with a ground truth in order to learn IoUness from a wide range of region overlaps. Thus, the RPN training objective becomes $\mathcal{L}_{\mathrm{RPN}} = \hat{o}\cdot (\mathcal{L}_{\mathrm{IoUness}} + \mathcal{L}_{\mathrm{reg}})$ , where $\mathcal{L}_{\mathrm{reg}}$ is the 2D box regression loss from [64]. The loss is weighted by $\hat{o}$ to prioritize candidates close to true objects. + +Cube Head. We extend Faster R-CNN with a new head, called cube head, to predict a 3D cuboid for each detected 2D object. The cube head inputs $7 \times 7$ feature maps pooled from the backbone for each predicted region and feeds them to 2 fully-connected (FC) layers with 1024 hidden dimensions. All 3D estimations in the cube head are category-specific. The cube head represents a 3D cuboid with 13 parameters each predicted by a final FC layer: + +- $[u, v]$ represent the projected 3D center on the image plane relative to the 2D RoI +- $z \in \mathbb{R}_+$ is the object's center depth in meters transformed from virtual depth $z_v$ (explained below) +- $[ \bar{w}, \bar{h}, \bar{l} ] \in \mathbb{R}_{+}^{3}$ are the log-normalized physical box dimensions in meters +- $\mathbf{p} \in \mathbb{R}^6$ is the continuous 6D [86] allocentric rotation +- $\mu \in \mathbb{R}_+$ is the predicted 3D uncertainty + +The above parameters form the final 3D box in camera view coordinates for each detected 2D object. The object's 3D center $\mathbf{X}$ is estimated from the predicted 2D projected + +center $[u,v]$ and depth $z$ via + +$$ +\mathbf {X} (u, v, z) = \left(\frac {z}{f _ {x}} \left(r _ {x} + u r _ {w} - p _ {x}\right), \frac {z}{f _ {y}} \left(r _ {y} + v r _ {h} - p _ {y}\right), z\right) \tag {1} +$$ + +where $[r_x, r_y, r_w, r_h]$ is the object's 2D box, $(f_x, f_y)$ are the camera's known focal lengths and $(p_x, p_y)$ the principal point. The 3D box dimensions $\mathbf{d}$ are derived from $[\bar{w}, \bar{h}, \bar{l}]$ which are log-normalized with category-specific precomputed means $(w_0, h_0, l_0)$ for width, height and length respectively, and are arranged into a diagonal matrix via + +$$ +\mathbf {d} (\bar {w}, \bar {h}, \bar {l}) = \operatorname {d i a g} \left(\exp (\bar {w}) w _ {0}, \exp (\bar {h}) h _ {0}, \exp (\bar {l}) l _ {0}\right) \tag {2} +$$ + +Finally, we derive the object's pose $\mathbf{R}(\mathbf{p})$ as a $3 \times 3$ rotation matrix based on a 6D parameterization (2 directional vectors) of $\mathbf{p}$ following [86] which is converted from allocentric to egocentric rotation similar to [39], defined formally in Appendix. The final 3D cuboid, defined by 8 corners, is + +$$ +B _ {\mathrm {3 D}} (u, v, z, \bar {w}, \bar {h}, \bar {l}, \mathbf {p}) = \mathbf {R} (\mathbf {p}) \mathbf {d} (\bar {w}, \bar {h}, \bar {l}) B _ {\text {u n i t}} + \mathbf {X} (u, v, z) \tag {3} +$$ + +where $B_{\mathrm{unit}}$ are the 8 corners of an axis-aligned unit cube centered at $(0,0,0)$ . Lastly, $\mu$ denotes a learned 3D uncertainty, which is mapped to a confidence at inference then joined with the classification score $s$ from the 2D box head to form the final score for the prediction, as $\sqrt{s \cdot \exp(-\mu)}$ . + +Training objective. Our training objective consists of 2D losses from the RPN and 2D box head and 3D losses from the cube head. The 3D objective compares each predicted 3D cuboid with its matched ground truth via a chamfer loss, treating the 8 box corners of the 3D boxes as point clouds, namely $\mathcal{L}_{\mathrm{3D}}^{\mathrm{all}} = \ell_{\mathrm{chamfer}}(B_{\mathrm{3D}}, B_{\mathrm{3D}}^{\mathrm{gt}})$ . Note that $\mathcal{L}_{\mathrm{3D}}^{\mathrm{all}}$ entangles all 3D variables via the box predictor $B_{\mathrm{3D}}$ (Eq. 3), such that that errors in variables may be ambiguous from one another. Thus, we isolate each variable group with separate disentangled losses, following [68]. The disentangled loss for each variable group substitutes all but its variables with the ground truth from Eq. 3 to create a pseudo box prediction. For example, the disentangled loss for the projected center $[u, v]$ produces a 3D box with all but $(u, v)$ replaced with the true values and then compares to the ground truth box, + +$$ +\mathcal {L} _ {\mathrm {3 D}} ^ {(u, v)} = \left\| B _ {\mathrm {3 D}} (u, v, z ^ {\mathrm {g t}}, \bar {w} ^ {\mathrm {g t}}, \bar {h} ^ {\mathrm {g t}}, \bar {l} ^ {\mathrm {g t}}, \mathbf {p} ^ {\mathrm {g t}}) - B _ {\mathrm {3 D}} ^ {\mathrm {g t}} \right\| _ {1} \tag {4} +$$ + +We use an $L_{1}$ loss for $\mathcal{L}_{\mathrm{3D}}^{(u,v)}$ , $\mathcal{L}_{\mathrm{3D}}^{(z)}$ and $\mathcal{L}_{\mathrm{3D}}^{(\bar{w},\bar{h},\bar{l})}$ and chamfer for $\mathcal{L}_{\mathrm{3D}}^{\mathbf{P}}$ to account for cuboid symmetry such that rotation matrices of Euler angles modulo $\pi$ produce the same noncanonical 3D cuboid. Losses in box coordinates have natural advantages over losses which directly compare variables to ground truths in that their gradients are appropriately weighted by the error as shown in [68]. The 3D objective is defined as, $\mathcal{L}_{\mathrm{3D}} = \mathcal{L}_{\mathrm{3D}}^{(u,v)} + \mathcal{L}_{\mathrm{3D}}^{(z)} + \mathcal{L}_{\mathrm{3D}}^{(\bar{w},\bar{h},\bar{l})} + \mathcal{L}_{\mathrm{3D}}^{\mathbf{P}} + \mathcal{L}_{\mathrm{3D}}^{\mathrm{all}}$ . + +The collective training objective of Cube R-CNN is, + +$$ +\mathcal {L} = \mathcal {L} _ {\mathrm {R P N}} + \mathcal {L} _ {\mathrm {2 D}} + \sqrt {2} \cdot \exp (- \mu) \cdot \mathcal {L} _ {\mathrm {3 D}} + \mu \tag {5} +$$ + +![](images/b994ceceef479499166490ea0d206666179a0b48729edfa20c1aaf24e501295d.jpg) +Figure 4. Varying camera intrinsics exaggerate scale-depth ambiguity – the same object at two different depths can project to the same image, as in (a) and (b). We address this by introducing a virtual camera which is invariant to intrinsics and transforms the object's depth to a virtual depth $z_v$ such that the effective image size $H_v$ and focal length $f_v$ are consistent across images. + +![](images/2078c047e41eabfaf45470df4b46e07f69028df1b006123e33e713c23eba49d7.jpg) + +$\mathcal{L}_{\mathrm{RPN}}$ is the RPN loss, described above, and $\mathcal{L}_{2\mathrm{D}}$ is the 2D box head loss from [64]. The 3D loss is weighted by the predicted 3D uncertainty (inspired by [52]), such that the model may trade a penalty to reduce the 3D loss when uncertain. In practice, $\mu$ is helpful for both improving the 3D ratings at inference and reducing the loss of hard samples in training. + +# 4.2. Virtual Depth + +A critical part of 3D object detection is predicting an object's depth in metric units. Estimating depth from visual cues requires an implicit mapping of 2D pixels to 3D distances, which is more ambiguous if camera intrinsics vary. Prior works are able to ignore this as they primarily train on images from a single sensor. OMNI3D contains images from many sensors and thus demonstrates large variations in camera intrinsics. We design Cube R-CNN to be robust to intrinsics by predicting the object's virtual depth $z_{v}$ , which projects the metric depth $z$ to an invariant camera space. + +Virtual depth scales the depth using the (known) camera intrinsics such that the effective image size and focal length are consistent across the dataset, illustrated in Fig. 4. The effective image properties are referred to as virtual image height $H_{v}$ and virtual focal length $f_{v}$ , and are both hyperparameters. If $z$ is the true metric depth of an object from an image with height $H$ and focal length $f$ , its virtual depth is defined as $z_{v} = z \cdot \frac{f_{v}}{f} \frac{H}{H_{v}}$ . The derivation is in the Appendix. + +Invariance to camera intrinsics via virtual depth also enables scale augmentations during training, since $H$ can vary without harming the consistency of $z_v$ . Data augmentations from image resizing tend to be critical for 2D models but are not often used in 3D methods [7, 50, 68] since if unaccounted for they increase scale-depth ambiguity. Virtual depth lessens such ambiguities and therefore enables powerful data augmentations in training. We empirically prove the two-fold effects of virtual depth in Sec. 5 + +![](images/a1a2d7a345fba5423768e23b6e43263027ad5e429ba8ca2ecb6e0bd1707e2621.jpg) + +![](images/a2dc74297760f80c0033139fd8f2a66ca1c5c049a5b35661ec6607549405f342.jpg) + +![](images/fe6baf6159df0942d2b2c58dc92dcab418722b3568ee670c0348462699d660e8.jpg) + +![](images/63865173e449343cb6a235b186fef61c7eb2831b82800b145e648ab7559c8b4a.jpg) + +![](images/9b2f18ec20a52ec69421e2c4747a97fa3ff7ac1e6c0e44d907e19ce1840c4000.jpg) + +![](images/994acec229be8a238f2e6ae3de3084ad3eff1edf9f1e7dc9153bdae35573ce04.jpg) + +![](images/1fe5ddef149919d8c5b2512ad32fd572416c68a60e7c3ed18fe48a020a463fdf.jpg) + +![](images/16cf37b188a3a95b8418357b413f7e0c4a64ce5f07b4c152025a192001691b48.jpg) + +![](images/7e97d86431114eced98487b0a9f167fd210658c5acd71e0788da9343891eceef.jpg) + +![](images/ed95d112221cbf933ef94b625c3e9c71ae08b6c3a7e52ddfe2ab54a65a4d3904.jpg) + +![](images/c5864368434eca5af31628f837fc8568b705e61a49f0d171a647acd10589eced.jpg) + +![](images/eb5549ccd6cbfcf2fdf7358d68b028373f4e940a742e804d7fc6f9e956e76eb7.jpg) + +![](images/62880ba4b80097575de00d90c051d8a0059a2d7951a901860f40d392abde6529.jpg) +Figure 5. Cube R-CNN predictions on OMNI3D test. For each example, we show the input image, the 3D predictions overlaid on the image and a top view with a base composed of $1\mathrm{m}\times 1\mathrm{m}$ tiles. See more results from images and videos in the Appendix. + +![](images/da91209b692c3de536abce28f727b2b9b52bea4afcb0c168baf45d40b4e3e456.jpg) + +![](images/5c3715f40fd44e583e3284afb2c8c619bd2bda3be2e27605386f9827378e1468.jpg) + +![](images/ad691737e8c087b90e5339938ce38bf914cb8e5308914dd316de4da9443ff3e7.jpg) + +![](images/78af7f71cfa0eb0c9f119b807ebafa86676166b3cf2cc4f2558e7274ebf03fb9.jpg) + +![](images/884996be08dfa5c2cedafd2a7fc444c85bbcb00f1e08495b4d65bf40eab8f572.jpg) + +# 5. Experiments + +We tackle image-based 3D object detection on OMNI3D and compare to prior best methods. We evaluate Cube R-CNN on existing 3D benchmarks with a single unified model. Finally, we prove the effectiveness of OMNI3D as a large-scale 3D object detection dataset by showing comprehensive cross-dataset generalization and its impact for pre-training. + +Implementation details. We implement Cube R-CNN using Detector2 [80] and PyTorch3D [61]. Our backbone is DLA34 [82]-FPN [43] pretrained on ImageNet [67]. We train for 128 epochs with a batch size of 192 images across 48 V100s. We use SGD with a learning rate of 0.12 which decays by a factor of 10 after $60\%$ and $80\%$ of training. During training, we use random data augmentation of horizontal flipping and scales $\in [0.50, 1.25]$ , enabled by virtual depth. Virtual camera parameters are set to $f_v = H_v = 512$ . Source code and models are publicly available. + +Metric. Following popular benchmarks for 2D and 3D recognition, we use average-precision (AP) as our 3D metric. Predictions are matched to ground truth by measuring their overlap using $\mathsf{I}\circ \mathsf{U}_{\mathsf{3D}}$ which computes the intersection-overunion of 3D cuboids. We compute a mean $\mathrm{AP}_{\mathrm{3D}}$ across all 50 categories in OMNI3D and over a range of $\mathsf{I}\circ \mathsf{U}_{\mathsf{3D}}$ thresholds $\tau \in [0.05,0.10,\dots ,0.50]$ . The range of $\tau$ is more relaxed than in 2D to account for the new dimension - see the Appendix for a comparison between 3D and 2D ranges for IoU. General 3D objects may be occluded by other objects or truncated by the image border, and can have arbitrarily small projections. Following [21], we ignore objects with high occlusion ( $>66\%$ ) or truncation ( $>33\%$ ), and tiny projected objects ( $< 6.25\%$ image height). Moreover, we also report $\mathrm{AP}_{\mathrm{3D}}$ at varying levels of depth $d$ as near: $0 < d \leq 10m$ , medium: $10m < d \leq 35m$ , far: $35m < d \leq \infty$ . + +Fast $\mathsf{IOU}_{\mathbf{3D}}$ . $\mathsf{IOU}_{\mathbf{3D}}$ compares two 3D cuboids by computing their intersection-over-union. Images usually have many objects and produce several predictions so $\mathsf{IOU}_{\mathbf{3D}}$ computations need to be fast. Prior implementations [21] approximate $\mathsf{IOU}_{\mathbf{3D}}$ by projecting 3D boxes on a ground plane and multiply the top-view 2D intersection with the box heights to compute a 3D volume. Such approximations become notably inaccurate when objects are not on a planar ground or have an arbitrary orientation (e.g. with nonzero pitch or roll). Objectron [2] provides an exact solution, but relies on external libraries [5, 74] implemented in $\mathbf{C}++$ and is not batched. We implement a new, fast and exact algorithm which computes the intersecting shape by representing cuboids as meshes and finds face intersections – described in Appendix. Our algorithm is batched with $\mathbf{C}++$ and CUDA support. Our implementation is $90\times$ faster than Objectron in $\mathbf{C}++$ , and $450\times$ faster in CUDA. As a result, evaluation on the large OMNI3D takes only seconds instead of several hours. + +# 5.1. Model Performance + +First, we ablate the design choices of Cube R-CNN. Then, we compare to state-of-the-art methods on existing benchmarks and OMNI3D and show that it performs superior or on par with prior works which are specialized for their respective domains. Our Cube R-CNN uses a single unified design to tackle general 3D object detection across domains. Fig. 5 shows qualitative predictions on the OMNI3D test set. + +Ablations. Table 1 ablates the features of Cube R-CNN on OMNI3D. We report $\mathrm{AP}_{3\mathrm{D}}$ performance at single IoU thresholds (0.25 and 0.50), and for near, medium and far objects. We further train and evaluate on domain-specific subsets, OMNI3DIN $(\mathrm{AP}_{3\mathrm{D}}^{\mathrm{IN}})$ and OMNI3DOUT $(\mathrm{AP}_{3\mathrm{D}}^{\mathrm{OUT}})$ to show how ablation trends hold for single domain scenarios. From Table 1 we draw a few standout observations: + +
Cube R-CNNAP3DAP253DAP503DAPnear3DAPmed3DAPfar3DAPIN3DAPOUT3D
w/o disentangled22.624.39.226.811.88.214.530.9
w/o IoUness22.223.68.826.411.18.314.331.0
w/o L3D20.221.86.426.412.17.413.829.1
w/o scale aug.20.221.58.023.59.86.812.328.1
w/o scale | virtual19.821.27.523.48.65.712.226.0
w/o virtual depth17.318.44.422.77.97.113.230.3
w/o uncertainty μ17.218.35.420.510.87.09.125.4
ours23.324.99.527.912.18.515.031.9
+ +Table 1. Cube R-CNN ablations. We report $\mathrm{AP}_{3\mathrm{D}}$ on OMNI3D, then at IoU thresholds 0.25 and 0.50, and for near, medium and far objects. We further report $\mathrm{AP}_{3\mathrm{D}}$ when both training and evaluating on OMNI3DIN and OMNI3DOUT subsets. + +
MethodfOMNI3DOUTOMNI3D
AP3DAPNU3DAPOUT3DAP3D
M3D-RPN [7]X10.417.913.7-
SMOKE [50]X25.420.419.59.6
FCOS3D [78]X14.620.917.69.8
PGD [77]X21.426.322.911.2
GUPNet [52]✓24.520.519.9-
M3D-RPN +vc✓16.2+5.820.4+2.917.0+3.3-
ImVoxelNet [66]✓23.523.421.59.4
SMOKE +vc✓25.9+0.520.4+0.020.0+0.510.4+0.8
FCOS3D +vc✓17.8+3.225.1+4.221.3+3.710.6+0.8
PGD +vc✓22.8+1.431.2+4.926.8+3.915.4+4.2
Cube R-CNN✓36.032.731.923.3
+ +Table 2. OMNI3D comparison with GUPNet [52], ImVoxelNet [66], M3D-RPN [7], FCOS3D [78], PGD [77], SMOKE [50]. We apply virtual camera to the last four $(+vc)$ and show the gains in purple. $f$ denotes if a method handles variance in focal length. + +Virtual depth is effective and improves $\mathrm{AP}_{3\mathrm{D}}$ by $+6\%$ , most noticeable in the full OMNI3D which has the largest variances in camera intrinsics. Enabled by virtual depth, scale augmentations during training increase $\mathrm{AP}_{3\mathrm{D}}$ by $+3.1\%$ on OMNI3D, $+2.7\%$ on OMNI3D $_{\text{IN}}$ and $+3.8\%$ on OMNI3D $_{\text{OUT}}$ . We find scale augmentation controlled without virtual depth to be harmful by $-2.5\%$ (rows 6 - 7). + +3D uncertainty boosts performance by about $+6\%$ on all OMNI3D subsets. Intuitively, it serves as a measure of confidence for the model's 3D predictions; removing it means that the model relies only on its 2D classification to score cuboids. If uncertainty is used only to scale samples in Eq. 5, but not at inference then $\mathrm{AP}_{3\mathrm{D}}$ still improves but by $+1.6\%$ . + +Table 1 also shows that the entangled $\mathcal{L}_{3\mathrm{D}}^{\mathrm{all}}$ loss from Eq. 5 boosts $\mathrm{AP}_{3\mathrm{D}}$ by $+3.1\%$ while disentangled losses contribute less $(+0.7\%)$ . Replacing objectness with IoUness in the RPN head improves $\mathrm{AP}_{3\mathrm{D}}$ on OMNI3D by $+1.1\%$ . + +Comparison to other methods. We compare Cube R-CNN to prior best approaches. We choose representative state-of-the-art methods designed for the urban and indoor domains, and evaluate on our proposed OMNI3D benchmark and single-dataset benchmarks, KITTI and SUN RGB-D. + +Comparisons on OMNI3D. Table 2 compares Cube R-CNN to M3D-RPN [7], a single-shot 3D anchor approach, FCOS3D [78] and its follow-up PGD [77], which use a fully convolutional one-stage model, GUPNet [52], which uses + +
MethodDAP703DAP70BEV
EasyMedHardEasyMedHard
SMOKE [50]X14.039.767.8420.8314.4912.75
ImVoxelNet [66]X17.1510.979.1525.1916.3713.58
PGD [77]X19.0511.769.3926.8916.5113.49
GUPNet [52]X22.2615.0213.1230.2921.1918.20
Cube R-CNNX23.5915.0112.5631.7021.2018.43
MonoDTR [30]✓21.9915.3912.7328.5920.3817.14
DD3D [59]✓23.1916.8714.3632.3523.4120.42
+ +Table 3. KITTI leaderboard results on car for image-based approaches. D denotes methods trained with extra depth supervision. + +2D-3D geometry and camera focal length to derive depth, SMOKE [50] which predicts 3D object center and offsets densely and ImVoxelNet [66], which uses intrinsics to unproject 2D image features to a 3D volume followed by 3D convolutions. These methods originally experiment on urban domains, so we compare on OMNI3DOUT and the full OMNI3D. We train each method using public code and report the best of multiple runs after tuning their hyper-parameters to ensure best performance. We extend all but GUPNet and ImVoxelNet with a virtual camera to handle varying intrinsics, denoted by $+vc$ . GUPNet and ImVoxelNet dynamically use per-image intrinsics, which can naturally account for camera variation. M3D-RPN and GUPNet do not support multi-node training, which makes scaling to OMNI3D difficult and are omitted. For OMNI3DOUT, we report $\mathrm{AP}_{3\mathrm{D}}^{\mathrm{OUT}}$ on its test set and its subparts, $\mathrm{AP}_{3\mathrm{D}}^{\mathrm{KIT}}$ and $\mathrm{AP}_{3\mathrm{D}}^{\mathrm{NU}}$ . + +From Table 2 we observe that our Cube R-CNN outperforms competitive methods on both $\mathrm{OMNI3D_{OUT}}$ and OMNI3D. The virtual camera extensions result in performance gains of varying degree for all approaches, e.g. $+4.2\%$ for PGD on OMNI3D, proving its impact as a general purpose feature for mixed dataset training for the task of 3D object detection. FCOS3D, SMOKE, and ImVoxelNet [66] struggle the most on OMNI3D, which is not surprising as they are typically tailored for dataset-specific model choices. + +Comparisons on KITTI. Table 3 shows results on KITTI's test set using their server [1]. We train Cube R-CNN on KITTI only. Cube R-CNN is a general purpose 3D object detector; its design is not tuned for the KITTI benchmark. Yet, it performs better or on par with recent best methods which are heavily tailored for KITTI and is only slightly worse to models trained with extra depth supervision. + +
MethodTrained onI0U3DAP3D
Total3D [58]SUN RGB-D23.3-
ImVoxelNet [66]SUN RGB-D-30.6
Cube R-CNNSUN RGB-D36.234.7
Cube R-CNNOMNI3DIN37.835.4
+ +Table 4. We compare to indoor models with common categories. For Total3D, we use oracle 2D detections to report $\mathsf{I}\circ \mathsf{U}_{\mathsf{3D}}$ ( $3^{\mathrm{rd}}$ col). For ImVoxelNet, we report the detection metric of $\mathrm{AP}_{\mathrm{3D}}$ ( $4^{\mathrm{th}}$ col). + +KITTI's $\mathrm{AP}_{3\mathrm{D}}$ is fundamentally similar to ours but sets a single $\mathsf{I}\circ \mathsf{U}_{3\mathrm{D}}$ threshold at 0.70, which behaves similar to a nearly perfect 2D threshold of 0.94 (see the Appendix). To remedy this, our $\mathrm{AP}_{3\mathrm{D}}$ is a mean over many thresholds, inspired by COCO [45]. In contrast, nuScenes proposes a mean AP based on 3D center distance but omits size and rotation, with settings generally tailored to the urban domain (e.g. size and rotation variations are less extreme for cars). We provide more analysis with this metric in the Appendix. + +Comparisons on SUN RGB-D. Table 4 compares Cube R-CNN to Total3D [58] and ImVoxelNet [66], two state-of-the-art methods on SUN RGB-D on 10 common categories. Total3D's public model requires 2D object boxes and categories as input. Hence, for a fair comparison, we use ground truth 2D detections as input to both our and their method, each trained using a ResNet34 [27] backbone and report mean IoU3D (3rd col). We compare to ImVoxelNet in the full 3D object detection setting and report AP3D (4th col). We compare Cube R-CNN from two respects, first trained on SUN RGB-D identical to the baselines, then trained on OMNI3DIN which subsumes SUN RGB-D. Our model outperforms Total3D by +12.9% and ImVoxelNet by +4.1% when trained on the same training set and increases the performance gap when trained on the larger OMNI3DIN. + +Zero-shot Performance. One commonly sought after property of detectors is zero-shot generalization to other datasets. Cube R-CNN outperforms all other methods at zero-shot. We find our model trained on KITTI achieves $12.7\%$ $\mathrm{AP}_{3\mathrm{D}}$ on nuScenes; the second best is M3D-RPN+vc with $10.7\%$ . Conversely, when training on nuScenes, our model achieves $20.2\%$ on KITTI; the second best is GUPNet with $17.3\%$ . + +# 5.2. The Impact of the OMNI3D Benchmark + +Sec. 5.1 analyzes the performance of Cube R-CNN for the task of 3D object detection. Now, we turn to OMNI3D and its impact as a large-scale benchmark. We show two use cases of OMNI3D: (a) a universal 3D dataset which integrates smaller ones, and (b) a pre-training dataset. + +OMNI3D as a universal dataset. We treat OMNI3D as a dataset which integrates smaller single ones and show its impact on each one. We train Cube R-CNN on OMNI3D and compare to single-dataset training in Table 5, for the indoor (left) and urban (right) domain. $\mathrm{AP}_{\mathrm{3D}}$ is reported on the category intersection (10 for indoor, 3 for outdoor) to ensure comparability. Training on OMNI3D and its domain-specific + +
Trained onAPH3DAPSUN3DAPAR3DTrained onAPKIT3DAPNU3D
Hypersim15.29.57.5KITTI37.112.7
SUN5.834.713.1nuScenes20.238.6
ARKit5.914.238.6OMNI3D37.835.8
OMNI3D19.032.638.2OMNI3DOUT42.439.0
OMNI3DIN17.835.441.2
+ +Table 5. Cross-dataset performance on intersecting categories (10 for indoor, 3 for outdoor) for comparable cross-evaluations. + +![](images/8fe75a37d19012db8085bce4fd4c08e79478b6133bc6adfa3132a832bcfe0cc7.jpg) +Figure 6. Pre-training on OMNI3D- vs. ImageNet. + +subsets, $\mathrm{OMN13D_{IN}}$ and $\mathrm{OMN13D_{OUT}}$ , results in higher performance compared to single-dataset training, signifying that our large OMN13D generalizes better and should be preferred over single dataset training. ARKit sees a $+2.6\%$ boost and KITTI $+5.3\%$ . Except for Hypersim, the domain-specific subsets tend to perform better on their domain, which is not surprising given their distinct properties (Fig. 2). + +OMNI3D as a pre-training dataset. Next, we demonstrate the utility of OMNI3D for pre-training. In this setting, an unseen dataset is used for finetuning from a model pre-trained on OMNI3D. The motivation is to determine how a large-scale 3D dataset could accelerate low-shot learning with minimum need for costly 3D annotations on a new dataset. We choose SUN RGB-D and KITTI as our unseen datasets given their popularity and small size. We pre-train Cube R-CNN on OMNI3D-, which removes them from OMNI3D, then finetune the models using a $\%$ of their training data. The curves in Fig. 6 show the model quickly gains its upper-bound performance at a small fraction of the training data, when pre-trained on OMNI3D- vs. ImageNet, without any few-shot training tricks. A model finetuned on only $5\%$ of its target can achieve $>70\%$ of the upper-bound performance. + +# 6. Conclusion + +We propose a large and diverse 3D object detection benchmark, OMNI3D, and a general purpose 3D object detector, Cube R-CNN. Models and data are publicly released. Our extensive analysis (Table 1, Fig. 5, Appendix) show the strength of general 3D object detection as well as the limitations of our method, e.g. localizing far away objects, and uncommon object types or contexts. Moreover, for accurate real-world 3D predictions, our method is limited by the assumption of known camera intrinsics, which may be minimized in future work using self-calibration techniques. + +# References + +[1] KITTI server. https://www.cvlibs.net/datasets/kitti/eval_object.php?obj_benchmark=3d. Accessed: 2022-11-01. 7 +[2] Adel Ahmadyan, Liangkai Zhang, Artsiom Ablavatski, Jianing Wei, and Matthias Grundmann. Objectron: A large scale dataset of object-centric videos in the wild with pose annotations. CVPR, 2021. 2, 3, 6 +[3] Armen Avetisyan, Manuel Dahnert, Angela Dai, Manolis Savva, Angel X Chang, and Matthias Nießner. Scan2CAD: Learning CAD model alignment in RGB-D scans. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 2614-2623, 2019. 2 +[4] Wentao Bao, Bin Xu, and Zhenzhong Chen. MonoFENet: Monocular 3D object detection with feature enhancement networks. TIP, 2019. 2 +[5] Bradford Barber, David Dobkin, and Hannu Huhdanpaa. The quickhull algorithm for convex hulls. ACM Transactions on Mathematical Software (TOMS), 1996. 6 +[6] Gilad Baruch, Zhuoyuan Chen, Afshin Dehghan, Tal Dimry, Yuri Feigin, Peter Fu, Thomas Gebauer, Brandon Joffe, Daniel Kurz, Arik Schwartz, and Elad Shulman. ARKitScenes - a diverse real-world dataset for 3D indoor scene understanding using mobile RGB-D data. In NeurIPS Datasets and Benchmarks Track, 2021. 2, 3 +[7] Garrick Brazil and Xiaoming Liu. M3D-RPN: Monocular 3D region proposal network for object detection. In ICCV, 2019. 1, 2, 5, 7 +[8] Garrick Brazil, Gerard Pons-Moll, Xiaoming Liu, and Bernt Schiele. Kinematic 3D object detection in monocular video. In ECCV, 2020. 2 +[9] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuScenes: A multimodal dataset for autonomous driving. In CVPR, 2020. 2, 3 +[10] Ming-Fang Chang, John Lambert, Patsorn Sangkloy, Jagjeet Singh, Slawomir Bak, Andrew Hartnett, De Wang, Peter Carr, Simon Lucey, Deva Ramanan, et al. Argoverse: 3D tracking and forecasting with rich maps. In CVPR, 2019. 2 +[11] Hansheng Chen, Yuyao Huang, Wei Tian, Zhong Gao, and Lu Xiong. MonoRUn: Monocular 3D object detection by reconstruction and uncertainty propagation. In CVPR, 2021. 2 +[12] Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jiarui Xu, Zheng Zhang, Dazhi Cheng, Chenchen Zhu, Tianheng Cheng, Qijie Zhao, Buyu Li, Xin Lu, Rui Zhu, Yue Wu, Jifeng Dai, Jingdong Wang, Jianping Shi, Wanli Ouyang, Chen Change Loy, and Dahua Lin. MMDetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155, 2019. 2 +[13] Xiaozhi Chen, Kaustav Kundu, Ziyu Zhang, Huimin Ma, Sanja Fidler, and Raquel Urtasun. Monocular 3D object detection for autonomous driving. In CVPR, 2016. 1, 2 +[14] Yongjian Chen, Lei Tai, Kai Sun, and Mingyang Li. MonoPair: Monocular 3D object detection using pairwise spatial relationships. In CVPR, 2020. 2 + +[15] Yi-Nan Chen, Hang Dai, and Yong Ding. Pseudo-stereo for monocular 3D object detection in autonomous driving. In CVPR, 2022. 2 +[16] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. ScanNet: Richly-annotated 3D reconstructions of indoor scenes. In CVPR, 2017. 2 +[17] Saumitro Dasgupta, Kuan Fang, Kevin Chen, and Silvio Savarese. Delay: Robust spatial layout estimation for cluttered indoor scenes. In CVPR, 2016. 2 +[18] Mingyu Ding, Yuqi Huo, Hongwei Yi, Zhe Wang, Jianping Shi, Zhiwu Lu, and Ping Luo. Learning depth-guided convolutions for monocular 3D object detection. In CVPRW, 2020. 2 +[19] Huan Fu, Mingming Gong, Chaohui Wang, Kayhan Bat-manghelich, and Dacheng Tao. Deep Ordinal Regression Network for Monocular Depth Estimation. In CVPR, 2018. 2 +[20] Nils Gählert, Nicolas Jourdan, Marius Cordts, Uwe Franke, and Joachim Denzler. Cityscapes 3D: Dataset and benchmark for 9 DOF vehicle detection. In CVPRW, 2020. 2 +[21] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the KITTI vision benchmark suite. In CVPR, 2012. 1, 2, 3, 6 +[22] Georgia Gkioxari, Jitendra Malik, and Justin Johnson. Mesh R-CNN. In ICCV, 2019. 2 +[23] Jiaqi Gu, Bojian Wu, Lubin Fan, Jianqiang Huang, Shen Cao, Zhiyu Xiang, and Xian-Sheng Hua. Homography loss for monocular 3D object detection. In CVPR, 2022. 2 +[24] Agrim Gupta, Piotr Dollar, and Ross Girshick. LVIS: A dataset for large vocabulary instance segmentation. In CVPR, 2019. 1, 3 +[25] Chenhang He, Jianqiang Huang, Xian-Sheng Hua, and Lei Zhang. Aug3D-RPN: Improving monocular 3D object detection by synthetic images with virtual depth. arXiv preprint arXiv:2107.13269, 2021. 2 +[26] Kaiming He, Georgia Gkioxari, Piotr Dollár, and Ross Girshick. Mask R-CNN. In ICCV, 2017. 1, 2, 4 +[27] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 8 +[28] Varsha Hedau, Derek Hoiem, and David Forsyth. Recovering the spatial layout of cluttered rooms. In ICCV, 2009. 2 +[29] John Houston, Guido Zuidhof, Luca Bergamini, Yawei Ye, Long Chen, Ashesh Jain, Sammy Omari, Vladimir Iglovikov, and Peter Ondruska. One thousand and one hours: Self-driving motion prediction dataset. arXiv preprint arXiv:2006.14480, 2020. 2 +[30] Kuan-Chih Huang, Tsung-Han Wu, Hung-Ting Su, and Winston Hsu. MonoDTR: Monocular 3D object detection with depth-aware transformer. In CVPR, 2022. 2, 7 +[31] Siyuan Huang, Siyuan Qi, Yinxue Xiao, Yixin Zhu, Ying Nian Wu, and Song-Chun Zhu. Cooperative holistic scene understanding: Unifying 3D object, layout, and camera pose estimation. In NeurIPS, 2018. 1, 2 +[32] Xinyu Huang, Peng Wang, Xinjing Cheng, Dingfu Zhou, Qichuan Geng, and Ruigang Yang. The apolloscape open dataset for autonomous driving and its application. TPAMI, 2019. 2 + +[33] R. Kesten, M. Usman, J. Houston, T. Pandya, K. Nadhamuni, A. Ferreira, M. Yuan, B. Low, A. Jain, P. Ondruska, S. Omari, S. Shah, A. Kulkarni, A. Kazakova, C. Tao, L. Platinsky, W. Jiang, and V. Shet. Level 5 perception dataset 2020. https://level-5.global/level5/data/, 2019.2 +[34] Dahun Kim, Tsung-Yi Lin, Anelia Angelova, In Kweon, and Weicheng Kuo. Learning open-world object proposals without learning to classify. RAL, 2022. 4 +[35] Jason Ku, Melissa Mozifian, Jungwook Lee, Ali Harakeh, and Steven Waslander. Joint 3D proposal generation and object detection from view aggregation. In IROS, 2018. 2 +[36] Nilesh Kulkarni, Ishan Misra, Shubham Tulsiani, and Abhinav Gupta. 3D-RelNet: Joint object and relational network for 3D prediction. In ICCV, 2019. 1, 2 +[37] Abhinav Kumar, Garrick Brazil, Enrique Corona, Armin Parchami, and Xiaoming Liu. DEVIANT: Depth EquiVarIAnt NeTwork for monocular 3D object detection. In ECCV, 2022. 2 +[38] Abhinav Kumar, Garrick Brazil, and Xiaoming Liu. GrooMeD-NMS: Grouped mathematically differentiable NMS for monocular 3D object detection. In CVPR, 2021. 2 +[39] Abhijit Kundu, Yin Li, and James Rehg. 3D-RCNN: Instance-level 3D object reconstruction via render-and-compare. In CVPR, 2018. 5 +[40] Hei Law and Jia Deng. CornerNet: Detecting objects as paired keypoints. In ECCV, 2018. 1 +[41] David C Lee, Martial Hebert, and Takeo Kanade. Geometric reasoning for single image structure recovery. In CVPR, 2009. 2 +[42] Peixuan Li, Huaici Zhao, Pengfei Liu, and Feidao Cao. RTM3D: Real-time monocular 3D detection from object keypoints for autonomous driving. In ECCV, 2020. 2 +[43] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In CVPR, 2017. 6 +[44] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In ICCV, 2017. 2 +[45] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dolkar, and C Lawrence Zitnick. Microsoft COCO: Common objects in context. In ECCV, 2014. 1, 3, 8 +[46] Ce Liu, Shuhang Gu, Luc Van Gool, and Radu Timofte. Deep line encoding for monocular 3D object detection and depth prediction. In BMVC, 2021. 2 +[47] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander Berg. SSD: Single shot multibox detector. In ECCV, 2016. 2 +[48] Xianpeng Liu, Nan Xue, and Tianfu Wu. Learning auxiliary monocular contexts helps monocular 3D object detection. In AAAI, 2022. 2 +[49] Yuxuan Liu, Yuan Yixuan, and Ming Liu. Ground-aware monocular 3D object detection for autonomous driving. RAL, 2021. 2 +[50] Zechen Liu, Zizhang Wu, and Roland Tóth. SMOKE: Single-stage monocular 3D object detection via keypoint estimation. In CVPRW, 2020. 1, 2, 5, 7 + +[51] Zongdai Liu, Dingfu Zhou, Feixiang Lu, Jin Fang, and Liangjun Zhang. Autoshape: Real-time shape-aware monocular 3D object detection. In ICCV, 2021. 2 +[52] Yan Lu, Xinzhu Ma, Lei Yang, Tianzhu Zhang, Yating Liu, Qi Chu, Junjie Yan, and Wanli Ouyang. Geometry uncertainty projection network for monocular 3D object detection. In ICCV, 2021. 1, 2, 5, 7 +[53] Xinzhu Ma, Shinan Liu, Zhiyi Xia, Hongwen Zhang, Xingyu Zeng, and Wanli Ouyang. Rethinking pseudo-lidar representation. In ECCV, 2020. 2 +[54] Xinzhu Ma, Zhihui Wang, Haojie Li, Pengbo Zhang, Wanli Ouyang, and Xin Fan. Accurate monocular 3D object detection via color-embedded 3D reconstruction for autonomous driving. In ICCV, 2019. 2 +[55] Xinzhu Ma, Yinmin Zhang, Dan Xu, Dongzhan Zhou, Shuai Yi, Haojie Li, and Wanli Ouyang. Delving into localization errors for monocular 3D object detection. In CVPR, 2021. 2 +[56] Arun Mallya and Svetlana Lazebnik. Learning informative edge maps for indoor scene layout prediction. In ICCV, 2015. 2 +[57] Arsalan Mousavian, Dragomir Anguelov, John Flynn, and Jana Kosecka. 3D bounding box estimation using deep learning and geometry. In CVPR, 2017. 1, 2 +[58] Yinyu Nie, Xiaoguang Han, Shihui Guo, Yujuan Zheng, Jian Chang, and Jian Zhang. Total3DUnderstanding: Joint layout, object pose and mesh reconstruction for indoor scenes from a single image. In CVPR, 2020. 1, 2, 8 +[59] Dennis Park, Rares Ambrus, Vitor Guizilini, Jie Li, and Adrien Gaidon. Is pseudo-lidar needed for monocular 3D object detection? In ICCV, 2021. 2, 7 +[60] Charles Qi, Wei Liu, Chenxia Wu, Hao Su, and Leonidas Guibas. Frustum pointnets for 3D object detection from RGB-D data. In CVPR, 2018. 2 +[61] Nikhila Ravi, Jeremy Reizenstein, David Novotny, Taylor Gordon, Wan-Yen Lo, Justin Johnson, and Georgia Gkioxari. Accelerating 3D deep learning with PyTorch3D. arXiv:2007.08501, 2020. 6 +[62] Cody Reading, Ali Harakeh, Julia Chae, and Steven Waslander. Categorical depth distribution network for monocular 3D object detection. In CVPR, 2021. 2 +[63] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In CVPR, 2016. 1, 2 +[64] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster R-CNN: Towards real-time object detection with region proposal networks. In NeurIPS, 2015. 1, 2, 4, 5 +[65] Mike Roberts, Jason Ramapuram, Anurag Ranjan, Atulit Kumar, Miguel Bautista, Nathan Paczan, Russ Webb, and Joshua Susskind. Hypersim: A photorealistic synthetic dataset for holistic indoor scene understanding. In ICCV, 2021. 2, 3 +[66] Danila Rukhovich, Anna Vorontsova, and Anton Konushin. ImVoxelNet: Image to voxels projection for monocular and multi-view general-purpose 3D object detection. In WACV, 2022. 2, 7, 8 +[67] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander Berg, and Li + +Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. IJCV, 2015. 6 +[68] Andrea Simonelli, Samuel Rota Bulo, Lorenzo Porzi, Manuel López-Antequera, and Peter Kontschieder. Disentangling monocular 3D object detection. In ICCV, 2019. 2, 4, 5 +[69] Andrea Simonelli, Samuel Rota Bulo, Lorenzo Porzi, Elisa Ricci, and Peter Kontschieder. Towards generalization across depth for monocular 3D object detection. In ECCV, 2020. 2 +[70] Shuran Song, Samuel Lichtenberg, and Jianxiong Xiao. SUN RGB-D: A RGB-D scene understanding benchmark suite. In CVPR, 2015. 1, 2, 3 +[71] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, Vijay Vasudevan, Wei Han, Jiquan Ngiam, Hang Zhao, Aleksei Timofeev, Scott Ettinger, Maxim Krivokon, Amy Gao, Aditya Joshi, Yu Zhang, Jonathon Shlens, Zhifeng Chen, and Dragomir Anguelov. Scalability in perception for autonomous driving: Waymo open dataset. In CVPR, 2020. 2 +[72] Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. FCOS: Fully convolutional one-stage object detection. In ICCV, 2019. 1, 2 +[73] Shubham Tulsiani, Saurabh Gupta, David Fouhey, Alexei Efros, and Jitendra Malik. Factoring shape, pose, and layout from the 2D image of a 3D scene. In CVPR, 2018. 1, 2 +[74] Pauli Virtanen, Ralf Gommers, Travis Oliphant, Matt Haberland, Tyler Reddy, David Cournapeau, Evgeni Burovski, Pearu Peterson, Warren Weckesser, Jonathan Bright, et al. Scipy 1.0: fundamental algorithms for scientific computing in python. Nature methods, 2020. 6 +[75] Li Wang, Liang Du, Xiaqing Ye, Yanwei Fu, Guodong Guo, Xiangyang Xue, Jianfeng Feng, and Li Zhang. Depth-conditioned dynamic message propagation for monocular 3D object detection. In CVPR, 2021. 2 +[76] Li Wang, Li Zhang, Yi Zhu, Zhi Zhang, Tong He, Mu Li, and Xiangyang Xue. Progressive coordinate transforms for monocular 3D object detection. In NeurIPS, 2021. 2 +[77] Tai Wang, Zhu Xinge, Jiangmiao Pang, and Dahua Lin. Probabilistic and geometric depth: Detecting objects in perspective. In CoRL, 2022. 2, 7 +[78] Tai Wang, Xinge Zhu, Jiangmiao Pang, and Dahua Lin. FCOS3D: Fully convolutional one-stage monocular 3D object detection. In ICCVW, 2021. 2, 7 +[79] Yan Wang, Wei-Lun Chao, Divyansh Garg, Bharath Hariharan, Mark Campbell, and Kilian Weinberger. Pseudo-lidar from visual depth estimation: Bridging the gap in 3D object detection for autonomous driving. In CVPR, 2019. 2 +[80] Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. Detector2. https://github.com/facebookresearch/detectron2, 2019. 2, 6 +[81] Yurong You, Yan Wang, Wei-Lun Chao, Divyansh Garg, Geoff Pleiss, Bharath Hariharan, Mark Campbell, and Kilian Weinberger. Pseudo-lidar++: Accurate depth for 3D object detection in autonomous driving. In ICLR, 2020. 2 +[82] Fisher Yu, Dequan Wang, Evan Shelhamer, and Trevor Darrell. Deep layer aggregation. In CVPR, 2018. 6 +[83] Yunpeng Zhang, Jiwen Lu, and Jie Zhou. Objects are different: Flexible monocular 3D object detection. In CVPR, 2021. 2 + +[84] Dingfu Zhou, Xibin Song, Yuchao Dai, Junbo Yin, Feixiang Lu, Miao Liao, Jin Fang, and Liangjun Zhang. IAFA: Instance-aware feature aggregation for 3D object detection from a single image. In ACCV, 2020. 2 +[85] Xingyi Zhou, Dequan Wang, and Philipp Krahenbuhl. Objects as points. In arXiv preprint arXiv:1904.07850, 2019. 2 +[86] Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. In CVPR, 2019. 4, 5 +[87] Yunsong Zhou, Yuan He, Hongzi Zhu, Cheng Wang, Hongyang Li, and Qinhong Jiang. Monocular 3D object detection: An extrinsic parameter free approach. In CVPR, 2021. 2 +[88] Zhikang Zou, Xiaqing Ye, Liang Du, Xianhui Cheng, Xiao Tan, Li Zhang, Jianfeng Feng, Xiangyang Xue, and Errui Ding. The devil is in the task: Exploiting reciprocal appearance-localization features for monocular 3D object detection. In ICCV, 2021. 2 \ No newline at end of file diff --git a/2023/Omni3D_ A Large Benchmark and Model for 3D Object Detection in the Wild/images.zip b/2023/Omni3D_ A Large Benchmark and Model for 3D Object Detection in the Wild/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..819f5441770b7059977f85fdd1c0fd693b9d65b2 --- /dev/null +++ b/2023/Omni3D_ A Large Benchmark and Model for 3D Object Detection in the Wild/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:700937ff81c9135b551817423e11e37619f5bd8712c444872116ced1b03ddf6e +size 621488 diff --git a/2023/Omni3D_ A Large Benchmark and Model for 3D Object Detection in the Wild/layout.json b/2023/Omni3D_ A Large Benchmark and Model for 3D Object Detection in the Wild/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..dbb7e489e5b3578cc8202ac74e32bc98a7f171fa --- /dev/null +++ b/2023/Omni3D_ A Large Benchmark and Model for 3D Object Detection in the Wild/layout.json @@ -0,0 +1,11894 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 54, + 103, + 539, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 103, + 539, + 120 + ], + "spans": [ + { + "bbox": [ + 54, + 103, + 539, + 120 + ], + "type": "text", + "content": "OMNI3D: A Large Benchmark and Model for 3D Object Detection in the Wild" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 120, + 142, + 457, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 142, + 457, + 175 + ], + "spans": [ + { + "bbox": [ + 120, + 142, + 457, + 175 + ], + "type": "text", + "content": "Garrick Brazil1 Abhinav Kumar2 Julian Straub1 Nikhila Ravi1 Justin Johnson1 Georgia Gkioxari3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 170, + 179, + 419, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 170, + 179, + 419, + 194 + ], + "spans": [ + { + "bbox": [ + 170, + 179, + 419, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 170, + 179, + 419, + 194 + ], + "type": "text", + "content": "Meta AI " + }, + { + "bbox": [ + 170, + 179, + 419, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 170, + 179, + 419, + 194 + ], + "type": "text", + "content": "Michigan State University " + }, + { + "bbox": [ + 170, + 179, + 419, + 194 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 170, + 179, + 419, + 194 + ], + "type": "text", + "content": "Caltech" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 50, + 224, + 189, + 342 + ], + "blocks": [ + { + "bbox": [ + 50, + 224, + 189, + 342 + ], + "lines": [ + { + "bbox": [ + 50, + 224, + 189, + 342 + ], + "spans": [ + { + "bbox": [ + 50, + 224, + 189, + 342 + ], + "type": "image", + "image_path": "cdc5e4b3a711848ba026a9f90d0ae7eb0318877f0b52fa7ff8fa6e6af19cd00d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 346, + 547, + 369 + ], + "lines": [ + { + "bbox": [ + 46, + 346, + 547, + 369 + ], + "spans": [ + { + "bbox": [ + 46, + 346, + 547, + 369 + ], + "type": "text", + "content": "Figure 1. Left: We introduce OMNI3D, a benchmark for 3D object detection which is larger and more diverse than popular 3D benchmarks. Right: We propose Cube R-CNN, which generalizes to unseen datasets (e.g. COCO [45]) and outperforms prior works on existing datasets." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 194, + 227, + 304, + 340 + ], + "blocks": [ + { + "bbox": [ + 194, + 227, + 304, + 340 + ], + "lines": [ + { + "bbox": [ + 194, + 227, + 304, + 340 + ], + "spans": [ + { + "bbox": [ + 194, + 227, + 304, + 340 + ], + "type": "image", + "image_path": "b2e4cd0d07e95acf9efe8fa561fe11ac96e94ea29c63c3de6bdc344f252986b5.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 306, + 227, + 433, + 341 + ], + "blocks": [ + { + "bbox": [ + 306, + 227, + 433, + 341 + ], + "lines": [ + { + "bbox": [ + 306, + 227, + 433, + 341 + ], + "spans": [ + { + "bbox": [ + 306, + 227, + 433, + 341 + ], + "type": "image", + "image_path": "d7c28d109f6884ad9edc9e375242ea1b28f20535f1caa7d0f9cd691613cd76d8.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 440, + 227, + 545, + 339 + ], + "blocks": [ + { + "bbox": [ + 440, + 227, + 545, + 339 + ], + "lines": [ + { + "bbox": [ + 440, + 227, + 545, + 339 + ], + "spans": [ + { + "bbox": [ + 440, + 227, + 545, + 339 + ], + "type": "image", + "image_path": "c54b7910c8ec5a1763fc91285dbf156997eae3fc5de7482371add9b80e36c733.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 143, + 389, + 191, + 402 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 389, + 191, + 402 + ], + "spans": [ + { + "bbox": [ + 143, + 389, + 191, + 402 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 416, + 290, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 416, + 290, + 668 + ], + "spans": [ + { + "bbox": [ + 46, + 416, + 290, + 668 + ], + "type": "text", + "content": "Recognizing scenes and objects in 3D from a single image is a longstanding goal of computer vision with applications in robotics and AR/VR. For 2D recognition, large datasets and scalable solutions have led to unprecedented advances. In 3D, existing benchmarks are small in size and approaches specialize in few object categories and specific domains, e.g. urban driving scenes. Motivated by the success of 2D recognition, we revisit the task of 3D object detection by introducing a large benchmark, called OMNI3D. OMNI3D re-purposes and combines existing datasets resulting in 234k images annotated with more than 3 million instances and 98 categories. 3D detection at such scale is challenging due to variations in camera intrinsic and the rich diversity of scene and object types. We propose a model, called Cube R-CNN, designed to generalize across camera and scene types with a unified approach. We show that Cube R-CNN outperforms prior works on the larger OMNI3D and existing benchmarks. Finally, we prove that OMNI3D is a powerful dataset for 3D object recognition and show that it improves single-dataset performance and can accelerate learning on new smaller datasets via pre-training.1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 307, + 389, + 387, + 402 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 389, + 387, + 402 + ], + "spans": [ + { + "bbox": [ + 307, + 389, + 387, + 402 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 411, + 547, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 411, + 547, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 411, + 547, + 544 + ], + "type": "text", + "content": "Understanding objects and their properties from single images is a longstanding problem in computer vision with applications in robotics and AR/VR. In the last decade, 2D object recognition [26,40,63,64,72] has made tremendous advances toward predicting objects on the image plane with the help of large datasets [24,45]. However, the world and its objects are three dimensional laid out in 3D space. Perceiving objects in 3D from 2D visual inputs poses new challenges framed by the task of 3D object detection. Here, the goal is to estimate a 3D location and 3D extent of each object in an image in the form of a tight oriented 3D bounding box." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 545, + 548, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 548, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 548, + 714 + ], + "type": "text", + "content": "Today 3D object detection is studied under two different lenses: for urban domains in the context of autonomous vehicles [7,13,50,52,57] or indoor scenes [31,36,58,73]. Despite the problem formulation being shared, methods share little insights between domains. Often approaches are tailored to work only for the domain in question. For instance, urban methods make assumptions about objects resting on a ground plane and model only yaw angles for 3D rotation. Indoor techniques may use a confined depth range (e.g. up to " + }, + { + "bbox": [ + 304, + 545, + 548, + 714 + ], + "type": "inline_equation", + "content": "6\\mathrm{m}" + }, + { + "bbox": [ + 304, + 545, + 548, + 714 + ], + "type": "text", + "content": " in [58]). These assumptions are generally not true in the real world. Moreover, the most popular benchmarks for image-based 3D object detection are small. Indoor SUN RGB-D [70] has " + }, + { + "bbox": [ + 304, + 545, + 548, + 714 + ], + "type": "inline_equation", + "content": "10\\mathrm{k}" + }, + { + "bbox": [ + 304, + 545, + 548, + 714 + ], + "type": "text", + "content": " images, urban KITTI [21] has " + }, + { + "bbox": [ + 304, + 545, + 548, + 714 + ], + "type": "inline_equation", + "content": "7\\mathrm{k}" + }, + { + "bbox": [ + 304, + 545, + 548, + 714 + ], + "type": "text", + "content": " images; 2D benchmarks like COCO [45] are " + }, + { + "bbox": [ + 304, + 545, + 548, + 714 + ], + "type": "inline_equation", + "content": "20\\times" + }, + { + "bbox": [ + 304, + 545, + 548, + 714 + ], + "type": "text", + "content": " larger." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 693, + 272, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 693, + 272, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 693, + 272, + 713 + ], + "type": "text", + "content": "1We release the OMNI3D benchmark and Cube R-CNN models at https://github.com/facebookresearch/omni3d." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "13154" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 300 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 300 + ], + "type": "text", + "content": "We address the absence of a general large-scale dataset for 3D object detection by introducing a large and diverse 3D benchmark called OMNI3D. OMNI3D is curated from publicly released datasets, SUN RBG-D [70], ARKitScenes [6], Hypersim [65], Objectron [2], KITTI [21] and nuScenes [9], and comprises 234k images with 3 million objects annotated with 3D boxes across 98 categories including chair, sofa, laptop, table, cup, shoes, pillow, books, car, person, etc. Sec. 3 describes the curation process which involves re-purposing the raw data and annotations from the aforementioned datasets, which originally target different applications. As shown in Fig. 1, OMNI3D is " + }, + { + "bbox": [ + 47, + 72, + 289, + 300 + ], + "type": "inline_equation", + "content": "20\\times" + }, + { + "bbox": [ + 47, + 72, + 289, + 300 + ], + "type": "text", + "content": " larger than existing popular benchmarks used for 3D detection, SUN RGB-D and KITTI. For efficient evaluation on the large OMNI3D, we introduce a new algorithm for intersection-over-union of 3D boxes which is " + }, + { + "bbox": [ + 47, + 72, + 289, + 300 + ], + "type": "inline_equation", + "content": "450\\times" + }, + { + "bbox": [ + 47, + 72, + 289, + 300 + ], + "type": "text", + "content": " faster than previous solutions [2]. We empirically prove the impact of OMNI3D as a large-scale dataset and show that it improves single-dataset performance by up to " + }, + { + "bbox": [ + 47, + 72, + 289, + 300 + ], + "type": "inline_equation", + "content": "5.3\\%" + }, + { + "bbox": [ + 47, + 72, + 289, + 300 + ], + "type": "text", + "content": " AP on urban and " + }, + { + "bbox": [ + 47, + 72, + 289, + 300 + ], + "type": "inline_equation", + "content": "3.8\\%" + }, + { + "bbox": [ + 47, + 72, + 289, + 300 + ], + "type": "text", + "content": " on indoor benchmarks." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 300, + 289, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 300, + 289, + 551 + ], + "spans": [ + { + "bbox": [ + 47, + 300, + 289, + 551 + ], + "type": "text", + "content": "On the large and diverse OMNI3D, we design a general and simple 3D object detector, called Cube R-CNN, inspired by advances in 2D and 3D recognition of recent years [22,52,64,68]. Cube R-CNN detects all objects and their 3D location, size and rotation end-to-end from a single image of any domain and for many object categories. Attributed to OMNI3D's diversity, our model shows strong generalization and outperforms prior works for indoor and urban domains with one unified model, as shown in Fig. 1. Learning from such diverse data comes with challenges as OMNI3D contains images of highly varying focal lengths which exaggerate scale-depth ambiguity (Fig. 4). We remedy this by operating on virtual depth which transforms object depth with the same virtual camera intrinsics across the dataset. An added benefit of virtual depth is that it allows the use of data augmentations (e.g. image rescaling) during training, which is a critical feature for 2D detection [12,80], and as we show, also for 3D. Our approach with one unified design outperforms prior best approaches in " + }, + { + "bbox": [ + 47, + 300, + 289, + 551 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{\\mathrm{3D}}" + }, + { + "bbox": [ + 47, + 300, + 289, + 551 + ], + "type": "text", + "content": ", ImVoxelNet [66] by " + }, + { + "bbox": [ + 47, + 300, + 289, + 551 + ], + "type": "inline_equation", + "content": "4.1\\%" + }, + { + "bbox": [ + 47, + 300, + 289, + 551 + ], + "type": "text", + "content": " on indoor SUN RGB-D, GUPNet [52] by " + }, + { + "bbox": [ + 47, + 300, + 289, + 551 + ], + "type": "inline_equation", + "content": "9.5\\%" + }, + { + "bbox": [ + 47, + 300, + 289, + 551 + ], + "type": "text", + "content": " on urban KITTI, and PGD [77] by " + }, + { + "bbox": [ + 47, + 300, + 289, + 551 + ], + "type": "inline_equation", + "content": "7.9\\%" + }, + { + "bbox": [ + 47, + 300, + 289, + 551 + ], + "type": "text", + "content": " on OMNI3D." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 59, + 552, + 195, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 552, + 195, + 563 + ], + "spans": [ + { + "bbox": [ + 59, + 552, + 195, + 563 + ], + "type": "text", + "content": "We summarize our contributions:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 573, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 59, + 573, + 287, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 573, + 287, + 609 + ], + "spans": [ + { + "bbox": [ + 59, + 573, + 287, + 609 + ], + "type": "text", + "content": "- We introduce OMNI3D, a benchmark for image-based 3D object detection sourced from existing 3D datasets, which is " + }, + { + "bbox": [ + 59, + 573, + 287, + 609 + ], + "type": "inline_equation", + "content": "20 \\times" + }, + { + "bbox": [ + 59, + 573, + 287, + 609 + ], + "type": "text", + "content": " larger than existing 3D benchmarks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 619, + 287, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 619, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 59, + 619, + 287, + 643 + ], + "type": "text", + "content": "- We implement a new algorithm for IoU of 3D boxes, which is " + }, + { + "bbox": [ + 59, + 619, + 287, + 643 + ], + "type": "inline_equation", + "content": "450 \\times" + }, + { + "bbox": [ + 59, + 619, + 287, + 643 + ], + "type": "text", + "content": " faster than prior solutions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 653, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 653, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 59, + 653, + 287, + 713 + ], + "type": "text", + "content": "- We design a general-purpose baseline method, Cube R-CNN, which tackles 3D object detection for many categories and across domains with a unified approach. We propose virtual depth to eliminate the ambiguity from varying camera focal lengths in OMNI3D." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 71, + 392, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 392, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 392, + 84 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 91, + 547, + 116 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 91, + 547, + 116 + ], + "spans": [ + { + "bbox": [ + 305, + 91, + 547, + 116 + ], + "type": "text", + "content": "Cube R-CNN and OMNI3D draw from key research advances in 2D and 3D object detection." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 119, + 547, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 119, + 547, + 179 + ], + "spans": [ + { + "bbox": [ + 305, + 119, + 547, + 179 + ], + "type": "text", + "content": "2D Object Detection. Here, methods include two-stage approaches [26, 64] which predict object regions with a region proposal network (RPN) and then refine them via an MLP. Single-stage detectors [44, 47, 63, 72, 85] omit the RPN and predict regions directly from the backbone." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 182, + 547, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 182, + 547, + 398 + ], + "spans": [ + { + "bbox": [ + 305, + 182, + 547, + 398 + ], + "type": "text", + "content": "3D Object Detection. Monocular 3D object detectors predict 3D cuboids from single input images. There is extensive work in the urban self-driving domain where the car class is at the epicenter [13, 19, 23, 25, 30, 46, 49, 54, 57, 62, 68, 69, 76, 78, 84]. CenterNet [85] predicts 3D depth and size from fully-convolutional center features, and is extended by [14, 37, 42, 48, 50–52, 55, 83, 84, 87]. M3D-RPN [7] trains an RPN with 3D anchors, enhanced further by [8, 18, 38, 75, 88]. FCOS3D [78] extends the anchorless FCOS [72] detector to predict 3D cuboids. Its successor PGD [77] furthers the approach with probabilistic depth uncertainty. Others use pseudo depth [4, 15, 53, 59, 79, 81] and explore depth and point-based LiDAR techniques [35, 60]. Similar to ours, [11, 68, 69] add a 3D head, specialized for urban scenes and objects, on two-stage Faster R-CNN. [25, 69] augment their training by synthetically generating depth and box-fitted views, coined as virtual views or depth. In our work, virtual depth aims at addressing varying focal lengths." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 399, + 547, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 399, + 547, + 531 + ], + "spans": [ + { + "bbox": [ + 305, + 399, + 547, + 531 + ], + "type": "text", + "content": "For indoor scenes, a vast line of work tackles room layout estimation [17,28,41,56]. Huang et al [31] predict 3D oriented bounding boxes for indoor objects. Factored3D [73] and 3D-RelNet [36] jointly predict object voxel shapes. Total3D [58] predicts 3D boxes and meshes by additionally training on datasets with annotated 3D shapes. ImVoxelNet [66] proposes domain-specific methods which share an underlying framework for processing volumes of 3D voxels. In contrast, we explore 3D object detection in its general form by tackling outdoor and indoor domains jointly in a single model and with a vocabulary of " + }, + { + "bbox": [ + 305, + 399, + 547, + 531 + ], + "type": "inline_equation", + "content": "5 \\times" + }, + { + "bbox": [ + 305, + 399, + 547, + 531 + ], + "type": "text", + "content": " more categories." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 533, + 548, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 533, + 548, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 533, + 548, + 713 + ], + "type": "text", + "content": "3D Datasets. KITTI [21] and SUN RGB-D [70] are popular datasets for 3D object detection on urban and indoor scenes respectively. Since 2019, 3D datasets have emerged, both for indoor [2, 3, 6, 16, 65] and outdoor [9, 10, 20, 29, 32, 33, 71]. In isolation, these datasets target different tasks and applications and have unique properties and biases, e.g. object and scene types, focal length, coordinate systems, etc. In this work, we unify existing representative datasets [2, 6, 9, 21, 65, 70]. We process the raw visual data, re-purpose their annotations, and carefully curate the union of their semantic labels in order to build a coherent large-scale benchmark, called OMNI3D. OMNI3D is " + }, + { + "bbox": [ + 305, + 533, + 548, + 713 + ], + "type": "inline_equation", + "content": "20 \\times" + }, + { + "bbox": [ + 305, + 533, + 548, + 713 + ], + "type": "text", + "content": " larger than widely-used benchmarks and notably more diverse. As such, new challenges arise stemming from the increased variance in visual domain, object rotation, size, layouts, and camera intrinsics." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "13155" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 68, + 272, + 177 + ], + "blocks": [ + { + "bbox": [ + 47, + 68, + 272, + 177 + ], + "lines": [ + { + "bbox": [ + 47, + 68, + 272, + 177 + ], + "spans": [ + { + "bbox": [ + 47, + 68, + 272, + 177 + ], + "type": "image", + "image_path": "789db86a26df8ee5fbd79eebf2d7ad9f27b912998be6b082816bdf1e6913a452.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 123, + 180, + 203, + 192 + ], + "lines": [ + { + "bbox": [ + 123, + 180, + 203, + 192 + ], + "spans": [ + { + "bbox": [ + 123, + 180, + 203, + 192 + ], + "type": "text", + "content": "(a) Spatial statistics" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 273, + 84, + 402, + 174 + ], + "blocks": [ + { + "bbox": [ + 273, + 84, + 402, + 174 + ], + "lines": [ + { + "bbox": [ + 273, + 84, + 402, + 174 + ], + "spans": [ + { + "bbox": [ + 273, + 84, + 402, + 174 + ], + "type": "image", + "image_path": "4063bb12980d1e73898db058fa5b56c72a5b320cca0603175613ba5e926cfce7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 296, + 180, + 386, + 190 + ], + "lines": [ + { + "bbox": [ + 296, + 180, + 386, + 190 + ], + "spans": [ + { + "bbox": [ + 296, + 180, + 386, + 190 + ], + "type": "text", + "content": "(b) 2D Scale statistics" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 45, + 194, + 547, + 217 + ], + "lines": [ + { + "bbox": [ + 45, + 194, + 547, + 217 + ], + "spans": [ + { + "bbox": [ + 45, + 194, + 547, + 217 + ], + "type": "text", + "content": "Figure 2. OMNI3D analysis. (a) distribution of object centers on (top) normalized image, XY-plane, and (bottom) normalized depth, XZ-plane, (b) relative 2D object scales, and (c) category frequency." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 403, + 85, + 534, + 174 + ], + "blocks": [ + { + "bbox": [ + 403, + 85, + 534, + 174 + ], + "lines": [ + { + "bbox": [ + 403, + 85, + 534, + 174 + ], + "spans": [ + { + "bbox": [ + 403, + 85, + 534, + 174 + ], + "type": "image", + "image_path": "0e975865c5d381f2f1f79180d494a13640ed937d56f993f333c2d1dd2d6c11f5.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 432, + 180, + 520, + 191 + ], + "lines": [ + { + "bbox": [ + 432, + 180, + 520, + 191 + ], + "spans": [ + { + "bbox": [ + 432, + 180, + 520, + 191 + ], + "type": "text", + "content": "(c) Category statistics" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 228, + 195, + 241 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 228, + 195, + 241 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 195, + 241 + ], + "type": "text", + "content": "3. The OMNI3D Benchmark" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 45, + 249, + 288, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 249, + 288, + 344 + ], + "spans": [ + { + "bbox": [ + 45, + 249, + 288, + 344 + ], + "type": "text", + "content": "The primary benchmarks for 3D object detection are small, focus on a few categories and are of a single domain. For instance, the popular KITTI [21] contains only urban scenes, has 7k images and 3 categories, with a focus on car. SUN RGB-D [70] has 10k images. The small size and lack of variance in 3D datasets is a stark difference to 2D counterparts, such as COCO [45] and LVIS [24], which have pioneered progress in 2D recognition." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 346, + 290, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 346, + 290, + 537 + ], + "spans": [ + { + "bbox": [ + 46, + 346, + 290, + 537 + ], + "type": "text", + "content": "We aim to bridge the gap to 2D by introducing OMNI3D, a large-scale and diverse benchmark for image-based 3D object detection consisting of " + }, + { + "bbox": [ + 46, + 346, + 290, + 537 + ], + "type": "inline_equation", + "content": "234\\mathrm{k}" + }, + { + "bbox": [ + 46, + 346, + 290, + 537 + ], + "type": "text", + "content": " images, 3 million labeled 3D bounding boxes, and 98 object categories. We source from recently released 3D datasets of urban (nuScenes [9] and KITTI [21]), indoor (SUN RGB-D [70], ARKitScenes [6] and Hypersim [65]), and general (Objectron [2]) scenes. Each of these datasets target different applications (e.g. point-cloud recognition or reconstruction in [6], inverse rendering in [65]), provide visual data in different forms (e.g. videos in [2, 6], rig captures in [9]) and annotate different object types. To build a coherent benchmark, we process the varying raw visual data, re-purpose their annotations to extract 3D cuboids in a unified 3D camera coordinate system, and carefully curate the final vocabulary. More details about the benchmark creation in the Appendix." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 537, + 288, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 537, + 288, + 597 + ], + "spans": [ + { + "bbox": [ + 46, + 537, + 288, + 597 + ], + "type": "text", + "content": "We analyze OMNI3D and show its rich spatial and semantic properties proving it is visually diverse, similar to 2D data, and highly challenging for 3D as depicted in Fig. 2. We show the value of OMNI3D for the task of 3D object detection with extensive quantitative analysis in Sec. 5." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 608, + 147, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 608, + 147, + 620 + ], + "spans": [ + { + "bbox": [ + 46, + 608, + 147, + 620 + ], + "type": "text", + "content": "3.1. Dataset Analysis" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": "Splits. We split the dataset into " + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "inline_equation", + "content": "175\\mathrm{k} / 19\\mathrm{k} / 39\\mathrm{k}" + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": " images for train/val/test respectively, consistent with original splits when available, and otherwise free of overlapping video sequences in splits. We denote indoor and outdoor subsets as " + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{OMN13D_{IN}}" + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": " (SUN RGB-D, Hypersim, ARKit), and " + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{OMN13D_{OUT}}" + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": " (KITTI, nuScenes). Objectron, with primarily close-up objects, is used only in the full OMN13D setting." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 229, + 547, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 229, + 547, + 361 + ], + "spans": [ + { + "bbox": [ + 304, + 229, + 547, + 361 + ], + "type": "text", + "content": "Layout statistics. Fig. 2(a) shows the distribution of object centers onto the image plane by projecting centroids on the XY-plane (top row), and the distribution of object depths by projecting centroids onto the XZ-plane (bottom row). We find that OMNI3D's spatial distribution has a center bias, similar to 2D datasets COCO and LVIS. Fig. 2(b) depicts the relative object size distribution, defined as the square root of object area divided by image area. Objects are more likely to be small in size similar to LVIS (Fig. 6c in [24]) suggesting that OMNI3D is also challenging for 2D detection, while objects in OMNI3DOUT are noticeably smaller." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 361, + 547, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 361, + 547, + 481 + ], + "spans": [ + { + "bbox": [ + 304, + 361, + 547, + 481 + ], + "type": "text", + "content": "The bottom row of Fig. 2(a) normalizes object depth in a " + }, + { + "bbox": [ + 304, + 361, + 547, + 481 + ], + "type": "inline_equation", + "content": "[0,20\\mathrm{m}]" + }, + { + "bbox": [ + 304, + 361, + 547, + 481 + ], + "type": "text", + "content": " range, chosen for visualization and satisfies " + }, + { + "bbox": [ + 304, + 361, + 547, + 481 + ], + "type": "inline_equation", + "content": "88\\%" + }, + { + "bbox": [ + 304, + 361, + 547, + 481 + ], + "type": "text", + "content": " of object instances; OMNI3D depth ranges as far as " + }, + { + "bbox": [ + 304, + 361, + 547, + 481 + ], + "type": "inline_equation", + "content": "300\\mathrm{m}" + }, + { + "bbox": [ + 304, + 361, + 547, + 481 + ], + "type": "text", + "content": ". We observe that the OMNI3D depth distribution is far more diverse than SUN RGB-D and KITTI, which are biased toward near or road-side objects respectively. See Appendix for each data source distribution plot. Fig. 2(a) demonstrates OMNI3D's rich diversity in spatial distribution and depth which suffers significantly less bias than existing 3D benchmarks and is comparable in complexity to 2D datasets." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 483, + 547, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 483, + 547, + 592 + ], + "spans": [ + { + "bbox": [ + 304, + 483, + 547, + 592 + ], + "type": "text", + "content": "2D and 3D correlation. A common assumption in urban scenes is that objects rest on a ground plane and appear smaller with depth. To verify if that is true generally, we compute correlations. We find that " + }, + { + "bbox": [ + 304, + 483, + 547, + 592 + ], + "type": "inline_equation", + "content": "2\\mathrm{D}y" + }, + { + "bbox": [ + 304, + 483, + 547, + 592 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 483, + 547, + 592 + ], + "type": "inline_equation", + "content": "3\\mathrm{D}z" + }, + { + "bbox": [ + 304, + 483, + 547, + 592 + ], + "type": "text", + "content": " are indeed fairly correlated in " + }, + { + "bbox": [ + 304, + 483, + 547, + 592 + ], + "type": "inline_equation", + "content": "\\mathrm{OMNI3D_{OUT}}" + }, + { + "bbox": [ + 304, + 483, + 547, + 592 + ], + "type": "text", + "content": " at 0.524, but significantly less in " + }, + { + "bbox": [ + 304, + 483, + 547, + 592 + ], + "type": "inline_equation", + "content": "\\mathrm{OMNI3D_{IN}}" + }, + { + "bbox": [ + 304, + 483, + 547, + 592 + ], + "type": "text", + "content": " at 0.006. Similarly, relative 2D object size " + }, + { + "bbox": [ + 304, + 483, + 547, + 592 + ], + "type": "inline_equation", + "content": "(\\sqrt{h\\cdot w})" + }, + { + "bbox": [ + 304, + 483, + 547, + 592 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 483, + 547, + 592 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 304, + 483, + 547, + 592 + ], + "type": "text", + "content": " correlation is 0.543 and 0.102 respectively. This confirms our claim that common assumptions in urban scenes are not generally true, making the task challenging." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 594, + 548, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 548, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 548, + 714 + ], + "type": "text", + "content": "Category statistics. Fig. 2(c) plots the distribution of instances across the 98 categories of OMNI3D. The long-tail suggests that low-shot recognition in both 2D and 3D will be critical for performance. In this work, we want to focus on large-scale and diverse 3D recognition, which is comparably unexplored. We therefore filter and focus on categories with at least 1000 instances. This leaves 50 diverse categories including chair, sofa, laptop, table, books, car, truck, pedestrian and more, 19 of which have more than 10k instances. We provide more per-category details in the Appendix." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13156" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 72, + 545, + 247 + ], + "blocks": [ + { + "bbox": [ + 48, + 72, + 545, + 247 + ], + "lines": [ + { + "bbox": [ + 48, + 72, + 545, + 247 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 545, + 247 + ], + "type": "image", + "image_path": "48250649f1de3140a3a8dbcf9325d3359327e7a602b7399a3a0aa1aabbe7005f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 250, + 547, + 273 + ], + "lines": [ + { + "bbox": [ + 46, + 250, + 547, + 273 + ], + "spans": [ + { + "bbox": [ + 46, + 250, + 547, + 273 + ], + "type": "text", + "content": "Figure 3. Overview. Cube R-CNN takes as input an RGB image, detects all objects in 2D and predicts their 3D cuboids " + }, + { + "bbox": [ + 46, + 250, + 547, + 273 + ], + "type": "inline_equation", + "content": "B_{3\\mathrm{D}}" + }, + { + "bbox": [ + 46, + 250, + 547, + 273 + ], + "type": "text", + "content": ". During training, the 3D corners of the cuboids are compared against 3D ground truth with the point-cloud chamfer distance." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 286, + 103, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 286, + 103, + 300 + ], + "spans": [ + { + "bbox": [ + 47, + 286, + 103, + 300 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 308, + 290, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 308, + 290, + 392 + ], + "spans": [ + { + "bbox": [ + 46, + 308, + 290, + 392 + ], + "type": "text", + "content": "Our goal is to design a simple and effective model for general 3D object detection. Hence, our approach is free of domain or object specific strategies. We design our 3D object detection framework by extending Faster R-CNN [64] with a 3D object head which predicts a cuboid per each detected 2D object. We refer to our method as Cube R-CNN. Figure 3 shows an overview of our approach." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 402, + 134, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 402, + 134, + 414 + ], + "spans": [ + { + "bbox": [ + 47, + 402, + 134, + 414 + ], + "type": "text", + "content": "4.1. Cube R-CNN" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 422, + 287, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 422, + 287, + 553 + ], + "spans": [ + { + "bbox": [ + 46, + 422, + 287, + 553 + ], + "type": "text", + "content": "Our model builds on Faster R-CNN [64], an end-to-end region-based object detection framework. Faster-RCNN consists of a backbone network, commonly a CNN, which embeds the input image into a higher-dimensional feature space. A region proposal network (RPN) predicts regions of interest (RoIs) representing object candidates in the image. A 2D box head inputs the backbone feature map and processes each RoI to predict a category and a more accurate 2D box. Faster R-CNN can be easily extended to tackle more tasks by adding task-specific heads, e.g. Mask R-CNN [26] adds a mask head to additionally output object silhouettes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 554, + 288, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 554, + 288, + 674 + ], + "spans": [ + { + "bbox": [ + 46, + 554, + 288, + 674 + ], + "type": "text", + "content": "For the task of 3D object detection, we extend Faster R-CNN by introducing a 3D detection head which predicts a 3D cuboid for each detected 2D object. Cube R-CNN extends Faster R-CNN in three ways: (a) we replace the binary classifier in RPN which predicts region objectness with a regressor that predicts IoUness, (b) we introduce a cube head which estimates the parameters to define a 3D cuboid for each detected object (similar in concept to [68]), and (c) we define a new training objective which incorporates a virtual depth for the task of 3D object detection." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 677, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 289, + 715 + ], + "type": "text", + "content": "IoUness. The role of RPN is two-fold: (a) it proposes RoIs by regressing 2D box coordinates from pre-computed anchors and (b) it classifies regions as object or not (object-" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 288, + 547, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 288, + 547, + 469 + ], + "spans": [ + { + "bbox": [ + 304, + 288, + 547, + 469 + ], + "type": "text", + "content": "ness). This is sensible in exhaustively labeled datasets where it can be reliably assessed if a region contains an object. However, OMNI3D combines many data sources with no guarantee that all instances of all classes are labeled. We overcome this by replacing objectness with IoUness, applied only to foreground. Similar to [34], a regressor predicts IoU between a RoI and a ground truth. Let " + }, + { + "bbox": [ + 304, + 288, + 547, + 469 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 304, + 288, + 547, + 469 + ], + "type": "text", + "content": " be the predicted IoU for a RoI and " + }, + { + "bbox": [ + 304, + 288, + 547, + 469 + ], + "type": "inline_equation", + "content": "\\hat{o}" + }, + { + "bbox": [ + 304, + 288, + 547, + 469 + ], + "type": "text", + "content": " be the 2D IoU between the region and its ground truth; we apply a binary cross-entropy (CE) loss " + }, + { + "bbox": [ + 304, + 288, + 547, + 469 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{IoUness}} = \\ell_{\\mathrm{CE}}(o,\\hat{o})" + }, + { + "bbox": [ + 304, + 288, + 547, + 469 + ], + "type": "text", + "content": ". We train on regions whose IoU exceeds 0.05 with a ground truth in order to learn IoUness from a wide range of region overlaps. Thus, the RPN training objective becomes " + }, + { + "bbox": [ + 304, + 288, + 547, + 469 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{RPN}} = \\hat{o}\\cdot (\\mathcal{L}_{\\mathrm{IoUness}} + \\mathcal{L}_{\\mathrm{reg}})" + }, + { + "bbox": [ + 304, + 288, + 547, + 469 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 288, + 547, + 469 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{reg}}" + }, + { + "bbox": [ + 304, + 288, + 547, + 469 + ], + "type": "text", + "content": " is the 2D box regression loss from [64]. The loss is weighted by " + }, + { + "bbox": [ + 304, + 288, + 547, + 469 + ], + "type": "inline_equation", + "content": "\\hat{o}" + }, + { + "bbox": [ + 304, + 288, + 547, + 469 + ], + "type": "text", + "content": " to prioritize candidates close to true objects." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 475, + 547, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 475, + 547, + 571 + ], + "spans": [ + { + "bbox": [ + 304, + 475, + 547, + 571 + ], + "type": "text", + "content": "Cube Head. We extend Faster R-CNN with a new head, called cube head, to predict a 3D cuboid for each detected 2D object. The cube head inputs " + }, + { + "bbox": [ + 304, + 475, + 547, + 571 + ], + "type": "inline_equation", + "content": "7 \\times 7" + }, + { + "bbox": [ + 304, + 475, + 547, + 571 + ], + "type": "text", + "content": " feature maps pooled from the backbone for each predicted region and feeds them to 2 fully-connected (FC) layers with 1024 hidden dimensions. All 3D estimations in the cube head are category-specific. The cube head represents a 3D cuboid with 13 parameters each predicted by a final FC layer:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 317, + 576, + 545, + 672 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 317, + 576, + 545, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 576, + 545, + 599 + ], + "spans": [ + { + "bbox": [ + 317, + 576, + 545, + 599 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 317, + 576, + 545, + 599 + ], + "type": "inline_equation", + "content": "[u, v]" + }, + { + "bbox": [ + 317, + 576, + 545, + 599 + ], + "type": "text", + "content": " represent the projected 3D center on the image plane relative to the 2D RoI" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 317, + 600, + 545, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 600, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 317, + 600, + 545, + 624 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 317, + 600, + 545, + 624 + ], + "type": "inline_equation", + "content": "z \\in \\mathbb{R}_+" + }, + { + "bbox": [ + 317, + 600, + 545, + 624 + ], + "type": "text", + "content": " is the object's center depth in meters transformed from virtual depth " + }, + { + "bbox": [ + 317, + 600, + 545, + 624 + ], + "type": "inline_equation", + "content": "z_v" + }, + { + "bbox": [ + 317, + 600, + 545, + 624 + ], + "type": "text", + "content": " (explained below)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 317, + 624, + 545, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 624, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 317, + 624, + 545, + 647 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 317, + 624, + 545, + 647 + ], + "type": "inline_equation", + "content": "[ \\bar{w}, \\bar{h}, \\bar{l} ] \\in \\mathbb{R}_{+}^{3}" + }, + { + "bbox": [ + 317, + 624, + 545, + 647 + ], + "type": "text", + "content": " are the log-normalized physical box dimensions in meters" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 317, + 647, + 541, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 647, + 541, + 660 + ], + "spans": [ + { + "bbox": [ + 317, + 647, + 541, + 660 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 317, + 647, + 541, + 660 + ], + "type": "inline_equation", + "content": "\\mathbf{p} \\in \\mathbb{R}^6" + }, + { + "bbox": [ + 317, + 647, + 541, + 660 + ], + "type": "text", + "content": " is the continuous 6D [86] allocentric rotation" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 660, + 486, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 660, + 486, + 672 + ], + "spans": [ + { + "bbox": [ + 317, + 660, + 486, + 672 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 317, + 660, + 486, + 672 + ], + "type": "inline_equation", + "content": "\\mu \\in \\mathbb{R}_+" + }, + { + "bbox": [ + 317, + 660, + 486, + 672 + ], + "type": "text", + "content": " is the predicted 3D uncertainty" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "content": "The above parameters form the final 3D box in camera view coordinates for each detected 2D object. The object's 3D center " + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "content": " is estimated from the predicted 2D projected" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13157" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 163, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 163, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 163, + 85 + ], + "type": "text", + "content": "center " + }, + { + "bbox": [ + 47, + 72, + 163, + 85 + ], + "type": "inline_equation", + "content": "[u,v]" + }, + { + "bbox": [ + 47, + 72, + 163, + 85 + ], + "type": "text", + "content": " and depth " + }, + { + "bbox": [ + 47, + 72, + 163, + 85 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 47, + 72, + 163, + 85 + ], + "type": "text", + "content": " via" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 91, + 287, + 118 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 118 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 118 + ], + "type": "interline_equation", + "content": "\\mathbf {X} (u, v, z) = \\left(\\frac {z}{f _ {x}} \\left(r _ {x} + u r _ {w} - p _ {x}\\right), \\frac {z}{f _ {y}} \\left(r _ {y} + v r _ {h} - p _ {y}\\right), z\\right) \\tag {1}", + "image_path": "12a14b29a5ca26f69881ea743b34e688f595662068ac70656801a4c3d6434f3e.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 118, + 289, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 118, + 289, + 192 + ], + "spans": [ + { + "bbox": [ + 47, + 118, + 289, + 192 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 118, + 289, + 192 + ], + "type": "inline_equation", + "content": "[r_x, r_y, r_w, r_h]" + }, + { + "bbox": [ + 47, + 118, + 289, + 192 + ], + "type": "text", + "content": " is the object's 2D box, " + }, + { + "bbox": [ + 47, + 118, + 289, + 192 + ], + "type": "inline_equation", + "content": "(f_x, f_y)" + }, + { + "bbox": [ + 47, + 118, + 289, + 192 + ], + "type": "text", + "content": " are the camera's known focal lengths and " + }, + { + "bbox": [ + 47, + 118, + 289, + 192 + ], + "type": "inline_equation", + "content": "(p_x, p_y)" + }, + { + "bbox": [ + 47, + 118, + 289, + 192 + ], + "type": "text", + "content": " the principal point. The 3D box dimensions " + }, + { + "bbox": [ + 47, + 118, + 289, + 192 + ], + "type": "inline_equation", + "content": "\\mathbf{d}" + }, + { + "bbox": [ + 47, + 118, + 289, + 192 + ], + "type": "text", + "content": " are derived from " + }, + { + "bbox": [ + 47, + 118, + 289, + 192 + ], + "type": "inline_equation", + "content": "[\\bar{w}, \\bar{h}, \\bar{l}]" + }, + { + "bbox": [ + 47, + 118, + 289, + 192 + ], + "type": "text", + "content": " which are log-normalized with category-specific precomputed means " + }, + { + "bbox": [ + 47, + 118, + 289, + 192 + ], + "type": "inline_equation", + "content": "(w_0, h_0, l_0)" + }, + { + "bbox": [ + 47, + 118, + 289, + 192 + ], + "type": "text", + "content": " for width, height and length respectively, and are arranged into a diagonal matrix via" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 197, + 287, + 211 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 197, + 287, + 211 + ], + "spans": [ + { + "bbox": [ + 69, + 197, + 287, + 211 + ], + "type": "interline_equation", + "content": "\\mathbf {d} (\\bar {w}, \\bar {h}, \\bar {l}) = \\operatorname {d i a g} \\left(\\exp (\\bar {w}) w _ {0}, \\exp (\\bar {h}) h _ {0}, \\exp (\\bar {l}) l _ {0}\\right) \\tag {2}", + "image_path": "a060780d3f321a9901dd06f501faf28d62e6e0dab1152b625681e78e4beb34f9.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 217, + 288, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 217, + 288, + 277 + ], + "spans": [ + { + "bbox": [ + 46, + 217, + 288, + 277 + ], + "type": "text", + "content": "Finally, we derive the object's pose " + }, + { + "bbox": [ + 46, + 217, + 288, + 277 + ], + "type": "inline_equation", + "content": "\\mathbf{R}(\\mathbf{p})" + }, + { + "bbox": [ + 46, + 217, + 288, + 277 + ], + "type": "text", + "content": " as a " + }, + { + "bbox": [ + 46, + 217, + 288, + 277 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 46, + 217, + 288, + 277 + ], + "type": "text", + "content": " rotation matrix based on a 6D parameterization (2 directional vectors) of " + }, + { + "bbox": [ + 46, + 217, + 288, + 277 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 46, + 217, + 288, + 277 + ], + "type": "text", + "content": " following [86] which is converted from allocentric to egocentric rotation similar to [39], defined formally in Appendix. The final 3D cuboid, defined by 8 corners, is" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 282, + 287, + 306 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 282, + 287, + 306 + ], + "spans": [ + { + "bbox": [ + 53, + 282, + 287, + 306 + ], + "type": "interline_equation", + "content": "B _ {\\mathrm {3 D}} (u, v, z, \\bar {w}, \\bar {h}, \\bar {l}, \\mathbf {p}) = \\mathbf {R} (\\mathbf {p}) \\mathbf {d} (\\bar {w}, \\bar {h}, \\bar {l}) B _ {\\text {u n i t}} + \\mathbf {X} (u, v, z) \\tag {3}", + "image_path": "091a4c2d3daf01908c59ae690b1e8238dc5ebb26587ebdd0f6b213a2b863e859.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 306, + 288, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 306, + 288, + 367 + ], + "spans": [ + { + "bbox": [ + 46, + 306, + 288, + 367 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 306, + 288, + 367 + ], + "type": "inline_equation", + "content": "B_{\\mathrm{unit}}" + }, + { + "bbox": [ + 46, + 306, + 288, + 367 + ], + "type": "text", + "content": " are the 8 corners of an axis-aligned unit cube centered at " + }, + { + "bbox": [ + 46, + 306, + 288, + 367 + ], + "type": "inline_equation", + "content": "(0,0,0)" + }, + { + "bbox": [ + 46, + 306, + 288, + 367 + ], + "type": "text", + "content": ". Lastly, " + }, + { + "bbox": [ + 46, + 306, + 288, + 367 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 46, + 306, + 288, + 367 + ], + "type": "text", + "content": " denotes a learned 3D uncertainty, which is mapped to a confidence at inference then joined with the classification score " + }, + { + "bbox": [ + 46, + 306, + 288, + 367 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 306, + 288, + 367 + ], + "type": "text", + "content": " from the 2D box head to form the final score for the prediction, as " + }, + { + "bbox": [ + 46, + 306, + 288, + 367 + ], + "type": "inline_equation", + "content": "\\sqrt{s \\cdot \\exp(-\\mu)}" + }, + { + "bbox": [ + 46, + 306, + 288, + 367 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 369, + 289, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 369, + 289, + 548 + ], + "spans": [ + { + "bbox": [ + 46, + 369, + 289, + 548 + ], + "type": "text", + "content": "Training objective. Our training objective consists of 2D losses from the RPN and 2D box head and 3D losses from the cube head. The 3D objective compares each predicted 3D cuboid with its matched ground truth via a chamfer loss, treating the 8 box corners of the 3D boxes as point clouds, namely " + }, + { + "bbox": [ + 46, + 369, + 289, + 548 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{3D}}^{\\mathrm{all}} = \\ell_{\\mathrm{chamfer}}(B_{\\mathrm{3D}}, B_{\\mathrm{3D}}^{\\mathrm{gt}})" + }, + { + "bbox": [ + 46, + 369, + 289, + 548 + ], + "type": "text", + "content": ". Note that " + }, + { + "bbox": [ + 46, + 369, + 289, + 548 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{3D}}^{\\mathrm{all}}" + }, + { + "bbox": [ + 46, + 369, + 289, + 548 + ], + "type": "text", + "content": " entangles all 3D variables via the box predictor " + }, + { + "bbox": [ + 46, + 369, + 289, + 548 + ], + "type": "inline_equation", + "content": "B_{\\mathrm{3D}}" + }, + { + "bbox": [ + 46, + 369, + 289, + 548 + ], + "type": "text", + "content": " (Eq. 3), such that that errors in variables may be ambiguous from one another. Thus, we isolate each variable group with separate disentangled losses, following [68]. The disentangled loss for each variable group substitutes all but its variables with the ground truth from Eq. 3 to create a pseudo box prediction. For example, the disentangled loss for the projected center " + }, + { + "bbox": [ + 46, + 369, + 289, + 548 + ], + "type": "inline_equation", + "content": "[u, v]" + }, + { + "bbox": [ + 46, + 369, + 289, + 548 + ], + "type": "text", + "content": " produces a 3D box with all but " + }, + { + "bbox": [ + 46, + 369, + 289, + 548 + ], + "type": "inline_equation", + "content": "(u, v)" + }, + { + "bbox": [ + 46, + 369, + 289, + 548 + ], + "type": "text", + "content": " replaced with the true values and then compares to the ground truth box," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 59, + 555, + 287, + 573 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 555, + 287, + 573 + ], + "spans": [ + { + "bbox": [ + 59, + 555, + 287, + 573 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {3 D}} ^ {(u, v)} = \\left\\| B _ {\\mathrm {3 D}} (u, v, z ^ {\\mathrm {g t}}, \\bar {w} ^ {\\mathrm {g t}}, \\bar {h} ^ {\\mathrm {g t}}, \\bar {l} ^ {\\mathrm {g t}}, \\mathbf {p} ^ {\\mathrm {g t}}) - B _ {\\mathrm {3 D}} ^ {\\mathrm {g t}} \\right\\| _ {1} \\tag {4}", + "image_path": "bd45b657b74d89b3de16b900b5a3390bb0dd19e36a94c1071ce82ac80ab786d1.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 579, + 287, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 579, + 287, + 681 + ], + "spans": [ + { + "bbox": [ + 46, + 579, + 287, + 681 + ], + "type": "text", + "content": "We use an " + }, + { + "bbox": [ + 46, + 579, + 287, + 681 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 46, + 579, + 287, + 681 + ], + "type": "text", + "content": " loss for " + }, + { + "bbox": [ + 46, + 579, + 287, + 681 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{3D}}^{(u,v)}" + }, + { + "bbox": [ + 46, + 579, + 287, + 681 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 579, + 287, + 681 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{3D}}^{(z)}" + }, + { + "bbox": [ + 46, + 579, + 287, + 681 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 579, + 287, + 681 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{3D}}^{(\\bar{w},\\bar{h},\\bar{l})}" + }, + { + "bbox": [ + 46, + 579, + 287, + 681 + ], + "type": "text", + "content": " and chamfer for " + }, + { + "bbox": [ + 46, + 579, + 287, + 681 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{3D}}^{\\mathbf{P}}" + }, + { + "bbox": [ + 46, + 579, + 287, + 681 + ], + "type": "text", + "content": " to account for cuboid symmetry such that rotation matrices of Euler angles modulo " + }, + { + "bbox": [ + 46, + 579, + 287, + 681 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 46, + 579, + 287, + 681 + ], + "type": "text", + "content": " produce the same noncanonical 3D cuboid. Losses in box coordinates have natural advantages over losses which directly compare variables to ground truths in that their gradients are appropriately weighted by the error as shown in [68]. The 3D objective is defined as, " + }, + { + "bbox": [ + 46, + 579, + 287, + 681 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{3D}} = \\mathcal{L}_{\\mathrm{3D}}^{(u,v)} + \\mathcal{L}_{\\mathrm{3D}}^{(z)} + \\mathcal{L}_{\\mathrm{3D}}^{(\\bar{w},\\bar{h},\\bar{l})} + \\mathcal{L}_{\\mathrm{3D}}^{\\mathbf{P}} + \\mathcal{L}_{\\mathrm{3D}}^{\\mathrm{all}}" + }, + { + "bbox": [ + 46, + 579, + 287, + 681 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 59, + 681, + 270, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 681, + 270, + 693 + ], + "spans": [ + { + "bbox": [ + 59, + 681, + 270, + 693 + ], + "type": "text", + "content": "The collective training objective of Cube R-CNN is," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 75, + 700, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 700, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 75, + 700, + 287, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\mathrm {R P N}} + \\mathcal {L} _ {\\mathrm {2 D}} + \\sqrt {2} \\cdot \\exp (- \\mu) \\cdot \\mathcal {L} _ {\\mathrm {3 D}} + \\mu \\tag {5}", + "image_path": "767af3ea29e1a37bb681149eb3e3a4edfea2fd2a9ec34bd04ac54cf6163395df.jpg" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 306, + 73, + 422, + 205 + ], + "blocks": [ + { + "bbox": [ + 306, + 73, + 422, + 205 + ], + "lines": [ + { + "bbox": [ + 306, + 73, + 422, + 205 + ], + "spans": [ + { + "bbox": [ + 306, + 73, + 422, + 205 + ], + "type": "image", + "image_path": "b994ceceef479499166490ea0d206666179a0b48729edfa20c1aaf24e501295d.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 207, + 547, + 275 + ], + "lines": [ + { + "bbox": [ + 304, + 207, + 547, + 275 + ], + "spans": [ + { + "bbox": [ + 304, + 207, + 547, + 275 + ], + "type": "text", + "content": "Figure 4. Varying camera intrinsics exaggerate scale-depth ambiguity – the same object at two different depths can project to the same image, as in (a) and (b). We address this by introducing a virtual camera which is invariant to intrinsics and transforms the object's depth to a virtual depth " + }, + { + "bbox": [ + 304, + 207, + 547, + 275 + ], + "type": "inline_equation", + "content": "z_v" + }, + { + "bbox": [ + 304, + 207, + 547, + 275 + ], + "type": "text", + "content": " such that the effective image size " + }, + { + "bbox": [ + 304, + 207, + 547, + 275 + ], + "type": "inline_equation", + "content": "H_v" + }, + { + "bbox": [ + 304, + 207, + 547, + 275 + ], + "type": "text", + "content": " and focal length " + }, + { + "bbox": [ + 304, + 207, + 547, + 275 + ], + "type": "inline_equation", + "content": "f_v" + }, + { + "bbox": [ + 304, + 207, + 547, + 275 + ], + "type": "text", + "content": " are consistent across images." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 427, + 74, + 545, + 205 + ], + "blocks": [ + { + "bbox": [ + 427, + 74, + 545, + 205 + ], + "lines": [ + { + "bbox": [ + 427, + 74, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 427, + 74, + 545, + 205 + ], + "type": "image", + "image_path": "2078c047e41eabfaf45470df4b46e07f69028df1b006123e33e713c23eba49d7.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 285, + 547, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 285, + 547, + 358 + ], + "spans": [ + { + "bbox": [ + 304, + 285, + 547, + 358 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{RPN}}" + }, + { + "bbox": [ + 304, + 285, + 547, + 358 + ], + "type": "text", + "content": " is the RPN loss, described above, and " + }, + { + "bbox": [ + 304, + 285, + 547, + 358 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{2\\mathrm{D}}" + }, + { + "bbox": [ + 304, + 285, + 547, + 358 + ], + "type": "text", + "content": " is the 2D box head loss from [64]. The 3D loss is weighted by the predicted 3D uncertainty (inspired by [52]), such that the model may trade a penalty to reduce the 3D loss when uncertain. In practice, " + }, + { + "bbox": [ + 304, + 285, + 547, + 358 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 304, + 285, + 547, + 358 + ], + "type": "text", + "content": " is helpful for both improving the 3D ratings at inference and reducing the loss of hard samples in training." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 369, + 394, + 382 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 369, + 394, + 382 + ], + "spans": [ + { + "bbox": [ + 306, + 369, + 394, + 382 + ], + "type": "text", + "content": "4.2. Virtual Depth" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 388, + 547, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 388, + 547, + 508 + ], + "spans": [ + { + "bbox": [ + 304, + 388, + 547, + 508 + ], + "type": "text", + "content": "A critical part of 3D object detection is predicting an object's depth in metric units. Estimating depth from visual cues requires an implicit mapping of 2D pixels to 3D distances, which is more ambiguous if camera intrinsics vary. Prior works are able to ignore this as they primarily train on images from a single sensor. OMNI3D contains images from many sensors and thus demonstrates large variations in camera intrinsics. We design Cube R-CNN to be robust to intrinsics by predicting the object's virtual depth " + }, + { + "bbox": [ + 304, + 388, + 547, + 508 + ], + "type": "inline_equation", + "content": "z_{v}" + }, + { + "bbox": [ + 304, + 388, + 547, + 508 + ], + "type": "text", + "content": ", which projects the metric depth " + }, + { + "bbox": [ + 304, + 388, + 547, + 508 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 304, + 388, + 547, + 508 + ], + "type": "text", + "content": " to an invariant camera space." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 509, + 547, + 607 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 509, + 547, + 607 + ], + "spans": [ + { + "bbox": [ + 304, + 509, + 547, + 607 + ], + "type": "text", + "content": "Virtual depth scales the depth using the (known) camera intrinsics such that the effective image size and focal length are consistent across the dataset, illustrated in Fig. 4. The effective image properties are referred to as virtual image height " + }, + { + "bbox": [ + 304, + 509, + 547, + 607 + ], + "type": "inline_equation", + "content": "H_{v}" + }, + { + "bbox": [ + 304, + 509, + 547, + 607 + ], + "type": "text", + "content": " and virtual focal length " + }, + { + "bbox": [ + 304, + 509, + 547, + 607 + ], + "type": "inline_equation", + "content": "f_{v}" + }, + { + "bbox": [ + 304, + 509, + 547, + 607 + ], + "type": "text", + "content": ", and are both hyperparameters. If " + }, + { + "bbox": [ + 304, + 509, + 547, + 607 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 304, + 509, + 547, + 607 + ], + "type": "text", + "content": " is the true metric depth of an object from an image with height " + }, + { + "bbox": [ + 304, + 509, + 547, + 607 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 304, + 509, + 547, + 607 + ], + "type": "text", + "content": " and focal length " + }, + { + "bbox": [ + 304, + 509, + 547, + 607 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 304, + 509, + 547, + 607 + ], + "type": "text", + "content": ", its virtual depth is defined as " + }, + { + "bbox": [ + 304, + 509, + 547, + 607 + ], + "type": "inline_equation", + "content": "z_{v} = z \\cdot \\frac{f_{v}}{f} \\frac{H}{H_{v}}" + }, + { + "bbox": [ + 304, + 509, + 547, + 607 + ], + "type": "text", + "content": ". The derivation is in the Appendix." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 606, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 606, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 606, + 547, + 714 + ], + "type": "text", + "content": "Invariance to camera intrinsics via virtual depth also enables scale augmentations during training, since " + }, + { + "bbox": [ + 304, + 606, + 547, + 714 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 304, + 606, + 547, + 714 + ], + "type": "text", + "content": " can vary without harming the consistency of " + }, + { + "bbox": [ + 304, + 606, + 547, + 714 + ], + "type": "inline_equation", + "content": "z_v" + }, + { + "bbox": [ + 304, + 606, + 547, + 714 + ], + "type": "text", + "content": ". Data augmentations from image resizing tend to be critical for 2D models but are not often used in 3D methods [7, 50, 68] since if unaccounted for they increase scale-depth ambiguity. Virtual depth lessens such ambiguities and therefore enables powerful data augmentations in training. We empirically prove the two-fold effects of virtual depth in Sec. 5" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13158" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 138, + 137 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 138, + 137 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 138, + 137 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 138, + 137 + ], + "type": "image", + "image_path": "a1a2d7a345fba5423768e23b6e43263027ad5e429ba8ca2ecb6e0bd1707e2621.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 139, + 70, + 228, + 137 + ], + "blocks": [ + { + "bbox": [ + 139, + 70, + 228, + 137 + ], + "lines": [ + { + "bbox": [ + 139, + 70, + 228, + 137 + ], + "spans": [ + { + "bbox": [ + 139, + 70, + 228, + 137 + ], + "type": "image", + "image_path": "a2dc74297760f80c0033139fd8f2a66ca1c5c049a5b35661ec6607549405f342.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 228, + 70, + 294, + 137 + ], + "blocks": [ + { + "bbox": [ + 228, + 70, + 294, + 137 + ], + "lines": [ + { + "bbox": [ + 228, + 70, + 294, + 137 + ], + "spans": [ + { + "bbox": [ + 228, + 70, + 294, + 137 + ], + "type": "image", + "image_path": "fe6baf6159df0942d2b2c58dc92dcab418722b3568ee670c0348462699d660e8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 294, + 70, + 385, + 137 + ], + "blocks": [ + { + "bbox": [ + 294, + 70, + 385, + 137 + ], + "lines": [ + { + "bbox": [ + 294, + 70, + 385, + 137 + ], + "spans": [ + { + "bbox": [ + 294, + 70, + 385, + 137 + ], + "type": "image", + "image_path": "63865173e449343cb6a235b186fef61c7eb2831b82800b145e648ab7559c8b4a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 385, + 71, + 474, + 137 + ], + "blocks": [ + { + "bbox": [ + 385, + 71, + 474, + 137 + ], + "lines": [ + { + "bbox": [ + 385, + 71, + 474, + 137 + ], + "spans": [ + { + "bbox": [ + 385, + 71, + 474, + 137 + ], + "type": "image", + "image_path": "9b2f18ec20a52ec69421e2c4747a97fa3ff7ac1e6c0e44d907e19ce1840c4000.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 474, + 71, + 541, + 137 + ], + "blocks": [ + { + "bbox": [ + 474, + 71, + 541, + 137 + ], + "lines": [ + { + "bbox": [ + 474, + 71, + 541, + 137 + ], + "spans": [ + { + "bbox": [ + 474, + 71, + 541, + 137 + ], + "type": "image", + "image_path": "994acec229be8a238f2e6ae3de3084ad3eff1edf9f1e7dc9153bdae35573ce04.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 49, + 137, + 138, + 205 + ], + "blocks": [ + { + "bbox": [ + 49, + 137, + 138, + 205 + ], + "lines": [ + { + "bbox": [ + 49, + 137, + 138, + 205 + ], + "spans": [ + { + "bbox": [ + 49, + 137, + 138, + 205 + ], + "type": "image", + "image_path": "1fe5ddef149919d8c5b2512ad32fd572416c68a60e7c3ed18fe48a020a463fdf.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 138, + 137, + 228, + 205 + ], + "blocks": [ + { + "bbox": [ + 138, + 137, + 228, + 205 + ], + "lines": [ + { + "bbox": [ + 138, + 137, + 228, + 205 + ], + "spans": [ + { + "bbox": [ + 138, + 137, + 228, + 205 + ], + "type": "image", + "image_path": "16cf37b188a3a95b8418357b413f7e0c4a64ce5f07b4c152025a192001691b48.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 228, + 137, + 294, + 205 + ], + "blocks": [ + { + "bbox": [ + 228, + 137, + 294, + 205 + ], + "lines": [ + { + "bbox": [ + 228, + 137, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 228, + 137, + 294, + 205 + ], + "type": "image", + "image_path": "7e97d86431114eced98487b0a9f167fd210658c5acd71e0788da9343891eceef.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 294, + 137, + 385, + 205 + ], + "blocks": [ + { + "bbox": [ + 294, + 137, + 385, + 205 + ], + "lines": [ + { + "bbox": [ + 294, + 137, + 385, + 205 + ], + "spans": [ + { + "bbox": [ + 294, + 137, + 385, + 205 + ], + "type": "image", + "image_path": "ed95d112221cbf933ef94b625c3e9c71ae08b6c3a7e52ddfe2ab54a65a4d3904.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 385, + 137, + 475, + 205 + ], + "blocks": [ + { + "bbox": [ + 385, + 137, + 475, + 205 + ], + "lines": [ + { + "bbox": [ + 385, + 137, + 475, + 205 + ], + "spans": [ + { + "bbox": [ + 385, + 137, + 475, + 205 + ], + "type": "image", + "image_path": "c5864368434eca5af31628f837fc8568b705e61a49f0d171a647acd10589eced.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 475, + 137, + 541, + 205 + ], + "blocks": [ + { + "bbox": [ + 475, + 137, + 541, + 205 + ], + "lines": [ + { + "bbox": [ + 475, + 137, + 541, + 205 + ], + "spans": [ + { + "bbox": [ + 475, + 137, + 541, + 205 + ], + "type": "image", + "image_path": "eb5549ccd6cbfcf2fdf7358d68b028373f4e940a742e804d7fc6f9e956e76eb7.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 50, + 205, + 145, + 258 + ], + "blocks": [ + { + "bbox": [ + 50, + 205, + 145, + 258 + ], + "lines": [ + { + "bbox": [ + 50, + 205, + 145, + 258 + ], + "spans": [ + { + "bbox": [ + 50, + 205, + 145, + 258 + ], + "type": "image", + "image_path": "62880ba4b80097575de00d90c051d8a0059a2d7951a901860f40d392abde6529.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 260, + 545, + 282 + ], + "lines": [ + { + "bbox": [ + 46, + 260, + 545, + 282 + ], + "spans": [ + { + "bbox": [ + 46, + 260, + 545, + 282 + ], + "type": "text", + "content": "Figure 5. Cube R-CNN predictions on OMNI3D test. For each example, we show the input image, the 3D predictions overlaid on the image and a top view with a base composed of " + }, + { + "bbox": [ + 46, + 260, + 545, + 282 + ], + "type": "inline_equation", + "content": "1\\mathrm{m}\\times 1\\mathrm{m}" + }, + { + "bbox": [ + 46, + 260, + 545, + 282 + ], + "type": "text", + "content": " tiles. See more results from images and videos in the Appendix." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 145, + 205, + 241, + 258 + ], + "blocks": [ + { + "bbox": [ + 145, + 205, + 241, + 258 + ], + "lines": [ + { + "bbox": [ + 145, + 205, + 241, + 258 + ], + "spans": [ + { + "bbox": [ + 145, + 205, + 241, + 258 + ], + "type": "image", + "image_path": "da91209b692c3de536abce28f727b2b9b52bea4afcb0c168baf45d40b4e3e456.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 242, + 205, + 294, + 258 + ], + "blocks": [ + { + "bbox": [ + 242, + 205, + 294, + 258 + ], + "lines": [ + { + "bbox": [ + 242, + 205, + 294, + 258 + ], + "spans": [ + { + "bbox": [ + 242, + 205, + 294, + 258 + ], + "type": "image", + "image_path": "5c3715f40fd44e583e3284afb2c8c619bd2bda3be2e27605386f9827378e1468.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 294, + 205, + 391, + 258 + ], + "blocks": [ + { + "bbox": [ + 294, + 205, + 391, + 258 + ], + "lines": [ + { + "bbox": [ + 294, + 205, + 391, + 258 + ], + "spans": [ + { + "bbox": [ + 294, + 205, + 391, + 258 + ], + "type": "image", + "image_path": "ad691737e8c087b90e5339938ce38bf914cb8e5308914dd316de4da9443ff3e7.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 391, + 205, + 487, + 258 + ], + "blocks": [ + { + "bbox": [ + 391, + 205, + 487, + 258 + ], + "lines": [ + { + "bbox": [ + 391, + 205, + 487, + 258 + ], + "spans": [ + { + "bbox": [ + 391, + 205, + 487, + 258 + ], + "type": "image", + "image_path": "78af7f71cfa0eb0c9f119b807ebafa86676166b3cf2cc4f2558e7274ebf03fb9.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 487, + 205, + 541, + 258 + ], + "blocks": [ + { + "bbox": [ + 487, + 205, + 541, + 258 + ], + "lines": [ + { + "bbox": [ + 487, + 205, + 541, + 258 + ], + "spans": [ + { + "bbox": [ + 487, + 205, + 541, + 258 + ], + "type": "image", + "image_path": "884996be08dfa5c2cedafd2a7fc444c85bbcb00f1e08495b4d65bf40eab8f572.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 47, + 293, + 127, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 293, + 127, + 306 + ], + "spans": [ + { + "bbox": [ + 47, + 293, + 127, + 306 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 46, + 317, + 288, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 317, + 288, + 389 + ], + "spans": [ + { + "bbox": [ + 46, + 317, + 288, + 389 + ], + "type": "text", + "content": "We tackle image-based 3D object detection on OMNI3D and compare to prior best methods. We evaluate Cube R-CNN on existing 3D benchmarks with a single unified model. Finally, we prove the effectiveness of OMNI3D as a large-scale 3D object detection dataset by showing comprehensive cross-dataset generalization and its impact for pre-training." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 46, + 396, + 288, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 396, + 288, + 515 + ], + "spans": [ + { + "bbox": [ + 46, + 396, + 288, + 515 + ], + "type": "text", + "content": "Implementation details. We implement Cube R-CNN using Detector2 [80] and PyTorch3D [61]. Our backbone is DLA34 [82]-FPN [43] pretrained on ImageNet [67]. We train for 128 epochs with a batch size of 192 images across 48 V100s. We use SGD with a learning rate of 0.12 which decays by a factor of 10 after " + }, + { + "bbox": [ + 46, + 396, + 288, + 515 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 46, + 396, + 288, + 515 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 396, + 288, + 515 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 46, + 396, + 288, + 515 + ], + "type": "text", + "content": " of training. During training, we use random data augmentation of horizontal flipping and scales " + }, + { + "bbox": [ + 46, + 396, + 288, + 515 + ], + "type": "inline_equation", + "content": "\\in [0.50, 1.25]" + }, + { + "bbox": [ + 46, + 396, + 288, + 515 + ], + "type": "text", + "content": ", enabled by virtual depth. Virtual camera parameters are set to " + }, + { + "bbox": [ + 46, + 396, + 288, + 515 + ], + "type": "inline_equation", + "content": "f_v = H_v = 512" + }, + { + "bbox": [ + 46, + 396, + 288, + 515 + ], + "type": "text", + "content": ". Source code and models are publicly available." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 46, + 521, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 289, + 713 + ], + "type": "text", + "content": "Metric. Following popular benchmarks for 2D and 3D recognition, we use average-precision (AP) as our 3D metric. Predictions are matched to ground truth by measuring their overlap using " + }, + { + "bbox": [ + 46, + 521, + 289, + 713 + ], + "type": "inline_equation", + "content": "\\mathsf{I}\\circ \\mathsf{U}_{\\mathsf{3D}}" + }, + { + "bbox": [ + 46, + 521, + 289, + 713 + ], + "type": "text", + "content": " which computes the intersection-overunion of 3D cuboids. We compute a mean " + }, + { + "bbox": [ + 46, + 521, + 289, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{\\mathrm{3D}}" + }, + { + "bbox": [ + 46, + 521, + 289, + 713 + ], + "type": "text", + "content": " across all 50 categories in OMNI3D and over a range of " + }, + { + "bbox": [ + 46, + 521, + 289, + 713 + ], + "type": "inline_equation", + "content": "\\mathsf{I}\\circ \\mathsf{U}_{\\mathsf{3D}}" + }, + { + "bbox": [ + 46, + 521, + 289, + 713 + ], + "type": "text", + "content": " thresholds " + }, + { + "bbox": [ + 46, + 521, + 289, + 713 + ], + "type": "inline_equation", + "content": "\\tau \\in [0.05,0.10,\\dots ,0.50]" + }, + { + "bbox": [ + 46, + 521, + 289, + 713 + ], + "type": "text", + "content": ". The range of " + }, + { + "bbox": [ + 46, + 521, + 289, + 713 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 46, + 521, + 289, + 713 + ], + "type": "text", + "content": " is more relaxed than in 2D to account for the new dimension - see the Appendix for a comparison between 3D and 2D ranges for IoU. General 3D objects may be occluded by other objects or truncated by the image border, and can have arbitrarily small projections. Following [21], we ignore objects with high occlusion (" + }, + { + "bbox": [ + 46, + 521, + 289, + 713 + ], + "type": "inline_equation", + "content": ">66\\%" + }, + { + "bbox": [ + 46, + 521, + 289, + 713 + ], + "type": "text", + "content": ") or truncation (" + }, + { + "bbox": [ + 46, + 521, + 289, + 713 + ], + "type": "inline_equation", + "content": ">33\\%" + }, + { + "bbox": [ + 46, + 521, + 289, + 713 + ], + "type": "text", + "content": "), and tiny projected objects (" + }, + { + "bbox": [ + 46, + 521, + 289, + 713 + ], + "type": "inline_equation", + "content": "< 6.25\\%" + }, + { + "bbox": [ + 46, + 521, + 289, + 713 + ], + "type": "text", + "content": " image height). Moreover, we also report " + }, + { + "bbox": [ + 46, + 521, + 289, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{\\mathrm{3D}}" + }, + { + "bbox": [ + 46, + 521, + 289, + 713 + ], + "type": "text", + "content": " at varying levels of depth " + }, + { + "bbox": [ + 46, + 521, + 289, + 713 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 521, + 289, + 713 + ], + "type": "text", + "content": " as near: " + }, + { + "bbox": [ + 46, + 521, + 289, + 713 + ], + "type": "inline_equation", + "content": "0 < d \\leq 10m" + }, + { + "bbox": [ + 46, + 521, + 289, + 713 + ], + "type": "text", + "content": ", medium: " + }, + { + "bbox": [ + 46, + 521, + 289, + 713 + ], + "type": "inline_equation", + "content": "10m < d \\leq 35m" + }, + { + "bbox": [ + 46, + 521, + 289, + 713 + ], + "type": "text", + "content": ", far: " + }, + { + "bbox": [ + 46, + 521, + 289, + 713 + ], + "type": "inline_equation", + "content": "35m < d \\leq \\infty" + }, + { + "bbox": [ + 46, + 521, + 289, + 713 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 294, + 547, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 294, + 547, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 294, + 547, + 510 + ], + "type": "text", + "content": "Fast " + }, + { + "bbox": [ + 304, + 294, + 547, + 510 + ], + "type": "inline_equation", + "content": "\\mathsf{IOU}_{\\mathbf{3D}}" + }, + { + "bbox": [ + 304, + 294, + 547, + 510 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 304, + 294, + 547, + 510 + ], + "type": "inline_equation", + "content": "\\mathsf{IOU}_{\\mathbf{3D}}" + }, + { + "bbox": [ + 304, + 294, + 547, + 510 + ], + "type": "text", + "content": " compares two 3D cuboids by computing their intersection-over-union. Images usually have many objects and produce several predictions so " + }, + { + "bbox": [ + 304, + 294, + 547, + 510 + ], + "type": "inline_equation", + "content": "\\mathsf{IOU}_{\\mathbf{3D}}" + }, + { + "bbox": [ + 304, + 294, + 547, + 510 + ], + "type": "text", + "content": " computations need to be fast. Prior implementations [21] approximate " + }, + { + "bbox": [ + 304, + 294, + 547, + 510 + ], + "type": "inline_equation", + "content": "\\mathsf{IOU}_{\\mathbf{3D}}" + }, + { + "bbox": [ + 304, + 294, + 547, + 510 + ], + "type": "text", + "content": " by projecting 3D boxes on a ground plane and multiply the top-view 2D intersection with the box heights to compute a 3D volume. Such approximations become notably inaccurate when objects are not on a planar ground or have an arbitrary orientation (e.g. with nonzero pitch or roll). Objectron [2] provides an exact solution, but relies on external libraries [5, 74] implemented in " + }, + { + "bbox": [ + 304, + 294, + 547, + 510 + ], + "type": "inline_equation", + "content": "\\mathbf{C}++" + }, + { + "bbox": [ + 304, + 294, + 547, + 510 + ], + "type": "text", + "content": " and is not batched. We implement a new, fast and exact algorithm which computes the intersecting shape by representing cuboids as meshes and finds face intersections – described in Appendix. Our algorithm is batched with " + }, + { + "bbox": [ + 304, + 294, + 547, + 510 + ], + "type": "inline_equation", + "content": "\\mathbf{C}++" + }, + { + "bbox": [ + 304, + 294, + 547, + 510 + ], + "type": "text", + "content": " and CUDA support. Our implementation is " + }, + { + "bbox": [ + 304, + 294, + 547, + 510 + ], + "type": "inline_equation", + "content": "90\\times" + }, + { + "bbox": [ + 304, + 294, + 547, + 510 + ], + "type": "text", + "content": " faster than Objectron in " + }, + { + "bbox": [ + 304, + 294, + 547, + 510 + ], + "type": "inline_equation", + "content": "\\mathbf{C}++" + }, + { + "bbox": [ + 304, + 294, + 547, + 510 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 294, + 547, + 510 + ], + "type": "inline_equation", + "content": "450\\times" + }, + { + "bbox": [ + 304, + 294, + 547, + 510 + ], + "type": "text", + "content": " faster in CUDA. As a result, evaluation on the large OMNI3D takes only seconds instead of several hours." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 306, + 521, + 421, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 521, + 421, + 533 + ], + "spans": [ + { + "bbox": [ + 306, + 521, + 421, + 533 + ], + "type": "text", + "content": "5.1. Model Performance" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 304, + 541, + 547, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 541, + 547, + 625 + ], + "spans": [ + { + "bbox": [ + 304, + 541, + 547, + 625 + ], + "type": "text", + "content": "First, we ablate the design choices of Cube R-CNN. Then, we compare to state-of-the-art methods on existing benchmarks and OMNI3D and show that it performs superior or on par with prior works which are specialized for their respective domains. Our Cube R-CNN uses a single unified design to tackle general 3D object detection across domains. Fig. 5 shows qualitative predictions on the OMNI3D test set." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 304, + 629, + 547, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 547, + 712 + ], + "type": "text", + "content": "Ablations. Table 1 ablates the features of Cube R-CNN on OMNI3D. We report " + }, + { + "bbox": [ + 304, + 629, + 547, + 712 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{3\\mathrm{D}}" + }, + { + "bbox": [ + 304, + 629, + 547, + 712 + ], + "type": "text", + "content": " performance at single IoU thresholds (0.25 and 0.50), and for near, medium and far objects. We further train and evaluate on domain-specific subsets, OMNI3DIN " + }, + { + "bbox": [ + 304, + 629, + 547, + 712 + ], + "type": "inline_equation", + "content": "(\\mathrm{AP}_{3\\mathrm{D}}^{\\mathrm{IN}})" + }, + { + "bbox": [ + 304, + 629, + 547, + 712 + ], + "type": "text", + "content": " and OMNI3DOUT " + }, + { + "bbox": [ + 304, + 629, + 547, + 712 + ], + "type": "inline_equation", + "content": "(\\mathrm{AP}_{3\\mathrm{D}}^{\\mathrm{OUT}})" + }, + { + "bbox": [ + 304, + 629, + 547, + 712 + ], + "type": "text", + "content": " to show how ablation trends hold for single domain scenarios. From Table 1 we draw a few standout observations:" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13159" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 111, + 68, + 482, + 173 + ], + "blocks": [ + { + "bbox": [ + 111, + 68, + 482, + 173 + ], + "lines": [ + { + "bbox": [ + 111, + 68, + 482, + 173 + ], + "spans": [ + { + "bbox": [ + 111, + 68, + 482, + 173 + ], + "type": "table", + "html": "
Cube R-CNNAP3DAP253DAP503DAPnear3DAPmed3DAPfar3DAPIN3DAPOUT3D
w/o disentangled22.624.39.226.811.88.214.530.9
w/o IoUness22.223.68.826.411.18.314.331.0
w/o L3D20.221.86.426.412.17.413.829.1
w/o scale aug.20.221.58.023.59.86.812.328.1
w/o scale | virtual19.821.27.523.48.65.712.226.0
w/o virtual depth17.318.44.422.77.97.113.230.3
w/o uncertainty μ17.218.35.420.510.87.09.125.4
ours23.324.99.527.912.18.515.031.9
", + "image_path": "fae1a06f10f7cdb927fd70106e23bdde063ae2755514e5f65c3d8b4ad935f682.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 50, + 198, + 291, + 331 + ], + "blocks": [ + { + "bbox": [ + 46, + 174, + 547, + 196 + ], + "lines": [ + { + "bbox": [ + 46, + 174, + 547, + 196 + ], + "spans": [ + { + "bbox": [ + 46, + 174, + 547, + 196 + ], + "type": "text", + "content": "Table 1. Cube R-CNN ablations. We report " + }, + { + "bbox": [ + 46, + 174, + 547, + 196 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{3\\mathrm{D}}" + }, + { + "bbox": [ + 46, + 174, + 547, + 196 + ], + "type": "text", + "content": " on OMNI3D, then at IoU thresholds 0.25 and 0.50, and for near, medium and far objects. We further report " + }, + { + "bbox": [ + 46, + 174, + 547, + 196 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{3\\mathrm{D}}" + }, + { + "bbox": [ + 46, + 174, + 547, + 196 + ], + "type": "text", + "content": " when both training and evaluating on OMNI3DIN and OMNI3DOUT subsets." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 198, + 291, + 331 + ], + "lines": [ + { + "bbox": [ + 50, + 198, + 291, + 331 + ], + "spans": [ + { + "bbox": [ + 50, + 198, + 291, + 331 + ], + "type": "table", + "html": "
MethodfOMNI3DOUTOMNI3D
AP3DAPNU3DAPOUT3DAP3D
M3D-RPN [7]X10.417.913.7-
SMOKE [50]X25.420.419.59.6
FCOS3D [78]X14.620.917.69.8
PGD [77]X21.426.322.911.2
GUPNet [52]✓24.520.519.9-
M3D-RPN +vc✓16.2+5.820.4+2.917.0+3.3-
ImVoxelNet [66]✓23.523.421.59.4
SMOKE +vc✓25.9+0.520.4+0.020.0+0.510.4+0.8
FCOS3D +vc✓17.8+3.225.1+4.221.3+3.710.6+0.8
PGD +vc✓22.8+1.431.2+4.926.8+3.915.4+4.2
Cube R-CNN✓36.032.731.923.3
", + "image_path": "8ad4cd841d71a02c837ea47b991d5118412c25059013b8c5ce73c1b15dedecb2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 46, + 331, + 288, + 376 + ], + "lines": [ + { + "bbox": [ + 46, + 331, + 288, + 376 + ], + "spans": [ + { + "bbox": [ + 46, + 331, + 288, + 376 + ], + "type": "text", + "content": "Table 2. OMNI3D comparison with GUPNet [52], ImVoxelNet [66], M3D-RPN [7], FCOS3D [78], PGD [77], SMOKE [50]. We apply virtual camera to the last four " + }, + { + "bbox": [ + 46, + 331, + 288, + 376 + ], + "type": "inline_equation", + "content": "(+vc)" + }, + { + "bbox": [ + 46, + 331, + 288, + 376 + ], + "type": "text", + "content": " and show the gains in purple. " + }, + { + "bbox": [ + 46, + 331, + 288, + 376 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 331, + 288, + 376 + ], + "type": "text", + "content": " denotes if a method handles variance in focal length." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 388, + 288, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 388, + 288, + 472 + ], + "spans": [ + { + "bbox": [ + 46, + 388, + 288, + 472 + ], + "type": "text", + "content": "Virtual depth is effective and improves " + }, + { + "bbox": [ + 46, + 388, + 288, + 472 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{3\\mathrm{D}}" + }, + { + "bbox": [ + 46, + 388, + 288, + 472 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 46, + 388, + 288, + 472 + ], + "type": "inline_equation", + "content": "+6\\%" + }, + { + "bbox": [ + 46, + 388, + 288, + 472 + ], + "type": "text", + "content": ", most noticeable in the full OMNI3D which has the largest variances in camera intrinsics. Enabled by virtual depth, scale augmentations during training increase " + }, + { + "bbox": [ + 46, + 388, + 288, + 472 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{3\\mathrm{D}}" + }, + { + "bbox": [ + 46, + 388, + 288, + 472 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 46, + 388, + 288, + 472 + ], + "type": "inline_equation", + "content": "+3.1\\%" + }, + { + "bbox": [ + 46, + 388, + 288, + 472 + ], + "type": "text", + "content": " on OMNI3D, " + }, + { + "bbox": [ + 46, + 388, + 288, + 472 + ], + "type": "inline_equation", + "content": "+2.7\\%" + }, + { + "bbox": [ + 46, + 388, + 288, + 472 + ], + "type": "text", + "content": " on OMNI3D" + }, + { + "bbox": [ + 46, + 388, + 288, + 472 + ], + "type": "inline_equation", + "content": "_{\\text{IN}}" + }, + { + "bbox": [ + 46, + 388, + 288, + 472 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 388, + 288, + 472 + ], + "type": "inline_equation", + "content": "+3.8\\%" + }, + { + "bbox": [ + 46, + 388, + 288, + 472 + ], + "type": "text", + "content": " on OMNI3D" + }, + { + "bbox": [ + 46, + 388, + 288, + 472 + ], + "type": "inline_equation", + "content": "_{\\text{OUT}}" + }, + { + "bbox": [ + 46, + 388, + 288, + 472 + ], + "type": "text", + "content": ". We find scale augmentation controlled without virtual depth to be harmful by " + }, + { + "bbox": [ + 46, + 388, + 288, + 472 + ], + "type": "inline_equation", + "content": "-2.5\\%" + }, + { + "bbox": [ + 46, + 388, + 288, + 472 + ], + "type": "text", + "content": " (rows 6 - 7)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 475, + 288, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 475, + 288, + 548 + ], + "spans": [ + { + "bbox": [ + 46, + 475, + 288, + 548 + ], + "type": "text", + "content": "3D uncertainty boosts performance by about " + }, + { + "bbox": [ + 46, + 475, + 288, + 548 + ], + "type": "inline_equation", + "content": "+6\\%" + }, + { + "bbox": [ + 46, + 475, + 288, + 548 + ], + "type": "text", + "content": " on all OMNI3D subsets. Intuitively, it serves as a measure of confidence for the model's 3D predictions; removing it means that the model relies only on its 2D classification to score cuboids. If uncertainty is used only to scale samples in Eq. 5, but not at inference then " + }, + { + "bbox": [ + 46, + 475, + 288, + 548 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{3\\mathrm{D}}" + }, + { + "bbox": [ + 46, + 475, + 288, + 548 + ], + "type": "text", + "content": " still improves but by " + }, + { + "bbox": [ + 46, + 475, + 288, + 548 + ], + "type": "inline_equation", + "content": "+1.6\\%" + }, + { + "bbox": [ + 46, + 475, + 288, + 548 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 548, + 288, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 548, + 288, + 597 + ], + "spans": [ + { + "bbox": [ + 46, + 548, + 288, + 597 + ], + "type": "text", + "content": "Table 1 also shows that the entangled " + }, + { + "bbox": [ + 46, + 548, + 288, + 597 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{3\\mathrm{D}}^{\\mathrm{all}}" + }, + { + "bbox": [ + 46, + 548, + 288, + 597 + ], + "type": "text", + "content": " loss from Eq. 5 boosts " + }, + { + "bbox": [ + 46, + 548, + 288, + 597 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{3\\mathrm{D}}" + }, + { + "bbox": [ + 46, + 548, + 288, + 597 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 46, + 548, + 288, + 597 + ], + "type": "inline_equation", + "content": "+3.1\\%" + }, + { + "bbox": [ + 46, + 548, + 288, + 597 + ], + "type": "text", + "content": " while disentangled losses contribute less " + }, + { + "bbox": [ + 46, + 548, + 288, + 597 + ], + "type": "inline_equation", + "content": "(+0.7\\%)" + }, + { + "bbox": [ + 46, + 548, + 288, + 597 + ], + "type": "text", + "content": ". Replacing objectness with IoUness in the RPN head improves " + }, + { + "bbox": [ + 46, + 548, + 288, + 597 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{3\\mathrm{D}}" + }, + { + "bbox": [ + 46, + 548, + 288, + 597 + ], + "type": "text", + "content": " on OMNI3D by " + }, + { + "bbox": [ + 46, + 548, + 288, + 597 + ], + "type": "inline_equation", + "content": "+1.1\\%" + }, + { + "bbox": [ + 46, + 548, + 288, + 597 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 601, + 288, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 601, + 288, + 661 + ], + "spans": [ + { + "bbox": [ + 46, + 601, + 288, + 661 + ], + "type": "text", + "content": "Comparison to other methods. We compare Cube R-CNN to prior best approaches. We choose representative state-of-the-art methods designed for the urban and indoor domains, and evaluate on our proposed OMNI3D benchmark and single-dataset benchmarks, KITTI and SUN RGB-D." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 665, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 289, + 714 + ], + "type": "text", + "content": "Comparisons on OMNI3D. Table 2 compares Cube R-CNN to M3D-RPN [7], a single-shot 3D anchor approach, FCOS3D [78] and its follow-up PGD [77], which use a fully convolutional one-stage model, GUPNet [52], which uses" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 309, + 198, + 550, + 292 + ], + "blocks": [ + { + "bbox": [ + 309, + 198, + 550, + 292 + ], + "lines": [ + { + "bbox": [ + 309, + 198, + 550, + 292 + ], + "spans": [ + { + "bbox": [ + 309, + 198, + 550, + 292 + ], + "type": "table", + "html": "
MethodDAP703DAP70BEV
EasyMedHardEasyMedHard
SMOKE [50]X14.039.767.8420.8314.4912.75
ImVoxelNet [66]X17.1510.979.1525.1916.3713.58
PGD [77]X19.0511.769.3926.8916.5113.49
GUPNet [52]X22.2615.0213.1230.2921.1918.20
Cube R-CNNX23.5915.0112.5631.7021.2018.43
MonoDTR [30]✓21.9915.3912.7328.5920.3817.14
DD3D [59]✓23.1916.8714.3632.3523.4120.42
", + "image_path": "eaf0d3cb0ed5d6ce271a8a50f06f51debe646a62a590620fb9fbdf595f084e2a.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 293, + 547, + 316 + ], + "lines": [ + { + "bbox": [ + 305, + 293, + 547, + 316 + ], + "spans": [ + { + "bbox": [ + 305, + 293, + 547, + 316 + ], + "type": "text", + "content": "Table 3. KITTI leaderboard results on car for image-based approaches. D denotes methods trained with extra depth supervision." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 327, + 547, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 327, + 547, + 518 + ], + "spans": [ + { + "bbox": [ + 304, + 327, + 547, + 518 + ], + "type": "text", + "content": "2D-3D geometry and camera focal length to derive depth, SMOKE [50] which predicts 3D object center and offsets densely and ImVoxelNet [66], which uses intrinsics to unproject 2D image features to a 3D volume followed by 3D convolutions. These methods originally experiment on urban domains, so we compare on OMNI3DOUT and the full OMNI3D. We train each method using public code and report the best of multiple runs after tuning their hyper-parameters to ensure best performance. We extend all but GUPNet and ImVoxelNet with a virtual camera to handle varying intrinsics, denoted by " + }, + { + "bbox": [ + 304, + 327, + 547, + 518 + ], + "type": "inline_equation", + "content": "+vc" + }, + { + "bbox": [ + 304, + 327, + 547, + 518 + ], + "type": "text", + "content": ". GUPNet and ImVoxelNet dynamically use per-image intrinsics, which can naturally account for camera variation. M3D-RPN and GUPNet do not support multi-node training, which makes scaling to OMNI3D difficult and are omitted. For OMNI3DOUT, we report " + }, + { + "bbox": [ + 304, + 327, + 547, + 518 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{3\\mathrm{D}}^{\\mathrm{OUT}}" + }, + { + "bbox": [ + 304, + 327, + 547, + 518 + ], + "type": "text", + "content": " on its test set and its subparts, " + }, + { + "bbox": [ + 304, + 327, + 547, + 518 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{3\\mathrm{D}}^{\\mathrm{KIT}}" + }, + { + "bbox": [ + 304, + 327, + 547, + 518 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 327, + 547, + 518 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{3\\mathrm{D}}^{\\mathrm{NU}}" + }, + { + "bbox": [ + 304, + 327, + 547, + 518 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 518, + 547, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 518, + 547, + 626 + ], + "spans": [ + { + "bbox": [ + 304, + 518, + 547, + 626 + ], + "type": "text", + "content": "From Table 2 we observe that our Cube R-CNN outperforms competitive methods on both " + }, + { + "bbox": [ + 304, + 518, + 547, + 626 + ], + "type": "inline_equation", + "content": "\\mathrm{OMNI3D_{OUT}}" + }, + { + "bbox": [ + 304, + 518, + 547, + 626 + ], + "type": "text", + "content": " and OMNI3D. The virtual camera extensions result in performance gains of varying degree for all approaches, e.g. " + }, + { + "bbox": [ + 304, + 518, + 547, + 626 + ], + "type": "inline_equation", + "content": "+4.2\\%" + }, + { + "bbox": [ + 304, + 518, + 547, + 626 + ], + "type": "text", + "content": " for PGD on OMNI3D, proving its impact as a general purpose feature for mixed dataset training for the task of 3D object detection. FCOS3D, SMOKE, and ImVoxelNet [66] struggle the most on OMNI3D, which is not surprising as they are typically tailored for dataset-specific model choices." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 547, + 713 + ], + "type": "text", + "content": "Comparisons on KITTI. Table 3 shows results on KITTI's test set using their server [1]. We train Cube R-CNN on KITTI only. Cube R-CNN is a general purpose 3D object detector; its design is not tuned for the KITTI benchmark. Yet, it performs better or on par with recent best methods which are heavily tailored for KITTI and is only slightly worse to models trained with extra depth supervision." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13160" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 53, + 70, + 279, + 123 + ], + "blocks": [ + { + "bbox": [ + 53, + 70, + 279, + 123 + ], + "lines": [ + { + "bbox": [ + 53, + 70, + 279, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 70, + 279, + 123 + ], + "type": "table", + "html": "
MethodTrained onI0U3DAP3D
Total3D [58]SUN RGB-D23.3-
ImVoxelNet [66]SUN RGB-D-30.6
Cube R-CNNSUN RGB-D36.234.7
Cube R-CNNOMNI3DIN37.835.4
", + "image_path": "0101700579859d3a07b5b845b6a241f90684d7647b283f535e2a6e23eb350e4d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 46, + 123, + 288, + 157 + ], + "lines": [ + { + "bbox": [ + 46, + 123, + 288, + 157 + ], + "spans": [ + { + "bbox": [ + 46, + 123, + 288, + 157 + ], + "type": "text", + "content": "Table 4. We compare to indoor models with common categories. For Total3D, we use oracle 2D detections to report " + }, + { + "bbox": [ + 46, + 123, + 288, + 157 + ], + "type": "inline_equation", + "content": "\\mathsf{I}\\circ \\mathsf{U}_{\\mathsf{3D}}" + }, + { + "bbox": [ + 46, + 123, + 288, + 157 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 46, + 123, + 288, + 157 + ], + "type": "inline_equation", + "content": "3^{\\mathrm{rd}}" + }, + { + "bbox": [ + 46, + 123, + 288, + 157 + ], + "type": "text", + "content": " col). For ImVoxelNet, we report the detection metric of " + }, + { + "bbox": [ + 46, + 123, + 288, + 157 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{\\mathrm{3D}}" + }, + { + "bbox": [ + 46, + 123, + 288, + 157 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 46, + 123, + 288, + 157 + ], + "type": "inline_equation", + "content": "4^{\\mathrm{th}}" + }, + { + "bbox": [ + 46, + 123, + 288, + 157 + ], + "type": "text", + "content": " col)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 165, + 288, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 165, + 288, + 274 + ], + "spans": [ + { + "bbox": [ + 46, + 165, + 288, + 274 + ], + "type": "text", + "content": "KITTI's " + }, + { + "bbox": [ + 46, + 165, + 288, + 274 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{3\\mathrm{D}}" + }, + { + "bbox": [ + 46, + 165, + 288, + 274 + ], + "type": "text", + "content": " is fundamentally similar to ours but sets a single " + }, + { + "bbox": [ + 46, + 165, + 288, + 274 + ], + "type": "inline_equation", + "content": "\\mathsf{I}\\circ \\mathsf{U}_{3\\mathrm{D}}" + }, + { + "bbox": [ + 46, + 165, + 288, + 274 + ], + "type": "text", + "content": " threshold at 0.70, which behaves similar to a nearly perfect 2D threshold of 0.94 (see the Appendix). To remedy this, our " + }, + { + "bbox": [ + 46, + 165, + 288, + 274 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{3\\mathrm{D}}" + }, + { + "bbox": [ + 46, + 165, + 288, + 274 + ], + "type": "text", + "content": " is a mean over many thresholds, inspired by COCO [45]. In contrast, nuScenes proposes a mean AP based on 3D center distance but omits size and rotation, with settings generally tailored to the urban domain (e.g. size and rotation variations are less extreme for cars). We provide more analysis with this metric in the Appendix." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 275, + 288, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 275, + 288, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 275, + 288, + 456 + ], + "type": "text", + "content": "Comparisons on SUN RGB-D. Table 4 compares Cube R-CNN to Total3D [58] and ImVoxelNet [66], two state-of-the-art methods on SUN RGB-D on 10 common categories. Total3D's public model requires 2D object boxes and categories as input. Hence, for a fair comparison, we use ground truth 2D detections as input to both our and their method, each trained using a ResNet34 [27] backbone and report mean IoU3D (3rd col). We compare to ImVoxelNet in the full 3D object detection setting and report AP3D (4th col). We compare Cube R-CNN from two respects, first trained on SUN RGB-D identical to the baselines, then trained on OMNI3DIN which subsumes SUN RGB-D. Our model outperforms Total3D by +12.9% and ImVoxelNet by +4.1% when trained on the same training set and increases the performance gap when trained on the larger OMNI3DIN." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 458, + 288, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 458, + 288, + 541 + ], + "spans": [ + { + "bbox": [ + 46, + 458, + 288, + 541 + ], + "type": "text", + "content": "Zero-shot Performance. One commonly sought after property of detectors is zero-shot generalization to other datasets. Cube R-CNN outperforms all other methods at zero-shot. We find our model trained on KITTI achieves " + }, + { + "bbox": [ + 46, + 458, + 288, + 541 + ], + "type": "inline_equation", + "content": "12.7\\%" + }, + { + "bbox": [ + 46, + 458, + 288, + 541 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{3\\mathrm{D}}" + }, + { + "bbox": [ + 46, + 458, + 288, + 541 + ], + "type": "text", + "content": " on nuScenes; the second best is M3D-RPN+vc with " + }, + { + "bbox": [ + 46, + 458, + 288, + 541 + ], + "type": "inline_equation", + "content": "10.7\\%" + }, + { + "bbox": [ + 46, + 458, + 288, + 541 + ], + "type": "text", + "content": ". Conversely, when training on nuScenes, our model achieves " + }, + { + "bbox": [ + 46, + 458, + 288, + 541 + ], + "type": "inline_equation", + "content": "20.2\\%" + }, + { + "bbox": [ + 46, + 458, + 288, + 541 + ], + "type": "text", + "content": " on KITTI; the second best is GUPNet with " + }, + { + "bbox": [ + 46, + 458, + 288, + 541 + ], + "type": "inline_equation", + "content": "17.3\\%" + }, + { + "bbox": [ + 46, + 458, + 288, + 541 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 548, + 251, + 561 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 548, + 251, + 561 + ], + "spans": [ + { + "bbox": [ + 47, + 548, + 251, + 561 + ], + "type": "text", + "content": "5.2. The Impact of the OMNI3D Benchmark" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 567, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 567, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 567, + 287, + 628 + ], + "type": "text", + "content": "Sec. 5.1 analyzes the performance of Cube R-CNN for the task of 3D object detection. Now, we turn to OMNI3D and its impact as a large-scale benchmark. We show two use cases of OMNI3D: (a) a universal 3D dataset which integrates smaller ones, and (b) a pre-training dataset." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "content": "OMNI3D as a universal dataset. We treat OMNI3D as a dataset which integrates smaller single ones and show its impact on each one. We train Cube R-CNN on OMNI3D and compare to single-dataset training in Table 5, for the indoor (left) and urban (right) domain. " + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{\\mathrm{3D}}" + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "content": " is reported on the category intersection (10 for indoor, 3 for outdoor) to ensure comparability. Training on OMNI3D and its domain-specific" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 309, + 69, + 541, + 133 + ], + "blocks": [ + { + "bbox": [ + 309, + 69, + 541, + 133 + ], + "lines": [ + { + "bbox": [ + 309, + 69, + 541, + 133 + ], + "spans": [ + { + "bbox": [ + 309, + 69, + 541, + 133 + ], + "type": "table", + "html": "
Trained onAPH3DAPSUN3DAPAR3DTrained onAPKIT3DAPNU3D
Hypersim15.29.57.5KITTI37.112.7
SUN5.834.713.1nuScenes20.238.6
ARKit5.914.238.6OMNI3D37.835.8
OMNI3D19.032.638.2OMNI3DOUT42.439.0
OMNI3DIN17.835.441.2
", + "image_path": "e0cc40fbdf096e9aa4486f6a78d3a405aaa34853c938ff167809c965102a0bf1.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 305, + 133, + 545, + 155 + ], + "lines": [ + { + "bbox": [ + 305, + 133, + 545, + 155 + ], + "spans": [ + { + "bbox": [ + 305, + 133, + 545, + 155 + ], + "type": "text", + "content": "Table 5. Cross-dataset performance on intersecting categories (10 for indoor, 3 for outdoor) for comparable cross-evaluations." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 326, + 163, + 523, + 273 + ], + "blocks": [ + { + "bbox": [ + 326, + 163, + 523, + 273 + ], + "lines": [ + { + "bbox": [ + 326, + 163, + 523, + 273 + ], + "spans": [ + { + "bbox": [ + 326, + 163, + 523, + 273 + ], + "type": "image", + "image_path": "8fe75a37d19012db8085bce4fd4c08e79478b6133bc6adfa3132a832bcfe0cc7.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 333, + 276, + 517, + 288 + ], + "lines": [ + { + "bbox": [ + 333, + 276, + 517, + 288 + ], + "spans": [ + { + "bbox": [ + 333, + 276, + 517, + 288 + ], + "type": "text", + "content": "Figure 6. Pre-training on OMNI3D- vs. ImageNet." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 297, + 547, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 297, + 547, + 381 + ], + "spans": [ + { + "bbox": [ + 304, + 297, + 547, + 381 + ], + "type": "text", + "content": "subsets, " + }, + { + "bbox": [ + 304, + 297, + 547, + 381 + ], + "type": "inline_equation", + "content": "\\mathrm{OMN13D_{IN}}" + }, + { + "bbox": [ + 304, + 297, + 547, + 381 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 297, + 547, + 381 + ], + "type": "inline_equation", + "content": "\\mathrm{OMN13D_{OUT}}" + }, + { + "bbox": [ + 304, + 297, + 547, + 381 + ], + "type": "text", + "content": ", results in higher performance compared to single-dataset training, signifying that our large OMN13D generalizes better and should be preferred over single dataset training. ARKit sees a " + }, + { + "bbox": [ + 304, + 297, + 547, + 381 + ], + "type": "inline_equation", + "content": "+2.6\\%" + }, + { + "bbox": [ + 304, + 297, + 547, + 381 + ], + "type": "text", + "content": " boost and KITTI " + }, + { + "bbox": [ + 304, + 297, + 547, + 381 + ], + "type": "inline_equation", + "content": "+5.3\\%" + }, + { + "bbox": [ + 304, + 297, + 547, + 381 + ], + "type": "text", + "content": ". Except for Hypersim, the domain-specific subsets tend to perform better on their domain, which is not surprising given their distinct properties (Fig. 2)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 384, + 547, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 384, + 547, + 564 + ], + "spans": [ + { + "bbox": [ + 304, + 384, + 547, + 564 + ], + "type": "text", + "content": "OMNI3D as a pre-training dataset. Next, we demonstrate the utility of OMNI3D for pre-training. In this setting, an unseen dataset is used for finetuning from a model pre-trained on OMNI3D. The motivation is to determine how a large-scale 3D dataset could accelerate low-shot learning with minimum need for costly 3D annotations on a new dataset. We choose SUN RGB-D and KITTI as our unseen datasets given their popularity and small size. We pre-train Cube R-CNN on OMNI3D-, which removes them from OMNI3D, then finetune the models using a " + }, + { + "bbox": [ + 304, + 384, + 547, + 564 + ], + "type": "inline_equation", + "content": "\\%" + }, + { + "bbox": [ + 304, + 384, + 547, + 564 + ], + "type": "text", + "content": " of their training data. The curves in Fig. 6 show the model quickly gains its upper-bound performance at a small fraction of the training data, when pre-trained on OMNI3D- vs. ImageNet, without any few-shot training tricks. A model finetuned on only " + }, + { + "bbox": [ + 304, + 384, + 547, + 564 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 304, + 384, + 547, + 564 + ], + "type": "text", + "content": " of its target can achieve " + }, + { + "bbox": [ + 304, + 384, + 547, + 564 + ], + "type": "inline_equation", + "content": ">70\\%" + }, + { + "bbox": [ + 304, + 384, + 547, + 564 + ], + "type": "text", + "content": " of the upper-bound performance." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 575, + 378, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 575, + 378, + 588 + ], + "spans": [ + { + "bbox": [ + 306, + 575, + 378, + 588 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": "We propose a large and diverse 3D object detection benchmark, OMNI3D, and a general purpose 3D object detector, Cube R-CNN. Models and data are publicly released. Our extensive analysis (Table 1, Fig. 5, Appendix) show the strength of general 3D object detection as well as the limitations of our method, e.g. localizing far away objects, and uncommon object types or contexts. Moreover, for accurate real-world 3D predictions, our method is limited by the assumption of known camera intrinsics, which may be minimized in future work using self-calibration techniques." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13161" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 91, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 288, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 288, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 288, + 123 + ], + "type": "text", + "content": "[1] KITTI server. https://www.cvlibs.net/datasets/kitti/eval_object.php?obj_benchmark=3d. Accessed: 2022-11-01. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 124, + 288, + 167 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 124, + 288, + 167 + ], + "spans": [ + { + "bbox": [ + 53, + 124, + 288, + 167 + ], + "type": "text", + "content": "[2] Adel Ahmadyan, Liangkai Zhang, Artsiom Ablavatski, Jianing Wei, and Matthias Grundmann. Objectron: A large scale dataset of object-centric videos in the wild with pose annotations. CVPR, 2021. 2, 3, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 168, + 288, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 168, + 288, + 224 + ], + "spans": [ + { + "bbox": [ + 53, + 168, + 288, + 224 + ], + "type": "text", + "content": "[3] Armen Avetisyan, Manuel Dahnert, Angela Dai, Manolis Savva, Angel X Chang, and Matthias Nießner. Scan2CAD: Learning CAD model alignment in RGB-D scans. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 2614-2623, 2019. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 224, + 288, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 224, + 288, + 257 + ], + "spans": [ + { + "bbox": [ + 53, + 224, + 288, + 257 + ], + "type": "text", + "content": "[4] Wentao Bao, Bin Xu, and Zhenzhong Chen. MonoFENet: Monocular 3D object detection with feature enhancement networks. TIP, 2019. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 258, + 288, + 291 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 258, + 288, + 291 + ], + "spans": [ + { + "bbox": [ + 53, + 258, + 288, + 291 + ], + "type": "text", + "content": "[5] Bradford Barber, David Dobkin, and Hannu Huhdanpaa. The quickhull algorithm for convex hulls. ACM Transactions on Mathematical Software (TOMS), 1996. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 291, + 288, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 291, + 288, + 357 + ], + "spans": [ + { + "bbox": [ + 53, + 291, + 288, + 357 + ], + "type": "text", + "content": "[6] Gilad Baruch, Zhuoyuan Chen, Afshin Dehghan, Tal Dimry, Yuri Feigin, Peter Fu, Thomas Gebauer, Brandon Joffe, Daniel Kurz, Arik Schwartz, and Elad Shulman. ARKitScenes - a diverse real-world dataset for 3D indoor scene understanding using mobile RGB-D data. In NeurIPS Datasets and Benchmarks Track, 2021. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 357, + 288, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 357, + 288, + 390 + ], + "spans": [ + { + "bbox": [ + 53, + 357, + 288, + 390 + ], + "type": "text", + "content": "[7] Garrick Brazil and Xiaoming Liu. M3D-RPN: Monocular 3D region proposal network for object detection. In ICCV, 2019. 1, 2, 5, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 391, + 288, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 391, + 288, + 424 + ], + "spans": [ + { + "bbox": [ + 53, + 391, + 288, + 424 + ], + "type": "text", + "content": "[8] Garrick Brazil, Gerard Pons-Moll, Xiaoming Liu, and Bernt Schiele. Kinematic 3D object detection in monocular video. In ECCV, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 424, + 288, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 424, + 288, + 468 + ], + "spans": [ + { + "bbox": [ + 53, + 424, + 288, + 468 + ], + "type": "text", + "content": "[9] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuScenes: A multimodal dataset for autonomous driving. In CVPR, 2020. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 469, + 288, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 469, + 288, + 514 + ], + "spans": [ + { + "bbox": [ + 47, + 469, + 288, + 514 + ], + "type": "text", + "content": "[10] Ming-Fang Chang, John Lambert, Patsorn Sangkloy, Jagjeet Singh, Slawomir Bak, Andrew Hartnett, De Wang, Peter Carr, Simon Lucey, Deva Ramanan, et al. Argoverse: 3D tracking and forecasting with rich maps. In CVPR, 2019. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 514, + 288, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 514, + 288, + 557 + ], + "spans": [ + { + "bbox": [ + 47, + 514, + 288, + 557 + ], + "type": "text", + "content": "[11] Hansheng Chen, Yuyao Huang, Wei Tian, Zhong Gao, and Lu Xiong. MonoRUn: Monocular 3D object detection by reconstruction and uncertainty propagation. In CVPR, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 558, + 288, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 558, + 288, + 646 + ], + "spans": [ + { + "bbox": [ + 47, + 558, + 288, + 646 + ], + "type": "text", + "content": "[12] Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jiarui Xu, Zheng Zhang, Dazhi Cheng, Chenchen Zhu, Tianheng Cheng, Qijie Zhao, Buyu Li, Xin Lu, Rui Zhu, Yue Wu, Jifeng Dai, Jingdong Wang, Jianping Shi, Wanli Ouyang, Chen Change Loy, and Dahua Lin. MMDetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155, 2019. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 646, + 288, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 646, + 288, + 680 + ], + "spans": [ + { + "bbox": [ + 47, + 646, + 288, + 680 + ], + "type": "text", + "content": "[13] Xiaozhi Chen, Kaustav Kundu, Ziyu Zhang, Huimin Ma, Sanja Fidler, and Raquel Urtasun. Monocular 3D object detection for autonomous driving. In CVPR, 2016. 1, 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 47, + 680, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 680, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 680, + 288, + 714 + ], + "type": "text", + "content": "[14] Yongjian Chen, Lei Tai, Kai Sun, and Mingyang Li. MonoPair: Monocular 3D object detection using pairwise spatial relationships. In CVPR, 2020. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 307, + 73, + 546, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 546, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 546, + 106 + ], + "type": "text", + "content": "[15] Yi-Nan Chen, Hang Dai, and Yong Ding. Pseudo-stereo for monocular 3D object detection in autonomous driving. In CVPR, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 106, + 547, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 106, + 547, + 149 + ], + "spans": [ + { + "bbox": [ + 307, + 106, + 547, + 149 + ], + "type": "text", + "content": "[16] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. ScanNet: Richly-annotated 3D reconstructions of indoor scenes. In CVPR, 2017. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 149, + 547, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 149, + 547, + 183 + ], + "spans": [ + { + "bbox": [ + 307, + 149, + 547, + 183 + ], + "type": "text", + "content": "[17] Saumitro Dasgupta, Kuan Fang, Kevin Chen, and Silvio Savarese. Delay: Robust spatial layout estimation for cluttered indoor scenes. In CVPR, 2016. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 183, + 547, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 183, + 547, + 226 + ], + "spans": [ + { + "bbox": [ + 307, + 183, + 547, + 226 + ], + "type": "text", + "content": "[18] Mingyu Ding, Yuqi Huo, Hongwei Yi, Zhe Wang, Jianping Shi, Zhiwu Lu, and Ping Luo. Learning depth-guided convolutions for monocular 3D object detection. In CVPRW, 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 226, + 546, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 226, + 546, + 260 + ], + "spans": [ + { + "bbox": [ + 307, + 226, + 546, + 260 + ], + "type": "text", + "content": "[19] Huan Fu, Mingming Gong, Chaohui Wang, Kayhan Bat-manghelich, and Dacheng Tao. Deep Ordinal Regression Network for Monocular Depth Estimation. In CVPR, 2018. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 260, + 546, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 260, + 546, + 293 + ], + "spans": [ + { + "bbox": [ + 307, + 260, + 546, + 293 + ], + "type": "text", + "content": "[20] Nils Gählert, Nicolas Jourdan, Marius Cordts, Uwe Franke, and Joachim Denzler. Cityscapes 3D: Dataset and benchmark for 9 DOF vehicle detection. In CVPRW, 2020. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 293, + 546, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 293, + 546, + 327 + ], + "spans": [ + { + "bbox": [ + 307, + 293, + 546, + 327 + ], + "type": "text", + "content": "[21] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the KITTI vision benchmark suite. In CVPR, 2012. 1, 2, 3, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 327, + 546, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 327, + 546, + 348 + ], + "spans": [ + { + "bbox": [ + 307, + 327, + 546, + 348 + ], + "type": "text", + "content": "[22] Georgia Gkioxari, Jitendra Malik, and Justin Johnson. Mesh R-CNN. In ICCV, 2019. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 348, + 546, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 348, + 546, + 382 + ], + "spans": [ + { + "bbox": [ + 307, + 348, + 546, + 382 + ], + "type": "text", + "content": "[23] Jiaqi Gu, Bojian Wu, Lubin Fan, Jianqiang Huang, Shen Cao, Zhiyu Xiang, and Xian-Sheng Hua. Homography loss for monocular 3D object detection. In CVPR, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 382, + 546, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 382, + 546, + 415 + ], + "spans": [ + { + "bbox": [ + 307, + 382, + 546, + 415 + ], + "type": "text", + "content": "[24] Agrim Gupta, Piotr Dollar, and Ross Girshick. LVIS: A dataset for large vocabulary instance segmentation. In CVPR, 2019. 1, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 415, + 546, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 415, + 546, + 459 + ], + "spans": [ + { + "bbox": [ + 307, + 415, + 546, + 459 + ], + "type": "text", + "content": "[25] Chenhang He, Jianqiang Huang, Xian-Sheng Hua, and Lei Zhang. Aug3D-RPN: Improving monocular 3D object detection by synthetic images with virtual depth. arXiv preprint arXiv:2107.13269, 2021. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 459, + 546, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 459, + 546, + 481 + ], + "spans": [ + { + "bbox": [ + 307, + 459, + 546, + 481 + ], + "type": "text", + "content": "[26] Kaiming He, Georgia Gkioxari, Piotr Dollár, and Ross Girshick. Mask R-CNN. In ICCV, 2017. 1, 2, 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 481, + 546, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 481, + 546, + 514 + ], + "spans": [ + { + "bbox": [ + 307, + 481, + 546, + 514 + ], + "type": "text", + "content": "[27] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 8" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 514, + 546, + 537 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 514, + 546, + 537 + ], + "spans": [ + { + "bbox": [ + 307, + 514, + 546, + 537 + ], + "type": "text", + "content": "[28] Varsha Hedau, Derek Hoiem, and David Forsyth. Recovering the spatial layout of cluttered rooms. In ICCV, 2009. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 537, + 546, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 537, + 546, + 591 + ], + "spans": [ + { + "bbox": [ + 307, + 537, + 546, + 591 + ], + "type": "text", + "content": "[29] John Houston, Guido Zuidhof, Luca Bergamini, Yawei Ye, Long Chen, Ashesh Jain, Sammy Omari, Vladimir Iglovikov, and Peter Ondruska. One thousand and one hours: Self-driving motion prediction dataset. arXiv preprint arXiv:2006.14480, 2020. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 591, + 546, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 591, + 546, + 624 + ], + "spans": [ + { + "bbox": [ + 307, + 591, + 546, + 624 + ], + "type": "text", + "content": "[30] Kuan-Chih Huang, Tsung-Han Wu, Hung-Ting Su, and Winston Hsu. MonoDTR: Monocular 3D object detection with depth-aware transformer. In CVPR, 2022. 2, 7" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 624, + 546, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 546, + 669 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 546, + 669 + ], + "type": "text", + "content": "[31] Siyuan Huang, Siyuan Qi, Yinxue Xiao, Yixin Zhu, Ying Nian Wu, and Song-Chun Zhu. Cooperative holistic scene understanding: Unifying 3D object, layout, and camera pose estimation. In NeurIPS, 2018. 1, 2" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 669, + 546, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 546, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 546, + 712 + ], + "type": "text", + "content": "[32] Xinyu Huang, Peng Wang, Xinjing Cheng, Dingfu Zhou, Qichuan Geng, and Ruigang Yang. The apolloscape open dataset for autonomous driving and its application. TPAMI, 2019. 2" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13162" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "type": "text", + "content": "[33] R. Kesten, M. Usman, J. Houston, T. Pandya, K. Nadhamuni, A. Ferreira, M. Yuan, B. Low, A. Jain, P. Ondruska, S. Omari, S. Shah, A. Kulkarni, A. Kazakova, C. Tao, L. Platinsky, W. Jiang, and V. Shet. Level 5 perception dataset 2020. https://level-5.global/level5/data/, 2019.2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 288, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 288, + 161 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 288, + 161 + ], + "type": "text", + "content": "[34] Dahun Kim, Tsung-Yi Lin, Anelia Angelova, In Kweon, and Weicheng Kuo. Learning open-world object proposals without learning to classify. RAL, 2022. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 288, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 288, + 195 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 288, + 195 + ], + "type": "text", + "content": "[35] Jason Ku, Melissa Mozifian, Jungwook Lee, Ali Harakeh, and Steven Waslander. Joint 3D proposal generation and object detection from view aggregation. In IROS, 2018. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 197, + 288, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 197, + 288, + 229 + ], + "spans": [ + { + "bbox": [ + 48, + 197, + 288, + 229 + ], + "type": "text", + "content": "[36] Nilesh Kulkarni, Ishan Misra, Shubham Tulsiani, and Abhinav Gupta. 3D-RelNet: Joint object and relational network for 3D prediction. In ICCV, 2019. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 230, + 288, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 230, + 288, + 273 + ], + "spans": [ + { + "bbox": [ + 48, + 230, + 288, + 273 + ], + "type": "text", + "content": "[37] Abhinav Kumar, Garrick Brazil, Enrique Corona, Armin Parchami, and Xiaoming Liu. DEVIANT: Depth EquiVarIAnt NeTwork for monocular 3D object detection. In ECCV, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 274, + 288, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 274, + 288, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 274, + 288, + 308 + ], + "type": "text", + "content": "[38] Abhinav Kumar, Garrick Brazil, and Xiaoming Liu. GrooMeD-NMS: Grouped mathematically differentiable NMS for monocular 3D object detection. In CVPR, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 308, + 288, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 308, + 288, + 341 + ], + "spans": [ + { + "bbox": [ + 48, + 308, + 288, + 341 + ], + "type": "text", + "content": "[39] Abhijit Kundu, Yin Li, and James Rehg. 3D-RCNN: Instance-level 3D object reconstruction via render-and-compare. In CVPR, 2018. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 342, + 288, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 342, + 288, + 365 + ], + "spans": [ + { + "bbox": [ + 48, + 342, + 288, + 365 + ], + "type": "text", + "content": "[40] Hei Law and Jia Deng. CornerNet: Detecting objects as paired keypoints. In ECCV, 2018. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 365, + 288, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 365, + 288, + 397 + ], + "spans": [ + { + "bbox": [ + 48, + 365, + 288, + 397 + ], + "type": "text", + "content": "[41] David C Lee, Martial Hebert, and Takeo Kanade. Geometric reasoning for single image structure recovery. In CVPR, 2009. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 399, + 288, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 288, + 432 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 288, + 432 + ], + "type": "text", + "content": "[42] Peixuan Li, Huaici Zhao, Pengfei Liu, and Feidao Cao. RTM3D: Real-time monocular 3D detection from object keypoints for autonomous driving. In ECCV, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 433, + 288, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 433, + 288, + 465 + ], + "spans": [ + { + "bbox": [ + 48, + 433, + 288, + 465 + ], + "type": "text", + "content": "[43] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In CVPR, 2017. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 467, + 288, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 467, + 288, + 498 + ], + "spans": [ + { + "bbox": [ + 48, + 467, + 288, + 498 + ], + "type": "text", + "content": "[44] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In ICCV, 2017. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 500, + 288, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 500, + 288, + 544 + ], + "spans": [ + { + "bbox": [ + 48, + 500, + 288, + 544 + ], + "type": "text", + "content": "[45] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dolkar, and C Lawrence Zitnick. Microsoft COCO: Common objects in context. In ECCV, 2014. 1, 3, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 545, + 288, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 545, + 288, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 545, + 288, + 578 + ], + "type": "text", + "content": "[46] Ce Liu, Shuhang Gu, Luc Van Gool, and Radu Timofte. Deep line encoding for monocular 3D object detection and depth prediction. In BMVC, 2021. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 579, + 288, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 579, + 288, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 579, + 288, + 612 + ], + "type": "text", + "content": "[47] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander Berg. SSD: Single shot multibox detector. In ECCV, 2016. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 613, + 288, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 288, + 644 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 288, + 644 + ], + "type": "text", + "content": "[48] Xianpeng Liu, Nan Xue, and Tianfu Wu. Learning auxiliary monocular contexts helps monocular 3D object detection. In AAAI, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 647, + 288, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 288, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 288, + 678 + ], + "type": "text", + "content": "[49] Yuxuan Liu, Yuan Yixuan, and Ming Liu. Ground-aware monocular 3D object detection for autonomous driving. RAL, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 680, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 288, + 712 + ], + "type": "text", + "content": "[50] Zechen Liu, Zizhang Wu, and Roland Tóth. SMOKE: Single-stage monocular 3D object detection via keypoint estimation. In CVPRW, 2020. 1, 2, 5, 7" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 546, + 713 + ], + "type": "list", + "angle": 0, + "index": 36, + "blocks": [ + { + "bbox": [ + 307, + 73, + 546, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 546, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 546, + 106 + ], + "type": "text", + "content": "[51] Zongdai Liu, Dingfu Zhou, Feixiang Lu, Jin Fang, and Liangjun Zhang. Autoshape: Real-time shape-aware monocular 3D object detection. In ICCV, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 107, + 546, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 546, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 546, + 150 + ], + "type": "text", + "content": "[52] Yan Lu, Xinzhu Ma, Lei Yang, Tianzhu Zhang, Yating Liu, Qi Chu, Junjie Yan, and Wanli Ouyang. Geometry uncertainty projection network for monocular 3D object detection. In ICCV, 2021. 1, 2, 5, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 152, + 546, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 546, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 546, + 184 + ], + "type": "text", + "content": "[53] Xinzhu Ma, Shinan Liu, Zhiyi Xia, Hongwen Zhang, Xingyu Zeng, and Wanli Ouyang. Rethinking pseudo-lidar representation. In ECCV, 2020. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 186, + 546, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 546, + 229 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 546, + 229 + ], + "type": "text", + "content": "[54] Xinzhu Ma, Zhihui Wang, Haojie Li, Pengbo Zhang, Wanli Ouyang, and Xin Fan. Accurate monocular 3D object detection via color-embedded 3D reconstruction for autonomous driving. In ICCV, 2019. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 231, + 546, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 231, + 546, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 231, + 546, + 262 + ], + "type": "text", + "content": "[55] Xinzhu Ma, Yinmin Zhang, Dan Xu, Dongzhan Zhou, Shuai Yi, Haojie Li, and Wanli Ouyang. Delving into localization errors for monocular 3D object detection. In CVPR, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 264, + 546, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 264, + 546, + 295 + ], + "spans": [ + { + "bbox": [ + 307, + 264, + 546, + 295 + ], + "type": "text", + "content": "[56] Arun Mallya and Svetlana Lazebnik. Learning informative edge maps for indoor scene layout prediction. In ICCV, 2015. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 297, + 546, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 297, + 546, + 331 + ], + "spans": [ + { + "bbox": [ + 307, + 297, + 546, + 331 + ], + "type": "text", + "content": "[57] Arsalan Mousavian, Dragomir Anguelov, John Flynn, and Jana Kosecka. 3D bounding box estimation using deep learning and geometry. In CVPR, 2017. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 332, + 546, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 332, + 546, + 376 + ], + "spans": [ + { + "bbox": [ + 307, + 332, + 546, + 376 + ], + "type": "text", + "content": "[58] Yinyu Nie, Xiaoguang Han, Shihui Guo, Yujuan Zheng, Jian Chang, and Jian Zhang. Total3DUnderstanding: Joint layout, object pose and mesh reconstruction for indoor scenes from a single image. In CVPR, 2020. 1, 2, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 376, + 546, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 376, + 546, + 409 + ], + "spans": [ + { + "bbox": [ + 307, + 376, + 546, + 409 + ], + "type": "text", + "content": "[59] Dennis Park, Rares Ambrus, Vitor Guizilini, Jie Li, and Adrien Gaidon. Is pseudo-lidar needed for monocular 3D object detection? In ICCV, 2021. 2, 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 411, + 546, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 411, + 546, + 443 + ], + "spans": [ + { + "bbox": [ + 307, + 411, + 546, + 443 + ], + "type": "text", + "content": "[60] Charles Qi, Wei Liu, Chenxia Wu, Hao Su, and Leonidas Guibas. Frustum pointnets for 3D object detection from RGB-D data. In CVPR, 2018. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 445, + 546, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 445, + 546, + 487 + ], + "spans": [ + { + "bbox": [ + 307, + 445, + 546, + 487 + ], + "type": "text", + "content": "[61] Nikhila Ravi, Jeremy Reizenstein, David Novotny, Taylor Gordon, Wan-Yen Lo, Justin Johnson, and Georgia Gkioxari. Accelerating 3D deep learning with PyTorch3D. arXiv:2007.08501, 2020. 6" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 489, + 546, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 489, + 546, + 521 + ], + "spans": [ + { + "bbox": [ + 307, + 489, + 546, + 521 + ], + "type": "text", + "content": "[62] Cody Reading, Ali Harakeh, Julia Chae, and Steven Waslander. Categorical depth distribution network for monocular 3D object detection. In CVPR, 2021. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 523, + 546, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 523, + 546, + 555 + ], + "spans": [ + { + "bbox": [ + 307, + 523, + 546, + 555 + ], + "type": "text", + "content": "[63] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In CVPR, 2016. 1, 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 557, + 546, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 557, + 546, + 590 + ], + "spans": [ + { + "bbox": [ + 307, + 557, + 546, + 590 + ], + "type": "text", + "content": "[64] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster R-CNN: Towards real-time object detection with region proposal networks. In NeurIPS, 2015. 1, 2, 4, 5" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 591, + 546, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 591, + 546, + 635 + ], + "spans": [ + { + "bbox": [ + 307, + 591, + 546, + 635 + ], + "type": "text", + "content": "[65] Mike Roberts, Jason Ramapuram, Anurag Ranjan, Atulit Kumar, Miguel Bautista, Nathan Paczan, Russ Webb, and Joshua Susskind. Hypersim: A photorealistic synthetic dataset for holistic indoor scene understanding. In ICCV, 2021. 2, 3" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 307, + 636, + 546, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 636, + 546, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 636, + 546, + 678 + ], + "type": "text", + "content": "[66] Danila Rukhovich, Anna Vorontsova, and Anton Konushin. ImVoxelNet: Image to voxels projection for monocular and multi-view general-purpose 3D object detection. In WACV, 2022. 2, 7, 8" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 307, + 680, + 546, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 546, + 713 + ], + "type": "text", + "content": "[67] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander Berg, and Li" + } + ] + } + ], + "index": 35 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13163" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "text", + "content": "Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. IJCV, 2015. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 128 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 128 + ], + "type": "text", + "content": "[68] Andrea Simonelli, Samuel Rota Bulo, Lorenzo Porzi, Manuel López-Antequera, and Peter Kontschieder. Disentangling monocular 3D object detection. In ICCV, 2019. 2, 4, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 129, + 287, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 161 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 161 + ], + "type": "text", + "content": "[69] Andrea Simonelli, Samuel Rota Bulo, Lorenzo Porzi, Elisa Ricci, and Peter Kontschieder. Towards generalization across depth for monocular 3D object detection. In ECCV, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 162, + 287, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 162, + 287, + 194 + ], + "spans": [ + { + "bbox": [ + 48, + 162, + 287, + 194 + ], + "type": "text", + "content": "[70] Shuran Song, Samuel Lichtenberg, and Jianxiong Xiao. SUN RGB-D: A RGB-D scene understanding benchmark suite. In CVPR, 2015. 1, 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 195, + 288, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 195, + 288, + 281 + ], + "spans": [ + { + "bbox": [ + 48, + 195, + 288, + 281 + ], + "type": "text", + "content": "[71] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, Vijay Vasudevan, Wei Han, Jiquan Ngiam, Hang Zhao, Aleksei Timofeev, Scott Ettinger, Maxim Krivokon, Amy Gao, Aditya Joshi, Yu Zhang, Jonathon Shlens, Zhifeng Chen, and Dragomir Anguelov. Scalability in perception for autonomous driving: Waymo open dataset. In CVPR, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 283, + 288, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 283, + 288, + 315 + ], + "spans": [ + { + "bbox": [ + 48, + 283, + 288, + 315 + ], + "type": "text", + "content": "[72] Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. FCOS: Fully convolutional one-stage object detection. In ICCV, 2019. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 316, + 287, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 316, + 287, + 348 + ], + "spans": [ + { + "bbox": [ + 48, + 316, + 287, + 348 + ], + "type": "text", + "content": "[73] Shubham Tulsiani, Saurabh Gupta, David Fouhey, Alexei Efros, and Jitendra Malik. Factoring shape, pose, and layout from the 2D image of a 3D scene. In CVPR, 2018. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 350, + 288, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 350, + 288, + 403 + ], + "spans": [ + { + "bbox": [ + 48, + 350, + 288, + 403 + ], + "type": "text", + "content": "[74] Pauli Virtanen, Ralf Gommers, Travis Oliphant, Matt Haberland, Tyler Reddy, David Cournapeau, Evgeni Burovski, Pearu Peterson, Warren Weckesser, Jonathan Bright, et al. Scipy 1.0: fundamental algorithms for scientific computing in python. Nature methods, 2020. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 404, + 288, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 404, + 288, + 448 + ], + "spans": [ + { + "bbox": [ + 48, + 404, + 288, + 448 + ], + "type": "text", + "content": "[75] Li Wang, Liang Du, Xiaqing Ye, Yanwei Fu, Guodong Guo, Xiangyang Xue, Jianfeng Feng, and Li Zhang. Depth-conditioned dynamic message propagation for monocular 3D object detection. In CVPR, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 449, + 288, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 449, + 288, + 480 + ], + "spans": [ + { + "bbox": [ + 48, + 449, + 288, + 480 + ], + "type": "text", + "content": "[76] Li Wang, Li Zhang, Yi Zhu, Zhi Zhang, Tong He, Mu Li, and Xiangyang Xue. Progressive coordinate transforms for monocular 3D object detection. In NeurIPS, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 482, + 288, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 482, + 288, + 514 + ], + "spans": [ + { + "bbox": [ + 48, + 482, + 288, + 514 + ], + "type": "text", + "content": "[77] Tai Wang, Zhu Xinge, Jiangmiao Pang, and Dahua Lin. Probabilistic and geometric depth: Detecting objects in perspective. In CoRL, 2022. 2, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 515, + 287, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 515, + 287, + 547 + ], + "spans": [ + { + "bbox": [ + 48, + 515, + 287, + 547 + ], + "type": "text", + "content": "[78] Tai Wang, Xinge Zhu, Jiangmiao Pang, and Dahua Lin. FCOS3D: Fully convolutional one-stage monocular 3D object detection. In ICCVW, 2021. 2, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 548, + 288, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 288, + 591 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 288, + 591 + ], + "type": "text", + "content": "[79] Yan Wang, Wei-Lun Chao, Divyansh Garg, Bharath Hariharan, Mark Campbell, and Kilian Weinberger. Pseudo-lidar from visual depth estimation: Bridging the gap in 3D object detection for autonomous driving. In CVPR, 2019. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 592, + 288, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 592, + 288, + 624 + ], + "spans": [ + { + "bbox": [ + 48, + 592, + 288, + 624 + ], + "type": "text", + "content": "[80] Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. Detector2. https://github.com/facebookresearch/detectron2, 2019. 2, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 625, + 288, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 625, + 288, + 669 + ], + "spans": [ + { + "bbox": [ + 48, + 625, + 288, + 669 + ], + "type": "text", + "content": "[81] Yurong You, Yan Wang, Wei-Lun Chao, Divyansh Garg, Geoff Pleiss, Bharath Hariharan, Mark Campbell, and Kilian Weinberger. Pseudo-lidar++: Accurate depth for 3D object detection in autonomous driving. In ICLR, 2020. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 670, + 288, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 288, + 691 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 288, + 691 + ], + "type": "text", + "content": "[82] Fisher Yu, Dequan Wang, Evan Shelhamer, and Trevor Darrell. Deep layer aggregation. In CVPR, 2018. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 692, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 692, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 692, + 288, + 713 + ], + "type": "text", + "content": "[83] Yunpeng Zhang, Jiwen Lu, and Jie Zhou. Objects are different: Flexible monocular 3D object detection. In CVPR, 2021. 2" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 274 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 116 + ], + "type": "text", + "content": "[84] Dingfu Zhou, Xibin Song, Yuchao Dai, Junbo Yin, Feixiang Lu, Miao Liao, Jin Fang, and Liangjun Zhang. IAFA: Instance-aware feature aggregation for 3D object detection from a single image. In ACCV, 2020. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 118, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 140 + ], + "type": "text", + "content": "[85] Xingyi Zhou, Dequan Wang, and Philipp Krahenbuhl. Objects as points. In arXiv preprint arXiv:1904.07850, 2019. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 141, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 141, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 308, + 141, + 545, + 173 + ], + "type": "text", + "content": "[86] Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. In CVPR, 2019. 4, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 175, + 547, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 175, + 547, + 217 + ], + "spans": [ + { + "bbox": [ + 308, + 175, + 547, + 217 + ], + "type": "text", + "content": "[87] Yunsong Zhou, Yuan He, Hongzi Zhu, Cheng Wang, Hongyang Li, and Qinhong Jiang. Monocular 3D object detection: An extrinsic parameter free approach. In CVPR, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 219, + 547, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 219, + 547, + 274 + ], + "spans": [ + { + "bbox": [ + 308, + 219, + 547, + 274 + ], + "type": "text", + "content": "[88] Zhikang Zou, Xiaqing Ye, Liang Du, Xianhui Cheng, Xiao Tan, Li Zhang, Jianfeng Feng, Xiangyang Xue, and Errui Ding. The devil is in the task: Exploiting reciprocal appearance-localization features for monocular 3D object detection. In ICCV, 2021. 2" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "13164" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/OmniAL_ A Unified CNN Framework for Unsupervised Anomaly Localization/390b7405-6295-46ac-87cb-a78d4a95dd63_content_list.json b/2023/OmniAL_ A Unified CNN Framework for Unsupervised Anomaly Localization/390b7405-6295-46ac-87cb-a78d4a95dd63_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f9cd352a4b749785eb104aef68d7ba72abdc0043 --- /dev/null +++ b/2023/OmniAL_ A Unified CNN Framework for Unsupervised Anomaly Localization/390b7405-6295-46ac-87cb-a78d4a95dd63_content_list.json @@ -0,0 +1,1388 @@ +[ + { + "type": "text", + "text": "OmniAL: A unified CNN framework for unsupervised anomaly localization", + "text_level": 1, + "bbox": [ + 104, + 130, + 866, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ying Zhao \nRicoh Software Research Center (Beijing) Co., Ltd.", + "bbox": [ + 279, + 181, + 689, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "zy_deepwhite.zy@hotmail.com", + "bbox": [ + 364, + 219, + 599, + 233 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c578b6fe0656ce94138d5b7e9fbf860c9e9cee3afb961f16fe8c0c0a857573c7.jpg", + "image_caption": [ + "(a). N models for N classes" + ], + "image_footnote": [], + "bbox": [ + 76, + 239, + 290, + 421 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3fb174b3e377479221cb887e62bdc6dd48c4d2cd5beffbe8f5d950db2709bfd3.jpg", + "image_caption": [ + "(b). One model for N classes" + ], + "image_footnote": [], + "bbox": [ + 334, + 241, + 552, + 421 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2c4b1b4d50b021c8efd9c7817168b581752878b475b92e121c46fe25429f5258.jpg", + "image_caption": [ + "(c). Proposed OmniAL", + "Figure 1. OmniAL is a unified framework for unsupervised anomaly localization. As shown in (c), OmniAL not only simplifies existing learning paradigm from (a) Separate: N models for N classes to (b) Unified: one model for N classes but also produces high quality anomaly localization results. Results in (a) and (b) are from separate and unified models of JNLD [38] on MVtecAD [1] dataset." + ], + "image_footnote": [], + "bbox": [ + 596, + 241, + 893, + 421 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 503, + 313, + 518 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Unsupervised anomaly localization and detection is crucial for industrial manufacturing processes due to the lack of anomalous samples. Recent unsupervised advances on industrial anomaly detection achieve high performance by training separate models for many different categories. The model storage and training time cost of this paradigm is high. Moreover, the setting of one-model-N-classes leads to fearful degradation of existing methods. In this paper, we propose a unified CNN framework for unsupervised anomaly localization, named OmniAL. This method conquers aforementioned problems by improving anomaly synthesis, reconstruction and localization. To prevent the model learning identical reconstruction, it trains the model with proposed panel-guided synthetic anomaly data rather than directly using normal data. It increases anomaly reconstruction error for multi-class distribution by using a network that is equipped with proposed Dilated Channel and Spatial Attention (DCSA) blocks. To better localize the anomaly regions, it employs proposed DiffNeck between reconstruction and localization sub-networks to explore multi-level differences. Experiments on 15-class MVTecAD and 12-class VisA datasets verify the advantage of proposed OmniAL that surpasses the state-of-the-art of unified models. On 15-class-MVTecAD/12-class-VisA, its single unified", + "bbox": [ + 75, + 537, + 473, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "model achieves 97.2/87.8 image-AUROC, 98.3/96.6 pixel-AUROC and 73.4/41.7 pixel-AP for anomaly detection and localization respectively. Besides that, we make the first attempt to conduct a comprehensive study on the robustness of unsupervised anomaly localization and detection methods against different level adversarial attacks. Experiential results show OmniAL has good application prospects for its superior performance.", + "bbox": [ + 496, + 505, + 893, + 627 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 647, + 630, + 662 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In real industrial scenarios, the location of anomaly [22, 26] reveals important information, such as defective types and degrees. It is essential not only to inspect whether a sample is defective but also to know where the specific anomaly regions are. Since anomaly appearance is inexhaustible, it is almost impossible and infeasible to collect and manually annotate all kinds of abnormal data. Thus, only normal samples are available for training a detector that is robust enough to find out unseen anomalies during inference phase. Considering the diversity of classes and various types of one class, the conventional training paradigm of N models for N classes, as shown in Fig.1a, may not be the best solution. The model storage and training time cost increase with the number of classes. As shown in Fig.1b, existing method severely degrades anomaly local-", + "bbox": [ + 496, + 674, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "3924", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/50640b258c33fab036f3f64210f80d9ab11981a8d04d725afa3a951cf2bbd3c1.jpg", + "image_caption": [ + "(a). Reconstruction-based anomaly localization", + "(1). MVtecAD" + ], + "image_footnote": [], + "bbox": [ + 86, + 93, + 230, + 284 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0ec8e8a7481a2e57af9bff2d53049336e23f72501496beb7ce13b8f5bcbf1f5b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 233, + 104, + 290, + 189 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ec5d871442bb5bcab9ab9878e954aa00b3f8a0ce6c7f0583e2ffa6188f909de6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 236, + 189, + 292, + 287 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e0cc6728ca0983fa01631e35b956d01e9db29ec95686c3a3db3274fa91cc64d5.jpg", + "image_caption": [ + "(2).JNLD Rec./Loc.", + "X Rec. X Loc." + ], + "image_footnote": [], + "bbox": [ + 299, + 104, + 406, + 189 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "", + "bbox": [ + 236, + 202, + 292, + 287 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/bf791d08e719ee28e038ef1f8b2f335d531853360f732cfcf22dc5adcaa4fc03.jpg", + "image_caption": [ + "sis unified models o", + "Figure 2. Problem analysis. The final failure may be caused by reconstruction and localization." + ], + "image_footnote": [], + "bbox": [ + 295, + 203, + 405, + 297 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/fbafbccb6a90f99a5d647c4eac91a3e55456db6cdd3c2c2471c88c42f73d10bb.jpg", + "image_caption": [ + "(3). OmniRec./Loc.", + "V Rec. V Loc." + ], + "image_footnote": [], + "bbox": [ + 410, + 104, + 519, + 189 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/08ecfb2baeb1600c30b74357fd1c65ee1e9263b9d8d47c4bd19d038347327100.jpg", + "image_caption": [ + "V Rec. V Loc." + ], + "image_footnote": [], + "bbox": [ + 410, + 203, + 517, + 289 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d45e0b87cbcd0bc1c30358911abc4aa3de3a3fb05baf8831ecc8f820ab106655.jpg", + "image_caption": [ + "(4).VisA", + "" + ], + "image_footnote": [], + "bbox": [ + 540, + 104, + 609, + 189 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b1f3fd5eccc8c573d5873458ff579f0d4d93ade00ee6bc2252ba388dbf7d7c23.jpg", + "image_caption": [ + "(5). JNLD Rec. / Loc.", + "X Rec." + ], + "image_footnote": [], + "bbox": [ + 614, + 104, + 751, + 189 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/1190e020d530cf7132c7690d369e4ea5b7dbc9dd3e941d2be6fcaf05854bc4d7.jpg", + "image_caption": [ + "V Rec.", + "Analysis unified models on VisA dataset." + ], + "image_footnote": [], + "bbox": [ + 614, + 203, + 751, + 287 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d78c4723537b92c5c4b8820ec491c98b090650f6b939f4b944656a7537d5be3f.jpg", + "image_caption": [ + "(6). OmniRec./Loc.", + "Rec.", + "√Loc." + ], + "image_footnote": [], + "bbox": [ + 754, + 104, + 890, + 189 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a18cc8e45a9a81061bcfd38dbb66af81ca883edb81cd645af2654895893a6ecd.jpg", + "image_caption": [ + "V Rec.", + "√ Loc." + ], + "image_footnote": [], + "bbox": [ + 754, + 203, + 890, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ization performance if the training paradigm changes to one model for N classes. Therefore, a robust unified framework for unsupervised anomaly localization is highly demanded for intelligent industrial.", + "bbox": [ + 75, + 343, + 468, + 402 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "With the limitation of available training data, many appealing unsupervised approaches [14, 18, 35, 38] using synthesized anomaly data are proposed. These approaches generate anomalous instances to inspire the anomaly detector to learn discriminative features. Their experiments show that the realiziness of generated anomalous instances had a strong impact on the quality of anomaly localization. However, none of these methods consider the training paradigm of one model for N classes. When switching to the unified training paradigm, they are more prone to learn an identical short-cut and fail to discriminate the anomaly.", + "bbox": [ + 75, + 407, + 468, + 574 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "With the normal and synthesized anomalous samples, recent unsupervised learning methods train a deep anomaly detector by either a distance-based [6, 14, 19-21, 27] or reconstruction-based [2, 9, 35, 36, 38] way. The reconstruction-based architectures [35, 38] are supposed to reconstruct normal images more accurately than the unseen anomalous. The anomaly localization is then calculated from the reconstruction error between the original and reconstructed versions of the input image, as shown in Fig.2a. The prediction of anomaly location is not only based on the reconstruction quality but also the ability of spotting the reconstruction error. The typical reconstruction-based method JNLD [38] learns a joint representation of an anomalous image and its anomaly-free reconstruction, while simultaneously learning a decision boundary between normal and simulated anomalous examples. As shown in Fig.2b and Fig.2c, under the unified setting, JNLD [38] fails to produce correct results either because of the reconstruction failure or the localization failure.", + "bbox": [ + 75, + 579, + 468, + 864 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To conquer aforementioned problems, we propose a novel unified framework OmniAL for effectively localiz-", + "bbox": [ + 76, + 869, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ing anomaly pixels of different classes only by using a single model. OmniAL uses a panel-guided anomaly synthesis method that controls the portion of normal and anomaly regions for each training sample. By doing this, OmniAL blocks the chance of learning identical shortcut from the source. To increase the anomaly reconstruction error for multi-class distribution, OmniAL constructs a reconstruction and a localization sub-networks that are equipped with proposed Dilated Channel and Spatial Attention (DCSA) blocks. To better localize the anomaly regions, OmniAL employs a DiffNeck module between the reconstruction and localization sub-networks to explore multi-level reconstruction errors. As shown in Fig.1 and Fig.2, OmniAL learns a single unified model for multiple classes that produces high quality reconstruction and precise anomaly localization. Furthermore, we conduct an exhaustive evaluation of reconstruction and localization performance against multi-level adversarial attacks.", + "bbox": [ + 496, + 343, + 890, + 613 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, we make following main contributions:", + "bbox": [ + 517, + 614, + 864, + 630 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We construct a unified CNN framework OmniAL for unsupervised anomaly localization that is equipped with proposed panel-guided anomaly synthesis, DCSA block, and DiffNeck module. OmniAL achieves superior performance for anomaly localization on challenging MVTecAD [1] and VisA [39] datasets compared to the state-of-the-art.", + "- By preventing model from learning identical reconstruction, our proposed panel-guided anomaly synthesis method also brings substantial improvement for existing methods under the unified setting. It boosts the image-AUROC/pixel-AUROC/pixel-AP from 88.7/87.1/49.4 to 92.5/94.5/57.4 for Draem [35].", + "- We make a comprehensive study on the robustness of separate/unified anomaly localization methods against different level adversarial attacks. Our synthesized" + ], + "bbox": [ + 516, + 638, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "3925", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/0609dd1ea28a4b3cfc8efc9cedeb888b7b397eab086b0021158a31d436678d94.jpg", + "image_caption": [ + "Figure 3. Framework of OmniAL. It consists of panel-guided anomaly synthesis, reconstruction and localization. Anomaly synthesis is based on anomaly panel and three variants of the Just Noticeable Distortion (JND) map. The synthetic anomaly is reconstructed into normal image and corresponding JND map by the Dilated Channel Spatial Attention (DCSA) modules equipped reconstruction sub-network. The localization sub-network with a DiffNeck module localizes the anomaly regions by exploring the difference between reconstructed and original data." + ], + "image_footnote": [], + "bbox": [ + 80, + 88, + 890, + 271 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "adversarial datasets exhibit strong attack capability against anomaly detection, reconstruction and localization, also helping to analyse the risks of existing methods.", + "bbox": [ + 107, + 354, + 470, + 414 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 433, + 218, + 450 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Anomaly synthesis. Due to the lack of anomaly samples, unsupervised learning methods are commonly used in industrial quality inspection. These methods are typically trained with normal data and synthesized anomalous that highly effect the performance. Therefore, how to synthesize the anomalous [7, 8, 14, 35, 37, 38] also draws extensive attention. CutPaste [14] learns representations by classifying normal data from the Cut-and-Paste augmentation. SPD [39] uses a smoothed version of CutPaste [14] augmentation. Instead of using simple regular shaped anomalous, Draem [35] simulates just-out-of-distribution anomalies having random shape and texture. To simulate photorealistic anomaly samples, JNLD [38] proposes a multiscale noticeable anomalous generation method based on just noticeable distortion [31]. The open-set supervised anomaly detection method DRA [7] adapts the popular CutMix [34] and the outlier exposure method [10] to generate pseudo anomalies from normal images for training. DSR [37] generates the anomalies at the feature level by sampling the learned quantized feature space, which allows a controlled generation of near-in-distribution anomalies. For the unified paradigm, we propose a panel-guided anomaly synthesis method that controls the portion of normal and anomaly regions for each training sample.", + "bbox": [ + 75, + 460, + 470, + 823 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Anomaly localization. Anomaly detection [22,26], also known as outlier detection or one-class classification, refers to the task of distinguishing defective image at the image-level from the majority of anomaly-free images. Anomaly localization (segmentation), on the other hand, aims to seg", + "bbox": [ + 75, + 824, + 468, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ment out the pixel-level anomaly regions. For industrial visual inspection, many recent methods [5, 6, 19] achieve high performance in anomaly detection but anomaly localization. SPADE [5] detects anomaly based on alignment between an anomalous image and a constant number of the similar normal images. It relies on K nearest neighbors of pixel-level feature pyramids extracted by pre-trained deep features. PSVDD [32] extends deep Support Vector Data Description (SVDD) to a patch-wise detection method. PaDim [6] also relies on ImageNet pretrained feature extractor with multiscale pyramid pooling. Instead of using time-costly clustering, it uses a well-known Mahalanobis distance metric [17] as an anomaly score. CutPaste [14] uses GradCAM [24] to get the defect localization. MKD [23] proposes to use multi-level features alignment to increase the discriminating capability of the Teacher-Student model on various types of abnormalities. PatchCore [19] combines patch-level embeddings from ImageNet models with an outlier detection model. However, PatchCore is inherently not suitable for unified setting (larger normal dataset) since it needs to build the coreset with all normal data. Draem [35] and JNLD [38] use segmentation sub-network to predict the defective regions. This paper differs from these previous works by focusing on the paradigm of only using a single unified model for N classes.", + "bbox": [ + 496, + 354, + 890, + 731 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unified anomaly localization. Recently, how to use a unified model to localize anomaly for different objects has already become researchers' concern. RegAD [11] trains a single generalizable model for few-shot anomaly detection, where a limited number of normal images are provided for each category at training. It employs SimSiam [3] with three spatial transformer network [13] blocks to solve the category-agnostic feature registration proxy task. It identifies anomalies by comparing the registered features of the test image and the corresponding normal images. UniAD [33] constructs a transformer with a layer-wise query de", + "bbox": [ + 496, + 734, + 892, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3926", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "coder and a neighbor masked attention module to model the multi-class distribution. It uses the layer-wise query decoder to intensify the use of query embedding. To avoid the information leak, it employs a neighbor masked attention module, where a feature point relates to neither itself nor its neighbors. In this paper, we propose a unified CNN framework for unsupervised anomaly localization. It manifests the problem-solving ability of fully convolutional networks.", + "bbox": [ + 75, + 90, + 472, + 212 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 229, + 168, + 244 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this work, we construct a unified CNN framework OmniAL for unsupervised anomaly localization under the practical setting of only using one model for N classes. Fig.3 shows the overview of proposed OmniAL. To train only with normal data, we firstly synthesize photo-realistic anomalous in the regions that are selected based 5 predefined panels. The synthesized anomaly samples are then used to train the OmniAL network that consists of an anomaly-free reconstruction and an anomaly localization sub-networks. The reconstruction and localization sub-networks are composed of alternative basic blocks and DCSA blocks. The reconstruction sub-network learns to recover the synthesized anomalous back to normal. The localization sub-network with a DiffNeck module localizes the anomaly regions by exploring the difference between reconstructed and original data.", + "bbox": [ + 75, + 256, + 472, + 500 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Panel-guided anomaly synthesis", + "text_level": 1, + "bbox": [ + 76, + 512, + 357, + 527 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Reconstruction-based anomaly localization methods rely on the hypothesis that the reconstruction errors in unseen anomalous regions are larger than the normal regions. The reconstruction models are trained with alternative normal and synthetic anomaly samples. For normal sample, the supervision ground truth is the direct copy of the input. Thus, directly using normal samples for training increases the chance of information leak and leads to identical reconstruction for any input. To prevent the learning of identical short-cut, the intuitive idea is that only using synthetic anomaly rather than normal data. However, without using normal samples for training leads to more false alarm. Then, the problem is transformed into how to better synthesis anomaly for unified model training.", + "bbox": [ + 75, + 537, + 468, + 748 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To solve the problem, we propose a panel-guided anomaly synthesis method that takes into account both normal and anomaly for training. As shown in Fig.4, we build 5 types of panels ('left', 'right', 'top', 'bottom', 'all') to control the portion of normal and anomaly regions. That is, we only generate the anomaly in one of the panel region for each sample in each iteration. Except the type of 'all', the area of panel region is also randomly adjusted in range of [0.5, 0.8]*ImageWidth and [0.5, 0.8]*ImageHeight. Given a panel, we synthesize anomaly in the selected region ac", + "bbox": [ + 75, + 750, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8a2c312adc127bd13d9679b228f5f314f456f94944c8e7c35fc5f2ec3338b76d.jpg", + "image_caption": [ + "Figure 4. Visualization of our panel-guided anomaly synthetic data. The anomaly source is selected either from augmented normal images (row1,2) or DTD [4] dataset (row3)." + ], + "image_footnote": [], + "bbox": [ + 501, + 88, + 893, + 241 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "cording to following equation.", + "bbox": [ + 500, + 290, + 702, + 305 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nA = \\sum_ {i = 1} ^ {3} \\left(W _ {i} T + (1 - W _ {i}) I\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 308, + 890, + 347 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Where $A$ is the synthesized anomaly image, $W_{i}$ is the pixelwise fusion weight map, $i$ indicates one of the 3 defect levels (easy, medium and hard), $I$ is the anomaly-free image, $T$ is the anomaly source image. To synthesize natural and diversity anomalies, $T$ is generated by applying a set of random augmentation (resize, crop, flip, color-jitter) to $I$ or image from the Describable Textures Dataset(DTD) [4].", + "bbox": [ + 496, + 349, + 890, + 455 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To get the 3-level defects, we first need to find out the baseline anomaly fusion weight map. Following JNLD [38], we use the JND [31] to guide the anomaly synthesis. JND [31] refers to the minimum visibility threshold of the Human Visual System, is useful in perceptual image/video processing systems. It reveals a perceptual threshold of intensity change in an image that can be noticed by the human vision system. We use a nonlinear additivity JND model to define the anomaly weight map $J$ as [25] that proposes a novel hybrid exposure weight measurement using the JND. Each level anomaly fusion weight map $W_{i}$ is defined as:", + "bbox": [ + 496, + 455, + 892, + 621 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nW _ {i} = \\alpha_ {i} * (J + \\beta_ {i}) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 625, + 636, + 890, + 652 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Where $J$ indicates the JND map, $\\beta_{i}$ is a bias term for different level of noticeable, $\\alpha_{i}$ is a smooth kernel applied to the mask and produces soft boundary between anomaly and normal regions. For the easier level anomaly, we use a smaller smooth kernel. Due to using random seeds, different level anomaly regions may overlap each other. Thus, directly using the masks as the ground truth for anomaly segmentation task may bring label inconsistency. Therefore, we refine and quantify the segmentation mask into 3 defect levels based on the structure similarity index measure (SSIM) [29] between the anomaly-free image $I$ and synthesized anomaly image $A$ .", + "bbox": [ + 496, + 657, + 890, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Anomaly reconstruction", + "text_level": 1, + "bbox": [ + 500, + 847, + 720, + 863 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Fig.3, our anomaly reconstruction subnetwork receives an input image and outputs reconstructed", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "3927", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "normal image and corresponding JND map. The reconstruction sub-network consists of an encoder and a decoder. Both of them are composed of alternating basic and DCSA blocks. The basic block extracts different levels of features with two consecutive $3 \\times 3$ convolution layer followed by normalization and ReLU activation layers. Considering the unified training paradigm, instance normalization [28] is more suitable than the broadly used batch normalization [12] in anomaly reconstruction. The reason is that batch normalization [12] normalizes using the information from the whole batch, while instance normalization [28] normalizes each feature map on its own. Under unified setting, each batch of training data consists of different classes that complicates the reconstruction problem. Instance normalization discards instance-specific contrast information and reduces undesirable instance-level variation. Batch normalization, meanwhile, lacks the ability to address the class inconsistency. Therefore, we use instance normalization in each basic block rather than batch normalization. Experimental results in Table 3 show that the performance is improved by swapping batch normalization with instance normalization.", + "bbox": [ + 76, + 90, + 472, + 421 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The feature maps extracted by the basic block are further integrated by the DCSA block that is inspired by the popular CBAM [30]. As shown in orange dash block of Fig.3, the DCSA block consists of channel and spatial attention sequential sub-modules. With average pooling and two consecutive 1x1 convolutions, the channel attention explores the correlations among feature channels. The correlations are add back to original feature maps to emphasize the important feature channels. The spatial attention further investigates the spatial correlation. It first aggregates spatial information by concatenating both average and maximum pooling of afore refined feature maps. Then, it spots the spatial importance of the aggregated features with a standard and a dilated $3 \\times 3$ convolution-norm-relu paths. It balances the portion of dilated path by a factor that is set as 0.1. The spatial attention map is obtained by summing up two paths of features. By adding the spatial attention map, the important regions in feature maps are highlighted. As shown in Table 3, the performance is further improved by using DCSA blocks.", + "bbox": [ + 76, + 424, + 472, + 727 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Anomaly localization", + "text_level": 1, + "bbox": [ + 76, + 739, + 276, + 755 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After reconstructing the normal and corresponding JND maps, we use a localization sub-network to highlight the anomaly regions. Instead of directly concatenating the reconstructed maps with input image, we firstly use a DiffNeck module to fully explore the difference between them. The pink dash block of Fig.3 illustrates the detailed structure of DiffNeck. Overall, DiffNeck extracts reconstruction difference among reconstructed normal map (Rec), reconstructed JND map (JND) and input image (Inp) in both sep", + "bbox": [ + 76, + 763, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "arate and unified ways. That is, DiffNeck uses basic blocks to extract the concatenated features of (Rec, Inp), (JND, Inp) and (Rec, JND, Inp) respectively. Then, all levels of differences are summed up and form a final descriptor that is fed to the localization sub-network. DiffNeck also balances the portion of the separate path by a factor that is set as 0.1.", + "bbox": [ + 496, + 90, + 893, + 195 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Similar with reconstruction sub-network, the localization sub-network also follows the encoder-decoder structure but with 6 scales (e1-e5 in Fig.3) and skip connections for corresponding scales. Scale-1 and scale-2 use standard convolution in all blocks. To use larger resolution feature maps and receptive field, the basic and DCSA blocks of the encoder contain dilated convolution layer from scale-3 to scale-6 with 2, 4, 8, 16 dilation rates respectively. Different with reconstruction sub-network, we use batch normalization in all blocks. Since we synthesize more realistic anomalies by considering easy, medium and hard level defects, localization sub-network predicts 3 levels of anomalous locations. To get the final pixel-level anomaly localization map, we combine them together by a sequence processing of SoftMax, smoothness and summing up. By doing this, the false alarm noise can be suppressed, and the weak true decision can be strengthened.", + "bbox": [ + 496, + 196, + 893, + 453 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The total target for training OmniAL is defined as", + "bbox": [ + 519, + 454, + 849, + 468 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L = L _ {2} \\left(J, J _ {r}\\right) + L _ {s s i m} \\left(J, J _ {r}\\right) \\tag {3} \\\\ + L _ {2} (I, I _ {r}) + L _ {s s i m} (I, I _ {r}) + L _ {f l} (S, S _ {s}) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 470, + 890, + 506 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Where $J$ is the ground truth JND map calculated from the anomaly-free image $I$ , $J_{r}$ and $I_{r}$ are reconstructed JND map and normal image, $S$ is the anomaly localization ground truth. Unlike JNLD [38], for both JND and normal reconstruction, we not only use MSE loss to supervise the pixel-to-pixel recovering but also the SSIM loss to yield plausible local consistency. To handle the unbalance of different types, we use focal loss [15] to supervise the predicted anomaly localization $S_{s}$ .", + "bbox": [ + 496, + 508, + 893, + 645 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experimental results", + "text_level": 1, + "bbox": [ + 500, + 656, + 700, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets and metrics", + "text_level": 1, + "bbox": [ + 500, + 681, + 696, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To demonstrate the effectiveness of proposed OmniAL, we conduct extensive experiments on the challenging MVTecAD [1] and VisA [39] datasets.", + "bbox": [ + 496, + 704, + 890, + 750 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "MVTecAD [1] dataset contains 10 object and 5 texture industrial products, such as bottle and leather. It consists of 3,629 normal images for training and 1,725 images for testing. There are 1,258 anomaly images of the testing set with pixel-level labelled various types of defects and the rest are normal images. Each class contains 60 to 320 color images with the resolution ranges from $700 \\times 700$ to $1024 \\times 1024$ pixels. In the testing set, defective appearance varies in different sizes, shapes and types, and most cases only contain a small fraction of anomalous pixels.", + "bbox": [ + 496, + 750, + 893, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "3928", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/964ab583725b9cd3f750bb1addf51c8e6058ecdb7681e32ca503ff6764068aab.jpg", + "table_caption": [ + "Table 1. Image AUROC comparison of our method OmniAL with the state-of-the-art on MVTecAD [1]." + ], + "table_footnote": [], + "table_body": "
CategoryPadim [6]CutPaste [14]MKD [23]Draem [35]JNLD [38]UniAD [33]OmniAL
Unified / Separate
bottle97.9 / 99.967.9 / 98.298.7 / 99.494.6 / 99.199.1 / 96.399.7 / 100100 / 99.4
cable70.9 / 92.769.2 / 81.278.2 / 89.261.8 / 94.790.6 / 98.895.2 / 97.698.2 / 97.6
capsule73.4 / 91.363.0 / 98.268.3 / 80.570.2 / 98.574.4 / 85.086.9 / 85.395.2 / 92.4
carpet93.8 / 99.893.6 / 93.969.8 / 79.395.9 / 95.577.1 / 97.899.8 / 99.998.7 / 99.6
grid73.9 / 96.793.2 / 10083.8 / 98.098.1 / 99.998.6 / 10098.2 / 98.599.9 / 100
hazelnut85.5 / 92.080.9 / 98.397.1 / 98.495.1 / 10090.8 / 10099.8 / 99.995.6 / 98.0
leather99.9 / 10093.4 / 10093.6 / 95.199.9 / 10097.0 / 100100 / 10099.0 / 97.6
metal nut88.0 / 98.760.0 / 99.964.9 / 73.688.9 / 98.793.3 / 99.699.2 / 99.099.2 / 99.9
pill68.8 / 93.371.4 / 94.979.7 / 82.769.0 / 98.982.7 / 94.693.7 / 88.397.2 / 97.7
screw56.9 / 85.885.2 / 88.775.6 / 83.393.3 / 93.981.8 / 95.887.5 / 91.988.0 / 81.0
tile93.3 / 98.188.6 / 94.689.5 / 91.698.3 / 99.699.2 / 10099.3 / 99.099.6 / 100
toothbrush95.3 / 96.163.9 / 99.475.3 / 92.282.8 / 100100 / 10094.2 / 95.0100 / 100
transistor86.6 / 97.457.9 / 96.173.4 / 85.683.9 / 93.190.3 / 93.099.8 / 10093.8 / 93.8
wood72.1 / 96.580.4 / 99.193.4 / 94.399.8 / 99.191.9 / 99.698.6 / 97.993.2 / 98.7
zipper79.7 / 90.393.5 / 99.987.4 / 93.299.1 / 10099.8 / 99.895.8 / 96.7100 / 100
average84.2 / 95.577.5 / 96.181.9 / 87.888.7 / 98.091.3 / 97.496.5 / 96.697.2 / 97.0
", + "bbox": [ + 83, + 103, + 888, + 327 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a8d84f6932e980a06ea66c06a0662f295b7e3d4cb3f0cc27a2024cab616e0626.jpg", + "table_caption": [ + "Table 2. Pixel AUROC comparison of our method OmniAL with the state-of-the-art on MVTecAD [1]." + ], + "table_footnote": [], + "table_body": "
CategoryPadim [6]PSVDD [32]MKD [23]Draem [35]JNLD [38]UniAD [33]OmniAL
Unified / Separate
bottle96.1 / 98.286.7 / 98.191.8 / 96.387.4 / 99.194.8 / 99.098.1 / 98.199.2 / 99.0
cable81.0 / 96.762.2 / 96.889.3 / 82.470.4 / 94.776.4 / 97.797.3 / 96.897.3 / 97.1
capsule96.9 / 98.683.1 / 95.888.3 / 95.949.2 / 94.357.0 / 92.798.5 / 97.996.9 / 92.2
carpet97.6 / 99.078.6 / 92.695.5 / 95.695.2 / 95.593.7 / 99.098.5 / 98.099.4 / 99.6
grid71.0 / 97.170.8 / 96.282.3 / 91.899.0 / 99.796.9 / 99.798.2 / 98.599.4 / 99.6
hazelnut96.3 / 98.197.4 / 97.591.2 / 94.696.0 / 99.785.9 / 99.496.5 / 94.698.4 / 98.6
leather84.8 / 99.093.5 / 97.496.7 / 98.198.6 / 98.687.0 / 99.598.8 / 98.399.3 / 99.7
metal nut84.8 / 97.396.0 / 98.064.2 / 86.472.6 / 99.597.4 / 99.594.8 / 95.799.1 / 99.1
pill87.7 / 95.796.5 / 95.169.7 / 89.690.0 / 97.691.2 / 96.695.0 / 95.198.9 / 98.6
screw94.1 / 98.474.3 / 95.792.1 / 96.089.3 / 97.687.0 / 99.798.3 / 97.498.0 / 97.2
tile80.5 / 94.192.1 / 91.485.3 / 82.898.1 / 99.294.7 / 99.691.8 / 91.899.0 / 99.4
toothbrush95.6 / 98.898.0 / 98.188.9 / 96.194.4 / 98.198.6 / 98.898.4 / 97.899.4 / 99.2
transistor92.3 / 97.678.5 / 97.071.7 / 76.573.1 / 90.983.6 / 92.197.9 / 98.793.3 / 91.7
wood89.1 / 94.180.7 / 90.880.5 / 84.896.2 / 96.488.7 / 96.393.2 / 93.497.4 / 96.9
zipper94.8 / 98.495.1 / 95.186.1 / 93.996.9 / 98.895.3 / 99.496.8 / 96.099.5 / 99.7
average89.5 / 97.485.6 / 95.784.9 / 97.087.1 / 97.388.6 / 97.996.8 / 96.698.3 / 97.8
", + "bbox": [ + 83, + 349, + 888, + 573 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/34fdf111d3907379b3e122603a3dee248a534c8bbc8ab22a9fa06fbc19ac369b.jpg", + "table_caption": [ + "Table 3. Ablation study on MVTecAD [1]. PA: Panel-guided anomaly synthesis, Rec: Reconstruction sub-network, Seg: Segmentation sub-network, BB: Basic Block, BN: Batch Norm, IN: Instance Norm, DC: Dilated Convolution, CA: Channel Attention, DSA: Dilated Spatial Attention, I: Image-level classification, P: Pixel-level localization." + ], + "table_footnote": [], + "table_body": "
PARec-BBDCSASeg-BBDiffNeckI-AUROCP-AUROCP-AP
-BNINDCCADSABNDC-Unified/Separate
-√--√-√--86.7/98.886.4/98.444.2/75.0
√√--√-√--94.1/—95.9/—68.3/—
√-√-√-√--95.8/—97.0/—69.1/—
√-√√√-√√-96.9/97.397.6/97.772.8/72.9
√-√√√√√√-94.7/96.596.9/97.869.8/74.9
√-√√√-√√√96.8/96.398.2/98.072.7/74.1
√-√√√√√√√97.2/97.098.3/97.873.4/73.5
", + "bbox": [ + 80, + 674, + 485, + 746 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/47647881ce1f3e1074ddd94c7bad44b73b20b6258f3e51c0abdeeaae95ef48d9.jpg", + "table_caption": [ + "Table 4. Ablation study on With/Without panel-guided anomaly synthesis for unified training on MVTecAD [1]." + ], + "table_footnote": [], + "table_body": "
Draem [35]JNLD [38]OmniAL
I-AUROC92.5/88.792.9/91.397.2/86.6
P-AUROC94.5/87.195.6/88.698.3/93.7
Pixel-AP57.4/49.463.1/46.673.4/54.4
", + "bbox": [ + 81, + 782, + 483, + 851 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "VisA [39] dataset consists of 10,821 high-resolution color images (9,621 normal and 1,200 anomalous samples) covering 12 objects in 3 domains, including complex struc", + "bbox": [ + 75, + 854, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ture, multiple instances and single instances. The anomalous images contain various flaws, including surface defects such as scratches, dents, color spots or crack, and structural defects like misplacement or missing parts. There are 5-20 images per defect type and an image may contain multiple defects. All images were acquired using a $4,000 \\times 6,000$ high-resolution RGB sensor. Example of each category and our corresponding anomaly reconstruction and localization results are shown in Fig.5a(MVTecAD) and Fig.5b(VisA).", + "bbox": [ + 496, + 593, + 890, + 731 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics For anomaly detection evaluation, the most common used metrics are Area Under the Receiver Operating Characteristic curve (AUROC) in both image-level and pixel-level. However, the pixel-AUROC is biased in favour of large anomalous regions and does not well reflect the pixel-level anomalous localization performance. Therefore, we additionally introduce the average precision (AP) to evaluate pixel-level anomaly localization performance. For reconstruction quality evaluation, we use peak signal to noise ratio (PSNR) and SSIM [29] that are commonly used for evaluating signal fidelity.", + "bbox": [ + 496, + 734, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "3929", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/5caf1edee28007a2aadb0fcbb2dc7ad4baaeaf010685510264c951598f7d887f.jpg", + "table_caption": [ + "Table 5. Performance comparison of our method OmniAL with the state-of-the-art on VisA [39]. (Separate/Unified)" + ], + "table_footnote": [], + "table_body": "
CategoryPadim+SPD [39]Draem [35]JNLD [38]OmniAL
I-AUCP-AUCI-AUCP-AUCP-API-AUCP-AUCP-API-AUCP-AUCP-AP
Complex structurePCB192.797.771.3/83.998.6/94.060.4/38.582.0/82.996.4/98.072.8/77.896.6/77.798.7/97.663.5/76.9
PCB287.997.289.7/81.792.5/94.13.5/13.396.3/79.191.9/95.031.2/37.799.4/81.083.2/93.92.8/31.6
PCB385.496.773.1/87.793.8/94.118.7/17.996.9/90.195.3/98.543.4/46.896.9/88.198.4/94.756.9/41.4
PCB499.189.291.3/87.195.8/72.332.7/13.194.8/96.296.1/97.537.4/29.797.4/95.398.5/97.138.4/33.2
Multiple instancesMacaroni185.798.870.3/68.695.8/89.88.2/8.094.3/90.598.8/93.325.9/14.296.9/92.698.9/98.67.6/7.1
Macaroni270.896.071.3/60.394.1/83.225.4/19.786.5/71.392.9/92.117.2/6.189.9/75.299.1/97.911.4/9.2
Capsules68.186.377.3/89.693.7/96.620.2/30.189.1/91.498.9/99.638.6/55.687.9/90.698.6/99.462.9/52.4
Candles89.197.382.3/70.287.0/82.627.9/12.689.1/85.494.8/94.525.9/27.485.1/86.890.5/95.829.2/24.6
Single instanceCashew90.586.194.2/67.394.7/68.541.2/7.096.0/82.596.3/94.143.7/39.497.1/88.698.9/95.077.3/47.4
Chewing gum99.396.993.4/90.097.5/92.740.9/59.098.5/96.099.4/98.974.7/81.694.9/96.498.7/99.082.9/79.5
Fryum89.888.0100/86.297.5/83.240.9/26.493.2/91.995.8/90.042.9/30.497.0/94.689.3/92.128.3/34.4
Pipe fryum95.695.494.1/87.181.8/72.323.7/13.196.0/87.597.0/92.544.1/31.891.4/86.199.1/98.269.1/62.6
Mean87.893.884.1/80.588.8/87.025.4/20.893.0/87.196.1/95.241.5/39.994.2/87.896.0/96.644.2/41.7
", + "bbox": [ + 81, + 85, + 890, + 263 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/61f42219a66eedb889bdab0543e37967c016d4231d383af80fb771d01a2b9b71.jpg", + "image_caption": [ + "(a). 15 categories of MVtecAD dataset" + ], + "image_footnote": [], + "bbox": [ + 81, + 270, + 486, + 520 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5d36cee805ebdc32748f1134b26f9deeafa13a2cb382e93460eebc9a3eb77251.jpg", + "image_caption": [ + "(b). 12 categories of VisA dataset", + "Figure 5. Qualitative illustration of our anomaly detection results. From top to bottom, the test images overlap with ground truth contours, the reconstructions and the anomaly localization map overlays are shown in rows." + ], + "image_footnote": [], + "bbox": [ + 488, + 271, + 890, + 521 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Implementation", + "text_level": 1, + "bbox": [ + 76, + 571, + 235, + 588 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To compare with existing methods, we train OmniAL both in unified and separate paradigm with a batch size of 20/12 images having size of $256 \\times 256 / 256 \\times 320$ for MVtec/VisA and pixel value range of [0, 1]. A single model is trained with batches that contain all-class samples for unified paradigm. On the contrary, the separate models are trained with corresponding class samples respectively. The Adam optimizer has an initial learning rate of 1e-4. To alleviate over-fitting, during training, the anomaly texture is alternatively selected either from augmented normal images or from the DTD [4] dataset.", + "bbox": [ + 75, + 595, + 470, + 761 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Comparison and performance", + "text_level": 1, + "bbox": [ + 76, + 770, + 344, + 787 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Performance on MVTecAD. We compare our proposed method with state-of-the-art methods recently reported on MVTecAD for both unified and separate models in anomaly detection and localization. Table 1 and 2 show our quantitative comparison with the state-of-the-art methods on the task of anomaly detection and localization. UniAD [33] is designed for the unified training paradigm while the others", + "bbox": [ + 75, + 794, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "are for the separate training scheme. The results of unified model of Draem [35] and JNLD [38] are got by retraining their models under the paradigm of one model for N classes. The others are reported from UniAD [33]. As shown in Table 1 and 2, for both image-level and pixel-level anomaly detection, most of the existing methods' performance drop drastically when the paradigm switch from training N separate models to a unified model. Comparing with the transformer-based method UniAD [33], our CNN-based method achieves $0.7\\%$ and $1.5\\%$ higher performance on image-AUROC and pixel-AUROC. Moreover, as shown in Table 4, OmniAL(73.4) surpasses SOTA reconstruction-based methods Draem(49.4) [35] and JNLD(46.6) [38] in pixel-AP with more than $24\\%$ .", + "bbox": [ + 496, + 571, + 892, + 785 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Performance on VisA. Comparing with MVTecAD, VisA is more difficult since it considers more complex structure and multiple misaligned instances scenes. Table 5 shows the superior performance of OmniAL comparing with the baseline separate method SPD [39] that proposes the VisA dataset and two reconstruction-based methods under the unified setting. The results of both separate and uni-", + "bbox": [ + 496, + 794, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "3930", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ca1fde8e5fbed17071294e052aca400e35e6324c9d8ae52145257d03dcfaeba0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 93, + 460, + 257 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/18faa221c1f5ecc57918c08b31601c43294e181a70f0c88f66df0eb8507778e9.jpg", + "image_caption": [ + "Figure 6. Robustness comparison of our method OmniAL with the state-of-the-art against adversarial attack." + ], + "image_footnote": [], + "bbox": [ + 78, + 258, + 460, + 436 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "fied model of Draem [35] and JNLD [38] are got by retraining on VisA. For separate training, OmniAL surpasses the best of them with $1.2\\%$ image-AUROC, $2.7\\%$ pixel-AP and similar pixel-AUROC. For unified training, OmniAL also surpasses the best of them with $0.7\\%$ image-AUROC, $1.4\\%$ pixel-AUROC and $1.8\\%$ pixel-AP.", + "bbox": [ + 75, + 476, + 467, + 566 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown in Fig.5, our reconstruction and localization achieve good performance in most categories. Even in the multi-instance scene, OmniAL still successfully recovers the anomaly region back to normal. The reconstructed images have the high-fidelity appearance with the inputs in normal regions and recover the anomalies as close as the expectation. The quantity evaluation of reconstruction is reported in Fig.6a. The bigger circles indicate the reconstruction performance on the clean normal MVTecAD [1]. Our OmniAL surpasses the Draem [35] and JNLD [38] with big gaps in both PSNR and SSIM scores under the unified setting.", + "bbox": [ + 75, + 566, + 467, + 748 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Adversarial robustness. Adversarial attacks are easily performed under the disguise of anomaly with additive noises and hardly arouse suspicion. Our anomaly synthesis method can be expeditiously expanded to generate adversarial anomaly samples with PGD [16] perturbations for robustness evaluation. To further evaluate the robustness of separate and unified models, we build up 3 adversarial datasets with PGD [16] perturbations $\\epsilon = \\{4 / 255, 8 / 255\\}$ and $\\epsilon = 8 / 255$ with targeted attack based on the anomaly-free training set from clean MVTecAD [1]. For realistic ad", + "bbox": [ + 75, + 750, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "versarial samples, the synthesized anomalies only appear in foreground of the object categories and the anomaly texture is cropped from the normal images. The generated adversarial datasets consist of $80\\%$ synthesized anomaly and $20\\%$ normal samples. Fig.6 illustrates the performance comparison in (a) reconstruction and (b) localization respectively. The names with 'sp' suffix indicate separate models while the others are unified models. OmniAL(green) achieves the best performance against all adversarial levels and has less degradation. All category adversarial samples and performance comparison are shown in supplementary material.", + "bbox": [ + 496, + 90, + 890, + 257 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablation study", + "text_level": 1, + "bbox": [ + 500, + 266, + 648, + 281 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3 demonstrates the effectiveness of proposed panel-guided anomaly synthesis, DCSA block and Diff-Neck module for both unified and separate training paradigm. We build a baseline by only using channel attention in DCSA blocks and achieve the best performance in separate setting but the worse performance in unified setting. By using panel-guided anomaly synthesis, we achieve about $10\\%$ image-level and $24\\%$ pixel-level performance improvement comparing with the baseline. The overall performance is further improved by using instance normalization rather than batch normalization in the reconstruction basic blocks. With the help of dilated convolution layer in basic blocks, we can preserve high resolution feature maps and get further improvement. By further using dilated spatial attention, the separate models performance gain $2\\%$ but the unified model decreases $3\\%$ in pixel-AP. Alternatively, the model using DiffNeck module gains $1.2\\%$ for separate setting and $0.6\\%$ for unified setting. Finally, by combining dilated spatial attention and DiffNeck together, we achieve the balanced performance in separate and unified settings and get the OmniAL. As shown in Table 4, proposed panel-guided anomaly synthesis also boosts existing reconstruction-based methods with a large margin. More ablation studies are shown in supplementary material.", + "bbox": [ + 496, + 287, + 890, + 652 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 664, + 617, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Considering the practical usages, we propose a unified CNN framework to localize anomalous for multiple classes with a single model. Extensive experiments on MVTecAD and VisA datasets verify the effectiveness of our proposed panel-guided anomaly synthesis, dilated channel and spatial attention (DCSA) block, and DiffNeck module. Especially, OmniAL produces more precisely anomaly reconstruction and localization results. The panel-guided anomaly synthesis can be easily used in existing methods and brings performance improvement under the unified training paradigm. Moreover, we make the first attempt to conduct a comprehensive study on the robustness of existing methods against different levels of adversarial for both separate and unified training paradigm.", + "bbox": [ + 496, + 688, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "3931", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. Mvtec AD - A comprehensive real-world dataset for unsupervised anomaly detection. In CVPR, pages 9592-9600, 2019. Computer Vision Foundation / IEEE. 1, 2, 5, 6, 8", + "[2] Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. Uninformed students: Student-teacher anomaly detection with discriminative latent embeddings. In CVPR, pages 4182–4191, 2020. Computer Vision Foundation / IEEE. 2", + "[3] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 15750-15758. Computer Vision Foundation / IEEE, 2021. 3", + "[4] Mircea Cimpoi, Subhansu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In CVPR, pages 3606-3613, 2014. IEEE Computer Society. 4, 7", + "[5] Niv Cohen and Yedid Hoshen. Sub-image anomaly detection with deep pyramid correspondences. CoRR, abs/2005.02357. 3", + "[6] Thomas Defard, Aleksandr Setkov, Angelique Loesch, and Romaric Audigier. Padim: A patch distribution modeling framework for anomaly detection and localization. In Alberto Del Bimbo, Rita Cucchiara, Stan Sclaroff, Giovanni Maria Farinella, Tao Mei, Marco Bertini, Hugo Jair Escalante, and Roberto Vezzani, editors, ICPR, volume 12664 of Lecture Notes in Computer Science, pages 475-489, 2020. Springer. 2, 3, 6", + "[7] Chouro Ding, Guansong Pang, and Chunhua Shen. Catching both gray and black swans: Open-set supervised anomaly detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 7378-7388. IEEE, 2022. 3", + "[8] Mariana-Iuliana Georgescu, Radu Tudor Ionescu, Fahad Shahbaz Khan, Marius Popescu, and Mubarak Shah. A background-agnostic framework with adversarial training for abnormal event detection in video. IEEE Trans. Pattern Anal. Mach. Intell., 44(9):4505-4523, 2022. 3", + "[9] Denis A. Gudovskiy, Shun Ishizaka, and Kazuki Kozuka. CFLOW-AD: real-time unsupervised anomaly detection with localization via conditional normalizing flows. In WACV, pages 1819–1828, 2022. IEEE. 2", + "[10] Dan Hendrycks, Mantas Mazeika, and Thomas G. Dietterich. Deep anomaly detection with outlier exposure. In 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019. OpenReview.net, 2019. 3", + "[11] Chaoqin Huang, Haoyan Guan, Aofan Jiang, Ya Zhang, Michael W. Spratling, and Yan-Feng Wang. Registration based few-shot anomaly detection. CoRR, abs/2207.07361, 2022. 3", + "[12] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In Francis R. Bach and David M. Blei, editors," + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Proceedings of the 32nd International Conference on Machine Learning, ICML 2015, Lille, France, 6-11 July 2015, volume 37 of JMLR Workshop and Conference Proceedings, pages 448-456. JMLR.org, 2015. 5", + "[13] Max Jaderberg, Karen Simonyan, Andrew Zisserman, and Koray Kavukcuoglu. Spatial transformer networks. In Corinna Cortes, Neil D. Lawrence, Daniel D. Lee, Masashi Sugiyama, and Roman Garnett, editors, Advances in Neural Information Processing Systems 28: Annual Conference on Neural Information Processing Systems 2015, December 7-12, 2015, Montreal, Quebec, Canada, pages 2017-2025, 2015. 3", + "[14] Chun-Liang Li, Kihyuk Sohn, Jinsung Yoon, and Tomas Pfister. Cutpaste: Self-supervised learning for anomaly detection and localization. In CVPR, pages 9664-9674, 2021. Computer Vision Foundation / IEEE. 2, 3, 6", + "[15] Tsung-Yi Lin, Priya Goyal, Ross B. Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. In CVPR, pages 2999-3007, 2017. IEEE Computer Society. 5", + "[16] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In ICLR, 2018. OpenReview.net. 8", + "[17] Prasanta Chandra Mahalanobis. On the generalized distance in statistics. National Institute of Science of India, 1936. 3", + "[18] Masoud PourReza, Bahram Mohammadi, Mostafa Khaki, Samir Bouindour, Hichem Snoussi, and Mohammad Sabokrou. G2D: generate to detect anomaly. In WACV, pages 2002–2011, 2021. IEEE. 2", + "[19] Karsten Roth, Latha Pemula, Joaquin Zepeda, Bernhard Scholkopf, Thomas Brox, and Peter Gehler. Towards total recall in industrial anomaly detection. In CVPR, pages 14318-14328, 2022. 2, 3", + "[20] Marco Rudolph, Bastian Wandt, and Bodo Rosenhahn. Same same but differnet: Semi-supervised defect detection with normalizing flows. In WACV, pages 1906-1915, 2021. IEEE. 2", + "[21] Marco Rudolph, Tom Wehrbein, Bodo Rosenhahn, and Bastian Wandt. Fully convolutional cross-scale-flows for image-based defect detection. In WACV, pages 1829-1838, 2022. IEEE. 2", + "[22] Mohammadreza Salehi, Hossein Mirzaei, Dan Hendrycks, Yixuan Li, Mohammad Hossein Rohban, and Mohammad Sabokrou. A unified survey on anomaly, novelty, open-set, and out-of-distribution detection: Solutions and future challenges. CoRR, abs/2110.14051. 1, 3", + "[23] Mohammadreza Salehi, Niousha Sadjadi, Soroosh Baselizadeh, Mohammad H. Rohban, and Hamid R. Rabiee. Multiresolution knowledge distillation for anomaly detection. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 14902-14912. Computer Vision Foundation / IEEE, 2021. 3, 6", + "[24] Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. Int. J. Comput. Vis., 128(2):336-359, 2020. 3" + ], + "bbox": [ + 503, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "3932", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Jianbing Shen, Ying Zhao, Shuicheng Yan, and Xuelong Li. Exposure fusion using boosting laplacian pyramid. IEEE Trans. Cybern., 44(9):1579-1590, 2014. 4", + "[26] Xian Tao, Xinyi Gong, Xin Zhang, Shaohua Yan, and Chandranath Adak. Deep learning for unsupervised anomaly localization in industrial images: A survey. CoRR, abs/2207.10298. 1, 3", + "[27] Chin-Chia Tsai, Tsung-Hsuan Wu, and Shang-Hong Lai. Multi-scale patch-based representation learning for image anomaly detection and segmentation. In WACV, pages 3065-3073, 2022. IEEE. 2", + "[28] Dmitry Ulyanov, Andrea Vedaldi, and Victor S. Lempitsky. Instance normalization: The missing ingredient for fast stylization. CoRR, abs/1607.08022, 2016. 5", + "[29] Zhou Wang, Alan C. Bovik, Hamid R. Sheikh, and Eero P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Trans. Image Process., 13(4):600-612, 2004. 4, 6", + "[30] Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon. CBAM: convolutional block attention module. In Vittorio Ferrari, Martial Hebert, Cristian Sminchisescu, and Yair Weiss, editors, Computer Vision - ECCV 2018 - 15th European Conference, Munich, Germany, September 8-14, 2018, Proceedings, Part VII, volume 11211 of Lecture Notes in Computer Science, pages 3-19. Springer, 2018. 5", + "[31] Jinjian Wu, Guangming Shi, Weisi Lin, Anmin Liu, and Fei Qi. Just noticeable difference estimation for images with free-energy principle. IEEE Trans. Multim., 15(7):1705-1710, 2013. 3, 4", + "[32] Jihun Yi and Sungroh Yoon. Patch SVDD: patch-level SVDD for anomaly detection and segmentation. In Hiroshi Ishikawa, Cheng-Lin Liu, Tomás Pajdla, and Jianbo Shi, editors, Computer Vision - ACCV 2020 - 15th Asian Conference on Computer Vision, Kyoto, Japan, November 30 - December 4, 2020, Revised Selected Papers, Part VI, volume 12627 of Lecture Notes in Computer Science, pages 375-390. Springer, 2020. 3, 6", + "[33] Zhiyuan You, Lei Cui, Yujun Shen, Kai Yang, Xin Lu, Yu Zheng, and Xinyi Le. A unified model for multi-class anomaly detection. CoRR, abs/2206.03687, 2022. 3, 6, 7", + "[34] Sangdoo Yun, Dongyoon Han, Sanghyuk Chun, Seong Joon Oh, Youngjoon Yoo, and Junsuk Choe. Cutmix: Regularization strategy to train strong classifiers with localizable features. In 2019 IEEE/CVF International Conference on Computer Vision, ICCV 2019, Seoul, Korea (South), October 27 - November 2, 2019, pages 6022-6031. IEEE, 2019. 3", + "[35] Vitjan Zavrtanik, Matej Kristan, and Danijel Skocaj. Dram - A discriminatively trained reconstruction embedding for surface anomaly detection. In ICCV, pages 8310-8319, 2021. IEEE. 2, 3, 6, 7, 8", + "[36] Vitjan Zavrtanik, Matej Kristan, and Danijel Skocaj. Reconstruction by inpainting for visual anomaly detection. Pattern Recognit., 112:107706, 2021. 2", + "[37] Vitjan Zavrtanik, Matej Kristan, and Danijel Skocaj. DSR - A dual subspace re-projection network for surface anomaly detection. In Shai Avidan, Gabriel J. Brostow, Moustapha Cisse, Giovanni Maria Farinella, and Tal Hassner, editors," + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Computer Vision - ECCV 2022 - 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXI, volume 13691 of Lecture Notes in Computer Science, pages 539-554. Springer, 2022. 3", + "[38] Ying Zhao. Just noticeable learning for unsupervised anomaly localization and detection. In ICME, pages In Press, 2022. 1, 2, 3, 4, 5, 6, 7, 8", + "[39] Yang Zou, Jongheon Jeong, Latha Pemula, Dongqing Zhang, and Onkar Dabeer. Spot-the-difference self-supervised pretraining for anomaly detection and segmentation. CoRR, abs/2207.14315, 2022. 2, 3, 5, 6, 7" + ], + "bbox": [ + 501, + 92, + 890, + 246 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "3933", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/OmniAL_ A Unified CNN Framework for Unsupervised Anomaly Localization/390b7405-6295-46ac-87cb-a78d4a95dd63_model.json b/2023/OmniAL_ A Unified CNN Framework for Unsupervised Anomaly Localization/390b7405-6295-46ac-87cb-a78d4a95dd63_model.json new file mode 100644 index 0000000000000000000000000000000000000000..dfef55783250412b7d826e24e2c80a944d9ed059 --- /dev/null +++ b/2023/OmniAL_ A Unified CNN Framework for Unsupervised Anomaly Localization/390b7405-6295-46ac-87cb-a78d4a95dd63_model.json @@ -0,0 +1,2090 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.105, + 0.131, + 0.867, + 0.152 + ], + "angle": 0, + "content": "OmniAL: A unified CNN framework for unsupervised anomaly localization" + }, + { + "type": "text", + "bbox": [ + 0.28, + 0.182, + 0.69, + 0.217 + ], + "angle": 0, + "content": "Ying Zhao \nRicoh Software Research Center (Beijing) Co., Ltd." + }, + { + "type": "text", + "bbox": [ + 0.365, + 0.22, + 0.601, + 0.234 + ], + "angle": 0, + "content": "zy_deepwhite.zy@hotmail.com" + }, + { + "type": "image", + "bbox": [ + 0.078, + 0.241, + 0.291, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.097, + 0.425, + 0.273, + 0.439 + ], + "angle": 0, + "content": "(a). N models for N classes" + }, + { + "type": "image", + "bbox": [ + 0.336, + 0.242, + 0.553, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.345, + 0.425, + 0.533, + 0.439 + ], + "angle": 0, + "content": "(b). One model for N classes" + }, + { + "type": "image", + "bbox": [ + 0.598, + 0.242, + 0.895, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.667, + 0.425, + 0.813, + 0.439 + ], + "angle": 0, + "content": "(c). Proposed OmniAL" + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.446, + 0.893, + 0.488 + ], + "angle": 0, + "content": "Figure 1. OmniAL is a unified framework for unsupervised anomaly localization. As shown in (c), OmniAL not only simplifies existing learning paradigm from (a) Separate: N models for N classes to (b) Unified: one model for N classes but also produces high quality anomaly localization results. Results in (a) and (b) are from separate and unified models of JNLD [38] on MVtecAD [1] dataset." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.505, + 0.314, + 0.52 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.539, + 0.474, + 0.903 + ], + "angle": 0, + "content": "Unsupervised anomaly localization and detection is crucial for industrial manufacturing processes due to the lack of anomalous samples. Recent unsupervised advances on industrial anomaly detection achieve high performance by training separate models for many different categories. The model storage and training time cost of this paradigm is high. Moreover, the setting of one-model-N-classes leads to fearful degradation of existing methods. In this paper, we propose a unified CNN framework for unsupervised anomaly localization, named OmniAL. This method conquers aforementioned problems by improving anomaly synthesis, reconstruction and localization. To prevent the model learning identical reconstruction, it trains the model with proposed panel-guided synthetic anomaly data rather than directly using normal data. It increases anomaly reconstruction error for multi-class distribution by using a network that is equipped with proposed Dilated Channel and Spatial Attention (DCSA) blocks. To better localize the anomaly regions, it employs proposed DiffNeck between reconstruction and localization sub-networks to explore multi-level differences. Experiments on 15-class MVTecAD and 12-class VisA datasets verify the advantage of proposed OmniAL that surpasses the state-of-the-art of unified models. On 15-class-MVTecAD/12-class-VisA, its single unified" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.506, + 0.895, + 0.628 + ], + "angle": 0, + "content": "model achieves 97.2/87.8 image-AUROC, 98.3/96.6 pixel-AUROC and 73.4/41.7 pixel-AP for anomaly detection and localization respectively. Besides that, we make the first attempt to conduct a comprehensive study on the robustness of unsupervised anomaly localization and detection methods against different level adversarial attacks. Experiential results show OmniAL has good application prospects for its superior performance." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.648, + 0.631, + 0.663 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.895, + 0.903 + ], + "angle": 0, + "content": "In real industrial scenarios, the location of anomaly [22, 26] reveals important information, such as defective types and degrees. It is essential not only to inspect whether a sample is defective but also to know where the specific anomaly regions are. Since anomaly appearance is inexhaustible, it is almost impossible and infeasible to collect and manually annotate all kinds of abnormal data. Thus, only normal samples are available for training a detector that is robust enough to find out unseen anomalies during inference phase. Considering the diversity of classes and various types of one class, the conventional training paradigm of N models for N classes, as shown in Fig.1a, may not be the best solution. The model storage and training time cost increase with the number of classes. As shown in Fig.1b, existing method severely degrades anomaly local-" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3924" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.094, + 0.231, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.087, + 0.291, + 0.23, + 0.316 + ], + "angle": 0, + "content": "(a). Reconstruction-based anomaly localization" + }, + { + "type": "image_caption", + "bbox": [ + 0.229, + 0.093, + 0.296, + 0.103 + ], + "angle": 0, + "content": "(1). MVtecAD" + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.106, + 0.292, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.306, + 0.093, + 0.403, + 0.103 + ], + "angle": 0, + "content": "(2).JNLD Rec./Loc." + }, + { + "type": "image", + "bbox": [ + 0.238, + 0.19, + 0.294, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.3, + 0.106, + 0.407, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.301, + 0.191, + 0.402, + 0.203 + ], + "angle": 0, + "content": "X Rec. X Loc." + }, + { + "type": "image", + "bbox": [ + 0.238, + 0.203, + 0.294, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.297, + 0.204, + 0.406, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.297, + 0.302, + 0.404, + 0.312 + ], + "angle": 0, + "content": "sis unified models o" + }, + { + "type": "image_caption", + "bbox": [ + 0.413, + 0.093, + 0.526, + 0.103 + ], + "angle": 0, + "content": "(3). OmniRec./Loc." + }, + { + "type": "image", + "bbox": [ + 0.411, + 0.105, + 0.52, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.415, + 0.191, + 0.516, + 0.203 + ], + "angle": 0, + "content": "V Rec. V Loc." + }, + { + "type": "image", + "bbox": [ + 0.411, + 0.204, + 0.519, + 0.29 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.415, + 0.29, + 0.515, + 0.301 + ], + "angle": 0, + "content": "V Rec. V Loc." + }, + { + "type": "image_caption", + "bbox": [ + 0.552, + 0.093, + 0.593, + 0.103 + ], + "angle": 0, + "content": "(4).VisA" + }, + { + "type": "image", + "bbox": [ + 0.542, + 0.106, + 0.611, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.548, + 0.191, + 0.612, + 0.203 + ], + "angle": 0, + "content": "" + }, + { + "type": "image_caption", + "bbox": [ + 0.629, + 0.093, + 0.726, + 0.103 + ], + "angle": 0, + "content": "(5). JNLD Rec. / Loc." + }, + { + "type": "image", + "bbox": [ + 0.616, + 0.106, + 0.752, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.627, + 0.191, + 0.676, + 0.203 + ], + "angle": 0, + "content": "X Rec." + }, + { + "type": "image", + "bbox": [ + 0.616, + 0.204, + 0.752, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.626, + 0.289, + 0.674, + 0.3 + ], + "angle": 0, + "content": "V Rec." + }, + { + "type": "image_caption", + "bbox": [ + 0.616, + 0.301, + 0.836, + 0.314 + ], + "angle": 0, + "content": "Analysis unified models on VisA dataset." + }, + { + "type": "image_caption", + "bbox": [ + 0.768, + 0.094, + 0.88, + 0.104 + ], + "angle": 0, + "content": "(6). OmniRec./Loc." + }, + { + "type": "image", + "bbox": [ + 0.756, + 0.106, + 0.892, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.764, + 0.191, + 0.813, + 0.202 + ], + "angle": 0, + "content": "Rec." + }, + { + "type": "image_caption", + "bbox": [ + 0.83, + 0.191, + 0.879, + 0.202 + ], + "angle": 0, + "content": "√Loc." + }, + { + "type": "image", + "bbox": [ + 0.756, + 0.204, + 0.892, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.764, + 0.289, + 0.813, + 0.3 + ], + "angle": 0, + "content": "V Rec." + }, + { + "type": "image_caption", + "bbox": [ + 0.83, + 0.289, + 0.879, + 0.299 + ], + "angle": 0, + "content": "√ Loc." + }, + { + "type": "image_caption", + "bbox": [ + 0.212, + 0.321, + 0.756, + 0.335 + ], + "angle": 0, + "content": "Figure 2. Problem analysis. The final failure may be caused by reconstruction and localization." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.344, + 0.469, + 0.404 + ], + "angle": 0, + "content": "ization performance if the training paradigm changes to one model for N classes. Therefore, a robust unified framework for unsupervised anomaly localization is highly demanded for intelligent industrial." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.409, + 0.469, + 0.575 + ], + "angle": 0, + "content": "With the limitation of available training data, many appealing unsupervised approaches [14, 18, 35, 38] using synthesized anomaly data are proposed. These approaches generate anomalous instances to inspire the anomaly detector to learn discriminative features. Their experiments show that the realiziness of generated anomalous instances had a strong impact on the quality of anomaly localization. However, none of these methods consider the training paradigm of one model for N classes. When switching to the unified training paradigm, they are more prone to learn an identical short-cut and fail to discriminate the anomaly." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.58, + 0.469, + 0.866 + ], + "angle": 0, + "content": "With the normal and synthesized anomalous samples, recent unsupervised learning methods train a deep anomaly detector by either a distance-based [6, 14, 19-21, 27] or reconstruction-based [2, 9, 35, 36, 38] way. The reconstruction-based architectures [35, 38] are supposed to reconstruct normal images more accurately than the unseen anomalous. The anomaly localization is then calculated from the reconstruction error between the original and reconstructed versions of the input image, as shown in Fig.2a. The prediction of anomaly location is not only based on the reconstruction quality but also the ability of spotting the reconstruction error. The typical reconstruction-based method JNLD [38] learns a joint representation of an anomalous image and its anomaly-free reconstruction, while simultaneously learning a decision boundary between normal and simulated anomalous examples. As shown in Fig.2b and Fig.2c, under the unified setting, JNLD [38] fails to produce correct results either because of the reconstruction failure or the localization failure." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.901 + ], + "angle": 0, + "content": "To conquer aforementioned problems, we propose a novel unified framework OmniAL for effectively localiz-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.344, + 0.892, + 0.614 + ], + "angle": 0, + "content": "ing anomaly pixels of different classes only by using a single model. OmniAL uses a panel-guided anomaly synthesis method that controls the portion of normal and anomaly regions for each training sample. By doing this, OmniAL blocks the chance of learning identical shortcut from the source. To increase the anomaly reconstruction error for multi-class distribution, OmniAL constructs a reconstruction and a localization sub-networks that are equipped with proposed Dilated Channel and Spatial Attention (DCSA) blocks. To better localize the anomaly regions, OmniAL employs a DiffNeck module between the reconstruction and localization sub-networks to explore multi-level reconstruction errors. As shown in Fig.1 and Fig.2, OmniAL learns a single unified model for multiple classes that produces high quality reconstruction and precise anomaly localization. Furthermore, we conduct an exhaustive evaluation of reconstruction and localization performance against multi-level adversarial attacks." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.616, + 0.866, + 0.631 + ], + "angle": 0, + "content": "In summary, we make following main contributions:" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.64, + 0.892, + 0.744 + ], + "angle": 0, + "content": "- We construct a unified CNN framework OmniAL for unsupervised anomaly localization that is equipped with proposed panel-guided anomaly synthesis, DCSA block, and DiffNeck module. OmniAL achieves superior performance for anomaly localization on challenging MVTecAD [1] and VisA [39] datasets compared to the state-of-the-art." + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.756, + 0.892, + 0.846 + ], + "angle": 0, + "content": "- By preventing model from learning identical reconstruction, our proposed panel-guided anomaly synthesis method also brings substantial improvement for existing methods under the unified setting. It boosts the image-AUROC/pixel-AUROC/pixel-AP from 88.7/87.1/49.4 to 92.5/94.5/57.4 for Draem [35]." + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.856, + 0.892, + 0.901 + ], + "angle": 0, + "content": "- We make a comprehensive study on the robustness of separate/unified anomaly localization methods against different level adversarial attacks. Our synthesized" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.64, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.515, + 0.957 + ], + "angle": 0, + "content": "3925" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.892, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.277, + 0.894, + 0.349 + ], + "angle": 0, + "content": "Figure 3. Framework of OmniAL. It consists of panel-guided anomaly synthesis, reconstruction and localization. Anomaly synthesis is based on anomaly panel and three variants of the Just Noticeable Distortion (JND) map. The synthetic anomaly is reconstructed into normal image and corresponding JND map by the Dilated Channel Spatial Attention (DCSA) modules equipped reconstruction sub-network. The localization sub-network with a DiffNeck module localizes the anomaly regions by exploring the difference between reconstructed and original data." + }, + { + "type": "text", + "bbox": [ + 0.108, + 0.355, + 0.471, + 0.415 + ], + "angle": 0, + "content": "adversarial datasets exhibit strong attack capability against anomaly detection, reconstruction and localization, also helping to analyse the risks of existing methods." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.434, + 0.22, + 0.451 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.462, + 0.471, + 0.824 + ], + "angle": 0, + "content": "Anomaly synthesis. Due to the lack of anomaly samples, unsupervised learning methods are commonly used in industrial quality inspection. These methods are typically trained with normal data and synthesized anomalous that highly effect the performance. Therefore, how to synthesize the anomalous [7, 8, 14, 35, 37, 38] also draws extensive attention. CutPaste [14] learns representations by classifying normal data from the Cut-and-Paste augmentation. SPD [39] uses a smoothed version of CutPaste [14] augmentation. Instead of using simple regular shaped anomalous, Draem [35] simulates just-out-of-distribution anomalies having random shape and texture. To simulate photorealistic anomaly samples, JNLD [38] proposes a multiscale noticeable anomalous generation method based on just noticeable distortion [31]. The open-set supervised anomaly detection method DRA [7] adapts the popular CutMix [34] and the outlier exposure method [10] to generate pseudo anomalies from normal images for training. DSR [37] generates the anomalies at the feature level by sampling the learned quantized feature space, which allows a controlled generation of near-in-distribution anomalies. For the unified paradigm, we propose a panel-guided anomaly synthesis method that controls the portion of normal and anomaly regions for each training sample." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Anomaly localization. Anomaly detection [22,26], also known as outlier detection or one-class classification, refers to the task of distinguishing defective image at the image-level from the majority of anomaly-free images. Anomaly localization (segmentation), on the other hand, aims to seg" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.355, + 0.892, + 0.732 + ], + "angle": 0, + "content": "ment out the pixel-level anomaly regions. For industrial visual inspection, many recent methods [5, 6, 19] achieve high performance in anomaly detection but anomaly localization. SPADE [5] detects anomaly based on alignment between an anomalous image and a constant number of the similar normal images. It relies on K nearest neighbors of pixel-level feature pyramids extracted by pre-trained deep features. PSVDD [32] extends deep Support Vector Data Description (SVDD) to a patch-wise detection method. PaDim [6] also relies on ImageNet pretrained feature extractor with multiscale pyramid pooling. Instead of using time-costly clustering, it uses a well-known Mahalanobis distance metric [17] as an anomaly score. CutPaste [14] uses GradCAM [24] to get the defect localization. MKD [23] proposes to use multi-level features alignment to increase the discriminating capability of the Teacher-Student model on various types of abnormalities. PatchCore [19] combines patch-level embeddings from ImageNet models with an outlier detection model. However, PatchCore is inherently not suitable for unified setting (larger normal dataset) since it needs to build the coreset with all normal data. Draem [35] and JNLD [38] use segmentation sub-network to predict the defective regions. This paper differs from these previous works by focusing on the paradigm of only using a single unified model for N classes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.893, + 0.903 + ], + "angle": 0, + "content": "Unified anomaly localization. Recently, how to use a unified model to localize anomaly for different objects has already become researchers' concern. RegAD [11] trains a single generalizable model for few-shot anomaly detection, where a limited number of normal images are provided for each category at training. It employs SimSiam [3] with three spatial transformer network [13] blocks to solve the category-agnostic feature registration proxy task. It identifies anomalies by comparing the registered features of the test image and the corresponding normal images. UniAD [33] constructs a transformer with a layer-wise query de" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3926" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.213 + ], + "angle": 0, + "content": "coder and a neighbor masked attention module to model the multi-class distribution. It uses the layer-wise query decoder to intensify the use of query embedding. To avoid the information leak, it employs a neighbor masked attention module, where a feature point relates to neither itself nor its neighbors. In this paper, we propose a unified CNN framework for unsupervised anomaly localization. It manifests the problem-solving ability of fully convolutional networks." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.231, + 0.169, + 0.246 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.257, + 0.473, + 0.501 + ], + "angle": 0, + "content": "In this work, we construct a unified CNN framework OmniAL for unsupervised anomaly localization under the practical setting of only using one model for N classes. Fig.3 shows the overview of proposed OmniAL. To train only with normal data, we firstly synthesize photo-realistic anomalous in the regions that are selected based 5 predefined panels. The synthesized anomaly samples are then used to train the OmniAL network that consists of an anomaly-free reconstruction and an anomaly localization sub-networks. The reconstruction and localization sub-networks are composed of alternative basic blocks and DCSA blocks. The reconstruction sub-network learns to recover the synthesized anomalous back to normal. The localization sub-network with a DiffNeck module localizes the anomaly regions by exploring the difference between reconstructed and original data." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.513, + 0.358, + 0.529 + ], + "angle": 0, + "content": "3.1. Panel-guided anomaly synthesis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.538, + 0.47, + 0.749 + ], + "angle": 0, + "content": "Reconstruction-based anomaly localization methods rely on the hypothesis that the reconstruction errors in unseen anomalous regions are larger than the normal regions. The reconstruction models are trained with alternative normal and synthetic anomaly samples. For normal sample, the supervision ground truth is the direct copy of the input. Thus, directly using normal samples for training increases the chance of information leak and leads to identical reconstruction for any input. To prevent the learning of identical short-cut, the intuitive idea is that only using synthetic anomaly rather than normal data. However, without using normal samples for training leads to more false alarm. Then, the problem is transformed into how to better synthesis anomaly for unified model training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.471, + 0.902 + ], + "angle": 0, + "content": "To solve the problem, we propose a panel-guided anomaly synthesis method that takes into account both normal and anomaly for training. As shown in Fig.4, we build 5 types of panels ('left', 'right', 'top', 'bottom', 'all') to control the portion of normal and anomaly regions. That is, we only generate the anomaly in one of the panel region for each sample in each iteration. Except the type of 'all', the area of panel region is also randomly adjusted in range of [0.5, 0.8]*ImageWidth and [0.5, 0.8]*ImageHeight. Given a panel, we synthesize anomaly in the selected region ac" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.089, + 0.895, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.242, + 0.892, + 0.283 + ], + "angle": 0, + "content": "Figure 4. Visualization of our panel-guided anomaly synthetic data. The anomaly source is selected either from augmented normal images (row1,2) or DTD [4] dataset (row3)." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.291, + 0.704, + 0.306 + ], + "angle": 0, + "content": "cording to following equation." + }, + { + "type": "equation", + "bbox": [ + 0.597, + 0.309, + 0.892, + 0.348 + ], + "angle": 0, + "content": "\\[\nA = \\sum_ {i = 1} ^ {3} \\left(W _ {i} T + (1 - W _ {i}) I\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.35, + 0.892, + 0.456 + ], + "angle": 0, + "content": "Where \\( A \\) is the synthesized anomaly image, \\( W_{i} \\) is the pixelwise fusion weight map, \\( i \\) indicates one of the 3 defect levels (easy, medium and hard), \\( I \\) is the anomaly-free image, \\( T \\) is the anomaly source image. To synthesize natural and diversity anomalies, \\( T \\) is generated by applying a set of random augmentation (resize, crop, flip, color-jitter) to \\( I \\) or image from the Describable Textures Dataset(DTD) [4]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.457, + 0.893, + 0.622 + ], + "angle": 0, + "content": "To get the 3-level defects, we first need to find out the baseline anomaly fusion weight map. Following JNLD [38], we use the JND [31] to guide the anomaly synthesis. JND [31] refers to the minimum visibility threshold of the Human Visual System, is useful in perceptual image/video processing systems. It reveals a perceptual threshold of intensity change in an image that can be noticed by the human vision system. We use a nonlinear additivity JND model to define the anomaly weight map \\(J\\) as [25] that proposes a novel hybrid exposure weight measurement using the JND. Each level anomaly fusion weight map \\(W_{i}\\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.627, + 0.637, + 0.891, + 0.654 + ], + "angle": 0, + "content": "\\[\nW _ {i} = \\alpha_ {i} * (J + \\beta_ {i}) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.658, + 0.892, + 0.84 + ], + "angle": 0, + "content": "Where \\( J \\) indicates the JND map, \\( \\beta_{i} \\) is a bias term for different level of noticeable, \\( \\alpha_{i} \\) is a smooth kernel applied to the mask and produces soft boundary between anomaly and normal regions. For the easier level anomaly, we use a smaller smooth kernel. Due to using random seeds, different level anomaly regions may overlap each other. Thus, directly using the masks as the ground truth for anomaly segmentation task may bring label inconsistency. Therefore, we refine and quantify the segmentation mask into 3 defect levels based on the structure similarity index measure (SSIM) [29] between the anomaly-free image \\( I \\) and synthesized anomaly image \\( A \\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.848, + 0.722, + 0.864 + ], + "angle": 0, + "content": "3.2. Anomaly reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "As shown in Fig.3, our anomaly reconstruction subnetwork receives an input image and outputs reconstructed" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3927" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.422 + ], + "angle": 0, + "content": "normal image and corresponding JND map. The reconstruction sub-network consists of an encoder and a decoder. Both of them are composed of alternating basic and DCSA blocks. The basic block extracts different levels of features with two consecutive \\(3 \\times 3\\) convolution layer followed by normalization and ReLU activation layers. Considering the unified training paradigm, instance normalization [28] is more suitable than the broadly used batch normalization [12] in anomaly reconstruction. The reason is that batch normalization [12] normalizes using the information from the whole batch, while instance normalization [28] normalizes each feature map on its own. Under unified setting, each batch of training data consists of different classes that complicates the reconstruction problem. Instance normalization discards instance-specific contrast information and reduces undesirable instance-level variation. Batch normalization, meanwhile, lacks the ability to address the class inconsistency. Therefore, we use instance normalization in each basic block rather than batch normalization. Experimental results in Table 3 show that the performance is improved by swapping batch normalization with instance normalization." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.425, + 0.473, + 0.728 + ], + "angle": 0, + "content": "The feature maps extracted by the basic block are further integrated by the DCSA block that is inspired by the popular CBAM [30]. As shown in orange dash block of Fig.3, the DCSA block consists of channel and spatial attention sequential sub-modules. With average pooling and two consecutive 1x1 convolutions, the channel attention explores the correlations among feature channels. The correlations are add back to original feature maps to emphasize the important feature channels. The spatial attention further investigates the spatial correlation. It first aggregates spatial information by concatenating both average and maximum pooling of afore refined feature maps. Then, it spots the spatial importance of the aggregated features with a standard and a dilated \\(3 \\times 3\\) convolution-norm-relu paths. It balances the portion of dilated path by a factor that is set as 0.1. The spatial attention map is obtained by summing up two paths of features. By adding the spatial attention map, the important regions in feature maps are highlighted. As shown in Table 3, the performance is further improved by using DCSA blocks." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.741, + 0.277, + 0.756 + ], + "angle": 0, + "content": "3.3. Anomaly localization" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.765, + 0.47, + 0.902 + ], + "angle": 0, + "content": "After reconstructing the normal and corresponding JND maps, we use a localization sub-network to highlight the anomaly regions. Instead of directly concatenating the reconstructed maps with input image, we firstly use a DiffNeck module to fully explore the difference between them. The pink dash block of Fig.3 illustrates the detailed structure of DiffNeck. Overall, DiffNeck extracts reconstruction difference among reconstructed normal map (Rec), reconstructed JND map (JND) and input image (Inp) in both sep" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.196 + ], + "angle": 0, + "content": "arate and unified ways. That is, DiffNeck uses basic blocks to extract the concatenated features of (Rec, Inp), (JND, Inp) and (Rec, JND, Inp) respectively. Then, all levels of differences are summed up and form a final descriptor that is fed to the localization sub-network. DiffNeck also balances the portion of the separate path by a factor that is set as 0.1." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.197, + 0.895, + 0.454 + ], + "angle": 0, + "content": "Similar with reconstruction sub-network, the localization sub-network also follows the encoder-decoder structure but with 6 scales (e1-e5 in Fig.3) and skip connections for corresponding scales. Scale-1 and scale-2 use standard convolution in all blocks. To use larger resolution feature maps and receptive field, the basic and DCSA blocks of the encoder contain dilated convolution layer from scale-3 to scale-6 with 2, 4, 8, 16 dilation rates respectively. Different with reconstruction sub-network, we use batch normalization in all blocks. Since we synthesize more realistic anomalies by considering easy, medium and hard level defects, localization sub-network predicts 3 levels of anomalous locations. To get the final pixel-level anomaly localization map, we combine them together by a sequence processing of SoftMax, smoothness and summing up. By doing this, the false alarm noise can be suppressed, and the weak true decision can be strengthened." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.455, + 0.851, + 0.469 + ], + "angle": 0, + "content": "The total target for training OmniAL is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.542, + 0.472, + 0.892, + 0.507 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L = L _ {2} \\left(J, J _ {r}\\right) + L _ {s s i m} \\left(J, J _ {r}\\right) \\tag {3} \\\\ + L _ {2} (I, I _ {r}) + L _ {s s i m} (I, I _ {r}) + L _ {f l} (S, S _ {s}) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.509, + 0.895, + 0.646 + ], + "angle": 0, + "content": "Where \\( J \\) is the ground truth JND map calculated from the anomaly-free image \\( I \\), \\( J_{r} \\) and \\( I_{r} \\) are reconstructed JND map and normal image, \\( S \\) is the anomaly localization ground truth. Unlike JNLD [38], for both JND and normal reconstruction, we not only use MSE loss to supervise the pixel-to-pixel recovering but also the SSIM loss to yield plausible local consistency. To handle the unbalance of different types, we use focal loss [15] to supervise the predicted anomaly localization \\( S_{s} \\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.657, + 0.702, + 0.674 + ], + "angle": 0, + "content": "4. Experimental results" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.682, + 0.697, + 0.697 + ], + "angle": 0, + "content": "4.1. Datasets and metrics" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.705, + 0.892, + 0.75 + ], + "angle": 0, + "content": "To demonstrate the effectiveness of proposed OmniAL, we conduct extensive experiments on the challenging MVTecAD [1] and VisA [39] datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.902 + ], + "angle": 0, + "content": "MVTecAD [1] dataset contains 10 object and 5 texture industrial products, such as bottle and leather. It consists of 3,629 normal images for training and 1,725 images for testing. There are 1,258 anomaly images of the testing set with pixel-level labelled various types of defects and the rest are normal images. Each class contains 60 to 320 color images with the resolution ranges from \\(700 \\times 700\\) to \\(1024 \\times 1024\\) pixels. In the testing set, defective appearance varies in different sizes, shapes and types, and most cases only contain a small fraction of anomalous pixels." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3928" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.176, + 0.09, + 0.794, + 0.103 + ], + "angle": 0, + "content": "Table 1. Image AUROC comparison of our method OmniAL with the state-of-the-art on MVTecAD [1]." + }, + { + "type": "table", + "bbox": [ + 0.084, + 0.104, + 0.89, + 0.328 + ], + "angle": 0, + "content": "
CategoryPadim [6]CutPaste [14]MKD [23]Draem [35]JNLD [38]UniAD [33]OmniAL
Unified / Separate
bottle97.9 / 99.967.9 / 98.298.7 / 99.494.6 / 99.199.1 / 96.399.7 / 100100 / 99.4
cable70.9 / 92.769.2 / 81.278.2 / 89.261.8 / 94.790.6 / 98.895.2 / 97.698.2 / 97.6
capsule73.4 / 91.363.0 / 98.268.3 / 80.570.2 / 98.574.4 / 85.086.9 / 85.395.2 / 92.4
carpet93.8 / 99.893.6 / 93.969.8 / 79.395.9 / 95.577.1 / 97.899.8 / 99.998.7 / 99.6
grid73.9 / 96.793.2 / 10083.8 / 98.098.1 / 99.998.6 / 10098.2 / 98.599.9 / 100
hazelnut85.5 / 92.080.9 / 98.397.1 / 98.495.1 / 10090.8 / 10099.8 / 99.995.6 / 98.0
leather99.9 / 10093.4 / 10093.6 / 95.199.9 / 10097.0 / 100100 / 10099.0 / 97.6
metal nut88.0 / 98.760.0 / 99.964.9 / 73.688.9 / 98.793.3 / 99.699.2 / 99.099.2 / 99.9
pill68.8 / 93.371.4 / 94.979.7 / 82.769.0 / 98.982.7 / 94.693.7 / 88.397.2 / 97.7
screw56.9 / 85.885.2 / 88.775.6 / 83.393.3 / 93.981.8 / 95.887.5 / 91.988.0 / 81.0
tile93.3 / 98.188.6 / 94.689.5 / 91.698.3 / 99.699.2 / 10099.3 / 99.099.6 / 100
toothbrush95.3 / 96.163.9 / 99.475.3 / 92.282.8 / 100100 / 10094.2 / 95.0100 / 100
transistor86.6 / 97.457.9 / 96.173.4 / 85.683.9 / 93.190.3 / 93.099.8 / 10093.8 / 93.8
wood72.1 / 96.580.4 / 99.193.4 / 94.399.8 / 99.191.9 / 99.698.6 / 97.993.2 / 98.7
zipper79.7 / 90.393.5 / 99.987.4 / 93.299.1 / 10099.8 / 99.895.8 / 96.7100 / 100
average84.2 / 95.577.5 / 96.181.9 / 87.888.7 / 98.091.3 / 97.496.5 / 96.697.2 / 97.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.179, + 0.335, + 0.79, + 0.348 + ], + "angle": 0, + "content": "Table 2. Pixel AUROC comparison of our method OmniAL with the state-of-the-art on MVTecAD [1]." + }, + { + "type": "table", + "bbox": [ + 0.084, + 0.35, + 0.89, + 0.574 + ], + "angle": 0, + "content": "
CategoryPadim [6]PSVDD [32]MKD [23]Draem [35]JNLD [38]UniAD [33]OmniAL
Unified / Separate
bottle96.1 / 98.286.7 / 98.191.8 / 96.387.4 / 99.194.8 / 99.098.1 / 98.199.2 / 99.0
cable81.0 / 96.762.2 / 96.889.3 / 82.470.4 / 94.776.4 / 97.797.3 / 96.897.3 / 97.1
capsule96.9 / 98.683.1 / 95.888.3 / 95.949.2 / 94.357.0 / 92.798.5 / 97.996.9 / 92.2
carpet97.6 / 99.078.6 / 92.695.5 / 95.695.2 / 95.593.7 / 99.098.5 / 98.099.4 / 99.6
grid71.0 / 97.170.8 / 96.282.3 / 91.899.0 / 99.796.9 / 99.798.2 / 98.599.4 / 99.6
hazelnut96.3 / 98.197.4 / 97.591.2 / 94.696.0 / 99.785.9 / 99.496.5 / 94.698.4 / 98.6
leather84.8 / 99.093.5 / 97.496.7 / 98.198.6 / 98.687.0 / 99.598.8 / 98.399.3 / 99.7
metal nut84.8 / 97.396.0 / 98.064.2 / 86.472.6 / 99.597.4 / 99.594.8 / 95.799.1 / 99.1
pill87.7 / 95.796.5 / 95.169.7 / 89.690.0 / 97.691.2 / 96.695.0 / 95.198.9 / 98.6
screw94.1 / 98.474.3 / 95.792.1 / 96.089.3 / 97.687.0 / 99.798.3 / 97.498.0 / 97.2
tile80.5 / 94.192.1 / 91.485.3 / 82.898.1 / 99.294.7 / 99.691.8 / 91.899.0 / 99.4
toothbrush95.6 / 98.898.0 / 98.188.9 / 96.194.4 / 98.198.6 / 98.898.4 / 97.899.4 / 99.2
transistor92.3 / 97.678.5 / 97.071.7 / 76.573.1 / 90.983.6 / 92.197.9 / 98.793.3 / 91.7
wood89.1 / 94.180.7 / 90.880.5 / 84.896.2 / 96.488.7 / 96.393.2 / 93.497.4 / 96.9
zipper94.8 / 98.495.1 / 95.186.1 / 93.996.9 / 98.895.3 / 99.496.8 / 96.099.5 / 99.7
average89.5 / 97.485.6 / 95.784.9 / 97.087.1 / 97.388.6 / 97.996.8 / 96.698.3 / 97.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.592, + 0.47, + 0.674 + ], + "angle": 0, + "content": "Table 3. Ablation study on MVTecAD [1]. PA: Panel-guided anomaly synthesis, Rec: Reconstruction sub-network, Seg: Segmentation sub-network, BB: Basic Block, BN: Batch Norm, IN: Instance Norm, DC: Dilated Convolution, CA: Channel Attention, DSA: Dilated Spatial Attention, I: Image-level classification, P: Pixel-level localization." + }, + { + "type": "table", + "bbox": [ + 0.081, + 0.675, + 0.486, + 0.747 + ], + "angle": 0, + "content": "
PARec-BBDCSASeg-BBDiffNeckI-AUROCP-AUROCP-AP
-BNINDCCADSABNDC-Unified/Separate
-√--√-√--86.7/98.886.4/98.444.2/75.0
√√--√-√--94.1/—95.9/—68.3/—
√-√-√-√--95.8/—97.0/—69.1/—
√-√√√-√√-96.9/97.397.6/97.772.8/72.9
√-√√√√√√-94.7/96.596.9/97.869.8/74.9
√-√√√-√√√96.8/96.398.2/98.072.7/74.1
√-√√√√√√√97.2/97.098.3/97.873.4/73.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.755, + 0.47, + 0.783 + ], + "angle": 0, + "content": "Table 4. Ablation study on With/Without panel-guided anomaly synthesis for unified training on MVTecAD [1]." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.784, + 0.485, + 0.852 + ], + "angle": 0, + "content": "
Draem [35]JNLD [38]OmniAL
I-AUROC92.5/88.792.9/91.397.2/86.6
P-AUROC94.5/87.195.6/88.698.3/93.7
Pixel-AP57.4/49.463.1/46.673.4/54.4
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.903 + ], + "angle": 0, + "content": "VisA [39] dataset consists of 10,821 high-resolution color images (9,621 normal and 1,200 anomalous samples) covering 12 objects in 3 domains, including complex struc" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.594, + 0.892, + 0.732 + ], + "angle": 0, + "content": "ture, multiple instances and single instances. The anomalous images contain various flaws, including surface defects such as scratches, dents, color spots or crack, and structural defects like misplacement or missing parts. There are 5-20 images per defect type and an image may contain multiple defects. All images were acquired using a \\(4,000 \\times 6,000\\) high-resolution RGB sensor. Example of each category and our corresponding anomaly reconstruction and localization results are shown in Fig.5a(MVTecAD) and Fig.5b(VisA)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Metrics For anomaly detection evaluation, the most common used metrics are Area Under the Receiver Operating Characteristic curve (AUROC) in both image-level and pixel-level. However, the pixel-AUROC is biased in favour of large anomalous regions and does not well reflect the pixel-level anomalous localization performance. Therefore, we additionally introduce the average precision (AP) to evaluate pixel-level anomaly localization performance. For reconstruction quality evaluation, we use peak signal to noise ratio (PSNR) and SSIM [29] that are commonly used for evaluating signal fidelity." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3929" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.142, + 0.072, + 0.827, + 0.086 + ], + "angle": 0, + "content": "Table 5. Performance comparison of our method OmniAL with the state-of-the-art on VisA [39]. (Separate/Unified)" + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.087, + 0.891, + 0.264 + ], + "angle": 0, + "content": "
CategoryPadim+SPD [39]Draem [35]JNLD [38]OmniAL
I-AUCP-AUCI-AUCP-AUCP-API-AUCP-AUCP-API-AUCP-AUCP-AP
Complex structurePCB192.797.771.3/83.998.6/94.060.4/38.582.0/82.996.4/98.072.8/77.896.6/77.798.7/97.663.5/76.9
PCB287.997.289.7/81.792.5/94.13.5/13.396.3/79.191.9/95.031.2/37.799.4/81.083.2/93.92.8/31.6
PCB385.496.773.1/87.793.8/94.118.7/17.996.9/90.195.3/98.543.4/46.896.9/88.198.4/94.756.9/41.4
PCB499.189.291.3/87.195.8/72.332.7/13.194.8/96.296.1/97.537.4/29.797.4/95.398.5/97.138.4/33.2
Multiple instancesMacaroni185.798.870.3/68.695.8/89.88.2/8.094.3/90.598.8/93.325.9/14.296.9/92.698.9/98.67.6/7.1
Macaroni270.896.071.3/60.394.1/83.225.4/19.786.5/71.392.9/92.117.2/6.189.9/75.299.1/97.911.4/9.2
Capsules68.186.377.3/89.693.7/96.620.2/30.189.1/91.498.9/99.638.6/55.687.9/90.698.6/99.462.9/52.4
Candles89.197.382.3/70.287.0/82.627.9/12.689.1/85.494.8/94.525.9/27.485.1/86.890.5/95.829.2/24.6
Single instanceCashew90.586.194.2/67.394.7/68.541.2/7.096.0/82.596.3/94.143.7/39.497.1/88.698.9/95.077.3/47.4
Chewing gum99.396.993.4/90.097.5/92.740.9/59.098.5/96.099.4/98.974.7/81.694.9/96.498.7/99.082.9/79.5
Fryum89.888.0100/86.297.5/83.240.9/26.493.2/91.995.8/90.042.9/30.497.0/94.689.3/92.128.3/34.4
Pipe fryum95.695.494.1/87.181.8/72.323.7/13.196.0/87.597.0/92.544.1/31.891.4/86.199.1/98.269.1/62.6
Mean87.893.884.1/80.588.8/87.025.4/20.893.0/87.196.1/95.241.5/39.994.2/87.896.0/96.644.2/41.7
" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.271, + 0.487, + 0.521 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.204, + 0.521, + 0.385, + 0.532 + ], + "angle": 0, + "content": "(a). 15 categories of MVtecAD dataset" + }, + { + "type": "image", + "bbox": [ + 0.489, + 0.272, + 0.891, + 0.522 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.618, + 0.521, + 0.774, + 0.532 + ], + "angle": 0, + "content": "(b). 12 categories of VisA dataset" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.536, + 0.893, + 0.565 + ], + "angle": 0, + "content": "Figure 5. Qualitative illustration of our anomaly detection results. From top to bottom, the test images overlap with ground truth contours, the reconstructions and the anomaly localization map overlays are shown in rows." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.572, + 0.236, + 0.589 + ], + "angle": 0, + "content": "4.2. Implementation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.596, + 0.471, + 0.762 + ], + "angle": 0, + "content": "To compare with existing methods, we train OmniAL both in unified and separate paradigm with a batch size of 20/12 images having size of \\(256 \\times 256 / 256 \\times 320\\) for MVtec/VisA and pixel value range of [0, 1]. A single model is trained with batches that contain all-class samples for unified paradigm. On the contrary, the separate models are trained with corresponding class samples respectively. The Adam optimizer has an initial learning rate of 1e-4. To alleviate over-fitting, during training, the anomaly texture is alternatively selected either from augmented normal images or from the DTD [4] dataset." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.771, + 0.345, + 0.789 + ], + "angle": 0, + "content": "4.3. Comparison and performance" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Performance on MVTecAD. We compare our proposed method with state-of-the-art methods recently reported on MVTecAD for both unified and separate models in anomaly detection and localization. Table 1 and 2 show our quantitative comparison with the state-of-the-art methods on the task of anomaly detection and localization. UniAD [33] is designed for the unified training paradigm while the others" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.573, + 0.893, + 0.786 + ], + "angle": 0, + "content": "are for the separate training scheme. The results of unified model of Draem [35] and JNLD [38] are got by retraining their models under the paradigm of one model for N classes. The others are reported from UniAD [33]. As shown in Table 1 and 2, for both image-level and pixel-level anomaly detection, most of the existing methods' performance drop drastically when the paradigm switch from training N separate models to a unified model. Comparing with the transformer-based method UniAD [33], our CNN-based method achieves \\(0.7\\%\\) and \\(1.5\\%\\) higher performance on image-AUROC and pixel-AUROC. Moreover, as shown in Table 4, OmniAL(73.4) surpasses SOTA reconstruction-based methods Draem(49.4) [35] and JNLD(46.6) [38] in pixel-AP with more than \\(24\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Performance on VisA. Comparing with MVTecAD, VisA is more difficult since it considers more complex structure and multiple misaligned instances scenes. Table 5 shows the superior performance of OmniAL comparing with the baseline separate method SPD [39] that proposes the VisA dataset and two reconstruction-based methods under the unified setting. The results of both separate and uni-" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3930" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.094, + 0.462, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.26, + 0.462, + 0.437 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.439, + 0.468, + 0.466 + ], + "angle": 0, + "content": "Figure 6. Robustness comparison of our method OmniAL with the state-of-the-art against adversarial attack." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.477, + 0.468, + 0.567 + ], + "angle": 0, + "content": "fied model of Draem [35] and JNLD [38] are got by retraining on VisA. For separate training, OmniAL surpasses the best of them with \\(1.2\\%\\) image-AUROC, \\(2.7\\%\\) pixel-AP and similar pixel-AUROC. For unified training, OmniAL also surpasses the best of them with \\(0.7\\%\\) image-AUROC, \\(1.4\\%\\) pixel-AUROC and \\(1.8\\%\\) pixel-AP." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.568, + 0.468, + 0.749 + ], + "angle": 0, + "content": "As shown in Fig.5, our reconstruction and localization achieve good performance in most categories. Even in the multi-instance scene, OmniAL still successfully recovers the anomaly region back to normal. The reconstructed images have the high-fidelity appearance with the inputs in normal regions and recover the anomalies as close as the expectation. The quantity evaluation of reconstruction is reported in Fig.6a. The bigger circles indicate the reconstruction performance on the clean normal MVTecAD [1]. Our OmniAL surpasses the Draem [35] and JNLD [38] with big gaps in both PSNR and SSIM scores under the unified setting." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Adversarial robustness. Adversarial attacks are easily performed under the disguise of anomaly with additive noises and hardly arouse suspicion. Our anomaly synthesis method can be expeditiously expanded to generate adversarial anomaly samples with PGD [16] perturbations for robustness evaluation. To further evaluate the robustness of separate and unified models, we build up 3 adversarial datasets with PGD [16] perturbations \\(\\epsilon = \\{4 / 255, 8 / 255\\}\\) and \\(\\epsilon = 8 / 255\\) with targeted attack based on the anomaly-free training set from clean MVTecAD [1]. For realistic ad" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.258 + ], + "angle": 0, + "content": "versarial samples, the synthesized anomalies only appear in foreground of the object categories and the anomaly texture is cropped from the normal images. The generated adversarial datasets consist of \\(80\\%\\) synthesized anomaly and \\(20\\%\\) normal samples. Fig.6 illustrates the performance comparison in (a) reconstruction and (b) localization respectively. The names with 'sp' suffix indicate separate models while the others are unified models. OmniAL(green) achieves the best performance against all adversarial levels and has less degradation. All category adversarial samples and performance comparison are shown in supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.267, + 0.649, + 0.282 + ], + "angle": 0, + "content": "4.4. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.289, + 0.892, + 0.653 + ], + "angle": 0, + "content": "Table 3 demonstrates the effectiveness of proposed panel-guided anomaly synthesis, DCSA block and Diff-Neck module for both unified and separate training paradigm. We build a baseline by only using channel attention in DCSA blocks and achieve the best performance in separate setting but the worse performance in unified setting. By using panel-guided anomaly synthesis, we achieve about \\(10\\%\\) image-level and \\(24\\%\\) pixel-level performance improvement comparing with the baseline. The overall performance is further improved by using instance normalization rather than batch normalization in the reconstruction basic blocks. With the help of dilated convolution layer in basic blocks, we can preserve high resolution feature maps and get further improvement. By further using dilated spatial attention, the separate models performance gain \\(2\\%\\) but the unified model decreases \\(3\\%\\) in pixel-AP. Alternatively, the model using DiffNeck module gains \\(1.2\\%\\) for separate setting and \\(0.6\\%\\) for unified setting. Finally, by combining dilated spatial attention and DiffNeck together, we achieve the balanced performance in separate and unified settings and get the OmniAL. As shown in Table 4, proposed panel-guided anomaly synthesis also boosts existing reconstruction-based methods with a large margin. More ablation studies are shown in supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.665, + 0.618, + 0.68 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Considering the practical usages, we propose a unified CNN framework to localize anomalous for multiple classes with a single model. Extensive experiments on MVTecAD and VisA datasets verify the effectiveness of our proposed panel-guided anomaly synthesis, dilated channel and spatial attention (DCSA) block, and DiffNeck module. Especially, OmniAL produces more precisely anomaly reconstruction and localization results. The panel-guided anomaly synthesis can be easily used in existing methods and brings performance improvement under the unified training paradigm. Moreover, we make the first attempt to conduct a comprehensive study on the robustness of existing methods against different levels of adversarial for both separate and unified training paradigm." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "3931" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.184 + ], + "angle": 0, + "content": "[1] Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. Mvtec AD - A comprehensive real-world dataset for unsupervised anomaly detection. In CVPR, pages 9592-9600, 2019. Computer Vision Foundation / IEEE. 1, 2, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.185, + 0.472, + 0.254 + ], + "angle": 0, + "content": "[2] Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. Uninformed students: Student-teacher anomaly detection with discriminative latent embeddings. In CVPR, pages 4182–4191, 2020. Computer Vision Foundation / IEEE. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.256, + 0.471, + 0.324 + ], + "angle": 0, + "content": "[3] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 15750-15758. Computer Vision Foundation / IEEE, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.326, + 0.47, + 0.381 + ], + "angle": 0, + "content": "[4] Mircea Cimpoi, Subhansu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In CVPR, pages 3606-3613, 2014. IEEE Computer Society. 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.382, + 0.469, + 0.422 + ], + "angle": 0, + "content": "[5] Niv Cohen and Yedid Hoshen. Sub-image anomaly detection with deep pyramid correspondences. CoRR, abs/2005.02357. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.424, + 0.47, + 0.535 + ], + "angle": 0, + "content": "[6] Thomas Defard, Aleksandr Setkov, Angelique Loesch, and Romaric Audigier. Padim: A patch distribution modeling framework for anomaly detection and localization. In Alberto Del Bimbo, Rita Cucchiara, Stan Sclaroff, Giovanni Maria Farinella, Tao Mei, Marco Bertini, Hugo Jair Escalante, and Roberto Vezzani, editors, ICPR, volume 12664 of Lecture Notes in Computer Science, pages 475-489, 2020. Springer. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.536, + 0.47, + 0.606 + ], + "angle": 0, + "content": "[7] Chouro Ding, Guansong Pang, and Chunhua Shen. Catching both gray and black swans: Open-set supervised anomaly detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 7378-7388. IEEE, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.607, + 0.469, + 0.675 + ], + "angle": 0, + "content": "[8] Mariana-Iuliana Georgescu, Radu Tudor Ionescu, Fahad Shahbaz Khan, Marius Popescu, and Mubarak Shah. A background-agnostic framework with adversarial training for abnormal event detection in video. IEEE Trans. Pattern Anal. Mach. Intell., 44(9):4505-4523, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.676, + 0.469, + 0.732 + ], + "angle": 0, + "content": "[9] Denis A. Gudovskiy, Shun Ishizaka, and Kazuki Kozuka. CFLOW-AD: real-time unsupervised anomaly detection with localization via conditional normalizing flows. In WACV, pages 1819–1828, 2022. IEEE. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.733, + 0.469, + 0.801 + ], + "angle": 0, + "content": "[10] Dan Hendrycks, Mantas Mazeika, and Thomas G. Dietterich. Deep anomaly detection with outlier exposure. In 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019. OpenReview.net, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.469, + 0.857 + ], + "angle": 0, + "content": "[11] Chaoqin Huang, Haoyan Guan, Aofan Jiang, Ya Zhang, Michael W. Spratling, and Yan-Feng Wang. Registration based few-shot anomaly detection. CoRR, abs/2207.07361, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[12] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In Francis R. Bach and David M. Blei, editors," + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.148 + ], + "angle": 0, + "content": "Proceedings of the 32nd International Conference on Machine Learning, ICML 2015, Lille, France, 6-11 July 2015, volume 37 of JMLR Workshop and Conference Proceedings, pages 448-456. JMLR.org, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.149, + 0.892, + 0.259 + ], + "angle": 0, + "content": "[13] Max Jaderberg, Karen Simonyan, Andrew Zisserman, and Koray Kavukcuoglu. Spatial transformer networks. In Corinna Cortes, Neil D. Lawrence, Daniel D. Lee, Masashi Sugiyama, and Roman Garnett, editors, Advances in Neural Information Processing Systems 28: Annual Conference on Neural Information Processing Systems 2015, December 7-12, 2015, Montreal, Quebec, Canada, pages 2017-2025, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.26, + 0.892, + 0.315 + ], + "angle": 0, + "content": "[14] Chun-Liang Li, Kihyuk Sohn, Jinsung Yoon, and Tomas Pfister. Cutpaste: Self-supervised learning for anomaly detection and localization. In CVPR, pages 9664-9674, 2021. Computer Vision Foundation / IEEE. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.316, + 0.892, + 0.357 + ], + "angle": 0, + "content": "[15] Tsung-Yi Lin, Priya Goyal, Ross B. Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. In CVPR, pages 2999-3007, 2017. IEEE Computer Society. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.358, + 0.892, + 0.412 + ], + "angle": 0, + "content": "[16] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In ICLR, 2018. OpenReview.net. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.413, + 0.892, + 0.44 + ], + "angle": 0, + "content": "[17] Prasanta Chandra Mahalanobis. On the generalized distance in statistics. National Institute of Science of India, 1936. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.442, + 0.892, + 0.496 + ], + "angle": 0, + "content": "[18] Masoud PourReza, Bahram Mohammadi, Mostafa Khaki, Samir Bouindour, Hichem Snoussi, and Mohammad Sabokrou. G2D: generate to detect anomaly. In WACV, pages 2002–2011, 2021. IEEE. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.498, + 0.892, + 0.552 + ], + "angle": 0, + "content": "[19] Karsten Roth, Latha Pemula, Joaquin Zepeda, Bernhard Scholkopf, Thomas Brox, and Peter Gehler. Towards total recall in industrial anomaly detection. In CVPR, pages 14318-14328, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.553, + 0.892, + 0.607 + ], + "angle": 0, + "content": "[20] Marco Rudolph, Bastian Wandt, and Bodo Rosenhahn. Same same but differnet: Semi-supervised defect detection with normalizing flows. In WACV, pages 1906-1915, 2021. IEEE. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.609, + 0.892, + 0.663 + ], + "angle": 0, + "content": "[21] Marco Rudolph, Tom Wehrbein, Bodo Rosenhahn, and Bastian Wandt. Fully convolutional cross-scale-flows for image-based defect detection. In WACV, pages 1829-1838, 2022. IEEE. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.665, + 0.892, + 0.734 + ], + "angle": 0, + "content": "[22] Mohammadreza Salehi, Hossein Mirzaei, Dan Hendrycks, Yixuan Li, Mohammad Hossein Rohban, and Mohammad Sabokrou. A unified survey on anomaly, novelty, open-set, and out-of-distribution detection: Solutions and future challenges. CoRR, abs/2110.14051. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.735, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[23] Mohammadreza Salehi, Niousha Sadjadi, Soroosh Baselizadeh, Mohammad H. Rohban, and Hamid R. Rabiee. Multiresolution knowledge distillation for anomaly detection. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 14902-14912. Computer Vision Foundation / IEEE, 2021. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[24] Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. Int. J. Comput. Vis., 128(2):336-359, 2020. 3" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3932" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[25] Jianbing Shen, Ying Zhao, Shuicheng Yan, and Xuelong Li. Exposure fusion using boosting laplacian pyramid. IEEE Trans. Cybern., 44(9):1579-1590, 2014. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.136, + 0.47, + 0.19 + ], + "angle": 0, + "content": "[26] Xian Tao, Xinyi Gong, Xin Zhang, Shaohua Yan, and Chandranath Adak. Deep learning for unsupervised anomaly localization in industrial images: A survey. CoRR, abs/2207.10298. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.193, + 0.471, + 0.247 + ], + "angle": 0, + "content": "[27] Chin-Chia Tsai, Tsung-Hsuan Wu, and Shang-Hong Lai. Multi-scale patch-based representation learning for image anomaly detection and segmentation. In WACV, pages 3065-3073, 2022. IEEE. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.25, + 0.47, + 0.29 + ], + "angle": 0, + "content": "[28] Dmitry Ulyanov, Andrea Vedaldi, and Victor S. Lempitsky. Instance normalization: The missing ingredient for fast stylization. CoRR, abs/1607.08022, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.293, + 0.47, + 0.348 + ], + "angle": 0, + "content": "[29] Zhou Wang, Alan C. Bovik, Hamid R. Sheikh, and Eero P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Trans. Image Process., 13(4):600-612, 2004. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.35, + 0.47, + 0.447 + ], + "angle": 0, + "content": "[30] Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon. CBAM: convolutional block attention module. In Vittorio Ferrari, Martial Hebert, Cristian Sminchisescu, and Yair Weiss, editors, Computer Vision - ECCV 2018 - 15th European Conference, Munich, Germany, September 8-14, 2018, Proceedings, Part VII, volume 11211 of Lecture Notes in Computer Science, pages 3-19. Springer, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.449, + 0.47, + 0.503 + ], + "angle": 0, + "content": "[31] Jinjian Wu, Guangming Shi, Weisi Lin, Anmin Liu, and Fei Qi. Just noticeable difference estimation for images with free-energy principle. IEEE Trans. Multim., 15(7):1705-1710, 2013. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.506, + 0.47, + 0.616 + ], + "angle": 0, + "content": "[32] Jihun Yi and Sungroh Yoon. Patch SVDD: patch-level SVDD for anomaly detection and segmentation. In Hiroshi Ishikawa, Cheng-Lin Liu, Tomás Pajdla, and Jianbo Shi, editors, Computer Vision - ACCV 2020 - 15th Asian Conference on Computer Vision, Kyoto, Japan, November 30 - December 4, 2020, Revised Selected Papers, Part VI, volume 12627 of Lecture Notes in Computer Science, pages 375-390. Springer, 2020. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.618, + 0.47, + 0.659 + ], + "angle": 0, + "content": "[33] Zhiyuan You, Lei Cui, Yujun Shen, Kai Yang, Xin Lu, Yu Zheng, and Xinyi Le. A unified model for multi-class anomaly detection. CoRR, abs/2206.03687, 2022. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.661, + 0.47, + 0.743 + ], + "angle": 0, + "content": "[34] Sangdoo Yun, Dongyoon Han, Sanghyuk Chun, Seong Joon Oh, Youngjoon Yoo, and Junsuk Choe. Cutmix: Regularization strategy to train strong classifiers with localizable features. In 2019 IEEE/CVF International Conference on Computer Vision, ICCV 2019, Seoul, Korea (South), October 27 - November 2, 2019, pages 6022-6031. IEEE, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.746, + 0.47, + 0.8 + ], + "angle": 0, + "content": "[35] Vitjan Zavrtanik, Matej Kristan, and Danijel Skocaj. Dram - A discriminatively trained reconstruction embedding for surface anomaly detection. In ICCV, pages 8310-8319, 2021. IEEE. 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.47, + 0.844 + ], + "angle": 0, + "content": "[36] Vitjan Zavrtanik, Matej Kristan, and Danijel Skocaj. Reconstruction by inpainting for visual anomaly detection. Pattern Recognit., 112:107706, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[37] Vitjan Zavrtanik, Matej Kristan, and Danijel Skocaj. DSR - A dual subspace re-projection network for surface anomaly detection. In Shai Avidan, Gabriel J. Brostow, Moustapha Cisse, Giovanni Maria Farinella, and Tal Hassner, editors," + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.148 + ], + "angle": 0, + "content": "Computer Vision - ECCV 2022 - 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXI, volume 13691 of Lecture Notes in Computer Science, pages 539-554. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[38] Ying Zhao. Just noticeable learning for unsupervised anomaly localization and detection. In ICME, pages In Press, 2022. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.193, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[39] Yang Zou, Jongheon Jeong, Latha Pemula, Dongqing Zhang, and Onkar Dabeer. Spot-the-difference self-supervised pretraining for anomaly detection and segmentation. CoRR, abs/2207.14315, 2022. 2, 3, 5, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3933" + } + ] +] \ No newline at end of file diff --git a/2023/OmniAL_ A Unified CNN Framework for Unsupervised Anomaly Localization/390b7405-6295-46ac-87cb-a78d4a95dd63_origin.pdf b/2023/OmniAL_ A Unified CNN Framework for Unsupervised Anomaly Localization/390b7405-6295-46ac-87cb-a78d4a95dd63_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..bf954595bbecea0f6f91b63eb5e08b27646d4f57 --- /dev/null +++ b/2023/OmniAL_ A Unified CNN Framework for Unsupervised Anomaly Localization/390b7405-6295-46ac-87cb-a78d4a95dd63_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7e4a2727c8c4420dafc059834f7cf39b35561358835c3fac5e1d3a69ee943a96 +size 7112254 diff --git a/2023/OmniAL_ A Unified CNN Framework for Unsupervised Anomaly Localization/full.md b/2023/OmniAL_ A Unified CNN Framework for Unsupervised Anomaly Localization/full.md new file mode 100644 index 0000000000000000000000000000000000000000..641cb6677390b0f505f7690e173fe54815ef881b --- /dev/null +++ b/2023/OmniAL_ A Unified CNN Framework for Unsupervised Anomaly Localization/full.md @@ -0,0 +1,276 @@ +# OmniAL: A unified CNN framework for unsupervised anomaly localization + +Ying Zhao +Ricoh Software Research Center (Beijing) Co., Ltd. + +zy_deepwhite.zy@hotmail.com + +![](images/c578b6fe0656ce94138d5b7e9fbf860c9e9cee3afb961f16fe8c0c0a857573c7.jpg) +(a). N models for N classes + +![](images/3fb174b3e377479221cb887e62bdc6dd48c4d2cd5beffbe8f5d950db2709bfd3.jpg) +(b). One model for N classes + +![](images/2c4b1b4d50b021c8efd9c7817168b581752878b475b92e121c46fe25429f5258.jpg) +(c). Proposed OmniAL +Figure 1. OmniAL is a unified framework for unsupervised anomaly localization. As shown in (c), OmniAL not only simplifies existing learning paradigm from (a) Separate: N models for N classes to (b) Unified: one model for N classes but also produces high quality anomaly localization results. Results in (a) and (b) are from separate and unified models of JNLD [38] on MVtecAD [1] dataset. + +# Abstract + +Unsupervised anomaly localization and detection is crucial for industrial manufacturing processes due to the lack of anomalous samples. Recent unsupervised advances on industrial anomaly detection achieve high performance by training separate models for many different categories. The model storage and training time cost of this paradigm is high. Moreover, the setting of one-model-N-classes leads to fearful degradation of existing methods. In this paper, we propose a unified CNN framework for unsupervised anomaly localization, named OmniAL. This method conquers aforementioned problems by improving anomaly synthesis, reconstruction and localization. To prevent the model learning identical reconstruction, it trains the model with proposed panel-guided synthetic anomaly data rather than directly using normal data. It increases anomaly reconstruction error for multi-class distribution by using a network that is equipped with proposed Dilated Channel and Spatial Attention (DCSA) blocks. To better localize the anomaly regions, it employs proposed DiffNeck between reconstruction and localization sub-networks to explore multi-level differences. Experiments on 15-class MVTecAD and 12-class VisA datasets verify the advantage of proposed OmniAL that surpasses the state-of-the-art of unified models. On 15-class-MVTecAD/12-class-VisA, its single unified + +model achieves 97.2/87.8 image-AUROC, 98.3/96.6 pixel-AUROC and 73.4/41.7 pixel-AP for anomaly detection and localization respectively. Besides that, we make the first attempt to conduct a comprehensive study on the robustness of unsupervised anomaly localization and detection methods against different level adversarial attacks. Experiential results show OmniAL has good application prospects for its superior performance. + +# 1. Introduction + +In real industrial scenarios, the location of anomaly [22, 26] reveals important information, such as defective types and degrees. It is essential not only to inspect whether a sample is defective but also to know where the specific anomaly regions are. Since anomaly appearance is inexhaustible, it is almost impossible and infeasible to collect and manually annotate all kinds of abnormal data. Thus, only normal samples are available for training a detector that is robust enough to find out unseen anomalies during inference phase. Considering the diversity of classes and various types of one class, the conventional training paradigm of N models for N classes, as shown in Fig.1a, may not be the best solution. The model storage and training time cost increase with the number of classes. As shown in Fig.1b, existing method severely degrades anomaly local- + +![](images/50640b258c33fab036f3f64210f80d9ab11981a8d04d725afa3a951cf2bbd3c1.jpg) +(a). Reconstruction-based anomaly localization +(1). MVtecAD + +![](images/0ec8e8a7481a2e57af9bff2d53049336e23f72501496beb7ce13b8f5bcbf1f5b.jpg) + +![](images/ec5d871442bb5bcab9ab9878e954aa00b3f8a0ce6c7f0583e2ffa6188f909de6.jpg) + +![](images/e0cc6728ca0983fa01631e35b956d01e9db29ec95686c3a3db3274fa91cc64d5.jpg) +(2).JNLD Rec./Loc. +X Rec. X Loc. + +![](images/bf791d08e719ee28e038ef1f8b2f335d531853360f732cfcf22dc5adcaa4fc03.jpg) +sis unified models o +Figure 2. Problem analysis. The final failure may be caused by reconstruction and localization. + +![](images/fbafbccb6a90f99a5d647c4eac91a3e55456db6cdd3c2c2471c88c42f73d10bb.jpg) +(3). OmniRec./Loc. +V Rec. V Loc. + +![](images/08ecfb2baeb1600c30b74357fd1c65ee1e9263b9d8d47c4bd19d038347327100.jpg) +V Rec. V Loc. + +![](images/d45e0b87cbcd0bc1c30358911abc4aa3de3a3fb05baf8831ecc8f820ab106655.jpg) +(4).VisA + +![](images/b1f3fd5eccc8c573d5873458ff579f0d4d93ade00ee6bc2252ba388dbf7d7c23.jpg) +(5). JNLD Rec. / Loc. +X Rec. + +![](images/1190e020d530cf7132c7690d369e4ea5b7dbc9dd3e941d2be6fcaf05854bc4d7.jpg) +V Rec. +Analysis unified models on VisA dataset. + +![](images/d78c4723537b92c5c4b8820ec491c98b090650f6b939f4b944656a7537d5be3f.jpg) +(6). OmniRec./Loc. +Rec. +√Loc. + +![](images/a18cc8e45a9a81061bcfd38dbb66af81ca883edb81cd645af2654895893a6ecd.jpg) +V Rec. +√ Loc. + +ization performance if the training paradigm changes to one model for N classes. Therefore, a robust unified framework for unsupervised anomaly localization is highly demanded for intelligent industrial. + +With the limitation of available training data, many appealing unsupervised approaches [14, 18, 35, 38] using synthesized anomaly data are proposed. These approaches generate anomalous instances to inspire the anomaly detector to learn discriminative features. Their experiments show that the realiziness of generated anomalous instances had a strong impact on the quality of anomaly localization. However, none of these methods consider the training paradigm of one model for N classes. When switching to the unified training paradigm, they are more prone to learn an identical short-cut and fail to discriminate the anomaly. + +With the normal and synthesized anomalous samples, recent unsupervised learning methods train a deep anomaly detector by either a distance-based [6, 14, 19-21, 27] or reconstruction-based [2, 9, 35, 36, 38] way. The reconstruction-based architectures [35, 38] are supposed to reconstruct normal images more accurately than the unseen anomalous. The anomaly localization is then calculated from the reconstruction error between the original and reconstructed versions of the input image, as shown in Fig.2a. The prediction of anomaly location is not only based on the reconstruction quality but also the ability of spotting the reconstruction error. The typical reconstruction-based method JNLD [38] learns a joint representation of an anomalous image and its anomaly-free reconstruction, while simultaneously learning a decision boundary between normal and simulated anomalous examples. As shown in Fig.2b and Fig.2c, under the unified setting, JNLD [38] fails to produce correct results either because of the reconstruction failure or the localization failure. + +To conquer aforementioned problems, we propose a novel unified framework OmniAL for effectively localiz- + +ing anomaly pixels of different classes only by using a single model. OmniAL uses a panel-guided anomaly synthesis method that controls the portion of normal and anomaly regions for each training sample. By doing this, OmniAL blocks the chance of learning identical shortcut from the source. To increase the anomaly reconstruction error for multi-class distribution, OmniAL constructs a reconstruction and a localization sub-networks that are equipped with proposed Dilated Channel and Spatial Attention (DCSA) blocks. To better localize the anomaly regions, OmniAL employs a DiffNeck module between the reconstruction and localization sub-networks to explore multi-level reconstruction errors. As shown in Fig.1 and Fig.2, OmniAL learns a single unified model for multiple classes that produces high quality reconstruction and precise anomaly localization. Furthermore, we conduct an exhaustive evaluation of reconstruction and localization performance against multi-level adversarial attacks. + +In summary, we make following main contributions: + +- We construct a unified CNN framework OmniAL for unsupervised anomaly localization that is equipped with proposed panel-guided anomaly synthesis, DCSA block, and DiffNeck module. OmniAL achieves superior performance for anomaly localization on challenging MVTecAD [1] and VisA [39] datasets compared to the state-of-the-art. +- By preventing model from learning identical reconstruction, our proposed panel-guided anomaly synthesis method also brings substantial improvement for existing methods under the unified setting. It boosts the image-AUROC/pixel-AUROC/pixel-AP from 88.7/87.1/49.4 to 92.5/94.5/57.4 for Draem [35]. +- We make a comprehensive study on the robustness of separate/unified anomaly localization methods against different level adversarial attacks. Our synthesized + +![](images/0609dd1ea28a4b3cfc8efc9cedeb888b7b397eab086b0021158a31d436678d94.jpg) +Figure 3. Framework of OmniAL. It consists of panel-guided anomaly synthesis, reconstruction and localization. Anomaly synthesis is based on anomaly panel and three variants of the Just Noticeable Distortion (JND) map. The synthetic anomaly is reconstructed into normal image and corresponding JND map by the Dilated Channel Spatial Attention (DCSA) modules equipped reconstruction sub-network. The localization sub-network with a DiffNeck module localizes the anomaly regions by exploring the difference between reconstructed and original data. + +adversarial datasets exhibit strong attack capability against anomaly detection, reconstruction and localization, also helping to analyse the risks of existing methods. + +# 2. Related Work + +Anomaly synthesis. Due to the lack of anomaly samples, unsupervised learning methods are commonly used in industrial quality inspection. These methods are typically trained with normal data and synthesized anomalous that highly effect the performance. Therefore, how to synthesize the anomalous [7, 8, 14, 35, 37, 38] also draws extensive attention. CutPaste [14] learns representations by classifying normal data from the Cut-and-Paste augmentation. SPD [39] uses a smoothed version of CutPaste [14] augmentation. Instead of using simple regular shaped anomalous, Draem [35] simulates just-out-of-distribution anomalies having random shape and texture. To simulate photorealistic anomaly samples, JNLD [38] proposes a multiscale noticeable anomalous generation method based on just noticeable distortion [31]. The open-set supervised anomaly detection method DRA [7] adapts the popular CutMix [34] and the outlier exposure method [10] to generate pseudo anomalies from normal images for training. DSR [37] generates the anomalies at the feature level by sampling the learned quantized feature space, which allows a controlled generation of near-in-distribution anomalies. For the unified paradigm, we propose a panel-guided anomaly synthesis method that controls the portion of normal and anomaly regions for each training sample. + +Anomaly localization. Anomaly detection [22,26], also known as outlier detection or one-class classification, refers to the task of distinguishing defective image at the image-level from the majority of anomaly-free images. Anomaly localization (segmentation), on the other hand, aims to seg + +ment out the pixel-level anomaly regions. For industrial visual inspection, many recent methods [5, 6, 19] achieve high performance in anomaly detection but anomaly localization. SPADE [5] detects anomaly based on alignment between an anomalous image and a constant number of the similar normal images. It relies on K nearest neighbors of pixel-level feature pyramids extracted by pre-trained deep features. PSVDD [32] extends deep Support Vector Data Description (SVDD) to a patch-wise detection method. PaDim [6] also relies on ImageNet pretrained feature extractor with multiscale pyramid pooling. Instead of using time-costly clustering, it uses a well-known Mahalanobis distance metric [17] as an anomaly score. CutPaste [14] uses GradCAM [24] to get the defect localization. MKD [23] proposes to use multi-level features alignment to increase the discriminating capability of the Teacher-Student model on various types of abnormalities. PatchCore [19] combines patch-level embeddings from ImageNet models with an outlier detection model. However, PatchCore is inherently not suitable for unified setting (larger normal dataset) since it needs to build the coreset with all normal data. Draem [35] and JNLD [38] use segmentation sub-network to predict the defective regions. This paper differs from these previous works by focusing on the paradigm of only using a single unified model for N classes. + +Unified anomaly localization. Recently, how to use a unified model to localize anomaly for different objects has already become researchers' concern. RegAD [11] trains a single generalizable model for few-shot anomaly detection, where a limited number of normal images are provided for each category at training. It employs SimSiam [3] with three spatial transformer network [13] blocks to solve the category-agnostic feature registration proxy task. It identifies anomalies by comparing the registered features of the test image and the corresponding normal images. UniAD [33] constructs a transformer with a layer-wise query de + +coder and a neighbor masked attention module to model the multi-class distribution. It uses the layer-wise query decoder to intensify the use of query embedding. To avoid the information leak, it employs a neighbor masked attention module, where a feature point relates to neither itself nor its neighbors. In this paper, we propose a unified CNN framework for unsupervised anomaly localization. It manifests the problem-solving ability of fully convolutional networks. + +# 3. Method + +In this work, we construct a unified CNN framework OmniAL for unsupervised anomaly localization under the practical setting of only using one model for N classes. Fig.3 shows the overview of proposed OmniAL. To train only with normal data, we firstly synthesize photo-realistic anomalous in the regions that are selected based 5 predefined panels. The synthesized anomaly samples are then used to train the OmniAL network that consists of an anomaly-free reconstruction and an anomaly localization sub-networks. The reconstruction and localization sub-networks are composed of alternative basic blocks and DCSA blocks. The reconstruction sub-network learns to recover the synthesized anomalous back to normal. The localization sub-network with a DiffNeck module localizes the anomaly regions by exploring the difference between reconstructed and original data. + +# 3.1. Panel-guided anomaly synthesis + +Reconstruction-based anomaly localization methods rely on the hypothesis that the reconstruction errors in unseen anomalous regions are larger than the normal regions. The reconstruction models are trained with alternative normal and synthetic anomaly samples. For normal sample, the supervision ground truth is the direct copy of the input. Thus, directly using normal samples for training increases the chance of information leak and leads to identical reconstruction for any input. To prevent the learning of identical short-cut, the intuitive idea is that only using synthetic anomaly rather than normal data. However, without using normal samples for training leads to more false alarm. Then, the problem is transformed into how to better synthesis anomaly for unified model training. + +To solve the problem, we propose a panel-guided anomaly synthesis method that takes into account both normal and anomaly for training. As shown in Fig.4, we build 5 types of panels ('left', 'right', 'top', 'bottom', 'all') to control the portion of normal and anomaly regions. That is, we only generate the anomaly in one of the panel region for each sample in each iteration. Except the type of 'all', the area of panel region is also randomly adjusted in range of [0.5, 0.8]*ImageWidth and [0.5, 0.8]*ImageHeight. Given a panel, we synthesize anomaly in the selected region ac + +![](images/8a2c312adc127bd13d9679b228f5f314f456f94944c8e7c35fc5f2ec3338b76d.jpg) +Figure 4. Visualization of our panel-guided anomaly synthetic data. The anomaly source is selected either from augmented normal images (row1,2) or DTD [4] dataset (row3). + +cording to following equation. + +$$ +A = \sum_ {i = 1} ^ {3} \left(W _ {i} T + (1 - W _ {i}) I\right) \tag {1} +$$ + +Where $A$ is the synthesized anomaly image, $W_{i}$ is the pixelwise fusion weight map, $i$ indicates one of the 3 defect levels (easy, medium and hard), $I$ is the anomaly-free image, $T$ is the anomaly source image. To synthesize natural and diversity anomalies, $T$ is generated by applying a set of random augmentation (resize, crop, flip, color-jitter) to $I$ or image from the Describable Textures Dataset(DTD) [4]. + +To get the 3-level defects, we first need to find out the baseline anomaly fusion weight map. Following JNLD [38], we use the JND [31] to guide the anomaly synthesis. JND [31] refers to the minimum visibility threshold of the Human Visual System, is useful in perceptual image/video processing systems. It reveals a perceptual threshold of intensity change in an image that can be noticed by the human vision system. We use a nonlinear additivity JND model to define the anomaly weight map $J$ as [25] that proposes a novel hybrid exposure weight measurement using the JND. Each level anomaly fusion weight map $W_{i}$ is defined as: + +$$ +W _ {i} = \alpha_ {i} * (J + \beta_ {i}) \tag {2} +$$ + +Where $J$ indicates the JND map, $\beta_{i}$ is a bias term for different level of noticeable, $\alpha_{i}$ is a smooth kernel applied to the mask and produces soft boundary between anomaly and normal regions. For the easier level anomaly, we use a smaller smooth kernel. Due to using random seeds, different level anomaly regions may overlap each other. Thus, directly using the masks as the ground truth for anomaly segmentation task may bring label inconsistency. Therefore, we refine and quantify the segmentation mask into 3 defect levels based on the structure similarity index measure (SSIM) [29] between the anomaly-free image $I$ and synthesized anomaly image $A$ . + +# 3.2. Anomaly reconstruction + +As shown in Fig.3, our anomaly reconstruction subnetwork receives an input image and outputs reconstructed + +normal image and corresponding JND map. The reconstruction sub-network consists of an encoder and a decoder. Both of them are composed of alternating basic and DCSA blocks. The basic block extracts different levels of features with two consecutive $3 \times 3$ convolution layer followed by normalization and ReLU activation layers. Considering the unified training paradigm, instance normalization [28] is more suitable than the broadly used batch normalization [12] in anomaly reconstruction. The reason is that batch normalization [12] normalizes using the information from the whole batch, while instance normalization [28] normalizes each feature map on its own. Under unified setting, each batch of training data consists of different classes that complicates the reconstruction problem. Instance normalization discards instance-specific contrast information and reduces undesirable instance-level variation. Batch normalization, meanwhile, lacks the ability to address the class inconsistency. Therefore, we use instance normalization in each basic block rather than batch normalization. Experimental results in Table 3 show that the performance is improved by swapping batch normalization with instance normalization. + +The feature maps extracted by the basic block are further integrated by the DCSA block that is inspired by the popular CBAM [30]. As shown in orange dash block of Fig.3, the DCSA block consists of channel and spatial attention sequential sub-modules. With average pooling and two consecutive 1x1 convolutions, the channel attention explores the correlations among feature channels. The correlations are add back to original feature maps to emphasize the important feature channels. The spatial attention further investigates the spatial correlation. It first aggregates spatial information by concatenating both average and maximum pooling of afore refined feature maps. Then, it spots the spatial importance of the aggregated features with a standard and a dilated $3 \times 3$ convolution-norm-relu paths. It balances the portion of dilated path by a factor that is set as 0.1. The spatial attention map is obtained by summing up two paths of features. By adding the spatial attention map, the important regions in feature maps are highlighted. As shown in Table 3, the performance is further improved by using DCSA blocks. + +# 3.3. Anomaly localization + +After reconstructing the normal and corresponding JND maps, we use a localization sub-network to highlight the anomaly regions. Instead of directly concatenating the reconstructed maps with input image, we firstly use a DiffNeck module to fully explore the difference between them. The pink dash block of Fig.3 illustrates the detailed structure of DiffNeck. Overall, DiffNeck extracts reconstruction difference among reconstructed normal map (Rec), reconstructed JND map (JND) and input image (Inp) in both sep + +arate and unified ways. That is, DiffNeck uses basic blocks to extract the concatenated features of (Rec, Inp), (JND, Inp) and (Rec, JND, Inp) respectively. Then, all levels of differences are summed up and form a final descriptor that is fed to the localization sub-network. DiffNeck also balances the portion of the separate path by a factor that is set as 0.1. + +Similar with reconstruction sub-network, the localization sub-network also follows the encoder-decoder structure but with 6 scales (e1-e5 in Fig.3) and skip connections for corresponding scales. Scale-1 and scale-2 use standard convolution in all blocks. To use larger resolution feature maps and receptive field, the basic and DCSA blocks of the encoder contain dilated convolution layer from scale-3 to scale-6 with 2, 4, 8, 16 dilation rates respectively. Different with reconstruction sub-network, we use batch normalization in all blocks. Since we synthesize more realistic anomalies by considering easy, medium and hard level defects, localization sub-network predicts 3 levels of anomalous locations. To get the final pixel-level anomaly localization map, we combine them together by a sequence processing of SoftMax, smoothness and summing up. By doing this, the false alarm noise can be suppressed, and the weak true decision can be strengthened. + +The total target for training OmniAL is defined as + +$$ +\begin{array}{l} L = L _ {2} \left(J, J _ {r}\right) + L _ {s s i m} \left(J, J _ {r}\right) \tag {3} \\ + L _ {2} (I, I _ {r}) + L _ {s s i m} (I, I _ {r}) + L _ {f l} (S, S _ {s}) \\ \end{array} +$$ + +Where $J$ is the ground truth JND map calculated from the anomaly-free image $I$ , $J_{r}$ and $I_{r}$ are reconstructed JND map and normal image, $S$ is the anomaly localization ground truth. Unlike JNLD [38], for both JND and normal reconstruction, we not only use MSE loss to supervise the pixel-to-pixel recovering but also the SSIM loss to yield plausible local consistency. To handle the unbalance of different types, we use focal loss [15] to supervise the predicted anomaly localization $S_{s}$ . + +# 4. Experimental results + +# 4.1. Datasets and metrics + +To demonstrate the effectiveness of proposed OmniAL, we conduct extensive experiments on the challenging MVTecAD [1] and VisA [39] datasets. + +MVTecAD [1] dataset contains 10 object and 5 texture industrial products, such as bottle and leather. It consists of 3,629 normal images for training and 1,725 images for testing. There are 1,258 anomaly images of the testing set with pixel-level labelled various types of defects and the rest are normal images. Each class contains 60 to 320 color images with the resolution ranges from $700 \times 700$ to $1024 \times 1024$ pixels. In the testing set, defective appearance varies in different sizes, shapes and types, and most cases only contain a small fraction of anomalous pixels. + +Table 1. Image AUROC comparison of our method OmniAL with the state-of-the-art on MVTecAD [1]. + +
CategoryPadim [6]CutPaste [14]MKD [23]Draem [35]JNLD [38]UniAD [33]OmniAL
Unified / Separate
bottle97.9 / 99.967.9 / 98.298.7 / 99.494.6 / 99.199.1 / 96.399.7 / 100100 / 99.4
cable70.9 / 92.769.2 / 81.278.2 / 89.261.8 / 94.790.6 / 98.895.2 / 97.698.2 / 97.6
capsule73.4 / 91.363.0 / 98.268.3 / 80.570.2 / 98.574.4 / 85.086.9 / 85.395.2 / 92.4
carpet93.8 / 99.893.6 / 93.969.8 / 79.395.9 / 95.577.1 / 97.899.8 / 99.998.7 / 99.6
grid73.9 / 96.793.2 / 10083.8 / 98.098.1 / 99.998.6 / 10098.2 / 98.599.9 / 100
hazelnut85.5 / 92.080.9 / 98.397.1 / 98.495.1 / 10090.8 / 10099.8 / 99.995.6 / 98.0
leather99.9 / 10093.4 / 10093.6 / 95.199.9 / 10097.0 / 100100 / 10099.0 / 97.6
metal nut88.0 / 98.760.0 / 99.964.9 / 73.688.9 / 98.793.3 / 99.699.2 / 99.099.2 / 99.9
pill68.8 / 93.371.4 / 94.979.7 / 82.769.0 / 98.982.7 / 94.693.7 / 88.397.2 / 97.7
screw56.9 / 85.885.2 / 88.775.6 / 83.393.3 / 93.981.8 / 95.887.5 / 91.988.0 / 81.0
tile93.3 / 98.188.6 / 94.689.5 / 91.698.3 / 99.699.2 / 10099.3 / 99.099.6 / 100
toothbrush95.3 / 96.163.9 / 99.475.3 / 92.282.8 / 100100 / 10094.2 / 95.0100 / 100
transistor86.6 / 97.457.9 / 96.173.4 / 85.683.9 / 93.190.3 / 93.099.8 / 10093.8 / 93.8
wood72.1 / 96.580.4 / 99.193.4 / 94.399.8 / 99.191.9 / 99.698.6 / 97.993.2 / 98.7
zipper79.7 / 90.393.5 / 99.987.4 / 93.299.1 / 10099.8 / 99.895.8 / 96.7100 / 100
average84.2 / 95.577.5 / 96.181.9 / 87.888.7 / 98.091.3 / 97.496.5 / 96.697.2 / 97.0
+ +Table 2. Pixel AUROC comparison of our method OmniAL with the state-of-the-art on MVTecAD [1]. + +
CategoryPadim [6]PSVDD [32]MKD [23]Draem [35]JNLD [38]UniAD [33]OmniAL
Unified / Separate
bottle96.1 / 98.286.7 / 98.191.8 / 96.387.4 / 99.194.8 / 99.098.1 / 98.199.2 / 99.0
cable81.0 / 96.762.2 / 96.889.3 / 82.470.4 / 94.776.4 / 97.797.3 / 96.897.3 / 97.1
capsule96.9 / 98.683.1 / 95.888.3 / 95.949.2 / 94.357.0 / 92.798.5 / 97.996.9 / 92.2
carpet97.6 / 99.078.6 / 92.695.5 / 95.695.2 / 95.593.7 / 99.098.5 / 98.099.4 / 99.6
grid71.0 / 97.170.8 / 96.282.3 / 91.899.0 / 99.796.9 / 99.798.2 / 98.599.4 / 99.6
hazelnut96.3 / 98.197.4 / 97.591.2 / 94.696.0 / 99.785.9 / 99.496.5 / 94.698.4 / 98.6
leather84.8 / 99.093.5 / 97.496.7 / 98.198.6 / 98.687.0 / 99.598.8 / 98.399.3 / 99.7
metal nut84.8 / 97.396.0 / 98.064.2 / 86.472.6 / 99.597.4 / 99.594.8 / 95.799.1 / 99.1
pill87.7 / 95.796.5 / 95.169.7 / 89.690.0 / 97.691.2 / 96.695.0 / 95.198.9 / 98.6
screw94.1 / 98.474.3 / 95.792.1 / 96.089.3 / 97.687.0 / 99.798.3 / 97.498.0 / 97.2
tile80.5 / 94.192.1 / 91.485.3 / 82.898.1 / 99.294.7 / 99.691.8 / 91.899.0 / 99.4
toothbrush95.6 / 98.898.0 / 98.188.9 / 96.194.4 / 98.198.6 / 98.898.4 / 97.899.4 / 99.2
transistor92.3 / 97.678.5 / 97.071.7 / 76.573.1 / 90.983.6 / 92.197.9 / 98.793.3 / 91.7
wood89.1 / 94.180.7 / 90.880.5 / 84.896.2 / 96.488.7 / 96.393.2 / 93.497.4 / 96.9
zipper94.8 / 98.495.1 / 95.186.1 / 93.996.9 / 98.895.3 / 99.496.8 / 96.099.5 / 99.7
average89.5 / 97.485.6 / 95.784.9 / 97.087.1 / 97.388.6 / 97.996.8 / 96.698.3 / 97.8
+ +Table 3. Ablation study on MVTecAD [1]. PA: Panel-guided anomaly synthesis, Rec: Reconstruction sub-network, Seg: Segmentation sub-network, BB: Basic Block, BN: Batch Norm, IN: Instance Norm, DC: Dilated Convolution, CA: Channel Attention, DSA: Dilated Spatial Attention, I: Image-level classification, P: Pixel-level localization. + +
PARec-BBDCSASeg-BBDiffNeckI-AUROCP-AUROCP-AP
-BNINDCCADSABNDC-Unified/Separate
-√--√-√--86.7/98.886.4/98.444.2/75.0
√√--√-√--94.1/—95.9/—68.3/—
√-√-√-√--95.8/—97.0/—69.1/—
√-√√√-√√-96.9/97.397.6/97.772.8/72.9
√-√√√√√√-94.7/96.596.9/97.869.8/74.9
√-√√√-√√√96.8/96.398.2/98.072.7/74.1
√-√√√√√√√97.2/97.098.3/97.873.4/73.5
+ +Table 4. Ablation study on With/Without panel-guided anomaly synthesis for unified training on MVTecAD [1]. + +
Draem [35]JNLD [38]OmniAL
I-AUROC92.5/88.792.9/91.397.2/86.6
P-AUROC94.5/87.195.6/88.698.3/93.7
Pixel-AP57.4/49.463.1/46.673.4/54.4
+ +VisA [39] dataset consists of 10,821 high-resolution color images (9,621 normal and 1,200 anomalous samples) covering 12 objects in 3 domains, including complex struc + +ture, multiple instances and single instances. The anomalous images contain various flaws, including surface defects such as scratches, dents, color spots or crack, and structural defects like misplacement or missing parts. There are 5-20 images per defect type and an image may contain multiple defects. All images were acquired using a $4,000 \times 6,000$ high-resolution RGB sensor. Example of each category and our corresponding anomaly reconstruction and localization results are shown in Fig.5a(MVTecAD) and Fig.5b(VisA). + +Metrics For anomaly detection evaluation, the most common used metrics are Area Under the Receiver Operating Characteristic curve (AUROC) in both image-level and pixel-level. However, the pixel-AUROC is biased in favour of large anomalous regions and does not well reflect the pixel-level anomalous localization performance. Therefore, we additionally introduce the average precision (AP) to evaluate pixel-level anomaly localization performance. For reconstruction quality evaluation, we use peak signal to noise ratio (PSNR) and SSIM [29] that are commonly used for evaluating signal fidelity. + +Table 5. Performance comparison of our method OmniAL with the state-of-the-art on VisA [39]. (Separate/Unified) + +
CategoryPadim+SPD [39]Draem [35]JNLD [38]OmniAL
I-AUCP-AUCI-AUCP-AUCP-API-AUCP-AUCP-API-AUCP-AUCP-AP
Complex structurePCB192.797.771.3/83.998.6/94.060.4/38.582.0/82.996.4/98.072.8/77.896.6/77.798.7/97.663.5/76.9
PCB287.997.289.7/81.792.5/94.13.5/13.396.3/79.191.9/95.031.2/37.799.4/81.083.2/93.92.8/31.6
PCB385.496.773.1/87.793.8/94.118.7/17.996.9/90.195.3/98.543.4/46.896.9/88.198.4/94.756.9/41.4
PCB499.189.291.3/87.195.8/72.332.7/13.194.8/96.296.1/97.537.4/29.797.4/95.398.5/97.138.4/33.2
Multiple instancesMacaroni185.798.870.3/68.695.8/89.88.2/8.094.3/90.598.8/93.325.9/14.296.9/92.698.9/98.67.6/7.1
Macaroni270.896.071.3/60.394.1/83.225.4/19.786.5/71.392.9/92.117.2/6.189.9/75.299.1/97.911.4/9.2
Capsules68.186.377.3/89.693.7/96.620.2/30.189.1/91.498.9/99.638.6/55.687.9/90.698.6/99.462.9/52.4
Candles89.197.382.3/70.287.0/82.627.9/12.689.1/85.494.8/94.525.9/27.485.1/86.890.5/95.829.2/24.6
Single instanceCashew90.586.194.2/67.394.7/68.541.2/7.096.0/82.596.3/94.143.7/39.497.1/88.698.9/95.077.3/47.4
Chewing gum99.396.993.4/90.097.5/92.740.9/59.098.5/96.099.4/98.974.7/81.694.9/96.498.7/99.082.9/79.5
Fryum89.888.0100/86.297.5/83.240.9/26.493.2/91.995.8/90.042.9/30.497.0/94.689.3/92.128.3/34.4
Pipe fryum95.695.494.1/87.181.8/72.323.7/13.196.0/87.597.0/92.544.1/31.891.4/86.199.1/98.269.1/62.6
Mean87.893.884.1/80.588.8/87.025.4/20.893.0/87.196.1/95.241.5/39.994.2/87.896.0/96.644.2/41.7
+ +![](images/61f42219a66eedb889bdab0543e37967c016d4231d383af80fb771d01a2b9b71.jpg) +(a). 15 categories of MVtecAD dataset + +![](images/5d36cee805ebdc32748f1134b26f9deeafa13a2cb382e93460eebc9a3eb77251.jpg) +(b). 12 categories of VisA dataset +Figure 5. Qualitative illustration of our anomaly detection results. From top to bottom, the test images overlap with ground truth contours, the reconstructions and the anomaly localization map overlays are shown in rows. + +# 4.2. Implementation + +To compare with existing methods, we train OmniAL both in unified and separate paradigm with a batch size of 20/12 images having size of $256 \times 256 / 256 \times 320$ for MVtec/VisA and pixel value range of [0, 1]. A single model is trained with batches that contain all-class samples for unified paradigm. On the contrary, the separate models are trained with corresponding class samples respectively. The Adam optimizer has an initial learning rate of 1e-4. To alleviate over-fitting, during training, the anomaly texture is alternatively selected either from augmented normal images or from the DTD [4] dataset. + +# 4.3. Comparison and performance + +Performance on MVTecAD. We compare our proposed method with state-of-the-art methods recently reported on MVTecAD for both unified and separate models in anomaly detection and localization. Table 1 and 2 show our quantitative comparison with the state-of-the-art methods on the task of anomaly detection and localization. UniAD [33] is designed for the unified training paradigm while the others + +are for the separate training scheme. The results of unified model of Draem [35] and JNLD [38] are got by retraining their models under the paradigm of one model for N classes. The others are reported from UniAD [33]. As shown in Table 1 and 2, for both image-level and pixel-level anomaly detection, most of the existing methods' performance drop drastically when the paradigm switch from training N separate models to a unified model. Comparing with the transformer-based method UniAD [33], our CNN-based method achieves $0.7\%$ and $1.5\%$ higher performance on image-AUROC and pixel-AUROC. Moreover, as shown in Table 4, OmniAL(73.4) surpasses SOTA reconstruction-based methods Draem(49.4) [35] and JNLD(46.6) [38] in pixel-AP with more than $24\%$ . + +Performance on VisA. Comparing with MVTecAD, VisA is more difficult since it considers more complex structure and multiple misaligned instances scenes. Table 5 shows the superior performance of OmniAL comparing with the baseline separate method SPD [39] that proposes the VisA dataset and two reconstruction-based methods under the unified setting. The results of both separate and uni- + +![](images/ca1fde8e5fbed17071294e052aca400e35e6324c9d8ae52145257d03dcfaeba0.jpg) + +![](images/18faa221c1f5ecc57918c08b31601c43294e181a70f0c88f66df0eb8507778e9.jpg) +Figure 6. Robustness comparison of our method OmniAL with the state-of-the-art against adversarial attack. + +fied model of Draem [35] and JNLD [38] are got by retraining on VisA. For separate training, OmniAL surpasses the best of them with $1.2\%$ image-AUROC, $2.7\%$ pixel-AP and similar pixel-AUROC. For unified training, OmniAL also surpasses the best of them with $0.7\%$ image-AUROC, $1.4\%$ pixel-AUROC and $1.8\%$ pixel-AP. + +As shown in Fig.5, our reconstruction and localization achieve good performance in most categories. Even in the multi-instance scene, OmniAL still successfully recovers the anomaly region back to normal. The reconstructed images have the high-fidelity appearance with the inputs in normal regions and recover the anomalies as close as the expectation. The quantity evaluation of reconstruction is reported in Fig.6a. The bigger circles indicate the reconstruction performance on the clean normal MVTecAD [1]. Our OmniAL surpasses the Draem [35] and JNLD [38] with big gaps in both PSNR and SSIM scores under the unified setting. + +Adversarial robustness. Adversarial attacks are easily performed under the disguise of anomaly with additive noises and hardly arouse suspicion. Our anomaly synthesis method can be expeditiously expanded to generate adversarial anomaly samples with PGD [16] perturbations for robustness evaluation. To further evaluate the robustness of separate and unified models, we build up 3 adversarial datasets with PGD [16] perturbations $\epsilon = \{4 / 255, 8 / 255\}$ and $\epsilon = 8 / 255$ with targeted attack based on the anomaly-free training set from clean MVTecAD [1]. For realistic ad + +versarial samples, the synthesized anomalies only appear in foreground of the object categories and the anomaly texture is cropped from the normal images. The generated adversarial datasets consist of $80\%$ synthesized anomaly and $20\%$ normal samples. Fig.6 illustrates the performance comparison in (a) reconstruction and (b) localization respectively. The names with 'sp' suffix indicate separate models while the others are unified models. OmniAL(green) achieves the best performance against all adversarial levels and has less degradation. All category adversarial samples and performance comparison are shown in supplementary material. + +# 4.4. Ablation study + +Table 3 demonstrates the effectiveness of proposed panel-guided anomaly synthesis, DCSA block and Diff-Neck module for both unified and separate training paradigm. We build a baseline by only using channel attention in DCSA blocks and achieve the best performance in separate setting but the worse performance in unified setting. By using panel-guided anomaly synthesis, we achieve about $10\%$ image-level and $24\%$ pixel-level performance improvement comparing with the baseline. The overall performance is further improved by using instance normalization rather than batch normalization in the reconstruction basic blocks. With the help of dilated convolution layer in basic blocks, we can preserve high resolution feature maps and get further improvement. By further using dilated spatial attention, the separate models performance gain $2\%$ but the unified model decreases $3\%$ in pixel-AP. Alternatively, the model using DiffNeck module gains $1.2\%$ for separate setting and $0.6\%$ for unified setting. Finally, by combining dilated spatial attention and DiffNeck together, we achieve the balanced performance in separate and unified settings and get the OmniAL. As shown in Table 4, proposed panel-guided anomaly synthesis also boosts existing reconstruction-based methods with a large margin. More ablation studies are shown in supplementary material. + +# 5. Conclusion + +Considering the practical usages, we propose a unified CNN framework to localize anomalous for multiple classes with a single model. Extensive experiments on MVTecAD and VisA datasets verify the effectiveness of our proposed panel-guided anomaly synthesis, dilated channel and spatial attention (DCSA) block, and DiffNeck module. Especially, OmniAL produces more precisely anomaly reconstruction and localization results. The panel-guided anomaly synthesis can be easily used in existing methods and brings performance improvement under the unified training paradigm. Moreover, we make the first attempt to conduct a comprehensive study on the robustness of existing methods against different levels of adversarial for both separate and unified training paradigm. + +# References + +[1] Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. Mvtec AD - A comprehensive real-world dataset for unsupervised anomaly detection. In CVPR, pages 9592-9600, 2019. Computer Vision Foundation / IEEE. 1, 2, 5, 6, 8 +[2] Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. Uninformed students: Student-teacher anomaly detection with discriminative latent embeddings. In CVPR, pages 4182–4191, 2020. Computer Vision Foundation / IEEE. 2 +[3] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 15750-15758. Computer Vision Foundation / IEEE, 2021. 3 +[4] Mircea Cimpoi, Subhansu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In CVPR, pages 3606-3613, 2014. IEEE Computer Society. 4, 7 +[5] Niv Cohen and Yedid Hoshen. Sub-image anomaly detection with deep pyramid correspondences. CoRR, abs/2005.02357. 3 +[6] Thomas Defard, Aleksandr Setkov, Angelique Loesch, and Romaric Audigier. Padim: A patch distribution modeling framework for anomaly detection and localization. In Alberto Del Bimbo, Rita Cucchiara, Stan Sclaroff, Giovanni Maria Farinella, Tao Mei, Marco Bertini, Hugo Jair Escalante, and Roberto Vezzani, editors, ICPR, volume 12664 of Lecture Notes in Computer Science, pages 475-489, 2020. Springer. 2, 3, 6 +[7] Chouro Ding, Guansong Pang, and Chunhua Shen. Catching both gray and black swans: Open-set supervised anomaly detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 7378-7388. IEEE, 2022. 3 +[8] Mariana-Iuliana Georgescu, Radu Tudor Ionescu, Fahad Shahbaz Khan, Marius Popescu, and Mubarak Shah. A background-agnostic framework with adversarial training for abnormal event detection in video. IEEE Trans. Pattern Anal. Mach. Intell., 44(9):4505-4523, 2022. 3 +[9] Denis A. Gudovskiy, Shun Ishizaka, and Kazuki Kozuka. CFLOW-AD: real-time unsupervised anomaly detection with localization via conditional normalizing flows. In WACV, pages 1819–1828, 2022. IEEE. 2 +[10] Dan Hendrycks, Mantas Mazeika, and Thomas G. Dietterich. Deep anomaly detection with outlier exposure. In 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019. OpenReview.net, 2019. 3 +[11] Chaoqin Huang, Haoyan Guan, Aofan Jiang, Ya Zhang, Michael W. Spratling, and Yan-Feng Wang. Registration based few-shot anomaly detection. CoRR, abs/2207.07361, 2022. 3 +[12] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In Francis R. Bach and David M. Blei, editors, + +Proceedings of the 32nd International Conference on Machine Learning, ICML 2015, Lille, France, 6-11 July 2015, volume 37 of JMLR Workshop and Conference Proceedings, pages 448-456. JMLR.org, 2015. 5 +[13] Max Jaderberg, Karen Simonyan, Andrew Zisserman, and Koray Kavukcuoglu. Spatial transformer networks. In Corinna Cortes, Neil D. Lawrence, Daniel D. Lee, Masashi Sugiyama, and Roman Garnett, editors, Advances in Neural Information Processing Systems 28: Annual Conference on Neural Information Processing Systems 2015, December 7-12, 2015, Montreal, Quebec, Canada, pages 2017-2025, 2015. 3 +[14] Chun-Liang Li, Kihyuk Sohn, Jinsung Yoon, and Tomas Pfister. Cutpaste: Self-supervised learning for anomaly detection and localization. In CVPR, pages 9664-9674, 2021. Computer Vision Foundation / IEEE. 2, 3, 6 +[15] Tsung-Yi Lin, Priya Goyal, Ross B. Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. In CVPR, pages 2999-3007, 2017. IEEE Computer Society. 5 +[16] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In ICLR, 2018. OpenReview.net. 8 +[17] Prasanta Chandra Mahalanobis. On the generalized distance in statistics. National Institute of Science of India, 1936. 3 +[18] Masoud PourReza, Bahram Mohammadi, Mostafa Khaki, Samir Bouindour, Hichem Snoussi, and Mohammad Sabokrou. G2D: generate to detect anomaly. In WACV, pages 2002–2011, 2021. IEEE. 2 +[19] Karsten Roth, Latha Pemula, Joaquin Zepeda, Bernhard Scholkopf, Thomas Brox, and Peter Gehler. Towards total recall in industrial anomaly detection. In CVPR, pages 14318-14328, 2022. 2, 3 +[20] Marco Rudolph, Bastian Wandt, and Bodo Rosenhahn. Same same but differnet: Semi-supervised defect detection with normalizing flows. In WACV, pages 1906-1915, 2021. IEEE. 2 +[21] Marco Rudolph, Tom Wehrbein, Bodo Rosenhahn, and Bastian Wandt. Fully convolutional cross-scale-flows for image-based defect detection. In WACV, pages 1829-1838, 2022. IEEE. 2 +[22] Mohammadreza Salehi, Hossein Mirzaei, Dan Hendrycks, Yixuan Li, Mohammad Hossein Rohban, and Mohammad Sabokrou. A unified survey on anomaly, novelty, open-set, and out-of-distribution detection: Solutions and future challenges. CoRR, abs/2110.14051. 1, 3 +[23] Mohammadreza Salehi, Niousha Sadjadi, Soroosh Baselizadeh, Mohammad H. Rohban, and Hamid R. Rabiee. Multiresolution knowledge distillation for anomaly detection. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 14902-14912. Computer Vision Foundation / IEEE, 2021. 3, 6 +[24] Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. Int. J. Comput. Vis., 128(2):336-359, 2020. 3 + +[25] Jianbing Shen, Ying Zhao, Shuicheng Yan, and Xuelong Li. Exposure fusion using boosting laplacian pyramid. IEEE Trans. Cybern., 44(9):1579-1590, 2014. 4 +[26] Xian Tao, Xinyi Gong, Xin Zhang, Shaohua Yan, and Chandranath Adak. Deep learning for unsupervised anomaly localization in industrial images: A survey. CoRR, abs/2207.10298. 1, 3 +[27] Chin-Chia Tsai, Tsung-Hsuan Wu, and Shang-Hong Lai. Multi-scale patch-based representation learning for image anomaly detection and segmentation. In WACV, pages 3065-3073, 2022. IEEE. 2 +[28] Dmitry Ulyanov, Andrea Vedaldi, and Victor S. Lempitsky. Instance normalization: The missing ingredient for fast stylization. CoRR, abs/1607.08022, 2016. 5 +[29] Zhou Wang, Alan C. Bovik, Hamid R. Sheikh, and Eero P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Trans. Image Process., 13(4):600-612, 2004. 4, 6 +[30] Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon. CBAM: convolutional block attention module. In Vittorio Ferrari, Martial Hebert, Cristian Sminchisescu, and Yair Weiss, editors, Computer Vision - ECCV 2018 - 15th European Conference, Munich, Germany, September 8-14, 2018, Proceedings, Part VII, volume 11211 of Lecture Notes in Computer Science, pages 3-19. Springer, 2018. 5 +[31] Jinjian Wu, Guangming Shi, Weisi Lin, Anmin Liu, and Fei Qi. Just noticeable difference estimation for images with free-energy principle. IEEE Trans. Multim., 15(7):1705-1710, 2013. 3, 4 +[32] Jihun Yi and Sungroh Yoon. Patch SVDD: patch-level SVDD for anomaly detection and segmentation. In Hiroshi Ishikawa, Cheng-Lin Liu, Tomás Pajdla, and Jianbo Shi, editors, Computer Vision - ACCV 2020 - 15th Asian Conference on Computer Vision, Kyoto, Japan, November 30 - December 4, 2020, Revised Selected Papers, Part VI, volume 12627 of Lecture Notes in Computer Science, pages 375-390. Springer, 2020. 3, 6 +[33] Zhiyuan You, Lei Cui, Yujun Shen, Kai Yang, Xin Lu, Yu Zheng, and Xinyi Le. A unified model for multi-class anomaly detection. CoRR, abs/2206.03687, 2022. 3, 6, 7 +[34] Sangdoo Yun, Dongyoon Han, Sanghyuk Chun, Seong Joon Oh, Youngjoon Yoo, and Junsuk Choe. Cutmix: Regularization strategy to train strong classifiers with localizable features. In 2019 IEEE/CVF International Conference on Computer Vision, ICCV 2019, Seoul, Korea (South), October 27 - November 2, 2019, pages 6022-6031. IEEE, 2019. 3 +[35] Vitjan Zavrtanik, Matej Kristan, and Danijel Skocaj. Dram - A discriminatively trained reconstruction embedding for surface anomaly detection. In ICCV, pages 8310-8319, 2021. IEEE. 2, 3, 6, 7, 8 +[36] Vitjan Zavrtanik, Matej Kristan, and Danijel Skocaj. Reconstruction by inpainting for visual anomaly detection. Pattern Recognit., 112:107706, 2021. 2 +[37] Vitjan Zavrtanik, Matej Kristan, and Danijel Skocaj. DSR - A dual subspace re-projection network for surface anomaly detection. In Shai Avidan, Gabriel J. Brostow, Moustapha Cisse, Giovanni Maria Farinella, and Tal Hassner, editors, + +Computer Vision - ECCV 2022 - 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXI, volume 13691 of Lecture Notes in Computer Science, pages 539-554. Springer, 2022. 3 +[38] Ying Zhao. Just noticeable learning for unsupervised anomaly localization and detection. In ICME, pages In Press, 2022. 1, 2, 3, 4, 5, 6, 7, 8 +[39] Yang Zou, Jongheon Jeong, Latha Pemula, Dongqing Zhang, and Onkar Dabeer. Spot-the-difference self-supervised pretraining for anomaly detection and segmentation. CoRR, abs/2207.14315, 2022. 2, 3, 5, 6, 7 \ No newline at end of file diff --git a/2023/OmniAL_ A Unified CNN Framework for Unsupervised Anomaly Localization/images.zip b/2023/OmniAL_ A Unified CNN Framework for Unsupervised Anomaly Localization/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..5471e371c69211fa8eac7d9749622609d77f2241 --- /dev/null +++ b/2023/OmniAL_ A Unified CNN Framework for Unsupervised Anomaly Localization/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:89bd76da7f078cdcf1fcca3e0e4a2763f7f33d108e49fce184bbdaf3ff301a35 +size 956428 diff --git a/2023/OmniAL_ A Unified CNN Framework for Unsupervised Anomaly Localization/layout.json b/2023/OmniAL_ A Unified CNN Framework for Unsupervised Anomaly Localization/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..2f328134db00d1594b07ead522021bd89d73b98f --- /dev/null +++ b/2023/OmniAL_ A Unified CNN Framework for Unsupervised Anomaly Localization/layout.json @@ -0,0 +1,7425 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 64, + 103, + 530, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 103, + 530, + 120 + ], + "spans": [ + { + "bbox": [ + 64, + 103, + 530, + 120 + ], + "type": "text", + "content": "OmniAL: A unified CNN framework for unsupervised anomaly localization" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 171, + 144, + 422, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 144, + 422, + 171 + ], + "spans": [ + { + "bbox": [ + 171, + 144, + 422, + 171 + ], + "type": "text", + "content": "Ying Zhao \nRicoh Software Research Center (Beijing) Co., Ltd." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 223, + 174, + 367, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 174, + 367, + 185 + ], + "spans": [ + { + "bbox": [ + 223, + 174, + 367, + 185 + ], + "type": "text", + "content": "zy_deepwhite.zy@hotmail.com" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 47, + 190, + 178, + 334 + ], + "blocks": [ + { + "bbox": [ + 47, + 190, + 178, + 334 + ], + "lines": [ + { + "bbox": [ + 47, + 190, + 178, + 334 + ], + "spans": [ + { + "bbox": [ + 47, + 190, + 178, + 334 + ], + "type": "image", + "image_path": "c578b6fe0656ce94138d5b7e9fbf860c9e9cee3afb961f16fe8c0c0a857573c7.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 59, + 336, + 167, + 347 + ], + "lines": [ + { + "bbox": [ + 59, + 336, + 167, + 347 + ], + "spans": [ + { + "bbox": [ + 59, + 336, + 167, + 347 + ], + "type": "text", + "content": "(a). N models for N classes" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 205, + 191, + 338, + 334 + ], + "blocks": [ + { + "bbox": [ + 205, + 191, + 338, + 334 + ], + "lines": [ + { + "bbox": [ + 205, + 191, + 338, + 334 + ], + "spans": [ + { + "bbox": [ + 205, + 191, + 338, + 334 + ], + "type": "image", + "image_path": "3fb174b3e377479221cb887e62bdc6dd48c4d2cd5beffbe8f5d950db2709bfd3.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 211, + 336, + 326, + 347 + ], + "lines": [ + { + "bbox": [ + 211, + 336, + 326, + 347 + ], + "spans": [ + { + "bbox": [ + 211, + 336, + 326, + 347 + ], + "type": "text", + "content": "(b). One model for N classes" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 365, + 191, + 547, + 334 + ], + "blocks": [ + { + "bbox": [ + 365, + 191, + 547, + 334 + ], + "lines": [ + { + "bbox": [ + 365, + 191, + 547, + 334 + ], + "spans": [ + { + "bbox": [ + 365, + 191, + 547, + 334 + ], + "type": "image", + "image_path": "2c4b1b4d50b021c8efd9c7817168b581752878b475b92e121c46fe25429f5258.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 408, + 336, + 497, + 347 + ], + "lines": [ + { + "bbox": [ + 408, + 336, + 497, + 347 + ], + "spans": [ + { + "bbox": [ + 408, + 336, + 497, + 347 + ], + "type": "text", + "content": "(c). Proposed OmniAL" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 45, + 353, + 546, + 386 + ], + "lines": [ + { + "bbox": [ + 45, + 353, + 546, + 386 + ], + "spans": [ + { + "bbox": [ + 45, + 353, + 546, + 386 + ], + "type": "text", + "content": "Figure 1. OmniAL is a unified framework for unsupervised anomaly localization. As shown in (c), OmniAL not only simplifies existing learning paradigm from (a) Separate: N models for N classes to (b) Unified: one model for N classes but also produces high quality anomaly localization results. Results in (a) and (b) are from separate and unified models of JNLD [38] on MVtecAD [1] dataset." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 143, + 399, + 192, + 411 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 399, + 192, + 411 + ], + "spans": [ + { + "bbox": [ + 143, + 399, + 192, + 411 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 426, + 290, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 290, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 290, + 715 + ], + "type": "text", + "content": "Unsupervised anomaly localization and detection is crucial for industrial manufacturing processes due to the lack of anomalous samples. Recent unsupervised advances on industrial anomaly detection achieve high performance by training separate models for many different categories. The model storage and training time cost of this paradigm is high. Moreover, the setting of one-model-N-classes leads to fearful degradation of existing methods. In this paper, we propose a unified CNN framework for unsupervised anomaly localization, named OmniAL. This method conquers aforementioned problems by improving anomaly synthesis, reconstruction and localization. To prevent the model learning identical reconstruction, it trains the model with proposed panel-guided synthetic anomaly data rather than directly using normal data. It increases anomaly reconstruction error for multi-class distribution by using a network that is equipped with proposed Dilated Channel and Spatial Attention (DCSA) blocks. To better localize the anomaly regions, it employs proposed DiffNeck between reconstruction and localization sub-networks to explore multi-level differences. Experiments on 15-class MVTecAD and 12-class VisA datasets verify the advantage of proposed OmniAL that surpasses the state-of-the-art of unified models. On 15-class-MVTecAD/12-class-VisA, its single unified" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 400, + 547, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 400, + 547, + 497 + ], + "spans": [ + { + "bbox": [ + 304, + 400, + 547, + 497 + ], + "type": "text", + "content": "model achieves 97.2/87.8 image-AUROC, 98.3/96.6 pixel-AUROC and 73.4/41.7 pixel-AP for anomaly detection and localization respectively. Besides that, we make the first attempt to conduct a comprehensive study on the robustness of unsupervised anomaly localization and detection methods against different level adversarial attacks. Experiential results show OmniAL has good application prospects for its superior performance." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 513, + 386, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 513, + 386, + 525 + ], + "spans": [ + { + "bbox": [ + 306, + 513, + 386, + 525 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 547, + 715 + ], + "type": "text", + "content": "In real industrial scenarios, the location of anomaly [22, 26] reveals important information, such as defective types and degrees. It is essential not only to inspect whether a sample is defective but also to know where the specific anomaly regions are. Since anomaly appearance is inexhaustible, it is almost impossible and infeasible to collect and manually annotate all kinds of abnormal data. Thus, only normal samples are available for training a detector that is robust enough to find out unseen anomalies during inference phase. Considering the diversity of classes and various types of one class, the conventional training paradigm of N models for N classes, as shown in Fig.1a, may not be the best solution. The model storage and training time cost increase with the number of classes. As shown in Fig.1b, existing method severely degrades anomaly local-" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3924" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 74, + 141, + 225 + ], + "blocks": [ + { + "bbox": [ + 53, + 74, + 141, + 225 + ], + "lines": [ + { + "bbox": [ + 53, + 74, + 141, + 225 + ], + "spans": [ + { + "bbox": [ + 53, + 74, + 141, + 225 + ], + "type": "image", + "image_path": "50640b258c33fab036f3f64210f80d9ab11981a8d04d725afa3a951cf2bbd3c1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 53, + 230, + 140, + 250 + ], + "lines": [ + { + "bbox": [ + 53, + 230, + 140, + 250 + ], + "spans": [ + { + "bbox": [ + 53, + 230, + 140, + 250 + ], + "type": "text", + "content": "(a). Reconstruction-based anomaly localization" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 140, + 73, + 181, + 81 + ], + "lines": [ + { + "bbox": [ + 140, + 73, + 181, + 81 + ], + "spans": [ + { + "bbox": [ + 140, + 73, + 181, + 81 + ], + "type": "text", + "content": "(1). MVtecAD" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 143, + 83, + 178, + 150 + ], + "blocks": [ + { + "bbox": [ + 143, + 83, + 178, + 150 + ], + "lines": [ + { + "bbox": [ + 143, + 83, + 178, + 150 + ], + "spans": [ + { + "bbox": [ + 143, + 83, + 178, + 150 + ], + "type": "image", + "image_path": "0ec8e8a7481a2e57af9bff2d53049336e23f72501496beb7ce13b8f5bcbf1f5b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 145, + 150, + 179, + 228 + ], + "blocks": [ + { + "bbox": [ + 145, + 150, + 179, + 228 + ], + "lines": [ + { + "bbox": [ + 145, + 150, + 179, + 228 + ], + "spans": [ + { + "bbox": [ + 145, + 150, + 179, + 228 + ], + "type": "image", + "image_path": "ec5d871442bb5bcab9ab9878e954aa00b3f8a0ce6c7f0583e2ffa6188f909de6.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 183, + 83, + 249, + 150 + ], + "blocks": [ + { + "bbox": [ + 187, + 73, + 246, + 81 + ], + "lines": [ + { + "bbox": [ + 187, + 73, + 246, + 81 + ], + "spans": [ + { + "bbox": [ + 187, + 73, + 246, + 81 + ], + "type": "text", + "content": "(2).JNLD Rec./Loc." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 183, + 83, + 249, + 150 + ], + "lines": [ + { + "bbox": [ + 183, + 83, + 249, + 150 + ], + "spans": [ + { + "bbox": [ + 183, + 83, + 249, + 150 + ], + "type": "image", + "image_path": "e0cc6728ca0983fa01631e35b956d01e9db29ec95686c3a3db3274fa91cc64d5.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 184, + 151, + 246, + 160 + ], + "lines": [ + { + "bbox": [ + 184, + 151, + 246, + 160 + ], + "spans": [ + { + "bbox": [ + 184, + 151, + 246, + 160 + ], + "type": "text", + "content": "X Rec. X Loc." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 145, + 160, + 179, + 228 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 160, + 179, + 228 + ], + "spans": [ + { + "bbox": [ + 145, + 160, + 179, + 228 + ], + "type": "image", + "image_path": "c80101def34cbbce53d3dc694b51a972d389cc69f43c906c54b2c1cb9150bf76.jpg" + } + ] + } + ], + "index": 8, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 181, + 161, + 248, + 236 + ], + "blocks": [ + { + "bbox": [ + 181, + 161, + 248, + 236 + ], + "lines": [ + { + "bbox": [ + 181, + 161, + 248, + 236 + ], + "spans": [ + { + "bbox": [ + 181, + 161, + 248, + 236 + ], + "type": "image", + "image_path": "bf791d08e719ee28e038ef1f8b2f335d531853360f732cfcf22dc5adcaa4fc03.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 181, + 239, + 247, + 247 + ], + "lines": [ + { + "bbox": [ + 181, + 239, + 247, + 247 + ], + "spans": [ + { + "bbox": [ + 181, + 239, + 247, + 247 + ], + "type": "text", + "content": "sis unified models o" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 129, + 254, + 462, + 265 + ], + "lines": [ + { + "bbox": [ + 129, + 254, + 462, + 265 + ], + "spans": [ + { + "bbox": [ + 129, + 254, + 462, + 265 + ], + "type": "text", + "content": "Figure 2. Problem analysis. The final failure may be caused by reconstruction and localization." + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 251, + 83, + 318, + 150 + ], + "blocks": [ + { + "bbox": [ + 252, + 73, + 321, + 81 + ], + "lines": [ + { + "bbox": [ + 252, + 73, + 321, + 81 + ], + "spans": [ + { + "bbox": [ + 252, + 73, + 321, + 81 + ], + "type": "text", + "content": "(3). OmniRec./Loc." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 251, + 83, + 318, + 150 + ], + "lines": [ + { + "bbox": [ + 251, + 83, + 318, + 150 + ], + "spans": [ + { + "bbox": [ + 251, + 83, + 318, + 150 + ], + "type": "image", + "image_path": "fbafbccb6a90f99a5d647c4eac91a3e55456db6cdd3c2c2471c88c42f73d10bb.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 253, + 151, + 315, + 160 + ], + "lines": [ + { + "bbox": [ + 253, + 151, + 315, + 160 + ], + "spans": [ + { + "bbox": [ + 253, + 151, + 315, + 160 + ], + "type": "text", + "content": "V Rec. V Loc." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 251, + 161, + 317, + 229 + ], + "blocks": [ + { + "bbox": [ + 251, + 161, + 317, + 229 + ], + "lines": [ + { + "bbox": [ + 251, + 161, + 317, + 229 + ], + "spans": [ + { + "bbox": [ + 251, + 161, + 317, + 229 + ], + "type": "image", + "image_path": "08ecfb2baeb1600c30b74357fd1c65ee1e9263b9d8d47c4bd19d038347327100.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 253, + 229, + 315, + 238 + ], + "lines": [ + { + "bbox": [ + 253, + 229, + 315, + 238 + ], + "spans": [ + { + "bbox": [ + 253, + 229, + 315, + 238 + ], + "type": "text", + "content": "V Rec. V Loc." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 331, + 83, + 373, + 150 + ], + "blocks": [ + { + "bbox": [ + 337, + 73, + 362, + 81 + ], + "lines": [ + { + "bbox": [ + 337, + 73, + 362, + 81 + ], + "spans": [ + { + "bbox": [ + 337, + 73, + 362, + 81 + ], + "type": "text", + "content": "(4).VisA" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 331, + 83, + 373, + 150 + ], + "lines": [ + { + "bbox": [ + 331, + 83, + 373, + 150 + ], + "spans": [ + { + "bbox": [ + 331, + 83, + 373, + 150 + ], + "type": "image", + "image_path": "d45e0b87cbcd0bc1c30358911abc4aa3de3a3fb05baf8831ecc8f820ab106655.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 335, + 151, + 374, + 160 + ], + "lines": [ + { + "bbox": [ + 335, + 151, + 374, + 160 + ], + "spans": [ + { + "bbox": [ + 335, + 151, + 374, + 160 + ], + "type": "text", + "content": "" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 376, + 83, + 460, + 150 + ], + "blocks": [ + { + "bbox": [ + 384, + 73, + 444, + 81 + ], + "lines": [ + { + "bbox": [ + 384, + 73, + 444, + 81 + ], + "spans": [ + { + "bbox": [ + 384, + 73, + 444, + 81 + ], + "type": "text", + "content": "(5). JNLD Rec. / Loc." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 376, + 83, + 460, + 150 + ], + "lines": [ + { + "bbox": [ + 376, + 83, + 460, + 150 + ], + "spans": [ + { + "bbox": [ + 376, + 83, + 460, + 150 + ], + "type": "image", + "image_path": "b1f3fd5eccc8c573d5873458ff579f0d4d93ade00ee6bc2252ba388dbf7d7c23.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 383, + 151, + 413, + 160 + ], + "lines": [ + { + "bbox": [ + 383, + 151, + 413, + 160 + ], + "spans": [ + { + "bbox": [ + 383, + 151, + 413, + 160 + ], + "type": "text", + "content": "X Rec." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 376, + 161, + 460, + 228 + ], + "blocks": [ + { + "bbox": [ + 376, + 161, + 460, + 228 + ], + "lines": [ + { + "bbox": [ + 376, + 161, + 460, + 228 + ], + "spans": [ + { + "bbox": [ + 376, + 161, + 460, + 228 + ], + "type": "image", + "image_path": "1190e020d530cf7132c7690d369e4ea5b7dbc9dd3e941d2be6fcaf05854bc4d7.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 383, + 228, + 412, + 237 + ], + "lines": [ + { + "bbox": [ + 383, + 228, + 412, + 237 + ], + "spans": [ + { + "bbox": [ + 383, + 228, + 412, + 237 + ], + "type": "text", + "content": "V Rec." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 376, + 238, + 511, + 248 + ], + "lines": [ + { + "bbox": [ + 376, + 238, + 511, + 248 + ], + "spans": [ + { + "bbox": [ + 376, + 238, + 511, + 248 + ], + "type": "text", + "content": "Analysis unified models on VisA dataset." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 462, + 83, + 545, + 150 + ], + "blocks": [ + { + "bbox": [ + 470, + 74, + 538, + 82 + ], + "lines": [ + { + "bbox": [ + 470, + 74, + 538, + 82 + ], + "spans": [ + { + "bbox": [ + 470, + 74, + 538, + 82 + ], + "type": "text", + "content": "(6). OmniRec./Loc." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 462, + 83, + 545, + 150 + ], + "lines": [ + { + "bbox": [ + 462, + 83, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 462, + 83, + 545, + 150 + ], + "type": "image", + "image_path": "d78c4723537b92c5c4b8820ec491c98b090650f6b939f4b944656a7537d5be3f.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 467, + 151, + 497, + 159 + ], + "lines": [ + { + "bbox": [ + 467, + 151, + 497, + 159 + ], + "spans": [ + { + "bbox": [ + 467, + 151, + 497, + 159 + ], + "type": "text", + "content": "Rec." + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 507, + 151, + 537, + 159 + ], + "lines": [ + { + "bbox": [ + 507, + 151, + 537, + 159 + ], + "spans": [ + { + "bbox": [ + 507, + 151, + 537, + 159 + ], + "type": "text", + "content": "√Loc." + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 462, + 161, + 545, + 228 + ], + "blocks": [ + { + "bbox": [ + 462, + 161, + 545, + 228 + ], + "lines": [ + { + "bbox": [ + 462, + 161, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 462, + 161, + 545, + 228 + ], + "type": "image", + "image_path": "a18cc8e45a9a81061bcfd38dbb66af81ca883edb81cd645af2654895893a6ecd.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 467, + 228, + 497, + 237 + ], + "lines": [ + { + "bbox": [ + 467, + 228, + 497, + 237 + ], + "spans": [ + { + "bbox": [ + 467, + 228, + 497, + 237 + ], + "type": "text", + "content": "V Rec." + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 507, + 228, + 537, + 236 + ], + "lines": [ + { + "bbox": [ + 507, + 228, + 537, + 236 + ], + "spans": [ + { + "bbox": [ + 507, + 228, + 537, + 236 + ], + "type": "text", + "content": "√ Loc." + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "bbox": [ + 46, + 272, + 287, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 272, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 46, + 272, + 287, + 319 + ], + "type": "text", + "content": "ization performance if the training paradigm changes to one model for N classes. Therefore, a robust unified framework for unsupervised anomaly localization is highly demanded for intelligent industrial." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 46, + 323, + 287, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 323, + 287, + 455 + ], + "spans": [ + { + "bbox": [ + 46, + 323, + 287, + 455 + ], + "type": "text", + "content": "With the limitation of available training data, many appealing unsupervised approaches [14, 18, 35, 38] using synthesized anomaly data are proposed. These approaches generate anomalous instances to inspire the anomaly detector to learn discriminative features. Their experiments show that the realiziness of generated anomalous instances had a strong impact on the quality of anomaly localization. However, none of these methods consider the training paradigm of one model for N classes. When switching to the unified training paradigm, they are more prone to learn an identical short-cut and fail to discriminate the anomaly." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 46, + 459, + 287, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 459, + 287, + 685 + ], + "spans": [ + { + "bbox": [ + 46, + 459, + 287, + 685 + ], + "type": "text", + "content": "With the normal and synthesized anomalous samples, recent unsupervised learning methods train a deep anomaly detector by either a distance-based [6, 14, 19-21, 27] or reconstruction-based [2, 9, 35, 36, 38] way. The reconstruction-based architectures [35, 38] are supposed to reconstruct normal images more accurately than the unseen anomalous. The anomaly localization is then calculated from the reconstruction error between the original and reconstructed versions of the input image, as shown in Fig.2a. The prediction of anomaly location is not only based on the reconstruction quality but also the ability of spotting the reconstruction error. The typical reconstruction-based method JNLD [38] learns a joint representation of an anomalous image and its anomaly-free reconstruction, while simultaneously learning a decision boundary between normal and simulated anomalous examples. As shown in Fig.2b and Fig.2c, under the unified setting, JNLD [38] fails to produce correct results either because of the reconstruction failure or the localization failure." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 713 + ], + "type": "text", + "content": "To conquer aforementioned problems, we propose a novel unified framework OmniAL for effectively localiz-" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 304, + 272, + 545, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 272, + 545, + 486 + ], + "spans": [ + { + "bbox": [ + 304, + 272, + 545, + 486 + ], + "type": "text", + "content": "ing anomaly pixels of different classes only by using a single model. OmniAL uses a panel-guided anomaly synthesis method that controls the portion of normal and anomaly regions for each training sample. By doing this, OmniAL blocks the chance of learning identical shortcut from the source. To increase the anomaly reconstruction error for multi-class distribution, OmniAL constructs a reconstruction and a localization sub-networks that are equipped with proposed Dilated Channel and Spatial Attention (DCSA) blocks. To better localize the anomaly regions, OmniAL employs a DiffNeck module between the reconstruction and localization sub-networks to explore multi-level reconstruction errors. As shown in Fig.1 and Fig.2, OmniAL learns a single unified model for multiple classes that produces high quality reconstruction and precise anomaly localization. Furthermore, we conduct an exhaustive evaluation of reconstruction and localization performance against multi-level adversarial attacks." + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 317, + 487, + 529, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 487, + 529, + 499 + ], + "spans": [ + { + "bbox": [ + 317, + 487, + 529, + 499 + ], + "type": "text", + "content": "In summary, we make following main contributions:" + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 316, + 506, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 42, + "blocks": [ + { + "bbox": [ + 316, + 506, + 545, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 506, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 316, + 506, + 545, + 589 + ], + "type": "text", + "content": "- We construct a unified CNN framework OmniAL for unsupervised anomaly localization that is equipped with proposed panel-guided anomaly synthesis, DCSA block, and DiffNeck module. OmniAL achieves superior performance for anomaly localization on challenging MVTecAD [1] and VisA [39] datasets compared to the state-of-the-art." + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 316, + 598, + 545, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 598, + 545, + 670 + ], + "spans": [ + { + "bbox": [ + 316, + 598, + 545, + 670 + ], + "type": "text", + "content": "- By preventing model from learning identical reconstruction, our proposed panel-guided anomaly synthesis method also brings substantial improvement for existing methods under the unified setting. It boosts the image-AUROC/pixel-AUROC/pixel-AP from 88.7/87.1/49.4 to 92.5/94.5/57.4 for Draem [35]." + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 316, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 677, + 545, + 713 + ], + "type": "text", + "content": "- We make a comprehensive study on the robustness of separate/unified anomaly localization methods against different level adversarial attacks. Our synthesized" + } + ] + } + ], + "index": 41 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3925" + } + ] + } + ], + "index": 43 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 545, + 215 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 545, + 215 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 545, + 215 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 545, + 215 + ], + "type": "image", + "image_path": "0609dd1ea28a4b3cfc8efc9cedeb888b7b397eab086b0021158a31d436678d94.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 219, + 547, + 276 + ], + "lines": [ + { + "bbox": [ + 46, + 219, + 547, + 276 + ], + "spans": [ + { + "bbox": [ + 46, + 219, + 547, + 276 + ], + "type": "text", + "content": "Figure 3. Framework of OmniAL. It consists of panel-guided anomaly synthesis, reconstruction and localization. Anomaly synthesis is based on anomaly panel and three variants of the Just Noticeable Distortion (JND) map. The synthetic anomaly is reconstructed into normal image and corresponding JND map by the Dilated Channel Spatial Attention (DCSA) modules equipped reconstruction sub-network. The localization sub-network with a DiffNeck module localizes the anomaly regions by exploring the difference between reconstructed and original data." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 66, + 281, + 288, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 281, + 288, + 328 + ], + "spans": [ + { + "bbox": [ + 66, + 281, + 288, + 328 + ], + "type": "text", + "content": "adversarial datasets exhibit strong attack capability against anomaly detection, reconstruction and localization, also helping to analyse the risks of existing methods." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 343, + 134, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 343, + 134, + 357 + ], + "spans": [ + { + "bbox": [ + 47, + 343, + 134, + 357 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 365, + 288, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 365, + 288, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 365, + 288, + 652 + ], + "type": "text", + "content": "Anomaly synthesis. Due to the lack of anomaly samples, unsupervised learning methods are commonly used in industrial quality inspection. These methods are typically trained with normal data and synthesized anomalous that highly effect the performance. Therefore, how to synthesize the anomalous [7, 8, 14, 35, 37, 38] also draws extensive attention. CutPaste [14] learns representations by classifying normal data from the Cut-and-Paste augmentation. SPD [39] uses a smoothed version of CutPaste [14] augmentation. Instead of using simple regular shaped anomalous, Draem [35] simulates just-out-of-distribution anomalies having random shape and texture. To simulate photorealistic anomaly samples, JNLD [38] proposes a multiscale noticeable anomalous generation method based on just noticeable distortion [31]. The open-set supervised anomaly detection method DRA [7] adapts the popular CutMix [34] and the outlier exposure method [10] to generate pseudo anomalies from normal images for training. DSR [37] generates the anomalies at the feature level by sampling the learned quantized feature space, which allows a controlled generation of near-in-distribution anomalies. For the unified paradigm, we propose a panel-guided anomaly synthesis method that controls the portion of normal and anomaly regions for each training sample." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 715 + ], + "type": "text", + "content": "Anomaly localization. Anomaly detection [22,26], also known as outlier detection or one-class classification, refers to the task of distinguishing defective image at the image-level from the majority of anomaly-free images. Anomaly localization (segmentation), on the other hand, aims to seg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 281, + 545, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 281, + 545, + 579 + ], + "spans": [ + { + "bbox": [ + 304, + 281, + 545, + 579 + ], + "type": "text", + "content": "ment out the pixel-level anomaly regions. For industrial visual inspection, many recent methods [5, 6, 19] achieve high performance in anomaly detection but anomaly localization. SPADE [5] detects anomaly based on alignment between an anomalous image and a constant number of the similar normal images. It relies on K nearest neighbors of pixel-level feature pyramids extracted by pre-trained deep features. PSVDD [32] extends deep Support Vector Data Description (SVDD) to a patch-wise detection method. PaDim [6] also relies on ImageNet pretrained feature extractor with multiscale pyramid pooling. Instead of using time-costly clustering, it uses a well-known Mahalanobis distance metric [17] as an anomaly score. CutPaste [14] uses GradCAM [24] to get the defect localization. MKD [23] proposes to use multi-level features alignment to increase the discriminating capability of the Teacher-Student model on various types of abnormalities. PatchCore [19] combines patch-level embeddings from ImageNet models with an outlier detection model. However, PatchCore is inherently not suitable for unified setting (larger normal dataset) since it needs to build the coreset with all normal data. Draem [35] and JNLD [38] use segmentation sub-network to predict the defective regions. This paper differs from these previous works by focusing on the paradigm of only using a single unified model for N classes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 582, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 546, + 715 + ], + "type": "text", + "content": "Unified anomaly localization. Recently, how to use a unified model to localize anomaly for different objects has already become researchers' concern. RegAD [11] trains a single generalizable model for few-shot anomaly detection, where a limited number of normal images are provided for each category at training. It employs SimSiam [3] with three spatial transformer network [13] blocks to solve the category-agnostic feature registration proxy task. It identifies anomalies by comparing the registered features of the test image and the corresponding normal images. UniAD [33] constructs a transformer with a layer-wise query de" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3926" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "content": "coder and a neighbor masked attention module to model the multi-class distribution. It uses the layer-wise query decoder to intensify the use of query embedding. To avoid the information leak, it employs a neighbor masked attention module, where a feature point relates to neither itself nor its neighbors. In this paper, we propose a unified CNN framework for unsupervised anomaly localization. It manifests the problem-solving ability of fully convolutional networks." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 182, + 103, + 194 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 182, + 103, + 194 + ], + "spans": [ + { + "bbox": [ + 47, + 182, + 103, + 194 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 203, + 289, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 203, + 289, + 396 + ], + "spans": [ + { + "bbox": [ + 46, + 203, + 289, + 396 + ], + "type": "text", + "content": "In this work, we construct a unified CNN framework OmniAL for unsupervised anomaly localization under the practical setting of only using one model for N classes. Fig.3 shows the overview of proposed OmniAL. To train only with normal data, we firstly synthesize photo-realistic anomalous in the regions that are selected based 5 predefined panels. The synthesized anomaly samples are then used to train the OmniAL network that consists of an anomaly-free reconstruction and an anomaly localization sub-networks. The reconstruction and localization sub-networks are composed of alternative basic blocks and DCSA blocks. The reconstruction sub-network learns to recover the synthesized anomalous back to normal. The localization sub-network with a DiffNeck module localizes the anomaly regions by exploring the difference between reconstructed and original data." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 406, + 219, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 406, + 219, + 418 + ], + "spans": [ + { + "bbox": [ + 47, + 406, + 219, + 418 + ], + "type": "text", + "content": "3.1. Panel-guided anomaly synthesis" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 426, + 287, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 287, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 287, + 593 + ], + "type": "text", + "content": "Reconstruction-based anomaly localization methods rely on the hypothesis that the reconstruction errors in unseen anomalous regions are larger than the normal regions. The reconstruction models are trained with alternative normal and synthetic anomaly samples. For normal sample, the supervision ground truth is the direct copy of the input. Thus, directly using normal samples for training increases the chance of information leak and leads to identical reconstruction for any input. To prevent the learning of identical short-cut, the intuitive idea is that only using synthetic anomaly rather than normal data. However, without using normal samples for training leads to more false alarm. Then, the problem is transformed into how to better synthesis anomaly for unified model training." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": "To solve the problem, we propose a panel-guided anomaly synthesis method that takes into account both normal and anomaly for training. As shown in Fig.4, we build 5 types of panels ('left', 'right', 'top', 'bottom', 'all') to control the portion of normal and anomaly regions. That is, we only generate the anomaly in one of the panel region for each sample in each iteration. Except the type of 'all', the area of panel region is also randomly adjusted in range of [0.5, 0.8]*ImageWidth and [0.5, 0.8]*ImageHeight. Given a panel, we synthesize anomaly in the selected region ac" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 307, + 70, + 547, + 191 + ], + "blocks": [ + { + "bbox": [ + 307, + 70, + 547, + 191 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 547, + 191 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 547, + 191 + ], + "type": "image", + "image_path": "8a2c312adc127bd13d9679b228f5f314f456f94944c8e7c35fc5f2ec3338b76d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 191, + 545, + 224 + ], + "lines": [ + { + "bbox": [ + 305, + 191, + 545, + 224 + ], + "spans": [ + { + "bbox": [ + 305, + 191, + 545, + 224 + ], + "type": "text", + "content": "Figure 4. Visualization of our panel-guided anomaly synthetic data. The anomaly source is selected either from augmented normal images (row1,2) or DTD [4] dataset (row3)." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 230, + 430, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 230, + 430, + 242 + ], + "spans": [ + { + "bbox": [ + 306, + 230, + 430, + 242 + ], + "type": "text", + "content": "cording to following equation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 365, + 244, + 545, + 275 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 244, + 545, + 275 + ], + "spans": [ + { + "bbox": [ + 365, + 244, + 545, + 275 + ], + "type": "interline_equation", + "content": "A = \\sum_ {i = 1} ^ {3} \\left(W _ {i} T + (1 - W _ {i}) I\\right) \\tag {1}", + "image_path": "7d4ba8878df209f6af2b7e515892906492b7996388695b23dc408ae074a4a62f.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 277, + 545, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 277, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 304, + 277, + 545, + 361 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 304, + 277, + 545, + 361 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 304, + 277, + 545, + 361 + ], + "type": "text", + "content": " is the synthesized anomaly image, " + }, + { + "bbox": [ + 304, + 277, + 545, + 361 + ], + "type": "inline_equation", + "content": "W_{i}" + }, + { + "bbox": [ + 304, + 277, + 545, + 361 + ], + "type": "text", + "content": " is the pixelwise fusion weight map, " + }, + { + "bbox": [ + 304, + 277, + 545, + 361 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 277, + 545, + 361 + ], + "type": "text", + "content": " indicates one of the 3 defect levels (easy, medium and hard), " + }, + { + "bbox": [ + 304, + 277, + 545, + 361 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 277, + 545, + 361 + ], + "type": "text", + "content": " is the anomaly-free image, " + }, + { + "bbox": [ + 304, + 277, + 545, + 361 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 277, + 545, + 361 + ], + "type": "text", + "content": " is the anomaly source image. To synthesize natural and diversity anomalies, " + }, + { + "bbox": [ + 304, + 277, + 545, + 361 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 277, + 545, + 361 + ], + "type": "text", + "content": " is generated by applying a set of random augmentation (resize, crop, flip, color-jitter) to " + }, + { + "bbox": [ + 304, + 277, + 545, + 361 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 277, + 545, + 361 + ], + "type": "text", + "content": " or image from the Describable Textures Dataset(DTD) [4]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 361, + 546, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 361, + 546, + 492 + ], + "spans": [ + { + "bbox": [ + 304, + 361, + 546, + 492 + ], + "type": "text", + "content": "To get the 3-level defects, we first need to find out the baseline anomaly fusion weight map. Following JNLD [38], we use the JND [31] to guide the anomaly synthesis. JND [31] refers to the minimum visibility threshold of the Human Visual System, is useful in perceptual image/video processing systems. It reveals a perceptual threshold of intensity change in an image that can be noticed by the human vision system. We use a nonlinear additivity JND model to define the anomaly weight map " + }, + { + "bbox": [ + 304, + 361, + 546, + 492 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 304, + 361, + 546, + 492 + ], + "type": "text", + "content": " as [25] that proposes a novel hybrid exposure weight measurement using the JND. Each level anomaly fusion weight map " + }, + { + "bbox": [ + 304, + 361, + 546, + 492 + ], + "type": "inline_equation", + "content": "W_{i}" + }, + { + "bbox": [ + 304, + 361, + 546, + 492 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 383, + 504, + 545, + 517 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 504, + 545, + 517 + ], + "spans": [ + { + "bbox": [ + 383, + 504, + 545, + 517 + ], + "type": "interline_equation", + "content": "W _ {i} = \\alpha_ {i} * (J + \\beta_ {i}) \\tag {2}", + "image_path": "98b0323c270721155302ba9eaa0de91344ba8aee40aa40408938ac01f4b88f96.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 521, + 545, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 665 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 304, + 521, + 545, + 665 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 304, + 521, + 545, + 665 + ], + "type": "text", + "content": " indicates the JND map, " + }, + { + "bbox": [ + 304, + 521, + 545, + 665 + ], + "type": "inline_equation", + "content": "\\beta_{i}" + }, + { + "bbox": [ + 304, + 521, + 545, + 665 + ], + "type": "text", + "content": " is a bias term for different level of noticeable, " + }, + { + "bbox": [ + 304, + 521, + 545, + 665 + ], + "type": "inline_equation", + "content": "\\alpha_{i}" + }, + { + "bbox": [ + 304, + 521, + 545, + 665 + ], + "type": "text", + "content": " is a smooth kernel applied to the mask and produces soft boundary between anomaly and normal regions. For the easier level anomaly, we use a smaller smooth kernel. Due to using random seeds, different level anomaly regions may overlap each other. Thus, directly using the masks as the ground truth for anomaly segmentation task may bring label inconsistency. Therefore, we refine and quantify the segmentation mask into 3 defect levels based on the structure similarity index measure (SSIM) [29] between the anomaly-free image " + }, + { + "bbox": [ + 304, + 521, + 545, + 665 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 521, + 545, + 665 + ], + "type": "text", + "content": " and synthesized anomaly image " + }, + { + "bbox": [ + 304, + 521, + 545, + 665 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 304, + 521, + 545, + 665 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 671, + 441, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 671, + 441, + 684 + ], + "spans": [ + { + "bbox": [ + 306, + 671, + 441, + 684 + ], + "type": "text", + "content": "3.2. Anomaly reconstruction" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "As shown in Fig.3, our anomaly reconstruction subnetwork receives an input image and outputs reconstructed" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3927" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 334 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 334 + ], + "type": "text", + "content": "normal image and corresponding JND map. The reconstruction sub-network consists of an encoder and a decoder. Both of them are composed of alternating basic and DCSA blocks. The basic block extracts different levels of features with two consecutive " + }, + { + "bbox": [ + 47, + 72, + 289, + 334 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 47, + 72, + 289, + 334 + ], + "type": "text", + "content": " convolution layer followed by normalization and ReLU activation layers. Considering the unified training paradigm, instance normalization [28] is more suitable than the broadly used batch normalization [12] in anomaly reconstruction. The reason is that batch normalization [12] normalizes using the information from the whole batch, while instance normalization [28] normalizes each feature map on its own. Under unified setting, each batch of training data consists of different classes that complicates the reconstruction problem. Instance normalization discards instance-specific contrast information and reduces undesirable instance-level variation. Batch normalization, meanwhile, lacks the ability to address the class inconsistency. Therefore, we use instance normalization in each basic block rather than batch normalization. Experimental results in Table 3 show that the performance is improved by swapping batch normalization with instance normalization." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 336, + 289, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 336, + 289, + 576 + ], + "spans": [ + { + "bbox": [ + 47, + 336, + 289, + 576 + ], + "type": "text", + "content": "The feature maps extracted by the basic block are further integrated by the DCSA block that is inspired by the popular CBAM [30]. As shown in orange dash block of Fig.3, the DCSA block consists of channel and spatial attention sequential sub-modules. With average pooling and two consecutive 1x1 convolutions, the channel attention explores the correlations among feature channels. The correlations are add back to original feature maps to emphasize the important feature channels. The spatial attention further investigates the spatial correlation. It first aggregates spatial information by concatenating both average and maximum pooling of afore refined feature maps. Then, it spots the spatial importance of the aggregated features with a standard and a dilated " + }, + { + "bbox": [ + 47, + 336, + 289, + 576 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 47, + 336, + 289, + 576 + ], + "type": "text", + "content": " convolution-norm-relu paths. It balances the portion of dilated path by a factor that is set as 0.1. The spatial attention map is obtained by summing up two paths of features. By adding the spatial attention map, the important regions in feature maps are highlighted. As shown in Table 3, the performance is further improved by using DCSA blocks." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 586, + 169, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 586, + 169, + 598 + ], + "spans": [ + { + "bbox": [ + 47, + 586, + 169, + 598 + ], + "type": "text", + "content": "3.3. Anomaly localization" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 605, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 605, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 605, + 287, + 714 + ], + "type": "text", + "content": "After reconstructing the normal and corresponding JND maps, we use a localization sub-network to highlight the anomaly regions. Instead of directly concatenating the reconstructed maps with input image, we firstly use a DiffNeck module to fully explore the difference between them. The pink dash block of Fig.3 illustrates the detailed structure of DiffNeck. Overall, DiffNeck extracts reconstruction difference among reconstructed normal map (Rec), reconstructed JND map (JND) and input image (Inp) in both sep" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 547, + 155 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 155 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 155 + ], + "type": "text", + "content": "arate and unified ways. That is, DiffNeck uses basic blocks to extract the concatenated features of (Rec, Inp), (JND, Inp) and (Rec, JND, Inp) respectively. Then, all levels of differences are summed up and form a final descriptor that is fed to the localization sub-network. DiffNeck also balances the portion of the separate path by a factor that is set as 0.1." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 156, + 547, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 156, + 547, + 359 + ], + "spans": [ + { + "bbox": [ + 304, + 156, + 547, + 359 + ], + "type": "text", + "content": "Similar with reconstruction sub-network, the localization sub-network also follows the encoder-decoder structure but with 6 scales (e1-e5 in Fig.3) and skip connections for corresponding scales. Scale-1 and scale-2 use standard convolution in all blocks. To use larger resolution feature maps and receptive field, the basic and DCSA blocks of the encoder contain dilated convolution layer from scale-3 to scale-6 with 2, 4, 8, 16 dilation rates respectively. Different with reconstruction sub-network, we use batch normalization in all blocks. Since we synthesize more realistic anomalies by considering easy, medium and hard level defects, localization sub-network predicts 3 levels of anomalous locations. To get the final pixel-level anomaly localization map, we combine them together by a sequence processing of SoftMax, smoothness and summing up. By doing this, the false alarm noise can be suppressed, and the weak true decision can be strengthened." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 318, + 360, + 520, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 360, + 520, + 371 + ], + "spans": [ + { + "bbox": [ + 318, + 360, + 520, + 371 + ], + "type": "text", + "content": "The total target for training OmniAL is defined as" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 331, + 373, + 545, + 401 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 373, + 545, + 401 + ], + "spans": [ + { + "bbox": [ + 331, + 373, + 545, + 401 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L = L _ {2} \\left(J, J _ {r}\\right) + L _ {s s i m} \\left(J, J _ {r}\\right) \\tag {3} \\\\ + L _ {2} (I, I _ {r}) + L _ {s s i m} (I, I _ {r}) + L _ {f l} (S, S _ {s}) \\\\ \\end{array}", + "image_path": "6ca48716d67bec568d522b1ab72357c63133cb18a232c8d19c4f0344708b5209.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 403, + 547, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 403, + 547, + 511 + ], + "spans": [ + { + "bbox": [ + 304, + 403, + 547, + 511 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 304, + 403, + 547, + 511 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 304, + 403, + 547, + 511 + ], + "type": "text", + "content": " is the ground truth JND map calculated from the anomaly-free image " + }, + { + "bbox": [ + 304, + 403, + 547, + 511 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 403, + 547, + 511 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 403, + 547, + 511 + ], + "type": "inline_equation", + "content": "J_{r}" + }, + { + "bbox": [ + 304, + 403, + 547, + 511 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 403, + 547, + 511 + ], + "type": "inline_equation", + "content": "I_{r}" + }, + { + "bbox": [ + 304, + 403, + 547, + 511 + ], + "type": "text", + "content": " are reconstructed JND map and normal image, " + }, + { + "bbox": [ + 304, + 403, + 547, + 511 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 403, + 547, + 511 + ], + "type": "text", + "content": " is the anomaly localization ground truth. Unlike JNLD [38], for both JND and normal reconstruction, we not only use MSE loss to supervise the pixel-to-pixel recovering but also the SSIM loss to yield plausible local consistency. To handle the unbalance of different types, we use focal loss [15] to supervise the predicted anomaly localization " + }, + { + "bbox": [ + 304, + 403, + 547, + 511 + ], + "type": "inline_equation", + "content": "S_{s}" + }, + { + "bbox": [ + 304, + 403, + 547, + 511 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 520, + 429, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 520, + 429, + 533 + ], + "spans": [ + { + "bbox": [ + 306, + 520, + 429, + 533 + ], + "type": "text", + "content": "4. Experimental results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 540, + 426, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 540, + 426, + 552 + ], + "spans": [ + { + "bbox": [ + 306, + 540, + 426, + 552 + ], + "type": "text", + "content": "4.1. Datasets and metrics" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 558, + 545, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 558, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 558, + 545, + 594 + ], + "type": "text", + "content": "To demonstrate the effectiveness of proposed OmniAL, we conduct extensive experiments on the challenging MVTecAD [1] and VisA [39] datasets." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "content": "MVTecAD [1] dataset contains 10 object and 5 texture industrial products, such as bottle and leather. It consists of 3,629 normal images for training and 1,725 images for testing. There are 1,258 anomaly images of the testing set with pixel-level labelled various types of defects and the rest are normal images. Each class contains 60 to 320 color images with the resolution ranges from " + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "inline_equation", + "content": "700 \\times 700" + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 304, + 594, + 547, + 714 + ], + "type": "text", + "content": " pixels. In the testing set, defective appearance varies in different sizes, shapes and types, and most cases only contain a small fraction of anomalous pixels." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3928" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 51, + 82, + 544, + 259 + ], + "blocks": [ + { + "bbox": [ + 107, + 71, + 485, + 81 + ], + "lines": [ + { + "bbox": [ + 107, + 71, + 485, + 81 + ], + "spans": [ + { + "bbox": [ + 107, + 71, + 485, + 81 + ], + "type": "text", + "content": "Table 1. Image AUROC comparison of our method OmniAL with the state-of-the-art on MVTecAD [1]." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 51, + 82, + 544, + 259 + ], + "lines": [ + { + "bbox": [ + 51, + 82, + 544, + 259 + ], + "spans": [ + { + "bbox": [ + 51, + 82, + 544, + 259 + ], + "type": "table", + "html": "
CategoryPadim [6]CutPaste [14]MKD [23]Draem [35]JNLD [38]UniAD [33]OmniAL
Unified / Separate
bottle97.9 / 99.967.9 / 98.298.7 / 99.494.6 / 99.199.1 / 96.399.7 / 100100 / 99.4
cable70.9 / 92.769.2 / 81.278.2 / 89.261.8 / 94.790.6 / 98.895.2 / 97.698.2 / 97.6
capsule73.4 / 91.363.0 / 98.268.3 / 80.570.2 / 98.574.4 / 85.086.9 / 85.395.2 / 92.4
carpet93.8 / 99.893.6 / 93.969.8 / 79.395.9 / 95.577.1 / 97.899.8 / 99.998.7 / 99.6
grid73.9 / 96.793.2 / 10083.8 / 98.098.1 / 99.998.6 / 10098.2 / 98.599.9 / 100
hazelnut85.5 / 92.080.9 / 98.397.1 / 98.495.1 / 10090.8 / 10099.8 / 99.995.6 / 98.0
leather99.9 / 10093.4 / 10093.6 / 95.199.9 / 10097.0 / 100100 / 10099.0 / 97.6
metal nut88.0 / 98.760.0 / 99.964.9 / 73.688.9 / 98.793.3 / 99.699.2 / 99.099.2 / 99.9
pill68.8 / 93.371.4 / 94.979.7 / 82.769.0 / 98.982.7 / 94.693.7 / 88.397.2 / 97.7
screw56.9 / 85.885.2 / 88.775.6 / 83.393.3 / 93.981.8 / 95.887.5 / 91.988.0 / 81.0
tile93.3 / 98.188.6 / 94.689.5 / 91.698.3 / 99.699.2 / 10099.3 / 99.099.6 / 100
toothbrush95.3 / 96.163.9 / 99.475.3 / 92.282.8 / 100100 / 10094.2 / 95.0100 / 100
transistor86.6 / 97.457.9 / 96.173.4 / 85.683.9 / 93.190.3 / 93.099.8 / 10093.8 / 93.8
wood72.1 / 96.580.4 / 99.193.4 / 94.399.8 / 99.191.9 / 99.698.6 / 97.993.2 / 98.7
zipper79.7 / 90.393.5 / 99.987.4 / 93.299.1 / 10099.8 / 99.895.8 / 96.7100 / 100
average84.2 / 95.577.5 / 96.181.9 / 87.888.7 / 98.091.3 / 97.496.5 / 96.697.2 / 97.0
", + "image_path": "964ab583725b9cd3f750bb1addf51c8e6058ecdb7681e32ca503ff6764068aab.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 51, + 277, + 544, + 454 + ], + "blocks": [ + { + "bbox": [ + 109, + 265, + 483, + 275 + ], + "lines": [ + { + "bbox": [ + 109, + 265, + 483, + 275 + ], + "spans": [ + { + "bbox": [ + 109, + 265, + 483, + 275 + ], + "type": "text", + "content": "Table 2. Pixel AUROC comparison of our method OmniAL with the state-of-the-art on MVTecAD [1]." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 51, + 277, + 544, + 454 + ], + "lines": [ + { + "bbox": [ + 51, + 277, + 544, + 454 + ], + "spans": [ + { + "bbox": [ + 51, + 277, + 544, + 454 + ], + "type": "table", + "html": "
CategoryPadim [6]PSVDD [32]MKD [23]Draem [35]JNLD [38]UniAD [33]OmniAL
Unified / Separate
bottle96.1 / 98.286.7 / 98.191.8 / 96.387.4 / 99.194.8 / 99.098.1 / 98.199.2 / 99.0
cable81.0 / 96.762.2 / 96.889.3 / 82.470.4 / 94.776.4 / 97.797.3 / 96.897.3 / 97.1
capsule96.9 / 98.683.1 / 95.888.3 / 95.949.2 / 94.357.0 / 92.798.5 / 97.996.9 / 92.2
carpet97.6 / 99.078.6 / 92.695.5 / 95.695.2 / 95.593.7 / 99.098.5 / 98.099.4 / 99.6
grid71.0 / 97.170.8 / 96.282.3 / 91.899.0 / 99.796.9 / 99.798.2 / 98.599.4 / 99.6
hazelnut96.3 / 98.197.4 / 97.591.2 / 94.696.0 / 99.785.9 / 99.496.5 / 94.698.4 / 98.6
leather84.8 / 99.093.5 / 97.496.7 / 98.198.6 / 98.687.0 / 99.598.8 / 98.399.3 / 99.7
metal nut84.8 / 97.396.0 / 98.064.2 / 86.472.6 / 99.597.4 / 99.594.8 / 95.799.1 / 99.1
pill87.7 / 95.796.5 / 95.169.7 / 89.690.0 / 97.691.2 / 96.695.0 / 95.198.9 / 98.6
screw94.1 / 98.474.3 / 95.792.1 / 96.089.3 / 97.687.0 / 99.798.3 / 97.498.0 / 97.2
tile80.5 / 94.192.1 / 91.485.3 / 82.898.1 / 99.294.7 / 99.691.8 / 91.899.0 / 99.4
toothbrush95.6 / 98.898.0 / 98.188.9 / 96.194.4 / 98.198.6 / 98.898.4 / 97.899.4 / 99.2
transistor92.3 / 97.678.5 / 97.071.7 / 76.573.1 / 90.983.6 / 92.197.9 / 98.793.3 / 91.7
wood89.1 / 94.180.7 / 90.880.5 / 84.896.2 / 96.488.7 / 96.393.2 / 93.497.4 / 96.9
zipper94.8 / 98.495.1 / 95.186.1 / 93.996.9 / 98.895.3 / 99.496.8 / 96.099.5 / 99.7
average89.5 / 97.485.6 / 95.784.9 / 97.087.1 / 97.388.6 / 97.996.8 / 96.698.3 / 97.8
", + "image_path": "a8d84f6932e980a06ea66c06a0662f295b7e3d4cb3f0cc27a2024cab616e0626.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 49, + 534, + 297, + 591 + ], + "blocks": [ + { + "bbox": [ + 46, + 468, + 287, + 533 + ], + "lines": [ + { + "bbox": [ + 46, + 468, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 468, + 287, + 533 + ], + "type": "text", + "content": "Table 3. Ablation study on MVTecAD [1]. PA: Panel-guided anomaly synthesis, Rec: Reconstruction sub-network, Seg: Segmentation sub-network, BB: Basic Block, BN: Batch Norm, IN: Instance Norm, DC: Dilated Convolution, CA: Channel Attention, DSA: Dilated Spatial Attention, I: Image-level classification, P: Pixel-level localization." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 49, + 534, + 297, + 591 + ], + "lines": [ + { + "bbox": [ + 49, + 534, + 297, + 591 + ], + "spans": [ + { + "bbox": [ + 49, + 534, + 297, + 591 + ], + "type": "table", + "html": "
PARec-BBDCSASeg-BBDiffNeckI-AUROCP-AUROCP-AP
-BNINDCCADSABNDC-Unified/Separate
-√--√-√--86.7/98.886.4/98.444.2/75.0
√√--√-√--94.1/—95.9/—68.3/—
√-√-√-√--95.8/—97.0/—69.1/—
√-√√√-√√-96.9/97.397.6/97.772.8/72.9
√-√√√√√√-94.7/96.596.9/97.869.8/74.9
√-√√√-√√√96.8/96.398.2/98.072.7/74.1
√-√√√√√√√97.2/97.098.3/97.873.4/73.5
", + "image_path": "34fdf111d3907379b3e122603a3dee248a534c8bbc8ab22a9fa06fbc19ac369b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 50, + 620, + 296, + 674 + ], + "blocks": [ + { + "bbox": [ + 46, + 597, + 287, + 620 + ], + "lines": [ + { + "bbox": [ + 46, + 597, + 287, + 620 + ], + "spans": [ + { + "bbox": [ + 46, + 597, + 287, + 620 + ], + "type": "text", + "content": "Table 4. Ablation study on With/Without panel-guided anomaly synthesis for unified training on MVTecAD [1]." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 620, + 296, + 674 + ], + "lines": [ + { + "bbox": [ + 50, + 620, + 296, + 674 + ], + "spans": [ + { + "bbox": [ + 50, + 620, + 296, + 674 + ], + "type": "table", + "html": "
Draem [35]JNLD [38]OmniAL
I-AUROC92.5/88.792.9/91.397.2/86.6
P-AUROC94.5/87.195.6/88.698.3/93.7
Pixel-AP57.4/49.463.1/46.673.4/54.4
", + "image_path": "47647881ce1f3e1074ddd94c7bad44b73b20b6258f3e51c0abdeeaae95ef48d9.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 715 + ], + "type": "text", + "content": "VisA [39] dataset consists of 10,821 high-resolution color images (9,621 normal and 1,200 anomalous samples) covering 12 objects in 3 domains, including complex struc" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 470, + 545, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 470, + 545, + 579 + ], + "spans": [ + { + "bbox": [ + 304, + 470, + 545, + 579 + ], + "type": "text", + "content": "ture, multiple instances and single instances. The anomalous images contain various flaws, including surface defects such as scratches, dents, color spots or crack, and structural defects like misplacement or missing parts. There are 5-20 images per defect type and an image may contain multiple defects. All images were acquired using a " + }, + { + "bbox": [ + 304, + 470, + 545, + 579 + ], + "type": "inline_equation", + "content": "4,000 \\times 6,000" + }, + { + "bbox": [ + 304, + 470, + 545, + 579 + ], + "type": "text", + "content": " high-resolution RGB sensor. Example of each category and our corresponding anomaly reconstruction and localization results are shown in Fig.5a(MVTecAD) and Fig.5b(VisA)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": "Metrics For anomaly detection evaluation, the most common used metrics are Area Under the Receiver Operating Characteristic curve (AUROC) in both image-level and pixel-level. However, the pixel-AUROC is biased in favour of large anomalous regions and does not well reflect the pixel-level anomalous localization performance. Therefore, we additionally introduce the average precision (AP) to evaluate pixel-level anomaly localization performance. For reconstruction quality evaluation, we use peak signal to noise ratio (PSNR) and SSIM [29] that are commonly used for evaluating signal fidelity." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3929" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 68, + 545, + 209 + ], + "blocks": [ + { + "bbox": [ + 86, + 57, + 506, + 68 + ], + "lines": [ + { + "bbox": [ + 86, + 57, + 506, + 68 + ], + "spans": [ + { + "bbox": [ + 86, + 57, + 506, + 68 + ], + "type": "text", + "content": "Table 5. Performance comparison of our method OmniAL with the state-of-the-art on VisA [39]. (Separate/Unified)" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 50, + 68, + 545, + 209 + ], + "lines": [ + { + "bbox": [ + 50, + 68, + 545, + 209 + ], + "spans": [ + { + "bbox": [ + 50, + 68, + 545, + 209 + ], + "type": "table", + "html": "
CategoryPadim+SPD [39]Draem [35]JNLD [38]OmniAL
I-AUCP-AUCI-AUCP-AUCP-API-AUCP-AUCP-API-AUCP-AUCP-AP
Complex structurePCB192.797.771.3/83.998.6/94.060.4/38.582.0/82.996.4/98.072.8/77.896.6/77.798.7/97.663.5/76.9
PCB287.997.289.7/81.792.5/94.13.5/13.396.3/79.191.9/95.031.2/37.799.4/81.083.2/93.92.8/31.6
PCB385.496.773.1/87.793.8/94.118.7/17.996.9/90.195.3/98.543.4/46.896.9/88.198.4/94.756.9/41.4
PCB499.189.291.3/87.195.8/72.332.7/13.194.8/96.296.1/97.537.4/29.797.4/95.398.5/97.138.4/33.2
Multiple instancesMacaroni185.798.870.3/68.695.8/89.88.2/8.094.3/90.598.8/93.325.9/14.296.9/92.698.9/98.67.6/7.1
Macaroni270.896.071.3/60.394.1/83.225.4/19.786.5/71.392.9/92.117.2/6.189.9/75.299.1/97.911.4/9.2
Capsules68.186.377.3/89.693.7/96.620.2/30.189.1/91.498.9/99.638.6/55.687.9/90.698.6/99.462.9/52.4
Candles89.197.382.3/70.287.0/82.627.9/12.689.1/85.494.8/94.525.9/27.485.1/86.890.5/95.829.2/24.6
Single instanceCashew90.586.194.2/67.394.7/68.541.2/7.096.0/82.596.3/94.143.7/39.497.1/88.698.9/95.077.3/47.4
Chewing gum99.396.993.4/90.097.5/92.740.9/59.098.5/96.099.4/98.974.7/81.694.9/96.498.7/99.082.9/79.5
Fryum89.888.0100/86.297.5/83.240.9/26.493.2/91.995.8/90.042.9/30.497.0/94.689.3/92.128.3/34.4
Pipe fryum95.695.494.1/87.181.8/72.323.7/13.196.0/87.597.0/92.544.1/31.891.4/86.199.1/98.269.1/62.6
Mean87.893.884.1/80.588.8/87.025.4/20.893.0/87.196.1/95.241.5/39.994.2/87.896.0/96.644.2/41.7
", + "image_path": "5caf1edee28007a2aadb0fcbb2dc7ad4baaeaf010685510264c951598f7d887f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 50, + 214, + 298, + 412 + ], + "blocks": [ + { + "bbox": [ + 50, + 214, + 298, + 412 + ], + "lines": [ + { + "bbox": [ + 50, + 214, + 298, + 412 + ], + "spans": [ + { + "bbox": [ + 50, + 214, + 298, + 412 + ], + "type": "image", + "image_path": "61f42219a66eedb889bdab0543e37967c016d4231d383af80fb771d01a2b9b71.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 124, + 412, + 235, + 421 + ], + "lines": [ + { + "bbox": [ + 124, + 412, + 235, + 421 + ], + "spans": [ + { + "bbox": [ + 124, + 412, + 235, + 421 + ], + "type": "text", + "content": "(a). 15 categories of MVtecAD dataset" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 299, + 215, + 545, + 413 + ], + "blocks": [ + { + "bbox": [ + 299, + 215, + 545, + 413 + ], + "lines": [ + { + "bbox": [ + 299, + 215, + 545, + 413 + ], + "spans": [ + { + "bbox": [ + 299, + 215, + 545, + 413 + ], + "type": "image", + "image_path": "5d36cee805ebdc32748f1134b26f9deeafa13a2cb382e93460eebc9a3eb77251.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 378, + 412, + 473, + 421 + ], + "lines": [ + { + "bbox": [ + 378, + 412, + 473, + 421 + ], + "spans": [ + { + "bbox": [ + 378, + 412, + 473, + 421 + ], + "type": "text", + "content": "(b). 12 categories of VisA dataset" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 424, + 546, + 447 + ], + "lines": [ + { + "bbox": [ + 46, + 424, + 546, + 447 + ], + "spans": [ + { + "bbox": [ + 46, + 424, + 546, + 447 + ], + "type": "text", + "content": "Figure 5. Qualitative illustration of our anomaly detection results. From top to bottom, the test images overlap with ground truth contours, the reconstructions and the anomaly localization map overlays are shown in rows." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 453, + 144, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 453, + 144, + 466 + ], + "spans": [ + { + "bbox": [ + 47, + 453, + 144, + 466 + ], + "type": "text", + "content": "4.2. Implementation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 472, + 288, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 472, + 288, + 603 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 288, + 603 + ], + "type": "text", + "content": "To compare with existing methods, we train OmniAL both in unified and separate paradigm with a batch size of 20/12 images having size of " + }, + { + "bbox": [ + 46, + 472, + 288, + 603 + ], + "type": "inline_equation", + "content": "256 \\times 256 / 256 \\times 320" + }, + { + "bbox": [ + 46, + 472, + 288, + 603 + ], + "type": "text", + "content": " for MVtec/VisA and pixel value range of [0, 1]. A single model is trained with batches that contain all-class samples for unified paradigm. On the contrary, the separate models are trained with corresponding class samples respectively. The Adam optimizer has an initial learning rate of 1e-4. To alleviate over-fitting, during training, the anomaly texture is alternatively selected either from augmented normal images or from the DTD [4] dataset." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 610, + 211, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 610, + 211, + 624 + ], + "spans": [ + { + "bbox": [ + 47, + 610, + 211, + 624 + ], + "type": "text", + "content": "4.3. Comparison and performance" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": "Performance on MVTecAD. We compare our proposed method with state-of-the-art methods recently reported on MVTecAD for both unified and separate models in anomaly detection and localization. Table 1 and 2 show our quantitative comparison with the state-of-the-art methods on the task of anomaly detection and localization. UniAD [33] is designed for the unified training paradigm while the others" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 453, + 546, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 453, + 546, + 622 + ], + "spans": [ + { + "bbox": [ + 304, + 453, + 546, + 622 + ], + "type": "text", + "content": "are for the separate training scheme. The results of unified model of Draem [35] and JNLD [38] are got by retraining their models under the paradigm of one model for N classes. The others are reported from UniAD [33]. As shown in Table 1 and 2, for both image-level and pixel-level anomaly detection, most of the existing methods' performance drop drastically when the paradigm switch from training N separate models to a unified model. Comparing with the transformer-based method UniAD [33], our CNN-based method achieves " + }, + { + "bbox": [ + 304, + 453, + 546, + 622 + ], + "type": "inline_equation", + "content": "0.7\\%" + }, + { + "bbox": [ + 304, + 453, + 546, + 622 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 453, + 546, + 622 + ], + "type": "inline_equation", + "content": "1.5\\%" + }, + { + "bbox": [ + 304, + 453, + 546, + 622 + ], + "type": "text", + "content": " higher performance on image-AUROC and pixel-AUROC. Moreover, as shown in Table 4, OmniAL(73.4) surpasses SOTA reconstruction-based methods Draem(49.4) [35] and JNLD(46.6) [38] in pixel-AP with more than " + }, + { + "bbox": [ + 304, + 453, + 546, + 622 + ], + "type": "inline_equation", + "content": "24\\%" + }, + { + "bbox": [ + 304, + 453, + 546, + 622 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": "Performance on VisA. Comparing with MVTecAD, VisA is more difficult since it considers more complex structure and multiple misaligned instances scenes. Table 5 shows the superior performance of OmniAL comparing with the baseline separate method SPD [39] that proposes the VisA dataset and two reconstruction-based methods under the unified setting. The results of both separate and uni-" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3930" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 74, + 282, + 204 + ], + "blocks": [ + { + "bbox": [ + 50, + 74, + 282, + 204 + ], + "lines": [ + { + "bbox": [ + 50, + 74, + 282, + 204 + ], + "spans": [ + { + "bbox": [ + 50, + 74, + 282, + 204 + ], + "type": "image", + "image_path": "ca1fde8e5fbed17071294e052aca400e35e6324c9d8ae52145257d03dcfaeba0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 48, + 205, + 282, + 346 + ], + "blocks": [ + { + "bbox": [ + 48, + 205, + 282, + 346 + ], + "lines": [ + { + "bbox": [ + 48, + 205, + 282, + 346 + ], + "spans": [ + { + "bbox": [ + 48, + 205, + 282, + 346 + ], + "type": "image", + "image_path": "18faa221c1f5ecc57918c08b31601c43294e181a70f0c88f66df0eb8507778e9.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 347, + 286, + 369 + ], + "lines": [ + { + "bbox": [ + 46, + 347, + 286, + 369 + ], + "spans": [ + { + "bbox": [ + 46, + 347, + 286, + 369 + ], + "type": "text", + "content": "Figure 6. Robustness comparison of our method OmniAL with the state-of-the-art against adversarial attack." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 377, + 286, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 377, + 286, + 449 + ], + "spans": [ + { + "bbox": [ + 46, + 377, + 286, + 449 + ], + "type": "text", + "content": "fied model of Draem [35] and JNLD [38] are got by retraining on VisA. For separate training, OmniAL surpasses the best of them with " + }, + { + "bbox": [ + 46, + 377, + 286, + 449 + ], + "type": "inline_equation", + "content": "1.2\\%" + }, + { + "bbox": [ + 46, + 377, + 286, + 449 + ], + "type": "text", + "content": " image-AUROC, " + }, + { + "bbox": [ + 46, + 377, + 286, + 449 + ], + "type": "inline_equation", + "content": "2.7\\%" + }, + { + "bbox": [ + 46, + 377, + 286, + 449 + ], + "type": "text", + "content": " pixel-AP and similar pixel-AUROC. For unified training, OmniAL also surpasses the best of them with " + }, + { + "bbox": [ + 46, + 377, + 286, + 449 + ], + "type": "inline_equation", + "content": "0.7\\%" + }, + { + "bbox": [ + 46, + 377, + 286, + 449 + ], + "type": "text", + "content": " image-AUROC, " + }, + { + "bbox": [ + 46, + 377, + 286, + 449 + ], + "type": "inline_equation", + "content": "1.4\\%" + }, + { + "bbox": [ + 46, + 377, + 286, + 449 + ], + "type": "text", + "content": " pixel-AUROC and " + }, + { + "bbox": [ + 46, + 377, + 286, + 449 + ], + "type": "inline_equation", + "content": "1.8\\%" + }, + { + "bbox": [ + 46, + 377, + 286, + 449 + ], + "type": "text", + "content": " pixel-AP." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 449, + 286, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 286, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 286, + 593 + ], + "type": "text", + "content": "As shown in Fig.5, our reconstruction and localization achieve good performance in most categories. Even in the multi-instance scene, OmniAL still successfully recovers the anomaly region back to normal. The reconstructed images have the high-fidelity appearance with the inputs in normal regions and recover the anomalies as close as the expectation. The quantity evaluation of reconstruction is reported in Fig.6a. The bigger circles indicate the reconstruction performance on the clean normal MVTecAD [1]. Our OmniAL surpasses the Draem [35] and JNLD [38] with big gaps in both PSNR and SSIM scores under the unified setting." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "content": "Adversarial robustness. Adversarial attacks are easily performed under the disguise of anomaly with additive noises and hardly arouse suspicion. Our anomaly synthesis method can be expeditiously expanded to generate adversarial anomaly samples with PGD [16] perturbations for robustness evaluation. To further evaluate the robustness of separate and unified models, we build up 3 adversarial datasets with PGD [16] perturbations " + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\epsilon = \\{4 / 255, 8 / 255\\}" + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\epsilon = 8 / 255" + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "content": " with targeted attack based on the anomaly-free training set from clean MVTecAD [1]. For realistic ad" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "text", + "content": "versarial samples, the synthesized anomalies only appear in foreground of the object categories and the anomaly texture is cropped from the normal images. The generated adversarial datasets consist of " + }, + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "text", + "content": " synthesized anomaly and " + }, + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 304, + 72, + 545, + 204 + ], + "type": "text", + "content": " normal samples. Fig.6 illustrates the performance comparison in (a) reconstruction and (b) localization respectively. The names with 'sp' suffix indicate separate models while the others are unified models. OmniAL(green) achieves the best performance against all adversarial levels and has less degradation. All category adversarial samples and performance comparison are shown in supplementary material." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 211, + 397, + 223 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 211, + 397, + 223 + ], + "spans": [ + { + "bbox": [ + 306, + 211, + 397, + 223 + ], + "type": "text", + "content": "4.4. Ablation study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 228, + 545, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 228, + 545, + 517 + ], + "spans": [ + { + "bbox": [ + 304, + 228, + 545, + 517 + ], + "type": "text", + "content": "Table 3 demonstrates the effectiveness of proposed panel-guided anomaly synthesis, DCSA block and Diff-Neck module for both unified and separate training paradigm. We build a baseline by only using channel attention in DCSA blocks and achieve the best performance in separate setting but the worse performance in unified setting. By using panel-guided anomaly synthesis, we achieve about " + }, + { + "bbox": [ + 304, + 228, + 545, + 517 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 304, + 228, + 545, + 517 + ], + "type": "text", + "content": " image-level and " + }, + { + "bbox": [ + 304, + 228, + 545, + 517 + ], + "type": "inline_equation", + "content": "24\\%" + }, + { + "bbox": [ + 304, + 228, + 545, + 517 + ], + "type": "text", + "content": " pixel-level performance improvement comparing with the baseline. The overall performance is further improved by using instance normalization rather than batch normalization in the reconstruction basic blocks. With the help of dilated convolution layer in basic blocks, we can preserve high resolution feature maps and get further improvement. By further using dilated spatial attention, the separate models performance gain " + }, + { + "bbox": [ + 304, + 228, + 545, + 517 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 304, + 228, + 545, + 517 + ], + "type": "text", + "content": " but the unified model decreases " + }, + { + "bbox": [ + 304, + 228, + 545, + 517 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 304, + 228, + 545, + 517 + ], + "type": "text", + "content": " in pixel-AP. Alternatively, the model using DiffNeck module gains " + }, + { + "bbox": [ + 304, + 228, + 545, + 517 + ], + "type": "inline_equation", + "content": "1.2\\%" + }, + { + "bbox": [ + 304, + 228, + 545, + 517 + ], + "type": "text", + "content": " for separate setting and " + }, + { + "bbox": [ + 304, + 228, + 545, + 517 + ], + "type": "inline_equation", + "content": "0.6\\%" + }, + { + "bbox": [ + 304, + 228, + 545, + 517 + ], + "type": "text", + "content": " for unified setting. Finally, by combining dilated spatial attention and DiffNeck together, we achieve the balanced performance in separate and unified settings and get the OmniAL. As shown in Table 4, proposed panel-guided anomaly synthesis also boosts existing reconstruction-based methods with a large margin. More ablation studies are shown in supplementary material." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 526, + 378, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 526, + 378, + 538 + ], + "spans": [ + { + "bbox": [ + 306, + 526, + 378, + 538 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 545, + 713 + ], + "type": "text", + "content": "Considering the practical usages, we propose a unified CNN framework to localize anomalous for multiple classes with a single model. Extensive experiments on MVTecAD and VisA datasets verify the effectiveness of our proposed panel-guided anomaly synthesis, dilated channel and spatial attention (DCSA) block, and DiffNeck module. Especially, OmniAL produces more precisely anomaly reconstruction and localization results. The panel-guided anomaly synthesis can be easily used in existing methods and brings performance improvement under the unified training paradigm. Moreover, we make the first attempt to conduct a comprehensive study on the robustness of existing methods against different levels of adversarial for both separate and unified training paradigm." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "3931" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 145 + ], + "type": "text", + "content": "[1] Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. Mvtec AD - A comprehensive real-world dataset for unsupervised anomaly detection. In CVPR, pages 9592-9600, 2019. Computer Vision Foundation / IEEE. 1, 2, 5, 6, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 146, + 288, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 146, + 288, + 201 + ], + "spans": [ + { + "bbox": [ + 53, + 146, + 288, + 201 + ], + "type": "text", + "content": "[2] Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. Uninformed students: Student-teacher anomaly detection with discriminative latent embeddings. In CVPR, pages 4182–4191, 2020. Computer Vision Foundation / IEEE. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 202, + 288, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 202, + 288, + 256 + ], + "spans": [ + { + "bbox": [ + 53, + 202, + 288, + 256 + ], + "type": "text", + "content": "[3] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 15750-15758. Computer Vision Foundation / IEEE, 2021. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 258, + 287, + 301 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 258, + 287, + 301 + ], + "spans": [ + { + "bbox": [ + 53, + 258, + 287, + 301 + ], + "type": "text", + "content": "[4] Mircea Cimpoi, Subhansu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In CVPR, pages 3606-3613, 2014. IEEE Computer Society. 4, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 302, + 287, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 302, + 287, + 334 + ], + "spans": [ + { + "bbox": [ + 53, + 302, + 287, + 334 + ], + "type": "text", + "content": "[5] Niv Cohen and Yedid Hoshen. Sub-image anomaly detection with deep pyramid correspondences. CoRR, abs/2005.02357. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 335, + 287, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 335, + 287, + 423 + ], + "spans": [ + { + "bbox": [ + 53, + 335, + 287, + 423 + ], + "type": "text", + "content": "[6] Thomas Defard, Aleksandr Setkov, Angelique Loesch, and Romaric Audigier. Padim: A patch distribution modeling framework for anomaly detection and localization. In Alberto Del Bimbo, Rita Cucchiara, Stan Sclaroff, Giovanni Maria Farinella, Tao Mei, Marco Bertini, Hugo Jair Escalante, and Roberto Vezzani, editors, ICPR, volume 12664 of Lecture Notes in Computer Science, pages 475-489, 2020. Springer. 2, 3, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 424, + 287, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 424, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 53, + 424, + 287, + 479 + ], + "type": "text", + "content": "[7] Chouro Ding, Guansong Pang, and Chunhua Shen. Catching both gray and black swans: Open-set supervised anomaly detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 7378-7388. IEEE, 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 480, + 287, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 480, + 287, + 534 + ], + "spans": [ + { + "bbox": [ + 53, + 480, + 287, + 534 + ], + "type": "text", + "content": "[8] Mariana-Iuliana Georgescu, Radu Tudor Ionescu, Fahad Shahbaz Khan, Marius Popescu, and Mubarak Shah. A background-agnostic framework with adversarial training for abnormal event detection in video. IEEE Trans. Pattern Anal. Mach. Intell., 44(9):4505-4523, 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 535, + 287, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 535, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 53, + 535, + 287, + 579 + ], + "type": "text", + "content": "[9] Denis A. Gudovskiy, Shun Ishizaka, and Kazuki Kozuka. CFLOW-AD: real-time unsupervised anomaly detection with localization via conditional normalizing flows. In WACV, pages 1819–1828, 2022. IEEE. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 580, + 287, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 580, + 287, + 634 + ], + "spans": [ + { + "bbox": [ + 48, + 580, + 287, + 634 + ], + "type": "text", + "content": "[10] Dan Hendrycks, Mantas Mazeika, and Thomas G. Dietterich. Deep anomaly detection with outlier exposure. In 7th International Conference on Learning Representations, ICLR 2019, New Orleans, LA, USA, May 6-9, 2019. OpenReview.net, 2019. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 678 + ], + "type": "text", + "content": "[11] Chaoqin Huang, Haoyan Guan, Aofan Jiang, Ya Zhang, Michael W. Spratling, and Yan-Feng Wang. Registration based few-shot anomaly detection. CoRR, abs/2207.07361, 2022. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 287, + 713 + ], + "type": "text", + "content": "[12] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In Francis R. Bach and David M. Blei, editors," + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 117 + ], + "type": "text", + "content": "Proceedings of the 32nd International Conference on Machine Learning, ICML 2015, Lille, France, 6-11 July 2015, volume 37 of JMLR Workshop and Conference Proceedings, pages 448-456. JMLR.org, 2015. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 118, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 118, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 308, + 118, + 545, + 205 + ], + "type": "text", + "content": "[13] Max Jaderberg, Karen Simonyan, Andrew Zisserman, and Koray Kavukcuoglu. Spatial transformer networks. In Corinna Cortes, Neil D. Lawrence, Daniel D. Lee, Masashi Sugiyama, and Roman Garnett, editors, Advances in Neural Information Processing Systems 28: Annual Conference on Neural Information Processing Systems 2015, December 7-12, 2015, Montreal, Quebec, Canada, pages 2017-2025, 2015. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 205, + 545, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 205, + 545, + 249 + ], + "spans": [ + { + "bbox": [ + 308, + 205, + 545, + 249 + ], + "type": "text", + "content": "[14] Chun-Liang Li, Kihyuk Sohn, Jinsung Yoon, and Tomas Pfister. Cutpaste: Self-supervised learning for anomaly detection and localization. In CVPR, pages 9664-9674, 2021. Computer Vision Foundation / IEEE. 2, 3, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 250, + 545, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 250, + 545, + 282 + ], + "spans": [ + { + "bbox": [ + 308, + 250, + 545, + 282 + ], + "type": "text", + "content": "[15] Tsung-Yi Lin, Priya Goyal, Ross B. Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. In CVPR, pages 2999-3007, 2017. IEEE Computer Society. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 283, + 545, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 283, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 308, + 283, + 545, + 326 + ], + "type": "text", + "content": "[16] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In ICLR, 2018. OpenReview.net. 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 327, + 545, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 327, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 308, + 327, + 545, + 348 + ], + "type": "text", + "content": "[17] Prasanta Chandra Mahalanobis. On the generalized distance in statistics. National Institute of Science of India, 1936. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 350, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 350, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 308, + 350, + 545, + 392 + ], + "type": "text", + "content": "[18] Masoud PourReza, Bahram Mohammadi, Mostafa Khaki, Samir Bouindour, Hichem Snoussi, and Mohammad Sabokrou. G2D: generate to detect anomaly. In WACV, pages 2002–2011, 2021. IEEE. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 394, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 394, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 308, + 394, + 545, + 437 + ], + "type": "text", + "content": "[19] Karsten Roth, Latha Pemula, Joaquin Zepeda, Bernhard Scholkopf, Thomas Brox, and Peter Gehler. Towards total recall in industrial anomaly detection. In CVPR, pages 14318-14328, 2022. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 437, + 545, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 437, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 308, + 437, + 545, + 480 + ], + "type": "text", + "content": "[20] Marco Rudolph, Bastian Wandt, and Bodo Rosenhahn. Same same but differnet: Semi-supervised defect detection with normalizing flows. In WACV, pages 1906-1915, 2021. IEEE. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 482, + 545, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 482, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 308, + 482, + 545, + 525 + ], + "type": "text", + "content": "[21] Marco Rudolph, Tom Wehrbein, Bodo Rosenhahn, and Bastian Wandt. Fully convolutional cross-scale-flows for image-based defect detection. In WACV, pages 1829-1838, 2022. IEEE. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 526, + 545, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 526, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 308, + 526, + 545, + 581 + ], + "type": "text", + "content": "[22] Mohammadreza Salehi, Hossein Mirzaei, Dan Hendrycks, Yixuan Li, Mohammad Hossein Rohban, and Mohammad Sabokrou. A unified survey on anomaly, novelty, open-set, and out-of-distribution detection: Solutions and future challenges. CoRR, abs/2110.14051. 1, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 582, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 582, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 308, + 582, + 545, + 657 + ], + "type": "text", + "content": "[23] Mohammadreza Salehi, Niousha Sadjadi, Soroosh Baselizadeh, Mohammad H. Rohban, and Hamid R. Rabiee. Multiresolution knowledge distillation for anomaly detection. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2021, virtual, June 19-25, 2021, pages 14902-14912. Computer Vision Foundation / IEEE, 2021. 3, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "text", + "content": "[24] Ramprasaath R. Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. Int. J. Comput. Vis., 128(2):336-359, 2020. 3" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3932" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[25] Jianbing Shen, Ying Zhao, Shuicheng Yan, and Xuelong Li. Exposure fusion using boosting laplacian pyramid. IEEE Trans. Cybern., 44(9):1579-1590, 2014. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 107, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 107, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 49, + 107, + 287, + 150 + ], + "type": "text", + "content": "[26] Xian Tao, Xinyi Gong, Xin Zhang, Shaohua Yan, and Chandranath Adak. Deep learning for unsupervised anomaly localization in industrial images: A survey. CoRR, abs/2207.10298. 1, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 288, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 288, + 195 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 288, + 195 + ], + "type": "text", + "content": "[27] Chin-Chia Tsai, Tsung-Hsuan Wu, and Shang-Hong Lai. Multi-scale patch-based representation learning for image anomaly detection and segmentation. In WACV, pages 3065-3073, 2022. IEEE. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 198, + 287, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 198, + 287, + 229 + ], + "spans": [ + { + "bbox": [ + 48, + 198, + 287, + 229 + ], + "type": "text", + "content": "[28] Dmitry Ulyanov, Andrea Vedaldi, and Victor S. Lempitsky. Instance normalization: The missing ingredient for fast stylization. CoRR, abs/1607.08022, 2016. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 232, + 287, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 232, + 287, + 275 + ], + "spans": [ + { + "bbox": [ + 48, + 232, + 287, + 275 + ], + "type": "text", + "content": "[29] Zhou Wang, Alan C. Bovik, Hamid R. Sheikh, and Eero P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Trans. Image Process., 13(4):600-612, 2004. 4, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 277, + 287, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 277, + 287, + 354 + ], + "spans": [ + { + "bbox": [ + 48, + 277, + 287, + 354 + ], + "type": "text", + "content": "[30] Sanghyun Woo, Jongchan Park, Joon-Young Lee, and In So Kweon. CBAM: convolutional block attention module. In Vittorio Ferrari, Martial Hebert, Cristian Sminchisescu, and Yair Weiss, editors, Computer Vision - ECCV 2018 - 15th European Conference, Munich, Germany, September 8-14, 2018, Proceedings, Part VII, volume 11211 of Lecture Notes in Computer Science, pages 3-19. Springer, 2018. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 355, + 287, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 355, + 287, + 398 + ], + "spans": [ + { + "bbox": [ + 48, + 355, + 287, + 398 + ], + "type": "text", + "content": "[31] Jinjian Wu, Guangming Shi, Weisi Lin, Anmin Liu, and Fei Qi. Just noticeable difference estimation for images with free-energy principle. IEEE Trans. Multim., 15(7):1705-1710, 2013. 3, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 400, + 287, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 400, + 287, + 487 + ], + "spans": [ + { + "bbox": [ + 48, + 400, + 287, + 487 + ], + "type": "text", + "content": "[32] Jihun Yi and Sungroh Yoon. Patch SVDD: patch-level SVDD for anomaly detection and segmentation. In Hiroshi Ishikawa, Cheng-Lin Liu, Tomás Pajdla, and Jianbo Shi, editors, Computer Vision - ACCV 2020 - 15th Asian Conference on Computer Vision, Kyoto, Japan, November 30 - December 4, 2020, Revised Selected Papers, Part VI, volume 12627 of Lecture Notes in Computer Science, pages 375-390. Springer, 2020. 3, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 489, + 287, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 489, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 48, + 489, + 287, + 521 + ], + "type": "text", + "content": "[33] Zhiyuan You, Lei Cui, Yujun Shen, Kai Yang, Xin Lu, Yu Zheng, and Xinyi Le. A unified model for multi-class anomaly detection. CoRR, abs/2206.03687, 2022. 3, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 523, + 287, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 523, + 287, + 588 + ], + "spans": [ + { + "bbox": [ + 48, + 523, + 287, + 588 + ], + "type": "text", + "content": "[34] Sangdoo Yun, Dongyoon Han, Sanghyuk Chun, Seong Joon Oh, Youngjoon Yoo, and Junsuk Choe. Cutmix: Regularization strategy to train strong classifiers with localizable features. In 2019 IEEE/CVF International Conference on Computer Vision, ICCV 2019, Seoul, Korea (South), October 27 - November 2, 2019, pages 6022-6031. IEEE, 2019. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 590, + 287, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 633 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 633 + ], + "type": "text", + "content": "[35] Vitjan Zavrtanik, Matej Kristan, and Danijel Skocaj. Dram - A discriminatively trained reconstruction embedding for surface anomaly detection. In ICCV, pages 8310-8319, 2021. IEEE. 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 635, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 668 + ], + "type": "text", + "content": "[36] Vitjan Zavrtanik, Matej Kristan, and Danijel Skocaj. Reconstruction by inpainting for visual anomaly detection. Pattern Recognit., 112:107706, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 712 + ], + "type": "text", + "content": "[37] Vitjan Zavrtanik, Matej Kristan, and Danijel Skocaj. DSR - A dual subspace re-projection network for surface anomaly detection. In Shai Avidan, Gabriel J. Brostow, Moustapha Cisse, Giovanni Maria Farinella, and Tal Hassner, editors," + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 195 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 117 + ], + "type": "text", + "content": "Computer Vision - ECCV 2022 - 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XXXI, volume 13691 of Lecture Notes in Computer Science, pages 539-554. Springer, 2022. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "text", + "content": "[38] Ying Zhao. Just noticeable learning for unsupervised anomaly localization and detection. In ICME, pages In Press, 2022. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 195 + ], + "type": "text", + "content": "[39] Yang Zou, Jongheon Jeong, Latha Pemula, Dongqing Zhang, and Onkar Dabeer. Spot-the-difference self-supervised pretraining for anomaly detection and segmentation. CoRR, abs/2207.14315, 2022. 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3933" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/OmniAvatar_ Geometry-Guided Controllable 3D Head Synthesis/13b82072-fc3f-4f4d-8e9e-44b54de26be7_content_list.json b/2023/OmniAvatar_ Geometry-Guided Controllable 3D Head Synthesis/13b82072-fc3f-4f4d-8e9e-44b54de26be7_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f48c8ce799d3cc2652ae218c9bfbc04f6ec7dd6b --- /dev/null +++ b/2023/OmniAvatar_ Geometry-Guided Controllable 3D Head Synthesis/13b82072-fc3f-4f4d-8e9e-44b54de26be7_content_list.json @@ -0,0 +1,1350 @@ +[ + { + "type": "text", + "text": "OmniAvatar: Geometry-Guided Controllable 3D Head Synthesis", + "text_level": 1, + "bbox": [ + 156, + 130, + 813, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hongyi Xu1 Guoxian Song1 Zihang Jiang1,2 Jianfeng Zhang1,2 Yichun Shi1 Jing Liu1 Wanchun Ma1 Jiashi Feng1 Linjie Luo1 \n1ByteDance Inc 2National University of Singapore \n{hongyixu, guoxian.song, zihang.jiang, jianfeng.zhang, yichun.shi, jing.liu, wanchun.ma, jshfeng, linjie.luo}@bytedance.com", + "bbox": [ + 148, + 178, + 818, + 270 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/09c331972bd1db073abd3f9aa8dca90f8d1112b24799daafd1a69425c8b01fb6.jpg", + "image_caption": [ + "Figure 1. Our model can synthesize diverse identity-preserved 3D heads with compelling dynamic details under full disentangled control over camera poses, facial expressions, head shapes, articulated neck and jaw poses (left). Our model can also reconstruct 3D heads from a single photo reference and enable multi-view-consistent head reenactment (right)." + ], + "image_footnote": [], + "bbox": [ + 78, + 303, + 895, + 532 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 594, + 313, + 609 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We present OmniAvatar, a novel geometry-guided 3D head synthesis model trained from in-the-wild unstructured images that is capable of synthesizing diverse identity-preserved 3D heads with compelling dynamic details under full disentangled control over camera poses, facial expressions, head shapes, articulated neck and jaw poses. To achieve such high level of disentangled control, we first explicitly define a novel semantic signed distance function (SDF) around a head geometry (FLAME) conditioned on the control parameters. This semantic SDF allows us to build a differentiable volumetric correspondence map from the observation space to a disentangled canonical space from all the control parameters. We then leverage the 3D-aware GAN framework (EG3D) to synthesize detailed shape and appearance of 3D full heads in the canonical space, followed by a volume rendering step guided by the volumetric correspondence map to output into the observation space. To ensure the control accuracy on the synthesized head", + "bbox": [ + 73, + 628, + 473, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "shapes and expressions, we introduce a geometry prior loss to conform to head SDF and a control loss to conform to the expression code. Further, we enhance the temporal realism with dynamic details conditioned upon varying expressions and joint poses. Our model can synthesize more preferable identity-preserved 3D heads with compelling dynamic details compared to the state-of-the-art methods both qualitatively and quantitatively. We also provide an ablation study to justify many of our system design choices.", + "bbox": [ + 496, + 595, + 893, + 733 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 767, + 632, + 782 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Photo-realistic face image synthesis, editing and animation attract significant interests in computer vision and graphics, with a wide range of important downstream applications in visual effects, digital avatars, telepresence and many others. With the advent of Generative Adversarial Networks (GANs) [15], remarkable progress has been achieved in face image synthesis by StyleGAN [23-25]", + "bbox": [ + 496, + 794, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "12814", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "as well as in semantic and style editing for face images [46, 54]. To manipulate and animate the expressions and poses in face images, many methods attempted to leverage 3D parametric face models, such as 3D Morphable Models (3DMMs) [6, 40], with StyleGAN-based synthesis models [10, 41, 53]. However, all these methods operate on 2D convolutional networks (CNNs) without explicitly enforcing the underlying 3D face structure. Therefore they cannot strictly maintain the 3D consistency when synthesizing faces under different poses and expressions.", + "bbox": [ + 75, + 90, + 472, + 242 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, a line of work has explored neural 3D representations by unsupervised training of 3D-aware GANs from in-the-wild unstructured images [7, 8, 11, 17, 37, 44, 45, 48, 61, 62, 67]. Among them, methods with generative Neural Radiance Fields (NeRFs) [33] have demonstrated striking quality and multi-view-consistent image synthesis [7, 11, 17, 37, 48]. The progress is largely due to the integration of the power of StyleGAN in photo-realistic image synthesis and NeRF representation in 3D scene modeling with view-consistent volumetric rendering. Nevertheless, these methods lack precise 3D control over the generated faces beyond camera pose, as well as the quality and consistency in control over other attributes, such as shape, expression, neck and jaw pose, leave much to be desired.", + "bbox": [ + 75, + 244, + 472, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we present OmniAvatar, a novel geometry-guided 3D head synthesis model trained from in-the-wild unstructured images. Our model can synthesize a wide range of 3D human heads with full control over camera poses, facial expressions, head shapes, articulated neck and jaw poses. To achieve such high level of disentangled control for 3D human head synthesis, we devise our model learning in two stages. We first define a novel semantic signed distance function (SDF) around a head geometry (i.e. FLAME [29]) conditioned on its control parameters. This semantic SDF fully distills rich 3D geometric prior knowledge from the statistical FLAME model and allows us to build a differentiable volumetric correspondence map from the observation space to a disentangled canonical space from all the control parameters. In the second training stage, we then leverage the state-of-the-art 3D GAN framework (EG3D [7]) to synthesize realistic shape and appearance of 3D human heads in the canonical space, including the modeling of hair and apparels. Following that, a volume rendering step is guided by the volumetric correspondence map to output the geometry and image in the observation space.", + "bbox": [ + 75, + 459, + 472, + 777 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To ensure the consistency of synthesized 3D head shape with controlling head geometry, we introduce a geometry prior loss to minimize the difference between the synthesized neural density field and the FLAME head SDF in observation space. Furthermore, to improve the control accuracy, we pre-train an image encoder of the control parameters and formulate a control loss to ensure synthesized images matching the input control code upon encoding. An-", + "bbox": [ + 75, + 779, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "other key aspect of synthesis realism is dynamic details such as wrinkles and varying shading as subjects change expressions and poses. To synthesize dynamic details, we propose to condition EG3D's triplane feature decoding with noised controlling expression.", + "bbox": [ + 496, + 90, + 893, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Compare to state-of-the-art methods, our method achieves superior synthesized image quality in terms of Frechet Inception Distance (FID) and Kernel Inception Distance (KID). Our method can consistently preserve the identity of synthesized subjects with compelling dynamic details while changing expressions and poses, outperforming prior methods both quantitatively and qualitatively.", + "bbox": [ + 496, + 167, + 893, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The contributions of our work can be summarized as:", + "bbox": [ + 517, + 272, + 872, + 287 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- A novel geometry-guided 3D GAN framework for high-quality 3D head synthesis with full control on camera poses, facial expressions, head shapes, articulated neck and jaw poses.", + "- A novel semantic SDF formulation that defines the volumetric correspondence map from observation space to canonical space and allows full disentanglement of control parameters in 3D GAN training.", + "- A geometric prior loss and a control loss to ensure the head shape and expression synthesis accuracy.", + "- A robust noised expression conditioning scheme to enable dynamic detail synthesis." + ], + "bbox": [ + 517, + 289, + 890, + 477 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 489, + 640, + 506 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D-Aware Generative Image Synthesis. Generative adversarial networks [15, 24, 25] gained popularity over the last decade due to their remarkable ability in photo-realistic image synthesis. Building on the success of 2D image-based GANs, recent works have extended the capabilities to view-consistent image synthesis with unsupervised learning from 2D single-view images. The key idea is to combine differential rendering with 3D scene representations, such as meshes [31, 50], point clouds [1, 28], voxels [34, 35, 57], and recently implicit neural representation [7, 8, 11, 17, 36, 37, 44, 48, 67]. We build our work on recent 3D GAN model by Chan et al [7] that uses an efficient triplane-based NeRF generation, combined with 2D CNN-based super-resolution. Even though 3D-aware GANs are able to control camera viewpoints, they lack precise 3D control over the other attributes such as shapes and expressions. In this work, we empower 3D-aware GANs with disentangled precise control over shapes and expressions.", + "bbox": [ + 496, + 515, + 893, + 787 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Controllable Face Image Synthesis. Considerable work [9, 10, 14, 27, 41, 52, 53] has been devoted to incorporate 3D priors of statistical face models, such as 3D Morphable Models (3DMMs) [6, 40], in controllable face synthesis and animation. Among them, DiscoFaceGAN [10] proposed imitative-contrastive learning to mimic the 3DMM rendering process by the generative model.", + "bbox": [ + 496, + 794, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "12815", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b04fe5cf49739ad18a3cd42dd0bec0dfdd222f7a0ed9e98008660542263f19a0.jpg", + "image_caption": [ + "Figure 2. Overview of our training framework. Stage 1: Trained from parameterized FLAME [29] mesh collections, a MLP-network $W$ maps a shape $\\alpha$ , expression $\\theta$ and articulated jaw and neck pose $\\theta$ into 3D point-to-point volumetric correspondences from observation to canonical space, together with a signed distance function of the corresponding FLAME head. Stage 2: Given a Gaussian latent code $\\mathbf{z}$ , our model generates a tri-plane represented 3D feature space of a canonical head, disentangled with shape and expression controls. The volume rendering is then guided by the volumetric correspondence field to map the decoded neural radiance field from the canonical to observation space. We condition the NeRF decoding with expression and joint pose for modeling dynamic details. A super-resolution module synthesizes the final high-resolution RGB image from the volume-rendered feature map. For fine-grained shape and expression control, we apply the FLAME SDF as geometric prior to the synthesized NeRF density, and self-supervise the image synthesis to commit to the target expression $\\beta$ and joint pose $\\theta$ by comparing the input code against the re-estimated values $\\hat{\\beta}, \\hat{\\theta}$ from synthesized images." + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 893, + 344 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A similar strategy has also been adopted with concurrent and follow-up works [14, 27, 41, 53]. However all of these approaches suffer from 3D inconsistency due to the use of 2D CNNs as image renderer. HeadNeRF [21] combines 3DMM with 3D NeRF representation, and is able to synthesize 3D heads conditioned on 3DMM attributes. However, the training relies on annotated multiview datasets whereas our approach learns the disentangled 3D head synthesis with only single-view images. There has been concurrent work [4, 49, 51, 58, 64] to ours in 3D-aware controllable face or full-body GANs. Differently from these approaches, we use a full-head parametric model FLAME [29], and fully exploit the spatial geometric prior knowledge beyond the surface deformation and skinning. We have also achieved fine-grained control with our novel losses and enhanced face animation with rich dynamic details.", + "bbox": [ + 75, + 496, + 472, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Controllable Neural Implicit Field of Face. Neural implicit functions [59], have emerged as a powerful continuous and differential representations of 3D scenes. Among them, Neural Radiance Field [3, 33] has been widely adopted due to its superiority in modeling complex scene details and synthesizing multiview images with inherited 3D consistency. While initial proposals have focused on static scene modeling, recent work has successfully demonstrated application of NeRF in modeling dynamic scenes [32,38,42,55,60]. In particular, dynamic neural radi-", + "bbox": [ + 75, + 750, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ance fields of human heads [13,18,22,38,39,56,66,68] have enabled photo-realistic head animation, often by conditioning with pose parameters or deforming the radiance field with 3D morphable models. However, they do not leverage a generative training paradigm and focus on scene-specific learning from video sequences or multiview images. In contrast, our model learns generative and controllable neural radiance fields from widely accessible single-view images.", + "bbox": [ + 496, + 496, + 893, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 631, + 591, + 647 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our goal is to build a geometry-guided 3D head synthesis model with full control of camera poses, face shapes and expressions, trained from in-the-wild unstructured image collections.", + "bbox": [ + 496, + 657, + 893, + 717 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 500, + 728, + 609, + 742 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Problem. To achieve our goal, we leverage a 3D-aware generator (EG3D [7]) for photo-realistic, multiview consistent image synthesis, while disentangle control of head geometric attributes from image generation with a 3D statistical head model (FLAME [29]). Specifically, given a random Gaussian-sampled latent code $\\mathbf{z}$ , a camera pose $\\mathbf{c}$ and a FLAME parameter $\\mathbf{p} = (\\alpha, \\beta, \\theta)$ consisting of shape $\\alpha$ , expression $\\beta$ , jaw and neck pose $\\theta$ , the generator $G$ synthesizes a photo-realistic human head image $I_{RGB}(\\mathbf{z}|\\mathbf{c},\\mathbf{p})$ with corresponding attributes as defined in $\\mathbf{p}$ .", + "bbox": [ + 496, + 751, + 893, + 904 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "12816", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "**Framework.** As illustrated in our pipeline 2, our controllable 3D-aware GAN is trained in two stages. From a large collection of 3D deformed FLAME meshes, we first pre-train a deformable semantic SDF around the FLAME geometry that builds a differential volumetric correspondence map from the observation to a predefined canonical space (Section 3.2.1). In the second stage, guided by the pre-trained volumetric mapping, we then deform the detailed 3D full heads synthesized in the disentangled canonical space to the desired shapes and expressions (Section 3.2.2). Fine-grained expression control is achieved by supervising image synthesis such that expressions estimated from the generated images is consistent with the input control (Section 3.2.3). Our approach further enhances temporal realism with dynamic details, such as dimples and wrinkles, synthesizing realistic shading and geometric variations as expression changes (Section 3.2.4).", + "bbox": [ + 75, + 90, + 472, + 349 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3D-Aware GAN Background. To ensure appearance consistency from different views, we choose EG3D [7] as our backbone for 3D-aware image synthesis. The generator $G$ takes a random latent code $\\mathbf{z}$ and conditioning camera label $\\hat{\\mathbf{c}}$ , and maps to a manifold of triplane features $\\mathcal{M}(\\mathbf{z},\\hat{\\mathbf{c}})$ . For presentation clarity, we absorb $\\hat{\\mathbf{c}}$ to $\\mathbf{z}$ and simply denote triplane as $\\mathcal{M}(\\mathbf{z})$ . A low-resolution feature map $I^{*}(\\mathbf{z}|\\mathbf{c})$ is then rendered from a desired camera pose $\\mathbf{c}$ by sampling the triplane features and integrating MLP-decoded neural radiance $(\\sigma ,\\mathbf{f})$ along camera rays. A super-resolution module is followed to modulate the feature map and synthesize the final RGB images at high resolution. We train $G$ and a dual discriminator $D$ with adversarial training.", + "bbox": [ + 75, + 357, + 472, + 554 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Controllable 3D-Aware Image Synthesis", + "text_level": 1, + "bbox": [ + 76, + 561, + 419, + 579 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To synthesize an image $\\mathbf{I}(\\mathbf{z}|\\mathbf{c},\\mathbf{p})$ with desired FLAME control $\\mathbf{p}$ , we leverage the EG3D framework to generate a triplane-based 3D volumetric feature space $\\mathcal{M}(\\mathbf{z})$ of a synthesized identity with canonical shape and expression. Guided by our pretrained volumetric correspondence map, we deform the synthesized feature volume into our target observation space, which is further decoded and volume rendered into high-fidelity head appearance and geometry with the target shape and expression. Our design explicitly disentangles the underlying geometric variations of changing shape and expression from canonical geometry and appearance synthesis. Following EG3D [7], we associate each training image with a set of camera parameters $\\mathbf{c}$ and control parameters $\\mathbf{p}$ , which are obtained from a nonlinear 2D landmarks-based optimization.", + "bbox": [ + 75, + 585, + 470, + 813 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.1 Semantic Signed Distance Function", + "text_level": 1, + "bbox": [ + 76, + 830, + 377, + 847 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For disentangled geometric modeling, we formulate an implicit semantic SDF representation $W(\\mathbf{x}|\\mathbf{p} = (\\alpha ,\\beta ,\\theta)) = (s,\\bar{\\mathbf{x}})$ , where $\\alpha ,\\beta$ are the linear shape and expression", + "bbox": [ + 76, + 854, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5375ab76ac4b310a1596b83d56b7d77f1d6b31faff19c0e4c5b218a5c3d05ea5.jpg", + "image_caption": [ + "Figure 3. We synthesize different identities with the same shape and expression code respectively in (a) and (b). We observe less shape variation with our geometric prior loss (a) and more consistent expression with our control loss (b). In (c), we show our expression-conditioned NeRF modeling builds rich dynamic details as the subject varies expressions." + ], + "image_footnote": [], + "bbox": [ + 500, + 88, + 911, + 349 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "blendshape coefficients, and $\\theta$ controls the rotation of a 3-DoF jaw and neck joint. Specifically, given a spatial point $\\mathbf{x}$ in observation space $\\mathcal{O}(\\mathbf{p})$ , $W$ returns its 3D correspondence point $\\bar{\\mathbf{x}}$ (i.e., semantics) in canonical space $\\mathcal{C}(\\bar{\\mathbf{p}})$ , with which we project and query the triplane features $\\mathcal{M}(\\mathbf{z})$ . Additionally it also computes the closest signed distance $s(\\mathbf{x}|\\mathbf{p})$ to the FLAME mesh surface $\\mathbf{S}(\\mathbf{p})$ . We illustrate the function in Figure 4. We co-learn the highly-correlated volumetric correspondence and SDF with the property that the signed distance is preserved between canonical and observation correspondence points as $s(\\bar{\\mathbf{x}}|\\bar{\\mathbf{p}}) = s(\\mathbf{x}|\\mathbf{p})$ .", + "bbox": [ + 496, + 453, + 890, + 619 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We learn $W(\\mathbf{x}|\\mathbf{p})$ with a large corpus of 3D FLAME meshes $\\mathbf{S}(\\mathbf{p})$ sampled from its parametric control space. Similar to IGR [16] and imGHUM [2], we model our implicit field as an MLP and optimize $W(\\mathbf{x}|\\mathbf{p})$ with losses,", + "bbox": [ + 496, + 619, + 890, + 680 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {i s o} = \\frac {1}{| N |} \\sum_ {\\mathbf {x} \\in N} (| s (\\mathbf {x} | \\mathbf {p}) | + | \\nabla s _ {\\mathbf {x}} (\\mathbf {x} | \\mathbf {p}) - \\mathbf {n} (\\mathbf {x} | \\mathbf {p}) |), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 700, + 890, + 738 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {e i k} = \\frac {1}{| F |} \\sum_ {\\mathbf {x} \\in F} \\| \\nabla s _ {\\mathbf {x}} (\\mathbf {x} | \\mathbf {p}) - 1 \\| _ {2}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 739, + 890, + 776 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {s e m} = \\frac {1}{| N |} \\sum_ {\\mathbf {x} \\in N} \\left(| \\bar {\\mathbf {x}} (\\mathbf {x} | \\mathbf {p}) - \\bar {\\mathbf {x}} ^ {*} (\\mathbf {x} | \\bar {\\mathbf {p}}) |\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 779, + 890, + 815 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $N,F$ are a batch of on and off surface samples. For the surface samples, the $L_{iso}$ encourages the signed distance values to be on the zero-level-set and the SDF gradient to be equal to the given surface normals $\\mathbf{n}$ . The Eikonal loss $L_{eik}$ is derived from [16] where the SDF is differen", + "bbox": [ + 500, + 824, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "12817", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/637a25f6e637a9c0a2a20adac5ac22048819dac91d8fd54ab84b4a827ea6aa24.jpg", + "image_caption": [ + "Figure 4. Illustration to semantic SDF learning." + ], + "image_footnote": [], + "bbox": [ + 153, + 88, + 401, + 247 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "tiable everywhere with gradient norm 1. The semantic loss $L_{sem}$ supervises the mapping of surface samples $\\mathbf{x} \\in F$ to the ground-truth correspondence points $\\bar{\\mathbf{x}}^*$ on the canonical FLAME surface, where $\\bar{\\mathbf{x}}^*$ and $\\mathbf{x}$ share the same barycentric coordinates. The $L_{eik}$ provides a geometric regularization over the volumetric SDF, whereas the $L_{iso}$ and $L_{sem}$ act as the boundary condition to the SDF and volumetric correspondence field respectively.", + "bbox": [ + 75, + 282, + 468, + 404 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To guide the canonical correspondence learning for off-the-surface points, one could try to minimize the signed distance difference between $s(\\bar{\\mathbf{x}}|\\bar{\\mathbf{p}})$ and $s(\\mathbf{x}|\\mathbf{p})$ . However, we note that the volume correspondence between the observation and canonical space is still ill-regularized, considering infinite number of points exist with the same signed distance value. We therefore reformulate volumetric correspondence field as $W(\\mathbf{x}|\\mathbf{p}) = (\\bar{s}(\\bar{\\mathbf{x}}), \\bar{\\mathbf{x}})$ , where the signed distance of an observation-space point is obtained by mapping it into its canonical correspondence $\\bar{\\mathbf{x}}$ and querying the pre-computed canonical SDF $\\bar{s}$ . Thus we only learn a correspondence field with which we can deform the canonical SDF to different FLAME configurations, and then supervise the deformed SDFs with the FLAME surface boundary, normals $(L_{iso})$ and Eikonal regularization $(L_{eik})$ . As such, even for off-the-surface samples, their canonical correspondences are well regularized in space with the geometric properties of signed distance functions via $L_{iso}$ and $L_{eik}$ . In contrast to explicit mesh deformation [4], our implicit volumetric correspondence is more accurate, differentiable, smooth everywhere and semantically consistent with the properties of SDF.", + "bbox": [ + 75, + 405, + 470, + 723 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.2 Canonical Generation with Geometric Prior", + "text_level": 1, + "bbox": [ + 76, + 739, + 437, + 753 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With our pretrained semantic SDF $W(\\mathbf{x}|\\mathbf{p})$ modeling the shape and expression variation, we leverage the triplane for 3d-aware generation of human heads with canonical shape and expression. In particular, to generate a neural radiance feature $(\\sigma, \\mathbf{f})$ for a point $\\mathbf{x}$ in observation space $\\mathcal{O}(\\mathbf{p})$ , we use our correspondence function $W$ to back warp $\\mathbf{x}$ into $\\bar{\\mathbf{x}}$ , with which we project and sample the canonical triplane features followed with a tiny MLP decoding.", + "bbox": [ + 75, + 763, + 468, + 883 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In spite of the control disentanglement, there is no ex", + "bbox": [ + 94, + 885, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "plicit loss that constrains the triplane-generated neural radiance density to conform to the shape and expression as defined in $\\mathbf{p}$ . Therefore, we guide the generation of neural radiance density field by minimizing its difference to the underlying FLAME head geometry represented with SDF, as", + "bbox": [ + 496, + 90, + 890, + 179 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {p r i o r} = \\frac {1}{| R |} \\sum_ {\\mathbf {x} \\in R} e ^ {- \\gamma | s (\\mathbf {x} | p) |} \\left| \\sigma \\left(\\mathbf {x} \\mid \\mathbf {z}, \\mathbf {p}\\right) - \\sigma^ {*} \\left(\\mathbf {x} \\mid \\mathbf {p}\\right) \\right|, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 185, + 890, + 223 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\sigma^ {*} (\\mathbf {x} | \\mathbf {p}) = \\frac {1}{\\kappa} \\cdot \\operatorname {S i g m o i d} \\left(\\frac {- s (\\mathbf {x} | \\mathbf {p})}{\\kappa}\\right) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 224, + 890, + 255 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $R$ is the stratified ray samples for volume rendering and $\\kappa$ is a learnable scalar controlling the density tightness around the SDF boundary. Following [37, 63], we convert SDF value $s(\\mathbf{x}|\\mathbf{p})$ to proxy 3D density $\\sigma^{*}(\\mathbf{x}|\\mathbf{p})$ assuming non-hollow surfaces. We decay the weights for our geometric prior loss $L_{prior}$ as the point moving away from the SDF boundary, allowing higher degrees of freedom in generation of residual geometries, such as hair and glasses. The geometric prior loss effectively guides the 3D head geometry learning but should not be overpowered which otherwise might lead to loss of geometric details.", + "bbox": [ + 496, + 261, + 890, + 426 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.3 Fine-Grained Expression Control", + "text_level": 1, + "bbox": [ + 498, + 444, + 787, + 459 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our geometric prior loss $L_{prior}$ provides local 3D pointwise guidance, and is able to well regularize the shape generation and achieve coarse-level expression control. However, for delicate expressions, such as eye blinks, $L_{prior}$ provides little supervision as the geometric variation is subtle. Moreover, for regions with complex correspondences, such as around the lips, it is challenging to guide the formation of correct expressions globally, just with point-wise geometric losses. To improve the control granularity, we propose an image-level supervision loss that requires a synthesized image $I_{RGB}(\\mathbf{z}|\\mathbf{c},\\mathbf{p})$ matching the target expression as defined in the input $\\mathbf{p}$ . Using our training images with estimated control labels $\\mathbf{p}$ , we first pretrain an image encoder $E(I_{RGB}) = (\\tilde{\\boldsymbol{\\beta}},\\tilde{\\boldsymbol{\\theta}})$ that regresses the expression coefficients $\\tilde{\\boldsymbol{\\beta}}$ and joint poses $\\tilde{\\boldsymbol{\\theta}}$ . During our 3D GAN training, we then apply our image-level control supervision as", + "bbox": [ + 496, + 468, + 890, + 710 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {e n c} = | \\tilde {\\boldsymbol {\\beta}} - \\boldsymbol {\\beta} | + | \\tilde {\\boldsymbol {\\theta}} - \\boldsymbol {\\theta} | + | \\mathbf {S} (\\boldsymbol {\\alpha}, \\tilde {\\boldsymbol {\\beta}}, \\tilde {\\boldsymbol {\\theta}}) - \\mathbf {S} (\\boldsymbol {\\alpha}, \\boldsymbol {\\beta}, \\boldsymbol {\\theta}) | + \\\\ + \\left| \\mathbf {J} \\mathbf {S} (\\boldsymbol {\\alpha}, \\tilde {\\boldsymbol {\\beta}}, \\tilde {\\boldsymbol {\\theta}}) - \\mathbf {J} \\mathbf {S} (\\boldsymbol {\\alpha}, \\boldsymbol {\\beta}, \\boldsymbol {\\theta}) \\right|, \\tag {6} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 503, + 731, + 890, + 771 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{S},\\mathbf{J}$ are the FLAME mesh vertices and 3D landmarks regressor respectively. While being straightforward for the first 2 terms, the last two terms in $L_{enc}$ penalize deviation of 3D vertex coordinates and surface landmarks after mesh decoding. We note that we do not supervise shape $\\alpha$ in $L_{enc}$ since our geometric prior loss $L_{prior}$ suffices in shape control already and also due to the ambiguity of shape scaling estimated from monocular images.", + "bbox": [ + 496, + 779, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "12818", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5c3066ab5d5b8315700701851a923395642694462106b4973e7799cd2ebb3d09.jpg", + "image_caption": [ + "Figure 5. Qualitative comparisons in expression control. Under a similar expression control, our method achieves the best perceptual quality with better identity-preservation (compared to GAN-Control [47] and 3DFaceShop [51]), more realistic dynamic details (compared to DiscoFaceGAN [10]) and higher expression control accuracy (compared to AntiFaceGAN [58], 3DFaceShop [51], HeadNeRF [21])." + ], + "image_footnote": [], + "bbox": [ + 86, + 88, + 354, + 402 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/583be0f7e1981a934d961356d719a966e245573b4936a53abc3f71932f68c92e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 383, + 87, + 468, + 402 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/cdccdc8738c51f4169406576d78e645eb54363672708f337e347e00ebf695203.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 488, + 88, + 620, + 402 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ff1b0720b4720eb7fc5a2bbd5872d0512799ca609fe9d427cc79e6b920c9ad9c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 622, + 88, + 753, + 402 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9d212897a27f4e529472d14ad6dd6482eae3e88b967af1c665fb2d16de795aa8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 754, + 88, + 887, + 402 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2.4 Dynamic Details Modeling", + "text_level": 1, + "bbox": [ + 76, + 481, + 318, + 498 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To this end, we have achieved controllable static image synthesis. However, one should observe the variation of shading and geometric details in a dynamic head motion, such as the appearance of dimples and eye wrinkles in smiling. We consider the appearance of dynamic details is highly correlated with the driving expression $(\\beta ,\\theta)$ . To model such temporal effects, one could try to condition the triplane generation on expression as $\\mathcal{M}(\\mathbf{z},\\beta ,\\theta)$ . However, such design results in entanglement of expression control with the generative latent code $\\mathbf{z}$ , inducing identity or appearance changes when varying expressions. It is also hard to synthesize images out of training distribution of expressions $(\\beta ,\\theta)$ . We therefore leave our triplane generation disentangled but when decoding neural feature field $(\\sigma ,\\mathbf{f})$ from sampled triplane features, we additionally condition the MLP-decoder on $\\beta$ and $\\theta$ . Specifically, we have", + "bbox": [ + 75, + 517, + 472, + 758 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n(\\sigma (\\mathbf {x}), \\mathbf {f} (\\mathbf {x}) | \\mathbf {z}, \\mathbf {p}) = \\Phi (\\mathcal {M} (\\bar {\\mathbf {x}} (\\mathbf {x} | \\mathbf {p}) | \\mathbf {z}), \\phi (\\boldsymbol {\\beta}, \\boldsymbol {\\theta})), \\quad (7)\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 782, + 468, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where both $\\Phi$ and $\\phi$ are tiny MLPs and $\\phi$ regresses an expression-dependent feature vector from $(\\beta, \\theta)$ after positional encoding. For better extrapolation to novel expressions and jaw poses, we add Gaussian noise to the conditioning parameters to prevent MLP overfitting.", + "bbox": [ + 75, + 824, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 479, + 633, + 498 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training and Dataset. Our training is devised into two stages. In the pretraining stage, we build our semantic SDF $W$ using a four 192-dimensional MLP with a collection of 150K FLAME meshes by Gaussian sampling of the control parameters $\\mathbf{p}$ . For each mesh, we sample 4K surface samples $N$ with surface normals $\\mathbf{n}$ and ground-truth correspondence point $\\bar{\\mathbf{x}}^*$ , and 4K off surface samples $F$ distributed uniformly inside the box-bounded volume. Our canonical mesh has a neutral shape $\\alpha = 0$ , expression $\\beta = 0$ and neck pose, but with an opening jaw at 10 degrees. We close the face connectivity between the lips for a watertight geometry and also as a proxy geometry for the mouth cavity modeling. We open the jaw at the canonical space such that the spatial geometry and appearance inside the mouth can be modeled with distinguishable triplane features. In the second training stage, we freeze the weights of $W$ for best efficiency and train our model on the FFHQ [24], a human face dataset with 70,000 in-the-wild images. For each image, we estimate the camera pose $\\mathbf{c}$ and FLAME parameters $\\mathbf{p}$ with a nonlinear optimization in fitting 2D landmarks, assuming zero root and neck rotation and camera located 1.0m away from the world origin. We refer to the supplementary materials for more details. We rebalance the FFHQ dataset by duplicating images with large head poses and jaw opening expressions. We note that we also use pairs of images and $(\\beta, \\theta)$ for fine-tuning of a light-weight image encoder", + "bbox": [ + 496, + 507, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "12819", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/3bcec1d3dfc5a935a21c4daea86416978f3f43809b8fe1861325886bac6d26bc.jpg", + "table_caption": [], + "table_footnote": [ + "(a) Baseline comparisons." + ], + "table_body": "
ModelFID ↓KID ↓DSα ↑DSβ,θ ↑DSc ↑IS ↑
DiscoFaceGAN-SR [10]42.60.04-27.20.110.59
GAN-Control [47]9.80.005-43.70.280.62
HeadNeRF [21]1630.1681.338.90.250.678
AniFaceGAN-SR [58]30.50.0242.011.10.040.73
3DFaceShop [51]24.80.0180.4564.20.0290.594
Ours5.70.00163.171.90.350.80
", + "bbox": [ + 107, + 87, + 540, + 189 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4aa5862c35481321bb14bb26e9435b13f604dff8b3e6a67d861a1e861fd949d2.jpg", + "table_caption": [], + "table_footnote": [ + "(b) Ablation study." + ], + "table_body": "
ModelOurs wo LencOurs wo LpriorOurs
FID ↓6.06.15.7
KID ↓0.001720.001740.0016
ASD (cm) ↓0.1240.1350.124
var(α) ↓0.0005250.0005770.000515
AED (cm) ↓0.12330.08130.0797
var(β, θ) ↓0.0047070.004060.00404
", + "bbox": [ + 568, + 83, + 903, + 186 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. (a) Our method outperforms the prior 2D and 3D controllable image synthesis methods in both image quality and control disentanglement. (b) Our method demonstrates the best control accuracy over shape and expression with geometric prior $L_{prior}$ and self-supervised reconstruction loss $L_{enc}$ .", + "bbox": [ + 75, + 212, + 893, + 255 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "$E$ , using a ResNet50 [19] backbone followed with a single-layer MLP.", + "bbox": [ + 76, + 267, + 468, + 297 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1. Qualitative Comparisons", + "text_level": 1, + "bbox": [ + 76, + 323, + 305, + 339 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Controlled Portrait Synthesis. Our framework enables high-fidelity head synthesis with disentangled control over camera pose, shape and expression, as shown in Figure 1. We compare our method with prior controllable portrait synthesis works include DiscoFaceGAN [10], GAN-Control [47], HeadNeRF [21], AntiFaceGAN [58] and 3DFaceShop [51]. The first two methods integrates 3DMM controls into 2D face synthesis. HeadNeRF [21] is a 3D parametric NeRF-based head model built from both indoor multiview image capture datasets and in-the-wild monocular image collections. AntiFaceGAN [58] and 3DFaceShop [51] are concurrent 3D-aware controllable GANs in face synthesis.", + "bbox": [ + 75, + 351, + 468, + 547 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We qualitatively compare with prior work in Figure 5 for expression control. GAN-Control shows high-quality expression manipulation but with noticeable identity variation with differences in hair, head shape contour and background, whereas DiscoFaceGAN better maintains identity but with lower perceptual quality, and lacks dynamic details as expression varies. The expression control space of 3DFaceShop [51] is sensitive, suffering from appearance changes even with minor expression variation. AniFaceGAN demonstrates visually-consistent expression editing but with limited resolution and image quality, e.g., with blurry artifacts in hair and teeth. HeadNeRF ensures decent consistency in image generation by rendering the conditional NeRF but lacks fine details with very limited perceptual quality. In contrast, our approach produces the most compelling images with consistent appearance and dynamic details under expression changes. Moreover, controllable neck pose is a key factor towards realistic video avatar in applications like talking head synthesis. As shown in Figure. 1, our method achieves explicit control of neck and jaw poses which are seldom explored in prior work. Please refer to supplementary material for qualitative comparison in view consistency and shape editing.", + "bbox": [ + 75, + 553, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Dynamic details. Our method presents highly-consistent image synthesis in control of shape, expression and camera poses, but also depicts temporal realism in portrait animation, credited to the modeling of expression-dependent dynamic details. As shown in the last row of Figure 3, the appearance of wrinkles around the mouth and eyes when transiting from a neural expression to smiling largely enhances the animation realism. In comparison, without explicitly modeling the dynamic details, the wrinkles are embedded in the appearance and do not vary with expressions.", + "bbox": [ + 496, + 267, + 890, + 419 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Expressive Head Synthesis with Extrapolated Controls. AntiFaceGAN [58] and 3DFaceShop [51] depict dynamic details as well since their generation of the neural fields is conditioned on the input expression latent code. However, their designs result in shape and expression entanglement with the appearance generation, as reflected in the identity changes as shown in Figure. 5. By embedding the controls in a Gaussian latent space with such as Variation AutoEncoder (VAE), their approaches sacrifice expressiveness. The synthesized image quality is also highly correlated to the distribution of training images with target expressions.", + "bbox": [ + 496, + 426, + 890, + 592 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In contrast, our tri-plane generation is explicitly disentangled from shape and expression control. Moreover, the volume deformation is independently learnt from the deformed FLAME mesh collections which offer abundant 3D geometric knowledge with largely augmented control space. Therefore we are much less dependent on the distribution of the training images and support better extrapolation to unseen novel expressions. In Figure. 1, we show high-quality synthesized head with extreme jaw and neck articulated movements which do not exist in our training images. Our expression control is also more expressive, supporting subtle expressions like eye blinks (Figure. 1 3).", + "bbox": [ + 496, + 594, + 890, + 776 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Quantitative Comparisons", + "text_level": 1, + "bbox": [ + 500, + 786, + 738, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Image Quality We measure the image quality with Frechet Inception Distance (FID) [20] and Kernel Inception Distance [5] between 50K randomly synthesized images and 50K randomly sampled real images at the resolution of $512 \\times 512$ . Since DiscoFaceGAN [10] and AniFaceGAN [58] only synthesize images at $256 \\times 256$ reso", + "bbox": [ + 496, + 810, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "12820", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "lution, we utilize a state-of-the-art super-resolution model, SwinIR [30] to upsample into $512 \\times 512$ for a fair comparison. As shown in Table. 1a, our method is superior in both FID and KID than all prior work, demonstrating the most compelling image quality. We note that the original EG3D has a slightly lower FID at 4.8 which is expected since we introduce controllability with additional loss regularization.", + "bbox": [ + 75, + 90, + 472, + 196 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Disentanglement To evaluate the disentangled controllability of our model over shape, expression and camera pose, we measure the disentanglement score [10] of synthesized images as $\\mathrm{DS}_{\\alpha}$ , $\\mathrm{DS}_{\\beta,\\theta}$ and $\\mathrm{DS}_{\\mathbf{c}}$ respectively. DS measures the stability of other factors when a single attribute is modified in image synthesis. We employ DECA [12] for estimation of FLAME parameters from generated images and calculate the variance of the estimated parameters $(\\alpha, \\{\\beta, \\theta\\}, \\mathbf{c})$ . Specifically the $\\mathrm{DS}_i$ is calculated as", + "bbox": [ + 75, + 207, + 472, + 344 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\nD S _ {i} = \\prod_ {j \\neq i} \\frac {\\operatorname {v a r} (i)}{\\operatorname {v a r} (j)}, i, j \\in \\{\\boldsymbol {\\alpha}, \\{\\boldsymbol {\\beta}, \\boldsymbol {\\theta} \\}, \\mathbf {c} \\}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 352, + 468, + 390 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Higher value of DS indicates better disentanglement. Additionally, we evaluate the identity similarity between pairs of synthesized images with random camera poses and expressions by calculating the cosine similarity of the face embeddings with a pre-trained face recognition module [26]. Our approach demonstrates the best disentanglement numerically over all prior work as indicated in Table. 1a.", + "bbox": [ + 75, + 398, + 472, + 505 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 76, + 513, + 243, + 527 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We ablate the efficacy of the individual component by removing it from our full pipeline. As shown in Figure. 3, we show the loss of control accuracy over shape and expression respectively when removing the geometric prior loss $L_{prior}$ and control loss $L_{enc}$ . Conditioning the neural radiance field on expression is also critical to the modeling of dynamic details. Numerically we validate the efficacy of $L_{prior}$ and $L_{enc}$ with Average Shape Distance (ASD) and Average Expression Distance (AED). From the prior distribution, we randomly sample 500 shapes and expressions, with which we synthesize images with 10 different identities. We then reconstruct FLAME parameters from the synthesized images and compare against the input control. We compute 3D per-vertex $L_{1}$ distance of FLAME meshes for ASD while we use 3D landmarks $L_{1}$ distance for AED calculation. Additionally we compute the variance of estimated shapes and expressions within each control group, with lower value indicating more precise control. The efficacy of $L_{prior}$ and $L_{enc}$ is well evidenced in Table. 1b, with even slight image quality improvements.", + "bbox": [ + 75, + 536, + 472, + 840 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Applications", + "text_level": 1, + "bbox": [ + 76, + 847, + 210, + 863 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Talking Head Video Generation. In Figure. 1, we showcase talking head video generation in controllable views", + "bbox": [ + 75, + 869, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "driven with animation sequences of FLAME. Thanks to the high control accuracy, our method is able to synthesize various talking head videos with the same head movements and expressions performed by different identities. Our method is expressive in depicting both large articulated neck and jaw movements and subtle facial expressions like eye blinks, with rich dynamic details. Shape manipulation is easily achievable as well by modifying the shape parameters. Please refer to our supplementary materials for more results in high resolution.", + "bbox": [ + 496, + 90, + 893, + 243 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Portrait Image Manipulation and Animation. As illustrated in Figure. 1, our model also supports 3D-aware face reenactment of a single-view portrait to a video sequence. To achieve that, we perform an optimization in the latent $Z+$ space [24] to find the corresponding latent embedding, with FLAME parameter and camera pose estimated from the input portrait. With a frozen generator, the optimization is performed by measuring the similarity between generated image and real image using the $L_{2}$ loss and LPIPS loss [65]. For better reconstruction quality, we alter the parameters of the tri-plane synthesis module with a fixed optimized latent code [43]. After that one can explicitly manipulate the portrait with a preserved identity and in a different camera pose and expression. With the expression codes $(\\beta, \\theta)$ reconstructed from a video sequence, we are also able to reenact the portrait to the video motion.", + "bbox": [ + 496, + 250, + 893, + 492 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Societal Impact. Our work focuses on improving the controllability of 3D-aware GANs in technical aspects and is not specifically designed for any malicious uses. This being said, we do see that the method could be potentially extended into controversial applications such as generating fake videos. Therefore, we believe that the synthesized images and videos should present themselves as synthetic.", + "bbox": [ + 496, + 500, + 893, + 608 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 618, + 619, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we introduce OmniAvatar, a novel 3D-aware generative model for synthesis of controllable high-fidelity human head. Our model achieves disentangled semantic control by factoring the generative process into 3D-aware canonical head synthesis and implicit volume deformation to target shapes and expressions. By learning a deformable signed distance function with canonical volumetric correspondence, we instill the geometric prior knowledge of a 3D statistical head model into the generation of neural radiance fields, enabling superior generative capability to novel shapes and expressions. Our approach demonstrates expressive and compelling talking head generation and portrait image animation with fine-grained control accuracy and temporal realism. We believe the proposed method presents an interesting direction for 3D avatar creation and animation, which sheds light on many potential downstream tasks.", + "bbox": [ + 496, + 643, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "12821", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas Guibas. Learning representations and generative models for 3d point clouds. In ICML, 2018. 2", + "[2] Thiemo Alldieck, Hongyi Xu, and Cristian Sminchisescu. imghum: Implicit generative models of 3d human shape and articulated pose. In ICCV, pages 5461-5470, 2021. 4", + "[3] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In iccv, pages 5855–5864, 2021. 3", + "[4] Alexander W Bergman, Petr Kellnhofer, Yifan Wang, Eric R Chan, David B Lindell, and Gordon Wetzstein. Generative neural articulated radiance fields. arXiv preprint arXiv:2206.14314, 2022. 3, 5", + "[5] Sai Bi, Kalyan Sunkavalli, Federico Perazzi, Eli Shechtman, Vladimir G Kim, and Ravi Ramamoorthi. Deep cg2real: Synthetic-to-real translation via image disentanglement. In iccv, pages 2730-2739, 2019. 7", + "[6] Volker Blanz and Thomas Vetter. A morphable model for the synthesis of 3d faces. In Proceedings of the 26th annual conference on Computer graphics and interactive techniques, 1999. 2", + "[7] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. CVPR, 2022. 2, 3, 4", + "[8] Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5799-5809, 2021. 2", + "[9] Anpei Chen, Ruiyang Liu, Ling Xie, Zhang Chen, Hao Su, and Jingyi Yu. Sofgan: A portrait image generator with dynamic styling. sigg, 41(1):1-26, 2022. 2", + "[10] Yu Deng, Jiaolong Yang, Dong Chen, Fang Wen, and Xin Tong. Disentangled and controllable face image generation via 3d imitative-contrastive learning. In cvpr, pages 5154–5163, 2020. 2, 6, 7, 8", + "[11] Yu Deng, Jiaolong Yang, Jianfeng Xiang, and Xin Tong. Gram: Generative radiance manifolds for 3d-aware image generation. CVPR, 2022. 2", + "[12] Yao Feng, Haiwen Feng, Michael J. Black, and Timo Bolkart. Learning an animatable detailed 3D face model from in-the-wild images. volume 40, 2021. 8", + "[13] Guy Gafni, Justus Thies, Michael Zollhofer, and Matthias Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8649-8658, 2021. 3", + "[14] Partha Ghosh, Pravir Singh Gupta, Roy Uziel, Anurag Ranjan, Michael J Black, and Timo Bolkart. *Gif: Generative interpretable faces*. In 3dv, pages 868-878. IEEE, 2020. 2, 3", + "[15] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and" + ], + "bbox": [ + 78, + 114, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Yoshua Bengio. Generative adversarial nets. In NeurIPS, 2014. 1, 2", + "[16] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. In ICML, pages 3569-3579. 2020. 4", + "[17] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. Stylenerf: A style-based 3d-aware generator for high-resolution image synthesis. CVPR, 2022. 2", + "[18] Yudong Guo, Keyu Chen, Sen Liang, Yong-Jin Liu, Hujun Bao, and Juyong Zhang. Ad-nerf: Audio driven neural radiance fields for talking head synthesis. In iccv, pages 5784-5794, 2021. 3", + "[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In cvpr, pages 770-778, 2016. 7", + "[20] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. NeurIPS, 2017. 7", + "[21] Yang Hong, Bo Peng, Haiyao Xiao, Ligang Liu, and Juyong Zhang. Headnerf: A real-time nerf-based parametric head model. In cvpr, pages 20374-20384, 2022. 3, 6, 7", + "[22] Kacper Kania, Kwang Moo Yi, Marek Kowalski, Tomasz Trzciński, and Andrea Tagliasacchi. Conerf: Controllable neural radiance fields. In cvpr, pages 18623-18632, 2022. 3", + "[23] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. NeurIPS, 2021. 1", + "[24] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, 2019. 1, 2, 6, 8", + "[25] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. In CVPR, 2020. 1, 2", + "[26] Minchul Kim, Anil K Jain, and Xiaoming Liu. Adaface: Quality adaptive margin for face recognition. In cvpr, pages 18750-18759, 2022. 8", + "[27] Marek Kowalski, Stephan J Garbin, Virginia Estellers, Tadas Baltrusaitis, Matthew Johnson, and Jamie Shotton. Config: Controllable neural face image generation. In $eccv$ , pages 299-315. Springer, 2020. 2, 3", + "[28] Ruihui Li, Xianzhi Li, Chi-Wing Fu, Daniel Cohen-Or, and Pheng-Ann Heng. Pu-gan: a point cloud upsampling adversarial network. In ICCV, 2019. 2", + "[29] Tianye Li, Timo Bolkart, Michael. J. Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4D scans. ACM Transactions on Graphics, 36(6):194:1-194:17, 2017. 2, 3", + "[30] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In iccv, pages 1833-1844, 2021. 8", + "[31] Yiyi Liao, Katja Schwarz, Lars Mescheder, and Andreas Geiger. Towards unsupervised learning of generative models for 3D controllable image synthesis. In CVPR, 2020. 2", + "[32] Lingjie Liu, Marc Habermann, Viktor Rudnev, Kripasindhu Sarkar, Jiatao Gu, and Christian Theobalt. Neural actor: Neural free-view synthesis of human actors with pose control. ACM Trans. on Graphics, 2021. 3" + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "12822", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[33] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 2, 3", + "[34] Thu Nguyen-Phuoc, Chuan Li, Lucas Theis, Christian Richardt, and Yong-Liang Yang. HoloGAN: Unsupervised learning of 3D representations from natural images. In ICCV, 2019. 2", + "[35] Thu Nguyen-Phuoc, Christian Richardt, Long Mai, Yong-Liang Yang, and Niloy Mitra. BlockGAN: Learning 3D object-aware scene representations from unlabelled images. In NeurIPS, 2020. 2", + "[36] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In CVPR, 2021. 2", + "[37] Roy Or-El, Xuan Luo, Mengyi Shan, Eli Shechtman, Jeong Joon Park, and Ira Kemelmacher-Shlizerman. Stylesdf: High-resolution 3d-consistent image and geometry generation. CVPR, 2022. 2, 5", + "[38] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In ICCV, 2021. 3", + "[39] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. ACM Transactions on Graphics, 2022. 3", + "[40] Pascal Paysan, Reinhard Knothe, Brian Amberg, Sami Romdhani, and Thomas Vetter. A 3d face model for pose and illumination invariant face recognition. In 2009 sixth IEEE international conference on advanced video and signal based surveillance, pages 296-301. IEEE, 2009. 2", + "[41] Jingtan Piao, Keqiang Sun, Quan Wang, Kwan-Yee Lin, and Hongsheng Li. Inverting generative adversarial renderer for face reconstruction. In cvpr, pages 15619-15628, 2021. 2, 3", + "[42] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In CVPR, 2021. 3", + "[43] Daniel Roich, Ron Mokady, Amit H Bermano, and Daniel Cohen-Or. Pivotal tuning for latent-based editing of real images. ACM Transactions on Graphics (TOG), 42(1):1-13, 2022. 8", + "[44] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. Graf: Generative radiance fields for 3d-aware image synthesis. NeurIPS, 2020. 2", + "[45] Yichun Shi, Divyansh Aggarwal, and Anil K Jain. Lifting 2d stylegan for 3d-aware face generation. In cvpr, pages 6258-6266, 2021. 2", + "[46] Yichun Shi, Xiao Yang, Yangyue Wan, and Xiaohui Shen. Semanticstylegan: Learning compositional generative priors for controllable image synthesis and editing. In CVPR, 2022. 2", + "[47] Alon Shoshan, Nadav Bhonker, Igor Kviatkovsky, and Gerard Medioni. Gan-control: Explicitly controllable gans. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14083-14093, 2021. 6, 7" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[48] Ivan Skorokhodov, Sergey Tulyakov, Yiqun Wang, and Peter Wonka. Epigraph: Rethinking training of 3d gans. arXiv preprint arXiv:2206.10535, 2022. 2", + "[49] Keqiang Sun, Shangzhe Wu, Zhaoyang Huang, Ning Zhang, Quan Wang, and HongSheng Li. Controllable 3d face synthesis with conditional generative occupancy fields. arXiv preprint arXiv:2206.08361, 2022. 3", + "[50] Attila Szabó, Givi Meishvili, and Paolo Favaro. Unsupervised generative 3D shape learning from natural images. arXiv, 2019. 2", + "[51] Junshu Tang, Bo Zhang, Binxin Yang, Ting Zhang, Dong Chen, Lizhuang Ma, and Fang Wen. Explicitly controllable 3d-aware portrait generation. arXiv preprint arXiv:2209.05434, 2022. 3, 6, 7", + "[52] Ayush Tewari, Mohamed Elgharib, Florian Bernard, Hans-Peter Seidel, Patrick Pérez, Michael Zollhöfer, and Christian Theobalt. Pie: Portrait image embedding for semantic control. sigg, 39(6):1-14, 2020. 2", + "[53] Ayush Tewari, Mohamed Elgharib, Gaurav Bharaj, Florian Bernard, Hans-Peter Seidel, Patrick Pérez, Michael Zollhofer, and Christian Theobalt. Stylerig: Rigging stylegan for 3d control over portrait images. In cvpr, pages 6142-6151, 2020. 2, 3", + "[54] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for stylegan image manipulation. ACM Trans. on Graphics, 2021. 2", + "[55] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Christoph Lassner, and Christian Theobalt. Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In ICCV, 2021. 3", + "[56] Ziyan Wang, Timur Bagautdinov, Stephen Lombardi, Tomas Simon, Jason Saragih, Jessica Hodgins, and Michael Zollhofer. Learning compositional radiance fields of dynamic human heads. In cvpr, pages 5704-5713, 2021. 3", + "[57] Jiajun Wu, Chengkai Zhang, Tianfan Xue, William T. Freeman, and Joshua B. Tenenbaum. Learning a probabilistic latent space of object shapes via 3D generative-adversarial modeling. In NeurIPS, 2016. 2", + "[58] Yue Wu, Yu Deng, Jiaolong Yang, Fangyun Wei, Qifeng Chen, and Xin Tong. Anifacegan: Animatable 3d-aware face image generation for video avatars. arXiv preprint arXiv:2210.06465, 2022. 3, 6, 7", + "[59] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond. In Computer Graphics Forum, volume 41, pages 641-676. Wiley Online Library, 2022. 3", + "[60] Hongyi Xu, Thiemo Alldieck, and Cristian Sminchisescu. H-nerf: Neural radiance fields for rendering and temporal reconstruction of humans in motion. NeurIPS, 2021. 3", + "[61] Xudong Xu, Xingang Pan, Dahua Lin, and Bo Dai. Generative occupancy fields for 3d surface-aware image synthesis. NeurIPS, 34:20683-20695, 2021. 2", + "[62] Yang Xue, Yuheng Li, Krishna Kumar Singh, and Yong Jae Lee. Giraffe hd: A high-resolution 3d-aware generative model. In CVPR, 2022. 2" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "12823", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[63] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. In NeurIPS, 2021. 5", + "[64] Jianfeng Zhang, Zihang Jiang, Dingdong Yang, Hongyi Xu, Yichun Shi, Guoxian Song, Zhongcong Xu, Xinchao Wang, and Jiashi Feng. Avatargen: a 3d generative model for animatable human avatars. arXiv preprint arXiv:2208.00561, 2022.3", + "[65] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In cvpr, pages 586-595, 2018. 8", + "[66] Yufeng Zheng, Victoria Fernández Abrevaya, Marcel C Bühler, Xu Chen, Michael J Black, and Otmar Hilliges. Im avatar: Implicit morphable head avatars from videos. In cvpr, pages 13545-13555, 2022. 3", + "[67] Peng Zhou, Lingxi Xie, Bingbing Ni, and Qi Tian. CIPS-3D: A 3D-Aware Generator of GANs Based on Conditionally-Independent Pixel Synthesis. arXiv, 2021. 2", + "[68] Yiyu Zhuang, Hao Zhu, Xusen Sun, and Xun Cao. Mofan-erf: Morphable facial neural radiance field. arXiv preprint arXiv:2112.02308, 2021. 3" + ], + "bbox": [ + 78, + 90, + 470, + 402 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "12824", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/OmniAvatar_ Geometry-Guided Controllable 3D Head Synthesis/13b82072-fc3f-4f4d-8e9e-44b54de26be7_model.json b/2023/OmniAvatar_ Geometry-Guided Controllable 3D Head Synthesis/13b82072-fc3f-4f4d-8e9e-44b54de26be7_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7015414569ba6f7418bbf2264bd1ad5889dcb21e --- /dev/null +++ b/2023/OmniAvatar_ Geometry-Guided Controllable 3D Head Synthesis/13b82072-fc3f-4f4d-8e9e-44b54de26be7_model.json @@ -0,0 +1,2103 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.158, + 0.131, + 0.815, + 0.154 + ], + "angle": 0, + "content": "OmniAvatar: Geometry-Guided Controllable 3D Head Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.149, + 0.179, + 0.82, + 0.271 + ], + "angle": 0, + "content": "Hongyi Xu1 Guoxian Song1 Zihang Jiang1,2 Jianfeng Zhang1,2 Yichun Shi1 Jing Liu1 Wanchun Ma1 Jiashi Feng1 Linjie Luo1 \n1ByteDance Inc 2National University of Singapore \n{hongyixu, guoxian.song, zihang.jiang, jianfeng.zhang, yichun.shi, jing.liu, wanchun.ma, jshfeng, linjie.luo}@bytedance.com" + }, + { + "type": "image", + "bbox": [ + 0.079, + 0.304, + 0.896, + 0.533 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.54, + 0.895, + 0.584 + ], + "angle": 0, + "content": "Figure 1. Our model can synthesize diverse identity-preserved 3D heads with compelling dynamic details under full disentangled control over camera poses, facial expressions, head shapes, articulated neck and jaw poses (left). Our model can also reconstruct 3D heads from a single photo reference and enable multi-view-consistent head reenactment (right)." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.595, + 0.314, + 0.611 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.629, + 0.474, + 0.903 + ], + "angle": 0, + "content": "We present OmniAvatar, a novel geometry-guided 3D head synthesis model trained from in-the-wild unstructured images that is capable of synthesizing diverse identity-preserved 3D heads with compelling dynamic details under full disentangled control over camera poses, facial expressions, head shapes, articulated neck and jaw poses. To achieve such high level of disentangled control, we first explicitly define a novel semantic signed distance function (SDF) around a head geometry (FLAME) conditioned on the control parameters. This semantic SDF allows us to build a differentiable volumetric correspondence map from the observation space to a disentangled canonical space from all the control parameters. We then leverage the 3D-aware GAN framework (EG3D) to synthesize detailed shape and appearance of 3D full heads in the canonical space, followed by a volume rendering step guided by the volumetric correspondence map to output into the observation space. To ensure the control accuracy on the synthesized head" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.597, + 0.895, + 0.734 + ], + "angle": 0, + "content": "shapes and expressions, we introduce a geometry prior loss to conform to head SDF and a control loss to conform to the expression code. Further, we enhance the temporal realism with dynamic details conditioned upon varying expressions and joint poses. Our model can synthesize more preferable identity-preserved 3D heads with compelling dynamic details compared to the state-of-the-art methods both qualitatively and quantitatively. We also provide an ablation study to justify many of our system design choices." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.768, + 0.633, + 0.784 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Photo-realistic face image synthesis, editing and animation attract significant interests in computer vision and graphics, with a wide range of important downstream applications in visual effects, digital avatars, telepresence and many others. With the advent of Generative Adversarial Networks (GANs) [15], remarkable progress has been achieved in face image synthesis by StyleGAN [23-25]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12814" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.243 + ], + "angle": 0, + "content": "as well as in semantic and style editing for face images [46, 54]. To manipulate and animate the expressions and poses in face images, many methods attempted to leverage 3D parametric face models, such as 3D Morphable Models (3DMMs) [6, 40], with StyleGAN-based synthesis models [10, 41, 53]. However, all these methods operate on 2D convolutional networks (CNNs) without explicitly enforcing the underlying 3D face structure. Therefore they cannot strictly maintain the 3D consistency when synthesizing faces under different poses and expressions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.246, + 0.473, + 0.457 + ], + "angle": 0, + "content": "Recently, a line of work has explored neural 3D representations by unsupervised training of 3D-aware GANs from in-the-wild unstructured images [7, 8, 11, 17, 37, 44, 45, 48, 61, 62, 67]. Among them, methods with generative Neural Radiance Fields (NeRFs) [33] have demonstrated striking quality and multi-view-consistent image synthesis [7, 11, 17, 37, 48]. The progress is largely due to the integration of the power of StyleGAN in photo-realistic image synthesis and NeRF representation in 3D scene modeling with view-consistent volumetric rendering. Nevertheless, these methods lack precise 3D control over the generated faces beyond camera pose, as well as the quality and consistency in control over other attributes, such as shape, expression, neck and jaw pose, leave much to be desired." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.46, + 0.473, + 0.779 + ], + "angle": 0, + "content": "In this work, we present OmniAvatar, a novel geometry-guided 3D head synthesis model trained from in-the-wild unstructured images. Our model can synthesize a wide range of 3D human heads with full control over camera poses, facial expressions, head shapes, articulated neck and jaw poses. To achieve such high level of disentangled control for 3D human head synthesis, we devise our model learning in two stages. We first define a novel semantic signed distance function (SDF) around a head geometry (i.e. FLAME [29]) conditioned on its control parameters. This semantic SDF fully distills rich 3D geometric prior knowledge from the statistical FLAME model and allows us to build a differentiable volumetric correspondence map from the observation space to a disentangled canonical space from all the control parameters. In the second training stage, we then leverage the state-of-the-art 3D GAN framework (EG3D [7]) to synthesize realistic shape and appearance of 3D human heads in the canonical space, including the modeling of hair and apparels. Following that, a volume rendering step is guided by the volumetric correspondence map to output the geometry and image in the observation space." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.473, + 0.903 + ], + "angle": 0, + "content": "To ensure the consistency of synthesized 3D head shape with controlling head geometry, we introduce a geometry prior loss to minimize the difference between the synthesized neural density field and the FLAME head SDF in observation space. Furthermore, to improve the control accuracy, we pre-train an image encoder of the control parameters and formulate a control loss to ensure synthesized images matching the input control code upon encoding. An-" + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.092, + 0.895, + 0.168 + ], + "angle": 0, + "content": "other key aspect of synthesis realism is dynamic details such as wrinkles and varying shading as subjects change expressions and poses. To synthesize dynamic details, we propose to condition EG3D's triplane feature decoding with noised controlling expression." + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.168, + 0.895, + 0.273 + ], + "angle": 0, + "content": "Compare to state-of-the-art methods, our method achieves superior synthesized image quality in terms of Frechet Inception Distance (FID) and Kernel Inception Distance (KID). Our method can consistently preserve the identity of synthesized subjects with compelling dynamic details while changing expressions and poses, outperforming prior methods both quantitatively and qualitatively." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.273, + 0.874, + 0.288 + ], + "angle": 0, + "content": "The contributions of our work can be summarized as:" + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.29, + 0.892, + 0.35 + ], + "angle": 0, + "content": "- A novel geometry-guided 3D GAN framework for high-quality 3D head synthesis with full control on camera poses, facial expressions, head shapes, articulated neck and jaw poses." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.352, + 0.892, + 0.413 + ], + "angle": 0, + "content": "- A novel semantic SDF formulation that defines the volumetric correspondence map from observation space to canonical space and allows full disentanglement of control parameters in 3D GAN training." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.415, + 0.891, + 0.446 + ], + "angle": 0, + "content": "- A geometric prior loss and a control loss to ensure the head shape and expression synthesis accuracy." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.448, + 0.891, + 0.478 + ], + "angle": 0, + "content": "- A robust noised expression conditioning scheme to enable dynamic detail synthesis." + }, + { + "type": "list", + "bbox": [ + 0.519, + 0.29, + 0.892, + 0.478 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.491, + 0.642, + 0.507 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.516, + 0.895, + 0.789 + ], + "angle": 0, + "content": "3D-Aware Generative Image Synthesis. Generative adversarial networks [15, 24, 25] gained popularity over the last decade due to their remarkable ability in photo-realistic image synthesis. Building on the success of 2D image-based GANs, recent works have extended the capabilities to view-consistent image synthesis with unsupervised learning from 2D single-view images. The key idea is to combine differential rendering with 3D scene representations, such as meshes [31, 50], point clouds [1, 28], voxels [34, 35, 57], and recently implicit neural representation [7, 8, 11, 17, 36, 37, 44, 48, 67]. We build our work on recent 3D GAN model by Chan et al [7] that uses an efficient triplane-based NeRF generation, combined with 2D CNN-based super-resolution. Even though 3D-aware GANs are able to control camera viewpoints, they lack precise 3D control over the other attributes such as shapes and expressions. In this work, we empower 3D-aware GANs with disentangled precise control over shapes and expressions." + }, + { + "type": "text", + "bbox": [ + 0.497, + 0.795, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Controllable Face Image Synthesis. Considerable work [9, 10, 14, 27, 41, 52, 53] has been devoted to incorporate 3D priors of statistical face models, such as 3D Morphable Models (3DMMs) [6, 40], in controllable face synthesis and animation. Among them, DiscoFaceGAN [10] proposed imitative-contrastive learning to mimic the 3DMM rendering process by the generative model." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12815" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.895, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.354, + 0.896, + 0.482 + ], + "angle": 0, + "content": "Figure 2. Overview of our training framework. Stage 1: Trained from parameterized FLAME [29] mesh collections, a MLP-network \\( W \\) maps a shape \\( \\alpha \\), expression \\( \\theta \\) and articulated jaw and neck pose \\( \\theta \\) into 3D point-to-point volumetric correspondences from observation to canonical space, together with a signed distance function of the corresponding FLAME head. Stage 2: Given a Gaussian latent code \\( \\mathbf{z} \\), our model generates a tri-plane represented 3D feature space of a canonical head, disentangled with shape and expression controls. The volume rendering is then guided by the volumetric correspondence field to map the decoded neural radiance field from the canonical to observation space. We condition the NeRF decoding with expression and joint pose for modeling dynamic details. A super-resolution module synthesizes the final high-resolution RGB image from the volume-rendered feature map. For fine-grained shape and expression control, we apply the FLAME SDF as geometric prior to the synthesized NeRF density, and self-supervise the image synthesis to commit to the target expression \\( \\beta \\) and joint pose \\( \\theta \\) by comparing the input code against the re-estimated values \\( \\hat{\\beta}, \\hat{\\theta} \\) from synthesized images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.497, + 0.473, + 0.74 + ], + "angle": 0, + "content": "A similar strategy has also been adopted with concurrent and follow-up works [14, 27, 41, 53]. However all of these approaches suffer from 3D inconsistency due to the use of 2D CNNs as image renderer. HeadNeRF [21] combines 3DMM with 3D NeRF representation, and is able to synthesize 3D heads conditioned on 3DMM attributes. However, the training relies on annotated multiview datasets whereas our approach learns the disentangled 3D head synthesis with only single-view images. There has been concurrent work [4, 49, 51, 58, 64] to ours in 3D-aware controllable face or full-body GANs. Differently from these approaches, we use a full-head parametric model FLAME [29], and fully exploit the spatial geometric prior knowledge beyond the surface deformation and skinning. We have also achieved fine-grained control with our novel losses and enhanced face animation with rich dynamic details." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Controllable Neural Implicit Field of Face. Neural implicit functions [59], have emerged as a powerful continuous and differential representations of 3D scenes. Among them, Neural Radiance Field [3, 33] has been widely adopted due to its superiority in modeling complex scene details and synthesizing multiview images with inherited 3D consistency. While initial proposals have focused on static scene modeling, recent work has successfully demonstrated application of NeRF in modeling dynamic scenes [32,38,42,55,60]. In particular, dynamic neural radi-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.497, + 0.895, + 0.62 + ], + "angle": 0, + "content": "ance fields of human heads [13,18,22,38,39,56,66,68] have enabled photo-realistic head animation, often by conditioning with pose parameters or deforming the radiance field with 3D morphable models. However, they do not leverage a generative training paradigm and focus on scene-specific learning from video sequences or multiview images. In contrast, our model learns generative and controllable neural radiance fields from widely accessible single-view images." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.632, + 0.593, + 0.648 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.658, + 0.894, + 0.718 + ], + "angle": 0, + "content": "Our goal is to build a geometry-guided 3D head synthesis model with full control of camera poses, face shapes and expressions, trained from in-the-wild unstructured image collections." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.729, + 0.611, + 0.743 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.752, + 0.895, + 0.905 + ], + "angle": 0, + "content": "Problem. To achieve our goal, we leverage a 3D-aware generator (EG3D [7]) for photo-realistic, multiview consistent image synthesis, while disentangle control of head geometric attributes from image generation with a 3D statistical head model (FLAME [29]). Specifically, given a random Gaussian-sampled latent code \\(\\mathbf{z}\\), a camera pose \\(\\mathbf{c}\\) and a FLAME parameter \\(\\mathbf{p} = (\\alpha, \\beta, \\theta)\\) consisting of shape \\(\\alpha\\), expression \\(\\beta\\), jaw and neck pose \\(\\theta\\), the generator \\(G\\) synthesizes a photo-realistic human head image \\(I_{RGB}(\\mathbf{z}|\\mathbf{c},\\mathbf{p})\\) with corresponding attributes as defined in \\(\\mathbf{p}\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12816" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.35 + ], + "angle": 0, + "content": "**Framework.** As illustrated in our pipeline 2, our controllable 3D-aware GAN is trained in two stages. From a large collection of 3D deformed FLAME meshes, we first pre-train a deformable semantic SDF around the FLAME geometry that builds a differential volumetric correspondence map from the observation to a predefined canonical space (Section 3.2.1). In the second stage, guided by the pre-trained volumetric mapping, we then deform the detailed 3D full heads synthesized in the disentangled canonical space to the desired shapes and expressions (Section 3.2.2). Fine-grained expression control is achieved by supervising image synthesis such that expressions estimated from the generated images is consistent with the input control (Section 3.2.3). Our approach further enhances temporal realism with dynamic details, such as dimples and wrinkles, synthesizing realistic shading and geometric variations as expression changes (Section 3.2.4)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.358, + 0.473, + 0.555 + ], + "angle": 0, + "content": "3D-Aware GAN Background. To ensure appearance consistency from different views, we choose EG3D [7] as our backbone for 3D-aware image synthesis. The generator \\(G\\) takes a random latent code \\(\\mathbf{z}\\) and conditioning camera label \\(\\hat{\\mathbf{c}}\\), and maps to a manifold of triplane features \\(\\mathcal{M}(\\mathbf{z},\\hat{\\mathbf{c}})\\). For presentation clarity, we absorb \\(\\hat{\\mathbf{c}}\\) to \\(\\mathbf{z}\\) and simply denote triplane as \\(\\mathcal{M}(\\mathbf{z})\\). A low-resolution feature map \\(I^{*}(\\mathbf{z}|\\mathbf{c})\\) is then rendered from a desired camera pose \\(\\mathbf{c}\\) by sampling the triplane features and integrating MLP-decoded neural radiance \\((\\sigma ,\\mathbf{f})\\) along camera rays. A super-resolution module is followed to modulate the feature map and synthesize the final RGB images at high resolution. We train \\(G\\) and a dual discriminator \\(D\\) with adversarial training." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.563, + 0.421, + 0.58 + ], + "angle": 0, + "content": "3.2. Controllable 3D-Aware Image Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.587, + 0.471, + 0.814 + ], + "angle": 0, + "content": "To synthesize an image \\(\\mathbf{I}(\\mathbf{z}|\\mathbf{c},\\mathbf{p})\\) with desired FLAME control \\(\\mathbf{p}\\), we leverage the EG3D framework to generate a triplane-based 3D volumetric feature space \\(\\mathcal{M}(\\mathbf{z})\\) of a synthesized identity with canonical shape and expression. Guided by our pretrained volumetric correspondence map, we deform the synthesized feature volume into our target observation space, which is further decoded and volume rendered into high-fidelity head appearance and geometry with the target shape and expression. Our design explicitly disentangles the underlying geometric variations of changing shape and expression from canonical geometry and appearance synthesis. Following EG3D [7], we associate each training image with a set of camera parameters \\(\\mathbf{c}\\) and control parameters \\(\\mathbf{p}\\), which are obtained from a nonlinear 2D landmarks-based optimization." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.832, + 0.379, + 0.848 + ], + "angle": 0, + "content": "3.2.1 Semantic Signed Distance Function" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.47, + 0.902 + ], + "angle": 0, + "content": "For disentangled geometric modeling, we formulate an implicit semantic SDF representation \\( W(\\mathbf{x}|\\mathbf{p} = (\\alpha ,\\beta ,\\theta)) = (s,\\bar{\\mathbf{x}}) \\), where \\( \\alpha ,\\beta \\) are the linear shape and expression" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.089, + 0.912, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.361, + 0.892, + 0.445 + ], + "angle": 0, + "content": "Figure 3. We synthesize different identities with the same shape and expression code respectively in (a) and (b). We observe less shape variation with our geometric prior loss (a) and more consistent expression with our control loss (b). In (c), we show our expression-conditioned NeRF modeling builds rich dynamic details as the subject varies expressions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.454, + 0.892, + 0.621 + ], + "angle": 0, + "content": "blendshape coefficients, and \\(\\theta\\) controls the rotation of a 3-DoF jaw and neck joint. Specifically, given a spatial point \\(\\mathbf{x}\\) in observation space \\(\\mathcal{O}(\\mathbf{p})\\), \\(W\\) returns its 3D correspondence point \\(\\bar{\\mathbf{x}}\\) (i.e., semantics) in canonical space \\(\\mathcal{C}(\\bar{\\mathbf{p}})\\), with which we project and query the triplane features \\(\\mathcal{M}(\\mathbf{z})\\). Additionally it also computes the closest signed distance \\(s(\\mathbf{x}|\\mathbf{p})\\) to the FLAME mesh surface \\(\\mathbf{S}(\\mathbf{p})\\). We illustrate the function in Figure 4. We co-learn the highly-correlated volumetric correspondence and SDF with the property that the signed distance is preserved between canonical and observation correspondence points as \\(s(\\bar{\\mathbf{x}}|\\bar{\\mathbf{p}}) = s(\\mathbf{x}|\\mathbf{p})\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.621, + 0.892, + 0.681 + ], + "angle": 0, + "content": "We learn \\( W(\\mathbf{x}|\\mathbf{p}) \\) with a large corpus of 3D FLAME meshes \\( \\mathbf{S}(\\mathbf{p}) \\) sampled from its parametric control space. Similar to IGR [16] and imGHUM [2], we model our implicit field as an MLP and optimize \\( W(\\mathbf{x}|\\mathbf{p}) \\) with losses," + }, + { + "type": "equation", + "bbox": [ + 0.511, + 0.702, + 0.892, + 0.739 + ], + "angle": 0, + "content": "\\[\nL _ {i s o} = \\frac {1}{| N |} \\sum_ {\\mathbf {x} \\in N} (| s (\\mathbf {x} | \\mathbf {p}) | + | \\nabla s _ {\\mathbf {x}} (\\mathbf {x} | \\mathbf {p}) - \\mathbf {n} (\\mathbf {x} | \\mathbf {p}) |), \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.741, + 0.891, + 0.777 + ], + "angle": 0, + "content": "\\[\nL _ {e i k} = \\frac {1}{| F |} \\sum_ {\\mathbf {x} \\in F} \\| \\nabla s _ {\\mathbf {x}} (\\mathbf {x} | \\mathbf {p}) - 1 \\| _ {2}, \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.564, + 0.78, + 0.891, + 0.816 + ], + "angle": 0, + "content": "\\[\nL _ {s e m} = \\frac {1}{| N |} \\sum_ {\\mathbf {x} \\in N} \\left(| \\bar {\\mathbf {x}} (\\mathbf {x} | \\mathbf {p}) - \\bar {\\mathbf {x}} ^ {*} (\\mathbf {x} | \\bar {\\mathbf {p}}) |\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.825, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where \\(N,F\\) are a batch of on and off surface samples. For the surface samples, the \\(L_{iso}\\) encourages the signed distance values to be on the zero-level-set and the SDF gradient to be equal to the given surface normals \\(\\mathbf{n}\\). The Eikonal loss \\(L_{eik}\\) is derived from [16] where the SDF is differen" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12817" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.154, + 0.089, + 0.403, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.131, + 0.258, + 0.414, + 0.273 + ], + "angle": 0, + "content": "Figure 4. Illustration to semantic SDF learning." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.284, + 0.47, + 0.405 + ], + "angle": 0, + "content": "tiable everywhere with gradient norm 1. The semantic loss \\( L_{sem} \\) supervises the mapping of surface samples \\( \\mathbf{x} \\in F \\) to the ground-truth correspondence points \\( \\bar{\\mathbf{x}}^* \\) on the canonical FLAME surface, where \\( \\bar{\\mathbf{x}}^* \\) and \\( \\mathbf{x} \\) share the same barycentric coordinates. The \\( L_{eik} \\) provides a geometric regularization over the volumetric SDF, whereas the \\( L_{iso} \\) and \\( L_{sem} \\) act as the boundary condition to the SDF and volumetric correspondence field respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.406, + 0.471, + 0.724 + ], + "angle": 0, + "content": "To guide the canonical correspondence learning for off-the-surface points, one could try to minimize the signed distance difference between \\( s(\\bar{\\mathbf{x}}|\\bar{\\mathbf{p}}) \\) and \\( s(\\mathbf{x}|\\mathbf{p}) \\). However, we note that the volume correspondence between the observation and canonical space is still ill-regularized, considering infinite number of points exist with the same signed distance value. We therefore reformulate volumetric correspondence field as \\( W(\\mathbf{x}|\\mathbf{p}) = (\\bar{s}(\\bar{\\mathbf{x}}), \\bar{\\mathbf{x}}) \\), where the signed distance of an observation-space point is obtained by mapping it into its canonical correspondence \\( \\bar{\\mathbf{x}} \\) and querying the pre-computed canonical SDF \\( \\bar{s} \\). Thus we only learn a correspondence field with which we can deform the canonical SDF to different FLAME configurations, and then supervise the deformed SDFs with the FLAME surface boundary, normals \\( (L_{iso}) \\) and Eikonal regularization \\( (L_{eik}) \\). As such, even for off-the-surface samples, their canonical correspondences are well regularized in space with the geometric properties of signed distance functions via \\( L_{iso} \\) and \\( L_{eik} \\). In contrast to explicit mesh deformation [4], our implicit volumetric correspondence is more accurate, differentiable, smooth everywhere and semantically consistent with the properties of SDF." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.74, + 0.439, + 0.755 + ], + "angle": 0, + "content": "3.2.2 Canonical Generation with Geometric Prior" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.764, + 0.469, + 0.884 + ], + "angle": 0, + "content": "With our pretrained semantic SDF \\( W(\\mathbf{x}|\\mathbf{p}) \\) modeling the shape and expression variation, we leverage the triplane for 3d-aware generation of human heads with canonical shape and expression. In particular, to generate a neural radiance feature \\( (\\sigma, \\mathbf{f}) \\) for a point \\( \\mathbf{x} \\) in observation space \\( \\mathcal{O}(\\mathbf{p}) \\), we use our correspondence function \\( W \\) to back warp \\( \\mathbf{x} \\) into \\( \\bar{\\mathbf{x}} \\), with which we project and sample the canonical triplane features followed with a tiny MLP decoding." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.886, + 0.469, + 0.901 + ], + "angle": 0, + "content": "In spite of the control disentanglement, there is no ex" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.18 + ], + "angle": 0, + "content": "plicit loss that constrains the triplane-generated neural radiance density to conform to the shape and expression as defined in \\(\\mathbf{p}\\). Therefore, we guide the generation of neural radiance density field by minimizing its difference to the underlying FLAME head geometry represented with SDF, as" + }, + { + "type": "equation", + "bbox": [ + 0.513, + 0.186, + 0.892, + 0.224 + ], + "angle": 0, + "content": "\\[\nL _ {p r i o r} = \\frac {1}{| R |} \\sum_ {\\mathbf {x} \\in R} e ^ {- \\gamma | s (\\mathbf {x} | p) |} \\left| \\sigma \\left(\\mathbf {x} \\mid \\mathbf {z}, \\mathbf {p}\\right) - \\sigma^ {*} \\left(\\mathbf {x} \\mid \\mathbf {p}\\right) \\right|, \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.226, + 0.892, + 0.256 + ], + "angle": 0, + "content": "\\[\n\\sigma^ {*} (\\mathbf {x} | \\mathbf {p}) = \\frac {1}{\\kappa} \\cdot \\operatorname {S i g m o i d} \\left(\\frac {- s (\\mathbf {x} | \\mathbf {p})}{\\kappa}\\right) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.262, + 0.892, + 0.428 + ], + "angle": 0, + "content": "where \\( R \\) is the stratified ray samples for volume rendering and \\( \\kappa \\) is a learnable scalar controlling the density tightness around the SDF boundary. Following [37, 63], we convert SDF value \\( s(\\mathbf{x}|\\mathbf{p}) \\) to proxy 3D density \\( \\sigma^{*}(\\mathbf{x}|\\mathbf{p}) \\) assuming non-hollow surfaces. We decay the weights for our geometric prior loss \\( L_{prior} \\) as the point moving away from the SDF boundary, allowing higher degrees of freedom in generation of residual geometries, such as hair and glasses. The geometric prior loss effectively guides the 3D head geometry learning but should not be overpowered which otherwise might lead to loss of geometric details." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.445, + 0.788, + 0.46 + ], + "angle": 0, + "content": "3.2.3 Fine-Grained Expression Control" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.469, + 0.892, + 0.711 + ], + "angle": 0, + "content": "Our geometric prior loss \\(L_{prior}\\) provides local 3D pointwise guidance, and is able to well regularize the shape generation and achieve coarse-level expression control. However, for delicate expressions, such as eye blinks, \\(L_{prior}\\) provides little supervision as the geometric variation is subtle. Moreover, for regions with complex correspondences, such as around the lips, it is challenging to guide the formation of correct expressions globally, just with point-wise geometric losses. To improve the control granularity, we propose an image-level supervision loss that requires a synthesized image \\(I_{RGB}(\\mathbf{z}|\\mathbf{c},\\mathbf{p})\\) matching the target expression as defined in the input \\(\\mathbf{p}\\). Using our training images with estimated control labels \\(\\mathbf{p}\\), we first pretrain an image encoder \\(E(I_{RGB}) = (\\tilde{\\boldsymbol{\\beta}},\\tilde{\\boldsymbol{\\theta}})\\) that regresses the expression coefficients \\(\\tilde{\\boldsymbol{\\beta}}\\) and joint poses \\(\\tilde{\\boldsymbol{\\theta}}\\). During our 3D GAN training, we then apply our image-level control supervision as" + }, + { + "type": "equation", + "bbox": [ + 0.504, + 0.732, + 0.891, + 0.772 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {e n c} = | \\tilde {\\boldsymbol {\\beta}} - \\boldsymbol {\\beta} | + | \\tilde {\\boldsymbol {\\theta}} - \\boldsymbol {\\theta} | + | \\mathbf {S} (\\boldsymbol {\\alpha}, \\tilde {\\boldsymbol {\\beta}}, \\tilde {\\boldsymbol {\\theta}}) - \\mathbf {S} (\\boldsymbol {\\alpha}, \\boldsymbol {\\beta}, \\boldsymbol {\\theta}) | + \\\\ + \\left| \\mathbf {J} \\mathbf {S} (\\boldsymbol {\\alpha}, \\tilde {\\boldsymbol {\\beta}}, \\tilde {\\boldsymbol {\\theta}}) - \\mathbf {J} \\mathbf {S} (\\boldsymbol {\\alpha}, \\boldsymbol {\\beta}, \\boldsymbol {\\theta}) \\right|, \\tag {6} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.78, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where \\(\\mathbf{S},\\mathbf{J}\\) are the FLAME mesh vertices and 3D landmarks regressor respectively. While being straightforward for the first 2 terms, the last two terms in \\(L_{enc}\\) penalize deviation of 3D vertex coordinates and surface landmarks after mesh decoding. We note that we do not supervise shape \\(\\alpha\\) in \\(L_{enc}\\) since our geometric prior loss \\(L_{prior}\\) suffices in shape control already and also due to the ambiguity of shape scaling estimated from monocular images." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12818" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.089, + 0.355, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.384, + 0.088, + 0.47, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.489, + 0.089, + 0.621, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.623, + 0.089, + 0.754, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.756, + 0.089, + 0.888, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.413, + 0.895, + 0.457 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparisons in expression control. Under a similar expression control, our method achieves the best perceptual quality with better identity-preservation (compared to GAN-Control [47] and 3DFaceShop [51]), more realistic dynamic details (compared to DiscoFaceGAN [10]) and higher expression control accuracy (compared to AntiFaceGAN [58], 3DFaceShop [51], HeadNeRF [21])." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.482, + 0.319, + 0.499 + ], + "angle": 0, + "content": "3.2.4 Dynamic Details Modeling" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.518, + 0.473, + 0.76 + ], + "angle": 0, + "content": "To this end, we have achieved controllable static image synthesis. However, one should observe the variation of shading and geometric details in a dynamic head motion, such as the appearance of dimples and eye wrinkles in smiling. We consider the appearance of dynamic details is highly correlated with the driving expression \\((\\beta ,\\theta)\\). To model such temporal effects, one could try to condition the triplane generation on expression as \\(\\mathcal{M}(\\mathbf{z},\\beta ,\\theta)\\). However, such design results in entanglement of expression control with the generative latent code \\(\\mathbf{z}\\), inducing identity or appearance changes when varying expressions. It is also hard to synthesize images out of training distribution of expressions \\((\\beta ,\\theta)\\). We therefore leave our triplane generation disentangled but when decoding neural feature field \\((\\sigma ,\\mathbf{f})\\) from sampled triplane features, we additionally condition the MLP-decoder on \\(\\beta\\) and \\(\\theta\\). Specifically, we have" + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.784, + 0.47, + 0.803 + ], + "angle": 0, + "content": "\\[\n(\\sigma (\\mathbf {x}), \\mathbf {f} (\\mathbf {x}) | \\mathbf {z}, \\mathbf {p}) = \\Phi (\\mathcal {M} (\\bar {\\mathbf {x}} (\\mathbf {x} | \\mathbf {p}) | \\mathbf {z}), \\phi (\\boldsymbol {\\beta}, \\boldsymbol {\\theta})), \\quad (7)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.471, + 0.903 + ], + "angle": 0, + "content": "where both \\(\\Phi\\) and \\(\\phi\\) are tiny MLPs and \\(\\phi\\) regresses an expression-dependent feature vector from \\((\\beta, \\theta)\\) after positional encoding. For better extrapolation to novel expressions and jaw poses, we add Gaussian noise to the conditioning parameters to prevent MLP overfitting." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.481, + 0.634, + 0.499 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.508, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Training and Dataset. Our training is devised into two stages. In the pretraining stage, we build our semantic SDF \\( W \\) using a four 192-dimensional MLP with a collection of 150K FLAME meshes by Gaussian sampling of the control parameters \\( \\mathbf{p} \\). For each mesh, we sample 4K surface samples \\( N \\) with surface normals \\( \\mathbf{n} \\) and ground-truth correspondence point \\( \\bar{\\mathbf{x}}^* \\), and 4K off surface samples \\( F \\) distributed uniformly inside the box-bounded volume. Our canonical mesh has a neutral shape \\( \\alpha = 0 \\), expression \\( \\beta = 0 \\) and neck pose, but with an opening jaw at 10 degrees. We close the face connectivity between the lips for a watertight geometry and also as a proxy geometry for the mouth cavity modeling. We open the jaw at the canonical space such that the spatial geometry and appearance inside the mouth can be modeled with distinguishable triplane features. In the second training stage, we freeze the weights of \\( W \\) for best efficiency and train our model on the FFHQ [24], a human face dataset with 70,000 in-the-wild images. For each image, we estimate the camera pose \\( \\mathbf{c} \\) and FLAME parameters \\( \\mathbf{p} \\) with a nonlinear optimization in fitting 2D landmarks, assuming zero root and neck rotation and camera located 1.0m away from the world origin. We refer to the supplementary materials for more details. We rebalance the FFHQ dataset by duplicating images with large head poses and jaw opening expressions. We note that we also use pairs of images and \\( (\\beta, \\theta) \\) for fine-tuning of a light-weight image encoder" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "12819" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.109, + 0.088, + 0.541, + 0.19 + ], + "angle": 0, + "content": "
ModelFID ↓KID ↓DSα ↑DSβ,θ ↑DSc ↑IS ↑
DiscoFaceGAN-SR [10]42.60.04-27.20.110.59
GAN-Control [47]9.80.005-43.70.280.62
HeadNeRF [21]1630.1681.338.90.250.678
AniFaceGAN-SR [58]30.50.0242.011.10.040.73
3DFaceShop [51]24.80.0180.4564.20.0290.594
Ours5.70.00163.171.90.350.80
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.254, + 0.193, + 0.393, + 0.205 + ], + "angle": 0, + "content": "(a) Baseline comparisons." + }, + { + "type": "table", + "bbox": [ + 0.569, + 0.084, + 0.905, + 0.187 + ], + "angle": 0, + "content": "
ModelOurs wo LencOurs wo LpriorOurs
FID ↓6.06.15.7
KID ↓0.001720.001740.0016
ASD (cm) ↓0.1240.1350.124
var(α) ↓0.0005250.0005770.000515
AED (cm) ↓0.12330.08130.0797
var(β, θ) ↓0.0047070.004060.00404
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.683, + 0.188, + 0.784, + 0.2 + ], + "angle": 0, + "content": "(b) Ablation study." + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.213, + 0.895, + 0.256 + ], + "angle": 0, + "content": "Table 1. (a) Our method outperforms the prior 2D and 3D controllable image synthesis methods in both image quality and control disentanglement. (b) Our method demonstrates the best control accuracy over shape and expression with geometric prior \\(L_{prior}\\) and self-supervised reconstruction loss \\(L_{enc}\\)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.268, + 0.469, + 0.299 + ], + "angle": 0, + "content": "\\(E\\), using a ResNet50 [19] backbone followed with a single-layer MLP." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.324, + 0.307, + 0.34 + ], + "angle": 0, + "content": "4.1. Qualitative Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.352, + 0.47, + 0.549 + ], + "angle": 0, + "content": "Controlled Portrait Synthesis. Our framework enables high-fidelity head synthesis with disentangled control over camera pose, shape and expression, as shown in Figure 1. We compare our method with prior controllable portrait synthesis works include DiscoFaceGAN [10], GAN-Control [47], HeadNeRF [21], AntiFaceGAN [58] and 3DFaceShop [51]. The first two methods integrates 3DMM controls into 2D face synthesis. HeadNeRF [21] is a 3D parametric NeRF-based head model built from both indoor multiview image capture datasets and in-the-wild monocular image collections. AntiFaceGAN [58] and 3DFaceShop [51] are concurrent 3D-aware controllable GANs in face synthesis." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.554, + 0.471, + 0.902 + ], + "angle": 0, + "content": "We qualitatively compare with prior work in Figure 5 for expression control. GAN-Control shows high-quality expression manipulation but with noticeable identity variation with differences in hair, head shape contour and background, whereas DiscoFaceGAN better maintains identity but with lower perceptual quality, and lacks dynamic details as expression varies. The expression control space of 3DFaceShop [51] is sensitive, suffering from appearance changes even with minor expression variation. AniFaceGAN demonstrates visually-consistent expression editing but with limited resolution and image quality, e.g., with blurry artifacts in hair and teeth. HeadNeRF ensures decent consistency in image generation by rendering the conditional NeRF but lacks fine details with very limited perceptual quality. In contrast, our approach produces the most compelling images with consistent appearance and dynamic details under expression changes. Moreover, controllable neck pose is a key factor towards realistic video avatar in applications like talking head synthesis. As shown in Figure. 1, our method achieves explicit control of neck and jaw poses which are seldom explored in prior work. Please refer to supplementary material for qualitative comparison in view consistency and shape editing." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.268, + 0.892, + 0.42 + ], + "angle": 0, + "content": "Dynamic details. Our method presents highly-consistent image synthesis in control of shape, expression and camera poses, but also depicts temporal realism in portrait animation, credited to the modeling of expression-dependent dynamic details. As shown in the last row of Figure 3, the appearance of wrinkles around the mouth and eyes when transiting from a neural expression to smiling largely enhances the animation realism. In comparison, without explicitly modeling the dynamic details, the wrinkles are embedded in the appearance and do not vary with expressions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.427, + 0.892, + 0.593 + ], + "angle": 0, + "content": "Expressive Head Synthesis with Extrapolated Controls. AntiFaceGAN [58] and 3DFaceShop [51] depict dynamic details as well since their generation of the neural fields is conditioned on the input expression latent code. However, their designs result in shape and expression entanglement with the appearance generation, as reflected in the identity changes as shown in Figure. 5. By embedding the controls in a Gaussian latent space with such as Variation AutoEncoder (VAE), their approaches sacrifice expressiveness. The synthesized image quality is also highly correlated to the distribution of training images with target expressions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.595, + 0.892, + 0.777 + ], + "angle": 0, + "content": "In contrast, our tri-plane generation is explicitly disentangled from shape and expression control. Moreover, the volume deformation is independently learnt from the deformed FLAME mesh collections which offer abundant 3D geometric knowledge with largely augmented control space. Therefore we are much less dependent on the distribution of the training images and support better extrapolation to unseen novel expressions. In Figure. 1, we show high-quality synthesized head with extreme jaw and neck articulated movements which do not exist in our training images. Our expression control is also more expressive, supporting subtle expressions like eye blinks (Figure. 1 3)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.787, + 0.74, + 0.803 + ], + "angle": 0, + "content": "4.2. Quantitative Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Image Quality We measure the image quality with Frechet Inception Distance (FID) [20] and Kernel Inception Distance [5] between 50K randomly synthesized images and 50K randomly sampled real images at the resolution of \\(512 \\times 512\\). Since DiscoFaceGAN [10] and AniFaceGAN [58] only synthesize images at \\(256 \\times 256\\) reso" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12820" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.198 + ], + "angle": 0, + "content": "lution, we utilize a state-of-the-art super-resolution model, SwinIR [30] to upsample into \\(512 \\times 512\\) for a fair comparison. As shown in Table. 1a, our method is superior in both FID and KID than all prior work, demonstrating the most compelling image quality. We note that the original EG3D has a slightly lower FID at 4.8 which is expected since we introduce controllability with additional loss regularization." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.208, + 0.473, + 0.345 + ], + "angle": 0, + "content": "Disentanglement To evaluate the disentangled controllability of our model over shape, expression and camera pose, we measure the disentanglement score [10] of synthesized images as \\(\\mathrm{DS}_{\\alpha}\\), \\(\\mathrm{DS}_{\\beta,\\theta}\\) and \\(\\mathrm{DS}_{\\mathbf{c}}\\) respectively. DS measures the stability of other factors when a single attribute is modified in image synthesis. We employ DECA [12] for estimation of FLAME parameters from generated images and calculate the variance of the estimated parameters \\((\\alpha, \\{\\beta, \\theta\\}, \\mathbf{c})\\). Specifically the \\(\\mathrm{DS}_i\\) is calculated as" + }, + { + "type": "equation", + "bbox": [ + 0.132, + 0.353, + 0.47, + 0.391 + ], + "angle": 0, + "content": "\\[\nD S _ {i} = \\prod_ {j \\neq i} \\frac {\\operatorname {v a r} (i)}{\\operatorname {v a r} (j)}, i, j \\in \\{\\boldsymbol {\\alpha}, \\{\\boldsymbol {\\beta}, \\boldsymbol {\\theta} \\}, \\mathbf {c} \\}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.399, + 0.473, + 0.506 + ], + "angle": 0, + "content": "Higher value of DS indicates better disentanglement. Additionally, we evaluate the identity similarity between pairs of synthesized images with random camera poses and expressions by calculating the cosine similarity of the face embeddings with a pre-trained face recognition module [26]. Our approach demonstrates the best disentanglement numerically over all prior work as indicated in Table. 1a." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.514, + 0.245, + 0.529 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.537, + 0.473, + 0.841 + ], + "angle": 0, + "content": "We ablate the efficacy of the individual component by removing it from our full pipeline. As shown in Figure. 3, we show the loss of control accuracy over shape and expression respectively when removing the geometric prior loss \\( L_{prior} \\) and control loss \\( L_{enc} \\). Conditioning the neural radiance field on expression is also critical to the modeling of dynamic details. Numerically we validate the efficacy of \\( L_{prior} \\) and \\( L_{enc} \\) with Average Shape Distance (ASD) and Average Expression Distance (AED). From the prior distribution, we randomly sample 500 shapes and expressions, with which we synthesize images with 10 different identities. We then reconstruct FLAME parameters from the synthesized images and compare against the input control. We compute 3D per-vertex \\( L_{1} \\) distance of FLAME meshes for ASD while we use 3D landmarks \\( L_{1} \\) distance for AED calculation. Additionally we compute the variance of estimated shapes and expressions within each control group, with lower value indicating more precise control. The efficacy of \\( L_{prior} \\) and \\( L_{enc} \\) is well evidenced in Table. 1b, with even slight image quality improvements." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.848, + 0.212, + 0.864 + ], + "angle": 0, + "content": "4.4. Applications" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Talking Head Video Generation. In Figure. 1, we showcase talking head video generation in controllable views" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.244 + ], + "angle": 0, + "content": "driven with animation sequences of FLAME. Thanks to the high control accuracy, our method is able to synthesize various talking head videos with the same head movements and expressions performed by different identities. Our method is expressive in depicting both large articulated neck and jaw movements and subtle facial expressions like eye blinks, with rich dynamic details. Shape manipulation is easily achievable as well by modifying the shape parameters. Please refer to our supplementary materials for more results in high resolution." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.251, + 0.895, + 0.493 + ], + "angle": 0, + "content": "Portrait Image Manipulation and Animation. As illustrated in Figure. 1, our model also supports 3D-aware face reenactment of a single-view portrait to a video sequence. To achieve that, we perform an optimization in the latent \\( Z+ \\) space [24] to find the corresponding latent embedding, with FLAME parameter and camera pose estimated from the input portrait. With a frozen generator, the optimization is performed by measuring the similarity between generated image and real image using the \\( L_{2} \\) loss and LPIPS loss [65]. For better reconstruction quality, we alter the parameters of the tri-plane synthesis module with a fixed optimized latent code [43]. After that one can explicitly manipulate the portrait with a preserved identity and in a different camera pose and expression. With the expression codes \\( (\\beta, \\theta) \\) reconstructed from a video sequence, we are also able to reenact the portrait to the video motion." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.501, + 0.894, + 0.609 + ], + "angle": 0, + "content": "Societal Impact. Our work focuses on improving the controllability of 3D-aware GANs in technical aspects and is not specifically designed for any malicious uses. This being said, we do see that the method could be potentially extended into controversial applications such as generating fake videos. Therefore, we believe that the synthesized images and videos should present themselves as synthetic." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.619, + 0.62, + 0.635 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.895, + 0.901 + ], + "angle": 0, + "content": "In this work, we introduce OmniAvatar, a novel 3D-aware generative model for synthesis of controllable high-fidelity human head. Our model achieves disentangled semantic control by factoring the generative process into 3D-aware canonical head synthesis and implicit volume deformation to target shapes and expressions. By learning a deformable signed distance function with canonical volumetric correspondence, we instill the geometric prior knowledge of a 3D statistical head model into the generation of neural radiance fields, enabling superior generative capability to novel shapes and expressions. Our approach demonstrates expressive and compelling talking head generation and portrait image animation with fine-grained control accuracy and temporal realism. We believe the proposed method presents an interesting direction for 3D avatar creation and animation, which sheds light on many potential downstream tasks." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "12821" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.157 + ], + "angle": 0, + "content": "[1] Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas Guibas. Learning representations and generative models for 3d point clouds. In ICML, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.16, + 0.47, + 0.2 + ], + "angle": 0, + "content": "[2] Thiemo Alldieck, Hongyi Xu, and Cristian Sminchisescu. imghum: Implicit generative models of 3d human shape and articulated pose. In ICCV, pages 5461-5470, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.203, + 0.47, + 0.258 + ], + "angle": 0, + "content": "[3] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In iccv, pages 5855–5864, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.26, + 0.469, + 0.314 + ], + "angle": 0, + "content": "[4] Alexander W Bergman, Petr Kellnhofer, Yifan Wang, Eric R Chan, David B Lindell, and Gordon Wetzstein. Generative neural articulated radiance fields. arXiv preprint arXiv:2206.14314, 2022. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.317, + 0.469, + 0.372 + ], + "angle": 0, + "content": "[5] Sai Bi, Kalyan Sunkavalli, Federico Perazzi, Eli Shechtman, Vladimir G Kim, and Ravi Ramamoorthi. Deep cg2real: Synthetic-to-real translation via image disentanglement. In iccv, pages 2730-2739, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.374, + 0.469, + 0.429 + ], + "angle": 0, + "content": "[6] Volker Blanz and Thomas Vetter. A morphable model for the synthesis of 3d faces. In Proceedings of the 26th annual conference on Computer graphics and interactive techniques, 1999. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.431, + 0.469, + 0.499 + ], + "angle": 0, + "content": "[7] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. CVPR, 2022. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.502, + 0.469, + 0.571 + ], + "angle": 0, + "content": "[8] Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5799-5809, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.573, + 0.469, + 0.614 + ], + "angle": 0, + "content": "[9] Anpei Chen, Ruiyang Liu, Ling Xie, Zhang Chen, Hao Su, and Jingyi Yu. Sofgan: A portrait image generator with dynamic styling. sigg, 41(1):1-26, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.616, + 0.469, + 0.67 + ], + "angle": 0, + "content": "[10] Yu Deng, Jiaolong Yang, Dong Chen, Fang Wen, and Xin Tong. Disentangled and controllable face image generation via 3d imitative-contrastive learning. In cvpr, pages 5154–5163, 2020. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.673, + 0.469, + 0.714 + ], + "angle": 0, + "content": "[11] Yu Deng, Jiaolong Yang, Jianfeng Xiang, and Xin Tong. Gram: Generative radiance manifolds for 3d-aware image generation. CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.716, + 0.469, + 0.757 + ], + "angle": 0, + "content": "[12] Yao Feng, Haiwen Feng, Michael J. Black, and Timo Bolkart. Learning an animatable detailed 3D face model from in-the-wild images. volume 40, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.76, + 0.469, + 0.828 + ], + "angle": 0, + "content": "[13] Guy Gafni, Justus Thies, Michael Zollhofer, and Matthias Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8649-8658, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[14] Partha Ghosh, Pravir Singh Gupta, Roy Uziel, Anurag Ranjan, Michael J Black, and Timo Bolkart. *Gif: Generative interpretable faces*. In 3dv, pages 868-878. IEEE, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.874, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[15] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.119 + ], + "angle": 0, + "content": "Yoshua Bengio. Generative adversarial nets. In NeurIPS, 2014. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[16] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. In ICML, pages 3569-3579. 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.164, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[17] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. Stylenerf: A style-based 3d-aware generator for high-resolution image synthesis. CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.206, + 0.892, + 0.258 + ], + "angle": 0, + "content": "[18] Yudong Guo, Keyu Chen, Sen Liang, Yong-Jin Liu, Hujun Bao, and Juyong Zhang. Ad-nerf: Audio driven neural radiance fields for talking head synthesis. In iccv, pages 5784-5794, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.261, + 0.892, + 0.3 + ], + "angle": 0, + "content": "[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In cvpr, pages 770-778, 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.303, + 0.892, + 0.356 + ], + "angle": 0, + "content": "[20] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. NeurIPS, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.359, + 0.892, + 0.399 + ], + "angle": 0, + "content": "[21] Yang Hong, Bo Peng, Haiyao Xiao, Ligang Liu, and Juyong Zhang. Headnerf: A real-time nerf-based parametric head model. In cvpr, pages 20374-20384, 2022. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.401, + 0.892, + 0.44 + ], + "angle": 0, + "content": "[22] Kacper Kania, Kwang Moo Yi, Marek Kowalski, Tomasz Trzciński, and Andrea Tagliasacchi. Conerf: Controllable neural radiance fields. In cvpr, pages 18623-18632, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.443, + 0.892, + 0.482 + ], + "angle": 0, + "content": "[23] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. NeurIPS, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.484, + 0.892, + 0.523 + ], + "angle": 0, + "content": "[24] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, 2019. 1, 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.526, + 0.892, + 0.566 + ], + "angle": 0, + "content": "[25] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. In CVPR, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.568, + 0.892, + 0.607 + ], + "angle": 0, + "content": "[26] Minchul Kim, Anil K Jain, and Xiaoming Liu. Adaface: Quality adaptive margin for face recognition. In cvpr, pages 18750-18759, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.61, + 0.892, + 0.664 + ], + "angle": 0, + "content": "[27] Marek Kowalski, Stephan J Garbin, Virginia Estellers, Tadas Baltrusaitis, Matthew Johnson, and Jamie Shotton. Config: Controllable neural face image generation. In \\(eccv\\), pages 299-315. Springer, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.665, + 0.892, + 0.705 + ], + "angle": 0, + "content": "[28] Ruihui Li, Xianzhi Li, Chi-Wing Fu, Daniel Cohen-Or, and Pheng-Ann Heng. Pu-gan: a point cloud upsampling adversarial network. In ICCV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.707, + 0.892, + 0.76 + ], + "angle": 0, + "content": "[29] Tianye Li, Timo Bolkart, Michael. J. Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4D scans. ACM Transactions on Graphics, 36(6):194:1-194:17, 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.763, + 0.892, + 0.803 + ], + "angle": 0, + "content": "[30] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In iccv, pages 1833-1844, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.805, + 0.892, + 0.845 + ], + "angle": 0, + "content": "[31] Yiyi Liao, Katja Schwarz, Lars Mescheder, and Andreas Geiger. Towards unsupervised learning of generative models for 3D controllable image synthesis. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.847, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[32] Lingjie Liu, Marc Habermann, Viktor Rudnev, Kripasindhu Sarkar, Jiatao Gu, and Christian Theobalt. Neural actor: Neural free-view synthesis of human actors with pose control. ACM Trans. on Graphics, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12822" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[33] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.47, + 0.204 + ], + "angle": 0, + "content": "[34] Thu Nguyen-Phuoc, Chuan Li, Lucas Theis, Christian Richardt, and Yong-Liang Yang. HoloGAN: Unsupervised learning of 3D representations from natural images. In ICCV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.206, + 0.469, + 0.261 + ], + "angle": 0, + "content": "[35] Thu Nguyen-Phuoc, Christian Richardt, Long Mai, Yong-Liang Yang, and Niloy Mitra. BlockGAN: Learning 3D object-aware scene representations from unlabelled images. In NeurIPS, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.263, + 0.469, + 0.303 + ], + "angle": 0, + "content": "[36] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.305, + 0.469, + 0.36 + ], + "angle": 0, + "content": "[37] Roy Or-El, Xuan Luo, Mengyi Shan, Eli Shechtman, Jeong Joon Park, and Ira Kemelmacher-Shlizerman. Stylesdf: High-resolution 3d-consistent image and geometry generation. CVPR, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.362, + 0.469, + 0.417 + ], + "angle": 0, + "content": "[38] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In ICCV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.419, + 0.469, + 0.489 + ], + "angle": 0, + "content": "[39] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. ACM Transactions on Graphics, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.49, + 0.469, + 0.559 + ], + "angle": 0, + "content": "[40] Pascal Paysan, Reinhard Knothe, Brian Amberg, Sami Romdhani, and Thomas Vetter. A 3d face model for pose and illumination invariant face recognition. In 2009 sixth IEEE international conference on advanced video and signal based surveillance, pages 296-301. IEEE, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.561, + 0.469, + 0.602 + ], + "angle": 0, + "content": "[41] Jingtan Piao, Keqiang Sun, Quan Wang, Kwan-Yee Lin, and Hongsheng Li. Inverting generative adversarial renderer for face reconstruction. In cvpr, pages 15619-15628, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.603, + 0.469, + 0.645 + ], + "angle": 0, + "content": "[42] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In CVPR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.646, + 0.469, + 0.7 + ], + "angle": 0, + "content": "[43] Daniel Roich, Ron Mokady, Amit H Bermano, and Daniel Cohen-Or. Pivotal tuning for latent-based editing of real images. ACM Transactions on Graphics (TOG), 42(1):1-13, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.702, + 0.469, + 0.744 + ], + "angle": 0, + "content": "[44] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. Graf: Generative radiance fields for 3d-aware image synthesis. NeurIPS, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.746, + 0.469, + 0.786 + ], + "angle": 0, + "content": "[45] Yichun Shi, Divyansh Aggarwal, and Anil K Jain. Lifting 2d stylegan for 3d-aware face generation. In cvpr, pages 6258-6266, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.469, + 0.842 + ], + "angle": 0, + "content": "[46] Yichun Shi, Xiao Yang, Yangyue Wan, and Xiaohui Shen. Semanticstylegan: Learning compositional generative priors for controllable image synthesis and editing. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[47] Alon Shoshan, Nadav Bhonker, Igor Kviatkovsky, and Gerard Medioni. Gan-control: Explicitly controllable gans. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14083-14093, 2021. 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "[48] Ivan Skorokhodov, Sergey Tulyakov, Yiqun Wang, and Peter Wonka. Epigraph: Rethinking training of 3d gans. arXiv preprint arXiv:2206.10535, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[49] Keqiang Sun, Shangzhe Wu, Zhaoyang Huang, Ning Zhang, Quan Wang, and HongSheng Li. Controllable 3d face synthesis with conditional generative occupancy fields. arXiv preprint arXiv:2206.08361, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.233 + ], + "angle": 0, + "content": "[50] Attila Szabó, Givi Meishvili, and Paolo Favaro. Unsupervised generative 3D shape learning from natural images. arXiv, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.235, + 0.892, + 0.29 + ], + "angle": 0, + "content": "[51] Junshu Tang, Bo Zhang, Binxin Yang, Ting Zhang, Dong Chen, Lizhuang Ma, and Fang Wen. Explicitly controllable 3d-aware portrait generation. arXiv preprint arXiv:2209.05434, 2022. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.292, + 0.892, + 0.347 + ], + "angle": 0, + "content": "[52] Ayush Tewari, Mohamed Elgharib, Florian Bernard, Hans-Peter Seidel, Patrick Pérez, Michael Zollhöfer, and Christian Theobalt. Pie: Portrait image embedding for semantic control. sigg, 39(6):1-14, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.349, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[53] Ayush Tewari, Mohamed Elgharib, Gaurav Bharaj, Florian Bernard, Hans-Peter Seidel, Patrick Pérez, Michael Zollhofer, and Christian Theobalt. Stylerig: Rigging stylegan for 3d control over portrait images. In cvpr, pages 6142-6151, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.419, + 0.892, + 0.461 + ], + "angle": 0, + "content": "[54] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for stylegan image manipulation. ACM Trans. on Graphics, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.462, + 0.892, + 0.53 + ], + "angle": 0, + "content": "[55] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Christoph Lassner, and Christian Theobalt. Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In ICCV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.532, + 0.892, + 0.588 + ], + "angle": 0, + "content": "[56] Ziyan Wang, Timur Bagautdinov, Stephen Lombardi, Tomas Simon, Jason Saragih, Jessica Hodgins, and Michael Zollhofer. Learning compositional radiance fields of dynamic human heads. In cvpr, pages 5704-5713, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.59, + 0.892, + 0.645 + ], + "angle": 0, + "content": "[57] Jiajun Wu, Chengkai Zhang, Tianfan Xue, William T. Freeman, and Joshua B. Tenenbaum. Learning a probabilistic latent space of object shapes via 3D generative-adversarial modeling. In NeurIPS, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.646, + 0.892, + 0.701 + ], + "angle": 0, + "content": "[58] Yue Wu, Yu Deng, Jiaolong Yang, Fangyun Wei, Qifeng Chen, and Xin Tong. Anifacegan: Animatable 3d-aware face image generation for video avatars. arXiv preprint arXiv:2210.06465, 2022. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.703, + 0.892, + 0.772 + ], + "angle": 0, + "content": "[59] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond. In Computer Graphics Forum, volume 41, pages 641-676. Wiley Online Library, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.773, + 0.892, + 0.814 + ], + "angle": 0, + "content": "[60] Hongyi Xu, Thiemo Alldieck, and Cristian Sminchisescu. H-nerf: Neural radiance fields for rendering and temporal reconstruction of humans in motion. NeurIPS, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.817, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[61] Xudong Xu, Xingang Pan, Dahua Lin, and Bo Dai. Generative occupancy fields for 3d surface-aware image synthesis. NeurIPS, 34:20683-20695, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.859, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[62] Yang Xue, Yuheng Li, Krishna Kumar Singh, and Yong Jae Lee. Giraffe hd: A high-resolution 3d-aware generative model. In CVPR, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "12823" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.133 + ], + "angle": 0, + "content": "[63] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. In NeurIPS, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.134, + 0.47, + 0.205 + ], + "angle": 0, + "content": "[64] Jianfeng Zhang, Zihang Jiang, Dingdong Yang, Hongyi Xu, Yichun Shi, Guoxian Song, Zhongcong Xu, Xinchao Wang, and Jiashi Feng. Avatargen: a 3d generative model for animatable human avatars. arXiv preprint arXiv:2208.00561, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.205, + 0.471, + 0.262 + ], + "angle": 0, + "content": "[65] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In cvpr, pages 586-595, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.262, + 0.471, + 0.319 + ], + "angle": 0, + "content": "[66] Yufeng Zheng, Victoria Fernández Abrevaya, Marcel C Bühler, Xu Chen, Michael J Black, and Otmar Hilliges. Im avatar: Implicit morphable head avatars from videos. In cvpr, pages 13545-13555, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.319, + 0.471, + 0.361 + ], + "angle": 0, + "content": "[67] Peng Zhou, Lingxi Xie, Bingbing Ni, and Qi Tian. CIPS-3D: A 3D-Aware Generator of GANs Based on Conditionally-Independent Pixel Synthesis. arXiv, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.361, + 0.471, + 0.404 + ], + "angle": 0, + "content": "[68] Yiyu Zhuang, Hao Zhu, Xusen Sun, and Xun Cao. Mofan-erf: Morphable facial neural radiance field. arXiv preprint arXiv:2112.02308, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "12824" + } + ] +] \ No newline at end of file diff --git a/2023/OmniAvatar_ Geometry-Guided Controllable 3D Head Synthesis/13b82072-fc3f-4f4d-8e9e-44b54de26be7_origin.pdf b/2023/OmniAvatar_ Geometry-Guided Controllable 3D Head Synthesis/13b82072-fc3f-4f4d-8e9e-44b54de26be7_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f5abc5d7d0dc12020458d8c3e6ef396a38f30a07 --- /dev/null +++ b/2023/OmniAvatar_ Geometry-Guided Controllable 3D Head Synthesis/13b82072-fc3f-4f4d-8e9e-44b54de26be7_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:115363faa7dead41a063c4eaebbadf670d41d7b5f441c2c997fb7ccf737d906d +size 8588061 diff --git a/2023/OmniAvatar_ Geometry-Guided Controllable 3D Head Synthesis/full.md b/2023/OmniAvatar_ Geometry-Guided Controllable 3D Head Synthesis/full.md new file mode 100644 index 0000000000000000000000000000000000000000..3b428df9f1b5186a7fc68bbcb25109a5ee8da82e --- /dev/null +++ b/2023/OmniAvatar_ Geometry-Guided Controllable 3D Head Synthesis/full.md @@ -0,0 +1,285 @@ +# OmniAvatar: Geometry-Guided Controllable 3D Head Synthesis + +Hongyi Xu1 Guoxian Song1 Zihang Jiang1,2 Jianfeng Zhang1,2 Yichun Shi1 Jing Liu1 Wanchun Ma1 Jiashi Feng1 Linjie Luo1 +1ByteDance Inc 2National University of Singapore +{hongyixu, guoxian.song, zihang.jiang, jianfeng.zhang, yichun.shi, jing.liu, wanchun.ma, jshfeng, linjie.luo}@bytedance.com + +![](images/09c331972bd1db073abd3f9aa8dca90f8d1112b24799daafd1a69425c8b01fb6.jpg) +Figure 1. Our model can synthesize diverse identity-preserved 3D heads with compelling dynamic details under full disentangled control over camera poses, facial expressions, head shapes, articulated neck and jaw poses (left). Our model can also reconstruct 3D heads from a single photo reference and enable multi-view-consistent head reenactment (right). + +# Abstract + +We present OmniAvatar, a novel geometry-guided 3D head synthesis model trained from in-the-wild unstructured images that is capable of synthesizing diverse identity-preserved 3D heads with compelling dynamic details under full disentangled control over camera poses, facial expressions, head shapes, articulated neck and jaw poses. To achieve such high level of disentangled control, we first explicitly define a novel semantic signed distance function (SDF) around a head geometry (FLAME) conditioned on the control parameters. This semantic SDF allows us to build a differentiable volumetric correspondence map from the observation space to a disentangled canonical space from all the control parameters. We then leverage the 3D-aware GAN framework (EG3D) to synthesize detailed shape and appearance of 3D full heads in the canonical space, followed by a volume rendering step guided by the volumetric correspondence map to output into the observation space. To ensure the control accuracy on the synthesized head + +shapes and expressions, we introduce a geometry prior loss to conform to head SDF and a control loss to conform to the expression code. Further, we enhance the temporal realism with dynamic details conditioned upon varying expressions and joint poses. Our model can synthesize more preferable identity-preserved 3D heads with compelling dynamic details compared to the state-of-the-art methods both qualitatively and quantitatively. We also provide an ablation study to justify many of our system design choices. + +# 1. Introduction + +Photo-realistic face image synthesis, editing and animation attract significant interests in computer vision and graphics, with a wide range of important downstream applications in visual effects, digital avatars, telepresence and many others. With the advent of Generative Adversarial Networks (GANs) [15], remarkable progress has been achieved in face image synthesis by StyleGAN [23-25] + +as well as in semantic and style editing for face images [46, 54]. To manipulate and animate the expressions and poses in face images, many methods attempted to leverage 3D parametric face models, such as 3D Morphable Models (3DMMs) [6, 40], with StyleGAN-based synthesis models [10, 41, 53]. However, all these methods operate on 2D convolutional networks (CNNs) without explicitly enforcing the underlying 3D face structure. Therefore they cannot strictly maintain the 3D consistency when synthesizing faces under different poses and expressions. + +Recently, a line of work has explored neural 3D representations by unsupervised training of 3D-aware GANs from in-the-wild unstructured images [7, 8, 11, 17, 37, 44, 45, 48, 61, 62, 67]. Among them, methods with generative Neural Radiance Fields (NeRFs) [33] have demonstrated striking quality and multi-view-consistent image synthesis [7, 11, 17, 37, 48]. The progress is largely due to the integration of the power of StyleGAN in photo-realistic image synthesis and NeRF representation in 3D scene modeling with view-consistent volumetric rendering. Nevertheless, these methods lack precise 3D control over the generated faces beyond camera pose, as well as the quality and consistency in control over other attributes, such as shape, expression, neck and jaw pose, leave much to be desired. + +In this work, we present OmniAvatar, a novel geometry-guided 3D head synthesis model trained from in-the-wild unstructured images. Our model can synthesize a wide range of 3D human heads with full control over camera poses, facial expressions, head shapes, articulated neck and jaw poses. To achieve such high level of disentangled control for 3D human head synthesis, we devise our model learning in two stages. We first define a novel semantic signed distance function (SDF) around a head geometry (i.e. FLAME [29]) conditioned on its control parameters. This semantic SDF fully distills rich 3D geometric prior knowledge from the statistical FLAME model and allows us to build a differentiable volumetric correspondence map from the observation space to a disentangled canonical space from all the control parameters. In the second training stage, we then leverage the state-of-the-art 3D GAN framework (EG3D [7]) to synthesize realistic shape and appearance of 3D human heads in the canonical space, including the modeling of hair and apparels. Following that, a volume rendering step is guided by the volumetric correspondence map to output the geometry and image in the observation space. + +To ensure the consistency of synthesized 3D head shape with controlling head geometry, we introduce a geometry prior loss to minimize the difference between the synthesized neural density field and the FLAME head SDF in observation space. Furthermore, to improve the control accuracy, we pre-train an image encoder of the control parameters and formulate a control loss to ensure synthesized images matching the input control code upon encoding. An- + +other key aspect of synthesis realism is dynamic details such as wrinkles and varying shading as subjects change expressions and poses. To synthesize dynamic details, we propose to condition EG3D's triplane feature decoding with noised controlling expression. + +Compare to state-of-the-art methods, our method achieves superior synthesized image quality in terms of Frechet Inception Distance (FID) and Kernel Inception Distance (KID). Our method can consistently preserve the identity of synthesized subjects with compelling dynamic details while changing expressions and poses, outperforming prior methods both quantitatively and qualitatively. + +The contributions of our work can be summarized as: + +- A novel geometry-guided 3D GAN framework for high-quality 3D head synthesis with full control on camera poses, facial expressions, head shapes, articulated neck and jaw poses. +- A novel semantic SDF formulation that defines the volumetric correspondence map from observation space to canonical space and allows full disentanglement of control parameters in 3D GAN training. +- A geometric prior loss and a control loss to ensure the head shape and expression synthesis accuracy. +- A robust noised expression conditioning scheme to enable dynamic detail synthesis. + +# 2. Related Work + +3D-Aware Generative Image Synthesis. Generative adversarial networks [15, 24, 25] gained popularity over the last decade due to their remarkable ability in photo-realistic image synthesis. Building on the success of 2D image-based GANs, recent works have extended the capabilities to view-consistent image synthesis with unsupervised learning from 2D single-view images. The key idea is to combine differential rendering with 3D scene representations, such as meshes [31, 50], point clouds [1, 28], voxels [34, 35, 57], and recently implicit neural representation [7, 8, 11, 17, 36, 37, 44, 48, 67]. We build our work on recent 3D GAN model by Chan et al [7] that uses an efficient triplane-based NeRF generation, combined with 2D CNN-based super-resolution. Even though 3D-aware GANs are able to control camera viewpoints, they lack precise 3D control over the other attributes such as shapes and expressions. In this work, we empower 3D-aware GANs with disentangled precise control over shapes and expressions. + +Controllable Face Image Synthesis. Considerable work [9, 10, 14, 27, 41, 52, 53] has been devoted to incorporate 3D priors of statistical face models, such as 3D Morphable Models (3DMMs) [6, 40], in controllable face synthesis and animation. Among them, DiscoFaceGAN [10] proposed imitative-contrastive learning to mimic the 3DMM rendering process by the generative model. + +![](images/b04fe5cf49739ad18a3cd42dd0bec0dfdd222f7a0ed9e98008660542263f19a0.jpg) +Figure 2. Overview of our training framework. Stage 1: Trained from parameterized FLAME [29] mesh collections, a MLP-network $W$ maps a shape $\alpha$ , expression $\theta$ and articulated jaw and neck pose $\theta$ into 3D point-to-point volumetric correspondences from observation to canonical space, together with a signed distance function of the corresponding FLAME head. Stage 2: Given a Gaussian latent code $\mathbf{z}$ , our model generates a tri-plane represented 3D feature space of a canonical head, disentangled with shape and expression controls. The volume rendering is then guided by the volumetric correspondence field to map the decoded neural radiance field from the canonical to observation space. We condition the NeRF decoding with expression and joint pose for modeling dynamic details. A super-resolution module synthesizes the final high-resolution RGB image from the volume-rendered feature map. For fine-grained shape and expression control, we apply the FLAME SDF as geometric prior to the synthesized NeRF density, and self-supervise the image synthesis to commit to the target expression $\beta$ and joint pose $\theta$ by comparing the input code against the re-estimated values $\hat{\beta}, \hat{\theta}$ from synthesized images. + +A similar strategy has also been adopted with concurrent and follow-up works [14, 27, 41, 53]. However all of these approaches suffer from 3D inconsistency due to the use of 2D CNNs as image renderer. HeadNeRF [21] combines 3DMM with 3D NeRF representation, and is able to synthesize 3D heads conditioned on 3DMM attributes. However, the training relies on annotated multiview datasets whereas our approach learns the disentangled 3D head synthesis with only single-view images. There has been concurrent work [4, 49, 51, 58, 64] to ours in 3D-aware controllable face or full-body GANs. Differently from these approaches, we use a full-head parametric model FLAME [29], and fully exploit the spatial geometric prior knowledge beyond the surface deformation and skinning. We have also achieved fine-grained control with our novel losses and enhanced face animation with rich dynamic details. + +Controllable Neural Implicit Field of Face. Neural implicit functions [59], have emerged as a powerful continuous and differential representations of 3D scenes. Among them, Neural Radiance Field [3, 33] has been widely adopted due to its superiority in modeling complex scene details and synthesizing multiview images with inherited 3D consistency. While initial proposals have focused on static scene modeling, recent work has successfully demonstrated application of NeRF in modeling dynamic scenes [32,38,42,55,60]. In particular, dynamic neural radi- + +ance fields of human heads [13,18,22,38,39,56,66,68] have enabled photo-realistic head animation, often by conditioning with pose parameters or deforming the radiance field with 3D morphable models. However, they do not leverage a generative training paradigm and focus on scene-specific learning from video sequences or multiview images. In contrast, our model learns generative and controllable neural radiance fields from widely accessible single-view images. + +# 3. Method + +Our goal is to build a geometry-guided 3D head synthesis model with full control of camera poses, face shapes and expressions, trained from in-the-wild unstructured image collections. + +# 3.1. Overview + +Problem. To achieve our goal, we leverage a 3D-aware generator (EG3D [7]) for photo-realistic, multiview consistent image synthesis, while disentangle control of head geometric attributes from image generation with a 3D statistical head model (FLAME [29]). Specifically, given a random Gaussian-sampled latent code $\mathbf{z}$ , a camera pose $\mathbf{c}$ and a FLAME parameter $\mathbf{p} = (\alpha, \beta, \theta)$ consisting of shape $\alpha$ , expression $\beta$ , jaw and neck pose $\theta$ , the generator $G$ synthesizes a photo-realistic human head image $I_{RGB}(\mathbf{z}|\mathbf{c},\mathbf{p})$ with corresponding attributes as defined in $\mathbf{p}$ . + +**Framework.** As illustrated in our pipeline 2, our controllable 3D-aware GAN is trained in two stages. From a large collection of 3D deformed FLAME meshes, we first pre-train a deformable semantic SDF around the FLAME geometry that builds a differential volumetric correspondence map from the observation to a predefined canonical space (Section 3.2.1). In the second stage, guided by the pre-trained volumetric mapping, we then deform the detailed 3D full heads synthesized in the disentangled canonical space to the desired shapes and expressions (Section 3.2.2). Fine-grained expression control is achieved by supervising image synthesis such that expressions estimated from the generated images is consistent with the input control (Section 3.2.3). Our approach further enhances temporal realism with dynamic details, such as dimples and wrinkles, synthesizing realistic shading and geometric variations as expression changes (Section 3.2.4). + +3D-Aware GAN Background. To ensure appearance consistency from different views, we choose EG3D [7] as our backbone for 3D-aware image synthesis. The generator $G$ takes a random latent code $\mathbf{z}$ and conditioning camera label $\hat{\mathbf{c}}$ , and maps to a manifold of triplane features $\mathcal{M}(\mathbf{z},\hat{\mathbf{c}})$ . For presentation clarity, we absorb $\hat{\mathbf{c}}$ to $\mathbf{z}$ and simply denote triplane as $\mathcal{M}(\mathbf{z})$ . A low-resolution feature map $I^{*}(\mathbf{z}|\mathbf{c})$ is then rendered from a desired camera pose $\mathbf{c}$ by sampling the triplane features and integrating MLP-decoded neural radiance $(\sigma ,\mathbf{f})$ along camera rays. A super-resolution module is followed to modulate the feature map and synthesize the final RGB images at high resolution. We train $G$ and a dual discriminator $D$ with adversarial training. + +# 3.2. Controllable 3D-Aware Image Synthesis + +To synthesize an image $\mathbf{I}(\mathbf{z}|\mathbf{c},\mathbf{p})$ with desired FLAME control $\mathbf{p}$ , we leverage the EG3D framework to generate a triplane-based 3D volumetric feature space $\mathcal{M}(\mathbf{z})$ of a synthesized identity with canonical shape and expression. Guided by our pretrained volumetric correspondence map, we deform the synthesized feature volume into our target observation space, which is further decoded and volume rendered into high-fidelity head appearance and geometry with the target shape and expression. Our design explicitly disentangles the underlying geometric variations of changing shape and expression from canonical geometry and appearance synthesis. Following EG3D [7], we associate each training image with a set of camera parameters $\mathbf{c}$ and control parameters $\mathbf{p}$ , which are obtained from a nonlinear 2D landmarks-based optimization. + +# 3.2.1 Semantic Signed Distance Function + +For disentangled geometric modeling, we formulate an implicit semantic SDF representation $W(\mathbf{x}|\mathbf{p} = (\alpha ,\beta ,\theta)) = (s,\bar{\mathbf{x}})$ , where $\alpha ,\beta$ are the linear shape and expression + +![](images/5375ab76ac4b310a1596b83d56b7d77f1d6b31faff19c0e4c5b218a5c3d05ea5.jpg) +Figure 3. We synthesize different identities with the same shape and expression code respectively in (a) and (b). We observe less shape variation with our geometric prior loss (a) and more consistent expression with our control loss (b). In (c), we show our expression-conditioned NeRF modeling builds rich dynamic details as the subject varies expressions. + +blendshape coefficients, and $\theta$ controls the rotation of a 3-DoF jaw and neck joint. Specifically, given a spatial point $\mathbf{x}$ in observation space $\mathcal{O}(\mathbf{p})$ , $W$ returns its 3D correspondence point $\bar{\mathbf{x}}$ (i.e., semantics) in canonical space $\mathcal{C}(\bar{\mathbf{p}})$ , with which we project and query the triplane features $\mathcal{M}(\mathbf{z})$ . Additionally it also computes the closest signed distance $s(\mathbf{x}|\mathbf{p})$ to the FLAME mesh surface $\mathbf{S}(\mathbf{p})$ . We illustrate the function in Figure 4. We co-learn the highly-correlated volumetric correspondence and SDF with the property that the signed distance is preserved between canonical and observation correspondence points as $s(\bar{\mathbf{x}}|\bar{\mathbf{p}}) = s(\mathbf{x}|\mathbf{p})$ . + +We learn $W(\mathbf{x}|\mathbf{p})$ with a large corpus of 3D FLAME meshes $\mathbf{S}(\mathbf{p})$ sampled from its parametric control space. Similar to IGR [16] and imGHUM [2], we model our implicit field as an MLP and optimize $W(\mathbf{x}|\mathbf{p})$ with losses, + +$$ +L _ {i s o} = \frac {1}{| N |} \sum_ {\mathbf {x} \in N} (| s (\mathbf {x} | \mathbf {p}) | + | \nabla s _ {\mathbf {x}} (\mathbf {x} | \mathbf {p}) - \mathbf {n} (\mathbf {x} | \mathbf {p}) |), \tag {1} +$$ + +$$ +L _ {e i k} = \frac {1}{| F |} \sum_ {\mathbf {x} \in F} \| \nabla s _ {\mathbf {x}} (\mathbf {x} | \mathbf {p}) - 1 \| _ {2}, \tag {2} +$$ + +$$ +L _ {s e m} = \frac {1}{| N |} \sum_ {\mathbf {x} \in N} \left(| \bar {\mathbf {x}} (\mathbf {x} | \mathbf {p}) - \bar {\mathbf {x}} ^ {*} (\mathbf {x} | \bar {\mathbf {p}}) |\right) \tag {3} +$$ + +where $N,F$ are a batch of on and off surface samples. For the surface samples, the $L_{iso}$ encourages the signed distance values to be on the zero-level-set and the SDF gradient to be equal to the given surface normals $\mathbf{n}$ . The Eikonal loss $L_{eik}$ is derived from [16] where the SDF is differen + +![](images/637a25f6e637a9c0a2a20adac5ac22048819dac91d8fd54ab84b4a827ea6aa24.jpg) +Figure 4. Illustration to semantic SDF learning. + +tiable everywhere with gradient norm 1. The semantic loss $L_{sem}$ supervises the mapping of surface samples $\mathbf{x} \in F$ to the ground-truth correspondence points $\bar{\mathbf{x}}^*$ on the canonical FLAME surface, where $\bar{\mathbf{x}}^*$ and $\mathbf{x}$ share the same barycentric coordinates. The $L_{eik}$ provides a geometric regularization over the volumetric SDF, whereas the $L_{iso}$ and $L_{sem}$ act as the boundary condition to the SDF and volumetric correspondence field respectively. + +To guide the canonical correspondence learning for off-the-surface points, one could try to minimize the signed distance difference between $s(\bar{\mathbf{x}}|\bar{\mathbf{p}})$ and $s(\mathbf{x}|\mathbf{p})$ . However, we note that the volume correspondence between the observation and canonical space is still ill-regularized, considering infinite number of points exist with the same signed distance value. We therefore reformulate volumetric correspondence field as $W(\mathbf{x}|\mathbf{p}) = (\bar{s}(\bar{\mathbf{x}}), \bar{\mathbf{x}})$ , where the signed distance of an observation-space point is obtained by mapping it into its canonical correspondence $\bar{\mathbf{x}}$ and querying the pre-computed canonical SDF $\bar{s}$ . Thus we only learn a correspondence field with which we can deform the canonical SDF to different FLAME configurations, and then supervise the deformed SDFs with the FLAME surface boundary, normals $(L_{iso})$ and Eikonal regularization $(L_{eik})$ . As such, even for off-the-surface samples, their canonical correspondences are well regularized in space with the geometric properties of signed distance functions via $L_{iso}$ and $L_{eik}$ . In contrast to explicit mesh deformation [4], our implicit volumetric correspondence is more accurate, differentiable, smooth everywhere and semantically consistent with the properties of SDF. + +# 3.2.2 Canonical Generation with Geometric Prior + +With our pretrained semantic SDF $W(\mathbf{x}|\mathbf{p})$ modeling the shape and expression variation, we leverage the triplane for 3d-aware generation of human heads with canonical shape and expression. In particular, to generate a neural radiance feature $(\sigma, \mathbf{f})$ for a point $\mathbf{x}$ in observation space $\mathcal{O}(\mathbf{p})$ , we use our correspondence function $W$ to back warp $\mathbf{x}$ into $\bar{\mathbf{x}}$ , with which we project and sample the canonical triplane features followed with a tiny MLP decoding. + +In spite of the control disentanglement, there is no ex + +plicit loss that constrains the triplane-generated neural radiance density to conform to the shape and expression as defined in $\mathbf{p}$ . Therefore, we guide the generation of neural radiance density field by minimizing its difference to the underlying FLAME head geometry represented with SDF, as + +$$ +L _ {p r i o r} = \frac {1}{| R |} \sum_ {\mathbf {x} \in R} e ^ {- \gamma | s (\mathbf {x} | p) |} \left| \sigma \left(\mathbf {x} \mid \mathbf {z}, \mathbf {p}\right) - \sigma^ {*} \left(\mathbf {x} \mid \mathbf {p}\right) \right|, \tag {4} +$$ + +$$ +\sigma^ {*} (\mathbf {x} | \mathbf {p}) = \frac {1}{\kappa} \cdot \operatorname {S i g m o i d} \left(\frac {- s (\mathbf {x} | \mathbf {p})}{\kappa}\right) \tag {5} +$$ + +where $R$ is the stratified ray samples for volume rendering and $\kappa$ is a learnable scalar controlling the density tightness around the SDF boundary. Following [37, 63], we convert SDF value $s(\mathbf{x}|\mathbf{p})$ to proxy 3D density $\sigma^{*}(\mathbf{x}|\mathbf{p})$ assuming non-hollow surfaces. We decay the weights for our geometric prior loss $L_{prior}$ as the point moving away from the SDF boundary, allowing higher degrees of freedom in generation of residual geometries, such as hair and glasses. The geometric prior loss effectively guides the 3D head geometry learning but should not be overpowered which otherwise might lead to loss of geometric details. + +# 3.2.3 Fine-Grained Expression Control + +Our geometric prior loss $L_{prior}$ provides local 3D pointwise guidance, and is able to well regularize the shape generation and achieve coarse-level expression control. However, for delicate expressions, such as eye blinks, $L_{prior}$ provides little supervision as the geometric variation is subtle. Moreover, for regions with complex correspondences, such as around the lips, it is challenging to guide the formation of correct expressions globally, just with point-wise geometric losses. To improve the control granularity, we propose an image-level supervision loss that requires a synthesized image $I_{RGB}(\mathbf{z}|\mathbf{c},\mathbf{p})$ matching the target expression as defined in the input $\mathbf{p}$ . Using our training images with estimated control labels $\mathbf{p}$ , we first pretrain an image encoder $E(I_{RGB}) = (\tilde{\boldsymbol{\beta}},\tilde{\boldsymbol{\theta}})$ that regresses the expression coefficients $\tilde{\boldsymbol{\beta}}$ and joint poses $\tilde{\boldsymbol{\theta}}$ . During our 3D GAN training, we then apply our image-level control supervision as + +$$ +\begin{array}{l} L _ {e n c} = | \tilde {\boldsymbol {\beta}} - \boldsymbol {\beta} | + | \tilde {\boldsymbol {\theta}} - \boldsymbol {\theta} | + | \mathbf {S} (\boldsymbol {\alpha}, \tilde {\boldsymbol {\beta}}, \tilde {\boldsymbol {\theta}}) - \mathbf {S} (\boldsymbol {\alpha}, \boldsymbol {\beta}, \boldsymbol {\theta}) | + \\ + \left| \mathbf {J} \mathbf {S} (\boldsymbol {\alpha}, \tilde {\boldsymbol {\beta}}, \tilde {\boldsymbol {\theta}}) - \mathbf {J} \mathbf {S} (\boldsymbol {\alpha}, \boldsymbol {\beta}, \boldsymbol {\theta}) \right|, \tag {6} \\ \end{array} +$$ + +where $\mathbf{S},\mathbf{J}$ are the FLAME mesh vertices and 3D landmarks regressor respectively. While being straightforward for the first 2 terms, the last two terms in $L_{enc}$ penalize deviation of 3D vertex coordinates and surface landmarks after mesh decoding. We note that we do not supervise shape $\alpha$ in $L_{enc}$ since our geometric prior loss $L_{prior}$ suffices in shape control already and also due to the ambiguity of shape scaling estimated from monocular images. + +![](images/5c3066ab5d5b8315700701851a923395642694462106b4973e7799cd2ebb3d09.jpg) +Figure 5. Qualitative comparisons in expression control. Under a similar expression control, our method achieves the best perceptual quality with better identity-preservation (compared to GAN-Control [47] and 3DFaceShop [51]), more realistic dynamic details (compared to DiscoFaceGAN [10]) and higher expression control accuracy (compared to AntiFaceGAN [58], 3DFaceShop [51], HeadNeRF [21]). + +![](images/583be0f7e1981a934d961356d719a966e245573b4936a53abc3f71932f68c92e.jpg) + +![](images/cdccdc8738c51f4169406576d78e645eb54363672708f337e347e00ebf695203.jpg) + +![](images/ff1b0720b4720eb7fc5a2bbd5872d0512799ca609fe9d427cc79e6b920c9ad9c.jpg) + +![](images/9d212897a27f4e529472d14ad6dd6482eae3e88b967af1c665fb2d16de795aa8.jpg) + +# 3.2.4 Dynamic Details Modeling + +To this end, we have achieved controllable static image synthesis. However, one should observe the variation of shading and geometric details in a dynamic head motion, such as the appearance of dimples and eye wrinkles in smiling. We consider the appearance of dynamic details is highly correlated with the driving expression $(\beta ,\theta)$ . To model such temporal effects, one could try to condition the triplane generation on expression as $\mathcal{M}(\mathbf{z},\beta ,\theta)$ . However, such design results in entanglement of expression control with the generative latent code $\mathbf{z}$ , inducing identity or appearance changes when varying expressions. It is also hard to synthesize images out of training distribution of expressions $(\beta ,\theta)$ . We therefore leave our triplane generation disentangled but when decoding neural feature field $(\sigma ,\mathbf{f})$ from sampled triplane features, we additionally condition the MLP-decoder on $\beta$ and $\theta$ . Specifically, we have + +$$ +(\sigma (\mathbf {x}), \mathbf {f} (\mathbf {x}) | \mathbf {z}, \mathbf {p}) = \Phi (\mathcal {M} (\bar {\mathbf {x}} (\mathbf {x} | \mathbf {p}) | \mathbf {z}), \phi (\boldsymbol {\beta}, \boldsymbol {\theta})), \quad (7) +$$ + +where both $\Phi$ and $\phi$ are tiny MLPs and $\phi$ regresses an expression-dependent feature vector from $(\beta, \theta)$ after positional encoding. For better extrapolation to novel expressions and jaw poses, we add Gaussian noise to the conditioning parameters to prevent MLP overfitting. + +# 4. Experiments + +Training and Dataset. Our training is devised into two stages. In the pretraining stage, we build our semantic SDF $W$ using a four 192-dimensional MLP with a collection of 150K FLAME meshes by Gaussian sampling of the control parameters $\mathbf{p}$ . For each mesh, we sample 4K surface samples $N$ with surface normals $\mathbf{n}$ and ground-truth correspondence point $\bar{\mathbf{x}}^*$ , and 4K off surface samples $F$ distributed uniformly inside the box-bounded volume. Our canonical mesh has a neutral shape $\alpha = 0$ , expression $\beta = 0$ and neck pose, but with an opening jaw at 10 degrees. We close the face connectivity between the lips for a watertight geometry and also as a proxy geometry for the mouth cavity modeling. We open the jaw at the canonical space such that the spatial geometry and appearance inside the mouth can be modeled with distinguishable triplane features. In the second training stage, we freeze the weights of $W$ for best efficiency and train our model on the FFHQ [24], a human face dataset with 70,000 in-the-wild images. For each image, we estimate the camera pose $\mathbf{c}$ and FLAME parameters $\mathbf{p}$ with a nonlinear optimization in fitting 2D landmarks, assuming zero root and neck rotation and camera located 1.0m away from the world origin. We refer to the supplementary materials for more details. We rebalance the FFHQ dataset by duplicating images with large head poses and jaw opening expressions. We note that we also use pairs of images and $(\beta, \theta)$ for fine-tuning of a light-weight image encoder + +
ModelFID ↓KID ↓DSα ↑DSβ,θ ↑DSc ↑IS ↑
DiscoFaceGAN-SR [10]42.60.04-27.20.110.59
GAN-Control [47]9.80.005-43.70.280.62
HeadNeRF [21]1630.1681.338.90.250.678
AniFaceGAN-SR [58]30.50.0242.011.10.040.73
3DFaceShop [51]24.80.0180.4564.20.0290.594
Ours5.70.00163.171.90.350.80
+ +(a) Baseline comparisons. + +
ModelOurs wo LencOurs wo LpriorOurs
FID ↓6.06.15.7
KID ↓0.001720.001740.0016
ASD (cm) ↓0.1240.1350.124
var(α) ↓0.0005250.0005770.000515
AED (cm) ↓0.12330.08130.0797
var(β, θ) ↓0.0047070.004060.00404
+ +(b) Ablation study. + +Table 1. (a) Our method outperforms the prior 2D and 3D controllable image synthesis methods in both image quality and control disentanglement. (b) Our method demonstrates the best control accuracy over shape and expression with geometric prior $L_{prior}$ and self-supervised reconstruction loss $L_{enc}$ . + +$E$ , using a ResNet50 [19] backbone followed with a single-layer MLP. + +# 4.1. Qualitative Comparisons + +Controlled Portrait Synthesis. Our framework enables high-fidelity head synthesis with disentangled control over camera pose, shape and expression, as shown in Figure 1. We compare our method with prior controllable portrait synthesis works include DiscoFaceGAN [10], GAN-Control [47], HeadNeRF [21], AntiFaceGAN [58] and 3DFaceShop [51]. The first two methods integrates 3DMM controls into 2D face synthesis. HeadNeRF [21] is a 3D parametric NeRF-based head model built from both indoor multiview image capture datasets and in-the-wild monocular image collections. AntiFaceGAN [58] and 3DFaceShop [51] are concurrent 3D-aware controllable GANs in face synthesis. + +We qualitatively compare with prior work in Figure 5 for expression control. GAN-Control shows high-quality expression manipulation but with noticeable identity variation with differences in hair, head shape contour and background, whereas DiscoFaceGAN better maintains identity but with lower perceptual quality, and lacks dynamic details as expression varies. The expression control space of 3DFaceShop [51] is sensitive, suffering from appearance changes even with minor expression variation. AniFaceGAN demonstrates visually-consistent expression editing but with limited resolution and image quality, e.g., with blurry artifacts in hair and teeth. HeadNeRF ensures decent consistency in image generation by rendering the conditional NeRF but lacks fine details with very limited perceptual quality. In contrast, our approach produces the most compelling images with consistent appearance and dynamic details under expression changes. Moreover, controllable neck pose is a key factor towards realistic video avatar in applications like talking head synthesis. As shown in Figure. 1, our method achieves explicit control of neck and jaw poses which are seldom explored in prior work. Please refer to supplementary material for qualitative comparison in view consistency and shape editing. + +Dynamic details. Our method presents highly-consistent image synthesis in control of shape, expression and camera poses, but also depicts temporal realism in portrait animation, credited to the modeling of expression-dependent dynamic details. As shown in the last row of Figure 3, the appearance of wrinkles around the mouth and eyes when transiting from a neural expression to smiling largely enhances the animation realism. In comparison, without explicitly modeling the dynamic details, the wrinkles are embedded in the appearance and do not vary with expressions. + +Expressive Head Synthesis with Extrapolated Controls. AntiFaceGAN [58] and 3DFaceShop [51] depict dynamic details as well since their generation of the neural fields is conditioned on the input expression latent code. However, their designs result in shape and expression entanglement with the appearance generation, as reflected in the identity changes as shown in Figure. 5. By embedding the controls in a Gaussian latent space with such as Variation AutoEncoder (VAE), their approaches sacrifice expressiveness. The synthesized image quality is also highly correlated to the distribution of training images with target expressions. + +In contrast, our tri-plane generation is explicitly disentangled from shape and expression control. Moreover, the volume deformation is independently learnt from the deformed FLAME mesh collections which offer abundant 3D geometric knowledge with largely augmented control space. Therefore we are much less dependent on the distribution of the training images and support better extrapolation to unseen novel expressions. In Figure. 1, we show high-quality synthesized head with extreme jaw and neck articulated movements which do not exist in our training images. Our expression control is also more expressive, supporting subtle expressions like eye blinks (Figure. 1 3). + +# 4.2. Quantitative Comparisons + +Image Quality We measure the image quality with Frechet Inception Distance (FID) [20] and Kernel Inception Distance [5] between 50K randomly synthesized images and 50K randomly sampled real images at the resolution of $512 \times 512$ . Since DiscoFaceGAN [10] and AniFaceGAN [58] only synthesize images at $256 \times 256$ reso + +lution, we utilize a state-of-the-art super-resolution model, SwinIR [30] to upsample into $512 \times 512$ for a fair comparison. As shown in Table. 1a, our method is superior in both FID and KID than all prior work, demonstrating the most compelling image quality. We note that the original EG3D has a slightly lower FID at 4.8 which is expected since we introduce controllability with additional loss regularization. + +Disentanglement To evaluate the disentangled controllability of our model over shape, expression and camera pose, we measure the disentanglement score [10] of synthesized images as $\mathrm{DS}_{\alpha}$ , $\mathrm{DS}_{\beta,\theta}$ and $\mathrm{DS}_{\mathbf{c}}$ respectively. DS measures the stability of other factors when a single attribute is modified in image synthesis. We employ DECA [12] for estimation of FLAME parameters from generated images and calculate the variance of the estimated parameters $(\alpha, \{\beta, \theta\}, \mathbf{c})$ . Specifically the $\mathrm{DS}_i$ is calculated as + +$$ +D S _ {i} = \prod_ {j \neq i} \frac {\operatorname {v a r} (i)}{\operatorname {v a r} (j)}, i, j \in \{\boldsymbol {\alpha}, \{\boldsymbol {\beta}, \boldsymbol {\theta} \}, \mathbf {c} \}. \tag {8} +$$ + +Higher value of DS indicates better disentanglement. Additionally, we evaluate the identity similarity between pairs of synthesized images with random camera poses and expressions by calculating the cosine similarity of the face embeddings with a pre-trained face recognition module [26]. Our approach demonstrates the best disentanglement numerically over all prior work as indicated in Table. 1a. + +# 4.3. Ablation Studies + +We ablate the efficacy of the individual component by removing it from our full pipeline. As shown in Figure. 3, we show the loss of control accuracy over shape and expression respectively when removing the geometric prior loss $L_{prior}$ and control loss $L_{enc}$ . Conditioning the neural radiance field on expression is also critical to the modeling of dynamic details. Numerically we validate the efficacy of $L_{prior}$ and $L_{enc}$ with Average Shape Distance (ASD) and Average Expression Distance (AED). From the prior distribution, we randomly sample 500 shapes and expressions, with which we synthesize images with 10 different identities. We then reconstruct FLAME parameters from the synthesized images and compare against the input control. We compute 3D per-vertex $L_{1}$ distance of FLAME meshes for ASD while we use 3D landmarks $L_{1}$ distance for AED calculation. Additionally we compute the variance of estimated shapes and expressions within each control group, with lower value indicating more precise control. The efficacy of $L_{prior}$ and $L_{enc}$ is well evidenced in Table. 1b, with even slight image quality improvements. + +# 4.4. Applications + +Talking Head Video Generation. In Figure. 1, we showcase talking head video generation in controllable views + +driven with animation sequences of FLAME. Thanks to the high control accuracy, our method is able to synthesize various talking head videos with the same head movements and expressions performed by different identities. Our method is expressive in depicting both large articulated neck and jaw movements and subtle facial expressions like eye blinks, with rich dynamic details. Shape manipulation is easily achievable as well by modifying the shape parameters. Please refer to our supplementary materials for more results in high resolution. + +Portrait Image Manipulation and Animation. As illustrated in Figure. 1, our model also supports 3D-aware face reenactment of a single-view portrait to a video sequence. To achieve that, we perform an optimization in the latent $Z+$ space [24] to find the corresponding latent embedding, with FLAME parameter and camera pose estimated from the input portrait. With a frozen generator, the optimization is performed by measuring the similarity between generated image and real image using the $L_{2}$ loss and LPIPS loss [65]. For better reconstruction quality, we alter the parameters of the tri-plane synthesis module with a fixed optimized latent code [43]. After that one can explicitly manipulate the portrait with a preserved identity and in a different camera pose and expression. With the expression codes $(\beta, \theta)$ reconstructed from a video sequence, we are also able to reenact the portrait to the video motion. + +Societal Impact. Our work focuses on improving the controllability of 3D-aware GANs in technical aspects and is not specifically designed for any malicious uses. This being said, we do see that the method could be potentially extended into controversial applications such as generating fake videos. Therefore, we believe that the synthesized images and videos should present themselves as synthetic. + +# 5. Conclusion + +In this work, we introduce OmniAvatar, a novel 3D-aware generative model for synthesis of controllable high-fidelity human head. Our model achieves disentangled semantic control by factoring the generative process into 3D-aware canonical head synthesis and implicit volume deformation to target shapes and expressions. By learning a deformable signed distance function with canonical volumetric correspondence, we instill the geometric prior knowledge of a 3D statistical head model into the generation of neural radiance fields, enabling superior generative capability to novel shapes and expressions. Our approach demonstrates expressive and compelling talking head generation and portrait image animation with fine-grained control accuracy and temporal realism. We believe the proposed method presents an interesting direction for 3D avatar creation and animation, which sheds light on many potential downstream tasks. + +# References + +[1] Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas Guibas. Learning representations and generative models for 3d point clouds. In ICML, 2018. 2 +[2] Thiemo Alldieck, Hongyi Xu, and Cristian Sminchisescu. imghum: Implicit generative models of 3d human shape and articulated pose. In ICCV, pages 5461-5470, 2021. 4 +[3] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In iccv, pages 5855–5864, 2021. 3 +[4] Alexander W Bergman, Petr Kellnhofer, Yifan Wang, Eric R Chan, David B Lindell, and Gordon Wetzstein. Generative neural articulated radiance fields. arXiv preprint arXiv:2206.14314, 2022. 3, 5 +[5] Sai Bi, Kalyan Sunkavalli, Federico Perazzi, Eli Shechtman, Vladimir G Kim, and Ravi Ramamoorthi. Deep cg2real: Synthetic-to-real translation via image disentanglement. In iccv, pages 2730-2739, 2019. 7 +[6] Volker Blanz and Thomas Vetter. A morphable model for the synthesis of 3d faces. In Proceedings of the 26th annual conference on Computer graphics and interactive techniques, 1999. 2 +[7] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. CVPR, 2022. 2, 3, 4 +[8] Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5799-5809, 2021. 2 +[9] Anpei Chen, Ruiyang Liu, Ling Xie, Zhang Chen, Hao Su, and Jingyi Yu. Sofgan: A portrait image generator with dynamic styling. sigg, 41(1):1-26, 2022. 2 +[10] Yu Deng, Jiaolong Yang, Dong Chen, Fang Wen, and Xin Tong. Disentangled and controllable face image generation via 3d imitative-contrastive learning. In cvpr, pages 5154–5163, 2020. 2, 6, 7, 8 +[11] Yu Deng, Jiaolong Yang, Jianfeng Xiang, and Xin Tong. Gram: Generative radiance manifolds for 3d-aware image generation. CVPR, 2022. 2 +[12] Yao Feng, Haiwen Feng, Michael J. Black, and Timo Bolkart. Learning an animatable detailed 3D face model from in-the-wild images. volume 40, 2021. 8 +[13] Guy Gafni, Justus Thies, Michael Zollhofer, and Matthias Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8649-8658, 2021. 3 +[14] Partha Ghosh, Pravir Singh Gupta, Roy Uziel, Anurag Ranjan, Michael J Black, and Timo Bolkart. *Gif: Generative interpretable faces*. In 3dv, pages 868-878. IEEE, 2020. 2, 3 +[15] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and + +Yoshua Bengio. Generative adversarial nets. In NeurIPS, 2014. 1, 2 +[16] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. In ICML, pages 3569-3579. 2020. 4 +[17] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. Stylenerf: A style-based 3d-aware generator for high-resolution image synthesis. CVPR, 2022. 2 +[18] Yudong Guo, Keyu Chen, Sen Liang, Yong-Jin Liu, Hujun Bao, and Juyong Zhang. Ad-nerf: Audio driven neural radiance fields for talking head synthesis. In iccv, pages 5784-5794, 2021. 3 +[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In cvpr, pages 770-778, 2016. 7 +[20] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. NeurIPS, 2017. 7 +[21] Yang Hong, Bo Peng, Haiyao Xiao, Ligang Liu, and Juyong Zhang. Headnerf: A real-time nerf-based parametric head model. In cvpr, pages 20374-20384, 2022. 3, 6, 7 +[22] Kacper Kania, Kwang Moo Yi, Marek Kowalski, Tomasz Trzciński, and Andrea Tagliasacchi. Conerf: Controllable neural radiance fields. In cvpr, pages 18623-18632, 2022. 3 +[23] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. NeurIPS, 2021. 1 +[24] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, 2019. 1, 2, 6, 8 +[25] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. In CVPR, 2020. 1, 2 +[26] Minchul Kim, Anil K Jain, and Xiaoming Liu. Adaface: Quality adaptive margin for face recognition. In cvpr, pages 18750-18759, 2022. 8 +[27] Marek Kowalski, Stephan J Garbin, Virginia Estellers, Tadas Baltrusaitis, Matthew Johnson, and Jamie Shotton. Config: Controllable neural face image generation. In $eccv$ , pages 299-315. Springer, 2020. 2, 3 +[28] Ruihui Li, Xianzhi Li, Chi-Wing Fu, Daniel Cohen-Or, and Pheng-Ann Heng. Pu-gan: a point cloud upsampling adversarial network. In ICCV, 2019. 2 +[29] Tianye Li, Timo Bolkart, Michael. J. Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4D scans. ACM Transactions on Graphics, 36(6):194:1-194:17, 2017. 2, 3 +[30] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In iccv, pages 1833-1844, 2021. 8 +[31] Yiyi Liao, Katja Schwarz, Lars Mescheder, and Andreas Geiger. Towards unsupervised learning of generative models for 3D controllable image synthesis. In CVPR, 2020. 2 +[32] Lingjie Liu, Marc Habermann, Viktor Rudnev, Kripasindhu Sarkar, Jiatao Gu, and Christian Theobalt. Neural actor: Neural free-view synthesis of human actors with pose control. ACM Trans. on Graphics, 2021. 3 + +[33] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 2, 3 +[34] Thu Nguyen-Phuoc, Chuan Li, Lucas Theis, Christian Richardt, and Yong-Liang Yang. HoloGAN: Unsupervised learning of 3D representations from natural images. In ICCV, 2019. 2 +[35] Thu Nguyen-Phuoc, Christian Richardt, Long Mai, Yong-Liang Yang, and Niloy Mitra. BlockGAN: Learning 3D object-aware scene representations from unlabelled images. In NeurIPS, 2020. 2 +[36] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In CVPR, 2021. 2 +[37] Roy Or-El, Xuan Luo, Mengyi Shan, Eli Shechtman, Jeong Joon Park, and Ira Kemelmacher-Shlizerman. Stylesdf: High-resolution 3d-consistent image and geometry generation. CVPR, 2022. 2, 5 +[38] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In ICCV, 2021. 3 +[39] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. ACM Transactions on Graphics, 2022. 3 +[40] Pascal Paysan, Reinhard Knothe, Brian Amberg, Sami Romdhani, and Thomas Vetter. A 3d face model for pose and illumination invariant face recognition. In 2009 sixth IEEE international conference on advanced video and signal based surveillance, pages 296-301. IEEE, 2009. 2 +[41] Jingtan Piao, Keqiang Sun, Quan Wang, Kwan-Yee Lin, and Hongsheng Li. Inverting generative adversarial renderer for face reconstruction. In cvpr, pages 15619-15628, 2021. 2, 3 +[42] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In CVPR, 2021. 3 +[43] Daniel Roich, Ron Mokady, Amit H Bermano, and Daniel Cohen-Or. Pivotal tuning for latent-based editing of real images. ACM Transactions on Graphics (TOG), 42(1):1-13, 2022. 8 +[44] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. Graf: Generative radiance fields for 3d-aware image synthesis. NeurIPS, 2020. 2 +[45] Yichun Shi, Divyansh Aggarwal, and Anil K Jain. Lifting 2d stylegan for 3d-aware face generation. In cvpr, pages 6258-6266, 2021. 2 +[46] Yichun Shi, Xiao Yang, Yangyue Wan, and Xiaohui Shen. Semanticstylegan: Learning compositional generative priors for controllable image synthesis and editing. In CVPR, 2022. 2 +[47] Alon Shoshan, Nadav Bhonker, Igor Kviatkovsky, and Gerard Medioni. Gan-control: Explicitly controllable gans. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14083-14093, 2021. 6, 7 + +[48] Ivan Skorokhodov, Sergey Tulyakov, Yiqun Wang, and Peter Wonka. Epigraph: Rethinking training of 3d gans. arXiv preprint arXiv:2206.10535, 2022. 2 +[49] Keqiang Sun, Shangzhe Wu, Zhaoyang Huang, Ning Zhang, Quan Wang, and HongSheng Li. Controllable 3d face synthesis with conditional generative occupancy fields. arXiv preprint arXiv:2206.08361, 2022. 3 +[50] Attila Szabó, Givi Meishvili, and Paolo Favaro. Unsupervised generative 3D shape learning from natural images. arXiv, 2019. 2 +[51] Junshu Tang, Bo Zhang, Binxin Yang, Ting Zhang, Dong Chen, Lizhuang Ma, and Fang Wen. Explicitly controllable 3d-aware portrait generation. arXiv preprint arXiv:2209.05434, 2022. 3, 6, 7 +[52] Ayush Tewari, Mohamed Elgharib, Florian Bernard, Hans-Peter Seidel, Patrick Pérez, Michael Zollhöfer, and Christian Theobalt. Pie: Portrait image embedding for semantic control. sigg, 39(6):1-14, 2020. 2 +[53] Ayush Tewari, Mohamed Elgharib, Gaurav Bharaj, Florian Bernard, Hans-Peter Seidel, Patrick Pérez, Michael Zollhofer, and Christian Theobalt. Stylerig: Rigging stylegan for 3d control over portrait images. In cvpr, pages 6142-6151, 2020. 2, 3 +[54] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for stylegan image manipulation. ACM Trans. on Graphics, 2021. 2 +[55] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Christoph Lassner, and Christian Theobalt. Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In ICCV, 2021. 3 +[56] Ziyan Wang, Timur Bagautdinov, Stephen Lombardi, Tomas Simon, Jason Saragih, Jessica Hodgins, and Michael Zollhofer. Learning compositional radiance fields of dynamic human heads. In cvpr, pages 5704-5713, 2021. 3 +[57] Jiajun Wu, Chengkai Zhang, Tianfan Xue, William T. Freeman, and Joshua B. Tenenbaum. Learning a probabilistic latent space of object shapes via 3D generative-adversarial modeling. In NeurIPS, 2016. 2 +[58] Yue Wu, Yu Deng, Jiaolong Yang, Fangyun Wei, Qifeng Chen, and Xin Tong. Anifacegan: Animatable 3d-aware face image generation for video avatars. arXiv preprint arXiv:2210.06465, 2022. 3, 6, 7 +[59] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond. In Computer Graphics Forum, volume 41, pages 641-676. Wiley Online Library, 2022. 3 +[60] Hongyi Xu, Thiemo Alldieck, and Cristian Sminchisescu. H-nerf: Neural radiance fields for rendering and temporal reconstruction of humans in motion. NeurIPS, 2021. 3 +[61] Xudong Xu, Xingang Pan, Dahua Lin, and Bo Dai. Generative occupancy fields for 3d surface-aware image synthesis. NeurIPS, 34:20683-20695, 2021. 2 +[62] Yang Xue, Yuheng Li, Krishna Kumar Singh, and Yong Jae Lee. Giraffe hd: A high-resolution 3d-aware generative model. In CVPR, 2022. 2 + +[63] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. In NeurIPS, 2021. 5 +[64] Jianfeng Zhang, Zihang Jiang, Dingdong Yang, Hongyi Xu, Yichun Shi, Guoxian Song, Zhongcong Xu, Xinchao Wang, and Jiashi Feng. Avatargen: a 3d generative model for animatable human avatars. arXiv preprint arXiv:2208.00561, 2022.3 +[65] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In cvpr, pages 586-595, 2018. 8 +[66] Yufeng Zheng, Victoria Fernández Abrevaya, Marcel C Bühler, Xu Chen, Michael J Black, and Otmar Hilliges. Im avatar: Implicit morphable head avatars from videos. In cvpr, pages 13545-13555, 2022. 3 +[67] Peng Zhou, Lingxi Xie, Bingbing Ni, and Qi Tian. CIPS-3D: A 3D-Aware Generator of GANs Based on Conditionally-Independent Pixel Synthesis. arXiv, 2021. 2 +[68] Yiyu Zhuang, Hao Zhu, Xusen Sun, and Xun Cao. Mofan-erf: Morphable facial neural radiance field. arXiv preprint arXiv:2112.02308, 2021. 3 \ No newline at end of file diff --git a/2023/OmniAvatar_ Geometry-Guided Controllable 3D Head Synthesis/images.zip b/2023/OmniAvatar_ Geometry-Guided Controllable 3D Head Synthesis/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..98f43a81681d669a8ceab576fdd99591d74b14e4 --- /dev/null +++ b/2023/OmniAvatar_ Geometry-Guided Controllable 3D Head Synthesis/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6ae6b8774886af3f02ea5a3676e0f05916c58a7fde58f21810fc6b70692de607 +size 506028 diff --git a/2023/OmniAvatar_ Geometry-Guided Controllable 3D Head Synthesis/layout.json b/2023/OmniAvatar_ Geometry-Guided Controllable 3D Head Synthesis/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..528f932b874b2fffbe96a351eb8242f2bbcd2448 --- /dev/null +++ b/2023/OmniAvatar_ Geometry-Guided Controllable 3D Head Synthesis/layout.json @@ -0,0 +1,9222 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 96, + 103, + 498, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 103, + 498, + 121 + ], + "spans": [ + { + "bbox": [ + 96, + 103, + 498, + 121 + ], + "type": "text", + "content": "OmniAvatar: Geometry-Guided Controllable 3D Head Synthesis" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 91, + 141, + 501, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 141, + 501, + 214 + ], + "spans": [ + { + "bbox": [ + 91, + 141, + 501, + 214 + ], + "type": "text", + "content": "Hongyi Xu1 Guoxian Song1 Zihang Jiang1,2 Jianfeng Zhang1,2 Yichun Shi1 Jing Liu1 Wanchun Ma1 Jiashi Feng1 Linjie Luo1 \n1ByteDance Inc 2National University of Singapore \n{hongyixu, guoxian.song, zihang.jiang, jianfeng.zhang, yichun.shi, jing.liu, wanchun.ma, jshfeng, linjie.luo}@bytedance.com" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 48, + 240, + 548, + 422 + ], + "blocks": [ + { + "bbox": [ + 48, + 240, + 548, + 422 + ], + "lines": [ + { + "bbox": [ + 48, + 240, + 548, + 422 + ], + "spans": [ + { + "bbox": [ + 48, + 240, + 548, + 422 + ], + "type": "image", + "image_path": "09c331972bd1db073abd3f9aa8dca90f8d1112b24799daafd1a69425c8b01fb6.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 427, + 547, + 462 + ], + "lines": [ + { + "bbox": [ + 46, + 427, + 547, + 462 + ], + "spans": [ + { + "bbox": [ + 46, + 427, + 547, + 462 + ], + "type": "text", + "content": "Figure 1. Our model can synthesize diverse identity-preserved 3D heads with compelling dynamic details under full disentangled control over camera poses, facial expressions, head shapes, articulated neck and jaw poses (left). Our model can also reconstruct 3D heads from a single photo reference and enable multi-view-consistent head reenactment (right)." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 471, + 192, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 471, + 192, + 483 + ], + "spans": [ + { + "bbox": [ + 143, + 471, + 192, + 483 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 45, + 498, + 290, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 498, + 290, + 715 + ], + "spans": [ + { + "bbox": [ + 45, + 498, + 290, + 715 + ], + "type": "text", + "content": "We present OmniAvatar, a novel geometry-guided 3D head synthesis model trained from in-the-wild unstructured images that is capable of synthesizing diverse identity-preserved 3D heads with compelling dynamic details under full disentangled control over camera poses, facial expressions, head shapes, articulated neck and jaw poses. To achieve such high level of disentangled control, we first explicitly define a novel semantic signed distance function (SDF) around a head geometry (FLAME) conditioned on the control parameters. This semantic SDF allows us to build a differentiable volumetric correspondence map from the observation space to a disentangled canonical space from all the control parameters. We then leverage the 3D-aware GAN framework (EG3D) to synthesize detailed shape and appearance of 3D full heads in the canonical space, followed by a volume rendering step guided by the volumetric correspondence map to output into the observation space. To ensure the control accuracy on the synthesized head" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 472, + 547, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 547, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 547, + 581 + ], + "type": "text", + "content": "shapes and expressions, we introduce a geometry prior loss to conform to head SDF and a control loss to conform to the expression code. Further, we enhance the temporal realism with dynamic details conditioned upon varying expressions and joint poses. Our model can synthesize more preferable identity-preserved 3D heads with compelling dynamic details compared to the state-of-the-art methods both qualitatively and quantitatively. We also provide an ablation study to justify many of our system design choices." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 608, + 387, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 608, + 387, + 620 + ], + "spans": [ + { + "bbox": [ + 306, + 608, + 387, + 620 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": "Photo-realistic face image synthesis, editing and animation attract significant interests in computer vision and graphics, with a wide range of important downstream applications in visual effects, digital avatars, telepresence and many others. With the advent of Generative Adversarial Networks (GANs) [15], remarkable progress has been achieved in face image synthesis by StyleGAN [23-25]" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12814" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "type": "text", + "content": "as well as in semantic and style editing for face images [46, 54]. To manipulate and animate the expressions and poses in face images, many methods attempted to leverage 3D parametric face models, such as 3D Morphable Models (3DMMs) [6, 40], with StyleGAN-based synthesis models [10, 41, 53]. However, all these methods operate on 2D convolutional networks (CNNs) without explicitly enforcing the underlying 3D face structure. Therefore they cannot strictly maintain the 3D consistency when synthesizing faces under different poses and expressions." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 194, + 289, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 194, + 289, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 194, + 289, + 361 + ], + "type": "text", + "content": "Recently, a line of work has explored neural 3D representations by unsupervised training of 3D-aware GANs from in-the-wild unstructured images [7, 8, 11, 17, 37, 44, 45, 48, 61, 62, 67]. Among them, methods with generative Neural Radiance Fields (NeRFs) [33] have demonstrated striking quality and multi-view-consistent image synthesis [7, 11, 17, 37, 48]. The progress is largely due to the integration of the power of StyleGAN in photo-realistic image synthesis and NeRF representation in 3D scene modeling with view-consistent volumetric rendering. Nevertheless, these methods lack precise 3D control over the generated faces beyond camera pose, as well as the quality and consistency in control over other attributes, such as shape, expression, neck and jaw pose, leave much to be desired." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 364, + 289, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 364, + 289, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 364, + 289, + 616 + ], + "type": "text", + "content": "In this work, we present OmniAvatar, a novel geometry-guided 3D head synthesis model trained from in-the-wild unstructured images. Our model can synthesize a wide range of 3D human heads with full control over camera poses, facial expressions, head shapes, articulated neck and jaw poses. To achieve such high level of disentangled control for 3D human head synthesis, we devise our model learning in two stages. We first define a novel semantic signed distance function (SDF) around a head geometry (i.e. FLAME [29]) conditioned on its control parameters. This semantic SDF fully distills rich 3D geometric prior knowledge from the statistical FLAME model and allows us to build a differentiable volumetric correspondence map from the observation space to a disentangled canonical space from all the control parameters. In the second training stage, we then leverage the state-of-the-art 3D GAN framework (EG3D [7]) to synthesize realistic shape and appearance of 3D human heads in the canonical space, including the modeling of hair and apparels. Following that, a volume rendering step is guided by the volumetric correspondence map to output the geometry and image in the observation space." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 289, + 715 + ], + "type": "text", + "content": "To ensure the consistency of synthesized 3D head shape with controlling head geometry, we introduce a geometry prior loss to minimize the difference between the synthesized neural density field and the FLAME head SDF in observation space. Furthermore, to improve the control accuracy, we pre-train an image encoder of the control parameters and formulate a control loss to ensure synthesized images matching the input control code upon encoding. An-" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 133 + ], + "type": "text", + "content": "other key aspect of synthesis realism is dynamic details such as wrinkles and varying shading as subjects change expressions and poses. To synthesize dynamic details, we propose to condition EG3D's triplane feature decoding with noised controlling expression." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 133, + 547, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 133, + 547, + 216 + ], + "spans": [ + { + "bbox": [ + 304, + 133, + 547, + 216 + ], + "type": "text", + "content": "Compare to state-of-the-art methods, our method achieves superior synthesized image quality in terms of Frechet Inception Distance (FID) and Kernel Inception Distance (KID). Our method can consistently preserve the identity of synthesized subjects with compelling dynamic details while changing expressions and poses, outperforming prior methods both quantitatively and qualitatively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 317, + 216, + 534, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 216, + 534, + 228 + ], + "spans": [ + { + "bbox": [ + 317, + 216, + 534, + 228 + ], + "type": "text", + "content": "The contributions of our work can be summarized as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 229, + 545, + 378 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 317, + 229, + 545, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 229, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 317, + 229, + 545, + 277 + ], + "type": "text", + "content": "- A novel geometry-guided 3D GAN framework for high-quality 3D head synthesis with full control on camera poses, facial expressions, head shapes, articulated neck and jaw poses." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 317, + 278, + 545, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 278, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 317, + 278, + 545, + 327 + ], + "type": "text", + "content": "- A novel semantic SDF formulation that defines the volumetric correspondence map from observation space to canonical space and allows full disentanglement of control parameters in 3D GAN training." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 317, + 328, + 545, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 328, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 317, + 328, + 545, + 353 + ], + "type": "text", + "content": "- A geometric prior loss and a control loss to ensure the head shape and expression synthesis accuracy." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 317, + 354, + 545, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 354, + 545, + 378 + ], + "spans": [ + { + "bbox": [ + 317, + 354, + 545, + 378 + ], + "type": "text", + "content": "- A robust noised expression conditioning scheme to enable dynamic detail synthesis." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 388, + 392, + 401 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 388, + 392, + 401 + ], + "spans": [ + { + "bbox": [ + 306, + 388, + 392, + 401 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 408, + 547, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 408, + 547, + 624 + ], + "spans": [ + { + "bbox": [ + 304, + 408, + 547, + 624 + ], + "type": "text", + "content": "3D-Aware Generative Image Synthesis. Generative adversarial networks [15, 24, 25] gained popularity over the last decade due to their remarkable ability in photo-realistic image synthesis. Building on the success of 2D image-based GANs, recent works have extended the capabilities to view-consistent image synthesis with unsupervised learning from 2D single-view images. The key idea is to combine differential rendering with 3D scene representations, such as meshes [31, 50], point clouds [1, 28], voxels [34, 35, 57], and recently implicit neural representation [7, 8, 11, 17, 36, 37, 44, 48, 67]. We build our work on recent 3D GAN model by Chan et al [7] that uses an efficient triplane-based NeRF generation, combined with 2D CNN-based super-resolution. Even though 3D-aware GANs are able to control camera viewpoints, they lack precise 3D control over the other attributes such as shapes and expressions. In this work, we empower 3D-aware GANs with disentangled precise control over shapes and expressions." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 547, + 715 + ], + "type": "text", + "content": "Controllable Face Image Synthesis. Considerable work [9, 10, 14, 27, 41, 52, 53] has been devoted to incorporate 3D priors of statistical face models, such as 3D Morphable Models (3DMMs) [6, 40], in controllable face synthesis and animation. Among them, DiscoFaceGAN [10] proposed imitative-contrastive learning to mimic the 3DMM rendering process by the generative model." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12815" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 547, + 273 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 547, + 273 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 547, + 273 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 547, + 273 + ], + "type": "image", + "image_path": "b04fe5cf49739ad18a3cd42dd0bec0dfdd222f7a0ed9e98008660542263f19a0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 280, + 548, + 381 + ], + "lines": [ + { + "bbox": [ + 46, + 280, + 548, + 381 + ], + "spans": [ + { + "bbox": [ + 46, + 280, + 548, + 381 + ], + "type": "text", + "content": "Figure 2. Overview of our training framework. Stage 1: Trained from parameterized FLAME [29] mesh collections, a MLP-network " + }, + { + "bbox": [ + 46, + 280, + 548, + 381 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 46, + 280, + 548, + 381 + ], + "type": "text", + "content": " maps a shape " + }, + { + "bbox": [ + 46, + 280, + 548, + 381 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 280, + 548, + 381 + ], + "type": "text", + "content": ", expression " + }, + { + "bbox": [ + 46, + 280, + 548, + 381 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 280, + 548, + 381 + ], + "type": "text", + "content": " and articulated jaw and neck pose " + }, + { + "bbox": [ + 46, + 280, + 548, + 381 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 280, + 548, + 381 + ], + "type": "text", + "content": " into 3D point-to-point volumetric correspondences from observation to canonical space, together with a signed distance function of the corresponding FLAME head. Stage 2: Given a Gaussian latent code " + }, + { + "bbox": [ + 46, + 280, + 548, + 381 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 46, + 280, + 548, + 381 + ], + "type": "text", + "content": ", our model generates a tri-plane represented 3D feature space of a canonical head, disentangled with shape and expression controls. The volume rendering is then guided by the volumetric correspondence field to map the decoded neural radiance field from the canonical to observation space. We condition the NeRF decoding with expression and joint pose for modeling dynamic details. A super-resolution module synthesizes the final high-resolution RGB image from the volume-rendered feature map. For fine-grained shape and expression control, we apply the FLAME SDF as geometric prior to the synthesized NeRF density, and self-supervise the image synthesis to commit to the target expression " + }, + { + "bbox": [ + 46, + 280, + 548, + 381 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 280, + 548, + 381 + ], + "type": "text", + "content": " and joint pose " + }, + { + "bbox": [ + 46, + 280, + 548, + 381 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 280, + 548, + 381 + ], + "type": "text", + "content": " by comparing the input code against the re-estimated values " + }, + { + "bbox": [ + 46, + 280, + 548, + 381 + ], + "type": "inline_equation", + "content": "\\hat{\\beta}, \\hat{\\theta}" + }, + { + "bbox": [ + 46, + 280, + 548, + 381 + ], + "type": "text", + "content": " from synthesized images." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 393, + 289, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 393, + 289, + 586 + ], + "spans": [ + { + "bbox": [ + 46, + 393, + 289, + 586 + ], + "type": "text", + "content": "A similar strategy has also been adopted with concurrent and follow-up works [14, 27, 41, 53]. However all of these approaches suffer from 3D inconsistency due to the use of 2D CNNs as image renderer. HeadNeRF [21] combines 3DMM with 3D NeRF representation, and is able to synthesize 3D heads conditioned on 3DMM attributes. However, the training relies on annotated multiview datasets whereas our approach learns the disentangled 3D head synthesis with only single-view images. There has been concurrent work [4, 49, 51, 58, 64] to ours in 3D-aware controllable face or full-body GANs. Differently from these approaches, we use a full-head parametric model FLAME [29], and fully exploit the spatial geometric prior knowledge beyond the surface deformation and skinning. We have also achieved fine-grained control with our novel losses and enhanced face animation with rich dynamic details." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": "Controllable Neural Implicit Field of Face. Neural implicit functions [59], have emerged as a powerful continuous and differential representations of 3D scenes. Among them, Neural Radiance Field [3, 33] has been widely adopted due to its superiority in modeling complex scene details and synthesizing multiview images with inherited 3D consistency. While initial proposals have focused on static scene modeling, recent work has successfully demonstrated application of NeRF in modeling dynamic scenes [32,38,42,55,60]. In particular, dynamic neural radi-" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 393, + 547, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 393, + 547, + 491 + ], + "spans": [ + { + "bbox": [ + 304, + 393, + 547, + 491 + ], + "type": "text", + "content": "ance fields of human heads [13,18,22,38,39,56,66,68] have enabled photo-realistic head animation, often by conditioning with pose parameters or deforming the radiance field with 3D morphable models. However, they do not leverage a generative training paradigm and focus on scene-specific learning from video sequences or multiview images. In contrast, our model learns generative and controllable neural radiance fields from widely accessible single-view images." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 500, + 362, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 500, + 362, + 513 + ], + "spans": [ + { + "bbox": [ + 306, + 500, + 362, + 513 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 521, + 547, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 547, + 568 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 547, + 568 + ], + "type": "text", + "content": "Our goal is to build a geometry-guided 3D head synthesis model with full control of camera poses, face shapes and expressions, trained from in-the-wild unstructured image collections." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 577, + 373, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 577, + 373, + 588 + ], + "spans": [ + { + "bbox": [ + 306, + 577, + 373, + 588 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 595, + 547, + 716 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 595, + 547, + 716 + ], + "spans": [ + { + "bbox": [ + 304, + 595, + 547, + 716 + ], + "type": "text", + "content": "Problem. To achieve our goal, we leverage a 3D-aware generator (EG3D [7]) for photo-realistic, multiview consistent image synthesis, while disentangle control of head geometric attributes from image generation with a 3D statistical head model (FLAME [29]). Specifically, given a random Gaussian-sampled latent code " + }, + { + "bbox": [ + 304, + 595, + 547, + 716 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 304, + 595, + 547, + 716 + ], + "type": "text", + "content": ", a camera pose " + }, + { + "bbox": [ + 304, + 595, + 547, + 716 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 304, + 595, + 547, + 716 + ], + "type": "text", + "content": " and a FLAME parameter " + }, + { + "bbox": [ + 304, + 595, + 547, + 716 + ], + "type": "inline_equation", + "content": "\\mathbf{p} = (\\alpha, \\beta, \\theta)" + }, + { + "bbox": [ + 304, + 595, + 547, + 716 + ], + "type": "text", + "content": " consisting of shape " + }, + { + "bbox": [ + 304, + 595, + 547, + 716 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 595, + 547, + 716 + ], + "type": "text", + "content": ", expression " + }, + { + "bbox": [ + 304, + 595, + 547, + 716 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 304, + 595, + 547, + 716 + ], + "type": "text", + "content": ", jaw and neck pose " + }, + { + "bbox": [ + 304, + 595, + 547, + 716 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 595, + 547, + 716 + ], + "type": "text", + "content": ", the generator " + }, + { + "bbox": [ + 304, + 595, + 547, + 716 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 304, + 595, + 547, + 716 + ], + "type": "text", + "content": " synthesizes a photo-realistic human head image " + }, + { + "bbox": [ + 304, + 595, + 547, + 716 + ], + "type": "inline_equation", + "content": "I_{RGB}(\\mathbf{z}|\\mathbf{c},\\mathbf{p})" + }, + { + "bbox": [ + 304, + 595, + 547, + 716 + ], + "type": "text", + "content": " with corresponding attributes as defined in " + }, + { + "bbox": [ + 304, + 595, + 547, + 716 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 304, + 595, + 547, + 716 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12816" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 277 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 277 + ], + "type": "text", + "content": "**Framework.** As illustrated in our pipeline 2, our controllable 3D-aware GAN is trained in two stages. From a large collection of 3D deformed FLAME meshes, we first pre-train a deformable semantic SDF around the FLAME geometry that builds a differential volumetric correspondence map from the observation to a predefined canonical space (Section 3.2.1). In the second stage, guided by the pre-trained volumetric mapping, we then deform the detailed 3D full heads synthesized in the disentangled canonical space to the desired shapes and expressions (Section 3.2.2). Fine-grained expression control is achieved by supervising image synthesis such that expressions estimated from the generated images is consistent with the input control (Section 3.2.3). Our approach further enhances temporal realism with dynamic details, such as dimples and wrinkles, synthesizing realistic shading and geometric variations as expression changes (Section 3.2.4)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 283, + 289, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 283, + 289, + 439 + ], + "spans": [ + { + "bbox": [ + 46, + 283, + 289, + 439 + ], + "type": "text", + "content": "3D-Aware GAN Background. To ensure appearance consistency from different views, we choose EG3D [7] as our backbone for 3D-aware image synthesis. The generator " + }, + { + "bbox": [ + 46, + 283, + 289, + 439 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 46, + 283, + 289, + 439 + ], + "type": "text", + "content": " takes a random latent code " + }, + { + "bbox": [ + 46, + 283, + 289, + 439 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 46, + 283, + 289, + 439 + ], + "type": "text", + "content": " and conditioning camera label " + }, + { + "bbox": [ + 46, + 283, + 289, + 439 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{c}}" + }, + { + "bbox": [ + 46, + 283, + 289, + 439 + ], + "type": "text", + "content": ", and maps to a manifold of triplane features " + }, + { + "bbox": [ + 46, + 283, + 289, + 439 + ], + "type": "inline_equation", + "content": "\\mathcal{M}(\\mathbf{z},\\hat{\\mathbf{c}})" + }, + { + "bbox": [ + 46, + 283, + 289, + 439 + ], + "type": "text", + "content": ". For presentation clarity, we absorb " + }, + { + "bbox": [ + 46, + 283, + 289, + 439 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{c}}" + }, + { + "bbox": [ + 46, + 283, + 289, + 439 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 283, + 289, + 439 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 46, + 283, + 289, + 439 + ], + "type": "text", + "content": " and simply denote triplane as " + }, + { + "bbox": [ + 46, + 283, + 289, + 439 + ], + "type": "inline_equation", + "content": "\\mathcal{M}(\\mathbf{z})" + }, + { + "bbox": [ + 46, + 283, + 289, + 439 + ], + "type": "text", + "content": ". A low-resolution feature map " + }, + { + "bbox": [ + 46, + 283, + 289, + 439 + ], + "type": "inline_equation", + "content": "I^{*}(\\mathbf{z}|\\mathbf{c})" + }, + { + "bbox": [ + 46, + 283, + 289, + 439 + ], + "type": "text", + "content": " is then rendered from a desired camera pose " + }, + { + "bbox": [ + 46, + 283, + 289, + 439 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 46, + 283, + 289, + 439 + ], + "type": "text", + "content": " by sampling the triplane features and integrating MLP-decoded neural radiance " + }, + { + "bbox": [ + 46, + 283, + 289, + 439 + ], + "type": "inline_equation", + "content": "(\\sigma ,\\mathbf{f})" + }, + { + "bbox": [ + 46, + 283, + 289, + 439 + ], + "type": "text", + "content": " along camera rays. A super-resolution module is followed to modulate the feature map and synthesize the final RGB images at high resolution. We train " + }, + { + "bbox": [ + 46, + 283, + 289, + 439 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 46, + 283, + 289, + 439 + ], + "type": "text", + "content": " and a dual discriminator " + }, + { + "bbox": [ + 46, + 283, + 289, + 439 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 283, + 289, + 439 + ], + "type": "text", + "content": " with adversarial training." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 445, + 257, + 459 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 445, + 257, + 459 + ], + "spans": [ + { + "bbox": [ + 47, + 445, + 257, + 459 + ], + "type": "text", + "content": "3.2. Controllable 3D-Aware Image Synthesis" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 464, + 288, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 464, + 288, + 644 + ], + "spans": [ + { + "bbox": [ + 46, + 464, + 288, + 644 + ], + "type": "text", + "content": "To synthesize an image " + }, + { + "bbox": [ + 46, + 464, + 288, + 644 + ], + "type": "inline_equation", + "content": "\\mathbf{I}(\\mathbf{z}|\\mathbf{c},\\mathbf{p})" + }, + { + "bbox": [ + 46, + 464, + 288, + 644 + ], + "type": "text", + "content": " with desired FLAME control " + }, + { + "bbox": [ + 46, + 464, + 288, + 644 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 46, + 464, + 288, + 644 + ], + "type": "text", + "content": ", we leverage the EG3D framework to generate a triplane-based 3D volumetric feature space " + }, + { + "bbox": [ + 46, + 464, + 288, + 644 + ], + "type": "inline_equation", + "content": "\\mathcal{M}(\\mathbf{z})" + }, + { + "bbox": [ + 46, + 464, + 288, + 644 + ], + "type": "text", + "content": " of a synthesized identity with canonical shape and expression. Guided by our pretrained volumetric correspondence map, we deform the synthesized feature volume into our target observation space, which is further decoded and volume rendered into high-fidelity head appearance and geometry with the target shape and expression. Our design explicitly disentangles the underlying geometric variations of changing shape and expression from canonical geometry and appearance synthesis. Following EG3D [7], we associate each training image with a set of camera parameters " + }, + { + "bbox": [ + 46, + 464, + 288, + 644 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 46, + 464, + 288, + 644 + ], + "type": "text", + "content": " and control parameters " + }, + { + "bbox": [ + 46, + 464, + 288, + 644 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 46, + 464, + 288, + 644 + ], + "type": "text", + "content": ", which are obtained from a nonlinear 2D landmarks-based optimization." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 658, + 231, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 231, + 671 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 231, + 671 + ], + "type": "text", + "content": "3.2.1 Semantic Signed Distance Function" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "content": "For disentangled geometric modeling, we formulate an implicit semantic SDF representation " + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "W(\\mathbf{x}|\\mathbf{p} = (\\alpha ,\\beta ,\\theta)) = (s,\\bar{\\mathbf{x}})" + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\alpha ,\\beta" + }, + { + "bbox": [ + 47, + 677, + 287, + 714 + ], + "type": "text", + "content": " are the linear shape and expression" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 306, + 70, + 558, + 277 + ], + "blocks": [ + { + "bbox": [ + 306, + 70, + 558, + 277 + ], + "lines": [ + { + "bbox": [ + 306, + 70, + 558, + 277 + ], + "spans": [ + { + "bbox": [ + 306, + 70, + 558, + 277 + ], + "type": "image", + "image_path": "5375ab76ac4b310a1596b83d56b7d77f1d6b31faff19c0e4c5b218a5c3d05ea5.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 285, + 545, + 352 + ], + "lines": [ + { + "bbox": [ + 304, + 285, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 304, + 285, + 545, + 352 + ], + "type": "text", + "content": "Figure 3. We synthesize different identities with the same shape and expression code respectively in (a) and (b). We observe less shape variation with our geometric prior loss (a) and more consistent expression with our control loss (b). In (c), we show our expression-conditioned NeRF modeling builds rich dynamic details as the subject varies expressions." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 359, + 545, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 359, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 304, + 359, + 545, + 491 + ], + "type": "text", + "content": "blendshape coefficients, and " + }, + { + "bbox": [ + 304, + 359, + 545, + 491 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 359, + 545, + 491 + ], + "type": "text", + "content": " controls the rotation of a 3-DoF jaw and neck joint. Specifically, given a spatial point " + }, + { + "bbox": [ + 304, + 359, + 545, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 359, + 545, + 491 + ], + "type": "text", + "content": " in observation space " + }, + { + "bbox": [ + 304, + 359, + 545, + 491 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(\\mathbf{p})" + }, + { + "bbox": [ + 304, + 359, + 545, + 491 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 359, + 545, + 491 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 304, + 359, + 545, + 491 + ], + "type": "text", + "content": " returns its 3D correspondence point " + }, + { + "bbox": [ + 304, + 359, + 545, + 491 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{x}}" + }, + { + "bbox": [ + 304, + 359, + 545, + 491 + ], + "type": "text", + "content": " (i.e., semantics) in canonical space " + }, + { + "bbox": [ + 304, + 359, + 545, + 491 + ], + "type": "inline_equation", + "content": "\\mathcal{C}(\\bar{\\mathbf{p}})" + }, + { + "bbox": [ + 304, + 359, + 545, + 491 + ], + "type": "text", + "content": ", with which we project and query the triplane features " + }, + { + "bbox": [ + 304, + 359, + 545, + 491 + ], + "type": "inline_equation", + "content": "\\mathcal{M}(\\mathbf{z})" + }, + { + "bbox": [ + 304, + 359, + 545, + 491 + ], + "type": "text", + "content": ". Additionally it also computes the closest signed distance " + }, + { + "bbox": [ + 304, + 359, + 545, + 491 + ], + "type": "inline_equation", + "content": "s(\\mathbf{x}|\\mathbf{p})" + }, + { + "bbox": [ + 304, + 359, + 545, + 491 + ], + "type": "text", + "content": " to the FLAME mesh surface " + }, + { + "bbox": [ + 304, + 359, + 545, + 491 + ], + "type": "inline_equation", + "content": "\\mathbf{S}(\\mathbf{p})" + }, + { + "bbox": [ + 304, + 359, + 545, + 491 + ], + "type": "text", + "content": ". We illustrate the function in Figure 4. We co-learn the highly-correlated volumetric correspondence and SDF with the property that the signed distance is preserved between canonical and observation correspondence points as " + }, + { + "bbox": [ + 304, + 359, + 545, + 491 + ], + "type": "inline_equation", + "content": "s(\\bar{\\mathbf{x}}|\\bar{\\mathbf{p}}) = s(\\mathbf{x}|\\mathbf{p})" + }, + { + "bbox": [ + 304, + 359, + 545, + 491 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 491, + 545, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 491, + 545, + 539 + ], + "spans": [ + { + "bbox": [ + 304, + 491, + 545, + 539 + ], + "type": "text", + "content": "We learn " + }, + { + "bbox": [ + 304, + 491, + 545, + 539 + ], + "type": "inline_equation", + "content": "W(\\mathbf{x}|\\mathbf{p})" + }, + { + "bbox": [ + 304, + 491, + 545, + 539 + ], + "type": "text", + "content": " with a large corpus of 3D FLAME meshes " + }, + { + "bbox": [ + 304, + 491, + 545, + 539 + ], + "type": "inline_equation", + "content": "\\mathbf{S}(\\mathbf{p})" + }, + { + "bbox": [ + 304, + 491, + 545, + 539 + ], + "type": "text", + "content": " sampled from its parametric control space. Similar to IGR [16] and imGHUM [2], we model our implicit field as an MLP and optimize " + }, + { + "bbox": [ + 304, + 491, + 545, + 539 + ], + "type": "inline_equation", + "content": "W(\\mathbf{x}|\\mathbf{p})" + }, + { + "bbox": [ + 304, + 491, + 545, + 539 + ], + "type": "text", + "content": " with losses," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 312, + 555, + 545, + 585 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 555, + 545, + 585 + ], + "spans": [ + { + "bbox": [ + 312, + 555, + 545, + 585 + ], + "type": "interline_equation", + "content": "L _ {i s o} = \\frac {1}{| N |} \\sum_ {\\mathbf {x} \\in N} (| s (\\mathbf {x} | \\mathbf {p}) | + | \\nabla s _ {\\mathbf {x}} (\\mathbf {x} | \\mathbf {p}) - \\mathbf {n} (\\mathbf {x} | \\mathbf {p}) |), \\tag {1}", + "image_path": "c39acae58049e7d91a2d6fa58ea319e5bcd7f70f6576e837f74aa5e072e1d05f.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 353, + 586, + 545, + 615 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 586, + 545, + 615 + ], + "spans": [ + { + "bbox": [ + 353, + 586, + 545, + 615 + ], + "type": "interline_equation", + "content": "L _ {e i k} = \\frac {1}{| F |} \\sum_ {\\mathbf {x} \\in F} \\| \\nabla s _ {\\mathbf {x}} (\\mathbf {x} | \\mathbf {p}) - 1 \\| _ {2}, \\tag {2}", + "image_path": "ea828f737203f906d1b3e4b46b7335962feb6ae815ace97c429143f0deabd6e2.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 345, + 617, + 545, + 646 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 617, + 545, + 646 + ], + "spans": [ + { + "bbox": [ + 345, + 617, + 545, + 646 + ], + "type": "interline_equation", + "content": "L _ {s e m} = \\frac {1}{| N |} \\sum_ {\\mathbf {x} \\in N} \\left(| \\bar {\\mathbf {x}} (\\mathbf {x} | \\mathbf {p}) - \\bar {\\mathbf {x}} ^ {*} (\\mathbf {x} | \\bar {\\mathbf {p}}) |\\right) \\tag {3}", + "image_path": "96de333e0e722a9f0095c7ecb87e03c5f42a2c8e03cb3b52cca3d097a585efa3.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 653, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 653, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 653, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 653, + 545, + 713 + ], + "type": "inline_equation", + "content": "N,F" + }, + { + "bbox": [ + 306, + 653, + 545, + 713 + ], + "type": "text", + "content": " are a batch of on and off surface samples. For the surface samples, the " + }, + { + "bbox": [ + 306, + 653, + 545, + 713 + ], + "type": "inline_equation", + "content": "L_{iso}" + }, + { + "bbox": [ + 306, + 653, + 545, + 713 + ], + "type": "text", + "content": " encourages the signed distance values to be on the zero-level-set and the SDF gradient to be equal to the given surface normals " + }, + { + "bbox": [ + 306, + 653, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 306, + 653, + 545, + 713 + ], + "type": "text", + "content": ". The Eikonal loss " + }, + { + "bbox": [ + 306, + 653, + 545, + 713 + ], + "type": "inline_equation", + "content": "L_{eik}" + }, + { + "bbox": [ + 306, + 653, + 545, + 713 + ], + "type": "text", + "content": " is derived from [16] where the SDF is differen" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12817" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 94, + 70, + 246, + 196 + ], + "blocks": [ + { + "bbox": [ + 94, + 70, + 246, + 196 + ], + "lines": [ + { + "bbox": [ + 94, + 70, + 246, + 196 + ], + "spans": [ + { + "bbox": [ + 94, + 70, + 246, + 196 + ], + "type": "image", + "image_path": "637a25f6e637a9c0a2a20adac5ac22048819dac91d8fd54ab84b4a827ea6aa24.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 80, + 204, + 253, + 216 + ], + "lines": [ + { + "bbox": [ + 80, + 204, + 253, + 216 + ], + "spans": [ + { + "bbox": [ + 80, + 204, + 253, + 216 + ], + "type": "text", + "content": "Figure 4. Illustration to semantic SDF learning." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 224, + 287, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 224, + 287, + 320 + ], + "spans": [ + { + "bbox": [ + 46, + 224, + 287, + 320 + ], + "type": "text", + "content": "tiable everywhere with gradient norm 1. The semantic loss " + }, + { + "bbox": [ + 46, + 224, + 287, + 320 + ], + "type": "inline_equation", + "content": "L_{sem}" + }, + { + "bbox": [ + 46, + 224, + 287, + 320 + ], + "type": "text", + "content": " supervises the mapping of surface samples " + }, + { + "bbox": [ + 46, + 224, + 287, + 320 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in F" + }, + { + "bbox": [ + 46, + 224, + 287, + 320 + ], + "type": "text", + "content": " to the ground-truth correspondence points " + }, + { + "bbox": [ + 46, + 224, + 287, + 320 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{x}}^*" + }, + { + "bbox": [ + 46, + 224, + 287, + 320 + ], + "type": "text", + "content": " on the canonical FLAME surface, where " + }, + { + "bbox": [ + 46, + 224, + 287, + 320 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{x}}^*" + }, + { + "bbox": [ + 46, + 224, + 287, + 320 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 224, + 287, + 320 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 46, + 224, + 287, + 320 + ], + "type": "text", + "content": " share the same barycentric coordinates. The " + }, + { + "bbox": [ + 46, + 224, + 287, + 320 + ], + "type": "inline_equation", + "content": "L_{eik}" + }, + { + "bbox": [ + 46, + 224, + 287, + 320 + ], + "type": "text", + "content": " provides a geometric regularization over the volumetric SDF, whereas the " + }, + { + "bbox": [ + 46, + 224, + 287, + 320 + ], + "type": "inline_equation", + "content": "L_{iso}" + }, + { + "bbox": [ + 46, + 224, + 287, + 320 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 224, + 287, + 320 + ], + "type": "inline_equation", + "content": "L_{sem}" + }, + { + "bbox": [ + 46, + 224, + 287, + 320 + ], + "type": "text", + "content": " act as the boundary condition to the SDF and volumetric correspondence field respectively." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 321, + 288, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 321, + 288, + 573 + ], + "spans": [ + { + "bbox": [ + 46, + 321, + 288, + 573 + ], + "type": "text", + "content": "To guide the canonical correspondence learning for off-the-surface points, one could try to minimize the signed distance difference between " + }, + { + "bbox": [ + 46, + 321, + 288, + 573 + ], + "type": "inline_equation", + "content": "s(\\bar{\\mathbf{x}}|\\bar{\\mathbf{p}})" + }, + { + "bbox": [ + 46, + 321, + 288, + 573 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 321, + 288, + 573 + ], + "type": "inline_equation", + "content": "s(\\mathbf{x}|\\mathbf{p})" + }, + { + "bbox": [ + 46, + 321, + 288, + 573 + ], + "type": "text", + "content": ". However, we note that the volume correspondence between the observation and canonical space is still ill-regularized, considering infinite number of points exist with the same signed distance value. We therefore reformulate volumetric correspondence field as " + }, + { + "bbox": [ + 46, + 321, + 288, + 573 + ], + "type": "inline_equation", + "content": "W(\\mathbf{x}|\\mathbf{p}) = (\\bar{s}(\\bar{\\mathbf{x}}), \\bar{\\mathbf{x}})" + }, + { + "bbox": [ + 46, + 321, + 288, + 573 + ], + "type": "text", + "content": ", where the signed distance of an observation-space point is obtained by mapping it into its canonical correspondence " + }, + { + "bbox": [ + 46, + 321, + 288, + 573 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{x}}" + }, + { + "bbox": [ + 46, + 321, + 288, + 573 + ], + "type": "text", + "content": " and querying the pre-computed canonical SDF " + }, + { + "bbox": [ + 46, + 321, + 288, + 573 + ], + "type": "inline_equation", + "content": "\\bar{s}" + }, + { + "bbox": [ + 46, + 321, + 288, + 573 + ], + "type": "text", + "content": ". Thus we only learn a correspondence field with which we can deform the canonical SDF to different FLAME configurations, and then supervise the deformed SDFs with the FLAME surface boundary, normals " + }, + { + "bbox": [ + 46, + 321, + 288, + 573 + ], + "type": "inline_equation", + "content": "(L_{iso})" + }, + { + "bbox": [ + 46, + 321, + 288, + 573 + ], + "type": "text", + "content": " and Eikonal regularization " + }, + { + "bbox": [ + 46, + 321, + 288, + 573 + ], + "type": "inline_equation", + "content": "(L_{eik})" + }, + { + "bbox": [ + 46, + 321, + 288, + 573 + ], + "type": "text", + "content": ". As such, even for off-the-surface samples, their canonical correspondences are well regularized in space with the geometric properties of signed distance functions via " + }, + { + "bbox": [ + 46, + 321, + 288, + 573 + ], + "type": "inline_equation", + "content": "L_{iso}" + }, + { + "bbox": [ + 46, + 321, + 288, + 573 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 321, + 288, + 573 + ], + "type": "inline_equation", + "content": "L_{eik}" + }, + { + "bbox": [ + 46, + 321, + 288, + 573 + ], + "type": "text", + "content": ". In contrast to explicit mesh deformation [4], our implicit volumetric correspondence is more accurate, differentiable, smooth everywhere and semantically consistent with the properties of SDF." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 586, + 268, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 586, + 268, + 597 + ], + "spans": [ + { + "bbox": [ + 47, + 586, + 268, + 597 + ], + "type": "text", + "content": "3.2.2 Canonical Generation with Geometric Prior" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 605, + 287, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 700 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 700 + ], + "type": "text", + "content": "With our pretrained semantic SDF " + }, + { + "bbox": [ + 46, + 605, + 287, + 700 + ], + "type": "inline_equation", + "content": "W(\\mathbf{x}|\\mathbf{p})" + }, + { + "bbox": [ + 46, + 605, + 287, + 700 + ], + "type": "text", + "content": " modeling the shape and expression variation, we leverage the triplane for 3d-aware generation of human heads with canonical shape and expression. In particular, to generate a neural radiance feature " + }, + { + "bbox": [ + 46, + 605, + 287, + 700 + ], + "type": "inline_equation", + "content": "(\\sigma, \\mathbf{f})" + }, + { + "bbox": [ + 46, + 605, + 287, + 700 + ], + "type": "text", + "content": " for a point " + }, + { + "bbox": [ + 46, + 605, + 287, + 700 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 46, + 605, + 287, + 700 + ], + "type": "text", + "content": " in observation space " + }, + { + "bbox": [ + 46, + 605, + 287, + 700 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(\\mathbf{p})" + }, + { + "bbox": [ + 46, + 605, + 287, + 700 + ], + "type": "text", + "content": ", we use our correspondence function " + }, + { + "bbox": [ + 46, + 605, + 287, + 700 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 46, + 605, + 287, + 700 + ], + "type": "text", + "content": " to back warp " + }, + { + "bbox": [ + 46, + 605, + 287, + 700 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 46, + 605, + 287, + 700 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 46, + 605, + 287, + 700 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{x}}" + }, + { + "bbox": [ + 46, + 605, + 287, + 700 + ], + "type": "text", + "content": ", with which we project and sample the canonical triplane features followed with a tiny MLP decoding." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 58, + 701, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 701, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 701, + 287, + 713 + ], + "type": "text", + "content": "In spite of the control disentanglement, there is no ex" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 545, + 142 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 142 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 142 + ], + "type": "text", + "content": "plicit loss that constrains the triplane-generated neural radiance density to conform to the shape and expression as defined in " + }, + { + "bbox": [ + 304, + 72, + 545, + 142 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 304, + 72, + 545, + 142 + ], + "type": "text", + "content": ". Therefore, we guide the generation of neural radiance density field by minimizing its difference to the underlying FLAME head geometry represented with SDF, as" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 147, + 545, + 177 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 147, + 545, + 177 + ], + "spans": [ + { + "bbox": [ + 313, + 147, + 545, + 177 + ], + "type": "interline_equation", + "content": "L _ {p r i o r} = \\frac {1}{| R |} \\sum_ {\\mathbf {x} \\in R} e ^ {- \\gamma | s (\\mathbf {x} | p) |} \\left| \\sigma \\left(\\mathbf {x} \\mid \\mathbf {z}, \\mathbf {p}\\right) - \\sigma^ {*} \\left(\\mathbf {x} \\mid \\mathbf {p}\\right) \\right|, \\tag {4}", + "image_path": "6d8dbe538e041c59bcc2d93876c87cef0aaae51a44ade1e33016463f48e8f057.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 353, + 178, + 545, + 202 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 178, + 545, + 202 + ], + "spans": [ + { + "bbox": [ + 353, + 178, + 545, + 202 + ], + "type": "interline_equation", + "content": "\\sigma^ {*} (\\mathbf {x} | \\mathbf {p}) = \\frac {1}{\\kappa} \\cdot \\operatorname {S i g m o i d} \\left(\\frac {- s (\\mathbf {x} | \\mathbf {p})}{\\kappa}\\right) \\tag {5}", + "image_path": "dbe31bfa7e29fe2b6d6f6c82146e6b4824f868badfb276d11e78e584226a3757.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 207, + 545, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 207, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 304, + 207, + 545, + 338 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 207, + 545, + 338 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 304, + 207, + 545, + 338 + ], + "type": "text", + "content": " is the stratified ray samples for volume rendering and " + }, + { + "bbox": [ + 304, + 207, + 545, + 338 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 304, + 207, + 545, + 338 + ], + "type": "text", + "content": " is a learnable scalar controlling the density tightness around the SDF boundary. Following [37, 63], we convert SDF value " + }, + { + "bbox": [ + 304, + 207, + 545, + 338 + ], + "type": "inline_equation", + "content": "s(\\mathbf{x}|\\mathbf{p})" + }, + { + "bbox": [ + 304, + 207, + 545, + 338 + ], + "type": "text", + "content": " to proxy 3D density " + }, + { + "bbox": [ + 304, + 207, + 545, + 338 + ], + "type": "inline_equation", + "content": "\\sigma^{*}(\\mathbf{x}|\\mathbf{p})" + }, + { + "bbox": [ + 304, + 207, + 545, + 338 + ], + "type": "text", + "content": " assuming non-hollow surfaces. We decay the weights for our geometric prior loss " + }, + { + "bbox": [ + 304, + 207, + 545, + 338 + ], + "type": "inline_equation", + "content": "L_{prior}" + }, + { + "bbox": [ + 304, + 207, + 545, + 338 + ], + "type": "text", + "content": " as the point moving away from the SDF boundary, allowing higher degrees of freedom in generation of residual geometries, such as hair and glasses. The geometric prior loss effectively guides the 3D head geometry learning but should not be overpowered which otherwise might lead to loss of geometric details." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 352, + 482, + 364 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 352, + 482, + 364 + ], + "spans": [ + { + "bbox": [ + 305, + 352, + 482, + 364 + ], + "type": "text", + "content": "3.2.3 Fine-Grained Expression Control" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 371, + 545, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 371, + 545, + 563 + ], + "spans": [ + { + "bbox": [ + 304, + 371, + 545, + 563 + ], + "type": "text", + "content": "Our geometric prior loss " + }, + { + "bbox": [ + 304, + 371, + 545, + 563 + ], + "type": "inline_equation", + "content": "L_{prior}" + }, + { + "bbox": [ + 304, + 371, + 545, + 563 + ], + "type": "text", + "content": " provides local 3D pointwise guidance, and is able to well regularize the shape generation and achieve coarse-level expression control. However, for delicate expressions, such as eye blinks, " + }, + { + "bbox": [ + 304, + 371, + 545, + 563 + ], + "type": "inline_equation", + "content": "L_{prior}" + }, + { + "bbox": [ + 304, + 371, + 545, + 563 + ], + "type": "text", + "content": " provides little supervision as the geometric variation is subtle. Moreover, for regions with complex correspondences, such as around the lips, it is challenging to guide the formation of correct expressions globally, just with point-wise geometric losses. To improve the control granularity, we propose an image-level supervision loss that requires a synthesized image " + }, + { + "bbox": [ + 304, + 371, + 545, + 563 + ], + "type": "inline_equation", + "content": "I_{RGB}(\\mathbf{z}|\\mathbf{c},\\mathbf{p})" + }, + { + "bbox": [ + 304, + 371, + 545, + 563 + ], + "type": "text", + "content": " matching the target expression as defined in the input " + }, + { + "bbox": [ + 304, + 371, + 545, + 563 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 304, + 371, + 545, + 563 + ], + "type": "text", + "content": ". Using our training images with estimated control labels " + }, + { + "bbox": [ + 304, + 371, + 545, + 563 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 304, + 371, + 545, + 563 + ], + "type": "text", + "content": ", we first pretrain an image encoder " + }, + { + "bbox": [ + 304, + 371, + 545, + 563 + ], + "type": "inline_equation", + "content": "E(I_{RGB}) = (\\tilde{\\boldsymbol{\\beta}},\\tilde{\\boldsymbol{\\theta}})" + }, + { + "bbox": [ + 304, + 371, + 545, + 563 + ], + "type": "text", + "content": " that regresses the expression coefficients " + }, + { + "bbox": [ + 304, + 371, + 545, + 563 + ], + "type": "inline_equation", + "content": "\\tilde{\\boldsymbol{\\beta}}" + }, + { + "bbox": [ + 304, + 371, + 545, + 563 + ], + "type": "text", + "content": " and joint poses " + }, + { + "bbox": [ + 304, + 371, + 545, + 563 + ], + "type": "inline_equation", + "content": "\\tilde{\\boldsymbol{\\theta}}" + }, + { + "bbox": [ + 304, + 371, + 545, + 563 + ], + "type": "text", + "content": ". During our 3D GAN training, we then apply our image-level control supervision as" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 308, + 579, + 545, + 611 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 579, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 308, + 579, + 545, + 611 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {e n c} = | \\tilde {\\boldsymbol {\\beta}} - \\boldsymbol {\\beta} | + | \\tilde {\\boldsymbol {\\theta}} - \\boldsymbol {\\theta} | + | \\mathbf {S} (\\boldsymbol {\\alpha}, \\tilde {\\boldsymbol {\\beta}}, \\tilde {\\boldsymbol {\\theta}}) - \\mathbf {S} (\\boldsymbol {\\alpha}, \\boldsymbol {\\beta}, \\boldsymbol {\\theta}) | + \\\\ + \\left| \\mathbf {J} \\mathbf {S} (\\boldsymbol {\\alpha}, \\tilde {\\boldsymbol {\\beta}}, \\tilde {\\boldsymbol {\\theta}}) - \\mathbf {J} \\mathbf {S} (\\boldsymbol {\\alpha}, \\boldsymbol {\\beta}, \\boldsymbol {\\theta}) \\right|, \\tag {6} \\\\ \\end{array}", + "image_path": "ab999e56bad743b29036b4d34dee140ccd9cf48518b8b0f27fb35ef056f773ff.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{S},\\mathbf{J}" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": " are the FLAME mesh vertices and 3D landmarks regressor respectively. While being straightforward for the first 2 terms, the last two terms in " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "L_{enc}" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": " penalize deviation of 3D vertex coordinates and surface landmarks after mesh decoding. We note that we do not supervise shape " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "L_{enc}" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": " since our geometric prior loss " + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "inline_equation", + "content": "L_{prior}" + }, + { + "bbox": [ + 304, + 617, + 545, + 713 + ], + "type": "text", + "content": " suffices in shape control already and also due to the ambiguity of shape scaling estimated from monocular images." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12818" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 70, + 217, + 319 + ], + "blocks": [ + { + "bbox": [ + 53, + 70, + 217, + 319 + ], + "lines": [ + { + "bbox": [ + 53, + 70, + 217, + 319 + ], + "spans": [ + { + "bbox": [ + 53, + 70, + 217, + 319 + ], + "type": "image", + "image_path": "5c3066ab5d5b8315700701851a923395642694462106b4973e7799cd2ebb3d09.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 327, + 547, + 361 + ], + "lines": [ + { + "bbox": [ + 46, + 327, + 547, + 361 + ], + "spans": [ + { + "bbox": [ + 46, + 327, + 547, + 361 + ], + "type": "text", + "content": "Figure 5. Qualitative comparisons in expression control. Under a similar expression control, our method achieves the best perceptual quality with better identity-preservation (compared to GAN-Control [47] and 3DFaceShop [51]), more realistic dynamic details (compared to DiscoFaceGAN [10]) and higher expression control accuracy (compared to AntiFaceGAN [58], 3DFaceShop [51], HeadNeRF [21])." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 235, + 69, + 287, + 319 + ], + "blocks": [ + { + "bbox": [ + 235, + 69, + 287, + 319 + ], + "lines": [ + { + "bbox": [ + 235, + 69, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 235, + 69, + 287, + 319 + ], + "type": "image", + "image_path": "583be0f7e1981a934d961356d719a966e245573b4936a53abc3f71932f68c92e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 299, + 70, + 380, + 319 + ], + "blocks": [ + { + "bbox": [ + 299, + 70, + 380, + 319 + ], + "lines": [ + { + "bbox": [ + 299, + 70, + 380, + 319 + ], + "spans": [ + { + "bbox": [ + 299, + 70, + 380, + 319 + ], + "type": "image", + "image_path": "cdccdc8738c51f4169406576d78e645eb54363672708f337e347e00ebf695203.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 381, + 70, + 461, + 319 + ], + "blocks": [ + { + "bbox": [ + 381, + 70, + 461, + 319 + ], + "lines": [ + { + "bbox": [ + 381, + 70, + 461, + 319 + ], + "spans": [ + { + "bbox": [ + 381, + 70, + 461, + 319 + ], + "type": "image", + "image_path": "ff1b0720b4720eb7fc5a2bbd5872d0512799ca609fe9d427cc79e6b920c9ad9c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 462, + 70, + 543, + 319 + ], + "blocks": [ + { + "bbox": [ + 462, + 70, + 543, + 319 + ], + "lines": [ + { + "bbox": [ + 462, + 70, + 543, + 319 + ], + "spans": [ + { + "bbox": [ + 462, + 70, + 543, + 319 + ], + "type": "image", + "image_path": "9d212897a27f4e529472d14ad6dd6482eae3e88b967af1c665fb2d16de795aa8.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 381, + 195, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 381, + 195, + 395 + ], + "spans": [ + { + "bbox": [ + 47, + 381, + 195, + 395 + ], + "type": "text", + "content": "3.2.4 Dynamic Details Modeling" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 410, + 289, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 410, + 289, + 601 + ], + "spans": [ + { + "bbox": [ + 46, + 410, + 289, + 601 + ], + "type": "text", + "content": "To this end, we have achieved controllable static image synthesis. However, one should observe the variation of shading and geometric details in a dynamic head motion, such as the appearance of dimples and eye wrinkles in smiling. We consider the appearance of dynamic details is highly correlated with the driving expression " + }, + { + "bbox": [ + 46, + 410, + 289, + 601 + ], + "type": "inline_equation", + "content": "(\\beta ,\\theta)" + }, + { + "bbox": [ + 46, + 410, + 289, + 601 + ], + "type": "text", + "content": ". To model such temporal effects, one could try to condition the triplane generation on expression as " + }, + { + "bbox": [ + 46, + 410, + 289, + 601 + ], + "type": "inline_equation", + "content": "\\mathcal{M}(\\mathbf{z},\\beta ,\\theta)" + }, + { + "bbox": [ + 46, + 410, + 289, + 601 + ], + "type": "text", + "content": ". However, such design results in entanglement of expression control with the generative latent code " + }, + { + "bbox": [ + 46, + 410, + 289, + 601 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 46, + 410, + 289, + 601 + ], + "type": "text", + "content": ", inducing identity or appearance changes when varying expressions. It is also hard to synthesize images out of training distribution of expressions " + }, + { + "bbox": [ + 46, + 410, + 289, + 601 + ], + "type": "inline_equation", + "content": "(\\beta ,\\theta)" + }, + { + "bbox": [ + 46, + 410, + 289, + 601 + ], + "type": "text", + "content": ". We therefore leave our triplane generation disentangled but when decoding neural feature field " + }, + { + "bbox": [ + 46, + 410, + 289, + 601 + ], + "type": "inline_equation", + "content": "(\\sigma ,\\mathbf{f})" + }, + { + "bbox": [ + 46, + 410, + 289, + 601 + ], + "type": "text", + "content": " from sampled triplane features, we additionally condition the MLP-decoder on " + }, + { + "bbox": [ + 46, + 410, + 289, + 601 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 46, + 410, + 289, + 601 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 410, + 289, + 601 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 46, + 410, + 289, + 601 + ], + "type": "text", + "content": ". Specifically, we have" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 63, + 620, + 287, + 635 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 620, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 63, + 620, + 287, + 635 + ], + "type": "interline_equation", + "content": "(\\sigma (\\mathbf {x}), \\mathbf {f} (\\mathbf {x}) | \\mathbf {z}, \\mathbf {p}) = \\Phi (\\mathcal {M} (\\bar {\\mathbf {x}} (\\mathbf {x} | \\mathbf {p}) | \\mathbf {z}), \\phi (\\boldsymbol {\\beta}, \\boldsymbol {\\theta})), \\quad (7)", + "image_path": "36141d52076f29da90146c977d1de777d67ad839e009899e15cb81f9de30740d.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": "where both " + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": " are tiny MLPs and " + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": " regresses an expression-dependent feature vector from " + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "inline_equation", + "content": "(\\beta, \\theta)" + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": " after positional encoding. For better extrapolation to novel expressions and jaw poses, we add Gaussian noise to the conditioning parameters to prevent MLP overfitting." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 380, + 388, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 380, + 388, + 395 + ], + "spans": [ + { + "bbox": [ + 306, + 380, + 388, + 395 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 402, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 402, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 402, + 547, + 715 + ], + "type": "text", + "content": "Training and Dataset. Our training is devised into two stages. In the pretraining stage, we build our semantic SDF " + }, + { + "bbox": [ + 304, + 402, + 547, + 715 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 304, + 402, + 547, + 715 + ], + "type": "text", + "content": " using a four 192-dimensional MLP with a collection of 150K FLAME meshes by Gaussian sampling of the control parameters " + }, + { + "bbox": [ + 304, + 402, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 304, + 402, + 547, + 715 + ], + "type": "text", + "content": ". For each mesh, we sample 4K surface samples " + }, + { + "bbox": [ + 304, + 402, + 547, + 715 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 402, + 547, + 715 + ], + "type": "text", + "content": " with surface normals " + }, + { + "bbox": [ + 304, + 402, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 304, + 402, + 547, + 715 + ], + "type": "text", + "content": " and ground-truth correspondence point " + }, + { + "bbox": [ + 304, + 402, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{x}}^*" + }, + { + "bbox": [ + 304, + 402, + 547, + 715 + ], + "type": "text", + "content": ", and 4K off surface samples " + }, + { + "bbox": [ + 304, + 402, + 547, + 715 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 304, + 402, + 547, + 715 + ], + "type": "text", + "content": " distributed uniformly inside the box-bounded volume. Our canonical mesh has a neutral shape " + }, + { + "bbox": [ + 304, + 402, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\alpha = 0" + }, + { + "bbox": [ + 304, + 402, + 547, + 715 + ], + "type": "text", + "content": ", expression " + }, + { + "bbox": [ + 304, + 402, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\beta = 0" + }, + { + "bbox": [ + 304, + 402, + 547, + 715 + ], + "type": "text", + "content": " and neck pose, but with an opening jaw at 10 degrees. We close the face connectivity between the lips for a watertight geometry and also as a proxy geometry for the mouth cavity modeling. We open the jaw at the canonical space such that the spatial geometry and appearance inside the mouth can be modeled with distinguishable triplane features. In the second training stage, we freeze the weights of " + }, + { + "bbox": [ + 304, + 402, + 547, + 715 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 304, + 402, + 547, + 715 + ], + "type": "text", + "content": " for best efficiency and train our model on the FFHQ [24], a human face dataset with 70,000 in-the-wild images. For each image, we estimate the camera pose " + }, + { + "bbox": [ + 304, + 402, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 304, + 402, + 547, + 715 + ], + "type": "text", + "content": " and FLAME parameters " + }, + { + "bbox": [ + 304, + 402, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 304, + 402, + 547, + 715 + ], + "type": "text", + "content": " with a nonlinear optimization in fitting 2D landmarks, assuming zero root and neck rotation and camera located 1.0m away from the world origin. We refer to the supplementary materials for more details. We rebalance the FFHQ dataset by duplicating images with large head poses and jaw opening expressions. We note that we also use pairs of images and " + }, + { + "bbox": [ + 304, + 402, + 547, + 715 + ], + "type": "inline_equation", + "content": "(\\beta, \\theta)" + }, + { + "bbox": [ + 304, + 402, + 547, + 715 + ], + "type": "text", + "content": " for fine-tuning of a light-weight image encoder" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "12819" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 66, + 69, + 331, + 150 + ], + "blocks": [ + { + "bbox": [ + 66, + 69, + 331, + 150 + ], + "lines": [ + { + "bbox": [ + 66, + 69, + 331, + 150 + ], + "spans": [ + { + "bbox": [ + 66, + 69, + 331, + 150 + ], + "type": "table", + "html": "
ModelFID ↓KID ↓DSα ↑DSβ,θ ↑DSc ↑IS ↑
DiscoFaceGAN-SR [10]42.60.04-27.20.110.59
GAN-Control [47]9.80.005-43.70.280.62
HeadNeRF [21]1630.1681.338.90.250.678
AniFaceGAN-SR [58]30.50.0242.011.10.040.73
3DFaceShop [51]24.80.0180.4564.20.0290.594
Ours5.70.00163.171.90.350.80
", + "image_path": "3bcec1d3dfc5a935a21c4daea86416978f3f43809b8fe1861325886bac6d26bc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 155, + 152, + 240, + 162 + ], + "lines": [ + { + "bbox": [ + 155, + 152, + 240, + 162 + ], + "spans": [ + { + "bbox": [ + 155, + 152, + 240, + 162 + ], + "type": "text", + "content": "(a) Baseline comparisons." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 348, + 66, + 553, + 148 + ], + "blocks": [ + { + "bbox": [ + 348, + 66, + 553, + 148 + ], + "lines": [ + { + "bbox": [ + 348, + 66, + 553, + 148 + ], + "spans": [ + { + "bbox": [ + 348, + 66, + 553, + 148 + ], + "type": "table", + "html": "
ModelOurs wo LencOurs wo LpriorOurs
FID ↓6.06.15.7
KID ↓0.001720.001740.0016
ASD (cm) ↓0.1240.1350.124
var(α) ↓0.0005250.0005770.000515
AED (cm) ↓0.12330.08130.0797
var(β, θ) ↓0.0047070.004060.00404
", + "image_path": "4aa5862c35481321bb14bb26e9435b13f604dff8b3e6a67d861a1e861fd949d2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 417, + 148, + 479, + 158 + ], + "lines": [ + { + "bbox": [ + 417, + 148, + 479, + 158 + ], + "spans": [ + { + "bbox": [ + 417, + 148, + 479, + 158 + ], + "type": "text", + "content": "(b) Ablation study." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 168, + 547, + 202 + ], + "lines": [ + { + "bbox": [ + 46, + 168, + 547, + 202 + ], + "spans": [ + { + "bbox": [ + 46, + 168, + 547, + 202 + ], + "type": "text", + "content": "Table 1. (a) Our method outperforms the prior 2D and 3D controllable image synthesis methods in both image quality and control disentanglement. (b) Our method demonstrates the best control accuracy over shape and expression with geometric prior " + }, + { + "bbox": [ + 46, + 168, + 547, + 202 + ], + "type": "inline_equation", + "content": "L_{prior}" + }, + { + "bbox": [ + 46, + 168, + 547, + 202 + ], + "type": "text", + "content": " and self-supervised reconstruction loss " + }, + { + "bbox": [ + 46, + 168, + 547, + 202 + ], + "type": "inline_equation", + "content": "L_{enc}" + }, + { + "bbox": [ + 46, + 168, + 547, + 202 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 212, + 287, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 212, + 287, + 236 + ], + "spans": [ + { + "bbox": [ + 47, + 212, + 287, + 236 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 47, + 212, + 287, + 236 + ], + "type": "text", + "content": ", using a ResNet50 [19] backbone followed with a single-layer MLP." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 256, + 187, + 269 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 256, + 187, + 269 + ], + "spans": [ + { + "bbox": [ + 47, + 256, + 187, + 269 + ], + "type": "text", + "content": "4.1. Qualitative Comparisons" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 278, + 287, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 278, + 287, + 434 + ], + "spans": [ + { + "bbox": [ + 46, + 278, + 287, + 434 + ], + "type": "text", + "content": "Controlled Portrait Synthesis. Our framework enables high-fidelity head synthesis with disentangled control over camera pose, shape and expression, as shown in Figure 1. We compare our method with prior controllable portrait synthesis works include DiscoFaceGAN [10], GAN-Control [47], HeadNeRF [21], AntiFaceGAN [58] and 3DFaceShop [51]. The first two methods integrates 3DMM controls into 2D face synthesis. HeadNeRF [21] is a 3D parametric NeRF-based head model built from both indoor multiview image capture datasets and in-the-wild monocular image collections. AntiFaceGAN [58] and 3DFaceShop [51] are concurrent 3D-aware controllable GANs in face synthesis." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 438, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 438, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 438, + 288, + 714 + ], + "type": "text", + "content": "We qualitatively compare with prior work in Figure 5 for expression control. GAN-Control shows high-quality expression manipulation but with noticeable identity variation with differences in hair, head shape contour and background, whereas DiscoFaceGAN better maintains identity but with lower perceptual quality, and lacks dynamic details as expression varies. The expression control space of 3DFaceShop [51] is sensitive, suffering from appearance changes even with minor expression variation. AniFaceGAN demonstrates visually-consistent expression editing but with limited resolution and image quality, e.g., with blurry artifacts in hair and teeth. HeadNeRF ensures decent consistency in image generation by rendering the conditional NeRF but lacks fine details with very limited perceptual quality. In contrast, our approach produces the most compelling images with consistent appearance and dynamic details under expression changes. Moreover, controllable neck pose is a key factor towards realistic video avatar in applications like talking head synthesis. As shown in Figure. 1, our method achieves explicit control of neck and jaw poses which are seldom explored in prior work. Please refer to supplementary material for qualitative comparison in view consistency and shape editing." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 212, + 545, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 212, + 545, + 332 + ], + "spans": [ + { + "bbox": [ + 304, + 212, + 545, + 332 + ], + "type": "text", + "content": "Dynamic details. Our method presents highly-consistent image synthesis in control of shape, expression and camera poses, but also depicts temporal realism in portrait animation, credited to the modeling of expression-dependent dynamic details. As shown in the last row of Figure 3, the appearance of wrinkles around the mouth and eyes when transiting from a neural expression to smiling largely enhances the animation realism. In comparison, without explicitly modeling the dynamic details, the wrinkles are embedded in the appearance and do not vary with expressions." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 338, + 545, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 338, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 304, + 338, + 545, + 469 + ], + "type": "text", + "content": "Expressive Head Synthesis with Extrapolated Controls. AntiFaceGAN [58] and 3DFaceShop [51] depict dynamic details as well since their generation of the neural fields is conditioned on the input expression latent code. However, their designs result in shape and expression entanglement with the appearance generation, as reflected in the identity changes as shown in Figure. 5. By embedding the controls in a Gaussian latent space with such as Variation AutoEncoder (VAE), their approaches sacrifice expressiveness. The synthesized image quality is also highly correlated to the distribution of training images with target expressions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 471, + 545, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 471, + 545, + 615 + ], + "spans": [ + { + "bbox": [ + 304, + 471, + 545, + 615 + ], + "type": "text", + "content": "In contrast, our tri-plane generation is explicitly disentangled from shape and expression control. Moreover, the volume deformation is independently learnt from the deformed FLAME mesh collections which offer abundant 3D geometric knowledge with largely augmented control space. Therefore we are much less dependent on the distribution of the training images and support better extrapolation to unseen novel expressions. In Figure. 1, we show high-quality synthesized head with extreme jaw and neck articulated movements which do not exist in our training images. Our expression control is also more expressive, supporting subtle expressions like eye blinks (Figure. 1 3)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 623, + 452, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 623, + 452, + 635 + ], + "spans": [ + { + "bbox": [ + 306, + 623, + 452, + 635 + ], + "type": "text", + "content": "4.2. Quantitative Comparisons" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": "Image Quality We measure the image quality with Frechet Inception Distance (FID) [20] and Kernel Inception Distance [5] between 50K randomly synthesized images and 50K randomly sampled real images at the resolution of " + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": ". Since DiscoFaceGAN [10] and AniFaceGAN [58] only synthesize images at " + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": " reso" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12820" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": "lution, we utilize a state-of-the-art super-resolution model, SwinIR [30] to upsample into " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": " for a fair comparison. As shown in Table. 1a, our method is superior in both FID and KID than all prior work, demonstrating the most compelling image quality. We note that the original EG3D has a slightly lower FID at 4.8 which is expected since we introduce controllability with additional loss regularization." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 164, + 289, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 164, + 289, + 273 + ], + "spans": [ + { + "bbox": [ + 46, + 164, + 289, + 273 + ], + "type": "text", + "content": "Disentanglement To evaluate the disentangled controllability of our model over shape, expression and camera pose, we measure the disentanglement score [10] of synthesized images as " + }, + { + "bbox": [ + 46, + 164, + 289, + 273 + ], + "type": "inline_equation", + "content": "\\mathrm{DS}_{\\alpha}" + }, + { + "bbox": [ + 46, + 164, + 289, + 273 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 164, + 289, + 273 + ], + "type": "inline_equation", + "content": "\\mathrm{DS}_{\\beta,\\theta}" + }, + { + "bbox": [ + 46, + 164, + 289, + 273 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 164, + 289, + 273 + ], + "type": "inline_equation", + "content": "\\mathrm{DS}_{\\mathbf{c}}" + }, + { + "bbox": [ + 46, + 164, + 289, + 273 + ], + "type": "text", + "content": " respectively. DS measures the stability of other factors when a single attribute is modified in image synthesis. We employ DECA [12] for estimation of FLAME parameters from generated images and calculate the variance of the estimated parameters " + }, + { + "bbox": [ + 46, + 164, + 289, + 273 + ], + "type": "inline_equation", + "content": "(\\alpha, \\{\\beta, \\theta\\}, \\mathbf{c})" + }, + { + "bbox": [ + 46, + 164, + 289, + 273 + ], + "type": "text", + "content": ". Specifically the " + }, + { + "bbox": [ + 46, + 164, + 289, + 273 + ], + "type": "inline_equation", + "content": "\\mathrm{DS}_i" + }, + { + "bbox": [ + 46, + 164, + 289, + 273 + ], + "type": "text", + "content": " is calculated as" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 80, + 279, + 287, + 309 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 279, + 287, + 309 + ], + "spans": [ + { + "bbox": [ + 80, + 279, + 287, + 309 + ], + "type": "interline_equation", + "content": "D S _ {i} = \\prod_ {j \\neq i} \\frac {\\operatorname {v a r} (i)}{\\operatorname {v a r} (j)}, i, j \\in \\{\\boldsymbol {\\alpha}, \\{\\boldsymbol {\\beta}, \\boldsymbol {\\theta} \\}, \\mathbf {c} \\}. \\tag {8}", + "image_path": "95a84816ed6d0c96b824e7edeea9978f95908bb13034300f819b141b62f399fe.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 316, + 289, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 316, + 289, + 400 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 289, + 400 + ], + "type": "text", + "content": "Higher value of DS indicates better disentanglement. Additionally, we evaluate the identity similarity between pairs of synthesized images with random camera poses and expressions by calculating the cosine similarity of the face embeddings with a pre-trained face recognition module [26]. Our approach demonstrates the best disentanglement numerically over all prior work as indicated in Table. 1a." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 407, + 149, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 407, + 149, + 418 + ], + "spans": [ + { + "bbox": [ + 47, + 407, + 149, + 418 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 425, + 289, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 425, + 289, + 666 + ], + "spans": [ + { + "bbox": [ + 46, + 425, + 289, + 666 + ], + "type": "text", + "content": "We ablate the efficacy of the individual component by removing it from our full pipeline. As shown in Figure. 3, we show the loss of control accuracy over shape and expression respectively when removing the geometric prior loss " + }, + { + "bbox": [ + 46, + 425, + 289, + 666 + ], + "type": "inline_equation", + "content": "L_{prior}" + }, + { + "bbox": [ + 46, + 425, + 289, + 666 + ], + "type": "text", + "content": " and control loss " + }, + { + "bbox": [ + 46, + 425, + 289, + 666 + ], + "type": "inline_equation", + "content": "L_{enc}" + }, + { + "bbox": [ + 46, + 425, + 289, + 666 + ], + "type": "text", + "content": ". Conditioning the neural radiance field on expression is also critical to the modeling of dynamic details. Numerically we validate the efficacy of " + }, + { + "bbox": [ + 46, + 425, + 289, + 666 + ], + "type": "inline_equation", + "content": "L_{prior}" + }, + { + "bbox": [ + 46, + 425, + 289, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 425, + 289, + 666 + ], + "type": "inline_equation", + "content": "L_{enc}" + }, + { + "bbox": [ + 46, + 425, + 289, + 666 + ], + "type": "text", + "content": " with Average Shape Distance (ASD) and Average Expression Distance (AED). From the prior distribution, we randomly sample 500 shapes and expressions, with which we synthesize images with 10 different identities. We then reconstruct FLAME parameters from the synthesized images and compare against the input control. We compute 3D per-vertex " + }, + { + "bbox": [ + 46, + 425, + 289, + 666 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 46, + 425, + 289, + 666 + ], + "type": "text", + "content": " distance of FLAME meshes for ASD while we use 3D landmarks " + }, + { + "bbox": [ + 46, + 425, + 289, + 666 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 46, + 425, + 289, + 666 + ], + "type": "text", + "content": " distance for AED calculation. Additionally we compute the variance of estimated shapes and expressions within each control group, with lower value indicating more precise control. The efficacy of " + }, + { + "bbox": [ + 46, + 425, + 289, + 666 + ], + "type": "inline_equation", + "content": "L_{prior}" + }, + { + "bbox": [ + 46, + 425, + 289, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 425, + 289, + 666 + ], + "type": "inline_equation", + "content": "L_{enc}" + }, + { + "bbox": [ + 46, + 425, + 289, + 666 + ], + "type": "text", + "content": " is well evidenced in Table. 1b, with even slight image quality improvements." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 671, + 129, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 129, + 684 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 129, + 684 + ], + "type": "text", + "content": "4.4. Applications" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "text", + "content": "Talking Head Video Generation. In Figure. 1, we showcase talking head video generation in controllable views" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 547, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 193 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 193 + ], + "type": "text", + "content": "driven with animation sequences of FLAME. Thanks to the high control accuracy, our method is able to synthesize various talking head videos with the same head movements and expressions performed by different identities. Our method is expressive in depicting both large articulated neck and jaw movements and subtle facial expressions like eye blinks, with rich dynamic details. Shape manipulation is easily achievable as well by modifying the shape parameters. Please refer to our supplementary materials for more results in high resolution." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 198, + 547, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 198, + 547, + 390 + ], + "spans": [ + { + "bbox": [ + 304, + 198, + 547, + 390 + ], + "type": "text", + "content": "Portrait Image Manipulation and Animation. As illustrated in Figure. 1, our model also supports 3D-aware face reenactment of a single-view portrait to a video sequence. To achieve that, we perform an optimization in the latent " + }, + { + "bbox": [ + 304, + 198, + 547, + 390 + ], + "type": "inline_equation", + "content": "Z+" + }, + { + "bbox": [ + 304, + 198, + 547, + 390 + ], + "type": "text", + "content": " space [24] to find the corresponding latent embedding, with FLAME parameter and camera pose estimated from the input portrait. With a frozen generator, the optimization is performed by measuring the similarity between generated image and real image using the " + }, + { + "bbox": [ + 304, + 198, + 547, + 390 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 304, + 198, + 547, + 390 + ], + "type": "text", + "content": " loss and LPIPS loss [65]. For better reconstruction quality, we alter the parameters of the tri-plane synthesis module with a fixed optimized latent code [43]. After that one can explicitly manipulate the portrait with a preserved identity and in a different camera pose and expression. With the expression codes " + }, + { + "bbox": [ + 304, + 198, + 547, + 390 + ], + "type": "inline_equation", + "content": "(\\beta, \\theta)" + }, + { + "bbox": [ + 304, + 198, + 547, + 390 + ], + "type": "text", + "content": " reconstructed from a video sequence, we are also able to reenact the portrait to the video motion." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 396, + 547, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 396, + 547, + 482 + ], + "spans": [ + { + "bbox": [ + 304, + 396, + 547, + 482 + ], + "type": "text", + "content": "Societal Impact. Our work focuses on improving the controllability of 3D-aware GANs in technical aspects and is not specifically designed for any malicious uses. This being said, we do see that the method could be potentially extended into controversial applications such as generating fake videos. Therefore, we believe that the synthesized images and videos should present themselves as synthetic." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 490, + 379, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 490, + 379, + 502 + ], + "spans": [ + { + "bbox": [ + 306, + 490, + 379, + 502 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 510, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 547, + 713 + ], + "type": "text", + "content": "In this work, we introduce OmniAvatar, a novel 3D-aware generative model for synthesis of controllable high-fidelity human head. Our model achieves disentangled semantic control by factoring the generative process into 3D-aware canonical head synthesis and implicit volume deformation to target shapes and expressions. By learning a deformable signed distance function with canonical volumetric correspondence, we instill the geometric prior knowledge of a 3D statistical head model into the generation of neural radiance fields, enabling superior generative capability to novel shapes and expressions. Our approach demonstrates expressive and compelling talking head generation and portrait image animation with fine-grained control accuracy and temporal realism. We believe the proposed method presents an interesting direction for 3D avatar creation and animation, which sheds light on many potential downstream tasks." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12821" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 124 + ], + "type": "text", + "content": "[1] Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas Guibas. Learning representations and generative models for 3d point clouds. In ICML, 2018. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 126, + 287, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 126, + 287, + 158 + ], + "spans": [ + { + "bbox": [ + 53, + 126, + 287, + 158 + ], + "type": "text", + "content": "[2] Thiemo Alldieck, Hongyi Xu, and Cristian Sminchisescu. imghum: Implicit generative models of 3d human shape and articulated pose. In ICCV, pages 5461-5470, 2021. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 160, + 287, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 160, + 287, + 204 + ], + "spans": [ + { + "bbox": [ + 53, + 160, + 287, + 204 + ], + "type": "text", + "content": "[3] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In iccv, pages 5855–5864, 2021. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 205, + 287, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 205, + 287, + 248 + ], + "spans": [ + { + "bbox": [ + 53, + 205, + 287, + 248 + ], + "type": "text", + "content": "[4] Alexander W Bergman, Petr Kellnhofer, Yifan Wang, Eric R Chan, David B Lindell, and Gordon Wetzstein. Generative neural articulated radiance fields. arXiv preprint arXiv:2206.14314, 2022. 3, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 251, + 287, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 251, + 287, + 294 + ], + "spans": [ + { + "bbox": [ + 53, + 251, + 287, + 294 + ], + "type": "text", + "content": "[5] Sai Bi, Kalyan Sunkavalli, Federico Perazzi, Eli Shechtman, Vladimir G Kim, and Ravi Ramamoorthi. Deep cg2real: Synthetic-to-real translation via image disentanglement. In iccv, pages 2730-2739, 2019. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 296, + 287, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 296, + 287, + 339 + ], + "spans": [ + { + "bbox": [ + 53, + 296, + 287, + 339 + ], + "type": "text", + "content": "[6] Volker Blanz and Thomas Vetter. A morphable model for the synthesis of 3d faces. In Proceedings of the 26th annual conference on Computer graphics and interactive techniques, 1999. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 341, + 287, + 395 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 341, + 287, + 395 + ], + "spans": [ + { + "bbox": [ + 53, + 341, + 287, + 395 + ], + "type": "text", + "content": "[7] Eric R Chan, Connor Z Lin, Matthew A Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, et al. Efficient geometry-aware 3d generative adversarial networks. CVPR, 2022. 2, 3, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 397, + 287, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 397, + 287, + 452 + ], + "spans": [ + { + "bbox": [ + 53, + 397, + 287, + 452 + ], + "type": "text", + "content": "[8] Eric R Chan, Marco Monteiro, Petr Kellnhofer, Jiajun Wu, and Gordon Wetzstein. pi-gan: Periodic implicit generative adversarial networks for 3d-aware image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5799-5809, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 453, + 287, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 453, + 287, + 486 + ], + "spans": [ + { + "bbox": [ + 53, + 453, + 287, + 486 + ], + "type": "text", + "content": "[9] Anpei Chen, Ruiyang Liu, Ling Xie, Zhang Chen, Hao Su, and Jingyi Yu. Sofgan: A portrait image generator with dynamic styling. sigg, 41(1):1-26, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 487, + 287, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 487, + 287, + 530 + ], + "spans": [ + { + "bbox": [ + 48, + 487, + 287, + 530 + ], + "type": "text", + "content": "[10] Yu Deng, Jiaolong Yang, Dong Chen, Fang Wen, and Xin Tong. Disentangled and controllable face image generation via 3d imitative-contrastive learning. In cvpr, pages 5154–5163, 2020. 2, 6, 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 533, + 287, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 533, + 287, + 565 + ], + "spans": [ + { + "bbox": [ + 48, + 533, + 287, + 565 + ], + "type": "text", + "content": "[11] Yu Deng, Jiaolong Yang, Jianfeng Xiang, and Xin Tong. Gram: Generative radiance manifolds for 3d-aware image generation. CVPR, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 567, + 287, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 287, + 599 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 287, + 599 + ], + "type": "text", + "content": "[12] Yao Feng, Haiwen Feng, Michael J. Black, and Timo Bolkart. Learning an animatable detailed 3D face model from in-the-wild images. volume 40, 2021. 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 601, + 287, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 287, + 655 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 287, + 655 + ], + "type": "text", + "content": "[13] Guy Gafni, Justus Thies, Michael Zollhofer, and Matthias Nießner. Dynamic neural radiance fields for monocular 4d facial avatar reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8649-8658, 2021. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 658, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 689 + ], + "type": "text", + "content": "[14] Partha Ghosh, Pravir Singh Gupta, Roy Uziel, Anurag Ranjan, Michael J Black, and Timo Bolkart. *Gif: Generative interpretable faces*. In 3dv, pages 868-878. IEEE, 2020. 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 692, + 287, + 713 + ], + "type": "text", + "content": "[15] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 94 + ], + "type": "text", + "content": "Yoshua Bengio. Generative adversarial nets. In NeurIPS, 2014. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 96, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 128 + ], + "type": "text", + "content": "[16] Amos Gropp, Lior Yariv, Niv Haim, Matan Atzmon, and Yaron Lipman. Implicit geometric regularization for learning shapes. In ICML, pages 3569-3579. 2020. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 129, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 129, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 308, + 129, + 545, + 161 + ], + "type": "text", + "content": "[17] Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. Stylenerf: A style-based 3d-aware generator for high-resolution image synthesis. CVPR, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 163, + 545, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 163, + 545, + 204 + ], + "spans": [ + { + "bbox": [ + 308, + 163, + 545, + 204 + ], + "type": "text", + "content": "[18] Yudong Guo, Keyu Chen, Sen Liang, Yong-Jin Liu, Hujun Bao, and Juyong Zhang. Ad-nerf: Audio driven neural radiance fields for talking head synthesis. In iccv, pages 5784-5794, 2021. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 206, + 545, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 206, + 545, + 237 + ], + "spans": [ + { + "bbox": [ + 308, + 206, + 545, + 237 + ], + "type": "text", + "content": "[19] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In cvpr, pages 770-778, 2016. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 239, + 545, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 239, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 308, + 239, + 545, + 281 + ], + "type": "text", + "content": "[20] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. NeurIPS, 2017. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 284, + 545, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 284, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 308, + 284, + 545, + 316 + ], + "type": "text", + "content": "[21] Yang Hong, Bo Peng, Haiyao Xiao, Ligang Liu, and Juyong Zhang. Headnerf: A real-time nerf-based parametric head model. In cvpr, pages 20374-20384, 2022. 3, 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 317, + 545, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 317, + 545, + 348 + ], + "spans": [ + { + "bbox": [ + 308, + 317, + 545, + 348 + ], + "type": "text", + "content": "[22] Kacper Kania, Kwang Moo Yi, Marek Kowalski, Tomasz Trzciński, and Andrea Tagliasacchi. Conerf: Controllable neural radiance fields. In cvpr, pages 18623-18632, 2022. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 350, + 545, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 350, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 308, + 350, + 545, + 381 + ], + "type": "text", + "content": "[23] Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Alias-free generative adversarial networks. NeurIPS, 2021. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 383, + 545, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 383, + 545, + 414 + ], + "spans": [ + { + "bbox": [ + 308, + 383, + 545, + 414 + ], + "type": "text", + "content": "[24] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, 2019. 1, 2, 6, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 416, + 545, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 416, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 308, + 416, + 545, + 448 + ], + "type": "text", + "content": "[25] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. In CVPR, 2020. 1, 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 449, + 545, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 449, + 545, + 480 + ], + "spans": [ + { + "bbox": [ + 308, + 449, + 545, + 480 + ], + "type": "text", + "content": "[26] Minchul Kim, Anil K Jain, and Xiaoming Liu. Adaface: Quality adaptive margin for face recognition. In cvpr, pages 18750-18759, 2022. 8" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 483, + 545, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 483, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 308, + 483, + 545, + 525 + ], + "type": "text", + "content": "[27] Marek Kowalski, Stephan J Garbin, Virginia Estellers, Tadas Baltrusaitis, Matthew Johnson, and Jamie Shotton. Config: Controllable neural face image generation. In " + }, + { + "bbox": [ + 308, + 483, + 545, + 525 + ], + "type": "inline_equation", + "content": "eccv" + }, + { + "bbox": [ + 308, + 483, + 545, + 525 + ], + "type": "text", + "content": ", pages 299-315. Springer, 2020. 2, 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 526, + 545, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 526, + 545, + 558 + ], + "spans": [ + { + "bbox": [ + 308, + 526, + 545, + 558 + ], + "type": "text", + "content": "[28] Ruihui Li, Xianzhi Li, Chi-Wing Fu, Daniel Cohen-Or, and Pheng-Ann Heng. Pu-gan: a point cloud upsampling adversarial network. In ICCV, 2019. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 559, + 545, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 559, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 308, + 559, + 545, + 601 + ], + "type": "text", + "content": "[29] Tianye Li, Timo Bolkart, Michael. J. Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4D scans. ACM Transactions on Graphics, 36(6):194:1-194:17, 2017. 2, 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 604, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 604, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 308, + 604, + 545, + 635 + ], + "type": "text", + "content": "[30] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In iccv, pages 1833-1844, 2021. 8" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 308, + 637, + 545, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 637, + 545, + 669 + ], + "spans": [ + { + "bbox": [ + 308, + 637, + 545, + 669 + ], + "type": "text", + "content": "[31] Yiyi Liao, Katja Schwarz, Lars Mescheder, and Andreas Geiger. Towards unsupervised learning of generative models for 3D controllable image synthesis. In CVPR, 2020. 2" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "text", + "content": "[32] Lingjie Liu, Marc Habermann, Viktor Rudnev, Kripasindhu Sarkar, Jiatao Gu, and Christian Theobalt. Neural actor: Neural free-view synthesis of human actors with pose control. ACM Trans. on Graphics, 2021. 3" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "12822" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[33] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 161 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 161 + ], + "type": "text", + "content": "[34] Thu Nguyen-Phuoc, Chuan Li, Lucas Theis, Christian Richardt, and Yong-Liang Yang. HoloGAN: Unsupervised learning of 3D representations from natural images. In ICCV, 2019. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 287, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 287, + 206 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 287, + 206 + ], + "type": "text", + "content": "[35] Thu Nguyen-Phuoc, Christian Richardt, Long Mai, Yong-Liang Yang, and Niloy Mitra. BlockGAN: Learning 3D object-aware scene representations from unlabelled images. In NeurIPS, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 208, + 287, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 208, + 287, + 239 + ], + "spans": [ + { + "bbox": [ + 48, + 208, + 287, + 239 + ], + "type": "text", + "content": "[36] Michael Niemeyer and Andreas Geiger. Giraffe: Representing scenes as compositional generative neural feature fields. In CVPR, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 241, + 287, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 241, + 287, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 241, + 287, + 285 + ], + "type": "text", + "content": "[37] Roy Or-El, Xuan Luo, Mengyi Shan, Eli Shechtman, Jeong Joon Park, and Ira Kemelmacher-Shlizerman. Stylesdf: High-resolution 3d-consistent image and geometry generation. CVPR, 2022. 2, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 286, + 287, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 286, + 287, + 330 + ], + "spans": [ + { + "bbox": [ + 48, + 286, + 287, + 330 + ], + "type": "text", + "content": "[38] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In ICCV, 2021. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 331, + 287, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 331, + 287, + 387 + ], + "spans": [ + { + "bbox": [ + 48, + 331, + 287, + 387 + ], + "type": "text", + "content": "[39] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. ACM Transactions on Graphics, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 388, + 287, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 287, + 442 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 287, + 442 + ], + "type": "text", + "content": "[40] Pascal Paysan, Reinhard Knothe, Brian Amberg, Sami Romdhani, and Thomas Vetter. A 3d face model for pose and illumination invariant face recognition. In 2009 sixth IEEE international conference on advanced video and signal based surveillance, pages 296-301. IEEE, 2009. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 444, + 287, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 444, + 287, + 476 + ], + "spans": [ + { + "bbox": [ + 48, + 444, + 287, + 476 + ], + "type": "text", + "content": "[41] Jingtan Piao, Keqiang Sun, Quan Wang, Kwan-Yee Lin, and Hongsheng Li. Inverting generative adversarial renderer for face reconstruction. In cvpr, pages 15619-15628, 2021. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 477, + 287, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 477, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 48, + 477, + 287, + 510 + ], + "type": "text", + "content": "[42] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In CVPR, 2021. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 511, + 287, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 511, + 287, + 554 + ], + "spans": [ + { + "bbox": [ + 48, + 511, + 287, + 554 + ], + "type": "text", + "content": "[43] Daniel Roich, Ron Mokady, Amit H Bermano, and Daniel Cohen-Or. Pivotal tuning for latent-based editing of real images. ACM Transactions on Graphics (TOG), 42(1):1-13, 2022. 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 555, + 287, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 555, + 287, + 589 + ], + "spans": [ + { + "bbox": [ + 48, + 555, + 287, + 589 + ], + "type": "text", + "content": "[44] Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. Graf: Generative radiance fields for 3d-aware image synthesis. NeurIPS, 2020. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 590, + 287, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 287, + 622 + ], + "type": "text", + "content": "[45] Yichun Shi, Divyansh Aggarwal, and Anil K Jain. Lifting 2d stylegan for 3d-aware face generation. In cvpr, pages 6258-6266, 2021. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 624, + 287, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 666 + ], + "type": "text", + "content": "[46] Yichun Shi, Xiao Yang, Yangyue Wan, and Xiaohui Shen. Semanticstylegan: Learning compositional generative priors for controllable image synthesis and editing. In CVPR, 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "type": "text", + "content": "[47] Alon Shoshan, Nadav Bhonker, Igor Kviatkovsky, and Gerard Medioni. Gan-control: Explicitly controllable gans. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14083-14093, 2021. 6, 7" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[48] Ivan Skorokhodov, Sergey Tulyakov, Yiqun Wang, and Peter Wonka. Epigraph: Rethinking training of 3d gans. arXiv preprint arXiv:2206.10535, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 150 + ], + "type": "text", + "content": "[49] Keqiang Sun, Shangzhe Wu, Zhaoyang Huang, Ning Zhang, Quan Wang, and HongSheng Li. Controllable 3d face synthesis with conditional generative occupancy fields. arXiv preprint arXiv:2206.08361, 2022. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 152, + 545, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 184 + ], + "type": "text", + "content": "[50] Attila Szabó, Givi Meishvili, and Paolo Favaro. Unsupervised generative 3D shape learning from natural images. arXiv, 2019. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 186, + 545, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 229 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 229 + ], + "type": "text", + "content": "[51] Junshu Tang, Bo Zhang, Binxin Yang, Ting Zhang, Dong Chen, Lizhuang Ma, and Fang Wen. Explicitly controllable 3d-aware portrait generation. arXiv preprint arXiv:2209.05434, 2022. 3, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 231, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 231, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 231, + 545, + 274 + ], + "type": "text", + "content": "[52] Ayush Tewari, Mohamed Elgharib, Florian Bernard, Hans-Peter Seidel, Patrick Pérez, Michael Zollhöfer, and Christian Theobalt. Pie: Portrait image embedding for semantic control. sigg, 39(6):1-14, 2020. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 276, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 276, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 307, + 276, + 545, + 330 + ], + "type": "text", + "content": "[53] Ayush Tewari, Mohamed Elgharib, Gaurav Bharaj, Florian Bernard, Hans-Peter Seidel, Patrick Pérez, Michael Zollhofer, and Christian Theobalt. Stylerig: Rigging stylegan for 3d control over portrait images. In cvpr, pages 6142-6151, 2020. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 331, + 545, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 331, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 307, + 331, + 545, + 365 + ], + "type": "text", + "content": "[54] Omer Tov, Yuval Alaluf, Yotam Nitzan, Or Patashnik, and Daniel Cohen-Or. Designing an encoder for stylegan image manipulation. ACM Trans. on Graphics, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 365, + 545, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 545, + 419 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 545, + 419 + ], + "type": "text", + "content": "[55] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Christoph Lassner, and Christian Theobalt. Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In ICCV, 2021. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 421, + 545, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 421, + 545, + 465 + ], + "spans": [ + { + "bbox": [ + 307, + 421, + 545, + 465 + ], + "type": "text", + "content": "[56] Ziyan Wang, Timur Bagautdinov, Stephen Lombardi, Tomas Simon, Jason Saragih, Jessica Hodgins, and Michael Zollhofer. Learning compositional radiance fields of dynamic human heads. In cvpr, pages 5704-5713, 2021. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 467, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 467, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 307, + 467, + 545, + 510 + ], + "type": "text", + "content": "[57] Jiajun Wu, Chengkai Zhang, Tianfan Xue, William T. Freeman, and Joshua B. Tenenbaum. Learning a probabilistic latent space of object shapes via 3D generative-adversarial modeling. In NeurIPS, 2016. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 511, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 511, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 307, + 511, + 545, + 555 + ], + "type": "text", + "content": "[58] Yue Wu, Yu Deng, Jiaolong Yang, Fangyun Wei, Qifeng Chen, and Xin Tong. Anifacegan: Animatable 3d-aware face image generation for video avatars. arXiv preprint arXiv:2210.06465, 2022. 3, 6, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 556, + 545, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 556, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 307, + 556, + 545, + 611 + ], + "type": "text", + "content": "[59] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond. In Computer Graphics Forum, volume 41, pages 641-676. Wiley Online Library, 2022. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 612, + 545, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 612, + 545, + 644 + ], + "spans": [ + { + "bbox": [ + 307, + 612, + 545, + 644 + ], + "type": "text", + "content": "[60] Hongyi Xu, Thiemo Alldieck, and Cristian Sminchisescu. H-nerf: Neural radiance fields for rendering and temporal reconstruction of humans in motion. NeurIPS, 2021. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 647, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 678 + ], + "type": "text", + "content": "[61] Xudong Xu, Xingang Pan, Dahua Lin, and Bo Dai. Generative occupancy fields for 3d surface-aware image synthesis. NeurIPS, 34:20683-20695, 2021. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 680, + 545, + 712 + ], + "type": "text", + "content": "[62] Yang Xue, Yuheng Li, Krishna Kumar Singh, and Yong Jae Lee. Giraffe hd: A high-resolution 3d-aware generative model. In CVPR, 2022. 2" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "12823" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 319 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "type": "text", + "content": "[63] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. In NeurIPS, 2021. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 106, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 106, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 106, + 287, + 162 + ], + "type": "text", + "content": "[64] Jianfeng Zhang, Zihang Jiang, Dingdong Yang, Hongyi Xu, Yichun Shi, Guoxian Song, Zhongcong Xu, Xinchao Wang, and Jiashi Feng. Avatargen: a 3d generative model for animatable human avatars. arXiv preprint arXiv:2208.00561, 2022.3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 162, + 288, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 162, + 288, + 207 + ], + "spans": [ + { + "bbox": [ + 48, + 162, + 288, + 207 + ], + "type": "text", + "content": "[65] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In cvpr, pages 586-595, 2018. 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 207, + 288, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 207, + 288, + 252 + ], + "spans": [ + { + "bbox": [ + 48, + 207, + 288, + 252 + ], + "type": "text", + "content": "[66] Yufeng Zheng, Victoria Fernández Abrevaya, Marcel C Bühler, Xu Chen, Michael J Black, and Otmar Hilliges. Im avatar: Implicit morphable head avatars from videos. In cvpr, pages 13545-13555, 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 252, + 288, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 252, + 288, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 252, + 288, + 285 + ], + "type": "text", + "content": "[67] Peng Zhou, Lingxi Xie, Bingbing Ni, and Qi Tian. CIPS-3D: A 3D-Aware Generator of GANs Based on Conditionally-Independent Pixel Synthesis. arXiv, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 285, + 288, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 285, + 288, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 285, + 288, + 319 + ], + "type": "text", + "content": "[68] Yiyu Zhuang, Hao Zhu, Xusen Sun, and Xun Cao. Mofan-erf: Morphable facial neural radiance field. arXiv preprint arXiv:2112.02308, 2021. 3" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "12824" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/OmniCity_ Omnipotent City Understanding With Multi-Level and Multi-View Images/1ecc869d-a020-42e0-bc1d-485949718b11_content_list.json b/2023/OmniCity_ Omnipotent City Understanding With Multi-Level and Multi-View Images/1ecc869d-a020-42e0-bc1d-485949718b11_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..211176072ddab8adbbe0a991b5df468bcf341614 --- /dev/null +++ b/2023/OmniCity_ Omnipotent City Understanding With Multi-Level and Multi-View Images/1ecc869d-a020-42e0-bc1d-485949718b11_content_list.json @@ -0,0 +1,1609 @@ +[ + { + "type": "text", + "text": "OmniCity: Omnipotent City Understanding with Multi-level and Multi-view Images", + "text_level": 1, + "bbox": [ + 261, + 130, + 709, + 175 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Weijia $\\mathsf{Li}^1$ , Yawen $\\mathsf{Lai}^2$ , Linning $\\mathsf{Xu}^3$ , Yuanbo $\\mathsf{Xiangli}^3$ , Jinhua $\\mathsf{Yu}^1$ ,", + "bbox": [ + 217, + 202, + 751, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Conghui He $^{2,4*}$ , Gui-Song Xia $^{5*}$ , Dahua Lin $^{3,4}$", + "bbox": [ + 305, + 220, + 661, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Sun Yat-Sen University, 2SenseTime Research, 3The Chinese University of Hong Kong,", + "bbox": [ + 135, + 238, + 834, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{4}$ Shanghai Artificial Intelligence Laboratory, $^{5}$ Wuhan University", + "bbox": [ + 230, + 256, + 738, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "liweij29@mail.sysu.edu.cn, alanlyawen@gmail.com, {x1020,xy019,dhlin}@ie.cuhk.edu.hk,", + "bbox": [ + 114, + 277, + 851, + 291 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "yujh56@mail2.sysu.edu.cn, heconghui@sensetime.com, guisong.xia@whu.edu.cn", + "bbox": [ + 156, + 295, + 803, + 309 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/17ab3d5166ad6ac2bd9135a4b1d471e714e762c8ce33ac287a2caadc2464873d.jpg", + "image_caption": [ + "Selected regions for image collection" + ], + "image_footnote": [], + "bbox": [ + 86, + 324, + 267, + 459 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8e739941d17d503c131a3a59c44244dacb429df05aefc4369495a32451a2e410.jpg", + "image_caption": [ + "Existing label maps and street-view", + "image viewpoints of a zoomed area" + ], + "image_footnote": [], + "bbox": [ + 89, + 470, + 267, + 606 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/df3096eba4b975b6853f5d50bfba1bda95ab2acbf866ed0745c1d15fe2e5ba9e.jpg", + "image_caption": [ + "Different satellite-level views" + ], + "image_footnote": [], + "bbox": [ + 285, + 325, + 426, + 412 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9baf49218cb34b86d8791dca9c85811d9039793c833f415a093c12d45dd68be1.jpg", + "image_caption": [ + "Transformation relations" + ], + "image_footnote": [], + "bbox": [ + 287, + 425, + 426, + 511 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6745b714f3acb034ea251eed71f2dd5b02da49c9c6df58714db8ba25f53e1b1b.jpg", + "image_caption": [ + "Different street-level views", + "Figure 1. An overview of our proposed OmniCity dataset. The satellite and street-level images of our dataset are collected in the six selected regions. The black dots along the streets denote the viewpoint locations and the land-lot colors indicate different fine-grained categories provided by the existing label maps. Note that all the images in the right part correspond to the same geo-location, and the intrinsic transformation relation between the satellite and street-level panorama view is demonstrated in the middle part." + ], + "image_footnote": [], + "bbox": [ + 285, + 525, + 424, + 612 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b0ed4e5dc56c5a0b4c188d8b334b72d54421f6bd02944c6245a0d6cf2c89a721.jpg", + "image_caption": [ + "Satellite-level images and annotations of multiple views" + ], + "image_footnote": [], + "bbox": [ + 444, + 325, + 883, + 411 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/73467ef54011833bcecb43d7d56ad6ca5320de447ed75077efae653eb7ef2902.jpg", + "image_caption": [ + "Street-level images and annotations of a panorama view" + ], + "image_footnote": [], + "bbox": [ + 442, + 424, + 883, + 510 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f6643e67e7001e20fcd2b4e68a133cce39629038f536021493c32683be155a65.jpg", + "image_caption": [ + "Street-level images and annotations of multiple mono-views" + ], + "image_footnote": [], + "bbox": [ + 442, + 525, + 883, + 612 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 728, + 313, + 744 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This paper presents OmniCity, a new dataset for omnipotent city understanding from multi-level and multi-view images. More precisely, OmniCity contains multi-view satellite images as well as street-level panorama and monoview images, constituting over 100K pixel-wise annotated images that are well-aligned and collected from 25K geolocations in New York City. To alleviate the substantial", + "bbox": [ + 75, + 762, + 470, + 867 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "pixel-wise annotation efforts, we propose an efficient street-view image annotation pipeline that leverages the existing label maps of satellite view and the transformation relations between different views (satellite, panorama, and mono-view). With the new OmniCity dataset, we provide benchmarks for a variety of tasks including building footprint extraction, height estimation, and building plane/instance/fine-grained segmentation. Compared with existing multi-level and multi-view benchmarks, OmniCity contains a larger number of images with richer annotation types and more views, provides more benchmark results", + "bbox": [ + 500, + 729, + 893, + 895 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding authors.", + "bbox": [ + 94, + 887, + 227, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "17397", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "of state-of-the-art models, and introduces a new task for fine-grained building instance segmentation on street-level panorama images. Moreover, OmniCity provides new problem settings for existing tasks, such as cross-view image matching, synthesis, segmentation, detection, etc., and facilitates the developing of new methods for large-scale city understanding, reconstruction, and simulation. The OmniCity dataset as well as the benchmarks will be released at https://city-super.github.io/omnicity/.", + "bbox": [ + 73, + 90, + 470, + 227 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 256, + 209, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Owning over a half global population and contributing the most economic growth, the city areas have been recorded and characterized by various data sources including satellite and aerial imagery, street-level imagery, LiDAR data, public maps, crowd-sourced data, etc. A great number of benchmarks have been proposed towards facilitating different vision tasks in city scene, of which the street-level imagery has been broadly used in multiple driving-related benchmarks [4, 10, 14, 18, 26, 31, 43]. The rich visual information of street-level imagery also enables complicated visual recognition tasks on specific categories, such as person detection [3], vehicle tracking and re-identification [32], and fine-grained land use classification [49].", + "bbox": [ + 73, + 281, + 468, + 477 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Nevertheless, constructing pixel-wise annotations for street-level imagery requires substantial human efforts, resulting in the small image quantity and limited view types of existing datasets, especially for the street-view panorama datasets with only hundreds of annotated images [36, 41, 42]. Regarding the annotation categories and levels, existing datasets mostly provide instance-level annotations for dynamic object categories in driving scenes. As a vital component for city understanding, the static objects such as buildings and roads take up a larger proportion of cities and remain a high consistency across the satellite and ground-level images. However, existing street-level datasets either provide pixel-wise building annotations without fine-grained semantic labels [4, 10, 14, 47] or provide fine-grained annotations at only bbox or image level [46, 49].", + "bbox": [ + 73, + 478, + 468, + 704 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Compared with street-level images, remote sensing images usually contain less visual information for conducting complicated tasks such as fine-grained land use segmentation and building function recognition. On the other hand, unlike the sparsely-distributed street-level images, remote sensing images have a dense spatial distribution and a worldwide coverage, which are well aligned with the open maps and government datasets at pixel level [22]. These existing maps and datasets contain a variety of satellite-level annotations for buildings (such as the footprint, land use, height, year built), roads (category and line coordinates), and other geographical objects, providing new perspectives for promoting novel city understanding datasets and tasks.", + "bbox": [ + 75, + 705, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, as illustrated in Figure 1, we construct an omnipotent city dataset unifying data sources from both satellite and street views, linked by geo-locations and urban planning data. Unlike existing city datasets that only support a limited number of tasks, OmniCity dataset incorporates rich geometric annotations and semantic meta data for each image, where multiple tasks can be conducted on. To leverage the existing map labels and the rich visual context from the street-level imagery, we propose an efficient pipeline for producing diverse street-level annotations. Based on this annotation pipeline, we built OmniCity, a dataset that contains over 100K annotated images collected from 25K geo-locations in New York City. We provide benchmark results on OmniCity for a variety of tasks, including building footprint extraction and height estimation on satellite images, as well as fine-grained/instance/plane segmentation of buildings on street-level panorama and mono-view images. To the best of our knowledge, this is the first work that involves fine-grained building instance segmentation on street-level panorama images. We also analyze the potential of OmniCity for promoting new tasks and methods with multi-level imagery.", + "bbox": [ + 496, + 90, + 890, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main contributions are summarized as follows:", + "bbox": [ + 517, + 422, + 857, + 436 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel pipeline for efficiently producing diverse pixel-wise annotations on street-level panorama and mono-view images.", + "- We build the OmniCity dataset, which contains well-aligned satellite and street-level images with a larger quantity, richer annotations and more views compared with existing datasets.", + "- We provide a series of benchmark experimental results for multiple tasks and data sources, and analyze the limitations of the current benchmarks on OmniCity.", + "- We discuss the potential of OmniCity for facilitating new methods and tasks for large-scale city understanding, reconstruction, and simulation." + ], + "bbox": [ + 519, + 450, + 890, + 675 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 500, + 702, + 635, + 715 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Datasets and methods for street-level tasks", + "text_level": 1, + "bbox": [ + 500, + 726, + 861, + 739 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As shown in Table 1, many street-level datasets have been proposed over the past few years. A large proportion of these datasets are designed for visual tasks in driving scene [4, 10, 14, 18, 26, 31, 44], such as 2D/3D object detection, semantic segmentation, object tracking, etc. Several street-level datasets are proposed for a specific object category, such as the EuroCity Persons dataset [3], the CityFlow dataset [32] for vehicle tracking and re-identification, etc. In addition to the above datasets containing only monoview images, some studies propose new datasets or methods", + "bbox": [ + 496, + 750, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "17398", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/9b15a05538a01d3533ba0aff045869826ae7cbf63b91e3356b9327b02169fda8.jpg", + "table_caption": [ + "Table 1. A comparison of OmniCity with existing city-related datasets of street-level, satellite-level and cross-level. The Street and Satellite (Sate.) columns show the available image view types. The last three columns indicate which level of tasks the dataset is designed for (semantic/instance/plane segmentation, object detection (bbox), and image classification), and whether the dataset contains fine-grained building attribute (Attri.) or height labels." + ], + "table_footnote": [], + "table_body": "
Dataset#ImagesStreetSate.Anno.Attri.Height
KITTI [14]15,000mono-semantic××
Cityscapes [10]25,000mono-semantic××
EuroCity [3]47,300mono-bbox××
WildPASS [42]500multi.-semantic××
PASS [41]400multi.-semantic××
HoliCity [47]6,300multi.-inst./plane××
SkyScapes [1]8,820-singlesemantic××
SpaceNet [38]60,000-multi.instance××
Christie et al. [9]11,000-singlesemantic×✓
Li et al. [21]3,300-singleinstance×✓
TorontoCity [36]Unknowmulti.multi.instance×✓
Wojna et al. [39]49,426monosingleimage✓×
OmniCity108,600multi.multi.inst./plane✓✓
", + "bbox": [ + 78, + 194, + 460, + 363 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "for semantic segmentation from panorama images, such as TorontoCity [36], PASS [41], and WildPASS [42]. However, such datasets require expensive annotation efforts and contains only hundreds of pixel-wise annotated panorama images in total. Besides the driving-related datasets, several recent studies target at the recognition tasks for static object categories. Zhu et al. [49] proposed a framework for fine-grained land use classification task using ground-level images. Zhang et al. [46] proposed a building recognition system for detecting their business entity information. HoliCity [47] proposed holistic 3D structure annotations generated from CAD models for panorama images, but it only conducted experiments on mono-view images and lacked the semantic information of buildings and roads.", + "bbox": [ + 75, + 377, + 468, + 588 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In summary, existing street-level datasets still have the following limitations. Regarding the image quantity and view types, most existing datasets require substantial annotation efforts and contain only a limited number of annotated images of a mono view. The image quantity of panorama datasets is even several orders of magnitude smaller than mono-view datasets. Regarding the annotation categories and levels, existing datasets mainly focus on dynamic or driving-related categories and lack in fine-grained annotations of an object category. Moreover, the datasets for static objects are lacking in pixel-level or fine-grained annotations. By contrast, our OmniCity contains over 100K satellite and street-level images of multiple views as well as the pixel-wise and fine-grained annotations, demonstrating superiority in annotation quantity, annotation type, and view type compared with existing datasets.", + "bbox": [ + 75, + 589, + 470, + 832 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Datasets and methods for satellite-level tasks", + "text_level": 1, + "bbox": [ + 76, + 845, + 455, + 861 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As a data source with a long time series and a large coverage, the satellite imagery has been broadly explored", + "bbox": [ + 75, + 869, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "for large-scale city understanding. Unlike the street-level datasets requiring manual annotations, the satellite imagery is already well aligned with existing label maps [20]. The OpenStreetMap (OSM) is one of the most broadly-used map data, which contains publicly available annotations of building footprints, heights, roads, land use, etc., of world wide. In addition, many public datasets provide rich information at a local scale. For example, the PLUTO dataset contains the block and lot information of the whole New York City, and each lot is associated with the land use, year built, number of floors, and other useful information. The Microsoft US building footprint dataset contains over a hundred million computer-generated building footprints. Several challenges provide manually labeled building footprints and the corresponding satellite images [11, 34, 38]. Moreover, some datasets provide fine-grained semantic categories of different objects [1] or building height annotations for 3D reconstruction tasks [9, 21].", + "bbox": [ + 496, + 90, + 890, + 362 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In summary, existing maps and datasets have provided substantial semantic and geometric information that is well aligned with the satellite imagery. In this work, we leverage these rich annotations and the transformation relations between different views to produce auxiliary information for street-level image annotation. Compared with existing work, OmniCity significantly reduces the human labeling efforts and provides more annotation types to enable omnipotent city understanding via multiple tasks and views.", + "bbox": [ + 496, + 363, + 890, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Datasets and methods for multi-level tasks", + "text_level": 1, + "bbox": [ + 498, + 508, + 859, + 523 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The ground-level images usually contain rich visual context that is not visible from the satellite or aerial imagery (e.g. building facade, the side of vegetations, etc.), while the spatial distribution is often sparse and unbalanced in different areas. By contrast, the remote sensing images have a much denser spatial distribution at a global scale, but the visual context is too limited to support complicated fine-grained tasks. Considering the above complementary characteristics, many datasets and methods have been proposed for cross-view scenarios. Cross-view visual recognition tasks integrate the two data types as the model input for predicting building functions, building ages, land use, tree species, etc. [2, 12, 37, 40], while cross-view image matching [17, 28, 33, 48] and image synthesis [25, 27, 30, 45] tasks take only one data type as the model input.", + "bbox": [ + 496, + 532, + 890, + 758 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In addition, several datasets are proposed for a variety of satellite and street-level tasks. The TorontoCity [36] contains a wide range of annotations including building height estimation, building instance segmentation, building footprint segmentation, road segmentation, etc., which are conducted on either satellite or street-level images and only adopt FCN and some CNN classification architectures as", + "bbox": [ + 496, + 758, + 890, + 866 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "1https://data.cityofnewyork.us/City-Government/Primary-Land-Use-Tax-Lot-Output-PLUTO-/64uk-42ks.", + "bbox": [ + 500, + 875, + 890, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "17399", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/7b23f45f7c52680205d1d6a47ce7f9a8b80cd7930a169cf4ef77f6e75d676507.jpg", + "image_caption": [ + "Figure 2. The annotation tool of OmniCity dataset. The left and right images show examples of one road scene with the label map (in the bottom-right window) displayed in land use mode, and crossroads scene with the label map in block-lot mode, respectively." + ], + "image_footnote": [], + "bbox": [ + 80, + 89, + 893, + 237 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the baseline models for all experiments. In a recent study [39], a holistic multi-view building analysis dataset is designed for multiple recognition tasks regarding facade material, number of floors, occupancy type, roof geometry, roof pitch, and construction type. The dataset contains only image-wise annotations for street-level classification tasks. The semantic categories of each building are decided by human annotators, resulting in more challenges and subjectivity to the annotation process compared with our work.", + "bbox": [ + 75, + 290, + 472, + 428 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In summary, for the existing cross-view studies designed for one specific task, the street-level images are not annotated and only supplement extra features to the satellite-level feature maps. By contrast, our OmniCity provides rich annotation types that are not contained in existing datasets, e.g. the pixel-wise building instances and fine-grained categories on street-level images, which can promote new methods to explore and leverage these annotations to improve the performance. In addition, OmniCity provides the benchmark experimental results of each task using more state-of-the-art models. With our efficient annotation pipeline, it also provides additional fine-grained pixel-wise annotations and benchmark results compared with [36] and [39].", + "bbox": [ + 75, + 438, + 472, + 636 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. Datasets", + "text_level": 1, + "bbox": [ + 76, + 681, + 171, + 698 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this work, we aim at building a dataset for omnipotent city understanding from multi-level and multi-view images. Our proposed OmniCity dataset contains 108,600 images of multiple views, which are collected from 25K geo-locations in New York City. Compared with existing datasets, OmniCity requires much fewer human efforts for street-level image annotation, contains more diverse annotation types for both 2D and 3D tasks, provides richer building semantics at instance segmentation level, and possesses higher scalability for new annotation type supplement and expansion to other cities. The details of data collection, annotation, and statistics are introduced as follows.", + "bbox": [ + 75, + 719, + 472, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Data collection", + "text_level": 1, + "bbox": [ + 500, + 290, + 650, + 306 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Figure 1, our OmniCity dataset is collected from six selected regions of New York. We download the panorama images in the six selected regions using google street view download 360, with a step distance of 65 meters. The regions for collecting the training and test samples are denoted by red and blue, respectively. We save the geographic coordinates, collection time, panorama id, north rotation, and zoom level for each panorama image. For each panorama site, we collect its corresponding google earth images of three acquisition dates according to the geographic coordinates, constituting three groups of satellite-level datasets with three cases of off-nadir view angles (small/medium/high, denoted by V1/V2/V3). For the annotation data sources, we collect the meta information from PLUTO and OpenStreetMap (OSM). The New York City is hierarchically formed by blocks, lots, and buildings. Each building can be identified by a specific block-lot id. In PLUTO, each lot (building) is associated with rich information, e.g. land use, year built, number of floors, etc. The OSM data contains footprint and height information while lacking in land use information for most buildings. Considering the characteristics of the two data sources, we align the land use attribute (from PLUTO) with the building footprint and height (from OSM) using the geographical coordinates. Overall, each building is assigned with a block-lot id, a land use category, a height value, and the geographical coordinates of a footprint polygon, which will be used as the reference label maps for panorama image annotation.", + "bbox": [ + 496, + 315, + 893, + 739 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Data annotation", + "text_level": 1, + "bbox": [ + 500, + 753, + 660, + 768 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Figure 2 shows the panorama image annotation tool proposed in our study. The street-level panorama and satellite images are naturally well-aligned at image-level between the central coordinates from satellite views and the camera pose of panorama views. The instance-level annotations of different views (e.g., building facade and footprint) are further aligned during the annotation process. According to the geo-transformation relation between different views [25],", + "bbox": [ + 496, + 780, + 893, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "17400", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "we calculate the angle between the north (upward) direction and each building footprint (in the satellite image), and the angle between the north direction (saved in the panorama file) and each building facade in the horizontal axis. The footprint and facade with the same angle value are aligned to the same building, providing the auxiliary information for the manual annotation process.", + "bbox": [ + 75, + 90, + 470, + 196 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The annotation pipeline includes four stages: (1) Image selection, i.e., select the panorama images that are essential to be annotated according to building coverage, occlusion extent, etc; (2) Segmentation annotation, i.e., adjust the floor/top line to fit the bottom/roof of each building, and add the boundary split line considering both auxiliary information and building appearance (e.g. texture discrepancy, doors, etc.); (3) Attribute assignment, i.e., add the attributes (instance ID, block-lot id and land use type) for each building plane; (4) Quality assessment, i.e., check the annotation quality and remove the unqualified images.", + "bbox": [ + 75, + 198, + 468, + 364 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In addition, the mono-view images and annotations are automatically generated from those of the panorama images via view transformation. For each panorama image, we select three views using three x-axis angles (-170, 10 and 170, in the range of [-180, 180]) and a fixed y-axis angle of 0. Then we design an image selection rule to filter out the unexpected images of which the buildings are distributed in only one side, building area proportion is smaller than $10\\%$ , or the building plane quantity is smaller than 2. The remaining images constitute the final mono-view dataset of our experiments. For satellite-level tasks, the annotations are already well-aligned with the meta information from PLUTO and OSM. Each building footprint on a satellite image is assigned with a block-lot id, a land use category, a height value, and the pixel coordinates of the footprint polygon.", + "bbox": [ + 75, + 364, + 470, + 592 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Statistics of the proposed dataset", + "text_level": 1, + "bbox": [ + 76, + 604, + 364, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The whole dataset contains three sub-datasets (in small, medium and large view angles) for satellite-level tasks and two sub-datasets (in panorama and mono views) for street-level tasks. For satellite-level datasets, the three sub-datasets of images collected from the 25K geo-locations constitute 75K images in total, which are cropped by $512 \\times 512$ pixels according to the coverage area of the corresponding panorama image. For the street-level datasets, the panorama dataset contains 18K images in $512 \\times 1024$ pixels, which are selected from the initial 25K panorama images during the image selection and quality assessment phases. Similarly, the mono-view dataset obtained via view transformation and image selection contains 156,00 monoview images in $512 \\times 512$ pixels. The ratio of train/test splits is set as 4:1 for all five sub-datasets. In OmniCity dataset, the initial land use categories of PLUTO with similar characteristics and low quantities are merged into one category, resulting in 7 land use categories in total. The", + "bbox": [ + 75, + 628, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "categories of $1/2$ family building ( $\\sim 17\\%$ ), walk-up building ( $\\sim 23\\%$ ) and the mixed residential/commercial building categories ( $\\sim 29\\%$ ) take up a larger proportion compared with the other four categories, i.e., elevator buildings ( $\\sim 7\\%$ ), office buildings ( $\\sim 10\\%$ ), industrial/transportation/utility buildings ( $\\sim 11\\%$ ) and others ( $\\sim 1\\%$ ). For the distribution of building height, most are between 1 and 25 meters while a small percentage ( $\\sim 1\\%$ ) reaches over 50 meters.", + "bbox": [ + 496, + 90, + 890, + 212 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Benchmark results", + "text_level": 1, + "bbox": [ + 498, + 229, + 683, + 244 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we provide a variety of benchmarks for multiple satellite and street-level tasks. The satellite-level tasks in our experiments include building footprint segmentation and height estimation. For both tasks, we conduct experiments on the satellite images with three view angles. For the street-level tasks, we conduct two instance segmentation tasks (i.e., land use and building instance segmentation) on the panorama images, and three instance segmentation tasks (i.e., land use / building instance / plane segmentation) on mono-view images. Please note that these are only preliminary experimental results on OmniCity dataset. More benchmarks of latest models and additional tasks will be continuously updated on OmniCity homepage.", + "bbox": [ + 496, + 257, + 890, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental setting", + "text_level": 1, + "bbox": [ + 498, + 468, + 697, + 484 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The experiments are mainly based on mmdetection [7] with the recommended hyper-parameter settings. We select Mask R-CNN [15] as the baseline method for segmentation tasks, and provide a comparison of different methods in Table 6, including Mask Scoring R-CNN (MS R-CNN) [19], Cascade Mask R-CNN (Cascade) [5], Content-Aware ReAssembly of FEatures (CARAFE) [35], and Hybrid Task Cascade (HTC) [6]. Specifically, we use ResNet-50 [16] with FPN [23] pre-trained on the ImageNet [29] as the backbone for all instance segmentation models. All models are trained on 8 NVIDIA Tesla V100 GPUs for 12 epochs, with a batch size of 16, a learning rate starting from 0.02 and decreasing by a factor of 0.1 from the $8^{th}$ to $11^{th}$ epoch, and the stochastic gradient descent (SGD) optimizer with a weight decay of $10^{-4}$ and a momentum of 0.9. For the height estimation task, we evaluate the performance of two widely-used monocular depth estimation methods on the satellite images of three view angles, i.e. Structure-Aware Residual Pyramid Network (SARPN) [8] and Deep Ordinal Regression Network (DORN) [13], which are trained on 4 NVIDIA Tesla V100 GPUs for 20 epochs. SARPN is trained with a batch size of 8, a learning rate starting from $10^{-4}$ and reduced by $10\\%$ every 5 epochs, and the Adam optimizer with $\\beta_{1} = 0.9$ , $\\beta_{2} = 0.999$ , and a weight decay of $10^{-4}$ . DORN is trained with a batch size of 4, a base learning rate of $10^{-4}$ and the power of 0.9, using SGD optimizer with a weight decay of 0.0005 and a momentum of 0.9.", + "bbox": [ + 496, + 493, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "17401", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e9d32052e9ca393c4f493081008bbb5ac67b6e756b5bc6a5058f74d9a2e20232.jpg", + "image_caption": [ + "Small view" + ], + "image_footnote": [], + "bbox": [ + 80, + 88, + 189, + 172 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2f658c1f3183d8110deec368391d8b8cac7c97b5d6d1b55270d41dc716fc88ee.jpg", + "image_caption": [ + "Medium view" + ], + "image_footnote": [], + "bbox": [ + 196, + 88, + 305, + 172 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/46baf51ab398638283631313bf8346e7752c3e33802b373f441d13fd9c4230b1.jpg", + "image_caption": [ + "Large view" + ], + "image_footnote": [], + "bbox": [ + 310, + 88, + 421, + 172 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0659340c6af97b34b589185c5a4a9a6896bb1bc4b9db2bb2f6325ca6dc12697d.jpg", + "image_caption": [ + "Satellite image" + ], + "image_footnote": [], + "bbox": [ + 80, + 186, + 189, + 272 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/73cc5a7d27527e98d435fac9e135f6d1de42946dcf42884ce9b62cc49c3992a7.jpg", + "image_caption": [ + "Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 196, + 186, + 305, + 272 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/93db0a7f5a73a502623284e9df92ef96db26b3ff6f289dcc57d285f7cbabad66.jpg", + "image_caption": [ + "DORN - small" + ], + "image_footnote": [], + "bbox": [ + 310, + 186, + 421, + 272 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/60ec9f7b70b57524dd1a16989d1f1a0e074a4793a2e388d10606610a238e9eed.jpg", + "image_caption": [ + "SARPN - small", + "Figure 3. Qualitative results of satellite-level tasks. The yellow/cyan/red polygons denote TP/FP/FN buildings." + ], + "image_footnote": [], + "bbox": [ + 78, + 287, + 189, + 375 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3eea1ca0371f98394bb37b58009e6f89d3aa955a02c71452177579b2ecf06f06.jpg", + "image_caption": [ + "SARPN - medium" + ], + "image_footnote": [], + "bbox": [ + 194, + 287, + 305, + 375 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8dc71005b97332c4f13564bb4f8e5f1faf3240499b3806775d41e2261bb93cb6.jpg", + "image_caption": [ + "SARPN - large" + ], + "image_footnote": [], + "bbox": [ + 310, + 287, + 421, + 375 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4ea60c47fe3bc4e1b3b546a2a26a9c8f261764e2ab0b228c57fbd8e5b4fbfde1.jpg", + "image_caption": [ + "High", + "Low" + ], + "image_footnote": [], + "bbox": [ + 426, + 290, + 455, + 375 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Experimental results of satellite-level tasks", + "text_level": 1, + "bbox": [ + 76, + 450, + 439, + 465 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2 shows the baseline results of the satellite-level instance segmentation task, which are evaluated using both COCO [24] and SpaceNet [11] evaluation metrics. The footprint segmentation performance is the best for satellite images with a small view angle, and deteriorates seriously when the view angle gets larger. For all three cases, the prediction score is the highest for buildings with large areas and the lowest for small buildings. Figure 3 provides a qualitative comparison of the footprint segmentation results on satellite images of three types of view angle. The large view angle results in great difficulties for extracting accurate footprint boundaries, due to the partial invisibility of building footprint, the serious shadow effects, etc. The height estimation performance is evaluated in terms of the mean absolute error (denoted by MAE), mean square error (denoted by MSE), and root mean square error (denoted by RMSE), which are commonly used metrics for depth estimation. All metrics are measured in meters at pixel level. Table 3 and Figure 3 show the height estimation results obtained from DORN [13] and SARPN [8] for satellite images with different view angles. Results demonstrate that DORN obtains better results compared with SARPN for all three cases. Both methods achieve the best performance for satellite images with a medium view angle (V2) compared with the other two cases. The footprint and roof have more overlaps and provide less building structure information for satellite images with a small view angle, while the shadow and parallax effect become more serious with", + "bbox": [ + 76, + 477, + 472, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2ed9af556c83367067b0d234c6d8ea3bcce0d06f4255dda2ce0c878db2044b87.jpg", + "table_caption": [ + "Table 2. Quantitative results of instance segmentation for satellite images with different view angles." + ], + "table_footnote": [], + "table_body": "
ViewMetrics of various thresholdsthreshold = 0.5
APAP50AP75APSAPMAPLPRF1
V129.766.023.515.933.936.776.966.371.2
V223.756.616.111.527.230.373.955.063.1
V318.951.49.69.121.525.370.751.759.7
", + "bbox": [ + 501, + 128, + 890, + 186 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/dc34dc59d2ea834c52840843e2af362b5db8344523e709d10dc904ba8e1f0573.jpg", + "table_caption": [ + "Table 3. Quantitative results of height estimation for satellite images with different view angles." + ], + "table_footnote": [], + "table_body": "
ViewSARPN [8]DORN [13]
MAEMSERMSEMAEMSERMSE
V116.18870.3429.5012.71670.5225.89
V213.75694.1726.3512.24628.0625.06
V315.32823.0128.6913.40730.6727.03
", + "bbox": [ + 519, + 234, + 870, + 300 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the increase of off-nadir view angle. The above aspects result in challenges for the accurate estimation of building height for satellite images with small and large view angles.", + "bbox": [ + 498, + 316, + 890, + 364 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Experimental results of street-level tasks", + "text_level": 1, + "bbox": [ + 500, + 376, + 846, + 392 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We analyze the performance of multiple segmentation tasks on street-level panorama and mono-view images. Table 4 and Figure 4 show the experimental results of panorama-view images on two segmentation tasks, of which the performance on building instance segmentation task (denoted by Instance Seg.) is significantly superior to the fine-grained land use segmentation task (denoted by Landuse Seg.). Table 5 and Figure 5 show the experimental results of mono-view images on three different tasks, i.e., landuse segmentation, instance segmentation, and plane segmentation. Similar to the results of panorama-view images, the baseline method achieves much higher scores for the two binary segmentation tasks (plane and instance segmentation) compared with the fine-grained land use segmentation task. The qualitative results also demonstrate that the baseline method has difficulties in identifying the accurate land use type of some building instances.", + "bbox": [ + 496, + 400, + 890, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 6 shows the land use segmentation results obtained from the five methods on street-level panorama images. HTC achieves the best performance for both overall and category metrics followed by MS R-CNN, which indicates that the cascade structure and mask scoring strategy can effectively improves the fine-grained segmentation performance of building instances. From the metrics of each category, we can find that C2, C4 and C5 (i.e., Walkup Buildings, Mixed Residential/Commercial, and Office Buildings) have better performance compared with C1, C3 and C6 (i.e., 1/2 Family Buildings, Elevator Buildings and Industrial/Transportation/Utility) for all methods, with the AP ranging from $20\\%$ to $40\\%$ . The category of Others has an extremely low AP score due to the small sample quantity. Compared with the residential buildings, the mixed Residential/Commercial and Office buildings have more special", + "bbox": [ + 496, + 659, + 892, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "17402", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b541a0ef87e31a184621fd24b40c1f94531f444feeae28e00863e223bbc3ced9.jpg", + "image_caption": [ + "Figure 4. Qualitative results of street-level tasks (i.e., instance segmentation and land use segmentation) on panorama images." + ], + "image_footnote": [], + "bbox": [ + 81, + 90, + 467, + 277 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/adcc651fbc976aba9cf00954caea7749ef10c377e87c9e7ef70a8600a57cc2ac.jpg", + "table_caption": [ + "Table 4. Quantitative results on street-level panorama images." + ], + "table_footnote": [], + "table_body": "
TaskAPAP50AP75APSAPMAPL
Landuse Seg.26.034.728.50.312.030.4
Instance Seg.66.786.572.51.740.274.1
", + "bbox": [ + 78, + 356, + 468, + 407 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/cf39b3dd9273127ad26bf077c76d709ee9c73f716031d667b4b5c589e0179db5.jpg", + "table_caption": [ + "Table 5. Quantitative results on street-level mono-view images." + ], + "table_footnote": [], + "table_body": "
TaskAPAP50AP75APSAPMAPL
Landuse Seg.23.932.126.70.310.627.5
Instance Seg.68.388.873.83.233.376.1
Plane Seg.65.187.471.05.040.773.8
", + "bbox": [ + 78, + 445, + 468, + 513 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "characteristics (i.e. the boundary between first and above floors, facade design, and building structure), contributing to the superior AP scores of these two categories. The performance of the above six categories is also concordant with the ratio of the sample quantity.", + "bbox": [ + 75, + 535, + 468, + 609 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Results analysis and discussions", + "text_level": 1, + "bbox": [ + 76, + 619, + 357, + 636 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section we summarize the limitations of existing methods for satellite and street-level tasks on our OmniCity dataset. For building footprint segmentation, the performance of existing methods get worse with the increasing of off-nadir view angle, which might due to the serious parallax and shadow effect on the satellite images with a large off-nadir view angle. For the height estimation task, most existing methods directly apply the monocular depth estimation methods to remote sensing scene. These methods usually produce poor results on the invisible side of the footprint boundary especially for high-rise buildings on very off-nadir images. In addition, most existing methods use deep neural networks to regress continuous values but the actual height values are discrete for building and non-building areas, resulting in difficulties for network training and extra efforts for converting the continuous prediction values into discrete height values via post-processing.", + "bbox": [ + 75, + 643, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3375a710fbf8843139a54b93e66df7291a245b83620c5ffd80638b67ef8a0f6a.jpg", + "image_caption": [ + "Figure 5. Qualitative results of street-level tasks (i.e., land use /instance/plane segmentation) on mono-view images." + ], + "image_footnote": [], + "bbox": [ + 506, + 90, + 890, + 335 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For the street-level tasks based on panorama images, existing methods target at general instance segmentation tasks for commonly-used datasets, e.g. COCO [24], CityScapes [10], BDD100K [44], etc. These datasets often have a single view and a narrow Field of View (FoV). However, for panorama images, the special properties such as the wide FoV covering full 360-degree in the horizontal direction, are not taken into consideration in the design of existing methods. For both mono-view and panorama images, existing methods have difficulties in accurately recognizing the building instance with a small area (e.g. buildings located in the side of the main parts), the land use categories with a small number of building instances, and the categories that are easily confused (e.g. 1/2 Family Buildings and Walk-Up Buildings, Mixed Residential/Commercial and Office Buildings). Figure 6 shows some typical failure cases of the current benchmark methods. New instance segmentation methods should be designed for solving the above limitations considering the characteristics of panorama images, building instances, fine-grained categories, etc.", + "bbox": [ + 496, + 402, + 892, + 705 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Potential of the OmniCity dataset", + "text_level": 1, + "bbox": [ + 500, + 723, + 805, + 739 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our proposed OmniCity dataset demonstrates great potential for facilitating city understanding, machine perception, and generative modeling researches in many aspects.", + "bbox": [ + 496, + 750, + 890, + 794 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "First, it can serve as a new dataset for the existing tasks such as image geo-localization/synthesis and segmentation/detection of buildings/trees/land use from cross-view images. OmniCity provides additional annotations that are not contained in existing datasets, e.g. the building instances and fine-grained categories on street-level images, which can promote new methods to explore and leverage", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "17403", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2b84bd86f766e4bbb845c8646f38910e09187e3ce5b7536e01593f9f393d9be9.jpg", + "table_caption": [ + "Table 6. Quantitative results of different methods for fine-grained land use segmentation on street-level panorama images." + ], + "table_footnote": [], + "table_body": "
MethodOverall MetricsMetrics of each category
APAP50AP75APSAPMAPLC1C2C3C4C5C6C7
Mask R-CNN [15]26.034.728.50.312.030.419.637.525.839.236.922.20.8
MS R-CNN [19]27.135.829.80.112.431.522.539.126.240.838.021.71.2
Cascade [5]25.933.828.30.211.430.52038.32538.536.722.10.3
CARAFE [35]25.934.528.50.111.930.219.637.324.939.937.221.50.8
HTC [6]27.235.729.90.312.432.020.838.727.239.938.424.51.2
", + "bbox": [ + 178, + 117, + 787, + 219 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the new annotations to improve the performance of these tasks. The annotations are organized in a unified version, which means multiple tasks can be performed on a single image, and thus can well support the multi-task learning setting. Additionally, since the building instances are directly linked with the urban planning data using block-lot id, it is easy to be enriched with more annotation types from other urban datasets, especially those for social and urban studies.", + "bbox": [ + 75, + 233, + 468, + 354 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Second, OmniCity provides a new application scenario or problem setting to existing tasks. For line segment detection and wireframe parsing tasks, existing datasets contains densely distributed line segment and wireframe labels, while our OmniCity focuses on the main line segments and wireframes on the outlines instead of the inner ones, resulting in a much sparser format. The serious shelters from the trees and vehicles also bring challenges to these tasks. New line segment detection and wireframe parsing methods should be designed for the OmniCity scenario.", + "bbox": [ + 75, + 354, + 468, + 505 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Moreover, OmniCity facilitates new tasks for city reconstruction and simulation. Treating each panorama as a unique city scene, a complete 3D model representing such a local scene is available, as also shown in the second window panel in our tool (Figure 2). These 3D models are stored in abstract vector formats, with clean and clear vertices indicating the building facades, streets, etc., which are suitable for many shape generation tasks represented with graphs.", + "bbox": [ + 75, + 506, + 468, + 628 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Finally, with the well-aligned satellite and street-level images as well as the various annotation types, novel city reconstruction tasks, e.g., 3D building reconstruction from cross-view images, can be derived for producing holistic 3D buildings with both fine-grained semantic category and precise geometry information (vector 3D model). Unlike the existing studies that only target at 3D reconstruction from monocular or multi-view remote sensing imagery, new methods should be designed to leverage the additional information from street-level images to improve 3D reconstruction and semantic prediction for OmniCity scenario.", + "bbox": [ + 75, + 628, + 468, + 794 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 814, + 194, + 829 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we have proposed OmniCity, a new dataset for omnipotent city understanding from satellite and street-level images of multiple views. The dataset contains over 100K images collected from 25K geo-locations in New", + "bbox": [ + 76, + 839, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fd54acc1ace974a66bc0e83340c9b91f98a941b8feb15c81f8abd604c016c358.jpg", + "image_caption": [ + "(a)GT:5-Office" + ], + "image_footnote": [], + "bbox": [ + 501, + 229, + 617, + 306 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/79fdc6e4eba59310541e8264db6edf336b6e06c37139b2ab64270f26042b255e.jpg", + "image_caption": [ + "(b) GT: 4-Mixed (three instances)" + ], + "image_footnote": [], + "bbox": [ + 620, + 229, + 754, + 306 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3e01033128d2fa9375d63e7f704ee13055bb5c2f160aefd991d59ef773c584c7.jpg", + "image_caption": [ + "(c) GT: 3-Elevator" + ], + "image_footnote": [], + "bbox": [ + 756, + 231, + 890, + 306 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/39b456a200f887e68fcf2d2b55a9cff239d9a73d5192af0d151aba7a3775f76a.jpg", + "image_caption": [ + "(d)GT:5-Office", + "Figure 6. Typical failure cases of the current benchmark methods." + ], + "image_footnote": [], + "bbox": [ + 501, + 319, + 602, + 395 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bfa689068c84c71966380a96dc8822117e5cf5d3507c03dee0d289e2131bb475.jpg", + "image_caption": [ + "(e)GT:5-Office" + ], + "image_footnote": [], + "bbox": [ + 602, + 319, + 689, + 395 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/dfdfa8efe3b89c78e888ce0eb8b0b70d8ec7ec2051f7fb45517d2807648b2ef0.jpg", + "image_caption": [ + "(f) GT: 7-Others" + ], + "image_footnote": [], + "bbox": [ + 691, + 319, + 785, + 395 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8d633283896b7e9e409109f760c2b09c0a7742772a950226c652cc462a03d087.jpg", + "image_caption": [ + "(g) GT: 1-Family" + ], + "image_footnote": [], + "bbox": [ + 787, + 319, + 890, + 395 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "York City, of which the annotations are generated from both existing label maps and our proposed annotation pipeline. We provide benchmark experimental results for multiple tasks and data sources based on state-of-the-art methods and analyze their limitations. We believe that OmniCity will not only promote new algorithms and application scenarios for existing tasks, but facilitate novel tasks for 3D city reconstruction and simulation. In our future work, we will keep updating the OmniCity dataset and the benchmarks in the following aspects. Owing to the proposed annotation pipeline and the unified annotations with a vector format and rich meta information (geo-locations, block-lot id, etc.), OmniCity can be efficiently supplemented with more properties of buildings and other geographical object types (roads, sidewalks, trees, green space, etc.), and extended to other cities of different countries. The benchmark results of more state-of-the-art models and new tasks will be provided accordingly. Based on the rich annotation and view types of OmniCity, we also plan to develop new methods for existing and novel tasks, such as object detection, instance segmentation, and 3D reconstruction from cross-view images.", + "bbox": [ + 496, + 469, + 890, + 787 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This project was funded in part by National Natural Science Foundation of China (No. 42201358, No. U22B2011 and No. 41820104006), Shanghai AI Laboratory, CUHK Interdisciplinary AI Research Institute, and the Centre for Perceptual and Interactive Intelligence (CPII) Ltd under the Innovation and Technology Commission (ITC)'s InnoHK.", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "17404", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Seyed Majid Azimi, Corentin Henry, Lars Sommer, Arne Schumann, and Eleonora Vig. Skyscapes fine-grained semantic understanding of aerial scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7393-7403, 2019.", + "[2] Sara Beery, Guanhang Wu, Trevor Edwards, Filip Pavetic, Bo Majewski, Shreyasee Mukherjee, Stanley Chan, John Morgan, Vivek Rathod, and Jonathan Huang. The auto arborist dataset: A large-scale benchmark for multiview urban forest monitoring under domain shift. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21294-21307, 2022.", + "[3] Markus Braun, Sebastian Krebs, Fabian B. Flohr, and Dariu M. Gavrila. Eurocity persons: A novel benchmark for person detection in traffic scenes. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-1, 2019.", + "[4] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020.", + "[5] Zhaowei Cai and Nuno Vasconcelos. Cascade r-cnn: High quality object detection and instance segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, page 1-1, 2019.", + "[6] Kai Chen, Jiangmiao Pang, Jiaqi Wang, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jianping Shi, Wanli Ouyang, et al. Hybrid task cascade for instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4974-4983, 2019.", + "[7] Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jiarui Xu, et al. Mmdetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155, 2019.", + "[8] Xiaotian Chen, Xuejin Chen, and Zheng-Jun Zha. Structure-aware residual pyramid network for monocular depth estimation. In Proceedings of the 28th International Joint Conference on Artificial Intelligence, pages 694-700, 2019.", + "[9] Gordon Christie, Rodrigo Rene Rai Munoz Abujder, Kevin Foster, Shea Hagstrom, Gregory D Hager, and Myron Z Brown. Learning geocentric object pose in oblique monocular images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14512-14520, 2020." + ], + "bbox": [ + 86, + 118, + 472, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[10] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3213-3223, 2016.", + "[11] Ilke Demir, Krzysztof Koperski, David Lindenbaum, Guan Pang, Jing Huang, Saikat Basu, Forest Hughes, Devis Tuia, and Ramesh Raskar. Deep globe 2018: A challenge to parse the earth through satellite images. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2018.", + "[12] Tian Feng, Quang-Trung Truong, Duc Thanh Nguyen, Jing Yu Koh, Lap-Fai Yu, Alexander Binder, and Sai-Kit Yeung. Urban zoning using higher-order markov random fields on multi-view imagery data. In Proceedings of the European Conference on Computer Vision (ECCV), pages 614–630, 2018.", + "[13] Huan Fu, Mingming Gong, Chaohui Wang, Kayhan Bat-manghelich, and Dacheng Tao. Deep ordinal regression network for monocular depth estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2002-2011, 2018.", + "[14] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE conference on computer vision and pattern recognition, pages 3354-3361. IEEE, 2012.", + "[15] Kaiming He, Georgia Gkioxari, Piotr Dollár, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision (CVPR), pages 2961-2969, 2017.", + "[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016.", + "[17] Sixing Hu, Mengdan Feng, Rang MH Nguyen, and Gim Hee Lee. Cvm-net: Cross-view matching network for image-based ground-to-aerial geo-localization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7258-7267, 2018.", + "[18] Xinyu Huang, Peng Wang, Xinjing Cheng, Dingfu Zhou, Qichuan Geng, and Ruigang Yang. The apolloscape open dataset for autonomous driving and its application. IEEE transactions on pattern analysis and machine intelligence, 42(10):2702-2719, 2019.", + "[19] Zhaojin Huang, Lichao Huang, Yongchao Gong, Chang Huang, and Xinggang Wang. Mask scoring r-cnn. In Pro-" + ], + "bbox": [ + 501, + 92, + 890, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "17405", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6409-6418, 2019.", + "[20] Weijia Li, Conghui He, Jiarui Fang, Juepeng Zheng, Hao-huan Fu, and Le Yu. Semantic segmentation-based building footprint extraction using very high-resolution satellite images and multi-source gis data. Remote Sensing, 11(4):403, 2019.", + "[21] Weijia Li, Lingxuan Meng, Jinwang Wang, Conghui He, Gui-Song Xia, and Dahua Lin. 3d building reconstruction from monocular remote sensing images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12548-12557, 2021.", + "[22] Weijia Li, Wenqian Zhao, Huaping Zhong, Conghui He, and Dahua Lin. Joint semantic-geometric learning for polygonal building segmentation. In AAAI, 2021.", + "[23] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition (CVPR), pages 2117-2125, 2017.", + "[24] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014.", + "[25] Xiaohu Lu, Zuoyue Li, Zhaopeng Cui, Martin R Oswald, Marc Pollefeys, and Rongjun Qin. Geometry-aware satellite-to-ground image synthesis for urban areas. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 859-867, 2020.", + "[26] Gerhard Neuhold, Tobias Ollmann, Samuel Rota Bulo, and Peter Kontschieder. The mapillary vistas dataset for semantic understanding of street scenes. In Proceedings of the IEEE international conference on computer vision, pages 4990-4999, 2017.", + "[27] Krishna Regmi and Ali Borji. Cross-view image synthesis using conditional gans. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3501-3510, 2018.", + "[28] Krishna Regmi and Mubarak Shah. Bridging the domain gap for ground-to-aerial image matching. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 470-479, 2019.", + "[29] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision (IJCV), 115(3):211-252, 2015." + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] Yujiao Shi, Dylan John Campbell, Xin Yu, and Hongdong Li. Geometry-guided street-view panorama synthesis from satellite imagery. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022.", + "[31] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2446-2454, 2020.", + "[32] Zheng Tang, Milind Naphade, Ming-Yu Liu, Xiaodong Yang, Stan Birchfield, Shuo Wang, Ratnesh Kumar, David Anastasiu, and Jenq-Neng Hwang. Cityflow: A city-scale benchmark for multi-target multi-camera vehicle tracking and re-identification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8797–8806, 2019.", + "[33] Yicong Tian, Chen Chen, and Mubarak Shah. Cross-view image matching for geo-localization in urban environments. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3608-3616, 2017.", + "[34] Adam Van Etten, Daniel Hogan, Jesus Martinez Manso, Jacob Shermeyer, Nicholas Weir, and Ryan Lewis. The multitemporal urban development spacenet dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6398-6407, 2021.", + "[35] Jiaqi Wang, Kai Chen, Rui Xu, Ziwei Liu, Chen Change Loy, and Dahua Lin. Carafe: Content-aware reassembly of features. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3007-3016, 2019.", + "[36] Shenlong Wang, Min Bai, Gellert Mattyus, Hang Chu, Wenjie Luo, Bin Yang, Justin Liang, Joel Cheverie, Sanja Fidler, and Raquel Urtasun. Toronto: Seeing the world with a million eyes. In 2017 IEEE International Conference on Computer Vision (ICCV), pages 3028-3036. IEEE, 2017.", + "[37] Jan D Wegner, Steven Branson, David Hall, Konrad Schindler, and Pietro Perona. Cataloging public objects using aerial and street-level images-urban trees. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6014–6023, 2016.", + "[38] Nicholas Weir, David Lindenbaum, Alexei Bastidas, Adam Van Etten, Sean McPherson, Jacob Shermeyer, Varun Kumar, and Hanlin Tang. Spacenet mvoi: A multi-view overhead imagery dataset. In Proceedings of the IEEE/cvf international conference on computer vision, pages 992-1001, 2019.", + "[39] Zbigniew Wojna, Krzysztof Maziarz, Łukasz Jocz, Robert Pałuba, Robert Kozikowski, and Iason Kokkinos. Holis-" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "17406", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tic multi-view building analysis in the wild with projection pooling. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 2870-2878, 2021.", + "[40] Scott Workman, M Usman Rafique, Hunter Blanton, and Nathan Jacobs. Revisiting near/remote sensing with geospatial attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1778-1787, 2022.", + "[41] Kailun Yang, Xinxin Hu, Luis M Bergasa, Eduardo Romera, and Kaiwei Wang. Pass: Panoramic annular semantic segmentation. IEEE Transactions on Intelligent Transportation Systems, 21(10):4171-4185, 2019.", + "[42] Kailun Yang, Xinxin Hu, and Rainer Stiefelhagen. Is context-aware cnn ready for the surroundings? panoramic semantic segmentation in the wild. IEEE Transactions on Image Processing, 30:1866-1881, 2021.", + "[43] Xiaoqing Ye, Mao Shu, Hanyu Li, Yifeng Shi, Yingying Li, Guangjie Wang, Xiao Tan, and Errui Ding. Rope3d: The roadside perception dataset for autonomous driving and monocular 3d object detection task. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21341-21350, 2022.", + "[44] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darryll. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2636-2645, 2020.", + "[45] Menghua Zhai, Zachary Bessinger, Scott Workman, and Nathan Jacobs. Predicting ground-level scene layout from aerial imagery. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 867-875, 2017.", + "[46] Chiqun Zhang, Dragomir Yankov, Chun-Ting Wu, Simon Shapiro, Jason Hong, and Wei Wu. What is that building? an end-to-end system for building recognition from streetside images. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pages 2425-2433, 2020.", + "[47] Yichao Zhou, Jingwei Huang, Xili Dai, Shichen Liu, Linjie Luo, Zhili Chen, and Yi Ma. Holicity: A city-scale data platform for learning holistic 3d structures. arXiv preprint arXiv:2008.03286, 2020.", + "[48] Sijie Zhu, Taojiannan Yang, and Chen Chen. Vigor: Cross-view image geo-localization beyond one-to-one retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3640-3649, 2021." + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 10 + }, + { + "type": "ref_text", + "text": "[49] Yi Zhu, Xueqing Deng, and Shawn Newsam. Fine-grained land use classification at the city scale using ground-level images. IEEE Transactions on Multimedia, 21(7):1825-1838, 2019.", + "bbox": [ + 501, + 90, + 890, + 154 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "17407", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/OmniCity_ Omnipotent City Understanding With Multi-Level and Multi-View Images/1ecc869d-a020-42e0-bc1d-485949718b11_model.json b/2023/OmniCity_ Omnipotent City Understanding With Multi-Level and Multi-View Images/1ecc869d-a020-42e0-bc1d-485949718b11_model.json new file mode 100644 index 0000000000000000000000000000000000000000..cf156039cb4d02507dabad931f5ff5646d5acb36 --- /dev/null +++ b/2023/OmniCity_ Omnipotent City Understanding With Multi-Level and Multi-View Images/1ecc869d-a020-42e0-bc1d-485949718b11_model.json @@ -0,0 +1,2422 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.263, + 0.131, + 0.71, + 0.176 + ], + "angle": 0, + "content": "OmniCity: Omnipotent City Understanding with Multi-level and Multi-view Images" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.203, + 0.753, + 0.221 + ], + "angle": 0, + "content": "Weijia \\(\\mathsf{Li}^1\\), Yawen \\(\\mathsf{Lai}^2\\), Linning \\(\\mathsf{Xu}^3\\), Yuanbo \\(\\mathsf{Xiangli}^3\\), Jinhua \\(\\mathsf{Yu}^1\\)," + }, + { + "type": "text", + "bbox": [ + 0.307, + 0.221, + 0.663, + 0.239 + ], + "angle": 0, + "content": "Conghui He\\(^{2,4*}\\), Gui-Song Xia\\(^{5*}\\), Dahua Lin\\(^{3,4}\\)" + }, + { + "type": "text", + "bbox": [ + 0.137, + 0.239, + 0.835, + 0.257 + ], + "angle": 0, + "content": "1Sun Yat-Sen University, 2SenseTime Research, 3The Chinese University of Hong Kong," + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.257, + 0.739, + 0.275 + ], + "angle": 0, + "content": "\\(^{4}\\)Shanghai Artificial Intelligence Laboratory, \\(^{5}\\)Wuhan University" + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.278, + 0.852, + 0.292 + ], + "angle": 0, + "content": "liweij29@mail.sysu.edu.cn, alanlyawen@gmail.com, {x1020,xy019,dhlin}@ie.cuhk.edu.hk," + }, + { + "type": "text", + "bbox": [ + 0.158, + 0.296, + 0.804, + 0.31 + ], + "angle": 0, + "content": "yujh56@mail2.sysu.edu.cn, heconghui@sensetime.com, guisong.xia@whu.edu.cn" + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.325, + 0.268, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.107, + 0.461, + 0.25, + 0.47 + ], + "angle": 0, + "content": "Selected regions for image collection" + }, + { + "type": "image", + "bbox": [ + 0.09, + 0.471, + 0.268, + 0.607 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.11, + 0.608, + 0.248, + 0.616 + ], + "angle": 0, + "content": "Existing label maps and street-view" + }, + { + "type": "image_caption", + "bbox": [ + 0.111, + 0.616, + 0.247, + 0.625 + ], + "angle": 0, + "content": "image viewpoints of a zoomed area" + }, + { + "type": "image", + "bbox": [ + 0.286, + 0.326, + 0.428, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.299, + 0.415, + 0.415, + 0.422 + ], + "angle": 0, + "content": "Different satellite-level views" + }, + { + "type": "image", + "bbox": [ + 0.288, + 0.426, + 0.427, + 0.512 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.308, + 0.514, + 0.404, + 0.521 + ], + "angle": 0, + "content": "Transformation relations" + }, + { + "type": "image", + "bbox": [ + 0.287, + 0.526, + 0.426, + 0.613 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.304, + 0.614, + 0.41, + 0.623 + ], + "angle": 0, + "content": "Different street-level views" + }, + { + "type": "image", + "bbox": [ + 0.445, + 0.326, + 0.884, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.556, + 0.413, + 0.771, + 0.422 + ], + "angle": 0, + "content": "Satellite-level images and annotations of multiple views" + }, + { + "type": "image", + "bbox": [ + 0.444, + 0.425, + 0.884, + 0.511 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.557, + 0.513, + 0.771, + 0.522 + ], + "angle": 0, + "content": "Street-level images and annotations of a panorama view" + }, + { + "type": "image", + "bbox": [ + 0.444, + 0.526, + 0.884, + 0.613 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.549, + 0.614, + 0.779, + 0.623 + ], + "angle": 0, + "content": "Street-level images and annotations of multiple mono-views" + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.641, + 0.894, + 0.698 + ], + "angle": 0, + "content": "Figure 1. An overview of our proposed OmniCity dataset. The satellite and street-level images of our dataset are collected in the six selected regions. The black dots along the streets denote the viewpoint locations and the land-lot colors indicate different fine-grained categories provided by the existing label maps. Note that all the images in the right part correspond to the same geo-location, and the intrinsic transformation relation between the satellite and street-level panorama view is demonstrated in the middle part." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.729, + 0.314, + 0.745 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.763, + 0.471, + 0.868 + ], + "angle": 0, + "content": "This paper presents OmniCity, a new dataset for omnipotent city understanding from multi-level and multi-view images. More precisely, OmniCity contains multi-view satellite images as well as street-level panorama and monoview images, constituting over 100K pixel-wise annotated images that are well-aligned and collected from 25K geolocations in New York City. To alleviate the substantial" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.731, + 0.894, + 0.896 + ], + "angle": 0, + "content": "pixel-wise annotation efforts, we propose an efficient street-view image annotation pipeline that leverages the existing label maps of satellite view and the transformation relations between different views (satellite, panorama, and mono-view). With the new OmniCity dataset, we provide benchmarks for a variety of tasks including building footprint extraction, height estimation, and building plane/instance/fine-grained segmentation. Compared with existing multi-level and multi-view benchmarks, OmniCity contains a larger number of images with richer annotation types and more views, provides more benchmark results" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.888, + 0.228, + 0.9 + ], + "angle": 0, + "content": "*Corresponding authors." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17397" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.075, + 0.092, + 0.472, + 0.228 + ], + "angle": 0, + "content": "of state-of-the-art models, and introduces a new task for fine-grained building instance segmentation on street-level panorama images. Moreover, OmniCity provides new problem settings for existing tasks, such as cross-view image matching, synthesis, segmentation, detection, etc., and facilitates the developing of new methods for large-scale city understanding, reconstruction, and simulation. The OmniCity dataset as well as the benchmarks will be released at https://city-super.github.io/omnicity/." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.257, + 0.21, + 0.273 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.282, + 0.47, + 0.478 + ], + "angle": 0, + "content": "Owning over a half global population and contributing the most economic growth, the city areas have been recorded and characterized by various data sources including satellite and aerial imagery, street-level imagery, LiDAR data, public maps, crowd-sourced data, etc. A great number of benchmarks have been proposed towards facilitating different vision tasks in city scene, of which the street-level imagery has been broadly used in multiple driving-related benchmarks [4, 10, 14, 18, 26, 31, 43]. The rich visual information of street-level imagery also enables complicated visual recognition tasks on specific categories, such as person detection [3], vehicle tracking and re-identification [32], and fine-grained land use classification [49]." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.479, + 0.47, + 0.705 + ], + "angle": 0, + "content": "Nevertheless, constructing pixel-wise annotations for street-level imagery requires substantial human efforts, resulting in the small image quantity and limited view types of existing datasets, especially for the street-view panorama datasets with only hundreds of annotated images [36, 41, 42]. Regarding the annotation categories and levels, existing datasets mostly provide instance-level annotations for dynamic object categories in driving scenes. As a vital component for city understanding, the static objects such as buildings and roads take up a larger proportion of cities and remain a high consistency across the satellite and ground-level images. However, existing street-level datasets either provide pixel-wise building annotations without fine-grained semantic labels [4, 10, 14, 47] or provide fine-grained annotations at only bbox or image level [46, 49]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.706, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Compared with street-level images, remote sensing images usually contain less visual information for conducting complicated tasks such as fine-grained land use segmentation and building function recognition. On the other hand, unlike the sparsely-distributed street-level images, remote sensing images have a dense spatial distribution and a worldwide coverage, which are well aligned with the open maps and government datasets at pixel level [22]. These existing maps and datasets contain a variety of satellite-level annotations for buildings (such as the footprint, land use, height, year built), roads (category and line coordinates), and other geographical objects, providing new perspectives for promoting novel city understanding datasets and tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.424 + ], + "angle": 0, + "content": "In this work, as illustrated in Figure 1, we construct an omnipotent city dataset unifying data sources from both satellite and street views, linked by geo-locations and urban planning data. Unlike existing city datasets that only support a limited number of tasks, OmniCity dataset incorporates rich geometric annotations and semantic meta data for each image, where multiple tasks can be conducted on. To leverage the existing map labels and the rich visual context from the street-level imagery, we propose an efficient pipeline for producing diverse street-level annotations. Based on this annotation pipeline, we built OmniCity, a dataset that contains over 100K annotated images collected from 25K geo-locations in New York City. We provide benchmark results on OmniCity for a variety of tasks, including building footprint extraction and height estimation on satellite images, as well as fine-grained/instance/plane segmentation of buildings on street-level panorama and mono-view images. To the best of our knowledge, this is the first work that involves fine-grained building instance segmentation on street-level panorama images. We also analyze the potential of OmniCity for promoting new tasks and methods with multi-level imagery." + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.424, + 0.858, + 0.438 + ], + "angle": 0, + "content": "Our main contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.451, + 0.892, + 0.496 + ], + "angle": 0, + "content": "- We propose a novel pipeline for efficiently producing diverse pixel-wise annotations on street-level panorama and mono-view images." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.506, + 0.892, + 0.567 + ], + "angle": 0, + "content": "- We build the OmniCity dataset, which contains well-aligned satellite and street-level images with a larger quantity, richer annotations and more views compared with existing datasets." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.577, + 0.892, + 0.622 + ], + "angle": 0, + "content": "- We provide a series of benchmark experimental results for multiple tasks and data sources, and analyze the limitations of the current benchmarks on OmniCity." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.631, + 0.892, + 0.676 + ], + "angle": 0, + "content": "- We discuss the potential of OmniCity for facilitating new methods and tasks for large-scale city understanding, reconstruction, and simulation." + }, + { + "type": "list", + "bbox": [ + 0.52, + 0.451, + 0.892, + 0.676 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.703, + 0.637, + 0.717 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.727, + 0.862, + 0.741 + ], + "angle": 0, + "content": "2.1. Datasets and methods for street-level tasks" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "As shown in Table 1, many street-level datasets have been proposed over the past few years. A large proportion of these datasets are designed for visual tasks in driving scene [4, 10, 14, 18, 26, 31, 44], such as 2D/3D object detection, semantic segmentation, object tracking, etc. Several street-level datasets are proposed for a specific object category, such as the EuroCity Persons dataset [3], the CityFlow dataset [32] for vehicle tracking and re-identification, etc. In addition to the above datasets containing only monoview images, some studies propose new datasets or methods" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17398" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.472, + 0.188 + ], + "angle": 0, + "content": "Table 1. A comparison of OmniCity with existing city-related datasets of street-level, satellite-level and cross-level. The Street and Satellite (Sate.) columns show the available image view types. The last three columns indicate which level of tasks the dataset is designed for (semantic/instance/plane segmentation, object detection (bbox), and image classification), and whether the dataset contains fine-grained building attribute (Attri.) or height labels." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.195, + 0.462, + 0.364 + ], + "angle": 0, + "content": "
Dataset#ImagesStreetSate.Anno.Attri.Height
KITTI [14]15,000mono-semantic××
Cityscapes [10]25,000mono-semantic××
EuroCity [3]47,300mono-bbox××
WildPASS [42]500multi.-semantic××
PASS [41]400multi.-semantic××
HoliCity [47]6,300multi.-inst./plane××
SkyScapes [1]8,820-singlesemantic××
SpaceNet [38]60,000-multi.instance××
Christie et al. [9]11,000-singlesemantic×✓
Li et al. [21]3,300-singleinstance×✓
TorontoCity [36]Unknowmulti.multi.instance×✓
Wojna et al. [39]49,426monosingleimage✓×
OmniCity108,600multi.multi.inst./plane✓✓
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.378, + 0.47, + 0.589 + ], + "angle": 0, + "content": "for semantic segmentation from panorama images, such as TorontoCity [36], PASS [41], and WildPASS [42]. However, such datasets require expensive annotation efforts and contains only hundreds of pixel-wise annotated panorama images in total. Besides the driving-related datasets, several recent studies target at the recognition tasks for static object categories. Zhu et al. [49] proposed a framework for fine-grained land use classification task using ground-level images. Zhang et al. [46] proposed a building recognition system for detecting their business entity information. HoliCity [47] proposed holistic 3D structure annotations generated from CAD models for panorama images, but it only conducted experiments on mono-view images and lacked the semantic information of buildings and roads." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.59, + 0.471, + 0.833 + ], + "angle": 0, + "content": "In summary, existing street-level datasets still have the following limitations. Regarding the image quantity and view types, most existing datasets require substantial annotation efforts and contain only a limited number of annotated images of a mono view. The image quantity of panorama datasets is even several orders of magnitude smaller than mono-view datasets. Regarding the annotation categories and levels, existing datasets mainly focus on dynamic or driving-related categories and lack in fine-grained annotations of an object category. Moreover, the datasets for static objects are lacking in pixel-level or fine-grained annotations. By contrast, our OmniCity contains over 100K satellite and street-level images of multiple views as well as the pixel-wise and fine-grained annotations, demonstrating superiority in annotation quantity, annotation type, and view type compared with existing datasets." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.846, + 0.456, + 0.862 + ], + "angle": 0, + "content": "2.2. Datasets and methods for satellite-level tasks" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.471, + 0.903 + ], + "angle": 0, + "content": "As a data source with a long time series and a large coverage, the satellite imagery has been broadly explored" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.091, + 0.892, + 0.363 + ], + "angle": 0, + "content": "for large-scale city understanding. Unlike the street-level datasets requiring manual annotations, the satellite imagery is already well aligned with existing label maps [20]. The OpenStreetMap (OSM) is one of the most broadly-used map data, which contains publicly available annotations of building footprints, heights, roads, land use, etc., of world wide. In addition, many public datasets provide rich information at a local scale. For example, the PLUTO dataset contains the block and lot information of the whole New York City, and each lot is associated with the land use, year built, number of floors, and other useful information. The Microsoft US building footprint dataset contains over a hundred million computer-generated building footprints. Several challenges provide manually labeled building footprints and the corresponding satellite images [11, 34, 38]. Moreover, some datasets provide fine-grained semantic categories of different objects [1] or building height annotations for 3D reconstruction tasks [9, 21]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.364, + 0.892, + 0.501 + ], + "angle": 0, + "content": "In summary, existing maps and datasets have provided substantial semantic and geometric information that is well aligned with the satellite imagery. In this work, we leverage these rich annotations and the transformation relations between different views to produce auxiliary information for street-level image annotation. Compared with existing work, OmniCity significantly reduces the human labeling efforts and provides more annotation types to enable omnipotent city understanding via multiple tasks and views." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.509, + 0.861, + 0.525 + ], + "angle": 0, + "content": "2.3. Datasets and methods for multi-level tasks" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.533, + 0.892, + 0.76 + ], + "angle": 0, + "content": "The ground-level images usually contain rich visual context that is not visible from the satellite or aerial imagery (e.g. building facade, the side of vegetations, etc.), while the spatial distribution is often sparse and unbalanced in different areas. By contrast, the remote sensing images have a much denser spatial distribution at a global scale, but the visual context is too limited to support complicated fine-grained tasks. Considering the above complementary characteristics, many datasets and methods have been proposed for cross-view scenarios. Cross-view visual recognition tasks integrate the two data types as the model input for predicting building functions, building ages, land use, tree species, etc. [2, 12, 37, 40], while cross-view image matching [17, 28, 33, 48] and image synthesis [25, 27, 30, 45] tasks take only one data type as the model input." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.76, + 0.892, + 0.867 + ], + "angle": 0, + "content": "In addition, several datasets are proposed for a variety of satellite and street-level tasks. The TorontoCity [36] contains a wide range of annotations including building height estimation, building instance segmentation, building footprint segmentation, road segmentation, etc., which are conducted on either satellite or street-level images and only adopt FCN and some CNN classification architectures as" + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.875, + 0.892, + 0.903 + ], + "angle": 0, + "content": "1https://data.cityofnewyork.us/City-Government/Primary-Land-Use-Tax-Lot-Output-PLUTO-/64uk-42ks." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "17399" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.09, + 0.894, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.247, + 0.893, + 0.276 + ], + "angle": 0, + "content": "Figure 2. The annotation tool of OmniCity dataset. The left and right images show examples of one road scene with the label map (in the bottom-right window) displayed in land use mode, and crossroads scene with the label map in block-lot mode, respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.291, + 0.473, + 0.429 + ], + "angle": 0, + "content": "the baseline models for all experiments. In a recent study [39], a holistic multi-view building analysis dataset is designed for multiple recognition tasks regarding facade material, number of floors, occupancy type, roof geometry, roof pitch, and construction type. The dataset contains only image-wise annotations for street-level classification tasks. The semantic categories of each building are decided by human annotators, resulting in more challenges and subjectivity to the annotation process compared with our work." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.439, + 0.473, + 0.637 + ], + "angle": 0, + "content": "In summary, for the existing cross-view studies designed for one specific task, the street-level images are not annotated and only supplement extra features to the satellite-level feature maps. By contrast, our OmniCity provides rich annotation types that are not contained in existing datasets, e.g. the pixel-wise building instances and fine-grained categories on street-level images, which can promote new methods to explore and leverage these annotations to improve the performance. In addition, OmniCity provides the benchmark experimental results of each task using more state-of-the-art models. With our efficient annotation pipeline, it also provides additional fine-grained pixel-wise annotations and benchmark results compared with [36] and [39]." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.683, + 0.173, + 0.699 + ], + "angle": 0, + "content": "3. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.473, + 0.901 + ], + "angle": 0, + "content": "In this work, we aim at building a dataset for omnipotent city understanding from multi-level and multi-view images. Our proposed OmniCity dataset contains 108,600 images of multiple views, which are collected from 25K geo-locations in New York City. Compared with existing datasets, OmniCity requires much fewer human efforts for street-level image annotation, contains more diverse annotation types for both 2D and 3D tasks, provides richer building semantics at instance segmentation level, and possesses higher scalability for new annotation type supplement and expansion to other cities. The details of data collection, annotation, and statistics are introduced as follows." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.291, + 0.651, + 0.307 + ], + "angle": 0, + "content": "3.1. Data collection" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.316, + 0.895, + 0.74 + ], + "angle": 0, + "content": "As shown in Figure 1, our OmniCity dataset is collected from six selected regions of New York. We download the panorama images in the six selected regions using google street view download 360, with a step distance of 65 meters. The regions for collecting the training and test samples are denoted by red and blue, respectively. We save the geographic coordinates, collection time, panorama id, north rotation, and zoom level for each panorama image. For each panorama site, we collect its corresponding google earth images of three acquisition dates according to the geographic coordinates, constituting three groups of satellite-level datasets with three cases of off-nadir view angles (small/medium/high, denoted by V1/V2/V3). For the annotation data sources, we collect the meta information from PLUTO and OpenStreetMap (OSM). The New York City is hierarchically formed by blocks, lots, and buildings. Each building can be identified by a specific block-lot id. In PLUTO, each lot (building) is associated with rich information, e.g. land use, year built, number of floors, etc. The OSM data contains footprint and height information while lacking in land use information for most buildings. Considering the characteristics of the two data sources, we align the land use attribute (from PLUTO) with the building footprint and height (from OSM) using the geographical coordinates. Overall, each building is assigned with a block-lot id, a land use category, a height value, and the geographical coordinates of a footprint polygon, which will be used as the reference label maps for panorama image annotation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.755, + 0.661, + 0.77 + ], + "angle": 0, + "content": "3.2. Data annotation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Figure 2 shows the panorama image annotation tool proposed in our study. The street-level panorama and satellite images are naturally well-aligned at image-level between the central coordinates from satellite views and the camera pose of panorama views. The instance-level annotations of different views (e.g., building facade and footprint) are further aligned during the annotation process. According to the geo-transformation relation between different views [25]," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17400" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.198 + ], + "angle": 0, + "content": "we calculate the angle between the north (upward) direction and each building footprint (in the satellite image), and the angle between the north direction (saved in the panorama file) and each building facade in the horizontal axis. The footprint and facade with the same angle value are aligned to the same building, providing the auxiliary information for the manual annotation process." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.199, + 0.47, + 0.365 + ], + "angle": 0, + "content": "The annotation pipeline includes four stages: (1) Image selection, i.e., select the panorama images that are essential to be annotated according to building coverage, occlusion extent, etc; (2) Segmentation annotation, i.e., adjust the floor/top line to fit the bottom/roof of each building, and add the boundary split line considering both auxiliary information and building appearance (e.g. texture discrepancy, doors, etc.); (3) Attribute assignment, i.e., add the attributes (instance ID, block-lot id and land use type) for each building plane; (4) Quality assessment, i.e., check the annotation quality and remove the unqualified images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.366, + 0.471, + 0.593 + ], + "angle": 0, + "content": "In addition, the mono-view images and annotations are automatically generated from those of the panorama images via view transformation. For each panorama image, we select three views using three x-axis angles (-170, 10 and 170, in the range of [-180, 180]) and a fixed y-axis angle of 0. Then we design an image selection rule to filter out the unexpected images of which the buildings are distributed in only one side, building area proportion is smaller than \\(10\\%\\), or the building plane quantity is smaller than 2. The remaining images constitute the final mono-view dataset of our experiments. For satellite-level tasks, the annotations are already well-aligned with the meta information from PLUTO and OSM. Each building footprint on a satellite image is assigned with a block-lot id, a land use category, a height value, and the pixel coordinates of the footprint polygon." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.605, + 0.365, + 0.621 + ], + "angle": 0, + "content": "3.3. Statistics of the proposed dataset" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.471, + 0.902 + ], + "angle": 0, + "content": "The whole dataset contains three sub-datasets (in small, medium and large view angles) for satellite-level tasks and two sub-datasets (in panorama and mono views) for street-level tasks. For satellite-level datasets, the three sub-datasets of images collected from the 25K geo-locations constitute 75K images in total, which are cropped by \\(512 \\times 512\\) pixels according to the coverage area of the corresponding panorama image. For the street-level datasets, the panorama dataset contains 18K images in \\(512 \\times 1024\\) pixels, which are selected from the initial 25K panorama images during the image selection and quality assessment phases. Similarly, the mono-view dataset obtained via view transformation and image selection contains 156,00 monoview images in \\(512 \\times 512\\) pixels. The ratio of train/test splits is set as 4:1 for all five sub-datasets. In OmniCity dataset, the initial land use categories of PLUTO with similar characteristics and low quantities are merged into one category, resulting in 7 land use categories in total. The" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.213 + ], + "angle": 0, + "content": "categories of \\(1/2\\) family building (\\(\\sim 17\\%\\)), walk-up building (\\(\\sim 23\\%\\)) and the mixed residential/commercial building categories (\\(\\sim 29\\%\\)) take up a larger proportion compared with the other four categories, i.e., elevator buildings (\\(\\sim 7\\%\\)), office buildings (\\(\\sim 10\\%\\)), industrial/transportation/utility buildings (\\(\\sim 11\\%\\)) and others (\\(\\sim 1\\%\\)). For the distribution of building height, most are between 1 and 25 meters while a small percentage (\\(\\sim 1\\%\\)) reaches over 50 meters." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.231, + 0.684, + 0.246 + ], + "angle": 0, + "content": "4. Benchmark results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.258, + 0.892, + 0.456 + ], + "angle": 0, + "content": "In this section, we provide a variety of benchmarks for multiple satellite and street-level tasks. The satellite-level tasks in our experiments include building footprint segmentation and height estimation. For both tasks, we conduct experiments on the satellite images with three view angles. For the street-level tasks, we conduct two instance segmentation tasks (i.e., land use and building instance segmentation) on the panorama images, and three instance segmentation tasks (i.e., land use / building instance / plane segmentation) on mono-view images. Please note that these are only preliminary experimental results on OmniCity dataset. More benchmarks of latest models and additional tasks will be continuously updated on OmniCity homepage." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.469, + 0.698, + 0.485 + ], + "angle": 0, + "content": "4.1. Experimental setting" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.494, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The experiments are mainly based on mmdetection [7] with the recommended hyper-parameter settings. We select Mask R-CNN [15] as the baseline method for segmentation tasks, and provide a comparison of different methods in Table 6, including Mask Scoring R-CNN (MS R-CNN) [19], Cascade Mask R-CNN (Cascade) [5], Content-Aware ReAssembly of FEatures (CARAFE) [35], and Hybrid Task Cascade (HTC) [6]. Specifically, we use ResNet-50 [16] with FPN [23] pre-trained on the ImageNet [29] as the backbone for all instance segmentation models. All models are trained on 8 NVIDIA Tesla V100 GPUs for 12 epochs, with a batch size of 16, a learning rate starting from 0.02 and decreasing by a factor of 0.1 from the \\(8^{th}\\) to \\(11^{th}\\) epoch, and the stochastic gradient descent (SGD) optimizer with a weight decay of \\(10^{-4}\\) and a momentum of 0.9. For the height estimation task, we evaluate the performance of two widely-used monocular depth estimation methods on the satellite images of three view angles, i.e. Structure-Aware Residual Pyramid Network (SARPN) [8] and Deep Ordinal Regression Network (DORN) [13], which are trained on 4 NVIDIA Tesla V100 GPUs for 20 epochs. SARPN is trained with a batch size of 8, a learning rate starting from \\(10^{-4}\\) and reduced by \\(10\\%\\) every 5 epochs, and the Adam optimizer with \\(\\beta_{1} = 0.9\\), \\(\\beta_{2} = 0.999\\), and a weight decay of \\(10^{-4}\\). DORN is trained with a batch size of 4, a base learning rate of \\(10^{-4}\\) and the power of 0.9, using SGD optimizer with a weight decay of 0.0005 and a momentum of 0.9." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "17401" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.089, + 0.191, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.103, + 0.174, + 0.168, + 0.185 + ], + "angle": 0, + "content": "Small view" + }, + { + "type": "image", + "bbox": [ + 0.197, + 0.089, + 0.306, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.212, + 0.174, + 0.292, + 0.185 + ], + "angle": 0, + "content": "Medium view" + }, + { + "type": "image", + "bbox": [ + 0.312, + 0.089, + 0.422, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.333, + 0.174, + 0.398, + 0.186 + ], + "angle": 0, + "content": "Large view" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.187, + 0.191, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.094, + 0.274, + 0.18, + 0.287 + ], + "angle": 0, + "content": "Satellite image" + }, + { + "type": "image", + "bbox": [ + 0.197, + 0.187, + 0.307, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.211, + 0.275, + 0.292, + 0.286 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image", + "bbox": [ + 0.312, + 0.187, + 0.422, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.327, + 0.275, + 0.41, + 0.286 + ], + "angle": 0, + "content": "DORN - small" + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.289, + 0.191, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.092, + 0.376, + 0.18, + 0.388 + ], + "angle": 0, + "content": "SARPN - small" + }, + { + "type": "image", + "bbox": [ + 0.196, + 0.289, + 0.307, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.2, + 0.376, + 0.304, + 0.388 + ], + "angle": 0, + "content": "SARPN - medium" + }, + { + "type": "image", + "bbox": [ + 0.312, + 0.289, + 0.422, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.321, + 0.376, + 0.411, + 0.389 + ], + "angle": 0, + "content": "SARPN - large" + }, + { + "type": "image_caption", + "bbox": [ + 0.429, + 0.277, + 0.458, + 0.288 + ], + "angle": 0, + "content": "High" + }, + { + "type": "image", + "bbox": [ + 0.428, + 0.291, + 0.456, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.43, + 0.377, + 0.457, + 0.388 + ], + "angle": 0, + "content": "Low" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.407, + 0.47, + 0.436 + ], + "angle": 0, + "content": "Figure 3. Qualitative results of satellite-level tasks. The yellow/cyan/red polygons denote TP/FP/FN buildings." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.451, + 0.44, + 0.467 + ], + "angle": 0, + "content": "4.2. Experimental results of satellite-level tasks" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.478, + 0.473, + 0.901 + ], + "angle": 0, + "content": "Table 2 shows the baseline results of the satellite-level instance segmentation task, which are evaluated using both COCO [24] and SpaceNet [11] evaluation metrics. The footprint segmentation performance is the best for satellite images with a small view angle, and deteriorates seriously when the view angle gets larger. For all three cases, the prediction score is the highest for buildings with large areas and the lowest for small buildings. Figure 3 provides a qualitative comparison of the footprint segmentation results on satellite images of three types of view angle. The large view angle results in great difficulties for extracting accurate footprint boundaries, due to the partial invisibility of building footprint, the serious shadow effects, etc. The height estimation performance is evaluated in terms of the mean absolute error (denoted by MAE), mean square error (denoted by MSE), and root mean square error (denoted by RMSE), which are commonly used metrics for depth estimation. All metrics are measured in meters at pixel level. Table 3 and Figure 3 show the height estimation results obtained from DORN [13] and SARPN [8] for satellite images with different view angles. Results demonstrate that DORN obtains better results compared with SARPN for all three cases. Both methods achieve the best performance for satellite images with a medium view angle (V2) compared with the other two cases. The footprint and roof have more overlaps and provide less building structure information for satellite images with a small view angle, while the shadow and parallax effect become more serious with" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.09, + 0.892, + 0.118 + ], + "angle": 0, + "content": "Table 2. Quantitative results of instance segmentation for satellite images with different view angles." + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.129, + 0.892, + 0.188 + ], + "angle": 0, + "content": "
ViewMetrics of various thresholdsthreshold = 0.5
APAP50AP75APSAPMAPLPRF1
V129.766.023.515.933.936.776.966.371.2
V223.756.616.111.527.230.373.955.063.1
V318.951.49.69.121.525.370.751.759.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.196, + 0.891, + 0.224 + ], + "angle": 0, + "content": "Table 3. Quantitative results of height estimation for satellite images with different view angles." + }, + { + "type": "table", + "bbox": [ + 0.52, + 0.235, + 0.871, + 0.301 + ], + "angle": 0, + "content": "
ViewSARPN [8]DORN [13]
MAEMSERMSEMAEMSERMSE
V116.18870.3429.5012.71670.5225.89
V213.75694.1726.3512.24628.0625.06
V315.32823.0128.6913.40730.6727.03
" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.318, + 0.892, + 0.365 + ], + "angle": 0, + "content": "the increase of off-nadir view angle. The above aspects result in challenges for the accurate estimation of building height for satellite images with small and large view angles." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.377, + 0.847, + 0.393 + ], + "angle": 0, + "content": "4.3. Experimental results of street-level tasks" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.401, + 0.892, + 0.658 + ], + "angle": 0, + "content": "We analyze the performance of multiple segmentation tasks on street-level panorama and mono-view images. Table 4 and Figure 4 show the experimental results of panorama-view images on two segmentation tasks, of which the performance on building instance segmentation task (denoted by Instance Seg.) is significantly superior to the fine-grained land use segmentation task (denoted by Landuse Seg.). Table 5 and Figure 5 show the experimental results of mono-view images on three different tasks, i.e., landuse segmentation, instance segmentation, and plane segmentation. Similar to the results of panorama-view images, the baseline method achieves much higher scores for the two binary segmentation tasks (plane and instance segmentation) compared with the fine-grained land use segmentation task. The qualitative results also demonstrate that the baseline method has difficulties in identifying the accurate land use type of some building instances." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Table 6 shows the land use segmentation results obtained from the five methods on street-level panorama images. HTC achieves the best performance for both overall and category metrics followed by MS R-CNN, which indicates that the cascade structure and mask scoring strategy can effectively improves the fine-grained segmentation performance of building instances. From the metrics of each category, we can find that C2, C4 and C5 (i.e., Walkup Buildings, Mixed Residential/Commercial, and Office Buildings) have better performance compared with C1, C3 and C6 (i.e., 1/2 Family Buildings, Elevator Buildings and Industrial/Transportation/Utility) for all methods, with the AP ranging from \\(20\\%\\) to \\(40\\%\\). The category of Others has an extremely low AP score due to the small sample quantity. Compared with the residential buildings, the mixed Residential/Commercial and Office buildings have more special" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17402" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.092, + 0.468, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.289, + 0.47, + 0.318 + ], + "angle": 0, + "content": "Figure 4. Qualitative results of street-level tasks (i.e., instance segmentation and land use segmentation) on panorama images." + }, + { + "type": "table_caption", + "bbox": [ + 0.089, + 0.33, + 0.456, + 0.345 + ], + "angle": 0, + "content": "Table 4. Quantitative results on street-level panorama images." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.357, + 0.469, + 0.409 + ], + "angle": 0, + "content": "
TaskAPAP50AP75APSAPMAPL
Landuse Seg.26.034.728.50.312.030.4
Instance Seg.66.786.572.51.740.274.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.085, + 0.42, + 0.461, + 0.434 + ], + "angle": 0, + "content": "Table 5. Quantitative results on street-level mono-view images." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.446, + 0.469, + 0.515 + ], + "angle": 0, + "content": "
TaskAPAP50AP75APSAPMAPL
Landuse Seg.23.932.126.70.310.627.5
Instance Seg.68.388.873.83.233.376.1
Plane Seg.65.187.471.05.040.773.8
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.536, + 0.47, + 0.611 + ], + "angle": 0, + "content": "characteristics (i.e. the boundary between first and above floors, facade design, and building structure), contributing to the superior AP scores of these two categories. The performance of the above six categories is also concordant with the ratio of the sample quantity." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.621, + 0.358, + 0.637 + ], + "angle": 0, + "content": "4.4. Results analysis and discussions" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.471, + 0.903 + ], + "angle": 0, + "content": "In this section we summarize the limitations of existing methods for satellite and street-level tasks on our OmniCity dataset. For building footprint segmentation, the performance of existing methods get worse with the increasing of off-nadir view angle, which might due to the serious parallax and shadow effect on the satellite images with a large off-nadir view angle. For the height estimation task, most existing methods directly apply the monocular depth estimation methods to remote sensing scene. These methods usually produce poor results on the invisible side of the footprint boundary especially for high-rise buildings on very off-nadir images. In addition, most existing methods use deep neural networks to regress continuous values but the actual height values are discrete for building and non-building areas, resulting in difficulties for network training and extra efforts for converting the continuous prediction values into discrete height values via post-processing." + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.092, + 0.892, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.351, + 0.892, + 0.382 + ], + "angle": 0, + "content": "Figure 5. Qualitative results of street-level tasks (i.e., land use /instance/plane segmentation) on mono-view images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.403, + 0.893, + 0.706 + ], + "angle": 0, + "content": "For the street-level tasks based on panorama images, existing methods target at general instance segmentation tasks for commonly-used datasets, e.g. COCO [24], CityScapes [10], BDD100K [44], etc. These datasets often have a single view and a narrow Field of View (FoV). However, for panorama images, the special properties such as the wide FoV covering full 360-degree in the horizontal direction, are not taken into consideration in the design of existing methods. For both mono-view and panorama images, existing methods have difficulties in accurately recognizing the building instance with a small area (e.g. buildings located in the side of the main parts), the land use categories with a small number of building instances, and the categories that are easily confused (e.g. 1/2 Family Buildings and Walk-Up Buildings, Mixed Residential/Commercial and Office Buildings). Figure 6 shows some typical failure cases of the current benchmark methods. New instance segmentation methods should be designed for solving the above limitations considering the characteristics of panorama images, building instances, fine-grained categories, etc." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.724, + 0.806, + 0.74 + ], + "angle": 0, + "content": "5. Potential of the OmniCity dataset" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.795 + ], + "angle": 0, + "content": "Our proposed OmniCity dataset demonstrates great potential for facilitating city understanding, machine perception, and generative modeling researches in many aspects." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "First, it can serve as a new dataset for the existing tasks such as image geo-localization/synthesis and segmentation/detection of buildings/trees/land use from cross-view images. OmniCity provides additional annotations that are not contained in existing datasets, e.g. the building instances and fine-grained categories on street-level images, which can promote new methods to explore and leverage" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17403" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.125, + 0.092, + 0.844, + 0.107 + ], + "angle": 0, + "content": "Table 6. Quantitative results of different methods for fine-grained land use segmentation on street-level panorama images." + }, + { + "type": "table", + "bbox": [ + 0.179, + 0.118, + 0.789, + 0.22 + ], + "angle": 0, + "content": "
MethodOverall MetricsMetrics of each category
APAP50AP75APSAPMAPLC1C2C3C4C5C6C7
Mask R-CNN [15]26.034.728.50.312.030.419.637.525.839.236.922.20.8
MS R-CNN [19]27.135.829.80.112.431.522.539.126.240.838.021.71.2
Cascade [5]25.933.828.30.211.430.52038.32538.536.722.10.3
CARAFE [35]25.934.528.50.111.930.219.637.324.939.937.221.50.8
HTC [6]27.235.729.90.312.432.020.838.727.239.938.424.51.2
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.234, + 0.47, + 0.355 + ], + "angle": 0, + "content": "the new annotations to improve the performance of these tasks. The annotations are organized in a unified version, which means multiple tasks can be performed on a single image, and thus can well support the multi-task learning setting. Additionally, since the building instances are directly linked with the urban planning data using block-lot id, it is easy to be enriched with more annotation types from other urban datasets, especially those for social and urban studies." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.356, + 0.47, + 0.506 + ], + "angle": 0, + "content": "Second, OmniCity provides a new application scenario or problem setting to existing tasks. For line segment detection and wireframe parsing tasks, existing datasets contains densely distributed line segment and wireframe labels, while our OmniCity focuses on the main line segments and wireframes on the outlines instead of the inner ones, resulting in a much sparser format. The serious shelters from the trees and vehicles also bring challenges to these tasks. New line segment detection and wireframe parsing methods should be designed for the OmniCity scenario." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.507, + 0.47, + 0.629 + ], + "angle": 0, + "content": "Moreover, OmniCity facilitates new tasks for city reconstruction and simulation. Treating each panorama as a unique city scene, a complete 3D model representing such a local scene is available, as also shown in the second window panel in our tool (Figure 2). These 3D models are stored in abstract vector formats, with clean and clear vertices indicating the building facades, streets, etc., which are suitable for many shape generation tasks represented with graphs." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.63, + 0.47, + 0.795 + ], + "angle": 0, + "content": "Finally, with the well-aligned satellite and street-level images as well as the various annotation types, novel city reconstruction tasks, e.g., 3D building reconstruction from cross-view images, can be derived for producing holistic 3D buildings with both fine-grained semantic category and precise geometry information (vector 3D model). Unlike the existing studies that only target at 3D reconstruction from monocular or multi-view remote sensing imagery, new methods should be designed to leverage the additional information from street-level images to improve 3D reconstruction and semantic prediction for OmniCity scenario." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.815, + 0.196, + 0.83 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.84, + 0.47, + 0.901 + ], + "angle": 0, + "content": "In this paper, we have proposed OmniCity, a new dataset for omnipotent city understanding from satellite and street-level images of multiple views. The dataset contains over 100K images collected from 25K geo-locations in New" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.23, + 0.619, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.531, + 0.308, + 0.591, + 0.316 + ], + "angle": 0, + "content": "(a)GT:5-Office" + }, + { + "type": "image", + "bbox": [ + 0.622, + 0.231, + 0.756, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.629, + 0.308, + 0.749, + 0.316 + ], + "angle": 0, + "content": "(b) GT: 4-Mixed (three instances)" + }, + { + "type": "image", + "bbox": [ + 0.758, + 0.232, + 0.892, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.791, + 0.308, + 0.859, + 0.316 + ], + "angle": 0, + "content": "(c) GT: 3-Elevator" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.32, + 0.603, + 0.396 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.522, + 0.397, + 0.583, + 0.405 + ], + "angle": 0, + "content": "(d)GT:5-Office" + }, + { + "type": "image", + "bbox": [ + 0.603, + 0.32, + 0.69, + 0.396 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.617, + 0.397, + 0.676, + 0.405 + ], + "angle": 0, + "content": "(e)GT:5-Office" + }, + { + "type": "image", + "bbox": [ + 0.692, + 0.32, + 0.787, + 0.396 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.709, + 0.397, + 0.768, + 0.405 + ], + "angle": 0, + "content": "(f) GT: 7-Others" + }, + { + "type": "image", + "bbox": [ + 0.788, + 0.32, + 0.892, + 0.396 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.807, + 0.397, + 0.871, + 0.406 + ], + "angle": 0, + "content": "(g) GT: 1-Family" + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.424, + 0.89, + 0.438 + ], + "angle": 0, + "content": "Figure 6. Typical failure cases of the current benchmark methods." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.47, + 0.892, + 0.788 + ], + "angle": 0, + "content": "York City, of which the annotations are generated from both existing label maps and our proposed annotation pipeline. We provide benchmark experimental results for multiple tasks and data sources based on state-of-the-art methods and analyze their limitations. We believe that OmniCity will not only promote new algorithms and application scenarios for existing tasks, but facilitate novel tasks for 3D city reconstruction and simulation. In our future work, we will keep updating the OmniCity dataset and the benchmarks in the following aspects. Owing to the proposed annotation pipeline and the unified annotations with a vector format and rich meta information (geo-locations, block-lot id, etc.), OmniCity can be efficiently supplemented with more properties of buildings and other geographical object types (roads, sidewalks, trees, green space, etc.), and extended to other cities of different countries. The benchmark results of more state-of-the-art models and new tasks will be provided accordingly. Based on the rich annotation and view types of OmniCity, we also plan to develop new methods for existing and novel tasks, such as object detection, instance segmentation, and 3D reconstruction from cross-view images." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Acknowledgements. This project was funded in part by National Natural Science Foundation of China (No. 42201358, No. U22B2011 and No. 41820104006), Shanghai AI Laboratory, CUHK Interdisciplinary AI Research Institute, and the Centre for Perceptual and Interactive Intelligence (CPII) Ltd under the Innovation and Technology Commission (ITC)'s InnoHK." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17404" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.119, + 0.473, + 0.2 + ], + "angle": 0, + "content": "[1] Seyed Majid Azimi, Corentin Henry, Lars Sommer, Arne Schumann, and Eleonora Vig. Skyscapes fine-grained semantic understanding of aerial scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7393-7403, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.205, + 0.472, + 0.319 + ], + "angle": 0, + "content": "[2] Sara Beery, Guanhang Wu, Trevor Edwards, Filip Pavetic, Bo Majewski, Shreyasee Mukherjee, Stanley Chan, John Morgan, Vivek Rathod, and Jonathan Huang. The auto arborist dataset: A large-scale benchmark for multiview urban forest monitoring under domain shift. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21294-21307, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.323, + 0.472, + 0.389 + ], + "angle": 0, + "content": "[3] Markus Braun, Sebastian Krebs, Fabian B. Flohr, and Dariu M. Gavrila. Eurocity persons: A novel benchmark for person detection in traffic scenes. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-1, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.392, + 0.472, + 0.49 + ], + "angle": 0, + "content": "[4] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.494, + 0.472, + 0.56 + ], + "angle": 0, + "content": "[5] Zhaowei Cai and Nuno Vasconcelos. Cascade r-cnn: High quality object detection and instance segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, page 1-1, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.564, + 0.472, + 0.661 + ], + "angle": 0, + "content": "[6] Kai Chen, Jiangmiao Pang, Jiaqi Wang, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jianping Shi, Wanli Ouyang, et al. Hybrid task cascade for instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4974-4983, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.665, + 0.472, + 0.731 + ], + "angle": 0, + "content": "[7] Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jiarui Xu, et al. Mmdetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.735, + 0.472, + 0.8 + ], + "angle": 0, + "content": "[8] Xiaotian Chen, Xuejin Chen, and Zheng-Jun Zha. Structure-aware residual pyramid network for monocular depth estimation. In Proceedings of the 28th International Joint Conference on Artificial Intelligence, pages 694-700, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.804, + 0.472, + 0.9 + ], + "angle": 0, + "content": "[9] Gordon Christie, Rodrigo Rene Rai Munoz Abujder, Kevin Foster, Shea Hagstrom, Gregory D Hager, and Myron Z Brown. Learning geocentric object pose in oblique monocular images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14512-14520, 2020." + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.119, + 0.473, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[10] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3213-3223, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.195, + 0.892, + 0.293 + ], + "angle": 0, + "content": "[11] Ilke Demir, Krzysztof Koperski, David Lindenbaum, Guan Pang, Jing Huang, Saikat Basu, Forest Hughes, Devis Tuia, and Ramesh Raskar. Deep globe 2018: A challenge to parse the earth through satellite images. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.298, + 0.892, + 0.396 + ], + "angle": 0, + "content": "[12] Tian Feng, Quang-Trung Truong, Duc Thanh Nguyen, Jing Yu Koh, Lap-Fai Yu, Alexander Binder, and Sai-Kit Yeung. Urban zoning using higher-order markov random fields on multi-view imagery data. In Proceedings of the European Conference on Computer Vision (ECCV), pages 614–630, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.402, + 0.892, + 0.483 + ], + "angle": 0, + "content": "[13] Huan Fu, Mingming Gong, Chaohui Wang, Kayhan Bat-manghelich, and Dacheng Tao. Deep ordinal regression network for monocular depth estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2002-2011, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.488, + 0.892, + 0.552 + ], + "angle": 0, + "content": "[14] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE conference on computer vision and pattern recognition, pages 3354-3361. IEEE, 2012." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.557, + 0.892, + 0.621 + ], + "angle": 0, + "content": "[15] Kaiming He, Georgia Gkioxari, Piotr Dollár, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision (CVPR), pages 2961-2969, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.628, + 0.892, + 0.692 + ], + "angle": 0, + "content": "[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.697, + 0.892, + 0.778 + ], + "angle": 0, + "content": "[17] Sixing Hu, Mengdan Feng, Rang MH Nguyen, and Gim Hee Lee. Cvm-net: Cross-view matching network for image-based ground-to-aerial geo-localization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7258-7267, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.784, + 0.892, + 0.864 + ], + "angle": 0, + "content": "[18] Xinyu Huang, Peng Wang, Xinjing Cheng, Dingfu Zhou, Qichuan Geng, and Ruigang Yang. The apolloscape open dataset for autonomous driving and its application. IEEE transactions on pattern analysis and machine intelligence, 42(10):2702-2719, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.87, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[19] Zhaojin Huang, Lichao Huang, Yongchao Gong, Chang Huang, and Xinggang Wang. Mask scoring r-cnn. In Pro-" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17405" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.47, + 0.123 + ], + "angle": 0, + "content": "ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6409-6418, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.127, + 0.471, + 0.207 + ], + "angle": 0, + "content": "[20] Weijia Li, Conghui He, Jiarui Fang, Juepeng Zheng, Hao-huan Fu, and Le Yu. Semantic segmentation-based building footprint extraction using very high-resolution satellite images and multi-source gis data. Remote Sensing, 11(4):403, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.211, + 0.472, + 0.293 + ], + "angle": 0, + "content": "[21] Weijia Li, Lingxuan Meng, Jinwang Wang, Conghui He, Gui-Song Xia, and Dahua Lin. 3d building reconstruction from monocular remote sensing images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12548-12557, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.296, + 0.471, + 0.343 + ], + "angle": 0, + "content": "[22] Weijia Li, Wenqian Zhao, Huaping Zhong, Conghui He, and Dahua Lin. Joint semantic-geometric learning for polygonal building segmentation. In AAAI, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.347, + 0.471, + 0.428 + ], + "angle": 0, + "content": "[23] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition (CVPR), pages 2117-2125, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.431, + 0.471, + 0.513 + ], + "angle": 0, + "content": "[24] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.516, + 0.472, + 0.597 + ], + "angle": 0, + "content": "[25] Xiaohu Lu, Zuoyue Li, Zhaopeng Cui, Martin R Oswald, Marc Pollefeys, and Rongjun Qin. Geometry-aware satellite-to-ground image synthesis for urban areas. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 859-867, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.601, + 0.471, + 0.68 + ], + "angle": 0, + "content": "[26] Gerhard Neuhold, Tobias Ollmann, Samuel Rota Bulo, and Peter Kontschieder. The mapillary vistas dataset for semantic understanding of street scenes. In Proceedings of the IEEE international conference on computer vision, pages 4990-4999, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.685, + 0.471, + 0.748 + ], + "angle": 0, + "content": "[27] Krishna Regmi and Ali Borji. Cross-view image synthesis using conditional gans. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3501-3510, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.753, + 0.471, + 0.817 + ], + "angle": 0, + "content": "[28] Krishna Regmi and Mubarak Shah. Bridging the domain gap for ground-to-aerial image matching. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 470-479, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.821, + 0.472, + 0.901 + ], + "angle": 0, + "content": "[29] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision (IJCV), 115(3):211-252, 2015." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.157 + ], + "angle": 0, + "content": "[30] Yujiao Shi, Dylan John Campbell, Xin Yu, and Hongdong Li. Geometry-guided street-view panorama synthesis from satellite imagery. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.16, + 0.894, + 0.258 + ], + "angle": 0, + "content": "[31] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2446-2454, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.262, + 0.893, + 0.375 + ], + "angle": 0, + "content": "[32] Zheng Tang, Milind Naphade, Ming-Yu Liu, Xiaodong Yang, Stan Birchfield, Shuo Wang, Ratnesh Kumar, David Anastasiu, and Jenq-Neng Hwang. Cityflow: A city-scale benchmark for multi-target multi-camera vehicle tracking and re-identification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8797–8806, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.379, + 0.893, + 0.444 + ], + "angle": 0, + "content": "[33] Yicong Tian, Chen Chen, and Mubarak Shah. Cross-view image matching for geo-localization in urban environments. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3608-3616, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.447, + 0.894, + 0.528 + ], + "angle": 0, + "content": "[34] Adam Van Etten, Daniel Hogan, Jesus Martinez Manso, Jacob Shermeyer, Nicholas Weir, and Ryan Lewis. The multitemporal urban development spacenet dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6398-6407, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.532, + 0.893, + 0.597 + ], + "angle": 0, + "content": "[35] Jiaqi Wang, Kai Chen, Rui Xu, Ziwei Liu, Chen Change Loy, and Dahua Lin. Carafe: Content-aware reassembly of features. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3007-3016, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.6, + 0.893, + 0.681 + ], + "angle": 0, + "content": "[36] Shenlong Wang, Min Bai, Gellert Mattyus, Hang Chu, Wenjie Luo, Bin Yang, Justin Liang, Joel Cheverie, Sanja Fidler, and Raquel Urtasun. Toronto: Seeing the world with a million eyes. In 2017 IEEE International Conference on Computer Vision (ICCV), pages 3028-3036. IEEE, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.684, + 0.893, + 0.765 + ], + "angle": 0, + "content": "[37] Jan D Wegner, Steven Branson, David Hall, Konrad Schindler, and Pietro Perona. Cataloging public objects using aerial and street-level images-urban trees. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6014–6023, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.769, + 0.893, + 0.866 + ], + "angle": 0, + "content": "[38] Nicholas Weir, David Lindenbaum, Alexei Bastidas, Adam Van Etten, Sean McPherson, Jacob Shermeyer, Varun Kumar, and Hanlin Tang. Spacenet mvoi: A multi-view overhead imagery dataset. In Proceedings of the IEEE/cvf international conference on computer vision, pages 992-1001, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.87, + 0.893, + 0.901 + ], + "angle": 0, + "content": "[39] Zbigniew Wojna, Krzysztof Maziarz, Łukasz Jocz, Robert Pałuba, Robert Kozikowski, and Iason Kokkinos. Holis-" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17406" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.14 + ], + "angle": 0, + "content": "tic multi-view building analysis in the wild with projection pooling. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 2870-2878, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.146, + 0.47, + 0.225 + ], + "angle": 0, + "content": "[40] Scott Workman, M Usman Rafique, Hunter Blanton, and Nathan Jacobs. Revisiting near/remote sensing with geospatial attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1778-1787, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.233, + 0.469, + 0.295 + ], + "angle": 0, + "content": "[41] Kailun Yang, Xinxin Hu, Luis M Bergasa, Eduardo Romera, and Kaiwei Wang. Pass: Panoramic annular semantic segmentation. IEEE Transactions on Intelligent Transportation Systems, 21(10):4171-4185, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.302, + 0.469, + 0.366 + ], + "angle": 0, + "content": "[42] Kailun Yang, Xinxin Hu, and Rainer Stiefelhagen. Is context-aware cnn ready for the surroundings? panoramic semantic segmentation in the wild. IEEE Transactions on Image Processing, 30:1866-1881, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.372, + 0.47, + 0.469 + ], + "angle": 0, + "content": "[43] Xiaoqing Ye, Mao Shu, Hanyu Li, Yifeng Shi, Yingying Li, Guangjie Wang, Xiao Tan, and Errui Ding. Rope3d: The roadside perception dataset for autonomous driving and monocular 3d object detection task. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21341-21350, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.475, + 0.469, + 0.571 + ], + "angle": 0, + "content": "[44] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darryll. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2636-2645, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.578, + 0.469, + 0.657 + ], + "angle": 0, + "content": "[45] Menghua Zhai, Zachary Bessinger, Scott Workman, and Nathan Jacobs. Predicting ground-level scene layout from aerial imagery. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 867-875, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.664, + 0.469, + 0.761 + ], + "angle": 0, + "content": "[46] Chiqun Zhang, Dragomir Yankov, Chun-Ting Wu, Simon Shapiro, Jason Hong, and Wei Wu. What is that building? an end-to-end system for building recognition from streetside images. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pages 2425-2433, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.767, + 0.469, + 0.83 + ], + "angle": 0, + "content": "[47] Yichao Zhou, Jingwei Huang, Xili Dai, Shichen Liu, Linjie Luo, Zhili Chen, and Yi Ma. Holicity: A city-scale data platform for learning holistic 3d structures. arXiv preprint arXiv:2008.03286, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.837, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[48] Sijie Zhu, Taojiannan Yang, and Chen Chen. Vigor: Cross-view image geo-localization beyond one-to-one retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3640-3649, 2021." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.155 + ], + "angle": 0, + "content": "[49] Yi Zhu, Xueqing Deng, and Shawn Newsam. Fine-grained land use classification at the city scale using ground-level images. IEEE Transactions on Multimedia, 21(7):1825-1838, 2019." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17407" + } + ] +] \ No newline at end of file diff --git a/2023/OmniCity_ Omnipotent City Understanding With Multi-Level and Multi-View Images/1ecc869d-a020-42e0-bc1d-485949718b11_origin.pdf b/2023/OmniCity_ Omnipotent City Understanding With Multi-Level and Multi-View Images/1ecc869d-a020-42e0-bc1d-485949718b11_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2d456861bf14bd38543c0026d21835dededc920a --- /dev/null +++ b/2023/OmniCity_ Omnipotent City Understanding With Multi-Level and Multi-View Images/1ecc869d-a020-42e0-bc1d-485949718b11_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3687905828135e3e054d859d4eee140cc2cf6e2c44f2a6cb7f522c8638590c32 +size 8979853 diff --git a/2023/OmniCity_ Omnipotent City Understanding With Multi-Level and Multi-View Images/full.md b/2023/OmniCity_ Omnipotent City Understanding With Multi-Level and Multi-View Images/full.md new file mode 100644 index 0000000000000000000000000000000000000000..e02d74b39e49b24cc2522bccc41a903a24cadb63 --- /dev/null +++ b/2023/OmniCity_ Omnipotent City Understanding With Multi-Level and Multi-View Images/full.md @@ -0,0 +1,312 @@ +# OmniCity: Omnipotent City Understanding with Multi-level and Multi-view Images + +Weijia $\mathsf{Li}^1$ , Yawen $\mathsf{Lai}^2$ , Linning $\mathsf{Xu}^3$ , Yuanbo $\mathsf{Xiangli}^3$ , Jinhua $\mathsf{Yu}^1$ , + +Conghui He $^{2,4*}$ , Gui-Song Xia $^{5*}$ , Dahua Lin $^{3,4}$ + +1Sun Yat-Sen University, 2SenseTime Research, 3The Chinese University of Hong Kong, + +$^{4}$ Shanghai Artificial Intelligence Laboratory, $^{5}$ Wuhan University + +liweij29@mail.sysu.edu.cn, alanlyawen@gmail.com, {x1020,xy019,dhlin}@ie.cuhk.edu.hk, + +yujh56@mail2.sysu.edu.cn, heconghui@sensetime.com, guisong.xia@whu.edu.cn + +![](images/17ab3d5166ad6ac2bd9135a4b1d471e714e762c8ce33ac287a2caadc2464873d.jpg) +Selected regions for image collection + +![](images/8e739941d17d503c131a3a59c44244dacb429df05aefc4369495a32451a2e410.jpg) +Existing label maps and street-view +image viewpoints of a zoomed area + +![](images/df3096eba4b975b6853f5d50bfba1bda95ab2acbf866ed0745c1d15fe2e5ba9e.jpg) +Different satellite-level views + +![](images/9baf49218cb34b86d8791dca9c85811d9039793c833f415a093c12d45dd68be1.jpg) +Transformation relations + +![](images/6745b714f3acb034ea251eed71f2dd5b02da49c9c6df58714db8ba25f53e1b1b.jpg) +Different street-level views +Figure 1. An overview of our proposed OmniCity dataset. The satellite and street-level images of our dataset are collected in the six selected regions. The black dots along the streets denote the viewpoint locations and the land-lot colors indicate different fine-grained categories provided by the existing label maps. Note that all the images in the right part correspond to the same geo-location, and the intrinsic transformation relation between the satellite and street-level panorama view is demonstrated in the middle part. + +![](images/b0ed4e5dc56c5a0b4c188d8b334b72d54421f6bd02944c6245a0d6cf2c89a721.jpg) +Satellite-level images and annotations of multiple views + +![](images/73467ef54011833bcecb43d7d56ad6ca5320de447ed75077efae653eb7ef2902.jpg) +Street-level images and annotations of a panorama view + +![](images/f6643e67e7001e20fcd2b4e68a133cce39629038f536021493c32683be155a65.jpg) +Street-level images and annotations of multiple mono-views + +# Abstract + +This paper presents OmniCity, a new dataset for omnipotent city understanding from multi-level and multi-view images. More precisely, OmniCity contains multi-view satellite images as well as street-level panorama and monoview images, constituting over 100K pixel-wise annotated images that are well-aligned and collected from 25K geolocations in New York City. To alleviate the substantial + +pixel-wise annotation efforts, we propose an efficient street-view image annotation pipeline that leverages the existing label maps of satellite view and the transformation relations between different views (satellite, panorama, and mono-view). With the new OmniCity dataset, we provide benchmarks for a variety of tasks including building footprint extraction, height estimation, and building plane/instance/fine-grained segmentation. Compared with existing multi-level and multi-view benchmarks, OmniCity contains a larger number of images with richer annotation types and more views, provides more benchmark results + +of state-of-the-art models, and introduces a new task for fine-grained building instance segmentation on street-level panorama images. Moreover, OmniCity provides new problem settings for existing tasks, such as cross-view image matching, synthesis, segmentation, detection, etc., and facilitates the developing of new methods for large-scale city understanding, reconstruction, and simulation. The OmniCity dataset as well as the benchmarks will be released at https://city-super.github.io/omnicity/. + +# 1. Introduction + +Owning over a half global population and contributing the most economic growth, the city areas have been recorded and characterized by various data sources including satellite and aerial imagery, street-level imagery, LiDAR data, public maps, crowd-sourced data, etc. A great number of benchmarks have been proposed towards facilitating different vision tasks in city scene, of which the street-level imagery has been broadly used in multiple driving-related benchmarks [4, 10, 14, 18, 26, 31, 43]. The rich visual information of street-level imagery also enables complicated visual recognition tasks on specific categories, such as person detection [3], vehicle tracking and re-identification [32], and fine-grained land use classification [49]. + +Nevertheless, constructing pixel-wise annotations for street-level imagery requires substantial human efforts, resulting in the small image quantity and limited view types of existing datasets, especially for the street-view panorama datasets with only hundreds of annotated images [36, 41, 42]. Regarding the annotation categories and levels, existing datasets mostly provide instance-level annotations for dynamic object categories in driving scenes. As a vital component for city understanding, the static objects such as buildings and roads take up a larger proportion of cities and remain a high consistency across the satellite and ground-level images. However, existing street-level datasets either provide pixel-wise building annotations without fine-grained semantic labels [4, 10, 14, 47] or provide fine-grained annotations at only bbox or image level [46, 49]. + +Compared with street-level images, remote sensing images usually contain less visual information for conducting complicated tasks such as fine-grained land use segmentation and building function recognition. On the other hand, unlike the sparsely-distributed street-level images, remote sensing images have a dense spatial distribution and a worldwide coverage, which are well aligned with the open maps and government datasets at pixel level [22]. These existing maps and datasets contain a variety of satellite-level annotations for buildings (such as the footprint, land use, height, year built), roads (category and line coordinates), and other geographical objects, providing new perspectives for promoting novel city understanding datasets and tasks. + +In this work, as illustrated in Figure 1, we construct an omnipotent city dataset unifying data sources from both satellite and street views, linked by geo-locations and urban planning data. Unlike existing city datasets that only support a limited number of tasks, OmniCity dataset incorporates rich geometric annotations and semantic meta data for each image, where multiple tasks can be conducted on. To leverage the existing map labels and the rich visual context from the street-level imagery, we propose an efficient pipeline for producing diverse street-level annotations. Based on this annotation pipeline, we built OmniCity, a dataset that contains over 100K annotated images collected from 25K geo-locations in New York City. We provide benchmark results on OmniCity for a variety of tasks, including building footprint extraction and height estimation on satellite images, as well as fine-grained/instance/plane segmentation of buildings on street-level panorama and mono-view images. To the best of our knowledge, this is the first work that involves fine-grained building instance segmentation on street-level panorama images. We also analyze the potential of OmniCity for promoting new tasks and methods with multi-level imagery. + +Our main contributions are summarized as follows: + +- We propose a novel pipeline for efficiently producing diverse pixel-wise annotations on street-level panorama and mono-view images. +- We build the OmniCity dataset, which contains well-aligned satellite and street-level images with a larger quantity, richer annotations and more views compared with existing datasets. +- We provide a series of benchmark experimental results for multiple tasks and data sources, and analyze the limitations of the current benchmarks on OmniCity. +- We discuss the potential of OmniCity for facilitating new methods and tasks for large-scale city understanding, reconstruction, and simulation. + +# 2. Related work + +# 2.1. Datasets and methods for street-level tasks + +As shown in Table 1, many street-level datasets have been proposed over the past few years. A large proportion of these datasets are designed for visual tasks in driving scene [4, 10, 14, 18, 26, 31, 44], such as 2D/3D object detection, semantic segmentation, object tracking, etc. Several street-level datasets are proposed for a specific object category, such as the EuroCity Persons dataset [3], the CityFlow dataset [32] for vehicle tracking and re-identification, etc. In addition to the above datasets containing only monoview images, some studies propose new datasets or methods + +Table 1. A comparison of OmniCity with existing city-related datasets of street-level, satellite-level and cross-level. The Street and Satellite (Sate.) columns show the available image view types. The last three columns indicate which level of tasks the dataset is designed for (semantic/instance/plane segmentation, object detection (bbox), and image classification), and whether the dataset contains fine-grained building attribute (Attri.) or height labels. + +
Dataset#ImagesStreetSate.Anno.Attri.Height
KITTI [14]15,000mono-semantic××
Cityscapes [10]25,000mono-semantic××
EuroCity [3]47,300mono-bbox××
WildPASS [42]500multi.-semantic××
PASS [41]400multi.-semantic××
HoliCity [47]6,300multi.-inst./plane××
SkyScapes [1]8,820-singlesemantic××
SpaceNet [38]60,000-multi.instance××
Christie et al. [9]11,000-singlesemantic×✓
Li et al. [21]3,300-singleinstance×✓
TorontoCity [36]Unknowmulti.multi.instance×✓
Wojna et al. [39]49,426monosingleimage✓×
OmniCity108,600multi.multi.inst./plane✓✓
+ +for semantic segmentation from panorama images, such as TorontoCity [36], PASS [41], and WildPASS [42]. However, such datasets require expensive annotation efforts and contains only hundreds of pixel-wise annotated panorama images in total. Besides the driving-related datasets, several recent studies target at the recognition tasks for static object categories. Zhu et al. [49] proposed a framework for fine-grained land use classification task using ground-level images. Zhang et al. [46] proposed a building recognition system for detecting their business entity information. HoliCity [47] proposed holistic 3D structure annotations generated from CAD models for panorama images, but it only conducted experiments on mono-view images and lacked the semantic information of buildings and roads. + +In summary, existing street-level datasets still have the following limitations. Regarding the image quantity and view types, most existing datasets require substantial annotation efforts and contain only a limited number of annotated images of a mono view. The image quantity of panorama datasets is even several orders of magnitude smaller than mono-view datasets. Regarding the annotation categories and levels, existing datasets mainly focus on dynamic or driving-related categories and lack in fine-grained annotations of an object category. Moreover, the datasets for static objects are lacking in pixel-level or fine-grained annotations. By contrast, our OmniCity contains over 100K satellite and street-level images of multiple views as well as the pixel-wise and fine-grained annotations, demonstrating superiority in annotation quantity, annotation type, and view type compared with existing datasets. + +# 2.2. Datasets and methods for satellite-level tasks + +As a data source with a long time series and a large coverage, the satellite imagery has been broadly explored + +for large-scale city understanding. Unlike the street-level datasets requiring manual annotations, the satellite imagery is already well aligned with existing label maps [20]. The OpenStreetMap (OSM) is one of the most broadly-used map data, which contains publicly available annotations of building footprints, heights, roads, land use, etc., of world wide. In addition, many public datasets provide rich information at a local scale. For example, the PLUTO dataset contains the block and lot information of the whole New York City, and each lot is associated with the land use, year built, number of floors, and other useful information. The Microsoft US building footprint dataset contains over a hundred million computer-generated building footprints. Several challenges provide manually labeled building footprints and the corresponding satellite images [11, 34, 38]. Moreover, some datasets provide fine-grained semantic categories of different objects [1] or building height annotations for 3D reconstruction tasks [9, 21]. + +In summary, existing maps and datasets have provided substantial semantic and geometric information that is well aligned with the satellite imagery. In this work, we leverage these rich annotations and the transformation relations between different views to produce auxiliary information for street-level image annotation. Compared with existing work, OmniCity significantly reduces the human labeling efforts and provides more annotation types to enable omnipotent city understanding via multiple tasks and views. + +# 2.3. Datasets and methods for multi-level tasks + +The ground-level images usually contain rich visual context that is not visible from the satellite or aerial imagery (e.g. building facade, the side of vegetations, etc.), while the spatial distribution is often sparse and unbalanced in different areas. By contrast, the remote sensing images have a much denser spatial distribution at a global scale, but the visual context is too limited to support complicated fine-grained tasks. Considering the above complementary characteristics, many datasets and methods have been proposed for cross-view scenarios. Cross-view visual recognition tasks integrate the two data types as the model input for predicting building functions, building ages, land use, tree species, etc. [2, 12, 37, 40], while cross-view image matching [17, 28, 33, 48] and image synthesis [25, 27, 30, 45] tasks take only one data type as the model input. + +In addition, several datasets are proposed for a variety of satellite and street-level tasks. The TorontoCity [36] contains a wide range of annotations including building height estimation, building instance segmentation, building footprint segmentation, road segmentation, etc., which are conducted on either satellite or street-level images and only adopt FCN and some CNN classification architectures as + +![](images/7b23f45f7c52680205d1d6a47ce7f9a8b80cd7930a169cf4ef77f6e75d676507.jpg) +Figure 2. The annotation tool of OmniCity dataset. The left and right images show examples of one road scene with the label map (in the bottom-right window) displayed in land use mode, and crossroads scene with the label map in block-lot mode, respectively. + +the baseline models for all experiments. In a recent study [39], a holistic multi-view building analysis dataset is designed for multiple recognition tasks regarding facade material, number of floors, occupancy type, roof geometry, roof pitch, and construction type. The dataset contains only image-wise annotations for street-level classification tasks. The semantic categories of each building are decided by human annotators, resulting in more challenges and subjectivity to the annotation process compared with our work. + +In summary, for the existing cross-view studies designed for one specific task, the street-level images are not annotated and only supplement extra features to the satellite-level feature maps. By contrast, our OmniCity provides rich annotation types that are not contained in existing datasets, e.g. the pixel-wise building instances and fine-grained categories on street-level images, which can promote new methods to explore and leverage these annotations to improve the performance. In addition, OmniCity provides the benchmark experimental results of each task using more state-of-the-art models. With our efficient annotation pipeline, it also provides additional fine-grained pixel-wise annotations and benchmark results compared with [36] and [39]. + +# 3. Datasets + +In this work, we aim at building a dataset for omnipotent city understanding from multi-level and multi-view images. Our proposed OmniCity dataset contains 108,600 images of multiple views, which are collected from 25K geo-locations in New York City. Compared with existing datasets, OmniCity requires much fewer human efforts for street-level image annotation, contains more diverse annotation types for both 2D and 3D tasks, provides richer building semantics at instance segmentation level, and possesses higher scalability for new annotation type supplement and expansion to other cities. The details of data collection, annotation, and statistics are introduced as follows. + +# 3.1. Data collection + +As shown in Figure 1, our OmniCity dataset is collected from six selected regions of New York. We download the panorama images in the six selected regions using google street view download 360, with a step distance of 65 meters. The regions for collecting the training and test samples are denoted by red and blue, respectively. We save the geographic coordinates, collection time, panorama id, north rotation, and zoom level for each panorama image. For each panorama site, we collect its corresponding google earth images of three acquisition dates according to the geographic coordinates, constituting three groups of satellite-level datasets with three cases of off-nadir view angles (small/medium/high, denoted by V1/V2/V3). For the annotation data sources, we collect the meta information from PLUTO and OpenStreetMap (OSM). The New York City is hierarchically formed by blocks, lots, and buildings. Each building can be identified by a specific block-lot id. In PLUTO, each lot (building) is associated with rich information, e.g. land use, year built, number of floors, etc. The OSM data contains footprint and height information while lacking in land use information for most buildings. Considering the characteristics of the two data sources, we align the land use attribute (from PLUTO) with the building footprint and height (from OSM) using the geographical coordinates. Overall, each building is assigned with a block-lot id, a land use category, a height value, and the geographical coordinates of a footprint polygon, which will be used as the reference label maps for panorama image annotation. + +# 3.2. Data annotation + +Figure 2 shows the panorama image annotation tool proposed in our study. The street-level panorama and satellite images are naturally well-aligned at image-level between the central coordinates from satellite views and the camera pose of panorama views. The instance-level annotations of different views (e.g., building facade and footprint) are further aligned during the annotation process. According to the geo-transformation relation between different views [25], + +we calculate the angle between the north (upward) direction and each building footprint (in the satellite image), and the angle between the north direction (saved in the panorama file) and each building facade in the horizontal axis. The footprint and facade with the same angle value are aligned to the same building, providing the auxiliary information for the manual annotation process. + +The annotation pipeline includes four stages: (1) Image selection, i.e., select the panorama images that are essential to be annotated according to building coverage, occlusion extent, etc; (2) Segmentation annotation, i.e., adjust the floor/top line to fit the bottom/roof of each building, and add the boundary split line considering both auxiliary information and building appearance (e.g. texture discrepancy, doors, etc.); (3) Attribute assignment, i.e., add the attributes (instance ID, block-lot id and land use type) for each building plane; (4) Quality assessment, i.e., check the annotation quality and remove the unqualified images. + +In addition, the mono-view images and annotations are automatically generated from those of the panorama images via view transformation. For each panorama image, we select three views using three x-axis angles (-170, 10 and 170, in the range of [-180, 180]) and a fixed y-axis angle of 0. Then we design an image selection rule to filter out the unexpected images of which the buildings are distributed in only one side, building area proportion is smaller than $10\%$ , or the building plane quantity is smaller than 2. The remaining images constitute the final mono-view dataset of our experiments. For satellite-level tasks, the annotations are already well-aligned with the meta information from PLUTO and OSM. Each building footprint on a satellite image is assigned with a block-lot id, a land use category, a height value, and the pixel coordinates of the footprint polygon. + +# 3.3. Statistics of the proposed dataset + +The whole dataset contains three sub-datasets (in small, medium and large view angles) for satellite-level tasks and two sub-datasets (in panorama and mono views) for street-level tasks. For satellite-level datasets, the three sub-datasets of images collected from the 25K geo-locations constitute 75K images in total, which are cropped by $512 \times 512$ pixels according to the coverage area of the corresponding panorama image. For the street-level datasets, the panorama dataset contains 18K images in $512 \times 1024$ pixels, which are selected from the initial 25K panorama images during the image selection and quality assessment phases. Similarly, the mono-view dataset obtained via view transformation and image selection contains 156,00 monoview images in $512 \times 512$ pixels. The ratio of train/test splits is set as 4:1 for all five sub-datasets. In OmniCity dataset, the initial land use categories of PLUTO with similar characteristics and low quantities are merged into one category, resulting in 7 land use categories in total. The + +categories of $1/2$ family building ( $\sim 17\%$ ), walk-up building ( $\sim 23\%$ ) and the mixed residential/commercial building categories ( $\sim 29\%$ ) take up a larger proportion compared with the other four categories, i.e., elevator buildings ( $\sim 7\%$ ), office buildings ( $\sim 10\%$ ), industrial/transportation/utility buildings ( $\sim 11\%$ ) and others ( $\sim 1\%$ ). For the distribution of building height, most are between 1 and 25 meters while a small percentage ( $\sim 1\%$ ) reaches over 50 meters. + +# 4. Benchmark results + +In this section, we provide a variety of benchmarks for multiple satellite and street-level tasks. The satellite-level tasks in our experiments include building footprint segmentation and height estimation. For both tasks, we conduct experiments on the satellite images with three view angles. For the street-level tasks, we conduct two instance segmentation tasks (i.e., land use and building instance segmentation) on the panorama images, and three instance segmentation tasks (i.e., land use / building instance / plane segmentation) on mono-view images. Please note that these are only preliminary experimental results on OmniCity dataset. More benchmarks of latest models and additional tasks will be continuously updated on OmniCity homepage. + +# 4.1. Experimental setting + +The experiments are mainly based on mmdetection [7] with the recommended hyper-parameter settings. We select Mask R-CNN [15] as the baseline method for segmentation tasks, and provide a comparison of different methods in Table 6, including Mask Scoring R-CNN (MS R-CNN) [19], Cascade Mask R-CNN (Cascade) [5], Content-Aware ReAssembly of FEatures (CARAFE) [35], and Hybrid Task Cascade (HTC) [6]. Specifically, we use ResNet-50 [16] with FPN [23] pre-trained on the ImageNet [29] as the backbone for all instance segmentation models. All models are trained on 8 NVIDIA Tesla V100 GPUs for 12 epochs, with a batch size of 16, a learning rate starting from 0.02 and decreasing by a factor of 0.1 from the $8^{th}$ to $11^{th}$ epoch, and the stochastic gradient descent (SGD) optimizer with a weight decay of $10^{-4}$ and a momentum of 0.9. For the height estimation task, we evaluate the performance of two widely-used monocular depth estimation methods on the satellite images of three view angles, i.e. Structure-Aware Residual Pyramid Network (SARPN) [8] and Deep Ordinal Regression Network (DORN) [13], which are trained on 4 NVIDIA Tesla V100 GPUs for 20 epochs. SARPN is trained with a batch size of 8, a learning rate starting from $10^{-4}$ and reduced by $10\%$ every 5 epochs, and the Adam optimizer with $\beta_{1} = 0.9$ , $\beta_{2} = 0.999$ , and a weight decay of $10^{-4}$ . DORN is trained with a batch size of 4, a base learning rate of $10^{-4}$ and the power of 0.9, using SGD optimizer with a weight decay of 0.0005 and a momentum of 0.9. + +![](images/e9d32052e9ca393c4f493081008bbb5ac67b6e756b5bc6a5058f74d9a2e20232.jpg) +Small view + +![](images/2f658c1f3183d8110deec368391d8b8cac7c97b5d6d1b55270d41dc716fc88ee.jpg) +Medium view + +![](images/46baf51ab398638283631313bf8346e7752c3e33802b373f441d13fd9c4230b1.jpg) +Large view + +![](images/0659340c6af97b34b589185c5a4a9a6896bb1bc4b9db2bb2f6325ca6dc12697d.jpg) +Satellite image + +![](images/73cc5a7d27527e98d435fac9e135f6d1de42946dcf42884ce9b62cc49c3992a7.jpg) +Ground Truth + +![](images/93db0a7f5a73a502623284e9df92ef96db26b3ff6f289dcc57d285f7cbabad66.jpg) +DORN - small + +![](images/60ec9f7b70b57524dd1a16989d1f1a0e074a4793a2e388d10606610a238e9eed.jpg) +SARPN - small +Figure 3. Qualitative results of satellite-level tasks. The yellow/cyan/red polygons denote TP/FP/FN buildings. + +![](images/3eea1ca0371f98394bb37b58009e6f89d3aa955a02c71452177579b2ecf06f06.jpg) +SARPN - medium + +![](images/8dc71005b97332c4f13564bb4f8e5f1faf3240499b3806775d41e2261bb93cb6.jpg) +SARPN - large + +![](images/4ea60c47fe3bc4e1b3b546a2a26a9c8f261764e2ab0b228c57fbd8e5b4fbfde1.jpg) +High +Low + +# 4.2. Experimental results of satellite-level tasks + +Table 2 shows the baseline results of the satellite-level instance segmentation task, which are evaluated using both COCO [24] and SpaceNet [11] evaluation metrics. The footprint segmentation performance is the best for satellite images with a small view angle, and deteriorates seriously when the view angle gets larger. For all three cases, the prediction score is the highest for buildings with large areas and the lowest for small buildings. Figure 3 provides a qualitative comparison of the footprint segmentation results on satellite images of three types of view angle. The large view angle results in great difficulties for extracting accurate footprint boundaries, due to the partial invisibility of building footprint, the serious shadow effects, etc. The height estimation performance is evaluated in terms of the mean absolute error (denoted by MAE), mean square error (denoted by MSE), and root mean square error (denoted by RMSE), which are commonly used metrics for depth estimation. All metrics are measured in meters at pixel level. Table 3 and Figure 3 show the height estimation results obtained from DORN [13] and SARPN [8] for satellite images with different view angles. Results demonstrate that DORN obtains better results compared with SARPN for all three cases. Both methods achieve the best performance for satellite images with a medium view angle (V2) compared with the other two cases. The footprint and roof have more overlaps and provide less building structure information for satellite images with a small view angle, while the shadow and parallax effect become more serious with + +Table 2. Quantitative results of instance segmentation for satellite images with different view angles. + +
ViewMetrics of various thresholdsthreshold = 0.5
APAP50AP75APSAPMAPLPRF1
V129.766.023.515.933.936.776.966.371.2
V223.756.616.111.527.230.373.955.063.1
V318.951.49.69.121.525.370.751.759.7
+ +Table 3. Quantitative results of height estimation for satellite images with different view angles. + +
ViewSARPN [8]DORN [13]
MAEMSERMSEMAEMSERMSE
V116.18870.3429.5012.71670.5225.89
V213.75694.1726.3512.24628.0625.06
V315.32823.0128.6913.40730.6727.03
+ +the increase of off-nadir view angle. The above aspects result in challenges for the accurate estimation of building height for satellite images with small and large view angles. + +# 4.3. Experimental results of street-level tasks + +We analyze the performance of multiple segmentation tasks on street-level panorama and mono-view images. Table 4 and Figure 4 show the experimental results of panorama-view images on two segmentation tasks, of which the performance on building instance segmentation task (denoted by Instance Seg.) is significantly superior to the fine-grained land use segmentation task (denoted by Landuse Seg.). Table 5 and Figure 5 show the experimental results of mono-view images on three different tasks, i.e., landuse segmentation, instance segmentation, and plane segmentation. Similar to the results of panorama-view images, the baseline method achieves much higher scores for the two binary segmentation tasks (plane and instance segmentation) compared with the fine-grained land use segmentation task. The qualitative results also demonstrate that the baseline method has difficulties in identifying the accurate land use type of some building instances. + +Table 6 shows the land use segmentation results obtained from the five methods on street-level panorama images. HTC achieves the best performance for both overall and category metrics followed by MS R-CNN, which indicates that the cascade structure and mask scoring strategy can effectively improves the fine-grained segmentation performance of building instances. From the metrics of each category, we can find that C2, C4 and C5 (i.e., Walkup Buildings, Mixed Residential/Commercial, and Office Buildings) have better performance compared with C1, C3 and C6 (i.e., 1/2 Family Buildings, Elevator Buildings and Industrial/Transportation/Utility) for all methods, with the AP ranging from $20\%$ to $40\%$ . The category of Others has an extremely low AP score due to the small sample quantity. Compared with the residential buildings, the mixed Residential/Commercial and Office buildings have more special + +![](images/b541a0ef87e31a184621fd24b40c1f94531f444feeae28e00863e223bbc3ced9.jpg) +Figure 4. Qualitative results of street-level tasks (i.e., instance segmentation and land use segmentation) on panorama images. + +Table 4. Quantitative results on street-level panorama images. + +
TaskAPAP50AP75APSAPMAPL
Landuse Seg.26.034.728.50.312.030.4
Instance Seg.66.786.572.51.740.274.1
+ +Table 5. Quantitative results on street-level mono-view images. + +
TaskAPAP50AP75APSAPMAPL
Landuse Seg.23.932.126.70.310.627.5
Instance Seg.68.388.873.83.233.376.1
Plane Seg.65.187.471.05.040.773.8
+ +characteristics (i.e. the boundary between first and above floors, facade design, and building structure), contributing to the superior AP scores of these two categories. The performance of the above six categories is also concordant with the ratio of the sample quantity. + +# 4.4. Results analysis and discussions + +In this section we summarize the limitations of existing methods for satellite and street-level tasks on our OmniCity dataset. For building footprint segmentation, the performance of existing methods get worse with the increasing of off-nadir view angle, which might due to the serious parallax and shadow effect on the satellite images with a large off-nadir view angle. For the height estimation task, most existing methods directly apply the monocular depth estimation methods to remote sensing scene. These methods usually produce poor results on the invisible side of the footprint boundary especially for high-rise buildings on very off-nadir images. In addition, most existing methods use deep neural networks to regress continuous values but the actual height values are discrete for building and non-building areas, resulting in difficulties for network training and extra efforts for converting the continuous prediction values into discrete height values via post-processing. + +![](images/3375a710fbf8843139a54b93e66df7291a245b83620c5ffd80638b67ef8a0f6a.jpg) +Figure 5. Qualitative results of street-level tasks (i.e., land use /instance/plane segmentation) on mono-view images. + +For the street-level tasks based on panorama images, existing methods target at general instance segmentation tasks for commonly-used datasets, e.g. COCO [24], CityScapes [10], BDD100K [44], etc. These datasets often have a single view and a narrow Field of View (FoV). However, for panorama images, the special properties such as the wide FoV covering full 360-degree in the horizontal direction, are not taken into consideration in the design of existing methods. For both mono-view and panorama images, existing methods have difficulties in accurately recognizing the building instance with a small area (e.g. buildings located in the side of the main parts), the land use categories with a small number of building instances, and the categories that are easily confused (e.g. 1/2 Family Buildings and Walk-Up Buildings, Mixed Residential/Commercial and Office Buildings). Figure 6 shows some typical failure cases of the current benchmark methods. New instance segmentation methods should be designed for solving the above limitations considering the characteristics of panorama images, building instances, fine-grained categories, etc. + +# 5. Potential of the OmniCity dataset + +Our proposed OmniCity dataset demonstrates great potential for facilitating city understanding, machine perception, and generative modeling researches in many aspects. + +First, it can serve as a new dataset for the existing tasks such as image geo-localization/synthesis and segmentation/detection of buildings/trees/land use from cross-view images. OmniCity provides additional annotations that are not contained in existing datasets, e.g. the building instances and fine-grained categories on street-level images, which can promote new methods to explore and leverage + +Table 6. Quantitative results of different methods for fine-grained land use segmentation on street-level panorama images. + +
MethodOverall MetricsMetrics of each category
APAP50AP75APSAPMAPLC1C2C3C4C5C6C7
Mask R-CNN [15]26.034.728.50.312.030.419.637.525.839.236.922.20.8
MS R-CNN [19]27.135.829.80.112.431.522.539.126.240.838.021.71.2
Cascade [5]25.933.828.30.211.430.52038.32538.536.722.10.3
CARAFE [35]25.934.528.50.111.930.219.637.324.939.937.221.50.8
HTC [6]27.235.729.90.312.432.020.838.727.239.938.424.51.2
+ +the new annotations to improve the performance of these tasks. The annotations are organized in a unified version, which means multiple tasks can be performed on a single image, and thus can well support the multi-task learning setting. Additionally, since the building instances are directly linked with the urban planning data using block-lot id, it is easy to be enriched with more annotation types from other urban datasets, especially those for social and urban studies. + +Second, OmniCity provides a new application scenario or problem setting to existing tasks. For line segment detection and wireframe parsing tasks, existing datasets contains densely distributed line segment and wireframe labels, while our OmniCity focuses on the main line segments and wireframes on the outlines instead of the inner ones, resulting in a much sparser format. The serious shelters from the trees and vehicles also bring challenges to these tasks. New line segment detection and wireframe parsing methods should be designed for the OmniCity scenario. + +Moreover, OmniCity facilitates new tasks for city reconstruction and simulation. Treating each panorama as a unique city scene, a complete 3D model representing such a local scene is available, as also shown in the second window panel in our tool (Figure 2). These 3D models are stored in abstract vector formats, with clean and clear vertices indicating the building facades, streets, etc., which are suitable for many shape generation tasks represented with graphs. + +Finally, with the well-aligned satellite and street-level images as well as the various annotation types, novel city reconstruction tasks, e.g., 3D building reconstruction from cross-view images, can be derived for producing holistic 3D buildings with both fine-grained semantic category and precise geometry information (vector 3D model). Unlike the existing studies that only target at 3D reconstruction from monocular or multi-view remote sensing imagery, new methods should be designed to leverage the additional information from street-level images to improve 3D reconstruction and semantic prediction for OmniCity scenario. + +# 6. Conclusion + +In this paper, we have proposed OmniCity, a new dataset for omnipotent city understanding from satellite and street-level images of multiple views. The dataset contains over 100K images collected from 25K geo-locations in New + +![](images/fd54acc1ace974a66bc0e83340c9b91f98a941b8feb15c81f8abd604c016c358.jpg) +(a)GT:5-Office + +![](images/79fdc6e4eba59310541e8264db6edf336b6e06c37139b2ab64270f26042b255e.jpg) +(b) GT: 4-Mixed (three instances) + +![](images/3e01033128d2fa9375d63e7f704ee13055bb5c2f160aefd991d59ef773c584c7.jpg) +(c) GT: 3-Elevator + +![](images/39b456a200f887e68fcf2d2b55a9cff239d9a73d5192af0d151aba7a3775f76a.jpg) +(d)GT:5-Office +Figure 6. Typical failure cases of the current benchmark methods. + +![](images/bfa689068c84c71966380a96dc8822117e5cf5d3507c03dee0d289e2131bb475.jpg) +(e)GT:5-Office + +![](images/dfdfa8efe3b89c78e888ce0eb8b0b70d8ec7ec2051f7fb45517d2807648b2ef0.jpg) +(f) GT: 7-Others + +![](images/8d633283896b7e9e409109f760c2b09c0a7742772a950226c652cc462a03d087.jpg) +(g) GT: 1-Family + +York City, of which the annotations are generated from both existing label maps and our proposed annotation pipeline. We provide benchmark experimental results for multiple tasks and data sources based on state-of-the-art methods and analyze their limitations. We believe that OmniCity will not only promote new algorithms and application scenarios for existing tasks, but facilitate novel tasks for 3D city reconstruction and simulation. In our future work, we will keep updating the OmniCity dataset and the benchmarks in the following aspects. Owing to the proposed annotation pipeline and the unified annotations with a vector format and rich meta information (geo-locations, block-lot id, etc.), OmniCity can be efficiently supplemented with more properties of buildings and other geographical object types (roads, sidewalks, trees, green space, etc.), and extended to other cities of different countries. The benchmark results of more state-of-the-art models and new tasks will be provided accordingly. Based on the rich annotation and view types of OmniCity, we also plan to develop new methods for existing and novel tasks, such as object detection, instance segmentation, and 3D reconstruction from cross-view images. + +Acknowledgements. This project was funded in part by National Natural Science Foundation of China (No. 42201358, No. U22B2011 and No. 41820104006), Shanghai AI Laboratory, CUHK Interdisciplinary AI Research Institute, and the Centre for Perceptual and Interactive Intelligence (CPII) Ltd under the Innovation and Technology Commission (ITC)'s InnoHK. + +# References + +[1] Seyed Majid Azimi, Corentin Henry, Lars Sommer, Arne Schumann, and Eleonora Vig. Skyscapes fine-grained semantic understanding of aerial scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7393-7403, 2019. +[2] Sara Beery, Guanhang Wu, Trevor Edwards, Filip Pavetic, Bo Majewski, Shreyasee Mukherjee, Stanley Chan, John Morgan, Vivek Rathod, and Jonathan Huang. The auto arborist dataset: A large-scale benchmark for multiview urban forest monitoring under domain shift. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21294-21307, 2022. +[3] Markus Braun, Sebastian Krebs, Fabian B. Flohr, and Dariu M. Gavrila. Eurocity persons: A novel benchmark for person detection in traffic scenes. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-1, 2019. +[4] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. +[5] Zhaowei Cai and Nuno Vasconcelos. Cascade r-cnn: High quality object detection and instance segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, page 1-1, 2019. +[6] Kai Chen, Jiangmiao Pang, Jiaqi Wang, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jianping Shi, Wanli Ouyang, et al. Hybrid task cascade for instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4974-4983, 2019. +[7] Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jiarui Xu, et al. Mmdetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155, 2019. +[8] Xiaotian Chen, Xuejin Chen, and Zheng-Jun Zha. Structure-aware residual pyramid network for monocular depth estimation. In Proceedings of the 28th International Joint Conference on Artificial Intelligence, pages 694-700, 2019. +[9] Gordon Christie, Rodrigo Rene Rai Munoz Abujder, Kevin Foster, Shea Hagstrom, Gregory D Hager, and Myron Z Brown. Learning geocentric object pose in oblique monocular images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14512-14520, 2020. + +[10] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3213-3223, 2016. +[11] Ilke Demir, Krzysztof Koperski, David Lindenbaum, Guan Pang, Jing Huang, Saikat Basu, Forest Hughes, Devis Tuia, and Ramesh Raskar. Deep globe 2018: A challenge to parse the earth through satellite images. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2018. +[12] Tian Feng, Quang-Trung Truong, Duc Thanh Nguyen, Jing Yu Koh, Lap-Fai Yu, Alexander Binder, and Sai-Kit Yeung. Urban zoning using higher-order markov random fields on multi-view imagery data. In Proceedings of the European Conference on Computer Vision (ECCV), pages 614–630, 2018. +[13] Huan Fu, Mingming Gong, Chaohui Wang, Kayhan Bat-manghelich, and Dacheng Tao. Deep ordinal regression network for monocular depth estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2002-2011, 2018. +[14] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE conference on computer vision and pattern recognition, pages 3354-3361. IEEE, 2012. +[15] Kaiming He, Georgia Gkioxari, Piotr Dollár, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision (CVPR), pages 2961-2969, 2017. +[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. +[17] Sixing Hu, Mengdan Feng, Rang MH Nguyen, and Gim Hee Lee. Cvm-net: Cross-view matching network for image-based ground-to-aerial geo-localization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7258-7267, 2018. +[18] Xinyu Huang, Peng Wang, Xinjing Cheng, Dingfu Zhou, Qichuan Geng, and Ruigang Yang. The apolloscape open dataset for autonomous driving and its application. IEEE transactions on pattern analysis and machine intelligence, 42(10):2702-2719, 2019. +[19] Zhaojin Huang, Lichao Huang, Yongchao Gong, Chang Huang, and Xinggang Wang. Mask scoring r-cnn. In Pro- + +ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6409-6418, 2019. +[20] Weijia Li, Conghui He, Jiarui Fang, Juepeng Zheng, Hao-huan Fu, and Le Yu. Semantic segmentation-based building footprint extraction using very high-resolution satellite images and multi-source gis data. Remote Sensing, 11(4):403, 2019. +[21] Weijia Li, Lingxuan Meng, Jinwang Wang, Conghui He, Gui-Song Xia, and Dahua Lin. 3d building reconstruction from monocular remote sensing images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12548-12557, 2021. +[22] Weijia Li, Wenqian Zhao, Huaping Zhong, Conghui He, and Dahua Lin. Joint semantic-geometric learning for polygonal building segmentation. In AAAI, 2021. +[23] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition (CVPR), pages 2117-2125, 2017. +[24] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. +[25] Xiaohu Lu, Zuoyue Li, Zhaopeng Cui, Martin R Oswald, Marc Pollefeys, and Rongjun Qin. Geometry-aware satellite-to-ground image synthesis for urban areas. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 859-867, 2020. +[26] Gerhard Neuhold, Tobias Ollmann, Samuel Rota Bulo, and Peter Kontschieder. The mapillary vistas dataset for semantic understanding of street scenes. In Proceedings of the IEEE international conference on computer vision, pages 4990-4999, 2017. +[27] Krishna Regmi and Ali Borji. Cross-view image synthesis using conditional gans. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3501-3510, 2018. +[28] Krishna Regmi and Mubarak Shah. Bridging the domain gap for ground-to-aerial image matching. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 470-479, 2019. +[29] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision (IJCV), 115(3):211-252, 2015. + +[30] Yujiao Shi, Dylan John Campbell, Xin Yu, and Hongdong Li. Geometry-guided street-view panorama synthesis from satellite imagery. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022. +[31] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2446-2454, 2020. +[32] Zheng Tang, Milind Naphade, Ming-Yu Liu, Xiaodong Yang, Stan Birchfield, Shuo Wang, Ratnesh Kumar, David Anastasiu, and Jenq-Neng Hwang. Cityflow: A city-scale benchmark for multi-target multi-camera vehicle tracking and re-identification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8797–8806, 2019. +[33] Yicong Tian, Chen Chen, and Mubarak Shah. Cross-view image matching for geo-localization in urban environments. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3608-3616, 2017. +[34] Adam Van Etten, Daniel Hogan, Jesus Martinez Manso, Jacob Shermeyer, Nicholas Weir, and Ryan Lewis. The multitemporal urban development spacenet dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6398-6407, 2021. +[35] Jiaqi Wang, Kai Chen, Rui Xu, Ziwei Liu, Chen Change Loy, and Dahua Lin. Carafe: Content-aware reassembly of features. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3007-3016, 2019. +[36] Shenlong Wang, Min Bai, Gellert Mattyus, Hang Chu, Wenjie Luo, Bin Yang, Justin Liang, Joel Cheverie, Sanja Fidler, and Raquel Urtasun. Toronto: Seeing the world with a million eyes. In 2017 IEEE International Conference on Computer Vision (ICCV), pages 3028-3036. IEEE, 2017. +[37] Jan D Wegner, Steven Branson, David Hall, Konrad Schindler, and Pietro Perona. Cataloging public objects using aerial and street-level images-urban trees. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6014–6023, 2016. +[38] Nicholas Weir, David Lindenbaum, Alexei Bastidas, Adam Van Etten, Sean McPherson, Jacob Shermeyer, Varun Kumar, and Hanlin Tang. Spacenet mvoi: A multi-view overhead imagery dataset. In Proceedings of the IEEE/cvf international conference on computer vision, pages 992-1001, 2019. +[39] Zbigniew Wojna, Krzysztof Maziarz, Łukasz Jocz, Robert Pałuba, Robert Kozikowski, and Iason Kokkinos. Holis- + +tic multi-view building analysis in the wild with projection pooling. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 2870-2878, 2021. +[40] Scott Workman, M Usman Rafique, Hunter Blanton, and Nathan Jacobs. Revisiting near/remote sensing with geospatial attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1778-1787, 2022. +[41] Kailun Yang, Xinxin Hu, Luis M Bergasa, Eduardo Romera, and Kaiwei Wang. Pass: Panoramic annular semantic segmentation. IEEE Transactions on Intelligent Transportation Systems, 21(10):4171-4185, 2019. +[42] Kailun Yang, Xinxin Hu, and Rainer Stiefelhagen. Is context-aware cnn ready for the surroundings? panoramic semantic segmentation in the wild. IEEE Transactions on Image Processing, 30:1866-1881, 2021. +[43] Xiaoqing Ye, Mao Shu, Hanyu Li, Yifeng Shi, Yingying Li, Guangjie Wang, Xiao Tan, and Errui Ding. Rope3d: The roadside perception dataset for autonomous driving and monocular 3d object detection task. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21341-21350, 2022. +[44] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darryll. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2636-2645, 2020. +[45] Menghua Zhai, Zachary Bessinger, Scott Workman, and Nathan Jacobs. Predicting ground-level scene layout from aerial imagery. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 867-875, 2017. +[46] Chiqun Zhang, Dragomir Yankov, Chun-Ting Wu, Simon Shapiro, Jason Hong, and Wei Wu. What is that building? an end-to-end system for building recognition from streetside images. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pages 2425-2433, 2020. +[47] Yichao Zhou, Jingwei Huang, Xili Dai, Shichen Liu, Linjie Luo, Zhili Chen, and Yi Ma. Holicity: A city-scale data platform for learning holistic 3d structures. arXiv preprint arXiv:2008.03286, 2020. +[48] Sijie Zhu, Taojiannan Yang, and Chen Chen. Vigor: Cross-view image geo-localization beyond one-to-one retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3640-3649, 2021. + +[49] Yi Zhu, Xueqing Deng, and Shawn Newsam. Fine-grained land use classification at the city scale using ground-level images. IEEE Transactions on Multimedia, 21(7):1825-1838, 2019. \ No newline at end of file diff --git a/2023/OmniCity_ Omnipotent City Understanding With Multi-Level and Multi-View Images/images.zip b/2023/OmniCity_ Omnipotent City Understanding With Multi-Level and Multi-View Images/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7d452b1256bc711839f272e56b51c181477f8bf7 --- /dev/null +++ b/2023/OmniCity_ Omnipotent City Understanding With Multi-Level and Multi-View Images/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8b688abb702880f51c429ed7b7cf3fa4cb3335626c8d4f40bedba7c352e2ca65 +size 720402 diff --git a/2023/OmniCity_ Omnipotent City Understanding With Multi-Level and Multi-View Images/layout.json b/2023/OmniCity_ Omnipotent City Understanding With Multi-Level and Multi-View Images/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..356edbb32202936bb593c6dd311270ba4053f9c5 --- /dev/null +++ b/2023/OmniCity_ Omnipotent City Understanding With Multi-Level and Multi-View Images/layout.json @@ -0,0 +1,8281 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 160, + 103, + 434, + 139 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 103, + 434, + 139 + ], + "spans": [ + { + "bbox": [ + 160, + 103, + 434, + 139 + ], + "type": "text", + "content": "OmniCity: Omnipotent City Understanding with Multi-level and Multi-view Images" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 160, + 460, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 160, + 460, + 175 + ], + "spans": [ + { + "bbox": [ + 133, + 160, + 460, + 175 + ], + "type": "text", + "content": "Weijia " + }, + { + "bbox": [ + 133, + 160, + 460, + 175 + ], + "type": "inline_equation", + "content": "\\mathsf{Li}^1" + }, + { + "bbox": [ + 133, + 160, + 460, + 175 + ], + "type": "text", + "content": ", Yawen " + }, + { + "bbox": [ + 133, + 160, + 460, + 175 + ], + "type": "inline_equation", + "content": "\\mathsf{Lai}^2" + }, + { + "bbox": [ + 133, + 160, + 460, + 175 + ], + "type": "text", + "content": ", Linning " + }, + { + "bbox": [ + 133, + 160, + 460, + 175 + ], + "type": "inline_equation", + "content": "\\mathsf{Xu}^3" + }, + { + "bbox": [ + 133, + 160, + 460, + 175 + ], + "type": "text", + "content": ", Yuanbo " + }, + { + "bbox": [ + 133, + 160, + 460, + 175 + ], + "type": "inline_equation", + "content": "\\mathsf{Xiangli}^3" + }, + { + "bbox": [ + 133, + 160, + 460, + 175 + ], + "type": "text", + "content": ", Jinhua " + }, + { + "bbox": [ + 133, + 160, + 460, + 175 + ], + "type": "inline_equation", + "content": "\\mathsf{Yu}^1" + }, + { + "bbox": [ + 133, + 160, + 460, + 175 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 187, + 175, + 405, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 175, + 405, + 189 + ], + "spans": [ + { + "bbox": [ + 187, + 175, + 405, + 189 + ], + "type": "text", + "content": "Conghui He" + }, + { + "bbox": [ + 187, + 175, + 405, + 189 + ], + "type": "inline_equation", + "content": "^{2,4*}" + }, + { + "bbox": [ + 187, + 175, + 405, + 189 + ], + "type": "text", + "content": ", Gui-Song Xia" + }, + { + "bbox": [ + 187, + 175, + 405, + 189 + ], + "type": "inline_equation", + "content": "^{5*}" + }, + { + "bbox": [ + 187, + 175, + 405, + 189 + ], + "type": "text", + "content": ", Dahua Lin" + }, + { + "bbox": [ + 187, + 175, + 405, + 189 + ], + "type": "inline_equation", + "content": "^{3,4}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 83, + 189, + 511, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 189, + 511, + 203 + ], + "spans": [ + { + "bbox": [ + 83, + 189, + 511, + 203 + ], + "type": "text", + "content": "1Sun Yat-Sen University, 2SenseTime Research, 3The Chinese University of Hong Kong," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 141, + 203, + 452, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 203, + 452, + 217 + ], + "spans": [ + { + "bbox": [ + 141, + 203, + 452, + 217 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 141, + 203, + 452, + 217 + ], + "type": "text", + "content": "Shanghai Artificial Intelligence Laboratory, " + }, + { + "bbox": [ + 141, + 203, + 452, + 217 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 141, + 203, + 452, + 217 + ], + "type": "text", + "content": "Wuhan University" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 70, + 220, + 521, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 220, + 521, + 231 + ], + "spans": [ + { + "bbox": [ + 70, + 220, + 521, + 231 + ], + "type": "text", + "content": "liweij29@mail.sysu.edu.cn, alanlyawen@gmail.com, {x1020,xy019,dhlin}@ie.cuhk.edu.hk," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 96, + 234, + 492, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 234, + 492, + 245 + ], + "spans": [ + { + "bbox": [ + 96, + 234, + 492, + 245 + ], + "type": "text", + "content": "yujh56@mail2.sysu.edu.cn, heconghui@sensetime.com, guisong.xia@whu.edu.cn" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 53, + 257, + 164, + 364 + ], + "blocks": [ + { + "bbox": [ + 53, + 257, + 164, + 364 + ], + "lines": [ + { + "bbox": [ + 53, + 257, + 164, + 364 + ], + "spans": [ + { + "bbox": [ + 53, + 257, + 164, + 364 + ], + "type": "image", + "image_path": "17ab3d5166ad6ac2bd9135a4b1d471e714e762c8ce33ac287a2caadc2464873d.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 65, + 365, + 153, + 372 + ], + "lines": [ + { + "bbox": [ + 65, + 365, + 153, + 372 + ], + "spans": [ + { + "bbox": [ + 65, + 365, + 153, + 372 + ], + "type": "text", + "content": "Selected regions for image collection" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 55, + 373, + 164, + 480 + ], + "blocks": [ + { + "bbox": [ + 55, + 373, + 164, + 480 + ], + "lines": [ + { + "bbox": [ + 55, + 373, + 164, + 480 + ], + "spans": [ + { + "bbox": [ + 55, + 373, + 164, + 480 + ], + "type": "image", + "image_path": "8e739941d17d503c131a3a59c44244dacb429df05aefc4369495a32451a2e410.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 481, + 151, + 487 + ], + "lines": [ + { + "bbox": [ + 67, + 481, + 151, + 487 + ], + "spans": [ + { + "bbox": [ + 67, + 481, + 151, + 487 + ], + "type": "text", + "content": "Existing label maps and street-view" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 67, + 487, + 151, + 495 + ], + "lines": [ + { + "bbox": [ + 67, + 487, + 151, + 495 + ], + "spans": [ + { + "bbox": [ + 67, + 487, + 151, + 495 + ], + "type": "text", + "content": "image viewpoints of a zoomed area" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 175, + 258, + 261, + 327 + ], + "blocks": [ + { + "bbox": [ + 175, + 258, + 261, + 327 + ], + "lines": [ + { + "bbox": [ + 175, + 258, + 261, + 327 + ], + "spans": [ + { + "bbox": [ + 175, + 258, + 261, + 327 + ], + "type": "image", + "image_path": "df3096eba4b975b6853f5d50bfba1bda95ab2acbf866ed0745c1d15fe2e5ba9e.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 182, + 328, + 253, + 334 + ], + "lines": [ + { + "bbox": [ + 182, + 328, + 253, + 334 + ], + "spans": [ + { + "bbox": [ + 182, + 328, + 253, + 334 + ], + "type": "text", + "content": "Different satellite-level views" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 176, + 337, + 261, + 405 + ], + "blocks": [ + { + "bbox": [ + 176, + 337, + 261, + 405 + ], + "lines": [ + { + "bbox": [ + 176, + 337, + 261, + 405 + ], + "spans": [ + { + "bbox": [ + 176, + 337, + 261, + 405 + ], + "type": "image", + "image_path": "9baf49218cb34b86d8791dca9c85811d9039793c833f415a093c12d45dd68be1.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 188, + 407, + 247, + 412 + ], + "lines": [ + { + "bbox": [ + 188, + 407, + 247, + 412 + ], + "spans": [ + { + "bbox": [ + 188, + 407, + 247, + 412 + ], + "type": "text", + "content": "Transformation relations" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 175, + 416, + 260, + 485 + ], + "blocks": [ + { + "bbox": [ + 175, + 416, + 260, + 485 + ], + "lines": [ + { + "bbox": [ + 175, + 416, + 260, + 485 + ], + "spans": [ + { + "bbox": [ + 175, + 416, + 260, + 485 + ], + "type": "image", + "image_path": "6745b714f3acb034ea251eed71f2dd5b02da49c9c6df58714db8ba25f53e1b1b.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 186, + 486, + 250, + 493 + ], + "lines": [ + { + "bbox": [ + 186, + 486, + 250, + 493 + ], + "spans": [ + { + "bbox": [ + 186, + 486, + 250, + 493 + ], + "type": "text", + "content": "Different street-level views" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 45, + 507, + 547, + 552 + ], + "lines": [ + { + "bbox": [ + 45, + 507, + 547, + 552 + ], + "spans": [ + { + "bbox": [ + 45, + 507, + 547, + 552 + ], + "type": "text", + "content": "Figure 1. An overview of our proposed OmniCity dataset. The satellite and street-level images of our dataset are collected in the six selected regions. The black dots along the streets denote the viewpoint locations and the land-lot colors indicate different fine-grained categories provided by the existing label maps. Note that all the images in the right part correspond to the same geo-location, and the intrinsic transformation relation between the satellite and street-level panorama view is demonstrated in the middle part." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 272, + 258, + 541, + 326 + ], + "blocks": [ + { + "bbox": [ + 272, + 258, + 541, + 326 + ], + "lines": [ + { + "bbox": [ + 272, + 258, + 541, + 326 + ], + "spans": [ + { + "bbox": [ + 272, + 258, + 541, + 326 + ], + "type": "image", + "image_path": "b0ed4e5dc56c5a0b4c188d8b334b72d54421f6bd02944c6245a0d6cf2c89a721.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 340, + 327, + 471, + 334 + ], + "lines": [ + { + "bbox": [ + 340, + 327, + 471, + 334 + ], + "spans": [ + { + "bbox": [ + 340, + 327, + 471, + 334 + ], + "type": "text", + "content": "Satellite-level images and annotations of multiple views" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 271, + 336, + 541, + 404 + ], + "blocks": [ + { + "bbox": [ + 271, + 336, + 541, + 404 + ], + "lines": [ + { + "bbox": [ + 271, + 336, + 541, + 404 + ], + "spans": [ + { + "bbox": [ + 271, + 336, + 541, + 404 + ], + "type": "image", + "image_path": "73467ef54011833bcecb43d7d56ad6ca5320de447ed75077efae653eb7ef2902.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 340, + 406, + 471, + 413 + ], + "lines": [ + { + "bbox": [ + 340, + 406, + 471, + 413 + ], + "spans": [ + { + "bbox": [ + 340, + 406, + 471, + 413 + ], + "type": "text", + "content": "Street-level images and annotations of a panorama view" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 271, + 416, + 541, + 485 + ], + "blocks": [ + { + "bbox": [ + 271, + 416, + 541, + 485 + ], + "lines": [ + { + "bbox": [ + 271, + 416, + 541, + 485 + ], + "spans": [ + { + "bbox": [ + 271, + 416, + 541, + 485 + ], + "type": "image", + "image_path": "f6643e67e7001e20fcd2b4e68a133cce39629038f536021493c32683be155a65.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 335, + 486, + 476, + 493 + ], + "lines": [ + { + "bbox": [ + 335, + 486, + 476, + 493 + ], + "spans": [ + { + "bbox": [ + 335, + 486, + 476, + 493 + ], + "type": "text", + "content": "Street-level images and annotations of multiple mono-views" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "bbox": [ + 143, + 577, + 192, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 577, + 192, + 590 + ], + "spans": [ + { + "bbox": [ + 143, + 577, + 192, + 590 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 46, + 604, + 288, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 604, + 288, + 687 + ], + "spans": [ + { + "bbox": [ + 46, + 604, + 288, + 687 + ], + "type": "text", + "content": "This paper presents OmniCity, a new dataset for omnipotent city understanding from multi-level and multi-view images. More precisely, OmniCity contains multi-view satellite images as well as street-level panorama and monoview images, constituting over 100K pixel-wise annotated images that are well-aligned and collected from 25K geolocations in New York City. To alleviate the substantial" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 306, + 578, + 547, + 709 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 578, + 547, + 709 + ], + "spans": [ + { + "bbox": [ + 306, + 578, + 547, + 709 + ], + "type": "text", + "content": "pixel-wise annotation efforts, we propose an efficient street-view image annotation pipeline that leverages the existing label maps of satellite view and the transformation relations between different views (satellite, panorama, and mono-view). With the new OmniCity dataset, we provide benchmarks for a variety of tasks including building footprint extraction, height estimation, and building plane/instance/fine-grained segmentation. Compared with existing multi-level and multi-view benchmarks, OmniCity contains a larger number of images with richer annotation types and more views, provides more benchmark results" + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 703, + 139, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 703, + 139, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 703, + 139, + 712 + ], + "type": "text", + "content": "*Corresponding authors." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17397" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 45, + 72, + 288, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 72, + 288, + 180 + ], + "spans": [ + { + "bbox": [ + 45, + 72, + 288, + 180 + ], + "type": "text", + "content": "of state-of-the-art models, and introduces a new task for fine-grained building instance segmentation on street-level panorama images. Moreover, OmniCity provides new problem settings for existing tasks, such as cross-view image matching, synthesis, segmentation, detection, etc., and facilitates the developing of new methods for large-scale city understanding, reconstruction, and simulation. The OmniCity dataset as well as the benchmarks will be released at https://city-super.github.io/omnicity/." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 203, + 128, + 216 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 203, + 128, + 216 + ], + "spans": [ + { + "bbox": [ + 47, + 203, + 128, + 216 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 45, + 223, + 287, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 223, + 287, + 378 + ], + "spans": [ + { + "bbox": [ + 45, + 223, + 287, + 378 + ], + "type": "text", + "content": "Owning over a half global population and contributing the most economic growth, the city areas have been recorded and characterized by various data sources including satellite and aerial imagery, street-level imagery, LiDAR data, public maps, crowd-sourced data, etc. A great number of benchmarks have been proposed towards facilitating different vision tasks in city scene, of which the street-level imagery has been broadly used in multiple driving-related benchmarks [4, 10, 14, 18, 26, 31, 43]. The rich visual information of street-level imagery also enables complicated visual recognition tasks on specific categories, such as person detection [3], vehicle tracking and re-identification [32], and fine-grained land use classification [49]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 379, + 287, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 379, + 287, + 558 + ], + "spans": [ + { + "bbox": [ + 45, + 379, + 287, + 558 + ], + "type": "text", + "content": "Nevertheless, constructing pixel-wise annotations for street-level imagery requires substantial human efforts, resulting in the small image quantity and limited view types of existing datasets, especially for the street-view panorama datasets with only hundreds of annotated images [36, 41, 42]. Regarding the annotation categories and levels, existing datasets mostly provide instance-level annotations for dynamic object categories in driving scenes. As a vital component for city understanding, the static objects such as buildings and roads take up a larger proportion of cities and remain a high consistency across the satellite and ground-level images. However, existing street-level datasets either provide pixel-wise building annotations without fine-grained semantic labels [4, 10, 14, 47] or provide fine-grained annotations at only bbox or image level [46, 49]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 559, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 559, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 559, + 288, + 714 + ], + "type": "text", + "content": "Compared with street-level images, remote sensing images usually contain less visual information for conducting complicated tasks such as fine-grained land use segmentation and building function recognition. On the other hand, unlike the sparsely-distributed street-level images, remote sensing images have a dense spatial distribution and a worldwide coverage, which are well aligned with the open maps and government datasets at pixel level [22]. These existing maps and datasets contain a variety of satellite-level annotations for buildings (such as the footprint, land use, height, year built), roads (category and line coordinates), and other geographical objects, providing new perspectives for promoting novel city understanding datasets and tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 335 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 335 + ], + "type": "text", + "content": "In this work, as illustrated in Figure 1, we construct an omnipotent city dataset unifying data sources from both satellite and street views, linked by geo-locations and urban planning data. Unlike existing city datasets that only support a limited number of tasks, OmniCity dataset incorporates rich geometric annotations and semantic meta data for each image, where multiple tasks can be conducted on. To leverage the existing map labels and the rich visual context from the street-level imagery, we propose an efficient pipeline for producing diverse street-level annotations. Based on this annotation pipeline, we built OmniCity, a dataset that contains over 100K annotated images collected from 25K geo-locations in New York City. We provide benchmark results on OmniCity for a variety of tasks, including building footprint extraction and height estimation on satellite images, as well as fine-grained/instance/plane segmentation of buildings on street-level panorama and mono-view images. To the best of our knowledge, this is the first work that involves fine-grained building instance segmentation on street-level panorama images. We also analyze the potential of OmniCity for promoting new tasks and methods with multi-level imagery." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 317, + 335, + 525, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 335, + 525, + 346 + ], + "spans": [ + { + "bbox": [ + 317, + 335, + 525, + 346 + ], + "type": "text", + "content": "Our main contributions are summarized as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 318, + 357, + 545, + 535 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 318, + 357, + 545, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 357, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 318, + 357, + 545, + 392 + ], + "type": "text", + "content": "- We propose a novel pipeline for efficiently producing diverse pixel-wise annotations on street-level panorama and mono-view images." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 318, + 400, + 545, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 400, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 318, + 400, + 545, + 449 + ], + "type": "text", + "content": "- We build the OmniCity dataset, which contains well-aligned satellite and street-level images with a larger quantity, richer annotations and more views compared with existing datasets." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 318, + 456, + 545, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 456, + 545, + 492 + ], + "spans": [ + { + "bbox": [ + 318, + 456, + 545, + 492 + ], + "type": "text", + "content": "- We provide a series of benchmark experimental results for multiple tasks and data sources, and analyze the limitations of the current benchmarks on OmniCity." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 318, + 499, + 545, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 499, + 545, + 535 + ], + "spans": [ + { + "bbox": [ + 318, + 499, + 545, + 535 + ], + "type": "text", + "content": "- We discuss the potential of OmniCity for facilitating new methods and tasks for large-scale city understanding, reconstruction, and simulation." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 556, + 389, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 556, + 389, + 567 + ], + "spans": [ + { + "bbox": [ + 306, + 556, + 389, + 567 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 575, + 527, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 575, + 527, + 586 + ], + "spans": [ + { + "bbox": [ + 306, + 575, + 527, + 586 + ], + "type": "text", + "content": "2.1. Datasets and methods for street-level tasks" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "As shown in Table 1, many street-level datasets have been proposed over the past few years. A large proportion of these datasets are designed for visual tasks in driving scene [4, 10, 14, 18, 26, 31, 44], such as 2D/3D object detection, semantic segmentation, object tracking, etc. Several street-level datasets are proposed for a specific object category, such as the EuroCity Persons dataset [3], the CityFlow dataset [32] for vehicle tracking and re-identification, etc. In addition to the above datasets containing only monoview images, some studies propose new datasets or methods" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17398" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 154, + 282, + 288 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 288, + 148 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 288, + 148 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 288, + 148 + ], + "type": "text", + "content": "Table 1. A comparison of OmniCity with existing city-related datasets of street-level, satellite-level and cross-level. The Street and Satellite (Sate.) columns show the available image view types. The last three columns indicate which level of tasks the dataset is designed for (semantic/instance/plane segmentation, object detection (bbox), and image classification), and whether the dataset contains fine-grained building attribute (Attri.) or height labels." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 154, + 282, + 288 + ], + "lines": [ + { + "bbox": [ + 48, + 154, + 282, + 288 + ], + "spans": [ + { + "bbox": [ + 48, + 154, + 282, + 288 + ], + "type": "table", + "html": "
Dataset#ImagesStreetSate.Anno.Attri.Height
KITTI [14]15,000mono-semantic××
Cityscapes [10]25,000mono-semantic××
EuroCity [3]47,300mono-bbox××
WildPASS [42]500multi.-semantic××
PASS [41]400multi.-semantic××
HoliCity [47]6,300multi.-inst./plane××
SkyScapes [1]8,820-singlesemantic××
SpaceNet [38]60,000-multi.instance××
Christie et al. [9]11,000-singlesemantic×✓
Li et al. [21]3,300-singleinstance×✓
TorontoCity [36]Unknowmulti.multi.instance×✓
Wojna et al. [39]49,426monosingleimage✓×
OmniCity108,600multi.multi.inst./plane✓✓
", + "image_path": "9b15a05538a01d3533ba0aff045869826ae7cbf63b91e3356b9327b02169fda8.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 299, + 287, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 299, + 287, + 466 + ], + "spans": [ + { + "bbox": [ + 46, + 299, + 287, + 466 + ], + "type": "text", + "content": "for semantic segmentation from panorama images, such as TorontoCity [36], PASS [41], and WildPASS [42]. However, such datasets require expensive annotation efforts and contains only hundreds of pixel-wise annotated panorama images in total. Besides the driving-related datasets, several recent studies target at the recognition tasks for static object categories. Zhu et al. [49] proposed a framework for fine-grained land use classification task using ground-level images. Zhang et al. [46] proposed a building recognition system for detecting their business entity information. HoliCity [47] proposed holistic 3D structure annotations generated from CAD models for panorama images, but it only conducted experiments on mono-view images and lacked the semantic information of buildings and roads." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 467, + 288, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 467, + 288, + 659 + ], + "spans": [ + { + "bbox": [ + 46, + 467, + 288, + 659 + ], + "type": "text", + "content": "In summary, existing street-level datasets still have the following limitations. Regarding the image quantity and view types, most existing datasets require substantial annotation efforts and contain only a limited number of annotated images of a mono view. The image quantity of panorama datasets is even several orders of magnitude smaller than mono-view datasets. Regarding the annotation categories and levels, existing datasets mainly focus on dynamic or driving-related categories and lack in fine-grained annotations of an object category. Moreover, the datasets for static objects are lacking in pixel-level or fine-grained annotations. By contrast, our OmniCity contains over 100K satellite and street-level images of multiple views as well as the pixel-wise and fine-grained annotations, demonstrating superiority in annotation quantity, annotation type, and view type compared with existing datasets." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 670, + 279, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 279, + 682 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 279, + 682 + ], + "type": "text", + "content": "2.2. Datasets and methods for satellite-level tasks" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "text", + "content": "As a data source with a long time series and a large coverage, the satellite imagery has been broadly explored" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 287 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 287 + ], + "type": "text", + "content": "for large-scale city understanding. Unlike the street-level datasets requiring manual annotations, the satellite imagery is already well aligned with existing label maps [20]. The OpenStreetMap (OSM) is one of the most broadly-used map data, which contains publicly available annotations of building footprints, heights, roads, land use, etc., of world wide. In addition, many public datasets provide rich information at a local scale. For example, the PLUTO dataset contains the block and lot information of the whole New York City, and each lot is associated with the land use, year built, number of floors, and other useful information. The Microsoft US building footprint dataset contains over a hundred million computer-generated building footprints. Several challenges provide manually labeled building footprints and the corresponding satellite images [11, 34, 38]. Moreover, some datasets provide fine-grained semantic categories of different objects [1] or building height annotations for 3D reconstruction tasks [9, 21]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 288, + 545, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 288, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 304, + 288, + 545, + 396 + ], + "type": "text", + "content": "In summary, existing maps and datasets have provided substantial semantic and geometric information that is well aligned with the satellite imagery. In this work, we leverage these rich annotations and the transformation relations between different views to produce auxiliary information for street-level image annotation. Compared with existing work, OmniCity significantly reduces the human labeling efforts and provides more annotation types to enable omnipotent city understanding via multiple tasks and views." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 403, + 526, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 403, + 526, + 415 + ], + "spans": [ + { + "bbox": [ + 305, + 403, + 526, + 415 + ], + "type": "text", + "content": "2.3. Datasets and methods for multi-level tasks" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 422, + 545, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 422, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 304, + 422, + 545, + 601 + ], + "type": "text", + "content": "The ground-level images usually contain rich visual context that is not visible from the satellite or aerial imagery (e.g. building facade, the side of vegetations, etc.), while the spatial distribution is often sparse and unbalanced in different areas. By contrast, the remote sensing images have a much denser spatial distribution at a global scale, but the visual context is too limited to support complicated fine-grained tasks. Considering the above complementary characteristics, many datasets and methods have been proposed for cross-view scenarios. Cross-view visual recognition tasks integrate the two data types as the model input for predicting building functions, building ages, land use, tree species, etc. [2, 12, 37, 40], while cross-view image matching [17, 28, 33, 48] and image synthesis [25, 27, 30, 45] tasks take only one data type as the model input." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 601, + 545, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 601, + 545, + 686 + ], + "spans": [ + { + "bbox": [ + 304, + 601, + 545, + 686 + ], + "type": "text", + "content": "In addition, several datasets are proposed for a variety of satellite and street-level tasks. The TorontoCity [36] contains a wide range of annotations including building height estimation, building instance segmentation, building footprint segmentation, road segmentation, etc., which are conducted on either satellite or street-level images and only adopt FCN and some CNN classification architectures as" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 306, + 693, + 545, + 715 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 693, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 306, + 693, + 545, + 715 + ], + "type": "text", + "content": "1https://data.cityofnewyork.us/City-Government/Primary-Land-Use-Tax-Lot-Output-PLUTO-/64uk-42ks." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "17399" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 71, + 547, + 188 + ], + "blocks": [ + { + "bbox": [ + 49, + 71, + 547, + 188 + ], + "lines": [ + { + "bbox": [ + 49, + 71, + 547, + 188 + ], + "spans": [ + { + "bbox": [ + 49, + 71, + 547, + 188 + ], + "type": "image", + "image_path": "7b23f45f7c52680205d1d6a47ce7f9a8b80cd7930a169cf4ef77f6e75d676507.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 195, + 546, + 218 + ], + "lines": [ + { + "bbox": [ + 46, + 195, + 546, + 218 + ], + "spans": [ + { + "bbox": [ + 46, + 195, + 546, + 218 + ], + "type": "text", + "content": "Figure 2. The annotation tool of OmniCity dataset. The left and right images show examples of one road scene with the label map (in the bottom-right window) displayed in land use mode, and crossroads scene with the label map in block-lot mode, respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 230, + 289, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 230, + 289, + 339 + ], + "spans": [ + { + "bbox": [ + 46, + 230, + 289, + 339 + ], + "type": "text", + "content": "the baseline models for all experiments. In a recent study [39], a holistic multi-view building analysis dataset is designed for multiple recognition tasks regarding facade material, number of floors, occupancy type, roof geometry, roof pitch, and construction type. The dataset contains only image-wise annotations for street-level classification tasks. The semantic categories of each building are decided by human annotators, resulting in more challenges and subjectivity to the annotation process compared with our work." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 347, + 289, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 347, + 289, + 504 + ], + "spans": [ + { + "bbox": [ + 46, + 347, + 289, + 504 + ], + "type": "text", + "content": "In summary, for the existing cross-view studies designed for one specific task, the street-level images are not annotated and only supplement extra features to the satellite-level feature maps. By contrast, our OmniCity provides rich annotation types that are not contained in existing datasets, e.g. the pixel-wise building instances and fine-grained categories on street-level images, which can promote new methods to explore and leverage these annotations to improve the performance. In addition, OmniCity provides the benchmark experimental results of each task using more state-of-the-art models. With our efficient annotation pipeline, it also provides additional fine-grained pixel-wise annotations and benchmark results compared with [36] and [39]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 540, + 105, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 540, + 105, + 553 + ], + "spans": [ + { + "bbox": [ + 47, + 540, + 105, + 553 + ], + "type": "text", + "content": "3. Datasets" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 570, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 289, + 713 + ], + "type": "text", + "content": "In this work, we aim at building a dataset for omnipotent city understanding from multi-level and multi-view images. Our proposed OmniCity dataset contains 108,600 images of multiple views, which are collected from 25K geo-locations in New York City. Compared with existing datasets, OmniCity requires much fewer human efforts for street-level image annotation, contains more diverse annotation types for both 2D and 3D tasks, provides richer building semantics at instance segmentation level, and possesses higher scalability for new annotation type supplement and expansion to other cities. The details of data collection, annotation, and statistics are introduced as follows." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 230, + 398, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 230, + 398, + 243 + ], + "spans": [ + { + "bbox": [ + 306, + 230, + 398, + 243 + ], + "type": "text", + "content": "3.1. Data collection" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 250, + 547, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 250, + 547, + 586 + ], + "spans": [ + { + "bbox": [ + 304, + 250, + 547, + 586 + ], + "type": "text", + "content": "As shown in Figure 1, our OmniCity dataset is collected from six selected regions of New York. We download the panorama images in the six selected regions using google street view download 360, with a step distance of 65 meters. The regions for collecting the training and test samples are denoted by red and blue, respectively. We save the geographic coordinates, collection time, panorama id, north rotation, and zoom level for each panorama image. For each panorama site, we collect its corresponding google earth images of three acquisition dates according to the geographic coordinates, constituting three groups of satellite-level datasets with three cases of off-nadir view angles (small/medium/high, denoted by V1/V2/V3). For the annotation data sources, we collect the meta information from PLUTO and OpenStreetMap (OSM). The New York City is hierarchically formed by blocks, lots, and buildings. Each building can be identified by a specific block-lot id. In PLUTO, each lot (building) is associated with rich information, e.g. land use, year built, number of floors, etc. The OSM data contains footprint and height information while lacking in land use information for most buildings. Considering the characteristics of the two data sources, we align the land use attribute (from PLUTO) with the building footprint and height (from OSM) using the geographical coordinates. Overall, each building is assigned with a block-lot id, a land use category, a height value, and the geographical coordinates of a footprint polygon, which will be used as the reference label maps for panorama image annotation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 597, + 404, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 597, + 404, + 609 + ], + "spans": [ + { + "bbox": [ + 306, + 597, + 404, + 609 + ], + "type": "text", + "content": "3.2. Data annotation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 618, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 547, + 715 + ], + "type": "text", + "content": "Figure 2 shows the panorama image annotation tool proposed in our study. The street-level panorama and satellite images are naturally well-aligned at image-level between the central coordinates from satellite views and the camera pose of panorama views. The instance-level annotations of different views (e.g., building facade and footprint) are further aligned during the annotation process. According to the geo-transformation relation between different views [25]," + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17400" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 156 + ], + "type": "text", + "content": "we calculate the angle between the north (upward) direction and each building footprint (in the satellite image), and the angle between the north direction (saved in the panorama file) and each building facade in the horizontal axis. The footprint and facade with the same angle value are aligned to the same building, providing the auxiliary information for the manual annotation process." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 157, + 287, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 157, + 287, + 289 + ], + "spans": [ + { + "bbox": [ + 46, + 157, + 287, + 289 + ], + "type": "text", + "content": "The annotation pipeline includes four stages: (1) Image selection, i.e., select the panorama images that are essential to be annotated according to building coverage, occlusion extent, etc; (2) Segmentation annotation, i.e., adjust the floor/top line to fit the bottom/roof of each building, and add the boundary split line considering both auxiliary information and building appearance (e.g. texture discrepancy, doors, etc.); (3) Attribute assignment, i.e., add the attributes (instance ID, block-lot id and land use type) for each building plane; (4) Quality assessment, i.e., check the annotation quality and remove the unqualified images." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 289, + 288, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 289, + 288, + 469 + ], + "spans": [ + { + "bbox": [ + 46, + 289, + 288, + 469 + ], + "type": "text", + "content": "In addition, the mono-view images and annotations are automatically generated from those of the panorama images via view transformation. For each panorama image, we select three views using three x-axis angles (-170, 10 and 170, in the range of [-180, 180]) and a fixed y-axis angle of 0. Then we design an image selection rule to filter out the unexpected images of which the buildings are distributed in only one side, building area proportion is smaller than " + }, + { + "bbox": [ + 46, + 289, + 288, + 469 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 289, + 288, + 469 + ], + "type": "text", + "content": ", or the building plane quantity is smaller than 2. The remaining images constitute the final mono-view dataset of our experiments. For satellite-level tasks, the annotations are already well-aligned with the meta information from PLUTO and OSM. Each building footprint on a satellite image is assigned with a block-lot id, a land use category, a height value, and the pixel coordinates of the footprint polygon." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 479, + 223, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 479, + 223, + 491 + ], + "spans": [ + { + "bbox": [ + 47, + 479, + 223, + 491 + ], + "type": "text", + "content": "3.3. Statistics of the proposed dataset" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 498, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 288, + 714 + ], + "type": "text", + "content": "The whole dataset contains three sub-datasets (in small, medium and large view angles) for satellite-level tasks and two sub-datasets (in panorama and mono views) for street-level tasks. For satellite-level datasets, the three sub-datasets of images collected from the 25K geo-locations constitute 75K images in total, which are cropped by " + }, + { + "bbox": [ + 46, + 498, + 288, + 714 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 46, + 498, + 288, + 714 + ], + "type": "text", + "content": " pixels according to the coverage area of the corresponding panorama image. For the street-level datasets, the panorama dataset contains 18K images in " + }, + { + "bbox": [ + 46, + 498, + 288, + 714 + ], + "type": "inline_equation", + "content": "512 \\times 1024" + }, + { + "bbox": [ + 46, + 498, + 288, + 714 + ], + "type": "text", + "content": " pixels, which are selected from the initial 25K panorama images during the image selection and quality assessment phases. Similarly, the mono-view dataset obtained via view transformation and image selection contains 156,00 monoview images in " + }, + { + "bbox": [ + 46, + 498, + 288, + 714 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 46, + 498, + 288, + 714 + ], + "type": "text", + "content": " pixels. The ratio of train/test splits is set as 4:1 for all five sub-datasets. In OmniCity dataset, the initial land use categories of PLUTO with similar characteristics and low quantities are merged into one category, resulting in 7 land use categories in total. The" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": "categories of " + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "1/2" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": " family building (" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "\\sim 17\\%" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": "), walk-up building (" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "\\sim 23\\%" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": ") and the mixed residential/commercial building categories (" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "\\sim 29\\%" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": ") take up a larger proportion compared with the other four categories, i.e., elevator buildings (" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "\\sim 7\\%" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": "), office buildings (" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "\\sim 10\\%" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": "), industrial/transportation/utility buildings (" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "\\sim 11\\%" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": ") and others (" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "\\sim 1\\%" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": "). For the distribution of building height, most are between 1 and 25 meters while a small percentage (" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "inline_equation", + "content": "\\sim 1\\%" + }, + { + "bbox": [ + 304, + 72, + 545, + 168 + ], + "type": "text", + "content": ") reaches over 50 meters." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 182, + 418, + 194 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 182, + 418, + 194 + ], + "spans": [ + { + "bbox": [ + 305, + 182, + 418, + 194 + ], + "type": "text", + "content": "4. Benchmark results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 204, + 545, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 204, + 545, + 361 + ], + "spans": [ + { + "bbox": [ + 304, + 204, + 545, + 361 + ], + "type": "text", + "content": "In this section, we provide a variety of benchmarks for multiple satellite and street-level tasks. The satellite-level tasks in our experiments include building footprint segmentation and height estimation. For both tasks, we conduct experiments on the satellite images with three view angles. For the street-level tasks, we conduct two instance segmentation tasks (i.e., land use and building instance segmentation) on the panorama images, and three instance segmentation tasks (i.e., land use / building instance / plane segmentation) on mono-view images. Please note that these are only preliminary experimental results on OmniCity dataset. More benchmarks of latest models and additional tasks will be continuously updated on OmniCity homepage." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 371, + 427, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 371, + 427, + 384 + ], + "spans": [ + { + "bbox": [ + 305, + 371, + 427, + 384 + ], + "type": "text", + "content": "4.1. Experimental setting" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 391, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 391, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 391, + 545, + 713 + ], + "type": "text", + "content": "The experiments are mainly based on mmdetection [7] with the recommended hyper-parameter settings. We select Mask R-CNN [15] as the baseline method for segmentation tasks, and provide a comparison of different methods in Table 6, including Mask Scoring R-CNN (MS R-CNN) [19], Cascade Mask R-CNN (Cascade) [5], Content-Aware ReAssembly of FEatures (CARAFE) [35], and Hybrid Task Cascade (HTC) [6]. Specifically, we use ResNet-50 [16] with FPN [23] pre-trained on the ImageNet [29] as the backbone for all instance segmentation models. All models are trained on 8 NVIDIA Tesla V100 GPUs for 12 epochs, with a batch size of 16, a learning rate starting from 0.02 and decreasing by a factor of 0.1 from the " + }, + { + "bbox": [ + 304, + 391, + 545, + 713 + ], + "type": "inline_equation", + "content": "8^{th}" + }, + { + "bbox": [ + 304, + 391, + 545, + 713 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 391, + 545, + 713 + ], + "type": "inline_equation", + "content": "11^{th}" + }, + { + "bbox": [ + 304, + 391, + 545, + 713 + ], + "type": "text", + "content": " epoch, and the stochastic gradient descent (SGD) optimizer with a weight decay of " + }, + { + "bbox": [ + 304, + 391, + 545, + 713 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 304, + 391, + 545, + 713 + ], + "type": "text", + "content": " and a momentum of 0.9. For the height estimation task, we evaluate the performance of two widely-used monocular depth estimation methods on the satellite images of three view angles, i.e. Structure-Aware Residual Pyramid Network (SARPN) [8] and Deep Ordinal Regression Network (DORN) [13], which are trained on 4 NVIDIA Tesla V100 GPUs for 20 epochs. SARPN is trained with a batch size of 8, a learning rate starting from " + }, + { + "bbox": [ + 304, + 391, + 545, + 713 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 304, + 391, + 545, + 713 + ], + "type": "text", + "content": " and reduced by " + }, + { + "bbox": [ + 304, + 391, + 545, + 713 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 304, + 391, + 545, + 713 + ], + "type": "text", + "content": " every 5 epochs, and the Adam optimizer with " + }, + { + "bbox": [ + 304, + 391, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\beta_{1} = 0.9" + }, + { + "bbox": [ + 304, + 391, + 545, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 391, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\beta_{2} = 0.999" + }, + { + "bbox": [ + 304, + 391, + 545, + 713 + ], + "type": "text", + "content": ", and a weight decay of " + }, + { + "bbox": [ + 304, + 391, + 545, + 713 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 304, + 391, + 545, + 713 + ], + "type": "text", + "content": ". DORN is trained with a batch size of 4, a base learning rate of " + }, + { + "bbox": [ + 304, + 391, + 545, + 713 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 304, + 391, + 545, + 713 + ], + "type": "text", + "content": " and the power of 0.9, using SGD optimizer with a weight decay of 0.0005 and a momentum of 0.9." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17401" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 70, + 116, + 137 + ], + "blocks": [ + { + "bbox": [ + 49, + 70, + 116, + 137 + ], + "lines": [ + { + "bbox": [ + 49, + 70, + 116, + 137 + ], + "spans": [ + { + "bbox": [ + 49, + 70, + 116, + 137 + ], + "type": "image", + "image_path": "e9d32052e9ca393c4f493081008bbb5ac67b6e756b5bc6a5058f74d9a2e20232.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 63, + 137, + 102, + 146 + ], + "lines": [ + { + "bbox": [ + 63, + 137, + 102, + 146 + ], + "spans": [ + { + "bbox": [ + 63, + 137, + 102, + 146 + ], + "type": "text", + "content": "Small view" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 120, + 70, + 187, + 137 + ], + "blocks": [ + { + "bbox": [ + 120, + 70, + 187, + 137 + ], + "lines": [ + { + "bbox": [ + 120, + 70, + 187, + 137 + ], + "spans": [ + { + "bbox": [ + 120, + 70, + 187, + 137 + ], + "type": "image", + "image_path": "2f658c1f3183d8110deec368391d8b8cac7c97b5d6d1b55270d41dc716fc88ee.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 129, + 137, + 178, + 146 + ], + "lines": [ + { + "bbox": [ + 129, + 137, + 178, + 146 + ], + "spans": [ + { + "bbox": [ + 129, + 137, + 178, + 146 + ], + "type": "text", + "content": "Medium view" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 190, + 70, + 258, + 137 + ], + "blocks": [ + { + "bbox": [ + 190, + 70, + 258, + 137 + ], + "lines": [ + { + "bbox": [ + 190, + 70, + 258, + 137 + ], + "spans": [ + { + "bbox": [ + 190, + 70, + 258, + 137 + ], + "type": "image", + "image_path": "46baf51ab398638283631313bf8346e7752c3e33802b373f441d13fd9c4230b1.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 203, + 137, + 243, + 147 + ], + "lines": [ + { + "bbox": [ + 203, + 137, + 243, + 147 + ], + "spans": [ + { + "bbox": [ + 203, + 137, + 243, + 147 + ], + "type": "text", + "content": "Large view" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 49, + 148, + 116, + 216 + ], + "blocks": [ + { + "bbox": [ + 49, + 148, + 116, + 216 + ], + "lines": [ + { + "bbox": [ + 49, + 148, + 116, + 216 + ], + "spans": [ + { + "bbox": [ + 49, + 148, + 116, + 216 + ], + "type": "image", + "image_path": "0659340c6af97b34b589185c5a4a9a6896bb1bc4b9db2bb2f6325ca6dc12697d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 57, + 217, + 110, + 227 + ], + "lines": [ + { + "bbox": [ + 57, + 217, + 110, + 227 + ], + "spans": [ + { + "bbox": [ + 57, + 217, + 110, + 227 + ], + "type": "text", + "content": "Satellite image" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 120, + 148, + 187, + 216 + ], + "blocks": [ + { + "bbox": [ + 120, + 148, + 187, + 216 + ], + "lines": [ + { + "bbox": [ + 120, + 148, + 187, + 216 + ], + "spans": [ + { + "bbox": [ + 120, + 148, + 187, + 216 + ], + "type": "image", + "image_path": "73cc5a7d27527e98d435fac9e135f6d1de42946dcf42884ce9b62cc49c3992a7.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 129, + 217, + 178, + 226 + ], + "lines": [ + { + "bbox": [ + 129, + 217, + 178, + 226 + ], + "spans": [ + { + "bbox": [ + 129, + 217, + 178, + 226 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 190, + 148, + 258, + 216 + ], + "blocks": [ + { + "bbox": [ + 190, + 148, + 258, + 216 + ], + "lines": [ + { + "bbox": [ + 190, + 148, + 258, + 216 + ], + "spans": [ + { + "bbox": [ + 190, + 148, + 258, + 216 + ], + "type": "image", + "image_path": "93db0a7f5a73a502623284e9df92ef96db26b3ff6f289dcc57d285f7cbabad66.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 200, + 217, + 250, + 226 + ], + "lines": [ + { + "bbox": [ + 200, + 217, + 250, + 226 + ], + "spans": [ + { + "bbox": [ + 200, + 217, + 250, + 226 + ], + "type": "text", + "content": "DORN - small" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 48, + 228, + 116, + 297 + ], + "blocks": [ + { + "bbox": [ + 48, + 228, + 116, + 297 + ], + "lines": [ + { + "bbox": [ + 48, + 228, + 116, + 297 + ], + "spans": [ + { + "bbox": [ + 48, + 228, + 116, + 297 + ], + "type": "image", + "image_path": "60ec9f7b70b57524dd1a16989d1f1a0e074a4793a2e388d10606610a238e9eed.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 56, + 297, + 110, + 307 + ], + "lines": [ + { + "bbox": [ + 56, + 297, + 110, + 307 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 110, + 307 + ], + "type": "text", + "content": "SARPN - small" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 322, + 287, + 345 + ], + "lines": [ + { + "bbox": [ + 46, + 322, + 287, + 345 + ], + "spans": [ + { + "bbox": [ + 46, + 322, + 287, + 345 + ], + "type": "text", + "content": "Figure 3. Qualitative results of satellite-level tasks. The yellow/cyan/red polygons denote TP/FP/FN buildings." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 119, + 228, + 187, + 297 + ], + "blocks": [ + { + "bbox": [ + 119, + 228, + 187, + 297 + ], + "lines": [ + { + "bbox": [ + 119, + 228, + 187, + 297 + ], + "spans": [ + { + "bbox": [ + 119, + 228, + 187, + 297 + ], + "type": "image", + "image_path": "3eea1ca0371f98394bb37b58009e6f89d3aa955a02c71452177579b2ecf06f06.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 122, + 297, + 186, + 307 + ], + "lines": [ + { + "bbox": [ + 122, + 297, + 186, + 307 + ], + "spans": [ + { + "bbox": [ + 122, + 297, + 186, + 307 + ], + "type": "text", + "content": "SARPN - medium" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 190, + 228, + 258, + 297 + ], + "blocks": [ + { + "bbox": [ + 190, + 228, + 258, + 297 + ], + "lines": [ + { + "bbox": [ + 190, + 228, + 258, + 297 + ], + "spans": [ + { + "bbox": [ + 190, + 228, + 258, + 297 + ], + "type": "image", + "image_path": "8dc71005b97332c4f13564bb4f8e5f1faf3240499b3806775d41e2261bb93cb6.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 196, + 297, + 251, + 308 + ], + "lines": [ + { + "bbox": [ + 196, + 297, + 251, + 308 + ], + "spans": [ + { + "bbox": [ + 196, + 297, + 251, + 308 + ], + "type": "text", + "content": "SARPN - large" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 261, + 230, + 279, + 297 + ], + "blocks": [ + { + "bbox": [ + 262, + 219, + 280, + 228 + ], + "lines": [ + { + "bbox": [ + 262, + 219, + 280, + 228 + ], + "spans": [ + { + "bbox": [ + 262, + 219, + 280, + 228 + ], + "type": "text", + "content": "High" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 261, + 230, + 279, + 297 + ], + "lines": [ + { + "bbox": [ + 261, + 230, + 279, + 297 + ], + "spans": [ + { + "bbox": [ + 261, + 230, + 279, + 297 + ], + "type": "image", + "image_path": "4ea60c47fe3bc4e1b3b546a2a26a9c8f261764e2ab0b228c57fbd8e5b4fbfde1.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 263, + 298, + 279, + 307 + ], + "lines": [ + { + "bbox": [ + 263, + 298, + 279, + 307 + ], + "spans": [ + { + "bbox": [ + 263, + 298, + 279, + 307 + ], + "type": "text", + "content": "Low" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 47, + 357, + 269, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 357, + 269, + 369 + ], + "spans": [ + { + "bbox": [ + 47, + 357, + 269, + 369 + ], + "type": "text", + "content": "4.2. Experimental results of satellite-level tasks" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 47, + 378, + 289, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 378, + 289, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 378, + 289, + 713 + ], + "type": "text", + "content": "Table 2 shows the baseline results of the satellite-level instance segmentation task, which are evaluated using both COCO [24] and SpaceNet [11] evaluation metrics. The footprint segmentation performance is the best for satellite images with a small view angle, and deteriorates seriously when the view angle gets larger. For all three cases, the prediction score is the highest for buildings with large areas and the lowest for small buildings. Figure 3 provides a qualitative comparison of the footprint segmentation results on satellite images of three types of view angle. The large view angle results in great difficulties for extracting accurate footprint boundaries, due to the partial invisibility of building footprint, the serious shadow effects, etc. The height estimation performance is evaluated in terms of the mean absolute error (denoted by MAE), mean square error (denoted by MSE), and root mean square error (denoted by RMSE), which are commonly used metrics for depth estimation. All metrics are measured in meters at pixel level. Table 3 and Figure 3 show the height estimation results obtained from DORN [13] and SARPN [8] for satellite images with different view angles. Results demonstrate that DORN obtains better results compared with SARPN for all three cases. Both methods achieve the best performance for satellite images with a medium view angle (V2) compared with the other two cases. The footprint and roof have more overlaps and provide less building structure information for satellite images with a small view angle, while the shadow and parallax effect become more serious with" + } + ] + } + ], + "index": 23 + }, + { + "type": "table", + "bbox": [ + 307, + 102, + 545, + 148 + ], + "blocks": [ + { + "bbox": [ + 306, + 71, + 545, + 93 + ], + "lines": [ + { + "bbox": [ + 306, + 71, + 545, + 93 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 545, + 93 + ], + "type": "text", + "content": "Table 2. Quantitative results of instance segmentation for satellite images with different view angles." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 102, + 545, + 148 + ], + "lines": [ + { + "bbox": [ + 307, + 102, + 545, + 148 + ], + "spans": [ + { + "bbox": [ + 307, + 102, + 545, + 148 + ], + "type": "table", + "html": "
ViewMetrics of various thresholdsthreshold = 0.5
APAP50AP75APSAPMAPLPRF1
V129.766.023.515.933.936.776.966.371.2
V223.756.616.111.527.230.373.955.063.1
V318.951.49.69.121.525.370.751.759.7
", + "image_path": "2ed9af556c83367067b0d234c6d8ea3bcce0d06f4255dda2ce0c878db2044b87.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "table_body" + } + ], + "index": 25 + }, + { + "type": "table", + "bbox": [ + 318, + 186, + 533, + 238 + ], + "blocks": [ + { + "bbox": [ + 306, + 155, + 545, + 177 + ], + "lines": [ + { + "bbox": [ + 306, + 155, + 545, + 177 + ], + "spans": [ + { + "bbox": [ + 306, + 155, + 545, + 177 + ], + "type": "text", + "content": "Table 3. Quantitative results of height estimation for satellite images with different view angles." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 186, + 533, + 238 + ], + "lines": [ + { + "bbox": [ + 318, + 186, + 533, + 238 + ], + "spans": [ + { + "bbox": [ + 318, + 186, + 533, + 238 + ], + "type": "table", + "html": "
ViewSARPN [8]DORN [13]
MAEMSERMSEMAEMSERMSE
V116.18870.3429.5012.71670.5225.89
V213.75694.1726.3512.24628.0625.06
V315.32823.0128.6913.40730.6727.03
", + "image_path": "dc34dc59d2ea834c52840843e2af362b5db8344523e709d10dc904ba8e1f0573.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "table_body" + } + ], + "index": 27 + }, + { + "bbox": [ + 305, + 251, + 545, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 251, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 305, + 251, + 545, + 289 + ], + "type": "text", + "content": "the increase of off-nadir view angle. The above aspects result in challenges for the accurate estimation of building height for satellite images with small and large view angles." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 306, + 298, + 518, + 311 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 298, + 518, + 311 + ], + "spans": [ + { + "bbox": [ + 306, + 298, + 518, + 311 + ], + "type": "text", + "content": "4.3. Experimental results of street-level tasks" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 304, + 317, + 545, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 317, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 317, + 545, + 521 + ], + "type": "text", + "content": "We analyze the performance of multiple segmentation tasks on street-level panorama and mono-view images. Table 4 and Figure 4 show the experimental results of panorama-view images on two segmentation tasks, of which the performance on building instance segmentation task (denoted by Instance Seg.) is significantly superior to the fine-grained land use segmentation task (denoted by Landuse Seg.). Table 5 and Figure 5 show the experimental results of mono-view images on three different tasks, i.e., landuse segmentation, instance segmentation, and plane segmentation. Similar to the results of panorama-view images, the baseline method achieves much higher scores for the two binary segmentation tasks (plane and instance segmentation) compared with the fine-grained land use segmentation task. The qualitative results also demonstrate that the baseline method has difficulties in identifying the accurate land use type of some building instances." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 304, + 522, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 546, + 713 + ], + "type": "text", + "content": "Table 6 shows the land use segmentation results obtained from the five methods on street-level panorama images. HTC achieves the best performance for both overall and category metrics followed by MS R-CNN, which indicates that the cascade structure and mask scoring strategy can effectively improves the fine-grained segmentation performance of building instances. From the metrics of each category, we can find that C2, C4 and C5 (i.e., Walkup Buildings, Mixed Residential/Commercial, and Office Buildings) have better performance compared with C1, C3 and C6 (i.e., 1/2 Family Buildings, Elevator Buildings and Industrial/Transportation/Utility) for all methods, with the AP ranging from " + }, + { + "bbox": [ + 304, + 522, + 546, + 713 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 304, + 522, + 546, + 713 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 522, + 546, + 713 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 304, + 522, + 546, + 713 + ], + "type": "text", + "content": ". The category of Others has an extremely low AP score due to the small sample quantity. Compared with the residential buildings, the mixed Residential/Commercial and Office buildings have more special" + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17402" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 72, + 286, + 220 + ], + "blocks": [ + { + "bbox": [ + 50, + 72, + 286, + 220 + ], + "lines": [ + { + "bbox": [ + 50, + 72, + 286, + 220 + ], + "spans": [ + { + "bbox": [ + 50, + 72, + 286, + 220 + ], + "type": "image", + "image_path": "b541a0ef87e31a184621fd24b40c1f94531f444feeae28e00863e223bbc3ced9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 228, + 287, + 251 + ], + "lines": [ + { + "bbox": [ + 47, + 228, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 47, + 228, + 287, + 251 + ], + "type": "text", + "content": "Figure 4. Qualitative results of street-level tasks (i.e., instance segmentation and land use segmentation) on panorama images." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 48, + 282, + 287, + 323 + ], + "blocks": [ + { + "bbox": [ + 54, + 261, + 279, + 273 + ], + "lines": [ + { + "bbox": [ + 54, + 261, + 279, + 273 + ], + "spans": [ + { + "bbox": [ + 54, + 261, + 279, + 273 + ], + "type": "text", + "content": "Table 4. Quantitative results on street-level panorama images." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 282, + 287, + 323 + ], + "lines": [ + { + "bbox": [ + 48, + 282, + 287, + 323 + ], + "spans": [ + { + "bbox": [ + 48, + 282, + 287, + 323 + ], + "type": "table", + "html": "
TaskAPAP50AP75APSAPMAPL
Landuse Seg.26.034.728.50.312.030.4
Instance Seg.66.786.572.51.740.274.1
", + "image_path": "adcc651fbc976aba9cf00954caea7749ef10c377e87c9e7ef70a8600a57cc2ac.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 48, + 353, + 287, + 407 + ], + "blocks": [ + { + "bbox": [ + 52, + 332, + 282, + 343 + ], + "lines": [ + { + "bbox": [ + 52, + 332, + 282, + 343 + ], + "spans": [ + { + "bbox": [ + 52, + 332, + 282, + 343 + ], + "type": "text", + "content": "Table 5. Quantitative results on street-level mono-view images." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 353, + 287, + 407 + ], + "lines": [ + { + "bbox": [ + 48, + 353, + 287, + 407 + ], + "spans": [ + { + "bbox": [ + 48, + 353, + 287, + 407 + ], + "type": "table", + "html": "
TaskAPAP50AP75APSAPMAPL
Landuse Seg.23.932.126.70.310.627.5
Instance Seg.68.388.873.83.233.376.1
Plane Seg.65.187.471.05.040.773.8
", + "image_path": "cf39b3dd9273127ad26bf077c76d709ee9c73f716031d667b4b5c589e0179db5.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 424, + 287, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 424, + 287, + 483 + ], + "spans": [ + { + "bbox": [ + 46, + 424, + 287, + 483 + ], + "type": "text", + "content": "characteristics (i.e. the boundary between first and above floors, facade design, and building structure), contributing to the superior AP scores of these two categories. The performance of the above six categories is also concordant with the ratio of the sample quantity." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 491, + 219, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 491, + 219, + 504 + ], + "spans": [ + { + "bbox": [ + 47, + 491, + 219, + 504 + ], + "type": "text", + "content": "4.4. Results analysis and discussions" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 510, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 288, + 715 + ], + "type": "text", + "content": "In this section we summarize the limitations of existing methods for satellite and street-level tasks on our OmniCity dataset. For building footprint segmentation, the performance of existing methods get worse with the increasing of off-nadir view angle, which might due to the serious parallax and shadow effect on the satellite images with a large off-nadir view angle. For the height estimation task, most existing methods directly apply the monocular depth estimation methods to remote sensing scene. These methods usually produce poor results on the invisible side of the footprint boundary especially for high-rise buildings on very off-nadir images. In addition, most existing methods use deep neural networks to regress continuous values but the actual height values are discrete for building and non-building areas, resulting in difficulties for network training and extra efforts for converting the continuous prediction values into discrete height values via post-processing." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 310, + 72, + 545, + 266 + ], + "blocks": [ + { + "bbox": [ + 310, + 72, + 545, + 266 + ], + "lines": [ + { + "bbox": [ + 310, + 72, + 545, + 266 + ], + "spans": [ + { + "bbox": [ + 310, + 72, + 545, + 266 + ], + "type": "image", + "image_path": "3375a710fbf8843139a54b93e66df7291a245b83620c5ffd80638b67ef8a0f6a.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 277, + 545, + 302 + ], + "lines": [ + { + "bbox": [ + 305, + 277, + 545, + 302 + ], + "spans": [ + { + "bbox": [ + 305, + 277, + 545, + 302 + ], + "type": "text", + "content": "Figure 5. Qualitative results of street-level tasks (i.e., land use /instance/plane segmentation) on mono-view images." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 319, + 546, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 319, + 546, + 559 + ], + "spans": [ + { + "bbox": [ + 304, + 319, + 546, + 559 + ], + "type": "text", + "content": "For the street-level tasks based on panorama images, existing methods target at general instance segmentation tasks for commonly-used datasets, e.g. COCO [24], CityScapes [10], BDD100K [44], etc. These datasets often have a single view and a narrow Field of View (FoV). However, for panorama images, the special properties such as the wide FoV covering full 360-degree in the horizontal direction, are not taken into consideration in the design of existing methods. For both mono-view and panorama images, existing methods have difficulties in accurately recognizing the building instance with a small area (e.g. buildings located in the side of the main parts), the land use categories with a small number of building instances, and the categories that are easily confused (e.g. 1/2 Family Buildings and Walk-Up Buildings, Mixed Residential/Commercial and Office Buildings). Figure 6 shows some typical failure cases of the current benchmark methods. New instance segmentation methods should be designed for solving the above limitations considering the characteristics of panorama images, building instances, fine-grained categories, etc." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 573, + 493, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 573, + 493, + 586 + ], + "spans": [ + { + "bbox": [ + 306, + 573, + 493, + 586 + ], + "type": "text", + "content": "5. Potential of the OmniCity dataset" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 594, + 545, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 629 + ], + "type": "text", + "content": "Our proposed OmniCity dataset demonstrates great potential for facilitating city understanding, machine perception, and generative modeling researches in many aspects." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "First, it can serve as a new dataset for the existing tasks such as image geo-localization/synthesis and segmentation/detection of buildings/trees/land use from cross-view images. OmniCity provides additional annotations that are not contained in existing datasets, e.g. the building instances and fine-grained categories on street-level images, which can promote new methods to explore and leverage" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17403" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 109, + 93, + 482, + 174 + ], + "blocks": [ + { + "bbox": [ + 76, + 72, + 516, + 84 + ], + "lines": [ + { + "bbox": [ + 76, + 72, + 516, + 84 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 516, + 84 + ], + "type": "text", + "content": "Table 6. Quantitative results of different methods for fine-grained land use segmentation on street-level panorama images." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 109, + 93, + 482, + 174 + ], + "lines": [ + { + "bbox": [ + 109, + 93, + 482, + 174 + ], + "spans": [ + { + "bbox": [ + 109, + 93, + 482, + 174 + ], + "type": "table", + "html": "
MethodOverall MetricsMetrics of each category
APAP50AP75APSAPMAPLC1C2C3C4C5C6C7
Mask R-CNN [15]26.034.728.50.312.030.419.637.525.839.236.922.20.8
MS R-CNN [19]27.135.829.80.112.431.522.539.126.240.838.021.71.2
Cascade [5]25.933.828.30.211.430.52038.32538.536.722.10.3
CARAFE [35]25.934.528.50.111.930.219.637.324.939.937.221.50.8
HTC [6]27.235.729.90.312.432.020.838.727.239.938.424.51.2
", + "image_path": "2b84bd86f766e4bbb845c8646f38910e09187e3ce5b7536e01593f9f393d9be9.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 185, + 287, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 185, + 287, + 281 + ], + "spans": [ + { + "bbox": [ + 46, + 185, + 287, + 281 + ], + "type": "text", + "content": "the new annotations to improve the performance of these tasks. The annotations are organized in a unified version, which means multiple tasks can be performed on a single image, and thus can well support the multi-task learning setting. Additionally, since the building instances are directly linked with the urban planning data using block-lot id, it is easy to be enriched with more annotation types from other urban datasets, especially those for social and urban studies." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 281, + 287, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 281, + 287, + 400 + ], + "spans": [ + { + "bbox": [ + 46, + 281, + 287, + 400 + ], + "type": "text", + "content": "Second, OmniCity provides a new application scenario or problem setting to existing tasks. For line segment detection and wireframe parsing tasks, existing datasets contains densely distributed line segment and wireframe labels, while our OmniCity focuses on the main line segments and wireframes on the outlines instead of the inner ones, resulting in a much sparser format. The serious shelters from the trees and vehicles also bring challenges to these tasks. New line segment detection and wireframe parsing methods should be designed for the OmniCity scenario." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 401, + 287, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 401, + 287, + 498 + ], + "spans": [ + { + "bbox": [ + 46, + 401, + 287, + 498 + ], + "type": "text", + "content": "Moreover, OmniCity facilitates new tasks for city reconstruction and simulation. Treating each panorama as a unique city scene, a complete 3D model representing such a local scene is available, as also shown in the second window panel in our tool (Figure 2). These 3D models are stored in abstract vector formats, with clean and clear vertices indicating the building facades, streets, etc., which are suitable for many shape generation tasks represented with graphs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 498, + 287, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 287, + 629 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 287, + 629 + ], + "type": "text", + "content": "Finally, with the well-aligned satellite and street-level images as well as the various annotation types, novel city reconstruction tasks, e.g., 3D building reconstruction from cross-view images, can be derived for producing holistic 3D buildings with both fine-grained semantic category and precise geometry information (vector 3D model). Unlike the existing studies that only target at 3D reconstruction from monocular or multi-view remote sensing imagery, new methods should be designed to leverage the additional information from street-level images to improve 3D reconstruction and semantic prediction for OmniCity scenario." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 645, + 119, + 657 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 645, + 119, + 657 + ], + "spans": [ + { + "bbox": [ + 47, + 645, + 119, + 657 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 665, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 665, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 665, + 287, + 713 + ], + "type": "text", + "content": "In this paper, we have proposed OmniCity, a new dataset for omnipotent city understanding from satellite and street-level images of multiple views. The dataset contains over 100K images collected from 25K geo-locations in New" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 307, + 182, + 378, + 243 + ], + "blocks": [ + { + "bbox": [ + 307, + 182, + 378, + 243 + ], + "lines": [ + { + "bbox": [ + 307, + 182, + 378, + 243 + ], + "spans": [ + { + "bbox": [ + 307, + 182, + 378, + 243 + ], + "type": "image", + "image_path": "fd54acc1ace974a66bc0e83340c9b91f98a941b8feb15c81f8abd604c016c358.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 324, + 243, + 361, + 250 + ], + "lines": [ + { + "bbox": [ + 324, + 243, + 361, + 250 + ], + "spans": [ + { + "bbox": [ + 324, + 243, + 361, + 250 + ], + "type": "text", + "content": "(a)GT:5-Office" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 380, + 182, + 462, + 243 + ], + "blocks": [ + { + "bbox": [ + 380, + 182, + 462, + 243 + ], + "lines": [ + { + "bbox": [ + 380, + 182, + 462, + 243 + ], + "spans": [ + { + "bbox": [ + 380, + 182, + 462, + 243 + ], + "type": "image", + "image_path": "79fdc6e4eba59310541e8264db6edf336b6e06c37139b2ab64270f26042b255e.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 384, + 243, + 458, + 250 + ], + "lines": [ + { + "bbox": [ + 384, + 243, + 458, + 250 + ], + "spans": [ + { + "bbox": [ + 384, + 243, + 458, + 250 + ], + "type": "text", + "content": "(b) GT: 4-Mixed (three instances)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 463, + 183, + 545, + 243 + ], + "blocks": [ + { + "bbox": [ + 463, + 183, + 545, + 243 + ], + "lines": [ + { + "bbox": [ + 463, + 183, + 545, + 243 + ], + "spans": [ + { + "bbox": [ + 463, + 183, + 545, + 243 + ], + "type": "image", + "image_path": "3e01033128d2fa9375d63e7f704ee13055bb5c2f160aefd991d59ef773c584c7.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 484, + 243, + 525, + 250 + ], + "lines": [ + { + "bbox": [ + 484, + 243, + 525, + 250 + ], + "spans": [ + { + "bbox": [ + 484, + 243, + 525, + 250 + ], + "type": "text", + "content": "(c) GT: 3-Elevator" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 307, + 253, + 369, + 313 + ], + "blocks": [ + { + "bbox": [ + 307, + 253, + 369, + 313 + ], + "lines": [ + { + "bbox": [ + 307, + 253, + 369, + 313 + ], + "spans": [ + { + "bbox": [ + 307, + 253, + 369, + 313 + ], + "type": "image", + "image_path": "39b456a200f887e68fcf2d2b55a9cff239d9a73d5192af0d151aba7a3775f76a.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 319, + 314, + 356, + 320 + ], + "lines": [ + { + "bbox": [ + 319, + 314, + 356, + 320 + ], + "spans": [ + { + "bbox": [ + 319, + 314, + 356, + 320 + ], + "type": "text", + "content": "(d)GT:5-Office" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 335, + 544, + 346 + ], + "lines": [ + { + "bbox": [ + 306, + 335, + 544, + 346 + ], + "spans": [ + { + "bbox": [ + 306, + 335, + 544, + 346 + ], + "type": "text", + "content": "Figure 6. Typical failure cases of the current benchmark methods." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 369, + 253, + 422, + 313 + ], + "blocks": [ + { + "bbox": [ + 369, + 253, + 422, + 313 + ], + "lines": [ + { + "bbox": [ + 369, + 253, + 422, + 313 + ], + "spans": [ + { + "bbox": [ + 369, + 253, + 422, + 313 + ], + "type": "image", + "image_path": "bfa689068c84c71966380a96dc8822117e5cf5d3507c03dee0d289e2131bb475.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 377, + 314, + 413, + 320 + ], + "lines": [ + { + "bbox": [ + 377, + 314, + 413, + 320 + ], + "spans": [ + { + "bbox": [ + 377, + 314, + 413, + 320 + ], + "type": "text", + "content": "(e)GT:5-Office" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 423, + 253, + 481, + 313 + ], + "blocks": [ + { + "bbox": [ + 423, + 253, + 481, + 313 + ], + "lines": [ + { + "bbox": [ + 423, + 253, + 481, + 313 + ], + "spans": [ + { + "bbox": [ + 423, + 253, + 481, + 313 + ], + "type": "image", + "image_path": "dfdfa8efe3b89c78e888ce0eb8b0b70d8ec7ec2051f7fb45517d2807648b2ef0.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 433, + 314, + 470, + 320 + ], + "lines": [ + { + "bbox": [ + 433, + 314, + 470, + 320 + ], + "spans": [ + { + "bbox": [ + 433, + 314, + 470, + 320 + ], + "type": "text", + "content": "(f) GT: 7-Others" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 482, + 253, + 545, + 313 + ], + "blocks": [ + { + "bbox": [ + 482, + 253, + 545, + 313 + ], + "lines": [ + { + "bbox": [ + 482, + 253, + 545, + 313 + ], + "spans": [ + { + "bbox": [ + 482, + 253, + 545, + 313 + ], + "type": "image", + "image_path": "8d633283896b7e9e409109f760c2b09c0a7742772a950226c652cc462a03d087.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 493, + 314, + 533, + 321 + ], + "lines": [ + { + "bbox": [ + 493, + 314, + 533, + 321 + ], + "spans": [ + { + "bbox": [ + 493, + 314, + 533, + 321 + ], + "type": "text", + "content": "(g) GT: 1-Family" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 372, + 545, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 372, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 304, + 372, + 545, + 624 + ], + "type": "text", + "content": "York City, of which the annotations are generated from both existing label maps and our proposed annotation pipeline. We provide benchmark experimental results for multiple tasks and data sources based on state-of-the-art methods and analyze their limitations. We believe that OmniCity will not only promote new algorithms and application scenarios for existing tasks, but facilitate novel tasks for 3D city reconstruction and simulation. In our future work, we will keep updating the OmniCity dataset and the benchmarks in the following aspects. Owing to the proposed annotation pipeline and the unified annotations with a vector format and rich meta information (geo-locations, block-lot id, etc.), OmniCity can be efficiently supplemented with more properties of buildings and other geographical object types (roads, sidewalks, trees, green space, etc.), and extended to other cities of different countries. The benchmark results of more state-of-the-art models and new tasks will be provided accordingly. Based on the rich annotation and view types of OmniCity, we also plan to develop new methods for existing and novel tasks, such as object detection, instance segmentation, and 3D reconstruction from cross-view images." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "Acknowledgements. This project was funded in part by National Natural Science Foundation of China (No. 42201358, No. U22B2011 and No. 41820104006), Shanghai AI Laboratory, CUHK Interdisciplinary AI Research Institute, and the Centre for Perceptual and Interactive Intelligence (CPII) Ltd under the Innovation and Technology Commission (ITC)'s InnoHK." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17404" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 94, + 289, + 712 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 53, + 94, + 289, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 94, + 289, + 158 + ], + "spans": [ + { + "bbox": [ + 53, + 94, + 289, + 158 + ], + "type": "text", + "content": "[1] Seyed Majid Azimi, Corentin Henry, Lars Sommer, Arne Schumann, and Eleonora Vig. Skyscapes fine-grained semantic understanding of aerial scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7393-7403, 2019." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 162, + 288, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 162, + 288, + 252 + ], + "spans": [ + { + "bbox": [ + 53, + 162, + 288, + 252 + ], + "type": "text", + "content": "[2] Sara Beery, Guanhang Wu, Trevor Edwards, Filip Pavetic, Bo Majewski, Shreyasee Mukherjee, Stanley Chan, John Morgan, Vivek Rathod, and Jonathan Huang. The auto arborist dataset: A large-scale benchmark for multiview urban forest monitoring under domain shift. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21294-21307, 2022." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 255, + 288, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 255, + 288, + 308 + ], + "spans": [ + { + "bbox": [ + 53, + 255, + 288, + 308 + ], + "type": "text", + "content": "[3] Markus Braun, Sebastian Krebs, Fabian B. Flohr, and Dariu M. Gavrila. Eurocity persons: A novel benchmark for person detection in traffic scenes. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-1, 2019." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 310, + 288, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 310, + 288, + 388 + ], + "spans": [ + { + "bbox": [ + 53, + 310, + 288, + 388 + ], + "type": "text", + "content": "[4] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 391, + 288, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 391, + 288, + 443 + ], + "spans": [ + { + "bbox": [ + 53, + 391, + 288, + 443 + ], + "type": "text", + "content": "[5] Zhaowei Cai and Nuno Vasconcelos. Cascade r-cnn: High quality object detection and instance segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, page 1-1, 2019." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 446, + 288, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 446, + 288, + 523 + ], + "spans": [ + { + "bbox": [ + 53, + 446, + 288, + 523 + ], + "type": "text", + "content": "[6] Kai Chen, Jiangmiao Pang, Jiaqi Wang, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jianping Shi, Wanli Ouyang, et al. Hybrid task cascade for instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4974-4983, 2019." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 526, + 288, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 526, + 288, + 578 + ], + "spans": [ + { + "bbox": [ + 53, + 526, + 288, + 578 + ], + "type": "text", + "content": "[7] Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jiarui Xu, et al. Mmdetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155, 2019." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 582, + 288, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 582, + 288, + 633 + ], + "spans": [ + { + "bbox": [ + 53, + 582, + 288, + 633 + ], + "type": "text", + "content": "[8] Xiaotian Chen, Xuejin Chen, and Zheng-Jun Zha. Structure-aware residual pyramid network for monocular depth estimation. In Proceedings of the 28th International Joint Conference on Artificial Intelligence, pages 694-700, 2019." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 636, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 636, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 53, + 636, + 288, + 712 + ], + "type": "text", + "content": "[9] Gordon Christie, Rodrigo Rene Rai Munoz Abujder, Kevin Foster, Shea Hagstrom, Gregory D Hager, and Myron Z Brown. Learning geocentric object pose in oblique monocular images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14512-14520, 2020." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 714 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 150 + ], + "type": "text", + "content": "[10] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3213-3223, 2016." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 307, + 154, + 545, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 154, + 545, + 232 + ], + "spans": [ + { + "bbox": [ + 307, + 154, + 545, + 232 + ], + "type": "text", + "content": "[11] Ilke Demir, Krzysztof Koperski, David Lindenbaum, Guan Pang, Jing Huang, Saikat Basu, Forest Hughes, Devis Tuia, and Ramesh Raskar. Deep globe 2018: A challenge to parse the earth through satellite images. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2018." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 307, + 236, + 545, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 236, + 545, + 313 + ], + "spans": [ + { + "bbox": [ + 307, + 236, + 545, + 313 + ], + "type": "text", + "content": "[12] Tian Feng, Quang-Trung Truong, Duc Thanh Nguyen, Jing Yu Koh, Lap-Fai Yu, Alexander Binder, and Sai-Kit Yeung. Urban zoning using higher-order markov random fields on multi-view imagery data. In Proceedings of the European Conference on Computer Vision (ECCV), pages 614–630, 2018." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 318, + 545, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 318, + 545, + 382 + ], + "spans": [ + { + "bbox": [ + 307, + 318, + 545, + 382 + ], + "type": "text", + "content": "[13] Huan Fu, Mingming Gong, Chaohui Wang, Kayhan Bat-manghelich, and Dacheng Tao. Deep ordinal regression network for monocular depth estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2002-2011, 2018." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 386, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 386, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 307, + 386, + 545, + 437 + ], + "type": "text", + "content": "[14] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE conference on computer vision and pattern recognition, pages 3354-3361. IEEE, 2012." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 441, + 545, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 441, + 545, + 491 + ], + "spans": [ + { + "bbox": [ + 307, + 441, + 545, + 491 + ], + "type": "text", + "content": "[15] Kaiming He, Georgia Gkioxari, Piotr Dollár, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision (CVPR), pages 2961-2969, 2017." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 497, + 545, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 497, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 307, + 497, + 545, + 548 + ], + "type": "text", + "content": "[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 552, + 545, + 616 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 552, + 545, + 616 + ], + "spans": [ + { + "bbox": [ + 307, + 552, + 545, + 616 + ], + "type": "text", + "content": "[17] Sixing Hu, Mengdan Feng, Rang MH Nguyen, and Gim Hee Lee. Cvm-net: Cross-view matching network for image-based ground-to-aerial geo-localization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7258-7267, 2018." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 620, + 545, + 684 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 620, + 545, + 684 + ], + "spans": [ + { + "bbox": [ + 307, + 620, + 545, + 684 + ], + "type": "text", + "content": "[18] Xinyu Huang, Peng Wang, Xinjing Cheng, Dingfu Zhou, Qichuan Geng, and Ruigang Yang. The apolloscape open dataset for autonomous driving and its application. IEEE transactions on pattern analysis and machine intelligence, 42(10):2702-2719, 2019." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 689, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 307, + 689, + 545, + 714 + ], + "type": "text", + "content": "[19] Zhaojin Huang, Lichao Huang, Yongchao Gong, Chang Huang, and Xinggang Wang. Mask scoring r-cnn. In Pro-" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17405" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 97 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 97 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 97 + ], + "type": "text", + "content": "ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6409-6418, 2019." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 100, + 288, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 100, + 288, + 163 + ], + "spans": [ + { + "bbox": [ + 48, + 100, + 288, + 163 + ], + "type": "text", + "content": "[20] Weijia Li, Conghui He, Jiarui Fang, Juepeng Zheng, Hao-huan Fu, and Le Yu. Semantic segmentation-based building footprint extraction using very high-resolution satellite images and multi-source gis data. Remote Sensing, 11(4):403, 2019." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 167, + 288, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 167, + 288, + 232 + ], + "spans": [ + { + "bbox": [ + 48, + 167, + 288, + 232 + ], + "type": "text", + "content": "[21] Weijia Li, Lingxuan Meng, Jinwang Wang, Conghui He, Gui-Song Xia, and Dahua Lin. 3d building reconstruction from monocular remote sensing images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12548-12557, 2021." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 234, + 288, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 234, + 288, + 271 + ], + "spans": [ + { + "bbox": [ + 48, + 234, + 288, + 271 + ], + "type": "text", + "content": "[22] Weijia Li, Wenqian Zhao, Huaping Zhong, Conghui He, and Dahua Lin. Joint semantic-geometric learning for polygonal building segmentation. In AAAI, 2021." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 274, + 288, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 274, + 288, + 338 + ], + "spans": [ + { + "bbox": [ + 48, + 274, + 288, + 338 + ], + "type": "text", + "content": "[23] Tsung-Yi Lin, Piotr Dólar, Ross Girshick, Kaiming He, Bharath Hariharan, and Serge Belongie. Feature pyramid networks for object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition (CVPR), pages 2117-2125, 2017." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 341, + 288, + 406 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 341, + 288, + 406 + ], + "spans": [ + { + "bbox": [ + 48, + 341, + 288, + 406 + ], + "type": "text", + "content": "[24] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 408, + 288, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 408, + 288, + 472 + ], + "spans": [ + { + "bbox": [ + 48, + 408, + 288, + 472 + ], + "type": "text", + "content": "[25] Xiaohu Lu, Zuoyue Li, Zhaopeng Cui, Martin R Oswald, Marc Pollefeys, and Rongjun Qin. Geometry-aware satellite-to-ground image synthesis for urban areas. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 859-867, 2020." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 475, + 288, + 538 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 475, + 288, + 538 + ], + "spans": [ + { + "bbox": [ + 48, + 475, + 288, + 538 + ], + "type": "text", + "content": "[26] Gerhard Neuhold, Tobias Ollmann, Samuel Rota Bulo, and Peter Kontschieder. The mapillary vistas dataset for semantic understanding of street scenes. In Proceedings of the IEEE international conference on computer vision, pages 4990-4999, 2017." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 542, + 288, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 542, + 288, + 592 + ], + "spans": [ + { + "bbox": [ + 48, + 542, + 288, + 592 + ], + "type": "text", + "content": "[27] Krishna Regmi and Ali Borji. Cross-view image synthesis using conditional gans. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3501-3510, 2018." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 596, + 288, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 596, + 288, + 647 + ], + "spans": [ + { + "bbox": [ + 48, + 596, + 288, + 647 + ], + "type": "text", + "content": "[28] Krishna Regmi and Mubarak Shah. Bridging the domain gap for ground-to-aerial image matching. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 470-479, 2019." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 650, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 650, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 650, + 288, + 713 + ], + "type": "text", + "content": "[29] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. International journal of computer vision (IJCV), 115(3):211-252, 2015." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 307, + 73, + 546, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 546, + 124 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 546, + 124 + ], + "type": "text", + "content": "[30] Yujiao Shi, Dylan John Campbell, Xin Yu, and Hongdong Li. Geometry-guided street-view panorama synthesis from satellite imagery. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 307, + 126, + 547, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 126, + 547, + 204 + ], + "spans": [ + { + "bbox": [ + 307, + 126, + 547, + 204 + ], + "type": "text", + "content": "[31] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2446-2454, 2020." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 207, + 546, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 207, + 546, + 297 + ], + "spans": [ + { + "bbox": [ + 307, + 207, + 546, + 297 + ], + "type": "text", + "content": "[32] Zheng Tang, Milind Naphade, Ming-Yu Liu, Xiaodong Yang, Stan Birchfield, Shuo Wang, Ratnesh Kumar, David Anastasiu, and Jenq-Neng Hwang. Cityflow: A city-scale benchmark for multi-target multi-camera vehicle tracking and re-identification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8797–8806, 2019." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 300, + 546, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 300, + 546, + 351 + ], + "spans": [ + { + "bbox": [ + 307, + 300, + 546, + 351 + ], + "type": "text", + "content": "[33] Yicong Tian, Chen Chen, and Mubarak Shah. Cross-view image matching for geo-localization in urban environments. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3608-3616, 2017." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 354, + 547, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 354, + 547, + 418 + ], + "spans": [ + { + "bbox": [ + 307, + 354, + 547, + 418 + ], + "type": "text", + "content": "[34] Adam Van Etten, Daniel Hogan, Jesus Martinez Manso, Jacob Shermeyer, Nicholas Weir, and Ryan Lewis. The multitemporal urban development spacenet dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6398-6407, 2021." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 421, + 546, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 421, + 546, + 472 + ], + "spans": [ + { + "bbox": [ + 307, + 421, + 546, + 472 + ], + "type": "text", + "content": "[35] Jiaqi Wang, Kai Chen, Rui Xu, Ziwei Liu, Chen Change Loy, and Dahua Lin. Carafe: Content-aware reassembly of features. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3007-3016, 2019." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 475, + 546, + 539 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 475, + 546, + 539 + ], + "spans": [ + { + "bbox": [ + 307, + 475, + 546, + 539 + ], + "type": "text", + "content": "[36] Shenlong Wang, Min Bai, Gellert Mattyus, Hang Chu, Wenjie Luo, Bin Yang, Justin Liang, Joel Cheverie, Sanja Fidler, and Raquel Urtasun. Toronto: Seeing the world with a million eyes. In 2017 IEEE International Conference on Computer Vision (ICCV), pages 3028-3036. IEEE, 2017." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 541, + 546, + 605 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 541, + 546, + 605 + ], + "spans": [ + { + "bbox": [ + 307, + 541, + 546, + 605 + ], + "type": "text", + "content": "[37] Jan D Wegner, Steven Branson, David Hall, Konrad Schindler, and Pietro Perona. Cataloging public objects using aerial and street-level images-urban trees. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6014–6023, 2016." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 609, + 546, + 685 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 609, + 546, + 685 + ], + "spans": [ + { + "bbox": [ + 307, + 609, + 546, + 685 + ], + "type": "text", + "content": "[38] Nicholas Weir, David Lindenbaum, Alexei Bastidas, Adam Van Etten, Sean McPherson, Jacob Shermeyer, Varun Kumar, and Hanlin Tang. Spacenet mvoi: A multi-view overhead imagery dataset. In Proceedings of the IEEE/cvf international conference on computer vision, pages 992-1001, 2019." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 689, + 546, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 689, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 689, + 546, + 713 + ], + "type": "text", + "content": "[39] Zbigniew Wojna, Krzysztof Maziarz, Łukasz Jocz, Robert Pałuba, Robert Kozikowski, and Iason Kokkinos. Holis-" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "17406" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 110 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 110 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 110 + ], + "type": "text", + "content": "tic multi-view building analysis in the wild with projection pooling. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 2870-2878, 2021." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 115, + 287, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 115, + 287, + 178 + ], + "spans": [ + { + "bbox": [ + 48, + 115, + 287, + 178 + ], + "type": "text", + "content": "[40] Scott Workman, M Usman Rafique, Hunter Blanton, and Nathan Jacobs. Revisiting near/remote sensing with geospatial attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1778-1787, 2022." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 184, + 287, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 184, + 287, + 233 + ], + "spans": [ + { + "bbox": [ + 48, + 184, + 287, + 233 + ], + "type": "text", + "content": "[41] Kailun Yang, Xinxin Hu, Luis M Bergasa, Eduardo Romera, and Kaiwei Wang. Pass: Panoramic annular semantic segmentation. IEEE Transactions on Intelligent Transportation Systems, 21(10):4171-4185, 2019." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 239, + 287, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 239, + 287, + 289 + ], + "spans": [ + { + "bbox": [ + 48, + 239, + 287, + 289 + ], + "type": "text", + "content": "[42] Kailun Yang, Xinxin Hu, and Rainer Stiefelhagen. Is context-aware cnn ready for the surroundings? panoramic semantic segmentation in the wild. IEEE Transactions on Image Processing, 30:1866-1881, 2021." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 294, + 287, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 294, + 287, + 371 + ], + "spans": [ + { + "bbox": [ + 48, + 294, + 287, + 371 + ], + "type": "text", + "content": "[43] Xiaoqing Ye, Mao Shu, Hanyu Li, Yifeng Shi, Yingying Li, Guangjie Wang, Xiao Tan, and Errui Ding. Rope3d: The roadside perception dataset for autonomous driving and monocular 3d object detection task. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21341-21350, 2022." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 376, + 287, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 376, + 287, + 452 + ], + "spans": [ + { + "bbox": [ + 48, + 376, + 287, + 452 + ], + "type": "text", + "content": "[44] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darryll. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2636-2645, 2020." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 457, + 287, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 457, + 287, + 520 + ], + "spans": [ + { + "bbox": [ + 48, + 457, + 287, + 520 + ], + "type": "text", + "content": "[45] Menghua Zhai, Zachary Bessinger, Scott Workman, and Nathan Jacobs. Predicting ground-level scene layout from aerial imagery. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 867-875, 2017." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 525, + 287, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 287, + 602 + ], + "type": "text", + "content": "[46] Chiqun Zhang, Dragomir Yankov, Chun-Ting Wu, Simon Shapiro, Jason Hong, and Wei Wu. What is that building? an end-to-end system for building recognition from streetside images. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pages 2425-2433, 2020." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 607, + 287, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 607, + 287, + 657 + ], + "spans": [ + { + "bbox": [ + 48, + 607, + 287, + 657 + ], + "type": "text", + "content": "[47] Yichao Zhou, Jingwei Huang, Xili Dai, Shichen Liu, Linjie Luo, Zhili Chen, and Yi Ma. Holicity: A city-scale data platform for learning holistic 3d structures. arXiv preprint arXiv:2008.03286, 2020." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 662, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 662, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 662, + 287, + 713 + ], + "type": "text", + "content": "[48] Sijie Zhu, Taojiannan Yang, and Chen Chen. Vigor: Cross-view image geo-localization beyond one-to-one retrieval. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3640-3649, 2021." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 545, + 122 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 545, + 122 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 545, + 122 + ], + "type": "text", + "content": "[49] Yi Zhu, Xueqing Deng, and Shawn Newsam. Fine-grained land use classification at the city scale using ground-level images. IEEE Transactions on Multimedia, 21(7):1825-1838, 2019." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "17407" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/OmniMAE_ Single Model Masked Pretraining on Images and Videos/2fdd9b20-eaac-4038-a247-56813890042d_content_list.json b/2023/OmniMAE_ Single Model Masked Pretraining on Images and Videos/2fdd9b20-eaac-4038-a247-56813890042d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ace7aa7967bbcdd91d64ede5932010b11c339fc3 --- /dev/null +++ b/2023/OmniMAE_ Single Model Masked Pretraining on Images and Videos/2fdd9b20-eaac-4038-a247-56813890042d_content_list.json @@ -0,0 +1,1359 @@ +[ + { + "type": "text", + "text": "OmniMAE: Single Model Masked Pretraining on Images and Videos", + "text_level": 1, + "bbox": [ + 138, + 130, + 831, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Rohit Girdhar* Alaaeldin El-Nouby* Mannat Singh* \nKalyan Vasudev Alwala* Armand Joulin Ishan Misra* \nFAIR, Meta AI", + "bbox": [ + 253, + 180, + 712, + 232 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://github.com/facebookresearch/omnivore", + "bbox": [ + 331, + 237, + 630, + 250 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/af9705937884c82df0e348223ab58e9199ed22c6ae7d522c3fe23126aa586415.jpg", + "image_caption": [ + "Figure 1. OmniMAE is a single model for images and videos that is trained using masked autoencoding [40]. We use a plain Vision Transformer [24] architecture but with spatio-temporal patches as input. At training, we 'patchify' the visual input (images or videos), and feed the encoder only a subset of the patches. The decoder reconstructs the pixels for the missing patches using the encoder's output. The encoder-decoder model is trained using a pixel reconstruction loss. After training, our single plain Transformer encoder performs competitively compared to specialized architectures on downstream image and video recognition tasks." + ], + "image_footnote": [], + "bbox": [ + 117, + 284, + 851, + 465 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 556, + 313, + 571 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Transformer-based architectures have become competitive across a variety of visual domains, most notably images and videos. While prior work studies these modalities in isolation, having a common architecture suggests that one can train a single unified model for multiple visual modalities. Prior attempts at unified modeling typically use architectures tailored for vision tasks, or obtain worse performance compared to single modality models. In this work, we show that masked autoencoding can be used to train a simple Vision Transformer on images and videos, without requiring any labeled data. This single model learns visual representations that are comparable to or better than single-modality representations on both image and video benchmarks, while using a much simpler architecture. Furthermore, this model can be learned by dropping $90\\%$ of the image and $95\\%$ of the video patches, enabling extremely fast training of huge model architectures. In particular, we show that our single ViT-Huge model can be finetuned to achieve $86.6\\%$ on ImageNet and $75.5\\%$ on the challenging Something Something-v2 video", + "bbox": [ + 75, + 588, + 473, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "benchmark, setting a new state-of-the-art.", + "bbox": [ + 500, + 556, + 777, + 571 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 602, + 630, + 618 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The Transformer architecture [78] is rapidly becoming competitive across the different visual modalities in Computer Vision, from images [24, 27, 55, 77], to 3D [57, 60, 89] and videos [2, 9, 27, 31, 32, 56]. This convergence toward a unified architecture naturally suggests that we should be able to train a single model that works across different visual modalities. However, recent attempts to train unified models either lead to worse performance compared to single modality models [53], or require the use of an alternative architecture [33], namely the Swin Transformer [55], with inductive biases tailored towards vision tasks. While specialized Transformer architectures for vision [27, 55, 56, 81] can offer better performance for visual modalities, they lose the generality and flexibility of the vanilla Transformer, making it harder to later model different domains like text, speech, 3D etc. in multi-modal architectures.", + "bbox": [ + 496, + 627, + 895, + 868 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we train a single vanilla Transformer that works for both images and videos, as illustrated in Figure 1.", + "bbox": [ + 500, + 869, + 893, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal technical contribution.", + "bbox": [ + 101, + 887, + 264, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "10406", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To this end, we leverage the findings of several recent works on the use of the masked pretraining [23] to greatly improve the training and performance of Transformers in the domain of images [6, 40, 80, 84], videos [76, 80, 82] or across text, audio and images [5]. We show that this masked pretraining is a viable strategy to pretrain a unified 'omnivorous' Transformer across visual modalities. In particular, we consider the Masked Auto-Encoding (MAE) approach [40] to train an Omnivorous visual encoder [33]. The resulting OmniMAE model learns from all the modalities with the same objective function and does not require any supervision.", + "bbox": [ + 75, + 90, + 472, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Using a masked pretraining objective has several advantages over supervised objectives [33, 53] or discriminative self-supervised objectives [13, 17, 41]. First, as opposed to supervised objectives, a general-purpose unsupervised loss does not require any human labeling effort. As a result, it is robust to biases introduced by a predefined set of labels [35]. Moreover, it does not require a multi-head architecture to incorporate supervision from each of the label spaces corresponding to each modality, which is hard to maintain and scale with new modalities. Second, although discriminative self-supervised methods produce superior frozen features compared to reconstruction objectives, they are non trivial to scale in model and data size [36]. Our masked pretraining objective is simple, efficient to train, and scales to different visual modalities with minimal changes.", + "bbox": [ + 75, + 268, + 472, + 496 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Contributions. (1) We show that the simple Vision Transformer architecture (ViT) [24] originally designed for images can naturally be applied on videos, and videos and images jointly. OmniMAE is a single ViT-based model for videos and images that outperforms architectures and models specifically designed for either modality. (2) Prior and concurrent work design self-supervised methods and architectures for either image or video and we find that these models do not transfer well across modalities. OmniMAE is the first single self-supervised model that achieves good performance on both modalities. (3) We show that our joint training using both images and videos enables us to use much higher masking ratios than any prior work for training MAE. Since ViT can processes only the non-masked input, we train OmniMAE models with only $10\\%$ of image and $5\\%$ of video patches. This enables us to train large (650M parameter) models with a $\\sim 7\\times$ and $\\sim 11\\times$ reduction in compute and memory on images and videos. (4) Finally, we propose improvements to the MAE training. We show that repeating samples in a mini-batch reduces dataloading (and thus training) time without loss in final transfer performance. Sample replication is particularly useful for masked pretraining as the unmasked patches are different across sample replicas. We also show that using a shallow shared decoder for videos and images leads to better performance while reducing the number of parameters by $2 - 4\\times$ .", + "bbox": [ + 75, + 507, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 89, + 640, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our work builds upon research in self-supervised learning, masked pretraining and unified modeling in computer vision. Self-supervised learning. In recent years, self-supervised approaches have been dominated by joint embedding methods which can rely on different objectives including contrastive [17,39,41,51,61,65,75,83], non-contrastive [7,18,26, 87], clustering [3, 11, 12, 85] or self-distillation [5, 13, 38, 92]. Such methods are trained to learn invariance to a set of predefined transformations which results in image descriptors with a strong linear probing and KNN performance. However, such methods can be challenging to scale since they can suffer from instabilities [19]. Additionally, the strongest performance is typically obtained with the help of augmentations like multi-crop [12, 13, 92] which can be hard to apply at scale due their compute and memory overhead.", + "bbox": [ + 496, + 114, + 893, + 340 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Masked pretraining. We build upon masked prediction methods where the representation is learned by predicting masked parts of the input. Such methods have recently gained popularity given their immense success in NLP. In particular, BERT [23] showed that masked language modeling by predicting a subset of masked words is a powerful pre-training objective and leads to impressive finetuning performance on various downstream tasks. In computer vision, input reconstruction methods have a rich history with non-linear PCA [49], sparse reconstruction [64], autoencoders [10, 30, 42, 50], RBMs [71] etc. Masked prediction methods can be viewed as a special case of denoising autoencoders [30, 79] where the input 'noise' is a masking function. An example of such a method that uses masking as noise is context encoders [68]. With the recent and rapid rise of Vision Transformers [24], masked prediction was revisited by multiple efforts. SiT [4] replaces variable sized patches in the image with random noise and trains a ViT model for reconstruction, among other objectives. BEiT [6] moves a step closer to BERT with replacing full patches with mask tokens and training a ViT encoder to predict the discrete visual words of masked patches using a cross-entropy loss. Masked prediction has also shown impressive performance for specialized vision transformer architectures such as MViT [27, 52, 82] and Swin transformer [54, 55, 84]. SimMIM [84] and MaskedFeat [82] predict pixel values and HOG features of the masked patches using a Swin-v2 [54] and MViT-v2 [52] backbones respectively. Finally, SplitMask [25] studied the interesting properties of masked prediction methods in terms of high sample efficiency.", + "bbox": [ + 496, + 340, + 895, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Of particular interest to OmniMAE, masked autoencoders (MAE) [40] demonstrated impressive scaling properties by utilizing a patch dropping strategy for masked patches accompanied with a high masking ratio of $75\\%$ . Under this setting, the encoder only process a small subset of the image patches followed by a relatively low capacity decoder which reconstructs the image pixels. This property is even more", + "bbox": [ + 496, + 795, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "10407", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "crucial for video representation learning given the large number of patches, and a concurrent work, ST-MAE [76], shows that MAE pretraining with an even higher masking ratio of $90\\%$ works well and obtains strong finetuning performance on downstream video recognition benchmarks. Notably, the efficiency gained by patch dropping is specific to vanilla ViTs, and multiscale architectures such as MViT and Swin are unable to benefit due to their design. Hence, we use the simple ViT as our architecture and show that it can be trained efficiently and jointly for images and videos using extremely high masking ratios (90-95% on both modalities), and yet perform competitively to specialized architectures like MViT [27] and Swin [55].", + "bbox": [ + 75, + 90, + 472, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unified modeling and multi-modal learning. Multi-modal learning in computer vision has a long history that includes training using images and text [15, 34, 47, 58, 59], video and optical flow [29, 72], and video and audio [1, 62, 63, 66]. The majority of such methods rely on training a separate backbone for each modality as well as the availability of alignment across modalities. More recently, Omnivore [33] was proposed for joint training of multiple modalities like images, videos and single-view 3D, attaining a strong performance in each of the modality specific benchmarks with a single shared trunk. PolyViT [53] co-trains a shared transformer encoder using images, videos and audio data and provides a competitive performance on various downstream tasks for each modality. The aforementioned methods differ compared to OmniMAE in that they are trained with supervised learning and require human annotations. BEVT [80] tackles BERT pre-training for videos and proposes that jointly pretraining using static images improves the finetuning performance of video recognition benchmarks. Unlike OmniMAE, BEVT uses the specialized Swin transformer architecture with separate decoder heads for images and videos. Thus, BEVT cannot drop patches while training which can limit its scalability. Furthermore, it relies on a tokenizer which must be trained apriori, and the tokenizer training itself can affect the model's performance.", + "bbox": [ + 75, + 299, + 472, + 679 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. OmniMAE", + "text_level": 1, + "bbox": [ + 76, + 727, + 194, + 742 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our goal is to pretrain a single unified model for images and videos. Rather than use specialized architectures tailored for a visual modality, we build upon the vanilla Vision Transformer (ViT) [24] architecture that has limited inductive biases for vision. For pretraining, we extend the simple self-supervised masked auto-encoding (MAE) approach [40]. The original architecture and pretraining method are tested only on images, and we show simple design decisions for a unified model.", + "bbox": [ + 75, + 763, + 470, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Training OmniMAE jointly on images and videos", + "text_level": 1, + "bbox": [ + 498, + 90, + 890, + 119 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We illustrate our method in Figure 1. For pretraining, we use an encoder-decoder architecture where the encoder only operates on a 'non-masked' subset of the input. The decoder predicts the pixel values for the entire input, i.e., masked and non-masked pixels. The model is trained to minimize the reconstruction error for the masked (unseen) part of the input. After pretraining, we evaluate the encoder by transfer learning (the decoder is discarded). Next, we describe the pretraining details.", + "bbox": [ + 496, + 128, + 893, + 265 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Images and videos as spatio-temporal patches. The input image or video can be represented as a 4D tensor of shape $T \\times H \\times W \\times 3$ where $T$ is the temporal dimension and $H, W$ are the spatial dimensions, and 3 represents the color channels. We treat images as being single-frame videos with $T = 1$ . The input is split into $N$ spatio-temporal patches, each of size $t \\times h \\times w \\times 3$ [33].", + "bbox": [ + 496, + 265, + 893, + 369 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Omnivorous visual encoder. We use an omnivorous [33] visual encoder that processes both images and video using the same parameters. The encoder operates on the $N$ spatiotemporal patches from the images and videos. The encoder can naturally handle variable number $N$ of patches from images and videos as it uses the Transformer architecture [78]. The encoder shares the same parameters for both image and video inputs and processes them in the same way to output per-patch embeddings.", + "bbox": [ + 496, + 369, + 893, + 507 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Pretraining. We convert the input into $N$ spatio-temporal patches and randomly mask $M$ of them. Following [40], we feed the non-masked $N - M$ patches (and their positions) to the encoder to produce per-patch embeddings. The encoder output is concatenated with $M$ copies of a learnable 'mask token' to obtain $N$ embeddings. We add the appropriate positional encoding to the $N$ embeddings and feed them into the decoder, which outputs $N \\times t \\times h \\times w \\times 3$ pixel values.", + "bbox": [ + 496, + 507, + 893, + 628 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Loss function and optimization. We minimize the reconstruction error between the decoder predictions and the input pixel values. The input pixel values are normalized to zero mean and unit variance [46] to get the targets for the loss. We minimize the $\\ell_2$ distance between the predictions and targets over the $M$ masked patches. At each point in the training, we sample a mini-batch of either images or video and compute the loss using the decoder predictions. We study the effect of different mini-batch construction strategies on the overall performance and speed of training in § 4.2.", + "bbox": [ + 496, + 628, + 893, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Dataset sample replication. Since we operate on a small subset of the input patches $(M \\ll N)$ , we find that our data reading speed is unable to keep up with the number of data points we can theoretically process in a single optimization step. This issue is even more pronounced for videos, where we spend significant compute to read and decode a full video, only to discard $>90\\%$ of it. Inspired from the success of RepeatAugment, which shows that replicating samples", + "bbox": [ + 496, + 780, + 893, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "10408", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d1ac4e9640794e89298e95775d504095633419cb539bbb20673e66969149a038.jpg", + "image_caption": [ + "Figure 2. OmniMAE on image and video downstream tasks. We finetune the MAE, ST-MAE, and OmniMAE models on image and video benchmarks. We use the ViT architecture with two model sizes: ViT-B and ViT-L. MAE has poor video recognition performance while ST-MAE's performance drops on image datasets. OmniMAE pretraining generalizes to both benchmarks. All models are trained for 800 epochs on the pretraining datasets. The image-only MAE model is inflated [14] to apply MAE to video recognition tasks. The input image is replicated to apply ST-MAE to image recognition benchmarks." + ], + "image_footnote": [], + "bbox": [ + 84, + 85, + 890, + 229 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "within a mini-batch is an effective technique to improve generalization [8, 12, 43], we replicate a single data point and apply different masks to it each time. Even with replicated samples, the non-masked input to the model is different due to random cropping and masking being different across the replicas. We show in § 4.2 that sample replication leads to gains in runtime without affecting the final transfer accuracy.", + "bbox": [ + 75, + 338, + 473, + 446 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Implementation Details", + "text_level": 1, + "bbox": [ + 76, + 455, + 294, + 470 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We note some of the salient implementation details and provide the complete details in Appendix B.", + "bbox": [ + 75, + 479, + 470, + 508 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Architecture. We use the ViT [24] architecture for the omnivorous encoder and experiment with its ViT-B, ViT-L, and ViT-H variants. We do not use the [CLS] token in the ViT models, yielding a small improvement in runtime without loss in performance. We use a Transformer decoder with 4 layers (8 for ViT-H) of 384, 512, and 512 embedding dimensions for ViT-B, ViT-L, and ViT-H, respectively. The decoder outputs the RGB colors for the pixels in all the input patches. We use sinusoidal positional encoding [78] for the patches in both the encoder and the decoder.", + "bbox": [ + 75, + 510, + 468, + 660 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Training details. We train our models with a mini-batch size of 2048. We resize the input images and videos spatially to $224 \\times 224$ pixels. For videos, we sample a clip of $T = 16$ frames at 6 FPS. We use a patch size of $2 \\times 16 \\times 16$ for ViT-B and ViT-L, and $2 \\times 14 \\times 14$ for ViT-H. Images are replicated temporally to meet the patch size.", + "bbox": [ + 75, + 662, + 468, + 753 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Masking the input. Compared to prior work, we use extremely high masking for pretraining and only $10\\%$ and $5\\%$ of the image and video patches are fed to the encoder for pretraining. We uniform randomly mask the patches for images and videos and ablate the masking hyperparameters in $\\S 4.2$ .", + "bbox": [ + 75, + 753, + 470, + 829 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 844, + 209, + 861 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Pretraining data. We pretrain representations by jointly training on images from the ImageNet (IN1K) [70] dataset", + "bbox": [ + 75, + 869, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and videos from the Something Something-v2 (SSv2) [37] dataset. We choose the SSv2 dataset over web video datasets like Kinetics-400 (K400) [48] for reproducibility, and given SSv2's challenging nature requiring temporal reasoning.", + "bbox": [ + 496, + 338, + 892, + 398 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Transfer learning datasets. We consider two image datasets: iNaturalist-2018 (iNat18) [44], a popular fine-grained recognition dataset, and Places-365 (P365) [91], a scene recognition benchmark. For video datasets, we focus on the popular K400 action recognition benchmark, and EPIC-Kitchens-100 (EK100) [22], an egocentric action recognition dataset. We report the top-1 classification accuracy for all transfer datasets. The details about the datasets can be found in Appendix A.", + "bbox": [ + 496, + 398, + 893, + 535 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. OmniMAE on Image and Video Recognition", + "text_level": 1, + "bbox": [ + 498, + 544, + 877, + 561 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We now evaluate the capabilities of OmniMAE's representations on both image and video recognition tasks.", + "bbox": [ + 496, + 568, + 893, + 598 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Baselines. We compare OmniMAE's representations trained jointly on images and videos to MAE, trained solely on images [40]. Additionally, we develop a video-only baseline, SpatioTemporal-MAE (ST-MAE), by training OmniMAE only on videos. For a fair comparison, all models are pretrained for 800 epochs on their respective datasets. The image-only MAE model is inflated [14] for finetuning on video tasks, see Appendix B.2 for details.", + "bbox": [ + 496, + 599, + 893, + 718 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Observations. Figure 2 presents the evaluation results for all the models on the image and video benchmarks. The modality specific MAE and ST-MAE models achieve good transfer performance when transferring to datasets which match the pretraining modality. However, both models show a degradation in performance when transferring to a different visual modality. MAE has lower video recognition performance, especially on egocentric videos from EK100 where the smaller ViT-B MAE model is $18.4\\%$ worse compared to ST-MAE. Similarly, compared to MAE, ST-MAE is $8.7\\%$ worse on the fine-grained classification task of iNat18. With a large model, MAE's performance on the image recognition", + "bbox": [ + 496, + 719, + 895, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "10409", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a1eb0f78d4fcbcc4d87c895bfbb0346df6d33657707c0c6c292d228dfcc0067c.jpg", + "image_caption": [ + "Figure 3. Different types of masking for images (left two) and videos. Causal and tube masking use the data's spatio-temporal structure. Random frame masking randomly masks frames in a video. Random masking randomly masks patches and is used by default for OmniMAE." + ], + "image_footnote": [], + "bbox": [ + 117, + 87, + 236, + 188 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e01de3061587e5a19d129ae2652a3c26da65b7a04d12868d46af82a2d4fa961c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 243, + 88, + 357, + 189 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1c6238b96d5ab1b44f95246d7e7b204b70e8aa02adc30ab3802282f60ba077da.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 370, + 88, + 480, + 189 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0df45b5cc7f205ade1731b919b40ecf46e1e6ee18b3660ed91fecd37458a9132.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 488, + 88, + 596, + 189 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/cafecf3af80989efae0adcac0170a585578b07837ebec5ff90539675851fd9cc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 604, + 88, + 714, + 189 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/cb30810cf07f8315653191565f6939e5ab2707f672c6cdc5b122fb735c911a17.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 725, + 88, + 846, + 190 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "benchmarks improves but the cross-modality performance degrades further. On EPIC-Kitchens-100 the large MAE model is more than $25\\%$ worse than ST-MAE.", + "bbox": [ + 75, + 258, + 468, + 303 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "When transferred to both image and video recognition tasks, OmniMAE performs favorably to the single modality baselines that use exactly the same architecture. OmniMAE also uses the same finetuning recipe as the single modality baselines. OmniMAE matches the video classification performance of ST-MAE and the image classification performance of MAE for both ViT-B and ViT-L, with OmniMAE's performance improving on both image and video recognition benchmarks with the larger model. These results suggest that our unified model of images and videos serves as a better pretraining across a wider variety of recognition tasks than single modality models.", + "bbox": [ + 75, + 306, + 468, + 488 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Qualitative results. Following [40], we re-train OmniMAE without normalizing the pixel targets to obtain easy to visualize RGB reconstructions. We visualize the predictions in Figure 4 using samples that are unseen during training: either val sets or different datasets altogether. OmniMAE makes reasonable predictions on the in-distribution ImageNet and SSv2 val sets, as well as the unseen, out-of-distribution K400 and EK100 datasets. As expected, the details in the reconstruction decrease when increasing the masking ratio. However, even with $95\\%$ masking, the model can reconstruct coarse details in the input, e.g., in ImageNet, the model reconstructs the coarse structure of the flower, vehicle, dog etc.", + "bbox": [ + 75, + 491, + 472, + 674 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3da07f163da4192e2af7cb84ade3b8943096683fa66e96192de72f49af0fb792.jpg", + "image_caption": [ + "Figure 5. Sample replication. We study the effect of repeating samples while training our model. In each case, we repeat a sample $n$ times within a mini-batch while fixing the overall mini-batch size and training updates. Replication leads to improved training speeds, especially on video without affecting the final performance." + ], + "image_footnote": [], + "bbox": [ + 84, + 718, + 267, + 815 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3437c6757e850f8d9b1e98baeab30c0df4a211cb589bfb6a02bb6f907a6ec27a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 272, + 720, + 457, + 814 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Analyzing OmniMAE", + "text_level": 1, + "bbox": [ + 500, + 258, + 702, + 273 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We train the ViT-B architecture jointly on the ImageNet and the SSv2 datasets for 800 epochs, where an epoch involves training over all samples from both the datasets. For these analysis experiments, we use the default masking hyperparameters of MAE for images (75%). For ST-MAE, due to redundancy across frames, we use \"tube\" dropping (i.e. dropping all patches at a given spatial location over time) with high masking ratio (90%) All hyperparameters for these experiments are in Appendix B.3. We evaluate all the pretrained models on ImageNet and SSv2 and report the top-1 classification accuracies in Table 1. Please see Appendix D for additional ablations.", + "bbox": [ + 496, + 286, + 893, + 467 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Extreme masking. We vary the ratio of masked patches in the input while training the model. We observe that videos benefit from significantly higher amounts of masking than images. Since video frames have a lot of redundant information, the resulting spatio-temporal patches are also redundant and thus higher masking leads to better self-supervised models. Unlike our experiments, prior work found that extremely high masking ratios lead to degradation in performance, for instance MAE [40] saw a significant degradation in performance when masking more than $75\\%$ of the patches. However, as we show in Table 1a, OmniMAE models can use extremely high masking of the input while learning good representations.", + "bbox": [ + 496, + 472, + 893, + 669 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Reduced compute due to extreme masking. Our models trained with $90\\%$ masking on images and $95\\%$ masking on videos yield good performance, while being trained with just 19 and 78 unmasked image and video patches respectively, for a patch size of $2 \\times 16 \\times 16$ . As we follow [40] and only pass unmasked patches to the encoder and have a lightweight decoder, extreme masking leads to a dramatically lower computational cost for training the encoder, and consequently the model as a whole. Compared to using all the patches, our masked autoencoder uses $5.9 \\times$ and $7.8 \\times$ fewer FLOPS for ViT-B on images and videos respectively, $7.1 \\times$ and $11.6 \\times$ for ViT-L, and $7.2 \\times$ and $11.3 \\times$ for ViT-H. Compared to MAE, on ViT-B, ViT-L and ViT-H, our higher masking leads to $1.8 \\times$ , $2.0 \\times$ and $2.0 \\times$ fewer FLOPS on images. On videos, compared to some concurrent works [28, 76] that use $90\\%$", + "bbox": [ + 496, + 674, + 893, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "10410", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/cd8793fdca259e545ca2662d09b583b0085b4ddc958153d157c003c3af0ed7bf.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
RatiosAccuracy
IN1KSSv2IN1KSSv2
75%75%82.667.2
75%90%82.867.9
75%95%82.868.1
90%75%82.867.7
90%90%82.668.5
90%95%82.868.6
", + "bbox": [ + 83, + 88, + 225, + 193 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/77d575a72ed00b2c52ebfe651baf85218543cd36f44e86d52c70f84977b3ba01.jpg", + "table_caption": [ + "(a) Masking ratios." + ], + "table_footnote": [], + "table_body": "
TypeAccuracy
IN1KSSv2IN1KSSv2
RandomTube82.867.9
CausalTube82.167.3
RandomFrame82.965.8
RandomCausal82.864.4
RandomRandom82.767.8
", + "bbox": [ + 245, + 104, + 403, + 193 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/544ddd2ec2d533d15491d90ac49e8006f0bb86824f1af01b9b449a64e2505b8b.jpg", + "table_caption": [ + "(b) Masking type." + ], + "table_footnote": [], + "table_body": "
#paramsIN1KSSv2
Common7.9M82.867.9
Separate33.0M82.667.7
L=23.8M82.767.9
L=47.9M82.867.9
L=814.5M82.868.0
d=3847.9M82.867.9
d=51213.0M82.768.1
", + "bbox": [ + 426, + 92, + 584, + 191 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "(c) Decoder capacity.", + "bbox": [ + 450, + 196, + 560, + 207 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1f7c489cc73d09c22ac440fb4ee50ef4e5f412ac731febf69a23e4d158a72d3b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 611, + 90, + 880, + 191 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "(d) Dataset ratio", + "bbox": [ + 704, + 196, + 790, + 205 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Ablations. The default setting for all ablations is highlighted in gray. (a) We vary the masking ratios of the input for pretraining the model. This allows us to use an extremely high masking ratio of $90\\%$ on images and $95\\%$ on videos. (b) OmniMAE works well across different masking types. Random masking makes no assumptions about the input patches. Tube masking masks random patches at the same spatial location across all frames. Causal masking for images masks the 'future' patches as determined by a raster left-to-right order. In videos, causal masking masks all the patches from future frames. Frame masking masks all patches for randomly selected frames. (c) Different decoder designs (common or separate) and capacities (number of layers $L$ and dimension $d$ of the MLP). A common decoder for both images and videos performs better than a separate decoder. Our final performance is robust to decoder capacity. (d) We vary the image/video dataset ratios by replicating the entire datasets by a factor. The number of training updates changes with such replication, and we observe that the model benefits from higher replication of videos.", + "bbox": [ + 75, + 219, + 893, + 345 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "masking, we obtain $1.3 \\times$ , $1.5 \\times$ and $1.4 \\times$ fewer FLOPS for ViT-B/L/H. Given the compute savings and strong performance, we choose $90 \\%$ and $95 \\%$ masking for images and videos respectively for our final models.", + "bbox": [ + 75, + 371, + 468, + 431 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Type of masking. We study the effect of the type of masking used in training our models. The different types of masking are illustrated in Figure 3. We experiment with Random masking which masks patches in the image or video randomly. For videos, we experiment with two types of masking that exploit the temporal structure. In Tube masking, we mask random patches at the same spatial location across all the frames. In Causal masking, we use a raster order masking, akin to generative image models [16, 67], i.e. the patches in the top-left of the image, and earlier frames of the video, are kept while the rest are masked. Finally, in Frame masking, we randomly mask some frames in the video, while keeping all patches from the unmasked frames. As seen in Table 1b, Random and Tube masking perform comparably well on both image and video tasks. We find that in case of Causal masking, the prediction task becomes exceedingly tough due to the uncertainty of the future, and in case of Frame masking, it becomes relatively easy due to the high redundancy of pixel values across frames. Hence in both these cases, the representation learned does not perform as well on video recognition tasks. Given the simplicity of random masking, we use that for both modalities.", + "bbox": [ + 75, + 431, + 470, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Decoder architecture. Since image and video prediction tasks may require specialized parameters, we study two settings: (1) the decoder parameters are shared for image and video pixel prediction; (2) two separate decoders are used for image and video prediction. For the latter we use our default decoder setting for videos (4 layers/384-D), however a larger 8-layer/512-D decoder for images as proposed in MAE [40]. The results in Table 1c show that using a shared decoder for both image and video prediction leads to better", + "bbox": [ + 75, + 763, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "transfer learning accuracy. A shared decoder also offers a $4 \\times$ reduction in the number of parameters compared to using separate decoders while also being simpler to implement.", + "bbox": [ + 496, + 371, + 890, + 416 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Decoder capacity. In Table 1c, we vary the decoder capacity and measure its impact on the final transfer learning accuracy. To change the decoder capacity, we vary the number of Transformer layers $L$ used and the dimension $d$ of the MLP used in the Transformer layers. Overall, the final transfer learning accuracy is robust to decoder capacity. A shallow decoder of 4 layers (8 for ViT-H) offers a good trade-off between the decoder size and final accuracies.", + "bbox": [ + 496, + 417, + 892, + 536 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset ratios. Since ImageNet and SSv2 have a different number of samples, we study the effect of varying the ratio of the datasets used in training. When increasing the ratio of the datasets, we measure a single epoch as training over the new oversampled set of samples from the datasets. We vary the relative ratio for ImageNet and Something Something-v2 by replicating only one dataset at a time as shown in Table 1d. We observe that increasing the relative dataset ratio has a positive effect on the final transfer accuracy for the oversampled dataset. This is expected as oversampling a dataset proportionally increases its corresponding parameter updates, thus making the representation better tuned for the dataset. We also observe that oversampling Something-Something-v2 twice as much as ImageNet leads to an improvement for the video transfer accuracy with no drop in the image transfer performance. Hence, OmniMAE is robust to changes in dataset sizes of individual modalities, and it suggests that longer training on both datasets can further improve the transfer learning performance of our models. For simplicity, by default we do not replicate any dataset for training our models.", + "bbox": [ + 496, + 537, + 893, + 853 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Sample replication. We replicate samples for both images and videos, and in each case measure a single epoch as training over the total samples, counting replicated samples, i.e.", + "bbox": [ + 496, + 854, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "10411", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e0236fdf699dda3c27e0ae084442361871d705c5422b55da07c16974fc4df7b5.jpg", + "image_caption": [ + "Figure 4. Reconstruction visualizations using OmniMAE on different video and image datasets. We show the model predictions for varying masking ratios of the input from $75\\%$ to $95\\%$ and the ground truth reference (Ref). OmniMAE is trained on ImageNet and SSv2 but the predictions generalize to other datasets like K400 and EK100. Please see the supplement for video visualizations." + ], + "image_footnote": [], + "bbox": [ + 71, + 88, + 883, + 652 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "replication maintains the same number of training iterations. We train models with different replication factors and show the final transfer accuracy and the normalized training time in Figs. 5a and 5b. Sample replication leads to faster training while maintaining or even improving the final transfer accuracy. Since a large portion of the input sample is masked, replicating the sample multiple number of times still provides enough learning signal for the model and does not lead to a degradation in performance. This becomes even more relevant for OmniMAE's final settings where we use", + "bbox": [ + 75, + 737, + 472, + 888 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "higher masking ratios. For video data, we use an optimized datalocator (see details in Appendix B) with asynchronous I/O and fast video decoding. Even with this optimized datalocator, sample replication leads to $20\\%$ faster training. We believe this is an important practical observation as video datalugging is often a bottleneck for training models.", + "bbox": [ + 496, + 737, + 893, + 829 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Comparison to Prior Work", + "text_level": 1, + "bbox": [ + 500, + 844, + 745, + 861 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Table 2, we compare OmniMAE's representation on our image and video classification benchmarks with other", + "bbox": [ + 500, + 869, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "10412", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/109784566343738d7688974f7ae5a4d678a8ddc5f23c0528e95fd9fb6a4a1fd0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodArch.Pretrain DataIN1KiNat18P365K400SSv2EK100
DINOViT-BIN1K82.872.6-XXX
iBOT [92]ViT-LIN1K84.8--XXX
MAE [40]ViT-BIN1K83.675.457.9XXX
MAE [40]ViT-LIN1K85.980.159.4XXX
MAE [40]ViT-HIN1K86.983.059.8XXX
BEiT [6]ViT-BIN1K83.472.3-XXX
BEiT [6]ViT-LIN1K85.2--XXX
VIMPAC [74]ViT-L/2†HowTo100MXXX77.468.1-
MaskedFeat [82]MViT-v2-LK400XXX84.3--
BEVT [80]Swin-BIN1K + K400XXX81.171.4-
OmniMAEViT-BIN1K + K40082.974.258.580.869.040.3
OmniMAEViT-BIN1K + SSv283.074.058.480.669.539.3
OmniMAEViT-LIN1K + SSv285.279.659.384.074.245.1
OmniMAEViT-HIN1K + SSv286.683.260.185.475.548.0
Concurrent methods specialized for videos
VideoMAE-16f [76]ViT-LSSv2XXX-74.2-
VideoMAE-16f [76]ViT-LK400XXX84.7--
MAE-Video [28]ViT-HK400XXX85.174.1-
", + "bbox": [ + 217, + 87, + 750, + 353 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 2. Comparing OmniMAE with prior self-supervised methods on image and video recognition. Prior work trains a specialized model for a particular visual modality, sometimes using specialized architectures. Our single OmniMAE model is pretrained jointly on images and videos for 1600 epochs (ViT-H for 2400) and performs competitively across all benchmarks while using a simple architecture and pretraining method, even outperforming concurrent work specialized for videos. ViT-L with half the MLP embedding dimension", + "bbox": [ + 75, + 368, + 893, + 426 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "state-of-the-art self-supervised methods. We focus on methods which use a Transformer backbone, ViT, or an architecture based on Transformers like Swin. To the best of our knowledge, prior work does not explore joint pre-training except for BEVT [80], which does masked image modeling followed by joint masked image and video modeling. Unlike OmniMAE, BEVT only focuses on video representations and does not evaluate on image recognition.", + "bbox": [ + 75, + 450, + 470, + 571 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "OmniMAE is pretrained for 1600 epochs (ViT-H for 2400 epochs). Training with IN1K and K400 or IN1K and SSv2 leads to similar results, so we use the latter when scaling model size. OmniMAE performs competitively for both image and video recognition when compared to the best models trained separately for either modality. OmniMAE also performs favorably compared to methods like MaskedFeat which are pretrained only for video with specialized architectures on larger video datasets. OmniMAE serves as a competitive initialization for transferring models on all modalities. More notably, OmniMAE's performance improves significantly with larger architectures. Using ViT-H, OmniMAE performs better or within the margin of error across all benchmarks. Notably, OmniMAE even outperforms concurrent works which use similar approaches while solely focusing on video representation learning.", + "bbox": [ + 75, + 571, + 470, + 815 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion and Future Work", + "text_level": 1, + "bbox": [ + 76, + 829, + 346, + 845 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present OmniMAE, a unified Transformer for images and videos that can be pretrained using masked autoencoding. OmniMAE uses a simple architecture with minimal", + "bbox": [ + 76, + 854, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "vision-specific changes and is competitive with specialized architectures and models tailored for images and videos. We believe such generic models and approaches are a critical and under explored area of representation learning. OmniMAE has not been empirically validated on other visual modalities like 3D, or non-visual modalities like audio. The simplicity and generality of OmniMAE can likely enable future multimodal systems that use shared parameters to model multiple modalities. We only studied masked autoencoder based pretraining strategies for OmniMAE since they are easier to train. However, pixel reconstruction methods do not learn good linearly separable features or similarity metrics [40]. These properties are essential when finetuning with limited labeled data. We believe that exploring other pretraining strategies, particularly in the context of joint modeling, will likely lead to improved models. Similarly, joint finetuning strategies [33] could be used to evaluate such unified models and lead to potentially even superior performance.", + "bbox": [ + 496, + 450, + 893, + 724 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ethical considerations. We study models for visual recognition on images and videos and our technical contributions are neutral from an ethics standpoint. We do not propose a new dataset or make claims about the suitability of our model on production data. Given their self-supervised nature, our learned representations are free from label bias, however they are still susceptible to the bias in the distribution of the visual data. We believe all ethical considerations that apply to visual recognition models equally apply to our model.", + "bbox": [ + 496, + 763, + 893, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "10413", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 80, + 90, + 171, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Relja Arandjelovic and Andrew Zisserman. Look, listen and learn. In ICCV, 2017.", + "[2] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lucic, and Cordelia Schmid. ViViT: A video vision transformer. In ICCV, 2021.", + "[3] Yuki Markus Asano, Christian Rupprecht, and Andrea Vedaldi. Self-labelling via simultaneous clustering and representation learning. In ICLR, 2020.", + "[4] Sara Atito, Muhammad Awais, and Josef Kittler. Sit: Self-supervised vision transformer. arXiv preprint arXiv:2104.03602, 2021.", + "[5] Alexei Baevski, Wei-Ning Hsu, Qiantong Xu, Arun Babu, Jiatao Gu, and Michael Auli. data2vec: A general framework for self-supervised learning in speech, vision and language. In ICML, 2022.", + "[6] Hangbo Bao, Li Dong, and Furu Wei. BEiT: Bert pre-training of image transformers. In ICLR, 2022.", + "[7] Adrien Bardes, Jean Ponce, and Yann LeCun. Vicreg: Variance-invariance-covariance regularization for self-supervised learning. In ICLR, 2022.", + "[8] Maxim Berman, Hervé Jégou, Andrea Vedaldi, Iasonas Kokkinos, and Matthijs Douze. Multigrain: a unified image embedding for classes and instances. arXiv preprint arXiv:1902.05509, 2019.", + "[9] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In ICML, 2021.", + "[10] Hervé Bourlard and Yves Kamp. Auto-association by multilayer perceptrons and singular value decomposition. Biological Cybernetics, 1988.", + "[11] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features. In ECCV, 2018.", + "[12] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. In NeurIPS, 2020.", + "[13] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021.", + "[14] João Carreira and Andrew Zisserman. Quo vadis, action recognition? A new model and the kinetics dataset. In CVPR, 2017.", + "[15] Lluis Castrejon, Yusuf Aytar, Carl Vondrick, Hamed Piriavash, and Antonio Torralba. Learning aligned cross-modal representations from weakly aligned data. In CVPR, 2016." + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In ICML, 2020.", + "[17] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020.", + "[18] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In CVPR, 2021.", + "[19] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In ICCV, 2021.", + "[20] Kevin Clark, Minh-Thang Luong, Quoc V Le, and Christopher D Manning. Electra: Pre-training text encoders as discriminators rather than generators. *ICLR*, 2020.", + "[21] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In CVPR, 2020.", + "[22] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Antonino Furnari, Evangelos Kazakos, Jian Ma, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, et al. Rescaling egocentric vision. IJCV, 2021.", + "[23] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding. In NAACL, 2018.", + "[24] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021.", + "[25] Alaaeldin El-Nouby, Gautier Izacard, Hugo Touvron, Ivan Laptev, Hervé Jegou, and Edouard Grave. Are large-scale datasets necessary for self-supervised pretraining? arXiv preprint arXiv:2112.10740, 2021.", + "[26] Aleksandr Ermolov, Aliaksandr Siarohin, Enver Sangineto, and Nicu Sebe. Whitening for self-supervised representation learning. In ICML, 2021.", + "[27] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. In ICCV, 2021.", + "[28] Christoph Feichtenhofer, Haoqi Fan, Yanghao Li, and Kaiming He. Masked autoencoders as spatiotemporal learners. arXiv preprint arXiv:2205.09113, 2022.", + "[29] Christoph Feichtenhofer, Axel Pinz, and Andrew Zisserman. Convolutional two-stream network fusion for video action recognition. In CVPR, 2016." + ], + "bbox": [ + 501, + 90, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "10414", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] Patrick Gallinari, Yann LeCun, Sylvie Thiria, and F Fogelman Soulie. Mémoires associatives distribuées: une comparaison (distributed associative memories: a comparison). In COGNITIVA. 1987.", + "[31] Rohit Girdhar, João Carreira, Carl Doersch, and Andrew Zisserman. Video action transformer network. In CVPR, 2019.", + "[32] Rohit Girdhar and Kristen Grauman. Anticipative Video Transformer. In ICCV, 2021.", + "[33] Rohit Girdhar, Mannat Singh, Nikhila Ravi, Laurens van der Maaten, Armand Joulin, and Ishan Misra. Omnivore: A Single Model for Many Visual Modalities. In CVPR, 2022.", + "[34] Yunchao Gong, Liwei Wang, Micah Hodosh, Julia Hockenmaier, and Svetlana Lazebnik. Improving image-sentence embeddings using large weakly annotated photo collections. In ECCV, 2014.", + "[35] Priya Goyal, Quentin Duval, Isaac Seessel, Mathilde Caron, Mannat Singh, Ishan Misra, Levent Sagun, Armand Joulin, and Piotr Bojanowski. Vision models are more robust and fair when pretrained on uncurated images without supervision. arXiv preprint arXiv:2202.08360, 2022.", + "[36] Priya Goyal, Dhruv Mahajan, Abhinav Gupta, and Ishan Misra. Scaling and benchmarking self-supervised visual representation learning. In ICCV, 2019.", + "[37] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, Florian Hoppe, Christian Thurau, Ingo Bax, and Roland Memisevic. The “something something” video database for learning and evaluating visual common sense. In ICCV, 2017.", + "[38] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. NeurIPS, 2020.", + "[39] Raia Hadsell, Sumit Chopra, and Yann LeCun. Dimensionality reduction by learning an invariant mapping. In CVPR, 2006.", + "[40] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, 2022.", + "[41] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, 2020.", + "[42] Geoffrey E Hinton and Richard Zemel. Autoencoders, minimum description length and helmholtz free energy. In NeurIPS, 1993." + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[43] Elad Hoffer, Tal Ben-Nun, Itay Hubara, Niv Giladi, Torsten Hoefler, and Daniel Soudry. Augment your batch: Improving generalization through instance repetition. In CVPR, 2020.", + "[44] Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alex Shepard, Hartwig Adam, Pietro Perona, and Serge Belongie. The inaturalist species classification and detection dataset. In CVPR, 2018.", + "[45] Gao Huang, Yu Sun, Zhuang Liu, Daniel Sedra, and Kilian Q Weinberger. Deep networks with stochastic depth. In ECCV, 2016.", + "[46] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In ICML, 2015.", + "[47] Andrej Karpathy and Li Fei-Fei. Deep visual-semantic alignments for generating image descriptions. In CVPR, 2015.", + "[48] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, AMustafa Suleyman, and Andrew Zisserman. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017.", + "[49] Mark A Kramer. Nonlinear principal component analysis using autoassociative neural networks. AIChE, 1991.", + "[50] Yann LeCun and Françoise Fogelman-Soulie. Modèles connexionnistes de l'apprentissage. Intellectica, 1987.", + "[51] Junnan Li, Pan Zhou, Caiming Xiong, and Steven C.H. Hoi. Prototypical contrastive learning of unsupervised representations. In ICLR, 2021.", + "[52] Yanghao Li, Chao-Yuan Wu, Haoqi Fan, Karttikeya Mangalam, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. Improved multiscale vision transformers for classification and detection. In CVPR, 2022.", + "[53] Valerii Likhosherstov, Anurag Arnab, Krzysztof Choromanski, Mario Lucic, Yi Tay, Adrian Weller, and Mostafa Dehghani. Polyvit: Co-training vision transformers on images, videos and audio. arXiv preprint arXiv:2111.12993, 2021.", + "[54] Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, Furu Wei, and Baining Guo. Swin transformer v2: Scaling up capacity and resolution. In CVPR, 2022.", + "[55] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021.", + "[56] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swin transformer. arXiv preprint arXiv:2106.13230, 2021." + ], + "bbox": [ + 501, + 90, + 893, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10415", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[57] Ze Liu, Zheng Zhang, Yue Cao, Han Hu, and Xin Tong. Group-free 3d object detection via transformers. In ICCV, 2021.", + "[58] Jiasen Lu, Vedanuj Goswami, Marcus Rohrbach, Devi Parikh, and Stefan Lee. 12-in-1: Multi-task vision and language representation learning. In CVPR, 2020.", + "[59] Antoine Miech, Jean-Baptiste Alayrac, Lucas Smaira, Ivan Laptev, Josef Sivic, and Andrew Zisserman. End-to-end learning of visual representations from uncu-rated instructional videos. In CVPR, 2020.", + "[60] Ishan Misra, Rohit Girdhar, and Armand Joulin. An End-to-End Transformer Model for 3D Object Detection. In ICCV, 2021.", + "[61] Ishan Misra and Laurens van der Maaten. Self-supervised learning of pretext-invariant representations. In CVPR, 2020.", + "[62] Pedro Morgado, Ishan Misra, and Nuno Vasconcelos. Robust audio-visual instance discrimination. In CVPR, 2021.", + "[63] Pedro Morgado, Nuno Vasconcelos, and Ishan Misra. Audio-visual instance discrimination with cross-modal agreement. In CVPR, 2021.", + "[64] B. A. Olshausen and D. J. Field. Emergence of simple-cell receptive field properties by learning a sparse code for natural images. Nature, 1996.", + "[65] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. In NeurIPS, 2018.", + "[66] Andrew Owens and Alexei A Efros. Audio-visual scene analysis with self-supervised multisensory features. In ECCV, 2018.", + "[67] Niki Parmar, Ashish Vaswani, Jakob Uszkoreit, Lukasz Kaiser, Noam Shazeer, Alexander Ku, and Dustin Tran. Image transformer. In ICML, 2018.", + "[68] Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A Efros. Context encoders: Feature learning by inpainting. In CVPR, 2016.", + "[69] Boris T Polyak and Anatoli B Juditsky. Acceleration of stochastic approximation by averaging. SIAM journal on control and optimization, 1992.", + "[70] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. IJCV, 2015.", + "[71] Ruslan Salakhutdinov and Geoffrey Hinton. Deep Boltzmann machines. In AISTATS, 2009.", + "[72] Karen Simonyan and Andrew Zisserman. Two-stream convolutional networks for action recognition in videos. In NeurIPS, 2014." + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[73] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In CVPR, 2016.", + "[74] Hao Tan, Jie Lei, Thomas Wolf, and Mohit Bansal. VIMPAC: Video pre-training via masked token prediction and contrastive learning. arXiv preprint arXiv:2106.11250, 2021.", + "[75] Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive multiview coding. arXiv preprint arXiv:1906.05849, 2019.", + "[76] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. arXiv preprint arXiv:2203.12602, 2022.", + "[77] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In ICML, 2021.", + "[78] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017.", + "[79] P. Vincent, H. Larochelle, Y. Bengio, and P.-A. Manzagol. Extracting and composing robust features with denoising autoencoders. In ICML, 2008.", + "[80] Rui Wang, Dongdong Chen, Zuxuan Wu, Yinpeng Chen, Xiyang Dai, Mengchen Liu, Yu-Gang Jiang, Luowei Zhou, and Lu Yuan. BEVT: Bert pretraining of video transformers. In CVPR, 2022.", + "[81] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In ICCV, 2021.", + "[82] Chen Wei, Haoqi Fan, Saining Xie, Chao-Yuan Wu, Alan Yuille, and Christoph Feichtenhofer. Masked feature prediction for self-supervised visual pre-training. In CVPR, 2022.", + "[83] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In CVPR, 2018.", + "[84] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jian-min Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In CVPR, 2022.", + "[85] Xueting Yan, Ishan Misra, Abhinav Gupta, Deepti Ghadiyaram, and Dhruv Mahajan. ClusterFit: Improving Generalization of Visual Representations. In CVPR, 2020." + ], + "bbox": [ + 501, + 90, + 893, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "10416", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[86] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In ICCV, 2019.", + "[87] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stephane Deny. Barlow twins: Self-supervised learning via redundancy reduction. In ICML, 2021.", + "[88] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In ICLR, 2018.", + "[89] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip Torr, and Vladlen Koltun. Point transformer. In ICCV, 2021.", + "[90] Zhun Zhong, Liang Zheng, Guoliang Kang, Shaozi Li, and Yi Yang. Random erasing data augmentation. In AAAI, 2020.", + "[91] Bolei Zhou, Agata Lapedriza, Jianxiong Xiao, Antonio Torralba, and Aude Oliva. Learning deep features for scene recognition using places database. In NeurIPS, 2014.", + "[92] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. ibot: Image bert pre-training with online tokenizer. In ICLR, 2022." + ], + "bbox": [ + 78, + 90, + 470, + 445 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "10417", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2023/OmniMAE_ Single Model Masked Pretraining on Images and Videos/2fdd9b20-eaac-4038-a247-56813890042d_model.json b/2023/OmniMAE_ Single Model Masked Pretraining on Images and Videos/2fdd9b20-eaac-4038-a247-56813890042d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f771fae1d0cb259c826c13b3f0333558b240b1ad --- /dev/null +++ b/2023/OmniMAE_ Single Model Masked Pretraining on Images and Videos/2fdd9b20-eaac-4038-a247-56813890042d_model.json @@ -0,0 +1,2303 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.139, + 0.131, + 0.833, + 0.154 + ], + "angle": 0, + "content": "OmniMAE: Single Model Masked Pretraining on Images and Videos" + }, + { + "type": "text", + "bbox": [ + 0.254, + 0.181, + 0.714, + 0.233 + ], + "angle": 0, + "content": "Rohit Girdhar* Alaaeldin El-Nouby* Mannat Singh* \nKalyan Vasudev Alwala* Armand Joulin Ishan Misra* \nFAIR, Meta AI" + }, + { + "type": "text", + "bbox": [ + 0.333, + 0.238, + 0.631, + 0.251 + ], + "angle": 0, + "content": "https://github.com/facebookresearch/omnivore" + }, + { + "type": "image", + "bbox": [ + 0.119, + 0.285, + 0.852, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.474, + 0.896, + 0.545 + ], + "angle": 0, + "content": "Figure 1. OmniMAE is a single model for images and videos that is trained using masked autoencoding [40]. We use a plain Vision Transformer [24] architecture but with spatio-temporal patches as input. At training, we 'patchify' the visual input (images or videos), and feed the encoder only a subset of the patches. The decoder reconstructs the pixels for the missing patches using the encoder's output. The encoder-decoder model is trained using a pixel reconstruction loss. After training, our single plain Transformer encoder performs competitively compared to specialized architectures on downstream image and video recognition tasks." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.557, + 0.314, + 0.572 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.589, + 0.474, + 0.876 + ], + "angle": 0, + "content": "Transformer-based architectures have become competitive across a variety of visual domains, most notably images and videos. While prior work studies these modalities in isolation, having a common architecture suggests that one can train a single unified model for multiple visual modalities. Prior attempts at unified modeling typically use architectures tailored for vision tasks, or obtain worse performance compared to single modality models. In this work, we show that masked autoencoding can be used to train a simple Vision Transformer on images and videos, without requiring any labeled data. This single model learns visual representations that are comparable to or better than single-modality representations on both image and video benchmarks, while using a much simpler architecture. Furthermore, this model can be learned by dropping \\(90\\%\\) of the image and \\(95\\%\\) of the video patches, enabling extremely fast training of huge model architectures. In particular, we show that our single ViT-Huge model can be finetuned to achieve \\(86.6\\%\\) on ImageNet and \\(75.5\\%\\) on the challenging Something Something-v2 video" + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.558, + 0.778, + 0.573 + ], + "angle": 0, + "content": "benchmark, setting a new state-of-the-art." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.603, + 0.631, + 0.619 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.628, + 0.896, + 0.869 + ], + "angle": 0, + "content": "The Transformer architecture [78] is rapidly becoming competitive across the different visual modalities in Computer Vision, from images [24, 27, 55, 77], to 3D [57, 60, 89] and videos [2, 9, 27, 31, 32, 56]. This convergence toward a unified architecture naturally suggests that we should be able to train a single model that works across different visual modalities. However, recent attempts to train unified models either lead to worse performance compared to single modality models [53], or require the use of an alternative architecture [33], namely the Swin Transformer [55], with inductive biases tailored towards vision tasks. While specialized Transformer architectures for vision [27, 55, 56, 81] can offer better performance for visual modalities, they lose the generality and flexibility of the vanilla Transformer, making it harder to later model different domains like text, speech, 3D etc. in multi-modal architectures." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.895, + 0.902 + ], + "angle": 0, + "content": "In this work, we train a single vanilla Transformer that works for both images and videos, as illustrated in Figure 1." + }, + { + "type": "page_footnote", + "bbox": [ + 0.102, + 0.888, + 0.265, + 0.901 + ], + "angle": 0, + "content": "*Equal technical contribution." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "10406" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.258 + ], + "angle": 0, + "content": "To this end, we leverage the findings of several recent works on the use of the masked pretraining [23] to greatly improve the training and performance of Transformers in the domain of images [6, 40, 80, 84], videos [76, 80, 82] or across text, audio and images [5]. We show that this masked pretraining is a viable strategy to pretrain a unified 'omnivorous' Transformer across visual modalities. In particular, we consider the Masked Auto-Encoding (MAE) approach [40] to train an Omnivorous visual encoder [33]. The resulting OmniMAE model learns from all the modalities with the same objective function and does not require any supervision." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.27, + 0.473, + 0.497 + ], + "angle": 0, + "content": "Using a masked pretraining objective has several advantages over supervised objectives [33, 53] or discriminative self-supervised objectives [13, 17, 41]. First, as opposed to supervised objectives, a general-purpose unsupervised loss does not require any human labeling effort. As a result, it is robust to biases introduced by a predefined set of labels [35]. Moreover, it does not require a multi-head architecture to incorporate supervision from each of the label spaces corresponding to each modality, which is hard to maintain and scale with new modalities. Second, although discriminative self-supervised methods produce superior frozen features compared to reconstruction objectives, they are non trivial to scale in model and data size [36]. Our masked pretraining objective is simple, efficient to train, and scales to different visual modalities with minimal changes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.508, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Contributions. (1) We show that the simple Vision Transformer architecture (ViT) [24] originally designed for images can naturally be applied on videos, and videos and images jointly. OmniMAE is a single ViT-based model for videos and images that outperforms architectures and models specifically designed for either modality. (2) Prior and concurrent work design self-supervised methods and architectures for either image or video and we find that these models do not transfer well across modalities. OmniMAE is the first single self-supervised model that achieves good performance on both modalities. (3) We show that our joint training using both images and videos enables us to use much higher masking ratios than any prior work for training MAE. Since ViT can processes only the non-masked input, we train OmniMAE models with only \\(10\\%\\) of image and \\(5\\%\\) of video patches. This enables us to train large (650M parameter) models with a \\(\\sim 7\\times\\) and \\(\\sim 11\\times\\) reduction in compute and memory on images and videos. (4) Finally, we propose improvements to the MAE training. We show that repeating samples in a mini-batch reduces dataloading (and thus training) time without loss in final transfer performance. Sample replication is particularly useful for masked pretraining as the unmasked patches are different across sample replicas. We also show that using a shallow shared decoder for videos and images leads to better performance while reducing the number of parameters by \\(2 - 4\\times\\)." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.09, + 0.642, + 0.107 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.116, + 0.895, + 0.342 + ], + "angle": 0, + "content": "Our work builds upon research in self-supervised learning, masked pretraining and unified modeling in computer vision. Self-supervised learning. In recent years, self-supervised approaches have been dominated by joint embedding methods which can rely on different objectives including contrastive [17,39,41,51,61,65,75,83], non-contrastive [7,18,26, 87], clustering [3, 11, 12, 85] or self-distillation [5, 13, 38, 92]. Such methods are trained to learn invariance to a set of predefined transformations which results in image descriptors with a strong linear probing and KNN performance. However, such methods can be challenging to scale since they can suffer from instabilities [19]. Additionally, the strongest performance is typically obtained with the help of augmentations like multi-crop [12, 13, 92] which can be hard to apply at scale due their compute and memory overhead." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.342, + 0.897, + 0.794 + ], + "angle": 0, + "content": "Masked pretraining. We build upon masked prediction methods where the representation is learned by predicting masked parts of the input. Such methods have recently gained popularity given their immense success in NLP. In particular, BERT [23] showed that masked language modeling by predicting a subset of masked words is a powerful pre-training objective and leads to impressive finetuning performance on various downstream tasks. In computer vision, input reconstruction methods have a rich history with non-linear PCA [49], sparse reconstruction [64], autoencoders [10, 30, 42, 50], RBMs [71] etc. Masked prediction methods can be viewed as a special case of denoising autoencoders [30, 79] where the input 'noise' is a masking function. An example of such a method that uses masking as noise is context encoders [68]. With the recent and rapid rise of Vision Transformers [24], masked prediction was revisited by multiple efforts. SiT [4] replaces variable sized patches in the image with random noise and trains a ViT model for reconstruction, among other objectives. BEiT [6] moves a step closer to BERT with replacing full patches with mask tokens and training a ViT encoder to predict the discrete visual words of masked patches using a cross-entropy loss. Masked prediction has also shown impressive performance for specialized vision transformer architectures such as MViT [27, 52, 82] and Swin transformer [54, 55, 84]. SimMIM [84] and MaskedFeat [82] predict pixel values and HOG features of the masked patches using a Swin-v2 [54] and MViT-v2 [52] backbones respectively. Finally, SplitMask [25] studied the interesting properties of masked prediction methods in terms of high sample efficiency." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Of particular interest to OmniMAE, masked autoencoders (MAE) [40] demonstrated impressive scaling properties by utilizing a patch dropping strategy for masked patches accompanied with a high masking ratio of \\(75\\%\\). Under this setting, the encoder only process a small subset of the image patches followed by a relatively low capacity decoder which reconstructs the image pixels. This property is even more" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "10407" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.288 + ], + "angle": 0, + "content": "crucial for video representation learning given the large number of patches, and a concurrent work, ST-MAE [76], shows that MAE pretraining with an even higher masking ratio of \\(90\\%\\) works well and obtains strong finetuning performance on downstream video recognition benchmarks. Notably, the efficiency gained by patch dropping is specific to vanilla ViTs, and multiscale architectures such as MViT and Swin are unable to benefit due to their design. Hence, we use the simple ViT as our architecture and show that it can be trained efficiently and jointly for images and videos using extremely high masking ratios (90-95% on both modalities), and yet perform competitively to specialized architectures like MViT [27] and Swin [55]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.3, + 0.473, + 0.68 + ], + "angle": 0, + "content": "Unified modeling and multi-modal learning. Multi-modal learning in computer vision has a long history that includes training using images and text [15, 34, 47, 58, 59], video and optical flow [29, 72], and video and audio [1, 62, 63, 66]. The majority of such methods rely on training a separate backbone for each modality as well as the availability of alignment across modalities. More recently, Omnivore [33] was proposed for joint training of multiple modalities like images, videos and single-view 3D, attaining a strong performance in each of the modality specific benchmarks with a single shared trunk. PolyViT [53] co-trains a shared transformer encoder using images, videos and audio data and provides a competitive performance on various downstream tasks for each modality. The aforementioned methods differ compared to OmniMAE in that they are trained with supervised learning and require human annotations. BEVT [80] tackles BERT pre-training for videos and proposes that jointly pretraining using static images improves the finetuning performance of video recognition benchmarks. Unlike OmniMAE, BEVT uses the specialized Swin transformer architecture with separate decoder heads for images and videos. Thus, BEVT cannot drop patches while training which can limit its scalability. Furthermore, it relies on a tokenizer which must be trained apriori, and the tokenizer training itself can affect the model's performance." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.728, + 0.196, + 0.743 + ], + "angle": 0, + "content": "3. OmniMAE" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.472, + 0.901 + ], + "angle": 0, + "content": "Our goal is to pretrain a single unified model for images and videos. Rather than use specialized architectures tailored for a visual modality, we build upon the vanilla Vision Transformer (ViT) [24] architecture that has limited inductive biases for vision. For pretraining, we extend the simple self-supervised masked auto-encoding (MAE) approach [40]. The original architecture and pretraining method are tested only on images, and we show simple design decisions for a unified model." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.091, + 0.892, + 0.121 + ], + "angle": 0, + "content": "3.1. Training OmniMAE jointly on images and videos" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.129, + 0.894, + 0.266 + ], + "angle": 0, + "content": "We illustrate our method in Figure 1. For pretraining, we use an encoder-decoder architecture where the encoder only operates on a 'non-masked' subset of the input. The decoder predicts the pixel values for the entire input, i.e., masked and non-masked pixels. The model is trained to minimize the reconstruction error for the masked (unseen) part of the input. After pretraining, we evaluate the encoder by transfer learning (the decoder is discarded). Next, we describe the pretraining details." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.266, + 0.895, + 0.37 + ], + "angle": 0, + "content": "Images and videos as spatio-temporal patches. The input image or video can be represented as a 4D tensor of shape \\( T \\times H \\times W \\times 3 \\) where \\( T \\) is the temporal dimension and \\( H, W \\) are the spatial dimensions, and 3 represents the color channels. We treat images as being single-frame videos with \\( T = 1 \\). The input is split into \\( N \\) spatio-temporal patches, each of size \\( t \\times h \\times w \\times 3 \\) [33]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.371, + 0.895, + 0.508 + ], + "angle": 0, + "content": "Omnivorous visual encoder. We use an omnivorous [33] visual encoder that processes both images and video using the same parameters. The encoder operates on the \\(N\\) spatiotemporal patches from the images and videos. The encoder can naturally handle variable number \\(N\\) of patches from images and videos as it uses the Transformer architecture [78]. The encoder shares the same parameters for both image and video inputs and processes them in the same way to output per-patch embeddings." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.508, + 0.894, + 0.629 + ], + "angle": 0, + "content": "Pretraining. We convert the input into \\( N \\) spatio-temporal patches and randomly mask \\( M \\) of them. Following [40], we feed the non-masked \\( N - M \\) patches (and their positions) to the encoder to produce per-patch embeddings. The encoder output is concatenated with \\( M \\) copies of a learnable 'mask token' to obtain \\( N \\) embeddings. We add the appropriate positional encoding to the \\( N \\) embeddings and feed them into the decoder, which outputs \\( N \\times t \\times h \\times w \\times 3 \\) pixel values." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.63, + 0.894, + 0.78 + ], + "angle": 0, + "content": "Loss function and optimization. We minimize the reconstruction error between the decoder predictions and the input pixel values. The input pixel values are normalized to zero mean and unit variance [46] to get the targets for the loss. We minimize the \\(\\ell_2\\) distance between the predictions and targets over the \\(M\\) masked patches. At each point in the training, we sample a mini-batch of either images or video and compute the loss using the decoder predictions. We study the effect of different mini-batch construction strategies on the overall performance and speed of training in § 4.2." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Dataset sample replication. Since we operate on a small subset of the input patches \\((M \\ll N)\\), we find that our data reading speed is unable to keep up with the number of data points we can theoretically process in a single optimization step. This issue is even more pronounced for videos, where we spend significant compute to read and decode a full video, only to discard \\(>90\\%\\) of it. Inspired from the success of RepeatAugment, which shows that replicating samples" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10408" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.085, + 0.087, + 0.891, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.243, + 0.896, + 0.314 + ], + "angle": 0, + "content": "Figure 2. OmniMAE on image and video downstream tasks. We finetune the MAE, ST-MAE, and OmniMAE models on image and video benchmarks. We use the ViT architecture with two model sizes: ViT-B and ViT-L. MAE has poor video recognition performance while ST-MAE's performance drops on image datasets. OmniMAE pretraining generalizes to both benchmarks. All models are trained for 800 epochs on the pretraining datasets. The image-only MAE model is inflated [14] to apply MAE to video recognition tasks. The input image is replicated to apply ST-MAE to image recognition benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.339, + 0.475, + 0.448 + ], + "angle": 0, + "content": "within a mini-batch is an effective technique to improve generalization [8, 12, 43], we replicate a single data point and apply different masks to it each time. Even with replicated samples, the non-masked input to the model is different due to random cropping and masking being different across the replicas. We show in § 4.2 that sample replication leads to gains in runtime without affecting the final transfer accuracy." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.456, + 0.295, + 0.472 + ], + "angle": 0, + "content": "3.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.48, + 0.471, + 0.51 + ], + "angle": 0, + "content": "We note some of the salient implementation details and provide the complete details in Appendix B." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.511, + 0.47, + 0.661 + ], + "angle": 0, + "content": "Architecture. We use the ViT [24] architecture for the omnivorous encoder and experiment with its ViT-B, ViT-L, and ViT-H variants. We do not use the [CLS] token in the ViT models, yielding a small improvement in runtime without loss in performance. We use a Transformer decoder with 4 layers (8 for ViT-H) of 384, 512, and 512 embedding dimensions for ViT-B, ViT-L, and ViT-H, respectively. The decoder outputs the RGB colors for the pixels in all the input patches. We use sinusoidal positional encoding [78] for the patches in both the encoder and the decoder." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.663, + 0.47, + 0.754 + ], + "angle": 0, + "content": "Training details. We train our models with a mini-batch size of 2048. We resize the input images and videos spatially to \\(224 \\times 224\\) pixels. For videos, we sample a clip of \\(T = 16\\) frames at 6 FPS. We use a patch size of \\(2 \\times 16 \\times 16\\) for ViT-B and ViT-L, and \\(2 \\times 14 \\times 14\\) for ViT-H. Images are replicated temporally to meet the patch size." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.754, + 0.471, + 0.83 + ], + "angle": 0, + "content": "Masking the input. Compared to prior work, we use extremely high masking for pretraining and only \\(10\\%\\) and \\(5\\%\\) of the image and video patches are fed to the encoder for pretraining. We uniform randomly mask the patches for images and videos and ablate the masking hyperparameters in \\(\\S 4.2\\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.845, + 0.21, + 0.862 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Pretraining data. We pretrain representations by jointly training on images from the ImageNet (IN1K) [70] dataset" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.339, + 0.893, + 0.4 + ], + "angle": 0, + "content": "and videos from the Something Something-v2 (SSv2) [37] dataset. We choose the SSv2 dataset over web video datasets like Kinetics-400 (K400) [48] for reproducibility, and given SSv2's challenging nature requiring temporal reasoning." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.4, + 0.895, + 0.536 + ], + "angle": 0, + "content": "Transfer learning datasets. We consider two image datasets: iNaturalist-2018 (iNat18) [44], a popular fine-grained recognition dataset, and Places-365 (P365) [91], a scene recognition benchmark. For video datasets, we focus on the popular K400 action recognition benchmark, and EPIC-Kitchens-100 (EK100) [22], an egocentric action recognition dataset. We report the top-1 classification accuracy for all transfer datasets. The details about the datasets can be found in Appendix A." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.545, + 0.878, + 0.562 + ], + "angle": 0, + "content": "4.1. OmniMAE on Image and Video Recognition" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.569, + 0.895, + 0.599 + ], + "angle": 0, + "content": "We now evaluate the capabilities of OmniMAE's representations on both image and video recognition tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.6, + 0.895, + 0.719 + ], + "angle": 0, + "content": "Baselines. We compare OmniMAE's representations trained jointly on images and videos to MAE, trained solely on images [40]. Additionally, we develop a video-only baseline, SpatioTemporal-MAE (ST-MAE), by training OmniMAE only on videos. For a fair comparison, all models are pretrained for 800 epochs on their respective datasets. The image-only MAE model is inflated [14] for finetuning on video tasks, see Appendix B.2 for details." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.896, + 0.902 + ], + "angle": 0, + "content": "Observations. Figure 2 presents the evaluation results for all the models on the image and video benchmarks. The modality specific MAE and ST-MAE models achieve good transfer performance when transferring to datasets which match the pretraining modality. However, both models show a degradation in performance when transferring to a different visual modality. MAE has lower video recognition performance, especially on egocentric videos from EK100 where the smaller ViT-B MAE model is \\(18.4\\%\\) worse compared to ST-MAE. Similarly, compared to MAE, ST-MAE is \\(8.7\\%\\) worse on the fine-grained classification task of iNat18. With a large model, MAE's performance on the image recognition" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "10409" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.119, + 0.088, + 0.237, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.244, + 0.089, + 0.359, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.371, + 0.089, + 0.482, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.489, + 0.089, + 0.598, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.605, + 0.089, + 0.715, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.727, + 0.089, + 0.848, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.205, + 0.894, + 0.233 + ], + "angle": 0, + "content": "Figure 3. Different types of masking for images (left two) and videos. Causal and tube masking use the data's spatio-temporal structure. Random frame masking randomly masks frames in a video. Random masking randomly masks patches and is used by default for OmniMAE." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.26, + 0.469, + 0.304 + ], + "angle": 0, + "content": "benchmarks improves but the cross-modality performance degrades further. On EPIC-Kitchens-100 the large MAE model is more than \\(25\\%\\) worse than ST-MAE." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.308, + 0.47, + 0.489 + ], + "angle": 0, + "content": "When transferred to both image and video recognition tasks, OmniMAE performs favorably to the single modality baselines that use exactly the same architecture. OmniMAE also uses the same finetuning recipe as the single modality baselines. OmniMAE matches the video classification performance of ST-MAE and the image classification performance of MAE for both ViT-B and ViT-L, with OmniMAE's performance improving on both image and video recognition benchmarks with the larger model. These results suggest that our unified model of images and videos serves as a better pretraining across a wider variety of recognition tasks than single modality models." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.492, + 0.473, + 0.675 + ], + "angle": 0, + "content": "Qualitative results. Following [40], we re-train OmniMAE without normalizing the pixel targets to obtain easy to visualize RGB reconstructions. We visualize the predictions in Figure 4 using samples that are unseen during training: either val sets or different datasets altogether. OmniMAE makes reasonable predictions on the in-distribution ImageNet and SSv2 val sets, as well as the unseen, out-of-distribution K400 and EK100 datasets. As expected, the details in the reconstruction decrease when increasing the masking ratio. However, even with \\(95\\%\\) masking, the model can reconstruct coarse details in the input, e.g., in ImageNet, the model reconstructs the coarse structure of the flower, vehicle, dog etc." + }, + { + "type": "image", + "bbox": [ + 0.085, + 0.719, + 0.268, + 0.816 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.722, + 0.458, + 0.815 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.827, + 0.47, + 0.897 + ], + "angle": 0, + "content": "Figure 5. Sample replication. We study the effect of repeating samples while training our model. In each case, we repeat a sample \\( n \\) times within a mini-batch while fixing the overall mini-batch size and training updates. Replication leads to improved training speeds, especially on video without affecting the final performance." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.259, + 0.704, + 0.275 + ], + "angle": 0, + "content": "4.2. Analyzing OmniMAE" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.287, + 0.895, + 0.468 + ], + "angle": 0, + "content": "We train the ViT-B architecture jointly on the ImageNet and the SSv2 datasets for 800 epochs, where an epoch involves training over all samples from both the datasets. For these analysis experiments, we use the default masking hyperparameters of MAE for images (75%). For ST-MAE, due to redundancy across frames, we use \"tube\" dropping (i.e. dropping all patches at a given spatial location over time) with high masking ratio (90%) All hyperparameters for these experiments are in Appendix B.3. We evaluate all the pretrained models on ImageNet and SSv2 and report the top-1 classification accuracies in Table 1. Please see Appendix D for additional ablations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.473, + 0.895, + 0.67 + ], + "angle": 0, + "content": "Extreme masking. We vary the ratio of masked patches in the input while training the model. We observe that videos benefit from significantly higher amounts of masking than images. Since video frames have a lot of redundant information, the resulting spatio-temporal patches are also redundant and thus higher masking leads to better self-supervised models. Unlike our experiments, prior work found that extremely high masking ratios lead to degradation in performance, for instance MAE [40] saw a significant degradation in performance when masking more than \\(75\\%\\) of the patches. However, as we show in Table 1a, OmniMAE models can use extremely high masking of the input while learning good representations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Reduced compute due to extreme masking. Our models trained with \\(90\\%\\) masking on images and \\(95\\%\\) masking on videos yield good performance, while being trained with just 19 and 78 unmasked image and video patches respectively, for a patch size of \\(2 \\times 16 \\times 16\\). As we follow [40] and only pass unmasked patches to the encoder and have a lightweight decoder, extreme masking leads to a dramatically lower computational cost for training the encoder, and consequently the model as a whole. Compared to using all the patches, our masked autoencoder uses \\(5.9 \\times\\) and \\(7.8 \\times\\) fewer FLOPS for ViT-B on images and videos respectively, \\(7.1 \\times\\) and \\(11.6 \\times\\) for ViT-L, and \\(7.2 \\times\\) and \\(11.3 \\times\\) for ViT-H. Compared to MAE, on ViT-B, ViT-L and ViT-H, our higher masking leads to \\(1.8 \\times\\), \\(2.0 \\times\\) and \\(2.0 \\times\\) fewer FLOPS on images. On videos, compared to some concurrent works [28, 76] that use \\(90\\%\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10410" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.084, + 0.089, + 0.226, + 0.194 + ], + "angle": 0, + "content": "
RatiosAccuracy
IN1KSSv2IN1KSSv2
75%75%82.667.2
75%90%82.867.9
75%95%82.868.1
90%75%82.867.7
90%90%82.668.5
90%95%82.868.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.103, + 0.197, + 0.203, + 0.208 + ], + "angle": 0, + "content": "(a) Masking ratios." + }, + { + "type": "table", + "bbox": [ + 0.246, + 0.105, + 0.404, + 0.194 + ], + "angle": 0, + "content": "
TypeAccuracy
IN1KSSv2IN1KSSv2
RandomTube82.867.9
CausalTube82.167.3
RandomFrame82.965.8
RandomCausal82.864.4
RandomRandom82.767.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.279, + 0.197, + 0.373, + 0.208 + ], + "angle": 0, + "content": "(b) Masking type." + }, + { + "type": "table", + "bbox": [ + 0.427, + 0.093, + 0.586, + 0.193 + ], + "angle": 0, + "content": "
#paramsIN1KSSv2
Common7.9M82.867.9
Separate33.0M82.667.7
L=23.8M82.767.9
L=47.9M82.867.9
L=814.5M82.868.0
d=3847.9M82.867.9
d=51213.0M82.768.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.452, + 0.197, + 0.562, + 0.208 + ], + "angle": 0, + "content": "(c) Decoder capacity." + }, + { + "type": "image", + "bbox": [ + 0.612, + 0.092, + 0.882, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "table_caption", + "bbox": [ + 0.705, + 0.197, + 0.792, + 0.207 + ], + "angle": 0, + "content": "(d) Dataset ratio" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.22, + 0.894, + 0.346 + ], + "angle": 0, + "content": "Table 1. Ablations. The default setting for all ablations is highlighted in gray. (a) We vary the masking ratios of the input for pretraining the model. This allows us to use an extremely high masking ratio of \\(90\\%\\) on images and \\(95\\%\\) on videos. (b) OmniMAE works well across different masking types. Random masking makes no assumptions about the input patches. Tube masking masks random patches at the same spatial location across all frames. Causal masking for images masks the 'future' patches as determined by a raster left-to-right order. In videos, causal masking masks all the patches from future frames. Frame masking masks all patches for randomly selected frames. (c) Different decoder designs (common or separate) and capacities (number of layers \\(L\\) and dimension \\(d\\) of the MLP). A common decoder for both images and videos performs better than a separate decoder. Our final performance is robust to decoder capacity. (d) We vary the image/video dataset ratios by replicating the entire datasets by a factor. The number of training updates changes with such replication, and we observe that the model benefits from higher replication of videos." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.372, + 0.47, + 0.432 + ], + "angle": 0, + "content": "masking, we obtain \\(1.3 \\times\\), \\(1.5 \\times\\) and \\(1.4 \\times\\) fewer FLOPS for ViT-B/L/H. Given the compute savings and strong performance, we choose \\(90 \\%\\) and \\(95 \\%\\) masking for images and videos respectively for our final models." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.433, + 0.471, + 0.764 + ], + "angle": 0, + "content": "Type of masking. We study the effect of the type of masking used in training our models. The different types of masking are illustrated in Figure 3. We experiment with Random masking which masks patches in the image or video randomly. For videos, we experiment with two types of masking that exploit the temporal structure. In Tube masking, we mask random patches at the same spatial location across all the frames. In Causal masking, we use a raster order masking, akin to generative image models [16, 67], i.e. the patches in the top-left of the image, and earlier frames of the video, are kept while the rest are masked. Finally, in Frame masking, we randomly mask some frames in the video, while keeping all patches from the unmasked frames. As seen in Table 1b, Random and Tube masking perform comparably well on both image and video tasks. We find that in case of Causal masking, the prediction task becomes exceedingly tough due to the uncertainty of the future, and in case of Frame masking, it becomes relatively easy due to the high redundancy of pixel values across frames. Hence in both these cases, the representation learned does not perform as well on video recognition tasks. Given the simplicity of random masking, we use that for both modalities." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Decoder architecture. Since image and video prediction tasks may require specialized parameters, we study two settings: (1) the decoder parameters are shared for image and video pixel prediction; (2) two separate decoders are used for image and video prediction. For the latter we use our default decoder setting for videos (4 layers/384-D), however a larger 8-layer/512-D decoder for images as proposed in MAE [40]. The results in Table 1c show that using a shared decoder for both image and video prediction leads to better" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.372, + 0.892, + 0.417 + ], + "angle": 0, + "content": "transfer learning accuracy. A shared decoder also offers a \\(4 \\times\\) reduction in the number of parameters compared to using separate decoders while also being simpler to implement." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.418, + 0.893, + 0.537 + ], + "angle": 0, + "content": "Decoder capacity. In Table 1c, we vary the decoder capacity and measure its impact on the final transfer learning accuracy. To change the decoder capacity, we vary the number of Transformer layers \\( L \\) used and the dimension \\( d \\) of the MLP used in the Transformer layers. Overall, the final transfer learning accuracy is robust to decoder capacity. A shallow decoder of 4 layers (8 for ViT-H) offers a good trade-off between the decoder size and final accuracies." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.539, + 0.895, + 0.854 + ], + "angle": 0, + "content": "Dataset ratios. Since ImageNet and SSv2 have a different number of samples, we study the effect of varying the ratio of the datasets used in training. When increasing the ratio of the datasets, we measure a single epoch as training over the new oversampled set of samples from the datasets. We vary the relative ratio for ImageNet and Something Something-v2 by replicating only one dataset at a time as shown in Table 1d. We observe that increasing the relative dataset ratio has a positive effect on the final transfer accuracy for the oversampled dataset. This is expected as oversampling a dataset proportionally increases its corresponding parameter updates, thus making the representation better tuned for the dataset. We also observe that oversampling Something-Something-v2 twice as much as ImageNet leads to an improvement for the video transfer accuracy with no drop in the image transfer performance. Hence, OmniMAE is robust to changes in dataset sizes of individual modalities, and it suggests that longer training on both datasets can further improve the transfer learning performance of our models. For simplicity, by default we do not replicate any dataset for training our models." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Sample replication. We replicate samples for both images and videos, and in each case measure a single epoch as training over the total samples, counting replicated samples, i.e." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "10411" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.073, + 0.089, + 0.885, + 0.653 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.669, + 0.895, + 0.711 + ], + "angle": 0, + "content": "Figure 4. Reconstruction visualizations using OmniMAE on different video and image datasets. We show the model predictions for varying masking ratios of the input from \\(75\\%\\) to \\(95\\%\\) and the ground truth reference (Ref). OmniMAE is trained on ImageNet and SSv2 but the predictions generalize to other datasets like K400 and EK100. Please see the supplement for video visualizations." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.738, + 0.473, + 0.89 + ], + "angle": 0, + "content": "replication maintains the same number of training iterations. We train models with different replication factors and show the final transfer accuracy and the normalized training time in Figs. 5a and 5b. Sample replication leads to faster training while maintaining or even improving the final transfer accuracy. Since a large portion of the input sample is masked, replicating the sample multiple number of times still provides enough learning signal for the model and does not lead to a degradation in performance. This becomes even more relevant for OmniMAE's final settings where we use" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.738, + 0.895, + 0.83 + ], + "angle": 0, + "content": "higher masking ratios. For video data, we use an optimized datalocator (see details in Appendix B) with asynchronous I/O and fast video decoding. Even with this optimized datalocator, sample replication leads to \\(20\\%\\) faster training. We believe this is an important practical observation as video datalugging is often a bottleneck for training models." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.845, + 0.746, + 0.862 + ], + "angle": 0, + "content": "4.3. Comparison to Prior Work" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.895, + 0.903 + ], + "angle": 0, + "content": "In Table 2, we compare OmniMAE's representation on our image and video classification benchmarks with other" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "10412" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.218, + 0.088, + 0.75, + 0.354 + ], + "angle": 0, + "content": "
MethodArch.Pretrain DataIN1KiNat18P365K400SSv2EK100
DINOViT-BIN1K82.872.6-XXX
iBOT [92]ViT-LIN1K84.8--XXX
MAE [40]ViT-BIN1K83.675.457.9XXX
MAE [40]ViT-LIN1K85.980.159.4XXX
MAE [40]ViT-HIN1K86.983.059.8XXX
BEiT [6]ViT-BIN1K83.472.3-XXX
BEiT [6]ViT-LIN1K85.2--XXX
VIMPAC [74]ViT-L/2†HowTo100MXXX77.468.1-
MaskedFeat [82]MViT-v2-LK400XXX84.3--
BEVT [80]Swin-BIN1K + K400XXX81.171.4-
OmniMAEViT-BIN1K + K40082.974.258.580.869.040.3
OmniMAEViT-BIN1K + SSv283.074.058.480.669.539.3
OmniMAEViT-LIN1K + SSv285.279.659.384.074.245.1
OmniMAEViT-HIN1K + SSv286.683.260.185.475.548.0
Concurrent methods specialized for videos
VideoMAE-16f [76]ViT-LSSv2XXX-74.2-
VideoMAE-16f [76]ViT-LK400XXX84.7--
MAE-Video [28]ViT-HK400XXX85.174.1-
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.369, + 0.895, + 0.427 + ], + "angle": 0, + "content": "Table 2. Comparing OmniMAE with prior self-supervised methods on image and video recognition. Prior work trains a specialized model for a particular visual modality, sometimes using specialized architectures. Our single OmniMAE model is pretrained jointly on images and videos for 1600 epochs (ViT-H for 2400) and performs competitively across all benchmarks while using a simple architecture and pretraining method, even outperforming concurrent work specialized for videos. ViT-L with half the MLP embedding dimension" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.451, + 0.471, + 0.572 + ], + "angle": 0, + "content": "state-of-the-art self-supervised methods. We focus on methods which use a Transformer backbone, ViT, or an architecture based on Transformers like Swin. To the best of our knowledge, prior work does not explore joint pre-training except for BEVT [80], which does masked image modeling followed by joint masked image and video modeling. Unlike OmniMAE, BEVT only focuses on video representations and does not evaluate on image recognition." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.573, + 0.472, + 0.816 + ], + "angle": 0, + "content": "OmniMAE is pretrained for 1600 epochs (ViT-H for 2400 epochs). Training with IN1K and K400 or IN1K and SSv2 leads to similar results, so we use the latter when scaling model size. OmniMAE performs competitively for both image and video recognition when compared to the best models trained separately for either modality. OmniMAE also performs favorably compared to methods like MaskedFeat which are pretrained only for video with specialized architectures on larger video datasets. OmniMAE serves as a competitive initialization for transferring models on all modalities. More notably, OmniMAE's performance improves significantly with larger architectures. Using ViT-H, OmniMAE performs better or within the margin of error across all benchmarks. Notably, OmniMAE even outperforms concurrent works which use similar approaches while solely focusing on video representation learning." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.83, + 0.348, + 0.847 + ], + "angle": 0, + "content": "5. Conclusion and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.472, + 0.903 + ], + "angle": 0, + "content": "We present OmniMAE, a unified Transformer for images and videos that can be pretrained using masked autoencoding. OmniMAE uses a simple architecture with minimal" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.451, + 0.895, + 0.725 + ], + "angle": 0, + "content": "vision-specific changes and is competitive with specialized architectures and models tailored for images and videos. We believe such generic models and approaches are a critical and under explored area of representation learning. OmniMAE has not been empirically validated on other visual modalities like 3D, or non-visual modalities like audio. The simplicity and generality of OmniMAE can likely enable future multimodal systems that use shared parameters to model multiple modalities. We only studied masked autoencoder based pretraining strategies for OmniMAE since they are easier to train. However, pixel reconstruction methods do not learn good linearly separable features or similarity metrics [40]. These properties are essential when finetuning with limited labeled data. We believe that exploring other pretraining strategies, particularly in the context of joint modeling, will likely lead to improved models. Similarly, joint finetuning strategies [33] could be used to evaluate such unified models and lead to potentially even superior performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Ethical considerations. We study models for visual recognition on images and videos and our technical contributions are neutral from an ethics standpoint. We do not propose a new dataset or make claims about the suitability of our model on production data. Given their self-supervised nature, our learned representations are free from label bias, however they are still susceptible to the bias in the distribution of the visual data. We believe all ethical considerations that apply to visual recognition models equally apply to our model." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "10413" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.081, + 0.091, + 0.173, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.116, + 0.471, + 0.145 + ], + "angle": 0, + "content": "[1] Relja Arandjelovic and Andrew Zisserman. Look, listen and learn. In ICCV, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.15, + 0.472, + 0.195 + ], + "angle": 0, + "content": "[2] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lucic, and Cordelia Schmid. ViViT: A video vision transformer. In ICCV, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.2, + 0.471, + 0.245 + ], + "angle": 0, + "content": "[3] Yuki Markus Asano, Christian Rupprecht, and Andrea Vedaldi. Self-labelling via simultaneous clustering and representation learning. In ICLR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.25, + 0.471, + 0.295 + ], + "angle": 0, + "content": "[4] Sara Atito, Muhammad Awais, and Josef Kittler. Sit: Self-supervised vision transformer. arXiv preprint arXiv:2104.03602, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.299, + 0.471, + 0.359 + ], + "angle": 0, + "content": "[5] Alexei Baevski, Wei-Ning Hsu, Qiantong Xu, Arun Babu, Jiatao Gu, and Michael Auli. data2vec: A general framework for self-supervised learning in speech, vision and language. In ICML, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.364, + 0.471, + 0.395 + ], + "angle": 0, + "content": "[6] Hangbo Bao, Li Dong, and Furu Wei. BEiT: Bert pre-training of image transformers. In ICLR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.399, + 0.471, + 0.444 + ], + "angle": 0, + "content": "[7] Adrien Bardes, Jean Ponce, and Yann LeCun. Vicreg: Variance-invariance-covariance regularization for self-supervised learning. In ICLR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.449, + 0.471, + 0.508 + ], + "angle": 0, + "content": "[8] Maxim Berman, Hervé Jégou, Andrea Vedaldi, Iasonas Kokkinos, and Matthijs Douze. Multigrain: a unified image embedding for classes and instances. arXiv preprint arXiv:1902.05509, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.513, + 0.471, + 0.558 + ], + "angle": 0, + "content": "[9] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In ICML, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.563, + 0.471, + 0.608 + ], + "angle": 0, + "content": "[10] Hervé Bourlard and Yves Kamp. Auto-association by multilayer perceptrons and singular value decomposition. Biological Cybernetics, 1988." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.612, + 0.471, + 0.657 + ], + "angle": 0, + "content": "[11] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features. In ECCV, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.662, + 0.471, + 0.722 + ], + "angle": 0, + "content": "[12] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. In NeurIPS, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.726, + 0.471, + 0.786 + ], + "angle": 0, + "content": "[13] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.791, + 0.471, + 0.836 + ], + "angle": 0, + "content": "[14] João Carreira and Andrew Zisserman. Quo vadis, action recognition? A new model and the kinetics dataset. In CVPR, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.84, + 0.471, + 0.9 + ], + "angle": 0, + "content": "[15] Lluis Castrejon, Yusuf Aytar, Carl Vondrick, Hamed Piriavash, and Antonio Torralba. Learning aligned cross-modal representations from weakly aligned data. In CVPR, 2016." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.894, + 0.137 + ], + "angle": 0, + "content": "[16] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In ICML, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.141, + 0.892, + 0.185 + ], + "angle": 0, + "content": "[17] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.189, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[18] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In CVPR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.222, + 0.894, + 0.266 + ], + "angle": 0, + "content": "[19] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In ICCV, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.271, + 0.894, + 0.328 + ], + "angle": 0, + "content": "[20] Kevin Clark, Minh-Thang Luong, Quoc V Le, and Christopher D Manning. Electra: Pre-training text encoders as discriminators rather than generators. *ICLR*, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.334, + 0.894, + 0.392 + ], + "angle": 0, + "content": "[21] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In CVPR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.397, + 0.894, + 0.471 + ], + "angle": 0, + "content": "[22] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Antonino Furnari, Evangelos Kazakos, Jian Ma, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, et al. Rescaling egocentric vision. IJCV, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.476, + 0.894, + 0.535 + ], + "angle": 0, + "content": "[23] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding. In NAACL, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.539, + 0.894, + 0.628 + ], + "angle": 0, + "content": "[24] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.633, + 0.894, + 0.693 + ], + "angle": 0, + "content": "[25] Alaaeldin El-Nouby, Gautier Izacard, Hugo Touvron, Ivan Laptev, Hervé Jegou, and Edouard Grave. Are large-scale datasets necessary for self-supervised pretraining? arXiv preprint arXiv:2112.10740, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.696, + 0.894, + 0.741 + ], + "angle": 0, + "content": "[26] Aleksandr Ermolov, Aliaksandr Siarohin, Enver Sangineto, and Nicu Sebe. Whitening for self-supervised representation learning. In ICML, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.745, + 0.894, + 0.803 + ], + "angle": 0, + "content": "[27] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. In ICCV, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.808, + 0.894, + 0.853 + ], + "angle": 0, + "content": "[28] Christoph Feichtenhofer, Haoqi Fan, Yanghao Li, and Kaiming He. Masked autoencoders as spatiotemporal learners. arXiv preprint arXiv:2205.09113, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.856, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[29] Christoph Feichtenhofer, Axel Pinz, and Andrew Zisserman. Convolutional two-stream network fusion for video action recognition. In CVPR, 2016." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10414" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.152 + ], + "angle": 0, + "content": "[30] Patrick Gallinari, Yann LeCun, Sylvie Thiria, and F Fogelman Soulie. Mémoires associatives distribuées: une comparaison (distributed associative memories: a comparison). In COGNITIVA. 1987." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.155, + 0.472, + 0.2 + ], + "angle": 0, + "content": "[31] Rohit Girdhar, João Carreira, Carl Doersch, and Andrew Zisserman. Video action transformer network. In CVPR, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.204, + 0.472, + 0.234 + ], + "angle": 0, + "content": "[32] Rohit Girdhar and Kristen Grauman. Anticipative Video Transformer. In ICCV, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.237, + 0.472, + 0.297 + ], + "angle": 0, + "content": "[33] Rohit Girdhar, Mannat Singh, Nikhila Ravi, Laurens van der Maaten, Armand Joulin, and Ishan Misra. Omnivore: A Single Model for Many Visual Modalities. In CVPR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.301, + 0.472, + 0.361 + ], + "angle": 0, + "content": "[34] Yunchao Gong, Liwei Wang, Micah Hodosh, Julia Hockenmaier, and Svetlana Lazebnik. Improving image-sentence embeddings using large weakly annotated photo collections. In ECCV, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.364, + 0.472, + 0.455 + ], + "angle": 0, + "content": "[35] Priya Goyal, Quentin Duval, Isaac Seessel, Mathilde Caron, Mannat Singh, Ishan Misra, Levent Sagun, Armand Joulin, and Piotr Bojanowski. Vision models are more robust and fair when pretrained on uncurated images without supervision. arXiv preprint arXiv:2202.08360, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.459, + 0.472, + 0.504 + ], + "angle": 0, + "content": "[36] Priya Goyal, Dhruv Mahajan, Abhinav Gupta, and Ishan Misra. Scaling and benchmarking self-supervised visual representation learning. In ICCV, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.507, + 0.472, + 0.613 + ], + "angle": 0, + "content": "[37] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, Florian Hoppe, Christian Thurau, Ingo Bax, and Roland Memisevic. The “something something” video database for learning and evaluating visual common sense. In ICCV, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.616, + 0.472, + 0.706 + ], + "angle": 0, + "content": "[38] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. NeurIPS, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.71, + 0.472, + 0.755 + ], + "angle": 0, + "content": "[39] Raia Hadsell, Sumit Chopra, and Yann LeCun. Dimensionality reduction by learning an invariant mapping. In CVPR, 2006." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.758, + 0.472, + 0.803 + ], + "angle": 0, + "content": "[40] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.807, + 0.472, + 0.853 + ], + "angle": 0, + "content": "[41] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.856, + 0.472, + 0.9 + ], + "angle": 0, + "content": "[42] Geoffrey E Hinton and Richard Zemel. Autoencoders, minimum description length and helmholtz free energy. In NeurIPS, 1993." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.894, + 0.151 + ], + "angle": 0, + "content": "[43] Elad Hoffer, Tal Ben-Nun, Itay Hubara, Niv Giladi, Torsten Hoefler, and Daniel Soudry. Augment your batch: Improving generalization through instance repetition. In CVPR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.155, + 0.893, + 0.215 + ], + "angle": 0, + "content": "[44] Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alex Shepard, Hartwig Adam, Pietro Perona, and Serge Belongie. The inaturalist species classification and detection dataset. In CVPR, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.219, + 0.893, + 0.263 + ], + "angle": 0, + "content": "[45] Gao Huang, Yu Sun, Zhuang Liu, Daniel Sedra, and Kilian Q Weinberger. Deep networks with stochastic depth. In ECCV, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.267, + 0.894, + 0.311 + ], + "angle": 0, + "content": "[46] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In ICML, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.315, + 0.893, + 0.359 + ], + "angle": 0, + "content": "[47] Andrej Karpathy and Li Fei-Fei. Deep visual-semantic alignments for generating image descriptions. In CVPR, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.363, + 0.894, + 0.453 + ], + "angle": 0, + "content": "[48] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, AMustafa Suleyman, and Andrew Zisserman. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.457, + 0.894, + 0.502 + ], + "angle": 0, + "content": "[49] Mark A Kramer. Nonlinear principal component analysis using autoassociative neural networks. AIChE, 1991." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.505, + 0.894, + 0.536 + ], + "angle": 0, + "content": "[50] Yann LeCun and Françoise Fogelman-Soulie. Modèles connexionnistes de l'apprentissage. Intellectica, 1987." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.539, + 0.893, + 0.583 + ], + "angle": 0, + "content": "[51] Junnan Li, Pan Zhou, Caiming Xiong, and Steven C.H. Hoi. Prototypical contrastive learning of unsupervised representations. In ICLR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.587, + 0.893, + 0.647 + ], + "angle": 0, + "content": "[52] Yanghao Li, Chao-Yuan Wu, Haoqi Fan, Karttikeya Mangalam, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. Improved multiscale vision transformers for classification and detection. In CVPR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.65, + 0.894, + 0.726 + ], + "angle": 0, + "content": "[53] Valerii Likhosherstov, Anurag Arnab, Krzysztof Choromanski, Mario Lucic, Yi Tay, Adrian Weller, and Mostafa Dehghani. Polyvit: Co-training vision transformers on images, videos and audio. arXiv preprint arXiv:2111.12993, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.729, + 0.893, + 0.789 + ], + "angle": 0, + "content": "[54] Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, Furu Wei, and Baining Guo. Swin transformer v2: Scaling up capacity and resolution. In CVPR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.792, + 0.893, + 0.853 + ], + "angle": 0, + "content": "[55] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.856, + 0.893, + 0.901 + ], + "angle": 0, + "content": "[56] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swin transformer. arXiv preprint arXiv:2106.13230, 2021." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10415" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.136 + ], + "angle": 0, + "content": "[57] Ze Liu, Zheng Zhang, Yue Cao, Han Hu, and Xin Tong. Group-free 3d object detection via transformers. In ICCV, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.141, + 0.47, + 0.186 + ], + "angle": 0, + "content": "[58] Jiasen Lu, Vedanuj Goswami, Marcus Rohrbach, Devi Parikh, and Stefan Lee. 12-in-1: Multi-task vision and language representation learning. In CVPR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.189, + 0.472, + 0.249 + ], + "angle": 0, + "content": "[59] Antoine Miech, Jean-Baptiste Alayrac, Lucas Smaira, Ivan Laptev, Josef Sivic, and Andrew Zisserman. End-to-end learning of visual representations from uncu-rated instructional videos. In CVPR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.254, + 0.472, + 0.298 + ], + "angle": 0, + "content": "[60] Ishan Misra, Rohit Girdhar, and Armand Joulin. An End-to-End Transformer Model for 3D Object Detection. In ICCV, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.302, + 0.472, + 0.347 + ], + "angle": 0, + "content": "[61] Ishan Misra and Laurens van der Maaten. Self-supervised learning of pretext-invariant representations. In CVPR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.352, + 0.472, + 0.395 + ], + "angle": 0, + "content": "[62] Pedro Morgado, Ishan Misra, and Nuno Vasconcelos. Robust audio-visual instance discrimination. In CVPR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.401, + 0.472, + 0.445 + ], + "angle": 0, + "content": "[63] Pedro Morgado, Nuno Vasconcelos, and Ishan Misra. Audio-visual instance discrimination with cross-modal agreement. In CVPR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.449, + 0.472, + 0.494 + ], + "angle": 0, + "content": "[64] B. A. Olshausen and D. J. Field. Emergence of simple-cell receptive field properties by learning a sparse code for natural images. Nature, 1996." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.499, + 0.472, + 0.543 + ], + "angle": 0, + "content": "[65] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. In NeurIPS, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.548, + 0.472, + 0.592 + ], + "angle": 0, + "content": "[66] Andrew Owens and Alexei A Efros. Audio-visual scene analysis with self-supervised multisensory features. In ECCV, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.596, + 0.472, + 0.641 + ], + "angle": 0, + "content": "[67] Niki Parmar, Ashish Vaswani, Jakob Uszkoreit, Lukasz Kaiser, Noam Shazeer, Alexander Ku, and Dustin Tran. Image transformer. In ICML, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.645, + 0.472, + 0.69 + ], + "angle": 0, + "content": "[68] Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A Efros. Context encoders: Feature learning by inpainting. In CVPR, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.694, + 0.472, + 0.739 + ], + "angle": 0, + "content": "[69] Boris T Polyak and Anatoli B Juditsky. Acceleration of stochastic approximation by averaging. SIAM journal on control and optimization, 1992." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.743, + 0.472, + 0.818 + ], + "angle": 0, + "content": "[70] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. IJCV, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.822, + 0.472, + 0.851 + ], + "angle": 0, + "content": "[71] Ruslan Salakhutdinov and Geoffrey Hinton. Deep Boltzmann machines. In AISTATS, 2009." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.856, + 0.472, + 0.9 + ], + "angle": 0, + "content": "[72] Karen Simonyan and Andrew Zisserman. Two-stream convolutional networks for action recognition in videos. In NeurIPS, 2014." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.894, + 0.15 + ], + "angle": 0, + "content": "[73] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In CVPR, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.156, + 0.894, + 0.216 + ], + "angle": 0, + "content": "[74] Hao Tan, Jie Lei, Thomas Wolf, and Mohit Bansal. VIMPAC: Video pre-training via masked token prediction and contrastive learning. arXiv preprint arXiv:2106.11250, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.222, + 0.894, + 0.266 + ], + "angle": 0, + "content": "[75] Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive multiview coding. arXiv preprint arXiv:1906.05849, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.272, + 0.894, + 0.331 + ], + "angle": 0, + "content": "[76] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. arXiv preprint arXiv:2203.12602, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.337, + 0.894, + 0.396 + ], + "angle": 0, + "content": "[77] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In ICML, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.402, + 0.894, + 0.461 + ], + "angle": 0, + "content": "[78] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.467, + 0.894, + 0.511 + ], + "angle": 0, + "content": "[79] P. Vincent, H. Larochelle, Y. Bengio, and P.-A. Manzagol. Extracting and composing robust features with denoising autoencoders. In ICML, 2008." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.516, + 0.894, + 0.576 + ], + "angle": 0, + "content": "[80] Rui Wang, Dongdong Chen, Zuxuan Wu, Yinpeng Chen, Xiyang Dai, Mengchen Liu, Yu-Gang Jiang, Luowei Zhou, and Lu Yuan. BEVT: Bert pretraining of video transformers. In CVPR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.581, + 0.894, + 0.655 + ], + "angle": 0, + "content": "[81] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In ICCV, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.661, + 0.894, + 0.72 + ], + "angle": 0, + "content": "[82] Chen Wei, Haoqi Fan, Saining Xie, Chao-Yuan Wu, Alan Yuille, and Christoph Feichtenhofer. Masked feature prediction for self-supervised visual pre-training. In CVPR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.726, + 0.894, + 0.77 + ], + "angle": 0, + "content": "[83] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In CVPR, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.776, + 0.894, + 0.835 + ], + "angle": 0, + "content": "[84] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jian-min Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In CVPR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.841, + 0.894, + 0.9 + ], + "angle": 0, + "content": "[85] Xueting Yan, Ishan Misra, Abhinav Gupta, Deepti Ghadiyaram, and Dhruv Mahajan. ClusterFit: Improving Generalization of Visual Representations. In CVPR, 2020." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.894, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10416" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.152 + ], + "angle": 0, + "content": "[86] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In ICCV, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.156, + 0.472, + 0.202 + ], + "angle": 0, + "content": "[87] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stephane Deny. Barlow twins: Self-supervised learning via redundancy reduction. In ICML, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.205, + 0.472, + 0.25 + ], + "angle": 0, + "content": "[88] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In ICLR, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.254, + 0.472, + 0.284 + ], + "angle": 0, + "content": "[89] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip Torr, and Vladlen Koltun. Point transformer. In ICCV, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.288, + 0.472, + 0.332 + ], + "angle": 0, + "content": "[90] Zhun Zhong, Liang Zheng, Guoliang Kang, Shaozi Li, and Yi Yang. Random erasing data augmentation. In AAAI, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.337, + 0.472, + 0.396 + ], + "angle": 0, + "content": "[91] Bolei Zhou, Agata Lapedriza, Jianxiong Xiao, Antonio Torralba, and Aude Oliva. Learning deep features for scene recognition using places database. In NeurIPS, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.401, + 0.472, + 0.446 + ], + "angle": 0, + "content": "[92] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. ibot: Image bert pre-training with online tokenizer. In ICLR, 2022." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10417" + } + ] +] \ No newline at end of file diff --git a/2023/OmniMAE_ Single Model Masked Pretraining on Images and Videos/2fdd9b20-eaac-4038-a247-56813890042d_origin.pdf b/2023/OmniMAE_ Single Model Masked Pretraining on Images and Videos/2fdd9b20-eaac-4038-a247-56813890042d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..1d559124554d72be5338a0db670e7a776d0f3de9 --- /dev/null +++ b/2023/OmniMAE_ Single Model Masked Pretraining on Images and Videos/2fdd9b20-eaac-4038-a247-56813890042d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e615b8bf57411f4921e7899780d2da557db4a5d885daf727fd22f3c82a187909 +size 3415960 diff --git a/2023/OmniMAE_ Single Model Masked Pretraining on Images and Videos/full.md b/2023/OmniMAE_ Single Model Masked Pretraining on Images and Videos/full.md new file mode 100644 index 0000000000000000000000000000000000000000..25365abe859279091a73ae191fe10380b6d58bd9 --- /dev/null +++ b/2023/OmniMAE_ Single Model Masked Pretraining on Images and Videos/full.md @@ -0,0 +1,281 @@ +# OmniMAE: Single Model Masked Pretraining on Images and Videos + +Rohit Girdhar* Alaaeldin El-Nouby* Mannat Singh* +Kalyan Vasudev Alwala* Armand Joulin Ishan Misra* +FAIR, Meta AI + +https://github.com/facebookresearch/omnivore + +![](images/af9705937884c82df0e348223ab58e9199ed22c6ae7d522c3fe23126aa586415.jpg) +Figure 1. OmniMAE is a single model for images and videos that is trained using masked autoencoding [40]. We use a plain Vision Transformer [24] architecture but with spatio-temporal patches as input. At training, we 'patchify' the visual input (images or videos), and feed the encoder only a subset of the patches. The decoder reconstructs the pixels for the missing patches using the encoder's output. The encoder-decoder model is trained using a pixel reconstruction loss. After training, our single plain Transformer encoder performs competitively compared to specialized architectures on downstream image and video recognition tasks. + +# Abstract + +Transformer-based architectures have become competitive across a variety of visual domains, most notably images and videos. While prior work studies these modalities in isolation, having a common architecture suggests that one can train a single unified model for multiple visual modalities. Prior attempts at unified modeling typically use architectures tailored for vision tasks, or obtain worse performance compared to single modality models. In this work, we show that masked autoencoding can be used to train a simple Vision Transformer on images and videos, without requiring any labeled data. This single model learns visual representations that are comparable to or better than single-modality representations on both image and video benchmarks, while using a much simpler architecture. Furthermore, this model can be learned by dropping $90\%$ of the image and $95\%$ of the video patches, enabling extremely fast training of huge model architectures. In particular, we show that our single ViT-Huge model can be finetuned to achieve $86.6\%$ on ImageNet and $75.5\%$ on the challenging Something Something-v2 video + +benchmark, setting a new state-of-the-art. + +# 1. Introduction + +The Transformer architecture [78] is rapidly becoming competitive across the different visual modalities in Computer Vision, from images [24, 27, 55, 77], to 3D [57, 60, 89] and videos [2, 9, 27, 31, 32, 56]. This convergence toward a unified architecture naturally suggests that we should be able to train a single model that works across different visual modalities. However, recent attempts to train unified models either lead to worse performance compared to single modality models [53], or require the use of an alternative architecture [33], namely the Swin Transformer [55], with inductive biases tailored towards vision tasks. While specialized Transformer architectures for vision [27, 55, 56, 81] can offer better performance for visual modalities, they lose the generality and flexibility of the vanilla Transformer, making it harder to later model different domains like text, speech, 3D etc. in multi-modal architectures. + +In this work, we train a single vanilla Transformer that works for both images and videos, as illustrated in Figure 1. + +To this end, we leverage the findings of several recent works on the use of the masked pretraining [23] to greatly improve the training and performance of Transformers in the domain of images [6, 40, 80, 84], videos [76, 80, 82] or across text, audio and images [5]. We show that this masked pretraining is a viable strategy to pretrain a unified 'omnivorous' Transformer across visual modalities. In particular, we consider the Masked Auto-Encoding (MAE) approach [40] to train an Omnivorous visual encoder [33]. The resulting OmniMAE model learns from all the modalities with the same objective function and does not require any supervision. + +Using a masked pretraining objective has several advantages over supervised objectives [33, 53] or discriminative self-supervised objectives [13, 17, 41]. First, as opposed to supervised objectives, a general-purpose unsupervised loss does not require any human labeling effort. As a result, it is robust to biases introduced by a predefined set of labels [35]. Moreover, it does not require a multi-head architecture to incorporate supervision from each of the label spaces corresponding to each modality, which is hard to maintain and scale with new modalities. Second, although discriminative self-supervised methods produce superior frozen features compared to reconstruction objectives, they are non trivial to scale in model and data size [36]. Our masked pretraining objective is simple, efficient to train, and scales to different visual modalities with minimal changes. + +Contributions. (1) We show that the simple Vision Transformer architecture (ViT) [24] originally designed for images can naturally be applied on videos, and videos and images jointly. OmniMAE is a single ViT-based model for videos and images that outperforms architectures and models specifically designed for either modality. (2) Prior and concurrent work design self-supervised methods and architectures for either image or video and we find that these models do not transfer well across modalities. OmniMAE is the first single self-supervised model that achieves good performance on both modalities. (3) We show that our joint training using both images and videos enables us to use much higher masking ratios than any prior work for training MAE. Since ViT can processes only the non-masked input, we train OmniMAE models with only $10\%$ of image and $5\%$ of video patches. This enables us to train large (650M parameter) models with a $\sim 7\times$ and $\sim 11\times$ reduction in compute and memory on images and videos. (4) Finally, we propose improvements to the MAE training. We show that repeating samples in a mini-batch reduces dataloading (and thus training) time without loss in final transfer performance. Sample replication is particularly useful for masked pretraining as the unmasked patches are different across sample replicas. We also show that using a shallow shared decoder for videos and images leads to better performance while reducing the number of parameters by $2 - 4\times$ . + +# 2. Related Work + +Our work builds upon research in self-supervised learning, masked pretraining and unified modeling in computer vision. Self-supervised learning. In recent years, self-supervised approaches have been dominated by joint embedding methods which can rely on different objectives including contrastive [17,39,41,51,61,65,75,83], non-contrastive [7,18,26, 87], clustering [3, 11, 12, 85] or self-distillation [5, 13, 38, 92]. Such methods are trained to learn invariance to a set of predefined transformations which results in image descriptors with a strong linear probing and KNN performance. However, such methods can be challenging to scale since they can suffer from instabilities [19]. Additionally, the strongest performance is typically obtained with the help of augmentations like multi-crop [12, 13, 92] which can be hard to apply at scale due their compute and memory overhead. + +Masked pretraining. We build upon masked prediction methods where the representation is learned by predicting masked parts of the input. Such methods have recently gained popularity given their immense success in NLP. In particular, BERT [23] showed that masked language modeling by predicting a subset of masked words is a powerful pre-training objective and leads to impressive finetuning performance on various downstream tasks. In computer vision, input reconstruction methods have a rich history with non-linear PCA [49], sparse reconstruction [64], autoencoders [10, 30, 42, 50], RBMs [71] etc. Masked prediction methods can be viewed as a special case of denoising autoencoders [30, 79] where the input 'noise' is a masking function. An example of such a method that uses masking as noise is context encoders [68]. With the recent and rapid rise of Vision Transformers [24], masked prediction was revisited by multiple efforts. SiT [4] replaces variable sized patches in the image with random noise and trains a ViT model for reconstruction, among other objectives. BEiT [6] moves a step closer to BERT with replacing full patches with mask tokens and training a ViT encoder to predict the discrete visual words of masked patches using a cross-entropy loss. Masked prediction has also shown impressive performance for specialized vision transformer architectures such as MViT [27, 52, 82] and Swin transformer [54, 55, 84]. SimMIM [84] and MaskedFeat [82] predict pixel values and HOG features of the masked patches using a Swin-v2 [54] and MViT-v2 [52] backbones respectively. Finally, SplitMask [25] studied the interesting properties of masked prediction methods in terms of high sample efficiency. + +Of particular interest to OmniMAE, masked autoencoders (MAE) [40] demonstrated impressive scaling properties by utilizing a patch dropping strategy for masked patches accompanied with a high masking ratio of $75\%$ . Under this setting, the encoder only process a small subset of the image patches followed by a relatively low capacity decoder which reconstructs the image pixels. This property is even more + +crucial for video representation learning given the large number of patches, and a concurrent work, ST-MAE [76], shows that MAE pretraining with an even higher masking ratio of $90\%$ works well and obtains strong finetuning performance on downstream video recognition benchmarks. Notably, the efficiency gained by patch dropping is specific to vanilla ViTs, and multiscale architectures such as MViT and Swin are unable to benefit due to their design. Hence, we use the simple ViT as our architecture and show that it can be trained efficiently and jointly for images and videos using extremely high masking ratios (90-95% on both modalities), and yet perform competitively to specialized architectures like MViT [27] and Swin [55]. + +Unified modeling and multi-modal learning. Multi-modal learning in computer vision has a long history that includes training using images and text [15, 34, 47, 58, 59], video and optical flow [29, 72], and video and audio [1, 62, 63, 66]. The majority of such methods rely on training a separate backbone for each modality as well as the availability of alignment across modalities. More recently, Omnivore [33] was proposed for joint training of multiple modalities like images, videos and single-view 3D, attaining a strong performance in each of the modality specific benchmarks with a single shared trunk. PolyViT [53] co-trains a shared transformer encoder using images, videos and audio data and provides a competitive performance on various downstream tasks for each modality. The aforementioned methods differ compared to OmniMAE in that they are trained with supervised learning and require human annotations. BEVT [80] tackles BERT pre-training for videos and proposes that jointly pretraining using static images improves the finetuning performance of video recognition benchmarks. Unlike OmniMAE, BEVT uses the specialized Swin transformer architecture with separate decoder heads for images and videos. Thus, BEVT cannot drop patches while training which can limit its scalability. Furthermore, it relies on a tokenizer which must be trained apriori, and the tokenizer training itself can affect the model's performance. + +# 3. OmniMAE + +Our goal is to pretrain a single unified model for images and videos. Rather than use specialized architectures tailored for a visual modality, we build upon the vanilla Vision Transformer (ViT) [24] architecture that has limited inductive biases for vision. For pretraining, we extend the simple self-supervised masked auto-encoding (MAE) approach [40]. The original architecture and pretraining method are tested only on images, and we show simple design decisions for a unified model. + +# 3.1. Training OmniMAE jointly on images and videos + +We illustrate our method in Figure 1. For pretraining, we use an encoder-decoder architecture where the encoder only operates on a 'non-masked' subset of the input. The decoder predicts the pixel values for the entire input, i.e., masked and non-masked pixels. The model is trained to minimize the reconstruction error for the masked (unseen) part of the input. After pretraining, we evaluate the encoder by transfer learning (the decoder is discarded). Next, we describe the pretraining details. + +Images and videos as spatio-temporal patches. The input image or video can be represented as a 4D tensor of shape $T \times H \times W \times 3$ where $T$ is the temporal dimension and $H, W$ are the spatial dimensions, and 3 represents the color channels. We treat images as being single-frame videos with $T = 1$ . The input is split into $N$ spatio-temporal patches, each of size $t \times h \times w \times 3$ [33]. + +Omnivorous visual encoder. We use an omnivorous [33] visual encoder that processes both images and video using the same parameters. The encoder operates on the $N$ spatiotemporal patches from the images and videos. The encoder can naturally handle variable number $N$ of patches from images and videos as it uses the Transformer architecture [78]. The encoder shares the same parameters for both image and video inputs and processes them in the same way to output per-patch embeddings. + +Pretraining. We convert the input into $N$ spatio-temporal patches and randomly mask $M$ of them. Following [40], we feed the non-masked $N - M$ patches (and their positions) to the encoder to produce per-patch embeddings. The encoder output is concatenated with $M$ copies of a learnable 'mask token' to obtain $N$ embeddings. We add the appropriate positional encoding to the $N$ embeddings and feed them into the decoder, which outputs $N \times t \times h \times w \times 3$ pixel values. + +Loss function and optimization. We minimize the reconstruction error between the decoder predictions and the input pixel values. The input pixel values are normalized to zero mean and unit variance [46] to get the targets for the loss. We minimize the $\ell_2$ distance between the predictions and targets over the $M$ masked patches. At each point in the training, we sample a mini-batch of either images or video and compute the loss using the decoder predictions. We study the effect of different mini-batch construction strategies on the overall performance and speed of training in § 4.2. + +Dataset sample replication. Since we operate on a small subset of the input patches $(M \ll N)$ , we find that our data reading speed is unable to keep up with the number of data points we can theoretically process in a single optimization step. This issue is even more pronounced for videos, where we spend significant compute to read and decode a full video, only to discard $>90\%$ of it. Inspired from the success of RepeatAugment, which shows that replicating samples + +![](images/d1ac4e9640794e89298e95775d504095633419cb539bbb20673e66969149a038.jpg) +Figure 2. OmniMAE on image and video downstream tasks. We finetune the MAE, ST-MAE, and OmniMAE models on image and video benchmarks. We use the ViT architecture with two model sizes: ViT-B and ViT-L. MAE has poor video recognition performance while ST-MAE's performance drops on image datasets. OmniMAE pretraining generalizes to both benchmarks. All models are trained for 800 epochs on the pretraining datasets. The image-only MAE model is inflated [14] to apply MAE to video recognition tasks. The input image is replicated to apply ST-MAE to image recognition benchmarks. + +within a mini-batch is an effective technique to improve generalization [8, 12, 43], we replicate a single data point and apply different masks to it each time. Even with replicated samples, the non-masked input to the model is different due to random cropping and masking being different across the replicas. We show in § 4.2 that sample replication leads to gains in runtime without affecting the final transfer accuracy. + +# 3.2. Implementation Details + +We note some of the salient implementation details and provide the complete details in Appendix B. + +Architecture. We use the ViT [24] architecture for the omnivorous encoder and experiment with its ViT-B, ViT-L, and ViT-H variants. We do not use the [CLS] token in the ViT models, yielding a small improvement in runtime without loss in performance. We use a Transformer decoder with 4 layers (8 for ViT-H) of 384, 512, and 512 embedding dimensions for ViT-B, ViT-L, and ViT-H, respectively. The decoder outputs the RGB colors for the pixels in all the input patches. We use sinusoidal positional encoding [78] for the patches in both the encoder and the decoder. + +Training details. We train our models with a mini-batch size of 2048. We resize the input images and videos spatially to $224 \times 224$ pixels. For videos, we sample a clip of $T = 16$ frames at 6 FPS. We use a patch size of $2 \times 16 \times 16$ for ViT-B and ViT-L, and $2 \times 14 \times 14$ for ViT-H. Images are replicated temporally to meet the patch size. + +Masking the input. Compared to prior work, we use extremely high masking for pretraining and only $10\%$ and $5\%$ of the image and video patches are fed to the encoder for pretraining. We uniform randomly mask the patches for images and videos and ablate the masking hyperparameters in $\S 4.2$ . + +# 4. Experiments + +Pretraining data. We pretrain representations by jointly training on images from the ImageNet (IN1K) [70] dataset + +and videos from the Something Something-v2 (SSv2) [37] dataset. We choose the SSv2 dataset over web video datasets like Kinetics-400 (K400) [48] for reproducibility, and given SSv2's challenging nature requiring temporal reasoning. + +Transfer learning datasets. We consider two image datasets: iNaturalist-2018 (iNat18) [44], a popular fine-grained recognition dataset, and Places-365 (P365) [91], a scene recognition benchmark. For video datasets, we focus on the popular K400 action recognition benchmark, and EPIC-Kitchens-100 (EK100) [22], an egocentric action recognition dataset. We report the top-1 classification accuracy for all transfer datasets. The details about the datasets can be found in Appendix A. + +# 4.1. OmniMAE on Image and Video Recognition + +We now evaluate the capabilities of OmniMAE's representations on both image and video recognition tasks. + +Baselines. We compare OmniMAE's representations trained jointly on images and videos to MAE, trained solely on images [40]. Additionally, we develop a video-only baseline, SpatioTemporal-MAE (ST-MAE), by training OmniMAE only on videos. For a fair comparison, all models are pretrained for 800 epochs on their respective datasets. The image-only MAE model is inflated [14] for finetuning on video tasks, see Appendix B.2 for details. + +Observations. Figure 2 presents the evaluation results for all the models on the image and video benchmarks. The modality specific MAE and ST-MAE models achieve good transfer performance when transferring to datasets which match the pretraining modality. However, both models show a degradation in performance when transferring to a different visual modality. MAE has lower video recognition performance, especially on egocentric videos from EK100 where the smaller ViT-B MAE model is $18.4\%$ worse compared to ST-MAE. Similarly, compared to MAE, ST-MAE is $8.7\%$ worse on the fine-grained classification task of iNat18. With a large model, MAE's performance on the image recognition + +![](images/a1eb0f78d4fcbcc4d87c895bfbb0346df6d33657707c0c6c292d228dfcc0067c.jpg) +Figure 3. Different types of masking for images (left two) and videos. Causal and tube masking use the data's spatio-temporal structure. Random frame masking randomly masks frames in a video. Random masking randomly masks patches and is used by default for OmniMAE. + +![](images/e01de3061587e5a19d129ae2652a3c26da65b7a04d12868d46af82a2d4fa961c.jpg) + +![](images/1c6238b96d5ab1b44f95246d7e7b204b70e8aa02adc30ab3802282f60ba077da.jpg) + +![](images/0df45b5cc7f205ade1731b919b40ecf46e1e6ee18b3660ed91fecd37458a9132.jpg) + +![](images/cafecf3af80989efae0adcac0170a585578b07837ebec5ff90539675851fd9cc.jpg) + +![](images/cb30810cf07f8315653191565f6939e5ab2707f672c6cdc5b122fb735c911a17.jpg) + +benchmarks improves but the cross-modality performance degrades further. On EPIC-Kitchens-100 the large MAE model is more than $25\%$ worse than ST-MAE. + +When transferred to both image and video recognition tasks, OmniMAE performs favorably to the single modality baselines that use exactly the same architecture. OmniMAE also uses the same finetuning recipe as the single modality baselines. OmniMAE matches the video classification performance of ST-MAE and the image classification performance of MAE for both ViT-B and ViT-L, with OmniMAE's performance improving on both image and video recognition benchmarks with the larger model. These results suggest that our unified model of images and videos serves as a better pretraining across a wider variety of recognition tasks than single modality models. + +Qualitative results. Following [40], we re-train OmniMAE without normalizing the pixel targets to obtain easy to visualize RGB reconstructions. We visualize the predictions in Figure 4 using samples that are unseen during training: either val sets or different datasets altogether. OmniMAE makes reasonable predictions on the in-distribution ImageNet and SSv2 val sets, as well as the unseen, out-of-distribution K400 and EK100 datasets. As expected, the details in the reconstruction decrease when increasing the masking ratio. However, even with $95\%$ masking, the model can reconstruct coarse details in the input, e.g., in ImageNet, the model reconstructs the coarse structure of the flower, vehicle, dog etc. + +![](images/3da07f163da4192e2af7cb84ade3b8943096683fa66e96192de72f49af0fb792.jpg) +Figure 5. Sample replication. We study the effect of repeating samples while training our model. In each case, we repeat a sample $n$ times within a mini-batch while fixing the overall mini-batch size and training updates. Replication leads to improved training speeds, especially on video without affecting the final performance. + +![](images/3437c6757e850f8d9b1e98baeab30c0df4a211cb589bfb6a02bb6f907a6ec27a.jpg) + +# 4.2. Analyzing OmniMAE + +We train the ViT-B architecture jointly on the ImageNet and the SSv2 datasets for 800 epochs, where an epoch involves training over all samples from both the datasets. For these analysis experiments, we use the default masking hyperparameters of MAE for images (75%). For ST-MAE, due to redundancy across frames, we use "tube" dropping (i.e. dropping all patches at a given spatial location over time) with high masking ratio (90%) All hyperparameters for these experiments are in Appendix B.3. We evaluate all the pretrained models on ImageNet and SSv2 and report the top-1 classification accuracies in Table 1. Please see Appendix D for additional ablations. + +Extreme masking. We vary the ratio of masked patches in the input while training the model. We observe that videos benefit from significantly higher amounts of masking than images. Since video frames have a lot of redundant information, the resulting spatio-temporal patches are also redundant and thus higher masking leads to better self-supervised models. Unlike our experiments, prior work found that extremely high masking ratios lead to degradation in performance, for instance MAE [40] saw a significant degradation in performance when masking more than $75\%$ of the patches. However, as we show in Table 1a, OmniMAE models can use extremely high masking of the input while learning good representations. + +Reduced compute due to extreme masking. Our models trained with $90\%$ masking on images and $95\%$ masking on videos yield good performance, while being trained with just 19 and 78 unmasked image and video patches respectively, for a patch size of $2 \times 16 \times 16$ . As we follow [40] and only pass unmasked patches to the encoder and have a lightweight decoder, extreme masking leads to a dramatically lower computational cost for training the encoder, and consequently the model as a whole. Compared to using all the patches, our masked autoencoder uses $5.9 \times$ and $7.8 \times$ fewer FLOPS for ViT-B on images and videos respectively, $7.1 \times$ and $11.6 \times$ for ViT-L, and $7.2 \times$ and $11.3 \times$ for ViT-H. Compared to MAE, on ViT-B, ViT-L and ViT-H, our higher masking leads to $1.8 \times$ , $2.0 \times$ and $2.0 \times$ fewer FLOPS on images. On videos, compared to some concurrent works [28, 76] that use $90\%$ + +
RatiosAccuracy
IN1KSSv2IN1KSSv2
75%75%82.667.2
75%90%82.867.9
75%95%82.868.1
90%75%82.867.7
90%90%82.668.5
90%95%82.868.6
+ +(a) Masking ratios. + +
TypeAccuracy
IN1KSSv2IN1KSSv2
RandomTube82.867.9
CausalTube82.167.3
RandomFrame82.965.8
RandomCausal82.864.4
RandomRandom82.767.8
+ +(b) Masking type. + +
#paramsIN1KSSv2
Common7.9M82.867.9
Separate33.0M82.667.7
L=23.8M82.767.9
L=47.9M82.867.9
L=814.5M82.868.0
d=3847.9M82.867.9
d=51213.0M82.768.1
+ +(c) Decoder capacity. + +![](images/1f7c489cc73d09c22ac440fb4ee50ef4e5f412ac731febf69a23e4d158a72d3b.jpg) + +(d) Dataset ratio + +Table 1. Ablations. The default setting for all ablations is highlighted in gray. (a) We vary the masking ratios of the input for pretraining the model. This allows us to use an extremely high masking ratio of $90\%$ on images and $95\%$ on videos. (b) OmniMAE works well across different masking types. Random masking makes no assumptions about the input patches. Tube masking masks random patches at the same spatial location across all frames. Causal masking for images masks the 'future' patches as determined by a raster left-to-right order. In videos, causal masking masks all the patches from future frames. Frame masking masks all patches for randomly selected frames. (c) Different decoder designs (common or separate) and capacities (number of layers $L$ and dimension $d$ of the MLP). A common decoder for both images and videos performs better than a separate decoder. Our final performance is robust to decoder capacity. (d) We vary the image/video dataset ratios by replicating the entire datasets by a factor. The number of training updates changes with such replication, and we observe that the model benefits from higher replication of videos. + +masking, we obtain $1.3 \times$ , $1.5 \times$ and $1.4 \times$ fewer FLOPS for ViT-B/L/H. Given the compute savings and strong performance, we choose $90 \%$ and $95 \%$ masking for images and videos respectively for our final models. + +Type of masking. We study the effect of the type of masking used in training our models. The different types of masking are illustrated in Figure 3. We experiment with Random masking which masks patches in the image or video randomly. For videos, we experiment with two types of masking that exploit the temporal structure. In Tube masking, we mask random patches at the same spatial location across all the frames. In Causal masking, we use a raster order masking, akin to generative image models [16, 67], i.e. the patches in the top-left of the image, and earlier frames of the video, are kept while the rest are masked. Finally, in Frame masking, we randomly mask some frames in the video, while keeping all patches from the unmasked frames. As seen in Table 1b, Random and Tube masking perform comparably well on both image and video tasks. We find that in case of Causal masking, the prediction task becomes exceedingly tough due to the uncertainty of the future, and in case of Frame masking, it becomes relatively easy due to the high redundancy of pixel values across frames. Hence in both these cases, the representation learned does not perform as well on video recognition tasks. Given the simplicity of random masking, we use that for both modalities. + +Decoder architecture. Since image and video prediction tasks may require specialized parameters, we study two settings: (1) the decoder parameters are shared for image and video pixel prediction; (2) two separate decoders are used for image and video prediction. For the latter we use our default decoder setting for videos (4 layers/384-D), however a larger 8-layer/512-D decoder for images as proposed in MAE [40]. The results in Table 1c show that using a shared decoder for both image and video prediction leads to better + +transfer learning accuracy. A shared decoder also offers a $4 \times$ reduction in the number of parameters compared to using separate decoders while also being simpler to implement. + +Decoder capacity. In Table 1c, we vary the decoder capacity and measure its impact on the final transfer learning accuracy. To change the decoder capacity, we vary the number of Transformer layers $L$ used and the dimension $d$ of the MLP used in the Transformer layers. Overall, the final transfer learning accuracy is robust to decoder capacity. A shallow decoder of 4 layers (8 for ViT-H) offers a good trade-off between the decoder size and final accuracies. + +Dataset ratios. Since ImageNet and SSv2 have a different number of samples, we study the effect of varying the ratio of the datasets used in training. When increasing the ratio of the datasets, we measure a single epoch as training over the new oversampled set of samples from the datasets. We vary the relative ratio for ImageNet and Something Something-v2 by replicating only one dataset at a time as shown in Table 1d. We observe that increasing the relative dataset ratio has a positive effect on the final transfer accuracy for the oversampled dataset. This is expected as oversampling a dataset proportionally increases its corresponding parameter updates, thus making the representation better tuned for the dataset. We also observe that oversampling Something-Something-v2 twice as much as ImageNet leads to an improvement for the video transfer accuracy with no drop in the image transfer performance. Hence, OmniMAE is robust to changes in dataset sizes of individual modalities, and it suggests that longer training on both datasets can further improve the transfer learning performance of our models. For simplicity, by default we do not replicate any dataset for training our models. + +Sample replication. We replicate samples for both images and videos, and in each case measure a single epoch as training over the total samples, counting replicated samples, i.e. + +![](images/e0236fdf699dda3c27e0ae084442361871d705c5422b55da07c16974fc4df7b5.jpg) +Figure 4. Reconstruction visualizations using OmniMAE on different video and image datasets. We show the model predictions for varying masking ratios of the input from $75\%$ to $95\%$ and the ground truth reference (Ref). OmniMAE is trained on ImageNet and SSv2 but the predictions generalize to other datasets like K400 and EK100. Please see the supplement for video visualizations. + +replication maintains the same number of training iterations. We train models with different replication factors and show the final transfer accuracy and the normalized training time in Figs. 5a and 5b. Sample replication leads to faster training while maintaining or even improving the final transfer accuracy. Since a large portion of the input sample is masked, replicating the sample multiple number of times still provides enough learning signal for the model and does not lead to a degradation in performance. This becomes even more relevant for OmniMAE's final settings where we use + +higher masking ratios. For video data, we use an optimized datalocator (see details in Appendix B) with asynchronous I/O and fast video decoding. Even with this optimized datalocator, sample replication leads to $20\%$ faster training. We believe this is an important practical observation as video datalugging is often a bottleneck for training models. + +# 4.3. Comparison to Prior Work + +In Table 2, we compare OmniMAE's representation on our image and video classification benchmarks with other + +
MethodArch.Pretrain DataIN1KiNat18P365K400SSv2EK100
DINOViT-BIN1K82.872.6-XXX
iBOT [92]ViT-LIN1K84.8--XXX
MAE [40]ViT-BIN1K83.675.457.9XXX
MAE [40]ViT-LIN1K85.980.159.4XXX
MAE [40]ViT-HIN1K86.983.059.8XXX
BEiT [6]ViT-BIN1K83.472.3-XXX
BEiT [6]ViT-LIN1K85.2--XXX
VIMPAC [74]ViT-L/2†HowTo100MXXX77.468.1-
MaskedFeat [82]MViT-v2-LK400XXX84.3--
BEVT [80]Swin-BIN1K + K400XXX81.171.4-
OmniMAEViT-BIN1K + K40082.974.258.580.869.040.3
OmniMAEViT-BIN1K + SSv283.074.058.480.669.539.3
OmniMAEViT-LIN1K + SSv285.279.659.384.074.245.1
OmniMAEViT-HIN1K + SSv286.683.260.185.475.548.0
Concurrent methods specialized for videos
VideoMAE-16f [76]ViT-LSSv2XXX-74.2-
VideoMAE-16f [76]ViT-LK400XXX84.7--
MAE-Video [28]ViT-HK400XXX85.174.1-
+ +Table 2. Comparing OmniMAE with prior self-supervised methods on image and video recognition. Prior work trains a specialized model for a particular visual modality, sometimes using specialized architectures. Our single OmniMAE model is pretrained jointly on images and videos for 1600 epochs (ViT-H for 2400) and performs competitively across all benchmarks while using a simple architecture and pretraining method, even outperforming concurrent work specialized for videos. ViT-L with half the MLP embedding dimension + +state-of-the-art self-supervised methods. We focus on methods which use a Transformer backbone, ViT, or an architecture based on Transformers like Swin. To the best of our knowledge, prior work does not explore joint pre-training except for BEVT [80], which does masked image modeling followed by joint masked image and video modeling. Unlike OmniMAE, BEVT only focuses on video representations and does not evaluate on image recognition. + +OmniMAE is pretrained for 1600 epochs (ViT-H for 2400 epochs). Training with IN1K and K400 or IN1K and SSv2 leads to similar results, so we use the latter when scaling model size. OmniMAE performs competitively for both image and video recognition when compared to the best models trained separately for either modality. OmniMAE also performs favorably compared to methods like MaskedFeat which are pretrained only for video with specialized architectures on larger video datasets. OmniMAE serves as a competitive initialization for transferring models on all modalities. More notably, OmniMAE's performance improves significantly with larger architectures. Using ViT-H, OmniMAE performs better or within the margin of error across all benchmarks. Notably, OmniMAE even outperforms concurrent works which use similar approaches while solely focusing on video representation learning. + +# 5. Conclusion and Future Work + +We present OmniMAE, a unified Transformer for images and videos that can be pretrained using masked autoencoding. OmniMAE uses a simple architecture with minimal + +vision-specific changes and is competitive with specialized architectures and models tailored for images and videos. We believe such generic models and approaches are a critical and under explored area of representation learning. OmniMAE has not been empirically validated on other visual modalities like 3D, or non-visual modalities like audio. The simplicity and generality of OmniMAE can likely enable future multimodal systems that use shared parameters to model multiple modalities. We only studied masked autoencoder based pretraining strategies for OmniMAE since they are easier to train. However, pixel reconstruction methods do not learn good linearly separable features or similarity metrics [40]. These properties are essential when finetuning with limited labeled data. We believe that exploring other pretraining strategies, particularly in the context of joint modeling, will likely lead to improved models. Similarly, joint finetuning strategies [33] could be used to evaluate such unified models and lead to potentially even superior performance. + +Ethical considerations. We study models for visual recognition on images and videos and our technical contributions are neutral from an ethics standpoint. We do not propose a new dataset or make claims about the suitability of our model on production data. Given their self-supervised nature, our learned representations are free from label bias, however they are still susceptible to the bias in the distribution of the visual data. We believe all ethical considerations that apply to visual recognition models equally apply to our model. + +# References + +[1] Relja Arandjelovic and Andrew Zisserman. Look, listen and learn. In ICCV, 2017. +[2] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lucic, and Cordelia Schmid. ViViT: A video vision transformer. In ICCV, 2021. +[3] Yuki Markus Asano, Christian Rupprecht, and Andrea Vedaldi. Self-labelling via simultaneous clustering and representation learning. In ICLR, 2020. +[4] Sara Atito, Muhammad Awais, and Josef Kittler. Sit: Self-supervised vision transformer. arXiv preprint arXiv:2104.03602, 2021. +[5] Alexei Baevski, Wei-Ning Hsu, Qiantong Xu, Arun Babu, Jiatao Gu, and Michael Auli. data2vec: A general framework for self-supervised learning in speech, vision and language. In ICML, 2022. +[6] Hangbo Bao, Li Dong, and Furu Wei. BEiT: Bert pre-training of image transformers. In ICLR, 2022. +[7] Adrien Bardes, Jean Ponce, and Yann LeCun. Vicreg: Variance-invariance-covariance regularization for self-supervised learning. In ICLR, 2022. +[8] Maxim Berman, Hervé Jégou, Andrea Vedaldi, Iasonas Kokkinos, and Matthijs Douze. Multigrain: a unified image embedding for classes and instances. arXiv preprint arXiv:1902.05509, 2019. +[9] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In ICML, 2021. +[10] Hervé Bourlard and Yves Kamp. Auto-association by multilayer perceptrons and singular value decomposition. Biological Cybernetics, 1988. +[11] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features. In ECCV, 2018. +[12] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. In NeurIPS, 2020. +[13] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. +[14] João Carreira and Andrew Zisserman. Quo vadis, action recognition? A new model and the kinetics dataset. In CVPR, 2017. +[15] Lluis Castrejon, Yusuf Aytar, Carl Vondrick, Hamed Piriavash, and Antonio Torralba. Learning aligned cross-modal representations from weakly aligned data. In CVPR, 2016. + +[16] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In ICML, 2020. +[17] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. +[18] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In CVPR, 2021. +[19] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In ICCV, 2021. +[20] Kevin Clark, Minh-Thang Luong, Quoc V Le, and Christopher D Manning. Electra: Pre-training text encoders as discriminators rather than generators. *ICLR*, 2020. +[21] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In CVPR, 2020. +[22] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Antonino Furnari, Evangelos Kazakos, Jian Ma, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, et al. Rescaling egocentric vision. IJCV, 2021. +[23] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding. In NAACL, 2018. +[24] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. +[25] Alaaeldin El-Nouby, Gautier Izacard, Hugo Touvron, Ivan Laptev, Hervé Jegou, and Edouard Grave. Are large-scale datasets necessary for self-supervised pretraining? arXiv preprint arXiv:2112.10740, 2021. +[26] Aleksandr Ermolov, Aliaksandr Siarohin, Enver Sangineto, and Nicu Sebe. Whitening for self-supervised representation learning. In ICML, 2021. +[27] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. In ICCV, 2021. +[28] Christoph Feichtenhofer, Haoqi Fan, Yanghao Li, and Kaiming He. Masked autoencoders as spatiotemporal learners. arXiv preprint arXiv:2205.09113, 2022. +[29] Christoph Feichtenhofer, Axel Pinz, and Andrew Zisserman. Convolutional two-stream network fusion for video action recognition. In CVPR, 2016. + +[30] Patrick Gallinari, Yann LeCun, Sylvie Thiria, and F Fogelman Soulie. Mémoires associatives distribuées: une comparaison (distributed associative memories: a comparison). In COGNITIVA. 1987. +[31] Rohit Girdhar, João Carreira, Carl Doersch, and Andrew Zisserman. Video action transformer network. In CVPR, 2019. +[32] Rohit Girdhar and Kristen Grauman. Anticipative Video Transformer. In ICCV, 2021. +[33] Rohit Girdhar, Mannat Singh, Nikhila Ravi, Laurens van der Maaten, Armand Joulin, and Ishan Misra. Omnivore: A Single Model for Many Visual Modalities. In CVPR, 2022. +[34] Yunchao Gong, Liwei Wang, Micah Hodosh, Julia Hockenmaier, and Svetlana Lazebnik. Improving image-sentence embeddings using large weakly annotated photo collections. In ECCV, 2014. +[35] Priya Goyal, Quentin Duval, Isaac Seessel, Mathilde Caron, Mannat Singh, Ishan Misra, Levent Sagun, Armand Joulin, and Piotr Bojanowski. Vision models are more robust and fair when pretrained on uncurated images without supervision. arXiv preprint arXiv:2202.08360, 2022. +[36] Priya Goyal, Dhruv Mahajan, Abhinav Gupta, and Ishan Misra. Scaling and benchmarking self-supervised visual representation learning. In ICCV, 2019. +[37] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, Florian Hoppe, Christian Thurau, Ingo Bax, and Roland Memisevic. The “something something” video database for learning and evaluating visual common sense. In ICCV, 2017. +[38] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. NeurIPS, 2020. +[39] Raia Hadsell, Sumit Chopra, and Yann LeCun. Dimensionality reduction by learning an invariant mapping. In CVPR, 2006. +[40] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, 2022. +[41] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, 2020. +[42] Geoffrey E Hinton and Richard Zemel. Autoencoders, minimum description length and helmholtz free energy. In NeurIPS, 1993. + +[43] Elad Hoffer, Tal Ben-Nun, Itay Hubara, Niv Giladi, Torsten Hoefler, and Daniel Soudry. Augment your batch: Improving generalization through instance repetition. In CVPR, 2020. +[44] Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alex Shepard, Hartwig Adam, Pietro Perona, and Serge Belongie. The inaturalist species classification and detection dataset. In CVPR, 2018. +[45] Gao Huang, Yu Sun, Zhuang Liu, Daniel Sedra, and Kilian Q Weinberger. Deep networks with stochastic depth. In ECCV, 2016. +[46] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In ICML, 2015. +[47] Andrej Karpathy and Li Fei-Fei. Deep visual-semantic alignments for generating image descriptions. In CVPR, 2015. +[48] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, AMustafa Suleyman, and Andrew Zisserman. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. +[49] Mark A Kramer. Nonlinear principal component analysis using autoassociative neural networks. AIChE, 1991. +[50] Yann LeCun and Françoise Fogelman-Soulie. Modèles connexionnistes de l'apprentissage. Intellectica, 1987. +[51] Junnan Li, Pan Zhou, Caiming Xiong, and Steven C.H. Hoi. Prototypical contrastive learning of unsupervised representations. In ICLR, 2021. +[52] Yanghao Li, Chao-Yuan Wu, Haoqi Fan, Karttikeya Mangalam, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. Improved multiscale vision transformers for classification and detection. In CVPR, 2022. +[53] Valerii Likhosherstov, Anurag Arnab, Krzysztof Choromanski, Mario Lucic, Yi Tay, Adrian Weller, and Mostafa Dehghani. Polyvit: Co-training vision transformers on images, videos and audio. arXiv preprint arXiv:2111.12993, 2021. +[54] Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, Furu Wei, and Baining Guo. Swin transformer v2: Scaling up capacity and resolution. In CVPR, 2022. +[55] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. +[56] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swin transformer. arXiv preprint arXiv:2106.13230, 2021. + +[57] Ze Liu, Zheng Zhang, Yue Cao, Han Hu, and Xin Tong. Group-free 3d object detection via transformers. In ICCV, 2021. +[58] Jiasen Lu, Vedanuj Goswami, Marcus Rohrbach, Devi Parikh, and Stefan Lee. 12-in-1: Multi-task vision and language representation learning. In CVPR, 2020. +[59] Antoine Miech, Jean-Baptiste Alayrac, Lucas Smaira, Ivan Laptev, Josef Sivic, and Andrew Zisserman. End-to-end learning of visual representations from uncu-rated instructional videos. In CVPR, 2020. +[60] Ishan Misra, Rohit Girdhar, and Armand Joulin. An End-to-End Transformer Model for 3D Object Detection. In ICCV, 2021. +[61] Ishan Misra and Laurens van der Maaten. Self-supervised learning of pretext-invariant representations. In CVPR, 2020. +[62] Pedro Morgado, Ishan Misra, and Nuno Vasconcelos. Robust audio-visual instance discrimination. In CVPR, 2021. +[63] Pedro Morgado, Nuno Vasconcelos, and Ishan Misra. Audio-visual instance discrimination with cross-modal agreement. In CVPR, 2021. +[64] B. A. Olshausen and D. J. Field. Emergence of simple-cell receptive field properties by learning a sparse code for natural images. Nature, 1996. +[65] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. In NeurIPS, 2018. +[66] Andrew Owens and Alexei A Efros. Audio-visual scene analysis with self-supervised multisensory features. In ECCV, 2018. +[67] Niki Parmar, Ashish Vaswani, Jakob Uszkoreit, Lukasz Kaiser, Noam Shazeer, Alexander Ku, and Dustin Tran. Image transformer. In ICML, 2018. +[68] Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A Efros. Context encoders: Feature learning by inpainting. In CVPR, 2016. +[69] Boris T Polyak and Anatoli B Juditsky. Acceleration of stochastic approximation by averaging. SIAM journal on control and optimization, 1992. +[70] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. IJCV, 2015. +[71] Ruslan Salakhutdinov and Geoffrey Hinton. Deep Boltzmann machines. In AISTATS, 2009. +[72] Karen Simonyan and Andrew Zisserman. Two-stream convolutional networks for action recognition in videos. In NeurIPS, 2014. + +[73] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In CVPR, 2016. +[74] Hao Tan, Jie Lei, Thomas Wolf, and Mohit Bansal. VIMPAC: Video pre-training via masked token prediction and contrastive learning. arXiv preprint arXiv:2106.11250, 2021. +[75] Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive multiview coding. arXiv preprint arXiv:1906.05849, 2019. +[76] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. arXiv preprint arXiv:2203.12602, 2022. +[77] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In ICML, 2021. +[78] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. +[79] P. Vincent, H. Larochelle, Y. Bengio, and P.-A. Manzagol. Extracting and composing robust features with denoising autoencoders. In ICML, 2008. +[80] Rui Wang, Dongdong Chen, Zuxuan Wu, Yinpeng Chen, Xiyang Dai, Mengchen Liu, Yu-Gang Jiang, Luowei Zhou, and Lu Yuan. BEVT: Bert pretraining of video transformers. In CVPR, 2022. +[81] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In ICCV, 2021. +[82] Chen Wei, Haoqi Fan, Saining Xie, Chao-Yuan Wu, Alan Yuille, and Christoph Feichtenhofer. Masked feature prediction for self-supervised visual pre-training. In CVPR, 2022. +[83] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In CVPR, 2018. +[84] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jian-min Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In CVPR, 2022. +[85] Xueting Yan, Ishan Misra, Abhinav Gupta, Deepti Ghadiyaram, and Dhruv Mahajan. ClusterFit: Improving Generalization of Visual Representations. In CVPR, 2020. + +[86] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In ICCV, 2019. +[87] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stephane Deny. Barlow twins: Self-supervised learning via redundancy reduction. In ICML, 2021. +[88] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In ICLR, 2018. +[89] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip Torr, and Vladlen Koltun. Point transformer. In ICCV, 2021. +[90] Zhun Zhong, Liang Zheng, Guoliang Kang, Shaozi Li, and Yi Yang. Random erasing data augmentation. In AAAI, 2020. +[91] Bolei Zhou, Agata Lapedriza, Jianxiong Xiao, Antonio Torralba, and Aude Oliva. Learning deep features for scene recognition using places database. In NeurIPS, 2014. +[92] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. ibot: Image bert pre-training with online tokenizer. In ICLR, 2022. \ No newline at end of file diff --git a/2023/OmniMAE_ Single Model Masked Pretraining on Images and Videos/images.zip b/2023/OmniMAE_ Single Model Masked Pretraining on Images and Videos/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..8d43a225ee374126b837757c1f22a16e416d3e0a --- /dev/null +++ b/2023/OmniMAE_ Single Model Masked Pretraining on Images and Videos/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c84024807a096a0587a7dc82ac518ea0982b7ebd09c55c1f4c6c16cd1325587f +size 660683 diff --git a/2023/OmniMAE_ Single Model Masked Pretraining on Images and Videos/layout.json b/2023/OmniMAE_ Single Model Masked Pretraining on Images and Videos/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e4a304d8e5cb2525b94dd40882b1d1c65239a745 --- /dev/null +++ b/2023/OmniMAE_ Single Model Masked Pretraining on Images and Videos/layout.json @@ -0,0 +1,8442 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 85, + 103, + 509, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 103, + 509, + 121 + ], + "spans": [ + { + "bbox": [ + 85, + 103, + 509, + 121 + ], + "type": "text", + "content": "OmniMAE: Single Model Masked Pretraining on Images and Videos" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 155, + 143, + 436, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 143, + 436, + 184 + ], + "spans": [ + { + "bbox": [ + 155, + 143, + 436, + 184 + ], + "type": "text", + "content": "Rohit Girdhar* Alaaeldin El-Nouby* Mannat Singh* \nKalyan Vasudev Alwala* Armand Joulin Ishan Misra* \nFAIR, Meta AI" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 203, + 188, + 386, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 203, + 188, + 386, + 198 + ], + "spans": [ + { + "bbox": [ + 203, + 188, + 386, + 198 + ], + "type": "text", + "content": "https://github.com/facebookresearch/omnivore" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 72, + 225, + 521, + 369 + ], + "blocks": [ + { + "bbox": [ + 72, + 225, + 521, + 369 + ], + "lines": [ + { + "bbox": [ + 72, + 225, + 521, + 369 + ], + "spans": [ + { + "bbox": [ + 72, + 225, + 521, + 369 + ], + "type": "image", + "image_path": "af9705937884c82df0e348223ab58e9199ed22c6ae7d522c3fe23126aa586415.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 375, + 548, + 431 + ], + "lines": [ + { + "bbox": [ + 46, + 375, + 548, + 431 + ], + "spans": [ + { + "bbox": [ + 46, + 375, + 548, + 431 + ], + "type": "text", + "content": "Figure 1. OmniMAE is a single model for images and videos that is trained using masked autoencoding [40]. We use a plain Vision Transformer [24] architecture but with spatio-temporal patches as input. At training, we 'patchify' the visual input (images or videos), and feed the encoder only a subset of the patches. The decoder reconstructs the pixels for the missing patches using the encoder's output. The encoder-decoder model is trained using a pixel reconstruction loss. After training, our single plain Transformer encoder performs competitively compared to specialized architectures on downstream image and video recognition tasks." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 441, + 192, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 441, + 192, + 453 + ], + "spans": [ + { + "bbox": [ + 143, + 441, + 192, + 453 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 466, + 290, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 466, + 290, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 466, + 290, + 693 + ], + "type": "text", + "content": "Transformer-based architectures have become competitive across a variety of visual domains, most notably images and videos. While prior work studies these modalities in isolation, having a common architecture suggests that one can train a single unified model for multiple visual modalities. Prior attempts at unified modeling typically use architectures tailored for vision tasks, or obtain worse performance compared to single modality models. In this work, we show that masked autoencoding can be used to train a simple Vision Transformer on images and videos, without requiring any labeled data. This single model learns visual representations that are comparable to or better than single-modality representations on both image and video benchmarks, while using a much simpler architecture. Furthermore, this model can be learned by dropping " + }, + { + "bbox": [ + 46, + 466, + 290, + 693 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 46, + 466, + 290, + 693 + ], + "type": "text", + "content": " of the image and " + }, + { + "bbox": [ + 46, + 466, + 290, + 693 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 46, + 466, + 290, + 693 + ], + "type": "text", + "content": " of the video patches, enabling extremely fast training of huge model architectures. In particular, we show that our single ViT-Huge model can be finetuned to achieve " + }, + { + "bbox": [ + 46, + 466, + 290, + 693 + ], + "type": "inline_equation", + "content": "86.6\\%" + }, + { + "bbox": [ + 46, + 466, + 290, + 693 + ], + "type": "text", + "content": " on ImageNet and " + }, + { + "bbox": [ + 46, + 466, + 290, + 693 + ], + "type": "inline_equation", + "content": "75.5\\%" + }, + { + "bbox": [ + 46, + 466, + 290, + 693 + ], + "type": "text", + "content": " on the challenging Something Something-v2 video" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 441, + 476, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 441, + 476, + 453 + ], + "spans": [ + { + "bbox": [ + 306, + 441, + 476, + 453 + ], + "type": "text", + "content": "benchmark, setting a new state-of-the-art." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 477, + 386, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 477, + 386, + 490 + ], + "spans": [ + { + "bbox": [ + 306, + 477, + 386, + 490 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 497, + 548, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 497, + 548, + 688 + ], + "spans": [ + { + "bbox": [ + 304, + 497, + 548, + 688 + ], + "type": "text", + "content": "The Transformer architecture [78] is rapidly becoming competitive across the different visual modalities in Computer Vision, from images [24, 27, 55, 77], to 3D [57, 60, 89] and videos [2, 9, 27, 31, 32, 56]. This convergence toward a unified architecture naturally suggests that we should be able to train a single model that works across different visual modalities. However, recent attempts to train unified models either lead to worse performance compared to single modality models [53], or require the use of an alternative architecture [33], namely the Swin Transformer [55], with inductive biases tailored towards vision tasks. While specialized Transformer architectures for vision [27, 55, 56, 81] can offer better performance for visual modalities, they lose the generality and flexibility of the vanilla Transformer, making it harder to later model different domains like text, speech, 3D etc. in multi-modal architectures." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 547, + 714 + ], + "type": "text", + "content": "In this work, we train a single vanilla Transformer that works for both images and videos, as illustrated in Figure 1." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 703, + 162, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 703, + 162, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 703, + 162, + 713 + ], + "type": "text", + "content": "*Equal technical contribution." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "10406" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": "To this end, we leverage the findings of several recent works on the use of the masked pretraining [23] to greatly improve the training and performance of Transformers in the domain of images [6, 40, 80, 84], videos [76, 80, 82] or across text, audio and images [5]. We show that this masked pretraining is a viable strategy to pretrain a unified 'omnivorous' Transformer across visual modalities. In particular, we consider the Masked Auto-Encoding (MAE) approach [40] to train an Omnivorous visual encoder [33]. The resulting OmniMAE model learns from all the modalities with the same objective function and does not require any supervision." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 213, + 289, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 213, + 289, + 393 + ], + "spans": [ + { + "bbox": [ + 46, + 213, + 289, + 393 + ], + "type": "text", + "content": "Using a masked pretraining objective has several advantages over supervised objectives [33, 53] or discriminative self-supervised objectives [13, 17, 41]. First, as opposed to supervised objectives, a general-purpose unsupervised loss does not require any human labeling effort. As a result, it is robust to biases introduced by a predefined set of labels [35]. Moreover, it does not require a multi-head architecture to incorporate supervision from each of the label spaces corresponding to each modality, which is hard to maintain and scale with new modalities. Second, although discriminative self-supervised methods produce superior frozen features compared to reconstruction objectives, they are non trivial to scale in model and data size [36]. Our masked pretraining objective is simple, efficient to train, and scales to different visual modalities with minimal changes." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 402, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 402, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 402, + 289, + 715 + ], + "type": "text", + "content": "Contributions. (1) We show that the simple Vision Transformer architecture (ViT) [24] originally designed for images can naturally be applied on videos, and videos and images jointly. OmniMAE is a single ViT-based model for videos and images that outperforms architectures and models specifically designed for either modality. (2) Prior and concurrent work design self-supervised methods and architectures for either image or video and we find that these models do not transfer well across modalities. OmniMAE is the first single self-supervised model that achieves good performance on both modalities. (3) We show that our joint training using both images and videos enables us to use much higher masking ratios than any prior work for training MAE. Since ViT can processes only the non-masked input, we train OmniMAE models with only " + }, + { + "bbox": [ + 46, + 402, + 289, + 715 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 402, + 289, + 715 + ], + "type": "text", + "content": " of image and " + }, + { + "bbox": [ + 46, + 402, + 289, + 715 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 46, + 402, + 289, + 715 + ], + "type": "text", + "content": " of video patches. This enables us to train large (650M parameter) models with a " + }, + { + "bbox": [ + 46, + 402, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\sim 7\\times" + }, + { + "bbox": [ + 46, + 402, + 289, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 402, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\sim 11\\times" + }, + { + "bbox": [ + 46, + 402, + 289, + 715 + ], + "type": "text", + "content": " reduction in compute and memory on images and videos. (4) Finally, we propose improvements to the MAE training. We show that repeating samples in a mini-batch reduces dataloading (and thus training) time without loss in final transfer performance. Sample replication is particularly useful for masked pretraining as the unmasked patches are different across sample replicas. We also show that using a shallow shared decoder for videos and images leads to better performance while reducing the number of parameters by " + }, + { + "bbox": [ + 46, + 402, + 289, + 715 + ], + "type": "inline_equation", + "content": "2 - 4\\times" + }, + { + "bbox": [ + 46, + 402, + 289, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 306, + 71, + 392, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 392, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 392, + 84 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 91, + 547, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 547, + 270 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 547, + 270 + ], + "type": "text", + "content": "Our work builds upon research in self-supervised learning, masked pretraining and unified modeling in computer vision. Self-supervised learning. In recent years, self-supervised approaches have been dominated by joint embedding methods which can rely on different objectives including contrastive [17,39,41,51,61,65,75,83], non-contrastive [7,18,26, 87], clustering [3, 11, 12, 85] or self-distillation [5, 13, 38, 92]. Such methods are trained to learn invariance to a set of predefined transformations which results in image descriptors with a strong linear probing and KNN performance. However, such methods can be challenging to scale since they can suffer from instabilities [19]. Additionally, the strongest performance is typically obtained with the help of augmentations like multi-crop [12, 13, 92] which can be hard to apply at scale due their compute and memory overhead." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 270, + 548, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 270, + 548, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 270, + 548, + 628 + ], + "type": "text", + "content": "Masked pretraining. We build upon masked prediction methods where the representation is learned by predicting masked parts of the input. Such methods have recently gained popularity given their immense success in NLP. In particular, BERT [23] showed that masked language modeling by predicting a subset of masked words is a powerful pre-training objective and leads to impressive finetuning performance on various downstream tasks. In computer vision, input reconstruction methods have a rich history with non-linear PCA [49], sparse reconstruction [64], autoencoders [10, 30, 42, 50], RBMs [71] etc. Masked prediction methods can be viewed as a special case of denoising autoencoders [30, 79] where the input 'noise' is a masking function. An example of such a method that uses masking as noise is context encoders [68]. With the recent and rapid rise of Vision Transformers [24], masked prediction was revisited by multiple efforts. SiT [4] replaces variable sized patches in the image with random noise and trains a ViT model for reconstruction, among other objectives. BEiT [6] moves a step closer to BERT with replacing full patches with mask tokens and training a ViT encoder to predict the discrete visual words of masked patches using a cross-entropy loss. Masked prediction has also shown impressive performance for specialized vision transformer architectures such as MViT [27, 52, 82] and Swin transformer [54, 55, 84]. SimMIM [84] and MaskedFeat [82] predict pixel values and HOG features of the masked patches using a Swin-v2 [54] and MViT-v2 [52] backbones respectively. Finally, SplitMask [25] studied the interesting properties of masked prediction methods in terms of high sample efficiency." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "content": "Of particular interest to OmniMAE, masked autoencoders (MAE) [40] demonstrated impressive scaling properties by utilizing a patch dropping strategy for masked patches accompanied with a high masking ratio of " + }, + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "content": ". Under this setting, the encoder only process a small subset of the image patches followed by a relatively low capacity decoder which reconstructs the image pixels. This property is even more" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "10407" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": "crucial for video representation learning given the large number of patches, and a concurrent work, ST-MAE [76], shows that MAE pretraining with an even higher masking ratio of " + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 46, + 72, + 289, + 228 + ], + "type": "text", + "content": " works well and obtains strong finetuning performance on downstream video recognition benchmarks. Notably, the efficiency gained by patch dropping is specific to vanilla ViTs, and multiscale architectures such as MViT and Swin are unable to benefit due to their design. Hence, we use the simple ViT as our architecture and show that it can be trained efficiently and jointly for images and videos using extremely high masking ratios (90-95% on both modalities), and yet perform competitively to specialized architectures like MViT [27] and Swin [55]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 237, + 289, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 237, + 289, + 538 + ], + "spans": [ + { + "bbox": [ + 46, + 237, + 289, + 538 + ], + "type": "text", + "content": "Unified modeling and multi-modal learning. Multi-modal learning in computer vision has a long history that includes training using images and text [15, 34, 47, 58, 59], video and optical flow [29, 72], and video and audio [1, 62, 63, 66]. The majority of such methods rely on training a separate backbone for each modality as well as the availability of alignment across modalities. More recently, Omnivore [33] was proposed for joint training of multiple modalities like images, videos and single-view 3D, attaining a strong performance in each of the modality specific benchmarks with a single shared trunk. PolyViT [53] co-trains a shared transformer encoder using images, videos and audio data and provides a competitive performance on various downstream tasks for each modality. The aforementioned methods differ compared to OmniMAE in that they are trained with supervised learning and require human annotations. BEVT [80] tackles BERT pre-training for videos and proposes that jointly pretraining using static images improves the finetuning performance of video recognition benchmarks. Unlike OmniMAE, BEVT uses the specialized Swin transformer architecture with separate decoder heads for images and videos. Thus, BEVT cannot drop patches while training which can limit its scalability. Furthermore, it relies on a tokenizer which must be trained apriori, and the tokenizer training itself can affect the model's performance." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 576, + 119, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 576, + 119, + 588 + ], + "spans": [ + { + "bbox": [ + 47, + 576, + 119, + 588 + ], + "type": "text", + "content": "3. OmniMAE" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 605, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 713 + ], + "type": "text", + "content": "Our goal is to pretrain a single unified model for images and videos. Rather than use specialized architectures tailored for a visual modality, we build upon the vanilla Vision Transformer (ViT) [24] architecture that has limited inductive biases for vision. For pretraining, we extend the simple self-supervised masked auto-encoding (MAE) approach [40]. The original architecture and pretraining method are tested only on images, and we show simple design decisions for a unified model." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 95 + ], + "type": "text", + "content": "3.1. Training OmniMAE jointly on images and videos" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 102, + 547, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 102, + 547, + 210 + ], + "spans": [ + { + "bbox": [ + 304, + 102, + 547, + 210 + ], + "type": "text", + "content": "We illustrate our method in Figure 1. For pretraining, we use an encoder-decoder architecture where the encoder only operates on a 'non-masked' subset of the input. The decoder predicts the pixel values for the entire input, i.e., masked and non-masked pixels. The model is trained to minimize the reconstruction error for the masked (unseen) part of the input. After pretraining, we evaluate the encoder by transfer learning (the decoder is discarded). Next, we describe the pretraining details." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 210, + 547, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 210, + 547, + 293 + ], + "spans": [ + { + "bbox": [ + 304, + 210, + 547, + 293 + ], + "type": "text", + "content": "Images and videos as spatio-temporal patches. The input image or video can be represented as a 4D tensor of shape " + }, + { + "bbox": [ + 304, + 210, + 547, + 293 + ], + "type": "inline_equation", + "content": "T \\times H \\times W \\times 3" + }, + { + "bbox": [ + 304, + 210, + 547, + 293 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 210, + 547, + 293 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 304, + 210, + 547, + 293 + ], + "type": "text", + "content": " is the temporal dimension and " + }, + { + "bbox": [ + 304, + 210, + 547, + 293 + ], + "type": "inline_equation", + "content": "H, W" + }, + { + "bbox": [ + 304, + 210, + 547, + 293 + ], + "type": "text", + "content": " are the spatial dimensions, and 3 represents the color channels. We treat images as being single-frame videos with " + }, + { + "bbox": [ + 304, + 210, + 547, + 293 + ], + "type": "inline_equation", + "content": "T = 1" + }, + { + "bbox": [ + 304, + 210, + 547, + 293 + ], + "type": "text", + "content": ". The input is split into " + }, + { + "bbox": [ + 304, + 210, + 547, + 293 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 210, + 547, + 293 + ], + "type": "text", + "content": " spatio-temporal patches, each of size " + }, + { + "bbox": [ + 304, + 210, + 547, + 293 + ], + "type": "inline_equation", + "content": "t \\times h \\times w \\times 3" + }, + { + "bbox": [ + 304, + 210, + 547, + 293 + ], + "type": "text", + "content": " [33]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 293, + 547, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 293, + 547, + 402 + ], + "spans": [ + { + "bbox": [ + 304, + 293, + 547, + 402 + ], + "type": "text", + "content": "Omnivorous visual encoder. We use an omnivorous [33] visual encoder that processes both images and video using the same parameters. The encoder operates on the " + }, + { + "bbox": [ + 304, + 293, + 547, + 402 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 293, + 547, + 402 + ], + "type": "text", + "content": " spatiotemporal patches from the images and videos. The encoder can naturally handle variable number " + }, + { + "bbox": [ + 304, + 293, + 547, + 402 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 293, + 547, + 402 + ], + "type": "text", + "content": " of patches from images and videos as it uses the Transformer architecture [78]. The encoder shares the same parameters for both image and video inputs and processes them in the same way to output per-patch embeddings." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 402, + 547, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 402, + 547, + 498 + ], + "spans": [ + { + "bbox": [ + 304, + 402, + 547, + 498 + ], + "type": "text", + "content": "Pretraining. We convert the input into " + }, + { + "bbox": [ + 304, + 402, + 547, + 498 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 402, + 547, + 498 + ], + "type": "text", + "content": " spatio-temporal patches and randomly mask " + }, + { + "bbox": [ + 304, + 402, + 547, + 498 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 402, + 547, + 498 + ], + "type": "text", + "content": " of them. Following [40], we feed the non-masked " + }, + { + "bbox": [ + 304, + 402, + 547, + 498 + ], + "type": "inline_equation", + "content": "N - M" + }, + { + "bbox": [ + 304, + 402, + 547, + 498 + ], + "type": "text", + "content": " patches (and their positions) to the encoder to produce per-patch embeddings. The encoder output is concatenated with " + }, + { + "bbox": [ + 304, + 402, + 547, + 498 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 402, + 547, + 498 + ], + "type": "text", + "content": " copies of a learnable 'mask token' to obtain " + }, + { + "bbox": [ + 304, + 402, + 547, + 498 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 402, + 547, + 498 + ], + "type": "text", + "content": " embeddings. We add the appropriate positional encoding to the " + }, + { + "bbox": [ + 304, + 402, + 547, + 498 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 402, + 547, + 498 + ], + "type": "text", + "content": " embeddings and feed them into the decoder, which outputs " + }, + { + "bbox": [ + 304, + 402, + 547, + 498 + ], + "type": "inline_equation", + "content": "N \\times t \\times h \\times w \\times 3" + }, + { + "bbox": [ + 304, + 402, + 547, + 498 + ], + "type": "text", + "content": " pixel values." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 498, + 547, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 547, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 547, + 617 + ], + "type": "text", + "content": "Loss function and optimization. We minimize the reconstruction error between the decoder predictions and the input pixel values. The input pixel values are normalized to zero mean and unit variance [46] to get the targets for the loss. We minimize the " + }, + { + "bbox": [ + 304, + 498, + 547, + 617 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 304, + 498, + 547, + 617 + ], + "type": "text", + "content": " distance between the predictions and targets over the " + }, + { + "bbox": [ + 304, + 498, + 547, + 617 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 304, + 498, + 547, + 617 + ], + "type": "text", + "content": " masked patches. At each point in the training, we sample a mini-batch of either images or video and compute the loss using the decoder predictions. We study the effect of different mini-batch construction strategies on the overall performance and speed of training in § 4.2." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 618, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 547, + 713 + ], + "type": "text", + "content": "Dataset sample replication. Since we operate on a small subset of the input patches " + }, + { + "bbox": [ + 304, + 618, + 547, + 713 + ], + "type": "inline_equation", + "content": "(M \\ll N)" + }, + { + "bbox": [ + 304, + 618, + 547, + 713 + ], + "type": "text", + "content": ", we find that our data reading speed is unable to keep up with the number of data points we can theoretically process in a single optimization step. This issue is even more pronounced for videos, where we spend significant compute to read and decode a full video, only to discard " + }, + { + "bbox": [ + 304, + 618, + 547, + 713 + ], + "type": "inline_equation", + "content": ">90\\%" + }, + { + "bbox": [ + 304, + 618, + 547, + 713 + ], + "type": "text", + "content": " of it. Inspired from the success of RepeatAugment, which shows that replicating samples" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10408" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 68, + 545, + 182 + ], + "blocks": [ + { + "bbox": [ + 52, + 68, + 545, + 182 + ], + "lines": [ + { + "bbox": [ + 52, + 68, + 545, + 182 + ], + "spans": [ + { + "bbox": [ + 52, + 68, + 545, + 182 + ], + "type": "image", + "image_path": "d1ac4e9640794e89298e95775d504095633419cb539bbb20673e66969149a038.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 192, + 548, + 248 + ], + "lines": [ + { + "bbox": [ + 46, + 192, + 548, + 248 + ], + "spans": [ + { + "bbox": [ + 46, + 192, + 548, + 248 + ], + "type": "text", + "content": "Figure 2. OmniMAE on image and video downstream tasks. We finetune the MAE, ST-MAE, and OmniMAE models on image and video benchmarks. We use the ViT architecture with two model sizes: ViT-B and ViT-L. MAE has poor video recognition performance while ST-MAE's performance drops on image datasets. OmniMAE pretraining generalizes to both benchmarks. All models are trained for 800 epochs on the pretraining datasets. The image-only MAE model is inflated [14] to apply MAE to video recognition tasks. The input image is replicated to apply ST-MAE to image recognition benchmarks." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 268, + 290, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 268, + 290, + 354 + ], + "spans": [ + { + "bbox": [ + 46, + 268, + 290, + 354 + ], + "type": "text", + "content": "within a mini-batch is an effective technique to improve generalization [8, 12, 43], we replicate a single data point and apply different masks to it each time. Even with replicated samples, the non-masked input to the model is different due to random cropping and masking being different across the replicas. We show in § 4.2 that sample replication leads to gains in runtime without affecting the final transfer accuracy." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 361, + 180, + 373 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 361, + 180, + 373 + ], + "spans": [ + { + "bbox": [ + 47, + 361, + 180, + 373 + ], + "type": "text", + "content": "3.2. Implementation Details" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 380, + 288, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 380, + 288, + 403 + ], + "spans": [ + { + "bbox": [ + 46, + 380, + 288, + 403 + ], + "type": "text", + "content": "We note some of the salient implementation details and provide the complete details in Appendix B." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 404, + 287, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 404, + 287, + 523 + ], + "spans": [ + { + "bbox": [ + 46, + 404, + 287, + 523 + ], + "type": "text", + "content": "Architecture. We use the ViT [24] architecture for the omnivorous encoder and experiment with its ViT-B, ViT-L, and ViT-H variants. We do not use the [CLS] token in the ViT models, yielding a small improvement in runtime without loss in performance. We use a Transformer decoder with 4 layers (8 for ViT-H) of 384, 512, and 512 embedding dimensions for ViT-B, ViT-L, and ViT-H, respectively. The decoder outputs the RGB colors for the pixels in all the input patches. We use sinusoidal positional encoding [78] for the patches in both the encoder and the decoder." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 525, + 287, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 525, + 287, + 597 + ], + "spans": [ + { + "bbox": [ + 46, + 525, + 287, + 597 + ], + "type": "text", + "content": "Training details. We train our models with a mini-batch size of 2048. We resize the input images and videos spatially to " + }, + { + "bbox": [ + 46, + 525, + 287, + 597 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 46, + 525, + 287, + 597 + ], + "type": "text", + "content": " pixels. For videos, we sample a clip of " + }, + { + "bbox": [ + 46, + 525, + 287, + 597 + ], + "type": "inline_equation", + "content": "T = 16" + }, + { + "bbox": [ + 46, + 525, + 287, + 597 + ], + "type": "text", + "content": " frames at 6 FPS. We use a patch size of " + }, + { + "bbox": [ + 46, + 525, + 287, + 597 + ], + "type": "inline_equation", + "content": "2 \\times 16 \\times 16" + }, + { + "bbox": [ + 46, + 525, + 287, + 597 + ], + "type": "text", + "content": " for ViT-B and ViT-L, and " + }, + { + "bbox": [ + 46, + 525, + 287, + 597 + ], + "type": "inline_equation", + "content": "2 \\times 14 \\times 14" + }, + { + "bbox": [ + 46, + 525, + 287, + 597 + ], + "type": "text", + "content": " for ViT-H. Images are replicated temporally to meet the patch size." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 597, + 288, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 597, + 288, + 657 + ], + "spans": [ + { + "bbox": [ + 46, + 597, + 288, + 657 + ], + "type": "text", + "content": "Masking the input. Compared to prior work, we use extremely high masking for pretraining and only " + }, + { + "bbox": [ + 46, + 597, + 288, + 657 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 597, + 288, + 657 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 597, + 288, + 657 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 46, + 597, + 288, + 657 + ], + "type": "text", + "content": " of the image and video patches are fed to the encoder for pretraining. We uniform randomly mask the patches for images and videos and ablate the masking hyperparameters in " + }, + { + "bbox": [ + 46, + 597, + 288, + 657 + ], + "type": "inline_equation", + "content": "\\S 4.2" + }, + { + "bbox": [ + 46, + 597, + 288, + 657 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 669, + 128, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 669, + 128, + 682 + ], + "spans": [ + { + "bbox": [ + 47, + 669, + 128, + 682 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 288, + 715 + ], + "type": "text", + "content": "Pretraining data. We pretrain representations by jointly training on images from the ImageNet (IN1K) [70] dataset" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 268, + 546, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 268, + 546, + 316 + ], + "spans": [ + { + "bbox": [ + 304, + 268, + 546, + 316 + ], + "type": "text", + "content": "and videos from the Something Something-v2 (SSv2) [37] dataset. We choose the SSv2 dataset over web video datasets like Kinetics-400 (K400) [48] for reproducibility, and given SSv2's challenging nature requiring temporal reasoning." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 316, + 547, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 316, + 547, + 424 + ], + "spans": [ + { + "bbox": [ + 304, + 316, + 547, + 424 + ], + "type": "text", + "content": "Transfer learning datasets. We consider two image datasets: iNaturalist-2018 (iNat18) [44], a popular fine-grained recognition dataset, and Places-365 (P365) [91], a scene recognition benchmark. For video datasets, we focus on the popular K400 action recognition benchmark, and EPIC-Kitchens-100 (EK100) [22], an egocentric action recognition dataset. We report the top-1 classification accuracy for all transfer datasets. The details about the datasets can be found in Appendix A." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 431, + 537, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 431, + 537, + 445 + ], + "spans": [ + { + "bbox": [ + 305, + 431, + 537, + 445 + ], + "type": "text", + "content": "4.1. OmniMAE on Image and Video Recognition" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 450, + 547, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 450, + 547, + 474 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 547, + 474 + ], + "type": "text", + "content": "We now evaluate the capabilities of OmniMAE's representations on both image and video recognition tasks." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 475, + 547, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 475, + 547, + 569 + ], + "spans": [ + { + "bbox": [ + 304, + 475, + 547, + 569 + ], + "type": "text", + "content": "Baselines. We compare OmniMAE's representations trained jointly on images and videos to MAE, trained solely on images [40]. Additionally, we develop a video-only baseline, SpatioTemporal-MAE (ST-MAE), by training OmniMAE only on videos. For a fair comparison, all models are pretrained for 800 epochs on their respective datasets. The image-only MAE model is inflated [14] for finetuning on video tasks, see Appendix B.2 for details." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 570, + 548, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 548, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 548, + 714 + ], + "type": "text", + "content": "Observations. Figure 2 presents the evaluation results for all the models on the image and video benchmarks. The modality specific MAE and ST-MAE models achieve good transfer performance when transferring to datasets which match the pretraining modality. However, both models show a degradation in performance when transferring to a different visual modality. MAE has lower video recognition performance, especially on egocentric videos from EK100 where the smaller ViT-B MAE model is " + }, + { + "bbox": [ + 304, + 570, + 548, + 714 + ], + "type": "inline_equation", + "content": "18.4\\%" + }, + { + "bbox": [ + 304, + 570, + 548, + 714 + ], + "type": "text", + "content": " worse compared to ST-MAE. Similarly, compared to MAE, ST-MAE is " + }, + { + "bbox": [ + 304, + 570, + 548, + 714 + ], + "type": "inline_equation", + "content": "8.7\\%" + }, + { + "bbox": [ + 304, + 570, + 548, + 714 + ], + "type": "text", + "content": " worse on the fine-grained classification task of iNat18. With a large model, MAE's performance on the image recognition" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "10409" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 69, + 145, + 149 + ], + "blocks": [ + { + "bbox": [ + 72, + 69, + 145, + 149 + ], + "lines": [ + { + "bbox": [ + 72, + 69, + 145, + 149 + ], + "spans": [ + { + "bbox": [ + 72, + 69, + 145, + 149 + ], + "type": "image", + "image_path": "a1eb0f78d4fcbcc4d87c895bfbb0346df6d33657707c0c6c292d228dfcc0067c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 162, + 547, + 184 + ], + "lines": [ + { + "bbox": [ + 46, + 162, + 547, + 184 + ], + "spans": [ + { + "bbox": [ + 46, + 162, + 547, + 184 + ], + "type": "text", + "content": "Figure 3. Different types of masking for images (left two) and videos. Causal and tube masking use the data's spatio-temporal structure. Random frame masking randomly masks frames in a video. Random masking randomly masks patches and is used by default for OmniMAE." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 149, + 70, + 219, + 150 + ], + "blocks": [ + { + "bbox": [ + 149, + 70, + 219, + 150 + ], + "lines": [ + { + "bbox": [ + 149, + 70, + 219, + 150 + ], + "spans": [ + { + "bbox": [ + 149, + 70, + 219, + 150 + ], + "type": "image", + "image_path": "e01de3061587e5a19d129ae2652a3c26da65b7a04d12868d46af82a2d4fa961c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 227, + 70, + 294, + 150 + ], + "blocks": [ + { + "bbox": [ + 227, + 70, + 294, + 150 + ], + "lines": [ + { + "bbox": [ + 227, + 70, + 294, + 150 + ], + "spans": [ + { + "bbox": [ + 227, + 70, + 294, + 150 + ], + "type": "image", + "image_path": "1c6238b96d5ab1b44f95246d7e7b204b70e8aa02adc30ab3802282f60ba077da.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 299, + 70, + 365, + 150 + ], + "blocks": [ + { + "bbox": [ + 299, + 70, + 365, + 150 + ], + "lines": [ + { + "bbox": [ + 299, + 70, + 365, + 150 + ], + "spans": [ + { + "bbox": [ + 299, + 70, + 365, + 150 + ], + "type": "image", + "image_path": "0df45b5cc7f205ade1731b919b40ecf46e1e6ee18b3660ed91fecd37458a9132.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 370, + 70, + 437, + 150 + ], + "blocks": [ + { + "bbox": [ + 370, + 70, + 437, + 150 + ], + "lines": [ + { + "bbox": [ + 370, + 70, + 437, + 150 + ], + "spans": [ + { + "bbox": [ + 370, + 70, + 437, + 150 + ], + "type": "image", + "image_path": "cafecf3af80989efae0adcac0170a585578b07837ebec5ff90539675851fd9cc.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 444, + 70, + 518, + 151 + ], + "blocks": [ + { + "bbox": [ + 444, + 70, + 518, + 151 + ], + "lines": [ + { + "bbox": [ + 444, + 70, + 518, + 151 + ], + "spans": [ + { + "bbox": [ + 444, + 70, + 518, + 151 + ], + "type": "image", + "image_path": "cb30810cf07f8315653191565f6939e5ab2707f672c6cdc5b122fb735c911a17.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 205, + 287, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 205, + 287, + 240 + ], + "spans": [ + { + "bbox": [ + 46, + 205, + 287, + 240 + ], + "type": "text", + "content": "benchmarks improves but the cross-modality performance degrades further. On EPIC-Kitchens-100 the large MAE model is more than " + }, + { + "bbox": [ + 46, + 205, + 287, + 240 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 46, + 205, + 287, + 240 + ], + "type": "text", + "content": " worse than ST-MAE." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 243, + 287, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 243, + 287, + 387 + ], + "spans": [ + { + "bbox": [ + 46, + 243, + 287, + 387 + ], + "type": "text", + "content": "When transferred to both image and video recognition tasks, OmniMAE performs favorably to the single modality baselines that use exactly the same architecture. OmniMAE also uses the same finetuning recipe as the single modality baselines. OmniMAE matches the video classification performance of ST-MAE and the image classification performance of MAE for both ViT-B and ViT-L, with OmniMAE's performance improving on both image and video recognition benchmarks with the larger model. These results suggest that our unified model of images and videos serves as a better pretraining across a wider variety of recognition tasks than single modality models." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 389, + 289, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 389, + 289, + 534 + ], + "spans": [ + { + "bbox": [ + 46, + 389, + 289, + 534 + ], + "type": "text", + "content": "Qualitative results. Following [40], we re-train OmniMAE without normalizing the pixel targets to obtain easy to visualize RGB reconstructions. We visualize the predictions in Figure 4 using samples that are unseen during training: either val sets or different datasets altogether. OmniMAE makes reasonable predictions on the in-distribution ImageNet and SSv2 val sets, as well as the unseen, out-of-distribution K400 and EK100 datasets. As expected, the details in the reconstruction decrease when increasing the masking ratio. However, even with " + }, + { + "bbox": [ + 46, + 389, + 289, + 534 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 46, + 389, + 289, + 534 + ], + "type": "text", + "content": " masking, the model can reconstruct coarse details in the input, e.g., in ImageNet, the model reconstructs the coarse structure of the flower, vehicle, dog etc." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 52, + 569, + 164, + 646 + ], + "blocks": [ + { + "bbox": [ + 52, + 569, + 164, + 646 + ], + "lines": [ + { + "bbox": [ + 52, + 569, + 164, + 646 + ], + "spans": [ + { + "bbox": [ + 52, + 569, + 164, + 646 + ], + "type": "image", + "image_path": "3da07f163da4192e2af7cb84ade3b8943096683fa66e96192de72f49af0fb792.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 654, + 287, + 710 + ], + "lines": [ + { + "bbox": [ + 46, + 654, + 287, + 710 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 287, + 710 + ], + "type": "text", + "content": "Figure 5. Sample replication. We study the effect of repeating samples while training our model. In each case, we repeat a sample " + }, + { + "bbox": [ + 46, + 654, + 287, + 710 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 654, + 287, + 710 + ], + "type": "text", + "content": " times within a mini-batch while fixing the overall mini-batch size and training updates. Replication leads to improved training speeds, especially on video without affecting the final performance." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 167, + 571, + 280, + 645 + ], + "blocks": [ + { + "bbox": [ + 167, + 571, + 280, + 645 + ], + "lines": [ + { + "bbox": [ + 167, + 571, + 280, + 645 + ], + "spans": [ + { + "bbox": [ + 167, + 571, + 280, + 645 + ], + "type": "image", + "image_path": "3437c6757e850f8d9b1e98baeab30c0df4a211cb589bfb6a02bb6f907a6ec27a.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 205, + 430, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 205, + 430, + 217 + ], + "spans": [ + { + "bbox": [ + 306, + 205, + 430, + 217 + ], + "type": "text", + "content": "4.2. Analyzing OmniMAE" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 227, + 547, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 227, + 547, + 370 + ], + "spans": [ + { + "bbox": [ + 304, + 227, + 547, + 370 + ], + "type": "text", + "content": "We train the ViT-B architecture jointly on the ImageNet and the SSv2 datasets for 800 epochs, where an epoch involves training over all samples from both the datasets. For these analysis experiments, we use the default masking hyperparameters of MAE for images (75%). For ST-MAE, due to redundancy across frames, we use \"tube\" dropping (i.e. dropping all patches at a given spatial location over time) with high masking ratio (90%) All hyperparameters for these experiments are in Appendix B.3. We evaluate all the pretrained models on ImageNet and SSv2 and report the top-1 classification accuracies in Table 1. Please see Appendix D for additional ablations." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 374, + 547, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 374, + 547, + 530 + ], + "spans": [ + { + "bbox": [ + 304, + 374, + 547, + 530 + ], + "type": "text", + "content": "Extreme masking. We vary the ratio of masked patches in the input while training the model. We observe that videos benefit from significantly higher amounts of masking than images. Since video frames have a lot of redundant information, the resulting spatio-temporal patches are also redundant and thus higher masking leads to better self-supervised models. Unlike our experiments, prior work found that extremely high masking ratios lead to degradation in performance, for instance MAE [40] saw a significant degradation in performance when masking more than " + }, + { + "bbox": [ + 304, + 374, + 547, + 530 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 304, + 374, + 547, + 530 + ], + "type": "text", + "content": " of the patches. However, as we show in Table 1a, OmniMAE models can use extremely high masking of the input while learning good representations." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 534, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 547, + 713 + ], + "type": "text", + "content": "Reduced compute due to extreme masking. Our models trained with " + }, + { + "bbox": [ + 304, + 534, + 547, + 713 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 304, + 534, + 547, + 713 + ], + "type": "text", + "content": " masking on images and " + }, + { + "bbox": [ + 304, + 534, + 547, + 713 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 304, + 534, + 547, + 713 + ], + "type": "text", + "content": " masking on videos yield good performance, while being trained with just 19 and 78 unmasked image and video patches respectively, for a patch size of " + }, + { + "bbox": [ + 304, + 534, + 547, + 713 + ], + "type": "inline_equation", + "content": "2 \\times 16 \\times 16" + }, + { + "bbox": [ + 304, + 534, + 547, + 713 + ], + "type": "text", + "content": ". As we follow [40] and only pass unmasked patches to the encoder and have a lightweight decoder, extreme masking leads to a dramatically lower computational cost for training the encoder, and consequently the model as a whole. Compared to using all the patches, our masked autoencoder uses " + }, + { + "bbox": [ + 304, + 534, + 547, + 713 + ], + "type": "inline_equation", + "content": "5.9 \\times" + }, + { + "bbox": [ + 304, + 534, + 547, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 534, + 547, + 713 + ], + "type": "inline_equation", + "content": "7.8 \\times" + }, + { + "bbox": [ + 304, + 534, + 547, + 713 + ], + "type": "text", + "content": " fewer FLOPS for ViT-B on images and videos respectively, " + }, + { + "bbox": [ + 304, + 534, + 547, + 713 + ], + "type": "inline_equation", + "content": "7.1 \\times" + }, + { + "bbox": [ + 304, + 534, + 547, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 534, + 547, + 713 + ], + "type": "inline_equation", + "content": "11.6 \\times" + }, + { + "bbox": [ + 304, + 534, + 547, + 713 + ], + "type": "text", + "content": " for ViT-L, and " + }, + { + "bbox": [ + 304, + 534, + 547, + 713 + ], + "type": "inline_equation", + "content": "7.2 \\times" + }, + { + "bbox": [ + 304, + 534, + 547, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 534, + 547, + 713 + ], + "type": "inline_equation", + "content": "11.3 \\times" + }, + { + "bbox": [ + 304, + 534, + 547, + 713 + ], + "type": "text", + "content": " for ViT-H. Compared to MAE, on ViT-B, ViT-L and ViT-H, our higher masking leads to " + }, + { + "bbox": [ + 304, + 534, + 547, + 713 + ], + "type": "inline_equation", + "content": "1.8 \\times" + }, + { + "bbox": [ + 304, + 534, + 547, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 534, + 547, + 713 + ], + "type": "inline_equation", + "content": "2.0 \\times" + }, + { + "bbox": [ + 304, + 534, + 547, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 534, + 547, + 713 + ], + "type": "inline_equation", + "content": "2.0 \\times" + }, + { + "bbox": [ + 304, + 534, + 547, + 713 + ], + "type": "text", + "content": " fewer FLOPS on images. On videos, compared to some concurrent works [28, 76] that use " + }, + { + "bbox": [ + 304, + 534, + 547, + 713 + ], + "type": "inline_equation", + "content": "90\\%" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10410" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 51, + 70, + 138, + 153 + ], + "blocks": [ + { + "bbox": [ + 51, + 70, + 138, + 153 + ], + "lines": [ + { + "bbox": [ + 51, + 70, + 138, + 153 + ], + "spans": [ + { + "bbox": [ + 51, + 70, + 138, + 153 + ], + "type": "table", + "html": "
RatiosAccuracy
IN1KSSv2IN1KSSv2
75%75%82.667.2
75%90%82.867.9
75%95%82.868.1
90%75%82.867.7
90%90%82.668.5
90%95%82.868.6
", + "image_path": "cd8793fdca259e545ca2662d09b583b0085b4ddc958153d157c003c3af0ed7bf.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 150, + 83, + 247, + 153 + ], + "blocks": [ + { + "bbox": [ + 63, + 156, + 124, + 164 + ], + "lines": [ + { + "bbox": [ + 63, + 156, + 124, + 164 + ], + "spans": [ + { + "bbox": [ + 63, + 156, + 124, + 164 + ], + "type": "text", + "content": "(a) Masking ratios." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 150, + 83, + 247, + 153 + ], + "lines": [ + { + "bbox": [ + 150, + 83, + 247, + 153 + ], + "spans": [ + { + "bbox": [ + 150, + 83, + 247, + 153 + ], + "type": "table", + "html": "
TypeAccuracy
IN1KSSv2IN1KSSv2
RandomTube82.867.9
CausalTube82.167.3
RandomFrame82.965.8
RandomCausal82.864.4
RandomRandom82.767.8
", + "image_path": "77d575a72ed00b2c52ebfe651baf85218543cd36f44e86d52c70f84977b3ba01.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 261, + 73, + 358, + 152 + ], + "blocks": [ + { + "bbox": [ + 170, + 156, + 228, + 164 + ], + "lines": [ + { + "bbox": [ + 170, + 156, + 228, + 164 + ], + "spans": [ + { + "bbox": [ + 170, + 156, + 228, + 164 + ], + "type": "text", + "content": "(b) Masking type." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 261, + 73, + 358, + 152 + ], + "lines": [ + { + "bbox": [ + 261, + 73, + 358, + 152 + ], + "spans": [ + { + "bbox": [ + 261, + 73, + 358, + 152 + ], + "type": "table", + "html": "
#paramsIN1KSSv2
Common7.9M82.867.9
Separate33.0M82.667.7
L=23.8M82.767.9
L=47.9M82.867.9
L=814.5M82.868.0
d=3847.9M82.867.9
d=51213.0M82.768.1
", + "image_path": "544ddd2ec2d533d15491d90ac49e8006f0bb86824f1af01b9b449a64e2505b8b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 276, + 156, + 343, + 164 + ], + "lines": [ + { + "bbox": [ + 276, + 156, + 343, + 164 + ], + "spans": [ + { + "bbox": [ + 276, + 156, + 343, + 164 + ], + "type": "text", + "content": "(c) Decoder capacity." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 374, + 72, + 539, + 152 + ], + "blocks": [ + { + "bbox": [ + 374, + 72, + 539, + 152 + ], + "lines": [ + { + "bbox": [ + 374, + 72, + 539, + 152 + ], + "spans": [ + { + "bbox": [ + 374, + 72, + 539, + 152 + ], + "type": "image", + "image_path": "1f7c489cc73d09c22ac440fb4ee50ef4e5f412ac731febf69a23e4d158a72d3b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 431, + 156, + 484, + 163 + ], + "lines": [ + { + "bbox": [ + 431, + 156, + 484, + 163 + ], + "spans": [ + { + "bbox": [ + 431, + 156, + 484, + 163 + ], + "type": "text", + "content": "(d) Dataset ratio" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 174, + 547, + 274 + ], + "lines": [ + { + "bbox": [ + 46, + 174, + 547, + 274 + ], + "spans": [ + { + "bbox": [ + 46, + 174, + 547, + 274 + ], + "type": "text", + "content": "Table 1. Ablations. The default setting for all ablations is highlighted in gray. (a) We vary the masking ratios of the input for pretraining the model. This allows us to use an extremely high masking ratio of " + }, + { + "bbox": [ + 46, + 174, + 547, + 274 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 46, + 174, + 547, + 274 + ], + "type": "text", + "content": " on images and " + }, + { + "bbox": [ + 46, + 174, + 547, + 274 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 46, + 174, + 547, + 274 + ], + "type": "text", + "content": " on videos. (b) OmniMAE works well across different masking types. Random masking makes no assumptions about the input patches. Tube masking masks random patches at the same spatial location across all frames. Causal masking for images masks the 'future' patches as determined by a raster left-to-right order. In videos, causal masking masks all the patches from future frames. Frame masking masks all patches for randomly selected frames. (c) Different decoder designs (common or separate) and capacities (number of layers " + }, + { + "bbox": [ + 46, + 174, + 547, + 274 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 46, + 174, + 547, + 274 + ], + "type": "text", + "content": " and dimension " + }, + { + "bbox": [ + 46, + 174, + 547, + 274 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 174, + 547, + 274 + ], + "type": "text", + "content": " of the MLP). A common decoder for both images and videos performs better than a separate decoder. Our final performance is robust to decoder capacity. (d) We vary the image/video dataset ratios by replicating the entire datasets by a factor. The number of training updates changes with such replication, and we observe that the model benefits from higher replication of videos." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 294, + 287, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 294, + 287, + 342 + ], + "spans": [ + { + "bbox": [ + 46, + 294, + 287, + 342 + ], + "type": "text", + "content": "masking, we obtain " + }, + { + "bbox": [ + 46, + 294, + 287, + 342 + ], + "type": "inline_equation", + "content": "1.3 \\times" + }, + { + "bbox": [ + 46, + 294, + 287, + 342 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 294, + 287, + 342 + ], + "type": "inline_equation", + "content": "1.5 \\times" + }, + { + "bbox": [ + 46, + 294, + 287, + 342 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 294, + 287, + 342 + ], + "type": "inline_equation", + "content": "1.4 \\times" + }, + { + "bbox": [ + 46, + 294, + 287, + 342 + ], + "type": "text", + "content": " fewer FLOPS for ViT-B/L/H. Given the compute savings and strong performance, we choose " + }, + { + "bbox": [ + 46, + 294, + 287, + 342 + ], + "type": "inline_equation", + "content": "90 \\%" + }, + { + "bbox": [ + 46, + 294, + 287, + 342 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 294, + 287, + 342 + ], + "type": "inline_equation", + "content": "95 \\%" + }, + { + "bbox": [ + 46, + 294, + 287, + 342 + ], + "type": "text", + "content": " masking for images and videos respectively for our final models." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 342, + 288, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 342, + 288, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 342, + 288, + 605 + ], + "type": "text", + "content": "Type of masking. We study the effect of the type of masking used in training our models. The different types of masking are illustrated in Figure 3. We experiment with Random masking which masks patches in the image or video randomly. For videos, we experiment with two types of masking that exploit the temporal structure. In Tube masking, we mask random patches at the same spatial location across all the frames. In Causal masking, we use a raster order masking, akin to generative image models [16, 67], i.e. the patches in the top-left of the image, and earlier frames of the video, are kept while the rest are masked. Finally, in Frame masking, we randomly mask some frames in the video, while keeping all patches from the unmasked frames. As seen in Table 1b, Random and Tube masking perform comparably well on both image and video tasks. We find that in case of Causal masking, the prediction task becomes exceedingly tough due to the uncertainty of the future, and in case of Frame masking, it becomes relatively easy due to the high redundancy of pixel values across frames. Hence in both these cases, the representation learned does not perform as well on video recognition tasks. Given the simplicity of random masking, we use that for both modalities." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 713 + ], + "type": "text", + "content": "Decoder architecture. Since image and video prediction tasks may require specialized parameters, we study two settings: (1) the decoder parameters are shared for image and video pixel prediction; (2) two separate decoders are used for image and video prediction. For the latter we use our default decoder setting for videos (4 layers/384-D), however a larger 8-layer/512-D decoder for images as proposed in MAE [40]. The results in Table 1c show that using a shared decoder for both image and video prediction leads to better" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 294, + 545, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 294, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 304, + 294, + 545, + 330 + ], + "type": "text", + "content": "transfer learning accuracy. A shared decoder also offers a " + }, + { + "bbox": [ + 304, + 294, + 545, + 330 + ], + "type": "inline_equation", + "content": "4 \\times" + }, + { + "bbox": [ + 304, + 294, + 545, + 330 + ], + "type": "text", + "content": " reduction in the number of parameters compared to using separate decoders while also being simpler to implement." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 331, + 546, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 331, + 546, + 425 + ], + "spans": [ + { + "bbox": [ + 304, + 331, + 546, + 425 + ], + "type": "text", + "content": "Decoder capacity. In Table 1c, we vary the decoder capacity and measure its impact on the final transfer learning accuracy. To change the decoder capacity, we vary the number of Transformer layers " + }, + { + "bbox": [ + 304, + 331, + 546, + 425 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 304, + 331, + 546, + 425 + ], + "type": "text", + "content": " used and the dimension " + }, + { + "bbox": [ + 304, + 331, + 546, + 425 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 304, + 331, + 546, + 425 + ], + "type": "text", + "content": " of the MLP used in the Transformer layers. Overall, the final transfer learning accuracy is robust to decoder capacity. A shallow decoder of 4 layers (8 for ViT-H) offers a good trade-off between the decoder size and final accuracies." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 426, + 547, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 547, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 547, + 676 + ], + "type": "text", + "content": "Dataset ratios. Since ImageNet and SSv2 have a different number of samples, we study the effect of varying the ratio of the datasets used in training. When increasing the ratio of the datasets, we measure a single epoch as training over the new oversampled set of samples from the datasets. We vary the relative ratio for ImageNet and Something Something-v2 by replicating only one dataset at a time as shown in Table 1d. We observe that increasing the relative dataset ratio has a positive effect on the final transfer accuracy for the oversampled dataset. This is expected as oversampling a dataset proportionally increases its corresponding parameter updates, thus making the representation better tuned for the dataset. We also observe that oversampling Something-Something-v2 twice as much as ImageNet leads to an improvement for the video transfer accuracy with no drop in the image transfer performance. Hence, OmniMAE is robust to changes in dataset sizes of individual modalities, and it suggests that longer training on both datasets can further improve the transfer learning performance of our models. For simplicity, by default we do not replicate any dataset for training our models." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "content": "Sample replication. We replicate samples for both images and videos, and in each case measure a single epoch as training over the total samples, counting replicated samples, i.e." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10411" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 44, + 70, + 541, + 517 + ], + "blocks": [ + { + "bbox": [ + 44, + 70, + 541, + 517 + ], + "lines": [ + { + "bbox": [ + 44, + 70, + 541, + 517 + ], + "spans": [ + { + "bbox": [ + 44, + 70, + 541, + 517 + ], + "type": "image", + "image_path": "e0236fdf699dda3c27e0ae084442361871d705c5422b55da07c16974fc4df7b5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 529, + 547, + 563 + ], + "lines": [ + { + "bbox": [ + 46, + 529, + 547, + 563 + ], + "spans": [ + { + "bbox": [ + 46, + 529, + 547, + 563 + ], + "type": "text", + "content": "Figure 4. Reconstruction visualizations using OmniMAE on different video and image datasets. We show the model predictions for varying masking ratios of the input from " + }, + { + "bbox": [ + 46, + 529, + 547, + 563 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 46, + 529, + 547, + 563 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 529, + 547, + 563 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 46, + 529, + 547, + 563 + ], + "type": "text", + "content": " and the ground truth reference (Ref). OmniMAE is trained on ImageNet and SSv2 but the predictions generalize to other datasets like K400 and EK100. Please see the supplement for video visualizations." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 584, + 289, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 584, + 289, + 704 + ], + "spans": [ + { + "bbox": [ + 46, + 584, + 289, + 704 + ], + "type": "text", + "content": "replication maintains the same number of training iterations. We train models with different replication factors and show the final transfer accuracy and the normalized training time in Figs. 5a and 5b. Sample replication leads to faster training while maintaining or even improving the final transfer accuracy. Since a large portion of the input sample is masked, replicating the sample multiple number of times still provides enough learning signal for the model and does not lead to a degradation in performance. This becomes even more relevant for OmniMAE's final settings where we use" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 304, + 584, + 547, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 584, + 547, + 657 + ], + "spans": [ + { + "bbox": [ + 304, + 584, + 547, + 657 + ], + "type": "text", + "content": "higher masking ratios. For video data, we use an optimized datalocator (see details in Appendix B) with asynchronous I/O and fast video decoding. Even with this optimized datalocator, sample replication leads to " + }, + { + "bbox": [ + 304, + 584, + 547, + 657 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 304, + 584, + 547, + 657 + ], + "type": "text", + "content": " faster training. We believe this is an important practical observation as video datalugging is often a bottleneck for training models." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 306, + 669, + 456, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 669, + 456, + 682 + ], + "spans": [ + { + "bbox": [ + 306, + 669, + 456, + 682 + ], + "type": "text", + "content": "4.3. Comparison to Prior Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 689, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 547, + 715 + ], + "type": "text", + "content": "In Table 2, we compare OmniMAE's representation on our image and video classification benchmarks with other" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "10412" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 133, + 69, + 459, + 280 + ], + "blocks": [ + { + "bbox": [ + 133, + 69, + 459, + 280 + ], + "lines": [ + { + "bbox": [ + 133, + 69, + 459, + 280 + ], + "spans": [ + { + "bbox": [ + 133, + 69, + 459, + 280 + ], + "type": "table", + "html": "
MethodArch.Pretrain DataIN1KiNat18P365K400SSv2EK100
DINOViT-BIN1K82.872.6-XXX
iBOT [92]ViT-LIN1K84.8--XXX
MAE [40]ViT-BIN1K83.675.457.9XXX
MAE [40]ViT-LIN1K85.980.159.4XXX
MAE [40]ViT-HIN1K86.983.059.8XXX
BEiT [6]ViT-BIN1K83.472.3-XXX
BEiT [6]ViT-LIN1K85.2--XXX
VIMPAC [74]ViT-L/2†HowTo100MXXX77.468.1-
MaskedFeat [82]MViT-v2-LK400XXX84.3--
BEVT [80]Swin-BIN1K + K400XXX81.171.4-
OmniMAEViT-BIN1K + K40082.974.258.580.869.040.3
OmniMAEViT-BIN1K + SSv283.074.058.480.669.539.3
OmniMAEViT-LIN1K + SSv285.279.659.384.074.245.1
OmniMAEViT-HIN1K + SSv286.683.260.185.475.548.0
Concurrent methods specialized for videos
VideoMAE-16f [76]ViT-LSSv2XXX-74.2-
VideoMAE-16f [76]ViT-LK400XXX84.7--
MAE-Video [28]ViT-HK400XXX85.174.1-
", + "image_path": "109784566343738d7688974f7ae5a4d678a8ddc5f23c0528e95fd9fb6a4a1fd0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 292, + 547, + 338 + ], + "lines": [ + { + "bbox": [ + 46, + 292, + 547, + 338 + ], + "spans": [ + { + "bbox": [ + 46, + 292, + 547, + 338 + ], + "type": "text", + "content": "Table 2. Comparing OmniMAE with prior self-supervised methods on image and video recognition. Prior work trains a specialized model for a particular visual modality, sometimes using specialized architectures. Our single OmniMAE model is pretrained jointly on images and videos for 1600 epochs (ViT-H for 2400) and performs competitively across all benchmarks while using a simple architecture and pretraining method, even outperforming concurrent work specialized for videos. ViT-L with half the MLP embedding dimension" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 357, + 288, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 357, + 288, + 453 + ], + "spans": [ + { + "bbox": [ + 46, + 357, + 288, + 453 + ], + "type": "text", + "content": "state-of-the-art self-supervised methods. We focus on methods which use a Transformer backbone, ViT, or an architecture based on Transformers like Swin. To the best of our knowledge, prior work does not explore joint pre-training except for BEVT [80], which does masked image modeling followed by joint masked image and video modeling. Unlike OmniMAE, BEVT only focuses on video representations and does not evaluate on image recognition." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 453, + 288, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 453, + 288, + 646 + ], + "spans": [ + { + "bbox": [ + 46, + 453, + 288, + 646 + ], + "type": "text", + "content": "OmniMAE is pretrained for 1600 epochs (ViT-H for 2400 epochs). Training with IN1K and K400 or IN1K and SSv2 leads to similar results, so we use the latter when scaling model size. OmniMAE performs competitively for both image and video recognition when compared to the best models trained separately for either modality. OmniMAE also performs favorably compared to methods like MaskedFeat which are pretrained only for video with specialized architectures on larger video datasets. OmniMAE serves as a competitive initialization for transferring models on all modalities. More notably, OmniMAE's performance improves significantly with larger architectures. Using ViT-H, OmniMAE performs better or within the margin of error across all benchmarks. Notably, OmniMAE even outperforms concurrent works which use similar approaches while solely focusing on video representation learning." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 657, + 212, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 657, + 212, + 670 + ], + "spans": [ + { + "bbox": [ + 47, + 657, + 212, + 670 + ], + "type": "text", + "content": "5. Conclusion and Future Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 288, + 715 + ], + "type": "text", + "content": "We present OmniMAE, a unified Transformer for images and videos that can be pretrained using masked autoencoding. OmniMAE uses a simple architecture with minimal" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 357, + 547, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 357, + 547, + 574 + ], + "spans": [ + { + "bbox": [ + 304, + 357, + 547, + 574 + ], + "type": "text", + "content": "vision-specific changes and is competitive with specialized architectures and models tailored for images and videos. We believe such generic models and approaches are a critical and under explored area of representation learning. OmniMAE has not been empirically validated on other visual modalities like 3D, or non-visual modalities like audio. The simplicity and generality of OmniMAE can likely enable future multimodal systems that use shared parameters to model multiple modalities. We only studied masked autoencoder based pretraining strategies for OmniMAE since they are easier to train. However, pixel reconstruction methods do not learn good linearly separable features or similarity metrics [40]. These properties are essential when finetuning with limited labeled data. We believe that exploring other pretraining strategies, particularly in the context of joint modeling, will likely lead to improved models. Similarly, joint finetuning strategies [33] could be used to evaluate such unified models and lead to potentially even superior performance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 715 + ], + "type": "text", + "content": "Ethical considerations. We study models for visual recognition on images and videos and our technical contributions are neutral from an ethics standpoint. We do not propose a new dataset or make claims about the suitability of our model on production data. Given their self-supervised nature, our learned representations are free from label bias, however they are still susceptible to the bias in the distribution of the visual data. We believe all ethical considerations that apply to visual recognition models equally apply to our model." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "10413" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 49, + 72, + 105, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 72, + 105, + 83 + ], + "spans": [ + { + "bbox": [ + 49, + 72, + 105, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 288, + 114 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 288, + 114 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 288, + 114 + ], + "type": "text", + "content": "[1] Relja Arandjelovic and Andrew Zisserman. Look, listen and learn. In ICCV, 2017." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 118, + 288, + 154 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 118, + 288, + 154 + ], + "spans": [ + { + "bbox": [ + 53, + 118, + 288, + 154 + ], + "type": "text", + "content": "[2] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lucic, and Cordelia Schmid. ViViT: A video vision transformer. In ICCV, 2021." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 158, + 288, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 158, + 288, + 194 + ], + "spans": [ + { + "bbox": [ + 54, + 158, + 288, + 194 + ], + "type": "text", + "content": "[3] Yuki Markus Asano, Christian Rupprecht, and Andrea Vedaldi. Self-labelling via simultaneous clustering and representation learning. In ICLR, 2020." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 198, + 288, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 198, + 288, + 233 + ], + "spans": [ + { + "bbox": [ + 54, + 198, + 288, + 233 + ], + "type": "text", + "content": "[4] Sara Atito, Muhammad Awais, and Josef Kittler. Sit: Self-supervised vision transformer. arXiv preprint arXiv:2104.03602, 2021." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 236, + 288, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 236, + 288, + 284 + ], + "spans": [ + { + "bbox": [ + 54, + 236, + 288, + 284 + ], + "type": "text", + "content": "[5] Alexei Baevski, Wei-Ning Hsu, Qiantong Xu, Arun Babu, Jiatao Gu, and Michael Auli. data2vec: A general framework for self-supervised learning in speech, vision and language. In ICML, 2022." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 288, + 288, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 288, + 288, + 312 + ], + "spans": [ + { + "bbox": [ + 54, + 288, + 288, + 312 + ], + "type": "text", + "content": "[6] Hangbo Bao, Li Dong, and Furu Wei. BEiT: Bert pre-training of image transformers. In ICLR, 2022." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 316, + 288, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 316, + 288, + 351 + ], + "spans": [ + { + "bbox": [ + 54, + 316, + 288, + 351 + ], + "type": "text", + "content": "[7] Adrien Bardes, Jean Ponce, and Yann LeCun. Vicreg: Variance-invariance-covariance regularization for self-supervised learning. In ICLR, 2022." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 355, + 288, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 355, + 288, + 402 + ], + "spans": [ + { + "bbox": [ + 54, + 355, + 288, + 402 + ], + "type": "text", + "content": "[8] Maxim Berman, Hervé Jégou, Andrea Vedaldi, Iasonas Kokkinos, and Matthijs Douze. Multigrain: a unified image embedding for classes and instances. arXiv preprint arXiv:1902.05509, 2019." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 406, + 288, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 406, + 288, + 441 + ], + "spans": [ + { + "bbox": [ + 54, + 406, + 288, + 441 + ], + "type": "text", + "content": "[9] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In ICML, 2021." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 445, + 288, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 445, + 288, + 481 + ], + "spans": [ + { + "bbox": [ + 48, + 445, + 288, + 481 + ], + "type": "text", + "content": "[10] Hervé Bourlard and Yves Kamp. Auto-association by multilayer perceptrons and singular value decomposition. Biological Cybernetics, 1988." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 484, + 288, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 484, + 288, + 520 + ], + "spans": [ + { + "bbox": [ + 48, + 484, + 288, + 520 + ], + "type": "text", + "content": "[11] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features. In ECCV, 2018." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 524, + 288, + 571 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 524, + 288, + 571 + ], + "spans": [ + { + "bbox": [ + 48, + 524, + 288, + 571 + ], + "type": "text", + "content": "[12] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. In NeurIPS, 2020." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 574, + 288, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 574, + 288, + 622 + ], + "spans": [ + { + "bbox": [ + 48, + 574, + 288, + 622 + ], + "type": "text", + "content": "[13] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 626, + 288, + 662 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 626, + 288, + 662 + ], + "spans": [ + { + "bbox": [ + 48, + 626, + 288, + 662 + ], + "type": "text", + "content": "[14] João Carreira and Andrew Zisserman. Quo vadis, action recognition? A new model and the kinetics dataset. In CVPR, 2017." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 665, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 665, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 665, + 288, + 712 + ], + "type": "text", + "content": "[15] Lluis Castrejon, Yusuf Aytar, Carl Vondrick, Hamed Piriavash, and Antonio Torralba. Learning aligned cross-modal representations from weakly aligned data. In CVPR, 2016." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 307, + 72, + 547, + 108 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 547, + 108 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 547, + 108 + ], + "type": "text", + "content": "[16] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In ICML, 2020." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 111, + 545, + 146 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 111, + 545, + 146 + ], + "spans": [ + { + "bbox": [ + 307, + 111, + 545, + 146 + ], + "type": "text", + "content": "[17] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 149, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 149, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 149, + 545, + 173 + ], + "type": "text", + "content": "[18] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In CVPR, 2021." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 175, + 547, + 210 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 547, + 210 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 547, + 210 + ], + "type": "text", + "content": "[19] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In ICCV, 2021." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 214, + 547, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 214, + 547, + 259 + ], + "spans": [ + { + "bbox": [ + 307, + 214, + 547, + 259 + ], + "type": "text", + "content": "[20] Kevin Clark, Minh-Thang Luong, Quoc V Le, and Christopher D Manning. Electra: Pre-training text encoders as discriminators rather than generators. *ICLR*, 2020." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 264, + 547, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 264, + 547, + 310 + ], + "spans": [ + { + "bbox": [ + 307, + 264, + 547, + 310 + ], + "type": "text", + "content": "[21] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In CVPR, 2020." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 314, + 547, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 314, + 547, + 373 + ], + "spans": [ + { + "bbox": [ + 307, + 314, + 547, + 373 + ], + "type": "text", + "content": "[22] Dima Damen, Hazel Doughty, Giovanni Maria Farinella, Antonino Furnari, Evangelos Kazakos, Jian Ma, Davide Moltisanti, Jonathan Munro, Toby Perrett, Will Price, et al. Rescaling egocentric vision. IJCV, 2021." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 376, + 547, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 376, + 547, + 423 + ], + "spans": [ + { + "bbox": [ + 307, + 376, + 547, + 423 + ], + "type": "text", + "content": "[23] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding. In NAACL, 2018." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 426, + 547, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 426, + 547, + 497 + ], + "spans": [ + { + "bbox": [ + 307, + 426, + 547, + 497 + ], + "type": "text", + "content": "[24] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 501, + 547, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 501, + 547, + 548 + ], + "spans": [ + { + "bbox": [ + 307, + 501, + 547, + 548 + ], + "type": "text", + "content": "[25] Alaaeldin El-Nouby, Gautier Izacard, Hugo Touvron, Ivan Laptev, Hervé Jegou, and Edouard Grave. Are large-scale datasets necessary for self-supervised pretraining? arXiv preprint arXiv:2112.10740, 2021." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 551, + 547, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 551, + 547, + 586 + ], + "spans": [ + { + "bbox": [ + 307, + 551, + 547, + 586 + ], + "type": "text", + "content": "[26] Aleksandr Ermolov, Aliaksandr Siarohin, Enver Sangineto, and Nicu Sebe. Whitening for self-supervised representation learning. In ICML, 2021." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 590, + 547, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 590, + 547, + 635 + ], + "spans": [ + { + "bbox": [ + 307, + 590, + 547, + 635 + ], + "type": "text", + "content": "[27] Haoqi Fan, Bo Xiong, Karttikeya Mangalam, Yanghao Li, Zhicheng Yan, Jitendra Malik, and Christoph Feichtenhofer. Multiscale vision transformers. In ICCV, 2021." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 639, + 547, + 675 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 639, + 547, + 675 + ], + "spans": [ + { + "bbox": [ + 307, + 639, + 547, + 675 + ], + "type": "text", + "content": "[28] Christoph Feichtenhofer, Haoqi Fan, Yanghao Li, and Kaiming He. Masked autoencoders as spatiotemporal learners. arXiv preprint arXiv:2205.09113, 2022." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 677, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 677, + 547, + 713 + ], + "type": "text", + "content": "[29] Christoph Feichtenhofer, Axel Pinz, and Andrew Zisserman. Convolutional two-stream network fusion for video action recognition. In CVPR, 2016." + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "10414" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 120 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 120 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 120 + ], + "type": "text", + "content": "[30] Patrick Gallinari, Yann LeCun, Sylvie Thiria, and F Fogelman Soulie. Mémoires associatives distribuées: une comparaison (distributed associative memories: a comparison). In COGNITIVA. 1987." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 122, + 288, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 122, + 288, + 158 + ], + "spans": [ + { + "bbox": [ + 48, + 122, + 288, + 158 + ], + "type": "text", + "content": "[31] Rohit Girdhar, João Carreira, Carl Doersch, and Andrew Zisserman. Video action transformer network. In CVPR, 2019." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 161, + 288, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 161, + 288, + 185 + ], + "spans": [ + { + "bbox": [ + 48, + 161, + 288, + 185 + ], + "type": "text", + "content": "[32] Rohit Girdhar and Kristen Grauman. Anticipative Video Transformer. In ICCV, 2021." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 187, + 288, + 235 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 187, + 288, + 235 + ], + "spans": [ + { + "bbox": [ + 48, + 187, + 288, + 235 + ], + "type": "text", + "content": "[33] Rohit Girdhar, Mannat Singh, Nikhila Ravi, Laurens van der Maaten, Armand Joulin, and Ishan Misra. Omnivore: A Single Model for Many Visual Modalities. In CVPR, 2022." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 238, + 288, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 238, + 288, + 285 + ], + "spans": [ + { + "bbox": [ + 48, + 238, + 288, + 285 + ], + "type": "text", + "content": "[34] Yunchao Gong, Liwei Wang, Micah Hodosh, Julia Hockenmaier, and Svetlana Lazebnik. Improving image-sentence embeddings using large weakly annotated photo collections. In ECCV, 2014." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 288, + 288, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 288, + 288, + 360 + ], + "spans": [ + { + "bbox": [ + 48, + 288, + 288, + 360 + ], + "type": "text", + "content": "[35] Priya Goyal, Quentin Duval, Isaac Seessel, Mathilde Caron, Mannat Singh, Ishan Misra, Levent Sagun, Armand Joulin, and Piotr Bojanowski. Vision models are more robust and fair when pretrained on uncurated images without supervision. arXiv preprint arXiv:2202.08360, 2022." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 363, + 288, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 363, + 288, + 399 + ], + "spans": [ + { + "bbox": [ + 48, + 363, + 288, + 399 + ], + "type": "text", + "content": "[36] Priya Goyal, Dhruv Mahajan, Abhinav Gupta, and Ishan Misra. Scaling and benchmarking self-supervised visual representation learning. In ICCV, 2019." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 401, + 288, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 401, + 288, + 485 + ], + "spans": [ + { + "bbox": [ + 48, + 401, + 288, + 485 + ], + "type": "text", + "content": "[37] Raghav Goyal, Samira Ebrahimi Kahou, Vincent Michalski, Joanna Materzynska, Susanne Westphal, Heuna Kim, Valentin Haenel, Ingo Fruend, Peter Yianilos, Moritz Mueller-Freitag, Florian Hoppe, Christian Thurau, Ingo Bax, and Roland Memisevic. The “something something” video database for learning and evaluating visual common sense. In ICCV, 2017." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 487, + 288, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 487, + 288, + 559 + ], + "spans": [ + { + "bbox": [ + 48, + 487, + 288, + 559 + ], + "type": "text", + "content": "[38] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. NeurIPS, 2020." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 562, + 288, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 562, + 288, + 597 + ], + "spans": [ + { + "bbox": [ + 48, + 562, + 288, + 597 + ], + "type": "text", + "content": "[39] Raia Hadsell, Sumit Chopra, and Yann LeCun. Dimensionality reduction by learning an invariant mapping. In CVPR, 2006." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 600, + 288, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 600, + 288, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 600, + 288, + 635 + ], + "type": "text", + "content": "[40] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dólár, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, 2022." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 639, + 288, + 675 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 639, + 288, + 675 + ], + "spans": [ + { + "bbox": [ + 48, + 639, + 288, + 675 + ], + "type": "text", + "content": "[41] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, 2020." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 677, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 677, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 677, + 288, + 712 + ], + "type": "text", + "content": "[42] Geoffrey E Hinton and Richard Zemel. Autoencoders, minimum description length and helmholtz free energy. In NeurIPS, 1993." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 307, + 72, + 547, + 119 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 547, + 119 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 547, + 119 + ], + "type": "text", + "content": "[43] Elad Hoffer, Tal Ben-Nun, Itay Hubara, Niv Giladi, Torsten Hoefler, and Daniel Soudry. Augment your batch: Improving generalization through instance repetition. In CVPR, 2020." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 122, + 546, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 122, + 546, + 170 + ], + "spans": [ + { + "bbox": [ + 307, + 122, + 546, + 170 + ], + "type": "text", + "content": "[44] Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alex Shepard, Hartwig Adam, Pietro Perona, and Serge Belongie. The inaturalist species classification and detection dataset. In CVPR, 2018." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 173, + 546, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 173, + 546, + 208 + ], + "spans": [ + { + "bbox": [ + 307, + 173, + 546, + 208 + ], + "type": "text", + "content": "[45] Gao Huang, Yu Sun, Zhuang Liu, Daniel Sedra, and Kilian Q Weinberger. Deep networks with stochastic depth. In ECCV, 2016." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 211, + 547, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 211, + 547, + 246 + ], + "spans": [ + { + "bbox": [ + 307, + 211, + 547, + 246 + ], + "type": "text", + "content": "[46] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In ICML, 2015." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 249, + 546, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 249, + 546, + 284 + ], + "spans": [ + { + "bbox": [ + 307, + 249, + 546, + 284 + ], + "type": "text", + "content": "[47] Andrej Karpathy and Li Fei-Fei. Deep visual-semantic alignments for generating image descriptions. In CVPR, 2015." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 287, + 547, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 287, + 547, + 358 + ], + "spans": [ + { + "bbox": [ + 307, + 287, + 547, + 358 + ], + "type": "text", + "content": "[48] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, AMustafa Suleyman, and Andrew Zisserman. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 361, + 547, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 361, + 547, + 397 + ], + "spans": [ + { + "bbox": [ + 307, + 361, + 547, + 397 + ], + "type": "text", + "content": "[49] Mark A Kramer. Nonlinear principal component analysis using autoassociative neural networks. AIChE, 1991." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 399, + 547, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 399, + 547, + 424 + ], + "spans": [ + { + "bbox": [ + 307, + 399, + 547, + 424 + ], + "type": "text", + "content": "[50] Yann LeCun and Françoise Fogelman-Soulie. Modèles connexionnistes de l'apprentissage. Intellectica, 1987." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 426, + 546, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 426, + 546, + 461 + ], + "spans": [ + { + "bbox": [ + 307, + 426, + 546, + 461 + ], + "type": "text", + "content": "[51] Junnan Li, Pan Zhou, Caiming Xiong, and Steven C.H. Hoi. Prototypical contrastive learning of unsupervised representations. In ICLR, 2021." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 464, + 546, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 464, + 546, + 512 + ], + "spans": [ + { + "bbox": [ + 307, + 464, + 546, + 512 + ], + "type": "text", + "content": "[52] Yanghao Li, Chao-Yuan Wu, Haoqi Fan, Karttikeya Mangalam, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. Improved multiscale vision transformers for classification and detection. In CVPR, 2022." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 514, + 547, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 514, + 547, + 574 + ], + "spans": [ + { + "bbox": [ + 307, + 514, + 547, + 574 + ], + "type": "text", + "content": "[53] Valerii Likhosherstov, Anurag Arnab, Krzysztof Choromanski, Mario Lucic, Yi Tay, Adrian Weller, and Mostafa Dehghani. Polyvit: Co-training vision transformers on images, videos and audio. arXiv preprint arXiv:2111.12993, 2021." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 577, + 546, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 577, + 546, + 624 + ], + "spans": [ + { + "bbox": [ + 307, + 577, + 546, + 624 + ], + "type": "text", + "content": "[54] Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, Furu Wei, and Baining Guo. Swin transformer v2: Scaling up capacity and resolution. In CVPR, 2022." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 627, + 546, + 675 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 627, + 546, + 675 + ], + "spans": [ + { + "bbox": [ + 307, + 627, + 546, + 675 + ], + "type": "text", + "content": "[55] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 677, + 546, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 677, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 677, + 546, + 713 + ], + "type": "text", + "content": "[56] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swin transformer. arXiv preprint arXiv:2106.13230, 2021." + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10415" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 107 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 107 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 107 + ], + "type": "text", + "content": "[57] Ze Liu, Zheng Zhang, Yue Cao, Han Hu, and Xin Tong. Group-free 3d object detection via transformers. In ICCV, 2021." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 111, + 287, + 147 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 111, + 287, + 147 + ], + "spans": [ + { + "bbox": [ + 49, + 111, + 287, + 147 + ], + "type": "text", + "content": "[58] Jiasen Lu, Vedanuj Goswami, Marcus Rohrbach, Devi Parikh, and Stefan Lee. 12-in-1: Multi-task vision and language representation learning. In CVPR, 2020." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 149, + 288, + 197 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 149, + 288, + 197 + ], + "spans": [ + { + "bbox": [ + 48, + 149, + 288, + 197 + ], + "type": "text", + "content": "[59] Antoine Miech, Jean-Baptiste Alayrac, Lucas Smaira, Ivan Laptev, Josef Sivic, and Andrew Zisserman. End-to-end learning of visual representations from uncu-rated instructional videos. In CVPR, 2020." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 201, + 288, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 201, + 288, + 236 + ], + "spans": [ + { + "bbox": [ + 48, + 201, + 288, + 236 + ], + "type": "text", + "content": "[60] Ishan Misra, Rohit Girdhar, and Armand Joulin. An End-to-End Transformer Model for 3D Object Detection. In ICCV, 2021." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 239, + 288, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 239, + 288, + 274 + ], + "spans": [ + { + "bbox": [ + 48, + 239, + 288, + 274 + ], + "type": "text", + "content": "[61] Ishan Misra and Laurens van der Maaten. Self-supervised learning of pretext-invariant representations. In CVPR, 2020." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 278, + 288, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 278, + 288, + 312 + ], + "spans": [ + { + "bbox": [ + 48, + 278, + 288, + 312 + ], + "type": "text", + "content": "[62] Pedro Morgado, Ishan Misra, and Nuno Vasconcelos. Robust audio-visual instance discrimination. In CVPR, 2021." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 317, + 288, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 317, + 288, + 352 + ], + "spans": [ + { + "bbox": [ + 48, + 317, + 288, + 352 + ], + "type": "text", + "content": "[63] Pedro Morgado, Nuno Vasconcelos, and Ishan Misra. Audio-visual instance discrimination with cross-modal agreement. In CVPR, 2021." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 355, + 288, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 355, + 288, + 391 + ], + "spans": [ + { + "bbox": [ + 48, + 355, + 288, + 391 + ], + "type": "text", + "content": "[64] B. A. Olshausen and D. J. Field. Emergence of simple-cell receptive field properties by learning a sparse code for natural images. Nature, 1996." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 395, + 288, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 395, + 288, + 430 + ], + "spans": [ + { + "bbox": [ + 48, + 395, + 288, + 430 + ], + "type": "text", + "content": "[65] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. In NeurIPS, 2018." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 434, + 288, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 434, + 288, + 468 + ], + "spans": [ + { + "bbox": [ + 48, + 434, + 288, + 468 + ], + "type": "text", + "content": "[66] Andrew Owens and Alexei A Efros. Audio-visual scene analysis with self-supervised multisensory features. In ECCV, 2018." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 472, + 288, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 472, + 288, + 507 + ], + "spans": [ + { + "bbox": [ + 48, + 472, + 288, + 507 + ], + "type": "text", + "content": "[67] Niki Parmar, Ashish Vaswani, Jakob Uszkoreit, Lukasz Kaiser, Noam Shazeer, Alexander Ku, and Dustin Tran. Image transformer. In ICML, 2018." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 510, + 288, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 510, + 288, + 546 + ], + "spans": [ + { + "bbox": [ + 48, + 510, + 288, + 546 + ], + "type": "text", + "content": "[68] Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A Efros. Context encoders: Feature learning by inpainting. In CVPR, 2016." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 549, + 288, + 585 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 549, + 288, + 585 + ], + "spans": [ + { + "bbox": [ + 48, + 549, + 288, + 585 + ], + "type": "text", + "content": "[69] Boris T Polyak and Anatoli B Juditsky. Acceleration of stochastic approximation by averaging. SIAM journal on control and optimization, 1992." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 588, + 288, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 588, + 288, + 647 + ], + "spans": [ + { + "bbox": [ + 48, + 588, + 288, + 647 + ], + "type": "text", + "content": "[70] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. IJCV, 2015." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 651, + 288, + 673 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 651, + 288, + 673 + ], + "spans": [ + { + "bbox": [ + 48, + 651, + 288, + 673 + ], + "type": "text", + "content": "[71] Ruslan Salakhutdinov and Geoffrey Hinton. Deep Boltzmann machines. In AISTATS, 2009." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 677, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 677, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 677, + 288, + 712 + ], + "type": "text", + "content": "[72] Karen Simonyan and Andrew Zisserman. Two-stream convolutional networks for action recognition in videos. In NeurIPS, 2014." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 307, + 72, + 547, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 547, + 118 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 547, + 118 + ], + "type": "text", + "content": "[73] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In CVPR, 2016." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 123, + 547, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 123, + 547, + 171 + ], + "spans": [ + { + "bbox": [ + 307, + 123, + 547, + 171 + ], + "type": "text", + "content": "[74] Hao Tan, Jie Lei, Thomas Wolf, and Mohit Bansal. VIMPAC: Video pre-training via masked token prediction and contrastive learning. arXiv preprint arXiv:2106.11250, 2021." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 175, + 547, + 210 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 547, + 210 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 547, + 210 + ], + "type": "text", + "content": "[75] Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive multiview coding. arXiv preprint arXiv:1906.05849, 2019." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 215, + 547, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 215, + 547, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 215, + 547, + 262 + ], + "type": "text", + "content": "[76] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. arXiv preprint arXiv:2203.12602, 2022." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 266, + 547, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 266, + 547, + 313 + ], + "spans": [ + { + "bbox": [ + 307, + 266, + 547, + 313 + ], + "type": "text", + "content": "[77] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In ICML, 2021." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 318, + 547, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 318, + 547, + 365 + ], + "spans": [ + { + "bbox": [ + 307, + 318, + 547, + 365 + ], + "type": "text", + "content": "[78] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 369, + 547, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 369, + 547, + 404 + ], + "spans": [ + { + "bbox": [ + 307, + 369, + 547, + 404 + ], + "type": "text", + "content": "[79] P. Vincent, H. Larochelle, Y. Bengio, and P.-A. Manzagol. Extracting and composing robust features with denoising autoencoders. In ICML, 2008." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 408, + 547, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 408, + 547, + 456 + ], + "spans": [ + { + "bbox": [ + 307, + 408, + 547, + 456 + ], + "type": "text", + "content": "[80] Rui Wang, Dongdong Chen, Zuxuan Wu, Yinpeng Chen, Xiyang Dai, Mengchen Liu, Yu-Gang Jiang, Luowei Zhou, and Lu Yuan. BEVT: Bert pretraining of video transformers. In CVPR, 2022." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 460, + 547, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 460, + 547, + 518 + ], + "spans": [ + { + "bbox": [ + 307, + 460, + 547, + 518 + ], + "type": "text", + "content": "[81] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In ICCV, 2021." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 523, + 547, + 570 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 523, + 547, + 570 + ], + "spans": [ + { + "bbox": [ + 307, + 523, + 547, + 570 + ], + "type": "text", + "content": "[82] Chen Wei, Haoqi Fan, Saining Xie, Chao-Yuan Wu, Alan Yuille, and Christoph Feichtenhofer. Masked feature prediction for self-supervised visual pre-training. In CVPR, 2022." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 574, + 547, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 574, + 547, + 609 + ], + "spans": [ + { + "bbox": [ + 307, + 574, + 547, + 609 + ], + "type": "text", + "content": "[83] Zhirong Wu, Yuanjun Xiong, Stella X Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In CVPR, 2018." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 614, + 547, + 661 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 614, + 547, + 661 + ], + "spans": [ + { + "bbox": [ + 307, + 614, + 547, + 661 + ], + "type": "text", + "content": "[84] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jian-min Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In CVPR, 2022." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 666, + 547, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 666, + 547, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 666, + 547, + 712 + ], + "type": "text", + "content": "[85] Xueting Yan, Ishan Misra, Abhinav Gupta, Deepti Ghadiyaram, and Dhruv Mahajan. ClusterFit: Improving Generalization of Visual Representations. In CVPR, 2020." + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10416" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 353 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 120 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 120 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 120 + ], + "type": "text", + "content": "[86] Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with localizable features. In ICCV, 2019." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 123, + 288, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 123, + 288, + 159 + ], + "spans": [ + { + "bbox": [ + 48, + 123, + 288, + 159 + ], + "type": "text", + "content": "[87] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stephane Deny. Barlow twins: Self-supervised learning via redundancy reduction. In ICML, 2021." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 162, + 288, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 162, + 288, + 198 + ], + "spans": [ + { + "bbox": [ + 48, + 162, + 288, + 198 + ], + "type": "text", + "content": "[88] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In ICLR, 2018." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 201, + 288, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 201, + 288, + 224 + ], + "spans": [ + { + "bbox": [ + 48, + 201, + 288, + 224 + ], + "type": "text", + "content": "[89] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip Torr, and Vladlen Koltun. Point transformer. In ICCV, 2021." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 228, + 288, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 228, + 288, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 228, + 288, + 262 + ], + "type": "text", + "content": "[90] Zhun Zhong, Liang Zheng, Guoliang Kang, Shaozi Li, and Yi Yang. Random erasing data augmentation. In AAAI, 2020." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 266, + 288, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 266, + 288, + 313 + ], + "spans": [ + { + "bbox": [ + 48, + 266, + 288, + 313 + ], + "type": "text", + "content": "[91] Bolei Zhou, Agata Lapedriza, Jianxiong Xiao, Antonio Torralba, and Aude Oliva. Learning deep features for scene recognition using places database. In NeurIPS, 2014." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 317, + 288, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 317, + 288, + 353 + ], + "spans": [ + { + "bbox": [ + 48, + 317, + 288, + 353 + ], + "type": "text", + "content": "[92] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. ibot: Image bert pre-training with online tokenizer. In ICLR, 2022." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10417" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/OmniObject3D_ Large-Vocabulary 3D Object Dataset for Realistic Perception, Reconstruction and Generation/243023a1-e64c-4bdc-a9c5-c30584f803ba_content_list.json b/2023/OmniObject3D_ Large-Vocabulary 3D Object Dataset for Realistic Perception, Reconstruction and Generation/243023a1-e64c-4bdc-a9c5-c30584f803ba_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..3d048010a3ab2d6583814b975bcc55217e3b1771 --- /dev/null +++ b/2023/OmniObject3D_ Large-Vocabulary 3D Object Dataset for Realistic Perception, Reconstruction and Generation/243023a1-e64c-4bdc-a9c5-c30584f803ba_content_list.json @@ -0,0 +1,1451 @@ +[ + { + "type": "text", + "text": "OmniObject3D: Large-Vocabulary 3D Object Dataset for Realistic Perception, Reconstruction and Generation", + "text_level": 1, + "bbox": [ + 194, + 130, + 776, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tong Wu $^{1,2}$ , Jiarui Zhang $^{1,3}$ , Xiao Fu $^{1}$ , Yuxin Wang $^{1,4}$ , Jiawei Ren $^{5}$ , Liang Pan $^{5}$ , Wayne Wu $^{1}$ , Lei Yang $^{1,3}$ , Jiaqi Wang $^{1}$ , Chen Qian $^{1}$ , Dahua Lin $^{1,2\\boxtimes}$ , Ziwei Liu $^{5\\boxtimes}$", + "bbox": [ + 194, + 202, + 831, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Shanghai Artificial Intelligence Laboratory, $^{2}$ The Chinese University of Hong Kong, $^{3}$ SenseTime Research, $^{4}$ Hong Kong University of Science and Technology, $^{5}$ S-Lab, Nanyang Technological University", + "bbox": [ + 88, + 239, + 939, + 275 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d056abb644ec104bb811689b4c3f1b8fab84ffe54f420c3ff8e80a763e512e92.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 101, + 286, + 669, + 522 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/91a4275a8b6f75345ad6b9a173b3caa5c79231e9ed119771bfea692e71cfc2dc.jpg", + "image_caption": [ + "Textured meshes" + ], + "image_footnote": [], + "bbox": [ + 684, + 292, + 836, + 354 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d78849eede2e862c95f04572aa4c1495954206ce24c266472b216583f5820702.jpg", + "image_caption": [ + "Point clouds Rendered images" + ], + "image_footnote": [], + "bbox": [ + 684, + 368, + 861, + 425 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9777ee296f0a8c41cf97ec290c7efdd6b42c764d2eed5c1ee91223b4f5bce0de.jpg", + "image_caption": [ + "Real-captured videos" + ], + "image_footnote": [], + "bbox": [ + 674, + 439, + 864, + 508 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/346d9e3b169997728a6217be4005801a4bb2a6c29de013bcc2e65406c124c1a4.jpg", + "image_caption": [ + "Figure 1. OmniObject3D is a large vocabulary 3D object dataset with massive high-quality real-scanned 3D objects and rich annotations. It supports various research topics, e.g., perception, novel view synthesis, neural surface reconstruction, and 3D generation." + ], + "image_footnote": [], + "bbox": [ + 99, + 523, + 867, + 612 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 655, + 313, + 671 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advances in modeling 3D objects mostly rely on synthetic datasets due to the lack of large-scale real-scanned 3D databases. To facilitate the development of 3D perception, reconstruction, and generation in the real world, we propose OmniObject3D, a large vocabulary 3D object dataset with massive high-quality real-scanned 3D objects. OmniObject3D has several appealing properties: 1) Large Vocabulary: It comprises 6,000 scanned objects in 190 daily categories, sharing common classes with popular 2D datasets (e.g., ImageNet and LVIS), benefiting the pursuit of generalizable 3D representations. 2) Rich Annotations: Each 3D object is captured with both 2D and", + "bbox": [ + 75, + 688, + 473, + 869 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D sensors, providing textured meshes, point clouds, multiview rendered images, and multiple real-captured videos. 3) Realistic Scans: The professional scanners support high-quality object scans with precise shapes and realistic appearances. With the vast exploration space offered by OmniObject3D, we carefully set up four evaluation tracks: a) robust 3D perception, b) novel-view synthesis, c) neural surface reconstruction, and d) 3D object generation. Extensive studies are performed on these four benchmarks, revealing new observations, challenges, and opportunities for future research in realistic 3D vision.", + "bbox": [ + 496, + 656, + 893, + 821 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 500, + 840, + 632, + 857 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sensing, understanding, and synthesizing realistic 3D objects is a long-standing problem in computer vision, with", + "bbox": [ + 500, + 869, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Corresponding authors. https://omniobject3d.github.io/", + "bbox": [ + 76, + 886, + 460, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "803", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "rapid progress emerging in recent years. However, a majority of the technical approaches rely on unrealistic synthetic datasets [6, 19, 64] due to the absence of a large-scale real-world 3D object database. However, the appearance and distribution gaps between synthetic and real data cannot be compensated for trivially, hindering their real-life applications. Therefore, it is imperative to equip the community with a large-scale and high-quality 3D object dataset from the real world, which can facilitate a variety of 3D vision tasks and downstream applications.", + "bbox": [ + 75, + 90, + 472, + 242 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent advances partially fulfill the requirements while still being unsatisfactory. As shown in Table 1, CO3D [48] contains 19k videos capturing objects from 50 MS-COCO categories, while only $20\\%$ of the videos are annotated with accurate point clouds reconstructed by COLMAP [50]. Moreover, they do not provide textured meshes. GSO [16] has 1k scanned objects while covering only 17 household classes. AKB-48 [33] focuses on robotics manipulation with 2k articulated object scans in 48 categories, but the focus on articulation leads to a relatively narrow semantic distribution, failing to support general 3D object research.", + "bbox": [ + 75, + 244, + 472, + 412 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To boost the research on general 3D object understanding and modeling, we present OmniObject3D: a large-vocabulary 3D object dataset with massive high-quality, real-scanned 3D objects. Our dataset has several appealing properties: 1) Large Vocabulary: It contains 6,000 high-quality textured meshes scanned from real-world objects, which, to the best of our knowledge, is the largest among real-world 3D object datasets with accurate 3D meshes. It comprises 190 daily categories, sharing common classes with popular 2D and 3D datasets (e.g., ImageNet [15], LVIS [25], and ShapeNet [6]), incorporating most daily object realms (See Figure 1 and Figure 2). 2) Rich Annotations: Each 3D object is captured with both 2D and 3D sensors, providing textured 3D meshes, sampled point clouds, posed multi-view images rendered by Blender [13], and real-captured video frames with foreground masks and COLMAP camera poses. 3) Realistic Scans: The object scans are of high fidelity thanks to the professional scanners, bearing precise shapes with geometric details and realistic appearance with high-frequency textures.", + "bbox": [ + 75, + 412, + 472, + 717 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Taking advantage of the vast exploration space offered by OmniObject3D, we carefully set up four evaluation tracks: a) robust 3D perception, b) novel-view synthesis, c) neural surface reconstruction, and d) 3D object generation. Extensive studies are performed on these benchmarks: First, the high-quality, real-world point clouds in OmniObject3D allow us to perform robust 3D perception analysis on both out-of-distribution (OOD) styles and corruptions, two major challenges in point cloud OOD generalization. Furthermore, we provide massive 3D models with multi-view images and precise 3D meshes for novel-view synthesis and neural surface reconstruction. The broad diver", + "bbox": [ + 75, + 719, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/e04f38f29771a66063f013ddf7670a63b67532ec91942b1fb3589a4335b5bda5.jpg", + "table_caption": [ + "Table 1. A comparison between OmniObject3D and other commonly-used 3D object datasets. $R^{LVis}$ denotes the ratio of the 1.2k LVIS [25] categories being covered." + ], + "table_footnote": [], + "table_body": "
DatasetRealFull MeshVideo# Objs# CatsR1vis (%)
ShapeNet [6]✓51k554.1
ModelNet [64]✓12k402.4
3D-Future [19]✓16k341.3
ABO [12]✓8k633.5
Toys4K [53]✓4k1057.7
CO3D V1 / V2 [48]✓✓19 / 40k504.2
DTU [1]✓✓124NA0
ScanObjectNN [56]✓15k151.3
GSO [16]✓✓1k170.9
AKB-48 [33]✓✓2k481.8
Ours✓✓✓6k19010.8
", + "bbox": [ + 504, + 130, + 901, + 290 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "sity in shapes and textures offers a comprehensive training and evaluation source for both scene-specific and generalizable algorithms. Finally, we equip the community with a database for large vocabulary and realistic 3D object generation, which pushes the boundary of existing state-of-the-art generation methods to real-world 3D objects. The four benchmarks reveal new observations, challenges, and opportunities for future research in realistic 3D vision.", + "bbox": [ + 496, + 297, + 893, + 420 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 500, + 436, + 650, + 452 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D Object Datasets. The acquisition of a large-scale realistic 3D database is usually expensive and challenging. Many widely-used 3D datasets prefer to collect synthetic CAD models from online repositories [6, 53, 64], for example, ShapeNet [6] has 51,300 3D CAD models in 55 categories, and ModelNet40 [64] consists of 12,311 models in 40 categories. Recent works, e.g., 3D-FUTURE [19] and ABO [12], introduce high-quality CAD models with rich geometric details and informative textures. However, due to the inevitable gap between synthetic and real objects, the community is still eager for a large-scale 3D object dataset in the real world. DTU [1] and BlendedMVS [69] are photo-realistic datasets designed for multi-view stereo benchmarks, while they are small in scale and lack category annotations. ScanObjectNN [56] is a real-world point cloud object dataset based on scanned indoor scenes, containing around 15,000 objects with colored point clouds in 15 categories. However, the point clouds are incomplete and noisy, and multiple objects usually co-exist in one scene. GSO [16] has 1,030 scanned objects with fine geometries and textures in 17 household items, and AKB-48 [33] focuses on robotics manipulation with 2,037 articulated object models in 48 articulated object categories. However, the relatively narrow semantic scope of GSO and AKB-48 hinders their applications for more general 3D research. CO3D v1 [48] contains 19,000 object-centric videos, while only $20\\%$ of them are annotated with accurate point clouds reconstructed by COLMAP [50], and they do not provide meshes or textures. In contrast, the proposed OmniOb", + "bbox": [ + 496, + 462, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "804", + "bbox": [ + 485, + 944, + 509, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a6195fa50c5f41afcc508a8931ff1a9b4ddf8bff89866401733e9bbf6c81687d.jpg", + "image_caption": [ + "Figure 2. Semantic distribution of the OmniObject3D dataset. It covers 190 daily categories with a long-tailed distribution, sharing common classes with popular 2D and 3D datasets." + ], + "image_footnote": [], + "bbox": [ + 98, + 88, + 648, + 257 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e19fc1e9e1bdc9b6e89a005fa17ee1b5f55474ecc4e6b2a1c1505ed18e97623d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 660, + 98, + 864, + 255 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "project3D dataset comprises 6,000 3D objects scanned by professional devices with meshes, textures, and multi-view photos in 190 categories, fulfilling the requirements of a wide range of research objectives. A detailed comparison is presented in Table 1.", + "bbox": [ + 75, + 295, + 470, + 369 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Robust 3D Perception. Robustness to out-of-distribution (OOD) data is important in point cloud perception. Two main challenges include OOD styles (e.g., differences between CAD models and real-world objects) and OOD corruptions (e.g., random point jittering or missing due to sensory inaccuracy). A branch of works [10, 30, 45, 60] studies the OOD corruptions and proposes standard corruption test suites [49, 55], while they fail to take account of OOD styles. Another branch of works [3, 48] evaluates the sim-to-real domain gap by training models on clean synthetic datasets [64] and testing them on noisy real-world test sets [56], while OOD styles and corruptions cannot be disentangled under this setting for an independent analysis. In this work, we leverage high-quality, real-world point clouds from OmniObject3D to systematically measure the robustness against the OOD style and OOD corruptions, providing the first benchmark for fine-grained evaluation of the point cloud perception robustness.", + "bbox": [ + 75, + 371, + 470, + 643 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Neural Radiance Field and Neural Surface Reconstruction. Neural radiance field (NeRF) [40] represents a scene with a fully-connected deep network (MLPs), which takes in hundreds of sampled points along each camera ray and outputs the predicted color and density. We can synthesize the image of an unseen view from a trained model via volume rendering. Inspired by the success of NeRF, massive follow-up efforts have been made to improve its quality [4, 5, 39, 57] and efficiency [7, 18, 42, 54]. A branch of works [8, 27, 35, 48, 59, 71] has also explored the generalization ability of NeRF-based frameworks, where they aim to learn priors from deep image features across multiple scenes. Beyond novel view synthesis, another trend of approaches [14, 43, 58, 63, 70] proposes to combine neural radiance field with implicit surface representations like Signed Distance Function (SDF), and they achieve accurate and mask-free surface reconstruction from multi-view im", + "bbox": [ + 75, + 643, + 470, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ages. Since dense camera views of scenes are sometimes unavailable, recent advances explore surface reconstruction from sparse views. They address the problem by exploiting generalizable priors cross scenes for a generic surface prediction [36] or taking advantage of the estimated geometry cues estimated by pre-trained networks [72]. OmniObject3D can serve as a large-scale benchmark with realistic photos and accurate meshes for both training and evaluation. The high diversity in shape and appearance offers an opportunity for pursuing more generalizable and robust novel view synthesis and surface reconstruction methods.", + "bbox": [ + 496, + 295, + 890, + 460 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D Object Generation. Early approaches [20,26,37,52,62] extend 2D generation frameworks to 3D voxels with a high computational cost. Some others adopt different 3D data formulations, e.g., point cloud [2,41,68,74], octree [29], and implicit representations [11,38]. However, it is nontrivial to generate complex and textured surfaces. Recent advances [9,21,44] explore the generation for textured 3D meshes, where GET3D [22] is a state-of-the-art approach that generates diverse meshes with rich geometry and textures in two branches. It is a promising but challenging task to train generative models on a large vocabulary and realistic dataset. We evaluate GET3D on our dataset and reveal several challenges and future opportunities.", + "bbox": [ + 496, + 462, + 890, + 657 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In supplementary materials, we present more detailed discussions on related works for different tracks.", + "bbox": [ + 498, + 657, + 890, + 688 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. The OmniObject3D Dataset", + "text_level": 1, + "bbox": [ + 500, + 703, + 759, + 722 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we describe the data collection, processing, and annotation pipeline of OmniObject3D. We also introduce the statistics and distribution of it.", + "bbox": [ + 498, + 729, + 890, + 773 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Data Collection, Processing, and Annotation", + "text_level": 1, + "bbox": [ + 500, + 786, + 874, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Category List Definition. In order to collect a large amount of 3D objects that are both commonly-distributed and highly diverse, we first pre-define a category list according to several popular 2D and 3D datasets [6,15,25,31, 32,51,64]. We cover most of the categories that lie within the application scope of the scanners and also dynamically", + "bbox": [ + 496, + 809, + 890, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "805", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "expand the list with reasonable new classes that are absent from the current list during collection. We end up with 190 widely-spread categories, which ensures a library with rich texture, geometry, and semantic information.", + "bbox": [ + 75, + 90, + 468, + 150 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Object Collection Pipeline. We then collect a variety of objects from each category and use professional 3D scanners to obtain high-resolution textured meshes. Specifically, we use the Shining 3D scanner1 and Artec Eva 3D scanner2 for objects in different scales. The scanning time varies with the properties of the object: it takes around 15 minutes to scan a small rigid object with a simple geometry (e.g., an apple, a toy), while it takes up to an hour to obtain a qualified 3D scan for non-rigid, complex, or large objects (e.g., a bed, a kite). For around $10\\%$ of the objects, we conduct common manipulations (e.g., taking a bite, cutting in pieces) to conform the natural instincts of them. The 3D scans can faithfully retain the real-world scale of each object, but their poses are not strictly aligned. We thus pre-define a canonical pose for each category and manually align the objects within a category. We then check the quality of each scan, and around $83\\%$ high-quality ones out of the total collection are finally reserved in the dataset.", + "bbox": [ + 75, + 152, + 472, + 422 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Image Rendering and Point Cloud Sampling. To support a variety of research topics like point cloud analysis, neural radiance fields, and 3D generation, we render multiview images and sample point clouds based on the collected 3D models. We use Blender [13] to render object-centric and photo-realistic multi-view images, together with accurate camera poses. The images are rendered from 100 random viewpoints sampled on the upper hemisphere at $800 \\times 800$ pixels. We also produce high-resolution mid-level cues like depth and normal for more research use. We then uniformly sample multi-resolution point clouds from each 3D model using the Open3D toolbox [75], with $2^{n}$ ( $n \\in \\{10, 11, 12, 13, 14\\}$ ) points in each point cloud, respectively. Besides the data existing in the dataset, we also provide a data generation pipeline. One can easily obtain new data with self-defined camera distributions, lighting, and point sampling methods to meet different requirements. Video Capturing and Annotation. After scanning each object, we capture its video with an iPhone 12 Pro mobile phone. The object is placed on or beside a calibration board, and each video covers a full $360^{\\circ}$ range around it. Square corners on the calibration board can be recognized by the QR Codes beside it, and we then filter out the blurry frames whose recognized corners are less than 8. We uniformly sample 200 frames, and then COLMAP [50], a well-known SfM pipeline, is applied to annotate the frames with camera poses. Finally, we use the scales of the calibration board in both the SfM coordinate space and the real world to recover the absolute scale of the SfM coordinate system.", + "bbox": [ + 75, + 424, + 472, + 864 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d6b968714584f3496f44ecb711309889bc0e7447485a1fa46be3725784a42863.jpg", + "image_caption": [ + "Figure 3. OmniObject3D provides the first clean real-world point cloud object dataset and allows fine-grained analysis on robustness to OOD styles and OOD corruptions.\"-C\": corrupted by common corruptions described in [49]" + ], + "image_footnote": [], + "bbox": [ + 527, + 95, + 870, + 183 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We also develop a two-stage matting pipeline based on the U $^2$ Net [47] and FBA [17] matting model to generate the foreground masks for all the frames. Please refer to supplementary materials for more implementation details.", + "bbox": [ + 498, + 244, + 893, + 305 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Statistics and Distribution", + "text_level": 1, + "bbox": [ + 500, + 315, + 736, + 330 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With 6,000 3D models in 190 categories, OmniObject3D exhibits a long-tailed distribution with an average of around 30 objects in each category, as shown in Figure 2. It shares many common categories with several famous 2D and 3D datasets [6, 15, 25, 31, 32, 51, 64]. For example, it covers 85 categories in ImageNet [15] and 130 categories in LVIS [25], which leads to the highest $R^{\\mathrm{vis}}$ in Table 1. Most of the categories are covered by the Open Images [31] image-level labels. It bears a huge diversity in shapes and appearances. The vast semantic and geometrical exploration spaces enable a wide range of research objectives.", + "bbox": [ + 496, + 338, + 893, + 505 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 500, + 518, + 633, + 535 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Robust 3D Perception", + "text_level": 1, + "bbox": [ + 500, + 544, + 702, + 560 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Object-level point cloud classification is one of the most fundamental tasks in point cloud perception. In this section, we show how OmniObject3D boosts robustness analysis of point cloud classification by disentangling the two critical out-of-distribution (OOD) challenges introduced in Sec. 2, i.e., OOD styles and OOD corruptions.", + "bbox": [ + 496, + 566, + 893, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Existing robustness evaluation utilizes clean synthetic datasets, e.g., ModelNet [64], for training and sets up two kinds of test sets for evaluation:", + "bbox": [ + 496, + 657, + 893, + 702 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1) Noisy real-world datasets, e.g., ScanObjectNN [56], which are cropped from real-world scenes. They are employed to measure the robustness of the sim-to-real domain gap. However, the gap couples both OOD styles and OOD corruptions simultaneously as the cropped point clouds are always noisy, making it impossible to analyze the two robustness challenges independently.", + "2) Corrupted synthetic datasets, e.g., ModelNet-C [49], which are artificially perturbed on top of clean synthetic datasets. The evaluation allows for detailed corruption analysis, but they do not reflect the robustness to OOD styles." + ], + "bbox": [ + 498, + 703, + 890, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "None of the existing robustness benchmarks allows for analyzing the robustness to both OOD styles and OOD cor", + "bbox": [ + 500, + 869, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "1https://www.einscan.com/", + "bbox": [ + 76, + 875, + 222, + 886 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "$^{2}$ https://www.artec3d.cn/", + "bbox": [ + 78, + 887, + 212, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "806", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/6b6a35fe0a6959e8c819485d424d8e69b56b00a0cd5bf73be131071237ca3a9c.jpg", + "table_caption": [ + "Table 2. Point cloud perception robustness analysis on OmniObject3D with different architecture designs. Models are trained on the ModelNet-40 dataset, with OAClean to be their overall accuracy on the standard ModelNet-40 test set. OAStyle on OmniObject3D evaluates the robustness to OOD styles. mCE on the corrupted OmniObject3D-C evaluates the robustness to OOD corruptions. Blue shadings indicate rankings. $\\dagger$ : results on ModelNet-C [49]. Full results are presented in the supplementary materials." + ], + "table_footnote": [], + "table_body": "
mCE†↓OA Clean ↑OA Style ↑mCE ↓
DGCNN [60]1.0000.9260.4481.000
PointNet [45]1.4220.9070.4660.969
PointNet++ [46]1.0720.9300.4071.066
RSCNN [34]1.1300.9230.3931.076
SimpleView [23]1.0470.9390.4760.990
GDANet [67]0.8920.9340.4970.920
PAConv [66]1.1040.9360.4031.073
CurveNet [65]0.9270.9380.5000.929
PCT [24]0.9250.9300.4590.940
RPC [49]0.8630.9300.4720.936
", + "bbox": [ + 76, + 200, + 480, + 367 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ruptions in fine granularity. OmniObject3D, on the other hand, as the first clean real-world point cloud object dataset, can help to address the issue. For models trained on ModelNet, we first evaluate their performance on OmniObject3D to examine OOD-style robustness. Then, we create OmniObject3D-C by corrupting OmniObject3D with common corruptions described in [49] to examine the OOD-corruption robustness. We show a complete robustness evaluation scheme in Figure 3. For evaluation metrics, we use the overall accuracy (OA) on OmniObject3D to measure the OOD-style robustness and use DGCNN normalized mCE [49] to measure the OOD-corruption robustness.", + "bbox": [ + 75, + 369, + 468, + 549 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We benchmark ten state-of-the-art point cloud classification models in Table 2. We observe that 1) performance on a clean test set has little correlation with OOD-style robustness. For example, SimpleView [23] achieves the best $\\mathrm{OA}_{\\mathrm{Clean}}$ but mediocre $\\mathrm{OA}_{\\mathrm{Style}}$ ; 2) Advanced point grouping, e.g., curve-based point grouping in CurveNet [65] and frequency-based point grouping in GDANet [67], are robust not only to OOD corruptions as pointed out in [49], but also to OOD styles; 3) OOD style + OOD corruption is a more challenging setting. In particular, RPC, the most robust architecture to OOD corruptions [49], shows inferior mCE. In summary, robust point cloud perception models against both OOD styles and OOD corruptions are still under-explored. Our dataset sheds new light on a comprehensive understanding of point cloud perception robustness. See more results in the supplementary materials.", + "bbox": [ + 75, + 550, + 468, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Novel View Synthesis", + "text_level": 1, + "bbox": [ + 76, + 801, + 277, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we study several representative methods on OmniObject3D for novel view synthesis (NVS) in two settings: 1) training on a single scene with densely captured images, which is the standard setting for NeRF [40]; 2) learning priors across scenes from our dataset to explore", + "bbox": [ + 75, + 825, + 468, + 902 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/05a7c2fe5cf700cbdae736191e0826fccc6e55ee2a0710520af141b0264a898a.jpg", + "table_caption": [ + "Table 3. Single-scene novel view synthesis results. Three metrics and their standard deviation (SD) across the training set." + ], + "table_footnote": [], + "table_body": "
MethodPSNR (↑) / SDSSIM (↑) / SDLPIPS (↓) / SD
NeRF [40]34.01 / 3.460.953 / 0.0290.068 / 0.061
mip-NeRF [4]39.86 / 4.580.974 / 0.0130.084 / 0.048
Plenoxels [18]41.04 / 6.840.982 / 0.0310.030 / 0.031
", + "bbox": [ + 501, + 116, + 901, + 186 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/c73cf3e4cbfc5a3e3e4777159082b0b18381bf9e19abcdfa97b87f0315b4f03b.jpg", + "table_caption": [ + "Table 4. Cross-scene novel view synthesis results on 10 categories. 'Cat.' and 'All*' denote training on each category and training on all categories except the 10 test ones, respectively." + ], + "table_footnote": [], + "table_body": "
MethodTrainPSNR (↑)SSIM (↑)LPIPS (↓)\\( {\\mathcal{L}}_{1}^{\\text{depth }}\\left( \\downarrow \\right) \\)
MVSNeRF [8]All*17.490.5440.4420.193
Cat.17.540.5420.4480.230
All*-ft.25.700.7540.2510.081
Cat.-ft.25.520.7500.2640.076
IBRNet [59]All*19.390.5690.3990.423
Cat.19.030.5510.4150.290
All*-ft.26.890.7920.2150.081
Cat.-ft.25.670.7600.2380.099
pixelNeRF [71]All*22.160.6920.3420.109
Cat.20.650.6760.3480.195
", + "bbox": [ + 503, + 224, + 903, + 392 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the generalization ability of NeRF-style models.", + "bbox": [ + 498, + 400, + 818, + 415 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Single-Scene NVs. We select three objects in each category for the experiments, randomly sampling 1/8 images as the hold-out test set. We involve NeRF [40], mip-NeRF [4], and a voxel-based system named Plenoxels [18] for comparison. As in Table 3, we find that Plenoxels achieve the best performance on average for PSNR, SSIM [61], and LPIPS [73]. There exists a clear margin for LPIPS between Plenoxels and the other two methods since voxel-based methods are especially good at modeling high-frequency appearance. We also present the standard deviation (SD) of results across all the training samples, where Plenoxels are relatively unstable compared to NeRF and mip-NeRF. We observe that Plenoxels introduce artifacts when encountering concave geometry (e.g., bowls, chairs) and suffer from an inaccurate density field modeling when the foreground object is dark. MLP-based methods are relatively robust against these difficult cases. In a nutshell, our dataset provides a library with a variety of shapes and appearances, allowing a comprehensive evaluation of different NVS methods. See the supplementary for more results with the iPhone videos, detailed analysis, and visualizations.", + "bbox": [ + 496, + 416, + 890, + 733 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Cross-Scene NVS. We conduct extensive experiments on novel view synthesis from sparse inputs by pixelNeRF [71], IBRNet [59] and MVSNeRF [8] in Table 4. We select 10 categories with the most various scenes as the test set. All metrics are averaged over 300 images. In the generalization setting, although not trained on the test category, $\\mathrm{MVSNeRF_{All^{*}}}$ is comparable to $\\mathrm{MVSNeRF_{Cat}}$ , $\\mathrm{IBRNet_{All^{*}}}$ and $\\mathrm{pixelNeRF_{All^{*}}}$ even outperforms the corresponding \"Cat\"s in all terms of visual metrics, especially on regular-shaped objects such as squash and apple. It confirms that OmniObject3D serves as an information-rich", + "bbox": [ + 496, + 734, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "807", + "bbox": [ + 485, + 944, + 509, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f271b8f08e6032bbe824a41c71e1411e752a0232f71bc4477990c945a3e4e4c6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 102, + 89, + 872, + 318 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/d53556d028b683893afe85ec9474085c75e82283f769f4e5e019a896be81664b.jpg", + "table_caption": [ + "Table 5. Dense-view surface reconstruction results." + ], + "table_footnote": [], + "table_body": "
MethodChamfer Distance ×103(↓)
HardMediumEasyAvg
NeuS [58]9.265.633.466.09
VolSDF [70]10.064.942.865.92
Voxurf [63]9.014.982.585.49
Avg9.445.192.975.83
", + "bbox": [ + 114, + 354, + 433, + 467 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "dataset that is beneficial for obtaining a strong generalizable prior on unseen scenes. Moreover, it is noteworthy that MVSNeRF and pixelNeRF with 'All* generate better underlying depth than those with 'Cat.', inferring generalizable methods can implicitly learn geometric cues though only trained from appearance in our dataset. It is reasonable that (1) IBRNet suffers more severely than the others in geometry under the scarcity of source context (only 3 views) as it is more suitable for dense-view generalization that complies with its view interpolation module. (2) MVSNeRF lags behind pixelNeRF on visual performance as we take 10 test frames widely distributed around the object in $360^{\\circ}$ by FPS sampling algorithm, where cost volume will be inaccurate on large-range viewpoint change. After further finetuning IBRNet for only around 10 minutes on each test scene, $\\mathrm{IBRNet}_{\\mathrm{All}^{*}\\text{-ft}}$ achieves the best view synthesis results, comparable to test-time optimized NVS methods on nearby views. It is promising to utilize the large-scale and category-prosperous OmniObject3D, to build a benchmark suite for evaluating diverse cross-scene NVS methods.", + "bbox": [ + 75, + 474, + 472, + 776 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Neural Surface Reconstruction", + "text_level": 1, + "bbox": [ + 76, + 786, + 352, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Precise surface reconstruction from multi-view images enables a broad range of applications. For a single scene with dense-view images, algorithms are expected to conduct accurate, robust, and efficient surface reconstruction. When only sparse-view images are available, it is crucial to learn generalizable priors from a set of scenes or use other", + "bbox": [ + 75, + 810, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/dfb3dafcdbf239dada19f6ddc3709d7fd1165f821c9539386b447cb9ea72e6f2.jpg", + "image_caption": [ + "Figure 4. Neural surface reconstruction results for both dense-view and sparse-view settings.", + "Figure 5. Performance distribution of dense-view surface reconstruction. The averaged results of the three methods is imbalanced. The colored area denotes a smoothed range of results." + ], + "image_footnote": [], + "bbox": [ + 522, + 342, + 867, + 489 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "geometric cues to assist reconstruction. Accordingly, we study the two settings for surface reconstruction methods.", + "bbox": [ + 496, + 536, + 890, + 565 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dense-View Surface Reconstruction. Under this setting, we include three representative methods. NeuS [58] and VolSDF [70] are two well-known approaches that bridge neural volume rendering with implicit surface representation. We also involve a voxel-based method called Voxsurf [63], which leverages a hybrid representation to achieve acceleration and fine geometry reconstruction.", + "bbox": [ + 496, + 566, + 890, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Previous approaches in this task mainly perform evaluations and comparisons on 15 scenes from the DTU [1] dataset, which is not comprehensive and robust enough to demonstrate the ability of the methods in different scenarios. In comparison, we select three objects per category to run each of the three methods above, leading to over 1,500 reconstructions in total. We calculate the Chamfer Distance (CD) between the reconstructed surface and the ground truth. The distribution of the results is shown in Figure 5. The average curve is imbalanced: hard categories usually include low-textured, concave, or complex shapes (e.g., bowls, vases, kennels, cabinets, and durians). We thus split the categories into three levels of \"difficulty\" based on the average curve, and the level-wise results are presented in Table 5. We can observe a clear margin among results in", + "bbox": [ + 496, + 674, + 892, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "808", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/89461208797dfd51a4f256ad0173c3246dc1edd12623c11b9ba91c07b5161f99.jpg", + "table_caption": [ + "Table 6. Sparse-view (3-view) surface reconstruction results." + ], + "table_footnote": [], + "table_body": "
MethodTrainChamfer Distance ×103(↓)
HardMediumEasyAvg
NeuS [58]Single29.3527.6224.7927.3
MonoSDF [72]Single35.1435.3532.7634.68
SparseNeuS [36]1 cat.34.0531.3231.1432.36
10 cats.30.7530.1128.3729.87
All cats.26.1326.0822.1325.00
Easy28.3926.6523.7626.48
Medium27.3826.6623.0825.87
Hard27.4226.9524.6326.47
MVSNeRF [8]All cats.56.6848.0948.7051.16
pixelNeRF [71]All cats.63.3159.9161.4761.56
", + "bbox": [ + 86, + 104, + 459, + 303 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "different levels for each method, indicating the split subsets to be generic and faithful.", + "bbox": [ + 75, + 306, + 467, + 335 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Sparse-View Surface Reconstruction. Dense-captured images of a scene are sometimes not available, so we also study the sparse-view scenario here. The following methods are included: NeuS [58] with sparse-view input; MonoSDF [72], which takes in geometry cues estimated by pre-trained models; SparseNeuS [36], a generic surface prediction pipeline that learns generalizable priors; pixelNeRF [71] and MVSNeRF [8] from Sec. 4.2, whose geometries are extracted from the density field. For NeuS, MonoSDF, and pixelNeRF, we use Farthest Point Sampling (FPS) to sample views that are most widely distributed; for SparseNeuS and MVSNeRF, we conduct FPS among the nearest 30 camera poses from a random reference view. We sample 3 views in all the experiments.", + "bbox": [ + 75, + 338, + 467, + 549 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The quantitative and qualitative comparisons are shown in Table 6 and Figure 4, respectively. We observe apparent artifacts in all the sparse-view reconstructed results. Among them, SparseNeuS trained on enough data demonstrates the best quantitative performance on average, and the pre-division on the train set does not result in a noticeable difference across difficulty levels. NeuS with sparse-view input achieves a surprisingly good performance. As shown in Figure 4, the FPS sampling equips NeuS with a coherent global shape for thin structures like case 3, while it encounters severe local geometry ambiguity like case 1. MonoSDF, on the contrary, partially overcomes the issue of ambiguity via the assistance of predicted geometry cues in case 1. However, it relies heavily on the accuracy of the estimated depth and normal and thus easily fails when the estimation is inaccurate (e.g., case 2 and 3). The surfaces extracted from generalized NeRF models, i.e., pixelNeRF and MVSNeRF, are of relatively low quality.", + "bbox": [ + 75, + 551, + 467, + 823 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In brief, the challenging problem of sparse-view surface reconstruction has not been solved well currently. OmniObject3D is a promising database to study generalizable surface reconstruction pipelines as well as strategies for a robust usage of estimated geometry cues for this track.", + "bbox": [ + 75, + 824, + 467, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7a4e5ca54b4d4298c303b4927bb22ea6dc5f3d941993e1c6fa6e38e2bd816e25.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 92, + 679, + 238 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8689dca28d26bf5462052488e2fff21143b2b450c96ec02d09ccc45eed566d5a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 691, + 92, + 888, + 239 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a66375ac3ccd9ce8f2115cc07b7f7488a65f96759c7bea379ee2050c5c8b185c.jpg", + "image_caption": [ + "Figure 6. The category distribution of the generated shapes. (a) shows a weak positive correlation between the number of generated shapes and training shapes per category. (b) visualizes the correlation matrix among different categories by Chamfer Distance between their mean shapes. (c) visualizes categories being clustered into eight groups by KMeans. (d) presents a clear training-generation relation in the group-level statistics." + ], + "image_footnote": [], + "bbox": [ + 516, + 241, + 679, + 387 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6e0ed030abc642b71c9765a4883a92e0c58a8df4e43565e37b2d59ac90178d43.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 241, + 879, + 387 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4.3D Object Generation", + "text_level": 1, + "bbox": [ + 500, + 497, + 705, + 513 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we adopt a state-of-the-art generative model that directly generates explicit textured 3D meshes, namely GET3D [22]. GET3D is originally evaluated on six categories (Car, Chair, Motorbike, Animal, House, and Human Body) with independent models trained on each category. The number of shapes per category ranges from 337 to 7,497. In comparison, OmniObject3D contains many more categories with fewer objects in each. As a result, it is natural to train each model with multiple categories.", + "bbox": [ + 496, + 521, + 890, + 656 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We first provide some qualitative results in Figure 7, where we show various generation results: The textures are rather realistic, and the shapes are coherent, enhanced by fine geometry details (e.g., the lychee and pineapple). We explore the latent space of the model and show interpolation results in Figure 8. We can observe a smooth transition across instances that are semantically different. We would like to further analyze the performance of the generative model trained on OmniObject3D from three aspects, i.e., semantic distribution, diversity, and quality.", + "bbox": [ + 496, + 657, + 890, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Semantic Distribution. We randomly select 100 categories to train an unconditional model jointly. We randomly generate 1,000 textured meshes at inference time and ask human experts to label them. Shapes with ambiguity are not counted. Figure 6 (a) shows that the generated shapes per category are highly imbalanced, exhibiting a weak positive", + "bbox": [ + 496, + 810, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "809", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/053c21d8c495415c591922523dff2f9a8c2fc954523d0bc10294d3739734be65.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 90, + 890, + 207 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/43af7c99aaee2740bf7d2ce347fecb8fca0dadfd3a6891a83c2088f299f4d5af.jpg", + "image_caption": [ + "Figure 7. Examples of the generated textured shapes rendered in Blender. OmniObject3D enables GET3D with realistic generations across a wide range of categories.", + "Figure 8. Shape interpolation. We interpolate both geometry and texture latent codes from left to right." + ], + "image_footnote": [], + "bbox": [ + 80, + 241, + 465, + 325 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "correlation with the training shape numbers. Actually, the categories are not independent but rather highly correlated. We calculate the \"mean shape\" for each category and visualize the Chamfer Distance among them in Figure 6 (b), which indicates that they can be further grouped. Regarding each matrix row as a feature vector, we use KMeans to cluster them into eight groups (Figure 6 (c)) and carry out group-level statistics in Figure 6 (d). It demonstrates a clear trend that the number of generated shapes increases along with or even faster than the number of training shapes in the group, revealing an enlarged semantic bias during generation. However, it also depends on the inner-group divergence. For example, Group 2 (883 shapes in 27 categories) has the largest number of training samples, while the high variation among its categories prevents it from dominating the generated shapes; Group 1 (587 shapes in 18 categories) has a relatively small divergence, which becomes the most popular in the generated shapes. We present more details in the supplementary material.", + "bbox": [ + 75, + 369, + 468, + 656 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Diversity and Quality. We select four representative data subsets for training and evaluation, namely fruits, furniture, toys, and Rand-100. We randomly split each subset into training (80%) and testing (20%). We leverage three evaluation metrics: for geometry, we use Chamfer Distance (CD) to compute the Coverage score (Cov) and Minimum Matching Distance (MMD), which focus on the diversity and quality of the shapes, respectively; for texture, we adopt the widely-used FID [28]. But the FID metrics with different test splits are not directly comparable, suffering from a large variance when the test set is small. We thus introduce FIDref for reference, which is the FID between the train and test set. The results are shown in Table 7. Furniture suffers from the lowest quality (MMD) since the small train set with 17 categories is a difficult training source. Fruits has the same number of categories while being 2.3 times", + "bbox": [ + 75, + 659, + 470, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/2d5af172fc9d839a116403ccc6ba1b084ec24048817b670f5f02fdf650ee1544.jpg", + "table_caption": [ + "Table 7. Quantitative evaluations on different data splits." + ], + "table_footnote": [], + "table_body": "
Split#Objs#CatsCov (% ↑)MMD (↓)FID (↓)FIDref
Furniture2651767.924.2787.3958.40
Fruits6101746.723.32105.3187.15
Toys339755.222.78122.7741.40
Rand-100295110061.703.8946.578.65
", + "bbox": [ + 504, + 253, + 893, + 335 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "larger in scale, and some fruits share a very similar structure, leading to relatively higher quality and lower diversity (Cov). Toys achieve the best quality by training on only 7 categories. Rand-100 is the most difficult case, and we can observe a trade-off between quality and diversity. Both the FID and $\\mathrm{FID}^{\\mathrm{ref}}$ are high for the first three subsets due to the small testing sets, while only 'Rand-100' is relatively low.", + "bbox": [ + 496, + 342, + 890, + 446 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In brief, training and evaluating generative models on a large-vocabulary and realistic dataset is a promising but challenging task. We reveal crucial problems like the semantic distribution bias and varying exploration difficulties in different groups. OmniObject3D serves as a great database for further examination in this area.", + "bbox": [ + 496, + 449, + 892, + 539 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion and Outlook", + "text_level": 1, + "bbox": [ + 500, + 551, + 728, + 568 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduce OmniObject3D, a large vocabulary 3D object dataset with massive high-quality real-scanned 3D objects, including 6,000 objects from 190 categories. It provides rich annotations, including textured 3D meshes, sampled point clouds, posed multi-view images rendered by Blender, and real-captured video frames with foreground masks and COLMAP camera poses. We set up four evaluation tracks, revealing new observations, challenges, and opportunities for future research in realistic 3D vision.", + "bbox": [ + 496, + 577, + 890, + 712 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We will regulate the usage of our data to avoid potential negative social impacts.", + "bbox": [ + 496, + 713, + 890, + 743 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. This project is funded by Shanghai AI Laboratory, CUHK Interdisciplinary AI Research Institute, the Centre for Perceptual and Interactive Intelligence (CPII) Ltd under the Innovation and Technology Commission (ITC)'s InnoHK, Hong Kong RGC Theme-based Research Scheme 2020/21 (No. T41-603/20-R), OpenXDLab, the Ministry of Education, Singapore, under its MOE AcRF Tier 2 (MOE-T2EP20221-0012), NTU NAP, and under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative.", + "bbox": [ + 496, + 744, + 890, + 893 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "810", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Henrik Aanaes, Rasmus Ramsbøl Jensen, George Vogiatzis, Engin Tola, and Anders Bjorholm Dahl. Large-scale data for multiple-view stereopsis. International Journal of Computer Vision (IJCV), 120(2):153-168, 2016. 2, 6", + "[2] Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas Guibas. Learning representations and generative models for 3d point clouds. In Proceedings of the International Conference on Machine learning (ICML), pages 40-49, 2018. 3", + "[3] Adel Ahmadyan, Liangkai Zhang, Artsiom Ablavatski, Jianing Wei, and Matthias Grundmann. Objectron: A large scale dataset of object-centric videos in the wild with pose annotations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7822-7831, 2021. 3", + "[4] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5855–5864, 2021. 3, 5", + "[5] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5470-5479, 2022. 3", + "[6] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv.org, 1512.03012, 2015. 2, 3, 4", + "[7] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 3", + "[8] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14124-14133, 2021. 3, 5, 7", + "[9] Wenzheng Chen, Huan Ling, Jun Gao, Edward Smith, Jaakko Lehtinen, Alec Jacobson, and Sanja Fidler. Learning to predict 3d objects with an interpolation-based differentiable renderer. In Advances in Neural Information Processing Systems (NIPS), volume 32, 2019. 3", + "[10] Yunlu Chen, Vincent Tao Hu, Efstratios Gavves, Thomas Mensink, Pascal Mettes, Pengwan Yang, and Cees GM Snoek. Pointmixup: Augmentation for point clouds. In Proceedings of the European Conference on Computer Vision (ECCV), pages 330-345, 2020. 3", + "[11] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5939-5948, 2019. 3" + ], + "bbox": [ + 78, + 116, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Jasmine Collins, Shubham Goel, Kenan Deng, Achleshwar Luthra, Leon Xu, Erhan Gundogdu, Xi Zhang, Tomas F Yago Vicente, Thomas Dideriksien, Himanshu Arora, et al. Abo: Dataset and benchmarks for real-world 3d object understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21126-21136, 2022. 2", + "[13] Blender Online Community. Blender - a 3d modelling and rendering package. 2018. 2, 4", + "[14] François Darmon, Bénédicte Bascle, Jean-Clément Devaux, Pascal Monasse, and Mathieu Aubry. Improving neural implicit surfaces geometry with patch warping. arXiv.org, 2112.09648, 2021. 3", + "[15] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 248–255, 2009. 2, 3, 4", + "[16] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas B McHugh, and Vincent Vanhoucke. Google scanned objects: A high-quality dataset of 3d scanned household items. arXiv.org, 2204.11918, 2022. 2", + "[17] Marco Forte and François Pitié. F, b, alpha matting. arXiv.org, 2003.07711, 2020. 4", + "[18] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5501–5510, 2022. 3, 5", + "[19] Huan Fu, Rongfei Jia, Lin Gao, Mingming Gong, Binqiang Zhao, Steve Maybank, and Dacheng Tao. 3d-future: 3d furniture shape with texture. International Journal of Computer Vision (IJCV), 129(12):3313-3337, 2021. 2", + "[20] Matheus Gadelha, Subhransu Maji, and Rui Wang. 3d shape induction from 2d views of multiple objects. In Proceedings of the International Conference on 3D Vision (3DV), pages 402-411, 2017. 3", + "[21] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. arXiv.org, 2209.11163, 2022.3", + "[22] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. In Advances in Neural Information Processing Systems (NIPS), 2022. 3, 7", + "[23] Ankit Goyal, Hei Law, Bowei Liu, Alejandro Newell, and Jia Deng. Revisiting point cloud shape classification with a simple and effective baseline. In Proceedings of the International Conference on Machine learning (ICML), pages 3809-3820, 2021. 5", + "[24] Meng-Hao Guo, Jun-Xiong Cai, Zheng-Ning Liu, Tai-Jiang Mu, Ralph R Martin, and Shi-Min Hu. Pct: Point cloud transformer. Computational Visual Media, 7(2):187-199, 2021. 5" + ], + "bbox": [ + 501, + 92, + 893, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "811", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Agrim Gupta, Piotr Dollar, and Ross Girshick. LVIS: A dataset for large vocabulary instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5356-5364, 2019. 2, 3, 4", + "[26] Philipp Henzler, Niloy J Mitra, and Tobias Ritschel. Escaping plato's cave: 3d shape from adversarial rendering. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9984-9993, 2019. 3", + "[27] Philipp Henzler, Jeremy Reizenstein, Patrick Labatut, Roman Shapovalov, Tobias Ritschel, Andrea Vedaldi, and David Novotny. Unsupervised learning of 3d object categories from videos in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4700-4709, 2021. 3", + "[28] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In Advances in Neural Information Processing Systems (NIPS), volume 30, 2017. 8", + "[29] Moritz Ibing, Gregor Kobsik, and Leif Kobbelt. Octree transformer: Autoregressive 3d shape generation on hierarchically structured sequences. arXiv.org, 2111.12480, 2022. 3", + "[30] Sihyeon Kim, Sanghyeok Lee, Dasol Hwang, Jaewon Lee, Seong Jae Hwang, and Hyunwoo J Kim. Point cloud augmentation with weighted local transformations. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 548-557, 2021. 3", + "[31] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4. International Journal of Computer Vision (IJCV), 128(7):1956-1981, 2020. 3, 4", + "[32] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Proceedings of the European Conference on Computer Vision (ECCV), pages 740-755, 2014. 3, 4", + "[33] Liu Liu, Wenqiang Xu, Haoyuan Fu, Sucheng Qian, Qiao-jun Yu, Yang Han, and Cewu Lu. Akb-48: A real-world articulated object knowledge base. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14809-14818, 2022. 2", + "[34] Yongcheng Liu, Bin Fan, Shiming Xiang, and Chunhong Pan. Relation-shape convolutional neural network for point cloud analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8895-8904, 2019. 5", + "[35] Yuan Liu, Sida Peng, Lingjie Liu, Qianqian Wang, Peng Wang, Christian Theobalt, Xiaowei Zhou, and Wenping Wang. Neural rays for occlusion-aware image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7824-7833, 2022. 3", + "[36] Xiaoxiao Long, Cheng Lin, Peng Wang, Taku Komura, and Wenping Wang. Sparseneus: Fast generalizable neural surface reconstruction from sparse views. arXiv.org, 2206.05737, 2022. 3, 7" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] Sebastian Lunz, Yingzhen Li, Andrew Fitzgibbon, and Nate Kushman. Inverse graphics gan: Learning to generate 3d shapes from unstructured 2d data. arXiv.org, 2002.12674, 2020.3", + "[38] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4460-4470, 2019. 3", + "[39] Ben Mildenhall, Peter Hedman, Ricardo Martin-Brualla, Pratul P Srinivasan, and Jonathan T Barron. Nerf in the dark: High dynamic range view synthesis from noisy raw images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16190-16199, 2022. 3", + "[40] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In Proceedings of the European Conference on Computer Vision (ECCV), pages 405-421, 2020. 3, 5", + "[41] Kaichun Mo, Paul Guerrero, Li Yi, Hao Su, Peter Wonka, Niloy J. Mitra, and Leonidas J. Guibas. Structurenet: Hierarchical graph networks for 3d shape generation. arXiv.org, 1908.00575, 2019. 3", + "[42] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics, 2022. 3", + "[43] Michael Oechsle, Songyou Peng, and Andreas Geiger. Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5589-5599, 2021. 3", + "[44] Dario Pavllo, Jonas Kohler, Thomas Hofmann, and Aurelien Lucchi. Learning generative models of textured 3d meshes from real-world images. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 13879-13889, 2021. 3", + "[45] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: deep learning on point sets for 3d classification and segmentation. corr abs/1612.00593 (2016). In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 652-660, 2017. 3, 5", + "[46] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In Advances in Neural Information Processing Systems (NIPS), volume 30, 2017. 5", + "[47] Xuebin Qin, Zichen Zhang, Chenyang Huang, Masood Dehghan, Osmar R Zaiane, and Martin Jagersand. U2-net: Going deeper with nested u-structure for salient object detection. Pattern Recognition, 106:107404, 2020. 4", + "[48] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10901-10911, 2021. 2, 3" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "812", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[49] Jiawei Ren, Liang Pan, and Ziwei Liu. Benchmarking and analyzing point cloud classification under corruptions. In Proceedings of the International Conference on Machine learning (ICML), 2022. 3, 4, 5", + "[50] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4104-4113, 2016. 2, 4", + "[51] Shuai Shao, Zeming Li, Tianyuan Zhang, Chao Peng, Gang Yu, Xiangyu Zhang, Jing Li, and Jian Sun. Objects365: A large-scale, high-quality dataset for object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8430-8439, 2019. 3, 4", + "[52] Edward J Smith and David Meger. Improved adversarial systems for 3d object generation and reconstruction. In Proceedings of the Conference on Robot Learning (CoRL), pages 87-96, 2017. 3", + "[53] Stefan Stojanov, Anh Thai, and James M Rehg. Using shape to categorize: Low-shot learning with an explicit shape bias. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1798-1808, 2021. 2", + "[54] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5459-5469, 2022. 3", + "[55] Saeid Asgari Taghanaki, Jieliang Luo, Ran Zhang, Ye Wang, Pradeep Kumar Jayaraman, and Krishna Murthy Jatavallabhula. Robustpointset: A dataset for benchmarking robustness of point cloud classifiers. arXiv.org, 2011.11572, 2020.3", + "[56] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 1588–1597, 2019. 2, 3, 4", + "[57] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T Barron, and Pratul P Srinivasan. Ref-nerf: Structured view-dependent appearance for neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5481-5490, 2022. 3", + "[58] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. In Advances in Neural Information Processing Systems (NIPS), volume 34, pages 27171-27183, 2021. 3, 6, 7", + "[59] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P Srinivasan, Howard Zhou, Jonathan T Barron, Ricardo Martin-Brualla, Noah Snavely, and Thomas Funkhouser. Ibrnet: Learning multi-view image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4690-4699, 2021. 3, 5", + "[60] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "graph cnn for learning on point clouds. ACM Transactions on Graphics, 38(5):1-12, 2019. 3, 5", + "[61] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing (TIP), 13(4):600-612, 2004. 5", + "[62] Jiajun Wu, Chengkai Zhang, Tianfan Xue, Bill Freeman, and Josh Tenenbaum. Learning a probabilistic latent space of object shapes via 3d generative-adversarial modeling. In Advances in Neural Information Processing Systems (NIPS), volume 29, 2016. 3", + "[63] Tong Wu, Jiaqi Wang, Xingang Pan, Xudong Xu, Christian Theobalt, Ziwei Liu, and Dahua Lin. Voxurf: Voxel-based efficient and accurate neural surface reconstruction. arXiv.org, 2208.12697, 2022. 3, 6", + "[64] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1912-1920, 2015. 2, 3, 4", + "[65] Tiange Xiang, Chaoyi Zhang, Yang Song, Jianhui Yu, and Weidong Cai. Walk in the cloud: Learning curves for point clouds shape analysis. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 915–924, 2021. 5", + "[66] Mutian Xu, Runyu Ding, Hengshuang Zhao, and Xiaojuan Qi. Paconv: Position adaptive convolution with dynamic kernel assembling on point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3173-3182, 2021. 5", + "[67] Mutian Xu, Junhao Zhang, Zhipeng Zhou, Mingye Xu, Xiaojuan Qi, and Yu Qiao. Learning geometry-disentangled representation for complementary understanding of 3d object point cloud. In Proceedings of the Conference on Artificial Intelligence (AAAI), volume 35, pages 3056-3064, 2021. 5", + "[68] Guandao Yang, Xun Huang, Zekun Hao, Ming-Yu Liu, Serge Belongie, and Bharath Hariharan. Pointflow: 3d point cloud generation with continuous normalizing flows. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4541-4550, 2019. 3", + "[69] Yao Yao, Zixin Luo, Shiwei Li, Jingyang Zhang, Yufan Ren, Lei Zhou, Tian Fang, and Long Quan. Blendedmvs: A large-scale dataset for generalized multi-view stereo networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1790-1799, 2020. 2", + "[70] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. In Advances in Neural Information Processing Systems (NIPS), volume 34, pages 4805-4815, 2021. 3, 6", + "[71] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4578-4587, 2021. 3, 5, 7" + ], + "bbox": [ + 503, + 92, + 890, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "813", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[72] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. In Advances in Neural Information Processing Systems (NIPS), 2022. 3, 7", + "[73] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 586-595, 2018. 5", + "[74] Linqi Zhou, Yilun Du, and Jiajun Wu. 3d shape generation and completion through point-voxel diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5826-5835, 2021. 3", + "[75] Qian-Yi Zhou, Jaesik Park, and Vladlen Koltun. Open3D: A modern library for 3D data processing. arXiv.org, 1801.09847, 2018. 4" + ], + "bbox": [ + 78, + 90, + 470, + 330 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "814", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2023/OmniObject3D_ Large-Vocabulary 3D Object Dataset for Realistic Perception, Reconstruction and Generation/243023a1-e64c-4bdc-a9c5-c30584f803ba_model.json b/2023/OmniObject3D_ Large-Vocabulary 3D Object Dataset for Realistic Perception, Reconstruction and Generation/243023a1-e64c-4bdc-a9c5-c30584f803ba_model.json new file mode 100644 index 0000000000000000000000000000000000000000..696c3232f24d3b52edb384b7f70894ed67ad9f5b --- /dev/null +++ b/2023/OmniObject3D_ Large-Vocabulary 3D Object Dataset for Realistic Perception, Reconstruction and Generation/243023a1-e64c-4bdc-a9c5-c30584f803ba_model.json @@ -0,0 +1,2347 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.196, + 0.131, + 0.777, + 0.177 + ], + "angle": 0, + "content": "OmniObject3D: Large-Vocabulary 3D Object Dataset for Realistic Perception, Reconstruction and Generation" + }, + { + "type": "text", + "bbox": [ + 0.196, + 0.203, + 0.832, + 0.239 + ], + "angle": 0, + "content": "Tong Wu\\(^{1,2}\\), Jiarui Zhang\\(^{1,3}\\), Xiao Fu\\(^{1}\\), Yuxin Wang\\(^{1,4}\\), Jiawei Ren\\(^{5}\\), Liang Pan\\(^{5}\\), Wayne Wu\\(^{1}\\), Lei Yang\\(^{1,3}\\), Jiaqi Wang\\(^{1}\\), Chen Qian\\(^{1}\\), Dahua Lin\\(^{1,2\\boxtimes}\\), Ziwei Liu\\(^{5\\boxtimes}\\)" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.24, + 0.941, + 0.276 + ], + "angle": 0, + "content": "\\(^{1}\\)Shanghai Artificial Intelligence Laboratory, \\(^{2}\\)The Chinese University of Hong Kong, \\(^{3}\\)SenseTime Research, \\(^{4}\\)Hong Kong University of Science and Technology, \\(^{5}\\)S-Lab, Nanyang Technological University" + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.287, + 0.671, + 0.523 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.685, + 0.293, + 0.837, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.73, + 0.357, + 0.812, + 0.365 + ], + "angle": 0, + "content": "Textured meshes" + }, + { + "type": "image", + "bbox": [ + 0.685, + 0.369, + 0.862, + 0.426 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.686, + 0.428, + 0.848, + 0.437 + ], + "angle": 0, + "content": "Point clouds Rendered images" + }, + { + "type": "image", + "bbox": [ + 0.676, + 0.44, + 0.865, + 0.51 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.721, + 0.512, + 0.822, + 0.522 + ], + "angle": 0, + "content": "Real-captured videos" + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.525, + 0.868, + 0.613 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.615, + 0.895, + 0.645 + ], + "angle": 0, + "content": "Figure 1. OmniObject3D is a large vocabulary 3D object dataset with massive high-quality real-scanned 3D objects and rich annotations. It supports various research topics, e.g., perception, novel view synthesis, neural surface reconstruction, and 3D generation." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.656, + 0.314, + 0.672 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.475, + 0.871 + ], + "angle": 0, + "content": "Recent advances in modeling 3D objects mostly rely on synthetic datasets due to the lack of large-scale real-scanned 3D databases. To facilitate the development of 3D perception, reconstruction, and generation in the real world, we propose OmniObject3D, a large vocabulary 3D object dataset with massive high-quality real-scanned 3D objects. OmniObject3D has several appealing properties: 1) Large Vocabulary: It comprises 6,000 scanned objects in 190 daily categories, sharing common classes with popular 2D datasets (e.g., ImageNet and LVIS), benefiting the pursuit of generalizable 3D representations. 2) Rich Annotations: Each 3D object is captured with both 2D and" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.657, + 0.895, + 0.823 + ], + "angle": 0, + "content": "3D sensors, providing textured meshes, point clouds, multiview rendered images, and multiple real-captured videos. 3) Realistic Scans: The professional scanners support high-quality object scans with precise shapes and realistic appearances. With the vast exploration space offered by OmniObject3D, we carefully set up four evaluation tracks: a) robust 3D perception, b) novel-view synthesis, c) neural surface reconstruction, and d) 3D object generation. Extensive studies are performed on these four benchmarks, revealing new observations, challenges, and opportunities for future research in realistic 3D vision." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.842, + 0.633, + 0.858 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.894, + 0.903 + ], + "angle": 0, + "content": "Sensing, understanding, and synthesizing realistic 3D objects is a long-standing problem in computer vision, with" + }, + { + "type": "page_footnote", + "bbox": [ + 0.078, + 0.887, + 0.462, + 0.902 + ], + "angle": 0, + "content": "Corresponding authors. https://omniobject3d.github.io/" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "803" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.243 + ], + "angle": 0, + "content": "rapid progress emerging in recent years. However, a majority of the technical approaches rely on unrealistic synthetic datasets [6, 19, 64] due to the absence of a large-scale real-world 3D object database. However, the appearance and distribution gaps between synthetic and real data cannot be compensated for trivially, hindering their real-life applications. Therefore, it is imperative to equip the community with a large-scale and high-quality 3D object dataset from the real world, which can facilitate a variety of 3D vision tasks and downstream applications." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.246, + 0.473, + 0.413 + ], + "angle": 0, + "content": "Recent advances partially fulfill the requirements while still being unsatisfactory. As shown in Table 1, CO3D [48] contains 19k videos capturing objects from 50 MS-COCO categories, while only \\(20\\%\\) of the videos are annotated with accurate point clouds reconstructed by COLMAP [50]. Moreover, they do not provide textured meshes. GSO [16] has 1k scanned objects while covering only 17 household classes. AKB-48 [33] focuses on robotics manipulation with 2k articulated object scans in 48 categories, but the focus on articulation leads to a relatively narrow semantic distribution, failing to support general 3D object research." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.414, + 0.473, + 0.718 + ], + "angle": 0, + "content": "To boost the research on general 3D object understanding and modeling, we present OmniObject3D: a large-vocabulary 3D object dataset with massive high-quality, real-scanned 3D objects. Our dataset has several appealing properties: 1) Large Vocabulary: It contains 6,000 high-quality textured meshes scanned from real-world objects, which, to the best of our knowledge, is the largest among real-world 3D object datasets with accurate 3D meshes. It comprises 190 daily categories, sharing common classes with popular 2D and 3D datasets (e.g., ImageNet [15], LVIS [25], and ShapeNet [6]), incorporating most daily object realms (See Figure 1 and Figure 2). 2) Rich Annotations: Each 3D object is captured with both 2D and 3D sensors, providing textured 3D meshes, sampled point clouds, posed multi-view images rendered by Blender [13], and real-captured video frames with foreground masks and COLMAP camera poses. 3) Realistic Scans: The object scans are of high fidelity thanks to the professional scanners, bearing precise shapes with geometric details and realistic appearance with high-frequency textures." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Taking advantage of the vast exploration space offered by OmniObject3D, we carefully set up four evaluation tracks: a) robust 3D perception, b) novel-view synthesis, c) neural surface reconstruction, and d) 3D object generation. Extensive studies are performed on these benchmarks: First, the high-quality, real-world point clouds in OmniObject3D allow us to perform robust 3D perception analysis on both out-of-distribution (OOD) styles and corruptions, two major challenges in point cloud OOD generalization. Furthermore, we provide massive 3D models with multi-view images and precise 3D meshes for novel-view synthesis and neural surface reconstruction. The broad diver" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.089, + 0.895, + 0.131 + ], + "angle": 0, + "content": "Table 1. A comparison between OmniObject3D and other commonly-used 3D object datasets. \\( R^{LVis} \\) denotes the ratio of the 1.2k LVIS [25] categories being covered." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.131, + 0.903, + 0.291 + ], + "angle": 0, + "content": "
DatasetRealFull MeshVideo# Objs# CatsR1vis (%)
ShapeNet [6]✓51k554.1
ModelNet [64]✓12k402.4
3D-Future [19]✓16k341.3
ABO [12]✓8k633.5
Toys4K [53]✓4k1057.7
CO3D V1 / V2 [48]✓✓19 / 40k504.2
DTU [1]✓✓124NA0
ScanObjectNN [56]✓15k151.3
GSO [16]✓✓1k170.9
AKB-48 [33]✓✓2k481.8
Ours✓✓✓6k19010.8
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.299, + 0.895, + 0.421 + ], + "angle": 0, + "content": "sity in shapes and textures offers a comprehensive training and evaluation source for both scene-specific and generalizable algorithms. Finally, we equip the community with a database for large vocabulary and realistic 3D object generation, which pushes the boundary of existing state-of-the-art generation methods to real-world 3D objects. The four benchmarks reveal new observations, challenges, and opportunities for future research in realistic 3D vision." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.437, + 0.651, + 0.453 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.463, + 0.895, + 0.903 + ], + "angle": 0, + "content": "3D Object Datasets. The acquisition of a large-scale realistic 3D database is usually expensive and challenging. Many widely-used 3D datasets prefer to collect synthetic CAD models from online repositories [6, 53, 64], for example, ShapeNet [6] has 51,300 3D CAD models in 55 categories, and ModelNet40 [64] consists of 12,311 models in 40 categories. Recent works, e.g., 3D-FUTURE [19] and ABO [12], introduce high-quality CAD models with rich geometric details and informative textures. However, due to the inevitable gap between synthetic and real objects, the community is still eager for a large-scale 3D object dataset in the real world. DTU [1] and BlendedMVS [69] are photo-realistic datasets designed for multi-view stereo benchmarks, while they are small in scale and lack category annotations. ScanObjectNN [56] is a real-world point cloud object dataset based on scanned indoor scenes, containing around 15,000 objects with colored point clouds in 15 categories. However, the point clouds are incomplete and noisy, and multiple objects usually co-exist in one scene. GSO [16] has 1,030 scanned objects with fine geometries and textures in 17 household items, and AKB-48 [33] focuses on robotics manipulation with 2,037 articulated object models in 48 articulated object categories. However, the relatively narrow semantic scope of GSO and AKB-48 hinders their applications for more general 3D research. CO3D v1 [48] contains 19,000 object-centric videos, while only \\(20\\%\\) of them are annotated with accurate point clouds reconstructed by COLMAP [50], and they do not provide meshes or textures. In contrast, the proposed OmniOb" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.945, + 0.511, + 0.956 + ], + "angle": 0, + "content": "804" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.089, + 0.65, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.661, + 0.099, + 0.865, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.26, + 0.892, + 0.289 + ], + "angle": 0, + "content": "Figure 2. Semantic distribution of the OmniObject3D dataset. It covers 190 daily categories with a long-tailed distribution, sharing common classes with popular 2D and 3D datasets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.296, + 0.471, + 0.37 + ], + "angle": 0, + "content": "project3D dataset comprises 6,000 3D objects scanned by professional devices with meshes, textures, and multi-view photos in 190 categories, fulfilling the requirements of a wide range of research objectives. A detailed comparison is presented in Table 1." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.372, + 0.471, + 0.644 + ], + "angle": 0, + "content": "Robust 3D Perception. Robustness to out-of-distribution (OOD) data is important in point cloud perception. Two main challenges include OOD styles (e.g., differences between CAD models and real-world objects) and OOD corruptions (e.g., random point jittering or missing due to sensory inaccuracy). A branch of works [10, 30, 45, 60] studies the OOD corruptions and proposes standard corruption test suites [49, 55], while they fail to take account of OOD styles. Another branch of works [3, 48] evaluates the sim-to-real domain gap by training models on clean synthetic datasets [64] and testing them on noisy real-world test sets [56], while OOD styles and corruptions cannot be disentangled under this setting for an independent analysis. In this work, we leverage high-quality, real-world point clouds from OmniObject3D to systematically measure the robustness against the OOD style and OOD corruptions, providing the first benchmark for fine-grained evaluation of the point cloud perception robustness." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.645, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Neural Radiance Field and Neural Surface Reconstruction. Neural radiance field (NeRF) [40] represents a scene with a fully-connected deep network (MLPs), which takes in hundreds of sampled points along each camera ray and outputs the predicted color and density. We can synthesize the image of an unseen view from a trained model via volume rendering. Inspired by the success of NeRF, massive follow-up efforts have been made to improve its quality [4, 5, 39, 57] and efficiency [7, 18, 42, 54]. A branch of works [8, 27, 35, 48, 59, 71] has also explored the generalization ability of NeRF-based frameworks, where they aim to learn priors from deep image features across multiple scenes. Beyond novel view synthesis, another trend of approaches [14, 43, 58, 63, 70] proposes to combine neural radiance field with implicit surface representations like Signed Distance Function (SDF), and they achieve accurate and mask-free surface reconstruction from multi-view im" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.296, + 0.892, + 0.461 + ], + "angle": 0, + "content": "ages. Since dense camera views of scenes are sometimes unavailable, recent advances explore surface reconstruction from sparse views. They address the problem by exploiting generalizable priors cross scenes for a generic surface prediction [36] or taking advantage of the estimated geometry cues estimated by pre-trained networks [72]. OmniObject3D can serve as a large-scale benchmark with realistic photos and accurate meshes for both training and evaluation. The high diversity in shape and appearance offers an opportunity for pursuing more generalizable and robust novel view synthesis and surface reconstruction methods." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.463, + 0.892, + 0.658 + ], + "angle": 0, + "content": "3D Object Generation. Early approaches [20,26,37,52,62] extend 2D generation frameworks to 3D voxels with a high computational cost. Some others adopt different 3D data formulations, e.g., point cloud [2,41,68,74], octree [29], and implicit representations [11,38]. However, it is nontrivial to generate complex and textured surfaces. Recent advances [9,21,44] explore the generation for textured 3D meshes, where GET3D [22] is a state-of-the-art approach that generates diverse meshes with rich geometry and textures in two branches. It is a promising but challenging task to train generative models on a large vocabulary and realistic dataset. We evaluate GET3D on our dataset and reveal several challenges and future opportunities." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.659, + 0.892, + 0.689 + ], + "angle": 0, + "content": "In supplementary materials, we present more detailed discussions on related works for different tracks." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.704, + 0.76, + 0.723 + ], + "angle": 0, + "content": "3. The OmniObject3D Dataset" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.731, + 0.892, + 0.775 + ], + "angle": 0, + "content": "In this section, we describe the data collection, processing, and annotation pipeline of OmniObject3D. We also introduce the statistics and distribution of it." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.787, + 0.875, + 0.803 + ], + "angle": 0, + "content": "3.1. Data Collection, Processing, and Annotation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.903 + ], + "angle": 0, + "content": "Category List Definition. In order to collect a large amount of 3D objects that are both commonly-distributed and highly diverse, we first pre-define a category list according to several popular 2D and 3D datasets [6,15,25,31, 32,51,64]. We cover most of the categories that lie within the application scope of the scanners and also dynamically" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "805" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.151 + ], + "angle": 0, + "content": "expand the list with reasonable new classes that are absent from the current list during collection. We end up with 190 widely-spread categories, which ensures a library with rich texture, geometry, and semantic information." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.153, + 0.473, + 0.424 + ], + "angle": 0, + "content": "Object Collection Pipeline. We then collect a variety of objects from each category and use professional 3D scanners to obtain high-resolution textured meshes. Specifically, we use the Shining 3D scanner1 and Artec Eva 3D scanner2 for objects in different scales. The scanning time varies with the properties of the object: it takes around 15 minutes to scan a small rigid object with a simple geometry (e.g., an apple, a toy), while it takes up to an hour to obtain a qualified 3D scan for non-rigid, complex, or large objects (e.g., a bed, a kite). For around \\(10\\%\\) of the objects, we conduct common manipulations (e.g., taking a bite, cutting in pieces) to conform the natural instincts of them. The 3D scans can faithfully retain the real-world scale of each object, but their poses are not strictly aligned. We thus pre-define a canonical pose for each category and manually align the objects within a category. We then check the quality of each scan, and around \\(83\\%\\) high-quality ones out of the total collection are finally reserved in the dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.425, + 0.473, + 0.865 + ], + "angle": 0, + "content": "Image Rendering and Point Cloud Sampling. To support a variety of research topics like point cloud analysis, neural radiance fields, and 3D generation, we render multiview images and sample point clouds based on the collected 3D models. We use Blender [13] to render object-centric and photo-realistic multi-view images, together with accurate camera poses. The images are rendered from 100 random viewpoints sampled on the upper hemisphere at \\(800 \\times 800\\) pixels. We also produce high-resolution mid-level cues like depth and normal for more research use. We then uniformly sample multi-resolution point clouds from each 3D model using the Open3D toolbox [75], with \\(2^{n}\\) (\\(n \\in \\{10, 11, 12, 13, 14\\}\\)) points in each point cloud, respectively. Besides the data existing in the dataset, we also provide a data generation pipeline. One can easily obtain new data with self-defined camera distributions, lighting, and point sampling methods to meet different requirements. Video Capturing and Annotation. After scanning each object, we capture its video with an iPhone 12 Pro mobile phone. The object is placed on or beside a calibration board, and each video covers a full \\(360^{\\circ}\\) range around it. Square corners on the calibration board can be recognized by the QR Codes beside it, and we then filter out the blurry frames whose recognized corners are less than 8. We uniformly sample 200 frames, and then COLMAP [50], a well-known SfM pipeline, is applied to annotate the frames with camera poses. Finally, we use the scales of the calibration board in both the SfM coordinate space and the real world to recover the absolute scale of the SfM coordinate system." + }, + { + "type": "image", + "bbox": [ + 0.529, + 0.096, + 0.872, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.187, + 0.895, + 0.243 + ], + "angle": 0, + "content": "Figure 3. OmniObject3D provides the first clean real-world point cloud object dataset and allows fine-grained analysis on robustness to OOD styles and OOD corruptions.\"-C\": corrupted by common corruptions described in [49]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.245, + 0.894, + 0.306 + ], + "angle": 0, + "content": "We also develop a two-stage matting pipeline based on the U\\(^2\\)Net [47] and FBA [17] matting model to generate the foreground masks for all the frames. Please refer to supplementary materials for more implementation details." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.316, + 0.738, + 0.331 + ], + "angle": 0, + "content": "3.2. Statistics and Distribution" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.339, + 0.895, + 0.506 + ], + "angle": 0, + "content": "With 6,000 3D models in 190 categories, OmniObject3D exhibits a long-tailed distribution with an average of around 30 objects in each category, as shown in Figure 2. It shares many common categories with several famous 2D and 3D datasets [6, 15, 25, 31, 32, 51, 64]. For example, it covers 85 categories in ImageNet [15] and 130 categories in LVIS [25], which leads to the highest \\( R^{\\mathrm{vis}} \\) in Table 1. Most of the categories are covered by the Open Images [31] image-level labels. It bears a huge diversity in shapes and appearances. The vast semantic and geometrical exploration spaces enable a wide range of research objectives." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.52, + 0.634, + 0.536 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.545, + 0.704, + 0.561 + ], + "angle": 0, + "content": "4.1. Robust 3D Perception" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.568, + 0.894, + 0.658 + ], + "angle": 0, + "content": "Object-level point cloud classification is one of the most fundamental tasks in point cloud perception. In this section, we show how OmniObject3D boosts robustness analysis of point cloud classification by disentangling the two critical out-of-distribution (OOD) challenges introduced in Sec. 2, i.e., OOD styles and OOD corruptions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.894, + 0.703 + ], + "angle": 0, + "content": "Existing robustness evaluation utilizes clean synthetic datasets, e.g., ModelNet [64], for training and sets up two kinds of test sets for evaluation:" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.704, + 0.892, + 0.81 + ], + "angle": 0, + "content": "1) Noisy real-world datasets, e.g., ScanObjectNN [56], which are cropped from real-world scenes. They are employed to measure the robustness of the sim-to-real domain gap. However, the gap couples both OOD styles and OOD corruptions simultaneously as the cropped point clouds are always noisy, making it impossible to analyze the two robustness challenges independently." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.81, + 0.892, + 0.871 + ], + "angle": 0, + "content": "2) Corrupted synthetic datasets, e.g., ModelNet-C [49], which are artificially perturbed on top of clean synthetic datasets. The evaluation allows for detailed corruption analysis, but they do not reflect the robustness to OOD styles." + }, + { + "type": "list", + "bbox": [ + 0.499, + 0.704, + 0.892, + 0.871 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "None of the existing robustness benchmarks allows for analyzing the robustness to both OOD styles and OOD cor" + }, + { + "type": "page_footnote", + "bbox": [ + 0.078, + 0.875, + 0.223, + 0.887 + ], + "angle": 0, + "content": "1https://www.einscan.com/" + }, + { + "type": "page_footnote", + "bbox": [ + 0.08, + 0.888, + 0.214, + 0.9 + ], + "angle": 0, + "content": "\\(^{2}\\)https://www.artec3d.cn/" + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.871, + 0.223, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "806" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.472, + 0.201 + ], + "angle": 0, + "content": "Table 2. Point cloud perception robustness analysis on OmniObject3D with different architecture designs. Models are trained on the ModelNet-40 dataset, with OAClean to be their overall accuracy on the standard ModelNet-40 test set. OAStyle on OmniObject3D evaluates the robustness to OOD styles. mCE on the corrupted OmniObject3D-C evaluates the robustness to OOD corruptions. Blue shadings indicate rankings. \\(\\dagger\\): results on ModelNet-C [49]. Full results are presented in the supplementary materials." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.201, + 0.481, + 0.368 + ], + "angle": 0, + "content": "
mCE†↓OA Clean ↑OA Style ↑mCE ↓
DGCNN [60]1.0000.9260.4481.000
PointNet [45]1.4220.9070.4660.969
PointNet++ [46]1.0720.9300.4071.066
RSCNN [34]1.1300.9230.3931.076
SimpleView [23]1.0470.9390.4760.990
GDANet [67]0.8920.9340.4970.920
PAConv [66]1.1040.9360.4031.073
CurveNet [65]0.9270.9380.5000.929
PCT [24]0.9250.9300.4590.940
RPC [49]0.8630.9300.4720.936
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.371, + 0.47, + 0.55 + ], + "angle": 0, + "content": "ruptions in fine granularity. OmniObject3D, on the other hand, as the first clean real-world point cloud object dataset, can help to address the issue. For models trained on ModelNet, we first evaluate their performance on OmniObject3D to examine OOD-style robustness. Then, we create OmniObject3D-C by corrupting OmniObject3D with common corruptions described in [49] to examine the OOD-corruption robustness. We show a complete robustness evaluation scheme in Figure 3. For evaluation metrics, we use the overall accuracy (OA) on OmniObject3D to measure the OOD-style robustness and use DGCNN normalized mCE [49] to measure the OOD-corruption robustness." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.551, + 0.47, + 0.793 + ], + "angle": 0, + "content": "We benchmark ten state-of-the-art point cloud classification models in Table 2. We observe that 1) performance on a clean test set has little correlation with OOD-style robustness. For example, SimpleView [23] achieves the best \\(\\mathrm{OA}_{\\mathrm{Clean}}\\) but mediocre \\(\\mathrm{OA}_{\\mathrm{Style}}\\); 2) Advanced point grouping, e.g., curve-based point grouping in CurveNet [65] and frequency-based point grouping in GDANet [67], are robust not only to OOD corruptions as pointed out in [49], but also to OOD styles; 3) OOD style + OOD corruption is a more challenging setting. In particular, RPC, the most robust architecture to OOD corruptions [49], shows inferior mCE. In summary, robust point cloud perception models against both OOD styles and OOD corruptions are still under-explored. Our dataset sheds new light on a comprehensive understanding of point cloud perception robustness. See more results in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.802, + 0.279, + 0.818 + ], + "angle": 0, + "content": "4.2. Novel View Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.47, + 0.903 + ], + "angle": 0, + "content": "In this section, we study several representative methods on OmniObject3D for novel view synthesis (NVS) in two settings: 1) training on a single scene with densely captured images, which is the standard setting for NeRF [40]; 2) learning priors across scenes from our dataset to explore" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.089, + 0.892, + 0.117 + ], + "angle": 0, + "content": "Table 3. Single-scene novel view synthesis results. Three metrics and their standard deviation (SD) across the training set." + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.117, + 0.902, + 0.187 + ], + "angle": 0, + "content": "
MethodPSNR (↑) / SDSSIM (↑) / SDLPIPS (↓) / SD
NeRF [40]34.01 / 3.460.953 / 0.0290.068 / 0.061
mip-NeRF [4]39.86 / 4.580.974 / 0.0130.084 / 0.048
Plenoxels [18]41.04 / 6.840.982 / 0.0310.030 / 0.031
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.187, + 0.894, + 0.226 + ], + "angle": 0, + "content": "Table 4. Cross-scene novel view synthesis results on 10 categories. 'Cat.' and 'All*' denote training on each category and training on all categories except the 10 test ones, respectively." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.226, + 0.905, + 0.393 + ], + "angle": 0, + "content": "
MethodTrainPSNR (↑)SSIM (↑)LPIPS (↓)\\( {\\mathcal{L}}_{1}^{\\text{depth }}\\left( \\downarrow \\right) \\)
MVSNeRF [8]All*17.490.5440.4420.193
Cat.17.540.5420.4480.230
All*-ft.25.700.7540.2510.081
Cat.-ft.25.520.7500.2640.076
IBRNet [59]All*19.390.5690.3990.423
Cat.19.030.5510.4150.290
All*-ft.26.890.7920.2150.081
Cat.-ft.25.670.7600.2380.099
pixelNeRF [71]All*22.160.6920.3420.109
Cat.20.650.6760.3480.195
" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.401, + 0.819, + 0.416 + ], + "angle": 0, + "content": "the generalization ability of NeRF-style models." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.417, + 0.892, + 0.734 + ], + "angle": 0, + "content": "Single-Scene NVs. We select three objects in each category for the experiments, randomly sampling 1/8 images as the hold-out test set. We involve NeRF [40], mip-NeRF [4], and a voxel-based system named Plenoxels [18] for comparison. As in Table 3, we find that Plenoxels achieve the best performance on average for PSNR, SSIM [61], and LPIPS [73]. There exists a clear margin for LPIPS between Plenoxels and the other two methods since voxel-based methods are especially good at modeling high-frequency appearance. We also present the standard deviation (SD) of results across all the training samples, where Plenoxels are relatively unstable compared to NeRF and mip-NeRF. We observe that Plenoxels introduce artifacts when encountering concave geometry (e.g., bowls, chairs) and suffer from an inaccurate density field modeling when the foreground object is dark. MLP-based methods are relatively robust against these difficult cases. In a nutshell, our dataset provides a library with a variety of shapes and appearances, allowing a comprehensive evaluation of different NVS methods. See the supplementary for more results with the iPhone videos, detailed analysis, and visualizations." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Cross-Scene NVS. We conduct extensive experiments on novel view synthesis from sparse inputs by pixelNeRF [71], IBRNet [59] and MVSNeRF [8] in Table 4. We select 10 categories with the most various scenes as the test set. All metrics are averaged over 300 images. In the generalization setting, although not trained on the test category, \\(\\mathrm{MVSNeRF_{All^{*}}}\\) is comparable to \\(\\mathrm{MVSNeRF_{Cat}}\\), \\(\\mathrm{IBRNet_{All^{*}}}\\) and \\(\\mathrm{pixelNeRF_{All^{*}}}\\) even outperforms the corresponding \"Cat\"s in all terms of visual metrics, especially on regular-shaped objects such as squash and apple. It confirms that OmniObject3D serves as an information-rich" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.945, + 0.511, + 0.956 + ], + "angle": 0, + "content": "807" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.09, + 0.874, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.194, + 0.328, + 0.774, + 0.342 + ], + "angle": 0, + "content": "Figure 4. Neural surface reconstruction results for both dense-view and sparse-view settings." + }, + { + "type": "table_caption", + "bbox": [ + 0.114, + 0.343, + 0.434, + 0.355 + ], + "angle": 0, + "content": "Table 5. Dense-view surface reconstruction results." + }, + { + "type": "table", + "bbox": [ + 0.115, + 0.356, + 0.434, + 0.468 + ], + "angle": 0, + "content": "
MethodChamfer Distance ×103(↓)
HardMediumEasyAvg
NeuS [58]9.265.633.466.09
VolSDF [70]10.064.942.865.92
Voxurf [63]9.014.982.585.49
Avg9.445.192.975.83
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.475, + 0.473, + 0.777 + ], + "angle": 0, + "content": "dataset that is beneficial for obtaining a strong generalizable prior on unseen scenes. Moreover, it is noteworthy that MVSNeRF and pixelNeRF with 'All* generate better underlying depth than those with 'Cat.', inferring generalizable methods can implicitly learn geometric cues though only trained from appearance in our dataset. It is reasonable that (1) IBRNet suffers more severely than the others in geometry under the scarcity of source context (only 3 views) as it is more suitable for dense-view generalization that complies with its view interpolation module. (2) MVSNeRF lags behind pixelNeRF on visual performance as we take 10 test frames widely distributed around the object in \\(360^{\\circ}\\) by FPS sampling algorithm, where cost volume will be inaccurate on large-range viewpoint change. After further finetuning IBRNet for only around 10 minutes on each test scene, \\(\\mathrm{IBRNet}_{\\mathrm{All}^{*}\\text{-ft}}\\) achieves the best view synthesis results, comparable to test-time optimized NVS methods on nearby views. It is promising to utilize the large-scale and category-prosperous OmniObject3D, to build a benchmark suite for evaluating diverse cross-scene NVS methods." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.787, + 0.353, + 0.803 + ], + "angle": 0, + "content": "4.3. Neural Surface Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Precise surface reconstruction from multi-view images enables a broad range of applications. For a single scene with dense-view images, algorithms are expected to conduct accurate, robust, and efficient surface reconstruction. When only sparse-view images are available, it is crucial to learn generalizable priors from a set of scenes or use other" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.343, + 0.868, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.49, + 0.892, + 0.532 + ], + "angle": 0, + "content": "Figure 5. Performance distribution of dense-view surface reconstruction. The averaged results of the three methods is imbalanced. The colored area denotes a smoothed range of results." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.537, + 0.892, + 0.566 + ], + "angle": 0, + "content": "geometric cues to assist reconstruction. Accordingly, we study the two settings for surface reconstruction methods." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.568, + 0.892, + 0.674 + ], + "angle": 0, + "content": "Dense-View Surface Reconstruction. Under this setting, we include three representative methods. NeuS [58] and VolSDF [70] are two well-known approaches that bridge neural volume rendering with implicit surface representation. We also involve a voxel-based method called Voxsurf [63], which leverages a hybrid representation to achieve acceleration and fine geometry reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Previous approaches in this task mainly perform evaluations and comparisons on 15 scenes from the DTU [1] dataset, which is not comprehensive and robust enough to demonstrate the ability of the methods in different scenarios. In comparison, we select three objects per category to run each of the three methods above, leading to over 1,500 reconstructions in total. We calculate the Chamfer Distance (CD) between the reconstructed surface and the ground truth. The distribution of the results is shown in Figure 5. The average curve is imbalanced: hard categories usually include low-textured, concave, or complex shapes (e.g., bowls, vases, kennels, cabinets, and durians). We thus split the categories into three levels of \"difficulty\" based on the average curve, and the level-wise results are presented in Table 5. We can observe a clear margin among results in" + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.956 + ], + "angle": 0, + "content": "808" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.085, + 0.09, + 0.462, + 0.103 + ], + "angle": 0, + "content": "Table 6. Sparse-view (3-view) surface reconstruction results." + }, + { + "type": "table", + "bbox": [ + 0.088, + 0.105, + 0.46, + 0.304 + ], + "angle": 0, + "content": "
MethodTrainChamfer Distance ×103(↓)
HardMediumEasyAvg
NeuS [58]Single29.3527.6224.7927.3
MonoSDF [72]Single35.1435.3532.7634.68
SparseNeuS [36]1 cat.34.0531.3231.1432.36
10 cats.30.7530.1128.3729.87
All cats.26.1326.0822.1325.00
Easy28.3926.6523.7626.48
Medium27.3826.6623.0825.87
Hard27.4226.9524.6326.47
MVSNeRF [8]All cats.56.6848.0948.7051.16
pixelNeRF [71]All cats.63.3159.9161.4761.56
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.307, + 0.468, + 0.337 + ], + "angle": 0, + "content": "different levels for each method, indicating the split subsets to be generic and faithful." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.339, + 0.468, + 0.55 + ], + "angle": 0, + "content": "Sparse-View Surface Reconstruction. Dense-captured images of a scene are sometimes not available, so we also study the sparse-view scenario here. The following methods are included: NeuS [58] with sparse-view input; MonoSDF [72], which takes in geometry cues estimated by pre-trained models; SparseNeuS [36], a generic surface prediction pipeline that learns generalizable priors; pixelNeRF [71] and MVSNeRF [8] from Sec. 4.2, whose geometries are extracted from the density field. For NeuS, MonoSDF, and pixelNeRF, we use Farthest Point Sampling (FPS) to sample views that are most widely distributed; for SparseNeuS and MVSNeRF, we conduct FPS among the nearest 30 camera poses from a random reference view. We sample 3 views in all the experiments." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.552, + 0.468, + 0.824 + ], + "angle": 0, + "content": "The quantitative and qualitative comparisons are shown in Table 6 and Figure 4, respectively. We observe apparent artifacts in all the sparse-view reconstructed results. Among them, SparseNeuS trained on enough data demonstrates the best quantitative performance on average, and the pre-division on the train set does not result in a noticeable difference across difficulty levels. NeuS with sparse-view input achieves a surprisingly good performance. As shown in Figure 4, the FPS sampling equips NeuS with a coherent global shape for thin structures like case 3, while it encounters severe local geometry ambiguity like case 1. MonoSDF, on the contrary, partially overcomes the issue of ambiguity via the assistance of predicted geometry cues in case 1. However, it relies heavily on the accuracy of the estimated depth and normal and thus easily fails when the estimation is inaccurate (e.g., case 2 and 3). The surfaces extracted from generalized NeRF models, i.e., pixelNeRF and MVSNeRF, are of relatively low quality." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.468, + 0.901 + ], + "angle": 0, + "content": "In brief, the challenging problem of sparse-view surface reconstruction has not been solved well currently. OmniObject3D is a promising database to study generalizable surface reconstruction pipelines as well as strategies for a robust usage of estimated geometry cues for this track." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.093, + 0.681, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.692, + 0.093, + 0.889, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.242, + 0.681, + 0.388 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.695, + 0.242, + 0.88, + 0.388 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.392, + 0.892, + 0.49 + ], + "angle": 0, + "content": "Figure 6. The category distribution of the generated shapes. (a) shows a weak positive correlation between the number of generated shapes and training shapes per category. (b) visualizes the correlation matrix among different categories by Chamfer Distance between their mean shapes. (c) visualizes categories being clustered into eight groups by KMeans. (d) presents a clear training-generation relation in the group-level statistics." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.498, + 0.706, + 0.514 + ], + "angle": 0, + "content": "4.4.3D Object Generation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.522, + 0.892, + 0.657 + ], + "angle": 0, + "content": "In this section, we adopt a state-of-the-art generative model that directly generates explicit textured 3D meshes, namely GET3D [22]. GET3D is originally evaluated on six categories (Car, Chair, Motorbike, Animal, House, and Human Body) with independent models trained on each category. The number of shapes per category ranges from 337 to 7,497. In comparison, OmniObject3D contains many more categories with fewer objects in each. As a result, it is natural to train each model with multiple categories." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.892, + 0.809 + ], + "angle": 0, + "content": "We first provide some qualitative results in Figure 7, where we show various generation results: The textures are rather realistic, and the shapes are coherent, enhanced by fine geometry details (e.g., the lychee and pineapple). We explore the latent space of the model and show interpolation results in Figure 8. We can observe a smooth transition across instances that are semantically different. We would like to further analyze the performance of the generative model trained on OmniObject3D from three aspects, i.e., semantic distribution, diversity, and quality." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Semantic Distribution. We randomly select 100 categories to train an unconditional model jointly. We randomly generate 1,000 textured meshes at inference time and ask human experts to label them. Shapes with ambiguity are not counted. Figure 6 (a) shows that the generated shapes per category are highly imbalanced, exhibiting a weak positive" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "809" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.092, + 0.891, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.212, + 0.893, + 0.24 + ], + "angle": 0, + "content": "Figure 7. Examples of the generated textured shapes rendered in Blender. OmniObject3D enables GET3D with realistic generations across a wide range of categories." + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.242, + 0.466, + 0.326 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.33, + 0.47, + 0.358 + ], + "angle": 0, + "content": "Figure 8. Shape interpolation. We interpolate both geometry and texture latent codes from left to right." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.371, + 0.47, + 0.657 + ], + "angle": 0, + "content": "correlation with the training shape numbers. Actually, the categories are not independent but rather highly correlated. We calculate the \"mean shape\" for each category and visualize the Chamfer Distance among them in Figure 6 (b), which indicates that they can be further grouped. Regarding each matrix row as a feature vector, we use KMeans to cluster them into eight groups (Figure 6 (c)) and carry out group-level statistics in Figure 6 (d). It demonstrates a clear trend that the number of generated shapes increases along with or even faster than the number of training shapes in the group, revealing an enlarged semantic bias during generation. However, it also depends on the inner-group divergence. For example, Group 2 (883 shapes in 27 categories) has the largest number of training samples, while the high variation among its categories prevents it from dominating the generated shapes; Group 1 (587 shapes in 18 categories) has a relatively small divergence, which becomes the most popular in the generated shapes. We present more details in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.66, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Diversity and Quality. We select four representative data subsets for training and evaluation, namely fruits, furniture, toys, and Rand-100. We randomly split each subset into training (80%) and testing (20%). We leverage three evaluation metrics: for geometry, we use Chamfer Distance (CD) to compute the Coverage score (Cov) and Minimum Matching Distance (MMD), which focus on the diversity and quality of the shapes, respectively; for texture, we adopt the widely-used FID [28]. But the FID metrics with different test splits are not directly comparable, suffering from a large variance when the test set is small. We thus introduce FIDref for reference, which is the FID between the train and test set. The results are shown in Table 7. Furniture suffers from the lowest quality (MMD) since the small train set with 17 categories is a difficult training source. Fruits has the same number of categories while being 2.3 times" + }, + { + "type": "table_caption", + "bbox": [ + 0.516, + 0.24, + 0.874, + 0.254 + ], + "angle": 0, + "content": "Table 7. Quantitative evaluations on different data splits." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.254, + 0.895, + 0.336 + ], + "angle": 0, + "content": "
Split#Objs#CatsCov (% ↑)MMD (↓)FID (↓)FIDref
Furniture2651767.924.2787.3958.40
Fruits6101746.723.32105.3187.15
Toys339755.222.78122.7741.40
Rand-100295110061.703.8946.578.65
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.343, + 0.892, + 0.448 + ], + "angle": 0, + "content": "larger in scale, and some fruits share a very similar structure, leading to relatively higher quality and lower diversity (Cov). Toys achieve the best quality by training on only 7 categories. Rand-100 is the most difficult case, and we can observe a trade-off between quality and diversity. Both the FID and \\(\\mathrm{FID}^{\\mathrm{ref}}\\) are high for the first three subsets due to the small testing sets, while only 'Rand-100' is relatively low." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.45, + 0.893, + 0.54 + ], + "angle": 0, + "content": "In brief, training and evaluating generative models on a large-vocabulary and realistic dataset is a promising but challenging task. We reveal crucial problems like the semantic distribution bias and varying exploration difficulties in different groups. OmniObject3D serves as a great database for further examination in this area." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.553, + 0.729, + 0.569 + ], + "angle": 0, + "content": "5. Conclusion and Outlook" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.578, + 0.892, + 0.713 + ], + "angle": 0, + "content": "We introduce OmniObject3D, a large vocabulary 3D object dataset with massive high-quality real-scanned 3D objects, including 6,000 objects from 190 categories. It provides rich annotations, including textured 3D meshes, sampled point clouds, posed multi-view images rendered by Blender, and real-captured video frames with foreground masks and COLMAP camera poses. We set up four evaluation tracks, revealing new observations, challenges, and opportunities for future research in realistic 3D vision." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.714, + 0.892, + 0.744 + ], + "angle": 0, + "content": "We will regulate the usage of our data to avoid potential negative social impacts." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.745, + 0.892, + 0.895 + ], + "angle": 0, + "content": "Acknowledgement. This project is funded by Shanghai AI Laboratory, CUHK Interdisciplinary AI Research Institute, the Centre for Perceptual and Interactive Intelligence (CPII) Ltd under the Innovation and Technology Commission (ITC)'s InnoHK, Hong Kong RGC Theme-based Research Scheme 2020/21 (No. T41-603/20-R), OpenXDLab, the Ministry of Education, Singapore, under its MOE AcRF Tier 2 (MOE-T2EP20221-0012), NTU NAP, and under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "810" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.117, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Henrik Aanaes, Rasmus Ramsbøl Jensen, George Vogiatzis, Engin Tola, and Anders Bjorholm Dahl. Large-scale data for multiple-view stereopsis. International Journal of Computer Vision (IJCV), 120(2):153-168, 2016. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.174, + 0.472, + 0.243 + ], + "angle": 0, + "content": "[2] Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas Guibas. Learning representations and generative models for 3d point clouds. In Proceedings of the International Conference on Machine learning (ICML), pages 40-49, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.246, + 0.472, + 0.327 + ], + "angle": 0, + "content": "[3] Adel Ahmadyan, Liangkai Zhang, Artsiom Ablavatski, Jianing Wei, and Matthias Grundmann. Objectron: A large scale dataset of object-centric videos in the wild with pose annotations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7822-7831, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.331, + 0.472, + 0.413 + ], + "angle": 0, + "content": "[4] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5855–5864, 2021. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.416, + 0.472, + 0.485 + ], + "angle": 0, + "content": "[5] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5470-5479, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.488, + 0.472, + 0.557 + ], + "angle": 0, + "content": "[6] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv.org, 1512.03012, 2015. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.56, + 0.472, + 0.613 + ], + "angle": 0, + "content": "[7] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.617, + 0.472, + 0.699 + ], + "angle": 0, + "content": "[8] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14124-14133, 2021. 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.703, + 0.472, + 0.771 + ], + "angle": 0, + "content": "[9] Wenzheng Chen, Huan Ling, Jun Gao, Edward Smith, Jaakko Lehtinen, Alec Jacobson, and Sanja Fidler. Learning to predict 3d objects with an interpolation-based differentiable renderer. In Advances in Neural Information Processing Systems (NIPS), volume 32, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.774, + 0.47, + 0.843 + ], + "angle": 0, + "content": "[10] Yunlu Chen, Vincent Tao Hu, Efstratios Gavves, Thomas Mensink, Pascal Mettes, Pengwan Yang, and Cees GM Snoek. Pointmixup: Augmentation for point clouds. In Proceedings of the European Conference on Computer Vision (ECCV), pages 330-345, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[11] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5939-5948, 2019. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.117, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.188 + ], + "angle": 0, + "content": "[12] Jasmine Collins, Shubham Goel, Kenan Deng, Achleshwar Luthra, Leon Xu, Erhan Gundogdu, Xi Zhang, Tomas F Yago Vicente, Thomas Dideriksien, Himanshu Arora, et al. Abo: Dataset and benchmarks for real-world 3d object understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21126-21136, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.191, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[13] Blender Online Community. Blender - a 3d modelling and rendering package. 2018. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[14] François Darmon, Bénédicte Bascle, Jean-Clément Devaux, Pascal Monasse, and Mathieu Aubry. Improving neural implicit surfaces geometry with patch warping. arXiv.org, 2112.09648, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.278, + 0.892, + 0.346 + ], + "angle": 0, + "content": "[15] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 248–255, 2009. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.349, + 0.892, + 0.416 + ], + "angle": 0, + "content": "[16] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas B McHugh, and Vincent Vanhoucke. Google scanned objects: A high-quality dataset of 3d scanned household items. arXiv.org, 2204.11918, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.419, + 0.892, + 0.446 + ], + "angle": 0, + "content": "[17] Marco Forte and François Pitié. F, b, alpha matting. arXiv.org, 2003.07711, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.449, + 0.894, + 0.518 + ], + "angle": 0, + "content": "[18] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5501–5510, 2022. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.52, + 0.892, + 0.574 + ], + "angle": 0, + "content": "[19] Huan Fu, Rongfei Jia, Lin Gao, Mingming Gong, Binqiang Zhao, Steve Maybank, and Dacheng Tao. 3d-future: 3d furniture shape with texture. International Journal of Computer Vision (IJCV), 129(12):3313-3337, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.577, + 0.892, + 0.63 + ], + "angle": 0, + "content": "[20] Matheus Gadelha, Subhransu Maji, and Rui Wang. 3d shape induction from 2d views of multiple objects. In Proceedings of the International Conference on 3D Vision (3DV), pages 402-411, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.633, + 0.892, + 0.701 + ], + "angle": 0, + "content": "[21] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. arXiv.org, 2209.11163, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.704, + 0.892, + 0.773 + ], + "angle": 0, + "content": "[22] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. In Advances in Neural Information Processing Systems (NIPS), 2022. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.775, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[23] Ankit Goyal, Hei Law, Bowei Liu, Alejandro Newell, and Jia Deng. Revisiting point cloud shape classification with a simple and effective baseline. In Proceedings of the International Conference on Machine learning (ICML), pages 3809-3820, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.846, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[24] Meng-Hao Guo, Jun-Xiong Cai, Zheng-Ning Liu, Tai-Jiang Mu, Ralph R Martin, and Shi-Min Hu. Pct: Point cloud transformer. Computational Visual Media, 7(2):187-199, 2021. 5" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "811" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.16 + ], + "angle": 0, + "content": "[25] Agrim Gupta, Piotr Dollar, and Ross Girshick. LVIS: A dataset for large vocabulary instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5356-5364, 2019. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.47, + 0.218 + ], + "angle": 0, + "content": "[26] Philipp Henzler, Niloy J Mitra, and Tobias Ritschel. Escaping plato's cave: 3d shape from adversarial rendering. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9984-9993, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.219, + 0.47, + 0.301 + ], + "angle": 0, + "content": "[27] Philipp Henzler, Jeremy Reizenstein, Patrick Labatut, Roman Shapovalov, Tobias Ritschel, Andrea Vedaldi, and David Novotny. Unsupervised learning of 3d object categories from videos in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4700-4709, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.302, + 0.47, + 0.37 + ], + "angle": 0, + "content": "[28] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In Advances in Neural Information Processing Systems (NIPS), volume 30, 2017. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.372, + 0.469, + 0.413 + ], + "angle": 0, + "content": "[29] Moritz Ibing, Gregor Kobsik, and Leif Kobbelt. Octree transformer: Autoregressive 3d shape generation on hierarchically structured sequences. arXiv.org, 2111.12480, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.414, + 0.469, + 0.483 + ], + "angle": 0, + "content": "[30] Sihyeon Kim, Sanghyeok Lee, Dasol Hwang, Jaewon Lee, Seong Jae Hwang, and Hyunwoo J Kim. Point cloud augmentation with weighted local transformations. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 548-557, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.484, + 0.469, + 0.552 + ], + "angle": 0, + "content": "[31] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4. International Journal of Computer Vision (IJCV), 128(7):1956-1981, 2020. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.553, + 0.469, + 0.621 + ], + "angle": 0, + "content": "[32] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Proceedings of the European Conference on Computer Vision (ECCV), pages 740-755, 2014. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.622, + 0.469, + 0.692 + ], + "angle": 0, + "content": "[33] Liu Liu, Wenqiang Xu, Haoyuan Fu, Sucheng Qian, Qiao-jun Yu, Yang Han, and Cewu Lu. Akb-48: A real-world articulated object knowledge base. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14809-14818, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.692, + 0.469, + 0.761 + ], + "angle": 0, + "content": "[34] Yongcheng Liu, Bin Fan, Shiming Xiang, and Chunhong Pan. Relation-shape convolutional neural network for point cloud analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8895-8904, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.762, + 0.469, + 0.844 + ], + "angle": 0, + "content": "[35] Yuan Liu, Sida Peng, Lingjie Liu, Qianqian Wang, Peng Wang, Christian Theobalt, Xiaowei Zhou, and Wenping Wang. Neural rays for occlusion-aware image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7824-7833, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[36] Xiaoxiao Long, Cheng Lin, Peng Wang, Taku Komura, and Wenping Wang. Sparseneus: Fast generalizable neural surface reconstruction from sparse views. arXiv.org, 2206.05737, 2022. 3, 7" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[37] Sebastian Lunz, Yingzhen Li, Andrew Fitzgibbon, and Nate Kushman. Inverse graphics gan: Learning to generate 3d shapes from unstructured 2d data. arXiv.org, 2002.12674, 2020.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[38] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4460-4470, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.892, + 0.303 + ], + "angle": 0, + "content": "[39] Ben Mildenhall, Peter Hedman, Ricardo Martin-Brualla, Pratul P Srinivasan, and Jonathan T Barron. Nerf in the dark: High dynamic range view synthesis from noisy raw images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16190-16199, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.305, + 0.892, + 0.375 + ], + "angle": 0, + "content": "[40] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In Proceedings of the European Conference on Computer Vision (ECCV), pages 405-421, 2020. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.377, + 0.892, + 0.431 + ], + "angle": 0, + "content": "[41] Kaichun Mo, Paul Guerrero, Li Yi, Hao Su, Peter Wonka, Niloy J. Mitra, and Leonidas J. Guibas. Structurenet: Hierarchical graph networks for 3d shape generation. arXiv.org, 1908.00575, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.433, + 0.892, + 0.488 + ], + "angle": 0, + "content": "[42] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.49, + 0.892, + 0.56 + ], + "angle": 0, + "content": "[43] Michael Oechsle, Songyou Peng, and Andreas Geiger. Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5589-5599, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.562, + 0.892, + 0.63 + ], + "angle": 0, + "content": "[44] Dario Pavllo, Jonas Kohler, Thomas Hofmann, and Aurelien Lucchi. Learning generative models of textured 3d meshes from real-world images. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 13879-13889, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.632, + 0.892, + 0.702 + ], + "angle": 0, + "content": "[45] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: deep learning on point sets for 3d classification and segmentation. corr abs/1612.00593 (2016). In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 652-660, 2017. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.704, + 0.892, + 0.759 + ], + "angle": 0, + "content": "[46] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In Advances in Neural Information Processing Systems (NIPS), volume 30, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.761, + 0.892, + 0.815 + ], + "angle": 0, + "content": "[47] Xuebin Qin, Zichen Zhang, Chenyang Huang, Masood Dehghan, Osmar R Zaiane, and Martin Jagersand. U2-net: Going deeper with nested u-structure for salient object detection. Pattern Recognition, 106:107404, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.818, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[48] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10901-10911, 2021. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "812" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[49] Jiawei Ren, Liang Pan, and Ziwei Liu. Benchmarking and analyzing point cloud classification under corruptions. In Proceedings of the International Conference on Machine learning (ICML), 2022. 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.47, + 0.204 + ], + "angle": 0, + "content": "[50] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4104-4113, 2016. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.205, + 0.469, + 0.273 + ], + "angle": 0, + "content": "[51] Shuai Shao, Zeming Li, Tianyuan Zhang, Chao Peng, Gang Yu, Xiangyu Zhang, Jing Li, and Jian Sun. Objects365: A large-scale, high-quality dataset for object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8430-8439, 2019. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.274, + 0.469, + 0.328 + ], + "angle": 0, + "content": "[52] Edward J Smith and David Meger. Improved adversarial systems for 3d object generation and reconstruction. In Proceedings of the Conference on Robot Learning (CoRL), pages 87-96, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.33, + 0.469, + 0.397 + ], + "angle": 0, + "content": "[53] Stefan Stojanov, Anh Thai, and James M Rehg. Using shape to categorize: Low-shot learning with an explicit shape bias. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1798-1808, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.4, + 0.469, + 0.468 + ], + "angle": 0, + "content": "[54] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5459-5469, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.469, + 0.469, + 0.525 + ], + "angle": 0, + "content": "[55] Saeid Asgari Taghanaki, Jieliang Luo, Ran Zhang, Ye Wang, Pradeep Kumar Jayaraman, and Krishna Murthy Jatavallabhula. Robustpointset: A dataset for benchmarking robustness of point cloud classifiers. arXiv.org, 2011.11572, 2020.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.526, + 0.469, + 0.607 + ], + "angle": 0, + "content": "[56] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 1588–1597, 2019. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.609, + 0.469, + 0.69 + ], + "angle": 0, + "content": "[57] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T Barron, and Pratul P Srinivasan. Ref-nerf: Structured view-dependent appearance for neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5481-5490, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.692, + 0.469, + 0.773 + ], + "angle": 0, + "content": "[58] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. In Advances in Neural Information Processing Systems (NIPS), volume 34, pages 27171-27183, 2021. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.776, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[59] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P Srinivasan, Howard Zhou, Jonathan T Barron, Ricardo Martin-Brualla, Noah Snavely, and Thomas Funkhouser. Ibrnet: Learning multi-view image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4690-4699, 2021. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.873, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[60] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "graph cnn for learning on point clouds. ACM Transactions on Graphics, 38(5):1-12, 2019. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.178 + ], + "angle": 0, + "content": "[61] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing (TIP), 13(4):600-612, 2004. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.18, + 0.892, + 0.247 + ], + "angle": 0, + "content": "[62] Jiajun Wu, Chengkai Zhang, Tianfan Xue, Bill Freeman, and Josh Tenenbaum. Learning a probabilistic latent space of object shapes via 3d generative-adversarial modeling. In Advances in Neural Information Processing Systems (NIPS), volume 29, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.25, + 0.892, + 0.305 + ], + "angle": 0, + "content": "[63] Tong Wu, Jiaqi Wang, Xingang Pan, Xudong Xu, Christian Theobalt, Ziwei Liu, and Dahua Lin. Voxurf: Voxel-based efficient and accurate neural surface reconstruction. arXiv.org, 2208.12697, 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.308, + 0.892, + 0.389 + ], + "angle": 0, + "content": "[64] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1912-1920, 2015. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.392, + 0.892, + 0.46 + ], + "angle": 0, + "content": "[65] Tiange Xiang, Chaoyi Zhang, Yang Song, Jianhui Yu, and Weidong Cai. Walk in the cloud: Learning curves for point clouds shape analysis. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 915–924, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.463, + 0.892, + 0.532 + ], + "angle": 0, + "content": "[66] Mutian Xu, Runyu Ding, Hengshuang Zhao, and Xiaojuan Qi. Paconv: Position adaptive convolution with dynamic kernel assembling on point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3173-3182, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.534, + 0.892, + 0.615 + ], + "angle": 0, + "content": "[67] Mutian Xu, Junhao Zhang, Zhipeng Zhou, Mingye Xu, Xiaojuan Qi, and Yu Qiao. Learning geometry-disentangled representation for complementary understanding of 3d object point cloud. In Proceedings of the Conference on Artificial Intelligence (AAAI), volume 35, pages 3056-3064, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.619, + 0.892, + 0.688 + ], + "angle": 0, + "content": "[68] Guandao Yang, Xun Huang, Zekun Hao, Ming-Yu Liu, Serge Belongie, and Bharath Hariharan. Pointflow: 3d point cloud generation with continuous normalizing flows. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4541-4550, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.69, + 0.892, + 0.771 + ], + "angle": 0, + "content": "[69] Yao Yao, Zixin Luo, Shiwei Li, Jingyang Zhang, Yufan Ren, Lei Zhou, Tian Fang, and Long Quan. Blendedmvs: A large-scale dataset for generalized multi-view stereo networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1790-1799, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.774, + 0.892, + 0.83 + ], + "angle": 0, + "content": "[70] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. In Advances in Neural Information Processing Systems (NIPS), volume 34, pages 4805-4815, 2021. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.832, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[71] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4578-4587, 2021. 3, 5, 7" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "813" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.16 + ], + "angle": 0, + "content": "[72] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. In Advances in Neural Information Processing Systems (NIPS), 2022. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.472, + 0.233 + ], + "angle": 0, + "content": "[73] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 586-595, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.234, + 0.472, + 0.289 + ], + "angle": 0, + "content": "[74] Linqi Zhou, Yilun Du, and Jiajun Wu. 3d shape generation and completion through point-voxel diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5826-5835, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.29, + 0.472, + 0.332 + ], + "angle": 0, + "content": "[75] Qian-Yi Zhou, Jaesik Park, and Vladlen Koltun. Open3D: A modern library for 3D data processing. arXiv.org, 1801.09847, 2018. 4" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.956 + ], + "angle": 0, + "content": "814" + } + ] +] \ No newline at end of file diff --git a/2023/OmniObject3D_ Large-Vocabulary 3D Object Dataset for Realistic Perception, Reconstruction and Generation/243023a1-e64c-4bdc-a9c5-c30584f803ba_origin.pdf b/2023/OmniObject3D_ Large-Vocabulary 3D Object Dataset for Realistic Perception, Reconstruction and Generation/243023a1-e64c-4bdc-a9c5-c30584f803ba_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..7d7ff1f394e8879edbd6b2e15b61904edbcde9bb --- /dev/null +++ b/2023/OmniObject3D_ Large-Vocabulary 3D Object Dataset for Realistic Perception, Reconstruction and Generation/243023a1-e64c-4bdc-a9c5-c30584f803ba_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9339b1def91d94f86b2220a6d9f524d72944b7c0b74365ac2d813bf53e6bf2f6 +size 5280747 diff --git a/2023/OmniObject3D_ Large-Vocabulary 3D Object Dataset for Realistic Perception, Reconstruction and Generation/full.md b/2023/OmniObject3D_ Large-Vocabulary 3D Object Dataset for Realistic Perception, Reconstruction and Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..66914526303b9a1dd343d56aa142018c62847fc6 --- /dev/null +++ b/2023/OmniObject3D_ Large-Vocabulary 3D Object Dataset for Realistic Perception, Reconstruction and Generation/full.md @@ -0,0 +1,288 @@ +# OmniObject3D: Large-Vocabulary 3D Object Dataset for Realistic Perception, Reconstruction and Generation + +Tong Wu $^{1,2}$ , Jiarui Zhang $^{1,3}$ , Xiao Fu $^{1}$ , Yuxin Wang $^{1,4}$ , Jiawei Ren $^{5}$ , Liang Pan $^{5}$ , Wayne Wu $^{1}$ , Lei Yang $^{1,3}$ , Jiaqi Wang $^{1}$ , Chen Qian $^{1}$ , Dahua Lin $^{1,2\boxtimes}$ , Ziwei Liu $^{5\boxtimes}$ + +$^{1}$ Shanghai Artificial Intelligence Laboratory, $^{2}$ The Chinese University of Hong Kong, $^{3}$ SenseTime Research, $^{4}$ Hong Kong University of Science and Technology, $^{5}$ S-Lab, Nanyang Technological University + +![](images/d056abb644ec104bb811689b4c3f1b8fab84ffe54f420c3ff8e80a763e512e92.jpg) + +![](images/91a4275a8b6f75345ad6b9a173b3caa5c79231e9ed119771bfea692e71cfc2dc.jpg) +Textured meshes + +![](images/d78849eede2e862c95f04572aa4c1495954206ce24c266472b216583f5820702.jpg) +Point clouds Rendered images + +![](images/9777ee296f0a8c41cf97ec290c7efdd6b42c764d2eed5c1ee91223b4f5bce0de.jpg) +Real-captured videos + +![](images/346d9e3b169997728a6217be4005801a4bb2a6c29de013bcc2e65406c124c1a4.jpg) +Figure 1. OmniObject3D is a large vocabulary 3D object dataset with massive high-quality real-scanned 3D objects and rich annotations. It supports various research topics, e.g., perception, novel view synthesis, neural surface reconstruction, and 3D generation. + +# Abstract + +Recent advances in modeling 3D objects mostly rely on synthetic datasets due to the lack of large-scale real-scanned 3D databases. To facilitate the development of 3D perception, reconstruction, and generation in the real world, we propose OmniObject3D, a large vocabulary 3D object dataset with massive high-quality real-scanned 3D objects. OmniObject3D has several appealing properties: 1) Large Vocabulary: It comprises 6,000 scanned objects in 190 daily categories, sharing common classes with popular 2D datasets (e.g., ImageNet and LVIS), benefiting the pursuit of generalizable 3D representations. 2) Rich Annotations: Each 3D object is captured with both 2D and + +3D sensors, providing textured meshes, point clouds, multiview rendered images, and multiple real-captured videos. 3) Realistic Scans: The professional scanners support high-quality object scans with precise shapes and realistic appearances. With the vast exploration space offered by OmniObject3D, we carefully set up four evaluation tracks: a) robust 3D perception, b) novel-view synthesis, c) neural surface reconstruction, and d) 3D object generation. Extensive studies are performed on these four benchmarks, revealing new observations, challenges, and opportunities for future research in realistic 3D vision. + +# 1. Introduction + +Sensing, understanding, and synthesizing realistic 3D objects is a long-standing problem in computer vision, with + +rapid progress emerging in recent years. However, a majority of the technical approaches rely on unrealistic synthetic datasets [6, 19, 64] due to the absence of a large-scale real-world 3D object database. However, the appearance and distribution gaps between synthetic and real data cannot be compensated for trivially, hindering their real-life applications. Therefore, it is imperative to equip the community with a large-scale and high-quality 3D object dataset from the real world, which can facilitate a variety of 3D vision tasks and downstream applications. + +Recent advances partially fulfill the requirements while still being unsatisfactory. As shown in Table 1, CO3D [48] contains 19k videos capturing objects from 50 MS-COCO categories, while only $20\%$ of the videos are annotated with accurate point clouds reconstructed by COLMAP [50]. Moreover, they do not provide textured meshes. GSO [16] has 1k scanned objects while covering only 17 household classes. AKB-48 [33] focuses on robotics manipulation with 2k articulated object scans in 48 categories, but the focus on articulation leads to a relatively narrow semantic distribution, failing to support general 3D object research. + +To boost the research on general 3D object understanding and modeling, we present OmniObject3D: a large-vocabulary 3D object dataset with massive high-quality, real-scanned 3D objects. Our dataset has several appealing properties: 1) Large Vocabulary: It contains 6,000 high-quality textured meshes scanned from real-world objects, which, to the best of our knowledge, is the largest among real-world 3D object datasets with accurate 3D meshes. It comprises 190 daily categories, sharing common classes with popular 2D and 3D datasets (e.g., ImageNet [15], LVIS [25], and ShapeNet [6]), incorporating most daily object realms (See Figure 1 and Figure 2). 2) Rich Annotations: Each 3D object is captured with both 2D and 3D sensors, providing textured 3D meshes, sampled point clouds, posed multi-view images rendered by Blender [13], and real-captured video frames with foreground masks and COLMAP camera poses. 3) Realistic Scans: The object scans are of high fidelity thanks to the professional scanners, bearing precise shapes with geometric details and realistic appearance with high-frequency textures. + +Taking advantage of the vast exploration space offered by OmniObject3D, we carefully set up four evaluation tracks: a) robust 3D perception, b) novel-view synthesis, c) neural surface reconstruction, and d) 3D object generation. Extensive studies are performed on these benchmarks: First, the high-quality, real-world point clouds in OmniObject3D allow us to perform robust 3D perception analysis on both out-of-distribution (OOD) styles and corruptions, two major challenges in point cloud OOD generalization. Furthermore, we provide massive 3D models with multi-view images and precise 3D meshes for novel-view synthesis and neural surface reconstruction. The broad diver + +Table 1. A comparison between OmniObject3D and other commonly-used 3D object datasets. $R^{LVis}$ denotes the ratio of the 1.2k LVIS [25] categories being covered. + +
DatasetRealFull MeshVideo# Objs# CatsR1vis (%)
ShapeNet [6]✓51k554.1
ModelNet [64]✓12k402.4
3D-Future [19]✓16k341.3
ABO [12]✓8k633.5
Toys4K [53]✓4k1057.7
CO3D V1 / V2 [48]✓✓19 / 40k504.2
DTU [1]✓✓124NA0
ScanObjectNN [56]✓15k151.3
GSO [16]✓✓1k170.9
AKB-48 [33]✓✓2k481.8
Ours✓✓✓6k19010.8
+ +sity in shapes and textures offers a comprehensive training and evaluation source for both scene-specific and generalizable algorithms. Finally, we equip the community with a database for large vocabulary and realistic 3D object generation, which pushes the boundary of existing state-of-the-art generation methods to real-world 3D objects. The four benchmarks reveal new observations, challenges, and opportunities for future research in realistic 3D vision. + +# 2. Related Works + +3D Object Datasets. The acquisition of a large-scale realistic 3D database is usually expensive and challenging. Many widely-used 3D datasets prefer to collect synthetic CAD models from online repositories [6, 53, 64], for example, ShapeNet [6] has 51,300 3D CAD models in 55 categories, and ModelNet40 [64] consists of 12,311 models in 40 categories. Recent works, e.g., 3D-FUTURE [19] and ABO [12], introduce high-quality CAD models with rich geometric details and informative textures. However, due to the inevitable gap between synthetic and real objects, the community is still eager for a large-scale 3D object dataset in the real world. DTU [1] and BlendedMVS [69] are photo-realistic datasets designed for multi-view stereo benchmarks, while they are small in scale and lack category annotations. ScanObjectNN [56] is a real-world point cloud object dataset based on scanned indoor scenes, containing around 15,000 objects with colored point clouds in 15 categories. However, the point clouds are incomplete and noisy, and multiple objects usually co-exist in one scene. GSO [16] has 1,030 scanned objects with fine geometries and textures in 17 household items, and AKB-48 [33] focuses on robotics manipulation with 2,037 articulated object models in 48 articulated object categories. However, the relatively narrow semantic scope of GSO and AKB-48 hinders their applications for more general 3D research. CO3D v1 [48] contains 19,000 object-centric videos, while only $20\%$ of them are annotated with accurate point clouds reconstructed by COLMAP [50], and they do not provide meshes or textures. In contrast, the proposed OmniOb + +![](images/a6195fa50c5f41afcc508a8931ff1a9b4ddf8bff89866401733e9bbf6c81687d.jpg) +Figure 2. Semantic distribution of the OmniObject3D dataset. It covers 190 daily categories with a long-tailed distribution, sharing common classes with popular 2D and 3D datasets. + +![](images/e19fc1e9e1bdc9b6e89a005fa17ee1b5f55474ecc4e6b2a1c1505ed18e97623d.jpg) + +project3D dataset comprises 6,000 3D objects scanned by professional devices with meshes, textures, and multi-view photos in 190 categories, fulfilling the requirements of a wide range of research objectives. A detailed comparison is presented in Table 1. + +Robust 3D Perception. Robustness to out-of-distribution (OOD) data is important in point cloud perception. Two main challenges include OOD styles (e.g., differences between CAD models and real-world objects) and OOD corruptions (e.g., random point jittering or missing due to sensory inaccuracy). A branch of works [10, 30, 45, 60] studies the OOD corruptions and proposes standard corruption test suites [49, 55], while they fail to take account of OOD styles. Another branch of works [3, 48] evaluates the sim-to-real domain gap by training models on clean synthetic datasets [64] and testing them on noisy real-world test sets [56], while OOD styles and corruptions cannot be disentangled under this setting for an independent analysis. In this work, we leverage high-quality, real-world point clouds from OmniObject3D to systematically measure the robustness against the OOD style and OOD corruptions, providing the first benchmark for fine-grained evaluation of the point cloud perception robustness. + +Neural Radiance Field and Neural Surface Reconstruction. Neural radiance field (NeRF) [40] represents a scene with a fully-connected deep network (MLPs), which takes in hundreds of sampled points along each camera ray and outputs the predicted color and density. We can synthesize the image of an unseen view from a trained model via volume rendering. Inspired by the success of NeRF, massive follow-up efforts have been made to improve its quality [4, 5, 39, 57] and efficiency [7, 18, 42, 54]. A branch of works [8, 27, 35, 48, 59, 71] has also explored the generalization ability of NeRF-based frameworks, where they aim to learn priors from deep image features across multiple scenes. Beyond novel view synthesis, another trend of approaches [14, 43, 58, 63, 70] proposes to combine neural radiance field with implicit surface representations like Signed Distance Function (SDF), and they achieve accurate and mask-free surface reconstruction from multi-view im + +ages. Since dense camera views of scenes are sometimes unavailable, recent advances explore surface reconstruction from sparse views. They address the problem by exploiting generalizable priors cross scenes for a generic surface prediction [36] or taking advantage of the estimated geometry cues estimated by pre-trained networks [72]. OmniObject3D can serve as a large-scale benchmark with realistic photos and accurate meshes for both training and evaluation. The high diversity in shape and appearance offers an opportunity for pursuing more generalizable and robust novel view synthesis and surface reconstruction methods. + +3D Object Generation. Early approaches [20,26,37,52,62] extend 2D generation frameworks to 3D voxels with a high computational cost. Some others adopt different 3D data formulations, e.g., point cloud [2,41,68,74], octree [29], and implicit representations [11,38]. However, it is nontrivial to generate complex and textured surfaces. Recent advances [9,21,44] explore the generation for textured 3D meshes, where GET3D [22] is a state-of-the-art approach that generates diverse meshes with rich geometry and textures in two branches. It is a promising but challenging task to train generative models on a large vocabulary and realistic dataset. We evaluate GET3D on our dataset and reveal several challenges and future opportunities. + +In supplementary materials, we present more detailed discussions on related works for different tracks. + +# 3. The OmniObject3D Dataset + +In this section, we describe the data collection, processing, and annotation pipeline of OmniObject3D. We also introduce the statistics and distribution of it. + +# 3.1. Data Collection, Processing, and Annotation + +Category List Definition. In order to collect a large amount of 3D objects that are both commonly-distributed and highly diverse, we first pre-define a category list according to several popular 2D and 3D datasets [6,15,25,31, 32,51,64]. We cover most of the categories that lie within the application scope of the scanners and also dynamically + +expand the list with reasonable new classes that are absent from the current list during collection. We end up with 190 widely-spread categories, which ensures a library with rich texture, geometry, and semantic information. + +Object Collection Pipeline. We then collect a variety of objects from each category and use professional 3D scanners to obtain high-resolution textured meshes. Specifically, we use the Shining 3D scanner1 and Artec Eva 3D scanner2 for objects in different scales. The scanning time varies with the properties of the object: it takes around 15 minutes to scan a small rigid object with a simple geometry (e.g., an apple, a toy), while it takes up to an hour to obtain a qualified 3D scan for non-rigid, complex, or large objects (e.g., a bed, a kite). For around $10\%$ of the objects, we conduct common manipulations (e.g., taking a bite, cutting in pieces) to conform the natural instincts of them. The 3D scans can faithfully retain the real-world scale of each object, but their poses are not strictly aligned. We thus pre-define a canonical pose for each category and manually align the objects within a category. We then check the quality of each scan, and around $83\%$ high-quality ones out of the total collection are finally reserved in the dataset. + +Image Rendering and Point Cloud Sampling. To support a variety of research topics like point cloud analysis, neural radiance fields, and 3D generation, we render multiview images and sample point clouds based on the collected 3D models. We use Blender [13] to render object-centric and photo-realistic multi-view images, together with accurate camera poses. The images are rendered from 100 random viewpoints sampled on the upper hemisphere at $800 \times 800$ pixels. We also produce high-resolution mid-level cues like depth and normal for more research use. We then uniformly sample multi-resolution point clouds from each 3D model using the Open3D toolbox [75], with $2^{n}$ ( $n \in \{10, 11, 12, 13, 14\}$ ) points in each point cloud, respectively. Besides the data existing in the dataset, we also provide a data generation pipeline. One can easily obtain new data with self-defined camera distributions, lighting, and point sampling methods to meet different requirements. Video Capturing and Annotation. After scanning each object, we capture its video with an iPhone 12 Pro mobile phone. The object is placed on or beside a calibration board, and each video covers a full $360^{\circ}$ range around it. Square corners on the calibration board can be recognized by the QR Codes beside it, and we then filter out the blurry frames whose recognized corners are less than 8. We uniformly sample 200 frames, and then COLMAP [50], a well-known SfM pipeline, is applied to annotate the frames with camera poses. Finally, we use the scales of the calibration board in both the SfM coordinate space and the real world to recover the absolute scale of the SfM coordinate system. + +![](images/d6b968714584f3496f44ecb711309889bc0e7447485a1fa46be3725784a42863.jpg) +Figure 3. OmniObject3D provides the first clean real-world point cloud object dataset and allows fine-grained analysis on robustness to OOD styles and OOD corruptions."-C": corrupted by common corruptions described in [49] + +We also develop a two-stage matting pipeline based on the U $^2$ Net [47] and FBA [17] matting model to generate the foreground masks for all the frames. Please refer to supplementary materials for more implementation details. + +# 3.2. Statistics and Distribution + +With 6,000 3D models in 190 categories, OmniObject3D exhibits a long-tailed distribution with an average of around 30 objects in each category, as shown in Figure 2. It shares many common categories with several famous 2D and 3D datasets [6, 15, 25, 31, 32, 51, 64]. For example, it covers 85 categories in ImageNet [15] and 130 categories in LVIS [25], which leads to the highest $R^{\mathrm{vis}}$ in Table 1. Most of the categories are covered by the Open Images [31] image-level labels. It bears a huge diversity in shapes and appearances. The vast semantic and geometrical exploration spaces enable a wide range of research objectives. + +# 4. Experiments + +# 4.1. Robust 3D Perception + +Object-level point cloud classification is one of the most fundamental tasks in point cloud perception. In this section, we show how OmniObject3D boosts robustness analysis of point cloud classification by disentangling the two critical out-of-distribution (OOD) challenges introduced in Sec. 2, i.e., OOD styles and OOD corruptions. + +Existing robustness evaluation utilizes clean synthetic datasets, e.g., ModelNet [64], for training and sets up two kinds of test sets for evaluation: + +1) Noisy real-world datasets, e.g., ScanObjectNN [56], which are cropped from real-world scenes. They are employed to measure the robustness of the sim-to-real domain gap. However, the gap couples both OOD styles and OOD corruptions simultaneously as the cropped point clouds are always noisy, making it impossible to analyze the two robustness challenges independently. +2) Corrupted synthetic datasets, e.g., ModelNet-C [49], which are artificially perturbed on top of clean synthetic datasets. The evaluation allows for detailed corruption analysis, but they do not reflect the robustness to OOD styles. + +None of the existing robustness benchmarks allows for analyzing the robustness to both OOD styles and OOD cor + +Table 2. Point cloud perception robustness analysis on OmniObject3D with different architecture designs. Models are trained on the ModelNet-40 dataset, with OAClean to be their overall accuracy on the standard ModelNet-40 test set. OAStyle on OmniObject3D evaluates the robustness to OOD styles. mCE on the corrupted OmniObject3D-C evaluates the robustness to OOD corruptions. Blue shadings indicate rankings. $\dagger$ : results on ModelNet-C [49]. Full results are presented in the supplementary materials. + +
mCE†↓OA Clean ↑OA Style ↑mCE ↓
DGCNN [60]1.0000.9260.4481.000
PointNet [45]1.4220.9070.4660.969
PointNet++ [46]1.0720.9300.4071.066
RSCNN [34]1.1300.9230.3931.076
SimpleView [23]1.0470.9390.4760.990
GDANet [67]0.8920.9340.4970.920
PAConv [66]1.1040.9360.4031.073
CurveNet [65]0.9270.9380.5000.929
PCT [24]0.9250.9300.4590.940
RPC [49]0.8630.9300.4720.936
+ +ruptions in fine granularity. OmniObject3D, on the other hand, as the first clean real-world point cloud object dataset, can help to address the issue. For models trained on ModelNet, we first evaluate their performance on OmniObject3D to examine OOD-style robustness. Then, we create OmniObject3D-C by corrupting OmniObject3D with common corruptions described in [49] to examine the OOD-corruption robustness. We show a complete robustness evaluation scheme in Figure 3. For evaluation metrics, we use the overall accuracy (OA) on OmniObject3D to measure the OOD-style robustness and use DGCNN normalized mCE [49] to measure the OOD-corruption robustness. + +We benchmark ten state-of-the-art point cloud classification models in Table 2. We observe that 1) performance on a clean test set has little correlation with OOD-style robustness. For example, SimpleView [23] achieves the best $\mathrm{OA}_{\mathrm{Clean}}$ but mediocre $\mathrm{OA}_{\mathrm{Style}}$ ; 2) Advanced point grouping, e.g., curve-based point grouping in CurveNet [65] and frequency-based point grouping in GDANet [67], are robust not only to OOD corruptions as pointed out in [49], but also to OOD styles; 3) OOD style + OOD corruption is a more challenging setting. In particular, RPC, the most robust architecture to OOD corruptions [49], shows inferior mCE. In summary, robust point cloud perception models against both OOD styles and OOD corruptions are still under-explored. Our dataset sheds new light on a comprehensive understanding of point cloud perception robustness. See more results in the supplementary materials. + +# 4.2. Novel View Synthesis + +In this section, we study several representative methods on OmniObject3D for novel view synthesis (NVS) in two settings: 1) training on a single scene with densely captured images, which is the standard setting for NeRF [40]; 2) learning priors across scenes from our dataset to explore + +Table 3. Single-scene novel view synthesis results. Three metrics and their standard deviation (SD) across the training set. + +
MethodPSNR (↑) / SDSSIM (↑) / SDLPIPS (↓) / SD
NeRF [40]34.01 / 3.460.953 / 0.0290.068 / 0.061
mip-NeRF [4]39.86 / 4.580.974 / 0.0130.084 / 0.048
Plenoxels [18]41.04 / 6.840.982 / 0.0310.030 / 0.031
+ +Table 4. Cross-scene novel view synthesis results on 10 categories. 'Cat.' and 'All*' denote training on each category and training on all categories except the 10 test ones, respectively. + +
MethodTrainPSNR (↑)SSIM (↑)LPIPS (↓)\( {\mathcal{L}}_{1}^{\text{depth }}\left( \downarrow \right) \)
MVSNeRF [8]All*17.490.5440.4420.193
Cat.17.540.5420.4480.230
All*-ft.25.700.7540.2510.081
Cat.-ft.25.520.7500.2640.076
IBRNet [59]All*19.390.5690.3990.423
Cat.19.030.5510.4150.290
All*-ft.26.890.7920.2150.081
Cat.-ft.25.670.7600.2380.099
pixelNeRF [71]All*22.160.6920.3420.109
Cat.20.650.6760.3480.195
+ +the generalization ability of NeRF-style models. + +Single-Scene NVs. We select three objects in each category for the experiments, randomly sampling 1/8 images as the hold-out test set. We involve NeRF [40], mip-NeRF [4], and a voxel-based system named Plenoxels [18] for comparison. As in Table 3, we find that Plenoxels achieve the best performance on average for PSNR, SSIM [61], and LPIPS [73]. There exists a clear margin for LPIPS between Plenoxels and the other two methods since voxel-based methods are especially good at modeling high-frequency appearance. We also present the standard deviation (SD) of results across all the training samples, where Plenoxels are relatively unstable compared to NeRF and mip-NeRF. We observe that Plenoxels introduce artifacts when encountering concave geometry (e.g., bowls, chairs) and suffer from an inaccurate density field modeling when the foreground object is dark. MLP-based methods are relatively robust against these difficult cases. In a nutshell, our dataset provides a library with a variety of shapes and appearances, allowing a comprehensive evaluation of different NVS methods. See the supplementary for more results with the iPhone videos, detailed analysis, and visualizations. + +Cross-Scene NVS. We conduct extensive experiments on novel view synthesis from sparse inputs by pixelNeRF [71], IBRNet [59] and MVSNeRF [8] in Table 4. We select 10 categories with the most various scenes as the test set. All metrics are averaged over 300 images. In the generalization setting, although not trained on the test category, $\mathrm{MVSNeRF_{All^{*}}}$ is comparable to $\mathrm{MVSNeRF_{Cat}}$ , $\mathrm{IBRNet_{All^{*}}}$ and $\mathrm{pixelNeRF_{All^{*}}}$ even outperforms the corresponding "Cat"s in all terms of visual metrics, especially on regular-shaped objects such as squash and apple. It confirms that OmniObject3D serves as an information-rich + +![](images/f271b8f08e6032bbe824a41c71e1411e752a0232f71bc4477990c945a3e4e4c6.jpg) + +Table 5. Dense-view surface reconstruction results. + +
MethodChamfer Distance ×103(↓)
HardMediumEasyAvg
NeuS [58]9.265.633.466.09
VolSDF [70]10.064.942.865.92
Voxurf [63]9.014.982.585.49
Avg9.445.192.975.83
+ +dataset that is beneficial for obtaining a strong generalizable prior on unseen scenes. Moreover, it is noteworthy that MVSNeRF and pixelNeRF with 'All* generate better underlying depth than those with 'Cat.', inferring generalizable methods can implicitly learn geometric cues though only trained from appearance in our dataset. It is reasonable that (1) IBRNet suffers more severely than the others in geometry under the scarcity of source context (only 3 views) as it is more suitable for dense-view generalization that complies with its view interpolation module. (2) MVSNeRF lags behind pixelNeRF on visual performance as we take 10 test frames widely distributed around the object in $360^{\circ}$ by FPS sampling algorithm, where cost volume will be inaccurate on large-range viewpoint change. After further finetuning IBRNet for only around 10 minutes on each test scene, $\mathrm{IBRNet}_{\mathrm{All}^{*}\text{-ft}}$ achieves the best view synthesis results, comparable to test-time optimized NVS methods on nearby views. It is promising to utilize the large-scale and category-prosperous OmniObject3D, to build a benchmark suite for evaluating diverse cross-scene NVS methods. + +# 4.3. Neural Surface Reconstruction + +Precise surface reconstruction from multi-view images enables a broad range of applications. For a single scene with dense-view images, algorithms are expected to conduct accurate, robust, and efficient surface reconstruction. When only sparse-view images are available, it is crucial to learn generalizable priors from a set of scenes or use other + +![](images/dfb3dafcdbf239dada19f6ddc3709d7fd1165f821c9539386b447cb9ea72e6f2.jpg) +Figure 4. Neural surface reconstruction results for both dense-view and sparse-view settings. +Figure 5. Performance distribution of dense-view surface reconstruction. The averaged results of the three methods is imbalanced. The colored area denotes a smoothed range of results. + +geometric cues to assist reconstruction. Accordingly, we study the two settings for surface reconstruction methods. + +Dense-View Surface Reconstruction. Under this setting, we include three representative methods. NeuS [58] and VolSDF [70] are two well-known approaches that bridge neural volume rendering with implicit surface representation. We also involve a voxel-based method called Voxsurf [63], which leverages a hybrid representation to achieve acceleration and fine geometry reconstruction. + +Previous approaches in this task mainly perform evaluations and comparisons on 15 scenes from the DTU [1] dataset, which is not comprehensive and robust enough to demonstrate the ability of the methods in different scenarios. In comparison, we select three objects per category to run each of the three methods above, leading to over 1,500 reconstructions in total. We calculate the Chamfer Distance (CD) between the reconstructed surface and the ground truth. The distribution of the results is shown in Figure 5. The average curve is imbalanced: hard categories usually include low-textured, concave, or complex shapes (e.g., bowls, vases, kennels, cabinets, and durians). We thus split the categories into three levels of "difficulty" based on the average curve, and the level-wise results are presented in Table 5. We can observe a clear margin among results in + +Table 6. Sparse-view (3-view) surface reconstruction results. + +
MethodTrainChamfer Distance ×103(↓)
HardMediumEasyAvg
NeuS [58]Single29.3527.6224.7927.3
MonoSDF [72]Single35.1435.3532.7634.68
SparseNeuS [36]1 cat.34.0531.3231.1432.36
10 cats.30.7530.1128.3729.87
All cats.26.1326.0822.1325.00
Easy28.3926.6523.7626.48
Medium27.3826.6623.0825.87
Hard27.4226.9524.6326.47
MVSNeRF [8]All cats.56.6848.0948.7051.16
pixelNeRF [71]All cats.63.3159.9161.4761.56
+ +different levels for each method, indicating the split subsets to be generic and faithful. + +Sparse-View Surface Reconstruction. Dense-captured images of a scene are sometimes not available, so we also study the sparse-view scenario here. The following methods are included: NeuS [58] with sparse-view input; MonoSDF [72], which takes in geometry cues estimated by pre-trained models; SparseNeuS [36], a generic surface prediction pipeline that learns generalizable priors; pixelNeRF [71] and MVSNeRF [8] from Sec. 4.2, whose geometries are extracted from the density field. For NeuS, MonoSDF, and pixelNeRF, we use Farthest Point Sampling (FPS) to sample views that are most widely distributed; for SparseNeuS and MVSNeRF, we conduct FPS among the nearest 30 camera poses from a random reference view. We sample 3 views in all the experiments. + +The quantitative and qualitative comparisons are shown in Table 6 and Figure 4, respectively. We observe apparent artifacts in all the sparse-view reconstructed results. Among them, SparseNeuS trained on enough data demonstrates the best quantitative performance on average, and the pre-division on the train set does not result in a noticeable difference across difficulty levels. NeuS with sparse-view input achieves a surprisingly good performance. As shown in Figure 4, the FPS sampling equips NeuS with a coherent global shape for thin structures like case 3, while it encounters severe local geometry ambiguity like case 1. MonoSDF, on the contrary, partially overcomes the issue of ambiguity via the assistance of predicted geometry cues in case 1. However, it relies heavily on the accuracy of the estimated depth and normal and thus easily fails when the estimation is inaccurate (e.g., case 2 and 3). The surfaces extracted from generalized NeRF models, i.e., pixelNeRF and MVSNeRF, are of relatively low quality. + +In brief, the challenging problem of sparse-view surface reconstruction has not been solved well currently. OmniObject3D is a promising database to study generalizable surface reconstruction pipelines as well as strategies for a robust usage of estimated geometry cues for this track. + +![](images/7a4e5ca54b4d4298c303b4927bb22ea6dc5f3d941993e1c6fa6e38e2bd816e25.jpg) + +![](images/8689dca28d26bf5462052488e2fff21143b2b450c96ec02d09ccc45eed566d5a.jpg) + +![](images/a66375ac3ccd9ce8f2115cc07b7f7488a65f96759c7bea379ee2050c5c8b185c.jpg) +Figure 6. The category distribution of the generated shapes. (a) shows a weak positive correlation between the number of generated shapes and training shapes per category. (b) visualizes the correlation matrix among different categories by Chamfer Distance between their mean shapes. (c) visualizes categories being clustered into eight groups by KMeans. (d) presents a clear training-generation relation in the group-level statistics. + +![](images/6e0ed030abc642b71c9765a4883a92e0c58a8df4e43565e37b2d59ac90178d43.jpg) + +# 4.4.3D Object Generation + +In this section, we adopt a state-of-the-art generative model that directly generates explicit textured 3D meshes, namely GET3D [22]. GET3D is originally evaluated on six categories (Car, Chair, Motorbike, Animal, House, and Human Body) with independent models trained on each category. The number of shapes per category ranges from 337 to 7,497. In comparison, OmniObject3D contains many more categories with fewer objects in each. As a result, it is natural to train each model with multiple categories. + +We first provide some qualitative results in Figure 7, where we show various generation results: The textures are rather realistic, and the shapes are coherent, enhanced by fine geometry details (e.g., the lychee and pineapple). We explore the latent space of the model and show interpolation results in Figure 8. We can observe a smooth transition across instances that are semantically different. We would like to further analyze the performance of the generative model trained on OmniObject3D from three aspects, i.e., semantic distribution, diversity, and quality. + +Semantic Distribution. We randomly select 100 categories to train an unconditional model jointly. We randomly generate 1,000 textured meshes at inference time and ask human experts to label them. Shapes with ambiguity are not counted. Figure 6 (a) shows that the generated shapes per category are highly imbalanced, exhibiting a weak positive + +![](images/053c21d8c495415c591922523dff2f9a8c2fc954523d0bc10294d3739734be65.jpg) + +![](images/43af7c99aaee2740bf7d2ce347fecb8fca0dadfd3a6891a83c2088f299f4d5af.jpg) +Figure 7. Examples of the generated textured shapes rendered in Blender. OmniObject3D enables GET3D with realistic generations across a wide range of categories. +Figure 8. Shape interpolation. We interpolate both geometry and texture latent codes from left to right. + +correlation with the training shape numbers. Actually, the categories are not independent but rather highly correlated. We calculate the "mean shape" for each category and visualize the Chamfer Distance among them in Figure 6 (b), which indicates that they can be further grouped. Regarding each matrix row as a feature vector, we use KMeans to cluster them into eight groups (Figure 6 (c)) and carry out group-level statistics in Figure 6 (d). It demonstrates a clear trend that the number of generated shapes increases along with or even faster than the number of training shapes in the group, revealing an enlarged semantic bias during generation. However, it also depends on the inner-group divergence. For example, Group 2 (883 shapes in 27 categories) has the largest number of training samples, while the high variation among its categories prevents it from dominating the generated shapes; Group 1 (587 shapes in 18 categories) has a relatively small divergence, which becomes the most popular in the generated shapes. We present more details in the supplementary material. + +Diversity and Quality. We select four representative data subsets for training and evaluation, namely fruits, furniture, toys, and Rand-100. We randomly split each subset into training (80%) and testing (20%). We leverage three evaluation metrics: for geometry, we use Chamfer Distance (CD) to compute the Coverage score (Cov) and Minimum Matching Distance (MMD), which focus on the diversity and quality of the shapes, respectively; for texture, we adopt the widely-used FID [28]. But the FID metrics with different test splits are not directly comparable, suffering from a large variance when the test set is small. We thus introduce FIDref for reference, which is the FID between the train and test set. The results are shown in Table 7. Furniture suffers from the lowest quality (MMD) since the small train set with 17 categories is a difficult training source. Fruits has the same number of categories while being 2.3 times + +Table 7. Quantitative evaluations on different data splits. + +
Split#Objs#CatsCov (% ↑)MMD (↓)FID (↓)FIDref
Furniture2651767.924.2787.3958.40
Fruits6101746.723.32105.3187.15
Toys339755.222.78122.7741.40
Rand-100295110061.703.8946.578.65
+ +larger in scale, and some fruits share a very similar structure, leading to relatively higher quality and lower diversity (Cov). Toys achieve the best quality by training on only 7 categories. Rand-100 is the most difficult case, and we can observe a trade-off between quality and diversity. Both the FID and $\mathrm{FID}^{\mathrm{ref}}$ are high for the first three subsets due to the small testing sets, while only 'Rand-100' is relatively low. + +In brief, training and evaluating generative models on a large-vocabulary and realistic dataset is a promising but challenging task. We reveal crucial problems like the semantic distribution bias and varying exploration difficulties in different groups. OmniObject3D serves as a great database for further examination in this area. + +# 5. Conclusion and Outlook + +We introduce OmniObject3D, a large vocabulary 3D object dataset with massive high-quality real-scanned 3D objects, including 6,000 objects from 190 categories. It provides rich annotations, including textured 3D meshes, sampled point clouds, posed multi-view images rendered by Blender, and real-captured video frames with foreground masks and COLMAP camera poses. We set up four evaluation tracks, revealing new observations, challenges, and opportunities for future research in realistic 3D vision. + +We will regulate the usage of our data to avoid potential negative social impacts. + +Acknowledgement. This project is funded by Shanghai AI Laboratory, CUHK Interdisciplinary AI Research Institute, the Centre for Perceptual and Interactive Intelligence (CPII) Ltd under the Innovation and Technology Commission (ITC)'s InnoHK, Hong Kong RGC Theme-based Research Scheme 2020/21 (No. T41-603/20-R), OpenXDLab, the Ministry of Education, Singapore, under its MOE AcRF Tier 2 (MOE-T2EP20221-0012), NTU NAP, and under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative. + +# References + +[1] Henrik Aanaes, Rasmus Ramsbøl Jensen, George Vogiatzis, Engin Tola, and Anders Bjorholm Dahl. Large-scale data for multiple-view stereopsis. International Journal of Computer Vision (IJCV), 120(2):153-168, 2016. 2, 6 +[2] Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas Guibas. Learning representations and generative models for 3d point clouds. In Proceedings of the International Conference on Machine learning (ICML), pages 40-49, 2018. 3 +[3] Adel Ahmadyan, Liangkai Zhang, Artsiom Ablavatski, Jianing Wei, and Matthias Grundmann. Objectron: A large scale dataset of object-centric videos in the wild with pose annotations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7822-7831, 2021. 3 +[4] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5855–5864, 2021. 3, 5 +[5] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5470-5479, 2022. 3 +[6] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv.org, 1512.03012, 2015. 2, 3, 4 +[7] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 3 +[8] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14124-14133, 2021. 3, 5, 7 +[9] Wenzheng Chen, Huan Ling, Jun Gao, Edward Smith, Jaakko Lehtinen, Alec Jacobson, and Sanja Fidler. Learning to predict 3d objects with an interpolation-based differentiable renderer. In Advances in Neural Information Processing Systems (NIPS), volume 32, 2019. 3 +[10] Yunlu Chen, Vincent Tao Hu, Efstratios Gavves, Thomas Mensink, Pascal Mettes, Pengwan Yang, and Cees GM Snoek. Pointmixup: Augmentation for point clouds. In Proceedings of the European Conference on Computer Vision (ECCV), pages 330-345, 2020. 3 +[11] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5939-5948, 2019. 3 + +[12] Jasmine Collins, Shubham Goel, Kenan Deng, Achleshwar Luthra, Leon Xu, Erhan Gundogdu, Xi Zhang, Tomas F Yago Vicente, Thomas Dideriksien, Himanshu Arora, et al. Abo: Dataset and benchmarks for real-world 3d object understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21126-21136, 2022. 2 +[13] Blender Online Community. Blender - a 3d modelling and rendering package. 2018. 2, 4 +[14] François Darmon, Bénédicte Bascle, Jean-Clément Devaux, Pascal Monasse, and Mathieu Aubry. Improving neural implicit surfaces geometry with patch warping. arXiv.org, 2112.09648, 2021. 3 +[15] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 248–255, 2009. 2, 3, 4 +[16] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas B McHugh, and Vincent Vanhoucke. Google scanned objects: A high-quality dataset of 3d scanned household items. arXiv.org, 2204.11918, 2022. 2 +[17] Marco Forte and François Pitié. F, b, alpha matting. arXiv.org, 2003.07711, 2020. 4 +[18] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5501–5510, 2022. 3, 5 +[19] Huan Fu, Rongfei Jia, Lin Gao, Mingming Gong, Binqiang Zhao, Steve Maybank, and Dacheng Tao. 3d-future: 3d furniture shape with texture. International Journal of Computer Vision (IJCV), 129(12):3313-3337, 2021. 2 +[20] Matheus Gadelha, Subhransu Maji, and Rui Wang. 3d shape induction from 2d views of multiple objects. In Proceedings of the International Conference on 3D Vision (3DV), pages 402-411, 2017. 3 +[21] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. arXiv.org, 2209.11163, 2022.3 +[22] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. In Advances in Neural Information Processing Systems (NIPS), 2022. 3, 7 +[23] Ankit Goyal, Hei Law, Bowei Liu, Alejandro Newell, and Jia Deng. Revisiting point cloud shape classification with a simple and effective baseline. In Proceedings of the International Conference on Machine learning (ICML), pages 3809-3820, 2021. 5 +[24] Meng-Hao Guo, Jun-Xiong Cai, Zheng-Ning Liu, Tai-Jiang Mu, Ralph R Martin, and Shi-Min Hu. Pct: Point cloud transformer. Computational Visual Media, 7(2):187-199, 2021. 5 + +[25] Agrim Gupta, Piotr Dollar, and Ross Girshick. LVIS: A dataset for large vocabulary instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5356-5364, 2019. 2, 3, 4 +[26] Philipp Henzler, Niloy J Mitra, and Tobias Ritschel. Escaping plato's cave: 3d shape from adversarial rendering. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9984-9993, 2019. 3 +[27] Philipp Henzler, Jeremy Reizenstein, Patrick Labatut, Roman Shapovalov, Tobias Ritschel, Andrea Vedaldi, and David Novotny. Unsupervised learning of 3d object categories from videos in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4700-4709, 2021. 3 +[28] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In Advances in Neural Information Processing Systems (NIPS), volume 30, 2017. 8 +[29] Moritz Ibing, Gregor Kobsik, and Leif Kobbelt. Octree transformer: Autoregressive 3d shape generation on hierarchically structured sequences. arXiv.org, 2111.12480, 2022. 3 +[30] Sihyeon Kim, Sanghyeok Lee, Dasol Hwang, Jaewon Lee, Seong Jae Hwang, and Hyunwoo J Kim. Point cloud augmentation with weighted local transformations. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 548-557, 2021. 3 +[31] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4. International Journal of Computer Vision (IJCV), 128(7):1956-1981, 2020. 3, 4 +[32] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Proceedings of the European Conference on Computer Vision (ECCV), pages 740-755, 2014. 3, 4 +[33] Liu Liu, Wenqiang Xu, Haoyuan Fu, Sucheng Qian, Qiao-jun Yu, Yang Han, and Cewu Lu. Akb-48: A real-world articulated object knowledge base. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14809-14818, 2022. 2 +[34] Yongcheng Liu, Bin Fan, Shiming Xiang, and Chunhong Pan. Relation-shape convolutional neural network for point cloud analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8895-8904, 2019. 5 +[35] Yuan Liu, Sida Peng, Lingjie Liu, Qianqian Wang, Peng Wang, Christian Theobalt, Xiaowei Zhou, and Wenping Wang. Neural rays for occlusion-aware image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7824-7833, 2022. 3 +[36] Xiaoxiao Long, Cheng Lin, Peng Wang, Taku Komura, and Wenping Wang. Sparseneus: Fast generalizable neural surface reconstruction from sparse views. arXiv.org, 2206.05737, 2022. 3, 7 + +[37] Sebastian Lunz, Yingzhen Li, Andrew Fitzgibbon, and Nate Kushman. Inverse graphics gan: Learning to generate 3d shapes from unstructured 2d data. arXiv.org, 2002.12674, 2020.3 +[38] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4460-4470, 2019. 3 +[39] Ben Mildenhall, Peter Hedman, Ricardo Martin-Brualla, Pratul P Srinivasan, and Jonathan T Barron. Nerf in the dark: High dynamic range view synthesis from noisy raw images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16190-16199, 2022. 3 +[40] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In Proceedings of the European Conference on Computer Vision (ECCV), pages 405-421, 2020. 3, 5 +[41] Kaichun Mo, Paul Guerrero, Li Yi, Hao Su, Peter Wonka, Niloy J. Mitra, and Leonidas J. Guibas. Structurenet: Hierarchical graph networks for 3d shape generation. arXiv.org, 1908.00575, 2019. 3 +[42] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics, 2022. 3 +[43] Michael Oechsle, Songyou Peng, and Andreas Geiger. Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5589-5599, 2021. 3 +[44] Dario Pavllo, Jonas Kohler, Thomas Hofmann, and Aurelien Lucchi. Learning generative models of textured 3d meshes from real-world images. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 13879-13889, 2021. 3 +[45] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: deep learning on point sets for 3d classification and segmentation. corr abs/1612.00593 (2016). In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 652-660, 2017. 3, 5 +[46] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In Advances in Neural Information Processing Systems (NIPS), volume 30, 2017. 5 +[47] Xuebin Qin, Zichen Zhang, Chenyang Huang, Masood Dehghan, Osmar R Zaiane, and Martin Jagersand. U2-net: Going deeper with nested u-structure for salient object detection. Pattern Recognition, 106:107404, 2020. 4 +[48] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10901-10911, 2021. 2, 3 + +[49] Jiawei Ren, Liang Pan, and Ziwei Liu. Benchmarking and analyzing point cloud classification under corruptions. In Proceedings of the International Conference on Machine learning (ICML), 2022. 3, 4, 5 +[50] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4104-4113, 2016. 2, 4 +[51] Shuai Shao, Zeming Li, Tianyuan Zhang, Chao Peng, Gang Yu, Xiangyu Zhang, Jing Li, and Jian Sun. Objects365: A large-scale, high-quality dataset for object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8430-8439, 2019. 3, 4 +[52] Edward J Smith and David Meger. Improved adversarial systems for 3d object generation and reconstruction. In Proceedings of the Conference on Robot Learning (CoRL), pages 87-96, 2017. 3 +[53] Stefan Stojanov, Anh Thai, and James M Rehg. Using shape to categorize: Low-shot learning with an explicit shape bias. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1798-1808, 2021. 2 +[54] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5459-5469, 2022. 3 +[55] Saeid Asgari Taghanaki, Jieliang Luo, Ran Zhang, Ye Wang, Pradeep Kumar Jayaraman, and Krishna Murthy Jatavallabhula. Robustpointset: A dataset for benchmarking robustness of point cloud classifiers. arXiv.org, 2011.11572, 2020.3 +[56] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 1588–1597, 2019. 2, 3, 4 +[57] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T Barron, and Pratul P Srinivasan. Ref-nerf: Structured view-dependent appearance for neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5481-5490, 2022. 3 +[58] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. In Advances in Neural Information Processing Systems (NIPS), volume 34, pages 27171-27183, 2021. 3, 6, 7 +[59] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P Srinivasan, Howard Zhou, Jonathan T Barron, Ricardo Martin-Brualla, Noah Snavely, and Thomas Funkhouser. Ibrnet: Learning multi-view image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4690-4699, 2021. 3, 5 +[60] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic + +graph cnn for learning on point clouds. ACM Transactions on Graphics, 38(5):1-12, 2019. 3, 5 +[61] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing (TIP), 13(4):600-612, 2004. 5 +[62] Jiajun Wu, Chengkai Zhang, Tianfan Xue, Bill Freeman, and Josh Tenenbaum. Learning a probabilistic latent space of object shapes via 3d generative-adversarial modeling. In Advances in Neural Information Processing Systems (NIPS), volume 29, 2016. 3 +[63] Tong Wu, Jiaqi Wang, Xingang Pan, Xudong Xu, Christian Theobalt, Ziwei Liu, and Dahua Lin. Voxurf: Voxel-based efficient and accurate neural surface reconstruction. arXiv.org, 2208.12697, 2022. 3, 6 +[64] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1912-1920, 2015. 2, 3, 4 +[65] Tiange Xiang, Chaoyi Zhang, Yang Song, Jianhui Yu, and Weidong Cai. Walk in the cloud: Learning curves for point clouds shape analysis. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 915–924, 2021. 5 +[66] Mutian Xu, Runyu Ding, Hengshuang Zhao, and Xiaojuan Qi. Paconv: Position adaptive convolution with dynamic kernel assembling on point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3173-3182, 2021. 5 +[67] Mutian Xu, Junhao Zhang, Zhipeng Zhou, Mingye Xu, Xiaojuan Qi, and Yu Qiao. Learning geometry-disentangled representation for complementary understanding of 3d object point cloud. In Proceedings of the Conference on Artificial Intelligence (AAAI), volume 35, pages 3056-3064, 2021. 5 +[68] Guandao Yang, Xun Huang, Zekun Hao, Ming-Yu Liu, Serge Belongie, and Bharath Hariharan. Pointflow: 3d point cloud generation with continuous normalizing flows. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4541-4550, 2019. 3 +[69] Yao Yao, Zixin Luo, Shiwei Li, Jingyang Zhang, Yufan Ren, Lei Zhou, Tian Fang, and Long Quan. Blendedmvs: A large-scale dataset for generalized multi-view stereo networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1790-1799, 2020. 2 +[70] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. In Advances in Neural Information Processing Systems (NIPS), volume 34, pages 4805-4815, 2021. 3, 6 +[71] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4578-4587, 2021. 3, 5, 7 + +[72] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. In Advances in Neural Information Processing Systems (NIPS), 2022. 3, 7 +[73] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 586-595, 2018. 5 +[74] Linqi Zhou, Yilun Du, and Jiajun Wu. 3d shape generation and completion through point-voxel diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5826-5835, 2021. 3 +[75] Qian-Yi Zhou, Jaesik Park, and Vladlen Koltun. Open3D: A modern library for 3D data processing. arXiv.org, 1801.09847, 2018. 4 \ No newline at end of file diff --git a/2023/OmniObject3D_ Large-Vocabulary 3D Object Dataset for Realistic Perception, Reconstruction and Generation/images.zip b/2023/OmniObject3D_ Large-Vocabulary 3D Object Dataset for Realistic Perception, Reconstruction and Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..0afec65145b89f15c75a68c5a6fab18c8d90481e --- /dev/null +++ b/2023/OmniObject3D_ Large-Vocabulary 3D Object Dataset for Realistic Perception, Reconstruction and Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:05b1df0119206514eb76b4b2996b16b95f734d224939b20ba075ff491b476f4e +size 637922 diff --git a/2023/OmniObject3D_ Large-Vocabulary 3D Object Dataset for Realistic Perception, Reconstruction and Generation/layout.json b/2023/OmniObject3D_ Large-Vocabulary 3D Object Dataset for Realistic Perception, Reconstruction and Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..07baca30e69b3bcbeb5e59637e1170512ef64d8a --- /dev/null +++ b/2023/OmniObject3D_ Large-Vocabulary 3D Object Dataset for Realistic Perception, Reconstruction and Generation/layout.json @@ -0,0 +1,7949 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 119, + 103, + 475, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 103, + 475, + 140 + ], + "spans": [ + { + "bbox": [ + 119, + 103, + 475, + 140 + ], + "type": "text", + "content": "OmniObject3D: Large-Vocabulary 3D Object Dataset for Realistic Perception, Reconstruction and Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 119, + 160, + 509, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 160, + 509, + 189 + ], + "spans": [ + { + "bbox": [ + 119, + 160, + 509, + 189 + ], + "type": "text", + "content": "Tong Wu" + }, + { + "bbox": [ + 119, + 160, + 509, + 189 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 119, + 160, + 509, + 189 + ], + "type": "text", + "content": ", Jiarui Zhang" + }, + { + "bbox": [ + 119, + 160, + 509, + 189 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 119, + 160, + 509, + 189 + ], + "type": "text", + "content": ", Xiao Fu" + }, + { + "bbox": [ + 119, + 160, + 509, + 189 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 119, + 160, + 509, + 189 + ], + "type": "text", + "content": ", Yuxin Wang" + }, + { + "bbox": [ + 119, + 160, + 509, + 189 + ], + "type": "inline_equation", + "content": "^{1,4}" + }, + { + "bbox": [ + 119, + 160, + 509, + 189 + ], + "type": "text", + "content": ", Jiawei Ren" + }, + { + "bbox": [ + 119, + 160, + 509, + 189 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 119, + 160, + 509, + 189 + ], + "type": "text", + "content": ", Liang Pan" + }, + { + "bbox": [ + 119, + 160, + 509, + 189 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 119, + 160, + 509, + 189 + ], + "type": "text", + "content": ", Wayne Wu" + }, + { + "bbox": [ + 119, + 160, + 509, + 189 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 119, + 160, + 509, + 189 + ], + "type": "text", + "content": ", Lei Yang" + }, + { + "bbox": [ + 119, + 160, + 509, + 189 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 119, + 160, + 509, + 189 + ], + "type": "text", + "content": ", Jiaqi Wang" + }, + { + "bbox": [ + 119, + 160, + 509, + 189 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 119, + 160, + 509, + 189 + ], + "type": "text", + "content": ", Chen Qian" + }, + { + "bbox": [ + 119, + 160, + 509, + 189 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 119, + 160, + 509, + 189 + ], + "type": "text", + "content": ", Dahua Lin" + }, + { + "bbox": [ + 119, + 160, + 509, + 189 + ], + "type": "inline_equation", + "content": "^{1,2\\boxtimes}" + }, + { + "bbox": [ + 119, + 160, + 509, + 189 + ], + "type": "text", + "content": ", Ziwei Liu" + }, + { + "bbox": [ + 119, + 160, + 509, + 189 + ], + "type": "inline_equation", + "content": "^{5\\boxtimes}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 190, + 575, + 218 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 190, + 575, + 218 + ], + "spans": [ + { + "bbox": [ + 54, + 190, + 575, + 218 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 54, + 190, + 575, + 218 + ], + "type": "text", + "content": "Shanghai Artificial Intelligence Laboratory, " + }, + { + "bbox": [ + 54, + 190, + 575, + 218 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 54, + 190, + 575, + 218 + ], + "type": "text", + "content": "The Chinese University of Hong Kong, " + }, + { + "bbox": [ + 54, + 190, + 575, + 218 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 54, + 190, + 575, + 218 + ], + "type": "text", + "content": "SenseTime Research, " + }, + { + "bbox": [ + 54, + 190, + 575, + 218 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 54, + 190, + 575, + 218 + ], + "type": "text", + "content": "Hong Kong University of Science and Technology, " + }, + { + "bbox": [ + 54, + 190, + 575, + 218 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 54, + 190, + 575, + 218 + ], + "type": "text", + "content": "S-Lab, Nanyang Technological University" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 62, + 227, + 410, + 414 + ], + "blocks": [ + { + "bbox": [ + 62, + 227, + 410, + 414 + ], + "lines": [ + { + "bbox": [ + 62, + 227, + 410, + 414 + ], + "spans": [ + { + "bbox": [ + 62, + 227, + 410, + 414 + ], + "type": "image", + "image_path": "d056abb644ec104bb811689b4c3f1b8fab84ffe54f420c3ff8e80a763e512e92.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 419, + 232, + 512, + 281 + ], + "blocks": [ + { + "bbox": [ + 419, + 232, + 512, + 281 + ], + "lines": [ + { + "bbox": [ + 419, + 232, + 512, + 281 + ], + "spans": [ + { + "bbox": [ + 419, + 232, + 512, + 281 + ], + "type": "image", + "image_path": "91a4275a8b6f75345ad6b9a173b3caa5c79231e9ed119771bfea692e71cfc2dc.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 446, + 282, + 496, + 289 + ], + "lines": [ + { + "bbox": [ + 446, + 282, + 496, + 289 + ], + "spans": [ + { + "bbox": [ + 446, + 282, + 496, + 289 + ], + "type": "text", + "content": "Textured meshes" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 419, + 292, + 527, + 337 + ], + "blocks": [ + { + "bbox": [ + 419, + 292, + 527, + 337 + ], + "lines": [ + { + "bbox": [ + 419, + 292, + 527, + 337 + ], + "spans": [ + { + "bbox": [ + 419, + 292, + 527, + 337 + ], + "type": "image", + "image_path": "d78849eede2e862c95f04572aa4c1495954206ce24c266472b216583f5820702.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 419, + 338, + 518, + 346 + ], + "lines": [ + { + "bbox": [ + 419, + 338, + 518, + 346 + ], + "spans": [ + { + "bbox": [ + 419, + 338, + 518, + 346 + ], + "type": "text", + "content": "Point clouds Rendered images" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 413, + 348, + 529, + 403 + ], + "blocks": [ + { + "bbox": [ + 413, + 348, + 529, + 403 + ], + "lines": [ + { + "bbox": [ + 413, + 348, + 529, + 403 + ], + "spans": [ + { + "bbox": [ + 413, + 348, + 529, + 403 + ], + "type": "image", + "image_path": "9777ee296f0a8c41cf97ec290c7efdd6b42c764d2eed5c1ee91223b4f5bce0de.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 441, + 405, + 503, + 413 + ], + "lines": [ + { + "bbox": [ + 441, + 405, + 503, + 413 + ], + "spans": [ + { + "bbox": [ + 441, + 405, + 503, + 413 + ], + "type": "text", + "content": "Real-captured videos" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 61, + 415, + 531, + 485 + ], + "blocks": [ + { + "bbox": [ + 61, + 415, + 531, + 485 + ], + "lines": [ + { + "bbox": [ + 61, + 415, + 531, + 485 + ], + "spans": [ + { + "bbox": [ + 61, + 415, + 531, + 485 + ], + "type": "image", + "image_path": "346d9e3b169997728a6217be4005801a4bb2a6c29de013bcc2e65406c124c1a4.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 487, + 547, + 510 + ], + "lines": [ + { + "bbox": [ + 46, + 487, + 547, + 510 + ], + "spans": [ + { + "bbox": [ + 46, + 487, + 547, + 510 + ], + "type": "text", + "content": "Figure 1. OmniObject3D is a large vocabulary 3D object dataset with massive high-quality real-scanned 3D objects and rich annotations. It supports various research topics, e.g., perception, novel view synthesis, neural surface reconstruction, and 3D generation." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 143, + 519, + 192, + 532 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 519, + 192, + 532 + ], + "spans": [ + { + "bbox": [ + 143, + 519, + 192, + 532 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 46, + 545, + 290, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 290, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 290, + 689 + ], + "type": "text", + "content": "Recent advances in modeling 3D objects mostly rely on synthetic datasets due to the lack of large-scale real-scanned 3D databases. To facilitate the development of 3D perception, reconstruction, and generation in the real world, we propose OmniObject3D, a large vocabulary 3D object dataset with massive high-quality real-scanned 3D objects. OmniObject3D has several appealing properties: 1) Large Vocabulary: It comprises 6,000 scanned objects in 190 daily categories, sharing common classes with popular 2D datasets (e.g., ImageNet and LVIS), benefiting the pursuit of generalizable 3D representations. 2) Rich Annotations: Each 3D object is captured with both 2D and" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 520, + 547, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 520, + 547, + 651 + ], + "spans": [ + { + "bbox": [ + 304, + 520, + 547, + 651 + ], + "type": "text", + "content": "3D sensors, providing textured meshes, point clouds, multiview rendered images, and multiple real-captured videos. 3) Realistic Scans: The professional scanners support high-quality object scans with precise shapes and realistic appearances. With the vast exploration space offered by OmniObject3D, we carefully set up four evaluation tracks: a) robust 3D perception, b) novel-view synthesis, c) neural surface reconstruction, and d) 3D object generation. Extensive studies are performed on these four benchmarks, revealing new observations, challenges, and opportunities for future research in realistic 3D vision." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 666, + 387, + 679 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 666, + 387, + 679 + ], + "spans": [ + { + "bbox": [ + 306, + 666, + 387, + 679 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 689, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 547, + 715 + ], + "type": "text", + "content": "Sensing, understanding, and synthesizing realistic 3D objects is a long-standing problem in computer vision, with" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 702, + 282, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 702, + 282, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 702, + 282, + 714 + ], + "type": "text", + "content": "Corresponding authors. https://omniobject3d.github.io/" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "803" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 192 + ], + "type": "text", + "content": "rapid progress emerging in recent years. However, a majority of the technical approaches rely on unrealistic synthetic datasets [6, 19, 64] due to the absence of a large-scale real-world 3D object database. However, the appearance and distribution gaps between synthetic and real data cannot be compensated for trivially, hindering their real-life applications. Therefore, it is imperative to equip the community with a large-scale and high-quality 3D object dataset from the real world, which can facilitate a variety of 3D vision tasks and downstream applications." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 194, + 289, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 194, + 289, + 327 + ], + "spans": [ + { + "bbox": [ + 46, + 194, + 289, + 327 + ], + "type": "text", + "content": "Recent advances partially fulfill the requirements while still being unsatisfactory. As shown in Table 1, CO3D [48] contains 19k videos capturing objects from 50 MS-COCO categories, while only " + }, + { + "bbox": [ + 46, + 194, + 289, + 327 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 46, + 194, + 289, + 327 + ], + "type": "text", + "content": " of the videos are annotated with accurate point clouds reconstructed by COLMAP [50]. Moreover, they do not provide textured meshes. GSO [16] has 1k scanned objects while covering only 17 household classes. AKB-48 [33] focuses on robotics manipulation with 2k articulated object scans in 48 categories, but the focus on articulation leads to a relatively narrow semantic distribution, failing to support general 3D object research." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 327, + 289, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 327, + 289, + 568 + ], + "spans": [ + { + "bbox": [ + 46, + 327, + 289, + 568 + ], + "type": "text", + "content": "To boost the research on general 3D object understanding and modeling, we present OmniObject3D: a large-vocabulary 3D object dataset with massive high-quality, real-scanned 3D objects. Our dataset has several appealing properties: 1) Large Vocabulary: It contains 6,000 high-quality textured meshes scanned from real-world objects, which, to the best of our knowledge, is the largest among real-world 3D object datasets with accurate 3D meshes. It comprises 190 daily categories, sharing common classes with popular 2D and 3D datasets (e.g., ImageNet [15], LVIS [25], and ShapeNet [6]), incorporating most daily object realms (See Figure 1 and Figure 2). 2) Rich Annotations: Each 3D object is captured with both 2D and 3D sensors, providing textured 3D meshes, sampled point clouds, posed multi-view images rendered by Blender [13], and real-captured video frames with foreground masks and COLMAP camera poses. 3) Realistic Scans: The object scans are of high fidelity thanks to the professional scanners, bearing precise shapes with geometric details and realistic appearance with high-frequency textures." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 289, + 715 + ], + "type": "text", + "content": "Taking advantage of the vast exploration space offered by OmniObject3D, we carefully set up four evaluation tracks: a) robust 3D perception, b) novel-view synthesis, c) neural surface reconstruction, and d) 3D object generation. Extensive studies are performed on these benchmarks: First, the high-quality, real-world point clouds in OmniObject3D allow us to perform robust 3D perception analysis on both out-of-distribution (OOD) styles and corruptions, two major challenges in point cloud OOD generalization. Furthermore, we provide massive 3D models with multi-view images and precise 3D meshes for novel-view synthesis and neural surface reconstruction. The broad diver" + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 309, + 103, + 552, + 230 + ], + "blocks": [ + { + "bbox": [ + 304, + 70, + 547, + 103 + ], + "lines": [ + { + "bbox": [ + 304, + 70, + 547, + 103 + ], + "spans": [ + { + "bbox": [ + 304, + 70, + 547, + 103 + ], + "type": "text", + "content": "Table 1. A comparison between OmniObject3D and other commonly-used 3D object datasets. " + }, + { + "bbox": [ + 304, + 70, + 547, + 103 + ], + "type": "inline_equation", + "content": "R^{LVis}" + }, + { + "bbox": [ + 304, + 70, + 547, + 103 + ], + "type": "text", + "content": " denotes the ratio of the 1.2k LVIS [25] categories being covered." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 103, + 552, + 230 + ], + "lines": [ + { + "bbox": [ + 309, + 103, + 552, + 230 + ], + "spans": [ + { + "bbox": [ + 309, + 103, + 552, + 230 + ], + "type": "table", + "html": "
DatasetRealFull MeshVideo# Objs# CatsR1vis (%)
ShapeNet [6]✓51k554.1
ModelNet [64]✓12k402.4
3D-Future [19]✓16k341.3
ABO [12]✓8k633.5
Toys4K [53]✓4k1057.7
CO3D V1 / V2 [48]✓✓19 / 40k504.2
DTU [1]✓✓124NA0
ScanObjectNN [56]✓15k151.3
GSO [16]✓✓1k170.9
AKB-48 [33]✓✓2k481.8
Ours✓✓✓6k19010.8
", + "image_path": "e04f38f29771a66063f013ddf7670a63b67532ec91942b1fb3589a4335b5bda5.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 236, + 547, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 236, + 547, + 333 + ], + "spans": [ + { + "bbox": [ + 304, + 236, + 547, + 333 + ], + "type": "text", + "content": "sity in shapes and textures offers a comprehensive training and evaluation source for both scene-specific and generalizable algorithms. Finally, we equip the community with a database for large vocabulary and realistic 3D object generation, which pushes the boundary of existing state-of-the-art generation methods to real-world 3D objects. The four benchmarks reveal new observations, challenges, and opportunities for future research in realistic 3D vision." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 346, + 398, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 346, + 398, + 358 + ], + "spans": [ + { + "bbox": [ + 306, + 346, + 398, + 358 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 366, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 366, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 366, + 547, + 715 + ], + "type": "text", + "content": "3D Object Datasets. The acquisition of a large-scale realistic 3D database is usually expensive and challenging. Many widely-used 3D datasets prefer to collect synthetic CAD models from online repositories [6, 53, 64], for example, ShapeNet [6] has 51,300 3D CAD models in 55 categories, and ModelNet40 [64] consists of 12,311 models in 40 categories. Recent works, e.g., 3D-FUTURE [19] and ABO [12], introduce high-quality CAD models with rich geometric details and informative textures. However, due to the inevitable gap between synthetic and real objects, the community is still eager for a large-scale 3D object dataset in the real world. DTU [1] and BlendedMVS [69] are photo-realistic datasets designed for multi-view stereo benchmarks, while they are small in scale and lack category annotations. ScanObjectNN [56] is a real-world point cloud object dataset based on scanned indoor scenes, containing around 15,000 objects with colored point clouds in 15 categories. However, the point clouds are incomplete and noisy, and multiple objects usually co-exist in one scene. GSO [16] has 1,030 scanned objects with fine geometries and textures in 17 household items, and AKB-48 [33] focuses on robotics manipulation with 2,037 articulated object models in 48 articulated object categories. However, the relatively narrow semantic scope of GSO and AKB-48 hinders their applications for more general 3D research. CO3D v1 [48] contains 19,000 object-centric videos, while only " + }, + { + "bbox": [ + 304, + 366, + 547, + 715 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 304, + 366, + 547, + 715 + ], + "type": "text", + "content": " of them are annotated with accurate point clouds reconstructed by COLMAP [50], and they do not provide meshes or textures. In contrast, the proposed OmniOb" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "type": "text", + "content": "804" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 70, + 397, + 204 + ], + "blocks": [ + { + "bbox": [ + 60, + 70, + 397, + 204 + ], + "lines": [ + { + "bbox": [ + 60, + 70, + 397, + 204 + ], + "spans": [ + { + "bbox": [ + 60, + 70, + 397, + 204 + ], + "type": "image", + "image_path": "a6195fa50c5f41afcc508a8931ff1a9b4ddf8bff89866401733e9bbf6c81687d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 205, + 545, + 228 + ], + "lines": [ + { + "bbox": [ + 46, + 205, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 46, + 205, + 545, + 228 + ], + "type": "text", + "content": "Figure 2. Semantic distribution of the OmniObject3D dataset. It covers 190 daily categories with a long-tailed distribution, sharing common classes with popular 2D and 3D datasets." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 404, + 78, + 529, + 202 + ], + "blocks": [ + { + "bbox": [ + 404, + 78, + 529, + 202 + ], + "lines": [ + { + "bbox": [ + 404, + 78, + 529, + 202 + ], + "spans": [ + { + "bbox": [ + 404, + 78, + 529, + 202 + ], + "type": "image", + "image_path": "e19fc1e9e1bdc9b6e89a005fa17ee1b5f55474ecc4e6b2a1c1505ed18e97623d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 234, + 288, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 234, + 288, + 293 + ], + "spans": [ + { + "bbox": [ + 46, + 234, + 288, + 293 + ], + "type": "text", + "content": "project3D dataset comprises 6,000 3D objects scanned by professional devices with meshes, textures, and multi-view photos in 190 categories, fulfilling the requirements of a wide range of research objectives. A detailed comparison is presented in Table 1." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 294, + 288, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 294, + 288, + 510 + ], + "spans": [ + { + "bbox": [ + 46, + 294, + 288, + 510 + ], + "type": "text", + "content": "Robust 3D Perception. Robustness to out-of-distribution (OOD) data is important in point cloud perception. Two main challenges include OOD styles (e.g., differences between CAD models and real-world objects) and OOD corruptions (e.g., random point jittering or missing due to sensory inaccuracy). A branch of works [10, 30, 45, 60] studies the OOD corruptions and proposes standard corruption test suites [49, 55], while they fail to take account of OOD styles. Another branch of works [3, 48] evaluates the sim-to-real domain gap by training models on clean synthetic datasets [64] and testing them on noisy real-world test sets [56], while OOD styles and corruptions cannot be disentangled under this setting for an independent analysis. In this work, we leverage high-quality, real-world point clouds from OmniObject3D to systematically measure the robustness against the OOD style and OOD corruptions, providing the first benchmark for fine-grained evaluation of the point cloud perception robustness." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 510, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 288, + 713 + ], + "type": "text", + "content": "Neural Radiance Field and Neural Surface Reconstruction. Neural radiance field (NeRF) [40] represents a scene with a fully-connected deep network (MLPs), which takes in hundreds of sampled points along each camera ray and outputs the predicted color and density. We can synthesize the image of an unseen view from a trained model via volume rendering. Inspired by the success of NeRF, massive follow-up efforts have been made to improve its quality [4, 5, 39, 57] and efficiency [7, 18, 42, 54]. A branch of works [8, 27, 35, 48, 59, 71] has also explored the generalization ability of NeRF-based frameworks, where they aim to learn priors from deep image features across multiple scenes. Beyond novel view synthesis, another trend of approaches [14, 43, 58, 63, 70] proposes to combine neural radiance field with implicit surface representations like Signed Distance Function (SDF), and they achieve accurate and mask-free surface reconstruction from multi-view im" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 234, + 545, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 234, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 304, + 234, + 545, + 365 + ], + "type": "text", + "content": "ages. Since dense camera views of scenes are sometimes unavailable, recent advances explore surface reconstruction from sparse views. They address the problem by exploiting generalizable priors cross scenes for a generic surface prediction [36] or taking advantage of the estimated geometry cues estimated by pre-trained networks [72]. OmniObject3D can serve as a large-scale benchmark with realistic photos and accurate meshes for both training and evaluation. The high diversity in shape and appearance offers an opportunity for pursuing more generalizable and robust novel view synthesis and surface reconstruction methods." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 366, + 545, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 366, + 545, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 366, + 545, + 521 + ], + "type": "text", + "content": "3D Object Generation. Early approaches [20,26,37,52,62] extend 2D generation frameworks to 3D voxels with a high computational cost. Some others adopt different 3D data formulations, e.g., point cloud [2,41,68,74], octree [29], and implicit representations [11,38]. However, it is nontrivial to generate complex and textured surfaces. Recent advances [9,21,44] explore the generation for textured 3D meshes, where GET3D [22] is a state-of-the-art approach that generates diverse meshes with rich geometry and textures in two branches. It is a promising but challenging task to train generative models on a large vocabulary and realistic dataset. We evaluate GET3D on our dataset and reveal several challenges and future opportunities." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 521, + 545, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 521, + 545, + 545 + ], + "spans": [ + { + "bbox": [ + 305, + 521, + 545, + 545 + ], + "type": "text", + "content": "In supplementary materials, we present more detailed discussions on related works for different tracks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 557, + 465, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 557, + 465, + 572 + ], + "spans": [ + { + "bbox": [ + 306, + 557, + 465, + 572 + ], + "type": "text", + "content": "3. The OmniObject3D Dataset" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 578, + 545, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 578, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 305, + 578, + 545, + 613 + ], + "type": "text", + "content": "In this section, we describe the data collection, processing, and annotation pipeline of OmniObject3D. We also introduce the statistics and distribution of it." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 623, + 535, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 623, + 535, + 635 + ], + "spans": [ + { + "bbox": [ + 306, + 623, + 535, + 635 + ], + "type": "text", + "content": "3.1. Data Collection, Processing, and Annotation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 641, + 545, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 715 + ], + "type": "text", + "content": "Category List Definition. In order to collect a large amount of 3D objects that are both commonly-distributed and highly diverse, we first pre-define a category list according to several popular 2D and 3D datasets [6,15,25,31, 32,51,64]. We cover most of the categories that lie within the application scope of the scanners and also dynamically" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "805" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 119 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 119 + ], + "type": "text", + "content": "expand the list with reasonable new classes that are absent from the current list during collection. We end up with 190 widely-spread categories, which ensures a library with rich texture, geometry, and semantic information." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 121, + 289, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 121, + 289, + 335 + ], + "spans": [ + { + "bbox": [ + 46, + 121, + 289, + 335 + ], + "type": "text", + "content": "Object Collection Pipeline. We then collect a variety of objects from each category and use professional 3D scanners to obtain high-resolution textured meshes. Specifically, we use the Shining 3D scanner1 and Artec Eva 3D scanner2 for objects in different scales. The scanning time varies with the properties of the object: it takes around 15 minutes to scan a small rigid object with a simple geometry (e.g., an apple, a toy), while it takes up to an hour to obtain a qualified 3D scan for non-rigid, complex, or large objects (e.g., a bed, a kite). For around " + }, + { + "bbox": [ + 46, + 121, + 289, + 335 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 121, + 289, + 335 + ], + "type": "text", + "content": " of the objects, we conduct common manipulations (e.g., taking a bite, cutting in pieces) to conform the natural instincts of them. The 3D scans can faithfully retain the real-world scale of each object, but their poses are not strictly aligned. We thus pre-define a canonical pose for each category and manually align the objects within a category. We then check the quality of each scan, and around " + }, + { + "bbox": [ + 46, + 121, + 289, + 335 + ], + "type": "inline_equation", + "content": "83\\%" + }, + { + "bbox": [ + 46, + 121, + 289, + 335 + ], + "type": "text", + "content": " high-quality ones out of the total collection are finally reserved in the dataset." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 336, + 289, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 336, + 289, + 685 + ], + "spans": [ + { + "bbox": [ + 46, + 336, + 289, + 685 + ], + "type": "text", + "content": "Image Rendering and Point Cloud Sampling. To support a variety of research topics like point cloud analysis, neural radiance fields, and 3D generation, we render multiview images and sample point clouds based on the collected 3D models. We use Blender [13] to render object-centric and photo-realistic multi-view images, together with accurate camera poses. The images are rendered from 100 random viewpoints sampled on the upper hemisphere at " + }, + { + "bbox": [ + 46, + 336, + 289, + 685 + ], + "type": "inline_equation", + "content": "800 \\times 800" + }, + { + "bbox": [ + 46, + 336, + 289, + 685 + ], + "type": "text", + "content": " pixels. We also produce high-resolution mid-level cues like depth and normal for more research use. We then uniformly sample multi-resolution point clouds from each 3D model using the Open3D toolbox [75], with " + }, + { + "bbox": [ + 46, + 336, + 289, + 685 + ], + "type": "inline_equation", + "content": "2^{n}" + }, + { + "bbox": [ + 46, + 336, + 289, + 685 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 46, + 336, + 289, + 685 + ], + "type": "inline_equation", + "content": "n \\in \\{10, 11, 12, 13, 14\\}" + }, + { + "bbox": [ + 46, + 336, + 289, + 685 + ], + "type": "text", + "content": ") points in each point cloud, respectively. Besides the data existing in the dataset, we also provide a data generation pipeline. One can easily obtain new data with self-defined camera distributions, lighting, and point sampling methods to meet different requirements. Video Capturing and Annotation. After scanning each object, we capture its video with an iPhone 12 Pro mobile phone. The object is placed on or beside a calibration board, and each video covers a full " + }, + { + "bbox": [ + 46, + 336, + 289, + 685 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 46, + 336, + 289, + 685 + ], + "type": "text", + "content": " range around it. Square corners on the calibration board can be recognized by the QR Codes beside it, and we then filter out the blurry frames whose recognized corners are less than 8. We uniformly sample 200 frames, and then COLMAP [50], a well-known SfM pipeline, is applied to annotate the frames with camera poses. Finally, we use the scales of the calibration board in both the SfM coordinate space and the real world to recover the absolute scale of the SfM coordinate system." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 323, + 76, + 533, + 145 + ], + "blocks": [ + { + "bbox": [ + 323, + 76, + 533, + 145 + ], + "lines": [ + { + "bbox": [ + 323, + 76, + 533, + 145 + ], + "spans": [ + { + "bbox": [ + 323, + 76, + 533, + 145 + ], + "type": "image", + "image_path": "d6b968714584f3496f44ecb711309889bc0e7447485a1fa46be3725784a42863.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 148, + 547, + 192 + ], + "lines": [ + { + "bbox": [ + 304, + 148, + 547, + 192 + ], + "spans": [ + { + "bbox": [ + 304, + 148, + 547, + 192 + ], + "type": "text", + "content": "Figure 3. OmniObject3D provides the first clean real-world point cloud object dataset and allows fine-grained analysis on robustness to OOD styles and OOD corruptions.\"-C\": corrupted by common corruptions described in [49]" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 305, + 194, + 547, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 194, + 547, + 242 + ], + "spans": [ + { + "bbox": [ + 305, + 194, + 547, + 242 + ], + "type": "text", + "content": "We also develop a two-stage matting pipeline based on the U" + }, + { + "bbox": [ + 305, + 194, + 547, + 242 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 305, + 194, + 547, + 242 + ], + "type": "text", + "content": "Net [47] and FBA [17] matting model to generate the foreground masks for all the frames. Please refer to supplementary materials for more implementation details." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 250, + 451, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 250, + 451, + 262 + ], + "spans": [ + { + "bbox": [ + 306, + 250, + 451, + 262 + ], + "type": "text", + "content": "3.2. Statistics and Distribution" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 268, + 547, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 268, + 547, + 400 + ], + "spans": [ + { + "bbox": [ + 304, + 268, + 547, + 400 + ], + "type": "text", + "content": "With 6,000 3D models in 190 categories, OmniObject3D exhibits a long-tailed distribution with an average of around 30 objects in each category, as shown in Figure 2. It shares many common categories with several famous 2D and 3D datasets [6, 15, 25, 31, 32, 51, 64]. For example, it covers 85 categories in ImageNet [15] and 130 categories in LVIS [25], which leads to the highest " + }, + { + "bbox": [ + 304, + 268, + 547, + 400 + ], + "type": "inline_equation", + "content": "R^{\\mathrm{vis}}" + }, + { + "bbox": [ + 304, + 268, + 547, + 400 + ], + "type": "text", + "content": " in Table 1. Most of the categories are covered by the Open Images [31] image-level labels. It bears a huge diversity in shapes and appearances. The vast semantic and geometrical exploration spaces enable a wide range of research objectives." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 411, + 388, + 424 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 411, + 388, + 424 + ], + "spans": [ + { + "bbox": [ + 306, + 411, + 388, + 424 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 431, + 430, + 444 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 431, + 430, + 444 + ], + "spans": [ + { + "bbox": [ + 306, + 431, + 430, + 444 + ], + "type": "text", + "content": "4.1. Robust 3D Perception" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 449, + 547, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 547, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 547, + 521 + ], + "type": "text", + "content": "Object-level point cloud classification is one of the most fundamental tasks in point cloud perception. In this section, we show how OmniObject3D boosts robustness analysis of point cloud classification by disentangling the two critical out-of-distribution (OOD) challenges introduced in Sec. 2, i.e., OOD styles and OOD corruptions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 521, + 547, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 547, + 556 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 547, + 556 + ], + "type": "text", + "content": "Existing robustness evaluation utilizes clean synthetic datasets, e.g., ModelNet [64], for training and sets up two kinds of test sets for evaluation:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 557, + 545, + 689 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 305, + 557, + 545, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 557, + 545, + 641 + ], + "spans": [ + { + "bbox": [ + 305, + 557, + 545, + 641 + ], + "type": "text", + "content": "1) Noisy real-world datasets, e.g., ScanObjectNN [56], which are cropped from real-world scenes. They are employed to measure the robustness of the sim-to-real domain gap. However, the gap couples both OOD styles and OOD corruptions simultaneously as the cropped point clouds are always noisy, making it impossible to analyze the two robustness challenges independently." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 641, + 545, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 641, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 305, + 641, + 545, + 689 + ], + "type": "text", + "content": "2) Corrupted synthetic datasets, e.g., ModelNet-C [49], which are artificially perturbed on top of clean synthetic datasets. The evaluation allows for detailed corruption analysis, but they do not reflect the robustness to OOD styles." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 713 + ], + "type": "text", + "content": "None of the existing robustness benchmarks allows for analyzing the robustness to both OOD styles and OOD cor" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 47, + 693, + 136, + 702 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 693, + 136, + 702 + ], + "spans": [ + { + "bbox": [ + 47, + 693, + 136, + 702 + ], + "type": "text", + "content": "1https://www.einscan.com/" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 703, + 130, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 703, + 130, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 703, + 130, + 712 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 48, + 703, + 130, + 712 + ], + "type": "text", + "content": "https://www.artec3d.cn/" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "806" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 159, + 294, + 291 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 288, + 159 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 288, + 159 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 288, + 159 + ], + "type": "text", + "content": "Table 2. Point cloud perception robustness analysis on OmniObject3D with different architecture designs. Models are trained on the ModelNet-40 dataset, with OAClean to be their overall accuracy on the standard ModelNet-40 test set. OAStyle on OmniObject3D evaluates the robustness to OOD styles. mCE on the corrupted OmniObject3D-C evaluates the robustness to OOD corruptions. Blue shadings indicate rankings. " + }, + { + "bbox": [ + 47, + 70, + 288, + 159 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 47, + 70, + 288, + 159 + ], + "type": "text", + "content": ": results on ModelNet-C [49]. Full results are presented in the supplementary materials." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 159, + 294, + 291 + ], + "lines": [ + { + "bbox": [ + 47, + 159, + 294, + 291 + ], + "spans": [ + { + "bbox": [ + 47, + 159, + 294, + 291 + ], + "type": "table", + "html": "
mCE†↓OA Clean ↑OA Style ↑mCE ↓
DGCNN [60]1.0000.9260.4481.000
PointNet [45]1.4220.9070.4660.969
PointNet++ [46]1.0720.9300.4071.066
RSCNN [34]1.1300.9230.3931.076
SimpleView [23]1.0470.9390.4760.990
GDANet [67]0.8920.9340.4970.920
PAConv [66]1.1040.9360.4031.073
CurveNet [65]0.9270.9380.5000.929
PCT [24]0.9250.9300.4590.940
RPC [49]0.8630.9300.4720.936
", + "image_path": "6b6a35fe0a6959e8c819485d424d8e69b56b00a0cd5bf73be131071237ca3a9c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 293, + 287, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 293, + 287, + 435 + ], + "spans": [ + { + "bbox": [ + 46, + 293, + 287, + 435 + ], + "type": "text", + "content": "ruptions in fine granularity. OmniObject3D, on the other hand, as the first clean real-world point cloud object dataset, can help to address the issue. For models trained on ModelNet, we first evaluate their performance on OmniObject3D to examine OOD-style robustness. Then, we create OmniObject3D-C by corrupting OmniObject3D with common corruptions described in [49] to examine the OOD-corruption robustness. We show a complete robustness evaluation scheme in Figure 3. For evaluation metrics, we use the overall accuracy (OA) on OmniObject3D to measure the OOD-style robustness and use DGCNN normalized mCE [49] to measure the OOD-corruption robustness." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 436, + 287, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 436, + 287, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 436, + 287, + 628 + ], + "type": "text", + "content": "We benchmark ten state-of-the-art point cloud classification models in Table 2. We observe that 1) performance on a clean test set has little correlation with OOD-style robustness. For example, SimpleView [23] achieves the best " + }, + { + "bbox": [ + 46, + 436, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\mathrm{OA}_{\\mathrm{Clean}}" + }, + { + "bbox": [ + 46, + 436, + 287, + 628 + ], + "type": "text", + "content": " but mediocre " + }, + { + "bbox": [ + 46, + 436, + 287, + 628 + ], + "type": "inline_equation", + "content": "\\mathrm{OA}_{\\mathrm{Style}}" + }, + { + "bbox": [ + 46, + 436, + 287, + 628 + ], + "type": "text", + "content": "; 2) Advanced point grouping, e.g., curve-based point grouping in CurveNet [65] and frequency-based point grouping in GDANet [67], are robust not only to OOD corruptions as pointed out in [49], but also to OOD styles; 3) OOD style + OOD corruption is a more challenging setting. In particular, RPC, the most robust architecture to OOD corruptions [49], shows inferior mCE. In summary, robust point cloud perception models against both OOD styles and OOD corruptions are still under-explored. Our dataset sheds new light on a comprehensive understanding of point cloud perception robustness. See more results in the supplementary materials." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 635, + 170, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 170, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 170, + 647 + ], + "type": "text", + "content": "4.2. Novel View Synthesis" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 654, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 287, + 715 + ], + "type": "text", + "content": "In this section, we study several representative methods on OmniObject3D for novel view synthesis (NVS) in two settings: 1) training on a single scene with densely captured images, which is the standard setting for NeRF [40]; 2) learning priors across scenes from our dataset to explore" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 307, + 92, + 552, + 148 + ], + "blocks": [ + { + "bbox": [ + 305, + 70, + 545, + 92 + ], + "lines": [ + { + "bbox": [ + 305, + 70, + 545, + 92 + ], + "spans": [ + { + "bbox": [ + 305, + 70, + 545, + 92 + ], + "type": "text", + "content": "Table 3. Single-scene novel view synthesis results. Three metrics and their standard deviation (SD) across the training set." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 92, + 552, + 148 + ], + "lines": [ + { + "bbox": [ + 307, + 92, + 552, + 148 + ], + "spans": [ + { + "bbox": [ + 307, + 92, + 552, + 148 + ], + "type": "table", + "html": "
MethodPSNR (↑) / SDSSIM (↑) / SDLPIPS (↓) / SD
NeRF [40]34.01 / 3.460.953 / 0.0290.068 / 0.061
mip-NeRF [4]39.86 / 4.580.974 / 0.0130.084 / 0.048
Plenoxels [18]41.04 / 6.840.982 / 0.0310.030 / 0.031
", + "image_path": "05a7c2fe5cf700cbdae736191e0826fccc6e55ee2a0710520af141b0264a898a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 308, + 178, + 553, + 311 + ], + "blocks": [ + { + "bbox": [ + 306, + 148, + 547, + 178 + ], + "lines": [ + { + "bbox": [ + 306, + 148, + 547, + 178 + ], + "spans": [ + { + "bbox": [ + 306, + 148, + 547, + 178 + ], + "type": "text", + "content": "Table 4. Cross-scene novel view synthesis results on 10 categories. 'Cat.' and 'All*' denote training on each category and training on all categories except the 10 test ones, respectively." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 178, + 553, + 311 + ], + "lines": [ + { + "bbox": [ + 308, + 178, + 553, + 311 + ], + "spans": [ + { + "bbox": [ + 308, + 178, + 553, + 311 + ], + "type": "table", + "html": "
MethodTrainPSNR (↑)SSIM (↑)LPIPS (↓)\\( {\\mathcal{L}}_{1}^{\\text{depth }}\\left( \\downarrow \\right) \\)
MVSNeRF [8]All*17.490.5440.4420.193
Cat.17.540.5420.4480.230
All*-ft.25.700.7540.2510.081
Cat.-ft.25.520.7500.2640.076
IBRNet [59]All*19.390.5690.3990.423
Cat.19.030.5510.4150.290
All*-ft.26.890.7920.2150.081
Cat.-ft.25.670.7600.2380.099
pixelNeRF [71]All*22.160.6920.3420.109
Cat.20.650.6760.3480.195
", + "image_path": "c73cf3e4cbfc5a3e3e4777159082b0b18381bf9e19abcdfa97b87f0315b4f03b.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 317, + 501, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 317, + 501, + 329 + ], + "spans": [ + { + "bbox": [ + 305, + 317, + 501, + 329 + ], + "type": "text", + "content": "the generalization ability of NeRF-style models." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 330, + 545, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 330, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 304, + 330, + 545, + 581 + ], + "type": "text", + "content": "Single-Scene NVs. We select three objects in each category for the experiments, randomly sampling 1/8 images as the hold-out test set. We involve NeRF [40], mip-NeRF [4], and a voxel-based system named Plenoxels [18] for comparison. As in Table 3, we find that Plenoxels achieve the best performance on average for PSNR, SSIM [61], and LPIPS [73]. There exists a clear margin for LPIPS between Plenoxels and the other two methods since voxel-based methods are especially good at modeling high-frequency appearance. We also present the standard deviation (SD) of results across all the training samples, where Plenoxels are relatively unstable compared to NeRF and mip-NeRF. We observe that Plenoxels introduce artifacts when encountering concave geometry (e.g., bowls, chairs) and suffer from an inaccurate density field modeling when the foreground object is dark. MLP-based methods are relatively robust against these difficult cases. In a nutshell, our dataset provides a library with a variety of shapes and appearances, allowing a comprehensive evaluation of different NVS methods. See the supplementary for more results with the iPhone videos, detailed analysis, and visualizations." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": "Cross-Scene NVS. We conduct extensive experiments on novel view synthesis from sparse inputs by pixelNeRF [71], IBRNet [59] and MVSNeRF [8] in Table 4. We select 10 categories with the most various scenes as the test set. All metrics are averaged over 300 images. In the generalization setting, although not trained on the test category, " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{MVSNeRF_{All^{*}}}" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " is comparable to " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{MVSNeRF_{Cat}}" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{IBRNet_{All^{*}}}" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{pixelNeRF_{All^{*}}}" + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": " even outperforms the corresponding \"Cat\"s in all terms of visual metrics, especially on regular-shaped objects such as squash and apple. It confirms that OmniObject3D serves as an information-rich" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "type": "text", + "content": "807" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 71, + 534, + 252 + ], + "blocks": [ + { + "bbox": [ + 63, + 71, + 534, + 252 + ], + "lines": [ + { + "bbox": [ + 63, + 71, + 534, + 252 + ], + "spans": [ + { + "bbox": [ + 63, + 71, + 534, + 252 + ], + "type": "image", + "image_path": "f271b8f08e6032bbe824a41c71e1411e752a0232f71bc4477990c945a3e4e4c6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 70, + 281, + 265, + 370 + ], + "blocks": [ + { + "bbox": [ + 69, + 271, + 265, + 281 + ], + "lines": [ + { + "bbox": [ + 69, + 271, + 265, + 281 + ], + "spans": [ + { + "bbox": [ + 69, + 271, + 265, + 281 + ], + "type": "text", + "content": "Table 5. Dense-view surface reconstruction results." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 70, + 281, + 265, + 370 + ], + "lines": [ + { + "bbox": [ + 70, + 281, + 265, + 370 + ], + "spans": [ + { + "bbox": [ + 70, + 281, + 265, + 370 + ], + "type": "table", + "html": "
MethodChamfer Distance ×103(↓)
HardMediumEasyAvg
NeuS [58]9.265.633.466.09
VolSDF [70]10.064.942.865.92
Voxurf [63]9.014.982.585.49
Avg9.445.192.975.83
", + "image_path": "d53556d028b683893afe85ec9474085c75e82283f769f4e5e019a896be81664b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 376, + 289, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 376, + 289, + 615 + ], + "spans": [ + { + "bbox": [ + 46, + 376, + 289, + 615 + ], + "type": "text", + "content": "dataset that is beneficial for obtaining a strong generalizable prior on unseen scenes. Moreover, it is noteworthy that MVSNeRF and pixelNeRF with 'All* generate better underlying depth than those with 'Cat.', inferring generalizable methods can implicitly learn geometric cues though only trained from appearance in our dataset. It is reasonable that (1) IBRNet suffers more severely than the others in geometry under the scarcity of source context (only 3 views) as it is more suitable for dense-view generalization that complies with its view interpolation module. (2) MVSNeRF lags behind pixelNeRF on visual performance as we take 10 test frames widely distributed around the object in " + }, + { + "bbox": [ + 46, + 376, + 289, + 615 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 46, + 376, + 289, + 615 + ], + "type": "text", + "content": " by FPS sampling algorithm, where cost volume will be inaccurate on large-range viewpoint change. After further finetuning IBRNet for only around 10 minutes on each test scene, " + }, + { + "bbox": [ + 46, + 376, + 289, + 615 + ], + "type": "inline_equation", + "content": "\\mathrm{IBRNet}_{\\mathrm{All}^{*}\\text{-ft}}" + }, + { + "bbox": [ + 46, + 376, + 289, + 615 + ], + "type": "text", + "content": " achieves the best view synthesis results, comparable to test-time optimized NVS methods on nearby views. It is promising to utilize the large-scale and category-prosperous OmniObject3D, to build a benchmark suite for evaluating diverse cross-scene NVS methods." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 623, + 216, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 623, + 216, + 635 + ], + "spans": [ + { + "bbox": [ + 47, + 623, + 216, + 635 + ], + "type": "text", + "content": "4.3. Neural Surface Reconstruction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "type": "text", + "content": "Precise surface reconstruction from multi-view images enables a broad range of applications. For a single scene with dense-view images, algorithms are expected to conduct accurate, robust, and efficient surface reconstruction. When only sparse-view images are available, it is crucial to learn generalizable priors from a set of scenes or use other" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 320, + 271, + 531, + 388 + ], + "blocks": [ + { + "bbox": [ + 118, + 259, + 473, + 270 + ], + "lines": [ + { + "bbox": [ + 118, + 259, + 473, + 270 + ], + "spans": [ + { + "bbox": [ + 118, + 259, + 473, + 270 + ], + "type": "text", + "content": "Figure 4. Neural surface reconstruction results for both dense-view and sparse-view settings." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 320, + 271, + 531, + 388 + ], + "lines": [ + { + "bbox": [ + 320, + 271, + 531, + 388 + ], + "spans": [ + { + "bbox": [ + 320, + 271, + 531, + 388 + ], + "type": "image", + "image_path": "dfb3dafcdbf239dada19f6ddc3709d7fd1165f821c9539386b447cb9ea72e6f2.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 388, + 545, + 421 + ], + "lines": [ + { + "bbox": [ + 304, + 388, + 545, + 421 + ], + "spans": [ + { + "bbox": [ + 304, + 388, + 545, + 421 + ], + "type": "text", + "content": "Figure 5. Performance distribution of dense-view surface reconstruction. The averaged results of the three methods is imbalanced. The colored area denotes a smoothed range of results." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 425, + 545, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 425, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 304, + 425, + 545, + 448 + ], + "type": "text", + "content": "geometric cues to assist reconstruction. Accordingly, we study the two settings for surface reconstruction methods." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 449, + 545, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 545, + 533 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 545, + 533 + ], + "type": "text", + "content": "Dense-View Surface Reconstruction. Under this setting, we include three representative methods. NeuS [58] and VolSDF [70] are two well-known approaches that bridge neural volume rendering with implicit surface representation. We also involve a voxel-based method called Voxsurf [63], which leverages a hybrid representation to achieve acceleration and fine geometry reconstruction." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 546, + 713 + ], + "type": "text", + "content": "Previous approaches in this task mainly perform evaluations and comparisons on 15 scenes from the DTU [1] dataset, which is not comprehensive and robust enough to demonstrate the ability of the methods in different scenarios. In comparison, we select three objects per category to run each of the three methods above, leading to over 1,500 reconstructions in total. We calculate the Chamfer Distance (CD) between the reconstructed surface and the ground truth. The distribution of the results is shown in Figure 5. The average curve is imbalanced: hard categories usually include low-textured, concave, or complex shapes (e.g., bowls, vases, kennels, cabinets, and durians). We thus split the categories into three levels of \"difficulty\" based on the average curve, and the level-wise results are presented in Table 5. We can observe a clear margin among results in" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "808" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 53, + 83, + 281, + 240 + ], + "blocks": [ + { + "bbox": [ + 52, + 71, + 282, + 81 + ], + "lines": [ + { + "bbox": [ + 52, + 71, + 282, + 81 + ], + "spans": [ + { + "bbox": [ + 52, + 71, + 282, + 81 + ], + "type": "text", + "content": "Table 6. Sparse-view (3-view) surface reconstruction results." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 53, + 83, + 281, + 240 + ], + "lines": [ + { + "bbox": [ + 53, + 83, + 281, + 240 + ], + "spans": [ + { + "bbox": [ + 53, + 83, + 281, + 240 + ], + "type": "table", + "html": "
MethodTrainChamfer Distance ×103(↓)
HardMediumEasyAvg
NeuS [58]Single29.3527.6224.7927.3
MonoSDF [72]Single35.1435.3532.7634.68
SparseNeuS [36]1 cat.34.0531.3231.1432.36
10 cats.30.7530.1128.3729.87
All cats.26.1326.0822.1325.00
Easy28.3926.6523.7626.48
Medium27.3826.6623.0825.87
Hard27.4226.9524.6326.47
MVSNeRF [8]All cats.56.6848.0948.7051.16
pixelNeRF [71]All cats.63.3159.9161.4761.56
", + "image_path": "89461208797dfd51a4f256ad0173c3246dc1edd12623c11b9ba91c07b5161f99.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 243, + 286, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 243, + 286, + 266 + ], + "spans": [ + { + "bbox": [ + 46, + 243, + 286, + 266 + ], + "type": "text", + "content": "different levels for each method, indicating the split subsets to be generic and faithful." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 268, + 286, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 268, + 286, + 435 + ], + "spans": [ + { + "bbox": [ + 46, + 268, + 286, + 435 + ], + "type": "text", + "content": "Sparse-View Surface Reconstruction. Dense-captured images of a scene are sometimes not available, so we also study the sparse-view scenario here. The following methods are included: NeuS [58] with sparse-view input; MonoSDF [72], which takes in geometry cues estimated by pre-trained models; SparseNeuS [36], a generic surface prediction pipeline that learns generalizable priors; pixelNeRF [71] and MVSNeRF [8] from Sec. 4.2, whose geometries are extracted from the density field. For NeuS, MonoSDF, and pixelNeRF, we use Farthest Point Sampling (FPS) to sample views that are most widely distributed; for SparseNeuS and MVSNeRF, we conduct FPS among the nearest 30 camera poses from a random reference view. We sample 3 views in all the experiments." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 437, + 286, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 437, + 286, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 437, + 286, + 652 + ], + "type": "text", + "content": "The quantitative and qualitative comparisons are shown in Table 6 and Figure 4, respectively. We observe apparent artifacts in all the sparse-view reconstructed results. Among them, SparseNeuS trained on enough data demonstrates the best quantitative performance on average, and the pre-division on the train set does not result in a noticeable difference across difficulty levels. NeuS with sparse-view input achieves a surprisingly good performance. As shown in Figure 4, the FPS sampling equips NeuS with a coherent global shape for thin structures like case 3, while it encounters severe local geometry ambiguity like case 1. MonoSDF, on the contrary, partially overcomes the issue of ambiguity via the assistance of predicted geometry cues in case 1. However, it relies heavily on the accuracy of the estimated depth and normal and thus easily fails when the estimation is inaccurate (e.g., case 2 and 3). The surfaces extracted from generalized NeRF models, i.e., pixelNeRF and MVSNeRF, are of relatively low quality." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 653, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 286, + 713 + ], + "type": "text", + "content": "In brief, the challenging problem of sparse-view surface reconstruction has not been solved well currently. OmniObject3D is a promising database to study generalizable surface reconstruction pipelines as well as strategies for a robust usage of estimated geometry cues for this track." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 307, + 73, + 416, + 189 + ], + "blocks": [ + { + "bbox": [ + 307, + 73, + 416, + 189 + ], + "lines": [ + { + "bbox": [ + 307, + 73, + 416, + 189 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 416, + 189 + ], + "type": "image", + "image_path": "7a4e5ca54b4d4298c303b4927bb22ea6dc5f3d941993e1c6fa6e38e2bd816e25.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 423, + 73, + 544, + 190 + ], + "blocks": [ + { + "bbox": [ + 423, + 73, + 544, + 190 + ], + "lines": [ + { + "bbox": [ + 423, + 73, + 544, + 190 + ], + "spans": [ + { + "bbox": [ + 423, + 73, + 544, + 190 + ], + "type": "image", + "image_path": "8689dca28d26bf5462052488e2fff21143b2b450c96ec02d09ccc45eed566d5a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 316, + 191, + 416, + 307 + ], + "blocks": [ + { + "bbox": [ + 316, + 191, + 416, + 307 + ], + "lines": [ + { + "bbox": [ + 316, + 191, + 416, + 307 + ], + "spans": [ + { + "bbox": [ + 316, + 191, + 416, + 307 + ], + "type": "image", + "image_path": "a66375ac3ccd9ce8f2115cc07b7f7488a65f96759c7bea379ee2050c5c8b185c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 310, + 545, + 388 + ], + "lines": [ + { + "bbox": [ + 304, + 310, + 545, + 388 + ], + "spans": [ + { + "bbox": [ + 304, + 310, + 545, + 388 + ], + "type": "text", + "content": "Figure 6. The category distribution of the generated shapes. (a) shows a weak positive correlation between the number of generated shapes and training shapes per category. (b) visualizes the correlation matrix among different categories by Chamfer Distance between their mean shapes. (c) visualizes categories being clustered into eight groups by KMeans. (d) presents a clear training-generation relation in the group-level statistics." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 425, + 191, + 538, + 307 + ], + "blocks": [ + { + "bbox": [ + 425, + 191, + 538, + 307 + ], + "lines": [ + { + "bbox": [ + 425, + 191, + 538, + 307 + ], + "spans": [ + { + "bbox": [ + 425, + 191, + 538, + 307 + ], + "type": "image", + "image_path": "6e0ed030abc642b71c9765a4883a92e0c58a8df4e43565e37b2d59ac90178d43.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 394, + 432, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 394, + 432, + 407 + ], + "spans": [ + { + "bbox": [ + 306, + 394, + 432, + 407 + ], + "type": "text", + "content": "4.4.3D Object Generation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 413, + 545, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 413, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 304, + 413, + 545, + 520 + ], + "type": "text", + "content": "In this section, we adopt a state-of-the-art generative model that directly generates explicit textured 3D meshes, namely GET3D [22]. GET3D is originally evaluated on six categories (Car, Chair, Motorbike, Animal, House, and Human Body) with independent models trained on each category. The number of shapes per category ranges from 337 to 7,497. In comparison, OmniObject3D contains many more categories with fewer objects in each. As a result, it is natural to train each model with multiple categories." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 521, + 545, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 640 + ], + "type": "text", + "content": "We first provide some qualitative results in Figure 7, where we show various generation results: The textures are rather realistic, and the shapes are coherent, enhanced by fine geometry details (e.g., the lychee and pineapple). We explore the latent space of the model and show interpolation results in Figure 8. We can observe a smooth transition across instances that are semantically different. We would like to further analyze the performance of the generative model trained on OmniObject3D from three aspects, i.e., semantic distribution, diversity, and quality." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 545, + 713 + ], + "type": "text", + "content": "Semantic Distribution. We randomly select 100 categories to train an unconditional model jointly. We randomly generate 1,000 textured meshes at inference time and ask human experts to label them. Shapes with ambiguity are not counted. Figure 6 (a) shows that the generated shapes per category are highly imbalanced, exhibiting a weak positive" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "809" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 72, + 545, + 164 + ], + "blocks": [ + { + "bbox": [ + 50, + 72, + 545, + 164 + ], + "lines": [ + { + "bbox": [ + 50, + 72, + 545, + 164 + ], + "spans": [ + { + "bbox": [ + 50, + 72, + 545, + 164 + ], + "type": "image", + "image_path": "053c21d8c495415c591922523dff2f9a8c2fc954523d0bc10294d3739734be65.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 49, + 191, + 285, + 258 + ], + "blocks": [ + { + "bbox": [ + 46, + 167, + 546, + 190 + ], + "lines": [ + { + "bbox": [ + 46, + 167, + 546, + 190 + ], + "spans": [ + { + "bbox": [ + 46, + 167, + 546, + 190 + ], + "type": "text", + "content": "Figure 7. Examples of the generated textured shapes rendered in Blender. OmniObject3D enables GET3D with realistic generations across a wide range of categories." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 49, + 191, + 285, + 258 + ], + "lines": [ + { + "bbox": [ + 49, + 191, + 285, + 258 + ], + "spans": [ + { + "bbox": [ + 49, + 191, + 285, + 258 + ], + "type": "image", + "image_path": "43af7c99aaee2740bf7d2ce347fecb8fca0dadfd3a6891a83c2088f299f4d5af.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 261, + 287, + 283 + ], + "lines": [ + { + "bbox": [ + 46, + 261, + 287, + 283 + ], + "spans": [ + { + "bbox": [ + 46, + 261, + 287, + 283 + ], + "type": "text", + "content": "Figure 8. Shape interpolation. We interpolate both geometry and texture latent codes from left to right." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 293, + 287, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 293, + 287, + 520 + ], + "spans": [ + { + "bbox": [ + 46, + 293, + 287, + 520 + ], + "type": "text", + "content": "correlation with the training shape numbers. Actually, the categories are not independent but rather highly correlated. We calculate the \"mean shape\" for each category and visualize the Chamfer Distance among them in Figure 6 (b), which indicates that they can be further grouped. Regarding each matrix row as a feature vector, we use KMeans to cluster them into eight groups (Figure 6 (c)) and carry out group-level statistics in Figure 6 (d). It demonstrates a clear trend that the number of generated shapes increases along with or even faster than the number of training shapes in the group, revealing an enlarged semantic bias during generation. However, it also depends on the inner-group divergence. For example, Group 2 (883 shapes in 27 categories) has the largest number of training samples, while the high variation among its categories prevents it from dominating the generated shapes; Group 1 (587 shapes in 18 categories) has a relatively small divergence, which becomes the most popular in the generated shapes. We present more details in the supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 522, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 522, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 522, + 288, + 713 + ], + "type": "text", + "content": "Diversity and Quality. We select four representative data subsets for training and evaluation, namely fruits, furniture, toys, and Rand-100. We randomly split each subset into training (80%) and testing (20%). We leverage three evaluation metrics: for geometry, we use Chamfer Distance (CD) to compute the Coverage score (Cov) and Minimum Matching Distance (MMD), which focus on the diversity and quality of the shapes, respectively; for texture, we adopt the widely-used FID [28]. But the FID metrics with different test splits are not directly comparable, suffering from a large variance when the test set is small. We thus introduce FIDref for reference, which is the FID between the train and test set. The results are shown in Table 7. Furniture suffers from the lowest quality (MMD) since the small train set with 17 categories is a difficult training source. Fruits has the same number of categories while being 2.3 times" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 309, + 201, + 547, + 266 + ], + "blocks": [ + { + "bbox": [ + 315, + 190, + 534, + 201 + ], + "lines": [ + { + "bbox": [ + 315, + 190, + 534, + 201 + ], + "spans": [ + { + "bbox": [ + 315, + 190, + 534, + 201 + ], + "type": "text", + "content": "Table 7. Quantitative evaluations on different data splits." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 201, + 547, + 266 + ], + "lines": [ + { + "bbox": [ + 309, + 201, + 547, + 266 + ], + "spans": [ + { + "bbox": [ + 309, + 201, + 547, + 266 + ], + "type": "table", + "html": "
Split#Objs#CatsCov (% ↑)MMD (↓)FID (↓)FIDref
Furniture2651767.924.2787.3958.40
Fruits6101746.723.32105.3187.15
Toys339755.222.78122.7741.40
Rand-100295110061.703.8946.578.65
", + "image_path": "2d5af172fc9d839a116403ccc6ba1b084ec24048817b670f5f02fdf650ee1544.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 271, + 545, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 271, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 304, + 271, + 545, + 354 + ], + "type": "text", + "content": "larger in scale, and some fruits share a very similar structure, leading to relatively higher quality and lower diversity (Cov). Toys achieve the best quality by training on only 7 categories. Rand-100 is the most difficult case, and we can observe a trade-off between quality and diversity. Both the FID and " + }, + { + "bbox": [ + 304, + 271, + 545, + 354 + ], + "type": "inline_equation", + "content": "\\mathrm{FID}^{\\mathrm{ref}}" + }, + { + "bbox": [ + 304, + 271, + 545, + 354 + ], + "type": "text", + "content": " are high for the first three subsets due to the small testing sets, while only 'Rand-100' is relatively low." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 356, + 546, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 356, + 546, + 427 + ], + "spans": [ + { + "bbox": [ + 304, + 356, + 546, + 427 + ], + "type": "text", + "content": "In brief, training and evaluating generative models on a large-vocabulary and realistic dataset is a promising but challenging task. We reveal crucial problems like the semantic distribution bias and varying exploration difficulties in different groups. OmniObject3D serves as a great database for further examination in this area." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 437, + 446, + 450 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 437, + 446, + 450 + ], + "spans": [ + { + "bbox": [ + 306, + 437, + 446, + 450 + ], + "type": "text", + "content": "5. Conclusion and Outlook" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 457, + 545, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 457, + 545, + 564 + ], + "spans": [ + { + "bbox": [ + 304, + 457, + 545, + 564 + ], + "type": "text", + "content": "We introduce OmniObject3D, a large vocabulary 3D object dataset with massive high-quality real-scanned 3D objects, including 6,000 objects from 190 categories. It provides rich annotations, including textured 3D meshes, sampled point clouds, posed multi-view images rendered by Blender, and real-captured video frames with foreground masks and COLMAP camera poses. We set up four evaluation tracks, revealing new observations, challenges, and opportunities for future research in realistic 3D vision." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 565, + 545, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 565, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 304, + 565, + 545, + 589 + ], + "type": "text", + "content": "We will regulate the usage of our data to avoid potential negative social impacts." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 590, + 545, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 590, + 545, + 708 + ], + "spans": [ + { + "bbox": [ + 304, + 590, + 545, + 708 + ], + "type": "text", + "content": "Acknowledgement. This project is funded by Shanghai AI Laboratory, CUHK Interdisciplinary AI Research Institute, the Centre for Perceptual and Interactive Intelligence (CPII) Ltd under the Innovation and Technology Commission (ITC)'s InnoHK, Hong Kong RGC Theme-based Research Scheme 2020/21 (No. T41-603/20-R), OpenXDLab, the Ministry of Education, Singapore, under its MOE AcRF Tier 2 (MOE-T2EP20221-0012), NTU NAP, and under the RIE2020 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) Funding Initiative." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "810" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 92, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 53, + 92, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 92, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 92, + 287, + 135 + ], + "type": "text", + "content": "[1] Henrik Aanaes, Rasmus Ramsbøl Jensen, George Vogiatzis, Engin Tola, and Anders Bjorholm Dahl. Large-scale data for multiple-view stereopsis. International Journal of Computer Vision (IJCV), 120(2):153-168, 2016. 2, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 288, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 288, + 192 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 288, + 192 + ], + "type": "text", + "content": "[2] Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas Guibas. Learning representations and generative models for 3d point clouds. In Proceedings of the International Conference on Machine learning (ICML), pages 40-49, 2018. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 194, + 288, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 194, + 288, + 258 + ], + "spans": [ + { + "bbox": [ + 53, + 194, + 288, + 258 + ], + "type": "text", + "content": "[3] Adel Ahmadyan, Liangkai Zhang, Artsiom Ablavatski, Jianing Wei, and Matthias Grundmann. Objectron: A large scale dataset of object-centric videos in the wild with pose annotations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7822-7831, 2021. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 262, + 288, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 262, + 288, + 327 + ], + "spans": [ + { + "bbox": [ + 53, + 262, + 288, + 327 + ], + "type": "text", + "content": "[4] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5855–5864, 2021. 3, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 329, + 288, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 329, + 288, + 384 + ], + "spans": [ + { + "bbox": [ + 53, + 329, + 288, + 384 + ], + "type": "text", + "content": "[5] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5470-5479, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 386, + 288, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 386, + 288, + 441 + ], + "spans": [ + { + "bbox": [ + 53, + 386, + 288, + 441 + ], + "type": "text", + "content": "[6] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv.org, 1512.03012, 2015. 2, 3, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 443, + 288, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 443, + 288, + 485 + ], + "spans": [ + { + "bbox": [ + 53, + 443, + 288, + 485 + ], + "type": "text", + "content": "[7] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 488, + 288, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 488, + 288, + 553 + ], + "spans": [ + { + "bbox": [ + 53, + 488, + 288, + 553 + ], + "type": "text", + "content": "[8] Anpei Chen, Zexiang Xu, Fuqiang Zhao, Xiaoshuai Zhang, Fanbo Xiang, Jingyi Yu, and Hao Su. Mvsnerf: Fast generalizable radiance field reconstruction from multi-view stereo. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14124-14133, 2021. 3, 5, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 556, + 288, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 556, + 288, + 610 + ], + "spans": [ + { + "bbox": [ + 53, + 556, + 288, + 610 + ], + "type": "text", + "content": "[9] Wenzheng Chen, Huan Ling, Jun Gao, Edward Smith, Jaakko Lehtinen, Alec Jacobson, and Sanja Fidler. Learning to predict 3d objects with an interpolation-based differentiable renderer. In Advances in Neural Information Processing Systems (NIPS), volume 32, 2019. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 613, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 287, + 667 + ], + "type": "text", + "content": "[10] Yunlu Chen, Vincent Tao Hu, Efstratios Gavves, Thomas Mensink, Pascal Mettes, Pengwan Yang, and Cees GM Snoek. Pointmixup: Augmentation for point clouds. In Proceedings of the European Conference on Computer Vision (ECCV), pages 330-345, 2020. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "text", + "content": "[11] Zhiqin Chen and Hao Zhang. Learning implicit fields for generative shape modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5939-5948, 2019. 3" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 712 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 148 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 148 + ], + "type": "text", + "content": "[12] Jasmine Collins, Shubham Goel, Kenan Deng, Achleshwar Luthra, Leon Xu, Erhan Gundogdu, Xi Zhang, Tomas F Yago Vicente, Thomas Dideriksien, Himanshu Arora, et al. Abo: Dataset and benchmarks for real-world 3d object understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21126-21136, 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 151, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 151, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 151, + 545, + 173 + ], + "type": "text", + "content": "[13] Blender Online Community. Blender - a 3d modelling and rendering package. 2018. 2, 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 217 + ], + "type": "text", + "content": "[14] François Darmon, Bénédicte Bascle, Jean-Clément Devaux, Pascal Monasse, and Mathieu Aubry. Improving neural implicit surfaces geometry with patch warping. arXiv.org, 2112.09648, 2021. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 220, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 220, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 307, + 220, + 545, + 274 + ], + "type": "text", + "content": "[15] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 248–255, 2009. 2, 3, 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 276, + 545, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 276, + 545, + 329 + ], + "spans": [ + { + "bbox": [ + 307, + 276, + 545, + 329 + ], + "type": "text", + "content": "[16] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Hickman, Krista Reymann, Thomas B McHugh, and Vincent Vanhoucke. Google scanned objects: A high-quality dataset of 3d scanned household items. arXiv.org, 2204.11918, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 331, + 545, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 331, + 545, + 353 + ], + "spans": [ + { + "bbox": [ + 307, + 331, + 545, + 353 + ], + "type": "text", + "content": "[17] Marco Forte and François Pitié. F, b, alpha matting. arXiv.org, 2003.07711, 2020. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 355, + 547, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 355, + 547, + 410 + ], + "spans": [ + { + "bbox": [ + 307, + 355, + 547, + 410 + ], + "type": "text", + "content": "[18] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5501–5510, 2022. 3, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 411, + 545, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 411, + 545, + 454 + ], + "spans": [ + { + "bbox": [ + 307, + 411, + 545, + 454 + ], + "type": "text", + "content": "[19] Huan Fu, Rongfei Jia, Lin Gao, Mingming Gong, Binqiang Zhao, Steve Maybank, and Dacheng Tao. 3d-future: 3d furniture shape with texture. International Journal of Computer Vision (IJCV), 129(12):3313-3337, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 456, + 545, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 456, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 307, + 456, + 545, + 498 + ], + "type": "text", + "content": "[20] Matheus Gadelha, Subhransu Maji, and Rui Wang. 3d shape induction from 2d views of multiple objects. In Proceedings of the International Conference on 3D Vision (3DV), pages 402-411, 2017. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 501, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 501, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 307, + 501, + 545, + 555 + ], + "type": "text", + "content": "[21] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. arXiv.org, 2209.11163, 2022.3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 557, + 545, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 557, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 307, + 557, + 545, + 612 + ], + "type": "text", + "content": "[22] Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3d: A generative model of high quality 3d textured shapes learned from images. In Advances in Neural Information Processing Systems (NIPS), 2022. 3, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 613, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 613, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 307, + 613, + 545, + 667 + ], + "type": "text", + "content": "[23] Ankit Goyal, Hei Law, Bowei Liu, Alejandro Newell, and Jia Deng. Revisiting point cloud shape classification with a simple and effective baseline. In Proceedings of the International Conference on Machine learning (ICML), pages 3809-3820, 2021. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 670, + 545, + 712 + ], + "type": "text", + "content": "[24] Meng-Hao Guo, Jun-Xiong Cai, Zheng-Ning Liu, Tai-Jiang Mu, Ralph R Martin, and Shi-Min Hu. Pct: Point cloud transformer. Computational Visual Media, 7(2):187-199, 2021. 5" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "811" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "type": "text", + "content": "[25] Agrim Gupta, Piotr Dollar, and Ross Girshick. LVIS: A dataset for large vocabulary instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5356-5364, 2019. 2, 3, 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 287, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 172 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 172 + ], + "type": "text", + "content": "[26] Philipp Henzler, Niloy J Mitra, and Tobias Ritschel. Escaping plato's cave: 3d shape from adversarial rendering. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 9984-9993, 2019. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 173, + 287, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 173, + 287, + 238 + ], + "spans": [ + { + "bbox": [ + 48, + 173, + 287, + 238 + ], + "type": "text", + "content": "[27] Philipp Henzler, Jeremy Reizenstein, Patrick Labatut, Roman Shapovalov, Tobias Ritschel, Andrea Vedaldi, and David Novotny. Unsupervised learning of 3d object categories from videos in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4700-4709, 2021. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 239, + 287, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 239, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 48, + 239, + 287, + 293 + ], + "type": "text", + "content": "[28] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In Advances in Neural Information Processing Systems (NIPS), volume 30, 2017. 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 294, + 287, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 294, + 287, + 327 + ], + "spans": [ + { + "bbox": [ + 48, + 294, + 287, + 327 + ], + "type": "text", + "content": "[29] Moritz Ibing, Gregor Kobsik, and Leif Kobbelt. Octree transformer: Autoregressive 3d shape generation on hierarchically structured sequences. arXiv.org, 2111.12480, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 327, + 287, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 327, + 287, + 382 + ], + "spans": [ + { + "bbox": [ + 48, + 327, + 287, + 382 + ], + "type": "text", + "content": "[30] Sihyeon Kim, Sanghyeok Lee, Dasol Hwang, Jaewon Lee, Seong Jae Hwang, and Hyunwoo J Kim. Point cloud augmentation with weighted local transformations. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 548-557, 2021. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 383, + 287, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 383, + 287, + 437 + ], + "spans": [ + { + "bbox": [ + 48, + 383, + 287, + 437 + ], + "type": "text", + "content": "[31] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4. International Journal of Computer Vision (IJCV), 128(7):1956-1981, 2020. 3, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 437, + 287, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 437, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 48, + 437, + 287, + 491 + ], + "type": "text", + "content": "[32] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Proceedings of the European Conference on Computer Vision (ECCV), pages 740-755, 2014. 3, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 492, + 287, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 492, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 48, + 492, + 287, + 548 + ], + "type": "text", + "content": "[33] Liu Liu, Wenqiang Xu, Haoyuan Fu, Sucheng Qian, Qiao-jun Yu, Yang Han, and Cewu Lu. Akb-48: A real-world articulated object knowledge base. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14809-14818, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 548, + 287, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 287, + 602 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 287, + 602 + ], + "type": "text", + "content": "[34] Yongcheng Liu, Bin Fan, Shiming Xiang, and Chunhong Pan. Relation-shape convolutional neural network for point cloud analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8895-8904, 2019. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 603, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 603, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 603, + 287, + 668 + ], + "type": "text", + "content": "[35] Yuan Liu, Sida Peng, Lingjie Liu, Qianqian Wang, Peng Wang, Christian Theobalt, Xiaowei Zhou, and Wenping Wang. Neural rays for occlusion-aware image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7824-7833, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "type": "text", + "content": "[36] Xiaoxiao Long, Cheng Lin, Peng Wang, Taku Komura, and Wenping Wang. Sparseneus: Fast generalizable neural surface reconstruction from sparse views. arXiv.org, 2206.05737, 2022. 3, 7" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[37] Sebastian Lunz, Yingzhen Li, Andrew Fitzgibbon, and Nate Kushman. Inverse graphics gan: Learning to generate 3d shapes from unstructured 2d data. arXiv.org, 2002.12674, 2020.3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 118, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 173 + ], + "type": "text", + "content": "[38] Lars Mescheder, Michael Oechsle, Michael Niemeyer, Sebastian Nowozin, and Andreas Geiger. Occupancy networks: Learning 3d reconstruction in function space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4460-4470, 2019. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 175, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 239 + ], + "type": "text", + "content": "[39] Ben Mildenhall, Peter Hedman, Ricardo Martin-Brualla, Pratul P Srinivasan, and Jonathan T Barron. Nerf in the dark: High dynamic range view synthesis from noisy raw images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16190-16199, 2022. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 241, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 241, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 307, + 241, + 545, + 297 + ], + "type": "text", + "content": "[40] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In Proceedings of the European Conference on Computer Vision (ECCV), pages 405-421, 2020. 3, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 298, + 545, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 298, + 545, + 341 + ], + "spans": [ + { + "bbox": [ + 307, + 298, + 545, + 341 + ], + "type": "text", + "content": "[41] Kaichun Mo, Paul Guerrero, Li Yi, Hao Su, Peter Wonka, Niloy J. Mitra, and Leonidas J. Guibas. Structurenet: Hierarchical graph networks for 3d shape generation. arXiv.org, 1908.00575, 2019. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 342, + 545, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 342, + 545, + 386 + ], + "spans": [ + { + "bbox": [ + 307, + 342, + 545, + 386 + ], + "type": "text", + "content": "[42] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Transactions on Graphics, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 388, + 545, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 388, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 307, + 388, + 545, + 443 + ], + "type": "text", + "content": "[43] Michael Oechsle, Songyou Peng, and Andreas Geiger. Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5589-5599, 2021. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 445, + 545, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 445, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 307, + 445, + 545, + 498 + ], + "type": "text", + "content": "[44] Dario Pavllo, Jonas Kohler, Thomas Hofmann, and Aurelien Lucchi. Learning generative models of textured 3d meshes from real-world images. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 13879-13889, 2021. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 500, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 500, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 307, + 500, + 545, + 555 + ], + "type": "text", + "content": "[45] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: deep learning on point sets for 3d classification and segmentation. corr abs/1612.00593 (2016). In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 652-660, 2017. 3, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 557, + 545, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 557, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 307, + 557, + 545, + 601 + ], + "type": "text", + "content": "[46] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. In Advances in Neural Information Processing Systems (NIPS), volume 30, 2017. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 602, + 545, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 602, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 307, + 602, + 545, + 645 + ], + "type": "text", + "content": "[47] Xuebin Qin, Zichen Zhang, Chenyang Huang, Masood Dehghan, Osmar R Zaiane, and Martin Jagersand. U2-net: Going deeper with nested u-structure for salient object detection. Pattern Recognition, 106:107404, 2020. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 647, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 712 + ], + "type": "text", + "content": "[48] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 10901-10911, 2021. 2, 3" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "812" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[49] Jiawei Ren, Liang Pan, and Ziwei Liu. Benchmarking and analyzing point cloud classification under corruptions. In Proceedings of the International Conference on Machine learning (ICML), 2022. 3, 4, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 287, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 287, + 161 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 287, + 161 + ], + "type": "text", + "content": "[50] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4104-4113, 2016. 2, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 162, + 287, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 162, + 287, + 216 + ], + "spans": [ + { + "bbox": [ + 49, + 162, + 287, + 216 + ], + "type": "text", + "content": "[51] Shuai Shao, Zeming Li, Tianyuan Zhang, Chao Peng, Gang Yu, Xiangyu Zhang, Jing Li, and Jian Sun. Objects365: A large-scale, high-quality dataset for object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8430-8439, 2019. 3, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 217, + 287, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 217, + 287, + 259 + ], + "spans": [ + { + "bbox": [ + 49, + 217, + 287, + 259 + ], + "type": "text", + "content": "[52] Edward J Smith and David Meger. Improved adversarial systems for 3d object generation and reconstruction. In Proceedings of the Conference on Robot Learning (CoRL), pages 87-96, 2017. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 261, + 287, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 261, + 287, + 314 + ], + "spans": [ + { + "bbox": [ + 49, + 261, + 287, + 314 + ], + "type": "text", + "content": "[53] Stefan Stojanov, Anh Thai, and James M Rehg. Using shape to categorize: Low-shot learning with an explicit shape bias. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1798-1808, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 316, + 287, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 316, + 287, + 370 + ], + "spans": [ + { + "bbox": [ + 49, + 316, + 287, + 370 + ], + "type": "text", + "content": "[54] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5459-5469, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 371, + 287, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 371, + 287, + 415 + ], + "spans": [ + { + "bbox": [ + 49, + 371, + 287, + 415 + ], + "type": "text", + "content": "[55] Saeid Asgari Taghanaki, Jieliang Luo, Ran Zhang, Ye Wang, Pradeep Kumar Jayaraman, and Krishna Murthy Jatavallabhula. Robustpointset: A dataset for benchmarking robustness of point cloud classifiers. arXiv.org, 2011.11572, 2020.3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 416, + 287, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 416, + 287, + 480 + ], + "spans": [ + { + "bbox": [ + 49, + 416, + 287, + 480 + ], + "type": "text", + "content": "[56] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 1588–1597, 2019. 2, 3, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 482, + 287, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 482, + 287, + 546 + ], + "spans": [ + { + "bbox": [ + 49, + 482, + 287, + 546 + ], + "type": "text", + "content": "[57] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T Barron, and Pratul P Srinivasan. Ref-nerf: Structured view-dependent appearance for neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5481-5490, 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 548, + 287, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 548, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 49, + 548, + 287, + 612 + ], + "type": "text", + "content": "[58] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. In Advances in Neural Information Processing Systems (NIPS), volume 34, pages 27171-27183, 2021. 3, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 614, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 614, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 49, + 614, + 287, + 689 + ], + "type": "text", + "content": "[59] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P Srinivasan, Howard Zhou, Jonathan T Barron, Ricardo Martin-Brualla, Noah Snavely, and Thomas Funkhouser. Ibrnet: Learning multi-view image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4690-4699, 2021. 3, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 691, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 691, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 691, + 287, + 712 + ], + "type": "text", + "content": "[60] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "graph cnn for learning on point clouds. ACM Transactions on Graphics, 38(5):1-12, 2019. 3, 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 308, + 96, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 140 + ], + "type": "text", + "content": "[61] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing (TIP), 13(4):600-612, 2004. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 142, + 545, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 142, + 545, + 195 + ], + "spans": [ + { + "bbox": [ + 308, + 142, + 545, + 195 + ], + "type": "text", + "content": "[62] Jiajun Wu, Chengkai Zhang, Tianfan Xue, Bill Freeman, and Josh Tenenbaum. Learning a probabilistic latent space of object shapes via 3d generative-adversarial modeling. In Advances in Neural Information Processing Systems (NIPS), volume 29, 2016. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 198, + 545, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 198, + 545, + 241 + ], + "spans": [ + { + "bbox": [ + 308, + 198, + 545, + 241 + ], + "type": "text", + "content": "[63] Tong Wu, Jiaqi Wang, Xingang Pan, Xudong Xu, Christian Theobalt, Ziwei Liu, and Dahua Lin. Voxurf: Voxel-based efficient and accurate neural surface reconstruction. arXiv.org, 2208.12697, 2022. 3, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 243, + 545, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 243, + 545, + 308 + ], + "spans": [ + { + "bbox": [ + 308, + 243, + 545, + 308 + ], + "type": "text", + "content": "[64] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1912-1920, 2015. 2, 3, 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 310, + 545, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 310, + 545, + 364 + ], + "spans": [ + { + "bbox": [ + 308, + 310, + 545, + 364 + ], + "type": "text", + "content": "[65] Tiange Xiang, Chaoyi Zhang, Yang Song, Jianhui Yu, and Weidong Cai. Walk in the cloud: Learning curves for point clouds shape analysis. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 915–924, 2021. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 366, + 545, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 366, + 545, + 421 + ], + "spans": [ + { + "bbox": [ + 308, + 366, + 545, + 421 + ], + "type": "text", + "content": "[66] Mutian Xu, Runyu Ding, Hengshuang Zhao, and Xiaojuan Qi. Paconv: Position adaptive convolution with dynamic kernel assembling on point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3173-3182, 2021. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 422, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 422, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 308, + 422, + 545, + 487 + ], + "type": "text", + "content": "[67] Mutian Xu, Junhao Zhang, Zhipeng Zhou, Mingye Xu, Xiaojuan Qi, and Yu Qiao. Learning geometry-disentangled representation for complementary understanding of 3d object point cloud. In Proceedings of the Conference on Artificial Intelligence (AAAI), volume 35, pages 3056-3064, 2021. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 490, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 490, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 308, + 490, + 545, + 544 + ], + "type": "text", + "content": "[68] Guandao Yang, Xun Huang, Zekun Hao, Ming-Yu Liu, Serge Belongie, and Bharath Hariharan. Pointflow: 3d point cloud generation with continuous normalizing flows. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4541-4550, 2019. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 546, + 545, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 546, + 545, + 610 + ], + "spans": [ + { + "bbox": [ + 308, + 546, + 545, + 610 + ], + "type": "text", + "content": "[69] Yao Yao, Zixin Luo, Shiwei Li, Jingyang Zhang, Yufan Ren, Lei Zhou, Tian Fang, and Long Quan. Blendedmvs: A large-scale dataset for generalized multi-view stereo networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1790-1799, 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 613, + 545, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 613, + 545, + 657 + ], + "spans": [ + { + "bbox": [ + 308, + 613, + 545, + 657 + ], + "type": "text", + "content": "[70] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. In Advances in Neural Information Processing Systems (NIPS), volume 34, pages 4805-4815, 2021. 3, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 658, + 545, + 712 + ], + "type": "text", + "content": "[71] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4578-4587, 2021. 3, 5, 7" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "813" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 262 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "type": "text", + "content": "[72] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. In Advances in Neural Information Processing Systems (NIPS), 2022. 3, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 288, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 288, + 184 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 288, + 184 + ], + "type": "text", + "content": "[73] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 586-595, 2018. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 185, + 288, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 185, + 288, + 228 + ], + "spans": [ + { + "bbox": [ + 48, + 185, + 288, + 228 + ], + "type": "text", + "content": "[74] Linqi Zhou, Yilun Du, and Jiajun Wu. 3d shape generation and completion through point-voxel diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 5826-5835, 2021. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 229, + 288, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 229, + 288, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 229, + 288, + 262 + ], + "type": "text", + "content": "[75] Qian-Yi Zhou, Jaesik Park, and Vladlen Koltun. Open3D: A modern library for 3D data processing. arXiv.org, 1801.09847, 2018. 4" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "814" + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/OmniVidar_ Omnidirectional Depth Estimation From Multi-Fisheye Images/e34f49e7-aba6-408f-9647-b38f210ab7c3_content_list.json b/2023/OmniVidar_ Omnidirectional Depth Estimation From Multi-Fisheye Images/e34f49e7-aba6-408f-9647-b38f210ab7c3_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..23a63cb8e948104b245f6391e8703dea7f4c0052 --- /dev/null +++ b/2023/OmniVidar_ Omnidirectional Depth Estimation From Multi-Fisheye Images/e34f49e7-aba6-408f-9647-b38f210ab7c3_content_list.json @@ -0,0 +1,1682 @@ +[ + { + "type": "text", + "text": "OmniVidar: Omnidirectional Depth Estimation from Multi-Fisheye Images", + "text_level": 1, + "bbox": [ + 104, + 130, + 867, + 154 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sheng Xie $^{1}$ , Daochuan Wang $^{2}$ , Yunhui Liu $^{1,2,3,*}$ \nHarbin Institute of Technology, Shenzhen $^{1}$ \nHong Kong Center for Logistics Robotics $^{2}$ , The Chinese University of Hong Kong $^{3}$ \nxiesheng@stu.hit.edu.cn, wangdaochuan@hotmail.com, yhliu@mae.cuhk.edu.hk", + "bbox": [ + 155, + 179, + 815, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 286, + 313, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Estimating depth from four large field of view (FoV) cameras has been a difficult and understudied problem. In this paper, we proposed a novel and simple system that can convert this difficult problem into easier binocular depth estimation. We name this system OmniVidar, as its results are similar to LiDAR, but rely only on vision. OmniVidar contains three components: (1) a new camera model to address the shortcomings of existing models, (2) a new multi-fisheye camera based epipolar rectification method for solving the image distortion and simplifying the depth estimation problem, (3) an improved binocular depth estimation network, which achieves a better balance between accuracy and efficiency. Unlike other omnidirectional stereo vision methods, OmniVidar does not contain any 3D convolution, so it can achieve higher resolution depth estimation at fast speed. Results demonstrate that OmniVidar outperforms all other methods in terms of accuracy and performance.", + "bbox": [ + 75, + 318, + 473, + 575 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 606, + 209, + 622 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Depth estimation from images is an important research field in computer vision, as it enables the acquisition of depth information with low-cost cameras for a wide range of applications. Traditional stereo cameras with pinhole lenses are limited in their FoV. However, many scenarios require an omnidirectional depth map, such as autonomous driving [42] and robot navigation [13, 43]. Although there are active sensors available that can provide omnidirectional depth information, such as LiDAR [25], their high cost makes them less accessible than stereo vision. Passive sensors, such as RGB cameras, are a common choice for depth estimation due to their low cost, lightweight, and low power consumption. To increase the FoV, fisheye lenses are often introduced into stereo vision setups.", + "bbox": [ + 75, + 632, + 468, + 843 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Over the past few decades, various methods have been proposed for depth estimation using fisheye cameras. These", + "bbox": [ + 76, + 845, + 468, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fa083f2bc73257d8f36f48ba7acb606040fcee25bc53124d9d7470654d4883f3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 285, + 692, + 414 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/18c7381f79988b5a4fe9a0e18fe7543499c067769a1eebdfc01ce33e36c8c321.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 285, + 890, + 414 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/268fff970f211c87e704f9b23d693668f8f8ac2cd30a504e4618a6190ca45e17.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 414, + 890, + 501 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b412a97201711db98241d7682ddc02c6d040baeb269d64fd1bd176fcc1137144.jpg", + "image_caption": [ + "Figure 1. Our prototype that built with four $250^{\\circ}$ fisheye cameras and its results of dense inverse distance map and cloud points. It can get great depth estimation results in real scenes and achieve real-time performance on modern GPU." + ], + "image_footnote": [], + "bbox": [ + 501, + 500, + 890, + 638 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "include the binocular fisheye system [31], up-down fisheye system [13], and catadiptric cameras [17, 23, 32]. However, all of these approaches have limitations. The binocular fisheye system cannot provide an omnidirectional perception. The up-down fisheye system and catadiptric cameras can offer horizontal $360^{\\circ}$ depth perception, but their vertical FoV is limited. Furthermore, catadiptric cameras tend to be bulky, challenging to calibrate, and prone to errors. It turns out that the best choice for omnidirectional depth estimation is a system consisting of four cameras with extremely wide FoV ( $>180^{\\circ}$ ). This system enables", + "bbox": [ + 496, + 734, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 94, + 886, + 220, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "21529", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$360^{\\circ}$ depth estimation both horizontally and vertically, and is light-weight, convenient to calibrate and maintain. Several studies [21,29,39,41] have shown excellent results with this approach. However, these methods often use many 3D convolutions, resulting in low efficiency. To adapt to limited memory, the images must be down-sampled, leading to a lower resolution of the output depth map. Furthermore, many of these approaches perform depth estimation using the original fisheye image without explicitly processing image distortion, which leads to the construction of a tedious and complicated cost volume and can result in reduced accuracy.", + "bbox": [ + 75, + 90, + 472, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Moreover, due to the complex imaging process of large FoV fisheye lenses, handling fisheye image distortion using mathematical formulas can be a challenging task. While several excellent large FoV fisheye camera models have been proposed [2, 18, 19, 27, 30, 36], our experiments have shown that none of these models can accurately approximate the imaging process, and we get less satisfactory depth estimation results when using them.", + "bbox": [ + 75, + 273, + 472, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Inspired by the above observations, we propose OmniVidar, a novel and simple multi-fisheye omnidirectional depth estimation system, as shown in Figure 2. OmniVidar contains three components. Firstly, we improve the DSCM [36] and propose a new camera model, named Triple Sphere Camera Model (TSCM), which can better approximate the imaging process and achieve the best accuracy. Then we propose an epipolar rectification algorithm designed for multi-fisheye camera system. We leverage a cubic-like projection approach to transform the four fisheye camera systems into four binocular camera systems and then conduct epipolar rectification on each binocular system. This method solves the distortion issue and reduces the complex multi-fisheye omnidirectional depth estimation problem to a much simpler binocular depth estimation problem. Finally, we design a lightweight binocular depth estimation network based on RAFT [35]. We add Transformer encoder [37] into feature extraction in RAFT to combine the advantages of Transformer and GRU, and it's easy to balance the accuracy and efficiency.", + "bbox": [ + 75, + 395, + 472, + 696 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We compare OmniVidar with existing methods on several datasets. The results demonstrate that our method outperforms all the others in terms of speed and accuracy, achieving State-of-The-Art performance.", + "bbox": [ + 75, + 698, + 470, + 758 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 773, + 218, + 789 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Multiview Fisheye Stereo", + "text_level": 1, + "bbox": [ + 76, + 800, + 307, + 816 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Gao et al. [13] mount two large FoV cameras $(245^{\\circ})$ vertically and reversely to obtain a $360^{\\circ}$ horizontal and $65^{\\circ}$ vertical FoV overlap and use this system in UAV for depth perception [43]. Won et al. [40] use four cameras with $220^{\\circ}$ FoV and propose SweepNet algorithm. They propose", + "bbox": [ + 75, + 824, + 470, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Sphere-Sweep method modified from Plane-Sweep [16] to construct the cost volume and then use the cost aggregation algorithm in SGM [15] to obtain depth. After this, Won et al. propose OmniMVS [39, 41], which uses 3D convolution instead of SGM for depth regression and achieves higher accuracy. Komatsu et al. [21] propose CrownConv360. They improve OmniMVS by using the icosahedral projection in the feature extraction and depth regression stage for distortion reduction. Instead of using neural network, Meuleman et al. [29] choose to investigate cost aggregation and propagation in traditional stereo matching. And they achieve real-time high-resolution depth estimation at 29 fps on the embedded device TX2. However, the algorithm is only suitable for short camera baseline.", + "bbox": [ + 496, + 90, + 893, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "All of the above works have more or fewer deficiencies. We propose OmniVidar, an omnidirectional depth estimation method with high accuracy, which can run in real-time and solve the fisheye image distortion problem with good generalization in various scenes.", + "bbox": [ + 496, + 303, + 893, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Fisheye Image Distortion", + "text_level": 1, + "bbox": [ + 500, + 391, + 728, + 407 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To solve the distortion problem, a camera model suitable for fisheye is required, and the distortion should be handled explicitly in depth estimation.", + "bbox": [ + 496, + 415, + 890, + 460 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Scaramuzzi et al. [30] propose a high-order polynomial-based camera projection model, which is highly generalizable and can be widely used for various catadioptric and fisheye cameras. But it has no closed-form solution, so the inverse projection equation needs to be fitted with a high-order polynomial, which introduces errors. Mei et al. [27] propose the MUCM, which is an improvement of the UCM [2] to achieve higher calibration accuracy and propose a novel intrinsic parameters initialization method. Usenko et al. [36] propose the DSCM, which has only one more intrinsic parameter than the MUCM model, but the accuracy is substantially improved.", + "bbox": [ + 496, + 460, + 893, + 643 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Su et al. [34] propose a mini-network that learns to modify the shape of filters based on location. Coors et al. [8] chose to use a camera projection model to shape the convolution kernel to directly address distortion. An alternative idea was chosen by Cohen et al. [7] who define a form of convolution on the Lie group $SO(3)$ that convolves over the rotation space rather than the translation space. Eder et al. [9] compared the performance of three approaches: equirectangular projection, cubic projection, and 7-level subdivided icosahedral projection, on depth estimation and semantic segmentation tasks. The results showed that icosahedral projection achieved the best results in all tasks. Komatsu et al. [21] used icosahedral projection in multi-fisheye omnidirectional stereo vision to improve the depth estimation results of OmniMVS. Eder et al. [10] proposed a tangent plane projection modified from the icosahedron projection.", + "bbox": [ + 496, + 643, + 893, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "21530", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/48c296a724e0d882bd1f0f98b4f5061e8fa03c27f3d291ac386dfef14525d566.jpg", + "image_caption": [ + "Figure 2. System overview. Our system contains three parts: multi-fisheye calibration, multi-fisheye epipolar rectification, and binocular depth estimation. In epipolar rectification, the images in the left column are original. The images in the middle column are the images after abstracting each fisheye camera into two pinhole cameras, where the two images connected by a red dotted arrow are a pair of binocular cameras. The images are rearranged and combined to obtain the right column." + ], + "image_footnote": [], + "bbox": [ + 76, + 80, + 890, + 292 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The above algorithms have a common feature that they are all designed for the monocular image. Thus they perform well on tasks that require only one camera, yet not much good for multi-fisheye tasks. Therefore, to address this problem, we propose a distortion removal method designed for multi-fisheye systems, which greatly simplifies the depth estimation problem at the same time.", + "bbox": [ + 75, + 383, + 472, + 491 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. RAFT in Stereo Vision", + "text_level": 1, + "bbox": [ + 76, + 501, + 289, + 515 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Zachary et al. [35] propose Recurrent All-Pairs Field Transforms (RAFT), a highly efficient and accurate deep network architecture for optical flow. RAFT uses modified GRU blocks to iteratively refine optical flow and shares weights between refinement units. Lahav et al. [24] apply the RAFT to binocular rectified stereo vision and introduce multi-level convolutional GRUs, which can more efficiently propagate information across the image. Wang et al. [38] modify RAFT and propose PVStereo which greatly outperforms other self-supervised stereo matching approaches. Li et al. [22] combine RAFT and Transformer and propose CREStereo. They add Transformer encoder in image feature extraction and repeat the cross-attention between left and right images in the iterative update stage. This method performs well in various binocular datasets with excellent generalization.", + "bbox": [ + 75, + 525, + 468, + 767 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We imitate CREStereo and improve it based on RAFT to obtain a binocular stereo vision depth estimation network with a better trade-off between efficiency and accuracy.", + "bbox": [ + 75, + 768, + 468, + 814 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methods", + "text_level": 1, + "bbox": [ + 76, + 828, + 174, + 844 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The overall flow of our method is shown in Figure 2 and is divided into three parts: camera calibration, epipolar rectification, and depth estimation.", + "bbox": [ + 75, + 854, + 468, + 902 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1718c5435412406974cc383527dbfba4d4f2145542d5570cf5c08cd8e0f339f3.jpg", + "image_caption": [ + "Figure 3. Our Triple Sphere Camera Model. First, point $P$ is projected onto the first sphere and then onto the second sphere, then the third. The second sphere is shifted w.r.t. the first sphere by $\\xi$ , and the third is shifted w.r.t. the second sphere by $\\lambda$ . Finally, the point is projected onto the image plane of a pinhole camera which is shifted by $\\frac{\\alpha}{1 - \\alpha}$ from the third sphere." + ], + "image_footnote": [], + "bbox": [ + 540, + 383, + 857, + 696 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Camera Calibration", + "text_level": 1, + "bbox": [ + 500, + 842, + 692, + 857 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We propose the Triple Sphere Camera Model, which can better fit the imaging process of large FoV fisheye cameras", + "bbox": [ + 498, + 869, + 892, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "21531", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and has a closed-form unprojection solution.", + "bbox": [ + 76, + 90, + 372, + 104 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Figure 3, our projection model considers that the incident light is refracted three times, and the displacements of the three unit spherical centers are $\\xi$ and $\\lambda$ . After three refractions, the incident light is finally projected to the image plane according to the pinhole camera model, and the displacement of the pinhole camera model's optical center from the third unit sphere is $\\frac{\\alpha}{1 - \\alpha}$ . Therefore, our model has totally 7 camera internal parameters: $f_{x}, f_{y}, c_{x}, c_{y}, \\xi, \\lambda, \\alpha$ . The projection equations are defined as follows.", + "bbox": [ + 75, + 106, + 468, + 256 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\pi (P, i) = \\frac {1}{\\zeta} \\left[ \\begin{array}{c c c} f _ {x} & 0 & c _ {x} \\\\ 0 & f _ {y} & c _ {y} \\end{array} \\right] \\left[ \\begin{array}{l} X \\\\ Y \\\\ \\zeta \\end{array} \\right] \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 125, + 263, + 468, + 314 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\zeta = Z + \\xi d _ {1} + \\lambda d _ {2} + \\frac {\\alpha}{1 - \\alpha} d _ {3} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 315, + 468, + 340 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nd _ {1} = \\sqrt {X ^ {2} + Y ^ {2} + Z ^ {2}} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 344, + 468, + 364 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nd _ {2} = \\sqrt {X ^ {2} + Y ^ {2} + (\\xi d _ {1} + Z) ^ {2}} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 367, + 468, + 386 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nd _ {3} = \\sqrt {X ^ {2} + Y ^ {2} + (\\lambda d _ {2} + \\xi d _ {1} + Z) ^ {2}} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 388, + 468, + 407 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $i$ is the vector of intrinsic parameters, $\\pi$ is the projection function.", + "bbox": [ + 76, + 419, + 467, + 446 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A set of 3D points that results in valid projection is expressed as follows:", + "bbox": [ + 76, + 449, + 467, + 478 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Omega = \\left\\{x \\in \\mathbb {R} ^ {3} | z > - w _ {2} d _ {1} \\right\\} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 489, + 468, + 508 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nw _ {2} = \\frac {\\xi + \\lambda + w _ {1}}{\\sqrt {1 + (\\xi + \\lambda) ^ {2} + 2 w _ {1} (\\xi + \\lambda)}} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 510, + 468, + 545 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nw _ {1} = \\frac {\\alpha}{1 - \\alpha} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 547, + 468, + 573 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The unprojection function is computed as follows:", + "bbox": [ + 96, + 583, + 429, + 598 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\pi^ {- 1} (p, i) = \\mu \\left[ \\begin{array}{l} \\eta \\gamma x \\\\ \\eta \\gamma y \\\\ m _ {z} \\end{array} \\right] - \\left[ \\begin{array}{l} 0 \\\\ 0 \\\\ \\xi \\end{array} \\right] \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 608, + 468, + 657 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mu = \\xi m _ {z} + \\sqrt {\\xi^ {2} m _ {z} ^ {2} - \\xi^ {2} + 1} \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 660, + 467, + 678 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nm _ {z} = \\eta (\\gamma - \\zeta) - \\lambda \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 137, + 681, + 468, + 696 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\eta = \\lambda (\\gamma - \\zeta) + \\sqrt {\\lambda^ {2} (\\gamma - \\zeta) ^ {2} - \\lambda^ {2} + 1} \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 700, + 468, + 719 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\gamma = \\frac {\\zeta + \\sqrt {1 + (1 - \\zeta^ {2}) \\left(x ^ {2} + y ^ {2}\\right)}}{x ^ {2} + y ^ {2} + 1} \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 723, + 468, + 756 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\zeta = \\left\\{ \\begin{array}{l l} \\frac {\\alpha}{1 - \\alpha} & \\text {i f} \\alpha \\leq 0. 5 \\\\ \\frac {1 - \\alpha}{\\alpha} & \\text {i f} \\alpha > 0. 5 \\end{array} \\right. \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 758, + 468, + 800 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $(x,y)$ is the normalized coordinate.", + "bbox": [ + 76, + 810, + 356, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We calibrate the intrinsic and extrinsic parameters of multi-fisheye camera system using a highly accurate planar checkerboard. The goal of the calibration is to minimize the reprojection error of the corner points in all pictures. For each image $n$ in the calibration sequence, the", + "bbox": [ + 75, + 825, + 468, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5be166fd2ac635187db3de74983f1bad2c25fc0222d0602184ff9cbd1aca0070.jpg", + "image_caption": [ + "Figure 4. The multi-fisheye epipolar rectification schematic. Each fisheye camera is converted into two pinhole cameras with an angle of $90^{\\circ}$ . Then the system becomes four binocular systems." + ], + "image_footnote": [], + "bbox": [ + 504, + 87, + 890, + 229 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "corner detector can obtain the projection point $u_{nk}$ of the $k$ th corner point $x_{k}$ . The coordinate of $u_{nk}$ is related to the camera intrinsic and extrinsic parameters. Let $s = [i, T_{cam_0}, T_{cam_1}, \\dots, T_{cam_n}]$ be the parameter to optimize. We can construct the nonlinear optimization problem as follows:", + "bbox": [ + 498, + 316, + 890, + 406 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ns = \\arg \\min _ {s ^ {\\prime}} \\sum_ {n = 0} ^ {N} \\sum_ {k \\in K} \\rho \\left(\\left(\\pi \\left(T _ {c a m _ {n}} x _ {k}, \\boldsymbol {i}\\right) - u _ {n k}\\right) ^ {2}\\right) \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 516, + 415, + 890, + 457 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $T_{cam_n} \\in SE(3)$ is the transformation from the coordinate frame of the calibration grid to the camera coordinate frame for image $n$ . $K$ is a set of detected corner points for the image $n$ and $\\rho$ is the robust Huber norm.", + "bbox": [ + 498, + 465, + 890, + 527 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Due to the highly nonlinear nature of the image, the above optimization problem requires a good initial value. We initialize the intrinsic parameters using the previously proposed method [27] and find initial poses using the UPnP algorithm [20]. After obtaining the initial values of the parameters, we completed the optimization operation using Ceres Solver [1].", + "bbox": [ + 498, + 527, + 890, + 632 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Multi-Fisheye Camera Epipolar Rectify", + "text_level": 1, + "bbox": [ + 500, + 641, + 839, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We propose a simple and effective undistortion method that cleverly transforms the omnidirectional depth estimation problem into a binocular depth estimation problem. And the method can be used in the system with any number of fisheye cameras. As shown in Figure 4, we abstract each fisheye camera as two pinhole cameras with $90^{\\circ}$ angle1. Then the four fisheye cameras system becomes four independent binocular systems. Each binocular systems can be rectified through epipolar rectification. We modified the classical epipolar rectification algorithm proposed by Fusiello [11] for our system.", + "bbox": [ + 496, + 664, + 890, + 830 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "1Although here we only show the horizontal binoculars, it is possible to construct binoculars facing both above and below to estimate the true omnidirectional depth. However, we consider that the above and below depth values are rarely used in practical applications. So we just use a limited $110^{\\circ}$ vertical FoV, which satisfies most of the application requirements.", + "bbox": [ + 500, + 838, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "21532", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d47d6dd39f947d590957077f36f3647181aa27b3741cb84ea7d73ba868e33b12.jpg", + "image_caption": [ + "Figure 5. The structure of our proposed binocular depth estimation network. A pair of stereo images are fed into feature extraction networks. Then the output features are calculated by self-attention and cross-attention, and after doing self-attention, the features of left image are used for context information for latter recurrent update stage. Then the left and right features are multiplied to construct cost volume. Finally use the disparity update model same as RAFT [35]." + ], + "image_footnote": [], + "bbox": [ + 76, + 85, + 893, + 330 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Firstly, the rotation matrix of the cameras in binocular system is calculated. Let the optical centers of the left camera and right camera be $C_1 = \\left[x_1 y_1 z_1\\right]^T$ , $C_2 = \\left[x_2 y_2 z_2\\right]^T$ , the rotation matrix can be calculated as follows:", + "bbox": [ + 76, + 434, + 468, + 515 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nR = \\left[ \\begin{array}{l l l} r _ {1} ^ {T} & r _ {2} ^ {T} & r _ {3} ^ {T} \\end{array} \\right] \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 526, + 467, + 542 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nr _ {1} = \\left[ \\begin{array}{l l l} x _ {2} - x _ {1} & y _ {2} - y _ {1} & z _ {2} - z _ {1} \\end{array} \\right]. n o r m a l i z e () \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 547, + 467, + 564 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nr _ {3} = \\left[ - \\left(z _ {2} - z _ {1}\\right) \\quad 0 \\quad x _ {2} - x _ {1} \\right]. n o r m a l i z e () \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 566, + 467, + 584 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nr _ {2} = r _ {3} \\times r _ {1} \\tag {19}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 588, + 467, + 601 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then remap the image for epipolar rectification. Taking the left camera in one binocular system as an example. Let the rotation matrix of the corresponding fisheye camera be $R$ , the optical center be $C$ . For a certain spatial point $P = (X, Y, Z, 1)$ , its projection point on normalized plane is $p = (u, v, 1)$ . The projection equation is shown below.", + "bbox": [ + 76, + 613, + 468, + 704 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mu \\left[ \\begin{array}{c} u \\\\ v \\\\ f (u, v) \\end{array} \\right] = [ R - R C ] \\left[ \\begin{array}{l} X \\\\ Y \\\\ Z \\\\ 1 \\end{array} \\right] \\tag {20}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 713, + 467, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Let the rotation matrix of the rectified pinhole camera be $R'$ , and the corresponding normalized plane coordinate be $(u', v', 1)$ . Then the projection equation is shown below.", + "bbox": [ + 76, + 785, + 468, + 830 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mu^ {\\prime} \\left[ \\begin{array}{l} u ^ {\\prime} \\\\ v ^ {\\prime} \\\\ 1 \\end{array} \\right] = \\left[ \\begin{array}{l l} R ^ {\\prime} & - R ^ {\\prime} C \\end{array} \\right] \\left[ \\begin{array}{l} X \\\\ Y \\\\ Z \\\\ 1 \\end{array} \\right] \\tag {21}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 840, + 467, + 905 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The relationship between fisheye image and virtual binocular image can be obtained, see equation (22).", + "bbox": [ + 500, + 434, + 890, + 464 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\lambda \\left[ \\begin{array}{l} u ^ {\\prime} \\\\ v ^ {\\prime} \\\\ 1 \\end{array} \\right] = \\left[ \\begin{array}{l l l} R ^ {\\prime} & - R ^ {\\prime} C \\end{array} \\right] \\left[ \\begin{array}{l l} R & - R C \\end{array} \\right] ^ {- 1} \\left[ \\begin{array}{c} u \\\\ v \\\\ f (u, v) \\end{array} \\right] \\tag {22}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 469, + 890, + 518 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "That is, for each pixel in the original fisheye image, the pixel coordinate in the pinhole image can be calculated. We can use pixel interpolation to obtain the rectified image.", + "bbox": [ + 498, + 522, + 890, + 568 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Figure 2, it can be seen that after epipolar rectification, the four fisheye images are converted into a pair of binocular images, where each image is stitched together from the four virtual pinhole camera images. A simple binocular depth estimation algorithm is then used.", + "bbox": [ + 498, + 569, + 890, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Stereo Depth Estimation", + "text_level": 1, + "bbox": [ + 500, + 650, + 723, + 666 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since our system needs to estimate the depth values of four binocular systems simultaneously, which is very computationally intensive, our primary goal is not to design a binocular algorithm with the highest accuracy, but to minimize the accuracy loss while ensuring real-time performance.", + "bbox": [ + 498, + 674, + 890, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The structure of our binocular depth estimation network is shown in Figure 5. We added Transformer [37] with self-attention and cross-attention in the image feature extraction stage. In binocular depth estimation, limited by computational performance and graphics card memory, many algorithms have to limit the matching search range of pixels. In this paper, we design an ingenious cost-volume calculation that can extend the estimable disparity range to the image width, while also reducing the computational complexity.", + "bbox": [ + 498, + 763, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "21533", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d0a7dd8fefda34d19958a3fd25d587fa57b38cdefe80187c3c556751ea4a75bc.jpg", + "image_caption": [ + "Figure 6. The depth estimation comparison between ours and other methods. The final row is the results from our method, and the column annotations indicate the dataset source." + ], + "image_footnote": [], + "bbox": [ + 83, + 85, + 874, + 429 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Given two images $I_1 \\in \\mathbb{R}^{H \\times W \\times C}$ , $I_2 \\in \\mathbb{R}^{H \\times W \\times C}$ , we need to compute the cost volume $C(I_1, I_2) \\in \\mathbb{R}^{H \\times W \\times W}$ . We found that the cost volume can be obtained by single matrix operation as follows.", + "bbox": [ + 76, + 494, + 468, + 555 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nC \\left(I _ {1}, I _ {2}\\right) = I _ {1} \\in \\mathbb {R} ^ {H \\times W \\times C} \\cdot I _ {2} ^ {T} \\in \\mathbb {R} ^ {H \\times C \\times W} \\tag {23}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 559, + 468, + 580 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To enable the network to learn information at multiple scales, we used the cost volume pooling opertaion in RAFT-Stereo [24].", + "bbox": [ + 76, + 584, + 468, + 630 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 641, + 209, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Implementation", + "text_level": 1, + "bbox": [ + 76, + 665, + 235, + 681 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our depth estimation network structure is shown in Figure 5. The number of channels in our network is identical to RAFT [35]. The feature encoder consists of 6 residual blocks, 2 at $1/2$ resolution, 2 at $1/4$ resolution, and 2 at $1/8$ resolution. The number of channels is 64, 128, and 256, respectively. The $1/8$ resolution feature map is then fed into the Transformer encoder [37]. We set the number of heads in attention to 8. The GRU iterative update part is identical to RAFT, and more details can be found in [35]. In addition, we replace all the normalization operators in RAFT with domain normalize [44].", + "bbox": [ + 75, + 688, + 468, + 854 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Since we convert the multi-fisheye cameras to four binocular cameras, we can use a large number of public binocular datasets for training, instead of multi-fisheye", + "bbox": [ + 75, + 854, + 470, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "datasets, which are very limited. Similar to the training strategy in CREStereo [22], we collected most of public binocular datasets, including SceneFlow [26], Sintel [4], CREStereo [22], KITTI [14, 28], InStereo2K [3], and Virtual KITTI [12]. The images are uniformly randomly cropped with a fixed resolution during the training process to solve the problem of different image resolutions in different datasets.", + "bbox": [ + 496, + 494, + 890, + 616 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We use a single NVIDIA RTX2080Ti for training, the batch size is set to 4, the learning rate is 0.0004, and we use Adam to train 200 epochs to get the final network weights.", + "bbox": [ + 496, + 617, + 890, + 662 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To test the performance of OmniVidar in real scenarios, we built a test setup. We used four leopard LI-USB30-OV10640-490-GMSL cameras with an image resolution of $1280 \\times 1024$ , paired with a fisheye lens with $250^{\\circ}$ FoV. The four cameras look in four directions, front and back, left and right, as shown in Figure 1. We use the device to capture indoor, corridor, and outdoor images, which can fully test the performance of OmniVidar in real scenes.", + "bbox": [ + 496, + 664, + 892, + 785 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Evaluation", + "text_level": 1, + "bbox": [ + 500, + 800, + 620, + 814 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We perform the comparison on the Omnihouse, OmniThings provided by Won et al. [40, 41], and the datasets provided by Meuleman et al. [29]. Our OmniVidar has only been trained on the binocular dataset mentioned in Section 4.1, not on the test dataset, and the methods involved in the", + "bbox": [ + 496, + 825, + 890, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "21534", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/35722dae7b36dfbbd41a702052c48a26270a272d6feffb11a605817f4693ea0e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodMeuleman et al. [29]Omnihouse [40]Omnithings [39]
bad 1.0RMSEMAEbad 1.0RMSEMAEbad 1.0RMSEMAE
OmniVidar(Ours)18.4990.9210.8034.8861.0021.04211.1990.9580.461
OmniMVS-ft [41]60.3722.3061.73226.3570.9290.70559.7532.2701.723
CrownConv360 [21]69.1194.0203.00585.55911.0776.72993.9469.5456.518
Meuleman et al. [29]44.1111.6641.27311.3040.6710.41065.3482.5501.941
", + "bbox": [ + 122, + 88, + 848, + 196 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "comparison all use the training method corresponding to the original paper.", + "bbox": [ + 75, + 247, + 468, + 277 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As can be seen in the Figure 6, although OmniVidar is not trained and fine-tuned on the above datasets, its depth estimation accuracy on these datasets far exceeds that of other methods, with accurate depth estimation in weakly textured regions.", + "bbox": [ + 75, + 277, + 470, + 354 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/63da8119d156432abeb078c8cc01f89e1c13774c748ad0ebe23efa11120b4e48.jpg", + "table_caption": [ + "Table 1. Quantitative comparison of algorithm accuracy on several datasets. The less the number is, the more accurate the method is." + ], + "table_footnote": [], + "table_body": "
MethodRuntimeMemoryParameters
OmniVidar (Ours)66 ms3.7 GB4.8 MB
OmniMVS-ft [41]110 ms6.8 GB43.7 MB
CrownConv360 [21]623 ms6.3 GB16.7 MB
", + "bbox": [ + 78, + 367, + 478, + 445 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The analysis of the quantitative results are shown in the Table 1 and Table 2. It can be seen that OmniVidar has the least memory consumption, the shortest time consumption and the highest accuracy compared to other deep learning based solutions. Compared with non-deep learning methods, the accuracy is much higher than it, despite the relatively low efficiency.", + "bbox": [ + 75, + 486, + 468, + 592 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/85cf9a42c4ee075faa9e176a0cb7ed10e703a63be27c925a84ed1d61cf06ca28.jpg", + "image_caption": [ + "Figure 7. Plot comparing the accuracy and efficiency evaluated in KITTI-2015 [28]. It can be seen that OmniVidar achieves real-time performance while maintain the accuracy." + ], + "image_footnote": [], + "bbox": [ + 81, + 604, + 465, + 698 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also compared the binocular depth estimation network in OmniVidar with other binocular depth estimation networks. We chose the KITTI-2015 dataset [28] for the evaluation and the results are shown in Figure 7. We compare our method with CREStereo [22], LEAStereo [6], PSMNet [5], EdgeStereo [33], RAFT-Stereo [24]. It can see that our scheme achieves a better balance between accuracy and efficiency.", + "bbox": [ + 75, + 780, + 470, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation", + "text_level": 1, + "bbox": [ + 500, + 246, + 602, + 261 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Camera Model: We compare the effect of omnidirectional depth estimation under several camera models.", + "bbox": [ + 498, + 270, + 890, + 301 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/31cd0fae314bb9ccffb1da0be9141bb5eb358bb1ab89f5367e316df0a8556e76.jpg", + "image_caption": [ + "Original fisheye image" + ], + "image_footnote": [], + "bbox": [ + 501, + 311, + 697, + 412 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/890c6965c93dcc230f404a60d00e96a5f7049674c33fd80d7b8c0fc18e6fa802.jpg", + "image_caption": [ + "Scaramuzza et al. [30]" + ], + "image_footnote": [], + "bbox": [ + 697, + 311, + 890, + 412 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5e3637eb032df4968b81aff04cae972c20c394d0371b66f2206e6a24a1bbd77f.jpg", + "image_caption": [ + "DSCM [36]", + "Figure 8. Depth estimation results using different camera projection models. Our TSCM can best remove the fisheye distortion and obtain the most accurate result." + ], + "image_footnote": [], + "bbox": [ + 501, + 428, + 699, + 531 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/87051fe40c3284da5f01fc05f5c0db868ca32308c169476eb8b97136391e35a1.jpg", + "image_caption": [ + "TSCM(Ours)" + ], + "image_footnote": [], + "bbox": [ + 697, + 428, + 890, + 531 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Figure 8, the distortion of the fisheye lens is not sufficiently handled in the DSCM [36] and the model proposed by Scaramuzza et al. [30], which affects the accuracy of depth estimation and leads to serious distortion of objects in the point cloud. In contrast, the quality of the point cloud obtained by the TSCM is much better, with flat wall surfaces and more accurate distance measurements.", + "bbox": [ + 496, + 616, + 890, + 722 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/878c214fba76d6b84ea654b595f2808e222e9619f9ee6f01e11edbf5084cf894.jpg", + "table_caption": [ + "Table 2. Efficiency comparison with other deep learning methods." + ], + "table_footnote": [], + "table_body": "
FoV2 sphere3 sphere (Ours)4 sphere
195°0.1530 px0.1528 px0.1528 px
220°0.2031 px0.1970 px0.1952 px
250°0.1365 px0.1255 px0.1222 px
", + "bbox": [ + 521, + 733, + 870, + 811 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Mean reprojection error for camera models with different number sphere.", + "bbox": [ + 498, + 820, + 890, + 849 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also compared the accuracy with different number of the refract sphere. Table 3 shows the results. It can be seen", + "bbox": [ + 498, + 869, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "21535", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1b4dbb8e5fda9efd61d3eb358a2ebefc510f5f02bf2bd4b3136c4c036ba3e295.jpg", + "image_caption": [ + "Figure 9. The images captured with our prototype and the inverse depth map and cloud points obtained by OmniVidar. It can be seen that OmniVidar has great generalization in various scenes, and performs well in low-textured areas." + ], + "image_footnote": [], + "bbox": [ + 76, + 85, + 870, + 549 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "that the camera model with 3 sphere can significantly improve the accuracy in large FoV scenes, while the improvement brought by more sphere is very limited.", + "bbox": [ + 75, + 626, + 468, + 672 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Distortion Handle: To verify whether the improvement of depth estimation accuracy is helped by removing the fisheye image distortion using epipolar rectification, we estimate the depth directly using the original fisheye image and compare it with the undistorted method. Unlike the undistorted binocular stereo, we use the Sphere-Sweep algorithm [40] to construct the cost volume. This step does not contain trainable parameters, so it does not affect the network itself and ensures the fairness of the comparison.", + "bbox": [ + 75, + 672, + 468, + 808 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The point cloud visualization of the results of the depth estimation of both is shown in Figure 8. It can be seen that when using original image, the depth information obtained from different cameras is not consistent. In contrast, the point cloud obtained by depth estimation after epipolar rectification shows the structural framework of the scene real-", + "bbox": [ + 75, + 810, + 468, + 898 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "istrically and accurately.", + "bbox": [ + 500, + 626, + 656, + 642 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Test in Real", + "text_level": 1, + "bbox": [ + 500, + 652, + 627, + 667 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We test the performance of the OmniVidar algorithm on a real dataset. Figure 9 shows the inverse depth maps and point clouds obtained by OmniVidar on the real dataset. It can be seen that our network has a strong generalization to obtain high-quality point clouds on real datasets.", + "bbox": [ + 496, + 676, + 890, + 753 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 768, + 617, + 784 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a novel, simple and effective system OmniVidar, to address the inefficiency of current omnidirectional depth estimation methods due to the extensive use of 3D convolution and the poor accuracy due to the lack of explicit handling of distortion in fisheye images. Our method outperforms all other methods in terms of time consumption and accuracy.", + "bbox": [ + 496, + 795, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "21536", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Sameer Agarwal, Keir Mierle, and The Ceres Solver Team. Ceres solver. http://ceres-solver.org. 4", + "[2] Simon Baker and Shree K. Nayar. A theory of single-viewpoint catadioptric image formation. Int. J. Comput. Vision, 35(2):175-196, nov 1999. 2", + "[3] Wei Bao, Wei Wang, Yuhua Xu, Yulan Guo, Siyu Hong, and Xiaohu Zhang. Instereo2k: a large real dataset for stereo matching in indoor scenes. Science China Information Sciences, 63(11):212101, 2019. 6", + "[4] D. J. Butler, J. Wulff, G. B. Stanley, and M. J. Black. A naturalistic open source movie for optical flow evaluation. In A. Fitzgibbon et al. (Eds.), editor, European Conf. on Computer Vision (ECCV), Part IV, LNCS 7577, pages 611-625. Springer-Verlag, Oct. 2012. 6", + "[5] Jia-Ren Chang and Yong-Sheng Chen. Pyramid stereo matching network. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5410-5418, 2018. 7", + "[6] Xuelian Cheng, Yiran Zhong, Mehrtash Harandi, Yuchao Dai, Xiaojun Chang, Tom Drummond, Hongdong Li, and Zongyuan Ge. Hierarchical neural architecture search for deep stereo matching. arXiv preprint arXiv:2010.13501, 2020.7", + "[7] Taco S. Cohen, Mario Geiger, Jonas Kohler, and Max Welling. Spherical CNNs. In International Conference on Learning Representations, 2018. 2", + "[8] Benjamin Coors, Alexandru Paul Condurache, and Andreas Geiger. Spherenet: Learning spherical representations for detection and classification in omnidirectional images. In Vittorio Ferrari, Martial Hebert, Cristian Sminchisescu, and Yair Weiss, editors, Computer Vision – ECCV 2018, pages 525–541. Springer International Publishing, 2018. 2", + "[9] Marc Eder and Jan-Michael Frahm. Convolutions on spherical images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, June 2019. 2", + "[10] Marc Eder, Mykhailo Shvets, John Lim, and Jan-Michael Frahm. Tangent images for mitigating spherical distortion. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12423-12431, 2020. 2", + "[11] Andrea Fusiello, Emanuele Trucco, and Alessandro Verri. A compact algorithm for rectification of stereo pairs. Machine Vision and Applications, 12(1):16-22, 2000. 4", + "[12] Adrien Gaidon, Qiao Wang, Yohann Cabon, and Eleonora Vig. Virtualworlds as proxy for multi-object tracking analysis. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4340-4349, 2016. 6", + "[13] Wenliang Gao and Shaojie Shen. Dual-fisheye omnidirectional stereo. In 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 6715-6722, 2017. 1, 2", + "[14] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, pages 3354-3361, 2012. 6" + ], + "bbox": [ + 78, + 114, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Heiko Hirschmuller. Stereo processing by semiglobal matching and mutual information. IEEE Transactions on Pattern Analysis and Machine Intelligence, 30(2):328-341, 2008. 2", + "[16] Christian Häne, Lionel Heng, Gim Hee Lee, Alexey Sizov, and Marc Pollefeys. Real-time direct dense matching on fisheye images using plane-sweeping stereo. In 2014 2nd International Conference on 3D Vision, volume 1, pages 57-64, 2014. 2", + "[17] Carlos Jaramillo, Roberto G. Valenti, Ling Guo, and Jizhong Xiao. Design and analysis of a single—camera omnistereo sensor for quadrotor micro aerial vehicles (mavs). Sensors, 16(2), 2016. 1", + "[18] J. Kannala and S.S. Brandt. A generic camera model and calibration method for conventional, wide-angle, and fish-eye lenses. IEEE Transactions on Pattern Analysis and Machine Intelligence, 28(8):1335-1340, 2006. 2", + "[19] Bogdan Khomutenko, Gaetan Garcia, and Philippe Martinet. An enhanced unified camera model. IEEE Robotics and Automation Letters, 1(1):137-144, 2016. 2", + "[20] Laurent Kneip, Hongdong Li, and Yongduek Seo. Upnp: An optimal o(n) solution to the absolute pose problem with universal applicability. In David Fleet, Tomas Pajdla, Bernt Schiele, and Tinne Tuytelaars, editors, Computer Vision - ECCV 2014, pages 127-142. Springer International Publishing, 2014. 4", + "[21] Ren Komatsu, Hiromitsu Fujii, Yusuke Tamura, Atsushi Yamashita, and Hajime Asama. $360^{\\circ}$ depth estimation from multiple fisheye images with origami crown representation of icosahedron. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 10092-10099, 2020. 2, 6, 7", + "[22] Jiankun Li, Peisen Wang, Pengfei Xiong, Tao Cai, Ziwei Yan, Lei Yang, Jiangyu Liu, Haoqiang Fan, and Shuaicheng Liu. Practical stereo matching via cascaded recurrent network with adaptive correlation, 2022. 3, 6, 7", + "[23] Weiming Li and Y. F. Li. Single-camera panoramic stereo imaging system with a fisheye lens and a convex mirror. Opt. Express, 19(7):5855-5867, Mar 2011. 1", + "[24] Lahav Lipson, Zachary Teed, and Jia Deng. Raft-stereo: Multilevel recurrent field transforms for stereo matching. In 2021 International Conference on 3D Vision (3DV), pages 218-227, 2021. 3, 6, 7", + "[25] Bin Lv, Hao Xu, Jianqing Wu, Yuan Tian, Sheng Tian, and Suoyao Feng. Revolution and rotation-based method for roadside lidar data integration. Optics & Laser Technology, 119:105571, 2019. 1", + "[26] Nikolaus Mayer, Eddy Ilg, Philip Häusser, Philipp Fischer, Daniel Cremers, Alexey Dosovitskiy, and Thomas Brox. A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4040-4048, 2016. 6", + "[27] Christopher Mei and Patrick Rives. Single view point omnidirectional camera calibration from planar grids. In Proceedings 2007 IEEE International Conference on Robotics and Automation, pages 3945-3950, 2007. 2, 4" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "21537", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] M. Menze, C. Heipke, and A. Geiger. Joint 3d estimation of vehicles and scene flow. ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences, II-3/W5:427-434, 2015. 6, 7", + "[29] Andreas Meuleman, Hyeonjoong Jang, Daniel S. Jeon, and Min H. Kim. Real-time sphere sweeping stereo from multiview fisheye images. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11418-11427, 2021. 2, 6, 7", + "[30] Davide Scaramuzza, Agostino Martinelli, and Roland Siegwart. A toolbox for easily calibrating omnidirectional cameras. In 2006 IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 5695-5701, 2006. 2, 7", + "[31] Johannes Schneider, Cyril Stachniss, and Wolfgang Forstner. On the accuracy of dense fisheye stereo. IEEE Robotics and Automation Letters, 1(1):227-234, 2016. 1", + "[32] Miriam Schonbein and Andreas Geiger. Omnidirectional 3d reconstruction in augmented manhattan worlds. In 2014 IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 716-723, 2014. 1", + "[33] Xiao Song, Xu Zhao, Hanwen Hu, and Liangji Fang. Edgestereo: A context integrated residual pyramid network for stereo matching. In C.V. Jawahar, Hongdong Li, Greg Mori, and Konrad Schindler, editors, Computer Vision – ACCV 2018, pages 20–35. Springer International Publishing, 2019. 7", + "[34] Yu-Chuan Su and Kristen Grauman. Kernel transformer networks for compact spherical convolution. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9434-9443, 2019. 2", + "[35] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. ECCV, 12347(1):402-419, 2020. 2, 3, 5, 6", + "[36] Vladyslav Usenko, Nikolaus Demmel, and Daniel Cremers. The double sphere camera model. In 2018 International Conference on 3D Vision (3DV), pages 552-560, 2018. 2, 7", + "[37] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Ilia Polosukhin. Attention is all you need. In Proceedings of the 31st International Conference on Neural Information Processing Systems, NIPS'17, pages 6000-6010, Red Hook, NY, USA, 2017. Curran Associates Inc. 2, 5, 6", + "[38] Hengli Wang, Rui Fan, Peide Cai, and Ming Liu. Pvstereo: Pyramid voting module for end-to-end self-supervised stereo matching. IEEE Robotics and Automation Letters, 6(3):4353-4360, 2021. 3", + "[39] Changhee Won, Jongbin Ryu, and Jongwoo Lim. Omnimvss: End-to-end learning for omnidirectional stereo matching. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 8986-8995, 2019. 2, 6, 7", + "[40] Changhee Won, Jongbin Ryu, and Jongwoo Lim. Sweepnet: Wide-baseline omnidirectional depth estimation. In 2019 International Conference on Robotics and Automation (ICRA), pages 6073-6079, 2019. 2, 6, 7, 8", + "[41] Changhee Won, Jongbin Ryu, and Jongwoo Lim. End-to-end learning for omnidirectional stereo matching with un" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "certainty prior. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(11):3850-3862, 2021. 2, 6, 7", + "[42] Changhee Won, Hochang Seok, Zhaopeng Cui, Marc Pollefeys, and Jongwoo Lim. Omnislam: Omnidirectional localization and dense mapping for wide-baseline multicamera systems. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 559-566, 2020. 1", + "[43] Hao Xu, Yichen Zhang, Boyu Zhou, Luqi Wang, Xinjie Yao, Guotao Meng, and Shaojie Shen. Omni-swarm: A decentralized omnidirectional visual-inertial-uwb state estimation system for aerial swarm. arXiv preprint arXiv:2103.04131, 2021. 1, 2", + "[44] Feihu Zhang, Xiaojuan Qi, Ruigang Yang, Victor Prisacariu, Benjamin Wah, and Philip Torr. Domain-invariant stereo matching networks. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, Computer Vision - ECCV 2020, pages 420-439. Springer International Publishing, 2020. 6" + ], + "bbox": [ + 501, + 92, + 890, + 344 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "21538", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/OmniVidar_ Omnidirectional Depth Estimation From Multi-Fisheye Images/e34f49e7-aba6-408f-9647-b38f210ab7c3_model.json b/2023/OmniVidar_ Omnidirectional Depth Estimation From Multi-Fisheye Images/e34f49e7-aba6-408f-9647-b38f210ab7c3_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7355e2d7579747ab159c6eba7b96f34e4fdf48c2 --- /dev/null +++ b/2023/OmniVidar_ Omnidirectional Depth Estimation From Multi-Fisheye Images/e34f49e7-aba6-408f-9647-b38f210ab7c3_model.json @@ -0,0 +1,2200 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.105, + 0.131, + 0.868, + 0.155 + ], + "angle": 0, + "content": "OmniVidar: Omnidirectional Depth Estimation from Multi-Fisheye Images" + }, + { + "type": "text", + "bbox": [ + 0.156, + 0.18, + 0.816, + 0.253 + ], + "angle": 0, + "content": "Sheng Xie\\(^{1}\\), Daochuan Wang\\(^{2}\\), Yunhui Liu\\(^{1,2,3,*}\\) \nHarbin Institute of Technology, Shenzhen\\(^{1}\\) \nHong Kong Center for Logistics Robotics\\(^{2}\\), The Chinese University of Hong Kong\\(^{3}\\) \nxiesheng@stu.hit.edu.cn, wangdaochuan@hotmail.com, yhliu@mae.cuhk.edu.hk" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.287, + 0.314, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.319, + 0.474, + 0.577 + ], + "angle": 0, + "content": "Estimating depth from four large field of view (FoV) cameras has been a difficult and understudied problem. In this paper, we proposed a novel and simple system that can convert this difficult problem into easier binocular depth estimation. We name this system OmniVidar, as its results are similar to LiDAR, but rely only on vision. OmniVidar contains three components: (1) a new camera model to address the shortcomings of existing models, (2) a new multi-fisheye camera based epipolar rectification method for solving the image distortion and simplifying the depth estimation problem, (3) an improved binocular depth estimation network, which achieves a better balance between accuracy and efficiency. Unlike other omnidirectional stereo vision methods, OmniVidar does not contain any 3D convolution, so it can achieve higher resolution depth estimation at fast speed. Results demonstrate that OmniVidar outperforms all other methods in terms of accuracy and performance." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.607, + 0.21, + 0.623 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.633, + 0.47, + 0.844 + ], + "angle": 0, + "content": "Depth estimation from images is an important research field in computer vision, as it enables the acquisition of depth information with low-cost cameras for a wide range of applications. Traditional stereo cameras with pinhole lenses are limited in their FoV. However, many scenarios require an omnidirectional depth map, such as autonomous driving [42] and robot navigation [13, 43]. Although there are active sensors available that can provide omnidirectional depth information, such as LiDAR [25], their high cost makes them less accessible than stereo vision. Passive sensors, such as RGB cameras, are a common choice for depth estimation due to their low cost, lightweight, and low power consumption. To increase the FoV, fisheye lenses are often introduced into stereo vision setups." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.846, + 0.47, + 0.876 + ], + "angle": 0, + "content": "Over the past few decades, various methods have been proposed for depth estimation using fisheye cameras. These" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.286, + 0.693, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.695, + 0.286, + 0.891, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.415, + 0.891, + 0.502 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.501, + 0.891, + 0.64 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.65, + 0.894, + 0.707 + ], + "angle": 0, + "content": "Figure 1. Our prototype that built with four \\(250^{\\circ}\\) fisheye cameras and its results of dense inverse distance map and cloud points. It can get great depth estimation results in real scenes and achieve real-time performance on modern GPU." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.895, + 0.903 + ], + "angle": 0, + "content": "include the binocular fisheye system [31], up-down fisheye system [13], and catadiptric cameras [17, 23, 32]. However, all of these approaches have limitations. The binocular fisheye system cannot provide an omnidirectional perception. The up-down fisheye system and catadiptric cameras can offer horizontal \\(360^{\\circ}\\) depth perception, but their vertical FoV is limited. Furthermore, catadiptric cameras tend to be bulky, challenging to calibrate, and prone to errors. It turns out that the best choice for omnidirectional depth estimation is a system consisting of four cameras with extremely wide FoV (\\(>180^{\\circ}\\)). This system enables" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.222, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21529" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.273 + ], + "angle": 0, + "content": "\\(360^{\\circ}\\) depth estimation both horizontally and vertically, and is light-weight, convenient to calibrate and maintain. Several studies [21,29,39,41] have shown excellent results with this approach. However, these methods often use many 3D convolutions, resulting in low efficiency. To adapt to limited memory, the images must be down-sampled, leading to a lower resolution of the output depth map. Furthermore, many of these approaches perform depth estimation using the original fisheye image without explicitly processing image distortion, which leads to the construction of a tedious and complicated cost volume and can result in reduced accuracy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.274, + 0.473, + 0.395 + ], + "angle": 0, + "content": "Moreover, due to the complex imaging process of large FoV fisheye lenses, handling fisheye image distortion using mathematical formulas can be a challenging task. While several excellent large FoV fisheye camera models have been proposed [2, 18, 19, 27, 30, 36], our experiments have shown that none of these models can accurately approximate the imaging process, and we get less satisfactory depth estimation results when using them." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.396, + 0.473, + 0.698 + ], + "angle": 0, + "content": "Inspired by the above observations, we propose OmniVidar, a novel and simple multi-fisheye omnidirectional depth estimation system, as shown in Figure 2. OmniVidar contains three components. Firstly, we improve the DSCM [36] and propose a new camera model, named Triple Sphere Camera Model (TSCM), which can better approximate the imaging process and achieve the best accuracy. Then we propose an epipolar rectification algorithm designed for multi-fisheye camera system. We leverage a cubic-like projection approach to transform the four fisheye camera systems into four binocular camera systems and then conduct epipolar rectification on each binocular system. This method solves the distortion issue and reduces the complex multi-fisheye omnidirectional depth estimation problem to a much simpler binocular depth estimation problem. Finally, we design a lightweight binocular depth estimation network based on RAFT [35]. We add Transformer encoder [37] into feature extraction in RAFT to combine the advantages of Transformer and GRU, and it's easy to balance the accuracy and efficiency." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.699, + 0.472, + 0.759 + ], + "angle": 0, + "content": "We compare OmniVidar with existing methods on several datasets. The results demonstrate that our method outperforms all the others in terms of speed and accuracy, achieving State-of-The-Art performance." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.775, + 0.22, + 0.79 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.801, + 0.308, + 0.817 + ], + "angle": 0, + "content": "2.1. Multiview Fisheye Stereo" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Gao et al. [13] mount two large FoV cameras \\((245^{\\circ})\\) vertically and reversely to obtain a \\(360^{\\circ}\\) horizontal and \\(65^{\\circ}\\) vertical FoV overlap and use this system in UAV for depth perception [43]. Won et al. [40] use four cameras with \\(220^{\\circ}\\) FoV and propose SweepNet algorithm. They propose" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.302 + ], + "angle": 0, + "content": "Sphere-Sweep method modified from Plane-Sweep [16] to construct the cost volume and then use the cost aggregation algorithm in SGM [15] to obtain depth. After this, Won et al. propose OmniMVS [39, 41], which uses 3D convolution instead of SGM for depth regression and achieves higher accuracy. Komatsu et al. [21] propose CrownConv360. They improve OmniMVS by using the icosahedral projection in the feature extraction and depth regression stage for distortion reduction. Instead of using neural network, Meuleman et al. [29] choose to investigate cost aggregation and propagation in traditional stereo matching. And they achieve real-time high-resolution depth estimation at 29 fps on the embedded device TX2. However, the algorithm is only suitable for short camera baseline." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.304, + 0.895, + 0.38 + ], + "angle": 0, + "content": "All of the above works have more or fewer deficiencies. We propose OmniVidar, an omnidirectional depth estimation method with high accuracy, which can run in real-time and solve the fisheye image distortion problem with good generalization in various scenes." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.392, + 0.73, + 0.408 + ], + "angle": 0, + "content": "2.2. Fisheye Image Distortion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.416, + 0.892, + 0.461 + ], + "angle": 0, + "content": "To solve the distortion problem, a camera model suitable for fisheye is required, and the distortion should be handled explicitly in depth estimation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.462, + 0.895, + 0.644 + ], + "angle": 0, + "content": "Scaramuzzi et al. [30] propose a high-order polynomial-based camera projection model, which is highly generalizable and can be widely used for various catadioptric and fisheye cameras. But it has no closed-form solution, so the inverse projection equation needs to be fitted with a high-order polynomial, which introduces errors. Mei et al. [27] propose the MUCM, which is an improvement of the UCM [2] to achieve higher calibration accuracy and propose a novel intrinsic parameters initialization method. Usenko et al. [36] propose the DSCM, which has only one more intrinsic parameter than the MUCM model, but the accuracy is substantially improved." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.645, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Su et al. [34] propose a mini-network that learns to modify the shape of filters based on location. Coors et al. [8] chose to use a camera projection model to shape the convolution kernel to directly address distortion. An alternative idea was chosen by Cohen et al. [7] who define a form of convolution on the Lie group \\(SO(3)\\) that convolves over the rotation space rather than the translation space. Eder et al. [9] compared the performance of three approaches: equirectangular projection, cubic projection, and 7-level subdivided icosahedral projection, on depth estimation and semantic segmentation tasks. The results showed that icosahedral projection achieved the best results in all tasks. Komatsu et al. [21] used icosahedral projection in multi-fisheye omnidirectional stereo vision to improve the depth estimation results of OmniMVS. Eder et al. [10] proposed a tangent plane projection modified from the icosahedron projection." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21530" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.081, + 0.892, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.302, + 0.897, + 0.361 + ], + "angle": 0, + "content": "Figure 2. System overview. Our system contains three parts: multi-fisheye calibration, multi-fisheye epipolar rectification, and binocular depth estimation. In epipolar rectification, the images in the left column are original. The images in the middle column are the images after abstracting each fisheye camera into two pinhole cameras, where the two images connected by a red dotted arrow are a pair of binocular cameras. The images are rearranged and combined to obtain the right column." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.385, + 0.473, + 0.492 + ], + "angle": 0, + "content": "The above algorithms have a common feature that they are all designed for the monocular image. Thus they perform well on tasks that require only one camera, yet not much good for multi-fisheye tasks. Therefore, to address this problem, we propose a distortion removal method designed for multi-fisheye systems, which greatly simplifies the depth estimation problem at the same time." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.502, + 0.29, + 0.516 + ], + "angle": 0, + "content": "2.3. RAFT in Stereo Vision" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.526, + 0.47, + 0.768 + ], + "angle": 0, + "content": "Zachary et al. [35] propose Recurrent All-Pairs Field Transforms (RAFT), a highly efficient and accurate deep network architecture for optical flow. RAFT uses modified GRU blocks to iteratively refine optical flow and shares weights between refinement units. Lahav et al. [24] apply the RAFT to binocular rectified stereo vision and introduce multi-level convolutional GRUs, which can more efficiently propagate information across the image. Wang et al. [38] modify RAFT and propose PVStereo which greatly outperforms other self-supervised stereo matching approaches. Li et al. [22] combine RAFT and Transformer and propose CREStereo. They add Transformer encoder in image feature extraction and repeat the cross-attention between left and right images in the iterative update stage. This method performs well in various binocular datasets with excellent generalization." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.769, + 0.47, + 0.815 + ], + "angle": 0, + "content": "We imitate CREStereo and improve it based on RAFT to obtain a binocular stereo vision depth estimation network with a better trade-off between efficiency and accuracy." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.829, + 0.176, + 0.845 + ], + "angle": 0, + "content": "3. Methods" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.903 + ], + "angle": 0, + "content": "The overall flow of our method is shown in Figure 2 and is divided into three parts: camera calibration, epipolar rectification, and depth estimation." + }, + { + "type": "image", + "bbox": [ + 0.541, + 0.385, + 0.858, + 0.698 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.723, + 0.895, + 0.81 + ], + "angle": 0, + "content": "Figure 3. Our Triple Sphere Camera Model. First, point \\( P \\) is projected onto the first sphere and then onto the second sphere, then the third. The second sphere is shifted w.r.t. the first sphere by \\( \\xi \\), and the third is shifted w.r.t. the second sphere by \\( \\lambda \\). Finally, the point is projected onto the image plane of a pinhole camera which is shifted by \\( \\frac{\\alpha}{1 - \\alpha} \\) from the third sphere." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.843, + 0.693, + 0.858 + ], + "angle": 0, + "content": "3.1. Camera Calibration" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.893, + 0.903 + ], + "angle": 0, + "content": "We propose the Triple Sphere Camera Model, which can better fit the imaging process of large FoV fisheye cameras" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.958 + ], + "angle": 0, + "content": "21531" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.373, + 0.106 + ], + "angle": 0, + "content": "and has a closed-form unprojection solution." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.107, + 0.47, + 0.257 + ], + "angle": 0, + "content": "As shown in Figure 3, our projection model considers that the incident light is refracted three times, and the displacements of the three unit spherical centers are \\(\\xi\\) and \\(\\lambda\\). After three refractions, the incident light is finally projected to the image plane according to the pinhole camera model, and the displacement of the pinhole camera model's optical center from the third unit sphere is \\(\\frac{\\alpha}{1 - \\alpha}\\). Therefore, our model has totally 7 camera internal parameters: \\(f_{x}, f_{y}, c_{x}, c_{y}, \\xi, \\lambda, \\alpha\\). The projection equations are defined as follows." + }, + { + "type": "equation", + "bbox": [ + 0.127, + 0.265, + 0.47, + 0.315 + ], + "angle": 0, + "content": "\\[\n\\pi (P, i) = \\frac {1}{\\zeta} \\left[ \\begin{array}{c c c} f _ {x} & 0 & c _ {x} \\\\ 0 & f _ {y} & c _ {y} \\end{array} \\right] \\left[ \\begin{array}{l} X \\\\ Y \\\\ \\zeta \\end{array} \\right] \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.165, + 0.316, + 0.469, + 0.342 + ], + "angle": 0, + "content": "\\[\n\\zeta = Z + \\xi d _ {1} + \\lambda d _ {2} + \\frac {\\alpha}{1 - \\alpha} d _ {3} \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.158, + 0.345, + 0.469, + 0.365 + ], + "angle": 0, + "content": "\\[\nd _ {1} = \\sqrt {X ^ {2} + Y ^ {2} + Z ^ {2}} \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.158, + 0.368, + 0.469, + 0.387 + ], + "angle": 0, + "content": "\\[\nd _ {2} = \\sqrt {X ^ {2} + Y ^ {2} + (\\xi d _ {1} + Z) ^ {2}} \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.158, + 0.39, + 0.469, + 0.409 + ], + "angle": 0, + "content": "\\[\nd _ {3} = \\sqrt {X ^ {2} + Y ^ {2} + (\\lambda d _ {2} + \\xi d _ {1} + Z) ^ {2}} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.42, + 0.468, + 0.448 + ], + "angle": 0, + "content": "where \\(i\\) is the vector of intrinsic parameters, \\(\\pi\\) is the projection function." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.45, + 0.468, + 0.479 + ], + "angle": 0, + "content": "A set of 3D points that results in valid projection is expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.155, + 0.49, + 0.469, + 0.509 + ], + "angle": 0, + "content": "\\[\n\\Omega = \\left\\{x \\in \\mathbb {R} ^ {3} | z > - w _ {2} d _ {1} \\right\\} \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.511, + 0.469, + 0.546 + ], + "angle": 0, + "content": "\\[\nw _ {2} = \\frac {\\xi + \\lambda + w _ {1}}{\\sqrt {1 + (\\xi + \\lambda) ^ {2} + 2 w _ {1} (\\xi + \\lambda)}} \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.548, + 0.469, + 0.574 + ], + "angle": 0, + "content": "\\[\nw _ {1} = \\frac {\\alpha}{1 - \\alpha} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.584, + 0.43, + 0.599 + ], + "angle": 0, + "content": "The unprojection function is computed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.098, + 0.609, + 0.469, + 0.659 + ], + "angle": 0, + "content": "\\[\n\\pi^ {- 1} (p, i) = \\mu \\left[ \\begin{array}{l} \\eta \\gamma x \\\\ \\eta \\gamma y \\\\ m _ {z} \\end{array} \\right] - \\left[ \\begin{array}{l} 0 \\\\ 0 \\\\ \\xi \\end{array} \\right] \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.661, + 0.468, + 0.679 + ], + "angle": 0, + "content": "\\[\n\\mu = \\xi m _ {z} + \\sqrt {\\xi^ {2} m _ {z} ^ {2} - \\xi^ {2} + 1} \\tag {10}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.138, + 0.682, + 0.469, + 0.698 + ], + "angle": 0, + "content": "\\[\nm _ {z} = \\eta (\\gamma - \\zeta) - \\lambda \\tag {11}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.702, + 0.469, + 0.72 + ], + "angle": 0, + "content": "\\[\n\\eta = \\lambda (\\gamma - \\zeta) + \\sqrt {\\lambda^ {2} (\\gamma - \\zeta) ^ {2} - \\lambda^ {2} + 1} \\tag {12}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.724, + 0.469, + 0.757 + ], + "angle": 0, + "content": "\\[\n\\gamma = \\frac {\\zeta + \\sqrt {1 + (1 - \\zeta^ {2}) \\left(x ^ {2} + y ^ {2}\\right)}}{x ^ {2} + y ^ {2} + 1} \\tag {13}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.76, + 0.469, + 0.801 + ], + "angle": 0, + "content": "\\[\n\\zeta = \\left\\{ \\begin{array}{l l} \\frac {\\alpha}{1 - \\alpha} & \\text {i f} \\alpha \\leq 0. 5 \\\\ \\frac {1 - \\alpha}{\\alpha} & \\text {i f} \\alpha > 0. 5 \\end{array} \\right. \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.811, + 0.357, + 0.825 + ], + "angle": 0, + "content": "where \\((x,y)\\) is the normalized coordinate." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.47, + 0.901 + ], + "angle": 0, + "content": "We calibrate the intrinsic and extrinsic parameters of multi-fisheye camera system using a highly accurate planar checkerboard. The goal of the calibration is to minimize the reprojection error of the corner points in all pictures. For each image \\( n \\) in the calibration sequence, the" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.088, + 0.892, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.252, + 0.892, + 0.294 + ], + "angle": 0, + "content": "Figure 4. The multi-fisheye epipolar rectification schematic. Each fisheye camera is converted into two pinhole cameras with an angle of \\(90^{\\circ}\\). Then the system becomes four binocular systems." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.318, + 0.892, + 0.407 + ], + "angle": 0, + "content": "corner detector can obtain the projection point \\( u_{nk} \\) of the \\( k \\)th corner point \\( x_{k} \\). The coordinate of \\( u_{nk} \\) is related to the camera intrinsic and extrinsic parameters. Let \\( s = [i, T_{cam_0}, T_{cam_1}, \\dots, T_{cam_n}] \\) be the parameter to optimize. We can construct the nonlinear optimization problem as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.517, + 0.416, + 0.891, + 0.458 + ], + "angle": 0, + "content": "\\[\ns = \\arg \\min _ {s ^ {\\prime}} \\sum_ {n = 0} ^ {N} \\sum_ {k \\in K} \\rho \\left(\\left(\\pi \\left(T _ {c a m _ {n}} x _ {k}, \\boldsymbol {i}\\right) - u _ {n k}\\right) ^ {2}\\right) \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.467, + 0.891, + 0.528 + ], + "angle": 0, + "content": "where \\( T_{cam_n} \\in SE(3) \\) is the transformation from the coordinate frame of the calibration grid to the camera coordinate frame for image \\( n \\). \\( K \\) is a set of detected corner points for the image \\( n \\) and \\( \\rho \\) is the robust Huber norm." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.528, + 0.892, + 0.633 + ], + "angle": 0, + "content": "Due to the highly nonlinear nature of the image, the above optimization problem requires a good initial value. We initialize the intrinsic parameters using the previously proposed method [27] and find initial poses using the UPnP algorithm [20]. After obtaining the initial values of the parameters, we completed the optimization operation using Ceres Solver [1]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.642, + 0.84, + 0.659 + ], + "angle": 0, + "content": "3.2. Multi-Fisheye Camera Epipolar Rectify" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.665, + 0.892, + 0.832 + ], + "angle": 0, + "content": "We propose a simple and effective undistortion method that cleverly transforms the omnidirectional depth estimation problem into a binocular depth estimation problem. And the method can be used in the system with any number of fisheye cameras. As shown in Figure 4, we abstract each fisheye camera as two pinhole cameras with \\(90^{\\circ}\\) angle1. Then the four fisheye cameras system becomes four independent binocular systems. Each binocular systems can be rectified through epipolar rectification. We modified the classical epipolar rectification algorithm proposed by Fusiello [11] for our system." + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.839, + 0.892, + 0.901 + ], + "angle": 0, + "content": "1Although here we only show the horizontal binoculars, it is possible to construct binoculars facing both above and below to estimate the true omnidirectional depth. However, we consider that the above and below depth values are rarely used in practical applications. So we just use a limited \\(110^{\\circ}\\) vertical FoV, which satisfies most of the application requirements." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21532" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.087, + 0.895, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.352, + 0.895, + 0.409 + ], + "angle": 0, + "content": "Figure 5. The structure of our proposed binocular depth estimation network. A pair of stereo images are fed into feature extraction networks. Then the output features are calculated by self-attention and cross-attention, and after doing self-attention, the features of left image are used for context information for latter recurrent update stage. Then the left and right features are multiplied to construct cost volume. Finally use the disparity update model same as RAFT [35]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.435, + 0.47, + 0.516 + ], + "angle": 0, + "content": "Firstly, the rotation matrix of the cameras in binocular system is calculated. Let the optical centers of the left camera and right camera be \\( C_1 = \\left[x_1 y_1 z_1\\right]^T \\), \\( C_2 = \\left[x_2 y_2 z_2\\right]^T \\), the rotation matrix can be calculated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.527, + 0.468, + 0.544 + ], + "angle": 0, + "content": "\\[\nR = \\left[ \\begin{array}{l l l} r _ {1} ^ {T} & r _ {2} ^ {T} & r _ {3} ^ {T} \\end{array} \\right] \\tag {16}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.548, + 0.468, + 0.565 + ], + "angle": 0, + "content": "\\[\nr _ {1} = \\left[ \\begin{array}{l l l} x _ {2} - x _ {1} & y _ {2} - y _ {1} & z _ {2} - z _ {1} \\end{array} \\right]. n o r m a l i z e () \\tag {17}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.568, + 0.468, + 0.585 + ], + "angle": 0, + "content": "\\[\nr _ {3} = \\left[ - \\left(z _ {2} - z _ {1}\\right) \\quad 0 \\quad x _ {2} - x _ {1} \\right]. n o r m a l i z e () \\tag {18}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.589, + 0.468, + 0.602 + ], + "angle": 0, + "content": "\\[\nr _ {2} = r _ {3} \\times r _ {1} \\tag {19}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.614, + 0.47, + 0.705 + ], + "angle": 0, + "content": "Then remap the image for epipolar rectification. Taking the left camera in one binocular system as an example. Let the rotation matrix of the corresponding fisheye camera be \\( R \\), the optical center be \\( C \\). For a certain spatial point \\( P = (X, Y, Z, 1) \\), its projection point on normalized plane is \\( p = (u, v, 1) \\). The projection equation is shown below." + }, + { + "type": "equation", + "bbox": [ + 0.162, + 0.714, + 0.468, + 0.778 + ], + "angle": 0, + "content": "\\[\n\\mu \\left[ \\begin{array}{c} u \\\\ v \\\\ f (u, v) \\end{array} \\right] = [ R - R C ] \\left[ \\begin{array}{l} X \\\\ Y \\\\ Z \\\\ 1 \\end{array} \\right] \\tag {20}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.786, + 0.47, + 0.832 + ], + "angle": 0, + "content": "Let the rotation matrix of the rectified pinhole camera be \\( R' \\), and the corresponding normalized plane coordinate be \\( (u', v', 1) \\). Then the projection equation is shown below." + }, + { + "type": "equation", + "bbox": [ + 0.171, + 0.841, + 0.468, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\mu^ {\\prime} \\left[ \\begin{array}{l} u ^ {\\prime} \\\\ v ^ {\\prime} \\\\ 1 \\end{array} \\right] = \\left[ \\begin{array}{l l} R ^ {\\prime} & - R ^ {\\prime} C \\end{array} \\right] \\left[ \\begin{array}{l} X \\\\ Y \\\\ Z \\\\ 1 \\end{array} \\right] \\tag {21}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.435, + 0.892, + 0.465 + ], + "angle": 0, + "content": "The relationship between fisheye image and virtual binocular image can be obtained, see equation (22)." + }, + { + "type": "equation", + "bbox": [ + 0.516, + 0.47, + 0.891, + 0.519 + ], + "angle": 0, + "content": "\\[\n\\lambda \\left[ \\begin{array}{l} u ^ {\\prime} \\\\ v ^ {\\prime} \\\\ 1 \\end{array} \\right] = \\left[ \\begin{array}{l l l} R ^ {\\prime} & - R ^ {\\prime} C \\end{array} \\right] \\left[ \\begin{array}{l l} R & - R C \\end{array} \\right] ^ {- 1} \\left[ \\begin{array}{c} u \\\\ v \\\\ f (u, v) \\end{array} \\right] \\tag {22}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.523, + 0.892, + 0.569 + ], + "angle": 0, + "content": "That is, for each pixel in the original fisheye image, the pixel coordinate in the pinhole image can be calculated. We can use pixel interpolation to obtain the rectified image." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.57, + 0.892, + 0.645 + ], + "angle": 0, + "content": "As shown in Figure 2, it can be seen that after epipolar rectification, the four fisheye images are converted into a pair of binocular images, where each image is stitched together from the four virtual pinhole camera images. A simple binocular depth estimation algorithm is then used." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.651, + 0.725, + 0.667 + ], + "angle": 0, + "content": "3.3. Stereo Depth Estimation" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.675, + 0.892, + 0.764 + ], + "angle": 0, + "content": "Since our system needs to estimate the depth values of four binocular systems simultaneously, which is very computationally intensive, our primary goal is not to design a binocular algorithm with the highest accuracy, but to minimize the accuracy loss while ensuring real-time performance." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The structure of our binocular depth estimation network is shown in Figure 5. We added Transformer [37] with self-attention and cross-attention in the image feature extraction stage. In binocular depth estimation, limited by computational performance and graphics card memory, many algorithms have to limit the matching search range of pixels. In this paper, we design an ingenious cost-volume calculation that can extend the estimable disparity range to the image width, while also reducing the computational complexity." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21533" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.084, + 0.086, + 0.875, + 0.43 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.441, + 0.895, + 0.469 + ], + "angle": 0, + "content": "Figure 6. The depth estimation comparison between ours and other methods. The final row is the results from our method, and the column annotations indicate the dataset source." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.495, + 0.47, + 0.556 + ], + "angle": 0, + "content": "Given two images \\( I_1 \\in \\mathbb{R}^{H \\times W \\times C} \\), \\( I_2 \\in \\mathbb{R}^{H \\times W \\times C} \\), we need to compute the cost volume \\( C(I_1, I_2) \\in \\mathbb{R}^{H \\times W \\times W} \\). We found that the cost volume can be obtained by single matrix operation as follows." + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.56, + 0.469, + 0.581 + ], + "angle": 0, + "content": "\\[\nC \\left(I _ {1}, I _ {2}\\right) = I _ {1} \\in \\mathbb {R} ^ {H \\times W \\times C} \\cdot I _ {2} ^ {T} \\in \\mathbb {R} ^ {H \\times C \\times W} \\tag {23}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.585, + 0.47, + 0.631 + ], + "angle": 0, + "content": "To enable the network to learn information at multiple scales, we used the cost volume pooling opertaion in RAFT-Stereo [24]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.642, + 0.21, + 0.658 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.666, + 0.236, + 0.682 + ], + "angle": 0, + "content": "4.1. Implementation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.47, + 0.856 + ], + "angle": 0, + "content": "Our depth estimation network structure is shown in Figure 5. The number of channels in our network is identical to RAFT [35]. The feature encoder consists of 6 residual blocks, 2 at \\(1/2\\) resolution, 2 at \\(1/4\\) resolution, and 2 at \\(1/8\\) resolution. The number of channels is 64, 128, and 256, respectively. The \\(1/8\\) resolution feature map is then fed into the Transformer encoder [37]. We set the number of heads in attention to 8. The GRU iterative update part is identical to RAFT, and more details can be found in [35]. In addition, we replace all the normalization operators in RAFT with domain normalize [44]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Since we convert the multi-fisheye cameras to four binocular cameras, we can use a large number of public binocular datasets for training, instead of multi-fisheye" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.496, + 0.892, + 0.617 + ], + "angle": 0, + "content": "datasets, which are very limited. Similar to the training strategy in CREStereo [22], we collected most of public binocular datasets, including SceneFlow [26], Sintel [4], CREStereo [22], KITTI [14, 28], InStereo2K [3], and Virtual KITTI [12]. The images are uniformly randomly cropped with a fixed resolution during the training process to solve the problem of different image resolutions in different datasets." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.618, + 0.892, + 0.664 + ], + "angle": 0, + "content": "We use a single NVIDIA RTX2080Ti for training, the batch size is set to 4, the learning rate is 0.0004, and we use Adam to train 200 epochs to get the final network weights." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.665, + 0.893, + 0.786 + ], + "angle": 0, + "content": "To test the performance of OmniVidar in real scenarios, we built a test setup. We used four leopard LI-USB30-OV10640-490-GMSL cameras with an image resolution of \\(1280 \\times 1024\\), paired with a fisheye lens with \\(250^{\\circ}\\) FoV. The four cameras look in four directions, front and back, left and right, as shown in Figure 1. We use the device to capture indoor, corridor, and outdoor images, which can fully test the performance of OmniVidar in real scenes." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.801, + 0.621, + 0.815 + ], + "angle": 0, + "content": "4.2. Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.892, + 0.9 + ], + "angle": 0, + "content": "We perform the comparison on the Omnihouse, OmniThings provided by Won et al. [40, 41], and the datasets provided by Meuleman et al. [29]. Our OmniVidar has only been trained on the binocular dataset mentioned in Section 4.1, not on the test dataset, and the methods involved in the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21534" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.124, + 0.089, + 0.849, + 0.198 + ], + "angle": 0, + "content": "
MethodMeuleman et al. [29]Omnihouse [40]Omnithings [39]
bad 1.0RMSEMAEbad 1.0RMSEMAEbad 1.0RMSEMAE
OmniVidar(Ours)18.4990.9210.8034.8861.0021.04211.1990.9580.461
OmniMVS-ft [41]60.3722.3061.73226.3570.9290.70559.7532.2701.723
CrownConv360 [21]69.1194.0203.00585.55911.0776.72993.9469.5456.518
Meuleman et al. [29]44.1111.6641.27311.3040.6710.41065.3482.5501.941
" + }, + { + "type": "table_caption", + "bbox": [ + 0.092, + 0.206, + 0.877, + 0.222 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison of algorithm accuracy on several datasets. The less the number is, the more accurate the method is." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.248, + 0.47, + 0.278 + ], + "angle": 0, + "content": "comparison all use the training method corresponding to the original paper." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.278, + 0.471, + 0.355 + ], + "angle": 0, + "content": "As can be seen in the Figure 6, although OmniVidar is not trained and fine-tuned on the above datasets, its depth estimation accuracy on these datasets far exceeds that of other methods, with accurate depth estimation in weakly textured regions." + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.368, + 0.48, + 0.446 + ], + "angle": 0, + "content": "
MethodRuntimeMemoryParameters
OmniVidar (Ours)66 ms3.7 GB4.8 MB
OmniMVS-ft [41]110 ms6.8 GB43.7 MB
CrownConv360 [21]623 ms6.3 GB16.7 MB
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.455, + 0.469, + 0.47 + ], + "angle": 0, + "content": "Table 2. Efficiency comparison with other deep learning methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.487, + 0.47, + 0.593 + ], + "angle": 0, + "content": "The analysis of the quantitative results are shown in the Table 1 and Table 2. It can be seen that OmniVidar has the least memory consumption, the shortest time consumption and the highest accuracy compared to other deep learning based solutions. Compared with non-deep learning methods, the accuracy is much higher than it, despite the relatively low efficiency." + }, + { + "type": "image", + "bbox": [ + 0.083, + 0.606, + 0.466, + 0.699 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.721, + 0.47, + 0.765 + ], + "angle": 0, + "content": "Figure 7. Plot comparing the accuracy and efficiency evaluated in KITTI-2015 [28]. It can be seen that OmniVidar achieves real-time performance while maintain the accuracy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.471, + 0.903 + ], + "angle": 0, + "content": "We also compared the binocular depth estimation network in OmniVidar with other binocular depth estimation networks. We chose the KITTI-2015 dataset [28] for the evaluation and the results are shown in Figure 7. We compare our method with CREStereo [22], LEAStereo [6], PSMNet [5], EdgeStereo [33], RAFT-Stereo [24]. It can see that our scheme achieves a better balance between accuracy and efficiency." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.247, + 0.604, + 0.262 + ], + "angle": 0, + "content": "4.3. Ablation" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.271, + 0.892, + 0.302 + ], + "angle": 0, + "content": "Camera Model: We compare the effect of omnidirectional depth estimation under several camera models." + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.313, + 0.698, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.536, + 0.415, + 0.674, + 0.428 + ], + "angle": 0, + "content": "Original fisheye image" + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.313, + 0.892, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.734, + 0.415, + 0.869, + 0.428 + ], + "angle": 0, + "content": "Scaramuzza et al. [30]" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.429, + 0.7, + 0.532 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.564, + 0.532, + 0.637, + 0.545 + ], + "angle": 0, + "content": "DSCM [36]" + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.429, + 0.892, + 0.532 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.754, + 0.532, + 0.834, + 0.544 + ], + "angle": 0, + "content": "TSCM(Ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.556, + 0.892, + 0.598 + ], + "angle": 0, + "content": "Figure 8. Depth estimation results using different camera projection models. Our TSCM can best remove the fisheye distortion and obtain the most accurate result." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.617, + 0.892, + 0.723 + ], + "angle": 0, + "content": "As shown in Figure 8, the distortion of the fisheye lens is not sufficiently handled in the DSCM [36] and the model proposed by Scaramuzza et al. [30], which affects the accuracy of depth estimation and leads to serious distortion of objects in the point cloud. In contrast, the quality of the point cloud obtained by the TSCM is much better, with flat wall surfaces and more accurate distance measurements." + }, + { + "type": "table", + "bbox": [ + 0.522, + 0.734, + 0.871, + 0.812 + ], + "angle": 0, + "content": "
FoV2 sphere3 sphere (Ours)4 sphere
195°0.1530 px0.1528 px0.1528 px
220°0.2031 px0.1970 px0.1952 px
250°0.1365 px0.1255 px0.1222 px
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.821, + 0.892, + 0.851 + ], + "angle": 0, + "content": "Table 3. Mean reprojection error for camera models with different number sphere." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.894, + 0.903 + ], + "angle": 0, + "content": "We also compared the accuracy with different number of the refract sphere. Table 3 shows the results. It can be seen" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21535" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.086, + 0.872, + 0.55 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.572, + 0.893, + 0.601 + ], + "angle": 0, + "content": "Figure 9. The images captured with our prototype and the inverse depth map and cloud points obtained by OmniVidar. It can be seen that OmniVidar has great generalization in various scenes, and performs well in low-textured areas." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.627, + 0.47, + 0.673 + ], + "angle": 0, + "content": "that the camera model with 3 sphere can significantly improve the accuracy in large FoV scenes, while the improvement brought by more sphere is very limited." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.673, + 0.47, + 0.809 + ], + "angle": 0, + "content": "Distortion Handle: To verify whether the improvement of depth estimation accuracy is helped by removing the fisheye image distortion using epipolar rectification, we estimate the depth directly using the original fisheye image and compare it with the undistorted method. Unlike the undistorted binocular stereo, we use the Sphere-Sweep algorithm [40] to construct the cost volume. This step does not contain trainable parameters, so it does not affect the network itself and ensures the fairness of the comparison." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.47, + 0.9 + ], + "angle": 0, + "content": "The point cloud visualization of the results of the depth estimation of both is shown in Figure 8. It can be seen that when using original image, the depth information obtained from different cameras is not consistent. In contrast, the point cloud obtained by depth estimation after epipolar rectification shows the structural framework of the scene real-" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.627, + 0.658, + 0.643 + ], + "angle": 0, + "content": "istrically and accurately." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.654, + 0.628, + 0.669 + ], + "angle": 0, + "content": "4.4. Test in Real" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.678, + 0.892, + 0.754 + ], + "angle": 0, + "content": "We test the performance of the OmniVidar algorithm on a real dataset. Figure 9 shows the inverse depth maps and point clouds obtained by OmniVidar on the real dataset. It can be seen that our network has a strong generalization to obtain high-quality point clouds on real datasets." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.769, + 0.619, + 0.785 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In this paper, we propose a novel, simple and effective system OmniVidar, to address the inefficiency of current omnidirectional depth estimation methods due to the extensive use of 3D convolution and the poor accuracy due to the lack of explicit handling of distortion in fisheye images. Our method outperforms all other methods in terms of time consumption and accuracy." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21536" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.468, + 0.143 + ], + "angle": 0, + "content": "[1] Sameer Agarwal, Keir Mierle, and The Ceres Solver Team. Ceres solver. http://ceres-solver.org. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.146, + 0.469, + 0.187 + ], + "angle": 0, + "content": "[2] Simon Baker and Shree K. Nayar. A theory of single-viewpoint catadioptric image formation. Int. J. Comput. Vision, 35(2):175-196, nov 1999. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.189, + 0.469, + 0.243 + ], + "angle": 0, + "content": "[3] Wei Bao, Wei Wang, Yuhua Xu, Yulan Guo, Siyu Hong, and Xiaohu Zhang. Instereo2k: a large real dataset for stereo matching in indoor scenes. Science China Information Sciences, 63(11):212101, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.246, + 0.469, + 0.315 + ], + "angle": 0, + "content": "[4] D. J. Butler, J. Wulff, G. B. Stanley, and M. J. Black. A naturalistic open source movie for optical flow evaluation. In A. Fitzgibbon et al. (Eds.), editor, European Conf. on Computer Vision (ECCV), Part IV, LNCS 7577, pages 611-625. Springer-Verlag, Oct. 2012. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.317, + 0.469, + 0.371 + ], + "angle": 0, + "content": "[5] Jia-Ren Chang and Yong-Sheng Chen. Pyramid stereo matching network. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5410-5418, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.374, + 0.469, + 0.442 + ], + "angle": 0, + "content": "[6] Xuelian Cheng, Yiran Zhong, Mehrtash Harandi, Yuchao Dai, Xiaojun Chang, Tom Drummond, Hongdong Li, and Zongyuan Ge. Hierarchical neural architecture search for deep stereo matching. arXiv preprint arXiv:2010.13501, 2020.7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.445, + 0.469, + 0.487 + ], + "angle": 0, + "content": "[7] Taco S. Cohen, Mario Geiger, Jonas Kohler, and Max Welling. Spherical CNNs. In International Conference on Learning Representations, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.489, + 0.469, + 0.572 + ], + "angle": 0, + "content": "[8] Benjamin Coors, Alexandru Paul Condurache, and Andreas Geiger. Spherenet: Learning spherical representations for detection and classification in omnidirectional images. In Vittorio Ferrari, Martial Hebert, Cristian Sminchisescu, and Yair Weiss, editors, Computer Vision – ECCV 2018, pages 525–541. Springer International Publishing, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.574, + 0.469, + 0.628 + ], + "angle": 0, + "content": "[9] Marc Eder and Jan-Michael Frahm. Convolutions on spherical images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, June 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.631, + 0.469, + 0.686 + ], + "angle": 0, + "content": "[10] Marc Eder, Mykhailo Shvets, John Lim, and Jan-Michael Frahm. Tangent images for mitigating spherical distortion. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12423-12431, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.688, + 0.469, + 0.729 + ], + "angle": 0, + "content": "[11] Andrea Fusiello, Emanuele Trucco, and Alessandro Verri. A compact algorithm for rectification of stereo pairs. Machine Vision and Applications, 12(1):16-22, 2000. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.732, + 0.469, + 0.786 + ], + "angle": 0, + "content": "[12] Adrien Gaidon, Qiao Wang, Yohann Cabon, and Eleonora Vig. Virtualworlds as proxy for multi-object tracking analysis. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4340-4349, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.789, + 0.469, + 0.842 + ], + "angle": 0, + "content": "[13] Wenliang Gao and Shaojie Shen. Dual-fisheye omnidirectional stereo. In 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 6715-6722, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[14] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, pages 3354-3361, 2012. 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "[15] Heiko Hirschmuller. Stereo processing by semiglobal matching and mutual information. IEEE Transactions on Pattern Analysis and Machine Intelligence, 30(2):328-341, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.137, + 0.892, + 0.205 + ], + "angle": 0, + "content": "[16] Christian Häne, Lionel Heng, Gim Hee Lee, Alexey Sizov, and Marc Pollefeys. Real-time direct dense matching on fisheye images using plane-sweeping stereo. In 2014 2nd International Conference on 3D Vision, volume 1, pages 57-64, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.209, + 0.892, + 0.264 + ], + "angle": 0, + "content": "[17] Carlos Jaramillo, Roberto G. Valenti, Ling Guo, and Jizhong Xiao. Design and analysis of a single—camera omnistereo sensor for quadrotor micro aerial vehicles (mavs). Sensors, 16(2), 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.267, + 0.892, + 0.322 + ], + "angle": 0, + "content": "[18] J. Kannala and S.S. Brandt. A generic camera model and calibration method for conventional, wide-angle, and fish-eye lenses. IEEE Transactions on Pattern Analysis and Machine Intelligence, 28(8):1335-1340, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.326, + 0.892, + 0.366 + ], + "angle": 0, + "content": "[19] Bogdan Khomutenko, Gaetan Garcia, and Philippe Martinet. An enhanced unified camera model. IEEE Robotics and Automation Letters, 1(1):137-144, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.369, + 0.892, + 0.452 + ], + "angle": 0, + "content": "[20] Laurent Kneip, Hongdong Li, and Yongduek Seo. Upnp: An optimal o(n) solution to the absolute pose problem with universal applicability. In David Fleet, Tomas Pajdla, Bernt Schiele, and Tinne Tuytelaars, editors, Computer Vision - ECCV 2014, pages 127-142. Springer International Publishing, 2014. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.455, + 0.892, + 0.537 + ], + "angle": 0, + "content": "[21] Ren Komatsu, Hiromitsu Fujii, Yusuke Tamura, Atsushi Yamashita, and Hajime Asama. \\(360^{\\circ}\\) depth estimation from multiple fisheye images with origami crown representation of icosahedron. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 10092-10099, 2020. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.541, + 0.892, + 0.596 + ], + "angle": 0, + "content": "[22] Jiankun Li, Peisen Wang, Pengfei Xiong, Tao Cai, Ziwei Yan, Lei Yang, Jiangyu Liu, Haoqiang Fan, and Shuaicheng Liu. Practical stereo matching via cascaded recurrent network with adaptive correlation, 2022. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.6, + 0.892, + 0.641 + ], + "angle": 0, + "content": "[23] Weiming Li and Y. F. Li. Single-camera panoramic stereo imaging system with a fisheye lens and a convex mirror. Opt. Express, 19(7):5855-5867, Mar 2011. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.644, + 0.892, + 0.698 + ], + "angle": 0, + "content": "[24] Lahav Lipson, Zachary Teed, and Jia Deng. Raft-stereo: Multilevel recurrent field transforms for stereo matching. In 2021 International Conference on 3D Vision (3DV), pages 218-227, 2021. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.702, + 0.892, + 0.756 + ], + "angle": 0, + "content": "[25] Bin Lv, Hao Xu, Jianqing Wu, Yuan Tian, Sheng Tian, and Suoyao Feng. Revolution and rotation-based method for roadside lidar data integration. Optics & Laser Technology, 119:105571, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.76, + 0.892, + 0.842 + ], + "angle": 0, + "content": "[26] Nikolaus Mayer, Eddy Ilg, Philip Häusser, Philipp Fischer, Daniel Cremers, Alexey Dosovitskiy, and Thomas Brox. A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4040-4048, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[27] Christopher Mei and Patrick Rives. Single view point omnidirectional camera calibration from planar grids. In Proceedings 2007 IEEE International Conference on Robotics and Automation, pages 3945-3950, 2007. 2, 4" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21537" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[28] M. Menze, C. Heipke, and A. Geiger. Joint 3d estimation of vehicles and scene flow. ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences, II-3/W5:427-434, 2015. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.47, + 0.218 + ], + "angle": 0, + "content": "[29] Andreas Meuleman, Hyeonjoong Jang, Daniel S. Jeon, and Min H. Kim. Real-time sphere sweeping stereo from multiview fisheye images. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11418-11427, 2021. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.22, + 0.469, + 0.276 + ], + "angle": 0, + "content": "[30] Davide Scaramuzza, Agostino Martinelli, and Roland Siegwart. A toolbox for easily calibrating omnidirectional cameras. In 2006 IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 5695-5701, 2006. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.277, + 0.469, + 0.318 + ], + "angle": 0, + "content": "[31] Johannes Schneider, Cyril Stachniss, and Wolfgang Forstner. On the accuracy of dense fisheye stereo. IEEE Robotics and Automation Letters, 1(1):227-234, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.32, + 0.469, + 0.375 + ], + "angle": 0, + "content": "[32] Miriam Schonbein and Andreas Geiger. Omnidirectional 3d reconstruction in augmented manhattan worlds. In 2014 IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 716-723, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.377, + 0.469, + 0.459 + ], + "angle": 0, + "content": "[33] Xiao Song, Xu Zhao, Hanwen Hu, and Liangji Fang. Edgestereo: A context integrated residual pyramid network for stereo matching. In C.V. Jawahar, Hongdong Li, Greg Mori, and Konrad Schindler, editors, Computer Vision – ACCV 2018, pages 20–35. Springer International Publishing, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.461, + 0.469, + 0.516 + ], + "angle": 0, + "content": "[34] Yu-Chuan Su and Kristen Grauman. Kernel transformer networks for compact spherical convolution. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9434-9443, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.518, + 0.469, + 0.559 + ], + "angle": 0, + "content": "[35] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. ECCV, 12347(1):402-419, 2020. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.561, + 0.469, + 0.615 + ], + "angle": 0, + "content": "[36] Vladyslav Usenko, Nikolaus Demmel, and Daniel Cremers. The double sphere camera model. In 2018 International Conference on 3D Vision (3DV), pages 552-560, 2018. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.618, + 0.469, + 0.7 + ], + "angle": 0, + "content": "[37] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Ilia Polosukhin. Attention is all you need. In Proceedings of the 31st International Conference on Neural Information Processing Systems, NIPS'17, pages 6000-6010, Red Hook, NY, USA, 2017. Curran Associates Inc. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.702, + 0.469, + 0.757 + ], + "angle": 0, + "content": "[38] Hengli Wang, Rui Fan, Peide Cai, and Ming Liu. Pvstereo: Pyramid voting module for end-to-end self-supervised stereo matching. IEEE Robotics and Automation Letters, 6(3):4353-4360, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.759, + 0.469, + 0.815 + ], + "angle": 0, + "content": "[39] Changhee Won, Jongbin Ryu, and Jongwoo Lim. Omnimvss: End-to-end learning for omnidirectional stereo matching. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 8986-8995, 2019. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.817, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[40] Changhee Won, Jongbin Ryu, and Jongwoo Lim. Sweepnet: Wide-baseline omnidirectional depth estimation. In 2019 International Conference on Robotics and Automation (ICRA), pages 6073-6079, 2019. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[41] Changhee Won, Jongbin Ryu, and Jongwoo Lim. End-to-end learning for omnidirectional stereo matching with un" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "certainty prior. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(11):3850-3862, 2021. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[42] Changhee Won, Hochang Seok, Zhaopeng Cui, Marc Pollefeys, and Jongwoo Lim. Omnislam: Omnidirectional localization and dense mapping for wide-baseline multicamera systems. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 559-566, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.26 + ], + "angle": 0, + "content": "[43] Hao Xu, Yichen Zhang, Boyu Zhou, Luqi Wang, Xinjie Yao, Guotao Meng, and Shaojie Shen. Omni-swarm: A decentralized omnidirectional visual-inertial-uwb state estimation system for aerial swarm. arXiv preprint arXiv:2103.04131, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.262, + 0.892, + 0.345 + ], + "angle": 0, + "content": "[44] Feihu Zhang, Xiaojuan Qi, Ruigang Yang, Victor Prisacariu, Benjamin Wah, and Philip Torr. Domain-invariant stereo matching networks. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, Computer Vision - ECCV 2020, pages 420-439. Springer International Publishing, 2020. 6" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "21538" + } + ] +] \ No newline at end of file diff --git a/2023/OmniVidar_ Omnidirectional Depth Estimation From Multi-Fisheye Images/e34f49e7-aba6-408f-9647-b38f210ab7c3_origin.pdf b/2023/OmniVidar_ Omnidirectional Depth Estimation From Multi-Fisheye Images/e34f49e7-aba6-408f-9647-b38f210ab7c3_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f6a0ee23deae60433ed5431838aed5271b018eee --- /dev/null +++ b/2023/OmniVidar_ Omnidirectional Depth Estimation From Multi-Fisheye Images/e34f49e7-aba6-408f-9647-b38f210ab7c3_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bfc6bb919ce97a154230b85055329ebb8e5d1c8f045bc0b436841dc5509ded3e +size 2560579 diff --git a/2023/OmniVidar_ Omnidirectional Depth Estimation From Multi-Fisheye Images/full.md b/2023/OmniVidar_ Omnidirectional Depth Estimation From Multi-Fisheye Images/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f9ef2563b71eae1312727bea2b89d4ea2dbb7aa0 --- /dev/null +++ b/2023/OmniVidar_ Omnidirectional Depth Estimation From Multi-Fisheye Images/full.md @@ -0,0 +1,354 @@ +# OmniVidar: Omnidirectional Depth Estimation from Multi-Fisheye Images + +Sheng Xie $^{1}$ , Daochuan Wang $^{2}$ , Yunhui Liu $^{1,2,3,*}$ +Harbin Institute of Technology, Shenzhen $^{1}$ +Hong Kong Center for Logistics Robotics $^{2}$ , The Chinese University of Hong Kong $^{3}$ +xiesheng@stu.hit.edu.cn, wangdaochuan@hotmail.com, yhliu@mae.cuhk.edu.hk + +# Abstract + +Estimating depth from four large field of view (FoV) cameras has been a difficult and understudied problem. In this paper, we proposed a novel and simple system that can convert this difficult problem into easier binocular depth estimation. We name this system OmniVidar, as its results are similar to LiDAR, but rely only on vision. OmniVidar contains three components: (1) a new camera model to address the shortcomings of existing models, (2) a new multi-fisheye camera based epipolar rectification method for solving the image distortion and simplifying the depth estimation problem, (3) an improved binocular depth estimation network, which achieves a better balance between accuracy and efficiency. Unlike other omnidirectional stereo vision methods, OmniVidar does not contain any 3D convolution, so it can achieve higher resolution depth estimation at fast speed. Results demonstrate that OmniVidar outperforms all other methods in terms of accuracy and performance. + +# 1. Introduction + +Depth estimation from images is an important research field in computer vision, as it enables the acquisition of depth information with low-cost cameras for a wide range of applications. Traditional stereo cameras with pinhole lenses are limited in their FoV. However, many scenarios require an omnidirectional depth map, such as autonomous driving [42] and robot navigation [13, 43]. Although there are active sensors available that can provide omnidirectional depth information, such as LiDAR [25], their high cost makes them less accessible than stereo vision. Passive sensors, such as RGB cameras, are a common choice for depth estimation due to their low cost, lightweight, and low power consumption. To increase the FoV, fisheye lenses are often introduced into stereo vision setups. + +Over the past few decades, various methods have been proposed for depth estimation using fisheye cameras. These + +![](images/fa083f2bc73257d8f36f48ba7acb606040fcee25bc53124d9d7470654d4883f3.jpg) + +![](images/18c7381f79988b5a4fe9a0e18fe7543499c067769a1eebdfc01ce33e36c8c321.jpg) + +![](images/268fff970f211c87e704f9b23d693668f8f8ac2cd30a504e4618a6190ca45e17.jpg) + +![](images/b412a97201711db98241d7682ddc02c6d040baeb269d64fd1bd176fcc1137144.jpg) +Figure 1. Our prototype that built with four $250^{\circ}$ fisheye cameras and its results of dense inverse distance map and cloud points. It can get great depth estimation results in real scenes and achieve real-time performance on modern GPU. + +include the binocular fisheye system [31], up-down fisheye system [13], and catadiptric cameras [17, 23, 32]. However, all of these approaches have limitations. The binocular fisheye system cannot provide an omnidirectional perception. The up-down fisheye system and catadiptric cameras can offer horizontal $360^{\circ}$ depth perception, but their vertical FoV is limited. Furthermore, catadiptric cameras tend to be bulky, challenging to calibrate, and prone to errors. It turns out that the best choice for omnidirectional depth estimation is a system consisting of four cameras with extremely wide FoV ( $>180^{\circ}$ ). This system enables + +$360^{\circ}$ depth estimation both horizontally and vertically, and is light-weight, convenient to calibrate and maintain. Several studies [21,29,39,41] have shown excellent results with this approach. However, these methods often use many 3D convolutions, resulting in low efficiency. To adapt to limited memory, the images must be down-sampled, leading to a lower resolution of the output depth map. Furthermore, many of these approaches perform depth estimation using the original fisheye image without explicitly processing image distortion, which leads to the construction of a tedious and complicated cost volume and can result in reduced accuracy. + +Moreover, due to the complex imaging process of large FoV fisheye lenses, handling fisheye image distortion using mathematical formulas can be a challenging task. While several excellent large FoV fisheye camera models have been proposed [2, 18, 19, 27, 30, 36], our experiments have shown that none of these models can accurately approximate the imaging process, and we get less satisfactory depth estimation results when using them. + +Inspired by the above observations, we propose OmniVidar, a novel and simple multi-fisheye omnidirectional depth estimation system, as shown in Figure 2. OmniVidar contains three components. Firstly, we improve the DSCM [36] and propose a new camera model, named Triple Sphere Camera Model (TSCM), which can better approximate the imaging process and achieve the best accuracy. Then we propose an epipolar rectification algorithm designed for multi-fisheye camera system. We leverage a cubic-like projection approach to transform the four fisheye camera systems into four binocular camera systems and then conduct epipolar rectification on each binocular system. This method solves the distortion issue and reduces the complex multi-fisheye omnidirectional depth estimation problem to a much simpler binocular depth estimation problem. Finally, we design a lightweight binocular depth estimation network based on RAFT [35]. We add Transformer encoder [37] into feature extraction in RAFT to combine the advantages of Transformer and GRU, and it's easy to balance the accuracy and efficiency. + +We compare OmniVidar with existing methods on several datasets. The results demonstrate that our method outperforms all the others in terms of speed and accuracy, achieving State-of-The-Art performance. + +# 2. Related Work + +# 2.1. Multiview Fisheye Stereo + +Gao et al. [13] mount two large FoV cameras $(245^{\circ})$ vertically and reversely to obtain a $360^{\circ}$ horizontal and $65^{\circ}$ vertical FoV overlap and use this system in UAV for depth perception [43]. Won et al. [40] use four cameras with $220^{\circ}$ FoV and propose SweepNet algorithm. They propose + +Sphere-Sweep method modified from Plane-Sweep [16] to construct the cost volume and then use the cost aggregation algorithm in SGM [15] to obtain depth. After this, Won et al. propose OmniMVS [39, 41], which uses 3D convolution instead of SGM for depth regression and achieves higher accuracy. Komatsu et al. [21] propose CrownConv360. They improve OmniMVS by using the icosahedral projection in the feature extraction and depth regression stage for distortion reduction. Instead of using neural network, Meuleman et al. [29] choose to investigate cost aggregation and propagation in traditional stereo matching. And they achieve real-time high-resolution depth estimation at 29 fps on the embedded device TX2. However, the algorithm is only suitable for short camera baseline. + +All of the above works have more or fewer deficiencies. We propose OmniVidar, an omnidirectional depth estimation method with high accuracy, which can run in real-time and solve the fisheye image distortion problem with good generalization in various scenes. + +# 2.2. Fisheye Image Distortion + +To solve the distortion problem, a camera model suitable for fisheye is required, and the distortion should be handled explicitly in depth estimation. + +Scaramuzzi et al. [30] propose a high-order polynomial-based camera projection model, which is highly generalizable and can be widely used for various catadioptric and fisheye cameras. But it has no closed-form solution, so the inverse projection equation needs to be fitted with a high-order polynomial, which introduces errors. Mei et al. [27] propose the MUCM, which is an improvement of the UCM [2] to achieve higher calibration accuracy and propose a novel intrinsic parameters initialization method. Usenko et al. [36] propose the DSCM, which has only one more intrinsic parameter than the MUCM model, but the accuracy is substantially improved. + +Su et al. [34] propose a mini-network that learns to modify the shape of filters based on location. Coors et al. [8] chose to use a camera projection model to shape the convolution kernel to directly address distortion. An alternative idea was chosen by Cohen et al. [7] who define a form of convolution on the Lie group $SO(3)$ that convolves over the rotation space rather than the translation space. Eder et al. [9] compared the performance of three approaches: equirectangular projection, cubic projection, and 7-level subdivided icosahedral projection, on depth estimation and semantic segmentation tasks. The results showed that icosahedral projection achieved the best results in all tasks. Komatsu et al. [21] used icosahedral projection in multi-fisheye omnidirectional stereo vision to improve the depth estimation results of OmniMVS. Eder et al. [10] proposed a tangent plane projection modified from the icosahedron projection. + +![](images/48c296a724e0d882bd1f0f98b4f5061e8fa03c27f3d291ac386dfef14525d566.jpg) +Figure 2. System overview. Our system contains three parts: multi-fisheye calibration, multi-fisheye epipolar rectification, and binocular depth estimation. In epipolar rectification, the images in the left column are original. The images in the middle column are the images after abstracting each fisheye camera into two pinhole cameras, where the two images connected by a red dotted arrow are a pair of binocular cameras. The images are rearranged and combined to obtain the right column. + +The above algorithms have a common feature that they are all designed for the monocular image. Thus they perform well on tasks that require only one camera, yet not much good for multi-fisheye tasks. Therefore, to address this problem, we propose a distortion removal method designed for multi-fisheye systems, which greatly simplifies the depth estimation problem at the same time. + +# 2.3. RAFT in Stereo Vision + +Zachary et al. [35] propose Recurrent All-Pairs Field Transforms (RAFT), a highly efficient and accurate deep network architecture for optical flow. RAFT uses modified GRU blocks to iteratively refine optical flow and shares weights between refinement units. Lahav et al. [24] apply the RAFT to binocular rectified stereo vision and introduce multi-level convolutional GRUs, which can more efficiently propagate information across the image. Wang et al. [38] modify RAFT and propose PVStereo which greatly outperforms other self-supervised stereo matching approaches. Li et al. [22] combine RAFT and Transformer and propose CREStereo. They add Transformer encoder in image feature extraction and repeat the cross-attention between left and right images in the iterative update stage. This method performs well in various binocular datasets with excellent generalization. + +We imitate CREStereo and improve it based on RAFT to obtain a binocular stereo vision depth estimation network with a better trade-off between efficiency and accuracy. + +# 3. Methods + +The overall flow of our method is shown in Figure 2 and is divided into three parts: camera calibration, epipolar rectification, and depth estimation. + +![](images/1718c5435412406974cc383527dbfba4d4f2145542d5570cf5c08cd8e0f339f3.jpg) +Figure 3. Our Triple Sphere Camera Model. First, point $P$ is projected onto the first sphere and then onto the second sphere, then the third. The second sphere is shifted w.r.t. the first sphere by $\xi$ , and the third is shifted w.r.t. the second sphere by $\lambda$ . Finally, the point is projected onto the image plane of a pinhole camera which is shifted by $\frac{\alpha}{1 - \alpha}$ from the third sphere. + +# 3.1. Camera Calibration + +We propose the Triple Sphere Camera Model, which can better fit the imaging process of large FoV fisheye cameras + +and has a closed-form unprojection solution. + +As shown in Figure 3, our projection model considers that the incident light is refracted three times, and the displacements of the three unit spherical centers are $\xi$ and $\lambda$ . After three refractions, the incident light is finally projected to the image plane according to the pinhole camera model, and the displacement of the pinhole camera model's optical center from the third unit sphere is $\frac{\alpha}{1 - \alpha}$ . Therefore, our model has totally 7 camera internal parameters: $f_{x}, f_{y}, c_{x}, c_{y}, \xi, \lambda, \alpha$ . The projection equations are defined as follows. + +$$ +\pi (P, i) = \frac {1}{\zeta} \left[ \begin{array}{c c c} f _ {x} & 0 & c _ {x} \\ 0 & f _ {y} & c _ {y} \end{array} \right] \left[ \begin{array}{l} X \\ Y \\ \zeta \end{array} \right] \tag {1} +$$ + +$$ +\zeta = Z + \xi d _ {1} + \lambda d _ {2} + \frac {\alpha}{1 - \alpha} d _ {3} \tag {2} +$$ + +$$ +d _ {1} = \sqrt {X ^ {2} + Y ^ {2} + Z ^ {2}} \tag {3} +$$ + +$$ +d _ {2} = \sqrt {X ^ {2} + Y ^ {2} + (\xi d _ {1} + Z) ^ {2}} \tag {4} +$$ + +$$ +d _ {3} = \sqrt {X ^ {2} + Y ^ {2} + (\lambda d _ {2} + \xi d _ {1} + Z) ^ {2}} \tag {5} +$$ + +where $i$ is the vector of intrinsic parameters, $\pi$ is the projection function. + +A set of 3D points that results in valid projection is expressed as follows: + +$$ +\Omega = \left\{x \in \mathbb {R} ^ {3} | z > - w _ {2} d _ {1} \right\} \tag {6} +$$ + +$$ +w _ {2} = \frac {\xi + \lambda + w _ {1}}{\sqrt {1 + (\xi + \lambda) ^ {2} + 2 w _ {1} (\xi + \lambda)}} \tag {7} +$$ + +$$ +w _ {1} = \frac {\alpha}{1 - \alpha} \tag {8} +$$ + +The unprojection function is computed as follows: + +$$ +\pi^ {- 1} (p, i) = \mu \left[ \begin{array}{l} \eta \gamma x \\ \eta \gamma y \\ m _ {z} \end{array} \right] - \left[ \begin{array}{l} 0 \\ 0 \\ \xi \end{array} \right] \tag {9} +$$ + +$$ +\mu = \xi m _ {z} + \sqrt {\xi^ {2} m _ {z} ^ {2} - \xi^ {2} + 1} \tag {10} +$$ + +$$ +m _ {z} = \eta (\gamma - \zeta) - \lambda \tag {11} +$$ + +$$ +\eta = \lambda (\gamma - \zeta) + \sqrt {\lambda^ {2} (\gamma - \zeta) ^ {2} - \lambda^ {2} + 1} \tag {12} +$$ + +$$ +\gamma = \frac {\zeta + \sqrt {1 + (1 - \zeta^ {2}) \left(x ^ {2} + y ^ {2}\right)}}{x ^ {2} + y ^ {2} + 1} \tag {13} +$$ + +$$ +\zeta = \left\{ \begin{array}{l l} \frac {\alpha}{1 - \alpha} & \text {i f} \alpha \leq 0. 5 \\ \frac {1 - \alpha}{\alpha} & \text {i f} \alpha > 0. 5 \end{array} \right. \tag {14} +$$ + +where $(x,y)$ is the normalized coordinate. + +We calibrate the intrinsic and extrinsic parameters of multi-fisheye camera system using a highly accurate planar checkerboard. The goal of the calibration is to minimize the reprojection error of the corner points in all pictures. For each image $n$ in the calibration sequence, the + +![](images/5be166fd2ac635187db3de74983f1bad2c25fc0222d0602184ff9cbd1aca0070.jpg) +Figure 4. The multi-fisheye epipolar rectification schematic. Each fisheye camera is converted into two pinhole cameras with an angle of $90^{\circ}$ . Then the system becomes four binocular systems. + +corner detector can obtain the projection point $u_{nk}$ of the $k$ th corner point $x_{k}$ . The coordinate of $u_{nk}$ is related to the camera intrinsic and extrinsic parameters. Let $s = [i, T_{cam_0}, T_{cam_1}, \dots, T_{cam_n}]$ be the parameter to optimize. We can construct the nonlinear optimization problem as follows: + +$$ +s = \arg \min _ {s ^ {\prime}} \sum_ {n = 0} ^ {N} \sum_ {k \in K} \rho \left(\left(\pi \left(T _ {c a m _ {n}} x _ {k}, \boldsymbol {i}\right) - u _ {n k}\right) ^ {2}\right) \tag {15} +$$ + +where $T_{cam_n} \in SE(3)$ is the transformation from the coordinate frame of the calibration grid to the camera coordinate frame for image $n$ . $K$ is a set of detected corner points for the image $n$ and $\rho$ is the robust Huber norm. + +Due to the highly nonlinear nature of the image, the above optimization problem requires a good initial value. We initialize the intrinsic parameters using the previously proposed method [27] and find initial poses using the UPnP algorithm [20]. After obtaining the initial values of the parameters, we completed the optimization operation using Ceres Solver [1]. + +# 3.2. Multi-Fisheye Camera Epipolar Rectify + +We propose a simple and effective undistortion method that cleverly transforms the omnidirectional depth estimation problem into a binocular depth estimation problem. And the method can be used in the system with any number of fisheye cameras. As shown in Figure 4, we abstract each fisheye camera as two pinhole cameras with $90^{\circ}$ angle1. Then the four fisheye cameras system becomes four independent binocular systems. Each binocular systems can be rectified through epipolar rectification. We modified the classical epipolar rectification algorithm proposed by Fusiello [11] for our system. + +![](images/d47d6dd39f947d590957077f36f3647181aa27b3741cb84ea7d73ba868e33b12.jpg) +Figure 5. The structure of our proposed binocular depth estimation network. A pair of stereo images are fed into feature extraction networks. Then the output features are calculated by self-attention and cross-attention, and after doing self-attention, the features of left image are used for context information for latter recurrent update stage. Then the left and right features are multiplied to construct cost volume. Finally use the disparity update model same as RAFT [35]. + +Firstly, the rotation matrix of the cameras in binocular system is calculated. Let the optical centers of the left camera and right camera be $C_1 = \left[x_1 y_1 z_1\right]^T$ , $C_2 = \left[x_2 y_2 z_2\right]^T$ , the rotation matrix can be calculated as follows: + +$$ +R = \left[ \begin{array}{l l l} r _ {1} ^ {T} & r _ {2} ^ {T} & r _ {3} ^ {T} \end{array} \right] \tag {16} +$$ + +$$ +r _ {1} = \left[ \begin{array}{l l l} x _ {2} - x _ {1} & y _ {2} - y _ {1} & z _ {2} - z _ {1} \end{array} \right]. n o r m a l i z e () \tag {17} +$$ + +$$ +r _ {3} = \left[ - \left(z _ {2} - z _ {1}\right) \quad 0 \quad x _ {2} - x _ {1} \right]. n o r m a l i z e () \tag {18} +$$ + +$$ +r _ {2} = r _ {3} \times r _ {1} \tag {19} +$$ + +Then remap the image for epipolar rectification. Taking the left camera in one binocular system as an example. Let the rotation matrix of the corresponding fisheye camera be $R$ , the optical center be $C$ . For a certain spatial point $P = (X, Y, Z, 1)$ , its projection point on normalized plane is $p = (u, v, 1)$ . The projection equation is shown below. + +$$ +\mu \left[ \begin{array}{c} u \\ v \\ f (u, v) \end{array} \right] = [ R - R C ] \left[ \begin{array}{l} X \\ Y \\ Z \\ 1 \end{array} \right] \tag {20} +$$ + +Let the rotation matrix of the rectified pinhole camera be $R'$ , and the corresponding normalized plane coordinate be $(u', v', 1)$ . Then the projection equation is shown below. + +$$ +\mu^ {\prime} \left[ \begin{array}{l} u ^ {\prime} \\ v ^ {\prime} \\ 1 \end{array} \right] = \left[ \begin{array}{l l} R ^ {\prime} & - R ^ {\prime} C \end{array} \right] \left[ \begin{array}{l} X \\ Y \\ Z \\ 1 \end{array} \right] \tag {21} +$$ + +The relationship between fisheye image and virtual binocular image can be obtained, see equation (22). + +$$ +\lambda \left[ \begin{array}{l} u ^ {\prime} \\ v ^ {\prime} \\ 1 \end{array} \right] = \left[ \begin{array}{l l l} R ^ {\prime} & - R ^ {\prime} C \end{array} \right] \left[ \begin{array}{l l} R & - R C \end{array} \right] ^ {- 1} \left[ \begin{array}{c} u \\ v \\ f (u, v) \end{array} \right] \tag {22} +$$ + +That is, for each pixel in the original fisheye image, the pixel coordinate in the pinhole image can be calculated. We can use pixel interpolation to obtain the rectified image. + +As shown in Figure 2, it can be seen that after epipolar rectification, the four fisheye images are converted into a pair of binocular images, where each image is stitched together from the four virtual pinhole camera images. A simple binocular depth estimation algorithm is then used. + +# 3.3. Stereo Depth Estimation + +Since our system needs to estimate the depth values of four binocular systems simultaneously, which is very computationally intensive, our primary goal is not to design a binocular algorithm with the highest accuracy, but to minimize the accuracy loss while ensuring real-time performance. + +The structure of our binocular depth estimation network is shown in Figure 5. We added Transformer [37] with self-attention and cross-attention in the image feature extraction stage. In binocular depth estimation, limited by computational performance and graphics card memory, many algorithms have to limit the matching search range of pixels. In this paper, we design an ingenious cost-volume calculation that can extend the estimable disparity range to the image width, while also reducing the computational complexity. + +![](images/d0a7dd8fefda34d19958a3fd25d587fa57b38cdefe80187c3c556751ea4a75bc.jpg) +Figure 6. The depth estimation comparison between ours and other methods. The final row is the results from our method, and the column annotations indicate the dataset source. + +Given two images $I_1 \in \mathbb{R}^{H \times W \times C}$ , $I_2 \in \mathbb{R}^{H \times W \times C}$ , we need to compute the cost volume $C(I_1, I_2) \in \mathbb{R}^{H \times W \times W}$ . We found that the cost volume can be obtained by single matrix operation as follows. + +$$ +C \left(I _ {1}, I _ {2}\right) = I _ {1} \in \mathbb {R} ^ {H \times W \times C} \cdot I _ {2} ^ {T} \in \mathbb {R} ^ {H \times C \times W} \tag {23} +$$ + +To enable the network to learn information at multiple scales, we used the cost volume pooling opertaion in RAFT-Stereo [24]. + +# 4. Experiments + +# 4.1. Implementation + +Our depth estimation network structure is shown in Figure 5. The number of channels in our network is identical to RAFT [35]. The feature encoder consists of 6 residual blocks, 2 at $1/2$ resolution, 2 at $1/4$ resolution, and 2 at $1/8$ resolution. The number of channels is 64, 128, and 256, respectively. The $1/8$ resolution feature map is then fed into the Transformer encoder [37]. We set the number of heads in attention to 8. The GRU iterative update part is identical to RAFT, and more details can be found in [35]. In addition, we replace all the normalization operators in RAFT with domain normalize [44]. + +Since we convert the multi-fisheye cameras to four binocular cameras, we can use a large number of public binocular datasets for training, instead of multi-fisheye + +datasets, which are very limited. Similar to the training strategy in CREStereo [22], we collected most of public binocular datasets, including SceneFlow [26], Sintel [4], CREStereo [22], KITTI [14, 28], InStereo2K [3], and Virtual KITTI [12]. The images are uniformly randomly cropped with a fixed resolution during the training process to solve the problem of different image resolutions in different datasets. + +We use a single NVIDIA RTX2080Ti for training, the batch size is set to 4, the learning rate is 0.0004, and we use Adam to train 200 epochs to get the final network weights. + +To test the performance of OmniVidar in real scenarios, we built a test setup. We used four leopard LI-USB30-OV10640-490-GMSL cameras with an image resolution of $1280 \times 1024$ , paired with a fisheye lens with $250^{\circ}$ FoV. The four cameras look in four directions, front and back, left and right, as shown in Figure 1. We use the device to capture indoor, corridor, and outdoor images, which can fully test the performance of OmniVidar in real scenes. + +# 4.2. Evaluation + +We perform the comparison on the Omnihouse, OmniThings provided by Won et al. [40, 41], and the datasets provided by Meuleman et al. [29]. Our OmniVidar has only been trained on the binocular dataset mentioned in Section 4.1, not on the test dataset, and the methods involved in the + +
MethodMeuleman et al. [29]Omnihouse [40]Omnithings [39]
bad 1.0RMSEMAEbad 1.0RMSEMAEbad 1.0RMSEMAE
OmniVidar(Ours)18.4990.9210.8034.8861.0021.04211.1990.9580.461
OmniMVS-ft [41]60.3722.3061.73226.3570.9290.70559.7532.2701.723
CrownConv360 [21]69.1194.0203.00585.55911.0776.72993.9469.5456.518
Meuleman et al. [29]44.1111.6641.27311.3040.6710.41065.3482.5501.941
+ +comparison all use the training method corresponding to the original paper. + +As can be seen in the Figure 6, although OmniVidar is not trained and fine-tuned on the above datasets, its depth estimation accuracy on these datasets far exceeds that of other methods, with accurate depth estimation in weakly textured regions. + +Table 1. Quantitative comparison of algorithm accuracy on several datasets. The less the number is, the more accurate the method is. + +
MethodRuntimeMemoryParameters
OmniVidar (Ours)66 ms3.7 GB4.8 MB
OmniMVS-ft [41]110 ms6.8 GB43.7 MB
CrownConv360 [21]623 ms6.3 GB16.7 MB
+ +The analysis of the quantitative results are shown in the Table 1 and Table 2. It can be seen that OmniVidar has the least memory consumption, the shortest time consumption and the highest accuracy compared to other deep learning based solutions. Compared with non-deep learning methods, the accuracy is much higher than it, despite the relatively low efficiency. + +![](images/85cf9a42c4ee075faa9e176a0cb7ed10e703a63be27c925a84ed1d61cf06ca28.jpg) +Figure 7. Plot comparing the accuracy and efficiency evaluated in KITTI-2015 [28]. It can be seen that OmniVidar achieves real-time performance while maintain the accuracy. + +We also compared the binocular depth estimation network in OmniVidar with other binocular depth estimation networks. We chose the KITTI-2015 dataset [28] for the evaluation and the results are shown in Figure 7. We compare our method with CREStereo [22], LEAStereo [6], PSMNet [5], EdgeStereo [33], RAFT-Stereo [24]. It can see that our scheme achieves a better balance between accuracy and efficiency. + +# 4.3. Ablation + +Camera Model: We compare the effect of omnidirectional depth estimation under several camera models. + +![](images/31cd0fae314bb9ccffb1da0be9141bb5eb358bb1ab89f5367e316df0a8556e76.jpg) +Original fisheye image + +![](images/890c6965c93dcc230f404a60d00e96a5f7049674c33fd80d7b8c0fc18e6fa802.jpg) +Scaramuzza et al. [30] + +![](images/5e3637eb032df4968b81aff04cae972c20c394d0371b66f2206e6a24a1bbd77f.jpg) +DSCM [36] +Figure 8. Depth estimation results using different camera projection models. Our TSCM can best remove the fisheye distortion and obtain the most accurate result. + +![](images/87051fe40c3284da5f01fc05f5c0db868ca32308c169476eb8b97136391e35a1.jpg) +TSCM(Ours) + +As shown in Figure 8, the distortion of the fisheye lens is not sufficiently handled in the DSCM [36] and the model proposed by Scaramuzza et al. [30], which affects the accuracy of depth estimation and leads to serious distortion of objects in the point cloud. In contrast, the quality of the point cloud obtained by the TSCM is much better, with flat wall surfaces and more accurate distance measurements. + +Table 2. Efficiency comparison with other deep learning methods. + +
FoV2 sphere3 sphere (Ours)4 sphere
195°0.1530 px0.1528 px0.1528 px
220°0.2031 px0.1970 px0.1952 px
250°0.1365 px0.1255 px0.1222 px
+ +Table 3. Mean reprojection error for camera models with different number sphere. + +We also compared the accuracy with different number of the refract sphere. Table 3 shows the results. It can be seen + +![](images/1b4dbb8e5fda9efd61d3eb358a2ebefc510f5f02bf2bd4b3136c4c036ba3e295.jpg) +Figure 9. The images captured with our prototype and the inverse depth map and cloud points obtained by OmniVidar. It can be seen that OmniVidar has great generalization in various scenes, and performs well in low-textured areas. + +that the camera model with 3 sphere can significantly improve the accuracy in large FoV scenes, while the improvement brought by more sphere is very limited. + +Distortion Handle: To verify whether the improvement of depth estimation accuracy is helped by removing the fisheye image distortion using epipolar rectification, we estimate the depth directly using the original fisheye image and compare it with the undistorted method. Unlike the undistorted binocular stereo, we use the Sphere-Sweep algorithm [40] to construct the cost volume. This step does not contain trainable parameters, so it does not affect the network itself and ensures the fairness of the comparison. + +The point cloud visualization of the results of the depth estimation of both is shown in Figure 8. It can be seen that when using original image, the depth information obtained from different cameras is not consistent. In contrast, the point cloud obtained by depth estimation after epipolar rectification shows the structural framework of the scene real- + +istrically and accurately. + +# 4.4. Test in Real + +We test the performance of the OmniVidar algorithm on a real dataset. Figure 9 shows the inverse depth maps and point clouds obtained by OmniVidar on the real dataset. It can be seen that our network has a strong generalization to obtain high-quality point clouds on real datasets. + +# 5. Conclusion + +In this paper, we propose a novel, simple and effective system OmniVidar, to address the inefficiency of current omnidirectional depth estimation methods due to the extensive use of 3D convolution and the poor accuracy due to the lack of explicit handling of distortion in fisheye images. Our method outperforms all other methods in terms of time consumption and accuracy. + +# References + +[1] Sameer Agarwal, Keir Mierle, and The Ceres Solver Team. Ceres solver. http://ceres-solver.org. 4 +[2] Simon Baker and Shree K. Nayar. A theory of single-viewpoint catadioptric image formation. Int. J. Comput. Vision, 35(2):175-196, nov 1999. 2 +[3] Wei Bao, Wei Wang, Yuhua Xu, Yulan Guo, Siyu Hong, and Xiaohu Zhang. Instereo2k: a large real dataset for stereo matching in indoor scenes. Science China Information Sciences, 63(11):212101, 2019. 6 +[4] D. J. Butler, J. Wulff, G. B. Stanley, and M. J. Black. A naturalistic open source movie for optical flow evaluation. In A. Fitzgibbon et al. (Eds.), editor, European Conf. on Computer Vision (ECCV), Part IV, LNCS 7577, pages 611-625. Springer-Verlag, Oct. 2012. 6 +[5] Jia-Ren Chang and Yong-Sheng Chen. Pyramid stereo matching network. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5410-5418, 2018. 7 +[6] Xuelian Cheng, Yiran Zhong, Mehrtash Harandi, Yuchao Dai, Xiaojun Chang, Tom Drummond, Hongdong Li, and Zongyuan Ge. Hierarchical neural architecture search for deep stereo matching. arXiv preprint arXiv:2010.13501, 2020.7 +[7] Taco S. Cohen, Mario Geiger, Jonas Kohler, and Max Welling. Spherical CNNs. In International Conference on Learning Representations, 2018. 2 +[8] Benjamin Coors, Alexandru Paul Condurache, and Andreas Geiger. Spherenet: Learning spherical representations for detection and classification in omnidirectional images. In Vittorio Ferrari, Martial Hebert, Cristian Sminchisescu, and Yair Weiss, editors, Computer Vision – ECCV 2018, pages 525–541. Springer International Publishing, 2018. 2 +[9] Marc Eder and Jan-Michael Frahm. Convolutions on spherical images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, June 2019. 2 +[10] Marc Eder, Mykhailo Shvets, John Lim, and Jan-Michael Frahm. Tangent images for mitigating spherical distortion. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12423-12431, 2020. 2 +[11] Andrea Fusiello, Emanuele Trucco, and Alessandro Verri. A compact algorithm for rectification of stereo pairs. Machine Vision and Applications, 12(1):16-22, 2000. 4 +[12] Adrien Gaidon, Qiao Wang, Yohann Cabon, and Eleonora Vig. Virtualworlds as proxy for multi-object tracking analysis. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4340-4349, 2016. 6 +[13] Wenliang Gao and Shaojie Shen. Dual-fisheye omnidirectional stereo. In 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 6715-6722, 2017. 1, 2 +[14] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, pages 3354-3361, 2012. 6 + +[15] Heiko Hirschmuller. Stereo processing by semiglobal matching and mutual information. IEEE Transactions on Pattern Analysis and Machine Intelligence, 30(2):328-341, 2008. 2 +[16] Christian Häne, Lionel Heng, Gim Hee Lee, Alexey Sizov, and Marc Pollefeys. Real-time direct dense matching on fisheye images using plane-sweeping stereo. In 2014 2nd International Conference on 3D Vision, volume 1, pages 57-64, 2014. 2 +[17] Carlos Jaramillo, Roberto G. Valenti, Ling Guo, and Jizhong Xiao. Design and analysis of a single—camera omnistereo sensor for quadrotor micro aerial vehicles (mavs). Sensors, 16(2), 2016. 1 +[18] J. Kannala and S.S. Brandt. A generic camera model and calibration method for conventional, wide-angle, and fish-eye lenses. IEEE Transactions on Pattern Analysis and Machine Intelligence, 28(8):1335-1340, 2006. 2 +[19] Bogdan Khomutenko, Gaetan Garcia, and Philippe Martinet. An enhanced unified camera model. IEEE Robotics and Automation Letters, 1(1):137-144, 2016. 2 +[20] Laurent Kneip, Hongdong Li, and Yongduek Seo. Upnp: An optimal o(n) solution to the absolute pose problem with universal applicability. In David Fleet, Tomas Pajdla, Bernt Schiele, and Tinne Tuytelaars, editors, Computer Vision - ECCV 2014, pages 127-142. Springer International Publishing, 2014. 4 +[21] Ren Komatsu, Hiromitsu Fujii, Yusuke Tamura, Atsushi Yamashita, and Hajime Asama. $360^{\circ}$ depth estimation from multiple fisheye images with origami crown representation of icosahedron. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 10092-10099, 2020. 2, 6, 7 +[22] Jiankun Li, Peisen Wang, Pengfei Xiong, Tao Cai, Ziwei Yan, Lei Yang, Jiangyu Liu, Haoqiang Fan, and Shuaicheng Liu. Practical stereo matching via cascaded recurrent network with adaptive correlation, 2022. 3, 6, 7 +[23] Weiming Li and Y. F. Li. Single-camera panoramic stereo imaging system with a fisheye lens and a convex mirror. Opt. Express, 19(7):5855-5867, Mar 2011. 1 +[24] Lahav Lipson, Zachary Teed, and Jia Deng. Raft-stereo: Multilevel recurrent field transforms for stereo matching. In 2021 International Conference on 3D Vision (3DV), pages 218-227, 2021. 3, 6, 7 +[25] Bin Lv, Hao Xu, Jianqing Wu, Yuan Tian, Sheng Tian, and Suoyao Feng. Revolution and rotation-based method for roadside lidar data integration. Optics & Laser Technology, 119:105571, 2019. 1 +[26] Nikolaus Mayer, Eddy Ilg, Philip Häusser, Philipp Fischer, Daniel Cremers, Alexey Dosovitskiy, and Thomas Brox. A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4040-4048, 2016. 6 +[27] Christopher Mei and Patrick Rives. Single view point omnidirectional camera calibration from planar grids. In Proceedings 2007 IEEE International Conference on Robotics and Automation, pages 3945-3950, 2007. 2, 4 + +[28] M. Menze, C. Heipke, and A. Geiger. Joint 3d estimation of vehicles and scene flow. ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences, II-3/W5:427-434, 2015. 6, 7 +[29] Andreas Meuleman, Hyeonjoong Jang, Daniel S. Jeon, and Min H. Kim. Real-time sphere sweeping stereo from multiview fisheye images. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11418-11427, 2021. 2, 6, 7 +[30] Davide Scaramuzza, Agostino Martinelli, and Roland Siegwart. A toolbox for easily calibrating omnidirectional cameras. In 2006 IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 5695-5701, 2006. 2, 7 +[31] Johannes Schneider, Cyril Stachniss, and Wolfgang Forstner. On the accuracy of dense fisheye stereo. IEEE Robotics and Automation Letters, 1(1):227-234, 2016. 1 +[32] Miriam Schonbein and Andreas Geiger. Omnidirectional 3d reconstruction in augmented manhattan worlds. In 2014 IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 716-723, 2014. 1 +[33] Xiao Song, Xu Zhao, Hanwen Hu, and Liangji Fang. Edgestereo: A context integrated residual pyramid network for stereo matching. In C.V. Jawahar, Hongdong Li, Greg Mori, and Konrad Schindler, editors, Computer Vision – ACCV 2018, pages 20–35. Springer International Publishing, 2019. 7 +[34] Yu-Chuan Su and Kristen Grauman. Kernel transformer networks for compact spherical convolution. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9434-9443, 2019. 2 +[35] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. ECCV, 12347(1):402-419, 2020. 2, 3, 5, 6 +[36] Vladyslav Usenko, Nikolaus Demmel, and Daniel Cremers. The double sphere camera model. In 2018 International Conference on 3D Vision (3DV), pages 552-560, 2018. 2, 7 +[37] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Ilia Polosukhin. Attention is all you need. In Proceedings of the 31st International Conference on Neural Information Processing Systems, NIPS'17, pages 6000-6010, Red Hook, NY, USA, 2017. Curran Associates Inc. 2, 5, 6 +[38] Hengli Wang, Rui Fan, Peide Cai, and Ming Liu. Pvstereo: Pyramid voting module for end-to-end self-supervised stereo matching. IEEE Robotics and Automation Letters, 6(3):4353-4360, 2021. 3 +[39] Changhee Won, Jongbin Ryu, and Jongwoo Lim. Omnimvss: End-to-end learning for omnidirectional stereo matching. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 8986-8995, 2019. 2, 6, 7 +[40] Changhee Won, Jongbin Ryu, and Jongwoo Lim. Sweepnet: Wide-baseline omnidirectional depth estimation. In 2019 International Conference on Robotics and Automation (ICRA), pages 6073-6079, 2019. 2, 6, 7, 8 +[41] Changhee Won, Jongbin Ryu, and Jongwoo Lim. End-to-end learning for omnidirectional stereo matching with un + +certainty prior. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(11):3850-3862, 2021. 2, 6, 7 +[42] Changhee Won, Hochang Seok, Zhaopeng Cui, Marc Pollefeys, and Jongwoo Lim. Omnislam: Omnidirectional localization and dense mapping for wide-baseline multicamera systems. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 559-566, 2020. 1 +[43] Hao Xu, Yichen Zhang, Boyu Zhou, Luqi Wang, Xinjie Yao, Guotao Meng, and Shaojie Shen. Omni-swarm: A decentralized omnidirectional visual-inertial-uwb state estimation system for aerial swarm. arXiv preprint arXiv:2103.04131, 2021. 1, 2 +[44] Feihu Zhang, Xiaojuan Qi, Ruigang Yang, Victor Prisacariu, Benjamin Wah, and Philip Torr. Domain-invariant stereo matching networks. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, Computer Vision - ECCV 2020, pages 420-439. Springer International Publishing, 2020. 6 \ No newline at end of file diff --git a/2023/OmniVidar_ Omnidirectional Depth Estimation From Multi-Fisheye Images/images.zip b/2023/OmniVidar_ Omnidirectional Depth Estimation From Multi-Fisheye Images/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..6d7e61243f06eeca3c2a3f678f2e6692fa4bc09c --- /dev/null +++ b/2023/OmniVidar_ Omnidirectional Depth Estimation From Multi-Fisheye Images/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cc9ba52e538ea8c0f45d20babdcb722711997d2f167e3a378dabc28f22bfd97e +size 943303 diff --git a/2023/OmniVidar_ Omnidirectional Depth Estimation From Multi-Fisheye Images/layout.json b/2023/OmniVidar_ Omnidirectional Depth Estimation From Multi-Fisheye Images/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..24e709be4ef4fcf563876b73524eeba97b06388c --- /dev/null +++ b/2023/OmniVidar_ Omnidirectional Depth Estimation From Multi-Fisheye Images/layout.json @@ -0,0 +1,7966 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 64, + 103, + 531, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 103, + 531, + 122 + ], + "spans": [ + { + "bbox": [ + 64, + 103, + 531, + 122 + ], + "type": "text", + "content": "OmniVidar: Omnidirectional Depth Estimation from Multi-Fisheye Images" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 95, + 142, + 499, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 142, + 499, + 200 + ], + "spans": [ + { + "bbox": [ + 95, + 142, + 499, + 200 + ], + "type": "text", + "content": "Sheng Xie" + }, + { + "bbox": [ + 95, + 142, + 499, + 200 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 95, + 142, + 499, + 200 + ], + "type": "text", + "content": ", Daochuan Wang" + }, + { + "bbox": [ + 95, + 142, + 499, + 200 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 95, + 142, + 499, + 200 + ], + "type": "text", + "content": ", Yunhui Liu" + }, + { + "bbox": [ + 95, + 142, + 499, + 200 + ], + "type": "inline_equation", + "content": "^{1,2,3,*}" + }, + { + "bbox": [ + 95, + 142, + 499, + 200 + ], + "type": "text", + "content": " \nHarbin Institute of Technology, Shenzhen" + }, + { + "bbox": [ + 95, + 142, + 499, + 200 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 95, + 142, + 499, + 200 + ], + "type": "text", + "content": " \nHong Kong Center for Logistics Robotics" + }, + { + "bbox": [ + 95, + 142, + 499, + 200 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 95, + 142, + 499, + 200 + ], + "type": "text", + "content": ", The Chinese University of Hong Kong" + }, + { + "bbox": [ + 95, + 142, + 499, + 200 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 95, + 142, + 499, + 200 + ], + "type": "text", + "content": " \nxiesheng@stu.hit.edu.cn, wangdaochuan@hotmail.com, yhliu@mae.cuhk.edu.hk" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "spans": [ + { + "bbox": [ + 143, + 227, + 192, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 252, + 290, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 252, + 290, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 252, + 290, + 456 + ], + "type": "text", + "content": "Estimating depth from four large field of view (FoV) cameras has been a difficult and understudied problem. In this paper, we proposed a novel and simple system that can convert this difficult problem into easier binocular depth estimation. We name this system OmniVidar, as its results are similar to LiDAR, but rely only on vision. OmniVidar contains three components: (1) a new camera model to address the shortcomings of existing models, (2) a new multi-fisheye camera based epipolar rectification method for solving the image distortion and simplifying the depth estimation problem, (3) an improved binocular depth estimation network, which achieves a better balance between accuracy and efficiency. Unlike other omnidirectional stereo vision methods, OmniVidar does not contain any 3D convolution, so it can achieve higher resolution depth estimation at fast speed. Results demonstrate that OmniVidar outperforms all other methods in terms of accuracy and performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 480, + 128, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 480, + 128, + 493 + ], + "spans": [ + { + "bbox": [ + 47, + 480, + 128, + 493 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 501, + 287, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 501, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 46, + 501, + 287, + 668 + ], + "type": "text", + "content": "Depth estimation from images is an important research field in computer vision, as it enables the acquisition of depth information with low-cost cameras for a wide range of applications. Traditional stereo cameras with pinhole lenses are limited in their FoV. However, many scenarios require an omnidirectional depth map, such as autonomous driving [42] and robot navigation [13, 43]. Although there are active sensors available that can provide omnidirectional depth information, such as LiDAR [25], their high cost makes them less accessible than stereo vision. Passive sensors, such as RGB cameras, are a common choice for depth estimation due to their low cost, lightweight, and low power consumption. To increase the FoV, fisheye lenses are often introduced into stereo vision setups." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 670, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 287, + 693 + ], + "type": "text", + "content": "Over the past few decades, various methods have been proposed for depth estimation using fisheye cameras. These" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 307, + 226, + 424, + 328 + ], + "blocks": [ + { + "bbox": [ + 307, + 226, + 424, + 328 + ], + "lines": [ + { + "bbox": [ + 307, + 226, + 424, + 328 + ], + "spans": [ + { + "bbox": [ + 307, + 226, + 424, + 328 + ], + "type": "image", + "image_path": "fa083f2bc73257d8f36f48ba7acb606040fcee25bc53124d9d7470654d4883f3.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 425, + 226, + 545, + 328 + ], + "blocks": [ + { + "bbox": [ + 425, + 226, + 545, + 328 + ], + "lines": [ + { + "bbox": [ + 425, + 226, + 545, + 328 + ], + "spans": [ + { + "bbox": [ + 425, + 226, + 545, + 328 + ], + "type": "image", + "image_path": "18c7381f79988b5a4fe9a0e18fe7543499c067769a1eebdfc01ce33e36c8c321.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 307, + 328, + 545, + 397 + ], + "blocks": [ + { + "bbox": [ + 307, + 328, + 545, + 397 + ], + "lines": [ + { + "bbox": [ + 307, + 328, + 545, + 397 + ], + "spans": [ + { + "bbox": [ + 307, + 328, + 545, + 397 + ], + "type": "image", + "image_path": "268fff970f211c87e704f9b23d693668f8f8ac2cd30a504e4618a6190ca45e17.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 307, + 396, + 545, + 506 + ], + "blocks": [ + { + "bbox": [ + 307, + 396, + 545, + 506 + ], + "lines": [ + { + "bbox": [ + 307, + 396, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 307, + 396, + 545, + 506 + ], + "type": "image", + "image_path": "b412a97201711db98241d7682ddc02c6d040baeb269d64fd1bd176fcc1137144.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 514, + 547, + 559 + ], + "lines": [ + { + "bbox": [ + 304, + 514, + 547, + 559 + ], + "spans": [ + { + "bbox": [ + 304, + 514, + 547, + 559 + ], + "type": "text", + "content": "Figure 1. Our prototype that built with four " + }, + { + "bbox": [ + 304, + 514, + 547, + 559 + ], + "type": "inline_equation", + "content": "250^{\\circ}" + }, + { + "bbox": [ + 304, + 514, + 547, + 559 + ], + "type": "text", + "content": " fisheye cameras and its results of dense inverse distance map and cloud points. It can get great depth estimation results in real scenes and achieve real-time performance on modern GPU." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": "include the binocular fisheye system [31], up-down fisheye system [13], and catadiptric cameras [17, 23, 32]. However, all of these approaches have limitations. The binocular fisheye system cannot provide an omnidirectional perception. The up-down fisheye system and catadiptric cameras can offer horizontal " + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": " depth perception, but their vertical FoV is limited. Furthermore, catadiptric cameras tend to be bulky, challenging to calibrate, and prone to errors. It turns out that the best choice for omnidirectional depth estimation is a system consisting of four cameras with extremely wide FoV (" + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "inline_equation", + "content": ">180^{\\circ}" + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": "). This system enables" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 135, + 712 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21529" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 46, + 72, + 289, + 216 + ], + "type": "text", + "content": " depth estimation both horizontally and vertically, and is light-weight, convenient to calibrate and maintain. Several studies [21,29,39,41] have shown excellent results with this approach. However, these methods often use many 3D convolutions, resulting in low efficiency. To adapt to limited memory, the images must be down-sampled, leading to a lower resolution of the output depth map. Furthermore, many of these approaches perform depth estimation using the original fisheye image without explicitly processing image distortion, which leads to the construction of a tedious and complicated cost volume and can result in reduced accuracy." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 217, + 289, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 217, + 289, + 312 + ], + "spans": [ + { + "bbox": [ + 46, + 217, + 289, + 312 + ], + "type": "text", + "content": "Moreover, due to the complex imaging process of large FoV fisheye lenses, handling fisheye image distortion using mathematical formulas can be a challenging task. While several excellent large FoV fisheye camera models have been proposed [2, 18, 19, 27, 30, 36], our experiments have shown that none of these models can accurately approximate the imaging process, and we get less satisfactory depth estimation results when using them." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 313, + 289, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 313, + 289, + 552 + ], + "spans": [ + { + "bbox": [ + 46, + 313, + 289, + 552 + ], + "type": "text", + "content": "Inspired by the above observations, we propose OmniVidar, a novel and simple multi-fisheye omnidirectional depth estimation system, as shown in Figure 2. OmniVidar contains three components. Firstly, we improve the DSCM [36] and propose a new camera model, named Triple Sphere Camera Model (TSCM), which can better approximate the imaging process and achieve the best accuracy. Then we propose an epipolar rectification algorithm designed for multi-fisheye camera system. We leverage a cubic-like projection approach to transform the four fisheye camera systems into four binocular camera systems and then conduct epipolar rectification on each binocular system. This method solves the distortion issue and reduces the complex multi-fisheye omnidirectional depth estimation problem to a much simpler binocular depth estimation problem. Finally, we design a lightweight binocular depth estimation network based on RAFT [35]. We add Transformer encoder [37] into feature extraction in RAFT to combine the advantages of Transformer and GRU, and it's easy to balance the accuracy and efficiency." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 553, + 288, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 553, + 288, + 601 + ], + "spans": [ + { + "bbox": [ + 46, + 553, + 288, + 601 + ], + "type": "text", + "content": "We compare OmniVidar with existing methods on several datasets. The results demonstrate that our method outperforms all the others in terms of speed and accuracy, achieving State-of-The-Art performance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 613, + 134, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 613, + 134, + 625 + ], + "spans": [ + { + "bbox": [ + 47, + 613, + 134, + 625 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 634, + 188, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 634, + 188, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 634, + 188, + 647 + ], + "type": "text", + "content": "2.1. Multiview Fisheye Stereo" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": "Gao et al. [13] mount two large FoV cameras " + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "inline_equation", + "content": "(245^{\\circ})" + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": " vertically and reversely to obtain a " + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": " horizontal and " + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "inline_equation", + "content": "65^{\\circ}" + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": " vertical FoV overlap and use this system in UAV for depth perception [43]. Won et al. [40] use four cameras with " + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "inline_equation", + "content": "220^{\\circ}" + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": " FoV and propose SweepNet algorithm. They propose" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 72, + 547, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 239 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 239 + ], + "type": "text", + "content": "Sphere-Sweep method modified from Plane-Sweep [16] to construct the cost volume and then use the cost aggregation algorithm in SGM [15] to obtain depth. After this, Won et al. propose OmniMVS [39, 41], which uses 3D convolution instead of SGM for depth regression and achieves higher accuracy. Komatsu et al. [21] propose CrownConv360. They improve OmniMVS by using the icosahedral projection in the feature extraction and depth regression stage for distortion reduction. Instead of using neural network, Meuleman et al. [29] choose to investigate cost aggregation and propagation in traditional stereo matching. And they achieve real-time high-resolution depth estimation at 29 fps on the embedded device TX2. However, the algorithm is only suitable for short camera baseline." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 240, + 547, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 240, + 547, + 300 + ], + "spans": [ + { + "bbox": [ + 304, + 240, + 547, + 300 + ], + "type": "text", + "content": "All of the above works have more or fewer deficiencies. We propose OmniVidar, an omnidirectional depth estimation method with high accuracy, which can run in real-time and solve the fisheye image distortion problem with good generalization in various scenes." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 310, + 446, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 310, + 446, + 323 + ], + "spans": [ + { + "bbox": [ + 306, + 310, + 446, + 323 + ], + "type": "text", + "content": "2.2. Fisheye Image Distortion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 329, + 545, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 329, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 304, + 329, + 545, + 365 + ], + "type": "text", + "content": "To solve the distortion problem, a camera model suitable for fisheye is required, and the distortion should be handled explicitly in depth estimation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 365, + 547, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 365, + 547, + 510 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 547, + 510 + ], + "type": "text", + "content": "Scaramuzzi et al. [30] propose a high-order polynomial-based camera projection model, which is highly generalizable and can be widely used for various catadioptric and fisheye cameras. But it has no closed-form solution, so the inverse projection equation needs to be fitted with a high-order polynomial, which introduces errors. Mei et al. [27] propose the MUCM, which is an improvement of the UCM [2] to achieve higher calibration accuracy and propose a novel intrinsic parameters initialization method. Usenko et al. [36] propose the DSCM, which has only one more intrinsic parameter than the MUCM model, but the accuracy is substantially improved." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 510, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 547, + 714 + ], + "type": "text", + "content": "Su et al. [34] propose a mini-network that learns to modify the shape of filters based on location. Coors et al. [8] chose to use a camera projection model to shape the convolution kernel to directly address distortion. An alternative idea was chosen by Cohen et al. [7] who define a form of convolution on the Lie group " + }, + { + "bbox": [ + 304, + 510, + 547, + 714 + ], + "type": "inline_equation", + "content": "SO(3)" + }, + { + "bbox": [ + 304, + 510, + 547, + 714 + ], + "type": "text", + "content": " that convolves over the rotation space rather than the translation space. Eder et al. [9] compared the performance of three approaches: equirectangular projection, cubic projection, and 7-level subdivided icosahedral projection, on depth estimation and semantic segmentation tasks. The results showed that icosahedral projection achieved the best results in all tasks. Komatsu et al. [21] used icosahedral projection in multi-fisheye omnidirectional stereo vision to improve the depth estimation results of OmniMVS. Eder et al. [10] proposed a tangent plane projection modified from the icosahedron projection." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21530" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 64, + 545, + 232 + ], + "blocks": [ + { + "bbox": [ + 47, + 64, + 545, + 232 + ], + "lines": [ + { + "bbox": [ + 47, + 64, + 545, + 232 + ], + "spans": [ + { + "bbox": [ + 47, + 64, + 545, + 232 + ], + "type": "image", + "image_path": "48c296a724e0d882bd1f0f98b4f5061e8fa03c27f3d291ac386dfef14525d566.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 239, + 548, + 285 + ], + "lines": [ + { + "bbox": [ + 46, + 239, + 548, + 285 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 548, + 285 + ], + "type": "text", + "content": "Figure 2. System overview. Our system contains three parts: multi-fisheye calibration, multi-fisheye epipolar rectification, and binocular depth estimation. In epipolar rectification, the images in the left column are original. The images in the middle column are the images after abstracting each fisheye camera into two pinhole cameras, where the two images connected by a red dotted arrow are a pair of binocular cameras. The images are rearranged and combined to obtain the right column." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 304, + 289, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 304, + 289, + 389 + ], + "spans": [ + { + "bbox": [ + 46, + 304, + 289, + 389 + ], + "type": "text", + "content": "The above algorithms have a common feature that they are all designed for the monocular image. Thus they perform well on tasks that require only one camera, yet not much good for multi-fisheye tasks. Therefore, to address this problem, we propose a distortion removal method designed for multi-fisheye systems, which greatly simplifies the depth estimation problem at the same time." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 397, + 177, + 408 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 397, + 177, + 408 + ], + "spans": [ + { + "bbox": [ + 47, + 397, + 177, + 408 + ], + "type": "text", + "content": "2.3. RAFT in Stereo Vision" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 416, + 287, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 416, + 287, + 608 + ], + "spans": [ + { + "bbox": [ + 46, + 416, + 287, + 608 + ], + "type": "text", + "content": "Zachary et al. [35] propose Recurrent All-Pairs Field Transforms (RAFT), a highly efficient and accurate deep network architecture for optical flow. RAFT uses modified GRU blocks to iteratively refine optical flow and shares weights between refinement units. Lahav et al. [24] apply the RAFT to binocular rectified stereo vision and introduce multi-level convolutional GRUs, which can more efficiently propagate information across the image. Wang et al. [38] modify RAFT and propose PVStereo which greatly outperforms other self-supervised stereo matching approaches. Li et al. [22] combine RAFT and Transformer and propose CREStereo. They add Transformer encoder in image feature extraction and repeat the cross-attention between left and right images in the iterative update stage. This method performs well in various binocular datasets with excellent generalization." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 609, + 287, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 609, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 46, + 609, + 287, + 645 + ], + "type": "text", + "content": "We imitate CREStereo and improve it based on RAFT to obtain a binocular stereo vision depth estimation network with a better trade-off between efficiency and accuracy." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 656, + 107, + 669 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 656, + 107, + 669 + ], + "spans": [ + { + "bbox": [ + 47, + 656, + 107, + 669 + ], + "type": "text", + "content": "3. Methods" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 715 + ], + "type": "text", + "content": "The overall flow of our method is shown in Figure 2 and is divided into three parts: camera calibration, epipolar rectification, and depth estimation." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 331, + 304, + 525, + 552 + ], + "blocks": [ + { + "bbox": [ + 331, + 304, + 525, + 552 + ], + "lines": [ + { + "bbox": [ + 331, + 304, + 525, + 552 + ], + "spans": [ + { + "bbox": [ + 331, + 304, + 525, + 552 + ], + "type": "image", + "image_path": "1718c5435412406974cc383527dbfba4d4f2145542d5570cf5c08cd8e0f339f3.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 572, + 547, + 641 + ], + "lines": [ + { + "bbox": [ + 304, + 572, + 547, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 572, + 547, + 641 + ], + "type": "text", + "content": "Figure 3. Our Triple Sphere Camera Model. First, point " + }, + { + "bbox": [ + 304, + 572, + 547, + 641 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 304, + 572, + 547, + 641 + ], + "type": "text", + "content": " is projected onto the first sphere and then onto the second sphere, then the third. The second sphere is shifted w.r.t. the first sphere by " + }, + { + "bbox": [ + 304, + 572, + 547, + 641 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 304, + 572, + 547, + 641 + ], + "type": "text", + "content": ", and the third is shifted w.r.t. the second sphere by " + }, + { + "bbox": [ + 304, + 572, + 547, + 641 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 572, + 547, + 641 + ], + "type": "text", + "content": ". Finally, the point is projected onto the image plane of a pinhole camera which is shifted by " + }, + { + "bbox": [ + 304, + 572, + 547, + 641 + ], + "type": "inline_equation", + "content": "\\frac{\\alpha}{1 - \\alpha}" + }, + { + "bbox": [ + 304, + 572, + 547, + 641 + ], + "type": "text", + "content": " from the third sphere." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 667, + 424, + 679 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 667, + 424, + 679 + ], + "spans": [ + { + "bbox": [ + 306, + 667, + 424, + 679 + ], + "type": "text", + "content": "3.1. Camera Calibration" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 689, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 546, + 715 + ], + "type": "text", + "content": "We propose the Triple Sphere Camera Model, which can better fit the imaging process of large FoV fisheye cameras" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "21531" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 228, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 228, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 228, + 83 + ], + "type": "text", + "content": "and has a closed-form unprojection solution." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 84, + 287, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 84, + 287, + 203 + ], + "spans": [ + { + "bbox": [ + 46, + 84, + 287, + 203 + ], + "type": "text", + "content": "As shown in Figure 3, our projection model considers that the incident light is refracted three times, and the displacements of the three unit spherical centers are " + }, + { + "bbox": [ + 46, + 84, + 287, + 203 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 46, + 84, + 287, + 203 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 84, + 287, + 203 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 84, + 287, + 203 + ], + "type": "text", + "content": ". After three refractions, the incident light is finally projected to the image plane according to the pinhole camera model, and the displacement of the pinhole camera model's optical center from the third unit sphere is " + }, + { + "bbox": [ + 46, + 84, + 287, + 203 + ], + "type": "inline_equation", + "content": "\\frac{\\alpha}{1 - \\alpha}" + }, + { + "bbox": [ + 46, + 84, + 287, + 203 + ], + "type": "text", + "content": ". Therefore, our model has totally 7 camera internal parameters: " + }, + { + "bbox": [ + 46, + 84, + 287, + 203 + ], + "type": "inline_equation", + "content": "f_{x}, f_{y}, c_{x}, c_{y}, \\xi, \\lambda, \\alpha" + }, + { + "bbox": [ + 46, + 84, + 287, + 203 + ], + "type": "text", + "content": ". The projection equations are defined as follows." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 77, + 209, + 287, + 249 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 209, + 287, + 249 + ], + "spans": [ + { + "bbox": [ + 77, + 209, + 287, + 249 + ], + "type": "interline_equation", + "content": "\\pi (P, i) = \\frac {1}{\\zeta} \\left[ \\begin{array}{c c c} f _ {x} & 0 & c _ {x} \\\\ 0 & f _ {y} & c _ {y} \\end{array} \\right] \\left[ \\begin{array}{l} X \\\\ Y \\\\ \\zeta \\end{array} \\right] \\tag {1}", + "image_path": "12afc60aafc19cbbb5e3578249e1f9842cde02ba7b269fcf660abe14d7a06f8e.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 100, + 250, + 287, + 270 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 250, + 287, + 270 + ], + "spans": [ + { + "bbox": [ + 100, + 250, + 287, + 270 + ], + "type": "interline_equation", + "content": "\\zeta = Z + \\xi d _ {1} + \\lambda d _ {2} + \\frac {\\alpha}{1 - \\alpha} d _ {3} \\tag {2}", + "image_path": "8a2855acb1f84a9d18352e7b6894ea4d2151f4e7c5f8beae17a035e9e70fb244.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 96, + 273, + 287, + 289 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 273, + 287, + 289 + ], + "spans": [ + { + "bbox": [ + 96, + 273, + 287, + 289 + ], + "type": "interline_equation", + "content": "d _ {1} = \\sqrt {X ^ {2} + Y ^ {2} + Z ^ {2}} \\tag {3}", + "image_path": "5aa9d85e4af99ec07b4d1eb9bb91b1570d9b86b85f2ce90a16118c225fb82760.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 96, + 291, + 287, + 306 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 291, + 287, + 306 + ], + "spans": [ + { + "bbox": [ + 96, + 291, + 287, + 306 + ], + "type": "interline_equation", + "content": "d _ {2} = \\sqrt {X ^ {2} + Y ^ {2} + (\\xi d _ {1} + Z) ^ {2}} \\tag {4}", + "image_path": "b050021c835a3f974b194b533fcf7c84a6fd5ed573b913fcaa7f195eaa74a409.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 96, + 308, + 287, + 323 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 308, + 287, + 323 + ], + "spans": [ + { + "bbox": [ + 96, + 308, + 287, + 323 + ], + "type": "interline_equation", + "content": "d _ {3} = \\sqrt {X ^ {2} + Y ^ {2} + (\\lambda d _ {2} + \\xi d _ {1} + Z) ^ {2}} \\tag {5}", + "image_path": "3584149a4e6eaa834272a5e5e8139fd9cc30cbfed9c36be9bb1601670d390f63.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 332, + 286, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 332, + 286, + 354 + ], + "spans": [ + { + "bbox": [ + 47, + 332, + 286, + 354 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 332, + 286, + 354 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 47, + 332, + 286, + 354 + ], + "type": "text", + "content": " is the vector of intrinsic parameters, " + }, + { + "bbox": [ + 47, + 332, + 286, + 354 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 47, + 332, + 286, + 354 + ], + "type": "text", + "content": " is the projection function." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 356, + 286, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 356, + 286, + 379 + ], + "spans": [ + { + "bbox": [ + 47, + 356, + 286, + 379 + ], + "type": "text", + "content": "A set of 3D points that results in valid projection is expressed as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 94, + 388, + 287, + 403 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 388, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 94, + 388, + 287, + 403 + ], + "type": "interline_equation", + "content": "\\Omega = \\left\\{x \\in \\mathbb {R} ^ {3} | z > - w _ {2} d _ {1} \\right\\} \\tag {6}", + "image_path": "08672176b89891f819167048f7cab432d4d938812013ec82f1952c81a490c98b.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 91, + 404, + 287, + 432 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 404, + 287, + 432 + ], + "spans": [ + { + "bbox": [ + 91, + 404, + 287, + 432 + ], + "type": "interline_equation", + "content": "w _ {2} = \\frac {\\xi + \\lambda + w _ {1}}{\\sqrt {1 + (\\xi + \\lambda) ^ {2} + 2 w _ {1} (\\xi + \\lambda)}} \\tag {7}", + "image_path": "dbcbbbf00e56657a3efc55283436f6b6ef3bd1f1df9a9144d8af3f9d56b29fe0.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 91, + 434, + 287, + 454 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 434, + 287, + 454 + ], + "spans": [ + { + "bbox": [ + 91, + 434, + 287, + 454 + ], + "type": "interline_equation", + "content": "w _ {1} = \\frac {\\alpha}{1 - \\alpha} \\tag {8}", + "image_path": "9f972be2817249ed1270c950ad9e8423d4071695a8d3d0cba3f18484f2eca9ea.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 59, + 462, + 263, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 462, + 263, + 474 + ], + "spans": [ + { + "bbox": [ + 59, + 462, + 263, + 474 + ], + "type": "text", + "content": "The unprojection function is computed as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 59, + 482, + 287, + 521 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 482, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 59, + 482, + 287, + 521 + ], + "type": "interline_equation", + "content": "\\pi^ {- 1} (p, i) = \\mu \\left[ \\begin{array}{l} \\eta \\gamma x \\\\ \\eta \\gamma y \\\\ m _ {z} \\end{array} \\right] - \\left[ \\begin{array}{l} 0 \\\\ 0 \\\\ \\xi \\end{array} \\right] \\tag {9}", + "image_path": "ef706bd1e7dc773b91266a74186487aab461794a8ef9143ec870eb8c55550b7b.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 91, + 523, + 286, + 537 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 523, + 286, + 537 + ], + "spans": [ + { + "bbox": [ + 91, + 523, + 286, + 537 + ], + "type": "interline_equation", + "content": "\\mu = \\xi m _ {z} + \\sqrt {\\xi^ {2} m _ {z} ^ {2} - \\xi^ {2} + 1} \\tag {10}", + "image_path": "0c028b71b9ef2667b7d1a9ff36906021d3640eda39dd7df50068fdf5f0d6736d.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 84, + 540, + 287, + 552 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 540, + 287, + 552 + ], + "spans": [ + { + "bbox": [ + 84, + 540, + 287, + 552 + ], + "type": "interline_equation", + "content": "m _ {z} = \\eta (\\gamma - \\zeta) - \\lambda \\tag {11}", + "image_path": "cfbd802409b9d8ab282e4a9e79a65267d3f4105334741db58f0d21c6a41669ee.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 91, + 555, + 287, + 570 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 555, + 287, + 570 + ], + "spans": [ + { + "bbox": [ + 91, + 555, + 287, + 570 + ], + "type": "interline_equation", + "content": "\\eta = \\lambda (\\gamma - \\zeta) + \\sqrt {\\lambda^ {2} (\\gamma - \\zeta) ^ {2} - \\lambda^ {2} + 1} \\tag {12}", + "image_path": "49af1dc161973aff58b455e7f3f94e316803d321558169b786611a33e9794cfc.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 91, + 573, + 287, + 599 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 573, + 287, + 599 + ], + "spans": [ + { + "bbox": [ + 91, + 573, + 287, + 599 + ], + "type": "interline_equation", + "content": "\\gamma = \\frac {\\zeta + \\sqrt {1 + (1 - \\zeta^ {2}) \\left(x ^ {2} + y ^ {2}\\right)}}{x ^ {2} + y ^ {2} + 1} \\tag {13}", + "image_path": "63069d0e06110b4e55c3f80338e6b5d93678b3bc9579699b603c20e791c21392.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 91, + 601, + 287, + 634 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 601, + 287, + 634 + ], + "spans": [ + { + "bbox": [ + 91, + 601, + 287, + 634 + ], + "type": "interline_equation", + "content": "\\zeta = \\left\\{ \\begin{array}{l l} \\frac {\\alpha}{1 - \\alpha} & \\text {i f} \\alpha \\leq 0. 5 \\\\ \\frac {1 - \\alpha}{\\alpha} & \\text {i f} \\alpha > 0. 5 \\end{array} \\right. \\tag {14}", + "image_path": "71b2606d16fa188a24e11c39722bc682c98c7224d2b723edb8f75f58dc434fa5.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 47, + 642, + 218, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 642, + 218, + 653 + ], + "spans": [ + { + "bbox": [ + 47, + 642, + 218, + 653 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 642, + 218, + 653 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 47, + 642, + 218, + 653 + ], + "type": "text", + "content": " is the normalized coordinate." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "text", + "content": "We calibrate the intrinsic and extrinsic parameters of multi-fisheye camera system using a highly accurate planar checkerboard. The goal of the calibration is to minimize the reprojection error of the corner points in all pictures. For each image " + }, + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 654, + 287, + 713 + ], + "type": "text", + "content": " in the calibration sequence, the" + } + ] + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 309, + 69, + 545, + 182 + ], + "blocks": [ + { + "bbox": [ + 309, + 69, + 545, + 182 + ], + "lines": [ + { + "bbox": [ + 309, + 69, + 545, + 182 + ], + "spans": [ + { + "bbox": [ + 309, + 69, + 545, + 182 + ], + "type": "image", + "image_path": "5be166fd2ac635187db3de74983f1bad2c25fc0222d0602184ff9cbd1aca0070.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 199, + 545, + 232 + ], + "lines": [ + { + "bbox": [ + 305, + 199, + 545, + 232 + ], + "spans": [ + { + "bbox": [ + 305, + 199, + 545, + 232 + ], + "type": "text", + "content": "Figure 4. The multi-fisheye epipolar rectification schematic. Each fisheye camera is converted into two pinhole cameras with an angle of " + }, + { + "bbox": [ + 305, + 199, + 545, + 232 + ], + "type": "inline_equation", + "content": "90^{\\circ}" + }, + { + "bbox": [ + 305, + 199, + 545, + 232 + ], + "type": "text", + "content": ". Then the system becomes four binocular systems." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "bbox": [ + 305, + 251, + 545, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 251, + 545, + 322 + ], + "spans": [ + { + "bbox": [ + 305, + 251, + 545, + 322 + ], + "type": "text", + "content": "corner detector can obtain the projection point " + }, + { + "bbox": [ + 305, + 251, + 545, + 322 + ], + "type": "inline_equation", + "content": "u_{nk}" + }, + { + "bbox": [ + 305, + 251, + 545, + 322 + ], + "type": "text", + "content": " of the " + }, + { + "bbox": [ + 305, + 251, + 545, + 322 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 305, + 251, + 545, + 322 + ], + "type": "text", + "content": "th corner point " + }, + { + "bbox": [ + 305, + 251, + 545, + 322 + ], + "type": "inline_equation", + "content": "x_{k}" + }, + { + "bbox": [ + 305, + 251, + 545, + 322 + ], + "type": "text", + "content": ". The coordinate of " + }, + { + "bbox": [ + 305, + 251, + 545, + 322 + ], + "type": "inline_equation", + "content": "u_{nk}" + }, + { + "bbox": [ + 305, + 251, + 545, + 322 + ], + "type": "text", + "content": " is related to the camera intrinsic and extrinsic parameters. Let " + }, + { + "bbox": [ + 305, + 251, + 545, + 322 + ], + "type": "inline_equation", + "content": "s = [i, T_{cam_0}, T_{cam_1}, \\dots, T_{cam_n}]" + }, + { + "bbox": [ + 305, + 251, + 545, + 322 + ], + "type": "text", + "content": " be the parameter to optimize. We can construct the nonlinear optimization problem as follows:" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 329, + 545, + 362 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 329, + 545, + 362 + ], + "spans": [ + { + "bbox": [ + 316, + 329, + 545, + 362 + ], + "type": "interline_equation", + "content": "s = \\arg \\min _ {s ^ {\\prime}} \\sum_ {n = 0} ^ {N} \\sum_ {k \\in K} \\rho \\left(\\left(\\pi \\left(T _ {c a m _ {n}} x _ {k}, \\boldsymbol {i}\\right) - u _ {n k}\\right) ^ {2}\\right) \\tag {15}", + "image_path": "99f1680bc9364c383c7f1c2fd948a26597f6e8f14a7b87b64319d46c351420fe.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 305, + 369, + 545, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 369, + 545, + 418 + ], + "spans": [ + { + "bbox": [ + 305, + 369, + 545, + 418 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 369, + 545, + 418 + ], + "type": "inline_equation", + "content": "T_{cam_n} \\in SE(3)" + }, + { + "bbox": [ + 305, + 369, + 545, + 418 + ], + "type": "text", + "content": " is the transformation from the coordinate frame of the calibration grid to the camera coordinate frame for image " + }, + { + "bbox": [ + 305, + 369, + 545, + 418 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 305, + 369, + 545, + 418 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 305, + 369, + 545, + 418 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 305, + 369, + 545, + 418 + ], + "type": "text", + "content": " is a set of detected corner points for the image " + }, + { + "bbox": [ + 305, + 369, + 545, + 418 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 305, + 369, + 545, + 418 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 369, + 545, + 418 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 305, + 369, + 545, + 418 + ], + "type": "text", + "content": " is the robust Huber norm." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 305, + 418, + 545, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 418, + 545, + 501 + ], + "spans": [ + { + "bbox": [ + 305, + 418, + 545, + 501 + ], + "type": "text", + "content": "Due to the highly nonlinear nature of the image, the above optimization problem requires a good initial value. We initialize the intrinsic parameters using the previously proposed method [27] and find initial poses using the UPnP algorithm [20]. After obtaining the initial values of the parameters, we completed the optimization operation using Ceres Solver [1]." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 306, + 508, + 514, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 508, + 514, + 521 + ], + "spans": [ + { + "bbox": [ + 306, + 508, + 514, + 521 + ], + "type": "text", + "content": "3.2. Multi-Fisheye Camera Epipolar Rectify" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 526, + 545, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 526, + 545, + 658 + ], + "spans": [ + { + "bbox": [ + 304, + 526, + 545, + 658 + ], + "type": "text", + "content": "We propose a simple and effective undistortion method that cleverly transforms the omnidirectional depth estimation problem into a binocular depth estimation problem. And the method can be used in the system with any number of fisheye cameras. As shown in Figure 4, we abstract each fisheye camera as two pinhole cameras with " + }, + { + "bbox": [ + 304, + 526, + 545, + 658 + ], + "type": "inline_equation", + "content": "90^{\\circ}" + }, + { + "bbox": [ + 304, + 526, + 545, + 658 + ], + "type": "text", + "content": " angle1. Then the four fisheye cameras system becomes four independent binocular systems. Each binocular systems can be rectified through epipolar rectification. We modified the classical epipolar rectification algorithm proposed by Fusiello [11] for our system." + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 306, + 664, + 545, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 664, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 664, + 545, + 713 + ], + "type": "text", + "content": "1Although here we only show the horizontal binoculars, it is possible to construct binoculars facing both above and below to estimate the true omnidirectional depth. However, we consider that the above and below depth values are rarely used in practical applications. So we just use a limited " + }, + { + "bbox": [ + 306, + 664, + 545, + 713 + ], + "type": "inline_equation", + "content": "110^{\\circ}" + }, + { + "bbox": [ + 306, + 664, + 545, + 713 + ], + "type": "text", + "content": " vertical FoV, which satisfies most of the application requirements." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21532" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 68, + 547, + 262 + ], + "blocks": [ + { + "bbox": [ + 47, + 68, + 547, + 262 + ], + "lines": [ + { + "bbox": [ + 47, + 68, + 547, + 262 + ], + "spans": [ + { + "bbox": [ + 47, + 68, + 547, + 262 + ], + "type": "image", + "image_path": "d47d6dd39f947d590957077f36f3647181aa27b3741cb84ea7d73ba868e33b12.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 278, + 547, + 323 + ], + "lines": [ + { + "bbox": [ + 46, + 278, + 547, + 323 + ], + "spans": [ + { + "bbox": [ + 46, + 278, + 547, + 323 + ], + "type": "text", + "content": "Figure 5. The structure of our proposed binocular depth estimation network. A pair of stereo images are fed into feature extraction networks. Then the output features are calculated by self-attention and cross-attention, and after doing self-attention, the features of left image are used for context information for latter recurrent update stage. Then the left and right features are multiplied to construct cost volume. Finally use the disparity update model same as RAFT [35]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 344, + 287, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 344, + 287, + 408 + ], + "spans": [ + { + "bbox": [ + 47, + 344, + 287, + 408 + ], + "type": "text", + "content": "Firstly, the rotation matrix of the cameras in binocular system is calculated. Let the optical centers of the left camera and right camera be " + }, + { + "bbox": [ + 47, + 344, + 287, + 408 + ], + "type": "inline_equation", + "content": "C_1 = \\left[x_1 y_1 z_1\\right]^T" + }, + { + "bbox": [ + 47, + 344, + 287, + 408 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 344, + 287, + 408 + ], + "type": "inline_equation", + "content": "C_2 = \\left[x_2 y_2 z_2\\right]^T" + }, + { + "bbox": [ + 47, + 344, + 287, + 408 + ], + "type": "text", + "content": ", the rotation matrix can be calculated as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 417, + 286, + 430 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 417, + 286, + 430 + ], + "spans": [ + { + "bbox": [ + 59, + 417, + 286, + 430 + ], + "type": "interline_equation", + "content": "R = \\left[ \\begin{array}{l l l} r _ {1} ^ {T} & r _ {2} ^ {T} & r _ {3} ^ {T} \\end{array} \\right] \\tag {16}", + "image_path": "f0e851e9975f2d8e61e1b1534ae2f22aa00903903b37faee494131f4153f19ab.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 434, + 286, + 447 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 434, + 286, + 447 + ], + "spans": [ + { + "bbox": [ + 58, + 434, + 286, + 447 + ], + "type": "interline_equation", + "content": "r _ {1} = \\left[ \\begin{array}{l l l} x _ {2} - x _ {1} & y _ {2} - y _ {1} & z _ {2} - z _ {1} \\end{array} \\right]. n o r m a l i z e () \\tag {17}", + "image_path": "bd5540ecbef06fc67c72c773d024f4faa12f56bb3fcaf31faa6dd1cae02d0ac2.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 449, + 286, + 463 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 449, + 286, + 463 + ], + "spans": [ + { + "bbox": [ + 58, + 449, + 286, + 463 + ], + "type": "interline_equation", + "content": "r _ {3} = \\left[ - \\left(z _ {2} - z _ {1}\\right) \\quad 0 \\quad x _ {2} - x _ {1} \\right]. n o r m a l i z e () \\tag {18}", + "image_path": "b27581a75bd60c7b63d53a373fde03256cfaf5c0b88db398aea4c79142e70cf9.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 58, + 466, + 286, + 476 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 466, + 286, + 476 + ], + "spans": [ + { + "bbox": [ + 58, + 466, + 286, + 476 + ], + "type": "interline_equation", + "content": "r _ {2} = r _ {3} \\times r _ {1} \\tag {19}", + "image_path": "09654356e189298d3b374b4018443237ec137e22239a43aec51e2bbafcd54fdd.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 486, + 287, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 486, + 287, + 558 + ], + "spans": [ + { + "bbox": [ + 47, + 486, + 287, + 558 + ], + "type": "text", + "content": "Then remap the image for epipolar rectification. Taking the left camera in one binocular system as an example. Let the rotation matrix of the corresponding fisheye camera be " + }, + { + "bbox": [ + 47, + 486, + 287, + 558 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 47, + 486, + 287, + 558 + ], + "type": "text", + "content": ", the optical center be " + }, + { + "bbox": [ + 47, + 486, + 287, + 558 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 47, + 486, + 287, + 558 + ], + "type": "text", + "content": ". For a certain spatial point " + }, + { + "bbox": [ + 47, + 486, + 287, + 558 + ], + "type": "inline_equation", + "content": "P = (X, Y, Z, 1)" + }, + { + "bbox": [ + 47, + 486, + 287, + 558 + ], + "type": "text", + "content": ", its projection point on normalized plane is " + }, + { + "bbox": [ + 47, + 486, + 287, + 558 + ], + "type": "inline_equation", + "content": "p = (u, v, 1)" + }, + { + "bbox": [ + 47, + 486, + 287, + 558 + ], + "type": "text", + "content": ". The projection equation is shown below." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 99, + 565, + 286, + 616 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 565, + 286, + 616 + ], + "spans": [ + { + "bbox": [ + 99, + 565, + 286, + 616 + ], + "type": "interline_equation", + "content": "\\mu \\left[ \\begin{array}{c} u \\\\ v \\\\ f (u, v) \\end{array} \\right] = [ R - R C ] \\left[ \\begin{array}{l} X \\\\ Y \\\\ Z \\\\ 1 \\end{array} \\right] \\tag {20}", + "image_path": "b8d3d0ef1df0a19074bd61cf38b1d4486d5641400dc7800e3f881acb0da31dff.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 622, + 287, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 622, + 287, + 658 + ], + "spans": [ + { + "bbox": [ + 47, + 622, + 287, + 658 + ], + "type": "text", + "content": "Let the rotation matrix of the rectified pinhole camera be " + }, + { + "bbox": [ + 47, + 622, + 287, + 658 + ], + "type": "inline_equation", + "content": "R'" + }, + { + "bbox": [ + 47, + 622, + 287, + 658 + ], + "type": "text", + "content": ", and the corresponding normalized plane coordinate be " + }, + { + "bbox": [ + 47, + 622, + 287, + 658 + ], + "type": "inline_equation", + "content": "(u', v', 1)" + }, + { + "bbox": [ + 47, + 622, + 287, + 658 + ], + "type": "text", + "content": ". Then the projection equation is shown below." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 104, + 666, + 286, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 666, + 286, + 717 + ], + "spans": [ + { + "bbox": [ + 104, + 666, + 286, + 717 + ], + "type": "interline_equation", + "content": "\\mu^ {\\prime} \\left[ \\begin{array}{l} u ^ {\\prime} \\\\ v ^ {\\prime} \\\\ 1 \\end{array} \\right] = \\left[ \\begin{array}{l l} R ^ {\\prime} & - R ^ {\\prime} C \\end{array} \\right] \\left[ \\begin{array}{l} X \\\\ Y \\\\ Z \\\\ 1 \\end{array} \\right] \\tag {21}", + "image_path": "f11afb8e3e7d9c84cc341c7a2d2bf30cfcc111ebda240c6ab594c2b47b4a4679.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 344, + 545, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 344, + 545, + 368 + ], + "spans": [ + { + "bbox": [ + 306, + 344, + 545, + 368 + ], + "type": "text", + "content": "The relationship between fisheye image and virtual binocular image can be obtained, see equation (22)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 315, + 372, + 545, + 411 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 372, + 545, + 411 + ], + "spans": [ + { + "bbox": [ + 315, + 372, + 545, + 411 + ], + "type": "interline_equation", + "content": "\\lambda \\left[ \\begin{array}{l} u ^ {\\prime} \\\\ v ^ {\\prime} \\\\ 1 \\end{array} \\right] = \\left[ \\begin{array}{l l l} R ^ {\\prime} & - R ^ {\\prime} C \\end{array} \\right] \\left[ \\begin{array}{l l} R & - R C \\end{array} \\right] ^ {- 1} \\left[ \\begin{array}{c} u \\\\ v \\\\ f (u, v) \\end{array} \\right] \\tag {22}", + "image_path": "bb835f8c4cc70ad3a9c69cec910b8144afae49a8ba548a914e26bd8e6ffa4aff.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 414, + 545, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 414, + 545, + 450 + ], + "spans": [ + { + "bbox": [ + 305, + 414, + 545, + 450 + ], + "type": "text", + "content": "That is, for each pixel in the original fisheye image, the pixel coordinate in the pinhole image can be calculated. We can use pixel interpolation to obtain the rectified image." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 451, + 545, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 451, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 305, + 451, + 545, + 510 + ], + "type": "text", + "content": "As shown in Figure 2, it can be seen that after epipolar rectification, the four fisheye images are converted into a pair of binocular images, where each image is stitched together from the four virtual pinhole camera images. A simple binocular depth estimation algorithm is then used." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 515, + 443, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 515, + 443, + 528 + ], + "spans": [ + { + "bbox": [ + 306, + 515, + 443, + 528 + ], + "type": "text", + "content": "3.3. Stereo Depth Estimation" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 534, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 534, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 305, + 534, + 545, + 605 + ], + "type": "text", + "content": "Since our system needs to estimate the depth values of four binocular systems simultaneously, which is very computationally intensive, our primary goal is not to design a binocular algorithm with the highest accuracy, but to minimize the accuracy loss while ensuring real-time performance." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 605, + 545, + 713 + ], + "type": "text", + "content": "The structure of our binocular depth estimation network is shown in Figure 5. We added Transformer [37] with self-attention and cross-attention in the image feature extraction stage. In binocular depth estimation, limited by computational performance and graphics card memory, many algorithms have to limit the matching search range of pixels. In this paper, we design an ingenious cost-volume calculation that can extend the estimable disparity range to the image width, while also reducing the computational complexity." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21533" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 51, + 68, + 535, + 340 + ], + "blocks": [ + { + "bbox": [ + 51, + 68, + 535, + 340 + ], + "lines": [ + { + "bbox": [ + 51, + 68, + 535, + 340 + ], + "spans": [ + { + "bbox": [ + 51, + 68, + 535, + 340 + ], + "type": "image", + "image_path": "d0a7dd8fefda34d19958a3fd25d587fa57b38cdefe80187c3c556751ea4a75bc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 349, + 547, + 371 + ], + "lines": [ + { + "bbox": [ + 46, + 349, + 547, + 371 + ], + "spans": [ + { + "bbox": [ + 46, + 349, + 547, + 371 + ], + "type": "text", + "content": "Figure 6. The depth estimation comparison between ours and other methods. The final row is the results from our method, and the column annotations indicate the dataset source." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 392, + 287, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 392, + 287, + 440 + ], + "spans": [ + { + "bbox": [ + 47, + 392, + 287, + 440 + ], + "type": "text", + "content": "Given two images " + }, + { + "bbox": [ + 47, + 392, + 287, + 440 + ], + "type": "inline_equation", + "content": "I_1 \\in \\mathbb{R}^{H \\times W \\times C}" + }, + { + "bbox": [ + 47, + 392, + 287, + 440 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 392, + 287, + 440 + ], + "type": "inline_equation", + "content": "I_2 \\in \\mathbb{R}^{H \\times W \\times C}" + }, + { + "bbox": [ + 47, + 392, + 287, + 440 + ], + "type": "text", + "content": ", we need to compute the cost volume " + }, + { + "bbox": [ + 47, + 392, + 287, + 440 + ], + "type": "inline_equation", + "content": "C(I_1, I_2) \\in \\mathbb{R}^{H \\times W \\times W}" + }, + { + "bbox": [ + 47, + 392, + 287, + 440 + ], + "type": "text", + "content": ". We found that the cost volume can be obtained by single matrix operation as follows." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 63, + 443, + 287, + 460 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 443, + 287, + 460 + ], + "spans": [ + { + "bbox": [ + 63, + 443, + 287, + 460 + ], + "type": "interline_equation", + "content": "C \\left(I _ {1}, I _ {2}\\right) = I _ {1} \\in \\mathbb {R} ^ {H \\times W \\times C} \\cdot I _ {2} ^ {T} \\in \\mathbb {R} ^ {H \\times C \\times W} \\tag {23}", + "image_path": "8472a0c62d848ad0d7bdffe72d1eeb0dd9c8064ab1314c0d58961dfad2962077.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 463, + 287, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 463, + 287, + 499 + ], + "spans": [ + { + "bbox": [ + 47, + 463, + 287, + 499 + ], + "type": "text", + "content": "To enable the network to learn information at multiple scales, we used the cost volume pooling opertaion in RAFT-Stereo [24]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 508, + 128, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 508, + 128, + 521 + ], + "spans": [ + { + "bbox": [ + 47, + 508, + 128, + 521 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 527, + 144, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 527, + 144, + 540 + ], + "spans": [ + { + "bbox": [ + 47, + 527, + 144, + 540 + ], + "type": "text", + "content": "4.1. Implementation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 545, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 287, + 677 + ], + "type": "text", + "content": "Our depth estimation network structure is shown in Figure 5. The number of channels in our network is identical to RAFT [35]. The feature encoder consists of 6 residual blocks, 2 at " + }, + { + "bbox": [ + 46, + 545, + 287, + 677 + ], + "type": "inline_equation", + "content": "1/2" + }, + { + "bbox": [ + 46, + 545, + 287, + 677 + ], + "type": "text", + "content": " resolution, 2 at " + }, + { + "bbox": [ + 46, + 545, + 287, + 677 + ], + "type": "inline_equation", + "content": "1/4" + }, + { + "bbox": [ + 46, + 545, + 287, + 677 + ], + "type": "text", + "content": " resolution, and 2 at " + }, + { + "bbox": [ + 46, + 545, + 287, + 677 + ], + "type": "inline_equation", + "content": "1/8" + }, + { + "bbox": [ + 46, + 545, + 287, + 677 + ], + "type": "text", + "content": " resolution. The number of channels is 64, 128, and 256, respectively. The " + }, + { + "bbox": [ + 46, + 545, + 287, + 677 + ], + "type": "inline_equation", + "content": "1/8" + }, + { + "bbox": [ + 46, + 545, + 287, + 677 + ], + "type": "text", + "content": " resolution feature map is then fed into the Transformer encoder [37]. We set the number of heads in attention to 8. The GRU iterative update part is identical to RAFT, and more details can be found in [35]. In addition, we replace all the normalization operators in RAFT with domain normalize [44]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "content": "Since we convert the multi-fisheye cameras to four binocular cameras, we can use a large number of public binocular datasets for training, instead of multi-fisheye" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 392, + 545, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 392, + 545, + 488 + ], + "spans": [ + { + "bbox": [ + 304, + 392, + 545, + 488 + ], + "type": "text", + "content": "datasets, which are very limited. Similar to the training strategy in CREStereo [22], we collected most of public binocular datasets, including SceneFlow [26], Sintel [4], CREStereo [22], KITTI [14, 28], InStereo2K [3], and Virtual KITTI [12]. The images are uniformly randomly cropped with a fixed resolution during the training process to solve the problem of different image resolutions in different datasets." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 489, + 545, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 489, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 304, + 489, + 545, + 525 + ], + "type": "text", + "content": "We use a single NVIDIA RTX2080Ti for training, the batch size is set to 4, the learning rate is 0.0004, and we use Adam to train 200 epochs to get the final network weights." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 526, + 546, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 526, + 546, + 622 + ], + "spans": [ + { + "bbox": [ + 304, + 526, + 546, + 622 + ], + "type": "text", + "content": "To test the performance of OmniVidar in real scenarios, we built a test setup. We used four leopard LI-USB30-OV10640-490-GMSL cameras with an image resolution of " + }, + { + "bbox": [ + 304, + 526, + 546, + 622 + ], + "type": "inline_equation", + "content": "1280 \\times 1024" + }, + { + "bbox": [ + 304, + 526, + 546, + 622 + ], + "type": "text", + "content": ", paired with a fisheye lens with " + }, + { + "bbox": [ + 304, + 526, + 546, + 622 + ], + "type": "inline_equation", + "content": "250^{\\circ}" + }, + { + "bbox": [ + 304, + 526, + 546, + 622 + ], + "type": "text", + "content": " FoV. The four cameras look in four directions, front and back, left and right, as shown in Figure 1. We use the device to capture indoor, corridor, and outdoor images, which can fully test the performance of OmniVidar in real scenes." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 634, + 380, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 634, + 380, + 645 + ], + "spans": [ + { + "bbox": [ + 306, + 634, + 380, + 645 + ], + "type": "text", + "content": "4.2. Evaluation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 545, + 712 + ], + "type": "text", + "content": "We perform the comparison on the Omnihouse, OmniThings provided by Won et al. [40, 41], and the datasets provided by Meuleman et al. [29]. Our OmniVidar has only been trained on the binocular dataset mentioned in Section 4.1, not on the test dataset, and the methods involved in the" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21534" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 75, + 70, + 519, + 156 + ], + "blocks": [ + { + "bbox": [ + 75, + 70, + 519, + 156 + ], + "lines": [ + { + "bbox": [ + 75, + 70, + 519, + 156 + ], + "spans": [ + { + "bbox": [ + 75, + 70, + 519, + 156 + ], + "type": "table", + "html": "
MethodMeuleman et al. [29]Omnihouse [40]Omnithings [39]
bad 1.0RMSEMAEbad 1.0RMSEMAEbad 1.0RMSEMAE
OmniVidar(Ours)18.4990.9210.8034.8861.0021.04211.1990.9580.461
OmniMVS-ft [41]60.3722.3061.73226.3570.9290.70559.7532.2701.723
CrownConv360 [21]69.1194.0203.00585.55911.0776.72993.9469.5456.518
Meuleman et al. [29]44.1111.6641.27311.3040.6710.41065.3482.5501.941
", + "image_path": "35722dae7b36dfbbd41a702052c48a26270a272d6feffb11a605817f4693ea0e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 196, + 287, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 196, + 287, + 220 + ], + "spans": [ + { + "bbox": [ + 46, + 196, + 287, + 220 + ], + "type": "text", + "content": "comparison all use the training method corresponding to the original paper." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 220, + 288, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 220, + 288, + 281 + ], + "spans": [ + { + "bbox": [ + 46, + 220, + 288, + 281 + ], + "type": "text", + "content": "As can be seen in the Figure 6, although OmniVidar is not trained and fine-tuned on the above datasets, its depth estimation accuracy on these datasets far exceeds that of other methods, with accurate depth estimation in weakly textured regions." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 48, + 291, + 293, + 353 + ], + "blocks": [ + { + "bbox": [ + 56, + 163, + 536, + 175 + ], + "lines": [ + { + "bbox": [ + 56, + 163, + 536, + 175 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 536, + 175 + ], + "type": "text", + "content": "Table 1. Quantitative comparison of algorithm accuracy on several datasets. The less the number is, the more accurate the method is." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 291, + 293, + 353 + ], + "lines": [ + { + "bbox": [ + 48, + 291, + 293, + 353 + ], + "spans": [ + { + "bbox": [ + 48, + 291, + 293, + 353 + ], + "type": "table", + "html": "
MethodRuntimeMemoryParameters
OmniVidar (Ours)66 ms3.7 GB4.8 MB
OmniMVS-ft [41]110 ms6.8 GB43.7 MB
CrownConv360 [21]623 ms6.3 GB16.7 MB
", + "image_path": "63da8119d156432abeb078c8cc01f89e1c13774c748ad0ebe23efa11120b4e48.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 385, + 287, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 385, + 287, + 469 + ], + "spans": [ + { + "bbox": [ + 46, + 385, + 287, + 469 + ], + "type": "text", + "content": "The analysis of the quantitative results are shown in the Table 1 and Table 2. It can be seen that OmniVidar has the least memory consumption, the shortest time consumption and the highest accuracy compared to other deep learning based solutions. Compared with non-deep learning methods, the accuracy is much higher than it, despite the relatively low efficiency." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 50, + 479, + 285, + 553 + ], + "blocks": [ + { + "bbox": [ + 50, + 479, + 285, + 553 + ], + "lines": [ + { + "bbox": [ + 50, + 479, + 285, + 553 + ], + "spans": [ + { + "bbox": [ + 50, + 479, + 285, + 553 + ], + "type": "image", + "image_path": "85cf9a42c4ee075faa9e176a0cb7ed10e703a63be27c925a84ed1d61cf06ca28.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 571, + 287, + 605 + ], + "lines": [ + { + "bbox": [ + 46, + 571, + 287, + 605 + ], + "spans": [ + { + "bbox": [ + 46, + 571, + 287, + 605 + ], + "type": "text", + "content": "Figure 7. Plot comparing the accuracy and efficiency evaluated in KITTI-2015 [28]. It can be seen that OmniVidar achieves real-time performance while maintain the accuracy." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "type": "text", + "content": "We also compared the binocular depth estimation network in OmniVidar with other binocular depth estimation networks. We chose the KITTI-2015 dataset [28] for the evaluation and the results are shown in Figure 7. We compare our method with CREStereo [22], LEAStereo [6], PSMNet [5], EdgeStereo [33], RAFT-Stereo [24]. It can see that our scheme achieves a better balance between accuracy and efficiency." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 195, + 369, + 207 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 195, + 369, + 207 + ], + "spans": [ + { + "bbox": [ + 306, + 195, + 369, + 207 + ], + "type": "text", + "content": "4.3. Ablation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 214, + 545, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 214, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 305, + 214, + 545, + 239 + ], + "type": "text", + "content": "Camera Model: We compare the effect of omnidirectional depth estimation under several camera models." + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 307, + 247, + 427, + 327 + ], + "blocks": [ + { + "bbox": [ + 307, + 247, + 427, + 327 + ], + "lines": [ + { + "bbox": [ + 307, + 247, + 427, + 327 + ], + "spans": [ + { + "bbox": [ + 307, + 247, + 427, + 327 + ], + "type": "image", + "image_path": "31cd0fae314bb9ccffb1da0be9141bb5eb358bb1ab89f5367e316df0a8556e76.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 328, + 328, + 412, + 338 + ], + "lines": [ + { + "bbox": [ + 328, + 328, + 412, + 338 + ], + "spans": [ + { + "bbox": [ + 328, + 328, + 412, + 338 + ], + "type": "text", + "content": "Original fisheye image" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 427, + 247, + 545, + 327 + ], + "blocks": [ + { + "bbox": [ + 427, + 247, + 545, + 327 + ], + "lines": [ + { + "bbox": [ + 427, + 247, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 427, + 247, + 545, + 327 + ], + "type": "image", + "image_path": "890c6965c93dcc230f404a60d00e96a5f7049674c33fd80d7b8c0fc18e6fa802.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 449, + 328, + 531, + 338 + ], + "lines": [ + { + "bbox": [ + 449, + 328, + 531, + 338 + ], + "spans": [ + { + "bbox": [ + 449, + 328, + 531, + 338 + ], + "type": "text", + "content": "Scaramuzza et al. [30]" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 307, + 339, + 428, + 421 + ], + "blocks": [ + { + "bbox": [ + 307, + 339, + 428, + 421 + ], + "lines": [ + { + "bbox": [ + 307, + 339, + 428, + 421 + ], + "spans": [ + { + "bbox": [ + 307, + 339, + 428, + 421 + ], + "type": "image", + "image_path": "5e3637eb032df4968b81aff04cae972c20c394d0371b66f2206e6a24a1bbd77f.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 345, + 421, + 389, + 431 + ], + "lines": [ + { + "bbox": [ + 345, + 421, + 389, + 431 + ], + "spans": [ + { + "bbox": [ + 345, + 421, + 389, + 431 + ], + "type": "text", + "content": "DSCM [36]" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 440, + 545, + 473 + ], + "lines": [ + { + "bbox": [ + 305, + 440, + 545, + 473 + ], + "spans": [ + { + "bbox": [ + 305, + 440, + 545, + 473 + ], + "type": "text", + "content": "Figure 8. Depth estimation results using different camera projection models. Our TSCM can best remove the fisheye distortion and obtain the most accurate result." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 427, + 339, + 545, + 421 + ], + "blocks": [ + { + "bbox": [ + 427, + 339, + 545, + 421 + ], + "lines": [ + { + "bbox": [ + 427, + 339, + 545, + 421 + ], + "spans": [ + { + "bbox": [ + 427, + 339, + 545, + 421 + ], + "type": "image", + "image_path": "87051fe40c3284da5f01fc05f5c0db868ca32308c169476eb8b97136391e35a1.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 461, + 421, + 510, + 430 + ], + "lines": [ + { + "bbox": [ + 461, + 421, + 510, + 430 + ], + "spans": [ + { + "bbox": [ + 461, + 421, + 510, + 430 + ], + "type": "text", + "content": "TSCM(Ours)" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 488, + 545, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 488, + 545, + 572 + ], + "spans": [ + { + "bbox": [ + 304, + 488, + 545, + 572 + ], + "type": "text", + "content": "As shown in Figure 8, the distortion of the fisheye lens is not sufficiently handled in the DSCM [36] and the model proposed by Scaramuzza et al. [30], which affects the accuracy of depth estimation and leads to serious distortion of objects in the point cloud. In contrast, the quality of the point cloud obtained by the TSCM is much better, with flat wall surfaces and more accurate distance measurements." + } + ] + } + ], + "index": 21 + }, + { + "type": "table", + "bbox": [ + 319, + 581, + 533, + 643 + ], + "blocks": [ + { + "bbox": [ + 47, + 360, + 287, + 372 + ], + "lines": [ + { + "bbox": [ + 47, + 360, + 287, + 372 + ], + "spans": [ + { + "bbox": [ + 47, + 360, + 287, + 372 + ], + "type": "text", + "content": "Table 2. Efficiency comparison with other deep learning methods." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 319, + 581, + 533, + 643 + ], + "lines": [ + { + "bbox": [ + 319, + 581, + 533, + 643 + ], + "spans": [ + { + "bbox": [ + 319, + 581, + 533, + 643 + ], + "type": "table", + "html": "
FoV2 sphere3 sphere (Ours)4 sphere
195°0.1530 px0.1528 px0.1528 px
220°0.2031 px0.1970 px0.1952 px
250°0.1365 px0.1255 px0.1222 px
", + "image_path": "878c214fba76d6b84ea654b595f2808e222e9619f9ee6f01e11edbf5084cf894.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "table_body" + } + ], + "index": 22 + }, + { + "bbox": [ + 305, + 650, + 545, + 673 + ], + "lines": [ + { + "bbox": [ + 305, + 650, + 545, + 673 + ], + "spans": [ + { + "bbox": [ + 305, + 650, + 545, + 673 + ], + "type": "text", + "content": "Table 3. Mean reprojection error for camera models with different number sphere." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 689, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 547, + 715 + ], + "type": "text", + "content": "We also compared the accuracy with different number of the refract sphere. Table 3 shows the results. It can be seen" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21535" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 68, + 533, + 435 + ], + "blocks": [ + { + "bbox": [ + 47, + 68, + 533, + 435 + ], + "lines": [ + { + "bbox": [ + 47, + 68, + 533, + 435 + ], + "spans": [ + { + "bbox": [ + 47, + 68, + 533, + 435 + ], + "type": "image", + "image_path": "1b4dbb8e5fda9efd61d3eb358a2ebefc510f5f02bf2bd4b3136c4c036ba3e295.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 453, + 546, + 475 + ], + "lines": [ + { + "bbox": [ + 46, + 453, + 546, + 475 + ], + "spans": [ + { + "bbox": [ + 46, + 453, + 546, + 475 + ], + "type": "text", + "content": "Figure 9. The images captured with our prototype and the inverse depth map and cloud points obtained by OmniVidar. It can be seen that OmniVidar has great generalization in various scenes, and performs well in low-textured areas." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 496, + 287, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 496, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 496, + 287, + 533 + ], + "type": "text", + "content": "that the camera model with 3 sphere can significantly improve the accuracy in large FoV scenes, while the improvement brought by more sphere is very limited." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 533, + 287, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 287, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 287, + 640 + ], + "type": "text", + "content": "Distortion Handle: To verify whether the improvement of depth estimation accuracy is helped by removing the fisheye image distortion using epipolar rectification, we estimate the depth directly using the original fisheye image and compare it with the undistorted method. Unlike the undistorted binocular stereo, we use the Sphere-Sweep algorithm [40] to construct the cost volume. This step does not contain trainable parameters, so it does not affect the network itself and ensures the fairness of the comparison." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 642, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 712 + ], + "type": "text", + "content": "The point cloud visualization of the results of the depth estimation of both is shown in Figure 8. It can be seen that when using original image, the depth information obtained from different cameras is not consistent. In contrast, the point cloud obtained by depth estimation after epipolar rectification shows the structural framework of the scene real-" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 496, + 402, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 496, + 402, + 509 + ], + "spans": [ + { + "bbox": [ + 306, + 496, + 402, + 509 + ], + "type": "text", + "content": "istrically and accurately." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 517, + 384, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 517, + 384, + 529 + ], + "spans": [ + { + "bbox": [ + 306, + 517, + 384, + 529 + ], + "type": "text", + "content": "4.4. Test in Real" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 536, + 545, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 536, + 545, + 597 + ], + "spans": [ + { + "bbox": [ + 304, + 536, + 545, + 597 + ], + "type": "text", + "content": "We test the performance of the OmniVidar algorithm on a real dataset. Figure 9 shows the inverse depth maps and point clouds obtained by OmniVidar on the real dataset. It can be seen that our network has a strong generalization to obtain high-quality point clouds on real datasets." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 609, + 378, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 609, + 378, + 621 + ], + "spans": [ + { + "bbox": [ + 306, + 609, + 378, + 621 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 545, + 713 + ], + "type": "text", + "content": "In this paper, we propose a novel, simple and effective system OmniVidar, to address the inefficiency of current omnidirectional depth estimation methods due to the extensive use of 3D convolution and the poor accuracy due to the lack of explicit handling of distortion in fisheye images. Our method outperforms all other methods in terms of time consumption and accuracy." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21536" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 286, + 113 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 286, + 113 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 286, + 113 + ], + "type": "text", + "content": "[1] Sameer Agarwal, Keir Mierle, and The Ceres Solver Team. Ceres solver. http://ceres-solver.org. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 115, + 287, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 115, + 287, + 148 + ], + "spans": [ + { + "bbox": [ + 53, + 115, + 287, + 148 + ], + "type": "text", + "content": "[2] Simon Baker and Shree K. Nayar. A theory of single-viewpoint catadioptric image formation. Int. J. Comput. Vision, 35(2):175-196, nov 1999. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 149, + 287, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 149, + 287, + 192 + ], + "spans": [ + { + "bbox": [ + 53, + 149, + 287, + 192 + ], + "type": "text", + "content": "[3] Wei Bao, Wei Wang, Yuhua Xu, Yulan Guo, Siyu Hong, and Xiaohu Zhang. Instereo2k: a large real dataset for stereo matching in indoor scenes. Science China Information Sciences, 63(11):212101, 2019. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 194, + 287, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 194, + 287, + 249 + ], + "spans": [ + { + "bbox": [ + 53, + 194, + 287, + 249 + ], + "type": "text", + "content": "[4] D. J. Butler, J. Wulff, G. B. Stanley, and M. J. Black. A naturalistic open source movie for optical flow evaluation. In A. Fitzgibbon et al. (Eds.), editor, European Conf. on Computer Vision (ECCV), Part IV, LNCS 7577, pages 611-625. Springer-Verlag, Oct. 2012. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 251, + 287, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 251, + 287, + 293 + ], + "spans": [ + { + "bbox": [ + 53, + 251, + 287, + 293 + ], + "type": "text", + "content": "[5] Jia-Ren Chang and Yong-Sheng Chen. Pyramid stereo matching network. In 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5410-5418, 2018. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 296, + 287, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 296, + 287, + 350 + ], + "spans": [ + { + "bbox": [ + 53, + 296, + 287, + 350 + ], + "type": "text", + "content": "[6] Xuelian Cheng, Yiran Zhong, Mehrtash Harandi, Yuchao Dai, Xiaojun Chang, Tom Drummond, Hongdong Li, and Zongyuan Ge. Hierarchical neural architecture search for deep stereo matching. arXiv preprint arXiv:2010.13501, 2020.7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 352, + 287, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 352, + 287, + 385 + ], + "spans": [ + { + "bbox": [ + 53, + 352, + 287, + 385 + ], + "type": "text", + "content": "[7] Taco S. Cohen, Mario Geiger, Jonas Kohler, and Max Welling. Spherical CNNs. In International Conference on Learning Representations, 2018. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 387, + 287, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 387, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 53, + 387, + 287, + 453 + ], + "type": "text", + "content": "[8] Benjamin Coors, Alexandru Paul Condurache, and Andreas Geiger. Spherenet: Learning spherical representations for detection and classification in omnidirectional images. In Vittorio Ferrari, Martial Hebert, Cristian Sminchisescu, and Yair Weiss, editors, Computer Vision – ECCV 2018, pages 525–541. Springer International Publishing, 2018. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 454, + 287, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 454, + 287, + 497 + ], + "spans": [ + { + "bbox": [ + 53, + 454, + 287, + 497 + ], + "type": "text", + "content": "[9] Marc Eder and Jan-Michael Frahm. Convolutions on spherical images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, June 2019. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 499, + 287, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 499, + 287, + 543 + ], + "spans": [ + { + "bbox": [ + 48, + 499, + 287, + 543 + ], + "type": "text", + "content": "[10] Marc Eder, Mykhailo Shvets, John Lim, and Jan-Michael Frahm. Tangent images for mitigating spherical distortion. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12423-12431, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 544, + 287, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 287, + 577 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 287, + 577 + ], + "type": "text", + "content": "[11] Andrea Fusiello, Emanuele Trucco, and Alessandro Verri. A compact algorithm for rectification of stereo pairs. Machine Vision and Applications, 12(1):16-22, 2000. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 579, + 287, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 579, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 48, + 579, + 287, + 622 + ], + "type": "text", + "content": "[12] Adrien Gaidon, Qiao Wang, Yohann Cabon, and Eleonora Vig. Virtualworlds as proxy for multi-object tracking analysis. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4340-4349, 2016. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 624, + 287, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 624, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 48, + 624, + 287, + 666 + ], + "type": "text", + "content": "[13] Wenliang Gao and Shaojie Shen. Dual-fisheye omnidirectional stereo. In 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 6715-6722, 2017. 1, 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 287, + 713 + ], + "type": "text", + "content": "[14] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE Conference on Computer Vision and Pattern Recognition, pages 3354-3361, 2012. 6" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[15] Heiko Hirschmuller. Stereo processing by semiglobal matching and mutual information. IEEE Transactions on Pattern Analysis and Machine Intelligence, 30(2):328-341, 2008. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 108, + 545, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 108, + 545, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 108, + 545, + 162 + ], + "type": "text", + "content": "[16] Christian Häne, Lionel Heng, Gim Hee Lee, Alexey Sizov, and Marc Pollefeys. Real-time direct dense matching on fisheye images using plane-sweeping stereo. In 2014 2nd International Conference on 3D Vision, volume 1, pages 57-64, 2014. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 165, + 545, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 165, + 545, + 209 + ], + "spans": [ + { + "bbox": [ + 307, + 165, + 545, + 209 + ], + "type": "text", + "content": "[17] Carlos Jaramillo, Roberto G. Valenti, Ling Guo, and Jizhong Xiao. Design and analysis of a single—camera omnistereo sensor for quadrotor micro aerial vehicles (mavs). Sensors, 16(2), 2016. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 211, + 545, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 211, + 545, + 255 + ], + "spans": [ + { + "bbox": [ + 308, + 211, + 545, + 255 + ], + "type": "text", + "content": "[18] J. Kannala and S.S. Brandt. A generic camera model and calibration method for conventional, wide-angle, and fish-eye lenses. IEEE Transactions on Pattern Analysis and Machine Intelligence, 28(8):1335-1340, 2006. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 258, + 545, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 258, + 545, + 289 + ], + "spans": [ + { + "bbox": [ + 308, + 258, + 545, + 289 + ], + "type": "text", + "content": "[19] Bogdan Khomutenko, Gaetan Garcia, and Philippe Martinet. An enhanced unified camera model. IEEE Robotics and Automation Letters, 1(1):137-144, 2016. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 292, + 545, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 292, + 545, + 357 + ], + "spans": [ + { + "bbox": [ + 308, + 292, + 545, + 357 + ], + "type": "text", + "content": "[20] Laurent Kneip, Hongdong Li, and Yongduek Seo. Upnp: An optimal o(n) solution to the absolute pose problem with universal applicability. In David Fleet, Tomas Pajdla, Bernt Schiele, and Tinne Tuytelaars, editors, Computer Vision - ECCV 2014, pages 127-142. Springer International Publishing, 2014. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 360, + 545, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 360, + 545, + 425 + ], + "spans": [ + { + "bbox": [ + 308, + 360, + 545, + 425 + ], + "type": "text", + "content": "[21] Ren Komatsu, Hiromitsu Fujii, Yusuke Tamura, Atsushi Yamashita, and Hajime Asama. " + }, + { + "bbox": [ + 308, + 360, + 545, + 425 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 308, + 360, + 545, + 425 + ], + "type": "text", + "content": " depth estimation from multiple fisheye images with origami crown representation of icosahedron. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 10092-10099, 2020. 2, 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 428, + 545, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 428, + 545, + 472 + ], + "spans": [ + { + "bbox": [ + 308, + 428, + 545, + 472 + ], + "type": "text", + "content": "[22] Jiankun Li, Peisen Wang, Pengfei Xiong, Tao Cai, Ziwei Yan, Lei Yang, Jiangyu Liu, Haoqiang Fan, and Shuaicheng Liu. Practical stereo matching via cascaded recurrent network with adaptive correlation, 2022. 3, 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 475, + 545, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 475, + 545, + 507 + ], + "spans": [ + { + "bbox": [ + 308, + 475, + 545, + 507 + ], + "type": "text", + "content": "[23] Weiming Li and Y. F. Li. Single-camera panoramic stereo imaging system with a fisheye lens and a convex mirror. Opt. Express, 19(7):5855-5867, Mar 2011. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 510, + 545, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 510, + 545, + 552 + ], + "spans": [ + { + "bbox": [ + 308, + 510, + 545, + 552 + ], + "type": "text", + "content": "[24] Lahav Lipson, Zachary Teed, and Jia Deng. Raft-stereo: Multilevel recurrent field transforms for stereo matching. In 2021 International Conference on 3D Vision (3DV), pages 218-227, 2021. 3, 6, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 555, + 545, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 555, + 545, + 598 + ], + "spans": [ + { + "bbox": [ + 308, + 555, + 545, + 598 + ], + "type": "text", + "content": "[25] Bin Lv, Hao Xu, Jianqing Wu, Yuan Tian, Sheng Tian, and Suoyao Feng. Revolution and rotation-based method for roadside lidar data integration. Optics & Laser Technology, 119:105571, 2019. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 601, + 545, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 601, + 545, + 666 + ], + "spans": [ + { + "bbox": [ + 308, + 601, + 545, + 666 + ], + "type": "text", + "content": "[26] Nikolaus Mayer, Eddy Ilg, Philip Häusser, Philipp Fischer, Daniel Cremers, Alexey Dosovitskiy, and Thomas Brox. A large dataset to train convolutional networks for disparity, optical flow, and scene flow estimation. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4040-4048, 2016. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 545, + 713 + ], + "type": "text", + "content": "[27] Christopher Mei and Patrick Rives. Single view point omnidirectional camera calibration from planar grids. In Proceedings 2007 IEEE International Conference on Robotics and Automation, pages 3945-3950, 2007. 2, 4" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21537" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[28] M. Menze, C. Heipke, and A. Geiger. Joint 3d estimation of vehicles and scene flow. ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences, II-3/W5:427-434, 2015. 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 172 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 172 + ], + "type": "text", + "content": "[29] Andreas Meuleman, Hyeonjoong Jang, Daniel S. Jeon, and Min H. Kim. Real-time sphere sweeping stereo from multiview fisheye images. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11418-11427, 2021. 2, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 174, + 287, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 174, + 287, + 218 + ], + "spans": [ + { + "bbox": [ + 48, + 174, + 287, + 218 + ], + "type": "text", + "content": "[30] Davide Scaramuzza, Agostino Martinelli, and Roland Siegwart. A toolbox for easily calibrating omnidirectional cameras. In 2006 IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 5695-5701, 2006. 2, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 219, + 287, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 219, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 48, + 219, + 287, + 251 + ], + "type": "text", + "content": "[31] Johannes Schneider, Cyril Stachniss, and Wolfgang Forstner. On the accuracy of dense fisheye stereo. IEEE Robotics and Automation Letters, 1(1):227-234, 2016. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 253, + 287, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 253, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 48, + 253, + 287, + 297 + ], + "type": "text", + "content": "[32] Miriam Schonbein and Andreas Geiger. Omnidirectional 3d reconstruction in augmented manhattan worlds. In 2014 IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 716-723, 2014. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 298, + 287, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 298, + 287, + 363 + ], + "spans": [ + { + "bbox": [ + 48, + 298, + 287, + 363 + ], + "type": "text", + "content": "[33] Xiao Song, Xu Zhao, Hanwen Hu, and Liangji Fang. Edgestereo: A context integrated residual pyramid network for stereo matching. In C.V. Jawahar, Hongdong Li, Greg Mori, and Konrad Schindler, editors, Computer Vision – ACCV 2018, pages 20–35. Springer International Publishing, 2019. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 365, + 287, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 365, + 287, + 408 + ], + "spans": [ + { + "bbox": [ + 48, + 365, + 287, + 408 + ], + "type": "text", + "content": "[34] Yu-Chuan Su and Kristen Grauman. Kernel transformer networks for compact spherical convolution. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9434-9443, 2019. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 410, + 287, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 410, + 287, + 442 + ], + "spans": [ + { + "bbox": [ + 48, + 410, + 287, + 442 + ], + "type": "text", + "content": "[35] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow. ECCV, 12347(1):402-419, 2020. 2, 3, 5, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 444, + 287, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 444, + 287, + 487 + ], + "spans": [ + { + "bbox": [ + 48, + 444, + 287, + 487 + ], + "type": "text", + "content": "[36] Vladyslav Usenko, Nikolaus Demmel, and Daniel Cremers. The double sphere camera model. In 2018 International Conference on 3D Vision (3DV), pages 552-560, 2018. 2, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 489, + 287, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 489, + 287, + 554 + ], + "spans": [ + { + "bbox": [ + 48, + 489, + 287, + 554 + ], + "type": "text", + "content": "[37] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Ilia Polosukhin. Attention is all you need. In Proceedings of the 31st International Conference on Neural Information Processing Systems, NIPS'17, pages 6000-6010, Red Hook, NY, USA, 2017. Curran Associates Inc. 2, 5, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 555, + 287, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 555, + 287, + 599 + ], + "spans": [ + { + "bbox": [ + 48, + 555, + 287, + 599 + ], + "type": "text", + "content": "[38] Hengli Wang, Rui Fan, Peide Cai, and Ming Liu. Pvstereo: Pyramid voting module for end-to-end self-supervised stereo matching. IEEE Robotics and Automation Letters, 6(3):4353-4360, 2021. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 601, + 287, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 287, + 645 + ], + "type": "text", + "content": "[39] Changhee Won, Jongbin Ryu, and Jongwoo Lim. Omnimvss: End-to-end learning for omnidirectional stereo matching. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 8986-8995, 2019. 2, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "text", + "content": "[40] Changhee Won, Jongbin Ryu, and Jongwoo Lim. Sweepnet: Wide-baseline omnidirectional depth estimation. In 2019 International Conference on Robotics and Automation (ICRA), pages 6073-6079, 2019. 2, 6, 7, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[41] Changhee Won, Jongbin Ryu, and Jongwoo Lim. End-to-end learning for omnidirectional stereo matching with un" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 273 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "certainty prior. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(11):3850-3862, 2021. 2, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 96, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 150 + ], + "type": "text", + "content": "[42] Changhee Won, Hochang Seok, Zhaopeng Cui, Marc Pollefeys, and Jongwoo Lim. Omnislam: Omnidirectional localization and dense mapping for wide-baseline multicamera systems. In 2020 IEEE International Conference on Robotics and Automation (ICRA), pages 559-566, 2020. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 152, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 205 + ], + "type": "text", + "content": "[43] Hao Xu, Yichen Zhang, Boyu Zhou, Luqi Wang, Xinjie Yao, Guotao Meng, and Shaojie Shen. Omni-swarm: A decentralized omnidirectional visual-inertial-uwb state estimation system for aerial swarm. arXiv preprint arXiv:2103.04131, 2021. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 207, + 545, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 207, + 545, + 273 + ], + "spans": [ + { + "bbox": [ + 307, + 207, + 545, + 273 + ], + "type": "text", + "content": "[44] Feihu Zhang, Xiaojuan Qi, Ruigang Yang, Victor Prisacariu, Benjamin Wah, and Philip Torr. Domain-invariant stereo matching networks. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, Computer Vision - ECCV 2020, pages 420-439. Springer International Publishing, 2020. 6" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "21538" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/Omnimatte3D_ Associating Objects and Their Effects in Unconstrained Monocular Video/5700f2ea-1d90-4ebb-a4e2-33ec303e3c7c_content_list.json b/2023/Omnimatte3D_ Associating Objects and Their Effects in Unconstrained Monocular Video/5700f2ea-1d90-4ebb-a4e2-33ec303e3c7c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6d6004b412126c5c6c6ca1fec340c6f3bca3650e --- /dev/null +++ b/2023/Omnimatte3D_ Associating Objects and Their Effects in Unconstrained Monocular Video/5700f2ea-1d90-4ebb-a4e2-33ec303e3c7c_content_list.json @@ -0,0 +1,2280 @@ +[ + { + "type": "text", + "text": "Omnimatte3D: Associating Objects and their Effects in Unconstrained Monocular Video", + "text_level": 1, + "bbox": [ + 218, + 130, + 751, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mohammed Suhail1,2", + "bbox": [ + 186, + 202, + 359, + 219 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Erika Lu4 Leonid Sigal1,2,3", + "bbox": [ + 354, + 203, + 483, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhengqi Li", + "bbox": [ + 519, + 203, + 620, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Noah Snavely4", + "bbox": [ + 660, + 203, + 779, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Forrester Cole", + "bbox": [ + 529, + 220, + 651, + 237 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1University of British Columbia", + "bbox": [ + 189, + 239, + 383, + 255 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Vector Institute for AI", + "bbox": [ + 397, + 239, + 537, + 255 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3Canada CIFAR AI Chair", + "bbox": [ + 555, + 241, + 709, + 255 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4Google", + "bbox": [ + 725, + 241, + 779, + 255 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/03bbd923cf0a354ef93ce235c54b6a0a99f3e51d471523a549ad38ddaedab5a6.jpg", + "image_caption": [ + "Input RGB" + ], + "image_footnote": [], + "bbox": [ + 107, + 260, + 250, + 452 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5dae43cac38bb1b03b87a1cbd6fce6b1709f2284bb6e2fb5cfbff9cc40a2afe8.jpg", + "image_caption": [ + "Input Masks" + ], + "image_footnote": [], + "bbox": [ + 250, + 260, + 392, + 452 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1d5c43233d2a14d35507f132cc7b44a884ea6381a5272c0499935ef815245e23.jpg", + "image_caption": [ + "Input Depth", + "Figure 1. Layer decomposition under strong camera parallax. Given an input video with unconstrained camera motion and approximate object masks and depth (left), our method estimates a layered representation composed of a background layer and object layers containing the subjects of interest and their associated effects (e.g. shadows). Results of combining object layers and background are shown on right." + ], + "image_footnote": [], + "bbox": [ + 393, + 260, + 535, + 452 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f44f80334f95e168cccff91f53db06d8f946618a8a5becc283c8a00d2e80e312.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 537, + 353, + 570, + 369 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e8374b32729bcb5ef8f552f80b4cce507e6056647ad005b27a5b7efac74d6f89.jpg", + "image_caption": [ + "Layer 1 + Background" + ], + "image_footnote": [], + "bbox": [ + 571, + 260, + 714, + 452 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b540f14db5bb39921df5bb5dc2daddc55588d46af2194e9130a6bef22bc9cd1f.jpg", + "image_caption": [ + "Layer 2 + Background" + ], + "image_footnote": [], + "bbox": [ + 715, + 260, + 857, + 452 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 525, + 313, + 540 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We propose a method to decompose a video into a background and a set of foreground layers, where the background captures stationary elements while the foreground layers capture moving objects along with their associated effects (e.g. shadows and reflections). Our approach is designed for unconstrained monocular videos, with an arbitrary camera and object motion. Prior work that tackles this problem assumes that the video can be mapped onto a fixed 2D canvas, severely limiting the possible space of camera motion. Instead, our method applies recent progress in monocular camera pose and depth estimation to create a full, RGBD video layer for the background, along with a video layer for each foreground object. To solve the underconstrained decomposition problem, we propose a new loss formulation based on multi-view consistency. We test our method on challenging videos with complex camera motion and show significant qualitative improvement over current approaches.", + "bbox": [ + 73, + 556, + 473, + 815 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 844, + 207, + 859 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Decomposing a video into meaningful layers (as show in Fig. 1) is a long-standing and complex problem [42] that has", + "bbox": [ + 75, + 869, + 468, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "seen a recent surge in progress with the application of deep neural networks [17, 24, 25, 46]. A challenging variant of layer decomposition is the omnimatte [25] task, which aims to separate an input video into a background and multiple foreground layers, each containing an object of interest along with its correlated effects such as shadows and reflections, thus enabling video editing applications such as object removal, background replacement and retiming. The original work on omnimatte [24, 25] introduced a self-supervised approach for decomposing a video by assuming the background can be unwrapped onto a single static 2D background canvas using homography warping. Following work [17, 46] relaxed the homography restriction, but maintained the necessity of unwrapping onto a 2D canvas.", + "bbox": [ + 496, + 527, + 893, + 738 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This 2D modelling, however, limits the applicability of these methods to camera motions that have limited or zero parallax; intuitively, only panning camera motions are allowed. If the camera's center of projection moves a significant distance over the video, 2D methods are unable to learn an accurate background and are forced to place the background detail in a foreground layer (Figure 2). Since camera parallax is quite common, 2D methods are severely restricted in practice.", + "bbox": [ + 496, + 743, + 893, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To handle camera parallax, we exploit another recent line", + "bbox": [ + 519, + 885, + 890, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "630", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/30b7cafafa870b70a4ca6193ce0fdc9d1717074df16336e5bc5abd603a32befc.jpg", + "image_caption": [ + "Figure 2. Existing methods fail on scenes with parallax. Top row: Omnimatte [25] fails due to inaccurate homography registration, reconstructing the majority of the video in the foreground layer. Middle row: Neural-Atlas [17] incorrectly places trees in the foreground layer and struggles to inpaint the person region. Bottom row: Our method obtains a clean background and captures the person and their shadow in the foreground layer." + ], + "image_footnote": [], + "bbox": [ + 84, + 89, + 467, + 250 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "of work on estimating 3D camera position and depth maps from casually-captured video with moving objects [18,26, 48, 49]. Instead of a 2D layer decomposition, we propose to learn a 3D background model that varies per-frame and includes inpainted RGB and depth in the regions occluded by the foreground. Allowing the background to vary per-frame, rather than assuming a global, static canvas, creates an ill-posed decomposition problem, as variation in each input pixel could be explained by a foreground object, background, or both. To resolve this ambiguity, we enforce that the background varies slowly over time with 3D multi-view constraints using dense depth and camera pose estimates [48]. As shown in Fig. 2, our model can successfully separate the layers obtaining a clean background RGB whilst capturing only the person and their shadow in the foreground layer.", + "bbox": [ + 75, + 387, + 470, + 616 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Contributions. Our main contribution is a novel video decomposition method capable of accurately separating an input video with complex object and camera motion (e.g., parallax) into a background and object layers. To address the ill-posed decomposition problem, we design regularization terms based on multi-view consistency and smoothness priors. The resulting model produces clean decompositions of real-world videos where previous methods fail.", + "bbox": [ + 75, + 618, + 470, + 739 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 76, + 753, + 218, + 768 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Layered video representations. There is a large body of work on decomposing videos into layers based on appearance and motion cues [4, 8, 16, 20, 36, 42]. Layered video representations have proven useful for a variety of applications, including view synthesis for static scenes [39, 50], reflection removal [1, 2], segmentation [9], game deconstruction [38], and text-based video editing [3]. Recently, Omnimatte [24, 25] aimed to decompose a video into layers", + "bbox": [ + 75, + 780, + 470, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "that group objects with their correlated effects, allowing the user to perform editing tasks such as object removal and retiming. However, as discussed in Section 1, these works use a constrained background model which limits their use case to panning camera motions only. Subsequent work [17,46] relaxes this assumption using general image warps, instead of homographies, but still cannot handle significant parallax; moreover, their goal is to learn a per-frame mapping to global sprites to enable applying consistent edits (e.g. style transfer) rather than targeting object removal or video stabilization.", + "bbox": [ + 496, + 90, + 893, + 243 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Monocular video depth. Multi-view stereo based methods [23,40] incorporate motion estimation and multi-view reconstruction to predict depth by combining information from multiple reference frames. Hybrid depth methods combine single-view depth estimates from a pretrained network with geometric reasoning to obtain coherent depth predictions. With depth maps from monocular video as initialization, these methods use correspondence to enforce consistency by either assuming static regions [26] or by modeling the scene flow [49]. Recent methods simultaneously estimate depth maps and camera poses [19, 48], exploiting the depth prior to resolve camera motion ambiguities. We rely on CasualSAM [48] to obtain camera poses and initial depth estimates as their method performs well on casually captured video.", + "bbox": [ + 496, + 265, + 895, + 477 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Video completion and object removal. Patch-based methods for video inpainting [7, 12, 14, 28, 43] complete the missing regions by borrowing pixels from other frames, relying on correspondence estimated by homography or optical flow. Learning-based works have explored the use of 3D convolutions [5, 41], attention mechanisms [21, 30, 47] and flow-based methods [10, 45] to achieve more realistic and consistent inpainting. These methods assume that the entire region to inpaint is specified by an input mask, whereas our method jointly learns where to inpaint the background while separating foreground objects and their correlated effects.", + "bbox": [ + 496, + 500, + 893, + 667 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "NeRF for dynamic scenes. With the recent success of neural fields [29,44] for view synthesis of static scenes, there has been interest in their adoption for dynamic scenes [11,22,27]. These works learn two NeRF models that decompose a scene into a static and dynamic region. However, these methods only support a single foreground layer and do not account for correlated effects such as shadows. ST-NeRF [15] proposed to learn separate NeRFs for each person to obtain an editable neural representation. Their method, however, requires each frame in the video to be captured from 16 different viewpoints simultaneously. Unlike NeRF methods, our method represents the video with multi-layer mesh geometry where object layers include associated effects, allowing editing as well as fast rerendering of the scene.", + "bbox": [ + 496, + 689, + 893, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "631", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/54af525eb386bfcbf3aafd896773ab954519b7222151a588440d3c72b57c61dc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 89, + 88, + 880, + 280 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4f825ba3e51340be93023f21cfac849c715535b074ff6b18c4c9d509fd56cc40.jpg", + "image_caption": [ + "Figure 3. Model Overview. Given a frame from the video along with an input mask, disparity and known camera pose, we first extract features by stacking the input and passing them though a 2D UNet architecture. These features are input to the background and foreground networks. The background network predicts the inpainted background RGB and disparity. The foreground networks predict an RGB-A image for each foreground layer. For scenes with multiple layers, the RGB-A image for each layer is predicted using independent networks." + ], + "image_footnote": [], + "bbox": [ + 96, + 282, + 875, + 401 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 76, + 501, + 168, + 517 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given an input video, with known camera poses and initial depth estimates, the goal is to predict a per-frame layer decomposition consisting of a background layer and $N$ foreground layers encompassing objects of interest along with their associated effects. To this end, we propose a new method that predicts a layered separation of an input video without relying on a 2D background canvas, thereby allowing for greater flexibility of camera motion. We show an overview of our model in Figure 3. We optimize our model to output layers that reconstruct a single target video subject to a projection consistency loss. For each frame in the video, we estimate a layered representation consisting of background and foreground object layers where the contents of the object layers are conditioned on their corresponding input mask. Post-optimization, depth for foreground objects can be extracted from the input depth to produce a Layered Depth Image (LDI). Unlike prior works that rely on deep priors for inpainting the depth and color of occluded regions in the background, our method relies on multi-view consistency losses to steer the inpainting process.", + "bbox": [ + 75, + 527, + 473, + 832 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Layer Decomposition Networks", + "text_level": 1, + "bbox": [ + 76, + 845, + 356, + 861 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We represent a frame at time $a$ in the video as $F_{a} = \\left(I_{a},D_{a},\\{M_{a}^{i}\\}_{i = 1}^{N}\\right)$ , where $I_{a}$ is the RGB frame, $D_{a}$ is the", + "bbox": [ + 76, + 869, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "disparity, and $\\{M_a^i\\}_{i = 1}^N$ are the input masks for $N$ objects in the video. We denote the camera extrinsics for the frame $F_{a}$ as $[R_a t_a]$ where $R_{a}$ is the rotation matrix and $t_a$ is the translation vector. The camera intrinsics are denoted as $K_{a} = [f_{x},f_{y},c_{x},c_{y}]$ where $f_{x}$ and $f_{y}$ are the focal lengths and $c_{x},c_{y}$ is the principal point.", + "bbox": [ + 496, + 502, + 890, + 594 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For each frame $(F_{a})$ in the input video, our method first extracts features by passing it to a feature extractor network modeled using a 2D UNet [35] architecture. The input to the feature extractor is constructed by concatenating input masks with the masked frame RGB and disparity along the channel dimension. The masked RGB and disparity are obtained as", + "bbox": [ + 496, + 594, + 890, + 685 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {Y} _ {a} = \\left[ Y _ {a} \\odot M _ {a} ^ {b g} \\| Y _ {a} \\odot M _ {a} ^ {1} \\| \\dots \\| Y _ {a} \\odot M _ {a} ^ {N} \\right] \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 690, + 890, + 710 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $Y_{a}$ is the concatenated RGB-D, $M_{a}^{bg} = 1 - \\bigcup_{i=1}^{N} M_{a}^{i}$ is an approximate mask for the background and $\\| \\cdot$ is the concatenation operation. These features act as input to both the background and foreground object networks.", + "bbox": [ + 496, + 717, + 890, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The background network (shown in green in Fig. 3) consists of a 2D UNet and a convolutional neural network (CNN) that predicts the inpainted background color $(\\mathcal{C}_a^{bg})$ and depth $(\\mathcal{D}_a^{bg})$ respectively. The foreground networks are separate CNNs that predict RGB $(\\mathcal{C}_a^i)$ and alpha $(\\mathcal{A}_a^i)$ images for each foreground layer independently. These predicted outputs, along with the input disparity, are combined to form an LDI (see supplement for details).", + "bbox": [ + 496, + 780, + 893, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "632", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The feature extractor and background RGB UNets consist of five down-sampling and up-sampling layers with three residual blocks at every resolution. The background disparity and foreground layer prediction networks are comprised of single convolutional layers.", + "bbox": [ + 76, + 90, + 468, + 167 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Differentiable Rendering", + "text_level": 1, + "bbox": [ + 76, + 174, + 305, + 191 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In addition to a reconstruction loss, our method relies on a projection consistency term (described in Section 3.3) that requires differentiable projections and renderings of the predicted background layer to different time steps. We achieve this using a differentiable rasterization framework [6].", + "bbox": [ + 76, + 196, + 468, + 272 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To project the predicted background layer from a source time step $a$ to target time step $b$ , we first construct a mesh $\\mathcal{M}_a = (\\mathcal{V}_a, \\mathcal{F}_a)$ , where $\\mathcal{V}_a$ and $\\mathcal{F}_a$ are the mesh vertices and faces. Each vertex in the mesh corresponds to an image pixel. To obtain the vertex coordinate for a pixel, we unproject its homogeneous coordinates $x = [i,j,1] \\in \\mathbf{RP}^2$ using the predicted disparity as", + "bbox": [ + 76, + 273, + 470, + 378 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nV _ {a} ^ {x} = d _ {i j} \\left(\\frac {\\left(i - c _ {x}\\right)}{f _ {x}}, \\frac {\\left(j - c _ {y}\\right)}{f _ {y}}, 1\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 155, + 385, + 468, + 419 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $d_{ij}$ is the depth at pixel $(i,j)$ . The attributes for the vertices are composed of the predicted RGB, an alpha masking consisting of all ones, and the pixel coordinates. The faces for the mesh are constructed by adding an edge between vertices of neighboring pixels (bottom right in Fig. 3). The view-projection matrix to render the predicted background from a target camera is computed as", + "bbox": [ + 76, + 425, + 472, + 532 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nH _ {a \\rightarrow b} = P _ {a} \\left[\\begin{array}{c c}R _ {a} ^ {T} R _ {b}&R _ {a} ^ {T} \\left(t _ {b} - t _ {a}\\right)\\\\\\mathbf {0}&1\\end{array}\\right] \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 150, + 537, + 468, + 571 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $P_{a}$ is the perspective matrix that maps camera-space to clip space. The view-projection matrix along with the estimated mesh is passed through a differentiable rasterizer that returns an RGB-A $(\\mathcal{C}_{a\\to b}^{bg},\\mathcal{A}_{a\\to b}^{bg})$ image corresponding to the projection of the background to the target time step.", + "bbox": [ + 76, + 577, + 468, + 652 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Losses", + "text_level": 1, + "bbox": [ + 76, + 660, + 163, + 674 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The total loss to train our model is:", + "bbox": [ + 96, + 683, + 331, + 696 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\lambda_ {1} \\mathcal {L} _ {\\text {r e c o n}} + \\lambda_ {2} \\mathcal {L} _ {\\text {p r o j}} + \\lambda_ {3} \\mathcal {L} _ {\\text {m a s k}} + \\lambda_ {4} \\mathcal {L} _ {\\text {d i s p}} + \\lambda_ {5} \\mathcal {L} _ {\\text {r e g}} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 705, + 468, + 723 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We describe each loss term in the following sections. Please see the supplement for values of the scalar weights $\\lambda$ .", + "bbox": [ + 76, + 729, + 468, + 761 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.1 Reconstruction Loss", + "text_level": 1, + "bbox": [ + 76, + 777, + 272, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For a source frame $F_{a}$ we reconstruct the image from the predicted background RGB and foreground RGBA layers using back-to-front compositing [33] (bottom left in Fig. 3). During training, we minimize the $\\mathcal{L}_2$ loss between the original RGB image and the reconstruction.", + "bbox": [ + 76, + 801, + 468, + 876 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r e c o n}} = \\left\\| \\left(I _ {a} - f \\left(\\mathcal {C} _ {a} ^ {b g}, \\left\\{\\mathcal {C} _ {a} ^ {i}, \\mathcal {A} _ {a} ^ {1} \\right\\} _ {i = 1} ^ {N}\\right) \\right\\| _ {2} \\right. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 883, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $f$ is the back-to-front over-compositing function.", + "bbox": [ + 498, + 90, + 867, + 106 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A large number of layer decompositions can reproduce the original frame and minimize the reconstruction loss. To guide the foreground layers to be semantically meaningful, we incorporate an $\\mathcal{L}_1$ loss between the predicted foreground alphas and the input masks.", + "bbox": [ + 498, + 106, + 890, + 181 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {m a s k}} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\| M _ {a} ^ {i} - \\mathcal {A} _ {a} ^ {i} \\| _ {1} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 191, + 890, + 232 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since we want the foreground alphas to be able to capture effects beyond the input mask, the weight for this loss is decayed as described in Sec. 3.4.", + "bbox": [ + 498, + 241, + 890, + 286 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.2 Projection Consistency Loss", + "text_level": 1, + "bbox": [ + 500, + 305, + 750, + 320 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The reconstruction loss on the over-composed image does not provide any gradients for the occluded region in the background. Prior works overcome this challenge by learning a background model that maps from a static 2D canvas. However, since our model predicts the background for each frame independently, we require additional regularization to prevent random artifacts or foreground details from appearing in occluded regions. We thus introduce a projection consistency loss on the background prediction. Each batch in our input is composed of three frames from the video. We denote these frames as $F_{a}$ , $F_{b}$ and $F_{c}$ . The frames $F_{a}$ and $F_{b}$ act as source images for the target frame $F_{c}$ . The criteria used for selecting the frames is described in Section 3.4.", + "bbox": [ + 498, + 328, + 893, + 523 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To estimate the projection consistency loss, we project the predicted background RGBs $(\\mathcal{C}_a^{bg}$ and $\\mathcal{C}_b^{bg})$ for the two source frames, $F_{a}$ and $F_{b}$ , using the method described in Sec. 3.2. We use the predicted disparities $(\\mathcal{D}_a^{bg}$ and $\\mathcal{D}_b^{bg})$ to construct the mesh and set the alpha values to one for all pixels. We then apply an $\\mathcal{L}_2$ loss between the projected source backgrounds and the predicted target background.", + "bbox": [ + 498, + 525, + 890, + 630 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {p n j} ^ {r g b} = \\frac {1}{2} \\sum_ {k \\in \\{a, b \\}} \\| \\mathcal {A} _ {k \\rightarrow c} ^ {b g} \\odot \\left(\\mathcal {C} _ {k \\rightarrow c} ^ {b g} - \\mathcal {C} _ {c} ^ {b g}\\right) \\| _ {2} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 549, + 638, + 890, + 678 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\odot$ denotes element wise multiplication. We mask the loss using the projected alphas $(\\mathcal{A}_{k\\to c}^{bg})$ to prevent noisy gradients from the boundaries of projection.", + "bbox": [ + 498, + 686, + 890, + 732 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Similarly, to encourage consistent inpainting of depth we apply a similar projection loss,", + "bbox": [ + 498, + 732, + 890, + 762 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {p r o j} ^ {c o o r d} = \\frac {1}{2} \\sum_ {k \\in \\{a, b \\}} \\| \\mathcal {A} _ {k \\rightarrow c} ^ {b g} \\odot \\left(\\mathcal {X} _ {k \\rightarrow c} ^ {b g} - \\mathcal {X} _ {c} ^ {b g}\\right) \\| _ {2} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 771, + 890, + 810 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{X}_c$ are the world coordinates for pixels in the frame $F_{c}$ obtained using the predicted background disparity $\\mathcal{D}_c^{bg}$ . $\\mathcal{X}_{k\\to c}^{bg}$ are the projections of the world coordinate from frame $F_{k}$ to $F_{c}$ and $\\mathcal{A}_{k\\to c}^{bg}$ are the projected alphas. The projection loss $\\mathcal{L}_{proj}$ is the sum of $\\mathcal{L}_{proj}^{rgb}$ and $\\mathcal{L}_{proj}^{coord}$ .", + "bbox": [ + 498, + 819, + 893, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "633", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.3 Alpha Regularization", + "text_level": 1, + "bbox": [ + 76, + 90, + 281, + 104 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "While the projection loss encourages the background predictions to be consistent across frames, it can also lead the model toward undesirable or degenerate solutions. For example, a model that predicts a constant color at every pixel for every frame will yield a projection loss of zero and essentially force the background to be captured in foreground layers to minimize the reconstruction loss. To prevent such solutions, we regularize the foreground alphas with an $\\mathcal{L}_1$ and approximate- $\\mathcal{L}_0$ term as in [25].", + "bbox": [ + 75, + 114, + 470, + 251 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s p} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\| \\mathcal {A} _ {a} ^ {i} \\| _ {1} + \\gamma \\| 2 \\cdot \\sigma (5 \\cdot \\mathcal {A} _ {a} ^ {i}) - 1 \\| _ {1} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 117, + 260, + 470, + 301 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\sigma$ is the sigmoid function and $\\gamma$ control the relative weight between the two loss terms.", + "bbox": [ + 75, + 311, + 468, + 339 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Additionally, we regularize the foreground alphas with a smoothness term to discourage them from capturing sharp details in the background which would otherwise arise due to the effects of depth imprecision or splatting in the rasterizer. The smoothness loss $(\\mathcal{L}_s^{\\text{alpha}})$ is computed using an $\\mathcal{L}_1$ loss on the second order gradients of the predicted alphas.", + "bbox": [ + 75, + 340, + 470, + 431 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.4 Disparity Loss", + "text_level": 1, + "bbox": [ + 76, + 449, + 233, + 464 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To distill the available depth information into our model, we apply an $\\mathcal{L}_2$ loss on the predicted background disparity.", + "bbox": [ + 75, + 473, + 468, + 503 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {d i s p}} = \\left\\| M _ {a} ^ {b g} \\odot \\left(D _ {a} - D _ {a} ^ {b g}\\right) \\right\\| _ {2} \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 512, + 468, + 531 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $M_{a}^{bg} = 1 - \\bigcup_{i=1}^{N} M_{a}^{i}$ is the mask indicating the approximate unoccluded background region. Additionally, to encourage smoothness in the depth inpainting we add an $L_{2}$ loss $(\\mathcal{L}_{s}^{disp})$ on the second-order gradients of the predicted disparity similar to $\\mathcal{L}_{s}^{alpha}$ . The sum of the losses $\\mathcal{L}_{sp}, \\mathcal{L}_{s}^{alpha}$ , and $\\mathcal{L}_{s}^{disp}$ forms the regularization loss $\\mathcal{L}_{reg}$ .", + "bbox": [ + 76, + 541, + 468, + 635 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Frame Selection", + "text_level": 1, + "bbox": [ + 76, + 642, + 238, + 656 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The prediction/inpainting of the background RGB and disparity is guided by the projection consistency loss. Intuitively, the source and target frames should be such that regions that are occluded in one are visible in the other whilst having a large overlap in the static regions. To identify such frame pairs we use an approximate frame overlap metric. For two frames $F_{a}$ and $F_{b}$ we estimate the metric by projecting a mesh from $F_{a}$ to $F_{b}$ . The attributes of the mesh vertices is set using $B_{a} = \\cup_{i} M_{a}^{i}$ where for a pixel $(i,j)$ , $B(i,j)$ indicates if the pixel is occupied by a foreground or background as dictated by the input mask. The alpha value for the rasterizer input $A_{a}^{bg}$ are set to all ones. The frame overlap metric is then estimated as", + "bbox": [ + 75, + 665, + 470, + 861 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nO _ {a \\rightarrow b} = \\frac {\\left\\| A _ {a \\rightarrow b} \\right\\|}{H \\cdot W} \\cdot \\frac {\\left\\| B _ {c} - A _ {a \\rightarrow b} \\right\\|}{\\left\\| B _ {c} \\right\\| + \\left\\| A _ {a \\rightarrow b} \\right\\|} \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 869, + 470, + 905 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For each frame in the video, we rank every other based on the frame overlap metric. During training, for a source frame $F_{a}$ , we choose the another frame $F_{b}$ randomly from the top 10 frames that maximize the overlap metric. The target frame $F_{c}$ is chosen to be midway between the two source frames.", + "bbox": [ + 498, + 90, + 890, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Detail Transfer", + "text_level": 1, + "bbox": [ + 500, + 175, + 653, + 189 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The optimized layers $\\mathcal{C}_k^{bg}$ and $\\mathcal{C}_k^i$ may miss high-frequency details present in the input video. Detail may be directly recovered from the input video using the transfer approach of Layered Neural Rendering [24], where the transmittance defined by the alpha maps $\\mathcal{A}_k^i$ controls the amount of detail transferred to each layer. In addition, the background depth map allows reprojection of detail from nearby frames, similar to [34], allowing some detail to be added in regions originally occluded by foreground elements. Figure 10 shows the effect of detail transfer on object removal.", + "bbox": [ + 496, + 198, + 890, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 498, + 363, + 633, + 380 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Training Details", + "text_level": 1, + "bbox": [ + 500, + 387, + 661, + 402 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Data preprocessing. We perform our experiments on videos from the DAVIS [32] dataset. We use CasualSAM [48] to obtain camera poses and initial depth estimates. The input masks are estimated using MaskRCNN [13].", + "bbox": [ + 498, + 411, + 893, + 470 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training Schedule. During the warm-up stage, the model is trained using the disparity $(\\mathcal{L}_{disp})$ and mask losses $(\\mathcal{L}_{mask})$ along with the smoothness regularizers, for initialization purposes. After the warm-up stage, the weights for these loss terms are decayed using a cosine schedule to allow the model to inpaint missing depths and include effects such as shadows in the foreground alpha layers. All other loss terms use a linear schedule starting from 0 during the warm-up stage and remain constant for the remainder of training.", + "bbox": [ + 496, + 474, + 890, + 611 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Layer Decomposition Results", + "text_level": 1, + "bbox": [ + 498, + 619, + 759, + 636 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Figure 4 shows examples of layer decomposition results for various real-world videos from the DAVIS [31] dataset. We show comparisons against Omnimatte [25] and NeuralAtlas [17] and observe clear improvements over these methods. On the Scooter-gray scene, the background predictions for the baselines are severely distorted due to complex camera motion. The incorrect background forces these methods to compensate by reconstructing the background details in the foreground layer. Our method is able to generate an accurate background and predict a much cleaner foreground layer, which includes the shadow. Similarly, on the Snowboard scene, inaccurate homography registration leads to incorrect background prediction by Omnimatte, particularly in the rocks in the upper right corner. Neural-Atlas shows some improvement in the background predictions but loses the rock texture and is unable to capture the shadow in the foreground layer. Our method successfully captures the shadow in the", + "bbox": [ + 496, + 643, + 893, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "634", + "bbox": [ + 485, + 944, + 509, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/20adf8f5fba2a5c5a79dbfe45ff6974ee9a96edd59dbca6968eb972bc5f4ae2e.jpg", + "image_caption": [ + "Snowboard" + ], + "image_footnote": [], + "bbox": [ + 140, + 123, + 313, + 191 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ed86914aadaa1209659d7dd4b48f40d8ffe86d1b98173febdccda68424ddb3c6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 140, + 193, + 312, + 260 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5fb735d45ad0dd6a681dbdaf8eb3a628d9883cab994b3f070b9b3242d4d00660.jpg", + "image_caption": [ + "Scooter-gray" + ], + "image_footnote": [], + "bbox": [ + 140, + 340, + 312, + 410 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e51a4bc01b214dc17b6484248745823f0596b286aba359b1463f10ef87c83b12.jpg", + "image_caption": [ + "(a) Input RGB and Mask" + ], + "image_footnote": [], + "bbox": [ + 140, + 411, + 312, + 477 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d15a9c773cf47cc85510cc60b16175da84cb626c0748e29cf8059483d4d20d40.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 328, + 88, + 500, + 157 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b36869128d160dd787176da9f8b7d946a1a27ee2d714a47f37e2124118f1b705.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 328, + 157, + 500, + 224 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/52b189b6a01711cee7f2d2c822e7dc22bb70fecd5be50c5c68e84491ab66c548.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 328, + 226, + 500, + 295 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f82c21b08afc888f82de1dcb8fd4443713cf76b3ec7242e92ddc0b45bdb0fcb6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 328, + 306, + 500, + 375 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2fe5c2c4b8766f0381d57de92c581e0e734a5065c0e423a883aa767ba11ab929.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 328, + 376, + 500, + 444 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8229a8044af90657caad6b08a4f61242f4ad14e943f43e2d07cd348ff295a9c4.jpg", + "image_caption": [ + "(b) Omnimatte [Lu, et al.]" + ], + "image_footnote": [], + "bbox": [ + 328, + 445, + 500, + 513 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c962a57bb6a79e84ae5eaea389c5cf41bbc81d793aff72b9086d91829c88cd91.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 88, + 674, + 157 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c3a02875d7a22924f919829172e4176765060e2cc1951539ddbf2f19e0f36228.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 157, + 674, + 224 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ccb7fa06775183f0cbbd8efbf29a89ae52b83731639930fe5d98df89fabca7cc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 226, + 674, + 295 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9ff863191447336726b224d0b71082af812d7617c27683a78a25bb740b9980ce.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 306, + 674, + 375 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/39eac2b1182121aa36d86bd7172b957f28d385bc21885266c7851e71256c205a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 376, + 674, + 444 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7669387a3aa52886bf58c2a63ea8300a0d746a350f65d91f60b3090070a1e22f.jpg", + "image_caption": [ + "(c) Neural Atlas [Kasten, et al.]" + ], + "image_footnote": [], + "bbox": [ + 504, + 445, + 674, + 513 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2c7b0948021ce7587f18b6518522ae401751a78740ed6ecfbd0ac78420e48312.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 679, + 88, + 851, + 157 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/54971f03d5c0a546679b9498640ec34cd6a725c035e006f598081049379b1e59.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 679, + 157, + 851, + 224 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/eb1f6d1701db8f3229140ddecbdd80b14701bc8e0c407250bae1df23c0e7f860.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 679, + 226, + 851, + 295 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/dbfd4b77c0ffb62f3aaae0d84b88e2e2a186b80c73b44cdda4b1142d2f970b8a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 679, + 306, + 851, + 375 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/62d37bc52c0eb1812811b2f496956b5f496f8256d57b00c549b2db2bc64ef9d7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 679, + 376, + 851, + 444 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/07fb06dd1c380a14ec2ce1849be95f2d4f66396c74f2b535967e400375d86772.jpg", + "image_caption": [ + "(d) Ours" + ], + "image_footnote": [], + "bbox": [ + 679, + 445, + 851, + 512 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d4f9dd9e564e34541bdd7e976e5b0eb19776fe84813bfe3dd58ad7c11c1200d9.jpg", + "image_caption": [ + "Longboard" + ], + "image_footnote": [], + "bbox": [ + 127, + 575, + 281, + 643 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5021088c22f6f892ecfdeea3b81d75ff5283d058110e7f64541ae792b90a16a3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 127, + 645, + 281, + 712 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b2047b7136227340aadfdd57aa5cb2bda93238cf25f66d55eec97316c667c5c2.jpg", + "image_caption": [ + "(a) Input RGB and Masks", + "Figure 4. Qualitative layer decomposition results on DAVIS. Top: examples of single-layer decomposition on Snowboard and Scooter-gray. For each example, we show the predicted background layer (row 1), predicted object layer RGBA (row 2), and alpha channel visualization (row 3). Bottom: example of multi-layer decomposition on the Longboard scene. On Longboard, we only compare against Omnimatte since Neural-Atlas does not support multi-layer decomposition. Our method produces stronger background inpainting results and more accurately disentangles the foreground objects and their effects from the background." + ], + "image_footnote": [], + "bbox": [ + 127, + 714, + 281, + 782 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/03d8b10fdb4617753fd969c2929f723ca25f25b6ffe5e254e82357a1f13538ba.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 357, + 575, + 493, + 642 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0c80a3ec8f783f02bdd536abdb2413053a35ff174138e70aa7183dd2e1638e95.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 287, + 645, + 424, + 712 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/759393cf1a10eb197c3cd49608101c0ae08aff0a6eb320ed288cd374dbb6df8e.jpg", + "image_caption": [ + "(b) Omnimatte [Lu, et al.]" + ], + "image_footnote": [], + "bbox": [ + 287, + 714, + 424, + 782 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4834d6c94a970b946c22997340df17c436a9bf9401054749d3d688acf379ceae.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 428, + 645, + 565, + 712 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/503bbc695f1eff1d08e50f2b75dabadaf16adcf4f4246a0002c23273f72a6a26.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 428, + 714, + 565, + 782 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2b2fa42908f225562330873f7b882c289e29c6b0b79fad6aa05627b6675dfbf1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 642, + 575, + 779, + 642 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/96681310eda1bc3ac9a1be98d63ac85993bb546fc5700c2fd0bd5da3fccf59f5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 571, + 645, + 709, + 712 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3d668f5836aa2544266770652979a6d6a98276e306d11384c134d3350d9fb2cc.jpg", + "image_caption": [ + "(c) Ours" + ], + "image_footnote": [], + "bbox": [ + 571, + 714, + 709, + 782 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/aff5a4b3e7d3a341bbccedb7a6985f4161ab287b910ba35a0736d05cb5f655a1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 712, + 645, + 851, + 712 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/020eb3e353a897fb9ea21d7eb73c737137b78ffe7e77beb1fe15927dc6340e69.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 712, + 714, + 851, + 782 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "635", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/712d558bba4b27d075dc19895bee7985952b200810d8cb2ea54e04c1a09712d4.jpg", + "image_caption": [ + "Figure 5. Object removal. Top row shows the input frames with the input masks highlighted in blue and green. Subsequent rows show the result of removing the green and blue objects respectively." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 457, + 304 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "foreground layer while accurately predicting the background color. The last row shows an example video (Longboard) with two object layers. For this scene, we compare against Omnimatte as Neural-Atlas does not support multi-layer decomposition. Omnimatte's foreground layer for the cyclist contains significant pollution from background elements. Additionally, despite severe occlusion, our method is able to plausibly inpaint the background as compared to Omnimatte. We provide additional results in the supplementary material.", + "bbox": [ + 75, + 387, + 470, + 523 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Object Removal", + "text_level": 1, + "bbox": [ + 76, + 534, + 238, + 551 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our predicted layers can be used to selectively remove some or all objects from a scene. Figure 5 shows an example of object removal using our layered representation on the Longboard video. Note that the input masks only account for the people and do not include attached objects such as the bicycle or longboard. Our method groups these objects with the correct person, allowing a user to easily remove different people from the scene without the undesirable effect of leaving behind attached objects.", + "bbox": [ + 75, + 558, + 468, + 695 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Depth-based Applications", + "text_level": 1, + "bbox": [ + 76, + 705, + 310, + 722 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "While our method does not directly produce foreground depth, the input depth can be transferred to the foreground layers using the predicted foreground matrices to construct a full Layered Depth Image (LDI) for each frame (see the supplement for details). This representation is useful for a variety of applications, including camera stabilization and synthetic defocus, which we show in the following sections.", + "bbox": [ + 75, + 729, + 468, + 835 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Camera Stabilization. For an input video captured with a shaky camera, we can stabilize the camera path by rendering the scene from a stable camera viewpoint at each time-step. Figure 6 shows results on rendering the Dancing-person [37]", + "bbox": [ + 75, + 839, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ad44e8c8ddf93a86611eede45f5fb9a7a162dff347471ff0c9e5a8fb62e44af3.jpg", + "image_caption": [ + "Figure 6. Camera stabilization. Our LDI representation can be re-rendered from new camera positions to produce stabilized or modified camera paths. Left: input video with camera dolly motion. Right: frames re-rendered from a stationary camera." + ], + "image_footnote": [], + "bbox": [ + 511, + 88, + 879, + 319 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "scene from a stationary camera (for a moving camera path, see supp. material). The explicit mesh representation obtained from our model facilitates editing, as they can be used with commercial video editing tools.", + "bbox": [ + 498, + 411, + 893, + 470 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Synthetic Defocus. Our foreground matte can be used to create a synthetic defocus (bokeh) effect (Fig. 7). To create this effect, we blur the background based on the depth map computed by CasualSAM [48], then composite our foreground RGBA layer on top of the blurred background. The single-layer depth map alone cannot support this effect, as it is insufficiently accurate near the subject's boundary,", + "bbox": [ + 498, + 474, + 893, + 582 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f75fb662a4e73d083449fe643b9580bae88f8775262e53ddcec8f4c332259744.jpg", + "image_caption": [ + "Original" + ], + "image_footnote": [], + "bbox": [ + 511, + 603, + 696, + 691 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6a87757ef4294670538aec93d049591b251d6bb032a63a9a7c8825159249de83.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 697, + 603, + 880, + 691 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e826f33b8cced0668e9bbbfa5515432b85c36cc08276d43225aaf680e95370a1.jpg", + "image_caption": [ + "CasualSAM [Zhang, et al.]" + ], + "image_footnote": [], + "bbox": [ + 513, + 702, + 696, + 789 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/72e9867a7ae975a0a3934631f112fd240385acde390c019957ec129f99807d35.jpg", + "image_caption": [ + "Omnimatte [Lu, et al.]", + "Figure 7. Synthetic defocus. Our foreground matte can enable a synthetic defocus effect (Ours). State-of-the-art single-layer depth maps (CasualSAM [48]) are not accurate near the subject's boundary, causing the goat's head to blur. The foreground matte from Omnimatte [25] contains background pixels, causing parts of the background to remain unblurred." + ], + "image_footnote": [], + "bbox": [ + 697, + 702, + 880, + 790 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "636", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a6a93ec25857b83686f5972fb8e71b5efd201b7d3ed3a548212ad4ab2099e059.jpg", + "image_caption": [ + "(a) Input RGB and Mask" + ], + "image_footnote": [], + "bbox": [ + 89, + 88, + 209, + 207 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a7d5844dfca74fef5014e5c75ac387fc8d04971f3615f38632f1fbed506bca6f.jpg", + "image_caption": [ + "(b) Random Frame Selection" + ], + "image_footnote": [], + "bbox": [ + 212, + 88, + 333, + 207 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ce2dfd7257cb489674c7bd36a6e90b9e817c46deb1dcc9fec04f6deb212ca26d.jpg", + "image_caption": [ + "(c) Our Frame Selection", + "Figure 8. Frame Selection Ablation. Results from random frame selection instead of the proposed heuristic. Unlike the full method, the foreground layer does not accurately capture the shadow." + ], + "image_footnote": [], + "bbox": [ + 336, + 88, + 455, + 207 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "causing important small features to be erroneously blurred (Fig. 7, bottom). The foreground matte from Omnimatte [25] is also insufficiently accurate, containing elements of the background that remain erroneously sharp when composited over the blurred background layer.", + "bbox": [ + 75, + 314, + 467, + 388 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Ablations", + "text_level": 1, + "bbox": [ + 76, + 398, + 187, + 412 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Frame Selection. We ablate our frame selection method 3.4 by training a model using random frame sampling. For a source frame $F_{a}$ , we randomly sample a shift value to obtain the second source $F_{b}$ . The target frame $(F_{c})$ is then chosen mid-way between the source frames. Figure 8 shows results from a model trained using the random frame selection on the hike scene. Compared to the model trained with our frame selection heuristic, random selection leads to most of the shadow being erroneously captured in the background layer. While it is possible to select shift values that can generate desirable results, the proposed frame selection heuristic removes the need to search over this hyperparameter.", + "bbox": [ + 75, + 421, + 468, + 603 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f176aa26e0cb83e883bba536b8648bce9b832ad330f239409c45702030ce3b90.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 614, + 207, + 676 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0d430f0b9ceeefbef2f54bfee616012061a11933a583540b161ce291d13e210d.jpg", + "image_caption": [ + "(a) Input RGB and Mask" + ], + "image_footnote": [], + "bbox": [ + 81, + 679, + 207, + 739 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3e7fc12ddd773e8dfb837b4846e963e9a5dfa8d001da51e3df3af0599a8fb752.jpg", + "image_caption": [ + "(b) Without Projection Consistency" + ], + "image_footnote": [], + "bbox": [ + 210, + 614, + 336, + 739 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/74be33b2385d4863398e60470138e5340da1da0d50d1565a7baf8c726d28dba7.jpg", + "image_caption": [ + "(c) Full Method" + ], + "image_footnote": [], + "bbox": [ + 339, + 614, + 464, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Projection Consistency Loss. We ablate the projection consistency loss in Fig. 9. As shown in column (b), the inpainting of the unobserved region fails without the projection", + "bbox": [ + 76, + 854, + 470, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5590f940c77393f901a6efdd9343c1bdf4acb35989ad34ae56c53b0ad852efd8.jpg", + "image_caption": [ + "Input", + "Figure 10. Detail Transfer Ablation. The background depth map allows reprojecting detail from nearby frames in time to further improve the visual quality of the predicted layers. From left to right: input frame, full result with detail transferred from frames $t$ , $t - 5$ , $t + 5$ , ablated result with detail only from frame $t$ (note blurry region under bicyclist), no background detail transfer." + ], + "image_footnote": [], + "bbox": [ + 501, + 89, + 599, + 186 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/64e0de7ae39545e803182657e07f1f0ece1372544ade00f9c77f314006e04847.jpg", + "image_caption": [ + "Frames t, t-5, t+5" + ], + "image_footnote": [], + "bbox": [ + 599, + 90, + 694, + 186 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4b80b9525ca417e0c73de097fc182652dc4d28dfdaa58544ea171b352e67f08e.jpg", + "image_caption": [ + "Frame t only" + ], + "image_footnote": [], + "bbox": [ + 697, + 90, + 790, + 186 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6477f9713fb018d1f68582956938a074f8e1dc032a213b77be8aa2c8a4aa45f8.jpg", + "image_caption": [ + "No detail transfer" + ], + "image_footnote": [], + "bbox": [ + 792, + 90, + 888, + 186 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "consistency term, as there is no reconstruction signal for that region. Additionally, the foreground layer is impacted due to the shadow being incorrectly placed in the background.", + "bbox": [ + 498, + 320, + 890, + 366 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Discussion and Limitations", + "text_level": 1, + "bbox": [ + 500, + 380, + 750, + 396 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present a new method for layer decomposition that separates a monocular video into a background and several object layers along with their associated effects. Unlike prior decomposition methods that rely on a static 2D background canvas, our method separates the video into layers on a per-frame basis, making it applicable to videos with unconstrained camera motion. To address the challenge of predicting a coherent background, we predict a 3D background and propose a multi-view consistency loss that enforces the background to only contain slowly moving or static details.", + "bbox": [ + 496, + 406, + 893, + 556 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our model can produce per-frame Layered Depth Images that allow for various depth-based editing applications.", + "bbox": [ + 498, + 558, + 890, + 588 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A limitation of our method is shown in Figure 11. Unlike the background, foreground layers are not reprojected to nearby frames and thus foreground objects are not inpainted when occluded. This limitation could be addressed by adding a projection consistency loss similar to Eq. (7) for the foreground layers; however, reprojection of dynamic elements is challenging and requires modelling the sceneflow, which we leave for future work.", + "bbox": [ + 496, + 588, + 892, + 708 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/48ee58420331082e8752bed6b56807f45c591c2adcd1db595d2486157d07f402.jpg", + "image_caption": [ + "Figure 9. Projection Loss Ablation. We show layer results from a model trained without the projection loss. The resulting background contains severe artifacts (column (b), top), and the foreground layer is missing the shadow of the car (column (b), bottom).", + "Input Frames" + ], + "image_footnote": [], + "bbox": [ + 504, + 720, + 630, + 818 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8d3fc30fe8bff3a281c7de34d00a2db1fde03ef36d9710a94cfc1305838a1886.jpg", + "image_caption": [ + "Output Background", + "Figure 11. Limitation: object occlusions. While the background layer is inpainted using nearby frames, our method cannot inpaint the object layers. When an object passes behind the stationary background layer (middle), it is erased from the object layer (right)." + ], + "image_footnote": [], + "bbox": [ + 635, + 720, + 761, + 819 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/eb29d5e3843f5d991717960c25c62dd8e48173befffaf502028ad5a3a47890c7.jpg", + "image_caption": [ + "Output Object Layer" + ], + "image_footnote": [], + "bbox": [ + 763, + 720, + 887, + 819 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "637", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jean-Baptiste Alayrac, João Carreira, and Andrew Zisserman. The visual centrifuge: Model-free layered video representations. In CVPR, 2019. 2", + "[2] Jean-Baptiste Alayrac, Joao Carreira, Relja Arandjelovic, and Andrew Zisserman. Controllable attention for structured layered video decomposition. In ICCV, 2019. 2", + "[3] Omer Bar-Tal, Dolev Ofri-Amar, Rafail Fridman, Yoni Kasten, and Tali Dekel. Text2live: Text-driven layered image and video editing. arXiv preprint arXiv:2204.02491, 2022. 2", + "[4] Gabriel J Brostow and Irfan A Essa. Motion based decomposing of video. In ICCV, 1999. 2", + "[5] Ya-Liang Chang, Zhe Yu Liu, Kuan-Ying Lee, and Winston Hsu. Free-form video inpainting with 3d gated convolution and temporal patchgan. In ICCV, pages 9066-9075, 2019. 2", + "[6] Forrester Cole, Kyle Genova, Avneesh Sud, Daniel Vlasic, and Zhoutong Zhang. Differentiable surface rendering via non-differentiable sampling. In ICCV, pages 6088-6097, 2021. 4", + "[7] Mounira Ebdelli, Olivier Le Meur, and Christine Guillemot. Video inpainting with short-term windows: application to object removal and error concealment. IEEE Transactions on Image Processing, 24(10):3034-3047, 2015. 2", + "[8] Matthieu Fradet, Patrick Pérez, and Philippe Robert. Semi-automatic motion segmentation with motion layer mosaics. In ECCV, 2008. 2", + "[9] Yossi Gandelsman, Assaf Shocher, and Michal Irani. \"Double-DIP\": Unsupervised image decomposition via coupled deep-image-priors. In CVPR, 2019. 2", + "[10] Chen Gao, Ayush Saraf, Jia-Bin Huang, and Johannes Kopf. Flow-edge guided video completion. In ECCV, 2020. 2", + "[11] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video. In ICCV, pages 5712-5721, 2021. 2", + "[12] Miguel Granados, James Tompkin, K Kim, Oliver Grau, Jan Kautz, and Christian Theobalt. How not to be seen—object removal from videos of crowded scenes. Computer Graphics Forum, 31(2pt1):219-228, 2012. 2", + "[13] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, pages 2961-2969, 2017. 5", + "[14] Jia-Bin Huang, Sing Bing Kang, Narendra Ahuja, and Johannes Kopf. Temporally coherent completion of dynamic video. ACM Transactions on Graphics (TOG), 35(6):1-11, 2016. 2", + "[15] Zhang Jiakai, Liu Xinhang, Ye Xinyi, Zhao Fuqiang, Zhang Yanshun, Wu Minye, Zhang Yingliang, Xu Lan, and Yu Jingyi. Editable free-viewpoint video using a layered neural representation. In ACM SIGGRAPH, 2021. 2", + "[16] Nebojsa Joic and Brendan J Frey. Learning flexible sprites in video layers. In Proceedings of the 2001 IEEE Computer Society Conference on Computer Vision and Pattern Recognition. CVPR 2001, volume 1, pages I-I. IEEE, 2001. 2", + "[17] Yoni Kasten, Dolev Ofri, Oliver Wang, and Tali Dekel. Layered neural atlases for consistent video editing. ACM Transactions on Graphics (TOG), 40(6):1-12, 2021. 1, 2, 5" + ], + "bbox": [ + 78, + 116, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[18] Johannes Kopf, Xuejian Rong, and Jia-Bin Huang. Robust consistent video depth estimation. In CVPR, pages 1611-1621, 2021. 2", + "[19] Johannes Kopf, Xuejian Rong, and Jia-Bin Huang. Robust consistent video depth estimation. In CVPR, 2021. 2", + "[20] M. Pawan Kumar, Philip H. S. Torr, and Andrew Zisserman. Learning layered motion segmentations of video. IJCV, 2008. 2", + "[21] Sungho Lee, Seoung Wug Oh, DaeYeun Won, and Seon Joo Kim. Copy-and-paste networks for deep video inpainting. In ICCV, pages 4413-4421, 2019. 2", + "[22] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In CVPR, pages 6498-6508, 2021. 2", + "[23] Chao Liu, Jinwei Gu, Kihwan Kim, Srinivasa G Narasimhan, and Jan Kautz. Neural rgb (r) d sensing: Depth and uncertainty from a video camera. In CVPR, pages 10986-10995, 2019. 2", + "[24] Erika Lu, Forrester Cole, Tali Dekel, Weidi Xie, Andrew Zisserman, David Salesin, William T Freeman, and Michael Rubinstein. Layered neural rendering for retiming people in video. In SIGGRAPH Asia, 2020. 1, 2, 5", + "[25] Erika Lu, Forrester Cole, Tali Dekel, Andrew Zisserman, William T. Freeman, and Michael Rubinstein. Omnimatte: Associating objects and their effects in video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4507-4515, June 2021. 1, 2, 5, 7, 8", + "[26] Xuan Luo, Jia-Bin Huang, Richard Szeliski, Kevin Matzen, and Johannes Kopf. Consistent video depth estimation. ACM Transactions on Graphics (ToG), 39(4):71-1, 2020. 2", + "[27] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In CVPR, pages 7210-7219, 2021. 2", + "[28] Yasuyuki Matsushita, Eyal Ofek, Weina Ge, Xiaou Tang, and Heung-Yeung Shum. Full-frame video stabilization with motion inpainting. IEEE Transactions on pattern analysis and Machine Intelligence, 28(7):1150-1163, 2006. 2", + "[29] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2", + "[30] Seoung Wug Oh, Sungho Lee, Joon-Young Lee, and Seon Joo Kim. Onion-peel networks for deep video completion. In ICCV, pages 4403–4412, 2019. 2", + "[31] F. Perazzi, J. Pont-Tuset, B. McWilliams, L. Van Gool, M. Gross, and A. Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In CVPR, 2016. 5", + "[32] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alexander Sorkine-Hornung, and Luc Van Gool. The 2017 DAVIS challenge on video object segmentation. arXiv:1704.00675, 2017. 5", + "[33] Thomas Porter and Tom Duff. Compositing digital images. SIGGRAPH Comput. Graph., 18(3):253-259, Jan. 1984. 4" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "638", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[34] Xuejian Rong, Jia-Bin Huang, Ayush Saraf, Changil Kim, and Johannes Kopf. Boosting view synthesis with residual transfer. CVPR, 2022. 5", + "[35] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional Networks for Biomedical Image Segmentation. MICCAI, 2015. 3", + "[36] Jonathan Shade, Steven Gortler, Li-wei He, and Richard Szeliski. Layered depth images. In Proceedings of the 25th annual conference on Computer graphics and interactive techniques, pages 231–242, 1998. 2", + "[37] Zhenmei Shi, Fuhao Shi, Wei-Sheng Lai, Chia-Kai Liang, and Yingyu Liang. Deep online fused video stabilization. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1250-1258, 2022. 7", + "[38] Dmitriy Smirnov, Michael Gharbi, Matthew Fisher, Vitor Guizilini, Alexei Efros, and Justin M Solomon. Marionette: Self-supervised sprite learning. Advances in Neural Information Processing Systems, 34, 2021. 2", + "[39] Pratul P. Srinivasan, Richard Tucker, Jonathan T. Barron, Ravi Ramamoorthi, Ren Ng, and Noah Snavely. Pushing the boundaries of view extrapolation with multiplane images. In CVPR, 2019. 2", + "[40] Zachary Teed and Jia Deng. Deepv2d: Video to depth with differentiable structure from motion. arXiv preprint arXiv:1812.04605, 2018. 2", + "[41] Chuan Wang, Haibin Huang, Xiaoguang Han, and Jue Wang. Video inpainting by jointly learning temporal structure and spatial details. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 5232-5239, 2019. 2" + ], + "bbox": [ + 78, + 90, + 470, + 511 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[42] John YA Wang and Edward H Adelson. Representing moving images with layers. IEEE transactions on image processing, 3(5):625-638, 1994. 1, 2", + "[43] Yonatan Wexler, Eli Shechtman, and Michal Irani. Space-time completion of video. PAMI, 2007. 2", + "[44] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond. Computer Graphics Forum, 2022. 2", + "[45] Rui Xu, Xiaoxiao Li, Bolei Zhou, and Chen Change Loy. Deep flow-guided video inpainting. In CVPR, pages 3723-3732, 2019. 2", + "[46] Vickie Ye, Zhengqi Li, Richard Tucker, Angjoo Kanazawa, and Noah Snavely. Deformable sprites for unsupervised video decomposition. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2022. 1, 2", + "[47] Kaidong Zhang, Jingjing Fu, and Dong Liu. Flow-guided transformer for video inpainting. In European Conference on Computer Vision, pages 74-90. Springer, 2022. 2", + "[48] Zhoutong Zhang, Forrester Cole, Zhengqi Li, Michael Rubinstein, Noah Snavely, and William T Freeman. Structure and motion from casual videos. In European Conference on Computer Vision, pages 20-37. Springer, 2022. 2, 5, 7", + "[49] Zhoutong Zhang, Forrester Cole, Richard Tucker, William T Freeman, and Tali Dekel. Consistent depth of moving objects in video. ACM Transactions on Graphics (TOG), 40(4):1-12, 2021. 2", + "[50] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images. In SIGGRAPH, 2018. 2" + ], + "bbox": [ + 501, + 92, + 893, + 511 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "639", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/Omnimatte3D_ Associating Objects and Their Effects in Unconstrained Monocular Video/5700f2ea-1d90-4ebb-a4e2-33ec303e3c7c_model.json b/2023/Omnimatte3D_ Associating Objects and Their Effects in Unconstrained Monocular Video/5700f2ea-1d90-4ebb-a4e2-33ec303e3c7c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c954b454b3330902e95b30542a1c54144be25fd4 --- /dev/null +++ b/2023/Omnimatte3D_ Associating Objects and Their Effects in Unconstrained Monocular Video/5700f2ea-1d90-4ebb-a4e2-33ec303e3c7c_model.json @@ -0,0 +1,3025 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.219, + 0.131, + 0.753, + 0.175 + ], + "angle": 0, + "content": "Omnimatte3D: Associating Objects and their Effects in Unconstrained Monocular Video" + }, + { + "type": "text", + "bbox": [ + 0.187, + 0.203, + 0.361, + 0.22 + ], + "angle": 0, + "content": "Mohammed Suhail1,2" + }, + { + "type": "text", + "bbox": [ + 0.356, + 0.204, + 0.484, + 0.239 + ], + "angle": 0, + "content": "2 Erika Lu4 Leonid Sigal1,2,3" + }, + { + "type": "text", + "bbox": [ + 0.521, + 0.204, + 0.621, + 0.222 + ], + "angle": 0, + "content": "Zhengqi Li" + }, + { + "type": "text", + "bbox": [ + 0.661, + 0.204, + 0.781, + 0.222 + ], + "angle": 0, + "content": "Noah Snavely4" + }, + { + "type": "text", + "bbox": [ + 0.53, + 0.222, + 0.653, + 0.238 + ], + "angle": 0, + "content": "Forrester Cole" + }, + { + "type": "text", + "bbox": [ + 0.191, + 0.241, + 0.384, + 0.256 + ], + "angle": 0, + "content": "1University of British Columbia" + }, + { + "type": "text", + "bbox": [ + 0.398, + 0.241, + 0.539, + 0.256 + ], + "angle": 0, + "content": "\\(^{2}\\)Vector Institute for AI" + }, + { + "type": "text", + "bbox": [ + 0.556, + 0.242, + 0.71, + 0.256 + ], + "angle": 0, + "content": "3Canada CIFAR AI Chair" + }, + { + "type": "text", + "bbox": [ + 0.727, + 0.242, + 0.78, + 0.256 + ], + "angle": 0, + "content": "4Google" + }, + { + "type": "image", + "bbox": [ + 0.108, + 0.261, + 0.25, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.153, + 0.455, + 0.205, + 0.466 + ], + "angle": 0, + "content": "Input RGB" + }, + { + "type": "image", + "bbox": [ + 0.251, + 0.261, + 0.393, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.292, + 0.455, + 0.351, + 0.466 + ], + "angle": 0, + "content": "Input Masks" + }, + { + "type": "image", + "bbox": [ + 0.395, + 0.261, + 0.536, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.443, + 0.455, + 0.503, + 0.466 + ], + "angle": 0, + "content": "Input Depth" + }, + { + "type": "image", + "bbox": [ + 0.538, + 0.354, + 0.571, + 0.371 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.573, + 0.261, + 0.715, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.591, + 0.455, + 0.697, + 0.466 + ], + "angle": 0, + "content": "Layer 1 + Background" + }, + { + "type": "image", + "bbox": [ + 0.717, + 0.261, + 0.859, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.734, + 0.455, + 0.84, + 0.466 + ], + "angle": 0, + "content": "Layer 2 + Background" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.471, + 0.893, + 0.514 + ], + "angle": 0, + "content": "Figure 1. Layer decomposition under strong camera parallax. Given an input video with unconstrained camera motion and approximate object masks and depth (left), our method estimates a layered representation composed of a background layer and object layers containing the subjects of interest and their associated effects (e.g. shadows). Results of combining object layers and background are shown on right." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.526, + 0.314, + 0.541 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.558, + 0.474, + 0.816 + ], + "angle": 0, + "content": "We propose a method to decompose a video into a background and a set of foreground layers, where the background captures stationary elements while the foreground layers capture moving objects along with their associated effects (e.g. shadows and reflections). Our approach is designed for unconstrained monocular videos, with an arbitrary camera and object motion. Prior work that tackles this problem assumes that the video can be mapped onto a fixed 2D canvas, severely limiting the possible space of camera motion. Instead, our method applies recent progress in monocular camera pose and depth estimation to create a full, RGBD video layer for the background, along with a video layer for each foreground object. To solve the underconstrained decomposition problem, we propose a new loss formulation based on multi-view consistency. We test our method on challenging videos with complex camera motion and show significant qualitative improvement over current approaches." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.845, + 0.208, + 0.86 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Decomposing a video into meaningful layers (as show in Fig. 1) is a long-standing and complex problem [42] that has" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.528, + 0.895, + 0.739 + ], + "angle": 0, + "content": "seen a recent surge in progress with the application of deep neural networks [17, 24, 25, 46]. A challenging variant of layer decomposition is the omnimatte [25] task, which aims to separate an input video into a background and multiple foreground layers, each containing an object of interest along with its correlated effects such as shadows and reflections, thus enabling video editing applications such as object removal, background replacement and retiming. The original work on omnimatte [24, 25] introduced a self-supervised approach for decomposing a video by assuming the background can be unwrapped onto a single static 2D background canvas using homography warping. Following work [17, 46] relaxed the homography restriction, but maintained the necessity of unwrapping onto a 2D canvas." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.744, + 0.895, + 0.88 + ], + "angle": 0, + "content": "This 2D modelling, however, limits the applicability of these methods to camera motions that have limited or zero parallax; intuitively, only panning camera motions are allowed. If the camera's center of projection moves a significant distance over the video, 2D methods are unable to learn an accurate background and are forced to place the background detail in a foreground layer (Figure 2). Since camera parallax is quite common, 2D methods are severely restricted in practice." + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.886, + 0.892, + 0.901 + ], + "angle": 0, + "content": "To handle camera parallax, we exploit another recent line" + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "630" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.086, + 0.09, + 0.468, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.263, + 0.472, + 0.364 + ], + "angle": 0, + "content": "Figure 2. Existing methods fail on scenes with parallax. Top row: Omnimatte [25] fails due to inaccurate homography registration, reconstructing the majority of the video in the foreground layer. Middle row: Neural-Atlas [17] incorrectly places trees in the foreground layer and struggles to inpaint the person region. Bottom row: Our method obtains a clean background and captures the person and their shadow in the foreground layer." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.388, + 0.472, + 0.617 + ], + "angle": 0, + "content": "of work on estimating 3D camera position and depth maps from casually-captured video with moving objects [18,26, 48, 49]. Instead of a 2D layer decomposition, we propose to learn a 3D background model that varies per-frame and includes inpainted RGB and depth in the regions occluded by the foreground. Allowing the background to vary per-frame, rather than assuming a global, static canvas, creates an ill-posed decomposition problem, as variation in each input pixel could be explained by a foreground object, background, or both. To resolve this ambiguity, we enforce that the background varies slowly over time with 3D multi-view constraints using dense depth and camera pose estimates [48]. As shown in Fig. 2, our model can successfully separate the layers obtaining a clean background RGB whilst capturing only the person and their shadow in the foreground layer." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.619, + 0.472, + 0.741 + ], + "angle": 0, + "content": "Contributions. Our main contribution is a novel video decomposition method capable of accurately separating an input video with complex object and camera motion (e.g., parallax) into a background and object layers. To address the ill-posed decomposition problem, we design regularization terms based on multi-view consistency and smoothness priors. The resulting model produces clean decompositions of real-world videos where previous methods fail." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.755, + 0.22, + 0.77 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Layered video representations. There is a large body of work on decomposing videos into layers based on appearance and motion cues [4, 8, 16, 20, 36, 42]. Layered video representations have proven useful for a variety of applications, including view synthesis for static scenes [39, 50], reflection removal [1, 2], segmentation [9], game deconstruction [38], and text-based video editing [3]. Recently, Omnimatte [24, 25] aimed to decompose a video into layers" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.244 + ], + "angle": 0, + "content": "that group objects with their correlated effects, allowing the user to perform editing tasks such as object removal and retiming. However, as discussed in Section 1, these works use a constrained background model which limits their use case to panning camera motions only. Subsequent work [17,46] relaxes this assumption using general image warps, instead of homographies, but still cannot handle significant parallax; moreover, their goal is to learn a per-frame mapping to global sprites to enable applying consistent edits (e.g. style transfer) rather than targeting object removal or video stabilization." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.266, + 0.897, + 0.478 + ], + "angle": 0, + "content": "Monocular video depth. Multi-view stereo based methods [23,40] incorporate motion estimation and multi-view reconstruction to predict depth by combining information from multiple reference frames. Hybrid depth methods combine single-view depth estimates from a pretrained network with geometric reasoning to obtain coherent depth predictions. With depth maps from monocular video as initialization, these methods use correspondence to enforce consistency by either assuming static regions [26] or by modeling the scene flow [49]. Recent methods simultaneously estimate depth maps and camera poses [19, 48], exploiting the depth prior to resolve camera motion ambiguities. We rely on CasualSAM [48] to obtain camera poses and initial depth estimates as their method performs well on casually captured video." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.501, + 0.895, + 0.668 + ], + "angle": 0, + "content": "Video completion and object removal. Patch-based methods for video inpainting [7, 12, 14, 28, 43] complete the missing regions by borrowing pixels from other frames, relying on correspondence estimated by homography or optical flow. Learning-based works have explored the use of 3D convolutions [5, 41], attention mechanisms [21, 30, 47] and flow-based methods [10, 45] to achieve more realistic and consistent inpainting. These methods assume that the entire region to inpaint is specified by an input mask, whereas our method jointly learns where to inpaint the background while separating foreground objects and their correlated effects." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.895, + 0.901 + ], + "angle": 0, + "content": "NeRF for dynamic scenes. With the recent success of neural fields [29,44] for view synthesis of static scenes, there has been interest in their adoption for dynamic scenes [11,22,27]. These works learn two NeRF models that decompose a scene into a static and dynamic region. However, these methods only support a single foreground layer and do not account for correlated effects such as shadows. ST-NeRF [15] proposed to learn separate NeRFs for each person to obtain an editable neural representation. Their method, however, requires each frame in the video to be captured from 16 different viewpoints simultaneously. Unlike NeRF methods, our method represents the video with multi-layer mesh geometry where object layers include associated effects, allowing editing as well as fast rerendering of the scene." + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "631" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.089, + 0.881, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.283, + 0.877, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.421, + 0.893, + 0.478 + ], + "angle": 0, + "content": "Figure 3. Model Overview. Given a frame from the video along with an input mask, disparity and known camera pose, we first extract features by stacking the input and passing them though a 2D UNet architecture. These features are input to the background and foreground networks. The background network predicts the inpainted background RGB and disparity. The foreground networks predict an RGB-A image for each foreground layer. For scenes with multiple layers, the RGB-A image for each layer is predicted using independent networks." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.502, + 0.169, + 0.518 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.529, + 0.474, + 0.833 + ], + "angle": 0, + "content": "Given an input video, with known camera poses and initial depth estimates, the goal is to predict a per-frame layer decomposition consisting of a background layer and \\(N\\) foreground layers encompassing objects of interest along with their associated effects. To this end, we propose a new method that predicts a layered separation of an input video without relying on a 2D background canvas, thereby allowing for greater flexibility of camera motion. We show an overview of our model in Figure 3. We optimize our model to output layers that reconstruct a single target video subject to a projection consistency loss. For each frame in the video, we estimate a layered representation consisting of background and foreground object layers where the contents of the object layers are conditioned on their corresponding input mask. Post-optimization, depth for foreground objects can be extracted from the input depth to produce a Layered Depth Image (LDI). Unlike prior works that rely on deep priors for inpainting the depth and color of occluded regions in the background, our method relies on multi-view consistency losses to steer the inpainting process." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.846, + 0.357, + 0.862 + ], + "angle": 0, + "content": "3.1. Layer Decomposition Networks" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.471, + 0.903 + ], + "angle": 0, + "content": "We represent a frame at time \\(a\\) in the video as \\(F_{a} = \\left(I_{a},D_{a},\\{M_{a}^{i}\\}_{i = 1}^{N}\\right)\\), where \\(I_{a}\\) is the RGB frame, \\(D_{a}\\) is the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.503, + 0.892, + 0.595 + ], + "angle": 0, + "content": "disparity, and \\(\\{M_a^i\\}_{i = 1}^N\\) are the input masks for \\(N\\) objects in the video. We denote the camera extrinsics for the frame \\(F_{a}\\) as \\([R_a t_a]\\) where \\(R_{a}\\) is the rotation matrix and \\(t_a\\) is the translation vector. The camera intrinsics are denoted as \\(K_{a} = [f_{x},f_{y},c_{x},c_{y}]\\) where \\(f_{x}\\) and \\(f_{y}\\) are the focal lengths and \\(c_{x},c_{y}\\) is the principal point." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.595, + 0.892, + 0.686 + ], + "angle": 0, + "content": "For each frame \\((F_{a})\\) in the input video, our method first extracts features by passing it to a feature extractor network modeled using a 2D UNet [35] architecture. The input to the feature extractor is constructed by concatenating input masks with the masked frame RGB and disparity along the channel dimension. The masked RGB and disparity are obtained as" + }, + { + "type": "equation", + "bbox": [ + 0.538, + 0.691, + 0.892, + 0.712 + ], + "angle": 0, + "content": "\\[\n\\tilde {Y} _ {a} = \\left[ Y _ {a} \\odot M _ {a} ^ {b g} \\| Y _ {a} \\odot M _ {a} ^ {1} \\| \\dots \\| Y _ {a} \\odot M _ {a} ^ {N} \\right] \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.718, + 0.892, + 0.78 + ], + "angle": 0, + "content": "where \\( Y_{a} \\) is the concatenated RGB-D, \\( M_{a}^{bg} = 1 - \\bigcup_{i=1}^{N} M_{a}^{i} \\) is an approximate mask for the background and \\( \\| \\cdot \\) is the concatenation operation. These features act as input to both the background and foreground object networks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.781, + 0.895, + 0.901 + ], + "angle": 0, + "content": "The background network (shown in green in Fig. 3) consists of a 2D UNet and a convolutional neural network (CNN) that predicts the inpainted background color \\((\\mathcal{C}_a^{bg})\\) and depth \\((\\mathcal{D}_a^{bg})\\) respectively. The foreground networks are separate CNNs that predict RGB \\((\\mathcal{C}_a^i)\\) and alpha \\((\\mathcal{A}_a^i)\\) images for each foreground layer independently. These predicted outputs, along with the input disparity, are combined to form an LDI (see supplement for details)." + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "632" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.168 + ], + "angle": 0, + "content": "The feature extractor and background RGB UNets consist of five down-sampling and up-sampling layers with three residual blocks at every resolution. The background disparity and foreground layer prediction networks are comprised of single convolutional layers." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.175, + 0.307, + 0.192 + ], + "angle": 0, + "content": "3.2. Differentiable Rendering" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.198, + 0.47, + 0.273 + ], + "angle": 0, + "content": "In addition to a reconstruction loss, our method relies on a projection consistency term (described in Section 3.3) that requires differentiable projections and renderings of the predicted background layer to different time steps. We achieve this using a differentiable rasterization framework [6]." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.274, + 0.471, + 0.38 + ], + "angle": 0, + "content": "To project the predicted background layer from a source time step \\(a\\) to target time step \\(b\\), we first construct a mesh \\(\\mathcal{M}_a = (\\mathcal{V}_a, \\mathcal{F}_a)\\), where \\(\\mathcal{V}_a\\) and \\(\\mathcal{F}_a\\) are the mesh vertices and faces. Each vertex in the mesh corresponds to an image pixel. To obtain the vertex coordinate for a pixel, we unproject its homogeneous coordinates \\(x = [i,j,1] \\in \\mathbf{RP}^2\\) using the predicted disparity as" + }, + { + "type": "equation", + "bbox": [ + 0.156, + 0.386, + 0.47, + 0.42 + ], + "angle": 0, + "content": "\\[\nV _ {a} ^ {x} = d _ {i j} \\left(\\frac {\\left(i - c _ {x}\\right)}{f _ {x}}, \\frac {\\left(j - c _ {y}\\right)}{f _ {y}}, 1\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.426, + 0.473, + 0.533 + ], + "angle": 0, + "content": "where \\(d_{ij}\\) is the depth at pixel \\((i,j)\\). The attributes for the vertices are composed of the predicted RGB, an alpha masking consisting of all ones, and the pixel coordinates. The faces for the mesh are constructed by adding an edge between vertices of neighboring pixels (bottom right in Fig. 3). The view-projection matrix to render the predicted background from a target camera is computed as" + }, + { + "type": "equation", + "bbox": [ + 0.151, + 0.538, + 0.47, + 0.572 + ], + "angle": 0, + "content": "\\[\nH _ {a \\rightarrow b} = P _ {a} \\left[\\begin{array}{c c}R _ {a} ^ {T} R _ {b}&R _ {a} ^ {T} \\left(t _ {b} - t _ {a}\\right)\\\\\\mathbf {0}&1\\end{array}\\right] \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.578, + 0.469, + 0.654 + ], + "angle": 0, + "content": "where \\(P_{a}\\) is the perspective matrix that maps camera-space to clip space. The view-projection matrix along with the estimated mesh is passed through a differentiable rasterizer that returns an RGB-A \\((\\mathcal{C}_{a\\to b}^{bg},\\mathcal{A}_{a\\to b}^{bg})\\) image corresponding to the projection of the background to the target time step." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.661, + 0.165, + 0.675 + ], + "angle": 0, + "content": "3.3. Losses" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.684, + 0.332, + 0.698 + ], + "angle": 0, + "content": "The total loss to train our model is:" + }, + { + "type": "equation", + "bbox": [ + 0.086, + 0.707, + 0.47, + 0.724 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\lambda_ {1} \\mathcal {L} _ {\\text {r e c o n}} + \\lambda_ {2} \\mathcal {L} _ {\\text {p r o j}} + \\lambda_ {3} \\mathcal {L} _ {\\text {m a s k}} + \\lambda_ {4} \\mathcal {L} _ {\\text {d i s p}} + \\lambda_ {5} \\mathcal {L} _ {\\text {r e g}} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.731, + 0.469, + 0.762 + ], + "angle": 0, + "content": "We describe each loss term in the following sections. Please see the supplement for values of the scalar weights \\(\\lambda\\)." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.779, + 0.274, + 0.793 + ], + "angle": 0, + "content": "3.3.1 Reconstruction Loss" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.803, + 0.47, + 0.877 + ], + "angle": 0, + "content": "For a source frame \\( F_{a} \\) we reconstruct the image from the predicted background RGB and foreground RGBA layers using back-to-front compositing [33] (bottom left in Fig. 3). During training, we minimize the \\( \\mathcal{L}_2 \\) loss between the original RGB image and the reconstruction." + }, + { + "type": "equation", + "bbox": [ + 0.139, + 0.884, + 0.47, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r e c o n}} = \\left\\| \\left(I _ {a} - f \\left(\\mathcal {C} _ {a} ^ {b g}, \\left\\{\\mathcal {C} _ {a} ^ {i}, \\mathcal {A} _ {a} ^ {1} \\right\\} _ {i = 1} ^ {N}\\right) \\right\\| _ {2} \\right. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.868, + 0.107 + ], + "angle": 0, + "content": "where \\(f\\) is the back-to-front over-compositing function." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.107, + 0.892, + 0.183 + ], + "angle": 0, + "content": "A large number of layer decompositions can reproduce the original frame and minimize the reconstruction loss. To guide the foreground layers to be semantically meaningful, we incorporate an \\(\\mathcal{L}_1\\) loss between the predicted foreground alphas and the input masks." + }, + { + "type": "equation", + "bbox": [ + 0.598, + 0.192, + 0.892, + 0.233 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {m a s k}} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\| M _ {a} ^ {i} - \\mathcal {A} _ {a} ^ {i} \\| _ {1} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.242, + 0.892, + 0.287 + ], + "angle": 0, + "content": "Since we want the foreground alphas to be able to capture effects beyond the input mask, the weight for this loss is decayed as described in Sec. 3.4." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.306, + 0.75, + 0.321 + ], + "angle": 0, + "content": "3.3.2 Projection Consistency Loss" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.329, + 0.894, + 0.525 + ], + "angle": 0, + "content": "The reconstruction loss on the over-composed image does not provide any gradients for the occluded region in the background. Prior works overcome this challenge by learning a background model that maps from a static 2D canvas. However, since our model predicts the background for each frame independently, we require additional regularization to prevent random artifacts or foreground details from appearing in occluded regions. We thus introduce a projection consistency loss on the background prediction. Each batch in our input is composed of three frames from the video. We denote these frames as \\( F_{a} \\), \\( F_{b} \\) and \\( F_{c} \\). The frames \\( F_{a} \\) and \\( F_{b} \\) act as source images for the target frame \\( F_{c} \\). The criteria used for selecting the frames is described in Section 3.4." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.526, + 0.892, + 0.631 + ], + "angle": 0, + "content": "To estimate the projection consistency loss, we project the predicted background RGBs \\((\\mathcal{C}_a^{bg}\\) and \\(\\mathcal{C}_b^{bg})\\) for the two source frames, \\(F_{a}\\) and \\(F_{b}\\), using the method described in Sec. 3.2. We use the predicted disparities \\((\\mathcal{D}_a^{bg}\\) and \\(\\mathcal{D}_b^{bg})\\) to construct the mesh and set the alpha values to one for all pixels. We then apply an \\(\\mathcal{L}_2\\) loss between the projected source backgrounds and the predicted target background." + }, + { + "type": "equation", + "bbox": [ + 0.55, + 0.64, + 0.892, + 0.679 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {p n j} ^ {r g b} = \\frac {1}{2} \\sum_ {k \\in \\{a, b \\}} \\| \\mathcal {A} _ {k \\rightarrow c} ^ {b g} \\odot \\left(\\mathcal {C} _ {k \\rightarrow c} ^ {b g} - \\mathcal {C} _ {c} ^ {b g}\\right) \\| _ {2} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.688, + 0.891, + 0.733 + ], + "angle": 0, + "content": "where \\(\\odot\\) denotes element wise multiplication. We mask the loss using the projected alphas \\((\\mathcal{A}_{k\\to c}^{bg})\\) to prevent noisy gradients from the boundaries of projection." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.733, + 0.891, + 0.763 + ], + "angle": 0, + "content": "Similarly, to encourage consistent inpainting of depth we apply a similar projection loss," + }, + { + "type": "equation", + "bbox": [ + 0.542, + 0.772, + 0.892, + 0.811 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {p r o j} ^ {c o o r d} = \\frac {1}{2} \\sum_ {k \\in \\{a, b \\}} \\| \\mathcal {A} _ {k \\rightarrow c} ^ {b g} \\odot \\left(\\mathcal {X} _ {k \\rightarrow c} ^ {b g} - \\mathcal {X} _ {c} ^ {b g}\\right) \\| _ {2} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.82, + 0.894, + 0.904 + ], + "angle": 0, + "content": "where \\(\\mathcal{X}_c\\) are the world coordinates for pixels in the frame \\(F_{c}\\) obtained using the predicted background disparity \\(\\mathcal{D}_c^{bg}\\). \\(\\mathcal{X}_{k\\to c}^{bg}\\) are the projections of the world coordinate from frame \\(F_{k}\\) to \\(F_{c}\\) and \\(\\mathcal{A}_{k\\to c}^{bg}\\) are the projected alphas. The projection loss \\(\\mathcal{L}_{proj}\\) is the sum of \\(\\mathcal{L}_{proj}^{rgb}\\) and \\(\\mathcal{L}_{proj}^{coord}\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "633" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.077, + 0.092, + 0.282, + 0.106 + ], + "angle": 0, + "content": "3.3.3 Alpha Regularization" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.115, + 0.471, + 0.252 + ], + "angle": 0, + "content": "While the projection loss encourages the background predictions to be consistent across frames, it can also lead the model toward undesirable or degenerate solutions. For example, a model that predicts a constant color at every pixel for every frame will yield a projection loss of zero and essentially force the background to be captured in foreground layers to minimize the reconstruction loss. To prevent such solutions, we regularize the foreground alphas with an \\(\\mathcal{L}_1\\) and approximate-\\(\\mathcal{L}_0\\) term as in [25]." + }, + { + "type": "equation", + "bbox": [ + 0.118, + 0.261, + 0.471, + 0.302 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s p} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\| \\mathcal {A} _ {a} ^ {i} \\| _ {1} + \\gamma \\| 2 \\cdot \\sigma (5 \\cdot \\mathcal {A} _ {a} ^ {i}) - 1 \\| _ {1} \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.312, + 0.469, + 0.34 + ], + "angle": 0, + "content": "where \\(\\sigma\\) is the sigmoid function and \\(\\gamma\\) control the relative weight between the two loss terms." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.342, + 0.471, + 0.432 + ], + "angle": 0, + "content": "Additionally, we regularize the foreground alphas with a smoothness term to discourage them from capturing sharp details in the background which would otherwise arise due to the effects of depth imprecision or splatting in the rasterizer. The smoothness loss \\((\\mathcal{L}_s^{\\text{alpha}})\\) is computed using an \\(\\mathcal{L}_1\\) loss on the second order gradients of the predicted alphas." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.45, + 0.234, + 0.465 + ], + "angle": 0, + "content": "3.3.4 Disparity Loss" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.474, + 0.469, + 0.504 + ], + "angle": 0, + "content": "To distill the available depth information into our model, we apply an \\(\\mathcal{L}_2\\) loss on the predicted background disparity." + }, + { + "type": "equation", + "bbox": [ + 0.167, + 0.513, + 0.47, + 0.532 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {d i s p}} = \\left\\| M _ {a} ^ {b g} \\odot \\left(D _ {a} - D _ {a} ^ {b g}\\right) \\right\\| _ {2} \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.542, + 0.47, + 0.636 + ], + "angle": 0, + "content": "where \\( M_{a}^{bg} = 1 - \\bigcup_{i=1}^{N} M_{a}^{i} \\) is the mask indicating the approximate unoccluded background region. Additionally, to encourage smoothness in the depth inpainting we add an \\( L_{2} \\) loss \\( (\\mathcal{L}_{s}^{disp}) \\) on the second-order gradients of the predicted disparity similar to \\( \\mathcal{L}_{s}^{alpha} \\). The sum of the losses \\( \\mathcal{L}_{sp}, \\mathcal{L}_{s}^{alpha} \\), and \\( \\mathcal{L}_{s}^{disp} \\) forms the regularization loss \\( \\mathcal{L}_{reg} \\)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.643, + 0.239, + 0.657 + ], + "angle": 0, + "content": "3.4. Frame Selection" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.666, + 0.471, + 0.862 + ], + "angle": 0, + "content": "The prediction/inpainting of the background RGB and disparity is guided by the projection consistency loss. Intuitively, the source and target frames should be such that regions that are occluded in one are visible in the other whilst having a large overlap in the static regions. To identify such frame pairs we use an approximate frame overlap metric. For two frames \\( F_{a} \\) and \\( F_{b} \\) we estimate the metric by projecting a mesh from \\( F_{a} \\) to \\( F_{b} \\). The attributes of the mesh vertices is set using \\( B_{a} = \\cup_{i} M_{a}^{i} \\) where for a pixel \\( (i,j) \\), \\( B(i,j) \\) indicates if the pixel is occupied by a foreground or background as dictated by the input mask. The alpha value for the rasterizer input \\( A_{a}^{bg} \\) are set to all ones. The frame overlap metric is then estimated as" + }, + { + "type": "equation", + "bbox": [ + 0.146, + 0.871, + 0.471, + 0.906 + ], + "angle": 0, + "content": "\\[\nO _ {a \\rightarrow b} = \\frac {\\left\\| A _ {a \\rightarrow b} \\right\\|}{H \\cdot W} \\cdot \\frac {\\left\\| B _ {c} - A _ {a \\rightarrow b} \\right\\|}{\\left\\| B _ {c} \\right\\| + \\left\\| A _ {a \\rightarrow b} \\right\\|} \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "For each frame in the video, we rank every other based on the frame overlap metric. During training, for a source frame \\( F_{a} \\), we choose the another frame \\( F_{b} \\) randomly from the top 10 frames that maximize the overlap metric. The target frame \\( F_{c} \\) is chosen to be midway between the two source frames." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.176, + 0.655, + 0.19 + ], + "angle": 0, + "content": "3.5. Detail Transfer" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.199, + 0.892, + 0.351 + ], + "angle": 0, + "content": "The optimized layers \\(\\mathcal{C}_k^{bg}\\) and \\(\\mathcal{C}_k^i\\) may miss high-frequency details present in the input video. Detail may be directly recovered from the input video using the transfer approach of Layered Neural Rendering [24], where the transmittance defined by the alpha maps \\(\\mathcal{A}_k^i\\) controls the amount of detail transferred to each layer. In addition, the background depth map allows reprojection of detail from nearby frames, similar to [34], allowing some detail to be added in regions originally occluded by foreground elements. Figure 10 shows the effect of detail transfer on object removal." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.364, + 0.634, + 0.381 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.388, + 0.662, + 0.404 + ], + "angle": 0, + "content": "4.1. Training Details" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.412, + 0.894, + 0.472 + ], + "angle": 0, + "content": "Data preprocessing. We perform our experiments on videos from the DAVIS [32] dataset. We use CasualSAM [48] to obtain camera poses and initial depth estimates. The input masks are estimated using MaskRCNN [13]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.476, + 0.892, + 0.612 + ], + "angle": 0, + "content": "Training Schedule. During the warm-up stage, the model is trained using the disparity \\((\\mathcal{L}_{disp})\\) and mask losses \\((\\mathcal{L}_{mask})\\) along with the smoothness regularizers, for initialization purposes. After the warm-up stage, the weights for these loss terms are decayed using a cosine schedule to allow the model to inpaint missing depths and include effects such as shadows in the foreground alpha layers. All other loss terms use a linear schedule starting from 0 during the warm-up stage and remain constant for the remainder of training." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.621, + 0.761, + 0.637 + ], + "angle": 0, + "content": "4.2. Layer Decomposition Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.644, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Figure 4 shows examples of layer decomposition results for various real-world videos from the DAVIS [31] dataset. We show comparisons against Omnimatte [25] and NeuralAtlas [17] and observe clear improvements over these methods. On the Scooter-gray scene, the background predictions for the baselines are severely distorted due to complex camera motion. The incorrect background forces these methods to compensate by reconstructing the background details in the foreground layer. Our method is able to generate an accurate background and predict a much cleaner foreground layer, which includes the shadow. Similarly, on the Snowboard scene, inaccurate homography registration leads to incorrect background prediction by Omnimatte, particularly in the rocks in the upper right corner. Neural-Atlas shows some improvement in the background predictions but loses the rock texture and is unable to capture the shadow in the foreground layer. Our method successfully captures the shadow in the" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.945, + 0.511, + 0.957 + ], + "angle": 0, + "content": "634" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.186, + 0.104, + 0.271, + 0.117 + ], + "angle": 0, + "content": "Snowboard" + }, + { + "type": "image", + "bbox": [ + 0.142, + 0.124, + 0.314, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.142, + 0.194, + 0.313, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.18, + 0.323, + 0.276, + 0.338 + ], + "angle": 0, + "content": "Scooter-gray" + }, + { + "type": "image", + "bbox": [ + 0.142, + 0.342, + 0.313, + 0.411 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.142, + 0.412, + 0.313, + 0.478 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.18, + 0.514, + 0.269, + 0.54 + ], + "angle": 0, + "content": "(a) Input RGB and Mask" + }, + { + "type": "image", + "bbox": [ + 0.329, + 0.089, + 0.5, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.33, + 0.159, + 0.5, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.33, + 0.227, + 0.5, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.329, + 0.307, + 0.5, + 0.376 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.329, + 0.377, + 0.5, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.329, + 0.446, + 0.5, + 0.514 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.366, + 0.516, + 0.463, + 0.542 + ], + "angle": 0, + "content": "(b) Omnimatte [Lu, et al.]" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.089, + 0.676, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.159, + 0.676, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.227, + 0.676, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.308, + 0.676, + 0.376 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.377, + 0.676, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.446, + 0.676, + 0.514 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.541, + 0.515, + 0.641, + 0.541 + ], + "angle": 0, + "content": "(c) Neural Atlas [Kasten, et al.]" + }, + { + "type": "image", + "bbox": [ + 0.681, + 0.089, + 0.852, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.681, + 0.159, + 0.852, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.681, + 0.227, + 0.852, + 0.296 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.681, + 0.307, + 0.852, + 0.376 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.681, + 0.377, + 0.852, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.681, + 0.446, + 0.852, + 0.513 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.739, + 0.515, + 0.793, + 0.527 + ], + "angle": 0, + "content": "(d) Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.173, + 0.56, + 0.255, + 0.574 + ], + "angle": 0, + "content": "Longboard" + }, + { + "type": "image", + "bbox": [ + 0.129, + 0.577, + 0.282, + 0.644 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.129, + 0.646, + 0.282, + 0.713 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.129, + 0.715, + 0.282, + 0.783 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.167, + 0.785, + 0.255, + 0.811 + ], + "angle": 0, + "content": "(a) Input RGB and Masks" + }, + { + "type": "image", + "bbox": [ + 0.359, + 0.577, + 0.495, + 0.643 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.289, + 0.646, + 0.425, + 0.713 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.289, + 0.715, + 0.425, + 0.783 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.347, + 0.785, + 0.51, + 0.798 + ], + "angle": 0, + "content": "(b) Omnimatte [Lu, et al.]" + }, + { + "type": "image", + "bbox": [ + 0.429, + 0.646, + 0.566, + 0.713 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.429, + 0.715, + 0.566, + 0.783 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.643, + 0.577, + 0.78, + 0.643 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.573, + 0.646, + 0.71, + 0.713 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.573, + 0.715, + 0.71, + 0.783 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.686, + 0.785, + 0.737, + 0.798 + ], + "angle": 0, + "content": "(c) Ours" + }, + { + "type": "image", + "bbox": [ + 0.714, + 0.646, + 0.852, + 0.713 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.714, + 0.715, + 0.852, + 0.783 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.825, + 0.895, + 0.896 + ], + "angle": 0, + "content": "Figure 4. Qualitative layer decomposition results on DAVIS. Top: examples of single-layer decomposition on Snowboard and Scooter-gray. For each example, we show the predicted background layer (row 1), predicted object layer RGBA (row 2), and alpha channel visualization (row 3). Bottom: example of multi-layer decomposition on the Longboard scene. On Longboard, we only compare against Omnimatte since Neural-Atlas does not support multi-layer decomposition. Our method produces stronger background inpainting results and more accurately disentangles the foreground objects and their effects from the background." + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "635" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.458, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.318, + 0.471, + 0.361 + ], + "angle": 0, + "content": "Figure 5. Object removal. Top row shows the input frames with the input masks highlighted in blue and green. Subsequent rows show the result of removing the green and blue objects respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.388, + 0.472, + 0.525 + ], + "angle": 0, + "content": "foreground layer while accurately predicting the background color. The last row shows an example video (Longboard) with two object layers. For this scene, we compare against Omnimatte as Neural-Atlas does not support multi-layer decomposition. Omnimatte's foreground layer for the cyclist contains significant pollution from background elements. Additionally, despite severe occlusion, our method is able to plausibly inpaint the background as compared to Omnimatte. We provide additional results in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.535, + 0.24, + 0.552 + ], + "angle": 0, + "content": "4.3. Object Removal" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.559, + 0.47, + 0.696 + ], + "angle": 0, + "content": "Our predicted layers can be used to selectively remove some or all objects from a scene. Figure 5 shows an example of object removal using our layered representation on the Longboard video. Note that the input masks only account for the people and do not include attached objects such as the bicycle or longboard. Our method groups these objects with the correct person, allowing a user to easily remove different people from the scene without the undesirable effect of leaving behind attached objects." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.707, + 0.312, + 0.723 + ], + "angle": 0, + "content": "4.4. Depth-based Applications" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.73, + 0.47, + 0.837 + ], + "angle": 0, + "content": "While our method does not directly produce foreground depth, the input depth can be transferred to the foreground layers using the predicted foreground matrices to construct a full Layered Depth Image (LDI) for each frame (see the supplement for details). This representation is useful for a variety of applications, including camera stabilization and synthetic defocus, which we show in the following sections." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Camera Stabilization. For an input video captured with a shaky camera, we can stabilize the camera path by rendering the scene from a stable camera viewpoint at each time-step. Figure 6 shows results on rendering the Dancing-person [37]" + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.089, + 0.88, + 0.32 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.329, + 0.895, + 0.385 + ], + "angle": 0, + "content": "Figure 6. Camera stabilization. Our LDI representation can be re-rendered from new camera positions to produce stabilized or modified camera paths. Left: input video with camera dolly motion. Right: frames re-rendered from a stationary camera." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.412, + 0.895, + 0.472 + ], + "angle": 0, + "content": "scene from a stationary camera (for a moving camera path, see supp. material). The explicit mesh representation obtained from our model facilitates editing, as they can be used with commercial video editing tools." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.476, + 0.895, + 0.583 + ], + "angle": 0, + "content": "Synthetic Defocus. Our foreground matte can be used to create a synthetic defocus (bokeh) effect (Fig. 7). To create this effect, we blur the background based on the depth map computed by CasualSAM [48], then composite our foreground RGBA layer on top of the blurred background. The single-layer depth map alone cannot support this effect, as it is insufficiently accurate near the subject's boundary," + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.604, + 0.697, + 0.692 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.58, + 0.692, + 0.626, + 0.703 + ], + "angle": 0, + "content": "Original" + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.604, + 0.881, + 0.693 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.777, + 0.693, + 0.806, + 0.703 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.703, + 0.697, + 0.79 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.54, + 0.791, + 0.683, + 0.803 + ], + "angle": 0, + "content": "CasualSAM [Zhang, et al.]" + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.703, + 0.881, + 0.791 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.725, + 0.791, + 0.848, + 0.803 + ], + "angle": 0, + "content": "Omnimatte [Lu, et al.]" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.814, + 0.895, + 0.897 + ], + "angle": 0, + "content": "Figure 7. Synthetic defocus. Our foreground matte can enable a synthetic defocus effect (Ours). State-of-the-art single-layer depth maps (CasualSAM [48]) are not accurate near the subject's boundary, causing the goat's head to blur. The foreground matte from Omnimatte [25] contains background pixels, causing parts of the background to remain unblurred." + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "636" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.089, + 0.21, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.11, + 0.21, + 0.188, + 0.233 + ], + "angle": 0, + "content": "(a) Input RGB and Mask" + }, + { + "type": "image", + "bbox": [ + 0.214, + 0.089, + 0.334, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.221, + 0.21, + 0.327, + 0.231 + ], + "angle": 0, + "content": "(b) Random Frame Selection" + }, + { + "type": "image", + "bbox": [ + 0.338, + 0.089, + 0.457, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.358, + 0.21, + 0.437, + 0.231 + ], + "angle": 0, + "content": "(c) Our Frame Selection" + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.246, + 0.471, + 0.288 + ], + "angle": 0, + "content": "Figure 8. Frame Selection Ablation. Results from random frame selection instead of the proposed heuristic. Unlike the full method, the foreground layer does not accurately capture the shadow." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.315, + 0.468, + 0.39 + ], + "angle": 0, + "content": "causing important small features to be erroneously blurred (Fig. 7, bottom). The foreground matte from Omnimatte [25] is also insufficiently accurate, containing elements of the background that remain erroneously sharp when composited over the blurred background layer." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.4, + 0.188, + 0.414 + ], + "angle": 0, + "content": "4.5. Ablations" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.422, + 0.47, + 0.604 + ], + "angle": 0, + "content": "Frame Selection. We ablate our frame selection method 3.4 by training a model using random frame sampling. For a source frame \\( F_{a} \\), we randomly sample a shift value to obtain the second source \\( F_{b} \\). The target frame \\( (F_{c}) \\) is then chosen mid-way between the source frames. Figure 8 shows results from a model trained using the random frame selection on the hike scene. Compared to the model trained with our frame selection heuristic, random selection leads to most of the shadow being erroneously captured in the background layer. While it is possible to select shift values that can generate desirable results, the proposed frame selection heuristic removes the need to search over this hyperparameter." + }, + { + "type": "image", + "bbox": [ + 0.083, + 0.616, + 0.209, + 0.678 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.083, + 0.68, + 0.208, + 0.741 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.103, + 0.742, + 0.184, + 0.766 + ], + "angle": 0, + "content": "(a) Input RGB and Mask" + }, + { + "type": "image", + "bbox": [ + 0.212, + 0.616, + 0.338, + 0.74 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.207, + 0.742, + 0.343, + 0.767 + ], + "angle": 0, + "content": "(b) Without Projection Consistency" + }, + { + "type": "image", + "bbox": [ + 0.34, + 0.616, + 0.465, + 0.74 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.357, + 0.742, + 0.448, + 0.753 + ], + "angle": 0, + "content": "(c) Full Method" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.78, + 0.47, + 0.836 + ], + "angle": 0, + "content": "Figure 9. Projection Loss Ablation. We show layer results from a model trained without the projection loss. The resulting background contains severe artifacts (column (b), top), and the foreground layer is missing the shadow of the car (column (b), bottom)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.856, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Projection Consistency Loss. We ablate the projection consistency loss in Fig. 9. As shown in column (b), the inpainting of the unobserved region fails without the projection" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.09, + 0.6, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.538, + 0.189, + 0.565, + 0.198 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.601, + 0.091, + 0.696, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.609, + 0.189, + 0.69, + 0.198 + ], + "angle": 0, + "content": "Frames t, t-5, t+5" + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.091, + 0.792, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.713, + 0.189, + 0.772, + 0.198 + ], + "angle": 0, + "content": "Frame t only" + }, + { + "type": "image", + "bbox": [ + 0.794, + 0.091, + 0.89, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.801, + 0.189, + 0.882, + 0.198 + ], + "angle": 0, + "content": "No detail transfer" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.21, + 0.892, + 0.294 + ], + "angle": 0, + "content": "Figure 10. Detail Transfer Ablation. The background depth map allows reprojecting detail from nearby frames in time to further improve the visual quality of the predicted layers. From left to right: input frame, full result with detail transferred from frames \\( t \\), \\( t - 5 \\), \\( t + 5 \\), ablated result with detail only from frame \\( t \\) (note blurry region under bicyclist), no background detail transfer." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.321, + 0.892, + 0.367 + ], + "angle": 0, + "content": "consistency term, as there is no reconstruction signal for that region. Additionally, the foreground layer is impacted due to the shadow being incorrectly placed in the background." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.381, + 0.75, + 0.397 + ], + "angle": 0, + "content": "5. Discussion and Limitations" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.407, + 0.894, + 0.558 + ], + "angle": 0, + "content": "We present a new method for layer decomposition that separates a monocular video into a background and several object layers along with their associated effects. Unlike prior decomposition methods that rely on a static 2D background canvas, our method separates the video into layers on a per-frame basis, making it applicable to videos with unconstrained camera motion. To address the challenge of predicting a coherent background, we predict a 3D background and propose a multi-view consistency loss that enforces the background to only contain slowly moving or static details." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.559, + 0.892, + 0.589 + ], + "angle": 0, + "content": "Our model can produce per-frame Layered Depth Images that allow for various depth-based editing applications." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.589, + 0.893, + 0.709 + ], + "angle": 0, + "content": "A limitation of our method is shown in Figure 11. Unlike the background, foreground layers are not reprojected to nearby frames and thus foreground objects are not inpainted when occluded. This limitation could be addressed by adding a projection consistency loss similar to Eq. (7) for the foreground layers; however, reprojection of dynamic elements is challenging and requires modelling the sceneflow, which we leave for future work." + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.721, + 0.632, + 0.819 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.539, + 0.82, + 0.599, + 0.828 + ], + "angle": 0, + "content": "Input Frames" + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.721, + 0.763, + 0.82 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.649, + 0.82, + 0.739, + 0.828 + ], + "angle": 0, + "content": "Output Background" + }, + { + "type": "image", + "bbox": [ + 0.764, + 0.722, + 0.888, + 0.82 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.779, + 0.82, + 0.872, + 0.828 + ], + "angle": 0, + "content": "Output Object Layer" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.841, + 0.892, + 0.898 + ], + "angle": 0, + "content": "Figure 11. Limitation: object occlusions. While the background layer is inpainted using nearby frames, our method cannot inpaint the object layers. When an object passes behind the stationary background layer (middle), it is erased from the object layer (right)." + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "637" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.117, + 0.472, + 0.157 + ], + "angle": 0, + "content": "[1] Jean-Baptiste Alayrac, João Carreira, and Andrew Zisserman. The visual centrifuge: Model-free layered video representations. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.16, + 0.472, + 0.202 + ], + "angle": 0, + "content": "[2] Jean-Baptiste Alayrac, Joao Carreira, Relja Arandjelovic, and Andrew Zisserman. Controllable attention for structured layered video decomposition. In ICCV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.204, + 0.472, + 0.247 + ], + "angle": 0, + "content": "[3] Omer Bar-Tal, Dolev Ofri-Amar, Rafail Fridman, Yoni Kasten, and Tali Dekel. Text2live: Text-driven layered image and video editing. arXiv preprint arXiv:2204.02491, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.249, + 0.472, + 0.275 + ], + "angle": 0, + "content": "[4] Gabriel J Brostow and Irfan A Essa. Motion based decomposing of video. In ICCV, 1999. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.278, + 0.472, + 0.32 + ], + "angle": 0, + "content": "[5] Ya-Liang Chang, Zhe Yu Liu, Kuan-Ying Lee, and Winston Hsu. Free-form video inpainting with 3d gated convolution and temporal patchgan. In ICCV, pages 9066-9075, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.322, + 0.472, + 0.376 + ], + "angle": 0, + "content": "[6] Forrester Cole, Kyle Genova, Avneesh Sud, Daniel Vlasic, and Zhoutong Zhang. Differentiable surface rendering via non-differentiable sampling. In ICCV, pages 6088-6097, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.379, + 0.472, + 0.435 + ], + "angle": 0, + "content": "[7] Mounira Ebdelli, Olivier Le Meur, and Christine Guillemot. Video inpainting with short-term windows: application to object removal and error concealment. IEEE Transactions on Image Processing, 24(10):3034-3047, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.437, + 0.472, + 0.478 + ], + "angle": 0, + "content": "[8] Matthieu Fradet, Patrick Pérez, and Philippe Robert. Semi-automatic motion segmentation with motion layer mosaics. In ECCV, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.481, + 0.472, + 0.522 + ], + "angle": 0, + "content": "[9] Yossi Gandelsman, Assaf Shocher, and Michal Irani. \"Double-DIP\": Unsupervised image decomposition via coupled deep-image-priors. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.525, + 0.472, + 0.552 + ], + "angle": 0, + "content": "[10] Chen Gao, Ayush Saraf, Jia-Bin Huang, and Johannes Kopf. Flow-edge guided video completion. In ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.555, + 0.472, + 0.596 + ], + "angle": 0, + "content": "[11] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video. In ICCV, pages 5712-5721, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.599, + 0.472, + 0.654 + ], + "angle": 0, + "content": "[12] Miguel Granados, James Tompkin, K Kim, Oliver Grau, Jan Kautz, and Christian Theobalt. How not to be seen—object removal from videos of crowded scenes. Computer Graphics Forum, 31(2pt1):219-228, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.656, + 0.472, + 0.684 + ], + "angle": 0, + "content": "[13] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, pages 2961-2969, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.686, + 0.472, + 0.74 + ], + "angle": 0, + "content": "[14] Jia-Bin Huang, Sing Bing Kang, Narendra Ahuja, and Johannes Kopf. Temporally coherent completion of dynamic video. ACM Transactions on Graphics (TOG), 35(6):1-11, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.744, + 0.472, + 0.799 + ], + "angle": 0, + "content": "[15] Zhang Jiakai, Liu Xinhang, Ye Xinyi, Zhao Fuqiang, Zhang Yanshun, Wu Minye, Zhang Yingliang, Xu Lan, and Yu Jingyi. Editable free-viewpoint video using a layered neural representation. In ACM SIGGRAPH, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.472, + 0.857 + ], + "angle": 0, + "content": "[16] Nebojsa Joic and Brendan J Frey. Learning flexible sprites in video layers. In Proceedings of the 2001 IEEE Computer Society Conference on Computer Vision and Pattern Recognition. CVPR 2001, volume 1, pages I-I. IEEE, 2001. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.472, + 0.901 + ], + "angle": 0, + "content": "[17] Yoni Kasten, Dolev Ofri, Oliver Wang, and Tali Dekel. Layered neural atlases for consistent video editing. ACM Transactions on Graphics (TOG), 40(6):1-12, 2021. 1, 2, 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.117, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.134 + ], + "angle": 0, + "content": "[18] Johannes Kopf, Xuejian Rong, and Jia-Bin Huang. Robust consistent video depth estimation. In CVPR, pages 1611-1621, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[19] Johannes Kopf, Xuejian Rong, and Jia-Bin Huang. Robust consistent video depth estimation. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.894, + 0.203 + ], + "angle": 0, + "content": "[20] M. Pawan Kumar, Philip H. S. Torr, and Andrew Zisserman. Learning layered motion segmentations of video. IJCV, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.205, + 0.892, + 0.245 + ], + "angle": 0, + "content": "[21] Sungho Lee, Seoung Wug Oh, DaeYeun Won, and Seon Joo Kim. Copy-and-paste networks for deep video inpainting. In ICCV, pages 4413-4421, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.247, + 0.894, + 0.288 + ], + "angle": 0, + "content": "[22] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In CVPR, pages 6498-6508, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.289, + 0.894, + 0.342 + ], + "angle": 0, + "content": "[23] Chao Liu, Jinwei Gu, Kihwan Kim, Srinivasa G Narasimhan, and Jan Kautz. Neural rgb (r) d sensing: Depth and uncertainty from a video camera. In CVPR, pages 10986-10995, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.344, + 0.892, + 0.399 + ], + "angle": 0, + "content": "[24] Erika Lu, Forrester Cole, Tali Dekel, Weidi Xie, Andrew Zisserman, David Salesin, William T Freeman, and Michael Rubinstein. Layered neural rendering for retiming people in video. In SIGGRAPH Asia, 2020. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.4, + 0.894, + 0.481 + ], + "angle": 0, + "content": "[25] Erika Lu, Forrester Cole, Tali Dekel, Andrew Zisserman, William T. Freeman, and Michael Rubinstein. Omnimatte: Associating objects and their effects in video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4507-4515, June 2021. 1, 2, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.484, + 0.892, + 0.524 + ], + "angle": 0, + "content": "[26] Xuan Luo, Jia-Bin Huang, Richard Szeliski, Kevin Matzen, and Johannes Kopf. Consistent video depth estimation. ACM Transactions on Graphics (ToG), 39(4):71-1, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.526, + 0.894, + 0.593 + ], + "angle": 0, + "content": "[27] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In CVPR, pages 7210-7219, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.595, + 0.892, + 0.65 + ], + "angle": 0, + "content": "[28] Yasuyuki Matsushita, Eyal Ofek, Weina Ge, Xiaou Tang, and Heung-Yeung Shum. Full-frame video stabilization with motion inpainting. IEEE Transactions on pattern analysis and Machine Intelligence, 28(7):1150-1163, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.651, + 0.894, + 0.718 + ], + "angle": 0, + "content": "[29] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.72, + 0.892, + 0.761 + ], + "angle": 0, + "content": "[30] Seoung Wug Oh, Sungho Lee, Joon-Young Lee, and Seon Joo Kim. Onion-peel networks for deep video completion. In ICCV, pages 4403–4412, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.762, + 0.892, + 0.815 + ], + "angle": 0, + "content": "[31] F. Perazzi, J. Pont-Tuset, B. McWilliams, L. Van Gool, M. Gross, and A. Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In CVPR, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.818, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[32] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alexander Sorkine-Hornung, and Luc Van Gool. The 2017 DAVIS challenge on video object segmentation. arXiv:1704.00675, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[33] Thomas Porter and Tom Duff. Compositing digital images. SIGGRAPH Comput. Graph., 18(3):253-259, Jan. 1984. 4" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "638" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.134 + ], + "angle": 0, + "content": "[34] Xuejian Rong, Jia-Bin Huang, Ayush Saraf, Changil Kim, and Johannes Kopf. Boosting view synthesis with residual transfer. CVPR, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.137, + 0.472, + 0.178 + ], + "angle": 0, + "content": "[35] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional Networks for Biomedical Image Segmentation. MICCAI, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.181, + 0.47, + 0.236 + ], + "angle": 0, + "content": "[36] Jonathan Shade, Steven Gortler, Li-wei He, and Richard Szeliski. Layered depth images. In Proceedings of the 25th annual conference on Computer graphics and interactive techniques, pages 231–242, 1998. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.239, + 0.471, + 0.294 + ], + "angle": 0, + "content": "[37] Zhenmei Shi, Fuhao Shi, Wei-Sheng Lai, Chia-Kai Liang, and Yingyu Liang. Deep online fused video stabilization. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1250-1258, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.297, + 0.471, + 0.352 + ], + "angle": 0, + "content": "[38] Dmitriy Smirnov, Michael Gharbi, Matthew Fisher, Vitor Guizilini, Alexei Efros, and Justin M Solomon. Marionette: Self-supervised sprite learning. Advances in Neural Information Processing Systems, 34, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.354, + 0.47, + 0.409 + ], + "angle": 0, + "content": "[39] Pratul P. Srinivasan, Richard Tucker, Jonathan T. Barron, Ravi Ramamoorthi, Ren Ng, and Noah Snavely. Pushing the boundaries of view extrapolation with multiplane images. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.412, + 0.47, + 0.453 + ], + "angle": 0, + "content": "[40] Zachary Teed and Jia Deng. Deepv2d: Video to depth with differentiable structure from motion. arXiv preprint arXiv:1812.04605, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.456, + 0.47, + 0.512 + ], + "angle": 0, + "content": "[41] Chuan Wang, Haibin Huang, Xiaoguang Han, and Jue Wang. Video inpainting by jointly learning temporal structure and spatial details. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 5232-5239, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.512 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.893, + 0.134 + ], + "angle": 0, + "content": "[42] John YA Wang and Edward H Adelson. Representing moving images with layers. IEEE transactions on image processing, 3(5):625-638, 1994. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[43] Yonatan Wexler, Eli Shechtman, and Michal Irani. Space-time completion of video. PAMI, 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.163, + 0.894, + 0.218 + ], + "angle": 0, + "content": "[44] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond. Computer Graphics Forum, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.219, + 0.894, + 0.259 + ], + "angle": 0, + "content": "[45] Rui Xu, Xiaoxiao Li, Bolei Zhou, and Chen Change Loy. Deep flow-guided video inpainting. In CVPR, pages 3723-3732, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.261, + 0.893, + 0.315 + ], + "angle": 0, + "content": "[46] Vickie Ye, Zhengqi Li, Richard Tucker, Angjoo Kanazawa, and Noah Snavely. Deformable sprites for unsupervised video decomposition. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.317, + 0.893, + 0.357 + ], + "angle": 0, + "content": "[47] Kaidong Zhang, Jingjing Fu, and Dong Liu. Flow-guided transformer for video inpainting. In European Conference on Computer Vision, pages 74-90. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.359, + 0.893, + 0.413 + ], + "angle": 0, + "content": "[48] Zhoutong Zhang, Forrester Cole, Zhengqi Li, Michael Rubinstein, Noah Snavely, and William T Freeman. Structure and motion from casual videos. In European Conference on Computer Vision, pages 20-37. Springer, 2022. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.414, + 0.893, + 0.468 + ], + "angle": 0, + "content": "[49] Zhoutong Zhang, Forrester Cole, Richard Tucker, William T Freeman, and Tali Dekel. Consistent depth of moving objects in video. ACM Transactions on Graphics (TOG), 40(4):1-12, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.47, + 0.893, + 0.512 + ], + "angle": 0, + "content": "[50] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images. In SIGGRAPH, 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.512 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.956 + ], + "angle": 0, + "content": "639" + } + ] +] \ No newline at end of file diff --git a/2023/Omnimatte3D_ Associating Objects and Their Effects in Unconstrained Monocular Video/5700f2ea-1d90-4ebb-a4e2-33ec303e3c7c_origin.pdf b/2023/Omnimatte3D_ Associating Objects and Their Effects in Unconstrained Monocular Video/5700f2ea-1d90-4ebb-a4e2-33ec303e3c7c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5f7616b59de075d67a16489b47afb4255de9d6a6 --- /dev/null +++ b/2023/Omnimatte3D_ Associating Objects and Their Effects in Unconstrained Monocular Video/5700f2ea-1d90-4ebb-a4e2-33ec303e3c7c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:999a19b1f4cd58488e804335184d52565610032e3e34ff797ec4fe70488ce711 +size 7327538 diff --git a/2023/Omnimatte3D_ Associating Objects and Their Effects in Unconstrained Monocular Video/full.md b/2023/Omnimatte3D_ Associating Objects and Their Effects in Unconstrained Monocular Video/full.md new file mode 100644 index 0000000000000000000000000000000000000000..1e4b1adc3088ad9dd1af5ed4a052233c4bebf7a3 --- /dev/null +++ b/2023/Omnimatte3D_ Associating Objects and Their Effects in Unconstrained Monocular Video/full.md @@ -0,0 +1,446 @@ +# Omnimatte3D: Associating Objects and their Effects in Unconstrained Monocular Video + +Mohammed Suhail1,2 + +2 Erika Lu4 Leonid Sigal1,2,3 + +Zhengqi Li + +Noah Snavely4 + +Forrester Cole + +1University of British Columbia + +$^{2}$ Vector Institute for AI + +3Canada CIFAR AI Chair + +4Google + +![](images/03bbd923cf0a354ef93ce235c54b6a0a99f3e51d471523a549ad38ddaedab5a6.jpg) +Input RGB + +![](images/5dae43cac38bb1b03b87a1cbd6fce6b1709f2284bb6e2fb5cfbff9cc40a2afe8.jpg) +Input Masks + +![](images/1d5c43233d2a14d35507f132cc7b44a884ea6381a5272c0499935ef815245e23.jpg) +Input Depth +Figure 1. Layer decomposition under strong camera parallax. Given an input video with unconstrained camera motion and approximate object masks and depth (left), our method estimates a layered representation composed of a background layer and object layers containing the subjects of interest and their associated effects (e.g. shadows). Results of combining object layers and background are shown on right. + +![](images/f44f80334f95e168cccff91f53db06d8f946618a8a5becc283c8a00d2e80e312.jpg) + +![](images/e8374b32729bcb5ef8f552f80b4cce507e6056647ad005b27a5b7efac74d6f89.jpg) +Layer 1 + Background + +![](images/b540f14db5bb39921df5bb5dc2daddc55588d46af2194e9130a6bef22bc9cd1f.jpg) +Layer 2 + Background + +# Abstract + +We propose a method to decompose a video into a background and a set of foreground layers, where the background captures stationary elements while the foreground layers capture moving objects along with their associated effects (e.g. shadows and reflections). Our approach is designed for unconstrained monocular videos, with an arbitrary camera and object motion. Prior work that tackles this problem assumes that the video can be mapped onto a fixed 2D canvas, severely limiting the possible space of camera motion. Instead, our method applies recent progress in monocular camera pose and depth estimation to create a full, RGBD video layer for the background, along with a video layer for each foreground object. To solve the underconstrained decomposition problem, we propose a new loss formulation based on multi-view consistency. We test our method on challenging videos with complex camera motion and show significant qualitative improvement over current approaches. + +# 1. Introduction + +Decomposing a video into meaningful layers (as show in Fig. 1) is a long-standing and complex problem [42] that has + +seen a recent surge in progress with the application of deep neural networks [17, 24, 25, 46]. A challenging variant of layer decomposition is the omnimatte [25] task, which aims to separate an input video into a background and multiple foreground layers, each containing an object of interest along with its correlated effects such as shadows and reflections, thus enabling video editing applications such as object removal, background replacement and retiming. The original work on omnimatte [24, 25] introduced a self-supervised approach for decomposing a video by assuming the background can be unwrapped onto a single static 2D background canvas using homography warping. Following work [17, 46] relaxed the homography restriction, but maintained the necessity of unwrapping onto a 2D canvas. + +This 2D modelling, however, limits the applicability of these methods to camera motions that have limited or zero parallax; intuitively, only panning camera motions are allowed. If the camera's center of projection moves a significant distance over the video, 2D methods are unable to learn an accurate background and are forced to place the background detail in a foreground layer (Figure 2). Since camera parallax is quite common, 2D methods are severely restricted in practice. + +To handle camera parallax, we exploit another recent line + +![](images/30b7cafafa870b70a4ca6193ce0fdc9d1717074df16336e5bc5abd603a32befc.jpg) +Figure 2. Existing methods fail on scenes with parallax. Top row: Omnimatte [25] fails due to inaccurate homography registration, reconstructing the majority of the video in the foreground layer. Middle row: Neural-Atlas [17] incorrectly places trees in the foreground layer and struggles to inpaint the person region. Bottom row: Our method obtains a clean background and captures the person and their shadow in the foreground layer. + +of work on estimating 3D camera position and depth maps from casually-captured video with moving objects [18,26, 48, 49]. Instead of a 2D layer decomposition, we propose to learn a 3D background model that varies per-frame and includes inpainted RGB and depth in the regions occluded by the foreground. Allowing the background to vary per-frame, rather than assuming a global, static canvas, creates an ill-posed decomposition problem, as variation in each input pixel could be explained by a foreground object, background, or both. To resolve this ambiguity, we enforce that the background varies slowly over time with 3D multi-view constraints using dense depth and camera pose estimates [48]. As shown in Fig. 2, our model can successfully separate the layers obtaining a clean background RGB whilst capturing only the person and their shadow in the foreground layer. + +Contributions. Our main contribution is a novel video decomposition method capable of accurately separating an input video with complex object and camera motion (e.g., parallax) into a background and object layers. To address the ill-posed decomposition problem, we design regularization terms based on multi-view consistency and smoothness priors. The resulting model produces clean decompositions of real-world videos where previous methods fail. + +# 2. Related Work + +Layered video representations. There is a large body of work on decomposing videos into layers based on appearance and motion cues [4, 8, 16, 20, 36, 42]. Layered video representations have proven useful for a variety of applications, including view synthesis for static scenes [39, 50], reflection removal [1, 2], segmentation [9], game deconstruction [38], and text-based video editing [3]. Recently, Omnimatte [24, 25] aimed to decompose a video into layers + +that group objects with their correlated effects, allowing the user to perform editing tasks such as object removal and retiming. However, as discussed in Section 1, these works use a constrained background model which limits their use case to panning camera motions only. Subsequent work [17,46] relaxes this assumption using general image warps, instead of homographies, but still cannot handle significant parallax; moreover, their goal is to learn a per-frame mapping to global sprites to enable applying consistent edits (e.g. style transfer) rather than targeting object removal or video stabilization. + +Monocular video depth. Multi-view stereo based methods [23,40] incorporate motion estimation and multi-view reconstruction to predict depth by combining information from multiple reference frames. Hybrid depth methods combine single-view depth estimates from a pretrained network with geometric reasoning to obtain coherent depth predictions. With depth maps from monocular video as initialization, these methods use correspondence to enforce consistency by either assuming static regions [26] or by modeling the scene flow [49]. Recent methods simultaneously estimate depth maps and camera poses [19, 48], exploiting the depth prior to resolve camera motion ambiguities. We rely on CasualSAM [48] to obtain camera poses and initial depth estimates as their method performs well on casually captured video. + +Video completion and object removal. Patch-based methods for video inpainting [7, 12, 14, 28, 43] complete the missing regions by borrowing pixels from other frames, relying on correspondence estimated by homography or optical flow. Learning-based works have explored the use of 3D convolutions [5, 41], attention mechanisms [21, 30, 47] and flow-based methods [10, 45] to achieve more realistic and consistent inpainting. These methods assume that the entire region to inpaint is specified by an input mask, whereas our method jointly learns where to inpaint the background while separating foreground objects and their correlated effects. + +NeRF for dynamic scenes. With the recent success of neural fields [29,44] for view synthesis of static scenes, there has been interest in their adoption for dynamic scenes [11,22,27]. These works learn two NeRF models that decompose a scene into a static and dynamic region. However, these methods only support a single foreground layer and do not account for correlated effects such as shadows. ST-NeRF [15] proposed to learn separate NeRFs for each person to obtain an editable neural representation. Their method, however, requires each frame in the video to be captured from 16 different viewpoints simultaneously. Unlike NeRF methods, our method represents the video with multi-layer mesh geometry where object layers include associated effects, allowing editing as well as fast rerendering of the scene. + +![](images/54af525eb386bfcbf3aafd896773ab954519b7222151a588440d3c72b57c61dc.jpg) + +![](images/4f825ba3e51340be93023f21cfac849c715535b074ff6b18c4c9d509fd56cc40.jpg) +Figure 3. Model Overview. Given a frame from the video along with an input mask, disparity and known camera pose, we first extract features by stacking the input and passing them though a 2D UNet architecture. These features are input to the background and foreground networks. The background network predicts the inpainted background RGB and disparity. The foreground networks predict an RGB-A image for each foreground layer. For scenes with multiple layers, the RGB-A image for each layer is predicted using independent networks. + +# 3. Method + +Given an input video, with known camera poses and initial depth estimates, the goal is to predict a per-frame layer decomposition consisting of a background layer and $N$ foreground layers encompassing objects of interest along with their associated effects. To this end, we propose a new method that predicts a layered separation of an input video without relying on a 2D background canvas, thereby allowing for greater flexibility of camera motion. We show an overview of our model in Figure 3. We optimize our model to output layers that reconstruct a single target video subject to a projection consistency loss. For each frame in the video, we estimate a layered representation consisting of background and foreground object layers where the contents of the object layers are conditioned on their corresponding input mask. Post-optimization, depth for foreground objects can be extracted from the input depth to produce a Layered Depth Image (LDI). Unlike prior works that rely on deep priors for inpainting the depth and color of occluded regions in the background, our method relies on multi-view consistency losses to steer the inpainting process. + +# 3.1. Layer Decomposition Networks + +We represent a frame at time $a$ in the video as $F_{a} = \left(I_{a},D_{a},\{M_{a}^{i}\}_{i = 1}^{N}\right)$ , where $I_{a}$ is the RGB frame, $D_{a}$ is the + +disparity, and $\{M_a^i\}_{i = 1}^N$ are the input masks for $N$ objects in the video. We denote the camera extrinsics for the frame $F_{a}$ as $[R_a t_a]$ where $R_{a}$ is the rotation matrix and $t_a$ is the translation vector. The camera intrinsics are denoted as $K_{a} = [f_{x},f_{y},c_{x},c_{y}]$ where $f_{x}$ and $f_{y}$ are the focal lengths and $c_{x},c_{y}$ is the principal point. + +For each frame $(F_{a})$ in the input video, our method first extracts features by passing it to a feature extractor network modeled using a 2D UNet [35] architecture. The input to the feature extractor is constructed by concatenating input masks with the masked frame RGB and disparity along the channel dimension. The masked RGB and disparity are obtained as + +$$ +\tilde {Y} _ {a} = \left[ Y _ {a} \odot M _ {a} ^ {b g} \| Y _ {a} \odot M _ {a} ^ {1} \| \dots \| Y _ {a} \odot M _ {a} ^ {N} \right] \tag {1} +$$ + +where $Y_{a}$ is the concatenated RGB-D, $M_{a}^{bg} = 1 - \bigcup_{i=1}^{N} M_{a}^{i}$ is an approximate mask for the background and $\| \cdot$ is the concatenation operation. These features act as input to both the background and foreground object networks. + +The background network (shown in green in Fig. 3) consists of a 2D UNet and a convolutional neural network (CNN) that predicts the inpainted background color $(\mathcal{C}_a^{bg})$ and depth $(\mathcal{D}_a^{bg})$ respectively. The foreground networks are separate CNNs that predict RGB $(\mathcal{C}_a^i)$ and alpha $(\mathcal{A}_a^i)$ images for each foreground layer independently. These predicted outputs, along with the input disparity, are combined to form an LDI (see supplement for details). + +The feature extractor and background RGB UNets consist of five down-sampling and up-sampling layers with three residual blocks at every resolution. The background disparity and foreground layer prediction networks are comprised of single convolutional layers. + +# 3.2. Differentiable Rendering + +In addition to a reconstruction loss, our method relies on a projection consistency term (described in Section 3.3) that requires differentiable projections and renderings of the predicted background layer to different time steps. We achieve this using a differentiable rasterization framework [6]. + +To project the predicted background layer from a source time step $a$ to target time step $b$ , we first construct a mesh $\mathcal{M}_a = (\mathcal{V}_a, \mathcal{F}_a)$ , where $\mathcal{V}_a$ and $\mathcal{F}_a$ are the mesh vertices and faces. Each vertex in the mesh corresponds to an image pixel. To obtain the vertex coordinate for a pixel, we unproject its homogeneous coordinates $x = [i,j,1] \in \mathbf{RP}^2$ using the predicted disparity as + +$$ +V _ {a} ^ {x} = d _ {i j} \left(\frac {\left(i - c _ {x}\right)}{f _ {x}}, \frac {\left(j - c _ {y}\right)}{f _ {y}}, 1\right) \tag {2} +$$ + +where $d_{ij}$ is the depth at pixel $(i,j)$ . The attributes for the vertices are composed of the predicted RGB, an alpha masking consisting of all ones, and the pixel coordinates. The faces for the mesh are constructed by adding an edge between vertices of neighboring pixels (bottom right in Fig. 3). The view-projection matrix to render the predicted background from a target camera is computed as + +$$ +H _ {a \rightarrow b} = P _ {a} \left[\begin{array}{c c}R _ {a} ^ {T} R _ {b}&R _ {a} ^ {T} \left(t _ {b} - t _ {a}\right)\\\mathbf {0}&1\end{array}\right] \tag {3} +$$ + +where $P_{a}$ is the perspective matrix that maps camera-space to clip space. The view-projection matrix along with the estimated mesh is passed through a differentiable rasterizer that returns an RGB-A $(\mathcal{C}_{a\to b}^{bg},\mathcal{A}_{a\to b}^{bg})$ image corresponding to the projection of the background to the target time step. + +# 3.3. Losses + +The total loss to train our model is: + +$$ +\mathcal {L} _ {\text {t o t a l}} = \lambda_ {1} \mathcal {L} _ {\text {r e c o n}} + \lambda_ {2} \mathcal {L} _ {\text {p r o j}} + \lambda_ {3} \mathcal {L} _ {\text {m a s k}} + \lambda_ {4} \mathcal {L} _ {\text {d i s p}} + \lambda_ {5} \mathcal {L} _ {\text {r e g}} \tag {4} +$$ + +We describe each loss term in the following sections. Please see the supplement for values of the scalar weights $\lambda$ . + +# 3.3.1 Reconstruction Loss + +For a source frame $F_{a}$ we reconstruct the image from the predicted background RGB and foreground RGBA layers using back-to-front compositing [33] (bottom left in Fig. 3). During training, we minimize the $\mathcal{L}_2$ loss between the original RGB image and the reconstruction. + +$$ +\mathcal {L} _ {\text {r e c o n}} = \left\| \left(I _ {a} - f \left(\mathcal {C} _ {a} ^ {b g}, \left\{\mathcal {C} _ {a} ^ {i}, \mathcal {A} _ {a} ^ {1} \right\} _ {i = 1} ^ {N}\right) \right\| _ {2} \right. \tag {5} +$$ + +where $f$ is the back-to-front over-compositing function. + +A large number of layer decompositions can reproduce the original frame and minimize the reconstruction loss. To guide the foreground layers to be semantically meaningful, we incorporate an $\mathcal{L}_1$ loss between the predicted foreground alphas and the input masks. + +$$ +\mathcal {L} _ {\text {m a s k}} = \frac {1}{N} \sum_ {i = 1} ^ {N} \| M _ {a} ^ {i} - \mathcal {A} _ {a} ^ {i} \| _ {1} \tag {6} +$$ + +Since we want the foreground alphas to be able to capture effects beyond the input mask, the weight for this loss is decayed as described in Sec. 3.4. + +# 3.3.2 Projection Consistency Loss + +The reconstruction loss on the over-composed image does not provide any gradients for the occluded region in the background. Prior works overcome this challenge by learning a background model that maps from a static 2D canvas. However, since our model predicts the background for each frame independently, we require additional regularization to prevent random artifacts or foreground details from appearing in occluded regions. We thus introduce a projection consistency loss on the background prediction. Each batch in our input is composed of three frames from the video. We denote these frames as $F_{a}$ , $F_{b}$ and $F_{c}$ . The frames $F_{a}$ and $F_{b}$ act as source images for the target frame $F_{c}$ . The criteria used for selecting the frames is described in Section 3.4. + +To estimate the projection consistency loss, we project the predicted background RGBs $(\mathcal{C}_a^{bg}$ and $\mathcal{C}_b^{bg})$ for the two source frames, $F_{a}$ and $F_{b}$ , using the method described in Sec. 3.2. We use the predicted disparities $(\mathcal{D}_a^{bg}$ and $\mathcal{D}_b^{bg})$ to construct the mesh and set the alpha values to one for all pixels. We then apply an $\mathcal{L}_2$ loss between the projected source backgrounds and the predicted target background. + +$$ +\mathcal {L} _ {p n j} ^ {r g b} = \frac {1}{2} \sum_ {k \in \{a, b \}} \| \mathcal {A} _ {k \rightarrow c} ^ {b g} \odot \left(\mathcal {C} _ {k \rightarrow c} ^ {b g} - \mathcal {C} _ {c} ^ {b g}\right) \| _ {2} \tag {7} +$$ + +where $\odot$ denotes element wise multiplication. We mask the loss using the projected alphas $(\mathcal{A}_{k\to c}^{bg})$ to prevent noisy gradients from the boundaries of projection. + +Similarly, to encourage consistent inpainting of depth we apply a similar projection loss, + +$$ +\mathcal {L} _ {p r o j} ^ {c o o r d} = \frac {1}{2} \sum_ {k \in \{a, b \}} \| \mathcal {A} _ {k \rightarrow c} ^ {b g} \odot \left(\mathcal {X} _ {k \rightarrow c} ^ {b g} - \mathcal {X} _ {c} ^ {b g}\right) \| _ {2} \tag {8} +$$ + +where $\mathcal{X}_c$ are the world coordinates for pixels in the frame $F_{c}$ obtained using the predicted background disparity $\mathcal{D}_c^{bg}$ . $\mathcal{X}_{k\to c}^{bg}$ are the projections of the world coordinate from frame $F_{k}$ to $F_{c}$ and $\mathcal{A}_{k\to c}^{bg}$ are the projected alphas. The projection loss $\mathcal{L}_{proj}$ is the sum of $\mathcal{L}_{proj}^{rgb}$ and $\mathcal{L}_{proj}^{coord}$ . + +# 3.3.3 Alpha Regularization + +While the projection loss encourages the background predictions to be consistent across frames, it can also lead the model toward undesirable or degenerate solutions. For example, a model that predicts a constant color at every pixel for every frame will yield a projection loss of zero and essentially force the background to be captured in foreground layers to minimize the reconstruction loss. To prevent such solutions, we regularize the foreground alphas with an $\mathcal{L}_1$ and approximate- $\mathcal{L}_0$ term as in [25]. + +$$ +\mathcal {L} _ {s p} = \frac {1}{N} \sum_ {i = 1} ^ {N} \| \mathcal {A} _ {a} ^ {i} \| _ {1} + \gamma \| 2 \cdot \sigma (5 \cdot \mathcal {A} _ {a} ^ {i}) - 1 \| _ {1} \tag {9} +$$ + +where $\sigma$ is the sigmoid function and $\gamma$ control the relative weight between the two loss terms. + +Additionally, we regularize the foreground alphas with a smoothness term to discourage them from capturing sharp details in the background which would otherwise arise due to the effects of depth imprecision or splatting in the rasterizer. The smoothness loss $(\mathcal{L}_s^{\text{alpha}})$ is computed using an $\mathcal{L}_1$ loss on the second order gradients of the predicted alphas. + +# 3.3.4 Disparity Loss + +To distill the available depth information into our model, we apply an $\mathcal{L}_2$ loss on the predicted background disparity. + +$$ +\mathcal {L} _ {\text {d i s p}} = \left\| M _ {a} ^ {b g} \odot \left(D _ {a} - D _ {a} ^ {b g}\right) \right\| _ {2} \tag {10} +$$ + +where $M_{a}^{bg} = 1 - \bigcup_{i=1}^{N} M_{a}^{i}$ is the mask indicating the approximate unoccluded background region. Additionally, to encourage smoothness in the depth inpainting we add an $L_{2}$ loss $(\mathcal{L}_{s}^{disp})$ on the second-order gradients of the predicted disparity similar to $\mathcal{L}_{s}^{alpha}$ . The sum of the losses $\mathcal{L}_{sp}, \mathcal{L}_{s}^{alpha}$ , and $\mathcal{L}_{s}^{disp}$ forms the regularization loss $\mathcal{L}_{reg}$ . + +# 3.4. Frame Selection + +The prediction/inpainting of the background RGB and disparity is guided by the projection consistency loss. Intuitively, the source and target frames should be such that regions that are occluded in one are visible in the other whilst having a large overlap in the static regions. To identify such frame pairs we use an approximate frame overlap metric. For two frames $F_{a}$ and $F_{b}$ we estimate the metric by projecting a mesh from $F_{a}$ to $F_{b}$ . The attributes of the mesh vertices is set using $B_{a} = \cup_{i} M_{a}^{i}$ where for a pixel $(i,j)$ , $B(i,j)$ indicates if the pixel is occupied by a foreground or background as dictated by the input mask. The alpha value for the rasterizer input $A_{a}^{bg}$ are set to all ones. The frame overlap metric is then estimated as + +$$ +O _ {a \rightarrow b} = \frac {\left\| A _ {a \rightarrow b} \right\|}{H \cdot W} \cdot \frac {\left\| B _ {c} - A _ {a \rightarrow b} \right\|}{\left\| B _ {c} \right\| + \left\| A _ {a \rightarrow b} \right\|} \tag {11} +$$ + +For each frame in the video, we rank every other based on the frame overlap metric. During training, for a source frame $F_{a}$ , we choose the another frame $F_{b}$ randomly from the top 10 frames that maximize the overlap metric. The target frame $F_{c}$ is chosen to be midway between the two source frames. + +# 3.5. Detail Transfer + +The optimized layers $\mathcal{C}_k^{bg}$ and $\mathcal{C}_k^i$ may miss high-frequency details present in the input video. Detail may be directly recovered from the input video using the transfer approach of Layered Neural Rendering [24], where the transmittance defined by the alpha maps $\mathcal{A}_k^i$ controls the amount of detail transferred to each layer. In addition, the background depth map allows reprojection of detail from nearby frames, similar to [34], allowing some detail to be added in regions originally occluded by foreground elements. Figure 10 shows the effect of detail transfer on object removal. + +# 4. Experiments + +# 4.1. Training Details + +Data preprocessing. We perform our experiments on videos from the DAVIS [32] dataset. We use CasualSAM [48] to obtain camera poses and initial depth estimates. The input masks are estimated using MaskRCNN [13]. + +Training Schedule. During the warm-up stage, the model is trained using the disparity $(\mathcal{L}_{disp})$ and mask losses $(\mathcal{L}_{mask})$ along with the smoothness regularizers, for initialization purposes. After the warm-up stage, the weights for these loss terms are decayed using a cosine schedule to allow the model to inpaint missing depths and include effects such as shadows in the foreground alpha layers. All other loss terms use a linear schedule starting from 0 during the warm-up stage and remain constant for the remainder of training. + +# 4.2. Layer Decomposition Results + +Figure 4 shows examples of layer decomposition results for various real-world videos from the DAVIS [31] dataset. We show comparisons against Omnimatte [25] and NeuralAtlas [17] and observe clear improvements over these methods. On the Scooter-gray scene, the background predictions for the baselines are severely distorted due to complex camera motion. The incorrect background forces these methods to compensate by reconstructing the background details in the foreground layer. Our method is able to generate an accurate background and predict a much cleaner foreground layer, which includes the shadow. Similarly, on the Snowboard scene, inaccurate homography registration leads to incorrect background prediction by Omnimatte, particularly in the rocks in the upper right corner. Neural-Atlas shows some improvement in the background predictions but loses the rock texture and is unable to capture the shadow in the foreground layer. Our method successfully captures the shadow in the + +![](images/20adf8f5fba2a5c5a79dbfe45ff6974ee9a96edd59dbca6968eb972bc5f4ae2e.jpg) +Snowboard + +![](images/ed86914aadaa1209659d7dd4b48f40d8ffe86d1b98173febdccda68424ddb3c6.jpg) + +![](images/5fb735d45ad0dd6a681dbdaf8eb3a628d9883cab994b3f070b9b3242d4d00660.jpg) +Scooter-gray + +![](images/e51a4bc01b214dc17b6484248745823f0596b286aba359b1463f10ef87c83b12.jpg) +(a) Input RGB and Mask + +![](images/d15a9c773cf47cc85510cc60b16175da84cb626c0748e29cf8059483d4d20d40.jpg) + +![](images/b36869128d160dd787176da9f8b7d946a1a27ee2d714a47f37e2124118f1b705.jpg) + +![](images/52b189b6a01711cee7f2d2c822e7dc22bb70fecd5be50c5c68e84491ab66c548.jpg) + +![](images/f82c21b08afc888f82de1dcb8fd4443713cf76b3ec7242e92ddc0b45bdb0fcb6.jpg) + +![](images/2fe5c2c4b8766f0381d57de92c581e0e734a5065c0e423a883aa767ba11ab929.jpg) + +![](images/8229a8044af90657caad6b08a4f61242f4ad14e943f43e2d07cd348ff295a9c4.jpg) +(b) Omnimatte [Lu, et al.] + +![](images/c962a57bb6a79e84ae5eaea389c5cf41bbc81d793aff72b9086d91829c88cd91.jpg) + +![](images/c3a02875d7a22924f919829172e4176765060e2cc1951539ddbf2f19e0f36228.jpg) + +![](images/ccb7fa06775183f0cbbd8efbf29a89ae52b83731639930fe5d98df89fabca7cc.jpg) + +![](images/9ff863191447336726b224d0b71082af812d7617c27683a78a25bb740b9980ce.jpg) + +![](images/39eac2b1182121aa36d86bd7172b957f28d385bc21885266c7851e71256c205a.jpg) + +![](images/7669387a3aa52886bf58c2a63ea8300a0d746a350f65d91f60b3090070a1e22f.jpg) +(c) Neural Atlas [Kasten, et al.] + +![](images/2c7b0948021ce7587f18b6518522ae401751a78740ed6ecfbd0ac78420e48312.jpg) + +![](images/54971f03d5c0a546679b9498640ec34cd6a725c035e006f598081049379b1e59.jpg) + +![](images/eb1f6d1701db8f3229140ddecbdd80b14701bc8e0c407250bae1df23c0e7f860.jpg) + +![](images/dbfd4b77c0ffb62f3aaae0d84b88e2e2a186b80c73b44cdda4b1142d2f970b8a.jpg) + +![](images/62d37bc52c0eb1812811b2f496956b5f496f8256d57b00c549b2db2bc64ef9d7.jpg) + +![](images/07fb06dd1c380a14ec2ce1849be95f2d4f66396c74f2b535967e400375d86772.jpg) +(d) Ours + +![](images/d4f9dd9e564e34541bdd7e976e5b0eb19776fe84813bfe3dd58ad7c11c1200d9.jpg) +Longboard + +![](images/5021088c22f6f892ecfdeea3b81d75ff5283d058110e7f64541ae792b90a16a3.jpg) + +![](images/b2047b7136227340aadfdd57aa5cb2bda93238cf25f66d55eec97316c667c5c2.jpg) +(a) Input RGB and Masks +Figure 4. Qualitative layer decomposition results on DAVIS. Top: examples of single-layer decomposition on Snowboard and Scooter-gray. For each example, we show the predicted background layer (row 1), predicted object layer RGBA (row 2), and alpha channel visualization (row 3). Bottom: example of multi-layer decomposition on the Longboard scene. On Longboard, we only compare against Omnimatte since Neural-Atlas does not support multi-layer decomposition. Our method produces stronger background inpainting results and more accurately disentangles the foreground objects and their effects from the background. + +![](images/03d8b10fdb4617753fd969c2929f723ca25f25b6ffe5e254e82357a1f13538ba.jpg) + +![](images/0c80a3ec8f783f02bdd536abdb2413053a35ff174138e70aa7183dd2e1638e95.jpg) + +![](images/759393cf1a10eb197c3cd49608101c0ae08aff0a6eb320ed288cd374dbb6df8e.jpg) +(b) Omnimatte [Lu, et al.] + +![](images/4834d6c94a970b946c22997340df17c436a9bf9401054749d3d688acf379ceae.jpg) + +![](images/503bbc695f1eff1d08e50f2b75dabadaf16adcf4f4246a0002c23273f72a6a26.jpg) + +![](images/2b2fa42908f225562330873f7b882c289e29c6b0b79fad6aa05627b6675dfbf1.jpg) + +![](images/96681310eda1bc3ac9a1be98d63ac85993bb546fc5700c2fd0bd5da3fccf59f5.jpg) + +![](images/3d668f5836aa2544266770652979a6d6a98276e306d11384c134d3350d9fb2cc.jpg) +(c) Ours + +![](images/aff5a4b3e7d3a341bbccedb7a6985f4161ab287b910ba35a0736d05cb5f655a1.jpg) + +![](images/020eb3e353a897fb9ea21d7eb73c737137b78ffe7e77beb1fe15927dc6340e69.jpg) + +![](images/712d558bba4b27d075dc19895bee7985952b200810d8cb2ea54e04c1a09712d4.jpg) +Figure 5. Object removal. Top row shows the input frames with the input masks highlighted in blue and green. Subsequent rows show the result of removing the green and blue objects respectively. + +foreground layer while accurately predicting the background color. The last row shows an example video (Longboard) with two object layers. For this scene, we compare against Omnimatte as Neural-Atlas does not support multi-layer decomposition. Omnimatte's foreground layer for the cyclist contains significant pollution from background elements. Additionally, despite severe occlusion, our method is able to plausibly inpaint the background as compared to Omnimatte. We provide additional results in the supplementary material. + +# 4.3. Object Removal + +Our predicted layers can be used to selectively remove some or all objects from a scene. Figure 5 shows an example of object removal using our layered representation on the Longboard video. Note that the input masks only account for the people and do not include attached objects such as the bicycle or longboard. Our method groups these objects with the correct person, allowing a user to easily remove different people from the scene without the undesirable effect of leaving behind attached objects. + +# 4.4. Depth-based Applications + +While our method does not directly produce foreground depth, the input depth can be transferred to the foreground layers using the predicted foreground matrices to construct a full Layered Depth Image (LDI) for each frame (see the supplement for details). This representation is useful for a variety of applications, including camera stabilization and synthetic defocus, which we show in the following sections. + +Camera Stabilization. For an input video captured with a shaky camera, we can stabilize the camera path by rendering the scene from a stable camera viewpoint at each time-step. Figure 6 shows results on rendering the Dancing-person [37] + +![](images/ad44e8c8ddf93a86611eede45f5fb9a7a162dff347471ff0c9e5a8fb62e44af3.jpg) +Figure 6. Camera stabilization. Our LDI representation can be re-rendered from new camera positions to produce stabilized or modified camera paths. Left: input video with camera dolly motion. Right: frames re-rendered from a stationary camera. + +scene from a stationary camera (for a moving camera path, see supp. material). The explicit mesh representation obtained from our model facilitates editing, as they can be used with commercial video editing tools. + +Synthetic Defocus. Our foreground matte can be used to create a synthetic defocus (bokeh) effect (Fig. 7). To create this effect, we blur the background based on the depth map computed by CasualSAM [48], then composite our foreground RGBA layer on top of the blurred background. The single-layer depth map alone cannot support this effect, as it is insufficiently accurate near the subject's boundary, + +![](images/f75fb662a4e73d083449fe643b9580bae88f8775262e53ddcec8f4c332259744.jpg) +Original + +![](images/6a87757ef4294670538aec93d049591b251d6bb032a63a9a7c8825159249de83.jpg) +Ours + +![](images/e826f33b8cced0668e9bbbfa5515432b85c36cc08276d43225aaf680e95370a1.jpg) +CasualSAM [Zhang, et al.] + +![](images/72e9867a7ae975a0a3934631f112fd240385acde390c019957ec129f99807d35.jpg) +Omnimatte [Lu, et al.] +Figure 7. Synthetic defocus. Our foreground matte can enable a synthetic defocus effect (Ours). State-of-the-art single-layer depth maps (CasualSAM [48]) are not accurate near the subject's boundary, causing the goat's head to blur. The foreground matte from Omnimatte [25] contains background pixels, causing parts of the background to remain unblurred. + +![](images/a6a93ec25857b83686f5972fb8e71b5efd201b7d3ed3a548212ad4ab2099e059.jpg) +(a) Input RGB and Mask + +![](images/a7d5844dfca74fef5014e5c75ac387fc8d04971f3615f38632f1fbed506bca6f.jpg) +(b) Random Frame Selection + +![](images/ce2dfd7257cb489674c7bd36a6e90b9e817c46deb1dcc9fec04f6deb212ca26d.jpg) +(c) Our Frame Selection +Figure 8. Frame Selection Ablation. Results from random frame selection instead of the proposed heuristic. Unlike the full method, the foreground layer does not accurately capture the shadow. + +causing important small features to be erroneously blurred (Fig. 7, bottom). The foreground matte from Omnimatte [25] is also insufficiently accurate, containing elements of the background that remain erroneously sharp when composited over the blurred background layer. + +# 4.5. Ablations + +Frame Selection. We ablate our frame selection method 3.4 by training a model using random frame sampling. For a source frame $F_{a}$ , we randomly sample a shift value to obtain the second source $F_{b}$ . The target frame $(F_{c})$ is then chosen mid-way between the source frames. Figure 8 shows results from a model trained using the random frame selection on the hike scene. Compared to the model trained with our frame selection heuristic, random selection leads to most of the shadow being erroneously captured in the background layer. While it is possible to select shift values that can generate desirable results, the proposed frame selection heuristic removes the need to search over this hyperparameter. + +![](images/f176aa26e0cb83e883bba536b8648bce9b832ad330f239409c45702030ce3b90.jpg) + +![](images/0d430f0b9ceeefbef2f54bfee616012061a11933a583540b161ce291d13e210d.jpg) +(a) Input RGB and Mask + +![](images/3e7fc12ddd773e8dfb837b4846e963e9a5dfa8d001da51e3df3af0599a8fb752.jpg) +(b) Without Projection Consistency + +![](images/74be33b2385d4863398e60470138e5340da1da0d50d1565a7baf8c726d28dba7.jpg) +(c) Full Method + +Projection Consistency Loss. We ablate the projection consistency loss in Fig. 9. As shown in column (b), the inpainting of the unobserved region fails without the projection + +![](images/5590f940c77393f901a6efdd9343c1bdf4acb35989ad34ae56c53b0ad852efd8.jpg) +Input +Figure 10. Detail Transfer Ablation. The background depth map allows reprojecting detail from nearby frames in time to further improve the visual quality of the predicted layers. From left to right: input frame, full result with detail transferred from frames $t$ , $t - 5$ , $t + 5$ , ablated result with detail only from frame $t$ (note blurry region under bicyclist), no background detail transfer. + +![](images/64e0de7ae39545e803182657e07f1f0ece1372544ade00f9c77f314006e04847.jpg) +Frames t, t-5, t+5 + +![](images/4b80b9525ca417e0c73de097fc182652dc4d28dfdaa58544ea171b352e67f08e.jpg) +Frame t only + +![](images/6477f9713fb018d1f68582956938a074f8e1dc032a213b77be8aa2c8a4aa45f8.jpg) +No detail transfer + +consistency term, as there is no reconstruction signal for that region. Additionally, the foreground layer is impacted due to the shadow being incorrectly placed in the background. + +# 5. Discussion and Limitations + +We present a new method for layer decomposition that separates a monocular video into a background and several object layers along with their associated effects. Unlike prior decomposition methods that rely on a static 2D background canvas, our method separates the video into layers on a per-frame basis, making it applicable to videos with unconstrained camera motion. To address the challenge of predicting a coherent background, we predict a 3D background and propose a multi-view consistency loss that enforces the background to only contain slowly moving or static details. + +Our model can produce per-frame Layered Depth Images that allow for various depth-based editing applications. + +A limitation of our method is shown in Figure 11. Unlike the background, foreground layers are not reprojected to nearby frames and thus foreground objects are not inpainted when occluded. This limitation could be addressed by adding a projection consistency loss similar to Eq. (7) for the foreground layers; however, reprojection of dynamic elements is challenging and requires modelling the sceneflow, which we leave for future work. + +![](images/48ee58420331082e8752bed6b56807f45c591c2adcd1db595d2486157d07f402.jpg) +Figure 9. Projection Loss Ablation. We show layer results from a model trained without the projection loss. The resulting background contains severe artifacts (column (b), top), and the foreground layer is missing the shadow of the car (column (b), bottom). +Input Frames + +![](images/8d3fc30fe8bff3a281c7de34d00a2db1fde03ef36d9710a94cfc1305838a1886.jpg) +Output Background +Figure 11. Limitation: object occlusions. While the background layer is inpainted using nearby frames, our method cannot inpaint the object layers. When an object passes behind the stationary background layer (middle), it is erased from the object layer (right). + +![](images/eb29d5e3843f5d991717960c25c62dd8e48173befffaf502028ad5a3a47890c7.jpg) +Output Object Layer + +# References + +[1] Jean-Baptiste Alayrac, João Carreira, and Andrew Zisserman. The visual centrifuge: Model-free layered video representations. In CVPR, 2019. 2 +[2] Jean-Baptiste Alayrac, Joao Carreira, Relja Arandjelovic, and Andrew Zisserman. Controllable attention for structured layered video decomposition. In ICCV, 2019. 2 +[3] Omer Bar-Tal, Dolev Ofri-Amar, Rafail Fridman, Yoni Kasten, and Tali Dekel. Text2live: Text-driven layered image and video editing. arXiv preprint arXiv:2204.02491, 2022. 2 +[4] Gabriel J Brostow and Irfan A Essa. Motion based decomposing of video. In ICCV, 1999. 2 +[5] Ya-Liang Chang, Zhe Yu Liu, Kuan-Ying Lee, and Winston Hsu. Free-form video inpainting with 3d gated convolution and temporal patchgan. In ICCV, pages 9066-9075, 2019. 2 +[6] Forrester Cole, Kyle Genova, Avneesh Sud, Daniel Vlasic, and Zhoutong Zhang. Differentiable surface rendering via non-differentiable sampling. In ICCV, pages 6088-6097, 2021. 4 +[7] Mounira Ebdelli, Olivier Le Meur, and Christine Guillemot. Video inpainting with short-term windows: application to object removal and error concealment. IEEE Transactions on Image Processing, 24(10):3034-3047, 2015. 2 +[8] Matthieu Fradet, Patrick Pérez, and Philippe Robert. Semi-automatic motion segmentation with motion layer mosaics. In ECCV, 2008. 2 +[9] Yossi Gandelsman, Assaf Shocher, and Michal Irani. "Double-DIP": Unsupervised image decomposition via coupled deep-image-priors. In CVPR, 2019. 2 +[10] Chen Gao, Ayush Saraf, Jia-Bin Huang, and Johannes Kopf. Flow-edge guided video completion. In ECCV, 2020. 2 +[11] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video. In ICCV, pages 5712-5721, 2021. 2 +[12] Miguel Granados, James Tompkin, K Kim, Oliver Grau, Jan Kautz, and Christian Theobalt. How not to be seen—object removal from videos of crowded scenes. Computer Graphics Forum, 31(2pt1):219-228, 2012. 2 +[13] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, pages 2961-2969, 2017. 5 +[14] Jia-Bin Huang, Sing Bing Kang, Narendra Ahuja, and Johannes Kopf. Temporally coherent completion of dynamic video. ACM Transactions on Graphics (TOG), 35(6):1-11, 2016. 2 +[15] Zhang Jiakai, Liu Xinhang, Ye Xinyi, Zhao Fuqiang, Zhang Yanshun, Wu Minye, Zhang Yingliang, Xu Lan, and Yu Jingyi. Editable free-viewpoint video using a layered neural representation. In ACM SIGGRAPH, 2021. 2 +[16] Nebojsa Joic and Brendan J Frey. Learning flexible sprites in video layers. In Proceedings of the 2001 IEEE Computer Society Conference on Computer Vision and Pattern Recognition. CVPR 2001, volume 1, pages I-I. IEEE, 2001. 2 +[17] Yoni Kasten, Dolev Ofri, Oliver Wang, and Tali Dekel. Layered neural atlases for consistent video editing. ACM Transactions on Graphics (TOG), 40(6):1-12, 2021. 1, 2, 5 + +[18] Johannes Kopf, Xuejian Rong, and Jia-Bin Huang. Robust consistent video depth estimation. In CVPR, pages 1611-1621, 2021. 2 +[19] Johannes Kopf, Xuejian Rong, and Jia-Bin Huang. Robust consistent video depth estimation. In CVPR, 2021. 2 +[20] M. Pawan Kumar, Philip H. S. Torr, and Andrew Zisserman. Learning layered motion segmentations of video. IJCV, 2008. 2 +[21] Sungho Lee, Seoung Wug Oh, DaeYeun Won, and Seon Joo Kim. Copy-and-paste networks for deep video inpainting. In ICCV, pages 4413-4421, 2019. 2 +[22] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In CVPR, pages 6498-6508, 2021. 2 +[23] Chao Liu, Jinwei Gu, Kihwan Kim, Srinivasa G Narasimhan, and Jan Kautz. Neural rgb (r) d sensing: Depth and uncertainty from a video camera. In CVPR, pages 10986-10995, 2019. 2 +[24] Erika Lu, Forrester Cole, Tali Dekel, Weidi Xie, Andrew Zisserman, David Salesin, William T Freeman, and Michael Rubinstein. Layered neural rendering for retiming people in video. In SIGGRAPH Asia, 2020. 1, 2, 5 +[25] Erika Lu, Forrester Cole, Tali Dekel, Andrew Zisserman, William T. Freeman, and Michael Rubinstein. Omnimatte: Associating objects and their effects in video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4507-4515, June 2021. 1, 2, 5, 7, 8 +[26] Xuan Luo, Jia-Bin Huang, Richard Szeliski, Kevin Matzen, and Johannes Kopf. Consistent video depth estimation. ACM Transactions on Graphics (ToG), 39(4):71-1, 2020. 2 +[27] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In CVPR, pages 7210-7219, 2021. 2 +[28] Yasuyuki Matsushita, Eyal Ofek, Weina Ge, Xiaou Tang, and Heung-Yeung Shum. Full-frame video stabilization with motion inpainting. IEEE Transactions on pattern analysis and Machine Intelligence, 28(7):1150-1163, 2006. 2 +[29] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2 +[30] Seoung Wug Oh, Sungho Lee, Joon-Young Lee, and Seon Joo Kim. Onion-peel networks for deep video completion. In ICCV, pages 4403–4412, 2019. 2 +[31] F. Perazzi, J. Pont-Tuset, B. McWilliams, L. Van Gool, M. Gross, and A. Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In CVPR, 2016. 5 +[32] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alexander Sorkine-Hornung, and Luc Van Gool. The 2017 DAVIS challenge on video object segmentation. arXiv:1704.00675, 2017. 5 +[33] Thomas Porter and Tom Duff. Compositing digital images. SIGGRAPH Comput. Graph., 18(3):253-259, Jan. 1984. 4 + +[34] Xuejian Rong, Jia-Bin Huang, Ayush Saraf, Changil Kim, and Johannes Kopf. Boosting view synthesis with residual transfer. CVPR, 2022. 5 +[35] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional Networks for Biomedical Image Segmentation. MICCAI, 2015. 3 +[36] Jonathan Shade, Steven Gortler, Li-wei He, and Richard Szeliski. Layered depth images. In Proceedings of the 25th annual conference on Computer graphics and interactive techniques, pages 231–242, 1998. 2 +[37] Zhenmei Shi, Fuhao Shi, Wei-Sheng Lai, Chia-Kai Liang, and Yingyu Liang. Deep online fused video stabilization. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1250-1258, 2022. 7 +[38] Dmitriy Smirnov, Michael Gharbi, Matthew Fisher, Vitor Guizilini, Alexei Efros, and Justin M Solomon. Marionette: Self-supervised sprite learning. Advances in Neural Information Processing Systems, 34, 2021. 2 +[39] Pratul P. Srinivasan, Richard Tucker, Jonathan T. Barron, Ravi Ramamoorthi, Ren Ng, and Noah Snavely. Pushing the boundaries of view extrapolation with multiplane images. In CVPR, 2019. 2 +[40] Zachary Teed and Jia Deng. Deepv2d: Video to depth with differentiable structure from motion. arXiv preprint arXiv:1812.04605, 2018. 2 +[41] Chuan Wang, Haibin Huang, Xiaoguang Han, and Jue Wang. Video inpainting by jointly learning temporal structure and spatial details. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 5232-5239, 2019. 2 + +[42] John YA Wang and Edward H Adelson. Representing moving images with layers. IEEE transactions on image processing, 3(5):625-638, 1994. 1, 2 +[43] Yonatan Wexler, Eli Shechtman, and Michal Irani. Space-time completion of video. PAMI, 2007. 2 +[44] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond. Computer Graphics Forum, 2022. 2 +[45] Rui Xu, Xiaoxiao Li, Bolei Zhou, and Chen Change Loy. Deep flow-guided video inpainting. In CVPR, pages 3723-3732, 2019. 2 +[46] Vickie Ye, Zhengqi Li, Richard Tucker, Angjoo Kanazawa, and Noah Snavely. Deformable sprites for unsupervised video decomposition. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2022. 1, 2 +[47] Kaidong Zhang, Jingjing Fu, and Dong Liu. Flow-guided transformer for video inpainting. In European Conference on Computer Vision, pages 74-90. Springer, 2022. 2 +[48] Zhoutong Zhang, Forrester Cole, Zhengqi Li, Michael Rubinstein, Noah Snavely, and William T Freeman. Structure and motion from casual videos. In European Conference on Computer Vision, pages 20-37. Springer, 2022. 2, 5, 7 +[49] Zhoutong Zhang, Forrester Cole, Richard Tucker, William T Freeman, and Tali Dekel. Consistent depth of moving objects in video. ACM Transactions on Graphics (TOG), 40(4):1-12, 2021. 2 +[50] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images. In SIGGRAPH, 2018. 2 \ No newline at end of file diff --git a/2023/Omnimatte3D_ Associating Objects and Their Effects in Unconstrained Monocular Video/images.zip b/2023/Omnimatte3D_ Associating Objects and Their Effects in Unconstrained Monocular Video/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d761b8b52c614b6971ee03ee06321c167df29487 --- /dev/null +++ b/2023/Omnimatte3D_ Associating Objects and Their Effects in Unconstrained Monocular Video/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8a39c95eeb83e58721a22f1fdb7a27106cd0f0da7158186bfe8bf8edbf4d9e04 +size 755332 diff --git a/2023/Omnimatte3D_ Associating Objects and Their Effects in Unconstrained Monocular Video/layout.json b/2023/Omnimatte3D_ Associating Objects and Their Effects in Unconstrained Monocular Video/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5b781817e16f1e6e553099d4f6815d134ade27a9 --- /dev/null +++ b/2023/Omnimatte3D_ Associating Objects and Their Effects in Unconstrained Monocular Video/layout.json @@ -0,0 +1,11838 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 134, + 103, + 460, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 103, + 460, + 138 + ], + "spans": [ + { + "bbox": [ + 134, + 103, + 460, + 138 + ], + "type": "text", + "content": "Omnimatte3D: Associating Objects and their Effects in Unconstrained Monocular Video" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 114, + 160, + 220, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 160, + 220, + 174 + ], + "spans": [ + { + "bbox": [ + 114, + 160, + 220, + 174 + ], + "type": "text", + "content": "Mohammed Suhail1,2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 217, + 161, + 296, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 161, + 296, + 189 + ], + "spans": [ + { + "bbox": [ + 217, + 161, + 296, + 189 + ], + "type": "text", + "content": "2 Erika Lu4 Leonid Sigal1,2,3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 318, + 161, + 380, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 161, + 380, + 175 + ], + "spans": [ + { + "bbox": [ + 318, + 161, + 380, + 175 + ], + "type": "text", + "content": "Zhengqi Li" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 404, + 161, + 477, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 404, + 161, + 477, + 175 + ], + "spans": [ + { + "bbox": [ + 404, + 161, + 477, + 175 + ], + "type": "text", + "content": "Noah Snavely4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 324, + 175, + 399, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 175, + 399, + 188 + ], + "spans": [ + { + "bbox": [ + 324, + 175, + 399, + 188 + ], + "type": "text", + "content": "Forrester Cole" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 116, + 190, + 235, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 190, + 235, + 202 + ], + "spans": [ + { + "bbox": [ + 116, + 190, + 235, + 202 + ], + "type": "text", + "content": "1University of British Columbia" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 243, + 190, + 329, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 243, + 190, + 329, + 202 + ], + "spans": [ + { + "bbox": [ + 243, + 190, + 329, + 202 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 243, + 190, + 329, + 202 + ], + "type": "text", + "content": "Vector Institute for AI" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 340, + 191, + 434, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 191, + 434, + 202 + ], + "spans": [ + { + "bbox": [ + 340, + 191, + 434, + 202 + ], + "type": "text", + "content": "3Canada CIFAR AI Chair" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 444, + 191, + 477, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 444, + 191, + 477, + 202 + ], + "spans": [ + { + "bbox": [ + 444, + 191, + 477, + 202 + ], + "type": "text", + "content": "4Google" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 66, + 206, + 153, + 358 + ], + "blocks": [ + { + "bbox": [ + 66, + 206, + 153, + 358 + ], + "lines": [ + { + "bbox": [ + 66, + 206, + 153, + 358 + ], + "spans": [ + { + "bbox": [ + 66, + 206, + 153, + 358 + ], + "type": "image", + "image_path": "03bbd923cf0a354ef93ce235c54b6a0a99f3e51d471523a549ad38ddaedab5a6.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 93, + 360, + 125, + 369 + ], + "lines": [ + { + "bbox": [ + 93, + 360, + 125, + 369 + ], + "spans": [ + { + "bbox": [ + 93, + 360, + 125, + 369 + ], + "type": "text", + "content": "Input RGB" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 153, + 206, + 240, + 358 + ], + "blocks": [ + { + "bbox": [ + 153, + 206, + 240, + 358 + ], + "lines": [ + { + "bbox": [ + 153, + 206, + 240, + 358 + ], + "spans": [ + { + "bbox": [ + 153, + 206, + 240, + 358 + ], + "type": "image", + "image_path": "5dae43cac38bb1b03b87a1cbd6fce6b1709f2284bb6e2fb5cfbff9cc40a2afe8.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 178, + 360, + 214, + 369 + ], + "lines": [ + { + "bbox": [ + 178, + 360, + 214, + 369 + ], + "spans": [ + { + "bbox": [ + 178, + 360, + 214, + 369 + ], + "type": "text", + "content": "Input Masks" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 241, + 206, + 328, + 358 + ], + "blocks": [ + { + "bbox": [ + 241, + 206, + 328, + 358 + ], + "lines": [ + { + "bbox": [ + 241, + 206, + 328, + 358 + ], + "spans": [ + { + "bbox": [ + 241, + 206, + 328, + 358 + ], + "type": "image", + "image_path": "1d5c43233d2a14d35507f132cc7b44a884ea6381a5272c0499935ef815245e23.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 271, + 360, + 307, + 369 + ], + "lines": [ + { + "bbox": [ + 271, + 360, + 307, + 369 + ], + "spans": [ + { + "bbox": [ + 271, + 360, + 307, + 369 + ], + "type": "text", + "content": "Input Depth" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 373, + 546, + 407 + ], + "lines": [ + { + "bbox": [ + 46, + 373, + 546, + 407 + ], + "spans": [ + { + "bbox": [ + 46, + 373, + 546, + 407 + ], + "type": "text", + "content": "Figure 1. Layer decomposition under strong camera parallax. Given an input video with unconstrained camera motion and approximate object masks and depth (left), our method estimates a layered representation composed of a background layer and object layers containing the subjects of interest and their associated effects (e.g. shadows). Results of combining object layers and background are shown on right." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 329, + 280, + 349, + 293 + ], + "blocks": [ + { + "bbox": [ + 329, + 280, + 349, + 293 + ], + "lines": [ + { + "bbox": [ + 329, + 280, + 349, + 293 + ], + "spans": [ + { + "bbox": [ + 329, + 280, + 349, + 293 + ], + "type": "image", + "image_path": "f44f80334f95e168cccff91f53db06d8f946618a8a5becc283c8a00d2e80e312.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 350, + 206, + 437, + 358 + ], + "blocks": [ + { + "bbox": [ + 350, + 206, + 437, + 358 + ], + "lines": [ + { + "bbox": [ + 350, + 206, + 437, + 358 + ], + "spans": [ + { + "bbox": [ + 350, + 206, + 437, + 358 + ], + "type": "image", + "image_path": "e8374b32729bcb5ef8f552f80b4cce507e6056647ad005b27a5b7efac74d6f89.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 361, + 360, + 426, + 369 + ], + "lines": [ + { + "bbox": [ + 361, + 360, + 426, + 369 + ], + "spans": [ + { + "bbox": [ + 361, + 360, + 426, + 369 + ], + "type": "text", + "content": "Layer 1 + Background" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 438, + 206, + 525, + 358 + ], + "blocks": [ + { + "bbox": [ + 438, + 206, + 525, + 358 + ], + "lines": [ + { + "bbox": [ + 438, + 206, + 525, + 358 + ], + "spans": [ + { + "bbox": [ + 438, + 206, + 525, + 358 + ], + "type": "image", + "image_path": "b540f14db5bb39921df5bb5dc2daddc55588d46af2194e9130a6bef22bc9cd1f.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 449, + 360, + 514, + 369 + ], + "lines": [ + { + "bbox": [ + 449, + 360, + 514, + 369 + ], + "spans": [ + { + "bbox": [ + 449, + 360, + 514, + 369 + ], + "type": "text", + "content": "Layer 2 + Background" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "bbox": [ + 143, + 416, + 192, + 428 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 416, + 192, + 428 + ], + "spans": [ + { + "bbox": [ + 143, + 416, + 192, + 428 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 45, + 441, + 290, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 441, + 290, + 646 + ], + "spans": [ + { + "bbox": [ + 45, + 441, + 290, + 646 + ], + "type": "text", + "content": "We propose a method to decompose a video into a background and a set of foreground layers, where the background captures stationary elements while the foreground layers capture moving objects along with their associated effects (e.g. shadows and reflections). Our approach is designed for unconstrained monocular videos, with an arbitrary camera and object motion. Prior work that tackles this problem assumes that the video can be mapped onto a fixed 2D canvas, severely limiting the possible space of camera motion. Instead, our method applies recent progress in monocular camera pose and depth estimation to create a full, RGBD video layer for the background, along with a video layer for each foreground object. To solve the underconstrained decomposition problem, we propose a new loss formulation based on multi-view consistency. We test our method on challenging videos with complex camera motion and show significant qualitative improvement over current approaches." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 47, + 669, + 127, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 669, + 127, + 681 + ], + "spans": [ + { + "bbox": [ + 47, + 669, + 127, + 681 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 715 + ], + "type": "text", + "content": "Decomposing a video into meaningful layers (as show in Fig. 1) is a long-standing and complex problem [42] that has" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 418, + 547, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 418, + 547, + 585 + ], + "spans": [ + { + "bbox": [ + 304, + 418, + 547, + 585 + ], + "type": "text", + "content": "seen a recent surge in progress with the application of deep neural networks [17, 24, 25, 46]. A challenging variant of layer decomposition is the omnimatte [25] task, which aims to separate an input video into a background and multiple foreground layers, each containing an object of interest along with its correlated effects such as shadows and reflections, thus enabling video editing applications such as object removal, background replacement and retiming. The original work on omnimatte [24, 25] introduced a self-supervised approach for decomposing a video by assuming the background can be unwrapped onto a single static 2D background canvas using homography warping. Following work [17, 46] relaxed the homography restriction, but maintained the necessity of unwrapping onto a 2D canvas." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 304, + 589, + 547, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 589, + 547, + 696 + ], + "spans": [ + { + "bbox": [ + 304, + 589, + 547, + 696 + ], + "type": "text", + "content": "This 2D modelling, however, limits the applicability of these methods to camera motions that have limited or zero parallax; intuitively, only panning camera motions are allowed. If the camera's center of projection moves a significant distance over the video, 2D methods are unable to learn an accurate background and are forced to place the background detail in a foreground layer (Figure 2). Since camera parallax is quite common, 2D methods are severely restricted in practice." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 318, + 701, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 701, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 318, + 701, + 545, + 713 + ], + "type": "text", + "content": "To handle camera parallax, we exploit another recent line" + } + ] + } + ], + "index": 30 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "630" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 71, + 286, + 198 + ], + "blocks": [ + { + "bbox": [ + 52, + 71, + 286, + 198 + ], + "lines": [ + { + "bbox": [ + 52, + 71, + 286, + 198 + ], + "spans": [ + { + "bbox": [ + 52, + 71, + 286, + 198 + ], + "type": "image", + "image_path": "30b7cafafa870b70a4ca6193ce0fdc9d1717074df16336e5bc5abd603a32befc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 208, + 288, + 288 + ], + "lines": [ + { + "bbox": [ + 46, + 208, + 288, + 288 + ], + "spans": [ + { + "bbox": [ + 46, + 208, + 288, + 288 + ], + "type": "text", + "content": "Figure 2. Existing methods fail on scenes with parallax. Top row: Omnimatte [25] fails due to inaccurate homography registration, reconstructing the majority of the video in the foreground layer. Middle row: Neural-Atlas [17] incorrectly places trees in the foreground layer and struggles to inpaint the person region. Bottom row: Our method obtains a clean background and captures the person and their shadow in the foreground layer." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 307, + 288, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 307, + 288, + 488 + ], + "spans": [ + { + "bbox": [ + 46, + 307, + 288, + 488 + ], + "type": "text", + "content": "of work on estimating 3D camera position and depth maps from casually-captured video with moving objects [18,26, 48, 49]. Instead of a 2D layer decomposition, we propose to learn a 3D background model that varies per-frame and includes inpainted RGB and depth in the regions occluded by the foreground. Allowing the background to vary per-frame, rather than assuming a global, static canvas, creates an ill-posed decomposition problem, as variation in each input pixel could be explained by a foreground object, background, or both. To resolve this ambiguity, we enforce that the background varies slowly over time with 3D multi-view constraints using dense depth and camera pose estimates [48]. As shown in Fig. 2, our model can successfully separate the layers obtaining a clean background RGB whilst capturing only the person and their shadow in the foreground layer." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 490, + 288, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 490, + 288, + 586 + ], + "spans": [ + { + "bbox": [ + 46, + 490, + 288, + 586 + ], + "type": "text", + "content": "Contributions. Our main contribution is a novel video decomposition method capable of accurately separating an input video with complex object and camera motion (e.g., parallax) into a background and object layers. To address the ill-posed decomposition problem, we design regularization terms based on multi-view consistency and smoothness priors. The resulting model produces clean decompositions of real-world videos where previous methods fail." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 597, + 134, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 597, + 134, + 609 + ], + "spans": [ + { + "bbox": [ + 47, + 597, + 134, + 609 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 288, + 715 + ], + "type": "text", + "content": "Layered video representations. There is a large body of work on decomposing videos into layers based on appearance and motion cues [4, 8, 16, 20, 36, 42]. Layered video representations have proven useful for a variety of applications, including view synthesis for static scenes [39, 50], reflection removal [1, 2], segmentation [9], game deconstruction [38], and text-based video editing [3]. Recently, Omnimatte [24, 25] aimed to decompose a video into layers" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 547, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 193 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 193 + ], + "type": "text", + "content": "that group objects with their correlated effects, allowing the user to perform editing tasks such as object removal and retiming. However, as discussed in Section 1, these works use a constrained background model which limits their use case to panning camera motions only. Subsequent work [17,46] relaxes this assumption using general image warps, instead of homographies, but still cannot handle significant parallax; moreover, their goal is to learn a per-frame mapping to global sprites to enable applying consistent edits (e.g. style transfer) rather than targeting object removal or video stabilization." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 210, + 548, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 210, + 548, + 378 + ], + "spans": [ + { + "bbox": [ + 304, + 210, + 548, + 378 + ], + "type": "text", + "content": "Monocular video depth. Multi-view stereo based methods [23,40] incorporate motion estimation and multi-view reconstruction to predict depth by combining information from multiple reference frames. Hybrid depth methods combine single-view depth estimates from a pretrained network with geometric reasoning to obtain coherent depth predictions. With depth maps from monocular video as initialization, these methods use correspondence to enforce consistency by either assuming static regions [26] or by modeling the scene flow [49]. Recent methods simultaneously estimate depth maps and camera poses [19, 48], exploiting the depth prior to resolve camera motion ambiguities. We rely on CasualSAM [48] to obtain camera poses and initial depth estimates as their method performs well on casually captured video." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 396, + 547, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 396, + 547, + 529 + ], + "spans": [ + { + "bbox": [ + 304, + 396, + 547, + 529 + ], + "type": "text", + "content": "Video completion and object removal. Patch-based methods for video inpainting [7, 12, 14, 28, 43] complete the missing regions by borrowing pixels from other frames, relying on correspondence estimated by homography or optical flow. Learning-based works have explored the use of 3D convolutions [5, 41], attention mechanisms [21, 30, 47] and flow-based methods [10, 45] to achieve more realistic and consistent inpainting. These methods assume that the entire region to inpaint is specified by an input mask, whereas our method jointly learns where to inpaint the background while separating foreground objects and their correlated effects." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 546, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 547, + 713 + ], + "type": "text", + "content": "NeRF for dynamic scenes. With the recent success of neural fields [29,44] for view synthesis of static scenes, there has been interest in their adoption for dynamic scenes [11,22,27]. These works learn two NeRF models that decompose a scene into a static and dynamic region. However, these methods only support a single foreground layer and do not account for correlated effects such as shadows. ST-NeRF [15] proposed to learn separate NeRFs for each person to obtain an editable neural representation. Their method, however, requires each frame in the video to be captured from 16 different viewpoints simultaneously. Unlike NeRF methods, our method represents the video with multi-layer mesh geometry where object layers include associated effects, allowing editing as well as fast rerendering of the scene." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "631" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 70, + 539, + 222 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 539, + 222 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 539, + 222 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 539, + 222 + ], + "type": "image", + "image_path": "54af525eb386bfcbf3aafd896773ab954519b7222151a588440d3c72b57c61dc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 59, + 224, + 536, + 318 + ], + "blocks": [ + { + "bbox": [ + 59, + 224, + 536, + 318 + ], + "lines": [ + { + "bbox": [ + 59, + 224, + 536, + 318 + ], + "spans": [ + { + "bbox": [ + 59, + 224, + 536, + 318 + ], + "type": "image", + "image_path": "4f825ba3e51340be93023f21cfac849c715535b074ff6b18c4c9d509fd56cc40.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 333, + 546, + 378 + ], + "lines": [ + { + "bbox": [ + 46, + 333, + 546, + 378 + ], + "spans": [ + { + "bbox": [ + 46, + 333, + 546, + 378 + ], + "type": "text", + "content": "Figure 3. Model Overview. Given a frame from the video along with an input mask, disparity and known camera pose, we first extract features by stacking the input and passing them though a 2D UNet architecture. These features are input to the background and foreground networks. The background network predicts the inpainted background RGB and disparity. The foreground networks predict an RGB-A image for each foreground layer. For scenes with multiple layers, the RGB-A image for each layer is predicted using independent networks." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 397, + 103, + 410 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 397, + 103, + 410 + ], + "spans": [ + { + "bbox": [ + 47, + 397, + 103, + 410 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 418, + 290, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 418, + 290, + 659 + ], + "spans": [ + { + "bbox": [ + 46, + 418, + 290, + 659 + ], + "type": "text", + "content": "Given an input video, with known camera poses and initial depth estimates, the goal is to predict a per-frame layer decomposition consisting of a background layer and " + }, + { + "bbox": [ + 46, + 418, + 290, + 659 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 418, + 290, + 659 + ], + "type": "text", + "content": " foreground layers encompassing objects of interest along with their associated effects. To this end, we propose a new method that predicts a layered separation of an input video without relying on a 2D background canvas, thereby allowing for greater flexibility of camera motion. We show an overview of our model in Figure 3. We optimize our model to output layers that reconstruct a single target video subject to a projection consistency loss. For each frame in the video, we estimate a layered representation consisting of background and foreground object layers where the contents of the object layers are conditioned on their corresponding input mask. Post-optimization, depth for foreground objects can be extracted from the input depth to produce a Layered Depth Image (LDI). Unlike prior works that rely on deep priors for inpainting the depth and color of occluded regions in the background, our method relies on multi-view consistency losses to steer the inpainting process." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 670, + 218, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 670, + 218, + 682 + ], + "spans": [ + { + "bbox": [ + 47, + 670, + 218, + 682 + ], + "type": "text", + "content": "3.1. Layer Decomposition Networks" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "content": "We represent a frame at time " + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "content": " in the video as " + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "inline_equation", + "content": "F_{a} = \\left(I_{a},D_{a},\\{M_{a}^{i}\\}_{i = 1}^{N}\\right)" + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "inline_equation", + "content": "I_{a}" + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "content": " is the RGB frame, " + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "inline_equation", + "content": "D_{a}" + }, + { + "bbox": [ + 47, + 689, + 288, + 715 + ], + "type": "text", + "content": " is the" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 398, + 545, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 398, + 545, + 471 + ], + "spans": [ + { + "bbox": [ + 304, + 398, + 545, + 471 + ], + "type": "text", + "content": "disparity, and " + }, + { + "bbox": [ + 304, + 398, + 545, + 471 + ], + "type": "inline_equation", + "content": "\\{M_a^i\\}_{i = 1}^N" + }, + { + "bbox": [ + 304, + 398, + 545, + 471 + ], + "type": "text", + "content": " are the input masks for " + }, + { + "bbox": [ + 304, + 398, + 545, + 471 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 398, + 545, + 471 + ], + "type": "text", + "content": " objects in the video. We denote the camera extrinsics for the frame " + }, + { + "bbox": [ + 304, + 398, + 545, + 471 + ], + "type": "inline_equation", + "content": "F_{a}" + }, + { + "bbox": [ + 304, + 398, + 545, + 471 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 304, + 398, + 545, + 471 + ], + "type": "inline_equation", + "content": "[R_a t_a]" + }, + { + "bbox": [ + 304, + 398, + 545, + 471 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 398, + 545, + 471 + ], + "type": "inline_equation", + "content": "R_{a}" + }, + { + "bbox": [ + 304, + 398, + 545, + 471 + ], + "type": "text", + "content": " is the rotation matrix and " + }, + { + "bbox": [ + 304, + 398, + 545, + 471 + ], + "type": "inline_equation", + "content": "t_a" + }, + { + "bbox": [ + 304, + 398, + 545, + 471 + ], + "type": "text", + "content": " is the translation vector. The camera intrinsics are denoted as " + }, + { + "bbox": [ + 304, + 398, + 545, + 471 + ], + "type": "inline_equation", + "content": "K_{a} = [f_{x},f_{y},c_{x},c_{y}]" + }, + { + "bbox": [ + 304, + 398, + 545, + 471 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 304, + 398, + 545, + 471 + ], + "type": "inline_equation", + "content": "f_{x}" + }, + { + "bbox": [ + 304, + 398, + 545, + 471 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 398, + 545, + 471 + ], + "type": "inline_equation", + "content": "f_{y}" + }, + { + "bbox": [ + 304, + 398, + 545, + 471 + ], + "type": "text", + "content": " are the focal lengths and " + }, + { + "bbox": [ + 304, + 398, + 545, + 471 + ], + "type": "inline_equation", + "content": "c_{x},c_{y}" + }, + { + "bbox": [ + 304, + 398, + 545, + 471 + ], + "type": "text", + "content": " is the principal point." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 471, + 545, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 471, + 545, + 543 + ], + "spans": [ + { + "bbox": [ + 304, + 471, + 545, + 543 + ], + "type": "text", + "content": "For each frame " + }, + { + "bbox": [ + 304, + 471, + 545, + 543 + ], + "type": "inline_equation", + "content": "(F_{a})" + }, + { + "bbox": [ + 304, + 471, + 545, + 543 + ], + "type": "text", + "content": " in the input video, our method first extracts features by passing it to a feature extractor network modeled using a 2D UNet [35] architecture. The input to the feature extractor is constructed by concatenating input masks with the masked frame RGB and disparity along the channel dimension. The masked RGB and disparity are obtained as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 329, + 547, + 545, + 563 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 547, + 545, + 563 + ], + "spans": [ + { + "bbox": [ + 329, + 547, + 545, + 563 + ], + "type": "interline_equation", + "content": "\\tilde {Y} _ {a} = \\left[ Y _ {a} \\odot M _ {a} ^ {b g} \\| Y _ {a} \\odot M _ {a} ^ {1} \\| \\dots \\| Y _ {a} \\odot M _ {a} ^ {N} \\right] \\tag {1}", + "image_path": "5ea7f16d90d3a122ffa5975ea66ec2851e4ea3e3fcee021d046bebfd318f5a09.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 568, + 545, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 568, + 545, + 617 + ], + "spans": [ + { + "bbox": [ + 304, + 568, + 545, + 617 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 568, + 545, + 617 + ], + "type": "inline_equation", + "content": "Y_{a}" + }, + { + "bbox": [ + 304, + 568, + 545, + 617 + ], + "type": "text", + "content": " is the concatenated RGB-D, " + }, + { + "bbox": [ + 304, + 568, + 545, + 617 + ], + "type": "inline_equation", + "content": "M_{a}^{bg} = 1 - \\bigcup_{i=1}^{N} M_{a}^{i}" + }, + { + "bbox": [ + 304, + 568, + 545, + 617 + ], + "type": "text", + "content": " is an approximate mask for the background and " + }, + { + "bbox": [ + 304, + 568, + 545, + 617 + ], + "type": "inline_equation", + "content": "\\| \\cdot" + }, + { + "bbox": [ + 304, + 568, + 545, + 617 + ], + "type": "text", + "content": " is the concatenation operation. These features act as input to both the background and foreground object networks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 618, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 618, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 618, + 547, + 713 + ], + "type": "text", + "content": "The background network (shown in green in Fig. 3) consists of a 2D UNet and a convolutional neural network (CNN) that predicts the inpainted background color " + }, + { + "bbox": [ + 304, + 618, + 547, + 713 + ], + "type": "inline_equation", + "content": "(\\mathcal{C}_a^{bg})" + }, + { + "bbox": [ + 304, + 618, + 547, + 713 + ], + "type": "text", + "content": " and depth " + }, + { + "bbox": [ + 304, + 618, + 547, + 713 + ], + "type": "inline_equation", + "content": "(\\mathcal{D}_a^{bg})" + }, + { + "bbox": [ + 304, + 618, + 547, + 713 + ], + "type": "text", + "content": " respectively. The foreground networks are separate CNNs that predict RGB " + }, + { + "bbox": [ + 304, + 618, + 547, + 713 + ], + "type": "inline_equation", + "content": "(\\mathcal{C}_a^i)" + }, + { + "bbox": [ + 304, + 618, + 547, + 713 + ], + "type": "text", + "content": " and alpha " + }, + { + "bbox": [ + 304, + 618, + 547, + 713 + ], + "type": "inline_equation", + "content": "(\\mathcal{A}_a^i)" + }, + { + "bbox": [ + 304, + 618, + 547, + 713 + ], + "type": "text", + "content": " images for each foreground layer independently. These predicted outputs, along with the input disparity, are combined to form an LDI (see supplement for details)." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "632" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 133 + ], + "type": "text", + "content": "The feature extractor and background RGB UNets consist of five down-sampling and up-sampling layers with three residual blocks at every resolution. The background disparity and foreground layer prediction networks are comprised of single convolutional layers." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 138, + 187, + 152 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 138, + 187, + 152 + ], + "spans": [ + { + "bbox": [ + 47, + 138, + 187, + 152 + ], + "type": "text", + "content": "3.2. Differentiable Rendering" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 156, + 287, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 156, + 287, + 216 + ], + "spans": [ + { + "bbox": [ + 47, + 156, + 287, + 216 + ], + "type": "text", + "content": "In addition to a reconstruction loss, our method relies on a projection consistency term (described in Section 3.3) that requires differentiable projections and renderings of the predicted background layer to different time steps. We achieve this using a differentiable rasterization framework [6]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 217, + 288, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 217, + 288, + 300 + ], + "spans": [ + { + "bbox": [ + 47, + 217, + 288, + 300 + ], + "type": "text", + "content": "To project the predicted background layer from a source time step " + }, + { + "bbox": [ + 47, + 217, + 288, + 300 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 47, + 217, + 288, + 300 + ], + "type": "text", + "content": " to target time step " + }, + { + "bbox": [ + 47, + 217, + 288, + 300 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 47, + 217, + 288, + 300 + ], + "type": "text", + "content": ", we first construct a mesh " + }, + { + "bbox": [ + 47, + 217, + 288, + 300 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_a = (\\mathcal{V}_a, \\mathcal{F}_a)" + }, + { + "bbox": [ + 47, + 217, + 288, + 300 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 47, + 217, + 288, + 300 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_a" + }, + { + "bbox": [ + 47, + 217, + 288, + 300 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 217, + 288, + 300 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_a" + }, + { + "bbox": [ + 47, + 217, + 288, + 300 + ], + "type": "text", + "content": " are the mesh vertices and faces. Each vertex in the mesh corresponds to an image pixel. To obtain the vertex coordinate for a pixel, we unproject its homogeneous coordinates " + }, + { + "bbox": [ + 47, + 217, + 288, + 300 + ], + "type": "inline_equation", + "content": "x = [i,j,1] \\in \\mathbf{RP}^2" + }, + { + "bbox": [ + 47, + 217, + 288, + 300 + ], + "type": "text", + "content": " using the predicted disparity as" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 95, + 305, + 287, + 332 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 305, + 287, + 332 + ], + "spans": [ + { + "bbox": [ + 95, + 305, + 287, + 332 + ], + "type": "interline_equation", + "content": "V _ {a} ^ {x} = d _ {i j} \\left(\\frac {\\left(i - c _ {x}\\right)}{f _ {x}}, \\frac {\\left(j - c _ {y}\\right)}{f _ {y}}, 1\\right) \\tag {2}", + "image_path": "a9c03c62a86e2530c93c7ce77dffaf3ff7d7837bbb40dcc2a98b10533af88e73.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 337, + 289, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 337, + 289, + 422 + ], + "spans": [ + { + "bbox": [ + 47, + 337, + 289, + 422 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 337, + 289, + 422 + ], + "type": "inline_equation", + "content": "d_{ij}" + }, + { + "bbox": [ + 47, + 337, + 289, + 422 + ], + "type": "text", + "content": " is the depth at pixel " + }, + { + "bbox": [ + 47, + 337, + 289, + 422 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 47, + 337, + 289, + 422 + ], + "type": "text", + "content": ". The attributes for the vertices are composed of the predicted RGB, an alpha masking consisting of all ones, and the pixel coordinates. The faces for the mesh are constructed by adding an edge between vertices of neighboring pixels (bottom right in Fig. 3). The view-projection matrix to render the predicted background from a target camera is computed as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 92, + 426, + 287, + 453 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 426, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 92, + 426, + 287, + 453 + ], + "type": "interline_equation", + "content": "H _ {a \\rightarrow b} = P _ {a} \\left[\\begin{array}{c c}R _ {a} ^ {T} R _ {b}&R _ {a} ^ {T} \\left(t _ {b} - t _ {a}\\right)\\\\\\mathbf {0}&1\\end{array}\\right] \\tag {3}", + "image_path": "c302ac060ac42f625fcd274cd992fbff45c8f42bed2743e42c58c3613274421d.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 457, + 287, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 457, + 287, + 517 + ], + "spans": [ + { + "bbox": [ + 47, + 457, + 287, + 517 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 457, + 287, + 517 + ], + "type": "inline_equation", + "content": "P_{a}" + }, + { + "bbox": [ + 47, + 457, + 287, + 517 + ], + "type": "text", + "content": " is the perspective matrix that maps camera-space to clip space. The view-projection matrix along with the estimated mesh is passed through a differentiable rasterizer that returns an RGB-A " + }, + { + "bbox": [ + 47, + 457, + 287, + 517 + ], + "type": "inline_equation", + "content": "(\\mathcal{C}_{a\\to b}^{bg},\\mathcal{A}_{a\\to b}^{bg})" + }, + { + "bbox": [ + 47, + 457, + 287, + 517 + ], + "type": "text", + "content": " image corresponding to the projection of the background to the target time step." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 523, + 100, + 534 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 523, + 100, + 534 + ], + "spans": [ + { + "bbox": [ + 47, + 523, + 100, + 534 + ], + "type": "text", + "content": "3.3. Losses" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 59, + 541, + 203, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 541, + 203, + 552 + ], + "spans": [ + { + "bbox": [ + 59, + 541, + 203, + 552 + ], + "type": "text", + "content": "The total loss to train our model is:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 52, + 559, + 287, + 573 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 559, + 287, + 573 + ], + "spans": [ + { + "bbox": [ + 52, + 559, + 287, + 573 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\lambda_ {1} \\mathcal {L} _ {\\text {r e c o n}} + \\lambda_ {2} \\mathcal {L} _ {\\text {p r o j}} + \\lambda_ {3} \\mathcal {L} _ {\\text {m a s k}} + \\lambda_ {4} \\mathcal {L} _ {\\text {d i s p}} + \\lambda_ {5} \\mathcal {L} _ {\\text {r e g}} \\tag {4}", + "image_path": "6cff436a2f4a53f04316f4758829a4185690b454d66e49cc73d949c4a8274d90.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 578, + 287, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 578, + 287, + 603 + ], + "spans": [ + { + "bbox": [ + 47, + 578, + 287, + 603 + ], + "type": "text", + "content": "We describe each loss term in the following sections. Please see the supplement for values of the scalar weights " + }, + { + "bbox": [ + 47, + 578, + 287, + 603 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 47, + 578, + 287, + 603 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 616, + 167, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 616, + 167, + 628 + ], + "spans": [ + { + "bbox": [ + 47, + 616, + 167, + 628 + ], + "type": "text", + "content": "3.3.1 Reconstruction Loss" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 635, + 287, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 635, + 287, + 694 + ], + "spans": [ + { + "bbox": [ + 47, + 635, + 287, + 694 + ], + "type": "text", + "content": "For a source frame " + }, + { + "bbox": [ + 47, + 635, + 287, + 694 + ], + "type": "inline_equation", + "content": "F_{a}" + }, + { + "bbox": [ + 47, + 635, + 287, + 694 + ], + "type": "text", + "content": " we reconstruct the image from the predicted background RGB and foreground RGBA layers using back-to-front compositing [33] (bottom left in Fig. 3). During training, we minimize the " + }, + { + "bbox": [ + 47, + 635, + 287, + 694 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_2" + }, + { + "bbox": [ + 47, + 635, + 287, + 694 + ], + "type": "text", + "content": " loss between the original RGB image and the reconstruction." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 85, + 700, + 287, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 700, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 85, + 700, + 287, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r e c o n}} = \\left\\| \\left(I _ {a} - f \\left(\\mathcal {C} _ {a} ^ {b g}, \\left\\{\\mathcal {C} _ {a} ^ {i}, \\mathcal {A} _ {a} ^ {1} \\right\\} _ {i = 1} ^ {N}\\right) \\right\\| _ {2} \\right. \\tag {5}", + "image_path": "85301d690fa157f9f7f34c059dd9b8930bf3d8f75aafdb4abd45ca69593da282.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 72, + 531, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 531, + 84 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 531, + 84 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 72, + 531, + 84 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 305, + 72, + 531, + 84 + ], + "type": "text", + "content": " is the back-to-front over-compositing function." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 84, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 84, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 305, + 84, + 545, + 144 + ], + "type": "text", + "content": "A large number of layer decompositions can reproduce the original frame and minimize the reconstruction loss. To guide the foreground layers to be semantically meaningful, we incorporate an " + }, + { + "bbox": [ + 305, + 84, + 545, + 144 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_1" + }, + { + "bbox": [ + 305, + 84, + 545, + 144 + ], + "type": "text", + "content": " loss between the predicted foreground alphas and the input masks." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 365, + 152, + 545, + 184 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 152, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 365, + 152, + 545, + 184 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {m a s k}} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\| M _ {a} ^ {i} - \\mathcal {A} _ {a} ^ {i} \\| _ {1} \\tag {6}", + "image_path": "7ba3453e3180cd8a4d1cd71e7af3c4732edfdde19f440c764eadab8de81430cd.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 191, + 545, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 191, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 305, + 191, + 545, + 227 + ], + "type": "text", + "content": "Since we want the foreground alphas to be able to capture effects beyond the input mask, the weight for this loss is decayed as described in Sec. 3.4." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 242, + 459, + 254 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 242, + 459, + 254 + ], + "spans": [ + { + "bbox": [ + 306, + 242, + 459, + 254 + ], + "type": "text", + "content": "3.3.2 Projection Consistency Loss" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 260, + 547, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 260, + 547, + 415 + ], + "spans": [ + { + "bbox": [ + 305, + 260, + 547, + 415 + ], + "type": "text", + "content": "The reconstruction loss on the over-composed image does not provide any gradients for the occluded region in the background. Prior works overcome this challenge by learning a background model that maps from a static 2D canvas. However, since our model predicts the background for each frame independently, we require additional regularization to prevent random artifacts or foreground details from appearing in occluded regions. We thus introduce a projection consistency loss on the background prediction. Each batch in our input is composed of three frames from the video. We denote these frames as " + }, + { + "bbox": [ + 305, + 260, + 547, + 415 + ], + "type": "inline_equation", + "content": "F_{a}" + }, + { + "bbox": [ + 305, + 260, + 547, + 415 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 260, + 547, + 415 + ], + "type": "inline_equation", + "content": "F_{b}" + }, + { + "bbox": [ + 305, + 260, + 547, + 415 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 260, + 547, + 415 + ], + "type": "inline_equation", + "content": "F_{c}" + }, + { + "bbox": [ + 305, + 260, + 547, + 415 + ], + "type": "text", + "content": ". The frames " + }, + { + "bbox": [ + 305, + 260, + 547, + 415 + ], + "type": "inline_equation", + "content": "F_{a}" + }, + { + "bbox": [ + 305, + 260, + 547, + 415 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 260, + 547, + 415 + ], + "type": "inline_equation", + "content": "F_{b}" + }, + { + "bbox": [ + 305, + 260, + 547, + 415 + ], + "type": "text", + "content": " act as source images for the target frame " + }, + { + "bbox": [ + 305, + 260, + 547, + 415 + ], + "type": "inline_equation", + "content": "F_{c}" + }, + { + "bbox": [ + 305, + 260, + 547, + 415 + ], + "type": "text", + "content": ". The criteria used for selecting the frames is described in Section 3.4." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 416, + 545, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 416, + 545, + 499 + ], + "spans": [ + { + "bbox": [ + 305, + 416, + 545, + 499 + ], + "type": "text", + "content": "To estimate the projection consistency loss, we project the predicted background RGBs " + }, + { + "bbox": [ + 305, + 416, + 545, + 499 + ], + "type": "inline_equation", + "content": "(\\mathcal{C}_a^{bg}" + }, + { + "bbox": [ + 305, + 416, + 545, + 499 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 416, + 545, + 499 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_b^{bg})" + }, + { + "bbox": [ + 305, + 416, + 545, + 499 + ], + "type": "text", + "content": " for the two source frames, " + }, + { + "bbox": [ + 305, + 416, + 545, + 499 + ], + "type": "inline_equation", + "content": "F_{a}" + }, + { + "bbox": [ + 305, + 416, + 545, + 499 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 416, + 545, + 499 + ], + "type": "inline_equation", + "content": "F_{b}" + }, + { + "bbox": [ + 305, + 416, + 545, + 499 + ], + "type": "text", + "content": ", using the method described in Sec. 3.2. We use the predicted disparities " + }, + { + "bbox": [ + 305, + 416, + 545, + 499 + ], + "type": "inline_equation", + "content": "(\\mathcal{D}_a^{bg}" + }, + { + "bbox": [ + 305, + 416, + 545, + 499 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 416, + 545, + 499 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_b^{bg})" + }, + { + "bbox": [ + 305, + 416, + 545, + 499 + ], + "type": "text", + "content": " to construct the mesh and set the alpha values to one for all pixels. We then apply an " + }, + { + "bbox": [ + 305, + 416, + 545, + 499 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_2" + }, + { + "bbox": [ + 305, + 416, + 545, + 499 + ], + "type": "text", + "content": " loss between the projected source backgrounds and the predicted target background." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 336, + 506, + 545, + 537 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 506, + 545, + 537 + ], + "spans": [ + { + "bbox": [ + 336, + 506, + 545, + 537 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {p n j} ^ {r g b} = \\frac {1}{2} \\sum_ {k \\in \\{a, b \\}} \\| \\mathcal {A} _ {k \\rightarrow c} ^ {b g} \\odot \\left(\\mathcal {C} _ {k \\rightarrow c} ^ {b g} - \\mathcal {C} _ {c} ^ {b g}\\right) \\| _ {2} \\tag {7}", + "image_path": "1f586eb161bb40b6ab7f6b8487a2b1380a07a35654157f93e76b75898c32ff97.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 305, + 544, + 545, + 580 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 544, + 545, + 580 + ], + "spans": [ + { + "bbox": [ + 305, + 544, + 545, + 580 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 544, + 545, + 580 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 305, + 544, + 545, + 580 + ], + "type": "text", + "content": " denotes element wise multiplication. We mask the loss using the projected alphas " + }, + { + "bbox": [ + 305, + 544, + 545, + 580 + ], + "type": "inline_equation", + "content": "(\\mathcal{A}_{k\\to c}^{bg})" + }, + { + "bbox": [ + 305, + 544, + 545, + 580 + ], + "type": "text", + "content": " to prevent noisy gradients from the boundaries of projection." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 305, + 580, + 545, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 580, + 545, + 604 + ], + "spans": [ + { + "bbox": [ + 305, + 580, + 545, + 604 + ], + "type": "text", + "content": "Similarly, to encourage consistent inpainting of depth we apply a similar projection loss," + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 331, + 611, + 545, + 642 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 611, + 545, + 642 + ], + "spans": [ + { + "bbox": [ + 331, + 611, + 545, + 642 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {p r o j} ^ {c o o r d} = \\frac {1}{2} \\sum_ {k \\in \\{a, b \\}} \\| \\mathcal {A} _ {k \\rightarrow c} ^ {b g} \\odot \\left(\\mathcal {X} _ {k \\rightarrow c} ^ {b g} - \\mathcal {X} _ {c} ^ {b g}\\right) \\| _ {2} \\tag {8}", + "image_path": "0cea751803144731ffcfbe87b44542b92f525d529a6dd5e338d761b01f000278.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 305, + 649, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 649, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 305, + 649, + 547, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 649, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_c" + }, + { + "bbox": [ + 305, + 649, + 547, + 715 + ], + "type": "text", + "content": " are the world coordinates for pixels in the frame " + }, + { + "bbox": [ + 305, + 649, + 547, + 715 + ], + "type": "inline_equation", + "content": "F_{c}" + }, + { + "bbox": [ + 305, + 649, + 547, + 715 + ], + "type": "text", + "content": " obtained using the predicted background disparity " + }, + { + "bbox": [ + 305, + 649, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_c^{bg}" + }, + { + "bbox": [ + 305, + 649, + 547, + 715 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 305, + 649, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_{k\\to c}^{bg}" + }, + { + "bbox": [ + 305, + 649, + 547, + 715 + ], + "type": "text", + "content": " are the projections of the world coordinate from frame " + }, + { + "bbox": [ + 305, + 649, + 547, + 715 + ], + "type": "inline_equation", + "content": "F_{k}" + }, + { + "bbox": [ + 305, + 649, + 547, + 715 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 305, + 649, + 547, + 715 + ], + "type": "inline_equation", + "content": "F_{c}" + }, + { + "bbox": [ + 305, + 649, + 547, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 649, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_{k\\to c}^{bg}" + }, + { + "bbox": [ + 305, + 649, + 547, + 715 + ], + "type": "text", + "content": " are the projected alphas. The projection loss " + }, + { + "bbox": [ + 305, + 649, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{proj}" + }, + { + "bbox": [ + 305, + 649, + 547, + 715 + ], + "type": "text", + "content": " is the sum of " + }, + { + "bbox": [ + 305, + 649, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{proj}^{rgb}" + }, + { + "bbox": [ + 305, + 649, + 547, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 305, + 649, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{proj}^{coord}" + }, + { + "bbox": [ + 305, + 649, + 547, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "633" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 172, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 172, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 172, + 83 + ], + "type": "text", + "content": "3.3.3 Alpha Regularization" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 91, + 288, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 91, + 288, + 199 + ], + "spans": [ + { + "bbox": [ + 46, + 91, + 288, + 199 + ], + "type": "text", + "content": "While the projection loss encourages the background predictions to be consistent across frames, it can also lead the model toward undesirable or degenerate solutions. For example, a model that predicts a constant color at every pixel for every frame will yield a projection loss of zero and essentially force the background to be captured in foreground layers to minimize the reconstruction loss. To prevent such solutions, we regularize the foreground alphas with an " + }, + { + "bbox": [ + 46, + 91, + 288, + 199 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_1" + }, + { + "bbox": [ + 46, + 91, + 288, + 199 + ], + "type": "text", + "content": " and approximate-" + }, + { + "bbox": [ + 46, + 91, + 288, + 199 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_0" + }, + { + "bbox": [ + 46, + 91, + 288, + 199 + ], + "type": "text", + "content": " term as in [25]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 72, + 206, + 288, + 239 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 206, + 288, + 239 + ], + "spans": [ + { + "bbox": [ + 72, + 206, + 288, + 239 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s p} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\| \\mathcal {A} _ {a} ^ {i} \\| _ {1} + \\gamma \\| 2 \\cdot \\sigma (5 \\cdot \\mathcal {A} _ {a} ^ {i}) - 1 \\| _ {1} \\tag {9}", + "image_path": "87088bae9a19b0241afda6059ff4c4193088e12203aa2a607a396e19f2da9f24.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 247, + 287, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 247, + 287, + 269 + ], + "spans": [ + { + "bbox": [ + 46, + 247, + 287, + 269 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 247, + 287, + 269 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 46, + 247, + 287, + 269 + ], + "type": "text", + "content": " is the sigmoid function and " + }, + { + "bbox": [ + 46, + 247, + 287, + 269 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 46, + 247, + 287, + 269 + ], + "type": "text", + "content": " control the relative weight between the two loss terms." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 270, + 288, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 270, + 288, + 342 + ], + "spans": [ + { + "bbox": [ + 46, + 270, + 288, + 342 + ], + "type": "text", + "content": "Additionally, we regularize the foreground alphas with a smoothness term to discourage them from capturing sharp details in the background which would otherwise arise due to the effects of depth imprecision or splatting in the rasterizer. The smoothness loss " + }, + { + "bbox": [ + 46, + 270, + 288, + 342 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_s^{\\text{alpha}})" + }, + { + "bbox": [ + 46, + 270, + 288, + 342 + ], + "type": "text", + "content": " is computed using an " + }, + { + "bbox": [ + 46, + 270, + 288, + 342 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_1" + }, + { + "bbox": [ + 46, + 270, + 288, + 342 + ], + "type": "text", + "content": " loss on the second order gradients of the predicted alphas." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 356, + 143, + 368 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 356, + 143, + 368 + ], + "spans": [ + { + "bbox": [ + 47, + 356, + 143, + 368 + ], + "type": "text", + "content": "3.3.4 Disparity Loss" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 375, + 287, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 375, + 287, + 399 + ], + "spans": [ + { + "bbox": [ + 46, + 375, + 287, + 399 + ], + "type": "text", + "content": "To distill the available depth information into our model, we apply an " + }, + { + "bbox": [ + 46, + 375, + 287, + 399 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_2" + }, + { + "bbox": [ + 46, + 375, + 287, + 399 + ], + "type": "text", + "content": " loss on the predicted background disparity." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 102, + 406, + 287, + 421 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 406, + 287, + 421 + ], + "spans": [ + { + "bbox": [ + 102, + 406, + 287, + 421 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {d i s p}} = \\left\\| M _ {a} ^ {b g} \\odot \\left(D _ {a} - D _ {a} ^ {b g}\\right) \\right\\| _ {2} \\tag {10}", + "image_path": "86518e37b1d03dd66d249db1e2f8d198398b4e49da695e799e9ff5a8c68dbe08.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 429, + 287, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 429, + 287, + 503 + ], + "spans": [ + { + "bbox": [ + 47, + 429, + 287, + 503 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 429, + 287, + 503 + ], + "type": "inline_equation", + "content": "M_{a}^{bg} = 1 - \\bigcup_{i=1}^{N} M_{a}^{i}" + }, + { + "bbox": [ + 47, + 429, + 287, + 503 + ], + "type": "text", + "content": " is the mask indicating the approximate unoccluded background region. Additionally, to encourage smoothness in the depth inpainting we add an " + }, + { + "bbox": [ + 47, + 429, + 287, + 503 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 47, + 429, + 287, + 503 + ], + "type": "text", + "content": " loss " + }, + { + "bbox": [ + 47, + 429, + 287, + 503 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{s}^{disp})" + }, + { + "bbox": [ + 47, + 429, + 287, + 503 + ], + "type": "text", + "content": " on the second-order gradients of the predicted disparity similar to " + }, + { + "bbox": [ + 47, + 429, + 287, + 503 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{s}^{alpha}" + }, + { + "bbox": [ + 47, + 429, + 287, + 503 + ], + "type": "text", + "content": ". The sum of the losses " + }, + { + "bbox": [ + 47, + 429, + 287, + 503 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{sp}, \\mathcal{L}_{s}^{alpha}" + }, + { + "bbox": [ + 47, + 429, + 287, + 503 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 47, + 429, + 287, + 503 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{s}^{disp}" + }, + { + "bbox": [ + 47, + 429, + 287, + 503 + ], + "type": "text", + "content": " forms the regularization loss " + }, + { + "bbox": [ + 47, + 429, + 287, + 503 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{reg}" + }, + { + "bbox": [ + 47, + 429, + 287, + 503 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 509, + 146, + 520 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 509, + 146, + 520 + ], + "spans": [ + { + "bbox": [ + 47, + 509, + 146, + 520 + ], + "type": "text", + "content": "3.4. Frame Selection" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 527, + 288, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 527, + 288, + 682 + ], + "spans": [ + { + "bbox": [ + 46, + 527, + 288, + 682 + ], + "type": "text", + "content": "The prediction/inpainting of the background RGB and disparity is guided by the projection consistency loss. Intuitively, the source and target frames should be such that regions that are occluded in one are visible in the other whilst having a large overlap in the static regions. To identify such frame pairs we use an approximate frame overlap metric. For two frames " + }, + { + "bbox": [ + 46, + 527, + 288, + 682 + ], + "type": "inline_equation", + "content": "F_{a}" + }, + { + "bbox": [ + 46, + 527, + 288, + 682 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 527, + 288, + 682 + ], + "type": "inline_equation", + "content": "F_{b}" + }, + { + "bbox": [ + 46, + 527, + 288, + 682 + ], + "type": "text", + "content": " we estimate the metric by projecting a mesh from " + }, + { + "bbox": [ + 46, + 527, + 288, + 682 + ], + "type": "inline_equation", + "content": "F_{a}" + }, + { + "bbox": [ + 46, + 527, + 288, + 682 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 527, + 288, + 682 + ], + "type": "inline_equation", + "content": "F_{b}" + }, + { + "bbox": [ + 46, + 527, + 288, + 682 + ], + "type": "text", + "content": ". The attributes of the mesh vertices is set using " + }, + { + "bbox": [ + 46, + 527, + 288, + 682 + ], + "type": "inline_equation", + "content": "B_{a} = \\cup_{i} M_{a}^{i}" + }, + { + "bbox": [ + 46, + 527, + 288, + 682 + ], + "type": "text", + "content": " where for a pixel " + }, + { + "bbox": [ + 46, + 527, + 288, + 682 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 46, + 527, + 288, + 682 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 527, + 288, + 682 + ], + "type": "inline_equation", + "content": "B(i,j)" + }, + { + "bbox": [ + 46, + 527, + 288, + 682 + ], + "type": "text", + "content": " indicates if the pixel is occupied by a foreground or background as dictated by the input mask. The alpha value for the rasterizer input " + }, + { + "bbox": [ + 46, + 527, + 288, + 682 + ], + "type": "inline_equation", + "content": "A_{a}^{bg}" + }, + { + "bbox": [ + 46, + 527, + 288, + 682 + ], + "type": "text", + "content": " are set to all ones. The frame overlap metric is then estimated as" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 89, + 689, + 288, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 689, + 288, + 717 + ], + "spans": [ + { + "bbox": [ + 89, + 689, + 288, + 717 + ], + "type": "interline_equation", + "content": "O _ {a \\rightarrow b} = \\frac {\\left\\| A _ {a \\rightarrow b} \\right\\|}{H \\cdot W} \\cdot \\frac {\\left\\| B _ {c} - A _ {a \\rightarrow b} \\right\\|}{\\left\\| B _ {c} \\right\\| + \\left\\| A _ {a \\rightarrow b} \\right\\|} \\tag {11}", + "image_path": "1a616de2d3fe47df237e313b88c7346b9143130641893cd39f3a2fce6688dafc.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "text", + "content": "For each frame in the video, we rank every other based on the frame overlap metric. During training, for a source frame " + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "F_{a}" + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "text", + "content": ", we choose the another frame " + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "F_{b}" + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "text", + "content": " randomly from the top 10 frames that maximize the overlap metric. The target frame " + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "inline_equation", + "content": "F_{c}" + }, + { + "bbox": [ + 305, + 72, + 545, + 133 + ], + "type": "text", + "content": " is chosen to be midway between the two source frames." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 139, + 400, + 150 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 139, + 400, + 150 + ], + "spans": [ + { + "bbox": [ + 306, + 139, + 400, + 150 + ], + "type": "text", + "content": "3.5. Detail Transfer" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 157, + 545, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 157, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 304, + 157, + 545, + 277 + ], + "type": "text", + "content": "The optimized layers " + }, + { + "bbox": [ + 304, + 157, + 545, + 277 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_k^{bg}" + }, + { + "bbox": [ + 304, + 157, + 545, + 277 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 157, + 545, + 277 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_k^i" + }, + { + "bbox": [ + 304, + 157, + 545, + 277 + ], + "type": "text", + "content": " may miss high-frequency details present in the input video. Detail may be directly recovered from the input video using the transfer approach of Layered Neural Rendering [24], where the transmittance defined by the alpha maps " + }, + { + "bbox": [ + 304, + 157, + 545, + 277 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_k^i" + }, + { + "bbox": [ + 304, + 157, + 545, + 277 + ], + "type": "text", + "content": " controls the amount of detail transferred to each layer. In addition, the background depth map allows reprojection of detail from nearby frames, similar to [34], allowing some detail to be added in regions originally occluded by foreground elements. Figure 10 shows the effect of detail transfer on object removal." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 288, + 388, + 301 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 288, + 388, + 301 + ], + "spans": [ + { + "bbox": [ + 305, + 288, + 388, + 301 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 307, + 405, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 307, + 405, + 319 + ], + "spans": [ + { + "bbox": [ + 306, + 307, + 405, + 319 + ], + "type": "text", + "content": "4.1. Training Details" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 326, + 547, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 326, + 547, + 373 + ], + "spans": [ + { + "bbox": [ + 305, + 326, + 547, + 373 + ], + "type": "text", + "content": "Data preprocessing. We perform our experiments on videos from the DAVIS [32] dataset. We use CasualSAM [48] to obtain camera poses and initial depth estimates. The input masks are estimated using MaskRCNN [13]." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 376, + 545, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 376, + 545, + 484 + ], + "spans": [ + { + "bbox": [ + 304, + 376, + 545, + 484 + ], + "type": "text", + "content": "Training Schedule. During the warm-up stage, the model is trained using the disparity " + }, + { + "bbox": [ + 304, + 376, + 545, + 484 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{disp})" + }, + { + "bbox": [ + 304, + 376, + 545, + 484 + ], + "type": "text", + "content": " and mask losses " + }, + { + "bbox": [ + 304, + 376, + 545, + 484 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{mask})" + }, + { + "bbox": [ + 304, + 376, + 545, + 484 + ], + "type": "text", + "content": " along with the smoothness regularizers, for initialization purposes. After the warm-up stage, the weights for these loss terms are decayed using a cosine schedule to allow the model to inpaint missing depths and include effects such as shadows in the foreground alpha layers. All other loss terms use a linear schedule starting from 0 during the warm-up stage and remain constant for the remainder of training." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 491, + 465, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 491, + 465, + 504 + ], + "spans": [ + { + "bbox": [ + 305, + 491, + 465, + 504 + ], + "type": "text", + "content": "4.2. Layer Decomposition Results" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 510, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 510, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 510, + 547, + 713 + ], + "type": "text", + "content": "Figure 4 shows examples of layer decomposition results for various real-world videos from the DAVIS [31] dataset. We show comparisons against Omnimatte [25] and NeuralAtlas [17] and observe clear improvements over these methods. On the Scooter-gray scene, the background predictions for the baselines are severely distorted due to complex camera motion. The incorrect background forces these methods to compensate by reconstructing the background details in the foreground layer. Our method is able to generate an accurate background and predict a much cleaner foreground layer, which includes the shadow. Similarly, on the Snowboard scene, inaccurate homography registration leads to incorrect background prediction by Omnimatte, particularly in the rocks in the upper right corner. Neural-Atlas shows some improvement in the background predictions but loses the rock texture and is unable to capture the shadow in the foreground layer. Our method successfully captures the shadow in the" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "type": "text", + "content": "634" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 86, + 98, + 192, + 152 + ], + "blocks": [ + { + "bbox": [ + 113, + 82, + 165, + 92 + ], + "lines": [ + { + "bbox": [ + 113, + 82, + 165, + 92 + ], + "spans": [ + { + "bbox": [ + 113, + 82, + 165, + 92 + ], + "type": "text", + "content": "Snowboard" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 86, + 98, + 192, + 152 + ], + "lines": [ + { + "bbox": [ + 86, + 98, + 192, + 152 + ], + "spans": [ + { + "bbox": [ + 86, + 98, + 192, + 152 + ], + "type": "image", + "image_path": "20adf8f5fba2a5c5a79dbfe45ff6974ee9a96edd59dbca6968eb972bc5f4ae2e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 86, + 153, + 191, + 206 + ], + "blocks": [ + { + "bbox": [ + 86, + 153, + 191, + 206 + ], + "lines": [ + { + "bbox": [ + 86, + 153, + 191, + 206 + ], + "spans": [ + { + "bbox": [ + 86, + 153, + 191, + 206 + ], + "type": "image", + "image_path": "ed86914aadaa1209659d7dd4b48f40d8ffe86d1b98173febdccda68424ddb3c6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 86, + 270, + 191, + 325 + ], + "blocks": [ + { + "bbox": [ + 110, + 255, + 168, + 267 + ], + "lines": [ + { + "bbox": [ + 110, + 255, + 168, + 267 + ], + "spans": [ + { + "bbox": [ + 110, + 255, + 168, + 267 + ], + "type": "text", + "content": "Scooter-gray" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 86, + 270, + 191, + 325 + ], + "lines": [ + { + "bbox": [ + 86, + 270, + 191, + 325 + ], + "spans": [ + { + "bbox": [ + 86, + 270, + 191, + 325 + ], + "type": "image", + "image_path": "5fb735d45ad0dd6a681dbdaf8eb3a628d9883cab994b3f070b9b3242d4d00660.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 86, + 326, + 191, + 378 + ], + "blocks": [ + { + "bbox": [ + 86, + 326, + 191, + 378 + ], + "lines": [ + { + "bbox": [ + 86, + 326, + 191, + 378 + ], + "spans": [ + { + "bbox": [ + 86, + 326, + 191, + 378 + ], + "type": "image", + "image_path": "e51a4bc01b214dc17b6484248745823f0596b286aba359b1463f10ef87c83b12.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 110, + 407, + 164, + 427 + ], + "lines": [ + { + "bbox": [ + 110, + 407, + 164, + 427 + ], + "spans": [ + { + "bbox": [ + 110, + 407, + 164, + 427 + ], + "type": "text", + "content": "(a) Input RGB and Mask" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 201, + 70, + 306, + 125 + ], + "blocks": [ + { + "bbox": [ + 201, + 70, + 306, + 125 + ], + "lines": [ + { + "bbox": [ + 201, + 70, + 306, + 125 + ], + "spans": [ + { + "bbox": [ + 201, + 70, + 306, + 125 + ], + "type": "image", + "image_path": "d15a9c773cf47cc85510cc60b16175da84cb626c0748e29cf8059483d4d20d40.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 201, + 125, + 306, + 178 + ], + "blocks": [ + { + "bbox": [ + 201, + 125, + 306, + 178 + ], + "lines": [ + { + "bbox": [ + 201, + 125, + 306, + 178 + ], + "spans": [ + { + "bbox": [ + 201, + 125, + 306, + 178 + ], + "type": "image", + "image_path": "b36869128d160dd787176da9f8b7d946a1a27ee2d714a47f37e2124118f1b705.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 201, + 179, + 306, + 234 + ], + "blocks": [ + { + "bbox": [ + 201, + 179, + 306, + 234 + ], + "lines": [ + { + "bbox": [ + 201, + 179, + 306, + 234 + ], + "spans": [ + { + "bbox": [ + 201, + 179, + 306, + 234 + ], + "type": "image", + "image_path": "52b189b6a01711cee7f2d2c822e7dc22bb70fecd5be50c5c68e84491ab66c548.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 201, + 243, + 306, + 297 + ], + "blocks": [ + { + "bbox": [ + 201, + 243, + 306, + 297 + ], + "lines": [ + { + "bbox": [ + 201, + 243, + 306, + 297 + ], + "spans": [ + { + "bbox": [ + 201, + 243, + 306, + 297 + ], + "type": "image", + "image_path": "f82c21b08afc888f82de1dcb8fd4443713cf76b3ec7242e92ddc0b45bdb0fcb6.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 201, + 298, + 306, + 352 + ], + "blocks": [ + { + "bbox": [ + 201, + 298, + 306, + 352 + ], + "lines": [ + { + "bbox": [ + 201, + 298, + 306, + 352 + ], + "spans": [ + { + "bbox": [ + 201, + 298, + 306, + 352 + ], + "type": "image", + "image_path": "2fe5c2c4b8766f0381d57de92c581e0e734a5065c0e423a883aa767ba11ab929.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 201, + 353, + 306, + 407 + ], + "blocks": [ + { + "bbox": [ + 201, + 353, + 306, + 407 + ], + "lines": [ + { + "bbox": [ + 201, + 353, + 306, + 407 + ], + "spans": [ + { + "bbox": [ + 201, + 353, + 306, + 407 + ], + "type": "image", + "image_path": "8229a8044af90657caad6b08a4f61242f4ad14e943f43e2d07cd348ff295a9c4.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 223, + 408, + 283, + 429 + ], + "lines": [ + { + "bbox": [ + 223, + 408, + 283, + 429 + ], + "spans": [ + { + "bbox": [ + 223, + 408, + 283, + 429 + ], + "type": "text", + "content": "(b) Omnimatte [Lu, et al.]" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 309, + 70, + 413, + 125 + ], + "blocks": [ + { + "bbox": [ + 309, + 70, + 413, + 125 + ], + "lines": [ + { + "bbox": [ + 309, + 70, + 413, + 125 + ], + "spans": [ + { + "bbox": [ + 309, + 70, + 413, + 125 + ], + "type": "image", + "image_path": "c962a57bb6a79e84ae5eaea389c5cf41bbc81d793aff72b9086d91829c88cd91.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 309, + 125, + 413, + 178 + ], + "blocks": [ + { + "bbox": [ + 309, + 125, + 413, + 178 + ], + "lines": [ + { + "bbox": [ + 309, + 125, + 413, + 178 + ], + "spans": [ + { + "bbox": [ + 309, + 125, + 413, + 178 + ], + "type": "image", + "image_path": "c3a02875d7a22924f919829172e4176765060e2cc1951539ddbf2f19e0f36228.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 309, + 179, + 413, + 234 + ], + "blocks": [ + { + "bbox": [ + 309, + 179, + 413, + 234 + ], + "lines": [ + { + "bbox": [ + 309, + 179, + 413, + 234 + ], + "spans": [ + { + "bbox": [ + 309, + 179, + 413, + 234 + ], + "type": "image", + "image_path": "ccb7fa06775183f0cbbd8efbf29a89ae52b83731639930fe5d98df89fabca7cc.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 309, + 243, + 413, + 297 + ], + "blocks": [ + { + "bbox": [ + 309, + 243, + 413, + 297 + ], + "lines": [ + { + "bbox": [ + 309, + 243, + 413, + 297 + ], + "spans": [ + { + "bbox": [ + 309, + 243, + 413, + 297 + ], + "type": "image", + "image_path": "9ff863191447336726b224d0b71082af812d7617c27683a78a25bb740b9980ce.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 309, + 298, + 413, + 352 + ], + "blocks": [ + { + "bbox": [ + 309, + 298, + 413, + 352 + ], + "lines": [ + { + "bbox": [ + 309, + 298, + 413, + 352 + ], + "spans": [ + { + "bbox": [ + 309, + 298, + 413, + 352 + ], + "type": "image", + "image_path": "39eac2b1182121aa36d86bd7172b957f28d385bc21885266c7851e71256c205a.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 309, + 353, + 413, + 407 + ], + "blocks": [ + { + "bbox": [ + 309, + 353, + 413, + 407 + ], + "lines": [ + { + "bbox": [ + 309, + 353, + 413, + 407 + ], + "spans": [ + { + "bbox": [ + 309, + 353, + 413, + 407 + ], + "type": "image", + "image_path": "7669387a3aa52886bf58c2a63ea8300a0d746a350f65d91f60b3090070a1e22f.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 331, + 407, + 392, + 428 + ], + "lines": [ + { + "bbox": [ + 331, + 407, + 392, + 428 + ], + "spans": [ + { + "bbox": [ + 331, + 407, + 392, + 428 + ], + "type": "text", + "content": "(c) Neural Atlas [Kasten, et al.]" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 416, + 70, + 521, + 125 + ], + "blocks": [ + { + "bbox": [ + 416, + 70, + 521, + 125 + ], + "lines": [ + { + "bbox": [ + 416, + 70, + 521, + 125 + ], + "spans": [ + { + "bbox": [ + 416, + 70, + 521, + 125 + ], + "type": "image", + "image_path": "2c7b0948021ce7587f18b6518522ae401751a78740ed6ecfbd0ac78420e48312.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 416, + 125, + 521, + 178 + ], + "blocks": [ + { + "bbox": [ + 416, + 125, + 521, + 178 + ], + "lines": [ + { + "bbox": [ + 416, + 125, + 521, + 178 + ], + "spans": [ + { + "bbox": [ + 416, + 125, + 521, + 178 + ], + "type": "image", + "image_path": "54971f03d5c0a546679b9498640ec34cd6a725c035e006f598081049379b1e59.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 416, + 179, + 521, + 234 + ], + "blocks": [ + { + "bbox": [ + 416, + 179, + 521, + 234 + ], + "lines": [ + { + "bbox": [ + 416, + 179, + 521, + 234 + ], + "spans": [ + { + "bbox": [ + 416, + 179, + 521, + 234 + ], + "type": "image", + "image_path": "eb1f6d1701db8f3229140ddecbdd80b14701bc8e0c407250bae1df23c0e7f860.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 416, + 243, + 521, + 297 + ], + "blocks": [ + { + "bbox": [ + 416, + 243, + 521, + 297 + ], + "lines": [ + { + "bbox": [ + 416, + 243, + 521, + 297 + ], + "spans": [ + { + "bbox": [ + 416, + 243, + 521, + 297 + ], + "type": "image", + "image_path": "dbfd4b77c0ffb62f3aaae0d84b88e2e2a186b80c73b44cdda4b1142d2f970b8a.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 416, + 298, + 521, + 352 + ], + "blocks": [ + { + "bbox": [ + 416, + 298, + 521, + 352 + ], + "lines": [ + { + "bbox": [ + 416, + 298, + 521, + 352 + ], + "spans": [ + { + "bbox": [ + 416, + 298, + 521, + 352 + ], + "type": "image", + "image_path": "62d37bc52c0eb1812811b2f496956b5f496f8256d57b00c549b2db2bc64ef9d7.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 416, + 353, + 521, + 406 + ], + "blocks": [ + { + "bbox": [ + 416, + 353, + 521, + 406 + ], + "lines": [ + { + "bbox": [ + 416, + 353, + 521, + 406 + ], + "spans": [ + { + "bbox": [ + 416, + 353, + 521, + 406 + ], + "type": "image", + "image_path": "07fb06dd1c380a14ec2ce1849be95f2d4f66396c74f2b535967e400375d86772.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 452, + 407, + 485, + 417 + ], + "lines": [ + { + "bbox": [ + 452, + 407, + 485, + 417 + ], + "spans": [ + { + "bbox": [ + 452, + 407, + 485, + 417 + ], + "type": "text", + "content": "(d) Ours" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 78, + 456, + 172, + 510 + ], + "blocks": [ + { + "bbox": [ + 105, + 443, + 156, + 454 + ], + "lines": [ + { + "bbox": [ + 105, + 443, + 156, + 454 + ], + "spans": [ + { + "bbox": [ + 105, + 443, + 156, + 454 + ], + "type": "text", + "content": "Longboard" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 78, + 456, + 172, + 510 + ], + "lines": [ + { + "bbox": [ + 78, + 456, + 172, + 510 + ], + "spans": [ + { + "bbox": [ + 78, + 456, + 172, + 510 + ], + "type": "image", + "image_path": "d4f9dd9e564e34541bdd7e976e5b0eb19776fe84813bfe3dd58ad7c11c1200d9.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 78, + 511, + 172, + 564 + ], + "blocks": [ + { + "bbox": [ + 78, + 511, + 172, + 564 + ], + "lines": [ + { + "bbox": [ + 78, + 511, + 172, + 564 + ], + "spans": [ + { + "bbox": [ + 78, + 511, + 172, + 564 + ], + "type": "image", + "image_path": "5021088c22f6f892ecfdeea3b81d75ff5283d058110e7f64541ae792b90a16a3.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 78, + 566, + 172, + 620 + ], + "blocks": [ + { + "bbox": [ + 78, + 566, + 172, + 620 + ], + "lines": [ + { + "bbox": [ + 78, + 566, + 172, + 620 + ], + "spans": [ + { + "bbox": [ + 78, + 566, + 172, + 620 + ], + "type": "image", + "image_path": "b2047b7136227340aadfdd57aa5cb2bda93238cf25f66d55eec97316c667c5c2.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 102, + 621, + 156, + 642 + ], + "lines": [ + { + "bbox": [ + 102, + 621, + 156, + 642 + ], + "spans": [ + { + "bbox": [ + 102, + 621, + 156, + 642 + ], + "type": "text", + "content": "(a) Input RGB and Masks" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 653, + 547, + 709 + ], + "lines": [ + { + "bbox": [ + 46, + 653, + 547, + 709 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 547, + 709 + ], + "type": "text", + "content": "Figure 4. Qualitative layer decomposition results on DAVIS. Top: examples of single-layer decomposition on Snowboard and Scooter-gray. For each example, we show the predicted background layer (row 1), predicted object layer RGBA (row 2), and alpha channel visualization (row 3). Bottom: example of multi-layer decomposition on the Longboard scene. On Longboard, we only compare against Omnimatte since Neural-Atlas does not support multi-layer decomposition. Our method produces stronger background inpainting results and more accurately disentangles the foreground objects and their effects from the background." + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_caption" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 219, + 456, + 302, + 509 + ], + "blocks": [ + { + "bbox": [ + 219, + 456, + 302, + 509 + ], + "lines": [ + { + "bbox": [ + 219, + 456, + 302, + 509 + ], + "spans": [ + { + "bbox": [ + 219, + 456, + 302, + 509 + ], + "type": "image", + "image_path": "03d8b10fdb4617753fd969c2929f723ca25f25b6ffe5e254e82357a1f13538ba.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 176, + 511, + 260, + 564 + ], + "blocks": [ + { + "bbox": [ + 176, + 511, + 260, + 564 + ], + "lines": [ + { + "bbox": [ + 176, + 511, + 260, + 564 + ], + "spans": [ + { + "bbox": [ + 176, + 511, + 260, + 564 + ], + "type": "image", + "image_path": "0c80a3ec8f783f02bdd536abdb2413053a35ff174138e70aa7183dd2e1638e95.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 176, + 566, + 260, + 620 + ], + "blocks": [ + { + "bbox": [ + 176, + 566, + 260, + 620 + ], + "lines": [ + { + "bbox": [ + 176, + 566, + 260, + 620 + ], + "spans": [ + { + "bbox": [ + 176, + 566, + 260, + 620 + ], + "type": "image", + "image_path": "759393cf1a10eb197c3cd49608101c0ae08aff0a6eb320ed288cd374dbb6df8e.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 212, + 621, + 312, + 632 + ], + "lines": [ + { + "bbox": [ + 212, + 621, + 312, + 632 + ], + "spans": [ + { + "bbox": [ + 212, + 621, + 312, + 632 + ], + "type": "text", + "content": "(b) Omnimatte [Lu, et al.]" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 262, + 511, + 346, + 564 + ], + "blocks": [ + { + "bbox": [ + 262, + 511, + 346, + 564 + ], + "lines": [ + { + "bbox": [ + 262, + 511, + 346, + 564 + ], + "spans": [ + { + "bbox": [ + 262, + 511, + 346, + 564 + ], + "type": "image", + "image_path": "4834d6c94a970b946c22997340df17c436a9bf9401054749d3d688acf379ceae.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 262, + 566, + 346, + 620 + ], + "blocks": [ + { + "bbox": [ + 262, + 566, + 346, + 620 + ], + "lines": [ + { + "bbox": [ + 262, + 566, + 346, + 620 + ], + "spans": [ + { + "bbox": [ + 262, + 566, + 346, + 620 + ], + "type": "image", + "image_path": "503bbc695f1eff1d08e50f2b75dabadaf16adcf4f4246a0002c23273f72a6a26.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 393, + 456, + 477, + 509 + ], + "blocks": [ + { + "bbox": [ + 393, + 456, + 477, + 509 + ], + "lines": [ + { + "bbox": [ + 393, + 456, + 477, + 509 + ], + "spans": [ + { + "bbox": [ + 393, + 456, + 477, + 509 + ], + "type": "image", + "image_path": "2b2fa42908f225562330873f7b882c289e29c6b0b79fad6aa05627b6675dfbf1.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 350, + 511, + 434, + 564 + ], + "blocks": [ + { + "bbox": [ + 350, + 511, + 434, + 564 + ], + "lines": [ + { + "bbox": [ + 350, + 511, + 434, + 564 + ], + "spans": [ + { + "bbox": [ + 350, + 511, + 434, + 564 + ], + "type": "image", + "image_path": "96681310eda1bc3ac9a1be98d63ac85993bb546fc5700c2fd0bd5da3fccf59f5.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 350, + 566, + 434, + 620 + ], + "blocks": [ + { + "bbox": [ + 350, + 566, + 434, + 620 + ], + "lines": [ + { + "bbox": [ + 350, + 566, + 434, + 620 + ], + "spans": [ + { + "bbox": [ + 350, + 566, + 434, + 620 + ], + "type": "image", + "image_path": "3d668f5836aa2544266770652979a6d6a98276e306d11384c134d3350d9fb2cc.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 419, + 621, + 451, + 632 + ], + "lines": [ + { + "bbox": [ + 419, + 621, + 451, + 632 + ], + "spans": [ + { + "bbox": [ + 419, + 621, + 451, + 632 + ], + "type": "text", + "content": "(c) Ours" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_caption" + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 436, + 511, + 521, + 564 + ], + "blocks": [ + { + "bbox": [ + 436, + 511, + 521, + 564 + ], + "lines": [ + { + "bbox": [ + 436, + 511, + 521, + 564 + ], + "spans": [ + { + "bbox": [ + 436, + 511, + 521, + 564 + ], + "type": "image", + "image_path": "aff5a4b3e7d3a341bbccedb7a6985f4161ab287b910ba35a0736d05cb5f655a1.jpg" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_body" + } + ], + "index": 43 + }, + { + "type": "image", + "bbox": [ + 436, + 566, + 521, + 620 + ], + "blocks": [ + { + "bbox": [ + 436, + 566, + 521, + 620 + ], + "lines": [ + { + "bbox": [ + 436, + 566, + 521, + 620 + ], + "spans": [ + { + "bbox": [ + 436, + 566, + 521, + 620 + ], + "type": "image", + "image_path": "020eb3e353a897fb9ea21d7eb73c737137b78ffe7e77beb1fe15927dc6340e69.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + } + ], + "index": 44 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "635" + } + ] + } + ], + "index": 46 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 280, + 241 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 280, + 241 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 280, + 241 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 280, + 241 + ], + "type": "image", + "image_path": "712d558bba4b27d075dc19895bee7985952b200810d8cb2ea54e04c1a09712d4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 251, + 288, + 285 + ], + "lines": [ + { + "bbox": [ + 46, + 251, + 288, + 285 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 288, + 285 + ], + "type": "text", + "content": "Figure 5. Object removal. Top row shows the input frames with the input masks highlighted in blue and green. Subsequent rows show the result of removing the green and blue objects respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 307, + 288, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 307, + 288, + 415 + ], + "spans": [ + { + "bbox": [ + 46, + 307, + 288, + 415 + ], + "type": "text", + "content": "foreground layer while accurately predicting the background color. The last row shows an example video (Longboard) with two object layers. For this scene, we compare against Omnimatte as Neural-Atlas does not support multi-layer decomposition. Omnimatte's foreground layer for the cyclist contains significant pollution from background elements. Additionally, despite severe occlusion, our method is able to plausibly inpaint the background as compared to Omnimatte. We provide additional results in the supplementary material." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 423, + 146, + 437 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 423, + 146, + 437 + ], + "spans": [ + { + "bbox": [ + 47, + 423, + 146, + 437 + ], + "type": "text", + "content": "4.3. Object Removal" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 442, + 287, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 442, + 287, + 551 + ], + "spans": [ + { + "bbox": [ + 46, + 442, + 287, + 551 + ], + "type": "text", + "content": "Our predicted layers can be used to selectively remove some or all objects from a scene. Figure 5 shows an example of object removal using our layered representation on the Longboard video. Note that the input masks only account for the people and do not include attached objects such as the bicycle or longboard. Our method groups these objects with the correct person, allowing a user to easily remove different people from the scene without the undesirable effect of leaving behind attached objects." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 559, + 190, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 559, + 190, + 572 + ], + "spans": [ + { + "bbox": [ + 47, + 559, + 190, + 572 + ], + "type": "text", + "content": "4.4. Depth-based Applications" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 578, + 287, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 578, + 287, + 662 + ], + "spans": [ + { + "bbox": [ + 46, + 578, + 287, + 662 + ], + "type": "text", + "content": "While our method does not directly produce foreground depth, the input depth can be transferred to the foreground layers using the predicted foreground matrices to construct a full Layered Depth Image (LDI) for each frame (see the supplement for details). This representation is useful for a variety of applications, including camera stabilization and synthetic defocus, which we show in the following sections." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 714 + ], + "type": "text", + "content": "Camera Stabilization. For an input video captured with a shaky camera, we can stabilize the camera path by rendering the scene from a stable camera viewpoint at each time-step. Figure 6 shows results on rendering the Dancing-person [37]" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 313, + 70, + 538, + 253 + ], + "blocks": [ + { + "bbox": [ + 313, + 70, + 538, + 253 + ], + "lines": [ + { + "bbox": [ + 313, + 70, + 538, + 253 + ], + "spans": [ + { + "bbox": [ + 313, + 70, + 538, + 253 + ], + "type": "image", + "image_path": "ad44e8c8ddf93a86611eede45f5fb9a7a162dff347471ff0c9e5a8fb62e44af3.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 260, + 547, + 304 + ], + "lines": [ + { + "bbox": [ + 305, + 260, + 547, + 304 + ], + "spans": [ + { + "bbox": [ + 305, + 260, + 547, + 304 + ], + "type": "text", + "content": "Figure 6. Camera stabilization. Our LDI representation can be re-rendered from new camera positions to produce stabilized or modified camera paths. Left: input video with camera dolly motion. Right: frames re-rendered from a stationary camera." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 326, + 547, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 326, + 547, + 373 + ], + "spans": [ + { + "bbox": [ + 305, + 326, + 547, + 373 + ], + "type": "text", + "content": "scene from a stationary camera (for a moving camera path, see supp. material). The explicit mesh representation obtained from our model facilitates editing, as they can be used with commercial video editing tools." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 376, + 547, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 376, + 547, + 461 + ], + "spans": [ + { + "bbox": [ + 305, + 376, + 547, + 461 + ], + "type": "text", + "content": "Synthetic Defocus. Our foreground matte can be used to create a synthetic defocus (bokeh) effect (Fig. 7). To create this effect, we blur the background based on the depth map computed by CasualSAM [48], then composite our foreground RGBA layer on top of the blurred background. The single-layer depth map alone cannot support this effect, as it is insufficiently accurate near the subject's boundary," + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 313, + 478, + 426, + 548 + ], + "blocks": [ + { + "bbox": [ + 313, + 478, + 426, + 548 + ], + "lines": [ + { + "bbox": [ + 313, + 478, + 426, + 548 + ], + "spans": [ + { + "bbox": [ + 313, + 478, + 426, + 548 + ], + "type": "image", + "image_path": "f75fb662a4e73d083449fe643b9580bae88f8775262e53ddcec8f4c332259744.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 354, + 548, + 383, + 556 + ], + "lines": [ + { + "bbox": [ + 354, + 548, + 383, + 556 + ], + "spans": [ + { + "bbox": [ + 354, + 548, + 383, + 556 + ], + "type": "text", + "content": "Original" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 427, + 478, + 539, + 548 + ], + "blocks": [ + { + "bbox": [ + 427, + 478, + 539, + 548 + ], + "lines": [ + { + "bbox": [ + 427, + 478, + 539, + 548 + ], + "spans": [ + { + "bbox": [ + 427, + 478, + 539, + 548 + ], + "type": "image", + "image_path": "6a87757ef4294670538aec93d049591b251d6bb032a63a9a7c8825159249de83.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 475, + 548, + 493, + 556 + ], + "lines": [ + { + "bbox": [ + 475, + 548, + 493, + 556 + ], + "spans": [ + { + "bbox": [ + 475, + 548, + 493, + 556 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 314, + 556, + 426, + 625 + ], + "blocks": [ + { + "bbox": [ + 314, + 556, + 426, + 625 + ], + "lines": [ + { + "bbox": [ + 314, + 556, + 426, + 625 + ], + "spans": [ + { + "bbox": [ + 314, + 556, + 426, + 625 + ], + "type": "image", + "image_path": "e826f33b8cced0668e9bbbfa5515432b85c36cc08276d43225aaf680e95370a1.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 330, + 626, + 417, + 635 + ], + "lines": [ + { + "bbox": [ + 330, + 626, + 417, + 635 + ], + "spans": [ + { + "bbox": [ + 330, + 626, + 417, + 635 + ], + "type": "text", + "content": "CasualSAM [Zhang, et al.]" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 427, + 556, + 539, + 626 + ], + "blocks": [ + { + "bbox": [ + 427, + 556, + 539, + 626 + ], + "lines": [ + { + "bbox": [ + 427, + 556, + 539, + 626 + ], + "spans": [ + { + "bbox": [ + 427, + 556, + 539, + 626 + ], + "type": "image", + "image_path": "72e9867a7ae975a0a3934631f112fd240385acde390c019957ec129f99807d35.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 443, + 626, + 518, + 635 + ], + "lines": [ + { + "bbox": [ + 443, + 626, + 518, + 635 + ], + "spans": [ + { + "bbox": [ + 443, + 626, + 518, + 635 + ], + "type": "text", + "content": "Omnimatte [Lu, et al.]" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 644, + 547, + 710 + ], + "lines": [ + { + "bbox": [ + 304, + 644, + 547, + 710 + ], + "spans": [ + { + "bbox": [ + 304, + 644, + 547, + 710 + ], + "type": "text", + "content": "Figure 7. Synthetic defocus. Our foreground matte can enable a synthetic defocus effect (Ours). State-of-the-art single-layer depth maps (CasualSAM [48]) are not accurate near the subject's boundary, causing the goat's head to blur. The foreground matte from Omnimatte [25] contains background pixels, causing parts of the background to remain unblurred." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "636" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 70, + 128, + 164 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 128, + 164 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 128, + 164 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 128, + 164 + ], + "type": "image", + "image_path": "a6a93ec25857b83686f5972fb8e71b5efd201b7d3ed3a548212ad4ab2099e059.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 166, + 115, + 184 + ], + "lines": [ + { + "bbox": [ + 67, + 166, + 115, + 184 + ], + "spans": [ + { + "bbox": [ + 67, + 166, + 115, + 184 + ], + "type": "text", + "content": "(a) Input RGB and Mask" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 130, + 70, + 204, + 164 + ], + "blocks": [ + { + "bbox": [ + 130, + 70, + 204, + 164 + ], + "lines": [ + { + "bbox": [ + 130, + 70, + 204, + 164 + ], + "spans": [ + { + "bbox": [ + 130, + 70, + 204, + 164 + ], + "type": "image", + "image_path": "a7d5844dfca74fef5014e5c75ac387fc8d04971f3615f38632f1fbed506bca6f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 135, + 166, + 200, + 182 + ], + "lines": [ + { + "bbox": [ + 135, + 166, + 200, + 182 + ], + "spans": [ + { + "bbox": [ + 135, + 166, + 200, + 182 + ], + "type": "text", + "content": "(b) Random Frame Selection" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 206, + 70, + 279, + 164 + ], + "blocks": [ + { + "bbox": [ + 206, + 70, + 279, + 164 + ], + "lines": [ + { + "bbox": [ + 206, + 70, + 279, + 164 + ], + "spans": [ + { + "bbox": [ + 206, + 70, + 279, + 164 + ], + "type": "image", + "image_path": "ce2dfd7257cb489674c7bd36a6e90b9e817c46deb1dcc9fec04f6deb212ca26d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 219, + 166, + 267, + 182 + ], + "lines": [ + { + "bbox": [ + 219, + 166, + 267, + 182 + ], + "spans": [ + { + "bbox": [ + 219, + 166, + 267, + 182 + ], + "type": "text", + "content": "(c) Our Frame Selection" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 47, + 194, + 288, + 228 + ], + "lines": [ + { + "bbox": [ + 47, + 194, + 288, + 228 + ], + "spans": [ + { + "bbox": [ + 47, + 194, + 288, + 228 + ], + "type": "text", + "content": "Figure 8. Frame Selection Ablation. Results from random frame selection instead of the proposed heuristic. Unlike the full method, the foreground layer does not accurately capture the shadow." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 249, + 286, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 249, + 286, + 308 + ], + "spans": [ + { + "bbox": [ + 46, + 249, + 286, + 308 + ], + "type": "text", + "content": "causing important small features to be erroneously blurred (Fig. 7, bottom). The foreground matte from Omnimatte [25] is also insufficiently accurate, containing elements of the background that remain erroneously sharp when composited over the blurred background layer." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 316, + 115, + 327 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 316, + 115, + 327 + ], + "spans": [ + { + "bbox": [ + 47, + 316, + 115, + 327 + ], + "type": "text", + "content": "4.5. Ablations" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 334, + 287, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 334, + 287, + 478 + ], + "spans": [ + { + "bbox": [ + 46, + 334, + 287, + 478 + ], + "type": "text", + "content": "Frame Selection. We ablate our frame selection method 3.4 by training a model using random frame sampling. For a source frame " + }, + { + "bbox": [ + 46, + 334, + 287, + 478 + ], + "type": "inline_equation", + "content": "F_{a}" + }, + { + "bbox": [ + 46, + 334, + 287, + 478 + ], + "type": "text", + "content": ", we randomly sample a shift value to obtain the second source " + }, + { + "bbox": [ + 46, + 334, + 287, + 478 + ], + "type": "inline_equation", + "content": "F_{b}" + }, + { + "bbox": [ + 46, + 334, + 287, + 478 + ], + "type": "text", + "content": ". The target frame " + }, + { + "bbox": [ + 46, + 334, + 287, + 478 + ], + "type": "inline_equation", + "content": "(F_{c})" + }, + { + "bbox": [ + 46, + 334, + 287, + 478 + ], + "type": "text", + "content": " is then chosen mid-way between the source frames. Figure 8 shows results from a model trained using the random frame selection on the hike scene. Compared to the model trained with our frame selection heuristic, random selection leads to most of the shadow being erroneously captured in the background layer. While it is possible to select shift values that can generate desirable results, the proposed frame selection heuristic removes the need to search over this hyperparameter." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 50, + 487, + 127, + 536 + ], + "blocks": [ + { + "bbox": [ + 50, + 487, + 127, + 536 + ], + "lines": [ + { + "bbox": [ + 50, + 487, + 127, + 536 + ], + "spans": [ + { + "bbox": [ + 50, + 487, + 127, + 536 + ], + "type": "image", + "image_path": "f176aa26e0cb83e883bba536b8648bce9b832ad330f239409c45702030ce3b90.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 50, + 538, + 127, + 586 + ], + "blocks": [ + { + "bbox": [ + 50, + 538, + 127, + 586 + ], + "lines": [ + { + "bbox": [ + 50, + 538, + 127, + 586 + ], + "spans": [ + { + "bbox": [ + 50, + 538, + 127, + 586 + ], + "type": "image", + "image_path": "0d430f0b9ceeefbef2f54bfee616012061a11933a583540b161ce291d13e210d.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 63, + 587, + 112, + 606 + ], + "lines": [ + { + "bbox": [ + 63, + 587, + 112, + 606 + ], + "spans": [ + { + "bbox": [ + 63, + 587, + 112, + 606 + ], + "type": "text", + "content": "(a) Input RGB and Mask" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 129, + 487, + 206, + 586 + ], + "blocks": [ + { + "bbox": [ + 129, + 487, + 206, + 586 + ], + "lines": [ + { + "bbox": [ + 129, + 487, + 206, + 586 + ], + "spans": [ + { + "bbox": [ + 129, + 487, + 206, + 586 + ], + "type": "image", + "image_path": "3e7fc12ddd773e8dfb837b4846e963e9a5dfa8d001da51e3df3af0599a8fb752.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 126, + 587, + 209, + 607 + ], + "lines": [ + { + "bbox": [ + 126, + 587, + 209, + 607 + ], + "spans": [ + { + "bbox": [ + 126, + 587, + 209, + 607 + ], + "type": "text", + "content": "(b) Without Projection Consistency" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 208, + 487, + 284, + 586 + ], + "blocks": [ + { + "bbox": [ + 208, + 487, + 284, + 586 + ], + "lines": [ + { + "bbox": [ + 208, + 487, + 284, + 586 + ], + "spans": [ + { + "bbox": [ + 208, + 487, + 284, + 586 + ], + "type": "image", + "image_path": "74be33b2385d4863398e60470138e5340da1da0d50d1565a7baf8c726d28dba7.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 218, + 587, + 274, + 596 + ], + "lines": [ + { + "bbox": [ + 218, + 587, + 274, + 596 + ], + "spans": [ + { + "bbox": [ + 218, + 587, + 274, + 596 + ], + "type": "text", + "content": "(c) Full Method" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 47, + 677, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 677, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 677, + 288, + 713 + ], + "type": "text", + "content": "Projection Consistency Loss. We ablate the projection consistency loss in Fig. 9. As shown in column (b), the inpainting of the unobserved region fails without the projection" + } + ] + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 307, + 71, + 367, + 148 + ], + "blocks": [ + { + "bbox": [ + 307, + 71, + 367, + 148 + ], + "lines": [ + { + "bbox": [ + 307, + 71, + 367, + 148 + ], + "spans": [ + { + "bbox": [ + 307, + 71, + 367, + 148 + ], + "type": "image", + "image_path": "5590f940c77393f901a6efdd9343c1bdf4acb35989ad34ae56c53b0ad852efd8.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 329, + 149, + 345, + 156 + ], + "lines": [ + { + "bbox": [ + 329, + 149, + 345, + 156 + ], + "spans": [ + { + "bbox": [ + 329, + 149, + 345, + 156 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 166, + 545, + 232 + ], + "lines": [ + { + "bbox": [ + 305, + 166, + 545, + 232 + ], + "spans": [ + { + "bbox": [ + 305, + 166, + 545, + 232 + ], + "type": "text", + "content": "Figure 10. Detail Transfer Ablation. The background depth map allows reprojecting detail from nearby frames in time to further improve the visual quality of the predicted layers. From left to right: input frame, full result with detail transferred from frames " + }, + { + "bbox": [ + 305, + 166, + 545, + 232 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 305, + 166, + 545, + 232 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 166, + 545, + 232 + ], + "type": "inline_equation", + "content": "t - 5" + }, + { + "bbox": [ + 305, + 166, + 545, + 232 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 305, + 166, + 545, + 232 + ], + "type": "inline_equation", + "content": "t + 5" + }, + { + "bbox": [ + 305, + 166, + 545, + 232 + ], + "type": "text", + "content": ", ablated result with detail only from frame " + }, + { + "bbox": [ + 305, + 166, + 545, + 232 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 305, + 166, + 545, + 232 + ], + "type": "text", + "content": " (note blurry region under bicyclist), no background detail transfer." + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 367, + 72, + 425, + 148 + ], + "blocks": [ + { + "bbox": [ + 367, + 72, + 425, + 148 + ], + "lines": [ + { + "bbox": [ + 367, + 72, + 425, + 148 + ], + "spans": [ + { + "bbox": [ + 367, + 72, + 425, + 148 + ], + "type": "image", + "image_path": "64e0de7ae39545e803182657e07f1f0ece1372544ade00f9c77f314006e04847.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 372, + 149, + 422, + 156 + ], + "lines": [ + { + "bbox": [ + 372, + 149, + 422, + 156 + ], + "spans": [ + { + "bbox": [ + 372, + 149, + 422, + 156 + ], + "type": "text", + "content": "Frames t, t-5, t+5" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 427, + 72, + 484, + 148 + ], + "blocks": [ + { + "bbox": [ + 427, + 72, + 484, + 148 + ], + "lines": [ + { + "bbox": [ + 427, + 72, + 484, + 148 + ], + "spans": [ + { + "bbox": [ + 427, + 72, + 484, + 148 + ], + "type": "image", + "image_path": "4b80b9525ca417e0c73de097fc182652dc4d28dfdaa58544ea171b352e67f08e.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 436, + 149, + 472, + 156 + ], + "lines": [ + { + "bbox": [ + 436, + 149, + 472, + 156 + ], + "spans": [ + { + "bbox": [ + 436, + 149, + 472, + 156 + ], + "type": "text", + "content": "Frame t only" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 485, + 72, + 544, + 148 + ], + "blocks": [ + { + "bbox": [ + 485, + 72, + 544, + 148 + ], + "lines": [ + { + "bbox": [ + 485, + 72, + 544, + 148 + ], + "spans": [ + { + "bbox": [ + 485, + 72, + 544, + 148 + ], + "type": "image", + "image_path": "6477f9713fb018d1f68582956938a074f8e1dc032a213b77be8aa2c8a4aa45f8.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 490, + 149, + 539, + 156 + ], + "lines": [ + { + "bbox": [ + 490, + 149, + 539, + 156 + ], + "spans": [ + { + "bbox": [ + 490, + 149, + 539, + 156 + ], + "type": "text", + "content": "No detail transfer" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "bbox": [ + 305, + 254, + 545, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 254, + 545, + 290 + ], + "spans": [ + { + "bbox": [ + 305, + 254, + 545, + 290 + ], + "type": "text", + "content": "consistency term, as there is no reconstruction signal for that region. Additionally, the foreground layer is impacted due to the shadow being incorrectly placed in the background." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 306, + 301, + 459, + 314 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 301, + 459, + 314 + ], + "spans": [ + { + "bbox": [ + 306, + 301, + 459, + 314 + ], + "type": "text", + "content": "5. Discussion and Limitations" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 304, + 322, + 547, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 322, + 547, + 441 + ], + "spans": [ + { + "bbox": [ + 304, + 322, + 547, + 441 + ], + "type": "text", + "content": "We present a new method for layer decomposition that separates a monocular video into a background and several object layers along with their associated effects. Unlike prior decomposition methods that rely on a static 2D background canvas, our method separates the video into layers on a per-frame basis, making it applicable to videos with unconstrained camera motion. To address the challenge of predicting a coherent background, we predict a 3D background and propose a multi-view consistency loss that enforces the background to only contain slowly moving or static details." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 305, + 442, + 545, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 442, + 545, + 466 + ], + "spans": [ + { + "bbox": [ + 305, + 442, + 545, + 466 + ], + "type": "text", + "content": "Our model can produce per-frame Layered Depth Images that allow for various depth-based editing applications." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 304, + 466, + 546, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 466, + 546, + 561 + ], + "spans": [ + { + "bbox": [ + 304, + 466, + 546, + 561 + ], + "type": "text", + "content": "A limitation of our method is shown in Figure 11. Unlike the background, foreground layers are not reprojected to nearby frames and thus foreground objects are not inpainted when occluded. This limitation could be addressed by adding a projection consistency loss similar to Eq. (7) for the foreground layers; however, reprojection of dynamic elements is challenging and requires modelling the sceneflow, which we leave for future work." + } + ] + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 309, + 571, + 386, + 648 + ], + "blocks": [ + { + "bbox": [ + 46, + 617, + 287, + 662 + ], + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 662 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 662 + ], + "type": "text", + "content": "Figure 9. Projection Loss Ablation. We show layer results from a model trained without the projection loss. The resulting background contains severe artifacts (column (b), top), and the foreground layer is missing the shadow of the car (column (b), bottom)." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 309, + 571, + 386, + 648 + ], + "lines": [ + { + "bbox": [ + 309, + 571, + 386, + 648 + ], + "spans": [ + { + "bbox": [ + 309, + 571, + 386, + 648 + ], + "type": "image", + "image_path": "48ee58420331082e8752bed6b56807f45c591c2adcd1db595d2486157d07f402.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 329, + 649, + 366, + 655 + ], + "lines": [ + { + "bbox": [ + 329, + 649, + 366, + 655 + ], + "spans": [ + { + "bbox": [ + 329, + 649, + 366, + 655 + ], + "type": "text", + "content": "Input Frames" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 389, + 571, + 466, + 649 + ], + "blocks": [ + { + "bbox": [ + 389, + 571, + 466, + 649 + ], + "lines": [ + { + "bbox": [ + 389, + 571, + 466, + 649 + ], + "spans": [ + { + "bbox": [ + 389, + 571, + 466, + 649 + ], + "type": "image", + "image_path": "8d3fc30fe8bff3a281c7de34d00a2db1fde03ef36d9710a94cfc1305838a1886.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 397, + 649, + 452, + 655 + ], + "lines": [ + { + "bbox": [ + 397, + 649, + 452, + 655 + ], + "spans": [ + { + "bbox": [ + 397, + 649, + 452, + 655 + ], + "type": "text", + "content": "Output Background" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 666, + 545, + 711 + ], + "lines": [ + { + "bbox": [ + 304, + 666, + 545, + 711 + ], + "spans": [ + { + "bbox": [ + 304, + 666, + 545, + 711 + ], + "type": "text", + "content": "Figure 11. Limitation: object occlusions. While the background layer is inpainted using nearby frames, our method cannot inpaint the object layers. When an object passes behind the stationary background layer (middle), it is erased from the object layer (right)." + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 467, + 571, + 543, + 649 + ], + "blocks": [ + { + "bbox": [ + 467, + 571, + 543, + 649 + ], + "lines": [ + { + "bbox": [ + 467, + 571, + 543, + 649 + ], + "spans": [ + { + "bbox": [ + 467, + 571, + 543, + 649 + ], + "type": "image", + "image_path": "eb29d5e3843f5d991717960c25c62dd8e48173befffaf502028ad5a3a47890c7.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 476, + 649, + 533, + 655 + ], + "lines": [ + { + "bbox": [ + 476, + 649, + 533, + 655 + ], + "spans": [ + { + "bbox": [ + 476, + 649, + 533, + 655 + ], + "type": "text", + "content": "Output Object Layer" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + } + ], + "index": 37 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "637" + } + ] + } + ], + "index": 40 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 92, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 53, + 92, + 288, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 92, + 288, + 124 + ], + "spans": [ + { + "bbox": [ + 53, + 92, + 288, + 124 + ], + "type": "text", + "content": "[1] Jean-Baptiste Alayrac, João Carreira, and Andrew Zisserman. The visual centrifuge: Model-free layered video representations. In CVPR, 2019. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 126, + 288, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 126, + 288, + 159 + ], + "spans": [ + { + "bbox": [ + 53, + 126, + 288, + 159 + ], + "type": "text", + "content": "[2] Jean-Baptiste Alayrac, Joao Carreira, Relja Arandjelovic, and Andrew Zisserman. Controllable attention for structured layered video decomposition. In ICCV, 2019. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 161, + 288, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 161, + 288, + 195 + ], + "spans": [ + { + "bbox": [ + 53, + 161, + 288, + 195 + ], + "type": "text", + "content": "[3] Omer Bar-Tal, Dolev Ofri-Amar, Rafail Fridman, Yoni Kasten, and Tali Dekel. Text2live: Text-driven layered image and video editing. arXiv preprint arXiv:2204.02491, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 197, + 288, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 197, + 288, + 217 + ], + "spans": [ + { + "bbox": [ + 53, + 197, + 288, + 217 + ], + "type": "text", + "content": "[4] Gabriel J Brostow and Irfan A Essa. Motion based decomposing of video. In ICCV, 1999. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 220, + 288, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 220, + 288, + 253 + ], + "spans": [ + { + "bbox": [ + 53, + 220, + 288, + 253 + ], + "type": "text", + "content": "[5] Ya-Liang Chang, Zhe Yu Liu, Kuan-Ying Lee, and Winston Hsu. Free-form video inpainting with 3d gated convolution and temporal patchgan. In ICCV, pages 9066-9075, 2019. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 255, + 288, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 255, + 288, + 297 + ], + "spans": [ + { + "bbox": [ + 53, + 255, + 288, + 297 + ], + "type": "text", + "content": "[6] Forrester Cole, Kyle Genova, Avneesh Sud, Daniel Vlasic, and Zhoutong Zhang. Differentiable surface rendering via non-differentiable sampling. In ICCV, pages 6088-6097, 2021. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 300, + 288, + 344 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 300, + 288, + 344 + ], + "spans": [ + { + "bbox": [ + 53, + 300, + 288, + 344 + ], + "type": "text", + "content": "[7] Mounira Ebdelli, Olivier Le Meur, and Christine Guillemot. Video inpainting with short-term windows: application to object removal and error concealment. IEEE Transactions on Image Processing, 24(10):3034-3047, 2015. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 346, + 288, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 346, + 288, + 378 + ], + "spans": [ + { + "bbox": [ + 53, + 346, + 288, + 378 + ], + "type": "text", + "content": "[8] Matthieu Fradet, Patrick Pérez, and Philippe Robert. Semi-automatic motion segmentation with motion layer mosaics. In ECCV, 2008. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 380, + 288, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 288, + 413 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 288, + 413 + ], + "type": "text", + "content": "[9] Yossi Gandelsman, Assaf Shocher, and Michal Irani. \"Double-DIP\": Unsupervised image decomposition via coupled deep-image-priors. In CVPR, 2019. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 415, + 288, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 415, + 288, + 437 + ], + "spans": [ + { + "bbox": [ + 48, + 415, + 288, + 437 + ], + "type": "text", + "content": "[10] Chen Gao, Ayush Saraf, Jia-Bin Huang, and Johannes Kopf. Flow-edge guided video completion. In ECCV, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 439, + 288, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 439, + 288, + 472 + ], + "spans": [ + { + "bbox": [ + 48, + 439, + 288, + 472 + ], + "type": "text", + "content": "[11] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video. In ICCV, pages 5712-5721, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 474, + 288, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 474, + 288, + 517 + ], + "spans": [ + { + "bbox": [ + 48, + 474, + 288, + 517 + ], + "type": "text", + "content": "[12] Miguel Granados, James Tompkin, K Kim, Oliver Grau, Jan Kautz, and Christian Theobalt. How not to be seen—object removal from videos of crowded scenes. Computer Graphics Forum, 31(2pt1):219-228, 2012. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 519, + 288, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 519, + 288, + 541 + ], + "spans": [ + { + "bbox": [ + 48, + 519, + 288, + 541 + ], + "type": "text", + "content": "[13] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, pages 2961-2969, 2017. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 543, + 288, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 543, + 288, + 586 + ], + "spans": [ + { + "bbox": [ + 48, + 543, + 288, + 586 + ], + "type": "text", + "content": "[14] Jia-Bin Huang, Sing Bing Kang, Narendra Ahuja, and Johannes Kopf. Temporally coherent completion of dynamic video. ACM Transactions on Graphics (TOG), 35(6):1-11, 2016. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 589, + 288, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 589, + 288, + 632 + ], + "spans": [ + { + "bbox": [ + 48, + 589, + 288, + 632 + ], + "type": "text", + "content": "[15] Zhang Jiakai, Liu Xinhang, Ye Xinyi, Zhao Fuqiang, Zhang Yanshun, Wu Minye, Zhang Yingliang, Xu Lan, and Yu Jingyi. Editable free-viewpoint video using a layered neural representation. In ACM SIGGRAPH, 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 635, + 288, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 288, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 288, + 678 + ], + "type": "text", + "content": "[16] Nebojsa Joic and Brendan J Frey. Learning flexible sprites in video layers. In Proceedings of the 2001 IEEE Computer Society Conference on Computer Vision and Pattern Recognition. CVPR 2001, volume 1, pages I-I. IEEE, 2001. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 680, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 288, + 713 + ], + "type": "text", + "content": "[17] Yoni Kasten, Dolev Ofri, Oliver Wang, and Tali Dekel. Layered neural atlases for consistent video editing. ACM Transactions on Graphics (TOG), 40(6):1-12, 2021. 1, 2, 5" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 106 + ], + "type": "text", + "content": "[18] Johannes Kopf, Xuejian Rong, and Jia-Bin Huang. Robust consistent video depth estimation. In CVPR, pages 1611-1621, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 107, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 128 + ], + "type": "text", + "content": "[19] Johannes Kopf, Xuejian Rong, and Jia-Bin Huang. Robust consistent video depth estimation. In CVPR, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 129, + 547, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 547, + 160 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 547, + 160 + ], + "type": "text", + "content": "[20] M. Pawan Kumar, Philip H. S. Torr, and Andrew Zisserman. Learning layered motion segmentations of video. IJCV, 2008. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 162, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 162, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 307, + 162, + 545, + 194 + ], + "type": "text", + "content": "[21] Sungho Lee, Seoung Wug Oh, DaeYeun Won, and Seon Joo Kim. Copy-and-paste networks for deep video inpainting. In ICCV, pages 4413-4421, 2019. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 195, + 547, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 195, + 547, + 228 + ], + "spans": [ + { + "bbox": [ + 307, + 195, + 547, + 228 + ], + "type": "text", + "content": "[22] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In CVPR, pages 6498-6508, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 228, + 547, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 228, + 547, + 270 + ], + "spans": [ + { + "bbox": [ + 307, + 228, + 547, + 270 + ], + "type": "text", + "content": "[23] Chao Liu, Jinwei Gu, Kihwan Kim, Srinivasa G Narasimhan, and Jan Kautz. Neural rgb (r) d sensing: Depth and uncertainty from a video camera. In CVPR, pages 10986-10995, 2019. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 272, + 545, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 272, + 545, + 316 + ], + "spans": [ + { + "bbox": [ + 307, + 272, + 545, + 316 + ], + "type": "text", + "content": "[24] Erika Lu, Forrester Cole, Tali Dekel, Weidi Xie, Andrew Zisserman, David Salesin, William T Freeman, and Michael Rubinstein. Layered neural rendering for retiming people in video. In SIGGRAPH Asia, 2020. 1, 2, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 316, + 547, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 316, + 547, + 380 + ], + "spans": [ + { + "bbox": [ + 307, + 316, + 547, + 380 + ], + "type": "text", + "content": "[25] Erika Lu, Forrester Cole, Tali Dekel, Andrew Zisserman, William T. Freeman, and Michael Rubinstein. Omnimatte: Associating objects and their effects in video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4507-4515, June 2021. 1, 2, 5, 7, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 383, + 545, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 383, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 307, + 383, + 545, + 415 + ], + "type": "text", + "content": "[26] Xuan Luo, Jia-Bin Huang, Richard Szeliski, Kevin Matzen, and Johannes Kopf. Consistent video depth estimation. ACM Transactions on Graphics (ToG), 39(4):71-1, 2020. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 416, + 547, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 416, + 547, + 469 + ], + "spans": [ + { + "bbox": [ + 307, + 416, + 547, + 469 + ], + "type": "text", + "content": "[27] Ricardo Martin-Brualla, Noha Radwan, Mehdi SM Sajjadi, Jonathan T Barron, Alexey Dosovitskiy, and Daniel Duckworth. Nerf in the wild: Neural radiance fields for unconstrained photo collections. In CVPR, pages 7210-7219, 2021. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 471, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 471, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 307, + 471, + 545, + 514 + ], + "type": "text", + "content": "[28] Yasuyuki Matsushita, Eyal Ofek, Weina Ge, Xiaou Tang, and Heung-Yeung Shum. Full-frame video stabilization with motion inpainting. IEEE Transactions on pattern analysis and Machine Intelligence, 28(7):1150-1163, 2006. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 515, + 547, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 515, + 547, + 568 + ], + "spans": [ + { + "bbox": [ + 307, + 515, + 547, + 568 + ], + "type": "text", + "content": "[29] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 570, + 545, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 570, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 307, + 570, + 545, + 602 + ], + "type": "text", + "content": "[30] Seoung Wug Oh, Sungho Lee, Joon-Young Lee, and Seon Joo Kim. Onion-peel networks for deep video completion. In ICCV, pages 4403–4412, 2019. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 603, + 545, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 603, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 307, + 603, + 545, + 645 + ], + "type": "text", + "content": "[31] F. Perazzi, J. Pont-Tuset, B. McWilliams, L. Van Gool, M. Gross, and A. Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In CVPR, 2016. 5" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 647, + 545, + 689 + ], + "type": "text", + "content": "[32] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alexander Sorkine-Hornung, and Luc Van Gool. The 2017 DAVIS challenge on video object segmentation. arXiv:1704.00675, 2017. 5" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 713 + ], + "type": "text", + "content": "[33] Thomas Porter and Tom Duff. Compositing digital images. SIGGRAPH Comput. Graph., 18(3):253-259, Jan. 1984. 4" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "638" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 405 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 106 + ], + "type": "text", + "content": "[34] Xuejian Rong, Jia-Bin Huang, Ayush Saraf, Changil Kim, and Johannes Kopf. Boosting view synthesis with residual transfer. CVPR, 2022. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 108, + 288, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 108, + 288, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 108, + 288, + 140 + ], + "type": "text", + "content": "[35] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional Networks for Biomedical Image Segmentation. MICCAI, 2015. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 143, + 287, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 143, + 287, + 186 + ], + "spans": [ + { + "bbox": [ + 48, + 143, + 287, + 186 + ], + "type": "text", + "content": "[36] Jonathan Shade, Steven Gortler, Li-wei He, and Richard Szeliski. Layered depth images. In Proceedings of the 25th annual conference on Computer graphics and interactive techniques, pages 231–242, 1998. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 189, + 288, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 189, + 288, + 232 + ], + "spans": [ + { + "bbox": [ + 48, + 189, + 288, + 232 + ], + "type": "text", + "content": "[37] Zhenmei Shi, Fuhao Shi, Wei-Sheng Lai, Chia-Kai Liang, and Yingyu Liang. Deep online fused video stabilization. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1250-1258, 2022. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 235, + 288, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 235, + 288, + 278 + ], + "spans": [ + { + "bbox": [ + 48, + 235, + 288, + 278 + ], + "type": "text", + "content": "[38] Dmitriy Smirnov, Michael Gharbi, Matthew Fisher, Vitor Guizilini, Alexei Efros, and Justin M Solomon. Marionette: Self-supervised sprite learning. Advances in Neural Information Processing Systems, 34, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 280, + 287, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 280, + 287, + 323 + ], + "spans": [ + { + "bbox": [ + 48, + 280, + 287, + 323 + ], + "type": "text", + "content": "[39] Pratul P. Srinivasan, Richard Tucker, Jonathan T. Barron, Ravi Ramamoorthi, Ren Ng, and Noah Snavely. Pushing the boundaries of view extrapolation with multiplane images. In CVPR, 2019. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 326, + 287, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 326, + 287, + 358 + ], + "spans": [ + { + "bbox": [ + 48, + 326, + 287, + 358 + ], + "type": "text", + "content": "[40] Zachary Teed and Jia Deng. Deepv2d: Video to depth with differentiable structure from motion. arXiv preprint arXiv:1812.04605, 2018. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 361, + 287, + 405 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 361, + 287, + 405 + ], + "spans": [ + { + "bbox": [ + 48, + 361, + 287, + 405 + ], + "type": "text", + "content": "[41] Chuan Wang, Haibin Huang, Xiaoguang Han, and Jue Wang. Video inpainting by jointly learning temporal structure and spatial details. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 5232-5239, 2019. 2" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 405 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 307, + 73, + 546, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 546, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 546, + 106 + ], + "type": "text", + "content": "[42] John YA Wang and Edward H Adelson. Representing moving images with layers. IEEE transactions on image processing, 3(5):625-638, 1994. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 307, + 107, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 128 + ], + "type": "text", + "content": "[43] Yonatan Wexler, Eli Shechtman, and Michal Irani. Space-time completion of video. PAMI, 2007. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 307, + 129, + 547, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 547, + 172 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 547, + 172 + ], + "type": "text", + "content": "[44] Yiheng Xie, Towaki Takikawa, Shunsuke Saito, Or Litany, Shiqin Yan, Numair Khan, Federico Tombari, James Tompkin, Vincent Sitzmann, and Srinath Sridhar. Neural fields in visual computing and beyond. Computer Graphics Forum, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 307, + 173, + 547, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 173, + 547, + 205 + ], + "spans": [ + { + "bbox": [ + 307, + 173, + 547, + 205 + ], + "type": "text", + "content": "[45] Rui Xu, Xiaoxiao Li, Bolei Zhou, and Chen Change Loy. Deep flow-guided video inpainting. In CVPR, pages 3723-3732, 2019. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 307, + 206, + 546, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 206, + 546, + 249 + ], + "spans": [ + { + "bbox": [ + 307, + 206, + 546, + 249 + ], + "type": "text", + "content": "[46] Vickie Ye, Zhengqi Li, Richard Tucker, Angjoo Kanazawa, and Noah Snavely. Deformable sprites for unsupervised video decomposition. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2022. 1, 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 251, + 546, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 251, + 546, + 282 + ], + "spans": [ + { + "bbox": [ + 307, + 251, + 546, + 282 + ], + "type": "text", + "content": "[47] Kaidong Zhang, Jingjing Fu, and Dong Liu. Flow-guided transformer for video inpainting. In European Conference on Computer Vision, pages 74-90. Springer, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 284, + 546, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 284, + 546, + 327 + ], + "spans": [ + { + "bbox": [ + 307, + 284, + 546, + 327 + ], + "type": "text", + "content": "[48] Zhoutong Zhang, Forrester Cole, Zhengqi Li, Michael Rubinstein, Noah Snavely, and William T Freeman. Structure and motion from casual videos. In European Conference on Computer Vision, pages 20-37. Springer, 2022. 2, 5, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 327, + 546, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 327, + 546, + 370 + ], + "spans": [ + { + "bbox": [ + 307, + 327, + 546, + 370 + ], + "type": "text", + "content": "[49] Zhoutong Zhang, Forrester Cole, Richard Tucker, William T Freeman, and Tali Dekel. Consistent depth of moving objects in video. ACM Transactions on Graphics (TOG), 40(4):1-12, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 372, + 546, + 405 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 372, + 546, + 405 + ], + "spans": [ + { + "bbox": [ + 307, + 372, + 546, + 405 + ], + "type": "text", + "content": "[50] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images. In SIGGRAPH, 2018. 2" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "639" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/On Calibrating Semantic Segmentation Models_ Analyses and an Algorithm/f35086d8-d1cb-49d4-972d-c5bbdb5412ed_content_list.json b/2023/On Calibrating Semantic Segmentation Models_ Analyses and an Algorithm/f35086d8-d1cb-49d4-972d-c5bbdb5412ed_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..7fc47b658383cb46dc020e59797068df789f1966 --- /dev/null +++ b/2023/On Calibrating Semantic Segmentation Models_ Analyses and an Algorithm/f35086d8-d1cb-49d4-972d-c5bbdb5412ed_content_list.json @@ -0,0 +1,1749 @@ +[ + { + "type": "text", + "text": "On Calibrating Semantic Segmentation Models: Analyses and An Algorithm", + "text_level": 1, + "bbox": [ + 99, + 130, + 870, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dongdong Wang \nUniversity of Central Florida \ndaniel.wang@knights.ucf.edu", + "bbox": [ + 135, + 181, + 374, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Boqing Gong Google Research bgong@google.com", + "bbox": [ + 416, + 181, + 560, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Liqiang Wang \nUniversity of Central Florida \nliqiang.wang@ucf.edu", + "bbox": [ + 604, + 181, + 836, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We study the problem of semantic segmentation calibration. Lots of solutions have been proposed to approach model miscalibration of confidence in image classification. However, to date, confidence calibration research on semantic segmentation is still limited. We provide a systematic study on the calibration of semantic segmentation models and propose a simple yet effective approach. First, we find that model capacity, crop size, multi-scale testing, and prediction correctness have impact on calibration. Among them, prediction correctness, especially misprediction, is more important to miscalibration due to over-confidence. Next, we propose a simple, unifying, and effective approach, namely selective scaling, by separating correct/incorrect prediction for scaling and more focusing on misprediction logit smoothing. Then, we study popular existing calibration methods and compare them with selective scaling on semantic segmentation calibration. We conduct extensive experiments with a variety of benchmarks on both indomain and domain-shift calibration and show that selective scaling consistently outperforms other methods.", + "bbox": [ + 75, + 301, + 473, + 603 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 650, + 209, + 666 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep neural networks (DNNs) have become the \"goto\" models in various computer vision tasks, such as image classification, object detection, semantic segmentation. However, recent work found that DNNs are often overconfident when they make mistakes [11], misleading downstream applications. To calibrate DNNs' confidence in prediction, researchers have developed a rich line of works on image classification using regularized training [30, 31, 41, 46], post-hoc processing [11, 19, 25, 36], and Bayesian modeling [15, 18, 21, 47], to name a few.", + "bbox": [ + 75, + 676, + 468, + 828 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, the extensive pursuit of image classification has left it unclear how to calibrate DNNs for other computer vision tasks and how well existing calibration methods gen", + "bbox": [ + 76, + 829, + 468, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/45c8dd131d6ad700c2a4266b9c343191802bafb14f5770c9a48d93be122a1cf1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 516, + 270, + 676, + 378 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7b7545543cb0e9d51e55ae6d60580fcc9d0f4f7eaf393c15c76f8b1bb9c0b672.jpg", + "image_caption": [ + "Confidence" + ], + "image_footnote": [], + "bbox": [ + 516, + 387, + 676, + 494 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4908090dac040b7671834c84d750998d01811490a309a5c4f1423056ccc836bd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 669, + 268, + 854, + 377 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/375aba9e23443008535c3e2200797845360559c534bb40053a39e30b99247d82.jpg", + "image_caption": [ + "Figure 1. Reliability diagrams [5] of visualized ECE for Segmenter [40]. We randomly select and evaluate fifty validation images from COCO-164K [1] and compare uncalibrated model (Uncal), selective scaling (Selective), temperature scaling (TempS), and ensembling (Ensemble). Smaller gap implies less ECE and better calibration. Ensembling yields comparable accuracy. Despite limited improvement by TempS and Ensemble, Selective exhibits better segmentation model calibration." + ], + "image_footnote": [], + "bbox": [ + 669, + 387, + 854, + 494 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "eralize to the tasks beyond image classification. In this paper, we conduct a comprehensive study of the calibration of deep semantic segmentation models.", + "bbox": [ + 496, + 641, + 893, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Semantic segmentation tags a semantic label to every pixel in an image. Over the past years, we have witnessed increasingly accurate DNN models [2, 7, 40, 49, 51, 53] for semantic segmentation over various benchmark datasets [1, 14, 38, 44, 50, 54]. The progress has benefited many downstream applications, such as medical imaging and diagnostics, autonomous driving, and robotics. While accuracy is essential for the applications, the segmentation models' uncertainties also provide crucial signals, especially for safety-critical applications — by a segmentation model's uncertainty, we refer to its confidence in the label it assigns to a pixel. For example, an autonomous driving system can use the uncertainty of a semantic segmentation model (e.g., about drivable areas) to make informed decisions.", + "bbox": [ + 496, + 689, + 895, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 808, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Code and models: https://github.com/dwang181/selectivecal", + "bbox": [ + 99, + 886, + 419, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "23652", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing work on the semantic segmentation uncertainty focuses on the medical image domain except [6]. Deep ensemble [8, 21, 27], MC-dropout [16], and stochastic processes [29] are introduced from image classification to calibrate a medical image segmentation model's uncertainty. These methods are applied to the training stage and some could reduce a segmentation model's accuracy. We instead focus on the post-hoc calibration methods that do not change the model's prediction, and our study covers various domains.", + "bbox": [ + 75, + 90, + 472, + 241 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Ding et al.'s work [6] is the most related to ours. While they proposed and validated a novel calibration method for semantic segmentation, our objective is two-fold. We first investigate the uncertainty in semantic segmentation from multiple perspectives, hoping to reveal the major challenges with the segmentation task's uncertainty. This comprehensive investigation and insights are necessary for future research, given the limited existing work in this area. We then propose a simple and effective post-hoc algorithm that outperforms and also can apply to almost all existing calibration methods initially designed for image classification, significantly improving their performance in semantic segmentation.", + "bbox": [ + 75, + 243, + 472, + 439 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To fill the gap of current research on semantic segmentation calibration, we conduct a systematic study. We analyze five state-of-the-art models [24, 40, 49, 51] with six existing calibration methods [6, 11, 19, 25, 27] over seven latest diverse benchmark datasets [1, 14, 38, 39, 44, 50, 54]. First, we investigate different factors to study how they affect miscalibration of semantic segmentation models through experiment. Given the observation, we find that model size, crop size, multi-scaling testing, and misprediction can affect calibration. In particular, segmentation model calibration error is more relevant to misprediction when model size is fixed. Next, we study different existing post-hoc calibrators and introduce our proposed selective scaling algorithm based upon the observation of misprediction. Our selective scaling optimizes ECE by reducing confidence of misprediction, which is reflected by more samples falling into lower confidence region (the far left bin) in Figure 1. Then, we carry out extensive experiments across different models, calibrators, and benchmark datasets. The experiments examine both in-domain calibration and domain-shift calibration to reveal calibration variability across different tasks.", + "bbox": [ + 75, + 441, + 472, + 758 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions can be summarized as follows:", + "bbox": [ + 96, + 761, + 421, + 775 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We conduct a systematic study on the calibration of semantic segmentation models and provide insights to segmentation model calibration. We find that larger models tend to be less calibrated. When model architecture is fixed, larger crop size and multi-scale testing help calibrate models. Moreover, misprediction is important to miscalibration.", + "- We compare different existing popular calibration methods on state-of-the-art semantic segmentation models," + ], + "bbox": [ + 75, + 777, + 468, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "and propose a simple but effective approach, selective scaling, which carries out separate scaling on correct and incorrect predictions.", + "bbox": [ + 496, + 90, + 890, + 136 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We conduct extensive experiments and justify the effectiveness of selective scaling, as shown in Figure 1. We also extend experiments from in-domain to domain-shift data and show that selective scaling consistently outperforms.", + "- We provide useful calibration observations as a comprehensive reference for further research on segmentation model calibration." + ], + "bbox": [ + 496, + 136, + 890, + 241 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 256, + 640, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "DNN calibration. The methods of calibrating DNNs can be categorized into three groups, regularized training, uncertainty estimation, and post-hoc calibration [9]. Regularized training focuses on calibrating DNN over training, such as [30,31,41,46]. Some methods estimate uncertainty using Bayesian methods [15, 18, 47] and deep ensemble [21, 26]. Post-hoc calibration conducts post-processing on the output of DNN to obtain well-calibrated results. While some post-hoc approaches exchange hurting accuracy for better calibration, like histogram binning [11, 36], some retain original accuracy of DNNs and only add calibration map to the last layer of DNNs to adjust probability distribution pattern. Several existing binary calibration maps are extended to multiclass calibration. For example, [11] extended temperature scaling and vector scaling aka Logistic scaling [11], to multiclassss calibration for DNN, and revealed temperature scaling is simple but more effective. To enrich calibration map, Dirichlet scaling [19] was proposed by extending Beta scaling [20] to multiclass DNN calibration. Recently, Meta-cal [25] integrated bipartite-ranking model with selective classification to improve calibration map for better calibrated model. Despite the success of these existing approaches in image classification calibration, there is no work to discuss their performance on semantic segmentation.", + "bbox": [ + 496, + 281, + 890, + 656 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Calibration beyond image classification. Several extensions based upon existing calibrators have been proposed to solve calibration beyond image classification. For example, [29] modeled predictive uncertainty with spatial correlation matrix in stochastic process and incorporated this information into model training for more calibrated segmentation model. [6] proposed local temperature scaling that extends temperature scaling to pixels, and thus, better captures spatial heterogeneity of temperature for a better calibrated segmentation model. [27] conducted model assembling to improve calibration and evaluate the improvement by both conventional and the proposed segment-level uncertainty metrics. [17] addressed semantic segmentation calibration through adversarial training a stochastic network which enables more accurate uncertainty estimation and robust prediction. [8] proposed a multi-headed Variational U-Net by", + "bbox": [ + 496, + 657, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "23653", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "combining ensemble modeling with variational inference and improves model training with better in-distribution calibration and out-of-distribution detection. [32] addressed visual object detector calibration with plug-and-play training-time calibration loss and improved uncertainty optimization.", + "bbox": [ + 75, + 90, + 468, + 179 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Domain-shift calibration. Domain generalization is an important challenge to DNN since model generalization to unseen data could be unpredictable when training and test data distributions differ [43]. In terms of model calibration, performance degradation can also be critical [13, 34]. To alleviate this problem, a variety of approaches have been proposed. [35] corrected covariant shift by importance weighting to achieve better calibrated models. [42] proposed a simple yet effective scheme by perturbing image for classification model calibration improvement. [45] proposed TransCal to improve image classification calibration degradation under the scenario of natural distribution shift. [10] leveraged multiple calibration domain to reduce disparity between source and target domain, thus improving model calibration on shift target domain. Despite some progress in domain-shift calibration, there is still no domain-shift study on semantic segmentation calibration.", + "bbox": [ + 75, + 181, + 470, + 436 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Semantic segmentation. Semantic segmentation is an important visual understanding task. Different from image classification, semantic segmentation focuses on pixel-level classification within an image. Due to this difference, semantic segmentation models are developed in consideration of spatial correlation between pixels. DeepLab family, as an important dilated convolution model, integrates multiscale spatial context into segmentation modeling [2]. SegFormer incorporates hierarchical architectures to encode multiscale features and performs efficient semantic segmentation through ViT structure [49]. Segmenter is a straightforward ViT extension to semantic segmentation by decoding global context of the encoded embedding features from ViT outputs [40]. K-net is a unifying framework for all segmentation tasks [51]. ConvNeXt is proposed as an alternative to ViT and shows accurate segmentation [24].", + "bbox": [ + 75, + 439, + 468, + 680 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminaries", + "text_level": 1, + "bbox": [ + 76, + 694, + 215, + 709 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Similar to image classification, semantic segmentation can be formulated to a multiclass classification problem with a deep neural network. Let $x \\in X$ and $y \\in Y$ denote an input and its label, respectively. A deep neural network $h(x) = (\\hat{y}, \\hat{p})$ yields $\\hat{y}$ as the predicted label with inference confidence $\\hat{p}$ . We expect a well-calibrated model to provide accurate prediction when its confidence is high.", + "bbox": [ + 75, + 719, + 468, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ECE. There are several metrics to measure a model's calibration, and one of the most popular and accepted metrics is expected calibration error (ECE) [33], which reflects the gap between predictive confidence and accuracy. The formal definition with continuous variable is as follows.", + "bbox": [ + 75, + 825, + 468, + 898 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nE C E = \\mathbb {E} _ {\\hat {p}} \\left[ | \\mathbb {P} (\\hat {y} = y | \\hat {p} = p) - p | \\right], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 90, + 890, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\hat{y}$ is the predicted label, $y$ is the true label, $p$ is expected confidence, and $\\mathbb{P}(\\hat{y} = y|\\hat{p} = p)$ is model predictive accuracy. The expectation is about the discrepancy between accuracy and confidence. A perfectly calibrated model has zero ECE.", + "bbox": [ + 496, + 114, + 890, + 189 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In practical problems, statistical binning is used to quantize continuous variables and estimate eq. (1) by equally binning the probability interval,", + "bbox": [ + 496, + 191, + 890, + 236 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\widehat {E C E} = \\sum_ {i = 1} ^ {m} \\frac {\\left| B _ {i} \\right|}{n} | \\operatorname {a c c} \\left(B _ {i}\\right) - \\operatorname {c o n f} \\left(B _ {i}\\right) |, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 558, + 244, + 890, + 282 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $n$ is the number of samples, $m$ is the number of bins, $B_{i}$ denotes a set of samples falling into the bin, and $\\mathrm{acc}(B_i)$ and $\\mathrm{conf}(B_i)$ are accuracy and confidence averaged over the samples in the bin. $\\widehat{ECE}$ can be visualized with the gaps in reliability diagram [5] as Figure 1.", + "bbox": [ + 496, + 291, + 890, + 368 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ECE in semantic segmentation. We extend ECE to semantic segmentation by considering each pixel as a sample. Instead of pooling all pixels of different images into a set, we calculate ECE over each image first before taking an average across images,", + "bbox": [ + 496, + 369, + 890, + 444 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nE C E = \\frac {1}{N} \\sum_ {I = 1} ^ {N} E C E _ {I}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 452, + 890, + 493 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $I$ is an inference image, and $N$ is the total number of images. Note that we adopt $ECE_{I}$ because an image-wise metric is more popular in segmentation model evaluation. Also, $ECE_{I}$ may significantly vary across images.", + "bbox": [ + 496, + 496, + 890, + 556 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Uncertainty of Segmentation Models", + "text_level": 1, + "bbox": [ + 498, + 570, + 830, + 587 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We analyze the uncertainty of five segmentation models, including Segmenter [40], SegFormer [49], Knet-DeepLab [51], Knet-SWIN [51], and ConvNeXt [24], with the ADE20K [54] dataset, a popular semantic segmentation benchmark with 20,210/2,000 training/validation images. The models are state-of-the-art and released by OpenMMLab [3]. Due to limit of space, we focus on Segmenter [40] and SegFormer [49] in this section and report other models' results in the supplementary materials.", + "bbox": [ + 496, + 595, + 890, + 731 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Larger models tending to be less calibrated", + "text_level": 1, + "bbox": [ + 498, + 741, + 864, + 757 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Among the models we analyze, we find that the model capacity is a primary factor determining their generalization accuracy. Large models often lead to high inference accuracy. Will it be valid for semantic segmentation models? To answer it, we conduct experiments on different model capacity by varying model depth and width. We select five variants of SegFormer [49], from B0 to B5 with a deeper and wider architecture, for model scaling-up study. From Figure 2a, we observe that the deeper and wider models,", + "bbox": [ + 496, + 763, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "23654", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/3e88ee4deacdd129bca04ad1635e7e4cbaeabdbb1d7222a19344e69b8a92d8cb.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 76, + 85, + 272, + 224 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2b2e60689a18d4b8cd8e136e78a94cfc3687455aeb8f753f0f0b274517864621.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 272, + 88, + 478, + 224 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/15cb2633b3fc3d4e217446b7d3159a186b32ddfc20714ba90f3969f1e95d9d9b.jpg", + "image_caption": [ + "(c)", + "Figure 2. The effect of model capacity, image crop size, multi-scale testing, and prediction correctness on miscalibration. Prediction correction separately computes the ECE of incorrect/correct predictions. Image-based ECE in eq.(3) is used." + ], + "image_footnote": [], + "bbox": [ + 480, + 88, + 697, + 220 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/818e12223a4d714e78606c2ac26fb979fc9b1d86c898fa15272a65e334167a0d.jpg", + "image_caption": [ + "(d)" + ], + "image_footnote": [], + "bbox": [ + 705, + 88, + 864, + 215 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "like B5, exhibit higher ECE (i.e., worse miscalibration) despite better mIoU. It is consistent with the observation for image classification from Figure 2 in [11], which is caused by the over-confidence during the later stage of NLL optimization [11]. This implies that more accurate prediction leads to worse calibration when it comes to varying model size.", + "bbox": [ + 75, + 285, + 468, + 388 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Larger crop size tending to be more calibrated", + "text_level": 1, + "bbox": [ + 76, + 402, + 468, + 419 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Image cropping is usually used to prepare input data for semantic segmentation models. Crop size determines model input resolution, i.e., the number of pixels for modeling. Larger crop size indicates larger input image, more pixels fed to model, and wider modeling region. Although larger crop size shows higher predictive accuracy, its effect on model calibration is unknown. We conduct experiments to explore their interaction. We select Segmenter-L [40] as a representative model with five crop sizes of 416, 448, 480, 512, and 640 to show the crop size scaling effect. From Figure 2b, we observe that larger crop size yields lower ECE and better calibration. We associate it with modeling field. When crop size increases, larger modeling field provides richer global context information which helps more accurately and confidently locate object boundary and position. This implies that more accurate prediction leads to better calibration in terms of varying crop size.", + "bbox": [ + 75, + 426, + 472, + 684 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.3. Multi-scaling testing improving calibration", + "text_level": 1, + "bbox": [ + 76, + 695, + 442, + 710 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Multi-scale testing is an important technique to boost semantic segmentation model. It usually resizes the input images into multiple scales, feed these resized images into the model, and average all inferences for final prediction. Predictive accuracy is improved after multi-scale testing because multi-scale features help address scale variation between different objects [52]. However, its effectiveness on semantic segmentation calibration is unknown. We conduct experiments with Segmenter-L and cropping size of $640 \\times 640$ of ADE20K to investigate the effect of multiscale testing. The rescale ratios include 0.5, 0.75, 1.0, 1.25, 1.5, and 1.75. Given Figure 2c, with multi-scale testing,", + "bbox": [ + 75, + 719, + 468, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ECE is reduced from 0.86 to 0.78 while mIoU increases from 0.515 to 0.535. This suggests that multi-scale testing enables both alleviating miscalibration and improving accuracy. The intuition is that the aggregation of multi-scale features improves the object boundary delineation and predictive confidence in objects. Also, this ensembling helps correct mispredictions and reduce miscalibration over the over-confidence. This implies that more accurate prediction leads to better calibration concerning multi-scaling testing.", + "bbox": [ + 496, + 284, + 890, + 421 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.4. Mispredictions affecting calibration", + "text_level": 1, + "bbox": [ + 498, + 436, + 808, + 452 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given eqs.(1) and (2), miscalibration results from discrepancy between predictive confidence and accuracy [33]. A perfectly calibrated model should yield predicted probability of 1, i.e., $100\\%$ confidence, for correct prediction, and predicted probability of 0, i.e., infinitesimal confidence, for misprediction. However, an optimal solution is hardly obtained in practice, which implies underconfidence in correct prediction and over-confidence in misprediction. How do these discrepancies affect the calibration property of a segmentation model? We conduct an experiment with SegFormer-B5 [49] and Segmenter-L [40] by grouping correct/incorrect predictions to separately compute their ECEs. This separation can more clearly reveal how much they contribute to final miscalibration. Figure 2d shows that the ECEs from incorrect predictions are significantly higher than those from correct predictions across both models. This implies that misprediction more contributes to miscalibration. This observation also justifies that over-confidence is a more critical problem for calibration [11].", + "bbox": [ + 496, + 460, + 890, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5. Calibrating Semantic Segmentation Model", + "text_level": 1, + "bbox": [ + 500, + 767, + 879, + 784 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We first introduce our post-hoc scaling algorithm in this section. Next, we briefly review the popular existing calibration methods, including four image classification post-hoc calibrators, one semantic segmentation calibrator, and one Bayesian modeling calibrator. We select these calibrators because they exhibit state-of-the-art calibration performance in different groups.", + "bbox": [ + 496, + 794, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "23655", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/fca2166d8e9df29985cc7cbd45e08601fd6379070d951534e431e1f3456f0cb1.jpg", + "image_caption": [ + "Figure 3. Segmented examples from ADE20K, COCO-164K, BDD100K, DAVIS, SpaceNet-7, BraTs-2017, and SYNTHIA (left to right)." + ], + "image_footnote": [], + "bbox": [ + 84, + 88, + 197, + 147 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ae4e04764fa12454d71acaf94b3cfa24c7df7adf90fd73449967c3fc95104527.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 205, + 88, + 305, + 147 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/63564311a8e8345e097787986fbb2191df5bc072a2609c459daa45ed956ec02b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 316, + 88, + 447, + 147 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/37a9fd1a71740da8aa98d661eebe0cee935db0b482016147044e77ddcbaa1dd4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 457, + 88, + 589, + 147 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/04a94016700a1cc17f11b5bd313a7b4b29c5f947a4e8f0f52ccf9095f864f7a4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 601, + 88, + 676, + 147 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ae00f1e2e89a8ce509fcd2713b816352814a6775d07b587bd4e86898757165bb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 686, + 88, + 746, + 147 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/908d5406d7c67a4a9858505546fd34934771c5740a35d8ff1518526b0a6c9612.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 756, + 89, + 880, + 147 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Selective Scaling", + "text_level": 1, + "bbox": [ + 76, + 185, + 240, + 200 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We propose selective scaling given that misprediction is more attributed to miscalibration (Figure 2d). Selective scaling is an extension of selective classification to model calibration. It introduces a binary classifier as a selector to categorize correct and incorrect predictions for separate scaling. For example, selected mispredictions' logits will be smoothed to alleviate the miscalibration from overconfidence.", + "bbox": [ + 75, + 208, + 468, + 328 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l l} \\hat {\\boldsymbol {p}} = \\sigma_ {S M} (\\boldsymbol {z} / T _ {1}), & \\text {i f} \\hat {y} \\neq y \\\\ \\hat {\\boldsymbol {p}} = \\sigma_ {S M} (\\boldsymbol {z} / T _ {2}), & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 337, + 468, + 376 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\hat{\\pmb{p}}$ is calibrated probability vector, $\\sigma_{SM}$ is softmax activation function, $\\pmb{z}$ is the logit vector before activation, $T_{1}$ is a temperature to smooth logit distribution, and $T_{2}$ is a temperature to sharpen logit distribution, so $T_{1} > T_{2}$ . For over-confidence problem, we set $T_{2}$ to 1. It can be extended to under-confidence problem with $T_{2}$ less than 1. The selection of $T_{1}$ relies on the selector accuracy. When its accuracy is higher, $T_{1}$ is larger. Our selective scaling optimizes ECE by reducing confidence of misprediction, which is reflected by more samples falling into the lower confidence region (the far left bin) in reliability diagram (Figure 1).", + "bbox": [ + 75, + 380, + 468, + 545 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Selective scheme was adopted for calibrator design in Meta-cal [25], but different from selective scaling. First, selective scaling does not degrade model accuracy while Meta-cal with miscoverage control does [25]. Different from Meta-cal, selective scaling is independent from entropy ranking model to determine classification threshold. Also, selective scaling obtains a selector without heuristic data resampling used in Meta-cal, which reduces hyperparameters. Selective scaling, to the best of our knowledge, is never used for semantic segmentation model calibration.", + "bbox": [ + 75, + 546, + 468, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2. Existing calibration methods", + "text_level": 1, + "bbox": [ + 76, + 705, + 333, + 722 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Temperature Scaling is a simple but effective approach for multi-classification model calibration [11]. The calibration is carried out with a single temperature parameter to scale logits for overfitting problem resolution.", + "bbox": [ + 75, + 729, + 468, + 790 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\boldsymbol {p}} = \\sigma_ {S M} (\\boldsymbol {z} / T), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 214, + 800, + 468, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $T$ is a scalar of temperature to scale logit vector $\\mathcal{Z}$", + "bbox": [ + 76, + 825, + 455, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Logistic Scaling is an extension of temperature scaling a.k.a. vector scaling [11]. The scaling model is formulated with linear transformation for more complex calibration map.", + "bbox": [ + 75, + 840, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\boldsymbol {p}} = \\sigma_ {S M} (w \\odot \\boldsymbol {z} + b), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 185, + 890, + 200 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $w$ and $b$ are two vectors to scale the logit vector $z$ .", + "bbox": [ + 500, + 203, + 874, + 218 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dirichlet Scaling is the extension of logistic scaling and derived with probability output distribution [19], which enriches calibration map for better optimization with Dirichlet distribution. We adopt linear parameterization [19] for Dirichlet scaling, which is formulated as follows.", + "bbox": [ + 498, + 219, + 890, + 292 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\boldsymbol {p}} = \\sigma_ {S M} (W \\cdot \\log (\\sigma_ {S M} (\\boldsymbol {z})) + b), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 305, + 890, + 321 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $W$ is a matrix and $b$ is a vector for linear parametrisation of the probability $\\sigma_{SM}(z)$ .", + "bbox": [ + 498, + 323, + 890, + 353 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Meta-cal is the scaling strategy derived from bipartite-ranking model and selective classification [25]. With the entropy threshold by a ranking model, probability outputs are separately processed. The probability output with the entropy smaller than threshold is scaled by temperature scaling; otherwise, the output is changed to random prediction. Obviously, it severely degrades model accuracy. Although [25] proposed coverage accuracy control for image classification, it is impractical to semantic segmentation because mIoU is hard to be tuned with calibrator training set.", + "bbox": [ + 498, + 354, + 890, + 503 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l l} \\hat {\\boldsymbol {p}} = \\mathbf {1} / k, & \\text {i f} - \\hat {p} \\log (\\hat {p}) > \\gamma , \\\\ \\hat {\\boldsymbol {p}} = \\sigma_ {S M} (\\boldsymbol {z} / T), & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 512, + 890, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $k$ is the number of classes, $\\gamma$ is the entropy threshold computed with a ranking model for calibration decision.", + "bbox": [ + 498, + 553, + 890, + 582 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Local Temperature Scaling (LTS) is the extension of temperature scaling to semantic segmentation [6]. It provides pixel-level heterogeneous scaling temperature by combing input images with output logit through CNN. Although it considers spatial heterogeneity over an image for scaling, the calibrator training relies on strong correlation between image features and temperature heterogeneity and more training data for feature extraction [6].", + "bbox": [ + 498, + 583, + 890, + 704 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\boldsymbol {p}} = \\sigma_ {S M} (\\boldsymbol {z} / (\\boldsymbol {T} (\\boldsymbol {z}, I)), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 710, + 890, + 727 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\pmb{T}(\\pmb{z}, I)$ is a pixel-wise temperature map by a convolution with segmented logit vector $\\pmb{z}$ and image $I$ .", + "bbox": [ + 498, + 732, + 890, + 762 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Ensembling is proposed to solve medical image segmentation uncertainty prediction [27]. It carries out calibration improvement with simple average of ensemble model, a simplified version of Bayesian inference [47]. Also, deep ensemble is a strong baseline for image classification calibration [21, 47]. We compare its performance with other post-hoc approaches.", + "bbox": [ + 498, + 763, + 890, + 868 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\boldsymbol {p}} = \\frac {1}{N} \\Sigma_ {n = 1} ^ {N} \\hat {\\boldsymbol {p}} ^ {n}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 637, + 875, + 890, + 905 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "23656", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $N$ is the number of ensemble members.", + "bbox": [ + 76, + 90, + 383, + 104 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6. Experiments", + "text_level": 1, + "bbox": [ + 76, + 119, + 209, + 137 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Models. We consider five recent state-of-the-art segmentation models which cover CNN and ViT architectures.", + "bbox": [ + 76, + 145, + 467, + 172 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. SegFormer [49] is a ViT-based encoder-decoder model based upon lightweight multilayer perception (MLP) decoders with pyramid architecture.", + "2. Segmenter [40] is an encoder-decoder model based exclusively on Transformer.", + "3. Knet [51] is a unified segmentation decoder module which includes two important variants, Knet-DeepLab and Knet-SWIN. Knet-DeepLab consists of ResNet-50 [12] backbone and DeepLab-V3 [2] decoder. Knet-SWIN is constructed with SWIN [23] backbone and UperNet [48] decoder.", + "4. ConvNeXt [24] is a CNN backbone, but shaped to ViT-like architecture for better scaling. We select ConvNeXt backbone with UperNet [48] decoder." + ], + "bbox": [ + 89, + 175, + 468, + 386 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For ADE20K, all models except Knet-DeepLab are trained with $640 \\times 640$ crop size. For BDD100K and SYNTHIA, the crop size is $512 \\times 1024$ . For other benchmarks, the crop size is $512 \\times 512$ . The batch size is set to 8.", + "bbox": [ + 76, + 387, + 468, + 446 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset. We examine existing calibration methods across six important benchmarks from various applications. Figure 3 illustrates the example images from benchmarks.", + "bbox": [ + 76, + 446, + 468, + 492 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Scene and stuff segmentation. We use ADE20K [54] and COCO-164K [1] as large-scale segmentation benchmark. ADE20K contains 150 object and stuff classes with 20,210/2,000 images in the training/validation set. COCO-164K, $\\sim 164\\mathrm{K}$ images for 91 stuff and 80 thing classes, includes $118\\mathrm{K} / 5\\mathrm{K}$ images in the training/validation set.", + "2. Autonomous driving. We choose BDD-100K [50], a latest benchmark for urban driving scene segmentation. BDD-100K contains 7K/1K $1280 \\times 720$ images in 19 classes for training/validation set. We select the validation set of CityScapes [4] as test set from target domain for domain-shift calibration assessment. CityScapes has 2975/500 $1024 \\times 2048$ images for training/validation set. BDD-100K and CityScapes share the same label space.", + "3. Video segmentation. We adopt DAVIS2016 [37] to assess calibration on temporal domain shift problems. DAVIS2016 contains 50 object and 1 background classes. We select 480p resolution video sequences to prepare temporal domain shift dataset. The total of the frames in each sequence ranges from 24 to 103. Each frame is an image segmented with a single object and background.", + "4. Remote sensing. We select SPACENET-7 (SN-7) [44], a binary small object identification benchmark, to examine spatial and temporal domain shift calibration." + ], + "bbox": [ + 89, + 493, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "SN-7 consists of 1,408 $1024 \\times 1024$ satellite images collected from monthly urban development across 101 cities. It contains 11,080,000 buildings over 41,000 $\\mathrm{km}^2$ observation area. Due to the development, images vary temporally; due to planning difference, the objects like the building styles differ by locations.", + "bbox": [ + 532, + 90, + 890, + 181 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "5. Medical imaging. We select BraTS2017 [14], a popular brain tumor segmentation dataset. BraTS2017 contains multimodal magnetic resonance imaging (MRI) scans from 1,210 patients, which consists of 3D voxels with 155 layers from top to bottom. Each layer is equivalent to a 2D semantic segmentation image with 3 tumor feature classes.", + "6. Simulation domain transfer. SYNTHIA [39] is a popular urban street synthetic image benchmark. We select its subset of RANDCITYSCAPES, which contains 9400 $1280 \\times 760$ images with 22 classes, as source domain for model training. We select CityScapes [4] as target domain for domain-shift calibration assessment." + ], + "bbox": [ + 511, + 181, + 890, + 392 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In-domain calibration. We select the validation set of each benchmark for in-domain calibration study. For ADE20K, COCO-164K, and BDD100K, we select 100/100/1800 images, 250/250/4500 images, and 50/50/900 images for calibrator training/validation/testing.", + "bbox": [ + 500, + 396, + 890, + 472 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Domain-shift calibration. We select the validation sets of COCO-164K and CityScapes as testing sets of ADE20K and BDD100K. For taxonomy incompatibility over datasets, like ADE20K-COCO164K, we follow the convention in [22] to merge and split label classes. For DAVIS, we merge all training and validation sequences and resplit them in half by frame. The first half of frames are used for training with 1731 images and the second half are split for validation/testing with 100/1624 images. For SN7, all training data are split into 1133/50/225 images for training/validation/testing set by location, and 698/50/660 images for training/validation/testing set by time. For BraTS, we select 800/50 images of $75^{th}$ layer scan for training/validation and 360 images of $100^{th}$ layer scan for testing. For SYNTHIA, we split training set into 9000/400 images for training/validation, and CityScapes' validation set for testing.", + "bbox": [ + 498, + 474, + 892, + 731 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Calibration methods. We select aforementioned six existing calibration methods to compare selective scaling. For selective scaling, selective model is simplified with three-layer MLP due to training data restriction. The inputs are pixel-wise predictive probabilities and the labels are binary which indicate correct/incorrect prediction. Moreover, Meta-cal with miscoverage control yields severe model degradation. We implement our extension by replacing random prediction with large temperature scaling to retain accuracy. For ensembling, we achieve a three-member ensemble with comparable accuracy for fair comparison.", + "bbox": [ + 496, + 734, + 890, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "23657", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/7ab62870cf3a7d7862c14c0e93e9711aad1a79d0f56713f197291f178ffc7560.jpg", + "table_caption": [ + "Table 1. Segmentation accuracy (mIoU) and calibration error (ECE) on different benchmarks. TempS, LogS, DirS, LTS, Ens., and Selective denote temperature, logistic, Dirichlet, local temperature, ensembling, and selective scaling, respectively. Ensembling is carried out by three models with reduced size for comparable mIoU. Meta-Cal* is implemented by our extension with large temperature." + ], + "table_footnote": [], + "table_body": "
DatasetModelmIoUUncalTempSLogSDirSMeta-Cal*LTSEns.Selective
ADE20KSegFormer-B5 [49]49.130.1110.1090.1100.1100.1030.1050.1090.086
ADE20KSegmenter-L [40]51.650.0870.0860.0860.0870.0810.0940.0860.069
ADE20KKnet-DeepLab [51]45.060.1110.1050.1070.1060.1020.1180.1100.095
ADE20KKnet-SWIN-L [51]52.460.0980.0940.0930.0970.0890.1340.0960.078
ADE20KConvNeXt-L [24]53.160.0970.0920.0940.0910.0880.1330.0940.082
COCO-164KSegFormer-B5 [49]45.780.1510.1490.1410.1510.1320.1510.1490.113
COCO-164KSegmenter-L [40]47.090.1520.1490.1490.1510.1300.1550.1500.109
COCO-164KKnet-DeepLab [51]37.240.1700.1700.1680.1710.1490.1720.1690.093
COCO-164KKnet-SWIN-L [51]46.490.1610.1590.1610.1600.1420.1620.1600.102
COCO-164KConvNeXt-L [24]46.480.1600.1570.1580.1590.1410.1620.1590.108
BDD100KSegFormer-B5 [49]65.080.0640.0550.0540.0530.0490.0690.0590.040
BDD100KSegmenter-L [40]61.330.0550.0450.0430.0420.0370.0710.0520.031
BDD100KKnet-DeepLab [51]62.890.0600.0490.0470.0480.0410.0630.0570.035
BDD100KKnet-SWIN-L [51]67.590.0650.0550.0540.0540.0490.0670.0630.040
BDD100KConvNeXt-L [24]67.260.0640.0540.0530.0560.0490.0650.0630.038
", + "bbox": [ + 147, + 133, + 823, + 330 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.1. Results", + "text_level": 1, + "bbox": [ + 76, + 354, + 171, + 369 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.1.1 In-domain calibration evaluation", + "text_level": 1, + "bbox": [ + 76, + 377, + 361, + 392 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We present ECEs for five models and compare selective scaling with six existing calibrators on three-run average. Table 1 shows that majority of calibrators yield very limited improvement in model calibration. However, surprisingly, selective scaling consistently outperforms these scaling approaches. This calibration gain is attributed to more scaling on mispredictions. Moreover, this improvement is more significant for scene and object benchmarks such as ADE20K and COCO-164K than street view datasets like BDD-100K.", + "bbox": [ + 75, + 401, + 468, + 537 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Furthermore, we find that BDD100K yields more calibrated models than other benchmarks. We associate it with larger crop size $(512 \\times 1024)$ and higher model accuracy. We also find that ensembling exhibits weaker calibration. We connect it with model accuracy since we restrain the ensemble accuracy for fair comparison, which limits its calibration performance. Moreover, across models, Segmenter exhibits better calibration. Since it is a Transformer-exclusive model, we link it to different spatial inductive bias which is speculated in [28].", + "bbox": [ + 75, + 537, + 468, + 689 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.1.2 Domain-shift calibration evaluation", + "text_level": 1, + "bbox": [ + 76, + 695, + 380, + 709 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We present ECEs for five models on seven sets of domain-shift experiments and compare selective scaling with six existing calibrators on three-run average. Table 2 shows that selective scaling still consistently outperforms by a large margin. It implies that selective scaling can be generalized to domain-shift problems. In particular, Meta-cal exhibits performance variation. We speculate that the hard threshold from entropy calculation limits separating shift data. In contrast, selective scaling is more consistent because binary classifier has better generalization to separating shift data.", + "bbox": [ + 75, + 719, + 468, + 869 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Across benchmarks, we find that label space shift causes more significant miscalibration. For example, ADE20K", + "bbox": [ + 76, + 869, + 468, + 902 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0a87e2d16e8da0f11dd59031680315baf5004f89dfe560ddbc58eaa3a36bd702.jpg", + "image_caption": [ + "Figure 4. The effect of selector accuracy on calibration. Misprediction detection accuracy denotes the percentage of misprediction detected by selector. The lower bound (optimal case) is $100\\%$ misprediction detected for scaling. The upper bound (worst case) is $0\\%$ , i.e., none misprediction detected is used for calibration." + ], + "image_footnote": [], + "bbox": [ + 535, + 349, + 687, + 473 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9fc311d02c318bf937c1bf3a0be095e56372c57a2f9b995c949127010388b99b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 705, + 351, + 857, + 473 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "to COCO-164K and SYNTHIA to CityScapes show worse ECE than other domain-shift experiments. We link it to poorer domain adaptation and model generalization degradation. Also, we observe that small object segmentation like SN-7, as a challenged task, can cause worse miscalibration. Except for SN-7, more accurate models are more calibrated by benchmarks. For SN-7, background pixels engage in training but are excluded from calibration calculation, which causes larger ECE. Within these models, we find that Segmenter is still more calibrated.", + "bbox": [ + 496, + 568, + 890, + 719 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.1.3 Ablation study", + "text_level": 1, + "bbox": [ + 500, + 729, + 656, + 744 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To analyze the effect of selector on selective scaling, we conduct ablation study by varying misprediction detection accuracy, integrating with different existing calibrators, and examining calibration error across regions.", + "bbox": [ + 496, + 748, + 890, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Selective scaling considering selection accuracy. Figure 4 shows that with the increase in selection accuracy, ECE decreases. This implies that when more mispredictions are detected and scaled, model calibration becomes better. It justifies the effectiveness of selective scaling.", + "bbox": [ + 496, + 809, + 890, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Selective scaling with existing calibrators. We also", + "bbox": [ + 517, + 885, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "23658", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/32dc9c82157ff9f61e931d18550cd62cae726c9d64b62a2cbcb345a055812329.jpg", + "table_caption": [ + "Table 2. Segmentation model accuracy (mIoU) and calibration error (ECE) on different benchmarks for domain-shift calibration assessment. InD and SD denotes in-domain and domain-shift test data. Meta-Cal* is our extension with large temperature scaling." + ], + "table_footnote": [], + "table_body": "
InDSDModelmIoUUncalTempSLogSDirSMeta-Cal*LTSEns.Selective
ADE20KCOCO-164KSegFormer-B5 [49]8.200.4670.4660.4680.4690.4150.4200.4650.331
ADE20KCOCO-164KSegmenter-L [40]9.600.3950.3940.3950.3940.3220.3310.3930.259
ADE20KCOCO-164KKnet-DeepLab [51]7.020.4470.4450.4460.4460.4030.4390.4440.351
ADE20KCOCO-164KKnet-SWIN-L [51]9.060.4640.4620.4630.4630.3790.4650.4620.327
ADE20KCOCO-164KConvNeXt-L [24]8.990.4610.4580.4590.4600.3770.4600.4590.320
BDD100KCityScapesSegFormer-B5 [49]63.050.0870.0830.0820.0840.0760.0850.0870.069
BDD100KCityScapesSegmenter-L [40]60.260.0620.0580.0590.0570.0550.0670.0600.048
BDD100KCityScapesKnet-DeepLab [51]61.270.0730.0700.0690.0710.0630.0730.0710.056
BDD100KCityScapesKnet-SWIN-L [51]67.460.0800.0780.0770.0770.0710.0790.0790.063
BDD100KCityScapesConvNeXt-L [24]67.180.0810.0790.0790.0770.0700.0800.0790.065
DAVIS-trainDAVIS-testSegFormer-B4 [49]89.330.0330.0310.0310.0310.0550.0980.0320.024
DAVIS-trainDAVIS-testSegmenter-B [40]81.350.0800.0770.0760.0750.1030.1210.0780.049
DAVIS-trainDAVIS-testKnet-DeepLab [51]83.120.0760.0720.0730.0730.0980.1070.0750.047
DAVIS-trainDAVIS-testKnet-SWIN-B [51]89.220.0450.0410.0410.0400.0650.1010.0440.029
DAVIS-trainDAVIS-testConvNeXt-B [24]89.000.0440.0400.0410.0420.0670.1000.0440.031
SN-7-SP-trainSN-7-SP-testSegFormer-B4 [49]57.230.6550.6360.6250.6340.7110.7010.6520.532
SN-7-SP-trainSN-7-SP-testSegmenter-B [40]51.980.7660.7510.7480.7400.7960.7890.7630.687
SN-7-SP-trainSN-7-SP-testKnet-DeepLab [51]55.580.7040.6880.6790.6810.7680.7200.6960.655
SN-7-SP-trainSN-7-SP-testKnet-SWIN-B [51]59.420.6090.5760.5590.5630.6310.6180.6050.504
SN-7-SP-trainSN-7-SP-testConvNeXt-B [24]58.690.6110.5700.5640.5660.6290.6330.6080.517
SN-7-TS-trainSN-7-TS-testSegFormer-B4 [49]54.390.7000.6730.6690.6730.7120.7230.6950.582
SN-7-TS-trainSN-7-TS-testSegmenter-B [40]50.130.7690.7450.7400.7360.7850.7770.7550.624
SN-7-TS-trainSN-7-TS-testKnet-DeepLab [51]56.730.6510.6220.6210.6190.6890.6750.6470.576
SN-7-TS-trainSN-7-TS-testKnet-SWIN-B [51]62.420.6190.5870.5790.5840.6300.6210.6090.551
SN-7-TS-trainSN-7-TS-testConvNeXt-B [24]62.140.6230.5780.5710.5730.6470.6380.6200.549
BraTS-trainBraTS-testSegFormer-B4 [49]45.810.1550.1370.1350.1350.1290.1500.1510.120
BraTS-trainBraTS-testSegmenter-B [40]44.570.1540.1490.1460.1470.1350.1420.1500.124
BraTS-trainBraTS-testKnet-DeepLab [51]46.050.1880.1790.1760.1760.1580.1870.1850.148
BraTS-trainBraTS-testKnet-SWIN-B [51]47.990.1550.1460.1430.1440.1320.1570.1520.123
BraTS-trainBraTS-testConvNeXt-B [24]48.400.1690.1590.1540.1550.1360.1680.1650.121
SYNTHIACityScapesSegFormer-B5 [49]33.120.2860.2640.2620.2630.2300.2880.2840.218
SYNTHIACityScapesSegmenter-L [40]32.330.2540.2350.2320.2300.2170.2800.2520.203
SYNTHIACityScapesKnet-DeepLab [51]30.190.3110.2980.2960.2960.2720.3050.3090.241
SYNTHIACityScapesKnet-SWIN-L [51]34.310.2810.2730.2700.2700.2390.2880.2790.226
SYNTHIACityScapesConvNeXt-L [24]34.790.2870.2670.2630.2640.2300.2990.2840.211
", + "bbox": [ + 119, + 119, + 848, + 556 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "find that selective scaling can boost existing calibrators, vice versa. Figure 4 shows that the integration with selective scaling helps existing calibrators to improve model calibration. When detection accuracy is lower, like $< 20\\%$ , existing calibrators can help improve selective scaling.", + "bbox": [ + 75, + 571, + 468, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Selective scaling across regions. Figure 5 shows segmentation calibration error statistics across regions. Top and bottom bars denote maximum and minimum ECE while top, central, and bottom lines of the box indicate $25\\%$ , $50\\%$ , and $75\\%$ of ECE distribution. It shows that: 1) boundary pixels are less calibrated than non-boundary pixels within an object bounding box; 2) selective scaling consistently yields lower mean ECE across both types of pixels.", + "bbox": [ + 75, + 648, + 470, + 770 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d25399fd587bb1b8d1de5c033566ec0cf1ae68ebf821dbe173ab23e45ed594b8.jpg", + "image_caption": [ + "Figure 5. SegFormer-B5 ECE comparison in ADE20K between boundary and non-boundary pixels among uncalibration, temperature scaling, and selective scaling. Crosses and dots show means and outliers. The numbers are the means." + ], + "image_footnote": [], + "bbox": [ + 107, + 784, + 433, + 849 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.2. Limitation", + "text_level": 1, + "bbox": [ + 500, + 571, + 619, + 585 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Logit smoothing can cause entropy increase because post-hoc over-confidence calibration, such as temperature scaling, is equivalent to entropy maximization under certain logit constraints [11]. For example, original Meta-cal leads to significant entropy increase due to randomizing certain prediction outputs. Our method alleviates this issue, but still yields higher entropy. To resolve it, careful temperature tuning is a trade-off scheme for entropy improvement.", + "bbox": [ + 496, + 595, + 890, + 717 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 724, + 617, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We analyzed semantic segmentation model calibration and proposed a simple but effective algorithm of selective scaling. We revealed the possible factors of segmentation model miscalibration. Given the analysis, we designed selective scaling to scale logits separately, and more focus on mispredicted logit smoothing. We performed extensive experiments on state-of-the-art models across various benchmarks. Both in-domain and domain-shift experiments justify the effectiveness of selective scaling. We also presented our observations for segmentation calibration research.", + "bbox": [ + 496, + 750, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "23659", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. Cocostuff: Thing and stuff classes in context. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1209-1218, 2018. 1, 2, 6", + "[2] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In ECCV, 2018. 1, 3, 6", + "[3] MMSegmentation Contributors. MMSegmentation: Openmmlab semantic segmentation toolbox and benchmark. https://github.com/open-mmlab/mmsegmentation, 2020.3", + "[4] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3213-3223, 2016. 6", + "[5] Morris H DeGroot and Stephen E Fienberg. The comparison and evaluation of forecasters. Journal of the Royal Statistical Society: Series D (The Statistician), 32(1-2):12-22, 1983. 1, 3", + "[6] Zhipeng Ding, Xu Han, Peirong Liu, and Marc Niethammer. Local temperature scaling for probability calibration. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6889-6899, 2021. 2, 5", + "[7] Mingyuan Fan, Shenqi Lai, Junshi Huang, Xiaoming Wei, Zhenhua Chai, Junfeng Luo, and Xiaolin Wei. Rethinking bisenet for real-time semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9716-9725, 2021. 1", + "[8] Moritz Fuchs, Camila Gonzalez, and Anirban Mukhopadhyay. Practical uncertainty quantification for brain tumor segmentation. In Medical Imaging with Deep Learning, 2021. 2", + "[9] Jakob Gawlikowski, Cedrique Rovile Njieutcheu Tassi, Mohsin Ali, Jongseok Lee, Matthias Humt, Jianxiang Feng, Anna Kruspe, Rudolph Triebel, Peter Jung, Ribana Roscher, et al. A survey of uncertainty in deep neural networks. arXiv preprint arXiv:2107.03342, 2021. 2", + "[10] Yunye Gong, Xiao Lin, Yi Yao, Thomas G Dietterich, Ajay Divakaran, and Melinda Gervasio. Confidence calibration for domain generalization under covariate shift. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8958-8967, 2021. 3", + "[11] Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q Weinberger. On calibration of modern neural networks. In International conference on machine learning, pages 1321-1330. PMLR, 2017. 1, 2, 4, 5, 8", + "[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6", + "[13] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. arXiv preprint arXiv:1903.12261, 2019. 3" + ], + "bbox": [ + 78, + 114, + 472, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Fabian Isensee, Philipp Kickingereder, Wolfgang Wick, Martin Bendszus, and Klaus H Maier-Hein. Brain tumor segmentation and radiomics survival prediction: Contribution to the brats 2017 challenge. In International MICCAI Brainlesion Workshop, pages 287–297. Springer, 2017. 1, 2, 6", + "[15] Pavel Izmailov, Wesley J Maddox, Polina Kirichenko, Timur Garipov, Dmitry Vetrov, and Andrew Gordon Wilson. Subspace inference for bayesian deep learning. In Uncertainty in Artificial Intelligence, pages 1169-1179. PMLR, 2020. 1, 2", + "[16] Alain Jungo and Mauricio Reyes. Assessing reliability and challenges of uncertainty estimations for medical image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 48-56. Springer, 2019. 2", + "[17] Elias Kassapis, Georgi Dikov, Deepak K Gupta, and Cedric Nugteren. Calibrated adversarial refinement for stochastic semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7057-7067, 2021. 2", + "[18] Alex Kendall and Yarin Gal. What uncertainties do we need in bayesian deep learning for computer vision? Advances in neural information processing systems, 30, 2017. 1, 2", + "[19] Meelis Kull, Miquel Perello Nieto, Markus Kangsepp, Telmo Silva Filho, Hao Song, and Peter Flach. Beyond temperature scaling: Obtaining well-calibrated multi-class probabilities with dirichlet calibration. Advances in neural information processing systems, 32, 2019. 1, 2, 5", + "[20] Meelis Kull, Telmo Silva Filho, and Peter Flach. Beta calibration: a well-founded and easily implemented improvement on logistic calibration for binary classifiers. In Artificial Intelligence and Statistics, pages 623-631. PMLR, 2017. 2", + "[21] Balaji Lakshminarayanan, Alexander Pritzel, and Charles Blundell. Simple and scalable predictive uncertainty estimation using deep ensembles. Advances in neural information processing systems, 30, 2017. 1, 2, 5", + "[22] John Lambert, Zhuang Liu, Ozan Sener, James Hays, and Vladlen Koltun. Mseg: A composite dataset for multi-domain semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2879-2888, 2020. 6", + "[23] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 6", + "[24] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3, 6, 7, 8", + "[25] Xingchen Ma and Matthew B Blaschko. Meta-cal: Well-controlled post-hoc calibration by ranking. In International Conference on Machine Learning, pages 7235-7245. PMLR, 2021. 1, 2, 5" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "23660", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Andrey Malinin, Bruno Mlodozeniec, and Mark Gales. Ensemble distribution distillation. arXiv preprint arXiv:1905.00076, 2019. 2", + "[27] Alireza Mehrtash, William M Wells, Clare M Tempany, Purang Abolmaesumi, and Tina Kapur. Confidence calibration and predictive uncertainty estimation for deep medical image segmentation. IEEE transactions on medical imaging, 39(12):3868-3878, 2020. 2, 5", + "[28] Matthias Minderer, Josip Djolonga, Rob Romijnders, Frances Hubis, Xiaohua Zhai, Neil Houlsby, Dustin Tran, and Mario Lucic. Revisiting the calibration of modern neural networks. Advances in Neural Information Processing Systems, 34:15682-15694, 2021. 7", + "[29] Miguel Monteiro, Loic Le Folgoc, Daniel Coelho de Castro, Nick Pawlowski, Bernardo Marques, Konstantinos Kamnitsas, Mark van der Wilk, and Ben Glocker. Stochastic segmentation networks: Modelling spatially correlated aleatoric uncertainty. Advances in Neural Information Processing Systems, 33:12756-12767, 2020. 2", + "[30] Jishnu Mukhoti, Viveka Kulharia, Amartya Sanyal, Stuart Golodetz, Philip Torr, and Puneet Dokania. Calibrating deep neural networks using focal loss. Advances in Neural Information Processing Systems, 33:15288-15299, 2020. 1, 2", + "[31] Rafael Müller, Simon Kornblith, and Geoffrey E Hinton. When does label smoothing help? Advances in neural information processing systems, 32, 2019. 1, 2", + "[32] Muhammad Akhtar Munir, Muhammad Haris Khan, M Saqib Sarfraz, and Mohsen Ali. Towards improving calibration in object detection under domain shift. arXiv preprint arXiv:2209.07601, 2022. 3", + "[33] Mahdi Pakdaman Naeini, Gregory Cooper, and Milos Hauskrecht. Obtaining well calibrated probabilities using bayesian binning. In Twenty-Ninth AAAI Conference on Artificial Intelligence, 2015. 3, 4", + "[34] Yaniv Ovadia, Emily Fertig, Jie Ren, Zachary Nado, David Sculley, Sebastian Nowozin, Joshua Dillon, Balaji Lakshminarayanan, and Jasper Snoek. Can you trust your model's uncertainty? evaluating predictive uncertainty under dataset shift. Advances in neural information processing systems, 32, 2019. 3", + "[35] Sangdon Park, Osbert Bastani, James Weimer, and Insup Lee. Calibrated prediction with covariate shift via unsupervised domain adaptation. In International Conference on Artificial Intelligence and Statistics, pages 3219-3229. PMLR, 2020. 3", + "[36] Kanil Patel, William Beluch, Bin Yang, Michael Pfeiffer, and Dan Zhang. Multi-class uncertainty calibration via mutual information maximization-based binning. arXiv preprint arXiv:2006.13092, 2020. 1, 2", + "[37] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 724-732, 2016. 6", + "[38] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017. 1, 2", + "[39] German Ros, Laura Sellart, Joanna Materzynska, David Vazquez, and Antonio M Lopez. The synthia dataset: A large collection of synthetic images for semantic segmentation of urban scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3234-3243, 2016. 2, 6", + "[40] Robin Strudel, Ricardo Garcia, Ivan Laptev, and Cordelia Schmid. Segmenter: Transformer for semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7262-7272, 2021. 1, 2, 3, 4, 6, 7, 8", + "[41] Sunil Thulasidasan, Gopinath Chennupati, Jeff A Bilmes, Tanmoy Bhattacharya, and Sarah Michalak. On mixup training: Improved calibration and predictive uncertainty for deep neural networks. Advances in Neural Information Processing Systems, 32, 2019. 1, 2", + "[42] Christian Tomani, Sebastian Gruber, Muhammed Ebrar Erdem, Daniel Cremers, and Florian Buettner. Post-hoc uncertainty calibration for domain drift scenarios. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10124-10132, 2021. 3", + "[43] Antonio Torralba and Alexei A Efros. Unbiased look at dataset bias. In CVPR 2011, pages 1521-1528. IEEE, 2011. 3", + "[44] Adam Van Etten, Daniel Hogan, Jesus Martinez Manso, Jacob Shermeyer, Nicholas Weir, and Ryan Lewis. The multitemporal urban development spacenet dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6398-6407, 2021. 1, 2, 6", + "[45] Ximei Wang, Mingsheng Long, Jianmin Wang, and Michael Jordan. Transferable calibration with lower bias and variance in domain adaptation. Advances in Neural Information Processing Systems, 33:19212-19223, 2020. 3", + "[46] Yezhen Wang, Bo Li, Tong Che, Kaiyang Zhou, Ziwei Liu, and Dongsheng Li. Energy-based open-world uncertainty modeling for confidence calibration. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9302-9311, 2021. 1, 2", + "[47] Andrew G Wilson and Pavel Izmailov. Bayesian deep learning and a probabilistic perspective of generalization. Advances in neural information processing systems, 33:4697-4708, 2020. 1, 2, 5", + "[48] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In Proceedings of the European conference on computer vision (ECCV), pages 418-434, 2018. 6", + "[49] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. arXiv preprint arXiv:2105.15203, 2021. 1, 2, 3, 4, 6, 7, 8", + "[50] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darryll. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2636-2645, 2020. 1, 2, 6" + ], + "bbox": [ + 503, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "23661", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[51] Wenwei Zhang, Jiangmiao Pang, Kai Chen, and Chen Change Loy. K-Net: Towards unified image segmentation. In NeurIPS, 2021. 1, 2, 3, 6, 7, 8", + "[52] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2881-2890, 2017. 4", + "[53] Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xiang, Philip HS Torr, et al. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. arXiv preprint arXiv:2012.15840, 2020. 1", + "[54] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. 1, 2, 3, 6" + ], + "bbox": [ + 78, + 90, + 467, + 330 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "23662", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2023/On Calibrating Semantic Segmentation Models_ Analyses and an Algorithm/f35086d8-d1cb-49d4-972d-c5bbdb5412ed_model.json b/2023/On Calibrating Semantic Segmentation Models_ Analyses and an Algorithm/f35086d8-d1cb-49d4-972d-c5bbdb5412ed_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6158243ae73e2ff665b9e798d843a8c6db3153b4 --- /dev/null +++ b/2023/On Calibrating Semantic Segmentation Models_ Analyses and an Algorithm/f35086d8-d1cb-49d4-972d-c5bbdb5412ed_model.json @@ -0,0 +1,2477 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.809, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.1, + 0.131, + 0.872, + 0.153 + ], + "angle": 0, + "content": "On Calibrating Semantic Segmentation Models: Analyses and An Algorithm" + }, + { + "type": "text", + "bbox": [ + 0.136, + 0.182, + 0.375, + 0.234 + ], + "angle": 0, + "content": "Dongdong Wang \nUniversity of Central Florida \ndaniel.wang@knights.ucf.edu" + }, + { + "type": "text", + "bbox": [ + 0.418, + 0.182, + 0.562, + 0.234 + ], + "angle": 0, + "content": "Boqing Gong Google Research bgong@google.com" + }, + { + "type": "text", + "bbox": [ + 0.606, + 0.182, + 0.837, + 0.234 + ], + "angle": 0, + "content": "Liqiang Wang \nUniversity of Central Florida \nliqiang.wang@ucf.edu" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.302, + 0.474, + 0.604 + ], + "angle": 0, + "content": "We study the problem of semantic segmentation calibration. Lots of solutions have been proposed to approach model miscalibration of confidence in image classification. However, to date, confidence calibration research on semantic segmentation is still limited. We provide a systematic study on the calibration of semantic segmentation models and propose a simple yet effective approach. First, we find that model capacity, crop size, multi-scale testing, and prediction correctness have impact on calibration. Among them, prediction correctness, especially misprediction, is more important to miscalibration due to over-confidence. Next, we propose a simple, unifying, and effective approach, namely selective scaling, by separating correct/incorrect prediction for scaling and more focusing on misprediction logit smoothing. Then, we study popular existing calibration methods and compare them with selective scaling on semantic segmentation calibration. We conduct extensive experiments with a variety of benchmarks on both indomain and domain-shift calibration and show that selective scaling consistently outperforms other methods." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.651, + 0.21, + 0.667 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.677, + 0.47, + 0.829 + ], + "angle": 0, + "content": "Deep neural networks (DNNs) have become the \"goto\" models in various computer vision tasks, such as image classification, object detection, semantic segmentation. However, recent work found that DNNs are often overconfident when they make mistakes [11], misleading downstream applications. To calibrate DNNs' confidence in prediction, researchers have developed a rich line of works on image classification using regularized training [30, 31, 41, 46], post-hoc processing [11, 19, 25, 36], and Bayesian modeling [15, 18, 21, 47], to name a few." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.83, + 0.47, + 0.876 + ], + "angle": 0, + "content": "However, the extensive pursuit of image classification has left it unclear how to calibrate DNNs for other computer vision tasks and how well existing calibration methods gen" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.271, + 0.677, + 0.379 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.388, + 0.677, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.671, + 0.27, + 0.856, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.671, + 0.388, + 0.856, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.671, + 0.496, + 0.735, + 0.506 + ], + "angle": 0, + "content": "Confidence" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.513, + 0.895, + 0.624 + ], + "angle": 0, + "content": "Figure 1. Reliability diagrams [5] of visualized ECE for Segmenter [40]. We randomly select and evaluate fifty validation images from COCO-164K [1] and compare uncalibrated model (Uncal), selective scaling (Selective), temperature scaling (TempS), and ensembling (Ensemble). Smaller gap implies less ECE and better calibration. Ensembling yields comparable accuracy. Despite limited improvement by TempS and Ensemble, Selective exhibits better segmentation model calibration." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.642, + 0.895, + 0.688 + ], + "angle": 0, + "content": "eralize to the tasks beyond image classification. In this paper, we conduct a comprehensive study of the calibration of deep semantic segmentation models." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.69, + 0.896, + 0.901 + ], + "angle": 0, + "content": "Semantic segmentation tags a semantic label to every pixel in an image. Over the past years, we have witnessed increasingly accurate DNN models [2, 7, 40, 49, 51, 53] for semantic segmentation over various benchmark datasets [1, 14, 38, 44, 50, 54]. The progress has benefited many downstream applications, such as medical imaging and diagnostics, autonomous driving, and robotics. While accuracy is essential for the applications, the segmentation models' uncertainties also provide crucial signals, especially for safety-critical applications — by a segmentation model's uncertainty, we refer to its confidence in the label it assigns to a pixel. For example, an autonomous driving system can use the uncertainty of a semantic segmentation model (e.g., about drivable areas) to make informed decisions." + }, + { + "type": "page_footnote", + "bbox": [ + 0.1, + 0.887, + 0.421, + 0.9 + ], + "angle": 0, + "content": "Code and models: https://github.com/dwang181/selectivecal" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23652" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.242 + ], + "angle": 0, + "content": "Existing work on the semantic segmentation uncertainty focuses on the medical image domain except [6]. Deep ensemble [8, 21, 27], MC-dropout [16], and stochastic processes [29] are introduced from image classification to calibrate a medical image segmentation model's uncertainty. These methods are applied to the training stage and some could reduce a segmentation model's accuracy. We instead focus on the post-hoc calibration methods that do not change the model's prediction, and our study covers various domains." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.244, + 0.473, + 0.44 + ], + "angle": 0, + "content": "Ding et al.'s work [6] is the most related to ours. While they proposed and validated a novel calibration method for semantic segmentation, our objective is two-fold. We first investigate the uncertainty in semantic segmentation from multiple perspectives, hoping to reveal the major challenges with the segmentation task's uncertainty. This comprehensive investigation and insights are necessary for future research, given the limited existing work in this area. We then propose a simple and effective post-hoc algorithm that outperforms and also can apply to almost all existing calibration methods initially designed for image classification, significantly improving their performance in semantic segmentation." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.442, + 0.473, + 0.76 + ], + "angle": 0, + "content": "To fill the gap of current research on semantic segmentation calibration, we conduct a systematic study. We analyze five state-of-the-art models [24, 40, 49, 51] with six existing calibration methods [6, 11, 19, 25, 27] over seven latest diverse benchmark datasets [1, 14, 38, 39, 44, 50, 54]. First, we investigate different factors to study how they affect miscalibration of semantic segmentation models through experiment. Given the observation, we find that model size, crop size, multi-scaling testing, and misprediction can affect calibration. In particular, segmentation model calibration error is more relevant to misprediction when model size is fixed. Next, we study different existing post-hoc calibrators and introduce our proposed selective scaling algorithm based upon the observation of misprediction. Our selective scaling optimizes ECE by reducing confidence of misprediction, which is reflected by more samples falling into lower confidence region (the far left bin) in Figure 1. Then, we carry out extensive experiments across different models, calibrators, and benchmark datasets. The experiments examine both in-domain calibration and domain-shift calibration to reveal calibration variability across different tasks." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.762, + 0.423, + 0.776 + ], + "angle": 0, + "content": "Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.778, + 0.47, + 0.869 + ], + "angle": 0, + "content": "- We conduct a systematic study on the calibration of semantic segmentation models and provide insights to segmentation model calibration. We find that larger models tend to be less calibrated. When model architecture is fixed, larger crop size and multi-scale testing help calibrate models. Moreover, misprediction is important to miscalibration." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.47, + 0.902 + ], + "angle": 0, + "content": "- We compare different existing popular calibration methods on state-of-the-art semantic segmentation models," + }, + { + "type": "list", + "bbox": [ + 0.076, + 0.778, + 0.47, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.137 + ], + "angle": 0, + "content": "and propose a simple but effective approach, selective scaling, which carries out separate scaling on correct and incorrect predictions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.137, + 0.892, + 0.198 + ], + "angle": 0, + "content": "- We conduct extensive experiments and justify the effectiveness of selective scaling, as shown in Figure 1. We also extend experiments from in-domain to domain-shift data and show that selective scaling consistently outperforms." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.198, + 0.892, + 0.242 + ], + "angle": 0, + "content": "- We provide useful calibration observations as a comprehensive reference for further research on segmentation model calibration." + }, + { + "type": "list", + "bbox": [ + 0.498, + 0.137, + 0.892, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.257, + 0.642, + 0.273 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.282, + 0.892, + 0.657 + ], + "angle": 0, + "content": "DNN calibration. The methods of calibrating DNNs can be categorized into three groups, regularized training, uncertainty estimation, and post-hoc calibration [9]. Regularized training focuses on calibrating DNN over training, such as [30,31,41,46]. Some methods estimate uncertainty using Bayesian methods [15, 18, 47] and deep ensemble [21, 26]. Post-hoc calibration conducts post-processing on the output of DNN to obtain well-calibrated results. While some post-hoc approaches exchange hurting accuracy for better calibration, like histogram binning [11, 36], some retain original accuracy of DNNs and only add calibration map to the last layer of DNNs to adjust probability distribution pattern. Several existing binary calibration maps are extended to multiclass calibration. For example, [11] extended temperature scaling and vector scaling aka Logistic scaling [11], to multiclassss calibration for DNN, and revealed temperature scaling is simple but more effective. To enrich calibration map, Dirichlet scaling [19] was proposed by extending Beta scaling [20] to multiclass DNN calibration. Recently, Meta-cal [25] integrated bipartite-ranking model with selective classification to improve calibration map for better calibrated model. Despite the success of these existing approaches in image classification calibration, there is no work to discuss their performance on semantic segmentation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.659, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Calibration beyond image classification. Several extensions based upon existing calibrators have been proposed to solve calibration beyond image classification. For example, [29] modeled predictive uncertainty with spatial correlation matrix in stochastic process and incorporated this information into model training for more calibrated segmentation model. [6] proposed local temperature scaling that extends temperature scaling to pixels, and thus, better captures spatial heterogeneity of temperature for a better calibrated segmentation model. [27] conducted model assembling to improve calibration and evaluate the improvement by both conventional and the proposed segment-level uncertainty metrics. [17] addressed semantic segmentation calibration through adversarial training a stochastic network which enables more accurate uncertainty estimation and robust prediction. [8] proposed a multi-headed Variational U-Net by" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23653" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.18 + ], + "angle": 0, + "content": "combining ensemble modeling with variational inference and improves model training with better in-distribution calibration and out-of-distribution detection. [32] addressed visual object detector calibration with plug-and-play training-time calibration loss and improved uncertainty optimization." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.182, + 0.471, + 0.438 + ], + "angle": 0, + "content": "Domain-shift calibration. Domain generalization is an important challenge to DNN since model generalization to unseen data could be unpredictable when training and test data distributions differ [43]. In terms of model calibration, performance degradation can also be critical [13, 34]. To alleviate this problem, a variety of approaches have been proposed. [35] corrected covariant shift by importance weighting to achieve better calibrated models. [42] proposed a simple yet effective scheme by perturbing image for classification model calibration improvement. [45] proposed TransCal to improve image classification calibration degradation under the scenario of natural distribution shift. [10] leveraged multiple calibration domain to reduce disparity between source and target domain, thus improving model calibration on shift target domain. Despite some progress in domain-shift calibration, there is still no domain-shift study on semantic segmentation calibration." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.44, + 0.47, + 0.681 + ], + "angle": 0, + "content": "Semantic segmentation. Semantic segmentation is an important visual understanding task. Different from image classification, semantic segmentation focuses on pixel-level classification within an image. Due to this difference, semantic segmentation models are developed in consideration of spatial correlation between pixels. DeepLab family, as an important dilated convolution model, integrates multiscale spatial context into segmentation modeling [2]. SegFormer incorporates hierarchical architectures to encode multiscale features and performs efficient semantic segmentation through ViT structure [49]. Segmenter is a straightforward ViT extension to semantic segmentation by decoding global context of the encoded embedding features from ViT outputs [40]. K-net is a unifying framework for all segmentation tasks [51]. ConvNeXt is proposed as an alternative to ViT and shows accurate segmentation [24]." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.695, + 0.216, + 0.71 + ], + "angle": 0, + "content": "3. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.469, + 0.825 + ], + "angle": 0, + "content": "Similar to image classification, semantic segmentation can be formulated to a multiclass classification problem with a deep neural network. Let \\( x \\in X \\) and \\( y \\in Y \\) denote an input and its label, respectively. A deep neural network \\( h(x) = (\\hat{y}, \\hat{p}) \\) yields \\( \\hat{y} \\) as the predicted label with inference confidence \\( \\hat{p} \\). We expect a well-calibrated model to provide accurate prediction when its confidence is high." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.826, + 0.469, + 0.899 + ], + "angle": 0, + "content": "ECE. There are several metrics to measure a model's calibration, and one of the most popular and accepted metrics is expected calibration error (ECE) [33], which reflects the gap between predictive confidence and accuracy. The formal definition with continuous variable is as follows." + }, + { + "type": "equation", + "bbox": [ + 0.574, + 0.092, + 0.892, + 0.108 + ], + "angle": 0, + "content": "\\[\nE C E = \\mathbb {E} _ {\\hat {p}} \\left[ | \\mathbb {P} (\\hat {y} = y | \\hat {p} = p) - p | \\right], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.116, + 0.892, + 0.19 + ], + "angle": 0, + "content": "where \\(\\hat{y}\\) is the predicted label, \\(y\\) is the true label, \\(p\\) is expected confidence, and \\(\\mathbb{P}(\\hat{y} = y|\\hat{p} = p)\\) is model predictive accuracy. The expectation is about the discrepancy between accuracy and confidence. A perfectly calibrated model has zero ECE." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.192, + 0.892, + 0.237 + ], + "angle": 0, + "content": "In practical problems, statistical binning is used to quantize continuous variables and estimate eq. (1) by equally binning the probability interval," + }, + { + "type": "equation", + "bbox": [ + 0.56, + 0.245, + 0.892, + 0.284 + ], + "angle": 0, + "content": "\\[\n\\widehat {E C E} = \\sum_ {i = 1} ^ {m} \\frac {\\left| B _ {i} \\right|}{n} | \\operatorname {a c c} \\left(B _ {i}\\right) - \\operatorname {c o n f} \\left(B _ {i}\\right) |, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.292, + 0.892, + 0.369 + ], + "angle": 0, + "content": "where \\( n \\) is the number of samples, \\( m \\) is the number of bins, \\( B_{i} \\) denotes a set of samples falling into the bin, and \\( \\mathrm{acc}(B_i) \\) and \\( \\mathrm{conf}(B_i) \\) are accuracy and confidence averaged over the samples in the bin. \\( \\widehat{ECE} \\) can be visualized with the gaps in reliability diagram [5] as Figure 1." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.37, + 0.892, + 0.445 + ], + "angle": 0, + "content": "ECE in semantic segmentation. We extend ECE to semantic segmentation by considering each pixel as a sample. Instead of pooling all pixels of different images into a set, we calculate ECE over each image first before taking an average across images," + }, + { + "type": "equation", + "bbox": [ + 0.615, + 0.453, + 0.892, + 0.494 + ], + "angle": 0, + "content": "\\[\nE C E = \\frac {1}{N} \\sum_ {I = 1} ^ {N} E C E _ {I}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.497, + 0.892, + 0.558 + ], + "angle": 0, + "content": "where \\(I\\) is an inference image, and \\(N\\) is the total number of images. Note that we adopt \\(ECE_{I}\\) because an image-wise metric is more popular in segmentation model evaluation. Also, \\(ECE_{I}\\) may significantly vary across images." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.571, + 0.831, + 0.588 + ], + "angle": 0, + "content": "4. Uncertainty of Segmentation Models" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.597, + 0.892, + 0.732 + ], + "angle": 0, + "content": "We analyze the uncertainty of five segmentation models, including Segmenter [40], SegFormer [49], Knet-DeepLab [51], Knet-SWIN [51], and ConvNeXt [24], with the ADE20K [54] dataset, a popular semantic segmentation benchmark with 20,210/2,000 training/validation images. The models are state-of-the-art and released by OpenMMLab [3]. Due to limit of space, we focus on Segmenter [40] and SegFormer [49] in this section and report other models' results in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.742, + 0.866, + 0.758 + ], + "angle": 0, + "content": "4.1. Larger models tending to be less calibrated" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Among the models we analyze, we find that the model capacity is a primary factor determining their generalization accuracy. Large models often lead to high inference accuracy. Will it be valid for semantic segmentation models? To answer it, we conduct experiments on different model capacity by varying model depth and width. We select five variants of SegFormer [49], from B0 to B5 with a deeper and wider architecture, for model scaling-up study. From Figure 2a, we observe that the deeper and wider models," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23654" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.087, + 0.273, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.174, + 0.229, + 0.191, + 0.24 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.089, + 0.48, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.381, + 0.229, + 0.399, + 0.241 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.481, + 0.089, + 0.699, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.589, + 0.229, + 0.606, + 0.241 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.706, + 0.089, + 0.865, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.79, + 0.229, + 0.807, + 0.241 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.244, + 0.892, + 0.272 + ], + "angle": 0, + "content": "Figure 2. The effect of model capacity, image crop size, multi-scale testing, and prediction correctness on miscalibration. Prediction correction separately computes the ECE of incorrect/correct predictions. Image-based ECE in eq.(3) is used." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.286, + 0.47, + 0.39 + ], + "angle": 0, + "content": "like B5, exhibit higher ECE (i.e., worse miscalibration) despite better mIoU. It is consistent with the observation for image classification from Figure 2 in [11], which is caused by the over-confidence during the later stage of NLL optimization [11]. This implies that more accurate prediction leads to worse calibration when it comes to varying model size." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.403, + 0.469, + 0.42 + ], + "angle": 0, + "content": "4.2. Larger crop size tending to be more calibrated" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.428, + 0.473, + 0.685 + ], + "angle": 0, + "content": "Image cropping is usually used to prepare input data for semantic segmentation models. Crop size determines model input resolution, i.e., the number of pixels for modeling. Larger crop size indicates larger input image, more pixels fed to model, and wider modeling region. Although larger crop size shows higher predictive accuracy, its effect on model calibration is unknown. We conduct experiments to explore their interaction. We select Segmenter-L [40] as a representative model with five crop sizes of 416, 448, 480, 512, and 640 to show the crop size scaling effect. From Figure 2b, we observe that larger crop size yields lower ECE and better calibration. We associate it with modeling field. When crop size increases, larger modeling field provides richer global context information which helps more accurately and confidently locate object boundary and position. This implies that more accurate prediction leads to better calibration in terms of varying crop size." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.696, + 0.444, + 0.712 + ], + "angle": 0, + "content": "4.3. Multi-scaling testing improving calibration" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Multi-scale testing is an important technique to boost semantic segmentation model. It usually resizes the input images into multiple scales, feed these resized images into the model, and average all inferences for final prediction. Predictive accuracy is improved after multi-scale testing because multi-scale features help address scale variation between different objects [52]. However, its effectiveness on semantic segmentation calibration is unknown. We conduct experiments with Segmenter-L and cropping size of \\(640 \\times 640\\) of ADE20K to investigate the effect of multiscale testing. The rescale ratios include 0.5, 0.75, 1.0, 1.25, 1.5, and 1.75. Given Figure 2c, with multi-scale testing," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.285, + 0.892, + 0.422 + ], + "angle": 0, + "content": "ECE is reduced from 0.86 to 0.78 while mIoU increases from 0.515 to 0.535. This suggests that multi-scale testing enables both alleviating miscalibration and improving accuracy. The intuition is that the aggregation of multi-scale features improves the object boundary delineation and predictive confidence in objects. Also, this ensembling helps correct mispredictions and reduce miscalibration over the over-confidence. This implies that more accurate prediction leads to better calibration concerning multi-scaling testing." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.437, + 0.81, + 0.453 + ], + "angle": 0, + "content": "4.4. Mispredictions affecting calibration" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.462, + 0.892, + 0.75 + ], + "angle": 0, + "content": "Given eqs.(1) and (2), miscalibration results from discrepancy between predictive confidence and accuracy [33]. A perfectly calibrated model should yield predicted probability of 1, i.e., \\(100\\%\\) confidence, for correct prediction, and predicted probability of 0, i.e., infinitesimal confidence, for misprediction. However, an optimal solution is hardly obtained in practice, which implies underconfidence in correct prediction and over-confidence in misprediction. How do these discrepancies affect the calibration property of a segmentation model? We conduct an experiment with SegFormer-B5 [49] and Segmenter-L [40] by grouping correct/incorrect predictions to separately compute their ECEs. This separation can more clearly reveal how much they contribute to final miscalibration. Figure 2d shows that the ECEs from incorrect predictions are significantly higher than those from correct predictions across both models. This implies that misprediction more contributes to miscalibration. This observation also justifies that over-confidence is a more critical problem for calibration [11]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.768, + 0.88, + 0.785 + ], + "angle": 0, + "content": "5. Calibrating Semantic Segmentation Model" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We first introduce our post-hoc scaling algorithm in this section. Next, we briefly review the popular existing calibration methods, including four image classification post-hoc calibrators, one semantic segmentation calibrator, and one Bayesian modeling calibrator. We select these calibrators because they exhibit state-of-the-art calibration performance in different groups." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23655" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.085, + 0.089, + 0.198, + 0.148 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.206, + 0.089, + 0.307, + 0.148 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.318, + 0.089, + 0.449, + 0.148 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.459, + 0.089, + 0.591, + 0.148 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.602, + 0.089, + 0.677, + 0.148 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.687, + 0.089, + 0.748, + 0.148 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.757, + 0.09, + 0.882, + 0.148 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.078, + 0.151, + 0.892, + 0.166 + ], + "angle": 0, + "content": "Figure 3. Segmented examples from ADE20K, COCO-164K, BDD100K, DAVIS, SpaceNet-7, BraTs-2017, and SYNTHIA (left to right)." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.186, + 0.241, + 0.202 + ], + "angle": 0, + "content": "5.1. Selective Scaling" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.209, + 0.47, + 0.329 + ], + "angle": 0, + "content": "We propose selective scaling given that misprediction is more attributed to miscalibration (Figure 2d). Selective scaling is an extension of selective classification to model calibration. It introduces a binary classifier as a selector to categorize correct and incorrect predictions for separate scaling. For example, selected mispredictions' logits will be smoothed to alleviate the miscalibration from overconfidence." + }, + { + "type": "equation", + "bbox": [ + 0.167, + 0.338, + 0.469, + 0.377 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l l} \\hat {\\boldsymbol {p}} = \\sigma_ {S M} (\\boldsymbol {z} / T _ {1}), & \\text {i f} \\hat {y} \\neq y \\\\ \\hat {\\boldsymbol {p}} = \\sigma_ {S M} (\\boldsymbol {z} / T _ {2}), & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.381, + 0.469, + 0.546 + ], + "angle": 0, + "content": "where \\(\\hat{\\pmb{p}}\\) is calibrated probability vector, \\(\\sigma_{SM}\\) is softmax activation function, \\(\\pmb{z}\\) is the logit vector before activation, \\(T_{1}\\) is a temperature to smooth logit distribution, and \\(T_{2}\\) is a temperature to sharpen logit distribution, so \\(T_{1} > T_{2}\\). For over-confidence problem, we set \\(T_{2}\\) to 1. It can be extended to under-confidence problem with \\(T_{2}\\) less than 1. The selection of \\(T_{1}\\) relies on the selector accuracy. When its accuracy is higher, \\(T_{1}\\) is larger. Our selective scaling optimizes ECE by reducing confidence of misprediction, which is reflected by more samples falling into the lower confidence region (the far left bin) in reliability diagram (Figure 1)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.547, + 0.47, + 0.698 + ], + "angle": 0, + "content": "Selective scheme was adopted for calibrator design in Meta-cal [25], but different from selective scaling. First, selective scaling does not degrade model accuracy while Meta-cal with miscoverage control does [25]. Different from Meta-cal, selective scaling is independent from entropy ranking model to determine classification threshold. Also, selective scaling obtains a selector without heuristic data resampling used in Meta-cal, which reduces hyperparameters. Selective scaling, to the best of our knowledge, is never used for semantic segmentation model calibration." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.707, + 0.334, + 0.723 + ], + "angle": 0, + "content": "5.2. Existing calibration methods" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.731, + 0.469, + 0.791 + ], + "angle": 0, + "content": "Temperature Scaling is a simple but effective approach for multi-classification model calibration [11]. The calibration is carried out with a single temperature parameter to scale logits for overfitting problem resolution." + }, + { + "type": "equation", + "bbox": [ + 0.215, + 0.801, + 0.469, + 0.817 + ], + "angle": 0, + "content": "\\[\n\\hat {\\boldsymbol {p}} = \\sigma_ {S M} (\\boldsymbol {z} / T), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.826, + 0.456, + 0.84 + ], + "angle": 0, + "content": "where \\(T\\) is a scalar of temperature to scale logit vector \\(\\mathcal{Z}\\)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.841, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Logistic Scaling is an extension of temperature scaling a.k.a. vector scaling [11]. The scaling model is formulated with linear transformation for more complex calibration map." + }, + { + "type": "equation", + "bbox": [ + 0.618, + 0.186, + 0.891, + 0.202 + ], + "angle": 0, + "content": "\\[\n\\hat {\\boldsymbol {p}} = \\sigma_ {S M} (w \\odot \\boldsymbol {z} + b), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.204, + 0.875, + 0.219 + ], + "angle": 0, + "content": "where \\(w\\) and \\(b\\) are two vectors to scale the logit vector \\(z\\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.22, + 0.892, + 0.294 + ], + "angle": 0, + "content": "Dirichlet Scaling is the extension of logistic scaling and derived with probability output distribution [19], which enriches calibration map for better optimization with Dirichlet distribution. We adopt linear parameterization [19] for Dirichlet scaling, which is formulated as follows." + }, + { + "type": "equation", + "bbox": [ + 0.579, + 0.306, + 0.891, + 0.322 + ], + "angle": 0, + "content": "\\[\n\\hat {\\boldsymbol {p}} = \\sigma_ {S M} (W \\cdot \\log (\\sigma_ {S M} (\\boldsymbol {z})) + b), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.324, + 0.891, + 0.354 + ], + "angle": 0, + "content": "where \\(W\\) is a matrix and \\(b\\) is a vector for linear parametrisation of the probability \\(\\sigma_{SM}(z)\\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.355, + 0.892, + 0.505 + ], + "angle": 0, + "content": "Meta-cal is the scaling strategy derived from bipartite-ranking model and selective classification [25]. With the entropy threshold by a ranking model, probability outputs are separately processed. The probability output with the entropy smaller than threshold is scaled by temperature scaling; otherwise, the output is changed to random prediction. Obviously, it severely degrades model accuracy. Although [25] proposed coverage accuracy control for image classification, it is impractical to semantic segmentation because mIoU is hard to be tuned with calibrator training set." + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.513, + 0.891, + 0.553 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l l} \\hat {\\boldsymbol {p}} = \\mathbf {1} / k, & \\text {i f} - \\hat {p} \\log (\\hat {p}) > \\gamma , \\\\ \\hat {\\boldsymbol {p}} = \\sigma_ {S M} (\\boldsymbol {z} / T), & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.554, + 0.891, + 0.583 + ], + "angle": 0, + "content": "where \\( k \\) is the number of classes, \\( \\gamma \\) is the entropy threshold computed with a ranking model for calibration decision." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.584, + 0.892, + 0.705 + ], + "angle": 0, + "content": "Local Temperature Scaling (LTS) is the extension of temperature scaling to semantic segmentation [6]. It provides pixel-level heterogeneous scaling temperature by combing input images with output logit through CNN. Although it considers spatial heterogeneity over an image for scaling, the calibrator training relies on strong correlation between image features and temperature heterogeneity and more training data for feature extraction [6]." + }, + { + "type": "equation", + "bbox": [ + 0.614, + 0.711, + 0.891, + 0.728 + ], + "angle": 0, + "content": "\\[\n\\hat {\\boldsymbol {p}} = \\sigma_ {S M} (\\boldsymbol {z} / (\\boldsymbol {T} (\\boldsymbol {z}, I)), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.733, + 0.891, + 0.763 + ], + "angle": 0, + "content": "where \\( \\pmb{T}(\\pmb{z}, I) \\) is a pixel-wise temperature map by a convolution with segmented logit vector \\( \\pmb{z} \\) and image \\( I \\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.764, + 0.892, + 0.869 + ], + "angle": 0, + "content": "Ensembling is proposed to solve medical image segmentation uncertainty prediction [27]. It carries out calibration improvement with simple average of ensemble model, a simplified version of Bayesian inference [47]. Also, deep ensemble is a strong baseline for image classification calibration [21, 47]. We compare its performance with other post-hoc approaches." + }, + { + "type": "equation", + "bbox": [ + 0.638, + 0.876, + 0.891, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\hat {\\boldsymbol {p}} = \\frac {1}{N} \\Sigma_ {n = 1} ^ {N} \\hat {\\boldsymbol {p}} ^ {n}, \\tag {10}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23656" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.078, + 0.092, + 0.385, + 0.105 + ], + "angle": 0, + "content": "where \\(N\\) is the number of ensemble members." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.121, + 0.21, + 0.138 + ], + "angle": 0, + "content": "6. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.146, + 0.468, + 0.174 + ], + "angle": 0, + "content": "Models. We consider five recent state-of-the-art segmentation models which cover CNN and ViT architectures." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.176, + 0.469, + 0.221 + ], + "angle": 0, + "content": "1. SegFormer [49] is a ViT-based encoder-decoder model based upon lightweight multilayer perception (MLP) decoders with pyramid architecture." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.222, + 0.469, + 0.251 + ], + "angle": 0, + "content": "2. Segmenter [40] is an encoder-decoder model based exclusively on Transformer." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.252, + 0.469, + 0.342 + ], + "angle": 0, + "content": "3. Knet [51] is a unified segmentation decoder module which includes two important variants, Knet-DeepLab and Knet-SWIN. Knet-DeepLab consists of ResNet-50 [12] backbone and DeepLab-V3 [2] decoder. Knet-SWIN is constructed with SWIN [23] backbone and UperNet [48] decoder." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.343, + 0.469, + 0.387 + ], + "angle": 0, + "content": "4. ConvNeXt [24] is a CNN backbone, but shaped to ViT-like architecture for better scaling. We select ConvNeXt backbone with UperNet [48] decoder." + }, + { + "type": "list", + "bbox": [ + 0.09, + 0.176, + 0.469, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.388, + 0.469, + 0.447 + ], + "angle": 0, + "content": "For ADE20K, all models except Knet-DeepLab are trained with \\(640 \\times 640\\) crop size. For BDD100K and SYNTHIA, the crop size is \\(512 \\times 1024\\). For other benchmarks, the crop size is \\(512 \\times 512\\). The batch size is set to 8." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.448, + 0.469, + 0.493 + ], + "angle": 0, + "content": "Dataset. We examine existing calibration methods across six important benchmarks from various applications. Figure 3 illustrates the example images from benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.494, + 0.469, + 0.599 + ], + "angle": 0, + "content": "1. Scene and stuff segmentation. We use ADE20K [54] and COCO-164K [1] as large-scale segmentation benchmark. ADE20K contains 150 object and stuff classes with 20,210/2,000 images in the training/validation set. COCO-164K, \\(\\sim 164\\mathrm{K}\\) images for 91 stuff and 80 thing classes, includes \\(118\\mathrm{K} / 5\\mathrm{K}\\) images in the training/validation set." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.469, + 0.734 + ], + "angle": 0, + "content": "2. Autonomous driving. We choose BDD-100K [50], a latest benchmark for urban driving scene segmentation. BDD-100K contains 7K/1K \\(1280 \\times 720\\) images in 19 classes for training/validation set. We select the validation set of CityScapes [4] as test set from target domain for domain-shift calibration assessment. CityScapes has 2975/500 \\(1024 \\times 2048\\) images for training/validation set. BDD-100K and CityScapes share the same label space." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.469, + 0.855 + ], + "angle": 0, + "content": "3. Video segmentation. We adopt DAVIS2016 [37] to assess calibration on temporal domain shift problems. DAVIS2016 contains 50 object and 1 background classes. We select 480p resolution video sequences to prepare temporal domain shift dataset. The total of the frames in each sequence ranges from 24 to 103. Each frame is an image segmented with a single object and background." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "4. Remote sensing. We select SPACENET-7 (SN-7) [44], a binary small object identification benchmark, to examine spatial and temporal domain shift calibration." + }, + { + "type": "list", + "bbox": [ + 0.09, + 0.494, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.092, + 0.892, + 0.182 + ], + "angle": 0, + "content": "SN-7 consists of 1,408 \\(1024 \\times 1024\\) satellite images collected from monthly urban development across 101 cities. It contains 11,080,000 buildings over 41,000 \\(\\mathrm{km}^2\\) observation area. Due to the development, images vary temporally; due to planning difference, the objects like the building styles differ by locations." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.183, + 0.892, + 0.287 + ], + "angle": 0, + "content": "5. Medical imaging. We select BraTS2017 [14], a popular brain tumor segmentation dataset. BraTS2017 contains multimodal magnetic resonance imaging (MRI) scans from 1,210 patients, which consists of 3D voxels with 155 layers from top to bottom. Each layer is equivalent to a 2D semantic segmentation image with 3 tumor feature classes." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.288, + 0.892, + 0.393 + ], + "angle": 0, + "content": "6. Simulation domain transfer. SYNTHIA [39] is a popular urban street synthetic image benchmark. We select its subset of RANDCITYSCAPES, which contains 9400 \\(1280 \\times 760\\) images with 22 classes, as source domain for model training. We select CityScapes [4] as target domain for domain-shift calibration assessment." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.183, + 0.892, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.397, + 0.892, + 0.473 + ], + "angle": 0, + "content": "In-domain calibration. We select the validation set of each benchmark for in-domain calibration study. For ADE20K, COCO-164K, and BDD100K, we select 100/100/1800 images, 250/250/4500 images, and 50/50/900 images for calibrator training/validation/testing." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.476, + 0.893, + 0.732 + ], + "angle": 0, + "content": "Domain-shift calibration. We select the validation sets of COCO-164K and CityScapes as testing sets of ADE20K and BDD100K. For taxonomy incompatibility over datasets, like ADE20K-COCO164K, we follow the convention in [22] to merge and split label classes. For DAVIS, we merge all training and validation sequences and resplit them in half by frame. The first half of frames are used for training with 1731 images and the second half are split for validation/testing with 100/1624 images. For SN7, all training data are split into 1133/50/225 images for training/validation/testing set by location, and 698/50/660 images for training/validation/testing set by time. For BraTS, we select 800/50 images of \\(75^{th}\\) layer scan for training/validation and 360 images of \\(100^{th}\\) layer scan for testing. For SYNTHIA, we split training set into 9000/400 images for training/validation, and CityScapes' validation set for testing." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Calibration methods. We select aforementioned six existing calibration methods to compare selective scaling. For selective scaling, selective model is simplified with three-layer MLP due to training data restriction. The inputs are pixel-wise predictive probabilities and the labels are binary which indicate correct/incorrect prediction. Moreover, Meta-cal with miscoverage control yields severe model degradation. We implement our extension by replacing random prediction with large temperature scaling to retain accuracy. For ensembling, we achieve a three-member ensemble with comparable accuracy for fair comparison." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23657" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.895, + 0.133 + ], + "angle": 0, + "content": "Table 1. Segmentation accuracy (mIoU) and calibration error (ECE) on different benchmarks. TempS, LogS, DirS, LTS, Ens., and Selective denote temperature, logistic, Dirichlet, local temperature, ensembling, and selective scaling, respectively. Ensembling is carried out by three models with reduced size for comparable mIoU. Meta-Cal* is implemented by our extension with large temperature." + }, + { + "type": "table", + "bbox": [ + 0.148, + 0.134, + 0.825, + 0.332 + ], + "angle": 0, + "content": "
DatasetModelmIoUUncalTempSLogSDirSMeta-Cal*LTSEns.Selective
ADE20KSegFormer-B5 [49]49.130.1110.1090.1100.1100.1030.1050.1090.086
ADE20KSegmenter-L [40]51.650.0870.0860.0860.0870.0810.0940.0860.069
ADE20KKnet-DeepLab [51]45.060.1110.1050.1070.1060.1020.1180.1100.095
ADE20KKnet-SWIN-L [51]52.460.0980.0940.0930.0970.0890.1340.0960.078
ADE20KConvNeXt-L [24]53.160.0970.0920.0940.0910.0880.1330.0940.082
COCO-164KSegFormer-B5 [49]45.780.1510.1490.1410.1510.1320.1510.1490.113
COCO-164KSegmenter-L [40]47.090.1520.1490.1490.1510.1300.1550.1500.109
COCO-164KKnet-DeepLab [51]37.240.1700.1700.1680.1710.1490.1720.1690.093
COCO-164KKnet-SWIN-L [51]46.490.1610.1590.1610.1600.1420.1620.1600.102
COCO-164KConvNeXt-L [24]46.480.1600.1570.1580.1590.1410.1620.1590.108
BDD100KSegFormer-B5 [49]65.080.0640.0550.0540.0530.0490.0690.0590.040
BDD100KSegmenter-L [40]61.330.0550.0450.0430.0420.0370.0710.0520.031
BDD100KKnet-DeepLab [51]62.890.0600.0490.0470.0480.0410.0630.0570.035
BDD100KKnet-SWIN-L [51]67.590.0650.0550.0540.0540.0490.0670.0630.040
BDD100KConvNeXt-L [24]67.260.0640.0540.0530.0560.0490.0650.0630.038
" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.355, + 0.172, + 0.37 + ], + "angle": 0, + "content": "6.1. Results" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.378, + 0.362, + 0.393 + ], + "angle": 0, + "content": "6.1.1 In-domain calibration evaluation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.402, + 0.469, + 0.538 + ], + "angle": 0, + "content": "We present ECEs for five models and compare selective scaling with six existing calibrators on three-run average. Table 1 shows that majority of calibrators yield very limited improvement in model calibration. However, surprisingly, selective scaling consistently outperforms these scaling approaches. This calibration gain is attributed to more scaling on mispredictions. Moreover, this improvement is more significant for scene and object benchmarks such as ADE20K and COCO-164K than street view datasets like BDD-100K." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.538, + 0.47, + 0.69 + ], + "angle": 0, + "content": "Furthermore, we find that BDD100K yields more calibrated models than other benchmarks. We associate it with larger crop size \\((512 \\times 1024)\\) and higher model accuracy. We also find that ensembling exhibits weaker calibration. We connect it with model accuracy since we restrain the ensemble accuracy for fair comparison, which limits its calibration performance. Moreover, across models, Segmenter exhibits better calibration. Since it is a Transformer-exclusive model, we link it to different spatial inductive bias which is speculated in [28]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.696, + 0.381, + 0.71 + ], + "angle": 0, + "content": "6.1.2 Domain-shift calibration evaluation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.469, + 0.87 + ], + "angle": 0, + "content": "We present ECEs for five models on seven sets of domain-shift experiments and compare selective scaling with six existing calibrators on three-run average. Table 2 shows that selective scaling still consistently outperforms by a large margin. It implies that selective scaling can be generalized to domain-shift problems. In particular, Meta-cal exhibits performance variation. We speculate that the hard threshold from entropy calculation limits separating shift data. In contrast, selective scaling is more consistent because binary classifier has better generalization to separating shift data." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Across benchmarks, we find that label space shift causes more significant miscalibration. For example, ADE20K" + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.351, + 0.688, + 0.474 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.706, + 0.352, + 0.858, + 0.474 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.484, + 0.892, + 0.554 + ], + "angle": 0, + "content": "Figure 4. The effect of selector accuracy on calibration. Misprediction detection accuracy denotes the percentage of misprediction detected by selector. The lower bound (optimal case) is \\(100\\%\\) misprediction detected for scaling. The upper bound (worst case) is \\(0\\%\\), i.e., none misprediction detected is used for calibration." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.569, + 0.892, + 0.72 + ], + "angle": 0, + "content": "to COCO-164K and SYNTHIA to CityScapes show worse ECE than other domain-shift experiments. We link it to poorer domain adaptation and model generalization degradation. Also, we observe that small object segmentation like SN-7, as a challenged task, can cause worse miscalibration. Except for SN-7, more accurate models are more calibrated by benchmarks. For SN-7, background pixels engage in training but are excluded from calibration calculation, which causes larger ECE. Within these models, we find that Segmenter is still more calibrated." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.73, + 0.658, + 0.745 + ], + "angle": 0, + "content": "6.1.3 Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.749, + 0.892, + 0.809 + ], + "angle": 0, + "content": "To analyze the effect of selector on selective scaling, we conduct ablation study by varying misprediction detection accuracy, integrating with different existing calibrators, and examining calibration error across regions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.892, + 0.885 + ], + "angle": 0, + "content": "Selective scaling considering selection accuracy. Figure 4 shows that with the increase in selection accuracy, ECE decreases. This implies that when more mispredictions are detected and scaled, model calibration becomes better. It justifies the effectiveness of selective scaling." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.886, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Selective scaling with existing calibrators. We also" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23658" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.09, + 0.892, + 0.118 + ], + "angle": 0, + "content": "Table 2. Segmentation model accuracy (mIoU) and calibration error (ECE) on different benchmarks for domain-shift calibration assessment. InD and SD denotes in-domain and domain-shift test data. Meta-Cal* is our extension with large temperature scaling." + }, + { + "type": "table", + "bbox": [ + 0.12, + 0.12, + 0.849, + 0.558 + ], + "angle": 0, + "content": "
InDSDModelmIoUUncalTempSLogSDirSMeta-Cal*LTSEns.Selective
ADE20KCOCO-164KSegFormer-B5 [49]8.200.4670.4660.4680.4690.4150.4200.4650.331
ADE20KCOCO-164KSegmenter-L [40]9.600.3950.3940.3950.3940.3220.3310.3930.259
ADE20KCOCO-164KKnet-DeepLab [51]7.020.4470.4450.4460.4460.4030.4390.4440.351
ADE20KCOCO-164KKnet-SWIN-L [51]9.060.4640.4620.4630.4630.3790.4650.4620.327
ADE20KCOCO-164KConvNeXt-L [24]8.990.4610.4580.4590.4600.3770.4600.4590.320
BDD100KCityScapesSegFormer-B5 [49]63.050.0870.0830.0820.0840.0760.0850.0870.069
BDD100KCityScapesSegmenter-L [40]60.260.0620.0580.0590.0570.0550.0670.0600.048
BDD100KCityScapesKnet-DeepLab [51]61.270.0730.0700.0690.0710.0630.0730.0710.056
BDD100KCityScapesKnet-SWIN-L [51]67.460.0800.0780.0770.0770.0710.0790.0790.063
BDD100KCityScapesConvNeXt-L [24]67.180.0810.0790.0790.0770.0700.0800.0790.065
DAVIS-trainDAVIS-testSegFormer-B4 [49]89.330.0330.0310.0310.0310.0550.0980.0320.024
DAVIS-trainDAVIS-testSegmenter-B [40]81.350.0800.0770.0760.0750.1030.1210.0780.049
DAVIS-trainDAVIS-testKnet-DeepLab [51]83.120.0760.0720.0730.0730.0980.1070.0750.047
DAVIS-trainDAVIS-testKnet-SWIN-B [51]89.220.0450.0410.0410.0400.0650.1010.0440.029
DAVIS-trainDAVIS-testConvNeXt-B [24]89.000.0440.0400.0410.0420.0670.1000.0440.031
SN-7-SP-trainSN-7-SP-testSegFormer-B4 [49]57.230.6550.6360.6250.6340.7110.7010.6520.532
SN-7-SP-trainSN-7-SP-testSegmenter-B [40]51.980.7660.7510.7480.7400.7960.7890.7630.687
SN-7-SP-trainSN-7-SP-testKnet-DeepLab [51]55.580.7040.6880.6790.6810.7680.7200.6960.655
SN-7-SP-trainSN-7-SP-testKnet-SWIN-B [51]59.420.6090.5760.5590.5630.6310.6180.6050.504
SN-7-SP-trainSN-7-SP-testConvNeXt-B [24]58.690.6110.5700.5640.5660.6290.6330.6080.517
SN-7-TS-trainSN-7-TS-testSegFormer-B4 [49]54.390.7000.6730.6690.6730.7120.7230.6950.582
SN-7-TS-trainSN-7-TS-testSegmenter-B [40]50.130.7690.7450.7400.7360.7850.7770.7550.624
SN-7-TS-trainSN-7-TS-testKnet-DeepLab [51]56.730.6510.6220.6210.6190.6890.6750.6470.576
SN-7-TS-trainSN-7-TS-testKnet-SWIN-B [51]62.420.6190.5870.5790.5840.6300.6210.6090.551
SN-7-TS-trainSN-7-TS-testConvNeXt-B [24]62.140.6230.5780.5710.5730.6470.6380.6200.549
BraTS-trainBraTS-testSegFormer-B4 [49]45.810.1550.1370.1350.1350.1290.1500.1510.120
BraTS-trainBraTS-testSegmenter-B [40]44.570.1540.1490.1460.1470.1350.1420.1500.124
BraTS-trainBraTS-testKnet-DeepLab [51]46.050.1880.1790.1760.1760.1580.1870.1850.148
BraTS-trainBraTS-testKnet-SWIN-B [51]47.990.1550.1460.1430.1440.1320.1570.1520.123
BraTS-trainBraTS-testConvNeXt-B [24]48.400.1690.1590.1540.1550.1360.1680.1650.121
SYNTHIACityScapesSegFormer-B5 [49]33.120.2860.2640.2620.2630.2300.2880.2840.218
SYNTHIACityScapesSegmenter-L [40]32.330.2540.2350.2320.2300.2170.2800.2520.203
SYNTHIACityScapesKnet-DeepLab [51]30.190.3110.2980.2960.2960.2720.3050.3090.241
SYNTHIACityScapesKnet-SWIN-L [51]34.310.2810.2730.2700.2700.2390.2880.2790.226
SYNTHIACityScapesConvNeXt-L [24]34.790.2870.2670.2630.2640.2300.2990.2840.211
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.573, + 0.47, + 0.649 + ], + "angle": 0, + "content": "find that selective scaling can boost existing calibrators, vice versa. Figure 4 shows that the integration with selective scaling helps existing calibrators to improve model calibration. When detection accuracy is lower, like \\(< 20\\%\\), existing calibrators can help improve selective scaling." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.65, + 0.471, + 0.771 + ], + "angle": 0, + "content": "Selective scaling across regions. Figure 5 shows segmentation calibration error statistics across regions. Top and bottom bars denote maximum and minimum ECE while top, central, and bottom lines of the box indicate \\(25\\%\\), \\(50\\%\\), and \\(75\\%\\) of ECE distribution. It shows that: 1) boundary pixels are less calibrated than non-boundary pixels within an object bounding box; 2) selective scaling consistently yields lower mean ECE across both types of pixels." + }, + { + "type": "image", + "bbox": [ + 0.109, + 0.785, + 0.434, + 0.85 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.854, + 0.47, + 0.909 + ], + "angle": 0, + "content": "Figure 5. SegFormer-B5 ECE comparison in ADE20K between boundary and non-boundary pixels among uncalibration, temperature scaling, and selective scaling. Crosses and dots show means and outliers. The numbers are the means." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.572, + 0.62, + 0.587 + ], + "angle": 0, + "content": "6.2. Limitation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.596, + 0.892, + 0.718 + ], + "angle": 0, + "content": "Logit smoothing can cause entropy increase because post-hoc over-confidence calibration, such as temperature scaling, is equivalent to entropy maximization under certain logit constraints [11]. For example, original Meta-cal leads to significant entropy increase due to randomizing certain prediction outputs. Our method alleviates this issue, but still yields higher entropy. To resolve it, careful temperature tuning is a trade-off scheme for entropy improvement." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.725, + 0.619, + 0.74 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We analyzed semantic segmentation model calibration and proposed a simple but effective algorithm of selective scaling. We revealed the possible factors of segmentation model miscalibration. Given the analysis, we designed selective scaling to scale logits separately, and more focus on mispredicted logit smoothing. We performed extensive experiments on state-of-the-art models across various benchmarks. Both in-domain and domain-shift experiments justify the effectiveness of selective scaling. We also presented our observations for segmentation calibration research." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23659" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.473, + 0.171 + ], + "angle": 0, + "content": "[1] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. Cocostuff: Thing and stuff classes in context. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1209-1218, 2018. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.172, + 0.472, + 0.226 + ], + "angle": 0, + "content": "[2] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In ECCV, 2018. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.228, + 0.472, + 0.284 + ], + "angle": 0, + "content": "[3] MMSegmentation Contributors. MMSegmentation: Openmmlab semantic segmentation toolbox and benchmark. https://github.com/open-mmlab/mmsegmentation, 2020.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.285, + 0.472, + 0.367 + ], + "angle": 0, + "content": "[4] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3213-3223, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.368, + 0.472, + 0.422 + ], + "angle": 0, + "content": "[5] Morris H DeGroot and Stephen E Fienberg. The comparison and evaluation of forecasters. Journal of the Royal Statistical Society: Series D (The Statistician), 32(1-2):12-22, 1983. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.424, + 0.472, + 0.48 + ], + "angle": 0, + "content": "[6] Zhipeng Ding, Xu Han, Peirong Liu, and Marc Niethammer. Local temperature scaling for probability calibration. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6889-6899, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.481, + 0.472, + 0.55 + ], + "angle": 0, + "content": "[7] Mingyuan Fan, Shenqi Lai, Junshi Huang, Xiaoming Wei, Zhenhua Chai, Junfeng Luo, and Xiaolin Wei. Rethinking bisenet for real-time semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9716-9725, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.551, + 0.472, + 0.605 + ], + "angle": 0, + "content": "[8] Moritz Fuchs, Camila Gonzalez, and Anirban Mukhopadhyay. Practical uncertainty quantification for brain tumor segmentation. In Medical Imaging with Deep Learning, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.607, + 0.472, + 0.676 + ], + "angle": 0, + "content": "[9] Jakob Gawlikowski, Cedrique Rovile Njieutcheu Tassi, Mohsin Ali, Jongseok Lee, Matthias Humt, Jianxiang Feng, Anna Kruspe, Rudolph Triebel, Peter Jung, Ribana Roscher, et al. A survey of uncertainty in deep neural networks. arXiv preprint arXiv:2107.03342, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.677, + 0.472, + 0.746 + ], + "angle": 0, + "content": "[10] Yunye Gong, Xiao Lin, Yi Yao, Thomas G Dietterich, Ajay Divakaran, and Melinda Gervasio. Confidence calibration for domain generalization under covariate shift. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8958-8967, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.747, + 0.472, + 0.802 + ], + "angle": 0, + "content": "[11] Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q Weinberger. On calibration of modern neural networks. In International conference on machine learning, pages 1321-1330. PMLR, 2017. 1, 2, 4, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.472, + 0.859 + ], + "angle": 0, + "content": "[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.472, + 0.902 + ], + "angle": 0, + "content": "[13] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. arXiv preprint arXiv:1903.12261, 2019. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.473, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.175 + ], + "angle": 0, + "content": "[14] Fabian Isensee, Philipp Kickingereder, Wolfgang Wick, Martin Bendszus, and Klaus H Maier-Hein. Brain tumor segmentation and radiomics survival prediction: Contribution to the brats 2017 challenge. In International MICCAI Brainlesion Workshop, pages 287–297. Springer, 2017. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.178, + 0.894, + 0.248 + ], + "angle": 0, + "content": "[15] Pavel Izmailov, Wesley J Maddox, Polina Kirichenko, Timur Garipov, Dmitry Vetrov, and Andrew Gordon Wilson. Subspace inference for bayesian deep learning. In Uncertainty in Artificial Intelligence, pages 1169-1179. PMLR, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.251, + 0.894, + 0.32 + ], + "angle": 0, + "content": "[16] Alain Jungo and Mauricio Reyes. Assessing reliability and challenges of uncertainty estimations for medical image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 48-56. Springer, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.323, + 0.894, + 0.392 + ], + "angle": 0, + "content": "[17] Elias Kassapis, Georgi Dikov, Deepak K Gupta, and Cedric Nugteren. Calibrated adversarial refinement for stochastic semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7057-7067, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.395, + 0.893, + 0.438 + ], + "angle": 0, + "content": "[18] Alex Kendall and Yarin Gal. What uncertainties do we need in bayesian deep learning for computer vision? Advances in neural information processing systems, 30, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.44, + 0.893, + 0.509 + ], + "angle": 0, + "content": "[19] Meelis Kull, Miquel Perello Nieto, Markus Kangsepp, Telmo Silva Filho, Hao Song, and Peter Flach. Beyond temperature scaling: Obtaining well-calibrated multi-class probabilities with dirichlet calibration. Advances in neural information processing systems, 32, 2019. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.512, + 0.893, + 0.568 + ], + "angle": 0, + "content": "[20] Meelis Kull, Telmo Silva Filho, and Peter Flach. Beta calibration: a well-founded and easily implemented improvement on logistic calibration for binary classifiers. In Artificial Intelligence and Statistics, pages 623-631. PMLR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.571, + 0.893, + 0.626 + ], + "angle": 0, + "content": "[21] Balaji Lakshminarayanan, Alexander Pritzel, and Charles Blundell. Simple and scalable predictive uncertainty estimation using deep ensembles. Advances in neural information processing systems, 30, 2017. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.629, + 0.894, + 0.699 + ], + "angle": 0, + "content": "[22] John Lambert, Zhuang Liu, Ozan Sener, James Hays, and Vladlen Koltun. Mseg: A composite dataset for multi-domain semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2879-2888, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.701, + 0.893, + 0.77 + ], + "angle": 0, + "content": "[23] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.773, + 0.893, + 0.842 + ], + "angle": 0, + "content": "[24] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.845, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[25] Xingchen Ma and Matthew B Blaschko. Meta-cal: Well-controlled post-hoc calibration by ranking. In International Conference on Machine Learning, pages 7235-7245. PMLR, 2021. 1, 2, 5" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "23660" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.133 + ], + "angle": 0, + "content": "[26] Andrey Malinin, Bruno Mlodozeniec, and Mark Gales. Ensemble distribution distillation. arXiv preprint arXiv:1905.00076, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.136, + 0.47, + 0.205 + ], + "angle": 0, + "content": "[27] Alireza Mehrtash, William M Wells, Clare M Tempany, Purang Abolmaesumi, and Tina Kapur. Confidence calibration and predictive uncertainty estimation for deep medical image segmentation. IEEE transactions on medical imaging, 39(12):3868-3878, 2020. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.207, + 0.469, + 0.275 + ], + "angle": 0, + "content": "[28] Matthias Minderer, Josip Djolonga, Rob Romijnders, Frances Hubis, Xiaohua Zhai, Neil Houlsby, Dustin Tran, and Mario Lucic. Revisiting the calibration of modern neural networks. Advances in Neural Information Processing Systems, 34:15682-15694, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.278, + 0.469, + 0.359 + ], + "angle": 0, + "content": "[29] Miguel Monteiro, Loic Le Folgoc, Daniel Coelho de Castro, Nick Pawlowski, Bernardo Marques, Konstantinos Kamnitsas, Mark van der Wilk, and Ben Glocker. Stochastic segmentation networks: Modelling spatially correlated aleatoric uncertainty. Advances in Neural Information Processing Systems, 33:12756-12767, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.362, + 0.469, + 0.417 + ], + "angle": 0, + "content": "[30] Jishnu Mukhoti, Viveka Kulharia, Amartya Sanyal, Stuart Golodetz, Philip Torr, and Puneet Dokania. Calibrating deep neural networks using focal loss. Advances in Neural Information Processing Systems, 33:15288-15299, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.42, + 0.469, + 0.46 + ], + "angle": 0, + "content": "[31] Rafael Müller, Simon Kornblith, and Geoffrey E Hinton. When does label smoothing help? Advances in neural information processing systems, 32, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.463, + 0.469, + 0.516 + ], + "angle": 0, + "content": "[32] Muhammad Akhtar Munir, Muhammad Haris Khan, M Saqib Sarfraz, and Mohsen Ali. Towards improving calibration in object detection under domain shift. arXiv preprint arXiv:2209.07601, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.519, + 0.469, + 0.574 + ], + "angle": 0, + "content": "[33] Mahdi Pakdaman Naeini, Gregory Cooper, and Milos Hauskrecht. Obtaining well calibrated probabilities using bayesian binning. In Twenty-Ninth AAAI Conference on Artificial Intelligence, 2015. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.577, + 0.469, + 0.658 + ], + "angle": 0, + "content": "[34] Yaniv Ovadia, Emily Fertig, Jie Ren, Zachary Nado, David Sculley, Sebastian Nowozin, Joshua Dillon, Balaji Lakshminarayanan, and Jasper Snoek. Can you trust your model's uncertainty? evaluating predictive uncertainty under dataset shift. Advances in neural information processing systems, 32, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.661, + 0.469, + 0.728 + ], + "angle": 0, + "content": "[35] Sangdon Park, Osbert Bastani, James Weimer, and Insup Lee. Calibrated prediction with covariate shift via unsupervised domain adaptation. In International Conference on Artificial Intelligence and Statistics, pages 3219-3229. PMLR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.732, + 0.469, + 0.786 + ], + "angle": 0, + "content": "[36] Kanil Patel, William Beluch, Bin Yang, Michael Pfeiffer, and Dan Zhang. Multi-class uncertainty calibration via mutual information maximization-based binning. arXiv preprint arXiv:2006.13092, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.789, + 0.469, + 0.87 + ], + "angle": 0, + "content": "[37] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 724-732, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.873, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[38] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.203 + ], + "angle": 0, + "content": "[39] German Ros, Laura Sellart, Joanna Materzynska, David Vazquez, and Antonio M Lopez. The synthia dataset: A large collection of synthetic images for semantic segmentation of urban scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3234-3243, 2016. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.205, + 0.892, + 0.272 + ], + "angle": 0, + "content": "[40] Robin Strudel, Ricardo Garcia, Ivan Laptev, and Cordelia Schmid. Segmenter: Transformer for semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7262-7272, 2021. 1, 2, 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.274, + 0.892, + 0.342 + ], + "angle": 0, + "content": "[41] Sunil Thulasidasan, Gopinath Chennupati, Jeff A Bilmes, Tanmoy Bhattacharya, and Sarah Michalak. On mixup training: Improved calibration and predictive uncertainty for deep neural networks. Advances in Neural Information Processing Systems, 32, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.344, + 0.892, + 0.412 + ], + "angle": 0, + "content": "[42] Christian Tomani, Sebastian Gruber, Muhammed Ebrar Erdem, Daniel Cremers, and Florian Buettner. Post-hoc uncertainty calibration for domain drift scenarios. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10124-10132, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.414, + 0.892, + 0.453 + ], + "angle": 0, + "content": "[43] Antonio Torralba and Alexei A Efros. Unbiased look at dataset bias. In CVPR 2011, pages 1521-1528. IEEE, 2011. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.456, + 0.892, + 0.524 + ], + "angle": 0, + "content": "[44] Adam Van Etten, Daniel Hogan, Jesus Martinez Manso, Jacob Shermeyer, Nicholas Weir, and Ryan Lewis. The multitemporal urban development spacenet dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6398-6407, 2021. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.526, + 0.892, + 0.579 + ], + "angle": 0, + "content": "[45] Ximei Wang, Mingsheng Long, Jianmin Wang, and Michael Jordan. Transferable calibration with lower bias and variance in domain adaptation. Advances in Neural Information Processing Systems, 33:19212-19223, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.581, + 0.892, + 0.649 + ], + "angle": 0, + "content": "[46] Yezhen Wang, Bo Li, Tong Che, Kaiyang Zhou, Ziwei Liu, and Dongsheng Li. Energy-based open-world uncertainty modeling for confidence calibration. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9302-9311, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.651, + 0.892, + 0.704 + ], + "angle": 0, + "content": "[47] Andrew G Wilson and Pavel Izmailov. Bayesian deep learning and a probabilistic perspective of generalization. Advances in neural information processing systems, 33:4697-4708, 2020. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.707, + 0.892, + 0.761 + ], + "angle": 0, + "content": "[48] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In Proceedings of the European conference on computer vision (ECCV), pages 418-434, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.763, + 0.892, + 0.817 + ], + "angle": 0, + "content": "[49] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. arXiv preprint arXiv:2105.15203, 2021. 1, 2, 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.818, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[50] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darryll. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2636-2645, 2020. 1, 2, 6" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23661" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.133 + ], + "angle": 0, + "content": "[51] Wenwei Zhang, Jiangmiao Pang, Kai Chen, and Chen Change Loy. K-Net: Towards unified image segmentation. In NeurIPS, 2021. 1, 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.136, + 0.468, + 0.19 + ], + "angle": 0, + "content": "[52] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2881-2890, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.193, + 0.468, + 0.261 + ], + "angle": 0, + "content": "[53] Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xiang, Philip HS Torr, et al. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. arXiv preprint arXiv:2012.15840, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.264, + 0.468, + 0.331 + ], + "angle": 0, + "content": "[54] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. 1, 2, 3, 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.468, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "23662" + } + ] +] \ No newline at end of file diff --git a/2023/On Calibrating Semantic Segmentation Models_ Analyses and an Algorithm/f35086d8-d1cb-49d4-972d-c5bbdb5412ed_origin.pdf b/2023/On Calibrating Semantic Segmentation Models_ Analyses and an Algorithm/f35086d8-d1cb-49d4-972d-c5bbdb5412ed_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e0a1e662f918be9c3a12b4187bdc28a554c6ed34 --- /dev/null +++ b/2023/On Calibrating Semantic Segmentation Models_ Analyses and an Algorithm/f35086d8-d1cb-49d4-972d-c5bbdb5412ed_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6d54ce627b01e90026c5ad3eafbca34bd74ea0595c2b70ab8dfa3e86048b2bb2 +size 1139112 diff --git a/2023/On Calibrating Semantic Segmentation Models_ Analyses and an Algorithm/full.md b/2023/On Calibrating Semantic Segmentation Models_ Analyses and an Algorithm/full.md new file mode 100644 index 0000000000000000000000000000000000000000..2e6fc81e9828933d7ef4659e55d296341a902607 --- /dev/null +++ b/2023/On Calibrating Semantic Segmentation Models_ Analyses and an Algorithm/full.md @@ -0,0 +1,352 @@ +# On Calibrating Semantic Segmentation Models: Analyses and An Algorithm + +Dongdong Wang +University of Central Florida +daniel.wang@knights.ucf.edu + +Boqing Gong Google Research bgong@google.com + +Liqiang Wang +University of Central Florida +liqiang.wang@ucf.edu + +# Abstract + +We study the problem of semantic segmentation calibration. Lots of solutions have been proposed to approach model miscalibration of confidence in image classification. However, to date, confidence calibration research on semantic segmentation is still limited. We provide a systematic study on the calibration of semantic segmentation models and propose a simple yet effective approach. First, we find that model capacity, crop size, multi-scale testing, and prediction correctness have impact on calibration. Among them, prediction correctness, especially misprediction, is more important to miscalibration due to over-confidence. Next, we propose a simple, unifying, and effective approach, namely selective scaling, by separating correct/incorrect prediction for scaling and more focusing on misprediction logit smoothing. Then, we study popular existing calibration methods and compare them with selective scaling on semantic segmentation calibration. We conduct extensive experiments with a variety of benchmarks on both indomain and domain-shift calibration and show that selective scaling consistently outperforms other methods. + +# 1. Introduction + +Deep neural networks (DNNs) have become the "goto" models in various computer vision tasks, such as image classification, object detection, semantic segmentation. However, recent work found that DNNs are often overconfident when they make mistakes [11], misleading downstream applications. To calibrate DNNs' confidence in prediction, researchers have developed a rich line of works on image classification using regularized training [30, 31, 41, 46], post-hoc processing [11, 19, 25, 36], and Bayesian modeling [15, 18, 21, 47], to name a few. + +However, the extensive pursuit of image classification has left it unclear how to calibrate DNNs for other computer vision tasks and how well existing calibration methods gen + +![](images/45c8dd131d6ad700c2a4266b9c343191802bafb14f5770c9a48d93be122a1cf1.jpg) + +![](images/7b7545543cb0e9d51e55ae6d60580fcc9d0f4f7eaf393c15c76f8b1bb9c0b672.jpg) +Confidence + +![](images/4908090dac040b7671834c84d750998d01811490a309a5c4f1423056ccc836bd.jpg) + +![](images/375aba9e23443008535c3e2200797845360559c534bb40053a39e30b99247d82.jpg) +Figure 1. Reliability diagrams [5] of visualized ECE for Segmenter [40]. We randomly select and evaluate fifty validation images from COCO-164K [1] and compare uncalibrated model (Uncal), selective scaling (Selective), temperature scaling (TempS), and ensembling (Ensemble). Smaller gap implies less ECE and better calibration. Ensembling yields comparable accuracy. Despite limited improvement by TempS and Ensemble, Selective exhibits better segmentation model calibration. + +eralize to the tasks beyond image classification. In this paper, we conduct a comprehensive study of the calibration of deep semantic segmentation models. + +Semantic segmentation tags a semantic label to every pixel in an image. Over the past years, we have witnessed increasingly accurate DNN models [2, 7, 40, 49, 51, 53] for semantic segmentation over various benchmark datasets [1, 14, 38, 44, 50, 54]. The progress has benefited many downstream applications, such as medical imaging and diagnostics, autonomous driving, and robotics. While accuracy is essential for the applications, the segmentation models' uncertainties also provide crucial signals, especially for safety-critical applications — by a segmentation model's uncertainty, we refer to its confidence in the label it assigns to a pixel. For example, an autonomous driving system can use the uncertainty of a semantic segmentation model (e.g., about drivable areas) to make informed decisions. + +Existing work on the semantic segmentation uncertainty focuses on the medical image domain except [6]. Deep ensemble [8, 21, 27], MC-dropout [16], and stochastic processes [29] are introduced from image classification to calibrate a medical image segmentation model's uncertainty. These methods are applied to the training stage and some could reduce a segmentation model's accuracy. We instead focus on the post-hoc calibration methods that do not change the model's prediction, and our study covers various domains. + +Ding et al.'s work [6] is the most related to ours. While they proposed and validated a novel calibration method for semantic segmentation, our objective is two-fold. We first investigate the uncertainty in semantic segmentation from multiple perspectives, hoping to reveal the major challenges with the segmentation task's uncertainty. This comprehensive investigation and insights are necessary for future research, given the limited existing work in this area. We then propose a simple and effective post-hoc algorithm that outperforms and also can apply to almost all existing calibration methods initially designed for image classification, significantly improving their performance in semantic segmentation. + +To fill the gap of current research on semantic segmentation calibration, we conduct a systematic study. We analyze five state-of-the-art models [24, 40, 49, 51] with six existing calibration methods [6, 11, 19, 25, 27] over seven latest diverse benchmark datasets [1, 14, 38, 39, 44, 50, 54]. First, we investigate different factors to study how they affect miscalibration of semantic segmentation models through experiment. Given the observation, we find that model size, crop size, multi-scaling testing, and misprediction can affect calibration. In particular, segmentation model calibration error is more relevant to misprediction when model size is fixed. Next, we study different existing post-hoc calibrators and introduce our proposed selective scaling algorithm based upon the observation of misprediction. Our selective scaling optimizes ECE by reducing confidence of misprediction, which is reflected by more samples falling into lower confidence region (the far left bin) in Figure 1. Then, we carry out extensive experiments across different models, calibrators, and benchmark datasets. The experiments examine both in-domain calibration and domain-shift calibration to reveal calibration variability across different tasks. + +Our contributions can be summarized as follows: + +- We conduct a systematic study on the calibration of semantic segmentation models and provide insights to segmentation model calibration. We find that larger models tend to be less calibrated. When model architecture is fixed, larger crop size and multi-scale testing help calibrate models. Moreover, misprediction is important to miscalibration. +- We compare different existing popular calibration methods on state-of-the-art semantic segmentation models, + +and propose a simple but effective approach, selective scaling, which carries out separate scaling on correct and incorrect predictions. + +- We conduct extensive experiments and justify the effectiveness of selective scaling, as shown in Figure 1. We also extend experiments from in-domain to domain-shift data and show that selective scaling consistently outperforms. +- We provide useful calibration observations as a comprehensive reference for further research on segmentation model calibration. + +# 2. Related Work + +DNN calibration. The methods of calibrating DNNs can be categorized into three groups, regularized training, uncertainty estimation, and post-hoc calibration [9]. Regularized training focuses on calibrating DNN over training, such as [30,31,41,46]. Some methods estimate uncertainty using Bayesian methods [15, 18, 47] and deep ensemble [21, 26]. Post-hoc calibration conducts post-processing on the output of DNN to obtain well-calibrated results. While some post-hoc approaches exchange hurting accuracy for better calibration, like histogram binning [11, 36], some retain original accuracy of DNNs and only add calibration map to the last layer of DNNs to adjust probability distribution pattern. Several existing binary calibration maps are extended to multiclass calibration. For example, [11] extended temperature scaling and vector scaling aka Logistic scaling [11], to multiclassss calibration for DNN, and revealed temperature scaling is simple but more effective. To enrich calibration map, Dirichlet scaling [19] was proposed by extending Beta scaling [20] to multiclass DNN calibration. Recently, Meta-cal [25] integrated bipartite-ranking model with selective classification to improve calibration map for better calibrated model. Despite the success of these existing approaches in image classification calibration, there is no work to discuss their performance on semantic segmentation. + +Calibration beyond image classification. Several extensions based upon existing calibrators have been proposed to solve calibration beyond image classification. For example, [29] modeled predictive uncertainty with spatial correlation matrix in stochastic process and incorporated this information into model training for more calibrated segmentation model. [6] proposed local temperature scaling that extends temperature scaling to pixels, and thus, better captures spatial heterogeneity of temperature for a better calibrated segmentation model. [27] conducted model assembling to improve calibration and evaluate the improvement by both conventional and the proposed segment-level uncertainty metrics. [17] addressed semantic segmentation calibration through adversarial training a stochastic network which enables more accurate uncertainty estimation and robust prediction. [8] proposed a multi-headed Variational U-Net by + +combining ensemble modeling with variational inference and improves model training with better in-distribution calibration and out-of-distribution detection. [32] addressed visual object detector calibration with plug-and-play training-time calibration loss and improved uncertainty optimization. + +Domain-shift calibration. Domain generalization is an important challenge to DNN since model generalization to unseen data could be unpredictable when training and test data distributions differ [43]. In terms of model calibration, performance degradation can also be critical [13, 34]. To alleviate this problem, a variety of approaches have been proposed. [35] corrected covariant shift by importance weighting to achieve better calibrated models. [42] proposed a simple yet effective scheme by perturbing image for classification model calibration improvement. [45] proposed TransCal to improve image classification calibration degradation under the scenario of natural distribution shift. [10] leveraged multiple calibration domain to reduce disparity between source and target domain, thus improving model calibration on shift target domain. Despite some progress in domain-shift calibration, there is still no domain-shift study on semantic segmentation calibration. + +Semantic segmentation. Semantic segmentation is an important visual understanding task. Different from image classification, semantic segmentation focuses on pixel-level classification within an image. Due to this difference, semantic segmentation models are developed in consideration of spatial correlation between pixels. DeepLab family, as an important dilated convolution model, integrates multiscale spatial context into segmentation modeling [2]. SegFormer incorporates hierarchical architectures to encode multiscale features and performs efficient semantic segmentation through ViT structure [49]. Segmenter is a straightforward ViT extension to semantic segmentation by decoding global context of the encoded embedding features from ViT outputs [40]. K-net is a unifying framework for all segmentation tasks [51]. ConvNeXt is proposed as an alternative to ViT and shows accurate segmentation [24]. + +# 3. Preliminaries + +Similar to image classification, semantic segmentation can be formulated to a multiclass classification problem with a deep neural network. Let $x \in X$ and $y \in Y$ denote an input and its label, respectively. A deep neural network $h(x) = (\hat{y}, \hat{p})$ yields $\hat{y}$ as the predicted label with inference confidence $\hat{p}$ . We expect a well-calibrated model to provide accurate prediction when its confidence is high. + +ECE. There are several metrics to measure a model's calibration, and one of the most popular and accepted metrics is expected calibration error (ECE) [33], which reflects the gap between predictive confidence and accuracy. The formal definition with continuous variable is as follows. + +$$ +E C E = \mathbb {E} _ {\hat {p}} \left[ | \mathbb {P} (\hat {y} = y | \hat {p} = p) - p | \right], \tag {1} +$$ + +where $\hat{y}$ is the predicted label, $y$ is the true label, $p$ is expected confidence, and $\mathbb{P}(\hat{y} = y|\hat{p} = p)$ is model predictive accuracy. The expectation is about the discrepancy between accuracy and confidence. A perfectly calibrated model has zero ECE. + +In practical problems, statistical binning is used to quantize continuous variables and estimate eq. (1) by equally binning the probability interval, + +$$ +\widehat {E C E} = \sum_ {i = 1} ^ {m} \frac {\left| B _ {i} \right|}{n} | \operatorname {a c c} \left(B _ {i}\right) - \operatorname {c o n f} \left(B _ {i}\right) |, \tag {2} +$$ + +where $n$ is the number of samples, $m$ is the number of bins, $B_{i}$ denotes a set of samples falling into the bin, and $\mathrm{acc}(B_i)$ and $\mathrm{conf}(B_i)$ are accuracy and confidence averaged over the samples in the bin. $\widehat{ECE}$ can be visualized with the gaps in reliability diagram [5] as Figure 1. + +ECE in semantic segmentation. We extend ECE to semantic segmentation by considering each pixel as a sample. Instead of pooling all pixels of different images into a set, we calculate ECE over each image first before taking an average across images, + +$$ +E C E = \frac {1}{N} \sum_ {I = 1} ^ {N} E C E _ {I}, \tag {3} +$$ + +where $I$ is an inference image, and $N$ is the total number of images. Note that we adopt $ECE_{I}$ because an image-wise metric is more popular in segmentation model evaluation. Also, $ECE_{I}$ may significantly vary across images. + +# 4. Uncertainty of Segmentation Models + +We analyze the uncertainty of five segmentation models, including Segmenter [40], SegFormer [49], Knet-DeepLab [51], Knet-SWIN [51], and ConvNeXt [24], with the ADE20K [54] dataset, a popular semantic segmentation benchmark with 20,210/2,000 training/validation images. The models are state-of-the-art and released by OpenMMLab [3]. Due to limit of space, we focus on Segmenter [40] and SegFormer [49] in this section and report other models' results in the supplementary materials. + +# 4.1. Larger models tending to be less calibrated + +Among the models we analyze, we find that the model capacity is a primary factor determining their generalization accuracy. Large models often lead to high inference accuracy. Will it be valid for semantic segmentation models? To answer it, we conduct experiments on different model capacity by varying model depth and width. We select five variants of SegFormer [49], from B0 to B5 with a deeper and wider architecture, for model scaling-up study. From Figure 2a, we observe that the deeper and wider models, + +![](images/3e88ee4deacdd129bca04ad1635e7e4cbaeabdbb1d7222a19344e69b8a92d8cb.jpg) +(a) + +![](images/2b2e60689a18d4b8cd8e136e78a94cfc3687455aeb8f753f0f0b274517864621.jpg) +(b) + +![](images/15cb2633b3fc3d4e217446b7d3159a186b32ddfc20714ba90f3969f1e95d9d9b.jpg) +(c) +Figure 2. The effect of model capacity, image crop size, multi-scale testing, and prediction correctness on miscalibration. Prediction correction separately computes the ECE of incorrect/correct predictions. Image-based ECE in eq.(3) is used. + +![](images/818e12223a4d714e78606c2ac26fb979fc9b1d86c898fa15272a65e334167a0d.jpg) +(d) + +like B5, exhibit higher ECE (i.e., worse miscalibration) despite better mIoU. It is consistent with the observation for image classification from Figure 2 in [11], which is caused by the over-confidence during the later stage of NLL optimization [11]. This implies that more accurate prediction leads to worse calibration when it comes to varying model size. + +# 4.2. Larger crop size tending to be more calibrated + +Image cropping is usually used to prepare input data for semantic segmentation models. Crop size determines model input resolution, i.e., the number of pixels for modeling. Larger crop size indicates larger input image, more pixels fed to model, and wider modeling region. Although larger crop size shows higher predictive accuracy, its effect on model calibration is unknown. We conduct experiments to explore their interaction. We select Segmenter-L [40] as a representative model with five crop sizes of 416, 448, 480, 512, and 640 to show the crop size scaling effect. From Figure 2b, we observe that larger crop size yields lower ECE and better calibration. We associate it with modeling field. When crop size increases, larger modeling field provides richer global context information which helps more accurately and confidently locate object boundary and position. This implies that more accurate prediction leads to better calibration in terms of varying crop size. + +# 4.3. Multi-scaling testing improving calibration + +Multi-scale testing is an important technique to boost semantic segmentation model. It usually resizes the input images into multiple scales, feed these resized images into the model, and average all inferences for final prediction. Predictive accuracy is improved after multi-scale testing because multi-scale features help address scale variation between different objects [52]. However, its effectiveness on semantic segmentation calibration is unknown. We conduct experiments with Segmenter-L and cropping size of $640 \times 640$ of ADE20K to investigate the effect of multiscale testing. The rescale ratios include 0.5, 0.75, 1.0, 1.25, 1.5, and 1.75. Given Figure 2c, with multi-scale testing, + +ECE is reduced from 0.86 to 0.78 while mIoU increases from 0.515 to 0.535. This suggests that multi-scale testing enables both alleviating miscalibration and improving accuracy. The intuition is that the aggregation of multi-scale features improves the object boundary delineation and predictive confidence in objects. Also, this ensembling helps correct mispredictions and reduce miscalibration over the over-confidence. This implies that more accurate prediction leads to better calibration concerning multi-scaling testing. + +# 4.4. Mispredictions affecting calibration + +Given eqs.(1) and (2), miscalibration results from discrepancy between predictive confidence and accuracy [33]. A perfectly calibrated model should yield predicted probability of 1, i.e., $100\%$ confidence, for correct prediction, and predicted probability of 0, i.e., infinitesimal confidence, for misprediction. However, an optimal solution is hardly obtained in practice, which implies underconfidence in correct prediction and over-confidence in misprediction. How do these discrepancies affect the calibration property of a segmentation model? We conduct an experiment with SegFormer-B5 [49] and Segmenter-L [40] by grouping correct/incorrect predictions to separately compute their ECEs. This separation can more clearly reveal how much they contribute to final miscalibration. Figure 2d shows that the ECEs from incorrect predictions are significantly higher than those from correct predictions across both models. This implies that misprediction more contributes to miscalibration. This observation also justifies that over-confidence is a more critical problem for calibration [11]. + +# 5. Calibrating Semantic Segmentation Model + +We first introduce our post-hoc scaling algorithm in this section. Next, we briefly review the popular existing calibration methods, including four image classification post-hoc calibrators, one semantic segmentation calibrator, and one Bayesian modeling calibrator. We select these calibrators because they exhibit state-of-the-art calibration performance in different groups. + +![](images/fca2166d8e9df29985cc7cbd45e08601fd6379070d951534e431e1f3456f0cb1.jpg) +Figure 3. Segmented examples from ADE20K, COCO-164K, BDD100K, DAVIS, SpaceNet-7, BraTs-2017, and SYNTHIA (left to right). + +![](images/ae4e04764fa12454d71acaf94b3cfa24c7df7adf90fd73449967c3fc95104527.jpg) + +![](images/63564311a8e8345e097787986fbb2191df5bc072a2609c459daa45ed956ec02b.jpg) + +![](images/37a9fd1a71740da8aa98d661eebe0cee935db0b482016147044e77ddcbaa1dd4.jpg) + +![](images/04a94016700a1cc17f11b5bd313a7b4b29c5f947a4e8f0f52ccf9095f864f7a4.jpg) + +![](images/ae00f1e2e89a8ce509fcd2713b816352814a6775d07b587bd4e86898757165bb.jpg) + +![](images/908d5406d7c67a4a9858505546fd34934771c5740a35d8ff1518526b0a6c9612.jpg) + +# 5.1. Selective Scaling + +We propose selective scaling given that misprediction is more attributed to miscalibration (Figure 2d). Selective scaling is an extension of selective classification to model calibration. It introduces a binary classifier as a selector to categorize correct and incorrect predictions for separate scaling. For example, selected mispredictions' logits will be smoothed to alleviate the miscalibration from overconfidence. + +$$ +\left\{ \begin{array}{l l} \hat {\boldsymbol {p}} = \sigma_ {S M} (\boldsymbol {z} / T _ {1}), & \text {i f} \hat {y} \neq y \\ \hat {\boldsymbol {p}} = \sigma_ {S M} (\boldsymbol {z} / T _ {2}), & \text {o t h e r w i s e} \end{array} \right. \tag {4} +$$ + +where $\hat{\pmb{p}}$ is calibrated probability vector, $\sigma_{SM}$ is softmax activation function, $\pmb{z}$ is the logit vector before activation, $T_{1}$ is a temperature to smooth logit distribution, and $T_{2}$ is a temperature to sharpen logit distribution, so $T_{1} > T_{2}$ . For over-confidence problem, we set $T_{2}$ to 1. It can be extended to under-confidence problem with $T_{2}$ less than 1. The selection of $T_{1}$ relies on the selector accuracy. When its accuracy is higher, $T_{1}$ is larger. Our selective scaling optimizes ECE by reducing confidence of misprediction, which is reflected by more samples falling into the lower confidence region (the far left bin) in reliability diagram (Figure 1). + +Selective scheme was adopted for calibrator design in Meta-cal [25], but different from selective scaling. First, selective scaling does not degrade model accuracy while Meta-cal with miscoverage control does [25]. Different from Meta-cal, selective scaling is independent from entropy ranking model to determine classification threshold. Also, selective scaling obtains a selector without heuristic data resampling used in Meta-cal, which reduces hyperparameters. Selective scaling, to the best of our knowledge, is never used for semantic segmentation model calibration. + +# 5.2. Existing calibration methods + +Temperature Scaling is a simple but effective approach for multi-classification model calibration [11]. The calibration is carried out with a single temperature parameter to scale logits for overfitting problem resolution. + +$$ +\hat {\boldsymbol {p}} = \sigma_ {S M} (\boldsymbol {z} / T), \tag {5} +$$ + +where $T$ is a scalar of temperature to scale logit vector $\mathcal{Z}$ + +Logistic Scaling is an extension of temperature scaling a.k.a. vector scaling [11]. The scaling model is formulated with linear transformation for more complex calibration map. + +$$ +\hat {\boldsymbol {p}} = \sigma_ {S M} (w \odot \boldsymbol {z} + b), \tag {6} +$$ + +where $w$ and $b$ are two vectors to scale the logit vector $z$ . + +Dirichlet Scaling is the extension of logistic scaling and derived with probability output distribution [19], which enriches calibration map for better optimization with Dirichlet distribution. We adopt linear parameterization [19] for Dirichlet scaling, which is formulated as follows. + +$$ +\hat {\boldsymbol {p}} = \sigma_ {S M} (W \cdot \log (\sigma_ {S M} (\boldsymbol {z})) + b), \tag {7} +$$ + +where $W$ is a matrix and $b$ is a vector for linear parametrisation of the probability $\sigma_{SM}(z)$ . + +Meta-cal is the scaling strategy derived from bipartite-ranking model and selective classification [25]. With the entropy threshold by a ranking model, probability outputs are separately processed. The probability output with the entropy smaller than threshold is scaled by temperature scaling; otherwise, the output is changed to random prediction. Obviously, it severely degrades model accuracy. Although [25] proposed coverage accuracy control for image classification, it is impractical to semantic segmentation because mIoU is hard to be tuned with calibrator training set. + +$$ +\left\{ \begin{array}{l l} \hat {\boldsymbol {p}} = \mathbf {1} / k, & \text {i f} - \hat {p} \log (\hat {p}) > \gamma , \\ \hat {\boldsymbol {p}} = \sigma_ {S M} (\boldsymbol {z} / T), & \text {o t h e r w i s e}, \end{array} \right. \tag {8} +$$ + +where $k$ is the number of classes, $\gamma$ is the entropy threshold computed with a ranking model for calibration decision. + +Local Temperature Scaling (LTS) is the extension of temperature scaling to semantic segmentation [6]. It provides pixel-level heterogeneous scaling temperature by combing input images with output logit through CNN. Although it considers spatial heterogeneity over an image for scaling, the calibrator training relies on strong correlation between image features and temperature heterogeneity and more training data for feature extraction [6]. + +$$ +\hat {\boldsymbol {p}} = \sigma_ {S M} (\boldsymbol {z} / (\boldsymbol {T} (\boldsymbol {z}, I)), \tag {9} +$$ + +where $\pmb{T}(\pmb{z}, I)$ is a pixel-wise temperature map by a convolution with segmented logit vector $\pmb{z}$ and image $I$ . + +Ensembling is proposed to solve medical image segmentation uncertainty prediction [27]. It carries out calibration improvement with simple average of ensemble model, a simplified version of Bayesian inference [47]. Also, deep ensemble is a strong baseline for image classification calibration [21, 47]. We compare its performance with other post-hoc approaches. + +$$ +\hat {\boldsymbol {p}} = \frac {1}{N} \Sigma_ {n = 1} ^ {N} \hat {\boldsymbol {p}} ^ {n}, \tag {10} +$$ + +where $N$ is the number of ensemble members. + +# 6. Experiments + +Models. We consider five recent state-of-the-art segmentation models which cover CNN and ViT architectures. + +1. SegFormer [49] is a ViT-based encoder-decoder model based upon lightweight multilayer perception (MLP) decoders with pyramid architecture. +2. Segmenter [40] is an encoder-decoder model based exclusively on Transformer. +3. Knet [51] is a unified segmentation decoder module which includes two important variants, Knet-DeepLab and Knet-SWIN. Knet-DeepLab consists of ResNet-50 [12] backbone and DeepLab-V3 [2] decoder. Knet-SWIN is constructed with SWIN [23] backbone and UperNet [48] decoder. +4. ConvNeXt [24] is a CNN backbone, but shaped to ViT-like architecture for better scaling. We select ConvNeXt backbone with UperNet [48] decoder. + +For ADE20K, all models except Knet-DeepLab are trained with $640 \times 640$ crop size. For BDD100K and SYNTHIA, the crop size is $512 \times 1024$ . For other benchmarks, the crop size is $512 \times 512$ . The batch size is set to 8. + +Dataset. We examine existing calibration methods across six important benchmarks from various applications. Figure 3 illustrates the example images from benchmarks. + +1. Scene and stuff segmentation. We use ADE20K [54] and COCO-164K [1] as large-scale segmentation benchmark. ADE20K contains 150 object and stuff classes with 20,210/2,000 images in the training/validation set. COCO-164K, $\sim 164\mathrm{K}$ images for 91 stuff and 80 thing classes, includes $118\mathrm{K} / 5\mathrm{K}$ images in the training/validation set. +2. Autonomous driving. We choose BDD-100K [50], a latest benchmark for urban driving scene segmentation. BDD-100K contains 7K/1K $1280 \times 720$ images in 19 classes for training/validation set. We select the validation set of CityScapes [4] as test set from target domain for domain-shift calibration assessment. CityScapes has 2975/500 $1024 \times 2048$ images for training/validation set. BDD-100K and CityScapes share the same label space. +3. Video segmentation. We adopt DAVIS2016 [37] to assess calibration on temporal domain shift problems. DAVIS2016 contains 50 object and 1 background classes. We select 480p resolution video sequences to prepare temporal domain shift dataset. The total of the frames in each sequence ranges from 24 to 103. Each frame is an image segmented with a single object and background. +4. Remote sensing. We select SPACENET-7 (SN-7) [44], a binary small object identification benchmark, to examine spatial and temporal domain shift calibration. + +SN-7 consists of 1,408 $1024 \times 1024$ satellite images collected from monthly urban development across 101 cities. It contains 11,080,000 buildings over 41,000 $\mathrm{km}^2$ observation area. Due to the development, images vary temporally; due to planning difference, the objects like the building styles differ by locations. + +5. Medical imaging. We select BraTS2017 [14], a popular brain tumor segmentation dataset. BraTS2017 contains multimodal magnetic resonance imaging (MRI) scans from 1,210 patients, which consists of 3D voxels with 155 layers from top to bottom. Each layer is equivalent to a 2D semantic segmentation image with 3 tumor feature classes. +6. Simulation domain transfer. SYNTHIA [39] is a popular urban street synthetic image benchmark. We select its subset of RANDCITYSCAPES, which contains 9400 $1280 \times 760$ images with 22 classes, as source domain for model training. We select CityScapes [4] as target domain for domain-shift calibration assessment. + +In-domain calibration. We select the validation set of each benchmark for in-domain calibration study. For ADE20K, COCO-164K, and BDD100K, we select 100/100/1800 images, 250/250/4500 images, and 50/50/900 images for calibrator training/validation/testing. + +Domain-shift calibration. We select the validation sets of COCO-164K and CityScapes as testing sets of ADE20K and BDD100K. For taxonomy incompatibility over datasets, like ADE20K-COCO164K, we follow the convention in [22] to merge and split label classes. For DAVIS, we merge all training and validation sequences and resplit them in half by frame. The first half of frames are used for training with 1731 images and the second half are split for validation/testing with 100/1624 images. For SN7, all training data are split into 1133/50/225 images for training/validation/testing set by location, and 698/50/660 images for training/validation/testing set by time. For BraTS, we select 800/50 images of $75^{th}$ layer scan for training/validation and 360 images of $100^{th}$ layer scan for testing. For SYNTHIA, we split training set into 9000/400 images for training/validation, and CityScapes' validation set for testing. + +Calibration methods. We select aforementioned six existing calibration methods to compare selective scaling. For selective scaling, selective model is simplified with three-layer MLP due to training data restriction. The inputs are pixel-wise predictive probabilities and the labels are binary which indicate correct/incorrect prediction. Moreover, Meta-cal with miscoverage control yields severe model degradation. We implement our extension by replacing random prediction with large temperature scaling to retain accuracy. For ensembling, we achieve a three-member ensemble with comparable accuracy for fair comparison. + +Table 1. Segmentation accuracy (mIoU) and calibration error (ECE) on different benchmarks. TempS, LogS, DirS, LTS, Ens., and Selective denote temperature, logistic, Dirichlet, local temperature, ensembling, and selective scaling, respectively. Ensembling is carried out by three models with reduced size for comparable mIoU. Meta-Cal* is implemented by our extension with large temperature. + +
DatasetModelmIoUUncalTempSLogSDirSMeta-Cal*LTSEns.Selective
ADE20KSegFormer-B5 [49]49.130.1110.1090.1100.1100.1030.1050.1090.086
ADE20KSegmenter-L [40]51.650.0870.0860.0860.0870.0810.0940.0860.069
ADE20KKnet-DeepLab [51]45.060.1110.1050.1070.1060.1020.1180.1100.095
ADE20KKnet-SWIN-L [51]52.460.0980.0940.0930.0970.0890.1340.0960.078
ADE20KConvNeXt-L [24]53.160.0970.0920.0940.0910.0880.1330.0940.082
COCO-164KSegFormer-B5 [49]45.780.1510.1490.1410.1510.1320.1510.1490.113
COCO-164KSegmenter-L [40]47.090.1520.1490.1490.1510.1300.1550.1500.109
COCO-164KKnet-DeepLab [51]37.240.1700.1700.1680.1710.1490.1720.1690.093
COCO-164KKnet-SWIN-L [51]46.490.1610.1590.1610.1600.1420.1620.1600.102
COCO-164KConvNeXt-L [24]46.480.1600.1570.1580.1590.1410.1620.1590.108
BDD100KSegFormer-B5 [49]65.080.0640.0550.0540.0530.0490.0690.0590.040
BDD100KSegmenter-L [40]61.330.0550.0450.0430.0420.0370.0710.0520.031
BDD100KKnet-DeepLab [51]62.890.0600.0490.0470.0480.0410.0630.0570.035
BDD100KKnet-SWIN-L [51]67.590.0650.0550.0540.0540.0490.0670.0630.040
BDD100KConvNeXt-L [24]67.260.0640.0540.0530.0560.0490.0650.0630.038
+ +# 6.1. Results + +# 6.1.1 In-domain calibration evaluation + +We present ECEs for five models and compare selective scaling with six existing calibrators on three-run average. Table 1 shows that majority of calibrators yield very limited improvement in model calibration. However, surprisingly, selective scaling consistently outperforms these scaling approaches. This calibration gain is attributed to more scaling on mispredictions. Moreover, this improvement is more significant for scene and object benchmarks such as ADE20K and COCO-164K than street view datasets like BDD-100K. + +Furthermore, we find that BDD100K yields more calibrated models than other benchmarks. We associate it with larger crop size $(512 \times 1024)$ and higher model accuracy. We also find that ensembling exhibits weaker calibration. We connect it with model accuracy since we restrain the ensemble accuracy for fair comparison, which limits its calibration performance. Moreover, across models, Segmenter exhibits better calibration. Since it is a Transformer-exclusive model, we link it to different spatial inductive bias which is speculated in [28]. + +# 6.1.2 Domain-shift calibration evaluation + +We present ECEs for five models on seven sets of domain-shift experiments and compare selective scaling with six existing calibrators on three-run average. Table 2 shows that selective scaling still consistently outperforms by a large margin. It implies that selective scaling can be generalized to domain-shift problems. In particular, Meta-cal exhibits performance variation. We speculate that the hard threshold from entropy calculation limits separating shift data. In contrast, selective scaling is more consistent because binary classifier has better generalization to separating shift data. + +Across benchmarks, we find that label space shift causes more significant miscalibration. For example, ADE20K + +![](images/0a87e2d16e8da0f11dd59031680315baf5004f89dfe560ddbc58eaa3a36bd702.jpg) +Figure 4. The effect of selector accuracy on calibration. Misprediction detection accuracy denotes the percentage of misprediction detected by selector. The lower bound (optimal case) is $100\%$ misprediction detected for scaling. The upper bound (worst case) is $0\%$ , i.e., none misprediction detected is used for calibration. + +![](images/9fc311d02c318bf937c1bf3a0be095e56372c57a2f9b995c949127010388b99b.jpg) + +to COCO-164K and SYNTHIA to CityScapes show worse ECE than other domain-shift experiments. We link it to poorer domain adaptation and model generalization degradation. Also, we observe that small object segmentation like SN-7, as a challenged task, can cause worse miscalibration. Except for SN-7, more accurate models are more calibrated by benchmarks. For SN-7, background pixels engage in training but are excluded from calibration calculation, which causes larger ECE. Within these models, we find that Segmenter is still more calibrated. + +# 6.1.3 Ablation study + +To analyze the effect of selector on selective scaling, we conduct ablation study by varying misprediction detection accuracy, integrating with different existing calibrators, and examining calibration error across regions. + +Selective scaling considering selection accuracy. Figure 4 shows that with the increase in selection accuracy, ECE decreases. This implies that when more mispredictions are detected and scaled, model calibration becomes better. It justifies the effectiveness of selective scaling. + +Selective scaling with existing calibrators. We also + +Table 2. Segmentation model accuracy (mIoU) and calibration error (ECE) on different benchmarks for domain-shift calibration assessment. InD and SD denotes in-domain and domain-shift test data. Meta-Cal* is our extension with large temperature scaling. + +
InDSDModelmIoUUncalTempSLogSDirSMeta-Cal*LTSEns.Selective
ADE20KCOCO-164KSegFormer-B5 [49]8.200.4670.4660.4680.4690.4150.4200.4650.331
ADE20KCOCO-164KSegmenter-L [40]9.600.3950.3940.3950.3940.3220.3310.3930.259
ADE20KCOCO-164KKnet-DeepLab [51]7.020.4470.4450.4460.4460.4030.4390.4440.351
ADE20KCOCO-164KKnet-SWIN-L [51]9.060.4640.4620.4630.4630.3790.4650.4620.327
ADE20KCOCO-164KConvNeXt-L [24]8.990.4610.4580.4590.4600.3770.4600.4590.320
BDD100KCityScapesSegFormer-B5 [49]63.050.0870.0830.0820.0840.0760.0850.0870.069
BDD100KCityScapesSegmenter-L [40]60.260.0620.0580.0590.0570.0550.0670.0600.048
BDD100KCityScapesKnet-DeepLab [51]61.270.0730.0700.0690.0710.0630.0730.0710.056
BDD100KCityScapesKnet-SWIN-L [51]67.460.0800.0780.0770.0770.0710.0790.0790.063
BDD100KCityScapesConvNeXt-L [24]67.180.0810.0790.0790.0770.0700.0800.0790.065
DAVIS-trainDAVIS-testSegFormer-B4 [49]89.330.0330.0310.0310.0310.0550.0980.0320.024
DAVIS-trainDAVIS-testSegmenter-B [40]81.350.0800.0770.0760.0750.1030.1210.0780.049
DAVIS-trainDAVIS-testKnet-DeepLab [51]83.120.0760.0720.0730.0730.0980.1070.0750.047
DAVIS-trainDAVIS-testKnet-SWIN-B [51]89.220.0450.0410.0410.0400.0650.1010.0440.029
DAVIS-trainDAVIS-testConvNeXt-B [24]89.000.0440.0400.0410.0420.0670.1000.0440.031
SN-7-SP-trainSN-7-SP-testSegFormer-B4 [49]57.230.6550.6360.6250.6340.7110.7010.6520.532
SN-7-SP-trainSN-7-SP-testSegmenter-B [40]51.980.7660.7510.7480.7400.7960.7890.7630.687
SN-7-SP-trainSN-7-SP-testKnet-DeepLab [51]55.580.7040.6880.6790.6810.7680.7200.6960.655
SN-7-SP-trainSN-7-SP-testKnet-SWIN-B [51]59.420.6090.5760.5590.5630.6310.6180.6050.504
SN-7-SP-trainSN-7-SP-testConvNeXt-B [24]58.690.6110.5700.5640.5660.6290.6330.6080.517
SN-7-TS-trainSN-7-TS-testSegFormer-B4 [49]54.390.7000.6730.6690.6730.7120.7230.6950.582
SN-7-TS-trainSN-7-TS-testSegmenter-B [40]50.130.7690.7450.7400.7360.7850.7770.7550.624
SN-7-TS-trainSN-7-TS-testKnet-DeepLab [51]56.730.6510.6220.6210.6190.6890.6750.6470.576
SN-7-TS-trainSN-7-TS-testKnet-SWIN-B [51]62.420.6190.5870.5790.5840.6300.6210.6090.551
SN-7-TS-trainSN-7-TS-testConvNeXt-B [24]62.140.6230.5780.5710.5730.6470.6380.6200.549
BraTS-trainBraTS-testSegFormer-B4 [49]45.810.1550.1370.1350.1350.1290.1500.1510.120
BraTS-trainBraTS-testSegmenter-B [40]44.570.1540.1490.1460.1470.1350.1420.1500.124
BraTS-trainBraTS-testKnet-DeepLab [51]46.050.1880.1790.1760.1760.1580.1870.1850.148
BraTS-trainBraTS-testKnet-SWIN-B [51]47.990.1550.1460.1430.1440.1320.1570.1520.123
BraTS-trainBraTS-testConvNeXt-B [24]48.400.1690.1590.1540.1550.1360.1680.1650.121
SYNTHIACityScapesSegFormer-B5 [49]33.120.2860.2640.2620.2630.2300.2880.2840.218
SYNTHIACityScapesSegmenter-L [40]32.330.2540.2350.2320.2300.2170.2800.2520.203
SYNTHIACityScapesKnet-DeepLab [51]30.190.3110.2980.2960.2960.2720.3050.3090.241
SYNTHIACityScapesKnet-SWIN-L [51]34.310.2810.2730.2700.2700.2390.2880.2790.226
SYNTHIACityScapesConvNeXt-L [24]34.790.2870.2670.2630.2640.2300.2990.2840.211
+ +find that selective scaling can boost existing calibrators, vice versa. Figure 4 shows that the integration with selective scaling helps existing calibrators to improve model calibration. When detection accuracy is lower, like $< 20\%$ , existing calibrators can help improve selective scaling. + +Selective scaling across regions. Figure 5 shows segmentation calibration error statistics across regions. Top and bottom bars denote maximum and minimum ECE while top, central, and bottom lines of the box indicate $25\%$ , $50\%$ , and $75\%$ of ECE distribution. It shows that: 1) boundary pixels are less calibrated than non-boundary pixels within an object bounding box; 2) selective scaling consistently yields lower mean ECE across both types of pixels. + +![](images/d25399fd587bb1b8d1de5c033566ec0cf1ae68ebf821dbe173ab23e45ed594b8.jpg) +Figure 5. SegFormer-B5 ECE comparison in ADE20K between boundary and non-boundary pixels among uncalibration, temperature scaling, and selective scaling. Crosses and dots show means and outliers. The numbers are the means. + +# 6.2. Limitation + +Logit smoothing can cause entropy increase because post-hoc over-confidence calibration, such as temperature scaling, is equivalent to entropy maximization under certain logit constraints [11]. For example, original Meta-cal leads to significant entropy increase due to randomizing certain prediction outputs. Our method alleviates this issue, but still yields higher entropy. To resolve it, careful temperature tuning is a trade-off scheme for entropy improvement. + +# 7. Conclusion + +We analyzed semantic segmentation model calibration and proposed a simple but effective algorithm of selective scaling. We revealed the possible factors of segmentation model miscalibration. Given the analysis, we designed selective scaling to scale logits separately, and more focus on mispredicted logit smoothing. We performed extensive experiments on state-of-the-art models across various benchmarks. Both in-domain and domain-shift experiments justify the effectiveness of selective scaling. We also presented our observations for segmentation calibration research. + +# References + +[1] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. Cocostuff: Thing and stuff classes in context. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1209-1218, 2018. 1, 2, 6 +[2] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In ECCV, 2018. 1, 3, 6 +[3] MMSegmentation Contributors. MMSegmentation: Openmmlab semantic segmentation toolbox and benchmark. https://github.com/open-mmlab/mmsegmentation, 2020.3 +[4] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3213-3223, 2016. 6 +[5] Morris H DeGroot and Stephen E Fienberg. The comparison and evaluation of forecasters. Journal of the Royal Statistical Society: Series D (The Statistician), 32(1-2):12-22, 1983. 1, 3 +[6] Zhipeng Ding, Xu Han, Peirong Liu, and Marc Niethammer. Local temperature scaling for probability calibration. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6889-6899, 2021. 2, 5 +[7] Mingyuan Fan, Shenqi Lai, Junshi Huang, Xiaoming Wei, Zhenhua Chai, Junfeng Luo, and Xiaolin Wei. Rethinking bisenet for real-time semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9716-9725, 2021. 1 +[8] Moritz Fuchs, Camila Gonzalez, and Anirban Mukhopadhyay. Practical uncertainty quantification for brain tumor segmentation. In Medical Imaging with Deep Learning, 2021. 2 +[9] Jakob Gawlikowski, Cedrique Rovile Njieutcheu Tassi, Mohsin Ali, Jongseok Lee, Matthias Humt, Jianxiang Feng, Anna Kruspe, Rudolph Triebel, Peter Jung, Ribana Roscher, et al. A survey of uncertainty in deep neural networks. arXiv preprint arXiv:2107.03342, 2021. 2 +[10] Yunye Gong, Xiao Lin, Yi Yao, Thomas G Dietterich, Ajay Divakaran, and Melinda Gervasio. Confidence calibration for domain generalization under covariate shift. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8958-8967, 2021. 3 +[11] Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q Weinberger. On calibration of modern neural networks. In International conference on machine learning, pages 1321-1330. PMLR, 2017. 1, 2, 4, 5, 8 +[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6 +[13] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. arXiv preprint arXiv:1903.12261, 2019. 3 + +[14] Fabian Isensee, Philipp Kickingereder, Wolfgang Wick, Martin Bendszus, and Klaus H Maier-Hein. Brain tumor segmentation and radiomics survival prediction: Contribution to the brats 2017 challenge. In International MICCAI Brainlesion Workshop, pages 287–297. Springer, 2017. 1, 2, 6 +[15] Pavel Izmailov, Wesley J Maddox, Polina Kirichenko, Timur Garipov, Dmitry Vetrov, and Andrew Gordon Wilson. Subspace inference for bayesian deep learning. In Uncertainty in Artificial Intelligence, pages 1169-1179. PMLR, 2020. 1, 2 +[16] Alain Jungo and Mauricio Reyes. Assessing reliability and challenges of uncertainty estimations for medical image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 48-56. Springer, 2019. 2 +[17] Elias Kassapis, Georgi Dikov, Deepak K Gupta, and Cedric Nugteren. Calibrated adversarial refinement for stochastic semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7057-7067, 2021. 2 +[18] Alex Kendall and Yarin Gal. What uncertainties do we need in bayesian deep learning for computer vision? Advances in neural information processing systems, 30, 2017. 1, 2 +[19] Meelis Kull, Miquel Perello Nieto, Markus Kangsepp, Telmo Silva Filho, Hao Song, and Peter Flach. Beyond temperature scaling: Obtaining well-calibrated multi-class probabilities with dirichlet calibration. Advances in neural information processing systems, 32, 2019. 1, 2, 5 +[20] Meelis Kull, Telmo Silva Filho, and Peter Flach. Beta calibration: a well-founded and easily implemented improvement on logistic calibration for binary classifiers. In Artificial Intelligence and Statistics, pages 623-631. PMLR, 2017. 2 +[21] Balaji Lakshminarayanan, Alexander Pritzel, and Charles Blundell. Simple and scalable predictive uncertainty estimation using deep ensembles. Advances in neural information processing systems, 30, 2017. 1, 2, 5 +[22] John Lambert, Zhuang Liu, Ozan Sener, James Hays, and Vladlen Koltun. Mseg: A composite dataset for multi-domain semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2879-2888, 2020. 6 +[23] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 6 +[24] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3, 6, 7, 8 +[25] Xingchen Ma and Matthew B Blaschko. Meta-cal: Well-controlled post-hoc calibration by ranking. In International Conference on Machine Learning, pages 7235-7245. PMLR, 2021. 1, 2, 5 + +[26] Andrey Malinin, Bruno Mlodozeniec, and Mark Gales. Ensemble distribution distillation. arXiv preprint arXiv:1905.00076, 2019. 2 +[27] Alireza Mehrtash, William M Wells, Clare M Tempany, Purang Abolmaesumi, and Tina Kapur. Confidence calibration and predictive uncertainty estimation for deep medical image segmentation. IEEE transactions on medical imaging, 39(12):3868-3878, 2020. 2, 5 +[28] Matthias Minderer, Josip Djolonga, Rob Romijnders, Frances Hubis, Xiaohua Zhai, Neil Houlsby, Dustin Tran, and Mario Lucic. Revisiting the calibration of modern neural networks. Advances in Neural Information Processing Systems, 34:15682-15694, 2021. 7 +[29] Miguel Monteiro, Loic Le Folgoc, Daniel Coelho de Castro, Nick Pawlowski, Bernardo Marques, Konstantinos Kamnitsas, Mark van der Wilk, and Ben Glocker. Stochastic segmentation networks: Modelling spatially correlated aleatoric uncertainty. Advances in Neural Information Processing Systems, 33:12756-12767, 2020. 2 +[30] Jishnu Mukhoti, Viveka Kulharia, Amartya Sanyal, Stuart Golodetz, Philip Torr, and Puneet Dokania. Calibrating deep neural networks using focal loss. Advances in Neural Information Processing Systems, 33:15288-15299, 2020. 1, 2 +[31] Rafael Müller, Simon Kornblith, and Geoffrey E Hinton. When does label smoothing help? Advances in neural information processing systems, 32, 2019. 1, 2 +[32] Muhammad Akhtar Munir, Muhammad Haris Khan, M Saqib Sarfraz, and Mohsen Ali. Towards improving calibration in object detection under domain shift. arXiv preprint arXiv:2209.07601, 2022. 3 +[33] Mahdi Pakdaman Naeini, Gregory Cooper, and Milos Hauskrecht. Obtaining well calibrated probabilities using bayesian binning. In Twenty-Ninth AAAI Conference on Artificial Intelligence, 2015. 3, 4 +[34] Yaniv Ovadia, Emily Fertig, Jie Ren, Zachary Nado, David Sculley, Sebastian Nowozin, Joshua Dillon, Balaji Lakshminarayanan, and Jasper Snoek. Can you trust your model's uncertainty? evaluating predictive uncertainty under dataset shift. Advances in neural information processing systems, 32, 2019. 3 +[35] Sangdon Park, Osbert Bastani, James Weimer, and Insup Lee. Calibrated prediction with covariate shift via unsupervised domain adaptation. In International Conference on Artificial Intelligence and Statistics, pages 3219-3229. PMLR, 2020. 3 +[36] Kanil Patel, William Beluch, Bin Yang, Michael Pfeiffer, and Dan Zhang. Multi-class uncertainty calibration via mutual information maximization-based binning. arXiv preprint arXiv:2006.13092, 2020. 1, 2 +[37] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 724-732, 2016. 6 +[38] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 + +davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017. 1, 2 +[39] German Ros, Laura Sellart, Joanna Materzynska, David Vazquez, and Antonio M Lopez. The synthia dataset: A large collection of synthetic images for semantic segmentation of urban scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3234-3243, 2016. 2, 6 +[40] Robin Strudel, Ricardo Garcia, Ivan Laptev, and Cordelia Schmid. Segmenter: Transformer for semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7262-7272, 2021. 1, 2, 3, 4, 6, 7, 8 +[41] Sunil Thulasidasan, Gopinath Chennupati, Jeff A Bilmes, Tanmoy Bhattacharya, and Sarah Michalak. On mixup training: Improved calibration and predictive uncertainty for deep neural networks. Advances in Neural Information Processing Systems, 32, 2019. 1, 2 +[42] Christian Tomani, Sebastian Gruber, Muhammed Ebrar Erdem, Daniel Cremers, and Florian Buettner. Post-hoc uncertainty calibration for domain drift scenarios. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10124-10132, 2021. 3 +[43] Antonio Torralba and Alexei A Efros. Unbiased look at dataset bias. In CVPR 2011, pages 1521-1528. IEEE, 2011. 3 +[44] Adam Van Etten, Daniel Hogan, Jesus Martinez Manso, Jacob Shermeyer, Nicholas Weir, and Ryan Lewis. The multitemporal urban development spacenet dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6398-6407, 2021. 1, 2, 6 +[45] Ximei Wang, Mingsheng Long, Jianmin Wang, and Michael Jordan. Transferable calibration with lower bias and variance in domain adaptation. Advances in Neural Information Processing Systems, 33:19212-19223, 2020. 3 +[46] Yezhen Wang, Bo Li, Tong Che, Kaiyang Zhou, Ziwei Liu, and Dongsheng Li. Energy-based open-world uncertainty modeling for confidence calibration. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9302-9311, 2021. 1, 2 +[47] Andrew G Wilson and Pavel Izmailov. Bayesian deep learning and a probabilistic perspective of generalization. Advances in neural information processing systems, 33:4697-4708, 2020. 1, 2, 5 +[48] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In Proceedings of the European conference on computer vision (ECCV), pages 418-434, 2018. 6 +[49] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. arXiv preprint arXiv:2105.15203, 2021. 1, 2, 3, 4, 6, 7, 8 +[50] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darryll. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2636-2645, 2020. 1, 2, 6 + +[51] Wenwei Zhang, Jiangmiao Pang, Kai Chen, and Chen Change Loy. K-Net: Towards unified image segmentation. In NeurIPS, 2021. 1, 2, 3, 6, 7, 8 +[52] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2881-2890, 2017. 4 +[53] Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xiang, Philip HS Torr, et al. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. arXiv preprint arXiv:2012.15840, 2020. 1 +[54] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. 1, 2, 3, 6 \ No newline at end of file diff --git a/2023/On Calibrating Semantic Segmentation Models_ Analyses and an Algorithm/images.zip b/2023/On Calibrating Semantic Segmentation Models_ Analyses and an Algorithm/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..4b06a026d913ab7738986709c2aa7cbbfbc86c07 --- /dev/null +++ b/2023/On Calibrating Semantic Segmentation Models_ Analyses and an Algorithm/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1f8bc45582fb5e96e920023c15fda2fcb663eff5c414ef0e6a92bfe1c9f0daed +size 686031 diff --git a/2023/On Calibrating Semantic Segmentation Models_ Analyses and an Algorithm/layout.json b/2023/On Calibrating Semantic Segmentation Models_ Analyses and an Algorithm/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b4a0fed14b37087e1bf5be6b6aea234264a5320a --- /dev/null +++ b/2023/On Calibrating Semantic Segmentation Models_ Analyses and an Algorithm/layout.json @@ -0,0 +1,8818 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 61, + 103, + 533, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 103, + 533, + 121 + ], + "spans": [ + { + "bbox": [ + 61, + 103, + 533, + 121 + ], + "type": "text", + "content": "On Calibrating Semantic Segmentation Models: Analyses and An Algorithm" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 83, + 144, + 229, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 144, + 229, + 185 + ], + "spans": [ + { + "bbox": [ + 83, + 144, + 229, + 185 + ], + "type": "text", + "content": "Dongdong Wang \nUniversity of Central Florida \ndaniel.wang@knights.ucf.edu" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 255, + 144, + 343, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 255, + 144, + 343, + 185 + ], + "spans": [ + { + "bbox": [ + 255, + 144, + 343, + 185 + ], + "type": "text", + "content": "Boqing Gong Google Research bgong@google.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 370, + 144, + 512, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 144, + 512, + 185 + ], + "spans": [ + { + "bbox": [ + 370, + 144, + 512, + 185 + ], + "type": "text", + "content": "Liqiang Wang \nUniversity of Central Florida \nliqiang.wang@ucf.edu" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 239, + 290, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 239, + 290, + 478 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 290, + 478 + ], + "type": "text", + "content": "We study the problem of semantic segmentation calibration. Lots of solutions have been proposed to approach model miscalibration of confidence in image classification. However, to date, confidence calibration research on semantic segmentation is still limited. We provide a systematic study on the calibration of semantic segmentation models and propose a simple yet effective approach. First, we find that model capacity, crop size, multi-scale testing, and prediction correctness have impact on calibration. Among them, prediction correctness, especially misprediction, is more important to miscalibration due to over-confidence. Next, we propose a simple, unifying, and effective approach, namely selective scaling, by separating correct/incorrect prediction for scaling and more focusing on misprediction logit smoothing. Then, we study popular existing calibration methods and compare them with selective scaling on semantic segmentation calibration. We conduct extensive experiments with a variety of benchmarks on both indomain and domain-shift calibration and show that selective scaling consistently outperforms other methods." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 515, + 128, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 515, + 128, + 528 + ], + "spans": [ + { + "bbox": [ + 47, + 515, + 128, + 528 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 536, + 287, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 536, + 287, + 656 + ], + "spans": [ + { + "bbox": [ + 46, + 536, + 287, + 656 + ], + "type": "text", + "content": "Deep neural networks (DNNs) have become the \"goto\" models in various computer vision tasks, such as image classification, object detection, semantic segmentation. However, recent work found that DNNs are often overconfident when they make mistakes [11], misleading downstream applications. To calibrate DNNs' confidence in prediction, researchers have developed a rich line of works on image classification using regularized training [30, 31, 41, 46], post-hoc processing [11, 19, 25, 36], and Bayesian modeling [15, 18, 21, 47], to name a few." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 657, + 287, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 657, + 287, + 693 + ], + "spans": [ + { + "bbox": [ + 47, + 657, + 287, + 693 + ], + "type": "text", + "content": "However, the extensive pursuit of image classification has left it unclear how to calibrate DNNs for other computer vision tasks and how well existing calibration methods gen" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 316, + 214, + 414, + 300 + ], + "blocks": [ + { + "bbox": [ + 316, + 214, + 414, + 300 + ], + "lines": [ + { + "bbox": [ + 316, + 214, + 414, + 300 + ], + "spans": [ + { + "bbox": [ + 316, + 214, + 414, + 300 + ], + "type": "image", + "image_path": "45c8dd131d6ad700c2a4266b9c343191802bafb14f5770c9a48d93be122a1cf1.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 316, + 307, + 414, + 392 + ], + "blocks": [ + { + "bbox": [ + 316, + 307, + 414, + 392 + ], + "lines": [ + { + "bbox": [ + 316, + 307, + 414, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 307, + 414, + 392 + ], + "type": "image", + "image_path": "7b7545543cb0e9d51e55ae6d60580fcc9d0f4f7eaf393c15c76f8b1bb9c0b672.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 410, + 392, + 449, + 400 + ], + "lines": [ + { + "bbox": [ + 410, + 392, + 449, + 400 + ], + "spans": [ + { + "bbox": [ + 410, + 392, + 449, + 400 + ], + "type": "text", + "content": "Confidence" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 410, + 213, + 523, + 299 + ], + "blocks": [ + { + "bbox": [ + 410, + 213, + 523, + 299 + ], + "lines": [ + { + "bbox": [ + 410, + 213, + 523, + 299 + ], + "spans": [ + { + "bbox": [ + 410, + 213, + 523, + 299 + ], + "type": "image", + "image_path": "4908090dac040b7671834c84d750998d01811490a309a5c4f1423056ccc836bd.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 410, + 307, + 523, + 392 + ], + "blocks": [ + { + "bbox": [ + 410, + 307, + 523, + 392 + ], + "lines": [ + { + "bbox": [ + 410, + 307, + 523, + 392 + ], + "spans": [ + { + "bbox": [ + 410, + 307, + 523, + 392 + ], + "type": "image", + "image_path": "375aba9e23443008535c3e2200797845360559c534bb40053a39e30b99247d82.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 406, + 547, + 494 + ], + "lines": [ + { + "bbox": [ + 304, + 406, + 547, + 494 + ], + "spans": [ + { + "bbox": [ + 304, + 406, + 547, + 494 + ], + "type": "text", + "content": "Figure 1. Reliability diagrams [5] of visualized ECE for Segmenter [40]. We randomly select and evaluate fifty validation images from COCO-164K [1] and compare uncalibrated model (Uncal), selective scaling (Selective), temperature scaling (TempS), and ensembling (Ensemble). Smaller gap implies less ECE and better calibration. Ensembling yields comparable accuracy. Despite limited improvement by TempS and Ensemble, Selective exhibits better segmentation model calibration." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 508, + 547, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 508, + 547, + 544 + ], + "spans": [ + { + "bbox": [ + 304, + 508, + 547, + 544 + ], + "type": "text", + "content": "eralize to the tasks beyond image classification. In this paper, we conduct a comprehensive study of the calibration of deep semantic segmentation models." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 546, + 548, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 546, + 548, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 546, + 548, + 713 + ], + "type": "text", + "content": "Semantic segmentation tags a semantic label to every pixel in an image. Over the past years, we have witnessed increasingly accurate DNN models [2, 7, 40, 49, 51, 53] for semantic segmentation over various benchmark datasets [1, 14, 38, 44, 50, 54]. The progress has benefited many downstream applications, such as medical imaging and diagnostics, autonomous driving, and robotics. While accuracy is essential for the applications, the segmentation models' uncertainties also provide crucial signals, especially for safety-critical applications — by a segmentation model's uncertainty, we refer to its confidence in the label it assigns to a pixel. For example, an autonomous driving system can use the uncertainty of a semantic segmentation model (e.g., about drivable areas) to make informed decisions." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 495, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 495, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 495, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 702, + 257, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 702, + 257, + 712 + ], + "spans": [ + { + "bbox": [ + 61, + 702, + 257, + 712 + ], + "type": "text", + "content": "Code and models: https://github.com/dwang181/selectivecal" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23652" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 191 + ], + "type": "text", + "content": "Existing work on the semantic segmentation uncertainty focuses on the medical image domain except [6]. Deep ensemble [8, 21, 27], MC-dropout [16], and stochastic processes [29] are introduced from image classification to calibrate a medical image segmentation model's uncertainty. These methods are applied to the training stage and some could reduce a segmentation model's accuracy. We instead focus on the post-hoc calibration methods that do not change the model's prediction, and our study covers various domains." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 193, + 289, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 193, + 289, + 348 + ], + "spans": [ + { + "bbox": [ + 46, + 193, + 289, + 348 + ], + "type": "text", + "content": "Ding et al.'s work [6] is the most related to ours. While they proposed and validated a novel calibration method for semantic segmentation, our objective is two-fold. We first investigate the uncertainty in semantic segmentation from multiple perspectives, hoping to reveal the major challenges with the segmentation task's uncertainty. This comprehensive investigation and insights are necessary for future research, given the limited existing work in this area. We then propose a simple and effective post-hoc algorithm that outperforms and also can apply to almost all existing calibration methods initially designed for image classification, significantly improving their performance in semantic segmentation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 350, + 289, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 350, + 289, + 601 + ], + "spans": [ + { + "bbox": [ + 46, + 350, + 289, + 601 + ], + "type": "text", + "content": "To fill the gap of current research on semantic segmentation calibration, we conduct a systematic study. We analyze five state-of-the-art models [24, 40, 49, 51] with six existing calibration methods [6, 11, 19, 25, 27] over seven latest diverse benchmark datasets [1, 14, 38, 39, 44, 50, 54]. First, we investigate different factors to study how they affect miscalibration of semantic segmentation models through experiment. Given the observation, we find that model size, crop size, multi-scaling testing, and misprediction can affect calibration. In particular, segmentation model calibration error is more relevant to misprediction when model size is fixed. Next, we study different existing post-hoc calibrators and introduce our proposed selective scaling algorithm based upon the observation of misprediction. Our selective scaling optimizes ECE by reducing confidence of misprediction, which is reflected by more samples falling into lower confidence region (the far left bin) in Figure 1. Then, we carry out extensive experiments across different models, calibrators, and benchmark datasets. The experiments examine both in-domain calibration and domain-shift calibration to reveal calibration variability across different tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 603, + 258, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 603, + 258, + 614 + ], + "spans": [ + { + "bbox": [ + 59, + 603, + 258, + 614 + ], + "type": "text", + "content": "Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 616, + 287, + 714 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 46, + 616, + 287, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 616, + 287, + 688 + ], + "spans": [ + { + "bbox": [ + 46, + 616, + 287, + 688 + ], + "type": "text", + "content": "- We conduct a systematic study on the calibration of semantic segmentation models and provide insights to segmentation model calibration. We find that larger models tend to be less calibrated. When model architecture is fixed, larger crop size and multi-scale testing help calibrate models. Moreover, misprediction is important to miscalibration." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 287, + 714 + ], + "type": "text", + "content": "- We compare different existing popular calibration methods on state-of-the-art semantic segmentation models," + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 108 + ], + "type": "text", + "content": "and propose a simple but effective approach, selective scaling, which carries out separate scaling on correct and incorrect predictions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 108, + 545, + 191 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 304, + 108, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 108, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 304, + 108, + 545, + 156 + ], + "type": "text", + "content": "- We conduct extensive experiments and justify the effectiveness of selective scaling, as shown in Figure 1. We also extend experiments from in-domain to domain-shift data and show that selective scaling consistently outperforms." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 156, + 545, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 156, + 545, + 191 + ], + "spans": [ + { + "bbox": [ + 304, + 156, + 545, + 191 + ], + "type": "text", + "content": "- We provide useful calibration observations as a comprehensive reference for further research on segmentation model calibration." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 203, + 392, + 216 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 203, + 392, + 216 + ], + "spans": [ + { + "bbox": [ + 306, + 203, + 392, + 216 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 223, + 545, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 223, + 545, + 520 + ], + "spans": [ + { + "bbox": [ + 304, + 223, + 545, + 520 + ], + "type": "text", + "content": "DNN calibration. The methods of calibrating DNNs can be categorized into three groups, regularized training, uncertainty estimation, and post-hoc calibration [9]. Regularized training focuses on calibrating DNN over training, such as [30,31,41,46]. Some methods estimate uncertainty using Bayesian methods [15, 18, 47] and deep ensemble [21, 26]. Post-hoc calibration conducts post-processing on the output of DNN to obtain well-calibrated results. While some post-hoc approaches exchange hurting accuracy for better calibration, like histogram binning [11, 36], some retain original accuracy of DNNs and only add calibration map to the last layer of DNNs to adjust probability distribution pattern. Several existing binary calibration maps are extended to multiclass calibration. For example, [11] extended temperature scaling and vector scaling aka Logistic scaling [11], to multiclassss calibration for DNN, and revealed temperature scaling is simple but more effective. To enrich calibration map, Dirichlet scaling [19] was proposed by extending Beta scaling [20] to multiclass DNN calibration. Recently, Meta-cal [25] integrated bipartite-ranking model with selective classification to improve calibration map for better calibrated model. Despite the success of these existing approaches in image classification calibration, there is no work to discuss their performance on semantic segmentation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 545, + 713 + ], + "type": "text", + "content": "Calibration beyond image classification. Several extensions based upon existing calibrators have been proposed to solve calibration beyond image classification. For example, [29] modeled predictive uncertainty with spatial correlation matrix in stochastic process and incorporated this information into model training for more calibrated segmentation model. [6] proposed local temperature scaling that extends temperature scaling to pixels, and thus, better captures spatial heterogeneity of temperature for a better calibrated segmentation model. [27] conducted model assembling to improve calibration and evaluate the improvement by both conventional and the proposed segment-level uncertainty metrics. [17] addressed semantic segmentation calibration through adversarial training a stochastic network which enables more accurate uncertainty estimation and robust prediction. [8] proposed a multi-headed Variational U-Net by" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23653" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 142 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 142 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 142 + ], + "type": "text", + "content": "combining ensemble modeling with variational inference and improves model training with better in-distribution calibration and out-of-distribution detection. [32] addressed visual object detector calibration with plug-and-play training-time calibration loss and improved uncertainty optimization." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 144, + 288, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 144, + 288, + 346 + ], + "spans": [ + { + "bbox": [ + 46, + 144, + 288, + 346 + ], + "type": "text", + "content": "Domain-shift calibration. Domain generalization is an important challenge to DNN since model generalization to unseen data could be unpredictable when training and test data distributions differ [43]. In terms of model calibration, performance degradation can also be critical [13, 34]. To alleviate this problem, a variety of approaches have been proposed. [35] corrected covariant shift by importance weighting to achieve better calibrated models. [42] proposed a simple yet effective scheme by perturbing image for classification model calibration improvement. [45] proposed TransCal to improve image classification calibration degradation under the scenario of natural distribution shift. [10] leveraged multiple calibration domain to reduce disparity between source and target domain, thus improving model calibration on shift target domain. Despite some progress in domain-shift calibration, there is still no domain-shift study on semantic segmentation calibration." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 348, + 287, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 348, + 287, + 539 + ], + "spans": [ + { + "bbox": [ + 46, + 348, + 287, + 539 + ], + "type": "text", + "content": "Semantic segmentation. Semantic segmentation is an important visual understanding task. Different from image classification, semantic segmentation focuses on pixel-level classification within an image. Due to this difference, semantic segmentation models are developed in consideration of spatial correlation between pixels. DeepLab family, as an important dilated convolution model, integrates multiscale spatial context into segmentation modeling [2]. SegFormer incorporates hierarchical architectures to encode multiscale features and performs efficient semantic segmentation through ViT structure [49]. Segmenter is a straightforward ViT extension to semantic segmentation by decoding global context of the encoded embedding features from ViT outputs [40]. K-net is a unifying framework for all segmentation tasks [51]. ConvNeXt is proposed as an alternative to ViT and shows accurate segmentation [24]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 550, + 132, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 550, + 132, + 562 + ], + "spans": [ + { + "bbox": [ + 47, + 550, + 132, + 562 + ], + "type": "text", + "content": "3. Preliminaries" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 570, + 287, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 653 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 653 + ], + "type": "text", + "content": "Similar to image classification, semantic segmentation can be formulated to a multiclass classification problem with a deep neural network. Let " + }, + { + "bbox": [ + 46, + 570, + 287, + 653 + ], + "type": "inline_equation", + "content": "x \\in X" + }, + { + "bbox": [ + 46, + 570, + 287, + 653 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 570, + 287, + 653 + ], + "type": "inline_equation", + "content": "y \\in Y" + }, + { + "bbox": [ + 46, + 570, + 287, + 653 + ], + "type": "text", + "content": " denote an input and its label, respectively. A deep neural network " + }, + { + "bbox": [ + 46, + 570, + 287, + 653 + ], + "type": "inline_equation", + "content": "h(x) = (\\hat{y}, \\hat{p})" + }, + { + "bbox": [ + 46, + 570, + 287, + 653 + ], + "type": "text", + "content": " yields " + }, + { + "bbox": [ + 46, + 570, + 287, + 653 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 46, + 570, + 287, + 653 + ], + "type": "text", + "content": " as the predicted label with inference confidence " + }, + { + "bbox": [ + 46, + 570, + 287, + 653 + ], + "type": "inline_equation", + "content": "\\hat{p}" + }, + { + "bbox": [ + 46, + 570, + 287, + 653 + ], + "type": "text", + "content": ". We expect a well-calibrated model to provide accurate prediction when its confidence is high." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 654, + 287, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 654, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 654, + 287, + 712 + ], + "type": "text", + "content": "ECE. There are several metrics to measure a model's calibration, and one of the most popular and accepted metrics is expected calibration error (ECE) [33], which reflects the gap between predictive confidence and accuracy. The formal definition with continuous variable is as follows." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 351, + 72, + 545, + 85 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 72, + 545, + 85 + ], + "spans": [ + { + "bbox": [ + 351, + 72, + 545, + 85 + ], + "type": "interline_equation", + "content": "E C E = \\mathbb {E} _ {\\hat {p}} \\left[ | \\mathbb {P} (\\hat {y} = y | \\hat {p} = p) - p | \\right], \\tag {1}", + "image_path": "fa8b79ff6601a971a647f14ea80be2046ab8b96873f0273fdfc1d34f0079288a.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 91, + 545, + 150 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 545, + 150 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 91, + 545, + 150 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 304, + 91, + 545, + 150 + ], + "type": "text", + "content": " is the predicted label, " + }, + { + "bbox": [ + 304, + 91, + 545, + 150 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 304, + 91, + 545, + 150 + ], + "type": "text", + "content": " is the true label, " + }, + { + "bbox": [ + 304, + 91, + 545, + 150 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 91, + 545, + 150 + ], + "type": "text", + "content": " is expected confidence, and " + }, + { + "bbox": [ + 304, + 91, + 545, + 150 + ], + "type": "inline_equation", + "content": "\\mathbb{P}(\\hat{y} = y|\\hat{p} = p)" + }, + { + "bbox": [ + 304, + 91, + 545, + 150 + ], + "type": "text", + "content": " is model predictive accuracy. The expectation is about the discrepancy between accuracy and confidence. A perfectly calibrated model has zero ECE." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 152, + 545, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 152, + 545, + 187 + ], + "spans": [ + { + "bbox": [ + 304, + 152, + 545, + 187 + ], + "type": "text", + "content": "In practical problems, statistical binning is used to quantize continuous variables and estimate eq. (1) by equally binning the probability interval," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 342, + 194, + 545, + 224 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 194, + 545, + 224 + ], + "spans": [ + { + "bbox": [ + 342, + 194, + 545, + 224 + ], + "type": "interline_equation", + "content": "\\widehat {E C E} = \\sum_ {i = 1} ^ {m} \\frac {\\left| B _ {i} \\right|}{n} | \\operatorname {a c c} \\left(B _ {i}\\right) - \\operatorname {c o n f} \\left(B _ {i}\\right) |, \\tag {2}", + "image_path": "a1dab0ed6c6dd3399ac535b69fc5007caf33081bb1ecb4a375d91e5097d0ee87.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 231, + 545, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 231, + 545, + 292 + ], + "spans": [ + { + "bbox": [ + 304, + 231, + 545, + 292 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 231, + 545, + 292 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 231, + 545, + 292 + ], + "type": "text", + "content": " is the number of samples, " + }, + { + "bbox": [ + 304, + 231, + 545, + 292 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 231, + 545, + 292 + ], + "type": "text", + "content": " is the number of bins, " + }, + { + "bbox": [ + 304, + 231, + 545, + 292 + ], + "type": "inline_equation", + "content": "B_{i}" + }, + { + "bbox": [ + 304, + 231, + 545, + 292 + ], + "type": "text", + "content": " denotes a set of samples falling into the bin, and " + }, + { + "bbox": [ + 304, + 231, + 545, + 292 + ], + "type": "inline_equation", + "content": "\\mathrm{acc}(B_i)" + }, + { + "bbox": [ + 304, + 231, + 545, + 292 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 231, + 545, + 292 + ], + "type": "inline_equation", + "content": "\\mathrm{conf}(B_i)" + }, + { + "bbox": [ + 304, + 231, + 545, + 292 + ], + "type": "text", + "content": " are accuracy and confidence averaged over the samples in the bin. " + }, + { + "bbox": [ + 304, + 231, + 545, + 292 + ], + "type": "inline_equation", + "content": "\\widehat{ECE}" + }, + { + "bbox": [ + 304, + 231, + 545, + 292 + ], + "type": "text", + "content": " can be visualized with the gaps in reliability diagram [5] as Figure 1." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 293, + 545, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 293, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 304, + 293, + 545, + 352 + ], + "type": "text", + "content": "ECE in semantic segmentation. We extend ECE to semantic segmentation by considering each pixel as a sample. Instead of pooling all pixels of different images into a set, we calculate ECE over each image first before taking an average across images," + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 376, + 358, + 545, + 391 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 358, + 545, + 391 + ], + "spans": [ + { + "bbox": [ + 376, + 358, + 545, + 391 + ], + "type": "interline_equation", + "content": "E C E = \\frac {1}{N} \\sum_ {I = 1} ^ {N} E C E _ {I}, \\tag {3}", + "image_path": "a559924e752a252496cd9f1b27413f84c65ae2d82f50ee63bb095c06023786c6.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 393, + 545, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 393, + 545, + 441 + ], + "spans": [ + { + "bbox": [ + 304, + 393, + 545, + 441 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 393, + 545, + 441 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 304, + 393, + 545, + 441 + ], + "type": "text", + "content": " is an inference image, and " + }, + { + "bbox": [ + 304, + 393, + 545, + 441 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 393, + 545, + 441 + ], + "type": "text", + "content": " is the total number of images. Note that we adopt " + }, + { + "bbox": [ + 304, + 393, + 545, + 441 + ], + "type": "inline_equation", + "content": "ECE_{I}" + }, + { + "bbox": [ + 304, + 393, + 545, + 441 + ], + "type": "text", + "content": " because an image-wise metric is more popular in segmentation model evaluation. Also, " + }, + { + "bbox": [ + 304, + 393, + 545, + 441 + ], + "type": "inline_equation", + "content": "ECE_{I}" + }, + { + "bbox": [ + 304, + 393, + 545, + 441 + ], + "type": "text", + "content": " may significantly vary across images." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 452, + 508, + 465 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 452, + 508, + 465 + ], + "spans": [ + { + "bbox": [ + 305, + 452, + 508, + 465 + ], + "type": "text", + "content": "4. Uncertainty of Segmentation Models" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 472, + 545, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 545, + 579 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 545, + 579 + ], + "type": "text", + "content": "We analyze the uncertainty of five segmentation models, including Segmenter [40], SegFormer [49], Knet-DeepLab [51], Knet-SWIN [51], and ConvNeXt [24], with the ADE20K [54] dataset, a popular semantic segmentation benchmark with 20,210/2,000 training/validation images. The models are state-of-the-art and released by OpenMMLab [3]. Due to limit of space, we focus on Segmenter [40] and SegFormer [49] in this section and report other models' results in the supplementary materials." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 587, + 529, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 587, + 529, + 600 + ], + "spans": [ + { + "bbox": [ + 305, + 587, + 529, + 600 + ], + "type": "text", + "content": "4.1. Larger models tending to be less calibrated" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 545, + 713 + ], + "type": "text", + "content": "Among the models we analyze, we find that the model capacity is a primary factor determining their generalization accuracy. Large models often lead to high inference accuracy. Will it be valid for semantic segmentation models? To answer it, we conduct experiments on different model capacity by varying model depth and width. We select five variants of SegFormer [49], from B0 to B5 with a deeper and wider architecture, for model scaling-up study. From Figure 2a, we observe that the deeper and wider models," + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23654" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 68, + 167, + 178 + ], + "blocks": [ + { + "bbox": [ + 47, + 68, + 167, + 178 + ], + "lines": [ + { + "bbox": [ + 47, + 68, + 167, + 178 + ], + "spans": [ + { + "bbox": [ + 47, + 68, + 167, + 178 + ], + "type": "image", + "image_path": "3e88ee4deacdd129bca04ad1635e7e4cbaeabdbb1d7222a19344e69b8a92d8cb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 106, + 181, + 116, + 190 + ], + "lines": [ + { + "bbox": [ + 106, + 181, + 116, + 190 + ], + "spans": [ + { + "bbox": [ + 106, + 181, + 116, + 190 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 167, + 70, + 293, + 178 + ], + "blocks": [ + { + "bbox": [ + 167, + 70, + 293, + 178 + ], + "lines": [ + { + "bbox": [ + 167, + 70, + 293, + 178 + ], + "spans": [ + { + "bbox": [ + 167, + 70, + 293, + 178 + ], + "type": "image", + "image_path": "2b2e60689a18d4b8cd8e136e78a94cfc3687455aeb8f753f0f0b274517864621.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 233, + 181, + 244, + 190 + ], + "lines": [ + { + "bbox": [ + 233, + 181, + 244, + 190 + ], + "spans": [ + { + "bbox": [ + 233, + 181, + 244, + 190 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 294, + 70, + 427, + 175 + ], + "blocks": [ + { + "bbox": [ + 294, + 70, + 427, + 175 + ], + "lines": [ + { + "bbox": [ + 294, + 70, + 427, + 175 + ], + "spans": [ + { + "bbox": [ + 294, + 70, + 427, + 175 + ], + "type": "image", + "image_path": "15cb2633b3fc3d4e217446b7d3159a186b32ddfc20714ba90f3969f1e95d9d9b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 360, + 181, + 370, + 190 + ], + "lines": [ + { + "bbox": [ + 360, + 181, + 370, + 190 + ], + "spans": [ + { + "bbox": [ + 360, + 181, + 370, + 190 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 193, + 545, + 215 + ], + "lines": [ + { + "bbox": [ + 46, + 193, + 545, + 215 + ], + "spans": [ + { + "bbox": [ + 46, + 193, + 545, + 215 + ], + "type": "text", + "content": "Figure 2. The effect of model capacity, image crop size, multi-scale testing, and prediction correctness on miscalibration. Prediction correction separately computes the ECE of incorrect/correct predictions. Image-based ECE in eq.(3) is used." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 432, + 70, + 529, + 171 + ], + "blocks": [ + { + "bbox": [ + 432, + 70, + 529, + 171 + ], + "lines": [ + { + "bbox": [ + 432, + 70, + 529, + 171 + ], + "spans": [ + { + "bbox": [ + 432, + 70, + 529, + 171 + ], + "type": "image", + "image_path": "818e12223a4d714e78606c2ac26fb979fc9b1d86c898fa15272a65e334167a0d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 483, + 181, + 493, + 190 + ], + "lines": [ + { + "bbox": [ + 483, + 181, + 493, + 190 + ], + "spans": [ + { + "bbox": [ + 483, + 181, + 493, + 190 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 226, + 287, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 226, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 46, + 226, + 287, + 308 + ], + "type": "text", + "content": "like B5, exhibit higher ECE (i.e., worse miscalibration) despite better mIoU. It is consistent with the observation for image classification from Figure 2 in [11], which is caused by the over-confidence during the later stage of NLL optimization [11]. This implies that more accurate prediction leads to worse calibration when it comes to varying model size." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 319, + 287, + 332 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 319, + 287, + 332 + ], + "spans": [ + { + "bbox": [ + 47, + 319, + 287, + 332 + ], + "type": "text", + "content": "4.2. Larger crop size tending to be more calibrated" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 338, + 289, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 338, + 289, + 542 + ], + "spans": [ + { + "bbox": [ + 46, + 338, + 289, + 542 + ], + "type": "text", + "content": "Image cropping is usually used to prepare input data for semantic segmentation models. Crop size determines model input resolution, i.e., the number of pixels for modeling. Larger crop size indicates larger input image, more pixels fed to model, and wider modeling region. Although larger crop size shows higher predictive accuracy, its effect on model calibration is unknown. We conduct experiments to explore their interaction. We select Segmenter-L [40] as a representative model with five crop sizes of 416, 448, 480, 512, and 640 to show the crop size scaling effect. From Figure 2b, we observe that larger crop size yields lower ECE and better calibration. We associate it with modeling field. When crop size increases, larger modeling field provides richer global context information which helps more accurately and confidently locate object boundary and position. This implies that more accurate prediction leads to better calibration in terms of varying crop size." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 47, + 551, + 271, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 551, + 271, + 563 + ], + "spans": [ + { + "bbox": [ + 47, + 551, + 271, + 563 + ], + "type": "text", + "content": "4.3. Multi-scaling testing improving calibration" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "type": "text", + "content": "Multi-scale testing is an important technique to boost semantic segmentation model. It usually resizes the input images into multiple scales, feed these resized images into the model, and average all inferences for final prediction. Predictive accuracy is improved after multi-scale testing because multi-scale features help address scale variation between different objects [52]. However, its effectiveness on semantic segmentation calibration is unknown. We conduct experiments with Segmenter-L and cropping size of " + }, + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "type": "inline_equation", + "content": "640 \\times 640" + }, + { + "bbox": [ + 46, + 570, + 287, + 714 + ], + "type": "text", + "content": " of ADE20K to investigate the effect of multiscale testing. The rescale ratios include 0.5, 0.75, 1.0, 1.25, 1.5, and 1.75. Given Figure 2c, with multi-scale testing," + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 225, + 545, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 225, + 545, + 334 + ], + "spans": [ + { + "bbox": [ + 304, + 225, + 545, + 334 + ], + "type": "text", + "content": "ECE is reduced from 0.86 to 0.78 while mIoU increases from 0.515 to 0.535. This suggests that multi-scale testing enables both alleviating miscalibration and improving accuracy. The intuition is that the aggregation of multi-scale features improves the object boundary delineation and predictive confidence in objects. Also, this ensembling helps correct mispredictions and reduce miscalibration over the over-confidence. This implies that more accurate prediction leads to better calibration concerning multi-scaling testing." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 346, + 495, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 346, + 495, + 358 + ], + "spans": [ + { + "bbox": [ + 305, + 346, + 495, + 358 + ], + "type": "text", + "content": "4.4. Mispredictions affecting calibration" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 365, + 545, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 365, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 365, + 545, + 594 + ], + "type": "text", + "content": "Given eqs.(1) and (2), miscalibration results from discrepancy between predictive confidence and accuracy [33]. A perfectly calibrated model should yield predicted probability of 1, i.e., " + }, + { + "bbox": [ + 304, + 365, + 545, + 594 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 304, + 365, + 545, + 594 + ], + "type": "text", + "content": " confidence, for correct prediction, and predicted probability of 0, i.e., infinitesimal confidence, for misprediction. However, an optimal solution is hardly obtained in practice, which implies underconfidence in correct prediction and over-confidence in misprediction. How do these discrepancies affect the calibration property of a segmentation model? We conduct an experiment with SegFormer-B5 [49] and Segmenter-L [40] by grouping correct/incorrect predictions to separately compute their ECEs. This separation can more clearly reveal how much they contribute to final miscalibration. Figure 2d shows that the ECEs from incorrect predictions are significantly higher than those from correct predictions across both models. This implies that misprediction more contributes to miscalibration. This observation also justifies that over-confidence is a more critical problem for calibration [11]." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 608, + 538, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 608, + 538, + 621 + ], + "spans": [ + { + "bbox": [ + 306, + 608, + 538, + 621 + ], + "type": "text", + "content": "5. Calibrating Semantic Segmentation Model" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": "We first introduce our post-hoc scaling algorithm in this section. Next, we briefly review the popular existing calibration methods, including four image classification post-hoc calibrators, one semantic segmentation calibrator, and one Bayesian modeling calibrator. We select these calibrators because they exhibit state-of-the-art calibration performance in different groups." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23655" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 70, + 121, + 117 + ], + "blocks": [ + { + "bbox": [ + 52, + 70, + 121, + 117 + ], + "lines": [ + { + "bbox": [ + 52, + 70, + 121, + 117 + ], + "spans": [ + { + "bbox": [ + 52, + 70, + 121, + 117 + ], + "type": "image", + "image_path": "fca2166d8e9df29985cc7cbd45e08601fd6379070d951534e431e1f3456f0cb1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 119, + 545, + 131 + ], + "lines": [ + { + "bbox": [ + 47, + 119, + 545, + 131 + ], + "spans": [ + { + "bbox": [ + 47, + 119, + 545, + 131 + ], + "type": "text", + "content": "Figure 3. Segmented examples from ADE20K, COCO-164K, BDD100K, DAVIS, SpaceNet-7, BraTs-2017, and SYNTHIA (left to right)." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 126, + 70, + 187, + 117 + ], + "blocks": [ + { + "bbox": [ + 126, + 70, + 187, + 117 + ], + "lines": [ + { + "bbox": [ + 126, + 70, + 187, + 117 + ], + "spans": [ + { + "bbox": [ + 126, + 70, + 187, + 117 + ], + "type": "image", + "image_path": "ae4e04764fa12454d71acaf94b3cfa24c7df7adf90fd73449967c3fc95104527.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 194, + 70, + 274, + 117 + ], + "blocks": [ + { + "bbox": [ + 194, + 70, + 274, + 117 + ], + "lines": [ + { + "bbox": [ + 194, + 70, + 274, + 117 + ], + "spans": [ + { + "bbox": [ + 194, + 70, + 274, + 117 + ], + "type": "image", + "image_path": "63564311a8e8345e097787986fbb2191df5bc072a2609c459daa45ed956ec02b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 280, + 70, + 361, + 117 + ], + "blocks": [ + { + "bbox": [ + 280, + 70, + 361, + 117 + ], + "lines": [ + { + "bbox": [ + 280, + 70, + 361, + 117 + ], + "spans": [ + { + "bbox": [ + 280, + 70, + 361, + 117 + ], + "type": "image", + "image_path": "37a9fd1a71740da8aa98d661eebe0cee935db0b482016147044e77ddcbaa1dd4.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 368, + 70, + 414, + 117 + ], + "blocks": [ + { + "bbox": [ + 368, + 70, + 414, + 117 + ], + "lines": [ + { + "bbox": [ + 368, + 70, + 414, + 117 + ], + "spans": [ + { + "bbox": [ + 368, + 70, + 414, + 117 + ], + "type": "image", + "image_path": "04a94016700a1cc17f11b5bd313a7b4b29c5f947a4e8f0f52ccf9095f864f7a4.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 420, + 70, + 457, + 117 + ], + "blocks": [ + { + "bbox": [ + 420, + 70, + 457, + 117 + ], + "lines": [ + { + "bbox": [ + 420, + 70, + 457, + 117 + ], + "spans": [ + { + "bbox": [ + 420, + 70, + 457, + 117 + ], + "type": "image", + "image_path": "ae00f1e2e89a8ce509fcd2713b816352814a6775d07b587bd4e86898757165bb.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 463, + 71, + 539, + 117 + ], + "blocks": [ + { + "bbox": [ + 463, + 71, + 539, + 117 + ], + "lines": [ + { + "bbox": [ + 463, + 71, + 539, + 117 + ], + "spans": [ + { + "bbox": [ + 463, + 71, + 539, + 117 + ], + "type": "image", + "image_path": "908d5406d7c67a4a9858505546fd34934771c5740a35d8ff1518526b0a6c9612.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 147, + 147, + 159 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 147, + 147, + 159 + ], + "spans": [ + { + "bbox": [ + 47, + 147, + 147, + 159 + ], + "type": "text", + "content": "5.1. Selective Scaling" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 165, + 287, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 165, + 287, + 260 + ], + "spans": [ + { + "bbox": [ + 46, + 165, + 287, + 260 + ], + "type": "text", + "content": "We propose selective scaling given that misprediction is more attributed to miscalibration (Figure 2d). Selective scaling is an extension of selective classification to model calibration. It introduces a binary classifier as a selector to categorize correct and incorrect predictions for separate scaling. For example, selected mispredictions' logits will be smoothed to alleviate the miscalibration from overconfidence." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 102, + 267, + 287, + 298 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 267, + 287, + 298 + ], + "spans": [ + { + "bbox": [ + 102, + 267, + 287, + 298 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l l} \\hat {\\boldsymbol {p}} = \\sigma_ {S M} (\\boldsymbol {z} / T _ {1}), & \\text {i f} \\hat {y} \\neq y \\\\ \\hat {\\boldsymbol {p}} = \\sigma_ {S M} (\\boldsymbol {z} / T _ {2}), & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {4}", + "image_path": "e15a15e2a0453ee892b74cd95a01df83272155f39f5882e4bfecf4e5f1735379.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 301, + 287, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 301, + 287, + 432 + ], + "spans": [ + { + "bbox": [ + 46, + 301, + 287, + 432 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 301, + 287, + 432 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{p}}" + }, + { + "bbox": [ + 46, + 301, + 287, + 432 + ], + "type": "text", + "content": " is calibrated probability vector, " + }, + { + "bbox": [ + 46, + 301, + 287, + 432 + ], + "type": "inline_equation", + "content": "\\sigma_{SM}" + }, + { + "bbox": [ + 46, + 301, + 287, + 432 + ], + "type": "text", + "content": " is softmax activation function, " + }, + { + "bbox": [ + 46, + 301, + 287, + 432 + ], + "type": "inline_equation", + "content": "\\pmb{z}" + }, + { + "bbox": [ + 46, + 301, + 287, + 432 + ], + "type": "text", + "content": " is the logit vector before activation, " + }, + { + "bbox": [ + 46, + 301, + 287, + 432 + ], + "type": "inline_equation", + "content": "T_{1}" + }, + { + "bbox": [ + 46, + 301, + 287, + 432 + ], + "type": "text", + "content": " is a temperature to smooth logit distribution, and " + }, + { + "bbox": [ + 46, + 301, + 287, + 432 + ], + "type": "inline_equation", + "content": "T_{2}" + }, + { + "bbox": [ + 46, + 301, + 287, + 432 + ], + "type": "text", + "content": " is a temperature to sharpen logit distribution, so " + }, + { + "bbox": [ + 46, + 301, + 287, + 432 + ], + "type": "inline_equation", + "content": "T_{1} > T_{2}" + }, + { + "bbox": [ + 46, + 301, + 287, + 432 + ], + "type": "text", + "content": ". For over-confidence problem, we set " + }, + { + "bbox": [ + 46, + 301, + 287, + 432 + ], + "type": "inline_equation", + "content": "T_{2}" + }, + { + "bbox": [ + 46, + 301, + 287, + 432 + ], + "type": "text", + "content": " to 1. It can be extended to under-confidence problem with " + }, + { + "bbox": [ + 46, + 301, + 287, + 432 + ], + "type": "inline_equation", + "content": "T_{2}" + }, + { + "bbox": [ + 46, + 301, + 287, + 432 + ], + "type": "text", + "content": " less than 1. The selection of " + }, + { + "bbox": [ + 46, + 301, + 287, + 432 + ], + "type": "inline_equation", + "content": "T_{1}" + }, + { + "bbox": [ + 46, + 301, + 287, + 432 + ], + "type": "text", + "content": " relies on the selector accuracy. When its accuracy is higher, " + }, + { + "bbox": [ + 46, + 301, + 287, + 432 + ], + "type": "inline_equation", + "content": "T_{1}" + }, + { + "bbox": [ + 46, + 301, + 287, + 432 + ], + "type": "text", + "content": " is larger. Our selective scaling optimizes ECE by reducing confidence of misprediction, which is reflected by more samples falling into the lower confidence region (the far left bin) in reliability diagram (Figure 1)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 433, + 287, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 433, + 287, + 552 + ], + "spans": [ + { + "bbox": [ + 46, + 433, + 287, + 552 + ], + "type": "text", + "content": "Selective scheme was adopted for calibrator design in Meta-cal [25], but different from selective scaling. First, selective scaling does not degrade model accuracy while Meta-cal with miscoverage control does [25]. Different from Meta-cal, selective scaling is independent from entropy ranking model to determine classification threshold. Also, selective scaling obtains a selector without heuristic data resampling used in Meta-cal, which reduces hyperparameters. Selective scaling, to the best of our knowledge, is never used for semantic segmentation model calibration." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 47, + 559, + 204, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 559, + 204, + 572 + ], + "spans": [ + { + "bbox": [ + 47, + 559, + 204, + 572 + ], + "type": "text", + "content": "5.2. Existing calibration methods" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 578, + 287, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 578, + 287, + 626 + ], + "spans": [ + { + "bbox": [ + 46, + 578, + 287, + 626 + ], + "type": "text", + "content": "Temperature Scaling is a simple but effective approach for multi-classification model calibration [11]. The calibration is carried out with a single temperature parameter to scale logits for overfitting problem resolution." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 131, + 634, + 287, + 647 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 634, + 287, + 647 + ], + "spans": [ + { + "bbox": [ + 131, + 634, + 287, + 647 + ], + "type": "interline_equation", + "content": "\\hat {\\boldsymbol {p}} = \\sigma_ {S M} (\\boldsymbol {z} / T), \\tag {5}", + "image_path": "d418d72df223226926be67a0c46c1450c51544fdd985db02ca200f7da62fd254.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 47, + 654, + 279, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 654, + 279, + 665 + ], + "spans": [ + { + "bbox": [ + 47, + 654, + 279, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 654, + 279, + 665 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 47, + 654, + 279, + 665 + ], + "type": "text", + "content": " is a scalar of temperature to scale logit vector " + }, + { + "bbox": [ + 47, + 654, + 279, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 46, + 666, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 666, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 666, + 287, + 713 + ], + "type": "text", + "content": "Logistic Scaling is an extension of temperature scaling a.k.a. vector scaling [11]. The scaling model is formulated with linear transformation for more complex calibration map." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 378, + 147, + 545, + 159 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 147, + 545, + 159 + ], + "spans": [ + { + "bbox": [ + 378, + 147, + 545, + 159 + ], + "type": "interline_equation", + "content": "\\hat {\\boldsymbol {p}} = \\sigma_ {S M} (w \\odot \\boldsymbol {z} + b), \\tag {6}", + "image_path": "a403ed5ec51453c05e8fa216c1a86ee2f50ef2aea14963ad15c5dc246ffc3f8d.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 306, + 161, + 535, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 161, + 535, + 173 + ], + "spans": [ + { + "bbox": [ + 306, + 161, + 535, + 173 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 306, + 161, + 535, + 173 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 306, + 161, + 535, + 173 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 306, + 161, + 535, + 173 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 306, + 161, + 535, + 173 + ], + "type": "text", + "content": " are two vectors to scale the logit vector " + }, + { + "bbox": [ + 306, + 161, + 535, + 173 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 306, + 161, + 535, + 173 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 174, + 545, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 174, + 545, + 232 + ], + "spans": [ + { + "bbox": [ + 305, + 174, + 545, + 232 + ], + "type": "text", + "content": "Dirichlet Scaling is the extension of logistic scaling and derived with probability output distribution [19], which enriches calibration map for better optimization with Dirichlet distribution. We adopt linear parameterization [19] for Dirichlet scaling, which is formulated as follows." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 354, + 242, + 545, + 255 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 242, + 545, + 255 + ], + "spans": [ + { + "bbox": [ + 354, + 242, + 545, + 255 + ], + "type": "interline_equation", + "content": "\\hat {\\boldsymbol {p}} = \\sigma_ {S M} (W \\cdot \\log (\\sigma_ {S M} (\\boldsymbol {z})) + b), \\tag {7}", + "image_path": "caa6f52b8072fbb3344d58bcd682cdad8b8a812d4b5a1cb4a53ca9f8fae140b5.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 305, + 256, + 545, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 256, + 545, + 280 + ], + "spans": [ + { + "bbox": [ + 305, + 256, + 545, + 280 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 256, + 545, + 280 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 305, + 256, + 545, + 280 + ], + "type": "text", + "content": " is a matrix and " + }, + { + "bbox": [ + 305, + 256, + 545, + 280 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 305, + 256, + 545, + 280 + ], + "type": "text", + "content": " is a vector for linear parametrisation of the probability " + }, + { + "bbox": [ + 305, + 256, + 545, + 280 + ], + "type": "inline_equation", + "content": "\\sigma_{SM}(z)" + }, + { + "bbox": [ + 305, + 256, + 545, + 280 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 305, + 281, + 545, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 281, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 305, + 281, + 545, + 399 + ], + "type": "text", + "content": "Meta-cal is the scaling strategy derived from bipartite-ranking model and selective classification [25]. With the entropy threshold by a ranking model, probability outputs are separately processed. The probability output with the entropy smaller than threshold is scaled by temperature scaling; otherwise, the output is changed to random prediction. Obviously, it severely degrades model accuracy. Although [25] proposed coverage accuracy control for image classification, it is impractical to semantic segmentation because mIoU is hard to be tuned with calibrator training set." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 344, + 406, + 545, + 437 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 406, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 344, + 406, + 545, + 437 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l l} \\hat {\\boldsymbol {p}} = \\mathbf {1} / k, & \\text {i f} - \\hat {p} \\log (\\hat {p}) > \\gamma , \\\\ \\hat {\\boldsymbol {p}} = \\sigma_ {S M} (\\boldsymbol {z} / T), & \\text {o t h e r w i s e}, \\end{array} \\right. \\tag {8}", + "image_path": "17acd7772421da6eb98cdea830d34a1e43ab9b80a778fec5c29ad30103347fa8.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 305, + 438, + 545, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 438, + 545, + 461 + ], + "spans": [ + { + "bbox": [ + 305, + 438, + 545, + 461 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 438, + 545, + 461 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 305, + 438, + 545, + 461 + ], + "type": "text", + "content": " is the number of classes, " + }, + { + "bbox": [ + 305, + 438, + 545, + 461 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 305, + 438, + 545, + 461 + ], + "type": "text", + "content": " is the entropy threshold computed with a ranking model for calibration decision." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 305, + 462, + 545, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 462, + 545, + 558 + ], + "spans": [ + { + "bbox": [ + 305, + 462, + 545, + 558 + ], + "type": "text", + "content": "Local Temperature Scaling (LTS) is the extension of temperature scaling to semantic segmentation [6]. It provides pixel-level heterogeneous scaling temperature by combing input images with output logit through CNN. Although it considers spatial heterogeneity over an image for scaling, the calibrator training relies on strong correlation between image features and temperature heterogeneity and more training data for feature extraction [6]." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 375, + 563, + 545, + 576 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 563, + 545, + 576 + ], + "spans": [ + { + "bbox": [ + 375, + 563, + 545, + 576 + ], + "type": "interline_equation", + "content": "\\hat {\\boldsymbol {p}} = \\sigma_ {S M} (\\boldsymbol {z} / (\\boldsymbol {T} (\\boldsymbol {z}, I)), \\tag {9}", + "image_path": "287336b24d5aa7d0f40bfea894369619d219474f58afa2d0c115a1f2d92ad65a.jpg" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 305, + 580, + 545, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 580, + 545, + 604 + ], + "spans": [ + { + "bbox": [ + 305, + 580, + 545, + 604 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 580, + 545, + 604 + ], + "type": "inline_equation", + "content": "\\pmb{T}(\\pmb{z}, I)" + }, + { + "bbox": [ + 305, + 580, + 545, + 604 + ], + "type": "text", + "content": " is a pixel-wise temperature map by a convolution with segmented logit vector " + }, + { + "bbox": [ + 305, + 580, + 545, + 604 + ], + "type": "inline_equation", + "content": "\\pmb{z}" + }, + { + "bbox": [ + 305, + 580, + 545, + 604 + ], + "type": "text", + "content": " and image " + }, + { + "bbox": [ + 305, + 580, + 545, + 604 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 305, + 580, + 545, + 604 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 305, + 605, + 545, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 605, + 545, + 688 + ], + "spans": [ + { + "bbox": [ + 305, + 605, + 545, + 688 + ], + "type": "text", + "content": "Ensembling is proposed to solve medical image segmentation uncertainty prediction [27]. It carries out calibration improvement with simple average of ensemble model, a simplified version of Bayesian inference [47]. Also, deep ensemble is a strong baseline for image classification calibration [21, 47]. We compare its performance with other post-hoc approaches." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 390, + 693, + 545, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 390, + 693, + 545, + 717 + ], + "spans": [ + { + "bbox": [ + 390, + 693, + 545, + 717 + ], + "type": "interline_equation", + "content": "\\hat {\\boldsymbol {p}} = \\frac {1}{N} \\Sigma_ {n = 1} ^ {N} \\hat {\\boldsymbol {p}} ^ {n}, \\tag {10}", + "image_path": "2f9f0522818f846cf942fbcf0ad47e311d5d7ce49d6a44fa8c8eb7fb9d5e9d71.jpg" + } + ] + } + ], + "index": 30 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23656" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 235, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 235, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 235, + 83 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 47, + 72, + 235, + 83 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 47, + 72, + 235, + 83 + ], + "type": "text", + "content": " is the number of ensemble members." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 95, + 128, + 109 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 95, + 128, + 109 + ], + "spans": [ + { + "bbox": [ + 47, + 95, + 128, + 109 + ], + "type": "text", + "content": "6. Experiments" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 115, + 286, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 115, + 286, + 137 + ], + "spans": [ + { + "bbox": [ + 47, + 115, + 286, + 137 + ], + "type": "text", + "content": "Models. We consider five recent state-of-the-art segmentation models which cover CNN and ViT architectures." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 139, + 287, + 306 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 55, + 139, + 287, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 139, + 287, + 175 + ], + "spans": [ + { + "bbox": [ + 55, + 139, + 287, + 175 + ], + "type": "text", + "content": "1. SegFormer [49] is a ViT-based encoder-decoder model based upon lightweight multilayer perception (MLP) decoders with pyramid architecture." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 175, + 287, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 175, + 287, + 198 + ], + "spans": [ + { + "bbox": [ + 55, + 175, + 287, + 198 + ], + "type": "text", + "content": "2. Segmenter [40] is an encoder-decoder model based exclusively on Transformer." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 199, + 287, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 199, + 287, + 270 + ], + "spans": [ + { + "bbox": [ + 55, + 199, + 287, + 270 + ], + "type": "text", + "content": "3. Knet [51] is a unified segmentation decoder module which includes two important variants, Knet-DeepLab and Knet-SWIN. Knet-DeepLab consists of ResNet-50 [12] backbone and DeepLab-V3 [2] decoder. Knet-SWIN is constructed with SWIN [23] backbone and UperNet [48] decoder." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 271, + 287, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 271, + 287, + 306 + ], + "spans": [ + { + "bbox": [ + 55, + 271, + 287, + 306 + ], + "type": "text", + "content": "4. ConvNeXt [24] is a CNN backbone, but shaped to ViT-like architecture for better scaling. We select ConvNeXt backbone with UperNet [48] decoder." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 47, + 307, + 287, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 307, + 287, + 354 + ], + "spans": [ + { + "bbox": [ + 47, + 307, + 287, + 354 + ], + "type": "text", + "content": "For ADE20K, all models except Knet-DeepLab are trained with " + }, + { + "bbox": [ + 47, + 307, + 287, + 354 + ], + "type": "inline_equation", + "content": "640 \\times 640" + }, + { + "bbox": [ + 47, + 307, + 287, + 354 + ], + "type": "text", + "content": " crop size. For BDD100K and SYNTHIA, the crop size is " + }, + { + "bbox": [ + 47, + 307, + 287, + 354 + ], + "type": "inline_equation", + "content": "512 \\times 1024" + }, + { + "bbox": [ + 47, + 307, + 287, + 354 + ], + "type": "text", + "content": ". For other benchmarks, the crop size is " + }, + { + "bbox": [ + 47, + 307, + 287, + 354 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 47, + 307, + 287, + 354 + ], + "type": "text", + "content": ". The batch size is set to 8." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 354, + 287, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 354, + 287, + 390 + ], + "spans": [ + { + "bbox": [ + 47, + 354, + 287, + 390 + ], + "type": "text", + "content": "Dataset. We examine existing calibration methods across six important benchmarks from various applications. Figure 3 illustrates the example images from benchmarks." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 391, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 55, + 391, + 287, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 391, + 287, + 474 + ], + "spans": [ + { + "bbox": [ + 55, + 391, + 287, + 474 + ], + "type": "text", + "content": "1. Scene and stuff segmentation. We use ADE20K [54] and COCO-164K [1] as large-scale segmentation benchmark. ADE20K contains 150 object and stuff classes with 20,210/2,000 images in the training/validation set. COCO-164K, " + }, + { + "bbox": [ + 55, + 391, + 287, + 474 + ], + "type": "inline_equation", + "content": "\\sim 164\\mathrm{K}" + }, + { + "bbox": [ + 55, + 391, + 287, + 474 + ], + "type": "text", + "content": " images for 91 stuff and 80 thing classes, includes " + }, + { + "bbox": [ + 55, + 391, + 287, + 474 + ], + "type": "inline_equation", + "content": "118\\mathrm{K} / 5\\mathrm{K}" + }, + { + "bbox": [ + 55, + 391, + 287, + 474 + ], + "type": "text", + "content": " images in the training/validation set." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 474, + 287, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 287, + 581 + ], + "type": "text", + "content": "2. Autonomous driving. We choose BDD-100K [50], a latest benchmark for urban driving scene segmentation. BDD-100K contains 7K/1K " + }, + { + "bbox": [ + 55, + 474, + 287, + 581 + ], + "type": "inline_equation", + "content": "1280 \\times 720" + }, + { + "bbox": [ + 55, + 474, + 287, + 581 + ], + "type": "text", + "content": " images in 19 classes for training/validation set. We select the validation set of CityScapes [4] as test set from target domain for domain-shift calibration assessment. CityScapes has 2975/500 " + }, + { + "bbox": [ + 55, + 474, + 287, + 581 + ], + "type": "inline_equation", + "content": "1024 \\times 2048" + }, + { + "bbox": [ + 55, + 474, + 287, + 581 + ], + "type": "text", + "content": " images for training/validation set. BDD-100K and CityScapes share the same label space." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 582, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 287, + 677 + ], + "type": "text", + "content": "3. Video segmentation. We adopt DAVIS2016 [37] to assess calibration on temporal domain shift problems. DAVIS2016 contains 50 object and 1 background classes. We select 480p resolution video sequences to prepare temporal domain shift dataset. The total of the frames in each sequence ranges from 24 to 103. Each frame is an image segmented with a single object and background." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 287, + 713 + ], + "type": "text", + "content": "4. Remote sensing. We select SPACENET-7 (SN-7) [44], a binary small object identification benchmark, to examine spatial and temporal domain shift calibration." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 326, + 72, + 545, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 72, + 545, + 144 + ], + "spans": [ + { + "bbox": [ + 326, + 72, + 545, + 144 + ], + "type": "text", + "content": "SN-7 consists of 1,408 " + }, + { + "bbox": [ + 326, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 326, + 72, + 545, + 144 + ], + "type": "text", + "content": " satellite images collected from monthly urban development across 101 cities. It contains 11,080,000 buildings over 41,000 " + }, + { + "bbox": [ + 326, + 72, + 545, + 144 + ], + "type": "inline_equation", + "content": "\\mathrm{km}^2" + }, + { + "bbox": [ + 326, + 72, + 545, + 144 + ], + "type": "text", + "content": " observation area. Due to the development, images vary temporally; due to planning difference, the objects like the building styles differ by locations." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 144, + 545, + 311 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 313, + 144, + 545, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 144, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 313, + 144, + 545, + 227 + ], + "type": "text", + "content": "5. Medical imaging. We select BraTS2017 [14], a popular brain tumor segmentation dataset. BraTS2017 contains multimodal magnetic resonance imaging (MRI) scans from 1,210 patients, which consists of 3D voxels with 155 layers from top to bottom. Each layer is equivalent to a 2D semantic segmentation image with 3 tumor feature classes." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 228, + 545, + 311 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 545, + 311 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 545, + 311 + ], + "type": "text", + "content": "6. Simulation domain transfer. SYNTHIA [39] is a popular urban street synthetic image benchmark. We select its subset of RANDCITYSCAPES, which contains 9400 " + }, + { + "bbox": [ + 313, + 228, + 545, + 311 + ], + "type": "inline_equation", + "content": "1280 \\times 760" + }, + { + "bbox": [ + 313, + 228, + 545, + 311 + ], + "type": "text", + "content": " images with 22 classes, as source domain for model training. We select CityScapes [4] as target domain for domain-shift calibration assessment." + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 314, + 545, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 314, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 306, + 314, + 545, + 374 + ], + "type": "text", + "content": "In-domain calibration. We select the validation set of each benchmark for in-domain calibration study. For ADE20K, COCO-164K, and BDD100K, we select 100/100/1800 images, 250/250/4500 images, and 50/50/900 images for calibrator training/validation/testing." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 305, + 376, + 546, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 376, + 546, + 579 + ], + "spans": [ + { + "bbox": [ + 305, + 376, + 546, + 579 + ], + "type": "text", + "content": "Domain-shift calibration. We select the validation sets of COCO-164K and CityScapes as testing sets of ADE20K and BDD100K. For taxonomy incompatibility over datasets, like ADE20K-COCO164K, we follow the convention in [22] to merge and split label classes. For DAVIS, we merge all training and validation sequences and resplit them in half by frame. The first half of frames are used for training with 1731 images and the second half are split for validation/testing with 100/1624 images. For SN7, all training data are split into 1133/50/225 images for training/validation/testing set by location, and 698/50/660 images for training/validation/testing set by time. For BraTS, we select 800/50 images of " + }, + { + "bbox": [ + 305, + 376, + 546, + 579 + ], + "type": "inline_equation", + "content": "75^{th}" + }, + { + "bbox": [ + 305, + 376, + 546, + 579 + ], + "type": "text", + "content": " layer scan for training/validation and 360 images of " + }, + { + "bbox": [ + 305, + 376, + 546, + 579 + ], + "type": "inline_equation", + "content": "100^{th}" + }, + { + "bbox": [ + 305, + 376, + 546, + 579 + ], + "type": "text", + "content": " layer scan for testing. For SYNTHIA, we split training set into 9000/400 images for training/validation, and CityScapes' validation set for testing." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 545, + 713 + ], + "type": "text", + "content": "Calibration methods. We select aforementioned six existing calibration methods to compare selective scaling. For selective scaling, selective model is simplified with three-layer MLP due to training data restriction. The inputs are pixel-wise predictive probabilities and the labels are binary which indicate correct/incorrect prediction. Moreover, Meta-cal with miscoverage control yields severe model degradation. We implement our extension by replacing random prediction with large temperature scaling to retain accuracy. For ensembling, we achieve a three-member ensemble with comparable accuracy for fair comparison." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23657" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 90, + 106, + 504, + 262 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "type": "text", + "content": "Table 1. Segmentation accuracy (mIoU) and calibration error (ECE) on different benchmarks. TempS, LogS, DirS, LTS, Ens., and Selective denote temperature, logistic, Dirichlet, local temperature, ensembling, and selective scaling, respectively. Ensembling is carried out by three models with reduced size for comparable mIoU. Meta-Cal* is implemented by our extension with large temperature." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 90, + 106, + 504, + 262 + ], + "lines": [ + { + "bbox": [ + 90, + 106, + 504, + 262 + ], + "spans": [ + { + "bbox": [ + 90, + 106, + 504, + 262 + ], + "type": "table", + "html": "
DatasetModelmIoUUncalTempSLogSDirSMeta-Cal*LTSEns.Selective
ADE20KSegFormer-B5 [49]49.130.1110.1090.1100.1100.1030.1050.1090.086
ADE20KSegmenter-L [40]51.650.0870.0860.0860.0870.0810.0940.0860.069
ADE20KKnet-DeepLab [51]45.060.1110.1050.1070.1060.1020.1180.1100.095
ADE20KKnet-SWIN-L [51]52.460.0980.0940.0930.0970.0890.1340.0960.078
ADE20KConvNeXt-L [24]53.160.0970.0920.0940.0910.0880.1330.0940.082
COCO-164KSegFormer-B5 [49]45.780.1510.1490.1410.1510.1320.1510.1490.113
COCO-164KSegmenter-L [40]47.090.1520.1490.1490.1510.1300.1550.1500.109
COCO-164KKnet-DeepLab [51]37.240.1700.1700.1680.1710.1490.1720.1690.093
COCO-164KKnet-SWIN-L [51]46.490.1610.1590.1610.1600.1420.1620.1600.102
COCO-164KConvNeXt-L [24]46.480.1600.1570.1580.1590.1410.1620.1590.108
BDD100KSegFormer-B5 [49]65.080.0640.0550.0540.0530.0490.0690.0590.040
BDD100KSegmenter-L [40]61.330.0550.0450.0430.0420.0370.0710.0520.031
BDD100KKnet-DeepLab [51]62.890.0600.0490.0470.0480.0410.0630.0570.035
BDD100KKnet-SWIN-L [51]67.590.0650.0550.0540.0540.0490.0670.0630.040
BDD100KConvNeXt-L [24]67.260.0640.0540.0530.0560.0490.0650.0630.038
", + "image_path": "7ab62870cf3a7d7862c14c0e93e9711aad1a79d0f56713f197291f178ffc7560.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 281, + 105, + 293 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 281, + 105, + 293 + ], + "spans": [ + { + "bbox": [ + 47, + 281, + 105, + 293 + ], + "type": "text", + "content": "6.1. Results" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 299, + 221, + 311 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 299, + 221, + 311 + ], + "spans": [ + { + "bbox": [ + 47, + 299, + 221, + 311 + ], + "type": "text", + "content": "6.1.1 In-domain calibration evaluation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 318, + 287, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 318, + 287, + 426 + ], + "spans": [ + { + "bbox": [ + 46, + 318, + 287, + 426 + ], + "type": "text", + "content": "We present ECEs for five models and compare selective scaling with six existing calibrators on three-run average. Table 1 shows that majority of calibrators yield very limited improvement in model calibration. However, surprisingly, selective scaling consistently outperforms these scaling approaches. This calibration gain is attributed to more scaling on mispredictions. Moreover, this improvement is more significant for scene and object benchmarks such as ADE20K and COCO-164K than street view datasets like BDD-100K." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 426, + 287, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 426, + 287, + 546 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 287, + 546 + ], + "type": "text", + "content": "Furthermore, we find that BDD100K yields more calibrated models than other benchmarks. We associate it with larger crop size " + }, + { + "bbox": [ + 46, + 426, + 287, + 546 + ], + "type": "inline_equation", + "content": "(512 \\times 1024)" + }, + { + "bbox": [ + 46, + 426, + 287, + 546 + ], + "type": "text", + "content": " and higher model accuracy. We also find that ensembling exhibits weaker calibration. We connect it with model accuracy since we restrain the ensemble accuracy for fair comparison, which limits its calibration performance. Moreover, across models, Segmenter exhibits better calibration. Since it is a Transformer-exclusive model, we link it to different spatial inductive bias which is speculated in [28]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 551, + 233, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 551, + 233, + 562 + ], + "spans": [ + { + "bbox": [ + 47, + 551, + 233, + 562 + ], + "type": "text", + "content": "6.1.2 Domain-shift calibration evaluation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 570, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 287, + 689 + ], + "type": "text", + "content": "We present ECEs for five models on seven sets of domain-shift experiments and compare selective scaling with six existing calibrators on three-run average. Table 2 shows that selective scaling still consistently outperforms by a large margin. It implies that selective scaling can be generalized to domain-shift problems. In particular, Meta-cal exhibits performance variation. We speculate that the hard threshold from entropy calculation limits separating shift data. In contrast, selective scaling is more consistent because binary classifier has better generalization to separating shift data." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 715 + ], + "type": "text", + "content": "Across benchmarks, we find that label space shift causes more significant miscalibration. For example, ADE20K" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 328, + 277, + 421, + 375 + ], + "blocks": [ + { + "bbox": [ + 328, + 277, + 421, + 375 + ], + "lines": [ + { + "bbox": [ + 328, + 277, + 421, + 375 + ], + "spans": [ + { + "bbox": [ + 328, + 277, + 421, + 375 + ], + "type": "image", + "image_path": "0a87e2d16e8da0f11dd59031680315baf5004f89dfe560ddbc58eaa3a36bd702.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 383, + 545, + 438 + ], + "lines": [ + { + "bbox": [ + 305, + 383, + 545, + 438 + ], + "spans": [ + { + "bbox": [ + 305, + 383, + 545, + 438 + ], + "type": "text", + "content": "Figure 4. The effect of selector accuracy on calibration. Misprediction detection accuracy denotes the percentage of misprediction detected by selector. The lower bound (optimal case) is " + }, + { + "bbox": [ + 305, + 383, + 545, + 438 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 305, + 383, + 545, + 438 + ], + "type": "text", + "content": " misprediction detected for scaling. The upper bound (worst case) is " + }, + { + "bbox": [ + 305, + 383, + 545, + 438 + ], + "type": "inline_equation", + "content": "0\\%" + }, + { + "bbox": [ + 305, + 383, + 545, + 438 + ], + "type": "text", + "content": ", i.e., none misprediction detected is used for calibration." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 432, + 278, + 525, + 375 + ], + "blocks": [ + { + "bbox": [ + 432, + 278, + 525, + 375 + ], + "lines": [ + { + "bbox": [ + 432, + 278, + 525, + 375 + ], + "spans": [ + { + "bbox": [ + 432, + 278, + 525, + 375 + ], + "type": "image", + "image_path": "9fc311d02c318bf937c1bf3a0be095e56372c57a2f9b995c949127010388b99b.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 450, + 545, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 450, + 545, + 570 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 545, + 570 + ], + "type": "text", + "content": "to COCO-164K and SYNTHIA to CityScapes show worse ECE than other domain-shift experiments. We link it to poorer domain adaptation and model generalization degradation. Also, we observe that small object segmentation like SN-7, as a challenged task, can cause worse miscalibration. Except for SN-7, more accurate models are more calibrated by benchmarks. For SN-7, background pixels engage in training but are excluded from calibration calculation, which causes larger ECE. Within these models, we find that Segmenter is still more calibrated." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 578, + 402, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 578, + 402, + 590 + ], + "spans": [ + { + "bbox": [ + 306, + 578, + 402, + 590 + ], + "type": "text", + "content": "6.1.3 Ablation study" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 593, + 545, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 593, + 545, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 593, + 545, + 640 + ], + "type": "text", + "content": "To analyze the effect of selector on selective scaling, we conduct ablation study by varying misprediction detection accuracy, integrating with different existing calibrators, and examining calibration error across regions." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 641, + 545, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 545, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 545, + 700 + ], + "type": "text", + "content": "Selective scaling considering selection accuracy. Figure 4 shows that with the increase in selection accuracy, ECE decreases. This implies that when more mispredictions are detected and scaled, model calibration becomes better. It justifies the effectiveness of selective scaling." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 701, + 545, + 713 + ], + "type": "text", + "content": "Selective scaling with existing calibrators. We also" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23658" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 73, + 95, + 519, + 441 + ], + "blocks": [ + { + "bbox": [ + 47, + 71, + 545, + 93 + ], + "lines": [ + { + "bbox": [ + 47, + 71, + 545, + 93 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 545, + 93 + ], + "type": "text", + "content": "Table 2. Segmentation model accuracy (mIoU) and calibration error (ECE) on different benchmarks for domain-shift calibration assessment. InD and SD denotes in-domain and domain-shift test data. Meta-Cal* is our extension with large temperature scaling." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 73, + 95, + 519, + 441 + ], + "lines": [ + { + "bbox": [ + 73, + 95, + 519, + 441 + ], + "spans": [ + { + "bbox": [ + 73, + 95, + 519, + 441 + ], + "type": "table", + "html": "
InDSDModelmIoUUncalTempSLogSDirSMeta-Cal*LTSEns.Selective
ADE20KCOCO-164KSegFormer-B5 [49]8.200.4670.4660.4680.4690.4150.4200.4650.331
ADE20KCOCO-164KSegmenter-L [40]9.600.3950.3940.3950.3940.3220.3310.3930.259
ADE20KCOCO-164KKnet-DeepLab [51]7.020.4470.4450.4460.4460.4030.4390.4440.351
ADE20KCOCO-164KKnet-SWIN-L [51]9.060.4640.4620.4630.4630.3790.4650.4620.327
ADE20KCOCO-164KConvNeXt-L [24]8.990.4610.4580.4590.4600.3770.4600.4590.320
BDD100KCityScapesSegFormer-B5 [49]63.050.0870.0830.0820.0840.0760.0850.0870.069
BDD100KCityScapesSegmenter-L [40]60.260.0620.0580.0590.0570.0550.0670.0600.048
BDD100KCityScapesKnet-DeepLab [51]61.270.0730.0700.0690.0710.0630.0730.0710.056
BDD100KCityScapesKnet-SWIN-L [51]67.460.0800.0780.0770.0770.0710.0790.0790.063
BDD100KCityScapesConvNeXt-L [24]67.180.0810.0790.0790.0770.0700.0800.0790.065
DAVIS-trainDAVIS-testSegFormer-B4 [49]89.330.0330.0310.0310.0310.0550.0980.0320.024
DAVIS-trainDAVIS-testSegmenter-B [40]81.350.0800.0770.0760.0750.1030.1210.0780.049
DAVIS-trainDAVIS-testKnet-DeepLab [51]83.120.0760.0720.0730.0730.0980.1070.0750.047
DAVIS-trainDAVIS-testKnet-SWIN-B [51]89.220.0450.0410.0410.0400.0650.1010.0440.029
DAVIS-trainDAVIS-testConvNeXt-B [24]89.000.0440.0400.0410.0420.0670.1000.0440.031
SN-7-SP-trainSN-7-SP-testSegFormer-B4 [49]57.230.6550.6360.6250.6340.7110.7010.6520.532
SN-7-SP-trainSN-7-SP-testSegmenter-B [40]51.980.7660.7510.7480.7400.7960.7890.7630.687
SN-7-SP-trainSN-7-SP-testKnet-DeepLab [51]55.580.7040.6880.6790.6810.7680.7200.6960.655
SN-7-SP-trainSN-7-SP-testKnet-SWIN-B [51]59.420.6090.5760.5590.5630.6310.6180.6050.504
SN-7-SP-trainSN-7-SP-testConvNeXt-B [24]58.690.6110.5700.5640.5660.6290.6330.6080.517
SN-7-TS-trainSN-7-TS-testSegFormer-B4 [49]54.390.7000.6730.6690.6730.7120.7230.6950.582
SN-7-TS-trainSN-7-TS-testSegmenter-B [40]50.130.7690.7450.7400.7360.7850.7770.7550.624
SN-7-TS-trainSN-7-TS-testKnet-DeepLab [51]56.730.6510.6220.6210.6190.6890.6750.6470.576
SN-7-TS-trainSN-7-TS-testKnet-SWIN-B [51]62.420.6190.5870.5790.5840.6300.6210.6090.551
SN-7-TS-trainSN-7-TS-testConvNeXt-B [24]62.140.6230.5780.5710.5730.6470.6380.6200.549
BraTS-trainBraTS-testSegFormer-B4 [49]45.810.1550.1370.1350.1350.1290.1500.1510.120
BraTS-trainBraTS-testSegmenter-B [40]44.570.1540.1490.1460.1470.1350.1420.1500.124
BraTS-trainBraTS-testKnet-DeepLab [51]46.050.1880.1790.1760.1760.1580.1870.1850.148
BraTS-trainBraTS-testKnet-SWIN-B [51]47.990.1550.1460.1430.1440.1320.1570.1520.123
BraTS-trainBraTS-testConvNeXt-B [24]48.400.1690.1590.1540.1550.1360.1680.1650.121
SYNTHIACityScapesSegFormer-B5 [49]33.120.2860.2640.2620.2630.2300.2880.2840.218
SYNTHIACityScapesSegmenter-L [40]32.330.2540.2350.2320.2300.2170.2800.2520.203
SYNTHIACityScapesKnet-DeepLab [51]30.190.3110.2980.2960.2960.2720.3050.3090.241
SYNTHIACityScapesKnet-SWIN-L [51]34.310.2810.2730.2700.2700.2390.2880.2790.226
SYNTHIACityScapesConvNeXt-L [24]34.790.2870.2670.2630.2640.2300.2990.2840.211
", + "image_path": "32dc9c82157ff9f61e931d18550cd62cae726c9d64b62a2cbcb345a055812329.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 453, + 287, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 453, + 287, + 514 + ], + "spans": [ + { + "bbox": [ + 46, + 453, + 287, + 514 + ], + "type": "text", + "content": "find that selective scaling can boost existing calibrators, vice versa. Figure 4 shows that the integration with selective scaling helps existing calibrators to improve model calibration. When detection accuracy is lower, like " + }, + { + "bbox": [ + 46, + 453, + 287, + 514 + ], + "type": "inline_equation", + "content": "< 20\\%" + }, + { + "bbox": [ + 46, + 453, + 287, + 514 + ], + "type": "text", + "content": ", existing calibrators can help improve selective scaling." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 514, + 288, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 514, + 288, + 610 + ], + "spans": [ + { + "bbox": [ + 46, + 514, + 288, + 610 + ], + "type": "text", + "content": "Selective scaling across regions. Figure 5 shows segmentation calibration error statistics across regions. Top and bottom bars denote maximum and minimum ECE while top, central, and bottom lines of the box indicate " + }, + { + "bbox": [ + 46, + 514, + 288, + 610 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 46, + 514, + 288, + 610 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 514, + 288, + 610 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 46, + 514, + 288, + 610 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 514, + 288, + 610 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 46, + 514, + 288, + 610 + ], + "type": "text", + "content": " of ECE distribution. It shows that: 1) boundary pixels are less calibrated than non-boundary pixels within an object bounding box; 2) selective scaling consistently yields lower mean ECE across both types of pixels." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 66, + 621, + 265, + 673 + ], + "blocks": [ + { + "bbox": [ + 66, + 621, + 265, + 673 + ], + "lines": [ + { + "bbox": [ + 66, + 621, + 265, + 673 + ], + "spans": [ + { + "bbox": [ + 66, + 621, + 265, + 673 + ], + "type": "image", + "image_path": "d25399fd587bb1b8d1de5c033566ec0cf1ae68ebf821dbe173ab23e45ed594b8.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 676, + 287, + 719 + ], + "lines": [ + { + "bbox": [ + 46, + 676, + 287, + 719 + ], + "spans": [ + { + "bbox": [ + 46, + 676, + 287, + 719 + ], + "type": "text", + "content": "Figure 5. SegFormer-B5 ECE comparison in ADE20K between boundary and non-boundary pixels among uncalibration, temperature scaling, and selective scaling. Crosses and dots show means and outliers. The numbers are the means." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 306, + 453, + 379, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 453, + 379, + 464 + ], + "spans": [ + { + "bbox": [ + 306, + 453, + 379, + 464 + ], + "type": "text", + "content": "6.2. Limitation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 472, + 545, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 545, + 568 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 545, + 568 + ], + "type": "text", + "content": "Logit smoothing can cause entropy increase because post-hoc over-confidence calibration, such as temperature scaling, is equivalent to entropy maximization under certain logit constraints [11]. For example, original Meta-cal leads to significant entropy increase due to randomizing certain prediction outputs. Our method alleviates this issue, but still yields higher entropy. To resolve it, careful temperature tuning is a trade-off scheme for entropy improvement." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 574, + 378, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 574, + 378, + 586 + ], + "spans": [ + { + "bbox": [ + 306, + 574, + 378, + 586 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "We analyzed semantic segmentation model calibration and proposed a simple but effective algorithm of selective scaling. We revealed the possible factors of segmentation model miscalibration. Given the analysis, we designed selective scaling to scale logits separately, and more focus on mispredicted logit smoothing. We performed extensive experiments on state-of-the-art models across various benchmarks. Both in-domain and domain-shift experiments justify the effectiveness of selective scaling. We also presented our observations for segmentation calibration research." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23659" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 289, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 91, + 289, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 289, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 289, + 135 + ], + "type": "text", + "content": "[1] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. Cocostuff: Thing and stuff classes in context. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1209-1218, 2018. 1, 2, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 136, + 288, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 136, + 288, + 178 + ], + "spans": [ + { + "bbox": [ + 53, + 136, + 288, + 178 + ], + "type": "text", + "content": "[2] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In ECCV, 2018. 1, 3, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 180, + 288, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 180, + 288, + 224 + ], + "spans": [ + { + "bbox": [ + 53, + 180, + 288, + 224 + ], + "type": "text", + "content": "[3] MMSegmentation Contributors. MMSegmentation: Openmmlab semantic segmentation toolbox and benchmark. https://github.com/open-mmlab/mmsegmentation, 2020.3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 225, + 288, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 225, + 288, + 290 + ], + "spans": [ + { + "bbox": [ + 53, + 225, + 288, + 290 + ], + "type": "text", + "content": "[4] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3213-3223, 2016. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 291, + 288, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 291, + 288, + 334 + ], + "spans": [ + { + "bbox": [ + 53, + 291, + 288, + 334 + ], + "type": "text", + "content": "[5] Morris H DeGroot and Stephen E Fienberg. The comparison and evaluation of forecasters. Journal of the Royal Statistical Society: Series D (The Statistician), 32(1-2):12-22, 1983. 1, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 335, + 288, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 335, + 288, + 380 + ], + "spans": [ + { + "bbox": [ + 53, + 335, + 288, + 380 + ], + "type": "text", + "content": "[6] Zhipeng Ding, Xu Han, Peirong Liu, and Marc Niethammer. Local temperature scaling for probability calibration. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6889-6899, 2021. 2, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 380, + 288, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 380, + 288, + 435 + ], + "spans": [ + { + "bbox": [ + 53, + 380, + 288, + 435 + ], + "type": "text", + "content": "[7] Mingyuan Fan, Shenqi Lai, Junshi Huang, Xiaoming Wei, Zhenhua Chai, Junfeng Luo, and Xiaolin Wei. Rethinking bisenet for real-time semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9716-9725, 2021. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 436, + 288, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 436, + 288, + 479 + ], + "spans": [ + { + "bbox": [ + 53, + 436, + 288, + 479 + ], + "type": "text", + "content": "[8] Moritz Fuchs, Camila Gonzalez, and Anirban Mukhopadhyay. Practical uncertainty quantification for brain tumor segmentation. In Medical Imaging with Deep Learning, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 480, + 288, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 480, + 288, + 535 + ], + "spans": [ + { + "bbox": [ + 53, + 480, + 288, + 535 + ], + "type": "text", + "content": "[9] Jakob Gawlikowski, Cedrique Rovile Njieutcheu Tassi, Mohsin Ali, Jongseok Lee, Matthias Humt, Jianxiang Feng, Anna Kruspe, Rudolph Triebel, Peter Jung, Ribana Roscher, et al. A survey of uncertainty in deep neural networks. arXiv preprint arXiv:2107.03342, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 536, + 288, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 288, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 288, + 590 + ], + "type": "text", + "content": "[10] Yunye Gong, Xiao Lin, Yi Yao, Thomas G Dietterich, Ajay Divakaran, and Melinda Gervasio. Confidence calibration for domain generalization under covariate shift. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8958-8967, 2021. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 591, + 288, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 591, + 288, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 591, + 288, + 635 + ], + "type": "text", + "content": "[11] Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q Weinberger. On calibration of modern neural networks. In International conference on machine learning, pages 1321-1330. PMLR, 2017. 1, 2, 4, 5, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 635, + 288, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 288, + 680 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 288, + 680 + ], + "type": "text", + "content": "[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 681, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 288, + 714 + ], + "type": "text", + "content": "[13] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. arXiv preprint arXiv:1903.12261, 2019. 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 138 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 138 + ], + "type": "text", + "content": "[14] Fabian Isensee, Philipp Kickingereder, Wolfgang Wick, Martin Bendszus, and Klaus H Maier-Hein. Brain tumor segmentation and radiomics survival prediction: Contribution to the brats 2017 challenge. In International MICCAI Brainlesion Workshop, pages 287–297. Springer, 2017. 1, 2, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 140, + 547, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 140, + 547, + 196 + ], + "spans": [ + { + "bbox": [ + 307, + 140, + 547, + 196 + ], + "type": "text", + "content": "[15] Pavel Izmailov, Wesley J Maddox, Polina Kirichenko, Timur Garipov, Dmitry Vetrov, and Andrew Gordon Wilson. Subspace inference for bayesian deep learning. In Uncertainty in Artificial Intelligence, pages 1169-1179. PMLR, 2020. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 198, + 547, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 198, + 547, + 253 + ], + "spans": [ + { + "bbox": [ + 307, + 198, + 547, + 253 + ], + "type": "text", + "content": "[16] Alain Jungo and Mauricio Reyes. Assessing reliability and challenges of uncertainty estimations for medical image segmentation. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 48-56. Springer, 2019. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 255, + 547, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 255, + 547, + 310 + ], + "spans": [ + { + "bbox": [ + 307, + 255, + 547, + 310 + ], + "type": "text", + "content": "[17] Elias Kassapis, Georgi Dikov, Deepak K Gupta, and Cedric Nugteren. Calibrated adversarial refinement for stochastic semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7057-7067, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 312, + 546, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 312, + 546, + 346 + ], + "spans": [ + { + "bbox": [ + 307, + 312, + 546, + 346 + ], + "type": "text", + "content": "[18] Alex Kendall and Yarin Gal. What uncertainties do we need in bayesian deep learning for computer vision? Advances in neural information processing systems, 30, 2017. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 348, + 546, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 348, + 546, + 403 + ], + "spans": [ + { + "bbox": [ + 307, + 348, + 546, + 403 + ], + "type": "text", + "content": "[19] Meelis Kull, Miquel Perello Nieto, Markus Kangsepp, Telmo Silva Filho, Hao Song, and Peter Flach. Beyond temperature scaling: Obtaining well-calibrated multi-class probabilities with dirichlet calibration. Advances in neural information processing systems, 32, 2019. 1, 2, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 405, + 546, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 405, + 546, + 449 + ], + "spans": [ + { + "bbox": [ + 307, + 405, + 546, + 449 + ], + "type": "text", + "content": "[20] Meelis Kull, Telmo Silva Filho, and Peter Flach. Beta calibration: a well-founded and easily implemented improvement on logistic calibration for binary classifiers. In Artificial Intelligence and Statistics, pages 623-631. PMLR, 2017. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 452, + 546, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 452, + 546, + 495 + ], + "spans": [ + { + "bbox": [ + 307, + 452, + 546, + 495 + ], + "type": "text", + "content": "[21] Balaji Lakshminarayanan, Alexander Pritzel, and Charles Blundell. Simple and scalable predictive uncertainty estimation using deep ensembles. Advances in neural information processing systems, 30, 2017. 1, 2, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 498, + 547, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 498, + 547, + 553 + ], + "spans": [ + { + "bbox": [ + 307, + 498, + 547, + 553 + ], + "type": "text", + "content": "[22] John Lambert, Zhuang Liu, Ozan Sener, James Hays, and Vladlen Koltun. Mseg: A composite dataset for multi-domain semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2879-2888, 2020. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 555, + 546, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 555, + 546, + 609 + ], + "spans": [ + { + "bbox": [ + 307, + 555, + 546, + 609 + ], + "type": "text", + "content": "[23] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10012-10022, 2021. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 612, + 546, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 612, + 546, + 666 + ], + "spans": [ + { + "bbox": [ + 307, + 612, + 546, + 666 + ], + "type": "text", + "content": "[24] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 669, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 669, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 669, + 547, + 713 + ], + "type": "text", + "content": "[25] Xingchen Ma and Matthew B Blaschko. Meta-cal: Well-controlled post-hoc calibration by ranking. In International Conference on Machine Learning, pages 7235-7245. PMLR, 2021. 1, 2, 5" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "23660" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 105 + ], + "type": "text", + "content": "[26] Andrey Malinin, Bruno Mlodozeniec, and Mark Gales. Ensemble distribution distillation. arXiv preprint arXiv:1905.00076, 2019. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 107, + 287, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 107, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 49, + 107, + 287, + 162 + ], + "type": "text", + "content": "[27] Alireza Mehrtash, William M Wells, Clare M Tempany, Purang Abolmaesumi, and Tina Kapur. Confidence calibration and predictive uncertainty estimation for deep medical image segmentation. IEEE transactions on medical imaging, 39(12):3868-3878, 2020. 2, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 163, + 287, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 163, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 49, + 163, + 287, + 217 + ], + "type": "text", + "content": "[28] Matthias Minderer, Josip Djolonga, Rob Romijnders, Frances Hubis, Xiaohua Zhai, Neil Houlsby, Dustin Tran, and Mario Lucic. Revisiting the calibration of modern neural networks. Advances in Neural Information Processing Systems, 34:15682-15694, 2021. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 220, + 287, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 220, + 287, + 284 + ], + "spans": [ + { + "bbox": [ + 49, + 220, + 287, + 284 + ], + "type": "text", + "content": "[29] Miguel Monteiro, Loic Le Folgoc, Daniel Coelho de Castro, Nick Pawlowski, Bernardo Marques, Konstantinos Kamnitsas, Mark van der Wilk, and Ben Glocker. Stochastic segmentation networks: Modelling spatially correlated aleatoric uncertainty. Advances in Neural Information Processing Systems, 33:12756-12767, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 286, + 287, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 286, + 287, + 330 + ], + "spans": [ + { + "bbox": [ + 49, + 286, + 287, + 330 + ], + "type": "text", + "content": "[30] Jishnu Mukhoti, Viveka Kulharia, Amartya Sanyal, Stuart Golodetz, Philip Torr, and Puneet Dokania. Calibrating deep neural networks using focal loss. Advances in Neural Information Processing Systems, 33:15288-15299, 2020. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 332, + 287, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 332, + 287, + 364 + ], + "spans": [ + { + "bbox": [ + 49, + 332, + 287, + 364 + ], + "type": "text", + "content": "[31] Rafael Müller, Simon Kornblith, and Geoffrey E Hinton. When does label smoothing help? Advances in neural information processing systems, 32, 2019. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 366, + 287, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 366, + 287, + 408 + ], + "spans": [ + { + "bbox": [ + 49, + 366, + 287, + 408 + ], + "type": "text", + "content": "[32] Muhammad Akhtar Munir, Muhammad Haris Khan, M Saqib Sarfraz, and Mohsen Ali. Towards improving calibration in object detection under domain shift. arXiv preprint arXiv:2209.07601, 2022. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 411, + 287, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 411, + 287, + 454 + ], + "spans": [ + { + "bbox": [ + 49, + 411, + 287, + 454 + ], + "type": "text", + "content": "[33] Mahdi Pakdaman Naeini, Gregory Cooper, and Milos Hauskrecht. Obtaining well calibrated probabilities using bayesian binning. In Twenty-Ninth AAAI Conference on Artificial Intelligence, 2015. 3, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 456, + 287, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 456, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 49, + 456, + 287, + 521 + ], + "type": "text", + "content": "[34] Yaniv Ovadia, Emily Fertig, Jie Ren, Zachary Nado, David Sculley, Sebastian Nowozin, Joshua Dillon, Balaji Lakshminarayanan, and Jasper Snoek. Can you trust your model's uncertainty? evaluating predictive uncertainty under dataset shift. Advances in neural information processing systems, 32, 2019. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 523, + 287, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 523, + 287, + 576 + ], + "spans": [ + { + "bbox": [ + 49, + 523, + 287, + 576 + ], + "type": "text", + "content": "[35] Sangdon Park, Osbert Bastani, James Weimer, and Insup Lee. Calibrated prediction with covariate shift via unsupervised domain adaptation. In International Conference on Artificial Intelligence and Statistics, pages 3219-3229. PMLR, 2020. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 579, + 287, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 579, + 287, + 622 + ], + "spans": [ + { + "bbox": [ + 49, + 579, + 287, + 622 + ], + "type": "text", + "content": "[36] Kanil Patel, William Beluch, Bin Yang, Michael Pfeiffer, and Dan Zhang. Multi-class uncertainty calibration via mutual information maximization-based binning. arXiv preprint arXiv:2006.13092, 2020. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 624, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 624, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 49, + 624, + 287, + 689 + ], + "type": "text", + "content": "[37] Federico Perazzi, Jordi Pont-Tuset, Brian McWilliams, Luc Van Gool, Markus Gross, and Alexander Sorkine-Hornung. A benchmark dataset and evaluation methodology for video object segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 724-732, 2016. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 691, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 691, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 691, + 287, + 712 + ], + "type": "text", + "content": "[38] Jordi Pont-Tuset, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "davis challenge on video object segmentation. arXiv preprint arXiv:1704.00675, 2017. 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 96, + 545, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 160 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 160 + ], + "type": "text", + "content": "[39] German Ros, Laura Sellart, Joanna Materzynska, David Vazquez, and Antonio M Lopez. The synthia dataset: A large collection of synthetic images for semantic segmentation of urban scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3234-3243, 2016. 2, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 162, + 545, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 162, + 545, + 215 + ], + "spans": [ + { + "bbox": [ + 308, + 162, + 545, + 215 + ], + "type": "text", + "content": "[40] Robin Strudel, Ricardo Garcia, Ivan Laptev, and Cordelia Schmid. Segmenter: Transformer for semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7262-7272, 2021. 1, 2, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 217, + 545, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 217, + 545, + 270 + ], + "spans": [ + { + "bbox": [ + 308, + 217, + 545, + 270 + ], + "type": "text", + "content": "[41] Sunil Thulasidasan, Gopinath Chennupati, Jeff A Bilmes, Tanmoy Bhattacharya, and Sarah Michalak. On mixup training: Improved calibration and predictive uncertainty for deep neural networks. Advances in Neural Information Processing Systems, 32, 2019. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 272, + 545, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 272, + 545, + 326 + ], + "spans": [ + { + "bbox": [ + 308, + 272, + 545, + 326 + ], + "type": "text", + "content": "[42] Christian Tomani, Sebastian Gruber, Muhammed Ebrar Erdem, Daniel Cremers, and Florian Buettner. Post-hoc uncertainty calibration for domain drift scenarios. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10124-10132, 2021. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 327, + 545, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 327, + 545, + 358 + ], + "spans": [ + { + "bbox": [ + 308, + 327, + 545, + 358 + ], + "type": "text", + "content": "[43] Antonio Torralba and Alexei A Efros. Unbiased look at dataset bias. In CVPR 2011, pages 1521-1528. IEEE, 2011. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 361, + 545, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 361, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 308, + 361, + 545, + 415 + ], + "type": "text", + "content": "[44] Adam Van Etten, Daniel Hogan, Jesus Martinez Manso, Jacob Shermeyer, Nicholas Weir, and Ryan Lewis. The multitemporal urban development spacenet dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6398-6407, 2021. 1, 2, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 416, + 545, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 416, + 545, + 458 + ], + "spans": [ + { + "bbox": [ + 308, + 416, + 545, + 458 + ], + "type": "text", + "content": "[45] Ximei Wang, Mingsheng Long, Jianmin Wang, and Michael Jordan. Transferable calibration with lower bias and variance in domain adaptation. Advances in Neural Information Processing Systems, 33:19212-19223, 2020. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 460, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 460, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 308, + 460, + 545, + 514 + ], + "type": "text", + "content": "[46] Yezhen Wang, Bo Li, Tong Che, Kaiyang Zhou, Ziwei Liu, and Dongsheng Li. Energy-based open-world uncertainty modeling for confidence calibration. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9302-9311, 2021. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 515, + 545, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 515, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 308, + 515, + 545, + 557 + ], + "type": "text", + "content": "[47] Andrew G Wilson and Pavel Izmailov. Bayesian deep learning and a probabilistic perspective of generalization. Advances in neural information processing systems, 33:4697-4708, 2020. 1, 2, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 559, + 545, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 559, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 308, + 559, + 545, + 602 + ], + "type": "text", + "content": "[48] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In Proceedings of the European conference on computer vision (ECCV), pages 418-434, 2018. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 604, + 545, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 604, + 545, + 647 + ], + "spans": [ + { + "bbox": [ + 308, + 604, + 545, + 647 + ], + "type": "text", + "content": "[49] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. arXiv preprint arXiv:2105.15203, 2021. 1, 2, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 647, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 712 + ], + "type": "text", + "content": "[50] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darryll. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2636-2645, 2020. 1, 2, 6" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23661" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 262 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 48, + 72, + 286, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 286, + 105 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 286, + 105 + ], + "type": "text", + "content": "[51] Wenwei Zhang, Jiangmiao Pang, Kai Chen, and Chen Change Loy. K-Net: Towards unified image segmentation. In NeurIPS, 2021. 1, 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 107, + 286, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 107, + 286, + 150 + ], + "spans": [ + { + "bbox": [ + 49, + 107, + 286, + 150 + ], + "type": "text", + "content": "[52] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2881-2890, 2017. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 152, + 286, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 152, + 286, + 206 + ], + "spans": [ + { + "bbox": [ + 49, + 152, + 286, + 206 + ], + "type": "text", + "content": "[53] Sixiao Zheng, Jiachen Lu, Hengshuang Zhao, Xiatian Zhu, Zekun Luo, Yabiao Wang, Yanwei Fu, Jianfeng Feng, Tao Xiang, Philip HS Torr, et al. Rethinking semantic segmentation from a sequence-to-sequence perspective with transformers. arXiv preprint arXiv:2012.15840, 2020. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 209, + 286, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 209, + 286, + 262 + ], + "spans": [ + { + "bbox": [ + 49, + 209, + 286, + 262 + ], + "type": "text", + "content": "[54] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. 1, 2, 3, 6" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23662" + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/On Data Scaling in Masked Image Modeling/23a2cca2-7ab3-48ca-b93b-e899d7bc87f3_content_list.json b/2023/On Data Scaling in Masked Image Modeling/23a2cca2-7ab3-48ca-b93b-e899d7bc87f3_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..40fc9ab0ad43aed43eb00507e41e27f94007bc76 --- /dev/null +++ b/2023/On Data Scaling in Masked Image Modeling/23a2cca2-7ab3-48ca-b93b-e899d7bc87f3_content_list.json @@ -0,0 +1,1328 @@ +[ + { + "type": "text", + "text": "On Data Scaling in Masked Image Modeling", + "text_level": 1, + "bbox": [ + 259, + 130, + 709, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhenda Xie $^{13}$ , Zheng Zhang $^{3\\dagger}$ , Yue Cao $^{3\\dagger}$ , Yutong Lin $^{23}$ , Yixuan Wei $^{13}$ , Qi Dai $^{3}$ , Han Hu $^{3}$ $^{1}$ Tsinghua University $^{2}$ Xi'an Jiaotong University $^{3}$ Microsoft Research Asia", + "bbox": [ + 129, + 179, + 836, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{t-zhxie, zhez, yuecao, t-yutonglin, t-yixuanwei, qi.dai, hanhu}@microsoft.com", + "bbox": [ + 138, + 219, + 828, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 268, + 313, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Scaling properties have been one of the central issues in self-supervised pre-training, especially the data scalability, which has successfully motivated the large-scale self-supervised pre-trained language models and endowed them with significant modeling capabilities. However, scaling properties seem to be unintentionally neglected in the recent trending studies on masked image modeling (MIM), and some arguments even suggest that MIM cannot benefit from large-scale data. In this work, we try to break down these preconceptions and systematically study the scaling behaviors of MIM through extensive experiments, with data ranging from $10\\%$ of ImageNet-1K to full ImageNet22K, model parameters ranging from 49-million to one-billion, and training length ranging from 125K to 500K iterations. And our main findings can be summarized in two folds: 1) masked image modeling remains demanding large-scale data in order to scale up computes and model parameters; 2) masked image modeling cannot benefit from more data under a non-overfitting scenario, which diverges from the previous observations in self-supervised pre-trained language models or supervised pre-trained vision models. In addition, we reveal several intriguing properties in MIM, such as high sample efficiency in large MIM models and strong correlation between pre-training validation loss and transfer performance. We hope that our findings could deepen the understanding of masked image modeling and facilitate future developments on large-scale vision models. Code and models will be available at https://github.com/microsoft/SimMIM.", + "bbox": [ + 75, + 300, + 472, + 738 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 766, + 207, + 781 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Masked Image Modeling (MIM) [3, 18, 44], which has recently emerged in the field of self-supervised visual pretraining, has attracted widespread interest and extensive applications throughout the community for unleashing the superior modeling capacity of attention-based Transformer", + "bbox": [ + 75, + 790, + 470, + 867 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "architectures [13, 26] and demonstrating excellent sample efficiency and impressive transfer performance on a variety of vision tasks. However, most recent practices are focused on the design of MIM methods, while the study of scaling properties of MIM is unintentionally neglected, especially the data scaling property, which successfully motivated the large-scale self-supervised pre-trained language models and endowed them with significant modeling capabilities. Although previous works [8, 10, 17, 37, 45] have explored several conclusions about the scaling properties of vision models, most of their findings were obtained under a supervised pre-training scheme or under a contrastive learning framework, so the extent to which these findings could be transferred to MIM still needs to be investigated.", + "bbox": [ + 500, + 268, + 893, + 481 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Meanwhile, with the emergence of Transformers [41] and masked language modeling (MLM) [12, 31], the systematic studies of scaling laws have already been explored in natural language processing field [21, 23, 35], which provided ample guidance for large models in recent years. The core finding drawn from scaling laws [21] for neural language models is that the performance has a power-law relationship with each of the three scale factors - model parameters $N$ , size of dataset $D$ , and amount of compute $C$ respectively - when not bottlenecked by the other two. This conclusion implies that better performance can be obtained by scaling up these three factors to the extent that the scaling laws are in effect, which led to the subsequent developments of large scale language models [16, 28, 29, 32, 33] that exhibit excellent modeling capabilities [4] on most language tasks. Therefore, it is natural to ask whether MIM possesses the same scaling signatures for vision models as the MLM method for language models, so that the scaling up of vision models can catch up with language models.", + "bbox": [ + 498, + 486, + 893, + 773 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Though masked image modeling and masked language modeling both belong to masked signal prediction, their property differences are also non-negligible due to the different nature of vision and language. That is, the images are highly redundant raw signals and words/sentences are semantically rich tokens, which may result in different abilities of data utilization, and it is thus debatable whether the observations from language models could be reproduced in vision models.", + "bbox": [ + 498, + 779, + 893, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "The work is done when Zhenda Xie, Yutong Lin, and Yixuan Wei are interns at Microsoft Research Asia. $\\dagger$ Project co-leaders.", + "bbox": [ + 75, + 875, + 470, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "10365", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/816aa38777e260b912428594c678e4cfda475d1e0447938cfaa211f93f9383f5.jpg", + "image_caption": [ + "Model Flops $\\times$ Training Iterations" + ], + "image_footnote": [], + "bbox": [ + 160, + 89, + 385, + 224 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/35bc9fb8857d9d4776c448afc9cffa445f8e3cc5f41903bd9d940b0b7dd520a1.jpg", + "image_caption": [ + "Dataset Size [D]", + "Millions of Images", + "Figure 1. The curves of validation loss of pre-training models w.r.t. the relative compute, dataset size and model size. MIM performance improves smoothly as we increase the relative compute and model size, but not improve when the dataset size is sufficient to prevent model from overfitting. We set the relative compute of SwinV2-S for 125K iterations as the value of 1. Best viewed in color." + ], + "image_footnote": [], + "bbox": [ + 387, + 90, + 596, + 215 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/540236580d56040df83fa97a68c9efde933d6d82e9c1d81cd63c544cfecd44c6.jpg", + "image_caption": [ + "Model Size [N]", + "Millions of Params" + ], + "image_footnote": [], + "bbox": [ + 601, + 92, + 807, + 215 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Moreover, recent studies [14, 38] have shown that using a small amount of training data in masked image modeling can achieve comparable performance to using large datasets, which also motivates our explorations as it disagrees with previous findings and intuitions.", + "bbox": [ + 75, + 311, + 470, + 387 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we systematically investigate the scaling properties, especially the data scaling capability of masked image modeling in terms of different dataset sizes ranging from $10\\%$ of ImageNet ( $\\sim 0.1$ million) to full ImageNet-22K ( $\\sim 14$ million), model sizes ranging from 49-million Swin-V2 Small to 1-billion Swin-V2 giant and training lengths ranging from 125K to 500K iterations. We use Swin Transformer V2 [25] as the vision encoder for its proven trainability of large models and applicability to a wide range of vision tasks, and adopt SimMIM [44] for masked image modeling pre-training because it has no restrictions on encoder architectures. We also conduct experiments with other MIM frameworks like MAE [18] and other vision encoder like the widely used ViT [13] to verify the generalizability of our findings. With these experimental setups, our main findings could be summarized into two folds:", + "bbox": [ + 75, + 392, + 470, + 633 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "i) Masked image modeling remains demanding large-scale data in order to scale up computes and model parameters. We empirically find that MIM performance has a power-law relationship with relative compute and model size when not bottlenecked by the dataset size (Figure 1). Besides, we observe that smaller datasets lead to severe overfitting phenomenon for training large models (Figure 2), and the size of the dataset to prevent model from overfitting increases clearly as the model increases (Figure 5-Left). Therefore, from the perspective of scaling up model, MIM still demands large-scale data. Furthermore, if we train large-scale models of different lengths, we find that relatively small datasets are adequate at shorter training lengths, but still suffer from overfitting at longer training lengths (Figure 5-Right), which further demonstrate the data scalability of MIM for scaling up compute.", + "bbox": [ + 75, + 638, + 470, + 881 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2) Masked image modeling cannot benefit from more data", + "bbox": [ + 96, + 885, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "under a non-overfitting scenario, which diverges from the previous observations in self-supervised pre-train language models or supervised pre-trained vision models. In MIM, we find that increasing the number of unique samples for a non-overfitting model does not provide additional benefits to performance (Figure 1). This behavior differs from previous observations in supervised vision transformers and self-supervised language models, where the model performance increases as the number of unique samples increases.", + "bbox": [ + 496, + 311, + 893, + 448 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In addition, we also demonstrate some intriguing properties about masked image modeling, such as larger models possessing higher sample efficiency, i.e., fewer optimization steps are required for larger models to achieve same performance (Section 4.2); and the consistency between transfer and test performance, i.e., test performance could be used to indicate the results on downstream tasks (Section 4.3). These observations also correspond to those in previous practices.", + "bbox": [ + 496, + 449, + 893, + 569 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "These findings on the one hand confirm the effect of MIM on scaling up model size and compute, and raise new concerns and challenges on the data scalability of MIM on the other. We hope that our findings could deepen the understanding of masked image modeling and facilitate future developments on large-scale vision models.", + "bbox": [ + 496, + 570, + 893, + 661 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 676, + 640, + 694 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Masked Image Modeling Masked Image Modeling learns representations by reconstructing the masked content of images, and its early exploration can be traced back to context encoder [30] and denoising autoencoder [42]. Recently, iGPT [6], BEiT [3], MAE [18] and SimMIM [44] recall this approach on training vision transformer. iGPT [6] sequentially predicted the pixels by auto-regressive manner. BEiT [3] proposed to predict the discrete visual tokens. MAE [18] and SimMIM [44] concurrently found predicting the raw pixels with a high masking ratio could work well. In this work, we use SimMIM as the default masked image modeling approach, because of its simplicity and no restrictions on the architecture of vision encoder like MAE.", + "bbox": [ + 496, + 703, + 893, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "10366", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2dcd36573cf2bed72f60350d216b6eab723dc27a8a33f9f5c846bf8ae71164d3.jpg", + "image_caption": [ + "Figure 2. Training loss of pre-training, validation loss of pre-training, and fine-tuning accuracy on ImageNet-1K of different model sizes, data sizes and training lengths, w.r.t. the relative pre-training compute. All models are SwinV2 models pre-trained with SimMIM. We set the training compute of SwinV2-S for 125K iterations as the value of 1. Bigger circles indicate larger models. Best viewed in color." + ], + "image_footnote": [], + "bbox": [ + 160, + 89, + 812, + 244 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Vision Transformers Transformer [41] was first applied to natural language processing and became the dominant architecture, and has recently attracted a lot of attention in computer vision. The pioneering work ViT [13] first shows that the transformer architecture works well in image classification when trained on large amounts of data. DeiT [39] proposed a better training recipe based on ViT and demonstrated that vision Transformer has promising performance when only using ImageNet-1K dataset. Swin Transformer [26] improves plain ViT by inducing the hierarchical architecture and non-overlapping local attention and successfully demonstrates the effectiveness of vision transformer on a wide range of vision tasks. Swin Transformer V2 [25] further addresses the training stability issue of [26] in model scaling and illustrates better performance than the original Swin Transformer, and thus we use it as the default vision encoder in this work.", + "bbox": [ + 75, + 310, + 472, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Scaling Vision Models Many works [25, 34, 37, 45] examine how to scale vision models, but most are more concerned with exploring the perspective of model architecture designs. For example, EfficientNet [37] extensively studied how model width, model depth and input resolution affect the convolutional neural networks; [34] proposed to scale vision model with sparse mixture-of-expert; [45] and [25] studied how to scale ViT and Swin Transformer, respectively. In addition, several studies [1, 2, 45] explored the aspect of data scaling under the pre-training fine-tuning paradigm. BiT [22] revisited the supervised pre-training on a wide range of data scales up to 1M images. SEER [17] studied the effectiveness of data scaling in the contrastive learning framework with up to one billion images. Recently, SplitMask [14] find that masked image modeling is robust to the size of pre-training data and challenges the data scaling capability of masked image modeling, which is most relevant to our work.", + "bbox": [ + 75, + 574, + 470, + 845 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Scaling Language Models Many works [20, 23, 35] systematically analyzed the scaling behavior on a relatively small scale. And [21] studied empirical scaling laws for", + "bbox": [ + 75, + 854, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "large scale neural language models and revealed that the loss scaled as a power-law with model size, dataset size, and the amount of compute used for training. Motivated by these laws, several large scale language models [4,16,28,29,32,33] were successfully trained and demonstrated excellent modeling capabilities on most language tasks. [36] established a comprehensive, large-scale benchmark designed to assist in quantifying and extrapolating the capabilities of large language models.", + "bbox": [ + 496, + 310, + 893, + 446 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Background and Experimental Setup", + "text_level": 1, + "bbox": [ + 498, + 458, + 836, + 476 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Masked Image Modeling", + "text_level": 1, + "bbox": [ + 500, + 483, + 725, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Masked image modeling is used to train the vision model by taking a corrupted image as input and predicting the content of the masked region as the target. In this study, we use SimMIM [44] as the default masked image modeling approach because of its simplicity and lack of restrictions on the architecture of the vision encoder. SimMIM consists of a visual encoder and an extremely lightweight prediction head of a linear layer for predicting the raw pixels of the corrupted images via $\\ell_2$ regression loss. To facilitate the implementation of the vision transformer, SimMIM adopts the patch-wise mask strategy with the masked patch size of $32 \\times 32$ and mask ratio of 0.6. To further alleviate the local dependency of raw pixels, we improved the SimMIM by normalizing the predicted target according to [15] with a sliding window of $47^2$ . As the result, a slight performance improvement is observed. In addition, we conduct experiments using MAE as the masked image modeling approach to verify the methodological generalizability of our findings. These experiments strictly follow the settings in [18].", + "bbox": [ + 496, + 506, + 893, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Architecture Specifications", + "text_level": 1, + "bbox": [ + 500, + 801, + 743, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We use Swin Transformer V2 [25] as the default vision encoder in this study. Thanks to its generality and scalability, we evaluate a series of SwinV2 models with a wide range of model sizes (the number of parameters ranges from $\\sim 50\\mathrm{M}$ to $\\sim 1\\mathrm{B}$ , and FLOPs range from $\\sim 9\\mathrm{G}$ to $\\sim 190\\mathrm{G}$ ) on", + "bbox": [ + 496, + 825, + 893, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "10367", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/dc85bc388912c22b2e844d6226f28208d54fecd13aebbde9e0c5ee10b85fc974.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelBase ChannelDepthHeadWindow SizeBackbone Params
pre-trainfine-tune
SwinV2-S96{2, 2, 18, 2}{3, 6, 12, 24}121449M
SwinV2-B128{2, 2, 18, 2}{4, 8, 16, 32}121487M
SwinV2-L192{2, 2, 18, 2}{6, 12, 24, 48}1214195M
SwinV2-H352{2, 2, 18, 2}{11, 22, 44, 88}1214655M
SwinV2-g448{2, 2, 18, 2}{14, 28, 56, 112}12141061M
", + "bbox": [ + 187, + 89, + 782, + 193 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/461e373fb847bfc9b643970d9e537a20d19fd9743d1da46fd051150a4ff5cfd1.jpg", + "table_caption": [ + "Table 1. Detailed architecture specifications. Note that, the model aliases used in our work are different from [25]." + ], + "table_footnote": [], + "table_body": "
IN1K (10%)IN1K (20%)IN1K (50%)IN100IN1K(100%)IN22K(100%)
#Classes\\( 1 \\times 10^{3} \\)\\( 1 \\times 10^{3} \\)\\( 1 \\times 10^{3} \\)\\( 1 \\times 10^{2} \\)\\( 1 \\times 10^{3} \\)\\( 2.18 \\times 10^{4} \\)
#Images\\( 1.28 \\times 10^{5} \\)\\( 2.56 \\times 10^{5} \\)\\( 6.41 \\times 10^{5} \\)\\( 1.27 \\times 10^{5} \\)\\( 1.28 \\times 10^{6} \\)\\( 1.42 \\times 10^{7} \\)
", + "bbox": [ + 160, + 233, + 810, + 282 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 2. Detailed dataset specifications used in the pre-training of masked image modeling.", + "bbox": [ + 212, + 292, + 753, + 306 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "multiple downstream tasks. The detailed model specifications are shown in Table 1. We use a new variant SwinV2-g (giant), with number of parameters between SwinV2-L and the 3-billion-parameter SwinV2-G (Giant) used in [25]. In addition, we conduct experiments with a series of ViT models [13] to prove the architectural generalizability of our findings. Specifically, we use ViT-B/16, ViT-L/16 and ViTH/14 according to the settings from [18].", + "bbox": [ + 75, + 321, + 470, + 444 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Pre-training Datasets", + "text_level": 1, + "bbox": [ + 76, + 452, + 277, + 468 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To study the effect of data size on masked image modeling, we build datasets with different sizes. We use the training set of ImageNet-1K and ImageNet-22K as two large-scale datasets, and randomly sample $10\\%$ , $20\\%$ , $50\\%$ of images in the ImageNet-1K training set as smaller datasets. By default, the images are uniformly sampled from each category. We also consider the sampling strategies could perform differently. To this end, we randomly sample 100 classes from ImageNet-1K as ImageNet-100, and compare it with ImageNet-1K $(10\\%)$ but find their training loss and fine-tuning performance are almost the same. The details and statistics of all pre-training datasets used in our study are shown in Table 2.", + "bbox": [ + 75, + 476, + 470, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Pre-training Details", + "text_level": 1, + "bbox": [ + 76, + 681, + 266, + 696 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To better compare the performance of models with different amounts of data under the same pre-training length, we use training iterations rather than training epochs and adopt the same hyper-parameters for all models with different sizes during pre-training. The total number of training iterations is in $\\{125\\mathrm{K}, 250\\mathrm{K}, 500\\mathrm{K}\\}$ and the batch size is set as 2048 for all experiments. In pre-training stage, we use the same hyper-parameters for all models, and the training details and hyper-parameters of pre-training are summarized in Appendix. Because of the excessive amount of experiments, we follow SimMIM [44] and also use the following two techniques for reducing the experimental overheads: First, we use the step learning rate scheduler in pre-training for", + "bbox": [ + 75, + 704, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "sharing the first training step among experiments with different training lengths. The first 7/8 training iterations are the first step and the last 1/8 training iterations are the second step with the learning rate ratio of 0.1 (i.e. learning rate is divided by 10 in the second step). Second, we adopt the input image size of $192^2$ and set the window size of 12. We improve the SimMIM by normalizing the predicted target according to [15] with a sliding window of $47^2$ and observe an improvement of 0.3 on top-1 accuracy of ImageNet-1K for the SwinV2-Large model. The same light data augmentation strategy as SimMIM is used: random resize cropping with a scale range of [0.67, 1], an aspect ratio range of [3/4, 4/3] and a random flipping with probability 0.5.", + "bbox": [ + 496, + 321, + 893, + 520 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.5. Fine-tuning Tasks", + "text_level": 1, + "bbox": [ + 500, + 531, + 671, + 547 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To extensively and accurately evaluate the performance of pre-trained models under different pre-training schedulers and datasets, a series of diverse and representative tasks including fine-tuning on ImageNet-1K, fine-grained image classification, object detection, instance segmentation, and semantic segmentation are selected for evaluation. Detailed setups of all experiments are illustrated in Appendix.", + "bbox": [ + 496, + 556, + 893, + 662 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ImageNet-1K We follow [3] to evaluate the quality of learnt representations by fine-tuning the pre-trained models on ImageNet-1K [11] image classification task, which is the most commonly used scenario and evaluation criterion for pre-trained models [18,44]. Different from pre-training, We adopt the image size with $224^2$ with window size of 14 in fine-tuning. The AdamW with batch size of 2048, base learning rate of 5e-3, weight decay of 0.05, $\\beta_{1}$ of 0.9 and $\\beta_{2}$ of 0.999 are used, and we adopt cosine learning rate scheduler. As larger models are more prone to overfitting, we fine-tune SwinV2-S/B/L for 100 epochs with 20 warm-up epochs and SwinV2-H/g for 50 epochs with 10 warm-up epochs, and decrease the layer decay as the model size increases. In addition, gradient clipping, stochastic depth, label smoothing and data augmentations (e.g. random crop,", + "bbox": [ + 496, + 674, + 893, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "10368", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f7af21374200d9e8f70bb2f5e77f76ea22f09720188570767effe034a4bedea8.jpg", + "image_caption": [ + "Figure 3. Left to right: Performances with SwinV2 models on (a) COCO object detection, (b) COCO instance segmentation, (c) iNaturalist-18, and (d) ADE20K semantic segmentation w.r.t. the relative compute. Note that the relative compute indicates the pre-training compute. We set the relative compute of SwinV2-S for 125K iterations as 1. Bigger circles indicate larger models. Best viewed in color." + ], + "image_footnote": [], + "bbox": [ + 158, + 90, + 812, + 233 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "rand erasing [47], rand augment [9], mixup [46], cutmix [46], etc.) are also used by following [44].", + "bbox": [ + 75, + 301, + 470, + 333 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "iNaturalist-18 iNaturalist [40] 2018 is a long-tailed fine-grained image classification dataset. As fine-tuning in ImageNet-1K, we also use the input image size of $224^2$ , window size of 14 and patch size of 4 in iNaturalist 2018. We fine-tune all models for 100 epochs with 20 warm-up epochs, and set layer decay to 0.8, 0.75 and 0.7 for SwinV2-S/B/L, respectively. The AdamW optimizer with cosine learning rate scheduler, batch size of 2048, base learning rate of 1.6e-2, weight decay of 0.1, $\\beta_{1}$ of 0.9 and $\\beta_{2}$ of 0.999 are used. In addition, we also adopt stochastic depth, label smoothing, gradient clipping and data augmentations in fine-tuning.", + "bbox": [ + 75, + 337, + 473, + 518 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "COCO Object Detection and Instance Segmentation [24]", + "text_level": 1, + "bbox": [ + 75, + 523, + 468, + 539 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use Mask R-CNN [19] for evaluation. We set the window size to 14 and patch size to 4. The AdamW optimizer with batch size of 32, base learning rate of 8e-5, weight decay of 0.05, $\\beta_{1}$ of 0.9, $\\beta_{2}$ of 0.999 and a step learning rate scheduler (step learning rate ratio of 0.1, step epochs are 27 and 33) are used. In training, the random cropping with crop size of [1024, 1024], large scale jittering with a range of [0.1, 2.0], random horizontal flip with probability 0.5, and stochastic depth regularization are used. In testing, all images are resized to (800, 1333) and keeping the aspect ratio unchanged.", + "bbox": [ + 75, + 539, + 470, + 705 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ADE20K Semantic Segmentation [48] Following [26], we use UPerNet [43] for evaluation. We set the window size to 20 and the patch size to 4. The AdamW optimizer with batch size of 32, base learning rate searched in a range of [1e-4, 3e-4], weight decay of 0.05, $\\beta_{1}$ of 0.9, $\\beta_{2}$ of 0.999 and a linear learning rate scheduler with a total of 80K iterations are used. Also, we use the layer decay of 0.95, 0.95, 0.9 for SwinV2-S/B/L, respectively. In training, the random cropping with crop size of [640, 640], scale jittering with a range of [0.5, 2.0], random horizontal flip with probability 0.5, random photometric distortion and stochastic", + "bbox": [ + 75, + 710, + 470, + 878 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "depth regularization of 0.1 are used. In testing, all images are evaluated by sliding window manner, and use the test image size of (2560, 640) and set sliding window stride to 426, following [26, 44].", + "bbox": [ + 500, + 301, + 893, + 363 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Results and Findings", + "text_level": 1, + "bbox": [ + 500, + 377, + 700, + 396 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We train numerous models with different training lengths, dataset sizes, and model sizes, and study how these factors affect the performance of masked image modeling. Figure 1 illustrates the validation loss of pre-training $^2$ with respect to the relative compute, dataset size and model size. Figure 2 illustrates the training loss of pre-training, validation loss of pre-training, and the fine-tuning top-1 accuracy of ImageNet-1K with respect to the relative compute. Based on these extensive experiments, we make the following observations:", + "bbox": [ + 498, + 404, + 893, + 541 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Data Scaling in Masked Image Modeling", + "text_level": 1, + "bbox": [ + 498, + 551, + 849, + 569 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Masked image modeling remains demanding for large-scale data. When with the high masking rate (e.g., $60\\%$ in our work), the masked image modeling is considered a very challenging training objective and has been found to be data efficient by previous literature [14,27], i.e., a comparable performance can be achieved with small datasets as with large datasets. However, Figure 2 shows that as the training cost increases, the training loss of some models drops significantly, and their validation loss rises significantly, even on using $50\\%$ images of ImageNet-1K (i.e., IN1K $(50\\%)$ ), indicating the overfitting phenomenon exists. In Figure 5-Left, we demonstrate that the size of the dataset to prevent model from overfitting increases clearly as the model increases. And significant decrease to the fine-tuning performance caused by overfitting could be observed in Figure 2 and Figure 3. Moreover, we measure the best fine-tuning performance of each model trained by different training schedulers in Figure 2-Right. We find the large models perform even worse than smaller models when small dataset is used for training. For", + "bbox": [ + 498, + 575, + 893, + 864 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "2The validation loss of pre-training is measured on the validation set of ImageNet-1K for all experiments.", + "bbox": [ + 500, + 875, + 893, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "1Our implementation based on MMDetection [5].", + "bbox": [ + 94, + 886, + 359, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "10369", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3fafda11fc16c0eb1c8d21d46c0bc528caf84840c45ad6ee96a365b19f2151ce.jpg", + "image_caption": [ + "Figure 4. Training loss of pre-training, validation loss of pre-training, and fine-tuning accuracy on ImageNet-1K of different model sizes, data sizes and training lengths, w.r.t. the relative pre-training compute. All models are ViT models pre-trained with MAE. We set the training compute of ViT-B/16 for 125K iterations as the value of 1. Bigger circles indicate larger models. Best viewed in color." + ], + "image_footnote": [], + "bbox": [ + 160, + 89, + 816, + 244 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "example, the best top-1 accuracy of SwinV2-H with IN1K (20%) is 84.4, worse than the best performance of SwinV2-L by 0.3. In addition, by comparing the best performance that can be obtained using different sizes of dataset, we find that using more data results in better performance. These observations suggest that masked image modeling does not alleviate the demands of large dataset.", + "bbox": [ + 75, + 310, + 473, + 416 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The training length matters. Larger models can benefit from more data at a longer training length. By comparing the performance of models pre-trained by different data sizes in Figure 2-Right, we find that the fine-tuning performance of the large models saturates more slowly with the increasing data size compared to the smaller models. For example, the SwinV2-S model pre-trained on IN1K (50%) has a very similar fine-tuning performance to the model pre-trained on IN1K (100%). In comparison, the performance difference between the SwinV2-H model pre-trained on IN1K (50%) and IN1K (100%) is near 0.5, which is a significant gap for ImageNet-1K classification.", + "bbox": [ + 75, + 424, + 470, + 604 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Furthermore, a comprehensive observation reveals that the improvements from using more data are not significant under short training lengths. As shown in Figure 5-Right, while there is a noticeable performance gap between SwinV2-H trained on IN1K (50%) and IN1K (100%) at a training length of 500K iterations, the gap is negligible at a training length of 125K iterations. This observation suggests that increasing the training length for larger models is critical to benefit from more data.", + "bbox": [ + 75, + 604, + 470, + 741 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Masked image modeling cannot benefit from more data under a non-overfitting scenario. As illustrated in Figure 1-Center, we plot the validation loss of pre-trained SwinV2 models with respect to different dataset sizes, and we demonstrate that for a particular model size and training length, there will be a certain dataset size that will keep it from overfitting. And the model cannot benefit from dataset larger than this size. This behavior differs from previous observations in supervised vision transformers [45] and self-supervised language models [21], where the model perfor", + "bbox": [ + 75, + 750, + 472, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "mance increases as the number of unique samples increases. We speculate that this may be related to the characteristics of the task itself, i.e., MIM provides a large number of training signals, making the model hard to learn more patterns from large-scale data.", + "bbox": [ + 498, + 310, + 893, + 385 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6c5ea36ba70b6898f3f5a6ab888582dda2f4ec0e3dc889ae444ff1d3b64f534e.jpg", + "image_caption": [ + "Figure 5. Left: Validation loss with respect to model size. The size of the dataset to prevent model from overfitting increases clearly as the model increases. Right: Validation loss with respect to training length using SwinV2-Huge. $50\\%$ subset of ImageNet-1K is sufficient for training 125K iterations, but not sufficient to prevent overfitting for 500K iterations. Best viewed in color." + ], + "image_footnote": [], + "bbox": [ + 501, + 397, + 885, + 536 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation on more tasks. In addition to ImageNet-1K image classification, we also evaluate the MIM pre-trained SwinV2-S, SwinV2-B and SwinV2-L on iNaturalist-18 fine-grained image classification, ADE20K semantic segmentation, and COCO object detection/segmentation. Figure 3 shows a similar pattern with ImageNet-1K (Figure. 2-Right) that as the training cost increases, some models have significantly performance drop. In addition, the smaller models rapidly reach saturation as the amount of data increases, while larger models can continuously benefit from more data after sufficient training. These results suggest that the conclusions drawn on ImageNet-1K are broadly applicable to other vision tasks.", + "bbox": [ + 496, + 654, + 895, + 849 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results with MAE and ViT Figure 4 demonstrates the results of ViT pre-trained with MAE. Similar to SwinV2 models pre-trained with SimMIM, we observe the same", + "bbox": [ + 498, + 854, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "10370", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "overfitting phenomenon when training with small datasets or large models, which makes MAE still demand for large-scale data. Besides, we could find that larger ViT models can also benefit from more data at a longer training lengths. These experiments verify the methodological and architectural generalizability of our results and findings.", + "bbox": [ + 75, + 90, + 470, + 183 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1825b07385a5a207ca6589331aedd1a04c57ba8cb6a4b2c3efb4b6722efaf239.jpg", + "image_caption": [ + "Figure 6. Validation loss and fine-tuning error rate on ImageNet-1K with respect to images processed during pre-training. Large models are more sample efficient than small models to achieve lower validation loss or fine-tuning error rate with fewer optimization steps. Best viewed in color." + ], + "image_footnote": [], + "bbox": [ + 81, + 196, + 468, + 335 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Larger Models Possess Higher Sample Efficiency", + "text_level": 1, + "bbox": [ + 76, + 431, + 468, + 463 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 6 shows the validation loss and fine-tuning error rate on ImageNet-1K with respect to the images processed (batch size times number of steps) during pre-training stage. Five Swin-V2 models with different sizes are pre-trained on full-set of ImageNet-1K. From these results, we observe that larger models possess higher sample efficiency, reaching the same level of validation loss or fine-tuning error rate on ImageNet with fewer optimization steps. These results indicate that larger models will continue to have better performance, and training larger models with fewer steps on sufficient data is a preferable choice. This observation is also in line with previous works on self-supervised language models [21] and supervised vision transformers [45].", + "bbox": [ + 75, + 470, + 470, + 667 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Correlation between Pre-training Losses and the Fine-tuning Performance", + "text_level": 1, + "bbox": [ + 76, + 680, + 468, + 712 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluating a pre-trained model by its fine-tuned performance on downstream tasks is costly. In supervised pretraining, the validation accuracy is used as the proxy indicator to evaluate the quality of the pre-trained models. While in previous studies [7] on other self-supervised learning approaches (e.g., contrastive learning), such a proxy indicator is lacking. In this study, we would like to explore whether the pre-training loss in the training of masked image modeling is a good indicator of its fine-tuning performance. We collect all pre-trained models and plot their training and validation loss curves on Figure 7. Interestingly, the correlations between pre-training losses and the fine-tuning performance", + "bbox": [ + 75, + 719, + 472, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/fc2e1d8063aaaae345f00c5ea8b26f04251ecebb8dbbe55c771a69d0e2db1185.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TaskOverfitNon-Overfit
Train LossVal LossTrain LossVal Loss
IN-1K+0.26-0.79-0.64-0.90
iNat-18+0.17-0.54-0.46-0.78
COCO OD+0.54-0.81-0.35-0.83
COCO IS+0.62-0.86-0.31-0.85
ADE-20K+0.75-0.91-0.14-0.90
", + "bbox": [ + 501, + 89, + 898, + 195 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "on multiple tasks could be observed with a phase transition around overfitting.", + "bbox": [ + 496, + 268, + 890, + 299 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Specifically, the correlation between training loss and fine-tuning performance is negative for the overfitting model (green circles) and positive for the non-overfitting model (red circles). The correlation between validation loss and fine-tuning performance is always negative, but the slope of their linear fit lines is significantly different.", + "bbox": [ + 496, + 301, + 890, + 393 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In addition, we further analyze the Pearson correlation coefficient between training loss and fine-tuning performance (Table 3), and find the validation loss has stronger correlation with fine-tuned performance than train loss for all cases, especially for non-overfitting models.", + "bbox": [ + 496, + 396, + 893, + 472 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Effects of Different Sizes of Decoders", + "text_level": 1, + "bbox": [ + 500, + 489, + 818, + 503 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We have studied the effects of encoder size from the data scaling perspective. Here, the effects of decoder size are further studied. We pre-train SwinV2-B models with decoder heads of different sizes on IN1K (20%), and Table 4 shows the results. Interestingly, although we find that the heavier decoder has lower training loss and higher validation loss than the linear decoder, indicating a more severe overfitting issue. But there is no decrease in its fine-tuning performance on ImageNet-1K than the linear decoder. This experiment shows that the decoder behaves very differently from the encoder, and we speculate that this is because the decoder \"blocks\" the damage to the encoder from overfitting.", + "bbox": [ + 496, + 515, + 890, + 696 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/dad3ed26861f664935c449d7669aca7ff3dff2682ef23b6e1db5eec57e906c41.jpg", + "table_caption": [ + "Table 3. Pearson correlation coefficients between pre-training losses (training and validation losses) and fine-tuning performances on five downstream tasks." + ], + "table_footnote": [], + "table_body": "
Decoder# ParamsTrain LossVal LossTop-1 Acc
linear90.0M0.460.4784.4
4-blocks140.4M0.440.4884.4
8-blocks190.8M0.410.5084.5
", + "bbox": [ + 503, + 714, + 888, + 781 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Results of different decoders, including converged training and validation losses of MIM pre-training, and fine-tuning performance (top-1 accuracy) on ImageNet-1K image classification. Encoders for all models are SwinV2-Base.", + "bbox": [ + 496, + 791, + 893, + 847 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "3The least squares method is used for linear fit.", + "bbox": [ + 514, + 886, + 767, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "10371", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dd935eddff1fb13ba7c9ab22ea78df540c1495b04b8c3d4259107d66a4d3fb8b.jpg", + "image_caption": [ + "Figure 7. The correlations between pre-training losses (training and validation losses) and the fine-tuning performances of (a) ImageNet-1k image classification, (b) iNat 2018 fine-grained classification, (c) COCO object detection, (d) COCO instance segmentation, (e) and ADE-20K semantic segmentation. Pre-training losses are highly correlated with fine-tuning performance on all five tasks. Red circles indicate the overfitting models and green circles indicate non-overfitting models. Best viewed in color." + ], + "image_footnote": [], + "bbox": [ + 160, + 88, + 816, + 290 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/330a994c98f4e3a44c7cd3cd2c3a8a43ce9088a52f4e8eebbf70682d30a138bf.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Dataset# ClassesTrain / Val LossTop-1 Acc
IN1K (10%)10000.351 / 0.51583.5
IN1001000.352 / 0.51183.4
", + "bbox": [ + 88, + 367, + 459, + 419 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Results on different dataset sampling strategies (ImageNet-1K (10%) and ImageNet-100) with same dataset size $\\left( {{1.28} \\times {10}^{5}}\\right.$ images for both), include converged training and validation losses of MIM pre-training, and fine-tuning performance (top-1 accuracy) on ImageNet-1K image classification.", + "bbox": [ + 75, + 430, + 470, + 501 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Impact of Different Dataset Sampling Strategies", + "text_level": 1, + "bbox": [ + 76, + 516, + 468, + 547 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We study different dataset sampling strategies by comparing the training behavior and fine-tuned performance of models pre-trained on IN1K (10%) and IN100. In IN1K (10%), the images are uniformly sampled from each category, and we randomly sample 100 categories from ImageNet-1K as IN100. Experiments are conducted on SwinV2-L with 500K training iterations. Table 5 shows the training loss, validation loss and fine-tuning top-1 accuracy of ImageNet-1K. For the two models pre-trained on IN1K (10%) and IN100, all three metrics are very similar. Figure 8 further illustrates the training dynamics of the two models, and we find both their training loss curves and validation loss curves are almost overlapping. These results show the disparity caused by different dataset sampling strategies is minor.", + "bbox": [ + 75, + 556, + 472, + 768 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 782, + 196, + 799 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In our work, we systematically study the data scaling capability of masked image modeling at different model sizes and training lengths. Based on the extensive experiments, we demonstrate that in order to scale up computes and model parameters, masked image modeling remains demanding for large-scale data due to the severe overfit", + "bbox": [ + 75, + 810, + 470, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/85a9bd2e8ee20d32e2de541ecf81af2ac5f1cac0bd2bdbf5f34ffa3b0f7338f4.jpg", + "image_caption": [ + "Figure 8. The training loss and validation loss of MIM pre-training with different dataset sampling strategies, ImageNet-1K (10%) and ImageNet-100. Best viewed in color." + ], + "image_footnote": [], + "bbox": [ + 501, + 368, + 893, + 492 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ting phenomenon when pre-training with small datasets of large models, which challenges the conclusions of previous literatures that a large dataset may not be necessary in masked image modeling. Besides, we also find that masked image modeling cannot benefit from more data under a nonoverfitting scenario, which diverges from the previous observations in large-scale pre-training language or vision models and raise new concerns and challenges on the data scalability of MIM. In addition, some intriguing properties of MIM are observed, such as larger models possessing higher sample efficiency and a strong correlation between the validation loss of masked image modeling and the fine-tuning performance. The former observation indicates that larger models will continue to have better performance and suggests that training larger models with fewer steps on sufficient data is a preferable option, while the latter observation suggests that validation loss can be considered as a good proxy indicator for evaluating pre-trained models, and makes it possible to reduce the experimental overhead of measuring models by fine-tuning. We hope that our findings could deepen the understanding of masked image modeling and facilitate future developments on large-scale vision models.", + "bbox": [ + 496, + 568, + 893, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "10372", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Ibrahim Alabdulmohsin, Behnam Neyshabur, and Xiaohua Zhai. Revisiting neural scaling laws in language and vision. In NeurIPS, 2022.", + "[2] Ibrahim M Alabdulmohsin and Mario Lucic. A near-optimal algorithm for debiasing trained machine learning models. Advances in Neural Information Processing Systems, 34:8072-8084, 2021.", + "[3] Hangbo Bao, Li Dong, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021.", + "[4] Tom B Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020.", + "[5] Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jiarui Xu, et al. Mmdetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155, 2019.", + "[6] Mark Chen, Alec Radford, Rewon Child, Jeff Wu, and Heewoo Jun. Generative pretraining from pixels. Advances in Neural Information Processing Systems, 2020.", + "[7] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. ICML, 2020.", + "[8] Ting Chen, Simon Kornblith, Kevin Swersky, Mohammad Norouzi, and Geoffrey Hinton. Big self-supervised models are strong semi-supervised learners, 2020.", + "[9] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 702-703, 2020.", + "[10] Zihang Dai, Hanxiao Liu, Quoc V. Le, and Mingxing Tan. Coatnet: Marrying convolution and attention for all data sizes, 2021.", + "[11] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248–255. IEEE, 2009.", + "[12] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018.", + "[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021.", + "[14] Alaaeldin El-Nouby, Gautier Izacard, Hugo Touvron, Ivan Laptev, Hervé Jegou, and Edouard Grave. Are large-scale datasets necessary for self-supervised pre-training? arXiv preprint arXiv:2112.10740, 2021." + ], + "bbox": [ + 78, + 116, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Yuxin Fang, Li Dong, Hangbo Bao, Xinggang Wang, and Furu Wei. Corrupted image modeling for self-supervised visual pre-training. arXiv preprint arXiv:2202.03382, 2022.", + "[16] William Fedus, Barret Zoph, and Noam Shazeer. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity, 2021.", + "[17] Priya Goyal, Mathilde Caron, Benjamin Lefaudeauux, Min Xu, Pengchao Wang, Vivek Pai, Mannat Singh, Vitaliy Liptchinsky, Ishan Misra, Armand Joulin, et al. Self-supervised pretraining of visual features in the wild. arXiv preprint arXiv:2103.01988, 2021.", + "[18] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dolkar, and Ross Girshick. Masked autoencoders are scalable vision learners. arXiv preprint arXiv:2111.06377, 2021.", + "[19] Kaiming He, Georgia Gkioxari, Piotr Dólár, and Ross Girshick. Mask r-cnn. In ICCV, pages 2961-2969, 2017.", + "[20] Joel Hestness, Sharan Narang, Newsha Ardalani, Gregory Diamos, Heewoo Jun, Hassan Kianinejad, Md Patwary, Mostofa Ali, Yang Yang, and Yanqi Zhou. Deep learning scaling is predictable, empirically. arXiv preprint arXiv:1712.00409, 2017.", + "[21] Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B. Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models, 2020.", + "[22] Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Joan Puigcerver, Jessica Yung, Sylvain Gelly, and Neil Houlsby. Big transfer (bit): General visual representation learning. In European conference on computer vision, pages 491-507. Springer, 2020.", + "[23] Zhuohan Li, Eric Wallace, Sheng Shen, Kevin Lin, Kurt Keutzer, Dan Klein, and Joey Gonzalez. Train big, then compress: Rethinking model size for efficient training and inference of transformers. In International Conference on Machine Learning, pages 5958-5968. PMLR, 2020.", + "[24] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014.", + "[25] Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, et al. Swin transformer v2: Scaling up capacity and resolution. arXiv preprint arXiv:2111.09883, 2021.", + "[26] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. arXiv preprint arXiv:2103.14030, 2021.", + "[27] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swin transformer, 2021.", + "[28] Microsoft. Turing-nlg: A 17-billion-parameter language model by microsoft, 2020.", + "[29] Microsoft. Using deepspeed and megatron to train megatron-turing nlg 530b, the world's largest and most powerful generative language model, 2021.", + "[30] Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A Efros. Context encoders: Feature learning by inpainting. In Proceedings of the IEEE conference on" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "10373", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "computer vision and pattern recognition, pages 2536-2544, 2016.", + "[31] Alec Radford, Karthik Narasimhan, Tim Salimans, and Ilya Sutskever. Improving language understanding by generative pre-training. 2018.", + "[32] Alec Radford, Jeff Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language models are unsupervised multitask learners. 2019.", + "[33] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research, 21(140):1-67, 2020.", + "[34] Carlos Riquelme, Joan Puigcerver, Basil Mustafa, Maxim Neumann, Rodolphe Jenatton, André Susano Pinto, Daniel Keysers, and Neil Houlsby. Scaling vision with sparse mixture of experts. arXiv preprint arXiv:2106.05974, 2021.", + "[35] Jonathan S Rosenfeld, Amir Rosenfeld, Yonatan Belinkov, and Nir Shavit. A constructive prediction of the generalization error across scales. arXiv preprint arXiv:1909.12673, 2019.", + "[36] Aarohi Srivastava, Abhinav Rastogi, Abhishek Rao, Abu Awal Md Shoeb, Abubakar Abid, Adam Fisch, Adam R Brown, Adam Santoro, Aditya Gupta, Adrià Garriga-Alonso, et al. Beyond the imitation game: Quantifying and extrapolating the capabilities of language models. arXiv preprint arXiv:2206.04615, 2022.", + "[37] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International conference on machine learning, pages 6105-6114. PMLR, 2019.", + "[38] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. arXiv preprint arXiv:2203.12602, 2022.", + "[39] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International Conference on Machine Learning, pages 10347-10357. PMLR, 2021.", + "[40] Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alex Shepard, Hartwig Adam, Pietro Perona, and Serge Belongie. The inaturalist species classification and detection dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8769-8778, 2018.", + "[41] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017.", + "[42] Pascal Vincent, Hugo Larochelle, Yoshua Bengio, and Pierre-Antoine Manzagol. Extracting and composing robust features with denoising autoencoders. In Proceedings of the 25th international conference on Machine learning, pages 1096-1103, 2008.", + "[43] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In Proceedings of the European Conference on Computer Vision (ECCV), pages 418-434, 2018." + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[44] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. arXiv preprint arXiv:2111.09886, 2021.", + "[45] Xiaohua Zhai, Alexander Kolesnikov, Neil Houlsby, and Lucas Beyer. Scaling vision transformers, 2021.", + "[46] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412, 2017.", + "[47] Zhun Zhong, Liang Zheng, Guoliang Kang, Shaozi Li, and Yi Yang. Random erasing data augmentation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 13001-13008, 2020.", + "[48] Bolei Zhou, Hang Zhao, Xavier Puig, Tete Xiao, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Semantic understanding of scenes through the ade20k dataset. International Journal on Computer Vision, 2018." + ], + "bbox": [ + 501, + 92, + 893, + 332 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10374", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/On Data Scaling in Masked Image Modeling/23a2cca2-7ab3-48ca-b93b-e899d7bc87f3_model.json b/2023/On Data Scaling in Masked Image Modeling/23a2cca2-7ab3-48ca-b93b-e899d7bc87f3_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f374a3116cf4af0ab245453628c74246752ffb86 --- /dev/null +++ b/2023/On Data Scaling in Masked Image Modeling/23a2cca2-7ab3-48ca-b93b-e899d7bc87f3_model.json @@ -0,0 +1,1914 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.26, + 0.131, + 0.71, + 0.154 + ], + "angle": 0, + "content": "On Data Scaling in Masked Image Modeling" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.18, + 0.838, + 0.217 + ], + "angle": 0, + "content": "Zhenda Xie\\(^{13}\\), Zheng Zhang\\(^{3\\dagger}\\), Yue Cao\\(^{3\\dagger}\\), Yutong Lin\\(^{23}\\), Yixuan Wei\\(^{13}\\), Qi Dai\\(^{3}\\), Han Hu\\(^{3}\\) \n\\(^{1}\\)Tsinghua University \\(^{2}\\)Xi'an Jiaotong University \\(^{3}\\)Microsoft Research Asia" + }, + { + "type": "text", + "bbox": [ + 0.139, + 0.22, + 0.829, + 0.234 + ], + "angle": 0, + "content": "{t-zhxie, zhez, yuecao, t-yutonglin, t-yixuanwei, qi.dai, hanhu}@microsoft.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.269, + 0.314, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.301, + 0.473, + 0.739 + ], + "angle": 0, + "content": "Scaling properties have been one of the central issues in self-supervised pre-training, especially the data scalability, which has successfully motivated the large-scale self-supervised pre-trained language models and endowed them with significant modeling capabilities. However, scaling properties seem to be unintentionally neglected in the recent trending studies on masked image modeling (MIM), and some arguments even suggest that MIM cannot benefit from large-scale data. In this work, we try to break down these preconceptions and systematically study the scaling behaviors of MIM through extensive experiments, with data ranging from \\(10\\%\\) of ImageNet-1K to full ImageNet22K, model parameters ranging from 49-million to one-billion, and training length ranging from 125K to 500K iterations. And our main findings can be summarized in two folds: 1) masked image modeling remains demanding large-scale data in order to scale up computes and model parameters; 2) masked image modeling cannot benefit from more data under a non-overfitting scenario, which diverges from the previous observations in self-supervised pre-trained language models or supervised pre-trained vision models. In addition, we reveal several intriguing properties in MIM, such as high sample efficiency in large MIM models and strong correlation between pre-training validation loss and transfer performance. We hope that our findings could deepen the understanding of masked image modeling and facilitate future developments on large-scale vision models. Code and models will be available at https://github.com/microsoft/SimMIM." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.767, + 0.208, + 0.782 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.791, + 0.471, + 0.868 + ], + "angle": 0, + "content": "Masked Image Modeling (MIM) [3, 18, 44], which has recently emerged in the field of self-supervised visual pretraining, has attracted widespread interest and extensive applications throughout the community for unleashing the superior modeling capacity of attention-based Transformer" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.27, + 0.895, + 0.482 + ], + "angle": 0, + "content": "architectures [13, 26] and demonstrating excellent sample efficiency and impressive transfer performance on a variety of vision tasks. However, most recent practices are focused on the design of MIM methods, while the study of scaling properties of MIM is unintentionally neglected, especially the data scaling property, which successfully motivated the large-scale self-supervised pre-trained language models and endowed them with significant modeling capabilities. Although previous works [8, 10, 17, 37, 45] have explored several conclusions about the scaling properties of vision models, most of their findings were obtained under a supervised pre-training scheme or under a contrastive learning framework, so the extent to which these findings could be transferred to MIM still needs to be investigated." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.487, + 0.895, + 0.775 + ], + "angle": 0, + "content": "Meanwhile, with the emergence of Transformers [41] and masked language modeling (MLM) [12, 31], the systematic studies of scaling laws have already been explored in natural language processing field [21, 23, 35], which provided ample guidance for large models in recent years. The core finding drawn from scaling laws [21] for neural language models is that the performance has a power-law relationship with each of the three scale factors - model parameters \\( N \\), size of dataset \\( D \\), and amount of compute \\( C \\) respectively - when not bottlenecked by the other two. This conclusion implies that better performance can be obtained by scaling up these three factors to the extent that the scaling laws are in effect, which led to the subsequent developments of large scale language models [16, 28, 29, 32, 33] that exhibit excellent modeling capabilities [4] on most language tasks. Therefore, it is natural to ask whether MIM possesses the same scaling signatures for vision models as the MLM method for language models, so that the scaling up of vision models can catch up with language models." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.78, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Though masked image modeling and masked language modeling both belong to masked signal prediction, their property differences are also non-negligible due to the different nature of vision and language. That is, the images are highly redundant raw signals and words/sentences are semantically rich tokens, which may result in different abilities of data utilization, and it is thus debatable whether the observations from language models could be reproduced in vision models." + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.875, + 0.471, + 0.901 + ], + "angle": 0, + "content": "The work is done when Zhenda Xie, Yutong Lin, and Yixuan Wei are interns at Microsoft Research Asia. \\(\\dagger\\) Project co-leaders." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10365" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.161, + 0.09, + 0.386, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.192, + 0.227, + 0.385, + 0.239 + ], + "angle": 0, + "content": "Model Flops \\(\\times\\) Training Iterations" + }, + { + "type": "image", + "bbox": [ + 0.388, + 0.091, + 0.597, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.448, + 0.217, + 0.554, + 0.228 + ], + "angle": 0, + "content": "Dataset Size [D]" + }, + { + "type": "image_caption", + "bbox": [ + 0.448, + 0.228, + 0.554, + 0.239 + ], + "angle": 0, + "content": "Millions of Images" + }, + { + "type": "image", + "bbox": [ + 0.602, + 0.093, + 0.808, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.665, + 0.217, + 0.756, + 0.228 + ], + "angle": 0, + "content": "Model Size [N]" + }, + { + "type": "image_caption", + "bbox": [ + 0.665, + 0.228, + 0.771, + 0.238 + ], + "angle": 0, + "content": "Millions of Params" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.256, + 0.895, + 0.3 + ], + "angle": 0, + "content": "Figure 1. The curves of validation loss of pre-training models w.r.t. the relative compute, dataset size and model size. MIM performance improves smoothly as we increase the relative compute and model size, but not improve when the dataset size is sufficient to prevent model from overfitting. We set the relative compute of SwinV2-S for 125K iterations as the value of 1. Best viewed in color." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.312, + 0.471, + 0.388 + ], + "angle": 0, + "content": "Moreover, recent studies [14, 38] have shown that using a small amount of training data in masked image modeling can achieve comparable performance to using large datasets, which also motivates our explorations as it disagrees with previous findings and intuitions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.393, + 0.471, + 0.634 + ], + "angle": 0, + "content": "In this paper, we systematically investigate the scaling properties, especially the data scaling capability of masked image modeling in terms of different dataset sizes ranging from \\(10\\%\\) of ImageNet (\\(\\sim 0.1\\) million) to full ImageNet-22K (\\(\\sim 14\\) million), model sizes ranging from 49-million Swin-V2 Small to 1-billion Swin-V2 giant and training lengths ranging from 125K to 500K iterations. We use Swin Transformer V2 [25] as the vision encoder for its proven trainability of large models and applicability to a wide range of vision tasks, and adopt SimMIM [44] for masked image modeling pre-training because it has no restrictions on encoder architectures. We also conduct experiments with other MIM frameworks like MAE [18] and other vision encoder like the widely used ViT [13] to verify the generalizability of our findings. With these experimental setups, our main findings could be summarized into two folds:" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.64, + 0.472, + 0.882 + ], + "angle": 0, + "content": "i) Masked image modeling remains demanding large-scale data in order to scale up computes and model parameters. We empirically find that MIM performance has a power-law relationship with relative compute and model size when not bottlenecked by the dataset size (Figure 1). Besides, we observe that smaller datasets lead to severe overfitting phenomenon for training large models (Figure 2), and the size of the dataset to prevent model from overfitting increases clearly as the model increases (Figure 5-Left). Therefore, from the perspective of scaling up model, MIM still demands large-scale data. Furthermore, if we train large-scale models of different lengths, we find that relatively small datasets are adequate at shorter training lengths, but still suffer from overfitting at longer training lengths (Figure 5-Right), which further demonstrate the data scalability of MIM for scaling up compute." + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.886, + 0.471, + 0.902 + ], + "angle": 0, + "content": "2) Masked image modeling cannot benefit from more data" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.312, + 0.895, + 0.449 + ], + "angle": 0, + "content": "under a non-overfitting scenario, which diverges from the previous observations in self-supervised pre-train language models or supervised pre-trained vision models. In MIM, we find that increasing the number of unique samples for a non-overfitting model does not provide additional benefits to performance (Figure 1). This behavior differs from previous observations in supervised vision transformers and self-supervised language models, where the model performance increases as the number of unique samples increases." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.45, + 0.895, + 0.57 + ], + "angle": 0, + "content": "In addition, we also demonstrate some intriguing properties about masked image modeling, such as larger models possessing higher sample efficiency, i.e., fewer optimization steps are required for larger models to achieve same performance (Section 4.2); and the consistency between transfer and test performance, i.e., test performance could be used to indicate the results on downstream tasks (Section 4.3). These observations also correspond to those in previous practices." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.571, + 0.895, + 0.662 + ], + "angle": 0, + "content": "These findings on the one hand confirm the effect of MIM on scaling up model size and compute, and raise new concerns and challenges on the data scalability of MIM on the other. We hope that our findings could deepen the understanding of masked image modeling and facilitate future developments on large-scale vision models." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.678, + 0.642, + 0.695 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.704, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Masked Image Modeling Masked Image Modeling learns representations by reconstructing the masked content of images, and its early exploration can be traced back to context encoder [30] and denoising autoencoder [42]. Recently, iGPT [6], BEiT [3], MAE [18] and SimMIM [44] recall this approach on training vision transformer. iGPT [6] sequentially predicted the pixels by auto-regressive manner. BEiT [3] proposed to predict the discrete visual tokens. MAE [18] and SimMIM [44] concurrently found predicting the raw pixels with a high masking ratio could work well. In this work, we use SimMIM as the default masked image modeling approach, because of its simplicity and no restrictions on the architecture of vision encoder like MAE." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10366" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.161, + 0.09, + 0.813, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.255, + 0.895, + 0.298 + ], + "angle": 0, + "content": "Figure 2. Training loss of pre-training, validation loss of pre-training, and fine-tuning accuracy on ImageNet-1K of different model sizes, data sizes and training lengths, w.r.t. the relative pre-training compute. All models are SwinV2 models pre-trained with SimMIM. We set the training compute of SwinV2-S for 125K iterations as the value of 1. Bigger circles indicate larger models. Best viewed in color." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.311, + 0.473, + 0.568 + ], + "angle": 0, + "content": "Vision Transformers Transformer [41] was first applied to natural language processing and became the dominant architecture, and has recently attracted a lot of attention in computer vision. The pioneering work ViT [13] first shows that the transformer architecture works well in image classification when trained on large amounts of data. DeiT [39] proposed a better training recipe based on ViT and demonstrated that vision Transformer has promising performance when only using ImageNet-1K dataset. Swin Transformer [26] improves plain ViT by inducing the hierarchical architecture and non-overlapping local attention and successfully demonstrates the effectiveness of vision transformer on a wide range of vision tasks. Swin Transformer V2 [25] further addresses the training stability issue of [26] in model scaling and illustrates better performance than the original Swin Transformer, and thus we use it as the default vision encoder in this work." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.575, + 0.472, + 0.846 + ], + "angle": 0, + "content": "Scaling Vision Models Many works [25, 34, 37, 45] examine how to scale vision models, but most are more concerned with exploring the perspective of model architecture designs. For example, EfficientNet [37] extensively studied how model width, model depth and input resolution affect the convolutional neural networks; [34] proposed to scale vision model with sparse mixture-of-expert; [45] and [25] studied how to scale ViT and Swin Transformer, respectively. In addition, several studies [1, 2, 45] explored the aspect of data scaling under the pre-training fine-tuning paradigm. BiT [22] revisited the supervised pre-training on a wide range of data scales up to 1M images. SEER [17] studied the effectiveness of data scaling in the contrastive learning framework with up to one billion images. Recently, SplitMask [14] find that masked image modeling is robust to the size of pre-training data and challenges the data scaling capability of masked image modeling, which is most relevant to our work." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Scaling Language Models Many works [20, 23, 35] systematically analyzed the scaling behavior on a relatively small scale. And [21] studied empirical scaling laws for" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.311, + 0.895, + 0.448 + ], + "angle": 0, + "content": "large scale neural language models and revealed that the loss scaled as a power-law with model size, dataset size, and the amount of compute used for training. Motivated by these laws, several large scale language models [4,16,28,29,32,33] were successfully trained and demonstrated excellent modeling capabilities on most language tasks. [36] established a comprehensive, large-scale benchmark designed to assist in quantifying and extrapolating the capabilities of large language models." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.459, + 0.837, + 0.477 + ], + "angle": 0, + "content": "3. Background and Experimental Setup" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.484, + 0.727, + 0.501 + ], + "angle": 0, + "content": "3.1. Masked Image Modeling" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.507, + 0.895, + 0.794 + ], + "angle": 0, + "content": "Masked image modeling is used to train the vision model by taking a corrupted image as input and predicting the content of the masked region as the target. In this study, we use SimMIM [44] as the default masked image modeling approach because of its simplicity and lack of restrictions on the architecture of the vision encoder. SimMIM consists of a visual encoder and an extremely lightweight prediction head of a linear layer for predicting the raw pixels of the corrupted images via \\(\\ell_2\\) regression loss. To facilitate the implementation of the vision transformer, SimMIM adopts the patch-wise mask strategy with the masked patch size of \\(32 \\times 32\\) and mask ratio of 0.6. To further alleviate the local dependency of raw pixels, we improved the SimMIM by normalizing the predicted target according to [15] with a sliding window of \\(47^2\\). As the result, a slight performance improvement is observed. In addition, we conduct experiments using MAE as the masked image modeling approach to verify the methodological generalizability of our findings. These experiments strictly follow the settings in [18]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.802, + 0.744, + 0.818 + ], + "angle": 0, + "content": "3.2. Architecture Specifications" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.895, + 0.901 + ], + "angle": 0, + "content": "We use Swin Transformer V2 [25] as the default vision encoder in this study. Thanks to its generality and scalability, we evaluate a series of SwinV2 models with a wide range of model sizes (the number of parameters ranges from \\(\\sim 50\\mathrm{M}\\) to \\(\\sim 1\\mathrm{B}\\), and FLOPs range from \\(\\sim 9\\mathrm{G}\\) to \\(\\sim 190\\mathrm{G}\\)) on" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10367" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.189, + 0.09, + 0.783, + 0.194 + ], + "angle": 0, + "content": "
ModelBase ChannelDepthHeadWindow SizeBackbone Params
pre-trainfine-tune
SwinV2-S96{2, 2, 18, 2}{3, 6, 12, 24}121449M
SwinV2-B128{2, 2, 18, 2}{4, 8, 16, 32}121487M
SwinV2-L192{2, 2, 18, 2}{6, 12, 24, 48}1214195M
SwinV2-H352{2, 2, 18, 2}{11, 22, 44, 88}1214655M
SwinV2-g448{2, 2, 18, 2}{14, 28, 56, 112}12141061M
" + }, + { + "type": "table_caption", + "bbox": [ + 0.147, + 0.205, + 0.821, + 0.22 + ], + "angle": 0, + "content": "Table 1. Detailed architecture specifications. Note that, the model aliases used in our work are different from [25]." + }, + { + "type": "table", + "bbox": [ + 0.161, + 0.234, + 0.812, + 0.283 + ], + "angle": 0, + "content": "
IN1K (10%)IN1K (20%)IN1K (50%)IN100IN1K(100%)IN22K(100%)
#Classes\\( 1 \\times 10^{3} \\)\\( 1 \\times 10^{3} \\)\\( 1 \\times 10^{3} \\)\\( 1 \\times 10^{2} \\)\\( 1 \\times 10^{3} \\)\\( 2.18 \\times 10^{4} \\)
#Images\\( 1.28 \\times 10^{5} \\)\\( 2.56 \\times 10^{5} \\)\\( 6.41 \\times 10^{5} \\)\\( 1.27 \\times 10^{5} \\)\\( 1.28 \\times 10^{6} \\)\\( 1.42 \\times 10^{7} \\)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.213, + 0.294, + 0.754, + 0.308 + ], + "angle": 0, + "content": "Table 2. Detailed dataset specifications used in the pre-training of masked image modeling." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.323, + 0.471, + 0.445 + ], + "angle": 0, + "content": "multiple downstream tasks. The detailed model specifications are shown in Table 1. We use a new variant SwinV2-g (giant), with number of parameters between SwinV2-L and the 3-billion-parameter SwinV2-G (Giant) used in [25]. In addition, we conduct experiments with a series of ViT models [13] to prove the architectural generalizability of our findings. Specifically, we use ViT-B/16, ViT-L/16 and ViTH/14 according to the settings from [18]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.453, + 0.279, + 0.469 + ], + "angle": 0, + "content": "3.3. Pre-training Datasets" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.477, + 0.471, + 0.673 + ], + "angle": 0, + "content": "To study the effect of data size on masked image modeling, we build datasets with different sizes. We use the training set of ImageNet-1K and ImageNet-22K as two large-scale datasets, and randomly sample \\(10\\%\\), \\(20\\%\\), \\(50\\%\\) of images in the ImageNet-1K training set as smaller datasets. By default, the images are uniformly sampled from each category. We also consider the sampling strategies could perform differently. To this end, we randomly sample 100 classes from ImageNet-1K as ImageNet-100, and compare it with ImageNet-1K \\((10\\%)\\) but find their training loss and fine-tuning performance are almost the same. The details and statistics of all pre-training datasets used in our study are shown in Table 2." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.682, + 0.267, + 0.698 + ], + "angle": 0, + "content": "3.4. Pre-training Details" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.472, + 0.903 + ], + "angle": 0, + "content": "To better compare the performance of models with different amounts of data under the same pre-training length, we use training iterations rather than training epochs and adopt the same hyper-parameters for all models with different sizes during pre-training. The total number of training iterations is in \\(\\{125\\mathrm{K}, 250\\mathrm{K}, 500\\mathrm{K}\\}\\) and the batch size is set as 2048 for all experiments. In pre-training stage, we use the same hyper-parameters for all models, and the training details and hyper-parameters of pre-training are summarized in Appendix. Because of the excessive amount of experiments, we follow SimMIM [44] and also use the following two techniques for reducing the experimental overheads: First, we use the step learning rate scheduler in pre-training for" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.323, + 0.895, + 0.521 + ], + "angle": 0, + "content": "sharing the first training step among experiments with different training lengths. The first 7/8 training iterations are the first step and the last 1/8 training iterations are the second step with the learning rate ratio of 0.1 (i.e. learning rate is divided by 10 in the second step). Second, we adopt the input image size of \\(192^2\\) and set the window size of 12. We improve the SimMIM by normalizing the predicted target according to [15] with a sliding window of \\(47^2\\) and observe an improvement of 0.3 on top-1 accuracy of ImageNet-1K for the SwinV2-Large model. The same light data augmentation strategy as SimMIM is used: random resize cropping with a scale range of [0.67, 1], an aspect ratio range of [3/4, 4/3] and a random flipping with probability 0.5." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.532, + 0.673, + 0.548 + ], + "angle": 0, + "content": "3.5. Fine-tuning Tasks" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.557, + 0.894, + 0.664 + ], + "angle": 0, + "content": "To extensively and accurately evaluate the performance of pre-trained models under different pre-training schedulers and datasets, a series of diverse and representative tasks including fine-tuning on ImageNet-1K, fine-grained image classification, object detection, instance segmentation, and semantic segmentation are selected for evaluation. Detailed setups of all experiments are illustrated in Appendix." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.895, + 0.902 + ], + "angle": 0, + "content": "ImageNet-1K We follow [3] to evaluate the quality of learnt representations by fine-tuning the pre-trained models on ImageNet-1K [11] image classification task, which is the most commonly used scenario and evaluation criterion for pre-trained models [18,44]. Different from pre-training, We adopt the image size with \\(224^2\\) with window size of 14 in fine-tuning. The AdamW with batch size of 2048, base learning rate of 5e-3, weight decay of 0.05, \\(\\beta_{1}\\) of 0.9 and \\(\\beta_{2}\\) of 0.999 are used, and we adopt cosine learning rate scheduler. As larger models are more prone to overfitting, we fine-tune SwinV2-S/B/L for 100 epochs with 20 warm-up epochs and SwinV2-H/g for 50 epochs with 10 warm-up epochs, and decrease the layer decay as the model size increases. In addition, gradient clipping, stochastic depth, label smoothing and data augmentations (e.g. random crop," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10368" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.16, + 0.091, + 0.813, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.246, + 0.895, + 0.29 + ], + "angle": 0, + "content": "Figure 3. Left to right: Performances with SwinV2 models on (a) COCO object detection, (b) COCO instance segmentation, (c) iNaturalist-18, and (d) ADE20K semantic segmentation w.r.t. the relative compute. Note that the relative compute indicates the pre-training compute. We set the relative compute of SwinV2-S for 125K iterations as 1. Bigger circles indicate larger models. Best viewed in color." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.302, + 0.471, + 0.334 + ], + "angle": 0, + "content": "rand erasing [47], rand augment [9], mixup [46], cutmix [46], etc.) are also used by following [44]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.338, + 0.474, + 0.52 + ], + "angle": 0, + "content": "iNaturalist-18 iNaturalist [40] 2018 is a long-tailed fine-grained image classification dataset. As fine-tuning in ImageNet-1K, we also use the input image size of \\(224^2\\), window size of 14 and patch size of 4 in iNaturalist 2018. We fine-tune all models for 100 epochs with 20 warm-up epochs, and set layer decay to 0.8, 0.75 and 0.7 for SwinV2-S/B/L, respectively. The AdamW optimizer with cosine learning rate scheduler, batch size of 2048, base learning rate of 1.6e-2, weight decay of 0.1, \\(\\beta_{1}\\) of 0.9 and \\(\\beta_{2}\\) of 0.999 are used. In addition, we also adopt stochastic depth, label smoothing, gradient clipping and data augmentations in fine-tuning." + }, + { + "type": "title", + "bbox": [ + 0.076, + 0.525, + 0.47, + 0.54 + ], + "angle": 0, + "content": "COCO Object Detection and Instance Segmentation [24]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.54, + 0.472, + 0.707 + ], + "angle": 0, + "content": "We use Mask R-CNN [19] for evaluation. We set the window size to 14 and patch size to 4. The AdamW optimizer with batch size of 32, base learning rate of 8e-5, weight decay of 0.05, \\(\\beta_{1}\\) of 0.9, \\(\\beta_{2}\\) of 0.999 and a step learning rate scheduler (step learning rate ratio of 0.1, step epochs are 27 and 33) are used. In training, the random cropping with crop size of [1024, 1024], large scale jittering with a range of [0.1, 2.0], random horizontal flip with probability 0.5, and stochastic depth regularization are used. In testing, all images are resized to (800, 1333) and keeping the aspect ratio unchanged." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.712, + 0.472, + 0.879 + ], + "angle": 0, + "content": "ADE20K Semantic Segmentation [48] Following [26], we use UPerNet [43] for evaluation. We set the window size to 20 and the patch size to 4. The AdamW optimizer with batch size of 32, base learning rate searched in a range of [1e-4, 3e-4], weight decay of 0.05, \\(\\beta_{1}\\) of 0.9, \\(\\beta_{2}\\) of 0.999 and a linear learning rate scheduler with a total of 80K iterations are used. Also, we use the layer decay of 0.95, 0.95, 0.9 for SwinV2-S/B/L, respectively. In training, the random cropping with crop size of [640, 640], scale jittering with a range of [0.5, 2.0], random horizontal flip with probability 0.5, random photometric distortion and stochastic" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.302, + 0.895, + 0.364 + ], + "angle": 0, + "content": "depth regularization of 0.1 are used. In testing, all images are evaluated by sliding window manner, and use the test image size of (2560, 640) and set sliding window stride to 426, following [26, 44]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.378, + 0.702, + 0.397 + ], + "angle": 0, + "content": "4. Results and Findings" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.405, + 0.895, + 0.542 + ], + "angle": 0, + "content": "We train numerous models with different training lengths, dataset sizes, and model sizes, and study how these factors affect the performance of masked image modeling. Figure 1 illustrates the validation loss of pre-training\\(^2\\) with respect to the relative compute, dataset size and model size. Figure 2 illustrates the training loss of pre-training, validation loss of pre-training, and the fine-tuning top-1 accuracy of ImageNet-1K with respect to the relative compute. Based on these extensive experiments, we make the following observations:" + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.552, + 0.85, + 0.57 + ], + "angle": 0, + "content": "4.1. Data Scaling in Masked Image Modeling" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.576, + 0.895, + 0.865 + ], + "angle": 0, + "content": "Masked image modeling remains demanding for large-scale data. When with the high masking rate (e.g., \\(60\\%\\) in our work), the masked image modeling is considered a very challenging training objective and has been found to be data efficient by previous literature [14,27], i.e., a comparable performance can be achieved with small datasets as with large datasets. However, Figure 2 shows that as the training cost increases, the training loss of some models drops significantly, and their validation loss rises significantly, even on using \\(50\\%\\) images of ImageNet-1K (i.e., IN1K \\((50\\%)\\)), indicating the overfitting phenomenon exists. In Figure 5-Left, we demonstrate that the size of the dataset to prevent model from overfitting increases clearly as the model increases. And significant decrease to the fine-tuning performance caused by overfitting could be observed in Figure 2 and Figure 3. Moreover, we measure the best fine-tuning performance of each model trained by different training schedulers in Figure 2-Right. We find the large models perform even worse than smaller models when small dataset is used for training. For" + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.875, + 0.895, + 0.901 + ], + "angle": 0, + "content": "2The validation loss of pre-training is measured on the validation set of ImageNet-1K for all experiments." + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.361, + 0.902 + ], + "angle": 0, + "content": "1Our implementation based on MMDetection [5]." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "10369" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.161, + 0.09, + 0.817, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.255, + 0.895, + 0.298 + ], + "angle": 0, + "content": "Figure 4. Training loss of pre-training, validation loss of pre-training, and fine-tuning accuracy on ImageNet-1K of different model sizes, data sizes and training lengths, w.r.t. the relative pre-training compute. All models are ViT models pre-trained with MAE. We set the training compute of ViT-B/16 for 125K iterations as the value of 1. Bigger circles indicate larger models. Best viewed in color." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.311, + 0.475, + 0.417 + ], + "angle": 0, + "content": "example, the best top-1 accuracy of SwinV2-H with IN1K (20%) is 84.4, worse than the best performance of SwinV2-L by 0.3. In addition, by comparing the best performance that can be obtained using different sizes of dataset, we find that using more data results in better performance. These observations suggest that masked image modeling does not alleviate the demands of large dataset." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.425, + 0.471, + 0.606 + ], + "angle": 0, + "content": "The training length matters. Larger models can benefit from more data at a longer training length. By comparing the performance of models pre-trained by different data sizes in Figure 2-Right, we find that the fine-tuning performance of the large models saturates more slowly with the increasing data size compared to the smaller models. For example, the SwinV2-S model pre-trained on IN1K (50%) has a very similar fine-tuning performance to the model pre-trained on IN1K (100%). In comparison, the performance difference between the SwinV2-H model pre-trained on IN1K (50%) and IN1K (100%) is near 0.5, which is a significant gap for ImageNet-1K classification." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.606, + 0.472, + 0.742 + ], + "angle": 0, + "content": "Furthermore, a comprehensive observation reveals that the improvements from using more data are not significant under short training lengths. As shown in Figure 5-Right, while there is a noticeable performance gap between SwinV2-H trained on IN1K (50%) and IN1K (100%) at a training length of 500K iterations, the gap is negligible at a training length of 125K iterations. This observation suggests that increasing the training length for larger models is critical to benefit from more data." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Masked image modeling cannot benefit from more data under a non-overfitting scenario. As illustrated in Figure 1-Center, we plot the validation loss of pre-trained SwinV2 models with respect to different dataset sizes, and we demonstrate that for a particular model size and training length, there will be a certain dataset size that will keep it from overfitting. And the model cannot benefit from dataset larger than this size. This behavior differs from previous observations in supervised vision transformers [45] and self-supervised language models [21], where the model perfor" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.311, + 0.895, + 0.386 + ], + "angle": 0, + "content": "mance increases as the number of unique samples increases. We speculate that this may be related to the characteristics of the task itself, i.e., MIM provides a large number of training signals, making the model hard to learn more patterns from large-scale data." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.398, + 0.887, + 0.537 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.551, + 0.895, + 0.634 + ], + "angle": 0, + "content": "Figure 5. Left: Validation loss with respect to model size. The size of the dataset to prevent model from overfitting increases clearly as the model increases. Right: Validation loss with respect to training length using SwinV2-Huge. \\(50\\%\\) subset of ImageNet-1K is sufficient for training 125K iterations, but not sufficient to prevent overfitting for 500K iterations. Best viewed in color." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.655, + 0.897, + 0.85 + ], + "angle": 0, + "content": "Evaluation on more tasks. In addition to ImageNet-1K image classification, we also evaluate the MIM pre-trained SwinV2-S, SwinV2-B and SwinV2-L on iNaturalist-18 fine-grained image classification, ADE20K semantic segmentation, and COCO object detection/segmentation. Figure 3 shows a similar pattern with ImageNet-1K (Figure. 2-Right) that as the training cost increases, some models have significantly performance drop. In addition, the smaller models rapidly reach saturation as the amount of data increases, while larger models can continuously benefit from more data after sufficient training. These results suggest that the conclusions drawn on ImageNet-1K are broadly applicable to other vision tasks." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Results with MAE and ViT Figure 4 demonstrates the results of ViT pre-trained with MAE. Similar to SwinV2 models pre-trained with SimMIM, we observe the same" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10370" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.472, + 0.184 + ], + "angle": 0, + "content": "overfitting phenomenon when training with small datasets or large models, which makes MAE still demand for large-scale data. Besides, we could find that larger ViT models can also benefit from more data at a longer training lengths. These experiments verify the methodological and architectural generalizability of our results and findings." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.198, + 0.47, + 0.336 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.346, + 0.473, + 0.416 + ], + "angle": 0, + "content": "Figure 6. Validation loss and fine-tuning error rate on ImageNet-1K with respect to images processed during pre-training. Large models are more sample efficient than small models to achieve lower validation loss or fine-tuning error rate with fewer optimization steps. Best viewed in color." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.433, + 0.47, + 0.464 + ], + "angle": 0, + "content": "4.2. Larger Models Possess Higher Sample Efficiency" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.472, + 0.472, + 0.669 + ], + "angle": 0, + "content": "Figure 6 shows the validation loss and fine-tuning error rate on ImageNet-1K with respect to the images processed (batch size times number of steps) during pre-training stage. Five Swin-V2 models with different sizes are pre-trained on full-set of ImageNet-1K. From these results, we observe that larger models possess higher sample efficiency, reaching the same level of validation loss or fine-tuning error rate on ImageNet with fewer optimization steps. These results indicate that larger models will continue to have better performance, and training larger models with fewer steps on sufficient data is a preferable choice. This observation is also in line with previous works on self-supervised language models [21] and supervised vision transformers [45]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.681, + 0.47, + 0.713 + ], + "angle": 0, + "content": "4.3. Correlation between Pre-training Losses and the Fine-tuning Performance" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Evaluating a pre-trained model by its fine-tuned performance on downstream tasks is costly. In supervised pretraining, the validation accuracy is used as the proxy indicator to evaluate the quality of the pre-trained models. While in previous studies [7] on other self-supervised learning approaches (e.g., contrastive learning), such a proxy indicator is lacking. In this study, we would like to explore whether the pre-training loss in the training of masked image modeling is a good indicator of its fine-tuning performance. We collect all pre-trained models and plot their training and validation loss curves on Figure 7. Interestingly, the correlations between pre-training losses and the fine-tuning performance" + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.09, + 0.899, + 0.196 + ], + "angle": 0, + "content": "
TaskOverfitNon-Overfit
Train LossVal LossTrain LossVal Loss
IN-1K+0.26-0.79-0.64-0.90
iNat-18+0.17-0.54-0.46-0.78
COCO OD+0.54-0.81-0.35-0.83
COCO IS+0.62-0.86-0.31-0.85
ADE-20K+0.75-0.91-0.14-0.90
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.208, + 0.892, + 0.248 + ], + "angle": 0, + "content": "Table 3. Pearson correlation coefficients between pre-training losses (training and validation losses) and fine-tuning performances on five downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.27, + 0.892, + 0.3 + ], + "angle": 0, + "content": "on multiple tasks could be observed with a phase transition around overfitting." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.303, + 0.892, + 0.394 + ], + "angle": 0, + "content": "Specifically, the correlation between training loss and fine-tuning performance is negative for the overfitting model (green circles) and positive for the non-overfitting model (red circles). The correlation between validation loss and fine-tuning performance is always negative, but the slope of their linear fit lines is significantly different." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.397, + 0.894, + 0.473 + ], + "angle": 0, + "content": "In addition, we further analyze the Pearson correlation coefficient between training loss and fine-tuning performance (Table 3), and find the validation loss has stronger correlation with fine-tuned performance than train loss for all cases, especially for non-overfitting models." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.49, + 0.82, + 0.505 + ], + "angle": 0, + "content": "4.4. Effects of Different Sizes of Decoders" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.516, + 0.892, + 0.698 + ], + "angle": 0, + "content": "We have studied the effects of encoder size from the data scaling perspective. Here, the effects of decoder size are further studied. We pre-train SwinV2-B models with decoder heads of different sizes on IN1K (20%), and Table 4 shows the results. Interestingly, although we find that the heavier decoder has lower training loss and higher validation loss than the linear decoder, indicating a more severe overfitting issue. But there is no decrease in its fine-tuning performance on ImageNet-1K than the linear decoder. This experiment shows that the decoder behaves very differently from the encoder, and we speculate that this is because the decoder \"blocks\" the damage to the encoder from overfitting." + }, + { + "type": "table", + "bbox": [ + 0.504, + 0.715, + 0.89, + 0.782 + ], + "angle": 0, + "content": "
Decoder# ParamsTrain LossVal LossTop-1 Acc
linear90.0M0.460.4784.4
4-blocks140.4M0.440.4884.4
8-blocks190.8M0.410.5084.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.498, + 0.792, + 0.894, + 0.848 + ], + "angle": 0, + "content": "Table 4. Results of different decoders, including converged training and validation losses of MIM pre-training, and fine-tuning performance (top-1 accuracy) on ImageNet-1K image classification. Encoders for all models are SwinV2-Base." + }, + { + "type": "page_footnote", + "bbox": [ + 0.516, + 0.887, + 0.768, + 0.9 + ], + "angle": 0, + "content": "3The least squares method is used for linear fit." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "10371" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.161, + 0.089, + 0.817, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.301, + 0.895, + 0.358 + ], + "angle": 0, + "content": "Figure 7. The correlations between pre-training losses (training and validation losses) and the fine-tuning performances of (a) ImageNet-1k image classification, (b) iNat 2018 fine-grained classification, (c) COCO object detection, (d) COCO instance segmentation, (e) and ADE-20K semantic segmentation. Pre-training losses are highly correlated with fine-tuning performance on all five tasks. Red circles indicate the overfitting models and green circles indicate non-overfitting models. Best viewed in color." + }, + { + "type": "table", + "bbox": [ + 0.089, + 0.368, + 0.46, + 0.42 + ], + "angle": 0, + "content": "
Dataset# ClassesTrain / Val LossTop-1 Acc
IN1K (10%)10000.351 / 0.51583.5
IN1001000.352 / 0.51183.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.431, + 0.472, + 0.502 + ], + "angle": 0, + "content": "Table 5. Results on different dataset sampling strategies (ImageNet-1K (10%) and ImageNet-100) with same dataset size \\( \\left( {{1.28} \\times {10}^{5}}\\right. \\) images for both), include converged training and validation losses of MIM pre-training, and fine-tuning performance (top-1 accuracy) on ImageNet-1K image classification." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.517, + 0.47, + 0.548 + ], + "angle": 0, + "content": "4.5. Impact of Different Dataset Sampling Strategies" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.557, + 0.473, + 0.769 + ], + "angle": 0, + "content": "We study different dataset sampling strategies by comparing the training behavior and fine-tuned performance of models pre-trained on IN1K (10%) and IN100. In IN1K (10%), the images are uniformly sampled from each category, and we randomly sample 100 categories from ImageNet-1K as IN100. Experiments are conducted on SwinV2-L with 500K training iterations. Table 5 shows the training loss, validation loss and fine-tuning top-1 accuracy of ImageNet-1K. For the two models pre-trained on IN1K (10%) and IN100, all three metrics are very similar. Figure 8 further illustrates the training dynamics of the two models, and we find both their training loss curves and validation loss curves are almost overlapping. These results show the disparity caused by different dataset sampling strategies is minor." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.784, + 0.197, + 0.8 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.472, + 0.903 + ], + "angle": 0, + "content": "In our work, we systematically study the data scaling capability of masked image modeling at different model sizes and training lengths. Based on the extensive experiments, we demonstrate that in order to scale up computes and model parameters, masked image modeling remains demanding for large-scale data due to the severe overfit" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.369, + 0.895, + 0.493 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.503, + 0.893, + 0.545 + ], + "angle": 0, + "content": "Figure 8. The training loss and validation loss of MIM pre-training with different dataset sampling strategies, ImageNet-1K (10%) and ImageNet-100. Best viewed in color." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.569, + 0.895, + 0.903 + ], + "angle": 0, + "content": "ting phenomenon when pre-training with small datasets of large models, which challenges the conclusions of previous literatures that a large dataset may not be necessary in masked image modeling. Besides, we also find that masked image modeling cannot benefit from more data under a nonoverfitting scenario, which diverges from the previous observations in large-scale pre-training language or vision models and raise new concerns and challenges on the data scalability of MIM. In addition, some intriguing properties of MIM are observed, such as larger models possessing higher sample efficiency and a strong correlation between the validation loss of masked image modeling and the fine-tuning performance. The former observation indicates that larger models will continue to have better performance and suggests that training larger models with fewer steps on sufficient data is a preferable option, while the latter observation suggests that validation loss can be considered as a good proxy indicator for evaluating pre-trained models, and makes it possible to reduce the experimental overhead of measuring models by fine-tuning. We hope that our findings could deepen the understanding of masked image modeling and facilitate future developments on large-scale vision models." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "10372" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.117, + 0.471, + 0.158 + ], + "angle": 0, + "content": "[1] Ibrahim Alabdulmohsin, Behnam Neyshabur, and Xiaohua Zhai. Revisiting neural scaling laws in language and vision. In NeurIPS, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.161, + 0.472, + 0.216 + ], + "angle": 0, + "content": "[2] Ibrahim M Alabdulmohsin and Mario Lucic. A near-optimal algorithm for debiasing trained machine learning models. Advances in Neural Information Processing Systems, 34:8072-8084, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.219, + 0.471, + 0.26 + ], + "angle": 0, + "content": "[3] Hangbo Bao, Li Dong, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.264, + 0.472, + 0.332 + ], + "angle": 0, + "content": "[4] Tom B Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.335, + 0.471, + 0.391 + ], + "angle": 0, + "content": "[5] Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jiarui Xu, et al. Mmdetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.394, + 0.471, + 0.435 + ], + "angle": 0, + "content": "[6] Mark Chen, Alec Radford, Rewon Child, Jeff Wu, and Heewoo Jun. Generative pretraining from pixels. Advances in Neural Information Processing Systems, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.438, + 0.471, + 0.48 + ], + "angle": 0, + "content": "[7] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. ICML, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.482, + 0.471, + 0.524 + ], + "angle": 0, + "content": "[8] Ting Chen, Simon Kornblith, Kevin Swersky, Mohammad Norouzi, and Geoffrey Hinton. Big self-supervised models are strong semi-supervised learners, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.527, + 0.472, + 0.596 + ], + "angle": 0, + "content": "[9] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 702-703, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.599, + 0.471, + 0.64 + ], + "angle": 0, + "content": "[10] Zihang Dai, Hanxiao Liu, Quoc V. Le, and Mingxing Tan. Coatnet: Marrying convolution and attention for all data sizes, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.643, + 0.471, + 0.685 + ], + "angle": 0, + "content": "[11] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248–255. IEEE, 2009." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.687, + 0.471, + 0.743 + ], + "angle": 0, + "content": "[12] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.746, + 0.471, + 0.842 + ], + "angle": 0, + "content": "[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[14] Alaaeldin El-Nouby, Gautier Izacard, Hugo Touvron, Ivan Laptev, Hervé Jegou, and Edouard Grave. Are large-scale datasets necessary for self-supervised pre-training? arXiv preprint arXiv:2112.10740, 2021." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.117, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[15] Yuxin Fang, Li Dong, Hangbo Bao, Xinggang Wang, and Furu Wei. Corrupted image modeling for self-supervised visual pre-training. arXiv preprint arXiv:2202.03382, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.176 + ], + "angle": 0, + "content": "[16] William Fedus, Barret Zoph, and Noam Shazeer. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.178, + 0.894, + 0.244 + ], + "angle": 0, + "content": "[17] Priya Goyal, Mathilde Caron, Benjamin Lefaudeauux, Min Xu, Pengchao Wang, Vivek Pai, Mannat Singh, Vitaliy Liptchinsky, Ishan Misra, Armand Joulin, et al. Self-supervised pretraining of visual features in the wild. arXiv preprint arXiv:2103.01988, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.247, + 0.892, + 0.287 + ], + "angle": 0, + "content": "[18] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dolkar, and Ross Girshick. Masked autoencoders are scalable vision learners. arXiv preprint arXiv:2111.06377, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.289, + 0.892, + 0.316 + ], + "angle": 0, + "content": "[19] Kaiming He, Georgia Gkioxari, Piotr Dólár, and Ross Girshick. Mask r-cnn. In ICCV, pages 2961-2969, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.317, + 0.892, + 0.383 + ], + "angle": 0, + "content": "[20] Joel Hestness, Sharan Narang, Newsha Ardalani, Gregory Diamos, Heewoo Jun, Hassan Kianinejad, Md Patwary, Mostofa Ali, Yang Yang, and Yanqi Zhou. Deep learning scaling is predictable, empirically. arXiv preprint arXiv:1712.00409, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.385, + 0.892, + 0.441 + ], + "angle": 0, + "content": "[21] Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B. Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.442, + 0.892, + 0.509 + ], + "angle": 0, + "content": "[22] Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Joan Puigcerver, Jessica Yung, Sylvain Gelly, and Neil Houlsby. Big transfer (bit): General visual representation learning. In European conference on computer vision, pages 491-507. Springer, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.511, + 0.892, + 0.579 + ], + "angle": 0, + "content": "[23] Zhuohan Li, Eric Wallace, Sheng Shen, Kevin Lin, Kurt Keutzer, Dan Klein, and Joey Gonzalez. Train big, then compress: Rethinking model size for efficient training and inference of transformers. In International Conference on Machine Learning, pages 5958-5968. PMLR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.581, + 0.892, + 0.649 + ], + "angle": 0, + "content": "[24] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.65, + 0.892, + 0.705 + ], + "angle": 0, + "content": "[25] Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, et al. Swin transformer v2: Scaling up capacity and resolution. arXiv preprint arXiv:2111.09883, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.707, + 0.892, + 0.761 + ], + "angle": 0, + "content": "[26] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. arXiv preprint arXiv:2103.14030, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.762, + 0.892, + 0.789 + ], + "angle": 0, + "content": "[27] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swin transformer, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.79, + 0.892, + 0.816 + ], + "angle": 0, + "content": "[28] Microsoft. Turing-nlg: A 17-billion-parameter language model by microsoft, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.818, + 0.892, + 0.859 + ], + "angle": 0, + "content": "[29] Microsoft. Using deepspeed and megatron to train megatron-turing nlg 530b, the world's largest and most powerful generative language model, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[30] Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A Efros. Context encoders: Feature learning by inpainting. In Proceedings of the IEEE conference on" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10373" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.47, + 0.119 + ], + "angle": 0, + "content": "computer vision and pattern recognition, pages 2536-2544, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.121, + 0.47, + 0.162 + ], + "angle": 0, + "content": "[31] Alec Radford, Karthik Narasimhan, Tim Salimans, and Ilya Sutskever. Improving language understanding by generative pre-training. 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.472, + 0.203 + ], + "angle": 0, + "content": "[32] Alec Radford, Jeff Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language models are unsupervised multitask learners. 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.205, + 0.471, + 0.273 + ], + "angle": 0, + "content": "[33] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research, 21(140):1-67, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.274, + 0.47, + 0.33 + ], + "angle": 0, + "content": "[34] Carlos Riquelme, Joan Puigcerver, Basil Mustafa, Maxim Neumann, Rodolphe Jenatton, André Susano Pinto, Daniel Keysers, and Neil Houlsby. Scaling vision with sparse mixture of experts. arXiv preprint arXiv:2106.05974, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.33, + 0.47, + 0.372 + ], + "angle": 0, + "content": "[35] Jonathan S Rosenfeld, Amir Rosenfeld, Yonatan Belinkov, and Nir Shavit. A constructive prediction of the generalization error across scales. arXiv preprint arXiv:1909.12673, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.372, + 0.471, + 0.455 + ], + "angle": 0, + "content": "[36] Aarohi Srivastava, Abhinav Rastogi, Abhishek Rao, Abu Awal Md Shoeb, Abubakar Abid, Adam Fisch, Adam R Brown, Adam Santoro, Aditya Gupta, Adrià Garriga-Alonso, et al. Beyond the imitation game: Quantifying and extrapolating the capabilities of language models. arXiv preprint arXiv:2206.04615, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.456, + 0.471, + 0.51 + ], + "angle": 0, + "content": "[37] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International conference on machine learning, pages 6105-6114. PMLR, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.511, + 0.471, + 0.566 + ], + "angle": 0, + "content": "[38] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. arXiv preprint arXiv:2203.12602, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.567, + 0.471, + 0.636 + ], + "angle": 0, + "content": "[39] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International Conference on Machine Learning, pages 10347-10357. PMLR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.637, + 0.471, + 0.718 + ], + "angle": 0, + "content": "[40] Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alex Shepard, Hartwig Adam, Pietro Perona, and Serge Belongie. The inaturalist species classification and detection dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8769-8778, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.72, + 0.471, + 0.775 + ], + "angle": 0, + "content": "[41] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.471, + 0.844 + ], + "angle": 0, + "content": "[42] Pascal Vincent, Hugo Larochelle, Yoshua Bengio, and Pierre-Antoine Manzagol. Extracting and composing robust features with denoising autoencoders. In Proceedings of the 25th international conference on Machine learning, pages 1096-1103, 2008." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[43] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In Proceedings of the European Conference on Computer Vision (ECCV), pages 418-434, 2018." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.147 + ], + "angle": 0, + "content": "[44] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. arXiv preprint arXiv:2111.09886, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.894, + 0.177 + ], + "angle": 0, + "content": "[45] Xiaohua Zhai, Alexander Kolesnikov, Neil Houlsby, and Lucas Beyer. Scaling vision transformers, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.894, + 0.22 + ], + "angle": 0, + "content": "[46] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.894, + 0.276 + ], + "angle": 0, + "content": "[47] Zhun Zhong, Liang Zheng, Guoliang Kang, Shaozi Li, and Yi Yang. Random erasing data augmentation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 13001-13008, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.277, + 0.894, + 0.333 + ], + "angle": 0, + "content": "[48] Bolei Zhou, Hang Zhao, Xavier Puig, Tete Xiao, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Semantic understanding of scenes through the ade20k dataset. International Journal on Computer Vision, 2018." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.333 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10374" + } + ] +] \ No newline at end of file diff --git a/2023/On Data Scaling in Masked Image Modeling/23a2cca2-7ab3-48ca-b93b-e899d7bc87f3_origin.pdf b/2023/On Data Scaling in Masked Image Modeling/23a2cca2-7ab3-48ca-b93b-e899d7bc87f3_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..15728589ca7979e9476e726184d0593d06865406 --- /dev/null +++ b/2023/On Data Scaling in Masked Image Modeling/23a2cca2-7ab3-48ca-b93b-e899d7bc87f3_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d856bf42e6b13d4f2f16f9ee2e72f223e080aac17f43c67ffbd154a36929f859 +size 2642967 diff --git a/2023/On Data Scaling in Masked Image Modeling/full.md b/2023/On Data Scaling in Masked Image Modeling/full.md new file mode 100644 index 0000000000000000000000000000000000000000..da7df84c21a9ccf8a88f190b11ac2808b6aabdbf --- /dev/null +++ b/2023/On Data Scaling in Masked Image Modeling/full.md @@ -0,0 +1,245 @@ +# On Data Scaling in Masked Image Modeling + +Zhenda Xie $^{13}$ , Zheng Zhang $^{3\dagger}$ , Yue Cao $^{3\dagger}$ , Yutong Lin $^{23}$ , Yixuan Wei $^{13}$ , Qi Dai $^{3}$ , Han Hu $^{3}$ $^{1}$ Tsinghua University $^{2}$ Xi'an Jiaotong University $^{3}$ Microsoft Research Asia + +{t-zhxie, zhez, yuecao, t-yutonglin, t-yixuanwei, qi.dai, hanhu}@microsoft.com + +# Abstract + +Scaling properties have been one of the central issues in self-supervised pre-training, especially the data scalability, which has successfully motivated the large-scale self-supervised pre-trained language models and endowed them with significant modeling capabilities. However, scaling properties seem to be unintentionally neglected in the recent trending studies on masked image modeling (MIM), and some arguments even suggest that MIM cannot benefit from large-scale data. In this work, we try to break down these preconceptions and systematically study the scaling behaviors of MIM through extensive experiments, with data ranging from $10\%$ of ImageNet-1K to full ImageNet22K, model parameters ranging from 49-million to one-billion, and training length ranging from 125K to 500K iterations. And our main findings can be summarized in two folds: 1) masked image modeling remains demanding large-scale data in order to scale up computes and model parameters; 2) masked image modeling cannot benefit from more data under a non-overfitting scenario, which diverges from the previous observations in self-supervised pre-trained language models or supervised pre-trained vision models. In addition, we reveal several intriguing properties in MIM, such as high sample efficiency in large MIM models and strong correlation between pre-training validation loss and transfer performance. We hope that our findings could deepen the understanding of masked image modeling and facilitate future developments on large-scale vision models. Code and models will be available at https://github.com/microsoft/SimMIM. + +# 1. Introduction + +Masked Image Modeling (MIM) [3, 18, 44], which has recently emerged in the field of self-supervised visual pretraining, has attracted widespread interest and extensive applications throughout the community for unleashing the superior modeling capacity of attention-based Transformer + +architectures [13, 26] and demonstrating excellent sample efficiency and impressive transfer performance on a variety of vision tasks. However, most recent practices are focused on the design of MIM methods, while the study of scaling properties of MIM is unintentionally neglected, especially the data scaling property, which successfully motivated the large-scale self-supervised pre-trained language models and endowed them with significant modeling capabilities. Although previous works [8, 10, 17, 37, 45] have explored several conclusions about the scaling properties of vision models, most of their findings were obtained under a supervised pre-training scheme or under a contrastive learning framework, so the extent to which these findings could be transferred to MIM still needs to be investigated. + +Meanwhile, with the emergence of Transformers [41] and masked language modeling (MLM) [12, 31], the systematic studies of scaling laws have already been explored in natural language processing field [21, 23, 35], which provided ample guidance for large models in recent years. The core finding drawn from scaling laws [21] for neural language models is that the performance has a power-law relationship with each of the three scale factors - model parameters $N$ , size of dataset $D$ , and amount of compute $C$ respectively - when not bottlenecked by the other two. This conclusion implies that better performance can be obtained by scaling up these three factors to the extent that the scaling laws are in effect, which led to the subsequent developments of large scale language models [16, 28, 29, 32, 33] that exhibit excellent modeling capabilities [4] on most language tasks. Therefore, it is natural to ask whether MIM possesses the same scaling signatures for vision models as the MLM method for language models, so that the scaling up of vision models can catch up with language models. + +Though masked image modeling and masked language modeling both belong to masked signal prediction, their property differences are also non-negligible due to the different nature of vision and language. That is, the images are highly redundant raw signals and words/sentences are semantically rich tokens, which may result in different abilities of data utilization, and it is thus debatable whether the observations from language models could be reproduced in vision models. + +![](images/816aa38777e260b912428594c678e4cfda475d1e0447938cfaa211f93f9383f5.jpg) +Model Flops $\times$ Training Iterations + +![](images/35bc9fb8857d9d4776c448afc9cffa445f8e3cc5f41903bd9d940b0b7dd520a1.jpg) +Dataset Size [D] +Millions of Images +Figure 1. The curves of validation loss of pre-training models w.r.t. the relative compute, dataset size and model size. MIM performance improves smoothly as we increase the relative compute and model size, but not improve when the dataset size is sufficient to prevent model from overfitting. We set the relative compute of SwinV2-S for 125K iterations as the value of 1. Best viewed in color. + +![](images/540236580d56040df83fa97a68c9efde933d6d82e9c1d81cd63c544cfecd44c6.jpg) +Model Size [N] +Millions of Params + +Moreover, recent studies [14, 38] have shown that using a small amount of training data in masked image modeling can achieve comparable performance to using large datasets, which also motivates our explorations as it disagrees with previous findings and intuitions. + +In this paper, we systematically investigate the scaling properties, especially the data scaling capability of masked image modeling in terms of different dataset sizes ranging from $10\%$ of ImageNet ( $\sim 0.1$ million) to full ImageNet-22K ( $\sim 14$ million), model sizes ranging from 49-million Swin-V2 Small to 1-billion Swin-V2 giant and training lengths ranging from 125K to 500K iterations. We use Swin Transformer V2 [25] as the vision encoder for its proven trainability of large models and applicability to a wide range of vision tasks, and adopt SimMIM [44] for masked image modeling pre-training because it has no restrictions on encoder architectures. We also conduct experiments with other MIM frameworks like MAE [18] and other vision encoder like the widely used ViT [13] to verify the generalizability of our findings. With these experimental setups, our main findings could be summarized into two folds: + +i) Masked image modeling remains demanding large-scale data in order to scale up computes and model parameters. We empirically find that MIM performance has a power-law relationship with relative compute and model size when not bottlenecked by the dataset size (Figure 1). Besides, we observe that smaller datasets lead to severe overfitting phenomenon for training large models (Figure 2), and the size of the dataset to prevent model from overfitting increases clearly as the model increases (Figure 5-Left). Therefore, from the perspective of scaling up model, MIM still demands large-scale data. Furthermore, if we train large-scale models of different lengths, we find that relatively small datasets are adequate at shorter training lengths, but still suffer from overfitting at longer training lengths (Figure 5-Right), which further demonstrate the data scalability of MIM for scaling up compute. + +2) Masked image modeling cannot benefit from more data + +under a non-overfitting scenario, which diverges from the previous observations in self-supervised pre-train language models or supervised pre-trained vision models. In MIM, we find that increasing the number of unique samples for a non-overfitting model does not provide additional benefits to performance (Figure 1). This behavior differs from previous observations in supervised vision transformers and self-supervised language models, where the model performance increases as the number of unique samples increases. + +In addition, we also demonstrate some intriguing properties about masked image modeling, such as larger models possessing higher sample efficiency, i.e., fewer optimization steps are required for larger models to achieve same performance (Section 4.2); and the consistency between transfer and test performance, i.e., test performance could be used to indicate the results on downstream tasks (Section 4.3). These observations also correspond to those in previous practices. + +These findings on the one hand confirm the effect of MIM on scaling up model size and compute, and raise new concerns and challenges on the data scalability of MIM on the other. We hope that our findings could deepen the understanding of masked image modeling and facilitate future developments on large-scale vision models. + +# 2. Related Work + +Masked Image Modeling Masked Image Modeling learns representations by reconstructing the masked content of images, and its early exploration can be traced back to context encoder [30] and denoising autoencoder [42]. Recently, iGPT [6], BEiT [3], MAE [18] and SimMIM [44] recall this approach on training vision transformer. iGPT [6] sequentially predicted the pixels by auto-regressive manner. BEiT [3] proposed to predict the discrete visual tokens. MAE [18] and SimMIM [44] concurrently found predicting the raw pixels with a high masking ratio could work well. In this work, we use SimMIM as the default masked image modeling approach, because of its simplicity and no restrictions on the architecture of vision encoder like MAE. + +![](images/2dcd36573cf2bed72f60350d216b6eab723dc27a8a33f9f5c846bf8ae71164d3.jpg) +Figure 2. Training loss of pre-training, validation loss of pre-training, and fine-tuning accuracy on ImageNet-1K of different model sizes, data sizes and training lengths, w.r.t. the relative pre-training compute. All models are SwinV2 models pre-trained with SimMIM. We set the training compute of SwinV2-S for 125K iterations as the value of 1. Bigger circles indicate larger models. Best viewed in color. + +Vision Transformers Transformer [41] was first applied to natural language processing and became the dominant architecture, and has recently attracted a lot of attention in computer vision. The pioneering work ViT [13] first shows that the transformer architecture works well in image classification when trained on large amounts of data. DeiT [39] proposed a better training recipe based on ViT and demonstrated that vision Transformer has promising performance when only using ImageNet-1K dataset. Swin Transformer [26] improves plain ViT by inducing the hierarchical architecture and non-overlapping local attention and successfully demonstrates the effectiveness of vision transformer on a wide range of vision tasks. Swin Transformer V2 [25] further addresses the training stability issue of [26] in model scaling and illustrates better performance than the original Swin Transformer, and thus we use it as the default vision encoder in this work. + +Scaling Vision Models Many works [25, 34, 37, 45] examine how to scale vision models, but most are more concerned with exploring the perspective of model architecture designs. For example, EfficientNet [37] extensively studied how model width, model depth and input resolution affect the convolutional neural networks; [34] proposed to scale vision model with sparse mixture-of-expert; [45] and [25] studied how to scale ViT and Swin Transformer, respectively. In addition, several studies [1, 2, 45] explored the aspect of data scaling under the pre-training fine-tuning paradigm. BiT [22] revisited the supervised pre-training on a wide range of data scales up to 1M images. SEER [17] studied the effectiveness of data scaling in the contrastive learning framework with up to one billion images. Recently, SplitMask [14] find that masked image modeling is robust to the size of pre-training data and challenges the data scaling capability of masked image modeling, which is most relevant to our work. + +Scaling Language Models Many works [20, 23, 35] systematically analyzed the scaling behavior on a relatively small scale. And [21] studied empirical scaling laws for + +large scale neural language models and revealed that the loss scaled as a power-law with model size, dataset size, and the amount of compute used for training. Motivated by these laws, several large scale language models [4,16,28,29,32,33] were successfully trained and demonstrated excellent modeling capabilities on most language tasks. [36] established a comprehensive, large-scale benchmark designed to assist in quantifying and extrapolating the capabilities of large language models. + +# 3. Background and Experimental Setup + +# 3.1. Masked Image Modeling + +Masked image modeling is used to train the vision model by taking a corrupted image as input and predicting the content of the masked region as the target. In this study, we use SimMIM [44] as the default masked image modeling approach because of its simplicity and lack of restrictions on the architecture of the vision encoder. SimMIM consists of a visual encoder and an extremely lightweight prediction head of a linear layer for predicting the raw pixels of the corrupted images via $\ell_2$ regression loss. To facilitate the implementation of the vision transformer, SimMIM adopts the patch-wise mask strategy with the masked patch size of $32 \times 32$ and mask ratio of 0.6. To further alleviate the local dependency of raw pixels, we improved the SimMIM by normalizing the predicted target according to [15] with a sliding window of $47^2$ . As the result, a slight performance improvement is observed. In addition, we conduct experiments using MAE as the masked image modeling approach to verify the methodological generalizability of our findings. These experiments strictly follow the settings in [18]. + +# 3.2. Architecture Specifications + +We use Swin Transformer V2 [25] as the default vision encoder in this study. Thanks to its generality and scalability, we evaluate a series of SwinV2 models with a wide range of model sizes (the number of parameters ranges from $\sim 50\mathrm{M}$ to $\sim 1\mathrm{B}$ , and FLOPs range from $\sim 9\mathrm{G}$ to $\sim 190\mathrm{G}$ ) on + +
ModelBase ChannelDepthHeadWindow SizeBackbone Params
pre-trainfine-tune
SwinV2-S96{2, 2, 18, 2}{3, 6, 12, 24}121449M
SwinV2-B128{2, 2, 18, 2}{4, 8, 16, 32}121487M
SwinV2-L192{2, 2, 18, 2}{6, 12, 24, 48}1214195M
SwinV2-H352{2, 2, 18, 2}{11, 22, 44, 88}1214655M
SwinV2-g448{2, 2, 18, 2}{14, 28, 56, 112}12141061M
+ +Table 1. Detailed architecture specifications. Note that, the model aliases used in our work are different from [25]. + +
IN1K (10%)IN1K (20%)IN1K (50%)IN100IN1K(100%)IN22K(100%)
#Classes\( 1 \times 10^{3} \)\( 1 \times 10^{3} \)\( 1 \times 10^{3} \)\( 1 \times 10^{2} \)\( 1 \times 10^{3} \)\( 2.18 \times 10^{4} \)
#Images\( 1.28 \times 10^{5} \)\( 2.56 \times 10^{5} \)\( 6.41 \times 10^{5} \)\( 1.27 \times 10^{5} \)\( 1.28 \times 10^{6} \)\( 1.42 \times 10^{7} \)
+ +Table 2. Detailed dataset specifications used in the pre-training of masked image modeling. + +multiple downstream tasks. The detailed model specifications are shown in Table 1. We use a new variant SwinV2-g (giant), with number of parameters between SwinV2-L and the 3-billion-parameter SwinV2-G (Giant) used in [25]. In addition, we conduct experiments with a series of ViT models [13] to prove the architectural generalizability of our findings. Specifically, we use ViT-B/16, ViT-L/16 and ViTH/14 according to the settings from [18]. + +# 3.3. Pre-training Datasets + +To study the effect of data size on masked image modeling, we build datasets with different sizes. We use the training set of ImageNet-1K and ImageNet-22K as two large-scale datasets, and randomly sample $10\%$ , $20\%$ , $50\%$ of images in the ImageNet-1K training set as smaller datasets. By default, the images are uniformly sampled from each category. We also consider the sampling strategies could perform differently. To this end, we randomly sample 100 classes from ImageNet-1K as ImageNet-100, and compare it with ImageNet-1K $(10\%)$ but find their training loss and fine-tuning performance are almost the same. The details and statistics of all pre-training datasets used in our study are shown in Table 2. + +# 3.4. Pre-training Details + +To better compare the performance of models with different amounts of data under the same pre-training length, we use training iterations rather than training epochs and adopt the same hyper-parameters for all models with different sizes during pre-training. The total number of training iterations is in $\{125\mathrm{K}, 250\mathrm{K}, 500\mathrm{K}\}$ and the batch size is set as 2048 for all experiments. In pre-training stage, we use the same hyper-parameters for all models, and the training details and hyper-parameters of pre-training are summarized in Appendix. Because of the excessive amount of experiments, we follow SimMIM [44] and also use the following two techniques for reducing the experimental overheads: First, we use the step learning rate scheduler in pre-training for + +sharing the first training step among experiments with different training lengths. The first 7/8 training iterations are the first step and the last 1/8 training iterations are the second step with the learning rate ratio of 0.1 (i.e. learning rate is divided by 10 in the second step). Second, we adopt the input image size of $192^2$ and set the window size of 12. We improve the SimMIM by normalizing the predicted target according to [15] with a sliding window of $47^2$ and observe an improvement of 0.3 on top-1 accuracy of ImageNet-1K for the SwinV2-Large model. The same light data augmentation strategy as SimMIM is used: random resize cropping with a scale range of [0.67, 1], an aspect ratio range of [3/4, 4/3] and a random flipping with probability 0.5. + +# 3.5. Fine-tuning Tasks + +To extensively and accurately evaluate the performance of pre-trained models under different pre-training schedulers and datasets, a series of diverse and representative tasks including fine-tuning on ImageNet-1K, fine-grained image classification, object detection, instance segmentation, and semantic segmentation are selected for evaluation. Detailed setups of all experiments are illustrated in Appendix. + +ImageNet-1K We follow [3] to evaluate the quality of learnt representations by fine-tuning the pre-trained models on ImageNet-1K [11] image classification task, which is the most commonly used scenario and evaluation criterion for pre-trained models [18,44]. Different from pre-training, We adopt the image size with $224^2$ with window size of 14 in fine-tuning. The AdamW with batch size of 2048, base learning rate of 5e-3, weight decay of 0.05, $\beta_{1}$ of 0.9 and $\beta_{2}$ of 0.999 are used, and we adopt cosine learning rate scheduler. As larger models are more prone to overfitting, we fine-tune SwinV2-S/B/L for 100 epochs with 20 warm-up epochs and SwinV2-H/g for 50 epochs with 10 warm-up epochs, and decrease the layer decay as the model size increases. In addition, gradient clipping, stochastic depth, label smoothing and data augmentations (e.g. random crop, + +![](images/f7af21374200d9e8f70bb2f5e77f76ea22f09720188570767effe034a4bedea8.jpg) +Figure 3. Left to right: Performances with SwinV2 models on (a) COCO object detection, (b) COCO instance segmentation, (c) iNaturalist-18, and (d) ADE20K semantic segmentation w.r.t. the relative compute. Note that the relative compute indicates the pre-training compute. We set the relative compute of SwinV2-S for 125K iterations as 1. Bigger circles indicate larger models. Best viewed in color. + +rand erasing [47], rand augment [9], mixup [46], cutmix [46], etc.) are also used by following [44]. + +iNaturalist-18 iNaturalist [40] 2018 is a long-tailed fine-grained image classification dataset. As fine-tuning in ImageNet-1K, we also use the input image size of $224^2$ , window size of 14 and patch size of 4 in iNaturalist 2018. We fine-tune all models for 100 epochs with 20 warm-up epochs, and set layer decay to 0.8, 0.75 and 0.7 for SwinV2-S/B/L, respectively. The AdamW optimizer with cosine learning rate scheduler, batch size of 2048, base learning rate of 1.6e-2, weight decay of 0.1, $\beta_{1}$ of 0.9 and $\beta_{2}$ of 0.999 are used. In addition, we also adopt stochastic depth, label smoothing, gradient clipping and data augmentations in fine-tuning. + +# COCO Object Detection and Instance Segmentation [24] + +We use Mask R-CNN [19] for evaluation. We set the window size to 14 and patch size to 4. The AdamW optimizer with batch size of 32, base learning rate of 8e-5, weight decay of 0.05, $\beta_{1}$ of 0.9, $\beta_{2}$ of 0.999 and a step learning rate scheduler (step learning rate ratio of 0.1, step epochs are 27 and 33) are used. In training, the random cropping with crop size of [1024, 1024], large scale jittering with a range of [0.1, 2.0], random horizontal flip with probability 0.5, and stochastic depth regularization are used. In testing, all images are resized to (800, 1333) and keeping the aspect ratio unchanged. + +ADE20K Semantic Segmentation [48] Following [26], we use UPerNet [43] for evaluation. We set the window size to 20 and the patch size to 4. The AdamW optimizer with batch size of 32, base learning rate searched in a range of [1e-4, 3e-4], weight decay of 0.05, $\beta_{1}$ of 0.9, $\beta_{2}$ of 0.999 and a linear learning rate scheduler with a total of 80K iterations are used. Also, we use the layer decay of 0.95, 0.95, 0.9 for SwinV2-S/B/L, respectively. In training, the random cropping with crop size of [640, 640], scale jittering with a range of [0.5, 2.0], random horizontal flip with probability 0.5, random photometric distortion and stochastic + +depth regularization of 0.1 are used. In testing, all images are evaluated by sliding window manner, and use the test image size of (2560, 640) and set sliding window stride to 426, following [26, 44]. + +# 4. Results and Findings + +We train numerous models with different training lengths, dataset sizes, and model sizes, and study how these factors affect the performance of masked image modeling. Figure 1 illustrates the validation loss of pre-training $^2$ with respect to the relative compute, dataset size and model size. Figure 2 illustrates the training loss of pre-training, validation loss of pre-training, and the fine-tuning top-1 accuracy of ImageNet-1K with respect to the relative compute. Based on these extensive experiments, we make the following observations: + +# 4.1. Data Scaling in Masked Image Modeling + +Masked image modeling remains demanding for large-scale data. When with the high masking rate (e.g., $60\%$ in our work), the masked image modeling is considered a very challenging training objective and has been found to be data efficient by previous literature [14,27], i.e., a comparable performance can be achieved with small datasets as with large datasets. However, Figure 2 shows that as the training cost increases, the training loss of some models drops significantly, and their validation loss rises significantly, even on using $50\%$ images of ImageNet-1K (i.e., IN1K $(50\%)$ ), indicating the overfitting phenomenon exists. In Figure 5-Left, we demonstrate that the size of the dataset to prevent model from overfitting increases clearly as the model increases. And significant decrease to the fine-tuning performance caused by overfitting could be observed in Figure 2 and Figure 3. Moreover, we measure the best fine-tuning performance of each model trained by different training schedulers in Figure 2-Right. We find the large models perform even worse than smaller models when small dataset is used for training. For + +![](images/3fafda11fc16c0eb1c8d21d46c0bc528caf84840c45ad6ee96a365b19f2151ce.jpg) +Figure 4. Training loss of pre-training, validation loss of pre-training, and fine-tuning accuracy on ImageNet-1K of different model sizes, data sizes and training lengths, w.r.t. the relative pre-training compute. All models are ViT models pre-trained with MAE. We set the training compute of ViT-B/16 for 125K iterations as the value of 1. Bigger circles indicate larger models. Best viewed in color. + +example, the best top-1 accuracy of SwinV2-H with IN1K (20%) is 84.4, worse than the best performance of SwinV2-L by 0.3. In addition, by comparing the best performance that can be obtained using different sizes of dataset, we find that using more data results in better performance. These observations suggest that masked image modeling does not alleviate the demands of large dataset. + +The training length matters. Larger models can benefit from more data at a longer training length. By comparing the performance of models pre-trained by different data sizes in Figure 2-Right, we find that the fine-tuning performance of the large models saturates more slowly with the increasing data size compared to the smaller models. For example, the SwinV2-S model pre-trained on IN1K (50%) has a very similar fine-tuning performance to the model pre-trained on IN1K (100%). In comparison, the performance difference between the SwinV2-H model pre-trained on IN1K (50%) and IN1K (100%) is near 0.5, which is a significant gap for ImageNet-1K classification. + +Furthermore, a comprehensive observation reveals that the improvements from using more data are not significant under short training lengths. As shown in Figure 5-Right, while there is a noticeable performance gap between SwinV2-H trained on IN1K (50%) and IN1K (100%) at a training length of 500K iterations, the gap is negligible at a training length of 125K iterations. This observation suggests that increasing the training length for larger models is critical to benefit from more data. + +Masked image modeling cannot benefit from more data under a non-overfitting scenario. As illustrated in Figure 1-Center, we plot the validation loss of pre-trained SwinV2 models with respect to different dataset sizes, and we demonstrate that for a particular model size and training length, there will be a certain dataset size that will keep it from overfitting. And the model cannot benefit from dataset larger than this size. This behavior differs from previous observations in supervised vision transformers [45] and self-supervised language models [21], where the model perfor + +mance increases as the number of unique samples increases. We speculate that this may be related to the characteristics of the task itself, i.e., MIM provides a large number of training signals, making the model hard to learn more patterns from large-scale data. + +![](images/6c5ea36ba70b6898f3f5a6ab888582dda2f4ec0e3dc889ae444ff1d3b64f534e.jpg) +Figure 5. Left: Validation loss with respect to model size. The size of the dataset to prevent model from overfitting increases clearly as the model increases. Right: Validation loss with respect to training length using SwinV2-Huge. $50\%$ subset of ImageNet-1K is sufficient for training 125K iterations, but not sufficient to prevent overfitting for 500K iterations. Best viewed in color. + +Evaluation on more tasks. In addition to ImageNet-1K image classification, we also evaluate the MIM pre-trained SwinV2-S, SwinV2-B and SwinV2-L on iNaturalist-18 fine-grained image classification, ADE20K semantic segmentation, and COCO object detection/segmentation. Figure 3 shows a similar pattern with ImageNet-1K (Figure. 2-Right) that as the training cost increases, some models have significantly performance drop. In addition, the smaller models rapidly reach saturation as the amount of data increases, while larger models can continuously benefit from more data after sufficient training. These results suggest that the conclusions drawn on ImageNet-1K are broadly applicable to other vision tasks. + +Results with MAE and ViT Figure 4 demonstrates the results of ViT pre-trained with MAE. Similar to SwinV2 models pre-trained with SimMIM, we observe the same + +overfitting phenomenon when training with small datasets or large models, which makes MAE still demand for large-scale data. Besides, we could find that larger ViT models can also benefit from more data at a longer training lengths. These experiments verify the methodological and architectural generalizability of our results and findings. + +![](images/1825b07385a5a207ca6589331aedd1a04c57ba8cb6a4b2c3efb4b6722efaf239.jpg) +Figure 6. Validation loss and fine-tuning error rate on ImageNet-1K with respect to images processed during pre-training. Large models are more sample efficient than small models to achieve lower validation loss or fine-tuning error rate with fewer optimization steps. Best viewed in color. + +# 4.2. Larger Models Possess Higher Sample Efficiency + +Figure 6 shows the validation loss and fine-tuning error rate on ImageNet-1K with respect to the images processed (batch size times number of steps) during pre-training stage. Five Swin-V2 models with different sizes are pre-trained on full-set of ImageNet-1K. From these results, we observe that larger models possess higher sample efficiency, reaching the same level of validation loss or fine-tuning error rate on ImageNet with fewer optimization steps. These results indicate that larger models will continue to have better performance, and training larger models with fewer steps on sufficient data is a preferable choice. This observation is also in line with previous works on self-supervised language models [21] and supervised vision transformers [45]. + +# 4.3. Correlation between Pre-training Losses and the Fine-tuning Performance + +Evaluating a pre-trained model by its fine-tuned performance on downstream tasks is costly. In supervised pretraining, the validation accuracy is used as the proxy indicator to evaluate the quality of the pre-trained models. While in previous studies [7] on other self-supervised learning approaches (e.g., contrastive learning), such a proxy indicator is lacking. In this study, we would like to explore whether the pre-training loss in the training of masked image modeling is a good indicator of its fine-tuning performance. We collect all pre-trained models and plot their training and validation loss curves on Figure 7. Interestingly, the correlations between pre-training losses and the fine-tuning performance + +
TaskOverfitNon-Overfit
Train LossVal LossTrain LossVal Loss
IN-1K+0.26-0.79-0.64-0.90
iNat-18+0.17-0.54-0.46-0.78
COCO OD+0.54-0.81-0.35-0.83
COCO IS+0.62-0.86-0.31-0.85
ADE-20K+0.75-0.91-0.14-0.90
+ +on multiple tasks could be observed with a phase transition around overfitting. + +Specifically, the correlation between training loss and fine-tuning performance is negative for the overfitting model (green circles) and positive for the non-overfitting model (red circles). The correlation between validation loss and fine-tuning performance is always negative, but the slope of their linear fit lines is significantly different. + +In addition, we further analyze the Pearson correlation coefficient between training loss and fine-tuning performance (Table 3), and find the validation loss has stronger correlation with fine-tuned performance than train loss for all cases, especially for non-overfitting models. + +# 4.4. Effects of Different Sizes of Decoders + +We have studied the effects of encoder size from the data scaling perspective. Here, the effects of decoder size are further studied. We pre-train SwinV2-B models with decoder heads of different sizes on IN1K (20%), and Table 4 shows the results. Interestingly, although we find that the heavier decoder has lower training loss and higher validation loss than the linear decoder, indicating a more severe overfitting issue. But there is no decrease in its fine-tuning performance on ImageNet-1K than the linear decoder. This experiment shows that the decoder behaves very differently from the encoder, and we speculate that this is because the decoder "blocks" the damage to the encoder from overfitting. + +Table 3. Pearson correlation coefficients between pre-training losses (training and validation losses) and fine-tuning performances on five downstream tasks. + +
Decoder# ParamsTrain LossVal LossTop-1 Acc
linear90.0M0.460.4784.4
4-blocks140.4M0.440.4884.4
8-blocks190.8M0.410.5084.5
+ +Table 4. Results of different decoders, including converged training and validation losses of MIM pre-training, and fine-tuning performance (top-1 accuracy) on ImageNet-1K image classification. Encoders for all models are SwinV2-Base. + +![](images/dd935eddff1fb13ba7c9ab22ea78df540c1495b04b8c3d4259107d66a4d3fb8b.jpg) +Figure 7. The correlations between pre-training losses (training and validation losses) and the fine-tuning performances of (a) ImageNet-1k image classification, (b) iNat 2018 fine-grained classification, (c) COCO object detection, (d) COCO instance segmentation, (e) and ADE-20K semantic segmentation. Pre-training losses are highly correlated with fine-tuning performance on all five tasks. Red circles indicate the overfitting models and green circles indicate non-overfitting models. Best viewed in color. + +
Dataset# ClassesTrain / Val LossTop-1 Acc
IN1K (10%)10000.351 / 0.51583.5
IN1001000.352 / 0.51183.4
+ +Table 5. Results on different dataset sampling strategies (ImageNet-1K (10%) and ImageNet-100) with same dataset size $\left( {{1.28} \times {10}^{5}}\right.$ images for both), include converged training and validation losses of MIM pre-training, and fine-tuning performance (top-1 accuracy) on ImageNet-1K image classification. + +# 4.5. Impact of Different Dataset Sampling Strategies + +We study different dataset sampling strategies by comparing the training behavior and fine-tuned performance of models pre-trained on IN1K (10%) and IN100. In IN1K (10%), the images are uniformly sampled from each category, and we randomly sample 100 categories from ImageNet-1K as IN100. Experiments are conducted on SwinV2-L with 500K training iterations. Table 5 shows the training loss, validation loss and fine-tuning top-1 accuracy of ImageNet-1K. For the two models pre-trained on IN1K (10%) and IN100, all three metrics are very similar. Figure 8 further illustrates the training dynamics of the two models, and we find both their training loss curves and validation loss curves are almost overlapping. These results show the disparity caused by different dataset sampling strategies is minor. + +# 5. Conclusion + +In our work, we systematically study the data scaling capability of masked image modeling at different model sizes and training lengths. Based on the extensive experiments, we demonstrate that in order to scale up computes and model parameters, masked image modeling remains demanding for large-scale data due to the severe overfit + +![](images/85a9bd2e8ee20d32e2de541ecf81af2ac5f1cac0bd2bdbf5f34ffa3b0f7338f4.jpg) +Figure 8. The training loss and validation loss of MIM pre-training with different dataset sampling strategies, ImageNet-1K (10%) and ImageNet-100. Best viewed in color. + +ting phenomenon when pre-training with small datasets of large models, which challenges the conclusions of previous literatures that a large dataset may not be necessary in masked image modeling. Besides, we also find that masked image modeling cannot benefit from more data under a nonoverfitting scenario, which diverges from the previous observations in large-scale pre-training language or vision models and raise new concerns and challenges on the data scalability of MIM. In addition, some intriguing properties of MIM are observed, such as larger models possessing higher sample efficiency and a strong correlation between the validation loss of masked image modeling and the fine-tuning performance. The former observation indicates that larger models will continue to have better performance and suggests that training larger models with fewer steps on sufficient data is a preferable option, while the latter observation suggests that validation loss can be considered as a good proxy indicator for evaluating pre-trained models, and makes it possible to reduce the experimental overhead of measuring models by fine-tuning. We hope that our findings could deepen the understanding of masked image modeling and facilitate future developments on large-scale vision models. + +# References + +[1] Ibrahim Alabdulmohsin, Behnam Neyshabur, and Xiaohua Zhai. Revisiting neural scaling laws in language and vision. In NeurIPS, 2022. +[2] Ibrahim M Alabdulmohsin and Mario Lucic. A near-optimal algorithm for debiasing trained machine learning models. Advances in Neural Information Processing Systems, 34:8072-8084, 2021. +[3] Hangbo Bao, Li Dong, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021. +[4] Tom B Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. +[5] Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jiarui Xu, et al. Mmdetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155, 2019. +[6] Mark Chen, Alec Radford, Rewon Child, Jeff Wu, and Heewoo Jun. Generative pretraining from pixels. Advances in Neural Information Processing Systems, 2020. +[7] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. ICML, 2020. +[8] Ting Chen, Simon Kornblith, Kevin Swersky, Mohammad Norouzi, and Geoffrey Hinton. Big self-supervised models are strong semi-supervised learners, 2020. +[9] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 702-703, 2020. +[10] Zihang Dai, Hanxiao Liu, Quoc V. Le, and Mingxing Tan. Coatnet: Marrying convolution and attention for all data sizes, 2021. +[11] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248–255. IEEE, 2009. +[12] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. +[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. +[14] Alaaeldin El-Nouby, Gautier Izacard, Hugo Touvron, Ivan Laptev, Hervé Jegou, and Edouard Grave. Are large-scale datasets necessary for self-supervised pre-training? arXiv preprint arXiv:2112.10740, 2021. + +[15] Yuxin Fang, Li Dong, Hangbo Bao, Xinggang Wang, and Furu Wei. Corrupted image modeling for self-supervised visual pre-training. arXiv preprint arXiv:2202.03382, 2022. +[16] William Fedus, Barret Zoph, and Noam Shazeer. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity, 2021. +[17] Priya Goyal, Mathilde Caron, Benjamin Lefaudeauux, Min Xu, Pengchao Wang, Vivek Pai, Mannat Singh, Vitaliy Liptchinsky, Ishan Misra, Armand Joulin, et al. Self-supervised pretraining of visual features in the wild. arXiv preprint arXiv:2103.01988, 2021. +[18] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dolkar, and Ross Girshick. Masked autoencoders are scalable vision learners. arXiv preprint arXiv:2111.06377, 2021. +[19] Kaiming He, Georgia Gkioxari, Piotr Dólár, and Ross Girshick. Mask r-cnn. In ICCV, pages 2961-2969, 2017. +[20] Joel Hestness, Sharan Narang, Newsha Ardalani, Gregory Diamos, Heewoo Jun, Hassan Kianinejad, Md Patwary, Mostofa Ali, Yang Yang, and Yanqi Zhou. Deep learning scaling is predictable, empirically. arXiv preprint arXiv:1712.00409, 2017. +[21] Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B. Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models, 2020. +[22] Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Joan Puigcerver, Jessica Yung, Sylvain Gelly, and Neil Houlsby. Big transfer (bit): General visual representation learning. In European conference on computer vision, pages 491-507. Springer, 2020. +[23] Zhuohan Li, Eric Wallace, Sheng Shen, Kevin Lin, Kurt Keutzer, Dan Klein, and Joey Gonzalez. Train big, then compress: Rethinking model size for efficient training and inference of transformers. In International Conference on Machine Learning, pages 5958-5968. PMLR, 2020. +[24] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. +[25] Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, et al. Swin transformer v2: Scaling up capacity and resolution. arXiv preprint arXiv:2111.09883, 2021. +[26] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. arXiv preprint arXiv:2103.14030, 2021. +[27] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swin transformer, 2021. +[28] Microsoft. Turing-nlg: A 17-billion-parameter language model by microsoft, 2020. +[29] Microsoft. Using deepspeed and megatron to train megatron-turing nlg 530b, the world's largest and most powerful generative language model, 2021. +[30] Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A Efros. Context encoders: Feature learning by inpainting. In Proceedings of the IEEE conference on + +computer vision and pattern recognition, pages 2536-2544, 2016. +[31] Alec Radford, Karthik Narasimhan, Tim Salimans, and Ilya Sutskever. Improving language understanding by generative pre-training. 2018. +[32] Alec Radford, Jeff Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language models are unsupervised multitask learners. 2019. +[33] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research, 21(140):1-67, 2020. +[34] Carlos Riquelme, Joan Puigcerver, Basil Mustafa, Maxim Neumann, Rodolphe Jenatton, André Susano Pinto, Daniel Keysers, and Neil Houlsby. Scaling vision with sparse mixture of experts. arXiv preprint arXiv:2106.05974, 2021. +[35] Jonathan S Rosenfeld, Amir Rosenfeld, Yonatan Belinkov, and Nir Shavit. A constructive prediction of the generalization error across scales. arXiv preprint arXiv:1909.12673, 2019. +[36] Aarohi Srivastava, Abhinav Rastogi, Abhishek Rao, Abu Awal Md Shoeb, Abubakar Abid, Adam Fisch, Adam R Brown, Adam Santoro, Aditya Gupta, Adrià Garriga-Alonso, et al. Beyond the imitation game: Quantifying and extrapolating the capabilities of language models. arXiv preprint arXiv:2206.04615, 2022. +[37] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International conference on machine learning, pages 6105-6114. PMLR, 2019. +[38] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. arXiv preprint arXiv:2203.12602, 2022. +[39] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International Conference on Machine Learning, pages 10347-10357. PMLR, 2021. +[40] Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alex Shepard, Hartwig Adam, Pietro Perona, and Serge Belongie. The inaturalist species classification and detection dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8769-8778, 2018. +[41] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. +[42] Pascal Vincent, Hugo Larochelle, Yoshua Bengio, and Pierre-Antoine Manzagol. Extracting and composing robust features with denoising autoencoders. In Proceedings of the 25th international conference on Machine learning, pages 1096-1103, 2008. +[43] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In Proceedings of the European Conference on Computer Vision (ECCV), pages 418-434, 2018. + +[44] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. arXiv preprint arXiv:2111.09886, 2021. +[45] Xiaohua Zhai, Alexander Kolesnikov, Neil Houlsby, and Lucas Beyer. Scaling vision transformers, 2021. +[46] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412, 2017. +[47] Zhun Zhong, Liang Zheng, Guoliang Kang, Shaozi Li, and Yi Yang. Random erasing data augmentation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 13001-13008, 2020. +[48] Bolei Zhou, Hang Zhao, Xavier Puig, Tete Xiao, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Semantic understanding of scenes through the ade20k dataset. International Journal on Computer Vision, 2018. \ No newline at end of file diff --git a/2023/On Data Scaling in Masked Image Modeling/images.zip b/2023/On Data Scaling in Masked Image Modeling/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..4cd57e93f6ea1b233db253f87b1825fa9d89d6de --- /dev/null +++ b/2023/On Data Scaling in Masked Image Modeling/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:622bc22f1bada07f609c9589a825663eb313dc87d0a39e9747117a0268937947 +size 428409 diff --git a/2023/On Data Scaling in Masked Image Modeling/layout.json b/2023/On Data Scaling in Masked Image Modeling/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5a3d7d3ea378556bdddb380a54998934588b3e33 --- /dev/null +++ b/2023/On Data Scaling in Masked Image Modeling/layout.json @@ -0,0 +1,6871 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 159, + 103, + 434, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 103, + 434, + 121 + ], + "spans": [ + { + "bbox": [ + 159, + 103, + 434, + 121 + ], + "type": "text", + "content": "On Data Scaling in Masked Image Modeling" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 79, + 142, + 512, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 142, + 512, + 171 + ], + "spans": [ + { + "bbox": [ + 79, + 142, + 512, + 171 + ], + "type": "text", + "content": "Zhenda Xie" + }, + { + "bbox": [ + 79, + 142, + 512, + 171 + ], + "type": "inline_equation", + "content": "^{13}" + }, + { + "bbox": [ + 79, + 142, + 512, + 171 + ], + "type": "text", + "content": ", Zheng Zhang" + }, + { + "bbox": [ + 79, + 142, + 512, + 171 + ], + "type": "inline_equation", + "content": "^{3\\dagger}" + }, + { + "bbox": [ + 79, + 142, + 512, + 171 + ], + "type": "text", + "content": ", Yue Cao" + }, + { + "bbox": [ + 79, + 142, + 512, + 171 + ], + "type": "inline_equation", + "content": "^{3\\dagger}" + }, + { + "bbox": [ + 79, + 142, + 512, + 171 + ], + "type": "text", + "content": ", Yutong Lin" + }, + { + "bbox": [ + 79, + 142, + 512, + 171 + ], + "type": "inline_equation", + "content": "^{23}" + }, + { + "bbox": [ + 79, + 142, + 512, + 171 + ], + "type": "text", + "content": ", Yixuan Wei" + }, + { + "bbox": [ + 79, + 142, + 512, + 171 + ], + "type": "inline_equation", + "content": "^{13}" + }, + { + "bbox": [ + 79, + 142, + 512, + 171 + ], + "type": "text", + "content": ", Qi Dai" + }, + { + "bbox": [ + 79, + 142, + 512, + 171 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 79, + 142, + 512, + 171 + ], + "type": "text", + "content": ", Han Hu" + }, + { + "bbox": [ + 79, + 142, + 512, + 171 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 79, + 142, + 512, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 79, + 142, + 512, + 171 + ], + "type": "text", + "content": "Tsinghua University " + }, + { + "bbox": [ + 79, + 142, + 512, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 79, + 142, + 512, + 171 + ], + "type": "text", + "content": "Xi'an Jiaotong University " + }, + { + "bbox": [ + 79, + 142, + 512, + 171 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 79, + 142, + 512, + 171 + ], + "type": "text", + "content": "Microsoft Research Asia" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 85, + 174, + 507, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 174, + 507, + 185 + ], + "spans": [ + { + "bbox": [ + 85, + 174, + 507, + 185 + ], + "type": "text", + "content": "{t-zhxie, zhez, yuecao, t-yutonglin, t-yixuanwei, qi.dai, hanhu}@microsoft.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 192, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 238, + 289, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 238, + 289, + 585 + ], + "spans": [ + { + "bbox": [ + 46, + 238, + 289, + 585 + ], + "type": "text", + "content": "Scaling properties have been one of the central issues in self-supervised pre-training, especially the data scalability, which has successfully motivated the large-scale self-supervised pre-trained language models and endowed them with significant modeling capabilities. However, scaling properties seem to be unintentionally neglected in the recent trending studies on masked image modeling (MIM), and some arguments even suggest that MIM cannot benefit from large-scale data. In this work, we try to break down these preconceptions and systematically study the scaling behaviors of MIM through extensive experiments, with data ranging from " + }, + { + "bbox": [ + 46, + 238, + 289, + 585 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 238, + 289, + 585 + ], + "type": "text", + "content": " of ImageNet-1K to full ImageNet22K, model parameters ranging from 49-million to one-billion, and training length ranging from 125K to 500K iterations. And our main findings can be summarized in two folds: 1) masked image modeling remains demanding large-scale data in order to scale up computes and model parameters; 2) masked image modeling cannot benefit from more data under a non-overfitting scenario, which diverges from the previous observations in self-supervised pre-trained language models or supervised pre-trained vision models. In addition, we reveal several intriguing properties in MIM, such as high sample efficiency in large MIM models and strong correlation between pre-training validation loss and transfer performance. We hope that our findings could deepen the understanding of masked image modeling and facilitate future developments on large-scale vision models. Code and models will be available at https://github.com/microsoft/SimMIM." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 607, + 127, + 619 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 607, + 127, + 619 + ], + "spans": [ + { + "bbox": [ + 47, + 607, + 127, + 619 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 626, + 288, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 626, + 288, + 687 + ], + "spans": [ + { + "bbox": [ + 46, + 626, + 288, + 687 + ], + "type": "text", + "content": "Masked Image Modeling (MIM) [3, 18, 44], which has recently emerged in the field of self-supervised visual pretraining, has attracted widespread interest and extensive applications throughout the community for unleashing the superior modeling capacity of attention-based Transformer" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 213, + 547, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 213, + 547, + 381 + ], + "spans": [ + { + "bbox": [ + 306, + 213, + 547, + 381 + ], + "type": "text", + "content": "architectures [13, 26] and demonstrating excellent sample efficiency and impressive transfer performance on a variety of vision tasks. However, most recent practices are focused on the design of MIM methods, while the study of scaling properties of MIM is unintentionally neglected, especially the data scaling property, which successfully motivated the large-scale self-supervised pre-trained language models and endowed them with significant modeling capabilities. Although previous works [8, 10, 17, 37, 45] have explored several conclusions about the scaling properties of vision models, most of their findings were obtained under a supervised pre-training scheme or under a contrastive learning framework, so the extent to which these findings could be transferred to MIM still needs to be investigated." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 385, + 547, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 385, + 547, + 613 + ], + "spans": [ + { + "bbox": [ + 305, + 385, + 547, + 613 + ], + "type": "text", + "content": "Meanwhile, with the emergence of Transformers [41] and masked language modeling (MLM) [12, 31], the systematic studies of scaling laws have already been explored in natural language processing field [21, 23, 35], which provided ample guidance for large models in recent years. The core finding drawn from scaling laws [21] for neural language models is that the performance has a power-law relationship with each of the three scale factors - model parameters " + }, + { + "bbox": [ + 305, + 385, + 547, + 613 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 305, + 385, + 547, + 613 + ], + "type": "text", + "content": ", size of dataset " + }, + { + "bbox": [ + 305, + 385, + 547, + 613 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 305, + 385, + 547, + 613 + ], + "type": "text", + "content": ", and amount of compute " + }, + { + "bbox": [ + 305, + 385, + 547, + 613 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 305, + 385, + 547, + 613 + ], + "type": "text", + "content": " respectively - when not bottlenecked by the other two. This conclusion implies that better performance can be obtained by scaling up these three factors to the extent that the scaling laws are in effect, which led to the subsequent developments of large scale language models [16, 28, 29, 32, 33] that exhibit excellent modeling capabilities [4] on most language tasks. Therefore, it is natural to ask whether MIM possesses the same scaling signatures for vision models as the MLM method for language models, so that the scaling up of vision models can catch up with language models." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 617, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 617, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 617, + 547, + 713 + ], + "type": "text", + "content": "Though masked image modeling and masked language modeling both belong to masked signal prediction, their property differences are also non-negligible due to the different nature of vision and language. That is, the images are highly redundant raw signals and words/sentences are semantically rich tokens, which may result in different abilities of data utilization, and it is thus debatable whether the observations from language models could be reproduced in vision models." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 693, + 288, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 693, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 693, + 288, + 713 + ], + "type": "text", + "content": "The work is done when Zhenda Xie, Yutong Lin, and Yixuan Wei are interns at Microsoft Research Asia. " + }, + { + "bbox": [ + 46, + 693, + 288, + 713 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 46, + 693, + 288, + 713 + ], + "type": "text", + "content": " Project co-leaders." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10365" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 98, + 71, + 236, + 178 + ], + "blocks": [ + { + "bbox": [ + 98, + 71, + 236, + 178 + ], + "lines": [ + { + "bbox": [ + 98, + 71, + 236, + 178 + ], + "spans": [ + { + "bbox": [ + 98, + 71, + 236, + 178 + ], + "type": "image", + "image_path": "816aa38777e260b912428594c678e4cfda475d1e0447938cfaa211f93f9383f5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 117, + 179, + 235, + 189 + ], + "lines": [ + { + "bbox": [ + 117, + 179, + 235, + 189 + ], + "spans": [ + { + "bbox": [ + 117, + 179, + 235, + 189 + ], + "type": "text", + "content": "Model Flops " + }, + { + "bbox": [ + 117, + 179, + 235, + 189 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 117, + 179, + 235, + 189 + ], + "type": "text", + "content": " Training Iterations" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 237, + 72, + 365, + 171 + ], + "blocks": [ + { + "bbox": [ + 237, + 72, + 365, + 171 + ], + "lines": [ + { + "bbox": [ + 237, + 72, + 365, + 171 + ], + "spans": [ + { + "bbox": [ + 237, + 72, + 365, + 171 + ], + "type": "image", + "image_path": "35bc9fb8857d9d4776c448afc9cffa445f8e3cc5f41903bd9d940b0b7dd520a1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 274, + 171, + 339, + 180 + ], + "lines": [ + { + "bbox": [ + 274, + 171, + 339, + 180 + ], + "spans": [ + { + "bbox": [ + 274, + 171, + 339, + 180 + ], + "type": "text", + "content": "Dataset Size [D]" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 274, + 180, + 339, + 189 + ], + "lines": [ + { + "bbox": [ + 274, + 180, + 339, + 189 + ], + "spans": [ + { + "bbox": [ + 274, + 180, + 339, + 189 + ], + "type": "text", + "content": "Millions of Images" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 202, + 547, + 237 + ], + "lines": [ + { + "bbox": [ + 46, + 202, + 547, + 237 + ], + "spans": [ + { + "bbox": [ + 46, + 202, + 547, + 237 + ], + "type": "text", + "content": "Figure 1. The curves of validation loss of pre-training models w.r.t. the relative compute, dataset size and model size. MIM performance improves smoothly as we increase the relative compute and model size, but not improve when the dataset size is sufficient to prevent model from overfitting. We set the relative compute of SwinV2-S for 125K iterations as the value of 1. Best viewed in color." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 368, + 73, + 494, + 171 + ], + "blocks": [ + { + "bbox": [ + 368, + 73, + 494, + 171 + ], + "lines": [ + { + "bbox": [ + 368, + 73, + 494, + 171 + ], + "spans": [ + { + "bbox": [ + 368, + 73, + 494, + 171 + ], + "type": "image", + "image_path": "540236580d56040df83fa97a68c9efde933d6d82e9c1d81cd63c544cfecd44c6.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 406, + 171, + 462, + 180 + ], + "lines": [ + { + "bbox": [ + 406, + 171, + 462, + 180 + ], + "spans": [ + { + "bbox": [ + 406, + 171, + 462, + 180 + ], + "type": "text", + "content": "Model Size [N]" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 406, + 180, + 471, + 188 + ], + "lines": [ + { + "bbox": [ + 406, + 180, + 471, + 188 + ], + "spans": [ + { + "bbox": [ + 406, + 180, + 471, + 188 + ], + "type": "text", + "content": "Millions of Params" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 247, + 288, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 247, + 288, + 307 + ], + "spans": [ + { + "bbox": [ + 46, + 247, + 288, + 307 + ], + "type": "text", + "content": "Moreover, recent studies [14, 38] have shown that using a small amount of training data in masked image modeling can achieve comparable performance to using large datasets, which also motivates our explorations as it disagrees with previous findings and intuitions." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 311, + 288, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 311, + 288, + 502 + ], + "spans": [ + { + "bbox": [ + 46, + 311, + 288, + 502 + ], + "type": "text", + "content": "In this paper, we systematically investigate the scaling properties, especially the data scaling capability of masked image modeling in terms of different dataset sizes ranging from " + }, + { + "bbox": [ + 46, + 311, + 288, + 502 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 311, + 288, + 502 + ], + "type": "text", + "content": " of ImageNet (" + }, + { + "bbox": [ + 46, + 311, + 288, + 502 + ], + "type": "inline_equation", + "content": "\\sim 0.1" + }, + { + "bbox": [ + 46, + 311, + 288, + 502 + ], + "type": "text", + "content": " million) to full ImageNet-22K (" + }, + { + "bbox": [ + 46, + 311, + 288, + 502 + ], + "type": "inline_equation", + "content": "\\sim 14" + }, + { + "bbox": [ + 46, + 311, + 288, + 502 + ], + "type": "text", + "content": " million), model sizes ranging from 49-million Swin-V2 Small to 1-billion Swin-V2 giant and training lengths ranging from 125K to 500K iterations. We use Swin Transformer V2 [25] as the vision encoder for its proven trainability of large models and applicability to a wide range of vision tasks, and adopt SimMIM [44] for masked image modeling pre-training because it has no restrictions on encoder architectures. We also conduct experiments with other MIM frameworks like MAE [18] and other vision encoder like the widely used ViT [13] to verify the generalizability of our findings. With these experimental setups, our main findings could be summarized into two folds:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 506, + 288, + 698 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 506, + 288, + 698 + ], + "spans": [ + { + "bbox": [ + 46, + 506, + 288, + 698 + ], + "type": "text", + "content": "i) Masked image modeling remains demanding large-scale data in order to scale up computes and model parameters. We empirically find that MIM performance has a power-law relationship with relative compute and model size when not bottlenecked by the dataset size (Figure 1). Besides, we observe that smaller datasets lead to severe overfitting phenomenon for training large models (Figure 2), and the size of the dataset to prevent model from overfitting increases clearly as the model increases (Figure 5-Left). Therefore, from the perspective of scaling up model, MIM still demands large-scale data. Furthermore, if we train large-scale models of different lengths, we find that relatively small datasets are adequate at shorter training lengths, but still suffer from overfitting at longer training lengths (Figure 5-Right), which further demonstrate the data scalability of MIM for scaling up compute." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 59, + 701, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 701, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 59, + 701, + 288, + 714 + ], + "type": "text", + "content": "2) Masked image modeling cannot benefit from more data" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 247, + 547, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 247, + 547, + 355 + ], + "spans": [ + { + "bbox": [ + 304, + 247, + 547, + 355 + ], + "type": "text", + "content": "under a non-overfitting scenario, which diverges from the previous observations in self-supervised pre-train language models or supervised pre-trained vision models. In MIM, we find that increasing the number of unique samples for a non-overfitting model does not provide additional benefits to performance (Figure 1). This behavior differs from previous observations in supervised vision transformers and self-supervised language models, where the model performance increases as the number of unique samples increases." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 356, + 547, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 356, + 547, + 451 + ], + "spans": [ + { + "bbox": [ + 304, + 356, + 547, + 451 + ], + "type": "text", + "content": "In addition, we also demonstrate some intriguing properties about masked image modeling, such as larger models possessing higher sample efficiency, i.e., fewer optimization steps are required for larger models to achieve same performance (Section 4.2); and the consistency between transfer and test performance, i.e., test performance could be used to indicate the results on downstream tasks (Section 4.3). These observations also correspond to those in previous practices." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 452, + 547, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 452, + 547, + 524 + ], + "spans": [ + { + "bbox": [ + 304, + 452, + 547, + 524 + ], + "type": "text", + "content": "These findings on the one hand confirm the effect of MIM on scaling up model size and compute, and raise new concerns and challenges on the data scalability of MIM on the other. We hope that our findings could deepen the understanding of masked image modeling and facilitate future developments on large-scale vision models." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 536, + 392, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 536, + 392, + 550 + ], + "spans": [ + { + "bbox": [ + 306, + 536, + 392, + 550 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 557, + 547, + 713 + ], + "type": "text", + "content": "Masked Image Modeling Masked Image Modeling learns representations by reconstructing the masked content of images, and its early exploration can be traced back to context encoder [30] and denoising autoencoder [42]. Recently, iGPT [6], BEiT [3], MAE [18] and SimMIM [44] recall this approach on training vision transformer. iGPT [6] sequentially predicted the pixels by auto-regressive manner. BEiT [3] proposed to predict the discrete visual tokens. MAE [18] and SimMIM [44] concurrently found predicting the raw pixels with a high masking ratio could work well. In this work, we use SimMIM as the default masked image modeling approach, because of its simplicity and no restrictions on the architecture of vision encoder like MAE." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10366" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 98, + 71, + 497, + 194 + ], + "blocks": [ + { + "bbox": [ + 98, + 71, + 497, + 194 + ], + "lines": [ + { + "bbox": [ + 98, + 71, + 497, + 194 + ], + "spans": [ + { + "bbox": [ + 98, + 71, + 497, + 194 + ], + "type": "image", + "image_path": "2dcd36573cf2bed72f60350d216b6eab723dc27a8a33f9f5c846bf8ae71164d3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 201, + 547, + 236 + ], + "lines": [ + { + "bbox": [ + 46, + 201, + 547, + 236 + ], + "spans": [ + { + "bbox": [ + 46, + 201, + 547, + 236 + ], + "type": "text", + "content": "Figure 2. Training loss of pre-training, validation loss of pre-training, and fine-tuning accuracy on ImageNet-1K of different model sizes, data sizes and training lengths, w.r.t. the relative pre-training compute. All models are SwinV2 models pre-trained with SimMIM. We set the training compute of SwinV2-S for 125K iterations as the value of 1. Bigger circles indicate larger models. Best viewed in color." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 246, + 289, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 246, + 289, + 449 + ], + "spans": [ + { + "bbox": [ + 46, + 246, + 289, + 449 + ], + "type": "text", + "content": "Vision Transformers Transformer [41] was first applied to natural language processing and became the dominant architecture, and has recently attracted a lot of attention in computer vision. The pioneering work ViT [13] first shows that the transformer architecture works well in image classification when trained on large amounts of data. DeiT [39] proposed a better training recipe based on ViT and demonstrated that vision Transformer has promising performance when only using ImageNet-1K dataset. Swin Transformer [26] improves plain ViT by inducing the hierarchical architecture and non-overlapping local attention and successfully demonstrates the effectiveness of vision transformer on a wide range of vision tasks. Swin Transformer V2 [25] further addresses the training stability issue of [26] in model scaling and illustrates better performance than the original Swin Transformer, and thus we use it as the default vision encoder in this work." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 455, + 288, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 455, + 288, + 670 + ], + "spans": [ + { + "bbox": [ + 46, + 455, + 288, + 670 + ], + "type": "text", + "content": "Scaling Vision Models Many works [25, 34, 37, 45] examine how to scale vision models, but most are more concerned with exploring the perspective of model architecture designs. For example, EfficientNet [37] extensively studied how model width, model depth and input resolution affect the convolutional neural networks; [34] proposed to scale vision model with sparse mixture-of-expert; [45] and [25] studied how to scale ViT and Swin Transformer, respectively. In addition, several studies [1, 2, 45] explored the aspect of data scaling under the pre-training fine-tuning paradigm. BiT [22] revisited the supervised pre-training on a wide range of data scales up to 1M images. SEER [17] studied the effectiveness of data scaling in the contrastive learning framework with up to one billion images. Recently, SplitMask [14] find that masked image modeling is robust to the size of pre-training data and challenges the data scaling capability of masked image modeling, which is most relevant to our work." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 677, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 289, + 715 + ], + "type": "text", + "content": "Scaling Language Models Many works [20, 23, 35] systematically analyzed the scaling behavior on a relatively small scale. And [21] studied empirical scaling laws for" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 246, + 547, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 246, + 547, + 354 + ], + "spans": [ + { + "bbox": [ + 304, + 246, + 547, + 354 + ], + "type": "text", + "content": "large scale neural language models and revealed that the loss scaled as a power-law with model size, dataset size, and the amount of compute used for training. Motivated by these laws, several large scale language models [4,16,28,29,32,33] were successfully trained and demonstrated excellent modeling capabilities on most language tasks. [36] established a comprehensive, large-scale benchmark designed to assist in quantifying and extrapolating the capabilities of large language models." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 363, + 512, + 377 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 363, + 512, + 377 + ], + "spans": [ + { + "bbox": [ + 305, + 363, + 512, + 377 + ], + "type": "text", + "content": "3. Background and Experimental Setup" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 383, + 444, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 383, + 444, + 396 + ], + "spans": [ + { + "bbox": [ + 306, + 383, + 444, + 396 + ], + "type": "text", + "content": "3.1. Masked Image Modeling" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 401, + 547, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 401, + 547, + 628 + ], + "spans": [ + { + "bbox": [ + 304, + 401, + 547, + 628 + ], + "type": "text", + "content": "Masked image modeling is used to train the vision model by taking a corrupted image as input and predicting the content of the masked region as the target. In this study, we use SimMIM [44] as the default masked image modeling approach because of its simplicity and lack of restrictions on the architecture of the vision encoder. SimMIM consists of a visual encoder and an extremely lightweight prediction head of a linear layer for predicting the raw pixels of the corrupted images via " + }, + { + "bbox": [ + 304, + 401, + 547, + 628 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 304, + 401, + 547, + 628 + ], + "type": "text", + "content": " regression loss. To facilitate the implementation of the vision transformer, SimMIM adopts the patch-wise mask strategy with the masked patch size of " + }, + { + "bbox": [ + 304, + 401, + 547, + 628 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 304, + 401, + 547, + 628 + ], + "type": "text", + "content": " and mask ratio of 0.6. To further alleviate the local dependency of raw pixels, we improved the SimMIM by normalizing the predicted target according to [15] with a sliding window of " + }, + { + "bbox": [ + 304, + 401, + 547, + 628 + ], + "type": "inline_equation", + "content": "47^2" + }, + { + "bbox": [ + 304, + 401, + 547, + 628 + ], + "type": "text", + "content": ". As the result, a slight performance improvement is observed. In addition, we conduct experiments using MAE as the masked image modeling approach to verify the methodological generalizability of our findings. These experiments strictly follow the settings in [18]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 635, + 455, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 635, + 455, + 647 + ], + "spans": [ + { + "bbox": [ + 306, + 635, + 455, + 647 + ], + "type": "text", + "content": "3.2. Architecture Specifications" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "content": "We use Swin Transformer V2 [25] as the default vision encoder in this study. Thanks to its generality and scalability, we evaluate a series of SwinV2 models with a wide range of model sizes (the number of parameters ranges from " + }, + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\sim 50\\mathrm{M}" + }, + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\sim 1\\mathrm{B}" + }, + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "content": ", and FLOPs range from " + }, + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\sim 9\\mathrm{G}" + }, + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\sim 190\\mathrm{G}" + }, + { + "bbox": [ + 304, + 654, + 547, + 713 + ], + "type": "text", + "content": ") on" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10367" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 115, + 71, + 479, + 153 + ], + "blocks": [ + { + "bbox": [ + 115, + 71, + 479, + 153 + ], + "lines": [ + { + "bbox": [ + 115, + 71, + 479, + 153 + ], + "spans": [ + { + "bbox": [ + 115, + 71, + 479, + 153 + ], + "type": "table", + "html": "
ModelBase ChannelDepthHeadWindow SizeBackbone Params
pre-trainfine-tune
SwinV2-S96{2, 2, 18, 2}{3, 6, 12, 24}121449M
SwinV2-B128{2, 2, 18, 2}{4, 8, 16, 32}121487M
SwinV2-L192{2, 2, 18, 2}{6, 12, 24, 48}1214195M
SwinV2-H352{2, 2, 18, 2}{11, 22, 44, 88}1214655M
SwinV2-g448{2, 2, 18, 2}{14, 28, 56, 112}12141061M
", + "image_path": "dc85bc388912c22b2e844d6226f28208d54fecd13aebbde9e0c5ee10b85fc974.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 98, + 185, + 496, + 224 + ], + "blocks": [ + { + "bbox": [ + 89, + 162, + 502, + 174 + ], + "lines": [ + { + "bbox": [ + 89, + 162, + 502, + 174 + ], + "spans": [ + { + "bbox": [ + 89, + 162, + 502, + 174 + ], + "type": "text", + "content": "Table 1. Detailed architecture specifications. Note that, the model aliases used in our work are different from [25]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 98, + 185, + 496, + 224 + ], + "lines": [ + { + "bbox": [ + 98, + 185, + 496, + 224 + ], + "spans": [ + { + "bbox": [ + 98, + 185, + 496, + 224 + ], + "type": "table", + "html": "
IN1K (10%)IN1K (20%)IN1K (50%)IN100IN1K(100%)IN22K(100%)
#Classes\\( 1 \\times 10^{3} \\)\\( 1 \\times 10^{3} \\)\\( 1 \\times 10^{3} \\)\\( 1 \\times 10^{2} \\)\\( 1 \\times 10^{3} \\)\\( 2.18 \\times 10^{4} \\)
#Images\\( 1.28 \\times 10^{5} \\)\\( 2.56 \\times 10^{5} \\)\\( 6.41 \\times 10^{5} \\)\\( 1.27 \\times 10^{5} \\)\\( 1.28 \\times 10^{6} \\)\\( 1.42 \\times 10^{7} \\)
", + "image_path": "461e373fb847bfc9b643970d9e537a20d19fd9743d1da46fd051150a4ff5cfd1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 130, + 232, + 461, + 243 + ], + "lines": [ + { + "bbox": [ + 130, + 232, + 461, + 243 + ], + "spans": [ + { + "bbox": [ + 130, + 232, + 461, + 243 + ], + "type": "text", + "content": "Table 2. Detailed dataset specifications used in the pre-training of masked image modeling." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 255, + 288, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 255, + 288, + 352 + ], + "spans": [ + { + "bbox": [ + 46, + 255, + 288, + 352 + ], + "type": "text", + "content": "multiple downstream tasks. The detailed model specifications are shown in Table 1. We use a new variant SwinV2-g (giant), with number of parameters between SwinV2-L and the 3-billion-parameter SwinV2-G (Giant) used in [25]. In addition, we conduct experiments with a series of ViT models [13] to prove the architectural generalizability of our findings. Specifically, we use ViT-B/16, ViT-L/16 and ViTH/14 according to the settings from [18]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 358, + 170, + 371 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 358, + 170, + 371 + ], + "spans": [ + { + "bbox": [ + 47, + 358, + 170, + 371 + ], + "type": "text", + "content": "3.3. Pre-training Datasets" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 377, + 288, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 377, + 288, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 377, + 288, + 533 + ], + "type": "text", + "content": "To study the effect of data size on masked image modeling, we build datasets with different sizes. We use the training set of ImageNet-1K and ImageNet-22K as two large-scale datasets, and randomly sample " + }, + { + "bbox": [ + 46, + 377, + 288, + 533 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 46, + 377, + 288, + 533 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 377, + 288, + 533 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 46, + 377, + 288, + 533 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 377, + 288, + 533 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 46, + 377, + 288, + 533 + ], + "type": "text", + "content": " of images in the ImageNet-1K training set as smaller datasets. By default, the images are uniformly sampled from each category. We also consider the sampling strategies could perform differently. To this end, we randomly sample 100 classes from ImageNet-1K as ImageNet-100, and compare it with ImageNet-1K " + }, + { + "bbox": [ + 46, + 377, + 288, + 533 + ], + "type": "inline_equation", + "content": "(10\\%)" + }, + { + "bbox": [ + 46, + 377, + 288, + 533 + ], + "type": "text", + "content": " but find their training loss and fine-tuning performance are almost the same. The details and statistics of all pre-training datasets used in our study are shown in Table 2." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 540, + 163, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 540, + 163, + 552 + ], + "spans": [ + { + "bbox": [ + 47, + 540, + 163, + 552 + ], + "type": "text", + "content": "3.4. Pre-training Details" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "content": "To better compare the performance of models with different amounts of data under the same pre-training length, we use training iterations rather than training epochs and adopt the same hyper-parameters for all models with different sizes during pre-training. The total number of training iterations is in " + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\{125\\mathrm{K}, 250\\mathrm{K}, 500\\mathrm{K}\\}" + }, + { + "bbox": [ + 46, + 558, + 288, + 715 + ], + "type": "text", + "content": " and the batch size is set as 2048 for all experiments. In pre-training stage, we use the same hyper-parameters for all models, and the training details and hyper-parameters of pre-training are summarized in Appendix. Because of the excessive amount of experiments, we follow SimMIM [44] and also use the following two techniques for reducing the experimental overheads: First, we use the step learning rate scheduler in pre-training for" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 255, + 547, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 255, + 547, + 412 + ], + "spans": [ + { + "bbox": [ + 304, + 255, + 547, + 412 + ], + "type": "text", + "content": "sharing the first training step among experiments with different training lengths. The first 7/8 training iterations are the first step and the last 1/8 training iterations are the second step with the learning rate ratio of 0.1 (i.e. learning rate is divided by 10 in the second step). Second, we adopt the input image size of " + }, + { + "bbox": [ + 304, + 255, + 547, + 412 + ], + "type": "inline_equation", + "content": "192^2" + }, + { + "bbox": [ + 304, + 255, + 547, + 412 + ], + "type": "text", + "content": " and set the window size of 12. We improve the SimMIM by normalizing the predicted target according to [15] with a sliding window of " + }, + { + "bbox": [ + 304, + 255, + 547, + 412 + ], + "type": "inline_equation", + "content": "47^2" + }, + { + "bbox": [ + 304, + 255, + 547, + 412 + ], + "type": "text", + "content": " and observe an improvement of 0.3 on top-1 accuracy of ImageNet-1K for the SwinV2-Large model. The same light data augmentation strategy as SimMIM is used: random resize cropping with a scale range of [0.67, 1], an aspect ratio range of [3/4, 4/3] and a random flipping with probability 0.5." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 421, + 411, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 421, + 411, + 434 + ], + "spans": [ + { + "bbox": [ + 306, + 421, + 411, + 434 + ], + "type": "text", + "content": "3.5. Fine-tuning Tasks" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 441, + 547, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 441, + 547, + 525 + ], + "spans": [ + { + "bbox": [ + 304, + 441, + 547, + 525 + ], + "type": "text", + "content": "To extensively and accurately evaluate the performance of pre-trained models under different pre-training schedulers and datasets, a series of diverse and representative tasks including fine-tuning on ImageNet-1K, fine-grained image classification, object detection, instance segmentation, and semantic segmentation are selected for evaluation. Detailed setups of all experiments are illustrated in Appendix." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 534, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 547, + 714 + ], + "type": "text", + "content": "ImageNet-1K We follow [3] to evaluate the quality of learnt representations by fine-tuning the pre-trained models on ImageNet-1K [11] image classification task, which is the most commonly used scenario and evaluation criterion for pre-trained models [18,44]. Different from pre-training, We adopt the image size with " + }, + { + "bbox": [ + 304, + 534, + 547, + 714 + ], + "type": "inline_equation", + "content": "224^2" + }, + { + "bbox": [ + 304, + 534, + 547, + 714 + ], + "type": "text", + "content": " with window size of 14 in fine-tuning. The AdamW with batch size of 2048, base learning rate of 5e-3, weight decay of 0.05, " + }, + { + "bbox": [ + 304, + 534, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\beta_{1}" + }, + { + "bbox": [ + 304, + 534, + 547, + 714 + ], + "type": "text", + "content": " of 0.9 and " + }, + { + "bbox": [ + 304, + 534, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\beta_{2}" + }, + { + "bbox": [ + 304, + 534, + 547, + 714 + ], + "type": "text", + "content": " of 0.999 are used, and we adopt cosine learning rate scheduler. As larger models are more prone to overfitting, we fine-tune SwinV2-S/B/L for 100 epochs with 20 warm-up epochs and SwinV2-H/g for 50 epochs with 10 warm-up epochs, and decrease the layer decay as the model size increases. In addition, gradient clipping, stochastic depth, label smoothing and data augmentations (e.g. random crop," + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10368" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 97, + 72, + 497, + 185 + ], + "blocks": [ + { + "bbox": [ + 97, + 72, + 497, + 185 + ], + "lines": [ + { + "bbox": [ + 97, + 72, + 497, + 185 + ], + "spans": [ + { + "bbox": [ + 97, + 72, + 497, + 185 + ], + "type": "image", + "image_path": "f7af21374200d9e8f70bb2f5e77f76ea22f09720188570767effe034a4bedea8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 194, + 547, + 229 + ], + "lines": [ + { + "bbox": [ + 46, + 194, + 547, + 229 + ], + "spans": [ + { + "bbox": [ + 46, + 194, + 547, + 229 + ], + "type": "text", + "content": "Figure 3. Left to right: Performances with SwinV2 models on (a) COCO object detection, (b) COCO instance segmentation, (c) iNaturalist-18, and (d) ADE20K semantic segmentation w.r.t. the relative compute. Note that the relative compute indicates the pre-training compute. We set the relative compute of SwinV2-S for 125K iterations as 1. Bigger circles indicate larger models. Best viewed in color." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 239, + 288, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 239, + 288, + 264 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 288, + 264 + ], + "type": "text", + "content": "rand erasing [47], rand augment [9], mixup [46], cutmix [46], etc.) are also used by following [44]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 267, + 290, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 267, + 290, + 411 + ], + "spans": [ + { + "bbox": [ + 46, + 267, + 290, + 411 + ], + "type": "text", + "content": "iNaturalist-18 iNaturalist [40] 2018 is a long-tailed fine-grained image classification dataset. As fine-tuning in ImageNet-1K, we also use the input image size of " + }, + { + "bbox": [ + 46, + 267, + 290, + 411 + ], + "type": "inline_equation", + "content": "224^2" + }, + { + "bbox": [ + 46, + 267, + 290, + 411 + ], + "type": "text", + "content": ", window size of 14 and patch size of 4 in iNaturalist 2018. We fine-tune all models for 100 epochs with 20 warm-up epochs, and set layer decay to 0.8, 0.75 and 0.7 for SwinV2-S/B/L, respectively. The AdamW optimizer with cosine learning rate scheduler, batch size of 2048, base learning rate of 1.6e-2, weight decay of 0.1, " + }, + { + "bbox": [ + 46, + 267, + 290, + 411 + ], + "type": "inline_equation", + "content": "\\beta_{1}" + }, + { + "bbox": [ + 46, + 267, + 290, + 411 + ], + "type": "text", + "content": " of 0.9 and " + }, + { + "bbox": [ + 46, + 267, + 290, + 411 + ], + "type": "inline_equation", + "content": "\\beta_{2}" + }, + { + "bbox": [ + 46, + 267, + 290, + 411 + ], + "type": "text", + "content": " of 0.999 are used. In addition, we also adopt stochastic depth, label smoothing, gradient clipping and data augmentations in fine-tuning." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 415, + 287, + 427 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 415, + 287, + 427 + ], + "spans": [ + { + "bbox": [ + 46, + 415, + 287, + 427 + ], + "type": "text", + "content": "COCO Object Detection and Instance Segmentation [24]" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 427, + 288, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 427, + 288, + 559 + ], + "spans": [ + { + "bbox": [ + 46, + 427, + 288, + 559 + ], + "type": "text", + "content": "We use Mask R-CNN [19] for evaluation. We set the window size to 14 and patch size to 4. The AdamW optimizer with batch size of 32, base learning rate of 8e-5, weight decay of 0.05, " + }, + { + "bbox": [ + 46, + 427, + 288, + 559 + ], + "type": "inline_equation", + "content": "\\beta_{1}" + }, + { + "bbox": [ + 46, + 427, + 288, + 559 + ], + "type": "text", + "content": " of 0.9, " + }, + { + "bbox": [ + 46, + 427, + 288, + 559 + ], + "type": "inline_equation", + "content": "\\beta_{2}" + }, + { + "bbox": [ + 46, + 427, + 288, + 559 + ], + "type": "text", + "content": " of 0.999 and a step learning rate scheduler (step learning rate ratio of 0.1, step epochs are 27 and 33) are used. In training, the random cropping with crop size of [1024, 1024], large scale jittering with a range of [0.1, 2.0], random horizontal flip with probability 0.5, and stochastic depth regularization are used. In testing, all images are resized to (800, 1333) and keeping the aspect ratio unchanged." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 563, + 288, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 563, + 288, + 696 + ], + "spans": [ + { + "bbox": [ + 46, + 563, + 288, + 696 + ], + "type": "text", + "content": "ADE20K Semantic Segmentation [48] Following [26], we use UPerNet [43] for evaluation. We set the window size to 20 and the patch size to 4. The AdamW optimizer with batch size of 32, base learning rate searched in a range of [1e-4, 3e-4], weight decay of 0.05, " + }, + { + "bbox": [ + 46, + 563, + 288, + 696 + ], + "type": "inline_equation", + "content": "\\beta_{1}" + }, + { + "bbox": [ + 46, + 563, + 288, + 696 + ], + "type": "text", + "content": " of 0.9, " + }, + { + "bbox": [ + 46, + 563, + 288, + 696 + ], + "type": "inline_equation", + "content": "\\beta_{2}" + }, + { + "bbox": [ + 46, + 563, + 288, + 696 + ], + "type": "text", + "content": " of 0.999 and a linear learning rate scheduler with a total of 80K iterations are used. Also, we use the layer decay of 0.95, 0.95, 0.9 for SwinV2-S/B/L, respectively. In training, the random cropping with crop size of [640, 640], scale jittering with a range of [0.5, 2.0], random horizontal flip with probability 0.5, random photometric distortion and stochastic" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 239, + 547, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 239, + 547, + 288 + ], + "spans": [ + { + "bbox": [ + 306, + 239, + 547, + 288 + ], + "type": "text", + "content": "depth regularization of 0.1 are used. In testing, all images are evaluated by sliding window manner, and use the test image size of (2560, 640) and set sliding window stride to 426, following [26, 44]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 299, + 429, + 314 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 299, + 429, + 314 + ], + "spans": [ + { + "bbox": [ + 306, + 299, + 429, + 314 + ], + "type": "text", + "content": "4. Results and Findings" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 320, + 547, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 320, + 547, + 429 + ], + "spans": [ + { + "bbox": [ + 305, + 320, + 547, + 429 + ], + "type": "text", + "content": "We train numerous models with different training lengths, dataset sizes, and model sizes, and study how these factors affect the performance of masked image modeling. Figure 1 illustrates the validation loss of pre-training" + }, + { + "bbox": [ + 305, + 320, + 547, + 429 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 305, + 320, + 547, + 429 + ], + "type": "text", + "content": " with respect to the relative compute, dataset size and model size. Figure 2 illustrates the training loss of pre-training, validation loss of pre-training, and the fine-tuning top-1 accuracy of ImageNet-1K with respect to the relative compute. Based on these extensive experiments, we make the following observations:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 437, + 520, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 437, + 520, + 451 + ], + "spans": [ + { + "bbox": [ + 305, + 437, + 520, + 451 + ], + "type": "text", + "content": "4.1. Data Scaling in Masked Image Modeling" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 456, + 547, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 456, + 547, + 685 + ], + "spans": [ + { + "bbox": [ + 305, + 456, + 547, + 685 + ], + "type": "text", + "content": "Masked image modeling remains demanding for large-scale data. When with the high masking rate (e.g., " + }, + { + "bbox": [ + 305, + 456, + 547, + 685 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 305, + 456, + 547, + 685 + ], + "type": "text", + "content": " in our work), the masked image modeling is considered a very challenging training objective and has been found to be data efficient by previous literature [14,27], i.e., a comparable performance can be achieved with small datasets as with large datasets. However, Figure 2 shows that as the training cost increases, the training loss of some models drops significantly, and their validation loss rises significantly, even on using " + }, + { + "bbox": [ + 305, + 456, + 547, + 685 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 305, + 456, + 547, + 685 + ], + "type": "text", + "content": " images of ImageNet-1K (i.e., IN1K " + }, + { + "bbox": [ + 305, + 456, + 547, + 685 + ], + "type": "inline_equation", + "content": "(50\\%)" + }, + { + "bbox": [ + 305, + 456, + 547, + 685 + ], + "type": "text", + "content": "), indicating the overfitting phenomenon exists. In Figure 5-Left, we demonstrate that the size of the dataset to prevent model from overfitting increases clearly as the model increases. And significant decrease to the fine-tuning performance caused by overfitting could be observed in Figure 2 and Figure 3. Moreover, we measure the best fine-tuning performance of each model trained by different training schedulers in Figure 2-Right. We find the large models perform even worse than smaller models when small dataset is used for training. For" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 306, + 693, + 547, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 693, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 693, + 547, + 713 + ], + "type": "text", + "content": "2The validation loss of pre-training is measured on the validation set of ImageNet-1K for all experiments." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 58, + 702, + 220, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 220, + 714 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 220, + 714 + ], + "type": "text", + "content": "1Our implementation based on MMDetection [5]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "10369" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 98, + 71, + 500, + 194 + ], + "blocks": [ + { + "bbox": [ + 98, + 71, + 500, + 194 + ], + "lines": [ + { + "bbox": [ + 98, + 71, + 500, + 194 + ], + "spans": [ + { + "bbox": [ + 98, + 71, + 500, + 194 + ], + "type": "image", + "image_path": "3fafda11fc16c0eb1c8d21d46c0bc528caf84840c45ad6ee96a365b19f2151ce.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 201, + 547, + 236 + ], + "lines": [ + { + "bbox": [ + 46, + 201, + 547, + 236 + ], + "spans": [ + { + "bbox": [ + 46, + 201, + 547, + 236 + ], + "type": "text", + "content": "Figure 4. Training loss of pre-training, validation loss of pre-training, and fine-tuning accuracy on ImageNet-1K of different model sizes, data sizes and training lengths, w.r.t. the relative pre-training compute. All models are ViT models pre-trained with MAE. We set the training compute of ViT-B/16 for 125K iterations as the value of 1. Bigger circles indicate larger models. Best viewed in color." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 246, + 290, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 246, + 290, + 330 + ], + "spans": [ + { + "bbox": [ + 46, + 246, + 290, + 330 + ], + "type": "text", + "content": "example, the best top-1 accuracy of SwinV2-H with IN1K (20%) is 84.4, worse than the best performance of SwinV2-L by 0.3. In addition, by comparing the best performance that can be obtained using different sizes of dataset, we find that using more data results in better performance. These observations suggest that masked image modeling does not alleviate the demands of large dataset." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 336, + 288, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 336, + 288, + 479 + ], + "spans": [ + { + "bbox": [ + 46, + 336, + 288, + 479 + ], + "type": "text", + "content": "The training length matters. Larger models can benefit from more data at a longer training length. By comparing the performance of models pre-trained by different data sizes in Figure 2-Right, we find that the fine-tuning performance of the large models saturates more slowly with the increasing data size compared to the smaller models. For example, the SwinV2-S model pre-trained on IN1K (50%) has a very similar fine-tuning performance to the model pre-trained on IN1K (100%). In comparison, the performance difference between the SwinV2-H model pre-trained on IN1K (50%) and IN1K (100%) is near 0.5, which is a significant gap for ImageNet-1K classification." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 479, + 288, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 479, + 288, + 587 + ], + "spans": [ + { + "bbox": [ + 46, + 479, + 288, + 587 + ], + "type": "text", + "content": "Furthermore, a comprehensive observation reveals that the improvements from using more data are not significant under short training lengths. As shown in Figure 5-Right, while there is a noticeable performance gap between SwinV2-H trained on IN1K (50%) and IN1K (100%) at a training length of 500K iterations, the gap is negligible at a training length of 125K iterations. This observation suggests that increasing the training length for larger models is critical to benefit from more data." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 289, + 715 + ], + "type": "text", + "content": "Masked image modeling cannot benefit from more data under a non-overfitting scenario. As illustrated in Figure 1-Center, we plot the validation loss of pre-trained SwinV2 models with respect to different dataset sizes, and we demonstrate that for a particular model size and training length, there will be a certain dataset size that will keep it from overfitting. And the model cannot benefit from dataset larger than this size. This behavior differs from previous observations in supervised vision transformers [45] and self-supervised language models [21], where the model perfor" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 305, + 246, + 547, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 246, + 547, + 305 + ], + "spans": [ + { + "bbox": [ + 305, + 246, + 547, + 305 + ], + "type": "text", + "content": "mance increases as the number of unique samples increases. We speculate that this may be related to the characteristics of the task itself, i.e., MIM provides a large number of training signals, making the model hard to learn more patterns from large-scale data." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 307, + 315, + 542, + 425 + ], + "blocks": [ + { + "bbox": [ + 307, + 315, + 542, + 425 + ], + "lines": [ + { + "bbox": [ + 307, + 315, + 542, + 425 + ], + "spans": [ + { + "bbox": [ + 307, + 315, + 542, + 425 + ], + "type": "image", + "image_path": "6c5ea36ba70b6898f3f5a6ab888582dda2f4ec0e3dc889ae444ff1d3b64f534e.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 436, + 547, + 502 + ], + "lines": [ + { + "bbox": [ + 304, + 436, + 547, + 502 + ], + "spans": [ + { + "bbox": [ + 304, + 436, + 547, + 502 + ], + "type": "text", + "content": "Figure 5. Left: Validation loss with respect to model size. The size of the dataset to prevent model from overfitting increases clearly as the model increases. Right: Validation loss with respect to training length using SwinV2-Huge. " + }, + { + "bbox": [ + 304, + 436, + 547, + 502 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 304, + 436, + 547, + 502 + ], + "type": "text", + "content": " subset of ImageNet-1K is sufficient for training 125K iterations, but not sufficient to prevent overfitting for 500K iterations. Best viewed in color." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 518, + 548, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 518, + 548, + 673 + ], + "spans": [ + { + "bbox": [ + 304, + 518, + 548, + 673 + ], + "type": "text", + "content": "Evaluation on more tasks. In addition to ImageNet-1K image classification, we also evaluate the MIM pre-trained SwinV2-S, SwinV2-B and SwinV2-L on iNaturalist-18 fine-grained image classification, ADE20K semantic segmentation, and COCO object detection/segmentation. Figure 3 shows a similar pattern with ImageNet-1K (Figure. 2-Right) that as the training cost increases, some models have significantly performance drop. In addition, the smaller models rapidly reach saturation as the amount of data increases, while larger models can continuously benefit from more data after sufficient training. These results suggest that the conclusions drawn on ImageNet-1K are broadly applicable to other vision tasks." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "text", + "content": "Results with MAE and ViT Figure 4 demonstrates the results of ViT pre-trained with MAE. Similar to SwinV2 models pre-trained with SimMIM, we observe the same" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10370" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 145 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 145 + ], + "type": "text", + "content": "overfitting phenomenon when training with small datasets or large models, which makes MAE still demand for large-scale data. Besides, we could find that larger ViT models can also benefit from more data at a longer training lengths. These experiments verify the methodological and architectural generalizability of our results and findings." + } + ] + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 50, + 156, + 287, + 266 + ], + "blocks": [ + { + "bbox": [ + 50, + 156, + 287, + 266 + ], + "lines": [ + { + "bbox": [ + 50, + 156, + 287, + 266 + ], + "spans": [ + { + "bbox": [ + 50, + 156, + 287, + 266 + ], + "type": "image", + "image_path": "1825b07385a5a207ca6589331aedd1a04c57ba8cb6a4b2c3efb4b6722efaf239.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 274, + 289, + 329 + ], + "lines": [ + { + "bbox": [ + 46, + 274, + 289, + 329 + ], + "spans": [ + { + "bbox": [ + 46, + 274, + 289, + 329 + ], + "type": "text", + "content": "Figure 6. Validation loss and fine-tuning error rate on ImageNet-1K with respect to images processed during pre-training. Large models are more sample efficient than small models to achieve lower validation loss or fine-tuning error rate with fewer optimization steps. Best viewed in color." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 342, + 287, + 367 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 342, + 287, + 367 + ], + "spans": [ + { + "bbox": [ + 47, + 342, + 287, + 367 + ], + "type": "text", + "content": "4.2. Larger Models Possess Higher Sample Efficiency" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 373, + 288, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 373, + 288, + 529 + ], + "spans": [ + { + "bbox": [ + 46, + 373, + 288, + 529 + ], + "type": "text", + "content": "Figure 6 shows the validation loss and fine-tuning error rate on ImageNet-1K with respect to the images processed (batch size times number of steps) during pre-training stage. Five Swin-V2 models with different sizes are pre-trained on full-set of ImageNet-1K. From these results, we observe that larger models possess higher sample efficiency, reaching the same level of validation loss or fine-tuning error rate on ImageNet with fewer optimization steps. These results indicate that larger models will continue to have better performance, and training larger models with fewer steps on sufficient data is a preferable choice. This observation is also in line with previous works on self-supervised language models [21] and supervised vision transformers [45]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 539, + 287, + 564 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 539, + 287, + 564 + ], + "spans": [ + { + "bbox": [ + 47, + 539, + 287, + 564 + ], + "type": "text", + "content": "4.3. Correlation between Pre-training Losses and the Fine-tuning Performance" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "content": "Evaluating a pre-trained model by its fine-tuned performance on downstream tasks is costly. In supervised pretraining, the validation accuracy is used as the proxy indicator to evaluate the quality of the pre-trained models. While in previous studies [7] on other self-supervised learning approaches (e.g., contrastive learning), such a proxy indicator is lacking. In this study, we would like to explore whether the pre-training loss in the training of masked image modeling is a good indicator of its fine-tuning performance. We collect all pre-trained models and plot their training and validation loss curves on Figure 7. Interestingly, the correlations between pre-training losses and the fine-tuning performance" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 307, + 71, + 550, + 155 + ], + "blocks": [ + { + "bbox": [ + 307, + 71, + 550, + 155 + ], + "lines": [ + { + "bbox": [ + 307, + 71, + 550, + 155 + ], + "spans": [ + { + "bbox": [ + 307, + 71, + 550, + 155 + ], + "type": "table", + "html": "
TaskOverfitNon-Overfit
Train LossVal LossTrain LossVal Loss
IN-1K+0.26-0.79-0.64-0.90
iNat-18+0.17-0.54-0.46-0.78
COCO OD+0.54-0.81-0.35-0.83
COCO IS+0.62-0.86-0.31-0.85
ADE-20K+0.75-0.91-0.14-0.90
", + "image_path": "fc2e1d8063aaaae345f00c5ea8b26f04251ecebb8dbbe55c771a69d0e2db1185.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 213, + 545, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 213, + 545, + 237 + ], + "spans": [ + { + "bbox": [ + 304, + 213, + 545, + 237 + ], + "type": "text", + "content": "on multiple tasks could be observed with a phase transition around overfitting." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 239, + 545, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 239, + 545, + 312 + ], + "spans": [ + { + "bbox": [ + 304, + 239, + 545, + 312 + ], + "type": "text", + "content": "Specifically, the correlation between training loss and fine-tuning performance is negative for the overfitting model (green circles) and positive for the non-overfitting model (red circles). The correlation between validation loss and fine-tuning performance is always negative, but the slope of their linear fit lines is significantly different." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 314, + 547, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 314, + 547, + 374 + ], + "spans": [ + { + "bbox": [ + 304, + 314, + 547, + 374 + ], + "type": "text", + "content": "In addition, we further analyze the Pearson correlation coefficient between training loss and fine-tuning performance (Table 3), and find the validation loss has stronger correlation with fine-tuned performance than train loss for all cases, especially for non-overfitting models." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 388, + 501, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 388, + 501, + 399 + ], + "spans": [ + { + "bbox": [ + 306, + 388, + 501, + 399 + ], + "type": "text", + "content": "4.4. Effects of Different Sizes of Decoders" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 408, + 545, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 408, + 545, + 552 + ], + "spans": [ + { + "bbox": [ + 304, + 408, + 545, + 552 + ], + "type": "text", + "content": "We have studied the effects of encoder size from the data scaling perspective. Here, the effects of decoder size are further studied. We pre-train SwinV2-B models with decoder heads of different sizes on IN1K (20%), and Table 4 shows the results. Interestingly, although we find that the heavier decoder has lower training loss and higher validation loss than the linear decoder, indicating a more severe overfitting issue. But there is no decrease in its fine-tuning performance on ImageNet-1K than the linear decoder. This experiment shows that the decoder behaves very differently from the encoder, and we speculate that this is because the decoder \"blocks\" the damage to the encoder from overfitting." + } + ] + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 308, + 566, + 544, + 619 + ], + "blocks": [ + { + "bbox": [ + 304, + 164, + 545, + 196 + ], + "lines": [ + { + "bbox": [ + 304, + 164, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 304, + 164, + 545, + 196 + ], + "type": "text", + "content": "Table 3. Pearson correlation coefficients between pre-training losses (training and validation losses) and fine-tuning performances on five downstream tasks." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 308, + 566, + 544, + 619 + ], + "lines": [ + { + "bbox": [ + 308, + 566, + 544, + 619 + ], + "spans": [ + { + "bbox": [ + 308, + 566, + 544, + 619 + ], + "type": "table", + "html": "
Decoder# ParamsTrain LossVal LossTop-1 Acc
linear90.0M0.460.4784.4
4-blocks140.4M0.440.4884.4
8-blocks190.8M0.410.5084.5
", + "image_path": "dad3ed26861f664935c449d7669aca7ff3dff2682ef23b6e1db5eec57e906c41.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 627, + 547, + 671 + ], + "lines": [ + { + "bbox": [ + 304, + 627, + 547, + 671 + ], + "spans": [ + { + "bbox": [ + 304, + 627, + 547, + 671 + ], + "type": "text", + "content": "Table 4. Results of different decoders, including converged training and validation losses of MIM pre-training, and fine-tuning performance (top-1 accuracy) on ImageNet-1K image classification. Encoders for all models are SwinV2-Base." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 315, + 702, + 470, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 702, + 470, + 712 + ], + "spans": [ + { + "bbox": [ + 315, + 702, + 470, + 712 + ], + "type": "text", + "content": "3The least squares method is used for linear fit." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10371" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 98, + 70, + 500, + 230 + ], + "blocks": [ + { + "bbox": [ + 98, + 70, + 500, + 230 + ], + "lines": [ + { + "bbox": [ + 98, + 70, + 500, + 230 + ], + "spans": [ + { + "bbox": [ + 98, + 70, + 500, + 230 + ], + "type": "image", + "image_path": "dd935eddff1fb13ba7c9ab22ea78df540c1495b04b8c3d4259107d66a4d3fb8b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 238, + 547, + 283 + ], + "lines": [ + { + "bbox": [ + 46, + 238, + 547, + 283 + ], + "spans": [ + { + "bbox": [ + 46, + 238, + 547, + 283 + ], + "type": "text", + "content": "Figure 7. The correlations between pre-training losses (training and validation losses) and the fine-tuning performances of (a) ImageNet-1k image classification, (b) iNat 2018 fine-grained classification, (c) COCO object detection, (d) COCO instance segmentation, (e) and ADE-20K semantic segmentation. Pre-training losses are highly correlated with fine-tuning performance on all five tasks. Red circles indicate the overfitting models and green circles indicate non-overfitting models. Best viewed in color." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 54, + 291, + 281, + 332 + ], + "blocks": [ + { + "bbox": [ + 54, + 291, + 281, + 332 + ], + "lines": [ + { + "bbox": [ + 54, + 291, + 281, + 332 + ], + "spans": [ + { + "bbox": [ + 54, + 291, + 281, + 332 + ], + "type": "table", + "html": "
Dataset# ClassesTrain / Val LossTop-1 Acc
IN1K (10%)10000.351 / 0.51583.5
IN1001000.352 / 0.51183.4
", + "image_path": "330a994c98f4e3a44c7cd3cd2c3a8a43ce9088a52f4e8eebbf70682d30a138bf.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 341, + 288, + 397 + ], + "lines": [ + { + "bbox": [ + 46, + 341, + 288, + 397 + ], + "spans": [ + { + "bbox": [ + 46, + 341, + 288, + 397 + ], + "type": "text", + "content": "Table 5. Results on different dataset sampling strategies (ImageNet-1K (10%) and ImageNet-100) with same dataset size " + }, + { + "bbox": [ + 46, + 341, + 288, + 397 + ], + "type": "inline_equation", + "content": "\\left( {{1.28} \\times {10}^{5}}\\right." + }, + { + "bbox": [ + 46, + 341, + 288, + 397 + ], + "type": "text", + "content": " images for both), include converged training and validation losses of MIM pre-training, and fine-tuning performance (top-1 accuracy) on ImageNet-1K image classification." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 409, + 287, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 409, + 287, + 434 + ], + "spans": [ + { + "bbox": [ + 47, + 409, + 287, + 434 + ], + "type": "text", + "content": "4.5. Impact of Different Dataset Sampling Strategies" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 441, + 289, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 441, + 289, + 609 + ], + "spans": [ + { + "bbox": [ + 46, + 441, + 289, + 609 + ], + "type": "text", + "content": "We study different dataset sampling strategies by comparing the training behavior and fine-tuned performance of models pre-trained on IN1K (10%) and IN100. In IN1K (10%), the images are uniformly sampled from each category, and we randomly sample 100 categories from ImageNet-1K as IN100. Experiments are conducted on SwinV2-L with 500K training iterations. Table 5 shows the training loss, validation loss and fine-tuning top-1 accuracy of ImageNet-1K. For the two models pre-trained on IN1K (10%) and IN100, all three metrics are very similar. Figure 8 further illustrates the training dynamics of the two models, and we find both their training loss curves and validation loss curves are almost overlapping. These results show the disparity caused by different dataset sampling strategies is minor." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 620, + 120, + 633 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 620, + 120, + 633 + ], + "spans": [ + { + "bbox": [ + 47, + 620, + 120, + 633 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 288, + 715 + ], + "type": "text", + "content": "In our work, we systematically study the data scaling capability of masked image modeling at different model sizes and training lengths. Based on the extensive experiments, we demonstrate that in order to scale up computes and model parameters, masked image modeling remains demanding for large-scale data due to the severe overfit" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 307, + 292, + 547, + 390 + ], + "blocks": [ + { + "bbox": [ + 307, + 292, + 547, + 390 + ], + "lines": [ + { + "bbox": [ + 307, + 292, + 547, + 390 + ], + "spans": [ + { + "bbox": [ + 307, + 292, + 547, + 390 + ], + "type": "image", + "image_path": "85a9bd2e8ee20d32e2de541ecf81af2ac5f1cac0bd2bdbf5f34ffa3b0f7338f4.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 398, + 546, + 431 + ], + "lines": [ + { + "bbox": [ + 305, + 398, + 546, + 431 + ], + "spans": [ + { + "bbox": [ + 305, + 398, + 546, + 431 + ], + "type": "text", + "content": "Figure 8. The training loss and validation loss of MIM pre-training with different dataset sampling strategies, ImageNet-1K (10%) and ImageNet-100. Best viewed in color." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 450, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 450, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 547, + 715 + ], + "type": "text", + "content": "ting phenomenon when pre-training with small datasets of large models, which challenges the conclusions of previous literatures that a large dataset may not be necessary in masked image modeling. Besides, we also find that masked image modeling cannot benefit from more data under a nonoverfitting scenario, which diverges from the previous observations in large-scale pre-training language or vision models and raise new concerns and challenges on the data scalability of MIM. In addition, some intriguing properties of MIM are observed, such as larger models possessing higher sample efficiency and a strong correlation between the validation loss of masked image modeling and the fine-tuning performance. The former observation indicates that larger models will continue to have better performance and suggests that training larger models with fewer steps on sufficient data is a preferable option, while the latter observation suggests that validation loss can be considered as a good proxy indicator for evaluating pre-trained models, and makes it possible to reduce the experimental overhead of measuring models by fine-tuning. We hope that our findings could deepen the understanding of masked image modeling and facilitate future developments on large-scale vision models." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "10372" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 92, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 92, + 288, + 125 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 92, + 288, + 125 + ], + "spans": [ + { + "bbox": [ + 53, + 92, + 288, + 125 + ], + "type": "text", + "content": "[1] Ibrahim Alabdulmohsin, Behnam Neyshabur, and Xiaohua Zhai. Revisiting neural scaling laws in language and vision. In NeurIPS, 2022." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 127, + 288, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 127, + 288, + 171 + ], + "spans": [ + { + "bbox": [ + 53, + 127, + 288, + 171 + ], + "type": "text", + "content": "[2] Ibrahim M Alabdulmohsin and Mario Lucic. A near-optimal algorithm for debiasing trained machine learning models. Advances in Neural Information Processing Systems, 34:8072-8084, 2021." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 173, + 288, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 173, + 288, + 205 + ], + "spans": [ + { + "bbox": [ + 53, + 173, + 288, + 205 + ], + "type": "text", + "content": "[3] Hangbo Bao, Li Dong, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 209, + 288, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 209, + 288, + 262 + ], + "spans": [ + { + "bbox": [ + 53, + 209, + 288, + 262 + ], + "type": "text", + "content": "[4] Tom B Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 265, + 288, + 309 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 265, + 288, + 309 + ], + "spans": [ + { + "bbox": [ + 53, + 265, + 288, + 309 + ], + "type": "text", + "content": "[5] Kai Chen, Jiaqi Wang, Jiangmiao Pang, Yuhang Cao, Yu Xiong, Xiaoxiao Li, Shuyang Sun, Wansen Feng, Ziwei Liu, Jiarui Xu, et al. Mmdetection: Open mmlab detection toolbox and benchmark. arXiv preprint arXiv:1906.07155, 2019." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 312, + 288, + 344 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 312, + 288, + 344 + ], + "spans": [ + { + "bbox": [ + 53, + 312, + 288, + 344 + ], + "type": "text", + "content": "[6] Mark Chen, Alec Radford, Rewon Child, Jeff Wu, and Heewoo Jun. Generative pretraining from pixels. Advances in Neural Information Processing Systems, 2020." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 346, + 288, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 346, + 288, + 380 + ], + "spans": [ + { + "bbox": [ + 53, + 346, + 288, + 380 + ], + "type": "text", + "content": "[7] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. ICML, 2020." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 381, + 288, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 381, + 288, + 415 + ], + "spans": [ + { + "bbox": [ + 53, + 381, + 288, + 415 + ], + "type": "text", + "content": "[8] Ting Chen, Simon Kornblith, Kevin Swersky, Mohammad Norouzi, and Geoffrey Hinton. Big self-supervised models are strong semi-supervised learners, 2020." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 417, + 288, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 417, + 288, + 472 + ], + "spans": [ + { + "bbox": [ + 53, + 417, + 288, + 472 + ], + "type": "text", + "content": "[9] Ekin D Cubuk, Barret Zoph, Jonathon Shlens, and Quoc V Le. Randaugment: Practical automated data augmentation with a reduced search space. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 702-703, 2020." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 474, + 288, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 474, + 288, + 506 + ], + "spans": [ + { + "bbox": [ + 48, + 474, + 288, + 506 + ], + "type": "text", + "content": "[10] Zihang Dai, Hanxiao Liu, Quoc V. Le, and Mingxing Tan. Coatnet: Marrying convolution and attention for all data sizes, 2021." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 509, + 288, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 509, + 288, + 542 + ], + "spans": [ + { + "bbox": [ + 48, + 509, + 288, + 542 + ], + "type": "text", + "content": "[11] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248–255. IEEE, 2009." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 544, + 288, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 288, + 588 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 288, + 588 + ], + "type": "text", + "content": "[12] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 590, + 288, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 590, + 288, + 666 + ], + "spans": [ + { + "bbox": [ + 48, + 590, + 288, + 666 + ], + "type": "text", + "content": "[13] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 669, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 288, + 713 + ], + "type": "text", + "content": "[14] Alaaeldin El-Nouby, Gautier Izacard, Hugo Touvron, Ivan Laptev, Hervé Jegou, and Edouard Grave. Are large-scale datasets necessary for self-supervised pre-training? arXiv preprint arXiv:2112.10740, 2021." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[15] Yuxin Fang, Li Dong, Hangbo Bao, Xinggang Wang, and Furu Wei. Corrupted image modeling for self-supervised visual pre-training. arXiv preprint arXiv:2202.03382, 2022." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 107, + 545, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 139 + ], + "type": "text", + "content": "[16] William Fedus, Barret Zoph, and Noam Shazeer. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity, 2021." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 140, + 547, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 140, + 547, + 193 + ], + "spans": [ + { + "bbox": [ + 308, + 140, + 547, + 193 + ], + "type": "text", + "content": "[17] Priya Goyal, Mathilde Caron, Benjamin Lefaudeauux, Min Xu, Pengchao Wang, Vivek Pai, Mannat Singh, Vitaliy Liptchinsky, Ishan Misra, Armand Joulin, et al. Self-supervised pretraining of visual features in the wild. arXiv preprint arXiv:2103.01988, 2021." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 195, + 545, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 195, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 308, + 195, + 545, + 227 + ], + "type": "text", + "content": "[18] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dolkar, and Ross Girshick. Masked autoencoders are scalable vision learners. arXiv preprint arXiv:2111.06377, 2021." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 228, + 545, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 228, + 545, + 250 + ], + "spans": [ + { + "bbox": [ + 308, + 228, + 545, + 250 + ], + "type": "text", + "content": "[19] Kaiming He, Georgia Gkioxari, Piotr Dólár, and Ross Girshick. Mask r-cnn. In ICCV, pages 2961-2969, 2017." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 251, + 545, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 251, + 545, + 303 + ], + "spans": [ + { + "bbox": [ + 308, + 251, + 545, + 303 + ], + "type": "text", + "content": "[20] Joel Hestness, Sharan Narang, Newsha Ardalani, Gregory Diamos, Heewoo Jun, Hassan Kianinejad, Md Patwary, Mostofa Ali, Yang Yang, and Yanqi Zhou. Deep learning scaling is predictable, empirically. arXiv preprint arXiv:1712.00409, 2017." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 304, + 545, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 304, + 545, + 349 + ], + "spans": [ + { + "bbox": [ + 308, + 304, + 545, + 349 + ], + "type": "text", + "content": "[21] Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B. Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models, 2020." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 350, + 545, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 350, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 308, + 350, + 545, + 403 + ], + "type": "text", + "content": "[22] Alexander Kolesnikov, Lucas Beyer, Xiaohua Zhai, Joan Puigcerver, Jessica Yung, Sylvain Gelly, and Neil Houlsby. Big transfer (bit): General visual representation learning. In European conference on computer vision, pages 491-507. Springer, 2020." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 404, + 545, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 404, + 545, + 458 + ], + "spans": [ + { + "bbox": [ + 308, + 404, + 545, + 458 + ], + "type": "text", + "content": "[23] Zhuohan Li, Eric Wallace, Sheng Shen, Kevin Lin, Kurt Keutzer, Dan Klein, and Joey Gonzalez. Train big, then compress: Rethinking model size for efficient training and inference of transformers. In International Conference on Machine Learning, pages 5958-5968. PMLR, 2020." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 460, + 545, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 460, + 545, + 514 + ], + "spans": [ + { + "bbox": [ + 308, + 460, + 545, + 514 + ], + "type": "text", + "content": "[24] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dálár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 514, + 545, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 514, + 545, + 558 + ], + "spans": [ + { + "bbox": [ + 308, + 514, + 545, + 558 + ], + "type": "text", + "content": "[25] Ze Liu, Han Hu, Yutong Lin, Zhuliang Yao, Zhenda Xie, Yixuan Wei, Jia Ning, Yue Cao, Zheng Zhang, Li Dong, et al. Swin transformer v2: Scaling up capacity and resolution. arXiv preprint arXiv:2111.09883, 2021." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 559, + 545, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 559, + 545, + 602 + ], + "spans": [ + { + "bbox": [ + 308, + 559, + 545, + 602 + ], + "type": "text", + "content": "[26] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. arXiv preprint arXiv:2103.14030, 2021." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 603, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 603, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 603, + 545, + 624 + ], + "type": "text", + "content": "[27] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swin transformer, 2021." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 625, + 545, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 625, + 545, + 646 + ], + "spans": [ + { + "bbox": [ + 308, + 625, + 545, + 646 + ], + "type": "text", + "content": "[28] Microsoft. Turing-nlg: A 17-billion-parameter language model by microsoft, 2020." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 647, + 545, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 680 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 680 + ], + "type": "text", + "content": "[29] Microsoft. Using deepspeed and megatron to train megatron-turing nlg 530b, the world's largest and most powerful generative language model, 2021." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "text", + "content": "[30] Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A Efros. Context encoders: Feature learning by inpainting. In Proceedings of the IEEE conference on" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10373" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 94 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 94 + ], + "type": "text", + "content": "computer vision and pattern recognition, pages 2536-2544, 2016." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 95, + 287, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 95, + 287, + 128 + ], + "spans": [ + { + "bbox": [ + 48, + 95, + 287, + 128 + ], + "type": "text", + "content": "[31] Alec Radford, Karthik Narasimhan, Tim Salimans, and Ilya Sutskever. Improving language understanding by generative pre-training. 2018." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 129, + 288, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 288, + 160 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 288, + 160 + ], + "type": "text", + "content": "[32] Alec Radford, Jeff Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language models are unsupervised multitask learners. 2019." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 162, + 288, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 162, + 288, + 216 + ], + "spans": [ + { + "bbox": [ + 48, + 162, + 288, + 216 + ], + "type": "text", + "content": "[33] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research, 21(140):1-67, 2020." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 217, + 287, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 217, + 287, + 261 + ], + "spans": [ + { + "bbox": [ + 48, + 217, + 287, + 261 + ], + "type": "text", + "content": "[34] Carlos Riquelme, Joan Puigcerver, Basil Mustafa, Maxim Neumann, Rodolphe Jenatton, André Susano Pinto, Daniel Keysers, and Neil Houlsby. Scaling vision with sparse mixture of experts. arXiv preprint arXiv:2106.05974, 2021." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 261, + 287, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 261, + 287, + 294 + ], + "spans": [ + { + "bbox": [ + 48, + 261, + 287, + 294 + ], + "type": "text", + "content": "[35] Jonathan S Rosenfeld, Amir Rosenfeld, Yonatan Belinkov, and Nir Shavit. A constructive prediction of the generalization error across scales. arXiv preprint arXiv:1909.12673, 2019." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 294, + 288, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 294, + 288, + 360 + ], + "spans": [ + { + "bbox": [ + 48, + 294, + 288, + 360 + ], + "type": "text", + "content": "[36] Aarohi Srivastava, Abhinav Rastogi, Abhishek Rao, Abu Awal Md Shoeb, Abubakar Abid, Adam Fisch, Adam R Brown, Adam Santoro, Aditya Gupta, Adrià Garriga-Alonso, et al. Beyond the imitation game: Quantifying and extrapolating the capabilities of language models. arXiv preprint arXiv:2206.04615, 2022." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 361, + 288, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 361, + 288, + 403 + ], + "spans": [ + { + "bbox": [ + 48, + 361, + 288, + 403 + ], + "type": "text", + "content": "[37] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International conference on machine learning, pages 6105-6114. PMLR, 2019." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 404, + 288, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 404, + 288, + 448 + ], + "spans": [ + { + "bbox": [ + 48, + 404, + 288, + 448 + ], + "type": "text", + "content": "[38] Zhan Tong, Yibing Song, Jue Wang, and Limin Wang. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training. arXiv preprint arXiv:2203.12602, 2022." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 449, + 288, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 449, + 288, + 503 + ], + "spans": [ + { + "bbox": [ + 48, + 449, + 288, + 503 + ], + "type": "text", + "content": "[39] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International Conference on Machine Learning, pages 10347-10357. PMLR, 2021." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 504, + 288, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 504, + 288, + 568 + ], + "spans": [ + { + "bbox": [ + 48, + 504, + 288, + 568 + ], + "type": "text", + "content": "[40] Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alex Shepard, Hartwig Adam, Pietro Perona, and Serge Belongie. The inaturalist species classification and detection dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 8769-8778, 2018." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 570, + 288, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 570, + 288, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 570, + 288, + 613 + ], + "type": "text", + "content": "[41] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 614, + 288, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 288, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 288, + 668 + ], + "type": "text", + "content": "[42] Pascal Vincent, Hugo Larochelle, Yoshua Bengio, and Pierre-Antoine Manzagol. Extracting and composing robust features with denoising autoencoders. In Proceedings of the 25th international conference on Machine learning, pages 1096-1103, 2008." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 669, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 288, + 713 + ], + "type": "text", + "content": "[43] Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, and Jian Sun. Unified perceptual parsing for scene understanding. In Proceedings of the European Conference on Computer Vision (ECCV), pages 418-434, 2018." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 263 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 116 + ], + "type": "text", + "content": "[44] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. arXiv preprint arXiv:2111.09886, 2021." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 547, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 547, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 547, + 140 + ], + "type": "text", + "content": "[45] Xiaohua Zhai, Alexander Kolesnikov, Neil Houlsby, and Lucas Beyer. Scaling vision transformers, 2021." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 141, + 547, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 547, + 174 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 547, + 174 + ], + "type": "text", + "content": "[46] Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412, 2017." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 175, + 547, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 547, + 218 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 547, + 218 + ], + "type": "text", + "content": "[47] Zhun Zhong, Liang Zheng, Guoliang Kang, Shaozi Li, and Yi Yang. Random erasing data augmentation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 13001-13008, 2020." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 219, + 547, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 219, + 547, + 263 + ], + "spans": [ + { + "bbox": [ + 307, + 219, + 547, + 263 + ], + "type": "text", + "content": "[48] Bolei Zhou, Hang Zhao, Xavier Puig, Tete Xiao, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Semantic understanding of scenes through the ade20k dataset. International Journal on Computer Vision, 2018." + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10374" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2023/On Distillation of Guided Diffusion Models/2a7f0724-3ab3-4aec-b0e2-13e799a34131_content_list.json b/2023/On Distillation of Guided Diffusion Models/2a7f0724-3ab3-4aec-b0e2-13e799a34131_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..72daaebfaf91e83840477e08ffa1dcb9e71fb2b9 --- /dev/null +++ b/2023/On Distillation of Guided Diffusion Models/2a7f0724-3ab3-4aec-b0e2-13e799a34131_content_list.json @@ -0,0 +1,2099 @@ +[ + { + "type": "text", + "text": "On Distillation of Guided Diffusion Models", + "text_level": 1, + "bbox": [ + 266, + 66, + 702, + 87 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chenlin Meng1", + "bbox": [ + 135, + 118, + 259, + 136 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "chenlin@cs.stanford.edu", + "bbox": [ + 96, + 138, + 302, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Robin Rombach", + "bbox": [ + 349, + 118, + 488, + 133 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "robin@stability.ai", + "bbox": [ + 338, + 138, + 500, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ruiqi Gao3", + "bbox": [ + 566, + 118, + 658, + 136 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ruiqig@google.com", + "bbox": [ + 537, + 138, + 689, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Diederik Kingma3", + "bbox": [ + 725, + 118, + 872, + 136 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "durk@google.com", + "bbox": [ + 733, + 138, + 867, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Stefano Ermon1", + "bbox": [ + 191, + 160, + 320, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ermon@cs.stanford.edu", + "bbox": [ + 163, + 181, + 351, + 194 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jonathan Ho3", + "bbox": [ + 437, + 161, + 547, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "jonathanho@google.com", + "bbox": [ + 401, + 181, + 588, + 195 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tim Salimans3", + "bbox": [ + 661, + 161, + 779, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "salimans@google.com", + "bbox": [ + 637, + 181, + 807, + 195 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Stanford University", + "bbox": [ + 114, + 202, + 281, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Stability AI & LMU Munich", + "bbox": [ + 320, + 203, + 558, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Google Research, Brain Team", + "bbox": [ + 596, + 203, + 849, + 220 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c2eeb0d1c8ee3d67de5f0bd8ebdb1d511976461768a02745b3ce32697ffeeb41.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 78, + 243, + 444, + 369 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6bc0a00ef4c5fb732f8a9a85c45ba6d9f335f764a52d73c374939fa732b3b43c.jpg", + "image_caption": [ + "Text-guided generation (2 steps)" + ], + "image_footnote": [], + "bbox": [ + 444, + 243, + 668, + 332 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fb7c068a1a3cd4193afa89c0fa80650a46e01b75ed9c93a1e7a5c754d1f53f51.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 668, + 243, + 777, + 332 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/05f585747c24d0b4496c7704edd85d5e06f031b142d4bf654d7e4d4f74b9d32a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 779, + 243, + 890, + 332 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/073499f830b61bcd4676870d20a763147b3120cd4955f99111068ee31a1d942d.jpg", + "image_caption": [ + "Text-guided generation (1 step)", + "Text-guided generation (4 steps)", + "Figure 1. Distilled Stable Diffusion samples generated by our method. Our two-stage distillation approach is able to generate realistic images using only 1 to 4 denoising steps on various tasks. Compared to standard classifier-free guided diffusion models, we reduce the total number of sampling steps by at least $20\\times$ ." + ], + "image_footnote": [], + "bbox": [ + 78, + 391, + 444, + 720 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/930d28fbb1d978339636729faa2bba17db923fb6442eff4e688f3e04265de9d3.jpg", + "image_caption": [ + "Class-conditional generation (1 step)" + ], + "image_footnote": [], + "bbox": [ + 444, + 354, + 555, + 400 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/761b62890495ac24622843713658135518820ba002808e144cf9e706c4bac63b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 557, + 354, + 666, + 400 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9d459bfa505fa2c1497c570b6921b2dcbab1629f9ef72ee07f7eb05569d69711.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 669, + 354, + 777, + 400 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bc9ccaa9a22410980952341e3ff2ef4ef800dca6aca46799c97caf8821aa29af.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 779, + 354, + 890, + 400 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/03db0c5d9eb6441c41986da2d6e5386a4bfcd34e759af4a7098f41f8cbd3aedc.jpg", + "image_caption": [ + "Input", + "Mask" + ], + "image_footnote": [], + "bbox": [ + 444, + 421, + 666, + 508 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3d549196b9cc2a5c0828462b35b7fd045c19ce2d6f5a135009ca75c236e28be3.jpg", + "image_caption": [ + "Result 1" + ], + "image_footnote": [], + "bbox": [ + 669, + 421, + 777, + 508 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/830501d9a57ebdc82776b758ec87a7be8ed0b3b746a57941d6c5d4f01a86ce20.jpg", + "image_caption": [ + "Result 2" + ], + "image_footnote": [], + "bbox": [ + 779, + 421, + 890, + 508 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e8067a0c257de4c5d754d4243c0bfd6104830da3c576bfcfb2530b2713bd7800.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 444, + 537, + 555, + 710 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8f7157f3ee46790d6efa8644ff1650d70db90df89d8f0c9ec99839a9b8f1f166.jpg", + "image_caption": [ + "Image inpainting (2 steps)", + "Output (different styles)", + "Image to image translation (3 steps)" + ], + "image_footnote": [], + "bbox": [ + 555, + 537, + 666, + 710 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bce2b8a74953cbb79b3031e8b93e5a714c9b756ef5513ec934346347561f0ae0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 666, + 537, + 779, + 710 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e33e987788fca892e98d8b4cd50ca2126e3bf56eeb23e98d3b3c4e22fb849e55.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 779, + 537, + 890, + 710 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 797, + 313, + 813 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Classifier-free guided diffusion models have recently been shown to be highly effective at high-resolution image generation, and they have been widely used in large-scale diffusion", + "bbox": [ + 76, + 830, + 470, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "frameworks including DALL·E 2, Stable Diffusion and Image. However, a downside of classifier-free guided diffusion models is that they are computationally expensive at inference time since they require evaluating two diffusion models, a class-conditional model and an unconditional model, tens to hundreds of times. To deal with this limitation, we pro", + "bbox": [ + 496, + 799, + 895, + 890 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation.", + "bbox": [ + 236, + 0, + 808, + 18 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "Except for this watermark, it is identical to the accepted version;", + "bbox": [ + 323, + 16, + 722, + 30 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 292, + 31, + 754, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Work partially done during an internship at Google", + "bbox": [ + 94, + 886, + 372, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "14297", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "pose an approach to distilling classifier-free guided diffusion models into models that are fast to sample from: Given a pre-trained classifier-free guided model, we first learn a single model to match the output of the combined conditional and unconditional models, and then we progressively distill that model to a diffusion model that requires much fewer sampling steps. For standard diffusion models trained on the pixel-space, our approach is able to generate images visually comparable to that of the original model using as few as 4 sampling steps on ImageNet 64x64 and CIFAR-10, achieving FID/IS scores comparable to that of the original model while being up to 256 times faster to sample from. For diffusion models trained on the latent-space (e.g., Stable Diffusion), our approach is able to generate high-fidelity images using as few as 1 to 4 denoising steps, accelerating inference by at least 10-fold compared to existing methods on ImageNet 256x256 and LAION datasets. We further demonstrate the effectiveness of our approach on text-guided image editing and inpainting, where our distilled model is able to generate high-quality results using as few as 2-4 denoising steps.", + "bbox": [ + 75, + 90, + 472, + 395 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 422, + 209, + 436 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Denoising diffusion probabilistic models (DDPMs) [4, 37, 39, 40] have achieved state-of-the-art performance on image generation [22, 26–28, 31], audio synthesis [11], molecular generation [44], and likelihood estimation [10]. Classifier-free guidance [6] further improves the sample quality of diffusion models and has been widely used in large-scale diffusion model frameworks including GLIDE [23], Stable Diffusion [28], DALL-E 2 [26], and Imagen [31]. However, one key limitation of classifier-free guidance is its low sampling efficiency—it requires evaluating two diffusion models tens to hundreds of times to generate one sample. This limitation has hindered the application of classifier-free guidance models in real-world settings. Although distillation approaches have been proposed for diffusion models [33, 38], these approaches are not directly applicable to classifier-free guided diffusion models. To deal with this issue, we propose a two-stage distillation approach to improving the sampling efficiency of classifier-free guided models. In the first stage, we introduce a single student model to match the combined output of the two diffusion models of the teacher. In the second stage, we progressively distill the model learned from the first stage to a fewer-step model using the approach introduced in [33]. Using our approach, a single distilled model is able to handle a wide range of different guidance strengths, allowing for the trade-off between sample quality and diversity efficiently. To sample from our model, we consider existing deterministic samplers in the literature [33, 38] and further propose a stochastic sampling process.", + "bbox": [ + 75, + 446, + 472, + 869 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our distillation framework can not only be applied to standard diffusion models trained on the pixel-space [4, 36, 39],", + "bbox": [ + 76, + 869, + 470, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "but also diffusion models trained on the latent-space of an autoencoder [28, 35] (e.g., Stable Diffusion [28]). For diffusion models directly trained on the pixel-space, our experiments on ImageNet 64x64 and CIFAR-10 show that the proposed distilled model can generate samples visually comparable to that of the teacher using only 4 steps and is able to achieve comparable FID/IS scores as the teacher model using as few as 4 to 16 steps on a wide range of guidance strengths (see Fig. 2). For diffusion model trained on the latent-space of an encoder [28, 35], our approach is able to achieve comparable visual quality to the base model using as few as 1 to 4 sampling steps (at least $10 \\times$ fewer steps than the base model) on ImageNet $256 \\times 256$ and LAION $512 \\times 512$ , matching the performance of the teacher (as evaluated by FID) with only 2-4 sampling steps. To the best of our knowledge, our work is the first to demonstrate the effectiveness of distillation for both pixel-space and latent-space classifier-free diffusion models. Finally, we apply our method to text-guided image inpainting and text-guided image editing tasks [20], where we reduce the total number of sampling steps to as few as 2-4 steps, demonstrating the potential of the proposed framework in style-transfer and image-editing applications [20, 41].", + "bbox": [ + 496, + 90, + 893, + 425 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e73a9203a22b09ac814fb8e05d548d163277f88bbb1752ecc5334f7bf4424773.jpg", + "image_caption": [ + "Figure 2. Class-conditional samples from our two-stage (deterministic) approach on ImageNet 64x64 for diffusion models trained on the pixel-space. By varying the guidance weight $w$ , our distilled model is able to trade-off between sample diversity and quality, while achieving good results using as few as one sampling step." + ], + "image_footnote": [], + "bbox": [ + 501, + 438, + 890, + 579 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Background on diffusion models", + "text_level": 1, + "bbox": [ + 500, + 672, + 797, + 689 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Given samples $\\mathbf{x}$ from a data distribution $p_{\\mathrm{data}}(\\mathbf{x})$ , noise scheduling functions $\\alpha_{t}$ and $\\sigma_{t}$ , we train a diffusion model $\\hat{\\mathbf{x}}_{\\theta}$ , with parameter $\\theta$ , via minimizing the weighted mean squared error [4, 36, 39, 40]", + "bbox": [ + 496, + 700, + 890, + 761 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} _ {t \\sim U [ 0, 1 ], \\mathbf {x} \\sim p _ {\\text {d a t a}} (\\mathbf {x}), \\mathbf {z} _ {t} \\sim q \\left(\\mathbf {z} _ {t} \\mid \\mathbf {x}\\right)} \\left[ \\omega \\left(\\lambda_ {t}\\right) \\right| | \\hat {\\mathbf {x}} _ {\\boldsymbol {\\theta}} \\left(\\mathbf {z} _ {t}\\right) - \\mathbf {x} | | _ {2} ^ {2} ], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 773, + 893, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $\\lambda_{t} = \\log [\\alpha_{t}^{2} / \\sigma_{t}^{2}]$ is a signal-to-noise ratio [10], $q(\\mathbf{z}_t|\\mathbf{x}) = \\mathcal{N}(\\mathbf{z}_t;\\alpha_t\\mathbf{x},\\sigma_t^2\\mathbf{I})$ and $\\omega (\\lambda_t)$ is a pre-specified weighting function [10].", + "bbox": [ + 496, + 806, + 890, + 853 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Once the diffusion model $\\hat{\\mathbf{x}}_{\\theta}$ is trained, one can use discrete-time DDIM sampler [38] to sample from the model. Specifically, the DDIM sampler starts with $\\mathbf{z}_1 \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{I})$", + "bbox": [ + 496, + 854, + 893, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "14298", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "and updates as follows", + "bbox": [ + 76, + 90, + 230, + 106 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {z} _ {s} = \\alpha_ {s} \\hat {\\mathbf {x}} _ {\\boldsymbol {\\theta}} (\\mathbf {z} _ {t}) + \\sigma_ {s} \\frac {\\mathbf {z} _ {t} - \\alpha_ {t} \\hat {\\mathbf {x}} _ {\\boldsymbol {\\theta}} (\\mathbf {z} _ {t})}{\\sigma_ {t}}, s = t - 1 / N \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 114, + 468, + 146 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "with $N$ the total number of sampling steps. The final sample will then be generated using $\\hat{\\mathbf{x}}_{\\theta}(\\mathbf{z}_0)$ .", + "bbox": [ + 75, + 152, + 468, + 183 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Classifier-free guidance Classifier-free guidance [6] is an effective approach shown to significantly improve the sample quality of class-conditioned diffusion models, and has been widely used in large-scale diffusion models including GLIDE [23], Stable Diffusion [28], DALL-E 2 [26] and Imagen [31]. Specifically, it introduces a guidance weight parameter $w \\in \\mathbb{R}^{\\geq 0}$ to trade-off between sample quality and diversity. To generate a sample, classifier-free guidance evaluates both a conditional diffusion model $\\hat{\\mathbf{x}}_{c,\\theta}$ , where $c$ is the context (e.g., class label, text prompt) to be conditioned on, and a jointly trained unconditional diffusion model $\\hat{\\mathbf{x}}_{\\theta}$ at each update step, using $\\hat{\\mathbf{x}}_{\\theta}^{w} = (1 + w)\\hat{\\mathbf{x}}_{c,\\theta} - w\\hat{\\mathbf{x}}_{\\theta}$ as the model prediction in Eq. (2). As each sampling update requires evaluating two diffusion models, sampling with classifier-free guidance is often expensive [6].", + "bbox": [ + 75, + 183, + 470, + 409 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Progressive distillation Our approach is inspired by progressive distillation [33], an effective method for improving the sampling speed of (unguided) diffusion models by repeated distillation. Until now, this method could not be directly applied to distilling classifier-free guided models or studied for samplers other than the deterministic DDIM sampler [33, 38]. In this paper we resolve these shortcomings.", + "bbox": [ + 75, + 410, + 470, + 515 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Latent diffusion models (LDMs) [21,24,28,35] increase the training and inference efficiency of diffusion models (directly learned on the pixel-space) by modeling images in the latent space of a pre-trained regularized autoencoder, where the latent representations are usually of lower dimensionality than the pixel-space. Latent diffusion models can be considered as an alternative to cascaded diffusion approaches [5], which rely on one or more super-resolution diffusion models to scale up a low-dimensional image to the desired target resolution.", + "bbox": [ + 75, + 515, + 470, + 665 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this work, we will apply our distillation framework to classifier-free guided diffusion models learned on both pixel-space [4, 36, 39] and latent-space [21, 24, 28, 35].", + "bbox": [ + 75, + 666, + 468, + 712 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Distilling a guided diffusion model", + "text_level": 1, + "bbox": [ + 76, + 724, + 390, + 742 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the following, we discuss our approach for distilling a classifier-free guided diffusion model [6] into a student model that requires fewer steps to sample from. Using a single distilled model conditioned on the guidance strength, our model can capture a wide range of classifier-free guidance levels, allowing for the trade-off between sample quality and diversity efficiently.", + "bbox": [ + 75, + 750, + 468, + 854 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a trained guided model $[\\hat{\\mathbf{x}}_{c,\\theta},\\hat{\\mathbf{x}}_{\\theta}]$ (teacher) either on the pixel-space or latent-space, our approach can be decomposed into two stages.", + "bbox": [ + 75, + 854, + 470, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Stage-one distillation", + "text_level": 1, + "bbox": [ + 500, + 90, + 697, + 106 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the first stage, we introduce a student model $\\hat{\\mathbf{x}}_{\\eta_1}(\\mathbf{z}_t,w)$ , with learnable parameter $\\pmb{\\eta}_{1}$ , to match the output of the teacher at any time-step $t\\in [0,1]$ . The student model can either be a continuous-time model [40] or a discrete-time model [4, 38] depending on whether the teacher model is discrete or continuous. For simplicity, in the following discussion, we assume both the student and teacher models are continuous as the algorithm for discrete models is almost identical.", + "bbox": [ + 496, + 113, + 893, + 247 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A key functionality of classifier-free guidance [6] is its ability to easily trade-off between sample quality and diversity, which is controlled by a \"guidance strength\" parameter. This property has demonstrated utility in real-world applications [6,23,26,28,31], where the optimal \"guidance strength\" is often a user preference. Thus, we would also want our distilled model to maintain this property. Given a range of guidance strengths $\\left[w_{\\min},w_{\\max}\\right]$ we are interested in, we optimize the student model using the following objective", + "bbox": [ + 496, + 250, + 893, + 387 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} _ {w \\sim p _ {w}, t \\sim U [ 0, 1 ], \\mathbf {x} \\sim p _ {\\mathrm {d a t a}} (\\mathbf {x})} \\left[ \\omega \\left(\\lambda_ {t}\\right) \\| \\hat {\\mathbf {x}} _ {\\boldsymbol {\\eta} _ {1}} \\left(\\mathbf {z} _ {t}, w\\right) - \\hat {\\mathbf {x}} _ {\\boldsymbol {\\theta}} ^ {w} \\left(\\mathbf {z} _ {t}\\right) \\| _ {2} ^ {2} \\right], \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 506, + 395, + 892, + 426 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\hat{\\mathbf{x}}_{\\pmb{\\theta}}^{w}(\\mathbf{z}_{t}) = (1 + w)\\hat{\\mathbf{x}}_{c,\\pmb{\\theta}}(\\mathbf{z}_{t}) - w\\hat{\\mathbf{x}}_{\\pmb{\\theta}}(\\mathbf{z}_{t}), \\mathbf{z}_{t} \\sim q(\\mathbf{z}_{t}|\\mathbf{x})$ and $p_w(w) = U[w_{\\min}, w_{\\max}]$ . Note that here, our distilled model $\\hat{\\mathbf{x}}_{\\pmb{\\eta}_1}(\\mathbf{z}_t, w)$ is also conditioned on the context $c$ (e.g., text prompt), but we drop the notation $c$ in the paper for simplicity. We provide the detailed training algorithm in Algorithm 1 in the supplement.", + "bbox": [ + 496, + 434, + 893, + 526 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To incorporate the guidance weight $w$ , we introduce a $w$ -conditioned model, where $w$ is fed as an input to the student model. To better capture the feature, we apply Fourier embedding to $w$ , which is then incorporated into the diffusion model backbone in a way similar to how the time-step was incorporated in [10, 33]. As initialization plays a key role in the performance [33], we initialize the student model with the same parameters as the conditional model of the teacher, except for the newly introduced parameters related to $w$ -conditioning. The model architecture we use is a U-Net model similar to the ones used in [6] for pixel-space diffusion models and [1, 28] for latent-space diffusion models. We use the same number of channels and attention as used in [6] and the open-sourced Stable Diffusion repository* for our experiments. We provide more details in the supplement.", + "bbox": [ + 496, + 527, + 893, + 753 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Stage-two distillation", + "text_level": 1, + "bbox": [ + 500, + 762, + 697, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the second stage, we consider a discrete time-step scenario and progressively distill the learned model from the first-stage $\\hat{\\mathbf{x}}_{\\eta_1}(\\mathbf{z}_t,w)$ into an fewer-step student model $\\hat{\\mathbf{x}}_{\\eta_2}(\\mathbf{z}_t,w)$ with learnable parameter $\\eta_{2}$ , by halving the number of sampling steps each time. Letting $N$ denote the number of sampling steps, given $w\\sim U[w_{\\min},w_{\\max}]$ and", + "bbox": [ + 496, + 786, + 893, + 878 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "*https://github.com/CompVis/stable-diffusion", + "bbox": [ + 516, + 887, + 864, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "14299", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/019b251179ebf0264ae30e436a0b7e9f5f3aafa287ebba6c38b489f2427f170f.jpg", + "image_caption": [ + "(a) 2 denoising steps" + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 176, + 290 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/98f99a856052c3f482501fe0669cb0c10b53bd98f17a18ce6e5906bc7babb438.jpg", + "image_caption": [ + "Figure 4. Text-guided generation on LAION (512x512) using our distilled Stable Diffusion model. Our model is able to generate high-quality image samples using 2, 4 or 8 denoising steps, significantly improving the inference efficiency of Stable Diffusion." + ], + "image_footnote": [], + "bbox": [ + 176, + 88, + 369, + 290 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/436fa4e547ed4588699e76e23a4412f052419a9cf226bfcc0136ebce27dbc277.jpg", + "image_caption": [ + "(b) 4 denoising steps" + ], + "image_footnote": [], + "bbox": [ + 370, + 88, + 614, + 290 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0cd6729a2cc48efc54e30037e083d383a4eda77f75bf42cc7ed659ecefeec9e6.jpg", + "image_caption": [ + "(c) 8 denoising steps" + ], + "image_footnote": [], + "bbox": [ + 616, + 88, + 877, + 290 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/4dfd2fda850903cabadf2eeb5d001c98f5e0f10c823a687410de2e7502df96e1.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
w=0w=0.3w=1w=4
MethodFID (↓)IS (↑)FID (↓)IS (↑)FID (↓)IS (↑)FID (↓)IS (↑)
Ours 1-step (D/S)22.74 / 26.9125.51 / 23.5514.85 / 18.4837.09 / 33.307.54 / 8.9275.19 / 67.8018.72 / 17.85157.46 / 148.97
Ours 4-step (D/S)4.14 / 3.9146.64 / 48.922.17 / 2.2469.64 / 73.737.95 / 8.51128.98 / 135.3626.45 / 27.33207.45 / 216.56
Ours 8-step (D/S)2.79 / 2.4450.72 / 55.032.05 / 2.3176.01 / 83.009.33 / 10.56136.47 / 147.3926.62 / 27.84203.47 / 219.89
Ours 16-step (D/S)2.44 / 2.1052.53 / 57.812.20 / 2.5679.47 / 87.509.99 / 11.63139.11 / 153.1726.53 / 27.69204.13 / 218.70
Single-w 1-step19.6124.0011.7036.956.6474.4119.857170.69
Single-w 4-step4.7938.772.3462.088.23118.5227.75219.64
Single-w 8-step3.3942.132.3268.769.69125.2027.67218.08
Single-w 16-step2.9743.632.5670.9710.34127.7027.40216.52
DDIM 16x2-step [38]7.6837.605.3360.839.53112.7521.56195.17
DDIM 32x2-step [38]5.0340.937.479.339.26126.2223.03213.23
DDIM 64x2-step [38]3.7443.165.529.519.53133.1723.64217.88
Teacher (DDIM 1024x2-step)2.9244.812.3674.839.84139.5023.94224.74
", + "bbox": [ + 81, + 347, + 890, + 508 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 1. ImageNet 64x64 distillation results for pixel-space diffusion models ( $w = 0$ refers to non-guided models). For our method, $D$ and $S$ stand for deterministic and stochastic sampler respectively. We observe that training the model conditioned on a guidance interval $w \\in [0,4]$ performs comparably with training a model on a fixed $w$ (see Single- $w$ ). Our approach significantly outperforms DDIM when using fewer steps, and is able to match the teacher performance using as few as 8 to 16 steps.", + "bbox": [ + 75, + 518, + 892, + 575 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$t \\in \\{1, \\dots, N\\}$ , we train the student model to match the output of two-step DDIM sampling of the teacher (i.e., from $t/N$ to $t - 0.5/N$ and from $t - 0.5/N$ to $t - 1/N$ ) in one step, following the approach of [33]. After distilling the $2N$ steps in the teacher model to $N$ steps in the student model, we can use the $N$ -step student model as the new teacher model, repeat the same procedure, and distill the teacher model into a $N/2$ -step student model. At each step, we initialize the student model with the parameters of the teacher. We provide the training algorithm and extra details in the supplementary material.", + "bbox": [ + 75, + 589, + 470, + 755 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. $N$ -step deterministic and stochastic sampling", + "text_level": 1, + "bbox": [ + 76, + 768, + 457, + 787 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Once the model $\\hat{\\mathbf{x}}_{\\eta_2}$ is trained, given a specified guidance strength $w \\in [w_{\\min}, w_{\\max}]$ , we can perform sampling via the DDIM update rule in Eq. (2). We note that given the distilled model $\\hat{\\mathbf{x}}_{\\eta_2}$ , this sampling procedure is deterministic given the initialization $\\mathbf{z}_1^w$ . In fact, we can also perform $N$ -step stochastic sampling: We apply one deterministic sampling step with two-times the original step-length (i.e.,", + "bbox": [ + 75, + 795, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the same as a $N/2$ -step deterministic sampler) and then perform one stochastic step backward (i.e., perturb with noise) using the original step-length, a process inspired by [9]. With $\\mathbf{z}_1^w \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{I})$ , we use the following update rule when $t > 1/N$", + "bbox": [ + 496, + 589, + 890, + 665 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {z} _ {k} ^ {w} = \\alpha_ {k} \\hat {\\mathbf {x}} _ {\\boldsymbol {\\eta} _ {2}} \\left(\\mathbf {z} _ {t} ^ {w}\\right) + \\sigma_ {k} \\frac {\\mathbf {z} _ {t} ^ {w} - \\alpha_ {t} \\hat {\\mathbf {x}} _ {\\boldsymbol {\\eta} _ {2}} ^ {w} \\left(\\mathbf {z} _ {t}\\right)}{\\sigma_ {t}}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 676, + 890, + 709 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\text {w h e r e} \\mathbf {z} _ {s} ^ {w} = \\left(\\alpha_ {s} / \\alpha_ {k}\\right) \\mathbf {z} _ {k} ^ {w} + \\sigma_ {s | k} \\epsilon , \\epsilon \\sim \\mathcal {N} (\\mathbf {0}, \\mathbf {I}); \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 710, + 890, + 729 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {z} _ {h} ^ {w} = \\alpha_ {h} \\hat {\\mathbf {x}} _ {\\boldsymbol {\\eta} _ {2}} \\left(\\mathbf {z} _ {s} ^ {w}\\right) + \\sigma_ {h} \\frac {\\mathbf {z} _ {s} ^ {w} - \\alpha_ {s} \\hat {\\mathbf {x}} _ {\\boldsymbol {\\eta} _ {2}} ^ {w} \\left(\\mathbf {z} _ {s} ^ {w}\\right)}{\\sigma_ {s}}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 731, + 890, + 763 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nw h e r e \\mathbf {z} _ {k} ^ {w} = \\left(\\alpha_ {k} / \\alpha_ {h}\\right) \\mathbf {z} _ {h} ^ {w} + \\sigma_ {k | h} \\epsilon , \\epsilon \\sim \\mathcal {N} (\\mathbf {0}, \\mathbf {I}). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 766, + 890, + 784 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the above equations, $h = t - 3 / N$ , $k = t - 2 / N$ , $s = t - 1 / N$ and $\\sigma_{a|b}^{2} = (1 - e^{\\lambda_{a} - \\lambda_{b}})\\sigma_{a}^{2}$ . When $t = 1 / N$ , we use deterministic update Eq. (2) to obtain $\\mathbf{z}_0^w$ from $\\mathbf{z}_{1 / N}^w$ . We provide an illustration of the process in Fig. 5, where the number of denoising steps is 4. We note that compared to the deterministic sampler, performing stochastic sampling requires evaluating the model at slightly different time-steps,", + "bbox": [ + 496, + 794, + 892, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "14300", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ba61fe9a0723a73bb3500aa6adbf6117174f16451908b263c1665bacb55124e2.jpg", + "image_caption": [ + "Figure 5. Sampling procedures of the distilled model where the number of denoising steps is 4." + ], + "image_footnote": [], + "bbox": [ + 84, + 75, + 468, + 179 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and would require small modifications to training algorithm for the edge cases. We provide the algorithm and more details in the supplementary material.", + "bbox": [ + 75, + 224, + 468, + 271 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 284, + 209, + 301 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we evaluate the performance of our distillation approach on pixel-space diffusion models (i.e. DDPM [4]) and latent-space diffusion models (i.e. Stable Diffusion [28]). We further apply our approach to text-guided image editing and inpainting tasks. Experiments show that our approach is able to achieve competitive performance while using as few as 2-4 steps on all tasks.", + "bbox": [ + 75, + 309, + 470, + 414 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Distillation for pixel-space guided models", + "text_level": 1, + "bbox": [ + 76, + 424, + 429, + 440 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this experiment, we consider class-conditional diffusion models directly trained on the pixel-space [4,6,33].", + "bbox": [ + 75, + 446, + 470, + 477 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Settings We focus on ImageNet 64x64 [30] and CIFAR-10 [12] as higher-resolution image generation in this scenario often relies on combining with other super-resolution techniques [5, 31]. We explore different ranges for the guidance weight and observe that all ranges work comparably and therefore use $[w_{min}, w_{max}] = [0, 4]$ for the experiments. The baselines we consider include DDPM ancestral sampling [4] and DDIM [38]. The teacher model we use is a $1024 \\times 2$ -step DDIM model, where the conditional and unconditional components both use 1024 DDIM denoising steps. To better understand how the guidance weight $w$ should be incorporated, we also include models trained using a single fixed $w$ as a baseline. We use the same pre-trained teacher model for all the methods for fair comparisons. Following [4, 6, 39], we use a U-Net [29, 39] architecture for the baselines, and the same U-Net backbone with the introduced $w$ -embedding for our two-step student models (see Sec. 3). Following [33], we use a v-prediction model for both datasets.", + "bbox": [ + 75, + 478, + 470, + 762 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Results We report the performance as evaluated in FID [3] and Inception scores (IS) [32] for all approaches on ImageNet 64x64 in Fig. 6 and Tab. 1 and provide extended ImageNet 64x64 and CIFAR-10 results in the supplement. We observe that our distilled model is able to match a teacher guided DDIM model with $1024 \\times 2$ sampling steps using only 4-16 steps, achieving a speedup for up to $256 \\times$ . We emphasize that, using our approach, a single distilled model is able to match the teacher performance on a wide range", + "bbox": [ + 75, + 763, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "of guidance strengths. This has not been achieved by any previous methods.", + "bbox": [ + 498, + 90, + 890, + 121 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Distillation for latent-space guided models", + "text_level": 1, + "bbox": [ + 498, + 132, + 857, + 148 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After demonstrating the effectiveness of our method on pixel-space class-guided diffusion models in Sec. 4.1, we now expand its scope to latent-space diffusion models. In the following sections, we show the effectiveness of our approach on Latent Diffusion [28] on a variety of tasks, including class-conditional generation, text-to-image generation, image inpainting and text-guided style-transfer [20].", + "bbox": [ + 496, + 155, + 893, + 260 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the following experiments, we use the open-sourced latent-space diffusion models [28] as the teacher models. As v-prediction teacher model tends to perform better than $\\epsilon$ -prediction model, we fine-tune the open-sourced $\\epsilon$ -prediction models into v-prediction teacher models. We provide more details in the supplementary material.", + "bbox": [ + 496, + 261, + 893, + 351 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2.1 Class-conditional generation", + "text_level": 1, + "bbox": [ + 498, + 372, + 751, + 387 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we apply our method to a class-conditional latent diffusion model pre-trained on ImageNet $256 \\times 256$ . We start from the DDIM teacher model with 512 sampling steps, and use the output as the target to train our distilled model. We use a batch size of 512 and uniformly sample the guidance strength $w \\in [w_{min} = 0, w_{max} = 14]$ during training.", + "bbox": [ + 496, + 396, + 893, + 501 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Results Empirically, we find that our distilled model is able to match the performance of the teacher model (originally trained on 1000 steps) in terms of FID scores while using only 2 or 4 sampling steps. We also achieve significantly better performance than DDIM when using 1-4 sampling steps (see Fig. 11). Qualitatively, we find that samples synthesized using a single denoising step still yield satisfying results, while the baseline fails to generate images with meaningful contents. We provide extra samples in the supplementary material.", + "bbox": [ + 496, + 502, + 893, + 654 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Similar to the pixel-based results in Fig. 6, we also observe the trade-off between sampling quality and diversity as measured by FID and Inception Score for our distilled latent diffusion model. Following Kynkänniemi et al [13], we further compute improved precision and recall metrics for this experiment in the appendix.", + "bbox": [ + 496, + 654, + 893, + 744 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2.2 Text-guided image generation", + "text_level": 1, + "bbox": [ + 498, + 763, + 759, + 780 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we focus on the text-guided Stable Diffusion model pretrained on subsets $\\dagger$ of LAION-5B [34] at a resolution $512 \\times 512$ . We then follow our two-stage approach introduced in Sec. 3 and distill the guided model in 3000 gradient updates into a $w$ -conditioned model using", + "bbox": [ + 496, + 787, + 893, + 864 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "$\\dagger$ https://github.com/CompVis/stable-diffusion/blob/main/Stable_Diffusion_v1_model/Card.md", + "bbox": [ + 500, + 875, + 893, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "14301", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/cbec320e32ef4c440a3a8913360b49b54d9efe22f3accccb03e69b5ce00c6d22.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 101, + 80, + 251, + 155 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/041b3f32f7b8297c6f864a31573be28d85a8c79f280ffd7c294753f35bfedc22.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 256, + 80, + 406, + 152 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ea5e4d98537a0cffaa6d98c47b80f3469a742385be7a41f9e98d3c400a873292.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 410, + 80, + 560, + 155 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fb707b4c85407ed7c68e73fdf99e27c94232faf70df8adc93f32a22f9ddc039b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 563, + 80, + 715, + 156 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/97ee8399aae499ac1037d73062cb681662c6d6d0066e44c1b5272ef2e23d7b13.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 718, + 80, + 870, + 155 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d0dd693b2a12bb672df55802208ae68a6927d17ee04fb407466baebb21e29e64.jpg", + "image_caption": [ + "$w = 0$" + ], + "image_footnote": [], + "bbox": [ + 101, + 157, + 251, + 226 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/02930878f9429bc9d7483765481e491c86abe4acda7a91cd5de85f01a0637430.jpg", + "image_caption": [ + "$w = 0.3$" + ], + "image_footnote": [], + "bbox": [ + 254, + 157, + 406, + 226 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a0f3e40986241e51c66d85b02b3f296aabecbbe7a66bfe69650e04a6d9e9cb90.jpg", + "image_caption": [ + "$w = 1$" + ], + "image_footnote": [], + "bbox": [ + 410, + 157, + 560, + 224 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3e78bed81ea1fb715c91732ccf22c11d97cf03d6ca52c8a36484e710a2610283.jpg", + "image_caption": [ + "$w = 2$" + ], + "image_footnote": [], + "bbox": [ + 563, + 157, + 715, + 224 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ece91057b2d35114a9a67470305a79bd85312c29203b98ca5fdee83b2f0f728a.jpg", + "image_caption": [ + "$w = 4$" + ], + "image_footnote": [], + "bbox": [ + 718, + 159, + 870, + 224 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/98f799da38b80d6a8b79674b305ab5c500be4cb173e8fe57d005af43ba25b5d7.jpg", + "image_caption": [ + "Figure 6. ImageNet $64 \\times 64$ sample quality evaluated by FID and IS scores. Our distilled model significantly outperforms the DDPM and DDIM baselines, and is able to match the performance of the teacher using as few as 8 to 16 steps. By varying $w$ , a single distilled model is able to capture the trade-off between sample diversity and quality.", + "Distilled Text-to-Image samples (4 steps)", + "Figure 7. Text-guided Stable Diffusion results. We distill the public Stable Diffusion model using the proposed pipeline, arriving at a model that achieves high sample quality using only four denoising steps (left). When sampling from the original model using four DDIM steps, the generated samples have clear artifacts (middle). When using eight DDIM steps, the results get better (right), but are still blurry and less consistent than the distilled results using fewer steps. More samples are provided in Fig. 4." + ], + "image_footnote": [], + "bbox": [ + 80, + 311, + 343, + 414 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f613993d643d5fa0a970a4460f3695f53d71bb718f2447bd5cdc014f1821b63d.jpg", + "image_caption": [ + "Native Text-to-Image samples (4 steps)" + ], + "image_footnote": [], + "bbox": [ + 352, + 311, + 616, + 412 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ff071da0624ff43e38d0107366b09d0da90636bd4f95afe9eb97d1145ec672fb.jpg", + "image_caption": [ + "Native Text-to-Image samples (8 steps)" + ], + "image_footnote": [], + "bbox": [ + 625, + 311, + 890, + 412 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$w \\in [w_{min} = 2, w_{max} = 14]$ , and a batch size of 512. Although we can condition on a broader range of $w$ for the distilled (student) model, the utility remains unclear as we typically do not exceed the normal guidance range when sampling with the teacher model. The final model is obtained by applying progressive distillation for 2000 training steps per stage, except when for the low-step regime of 1,2, and 4 steps, where we train for 20000 gradient updates. A detailed analysis of the convergence properties of this model in the supplement.", + "bbox": [ + 75, + 493, + 472, + 643 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/6ae865654f4f93277c93f96fc2e22483dd1dfa6812da7cb10baa77824e2305f4.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Method2-step4-step8-step
DPM [16]98.9/0.2034.3/0.2931.7/0.32
DPM++ [18]98.8/0.2034.1/0.2925.6/0.32
Ours37.3/0.2726.0/0.3026.9/0.30
", + "bbox": [ + 135, + 645, + 408, + 703 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. FID/CLIP scores on LAION 512X512 ( $w = 8.0$ ). We point out that DPM and DPM++ use both the conditional and unconditional components for sampling. Depending on the implementation, this either requires higher peak memory or two times more sampling steps.", + "bbox": [ + 75, + 705, + 470, + 775 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results We present samples in Fig. 4. We evaluate the resulting model both qualitatively and quantitatively. For the latter analysis, we follow [31] and evaluate CLIP [25] and FID scores to assess text-image alignment and quality, respectively. We use the open-sourced ViT-g/14 [7] CLIP model for evaluation. The quantitative results in Fig. 10 show that our method can significantly increase the performance in both metrics over DDIM sampling on the base model for", + "bbox": [ + 75, + 780, + 470, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "2 and 4 sampling steps. For 8 steps, these metrics do not show a significant difference. However, when considering the corresponding samples in Fig. 7 we can observe a stark difference in terms of visual image quality. In contrast to the 8-step DDIM samples from the original model, the distilled samples are sharper and more coherent. We hypothesize that FID and CLIP do not fully capture these differences in our evaluation setting on COCO2017 [14], where we used 5000 random captions from the validation set. We further compute the FID and CLIP scores for our distilled LAION $512 \\times 512$ model and compare them with the DPM [16] and $\\mathrm{DPM}++$ [18] solver in Tab. 2. We observe that our method is able to achieve significantly better performance when the denoising step is 2 or 4. Furthermore, we stress that stage-one of our method already decreases the number of function evaluations by a factor of 2, as we distill the classifier-free guidance step into a single model. Depending on the exact implementation (batched vs. sequential network evaluation), this either decreases peak memory or sampling time compared to existing solvers [16, 18, 38].", + "bbox": [ + 496, + 493, + 893, + 796 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.3 Text-guided image-to-image translation", + "text_level": 1, + "bbox": [ + 500, + 816, + 828, + 830 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we perform experiments on text-guided image-to-image translation with SDEdit [20] using our distilled model from Sec. 4.2.2. Following SDEdit [20], we perform stochastic encoding in the latent space, but instead", + "bbox": [ + 496, + 839, + 893, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "14302", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fd4a576fa344ffbe1d42def9d273748724b60d646b3109add5004344f748410a.jpg", + "image_caption": [ + "Input", + "Watercolor", + "Art station", + "Photograph", + "Input", + "Watercolor", + "Art station painting" + ], + "image_footnote": [], + "bbox": [ + 80, + 71, + 890, + 136 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5d49590896898c6051dd3950f862aec21dee3a99a02e14257edb9cfc27997da3.jpg", + "image_caption": [ + "Input", + "Girl" + ], + "image_footnote": [], + "bbox": [ + 81, + 151, + 888, + 234 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/649e891f3bcdf8b550572949be033f37aa578bc4ad79ac96b35516b60d6fa51a.jpg", + "image_caption": [ + "Oil painting", + "Boy", + "Anime", + "3D render", + "Anime", + "3D render", + "Male", + "Van Gogh", + "Watercolor", + "Art station" + ], + "image_footnote": [], + "bbox": [ + 81, + 244, + 888, + 329 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e6a319a1aee460d2faec76fc15904b4a758b5e52c5ae49bc288a0f389dc17f09.jpg", + "image_caption": [ + "Figure 8. Text-guided image-to-image translation [20] with the distilled Stable Diffusion model (3 denoising steps). We observe that our model is able to generate high-quality and faithful outputs using only 3 denoising steps.", + "Input" + ], + "image_footnote": [], + "bbox": [ + 81, + 393, + 888, + 484 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/de5653f9a17f5cb852113031736c4f643bd0f5ace4f4dc79472ff83ba3e7eaf2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 81, + 496, + 888, + 563 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b8250917f941b6991c587ce05a0b56d3e8fcbfb4bd96b3861a6272d3865df53c.jpg", + "image_caption": [ + "Figure 10. FID and CLIP ViT-g/14 score for text-to-image generation at $512 \\times 512$ px using the distilled Stable Diffusion model. The results are evaluated on 5000 captions from the COCO2017 [14] validation set. Our distilled latent diffusion model is able to generate high-quality image samples using significantly less sampling steps than the original model while achieving similar or better FID and CLIP scores, especially in the low-step regime." + ], + "image_footnote": [], + "bbox": [ + 81, + 563, + 269, + 645 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f5ef46df91d8617dd752c0e3daeb21932ce5f8e5208038bcd5f76011355e256d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 274, + 563, + 462, + 645 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "use the deterministic sampler of the distilled model to perform deterministic decoding. We consider input image and text of various kinds and provide qualitative results in Fig. 8. We observe that our distilled model generates high-quality style-transfer results using as few as 3 denoising steps. We provide more analysis on the trade-off between sample quality, controllability and efficiency in the supplement.", + "bbox": [ + 75, + 794, + 470, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/280b6991cfac975dd2c8905ef3a699ba24d9805d3babb9bdcb5e47cffbb77c31.jpg", + "image_caption": [ + "Figure 11. FID and Inception Score for class-conditional image generation on ImageNet $(256 \\times 256)$ with distilled latent diffusion. The results are evaluated on 5000 samples. Our distilled latent diffusion model is able to generate high-quality image samples using significantly less sampling steps (up to a factor of 16) than the original model while achieving similar or better FID scores." + ], + "image_footnote": [], + "bbox": [ + 501, + 563, + 687, + 645 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e53a1bc88b4ad86e626365445a352c30deccb744a0b679bd88603736daed3adb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 563, + 887, + 645 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2.4 Image inpainting", + "text_level": 1, + "bbox": [ + 500, + 767, + 671, + 782 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we apply our approach to a pre-trained image inpainting latent diffusion model. We use the open-source Stable Diffusion Inpainting\\* image-inpainting model. This model is a fine-tuned version of the pure text-to-image Stable Diffusion model from above, where additional input channels", + "bbox": [ + 496, + 790, + 890, + 866 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "https://huggingface.co/runwayml/stable-diffusion-inpainting", + "bbox": [ + 500, + 875, + 893, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "14303", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2da267968cdc011b6132827c27e3bf9ff0ba5b7f4a3ca5f167fb7c41033eeeca.jpg", + "image_caption": [ + "Figure 12. Style transfer comparison on ImageNet 64x64 for pixel-space models. For our approach, we use a distilled encoder and decoder. For the baseline, we encode and decode using DDIM. We use $w = 0$ and 16 sampling steps for both the encoder and decoder. We observe that our method achieves more realistic outputs." + ], + "image_footnote": [], + "bbox": [ + 78, + 88, + 467, + 160 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "were added to process masks and masked images.", + "bbox": [ + 76, + 268, + 406, + 282 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We use the same distillation algorithm as used in the previous section. For training, we start from the v-prediction teacher model sampled with 512 DDIM steps, and use the output as the target to optimize our student model. We present qualitative results in Fig. 9, demonstrating the potential of our method for fast, real-world image editing applications. For additional training details and a quantitative evaluation, see the supplementary.", + "bbox": [ + 75, + 282, + 470, + 404 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Progressive distillation for encoding", + "text_level": 1, + "bbox": [ + 76, + 414, + 387, + 430 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this experiment, we explore distilling the encoding process for the teacher model and perform experiments on style-transfer in a setting similar to [41]. We focus on pixel-space diffusion models pre-trained on ImageNet $64 \\times 64$ . Specifically, to perform style-transfer between two domains $A$ and $B$ , we encode the image from domain- $A$ using a diffusion model trained on domain- $A$ , and then decode with a diffusion model trained on domain- $B$ . As the encoding process can be understood as reversing the DDIM sampling process, we perform distillation for both the encoder and decoder with classifier-free guidance, and compare with a DDIM encoder and decoder in Fig. 12. We also explore how modifying the guidance strength $w$ can impact the performance and provide more details in the supplementary material.", + "bbox": [ + 75, + 438, + 470, + 662 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Related Work", + "text_level": 1, + "bbox": [ + 76, + 678, + 218, + 694 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our approach is related to existing works on improving the sampling speed of diffusion models [4, 37, 40]. For instance, denoising diffusion implicit models (DDIM [38]), probability flow sampler [40], fast SDE integrators [8] have been proposed to improve the sampling speed of diffusion models. Other works develop higher-order solvers [17], exponential integrators [15], and dynamic programming based approach [43] to accelerating sampling speed. However, none of these approaches have achieved comparable performance as our method on distilling classifier-free guided diffusion models.", + "bbox": [ + 75, + 704, + 470, + 868 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Existing distillation-based methods for diffusion models are mainly designed for non-classifier-free guided diffusion", + "bbox": [ + 76, + 869, + 470, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "models. For instance, [19] proposes to predict the data from noise in one single step by inverting a deterministic encoding of DDIM, [2] proposes to achieve faster sampling speed by distilling higher order solvers into an additional prediction head of the neural network backbone [2]. Progressive distillation [33] is perhaps the most relevant work. Specifically, it proposes to progressively distill a pre-trained diffusion model into a fewer-step student model with the same model architecture. However, none of these approaches are directly applicable or have been applied to classifier-free guided diffusion models. They are also unable to capture a range of different guidance strengths using one single distilled model. On the contrary, by incorporating the guidance strength into the model architecture and training the model using a two-stage procedure, our approach is able to match the performance of the teacher model on a wide range of guidance strength using one single model. Using our method, one single model can capture the trade-off between sample quality and diversity, enabling the real-world application of classifier-free guided diffusion models, where the guidance strength is often specified by users. Moreover, none of the above distillation approaches have been applied to or shown effectiveness for latent-space text-to-image models. Finally, most fast sampling approaches [33, 38, 40] only consider using deterministic sampling schemes to improve the sampling speed. In this work, we further develop an effective stochastic sampling approach to sample from the distilled models.", + "bbox": [ + 496, + 90, + 893, + 512 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 530, + 617, + 545 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a distillation approach for guided diffusion models [6]. Our two-stage approach allows us to significantly speed up popular but relatively inefficient guided diffusion models. We show that our approach can reduce the inference cost of classifier-free guided pixel-space and latent-space diffusion models by at least an order of magnitude. Empirically, we show that our approach is able to produce visually appealing results with only 2 steps, achieving a comparable FID score to the teacher with as few as 4 to 8 steps. We further demonstrate practical applications of our distillation approach to text-guided image-to-image translation and inpainting tasks. We hope that by significantly reducing the inference cost of classifier-free guided diffusion models, our method will promote creative applications as well as the wider adoption of image generation systems. In the future work, we aim to further improve the performance in the two and one sampling step regimes.", + "bbox": [ + 496, + 556, + 893, + 814 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 500, + 829, + 666, + 845 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We thank the anonymous reviewers for their insightful discussions and feedback. All experiments on Stable Diffusion are supported by Stability AI.", + "bbox": [ + 498, + 854, + 893, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "14304", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 174, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Prafulla Dhariwal and Alex Nichol. Diffusion models beat GANs on image synthesis. arXiv preprint arXiv:2105.05233, 2021. 3", + "[2] Tim Dockhorn, Arash Vahdat, and Karsten Kreis. Genie: Higher-order denoising diffusion solvers. In Advances in Neural Information Processing Systems, 2022. 8", + "[3] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs trained by a two time-scale update rule converge to a local Nash equilibrium. In Advances in Neural Information Processing Systems, pages 6626-6637, 2017. 5", + "[4] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems, pages 6840-6851, 2020. 2, 3, 5, 8", + "[5] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. arXiv preprint arXiv:2106.15282, 2021. 3, 5", + "[6] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 2, 3, 5, 8, 11", + "[7] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, July 2021. If you use this software, please cite it as below. 6, 24", + "[8] Alexia Jolicoeur-Martineau, Ke Li, Rémi Piché-Taillefer, Tal Kachman, and Ioannis Mitliagkas. Gotta go fast when generating data with score-based models. arXiv preprint arXiv:2105.14080, 2021. 8", + "[9] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. arXiv preprint arXiv:2206.00364, 2022. 4", + "[10] Diederik P Kingma, Tim Salimans, Ben Poole, and Jonathan Ho. Variational diffusion models. arXiv preprint arXiv:2107.00630, 2021. 2, 3, 11, 22, 23, 25", + "[11] Zhifeng Kong, Wei Ping, Jiaji Huang, Kexin Zhao, and Bryan Catanzaro. DiffWave: A Versatile Diffusion Model for Audio Synthesis. International Conference on Learning Representations, 2021. 2", + "[12] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 5", + "[13] Tuomas Kynkänniemi, Tero Karras, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Improved precision and recall metric for assessing generative models. Advances in Neural Information Processing Systems, 32, 2019. 5, 24", + "[14] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 6, 7, 24", + "[15] Luping Liu, Yi Ren, Zhijie Lin, and Zhou Zhao. Pseudo numerical methods for diffusion models on manifolds. In International Conference on Learning Representations, 2022. 8" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps, 2022. 6, 23, 24", + "[17] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. arXiv:2206.00927, 2022. 8, 11", + "[18] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models, 2022. 6, 11, 13, 14, 23, 24, 28", + "[19] Eric Luhman and Troy Luhman. Knowledge distillation in iterative generative models for improved sampling speed. arXiv preprint arXiv:2101.02388, 2021. 8", + "[20] Chenlin Meng, Yang Song, Jiaming Song, Jiajun Wu, JunYan Zhu, and Stefano Ermon. Sdedit: Image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 2, 5, 6, 7, 25, 26, 27", + "[21] Gautam Mittal, Jesse Engel, Curtis Hawthorne, and Ian Simon. Symbolic music generation with diffusion models. arXiv preprint arXiv:2103.16091, 2021. 3", + "[22] Alex Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. International Conference on Machine Learning, 2021. 2", + "[23] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 2, 3", + "[24] Konpat Preechakul, Nattanat Chathee, Suttisak Wizadwongsa, and Supasorn Suwajanakorn. Diffusion autoencoders: Toward a meaningful and decodable representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10619-10629, 2022. 3", + "[25] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 6, 24", + "[26] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2, 3", + "[27] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 2", + "[28] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2, 3, 5", + "[29] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation." + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "14305", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "In International Conference on Medical image computing and computer-assisted intervention, pages 234-241. Springer, 2015. 5", + "[30] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. ImageNet large scale visual recognition challenge. International Journal of Computer Vision, 115(3):211-252, 2015. 5", + "[31] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 2, 3, 5, 6", + "[32] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training GANs. In Advances in Neural Information Processing Systems, pages 2234-2242, 2016. 5", + "[33] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. arXiv preprint arXiv:2202.00512, 2022. 2, 3, 4, 5, 8, 11, 19, 22, 23, 25", + "[34] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. arXiv preprint arXiv:2210.08402, 2022. 5", + "[35] Abhishek Sinha, Jiaming Song, Chenlin Meng, and Stefano Ermon. D2c: Diffusion-decoding models for few-shot conditional generation. Advances in Neural Information Processing Systems, 34:12533-12548, 2021. 2, 3", + "[36] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256-2265, 2015. 2, 3", + "[37] Jascha Sohl-Dickstein, Eric A Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. arXiv preprint arXiv:1503.03585, March 2015. 2, 8", + "[38] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. International Conference on Learning Representations, 2021. 2, 3, 4, 5, 6, 8, 11, 13, 14, 16, 22, 24, 28", + "[39] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. In Advances in Neural Information Processing Systems, pages 11895-11907, 2019. 2, 3, 5", + "[40] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. International Conference on Learning Representations, 2021. 2, 3, 8", + "[41] Xuan Su, Jiaming Song, Chenlin Meng, and Stefano Ermon. Dual diffusion implicit bridges for image-to-image translation. arXiv preprint arXiv:2203.08382, 2022. 2, 8, 12", + "[42] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Arsenii Ashukha, Aleksei Silvestrov," + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with fourier convolutions. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2149–2159, 2022. 25", + "[43] Daniel Watson, William Chan, Jonathan Ho, and Mohammad Norouzi. Learning fast samplers for diffusion models by differentiating through sample quality. In International Conference on Learning Representations, 2022. 8", + "[44] Minkai Xu, Lantao Yu, Yang Song, Chence Shi, Stefano Ermon, and Jian Tang. Geodiff: A geometric diffusion model for molecular conformation generation. arXiv preprint arXiv:2203.02923, 2022. 2", + "[45] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2022. 24" + ], + "bbox": [ + 501, + 90, + 893, + 344 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "14306", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2023/On Distillation of Guided Diffusion Models/2a7f0724-3ab3-4aec-b0e2-13e799a34131_model.json b/2023/On Distillation of Guided Diffusion Models/2a7f0724-3ab3-4aec-b0e2-13e799a34131_model.json new file mode 100644 index 0000000000000000000000000000000000000000..80550e5a5e281c588b4e2edae185e30c4c68a3dd --- /dev/null +++ b/2023/On Distillation of Guided Diffusion Models/2a7f0724-3ab3-4aec-b0e2-13e799a34131_model.json @@ -0,0 +1,2959 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.809, + 0.02 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + }, + { + "type": "header", + "bbox": [ + 0.325, + 0.017, + 0.723, + 0.031 + ], + "angle": 0, + "content": "Except for this watermark, it is identical to the accepted version;" + }, + { + "type": "header", + "bbox": [ + 0.294, + 0.032, + 0.755, + 0.047 + ], + "angle": 0, + "content": "the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.267, + 0.068, + 0.704, + 0.088 + ], + "angle": 0, + "content": "On Distillation of Guided Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.136, + 0.119, + 0.26, + 0.137 + ], + "angle": 0, + "content": "Chenlin Meng1" + }, + { + "type": "text", + "bbox": [ + 0.097, + 0.14, + 0.303, + 0.153 + ], + "angle": 0, + "content": "chenlin@cs.stanford.edu" + }, + { + "type": "text", + "bbox": [ + 0.35, + 0.119, + 0.49, + 0.135 + ], + "angle": 0, + "content": "Robin Rombach" + }, + { + "type": "text", + "bbox": [ + 0.339, + 0.14, + 0.5, + 0.153 + ], + "angle": 0, + "content": "robin@stability.ai" + }, + { + "type": "text", + "bbox": [ + 0.567, + 0.119, + 0.66, + 0.137 + ], + "angle": 0, + "content": "Ruiqi Gao3" + }, + { + "type": "text", + "bbox": [ + 0.539, + 0.14, + 0.691, + 0.153 + ], + "angle": 0, + "content": "ruiqig@google.com" + }, + { + "type": "text", + "bbox": [ + 0.727, + 0.119, + 0.874, + 0.137 + ], + "angle": 0, + "content": "Diederik Kingma3" + }, + { + "type": "text", + "bbox": [ + 0.734, + 0.14, + 0.869, + 0.153 + ], + "angle": 0, + "content": "durk@google.com" + }, + { + "type": "text", + "bbox": [ + 0.192, + 0.161, + 0.321, + 0.178 + ], + "angle": 0, + "content": "Stefano Ermon1" + }, + { + "type": "text", + "bbox": [ + 0.164, + 0.183, + 0.352, + 0.195 + ], + "angle": 0, + "content": "ermon@cs.stanford.edu" + }, + { + "type": "text", + "bbox": [ + 0.439, + 0.162, + 0.549, + 0.178 + ], + "angle": 0, + "content": "Jonathan Ho3" + }, + { + "type": "text", + "bbox": [ + 0.402, + 0.183, + 0.589, + 0.196 + ], + "angle": 0, + "content": "jonathanho@google.com" + }, + { + "type": "text", + "bbox": [ + 0.662, + 0.162, + 0.781, + 0.178 + ], + "angle": 0, + "content": "Tim Salimans3" + }, + { + "type": "text", + "bbox": [ + 0.638, + 0.183, + 0.808, + 0.196 + ], + "angle": 0, + "content": "salimans@google.com" + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.203, + 0.282, + 0.222 + ], + "angle": 0, + "content": "\\(^{1}\\)Stanford University" + }, + { + "type": "text", + "bbox": [ + 0.321, + 0.204, + 0.56, + 0.222 + ], + "angle": 0, + "content": "\\(^{2}\\)Stability AI & LMU Munich" + }, + { + "type": "text", + "bbox": [ + 0.597, + 0.204, + 0.851, + 0.222 + ], + "angle": 0, + "content": "3 Google Research, Brain Team" + }, + { + "type": "image", + "bbox": [ + 0.08, + 0.244, + 0.445, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.16, + 0.376, + 0.353, + 0.391 + ], + "angle": 0, + "content": "Text-guided generation (1 step)" + }, + { + "type": "image", + "bbox": [ + 0.446, + 0.244, + 0.669, + 0.333 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.669, + 0.244, + 0.779, + 0.333 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.78, + 0.244, + 0.892, + 0.333 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.571, + 0.336, + 0.768, + 0.351 + ], + "angle": 0, + "content": "Text-guided generation (2 steps)" + }, + { + "type": "image", + "bbox": [ + 0.079, + 0.392, + 0.445, + 0.722 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.16, + 0.725, + 0.359, + 0.738 + ], + "angle": 0, + "content": "Text-guided generation (4 steps)" + }, + { + "type": "image", + "bbox": [ + 0.445, + 0.355, + 0.557, + 0.401 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.558, + 0.356, + 0.668, + 0.401 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.67, + 0.356, + 0.779, + 0.401 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.78, + 0.356, + 0.892, + 0.401 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.558, + 0.405, + 0.783, + 0.419 + ], + "angle": 0, + "content": "Class-conditional generation (1 step)" + }, + { + "type": "image", + "bbox": [ + 0.446, + 0.422, + 0.668, + 0.509 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.67, + 0.422, + 0.779, + 0.509 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.78, + 0.422, + 0.892, + 0.509 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.488, + 0.511, + 0.517, + 0.522 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image_caption", + "bbox": [ + 0.599, + 0.511, + 0.629, + 0.521 + ], + "angle": 0, + "content": "Mask" + }, + { + "type": "image_caption", + "bbox": [ + 0.704, + 0.511, + 0.746, + 0.521 + ], + "angle": 0, + "content": "Result 1" + }, + { + "type": "image_caption", + "bbox": [ + 0.815, + 0.511, + 0.858, + 0.521 + ], + "angle": 0, + "content": "Result 2" + }, + { + "type": "image_caption", + "bbox": [ + 0.59, + 0.524, + 0.75, + 0.537 + ], + "angle": 0, + "content": "Image inpainting (2 steps)" + }, + { + "type": "image", + "bbox": [ + 0.446, + 0.538, + 0.556, + 0.712 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.557, + 0.538, + 0.668, + 0.712 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.668, + 0.538, + 0.781, + 0.712 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.781, + 0.538, + 0.892, + 0.712 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.489, + 0.713, + 0.518, + 0.724 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image_caption", + "bbox": [ + 0.667, + 0.713, + 0.788, + 0.724 + ], + "angle": 0, + "content": "Output (different styles)" + }, + { + "type": "image_caption", + "bbox": [ + 0.562, + 0.725, + 0.783, + 0.739 + ], + "angle": 0, + "content": "Image to image translation (3 steps)" + }, + { + "type": "image_caption", + "bbox": [ + 0.075, + 0.748, + 0.895, + 0.791 + ], + "angle": 0, + "content": "Figure 1. Distilled Stable Diffusion samples generated by our method. Our two-stage distillation approach is able to generate realistic images using only 1 to 4 denoising steps on various tasks. Compared to standard classifier-free guided diffusion models, we reduce the total number of sampling steps by at least \\(20\\times\\)." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.798, + 0.314, + 0.814 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.831, + 0.471, + 0.876 + ], + "angle": 0, + "content": "Classifier-free guided diffusion models have recently been shown to be highly effective at high-resolution image generation, and they have been widely used in large-scale diffusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.8, + 0.896, + 0.891 + ], + "angle": 0, + "content": "frameworks including DALL·E 2, Stable Diffusion and Image. However, a downside of classifier-free guided diffusion models is that they are computationally expensive at inference time since they require evaluating two diffusion models, a class-conditional model and an unconditional model, tens to hundreds of times. To deal with this limitation, we pro" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.887, + 0.374, + 0.901 + ], + "angle": 0, + "content": "*Work partially done during an internship at Google" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14297" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.396 + ], + "angle": 0, + "content": "pose an approach to distilling classifier-free guided diffusion models into models that are fast to sample from: Given a pre-trained classifier-free guided model, we first learn a single model to match the output of the combined conditional and unconditional models, and then we progressively distill that model to a diffusion model that requires much fewer sampling steps. For standard diffusion models trained on the pixel-space, our approach is able to generate images visually comparable to that of the original model using as few as 4 sampling steps on ImageNet 64x64 and CIFAR-10, achieving FID/IS scores comparable to that of the original model while being up to 256 times faster to sample from. For diffusion models trained on the latent-space (e.g., Stable Diffusion), our approach is able to generate high-fidelity images using as few as 1 to 4 denoising steps, accelerating inference by at least 10-fold compared to existing methods on ImageNet 256x256 and LAION datasets. We further demonstrate the effectiveness of our approach on text-guided image editing and inpainting, where our distilled model is able to generate high-quality results using as few as 2-4 denoising steps." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.423, + 0.21, + 0.438 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.448, + 0.473, + 0.87 + ], + "angle": 0, + "content": "Denoising diffusion probabilistic models (DDPMs) [4, 37, 39, 40] have achieved state-of-the-art performance on image generation [22, 26–28, 31], audio synthesis [11], molecular generation [44], and likelihood estimation [10]. Classifier-free guidance [6] further improves the sample quality of diffusion models and has been widely used in large-scale diffusion model frameworks including GLIDE [23], Stable Diffusion [28], DALL-E 2 [26], and Imagen [31]. However, one key limitation of classifier-free guidance is its low sampling efficiency—it requires evaluating two diffusion models tens to hundreds of times to generate one sample. This limitation has hindered the application of classifier-free guidance models in real-world settings. Although distillation approaches have been proposed for diffusion models [33, 38], these approaches are not directly applicable to classifier-free guided diffusion models. To deal with this issue, we propose a two-stage distillation approach to improving the sampling efficiency of classifier-free guided models. In the first stage, we introduce a single student model to match the combined output of the two diffusion models of the teacher. In the second stage, we progressively distill the model learned from the first stage to a fewer-step model using the approach introduced in [33]. Using our approach, a single distilled model is able to handle a wide range of different guidance strengths, allowing for the trade-off between sample quality and diversity efficiently. To sample from our model, we consider existing deterministic samplers in the literature [33, 38] and further propose a stochastic sampling process." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.472, + 0.902 + ], + "angle": 0, + "content": "Our distillation framework can not only be applied to standard diffusion models trained on the pixel-space [4, 36, 39]," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.426 + ], + "angle": 0, + "content": "but also diffusion models trained on the latent-space of an autoencoder [28, 35] (e.g., Stable Diffusion [28]). For diffusion models directly trained on the pixel-space, our experiments on ImageNet 64x64 and CIFAR-10 show that the proposed distilled model can generate samples visually comparable to that of the teacher using only 4 steps and is able to achieve comparable FID/IS scores as the teacher model using as few as 4 to 16 steps on a wide range of guidance strengths (see Fig. 2). For diffusion model trained on the latent-space of an encoder [28, 35], our approach is able to achieve comparable visual quality to the base model using as few as 1 to 4 sampling steps (at least \\(10 \\times\\) fewer steps than the base model) on ImageNet \\(256 \\times 256\\) and LAION \\(512 \\times 512\\), matching the performance of the teacher (as evaluated by FID) with only 2-4 sampling steps. To the best of our knowledge, our work is the first to demonstrate the effectiveness of distillation for both pixel-space and latent-space classifier-free diffusion models. Finally, we apply our method to text-guided image inpainting and text-guided image editing tasks [20], where we reduce the total number of sampling steps to as few as 2-4 steps, demonstrating the potential of the proposed framework in style-transfer and image-editing applications [20, 41]." + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.439, + 0.892, + 0.58 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.589, + 0.895, + 0.66 + ], + "angle": 0, + "content": "Figure 2. Class-conditional samples from our two-stage (deterministic) approach on ImageNet 64x64 for diffusion models trained on the pixel-space. By varying the guidance weight \\( w \\), our distilled model is able to trade-off between sample diversity and quality, while achieving good results using as few as one sampling step." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.673, + 0.798, + 0.69 + ], + "angle": 0, + "content": "2. Background on diffusion models" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.701, + 0.892, + 0.762 + ], + "angle": 0, + "content": "Given samples \\(\\mathbf{x}\\) from a data distribution \\(p_{\\mathrm{data}}(\\mathbf{x})\\), noise scheduling functions \\(\\alpha_{t}\\) and \\(\\sigma_{t}\\), we train a diffusion model \\(\\hat{\\mathbf{x}}_{\\theta}\\), with parameter \\(\\theta\\), via minimizing the weighted mean squared error [4, 36, 39, 40]" + }, + { + "type": "equation", + "bbox": [ + 0.513, + 0.775, + 0.894, + 0.794 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} _ {t \\sim U [ 0, 1 ], \\mathbf {x} \\sim p _ {\\text {d a t a}} (\\mathbf {x}), \\mathbf {z} _ {t} \\sim q \\left(\\mathbf {z} _ {t} \\mid \\mathbf {x}\\right)} \\left[ \\omega \\left(\\lambda_ {t}\\right) \\right| | \\hat {\\mathbf {x}} _ {\\boldsymbol {\\theta}} \\left(\\mathbf {z} _ {t}\\right) - \\mathbf {x} | | _ {2} ^ {2} ], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.808, + 0.892, + 0.854 + ], + "angle": 0, + "content": "where \\(\\lambda_{t} = \\log [\\alpha_{t}^{2} / \\sigma_{t}^{2}]\\) is a signal-to-noise ratio [10], \\(q(\\mathbf{z}_t|\\mathbf{x}) = \\mathcal{N}(\\mathbf{z}_t;\\alpha_t\\mathbf{x},\\sigma_t^2\\mathbf{I})\\) and \\(\\omega (\\lambda_t)\\) is a pre-specified weighting function [10]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.895, + 0.902 + ], + "angle": 0, + "content": "Once the diffusion model \\(\\hat{\\mathbf{x}}_{\\theta}\\) is trained, one can use discrete-time DDIM sampler [38] to sample from the model. Specifically, the DDIM sampler starts with \\(\\mathbf{z}_1 \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{I})\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14298" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.078, + 0.092, + 0.232, + 0.107 + ], + "angle": 0, + "content": "and updates as follows" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.115, + 0.47, + 0.147 + ], + "angle": 0, + "content": "\\[\n\\mathbf {z} _ {s} = \\alpha_ {s} \\hat {\\mathbf {x}} _ {\\boldsymbol {\\theta}} (\\mathbf {z} _ {t}) + \\sigma_ {s} \\frac {\\mathbf {z} _ {t} - \\alpha_ {t} \\hat {\\mathbf {x}} _ {\\boldsymbol {\\theta}} (\\mathbf {z} _ {t})}{\\sigma_ {t}}, s = t - 1 / N \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.154, + 0.47, + 0.184 + ], + "angle": 0, + "content": "with \\(N\\) the total number of sampling steps. The final sample will then be generated using \\(\\hat{\\mathbf{x}}_{\\theta}(\\mathbf{z}_0)\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.184, + 0.471, + 0.41 + ], + "angle": 0, + "content": "Classifier-free guidance Classifier-free guidance [6] is an effective approach shown to significantly improve the sample quality of class-conditioned diffusion models, and has been widely used in large-scale diffusion models including GLIDE [23], Stable Diffusion [28], DALL-E 2 [26] and Imagen [31]. Specifically, it introduces a guidance weight parameter \\( w \\in \\mathbb{R}^{\\geq 0} \\) to trade-off between sample quality and diversity. To generate a sample, classifier-free guidance evaluates both a conditional diffusion model \\( \\hat{\\mathbf{x}}_{c,\\theta} \\), where \\( c \\) is the context (e.g., class label, text prompt) to be conditioned on, and a jointly trained unconditional diffusion model \\( \\hat{\\mathbf{x}}_{\\theta} \\) at each update step, using \\( \\hat{\\mathbf{x}}_{\\theta}^{w} = (1 + w)\\hat{\\mathbf{x}}_{c,\\theta} - w\\hat{\\mathbf{x}}_{\\theta} \\) as the model prediction in Eq. (2). As each sampling update requires evaluating two diffusion models, sampling with classifier-free guidance is often expensive [6]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.411, + 0.472, + 0.516 + ], + "angle": 0, + "content": "Progressive distillation Our approach is inspired by progressive distillation [33], an effective method for improving the sampling speed of (unguided) diffusion models by repeated distillation. Until now, this method could not be directly applied to distilling classifier-free guided models or studied for samplers other than the deterministic DDIM sampler [33, 38]. In this paper we resolve these shortcomings." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.516, + 0.472, + 0.666 + ], + "angle": 0, + "content": "Latent diffusion models (LDMs) [21,24,28,35] increase the training and inference efficiency of diffusion models (directly learned on the pixel-space) by modeling images in the latent space of a pre-trained regularized autoencoder, where the latent representations are usually of lower dimensionality than the pixel-space. Latent diffusion models can be considered as an alternative to cascaded diffusion approaches [5], which rely on one or more super-resolution diffusion models to scale up a low-dimensional image to the desired target resolution." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.667, + 0.47, + 0.713 + ], + "angle": 0, + "content": "In this work, we will apply our distillation framework to classifier-free guided diffusion models learned on both pixel-space [4, 36, 39] and latent-space [21, 24, 28, 35]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.725, + 0.391, + 0.743 + ], + "angle": 0, + "content": "3. Distilling a guided diffusion model" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.47, + 0.856 + ], + "angle": 0, + "content": "In the following, we discuss our approach for distilling a classifier-free guided diffusion model [6] into a student model that requires fewer steps to sample from. Using a single distilled model conditioned on the guidance strength, our model can capture a wide range of classifier-free guidance levels, allowing for the trade-off between sample quality and diversity efficiently." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.472, + 0.902 + ], + "angle": 0, + "content": "Given a trained guided model \\([\\hat{\\mathbf{x}}_{c,\\theta},\\hat{\\mathbf{x}}_{\\theta}]\\) (teacher) either on the pixel-space or latent-space, our approach can be decomposed into two stages." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.092, + 0.698, + 0.107 + ], + "angle": 0, + "content": "3.1. Stage-one distillation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.114, + 0.895, + 0.248 + ], + "angle": 0, + "content": "In the first stage, we introduce a student model \\(\\hat{\\mathbf{x}}_{\\eta_1}(\\mathbf{z}_t,w)\\), with learnable parameter \\(\\pmb{\\eta}_{1}\\), to match the output of the teacher at any time-step \\(t\\in [0,1]\\). The student model can either be a continuous-time model [40] or a discrete-time model [4, 38] depending on whether the teacher model is discrete or continuous. For simplicity, in the following discussion, we assume both the student and teacher models are continuous as the algorithm for discrete models is almost identical." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.25, + 0.895, + 0.388 + ], + "angle": 0, + "content": "A key functionality of classifier-free guidance [6] is its ability to easily trade-off between sample quality and diversity, which is controlled by a \"guidance strength\" parameter. This property has demonstrated utility in real-world applications [6,23,26,28,31], where the optimal \"guidance strength\" is often a user preference. Thus, we would also want our distilled model to maintain this property. Given a range of guidance strengths \\(\\left[w_{\\min},w_{\\max}\\right]\\) we are interested in, we optimize the student model using the following objective" + }, + { + "type": "equation", + "bbox": [ + 0.507, + 0.396, + 0.893, + 0.427 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} _ {w \\sim p _ {w}, t \\sim U [ 0, 1 ], \\mathbf {x} \\sim p _ {\\mathrm {d a t a}} (\\mathbf {x})} \\left[ \\omega \\left(\\lambda_ {t}\\right) \\| \\hat {\\mathbf {x}} _ {\\boldsymbol {\\eta} _ {1}} \\left(\\mathbf {z} _ {t}, w\\right) - \\hat {\\mathbf {x}} _ {\\boldsymbol {\\theta}} ^ {w} \\left(\\mathbf {z} _ {t}\\right) \\| _ {2} ^ {2} \\right], \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.435, + 0.894, + 0.527 + ], + "angle": 0, + "content": "where \\(\\hat{\\mathbf{x}}_{\\pmb{\\theta}}^{w}(\\mathbf{z}_{t}) = (1 + w)\\hat{\\mathbf{x}}_{c,\\pmb{\\theta}}(\\mathbf{z}_{t}) - w\\hat{\\mathbf{x}}_{\\pmb{\\theta}}(\\mathbf{z}_{t}), \\mathbf{z}_{t} \\sim q(\\mathbf{z}_{t}|\\mathbf{x})\\) and \\(p_w(w) = U[w_{\\min}, w_{\\max}]\\). Note that here, our distilled model \\(\\hat{\\mathbf{x}}_{\\pmb{\\eta}_1}(\\mathbf{z}_t, w)\\) is also conditioned on the context \\(c\\) (e.g., text prompt), but we drop the notation \\(c\\) in the paper for simplicity. We provide the detailed training algorithm in Algorithm 1 in the supplement." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.528, + 0.895, + 0.754 + ], + "angle": 0, + "content": "To incorporate the guidance weight \\( w \\), we introduce a \\( w \\)-conditioned model, where \\( w \\) is fed as an input to the student model. To better capture the feature, we apply Fourier embedding to \\( w \\), which is then incorporated into the diffusion model backbone in a way similar to how the time-step was incorporated in [10, 33]. As initialization plays a key role in the performance [33], we initialize the student model with the same parameters as the conditional model of the teacher, except for the newly introduced parameters related to \\( w \\)-conditioning. The model architecture we use is a U-Net model similar to the ones used in [6] for pixel-space diffusion models and [1, 28] for latent-space diffusion models. We use the same number of channels and attention as used in [6] and the open-sourced Stable Diffusion repository* for our experiments. We provide more details in the supplement." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.763, + 0.699, + 0.779 + ], + "angle": 0, + "content": "3.2. Stage-two distillation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.787, + 0.894, + 0.879 + ], + "angle": 0, + "content": "In the second stage, we consider a discrete time-step scenario and progressively distill the learned model from the first-stage \\(\\hat{\\mathbf{x}}_{\\eta_1}(\\mathbf{z}_t,w)\\) into an fewer-step student model \\(\\hat{\\mathbf{x}}_{\\eta_2}(\\mathbf{z}_t,w)\\) with learnable parameter \\(\\eta_{2}\\), by halving the number of sampling steps each time. Letting \\(N\\) denote the number of sampling steps, given \\(w\\sim U[w_{\\min},w_{\\max}]\\) and" + }, + { + "type": "page_footnote", + "bbox": [ + 0.517, + 0.888, + 0.866, + 0.901 + ], + "angle": 0, + "content": "*https://github.com/CompVis/stable-diffusion" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14299" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.089, + 0.177, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.168, + 0.294, + 0.277, + 0.306 + ], + "angle": 0, + "content": "(a) 2 denoising steps" + }, + { + "type": "image", + "bbox": [ + 0.178, + 0.089, + 0.37, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.371, + 0.089, + 0.616, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.43, + 0.294, + 0.541, + 0.306 + ], + "angle": 0, + "content": "(b) 4 denoising steps" + }, + { + "type": "image", + "bbox": [ + 0.617, + 0.089, + 0.878, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.693, + 0.294, + 0.803, + 0.306 + ], + "angle": 0, + "content": "(c) 8 denoising steps" + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.318, + 0.892, + 0.346 + ], + "angle": 0, + "content": "Figure 4. Text-guided generation on LAION (512x512) using our distilled Stable Diffusion model. Our model is able to generate high-quality image samples using 2, 4 or 8 denoising steps, significantly improving the inference efficiency of Stable Diffusion." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.348, + 0.891, + 0.51 + ], + "angle": 0, + "content": "
w=0w=0.3w=1w=4
MethodFID (↓)IS (↑)FID (↓)IS (↑)FID (↓)IS (↑)FID (↓)IS (↑)
Ours 1-step (D/S)22.74 / 26.9125.51 / 23.5514.85 / 18.4837.09 / 33.307.54 / 8.9275.19 / 67.8018.72 / 17.85157.46 / 148.97
Ours 4-step (D/S)4.14 / 3.9146.64 / 48.922.17 / 2.2469.64 / 73.737.95 / 8.51128.98 / 135.3626.45 / 27.33207.45 / 216.56
Ours 8-step (D/S)2.79 / 2.4450.72 / 55.032.05 / 2.3176.01 / 83.009.33 / 10.56136.47 / 147.3926.62 / 27.84203.47 / 219.89
Ours 16-step (D/S)2.44 / 2.1052.53 / 57.812.20 / 2.5679.47 / 87.509.99 / 11.63139.11 / 153.1726.53 / 27.69204.13 / 218.70
Single-w 1-step19.6124.0011.7036.956.6474.4119.857170.69
Single-w 4-step4.7938.772.3462.088.23118.5227.75219.64
Single-w 8-step3.3942.132.3268.769.69125.2027.67218.08
Single-w 16-step2.9743.632.5670.9710.34127.7027.40216.52
DDIM 16x2-step [38]7.6837.605.3360.839.53112.7521.56195.17
DDIM 32x2-step [38]5.0340.937.479.339.26126.2223.03213.23
DDIM 64x2-step [38]3.7443.165.529.519.53133.1723.64217.88
Teacher (DDIM 1024x2-step)2.9244.812.3674.839.84139.5023.94224.74
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.52, + 0.893, + 0.577 + ], + "angle": 0, + "content": "Table 1. ImageNet 64x64 distillation results for pixel-space diffusion models (\\( w = 0 \\) refers to non-guided models). For our method, \\( D \\) and \\( S \\) stand for deterministic and stochastic sampler respectively. We observe that training the model conditioned on a guidance interval \\( w \\in [0,4] \\) performs comparably with training a model on a fixed \\( w \\) (see Single-\\( w \\)). Our approach significantly outperforms DDIM when using fewer steps, and is able to match the teacher performance using as few as 8 to 16 steps." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.59, + 0.471, + 0.756 + ], + "angle": 0, + "content": "\\(t \\in \\{1, \\dots, N\\}\\), we train the student model to match the output of two-step DDIM sampling of the teacher (i.e., from \\(t/N\\) to \\(t - 0.5/N\\) and from \\(t - 0.5/N\\) to \\(t - 1/N\\)) in one step, following the approach of [33]. After distilling the \\(2N\\) steps in the teacher model to \\(N\\) steps in the student model, we can use the \\(N\\)-step student model as the new teacher model, repeat the same procedure, and distill the teacher model into a \\(N/2\\)-step student model. At each step, we initialize the student model with the parameters of the teacher. We provide the training algorithm and extra details in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.77, + 0.458, + 0.788 + ], + "angle": 0, + "content": "3.3. \\(N\\)-step deterministic and stochastic sampling" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Once the model \\(\\hat{\\mathbf{x}}_{\\eta_2}\\) is trained, given a specified guidance strength \\(w \\in [w_{\\min}, w_{\\max}]\\), we can perform sampling via the DDIM update rule in Eq. (2). We note that given the distilled model \\(\\hat{\\mathbf{x}}_{\\eta_2}\\), this sampling procedure is deterministic given the initialization \\(\\mathbf{z}_1^w\\). In fact, we can also perform \\(N\\)-step stochastic sampling: We apply one deterministic sampling step with two-times the original step-length (i.e.," + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.59, + 0.892, + 0.666 + ], + "angle": 0, + "content": "the same as a \\(N/2\\)-step deterministic sampler) and then perform one stochastic step backward (i.e., perturb with noise) using the original step-length, a process inspired by [9]. With \\(\\mathbf{z}_1^w \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{I})\\), we use the following update rule when \\(t > 1/N\\)" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.677, + 0.892, + 0.71 + ], + "angle": 0, + "content": "\\[\n\\mathbf {z} _ {k} ^ {w} = \\alpha_ {k} \\hat {\\mathbf {x}} _ {\\boldsymbol {\\eta} _ {2}} \\left(\\mathbf {z} _ {t} ^ {w}\\right) + \\sigma_ {k} \\frac {\\mathbf {z} _ {t} ^ {w} - \\alpha_ {t} \\hat {\\mathbf {x}} _ {\\boldsymbol {\\eta} _ {2}} ^ {w} \\left(\\mathbf {z} _ {t}\\right)}{\\sigma_ {t}}, \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.712, + 0.892, + 0.73 + ], + "angle": 0, + "content": "\\[\n\\text {w h e r e} \\mathbf {z} _ {s} ^ {w} = \\left(\\alpha_ {s} / \\alpha_ {k}\\right) \\mathbf {z} _ {k} ^ {w} + \\sigma_ {s | k} \\epsilon , \\epsilon \\sim \\mathcal {N} (\\mathbf {0}, \\mathbf {I}); \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.732, + 0.892, + 0.765 + ], + "angle": 0, + "content": "\\[\n\\mathbf {z} _ {h} ^ {w} = \\alpha_ {h} \\hat {\\mathbf {x}} _ {\\boldsymbol {\\eta} _ {2}} \\left(\\mathbf {z} _ {s} ^ {w}\\right) + \\sigma_ {h} \\frac {\\mathbf {z} _ {s} ^ {w} - \\alpha_ {s} \\hat {\\mathbf {x}} _ {\\boldsymbol {\\eta} _ {2}} ^ {w} \\left(\\mathbf {z} _ {s} ^ {w}\\right)}{\\sigma_ {s}}, \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.767, + 0.892, + 0.785 + ], + "angle": 0, + "content": "\\[\nw h e r e \\mathbf {z} _ {k} ^ {w} = \\left(\\alpha_ {k} / \\alpha_ {h}\\right) \\mathbf {z} _ {h} ^ {w} + \\sigma_ {k | h} \\epsilon , \\epsilon \\sim \\mathcal {N} (\\mathbf {0}, \\mathbf {I}). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.893, + 0.902 + ], + "angle": 0, + "content": "In the above equations, \\( h = t - 3 / N \\), \\( k = t - 2 / N \\), \\( s = t - 1 / N \\) and \\( \\sigma_{a|b}^{2} = (1 - e^{\\lambda_{a} - \\lambda_{b}})\\sigma_{a}^{2} \\). When \\( t = 1 / N \\), we use deterministic update Eq. (2) to obtain \\( \\mathbf{z}_0^w \\) from \\( \\mathbf{z}_{1 / N}^w \\). We provide an illustration of the process in Fig. 5, where the number of denoising steps is 4. We note that compared to the deterministic sampler, performing stochastic sampling requires evaluating the model at slightly different time-steps," + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14300" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.085, + 0.076, + 0.47, + 0.18 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.189, + 0.47, + 0.218 + ], + "angle": 0, + "content": "Figure 5. Sampling procedures of the distilled model where the number of denoising steps is 4." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.225, + 0.469, + 0.272 + ], + "angle": 0, + "content": "and would require small modifications to training algorithm for the edge cases. We provide the algorithm and more details in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.285, + 0.21, + 0.302 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.31, + 0.471, + 0.415 + ], + "angle": 0, + "content": "In this section, we evaluate the performance of our distillation approach on pixel-space diffusion models (i.e. DDPM [4]) and latent-space diffusion models (i.e. Stable Diffusion [28]). We further apply our approach to text-guided image editing and inpainting tasks. Experiments show that our approach is able to achieve competitive performance while using as few as 2-4 steps on all tasks." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.425, + 0.43, + 0.441 + ], + "angle": 0, + "content": "4.1. Distillation for pixel-space guided models" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.448, + 0.471, + 0.478 + ], + "angle": 0, + "content": "In this experiment, we consider class-conditional diffusion models directly trained on the pixel-space [4,6,33]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.479, + 0.471, + 0.763 + ], + "angle": 0, + "content": "Settings We focus on ImageNet 64x64 [30] and CIFAR-10 [12] as higher-resolution image generation in this scenario often relies on combining with other super-resolution techniques [5, 31]. We explore different ranges for the guidance weight and observe that all ranges work comparably and therefore use \\([w_{min}, w_{max}] = [0, 4]\\) for the experiments. The baselines we consider include DDPM ancestral sampling [4] and DDIM [38]. The teacher model we use is a \\(1024 \\times 2\\)-step DDIM model, where the conditional and unconditional components both use 1024 DDIM denoising steps. To better understand how the guidance weight \\(w\\) should be incorporated, we also include models trained using a single fixed \\(w\\) as a baseline. We use the same pre-trained teacher model for all the methods for fair comparisons. Following [4, 6, 39], we use a U-Net [29, 39] architecture for the baselines, and the same U-Net backbone with the introduced \\(w\\)-embedding for our two-step student models (see Sec. 3). Following [33], we use a v-prediction model for both datasets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Results We report the performance as evaluated in FID [3] and Inception scores (IS) [32] for all approaches on ImageNet 64x64 in Fig. 6 and Tab. 1 and provide extended ImageNet 64x64 and CIFAR-10 results in the supplement. We observe that our distilled model is able to match a teacher guided DDIM model with \\(1024 \\times 2\\) sampling steps using only 4-16 steps, achieving a speedup for up to \\(256 \\times\\). We emphasize that, using our approach, a single distilled model is able to match the teacher performance on a wide range" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "of guidance strengths. This has not been achieved by any previous methods." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.133, + 0.859, + 0.149 + ], + "angle": 0, + "content": "4.2. Distillation for latent-space guided models" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.156, + 0.894, + 0.261 + ], + "angle": 0, + "content": "After demonstrating the effectiveness of our method on pixel-space class-guided diffusion models in Sec. 4.1, we now expand its scope to latent-space diffusion models. In the following sections, we show the effectiveness of our approach on Latent Diffusion [28] on a variety of tasks, including class-conditional generation, text-to-image generation, image inpainting and text-guided style-transfer [20]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.262, + 0.894, + 0.352 + ], + "angle": 0, + "content": "In the following experiments, we use the open-sourced latent-space diffusion models [28] as the teacher models. As v-prediction teacher model tends to perform better than \\(\\epsilon\\)-prediction model, we fine-tune the open-sourced \\(\\epsilon\\)-prediction models into v-prediction teacher models. We provide more details in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.373, + 0.753, + 0.388 + ], + "angle": 0, + "content": "4.2.1 Class-conditional generation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.397, + 0.894, + 0.502 + ], + "angle": 0, + "content": "In this section, we apply our method to a class-conditional latent diffusion model pre-trained on ImageNet \\(256 \\times 256\\). We start from the DDIM teacher model with 512 sampling steps, and use the output as the target to train our distilled model. We use a batch size of 512 and uniformly sample the guidance strength \\(w \\in [w_{min} = 0, w_{max} = 14]\\) during training." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.503, + 0.894, + 0.655 + ], + "angle": 0, + "content": "Results Empirically, we find that our distilled model is able to match the performance of the teacher model (originally trained on 1000 steps) in terms of FID scores while using only 2 or 4 sampling steps. We also achieve significantly better performance than DDIM when using 1-4 sampling steps (see Fig. 11). Qualitatively, we find that samples synthesized using a single denoising step still yield satisfying results, while the baseline fails to generate images with meaningful contents. We provide extra samples in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.655, + 0.894, + 0.745 + ], + "angle": 0, + "content": "Similar to the pixel-based results in Fig. 6, we also observe the trade-off between sampling quality and diversity as measured by FID and Inception Score for our distilled latent diffusion model. Following Kynkänniemi et al [13], we further compute improved precision and recall metrics for this experiment in the appendix." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.765, + 0.761, + 0.781 + ], + "angle": 0, + "content": "4.2.2 Text-guided image generation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.789, + 0.894, + 0.866 + ], + "angle": 0, + "content": "In this section, we focus on the text-guided Stable Diffusion model pretrained on subsets \\(\\dagger\\) of LAION-5B [34] at a resolution \\(512 \\times 512\\). We then follow our two-stage approach introduced in Sec. 3 and distill the guided model in 3000 gradient updates into a \\(w\\)-conditioned model using" + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.876, + 0.894, + 0.901 + ], + "angle": 0, + "content": "\\(\\dagger\\)https://github.com/CompVis/stable-diffusion/blob/main/Stable_Diffusion_v1_model/Card.md" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "14301" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.082, + 0.252, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.257, + 0.081, + 0.408, + 0.154 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.411, + 0.081, + 0.561, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.565, + 0.082, + 0.716, + 0.157 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.72, + 0.082, + 0.871, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.158, + 0.253, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.168, + 0.229, + 0.204, + 0.239 + ], + "angle": 0, + "content": "\\(w = 0\\)" + }, + { + "type": "image", + "bbox": [ + 0.256, + 0.158, + 0.407, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.319, + 0.228, + 0.365, + 0.239 + ], + "angle": 0, + "content": "\\(w = 0.3\\)" + }, + { + "type": "image", + "bbox": [ + 0.411, + 0.159, + 0.561, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.48, + 0.228, + 0.516, + 0.239 + ], + "angle": 0, + "content": "\\(w = 1\\)" + }, + { + "type": "image", + "bbox": [ + 0.565, + 0.159, + 0.716, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.638, + 0.228, + 0.674, + 0.239 + ], + "angle": 0, + "content": "\\(w = 2\\)" + }, + { + "type": "image", + "bbox": [ + 0.72, + 0.16, + 0.871, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.787, + 0.228, + 0.823, + 0.239 + ], + "angle": 0, + "content": "\\(w = 4\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.252, + 0.895, + 0.294 + ], + "angle": 0, + "content": "Figure 6. ImageNet \\(64 \\times 64\\) sample quality evaluated by FID and IS scores. Our distilled model significantly outperforms the DDPM and DDIM baselines, and is able to match the performance of the teacher using as few as 8 to 16 steps. By varying \\(w\\), a single distilled model is able to capture the trade-off between sample diversity and quality." + }, + { + "type": "image_caption", + "bbox": [ + 0.121, + 0.301, + 0.305, + 0.311 + ], + "angle": 0, + "content": "Distilled Text-to-Image samples (4 steps)" + }, + { + "type": "image", + "bbox": [ + 0.081, + 0.312, + 0.344, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.398, + 0.302, + 0.573, + 0.312 + ], + "angle": 0, + "content": "Native Text-to-Image samples (4 steps)" + }, + { + "type": "image", + "bbox": [ + 0.354, + 0.312, + 0.617, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.671, + 0.301, + 0.847, + 0.311 + ], + "angle": 0, + "content": "Native Text-to-Image samples (8 steps)" + }, + { + "type": "image", + "bbox": [ + 0.627, + 0.312, + 0.892, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.426, + 0.894, + 0.482 + ], + "angle": 0, + "content": "Figure 7. Text-guided Stable Diffusion results. We distill the public Stable Diffusion model using the proposed pipeline, arriving at a model that achieves high sample quality using only four denoising steps (left). When sampling from the original model using four DDIM steps, the generated samples have clear artifacts (middle). When using eight DDIM steps, the results get better (right), but are still blurry and less consistent than the distilled results using fewer steps. More samples are provided in Fig. 4." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.494, + 0.473, + 0.645 + ], + "angle": 0, + "content": "\\(w \\in [w_{min} = 2, w_{max} = 14]\\), and a batch size of 512. Although we can condition on a broader range of \\(w\\) for the distilled (student) model, the utility remains unclear as we typically do not exceed the normal guidance range when sampling with the teacher model. The final model is obtained by applying progressive distillation for 2000 training steps per stage, except when for the low-step regime of 1,2, and 4 steps, where we train for 20000 gradient updates. A detailed analysis of the convergence properties of this model in the supplement." + }, + { + "type": "table", + "bbox": [ + 0.137, + 0.646, + 0.409, + 0.704 + ], + "angle": 0, + "content": "
Method2-step4-step8-step
DPM [16]98.9/0.2034.3/0.2931.7/0.32
DPM++ [18]98.8/0.2034.1/0.2925.6/0.32
Ours37.3/0.2726.0/0.3026.9/0.30
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.706, + 0.471, + 0.776 + ], + "angle": 0, + "content": "Table 2. FID/CLIP scores on LAION 512X512 (\\( w = 8.0 \\)). We point out that DPM and DPM++ use both the conditional and unconditional components for sampling. Depending on the implementation, this either requires higher peak memory or two times more sampling steps." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Results We present samples in Fig. 4. We evaluate the resulting model both qualitatively and quantitatively. For the latter analysis, we follow [31] and evaluate CLIP [25] and FID scores to assess text-image alignment and quality, respectively. We use the open-sourced ViT-g/14 [7] CLIP model for evaluation. The quantitative results in Fig. 10 show that our method can significantly increase the performance in both metrics over DDIM sampling on the base model for" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.494, + 0.895, + 0.797 + ], + "angle": 0, + "content": "2 and 4 sampling steps. For 8 steps, these metrics do not show a significant difference. However, when considering the corresponding samples in Fig. 7 we can observe a stark difference in terms of visual image quality. In contrast to the 8-step DDIM samples from the original model, the distilled samples are sharper and more coherent. We hypothesize that FID and CLIP do not fully capture these differences in our evaluation setting on COCO2017 [14], where we used 5000 random captions from the validation set. We further compute the FID and CLIP scores for our distilled LAION \\(512 \\times 512\\) model and compare them with the DPM [16] and \\(\\mathrm{DPM}++\\) [18] solver in Tab. 2. We observe that our method is able to achieve significantly better performance when the denoising step is 2 or 4. Furthermore, we stress that stage-one of our method already decreases the number of function evaluations by a factor of 2, as we distill the classifier-free guidance step into a single model. Depending on the exact implementation (batched vs. sequential network evaluation), this either decreases peak memory or sampling time compared to existing solvers [16, 18, 38]." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.817, + 0.83, + 0.832 + ], + "angle": 0, + "content": "4.2.3 Text-guided image-to-image translation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.895, + 0.903 + ], + "angle": 0, + "content": "In this section, we perform experiments on text-guided image-to-image translation with SDEdit [20] using our distilled model from Sec. 4.2.2. Following SDEdit [20], we perform stochastic encoding in the latent space, but instead" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14302" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.081, + 0.072, + 0.891, + 0.137 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.106, + 0.139, + 0.14, + 0.15 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image_caption", + "bbox": [ + 0.171, + 0.139, + 0.237, + 0.15 + ], + "angle": 0, + "content": "Watercolor" + }, + { + "type": "image_caption", + "bbox": [ + 0.253, + 0.139, + 0.316, + 0.15 + ], + "angle": 0, + "content": "Art station" + }, + { + "type": "image_caption", + "bbox": [ + 0.33, + 0.139, + 0.399, + 0.15 + ], + "angle": 0, + "content": "Photograph" + }, + { + "type": "image_caption", + "bbox": [ + 0.469, + 0.139, + 0.503, + 0.15 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image_caption", + "bbox": [ + 0.609, + 0.139, + 0.676, + 0.15 + ], + "angle": 0, + "content": "Watercolor" + }, + { + "type": "image_caption", + "bbox": [ + 0.75, + 0.139, + 0.866, + 0.15 + ], + "angle": 0, + "content": "Art station painting" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.152, + 0.89, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.116, + 0.232, + 0.15, + 0.244 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image_caption", + "bbox": [ + 0.273, + 0.235, + 0.297, + 0.245 + ], + "angle": 0, + "content": "Girl" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.246, + 0.89, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.097, + 0.338, + 0.17, + 0.35 + ], + "angle": 0, + "content": "Oil painting" + }, + { + "type": "image_caption", + "bbox": [ + 0.224, + 0.33, + 0.249, + 0.34 + ], + "angle": 0, + "content": "Boy" + }, + { + "type": "image_caption", + "bbox": [ + 0.214, + 0.34, + 0.255, + 0.35 + ], + "angle": 0, + "content": "Anime" + }, + { + "type": "image_caption", + "bbox": [ + 0.304, + 0.338, + 0.365, + 0.348 + ], + "angle": 0, + "content": "3D render" + }, + { + "type": "image_caption", + "bbox": [ + 0.416, + 0.338, + 0.456, + 0.349 + ], + "angle": 0, + "content": "Anime" + }, + { + "type": "image_caption", + "bbox": [ + 0.504, + 0.338, + 0.566, + 0.349 + ], + "angle": 0, + "content": "3D render" + }, + { + "type": "image_caption", + "bbox": [ + 0.624, + 0.329, + 0.654, + 0.339 + ], + "angle": 0, + "content": "Male" + }, + { + "type": "image_caption", + "bbox": [ + 0.607, + 0.339, + 0.666, + 0.35 + ], + "angle": 0, + "content": "Van Gogh" + }, + { + "type": "image_caption", + "bbox": [ + 0.704, + 0.338, + 0.771, + 0.349 + ], + "angle": 0, + "content": "Watercolor" + }, + { + "type": "image_caption", + "bbox": [ + 0.806, + 0.338, + 0.871, + 0.349 + ], + "angle": 0, + "content": "Art station" + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.363, + 0.893, + 0.391 + ], + "angle": 0, + "content": "Figure 8. Text-guided image-to-image translation [20] with the distilled Stable Diffusion model (3 denoising steps). We observe that our model is able to generate high-quality and faithful outputs using only 3 denoising steps." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.395, + 0.89, + 0.486 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.128, + 0.487, + 0.153, + 0.496 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.497, + 0.89, + 0.564 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.564, + 0.27, + 0.646 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.564, + 0.464, + 0.646 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.658, + 0.47, + 0.755 + ], + "angle": 0, + "content": "Figure 10. FID and CLIP ViT-g/14 score for text-to-image generation at \\(512 \\times 512\\) px using the distilled Stable Diffusion model. The results are evaluated on 5000 captions from the COCO2017 [14] validation set. Our distilled latent diffusion model is able to generate high-quality image samples using significantly less sampling steps than the original model while achieving similar or better FID and CLIP scores, especially in the low-step regime." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.471, + 0.901 + ], + "angle": 0, + "content": "use the deterministic sampler of the distilled model to perform deterministic decoding. We consider input image and text of various kinds and provide qualitative results in Fig. 8. We observe that our distilled model generates high-quality style-transfer results using as few as 3 denoising steps. We provide more analysis on the trade-off between sample quality, controllability and efficiency in the supplement." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.564, + 0.688, + 0.646 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.695, + 0.564, + 0.888, + 0.646 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.658, + 0.893, + 0.741 + ], + "angle": 0, + "content": "Figure 11. FID and Inception Score for class-conditional image generation on ImageNet \\((256 \\times 256)\\) with distilled latent diffusion. The results are evaluated on 5000 samples. Our distilled latent diffusion model is able to generate high-quality image samples using significantly less sampling steps (up to a factor of 16) than the original model while achieving similar or better FID scores." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.768, + 0.673, + 0.783 + ], + "angle": 0, + "content": "4.2.4 Image inpainting" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.791, + 0.892, + 0.867 + ], + "angle": 0, + "content": "In this section, we apply our approach to a pre-trained image inpainting latent diffusion model. We use the open-source Stable Diffusion Inpainting\\* image-inpainting model. This model is a fine-tuned version of the pure text-to-image Stable Diffusion model from above, where additional input channels" + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.876, + 0.894, + 0.901 + ], + "angle": 0, + "content": "https://huggingface.co/runwayml/stable-diffusion-inpainting" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14303" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.089, + 0.468, + 0.161 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.171, + 0.472, + 0.242 + ], + "angle": 0, + "content": "Figure 12. Style transfer comparison on ImageNet 64x64 for pixel-space models. For our approach, we use a distilled encoder and decoder. For the baseline, we encode and decode using DDIM. We use \\( w = 0 \\) and 16 sampling steps for both the encoder and decoder. We observe that our method achieves more realistic outputs." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.269, + 0.407, + 0.283 + ], + "angle": 0, + "content": "were added to process masks and masked images." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.283, + 0.471, + 0.405 + ], + "angle": 0, + "content": "We use the same distillation algorithm as used in the previous section. For training, we start from the v-prediction teacher model sampled with 512 DDIM steps, and use the output as the target to optimize our student model. We present qualitative results in Fig. 9, demonstrating the potential of our method for fast, real-world image editing applications. For additional training details and a quantitative evaluation, see the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.415, + 0.388, + 0.431 + ], + "angle": 0, + "content": "4.3. Progressive distillation for encoding" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.439, + 0.472, + 0.664 + ], + "angle": 0, + "content": "In this experiment, we explore distilling the encoding process for the teacher model and perform experiments on style-transfer in a setting similar to [41]. We focus on pixel-space diffusion models pre-trained on ImageNet \\(64 \\times 64\\). Specifically, to perform style-transfer between two domains \\(A\\) and \\(B\\), we encode the image from domain-\\(A\\) using a diffusion model trained on domain-\\(A\\), and then decode with a diffusion model trained on domain-\\(B\\). As the encoding process can be understood as reversing the DDIM sampling process, we perform distillation for both the encoder and decoder with classifier-free guidance, and compare with a DDIM encoder and decoder in Fig. 12. We also explore how modifying the guidance strength \\(w\\) can impact the performance and provide more details in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.679, + 0.219, + 0.695 + ], + "angle": 0, + "content": "5. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.705, + 0.471, + 0.869 + ], + "angle": 0, + "content": "Our approach is related to existing works on improving the sampling speed of diffusion models [4, 37, 40]. For instance, denoising diffusion implicit models (DDIM [38]), probability flow sampler [40], fast SDE integrators [8] have been proposed to improve the sampling speed of diffusion models. Other works develop higher-order solvers [17], exponential integrators [15], and dynamic programming based approach [43] to accelerating sampling speed. However, none of these approaches have achieved comparable performance as our method on distilling classifier-free guided diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Existing distillation-based methods for diffusion models are mainly designed for non-classifier-free guided diffusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.513 + ], + "angle": 0, + "content": "models. For instance, [19] proposes to predict the data from noise in one single step by inverting a deterministic encoding of DDIM, [2] proposes to achieve faster sampling speed by distilling higher order solvers into an additional prediction head of the neural network backbone [2]. Progressive distillation [33] is perhaps the most relevant work. Specifically, it proposes to progressively distill a pre-trained diffusion model into a fewer-step student model with the same model architecture. However, none of these approaches are directly applicable or have been applied to classifier-free guided diffusion models. They are also unable to capture a range of different guidance strengths using one single distilled model. On the contrary, by incorporating the guidance strength into the model architecture and training the model using a two-stage procedure, our approach is able to match the performance of the teacher model on a wide range of guidance strength using one single model. Using our method, one single model can capture the trade-off between sample quality and diversity, enabling the real-world application of classifier-free guided diffusion models, where the guidance strength is often specified by users. Moreover, none of the above distillation approaches have been applied to or shown effectiveness for latent-space text-to-image models. Finally, most fast sampling approaches [33, 38, 40] only consider using deterministic sampling schemes to improve the sampling speed. In this work, we further develop an effective stochastic sampling approach to sample from the distilled models." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.531, + 0.619, + 0.546 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.557, + 0.895, + 0.815 + ], + "angle": 0, + "content": "In this paper, we propose a distillation approach for guided diffusion models [6]. Our two-stage approach allows us to significantly speed up popular but relatively inefficient guided diffusion models. We show that our approach can reduce the inference cost of classifier-free guided pixel-space and latent-space diffusion models by at least an order of magnitude. Empirically, we show that our approach is able to produce visually appealing results with only 2 steps, achieving a comparable FID score to the teacher with as few as 4 to 8 steps. We further demonstrate practical applications of our distillation approach to text-guided image-to-image translation and inpainting tasks. We hope that by significantly reducing the inference cost of classifier-free guided diffusion models, our method will promote creative applications as well as the wider adoption of image generation systems. In the future work, we aim to further improve the performance in the two and one sampling step regimes." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.83, + 0.667, + 0.846 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.894, + 0.901 + ], + "angle": 0, + "content": "We thank the anonymous reviewers for their insightful discussions and feedback. All experiments on Stable Diffusion are supported by Stability AI." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14304" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.175, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.116, + 0.47, + 0.156 + ], + "angle": 0, + "content": "[1] Prafulla Dhariwal and Alex Nichol. Diffusion models beat GANs on image synthesis. arXiv preprint arXiv:2105.05233, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.159, + 0.472, + 0.2 + ], + "angle": 0, + "content": "[2] Tim Dockhorn, Arash Vahdat, and Karsten Kreis. Genie: Higher-order denoising diffusion solvers. In Advances in Neural Information Processing Systems, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.203, + 0.472, + 0.27 + ], + "angle": 0, + "content": "[3] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs trained by a two time-scale update rule converge to a local Nash equilibrium. In Advances in Neural Information Processing Systems, pages 6626-6637, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.273, + 0.472, + 0.315 + ], + "angle": 0, + "content": "[4] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems, pages 6840-6851, 2020. 2, 3, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.317, + 0.472, + 0.371 + ], + "angle": 0, + "content": "[5] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. arXiv preprint arXiv:2106.15282, 2021. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.374, + 0.472, + 0.413 + ], + "angle": 0, + "content": "[6] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 2, 3, 5, 8, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.416, + 0.472, + 0.486 + ], + "angle": 0, + "content": "[7] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, July 2021. If you use this software, please cite it as below. 6, 24" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.488, + 0.472, + 0.543 + ], + "angle": 0, + "content": "[8] Alexia Jolicoeur-Martineau, Ke Li, Rémi Piché-Taillefer, Tal Kachman, and Ioannis Mitliagkas. Gotta go fast when generating data with score-based models. arXiv preprint arXiv:2105.14080, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.545, + 0.472, + 0.586 + ], + "angle": 0, + "content": "[9] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. arXiv preprint arXiv:2206.00364, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.588, + 0.472, + 0.629 + ], + "angle": 0, + "content": "[10] Diederik P Kingma, Tim Salimans, Ben Poole, and Jonathan Ho. Variational diffusion models. arXiv preprint arXiv:2107.00630, 2021. 2, 3, 11, 22, 23, 25" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.631, + 0.472, + 0.685 + ], + "angle": 0, + "content": "[11] Zhifeng Kong, Wei Ping, Jiaji Huang, Kexin Zhao, and Bryan Catanzaro. DiffWave: A Versatile Diffusion Model for Audio Synthesis. International Conference on Learning Representations, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.688, + 0.472, + 0.716 + ], + "angle": 0, + "content": "[12] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.718, + 0.472, + 0.773 + ], + "angle": 0, + "content": "[13] Tuomas Kynkänniemi, Tero Karras, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Improved precision and recall metric for assessing generative models. Advances in Neural Information Processing Systems, 32, 2019. 5, 24" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.775, + 0.472, + 0.843 + ], + "angle": 0, + "content": "[14] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 6, 7, 24" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.472, + 0.899 + ], + "angle": 0, + "content": "[15] Luping Liu, Yi Ren, Zhijie Lin, and Zhou Zhao. Pseudo numerical methods for diffusion models on manifolds. In International Conference on Learning Representations, 2022. 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.116, + 0.472, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.146 + ], + "angle": 0, + "content": "[16] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps, 2022. 6, 23, 24" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.894, + 0.204 + ], + "angle": 0, + "content": "[17] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. arXiv:2206.00927, 2022. 8, 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.894, + 0.261 + ], + "angle": 0, + "content": "[18] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models, 2022. 6, 11, 13, 14, 23, 24, 28" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.264, + 0.894, + 0.305 + ], + "angle": 0, + "content": "[19] Eric Luhman and Troy Luhman. Knowledge distillation in iterative generative models for improved sampling speed. arXiv preprint arXiv:2101.02388, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.307, + 0.894, + 0.36 + ], + "angle": 0, + "content": "[20] Chenlin Meng, Yang Song, Jiaming Song, Jiajun Wu, JunYan Zhu, and Stefano Ermon. Sdedit: Image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 2, 5, 6, 7, 25, 26, 27" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.363, + 0.894, + 0.404 + ], + "angle": 0, + "content": "[21] Gautam Mittal, Jesse Engel, Curtis Hawthorne, and Ian Simon. Symbolic music generation with diffusion models. arXiv preprint arXiv:2103.16091, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.407, + 0.894, + 0.448 + ], + "angle": 0, + "content": "[22] Alex Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. International Conference on Machine Learning, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.449, + 0.894, + 0.517 + ], + "angle": 0, + "content": "[23] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.52, + 0.894, + 0.589 + ], + "angle": 0, + "content": "[24] Konpat Preechakul, Nattanat Chathee, Suttisak Wizadwongsa, and Supasorn Suwajanakorn. Diffusion autoencoders: Toward a meaningful and decodable representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10619-10629, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.591, + 0.894, + 0.673 + ], + "angle": 0, + "content": "[25] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 6, 24" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.676, + 0.894, + 0.729 + ], + "angle": 0, + "content": "[26] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.733, + 0.894, + 0.799 + ], + "angle": 0, + "content": "[27] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.803, + 0.894, + 0.871 + ], + "angle": 0, + "content": "[28] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.874, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[29] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14305" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.472, + 0.133 + ], + "angle": 0, + "content": "In International Conference on Medical image computing and computer-assisted intervention, pages 234-241. Springer, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.136, + 0.472, + 0.205 + ], + "angle": 0, + "content": "[30] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. ImageNet large scale visual recognition challenge. International Journal of Computer Vision, 115(3):211-252, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.207, + 0.472, + 0.289 + ], + "angle": 0, + "content": "[31] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.291, + 0.472, + 0.347 + ], + "angle": 0, + "content": "[32] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training GANs. In Advances in Neural Information Processing Systems, pages 2234-2242, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.349, + 0.472, + 0.389 + ], + "angle": 0, + "content": "[33] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. arXiv preprint arXiv:2202.00512, 2022. 2, 3, 4, 5, 8, 11, 19, 22, 23, 25" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.391, + 0.472, + 0.473 + ], + "angle": 0, + "content": "[34] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. arXiv preprint arXiv:2210.08402, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.475, + 0.472, + 0.531 + ], + "angle": 0, + "content": "[35] Abhishek Sinha, Jiaming Song, Chenlin Meng, and Stefano Ermon. D2c: Diffusion-decoding models for few-shot conditional generation. Advances in Neural Information Processing Systems, 34:12533-12548, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.533, + 0.472, + 0.588 + ], + "angle": 0, + "content": "[36] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256-2265, 2015. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.59, + 0.472, + 0.644 + ], + "angle": 0, + "content": "[37] Jascha Sohl-Dickstein, Eric A Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. arXiv preprint arXiv:1503.03585, March 2015. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.646, + 0.472, + 0.701 + ], + "angle": 0, + "content": "[38] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. International Conference on Learning Representations, 2021. 2, 3, 4, 5, 6, 8, 11, 13, 14, 16, 22, 24, 28" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.703, + 0.472, + 0.758 + ], + "angle": 0, + "content": "[39] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. In Advances in Neural Information Processing Systems, pages 11895-11907, 2019. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.76, + 0.472, + 0.828 + ], + "angle": 0, + "content": "[40] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. International Conference on Learning Representations, 2021. 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.472, + 0.872 + ], + "angle": 0, + "content": "[41] Xuan Su, Jiaming Song, Chenlin Meng, and Stefano Ermon. Dual diffusion implicit bridges for image-to-image translation. arXiv preprint arXiv:2203.08382, 2022. 2, 8, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.874, + 0.472, + 0.901 + ], + "angle": 0, + "content": "[42] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Arsenii Ashukha, Aleksei Silvestrov," + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.092, + 0.895, + 0.161 + ], + "angle": 0, + "content": "Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with fourier convolutions. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2149–2159, 2022. 25" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.895, + 0.219 + ], + "angle": 0, + "content": "[43] Daniel Watson, William Chan, Jonathan Ho, and Mohammad Norouzi. Learning fast samplers for diffusion models by differentiating through sample quality. In International Conference on Learning Representations, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.895, + 0.275 + ], + "angle": 0, + "content": "[44] Minkai Xu, Lantao Yu, Yang Song, Chence Shi, Stefano Ermon, and Jian Tang. Geodiff: A geometric diffusion model for molecular conformation generation. arXiv preprint arXiv:2203.02923, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.277, + 0.895, + 0.345 + ], + "angle": 0, + "content": "[45] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2022. 24" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.895, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14306" + } + ] +] \ No newline at end of file diff --git a/2023/On Distillation of Guided Diffusion Models/2a7f0724-3ab3-4aec-b0e2-13e799a34131_origin.pdf b/2023/On Distillation of Guided Diffusion Models/2a7f0724-3ab3-4aec-b0e2-13e799a34131_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8eceadce0143b868dc5435e5ed0b56544a991b85 --- /dev/null +++ b/2023/On Distillation of Guided Diffusion Models/2a7f0724-3ab3-4aec-b0e2-13e799a34131_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3242ec5dd54198f4e0980bf9dee78cd2c1ad2ee759e72178e1a6c17f9659eec4 +size 9180238 diff --git a/2023/On Distillation of Guided Diffusion Models/full.md b/2023/On Distillation of Guided Diffusion Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..de95571ce175703d9ee8cd8668152b2e5171a393 --- /dev/null +++ b/2023/On Distillation of Guided Diffusion Models/full.md @@ -0,0 +1,407 @@ +# On Distillation of Guided Diffusion Models + +Chenlin Meng1 + +chenlin@cs.stanford.edu + +Robin Rombach + +robin@stability.ai + +Ruiqi Gao3 + +ruiqig@google.com + +Diederik Kingma3 + +durk@google.com + +Stefano Ermon1 + +ermon@cs.stanford.edu + +Jonathan Ho3 + +jonathanho@google.com + +Tim Salimans3 + +salimans@google.com + +$^{1}$ Stanford University + +$^{2}$ Stability AI & LMU Munich + +3 Google Research, Brain Team + +![](images/c2eeb0d1c8ee3d67de5f0bd8ebdb1d511976461768a02745b3ce32697ffeeb41.jpg) + +![](images/6bc0a00ef4c5fb732f8a9a85c45ba6d9f335f764a52d73c374939fa732b3b43c.jpg) +Text-guided generation (2 steps) + +![](images/fb7c068a1a3cd4193afa89c0fa80650a46e01b75ed9c93a1e7a5c754d1f53f51.jpg) + +![](images/05f585747c24d0b4496c7704edd85d5e06f031b142d4bf654d7e4d4f74b9d32a.jpg) + +![](images/073499f830b61bcd4676870d20a763147b3120cd4955f99111068ee31a1d942d.jpg) +Text-guided generation (1 step) +Text-guided generation (4 steps) +Figure 1. Distilled Stable Diffusion samples generated by our method. Our two-stage distillation approach is able to generate realistic images using only 1 to 4 denoising steps on various tasks. Compared to standard classifier-free guided diffusion models, we reduce the total number of sampling steps by at least $20\times$ . + +![](images/930d28fbb1d978339636729faa2bba17db923fb6442eff4e688f3e04265de9d3.jpg) +Class-conditional generation (1 step) + +![](images/761b62890495ac24622843713658135518820ba002808e144cf9e706c4bac63b.jpg) + +![](images/9d459bfa505fa2c1497c570b6921b2dcbab1629f9ef72ee07f7eb05569d69711.jpg) + +![](images/bc9ccaa9a22410980952341e3ff2ef4ef800dca6aca46799c97caf8821aa29af.jpg) + +![](images/03db0c5d9eb6441c41986da2d6e5386a4bfcd34e759af4a7098f41f8cbd3aedc.jpg) +Input +Mask + +![](images/3d549196b9cc2a5c0828462b35b7fd045c19ce2d6f5a135009ca75c236e28be3.jpg) +Result 1 + +![](images/830501d9a57ebdc82776b758ec87a7be8ed0b3b746a57941d6c5d4f01a86ce20.jpg) +Result 2 + +![](images/e8067a0c257de4c5d754d4243c0bfd6104830da3c576bfcfb2530b2713bd7800.jpg) +Input + +![](images/8f7157f3ee46790d6efa8644ff1650d70db90df89d8f0c9ec99839a9b8f1f166.jpg) +Image inpainting (2 steps) +Output (different styles) +Image to image translation (3 steps) + +![](images/bce2b8a74953cbb79b3031e8b93e5a714c9b756ef5513ec934346347561f0ae0.jpg) + +![](images/e33e987788fca892e98d8b4cd50ca2126e3bf56eeb23e98d3b3c4e22fb849e55.jpg) + +# Abstract + +Classifier-free guided diffusion models have recently been shown to be highly effective at high-resolution image generation, and they have been widely used in large-scale diffusion + +frameworks including DALL·E 2, Stable Diffusion and Image. However, a downside of classifier-free guided diffusion models is that they are computationally expensive at inference time since they require evaluating two diffusion models, a class-conditional model and an unconditional model, tens to hundreds of times. To deal with this limitation, we pro + +pose an approach to distilling classifier-free guided diffusion models into models that are fast to sample from: Given a pre-trained classifier-free guided model, we first learn a single model to match the output of the combined conditional and unconditional models, and then we progressively distill that model to a diffusion model that requires much fewer sampling steps. For standard diffusion models trained on the pixel-space, our approach is able to generate images visually comparable to that of the original model using as few as 4 sampling steps on ImageNet 64x64 and CIFAR-10, achieving FID/IS scores comparable to that of the original model while being up to 256 times faster to sample from. For diffusion models trained on the latent-space (e.g., Stable Diffusion), our approach is able to generate high-fidelity images using as few as 1 to 4 denoising steps, accelerating inference by at least 10-fold compared to existing methods on ImageNet 256x256 and LAION datasets. We further demonstrate the effectiveness of our approach on text-guided image editing and inpainting, where our distilled model is able to generate high-quality results using as few as 2-4 denoising steps. + +# 1. Introduction + +Denoising diffusion probabilistic models (DDPMs) [4, 37, 39, 40] have achieved state-of-the-art performance on image generation [22, 26–28, 31], audio synthesis [11], molecular generation [44], and likelihood estimation [10]. Classifier-free guidance [6] further improves the sample quality of diffusion models and has been widely used in large-scale diffusion model frameworks including GLIDE [23], Stable Diffusion [28], DALL-E 2 [26], and Imagen [31]. However, one key limitation of classifier-free guidance is its low sampling efficiency—it requires evaluating two diffusion models tens to hundreds of times to generate one sample. This limitation has hindered the application of classifier-free guidance models in real-world settings. Although distillation approaches have been proposed for diffusion models [33, 38], these approaches are not directly applicable to classifier-free guided diffusion models. To deal with this issue, we propose a two-stage distillation approach to improving the sampling efficiency of classifier-free guided models. In the first stage, we introduce a single student model to match the combined output of the two diffusion models of the teacher. In the second stage, we progressively distill the model learned from the first stage to a fewer-step model using the approach introduced in [33]. Using our approach, a single distilled model is able to handle a wide range of different guidance strengths, allowing for the trade-off between sample quality and diversity efficiently. To sample from our model, we consider existing deterministic samplers in the literature [33, 38] and further propose a stochastic sampling process. + +Our distillation framework can not only be applied to standard diffusion models trained on the pixel-space [4, 36, 39], + +but also diffusion models trained on the latent-space of an autoencoder [28, 35] (e.g., Stable Diffusion [28]). For diffusion models directly trained on the pixel-space, our experiments on ImageNet 64x64 and CIFAR-10 show that the proposed distilled model can generate samples visually comparable to that of the teacher using only 4 steps and is able to achieve comparable FID/IS scores as the teacher model using as few as 4 to 16 steps on a wide range of guidance strengths (see Fig. 2). For diffusion model trained on the latent-space of an encoder [28, 35], our approach is able to achieve comparable visual quality to the base model using as few as 1 to 4 sampling steps (at least $10 \times$ fewer steps than the base model) on ImageNet $256 \times 256$ and LAION $512 \times 512$ , matching the performance of the teacher (as evaluated by FID) with only 2-4 sampling steps. To the best of our knowledge, our work is the first to demonstrate the effectiveness of distillation for both pixel-space and latent-space classifier-free diffusion models. Finally, we apply our method to text-guided image inpainting and text-guided image editing tasks [20], where we reduce the total number of sampling steps to as few as 2-4 steps, demonstrating the potential of the proposed framework in style-transfer and image-editing applications [20, 41]. + +![](images/e73a9203a22b09ac814fb8e05d548d163277f88bbb1752ecc5334f7bf4424773.jpg) +Figure 2. Class-conditional samples from our two-stage (deterministic) approach on ImageNet 64x64 for diffusion models trained on the pixel-space. By varying the guidance weight $w$ , our distilled model is able to trade-off between sample diversity and quality, while achieving good results using as few as one sampling step. + +# 2. Background on diffusion models + +Given samples $\mathbf{x}$ from a data distribution $p_{\mathrm{data}}(\mathbf{x})$ , noise scheduling functions $\alpha_{t}$ and $\sigma_{t}$ , we train a diffusion model $\hat{\mathbf{x}}_{\theta}$ , with parameter $\theta$ , via minimizing the weighted mean squared error [4, 36, 39, 40] + +$$ +\mathbb {E} _ {t \sim U [ 0, 1 ], \mathbf {x} \sim p _ {\text {d a t a}} (\mathbf {x}), \mathbf {z} _ {t} \sim q \left(\mathbf {z} _ {t} \mid \mathbf {x}\right)} \left[ \omega \left(\lambda_ {t}\right) \right| | \hat {\mathbf {x}} _ {\boldsymbol {\theta}} \left(\mathbf {z} _ {t}\right) - \mathbf {x} | | _ {2} ^ {2} ], \tag {1} +$$ + +where $\lambda_{t} = \log [\alpha_{t}^{2} / \sigma_{t}^{2}]$ is a signal-to-noise ratio [10], $q(\mathbf{z}_t|\mathbf{x}) = \mathcal{N}(\mathbf{z}_t;\alpha_t\mathbf{x},\sigma_t^2\mathbf{I})$ and $\omega (\lambda_t)$ is a pre-specified weighting function [10]. + +Once the diffusion model $\hat{\mathbf{x}}_{\theta}$ is trained, one can use discrete-time DDIM sampler [38] to sample from the model. Specifically, the DDIM sampler starts with $\mathbf{z}_1 \sim \mathcal{N}(\mathbf{0}, \mathbf{I})$ + +and updates as follows + +$$ +\mathbf {z} _ {s} = \alpha_ {s} \hat {\mathbf {x}} _ {\boldsymbol {\theta}} (\mathbf {z} _ {t}) + \sigma_ {s} \frac {\mathbf {z} _ {t} - \alpha_ {t} \hat {\mathbf {x}} _ {\boldsymbol {\theta}} (\mathbf {z} _ {t})}{\sigma_ {t}}, s = t - 1 / N \tag {2} +$$ + +with $N$ the total number of sampling steps. The final sample will then be generated using $\hat{\mathbf{x}}_{\theta}(\mathbf{z}_0)$ . + +Classifier-free guidance Classifier-free guidance [6] is an effective approach shown to significantly improve the sample quality of class-conditioned diffusion models, and has been widely used in large-scale diffusion models including GLIDE [23], Stable Diffusion [28], DALL-E 2 [26] and Imagen [31]. Specifically, it introduces a guidance weight parameter $w \in \mathbb{R}^{\geq 0}$ to trade-off between sample quality and diversity. To generate a sample, classifier-free guidance evaluates both a conditional diffusion model $\hat{\mathbf{x}}_{c,\theta}$ , where $c$ is the context (e.g., class label, text prompt) to be conditioned on, and a jointly trained unconditional diffusion model $\hat{\mathbf{x}}_{\theta}$ at each update step, using $\hat{\mathbf{x}}_{\theta}^{w} = (1 + w)\hat{\mathbf{x}}_{c,\theta} - w\hat{\mathbf{x}}_{\theta}$ as the model prediction in Eq. (2). As each sampling update requires evaluating two diffusion models, sampling with classifier-free guidance is often expensive [6]. + +Progressive distillation Our approach is inspired by progressive distillation [33], an effective method for improving the sampling speed of (unguided) diffusion models by repeated distillation. Until now, this method could not be directly applied to distilling classifier-free guided models or studied for samplers other than the deterministic DDIM sampler [33, 38]. In this paper we resolve these shortcomings. + +Latent diffusion models (LDMs) [21,24,28,35] increase the training and inference efficiency of diffusion models (directly learned on the pixel-space) by modeling images in the latent space of a pre-trained regularized autoencoder, where the latent representations are usually of lower dimensionality than the pixel-space. Latent diffusion models can be considered as an alternative to cascaded diffusion approaches [5], which rely on one or more super-resolution diffusion models to scale up a low-dimensional image to the desired target resolution. + +In this work, we will apply our distillation framework to classifier-free guided diffusion models learned on both pixel-space [4, 36, 39] and latent-space [21, 24, 28, 35]. + +# 3. Distilling a guided diffusion model + +In the following, we discuss our approach for distilling a classifier-free guided diffusion model [6] into a student model that requires fewer steps to sample from. Using a single distilled model conditioned on the guidance strength, our model can capture a wide range of classifier-free guidance levels, allowing for the trade-off between sample quality and diversity efficiently. + +Given a trained guided model $[\hat{\mathbf{x}}_{c,\theta},\hat{\mathbf{x}}_{\theta}]$ (teacher) either on the pixel-space or latent-space, our approach can be decomposed into two stages. + +# 3.1. Stage-one distillation + +In the first stage, we introduce a student model $\hat{\mathbf{x}}_{\eta_1}(\mathbf{z}_t,w)$ , with learnable parameter $\pmb{\eta}_{1}$ , to match the output of the teacher at any time-step $t\in [0,1]$ . The student model can either be a continuous-time model [40] or a discrete-time model [4, 38] depending on whether the teacher model is discrete or continuous. For simplicity, in the following discussion, we assume both the student and teacher models are continuous as the algorithm for discrete models is almost identical. + +A key functionality of classifier-free guidance [6] is its ability to easily trade-off between sample quality and diversity, which is controlled by a "guidance strength" parameter. This property has demonstrated utility in real-world applications [6,23,26,28,31], where the optimal "guidance strength" is often a user preference. Thus, we would also want our distilled model to maintain this property. Given a range of guidance strengths $\left[w_{\min},w_{\max}\right]$ we are interested in, we optimize the student model using the following objective + +$$ +\mathbb {E} _ {w \sim p _ {w}, t \sim U [ 0, 1 ], \mathbf {x} \sim p _ {\mathrm {d a t a}} (\mathbf {x})} \left[ \omega \left(\lambda_ {t}\right) \| \hat {\mathbf {x}} _ {\boldsymbol {\eta} _ {1}} \left(\mathbf {z} _ {t}, w\right) - \hat {\mathbf {x}} _ {\boldsymbol {\theta}} ^ {w} \left(\mathbf {z} _ {t}\right) \| _ {2} ^ {2} \right], \tag {3} +$$ + +where $\hat{\mathbf{x}}_{\pmb{\theta}}^{w}(\mathbf{z}_{t}) = (1 + w)\hat{\mathbf{x}}_{c,\pmb{\theta}}(\mathbf{z}_{t}) - w\hat{\mathbf{x}}_{\pmb{\theta}}(\mathbf{z}_{t}), \mathbf{z}_{t} \sim q(\mathbf{z}_{t}|\mathbf{x})$ and $p_w(w) = U[w_{\min}, w_{\max}]$ . Note that here, our distilled model $\hat{\mathbf{x}}_{\pmb{\eta}_1}(\mathbf{z}_t, w)$ is also conditioned on the context $c$ (e.g., text prompt), but we drop the notation $c$ in the paper for simplicity. We provide the detailed training algorithm in Algorithm 1 in the supplement. + +To incorporate the guidance weight $w$ , we introduce a $w$ -conditioned model, where $w$ is fed as an input to the student model. To better capture the feature, we apply Fourier embedding to $w$ , which is then incorporated into the diffusion model backbone in a way similar to how the time-step was incorporated in [10, 33]. As initialization plays a key role in the performance [33], we initialize the student model with the same parameters as the conditional model of the teacher, except for the newly introduced parameters related to $w$ -conditioning. The model architecture we use is a U-Net model similar to the ones used in [6] for pixel-space diffusion models and [1, 28] for latent-space diffusion models. We use the same number of channels and attention as used in [6] and the open-sourced Stable Diffusion repository* for our experiments. We provide more details in the supplement. + +# 3.2. Stage-two distillation + +In the second stage, we consider a discrete time-step scenario and progressively distill the learned model from the first-stage $\hat{\mathbf{x}}_{\eta_1}(\mathbf{z}_t,w)$ into an fewer-step student model $\hat{\mathbf{x}}_{\eta_2}(\mathbf{z}_t,w)$ with learnable parameter $\eta_{2}$ , by halving the number of sampling steps each time. Letting $N$ denote the number of sampling steps, given $w\sim U[w_{\min},w_{\max}]$ and + +![](images/019b251179ebf0264ae30e436a0b7e9f5f3aafa287ebba6c38b489f2427f170f.jpg) +(a) 2 denoising steps + +![](images/98f99a856052c3f482501fe0669cb0c10b53bd98f17a18ce6e5906bc7babb438.jpg) +Figure 4. Text-guided generation on LAION (512x512) using our distilled Stable Diffusion model. Our model is able to generate high-quality image samples using 2, 4 or 8 denoising steps, significantly improving the inference efficiency of Stable Diffusion. + +![](images/436fa4e547ed4588699e76e23a4412f052419a9cf226bfcc0136ebce27dbc277.jpg) +(b) 4 denoising steps + +![](images/0cd6729a2cc48efc54e30037e083d383a4eda77f75bf42cc7ed659ecefeec9e6.jpg) +(c) 8 denoising steps + +
w=0w=0.3w=1w=4
MethodFID (↓)IS (↑)FID (↓)IS (↑)FID (↓)IS (↑)FID (↓)IS (↑)
Ours 1-step (D/S)22.74 / 26.9125.51 / 23.5514.85 / 18.4837.09 / 33.307.54 / 8.9275.19 / 67.8018.72 / 17.85157.46 / 148.97
Ours 4-step (D/S)4.14 / 3.9146.64 / 48.922.17 / 2.2469.64 / 73.737.95 / 8.51128.98 / 135.3626.45 / 27.33207.45 / 216.56
Ours 8-step (D/S)2.79 / 2.4450.72 / 55.032.05 / 2.3176.01 / 83.009.33 / 10.56136.47 / 147.3926.62 / 27.84203.47 / 219.89
Ours 16-step (D/S)2.44 / 2.1052.53 / 57.812.20 / 2.5679.47 / 87.509.99 / 11.63139.11 / 153.1726.53 / 27.69204.13 / 218.70
Single-w 1-step19.6124.0011.7036.956.6474.4119.857170.69
Single-w 4-step4.7938.772.3462.088.23118.5227.75219.64
Single-w 8-step3.3942.132.3268.769.69125.2027.67218.08
Single-w 16-step2.9743.632.5670.9710.34127.7027.40216.52
DDIM 16x2-step [38]7.6837.605.3360.839.53112.7521.56195.17
DDIM 32x2-step [38]5.0340.937.479.339.26126.2223.03213.23
DDIM 64x2-step [38]3.7443.165.529.519.53133.1723.64217.88
Teacher (DDIM 1024x2-step)2.9244.812.3674.839.84139.5023.94224.74
+ +Table 1. ImageNet 64x64 distillation results for pixel-space diffusion models ( $w = 0$ refers to non-guided models). For our method, $D$ and $S$ stand for deterministic and stochastic sampler respectively. We observe that training the model conditioned on a guidance interval $w \in [0,4]$ performs comparably with training a model on a fixed $w$ (see Single- $w$ ). Our approach significantly outperforms DDIM when using fewer steps, and is able to match the teacher performance using as few as 8 to 16 steps. + +$t \in \{1, \dots, N\}$ , we train the student model to match the output of two-step DDIM sampling of the teacher (i.e., from $t/N$ to $t - 0.5/N$ and from $t - 0.5/N$ to $t - 1/N$ ) in one step, following the approach of [33]. After distilling the $2N$ steps in the teacher model to $N$ steps in the student model, we can use the $N$ -step student model as the new teacher model, repeat the same procedure, and distill the teacher model into a $N/2$ -step student model. At each step, we initialize the student model with the parameters of the teacher. We provide the training algorithm and extra details in the supplementary material. + +# 3.3. $N$ -step deterministic and stochastic sampling + +Once the model $\hat{\mathbf{x}}_{\eta_2}$ is trained, given a specified guidance strength $w \in [w_{\min}, w_{\max}]$ , we can perform sampling via the DDIM update rule in Eq. (2). We note that given the distilled model $\hat{\mathbf{x}}_{\eta_2}$ , this sampling procedure is deterministic given the initialization $\mathbf{z}_1^w$ . In fact, we can also perform $N$ -step stochastic sampling: We apply one deterministic sampling step with two-times the original step-length (i.e., + +the same as a $N/2$ -step deterministic sampler) and then perform one stochastic step backward (i.e., perturb with noise) using the original step-length, a process inspired by [9]. With $\mathbf{z}_1^w \sim \mathcal{N}(\mathbf{0}, \mathbf{I})$ , we use the following update rule when $t > 1/N$ + +$$ +\mathbf {z} _ {k} ^ {w} = \alpha_ {k} \hat {\mathbf {x}} _ {\boldsymbol {\eta} _ {2}} \left(\mathbf {z} _ {t} ^ {w}\right) + \sigma_ {k} \frac {\mathbf {z} _ {t} ^ {w} - \alpha_ {t} \hat {\mathbf {x}} _ {\boldsymbol {\eta} _ {2}} ^ {w} \left(\mathbf {z} _ {t}\right)}{\sigma_ {t}}, \tag {4} +$$ + +$$ +\text {w h e r e} \mathbf {z} _ {s} ^ {w} = \left(\alpha_ {s} / \alpha_ {k}\right) \mathbf {z} _ {k} ^ {w} + \sigma_ {s | k} \epsilon , \epsilon \sim \mathcal {N} (\mathbf {0}, \mathbf {I}); \tag {5} +$$ + +$$ +\mathbf {z} _ {h} ^ {w} = \alpha_ {h} \hat {\mathbf {x}} _ {\boldsymbol {\eta} _ {2}} \left(\mathbf {z} _ {s} ^ {w}\right) + \sigma_ {h} \frac {\mathbf {z} _ {s} ^ {w} - \alpha_ {s} \hat {\mathbf {x}} _ {\boldsymbol {\eta} _ {2}} ^ {w} \left(\mathbf {z} _ {s} ^ {w}\right)}{\sigma_ {s}}, \tag {6} +$$ + +$$ +w h e r e \mathbf {z} _ {k} ^ {w} = \left(\alpha_ {k} / \alpha_ {h}\right) \mathbf {z} _ {h} ^ {w} + \sigma_ {k | h} \epsilon , \epsilon \sim \mathcal {N} (\mathbf {0}, \mathbf {I}). \tag {7} +$$ + +In the above equations, $h = t - 3 / N$ , $k = t - 2 / N$ , $s = t - 1 / N$ and $\sigma_{a|b}^{2} = (1 - e^{\lambda_{a} - \lambda_{b}})\sigma_{a}^{2}$ . When $t = 1 / N$ , we use deterministic update Eq. (2) to obtain $\mathbf{z}_0^w$ from $\mathbf{z}_{1 / N}^w$ . We provide an illustration of the process in Fig. 5, where the number of denoising steps is 4. We note that compared to the deterministic sampler, performing stochastic sampling requires evaluating the model at slightly different time-steps, + +![](images/ba61fe9a0723a73bb3500aa6adbf6117174f16451908b263c1665bacb55124e2.jpg) +Figure 5. Sampling procedures of the distilled model where the number of denoising steps is 4. + +and would require small modifications to training algorithm for the edge cases. We provide the algorithm and more details in the supplementary material. + +# 4. Experiments + +In this section, we evaluate the performance of our distillation approach on pixel-space diffusion models (i.e. DDPM [4]) and latent-space diffusion models (i.e. Stable Diffusion [28]). We further apply our approach to text-guided image editing and inpainting tasks. Experiments show that our approach is able to achieve competitive performance while using as few as 2-4 steps on all tasks. + +# 4.1. Distillation for pixel-space guided models + +In this experiment, we consider class-conditional diffusion models directly trained on the pixel-space [4,6,33]. + +Settings We focus on ImageNet 64x64 [30] and CIFAR-10 [12] as higher-resolution image generation in this scenario often relies on combining with other super-resolution techniques [5, 31]. We explore different ranges for the guidance weight and observe that all ranges work comparably and therefore use $[w_{min}, w_{max}] = [0, 4]$ for the experiments. The baselines we consider include DDPM ancestral sampling [4] and DDIM [38]. The teacher model we use is a $1024 \times 2$ -step DDIM model, where the conditional and unconditional components both use 1024 DDIM denoising steps. To better understand how the guidance weight $w$ should be incorporated, we also include models trained using a single fixed $w$ as a baseline. We use the same pre-trained teacher model for all the methods for fair comparisons. Following [4, 6, 39], we use a U-Net [29, 39] architecture for the baselines, and the same U-Net backbone with the introduced $w$ -embedding for our two-step student models (see Sec. 3). Following [33], we use a v-prediction model for both datasets. + +Results We report the performance as evaluated in FID [3] and Inception scores (IS) [32] for all approaches on ImageNet 64x64 in Fig. 6 and Tab. 1 and provide extended ImageNet 64x64 and CIFAR-10 results in the supplement. We observe that our distilled model is able to match a teacher guided DDIM model with $1024 \times 2$ sampling steps using only 4-16 steps, achieving a speedup for up to $256 \times$ . We emphasize that, using our approach, a single distilled model is able to match the teacher performance on a wide range + +of guidance strengths. This has not been achieved by any previous methods. + +# 4.2. Distillation for latent-space guided models + +After demonstrating the effectiveness of our method on pixel-space class-guided diffusion models in Sec. 4.1, we now expand its scope to latent-space diffusion models. In the following sections, we show the effectiveness of our approach on Latent Diffusion [28] on a variety of tasks, including class-conditional generation, text-to-image generation, image inpainting and text-guided style-transfer [20]. + +In the following experiments, we use the open-sourced latent-space diffusion models [28] as the teacher models. As v-prediction teacher model tends to perform better than $\epsilon$ -prediction model, we fine-tune the open-sourced $\epsilon$ -prediction models into v-prediction teacher models. We provide more details in the supplementary material. + +# 4.2.1 Class-conditional generation + +In this section, we apply our method to a class-conditional latent diffusion model pre-trained on ImageNet $256 \times 256$ . We start from the DDIM teacher model with 512 sampling steps, and use the output as the target to train our distilled model. We use a batch size of 512 and uniformly sample the guidance strength $w \in [w_{min} = 0, w_{max} = 14]$ during training. + +Results Empirically, we find that our distilled model is able to match the performance of the teacher model (originally trained on 1000 steps) in terms of FID scores while using only 2 or 4 sampling steps. We also achieve significantly better performance than DDIM when using 1-4 sampling steps (see Fig. 11). Qualitatively, we find that samples synthesized using a single denoising step still yield satisfying results, while the baseline fails to generate images with meaningful contents. We provide extra samples in the supplementary material. + +Similar to the pixel-based results in Fig. 6, we also observe the trade-off between sampling quality and diversity as measured by FID and Inception Score for our distilled latent diffusion model. Following Kynkänniemi et al [13], we further compute improved precision and recall metrics for this experiment in the appendix. + +# 4.2.2 Text-guided image generation + +In this section, we focus on the text-guided Stable Diffusion model pretrained on subsets $\dagger$ of LAION-5B [34] at a resolution $512 \times 512$ . We then follow our two-stage approach introduced in Sec. 3 and distill the guided model in 3000 gradient updates into a $w$ -conditioned model using + +![](images/cbec320e32ef4c440a3a8913360b49b54d9efe22f3accccb03e69b5ce00c6d22.jpg) + +![](images/041b3f32f7b8297c6f864a31573be28d85a8c79f280ffd7c294753f35bfedc22.jpg) + +![](images/ea5e4d98537a0cffaa6d98c47b80f3469a742385be7a41f9e98d3c400a873292.jpg) + +![](images/fb707b4c85407ed7c68e73fdf99e27c94232faf70df8adc93f32a22f9ddc039b.jpg) + +![](images/97ee8399aae499ac1037d73062cb681662c6d6d0066e44c1b5272ef2e23d7b13.jpg) + +![](images/d0dd693b2a12bb672df55802208ae68a6927d17ee04fb407466baebb21e29e64.jpg) +$w = 0$ + +![](images/02930878f9429bc9d7483765481e491c86abe4acda7a91cd5de85f01a0637430.jpg) +$w = 0.3$ + +![](images/a0f3e40986241e51c66d85b02b3f296aabecbbe7a66bfe69650e04a6d9e9cb90.jpg) +$w = 1$ + +![](images/3e78bed81ea1fb715c91732ccf22c11d97cf03d6ca52c8a36484e710a2610283.jpg) +$w = 2$ + +![](images/ece91057b2d35114a9a67470305a79bd85312c29203b98ca5fdee83b2f0f728a.jpg) +$w = 4$ + +![](images/98f799da38b80d6a8b79674b305ab5c500be4cb173e8fe57d005af43ba25b5d7.jpg) +Figure 6. ImageNet $64 \times 64$ sample quality evaluated by FID and IS scores. Our distilled model significantly outperforms the DDPM and DDIM baselines, and is able to match the performance of the teacher using as few as 8 to 16 steps. By varying $w$ , a single distilled model is able to capture the trade-off between sample diversity and quality. +Distilled Text-to-Image samples (4 steps) +Figure 7. Text-guided Stable Diffusion results. We distill the public Stable Diffusion model using the proposed pipeline, arriving at a model that achieves high sample quality using only four denoising steps (left). When sampling from the original model using four DDIM steps, the generated samples have clear artifacts (middle). When using eight DDIM steps, the results get better (right), but are still blurry and less consistent than the distilled results using fewer steps. More samples are provided in Fig. 4. + +![](images/f613993d643d5fa0a970a4460f3695f53d71bb718f2447bd5cdc014f1821b63d.jpg) +Native Text-to-Image samples (4 steps) + +![](images/ff071da0624ff43e38d0107366b09d0da90636bd4f95afe9eb97d1145ec672fb.jpg) +Native Text-to-Image samples (8 steps) + +$w \in [w_{min} = 2, w_{max} = 14]$ , and a batch size of 512. Although we can condition on a broader range of $w$ for the distilled (student) model, the utility remains unclear as we typically do not exceed the normal guidance range when sampling with the teacher model. The final model is obtained by applying progressive distillation for 2000 training steps per stage, except when for the low-step regime of 1,2, and 4 steps, where we train for 20000 gradient updates. A detailed analysis of the convergence properties of this model in the supplement. + +
Method2-step4-step8-step
DPM [16]98.9/0.2034.3/0.2931.7/0.32
DPM++ [18]98.8/0.2034.1/0.2925.6/0.32
Ours37.3/0.2726.0/0.3026.9/0.30
+ +Table 2. FID/CLIP scores on LAION 512X512 ( $w = 8.0$ ). We point out that DPM and DPM++ use both the conditional and unconditional components for sampling. Depending on the implementation, this either requires higher peak memory or two times more sampling steps. + +Results We present samples in Fig. 4. We evaluate the resulting model both qualitatively and quantitatively. For the latter analysis, we follow [31] and evaluate CLIP [25] and FID scores to assess text-image alignment and quality, respectively. We use the open-sourced ViT-g/14 [7] CLIP model for evaluation. The quantitative results in Fig. 10 show that our method can significantly increase the performance in both metrics over DDIM sampling on the base model for + +2 and 4 sampling steps. For 8 steps, these metrics do not show a significant difference. However, when considering the corresponding samples in Fig. 7 we can observe a stark difference in terms of visual image quality. In contrast to the 8-step DDIM samples from the original model, the distilled samples are sharper and more coherent. We hypothesize that FID and CLIP do not fully capture these differences in our evaluation setting on COCO2017 [14], where we used 5000 random captions from the validation set. We further compute the FID and CLIP scores for our distilled LAION $512 \times 512$ model and compare them with the DPM [16] and $\mathrm{DPM}++$ [18] solver in Tab. 2. We observe that our method is able to achieve significantly better performance when the denoising step is 2 or 4. Furthermore, we stress that stage-one of our method already decreases the number of function evaluations by a factor of 2, as we distill the classifier-free guidance step into a single model. Depending on the exact implementation (batched vs. sequential network evaluation), this either decreases peak memory or sampling time compared to existing solvers [16, 18, 38]. + +# 4.2.3 Text-guided image-to-image translation + +In this section, we perform experiments on text-guided image-to-image translation with SDEdit [20] using our distilled model from Sec. 4.2.2. Following SDEdit [20], we perform stochastic encoding in the latent space, but instead + +![](images/fd4a576fa344ffbe1d42def9d273748724b60d646b3109add5004344f748410a.jpg) +Input +Watercolor +Art station +Photograph +Input +Watercolor +Art station painting + +![](images/5d49590896898c6051dd3950f862aec21dee3a99a02e14257edb9cfc27997da3.jpg) +Input +Girl + +![](images/649e891f3bcdf8b550572949be033f37aa578bc4ad79ac96b35516b60d6fa51a.jpg) +Oil painting +Boy +Anime +3D render +Anime +3D render +Male +Van Gogh +Watercolor +Art station + +![](images/e6a319a1aee460d2faec76fc15904b4a758b5e52c5ae49bc288a0f389dc17f09.jpg) +Figure 8. Text-guided image-to-image translation [20] with the distilled Stable Diffusion model (3 denoising steps). We observe that our model is able to generate high-quality and faithful outputs using only 3 denoising steps. +Input + +![](images/de5653f9a17f5cb852113031736c4f643bd0f5ace4f4dc79472ff83ba3e7eaf2.jpg) + +![](images/b8250917f941b6991c587ce05a0b56d3e8fcbfb4bd96b3861a6272d3865df53c.jpg) +Figure 10. FID and CLIP ViT-g/14 score for text-to-image generation at $512 \times 512$ px using the distilled Stable Diffusion model. The results are evaluated on 5000 captions from the COCO2017 [14] validation set. Our distilled latent diffusion model is able to generate high-quality image samples using significantly less sampling steps than the original model while achieving similar or better FID and CLIP scores, especially in the low-step regime. + +![](images/f5ef46df91d8617dd752c0e3daeb21932ce5f8e5208038bcd5f76011355e256d.jpg) + +use the deterministic sampler of the distilled model to perform deterministic decoding. We consider input image and text of various kinds and provide qualitative results in Fig. 8. We observe that our distilled model generates high-quality style-transfer results using as few as 3 denoising steps. We provide more analysis on the trade-off between sample quality, controllability and efficiency in the supplement. + +![](images/280b6991cfac975dd2c8905ef3a699ba24d9805d3babb9bdcb5e47cffbb77c31.jpg) +Figure 11. FID and Inception Score for class-conditional image generation on ImageNet $(256 \times 256)$ with distilled latent diffusion. The results are evaluated on 5000 samples. Our distilled latent diffusion model is able to generate high-quality image samples using significantly less sampling steps (up to a factor of 16) than the original model while achieving similar or better FID scores. + +![](images/e53a1bc88b4ad86e626365445a352c30deccb744a0b679bd88603736daed3adb.jpg) + +# 4.2.4 Image inpainting + +In this section, we apply our approach to a pre-trained image inpainting latent diffusion model. We use the open-source Stable Diffusion Inpainting\* image-inpainting model. This model is a fine-tuned version of the pure text-to-image Stable Diffusion model from above, where additional input channels + +![](images/2da267968cdc011b6132827c27e3bf9ff0ba5b7f4a3ca5f167fb7c41033eeeca.jpg) +Figure 12. Style transfer comparison on ImageNet 64x64 for pixel-space models. For our approach, we use a distilled encoder and decoder. For the baseline, we encode and decode using DDIM. We use $w = 0$ and 16 sampling steps for both the encoder and decoder. We observe that our method achieves more realistic outputs. + +were added to process masks and masked images. + +We use the same distillation algorithm as used in the previous section. For training, we start from the v-prediction teacher model sampled with 512 DDIM steps, and use the output as the target to optimize our student model. We present qualitative results in Fig. 9, demonstrating the potential of our method for fast, real-world image editing applications. For additional training details and a quantitative evaluation, see the supplementary. + +# 4.3. Progressive distillation for encoding + +In this experiment, we explore distilling the encoding process for the teacher model and perform experiments on style-transfer in a setting similar to [41]. We focus on pixel-space diffusion models pre-trained on ImageNet $64 \times 64$ . Specifically, to perform style-transfer between two domains $A$ and $B$ , we encode the image from domain- $A$ using a diffusion model trained on domain- $A$ , and then decode with a diffusion model trained on domain- $B$ . As the encoding process can be understood as reversing the DDIM sampling process, we perform distillation for both the encoder and decoder with classifier-free guidance, and compare with a DDIM encoder and decoder in Fig. 12. We also explore how modifying the guidance strength $w$ can impact the performance and provide more details in the supplementary material. + +# 5. Related Work + +Our approach is related to existing works on improving the sampling speed of diffusion models [4, 37, 40]. For instance, denoising diffusion implicit models (DDIM [38]), probability flow sampler [40], fast SDE integrators [8] have been proposed to improve the sampling speed of diffusion models. Other works develop higher-order solvers [17], exponential integrators [15], and dynamic programming based approach [43] to accelerating sampling speed. However, none of these approaches have achieved comparable performance as our method on distilling classifier-free guided diffusion models. + +Existing distillation-based methods for diffusion models are mainly designed for non-classifier-free guided diffusion + +models. For instance, [19] proposes to predict the data from noise in one single step by inverting a deterministic encoding of DDIM, [2] proposes to achieve faster sampling speed by distilling higher order solvers into an additional prediction head of the neural network backbone [2]. Progressive distillation [33] is perhaps the most relevant work. Specifically, it proposes to progressively distill a pre-trained diffusion model into a fewer-step student model with the same model architecture. However, none of these approaches are directly applicable or have been applied to classifier-free guided diffusion models. They are also unable to capture a range of different guidance strengths using one single distilled model. On the contrary, by incorporating the guidance strength into the model architecture and training the model using a two-stage procedure, our approach is able to match the performance of the teacher model on a wide range of guidance strength using one single model. Using our method, one single model can capture the trade-off between sample quality and diversity, enabling the real-world application of classifier-free guided diffusion models, where the guidance strength is often specified by users. Moreover, none of the above distillation approaches have been applied to or shown effectiveness for latent-space text-to-image models. Finally, most fast sampling approaches [33, 38, 40] only consider using deterministic sampling schemes to improve the sampling speed. In this work, we further develop an effective stochastic sampling approach to sample from the distilled models. + +# 6. Conclusion + +In this paper, we propose a distillation approach for guided diffusion models [6]. Our two-stage approach allows us to significantly speed up popular but relatively inefficient guided diffusion models. We show that our approach can reduce the inference cost of classifier-free guided pixel-space and latent-space diffusion models by at least an order of magnitude. Empirically, we show that our approach is able to produce visually appealing results with only 2 steps, achieving a comparable FID score to the teacher with as few as 4 to 8 steps. We further demonstrate practical applications of our distillation approach to text-guided image-to-image translation and inpainting tasks. We hope that by significantly reducing the inference cost of classifier-free guided diffusion models, our method will promote creative applications as well as the wider adoption of image generation systems. In the future work, we aim to further improve the performance in the two and one sampling step regimes. + +# Acknowledgements + +We thank the anonymous reviewers for their insightful discussions and feedback. All experiments on Stable Diffusion are supported by Stability AI. + +# References + +[1] Prafulla Dhariwal and Alex Nichol. Diffusion models beat GANs on image synthesis. arXiv preprint arXiv:2105.05233, 2021. 3 +[2] Tim Dockhorn, Arash Vahdat, and Karsten Kreis. Genie: Higher-order denoising diffusion solvers. In Advances in Neural Information Processing Systems, 2022. 8 +[3] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs trained by a two time-scale update rule converge to a local Nash equilibrium. In Advances in Neural Information Processing Systems, pages 6626-6637, 2017. 5 +[4] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems, pages 6840-6851, 2020. 2, 3, 5, 8 +[5] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. arXiv preprint arXiv:2106.15282, 2021. 3, 5 +[6] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 2, 3, 5, 8, 11 +[7] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, July 2021. If you use this software, please cite it as below. 6, 24 +[8] Alexia Jolicoeur-Martineau, Ke Li, Rémi Piché-Taillefer, Tal Kachman, and Ioannis Mitliagkas. Gotta go fast when generating data with score-based models. arXiv preprint arXiv:2105.14080, 2021. 8 +[9] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. arXiv preprint arXiv:2206.00364, 2022. 4 +[10] Diederik P Kingma, Tim Salimans, Ben Poole, and Jonathan Ho. Variational diffusion models. arXiv preprint arXiv:2107.00630, 2021. 2, 3, 11, 22, 23, 25 +[11] Zhifeng Kong, Wei Ping, Jiaji Huang, Kexin Zhao, and Bryan Catanzaro. DiffWave: A Versatile Diffusion Model for Audio Synthesis. International Conference on Learning Representations, 2021. 2 +[12] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 5 +[13] Tuomas Kynkänniemi, Tero Karras, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Improved precision and recall metric for assessing generative models. Advances in Neural Information Processing Systems, 32, 2019. 5, 24 +[14] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 6, 7, 24 +[15] Luping Liu, Yi Ren, Zhijie Lin, and Zhou Zhao. Pseudo numerical methods for diffusion models on manifolds. In International Conference on Learning Representations, 2022. 8 + +[16] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps, 2022. 6, 23, 24 +[17] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. arXiv:2206.00927, 2022. 8, 11 +[18] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models, 2022. 6, 11, 13, 14, 23, 24, 28 +[19] Eric Luhman and Troy Luhman. Knowledge distillation in iterative generative models for improved sampling speed. arXiv preprint arXiv:2101.02388, 2021. 8 +[20] Chenlin Meng, Yang Song, Jiaming Song, Jiajun Wu, JunYan Zhu, and Stefano Ermon. Sdedit: Image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 2, 5, 6, 7, 25, 26, 27 +[21] Gautam Mittal, Jesse Engel, Curtis Hawthorne, and Ian Simon. Symbolic music generation with diffusion models. arXiv preprint arXiv:2103.16091, 2021. 3 +[22] Alex Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. International Conference on Machine Learning, 2021. 2 +[23] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 2, 3 +[24] Konpat Preechakul, Nattanat Chathee, Suttisak Wizadwongsa, and Supasorn Suwajanakorn. Diffusion autoencoders: Toward a meaningful and decodable representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10619-10629, 2022. 3 +[25] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 6, 24 +[26] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2, 3 +[27] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 2 +[28] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2, 3, 5 +[29] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. + +In International Conference on Medical image computing and computer-assisted intervention, pages 234-241. Springer, 2015. 5 +[30] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. ImageNet large scale visual recognition challenge. International Journal of Computer Vision, 115(3):211-252, 2015. 5 +[31] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 2, 3, 5, 6 +[32] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training GANs. In Advances in Neural Information Processing Systems, pages 2234-2242, 2016. 5 +[33] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. arXiv preprint arXiv:2202.00512, 2022. 2, 3, 4, 5, 8, 11, 19, 22, 23, 25 +[34] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. arXiv preprint arXiv:2210.08402, 2022. 5 +[35] Abhishek Sinha, Jiaming Song, Chenlin Meng, and Stefano Ermon. D2c: Diffusion-decoding models for few-shot conditional generation. Advances in Neural Information Processing Systems, 34:12533-12548, 2021. 2, 3 +[36] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256-2265, 2015. 2, 3 +[37] Jascha Sohl-Dickstein, Eric A Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. arXiv preprint arXiv:1503.03585, March 2015. 2, 8 +[38] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. International Conference on Learning Representations, 2021. 2, 3, 4, 5, 6, 8, 11, 13, 14, 16, 22, 24, 28 +[39] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. In Advances in Neural Information Processing Systems, pages 11895-11907, 2019. 2, 3, 5 +[40] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. International Conference on Learning Representations, 2021. 2, 3, 8 +[41] Xuan Su, Jiaming Song, Chenlin Meng, and Stefano Ermon. Dual diffusion implicit bridges for image-to-image translation. arXiv preprint arXiv:2203.08382, 2022. 2, 8, 12 +[42] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Arsenii Ashukha, Aleksei Silvestrov, + +Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with fourier convolutions. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2149–2159, 2022. 25 +[43] Daniel Watson, William Chan, Jonathan Ho, and Mohammad Norouzi. Learning fast samplers for diffusion models by differentiating through sample quality. In International Conference on Learning Representations, 2022. 8 +[44] Minkai Xu, Lantao Yu, Yang Song, Chence Shi, Stefano Ermon, and Jian Tang. Geodiff: A geometric diffusion model for molecular conformation generation. arXiv preprint arXiv:2203.02923, 2022. 2 +[45] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2022. 24 \ No newline at end of file diff --git a/2023/On Distillation of Guided Diffusion Models/images.zip b/2023/On Distillation of Guided Diffusion Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..24abf96cb108e04866b7bbcbcf31624083687332 --- /dev/null +++ b/2023/On Distillation of Guided Diffusion Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d7da07bc83b6f9b599a9d4f31de2e1a0571a1f1836c35fc515b2ba9241596cf8 +size 1058792 diff --git a/2023/On Distillation of Guided Diffusion Models/layout.json b/2023/On Distillation of Guided Diffusion Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..1df672100efc02a646ac974a638c6225159f9ac1 --- /dev/null +++ b/2023/On Distillation of Guided Diffusion Models/layout.json @@ -0,0 +1,11514 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 163, + 53, + 430, + 69 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 163, + 53, + 430, + 69 + ], + "spans": [ + { + "bbox": [ + 163, + 53, + 430, + 69 + ], + "type": "text", + "content": "On Distillation of Guided Diffusion Models" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 83, + 94, + 159, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 94, + 159, + 108 + ], + "spans": [ + { + "bbox": [ + 83, + 94, + 159, + 108 + ], + "type": "text", + "content": "Chenlin Meng1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 59, + 110, + 185, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 110, + 185, + 121 + ], + "spans": [ + { + "bbox": [ + 59, + 110, + 185, + 121 + ], + "type": "text", + "content": "chenlin@cs.stanford.edu" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 214, + 94, + 299, + 106 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 94, + 299, + 106 + ], + "spans": [ + { + "bbox": [ + 214, + 94, + 299, + 106 + ], + "type": "text", + "content": "Robin Rombach" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 207, + 110, + 306, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 207, + 110, + 306, + 121 + ], + "spans": [ + { + "bbox": [ + 207, + 110, + 306, + 121 + ], + "type": "text", + "content": "robin@stability.ai" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 347, + 94, + 403, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 94, + 403, + 108 + ], + "spans": [ + { + "bbox": [ + 347, + 94, + 403, + 108 + ], + "type": "text", + "content": "Ruiqi Gao3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 329, + 110, + 422, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 110, + 422, + 121 + ], + "spans": [ + { + "bbox": [ + 329, + 110, + 422, + 121 + ], + "type": "text", + "content": "ruiqig@google.com" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 444, + 94, + 534, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 444, + 94, + 534, + 108 + ], + "spans": [ + { + "bbox": [ + 444, + 94, + 534, + 108 + ], + "type": "text", + "content": "Diederik Kingma3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 449, + 110, + 531, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 449, + 110, + 531, + 121 + ], + "spans": [ + { + "bbox": [ + 449, + 110, + 531, + 121 + ], + "type": "text", + "content": "durk@google.com" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 117, + 127, + 196, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 127, + 196, + 140 + ], + "spans": [ + { + "bbox": [ + 117, + 127, + 196, + 140 + ], + "type": "text", + "content": "Stefano Ermon1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 100, + 144, + 215, + 154 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 144, + 215, + 154 + ], + "spans": [ + { + "bbox": [ + 100, + 144, + 215, + 154 + ], + "type": "text", + "content": "ermon@cs.stanford.edu" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 268, + 128, + 335, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 268, + 128, + 335, + 140 + ], + "spans": [ + { + "bbox": [ + 268, + 128, + 335, + 140 + ], + "type": "text", + "content": "Jonathan Ho3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 246, + 144, + 360, + 155 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 144, + 360, + 155 + ], + "spans": [ + { + "bbox": [ + 246, + 144, + 360, + 155 + ], + "type": "text", + "content": "jonathanho@google.com" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 405, + 128, + 477, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 405, + 128, + 477, + 140 + ], + "spans": [ + { + "bbox": [ + 405, + 128, + 477, + 140 + ], + "type": "text", + "content": "Tim Salimans3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 390, + 144, + 494, + 155 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 390, + 144, + 494, + 155 + ], + "spans": [ + { + "bbox": [ + 390, + 144, + 494, + 155 + ], + "type": "text", + "content": "salimans@google.com" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 70, + 160, + 172, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 160, + 172, + 175 + ], + "spans": [ + { + "bbox": [ + 70, + 160, + 172, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 70, + 160, + 172, + 175 + ], + "type": "text", + "content": "Stanford University" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 196, + 161, + 342, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 161, + 342, + 175 + ], + "spans": [ + { + "bbox": [ + 196, + 161, + 342, + 175 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 196, + 161, + 342, + 175 + ], + "type": "text", + "content": "Stability AI & LMU Munich" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 365, + 161, + 520, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 161, + 520, + 175 + ], + "spans": [ + { + "bbox": [ + 365, + 161, + 520, + 175 + ], + "type": "text", + "content": "3 Google Research, Brain Team" + } + ] + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 48, + 193, + 272, + 293 + ], + "blocks": [ + { + "bbox": [ + 48, + 193, + 272, + 293 + ], + "lines": [ + { + "bbox": [ + 48, + 193, + 272, + 293 + ], + "spans": [ + { + "bbox": [ + 48, + 193, + 272, + 293 + ], + "type": "image", + "image_path": "c2eeb0d1c8ee3d67de5f0bd8ebdb1d511976461768a02745b3ce32697ffeeb41.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 272, + 193, + 409, + 263 + ], + "blocks": [ + { + "bbox": [ + 272, + 193, + 409, + 263 + ], + "lines": [ + { + "bbox": [ + 272, + 193, + 409, + 263 + ], + "spans": [ + { + "bbox": [ + 272, + 193, + 409, + 263 + ], + "type": "image", + "image_path": "6bc0a00ef4c5fb732f8a9a85c45ba6d9f335f764a52d73c374939fa732b3b43c.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 349, + 266, + 470, + 277 + ], + "lines": [ + { + "bbox": [ + 349, + 266, + 470, + 277 + ], + "spans": [ + { + "bbox": [ + 349, + 266, + 470, + 277 + ], + "type": "text", + "content": "Text-guided generation (2 steps)" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 409, + 193, + 476, + 263 + ], + "blocks": [ + { + "bbox": [ + 409, + 193, + 476, + 263 + ], + "lines": [ + { + "bbox": [ + 409, + 193, + 476, + 263 + ], + "spans": [ + { + "bbox": [ + 409, + 193, + 476, + 263 + ], + "type": "image", + "image_path": "fb7c068a1a3cd4193afa89c0fa80650a46e01b75ed9c93a1e7a5c754d1f53f51.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 477, + 193, + 545, + 263 + ], + "blocks": [ + { + "bbox": [ + 477, + 193, + 545, + 263 + ], + "lines": [ + { + "bbox": [ + 477, + 193, + 545, + 263 + ], + "spans": [ + { + "bbox": [ + 477, + 193, + 545, + 263 + ], + "type": "image", + "image_path": "05f585747c24d0b4496c7704edd85d5e06f031b142d4bf654d7e4d4f74b9d32a.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 48, + 310, + 272, + 571 + ], + "blocks": [ + { + "bbox": [ + 97, + 297, + 216, + 309 + ], + "lines": [ + { + "bbox": [ + 97, + 297, + 216, + 309 + ], + "spans": [ + { + "bbox": [ + 97, + 297, + 216, + 309 + ], + "type": "text", + "content": "Text-guided generation (1 step)" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 48, + 310, + 272, + 571 + ], + "lines": [ + { + "bbox": [ + 48, + 310, + 272, + 571 + ], + "spans": [ + { + "bbox": [ + 48, + 310, + 272, + 571 + ], + "type": "image", + "image_path": "073499f830b61bcd4676870d20a763147b3120cd4955f99111068ee31a1d942d.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 97, + 574, + 219, + 584 + ], + "lines": [ + { + "bbox": [ + 97, + 574, + 219, + 584 + ], + "spans": [ + { + "bbox": [ + 97, + 574, + 219, + 584 + ], + "type": "text", + "content": "Text-guided generation (4 steps)" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 45, + 592, + 547, + 626 + ], + "lines": [ + { + "bbox": [ + 45, + 592, + 547, + 626 + ], + "spans": [ + { + "bbox": [ + 45, + 592, + 547, + 626 + ], + "type": "text", + "content": "Figure 1. Distilled Stable Diffusion samples generated by our method. Our two-stage distillation approach is able to generate realistic images using only 1 to 4 denoising steps on various tasks. Compared to standard classifier-free guided diffusion models, we reduce the total number of sampling steps by at least " + }, + { + "bbox": [ + 45, + 592, + 547, + 626 + ], + "type": "inline_equation", + "content": "20\\times" + }, + { + "bbox": [ + 45, + 592, + 547, + 626 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 272, + 281, + 340, + 317 + ], + "blocks": [ + { + "bbox": [ + 272, + 281, + 340, + 317 + ], + "lines": [ + { + "bbox": [ + 272, + 281, + 340, + 317 + ], + "spans": [ + { + "bbox": [ + 272, + 281, + 340, + 317 + ], + "type": "image", + "image_path": "930d28fbb1d978339636729faa2bba17db923fb6442eff4e688f3e04265de9d3.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 341, + 320, + 479, + 331 + ], + "lines": [ + { + "bbox": [ + 341, + 320, + 479, + 331 + ], + "spans": [ + { + "bbox": [ + 341, + 320, + 479, + 331 + ], + "type": "text", + "content": "Class-conditional generation (1 step)" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 341, + 281, + 408, + 317 + ], + "blocks": [ + { + "bbox": [ + 341, + 281, + 408, + 317 + ], + "lines": [ + { + "bbox": [ + 341, + 281, + 408, + 317 + ], + "spans": [ + { + "bbox": [ + 341, + 281, + 408, + 317 + ], + "type": "image", + "image_path": "761b62890495ac24622843713658135518820ba002808e144cf9e706c4bac63b.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 410, + 281, + 476, + 317 + ], + "blocks": [ + { + "bbox": [ + 410, + 281, + 476, + 317 + ], + "lines": [ + { + "bbox": [ + 410, + 281, + 476, + 317 + ], + "spans": [ + { + "bbox": [ + 410, + 281, + 476, + 317 + ], + "type": "image", + "image_path": "9d459bfa505fa2c1497c570b6921b2dcbab1629f9ef72ee07f7eb05569d69711.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 477, + 281, + 545, + 317 + ], + "blocks": [ + { + "bbox": [ + 477, + 281, + 545, + 317 + ], + "lines": [ + { + "bbox": [ + 477, + 281, + 545, + 317 + ], + "spans": [ + { + "bbox": [ + 477, + 281, + 545, + 317 + ], + "type": "image", + "image_path": "bc9ccaa9a22410980952341e3ff2ef4ef800dca6aca46799c97caf8821aa29af.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 272, + 334, + 408, + 403 + ], + "blocks": [ + { + "bbox": [ + 272, + 334, + 408, + 403 + ], + "lines": [ + { + "bbox": [ + 272, + 334, + 408, + 403 + ], + "spans": [ + { + "bbox": [ + 272, + 334, + 408, + 403 + ], + "type": "image", + "image_path": "03db0c5d9eb6441c41986da2d6e5386a4bfcd34e759af4a7098f41f8cbd3aedc.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 298, + 404, + 316, + 413 + ], + "lines": [ + { + "bbox": [ + 298, + 404, + 316, + 413 + ], + "spans": [ + { + "bbox": [ + 298, + 404, + 316, + 413 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 366, + 404, + 384, + 412 + ], + "lines": [ + { + "bbox": [ + 366, + 404, + 384, + 412 + ], + "spans": [ + { + "bbox": [ + 366, + 404, + 384, + 412 + ], + "type": "text", + "content": "Mask" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 410, + 334, + 476, + 403 + ], + "blocks": [ + { + "bbox": [ + 410, + 334, + 476, + 403 + ], + "lines": [ + { + "bbox": [ + 410, + 334, + 476, + 403 + ], + "spans": [ + { + "bbox": [ + 410, + 334, + 476, + 403 + ], + "type": "image", + "image_path": "3d549196b9cc2a5c0828462b35b7fd045c19ce2d6f5a135009ca75c236e28be3.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 430, + 404, + 456, + 412 + ], + "lines": [ + { + "bbox": [ + 430, + 404, + 456, + 412 + ], + "spans": [ + { + "bbox": [ + 430, + 404, + 456, + 412 + ], + "type": "text", + "content": "Result 1" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 477, + 334, + 545, + 403 + ], + "blocks": [ + { + "bbox": [ + 477, + 334, + 545, + 403 + ], + "lines": [ + { + "bbox": [ + 477, + 334, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 477, + 334, + 545, + 403 + ], + "type": "image", + "image_path": "830501d9a57ebdc82776b758ec87a7be8ed0b3b746a57941d6c5d4f01a86ce20.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 498, + 404, + 525, + 412 + ], + "lines": [ + { + "bbox": [ + 498, + 404, + 525, + 412 + ], + "spans": [ + { + "bbox": [ + 498, + 404, + 525, + 412 + ], + "type": "text", + "content": "Result 2" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_caption" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 272, + 426, + 340, + 563 + ], + "blocks": [ + { + "bbox": [ + 272, + 426, + 340, + 563 + ], + "lines": [ + { + "bbox": [ + 272, + 426, + 340, + 563 + ], + "spans": [ + { + "bbox": [ + 272, + 426, + 340, + 563 + ], + "type": "image", + "image_path": "e8067a0c257de4c5d754d4243c0bfd6104830da3c576bfcfb2530b2713bd7800.jpg" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 299, + 564, + 317, + 573 + ], + "lines": [ + { + "bbox": [ + 299, + 564, + 317, + 573 + ], + "spans": [ + { + "bbox": [ + 299, + 564, + 317, + 573 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_caption" + } + ], + "index": 43 + }, + { + "type": "image", + "bbox": [ + 340, + 426, + 408, + 563 + ], + "blocks": [ + { + "bbox": [ + 361, + 415, + 459, + 425 + ], + "lines": [ + { + "bbox": [ + 361, + 415, + 459, + 425 + ], + "spans": [ + { + "bbox": [ + 361, + 415, + 459, + 425 + ], + "type": "text", + "content": "Image inpainting (2 steps)" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 340, + 426, + 408, + 563 + ], + "lines": [ + { + "bbox": [ + 340, + 426, + 408, + 563 + ], + "spans": [ + { + "bbox": [ + 340, + 426, + 408, + 563 + ], + "type": "image", + "image_path": "8f7157f3ee46790d6efa8644ff1650d70db90df89d8f0c9ec99839a9b8f1f166.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 408, + 564, + 482, + 573 + ], + "lines": [ + { + "bbox": [ + 408, + 564, + 482, + 573 + ], + "spans": [ + { + "bbox": [ + 408, + 564, + 482, + 573 + ], + "type": "text", + "content": "Output (different styles)" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 343, + 574, + 479, + 585 + ], + "lines": [ + { + "bbox": [ + 343, + 574, + 479, + 585 + ], + "spans": [ + { + "bbox": [ + 343, + 574, + 479, + 585 + ], + "type": "text", + "content": "Image to image translation (3 steps)" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_caption" + } + ], + "index": 44 + }, + { + "type": "image", + "bbox": [ + 408, + 426, + 477, + 563 + ], + "blocks": [ + { + "bbox": [ + 408, + 426, + 477, + 563 + ], + "lines": [ + { + "bbox": [ + 408, + 426, + 477, + 563 + ], + "spans": [ + { + "bbox": [ + 408, + 426, + 477, + 563 + ], + "type": "image", + "image_path": "bce2b8a74953cbb79b3031e8b93e5a714c9b756ef5513ec934346347561f0ae0.jpg" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_body" + } + ], + "index": 45 + }, + { + "type": "image", + "bbox": [ + 477, + 426, + 545, + 563 + ], + "blocks": [ + { + "bbox": [ + 477, + 426, + 545, + 563 + ], + "lines": [ + { + "bbox": [ + 477, + 426, + 545, + 563 + ], + "spans": [ + { + "bbox": [ + 477, + 426, + 545, + 563 + ], + "type": "image", + "image_path": "e33e987788fca892e98d8b4cd50ca2126e3bf56eeb23e98d3b3c4e22fb849e55.jpg" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_body" + } + ], + "index": 46 + }, + { + "bbox": [ + 143, + 632, + 192, + 644 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 632, + 192, + 644 + ], + "spans": [ + { + "bbox": [ + 143, + 632, + 192, + 644 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 51 + }, + { + "bbox": [ + 47, + 658, + 288, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 288, + 693 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 288, + 693 + ], + "type": "text", + "content": "Classifier-free guided diffusion models have recently been shown to be highly effective at high-resolution image generation, and they have been widely used in large-scale diffusion" + } + ] + } + ], + "index": 52 + }, + { + "bbox": [ + 304, + 633, + 548, + 705 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 633, + 548, + 705 + ], + "spans": [ + { + "bbox": [ + 304, + 633, + 548, + 705 + ], + "type": "text", + "content": "frameworks including DALL·E 2, Stable Diffusion and Image. However, a downside of classifier-free guided diffusion models is that they are computationally expensive at inference time since they require evaluating two diffusion models, a class-conditional model and an unconditional model, tens to hundreds of times. To deal with this limitation, we pro" + } + ] + } + ], + "index": 53 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 495, + 15 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 495, + 15 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 495, + 15 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 198, + 13, + 442, + 24 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 13, + 442, + 24 + ], + "spans": [ + { + "bbox": [ + 198, + 13, + 442, + 24 + ], + "type": "text", + "content": "Except for this watermark, it is identical to the accepted version;" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "spans": [ + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "type": "text", + "content": "the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 702, + 228, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 228, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 228, + 713 + ], + "type": "text", + "content": "*Work partially done during an internship at Google" + } + ] + } + ], + "index": 54 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14297" + } + ] + } + ], + "index": 55 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 313 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 313 + ], + "type": "text", + "content": "pose an approach to distilling classifier-free guided diffusion models into models that are fast to sample from: Given a pre-trained classifier-free guided model, we first learn a single model to match the output of the combined conditional and unconditional models, and then we progressively distill that model to a diffusion model that requires much fewer sampling steps. For standard diffusion models trained on the pixel-space, our approach is able to generate images visually comparable to that of the original model using as few as 4 sampling steps on ImageNet 64x64 and CIFAR-10, achieving FID/IS scores comparable to that of the original model while being up to 256 times faster to sample from. For diffusion models trained on the latent-space (e.g., Stable Diffusion), our approach is able to generate high-fidelity images using as few as 1 to 4 denoising steps, accelerating inference by at least 10-fold compared to existing methods on ImageNet 256x256 and LAION datasets. We further demonstrate the effectiveness of our approach on text-guided image editing and inpainting, where our distilled model is able to generate high-quality results using as few as 2-4 denoising steps." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 335, + 128, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 335, + 128, + 346 + ], + "spans": [ + { + "bbox": [ + 47, + 335, + 128, + 346 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 354, + 289, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 354, + 289, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 289, + 689 + ], + "type": "text", + "content": "Denoising diffusion probabilistic models (DDPMs) [4, 37, 39, 40] have achieved state-of-the-art performance on image generation [22, 26–28, 31], audio synthesis [11], molecular generation [44], and likelihood estimation [10]. Classifier-free guidance [6] further improves the sample quality of diffusion models and has been widely used in large-scale diffusion model frameworks including GLIDE [23], Stable Diffusion [28], DALL-E 2 [26], and Imagen [31]. However, one key limitation of classifier-free guidance is its low sampling efficiency—it requires evaluating two diffusion models tens to hundreds of times to generate one sample. This limitation has hindered the application of classifier-free guidance models in real-world settings. Although distillation approaches have been proposed for diffusion models [33, 38], these approaches are not directly applicable to classifier-free guided diffusion models. To deal with this issue, we propose a two-stage distillation approach to improving the sampling efficiency of classifier-free guided models. In the first stage, we introduce a single student model to match the combined output of the two diffusion models of the teacher. In the second stage, we progressively distill the model learned from the first stage to a fewer-step model using the approach introduced in [33]. Using our approach, a single distilled model is able to handle a wide range of different guidance strengths, allowing for the trade-off between sample quality and diversity efficiently. To sample from our model, we consider existing deterministic samplers in the literature [33, 38] and further propose a stochastic sampling process." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "content": "Our distillation framework can not only be applied to standard diffusion models trained on the pixel-space [4, 36, 39]," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 72, + 547, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 337 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 337 + ], + "type": "text", + "content": "but also diffusion models trained on the latent-space of an autoencoder [28, 35] (e.g., Stable Diffusion [28]). For diffusion models directly trained on the pixel-space, our experiments on ImageNet 64x64 and CIFAR-10 show that the proposed distilled model can generate samples visually comparable to that of the teacher using only 4 steps and is able to achieve comparable FID/IS scores as the teacher model using as few as 4 to 16 steps on a wide range of guidance strengths (see Fig. 2). For diffusion model trained on the latent-space of an encoder [28, 35], our approach is able to achieve comparable visual quality to the base model using as few as 1 to 4 sampling steps (at least " + }, + { + "bbox": [ + 304, + 72, + 547, + 337 + ], + "type": "inline_equation", + "content": "10 \\times" + }, + { + "bbox": [ + 304, + 72, + 547, + 337 + ], + "type": "text", + "content": " fewer steps than the base model) on ImageNet " + }, + { + "bbox": [ + 304, + 72, + 547, + 337 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 304, + 72, + 547, + 337 + ], + "type": "text", + "content": " and LAION " + }, + { + "bbox": [ + 304, + 72, + 547, + 337 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 304, + 72, + 547, + 337 + ], + "type": "text", + "content": ", matching the performance of the teacher (as evaluated by FID) with only 2-4 sampling steps. To the best of our knowledge, our work is the first to demonstrate the effectiveness of distillation for both pixel-space and latent-space classifier-free diffusion models. Finally, we apply our method to text-guided image inpainting and text-guided image editing tasks [20], where we reduce the total number of sampling steps to as few as 2-4 steps, demonstrating the potential of the proposed framework in style-transfer and image-editing applications [20, 41]." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 307, + 347, + 545, + 459 + ], + "blocks": [ + { + "bbox": [ + 307, + 347, + 545, + 459 + ], + "lines": [ + { + "bbox": [ + 307, + 347, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 307, + 347, + 545, + 459 + ], + "type": "image", + "image_path": "e73a9203a22b09ac814fb8e05d548d163277f88bbb1752ecc5334f7bf4424773.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 466, + 547, + 522 + ], + "lines": [ + { + "bbox": [ + 304, + 466, + 547, + 522 + ], + "spans": [ + { + "bbox": [ + 304, + 466, + 547, + 522 + ], + "type": "text", + "content": "Figure 2. Class-conditional samples from our two-stage (deterministic) approach on ImageNet 64x64 for diffusion models trained on the pixel-space. By varying the guidance weight " + }, + { + "bbox": [ + 304, + 466, + 547, + 522 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 304, + 466, + 547, + 522 + ], + "type": "text", + "content": ", our distilled model is able to trade-off between sample diversity and quality, while achieving good results using as few as one sampling step." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 533, + 488, + 546 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 533, + 488, + 546 + ], + "spans": [ + { + "bbox": [ + 306, + 533, + 488, + 546 + ], + "type": "text", + "content": "2. Background on diffusion models" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 555, + 545, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 555, + 545, + 603 + ], + "spans": [ + { + "bbox": [ + 304, + 555, + 545, + 603 + ], + "type": "text", + "content": "Given samples " + }, + { + "bbox": [ + 304, + 555, + 545, + 603 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 304, + 555, + 545, + 603 + ], + "type": "text", + "content": " from a data distribution " + }, + { + "bbox": [ + 304, + 555, + 545, + 603 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{data}}(\\mathbf{x})" + }, + { + "bbox": [ + 304, + 555, + 545, + 603 + ], + "type": "text", + "content": ", noise scheduling functions " + }, + { + "bbox": [ + 304, + 555, + 545, + 603 + ], + "type": "inline_equation", + "content": "\\alpha_{t}" + }, + { + "bbox": [ + 304, + 555, + 545, + 603 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 555, + 545, + 603 + ], + "type": "inline_equation", + "content": "\\sigma_{t}" + }, + { + "bbox": [ + 304, + 555, + 545, + 603 + ], + "type": "text", + "content": ", we train a diffusion model " + }, + { + "bbox": [ + 304, + 555, + 545, + 603 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_{\\theta}" + }, + { + "bbox": [ + 304, + 555, + 545, + 603 + ], + "type": "text", + "content": ", with parameter " + }, + { + "bbox": [ + 304, + 555, + 545, + 603 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 555, + 545, + 603 + ], + "type": "text", + "content": ", via minimizing the weighted mean squared error [4, 36, 39, 40]" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 613, + 547, + 628 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 613, + 547, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 613, + 547, + 628 + ], + "type": "interline_equation", + "content": "\\mathbb {E} _ {t \\sim U [ 0, 1 ], \\mathbf {x} \\sim p _ {\\text {d a t a}} (\\mathbf {x}), \\mathbf {z} _ {t} \\sim q \\left(\\mathbf {z} _ {t} \\mid \\mathbf {x}\\right)} \\left[ \\omega \\left(\\lambda_ {t}\\right) \\right| | \\hat {\\mathbf {x}} _ {\\boldsymbol {\\theta}} \\left(\\mathbf {z} _ {t}\\right) - \\mathbf {x} | | _ {2} ^ {2} ], \\tag {1}", + "image_path": "514e38b3a6dc388e0924aa527d5654ac954d14dd359f02b8b763a4feca58cee9.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 639, + 545, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 639, + 545, + 676 + ], + "spans": [ + { + "bbox": [ + 304, + 639, + 545, + 676 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 639, + 545, + 676 + ], + "type": "inline_equation", + "content": "\\lambda_{t} = \\log [\\alpha_{t}^{2} / \\sigma_{t}^{2}]" + }, + { + "bbox": [ + 304, + 639, + 545, + 676 + ], + "type": "text", + "content": " is a signal-to-noise ratio [10], " + }, + { + "bbox": [ + 304, + 639, + 545, + 676 + ], + "type": "inline_equation", + "content": "q(\\mathbf{z}_t|\\mathbf{x}) = \\mathcal{N}(\\mathbf{z}_t;\\alpha_t\\mathbf{x},\\sigma_t^2\\mathbf{I})" + }, + { + "bbox": [ + 304, + 639, + 545, + 676 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 639, + 545, + 676 + ], + "type": "inline_equation", + "content": "\\omega (\\lambda_t)" + }, + { + "bbox": [ + 304, + 639, + 545, + 676 + ], + "type": "text", + "content": " is a pre-specified weighting function [10]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 677, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 714 + ], + "type": "text", + "content": "Once the diffusion model " + }, + { + "bbox": [ + 304, + 677, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_{\\theta}" + }, + { + "bbox": [ + 304, + 677, + 547, + 714 + ], + "type": "text", + "content": " is trained, one can use discrete-time DDIM sampler [38] to sample from the model. Specifically, the DDIM sampler starts with " + }, + { + "bbox": [ + 304, + 677, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_1 \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{I})" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14298" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 141, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 141, + 84 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 141, + 84 + ], + "type": "text", + "content": "and updates as follows" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 58, + 91, + 287, + 116 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 91, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 58, + 91, + 287, + 116 + ], + "type": "interline_equation", + "content": "\\mathbf {z} _ {s} = \\alpha_ {s} \\hat {\\mathbf {x}} _ {\\boldsymbol {\\theta}} (\\mathbf {z} _ {t}) + \\sigma_ {s} \\frac {\\mathbf {z} _ {t} - \\alpha_ {t} \\hat {\\mathbf {x}} _ {\\boldsymbol {\\theta}} (\\mathbf {z} _ {t})}{\\sigma_ {t}}, s = t - 1 / N \\tag {2}", + "image_path": "7bf649a29fe71ff4d6f6b5c0b32bcde46a027a640b780f1009b27bd92a3c031c.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 121, + 287, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 121, + 287, + 145 + ], + "spans": [ + { + "bbox": [ + 46, + 121, + 287, + 145 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 46, + 121, + 287, + 145 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 121, + 287, + 145 + ], + "type": "text", + "content": " the total number of sampling steps. The final sample will then be generated using " + }, + { + "bbox": [ + 46, + 121, + 287, + 145 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_{\\theta}(\\mathbf{z}_0)" + }, + { + "bbox": [ + 46, + 121, + 287, + 145 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 145, + 288, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 145, + 288, + 324 + ], + "spans": [ + { + "bbox": [ + 46, + 145, + 288, + 324 + ], + "type": "text", + "content": "Classifier-free guidance Classifier-free guidance [6] is an effective approach shown to significantly improve the sample quality of class-conditioned diffusion models, and has been widely used in large-scale diffusion models including GLIDE [23], Stable Diffusion [28], DALL-E 2 [26] and Imagen [31]. Specifically, it introduces a guidance weight parameter " + }, + { + "bbox": [ + 46, + 145, + 288, + 324 + ], + "type": "inline_equation", + "content": "w \\in \\mathbb{R}^{\\geq 0}" + }, + { + "bbox": [ + 46, + 145, + 288, + 324 + ], + "type": "text", + "content": " to trade-off between sample quality and diversity. To generate a sample, classifier-free guidance evaluates both a conditional diffusion model " + }, + { + "bbox": [ + 46, + 145, + 288, + 324 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_{c,\\theta}" + }, + { + "bbox": [ + 46, + 145, + 288, + 324 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 145, + 288, + 324 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 145, + 288, + 324 + ], + "type": "text", + "content": " is the context (e.g., class label, text prompt) to be conditioned on, and a jointly trained unconditional diffusion model " + }, + { + "bbox": [ + 46, + 145, + 288, + 324 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_{\\theta}" + }, + { + "bbox": [ + 46, + 145, + 288, + 324 + ], + "type": "text", + "content": " at each update step, using " + }, + { + "bbox": [ + 46, + 145, + 288, + 324 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_{\\theta}^{w} = (1 + w)\\hat{\\mathbf{x}}_{c,\\theta} - w\\hat{\\mathbf{x}}_{\\theta}" + }, + { + "bbox": [ + 46, + 145, + 288, + 324 + ], + "type": "text", + "content": " as the model prediction in Eq. (2). As each sampling update requires evaluating two diffusion models, sampling with classifier-free guidance is often expensive [6]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 325, + 288, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 325, + 288, + 408 + ], + "spans": [ + { + "bbox": [ + 46, + 325, + 288, + 408 + ], + "type": "text", + "content": "Progressive distillation Our approach is inspired by progressive distillation [33], an effective method for improving the sampling speed of (unguided) diffusion models by repeated distillation. Until now, this method could not be directly applied to distilling classifier-free guided models or studied for samplers other than the deterministic DDIM sampler [33, 38]. In this paper we resolve these shortcomings." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 408, + 288, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 408, + 288, + 527 + ], + "spans": [ + { + "bbox": [ + 46, + 408, + 288, + 527 + ], + "type": "text", + "content": "Latent diffusion models (LDMs) [21,24,28,35] increase the training and inference efficiency of diffusion models (directly learned on the pixel-space) by modeling images in the latent space of a pre-trained regularized autoencoder, where the latent representations are usually of lower dimensionality than the pixel-space. Latent diffusion models can be considered as an alternative to cascaded diffusion approaches [5], which rely on one or more super-resolution diffusion models to scale up a low-dimensional image to the desired target resolution." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 528, + 287, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 528, + 287, + 564 + ], + "spans": [ + { + "bbox": [ + 46, + 528, + 287, + 564 + ], + "type": "text", + "content": "In this work, we will apply our distillation framework to classifier-free guided diffusion models learned on both pixel-space [4, 36, 39] and latent-space [21, 24, 28, 35]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 574, + 239, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 574, + 239, + 588 + ], + "spans": [ + { + "bbox": [ + 47, + 574, + 239, + 588 + ], + "type": "text", + "content": "3. Distilling a guided diffusion model" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 594, + 287, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 287, + 677 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 287, + 677 + ], + "type": "text", + "content": "In the following, we discuss our approach for distilling a classifier-free guided diffusion model [6] into a student model that requires fewer steps to sample from. Using a single distilled model conditioned on the guidance strength, our model can capture a wide range of classifier-free guidance levels, allowing for the trade-off between sample quality and diversity efficiently." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "text", + "content": "Given a trained guided model " + }, + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "inline_equation", + "content": "[\\hat{\\mathbf{x}}_{c,\\theta},\\hat{\\mathbf{x}}_{\\theta}]" + }, + { + "bbox": [ + 46, + 677, + 288, + 714 + ], + "type": "text", + "content": " (teacher) either on the pixel-space or latent-space, our approach can be decomposed into two stages." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 72, + 427, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 427, + 84 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 427, + 84 + ], + "type": "text", + "content": "3.1. Stage-one distillation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 90, + 547, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 90, + 547, + 196 + ], + "spans": [ + { + "bbox": [ + 304, + 90, + 547, + 196 + ], + "type": "text", + "content": "In the first stage, we introduce a student model " + }, + { + "bbox": [ + 304, + 90, + 547, + 196 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_{\\eta_1}(\\mathbf{z}_t,w)" + }, + { + "bbox": [ + 304, + 90, + 547, + 196 + ], + "type": "text", + "content": ", with learnable parameter " + }, + { + "bbox": [ + 304, + 90, + 547, + 196 + ], + "type": "inline_equation", + "content": "\\pmb{\\eta}_{1}" + }, + { + "bbox": [ + 304, + 90, + 547, + 196 + ], + "type": "text", + "content": ", to match the output of the teacher at any time-step " + }, + { + "bbox": [ + 304, + 90, + 547, + 196 + ], + "type": "inline_equation", + "content": "t\\in [0,1]" + }, + { + "bbox": [ + 304, + 90, + 547, + 196 + ], + "type": "text", + "content": ". The student model can either be a continuous-time model [40] or a discrete-time model [4, 38] depending on whether the teacher model is discrete or continuous. For simplicity, in the following discussion, we assume both the student and teacher models are continuous as the algorithm for discrete models is almost identical." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 198, + 547, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 198, + 547, + 307 + ], + "spans": [ + { + "bbox": [ + 304, + 198, + 547, + 307 + ], + "type": "text", + "content": "A key functionality of classifier-free guidance [6] is its ability to easily trade-off between sample quality and diversity, which is controlled by a \"guidance strength\" parameter. This property has demonstrated utility in real-world applications [6,23,26,28,31], where the optimal \"guidance strength\" is often a user preference. Thus, we would also want our distilled model to maintain this property. Given a range of guidance strengths " + }, + { + "bbox": [ + 304, + 198, + 547, + 307 + ], + "type": "inline_equation", + "content": "\\left[w_{\\min},w_{\\max}\\right]" + }, + { + "bbox": [ + 304, + 198, + 547, + 307 + ], + "type": "text", + "content": " we are interested in, we optimize the student model using the following objective" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 310, + 313, + 546, + 338 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 313, + 546, + 338 + ], + "spans": [ + { + "bbox": [ + 310, + 313, + 546, + 338 + ], + "type": "interline_equation", + "content": "\\mathbb {E} _ {w \\sim p _ {w}, t \\sim U [ 0, 1 ], \\mathbf {x} \\sim p _ {\\mathrm {d a t a}} (\\mathbf {x})} \\left[ \\omega \\left(\\lambda_ {t}\\right) \\| \\hat {\\mathbf {x}} _ {\\boldsymbol {\\eta} _ {1}} \\left(\\mathbf {z} _ {t}, w\\right) - \\hat {\\mathbf {x}} _ {\\boldsymbol {\\theta}} ^ {w} \\left(\\mathbf {z} _ {t}\\right) \\| _ {2} ^ {2} \\right], \\tag {3}", + "image_path": "35ec48582e2af6f4ea17fa381fc14c2c092a7e1ee2d8f02a93db225a268626b9.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 344, + 547, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 344, + 547, + 417 + ], + "spans": [ + { + "bbox": [ + 304, + 344, + 547, + 417 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 344, + 547, + 417 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_{\\pmb{\\theta}}^{w}(\\mathbf{z}_{t}) = (1 + w)\\hat{\\mathbf{x}}_{c,\\pmb{\\theta}}(\\mathbf{z}_{t}) - w\\hat{\\mathbf{x}}_{\\pmb{\\theta}}(\\mathbf{z}_{t}), \\mathbf{z}_{t} \\sim q(\\mathbf{z}_{t}|\\mathbf{x})" + }, + { + "bbox": [ + 304, + 344, + 547, + 417 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 344, + 547, + 417 + ], + "type": "inline_equation", + "content": "p_w(w) = U[w_{\\min}, w_{\\max}]" + }, + { + "bbox": [ + 304, + 344, + 547, + 417 + ], + "type": "text", + "content": ". Note that here, our distilled model " + }, + { + "bbox": [ + 304, + 344, + 547, + 417 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_{\\pmb{\\eta}_1}(\\mathbf{z}_t, w)" + }, + { + "bbox": [ + 304, + 344, + 547, + 417 + ], + "type": "text", + "content": " is also conditioned on the context " + }, + { + "bbox": [ + 304, + 344, + 547, + 417 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 344, + 547, + 417 + ], + "type": "text", + "content": " (e.g., text prompt), but we drop the notation " + }, + { + "bbox": [ + 304, + 344, + 547, + 417 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 344, + 547, + 417 + ], + "type": "text", + "content": " in the paper for simplicity. We provide the detailed training algorithm in Algorithm 1 in the supplement." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 418, + 547, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 418, + 547, + 597 + ], + "spans": [ + { + "bbox": [ + 304, + 418, + 547, + 597 + ], + "type": "text", + "content": "To incorporate the guidance weight " + }, + { + "bbox": [ + 304, + 418, + 547, + 597 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 304, + 418, + 547, + 597 + ], + "type": "text", + "content": ", we introduce a " + }, + { + "bbox": [ + 304, + 418, + 547, + 597 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 304, + 418, + 547, + 597 + ], + "type": "text", + "content": "-conditioned model, where " + }, + { + "bbox": [ + 304, + 418, + 547, + 597 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 304, + 418, + 547, + 597 + ], + "type": "text", + "content": " is fed as an input to the student model. To better capture the feature, we apply Fourier embedding to " + }, + { + "bbox": [ + 304, + 418, + 547, + 597 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 304, + 418, + 547, + 597 + ], + "type": "text", + "content": ", which is then incorporated into the diffusion model backbone in a way similar to how the time-step was incorporated in [10, 33]. As initialization plays a key role in the performance [33], we initialize the student model with the same parameters as the conditional model of the teacher, except for the newly introduced parameters related to " + }, + { + "bbox": [ + 304, + 418, + 547, + 597 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 304, + 418, + 547, + 597 + ], + "type": "text", + "content": "-conditioning. The model architecture we use is a U-Net model similar to the ones used in [6] for pixel-space diffusion models and [1, 28] for latent-space diffusion models. We use the same number of channels and attention as used in [6] and the open-sourced Stable Diffusion repository* for our experiments. We provide more details in the supplement." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 604, + 427, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 604, + 427, + 616 + ], + "spans": [ + { + "bbox": [ + 306, + 604, + 427, + 616 + ], + "type": "text", + "content": "3.2. Stage-two distillation" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 623, + 547, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 623, + 547, + 696 + ], + "spans": [ + { + "bbox": [ + 304, + 623, + 547, + 696 + ], + "type": "text", + "content": "In the second stage, we consider a discrete time-step scenario and progressively distill the learned model from the first-stage " + }, + { + "bbox": [ + 304, + 623, + 547, + 696 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_{\\eta_1}(\\mathbf{z}_t,w)" + }, + { + "bbox": [ + 304, + 623, + 547, + 696 + ], + "type": "text", + "content": " into an fewer-step student model " + }, + { + "bbox": [ + 304, + 623, + 547, + 696 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_{\\eta_2}(\\mathbf{z}_t,w)" + }, + { + "bbox": [ + 304, + 623, + 547, + 696 + ], + "type": "text", + "content": " with learnable parameter " + }, + { + "bbox": [ + 304, + 623, + 547, + 696 + ], + "type": "inline_equation", + "content": "\\eta_{2}" + }, + { + "bbox": [ + 304, + 623, + 547, + 696 + ], + "type": "text", + "content": ", by halving the number of sampling steps each time. Letting " + }, + { + "bbox": [ + 304, + 623, + 547, + 696 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 623, + 547, + 696 + ], + "type": "text", + "content": " denote the number of sampling steps, given " + }, + { + "bbox": [ + 304, + 623, + 547, + 696 + ], + "type": "inline_equation", + "content": "w\\sim U[w_{\\min},w_{\\max}]" + }, + { + "bbox": [ + 304, + 623, + 547, + 696 + ], + "type": "text", + "content": " and" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 316, + 703, + 529, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 703, + 529, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 703, + 529, + 713 + ], + "type": "text", + "content": "*https://github.com/CompVis/stable-diffusion" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14299" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 108, + 230 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 108, + 230 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 108, + 230 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 108, + 230 + ], + "type": "image", + "image_path": "019b251179ebf0264ae30e436a0b7e9f5f3aafa287ebba6c38b489f2427f170f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 102, + 232, + 169, + 242 + ], + "lines": [ + { + "bbox": [ + 102, + 232, + 169, + 242 + ], + "spans": [ + { + "bbox": [ + 102, + 232, + 169, + 242 + ], + "type": "text", + "content": "(a) 2 denoising steps" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 108, + 70, + 226, + 230 + ], + "blocks": [ + { + "bbox": [ + 108, + 70, + 226, + 230 + ], + "lines": [ + { + "bbox": [ + 108, + 70, + 226, + 230 + ], + "spans": [ + { + "bbox": [ + 108, + 70, + 226, + 230 + ], + "type": "image", + "image_path": "98f99a856052c3f482501fe0669cb0c10b53bd98f17a18ce6e5906bc7babb438.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 251, + 545, + 274 + ], + "lines": [ + { + "bbox": [ + 47, + 251, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 47, + 251, + 545, + 274 + ], + "type": "text", + "content": "Figure 4. Text-guided generation on LAION (512x512) using our distilled Stable Diffusion model. Our model is able to generate high-quality image samples using 2, 4 or 8 denoising steps, significantly improving the inference efficiency of Stable Diffusion." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 227, + 70, + 376, + 230 + ], + "blocks": [ + { + "bbox": [ + 227, + 70, + 376, + 230 + ], + "lines": [ + { + "bbox": [ + 227, + 70, + 376, + 230 + ], + "spans": [ + { + "bbox": [ + 227, + 70, + 376, + 230 + ], + "type": "image", + "image_path": "436fa4e547ed4588699e76e23a4412f052419a9cf226bfcc0136ebce27dbc277.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 263, + 232, + 331, + 242 + ], + "lines": [ + { + "bbox": [ + 263, + 232, + 331, + 242 + ], + "spans": [ + { + "bbox": [ + 263, + 232, + 331, + 242 + ], + "type": "text", + "content": "(b) 4 denoising steps" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 377, + 70, + 537, + 230 + ], + "blocks": [ + { + "bbox": [ + 377, + 70, + 537, + 230 + ], + "lines": [ + { + "bbox": [ + 377, + 70, + 537, + 230 + ], + "spans": [ + { + "bbox": [ + 377, + 70, + 537, + 230 + ], + "type": "image", + "image_path": "0cd6729a2cc48efc54e30037e083d383a4eda77f75bf42cc7ed659ecefeec9e6.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 424, + 232, + 491, + 242 + ], + "lines": [ + { + "bbox": [ + 424, + 232, + 491, + 242 + ], + "spans": [ + { + "bbox": [ + 424, + 232, + 491, + 242 + ], + "type": "text", + "content": "(c) 8 denoising steps" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 50, + 275, + 545, + 403 + ], + "blocks": [ + { + "bbox": [ + 50, + 275, + 545, + 403 + ], + "lines": [ + { + "bbox": [ + 50, + 275, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 50, + 275, + 545, + 403 + ], + "type": "table", + "html": "
w=0w=0.3w=1w=4
MethodFID (↓)IS (↑)FID (↓)IS (↑)FID (↓)IS (↑)FID (↓)IS (↑)
Ours 1-step (D/S)22.74 / 26.9125.51 / 23.5514.85 / 18.4837.09 / 33.307.54 / 8.9275.19 / 67.8018.72 / 17.85157.46 / 148.97
Ours 4-step (D/S)4.14 / 3.9146.64 / 48.922.17 / 2.2469.64 / 73.737.95 / 8.51128.98 / 135.3626.45 / 27.33207.45 / 216.56
Ours 8-step (D/S)2.79 / 2.4450.72 / 55.032.05 / 2.3176.01 / 83.009.33 / 10.56136.47 / 147.3926.62 / 27.84203.47 / 219.89
Ours 16-step (D/S)2.44 / 2.1052.53 / 57.812.20 / 2.5679.47 / 87.509.99 / 11.63139.11 / 153.1726.53 / 27.69204.13 / 218.70
Single-w 1-step19.6124.0011.7036.956.6474.4119.857170.69
Single-w 4-step4.7938.772.3462.088.23118.5227.75219.64
Single-w 8-step3.3942.132.3268.769.69125.2027.67218.08
Single-w 16-step2.9743.632.5670.9710.34127.7027.40216.52
DDIM 16x2-step [38]7.6837.605.3360.839.53112.7521.56195.17
DDIM 32x2-step [38]5.0340.937.479.339.26126.2223.03213.23
DDIM 64x2-step [38]3.7443.165.529.519.53133.1723.64217.88
Teacher (DDIM 1024x2-step)2.9244.812.3674.839.84139.5023.94224.74
", + "image_path": "4dfd2fda850903cabadf2eeb5d001c98f5e0f10c823a687410de2e7502df96e1.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 411, + 546, + 456 + ], + "lines": [ + { + "bbox": [ + 46, + 411, + 546, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 411, + 546, + 456 + ], + "type": "text", + "content": "Table 1. ImageNet 64x64 distillation results for pixel-space diffusion models (" + }, + { + "bbox": [ + 46, + 411, + 546, + 456 + ], + "type": "inline_equation", + "content": "w = 0" + }, + { + "bbox": [ + 46, + 411, + 546, + 456 + ], + "type": "text", + "content": " refers to non-guided models). For our method, " + }, + { + "bbox": [ + 46, + 411, + 546, + 456 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 46, + 411, + 546, + 456 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 411, + 546, + 456 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 46, + 411, + 546, + 456 + ], + "type": "text", + "content": " stand for deterministic and stochastic sampler respectively. We observe that training the model conditioned on a guidance interval " + }, + { + "bbox": [ + 46, + 411, + 546, + 456 + ], + "type": "inline_equation", + "content": "w \\in [0,4]" + }, + { + "bbox": [ + 46, + 411, + 546, + 456 + ], + "type": "text", + "content": " performs comparably with training a model on a fixed " + }, + { + "bbox": [ + 46, + 411, + 546, + 456 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 46, + 411, + 546, + 456 + ], + "type": "text", + "content": " (see Single-" + }, + { + "bbox": [ + 46, + 411, + 546, + 456 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 46, + 411, + 546, + 456 + ], + "type": "text", + "content": "). Our approach significantly outperforms DDIM when using fewer steps, and is able to match the teacher performance using as few as 8 to 16 steps." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 467, + 288, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 467, + 288, + 598 + ], + "spans": [ + { + "bbox": [ + 46, + 467, + 288, + 598 + ], + "type": "inline_equation", + "content": "t \\in \\{1, \\dots, N\\}" + }, + { + "bbox": [ + 46, + 467, + 288, + 598 + ], + "type": "text", + "content": ", we train the student model to match the output of two-step DDIM sampling of the teacher (i.e., from " + }, + { + "bbox": [ + 46, + 467, + 288, + 598 + ], + "type": "inline_equation", + "content": "t/N" + }, + { + "bbox": [ + 46, + 467, + 288, + 598 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 467, + 288, + 598 + ], + "type": "inline_equation", + "content": "t - 0.5/N" + }, + { + "bbox": [ + 46, + 467, + 288, + 598 + ], + "type": "text", + "content": " and from " + }, + { + "bbox": [ + 46, + 467, + 288, + 598 + ], + "type": "inline_equation", + "content": "t - 0.5/N" + }, + { + "bbox": [ + 46, + 467, + 288, + 598 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 467, + 288, + 598 + ], + "type": "inline_equation", + "content": "t - 1/N" + }, + { + "bbox": [ + 46, + 467, + 288, + 598 + ], + "type": "text", + "content": ") in one step, following the approach of [33]. After distilling the " + }, + { + "bbox": [ + 46, + 467, + 288, + 598 + ], + "type": "inline_equation", + "content": "2N" + }, + { + "bbox": [ + 46, + 467, + 288, + 598 + ], + "type": "text", + "content": " steps in the teacher model to " + }, + { + "bbox": [ + 46, + 467, + 288, + 598 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 467, + 288, + 598 + ], + "type": "text", + "content": " steps in the student model, we can use the " + }, + { + "bbox": [ + 46, + 467, + 288, + 598 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 467, + 288, + 598 + ], + "type": "text", + "content": "-step student model as the new teacher model, repeat the same procedure, and distill the teacher model into a " + }, + { + "bbox": [ + 46, + 467, + 288, + 598 + ], + "type": "inline_equation", + "content": "N/2" + }, + { + "bbox": [ + 46, + 467, + 288, + 598 + ], + "type": "text", + "content": "-step student model. At each step, we initialize the student model with the parameters of the teacher. We provide the training algorithm and extra details in the supplementary material." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 609, + 280, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 609, + 280, + 624 + ], + "spans": [ + { + "bbox": [ + 47, + 609, + 280, + 624 + ], + "type": "text", + "content": "3.3. " + }, + { + "bbox": [ + 47, + 609, + 280, + 624 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 47, + 609, + 280, + 624 + ], + "type": "text", + "content": "-step deterministic and stochastic sampling" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "content": "Once the model " + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_{\\eta_2}" + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "content": " is trained, given a specified guidance strength " + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "inline_equation", + "content": "w \\in [w_{\\min}, w_{\\max}]" + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "content": ", we can perform sampling via the DDIM update rule in Eq. (2). We note that given the distilled model " + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_{\\eta_2}" + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "content": ", this sampling procedure is deterministic given the initialization " + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_1^w" + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "content": ". In fact, we can also perform " + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 630, + 288, + 715 + ], + "type": "text", + "content": "-step stochastic sampling: We apply one deterministic sampling step with two-times the original step-length (i.e.," + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 467, + 545, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 467, + 545, + 527 + ], + "spans": [ + { + "bbox": [ + 304, + 467, + 545, + 527 + ], + "type": "text", + "content": "the same as a " + }, + { + "bbox": [ + 304, + 467, + 545, + 527 + ], + "type": "inline_equation", + "content": "N/2" + }, + { + "bbox": [ + 304, + 467, + 545, + 527 + ], + "type": "text", + "content": "-step deterministic sampler) and then perform one stochastic step backward (i.e., perturb with noise) using the original step-length, a process inspired by [9]. With " + }, + { + "bbox": [ + 304, + 467, + 545, + 527 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_1^w \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{I})" + }, + { + "bbox": [ + 304, + 467, + 545, + 527 + ], + "type": "text", + "content": ", we use the following update rule when " + }, + { + "bbox": [ + 304, + 467, + 545, + 527 + ], + "type": "inline_equation", + "content": "t > 1/N" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 347, + 536, + 545, + 562 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 536, + 545, + 562 + ], + "spans": [ + { + "bbox": [ + 347, + 536, + 545, + 562 + ], + "type": "interline_equation", + "content": "\\mathbf {z} _ {k} ^ {w} = \\alpha_ {k} \\hat {\\mathbf {x}} _ {\\boldsymbol {\\eta} _ {2}} \\left(\\mathbf {z} _ {t} ^ {w}\\right) + \\sigma_ {k} \\frac {\\mathbf {z} _ {t} ^ {w} - \\alpha_ {t} \\hat {\\mathbf {x}} _ {\\boldsymbol {\\eta} _ {2}} ^ {w} \\left(\\mathbf {z} _ {t}\\right)}{\\sigma_ {t}}, \\tag {4}", + "image_path": "aa48962869c672d71dc42aae139aac7bcc0f370f297cba2087484cced7f80d14.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 321, + 563, + 545, + 578 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 563, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 321, + 563, + 545, + 578 + ], + "type": "interline_equation", + "content": "\\text {w h e r e} \\mathbf {z} _ {s} ^ {w} = \\left(\\alpha_ {s} / \\alpha_ {k}\\right) \\mathbf {z} _ {k} ^ {w} + \\sigma_ {s | k} \\epsilon , \\epsilon \\sim \\mathcal {N} (\\mathbf {0}, \\mathbf {I}); \\tag {5}", + "image_path": "aa002c73c2fa804864352bd4b6d887dd42187781a63061ce22a3e6cfb947926b.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 347, + 579, + 545, + 605 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 579, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 347, + 579, + 545, + 605 + ], + "type": "interline_equation", + "content": "\\mathbf {z} _ {h} ^ {w} = \\alpha_ {h} \\hat {\\mathbf {x}} _ {\\boldsymbol {\\eta} _ {2}} \\left(\\mathbf {z} _ {s} ^ {w}\\right) + \\sigma_ {h} \\frac {\\mathbf {z} _ {s} ^ {w} - \\alpha_ {s} \\hat {\\mathbf {x}} _ {\\boldsymbol {\\eta} _ {2}} ^ {w} \\left(\\mathbf {z} _ {s} ^ {w}\\right)}{\\sigma_ {s}}, \\tag {6}", + "image_path": "70043f967f0581be37bb6621d0071d9248463b4384eb4da64c0aaccc21b9eef3.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 321, + 607, + 545, + 621 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 607, + 545, + 621 + ], + "spans": [ + { + "bbox": [ + 321, + 607, + 545, + 621 + ], + "type": "interline_equation", + "content": "w h e r e \\mathbf {z} _ {k} ^ {w} = \\left(\\alpha_ {k} / \\alpha_ {h}\\right) \\mathbf {z} _ {h} ^ {w} + \\sigma_ {k | h} \\epsilon , \\epsilon \\sim \\mathcal {N} (\\mathbf {0}, \\mathbf {I}). \\tag {7}", + "image_path": "2bce5c6ca71df58ab241a8dc71ac1c8fe8ebff6ae8e0a9af05f26db53f2a1418.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 629, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 546, + 714 + ], + "type": "text", + "content": "In the above equations, " + }, + { + "bbox": [ + 304, + 629, + 546, + 714 + ], + "type": "inline_equation", + "content": "h = t - 3 / N" + }, + { + "bbox": [ + 304, + 629, + 546, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 629, + 546, + 714 + ], + "type": "inline_equation", + "content": "k = t - 2 / N" + }, + { + "bbox": [ + 304, + 629, + 546, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 629, + 546, + 714 + ], + "type": "inline_equation", + "content": "s = t - 1 / N" + }, + { + "bbox": [ + 304, + 629, + 546, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 629, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\sigma_{a|b}^{2} = (1 - e^{\\lambda_{a} - \\lambda_{b}})\\sigma_{a}^{2}" + }, + { + "bbox": [ + 304, + 629, + 546, + 714 + ], + "type": "text", + "content": ". When " + }, + { + "bbox": [ + 304, + 629, + 546, + 714 + ], + "type": "inline_equation", + "content": "t = 1 / N" + }, + { + "bbox": [ + 304, + 629, + 546, + 714 + ], + "type": "text", + "content": ", we use deterministic update Eq. (2) to obtain " + }, + { + "bbox": [ + 304, + 629, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_0^w" + }, + { + "bbox": [ + 304, + 629, + 546, + 714 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 304, + 629, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{1 / N}^w" + }, + { + "bbox": [ + 304, + 629, + 546, + 714 + ], + "type": "text", + "content": ". We provide an illustration of the process in Fig. 5, where the number of denoising steps is 4. We note that compared to the deterministic sampler, performing stochastic sampling requires evaluating the model at slightly different time-steps," + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "14300" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 52, + 60, + 287, + 142 + ], + "blocks": [ + { + "bbox": [ + 52, + 60, + 287, + 142 + ], + "lines": [ + { + "bbox": [ + 52, + 60, + 287, + 142 + ], + "spans": [ + { + "bbox": [ + 52, + 60, + 287, + 142 + ], + "type": "image", + "image_path": "ba61fe9a0723a73bb3500aa6adbf6117174f16451908b263c1665bacb55124e2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 149, + 287, + 172 + ], + "lines": [ + { + "bbox": [ + 46, + 149, + 287, + 172 + ], + "spans": [ + { + "bbox": [ + 46, + 149, + 287, + 172 + ], + "type": "text", + "content": "Figure 5. Sampling procedures of the distilled model where the number of denoising steps is 4." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 178, + 287, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 178, + 287, + 215 + ], + "spans": [ + { + "bbox": [ + 46, + 178, + 287, + 215 + ], + "type": "text", + "content": "and would require small modifications to training algorithm for the edge cases. We provide the algorithm and more details in the supplementary material." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 225, + 128, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 225, + 128, + 239 + ], + "spans": [ + { + "bbox": [ + 47, + 225, + 128, + 239 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 245, + 288, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 245, + 288, + 328 + ], + "spans": [ + { + "bbox": [ + 46, + 245, + 288, + 328 + ], + "type": "text", + "content": "In this section, we evaluate the performance of our distillation approach on pixel-space diffusion models (i.e. DDPM [4]) and latent-space diffusion models (i.e. Stable Diffusion [28]). We further apply our approach to text-guided image editing and inpainting tasks. Experiments show that our approach is able to achieve competitive performance while using as few as 2-4 steps on all tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 336, + 263, + 349 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 336, + 263, + 349 + ], + "spans": [ + { + "bbox": [ + 47, + 336, + 263, + 349 + ], + "type": "text", + "content": "4.1. Distillation for pixel-space guided models" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 354, + 288, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 354, + 288, + 378 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 288, + 378 + ], + "type": "text", + "content": "In this experiment, we consider class-conditional diffusion models directly trained on the pixel-space [4,6,33]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 379, + 288, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 379, + 288, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 379, + 288, + 604 + ], + "type": "text", + "content": "Settings We focus on ImageNet 64x64 [30] and CIFAR-10 [12] as higher-resolution image generation in this scenario often relies on combining with other super-resolution techniques [5, 31]. We explore different ranges for the guidance weight and observe that all ranges work comparably and therefore use " + }, + { + "bbox": [ + 46, + 379, + 288, + 604 + ], + "type": "inline_equation", + "content": "[w_{min}, w_{max}] = [0, 4]" + }, + { + "bbox": [ + 46, + 379, + 288, + 604 + ], + "type": "text", + "content": " for the experiments. The baselines we consider include DDPM ancestral sampling [4] and DDIM [38]. The teacher model we use is a " + }, + { + "bbox": [ + 46, + 379, + 288, + 604 + ], + "type": "inline_equation", + "content": "1024 \\times 2" + }, + { + "bbox": [ + 46, + 379, + 288, + 604 + ], + "type": "text", + "content": "-step DDIM model, where the conditional and unconditional components both use 1024 DDIM denoising steps. To better understand how the guidance weight " + }, + { + "bbox": [ + 46, + 379, + 288, + 604 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 46, + 379, + 288, + 604 + ], + "type": "text", + "content": " should be incorporated, we also include models trained using a single fixed " + }, + { + "bbox": [ + 46, + 379, + 288, + 604 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 46, + 379, + 288, + 604 + ], + "type": "text", + "content": " as a baseline. We use the same pre-trained teacher model for all the methods for fair comparisons. Following [4, 6, 39], we use a U-Net [29, 39] architecture for the baselines, and the same U-Net backbone with the introduced " + }, + { + "bbox": [ + 46, + 379, + 288, + 604 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 46, + 379, + 288, + 604 + ], + "type": "text", + "content": "-embedding for our two-step student models (see Sec. 3). Following [33], we use a v-prediction model for both datasets." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "text", + "content": "Results We report the performance as evaluated in FID [3] and Inception scores (IS) [32] for all approaches on ImageNet 64x64 in Fig. 6 and Tab. 1 and provide extended ImageNet 64x64 and CIFAR-10 results in the supplement. We observe that our distilled model is able to match a teacher guided DDIM model with " + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "inline_equation", + "content": "1024 \\times 2" + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "text", + "content": " sampling steps using only 4-16 steps, achieving a speedup for up to " + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "inline_equation", + "content": "256 \\times" + }, + { + "bbox": [ + 46, + 605, + 288, + 714 + ], + "type": "text", + "content": ". We emphasize that, using our approach, a single distilled model is able to match the teacher performance on a wide range" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": "of guidance strengths. This has not been achieved by any previous methods." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 105, + 525, + 118 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 105, + 525, + 118 + ], + "spans": [ + { + "bbox": [ + 305, + 105, + 525, + 118 + ], + "type": "text", + "content": "4.2. Distillation for latent-space guided models" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 123, + 547, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 123, + 547, + 206 + ], + "spans": [ + { + "bbox": [ + 304, + 123, + 547, + 206 + ], + "type": "text", + "content": "After demonstrating the effectiveness of our method on pixel-space class-guided diffusion models in Sec. 4.1, we now expand its scope to latent-space diffusion models. In the following sections, we show the effectiveness of our approach on Latent Diffusion [28] on a variety of tasks, including class-conditional generation, text-to-image generation, image inpainting and text-guided style-transfer [20]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 207, + 547, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 207, + 547, + 278 + ], + "spans": [ + { + "bbox": [ + 304, + 207, + 547, + 278 + ], + "type": "text", + "content": "In the following experiments, we use the open-sourced latent-space diffusion models [28] as the teacher models. As v-prediction teacher model tends to perform better than " + }, + { + "bbox": [ + 304, + 207, + 547, + 278 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 304, + 207, + 547, + 278 + ], + "type": "text", + "content": "-prediction model, we fine-tune the open-sourced " + }, + { + "bbox": [ + 304, + 207, + 547, + 278 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 304, + 207, + 547, + 278 + ], + "type": "text", + "content": "-prediction models into v-prediction teacher models. We provide more details in the supplementary material." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 295, + 460, + 307 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 295, + 460, + 307 + ], + "spans": [ + { + "bbox": [ + 305, + 295, + 460, + 307 + ], + "type": "text", + "content": "4.2.1 Class-conditional generation" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 314, + 547, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 314, + 547, + 397 + ], + "spans": [ + { + "bbox": [ + 304, + 314, + 547, + 397 + ], + "type": "text", + "content": "In this section, we apply our method to a class-conditional latent diffusion model pre-trained on ImageNet " + }, + { + "bbox": [ + 304, + 314, + 547, + 397 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 304, + 314, + 547, + 397 + ], + "type": "text", + "content": ". We start from the DDIM teacher model with 512 sampling steps, and use the output as the target to train our distilled model. We use a batch size of 512 and uniformly sample the guidance strength " + }, + { + "bbox": [ + 304, + 314, + 547, + 397 + ], + "type": "inline_equation", + "content": "w \\in [w_{min} = 0, w_{max} = 14]" + }, + { + "bbox": [ + 304, + 314, + 547, + 397 + ], + "type": "text", + "content": " during training." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 398, + 547, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 398, + 547, + 518 + ], + "spans": [ + { + "bbox": [ + 304, + 398, + 547, + 518 + ], + "type": "text", + "content": "Results Empirically, we find that our distilled model is able to match the performance of the teacher model (originally trained on 1000 steps) in terms of FID scores while using only 2 or 4 sampling steps. We also achieve significantly better performance than DDIM when using 1-4 sampling steps (see Fig. 11). Qualitatively, we find that samples synthesized using a single denoising step still yield satisfying results, while the baseline fails to generate images with meaningful contents. We provide extra samples in the supplementary material." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 518, + 547, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 518, + 547, + 590 + ], + "spans": [ + { + "bbox": [ + 304, + 518, + 547, + 590 + ], + "type": "text", + "content": "Similar to the pixel-based results in Fig. 6, we also observe the trade-off between sampling quality and diversity as measured by FID and Inception Score for our distilled latent diffusion model. Following Kynkänniemi et al [13], we further compute improved precision and recall metrics for this experiment in the appendix." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 605, + 465, + 618 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 605, + 465, + 618 + ], + "spans": [ + { + "bbox": [ + 305, + 605, + 465, + 618 + ], + "type": "text", + "content": "4.2.2 Text-guided image generation" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 624, + 547, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 624, + 547, + 685 + ], + "spans": [ + { + "bbox": [ + 304, + 624, + 547, + 685 + ], + "type": "text", + "content": "In this section, we focus on the text-guided Stable Diffusion model pretrained on subsets " + }, + { + "bbox": [ + 304, + 624, + 547, + 685 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 304, + 624, + 547, + 685 + ], + "type": "text", + "content": " of LAION-5B [34] at a resolution " + }, + { + "bbox": [ + 304, + 624, + 547, + 685 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 304, + 624, + 547, + 685 + ], + "type": "text", + "content": ". We then follow our two-stage approach introduced in Sec. 3 and distill the guided model in 3000 gradient updates into a " + }, + { + "bbox": [ + 304, + 624, + 547, + 685 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 304, + 624, + 547, + 685 + ], + "type": "text", + "content": "-conditioned model using" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 306, + 693, + 547, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 693, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 693, + 547, + 713 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 306, + 693, + 547, + 713 + ], + "type": "text", + "content": "https://github.com/CompVis/stable-diffusion/blob/main/Stable_Diffusion_v1_model/Card.md" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14301" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 64, + 154, + 123 + ], + "blocks": [ + { + "bbox": [ + 62, + 64, + 154, + 123 + ], + "lines": [ + { + "bbox": [ + 62, + 64, + 154, + 123 + ], + "spans": [ + { + "bbox": [ + 62, + 64, + 154, + 123 + ], + "type": "image", + "image_path": "cbec320e32ef4c440a3a8913360b49b54d9efe22f3accccb03e69b5ce00c6d22.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 157, + 64, + 249, + 121 + ], + "blocks": [ + { + "bbox": [ + 157, + 64, + 249, + 121 + ], + "lines": [ + { + "bbox": [ + 157, + 64, + 249, + 121 + ], + "spans": [ + { + "bbox": [ + 157, + 64, + 249, + 121 + ], + "type": "image", + "image_path": "041b3f32f7b8297c6f864a31573be28d85a8c79f280ffd7c294753f35bfedc22.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 251, + 64, + 343, + 123 + ], + "blocks": [ + { + "bbox": [ + 251, + 64, + 343, + 123 + ], + "lines": [ + { + "bbox": [ + 251, + 64, + 343, + 123 + ], + "spans": [ + { + "bbox": [ + 251, + 64, + 343, + 123 + ], + "type": "image", + "image_path": "ea5e4d98537a0cffaa6d98c47b80f3469a742385be7a41f9e98d3c400a873292.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 345, + 64, + 438, + 124 + ], + "blocks": [ + { + "bbox": [ + 345, + 64, + 438, + 124 + ], + "lines": [ + { + "bbox": [ + 345, + 64, + 438, + 124 + ], + "spans": [ + { + "bbox": [ + 345, + 64, + 438, + 124 + ], + "type": "image", + "image_path": "fb707b4c85407ed7c68e73fdf99e27c94232faf70df8adc93f32a22f9ddc039b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 440, + 64, + 533, + 123 + ], + "blocks": [ + { + "bbox": [ + 440, + 64, + 533, + 123 + ], + "lines": [ + { + "bbox": [ + 440, + 64, + 533, + 123 + ], + "spans": [ + { + "bbox": [ + 440, + 64, + 533, + 123 + ], + "type": "image", + "image_path": "97ee8399aae499ac1037d73062cb681662c6d6d0066e44c1b5272ef2e23d7b13.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 62, + 125, + 154, + 179 + ], + "blocks": [ + { + "bbox": [ + 62, + 125, + 154, + 179 + ], + "lines": [ + { + "bbox": [ + 62, + 125, + 154, + 179 + ], + "spans": [ + { + "bbox": [ + 62, + 125, + 154, + 179 + ], + "type": "image", + "image_path": "d0dd693b2a12bb672df55802208ae68a6927d17ee04fb407466baebb21e29e64.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 102, + 181, + 124, + 189 + ], + "lines": [ + { + "bbox": [ + 102, + 181, + 124, + 189 + ], + "spans": [ + { + "bbox": [ + 102, + 181, + 124, + 189 + ], + "type": "inline_equation", + "content": "w = 0" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 156, + 125, + 249, + 179 + ], + "blocks": [ + { + "bbox": [ + 156, + 125, + 249, + 179 + ], + "lines": [ + { + "bbox": [ + 156, + 125, + 249, + 179 + ], + "spans": [ + { + "bbox": [ + 156, + 125, + 249, + 179 + ], + "type": "image", + "image_path": "02930878f9429bc9d7483765481e491c86abe4acda7a91cd5de85f01a0637430.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 195, + 180, + 223, + 189 + ], + "lines": [ + { + "bbox": [ + 195, + 180, + 223, + 189 + ], + "spans": [ + { + "bbox": [ + 195, + 180, + 223, + 189 + ], + "type": "inline_equation", + "content": "w = 0.3" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 251, + 125, + 343, + 178 + ], + "blocks": [ + { + "bbox": [ + 251, + 125, + 343, + 178 + ], + "lines": [ + { + "bbox": [ + 251, + 125, + 343, + 178 + ], + "spans": [ + { + "bbox": [ + 251, + 125, + 343, + 178 + ], + "type": "image", + "image_path": "a0f3e40986241e51c66d85b02b3f296aabecbbe7a66bfe69650e04a6d9e9cb90.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 293, + 180, + 315, + 189 + ], + "lines": [ + { + "bbox": [ + 293, + 180, + 315, + 189 + ], + "spans": [ + { + "bbox": [ + 293, + 180, + 315, + 189 + ], + "type": "inline_equation", + "content": "w = 1" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 345, + 125, + 438, + 178 + ], + "blocks": [ + { + "bbox": [ + 345, + 125, + 438, + 178 + ], + "lines": [ + { + "bbox": [ + 345, + 125, + 438, + 178 + ], + "spans": [ + { + "bbox": [ + 345, + 125, + 438, + 178 + ], + "type": "image", + "image_path": "3e78bed81ea1fb715c91732ccf22c11d97cf03d6ca52c8a36484e710a2610283.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 390, + 180, + 412, + 189 + ], + "lines": [ + { + "bbox": [ + 390, + 180, + 412, + 189 + ], + "spans": [ + { + "bbox": [ + 390, + 180, + 412, + 189 + ], + "type": "inline_equation", + "content": "w = 2" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 440, + 126, + 533, + 178 + ], + "blocks": [ + { + "bbox": [ + 440, + 126, + 533, + 178 + ], + "lines": [ + { + "bbox": [ + 440, + 126, + 533, + 178 + ], + "spans": [ + { + "bbox": [ + 440, + 126, + 533, + 178 + ], + "type": "image", + "image_path": "ece91057b2d35114a9a67470305a79bd85312c29203b98ca5fdee83b2f0f728a.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 481, + 180, + 503, + 189 + ], + "lines": [ + { + "bbox": [ + 481, + 180, + 503, + 189 + ], + "spans": [ + { + "bbox": [ + 481, + 180, + 503, + 189 + ], + "type": "inline_equation", + "content": "w = 4" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 49, + 247, + 210, + 328 + ], + "blocks": [ + { + "bbox": [ + 46, + 199, + 547, + 232 + ], + "lines": [ + { + "bbox": [ + 46, + 199, + 547, + 232 + ], + "spans": [ + { + "bbox": [ + 46, + 199, + 547, + 232 + ], + "type": "text", + "content": "Figure 6. ImageNet " + }, + { + "bbox": [ + 46, + 199, + 547, + 232 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 46, + 199, + 547, + 232 + ], + "type": "text", + "content": " sample quality evaluated by FID and IS scores. Our distilled model significantly outperforms the DDPM and DDIM baselines, and is able to match the performance of the teacher using as few as 8 to 16 steps. By varying " + }, + { + "bbox": [ + 46, + 199, + 547, + 232 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 46, + 199, + 547, + 232 + ], + "type": "text", + "content": ", a single distilled model is able to capture the trade-off between sample diversity and quality." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 74, + 238, + 186, + 246 + ], + "lines": [ + { + "bbox": [ + 74, + 238, + 186, + 246 + ], + "spans": [ + { + "bbox": [ + 74, + 238, + 186, + 246 + ], + "type": "text", + "content": "Distilled Text-to-Image samples (4 steps)" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 49, + 247, + 210, + 328 + ], + "lines": [ + { + "bbox": [ + 49, + 247, + 210, + 328 + ], + "spans": [ + { + "bbox": [ + 49, + 247, + 210, + 328 + ], + "type": "image", + "image_path": "98f799da38b80d6a8b79674b305ab5c500be4cb173e8fe57d005af43ba25b5d7.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 337, + 547, + 381 + ], + "lines": [ + { + "bbox": [ + 46, + 337, + 547, + 381 + ], + "spans": [ + { + "bbox": [ + 46, + 337, + 547, + 381 + ], + "type": "text", + "content": "Figure 7. Text-guided Stable Diffusion results. We distill the public Stable Diffusion model using the proposed pipeline, arriving at a model that achieves high sample quality using only four denoising steps (left). When sampling from the original model using four DDIM steps, the generated samples have clear artifacts (middle). When using eight DDIM steps, the results get better (right), but are still blurry and less consistent than the distilled results using fewer steps. More samples are provided in Fig. 4." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 216, + 247, + 377, + 327 + ], + "blocks": [ + { + "bbox": [ + 243, + 239, + 350, + 247 + ], + "lines": [ + { + "bbox": [ + 243, + 239, + 350, + 247 + ], + "spans": [ + { + "bbox": [ + 243, + 239, + 350, + 247 + ], + "type": "text", + "content": "Native Text-to-Image samples (4 steps)" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 216, + 247, + 377, + 327 + ], + "lines": [ + { + "bbox": [ + 216, + 247, + 377, + 327 + ], + "spans": [ + { + "bbox": [ + 216, + 247, + 377, + 327 + ], + "type": "image", + "image_path": "f613993d643d5fa0a970a4460f3695f53d71bb718f2447bd5cdc014f1821b63d.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 383, + 247, + 545, + 327 + ], + "blocks": [ + { + "bbox": [ + 410, + 238, + 518, + 246 + ], + "lines": [ + { + "bbox": [ + 410, + 238, + 518, + 246 + ], + "spans": [ + { + "bbox": [ + 410, + 238, + 518, + 246 + ], + "type": "text", + "content": "Native Text-to-Image samples (8 steps)" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 383, + 247, + 545, + 327 + ], + "lines": [ + { + "bbox": [ + 383, + 247, + 545, + 327 + ], + "spans": [ + { + "bbox": [ + 383, + 247, + 545, + 327 + ], + "type": "image", + "image_path": "ff071da0624ff43e38d0107366b09d0da90636bd4f95afe9eb97d1145ec672fb.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "bbox": [ + 46, + 391, + 289, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 391, + 289, + 510 + ], + "spans": [ + { + "bbox": [ + 46, + 391, + 289, + 510 + ], + "type": "inline_equation", + "content": "w \\in [w_{min} = 2, w_{max} = 14]" + }, + { + "bbox": [ + 46, + 391, + 289, + 510 + ], + "type": "text", + "content": ", and a batch size of 512. Although we can condition on a broader range of " + }, + { + "bbox": [ + 46, + 391, + 289, + 510 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 46, + 391, + 289, + 510 + ], + "type": "text", + "content": " for the distilled (student) model, the utility remains unclear as we typically do not exceed the normal guidance range when sampling with the teacher model. The final model is obtained by applying progressive distillation for 2000 training steps per stage, except when for the low-step regime of 1,2, and 4 steps, where we train for 20000 gradient updates. A detailed analysis of the convergence properties of this model in the supplement." + } + ] + } + ], + "index": 23 + }, + { + "type": "table", + "bbox": [ + 83, + 511, + 250, + 557 + ], + "blocks": [ + { + "bbox": [ + 83, + 511, + 250, + 557 + ], + "lines": [ + { + "bbox": [ + 83, + 511, + 250, + 557 + ], + "spans": [ + { + "bbox": [ + 83, + 511, + 250, + 557 + ], + "type": "table", + "html": "
Method2-step4-step8-step
DPM [16]98.9/0.2034.3/0.2931.7/0.32
DPM++ [18]98.8/0.2034.1/0.2925.6/0.32
Ours37.3/0.2726.0/0.3026.9/0.30
", + "image_path": "6ae865654f4f93277c93f96fc2e22483dd1dfa6812da7cb10baa77824e2305f4.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "table_body" + } + ], + "index": 24 + }, + { + "bbox": [ + 46, + 559, + 288, + 614 + ], + "lines": [ + { + "bbox": [ + 46, + 559, + 288, + 614 + ], + "spans": [ + { + "bbox": [ + 46, + 559, + 288, + 614 + ], + "type": "text", + "content": "Table 2. FID/CLIP scores on LAION 512X512 (" + }, + { + "bbox": [ + 46, + 559, + 288, + 614 + ], + "type": "inline_equation", + "content": "w = 8.0" + }, + { + "bbox": [ + 46, + 559, + 288, + 614 + ], + "type": "text", + "content": "). We point out that DPM and DPM++ use both the conditional and unconditional components for sampling. Depending on the implementation, this either requires higher peak memory or two times more sampling steps." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 618, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 288, + 713 + ], + "type": "text", + "content": "Results We present samples in Fig. 4. We evaluate the resulting model both qualitatively and quantitatively. For the latter analysis, we follow [31] and evaluate CLIP [25] and FID scores to assess text-image alignment and quality, respectively. We use the open-sourced ViT-g/14 [7] CLIP model for evaluation. The quantitative results in Fig. 10 show that our method can significantly increase the performance in both metrics over DDIM sampling on the base model for" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 391, + 547, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 391, + 547, + 631 + ], + "spans": [ + { + "bbox": [ + 304, + 391, + 547, + 631 + ], + "type": "text", + "content": "2 and 4 sampling steps. For 8 steps, these metrics do not show a significant difference. However, when considering the corresponding samples in Fig. 7 we can observe a stark difference in terms of visual image quality. In contrast to the 8-step DDIM samples from the original model, the distilled samples are sharper and more coherent. We hypothesize that FID and CLIP do not fully capture these differences in our evaluation setting on COCO2017 [14], where we used 5000 random captions from the validation set. We further compute the FID and CLIP scores for our distilled LAION " + }, + { + "bbox": [ + 304, + 391, + 547, + 631 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 304, + 391, + 547, + 631 + ], + "type": "text", + "content": " model and compare them with the DPM [16] and " + }, + { + "bbox": [ + 304, + 391, + 547, + 631 + ], + "type": "inline_equation", + "content": "\\mathrm{DPM}++" + }, + { + "bbox": [ + 304, + 391, + 547, + 631 + ], + "type": "text", + "content": " [18] solver in Tab. 2. We observe that our method is able to achieve significantly better performance when the denoising step is 2 or 4. Furthermore, we stress that stage-one of our method already decreases the number of function evaluations by a factor of 2, as we distill the classifier-free guidance step into a single model. Depending on the exact implementation (batched vs. sequential network evaluation), this either decreases peak memory or sampling time compared to existing solvers [16, 18, 38]." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 306, + 647, + 507, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 647, + 507, + 658 + ], + "spans": [ + { + "bbox": [ + 306, + 647, + 507, + 658 + ], + "type": "text", + "content": "4.2.3 Text-guided image-to-image translation" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": "In this section, we perform experiments on text-guided image-to-image translation with SDEdit [20] using our distilled model from Sec. 4.2.2. Following SDEdit [20], we perform stochastic encoding in the latent space, but instead" + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14302" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 49, + 57, + 545, + 108 + ], + "blocks": [ + { + "bbox": [ + 49, + 57, + 545, + 108 + ], + "lines": [ + { + "bbox": [ + 49, + 57, + 545, + 108 + ], + "spans": [ + { + "bbox": [ + 49, + 57, + 545, + 108 + ], + "type": "image", + "image_path": "fd4a576fa344ffbe1d42def9d273748724b60d646b3109add5004344f748410a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 64, + 110, + 85, + 118 + ], + "lines": [ + { + "bbox": [ + 64, + 110, + 85, + 118 + ], + "spans": [ + { + "bbox": [ + 64, + 110, + 85, + 118 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 104, + 110, + 145, + 118 + ], + "lines": [ + { + "bbox": [ + 104, + 110, + 145, + 118 + ], + "spans": [ + { + "bbox": [ + 104, + 110, + 145, + 118 + ], + "type": "text", + "content": "Watercolor" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 154, + 110, + 193, + 118 + ], + "lines": [ + { + "bbox": [ + 154, + 110, + 193, + 118 + ], + "spans": [ + { + "bbox": [ + 154, + 110, + 193, + 118 + ], + "type": "text", + "content": "Art station" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 201, + 110, + 244, + 118 + ], + "lines": [ + { + "bbox": [ + 201, + 110, + 244, + 118 + ], + "spans": [ + { + "bbox": [ + 201, + 110, + 244, + 118 + ], + "type": "text", + "content": "Photograph" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 287, + 110, + 307, + 118 + ], + "lines": [ + { + "bbox": [ + 287, + 110, + 307, + 118 + ], + "spans": [ + { + "bbox": [ + 287, + 110, + 307, + 118 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 372, + 110, + 413, + 118 + ], + "lines": [ + { + "bbox": [ + 372, + 110, + 413, + 118 + ], + "spans": [ + { + "bbox": [ + 372, + 110, + 413, + 118 + ], + "type": "text", + "content": "Watercolor" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 459, + 110, + 529, + 118 + ], + "lines": [ + { + "bbox": [ + 459, + 110, + 529, + 118 + ], + "spans": [ + { + "bbox": [ + 459, + 110, + 529, + 118 + ], + "type": "text", + "content": "Art station painting" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 50, + 120, + 544, + 186 + ], + "blocks": [ + { + "bbox": [ + 50, + 120, + 544, + 186 + ], + "lines": [ + { + "bbox": [ + 50, + 120, + 544, + 186 + ], + "spans": [ + { + "bbox": [ + 50, + 120, + 544, + 186 + ], + "type": "image", + "image_path": "5d49590896898c6051dd3950f862aec21dee3a99a02e14257edb9cfc27997da3.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 70, + 183, + 91, + 193 + ], + "lines": [ + { + "bbox": [ + 70, + 183, + 91, + 193 + ], + "spans": [ + { + "bbox": [ + 70, + 183, + 91, + 193 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 167, + 186, + 181, + 194 + ], + "lines": [ + { + "bbox": [ + 167, + 186, + 181, + 194 + ], + "spans": [ + { + "bbox": [ + 167, + 186, + 181, + 194 + ], + "type": "text", + "content": "Girl" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 50, + 194, + 544, + 261 + ], + "blocks": [ + { + "bbox": [ + 50, + 194, + 544, + 261 + ], + "lines": [ + { + "bbox": [ + 50, + 194, + 544, + 261 + ], + "spans": [ + { + "bbox": [ + 50, + 194, + 544, + 261 + ], + "type": "image", + "image_path": "649e891f3bcdf8b550572949be033f37aa578bc4ad79ac96b35516b60d6fa51a.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 59, + 267, + 104, + 277 + ], + "lines": [ + { + "bbox": [ + 59, + 267, + 104, + 277 + ], + "spans": [ + { + "bbox": [ + 59, + 267, + 104, + 277 + ], + "type": "text", + "content": "Oil painting" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 137, + 261, + 152, + 269 + ], + "lines": [ + { + "bbox": [ + 137, + 261, + 152, + 269 + ], + "spans": [ + { + "bbox": [ + 137, + 261, + 152, + 269 + ], + "type": "text", + "content": "Boy" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 130, + 269, + 156, + 277 + ], + "lines": [ + { + "bbox": [ + 130, + 269, + 156, + 277 + ], + "spans": [ + { + "bbox": [ + 130, + 269, + 156, + 277 + ], + "type": "text", + "content": "Anime" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 186, + 267, + 223, + 275 + ], + "lines": [ + { + "bbox": [ + 186, + 267, + 223, + 275 + ], + "spans": [ + { + "bbox": [ + 186, + 267, + 223, + 275 + ], + "type": "text", + "content": "3D render" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 254, + 267, + 279, + 276 + ], + "lines": [ + { + "bbox": [ + 254, + 267, + 279, + 276 + ], + "spans": [ + { + "bbox": [ + 254, + 267, + 279, + 276 + ], + "type": "text", + "content": "Anime" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 308, + 267, + 346, + 276 + ], + "lines": [ + { + "bbox": [ + 308, + 267, + 346, + 276 + ], + "spans": [ + { + "bbox": [ + 308, + 267, + 346, + 276 + ], + "type": "text", + "content": "3D render" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 381, + 260, + 400, + 268 + ], + "lines": [ + { + "bbox": [ + 381, + 260, + 400, + 268 + ], + "spans": [ + { + "bbox": [ + 381, + 260, + 400, + 268 + ], + "type": "text", + "content": "Male" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 371, + 268, + 407, + 277 + ], + "lines": [ + { + "bbox": [ + 371, + 268, + 407, + 277 + ], + "spans": [ + { + "bbox": [ + 371, + 268, + 407, + 277 + ], + "type": "text", + "content": "Van Gogh" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 430, + 267, + 471, + 276 + ], + "lines": [ + { + "bbox": [ + 430, + 267, + 471, + 276 + ], + "spans": [ + { + "bbox": [ + 430, + 267, + 471, + 276 + ], + "type": "text", + "content": "Watercolor" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 493, + 267, + 533, + 276 + ], + "lines": [ + { + "bbox": [ + 493, + 267, + 533, + 276 + ], + "spans": [ + { + "bbox": [ + 493, + 267, + 533, + 276 + ], + "type": "text", + "content": "Art station" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 50, + 312, + 544, + 384 + ], + "blocks": [ + { + "bbox": [ + 47, + 287, + 546, + 309 + ], + "lines": [ + { + "bbox": [ + 47, + 287, + 546, + 309 + ], + "spans": [ + { + "bbox": [ + 47, + 287, + 546, + 309 + ], + "type": "text", + "content": "Figure 8. Text-guided image-to-image translation [20] with the distilled Stable Diffusion model (3 denoising steps). We observe that our model is able to generate high-quality and faithful outputs using only 3 denoising steps." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 50, + 312, + 544, + 384 + ], + "lines": [ + { + "bbox": [ + 50, + 312, + 544, + 384 + ], + "spans": [ + { + "bbox": [ + 50, + 312, + 544, + 384 + ], + "type": "image", + "image_path": "e6a319a1aee460d2faec76fc15904b4a758b5e52c5ae49bc288a0f389dc17f09.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 78, + 385, + 93, + 392 + ], + "lines": [ + { + "bbox": [ + 78, + 385, + 93, + 392 + ], + "spans": [ + { + "bbox": [ + 78, + 385, + 93, + 392 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 50, + 393, + 544, + 446 + ], + "blocks": [ + { + "bbox": [ + 50, + 393, + 544, + 446 + ], + "lines": [ + { + "bbox": [ + 50, + 393, + 544, + 446 + ], + "spans": [ + { + "bbox": [ + 50, + 393, + 544, + 446 + ], + "type": "image", + "image_path": "de5653f9a17f5cb852113031736c4f643bd0f5ace4f4dc79472ff83ba3e7eaf2.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 50, + 446, + 165, + 511 + ], + "blocks": [ + { + "bbox": [ + 50, + 446, + 165, + 511 + ], + "lines": [ + { + "bbox": [ + 50, + 446, + 165, + 511 + ], + "spans": [ + { + "bbox": [ + 50, + 446, + 165, + 511 + ], + "type": "image", + "image_path": "b8250917f941b6991c587ce05a0b56d3e8fcbfb4bd96b3861a6272d3865df53c.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 521, + 287, + 597 + ], + "lines": [ + { + "bbox": [ + 46, + 521, + 287, + 597 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 287, + 597 + ], + "type": "text", + "content": "Figure 10. FID and CLIP ViT-g/14 score for text-to-image generation at " + }, + { + "bbox": [ + 46, + 521, + 287, + 597 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 46, + 521, + 287, + 597 + ], + "type": "text", + "content": " px using the distilled Stable Diffusion model. The results are evaluated on 5000 captions from the COCO2017 [14] validation set. Our distilled latent diffusion model is able to generate high-quality image samples using significantly less sampling steps than the original model while achieving similar or better FID and CLIP scores, especially in the low-step regime." + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 168, + 446, + 283, + 511 + ], + "blocks": [ + { + "bbox": [ + 168, + 446, + 283, + 511 + ], + "lines": [ + { + "bbox": [ + 168, + 446, + 283, + 511 + ], + "spans": [ + { + "bbox": [ + 168, + 446, + 283, + 511 + ], + "type": "image", + "image_path": "f5ef46df91d8617dd752c0e3daeb21932ce5f8e5208038bcd5f76011355e256d.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 713 + ], + "type": "text", + "content": "use the deterministic sampler of the distilled model to perform deterministic decoding. We consider input image and text of various kinds and provide qualitative results in Fig. 8. We observe that our distilled model generates high-quality style-transfer results using as few as 3 denoising steps. We provide more analysis on the trade-off between sample quality, controllability and efficiency in the supplement." + } + ] + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 307, + 446, + 421, + 511 + ], + "blocks": [ + { + "bbox": [ + 307, + 446, + 421, + 511 + ], + "lines": [ + { + "bbox": [ + 307, + 446, + 421, + 511 + ], + "spans": [ + { + "bbox": [ + 307, + 446, + 421, + 511 + ], + "type": "image", + "image_path": "280b6991cfac975dd2c8905ef3a699ba24d9805d3babb9bdcb5e47cffbb77c31.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 521, + 546, + 586 + ], + "lines": [ + { + "bbox": [ + 304, + 521, + 546, + 586 + ], + "spans": [ + { + "bbox": [ + 304, + 521, + 546, + 586 + ], + "type": "text", + "content": "Figure 11. FID and Inception Score for class-conditional image generation on ImageNet " + }, + { + "bbox": [ + 304, + 521, + 546, + 586 + ], + "type": "inline_equation", + "content": "(256 \\times 256)" + }, + { + "bbox": [ + 304, + 521, + 546, + 586 + ], + "type": "text", + "content": " with distilled latent diffusion. The results are evaluated on 5000 samples. Our distilled latent diffusion model is able to generate high-quality image samples using significantly less sampling steps (up to a factor of 16) than the original model while achieving similar or better FID scores." + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 425, + 446, + 543, + 511 + ], + "blocks": [ + { + "bbox": [ + 425, + 446, + 543, + 511 + ], + "lines": [ + { + "bbox": [ + 425, + 446, + 543, + 511 + ], + "spans": [ + { + "bbox": [ + 425, + 446, + 543, + 511 + ], + "type": "image", + "image_path": "e53a1bc88b4ad86e626365445a352c30deccb744a0b679bd88603736daed3adb.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "bbox": [ + 306, + 608, + 411, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 608, + 411, + 620 + ], + "spans": [ + { + "bbox": [ + 306, + 608, + 411, + 620 + ], + "type": "text", + "content": "4.2.4 Image inpainting" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 304, + 626, + 545, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 626, + 545, + 686 + ], + "spans": [ + { + "bbox": [ + 304, + 626, + 545, + 686 + ], + "type": "text", + "content": "In this section, we apply our approach to a pre-trained image inpainting latent diffusion model. We use the open-source Stable Diffusion Inpainting\\* image-inpainting model. This model is a fine-tuned version of the pure text-to-image Stable Diffusion model from above, where additional input channels" + } + ] + } + ], + "index": 34 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 306, + 693, + 547, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 693, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 693, + 547, + 713 + ], + "type": "text", + "content": "https://huggingface.co/runwayml/stable-diffusion-inpainting" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14303" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 70, + 286, + 127 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 286, + 127 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 286, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 286, + 127 + ], + "type": "image", + "image_path": "2da267968cdc011b6132827c27e3bf9ff0ba5b7f4a3ca5f167fb7c41033eeeca.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 135, + 288, + 191 + ], + "lines": [ + { + "bbox": [ + 47, + 135, + 288, + 191 + ], + "spans": [ + { + "bbox": [ + 47, + 135, + 288, + 191 + ], + "type": "text", + "content": "Figure 12. Style transfer comparison on ImageNet 64x64 for pixel-space models. For our approach, we use a distilled encoder and decoder. For the baseline, we encode and decode using DDIM. We use " + }, + { + "bbox": [ + 47, + 135, + 288, + 191 + ], + "type": "inline_equation", + "content": "w = 0" + }, + { + "bbox": [ + 47, + 135, + 288, + 191 + ], + "type": "text", + "content": " and 16 sampling steps for both the encoder and decoder. We observe that our method achieves more realistic outputs." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 213, + 249, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 213, + 249, + 224 + ], + "spans": [ + { + "bbox": [ + 47, + 213, + 249, + 224 + ], + "type": "text", + "content": "were added to process masks and masked images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 224, + 288, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 224, + 288, + 320 + ], + "spans": [ + { + "bbox": [ + 46, + 224, + 288, + 320 + ], + "type": "text", + "content": "We use the same distillation algorithm as used in the previous section. For training, we start from the v-prediction teacher model sampled with 512 DDIM steps, and use the output as the target to optimize our student model. We present qualitative results in Fig. 9, demonstrating the potential of our method for fast, real-world image editing applications. For additional training details and a quantitative evaluation, see the supplementary." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 328, + 237, + 341 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 328, + 237, + 341 + ], + "spans": [ + { + "bbox": [ + 47, + 328, + 237, + 341 + ], + "type": "text", + "content": "4.3. Progressive distillation for encoding" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 347, + 288, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 347, + 288, + 525 + ], + "spans": [ + { + "bbox": [ + 46, + 347, + 288, + 525 + ], + "type": "text", + "content": "In this experiment, we explore distilling the encoding process for the teacher model and perform experiments on style-transfer in a setting similar to [41]. We focus on pixel-space diffusion models pre-trained on ImageNet " + }, + { + "bbox": [ + 46, + 347, + 288, + 525 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 46, + 347, + 288, + 525 + ], + "type": "text", + "content": ". Specifically, to perform style-transfer between two domains " + }, + { + "bbox": [ + 46, + 347, + 288, + 525 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 46, + 347, + 288, + 525 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 347, + 288, + 525 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 46, + 347, + 288, + 525 + ], + "type": "text", + "content": ", we encode the image from domain-" + }, + { + "bbox": [ + 46, + 347, + 288, + 525 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 46, + 347, + 288, + 525 + ], + "type": "text", + "content": " using a diffusion model trained on domain-" + }, + { + "bbox": [ + 46, + 347, + 288, + 525 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 46, + 347, + 288, + 525 + ], + "type": "text", + "content": ", and then decode with a diffusion model trained on domain-" + }, + { + "bbox": [ + 46, + 347, + 288, + 525 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 46, + 347, + 288, + 525 + ], + "type": "text", + "content": ". As the encoding process can be understood as reversing the DDIM sampling process, we perform distillation for both the encoder and decoder with classifier-free guidance, and compare with a DDIM encoder and decoder in Fig. 12. We also explore how modifying the guidance strength " + }, + { + "bbox": [ + 46, + 347, + 288, + 525 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 46, + 347, + 288, + 525 + ], + "type": "text", + "content": " can impact the performance and provide more details in the supplementary material." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 537, + 134, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 537, + 134, + 550 + ], + "spans": [ + { + "bbox": [ + 47, + 537, + 134, + 550 + ], + "type": "text", + "content": "5. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 558, + 288, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 558, + 288, + 688 + ], + "spans": [ + { + "bbox": [ + 46, + 558, + 288, + 688 + ], + "type": "text", + "content": "Our approach is related to existing works on improving the sampling speed of diffusion models [4, 37, 40]. For instance, denoising diffusion implicit models (DDIM [38]), probability flow sampler [40], fast SDE integrators [8] have been proposed to improve the sampling speed of diffusion models. Other works develop higher-order solvers [17], exponential integrators [15], and dynamic programming based approach [43] to accelerating sampling speed. However, none of these approaches have achieved comparable performance as our method on distilling classifier-free guided diffusion models." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 689, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 713 + ], + "type": "text", + "content": "Existing distillation-based methods for diffusion models are mainly designed for non-classifier-free guided diffusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 72, + 547, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 406 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 406 + ], + "type": "text", + "content": "models. For instance, [19] proposes to predict the data from noise in one single step by inverting a deterministic encoding of DDIM, [2] proposes to achieve faster sampling speed by distilling higher order solvers into an additional prediction head of the neural network backbone [2]. Progressive distillation [33] is perhaps the most relevant work. Specifically, it proposes to progressively distill a pre-trained diffusion model into a fewer-step student model with the same model architecture. However, none of these approaches are directly applicable or have been applied to classifier-free guided diffusion models. They are also unable to capture a range of different guidance strengths using one single distilled model. On the contrary, by incorporating the guidance strength into the model architecture and training the model using a two-stage procedure, our approach is able to match the performance of the teacher model on a wide range of guidance strength using one single model. Using our method, one single model can capture the trade-off between sample quality and diversity, enabling the real-world application of classifier-free guided diffusion models, where the guidance strength is often specified by users. Moreover, none of the above distillation approaches have been applied to or shown effectiveness for latent-space text-to-image models. Finally, most fast sampling approaches [33, 38, 40] only consider using deterministic sampling schemes to improve the sampling speed. In this work, we further develop an effective stochastic sampling approach to sample from the distilled models." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 420, + 378, + 432 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 420, + 378, + 432 + ], + "spans": [ + { + "bbox": [ + 306, + 420, + 378, + 432 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 441, + 547, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 441, + 547, + 645 + ], + "spans": [ + { + "bbox": [ + 304, + 441, + 547, + 645 + ], + "type": "text", + "content": "In this paper, we propose a distillation approach for guided diffusion models [6]. Our two-stage approach allows us to significantly speed up popular but relatively inefficient guided diffusion models. We show that our approach can reduce the inference cost of classifier-free guided pixel-space and latent-space diffusion models by at least an order of magnitude. Empirically, we show that our approach is able to produce visually appealing results with only 2 steps, achieving a comparable FID score to the teacher with as few as 4 to 8 steps. We further demonstrate practical applications of our distillation approach to text-guided image-to-image translation and inpainting tasks. We hope that by significantly reducing the inference cost of classifier-free guided diffusion models, our method will promote creative applications as well as the wider adoption of image generation systems. In the future work, we aim to further improve the performance in the two and one sampling step regimes." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 657, + 408, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 657, + 408, + 670 + ], + "spans": [ + { + "bbox": [ + 306, + 657, + 408, + 670 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 547, + 713 + ], + "type": "text", + "content": "We thank the anonymous reviewers for their insightful discussions and feedback. All experiments on Stable Diffusion are supported by Stability AI." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14304" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 107, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 123 + ], + "type": "text", + "content": "[1] Prafulla Dhariwal and Alex Nichol. Diffusion models beat GANs on image synthesis. arXiv preprint arXiv:2105.05233, 2021. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 125, + 288, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 125, + 288, + 158 + ], + "spans": [ + { + "bbox": [ + 53, + 125, + 288, + 158 + ], + "type": "text", + "content": "[2] Tim Dockhorn, Arash Vahdat, and Karsten Kreis. Genie: Higher-order denoising diffusion solvers. In Advances in Neural Information Processing Systems, 2022. 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 160, + 288, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 160, + 288, + 213 + ], + "spans": [ + { + "bbox": [ + 53, + 160, + 288, + 213 + ], + "type": "text", + "content": "[3] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. GANs trained by a two time-scale update rule converge to a local Nash equilibrium. In Advances in Neural Information Processing Systems, pages 6626-6637, 2017. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 216, + 288, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 216, + 288, + 249 + ], + "spans": [ + { + "bbox": [ + 53, + 216, + 288, + 249 + ], + "type": "text", + "content": "[4] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems, pages 6840-6851, 2020. 2, 3, 5, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 251, + 288, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 251, + 288, + 293 + ], + "spans": [ + { + "bbox": [ + 53, + 251, + 288, + 293 + ], + "type": "text", + "content": "[5] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. arXiv preprint arXiv:2106.15282, 2021. 3, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 296, + 288, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 296, + 288, + 327 + ], + "spans": [ + { + "bbox": [ + 53, + 296, + 288, + 327 + ], + "type": "text", + "content": "[6] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 2, 3, 5, 8, 11" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 329, + 288, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 329, + 288, + 384 + ], + "spans": [ + { + "bbox": [ + 53, + 329, + 288, + 384 + ], + "type": "text", + "content": "[7] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, July 2021. If you use this software, please cite it as below. 6, 24" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 386, + 288, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 386, + 288, + 430 + ], + "spans": [ + { + "bbox": [ + 53, + 386, + 288, + 430 + ], + "type": "text", + "content": "[8] Alexia Jolicoeur-Martineau, Ke Li, Rémi Piché-Taillefer, Tal Kachman, and Ioannis Mitliagkas. Gotta go fast when generating data with score-based models. arXiv preprint arXiv:2105.14080, 2021. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 431, + 288, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 431, + 288, + 464 + ], + "spans": [ + { + "bbox": [ + 53, + 431, + 288, + 464 + ], + "type": "text", + "content": "[9] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. arXiv preprint arXiv:2206.00364, 2022. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 465, + 288, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 465, + 288, + 498 + ], + "spans": [ + { + "bbox": [ + 48, + 465, + 288, + 498 + ], + "type": "text", + "content": "[10] Diederik P Kingma, Tim Salimans, Ben Poole, and Jonathan Ho. Variational diffusion models. arXiv preprint arXiv:2107.00630, 2021. 2, 3, 11, 22, 23, 25" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 499, + 288, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 499, + 288, + 542 + ], + "spans": [ + { + "bbox": [ + 48, + 499, + 288, + 542 + ], + "type": "text", + "content": "[11] Zhifeng Kong, Wei Ping, Jiaji Huang, Kexin Zhao, and Bryan Catanzaro. DiffWave: A Versatile Diffusion Model for Audio Synthesis. International Conference on Learning Representations, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 544, + 288, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 544, + 288, + 567 + ], + "spans": [ + { + "bbox": [ + 48, + 544, + 288, + 567 + ], + "type": "text", + "content": "[12] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 568, + 288, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 568, + 288, + 612 + ], + "spans": [ + { + "bbox": [ + 48, + 568, + 288, + 612 + ], + "type": "text", + "content": "[13] Tuomas Kynkänniemi, Tero Karras, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Improved precision and recall metric for assessing generative models. Advances in Neural Information Processing Systems, 32, 2019. 5, 24" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 613, + 288, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 288, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 288, + 667 + ], + "type": "text", + "content": "[14] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740-755. Springer, 2014. 6, 7, 24" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "type": "text", + "content": "[15] Luping Liu, Yi Ren, Zhijie Lin, and Zhou Zhao. Pseudo numerical methods for diffusion models on manifolds. In International Conference on Learning Representations, 2022. 8" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 115 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 115 + ], + "type": "text", + "content": "[16] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps, 2022. 6, 23, 24" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 118, + 547, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 547, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 547, + 161 + ], + "type": "text", + "content": "[17] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. arXiv:2206.00927, 2022. 8, 11" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 163, + 547, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 547, + 206 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 547, + 206 + ], + "type": "text", + "content": "[18] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models, 2022. 6, 11, 13, 14, 23, 24, 28" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 209, + 547, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 209, + 547, + 241 + ], + "spans": [ + { + "bbox": [ + 307, + 209, + 547, + 241 + ], + "type": "text", + "content": "[19] Eric Luhman and Troy Luhman. Knowledge distillation in iterative generative models for improved sampling speed. arXiv preprint arXiv:2101.02388, 2021. 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 243, + 547, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 243, + 547, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 243, + 547, + 285 + ], + "type": "text", + "content": "[20] Chenlin Meng, Yang Song, Jiaming Song, Jiajun Wu, JunYan Zhu, and Stefano Ermon. Sdedit: Image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 2, 5, 6, 7, 25, 26, 27" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 287, + 547, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 287, + 547, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 287, + 547, + 319 + ], + "type": "text", + "content": "[21] Gautam Mittal, Jesse Engel, Curtis Hawthorne, and Ian Simon. Symbolic music generation with diffusion models. arXiv preprint arXiv:2103.16091, 2021. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 322, + 547, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 322, + 547, + 354 + ], + "spans": [ + { + "bbox": [ + 307, + 322, + 547, + 354 + ], + "type": "text", + "content": "[22] Alex Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. International Conference on Machine Learning, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 355, + 547, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 355, + 547, + 409 + ], + "spans": [ + { + "bbox": [ + 307, + 355, + 547, + 409 + ], + "type": "text", + "content": "[23] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 411, + 547, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 411, + 547, + 466 + ], + "spans": [ + { + "bbox": [ + 307, + 411, + 547, + 466 + ], + "type": "text", + "content": "[24] Konpat Preechakul, Nattanat Chathee, Suttisak Wizadwongsa, and Supasorn Suwajanakorn. Diffusion autoencoders: Toward a meaningful and decodable representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10619-10629, 2022. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 468, + 547, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 468, + 547, + 533 + ], + "spans": [ + { + "bbox": [ + 307, + 468, + 547, + 533 + ], + "type": "text", + "content": "[25] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763. PMLR, 2021. 6, 24" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 535, + 547, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 535, + 547, + 577 + ], + "spans": [ + { + "bbox": [ + 307, + 535, + 547, + 577 + ], + "type": "text", + "content": "[26] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2, 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 580, + 547, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 580, + 547, + 632 + ], + "spans": [ + { + "bbox": [ + 307, + 580, + 547, + 632 + ], + "type": "text", + "content": "[27] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 635, + 547, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 635, + 547, + 689 + ], + "spans": [ + { + "bbox": [ + 307, + 635, + 547, + 689 + ], + "type": "text", + "content": "[28] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 2, 3, 5" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 692, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 692, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 692, + 547, + 713 + ], + "type": "text", + "content": "[29] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation." + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14305" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 66, + 72, + 288, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 288, + 105 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 288, + 105 + ], + "type": "text", + "content": "In International Conference on Medical image computing and computer-assisted intervention, pages 234-241. Springer, 2015. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 107, + 288, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 107, + 288, + 162 + ], + "spans": [ + { + "bbox": [ + 48, + 107, + 288, + 162 + ], + "type": "text", + "content": "[30] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. ImageNet large scale visual recognition challenge. International Journal of Computer Vision, 115(3):211-252, 2015. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 288, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 288, + 228 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 288, + 228 + ], + "type": "text", + "content": "[31] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models with deep language understanding. arXiv preprint arXiv:2205.11487, 2022. 2, 3, 5, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 230, + 288, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 230, + 288, + 274 + ], + "spans": [ + { + "bbox": [ + 48, + 230, + 288, + 274 + ], + "type": "text", + "content": "[32] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training GANs. In Advances in Neural Information Processing Systems, pages 2234-2242, 2016. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 276, + 288, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 276, + 288, + 308 + ], + "spans": [ + { + "bbox": [ + 48, + 276, + 288, + 308 + ], + "type": "text", + "content": "[33] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. arXiv preprint arXiv:2202.00512, 2022. 2, 3, 4, 5, 8, 11, 19, 22, 23, 25" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 309, + 288, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 309, + 288, + 374 + ], + "spans": [ + { + "bbox": [ + 48, + 309, + 288, + 374 + ], + "type": "text", + "content": "[34] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. arXiv preprint arXiv:2210.08402, 2022. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 376, + 288, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 376, + 288, + 420 + ], + "spans": [ + { + "bbox": [ + 48, + 376, + 288, + 420 + ], + "type": "text", + "content": "[35] Abhishek Sinha, Jiaming Song, Chenlin Meng, and Stefano Ermon. D2c: Diffusion-decoding models for few-shot conditional generation. Advances in Neural Information Processing Systems, 34:12533-12548, 2021. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 422, + 288, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 422, + 288, + 465 + ], + "spans": [ + { + "bbox": [ + 48, + 422, + 288, + 465 + ], + "type": "text", + "content": "[36] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, pages 2256-2265, 2015. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 467, + 288, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 467, + 288, + 510 + ], + "spans": [ + { + "bbox": [ + 48, + 467, + 288, + 510 + ], + "type": "text", + "content": "[37] Jascha Sohl-Dickstein, Eric A Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. arXiv preprint arXiv:1503.03585, March 2015. 2, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 511, + 288, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 511, + 288, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 511, + 288, + 555 + ], + "type": "text", + "content": "[38] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. International Conference on Learning Representations, 2021. 2, 3, 4, 5, 6, 8, 11, 13, 14, 16, 22, 24, 28" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 556, + 288, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 556, + 288, + 600 + ], + "spans": [ + { + "bbox": [ + 48, + 556, + 288, + 600 + ], + "type": "text", + "content": "[39] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. In Advances in Neural Information Processing Systems, pages 11895-11907, 2019. 2, 3, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 601, + 288, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 288, + 655 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 288, + 655 + ], + "type": "text", + "content": "[40] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. International Conference on Learning Representations, 2021. 2, 3, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 288, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 288, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 288, + 690 + ], + "type": "text", + "content": "[41] Xuan Su, Jiaming Song, Chenlin Meng, and Stefano Ermon. Dual diffusion implicit bridges for image-to-image translation. arXiv preprint arXiv:2203.08382, 2022. 2, 8, 12" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 692, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 692, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 692, + 288, + 713 + ], + "type": "text", + "content": "[42] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Arsenii Ashukha, Aleksei Silvestrov," + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 547, + 273 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 326, + 72, + 547, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 72, + 547, + 127 + ], + "spans": [ + { + "bbox": [ + 326, + 72, + 547, + 127 + ], + "type": "text", + "content": "Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with fourier convolutions. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2149–2159, 2022. 25" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 129, + 547, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 547, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 547, + 173 + ], + "type": "text", + "content": "[43] Daniel Watson, William Chan, Jonathan Ho, and Mohammad Norouzi. Learning fast samplers for diffusion models by differentiating through sample quality. In International Conference on Learning Representations, 2022. 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 175, + 547, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 547, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 547, + 217 + ], + "type": "text", + "content": "[44] Minkai Xu, Lantao Yu, Yang Song, Chence Shi, Stefano Ermon, and Jian Tang. Geodiff: A geometric diffusion model for molecular conformation generation. arXiv preprint arXiv:2203.02923, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 219, + 547, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 219, + 547, + 273 + ], + "spans": [ + { + "bbox": [ + 307, + 219, + 547, + 273 + ], + "type": "text", + "content": "[45] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2022. 24" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "14306" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file